diff --git a/.gitattributes b/.gitattributes index b556a2be7c8a1f72f86c097ff99db6a57a55c051..76f2bc8425827fc9fd183655498b99bde641d504 100644 --- a/.gitattributes +++ b/.gitattributes @@ -860,3 +860,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text 2025/Expressive[[:space:]]Talking[[:space:]]Human[[:space:]]from[[:space:]]Single-Image[[:space:]]with[[:space:]]Imperfect[[:space:]]Priors/d3058941-5a3c-47b1-a217-ebdb5b94b775_origin.pdf filter=lfs diff=lfs merge=lfs -text 2025/Extending[[:space:]]Foundational[[:space:]]Monocular[[:space:]]Depth[[:space:]]Estimators[[:space:]]to[[:space:]]Fisheye[[:space:]]Cameras[[:space:]]with[[:space:]]Calibration[[:space:]]Tokens/ff8cd9c3-7a4b-4914-bf82-9a51ee69f9ce_origin.pdf filter=lfs diff=lfs merge=lfs -text 2025/External[[:space:]]Knowledge[[:space:]]Injection[[:space:]]for[[:space:]]CLIP-Based[[:space:]]Class-Incremental[[:space:]]Learning/59db92d7-ff41-4079-a08f-c630a2445462_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Extrapolated[[:space:]]Urban[[:space:]]View[[:space:]]Synthesis[[:space:]]Benchmark/e65fa038-3ae6-4620-9f47-ad1a63db0abf_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/F-Bench_[[:space:]]Rethinking[[:space:]]Human[[:space:]]Preference[[:space:]]Evaluation[[:space:]]Metrics[[:space:]]for[[:space:]]Benchmarking[[:space:]]Face[[:space:]]Generation,[[:space:]]Customization,[[:space:]]and[[:space:]]Restoration/86d5740f-dff5-47ab-aa67-761b97593a44_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/FALCON_[[:space:]]Resolving[[:space:]]Visual[[:space:]]Redundancy[[:space:]]and[[:space:]]Fragmentation[[:space:]]in[[:space:]]High-resolution[[:space:]]Multimodal[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]via[[:space:]]Visual[[:space:]]Registers/3c258128-6a38-46a9-bf2a-aabe422e5840_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/FA_[[:space:]]Forced[[:space:]]Prompt[[:space:]]Learning[[:space:]]of[[:space:]]Vision-Language[[:space:]]Models[[:space:]]for[[:space:]]Out-of-Distribution[[:space:]]Detection/8ae5a438-b943-4e71-b751-475d9f24c3b7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/FB-Diff_[[:space:]]Fourier[[:space:]]Basis-guided[[:space:]]Diffusion[[:space:]]for[[:space:]]Temporal[[:space:]]Interpolation[[:space:]]of[[:space:]]4D[[:space:]]Medical[[:space:]]Imaging/c8ad7765-483e-400a-90bd-72090b0597d0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/FDPT_[[:space:]]Federated[[:space:]]Discrete[[:space:]]Prompt[[:space:]]Tuning[[:space:]]for[[:space:]]Black-Box[[:space:]]Visual-Language[[:space:]]Models/5d5a1b2f-9957-4fb4-b461-7bf3e1ebccfb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/FE-CLIP_[[:space:]]Frequency[[:space:]]Enhanced[[:space:]]CLIP[[:space:]]Model[[:space:]]for[[:space:]]Zero-Shot[[:space:]]Anomaly[[:space:]]Detection[[:space:]]and[[:space:]]Segmentation/e20fdad0-0c02-434f-a4f5-0e729b38aab0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/FED-PsyAU_[[:space:]]Privacy-Preserving[[:space:]]Micro-Expression[[:space:]]Recognition[[:space:]]via[[:space:]]Psychological[[:space:]]AU[[:space:]]Coordination[[:space:]]and[[:space:]]Dynamic[[:space:]]Facial[[:space:]]Motion[[:space:]]Modeling/e27ad4a5-3106-4216-bcea-a95709cd5f65_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/FEVER-OOD_[[:space:]]Free[[:space:]]Energy[[:space:]]Vulnerability[[:space:]]Elimination[[:space:]]for[[:space:]]Robust[[:space:]]Out-of-Distribution[[:space:]]Detection/0c16965c-d6fa-4547-a6ba-ab34512ce472_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/FG-OrIU_[[:space:]]Towards[[:space:]]Better[[:space:]]Forgetting[[:space:]]via[[:space:]]Feature-Gradient[[:space:]]Orthogonality[[:space:]]for[[:space:]]Incremental[[:space:]]Unlearning/4a750062-438a-4c8c-b4db-9a6576c3de7a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/FICGen_[[:space:]]Frequency-Inspired[[:space:]]Contextual[[:space:]]Disentanglement[[:space:]]for[[:space:]]Layout-driven[[:space:]]Degraded[[:space:]]Image[[:space:]]Generation/24509018-dca5-4420-ad1a-dd11a5974d9d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/FIND_[[:space:]]Few-Shot[[:space:]]Anomaly[[:space:]]Inspection[[:space:]]with[[:space:]]Normal-Only[[:space:]]Multi-Modal[[:space:]]Data/94dd2e05-5a4d-4859-bd44-9496d30c1756_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/FLOAT_[[:space:]]Generative[[:space:]]Motion[[:space:]]Latent[[:space:]]Flow[[:space:]]Matching[[:space:]]for[[:space:]]Audio-driven[[:space:]]Talking[[:space:]]Portrait/5eb040a4-3c63-4680-a42f-da4ce86274a7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/FLOSS_[[:space:]]Free[[:space:]]Lunch[[:space:]]in[[:space:]]Open-vocabulary[[:space:]]Semantic[[:space:]]Segmentation/759adae1-3d57-4b38-8780-a8c8f1ca245e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/FLSeg_[[:space:]]Enhancing[[:space:]]Privacy[[:space:]]and[[:space:]]Robustness[[:space:]]in[[:space:]]Federated[[:space:]]Learning[[:space:]]under[[:space:]]Heterogeneous[[:space:]]Data[[:space:]]via[[:space:]]Model[[:space:]]Segmentation/a4f90945-afa0-4f01-8edd-1b86f5d9d098_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/FOLDER_[[:space:]]Accelerating[[:space:]]Multi-Modal[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]with[[:space:]]Enhanced[[:space:]]Performance/7a17b26d-c0b2-4285-9105-7de16ee42381_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/FPEM_[[:space:]]Face[[:space:]]Prior[[:space:]]Enhanced[[:space:]]Facial[[:space:]]Attractiveness[[:space:]]Prediction[[:space:]]for[[:space:]]Live[[:space:]]Videos[[:space:]]with[[:space:]]Face[[:space:]]Retouching/f47f8d6d-8a6c-491b-8f19-08123eb17e5c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/FREE-Merging_[[:space:]]Fourier[[:space:]]Transform[[:space:]]for[[:space:]]Efficient[[:space:]]Model[[:space:]]Merging/f3aa67c5-53d3-463b-8730-22ad4f95fe7f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/FRET_[[:space:]]Feature[[:space:]]Redundancy[[:space:]]Elimination[[:space:]]for[[:space:]]Test[[:space:]]Time[[:space:]]Adaptation/013e8d33-0504-4dd6-a3e9-326514c12e30_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/FROSS_[[:space:]]Faster-Than-Real-Time[[:space:]]Online[[:space:]]3D[[:space:]]Semantic[[:space:]]Scene[[:space:]]Graph[[:space:]]Generation[[:space:]]from[[:space:]]RGB-D[[:space:]]Images/f4b25ed8-2175-4e29-a597-e457f4795510_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/FVGen_[[:space:]]Accelerating[[:space:]]Novel-View[[:space:]]Synthesis[[:space:]]with[[:space:]]Adversarial[[:space:]]Video[[:space:]]Diffusion[[:space:]]Distillation/381cfeb4-88c2-44b1-a694-860cc7f57161_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/FW-Merging_[[:space:]]Scaling[[:space:]]Model[[:space:]]Merging[[:space:]]with[[:space:]]Frank-Wolfe[[:space:]]Optimization/56e48ade-9cb7-4688-a0be-edb71b4d459b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Face[[:space:]]Retouching[[:space:]]with[[:space:]]Diffusion[[:space:]]Data[[:space:]]Generation[[:space:]]and[[:space:]]Spectral[[:space:]]Restorement/2a12f446-c130-4e36-bc58-dba590a7c6d7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/FaceCraft4D_[[:space:]]Animated[[:space:]]3D[[:space:]]Facial[[:space:]]Avatar[[:space:]]Generation[[:space:]]from[[:space:]]a[[:space:]]Single[[:space:]]Image/c512f2aa-6fbd-4940-89be-2245ba51ffe2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/FaceLift_[[:space:]]Learning[[:space:]]Generalizable[[:space:]]Single[[:space:]]Image[[:space:]]3D[[:space:]]Face[[:space:]]Reconstruction[[:space:]]from[[:space:]]Synthetic[[:space:]]Heads/1763108f-f851-4f26-ba5a-510a2c051a36_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/FaceShield_[[:space:]]Defending[[:space:]]Facial[[:space:]]Image[[:space:]]against[[:space:]]Deepfake[[:space:]]Threats/4436b2e6-eb11-4750-8df1-5fa6df4f36fe_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/FaceXFormer_[[:space:]]A[[:space:]]Unified[[:space:]]Transformer[[:space:]]for[[:space:]]Facial[[:space:]]Analysis/43f275cd-265b-414a-a63d-4d3a6e26c5b2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Factorized[[:space:]]Learning[[:space:]]for[[:space:]]Temporally[[:space:]]Grounded[[:space:]]Video-Language[[:space:]]Models/0391813b-b98c-4912-884e-a194864058af_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Failure[[:space:]]Cases[[:space:]]Are[[:space:]]Better[[:space:]]Learned[[:space:]]But[[:space:]]Boundary[[:space:]]Says[[:space:]]Sorry_[[:space:]]Facilitating[[:space:]]Smooth[[:space:]]Perception[[:space:]]Change[[:space:]]for[[:space:]]Accuracy-Robustness[[:space:]]Trade-Off[[:space:]]in[[:space:]]Adversarial[[:space:]]Training/9886c220-8af0-438e-8e12-4de3189d4e86_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Fair[[:space:]]Generation[[:space:]]without[[:space:]]Unfair[[:space:]]Distortions_[[:space:]]Debiasing[[:space:]]Text-to-Image[[:space:]]Generation[[:space:]]with[[:space:]]Entanglement-Free[[:space:]]Attention/6ef80803-cc14-446f-a75b-f2593f51cc3f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/FairGen_[[:space:]]Enhancing[[:space:]]Fairness[[:space:]]in[[:space:]]Text-to-Image[[:space:]]Diffusion[[:space:]]Models[[:space:]]via[[:space:]]Self-Discovering[[:space:]]Latent[[:space:]]Directions/207183e6-db8a-436f-8e4f-faec1fb9d3a5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/FairHuman_[[:space:]]Boosting[[:space:]]Hand[[:space:]]and[[:space:]]Face[[:space:]]Quality[[:space:]]in[[:space:]]Human[[:space:]]Image[[:space:]]Generation[[:space:]]with[[:space:]]Minimum[[:space:]]Potential[[:space:]]Delay[[:space:]]Fairness[[:space:]]in[[:space:]]Diffusion[[:space:]]Models/c1cadd0b-ab93-4bff-a36b-6e24c8c2c52f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/FakeRadar_[[:space:]]Probing[[:space:]]Forgery[[:space:]]Outliers[[:space:]]to[[:space:]]Detect[[:space:]]Unknown[[:space:]]Deepfake[[:space:]]Videos/28a1e3bc-e2b9-40da-b7fe-fcc0d6369340_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Fast[[:space:]]Globally[[:space:]]Optimal[[:space:]]and[[:space:]]Geometrically[[:space:]]Consistent[[:space:]]3D[[:space:]]Shape[[:space:]]Matching/cbbb1f39-340b-4752-9244-371644b15edf_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Fast[[:space:]]Image[[:space:]]Super-Resolution[[:space:]]via[[:space:]]Consistency[[:space:]]Rectified[[:space:]]Flow/5e42c446-52f8-4769-b738-a2b096f3a7a5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/FastJSMA_[[:space:]]Accelerating[[:space:]]Jacobian-based[[:space:]]Saliency[[:space:]]Map[[:space:]]Attacks[[:space:]]through[[:space:]]Gradient[[:space:]]Decoupling/caddbdfc-9052-49a6-8b20-7e4e162ec5a3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/FastPoint_[[:space:]]Accelerating[[:space:]]3D[[:space:]]Point[[:space:]]Cloud[[:space:]]Model[[:space:]]Inference[[:space:]]via[[:space:]]Sample[[:space:]]Point[[:space:]]Distance[[:space:]]Prediction/670cb46f-786b-455d-998f-328c26d7e47d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/FastVAR_[[:space:]]Linear[[:space:]]Visual[[:space:]]Autoregressive[[:space:]]Modeling[[:space:]]via[[:space:]]Cached[[:space:]]Token[[:space:]]Pruning/70725db1-fcd5-4496-a95d-07038d174933_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Faster[[:space:]]and[[:space:]]Better[[:space:]]3D[[:space:]]Splatting[[:space:]]via[[:space:]]Group[[:space:]]Training/21033709-b705-4347-a1bb-c3e7a58591e2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Feather[[:space:]]the[[:space:]]Throttle_[[:space:]]Revisiting[[:space:]]Visual[[:space:]]Token[[:space:]]Pruning[[:space:]]for[[:space:]]Vision-Language[[:space:]]Model[[:space:]]Acceleration/eb2a11ee-50f9-4c43-8cc0-429a524795eb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Feature[[:space:]]Coding[[:space:]]in[[:space:]]the[[:space:]]Era[[:space:]]of[[:space:]]Large[[:space:]]Models_[[:space:]]Dataset,[[:space:]]Test[[:space:]]Conditions,[[:space:]]and[[:space:]]Benchmark/83895aed-d59c-465f-afa3-c8ed37ae4a4b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Feature[[:space:]]Decomposition-Recomposition[[:space:]]in[[:space:]]Large[[:space:]]Vision-Language[[:space:]]Model[[:space:]]for[[:space:]]Few-Shot[[:space:]]Class-Incremental[[:space:]]Learning/df51b4e9-12ac-40bf-8801-dbd6097351a5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Feature[[:space:]]Extraction[[:space:]]and[[:space:]]Representation[[:space:]]of[[:space:]]Pre-training[[:space:]]Point[[:space:]]Cloud[[:space:]]Based[[:space:]]on[[:space:]]Diffusion[[:space:]]Models/0c060285-4bbf-430f-b421-79f016d6dd03_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Feature[[:space:]]Purification[[:space:]]Matters_[[:space:]]Suppressing[[:space:]]Outlier[[:space:]]Propagation[[:space:]]for[[:space:]]Training-Free[[:space:]]Open-Vocabulary[[:space:]]Semantic[[:space:]]Segmentation/4f979197-bed4-425a-a1d2-c2b5ca2160b9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/FedAGC_[[:space:]]Federated[[:space:]]Continual[[:space:]]Learning[[:space:]]with[[:space:]]Asymmetric[[:space:]]Gradient[[:space:]]Correction/69b6bc67-ec8c-4e1a-b957-09f50cc21aab_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/FedDifRC_[[:space:]]Unlocking[[:space:]]the[[:space:]]Potential[[:space:]]of[[:space:]]Text-to-Image[[:space:]]Diffusion[[:space:]]Models[[:space:]]in[[:space:]]Heterogeneous[[:space:]]Federated[[:space:]]Learning/d48a8fed-132a-4154-9395-1fe6c487ac0e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/FedMVP_[[:space:]]Federated[[:space:]]Multimodal[[:space:]]Visual[[:space:]]Prompt[[:space:]]Tuning[[:space:]]for[[:space:]]Vision-Language[[:space:]]Models/9a8979ca-f59a-40f4-a042-065bd341c994_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/FedMeNF_[[:space:]]Privacy-Preserving[[:space:]]Federated[[:space:]]Meta-Learning[[:space:]]for[[:space:]]Neural[[:space:]]Fields/97cb26d5-5659-4687-b57e-cc32fd64dcc4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/FedPall_[[:space:]]Prototype-based[[:space:]]Adversarial[[:space:]]and[[:space:]]Collaborative[[:space:]]Learning[[:space:]]for[[:space:]]Federated[[:space:]]Learning[[:space:]]with[[:space:]]Feature[[:space:]]Drift/f6c6ce84-c8dc-4de0-bad0-3f6c4ed86bd8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/FedVLA_[[:space:]]Federated[[:space:]]Vision-Language-Action[[:space:]]Learning[[:space:]]with[[:space:]]Dual[[:space:]]Gating[[:space:]]Mixture-of-Experts[[:space:]]for[[:space:]]Robotic[[:space:]]Manipulation/9830aa12-506c-4606-9232-07331ac373a7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/FedWSQ_[[:space:]]Efficient[[:space:]]Federated[[:space:]]Learning[[:space:]]with[[:space:]]Weight[[:space:]]Standardization[[:space:]]and[[:space:]]Distribution-Aware[[:space:]]Non-Uniform[[:space:]]Quantization/1f166dee-20e5-452e-8e86-86019ce8d0bb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/FedXDS_[[:space:]]Leveraging[[:space:]]Model[[:space:]]Attribution[[:space:]]Methods[[:space:]]to[[:space:]]counteract[[:space:]]Data[[:space:]]Heterogeneity[[:space:]]in[[:space:]]Federated[[:space:]]Learning/3d4823b4-ec38-431a-8ed1-cb351135854c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Federated[[:space:]]Continual[[:space:]]Instruction[[:space:]]Tuning/c95b9507-af86-4883-a6e8-057630ebe555_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Federated[[:space:]]Continuous[[:space:]]Category[[:space:]]Discovery[[:space:]]and[[:space:]]Learning/7332a50f-74a4-4833-9616-ad03014bd631_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Federated[[:space:]]Domain[[:space:]]Generalization[[:space:]]with[[:space:]]Domain-specific[[:space:]]Soft[[:space:]]Prompts[[:space:]]Generation/b70289dd-c48a-4e53-8732-8d816d04b0c4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Federated[[:space:]]Prompt-Tuning[[:space:]]with[[:space:]]Heterogeneous[[:space:]]and[[:space:]]Incomplete[[:space:]]Multimodal[[:space:]]Client[[:space:]]Data/c4864325-3bcd-4609-87ea-0a89ff648c25_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Federated[[:space:]]Representation[[:space:]]Angle[[:space:]]Learning/f94f188a-83f1-436a-a6d8-36878318e078_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Feed-Forward[[:space:]]SceneDINO[[:space:]]for[[:space:]]Unsupervised[[:space:]]Semantic[[:space:]]Scene[[:space:]]Completion/0cb8dcd6-d20d-4d18-a810-629c106f3cdd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Few-Shot[[:space:]]Image[[:space:]]Quality[[:space:]]Assessment[[:space:]]via[[:space:]]Adaptation[[:space:]]of[[:space:]]Vision-Language[[:space:]]Models/89cdaf16-5828-4989-b825-83012badf535_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Few-Shot[[:space:]]Pattern[[:space:]]Detection[[:space:]]via[[:space:]]Template[[:space:]]Matching[[:space:]]and[[:space:]]Regression/9a1fc29b-9538-4897-943a-30b2211079d1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Fewer[[:space:]]Denoising[[:space:]]Steps[[:space:]]or[[:space:]]Cheaper[[:space:]]Per-Step[[:space:]]Inference_[[:space:]]Towards[[:space:]]Compute-Optimal[[:space:]]Diffusion[[:space:]]Model[[:space:]]Deployment/98cafbf9-4954-46a4-b5af-59f0bca621f7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/FiVE-Bench_[[:space:]]A[[:space:]]Fine-grained[[:space:]]Video[[:space:]]Editing[[:space:]]Benchmark[[:space:]]for[[:space:]]Evaluating[[:space:]]Emerging[[:space:]]Diffusion[[:space:]]and[[:space:]]Rectified[[:space:]]Flow[[:space:]]Models/61b80164-0e5a-42b8-b65d-d1598a219bbd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/FiffDepth_[[:space:]]Feed-forward[[:space:]]Transformation[[:space:]]of[[:space:]]Diffusion-Based[[:space:]]Generators[[:space:]]for[[:space:]]Detailed[[:space:]]Depth[[:space:]]Estimation/3c8dc7fb-98b2-4d77-857e-35d78d6eaaf6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/FinMMR_[[:space:]]Make[[:space:]]Financial[[:space:]]Numerical[[:space:]]Reasoning[[:space:]]More[[:space:]]Multimodal,[[:space:]]Comprehensive,[[:space:]]and[[:space:]]Challenging/9bdd874d-4c07-4812-a904-cd4ae1c8c840_origin.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/2025/Extrapolated Urban View Synthesis Benchmark/e65fa038-3ae6-4620-9f47-ad1a63db0abf_content_list.json b/2025/Extrapolated Urban View Synthesis Benchmark/e65fa038-3ae6-4620-9f47-ad1a63db0abf_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d4a58c4237d731b681ac823cff799c152be4c39f --- /dev/null +++ b/2025/Extrapolated Urban View Synthesis Benchmark/e65fa038-3ae6-4620-9f47-ad1a63db0abf_content_list.json @@ -0,0 +1,1342 @@ +[ + { + "type": "text", + "text": "Extrapolated Urban View Synthesis Benchmark", + "text_level": 1, + "bbox": [ + 254, + 130, + 741, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiangyu Han $^{1,3*}$ Zhen Jia $^{1*}$ Boyi Li $^{2}$ Yan Wang $^{2}$ Boris Ivanovic $^{2}$ Yurong You $^{2}$ Lingjie Liu $^{3}$ Yue Wang $^{2,4}$ Marco Pavone $^{2,5}$ Chen Feng $^{1}$ Yiming Li $^{1,2\\dagger}$ $^{1}$ NYU ${}^{2}$ NVIDIA ${}^{3}$ UPenn ${}^{4}$ USC ${}^{5}$ Stanford", + "bbox": [ + 147, + 178, + 849, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://ai4ce.github.io/EUVS-Benchmark", + "bbox": [ + 326, + 243, + 665, + 258 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2750e1de8e2e50beca42baa1099958d524968a043102410c836203c281dd0c6d.jpg", + "image_caption": [ + "Figure 1. Our key contributions. Previous evaluations for urban view synthesis have primarily focused on interpolated poses, as the lack of ground truth data has made it challenging to evaluate extrapolated poses. We address this gap by providing real-world data that enables both quantitative and qualitative evaluations of extrapolated view synthesis in urban scenes. The quantitative results reveal a significant performance drop in 3D Gaussian Splatting [25] when handling extrapolated views, highlighting the need for more robust NVS methods." + ], + "image_footnote": [], + "bbox": [ + 99, + 273, + 344, + 525 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fc9a854c63e2bee516fd917f4f42336948855da87e0a6a0784af664d4c075558.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 273, + 488, + 506 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ce57cae050600ffb83d7b56b896254455bb4ea9f953c02cde065b1107a85d519.jpg", + "image_caption": [ + "GT is available in previous benchmark" + ], + "image_footnote": [], + "bbox": [ + 398, + 508, + 419, + 525 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8e20560ca42402c382e700e53a8b707c1a96b5c1dba871c7051a5da554f6c2af.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 273, + 707, + 506 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/60d60ba9ab5eafe33b3d1185866f863d630b62c9e743fb06fe8a2c06ff73990e.jpg", + "image_caption": [ + "GT is available in our benchmark" + ], + "image_footnote": [], + "bbox": [ + 658, + 508, + 679, + 525 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/59d0e3a7f01289ad2105a343148fef98d14a6eaa7349ffc290cc04ae639af593.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 273, + 890, + 506 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 604, + 326, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Photorealistic simulators are essential for the training and evaluation of vision-centric autonomous vehicles (AVs). At their core is Novel View Synthesis (NVS), a crucial capability that generates diverse unseen viewpoints to accommodate the broad and continuous pose distribution of AVs. Recent advances in radiance fields, such as 3D Gaussian Splatting, achieve photorealistic rendering at real-time speeds and have been widely used in modeling large-scale driving scenes. However, their performance is commonly evaluated using an interpolated setup with highly correlated training and test views. In contrast, extrapolation, where test views largely deviate from training views, remains underexplored, limiting progress in generalizable simulation technology. To address this gap, we leverage publicly available AV datasets with multiple traversals,", + "bbox": [ + 88, + 637, + 485, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "multiple vehicles, and multiple cameras to build the first Extrapolated Urban View Synthesis (EUVS) benchmark. Meanwhile, we conduct both quantitative and qualitative evaluations of state-of-the-art NVS methods across different evaluation settings. Our results show that current NVS methods are prone to overfitting to training views. Besides, incorporating diffusion priors and improving geometry cannot fundamentally improve NVS under large view changes, highlighting the need for more robust approaches and large-scale training. We have released the data to help advance self-driving and urban robotics simulation technology.", + "bbox": [ + 511, + 606, + 908, + 773 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 796, + 645, + 811 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The development of vision-centric autonomous vehicles (AVs) relies heavily on photorealistic simulators, which provide controlled, reproducible, and scalable environments for training and evaluation of driving models [13, 44, 56]. These simulators enable AVs to learn and adapt to a variety", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 107, + 875, + 220, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28718", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of real-world scenarios, from crowded urban streets to adverse weather conditions, without the logistical and safety concerns of physical road testing. At the heart of these simulators is the capability for Novel View Synthesis (NVS)—a key technology that generates realistic images of unseen viewpoints, simulating the continuous changes in perspective that occur as AVs navigate through urban environments.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent advancements in radiance fields, particularly methods based on 3D Gaussian Splatting [25], have significantly improved the realism and efficiency of NVS. These approaches [8, 53, 64, 66] can produce photorealistic renderings at real-time speeds, making them highly attractive for large-scale driving scene simulation. However, despite their impressive results, the evaluation of NVS methods has predominantly focused on interpolated scenarios, where training and test viewpoints are closely related. While interpolation tests are valuable for assessing local consistency, they fall short in addressing the more critical challenge of extrapolation—where test viewpoints differ significantly from the training data. As shown in Figure 1, the interpolation test set demonstrates strong performance, with metrics such as PSNR, SSIM, and LPIPS remaining very close to the training set values. In contrast, the extrapolation test set, which includes additional translation and rotation changes relative to the training set, exhibits notable drops in performance. Specifically, the metric decreases relative to the training set are $28\\%$ for PSNR, $22\\%$ for SSIM, and $50\\%$ for LPIPS. These results underscore the urgent need to explore and advance extrapolated view synthesis in complex urban scenes, as real-world driving often involves encountering scenarios with significant viewpoint shifts and diverse spatial transformations that deviate from training distributions. Several recent studies [20, 23] have investigated the generalization capabilities of NVS in 3D Gaussian Splatting. Although they show promising qualitative results, there is no comprehensive quantitative analysis due to the absence of standardized datasets. Moreover, their evaluations are primarily limited to specific scenarios or use cases, without investigating varying evaluation settings based on the degree of extrapolation. This gap underscores the urgent need for a benchmark that offers diverse and challenging datasets, enabling a rigorous and systematic assessment of NVS methods.", + "bbox": [ + 91, + 200, + 483, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To establish a common platform for assessing the robustness of NVS methods, we introduce a comprehensive benchmark for quantitatively and qualitatively evaluating extrapolated novel view synthesis in large-scale urban scenes. Our benchmark leverages publicly available datasets, including NuPlan [4], MARS [30], and Argoverse2 [46], which feature multi-traversal, multi-agent and multi-camera sensory recordings. Multi-traversal data consists of asynchronous traversals of the same location, while multi-agent data refers to data collected simultaneously", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "from multiple vehicles within the same area. These data provide diverse camera poses within a 3D scene, enabling the training and evaluation of extrapolated view synthesis in outdoor environments. For the experimental setup, we define three evaluation settings: (1) translation only, (2) rotation only, and (3) translation + rotation, as shown in Figure 4. In autonomous driving scenarios, Setting 1 corresponds to maneuvers such as lane changes, Setting 2 involves switching between cameras facing different directions, and Setting 3 addresses complex intersections, such as crossroads with diverse traversal paths. These settings represent common challenges in autonomous driving, and addressing them enables the synthesis of complete scenes from sparse image observations.", + "bbox": [ + 511, + 90, + 903, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct pose estimation and sparse reconstruction using COLMAP [39], which facilitates the initialization of Gaussian Splatting. We then evaluate state-of-the-art Gaussian Splatting-based approaches across each evaluation setting, identifying performance gaps both qualitatively and quantitatively in extrapolated urban view synthesis.", + "bbox": [ + 511, + 301, + 903, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our main contributions are as follows:", + "bbox": [ + 532, + 393, + 870, + 407 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We initiate the first comprehensive quantitative and qualitative study on the Extrapolated Urban View Synthesis (EUVS) problem, supported by a robust evaluation framework that categorizes evaluation settings (translation-only, rotation-only, and translation + rotation) while assessing performance using diverse metrics, including reconstruction accuracy and visual fidelity.", + "- We construct a novel dataset by integrating multitraversal, multi-agent, and multi-camera data from publicly available resources, totaling 90,810 frames across 345 videos. Our dataset effectively addresses the limitations of existing benchmarks, enabling rigorous and robust evaluation for extrapolated urban view synthesis.", + "- We benchmark state-of-the-art Gaussian Splitting-based and NeRF-based models and analyze key factors that influence the performance of extrapolated NVs, laying a solid foundation for future advancements in this challenging task. Data and code will be released upon acceptance." + ], + "bbox": [ + 513, + 407, + 903, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 694, + 661, + 709 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extrapolated View Synthesis. Extrapolated view synthesis aims to generate novel views beyond observed perspectives, addressing challenges in visual coherence for unseen regions. RapNeRF [60] proposes a random ray-casting policy that enables training on unseen views based on visible ones. Following work [54] enhances this approach by incorporating holistic priors. Besides, some generalizable models [6, 9, 45] have emerged, capable of generating extrapolated novel views from a limited number of input images. While these methods are designed for indoor scenes, several works address extrapolated view synthesis in outdoor driving scenarios, which typically involve forward-facing", + "bbox": [ + 511, + 719, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28719", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "cameras and unbounded environments. To tackle the Setting 1 challenge in our benchmark and address the scarcity of lane change data, GGS [20] introduces a novel virtual lane generation module. In parallel, AutoSplat [26] tackles lane change in dynamic scenes by applying geometric and reflected consistency constraints. To address the Setting 1 or 2 challenge, FreeSim [15], VEGS [23], and SGD [58] enhance 3DGS [25] with diffusion priors to improve generalization ability. Yet existing methods suffer from two major limitations: (1) a lack of real data for quantitative evaluation, which confines them to quantitative analysis, and (2) a narrow focus on a specific setting in our benchmark, preventing a comprehensive and systematic exploration.", + "bbox": [ + 89, + 90, + 480, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Gaussian Splatting. Recent advances in radiance fields, particularly NeRF [34] and 3DGS [25], have garnered significant attention due to their impressive advancements in NVs. NeRF employs an implicit representation through a multi-layer perceptron (MLP). Furthermore, 3DGS explicitly represents scenes using anisotropic 3D Gaussian ellipsoids, enabling high-quality real-time rendering. Several works have addressed issues such as difficulties with reflective surfaces [24], aliasing [57], etc. However, urban scenes introduce unique challenges due to their unbounded and dynamic nature. To address the challenge, several works separate dynamic and static elements in the scene by leveraging a composite dynamic Gaussian graph [53, 66], optical flow prediction [55, 64], etc. PVG [8] presents a unified representation model that simultaneously incorporates both dynamic and static components without relying on priors. To achieve realistic geometry and efficient rendering, 2DGS [22] collapses 3D Gaussians onto 2D planes, while hybrid approaches [41, 48] combine different Gaussians to better capture region-specific features. In summary, current urban NVS methods primarily focus on effectively handling dynamic elements and enhancing geometry representation, while the challenge of extrapolated view synthesis remains largely underexplored.", + "bbox": [ + 91, + 292, + 483, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Autonomous Driving Simulators. Current simulators focus on three key challenges: parameter initialization [16, 42], traffic simulation [28, 51, 63], and sensor simulation [1, 13, 18, 27, 56, 62]. Sensor simulation is crucial for generating realistic sensory data that AVs depend on for perception and decision-making. Early sensor simulators [13, 40, 50] provide simulated environments that are valuable for research but lack visual realism. Recent studies have focused on data-driven simulators that extract data from real-world driving logs, creating more realistic and adaptable environments. These methods can be classified into two categories: generation-based and reconstruction-based approaches. The former rely on inputs such as text, video, and other data sources for simulation, supported by world models or prior knowledge [18, 27, 62]. Reconstruction-based simulations leverage real-world data", + "bbox": [ + 89, + 659, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to ensure both visual fidelity and geometric consistency [47, 49, 56]. UniSim [56] is a pioneering example of this approach, utilizing NeRF-based scene representation to create dynamic scenes with geometric information that are both editable and controllable. Extrapolated view synthesis is essential for these simulators, as it enables the generation of realistic and consistent views from diverse angles.", + "bbox": [ + 511, + 90, + 906, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Autonomous Driving Datasets. High-quality datasets play a vital role in advancing autonomous driving research. The KITTI dataset [19], released in 2012, marked a major milestone, significantly accelerating advances in AVs [17, 33, 43]. Since then, many influential autonomous vehicle datasets have been developed to tackle challenges like adverse weather conditions [37], multimodal fusion [3, 4], repeated driving [4, 12], collaborative driving [29, 30, 52], and motion prediction [5, 14, 46], etc. We leverage publicly available datasets with multi-traversal, multi-agent, and multi-camera recordings, enabling a comprehensive and robust evaluation of extrapolated urban view synthesis.", + "bbox": [ + 511, + 198, + 908, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. The EUVS Benchmark", + "text_level": 1, + "bbox": [ + 511, + 393, + 732, + 410 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Dataset Curation", + "text_level": 1, + "bbox": [ + 511, + 419, + 683, + 434 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We leverage three publicly available, community-verified autonomous driving datasets—nuPlan [4], Argoverse 2 [46], and MARS [30]—to enhance adoption and foster trust through their established reliability. nuPlan [4] provides 1,200 hours of driving data from four cities, serving as the first large-scale planning benchmark. Argoverse 2 [46] supports multimodal perception and forecasting with 1,000 annotated 3D scenarios, 20,000 unlabeled lidar sequences, and 250,000 motion forecasting cases. MARS [30] enables collaborative driving research with multi-agent and multitraversal scenarios. The original datasets were not designed for evaluating extrapolated view synthesis, which requires our significant labor in data processing— $300+$ hours of manual traversal selection and $800+$ hours of computing to run COLMAP. By integrating these datasets, our benchmark enables the evaluation of view synthesis across diverse and realistic urban environments under varying conditions. Figure 3 illustrates the distribution of the integrated datasets.", + "bbox": [ + 511, + 441, + 908, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Evaluation Framework", + "text_level": 1, + "bbox": [ + 511, + 726, + 728, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To systematically assess model performance in extrapolated urban view synthesis, our evaluation framework incorporates three evaluation settings and three data configurations. Data configurations include multi-traversal, multi-agent, and multi-camera, while evaluation settings are categorized into (1) translation only, (2) rotation only, and (3) translation + rotation, as illustrated in Figure 4.", + "bbox": [ + 511, + 750, + 906, + 853 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Setting 1. The translation-only experimental setup involves scenarios where the vehicle's position shifts without any change in orientation. This scenario is commonly observed", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28720", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1709a7854bc488c0ae02b1c9a19f4299eefca5241f7ba846a7195561fb2aa78d.jpg", + "image_caption": [ + "Figure 2. Dataset visualization. Our dataset features diverse scenes across various locations in different cities, sourced from multiple datasets. Typical driving scenarios include maneuvers such as lane changes, cross intersections, and T-junctions. Top: Each column displays images captured at the same location by different agents or traversals. Bottom: Each image displays the COLMAP points at a specific location, along with the corresponding camera poses." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 898, + 415 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/686306d5c9869386ddea58a0a3dd01fc160eb04be8c4e3c8449e63501b05a228.jpg", + "image_caption": [ + "Figure 3. Dataset distribution. Our dataset comprises 90,810 frames distributed over 104 cases, capturing a diverse array of multi-traversal paths, multi-agent interactions, and multi-camera perspectives across varying evaluation settings." + ], + "image_footnote": [], + "bbox": [ + 96, + 486, + 888, + 570 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in lane changes. We use traversals from different lanes in multi-traversal data, focusing on the three front cameras. The data is sourced from nuPlan [4] and Argoverse 2 [46].", + "bbox": [ + 89, + 622, + 482, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Setting 2. The second setting, rotation only, evaluates models on views with significant orientation changes. In vision-centric autonomous vehicles, this corresponds to transitions between cameras capturing different directions. We leverage multi-camera data from nuPlan [4], training on three forward-facing and three rear-facing cameras, while evaluating on two side-facing cameras.", + "bbox": [ + 89, + 670, + 482, + 776 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Setting 3. The third setting, combining translation and rotation, includes both positional shifts and orientation changes, posing the greatest challenge for NVS. To address this, we utilize multi-traversal driving data collected from the same location but across different traversal routes. For example, the training and test sets may include routes that approach an intersection from different directions. Typical route combinations feature scenarios such as intersections,", + "bbox": [ + 89, + 779, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "T-junctions, and Y-junctions, as shown in Figure 2, ensuring diversity and comprehensive evaluation. The data for this setting is from MARS [30] and Argoverse 2 [46].", + "bbox": [ + 511, + 622, + 906, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Algorithm Overview", + "text_level": 1, + "bbox": [ + 511, + 680, + 707, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Vanilla 3D Gaussian Splatting. 3D Gaussian Splatting (3DGS) [25] leverage 3D Gaussians to explicitly represent the scene, which achieves high quality while offering real-time rendering by avoiding unnecessary computation in the empty space. Building on this, 3DGM [31] leverages multitraversal consensus to differentiate transient and permanent elements, enabling joint 2D segmentation and 3D mapping without using any human supervision.", + "bbox": [ + 511, + 702, + 906, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Planar-based and Geometry Refined Gaussian Splat-ting. GaussianPro [11] builds on 3DGS [25] by introducing multi-frame geometric optimization, which guides the densi-sification of 3D Gaussians, enhancing scene consistency in complex geometries. It further refines geometry by encour", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28721", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/dce74e08eb5f1a159d2d3546b7f157deb1bfd9ffda30b71b6febb0bc60b72ed1.jpg", + "image_caption": [ + "(a) Evaluation Setting 1: Translation Only" + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 895, + 210 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dcd886910944c08c09424ea1e29b545ff39f8ff761dbbd9bbe7bd98c2d7a4e29.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 227, + 895, + 340 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cb960acf25f7e663d00c9736e853be8325d6acbf1a823ac3560a89ba1df9b371.jpg", + "image_caption": [ + "(b) Evaluation Setting 2: Rotation Only", + "(c) Evaluation Setting 3: Translation + Rotation", + "Figure 4. Qualitative and quantitative results across three evaluation settings. The performance drop from interpolation to extrapolation is significant in both qualitative and quantitative comparison. Different testing settings have distinct scenario characteristics, enabling the evaluation of a method's capabilities from various aspects, thus systematically assessing the overall performance of reconstruction algorithms, including geometric accuracy, hallucination ability, view consistency, and depth precision, etc." + ], + "image_footnote": [], + "bbox": [ + 96, + 362, + 895, + 478 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "aging Gaussian primitives to adopt flat structures. Similarly, 2DGS [22] projects the 3D volume into a set of 2D oriented planar Gaussian disks, enabling high-fidelity surface reconstruction. PGSR [7] introduces an unbiased depth rendering method and integrates single-view geometric, multiview photometric, and geometric regularization techniques to improve global geometry accuracy.", + "bbox": [ + 88, + 574, + 482, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Gaussian Splatting with Diffusion Priors. VEGS [23] introduces a novel view generalization approach that harnesses pre-extracted surface normals to align 3D Gaussians while generating augmented camera views guided by diffusion priors. These diffusion priors serve a dual purpose: providing denoising loss guidance and supervising the training of augmented cameras. This process effectively mitigates floating artifacts and fragmented geometries, resulting in more accurate and coherent 3D representations.", + "bbox": [ + 88, + 684, + 482, + 820 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Feature-Enhanced Gaussian Splatting. Feature 3DGS [65] extends 3D Gaussian Splatting with a Parallel N-dimensional Gaussian Rasterizer, allowing simultaneous rendering of radiance fields and high-dimensional semantic features. By embedding semantic features directly into 3D", + "bbox": [ + 89, + 825, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Gaussians, the approach enhances optimization, enabling better correspondence with scene semantics and achieving more detailed and accurate spatial representations.", + "bbox": [ + 511, + 574, + 906, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "NeRF-based Method. Instant-NGP [35] uses a multiresolution hash encoding to map spatial coordinates into compact latent representations via hash tables. It efficiently encodes high-frequency details by combining trainable feature vectors with interpolation, enabling adaptive and scalable input encodings. Zip-NeRF [2] leverages multisampling with isotropic Gaussians for scale-aware features and introduces a smooth anti-aliasing loss to address z-aliasing. In addition, it incorporates a novel distance normalization technique to better manage close and distant objects, achieving high-quality rendering and fast training.", + "bbox": [ + 511, + 621, + 908, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 511, + 805, + 638, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 511, + 830, + 689, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. All 3DGS-based methods are initialized only from sparse points obtained by COLMAP [39] and exclude lidar points. We employ Grounded-SAM-2", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28722", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/8762fe8aee21c2d60ceb89757fa613178e7395fb6bfa8c2867fa41b4fc7afd75.jpg", + "table_caption": [ + "Table 1. Quantitative rendering results across three evaluation settings. In. denotes interpolation, while Ex. represents extrapolation. Different baselines excel in different settings, reflecting the comprehensiveness and completeness of the evaluation protocol." + ], + "table_footnote": [], + "table_body": "
MethodPSNR↑SSIM↑LPIPS↓Feat Cos Sim ↑
In.Ex.DropIn.Ex.DropIn.Ex.DropIn.Ex.Drop
Setting 13DGS [25]21.3616.3723.4%0.82750.720313.0%0.20410.259927.3%0.68280.603911.6%
3DGM [31]20.9616.3522.0%0.82930.724812.6%0.20030.254226.9%0.68020.608710.5%
GSPro [11]21.5116.3923.8%0.83100.718913.5%0.18040.245035.8%0.70810.613013.4%
VEGS [23]21.2615.8825.3%0.81070.704713.1%0.24980.306222.6%0.63230.552112.7%
PGSR [7]20.5716.3220.7%0.81040.710212.4%0.22620.273320.8%0.65150.584810.2%
2DGS [22]20.8716.3021.9%0.80760.710312.0%0.24380.289018.5%0.62560.56449.8%
Feature 3DGS [65]21.0216.0123.8%0.80960.724310.5%0.18760.257537.3%0.69580.612212.0%
Zip-NeRF [2]19.6814.0628.6%0.78560.691712.0%0.27110.341826.1%0.63180.554212.3%
Instant-NGP [35]18.7712.6532.6%0.76310.625218.1%0.48740.593821.8%0.54650.483711.5%
AVERAGE20.6715.5924.6%0.80830.704612.8%0.25010.313425.3%0.65050.577411.2%
Setting 23DGS [25]25.7519.5324.2%0.87660.751114.3%0.15360.266873.7%0.73270.631913.8%
3DGM [31]25.7518.7827.1%0.87860.746415.0%0.15560.281380.8%0.72780.634412.8%
GSPro [11]26.4219.3926.6%0.88210.747015.3%0.13290.224669.0%0.75230.648713.8%
VEGS [23]24.5423.334.9%0.83660.79495.0%0.23010.281122.2%0.65950.61337.0%
PGSR [7]24.5318.3825.1%0.86120.711917.3%0.15550.253262.8%0.72000.581719.2%
2DGS [22]25.1518.8325.1%0.85780.720416.0%0.17560.291766.1%0.68980.578516.1%
Feature 3DGS [65]24.9119.5921.4%0.88000.786410.6%0.13770.227865.4%0.74270.646413.0%
Zip-NeRF [2]29.0617.3640.3%0.86600.671522.5%0.20780.358272.4%0.74790.584321.9%
Instant-NGP [35]25.6117.1533.0%0.85960.721216.1%0.33400.517154.8%0.72540.618214.8%
AVERAGE25.7519.1525.6%0.86650.739014.7%0.18700.300262.5%0.72200.615314.8%
Setting 33DGS [25]21.2214.9929.4%0.85500.716916.1%0.22520.405079.8%0.70020.477431.8%
3DGM [31]20.6214.6029.2%0.85430.723315.3%0.22540.404979.6%0.68740.467232.0%
GSPro [11]21.5814.8231.3%0.86340.699619.0%0.20100.387792.9%0.70930.454136.0%
VEGS [23]21.1314.2532.6%0.82660.647521.7%0.23590.442287.5%0.67850.444234.5%
PGSR [7]19.6014.1727.7%0.82380.698415.2%0.29340.436348.7%0.58670.378735.5%
2DGS [22]17.3511.3634.5%0.75680.544728.0%0.42960.545927.1%0.35520.232734.5%
Feature 3DGS [65]21.8814.3334.5%0.86430.638626.1%0.14000.3816172.6%0.74110.466937.0%
Zip-NeRF [2]20.6114.4230.0%0.83830.656521.7%0.21970.4546106.9%0.71080.364548.7%
Instant-NGP [35]19.6314.3926.7%0.81790.710413.1%0.49560.659233.0%0.60830.415731.7%
AVERAGE20.4014.1530.6%0.83340.670719.5%0.27400.457570.0%0.64190.411335.9%
", + "bbox": [ + 91, + 127, + 903, + 547 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6d06a4968b2da68a66458dcb048216312f128e1bcba7c93697539170691d92d4.jpg", + "table_caption": [ + "Table 2. Quantitative comparison of depth evaluation in extrapolated views in Setting 1. 3DGM [31] demonstrates superior performance on most evaluation metrics, while VEGS [23] and GSPro [11] excel in SqRel and Delta1, respectively." + ], + "table_footnote": [], + "table_body": "
BaselineAbsRel ↓RMSE ↓SqRel ↓Delta1 ↑Delta2 ↑Delta3 ↑
3DGS [25]0.36114.4410.410.6490.8240.895
3DGM [31]0.30113.938.9060.6510.8460.915
PGSR [7]0.36617.5721.500.7590.8340.883
GSPro [11]0.35519.6632.010.6430.8390.909
VEGS [23]0.36815.008.3980.4410.6910.827
", + "bbox": [ + 94, + 627, + 473, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "[38] to mask out pixels corresponding to potentially movable objects during both training and evaluation and also exclude them from the initialization of 3DGS.", + "bbox": [ + 89, + 732, + 482, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We use three widely-used metrics to evaluate visual quality: peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and learned perceptual image patch similarity (LPIPS) [61]. We also employ DINOv2 [36] feature cosine similarity to evaluate image quality in latent space. For geometry evaluation, we use depth metrics, including RMSE and $\\delta_{1.25}$ . Additionally, we incorporate the Fréchet Inception Distance (FID) [21] to", + "bbox": [ + 89, + 779, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9322a9ccabd85634e9bafa9debda576eff1aadb37c39e6137bb878f22d5237c2.jpg", + "table_caption": [ + "Table 3. Perceptual quality metric of VEGS [23]. Our benchmark also incorporates the Fréchet Inception Distance (FID) to evaluate the inpainting ability of models with generative priors." + ], + "table_footnote": [], + "table_body": "
Setting 1Setting 2Setting 3AVERAGE
FID (In.) ↓46.333.546.142.0
FID (Ex.) ↓87.467.2132.495.7
FID (Drop)89%101%187%126.7%
", + "bbox": [ + 516, + 613, + 897, + 679 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "evaluate the inpainting ability of models with strong priors.", + "bbox": [ + 511, + 695, + 903, + 712 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Experimental Results", + "text_level": 1, + "bbox": [ + 511, + 724, + 718, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 presents the quantitative results across Settings 1-3, while Figure 5 illustrates the qualitative outcomes on the extrapolated test set. Table 3 and Table 2 present the perceptual quality and depth metrics. The results indicate that, while the metrics perform relatively well in the interpolated test set, there is a significant drop in performance in the extrapolated test set across all baselines.", + "bbox": [ + 511, + 748, + 906, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Setting 1: Translation-only. In Setting 1, training views fully cover test views with moderate translational changes. (1) Results show a consistent drop in performance from", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28723", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5481253f8bc9af45aa50f2778f2efc58a4a90a023e48c8b55a1490a87f41c2d2.jpg", + "image_caption": [ + "Setting 1: Translation Only" + ], + "image_footnote": [], + "bbox": [ + 99, + 90, + 395, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/544945f12614aa72971e9dcd893b56c88f38ebbcc03e182571dfcb3fd272fe91.jpg", + "image_caption": [ + "Setting 2: Rotation Only" + ], + "image_footnote": [], + "bbox": [ + 398, + 90, + 620, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a281b5f540c7d598c32ec069ca3181a039727fc9a85a60f86c9fbaf2d2c6fcb8.jpg", + "image_caption": [ + "Setting 3: Rotation + Translation", + "Figure 5. Qualitative comparison of extrapolated view synthesis across different settings. For each setting, results from different methods are compared against the ground truth. Red boxes highlight areas where methods are limited in capturing fine details, such as road surfaces, sky regions, or object boundaries, demonstrating the challenges faced by each approach under varying movement complexities." + ], + "image_footnote": [], + "bbox": [ + 624, + 90, + 890, + 398 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "interpolation to extrapolation across all metrics, highlighting the challenge of generalization to views with variations. Relative drops vary by metric: PSNR falls $23 - 25\\%$ (e.g., GSPro [11]: $21.51 \\rightarrow 16.39$ ), with SSIM and LPIPS showing similar declines. (2) On the extrapolative test set, methods perform comparably: GSPro tops PSNR (16.39) and minimizes LPIPS (0.2450), while 3DGM [31] leads in SSIM (0.7248). Both yield similar feature cosine similarity; VEGS [23] and PGSR [7] lag slightly. Results suggest that although GSPro marginally outperforms others, differences are small, underscoring the need for improved solutions.", + "bbox": [ + 89, + 469, + 482, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Setting 2: Rotation-only. In Setting 2, training views provide extensive coverage of the surrounding scene. However, most methods show poor generalization, with PSNR dropping by $22.75\\%$ . Rotation changes are particularly challenging in textured regions (e.g., trees and intricate details), often resulting in blurring, while regions perpendicular to the vehicle remain difficult to capture. Additionally, distant areas pose reconstruction challenges, frequently leading to missing buildings and sky blackouts. Among the evaluated baselines, VEGS [23] and GSPro [11] stand out as the best-performing baselines in this setting: VEGS utilizes its diffusion prior for inpainting and refining missing regions, while GSPro's robust geometry handling improves generalization.", + "bbox": [ + 89, + 640, + 482, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Setting 3: Translation + Rotation. In Setting 3, the view changes are the largest. (1) All methods exhibit notable performance drops from interpolation to extrapolation across metrics. For instance, 3DGS [25] experiences a PSNR", + "bbox": [ + 89, + 839, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "drop of $29.4\\%$ ( $21.22 \\rightarrow 14.99$ ), while GSPro [11] undergoes a similar drop of $31.3\\%$ ( $21.58 \\rightarrow 14.82$ ). (2) Feature 3DGS [65] and 3DGM [31] stand out as leading methods, excelling in LPIPS (0.3816) and SSIM (0.7233), respectively. However, overall performance remains limited, with PSNR consistently falling below 15, highlighting significant room for improvement in generating high-fidelity outputs for extrapolated views.", + "bbox": [ + 511, + 469, + 906, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Discussions", + "text_level": 1, + "bbox": [ + 511, + 603, + 635, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Lighting Inconsistency Handling. Lighting inconsistencies are a widespread challenge in multi-traversal datasets due to varying illumination and weather conditions. To address this, we have taken specific measures. (1) Importantly, our multi-traversal dataset is manually curated to ensure that the lighting across images appears consistent to the eye. This step helps reduce the impact of extreme lighting changes and ensures a fairer evaluation of our method under controlled conditions. (2) One potential approach is to incorporate an appearance embedding for each image. We experiment with the Gaussian in the Wild (GS-W) baseline [59]. GS-W replaces traditional spherical harmonic-based color modeling with a method that separates intrinsic properties of each Gaussian point from dynamic appearance features of each image. This approach captures the stable, inherent appearance of objects while accommodating dynamic factors like highlights and shadows. Despite these efforts, our results, especially in Setting 3, still exhibit", + "bbox": [ + 511, + 628, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28724", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/da6c60a73c394fcedf097463a5914d3d68971797e22304e6938efca9efb646be.jpg", + "image_caption": [ + "(a) Mitigating lighting inconsistency by camera embedding." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 491, + 262 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9739511aa7f4910e80e89a9e7eb4c373e6276bbd22fcc5af6056f54a9bad3e71.jpg", + "image_caption": [ + "(b) Qualitative dynamic scenes rendering comparison.", + "Figure 6. Qualitative results of dynamic baseline and lighting handling. We mitigate the lighting inconsistency issue by carefully selecting traversals under similar lighting conditions, and it can be further alleviated by introducing camera embeddings. We provide the quantitative evaluation of dynamic scene reconstruction methods in Setting 2." + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 900, + 262 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9bd9aca3939a6630077de23f5d942da933d8f5e31ba8ae7cc6252053785e3eac.jpg", + "table_caption": [ + "Table 4. Quantitative performance of GS-W [59] in different settings. After learning the lighting features, the interpolated and extrapolated test metrics show significant improvement compared to other baselines. However, there is still a considerable drop from interpolation to extrapolation." + ], + "table_footnote": [], + "table_body": "
SettingPSNR ↑SSIM ↑LPIPS ↓Feat Cos Sim ↑
In.Ex.In.Ex.In.Ex.In.Ex.
Setting 128.1520.220.890.780.150.230.710.64
Setting 230.1021.210.910.820.130.200.760.67
Setting 328.6219.360.870.730.150.310.740.50
", + "bbox": [ + 101, + 421, + 467, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "a significant performance drop, as shown in Table 4 and Figure 6a, underscoring the limitations of Gaussian-based models in handling extrapolated, unseen scenarios.", + "bbox": [ + 89, + 516, + 482, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Dynamic Scenes. In addition to the static scene reconstruction methods that we evaluated in Table 1, some dynamic scene reconstruction techniques have recently emerged. We conduct evaluation for Setting 2 to assess these approaches, e.g., OmniRe [10]. It organizes rigid-deformable nodes and background nodes to capture dynamic scene structures and employs SMPL [32] for non-rigid object modeling. As illustrated in Figure 6b, the rendering results reveal a significant performance gap between the training and extrapolated test cameras. In the extrapolated test views, objects such as trees and stakes lose texture and geometric details, resulting in noticeably blurry outputs, whereas the training views maintain high fidelity. As shown in Table 5, the reconstruction metrics indicate an average drop of $25\\%$ when transitioning from interpolated to extrapolated settings. The results highlight the challenges of extrapolated view synthesis in dynamic scenes and the need for further research.", + "bbox": [ + 89, + 561, + 482, + 818 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions, Limitations and Future Work", + "text_level": 1, + "bbox": [ + 89, + 829, + 475, + 845 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conclusions. We introduce the first benchmark enabling quantitative evaluation of extrapolated view synthesis, advancing photorealistic simulation for self-driving", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b2139876ec7719c95af06b6d0d6930f0c541c57a706658d91c4bbc8487ee49b9.jpg", + "table_caption": [ + "Table 5. Quantitative performance of OmniRe [10] in Setting 2. The experiment uses a single traversal with dynamic objects, showing a noticeable drop from interpolation to extrapolation." + ], + "table_footnote": [], + "table_body": "
SettingPSNR ↑SSIM ↑LPIPS ↓Feat Cos Sim ↑
In.Ex.In.Ex.In.Ex.In.Ex.
Setting 219.7815.320.650.450.380.530.730.58
", + "bbox": [ + 524, + 393, + 888, + 452 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and robotics. Our benchmark integrates real-world multi-traversal, multi-agent, and multi-camera data, categorizes scenes into different evaluation settings, and evaluates state-of-the-art NVS models. Experimental results reveal that while some methods address specific challenges, current models demonstrate limited generalization, with significant overfitting to training views and suboptimal performance in extrapolated view synthesis. To support further research, we will release the dataset and benchmark, addressing the long-standing data scarcity and providing evaluation protocols. We believe the EUVS benchmark will catalyze meaningful advancements in self-driving and robotics innovation.", + "bbox": [ + 511, + 463, + 906, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and Future Work. Our benchmark does have limitations. First, although we provide the ground truth to enable quantitative evaluation for static scenes in all settings and the foreground objects evaluation in Setting 2, we lack the foreground evaluation in Settings 1 and 3. Future work will aim to expand our work to include the dynamic object evaluation in Settings 1 and 3 using multi-agent data. Secondly, although we carefully manually selected and ensured that the test trajectory viewpoints are well covered by training trajectories, a more in-depth evaluation of how observed and unseen regions differ is left for future work.", + "bbox": [ + 511, + 643, + 908, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment. This work was supported in part through NSF grants 2238968 and 2121391, and the NYU IT High Performance Computing resources, services, and staff expertise. Yiming Li is supported by NVIDIA Graduate Fellowship (2024-2025).", + "bbox": [ + 511, + 810, + 908, + 886 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28725", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Alexander Amini, Tsun-Hsuan Wang, Igor Gilitschenski, Wilko Schwarting, Zhijian Liu, Song Han, Sertac Karaman, and Daniela Rus. Vista 2.0: An open, data-driven simulator for multimodal sensing and policy learning for autonomous vehicles. In 2022 International Conference on Robotics and Automation (ICRA), pages 2419-2426. IEEE, 2022. 3", + "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023. 5, 6", + "[3] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 3", + "[4] Holger Caesar, Juraj Kabzan, Kok Seang Tan, Whye Kit Fong, Eric Wolff, Alex Lang, Luke Fletcher, Oscar Beijbom, and Sammy Omari. nuplan: A closed-loop ml-based planning benchmark for autonomous vehicles. arXiv preprint arXiv:2106.11810, 2021. 2, 3, 4", + "[5] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3d tracking and forecasting with rich maps. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8748-8757, 2019. 3", + "[6] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19467, 2024. 2", + "[7] Danpeng Chen, Hai Li, Weicai Ye, Yifan Wang, Weijian Xie, Shangjin Zhai, Nan Wang, Haomin Liu, Hujun Bao, and Guofeng Zhang. Pgsr: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction. arXiv preprint arXiv:2406.06521, 2024. 5, 6, 7", + "[8] Yurui Chen, Chun Gu, Junzhe Jiang, Xiatian Zhu, and Li Zhang. Periodic vibration gaussian: Dynamic urban scene reconstruction and real-time rendering. arXiv preprint arXiv:2311.18561, 2023. 2, 3", + "[9] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splattering from sparse multi-view images. arXiv preprint arXiv:2403.14627, 2024. 2", + "[10] Ziyu Chen, Jiawei Yang, Jiahui Huang, Riccardo de Lutio, Janick Martinez Esturo, Boris Ivanovic, Or Litany, Zan Gojcic, Sanja Fidler, Marco Pavone, et al. Omnire: Omni urban scene reconstruction. arXiv preprint arXiv:2408.16760, 2024. 8", + "[11] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussian-pro: 3d gaussian splattering with progressive propagation. In" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Forty-first International Conference on Machine Learning, 2024. 4, 6, 7", + "[12] Carlos A Diaz-Ruiz, Youya Xia, Yurong You, Jose Nino, Junan Chen, Josephine Monica, Xiangyu Chen, Katie Luo, Yan Wang, Marc Emond, et al. Ithaca365: Dataset and driving perception under repeated and challenging weather conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21383-21392, 2022. 3", + "[13] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 1, 3", + "[14] Scott Ettinger, Shuyang Cheng, Benjamin Caine, Chenxi Liu, Hang Zhao, Sabeek Pradhan, Yuning Chai, Ben Sapp, Charles R Qi, Yin Zhou, et al. Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9710-9719, 2021. 3", + "[15] Lue Fan, Hao Zhang, Qitai Wang, Hongsheng Li, and Zhaoxiang Zhang. Freesim: Toward free-viewpoint camera simulation in driving scenes, 2024. 3", + "[16] Lan Feng, Quanyi Li, Zhenghao Peng, Shuhan Tan, and Bolei Zhou. Trafficgen: Learning to generate diverse and realistic traffic scenarios. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 3567-3575. IEEE, 2023. 3", + "[17] Jannik Fritsch, Tobias Kuehnl, and Andreas Geiger. A new performance measure and evaluation benchmark for road detection algorithms. In 16th International IEEE Conference on Intelligent Transportation Systems (ITSC 2013), pages 1693-1700. IEEE, 2013. 3", + "[18] Shenyuan Gao, Jiazhi Yang, Li Chen, Kashyap Chitta, Yihang Qiu, Andreas Geiger, Jun Zhang, and Hongyang Li. Vista: A generalizable driving world model with high fidelity and versatile controllability. arXiv preprint arXiv:2405.17398, 2024. 3", + "[19] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 3", + "[20] Huasong Han, Kaixuan Zhou, Xiaoxiao Long, Yusen Wang, and Chunxia Xiao. Ggs: Generalizable gaussian splatting for lane switching in autonomous driving. arXiv preprint arXiv:2409.02382, 2024. 2, 3", + "[21] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6", + "[22] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3, 5, 6", + "[23] Sungwon Hwang, Min-Jung Kim, Taewoong Kang, Jayeon Kang, and Jaegul Choo. Vegs: View extrapolation of urban scenes in 3d gaussian splatting using learned priors. arXiv preprint arXiv:2407.02945, 2024. 2, 3, 5, 6, 7" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28726", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5322-5332, 2024. 3", + "[25] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 2, 3, 4, 6, 7", + "[26] Mustafa Khan, Hamidreza Fazlali, Dhruv Sharma, Tongtong Cao, Dongfeng Bai, Yuan Ren, and Bingbing Liu. Autosplat: Constrained gaussian splatting for autonomous driving scene reconstruction. arXiv preprint arXiv:2407.02598, 2024. 3", + "[27] Seung Wook Kim, Jonah Philion, Antonio Torralba, and Sanja Fidler. Drivegan: Towards a controllable high-quality neural simulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5820-5829, 2021. 3", + "[28] Quanyi Li, Zhenghao Mark Peng, Lan Feng, Zhizheng Liu, Chenda Duan, Wenjie Mo, and Bolei Zhou. ScenarioNet: Open-source platform for large-scale traffic scenario simulation and modeling. Advances in neural information processing systems, 36, 2024. 3", + "[29] Yiming Li, Dekun Ma, Ziyan An, Zixun Wang, Yiqi Zhong, Siheng Chen, and Chen Feng. V2x-sim: Multi-agent collaborative perception dataset and benchmark for autonomous driving. IEEE Robotics and Automation Letters, 7(4): 10914-10921, 2022. 3", + "[30] Yiming Li, Zhiheng Li, Nuo Chen, Moonjun Gong, Zonglin Lyu, Zehong Wang, Peili Jiang, and Chen Feng. Multiagent multitraversal multimodal self-driving: Open mars dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22041-22051, 2024. 2, 3, 4", + "[31] Yiming Li, Zehong Wang, Yue Wang, Zhiding Yu, Zan Gojcic, Marco Pavone, Chen Feng, and Jose M. Alvarez. Memorize what matters: Emergent scene decomposition from multitraverse. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 4, 6, 7", + "[32] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1-248:16, 2015. 8", + "[33] Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3061-3070, 2015. 3", + "[34] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3", + "[35] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 5, 6" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6", + "[37] Matthew Pitropov, Danson Evan Garcia, Jason Rebello, Michael Smart, Carlos Wang, Krzysztof Czarnecki, and Steven Waslander. Canadian adverse driving conditions dataset. The International Journal of Robotics Research, 40 (4-5):681-690, 2021. 3", + "[38] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dólar, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos, 2024. 6", + "[39] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 5", + "[40] Shital Shah, Debadeepta Dey, Chris Lovett, and Ashish Kapoor. Airsim: High-fidelity visual and physical simulation for autonomous vehicles. In Field and Service Robotics: Results of the 11th International Conference, pages 621-635. Springer, 2018. 3", + "[41] Xi Shi, Lingli Chen, Peng Wei, Xi Wu, Tian Jiang, Yonggang Luo, and Lecheng Xie. Dhgs: Decoupled hybrid gaussian splatting for driving scene. arXiv preprint arXiv:2407.16600, 2024.3", + "[42] Shuhan Tan, Kelvin Wong, Shenlong Wang, Sivabalan Manivasagam, Mengye Ren, and Raquel Urtasun. Scenegen: Learning to generate realistic traffic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 892-901, 2021. 3", + "[43] Izzeddin Teeti, Valentina Musat, Salman Khan, Alexander Rast, Fabio Cuzzolin, and Andrew Bradley. Vision in adverse weather: Augmentation using cyclegans with various object detectors for robust perception in autonomous racing. arXiv preprint arXiv:2201.03246, v3, 2023. 3", + "[44] Adam Tondersski, Carl Lindstrom, Georg Hess, William Ljungbergh, Lennart Svensson, and Christoffer Petersson. Neurad: Neural rendering for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14895-14904, 2024. 1", + "[45] Christopher Wewer, Kevin Raj, Eddy Ilg, Bernt Schiele, and Jan Eric Lenssen. latentsplat: Autoencoding variational gaussians for fast generalizable 3d reconstruction. arXiv preprint arXiv:2403.16292, 2024. 2", + "[46] Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, et al. Argoverse 2: Next generation datasets for self-driving perception and forecasting. arXiv preprint arXiv:2301.00493, 2023. 2, 3, 4", + "[47] Chenming Wu, Jiadai Sun, Zhelun Shen, and Liangjun Zhang. Mapnerf: Incorporating map priors into neural ra" + ], + "bbox": [ + 516, + 90, + 905, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28727", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "diance fields for driving view simulation. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7082-7088. IEEE, 2023. 3", + "[48] Ke Wu, Kaizhao Zhang, Zhiwei Zhang, Shanshuai Yuan, Muer Tie, Julong Wei, Zijun Xu, Jieru Zhao, Zhongxue Gan, and Wenchao Ding. Hgs-mapping: Online dense mapping using hybrid gaussian representation in urban scenes. arXiv preprint arXiv:2403.20159, 2024. 3", + "[49] Zirui Wu, Tianyu Liu, Liyi Luo, Zhide Zhong, Jianteng Chen, Hongmin Xiao, Chao Hou, Haozhe Lou, Yuantao Chen, Runyi Yang, et al. Mars: An instance-aware, modular and realistic simulator for autonomous driving. In CAAI International Conference on Artificial Intelligence, pages 3-15. Springer, 2023. 3", + "[50] Bernhard Wymann, Eric Espie, Christophe Guionneau, Christos Dimitrakakis, Rémi Coulom, and Andrew Sumner. Torcs, the open racing car simulator. Software available at http://torcs.sourceforge.net, 4(6):2, 2000. 3", + "[51] Danfei Xu, Yuxiao Chen, Boris Ivanovic, and Marco Pavone. Bits: Bi-level imitation for traffic simulation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2929-2936. IEEE, 2023. 3", + "[52] Runsheng Xu, Xin Xia, Jinlong Li, Hanzhao Li, Shuo Zhang, Zhengzhong Tu, Zonglin Meng, Hao Xiang, Xiaoyu Dong, Rui Song, et al. V2v4real: A real-world large-scale dataset for vehicle-to-vehicle cooperative perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13712-13722, 2023. 3", + "[53] Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians for modeling dynamic urban scenes. arXiv preprint arXiv:2401.01339, 2024. 2, 3", + "[54] Chen Yang, Peihao Li, Zanwei Zhou, Shanxin Yuan, Bingbing Liu, Xiaokang Yang, Weichao Qiu, and Wei Shen. Ner-fvs: Neural radiance fields for free view synthesis via geometry scaffolds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16549-16558, 2023. 2", + "[55] Jiawei Yang, Boris Ivanovic, Or Litany, Xinshuo Weng, Seung Wook Kim, Boyi Li, Tong Che, Danfei Xu, Sanja Fidler, Marco Pavone, et al. Emernerf: Emergent spatial-temporal scene decomposition via self-supervision. arXiv preprint arXiv:2311.02077, 2023. 3", + "[56] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1389-1399, 2023. 1, 3", + "[57] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19447-19456, 2024. 3", + "[58] Zhongrui Yu, Haoran Wang, Jinze Yang, Hanzhang Wang, Zeke Xie, Yunfeng Cai, Jiale Cao, Zhong Ji, and Mingming" + ], + "bbox": [ + 91, + 90, + 482, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Sun. Sgd: Street view synthesis with gaussian splatting and diffusion prior. arXiv preprint arXiv:2403.20079, 2024. 3", + "[59] Dongbin Zhang, Chuming Wang, Weitao Wang, Peihao Li, Minghan Qin, and Haoqian Wang. Gaussian in the wild: 3d gaussian splatting for unconstrained image collections. arXiv preprint arXiv:2403.15704, 2024. 7, 8", + "[60] Jian Zhang, Yuanqing Zhang, Huan Fu, Xiaowei Zhou, Bowen Cai, Jinchi Huang, Rongfei Jia, Binqiang Zhao, and Xing Tang. Ray priors through reprojection: Improving neural radiance fields for novel view extrapolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18376-18386, 2022. 2", + "[61] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6", + "[62] Guosheng Zhao, Chaojun Ni, Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, Boyuan Wang, Youyi Zhang, Wenjun Mei, and Xingang Wang. Drivedreamer4d: World models are effective data machines for 4d driving scene representation. arXiv preprint arXiv:2410.13571, 2024. 3", + "[63] Ziyuan Zhong, Davis Rempe, Danfei Xu, Yuxiao Chen, Sushant Veer, Tong Che, Baishakhi Ray, and Marco Pavone. Guided conditional diffusion for controllable traffic simulation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 3560-3566. IEEE, 2023. 3", + "[64] Hongyu Zhou, Jiahao Shao, Lu Xu, Dongfeng Bai, Weichao Qiu, Bingbing Liu, Yue Wang, Andreas Geiger, and Yiyi Liao. Hugs: Holistic urban 3d scene understanding via gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21336-21345, 2024. 2, 3", + "[65] Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Zehao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21676-21685, 2024. 5, 6, 7", + "[66] Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21634-21643, 2024. 2, 3" + ], + "bbox": [ + 516, + 92, + 903, + 752 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "28728", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Extrapolated Urban View Synthesis Benchmark/e65fa038-3ae6-4620-9f47-ad1a63db0abf_model.json b/2025/Extrapolated Urban View Synthesis Benchmark/e65fa038-3ae6-4620-9f47-ad1a63db0abf_model.json new file mode 100644 index 0000000000000000000000000000000000000000..858b2dd50d034e59ae04cd4241d1a4be2c4875c7 --- /dev/null +++ b/2025/Extrapolated Urban View Synthesis Benchmark/e65fa038-3ae6-4620-9f47-ad1a63db0abf_model.json @@ -0,0 +1,2213 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.256, + 0.131, + 0.743, + 0.154 + ], + "angle": 0, + "content": "Extrapolated Urban View Synthesis Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.148, + 0.179, + 0.85, + 0.242 + ], + "angle": 0, + "content": "Xiangyu Han\\(^{1,3*}\\) Zhen Jia\\(^{1*}\\) Boyi Li\\(^{2}\\) Yan Wang\\(^{2}\\) Boris Ivanovic\\(^{2}\\) Yurong You\\(^{2}\\) Lingjie Liu\\(^{3}\\) Yue Wang\\(^{2,4}\\) Marco Pavone\\(^{2,5}\\) Chen Feng\\(^{1}\\) Yiming Li\\(^{1,2\\dagger}\\) \n\\(^{1}\\)NYU \\({}^{2}\\)NVIDIA \\({}^{3}\\)UPenn \\({}^{4}\\)USC \\({}^{5}\\)Stanford" + }, + { + "type": "text", + "bbox": [ + 0.327, + 0.244, + 0.666, + 0.259 + ], + "angle": 0, + "content": "https://ai4ce.github.io/EUVS-Benchmark" + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.274, + 0.345, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.274, + 0.49, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.509, + 0.421, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.421, + 0.511, + 0.6, + 0.523 + ], + "angle": 0, + "content": "GT is available in previous benchmark" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.274, + 0.709, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.659, + 0.509, + 0.681, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.682, + 0.513, + 0.838, + 0.523 + ], + "angle": 0, + "content": "GT is available in our benchmark" + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.275, + 0.891, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.537, + 0.908, + 0.594 + ], + "angle": 0, + "content": "Figure 1. Our key contributions. Previous evaluations for urban view synthesis have primarily focused on interpolated poses, as the lack of ground truth data has made it challenging to evaluate extrapolated poses. We address this gap by providing real-world data that enables both quantitative and qualitative evaluations of extrapolated view synthesis in urban scenes. The quantitative results reveal a significant performance drop in 3D Gaussian Splatting [25] when handling extrapolated views, highlighting the need for more robust NVS methods." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.606, + 0.327, + 0.621 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.638, + 0.486, + 0.865 + ], + "angle": 0, + "content": "Photorealistic simulators are essential for the training and evaluation of vision-centric autonomous vehicles (AVs). At their core is Novel View Synthesis (NVS), a crucial capability that generates diverse unseen viewpoints to accommodate the broad and continuous pose distribution of AVs. Recent advances in radiance fields, such as 3D Gaussian Splatting, achieve photorealistic rendering at real-time speeds and have been widely used in modeling large-scale driving scenes. However, their performance is commonly evaluated using an interpolated setup with highly correlated training and test views. In contrast, extrapolation, where test views largely deviate from training views, remains underexplored, limiting progress in generalizable simulation technology. To address this gap, we leverage publicly available AV datasets with multiple traversals," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.607, + 0.909, + 0.775 + ], + "angle": 0, + "content": "multiple vehicles, and multiple cameras to build the first Extrapolated Urban View Synthesis (EUVS) benchmark. Meanwhile, we conduct both quantitative and qualitative evaluations of state-of-the-art NVS methods across different evaluation settings. Our results show that current NVS methods are prone to overfitting to training views. Besides, incorporating diffusion priors and improving geometry cannot fundamentally improve NVS under large view changes, highlighting the need for more robust approaches and large-scale training. We have released the data to help advance self-driving and urban robotics simulation technology." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.797, + 0.646, + 0.813 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The development of vision-centric autonomous vehicles (AVs) relies heavily on photorealistic simulators, which provide controlled, reproducible, and scalable environments for training and evaluation of driving models [13, 44, 56]. These simulators enable AVs to learn and adapt to a variety" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.221, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.236, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28718" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "of real-world scenarios, from crowded urban streets to adverse weather conditions, without the logistical and safety concerns of physical road testing. At the heart of these simulators is the capability for Novel View Synthesis (NVS)—a key technology that generates realistic images of unseen viewpoints, simulating the continuous changes in perspective that occur as AVs navigate through urban environments." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.201, + 0.485, + 0.745 + ], + "angle": 0, + "content": "Recent advancements in radiance fields, particularly methods based on 3D Gaussian Splatting [25], have significantly improved the realism and efficiency of NVS. These approaches [8, 53, 64, 66] can produce photorealistic renderings at real-time speeds, making them highly attractive for large-scale driving scene simulation. However, despite their impressive results, the evaluation of NVS methods has predominantly focused on interpolated scenarios, where training and test viewpoints are closely related. While interpolation tests are valuable for assessing local consistency, they fall short in addressing the more critical challenge of extrapolation—where test viewpoints differ significantly from the training data. As shown in Figure 1, the interpolation test set demonstrates strong performance, with metrics such as PSNR, SSIM, and LPIPS remaining very close to the training set values. In contrast, the extrapolation test set, which includes additional translation and rotation changes relative to the training set, exhibits notable drops in performance. Specifically, the metric decreases relative to the training set are \\(28\\%\\) for PSNR, \\(22\\%\\) for SSIM, and \\(50\\%\\) for LPIPS. These results underscore the urgent need to explore and advance extrapolated view synthesis in complex urban scenes, as real-world driving often involves encountering scenarios with significant viewpoint shifts and diverse spatial transformations that deviate from training distributions. Several recent studies [20, 23] have investigated the generalization capabilities of NVS in 3D Gaussian Splatting. Although they show promising qualitative results, there is no comprehensive quantitative analysis due to the absence of standardized datasets. Moreover, their evaluations are primarily limited to specific scenarios or use cases, without investigating varying evaluation settings based on the degree of extrapolation. This gap underscores the urgent need for a benchmark that offers diverse and challenging datasets, enabling a rigorous and systematic assessment of NVS methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To establish a common platform for assessing the robustness of NVS methods, we introduce a comprehensive benchmark for quantitatively and qualitatively evaluating extrapolated novel view synthesis in large-scale urban scenes. Our benchmark leverages publicly available datasets, including NuPlan [4], MARS [30], and Argoverse2 [46], which feature multi-traversal, multi-agent and multi-camera sensory recordings. Multi-traversal data consists of asynchronous traversals of the same location, while multi-agent data refers to data collected simultaneously" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.302 + ], + "angle": 0, + "content": "from multiple vehicles within the same area. These data provide diverse camera poses within a 3D scene, enabling the training and evaluation of extrapolated view synthesis in outdoor environments. For the experimental setup, we define three evaluation settings: (1) translation only, (2) rotation only, and (3) translation + rotation, as shown in Figure 4. In autonomous driving scenarios, Setting 1 corresponds to maneuvers such as lane changes, Setting 2 involves switching between cameras facing different directions, and Setting 3 addresses complex intersections, such as crossroads with diverse traversal paths. These settings represent common challenges in autonomous driving, and addressing them enables the synthesis of complete scenes from sparse image observations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.303, + 0.905, + 0.394 + ], + "angle": 0, + "content": "We conduct pose estimation and sparse reconstruction using COLMAP [39], which facilitates the initialization of Gaussian Splatting. We then evaluate state-of-the-art Gaussian Splatting-based approaches across each evaluation setting, identifying performance gaps both qualitatively and quantitatively in extrapolated urban view synthesis." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.395, + 0.871, + 0.408 + ], + "angle": 0, + "content": "In summary, our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.409, + 0.905, + 0.514 + ], + "angle": 0, + "content": "- We initiate the first comprehensive quantitative and qualitative study on the Extrapolated Urban View Synthesis (EUVS) problem, supported by a robust evaluation framework that categorizes evaluation settings (translation-only, rotation-only, and translation + rotation) while assessing performance using diverse metrics, including reconstruction accuracy and visual fidelity." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.515, + 0.905, + 0.605 + ], + "angle": 0, + "content": "- We construct a novel dataset by integrating multitraversal, multi-agent, and multi-camera data from publicly available resources, totaling 90,810 frames across 345 videos. Our dataset effectively addresses the limitations of existing benchmarks, enabling rigorous and robust evaluation for extrapolated urban view synthesis." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.606, + 0.905, + 0.68 + ], + "angle": 0, + "content": "- We benchmark state-of-the-art Gaussian Splitting-based and NeRF-based models and analyze key factors that influence the performance of extrapolated NVs, laying a solid foundation for future advancements in this challenging task. Data and code will be released upon acceptance." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.409, + 0.905, + 0.68 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.695, + 0.663, + 0.71 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Extrapolated View Synthesis. Extrapolated view synthesis aims to generate novel views beyond observed perspectives, addressing challenges in visual coherence for unseen regions. RapNeRF [60] proposes a random ray-casting policy that enables training on unseen views based on visible ones. Following work [54] enhances this approach by incorporating holistic priors. Besides, some generalizable models [6, 9, 45] have emerged, capable of generating extrapolated novel views from a limited number of input images. While these methods are designed for indoor scenes, several works address extrapolated view synthesis in outdoor driving scenarios, which typically involve forward-facing" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28719" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.288 + ], + "angle": 0, + "content": "cameras and unbounded environments. To tackle the Setting 1 challenge in our benchmark and address the scarcity of lane change data, GGS [20] introduces a novel virtual lane generation module. In parallel, AutoSplat [26] tackles lane change in dynamic scenes by applying geometric and reflected consistency constraints. To address the Setting 1 or 2 challenge, FreeSim [15], VEGS [23], and SGD [58] enhance 3DGS [25] with diffusion priors to improve generalization ability. Yet existing methods suffer from two major limitations: (1) a lack of real data for quantitative evaluation, which confines them to quantitative analysis, and (2) a narrow focus on a specific setting in our benchmark, preventing a comprehensive and systematic exploration." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.293, + 0.485, + 0.655 + ], + "angle": 0, + "content": "3D Gaussian Splatting. Recent advances in radiance fields, particularly NeRF [34] and 3DGS [25], have garnered significant attention due to their impressive advancements in NVs. NeRF employs an implicit representation through a multi-layer perceptron (MLP). Furthermore, 3DGS explicitly represents scenes using anisotropic 3D Gaussian ellipsoids, enabling high-quality real-time rendering. Several works have addressed issues such as difficulties with reflective surfaces [24], aliasing [57], etc. However, urban scenes introduce unique challenges due to their unbounded and dynamic nature. To address the challenge, several works separate dynamic and static elements in the scene by leveraging a composite dynamic Gaussian graph [53, 66], optical flow prediction [55, 64], etc. PVG [8] presents a unified representation model that simultaneously incorporates both dynamic and static components without relying on priors. To achieve realistic geometry and efficient rendering, 2DGS [22] collapses 3D Gaussians onto 2D planes, while hybrid approaches [41, 48] combine different Gaussians to better capture region-specific features. In summary, current urban NVS methods primarily focus on effectively handling dynamic elements and enhancing geometry representation, while the challenge of extrapolated view synthesis remains largely underexplored." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Autonomous Driving Simulators. Current simulators focus on three key challenges: parameter initialization [16, 42], traffic simulation [28, 51, 63], and sensor simulation [1, 13, 18, 27, 56, 62]. Sensor simulation is crucial for generating realistic sensory data that AVs depend on for perception and decision-making. Early sensor simulators [13, 40, 50] provide simulated environments that are valuable for research but lack visual realism. Recent studies have focused on data-driven simulators that extract data from real-world driving logs, creating more realistic and adaptable environments. These methods can be classified into two categories: generation-based and reconstruction-based approaches. The former rely on inputs such as text, video, and other data sources for simulation, supported by world models or prior knowledge [18, 27, 62]. Reconstruction-based simulations leverage real-world data" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.198 + ], + "angle": 0, + "content": "to ensure both visual fidelity and geometric consistency [47, 49, 56]. UniSim [56] is a pioneering example of this approach, utilizing NeRF-based scene representation to create dynamic scenes with geometric information that are both editable and controllable. Extrapolated view synthesis is essential for these simulators, as it enables the generation of realistic and consistent views from diverse angles." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.199, + 0.909, + 0.38 + ], + "angle": 0, + "content": "Autonomous Driving Datasets. High-quality datasets play a vital role in advancing autonomous driving research. The KITTI dataset [19], released in 2012, marked a major milestone, significantly accelerating advances in AVs [17, 33, 43]. Since then, many influential autonomous vehicle datasets have been developed to tackle challenges like adverse weather conditions [37], multimodal fusion [3, 4], repeated driving [4, 12], collaborative driving [29, 30, 52], and motion prediction [5, 14, 46], etc. We leverage publicly available datasets with multi-traversal, multi-agent, and multi-camera recordings, enabling a comprehensive and robust evaluation of extrapolated urban view synthesis." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.395, + 0.733, + 0.411 + ], + "angle": 0, + "content": "3. The EUVS Benchmark" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.42, + 0.684, + 0.435 + ], + "angle": 0, + "content": "3.1. Dataset Curation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.443, + 0.909, + 0.716 + ], + "angle": 0, + "content": "We leverage three publicly available, community-verified autonomous driving datasets—nuPlan [4], Argoverse 2 [46], and MARS [30]—to enhance adoption and foster trust through their established reliability. nuPlan [4] provides 1,200 hours of driving data from four cities, serving as the first large-scale planning benchmark. Argoverse 2 [46] supports multimodal perception and forecasting with 1,000 annotated 3D scenarios, 20,000 unlabeled lidar sequences, and 250,000 motion forecasting cases. MARS [30] enables collaborative driving research with multi-agent and multitraversal scenarios. The original datasets were not designed for evaluating extrapolated view synthesis, which requires our significant labor in data processing—\\(300+\\) hours of manual traversal selection and \\(800+\\) hours of computing to run COLMAP. By integrating these datasets, our benchmark enables the evaluation of view synthesis across diverse and realistic urban environments under varying conditions. Figure 3 illustrates the distribution of the integrated datasets." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.727, + 0.73, + 0.741 + ], + "angle": 0, + "content": "3.2. Evaluation Framework" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.854 + ], + "angle": 0, + "content": "To systematically assess model performance in extrapolated urban view synthesis, our evaluation framework incorporates three evaluation settings and three data configurations. Data configurations include multi-traversal, multi-agent, and multi-camera, while evaluation settings are categorized into (1) translation only, (2) rotation only, and (3) translation + rotation, as illustrated in Figure 4." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.901 + ], + "angle": 0, + "content": "Setting 1. The translation-only experimental setup involves scenarios where the vehicle's position shifts without any change in orientation. This scenario is commonly observed" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28720" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.9, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.423, + 0.907, + 0.479 + ], + "angle": 0, + "content": "Figure 2. Dataset visualization. Our dataset features diverse scenes across various locations in different cities, sourced from multiple datasets. Typical driving scenarios include maneuvers such as lane changes, cross intersections, and T-junctions. Top: Each column displays images captured at the same location by different agents or traversals. Bottom: Each image displays the COLMAP points at a specific location, along with the corresponding camera poses." + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.487, + 0.889, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.579, + 0.908, + 0.608 + ], + "angle": 0, + "content": "Figure 3. Dataset distribution. Our dataset comprises 90,810 frames distributed over 104 cases, capturing a diverse array of multi-traversal paths, multi-agent interactions, and multi-camera perspectives across varying evaluation settings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.623, + 0.483, + 0.669 + ], + "angle": 0, + "content": "in lane changes. We use traversals from different lanes in multi-traversal data, focusing on the three front cameras. The data is sourced from nuPlan [4] and Argoverse 2 [46]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.671, + 0.483, + 0.777 + ], + "angle": 0, + "content": "Setting 2. The second setting, rotation only, evaluates models on views with significant orientation changes. In vision-centric autonomous vehicles, this corresponds to transitions between cameras capturing different directions. We leverage multi-camera data from nuPlan [4], training on three forward-facing and three rear-facing cameras, while evaluating on two side-facing cameras." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Setting 3. The third setting, combining translation and rotation, includes both positional shifts and orientation changes, posing the greatest challenge for NVS. To address this, we utilize multi-traversal driving data collected from the same location but across different traversal routes. For example, the training and test sets may include routes that approach an intersection from different directions. Typical route combinations feature scenarios such as intersections," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.623, + 0.907, + 0.669 + ], + "angle": 0, + "content": "T-junctions, and Y-junctions, as shown in Figure 2, ensuring diversity and comprehensive evaluation. The data for this setting is from MARS [30] and Argoverse 2 [46]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.681, + 0.709, + 0.697 + ], + "angle": 0, + "content": "3.3. Algorithm Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.703, + 0.907, + 0.825 + ], + "angle": 0, + "content": "Vanilla 3D Gaussian Splatting. 3D Gaussian Splatting (3DGS) [25] leverage 3D Gaussians to explicitly represent the scene, which achieves high quality while offering real-time rendering by avoiding unnecessary computation in the empty space. Building on this, 3DGM [31] leverages multitraversal consensus to differentiate transient and permanent elements, enabling joint 2D segmentation and 3D mapping without using any human supervision." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Planar-based and Geometry Refined Gaussian Splat-ting. GaussianPro [11] builds on 3DGS [25] by introducing multi-frame geometric optimization, which guides the densi-sification of 3D Gaussians, enhancing scene consistency in complex geometries. It further refines geometry by encour" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28721" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.09, + 0.896, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.342, + 0.212, + 0.63, + 0.227 + ], + "angle": 0, + "content": "(a) Evaluation Setting 1: Translation Only" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.228, + 0.896, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.351, + 0.345, + 0.621, + 0.361 + ], + "angle": 0, + "content": "(b) Evaluation Setting 2: Rotation Only" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.363, + 0.896, + 0.479 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.334, + 0.48, + 0.66, + 0.495 + ], + "angle": 0, + "content": "(c) Evaluation Setting 3: Translation + Rotation" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.501, + 0.908, + 0.558 + ], + "angle": 0, + "content": "Figure 4. Qualitative and quantitative results across three evaluation settings. The performance drop from interpolation to extrapolation is significant in both qualitative and quantitative comparison. Different testing settings have distinct scenario characteristics, enabling the evaluation of a method's capabilities from various aspects, thus systematically assessing the overall performance of reconstruction algorithms, including geometric accuracy, hallucination ability, view consistency, and depth precision, etc." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.575, + 0.483, + 0.681 + ], + "angle": 0, + "content": "aging Gaussian primitives to adopt flat structures. Similarly, 2DGS [22] projects the 3D volume into a set of 2D oriented planar Gaussian disks, enabling high-fidelity surface reconstruction. PGSR [7] introduces an unbiased depth rendering method and integrates single-view geometric, multiview photometric, and geometric regularization techniques to improve global geometry accuracy." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.685, + 0.483, + 0.821 + ], + "angle": 0, + "content": "Gaussian Splatting with Diffusion Priors. VEGS [23] introduces a novel view generalization approach that harnesses pre-extracted surface normals to align 3D Gaussians while generating augmented camera views guided by diffusion priors. These diffusion priors serve a dual purpose: providing denoising loss guidance and supervising the training of augmented cameras. This process effectively mitigates floating artifacts and fragmented geometries, resulting in more accurate and coherent 3D representations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Feature-Enhanced Gaussian Splatting. Feature 3DGS [65] extends 3D Gaussian Splatting with a Parallel N-dimensional Gaussian Rasterizer, allowing simultaneous rendering of radiance fields and high-dimensional semantic features. By embedding semantic features directly into 3D" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.575, + 0.907, + 0.62 + ], + "angle": 0, + "content": "Gaussians, the approach enhances optimization, enabling better correspondence with scene semantics and achieving more detailed and accurate spatial representations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.622, + 0.909, + 0.789 + ], + "angle": 0, + "content": "NeRF-based Method. Instant-NGP [35] uses a multiresolution hash encoding to map spatial coordinates into compact latent representations via hash tables. It efficiently encodes high-frequency details by combining trainable feature vectors with interpolation, enabling adaptive and scalable input encodings. Zip-NeRF [2] leverages multisampling with isotropic Gaussians for scale-aware features and introduces a smooth anti-aliasing loss to address z-aliasing. In addition, it incorporates a novel distance normalization technique to better manage close and distant objects, achieving high-quality rendering and fast training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.806, + 0.64, + 0.823 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.832, + 0.691, + 0.849 + ], + "angle": 0, + "content": "4.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. All 3DGS-based methods are initialized only from sparse points obtained by COLMAP [39] and exclude lidar points. We employ Grounded-SAM-2" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "28722" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.119 + ], + "angle": 0, + "content": "Table 1. Quantitative rendering results across three evaluation settings. In. denotes interpolation, while Ex. represents extrapolation. Different baselines excel in different settings, reflecting the comprehensiveness and completeness of the evaluation protocol." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.128, + 0.905, + 0.548 + ], + "angle": 0, + "content": "
MethodPSNR↑SSIM↑LPIPS↓Feat Cos Sim ↑
In.Ex.DropIn.Ex.DropIn.Ex.DropIn.Ex.Drop
Setting 13DGS [25]21.3616.3723.4%0.82750.720313.0%0.20410.259927.3%0.68280.603911.6%
3DGM [31]20.9616.3522.0%0.82930.724812.6%0.20030.254226.9%0.68020.608710.5%
GSPro [11]21.5116.3923.8%0.83100.718913.5%0.18040.245035.8%0.70810.613013.4%
VEGS [23]21.2615.8825.3%0.81070.704713.1%0.24980.306222.6%0.63230.552112.7%
PGSR [7]20.5716.3220.7%0.81040.710212.4%0.22620.273320.8%0.65150.584810.2%
2DGS [22]20.8716.3021.9%0.80760.710312.0%0.24380.289018.5%0.62560.56449.8%
Feature 3DGS [65]21.0216.0123.8%0.80960.724310.5%0.18760.257537.3%0.69580.612212.0%
Zip-NeRF [2]19.6814.0628.6%0.78560.691712.0%0.27110.341826.1%0.63180.554212.3%
Instant-NGP [35]18.7712.6532.6%0.76310.625218.1%0.48740.593821.8%0.54650.483711.5%
AVERAGE20.6715.5924.6%0.80830.704612.8%0.25010.313425.3%0.65050.577411.2%
Setting 23DGS [25]25.7519.5324.2%0.87660.751114.3%0.15360.266873.7%0.73270.631913.8%
3DGM [31]25.7518.7827.1%0.87860.746415.0%0.15560.281380.8%0.72780.634412.8%
GSPro [11]26.4219.3926.6%0.88210.747015.3%0.13290.224669.0%0.75230.648713.8%
VEGS [23]24.5423.334.9%0.83660.79495.0%0.23010.281122.2%0.65950.61337.0%
PGSR [7]24.5318.3825.1%0.86120.711917.3%0.15550.253262.8%0.72000.581719.2%
2DGS [22]25.1518.8325.1%0.85780.720416.0%0.17560.291766.1%0.68980.578516.1%
Feature 3DGS [65]24.9119.5921.4%0.88000.786410.6%0.13770.227865.4%0.74270.646413.0%
Zip-NeRF [2]29.0617.3640.3%0.86600.671522.5%0.20780.358272.4%0.74790.584321.9%
Instant-NGP [35]25.6117.1533.0%0.85960.721216.1%0.33400.517154.8%0.72540.618214.8%
AVERAGE25.7519.1525.6%0.86650.739014.7%0.18700.300262.5%0.72200.615314.8%
Setting 33DGS [25]21.2214.9929.4%0.85500.716916.1%0.22520.405079.8%0.70020.477431.8%
3DGM [31]20.6214.6029.2%0.85430.723315.3%0.22540.404979.6%0.68740.467232.0%
GSPro [11]21.5814.8231.3%0.86340.699619.0%0.20100.387792.9%0.70930.454136.0%
VEGS [23]21.1314.2532.6%0.82660.647521.7%0.23590.442287.5%0.67850.444234.5%
PGSR [7]19.6014.1727.7%0.82380.698415.2%0.29340.436348.7%0.58670.378735.5%
2DGS [22]17.3511.3634.5%0.75680.544728.0%0.42960.545927.1%0.35520.232734.5%
Feature 3DGS [65]21.8814.3334.5%0.86430.638626.1%0.14000.3816172.6%0.74110.466937.0%
Zip-NeRF [2]20.6114.4230.0%0.83830.656521.7%0.21970.4546106.9%0.71080.364548.7%
Instant-NGP [35]19.6314.3926.7%0.81790.710413.1%0.49560.659233.0%0.60830.415731.7%
AVERAGE20.4014.1530.6%0.83340.670719.5%0.27400.457570.0%0.64190.411335.9%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.561, + 0.486, + 0.618 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison of depth evaluation in extrapolated views in Setting 1. 3DGM [31] demonstrates superior performance on most evaluation metrics, while VEGS [23] and GSPro [11] excel in SqRel and Delta1, respectively." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.628, + 0.475, + 0.719 + ], + "angle": 0, + "content": "
BaselineAbsRel ↓RMSE ↓SqRel ↓Delta1 ↑Delta2 ↑Delta3 ↑
3DGS [25]0.36114.4410.410.6490.8240.895
3DGM [31]0.30113.938.9060.6510.8460.915
PGSR [7]0.36617.5721.500.7590.8340.883
GSPro [11]0.35519.6632.010.6430.8390.909
VEGS [23]0.36815.008.3980.4410.6910.827
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.733, + 0.483, + 0.778 + ], + "angle": 0, + "content": "[38] to mask out pixels corresponding to potentially movable objects during both training and evaluation and also exclude them from the initialization of 3DGS." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Evaluation Metrics. We use three widely-used metrics to evaluate visual quality: peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and learned perceptual image patch similarity (LPIPS) [61]. We also employ DINOv2 [36] feature cosine similarity to evaluate image quality in latent space. For geometry evaluation, we use depth metrics, including RMSE and \\(\\delta_{1.25}\\). Additionally, we incorporate the Fréchet Inception Distance (FID) [21] to" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.561, + 0.907, + 0.604 + ], + "angle": 0, + "content": "Table 3. Perceptual quality metric of VEGS [23]. Our benchmark also incorporates the Fréchet Inception Distance (FID) to evaluate the inpainting ability of models with generative priors." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.614, + 0.898, + 0.68 + ], + "angle": 0, + "content": "
Setting 1Setting 2Setting 3AVERAGE
FID (In.) ↓46.333.546.142.0
FID (Ex.) ↓87.467.2132.495.7
FID (Drop)89%101%187%126.7%
" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.696, + 0.905, + 0.713 + ], + "angle": 0, + "content": "evaluate the inpainting ability of models with strong priors." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.725, + 0.719, + 0.74 + ], + "angle": 0, + "content": "4.2. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.749, + 0.907, + 0.854 + ], + "angle": 0, + "content": "Table 1 presents the quantitative results across Settings 1-3, while Figure 5 illustrates the qualitative outcomes on the extrapolated test set. Table 3 and Table 2 present the perceptual quality and depth metrics. The results indicate that, while the metrics perform relatively well in the interpolated test set, there is a significant drop in performance in the extrapolated test set across all baselines." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Setting 1: Translation-only. In Setting 1, training views fully cover test views with moderate translational changes. (1) Results show a consistent drop in performance from" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28723" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.091, + 0.396, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.158, + 0.401, + 0.337, + 0.416 + ], + "angle": 0, + "content": "Setting 1: Translation Only" + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.091, + 0.622, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.43, + 0.402, + 0.59, + 0.417 + ], + "angle": 0, + "content": "Setting 2: Rotation Only" + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.091, + 0.892, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.651, + 0.401, + 0.863, + 0.415 + ], + "angle": 0, + "content": "Setting 3: Rotation + Translation" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.42, + 0.907, + 0.463 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison of extrapolated view synthesis across different settings. For each setting, results from different methods are compared against the ground truth. Red boxes highlight areas where methods are limited in capturing fine details, such as road surfaces, sky regions, or object boundaries, demonstrating the challenges faced by each approach under varying movement complexities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.47, + 0.483, + 0.637 + ], + "angle": 0, + "content": "interpolation to extrapolation across all metrics, highlighting the challenge of generalization to views with variations. Relative drops vary by metric: PSNR falls \\(23 - 25\\%\\) (e.g., GSPro [11]: \\(21.51 \\rightarrow 16.39\\)), with SSIM and LPIPS showing similar declines. (2) On the extrapolative test set, methods perform comparably: GSPro tops PSNR (16.39) and minimizes LPIPS (0.2450), while 3DGM [31] leads in SSIM (0.7248). Both yield similar feature cosine similarity; VEGS [23] and PGSR [7] lag slightly. Results suggest that although GSPro marginally outperforms others, differences are small, underscoring the need for improved solutions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.641, + 0.483, + 0.837 + ], + "angle": 0, + "content": "Setting 2: Rotation-only. In Setting 2, training views provide extensive coverage of the surrounding scene. However, most methods show poor generalization, with PSNR dropping by \\(22.75\\%\\). Rotation changes are particularly challenging in textured regions (e.g., trees and intricate details), often resulting in blurring, while regions perpendicular to the vehicle remain difficult to capture. Additionally, distant areas pose reconstruction challenges, frequently leading to missing buildings and sky blackouts. Among the evaluated baselines, VEGS [23] and GSPro [11] stand out as the best-performing baselines in this setting: VEGS utilizes its diffusion prior for inpainting and refining missing regions, while GSPro's robust geometry handling improves generalization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Setting 3: Translation + Rotation. In Setting 3, the view changes are the largest. (1) All methods exhibit notable performance drops from interpolation to extrapolation across metrics. For instance, 3DGS [25] experiences a PSNR" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.47, + 0.907, + 0.592 + ], + "angle": 0, + "content": "drop of \\(29.4\\%\\) (\\(21.22 \\rightarrow 14.99\\)), while GSPro [11] undergoes a similar drop of \\(31.3\\%\\) (\\(21.58 \\rightarrow 14.82\\)). (2) Feature 3DGS [65] and 3DGM [31] stand out as leading methods, excelling in LPIPS (0.3816) and SSIM (0.7233), respectively. However, overall performance remains limited, with PSNR consistently falling below 15, highlighting significant room for improvement in generating high-fidelity outputs for extrapolated views." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.604, + 0.636, + 0.62 + ], + "angle": 0, + "content": "5. Discussions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Lighting Inconsistency Handling. Lighting inconsistencies are a widespread challenge in multi-traversal datasets due to varying illumination and weather conditions. To address this, we have taken specific measures. (1) Importantly, our multi-traversal dataset is manually curated to ensure that the lighting across images appears consistent to the eye. This step helps reduce the impact of extreme lighting changes and ensures a fairer evaluation of our method under controlled conditions. (2) One potential approach is to incorporate an appearance embedding for each image. We experiment with the Gaussian in the Wild (GS-W) baseline [59]. GS-W replaces traditional spherical harmonic-based color modeling with a method that separates intrinsic properties of each Gaussian point from dynamic appearance features of each image. This approach captures the stable, inherent appearance of objects while accommodating dynamic factors like highlights and shadows. Despite these efforts, our results, especially in Setting 3, still exhibit" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28724" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.493, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.134, + 0.266, + 0.45, + 0.278 + ], + "angle": 0, + "content": "(a) Mitigating lighting inconsistency by camera embedding." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.089, + 0.901, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.558, + 0.266, + 0.843, + 0.278 + ], + "angle": 0, + "content": "(b) Qualitative dynamic scenes rendering comparison." + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.285, + 0.908, + 0.327 + ], + "angle": 0, + "content": "Figure 6. Qualitative results of dynamic baseline and lighting handling. We mitigate the lighting inconsistency issue by carefully selecting traversals under similar lighting conditions, and it can be further alleviated by introducing camera embeddings. We provide the quantitative evaluation of dynamic scene reconstruction methods in Setting 2." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.341, + 0.483, + 0.41 + ], + "angle": 0, + "content": "Table 4. Quantitative performance of GS-W [59] in different settings. After learning the lighting features, the interpolated and extrapolated test metrics show significant improvement compared to other baselines. However, there is still a considerable drop from interpolation to extrapolation." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.422, + 0.468, + 0.505 + ], + "angle": 0, + "content": "
SettingPSNR ↑SSIM ↑LPIPS ↓Feat Cos Sim ↑
In.Ex.In.Ex.In.Ex.In.Ex.
Setting 128.1520.220.890.780.150.230.710.64
Setting 230.1021.210.910.820.130.200.760.67
Setting 328.6219.360.870.730.150.310.740.50
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.517, + 0.483, + 0.562 + ], + "angle": 0, + "content": "a significant performance drop, as shown in Table 4 and Figure 6a, underscoring the limitations of Gaussian-based models in handling extrapolated, unseen scenarios." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.563, + 0.483, + 0.819 + ], + "angle": 0, + "content": "Dynamic Scenes. In addition to the static scene reconstruction methods that we evaluated in Table 1, some dynamic scene reconstruction techniques have recently emerged. We conduct evaluation for Setting 2 to assess these approaches, e.g., OmniRe [10]. It organizes rigid-deformable nodes and background nodes to capture dynamic scene structures and employs SMPL [32] for non-rigid object modeling. As illustrated in Figure 6b, the rendering results reveal a significant performance gap between the training and extrapolated test cameras. In the extrapolated test views, objects such as trees and stakes lose texture and geometric details, resulting in noticeably blurry outputs, whereas the training views maintain high fidelity. As shown in Table 5, the reconstruction metrics indicate an average drop of \\(25\\%\\) when transitioning from interpolated to extrapolated settings. The results highlight the challenges of extrapolated view synthesis in dynamic scenes and the need for further research." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.83, + 0.476, + 0.847 + ], + "angle": 0, + "content": "6. Conclusions, Limitations and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Conclusions. We introduce the first benchmark enabling quantitative evaluation of extrapolated view synthesis, advancing photorealistic simulation for self-driving" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.341, + 0.906, + 0.383 + ], + "angle": 0, + "content": "Table 5. Quantitative performance of OmniRe [10] in Setting 2. The experiment uses a single traversal with dynamic objects, showing a noticeable drop from interpolation to extrapolation." + }, + { + "type": "table", + "bbox": [ + 0.525, + 0.394, + 0.89, + 0.453 + ], + "angle": 0, + "content": "
SettingPSNR ↑SSIM ↑LPIPS ↓Feat Cos Sim ↑
In.Ex.In.Ex.In.Ex.In.Ex.
Setting 219.7815.320.650.450.380.530.730.58
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.464, + 0.907, + 0.644 + ], + "angle": 0, + "content": "and robotics. Our benchmark integrates real-world multi-traversal, multi-agent, and multi-camera data, categorizes scenes into different evaluation settings, and evaluates state-of-the-art NVS models. Experimental results reveal that while some methods address specific challenges, current models demonstrate limited generalization, with significant overfitting to training views and suboptimal performance in extrapolated view synthesis. To support further research, we will release the dataset and benchmark, addressing the long-standing data scarcity and providing evaluation protocols. We believe the EUVS benchmark will catalyze meaningful advancements in self-driving and robotics innovation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.909, + 0.81 + ], + "angle": 0, + "content": "Limitations and Future Work. Our benchmark does have limitations. First, although we provide the ground truth to enable quantitative evaluation for static scenes in all settings and the foreground objects evaluation in Setting 2, we lack the foreground evaluation in Settings 1 and 3. Future work will aim to expand our work to include the dynamic object evaluation in Settings 1 and 3 using multi-agent data. Secondly, although we carefully manually selected and ensured that the test trajectory viewpoints are well covered by training trajectories, a more in-depth evaluation of how observed and unseen regions differ is left for future work." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.887 + ], + "angle": 0, + "content": "Acknowledgment. This work was supported in part through NSF grants 2238968 and 2121391, and the NYU IT High Performance Computing resources, services, and staff expertise. Yiming Li is supported by NVIDIA Graduate Fellowship (2024-2025)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28725" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.198 + ], + "angle": 0, + "content": "[1] Alexander Amini, Tsun-Hsuan Wang, Igor Gilitschenski, Wilko Schwarting, Zhijian Liu, Song Han, Sertac Karaman, and Daniela Rus. Vista 2.0: An open, data-driven simulator for multimodal sensing and policy learning for autonomous vehicles. In 2022 International Conference on Robotics and Automation (ICRA), pages 2419-2426. IEEE, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.2, + 0.484, + 0.268 + ], + "angle": 0, + "content": "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.27, + 0.484, + 0.352 + ], + "angle": 0, + "content": "[3] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.354, + 0.484, + 0.422 + ], + "angle": 0, + "content": "[4] Holger Caesar, Juraj Kabzan, Kok Seang Tan, Whye Kit Fong, Eric Wolff, Alex Lang, Luke Fletcher, Oscar Beijbom, and Sammy Omari. nuplan: A closed-loop ml-based planning benchmark for autonomous vehicles. arXiv preprint arXiv:2106.11810, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.424, + 0.484, + 0.507 + ], + "angle": 0, + "content": "[5] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3d tracking and forecasting with rich maps. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8748-8757, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.508, + 0.484, + 0.588 + ], + "angle": 0, + "content": "[6] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19467, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.592, + 0.484, + 0.661 + ], + "angle": 0, + "content": "[7] Danpeng Chen, Hai Li, Weicai Ye, Yifan Wang, Weijian Xie, Shangjin Zhai, Nan Wang, Haomin Liu, Hujun Bao, and Guofeng Zhang. Pgsr: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction. arXiv preprint arXiv:2406.06521, 2024. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.663, + 0.484, + 0.717 + ], + "angle": 0, + "content": "[8] Yurui Chen, Chun Gu, Junzhe Jiang, Xiatian Zhu, and Li Zhang. Periodic vibration gaussian: Dynamic urban scene reconstruction and real-time rendering. arXiv preprint arXiv:2311.18561, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.719, + 0.484, + 0.786 + ], + "angle": 0, + "content": "[9] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splattering from sparse multi-view images. arXiv preprint arXiv:2403.14627, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.483, + 0.856 + ], + "angle": 0, + "content": "[10] Ziyu Chen, Jiawei Yang, Jiahui Huang, Riccardo de Lutio, Janick Martinez Esturo, Boris Ivanovic, Or Litany, Zan Gojcic, Sanja Fidler, Marco Pavone, et al. Omnire: Omni urban scene reconstruction. arXiv preprint arXiv:2408.16760, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[11] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussian-pro: 3d gaussian splattering with progressive propagation. In" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "Forty-first International Conference on Machine Learning, 2024. 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.216 + ], + "angle": 0, + "content": "[12] Carlos A Diaz-Ruiz, Youya Xia, Yurong You, Jose Nino, Junan Chen, Josephine Monica, Xiangyu Chen, Katie Luo, Yan Wang, Marc Emond, et al. Ithaca365: Dataset and driving perception under repeated and challenging weather conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21383-21392, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[13] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.356 + ], + "angle": 0, + "content": "[14] Scott Ettinger, Shuyang Cheng, Benjamin Caine, Chenxi Liu, Hang Zhao, Sabeek Pradhan, Yuning Chai, Ben Sapp, Charles R Qi, Yin Zhou, et al. Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9710-9719, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.904, + 0.398 + ], + "angle": 0, + "content": "[15] Lue Fan, Hao Zhang, Qitai Wang, Hongsheng Li, and Zhaoxiang Zhang. Freesim: Toward free-viewpoint camera simulation in driving scenes, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.4, + 0.905, + 0.467 + ], + "angle": 0, + "content": "[16] Lan Feng, Quanyi Li, Zhenghao Peng, Shuhan Tan, and Bolei Zhou. Trafficgen: Learning to generate diverse and realistic traffic scenarios. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 3567-3575. IEEE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.536 + ], + "angle": 0, + "content": "[17] Jannik Fritsch, Tobias Kuehnl, and Andreas Geiger. A new performance measure and evaluation benchmark for road detection algorithms. In 16th International IEEE Conference on Intelligent Transportation Systems (ITSC 2013), pages 1693-1700. IEEE, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.539, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[18] Shenyuan Gao, Jiazhi Yang, Li Chen, Kashyap Chitta, Yihang Qiu, Andreas Geiger, Jun Zhang, and Hongyang Li. Vista: A generalizable driving world model with high fidelity and versatile controllability. arXiv preprint arXiv:2405.17398, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.609, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[19] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[20] Huasong Han, Kaixuan Zhou, Xiaoxiao Long, Yusen Wang, and Chunxia Xiao. Ggs: Generalizable gaussian splatting for lane switching in autonomous driving. arXiv preprint arXiv:2409.02382, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.72, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[21] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[22] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[23] Sungwon Hwang, Min-Jung Kim, Taewoong Kang, Jayeon Kang, and Jaegul Choo. Vegs: View extrapolation of urban scenes in 3d gaussian splatting using learned priors. arXiv preprint arXiv:2407.02945, 2024. 2, 3, 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28726" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.176 + ], + "angle": 0, + "content": "[24] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5322-5332, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.484, + 0.234 + ], + "angle": 0, + "content": "[25] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.484, + 0.292 + ], + "angle": 0, + "content": "[26] Mustafa Khan, Hamidreza Fazlali, Dhruv Sharma, Tongtong Cao, Dongfeng Bai, Yuan Ren, and Bingbing Liu. Autosplat: Constrained gaussian splatting for autonomous driving scene reconstruction. arXiv preprint arXiv:2407.02598, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.293, + 0.484, + 0.362 + ], + "angle": 0, + "content": "[27] Seung Wook Kim, Jonah Philion, Antonio Torralba, and Sanja Fidler. Drivegan: Towards a controllable high-quality neural simulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5820-5829, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.363, + 0.484, + 0.434 + ], + "angle": 0, + "content": "[28] Quanyi Li, Zhenghao Mark Peng, Lan Feng, Zhizheng Liu, Chenda Duan, Wenjie Mo, and Bolei Zhou. ScenarioNet: Open-source platform for large-scale traffic scenario simulation and modeling. Advances in neural information processing systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.435, + 0.484, + 0.503 + ], + "angle": 0, + "content": "[29] Yiming Li, Dekun Ma, Ziyan An, Zixun Wang, Yiqi Zhong, Siheng Chen, and Chen Feng. V2x-sim: Multi-agent collaborative perception dataset and benchmark for autonomous driving. IEEE Robotics and Automation Letters, 7(4): 10914-10921, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.484, + 0.587 + ], + "angle": 0, + "content": "[30] Yiming Li, Zhiheng Li, Nuo Chen, Moonjun Gong, Zonglin Lyu, Zehong Wang, Peili Jiang, and Chen Feng. Multiagent multitraversal multimodal self-driving: Open mars dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22041-22051, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.589, + 0.484, + 0.659 + ], + "angle": 0, + "content": "[31] Yiming Li, Zehong Wang, Yue Wang, Zhiding Yu, Zan Gojcic, Marco Pavone, Chen Feng, and Jose M. Alvarez. Memorize what matters: Emergent scene decomposition from multitraverse. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.484, + 0.717 + ], + "angle": 0, + "content": "[32] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1-248:16, 2015. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[33] Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3061-3070, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.484, + 0.843 + ], + "angle": 0, + "content": "[34] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[35] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.163 + ], + "angle": 0, + "content": "[36] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.906, + 0.232 + ], + "angle": 0, + "content": "[37] Matthew Pitropov, Danson Evan Garcia, Jason Rebello, Michael Smart, Carlos Wang, Krzysztof Czarnecki, and Steven Waslander. Canadian adverse driving conditions dataset. The International Journal of Robotics Research, 40 (4-5):681-690, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.906, + 0.328 + ], + "angle": 0, + "content": "[38] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dólar, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.329, + 0.906, + 0.383 + ], + "angle": 0, + "content": "[39] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.385, + 0.906, + 0.455 + ], + "angle": 0, + "content": "[40] Shital Shah, Debadeepta Dey, Chris Lovett, and Ashish Kapoor. Airsim: High-fidelity visual and physical simulation for autonomous vehicles. In Field and Service Robotics: Results of the 11th International Conference, pages 621-635. Springer, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.906, + 0.51 + ], + "angle": 0, + "content": "[41] Xi Shi, Lingli Chen, Peng Wei, Xi Wu, Tian Jiang, Yonggang Luo, and Lecheng Xie. Dhgs: Decoupled hybrid gaussian splatting for driving scene. arXiv preprint arXiv:2407.16600, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.906, + 0.58 + ], + "angle": 0, + "content": "[42] Shuhan Tan, Kelvin Wong, Shenlong Wang, Sivabalan Manivasagam, Mengye Ren, and Raquel Urtasun. Scenegen: Learning to generate realistic traffic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 892-901, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.906, + 0.65 + ], + "angle": 0, + "content": "[43] Izzeddin Teeti, Valentina Musat, Salman Khan, Alexander Rast, Fabio Cuzzolin, and Andrew Bradley. Vision in adverse weather: Augmentation using cyclegans with various object detectors for robust perception in autonomous racing. arXiv preprint arXiv:2201.03246, v3, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.906, + 0.733 + ], + "angle": 0, + "content": "[44] Adam Tondersski, Carl Lindstrom, Georg Hess, William Ljungbergh, Lennart Svensson, and Christoffer Petersson. Neurad: Neural rendering for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14895-14904, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.734, + 0.906, + 0.79 + ], + "angle": 0, + "content": "[45] Christopher Wewer, Kevin Raj, Eddy Ilg, Bernt Schiele, and Jan Eric Lenssen. latentsplat: Autoencoding variational gaussians for fast generalizable 3d reconstruction. arXiv preprint arXiv:2403.16292, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.791, + 0.906, + 0.873 + ], + "angle": 0, + "content": "[46] Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, et al. Argoverse 2: Next generation datasets for self-driving perception and forecasting. arXiv preprint arXiv:2301.00493, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[47] Chenming Wu, Jiadai Sun, Zhelun Shen, and Liangjun Zhang. Mapnerf: Incorporating map priors into neural ra" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28727" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.483, + 0.134 + ], + "angle": 0, + "content": "diance fields for driving view simulation. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7082-7088. IEEE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.207 + ], + "angle": 0, + "content": "[48] Ke Wu, Kaizhao Zhang, Zhiwei Zhang, Shanshuai Yuan, Muer Tie, Julong Wei, Zijun Xu, Jieru Zhao, Zhongxue Gan, and Wenchao Ding. Hgs-mapping: Online dense mapping using hybrid gaussian representation in urban scenes. arXiv preprint arXiv:2403.20159, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.483, + 0.29 + ], + "angle": 0, + "content": "[49] Zirui Wu, Tianyu Liu, Liyi Luo, Zhide Zhong, Jianteng Chen, Hongmin Xiao, Chao Hou, Haozhe Lou, Yuantao Chen, Runyi Yang, et al. Mars: An instance-aware, modular and realistic simulator for autonomous driving. In CAAI International Conference on Artificial Intelligence, pages 3-15. Springer, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.483, + 0.348 + ], + "angle": 0, + "content": "[50] Bernhard Wymann, Eric Espie, Christophe Guionneau, Christos Dimitrakakis, Rémi Coulom, and Andrew Sumner. Torcs, the open racing car simulator. Software available at http://torcs.sourceforge.net, 4(6):2, 2000. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.483, + 0.404 + ], + "angle": 0, + "content": "[51] Danfei Xu, Yuxiao Chen, Boris Ivanovic, and Marco Pavone. Bits: Bi-level imitation for traffic simulation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2929-2936. IEEE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.483, + 0.502 + ], + "angle": 0, + "content": "[52] Runsheng Xu, Xin Xia, Jinlong Li, Hanzhao Li, Shuo Zhang, Zhengzhong Tu, Zonglin Meng, Hao Xiang, Xiaoyu Dong, Rui Song, et al. V2v4real: A real-world large-scale dataset for vehicle-to-vehicle cooperative perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13712-13722, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.483, + 0.56 + ], + "angle": 0, + "content": "[53] Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians for modeling dynamic urban scenes. arXiv preprint arXiv:2401.01339, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.483, + 0.644 + ], + "angle": 0, + "content": "[54] Chen Yang, Peihao Li, Zanwei Zhou, Shanxin Yuan, Bingbing Liu, Xiaokang Yang, Weichao Qiu, and Wei Shen. Ner-fvs: Neural radiance fields for free view synthesis via geometry scaffolds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16549-16558, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.483, + 0.716 + ], + "angle": 0, + "content": "[55] Jiawei Yang, Boris Ivanovic, Or Litany, Xinshuo Weng, Seung Wook Kim, Boyi Li, Tong Che, Danfei Xu, Sanja Fidler, Marco Pavone, et al. Emernerf: Emergent spatial-temporal scene decomposition via self-supervision. arXiv preprint arXiv:2311.02077, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[56] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1389-1399, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[57] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19447-19456, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[58] Zhongrui Yu, Haoran Wang, Jinze Yang, Hanzhang Wang, Zeke Xie, Yunfeng Cai, Jiale Cao, Zhong Ji, and Mingming" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "Sun. Sgd: Street view synthesis with gaussian splatting and diffusion prior. arXiv preprint arXiv:2403.20079, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[59] Dongbin Zhang, Chuming Wang, Weitao Wang, Peihao Li, Minghan Qin, and Haoqian Wang. Gaussian in the wild: 3d gaussian splatting for unconstrained image collections. arXiv preprint arXiv:2403.15704, 2024. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[60] Jian Zhang, Yuanqing Zhang, Huan Fu, Xiaowei Zhou, Bowen Cai, Jinchi Huang, Rongfei Jia, Binqiang Zhao, and Xing Tang. Ray priors through reprojection: Improving neural radiance fields for novel view extrapolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18376-18386, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[61] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.348, + 0.905, + 0.416 + ], + "angle": 0, + "content": "[62] Guosheng Zhao, Chaojun Ni, Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, Boyuan Wang, Youyi Zhang, Wenjun Mei, and Xingang Wang. Drivedreamer4d: World models are effective data machines for 4d driving scene representation. arXiv preprint arXiv:2410.13571, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.905, + 0.487 + ], + "angle": 0, + "content": "[63] Ziyuan Zhong, Davis Rempe, Danfei Xu, Yuxiao Chen, Sushant Veer, Tong Che, Baishakhi Ray, and Marco Pavone. Guided conditional diffusion for controllable traffic simulation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 3560-3566. IEEE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.489, + 0.905, + 0.57 + ], + "angle": 0, + "content": "[64] Hongyu Zhou, Jiahao Shao, Lu Xu, Dongfeng Bai, Weichao Qiu, Bingbing Liu, Yue Wang, Andreas Geiger, and Yiyi Liao. Hugs: Holistic urban 3d scene understanding via gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21336-21345, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.572, + 0.905, + 0.668 + ], + "angle": 0, + "content": "[65] Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Zehao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21676-21685, 2024. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.67, + 0.905, + 0.753 + ], + "angle": 0, + "content": "[66] Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21634-21643, 2024. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.753 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28728" + } + ] +] \ No newline at end of file diff --git a/2025/Extrapolated Urban View Synthesis Benchmark/e65fa038-3ae6-4620-9f47-ad1a63db0abf_origin.pdf b/2025/Extrapolated Urban View Synthesis Benchmark/e65fa038-3ae6-4620-9f47-ad1a63db0abf_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..06181362cc629a2c061c0539adbeed9f687e0b7f --- /dev/null +++ b/2025/Extrapolated Urban View Synthesis Benchmark/e65fa038-3ae6-4620-9f47-ad1a63db0abf_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b8ce94aac7014df0433e2b519f15f61fd2c66e37029b18860bbc5656eccb0041 +size 35722908 diff --git a/2025/Extrapolated Urban View Synthesis Benchmark/full.md b/2025/Extrapolated Urban View Synthesis Benchmark/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a898e81f6cd1dbae84726c2ffe8f6e500b5a8e11 --- /dev/null +++ b/2025/Extrapolated Urban View Synthesis Benchmark/full.md @@ -0,0 +1,270 @@ +# Extrapolated Urban View Synthesis Benchmark + +Xiangyu Han $^{1,3*}$ Zhen Jia $^{1*}$ Boyi Li $^{2}$ Yan Wang $^{2}$ Boris Ivanovic $^{2}$ Yurong You $^{2}$ Lingjie Liu $^{3}$ Yue Wang $^{2,4}$ Marco Pavone $^{2,5}$ Chen Feng $^{1}$ Yiming Li $^{1,2\dagger}$ $^{1}$ NYU ${}^{2}$ NVIDIA ${}^{3}$ UPenn ${}^{4}$ USC ${}^{5}$ Stanford + +https://ai4ce.github.io/EUVS-Benchmark + +![](images/2750e1de8e2e50beca42baa1099958d524968a043102410c836203c281dd0c6d.jpg) +Figure 1. Our key contributions. Previous evaluations for urban view synthesis have primarily focused on interpolated poses, as the lack of ground truth data has made it challenging to evaluate extrapolated poses. We address this gap by providing real-world data that enables both quantitative and qualitative evaluations of extrapolated view synthesis in urban scenes. The quantitative results reveal a significant performance drop in 3D Gaussian Splatting [25] when handling extrapolated views, highlighting the need for more robust NVS methods. + +![](images/fc9a854c63e2bee516fd917f4f42336948855da87e0a6a0784af664d4c075558.jpg) + +![](images/ce57cae050600ffb83d7b56b896254455bb4ea9f953c02cde065b1107a85d519.jpg) +GT is available in previous benchmark + +![](images/8e20560ca42402c382e700e53a8b707c1a96b5c1dba871c7051a5da554f6c2af.jpg) + +![](images/60d60ba9ab5eafe33b3d1185866f863d630b62c9e743fb06fe8a2c06ff73990e.jpg) +GT is available in our benchmark + +![](images/59d0e3a7f01289ad2105a343148fef98d14a6eaa7349ffc290cc04ae639af593.jpg) + +# Abstract + +Photorealistic simulators are essential for the training and evaluation of vision-centric autonomous vehicles (AVs). At their core is Novel View Synthesis (NVS), a crucial capability that generates diverse unseen viewpoints to accommodate the broad and continuous pose distribution of AVs. Recent advances in radiance fields, such as 3D Gaussian Splatting, achieve photorealistic rendering at real-time speeds and have been widely used in modeling large-scale driving scenes. However, their performance is commonly evaluated using an interpolated setup with highly correlated training and test views. In contrast, extrapolation, where test views largely deviate from training views, remains underexplored, limiting progress in generalizable simulation technology. To address this gap, we leverage publicly available AV datasets with multiple traversals, + +multiple vehicles, and multiple cameras to build the first Extrapolated Urban View Synthesis (EUVS) benchmark. Meanwhile, we conduct both quantitative and qualitative evaluations of state-of-the-art NVS methods across different evaluation settings. Our results show that current NVS methods are prone to overfitting to training views. Besides, incorporating diffusion priors and improving geometry cannot fundamentally improve NVS under large view changes, highlighting the need for more robust approaches and large-scale training. We have released the data to help advance self-driving and urban robotics simulation technology. + +# 1. Introduction + +The development of vision-centric autonomous vehicles (AVs) relies heavily on photorealistic simulators, which provide controlled, reproducible, and scalable environments for training and evaluation of driving models [13, 44, 56]. These simulators enable AVs to learn and adapt to a variety + +of real-world scenarios, from crowded urban streets to adverse weather conditions, without the logistical and safety concerns of physical road testing. At the heart of these simulators is the capability for Novel View Synthesis (NVS)—a key technology that generates realistic images of unseen viewpoints, simulating the continuous changes in perspective that occur as AVs navigate through urban environments. + +Recent advancements in radiance fields, particularly methods based on 3D Gaussian Splatting [25], have significantly improved the realism and efficiency of NVS. These approaches [8, 53, 64, 66] can produce photorealistic renderings at real-time speeds, making them highly attractive for large-scale driving scene simulation. However, despite their impressive results, the evaluation of NVS methods has predominantly focused on interpolated scenarios, where training and test viewpoints are closely related. While interpolation tests are valuable for assessing local consistency, they fall short in addressing the more critical challenge of extrapolation—where test viewpoints differ significantly from the training data. As shown in Figure 1, the interpolation test set demonstrates strong performance, with metrics such as PSNR, SSIM, and LPIPS remaining very close to the training set values. In contrast, the extrapolation test set, which includes additional translation and rotation changes relative to the training set, exhibits notable drops in performance. Specifically, the metric decreases relative to the training set are $28\%$ for PSNR, $22\%$ for SSIM, and $50\%$ for LPIPS. These results underscore the urgent need to explore and advance extrapolated view synthesis in complex urban scenes, as real-world driving often involves encountering scenarios with significant viewpoint shifts and diverse spatial transformations that deviate from training distributions. Several recent studies [20, 23] have investigated the generalization capabilities of NVS in 3D Gaussian Splatting. Although they show promising qualitative results, there is no comprehensive quantitative analysis due to the absence of standardized datasets. Moreover, their evaluations are primarily limited to specific scenarios or use cases, without investigating varying evaluation settings based on the degree of extrapolation. This gap underscores the urgent need for a benchmark that offers diverse and challenging datasets, enabling a rigorous and systematic assessment of NVS methods. + +To establish a common platform for assessing the robustness of NVS methods, we introduce a comprehensive benchmark for quantitatively and qualitatively evaluating extrapolated novel view synthesis in large-scale urban scenes. Our benchmark leverages publicly available datasets, including NuPlan [4], MARS [30], and Argoverse2 [46], which feature multi-traversal, multi-agent and multi-camera sensory recordings. Multi-traversal data consists of asynchronous traversals of the same location, while multi-agent data refers to data collected simultaneously + +from multiple vehicles within the same area. These data provide diverse camera poses within a 3D scene, enabling the training and evaluation of extrapolated view synthesis in outdoor environments. For the experimental setup, we define three evaluation settings: (1) translation only, (2) rotation only, and (3) translation + rotation, as shown in Figure 4. In autonomous driving scenarios, Setting 1 corresponds to maneuvers such as lane changes, Setting 2 involves switching between cameras facing different directions, and Setting 3 addresses complex intersections, such as crossroads with diverse traversal paths. These settings represent common challenges in autonomous driving, and addressing them enables the synthesis of complete scenes from sparse image observations. + +We conduct pose estimation and sparse reconstruction using COLMAP [39], which facilitates the initialization of Gaussian Splatting. We then evaluate state-of-the-art Gaussian Splatting-based approaches across each evaluation setting, identifying performance gaps both qualitatively and quantitatively in extrapolated urban view synthesis. + +In summary, our main contributions are as follows: + +- We initiate the first comprehensive quantitative and qualitative study on the Extrapolated Urban View Synthesis (EUVS) problem, supported by a robust evaluation framework that categorizes evaluation settings (translation-only, rotation-only, and translation + rotation) while assessing performance using diverse metrics, including reconstruction accuracy and visual fidelity. +- We construct a novel dataset by integrating multitraversal, multi-agent, and multi-camera data from publicly available resources, totaling 90,810 frames across 345 videos. Our dataset effectively addresses the limitations of existing benchmarks, enabling rigorous and robust evaluation for extrapolated urban view synthesis. +- We benchmark state-of-the-art Gaussian Splitting-based and NeRF-based models and analyze key factors that influence the performance of extrapolated NVs, laying a solid foundation for future advancements in this challenging task. Data and code will be released upon acceptance. + +# 2. Related Works + +Extrapolated View Synthesis. Extrapolated view synthesis aims to generate novel views beyond observed perspectives, addressing challenges in visual coherence for unseen regions. RapNeRF [60] proposes a random ray-casting policy that enables training on unseen views based on visible ones. Following work [54] enhances this approach by incorporating holistic priors. Besides, some generalizable models [6, 9, 45] have emerged, capable of generating extrapolated novel views from a limited number of input images. While these methods are designed for indoor scenes, several works address extrapolated view synthesis in outdoor driving scenarios, which typically involve forward-facing + +cameras and unbounded environments. To tackle the Setting 1 challenge in our benchmark and address the scarcity of lane change data, GGS [20] introduces a novel virtual lane generation module. In parallel, AutoSplat [26] tackles lane change in dynamic scenes by applying geometric and reflected consistency constraints. To address the Setting 1 or 2 challenge, FreeSim [15], VEGS [23], and SGD [58] enhance 3DGS [25] with diffusion priors to improve generalization ability. Yet existing methods suffer from two major limitations: (1) a lack of real data for quantitative evaluation, which confines them to quantitative analysis, and (2) a narrow focus on a specific setting in our benchmark, preventing a comprehensive and systematic exploration. + +3D Gaussian Splatting. Recent advances in radiance fields, particularly NeRF [34] and 3DGS [25], have garnered significant attention due to their impressive advancements in NVs. NeRF employs an implicit representation through a multi-layer perceptron (MLP). Furthermore, 3DGS explicitly represents scenes using anisotropic 3D Gaussian ellipsoids, enabling high-quality real-time rendering. Several works have addressed issues such as difficulties with reflective surfaces [24], aliasing [57], etc. However, urban scenes introduce unique challenges due to their unbounded and dynamic nature. To address the challenge, several works separate dynamic and static elements in the scene by leveraging a composite dynamic Gaussian graph [53, 66], optical flow prediction [55, 64], etc. PVG [8] presents a unified representation model that simultaneously incorporates both dynamic and static components without relying on priors. To achieve realistic geometry and efficient rendering, 2DGS [22] collapses 3D Gaussians onto 2D planes, while hybrid approaches [41, 48] combine different Gaussians to better capture region-specific features. In summary, current urban NVS methods primarily focus on effectively handling dynamic elements and enhancing geometry representation, while the challenge of extrapolated view synthesis remains largely underexplored. + +Autonomous Driving Simulators. Current simulators focus on three key challenges: parameter initialization [16, 42], traffic simulation [28, 51, 63], and sensor simulation [1, 13, 18, 27, 56, 62]. Sensor simulation is crucial for generating realistic sensory data that AVs depend on for perception and decision-making. Early sensor simulators [13, 40, 50] provide simulated environments that are valuable for research but lack visual realism. Recent studies have focused on data-driven simulators that extract data from real-world driving logs, creating more realistic and adaptable environments. These methods can be classified into two categories: generation-based and reconstruction-based approaches. The former rely on inputs such as text, video, and other data sources for simulation, supported by world models or prior knowledge [18, 27, 62]. Reconstruction-based simulations leverage real-world data + +to ensure both visual fidelity and geometric consistency [47, 49, 56]. UniSim [56] is a pioneering example of this approach, utilizing NeRF-based scene representation to create dynamic scenes with geometric information that are both editable and controllable. Extrapolated view synthesis is essential for these simulators, as it enables the generation of realistic and consistent views from diverse angles. + +Autonomous Driving Datasets. High-quality datasets play a vital role in advancing autonomous driving research. The KITTI dataset [19], released in 2012, marked a major milestone, significantly accelerating advances in AVs [17, 33, 43]. Since then, many influential autonomous vehicle datasets have been developed to tackle challenges like adverse weather conditions [37], multimodal fusion [3, 4], repeated driving [4, 12], collaborative driving [29, 30, 52], and motion prediction [5, 14, 46], etc. We leverage publicly available datasets with multi-traversal, multi-agent, and multi-camera recordings, enabling a comprehensive and robust evaluation of extrapolated urban view synthesis. + +# 3. The EUVS Benchmark + +# 3.1. Dataset Curation + +We leverage three publicly available, community-verified autonomous driving datasets—nuPlan [4], Argoverse 2 [46], and MARS [30]—to enhance adoption and foster trust through their established reliability. nuPlan [4] provides 1,200 hours of driving data from four cities, serving as the first large-scale planning benchmark. Argoverse 2 [46] supports multimodal perception and forecasting with 1,000 annotated 3D scenarios, 20,000 unlabeled lidar sequences, and 250,000 motion forecasting cases. MARS [30] enables collaborative driving research with multi-agent and multitraversal scenarios. The original datasets were not designed for evaluating extrapolated view synthesis, which requires our significant labor in data processing— $300+$ hours of manual traversal selection and $800+$ hours of computing to run COLMAP. By integrating these datasets, our benchmark enables the evaluation of view synthesis across diverse and realistic urban environments under varying conditions. Figure 3 illustrates the distribution of the integrated datasets. + +# 3.2. Evaluation Framework + +To systematically assess model performance in extrapolated urban view synthesis, our evaluation framework incorporates three evaluation settings and three data configurations. Data configurations include multi-traversal, multi-agent, and multi-camera, while evaluation settings are categorized into (1) translation only, (2) rotation only, and (3) translation + rotation, as illustrated in Figure 4. + +Setting 1. The translation-only experimental setup involves scenarios where the vehicle's position shifts without any change in orientation. This scenario is commonly observed + +![](images/1709a7854bc488c0ae02b1c9a19f4299eefca5241f7ba846a7195561fb2aa78d.jpg) +Figure 2. Dataset visualization. Our dataset features diverse scenes across various locations in different cities, sourced from multiple datasets. Typical driving scenarios include maneuvers such as lane changes, cross intersections, and T-junctions. Top: Each column displays images captured at the same location by different agents or traversals. Bottom: Each image displays the COLMAP points at a specific location, along with the corresponding camera poses. + +![](images/686306d5c9869386ddea58a0a3dd01fc160eb04be8c4e3c8449e63501b05a228.jpg) +Figure 3. Dataset distribution. Our dataset comprises 90,810 frames distributed over 104 cases, capturing a diverse array of multi-traversal paths, multi-agent interactions, and multi-camera perspectives across varying evaluation settings. + +in lane changes. We use traversals from different lanes in multi-traversal data, focusing on the three front cameras. The data is sourced from nuPlan [4] and Argoverse 2 [46]. + +Setting 2. The second setting, rotation only, evaluates models on views with significant orientation changes. In vision-centric autonomous vehicles, this corresponds to transitions between cameras capturing different directions. We leverage multi-camera data from nuPlan [4], training on three forward-facing and three rear-facing cameras, while evaluating on two side-facing cameras. + +Setting 3. The third setting, combining translation and rotation, includes both positional shifts and orientation changes, posing the greatest challenge for NVS. To address this, we utilize multi-traversal driving data collected from the same location but across different traversal routes. For example, the training and test sets may include routes that approach an intersection from different directions. Typical route combinations feature scenarios such as intersections, + +T-junctions, and Y-junctions, as shown in Figure 2, ensuring diversity and comprehensive evaluation. The data for this setting is from MARS [30] and Argoverse 2 [46]. + +# 3.3. Algorithm Overview + +Vanilla 3D Gaussian Splatting. 3D Gaussian Splatting (3DGS) [25] leverage 3D Gaussians to explicitly represent the scene, which achieves high quality while offering real-time rendering by avoiding unnecessary computation in the empty space. Building on this, 3DGM [31] leverages multitraversal consensus to differentiate transient and permanent elements, enabling joint 2D segmentation and 3D mapping without using any human supervision. + +Planar-based and Geometry Refined Gaussian Splat-ting. GaussianPro [11] builds on 3DGS [25] by introducing multi-frame geometric optimization, which guides the densi-sification of 3D Gaussians, enhancing scene consistency in complex geometries. It further refines geometry by encour + +![](images/dce74e08eb5f1a159d2d3546b7f157deb1bfd9ffda30b71b6febb0bc60b72ed1.jpg) +(a) Evaluation Setting 1: Translation Only + +![](images/dcd886910944c08c09424ea1e29b545ff39f8ff761dbbd9bbe7bd98c2d7a4e29.jpg) + +![](images/cb960acf25f7e663d00c9736e853be8325d6acbf1a823ac3560a89ba1df9b371.jpg) +(b) Evaluation Setting 2: Rotation Only +(c) Evaluation Setting 3: Translation + Rotation +Figure 4. Qualitative and quantitative results across three evaluation settings. The performance drop from interpolation to extrapolation is significant in both qualitative and quantitative comparison. Different testing settings have distinct scenario characteristics, enabling the evaluation of a method's capabilities from various aspects, thus systematically assessing the overall performance of reconstruction algorithms, including geometric accuracy, hallucination ability, view consistency, and depth precision, etc. + +aging Gaussian primitives to adopt flat structures. Similarly, 2DGS [22] projects the 3D volume into a set of 2D oriented planar Gaussian disks, enabling high-fidelity surface reconstruction. PGSR [7] introduces an unbiased depth rendering method and integrates single-view geometric, multiview photometric, and geometric regularization techniques to improve global geometry accuracy. + +Gaussian Splatting with Diffusion Priors. VEGS [23] introduces a novel view generalization approach that harnesses pre-extracted surface normals to align 3D Gaussians while generating augmented camera views guided by diffusion priors. These diffusion priors serve a dual purpose: providing denoising loss guidance and supervising the training of augmented cameras. This process effectively mitigates floating artifacts and fragmented geometries, resulting in more accurate and coherent 3D representations. + +Feature-Enhanced Gaussian Splatting. Feature 3DGS [65] extends 3D Gaussian Splatting with a Parallel N-dimensional Gaussian Rasterizer, allowing simultaneous rendering of radiance fields and high-dimensional semantic features. By embedding semantic features directly into 3D + +Gaussians, the approach enhances optimization, enabling better correspondence with scene semantics and achieving more detailed and accurate spatial representations. + +NeRF-based Method. Instant-NGP [35] uses a multiresolution hash encoding to map spatial coordinates into compact latent representations via hash tables. It efficiently encodes high-frequency details by combining trainable feature vectors with interpolation, enabling adaptive and scalable input encodings. Zip-NeRF [2] leverages multisampling with isotropic Gaussians for scale-aware features and introduces a smooth anti-aliasing loss to address z-aliasing. In addition, it incorporates a novel distance normalization technique to better manage close and distant objects, achieving high-quality rendering and fast training. + +# 4. Experiment + +# 4.1. Experiment Setup + +Implementation Details. All 3DGS-based methods are initialized only from sparse points obtained by COLMAP [39] and exclude lidar points. We employ Grounded-SAM-2 + +Table 1. Quantitative rendering results across three evaluation settings. In. denotes interpolation, while Ex. represents extrapolation. Different baselines excel in different settings, reflecting the comprehensiveness and completeness of the evaluation protocol. + +
MethodPSNR↑SSIM↑LPIPS↓Feat Cos Sim ↑
In.Ex.DropIn.Ex.DropIn.Ex.DropIn.Ex.Drop
Setting 13DGS [25]21.3616.3723.4%0.82750.720313.0%0.20410.259927.3%0.68280.603911.6%
3DGM [31]20.9616.3522.0%0.82930.724812.6%0.20030.254226.9%0.68020.608710.5%
GSPro [11]21.5116.3923.8%0.83100.718913.5%0.18040.245035.8%0.70810.613013.4%
VEGS [23]21.2615.8825.3%0.81070.704713.1%0.24980.306222.6%0.63230.552112.7%
PGSR [7]20.5716.3220.7%0.81040.710212.4%0.22620.273320.8%0.65150.584810.2%
2DGS [22]20.8716.3021.9%0.80760.710312.0%0.24380.289018.5%0.62560.56449.8%
Feature 3DGS [65]21.0216.0123.8%0.80960.724310.5%0.18760.257537.3%0.69580.612212.0%
Zip-NeRF [2]19.6814.0628.6%0.78560.691712.0%0.27110.341826.1%0.63180.554212.3%
Instant-NGP [35]18.7712.6532.6%0.76310.625218.1%0.48740.593821.8%0.54650.483711.5%
AVERAGE20.6715.5924.6%0.80830.704612.8%0.25010.313425.3%0.65050.577411.2%
Setting 23DGS [25]25.7519.5324.2%0.87660.751114.3%0.15360.266873.7%0.73270.631913.8%
3DGM [31]25.7518.7827.1%0.87860.746415.0%0.15560.281380.8%0.72780.634412.8%
GSPro [11]26.4219.3926.6%0.88210.747015.3%0.13290.224669.0%0.75230.648713.8%
VEGS [23]24.5423.334.9%0.83660.79495.0%0.23010.281122.2%0.65950.61337.0%
PGSR [7]24.5318.3825.1%0.86120.711917.3%0.15550.253262.8%0.72000.581719.2%
2DGS [22]25.1518.8325.1%0.85780.720416.0%0.17560.291766.1%0.68980.578516.1%
Feature 3DGS [65]24.9119.5921.4%0.88000.786410.6%0.13770.227865.4%0.74270.646413.0%
Zip-NeRF [2]29.0617.3640.3%0.86600.671522.5%0.20780.358272.4%0.74790.584321.9%
Instant-NGP [35]25.6117.1533.0%0.85960.721216.1%0.33400.517154.8%0.72540.618214.8%
AVERAGE25.7519.1525.6%0.86650.739014.7%0.18700.300262.5%0.72200.615314.8%
Setting 33DGS [25]21.2214.9929.4%0.85500.716916.1%0.22520.405079.8%0.70020.477431.8%
3DGM [31]20.6214.6029.2%0.85430.723315.3%0.22540.404979.6%0.68740.467232.0%
GSPro [11]21.5814.8231.3%0.86340.699619.0%0.20100.387792.9%0.70930.454136.0%
VEGS [23]21.1314.2532.6%0.82660.647521.7%0.23590.442287.5%0.67850.444234.5%
PGSR [7]19.6014.1727.7%0.82380.698415.2%0.29340.436348.7%0.58670.378735.5%
2DGS [22]17.3511.3634.5%0.75680.544728.0%0.42960.545927.1%0.35520.232734.5%
Feature 3DGS [65]21.8814.3334.5%0.86430.638626.1%0.14000.3816172.6%0.74110.466937.0%
Zip-NeRF [2]20.6114.4230.0%0.83830.656521.7%0.21970.4546106.9%0.71080.364548.7%
Instant-NGP [35]19.6314.3926.7%0.81790.710413.1%0.49560.659233.0%0.60830.415731.7%
AVERAGE20.4014.1530.6%0.83340.670719.5%0.27400.457570.0%0.64190.411335.9%
+ +Table 2. Quantitative comparison of depth evaluation in extrapolated views in Setting 1. 3DGM [31] demonstrates superior performance on most evaluation metrics, while VEGS [23] and GSPro [11] excel in SqRel and Delta1, respectively. + +
BaselineAbsRel ↓RMSE ↓SqRel ↓Delta1 ↑Delta2 ↑Delta3 ↑
3DGS [25]0.36114.4410.410.6490.8240.895
3DGM [31]0.30113.938.9060.6510.8460.915
PGSR [7]0.36617.5721.500.7590.8340.883
GSPro [11]0.35519.6632.010.6430.8390.909
VEGS [23]0.36815.008.3980.4410.6910.827
+ +[38] to mask out pixels corresponding to potentially movable objects during both training and evaluation and also exclude them from the initialization of 3DGS. + +Evaluation Metrics. We use three widely-used metrics to evaluate visual quality: peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and learned perceptual image patch similarity (LPIPS) [61]. We also employ DINOv2 [36] feature cosine similarity to evaluate image quality in latent space. For geometry evaluation, we use depth metrics, including RMSE and $\delta_{1.25}$ . Additionally, we incorporate the Fréchet Inception Distance (FID) [21] to + +Table 3. Perceptual quality metric of VEGS [23]. Our benchmark also incorporates the Fréchet Inception Distance (FID) to evaluate the inpainting ability of models with generative priors. + +
Setting 1Setting 2Setting 3AVERAGE
FID (In.) ↓46.333.546.142.0
FID (Ex.) ↓87.467.2132.495.7
FID (Drop)89%101%187%126.7%
+ +evaluate the inpainting ability of models with strong priors. + +# 4.2. Experimental Results + +Table 1 presents the quantitative results across Settings 1-3, while Figure 5 illustrates the qualitative outcomes on the extrapolated test set. Table 3 and Table 2 present the perceptual quality and depth metrics. The results indicate that, while the metrics perform relatively well in the interpolated test set, there is a significant drop in performance in the extrapolated test set across all baselines. + +Setting 1: Translation-only. In Setting 1, training views fully cover test views with moderate translational changes. (1) Results show a consistent drop in performance from + +![](images/5481253f8bc9af45aa50f2778f2efc58a4a90a023e48c8b55a1490a87f41c2d2.jpg) +Setting 1: Translation Only + +![](images/544945f12614aa72971e9dcd893b56c88f38ebbcc03e182571dfcb3fd272fe91.jpg) +Setting 2: Rotation Only + +![](images/a281b5f540c7d598c32ec069ca3181a039727fc9a85a60f86c9fbaf2d2c6fcb8.jpg) +Setting 3: Rotation + Translation +Figure 5. Qualitative comparison of extrapolated view synthesis across different settings. For each setting, results from different methods are compared against the ground truth. Red boxes highlight areas where methods are limited in capturing fine details, such as road surfaces, sky regions, or object boundaries, demonstrating the challenges faced by each approach under varying movement complexities. + +interpolation to extrapolation across all metrics, highlighting the challenge of generalization to views with variations. Relative drops vary by metric: PSNR falls $23 - 25\%$ (e.g., GSPro [11]: $21.51 \rightarrow 16.39$ ), with SSIM and LPIPS showing similar declines. (2) On the extrapolative test set, methods perform comparably: GSPro tops PSNR (16.39) and minimizes LPIPS (0.2450), while 3DGM [31] leads in SSIM (0.7248). Both yield similar feature cosine similarity; VEGS [23] and PGSR [7] lag slightly. Results suggest that although GSPro marginally outperforms others, differences are small, underscoring the need for improved solutions. + +Setting 2: Rotation-only. In Setting 2, training views provide extensive coverage of the surrounding scene. However, most methods show poor generalization, with PSNR dropping by $22.75\%$ . Rotation changes are particularly challenging in textured regions (e.g., trees and intricate details), often resulting in blurring, while regions perpendicular to the vehicle remain difficult to capture. Additionally, distant areas pose reconstruction challenges, frequently leading to missing buildings and sky blackouts. Among the evaluated baselines, VEGS [23] and GSPro [11] stand out as the best-performing baselines in this setting: VEGS utilizes its diffusion prior for inpainting and refining missing regions, while GSPro's robust geometry handling improves generalization. + +Setting 3: Translation + Rotation. In Setting 3, the view changes are the largest. (1) All methods exhibit notable performance drops from interpolation to extrapolation across metrics. For instance, 3DGS [25] experiences a PSNR + +drop of $29.4\%$ ( $21.22 \rightarrow 14.99$ ), while GSPro [11] undergoes a similar drop of $31.3\%$ ( $21.58 \rightarrow 14.82$ ). (2) Feature 3DGS [65] and 3DGM [31] stand out as leading methods, excelling in LPIPS (0.3816) and SSIM (0.7233), respectively. However, overall performance remains limited, with PSNR consistently falling below 15, highlighting significant room for improvement in generating high-fidelity outputs for extrapolated views. + +# 5. Discussions + +Lighting Inconsistency Handling. Lighting inconsistencies are a widespread challenge in multi-traversal datasets due to varying illumination and weather conditions. To address this, we have taken specific measures. (1) Importantly, our multi-traversal dataset is manually curated to ensure that the lighting across images appears consistent to the eye. This step helps reduce the impact of extreme lighting changes and ensures a fairer evaluation of our method under controlled conditions. (2) One potential approach is to incorporate an appearance embedding for each image. We experiment with the Gaussian in the Wild (GS-W) baseline [59]. GS-W replaces traditional spherical harmonic-based color modeling with a method that separates intrinsic properties of each Gaussian point from dynamic appearance features of each image. This approach captures the stable, inherent appearance of objects while accommodating dynamic factors like highlights and shadows. Despite these efforts, our results, especially in Setting 3, still exhibit + +![](images/da6c60a73c394fcedf097463a5914d3d68971797e22304e6938efca9efb646be.jpg) +(a) Mitigating lighting inconsistency by camera embedding. + +![](images/9739511aa7f4910e80e89a9e7eb4c373e6276bbd22fcc5af6056f54a9bad3e71.jpg) +(b) Qualitative dynamic scenes rendering comparison. +Figure 6. Qualitative results of dynamic baseline and lighting handling. We mitigate the lighting inconsistency issue by carefully selecting traversals under similar lighting conditions, and it can be further alleviated by introducing camera embeddings. We provide the quantitative evaluation of dynamic scene reconstruction methods in Setting 2. + +Table 4. Quantitative performance of GS-W [59] in different settings. After learning the lighting features, the interpolated and extrapolated test metrics show significant improvement compared to other baselines. However, there is still a considerable drop from interpolation to extrapolation. + +
SettingPSNR ↑SSIM ↑LPIPS ↓Feat Cos Sim ↑
In.Ex.In.Ex.In.Ex.In.Ex.
Setting 128.1520.220.890.780.150.230.710.64
Setting 230.1021.210.910.820.130.200.760.67
Setting 328.6219.360.870.730.150.310.740.50
+ +a significant performance drop, as shown in Table 4 and Figure 6a, underscoring the limitations of Gaussian-based models in handling extrapolated, unseen scenarios. + +Dynamic Scenes. In addition to the static scene reconstruction methods that we evaluated in Table 1, some dynamic scene reconstruction techniques have recently emerged. We conduct evaluation for Setting 2 to assess these approaches, e.g., OmniRe [10]. It organizes rigid-deformable nodes and background nodes to capture dynamic scene structures and employs SMPL [32] for non-rigid object modeling. As illustrated in Figure 6b, the rendering results reveal a significant performance gap between the training and extrapolated test cameras. In the extrapolated test views, objects such as trees and stakes lose texture and geometric details, resulting in noticeably blurry outputs, whereas the training views maintain high fidelity. As shown in Table 5, the reconstruction metrics indicate an average drop of $25\%$ when transitioning from interpolated to extrapolated settings. The results highlight the challenges of extrapolated view synthesis in dynamic scenes and the need for further research. + +# 6. Conclusions, Limitations and Future Work + +Conclusions. We introduce the first benchmark enabling quantitative evaluation of extrapolated view synthesis, advancing photorealistic simulation for self-driving + +Table 5. Quantitative performance of OmniRe [10] in Setting 2. The experiment uses a single traversal with dynamic objects, showing a noticeable drop from interpolation to extrapolation. + +
SettingPSNR ↑SSIM ↑LPIPS ↓Feat Cos Sim ↑
In.Ex.In.Ex.In.Ex.In.Ex.
Setting 219.7815.320.650.450.380.530.730.58
+ +and robotics. Our benchmark integrates real-world multi-traversal, multi-agent, and multi-camera data, categorizes scenes into different evaluation settings, and evaluates state-of-the-art NVS models. Experimental results reveal that while some methods address specific challenges, current models demonstrate limited generalization, with significant overfitting to training views and suboptimal performance in extrapolated view synthesis. To support further research, we will release the dataset and benchmark, addressing the long-standing data scarcity and providing evaluation protocols. We believe the EUVS benchmark will catalyze meaningful advancements in self-driving and robotics innovation. + +Limitations and Future Work. Our benchmark does have limitations. First, although we provide the ground truth to enable quantitative evaluation for static scenes in all settings and the foreground objects evaluation in Setting 2, we lack the foreground evaluation in Settings 1 and 3. Future work will aim to expand our work to include the dynamic object evaluation in Settings 1 and 3 using multi-agent data. Secondly, although we carefully manually selected and ensured that the test trajectory viewpoints are well covered by training trajectories, a more in-depth evaluation of how observed and unseen regions differ is left for future work. + +Acknowledgment. This work was supported in part through NSF grants 2238968 and 2121391, and the NYU IT High Performance Computing resources, services, and staff expertise. Yiming Li is supported by NVIDIA Graduate Fellowship (2024-2025). + +# References + +[1] Alexander Amini, Tsun-Hsuan Wang, Igor Gilitschenski, Wilko Schwarting, Zhijian Liu, Song Han, Sertac Karaman, and Daniela Rus. Vista 2.0: An open, data-driven simulator for multimodal sensing and policy learning for autonomous vehicles. In 2022 International Conference on Robotics and Automation (ICRA), pages 2419-2426. IEEE, 2022. 3 +[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023. 5, 6 +[3] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 3 +[4] Holger Caesar, Juraj Kabzan, Kok Seang Tan, Whye Kit Fong, Eric Wolff, Alex Lang, Luke Fletcher, Oscar Beijbom, and Sammy Omari. nuplan: A closed-loop ml-based planning benchmark for autonomous vehicles. arXiv preprint arXiv:2106.11810, 2021. 2, 3, 4 +[5] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3d tracking and forecasting with rich maps. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8748-8757, 2019. 3 +[6] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19467, 2024. 2 +[7] Danpeng Chen, Hai Li, Weicai Ye, Yifan Wang, Weijian Xie, Shangjin Zhai, Nan Wang, Haomin Liu, Hujun Bao, and Guofeng Zhang. Pgsr: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction. arXiv preprint arXiv:2406.06521, 2024. 5, 6, 7 +[8] Yurui Chen, Chun Gu, Junzhe Jiang, Xiatian Zhu, and Li Zhang. Periodic vibration gaussian: Dynamic urban scene reconstruction and real-time rendering. arXiv preprint arXiv:2311.18561, 2023. 2, 3 +[9] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splattering from sparse multi-view images. arXiv preprint arXiv:2403.14627, 2024. 2 +[10] Ziyu Chen, Jiawei Yang, Jiahui Huang, Riccardo de Lutio, Janick Martinez Esturo, Boris Ivanovic, Or Litany, Zan Gojcic, Sanja Fidler, Marco Pavone, et al. Omnire: Omni urban scene reconstruction. arXiv preprint arXiv:2408.16760, 2024. 8 +[11] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussian-pro: 3d gaussian splattering with progressive propagation. In + +Forty-first International Conference on Machine Learning, 2024. 4, 6, 7 +[12] Carlos A Diaz-Ruiz, Youya Xia, Yurong You, Jose Nino, Junan Chen, Josephine Monica, Xiangyu Chen, Katie Luo, Yan Wang, Marc Emond, et al. Ithaca365: Dataset and driving perception under repeated and challenging weather conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21383-21392, 2022. 3 +[13] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 1, 3 +[14] Scott Ettinger, Shuyang Cheng, Benjamin Caine, Chenxi Liu, Hang Zhao, Sabeek Pradhan, Yuning Chai, Ben Sapp, Charles R Qi, Yin Zhou, et al. Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9710-9719, 2021. 3 +[15] Lue Fan, Hao Zhang, Qitai Wang, Hongsheng Li, and Zhaoxiang Zhang. Freesim: Toward free-viewpoint camera simulation in driving scenes, 2024. 3 +[16] Lan Feng, Quanyi Li, Zhenghao Peng, Shuhan Tan, and Bolei Zhou. Trafficgen: Learning to generate diverse and realistic traffic scenarios. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 3567-3575. IEEE, 2023. 3 +[17] Jannik Fritsch, Tobias Kuehnl, and Andreas Geiger. A new performance measure and evaluation benchmark for road detection algorithms. In 16th International IEEE Conference on Intelligent Transportation Systems (ITSC 2013), pages 1693-1700. IEEE, 2013. 3 +[18] Shenyuan Gao, Jiazhi Yang, Li Chen, Kashyap Chitta, Yihang Qiu, Andreas Geiger, Jun Zhang, and Hongyang Li. Vista: A generalizable driving world model with high fidelity and versatile controllability. arXiv preprint arXiv:2405.17398, 2024. 3 +[19] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 3 +[20] Huasong Han, Kaixuan Zhou, Xiaoxiao Long, Yusen Wang, and Chunxia Xiao. Ggs: Generalizable gaussian splatting for lane switching in autonomous driving. arXiv preprint arXiv:2409.02382, 2024. 2, 3 +[21] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6 +[22] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3, 5, 6 +[23] Sungwon Hwang, Min-Jung Kim, Taewoong Kang, Jayeon Kang, and Jaegul Choo. Vegs: View extrapolation of urban scenes in 3d gaussian splatting using learned priors. arXiv preprint arXiv:2407.02945, 2024. 2, 3, 5, 6, 7 + +[24] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5322-5332, 2024. 3 +[25] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 2, 3, 4, 6, 7 +[26] Mustafa Khan, Hamidreza Fazlali, Dhruv Sharma, Tongtong Cao, Dongfeng Bai, Yuan Ren, and Bingbing Liu. Autosplat: Constrained gaussian splatting for autonomous driving scene reconstruction. arXiv preprint arXiv:2407.02598, 2024. 3 +[27] Seung Wook Kim, Jonah Philion, Antonio Torralba, and Sanja Fidler. Drivegan: Towards a controllable high-quality neural simulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5820-5829, 2021. 3 +[28] Quanyi Li, Zhenghao Mark Peng, Lan Feng, Zhizheng Liu, Chenda Duan, Wenjie Mo, and Bolei Zhou. ScenarioNet: Open-source platform for large-scale traffic scenario simulation and modeling. Advances in neural information processing systems, 36, 2024. 3 +[29] Yiming Li, Dekun Ma, Ziyan An, Zixun Wang, Yiqi Zhong, Siheng Chen, and Chen Feng. V2x-sim: Multi-agent collaborative perception dataset and benchmark for autonomous driving. IEEE Robotics and Automation Letters, 7(4): 10914-10921, 2022. 3 +[30] Yiming Li, Zhiheng Li, Nuo Chen, Moonjun Gong, Zonglin Lyu, Zehong Wang, Peili Jiang, and Chen Feng. Multiagent multitraversal multimodal self-driving: Open mars dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22041-22051, 2024. 2, 3, 4 +[31] Yiming Li, Zehong Wang, Yue Wang, Zhiding Yu, Zan Gojcic, Marco Pavone, Chen Feng, and Jose M. Alvarez. Memorize what matters: Emergent scene decomposition from multitraverse. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 4, 6, 7 +[32] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1-248:16, 2015. 8 +[33] Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3061-3070, 2015. 3 +[34] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3 +[35] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 5, 6 + +[36] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6 +[37] Matthew Pitropov, Danson Evan Garcia, Jason Rebello, Michael Smart, Carlos Wang, Krzysztof Czarnecki, and Steven Waslander. Canadian adverse driving conditions dataset. The International Journal of Robotics Research, 40 (4-5):681-690, 2021. 3 +[38] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dólar, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos, 2024. 6 +[39] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 5 +[40] Shital Shah, Debadeepta Dey, Chris Lovett, and Ashish Kapoor. Airsim: High-fidelity visual and physical simulation for autonomous vehicles. In Field and Service Robotics: Results of the 11th International Conference, pages 621-635. Springer, 2018. 3 +[41] Xi Shi, Lingli Chen, Peng Wei, Xi Wu, Tian Jiang, Yonggang Luo, and Lecheng Xie. Dhgs: Decoupled hybrid gaussian splatting for driving scene. arXiv preprint arXiv:2407.16600, 2024.3 +[42] Shuhan Tan, Kelvin Wong, Shenlong Wang, Sivabalan Manivasagam, Mengye Ren, and Raquel Urtasun. Scenegen: Learning to generate realistic traffic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 892-901, 2021. 3 +[43] Izzeddin Teeti, Valentina Musat, Salman Khan, Alexander Rast, Fabio Cuzzolin, and Andrew Bradley. Vision in adverse weather: Augmentation using cyclegans with various object detectors for robust perception in autonomous racing. arXiv preprint arXiv:2201.03246, v3, 2023. 3 +[44] Adam Tondersski, Carl Lindstrom, Georg Hess, William Ljungbergh, Lennart Svensson, and Christoffer Petersson. Neurad: Neural rendering for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14895-14904, 2024. 1 +[45] Christopher Wewer, Kevin Raj, Eddy Ilg, Bernt Schiele, and Jan Eric Lenssen. latentsplat: Autoencoding variational gaussians for fast generalizable 3d reconstruction. arXiv preprint arXiv:2403.16292, 2024. 2 +[46] Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, et al. Argoverse 2: Next generation datasets for self-driving perception and forecasting. arXiv preprint arXiv:2301.00493, 2023. 2, 3, 4 +[47] Chenming Wu, Jiadai Sun, Zhelun Shen, and Liangjun Zhang. Mapnerf: Incorporating map priors into neural ra + +diance fields for driving view simulation. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7082-7088. IEEE, 2023. 3 +[48] Ke Wu, Kaizhao Zhang, Zhiwei Zhang, Shanshuai Yuan, Muer Tie, Julong Wei, Zijun Xu, Jieru Zhao, Zhongxue Gan, and Wenchao Ding. Hgs-mapping: Online dense mapping using hybrid gaussian representation in urban scenes. arXiv preprint arXiv:2403.20159, 2024. 3 +[49] Zirui Wu, Tianyu Liu, Liyi Luo, Zhide Zhong, Jianteng Chen, Hongmin Xiao, Chao Hou, Haozhe Lou, Yuantao Chen, Runyi Yang, et al. Mars: An instance-aware, modular and realistic simulator for autonomous driving. In CAAI International Conference on Artificial Intelligence, pages 3-15. Springer, 2023. 3 +[50] Bernhard Wymann, Eric Espie, Christophe Guionneau, Christos Dimitrakakis, Rémi Coulom, and Andrew Sumner. Torcs, the open racing car simulator. Software available at http://torcs.sourceforge.net, 4(6):2, 2000. 3 +[51] Danfei Xu, Yuxiao Chen, Boris Ivanovic, and Marco Pavone. Bits: Bi-level imitation for traffic simulation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2929-2936. IEEE, 2023. 3 +[52] Runsheng Xu, Xin Xia, Jinlong Li, Hanzhao Li, Shuo Zhang, Zhengzhong Tu, Zonglin Meng, Hao Xiang, Xiaoyu Dong, Rui Song, et al. V2v4real: A real-world large-scale dataset for vehicle-to-vehicle cooperative perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13712-13722, 2023. 3 +[53] Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians for modeling dynamic urban scenes. arXiv preprint arXiv:2401.01339, 2024. 2, 3 +[54] Chen Yang, Peihao Li, Zanwei Zhou, Shanxin Yuan, Bingbing Liu, Xiaokang Yang, Weichao Qiu, and Wei Shen. Ner-fvs: Neural radiance fields for free view synthesis via geometry scaffolds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16549-16558, 2023. 2 +[55] Jiawei Yang, Boris Ivanovic, Or Litany, Xinshuo Weng, Seung Wook Kim, Boyi Li, Tong Che, Danfei Xu, Sanja Fidler, Marco Pavone, et al. Emernerf: Emergent spatial-temporal scene decomposition via self-supervision. arXiv preprint arXiv:2311.02077, 2023. 3 +[56] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1389-1399, 2023. 1, 3 +[57] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19447-19456, 2024. 3 +[58] Zhongrui Yu, Haoran Wang, Jinze Yang, Hanzhang Wang, Zeke Xie, Yunfeng Cai, Jiale Cao, Zhong Ji, and Mingming + +Sun. Sgd: Street view synthesis with gaussian splatting and diffusion prior. arXiv preprint arXiv:2403.20079, 2024. 3 +[59] Dongbin Zhang, Chuming Wang, Weitao Wang, Peihao Li, Minghan Qin, and Haoqian Wang. Gaussian in the wild: 3d gaussian splatting for unconstrained image collections. arXiv preprint arXiv:2403.15704, 2024. 7, 8 +[60] Jian Zhang, Yuanqing Zhang, Huan Fu, Xiaowei Zhou, Bowen Cai, Jinchi Huang, Rongfei Jia, Binqiang Zhao, and Xing Tang. Ray priors through reprojection: Improving neural radiance fields for novel view extrapolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18376-18386, 2022. 2 +[61] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6 +[62] Guosheng Zhao, Chaojun Ni, Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, Boyuan Wang, Youyi Zhang, Wenjun Mei, and Xingang Wang. Drivedreamer4d: World models are effective data machines for 4d driving scene representation. arXiv preprint arXiv:2410.13571, 2024. 3 +[63] Ziyuan Zhong, Davis Rempe, Danfei Xu, Yuxiao Chen, Sushant Veer, Tong Che, Baishakhi Ray, and Marco Pavone. Guided conditional diffusion for controllable traffic simulation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 3560-3566. IEEE, 2023. 3 +[64] Hongyu Zhou, Jiahao Shao, Lu Xu, Dongfeng Bai, Weichao Qiu, Bingbing Liu, Yue Wang, Andreas Geiger, and Yiyi Liao. Hugs: Holistic urban 3d scene understanding via gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21336-21345, 2024. 2, 3 +[65] Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Zehao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21676-21685, 2024. 5, 6, 7 +[66] Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21634-21643, 2024. 2, 3 \ No newline at end of file diff --git a/2025/Extrapolated Urban View Synthesis Benchmark/images.zip b/2025/Extrapolated Urban View Synthesis Benchmark/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5ef817381a0c2c3c8c470e2c320378e7b55f8308 --- /dev/null +++ b/2025/Extrapolated Urban View Synthesis Benchmark/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:40de74a35f4a427f2271f684dfd4673b5a770959315cccac9e08567ff6620431 +size 1275364 diff --git a/2025/Extrapolated Urban View Synthesis Benchmark/layout.json b/2025/Extrapolated Urban View Synthesis Benchmark/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a22892ec863b9bcc22fc6d9a94c6c93732a984c8 --- /dev/null +++ b/2025/Extrapolated Urban View Synthesis Benchmark/layout.json @@ -0,0 +1,7375 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 156, + 103, + 454, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 103, + 454, + 121 + ], + "spans": [ + { + "bbox": [ + 156, + 103, + 454, + 121 + ], + "type": "text", + "content": "Extrapolated Urban View Synthesis Benchmark" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 90, + 141, + 520, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 141, + 520, + 191 + ], + "spans": [ + { + "bbox": [ + 90, + 141, + 520, + 191 + ], + "type": "text", + "content": "Xiangyu Han" + }, + { + "bbox": [ + 90, + 141, + 520, + 191 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 90, + 141, + 520, + 191 + ], + "type": "text", + "content": " Zhen Jia" + }, + { + "bbox": [ + 90, + 141, + 520, + 191 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 90, + 141, + 520, + 191 + ], + "type": "text", + "content": " Boyi Li" + }, + { + "bbox": [ + 90, + 141, + 520, + 191 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 90, + 141, + 520, + 191 + ], + "type": "text", + "content": " Yan Wang" + }, + { + "bbox": [ + 90, + 141, + 520, + 191 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 90, + 141, + 520, + 191 + ], + "type": "text", + "content": " Boris Ivanovic" + }, + { + "bbox": [ + 90, + 141, + 520, + 191 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 90, + 141, + 520, + 191 + ], + "type": "text", + "content": " Yurong You" + }, + { + "bbox": [ + 90, + 141, + 520, + 191 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 90, + 141, + 520, + 191 + ], + "type": "text", + "content": " Lingjie Liu" + }, + { + "bbox": [ + 90, + 141, + 520, + 191 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 90, + 141, + 520, + 191 + ], + "type": "text", + "content": " Yue Wang" + }, + { + "bbox": [ + 90, + 141, + 520, + 191 + ], + "type": "inline_equation", + "content": "^{2,4}" + }, + { + "bbox": [ + 90, + 141, + 520, + 191 + ], + "type": "text", + "content": " Marco Pavone" + }, + { + "bbox": [ + 90, + 141, + 520, + 191 + ], + "type": "inline_equation", + "content": "^{2,5}" + }, + { + "bbox": [ + 90, + 141, + 520, + 191 + ], + "type": "text", + "content": " Chen Feng" + }, + { + "bbox": [ + 90, + 141, + 520, + 191 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 90, + 141, + 520, + 191 + ], + "type": "text", + "content": " Yiming Li" + }, + { + "bbox": [ + 90, + 141, + 520, + 191 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + }, + { + "bbox": [ + 90, + 141, + 520, + 191 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 90, + 141, + 520, + 191 + ], + "type": "text", + "content": "NYU " + }, + { + "bbox": [ + 90, + 141, + 520, + 191 + ], + "type": "inline_equation", + "content": "{}^{2}" + }, + { + "bbox": [ + 90, + 141, + 520, + 191 + ], + "type": "text", + "content": "NVIDIA " + }, + { + "bbox": [ + 90, + 141, + 520, + 191 + ], + "type": "inline_equation", + "content": "{}^{3}" + }, + { + "bbox": [ + 90, + 141, + 520, + 191 + ], + "type": "text", + "content": "UPenn " + }, + { + "bbox": [ + 90, + 141, + 520, + 191 + ], + "type": "inline_equation", + "content": "{}^{4}" + }, + { + "bbox": [ + 90, + 141, + 520, + 191 + ], + "type": "text", + "content": "USC " + }, + { + "bbox": [ + 90, + 141, + 520, + 191 + ], + "type": "inline_equation", + "content": "{}^{5}" + }, + { + "bbox": [ + 90, + 141, + 520, + 191 + ], + "type": "text", + "content": "Stanford" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 200, + 193, + 407, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 193, + 407, + 205 + ], + "spans": [ + { + "bbox": [ + 200, + 193, + 407, + 205 + ], + "type": "text", + "content": "https://ai4ce.github.io/EUVS-Benchmark" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 61, + 217, + 211, + 416 + ], + "blocks": [ + { + "bbox": [ + 61, + 217, + 211, + 416 + ], + "lines": [ + { + "bbox": [ + 61, + 217, + 211, + 416 + ], + "spans": [ + { + "bbox": [ + 61, + 217, + 211, + 416 + ], + "type": "image", + "image_path": "2750e1de8e2e50beca42baa1099958d524968a043102410c836203c281dd0c6d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 425, + 555, + 470 + ], + "lines": [ + { + "bbox": [ + 54, + 425, + 555, + 470 + ], + "spans": [ + { + "bbox": [ + 54, + 425, + 555, + 470 + ], + "type": "text", + "content": "Figure 1. Our key contributions. Previous evaluations for urban view synthesis have primarily focused on interpolated poses, as the lack of ground truth data has made it challenging to evaluate extrapolated poses. We address this gap by providing real-world data that enables both quantitative and qualitative evaluations of extrapolated view synthesis in urban scenes. The quantitative results reveal a significant performance drop in 3D Gaussian Splatting [25] when handling extrapolated views, highlighting the need for more robust NVS methods." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 219, + 217, + 299, + 401 + ], + "blocks": [ + { + "bbox": [ + 219, + 217, + 299, + 401 + ], + "lines": [ + { + "bbox": [ + 219, + 217, + 299, + 401 + ], + "spans": [ + { + "bbox": [ + 219, + 217, + 299, + 401 + ], + "type": "image", + "image_path": "fc9a854c63e2bee516fd917f4f42336948855da87e0a6a0784af664d4c075558.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 244, + 403, + 257, + 416 + ], + "blocks": [ + { + "bbox": [ + 244, + 403, + 257, + 416 + ], + "lines": [ + { + "bbox": [ + 244, + 403, + 257, + 416 + ], + "spans": [ + { + "bbox": [ + 244, + 403, + 257, + 416 + ], + "type": "image", + "image_path": "ce57cae050600ffb83d7b56b896254455bb4ea9f953c02cde065b1107a85d519.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 257, + 404, + 367, + 414 + ], + "lines": [ + { + "bbox": [ + 257, + 404, + 367, + 414 + ], + "spans": [ + { + "bbox": [ + 257, + 404, + 367, + 414 + ], + "type": "text", + "content": "GT is available in previous benchmark" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 309, + 217, + 433, + 401 + ], + "blocks": [ + { + "bbox": [ + 309, + 217, + 433, + 401 + ], + "lines": [ + { + "bbox": [ + 309, + 217, + 433, + 401 + ], + "spans": [ + { + "bbox": [ + 309, + 217, + 433, + 401 + ], + "type": "image", + "image_path": "8e20560ca42402c382e700e53a8b707c1a96b5c1dba871c7051a5da554f6c2af.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 403, + 403, + 416, + 416 + ], + "blocks": [ + { + "bbox": [ + 403, + 403, + 416, + 416 + ], + "lines": [ + { + "bbox": [ + 403, + 403, + 416, + 416 + ], + "spans": [ + { + "bbox": [ + 403, + 403, + 416, + 416 + ], + "type": "image", + "image_path": "60d60ba9ab5eafe33b3d1185866f863d630b62c9e743fb06fe8a2c06ff73990e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 417, + 406, + 512, + 414 + ], + "lines": [ + { + "bbox": [ + 417, + 406, + 512, + 414 + ], + "spans": [ + { + "bbox": [ + 417, + 406, + 512, + 414 + ], + "type": "text", + "content": "GT is available in our benchmark" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 440, + 217, + 545, + 401 + ], + "blocks": [ + { + "bbox": [ + 440, + 217, + 545, + 401 + ], + "lines": [ + { + "bbox": [ + 440, + 217, + 545, + 401 + ], + "spans": [ + { + "bbox": [ + 440, + 217, + 545, + 401 + ], + "type": "image", + "image_path": "59d0e3a7f01289ad2105a343148fef98d14a6eaa7349ffc290cc04ae639af593.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 151, + 479, + 200, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 479, + 200, + 491 + ], + "spans": [ + { + "bbox": [ + 151, + 479, + 200, + 491 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 54, + 505, + 297, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 505, + 297, + 685 + ], + "spans": [ + { + "bbox": [ + 54, + 505, + 297, + 685 + ], + "type": "text", + "content": "Photorealistic simulators are essential for the training and evaluation of vision-centric autonomous vehicles (AVs). At their core is Novel View Synthesis (NVS), a crucial capability that generates diverse unseen viewpoints to accommodate the broad and continuous pose distribution of AVs. Recent advances in radiance fields, such as 3D Gaussian Splatting, achieve photorealistic rendering at real-time speeds and have been widely used in modeling large-scale driving scenes. However, their performance is commonly evaluated using an interpolated setup with highly correlated training and test views. In contrast, extrapolation, where test views largely deviate from training views, remains underexplored, limiting progress in generalizable simulation technology. To address this gap, we leverage publicly available AV datasets with multiple traversals," + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 480, + 556, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 556, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 556, + 613 + ], + "type": "text", + "content": "multiple vehicles, and multiple cameras to build the first Extrapolated Urban View Synthesis (EUVS) benchmark. Meanwhile, we conduct both quantitative and qualitative evaluations of state-of-the-art NVS methods across different evaluation settings. Our results show that current NVS methods are prone to overfitting to training views. Besides, incorporating diffusion priors and improving geometry cannot fundamentally improve NVS under large view changes, highlighting the need for more robust approaches and large-scale training. We have released the data to help advance self-driving and urban robotics simulation technology." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 631, + 395, + 643 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 631, + 395, + 643 + ], + "spans": [ + { + "bbox": [ + 314, + 631, + 395, + 643 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "The development of vision-centric autonomous vehicles (AVs) relies heavily on photorealistic simulators, which provide controlled, reproducible, and scalable environments for training and evaluation of driving models [13, 44, 56]. These simulators enable AVs to learn and adapt to a variety" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28718" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "of real-world scenarios, from crowded urban streets to adverse weather conditions, without the logistical and safety concerns of physical road testing. At the heart of these simulators is the capability for Novel View Synthesis (NVS)—a key technology that generates realistic images of unseen viewpoints, simulating the continuous changes in perspective that occur as AVs navigate through urban environments." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 159, + 296, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 159, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 159, + 296, + 590 + ], + "type": "text", + "content": "Recent advancements in radiance fields, particularly methods based on 3D Gaussian Splatting [25], have significantly improved the realism and efficiency of NVS. These approaches [8, 53, 64, 66] can produce photorealistic renderings at real-time speeds, making them highly attractive for large-scale driving scene simulation. However, despite their impressive results, the evaluation of NVS methods has predominantly focused on interpolated scenarios, where training and test viewpoints are closely related. While interpolation tests are valuable for assessing local consistency, they fall short in addressing the more critical challenge of extrapolation—where test viewpoints differ significantly from the training data. As shown in Figure 1, the interpolation test set demonstrates strong performance, with metrics such as PSNR, SSIM, and LPIPS remaining very close to the training set values. In contrast, the extrapolation test set, which includes additional translation and rotation changes relative to the training set, exhibits notable drops in performance. Specifically, the metric decreases relative to the training set are " + }, + { + "bbox": [ + 56, + 159, + 296, + 590 + ], + "type": "inline_equation", + "content": "28\\%" + }, + { + "bbox": [ + 56, + 159, + 296, + 590 + ], + "type": "text", + "content": " for PSNR, " + }, + { + "bbox": [ + 56, + 159, + 296, + 590 + ], + "type": "inline_equation", + "content": "22\\%" + }, + { + "bbox": [ + 56, + 159, + 296, + 590 + ], + "type": "text", + "content": " for SSIM, and " + }, + { + "bbox": [ + 56, + 159, + 296, + 590 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 56, + 159, + 296, + 590 + ], + "type": "text", + "content": " for LPIPS. These results underscore the urgent need to explore and advance extrapolated view synthesis in complex urban scenes, as real-world driving often involves encountering scenarios with significant viewpoint shifts and diverse spatial transformations that deviate from training distributions. Several recent studies [20, 23] have investigated the generalization capabilities of NVS in 3D Gaussian Splatting. Although they show promising qualitative results, there is no comprehensive quantitative analysis due to the absence of standardized datasets. Moreover, their evaluations are primarily limited to specific scenarios or use cases, without investigating varying evaluation settings based on the degree of extrapolation. This gap underscores the urgent need for a benchmark that offers diverse and challenging datasets, enabling a rigorous and systematic assessment of NVS methods." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "To establish a common platform for assessing the robustness of NVS methods, we introduce a comprehensive benchmark for quantitatively and qualitatively evaluating extrapolated novel view synthesis in large-scale urban scenes. Our benchmark leverages publicly available datasets, including NuPlan [4], MARS [30], and Argoverse2 [46], which feature multi-traversal, multi-agent and multi-camera sensory recordings. Multi-traversal data consists of asynchronous traversals of the same location, while multi-agent data refers to data collected simultaneously" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "content": "from multiple vehicles within the same area. These data provide diverse camera poses within a 3D scene, enabling the training and evaluation of extrapolated view synthesis in outdoor environments. For the experimental setup, we define three evaluation settings: (1) translation only, (2) rotation only, and (3) translation + rotation, as shown in Figure 4. In autonomous driving scenarios, Setting 1 corresponds to maneuvers such as lane changes, Setting 2 involves switching between cameras facing different directions, and Setting 3 addresses complex intersections, such as crossroads with diverse traversal paths. These settings represent common challenges in autonomous driving, and addressing them enables the synthesis of complete scenes from sparse image observations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 239, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 553, + 312 + ], + "type": "text", + "content": "We conduct pose estimation and sparse reconstruction using COLMAP [39], which facilitates the initialization of Gaussian Splatting. We then evaluate state-of-the-art Gaussian Splatting-based approaches across each evaluation setting, identifying performance gaps both qualitatively and quantitatively in extrapolated urban view synthesis." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 326, + 312, + 533, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 312, + 533, + 323 + ], + "spans": [ + { + "bbox": [ + 326, + 312, + 533, + 323 + ], + "type": "text", + "content": "In summary, our main contributions are as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 323, + 553, + 538 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 323, + 553, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 323, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 314, + 323, + 553, + 407 + ], + "type": "text", + "content": "- We initiate the first comprehensive quantitative and qualitative study on the Extrapolated Urban View Synthesis (EUVS) problem, supported by a robust evaluation framework that categorizes evaluation settings (translation-only, rotation-only, and translation + rotation) while assessing performance using diverse metrics, including reconstruction accuracy and visual fidelity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 407, + 553, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 407, + 553, + 479 + ], + "spans": [ + { + "bbox": [ + 314, + 407, + 553, + 479 + ], + "type": "text", + "content": "- We construct a novel dataset by integrating multitraversal, multi-agent, and multi-camera data from publicly available resources, totaling 90,810 frames across 345 videos. Our dataset effectively addresses the limitations of existing benchmarks, enabling rigorous and robust evaluation for extrapolated urban view synthesis." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 479, + 553, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 479, + 553, + 538 + ], + "spans": [ + { + "bbox": [ + 314, + 479, + 553, + 538 + ], + "type": "text", + "content": "- We benchmark state-of-the-art Gaussian Splitting-based and NeRF-based models and analyze key factors that influence the performance of extrapolated NVs, laying a solid foundation for future advancements in this challenging task. Data and code will be released upon acceptance." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 550, + 405, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 550, + 405, + 562 + ], + "spans": [ + { + "bbox": [ + 314, + 550, + 405, + 562 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "type": "text", + "content": "Extrapolated View Synthesis. Extrapolated view synthesis aims to generate novel views beyond observed perspectives, addressing challenges in visual coherence for unseen regions. RapNeRF [60] proposes a random ray-casting policy that enables training on unseen views based on visible ones. Following work [54] enhances this approach by incorporating holistic priors. Besides, some generalizable models [6, 9, 45] have emerged, capable of generating extrapolated novel views from a limited number of input images. While these methods are designed for indoor scenes, several works address extrapolated view synthesis in outdoor driving scenarios, which typically involve forward-facing" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28719" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": "cameras and unbounded environments. To tackle the Setting 1 challenge in our benchmark and address the scarcity of lane change data, GGS [20] introduces a novel virtual lane generation module. In parallel, AutoSplat [26] tackles lane change in dynamic scenes by applying geometric and reflected consistency constraints. To address the Setting 1 or 2 challenge, FreeSim [15], VEGS [23], and SGD [58] enhance 3DGS [25] with diffusion priors to improve generalization ability. Yet existing methods suffer from two major limitations: (1) a lack of real data for quantitative evaluation, which confines them to quantitative analysis, and (2) a narrow focus on a specific setting in our benchmark, preventing a comprehensive and systematic exploration." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 232, + 296, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 296, + 518 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 296, + 518 + ], + "type": "text", + "content": "3D Gaussian Splatting. Recent advances in radiance fields, particularly NeRF [34] and 3DGS [25], have garnered significant attention due to their impressive advancements in NVs. NeRF employs an implicit representation through a multi-layer perceptron (MLP). Furthermore, 3DGS explicitly represents scenes using anisotropic 3D Gaussian ellipsoids, enabling high-quality real-time rendering. Several works have addressed issues such as difficulties with reflective surfaces [24], aliasing [57], etc. However, urban scenes introduce unique challenges due to their unbounded and dynamic nature. To address the challenge, several works separate dynamic and static elements in the scene by leveraging a composite dynamic Gaussian graph [53, 66], optical flow prediction [55, 64], etc. PVG [8] presents a unified representation model that simultaneously incorporates both dynamic and static components without relying on priors. To achieve realistic geometry and efficient rendering, 2DGS [22] collapses 3D Gaussians onto 2D planes, while hybrid approaches [41, 48] combine different Gaussians to better capture region-specific features. In summary, current urban NVS methods primarily focus on effectively handling dynamic elements and enhancing geometry representation, while the challenge of extrapolated view synthesis remains largely underexplored." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": "Autonomous Driving Simulators. Current simulators focus on three key challenges: parameter initialization [16, 42], traffic simulation [28, 51, 63], and sensor simulation [1, 13, 18, 27, 56, 62]. Sensor simulation is crucial for generating realistic sensory data that AVs depend on for perception and decision-making. Early sensor simulators [13, 40, 50] provide simulated environments that are valuable for research but lack visual realism. Recent studies have focused on data-driven simulators that extract data from real-world driving logs, creating more realistic and adaptable environments. These methods can be classified into two categories: generation-based and reconstruction-based approaches. The former rely on inputs such as text, video, and other data sources for simulation, supported by world models or prior knowledge [18, 27, 62]. Reconstruction-based simulations leverage real-world data" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": "to ensure both visual fidelity and geometric consistency [47, 49, 56]. UniSim [56] is a pioneering example of this approach, utilizing NeRF-based scene representation to create dynamic scenes with geometric information that are both editable and controllable. Extrapolated view synthesis is essential for these simulators, as it enables the generation of realistic and consistent views from diverse angles." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 157, + 556, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 556, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 556, + 300 + ], + "type": "text", + "content": "Autonomous Driving Datasets. High-quality datasets play a vital role in advancing autonomous driving research. The KITTI dataset [19], released in 2012, marked a major milestone, significantly accelerating advances in AVs [17, 33, 43]. Since then, many influential autonomous vehicle datasets have been developed to tackle challenges like adverse weather conditions [37], multimodal fusion [3, 4], repeated driving [4, 12], collaborative driving [29, 30, 52], and motion prediction [5, 14, 46], etc. We leverage publicly available datasets with multi-traversal, multi-agent, and multi-camera recordings, enabling a comprehensive and robust evaluation of extrapolated urban view synthesis." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 312, + 448, + 325 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 448, + 325 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 448, + 325 + ], + "type": "text", + "content": "3. The EUVS Benchmark" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 332, + 418, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 332, + 418, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 418, + 344 + ], + "type": "text", + "content": "3.1. Dataset Curation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 350, + 556, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 556, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 556, + 567 + ], + "type": "text", + "content": "We leverage three publicly available, community-verified autonomous driving datasets—nuPlan [4], Argoverse 2 [46], and MARS [30]—to enhance adoption and foster trust through their established reliability. nuPlan [4] provides 1,200 hours of driving data from four cities, serving as the first large-scale planning benchmark. Argoverse 2 [46] supports multimodal perception and forecasting with 1,000 annotated 3D scenarios, 20,000 unlabeled lidar sequences, and 250,000 motion forecasting cases. MARS [30] enables collaborative driving research with multi-agent and multitraversal scenarios. The original datasets were not designed for evaluating extrapolated view synthesis, which requires our significant labor in data processing—" + }, + { + "bbox": [ + 313, + 350, + 556, + 567 + ], + "type": "inline_equation", + "content": "300+" + }, + { + "bbox": [ + 313, + 350, + 556, + 567 + ], + "type": "text", + "content": " hours of manual traversal selection and " + }, + { + "bbox": [ + 313, + 350, + 556, + 567 + ], + "type": "inline_equation", + "content": "800+" + }, + { + "bbox": [ + 313, + 350, + 556, + 567 + ], + "type": "text", + "content": " hours of computing to run COLMAP. By integrating these datasets, our benchmark enables the evaluation of view synthesis across diverse and realistic urban environments under varying conditions. Figure 3 illustrates the distribution of the integrated datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 575, + 446, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 575, + 446, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 575, + 446, + 586 + ], + "type": "text", + "content": "3.2. Evaluation Framework" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 594, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 676 + ], + "type": "text", + "content": "To systematically assess model performance in extrapolated urban view synthesis, our evaluation framework incorporates three evaluation settings and three data configurations. Data configurations include multi-traversal, multi-agent, and multi-camera, while evaluation settings are categorized into (1) translation only, (2) rotation only, and (3) translation + rotation, as illustrated in Figure 4." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Setting 1. The translation-only experimental setup involves scenarios where the vehicle's position shifts without any change in orientation. This scenario is commonly observed" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28720" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 550, + 329 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 550, + 329 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 550, + 329 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 550, + 329 + ], + "type": "image", + "image_path": "1709a7854bc488c0ae02b1c9a19f4299eefca5241f7ba846a7195561fb2aa78d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 335, + 555, + 379 + ], + "lines": [ + { + "bbox": [ + 55, + 335, + 555, + 379 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 555, + 379 + ], + "type": "text", + "content": "Figure 2. Dataset visualization. Our dataset features diverse scenes across various locations in different cities, sourced from multiple datasets. Typical driving scenarios include maneuvers such as lane changes, cross intersections, and T-junctions. Top: Each column displays images captured at the same location by different agents or traversals. Bottom: Each image displays the COLMAP points at a specific location, along with the corresponding camera poses." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 385, + 544, + 452 + ], + "blocks": [ + { + "bbox": [ + 59, + 385, + 544, + 452 + ], + "lines": [ + { + "bbox": [ + 59, + 385, + 544, + 452 + ], + "spans": [ + { + "bbox": [ + 59, + 385, + 544, + 452 + ], + "type": "image", + "image_path": "686306d5c9869386ddea58a0a3dd01fc160eb04be8c4e3c8449e63501b05a228.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 458, + 555, + 481 + ], + "lines": [ + { + "bbox": [ + 55, + 458, + 555, + 481 + ], + "spans": [ + { + "bbox": [ + 55, + 458, + 555, + 481 + ], + "type": "text", + "content": "Figure 3. Dataset distribution. Our dataset comprises 90,810 frames distributed over 104 cases, capturing a diverse array of multi-traversal paths, multi-agent interactions, and multi-camera perspectives across varying evaluation settings." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 493, + 295, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 493, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 493, + 295, + 529 + ], + "type": "text", + "content": "in lane changes. We use traversals from different lanes in multi-traversal data, focusing on the three front cameras. The data is sourced from nuPlan [4] and Argoverse 2 [46]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 531, + 295, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 531, + 295, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 531, + 295, + 615 + ], + "type": "text", + "content": "Setting 2. The second setting, rotation only, evaluates models on views with significant orientation changes. In vision-centric autonomous vehicles, this corresponds to transitions between cameras capturing different directions. We leverage multi-camera data from nuPlan [4], training on three forward-facing and three rear-facing cameras, while evaluating on two side-facing cameras." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": "Setting 3. The third setting, combining translation and rotation, includes both positional shifts and orientation changes, posing the greatest challenge for NVS. To address this, we utilize multi-traversal driving data collected from the same location but across different traversal routes. For example, the training and test sets may include routes that approach an intersection from different directions. Typical route combinations feature scenarios such as intersections," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 493, + 555, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 493, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 493, + 555, + 529 + ], + "type": "text", + "content": "T-junctions, and Y-junctions, as shown in Figure 2, ensuring diversity and comprehensive evaluation. The data for this setting is from MARS [30] and Argoverse 2 [46]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 539, + 433, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 539, + 433, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 433, + 552 + ], + "type": "text", + "content": "3.3. Algorithm Overview" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 556, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 555, + 653 + ], + "type": "text", + "content": "Vanilla 3D Gaussian Splatting. 3D Gaussian Splatting (3DGS) [25] leverage 3D Gaussians to explicitly represent the scene, which achieves high quality while offering real-time rendering by avoiding unnecessary computation in the empty space. Building on this, 3DGM [31] leverages multitraversal consensus to differentiate transient and permanent elements, enabling joint 2D segmentation and 3D mapping without using any human supervision." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "Planar-based and Geometry Refined Gaussian Splat-ting. GaussianPro [11] builds on 3DGS [25] by introducing multi-frame geometric optimization, which guides the densi-sification of 3D Gaussians, enhancing scene consistency in complex geometries. It further refines geometry by encour" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28721" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 548, + 167 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 548, + 167 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 548, + 167 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 548, + 167 + ], + "type": "image", + "image_path": "dce74e08eb5f1a159d2d3546b7f157deb1bfd9ffda30b71b6febb0bc60b72ed1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 209, + 167, + 385, + 179 + ], + "lines": [ + { + "bbox": [ + 209, + 167, + 385, + 179 + ], + "spans": [ + { + "bbox": [ + 209, + 167, + 385, + 179 + ], + "type": "text", + "content": "(a) Evaluation Setting 1: Translation Only" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 180, + 548, + 270 + ], + "blocks": [ + { + "bbox": [ + 59, + 180, + 548, + 270 + ], + "lines": [ + { + "bbox": [ + 59, + 180, + 548, + 270 + ], + "spans": [ + { + "bbox": [ + 59, + 180, + 548, + 270 + ], + "type": "image", + "image_path": "dcd886910944c08c09424ea1e29b545ff39f8ff761dbbd9bbe7bd98c2d7a4e29.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 59, + 287, + 548, + 379 + ], + "blocks": [ + { + "bbox": [ + 214, + 273, + 380, + 285 + ], + "lines": [ + { + "bbox": [ + 214, + 273, + 380, + 285 + ], + "spans": [ + { + "bbox": [ + 214, + 273, + 380, + 285 + ], + "type": "text", + "content": "(b) Evaluation Setting 2: Rotation Only" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 287, + 548, + 379 + ], + "lines": [ + { + "bbox": [ + 59, + 287, + 548, + 379 + ], + "spans": [ + { + "bbox": [ + 59, + 287, + 548, + 379 + ], + "type": "image", + "image_path": "cb960acf25f7e663d00c9736e853be8325d6acbf1a823ac3560a89ba1df9b371.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 204, + 380, + 403, + 392 + ], + "lines": [ + { + "bbox": [ + 204, + 380, + 403, + 392 + ], + "spans": [ + { + "bbox": [ + 204, + 380, + 403, + 392 + ], + "type": "text", + "content": "(c) Evaluation Setting 3: Translation + Rotation" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 396, + 555, + 441 + ], + "lines": [ + { + "bbox": [ + 54, + 396, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 54, + 396, + 555, + 441 + ], + "type": "text", + "content": "Figure 4. Qualitative and quantitative results across three evaluation settings. The performance drop from interpolation to extrapolation is significant in both qualitative and quantitative comparison. Different testing settings have distinct scenario characteristics, enabling the evaluation of a method's capabilities from various aspects, thus systematically assessing the overall performance of reconstruction algorithms, including geometric accuracy, hallucination ability, view consistency, and depth precision, etc." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 455, + 295, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 455, + 295, + 539 + ], + "spans": [ + { + "bbox": [ + 54, + 455, + 295, + 539 + ], + "type": "text", + "content": "aging Gaussian primitives to adopt flat structures. Similarly, 2DGS [22] projects the 3D volume into a set of 2D oriented planar Gaussian disks, enabling high-fidelity surface reconstruction. PGSR [7] introduces an unbiased depth rendering method and integrates single-view geometric, multiview photometric, and geometric regularization techniques to improve global geometry accuracy." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 542, + 295, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 542, + 295, + 650 + ], + "spans": [ + { + "bbox": [ + 54, + 542, + 295, + 650 + ], + "type": "text", + "content": "Gaussian Splatting with Diffusion Priors. VEGS [23] introduces a novel view generalization approach that harnesses pre-extracted surface normals to align 3D Gaussians while generating augmented camera views guided by diffusion priors. These diffusion priors serve a dual purpose: providing denoising loss guidance and supervising the training of augmented cameras. This process effectively mitigates floating artifacts and fragmented geometries, resulting in more accurate and coherent 3D representations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "content": "Feature-Enhanced Gaussian Splatting. Feature 3DGS [65] extends 3D Gaussian Splatting with a Parallel N-dimensional Gaussian Rasterizer, allowing simultaneous rendering of radiance fields and high-dimensional semantic features. By embedding semantic features directly into 3D" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 455, + 555, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 555, + 491 + ], + "type": "text", + "content": "Gaussians, the approach enhances optimization, enabling better correspondence with scene semantics and achieving more detailed and accurate spatial representations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 492, + 556, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 492, + 556, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 492, + 556, + 624 + ], + "type": "text", + "content": "NeRF-based Method. Instant-NGP [35] uses a multiresolution hash encoding to map spatial coordinates into compact latent representations via hash tables. It efficiently encodes high-frequency details by combining trainable feature vectors with interpolation, enabling adaptive and scalable input encodings. Zip-NeRF [2] leverages multisampling with isotropic Gaussians for scale-aware features and introduces a smooth anti-aliasing loss to address z-aliasing. In addition, it incorporates a novel distance normalization technique to better manage close and distant objects, achieving high-quality rendering and fast training." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 638, + 391, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 638, + 391, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 638, + 391, + 651 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 658, + 422, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 422, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 422, + 672 + ], + "type": "text", + "content": "4.1. Experiment Setup" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Implementation Details. All 3DGS-based methods are initialized only from sparse points obtained by COLMAP [39] and exclude lidar points. We employ Grounded-SAM-2" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28722" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 101, + 553, + 434 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Quantitative rendering results across three evaluation settings. In. denotes interpolation, while Ex. represents extrapolation. Different baselines excel in different settings, reflecting the comprehensiveness and completeness of the evaluation protocol." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 101, + 553, + 434 + ], + "lines": [ + { + "bbox": [ + 56, + 101, + 553, + 434 + ], + "spans": [ + { + "bbox": [ + 56, + 101, + 553, + 434 + ], + "type": "table", + "html": "
MethodPSNR↑SSIM↑LPIPS↓Feat Cos Sim ↑
In.Ex.DropIn.Ex.DropIn.Ex.DropIn.Ex.Drop
Setting 13DGS [25]21.3616.3723.4%0.82750.720313.0%0.20410.259927.3%0.68280.603911.6%
3DGM [31]20.9616.3522.0%0.82930.724812.6%0.20030.254226.9%0.68020.608710.5%
GSPro [11]21.5116.3923.8%0.83100.718913.5%0.18040.245035.8%0.70810.613013.4%
VEGS [23]21.2615.8825.3%0.81070.704713.1%0.24980.306222.6%0.63230.552112.7%
PGSR [7]20.5716.3220.7%0.81040.710212.4%0.22620.273320.8%0.65150.584810.2%
2DGS [22]20.8716.3021.9%0.80760.710312.0%0.24380.289018.5%0.62560.56449.8%
Feature 3DGS [65]21.0216.0123.8%0.80960.724310.5%0.18760.257537.3%0.69580.612212.0%
Zip-NeRF [2]19.6814.0628.6%0.78560.691712.0%0.27110.341826.1%0.63180.554212.3%
Instant-NGP [35]18.7712.6532.6%0.76310.625218.1%0.48740.593821.8%0.54650.483711.5%
AVERAGE20.6715.5924.6%0.80830.704612.8%0.25010.313425.3%0.65050.577411.2%
Setting 23DGS [25]25.7519.5324.2%0.87660.751114.3%0.15360.266873.7%0.73270.631913.8%
3DGM [31]25.7518.7827.1%0.87860.746415.0%0.15560.281380.8%0.72780.634412.8%
GSPro [11]26.4219.3926.6%0.88210.747015.3%0.13290.224669.0%0.75230.648713.8%
VEGS [23]24.5423.334.9%0.83660.79495.0%0.23010.281122.2%0.65950.61337.0%
PGSR [7]24.5318.3825.1%0.86120.711917.3%0.15550.253262.8%0.72000.581719.2%
2DGS [22]25.1518.8325.1%0.85780.720416.0%0.17560.291766.1%0.68980.578516.1%
Feature 3DGS [65]24.9119.5921.4%0.88000.786410.6%0.13770.227865.4%0.74270.646413.0%
Zip-NeRF [2]29.0617.3640.3%0.86600.671522.5%0.20780.358272.4%0.74790.584321.9%
Instant-NGP [35]25.6117.1533.0%0.85960.721216.1%0.33400.517154.8%0.72540.618214.8%
AVERAGE25.7519.1525.6%0.86650.739014.7%0.18700.300262.5%0.72200.615314.8%
Setting 33DGS [25]21.2214.9929.4%0.85500.716916.1%0.22520.405079.8%0.70020.477431.8%
3DGM [31]20.6214.6029.2%0.85430.723315.3%0.22540.404979.6%0.68740.467232.0%
GSPro [11]21.5814.8231.3%0.86340.699619.0%0.20100.387792.9%0.70930.454136.0%
VEGS [23]21.1314.2532.6%0.82660.647521.7%0.23590.442287.5%0.67850.444234.5%
PGSR [7]19.6014.1727.7%0.82380.698415.2%0.29340.436348.7%0.58670.378735.5%
2DGS [22]17.3511.3634.5%0.75680.544728.0%0.42960.545927.1%0.35520.232734.5%
Feature 3DGS [65]21.8814.3334.5%0.86430.638626.1%0.14000.3816172.6%0.74110.466937.0%
Zip-NeRF [2]20.6114.4230.0%0.83830.656521.7%0.21970.4546106.9%0.71080.364548.7%
Instant-NGP [35]19.6314.3926.7%0.81790.710413.1%0.49560.659233.0%0.60830.415731.7%
AVERAGE20.4014.1530.6%0.83340.670719.5%0.27400.457570.0%0.64190.411335.9%
", + "image_path": "8762fe8aee21c2d60ceb89757fa613178e7395fb6bfa8c2867fa41b4fc7afd75.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 58, + 497, + 290, + 569 + ], + "blocks": [ + { + "bbox": [ + 55, + 444, + 297, + 489 + ], + "lines": [ + { + "bbox": [ + 55, + 444, + 297, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 444, + 297, + 489 + ], + "type": "text", + "content": "Table 2. Quantitative comparison of depth evaluation in extrapolated views in Setting 1. 3DGM [31] demonstrates superior performance on most evaluation metrics, while VEGS [23] and GSPro [11] excel in SqRel and Delta1, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 497, + 290, + 569 + ], + "lines": [ + { + "bbox": [ + 58, + 497, + 290, + 569 + ], + "spans": [ + { + "bbox": [ + 58, + 497, + 290, + 569 + ], + "type": "table", + "html": "
BaselineAbsRel ↓RMSE ↓SqRel ↓Delta1 ↑Delta2 ↑Delta3 ↑
3DGS [25]0.36114.4410.410.6490.8240.895
3DGM [31]0.30113.938.9060.6510.8460.915
PGSR [7]0.36617.5721.500.7590.8340.883
GSPro [11]0.35519.6632.010.6430.8390.909
VEGS [23]0.36815.008.3980.4410.6910.827
", + "image_path": "6d06a4968b2da68a66458dcb048216312f128e1bcba7c93697539170691d92d4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 580, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 295, + 616 + ], + "type": "text", + "content": "[38] to mask out pixels corresponding to potentially movable objects during both training and evaluation and also exclude them from the initialization of 3DGS." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": "Evaluation Metrics. We use three widely-used metrics to evaluate visual quality: peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and learned perceptual image patch similarity (LPIPS) [61]. We also employ DINOv2 [36] feature cosine similarity to evaluate image quality in latent space. For geometry evaluation, we use depth metrics, including RMSE and " + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\delta_{1.25}" + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": ". Additionally, we incorporate the Fréchet Inception Distance (FID) [21] to" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 316, + 486, + 549, + 538 + ], + "blocks": [ + { + "bbox": [ + 313, + 444, + 555, + 478 + ], + "lines": [ + { + "bbox": [ + 313, + 444, + 555, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 555, + 478 + ], + "type": "text", + "content": "Table 3. Perceptual quality metric of VEGS [23]. Our benchmark also incorporates the Fréchet Inception Distance (FID) to evaluate the inpainting ability of models with generative priors." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 486, + 549, + 538 + ], + "lines": [ + { + "bbox": [ + 316, + 486, + 549, + 538 + ], + "spans": [ + { + "bbox": [ + 316, + 486, + 549, + 538 + ], + "type": "table", + "html": "
Setting 1Setting 2Setting 3AVERAGE
FID (In.) ↓46.333.546.142.0
FID (Ex.) ↓87.467.2132.495.7
FID (Drop)89%101%187%126.7%
", + "image_path": "9322a9ccabd85634e9bafa9debda576eff1aadb37c39e6137bb878f22d5237c2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 551, + 553, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 551, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 551, + 553, + 564 + ], + "type": "text", + "content": "evaluate the inpainting ability of models with strong priors." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 574, + 440, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 440, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 440, + 586 + ], + "type": "text", + "content": "4.2. Experimental Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 593, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 593, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 593, + 555, + 676 + ], + "type": "text", + "content": "Table 1 presents the quantitative results across Settings 1-3, while Figure 5 illustrates the qualitative outcomes on the extrapolated test set. Table 3 and Table 2 present the perceptual quality and depth metrics. The results indicate that, while the metrics perform relatively well in the interpolated test set, there is a significant drop in performance in the extrapolated test set across all baselines." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "Setting 1: Translation-only. In Setting 1, training views fully cover test views with moderate translational changes. (1) Results show a consistent drop in performance from" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28723" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 72, + 242, + 316 + ], + "blocks": [ + { + "bbox": [ + 61, + 72, + 242, + 316 + ], + "lines": [ + { + "bbox": [ + 61, + 72, + 242, + 316 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 242, + 316 + ], + "type": "image", + "image_path": "5481253f8bc9af45aa50f2778f2efc58a4a90a023e48c8b55a1490a87f41c2d2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 96, + 317, + 206, + 329 + ], + "lines": [ + { + "bbox": [ + 96, + 317, + 206, + 329 + ], + "spans": [ + { + "bbox": [ + 96, + 317, + 206, + 329 + ], + "type": "text", + "content": "Setting 1: Translation Only" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 244, + 72, + 380, + 316 + ], + "blocks": [ + { + "bbox": [ + 244, + 72, + 380, + 316 + ], + "lines": [ + { + "bbox": [ + 244, + 72, + 380, + 316 + ], + "spans": [ + { + "bbox": [ + 244, + 72, + 380, + 316 + ], + "type": "image", + "image_path": "544945f12614aa72971e9dcd893b56c88f38ebbcc03e182571dfcb3fd272fe91.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 263, + 318, + 361, + 330 + ], + "lines": [ + { + "bbox": [ + 263, + 318, + 361, + 330 + ], + "spans": [ + { + "bbox": [ + 263, + 318, + 361, + 330 + ], + "type": "text", + "content": "Setting 2: Rotation Only" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 382, + 72, + 545, + 316 + ], + "blocks": [ + { + "bbox": [ + 382, + 72, + 545, + 316 + ], + "lines": [ + { + "bbox": [ + 382, + 72, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 382, + 72, + 545, + 316 + ], + "type": "image", + "image_path": "a281b5f540c7d598c32ec069ca3181a039727fc9a85a60f86c9fbaf2d2c6fcb8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 398, + 317, + 528, + 328 + ], + "lines": [ + { + "bbox": [ + 398, + 317, + 528, + 328 + ], + "spans": [ + { + "bbox": [ + 398, + 317, + 528, + 328 + ], + "type": "text", + "content": "Setting 3: Rotation + Translation" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 332, + 555, + 366 + ], + "lines": [ + { + "bbox": [ + 55, + 332, + 555, + 366 + ], + "spans": [ + { + "bbox": [ + 55, + 332, + 555, + 366 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison of extrapolated view synthesis across different settings. For each setting, results from different methods are compared against the ground truth. Red boxes highlight areas where methods are limited in capturing fine details, such as road surfaces, sky regions, or object boundaries, demonstrating the challenges faced by each approach under varying movement complexities." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 372, + 295, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 372, + 295, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 295, + 504 + ], + "type": "text", + "content": "interpolation to extrapolation across all metrics, highlighting the challenge of generalization to views with variations. Relative drops vary by metric: PSNR falls " + }, + { + "bbox": [ + 55, + 372, + 295, + 504 + ], + "type": "inline_equation", + "content": "23 - 25\\%" + }, + { + "bbox": [ + 55, + 372, + 295, + 504 + ], + "type": "text", + "content": " (e.g., GSPro [11]: " + }, + { + "bbox": [ + 55, + 372, + 295, + 504 + ], + "type": "inline_equation", + "content": "21.51 \\rightarrow 16.39" + }, + { + "bbox": [ + 55, + 372, + 295, + 504 + ], + "type": "text", + "content": "), with SSIM and LPIPS showing similar declines. (2) On the extrapolative test set, methods perform comparably: GSPro tops PSNR (16.39) and minimizes LPIPS (0.2450), while 3DGM [31] leads in SSIM (0.7248). Both yield similar feature cosine similarity; VEGS [23] and PGSR [7] lag slightly. Results suggest that although GSPro marginally outperforms others, differences are small, underscoring the need for improved solutions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 507, + 295, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 507, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 507, + 295, + 662 + ], + "type": "text", + "content": "Setting 2: Rotation-only. In Setting 2, training views provide extensive coverage of the surrounding scene. However, most methods show poor generalization, with PSNR dropping by " + }, + { + "bbox": [ + 55, + 507, + 295, + 662 + ], + "type": "inline_equation", + "content": "22.75\\%" + }, + { + "bbox": [ + 55, + 507, + 295, + 662 + ], + "type": "text", + "content": ". Rotation changes are particularly challenging in textured regions (e.g., trees and intricate details), often resulting in blurring, while regions perpendicular to the vehicle remain difficult to capture. Additionally, distant areas pose reconstruction challenges, frequently leading to missing buildings and sky blackouts. Among the evaluated baselines, VEGS [23] and GSPro [11] stand out as the best-performing baselines in this setting: VEGS utilizes its diffusion prior for inpainting and refining missing regions, while GSPro's robust geometry handling improves generalization." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "content": "Setting 3: Translation + Rotation. In Setting 3, the view changes are the largest. (1) All methods exhibit notable performance drops from interpolation to extrapolation across metrics. For instance, 3DGS [25] experiences a PSNR" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 372, + 555, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 555, + 468 + ], + "type": "text", + "content": "drop of " + }, + { + "bbox": [ + 313, + 372, + 555, + 468 + ], + "type": "inline_equation", + "content": "29.4\\%" + }, + { + "bbox": [ + 313, + 372, + 555, + 468 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 372, + 555, + 468 + ], + "type": "inline_equation", + "content": "21.22 \\rightarrow 14.99" + }, + { + "bbox": [ + 313, + 372, + 555, + 468 + ], + "type": "text", + "content": "), while GSPro [11] undergoes a similar drop of " + }, + { + "bbox": [ + 313, + 372, + 555, + 468 + ], + "type": "inline_equation", + "content": "31.3\\%" + }, + { + "bbox": [ + 313, + 372, + 555, + 468 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 372, + 555, + 468 + ], + "type": "inline_equation", + "content": "21.58 \\rightarrow 14.82" + }, + { + "bbox": [ + 313, + 372, + 555, + 468 + ], + "type": "text", + "content": "). (2) Feature 3DGS [65] and 3DGM [31] stand out as leading methods, excelling in LPIPS (0.3816) and SSIM (0.7233), respectively. However, overall performance remains limited, with PSNR consistently falling below 15, highlighting significant room for improvement in generating high-fidelity outputs for extrapolated views." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 478, + 389, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 478, + 389, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 389, + 491 + ], + "type": "text", + "content": "5. Discussions" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 498, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 713 + ], + "type": "text", + "content": "Lighting Inconsistency Handling. Lighting inconsistencies are a widespread challenge in multi-traversal datasets due to varying illumination and weather conditions. To address this, we have taken specific measures. (1) Importantly, our multi-traversal dataset is manually curated to ensure that the lighting across images appears consistent to the eye. This step helps reduce the impact of extreme lighting changes and ensures a fairer evaluation of our method under controlled conditions. (2) One potential approach is to incorporate an appearance embedding for each image. We experiment with the Gaussian in the Wild (GS-W) baseline [59]. GS-W replaces traditional spherical harmonic-based color modeling with a method that separates intrinsic properties of each Gaussian point from dynamic appearance features of each image. This approach captures the stable, inherent appearance of objects while accommodating dynamic factors like highlights and shadows. Despite these efforts, our results, especially in Setting 3, still exhibit" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28724" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 301, + 208 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 301, + 208 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 301, + 208 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 301, + 208 + ], + "type": "image", + "image_path": "da6c60a73c394fcedf097463a5914d3d68971797e22304e6938efca9efb646be.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 210, + 275, + 220 + ], + "lines": [ + { + "bbox": [ + 82, + 210, + 275, + 220 + ], + "spans": [ + { + "bbox": [ + 82, + 210, + 275, + 220 + ], + "type": "text", + "content": "(a) Mitigating lighting inconsistency by camera embedding." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 551, + 208 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 551, + 208 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 551, + 208 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 551, + 208 + ], + "type": "image", + "image_path": "9739511aa7f4910e80e89a9e7eb4c373e6276bbd22fcc5af6056f54a9bad3e71.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 341, + 210, + 515, + 220 + ], + "lines": [ + { + "bbox": [ + 341, + 210, + 515, + 220 + ], + "spans": [ + { + "bbox": [ + 341, + 210, + 515, + 220 + ], + "type": "text", + "content": "(b) Qualitative dynamic scenes rendering comparison." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 225, + 555, + 258 + ], + "lines": [ + { + "bbox": [ + 54, + 225, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 54, + 225, + 555, + 258 + ], + "type": "text", + "content": "Figure 6. Qualitative results of dynamic baseline and lighting handling. We mitigate the lighting inconsistency issue by carefully selecting traversals under similar lighting conditions, and it can be further alleviated by introducing camera embeddings. We provide the quantitative evaluation of dynamic scene reconstruction methods in Setting 2." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 62, + 334, + 286, + 399 + ], + "blocks": [ + { + "bbox": [ + 55, + 270, + 295, + 324 + ], + "lines": [ + { + "bbox": [ + 55, + 270, + 295, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 295, + 324 + ], + "type": "text", + "content": "Table 4. Quantitative performance of GS-W [59] in different settings. After learning the lighting features, the interpolated and extrapolated test metrics show significant improvement compared to other baselines. However, there is still a considerable drop from interpolation to extrapolation." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 334, + 286, + 399 + ], + "lines": [ + { + "bbox": [ + 62, + 334, + 286, + 399 + ], + "spans": [ + { + "bbox": [ + 62, + 334, + 286, + 399 + ], + "type": "table", + "html": "
SettingPSNR ↑SSIM ↑LPIPS ↓Feat Cos Sim ↑
In.Ex.In.Ex.In.Ex.In.Ex.
Setting 128.1520.220.890.780.150.230.710.64
Setting 230.1021.210.910.820.130.200.760.67
Setting 328.6219.360.870.730.150.310.740.50
", + "image_path": "9bd9aca3939a6630077de23f5d942da933d8f5e31ba8ae7cc6252053785e3eac.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 409, + 295, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 409, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 295, + 445 + ], + "type": "text", + "content": "a significant performance drop, as shown in Table 4 and Figure 6a, underscoring the limitations of Gaussian-based models in handling extrapolated, unseen scenarios." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 445, + 295, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 295, + 648 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 295, + 648 + ], + "type": "text", + "content": "Dynamic Scenes. In addition to the static scene reconstruction methods that we evaluated in Table 1, some dynamic scene reconstruction techniques have recently emerged. We conduct evaluation for Setting 2 to assess these approaches, e.g., OmniRe [10]. It organizes rigid-deformable nodes and background nodes to capture dynamic scene structures and employs SMPL [32] for non-rigid object modeling. As illustrated in Figure 6b, the rendering results reveal a significant performance gap between the training and extrapolated test cameras. In the extrapolated test views, objects such as trees and stakes lose texture and geometric details, resulting in noticeably blurry outputs, whereas the training views maintain high fidelity. As shown in Table 5, the reconstruction metrics indicate an average drop of " + }, + { + "bbox": [ + 55, + 445, + 295, + 648 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 55, + 445, + 295, + 648 + ], + "type": "text", + "content": " when transitioning from interpolated to extrapolated settings. The results highlight the challenges of extrapolated view synthesis in dynamic scenes and the need for further research." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 657, + 291, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 291, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 291, + 670 + ], + "type": "text", + "content": "6. Conclusions, Limitations and Future Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "Conclusions. We introduce the first benchmark enabling quantitative evaluation of extrapolated view synthesis, advancing photorealistic simulation for self-driving" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 321, + 312, + 544, + 358 + ], + "blocks": [ + { + "bbox": [ + 313, + 270, + 554, + 303 + ], + "lines": [ + { + "bbox": [ + 313, + 270, + 554, + 303 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 554, + 303 + ], + "type": "text", + "content": "Table 5. Quantitative performance of OmniRe [10] in Setting 2. The experiment uses a single traversal with dynamic objects, showing a noticeable drop from interpolation to extrapolation." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 312, + 544, + 358 + ], + "lines": [ + { + "bbox": [ + 321, + 312, + 544, + 358 + ], + "spans": [ + { + "bbox": [ + 321, + 312, + 544, + 358 + ], + "type": "table", + "html": "
SettingPSNR ↑SSIM ↑LPIPS ↓Feat Cos Sim ↑
In.Ex.In.Ex.In.Ex.In.Ex.
Setting 219.7815.320.650.450.380.530.730.58
", + "image_path": "b2139876ec7719c95af06b6d0d6930f0c541c57a706658d91c4bbc8487ee49b9.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 367, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 555, + 510 + ], + "type": "text", + "content": "and robotics. Our benchmark integrates real-world multi-traversal, multi-agent, and multi-camera data, categorizes scenes into different evaluation settings, and evaluates state-of-the-art NVS models. Experimental results reveal that while some methods address specific challenges, current models demonstrate limited generalization, with significant overfitting to training views and suboptimal performance in extrapolated view synthesis. To support further research, we will release the dataset and benchmark, addressing the long-standing data scarcity and providing evaluation protocols. We believe the EUVS benchmark will catalyze meaningful advancements in self-driving and robotics innovation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 510, + 556, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 641 + ], + "type": "text", + "content": "Limitations and Future Work. Our benchmark does have limitations. First, although we provide the ground truth to enable quantitative evaluation for static scenes in all settings and the foreground objects evaluation in Setting 2, we lack the foreground evaluation in Settings 1 and 3. Future work will aim to expand our work to include the dynamic object evaluation in Settings 1 and 3 using multi-agent data. Secondly, although we carefully manually selected and ensured that the test trajectory viewpoints are well covered by training trajectories, a more in-depth evaluation of how observed and unseen regions differ is left for future work." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 642, + 556, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 702 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 702 + ], + "type": "text", + "content": "Acknowledgment. This work was supported in part through NSF grants 2238968 and 2121391, and the NYU IT High Performance Computing resources, services, and staff expertise. Yiming Li is supported by NVIDIA Graduate Fellowship (2024-2025)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28725" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "text", + "content": "[1] Alexander Amini, Tsun-Hsuan Wang, Igor Gilitschenski, Wilko Schwarting, Zhijian Liu, Song Han, Sertac Karaman, and Daniela Rus. Vista 2.0: An open, data-driven simulator for multimodal sensing and policy learning for autonomous vehicles. In 2022 International Conference on Robotics and Automation (ICRA), pages 2419-2426. IEEE, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 158, + 296, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 158, + 296, + 212 + ], + "spans": [ + { + "bbox": [ + 61, + 158, + 296, + 212 + ], + "type": "text", + "content": "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023. 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 213, + 296, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 213, + 296, + 278 + ], + "spans": [ + { + "bbox": [ + 62, + 213, + 296, + 278 + ], + "type": "text", + "content": "[3] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 280, + 296, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 280, + 296, + 334 + ], + "spans": [ + { + "bbox": [ + 62, + 280, + 296, + 334 + ], + "type": "text", + "content": "[4] Holger Caesar, Juraj Kabzan, Kok Seang Tan, Whye Kit Fong, Eric Wolff, Alex Lang, Luke Fletcher, Oscar Beijbom, and Sammy Omari. nuplan: A closed-loop ml-based planning benchmark for autonomous vehicles. arXiv preprint arXiv:2106.11810, 2021. 2, 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 335, + 296, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 335, + 296, + 401 + ], + "spans": [ + { + "bbox": [ + 62, + 335, + 296, + 401 + ], + "type": "text", + "content": "[5] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3d tracking and forecasting with rich maps. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8748-8757, 2019. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 402, + 296, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 402, + 296, + 465 + ], + "spans": [ + { + "bbox": [ + 62, + 402, + 296, + 465 + ], + "type": "text", + "content": "[6] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19467, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 468, + 296, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 468, + 296, + 523 + ], + "spans": [ + { + "bbox": [ + 62, + 468, + 296, + 523 + ], + "type": "text", + "content": "[7] Danpeng Chen, Hai Li, Weicai Ye, Yifan Wang, Weijian Xie, Shangjin Zhai, Nan Wang, Haomin Liu, Hujun Bao, and Guofeng Zhang. Pgsr: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction. arXiv preprint arXiv:2406.06521, 2024. 5, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 525, + 296, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 525, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 62, + 525, + 296, + 567 + ], + "type": "text", + "content": "[8] Yurui Chen, Chun Gu, Junzhe Jiang, Xiatian Zhu, and Li Zhang. Periodic vibration gaussian: Dynamic urban scene reconstruction and real-time rendering. arXiv preprint arXiv:2311.18561, 2023. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 569, + 296, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 569, + 296, + 622 + ], + "spans": [ + { + "bbox": [ + 62, + 569, + 296, + 622 + ], + "type": "text", + "content": "[9] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splattering from sparse multi-view images. arXiv preprint arXiv:2403.14627, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 624, + 295, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 677 + ], + "type": "text", + "content": "[10] Ziyu Chen, Jiawei Yang, Jiahui Huang, Riccardo de Lutio, Janick Martinez Esturo, Boris Ivanovic, Or Litany, Zan Gojcic, Sanja Fidler, Marco Pavone, et al. Omnire: Omni urban scene reconstruction. arXiv preprint arXiv:2408.16760, 2024. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "type": "text", + "content": "[11] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussian-pro: 3d gaussian splattering with progressive propagation. In" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "Forty-first International Conference on Machine Learning, 2024. 4, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 96, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 171 + ], + "type": "text", + "content": "[12] Carlos A Diaz-Ruiz, Youya Xia, Yurong You, Jose Nino, Junan Chen, Josephine Monica, Xiangyu Chen, Katie Luo, Yan Wang, Marc Emond, et al. Ithaca365: Dataset and driving perception under repeated and challenging weather conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21383-21392, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 173, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 553, + 215 + ], + "type": "text", + "content": "[13] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 1, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 217, + 553, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 281 + ], + "type": "text", + "content": "[14] Scott Ettinger, Shuyang Cheng, Benjamin Caine, Chenxi Liu, Hang Zhao, Sabeek Pradhan, Yuning Chai, Ben Sapp, Charles R Qi, Yin Zhou, et al. Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9710-9719, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 283, + 553, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 553, + 315 + ], + "type": "text", + "content": "[15] Lue Fan, Hao Zhang, Qitai Wang, Hongsheng Li, and Zhaoxiang Zhang. Freesim: Toward free-viewpoint camera simulation in driving scenes, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 316, + 553, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 316, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 317, + 316, + 553, + 369 + ], + "type": "text", + "content": "[16] Lan Feng, Quanyi Li, Zhenghao Peng, Shuhan Tan, and Bolei Zhou. Trafficgen: Learning to generate diverse and realistic traffic scenarios. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 3567-3575. IEEE, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 372, + 553, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 424 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 424 + ], + "type": "text", + "content": "[17] Jannik Fritsch, Tobias Kuehnl, and Andreas Geiger. A new performance measure and evaluation benchmark for road detection algorithms. In 16th International IEEE Conference on Intelligent Transportation Systems (ITSC 2013), pages 1693-1700. IEEE, 2013. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 426, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 426, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 317, + 426, + 553, + 480 + ], + "type": "text", + "content": "[18] Shenyuan Gao, Jiazhi Yang, Li Chen, Kashyap Chitta, Yihang Qiu, Andreas Geiger, Jun Zhang, and Hongyang Li. Vista: A generalizable driving world model with high fidelity and versatile controllability. arXiv preprint arXiv:2405.17398, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 482, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 482, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 482, + 553, + 525 + ], + "type": "text", + "content": "[19] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 526, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 553, + 568 + ], + "type": "text", + "content": "[20] Huasong Han, Kaixuan Zhou, Xiaoxiao Long, Yusen Wang, and Chunxia Xiao. Ggs: Generalizable gaussian splatting for lane switching in autonomous driving. arXiv preprint arXiv:2409.02382, 2024. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 570, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 570, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 317, + 570, + 553, + 623 + ], + "type": "text", + "content": "[21] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "type": "text", + "content": "[22] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3, 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[23] Sungwon Hwang, Min-Jung Kim, Taewoong Kang, Jayeon Kang, and Jaegul Choo. Vegs: View extrapolation of urban scenes in 3d gaussian splatting using learned priors. arXiv preprint arXiv:2407.02945, 2024. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "28726" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 139 + ], + "type": "text", + "content": "[24] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5322-5332, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 296, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 296, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 296, + 185 + ], + "type": "text", + "content": "[25] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 186, + 296, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 296, + 231 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 296, + 231 + ], + "type": "text", + "content": "[26] Mustafa Khan, Hamidreza Fazlali, Dhruv Sharma, Tongtong Cao, Dongfeng Bai, Yuan Ren, and Bingbing Liu. Autosplat: Constrained gaussian splatting for autonomous driving scene reconstruction. arXiv preprint arXiv:2407.02598, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 232, + 296, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 296, + 286 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 296, + 286 + ], + "type": "text", + "content": "[27] Seung Wook Kim, Jonah Philion, Antonio Torralba, and Sanja Fidler. Drivegan: Towards a controllable high-quality neural simulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5820-5829, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 287, + 296, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 287, + 296, + 343 + ], + "spans": [ + { + "bbox": [ + 56, + 287, + 296, + 343 + ], + "type": "text", + "content": "[28] Quanyi Li, Zhenghao Mark Peng, Lan Feng, Zhizheng Liu, Chenda Duan, Wenjie Mo, and Bolei Zhou. ScenarioNet: Open-source platform for large-scale traffic scenario simulation and modeling. Advances in neural information processing systems, 36, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 344, + 296, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 344, + 296, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 344, + 296, + 398 + ], + "type": "text", + "content": "[29] Yiming Li, Dekun Ma, Ziyan An, Zixun Wang, Yiqi Zhong, Siheng Chen, and Chen Feng. V2x-sim: Multi-agent collaborative perception dataset and benchmark for autonomous driving. IEEE Robotics and Automation Letters, 7(4): 10914-10921, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 399, + 296, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 296, + 464 + ], + "type": "text", + "content": "[30] Yiming Li, Zhiheng Li, Nuo Chen, Moonjun Gong, Zonglin Lyu, Zehong Wang, Peili Jiang, and Chen Feng. Multiagent multitraversal multimodal self-driving: Open mars dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22041-22051, 2024. 2, 3, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 466, + 296, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 466, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 466, + 296, + 521 + ], + "type": "text", + "content": "[31] Yiming Li, Zehong Wang, Yue Wang, Zhiding Yu, Zan Gojcic, Marco Pavone, Chen Feng, and Jose M. Alvarez. Memorize what matters: Emergent scene decomposition from multitraverse. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 4, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 523, + 296, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 296, + 567 + ], + "type": "text", + "content": "[32] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1-248:16, 2015. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 568, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 296, + 612 + ], + "type": "text", + "content": "[33] Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3061-3070, 2015. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 613, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 296, + 667 + ], + "type": "text", + "content": "[34] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 669, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 296, + 713 + ], + "type": "text", + "content": "[35] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 5, 6" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 129 + ], + "type": "text", + "content": "[36] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 129, + 554, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 554, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 554, + 183 + ], + "type": "text", + "content": "[37] Matthew Pitropov, Danson Evan Garcia, Jason Rebello, Michael Smart, Carlos Wang, Krzysztof Czarnecki, and Steven Waslander. Canadian adverse driving conditions dataset. The International Journal of Robotics Research, 40 (4-5):681-690, 2021. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 184, + 554, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 554, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 554, + 259 + ], + "type": "text", + "content": "[38] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dólar, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos, 2024. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 260, + 554, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 260, + 554, + 303 + ], + "spans": [ + { + "bbox": [ + 316, + 260, + 554, + 303 + ], + "type": "text", + "content": "[39] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 304, + 554, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 554, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 554, + 360 + ], + "type": "text", + "content": "[40] Shital Shah, Debadeepta Dey, Chris Lovett, and Ashish Kapoor. Airsim: High-fidelity visual and physical simulation for autonomous vehicles. In Field and Service Robotics: Results of the 11th International Conference, pages 621-635. Springer, 2018. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 361, + 554, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 554, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 554, + 403 + ], + "type": "text", + "content": "[41] Xi Shi, Lingli Chen, Peng Wei, Xi Wu, Tian Jiang, Yonggang Luo, and Lecheng Xie. Dhgs: Decoupled hybrid gaussian splatting for driving scene. arXiv preprint arXiv:2407.16600, 2024.3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 404, + 554, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 554, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 554, + 459 + ], + "type": "text", + "content": "[42] Shuhan Tan, Kelvin Wong, Shenlong Wang, Sivabalan Manivasagam, Mengye Ren, and Raquel Urtasun. Scenegen: Learning to generate realistic traffic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 892-901, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 460, + 554, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 554, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 554, + 514 + ], + "type": "text", + "content": "[43] Izzeddin Teeti, Valentina Musat, Salman Khan, Alexander Rast, Fabio Cuzzolin, and Andrew Bradley. Vision in adverse weather: Augmentation using cyclegans with various object detectors for robust perception in autonomous racing. arXiv preprint arXiv:2201.03246, v3, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 515, + 554, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 554, + 580 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 554, + 580 + ], + "type": "text", + "content": "[44] Adam Tondersski, Carl Lindstrom, Georg Hess, William Ljungbergh, Lennart Svensson, and Christoffer Petersson. Neurad: Neural rendering for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14895-14904, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 581, + 554, + 625 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 581, + 554, + 625 + ], + "spans": [ + { + "bbox": [ + 316, + 581, + 554, + 625 + ], + "type": "text", + "content": "[45] Christopher Wewer, Kevin Raj, Eddy Ilg, Bernt Schiele, and Jan Eric Lenssen. latentsplat: Autoencoding variational gaussians for fast generalizable 3d reconstruction. arXiv preprint arXiv:2403.16292, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 626, + 554, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 626, + 554, + 691 + ], + "spans": [ + { + "bbox": [ + 316, + 626, + 554, + 691 + ], + "type": "text", + "content": "[46] Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, et al. Argoverse 2: Next generation datasets for self-driving perception and forecasting. arXiv preprint arXiv:2301.00493, 2023. 2, 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "type": "text", + "content": "[47] Chenming Wu, Jiadai Sun, Zhelun Shen, and Liangjun Zhang. Mapnerf: Incorporating map priors into neural ra" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28727" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 295, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 295, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 295, + 106 + ], + "type": "text", + "content": "diance fields for driving view simulation. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7082-7088. IEEE, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 163 + ], + "type": "text", + "content": "[48] Ke Wu, Kaizhao Zhang, Zhiwei Zhang, Shanshuai Yuan, Muer Tie, Julong Wei, Zijun Xu, Jieru Zhao, Zhongxue Gan, and Wenchao Ding. Hgs-mapping: Online dense mapping using hybrid gaussian representation in urban scenes. arXiv preprint arXiv:2403.20159, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 164, + 295, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 295, + 229 + ], + "type": "text", + "content": "[49] Zirui Wu, Tianyu Liu, Liyi Luo, Zhide Zhong, Jianteng Chen, Hongmin Xiao, Chao Hou, Haozhe Lou, Yuantao Chen, Runyi Yang, et al. Mars: An instance-aware, modular and realistic simulator for autonomous driving. In CAAI International Conference on Artificial Intelligence, pages 3-15. Springer, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 231, + 295, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 295, + 275 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 295, + 275 + ], + "type": "text", + "content": "[50] Bernhard Wymann, Eric Espie, Christophe Guionneau, Christos Dimitrakakis, Rémi Coulom, and Andrew Sumner. Torcs, the open racing car simulator. Software available at http://torcs.sourceforge.net, 4(6):2, 2000. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 295, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 295, + 319 + ], + "type": "text", + "content": "[51] Danfei Xu, Yuxiao Chen, Boris Ivanovic, and Marco Pavone. Bits: Bi-level imitation for traffic simulation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2929-2936. IEEE, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 321, + 295, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 295, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 295, + 397 + ], + "type": "text", + "content": "[52] Runsheng Xu, Xin Xia, Jinlong Li, Hanzhao Li, Shuo Zhang, Zhengzhong Tu, Zonglin Meng, Hao Xiang, Xiaoyu Dong, Rui Song, et al. V2v4real: A real-world large-scale dataset for vehicle-to-vehicle cooperative perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13712-13722, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 399, + 295, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 295, + 443 + ], + "type": "text", + "content": "[53] Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians for modeling dynamic urban scenes. arXiv preprint arXiv:2401.01339, 2024. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 445, + 295, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 295, + 510 + ], + "type": "text", + "content": "[54] Chen Yang, Peihao Li, Zanwei Zhou, Shanxin Yuan, Bingbing Liu, Xiaokang Yang, Weichao Qiu, and Wei Shen. Ner-fvs: Neural radiance fields for free view synthesis via geometry scaffolds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16549-16558, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 512, + 295, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 295, + 567 + ], + "type": "text", + "content": "[55] Jiawei Yang, Boris Ivanovic, Or Litany, Xinshuo Weng, Seung Wook Kim, Boyi Li, Tong Che, Danfei Xu, Sanja Fidler, Marco Pavone, et al. Emernerf: Emergent spatial-temporal scene decomposition via self-supervision. arXiv preprint arXiv:2311.02077, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 568, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 295, + 633 + ], + "type": "text", + "content": "[56] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1389-1399, 2023. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 635, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 295, + 689 + ], + "type": "text", + "content": "[57] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19447-19456, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 295, + 714 + ], + "type": "text", + "content": "[58] Zhongrui Yu, Haoran Wang, Jinze Yang, Hanzhang Wang, Zeke Xie, Yunfeng Cai, Jiale Cao, Zhong Ji, and Mingming" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 596 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Sun. Sgd: Street view synthesis with gaussian splatting and diffusion prior. arXiv preprint arXiv:2403.20079, 2024. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 97, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 140 + ], + "type": "text", + "content": "[59] Dongbin Zhang, Chuming Wang, Weitao Wang, Peihao Li, Minghan Qin, and Haoqian Wang. Gaussian in the wild: 3d gaussian splatting for unconstrained image collections. arXiv preprint arXiv:2403.15704, 2024. 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 141, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 217 + ], + "type": "text", + "content": "[60] Jian Zhang, Yuanqing Zhang, Huan Fu, Xiaowei Zhou, Bowen Cai, Jinchi Huang, Rongfei Jia, Binqiang Zhao, and Xing Tang. Ray priors through reprojection: Improving neural radiance fields for novel view extrapolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18376-18386, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 219, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 274 + ], + "type": "text", + "content": "[61] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 275, + 553, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 275, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 275, + 553, + 329 + ], + "type": "text", + "content": "[62] Guosheng Zhao, Chaojun Ni, Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, Boyuan Wang, Youyi Zhang, Wenjun Mei, and Xingang Wang. Drivedreamer4d: World models are effective data machines for 4d driving scene representation. arXiv preprint arXiv:2410.13571, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 331, + 553, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 385 + ], + "type": "text", + "content": "[63] Ziyuan Zhong, Davis Rempe, Danfei Xu, Yuxiao Chen, Sushant Veer, Tong Che, Baishakhi Ray, and Marco Pavone. Guided conditional diffusion for controllable traffic simulation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 3560-3566. IEEE, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 387, + 553, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 387, + 553, + 451 + ], + "spans": [ + { + "bbox": [ + 316, + 387, + 553, + 451 + ], + "type": "text", + "content": "[64] Hongyu Zhou, Jiahao Shao, Lu Xu, Dongfeng Bai, Weichao Qiu, Bingbing Liu, Yue Wang, Andreas Geiger, and Yiyi Liao. Hugs: Holistic urban 3d scene understanding via gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21336-21345, 2024. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 453, + 553, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 453, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 316, + 453, + 553, + 529 + ], + "type": "text", + "content": "[65] Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Zehao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21676-21685, 2024. 5, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 530, + 553, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 530, + 553, + 596 + ], + "spans": [ + { + "bbox": [ + 316, + 530, + 553, + 596 + ], + "type": "text", + "content": "[66] Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21634-21643, 2024. 2, 3" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28728" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/F-Bench_ Rethinking Human Preference Evaluation Metrics for Benchmarking Face Generation, Customization, and Restoration/86d5740f-dff5-47ab-aa67-761b97593a44_content_list.json b/2025/F-Bench_ Rethinking Human Preference Evaluation Metrics for Benchmarking Face Generation, Customization, and Restoration/86d5740f-dff5-47ab-aa67-761b97593a44_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cf93b3980b5e8a3eccc2033e1cc9bf338c278b33 --- /dev/null +++ b/2025/F-Bench_ Rethinking Human Preference Evaluation Metrics for Benchmarking Face Generation, Customization, and Restoration/86d5740f-dff5-47ab-aa67-761b97593a44_content_list.json @@ -0,0 +1,1905 @@ +[ + { + "type": "text", + "text": "F-Bench: Rethinking Human Preference Evaluation Metrics for Benchmarking Face Generation, Customization, and Restoration", + "text_level": 1, + "bbox": [ + 99, + 128, + 898, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lu Liu $^{1,*}$ , Huiyu Duan $^{1,*}$ , Qiang Hu $^{1,\\dagger}$ , Liu Yang $^{1}$ , Chunlei Cai $^{2}$", + "bbox": [ + 241, + 202, + 751, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tianxiao Ye $^{2}$ , Huayu Liu $^{1}$ , Xiaoyun Zhang $^{1}$ , Guangtao Zhai $^{1}$", + "bbox": [ + 256, + 220, + 738, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Shanghai Jiao Tong University, Shanghai, China 2Bilibili Inc., Shanghai, China", + "bbox": [ + 186, + 239, + 812, + 257 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/be54ea568985a727a2cc4028c6a24db27d8ed7043d43028000e0e447b132f563.jpg", + "image_caption": [ + "Figure 1. Data construction pipeline of $F$ -bench and content overview of FaceQ database. From left to right, (a) Diverse input source data for face generation, customization, and restoration. (b) $F$ -Bench benchmarks 29 face generative models, including 14 face generation models, 6 face customization models, and 9 face restoration models. (c) Multi-dimensional subjective preference evaluation in $F$ -Bench. (d) 3D scatter plots of MOS distributions across FaceQ three subsets." + ], + "image_footnote": [], + "bbox": [ + 94, + 266, + 330, + 483 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/50a31b7cb83ff8691ad34b7636a1481484922af6f46b1659ee084656b72f3aeb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 330, + 267, + 665, + 479 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cf85e2bc4cae21c3c7775c8426f5183047677a9fc1bbce0840f052dde717a00a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 665, + 267, + 903, + 482 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 549, + 326, + 565 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent artificial intelligence (AI) generative models have demonstrated remarkable capabilities in image production, and have been widely applied to face image generation, customization, and restoration. However, many AI-generated faces (AIGFs) still suffer from issues such as unique distortions, unrealistic details, and unexpected identity shifts, underscoring the need for a comprehensive quality evaluation method for AIGFs. To this end, we introduce FaceQ, the first comprehensive AI-generated Face image database with fine-grained Quality annotations aligned with human preferences, which consists of 12K images and 491K ratings across multiple dimensions. Using the FaceQ database, we establish F-Bench, a benchmark for comparing and evaluating face generation, customization, and restoration models, highlighting strengths and weaknesses across various prompts and evaluation dimensions. Additionally, we assess the performance of existing image quality assessment (IQA) methods on FaceQ, and further propose a large multimodal model (LMM) based Face quality Evaluator (F", + "bbox": [ + 88, + 583, + 485, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Eval) to accurately assess the multi-dimensional quality of generated faces in a one-for-all manner. Extensive experimental results demonstrate the state-of-the-art performance of our F-Eval. The project page is: https://media-sjtu.github.io/F-Bench/.", + "bbox": [ + 511, + 550, + 908, + 626 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 660, + 643, + 676 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generative models, including VAE [47], GAN [26], and diffusion models [33, 74, 80], have made remarkable progress in recent years, and have been extended to the fields of text-to-image generation [3], text-driven image editing [44], image restoration [99], etc, leading to high-quality artificial intelligence generated images (AIGIs). As an important aspect of AIGIs, many studies have also investigated face generation [100], face customization [57, 108], and face restoration [62, 87], producing creative, high-quality face images. Despite these advances, many AI-generated faces (AIGFs) still suffer from perceptual deficiencies.", + "bbox": [ + 509, + 686, + 906, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Many natural face image quality assessment databases have been established in the literature [48, 51, 59, 86]. However, natural images are mainly degraded by distor", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution. † Corresponding author.", + "bbox": [ + 114, + 886, + 359, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10982", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/242cd0ee9f6c552fbdf553649873ffa2f11f98908744e1f2698cbb8ec4ab3d82.jpg", + "table_caption": [ + "Table 1. Comparison between FaceQ and existing AIGC quality assessment databases (top) as well as natural face quality assessment databases (bottom). FaceQ is the first large-scale AI-generated, customized and restored face quality assessment database." + ], + "table_footnote": [], + "table_body": "
DatabaseDomainSourceImagesScoresDimensions
AGIQA-3K (TCSVT2024) [51]GeneralAI-generated2,982MOSPerception, Alignment
AIGCIQA (CVPR2024) [86]GeneralAI-generated2,400MOSQuality, Authenticity, Correspondence
RichHF-18K (CVPR2024) [59]GeneralAI-generated17,760MOSPlausibility, Aesthetics, Text-image Alignment, Overall
PIQ23 (CVPR2023) [13]FaceReal5,116PairOverall, Details, Exposure
FaceQFaceAI-generated, customized and restored12,255MOSQuality, Authenticity, ID Fidelity, Correspondence
", + "bbox": [ + 89, + 116, + 903, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tions such as noise, blur, compression [23, 66, 110], etc., while AIGIs generally suffer from unique distortions including unrealistic structures and text-image inconsistencies [59, 86], etc. Recently, many studies have investigated human preferences for AIGIs [39, 50, 59, 86, 103]. However, these general AIGI quality assessment works may not be well aligned with human preferences on AIGFs, since the perceptual quality of AIGFs can be significantly affected by small deviations in facial attributes,(e.g., eye size, skin texture, identity shifts), while the number of AIGFs in these AIGI databases is limited. Moreover, current AIGI quality assessment works mainly focus on the image generation, while human preferences for face customization have rarely been considered, where identity fidelity is significant.", + "bbox": [ + 88, + 199, + 480, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The above limitation highlights the necessity of establishing a comprehensive quality evaluator for AIGFs. To this end, we propose FaceQ, the first AI-generated Face image Quality assessment database with three popular tasks. Specifically, FaceQ contains $12\\mathrm{K}+$ images obtained from 29 generative models across three tasks: 1) face generation, 2) face customization, and 3) face restoration, as shown in Fig.1. Based on the generated images, we collect $491\\mathrm{K}+$ quality ratings from multiple dimensions including quality, authenticity, identity (ID) fidelity, and text-image correspondence, respectively, finally forming $32\\mathrm{K}+$ mean opinion scores (MOSs). Using FaceQ, we establish a human preference benchmark, i.e., F-Bench, for face generation, customization, and restoration models, and thoroughly analyze the generation performance across different evaluation perspectives, models, and categories. In addition, we also compare the correlation of existing image quality assessment (IQA), face quality assessment (FQA), AIGI quality assessment, and preference evaluation metrics with human preferences.", + "bbox": [ + 91, + 412, + 483, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, we present F-Eval, the first generated Face image quality Evaluation method designed to perform comprehensive multi-dimensional quality evaluation for generated, customized, and restorted faces in a one-for-all manner. Specifically, F-Eval leverages large multimodal models (LMMs) [11] and adopts instruction tuning [63] for preference-related vision-language adaption. We also apply low-rank adaptation (LoRA) [35] techniques for more efficient feature refining. Extensive experimental results demonstrate that F-Eval outperforms existing quality assessment methods in terms of multiple dimensions relevant to human preference for all three tasks. The main contribu", + "bbox": [ + 89, + 719, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tions of our work are highlighted as follows:", + "bbox": [ + 511, + 199, + 807, + 215 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We construct FaceQ, the first generated face quality assessment database consisting $491\\mathrm{K}+$ human preference ratings from quality, authenticity, ID fidelity, correspondence dimensions, on $12\\mathrm{K}+$ generative face images across three tasks.", + "- We establish F-Bench, a comprehensive human preference benchmark for 29 face generation, customization, and restoration methods. It also extensively evaluates the alignment of 26 existing quality assessment methods with human preferences.", + "- We present F-Eval, a LMM-based quality assessment model to predict multi-dimensional human preference on generated faces. Extensive experiments demonstrate the state-of-art QA performance of the proposed F-Eval." + ], + "bbox": [ + 511, + 220, + 903, + 431 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 450, + 655, + 465 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Face Generation & Customization & Restoration. (1) For face generation models, earlier works such as GANs [26, 43, 85] have demonstrated significant influences. Recently, diffusion models [73, 74, 77] have achieved rapid developments in text-to-image generation. These models can be categorized into pixel-space diffusion models [1, 72, 73, 76], and latent space diffusion models [17, 58, 71, 74]. (2) For face customization, this is a recently defined task that aims to create images of a specific person with the input of an identity reference image and prompts [98]. It can be divided into two categories including conventional test-time optimization [24, 49, 75] and tuning-free customization methods utilizing pre-trained diffusion models [20, 30, 57, 91, 101, 108]. (3) For face restoration, which aims to restore high-quality face images from their low-quality counterparts. Some traditional restoration methods utilize geometric priors [16, 19, 46] or reference priors [54, 55] while in recent years, exploring generative priors such as GAN priors [15, 28, 94, 107, 113] and diffusion priors [36, 37, 62, 84, 87, 97, 109] has become a prevalent trend. Despite these improvements, the generated, customized, and restored images may suffer from various quality issues, highlighting the necessity of evaluators.", + "bbox": [ + 511, + 477, + 906, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Face Image Quality Assessment Database. Existing face image quality assessment (FIQA) can be divided into Biometric FIQA [12, 31, 60, 69] and Generic FIQA [13, 18, 82, 92]. BFIQA methods are most developed to evaluate the biometric utility for face recognition. GFIQA meth", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10983", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0ca6f16b7c42b7cffbe0d97807f01593ad18168ac065d62cf16b2ebdca062ecf.jpg", + "image_caption": [ + "Figure 2. Rating comparisons of eight dimensions. Each column presents a pair of intuitive examples of each dimension, with red indicating the better rating and blue indicating the worse one. From left to right, the subsets are face generation, face customization, and face restoration subsets. The last row displays the corresponding prompts, reference image-prompt pairs, and the GT-LQ image pairs." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 395, + 256 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/15cf956e90dd11acd6ae04bce2981f7b38875f0c25c21bc20e03bb64418aa45d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 88, + 694, + 256 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/959bcd90fe7fba8303ac06245c48e7d7a9e96871c01f3bfb428c7c084619b239.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 88, + 905, + 256 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ods [82] focus on perceptual degradation in face restoration task. Both of these databases consist of natural face images, which has distinctly different nature with generated faces.", + "bbox": [ + 89, + 303, + 480, + 348 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "AIGC Quality Assessment Database. In recent years, plenty of researchers have raised interest in assessing AI-generated images(AIGIs) quality assessment [23, 40, 48, 51, 86, 104] by building various databases as shown in Tab 1. All of these existing database contain various categories of general AIGI images, leaving a notable gap in datasets dedicated to AI-generated face images.", + "bbox": [ + 89, + 349, + 482, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. FaceQ Database Construction", + "text_level": 1, + "bbox": [ + 89, + 467, + 364, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce the proposed generative face quality assessment database FaceQ. The database includes 12,255 images and 491,130 human opinion ratings. Our FaceQ demonstrates the merits of comprehensive generated visual content and multi-dimensional labeling, which can significantly facilitate the advancement of AIGC research.", + "bbox": [ + 89, + 493, + 483, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Data Collection", + "text_level": 1, + "bbox": [ + 89, + 592, + 246, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1. FaceQ-Gen Subset Collection", + "text_level": 1, + "bbox": [ + 89, + 614, + 339, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prompt Sources. The prompts in FaceQ-Gen subset are sourced from MS-COCO [61] [57] and GPT-4o [10]. These face-centric prompts can be categorized into nine types: General, Accessories, Clothes, Action, View, Expression, Background, Style, and Facial Attributes.", + "bbox": [ + 89, + 633, + 482, + 708 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Models Collection. We collect 14 representative face generation models, including Deep Floyd [1], Stable Diffusion V1.5 [74], Stable Diffusion V2.1 [74], SD-XL [71], Realistic Vision [8], Stable Cascade [70], ProtoVision V6.6 [7], SD3 [9], Playground V2.5 [52], Kolors [6], PixArt-alpha [17], Hunyuan [58], and Flux-dev [3]. Negative prompts and safe sensors are enabled. In total, FaceQ-Gen includes 4,032 generated face images.", + "bbox": [ + 89, + 709, + 482, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.2. FaceQ-Cus Subset Collection", + "text_level": 1, + "bbox": [ + 89, + 835, + 336, + 851 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Identity and Prompt Sources. Face customization relies on a reference identity image and a guiding prompt. The reference images in FaceQ-Cus are partially self-collected,", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with the rest sourced from [57]. The prompts are selected and filtered from the FaceQ-Gen prompt list.", + "bbox": [ + 511, + 303, + 906, + 334 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Models Collection. We collect six representative face customization methods, including ELITE [98], FastComposer [101], IP-Adapter-FaceID [4], InstantID [91], IP-Adapter-Face-ID-Plus [5], and PhotoMaker V2 [57]. Here, we use single reference image as input. In total, FaceQ-Cus includes 4,200 generated face images.", + "bbox": [ + 511, + 335, + 906, + 426 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.3. FaceQ-Res Subset Collection", + "text_level": 1, + "bbox": [ + 511, + 436, + 756, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Low Quality Images Source. (1) Synthetic: We construct two degraded pipelines following previous works [15, 93, 95, 97] to simulate real-world degradations on both indomain data from VFHQ [90], CelebRef [56] and out-of-domain self-collected high-quality Asian face images (2) Real-world: Real-world low-quality images are sourced from various public datasets, including AgeAB [68], CelebChild [28], Wider [105], LFW [38] and MegaFace [45], ensuring the degradation diversity.", + "bbox": [ + 511, + 455, + 905, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Models Collection and Face Image Restoration. We evaluate both diffusion-prior-based methods such as DiffBIR [62], DifFace [109], StableSR [87], and GAN or Transformer-based methods GFPGAN [94], GPEN [107], CodeFormer [113], VQFR [28] and Face-SPARNet [15]. Each method is benchmarked on both the $F$ -Res synthetic dataset and the real-world dataset. In total, FaceQ-Res include 4023 restored images, with 2007 images in the real-world case and 2016 images in the synthetic case.", + "bbox": [ + 511, + 593, + 905, + 729 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Human Preference Dimension Design", + "text_level": 1, + "bbox": [ + 511, + 741, + 836, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The evaluation dimensions of generated faces need specific design to adequately reflect human preference, given the complexity of perceptual factors involved. To the end, we decompose \"quality\" into four core dimensions: Quality, Authenticity, ID Fidelity, and Correspondence.", + "bbox": [ + 511, + 763, + 905, + 840 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Quality. This dimension evaluates the overall perceptual quality of the image, considering factors such as color accuracy, blur, noise, and artifacts. It is used in all tasks.", + "bbox": [ + 511, + 840, + 903, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Authenticity. This dimension evaluates how closely the", + "bbox": [ + 511, + 885, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10984", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/906cf53f1707bfcabc0ffea135230ee9f39b5e4307d42b8189ee2f896f3f2245.jpg", + "image_caption": [ + "(a) Face generation MOS distribution" + ], + "image_footnote": [], + "bbox": [ + 94, + 104, + 357, + 172 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f75d4c146079fe3a3a587bb40ab780de22b24212ea513086b7e2ba41df47cde3.jpg", + "image_caption": [ + "(b) Face customization MOS distribution" + ], + "image_footnote": [], + "bbox": [ + 349, + 104, + 629, + 174 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e6fc96cb02e178976b89dc1fedc2b07e1fc527d99f6a9e0dbe615ac2b0e0470d.jpg", + "image_caption": [ + "(c) Face restoration MOS distribution" + ], + "image_footnote": [], + "bbox": [ + 629, + 104, + 903, + 174 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3c50297297cdc84bc438869a8f6c868263c2dfd6249d458e8a5464371dc5356b.jpg", + "image_caption": [ + "(d) Hunyuan-DiT", + "Figure 3. Mean Opinion Score (MOS) distribution histograms and kernel density curves. (a)-(c) MOS distributions for three subsets. (d)-(h) Model-wise MOS distributions. MOS Distributions of other methods are provided in the supplementary material." + ], + "image_footnote": [], + "bbox": [ + 96, + 186, + 251, + 229 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5c5a432204711046296b8eb590eb28e6cfa3397c63fa6bd83fcdec4b96b1c1c7.jpg", + "image_caption": [ + "(e) SD3" + ], + "image_footnote": [], + "bbox": [ + 251, + 186, + 413, + 229 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/79d7ab75f175a10401291e5733ca09c3a9a45d4049bcae52a7f8d3669e2ca469.jpg", + "image_caption": [ + "(f) InstantID" + ], + "image_footnote": [], + "bbox": [ + 416, + 186, + 576, + 229 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/96e20e2700d895b48c3bae974748669a32f3565913b33186b0a6b646c5450d6c.jpg", + "image_caption": [ + "(g) VQFR (real world)" + ], + "image_footnote": [], + "bbox": [ + 576, + 186, + 736, + 229 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5549685e6ba630acb1c517d5728a32ae0ca6ceee01efa7ff46dfa9400665b798.jpg", + "image_caption": [ + "(h) StableSR (synthetic)" + ], + "image_footnote": [], + "bbox": [ + 736, + 186, + 898, + 229 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "generated image resembles a natural, real-life photograph, with particular emphasis on realistic skin texture, facial details, and wrinkles. It is used in face generation task.", + "bbox": [ + 89, + 282, + 482, + 325 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ID Fidelity. This dimension evaluates how well the identity of the reference image is preserved in the customized or restored image in face customization and restoration tasks.", + "bbox": [ + 89, + 327, + 482, + 372 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Correspondence. This dimension evaluates text-to-image alignment in face generation and customization tasks.", + "bbox": [ + 89, + 373, + 482, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Subjective Experiment", + "text_level": 1, + "bbox": [ + 89, + 414, + 302, + 430 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Main Assessment. To ensure a comprehensive, fair, and reliable subjective assessment, 180 participants with normal or corrected-to-normal eyesight are recruited for the experiment. They first take expert-led tutorials before the main assessment. In the main assessment, participants are asked to rate the image by dragging a sliding window from 0 to 5 with a two-digit decimal from different dimensions. All images are displayed on 27-inch 4K Dell monitors, randomly presented, at their original resolution under standard lighting conditions. The experiment follows the ITU-R BT.500-14 guidelines [78] for subjective evaluations.", + "bbox": [ + 89, + 436, + 483, + 602 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Subjective Score Processing. To eliminate the inter-annotator variability, outlier detection based on Kurtosis is conducted with a rejection rate of $3\\%$ [78]. As a result, each image is rated by at least 15 subjects. These valid ratings are converted and scaled into Z-scores ranging from [0, 100] with the following formulas:", + "bbox": [ + 89, + 604, + 483, + 695 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz _ {i j} = \\frac {r _ {i j} - \\mu_ {i}}{\\sigma_ {i}}, \\quad z _ {i j} ^ {\\prime} = \\frac {1 0 0 (z _ {i j} + 3)}{6}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 705, + 482, + 738 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {i} = \\frac {1}{M _ {i}} \\sum_ {j = 1} ^ {M _ {i}} r _ {i j}, \\sigma_ {i} = \\sqrt {\\frac {1}{M _ {i} - 1} \\sum_ {j = 1} ^ {M _ {i}} (r _ {i j} - \\mu_ {i}) ^ {2}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 765, + 482, + 813 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $r_{ij}$ represents the original ratings provided by the $i$ -th evaluator for the $j$ -th image, $M_i$ denotes the total number of images that were assessed by the $i$ -th participant in this group. Finally, the mean opinion score (MOS) of the image $j$ ( $MOS_j$ ) is computed by averaging the rescaled", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "z-scores as follows:", + "bbox": [ + 513, + 282, + 645, + 296 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM O S _ {j} = \\frac {1}{N _ {j}} \\sum_ {i = 1} ^ {N _ {j}} z _ {i j} ^ {\\prime} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 306, + 905, + 348 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N_{j}$ denotes the number of valid subjects for image $j$ , and $z_{ij}^{\\prime}$ denotes rescaled z-scores.", + "bbox": [ + 511, + 358, + 903, + 391 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. F-Bench: Benchmarking Current Generative Face Models", + "text_level": 1, + "bbox": [ + 511, + 402, + 903, + 435 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we establish a human preference benchmark F-Bench for today's state-of-the-art face generation, customization, and restoration methods, from quality, authenticity, ID fidelity, and correspondence dimensions. This benchmark analyzes the strengths and weaknesses, providing new insights and potential directions for future work.", + "bbox": [ + 511, + 445, + 905, + 536 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Perspective Analysis", + "text_level": 1, + "bbox": [ + 511, + 545, + 705, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig.3 demonstrate the MOS distribution across eight dimensions in the three subsets, highlighting distinct variations between dimensions. (1) In the face generation task, the Correspondence score is the highest among the three dimensions, indicating that existing generative models perform well in aligning with prompts. The distribution of Authenticity is skewed towards lower scores, suggesting that current face generation models still fall short in generating photorealistic face images that satisfy human preferences. (2) In face customization, models perform the worst in the Quality dimension overall, demonstrating the difficulty of injecting identity information without compromising image quality. Identity Fidelity shows a bimodal distribution, reflecting the dichotomous nature of human judgment regarding whether the generated face accurately represents the same identity. (3) In face restoration, models perform better in Quality than in Identity Fidelity.", + "bbox": [ + 511, + 566, + 906, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Model-wise Comparison", + "text_level": 1, + "bbox": [ + 511, + 833, + 736, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Face Generation. Fig.4a presents the average MOS scores across three dimensions: Quality, Authenticity, and Correspondence. Correspondence scores show a steady increase", + "bbox": [ + 511, + 854, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10985", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/303eb0332611bac5a2f83daa16a9a9abd49abc0fce780e81bf45e786228ec912.jpg", + "image_caption": [ + "Figure 4. Average MOS score comparison across all models and dimensions. (a) Face generation. (b) Face customization. (c) Face restoration. The models are arranged in a clockwise order by release date." + ], + "image_footnote": [], + "bbox": [ + 161, + 88, + 387, + 195 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f59aec7a75dddcb25645866c238883c65a864408f00f39ecea66cc85f857374e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 88, + 653, + 196 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/23104c2b45121fa81dc3d3fde8bf89c52a2708806106f4cf5126a2ed3de561b8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 663, + 88, + 893, + 196 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/58c75820987283522713a3ba1b67a33e0791394c6a0c2832d19922b94416984a.jpg", + "image_caption": [ + "Figure 5. Violin plots of quality, authenticity and correspondence scores in FaceQ-Gen subset across nine prompt categories." + ], + "image_footnote": [], + "bbox": [ + 91, + 233, + 903, + 338 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "over time, indicating consistent improvement in text comprehension. Quality scores generally rise with the release year, although Hunyuan [58] and Dreamlike [2] exhibit deviations. Hunyuan [58] oversmoothes the facial details and performs worse compared to its contemporaries, while Dreamlike [2] performs the opposite. Authenticity scores exhibit the greatest variation among models, suggesting that current face generation models place insufficient emphasis on authenticity. Except for Realistic Vision [8], SD3 [9], Flux-dev [3], and Dreamlike [2], most models lag in authenticity relative to quality, highlighting a common phenomenon of high-quality but less realistic outputs in generative models. Fig.3d,e give some examples of MOS distributions of different methods, revealing a reasonable fine-grained performance difference.", + "bbox": [ + 88, + 361, + 482, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Face Customization. Fig.4b is the performance radar chart for face customization models. IP-Adapter-FaceID-Plus[5] achieves the best performance across all dimensions, followed by PhotoMaker [57]. This highlights their great robustness in identity encoders. ELITE [98] scores the lowest in all dimensions. IP-Adapter-FaceID [4] produces high-quality images but lacks accuracy in Identity Fidelity. This may be because the base model [71] introduces many AI artifacts, which affect both Identity Fidelity and Quality. InstantID [91] and FastComposer [101] score well in Identity Fidelity but underperform significantly in Quality and Correspondence, indicating strong identity preservation but limited text-driven editing capabilities. Fig.3h illustrate the model-wise MOS distributions, revealing similar trends to those observed in the radar chart.", + "bbox": [ + 89, + 587, + 482, + 813 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Face Restoration. Fig.4c displays the average performance comparison of nine face restoration models. DR2 [97] exhibits the weakest performance in both Quality and Identity fidelity. This may be due to the degradation removal process of DR2 [97], which results in the loss of fine de", + "bbox": [ + 89, + 814, + 482, + 888 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tails and excessive smoothing. In contrast, StableSR [89] and CodeFormer [113] achieve the best overall performance among the nine face restoration models, successfully preserving both intricate details and accurate identity representations. When comparing the performance on real-world and synthetic data, most methods struggle to achieve satisfactory results on complex real-world degradations, except for SAPRNet [15] and DiffBIR [62]. StableSR [89] and GFPGAN [94] show strong robustness, performing consistently well on both real-world and synthetic data.", + "bbox": [ + 511, + 361, + 906, + 512 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Class-wise Comparison", + "text_level": 1, + "bbox": [ + 511, + 518, + 728, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the face generation task, we analyze the Mean Opinion Score (MOS) distributions of Quality, Authenticity, and Correspondence across nine prompt categories, as shown in Fig.5. The violin plots highlight distinct performance patterns across these categories. The Action, Expression, and Facial Attributes categories exhibit lower MOS values in all three dimensions, with notable variability. In the Action category, limited visibility of the face due to body movements affects the generation of fine facial details, resulting in reduced scores, especially in Quality and Correspondence. The Expression category poses challenges in maintaining spatial relationships and physical constraints, leading to distortion and artifacts in expressions with exaggerated facial movements, such as laughing or crying, thereby impacting both Quality and Authenticity. For Facial Attributes, generating images with precise attributes (e.g., specific eye color, hair style) is challenging, particularly when multiple attributes are specified, which lowers the scores. In contrast, the General and Background categories achieve the highest MOS scores across all dimensions, as simpler or less specific prompts are easier for the model to interpret accurately. The Clothes and Accessories categories also perform well, indicating that the model handles clothes and accessory prompts effectively. Overall, prompts with specific or", + "bbox": [ + 511, + 537, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10986", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ea49b49f9d03c0009d83adb59ba0355c21572a7476d5ebc10ade03953940f09e.jpg", + "image_caption": [ + "Figure 6. The overall framework of F-Eval. F-Eval can evaluate quality, authenticity, correspondence, and identity fidelity in a one-for-all framework. It can process both single and paired images, along with prompts, to produce quality scores. It consists of three encoders, including a vision encoder, a face encoder, and a text tokenizer to process multi-modal inputs. These features are projected into the same space by trained projectors. A pre-trained large language model is utilized to fuse the features while fine-tuned with four LoRA experts. Specific LoRA will be activated by dimension ID, which is classified by a trainable router." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 341, + 305 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/eabc3a39c23181b6f185bfa475841063d2fa90b82c807841ba30c66e7609deb8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 88, + 903, + 305 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "complex attribute requirements tend to result in lower MOS scores, while general or simple prompts yield consistently high scores across dimensions. This can be attributed to the absence of high-quality fine-grained text-face pair datasets during the training of generative models.", + "bbox": [ + 89, + 380, + 483, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. F-Eval Method", + "text_level": 1, + "bbox": [ + 89, + 467, + 243, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we introduce our one-for-all generated face quality assessment method, F-Eval, towards predicting quality scores from multiple dimensions for three tasks using one model.", + "bbox": [ + 89, + 493, + 483, + 554 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Overall Architecture", + "text_level": 1, + "bbox": [ + 89, + 561, + 287, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The overall architecture of F-Eval is depicted in Fig. 6. As shown on the left, F-Eval can assess four dimensions: quality, authenticity, text-image correspondence, and identity fidelity, with multimodal inputs, i.e.text prompts, image and optional reference image, to output quality-related answers. On the right, the input image(s) can originate from all three tasks. F-Eval employs a vision encoder and a face encoder to extract visual features, with projectors projecting both into a unified language space. User prompts are encoded into textual embeddings through a tokenizer. These multimodal features are concatenated and fed into a pre-trained large language model for the final output.", + "bbox": [ + 89, + 583, + 483, + 765 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Model Design", + "text_level": 1, + "bbox": [ + 89, + 772, + 233, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Vision Encoding. For input image(s), F-Eval adopt a vision encoder to extract global features and a face encoder captures face features related to facial structure and identity. Specifically, the vision encoder, Qwen-VL-2.5-ViT [11], is tuned on FaceQ to adjust different dimensions. Face encoder is Arcface [22], which is frozen. Two two-layer multi-layer perception projectors then map the features.", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Feature Fusion via the LLM. For input prompts, it is tokenized into text tokens and then concatenated with visual tokens. Here, Qwen-VL2.5-7B [11] is employed as the pretrained LLM backbone to combine visual tokens and text tokens to perform multi-modal learning.. Finally, the output features of the LLM are decoded with a text decoder and then projected to the prompt-corresponded quality space.", + "bbox": [ + 511, + 380, + 906, + 488 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Multi-Expert LoRA Tuning", + "text_level": 1, + "bbox": [ + 511, + 496, + 764, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Instruction Tuning. Previous IQA methods are often designed for specific dimensions [106, 112], addressing only one aspect of the problem or requiring retraining of model weights for different dimensions. Leveraging the generalization ability of LLMs, instruction tuning allows for the integration of different dimensions within a unified framework. Therefore, we construct diverse instruction data for the four facial quality evaluation dimensions in the FaceQ dataset and fine-tune F-Eval by adjusting the visual encoder, projector, and LLM parameters to adapt it to the specific task. As a result, F-Eval is capable of evaluating facial quality using unified weights.", + "bbox": [ + 511, + 518, + 906, + 699 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Mixture of LoRA Experts (MoLE). Tuning a holistic quality model for different dimensions using a single LoRA expert introduces inherent interference between dimensions. Hence, F-Eval integrates four LoRA experts for four dimensions when tuning vision encoder and LLM, constituting MoLE modules. A router is trained to classify input prompts into the four dimensions and output the corresponding dimension ID $k$ . Specifically, for a frozen linear layer $h = Wx$ with input $x \\in \\mathbb{R}^{d_i}$ and weight matrix $W \\in \\mathbb{R}^{d_o \\times d_i}$ , a LoRA expert is trained to learn a low-rank decomposed update:", + "bbox": [ + 511, + 699, + 908, + 867 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nh = W x + \\Delta W x = W x + \\frac {\\alpha}{r} B A x, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 877, + 906, + 905 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10987", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2404cf1bdcbb2950b5e72e8a7e17a7c42e44040aba12ba64fcc397b03b4e2d71.jpg", + "table_caption": [ + "Table 2. Performance comparison on FaceQ-Gen and FaceQ-Cus subsets. $\\spadesuit$ , $\\clubsuit$ , $\\diamond$ , and $\\heartsuit$ denote traditional IQA models, face IQA models, classical deep learning-based IQA models, and AIGC IQA models respectively. * indicates fine-tuned results." + ], + "table_footnote": [], + "table_body": "
TaskFace GenerationFace Customization
DimensionQualityAuthenticityCorrespondenceQualityID FidelityCorrespondence
ModelSRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑
◆NIQE [67]0.17260.11420.25530.08690.05710.13460.07520.05070.10980.13650.08990.14360.14900.09910.13100.11860.07840.1245
◆ILNIQE [111]0.19810.13190.31590.06750.04450.15050.08890.05910.16650.20320.13270.24790.21920.14520.27780.18780.12510.2275
◆HOSA [102]0.25370.17080.29720.11030.07280.14460.1220.08140.14950.10190.06880.05970.07530.05060.05520.10360.06960.0748
◆BPRI-PSS [64]0.23620.15950.24130.17000.11420.17720.16730.11210.16470.02380.01660.03140.14250.09720.17580.11000.07290.1067
◆FISBLIM [27]0.09010.05950.04330.00280.00030.01750.10000.06650.01960.16140.10240.16900.02560.01730.01260.08520.05590.0854
◆ArcFace[22]N/AN/AN/AN/AN/AN/AN/AN/AN/A0.30990.20940.36050.50620.34390.55720.18370.12430.2244
◆SER-FIQ[83]0.15520.10340.15610.04700.03140.05750.11300.0750.1110-0.1241-0.080-0.1241-0.0109-0.00780.0042-0.1680-0.1121-0.1911
◆DSL-FIQA[18]0.56060.39270.59450.31560.20960.39600.37060.25140.43550.36150.24500.36510.24290.16520.25190.10400.07030.0974
◆IFQA[41]0.23130.15490.32610.15970.10620.20970.15970.10770.22600.24010.16140.22200.15030.09730.17390.37310.25460.3941
◆CNNIAQ*[42]0.42190.28950.35100.30370.20540.29730.28240.19310.25950.62440.43760.66200.65110.44410.63810.37650.25450.3264
◇VGG16*[79]0.58460.41410.60060.50100.35080.51810.43490.29940.47230.72170.53240.78570.67350.47170.66360.56350.39320.5589
◇VGG19*[79]0.57280.40430.56440.46810.32480.44790.41710.28730.47540.79520.60310.80960.75440.54170.75550.65850.47420.6465
◇ResNet18*[29]0.61500.44170.66600.46000.31600.46040.48290.33600.46280.79660.60130.80060.74730.53350.76560.63080.44880.6191
◇ResNet34*[29]0.60920.43540.65370.55640.39490.54010.44100.30650.52070.81410.62110.81510.79690.58780.81570.57080.40000.5810
◇HyperIAQ*[81]0.63330.45450.65910.61060.44060.60420.42350.29120.42400.85240.66310.82050.84190.64630.85880.66280.47570.6454
◇TRES*[25]0.77660.58670.80470.64580.47030.65790.57080.40380.62700.88830.70830.89510.83300.64090.86580.73930.54970.7400
◆MANIAQ*[106]0.78710.59510.81500.73080.54360.77940.62780.44430.70660.89700.71850.89520.85900.66540.87960.78490.59410.7822
◇PromptIAQ*[14]0.78230.59070.82450.72320.53270.75580.65140.46980.70650.90710.73150.88210.86930.67920.88760.83480.63420.8281
♥CLIPScore[32]0.09670.06420.13150.06400.04080.08170.23960.15910.27330.31260.21060.31850.18160.12180.18110.70080.50980.6986
♥BLIPScore[53]0.13060.08720.19290.09350.06130.15690.20510.13640.29630.30730.20630.31360.12130.07960.12660.67070.47970.6123
♥ImageReward [103]0.38490.26310.43580.26060.17360.31810.51550.35420.58710.32320.21660.31960.19050.12700.18390.77140.56730.7740
♥MINTIAQ[88]0.70020.51600.80470.62170.44500.68260.54910.38910.65840.53780.37750.55740.40880.28160.40410.28860.19320.2959
♥MINTIAQ*[88]0.83120.64740.89740.81770.63060.85110.79080.59910.86670.89040.70920.88160.85240.66970.87550.83910.64130.8318
F-Eval (Ours)0.84860.66700.90850.83120.65850.85780.84710.66370.91060.94620.79610.94610.91880.76400.93220.94600.79590.9457
", + "bbox": [ + 93, + 121, + 906, + 434 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $A\\in \\mathbb{R}^{r\\times d_i}$ and $B\\in \\mathbb{R}^{d_o\\times r}$ are the low-rank matrices, $r\\ll \\min (d_o,d_i)$ is the rank, and $\\alpha$ is scaling factor. For prompt of certain dimension, the corresponding expert is activated based on the dimension ID $k$ , which can be formulated as:", + "bbox": [ + 89, + 438, + 482, + 513 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nE _ {k} (x) = \\frac {\\alpha_ {k}}{r _ {k}} B _ {k} A _ {k} x, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 510, + 482, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $E_{k}(\\cdot)$ is the k-th LoRA expert.", + "bbox": [ + 89, + 542, + 339, + 558 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 89, + 570, + 223, + 588 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.1. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 594, + 307, + 611 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The LLM used in F-Eval is Qwen-2.5VL [11], with an input token channel dimension of 3584. The vision encoder, face encoder, and LLM are all frozen during training. The ranks of LoRA experts are 16,32,16,16, respectively. The LoRA alphas are set equal to LoRA ranks. LoRA dropout is set as 0.05. The model is trained on 2 NVIDIA GeForce RTX 3090 GPUs with flash-attn [21] enabled. We use AdamW Optimizer to train the network for 10 epochs with a batch size of 4 and learning rate $1e^{-4}$ .", + "bbox": [ + 89, + 616, + 483, + 753 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2. Comparison with existing QA Models", + "text_level": 1, + "bbox": [ + 89, + 757, + 415, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct comprehensive experiments on the proposed FaceQ database to evaluate the effectiveness of both current quality assessment methods and our proposed methods, as shown in Tab.2, Tab.3, and Tab.4. The evaluation metrics are Spearman rank-order correlation coefficient (SRCC), Kendall rank order correlation coefficient (KRCC), and Pearson linear correlation coefficient (PLCC), which assess the correlation between predicted scores by the QA", + "bbox": [ + 89, + 780, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "methods and the MOS scores provided by humans. Higher metrics indicated better alignment with human preferences. The comparison QA models falls into four categories: 1) Traditional IQA Models. NR-IQA models include NIQE [67], ILNIQE [111], HOSA [102], BPRI-PSS [64], FISBLIM [27]. FR-IQA models include LPIPS [112], SSIM [96] and PSNR [34]. 2) Classical Deep learning-based IQA Models. This category includes CNNIQA [42], VGG16 [79], VGG19 [79], ResNet18 [29], ResNet34 [29], Hyper-IQA [81], TReS [25], MANIQA [106], PromptIQA[14]. All models are retrained on FaceQ with an 80:20 training/testing split. 3) AIGC IQA Models. This category includes pre-trained vision-language models, including CLIP-Score [32], BLIPScore [53], zero-shot metric ImageReward [103], and the AIGC-specific QA model MINTIQA [88]. 4) Face IQA Models. This category includes SER-FIQ [83], DSL-FIQA [18], ArcFace [22], IFQA[41].", + "bbox": [ + 511, + 439, + 906, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.3. Performance Analysis", + "text_level": 1, + "bbox": [ + 511, + 709, + 718, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Traditional IQA Models. Traditional IQA methods, including both NR-IQA and FR-IQA models, generally perform poorly across all tasks and dimensions, with low SRCC, KRCC, and PLCC scores.", + "bbox": [ + 511, + 734, + 905, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Classical Deep Learning-based IQA Models. After being trained on our dataset, deep learning-based IQA models, such as MANIQA [106] and HyperIQA [81], significantly outperform traditional models in terms of quality and authenticity. However, these methods face challenges in text understanding, leading to unsatisfactory correlations.", + "bbox": [ + 511, + 795, + 905, + 885 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "AIGC IQA Models. CLIPScore [32] BLIPScore [53], and", + "bbox": [ + 511, + 885, + 905, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10988", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/65b474d478b7053cea0a3ed0d9efa13383f5095e8a8cf4b7d99eff41fd4c767b.jpg", + "table_caption": [ + "Table 3. Performance comparison on FaceQ-Res synthetic subset. $\\spadesuit$ , $\\heartsuit$ , $\\clubsuit$ , and $\\diamond$ denote traditional NR-IQA models, FR-IQA models, face IQA models, and classical deep learning-based IQA models respectively. * indicates fine-tuned results." + ], + "table_footnote": [], + "table_body": "
Dimension ModelQualityID Fidelity
SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑
▲BMPRI[65]0.21610.14280.18620.11130.07390.1069
▲NIQE[67]0.17980.12130.16420.21930.14800.2118
▲ILNIQE[111]0.25050.17270.31800.26960.18570.3321
▲HOSA[102]0.37580.25220.37480.34180.34550.2301
▲BPII-LSSn [64]0.13980.09570.09990.12440.08370.0973
▲BPII [64]0.16370.10650.14800.01950.01210.0281
▲FISBLIM [27]0.15890.10890.20380.10540.07090.1715
\\( \\diamondsuit \\)LPIPS[112]0.39090.26580.46430.43720.30180.4870
\\( \\diamondsuit \\)SSIM[96]0.10170.06750.04880.09330.06010.0504
\\( \\diamondsuit \\)PSNR [34]0.01980.01130.05360.09860.06620.1212
\\( \\diamondsuit \\)ArcFace[22]0.34660.23270.37740.57280.40300.5762
\\( \\diamondsuit \\)SER-FIQ [25]0.38130.25840.37460.39510.26410.3707
\\( \\diamondsuit \\)DSL-FIQA [18]0.63870.45080.65280.48820.33600.5379
\\( \\diamondsuit \\)IFQA [41]0.22730.15270.33480.32140.21850.3874
\\( \\diamondsuit \\)CNNIA* [42]0.46610.31970.38360.46780.32160.4288
\\( \\diamondsuit \\)ResNet18* [29]0.72540.52950.69820.66620.48020.6527
\\( \\diamondsuit \\)ResNet34* [29]0.66930.49160.65500.59050.41700.6084
\\( \\diamondsuit \\)VGG16* [79]0.74870.55130.71010.61770.44130.6381
\\( \\diamondsuit \\)VGG19* [79]0.69940.51260.67220.60380.42780.5999
\\( \\diamondsuit \\)HyperIA* [81]0.82690.63230.82280.76150.55850.7471
\\( \\diamondsuit \\)MANIA* [106]0.82870.64030.87780.77040.56600.7619
\\( \\diamondsuit \\)TReS* [25]0.86560.68300.86220.79510.58790.7829
\\( \\diamondsuit \\)PromptIA* [14]0.85190.66610.87330.82200.62750.8254
F-Eval (Ours)0.86920.68550.90090.85070.67310.8726
", + "bbox": [ + 91, + 146, + 480, + 395 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/23070b0e5dae106f1be8f9cb016e40e7ca63b3bdba0a1cb43a109df1d31d079d.jpg", + "table_caption": [ + "Table 4. Performance comparison on FaceQ-Res real-world subset. $\\spadesuit$ , $\\clubsuit$ , and $\\diamond$ denote traditional IQA models, face IQA models, and classical deep learning-based IQA models respectively. * indicates fine-tuned results." + ], + "table_footnote": [], + "table_body": "
Dimension ModelQualityID Fidelity
SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑
▲BMPRII[65]0.15820.10930.27990.01490.01080.1386
▲NIQE[67]0.17830.11860.19620.12450.08100.1701
▲ILNIQE[111]0.22300.16100.22910.19640.13620.2340
▲HOSA[102]0.39270.42770.26540.33280.22340.3813
▲BPRI-LSSn [64]0.27150.18400.30220.20590.13800.2475
▲BPRI [64]0.19200.13130.22000.08460.05700.1136
▲FISBLIM [27]0.25120.17050.27250.23410.15810.2699
▲ArcFace[22]0.23710.15650.41830.49560.34700.6330
▲SER-FIQ [25]-0.0482-0.02990.10230.04210.02990.1584
▲DSL-FIQA [18]0.65090.46410.64940.48400.33620.5118
▲IFQA [41]0.02510.02100.12210.08930.06030.2086
◇CNNIQA* [42]0.36500.25060.47050.32520.21970.3112
◇VGG16* [79]0.49960.34390.42690.41170.28590.4622
◇VGG19* [79]0.58710.41810.51330.30540.20440.3594
◇ResNet18* [29]0.59770.42210.57220.55540.39090.6161
◇ResNet34* [29]0.62590.44750.61680.50360.35390.6136
◇HyperiQA* [81]0.79260.59650.79260.77830.58450.8163
◇MANIQA* [106]0.83560.64820.84670.75970.56550.8081
◇TReS* [25]0.82870.63910.85010.72870.54410.8009
◇PromptIQA* [14]0.81150.61860.86450.73980.55260.8297
F-Eval(Ours)0.84480.65770.87050.79570.60570.8366
", + "bbox": [ + 91, + 455, + 480, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ImageReward [103] exhibit weak correlations with human preference, likely due to their focus on high-level semantics rather than low-level details crucial for image quality assessment. MINTIQA [88] achieves best performance among current AIGC IQA methods, but its performance is still unsatisfactory, particularly on the customization task. Fine-tuning on our dataset leads to significant improvements, highlighting the great potential of FaceQ.", + "bbox": [ + 88, + 688, + 482, + 810 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Face IQA Models. ArcFace similarity [22] demonstrates its ability to extract identity-related features. However, it lacks perceptual sensitivity for image quality, leading to poor performance in quality assessment. DSL-FIQA [18], as the latest method, achieves the best performance in predicting quality scores among face IQA methods, but it is", + "bbox": [ + 88, + 810, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "still far from satisfactory. SER-FIQ fails to align with human perception across most dimensions and tasks. Experiments confirm that these natural face evaluation algorithms struggle to assess the complex characteristics of AIGFs, highlighting the need for an effective AIGF evaluator.", + "bbox": [ + 511, + 90, + 903, + 166 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "F-Eval (Ours). Our method achieves optimal performance across all tasks and dimensions, with the most significant gain in the customization task, as existing methods rarely focus human preferences in face customization. Among the four dimensions, the improvements in correspondence and ID fidelity stand out, thanks to dual-stream vision encoding and the powerful LLM backbone for text understanding and modality fusion. The MoLE technique allows the model to learn distinct features for each dimension.", + "bbox": [ + 511, + 167, + 906, + 301 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.4. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 306, + 666, + 323 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conducted the ablation study to validate the utility of the core components of F-Eval in Tab. 5.", + "bbox": [ + 511, + 329, + 903, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of Vision Encoding. 1) The effectiveness of the face encoder is demonstrated by comparing the second and last rows. Without the face encoder, a significant performance drop is observed, indicating that the extracted features help the LLM interpret face images. 2) Removing the projector leads to a dramatic performance drop.", + "bbox": [ + 511, + 359, + 905, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of MoLE. We conducted three sets of LoRA adaptation experiments: fine-tuning a single LoRA for all dimensions and tasks, fine-tuning three LoRA experts for three tasks, and fine-tuning four LoRA experts for four dimensions. The results show that fine-tuning multiple dimensions with a single LoRA leads to performance deterioration, which further confirms the independence of the characteristics across different dimensions.", + "bbox": [ + 511, + 450, + 905, + 570 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9ebdecafbdb12c7efb3beb5e68bbe7d6c72c20c390f185fb5b9be00ebc00d9e6.jpg", + "table_caption": [ + "Table 5. Ablation study of the proposed F-Eval method. Here only present SRCC due to space limitations." + ], + "table_footnote": [], + "table_body": "
Feature & StrategyFace GenerationFace CustomizationFace Restoration synFace Restoration rw
VisionProjectorFaceSingle LoRATask LoRADim. LoRAQualityAuthen.Corres.QualityID.Corres.QualityID.Corres.QualityID.QualityID.
✓✓✓0.8080.7500.8460.9240.8430.9270.7770.7760.7760.7070.748
✓✓✓0.8150.7520.8350.9240.8550.9210.8120.7830.7160.603
✓✓✓✓✓0.7930.7760.8190.9410.8850.9420.8380.8230.8020.671
✓✓✓✓0.8230.7800.8380.9420.8730.9420.8240.8260.7540.638
✓✓✓✓0.8490.8310.8470.9460.9190.9460.8690.8510.8450.796
", + "bbox": [ + 516, + 603, + 903, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 664, + 633, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we conduct a pioneering generative face image quality assessment study. Specifically, we introduce FaceQ, the first generated face quality assessment database comprising 12,255 face images and 491,130 human preference ratings. Then, we establish F-Bench, a benchmark for evaluating the capabilities of existing face generation, customization, and restoration models, while also reveals the limitations of current Image Quality Assessment (IQA) and Face Quality Assessment (FQA) methods. Furthermore, we propose an evaluation model F-Eval to accurately predict human preferences on generated face images, which demonstrates state-of-the-art performance. We believe this work will contribute to advancing generative models and help overcome current limitations in AIGC community.", + "bbox": [ + 511, + 689, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10989", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported by the National Natural Science Foundation of China under Grants 62271308, 62401365, 62225112, 62132006, U24A20220, STCSM under Grants 24ZR1432000, 24511106902, 24511106900, 22511105700, 22DZ2229005, 111 plan under Grant Number BP0719010, China Postdoctoral Science Foundation under Grants BX20250411, 2025M773473, and State Key Laboratory of UHD Video and Audio Production and Presentation.", + "bbox": [ + 89, + 114, + 483, + 250 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 263, + 187, + 279 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Deep Floyd. https://github.com/deep-floyd/IF, Accessed: 2024-10-03. 2, 3", + "[2] Dreamlike V2. https://huggingface.co/dreamlike-art/dreamlike-photoreal-2.0, Accessed: 2024-10-03.5", + "[3] Flux-dev. https://huggingface.co/black-forest-labs/FLUX.1-dev, Accessed: 2024-10-03.1,3,5", + "[4] IP-Adapter-FaceID-SDXL. https://huggingface.co/h94/IP-Adapter, Accessed: 2024-10-03. 3, 5", + "[5] IP-Adapter-FaceID-Plus. https://huggingface.co/h94/IP-Adapter-FaceID, Accessed: 2024-10-03.3,5", + "[6] Kcolors. https://huggingface.co/Kwai-Kcolors/Kcolors, Accessed: 2024-10-03. 3", + "[7] ProtoVision V6.6. https://huggingface.co/stablediffusionapi/protovision-xl-v6.6, Accessed: 2024-10-03. 3", + "[8] Realistic Vision V5.1. https://huggingface.co/SG161222/Realistic_Vision_V5.1_noVAE, Accessed: 2024-10-03. 3, 5", + "[9] SD3. https://huggingface.co/stabilityai/stable-diffusion-3-medium, Accessed: 2024-10-03. 3,5", + "[10] GPT-4o. https://chatgpt.com/, Accessed: 2024-10-10.3", + "[11] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. 2, 6, 7", + "[12] Fadi Boutros, Meiling Fang, Marcel Klemt, Biying Fu, and Naser Damer. Cr-fiqa: face image quality assessment by learning sample relative classifiability. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5836-5845, 2023. 2", + "[13] Nicolas Chahine, Stefania Calarasanu, Davide Garcia-Civiero, Theo Cayla, Sira Ferradans, and Jean Ponce. An image quality assessment dataset for portraits. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9968-9978, 2023. 2", + "[14] Chen et al. Promptia: Boosting the performance and generalization for no-reference image quality assessment via prompts. In ECCV, 2024. 7, 8" + ], + "bbox": [ + 101, + 289, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Chaofeng Chen, Dihong Gong, Hao Wang, Zhifeng Li, and Kwan-Yee K Wong. Learning spatial attention for face super-resolution. IEEE Transactions on Image Processing (TIP), 30:1219-1231, 2020. 2, 3, 5", + "[16] Chaofeng Chen, Xiaoming Li, Lingbo Yang, Xianhui Lin, Lei Zhang, and Kwan-Yee K Wong. Progressive semantic-aware style transformation for blind face restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11896-11905, 2021. 2", + "[17] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arxiv:2310.00426, 2023. 2, 3", + "[18] Wei-Ting Chen, Gurunandan Krishnan, Qiang Gao, SyYen Kuo, Sizhou Ma, and Jian Wang. Dsl-fqa: Assessing facial image quality via dual-set degradation learning and landmark-guided transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2931-2941, 2024. 2, 7, 8", + "[19] Yu Chen, Ying Tai, Xiaoming Liu, Chunhua Shen, and Jian Yang. Fsrnet: End-to-end learning face super-resolution with facial priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2492-2501, 2018. 2", + "[20] Siying Cui, Jiankang Deng, Jia Guo, Xiang An, Yongle Zhao, Xinyu Wei, and Ziyong Feng. Idapter: Learning mixed features for tuning-free personalization of text-to-image models. arXiv preprint arXiv:2403.13535, 2024. 2", + "[21] Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning. In Proceedings of the International Conference on Learning Representations (ICLR). 7", + "[22] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4690-4699, 2019. 6, 7, 8", + "[23] Huiyu Duan, Qiang Hu, Jiarui Wang, Liu Yang, Zitong Xu, Lu Liu, Xiongkuo Min, Chunlei Cai, Tianxiao Ye, Xiaoyun Zhang, et al. Fine-grained user generated content video quality assessment. In m, 2025. 2, 3", + "[24] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2", + "[25] S. Alireza Golestaneh, Saba Dadsetan, and Kris M. Kitani. No-reference image quality assessment via transformers, relative ranking, and self-consistency. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1220-1230, 2022. 7, 8", + "[26] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Pro" + ], + "bbox": [ + 522, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10990", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ceedings of the Advances in Neural Information Processing Systems, 2014. 1, 2", + "[27] Ke Gu, Guangtao Zhai, Min Liu, Xiaokang Yang, Wenjun Zhang, Xianghui Sun, Wanhong Chen, and Ying Zuo. Fisblim: A five-step blind metric for quality assessment of multiply distorted images. In Proceedings of the SiPS, pages 241-246, 2013. 7, 8", + "[28] Yuchao Gu, Xintao Wang, Liangbin Xie, Chao Dong, Gen Li, Ying Shan, and Ming-Ming Cheng. Vqfr: Blind face restoration with vector-quantized dictionary and parallel decoder. arXiv preprint arXiv:2205.06803, 2022. 2, 3", + "[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 7, 8", + "[30] Xuanhua He, Quande Liu, Shengju Qian, Xin Wang, Tao Hu, Ke Cao, Keyu Yan, Man Zhou, and Jie Zhang. Id-animator: Zero-shot identity-preserving human video generation. arXiv preprint arXiv:2404.15275, 2024. 2", + "[31] Javier Hernandez-Ortega, Javier Galbally, Julian Fierrez, Rudolf Haraksim, and Laurent Beslay. Faceqnet: Quality assessment for face recognition based on deep learning. In 2019 International Conference on Biometrics (ICB), pages 1–8. IEEE, 2019. 2", + "[32] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 7", + "[33] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 33: 6840-6851, 2020. 1", + "[34] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssim. In Proceedings of the IEEE International Conference on Pattern Recognition (ICPR), pages 2366-2369. IEEE, 2010. 7, 8", + "[35] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. 2", + "[36] Qiang Hu, Qihan He, Houqiang Zhong, Guo Lu, Xiaoyun Zhang, Guangtao Zhai, and Yanfeng Wang. Varfvv: View-adaptive real-time interactive free-view video streaming with edge computing. IEEE Journal on Selected Areas in Communications, 2025. 2", + "[37] Qiang Hu, Zihan Zheng, Houqiang Zhong, Sihua Fu, Li Song, Xiaoyun Zhang, Guangtao Zhai, and Yanfeng Wang. 4dgc: Rate-aware 4d gaussian compression for efficient streamable free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 875-885, 2025. 2", + "[38] Gary B. Huang, Manu Ramesh, Tamara Berg, and Erik Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. Technical Report 07-49, University of Massachusetts, Amherst, 2007. 3" + ], + "bbox": [ + 99, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 36:78723-78747, 2023. 2", + "[40] Haoyun Jiang, Fei Huang, Qiang Hu, Minmin Sun, Shuai Xiao, Yong Li, Junyang Lin, Jiangchao Yao, et al. Catekv: On sequential consistency for long-context llm inference acceleration. In Proceedings of International Conference on Machine Learning (ICML), 2025. 3", + "[41] Byungho Jo, Donghyeon Cho, In Kyu Park, and Sungeun Hong. Ifqa: Interpretable face quality assessment. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 3444-3453, 2023. 7, 8", + "[42] Le Kang, Peng Ye, Yi Li, and David Doermann. Convolutional neural networks for no-reference image quality assessment. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014. 7, 8", + "[43] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 4401-4410, 2019. 2", + "[44] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6007-6017, 2023. 1", + "[45] Ira Kemelmacher-Shlizerman, Steven M Seitz, Daniel Miller, and Evan Brossard. The megaface benchmark: 1 million faces for recognition at scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4873-4882, 2016. 3", + "[46] Deokyun Kim, Minseon Kim, Gihyun Kwon, and Dae-Shik Kim. Progressive face super-resolution via attention to facial landmark. arXiv preprint arXiv:1908.08239, 2019. 2", + "[47] Diederik P Kingma, Max Welling, et al. Auto-encoding variational bayes. 2013. 1", + "[48] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Matiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. Proceedings of Advances in neural information processing systems (Neurips), 36:36652-36663, 2023. 1, 3", + "[49] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1931-1941, 2023. 2", + "[50] Li et al. Agiqa-3k: An open database for ai-generated image quality assessment. arXiv preprint arXiv:2306.04717, 2023. 2", + "[51] Chunyi Li, Zicheng Zhang, Haoning Wu, Wei Sun, Xiongkuo Min, Xiaohong Liu, Guangtao Zhai, and Weisi Lin. Agiaq-3k: An open database for ai-generated image quality assessment. arXiv preprint arXiv: 2306.04717, 2023. 1, 2, 3" + ], + "bbox": [ + 522, + 92, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10991", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Linmiao Xu, and Suhail Doshi. Playground v2.5: Three insights towards enhancing aesthetic quality in text-to-image generation, 2024. 3", + "[53] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proceedings of International Conference on Machine learning (ICML), pages 12888-12900, 2022. 7", + "[54] Xiaoming Li, Ming Liu, Yuting Ye, Wangmeng Zuo, Liang Lin, and Ruigang Yang. Learning warped guidance for blind face restoration. In The European Conference on Computer Vision (ECCV), 2018. 2", + "[55] Xiaoming Li, Wenyu Li, Dongwei Ren, Hongzhi Zhang, Meng Wang, and Wangmeng Zuo. Enhanced blind face restoration with multi-exemplar images and adaptive spatial feature fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2706-2715, 2020. 2", + "[56] Xiaoming Li, Shiguang Zhang, Shangchen Zhou, Lei Zhang, and Wangmeng Zuo. Learning dual memory dictionaries for blind face restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2022. 3", + "[57] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, Ming-Ming Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. arXiv preprint arXiv:2312.04461, 2023. 1, 2, 3, 5", + "[58] Zhimin Li, Jianwei Zhang, Qin Lin, Jiangfeng Xiong, Yanxin Long, Xinchi Deng, Yingfang Zhang, Xingchao Liu, Minbin Huang, Zedong Xiao, et al. Hunyuan-dit: A powerful multi-resolution diffusion transformer with fine-grained chinese understanding. arXiv preprint arXiv:2405.08748, 2024. 2, 3, 5", + "[59] Youwei Liang, Junfeng He, Gang Li, Peizhao Li, Arseniy Klimovskiy, Nicholas Carolan, Jiao Sun, Jordi Pont-Tuset, Sarah Young, Feng Yang, et al. Rich human feedback for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19401-19411, 2024. 1, 2", + "[60] Zhang Lijun, Shao Xiaohu, Yang Fei, Deng Pingling, Zhou Xiangdong, and Shi Yu. Multi-branch face quality assessment for face recognition. In 2019 IEEE 19th International Conference on Communication Technology (ICCT), pages 1659–1664. IEEE, 2019. 2", + "[61] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of European conference on computer vision (ECCV), pages 740-755. Springer, 2014. 3", + "[62] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Ben Fei, Bo Dai, Wanli Ouyang, Yu Qiao, and Chao Dong. Diff-bir: Towards blind image restoration with generative diffusion prior. arXiv preprint arXiv:2308.15070, 2023. 1, 2, 3, 5", + "[63] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Proceedings of" + ], + "bbox": [ + 99, + 92, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the Advances in Neural Information Processing Systems (NeurIPS), pages 34892-34916, 2023. 2", + "[64] Xiongkuo Min, Ke Gu, Guangtao Zhai, Jing Liu, Xiaokang Yang, and Chang Wen Chen. Blind quality assessment based on pseudo-reference image. IEEE Transactions on Multimedia (TMM), 20(8):2049-2062, 2018. 7, 8", + "[65] Xiongkuo Min, Guangtao Zhai, Ke Gu, Yutao Liu, and Xiaokang Yang. Blind image quality estimation via distortion aggravation. IEEE Transactions on Broadcasting, 64(2): 508-517, 2018. 8", + "[66] Xiongkuo Min, Huiyu Duan, Wei Sun, Yucheng Zhu, and Guangtao Zhai. Perceptual video quality assessment: A survey. Science China Information Sciences, 67(11): 211301, 2024. 2", + "[67] Anish Mittal, Rajiv Soundararajan, and Alan C. Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal Processing Letters, 20(3):209-212, 2013. 7, 8", + "[68] Stylianos Moschoglou, Athanasios Papaioannou, Christos Sagonas, Jiankang Deng, Irene Kotsia, and Stefanos Zafeiriou. Agedb: the first manually collected, in-the-wild age database. In proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 51-59, 2017. 3", + "[69] Fu-Zhao Ou, Xingyu Chen, Ruixin Zhang, Yuge Huang, Shaoxin Li, Jilin Li, Yong Li, Liujuan Cao, and Yuan-Gen Wang. Sdd-fiqa: Unsupervised face image quality assessment with similarity distribution distance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7670–7679, 2021. 2", + "[70] Pablo Pernias, Dominic Rampas, Mats L Richter, Christopher J Pal, and Marc Aubreville. Würstchen: An efficient architecture for large-scale text-to-image diffusion models. arXiv preprint arXiv:2306.00637, 2023. 3", + "[71] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 3, 5", + "[72] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation, 2021. 2", + "[73] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2", + "[74] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 2, 3", + "[75] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22500-22510, 2023. 2", + "[76] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed" + ], + "bbox": [ + 524, + 93, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "10992", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding, 2022. 2", + "[77] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 35:36479-36494, 2022. 2", + "[78] B Series. Methodology for the subjective assessment of the quality of television pictures. Recommendation ITU-R BT, 500(13), 2012. 4", + "[79] Karen Simonyan and Andrew Zisserman. Very Deep Convolutional Networks for Large-Scale Image Recognition. arXiv e-prints, art. arXiv:1409.1556, 2014. 7, 8", + "[80] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv:2010.02502, 2020. 1", + "[81] Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3664-3673, 2020. 7, 8", + "[82] Shaolin Su, Hanhe Lin, Vlad Hosu, Oliver Wiedemann, Jinqiu Sun, Yu Zhu, Hantao Liu, Yanning Zhang, and Dietmar Saupe. Going the extra mile in face image quality assessment: A novel database and model. IEEE Transactions on Multimedia, 2023. 2, 3", + "[83] Philipp Terhorst, Jan Niklas Kolf, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. Ser-fiq: Unsupervised estimation of face image quality based on stochastic embedding robustness. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5651-5660, 2020. 7", + "[84] Yuan Tian, Kaiyuan Ji, Rongzhao Zhang, Yankai Jiang, Chunyi Li, Xiaosong Wang, and Guangtao Zhai. Towards all-in-one medical image re-identification. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 30774–30786, 2025. 2", + "[85] Yuri Viazovetskyi, Vladimir Ivashkin, and Evgeny Kashin. Stylegan2 distillation for feed-forward image manipulation. Proceedings of the European conference on computer vision (ECCV), pages 170-186, 2020. 2", + "[86] Jiarui Wang, Huiyu Duan, Jing Liu, Shi Chen, Xiongkuo Min, and Guangtao Zhai. Aigciqa2023: A large-scale image quality assessment database for ai generated images: from the perspectives of quality, authenticity and correspondence. 2023. 1, 2, 3", + "[87] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. arXiv preprint arXiv:2305.07015, 2023. 1, 2, 3", + "[88] Jiarui Wang, Huiyu Duan, Guangtao Zhai, and Xiongkuo Min. Understanding and evaluating human preferences for ai generated images with instruction tuning. arXiv preprint arXiv:2405.07346, 2024. 7, 8" + ], + "bbox": [ + 99, + 90, + 483, + 898 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[89] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin C.K. Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. 2024. 5", + "[90] Liangbin Xie. Xintao Wang, Honglun Zhang, Chao Dong, and Ying Shan. Vfhq: A high-quality dataset and benchmark for video face super-resolution, 2022. 3", + "[91] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 2, 3, 5", + "[92] Tao Wang, Kaihao Zhang, Xuanxi Chen, Wenhan Luo, Jiankang Deng, Tong Lu, Xiaochun Cao, Wei Liu, Hongdong Li, and Stefanos Zafeiriou. A survey of deep face restoration: Denoise, super-resolution, deblur, artifact removal. arXiv preprint arXiv:2211.02831, 2022. 2", + "[93] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In International Conference on Computer Vision Workshops (ICCVW). 3", + "[94] Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. Towards real-world blind face restoration with generative facial prior. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 3, 5", + "[95] Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. Towards real-world blind face restoration with generative facial prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3", + "[96] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing (TIP), 13(4):600-612, 2004. 7, 8", + "[97] Zhixin Wang, Ziying Zhang, Xiaoyun Zhang, Huangjie Zheng, Mingyuan Zhou, Ya Zhang, and Yanfeng Wang. Dr2: Diffusion-based robust degradation remover for blind face restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1704-1713, 2023. 2, 3, 5", + "[98] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15943-15953, 2023. 2, 3, 5", + "[99] Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xinglong Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. Diffir: Efficient diffusion model for image restoration. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13095-13105, 2023. 1", + "[100] Weihao Xia, Yujiu Yang, Jing-Hao Xue, and Baoyuan Wu. Tedigan: Text-guided diverse face image generation and manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2256-2265, 2021. 1", + "[101] Guangxuan Xiao, Tianwei Yin, William T Freeman, Frédo Durand, and Song Han. Fastcomposer: Tuning-free multisubject image generation with localized attention. Inter-" + ], + "bbox": [ + 514, + 92, + 903, + 900 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "10993", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "national Journal of Computer Vision (IJCV), 133(3):1175-1194, 2025. 2, 3, 5", + "[102] Jingtao Xu, Peng Ye, Qiaohong Li, Haiqing Du, Yong Liu, and David S. Doermann. Blind image quality assessment based on high order statistics aggregation. IEEE Transactions on Image Processing (TIP), 25:4444-4457, 2016. 7, 8", + "[103] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 36, 2024. 2, 7, 8", + "[104] Liu Yang, Huiyu Duan, Long Teng, Yucheng Zhu, Xiaohong Liu, Menghan Hu, Xiongkuo Min, Guangtao Zhai, and Patrick Le Callet. Aigcoiaq2024: Perceptual quality assessment of ai generated omnidirectional images. In 2024 IEEE International Conference on Image Processing (ICIP), pages 1239-1245. IEEE, 2024. 3", + "[105] Shuo Yang, Ping Luo, Chen-Change Loy, and Xiaou Tang. Wider face: A face detection benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5525–5533, 2016. 3", + "[106] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. MANIQA: Multi-dimension Attention Network for No-Reference Image Quality Assessment. arXiv e-prints, art. arXiv:2204.08958, 2022. 6, 7, 8", + "[107] Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. Gan prior embedded network for blind face restoration in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 672-681, 2021. 2, 3", + "[108] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 1, 2", + "[109] Zongsheng Yue and Chen Change Loy. Difface: Blind face restoration with diffused error contraction. arXiv preprint arXiv:2212.06512, 2022. 2, 3", + "[110] Guangtao Zhai and Xiongkuo Min. Perceptual image quality assessment: a survey. Science China Information Sciences, 63:1-52, 2020. 2", + "[111] Lin Zhang, Lei Zhang, and Alan C. Bovik. A feature-enriched completely blind image quality evaluator. IEEE Transactions on Image Processing (TIP), 24(8):2579-2591, 2015. 7, 8", + "[112] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 6, 7, 8", + "[113] Shangchen Zhou, Kelvin Chan, Chongyi Li, and Chen Change Loy. Towards robust blind face restoration with codebook lookup transformer. Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 35:30599-30611, 2022. 2, 3, 5" + ], + "bbox": [ + 91, + 90, + 483, + 896 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "10994", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2025/F-Bench_ Rethinking Human Preference Evaluation Metrics for Benchmarking Face Generation, Customization, and Restoration/86d5740f-dff5-47ab-aa67-761b97593a44_model.json b/2025/F-Bench_ Rethinking Human Preference Evaluation Metrics for Benchmarking Face Generation, Customization, and Restoration/86d5740f-dff5-47ab-aa67-761b97593a44_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5301437178a2ac1d45da1dde14911c86dd96bcf4 --- /dev/null +++ b/2025/F-Bench_ Rethinking Human Preference Evaluation Metrics for Benchmarking Face Generation, Customization, and Restoration/86d5740f-dff5-47ab-aa67-761b97593a44_model.json @@ -0,0 +1,3196 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.1, + 0.13, + 0.9, + 0.175 + ], + "angle": 0, + "content": "F-Bench: Rethinking Human Preference Evaluation Metrics for Benchmarking Face Generation, Customization, and Restoration" + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.203, + 0.753, + 0.221 + ], + "angle": 0, + "content": "Lu Liu\\(^{1,*}\\), Huiyu Duan\\(^{1,*}\\), Qiang Hu\\(^{1,\\dagger}\\), Liu Yang\\(^{1}\\), Chunlei Cai\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.222, + 0.739, + 0.239 + ], + "angle": 0, + "content": "Tianxiao Ye\\(^{2}\\), Huayu Liu\\(^{1}\\), Xiaoyun Zhang\\(^{1}\\), Guangtao Zhai\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.187, + 0.24, + 0.813, + 0.258 + ], + "angle": 0, + "content": "1Shanghai Jiao Tong University, Shanghai, China 2Bilibili Inc., Shanghai, China" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.267, + 0.331, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.268, + 0.666, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.666, + 0.268, + 0.904, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.486, + 0.907, + 0.539 + ], + "angle": 0, + "content": "Figure 1. Data construction pipeline of \\( F \\)-bench and content overview of FaceQ database. From left to right, (a) Diverse input source data for face generation, customization, and restoration. (b) \\( F \\)-Bench benchmarks 29 face generative models, including 14 face generation models, 6 face customization models, and 9 face restoration models. (c) Multi-dimensional subjective preference evaluation in \\( F \\)-Bench. (d) 3D scatter plots of MOS distributions across FaceQ three subsets." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.55, + 0.327, + 0.566 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.584, + 0.486, + 0.871 + ], + "angle": 0, + "content": "Recent artificial intelligence (AI) generative models have demonstrated remarkable capabilities in image production, and have been widely applied to face image generation, customization, and restoration. However, many AI-generated faces (AIGFs) still suffer from issues such as unique distortions, unrealistic details, and unexpected identity shifts, underscoring the need for a comprehensive quality evaluation method for AIGFs. To this end, we introduce FaceQ, the first comprehensive AI-generated Face image database with fine-grained Quality annotations aligned with human preferences, which consists of 12K images and 491K ratings across multiple dimensions. Using the FaceQ database, we establish F-Bench, a benchmark for comparing and evaluating face generation, customization, and restoration models, highlighting strengths and weaknesses across various prompts and evaluation dimensions. Additionally, we assess the performance of existing image quality assessment (IQA) methods on FaceQ, and further propose a large multimodal model (LMM) based Face quality Evaluator (F" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.551, + 0.909, + 0.627 + ], + "angle": 0, + "content": "Eval) to accurately assess the multi-dimensional quality of generated faces in a one-for-all manner. Extensive experimental results demonstrate the state-of-the-art performance of our F-Eval. The project page is: https://media-sjtu.github.io/F-Bench/." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.661, + 0.645, + 0.677 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.687, + 0.907, + 0.854 + ], + "angle": 0, + "content": "Generative models, including VAE [47], GAN [26], and diffusion models [33, 74, 80], have made remarkable progress in recent years, and have been extended to the fields of text-to-image generation [3], text-driven image editing [44], image restoration [99], etc, leading to high-quality artificial intelligence generated images (AIGIs). As an important aspect of AIGIs, many studies have also investigated face generation [100], face customization [57, 108], and face restoration [62, 87], producing creative, high-quality face images. Despite these advances, many AI-generated faces (AIGFs) still suffer from perceptual deficiencies." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Many natural face image quality assessment databases have been established in the literature [48, 51, 59, 86]. However, natural images are mainly degraded by distor" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.887, + 0.361, + 0.901 + ], + "angle": 0, + "content": "* Equal contribution. † Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10982" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.09, + 0.908, + 0.117 + ], + "angle": 0, + "content": "Table 1. Comparison between FaceQ and existing AIGC quality assessment databases (top) as well as natural face quality assessment databases (bottom). FaceQ is the first large-scale AI-generated, customized and restored face quality assessment database." + }, + { + "type": "table", + "bbox": [ + 0.09, + 0.117, + 0.905, + 0.197 + ], + "angle": 0, + "content": "
DatabaseDomainSourceImagesScoresDimensions
AGIQA-3K (TCSVT2024) [51]GeneralAI-generated2,982MOSPerception, Alignment
AIGCIQA (CVPR2024) [86]GeneralAI-generated2,400MOSQuality, Authenticity, Correspondence
RichHF-18K (CVPR2024) [59]GeneralAI-generated17,760MOSPlausibility, Aesthetics, Text-image Alignment, Overall
PIQ23 (CVPR2023) [13]FaceReal5,116PairOverall, Details, Exposure
FaceQFaceAI-generated, customized and restored12,255MOSQuality, Authenticity, ID Fidelity, Correspondence
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.2, + 0.482, + 0.412 + ], + "angle": 0, + "content": "tions such as noise, blur, compression [23, 66, 110], etc., while AIGIs generally suffer from unique distortions including unrealistic structures and text-image inconsistencies [59, 86], etc. Recently, many studies have investigated human preferences for AIGIs [39, 50, 59, 86, 103]. However, these general AIGI quality assessment works may not be well aligned with human preferences on AIGFs, since the perceptual quality of AIGFs can be significantly affected by small deviations in facial attributes,(e.g., eye size, skin texture, identity shifts), while the number of AIGFs in these AIGI databases is limited. Moreover, current AIGI quality assessment works mainly focus on the image generation, while human preferences for face customization have rarely been considered, where identity fidelity is significant." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.414, + 0.485, + 0.718 + ], + "angle": 0, + "content": "The above limitation highlights the necessity of establishing a comprehensive quality evaluator for AIGFs. To this end, we propose FaceQ, the first AI-generated Face image Quality assessment database with three popular tasks. Specifically, FaceQ contains \\(12\\mathrm{K}+\\) images obtained from 29 generative models across three tasks: 1) face generation, 2) face customization, and 3) face restoration, as shown in Fig.1. Based on the generated images, we collect \\(491\\mathrm{K}+\\) quality ratings from multiple dimensions including quality, authenticity, identity (ID) fidelity, and text-image correspondence, respectively, finally forming \\(32\\mathrm{K}+\\) mean opinion scores (MOSs). Using FaceQ, we establish a human preference benchmark, i.e., F-Bench, for face generation, customization, and restoration models, and thoroughly analyze the generation performance across different evaluation perspectives, models, and categories. In addition, we also compare the correlation of existing image quality assessment (IQA), face quality assessment (FQA), AIGI quality assessment, and preference evaluation metrics with human preferences." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Furthermore, we present F-Eval, the first generated Face image quality Evaluation method designed to perform comprehensive multi-dimensional quality evaluation for generated, customized, and restorted faces in a one-for-all manner. Specifically, F-Eval leverages large multimodal models (LMMs) [11] and adopts instruction tuning [63] for preference-related vision-language adaption. We also apply low-rank adaptation (LoRA) [35] techniques for more efficient feature refining. Extensive experimental results demonstrate that F-Eval outperforms existing quality assessment methods in terms of multiple dimensions relevant to human preference for all three tasks. The main contribu" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.2, + 0.808, + 0.216 + ], + "angle": 0, + "content": "tions of our work are highlighted as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.222, + 0.905, + 0.295 + ], + "angle": 0, + "content": "- We construct FaceQ, the first generated face quality assessment database consisting \\(491\\mathrm{K}+\\) human preference ratings from quality, authenticity, ID fidelity, correspondence dimensions, on \\(12\\mathrm{K}+\\) generative face images across three tasks." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.297, + 0.905, + 0.371 + ], + "angle": 0, + "content": "- We establish F-Bench, a comprehensive human preference benchmark for 29 face generation, customization, and restoration methods. It also extensively evaluates the alignment of 26 existing quality assessment methods with human preferences." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.372, + 0.905, + 0.432 + ], + "angle": 0, + "content": "- We present F-Eval, a LMM-based quality assessment model to predict multi-dimensional human preference on generated faces. Extensive experiments demonstrate the state-of-art QA performance of the proposed F-Eval." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.222, + 0.905, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.451, + 0.656, + 0.467 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.907, + 0.825 + ], + "angle": 0, + "content": "Face Generation & Customization & Restoration. (1) For face generation models, earlier works such as GANs [26, 43, 85] have demonstrated significant influences. Recently, diffusion models [73, 74, 77] have achieved rapid developments in text-to-image generation. These models can be categorized into pixel-space diffusion models [1, 72, 73, 76], and latent space diffusion models [17, 58, 71, 74]. (2) For face customization, this is a recently defined task that aims to create images of a specific person with the input of an identity reference image and prompts [98]. It can be divided into two categories including conventional test-time optimization [24, 49, 75] and tuning-free customization methods utilizing pre-trained diffusion models [20, 30, 57, 91, 101, 108]. (3) For face restoration, which aims to restore high-quality face images from their low-quality counterparts. Some traditional restoration methods utilize geometric priors [16, 19, 46] or reference priors [54, 55] while in recent years, exploring generative priors such as GAN priors [15, 28, 94, 107, 113] and diffusion priors [36, 37, 62, 84, 87, 97, 109] has become a prevalent trend. Despite these improvements, the generated, customized, and restored images may suffer from various quality issues, highlighting the necessity of evaluators." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Face Image Quality Assessment Database. Existing face image quality assessment (FIQA) can be divided into Biometric FIQA [12, 31, 60, 69] and Generic FIQA [13, 18, 82, 92]. BFIQA methods are most developed to evaluate the biometric utility for face recognition. GFIQA meth" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10983" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.396, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.4, + 0.089, + 0.695, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.089, + 0.906, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.258, + 0.907, + 0.3 + ], + "angle": 0, + "content": "Figure 2. Rating comparisons of eight dimensions. Each column presents a pair of intuitive examples of each dimension, with red indicating the better rating and blue indicating the worse one. From left to right, the subsets are face generation, face customization, and face restoration subsets. The last row displays the corresponding prompts, reference image-prompt pairs, and the GT-LQ image pairs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.304, + 0.482, + 0.349 + ], + "angle": 0, + "content": "ods [82] focus on perceptual degradation in face restoration task. Both of these databases consist of natural face images, which has distinctly different nature with generated faces." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.35, + 0.483, + 0.456 + ], + "angle": 0, + "content": "AIGC Quality Assessment Database. In recent years, plenty of researchers have raised interest in assessing AI-generated images(AIGIs) quality assessment [23, 40, 48, 51, 86, 104] by building various databases as shown in Tab 1. All of these existing database contain various categories of general AIGI images, leaving a notable gap in datasets dedicated to AI-generated face images." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.468, + 0.365, + 0.485 + ], + "angle": 0, + "content": "3. FaceQ Database Construction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.494, + 0.484, + 0.584 + ], + "angle": 0, + "content": "In this section, we introduce the proposed generative face quality assessment database FaceQ. The database includes 12,255 images and 491,130 human opinion ratings. Our FaceQ demonstrates the merits of comprehensive generated visual content and multi-dimensional labeling, which can significantly facilitate the advancement of AIGC research." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.593, + 0.247, + 0.607 + ], + "angle": 0, + "content": "3.1. Data Collection" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.615, + 0.34, + 0.63 + ], + "angle": 0, + "content": "3.1.1. FaceQ-Gen Subset Collection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.634, + 0.483, + 0.709 + ], + "angle": 0, + "content": "Prompt Sources. The prompts in FaceQ-Gen subset are sourced from MS-COCO [61] [57] and GPT-4o [10]. These face-centric prompts can be categorized into nine types: General, Accessories, Clothes, Action, View, Expression, Background, Style, and Facial Attributes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.71, + 0.483, + 0.831 + ], + "angle": 0, + "content": "Models Collection. We collect 14 representative face generation models, including Deep Floyd [1], Stable Diffusion V1.5 [74], Stable Diffusion V2.1 [74], SD-XL [71], Realistic Vision [8], Stable Cascade [70], ProtoVision V6.6 [7], SD3 [9], Playground V2.5 [52], Kolors [6], PixArt-alpha [17], Hunyuan [58], and Flux-dev [3]. Negative prompts and safe sensors are enabled. In total, FaceQ-Gen includes 4,032 generated face images." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.837, + 0.338, + 0.852 + ], + "angle": 0, + "content": "3.1.2. FaceQ-Cus Subset Collection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Identity and Prompt Sources. Face customization relies on a reference identity image and a guiding prompt. The reference images in FaceQ-Cus are partially self-collected," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.304, + 0.907, + 0.335 + ], + "angle": 0, + "content": "with the rest sourced from [57]. The prompts are selected and filtered from the FaceQ-Gen prompt list." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.336, + 0.907, + 0.427 + ], + "angle": 0, + "content": "Models Collection. We collect six representative face customization methods, including ELITE [98], FastComposer [101], IP-Adapter-FaceID [4], InstantID [91], IP-Adapter-Face-ID-Plus [5], and PhotoMaker V2 [57]. Here, we use single reference image as input. In total, FaceQ-Cus includes 4,200 generated face images." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.437, + 0.758, + 0.451 + ], + "angle": 0, + "content": "3.1.3. FaceQ-Res Subset Collection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.457, + 0.906, + 0.592 + ], + "angle": 0, + "content": "Low Quality Images Source. (1) Synthetic: We construct two degraded pipelines following previous works [15, 93, 95, 97] to simulate real-world degradations on both indomain data from VFHQ [90], CelebRef [56] and out-of-domain self-collected high-quality Asian face images (2) Real-world: Real-world low-quality images are sourced from various public datasets, including AgeAB [68], CelebChild [28], Wider [105], LFW [38] and MegaFace [45], ensuring the degradation diversity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.594, + 0.906, + 0.73 + ], + "angle": 0, + "content": "Models Collection and Face Image Restoration. We evaluate both diffusion-prior-based methods such as DiffBIR [62], DifFace [109], StableSR [87], and GAN or Transformer-based methods GFPGAN [94], GPEN [107], CodeFormer [113], VQFR [28] and Face-SPARNet [15]. Each method is benchmarked on both the \\(F\\)-Res synthetic dataset and the real-world dataset. In total, FaceQ-Res include 4023 restored images, with 2007 images in the real-world case and 2016 images in the synthetic case." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.742, + 0.838, + 0.759 + ], + "angle": 0, + "content": "3.2. Human Preference Dimension Design" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.841 + ], + "angle": 0, + "content": "The evaluation dimensions of generated faces need specific design to adequately reflect human preference, given the complexity of perceptual factors involved. To the end, we decompose \"quality\" into four core dimensions: Quality, Authenticity, ID Fidelity, and Correspondence." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.905, + 0.885 + ], + "angle": 0, + "content": "Quality. This dimension evaluates the overall perceptual quality of the image, considering factors such as color accuracy, blur, noise, and artifacts. It is used in all tasks." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Authenticity. This dimension evaluates how closely the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10984" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.105, + 0.359, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.168, + 0.174, + 0.325, + 0.184 + ], + "angle": 0, + "content": "(a) Face generation MOS distribution" + }, + { + "type": "image", + "bbox": [ + 0.35, + 0.105, + 0.63, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.416, + 0.175, + 0.588, + 0.184 + ], + "angle": 0, + "content": "(b) Face customization MOS distribution" + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.105, + 0.905, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.693, + 0.175, + 0.852, + 0.184 + ], + "angle": 0, + "content": "(c) Face restoration MOS distribution" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.187, + 0.252, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.139, + 0.23, + 0.212, + 0.24 + ], + "angle": 0, + "content": "(d) Hunyuan-DiT" + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.187, + 0.415, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.319, + 0.23, + 0.352, + 0.24 + ], + "angle": 0, + "content": "(e) SD3" + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.187, + 0.577, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.479, + 0.23, + 0.532, + 0.239 + ], + "angle": 0, + "content": "(f) InstantID" + }, + { + "type": "image", + "bbox": [ + 0.577, + 0.187, + 0.738, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.61, + 0.23, + 0.703, + 0.24 + ], + "angle": 0, + "content": "(g) VQFR (real world)" + }, + { + "type": "image", + "bbox": [ + 0.738, + 0.187, + 0.9, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.23, + 0.866, + 0.24 + ], + "angle": 0, + "content": "(h) StableSR (synthetic)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.242, + 0.905, + 0.269 + ], + "angle": 0, + "content": "Figure 3. Mean Opinion Score (MOS) distribution histograms and kernel density curves. (a)-(c) MOS distributions for three subsets. (d)-(h) Model-wise MOS distributions. MOS Distributions of other methods are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.283, + 0.483, + 0.327 + ], + "angle": 0, + "content": "generated image resembles a natural, real-life photograph, with particular emphasis on realistic skin texture, facial details, and wrinkles. It is used in face generation task." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.328, + 0.483, + 0.373 + ], + "angle": 0, + "content": "ID Fidelity. This dimension evaluates how well the identity of the reference image is preserved in the customized or restored image in face customization and restoration tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.374, + 0.483, + 0.404 + ], + "angle": 0, + "content": "Correspondence. This dimension evaluates text-to-image alignment in face generation and customization tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.415, + 0.303, + 0.431 + ], + "angle": 0, + "content": "3.3. Subjective Experiment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.438, + 0.484, + 0.603 + ], + "angle": 0, + "content": "Main Assessment. To ensure a comprehensive, fair, and reliable subjective assessment, 180 participants with normal or corrected-to-normal eyesight are recruited for the experiment. They first take expert-led tutorials before the main assessment. In the main assessment, participants are asked to rate the image by dragging a sliding window from 0 to 5 with a two-digit decimal from different dimensions. All images are displayed on 27-inch 4K Dell monitors, randomly presented, at their original resolution under standard lighting conditions. The experiment follows the ITU-R BT.500-14 guidelines [78] for subjective evaluations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.605, + 0.484, + 0.696 + ], + "angle": 0, + "content": "Subjective Score Processing. To eliminate the inter-annotator variability, outlier detection based on Kurtosis is conducted with a rejection rate of \\(3\\%\\) [78]. As a result, each image is rated by at least 15 subjects. These valid ratings are converted and scaled into Z-scores ranging from [0, 100] with the following formulas:" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.707, + 0.483, + 0.739 + ], + "angle": 0, + "content": "\\[\nz _ {i j} = \\frac {r _ {i j} - \\mu_ {i}}{\\sigma_ {i}}, \\quad z _ {i j} ^ {\\prime} = \\frac {1 0 0 (z _ {i j} + 3)}{6}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.766, + 0.483, + 0.814 + ], + "angle": 0, + "content": "\\[\n\\mu_ {i} = \\frac {1}{M _ {i}} \\sum_ {j = 1} ^ {M _ {i}} r _ {i j}, \\sigma_ {i} = \\sqrt {\\frac {1}{M _ {i} - 1} \\sum_ {j = 1} ^ {M _ {i}} (r _ {i j} - \\mu_ {i}) ^ {2}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Here, \\( r_{ij} \\) represents the original ratings provided by the \\( i \\)-th evaluator for the \\( j \\)-th image, \\( M_i \\) denotes the total number of images that were assessed by the \\( i \\)-th participant in this group. Finally, the mean opinion score (MOS) of the image \\( j \\) (\\( MOS_j \\)) is computed by averaging the rescaled" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.283, + 0.647, + 0.297 + ], + "angle": 0, + "content": "z-scores as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.307, + 0.906, + 0.349 + ], + "angle": 0, + "content": "\\[\nM O S _ {j} = \\frac {1}{N _ {j}} \\sum_ {i = 1} ^ {N _ {j}} z _ {i j} ^ {\\prime} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.359, + 0.905, + 0.392 + ], + "angle": 0, + "content": "where \\(N_{j}\\) denotes the number of valid subjects for image \\(j\\), and \\(z_{ij}^{\\prime}\\) denotes rescaled z-scores." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.403, + 0.905, + 0.436 + ], + "angle": 0, + "content": "4. F-Bench: Benchmarking Current Generative Face Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.446, + 0.906, + 0.537 + ], + "angle": 0, + "content": "In this section, we establish a human preference benchmark F-Bench for today's state-of-the-art face generation, customization, and restoration methods, from quality, authenticity, ID fidelity, and correspondence dimensions. This benchmark analyzes the strengths and weaknesses, providing new insights and potential directions for future work." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.546, + 0.707, + 0.562 + ], + "angle": 0, + "content": "4.1. Perspective Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.907, + 0.825 + ], + "angle": 0, + "content": "Fig.3 demonstrate the MOS distribution across eight dimensions in the three subsets, highlighting distinct variations between dimensions. (1) In the face generation task, the Correspondence score is the highest among the three dimensions, indicating that existing generative models perform well in aligning with prompts. The distribution of Authenticity is skewed towards lower scores, suggesting that current face generation models still fall short in generating photorealistic face images that satisfy human preferences. (2) In face customization, models perform the worst in the Quality dimension overall, demonstrating the difficulty of injecting identity information without compromising image quality. Identity Fidelity shows a bimodal distribution, reflecting the dichotomous nature of human judgment regarding whether the generated face accurately represents the same identity. (3) In face restoration, models perform better in Quality than in Identity Fidelity." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.834, + 0.738, + 0.85 + ], + "angle": 0, + "content": "4.2. Model-wise Comparison" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Face Generation. Fig.4a presents the average MOS scores across three dimensions: Quality, Authenticity, and Correspondence. Correspondence scores show a steady increase" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10985" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.163, + 0.089, + 0.388, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.089, + 0.655, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.089, + 0.895, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.202, + 0.908, + 0.23 + ], + "angle": 0, + "content": "Figure 4. Average MOS score comparison across all models and dimensions. (a) Face generation. (b) Face customization. (c) Face restoration. The models are arranged in a clockwise order by release date." + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.234, + 0.905, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.109, + 0.34, + 0.886, + 0.355 + ], + "angle": 0, + "content": "Figure 5. Violin plots of quality, authenticity and correspondence scores in FaceQ-Gen subset across nine prompt categories." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.362, + 0.483, + 0.587 + ], + "angle": 0, + "content": "over time, indicating consistent improvement in text comprehension. Quality scores generally rise with the release year, although Hunyuan [58] and Dreamlike [2] exhibit deviations. Hunyuan [58] oversmoothes the facial details and performs worse compared to its contemporaries, while Dreamlike [2] performs the opposite. Authenticity scores exhibit the greatest variation among models, suggesting that current face generation models place insufficient emphasis on authenticity. Except for Realistic Vision [8], SD3 [9], Flux-dev [3], and Dreamlike [2], most models lag in authenticity relative to quality, highlighting a common phenomenon of high-quality but less realistic outputs in generative models. Fig.3d,e give some examples of MOS distributions of different methods, revealing a reasonable fine-grained performance difference." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.588, + 0.483, + 0.814 + ], + "angle": 0, + "content": "Face Customization. Fig.4b is the performance radar chart for face customization models. IP-Adapter-FaceID-Plus[5] achieves the best performance across all dimensions, followed by PhotoMaker [57]. This highlights their great robustness in identity encoders. ELITE [98] scores the lowest in all dimensions. IP-Adapter-FaceID [4] produces high-quality images but lacks accuracy in Identity Fidelity. This may be because the base model [71] introduces many AI artifacts, which affect both Identity Fidelity and Quality. InstantID [91] and FastComposer [101] score well in Identity Fidelity but underperform significantly in Quality and Correspondence, indicating strong identity preservation but limited text-driven editing capabilities. Fig.3h illustrate the model-wise MOS distributions, revealing similar trends to those observed in the radar chart." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.815, + 0.483, + 0.89 + ], + "angle": 0, + "content": "Face Restoration. Fig.4c displays the average performance comparison of nine face restoration models. DR2 [97] exhibits the weakest performance in both Quality and Identity fidelity. This may be due to the degradation removal process of DR2 [97], which results in the loss of fine de" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.362, + 0.907, + 0.513 + ], + "angle": 0, + "content": "tails and excessive smoothing. In contrast, StableSR [89] and CodeFormer [113] achieve the best overall performance among the nine face restoration models, successfully preserving both intricate details and accurate identity representations. When comparing the performance on real-world and synthetic data, most methods struggle to achieve satisfactory results on complex real-world degradations, except for SAPRNet [15] and DiffBIR [62]. StableSR [89] and GFPGAN [94] show strong robustness, performing consistently well on both real-world and synthetic data." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.519, + 0.73, + 0.535 + ], + "angle": 0, + "content": "4.3. Class-wise Comparison" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.909, + 0.903 + ], + "angle": 0, + "content": "For the face generation task, we analyze the Mean Opinion Score (MOS) distributions of Quality, Authenticity, and Correspondence across nine prompt categories, as shown in Fig.5. The violin plots highlight distinct performance patterns across these categories. The Action, Expression, and Facial Attributes categories exhibit lower MOS values in all three dimensions, with notable variability. In the Action category, limited visibility of the face due to body movements affects the generation of fine facial details, resulting in reduced scores, especially in Quality and Correspondence. The Expression category poses challenges in maintaining spatial relationships and physical constraints, leading to distortion and artifacts in expressions with exaggerated facial movements, such as laughing or crying, thereby impacting both Quality and Authenticity. For Facial Attributes, generating images with precise attributes (e.g., specific eye color, hair style) is challenging, particularly when multiple attributes are specified, which lowers the scores. In contrast, the General and Background categories achieve the highest MOS scores across all dimensions, as simpler or less specific prompts are easier for the model to interpret accurately. The Clothes and Accessories categories also perform well, indicating that the model handles clothes and accessory prompts effectively. Overall, prompts with specific or" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10986" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.342, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.359, + 0.089, + 0.905, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.307, + 0.907, + 0.378 + ], + "angle": 0, + "content": "Figure 6. The overall framework of F-Eval. F-Eval can evaluate quality, authenticity, correspondence, and identity fidelity in a one-for-all framework. It can process both single and paired images, along with prompts, to produce quality scores. It consists of three encoders, including a vision encoder, a face encoder, and a text tokenizer to process multi-modal inputs. These features are projected into the same space by trained projectors. A pre-trained large language model is utilized to fuse the features while fine-tuned with four LoRA experts. Specific LoRA will be activated by dimension ID, which is classified by a trainable router." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.381, + 0.484, + 0.457 + ], + "angle": 0, + "content": "complex attribute requirements tend to result in lower MOS scores, while general or simple prompts yield consistently high scores across dimensions. This can be attributed to the absence of high-quality fine-grained text-face pair datasets during the training of generative models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.468, + 0.245, + 0.485 + ], + "angle": 0, + "content": "5. F-Eval Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.494, + 0.484, + 0.555 + ], + "angle": 0, + "content": "In this section, we introduce our one-for-all generated face quality assessment method, F-Eval, towards predicting quality scores from multiple dimensions for three tasks using one model." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.562, + 0.288, + 0.577 + ], + "angle": 0, + "content": "5.1. Overall Architecture" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.484, + 0.766 + ], + "angle": 0, + "content": "The overall architecture of F-Eval is depicted in Fig. 6. As shown on the left, F-Eval can assess four dimensions: quality, authenticity, text-image correspondence, and identity fidelity, with multimodal inputs, i.e.text prompts, image and optional reference image, to output quality-related answers. On the right, the input image(s) can originate from all three tasks. F-Eval employs a vision encoder and a face encoder to extract visual features, with projectors projecting both into a unified language space. User prompts are encoded into textual embeddings through a tokenizer. These multimodal features are concatenated and fed into a pre-trained large language model for the final output." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.773, + 0.234, + 0.79 + ], + "angle": 0, + "content": "5.2. Model Design" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Vision Encoding. For input image(s), F-Eval adopt a vision encoder to extract global features and a face encoder captures face features related to facial structure and identity. Specifically, the vision encoder, Qwen-VL-2.5-ViT [11], is tuned on FaceQ to adjust different dimensions. Face encoder is Arcface [22], which is frozen. Two two-layer multi-layer perception projectors then map the features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.381, + 0.907, + 0.489 + ], + "angle": 0, + "content": "Feature Fusion via the LLM. For input prompts, it is tokenized into text tokens and then concatenated with visual tokens. Here, Qwen-VL2.5-7B [11] is employed as the pretrained LLM backbone to combine visual tokens and text tokens to perform multi-modal learning.. Finally, the output features of the LLM are decoded with a text decoder and then projected to the prompt-corresponded quality space." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.497, + 0.766, + 0.514 + ], + "angle": 0, + "content": "5.3. Multi-Expert LoRA Tuning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.519, + 0.907, + 0.7 + ], + "angle": 0, + "content": "Instruction Tuning. Previous IQA methods are often designed for specific dimensions [106, 112], addressing only one aspect of the problem or requiring retraining of model weights for different dimensions. Leveraging the generalization ability of LLMs, instruction tuning allows for the integration of different dimensions within a unified framework. Therefore, we construct diverse instruction data for the four facial quality evaluation dimensions in the FaceQ dataset and fine-tune F-Eval by adjusting the visual encoder, projector, and LLM parameters to adapt it to the specific task. As a result, F-Eval is capable of evaluating facial quality using unified weights." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.7, + 0.909, + 0.868 + ], + "angle": 0, + "content": "Mixture of LoRA Experts (MoLE). Tuning a holistic quality model for different dimensions using a single LoRA expert introduces inherent interference between dimensions. Hence, F-Eval integrates four LoRA experts for four dimensions when tuning vision encoder and LLM, constituting MoLE modules. A router is trained to classify input prompts into the four dimensions and output the corresponding dimension ID \\( k \\). Specifically, for a frozen linear layer \\( h = Wx \\) with input \\( x \\in \\mathbb{R}^{d_i} \\) and weight matrix \\( W \\in \\mathbb{R}^{d_o \\times d_i} \\), a LoRA expert is trained to learn a low-rank decomposed update:" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.878, + 0.907, + 0.906 + ], + "angle": 0, + "content": "\\[\nh = W x + \\Delta W x = W x + \\frac {\\alpha}{r} B A x, \\tag {4}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10987" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 2. Performance comparison on FaceQ-Gen and FaceQ-Cus subsets. \\(\\spadesuit\\), \\(\\clubsuit\\), \\(\\diamond\\), and \\(\\heartsuit\\) denote traditional IQA models, face IQA models, classical deep learning-based IQA models, and AIGC IQA models respectively. * indicates fine-tuned results." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.122, + 0.907, + 0.435 + ], + "angle": 0, + "content": "
TaskFace GenerationFace Customization
DimensionQualityAuthenticityCorrespondenceQualityID FidelityCorrespondence
ModelSRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑
◆NIQE [67]0.17260.11420.25530.08690.05710.13460.07520.05070.10980.13650.08990.14360.14900.09910.13100.11860.07840.1245
◆ILNIQE [111]0.19810.13190.31590.06750.04450.15050.08890.05910.16650.20320.13270.24790.21920.14520.27780.18780.12510.2275
◆HOSA [102]0.25370.17080.29720.11030.07280.14460.1220.08140.14950.10190.06880.05970.07530.05060.05520.10360.06960.0748
◆BPRI-PSS [64]0.23620.15950.24130.17000.11420.17720.16730.11210.16470.02380.01660.03140.14250.09720.17580.11000.07290.1067
◆FISBLIM [27]0.09010.05950.04330.00280.00030.01750.10000.06650.01960.16140.10240.16900.02560.01730.01260.08520.05590.0854
◆ArcFace[22]N/AN/AN/AN/AN/AN/AN/AN/AN/A0.30990.20940.36050.50620.34390.55720.18370.12430.2244
◆SER-FIQ[83]0.15520.10340.15610.04700.03140.05750.11300.0750.1110-0.1241-0.080-0.1241-0.0109-0.00780.0042-0.1680-0.1121-0.1911
◆DSL-FIQA[18]0.56060.39270.59450.31560.20960.39600.37060.25140.43550.36150.24500.36510.24290.16520.25190.10400.07030.0974
◆IFQA[41]0.23130.15490.32610.15970.10620.20970.15970.10770.22600.24010.16140.22200.15030.09730.17390.37310.25460.3941
◆CNNIAQ*[42]0.42190.28950.35100.30370.20540.29730.28240.19310.25950.62440.43760.66200.65110.44410.63810.37650.25450.3264
◇VGG16*[79]0.58460.41410.60060.50100.35080.51810.43490.29940.47230.72170.53240.78570.67350.47170.66360.56350.39320.5589
◇VGG19*[79]0.57280.40430.56440.46810.32480.44790.41710.28730.47540.79520.60310.80960.75440.54170.75550.65850.47420.6465
◇ResNet18*[29]0.61500.44170.66600.46000.31600.46040.48290.33600.46280.79660.60130.80060.74730.53350.76560.63080.44880.6191
◇ResNet34*[29]0.60920.43540.65370.55640.39490.54010.44100.30650.52070.81410.62110.81510.79690.58780.81570.57080.40000.5810
◇HyperIAQ*[81]0.63330.45450.65910.61060.44060.60420.42350.29120.42400.85240.66310.82050.84190.64630.85880.66280.47570.6454
◇TRES*[25]0.77660.58670.80470.64580.47030.65790.57080.40380.62700.88830.70830.89510.83300.64090.86580.73930.54970.7400
◆MANIAQ*[106]0.78710.59510.81500.73080.54360.77940.62780.44430.70660.89700.71850.89520.85900.66540.87960.78490.59410.7822
◇PromptIAQ*[14]0.78230.59070.82450.72320.53270.75580.65140.46980.70650.90710.73150.88210.86930.67920.88760.83480.63420.8281
♥CLIPScore[32]0.09670.06420.13150.06400.04080.08170.23960.15910.27330.31260.21060.31850.18160.12180.18110.70080.50980.6986
♥BLIPScore[53]0.13060.08720.19290.09350.06130.15690.20510.13640.29630.30730.20630.31360.12130.07960.12660.67070.47970.6123
♥ImageReward [103]0.38490.26310.43580.26060.17360.31810.51550.35420.58710.32320.21660.31960.19050.12700.18390.77140.56730.7740
♥MINTIAQ[88]0.70020.51600.80470.62170.44500.68260.54910.38910.65840.53780.37750.55740.40880.28160.40410.28860.19320.2959
♥MINTIAQ*[88]0.83120.64740.89740.81770.63060.85110.79080.59910.86670.89040.70920.88160.85240.66970.87550.83910.64130.8318
F-Eval (Ours)0.84860.66700.90850.83120.65850.85780.84710.66370.91060.94620.79610.94610.91880.76400.93220.94600.79590.9457
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.439, + 0.483, + 0.514 + ], + "angle": 0, + "content": "where \\(A\\in \\mathbb{R}^{r\\times d_i}\\) and \\(B\\in \\mathbb{R}^{d_o\\times r}\\) are the low-rank matrices, \\(r\\ll \\min (d_o,d_i)\\) is the rank, and \\(\\alpha\\) is scaling factor. For prompt of certain dimension, the corresponding expert is activated based on the dimension ID \\(k\\), which can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.214, + 0.511, + 0.483, + 0.539 + ], + "angle": 0, + "content": "\\[\nE _ {k} (x) = \\frac {\\alpha_ {k}}{r _ {k}} B _ {k} A _ {k} x, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.543, + 0.341, + 0.559 + ], + "angle": 0, + "content": "where \\(E_{k}(\\cdot)\\) is the k-th LoRA expert." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.571, + 0.224, + 0.589 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.595, + 0.308, + 0.612 + ], + "angle": 0, + "content": "6.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.617, + 0.484, + 0.754 + ], + "angle": 0, + "content": "The LLM used in F-Eval is Qwen-2.5VL [11], with an input token channel dimension of 3584. The vision encoder, face encoder, and LLM are all frozen during training. The ranks of LoRA experts are 16,32,16,16, respectively. The LoRA alphas are set equal to LoRA ranks. LoRA dropout is set as 0.05. The model is trained on 2 NVIDIA GeForce RTX 3090 GPUs with flash-attn [21] enabled. We use AdamW Optimizer to train the network for 10 epochs with a batch size of 4 and learning rate \\( 1e^{-4} \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.758, + 0.416, + 0.775 + ], + "angle": 0, + "content": "6.2. Comparison with existing QA Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We conduct comprehensive experiments on the proposed FaceQ database to evaluate the effectiveness of both current quality assessment methods and our proposed methods, as shown in Tab.2, Tab.3, and Tab.4. The evaluation metrics are Spearman rank-order correlation coefficient (SRCC), Kendall rank order correlation coefficient (KRCC), and Pearson linear correlation coefficient (PLCC), which assess the correlation between predicted scores by the QA" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.44, + 0.907, + 0.698 + ], + "angle": 0, + "content": "methods and the MOS scores provided by humans. Higher metrics indicated better alignment with human preferences. The comparison QA models falls into four categories: 1) Traditional IQA Models. NR-IQA models include NIQE [67], ILNIQE [111], HOSA [102], BPRI-PSS [64], FISBLIM [27]. FR-IQA models include LPIPS [112], SSIM [96] and PSNR [34]. 2) Classical Deep learning-based IQA Models. This category includes CNNIQA [42], VGG16 [79], VGG19 [79], ResNet18 [29], ResNet34 [29], Hyper-IQA [81], TReS [25], MANIQA [106], PromptIQA[14]. All models are retrained on FaceQ with an 80:20 training/testing split. 3) AIGC IQA Models. This category includes pre-trained vision-language models, including CLIP-Score [32], BLIPScore [53], zero-shot metric ImageReward [103], and the AIGC-specific QA model MINTIQA [88]. 4) Face IQA Models. This category includes SER-FIQ [83], DSL-FIQA [18], ArcFace [22], IFQA[41]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.71, + 0.72, + 0.727 + ], + "angle": 0, + "content": "6.3. Performance Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.795 + ], + "angle": 0, + "content": "Traditional IQA Models. Traditional IQA methods, including both NR-IQA and FR-IQA models, generally perform poorly across all tasks and dimensions, with low SRCC, KRCC, and PLCC scores." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.886 + ], + "angle": 0, + "content": "Classical Deep Learning-based IQA Models. After being trained on our dataset, deep learning-based IQA models, such as MANIQA [106] and HyperIQA [81], significantly outperform traditional models in terms of quality and authenticity. However, these methods face challenges in text understanding, leading to unsatisfactory correlations." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.906, + 0.902 + ], + "angle": 0, + "content": "AIGC IQA Models. CLIPScore [32] BLIPScore [53], and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10988" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.09, + 0.483, + 0.145 + ], + "angle": 0, + "content": "Table 3. Performance comparison on FaceQ-Res synthetic subset. \\(\\spadesuit\\), \\(\\heartsuit\\), \\(\\clubsuit\\), and \\(\\diamond\\) denote traditional NR-IQA models, FR-IQA models, face IQA models, and classical deep learning-based IQA models respectively. * indicates fine-tuned results." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.147, + 0.481, + 0.396 + ], + "angle": 0, + "content": "
Dimension ModelQualityID Fidelity
SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑
▲BMPRI[65]0.21610.14280.18620.11130.07390.1069
▲NIQE[67]0.17980.12130.16420.21930.14800.2118
▲ILNIQE[111]0.25050.17270.31800.26960.18570.3321
▲HOSA[102]0.37580.25220.37480.34180.34550.2301
▲BPII-LSSn [64]0.13980.09570.09990.12440.08370.0973
▲BPII [64]0.16370.10650.14800.01950.01210.0281
▲FISBLIM [27]0.15890.10890.20380.10540.07090.1715
\\( \\diamondsuit \\)LPIPS[112]0.39090.26580.46430.43720.30180.4870
\\( \\diamondsuit \\)SSIM[96]0.10170.06750.04880.09330.06010.0504
\\( \\diamondsuit \\)PSNR [34]0.01980.01130.05360.09860.06620.1212
\\( \\diamondsuit \\)ArcFace[22]0.34660.23270.37740.57280.40300.5762
\\( \\diamondsuit \\)SER-FIQ [25]0.38130.25840.37460.39510.26410.3707
\\( \\diamondsuit \\)DSL-FIQA [18]0.63870.45080.65280.48820.33600.5379
\\( \\diamondsuit \\)IFQA [41]0.22730.15270.33480.32140.21850.3874
\\( \\diamondsuit \\)CNNIA* [42]0.46610.31970.38360.46780.32160.4288
\\( \\diamondsuit \\)ResNet18* [29]0.72540.52950.69820.66620.48020.6527
\\( \\diamondsuit \\)ResNet34* [29]0.66930.49160.65500.59050.41700.6084
\\( \\diamondsuit \\)VGG16* [79]0.74870.55130.71010.61770.44130.6381
\\( \\diamondsuit \\)VGG19* [79]0.69940.51260.67220.60380.42780.5999
\\( \\diamondsuit \\)HyperIA* [81]0.82690.63230.82280.76150.55850.7471
\\( \\diamondsuit \\)MANIA* [106]0.82870.64030.87780.77040.56600.7619
\\( \\diamondsuit \\)TReS* [25]0.86560.68300.86220.79510.58790.7829
\\( \\diamondsuit \\)PromptIA* [14]0.85190.66610.87330.82200.62750.8254
F-Eval (Ours)0.86920.68550.90090.85070.67310.8726
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.4, + 0.483, + 0.455 + ], + "angle": 0, + "content": "Table 4. Performance comparison on FaceQ-Res real-world subset. \\(\\spadesuit\\), \\(\\clubsuit\\), and \\(\\diamond\\) denote traditional IQA models, face IQA models, and classical deep learning-based IQA models respectively. * indicates fine-tuned results." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.457, + 0.481, + 0.678 + ], + "angle": 0, + "content": "
Dimension ModelQualityID Fidelity
SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑
▲BMPRII[65]0.15820.10930.27990.01490.01080.1386
▲NIQE[67]0.17830.11860.19620.12450.08100.1701
▲ILNIQE[111]0.22300.16100.22910.19640.13620.2340
▲HOSA[102]0.39270.42770.26540.33280.22340.3813
▲BPRI-LSSn [64]0.27150.18400.30220.20590.13800.2475
▲BPRI [64]0.19200.13130.22000.08460.05700.1136
▲FISBLIM [27]0.25120.17050.27250.23410.15810.2699
▲ArcFace[22]0.23710.15650.41830.49560.34700.6330
▲SER-FIQ [25]-0.0482-0.02990.10230.04210.02990.1584
▲DSL-FIQA [18]0.65090.46410.64940.48400.33620.5118
▲IFQA [41]0.02510.02100.12210.08930.06030.2086
◇CNNIQA* [42]0.36500.25060.47050.32520.21970.3112
◇VGG16* [79]0.49960.34390.42690.41170.28590.4622
◇VGG19* [79]0.58710.41810.51330.30540.20440.3594
◇ResNet18* [29]0.59770.42210.57220.55540.39090.6161
◇ResNet34* [29]0.62590.44750.61680.50360.35390.6136
◇HyperiQA* [81]0.79260.59650.79260.77830.58450.8163
◇MANIQA* [106]0.83560.64820.84670.75970.56550.8081
◇TReS* [25]0.82870.63910.85010.72870.54410.8009
◇PromptIQA* [14]0.81150.61860.86450.73980.55260.8297
F-Eval(Ours)0.84480.65770.87050.79570.60570.8366
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.689, + 0.483, + 0.811 + ], + "angle": 0, + "content": "ImageReward [103] exhibit weak correlations with human preference, likely due to their focus on high-level semantics rather than low-level details crucial for image quality assessment. MINTIQA [88] achieves best performance among current AIGC IQA methods, but its performance is still unsatisfactory, particularly on the customization task. Fine-tuning on our dataset leads to significant improvements, highlighting the great potential of FaceQ." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Face IQA Models. ArcFace similarity [22] demonstrates its ability to extract identity-related features. However, it lacks perceptual sensitivity for image quality, leading to poor performance in quality assessment. DSL-FIQA [18], as the latest method, achieves the best performance in predicting quality scores among face IQA methods, but it is" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.167 + ], + "angle": 0, + "content": "still far from satisfactory. SER-FIQ fails to align with human perception across most dimensions and tasks. Experiments confirm that these natural face evaluation algorithms struggle to assess the complex characteristics of AIGFs, highlighting the need for an effective AIGF evaluator." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.168, + 0.907, + 0.303 + ], + "angle": 0, + "content": "F-Eval (Ours). Our method achieves optimal performance across all tasks and dimensions, with the most significant gain in the customization task, as existing methods rarely focus human preferences in face customization. Among the four dimensions, the improvements in correspondence and ID fidelity stand out, thanks to dual-stream vision encoding and the powerful LLM backbone for text understanding and modality fusion. The MoLE technique allows the model to learn distinct features for each dimension." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.308, + 0.667, + 0.324 + ], + "angle": 0, + "content": "6.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.33, + 0.905, + 0.36 + ], + "angle": 0, + "content": "We conducted the ablation study to validate the utility of the core components of F-Eval in Tab. 5." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.361, + 0.906, + 0.451 + ], + "angle": 0, + "content": "Effectiveness of Vision Encoding. 1) The effectiveness of the face encoder is demonstrated by comparing the second and last rows. Without the face encoder, a significant performance drop is observed, indicating that the extracted features help the LLM interpret face images. 2) Removing the projector leads to a dramatic performance drop." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.452, + 0.906, + 0.571 + ], + "angle": 0, + "content": "Effectiveness of MoLE. We conducted three sets of LoRA adaptation experiments: fine-tuning a single LoRA for all dimensions and tasks, fine-tuning three LoRA experts for three tasks, and fine-tuning four LoRA experts for four dimensions. The results show that fine-tuning multiple dimensions with a single LoRA leads to performance deterioration, which further confirms the independence of the characteristics across different dimensions." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.574, + 0.905, + 0.603 + ], + "angle": 0, + "content": "Table 5. Ablation study of the proposed F-Eval method. Here only present SRCC due to space limitations." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.651 + ], + "angle": 0, + "content": "
Feature & StrategyFace GenerationFace CustomizationFace Restoration synFace Restoration rw
VisionProjectorFaceSingle LoRATask LoRADim. LoRAQualityAuthen.Corres.QualityID.Corres.QualityID.Corres.QualityID.QualityID.
✓✓✓0.8080.7500.8460.9240.8430.9270.7770.7760.7760.7070.748
✓✓✓0.8150.7520.8350.9240.8550.9210.8120.7830.7160.603
✓✓✓✓✓0.7930.7760.8190.9410.8850.9420.8380.8230.8020.671
✓✓✓✓0.8230.7800.8380.9420.8730.9420.8240.8260.7540.638
✓✓✓✓0.8490.8310.8470.9460.9190.9460.8690.8510.8450.796
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.665, + 0.634, + 0.68 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this work, we conduct a pioneering generative face image quality assessment study. Specifically, we introduce FaceQ, the first generated face quality assessment database comprising 12,255 face images and 491,130 human preference ratings. Then, we establish F-Bench, a benchmark for evaluating the capabilities of existing face generation, customization, and restoration models, while also reveals the limitations of current Image Quality Assessment (IQA) and Face Quality Assessment (FQA) methods. Furthermore, we propose an evaluation model F-Eval to accurately predict human preferences on generated face images, which demonstrates state-of-the-art performance. We believe this work will contribute to advancing generative models and help overcome current limitations in AIGC community." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "10989" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.251 + ], + "angle": 0, + "content": "This work is supported by the National Natural Science Foundation of China under Grants 62271308, 62401365, 62225112, 62132006, U24A20220, STCSM under Grants 24ZR1432000, 24511106902, 24511106900, 22511105700, 22DZ2229005, 111 plan under Grant Number BP0719010, China Postdoctoral Science Foundation under Grants BX20250411, 2025M773473, and State Key Laboratory of UHD Video and Audio Production and Presentation." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.265, + 0.188, + 0.28 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.29, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[1] Deep Floyd. https://github.com/deep-floyd/IF, Accessed: 2024-10-03. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.319, + 0.484, + 0.358 + ], + "angle": 0, + "content": "[2] Dreamlike V2. https://huggingface.co/dreamlike-art/dreamlike-photoreal-2.0, Accessed: 2024-10-03.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.362, + 0.482, + 0.401 + ], + "angle": 0, + "content": "[3] Flux-dev. https://huggingface.co/black-forest-labs/FLUX.1-dev, Accessed: 2024-10-03.1,3,5" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.405, + 0.48, + 0.432 + ], + "angle": 0, + "content": "[4] IP-Adapter-FaceID-SDXL. https://huggingface.co/h94/IP-Adapter, Accessed: 2024-10-03. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.434, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[5] IP-Adapter-FaceID-Plus. https://huggingface.co/h94/IP-Adapter-FaceID, Accessed: 2024-10-03.3,5" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.477, + 0.482, + 0.503 + ], + "angle": 0, + "content": "[6] Kcolors. https://huggingface.co/Kwai-Kcolors/Kcolors, Accessed: 2024-10-03. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.506, + 0.482, + 0.545 + ], + "angle": 0, + "content": "[7] ProtoVision V6.6. https://huggingface.co/stablediffusionapi/protovision-xl-v6.6, Accessed: 2024-10-03. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.548, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[8] Realistic Vision V5.1. https://huggingface.co/SG161222/Realistic_Vision_V5.1_noVAE, Accessed: 2024-10-03. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.591, + 0.482, + 0.63 + ], + "angle": 0, + "content": "[9] SD3. https://huggingface.co/stabilityai/stable-diffusion-3-medium, Accessed: 2024-10-03. 3,5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.633, + 0.482, + 0.659 + ], + "angle": 0, + "content": "[10] GPT-4o. https://chatgpt.com/, Accessed: 2024-10-10.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.662, + 0.482, + 0.716 + ], + "angle": 0, + "content": "[11] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.719, + 0.484, + 0.787 + ], + "angle": 0, + "content": "[12] Fadi Boutros, Meiling Fang, Marcel Klemt, Biying Fu, and Naser Damer. Cr-fiqa: face image quality assessment by learning sample relative classifiability. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5836-5845, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.789, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[13] Nicolas Chahine, Stefania Calarasanu, Davide Garcia-Civiero, Theo Cayla, Sira Ferradans, and Jean Ponce. An image quality assessment dataset for portraits. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9968-9978, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[14] Chen et al. Promptia: Boosting the performance and generalization for no-reference image quality assessment via prompts. In ECCV, 2024. 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.102, + 0.29, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[15] Chaofeng Chen, Dihong Gong, Hao Wang, Zhifeng Li, and Kwan-Yee K Wong. Learning spatial attention for face super-resolution. IEEE Transactions on Image Processing (TIP), 30:1219-1231, 2020. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.151, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[16] Chaofeng Chen, Xiaoming Li, Lingbo Yang, Xianhui Lin, Lei Zhang, and Kwan-Yee K Wong. Progressive semantic-aware style transformation for blind face restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11896-11905, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.235, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[17] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arxiv:2310.00426, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.307, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[18] Wei-Ting Chen, Gurunandan Krishnan, Qiang Gao, SyYen Kuo, Sizhou Ma, and Jian Wang. Dsl-fqa: Assessing facial image quality via dual-set degradation learning and landmark-guided transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2931-2941, 2024. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.391, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[19] Yu Chen, Ying Tai, Xiaoming Liu, Chunhua Shen, and Jian Yang. Fsrnet: End-to-end learning face super-resolution with facial priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2492-2501, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.462, + 0.904, + 0.529 + ], + "angle": 0, + "content": "[20] Siying Cui, Jiankang Deng, Jia Guo, Xiang An, Yongle Zhao, Xinyu Wei, and Ziyong Feng. Idapter: Learning mixed features for tuning-free personalization of text-to-image models. arXiv preprint arXiv:2403.13535, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.533, + 0.905, + 0.585 + ], + "angle": 0, + "content": "[21] Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning. In Proceedings of the International Conference on Learning Representations (ICLR). 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.59, + 0.905, + 0.658 + ], + "angle": 0, + "content": "[22] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4690-4699, 2019. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.661, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[23] Huiyu Duan, Qiang Hu, Jiarui Wang, Liu Yang, Zitong Xu, Lu Liu, Xiongkuo Min, Chunlei Cai, Tianxiao Ye, Xiaoyun Zhang, et al. Fine-grained user generated content video quality assessment. In m, 2025. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.718, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[24] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.789, + 0.907, + 0.857 + ], + "angle": 0, + "content": "[25] S. Alireza Golestaneh, Saba Dadsetan, and Kris M. Kitani. No-reference image quality assessment via transformers, relative ranking, and self-consistency. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1220-1230, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Pro" + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10990" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "ceedings of the Advances in Neural Information Processing Systems, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.122, + 0.483, + 0.191 + ], + "angle": 0, + "content": "[27] Ke Gu, Guangtao Zhai, Min Liu, Xiaokang Yang, Wenjun Zhang, Xianghui Sun, Wanhong Chen, and Ying Zuo. Fisblim: A five-step blind metric for quality assessment of multiply distorted images. In Proceedings of the SiPS, pages 241-246, 2013. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.193, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[28] Yuchao Gu, Xintao Wang, Liangbin Xie, Chao Dong, Gen Li, Ying Shan, and Ming-Ming Cheng. Vqfr: Blind face restoration with vector-quantized dictionary and parallel decoder. arXiv preprint arXiv:2205.06803, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.25, + 0.483, + 0.305 + ], + "angle": 0, + "content": "[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.307, + 0.482, + 0.362 + ], + "angle": 0, + "content": "[30] Xuanhua He, Quande Liu, Shengju Qian, Xin Wang, Tao Hu, Ke Cao, Keyu Yan, Man Zhou, and Jie Zhang. Id-animator: Zero-shot identity-preserving human video generation. arXiv preprint arXiv:2404.15275, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.364, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[31] Javier Hernandez-Ortega, Javier Galbally, Julian Fierrez, Rudolf Haraksim, and Laurent Beslay. Faceqnet: Quality assessment for face recognition based on deep learning. In 2019 International Conference on Biometrics (ICB), pages 1–8. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.435, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[32] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.491, + 0.482, + 0.546 + ], + "angle": 0, + "content": "[33] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 33: 6840-6851, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.549, + 0.482, + 0.603 + ], + "angle": 0, + "content": "[34] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssim. In Proceedings of the IEEE International Conference on Pattern Recognition (ICPR), pages 2366-2369. IEEE, 2010. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.606, + 0.482, + 0.674 + ], + "angle": 0, + "content": "[35] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.676, + 0.482, + 0.745 + ], + "angle": 0, + "content": "[36] Qiang Hu, Qihan He, Houqiang Zhong, Guo Lu, Xiaoyun Zhang, Guangtao Zhai, and Yanfeng Wang. Varfvv: View-adaptive real-time interactive free-view video streaming with edge computing. IEEE Journal on Selected Areas in Communications, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.747, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[37] Qiang Hu, Zihan Zheng, Houqiang Zhong, Sihua Fu, Li Song, Xiaoyun Zhang, Guangtao Zhai, and Yanfeng Wang. 4dgc: Rate-aware 4d gaussian compression for efficient streamable free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 875-885, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[38] Gary B. Huang, Manu Ramesh, Tamara Berg, and Erik Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. Technical Report 07-49, University of Massachusetts, Amherst, 2007. 3" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[39] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 36:78723-78747, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.163, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[40] Haoyun Jiang, Fei Huang, Qiang Hu, Minmin Sun, Shuai Xiao, Yong Li, Junyang Lin, Jiangchao Yao, et al. Catekv: On sequential consistency for long-context llm inference acceleration. In Proceedings of International Conference on Machine Learning (ICML), 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.233, + 0.905, + 0.3 + ], + "angle": 0, + "content": "[41] Byungho Jo, Donghyeon Cho, In Kyu Park, and Sungeun Hong. Ifqa: Interpretable face quality assessment. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 3444-3453, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.302, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[42] Le Kang, Peng Ye, Yi Li, and David Doermann. Convolutional neural networks for no-reference image quality assessment. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.358, + 0.905, + 0.425 + ], + "angle": 0, + "content": "[43] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 4401-4410, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.427, + 0.905, + 0.509 + ], + "angle": 0, + "content": "[44] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6007-6017, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.511, + 0.907, + 0.58 + ], + "angle": 0, + "content": "[45] Ira Kemelmacher-Shlizerman, Steven M Seitz, Daniel Miller, and Evan Brossard. The megaface benchmark: 1 million faces for recognition at scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4873-4882, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.581, + 0.905, + 0.622 + ], + "angle": 0, + "content": "[46] Deokyun Kim, Minseon Kim, Gihyun Kwon, and Dae-Shik Kim. Progressive face super-resolution via attention to facial landmark. arXiv preprint arXiv:1908.08239, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.623, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[47] Diederik P Kingma, Max Welling, et al. Auto-encoding variational bayes. 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.651, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[48] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Matiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. Proceedings of Advances in neural information processing systems (Neurips), 36:36652-36663, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.721, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[49] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1931-1941, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.79, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[50] Li et al. Agiqa-3k: An open database for ai-generated image quality assessment. arXiv preprint arXiv:2306.04717, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[51] Chunyi Li, Zicheng Zhang, Haoning Wu, Wei Sun, Xiongkuo Min, Xiaohong Liu, Guangtao Zhai, and Weisi Lin. Agiaq-3k: An open database for ai-generated image quality assessment. arXiv preprint arXiv: 2306.04717, 2023. 1, 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10991" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[52] Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Linmiao Xu, and Suhail Doshi. Playground v2.5: Three insights towards enhancing aesthetic quality in text-to-image generation, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.15, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[53] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proceedings of International Conference on Machine learning (ICML), pages 12888-12900, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.221, + 0.483, + 0.276 + ], + "angle": 0, + "content": "[54] Xiaoming Li, Ming Liu, Yuting Ye, Wangmeng Zuo, Liang Lin, and Ruigang Yang. Learning warped guidance for blind face restoration. In The European Conference on Computer Vision (ECCV), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.278, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[55] Xiaoming Li, Wenyu Li, Dongwei Ren, Hongzhi Zhang, Meng Wang, and Wangmeng Zuo. Enhanced blind face restoration with multi-exemplar images and adaptive spatial feature fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2706-2715, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.363, + 0.482, + 0.43 + ], + "angle": 0, + "content": "[56] Xiaoming Li, Shiguang Zhang, Shangchen Zhou, Lei Zhang, and Wangmeng Zuo. Learning dual memory dictionaries for blind face restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.434, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[57] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, Ming-Ming Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. arXiv preprint arXiv:2312.04461, 2023. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.491, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[58] Zhimin Li, Jianwei Zhang, Qin Lin, Jiangfeng Xiong, Yanxin Long, Xinchi Deng, Yingfang Zhang, Xingchao Liu, Minbin Huang, Zedong Xiao, et al. Hunyuan-dit: A powerful multi-resolution diffusion transformer with fine-grained chinese understanding. arXiv preprint arXiv:2405.08748, 2024. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.576, + 0.482, + 0.658 + ], + "angle": 0, + "content": "[59] Youwei Liang, Junfeng He, Gang Li, Peizhao Li, Arseniy Klimovskiy, Nicholas Carolan, Jiao Sun, Jordi Pont-Tuset, Sarah Young, Feng Yang, et al. Rich human feedback for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19401-19411, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.661, + 0.482, + 0.728 + ], + "angle": 0, + "content": "[60] Zhang Lijun, Shao Xiaohu, Yang Fei, Deng Pingling, Zhou Xiangdong, and Shi Yu. Multi-branch face quality assessment for face recognition. In 2019 IEEE 19th International Conference on Communication Technology (ICCT), pages 1659–1664. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.732, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[61] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of European conference on computer vision (ECCV), pages 740-755. Springer, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.803, + 0.482, + 0.869 + ], + "angle": 0, + "content": "[62] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Ben Fei, Bo Dai, Wanli Ouyang, Yu Qiao, and Chao Dong. Diff-bir: Towards blind image restoration with generative diffusion prior. arXiv preprint arXiv:2308.15070, 2023. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[63] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.094, + 0.905, + 0.12 + ], + "angle": 0, + "content": "the Advances in Neural Information Processing Systems (NeurIPS), pages 34892-34916, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[64] Xiongkuo Min, Ke Gu, Guangtao Zhai, Jing Liu, Xiaokang Yang, and Chang Wen Chen. Blind quality assessment based on pseudo-reference image. IEEE Transactions on Multimedia (TMM), 20(8):2049-2062, 2018. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.178, + 0.905, + 0.231 + ], + "angle": 0, + "content": "[65] Xiongkuo Min, Guangtao Zhai, Ke Gu, Yutao Liu, and Xiaokang Yang. Blind image quality estimation via distortion aggravation. IEEE Transactions on Broadcasting, 64(2): 508-517, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.233, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[66] Xiongkuo Min, Huiyu Duan, Wei Sun, Yucheng Zhu, and Guangtao Zhai. Perceptual video quality assessment: A survey. Science China Information Sciences, 67(11): 211301, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.289, + 0.905, + 0.329 + ], + "angle": 0, + "content": "[67] Anish Mittal, Rajiv Soundararajan, and Alan C. Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal Processing Letters, 20(3):209-212, 2013. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.331, + 0.905, + 0.412 + ], + "angle": 0, + "content": "[68] Stylianos Moschoglou, Athanasios Papaioannou, Christos Sagonas, Jiankang Deng, Irene Kotsia, and Stefanos Zafeiriou. Agedb: the first manually collected, in-the-wild age database. In proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 51-59, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.414, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[69] Fu-Zhao Ou, Xingyu Chen, Ruixin Zhang, Yuge Huang, Shaoxin Li, Jilin Li, Yong Li, Liujuan Cao, and Yuan-Gen Wang. Sdd-fiqa: Unsupervised face image quality assessment with similarity distribution distance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7670–7679, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.498, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[70] Pablo Pernias, Dominic Rampas, Mats L Richter, Christopher J Pal, and Marc Aubreville. Würstchen: An efficient architecture for large-scale text-to-image diffusion models. arXiv preprint arXiv:2306.00637, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.554, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[71] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.623, + 0.905, + 0.664 + ], + "angle": 0, + "content": "[72] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.665, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[73] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.721, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[74] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.79, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[75] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22500-22510, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[76] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed" + }, + { + "type": "list", + "bbox": [ + 0.525, + 0.094, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10992" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.092, + 0.484, + 0.161 + ], + "angle": 0, + "content": "Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.162, + 0.484, + 0.259 + ], + "angle": 0, + "content": "[77] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 35:36479-36494, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.26, + 0.483, + 0.3 + ], + "angle": 0, + "content": "[78] B Series. Methodology for the subjective assessment of the quality of television pictures. Recommendation ITU-R BT, 500(13), 2012. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.302, + 0.483, + 0.343 + ], + "angle": 0, + "content": "[79] Karen Simonyan and Andrew Zisserman. Very Deep Convolutional Networks for Large-Scale Image Recognition. arXiv e-prints, art. arXiv:1409.1556, 2014. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.344, + 0.483, + 0.371 + ], + "angle": 0, + "content": "[80] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv:2010.02502, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.373, + 0.483, + 0.44 + ], + "angle": 0, + "content": "[81] Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3664-3673, 2020. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.442, + 0.483, + 0.509 + ], + "angle": 0, + "content": "[82] Shaolin Su, Hanhe Lin, Vlad Hosu, Oliver Wiedemann, Jinqiu Sun, Yu Zhu, Hantao Liu, Yanning Zhang, and Dietmar Saupe. Going the extra mile in face image quality assessment: A novel database and model. IEEE Transactions on Multimedia, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.511, + 0.483, + 0.594 + ], + "angle": 0, + "content": "[83] Philipp Terhorst, Jan Niklas Kolf, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. Ser-fiq: Unsupervised estimation of face image quality based on stochastic embedding robustness. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5651-5660, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.595, + 0.483, + 0.663 + ], + "angle": 0, + "content": "[84] Yuan Tian, Kaiyuan Ji, Rongzhao Zhang, Yankai Jiang, Chunyi Li, Xiaosong Wang, and Guangtao Zhai. Towards all-in-one medical image re-identification. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 30774–30786, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.665, + 0.483, + 0.719 + ], + "angle": 0, + "content": "[85] Yuri Viazovetskyi, Vladimir Ivashkin, and Evgeny Kashin. Stylegan2 distillation for feed-forward image manipulation. Proceedings of the European conference on computer vision (ECCV), pages 170-186, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.721, + 0.483, + 0.788 + ], + "angle": 0, + "content": "[86] Jiarui Wang, Huiyu Duan, Jing Liu, Shi Chen, Xiongkuo Min, and Guangtao Zhai. Aigciqa2023: A large-scale image quality assessment database for ai generated images: from the perspectives of quality, authenticity and correspondence. 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.79, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[87] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. arXiv preprint arXiv:2305.07015, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.846, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[88] Jiarui Wang, Huiyu Duan, Guangtao Zhai, and Xiongkuo Min. Understanding and evaluating human preferences for ai generated images with instruction tuning. arXiv preprint arXiv:2405.07346, 2024. 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.092, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[89] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin C.K. Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.149, + 0.905, + 0.191 + ], + "angle": 0, + "content": "[90] Liangbin Xie. Xintao Wang, Honglun Zhang, Chao Dong, and Ying Shan. Vfhq: A high-quality dataset and benchmark for video face super-resolution, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.193, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[91] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.25, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[92] Tao Wang, Kaihao Zhang, Xuanxi Chen, Wenhan Luo, Jiankang Deng, Tong Lu, Xiaochun Cao, Wei Liu, Hongdong Li, and Stefanos Zafeiriou. A survey of deep face restoration: Denoise, super-resolution, deblur, artifact removal. arXiv preprint arXiv:2211.02831, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.321, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[93] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In International Conference on Computer Vision Workshops (ICCVW). 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.377, + 0.905, + 0.433 + ], + "angle": 0, + "content": "[94] Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. Towards real-world blind face restoration with generative facial prior. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.435, + 0.905, + 0.49 + ], + "angle": 0, + "content": "[95] Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. Towards real-world blind face restoration with generative facial prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.492, + 0.905, + 0.546 + ], + "angle": 0, + "content": "[96] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing (TIP), 13(4):600-612, 2004. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.549, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[97] Zhixin Wang, Ziying Zhang, Xiaoyun Zhang, Huangjie Zheng, Mingyuan Zhou, Ya Zhang, and Yanfeng Wang. Dr2: Diffusion-based robust degradation remover for blind face restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1704-1713, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.633, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[98] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15943-15953, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.718, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[99] Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xinglong Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. Diffir: Efficient diffusion model for image restoration. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13095-13105, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.789, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[100] Weihao Xia, Yujiu Yang, Jing-Hao Xue, and Baoyuan Wu. Tedigan: Text-guided diverse face image generation and manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2256-2265, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[101] Guangxuan Xiao, Tianwei Yin, William T Freeman, Frédo Durand, and Song Han. Fastcomposer: Tuning-free multisubject image generation with localized attention. Inter-" + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10993" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "national Journal of Computer Vision (IJCV), 133(3):1175-1194, 2025. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.189 + ], + "angle": 0, + "content": "[102] Jingtao Xu, Peng Ye, Qiaohong Li, Haiqing Du, Yong Liu, and David S. Doermann. Blind image quality assessment based on high order statistics aggregation. IEEE Transactions on Image Processing (TIP), 25:4444-4457, 2016. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.483, + 0.274 + ], + "angle": 0, + "content": "[103] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 36, 2024. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.483, + 0.359 + ], + "angle": 0, + "content": "[104] Liu Yang, Huiyu Duan, Long Teng, Yucheng Zhu, Xiaohong Liu, Menghan Hu, Xiongkuo Min, Guangtao Zhai, and Patrick Le Callet. Aigcoiaq2024: Perceptual quality assessment of ai generated omnidirectional images. In 2024 IEEE International Conference on Image Processing (ICIP), pages 1239-1245. IEEE, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.361, + 0.484, + 0.416 + ], + "angle": 0, + "content": "[105] Shuo Yang, Ping Luo, Chen-Change Loy, and Xiaou Tang. Wider face: A face detection benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5525–5533, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.418, + 0.483, + 0.486 + ], + "angle": 0, + "content": "[106] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. MANIQA: Multi-dimension Attention Network for No-Reference Image Quality Assessment. arXiv e-prints, art. arXiv:2204.08958, 2022. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.488, + 0.483, + 0.556 + ], + "angle": 0, + "content": "[107] Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. Gan prior embedded network for blind face restoration in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 672-681, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.558, + 0.483, + 0.613 + ], + "angle": 0, + "content": "[108] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.615, + 0.483, + 0.655 + ], + "angle": 0, + "content": "[109] Zongsheng Yue and Chen Change Loy. Difface: Blind face restoration with diffused error contraction. arXiv preprint arXiv:2212.06512, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.657, + 0.483, + 0.698 + ], + "angle": 0, + "content": "[110] Guangtao Zhai and Xiongkuo Min. Perceptual image quality assessment: a survey. Science China Information Sciences, 63:1-52, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.7, + 0.483, + 0.755 + ], + "angle": 0, + "content": "[111] Lin Zhang, Lei Zhang, and Alan C. Bovik. A feature-enriched completely blind image quality evaluator. IEEE Transactions on Image Processing (TIP), 24(8):2579-2591, 2015. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.757, + 0.483, + 0.826 + ], + "angle": 0, + "content": "[112] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.828, + 0.483, + 0.897 + ], + "angle": 0, + "content": "[113] Shangchen Zhou, Kelvin Chan, Chongyi Li, and Chen Change Loy. Towards robust blind face restoration with codebook lookup transformer. Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 35:30599-30611, 2022. 2, 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.897 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10994" + } + ] +] \ No newline at end of file diff --git a/2025/F-Bench_ Rethinking Human Preference Evaluation Metrics for Benchmarking Face Generation, Customization, and Restoration/86d5740f-dff5-47ab-aa67-761b97593a44_origin.pdf b/2025/F-Bench_ Rethinking Human Preference Evaluation Metrics for Benchmarking Face Generation, Customization, and Restoration/86d5740f-dff5-47ab-aa67-761b97593a44_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e4cb288cb7ed502116c1d5abc405afb3613fea53 --- /dev/null +++ b/2025/F-Bench_ Rethinking Human Preference Evaluation Metrics for Benchmarking Face Generation, Customization, and Restoration/86d5740f-dff5-47ab-aa67-761b97593a44_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d84d290b7dbd0f912ddcde2d40df7273d304ed2048d7b65e4a8b199676e57c9a +size 21396946 diff --git a/2025/F-Bench_ Rethinking Human Preference Evaluation Metrics for Benchmarking Face Generation, Customization, and Restoration/full.md b/2025/F-Bench_ Rethinking Human Preference Evaluation Metrics for Benchmarking Face Generation, Customization, and Restoration/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0646e65804f5aebfb5e1648c703ba39ab945c772 --- /dev/null +++ b/2025/F-Bench_ Rethinking Human Preference Evaluation Metrics for Benchmarking Face Generation, Customization, and Restoration/full.md @@ -0,0 +1,409 @@ +# F-Bench: Rethinking Human Preference Evaluation Metrics for Benchmarking Face Generation, Customization, and Restoration + +Lu Liu $^{1,*}$ , Huiyu Duan $^{1,*}$ , Qiang Hu $^{1,\dagger}$ , Liu Yang $^{1}$ , Chunlei Cai $^{2}$ + +Tianxiao Ye $^{2}$ , Huayu Liu $^{1}$ , Xiaoyun Zhang $^{1}$ , Guangtao Zhai $^{1}$ + +1Shanghai Jiao Tong University, Shanghai, China 2Bilibili Inc., Shanghai, China + +![](images/be54ea568985a727a2cc4028c6a24db27d8ed7043d43028000e0e447b132f563.jpg) +Figure 1. Data construction pipeline of $F$ -bench and content overview of FaceQ database. From left to right, (a) Diverse input source data for face generation, customization, and restoration. (b) $F$ -Bench benchmarks 29 face generative models, including 14 face generation models, 6 face customization models, and 9 face restoration models. (c) Multi-dimensional subjective preference evaluation in $F$ -Bench. (d) 3D scatter plots of MOS distributions across FaceQ three subsets. + +![](images/50a31b7cb83ff8691ad34b7636a1481484922af6f46b1659ee084656b72f3aeb.jpg) + +![](images/cf85e2bc4cae21c3c7775c8426f5183047677a9fc1bbce0840f052dde717a00a.jpg) + +# Abstract + +Recent artificial intelligence (AI) generative models have demonstrated remarkable capabilities in image production, and have been widely applied to face image generation, customization, and restoration. However, many AI-generated faces (AIGFs) still suffer from issues such as unique distortions, unrealistic details, and unexpected identity shifts, underscoring the need for a comprehensive quality evaluation method for AIGFs. To this end, we introduce FaceQ, the first comprehensive AI-generated Face image database with fine-grained Quality annotations aligned with human preferences, which consists of 12K images and 491K ratings across multiple dimensions. Using the FaceQ database, we establish F-Bench, a benchmark for comparing and evaluating face generation, customization, and restoration models, highlighting strengths and weaknesses across various prompts and evaluation dimensions. Additionally, we assess the performance of existing image quality assessment (IQA) methods on FaceQ, and further propose a large multimodal model (LMM) based Face quality Evaluator (F + +Eval) to accurately assess the multi-dimensional quality of generated faces in a one-for-all manner. Extensive experimental results demonstrate the state-of-the-art performance of our F-Eval. The project page is: https://media-sjtu.github.io/F-Bench/. + +# 1. Introduction + +Generative models, including VAE [47], GAN [26], and diffusion models [33, 74, 80], have made remarkable progress in recent years, and have been extended to the fields of text-to-image generation [3], text-driven image editing [44], image restoration [99], etc, leading to high-quality artificial intelligence generated images (AIGIs). As an important aspect of AIGIs, many studies have also investigated face generation [100], face customization [57, 108], and face restoration [62, 87], producing creative, high-quality face images. Despite these advances, many AI-generated faces (AIGFs) still suffer from perceptual deficiencies. + +Many natural face image quality assessment databases have been established in the literature [48, 51, 59, 86]. However, natural images are mainly degraded by distor + +Table 1. Comparison between FaceQ and existing AIGC quality assessment databases (top) as well as natural face quality assessment databases (bottom). FaceQ is the first large-scale AI-generated, customized and restored face quality assessment database. + +
DatabaseDomainSourceImagesScoresDimensions
AGIQA-3K (TCSVT2024) [51]GeneralAI-generated2,982MOSPerception, Alignment
AIGCIQA (CVPR2024) [86]GeneralAI-generated2,400MOSQuality, Authenticity, Correspondence
RichHF-18K (CVPR2024) [59]GeneralAI-generated17,760MOSPlausibility, Aesthetics, Text-image Alignment, Overall
PIQ23 (CVPR2023) [13]FaceReal5,116PairOverall, Details, Exposure
FaceQFaceAI-generated, customized and restored12,255MOSQuality, Authenticity, ID Fidelity, Correspondence
+ +tions such as noise, blur, compression [23, 66, 110], etc., while AIGIs generally suffer from unique distortions including unrealistic structures and text-image inconsistencies [59, 86], etc. Recently, many studies have investigated human preferences for AIGIs [39, 50, 59, 86, 103]. However, these general AIGI quality assessment works may not be well aligned with human preferences on AIGFs, since the perceptual quality of AIGFs can be significantly affected by small deviations in facial attributes,(e.g., eye size, skin texture, identity shifts), while the number of AIGFs in these AIGI databases is limited. Moreover, current AIGI quality assessment works mainly focus on the image generation, while human preferences for face customization have rarely been considered, where identity fidelity is significant. + +The above limitation highlights the necessity of establishing a comprehensive quality evaluator for AIGFs. To this end, we propose FaceQ, the first AI-generated Face image Quality assessment database with three popular tasks. Specifically, FaceQ contains $12\mathrm{K}+$ images obtained from 29 generative models across three tasks: 1) face generation, 2) face customization, and 3) face restoration, as shown in Fig.1. Based on the generated images, we collect $491\mathrm{K}+$ quality ratings from multiple dimensions including quality, authenticity, identity (ID) fidelity, and text-image correspondence, respectively, finally forming $32\mathrm{K}+$ mean opinion scores (MOSs). Using FaceQ, we establish a human preference benchmark, i.e., F-Bench, for face generation, customization, and restoration models, and thoroughly analyze the generation performance across different evaluation perspectives, models, and categories. In addition, we also compare the correlation of existing image quality assessment (IQA), face quality assessment (FQA), AIGI quality assessment, and preference evaluation metrics with human preferences. + +Furthermore, we present F-Eval, the first generated Face image quality Evaluation method designed to perform comprehensive multi-dimensional quality evaluation for generated, customized, and restorted faces in a one-for-all manner. Specifically, F-Eval leverages large multimodal models (LMMs) [11] and adopts instruction tuning [63] for preference-related vision-language adaption. We also apply low-rank adaptation (LoRA) [35] techniques for more efficient feature refining. Extensive experimental results demonstrate that F-Eval outperforms existing quality assessment methods in terms of multiple dimensions relevant to human preference for all three tasks. The main contribu + +tions of our work are highlighted as follows: + +- We construct FaceQ, the first generated face quality assessment database consisting $491\mathrm{K}+$ human preference ratings from quality, authenticity, ID fidelity, correspondence dimensions, on $12\mathrm{K}+$ generative face images across three tasks. +- We establish F-Bench, a comprehensive human preference benchmark for 29 face generation, customization, and restoration methods. It also extensively evaluates the alignment of 26 existing quality assessment methods with human preferences. +- We present F-Eval, a LMM-based quality assessment model to predict multi-dimensional human preference on generated faces. Extensive experiments demonstrate the state-of-art QA performance of the proposed F-Eval. + +# 2. Related Work + +Face Generation & Customization & Restoration. (1) For face generation models, earlier works such as GANs [26, 43, 85] have demonstrated significant influences. Recently, diffusion models [73, 74, 77] have achieved rapid developments in text-to-image generation. These models can be categorized into pixel-space diffusion models [1, 72, 73, 76], and latent space diffusion models [17, 58, 71, 74]. (2) For face customization, this is a recently defined task that aims to create images of a specific person with the input of an identity reference image and prompts [98]. It can be divided into two categories including conventional test-time optimization [24, 49, 75] and tuning-free customization methods utilizing pre-trained diffusion models [20, 30, 57, 91, 101, 108]. (3) For face restoration, which aims to restore high-quality face images from their low-quality counterparts. Some traditional restoration methods utilize geometric priors [16, 19, 46] or reference priors [54, 55] while in recent years, exploring generative priors such as GAN priors [15, 28, 94, 107, 113] and diffusion priors [36, 37, 62, 84, 87, 97, 109] has become a prevalent trend. Despite these improvements, the generated, customized, and restored images may suffer from various quality issues, highlighting the necessity of evaluators. + +Face Image Quality Assessment Database. Existing face image quality assessment (FIQA) can be divided into Biometric FIQA [12, 31, 60, 69] and Generic FIQA [13, 18, 82, 92]. BFIQA methods are most developed to evaluate the biometric utility for face recognition. GFIQA meth + +![](images/0ca6f16b7c42b7cffbe0d97807f01593ad18168ac065d62cf16b2ebdca062ecf.jpg) +Figure 2. Rating comparisons of eight dimensions. Each column presents a pair of intuitive examples of each dimension, with red indicating the better rating and blue indicating the worse one. From left to right, the subsets are face generation, face customization, and face restoration subsets. The last row displays the corresponding prompts, reference image-prompt pairs, and the GT-LQ image pairs. + +![](images/15cf956e90dd11acd6ae04bce2981f7b38875f0c25c21bc20e03bb64418aa45d.jpg) + +![](images/959bcd90fe7fba8303ac06245c48e7d7a9e96871c01f3bfb428c7c084619b239.jpg) + +ods [82] focus on perceptual degradation in face restoration task. Both of these databases consist of natural face images, which has distinctly different nature with generated faces. + +AIGC Quality Assessment Database. In recent years, plenty of researchers have raised interest in assessing AI-generated images(AIGIs) quality assessment [23, 40, 48, 51, 86, 104] by building various databases as shown in Tab 1. All of these existing database contain various categories of general AIGI images, leaving a notable gap in datasets dedicated to AI-generated face images. + +# 3. FaceQ Database Construction + +In this section, we introduce the proposed generative face quality assessment database FaceQ. The database includes 12,255 images and 491,130 human opinion ratings. Our FaceQ demonstrates the merits of comprehensive generated visual content and multi-dimensional labeling, which can significantly facilitate the advancement of AIGC research. + +# 3.1. Data Collection + +# 3.1.1. FaceQ-Gen Subset Collection + +Prompt Sources. The prompts in FaceQ-Gen subset are sourced from MS-COCO [61] [57] and GPT-4o [10]. These face-centric prompts can be categorized into nine types: General, Accessories, Clothes, Action, View, Expression, Background, Style, and Facial Attributes. + +Models Collection. We collect 14 representative face generation models, including Deep Floyd [1], Stable Diffusion V1.5 [74], Stable Diffusion V2.1 [74], SD-XL [71], Realistic Vision [8], Stable Cascade [70], ProtoVision V6.6 [7], SD3 [9], Playground V2.5 [52], Kolors [6], PixArt-alpha [17], Hunyuan [58], and Flux-dev [3]. Negative prompts and safe sensors are enabled. In total, FaceQ-Gen includes 4,032 generated face images. + +# 3.1.2. FaceQ-Cus Subset Collection + +Identity and Prompt Sources. Face customization relies on a reference identity image and a guiding prompt. The reference images in FaceQ-Cus are partially self-collected, + +with the rest sourced from [57]. The prompts are selected and filtered from the FaceQ-Gen prompt list. + +Models Collection. We collect six representative face customization methods, including ELITE [98], FastComposer [101], IP-Adapter-FaceID [4], InstantID [91], IP-Adapter-Face-ID-Plus [5], and PhotoMaker V2 [57]. Here, we use single reference image as input. In total, FaceQ-Cus includes 4,200 generated face images. + +# 3.1.3. FaceQ-Res Subset Collection + +Low Quality Images Source. (1) Synthetic: We construct two degraded pipelines following previous works [15, 93, 95, 97] to simulate real-world degradations on both indomain data from VFHQ [90], CelebRef [56] and out-of-domain self-collected high-quality Asian face images (2) Real-world: Real-world low-quality images are sourced from various public datasets, including AgeAB [68], CelebChild [28], Wider [105], LFW [38] and MegaFace [45], ensuring the degradation diversity. + +Models Collection and Face Image Restoration. We evaluate both diffusion-prior-based methods such as DiffBIR [62], DifFace [109], StableSR [87], and GAN or Transformer-based methods GFPGAN [94], GPEN [107], CodeFormer [113], VQFR [28] and Face-SPARNet [15]. Each method is benchmarked on both the $F$ -Res synthetic dataset and the real-world dataset. In total, FaceQ-Res include 4023 restored images, with 2007 images in the real-world case and 2016 images in the synthetic case. + +# 3.2. Human Preference Dimension Design + +The evaluation dimensions of generated faces need specific design to adequately reflect human preference, given the complexity of perceptual factors involved. To the end, we decompose "quality" into four core dimensions: Quality, Authenticity, ID Fidelity, and Correspondence. + +Quality. This dimension evaluates the overall perceptual quality of the image, considering factors such as color accuracy, blur, noise, and artifacts. It is used in all tasks. + +Authenticity. This dimension evaluates how closely the + +![](images/906cf53f1707bfcabc0ffea135230ee9f39b5e4307d42b8189ee2f896f3f2245.jpg) +(a) Face generation MOS distribution + +![](images/f75d4c146079fe3a3a587bb40ab780de22b24212ea513086b7e2ba41df47cde3.jpg) +(b) Face customization MOS distribution + +![](images/e6fc96cb02e178976b89dc1fedc2b07e1fc527d99f6a9e0dbe615ac2b0e0470d.jpg) +(c) Face restoration MOS distribution + +![](images/3c50297297cdc84bc438869a8f6c868263c2dfd6249d458e8a5464371dc5356b.jpg) +(d) Hunyuan-DiT +Figure 3. Mean Opinion Score (MOS) distribution histograms and kernel density curves. (a)-(c) MOS distributions for three subsets. (d)-(h) Model-wise MOS distributions. MOS Distributions of other methods are provided in the supplementary material. + +![](images/5c5a432204711046296b8eb590eb28e6cfa3397c63fa6bd83fcdec4b96b1c1c7.jpg) +(e) SD3 + +![](images/79d7ab75f175a10401291e5733ca09c3a9a45d4049bcae52a7f8d3669e2ca469.jpg) +(f) InstantID + +![](images/96e20e2700d895b48c3bae974748669a32f3565913b33186b0a6b646c5450d6c.jpg) +(g) VQFR (real world) + +![](images/5549685e6ba630acb1c517d5728a32ae0ca6ceee01efa7ff46dfa9400665b798.jpg) +(h) StableSR (synthetic) + +generated image resembles a natural, real-life photograph, with particular emphasis on realistic skin texture, facial details, and wrinkles. It is used in face generation task. + +ID Fidelity. This dimension evaluates how well the identity of the reference image is preserved in the customized or restored image in face customization and restoration tasks. + +Correspondence. This dimension evaluates text-to-image alignment in face generation and customization tasks. + +# 3.3. Subjective Experiment + +Main Assessment. To ensure a comprehensive, fair, and reliable subjective assessment, 180 participants with normal or corrected-to-normal eyesight are recruited for the experiment. They first take expert-led tutorials before the main assessment. In the main assessment, participants are asked to rate the image by dragging a sliding window from 0 to 5 with a two-digit decimal from different dimensions. All images are displayed on 27-inch 4K Dell monitors, randomly presented, at their original resolution under standard lighting conditions. The experiment follows the ITU-R BT.500-14 guidelines [78] for subjective evaluations. + +Subjective Score Processing. To eliminate the inter-annotator variability, outlier detection based on Kurtosis is conducted with a rejection rate of $3\%$ [78]. As a result, each image is rated by at least 15 subjects. These valid ratings are converted and scaled into Z-scores ranging from [0, 100] with the following formulas: + +$$ +z _ {i j} = \frac {r _ {i j} - \mu_ {i}}{\sigma_ {i}}, \quad z _ {i j} ^ {\prime} = \frac {1 0 0 (z _ {i j} + 3)}{6}, \tag {1} +$$ + +$$ +\mu_ {i} = \frac {1}{M _ {i}} \sum_ {j = 1} ^ {M _ {i}} r _ {i j}, \sigma_ {i} = \sqrt {\frac {1}{M _ {i} - 1} \sum_ {j = 1} ^ {M _ {i}} (r _ {i j} - \mu_ {i}) ^ {2}}, \tag {2} +$$ + +Here, $r_{ij}$ represents the original ratings provided by the $i$ -th evaluator for the $j$ -th image, $M_i$ denotes the total number of images that were assessed by the $i$ -th participant in this group. Finally, the mean opinion score (MOS) of the image $j$ ( $MOS_j$ ) is computed by averaging the rescaled + +z-scores as follows: + +$$ +M O S _ {j} = \frac {1}{N _ {j}} \sum_ {i = 1} ^ {N _ {j}} z _ {i j} ^ {\prime} \tag {3} +$$ + +where $N_{j}$ denotes the number of valid subjects for image $j$ , and $z_{ij}^{\prime}$ denotes rescaled z-scores. + +# 4. F-Bench: Benchmarking Current Generative Face Models + +In this section, we establish a human preference benchmark F-Bench for today's state-of-the-art face generation, customization, and restoration methods, from quality, authenticity, ID fidelity, and correspondence dimensions. This benchmark analyzes the strengths and weaknesses, providing new insights and potential directions for future work. + +# 4.1. Perspective Analysis + +Fig.3 demonstrate the MOS distribution across eight dimensions in the three subsets, highlighting distinct variations between dimensions. (1) In the face generation task, the Correspondence score is the highest among the three dimensions, indicating that existing generative models perform well in aligning with prompts. The distribution of Authenticity is skewed towards lower scores, suggesting that current face generation models still fall short in generating photorealistic face images that satisfy human preferences. (2) In face customization, models perform the worst in the Quality dimension overall, demonstrating the difficulty of injecting identity information without compromising image quality. Identity Fidelity shows a bimodal distribution, reflecting the dichotomous nature of human judgment regarding whether the generated face accurately represents the same identity. (3) In face restoration, models perform better in Quality than in Identity Fidelity. + +# 4.2. Model-wise Comparison + +Face Generation. Fig.4a presents the average MOS scores across three dimensions: Quality, Authenticity, and Correspondence. Correspondence scores show a steady increase + +![](images/303eb0332611bac5a2f83daa16a9a9abd49abc0fce780e81bf45e786228ec912.jpg) +Figure 4. Average MOS score comparison across all models and dimensions. (a) Face generation. (b) Face customization. (c) Face restoration. The models are arranged in a clockwise order by release date. + +![](images/f59aec7a75dddcb25645866c238883c65a864408f00f39ecea66cc85f857374e.jpg) + +![](images/23104c2b45121fa81dc3d3fde8bf89c52a2708806106f4cf5126a2ed3de561b8.jpg) + +![](images/58c75820987283522713a3ba1b67a33e0791394c6a0c2832d19922b94416984a.jpg) +Figure 5. Violin plots of quality, authenticity and correspondence scores in FaceQ-Gen subset across nine prompt categories. + +over time, indicating consistent improvement in text comprehension. Quality scores generally rise with the release year, although Hunyuan [58] and Dreamlike [2] exhibit deviations. Hunyuan [58] oversmoothes the facial details and performs worse compared to its contemporaries, while Dreamlike [2] performs the opposite. Authenticity scores exhibit the greatest variation among models, suggesting that current face generation models place insufficient emphasis on authenticity. Except for Realistic Vision [8], SD3 [9], Flux-dev [3], and Dreamlike [2], most models lag in authenticity relative to quality, highlighting a common phenomenon of high-quality but less realistic outputs in generative models. Fig.3d,e give some examples of MOS distributions of different methods, revealing a reasonable fine-grained performance difference. + +Face Customization. Fig.4b is the performance radar chart for face customization models. IP-Adapter-FaceID-Plus[5] achieves the best performance across all dimensions, followed by PhotoMaker [57]. This highlights their great robustness in identity encoders. ELITE [98] scores the lowest in all dimensions. IP-Adapter-FaceID [4] produces high-quality images but lacks accuracy in Identity Fidelity. This may be because the base model [71] introduces many AI artifacts, which affect both Identity Fidelity and Quality. InstantID [91] and FastComposer [101] score well in Identity Fidelity but underperform significantly in Quality and Correspondence, indicating strong identity preservation but limited text-driven editing capabilities. Fig.3h illustrate the model-wise MOS distributions, revealing similar trends to those observed in the radar chart. + +Face Restoration. Fig.4c displays the average performance comparison of nine face restoration models. DR2 [97] exhibits the weakest performance in both Quality and Identity fidelity. This may be due to the degradation removal process of DR2 [97], which results in the loss of fine de + +tails and excessive smoothing. In contrast, StableSR [89] and CodeFormer [113] achieve the best overall performance among the nine face restoration models, successfully preserving both intricate details and accurate identity representations. When comparing the performance on real-world and synthetic data, most methods struggle to achieve satisfactory results on complex real-world degradations, except for SAPRNet [15] and DiffBIR [62]. StableSR [89] and GFPGAN [94] show strong robustness, performing consistently well on both real-world and synthetic data. + +# 4.3. Class-wise Comparison + +For the face generation task, we analyze the Mean Opinion Score (MOS) distributions of Quality, Authenticity, and Correspondence across nine prompt categories, as shown in Fig.5. The violin plots highlight distinct performance patterns across these categories. The Action, Expression, and Facial Attributes categories exhibit lower MOS values in all three dimensions, with notable variability. In the Action category, limited visibility of the face due to body movements affects the generation of fine facial details, resulting in reduced scores, especially in Quality and Correspondence. The Expression category poses challenges in maintaining spatial relationships and physical constraints, leading to distortion and artifacts in expressions with exaggerated facial movements, such as laughing or crying, thereby impacting both Quality and Authenticity. For Facial Attributes, generating images with precise attributes (e.g., specific eye color, hair style) is challenging, particularly when multiple attributes are specified, which lowers the scores. In contrast, the General and Background categories achieve the highest MOS scores across all dimensions, as simpler or less specific prompts are easier for the model to interpret accurately. The Clothes and Accessories categories also perform well, indicating that the model handles clothes and accessory prompts effectively. Overall, prompts with specific or + +![](images/ea49b49f9d03c0009d83adb59ba0355c21572a7476d5ebc10ade03953940f09e.jpg) +Figure 6. The overall framework of F-Eval. F-Eval can evaluate quality, authenticity, correspondence, and identity fidelity in a one-for-all framework. It can process both single and paired images, along with prompts, to produce quality scores. It consists of three encoders, including a vision encoder, a face encoder, and a text tokenizer to process multi-modal inputs. These features are projected into the same space by trained projectors. A pre-trained large language model is utilized to fuse the features while fine-tuned with four LoRA experts. Specific LoRA will be activated by dimension ID, which is classified by a trainable router. + +![](images/eabc3a39c23181b6f185bfa475841063d2fa90b82c807841ba30c66e7609deb8.jpg) + +complex attribute requirements tend to result in lower MOS scores, while general or simple prompts yield consistently high scores across dimensions. This can be attributed to the absence of high-quality fine-grained text-face pair datasets during the training of generative models. + +# 5. F-Eval Method + +In this section, we introduce our one-for-all generated face quality assessment method, F-Eval, towards predicting quality scores from multiple dimensions for three tasks using one model. + +# 5.1. Overall Architecture + +The overall architecture of F-Eval is depicted in Fig. 6. As shown on the left, F-Eval can assess four dimensions: quality, authenticity, text-image correspondence, and identity fidelity, with multimodal inputs, i.e.text prompts, image and optional reference image, to output quality-related answers. On the right, the input image(s) can originate from all three tasks. F-Eval employs a vision encoder and a face encoder to extract visual features, with projectors projecting both into a unified language space. User prompts are encoded into textual embeddings through a tokenizer. These multimodal features are concatenated and fed into a pre-trained large language model for the final output. + +# 5.2. Model Design + +Vision Encoding. For input image(s), F-Eval adopt a vision encoder to extract global features and a face encoder captures face features related to facial structure and identity. Specifically, the vision encoder, Qwen-VL-2.5-ViT [11], is tuned on FaceQ to adjust different dimensions. Face encoder is Arcface [22], which is frozen. Two two-layer multi-layer perception projectors then map the features. + +Feature Fusion via the LLM. For input prompts, it is tokenized into text tokens and then concatenated with visual tokens. Here, Qwen-VL2.5-7B [11] is employed as the pretrained LLM backbone to combine visual tokens and text tokens to perform multi-modal learning.. Finally, the output features of the LLM are decoded with a text decoder and then projected to the prompt-corresponded quality space. + +# 5.3. Multi-Expert LoRA Tuning + +Instruction Tuning. Previous IQA methods are often designed for specific dimensions [106, 112], addressing only one aspect of the problem or requiring retraining of model weights for different dimensions. Leveraging the generalization ability of LLMs, instruction tuning allows for the integration of different dimensions within a unified framework. Therefore, we construct diverse instruction data for the four facial quality evaluation dimensions in the FaceQ dataset and fine-tune F-Eval by adjusting the visual encoder, projector, and LLM parameters to adapt it to the specific task. As a result, F-Eval is capable of evaluating facial quality using unified weights. + +Mixture of LoRA Experts (MoLE). Tuning a holistic quality model for different dimensions using a single LoRA expert introduces inherent interference between dimensions. Hence, F-Eval integrates four LoRA experts for four dimensions when tuning vision encoder and LLM, constituting MoLE modules. A router is trained to classify input prompts into the four dimensions and output the corresponding dimension ID $k$ . Specifically, for a frozen linear layer $h = Wx$ with input $x \in \mathbb{R}^{d_i}$ and weight matrix $W \in \mathbb{R}^{d_o \times d_i}$ , a LoRA expert is trained to learn a low-rank decomposed update: + +$$ +h = W x + \Delta W x = W x + \frac {\alpha}{r} B A x, \tag {4} +$$ + +Table 2. Performance comparison on FaceQ-Gen and FaceQ-Cus subsets. $\spadesuit$ , $\clubsuit$ , $\diamond$ , and $\heartsuit$ denote traditional IQA models, face IQA models, classical deep learning-based IQA models, and AIGC IQA models respectively. * indicates fine-tuned results. + +
TaskFace GenerationFace Customization
DimensionQualityAuthenticityCorrespondenceQualityID FidelityCorrespondence
ModelSRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑
◆NIQE [67]0.17260.11420.25530.08690.05710.13460.07520.05070.10980.13650.08990.14360.14900.09910.13100.11860.07840.1245
◆ILNIQE [111]0.19810.13190.31590.06750.04450.15050.08890.05910.16650.20320.13270.24790.21920.14520.27780.18780.12510.2275
◆HOSA [102]0.25370.17080.29720.11030.07280.14460.1220.08140.14950.10190.06880.05970.07530.05060.05520.10360.06960.0748
◆BPRI-PSS [64]0.23620.15950.24130.17000.11420.17720.16730.11210.16470.02380.01660.03140.14250.09720.17580.11000.07290.1067
◆FISBLIM [27]0.09010.05950.04330.00280.00030.01750.10000.06650.01960.16140.10240.16900.02560.01730.01260.08520.05590.0854
◆ArcFace[22]N/AN/AN/AN/AN/AN/AN/AN/AN/A0.30990.20940.36050.50620.34390.55720.18370.12430.2244
◆SER-FIQ[83]0.15520.10340.15610.04700.03140.05750.11300.0750.1110-0.1241-0.080-0.1241-0.0109-0.00780.0042-0.1680-0.1121-0.1911
◆DSL-FIQA[18]0.56060.39270.59450.31560.20960.39600.37060.25140.43550.36150.24500.36510.24290.16520.25190.10400.07030.0974
◆IFQA[41]0.23130.15490.32610.15970.10620.20970.15970.10770.22600.24010.16140.22200.15030.09730.17390.37310.25460.3941
◆CNNIAQ*[42]0.42190.28950.35100.30370.20540.29730.28240.19310.25950.62440.43760.66200.65110.44410.63810.37650.25450.3264
◇VGG16*[79]0.58460.41410.60060.50100.35080.51810.43490.29940.47230.72170.53240.78570.67350.47170.66360.56350.39320.5589
◇VGG19*[79]0.57280.40430.56440.46810.32480.44790.41710.28730.47540.79520.60310.80960.75440.54170.75550.65850.47420.6465
◇ResNet18*[29]0.61500.44170.66600.46000.31600.46040.48290.33600.46280.79660.60130.80060.74730.53350.76560.63080.44880.6191
◇ResNet34*[29]0.60920.43540.65370.55640.39490.54010.44100.30650.52070.81410.62110.81510.79690.58780.81570.57080.40000.5810
◇HyperIAQ*[81]0.63330.45450.65910.61060.44060.60420.42350.29120.42400.85240.66310.82050.84190.64630.85880.66280.47570.6454
◇TRES*[25]0.77660.58670.80470.64580.47030.65790.57080.40380.62700.88830.70830.89510.83300.64090.86580.73930.54970.7400
◆MANIAQ*[106]0.78710.59510.81500.73080.54360.77940.62780.44430.70660.89700.71850.89520.85900.66540.87960.78490.59410.7822
◇PromptIAQ*[14]0.78230.59070.82450.72320.53270.75580.65140.46980.70650.90710.73150.88210.86930.67920.88760.83480.63420.8281
♥CLIPScore[32]0.09670.06420.13150.06400.04080.08170.23960.15910.27330.31260.21060.31850.18160.12180.18110.70080.50980.6986
♥BLIPScore[53]0.13060.08720.19290.09350.06130.15690.20510.13640.29630.30730.20630.31360.12130.07960.12660.67070.47970.6123
♥ImageReward [103]0.38490.26310.43580.26060.17360.31810.51550.35420.58710.32320.21660.31960.19050.12700.18390.77140.56730.7740
♥MINTIAQ[88]0.70020.51600.80470.62170.44500.68260.54910.38910.65840.53780.37750.55740.40880.28160.40410.28860.19320.2959
♥MINTIAQ*[88]0.83120.64740.89740.81770.63060.85110.79080.59910.86670.89040.70920.88160.85240.66970.87550.83910.64130.8318
F-Eval (Ours)0.84860.66700.90850.83120.65850.85780.84710.66370.91060.94620.79610.94610.91880.76400.93220.94600.79590.9457
+ +where $A\in \mathbb{R}^{r\times d_i}$ and $B\in \mathbb{R}^{d_o\times r}$ are the low-rank matrices, $r\ll \min (d_o,d_i)$ is the rank, and $\alpha$ is scaling factor. For prompt of certain dimension, the corresponding expert is activated based on the dimension ID $k$ , which can be formulated as: + +$$ +E _ {k} (x) = \frac {\alpha_ {k}}{r _ {k}} B _ {k} A _ {k} x, \tag {5} +$$ + +where $E_{k}(\cdot)$ is the k-th LoRA expert. + +# 6. Experiments + +# 6.1. Implementation Details + +The LLM used in F-Eval is Qwen-2.5VL [11], with an input token channel dimension of 3584. The vision encoder, face encoder, and LLM are all frozen during training. The ranks of LoRA experts are 16,32,16,16, respectively. The LoRA alphas are set equal to LoRA ranks. LoRA dropout is set as 0.05. The model is trained on 2 NVIDIA GeForce RTX 3090 GPUs with flash-attn [21] enabled. We use AdamW Optimizer to train the network for 10 epochs with a batch size of 4 and learning rate $1e^{-4}$ . + +# 6.2. Comparison with existing QA Models + +We conduct comprehensive experiments on the proposed FaceQ database to evaluate the effectiveness of both current quality assessment methods and our proposed methods, as shown in Tab.2, Tab.3, and Tab.4. The evaluation metrics are Spearman rank-order correlation coefficient (SRCC), Kendall rank order correlation coefficient (KRCC), and Pearson linear correlation coefficient (PLCC), which assess the correlation between predicted scores by the QA + +methods and the MOS scores provided by humans. Higher metrics indicated better alignment with human preferences. The comparison QA models falls into four categories: 1) Traditional IQA Models. NR-IQA models include NIQE [67], ILNIQE [111], HOSA [102], BPRI-PSS [64], FISBLIM [27]. FR-IQA models include LPIPS [112], SSIM [96] and PSNR [34]. 2) Classical Deep learning-based IQA Models. This category includes CNNIQA [42], VGG16 [79], VGG19 [79], ResNet18 [29], ResNet34 [29], Hyper-IQA [81], TReS [25], MANIQA [106], PromptIQA[14]. All models are retrained on FaceQ with an 80:20 training/testing split. 3) AIGC IQA Models. This category includes pre-trained vision-language models, including CLIP-Score [32], BLIPScore [53], zero-shot metric ImageReward [103], and the AIGC-specific QA model MINTIQA [88]. 4) Face IQA Models. This category includes SER-FIQ [83], DSL-FIQA [18], ArcFace [22], IFQA[41]. + +# 6.3. Performance Analysis + +Traditional IQA Models. Traditional IQA methods, including both NR-IQA and FR-IQA models, generally perform poorly across all tasks and dimensions, with low SRCC, KRCC, and PLCC scores. + +Classical Deep Learning-based IQA Models. After being trained on our dataset, deep learning-based IQA models, such as MANIQA [106] and HyperIQA [81], significantly outperform traditional models in terms of quality and authenticity. However, these methods face challenges in text understanding, leading to unsatisfactory correlations. + +AIGC IQA Models. CLIPScore [32] BLIPScore [53], and + +Table 3. Performance comparison on FaceQ-Res synthetic subset. $\spadesuit$ , $\heartsuit$ , $\clubsuit$ , and $\diamond$ denote traditional NR-IQA models, FR-IQA models, face IQA models, and classical deep learning-based IQA models respectively. * indicates fine-tuned results. + +
Dimension ModelQualityID Fidelity
SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑
▲BMPRI[65]0.21610.14280.18620.11130.07390.1069
▲NIQE[67]0.17980.12130.16420.21930.14800.2118
▲ILNIQE[111]0.25050.17270.31800.26960.18570.3321
▲HOSA[102]0.37580.25220.37480.34180.34550.2301
▲BPII-LSSn [64]0.13980.09570.09990.12440.08370.0973
▲BPII [64]0.16370.10650.14800.01950.01210.0281
▲FISBLIM [27]0.15890.10890.20380.10540.07090.1715
\( \diamondsuit \)LPIPS[112]0.39090.26580.46430.43720.30180.4870
\( \diamondsuit \)SSIM[96]0.10170.06750.04880.09330.06010.0504
\( \diamondsuit \)PSNR [34]0.01980.01130.05360.09860.06620.1212
\( \diamondsuit \)ArcFace[22]0.34660.23270.37740.57280.40300.5762
\( \diamondsuit \)SER-FIQ [25]0.38130.25840.37460.39510.26410.3707
\( \diamondsuit \)DSL-FIQA [18]0.63870.45080.65280.48820.33600.5379
\( \diamondsuit \)IFQA [41]0.22730.15270.33480.32140.21850.3874
\( \diamondsuit \)CNNIA* [42]0.46610.31970.38360.46780.32160.4288
\( \diamondsuit \)ResNet18* [29]0.72540.52950.69820.66620.48020.6527
\( \diamondsuit \)ResNet34* [29]0.66930.49160.65500.59050.41700.6084
\( \diamondsuit \)VGG16* [79]0.74870.55130.71010.61770.44130.6381
\( \diamondsuit \)VGG19* [79]0.69940.51260.67220.60380.42780.5999
\( \diamondsuit \)HyperIA* [81]0.82690.63230.82280.76150.55850.7471
\( \diamondsuit \)MANIA* [106]0.82870.64030.87780.77040.56600.7619
\( \diamondsuit \)TReS* [25]0.86560.68300.86220.79510.58790.7829
\( \diamondsuit \)PromptIA* [14]0.85190.66610.87330.82200.62750.8254
F-Eval (Ours)0.86920.68550.90090.85070.67310.8726
+ +Table 4. Performance comparison on FaceQ-Res real-world subset. $\spadesuit$ , $\clubsuit$ , and $\diamond$ denote traditional IQA models, face IQA models, and classical deep learning-based IQA models respectively. * indicates fine-tuned results. + +
Dimension ModelQualityID Fidelity
SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑
▲BMPRII[65]0.15820.10930.27990.01490.01080.1386
▲NIQE[67]0.17830.11860.19620.12450.08100.1701
▲ILNIQE[111]0.22300.16100.22910.19640.13620.2340
▲HOSA[102]0.39270.42770.26540.33280.22340.3813
▲BPRI-LSSn [64]0.27150.18400.30220.20590.13800.2475
▲BPRI [64]0.19200.13130.22000.08460.05700.1136
▲FISBLIM [27]0.25120.17050.27250.23410.15810.2699
▲ArcFace[22]0.23710.15650.41830.49560.34700.6330
▲SER-FIQ [25]-0.0482-0.02990.10230.04210.02990.1584
▲DSL-FIQA [18]0.65090.46410.64940.48400.33620.5118
▲IFQA [41]0.02510.02100.12210.08930.06030.2086
◇CNNIQA* [42]0.36500.25060.47050.32520.21970.3112
◇VGG16* [79]0.49960.34390.42690.41170.28590.4622
◇VGG19* [79]0.58710.41810.51330.30540.20440.3594
◇ResNet18* [29]0.59770.42210.57220.55540.39090.6161
◇ResNet34* [29]0.62590.44750.61680.50360.35390.6136
◇HyperiQA* [81]0.79260.59650.79260.77830.58450.8163
◇MANIQA* [106]0.83560.64820.84670.75970.56550.8081
◇TReS* [25]0.82870.63910.85010.72870.54410.8009
◇PromptIQA* [14]0.81150.61860.86450.73980.55260.8297
F-Eval(Ours)0.84480.65770.87050.79570.60570.8366
+ +ImageReward [103] exhibit weak correlations with human preference, likely due to their focus on high-level semantics rather than low-level details crucial for image quality assessment. MINTIQA [88] achieves best performance among current AIGC IQA methods, but its performance is still unsatisfactory, particularly on the customization task. Fine-tuning on our dataset leads to significant improvements, highlighting the great potential of FaceQ. + +Face IQA Models. ArcFace similarity [22] demonstrates its ability to extract identity-related features. However, it lacks perceptual sensitivity for image quality, leading to poor performance in quality assessment. DSL-FIQA [18], as the latest method, achieves the best performance in predicting quality scores among face IQA methods, but it is + +still far from satisfactory. SER-FIQ fails to align with human perception across most dimensions and tasks. Experiments confirm that these natural face evaluation algorithms struggle to assess the complex characteristics of AIGFs, highlighting the need for an effective AIGF evaluator. + +F-Eval (Ours). Our method achieves optimal performance across all tasks and dimensions, with the most significant gain in the customization task, as existing methods rarely focus human preferences in face customization. Among the four dimensions, the improvements in correspondence and ID fidelity stand out, thanks to dual-stream vision encoding and the powerful LLM backbone for text understanding and modality fusion. The MoLE technique allows the model to learn distinct features for each dimension. + +# 6.4. Ablation Study + +We conducted the ablation study to validate the utility of the core components of F-Eval in Tab. 5. + +Effectiveness of Vision Encoding. 1) The effectiveness of the face encoder is demonstrated by comparing the second and last rows. Without the face encoder, a significant performance drop is observed, indicating that the extracted features help the LLM interpret face images. 2) Removing the projector leads to a dramatic performance drop. + +Effectiveness of MoLE. We conducted three sets of LoRA adaptation experiments: fine-tuning a single LoRA for all dimensions and tasks, fine-tuning three LoRA experts for three tasks, and fine-tuning four LoRA experts for four dimensions. The results show that fine-tuning multiple dimensions with a single LoRA leads to performance deterioration, which further confirms the independence of the characteristics across different dimensions. + +Table 5. Ablation study of the proposed F-Eval method. Here only present SRCC due to space limitations. + +
Feature & StrategyFace GenerationFace CustomizationFace Restoration synFace Restoration rw
VisionProjectorFaceSingle LoRATask LoRADim. LoRAQualityAuthen.Corres.QualityID.Corres.QualityID.Corres.QualityID.QualityID.
✓✓✓0.8080.7500.8460.9240.8430.9270.7770.7760.7760.7070.748
✓✓✓0.8150.7520.8350.9240.8550.9210.8120.7830.7160.603
✓✓✓✓✓0.7930.7760.8190.9410.8850.9420.8380.8230.8020.671
✓✓✓✓0.8230.7800.8380.9420.8730.9420.8240.8260.7540.638
✓✓✓✓0.8490.8310.8470.9460.9190.9460.8690.8510.8450.796
+ +# 7. Conclusion + +In this work, we conduct a pioneering generative face image quality assessment study. Specifically, we introduce FaceQ, the first generated face quality assessment database comprising 12,255 face images and 491,130 human preference ratings. Then, we establish F-Bench, a benchmark for evaluating the capabilities of existing face generation, customization, and restoration models, while also reveals the limitations of current Image Quality Assessment (IQA) and Face Quality Assessment (FQA) methods. Furthermore, we propose an evaluation model F-Eval to accurately predict human preferences on generated face images, which demonstrates state-of-the-art performance. We believe this work will contribute to advancing generative models and help overcome current limitations in AIGC community. + +# Acknowledgment + +This work is supported by the National Natural Science Foundation of China under Grants 62271308, 62401365, 62225112, 62132006, U24A20220, STCSM under Grants 24ZR1432000, 24511106902, 24511106900, 22511105700, 22DZ2229005, 111 plan under Grant Number BP0719010, China Postdoctoral Science Foundation under Grants BX20250411, 2025M773473, and State Key Laboratory of UHD Video and Audio Production and Presentation. + +# References + +[1] Deep Floyd. https://github.com/deep-floyd/IF, Accessed: 2024-10-03. 2, 3 +[2] Dreamlike V2. https://huggingface.co/dreamlike-art/dreamlike-photoreal-2.0, Accessed: 2024-10-03.5 +[3] Flux-dev. https://huggingface.co/black-forest-labs/FLUX.1-dev, Accessed: 2024-10-03.1,3,5 +[4] IP-Adapter-FaceID-SDXL. https://huggingface.co/h94/IP-Adapter, Accessed: 2024-10-03. 3, 5 +[5] IP-Adapter-FaceID-Plus. https://huggingface.co/h94/IP-Adapter-FaceID, Accessed: 2024-10-03.3,5 +[6] Kcolors. https://huggingface.co/Kwai-Kcolors/Kcolors, Accessed: 2024-10-03. 3 +[7] ProtoVision V6.6. https://huggingface.co/stablediffusionapi/protovision-xl-v6.6, Accessed: 2024-10-03. 3 +[8] Realistic Vision V5.1. https://huggingface.co/SG161222/Realistic_Vision_V5.1_noVAE, Accessed: 2024-10-03. 3, 5 +[9] SD3. https://huggingface.co/stabilityai/stable-diffusion-3-medium, Accessed: 2024-10-03. 3,5 +[10] GPT-4o. https://chatgpt.com/, Accessed: 2024-10-10.3 +[11] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. 2, 6, 7 +[12] Fadi Boutros, Meiling Fang, Marcel Klemt, Biying Fu, and Naser Damer. Cr-fiqa: face image quality assessment by learning sample relative classifiability. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5836-5845, 2023. 2 +[13] Nicolas Chahine, Stefania Calarasanu, Davide Garcia-Civiero, Theo Cayla, Sira Ferradans, and Jean Ponce. An image quality assessment dataset for portraits. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9968-9978, 2023. 2 +[14] Chen et al. Promptia: Boosting the performance and generalization for no-reference image quality assessment via prompts. In ECCV, 2024. 7, 8 + +[15] Chaofeng Chen, Dihong Gong, Hao Wang, Zhifeng Li, and Kwan-Yee K Wong. Learning spatial attention for face super-resolution. IEEE Transactions on Image Processing (TIP), 30:1219-1231, 2020. 2, 3, 5 +[16] Chaofeng Chen, Xiaoming Li, Lingbo Yang, Xianhui Lin, Lei Zhang, and Kwan-Yee K Wong. Progressive semantic-aware style transformation for blind face restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11896-11905, 2021. 2 +[17] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arxiv:2310.00426, 2023. 2, 3 +[18] Wei-Ting Chen, Gurunandan Krishnan, Qiang Gao, SyYen Kuo, Sizhou Ma, and Jian Wang. Dsl-fqa: Assessing facial image quality via dual-set degradation learning and landmark-guided transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2931-2941, 2024. 2, 7, 8 +[19] Yu Chen, Ying Tai, Xiaoming Liu, Chunhua Shen, and Jian Yang. Fsrnet: End-to-end learning face super-resolution with facial priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2492-2501, 2018. 2 +[20] Siying Cui, Jiankang Deng, Jia Guo, Xiang An, Yongle Zhao, Xinyu Wei, and Ziyong Feng. Idapter: Learning mixed features for tuning-free personalization of text-to-image models. arXiv preprint arXiv:2403.13535, 2024. 2 +[21] Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning. In Proceedings of the International Conference on Learning Representations (ICLR). 7 +[22] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4690-4699, 2019. 6, 7, 8 +[23] Huiyu Duan, Qiang Hu, Jiarui Wang, Liu Yang, Zitong Xu, Lu Liu, Xiongkuo Min, Chunlei Cai, Tianxiao Ye, Xiaoyun Zhang, et al. Fine-grained user generated content video quality assessment. In m, 2025. 2, 3 +[24] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2 +[25] S. Alireza Golestaneh, Saba Dadsetan, and Kris M. Kitani. No-reference image quality assessment via transformers, relative ranking, and self-consistency. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1220-1230, 2022. 7, 8 +[26] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Pro + +ceedings of the Advances in Neural Information Processing Systems, 2014. 1, 2 +[27] Ke Gu, Guangtao Zhai, Min Liu, Xiaokang Yang, Wenjun Zhang, Xianghui Sun, Wanhong Chen, and Ying Zuo. Fisblim: A five-step blind metric for quality assessment of multiply distorted images. In Proceedings of the SiPS, pages 241-246, 2013. 7, 8 +[28] Yuchao Gu, Xintao Wang, Liangbin Xie, Chao Dong, Gen Li, Ying Shan, and Ming-Ming Cheng. Vqfr: Blind face restoration with vector-quantized dictionary and parallel decoder. arXiv preprint arXiv:2205.06803, 2022. 2, 3 +[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 7, 8 +[30] Xuanhua He, Quande Liu, Shengju Qian, Xin Wang, Tao Hu, Ke Cao, Keyu Yan, Man Zhou, and Jie Zhang. Id-animator: Zero-shot identity-preserving human video generation. arXiv preprint arXiv:2404.15275, 2024. 2 +[31] Javier Hernandez-Ortega, Javier Galbally, Julian Fierrez, Rudolf Haraksim, and Laurent Beslay. Faceqnet: Quality assessment for face recognition based on deep learning. In 2019 International Conference on Biometrics (ICB), pages 1–8. IEEE, 2019. 2 +[32] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 7 +[33] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 33: 6840-6851, 2020. 1 +[34] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssim. In Proceedings of the IEEE International Conference on Pattern Recognition (ICPR), pages 2366-2369. IEEE, 2010. 7, 8 +[35] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. 2 +[36] Qiang Hu, Qihan He, Houqiang Zhong, Guo Lu, Xiaoyun Zhang, Guangtao Zhai, and Yanfeng Wang. Varfvv: View-adaptive real-time interactive free-view video streaming with edge computing. IEEE Journal on Selected Areas in Communications, 2025. 2 +[37] Qiang Hu, Zihan Zheng, Houqiang Zhong, Sihua Fu, Li Song, Xiaoyun Zhang, Guangtao Zhai, and Yanfeng Wang. 4dgc: Rate-aware 4d gaussian compression for efficient streamable free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 875-885, 2025. 2 +[38] Gary B. Huang, Manu Ramesh, Tamara Berg, and Erik Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. Technical Report 07-49, University of Massachusetts, Amherst, 2007. 3 + +[39] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 36:78723-78747, 2023. 2 +[40] Haoyun Jiang, Fei Huang, Qiang Hu, Minmin Sun, Shuai Xiao, Yong Li, Junyang Lin, Jiangchao Yao, et al. Catekv: On sequential consistency for long-context llm inference acceleration. In Proceedings of International Conference on Machine Learning (ICML), 2025. 3 +[41] Byungho Jo, Donghyeon Cho, In Kyu Park, and Sungeun Hong. Ifqa: Interpretable face quality assessment. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 3444-3453, 2023. 7, 8 +[42] Le Kang, Peng Ye, Yi Li, and David Doermann. Convolutional neural networks for no-reference image quality assessment. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014. 7, 8 +[43] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 4401-4410, 2019. 2 +[44] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6007-6017, 2023. 1 +[45] Ira Kemelmacher-Shlizerman, Steven M Seitz, Daniel Miller, and Evan Brossard. The megaface benchmark: 1 million faces for recognition at scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4873-4882, 2016. 3 +[46] Deokyun Kim, Minseon Kim, Gihyun Kwon, and Dae-Shik Kim. Progressive face super-resolution via attention to facial landmark. arXiv preprint arXiv:1908.08239, 2019. 2 +[47] Diederik P Kingma, Max Welling, et al. Auto-encoding variational bayes. 2013. 1 +[48] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Matiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. Proceedings of Advances in neural information processing systems (Neurips), 36:36652-36663, 2023. 1, 3 +[49] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1931-1941, 2023. 2 +[50] Li et al. Agiqa-3k: An open database for ai-generated image quality assessment. arXiv preprint arXiv:2306.04717, 2023. 2 +[51] Chunyi Li, Zicheng Zhang, Haoning Wu, Wei Sun, Xiongkuo Min, Xiaohong Liu, Guangtao Zhai, and Weisi Lin. Agiaq-3k: An open database for ai-generated image quality assessment. arXiv preprint arXiv: 2306.04717, 2023. 1, 2, 3 + +[52] Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Linmiao Xu, and Suhail Doshi. Playground v2.5: Three insights towards enhancing aesthetic quality in text-to-image generation, 2024. 3 +[53] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proceedings of International Conference on Machine learning (ICML), pages 12888-12900, 2022. 7 +[54] Xiaoming Li, Ming Liu, Yuting Ye, Wangmeng Zuo, Liang Lin, and Ruigang Yang. Learning warped guidance for blind face restoration. In The European Conference on Computer Vision (ECCV), 2018. 2 +[55] Xiaoming Li, Wenyu Li, Dongwei Ren, Hongzhi Zhang, Meng Wang, and Wangmeng Zuo. Enhanced blind face restoration with multi-exemplar images and adaptive spatial feature fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2706-2715, 2020. 2 +[56] Xiaoming Li, Shiguang Zhang, Shangchen Zhou, Lei Zhang, and Wangmeng Zuo. Learning dual memory dictionaries for blind face restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2022. 3 +[57] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, Ming-Ming Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. arXiv preprint arXiv:2312.04461, 2023. 1, 2, 3, 5 +[58] Zhimin Li, Jianwei Zhang, Qin Lin, Jiangfeng Xiong, Yanxin Long, Xinchi Deng, Yingfang Zhang, Xingchao Liu, Minbin Huang, Zedong Xiao, et al. Hunyuan-dit: A powerful multi-resolution diffusion transformer with fine-grained chinese understanding. arXiv preprint arXiv:2405.08748, 2024. 2, 3, 5 +[59] Youwei Liang, Junfeng He, Gang Li, Peizhao Li, Arseniy Klimovskiy, Nicholas Carolan, Jiao Sun, Jordi Pont-Tuset, Sarah Young, Feng Yang, et al. Rich human feedback for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19401-19411, 2024. 1, 2 +[60] Zhang Lijun, Shao Xiaohu, Yang Fei, Deng Pingling, Zhou Xiangdong, and Shi Yu. Multi-branch face quality assessment for face recognition. In 2019 IEEE 19th International Conference on Communication Technology (ICCT), pages 1659–1664. IEEE, 2019. 2 +[61] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of European conference on computer vision (ECCV), pages 740-755. Springer, 2014. 3 +[62] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Ben Fei, Bo Dai, Wanli Ouyang, Yu Qiao, and Chao Dong. Diff-bir: Towards blind image restoration with generative diffusion prior. arXiv preprint arXiv:2308.15070, 2023. 1, 2, 3, 5 +[63] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Proceedings of + +the Advances in Neural Information Processing Systems (NeurIPS), pages 34892-34916, 2023. 2 +[64] Xiongkuo Min, Ke Gu, Guangtao Zhai, Jing Liu, Xiaokang Yang, and Chang Wen Chen. Blind quality assessment based on pseudo-reference image. IEEE Transactions on Multimedia (TMM), 20(8):2049-2062, 2018. 7, 8 +[65] Xiongkuo Min, Guangtao Zhai, Ke Gu, Yutao Liu, and Xiaokang Yang. Blind image quality estimation via distortion aggravation. IEEE Transactions on Broadcasting, 64(2): 508-517, 2018. 8 +[66] Xiongkuo Min, Huiyu Duan, Wei Sun, Yucheng Zhu, and Guangtao Zhai. Perceptual video quality assessment: A survey. Science China Information Sciences, 67(11): 211301, 2024. 2 +[67] Anish Mittal, Rajiv Soundararajan, and Alan C. Bovik. Making a "completely blind" image quality analyzer. IEEE Signal Processing Letters, 20(3):209-212, 2013. 7, 8 +[68] Stylianos Moschoglou, Athanasios Papaioannou, Christos Sagonas, Jiankang Deng, Irene Kotsia, and Stefanos Zafeiriou. Agedb: the first manually collected, in-the-wild age database. In proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 51-59, 2017. 3 +[69] Fu-Zhao Ou, Xingyu Chen, Ruixin Zhang, Yuge Huang, Shaoxin Li, Jilin Li, Yong Li, Liujuan Cao, and Yuan-Gen Wang. Sdd-fiqa: Unsupervised face image quality assessment with similarity distribution distance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7670–7679, 2021. 2 +[70] Pablo Pernias, Dominic Rampas, Mats L Richter, Christopher J Pal, and Marc Aubreville. Würstchen: An efficient architecture for large-scale text-to-image diffusion models. arXiv preprint arXiv:2306.00637, 2023. 3 +[71] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 3, 5 +[72] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation, 2021. 2 +[73] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2 +[74] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 2, 3 +[75] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22500-22510, 2023. 2 +[76] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed + +Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding, 2022. 2 +[77] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 35:36479-36494, 2022. 2 +[78] B Series. Methodology for the subjective assessment of the quality of television pictures. Recommendation ITU-R BT, 500(13), 2012. 4 +[79] Karen Simonyan and Andrew Zisserman. Very Deep Convolutional Networks for Large-Scale Image Recognition. arXiv e-prints, art. arXiv:1409.1556, 2014. 7, 8 +[80] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv:2010.02502, 2020. 1 +[81] Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3664-3673, 2020. 7, 8 +[82] Shaolin Su, Hanhe Lin, Vlad Hosu, Oliver Wiedemann, Jinqiu Sun, Yu Zhu, Hantao Liu, Yanning Zhang, and Dietmar Saupe. Going the extra mile in face image quality assessment: A novel database and model. IEEE Transactions on Multimedia, 2023. 2, 3 +[83] Philipp Terhorst, Jan Niklas Kolf, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. Ser-fiq: Unsupervised estimation of face image quality based on stochastic embedding robustness. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5651-5660, 2020. 7 +[84] Yuan Tian, Kaiyuan Ji, Rongzhao Zhang, Yankai Jiang, Chunyi Li, Xiaosong Wang, and Guangtao Zhai. Towards all-in-one medical image re-identification. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 30774–30786, 2025. 2 +[85] Yuri Viazovetskyi, Vladimir Ivashkin, and Evgeny Kashin. Stylegan2 distillation for feed-forward image manipulation. Proceedings of the European conference on computer vision (ECCV), pages 170-186, 2020. 2 +[86] Jiarui Wang, Huiyu Duan, Jing Liu, Shi Chen, Xiongkuo Min, and Guangtao Zhai. Aigciqa2023: A large-scale image quality assessment database for ai generated images: from the perspectives of quality, authenticity and correspondence. 2023. 1, 2, 3 +[87] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. arXiv preprint arXiv:2305.07015, 2023. 1, 2, 3 +[88] Jiarui Wang, Huiyu Duan, Guangtao Zhai, and Xiongkuo Min. Understanding and evaluating human preferences for ai generated images with instruction tuning. arXiv preprint arXiv:2405.07346, 2024. 7, 8 + +[89] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin C.K. Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. 2024. 5 +[90] Liangbin Xie. Xintao Wang, Honglun Zhang, Chao Dong, and Ying Shan. Vfhq: A high-quality dataset and benchmark for video face super-resolution, 2022. 3 +[91] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 2, 3, 5 +[92] Tao Wang, Kaihao Zhang, Xuanxi Chen, Wenhan Luo, Jiankang Deng, Tong Lu, Xiaochun Cao, Wei Liu, Hongdong Li, and Stefanos Zafeiriou. A survey of deep face restoration: Denoise, super-resolution, deblur, artifact removal. arXiv preprint arXiv:2211.02831, 2022. 2 +[93] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In International Conference on Computer Vision Workshops (ICCVW). 3 +[94] Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. Towards real-world blind face restoration with generative facial prior. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 3, 5 +[95] Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. Towards real-world blind face restoration with generative facial prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3 +[96] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing (TIP), 13(4):600-612, 2004. 7, 8 +[97] Zhixin Wang, Ziying Zhang, Xiaoyun Zhang, Huangjie Zheng, Mingyuan Zhou, Ya Zhang, and Yanfeng Wang. Dr2: Diffusion-based robust degradation remover for blind face restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1704-1713, 2023. 2, 3, 5 +[98] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15943-15953, 2023. 2, 3, 5 +[99] Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xinglong Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. Diffir: Efficient diffusion model for image restoration. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13095-13105, 2023. 1 +[100] Weihao Xia, Yujiu Yang, Jing-Hao Xue, and Baoyuan Wu. Tedigan: Text-guided diverse face image generation and manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2256-2265, 2021. 1 +[101] Guangxuan Xiao, Tianwei Yin, William T Freeman, Frédo Durand, and Song Han. Fastcomposer: Tuning-free multisubject image generation with localized attention. Inter- + +national Journal of Computer Vision (IJCV), 133(3):1175-1194, 2025. 2, 3, 5 +[102] Jingtao Xu, Peng Ye, Qiaohong Li, Haiqing Du, Yong Liu, and David S. Doermann. Blind image quality assessment based on high order statistics aggregation. IEEE Transactions on Image Processing (TIP), 25:4444-4457, 2016. 7, 8 +[103] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 36, 2024. 2, 7, 8 +[104] Liu Yang, Huiyu Duan, Long Teng, Yucheng Zhu, Xiaohong Liu, Menghan Hu, Xiongkuo Min, Guangtao Zhai, and Patrick Le Callet. Aigcoiaq2024: Perceptual quality assessment of ai generated omnidirectional images. In 2024 IEEE International Conference on Image Processing (ICIP), pages 1239-1245. IEEE, 2024. 3 +[105] Shuo Yang, Ping Luo, Chen-Change Loy, and Xiaou Tang. Wider face: A face detection benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5525–5533, 2016. 3 +[106] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. MANIQA: Multi-dimension Attention Network for No-Reference Image Quality Assessment. arXiv e-prints, art. arXiv:2204.08958, 2022. 6, 7, 8 +[107] Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. Gan prior embedded network for blind face restoration in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 672-681, 2021. 2, 3 +[108] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 1, 2 +[109] Zongsheng Yue and Chen Change Loy. Difface: Blind face restoration with diffused error contraction. arXiv preprint arXiv:2212.06512, 2022. 2, 3 +[110] Guangtao Zhai and Xiongkuo Min. Perceptual image quality assessment: a survey. Science China Information Sciences, 63:1-52, 2020. 2 +[111] Lin Zhang, Lei Zhang, and Alan C. Bovik. A feature-enriched completely blind image quality evaluator. IEEE Transactions on Image Processing (TIP), 24(8):2579-2591, 2015. 7, 8 +[112] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 6, 7, 8 +[113] Shangchen Zhou, Kelvin Chan, Chongyi Li, and Chen Change Loy. Towards robust blind face restoration with codebook lookup transformer. Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 35:30599-30611, 2022. 2, 3, 5 \ No newline at end of file diff --git a/2025/F-Bench_ Rethinking Human Preference Evaluation Metrics for Benchmarking Face Generation, Customization, and Restoration/images.zip b/2025/F-Bench_ Rethinking Human Preference Evaluation Metrics for Benchmarking Face Generation, Customization, and Restoration/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f31997dfec4a5fb94e15d9cd30e149c65f53b1f4 --- /dev/null +++ b/2025/F-Bench_ Rethinking Human Preference Evaluation Metrics for Benchmarking Face Generation, Customization, and Restoration/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b6565ed2a814517d6ce9d9bbbcbc547a72cea67bc989e8ff04a2756ebdbbadf6 +size 1051504 diff --git a/2025/F-Bench_ Rethinking Human Preference Evaluation Metrics for Benchmarking Face Generation, Customization, and Restoration/layout.json b/2025/F-Bench_ Rethinking Human Preference Evaluation Metrics for Benchmarking Face Generation, Customization, and Restoration/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..069f38b370857fb87e3c08851cb40986ede79247 --- /dev/null +++ b/2025/F-Bench_ Rethinking Human Preference Evaluation Metrics for Benchmarking Face Generation, Customization, and Restoration/layout.json @@ -0,0 +1,10773 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 61, + 102, + 550, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 102, + 550, + 138 + ], + "spans": [ + { + "bbox": [ + 61, + 102, + 550, + 138 + ], + "type": "text", + "content": "F-Bench: Rethinking Human Preference Evaluation Metrics for Benchmarking Face Generation, Customization, and Restoration" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 148, + 160, + 460, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 160, + 460, + 175 + ], + "spans": [ + { + "bbox": [ + 148, + 160, + 460, + 175 + ], + "type": "text", + "content": "Lu Liu" + }, + { + "bbox": [ + 148, + 160, + 460, + 175 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 148, + 160, + 460, + 175 + ], + "type": "text", + "content": ", Huiyu Duan" + }, + { + "bbox": [ + 148, + 160, + 460, + 175 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 148, + 160, + 460, + 175 + ], + "type": "text", + "content": ", Qiang Hu" + }, + { + "bbox": [ + 148, + 160, + 460, + 175 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + }, + { + "bbox": [ + 148, + 160, + 460, + 175 + ], + "type": "text", + "content": ", Liu Yang" + }, + { + "bbox": [ + 148, + 160, + 460, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 148, + 160, + 460, + 175 + ], + "type": "text", + "content": ", Chunlei Cai" + }, + { + "bbox": [ + 148, + 160, + 460, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 157, + 175, + 452, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 175, + 452, + 189 + ], + "spans": [ + { + "bbox": [ + 157, + 175, + 452, + 189 + ], + "type": "text", + "content": "Tianxiao Ye" + }, + { + "bbox": [ + 157, + 175, + 452, + 189 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 157, + 175, + 452, + 189 + ], + "type": "text", + "content": ", Huayu Liu" + }, + { + "bbox": [ + 157, + 175, + 452, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 157, + 175, + 452, + 189 + ], + "type": "text", + "content": ", Xiaoyun Zhang" + }, + { + "bbox": [ + 157, + 175, + 452, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 157, + 175, + 452, + 189 + ], + "type": "text", + "content": ", Guangtao Zhai" + }, + { + "bbox": [ + 157, + 175, + 452, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 114, + 190, + 497, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 190, + 497, + 204 + ], + "spans": [ + { + "bbox": [ + 114, + 190, + 497, + 204 + ], + "type": "text", + "content": "1Shanghai Jiao Tong University, Shanghai, China 2Bilibili Inc., Shanghai, China" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 58, + 211, + 202, + 383 + ], + "blocks": [ + { + "bbox": [ + 58, + 211, + 202, + 383 + ], + "lines": [ + { + "bbox": [ + 58, + 211, + 202, + 383 + ], + "spans": [ + { + "bbox": [ + 58, + 211, + 202, + 383 + ], + "type": "image", + "image_path": "be54ea568985a727a2cc4028c6a24db27d8ed7043d43028000e0e447b132f563.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 384, + 555, + 426 + ], + "lines": [ + { + "bbox": [ + 55, + 384, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 555, + 426 + ], + "type": "text", + "content": "Figure 1. Data construction pipeline of " + }, + { + "bbox": [ + 55, + 384, + 555, + 426 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 384, + 555, + 426 + ], + "type": "text", + "content": "-bench and content overview of FaceQ database. From left to right, (a) Diverse input source data for face generation, customization, and restoration. (b) " + }, + { + "bbox": [ + 55, + 384, + 555, + 426 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 384, + 555, + 426 + ], + "type": "text", + "content": "-Bench benchmarks 29 face generative models, including 14 face generation models, 6 face customization models, and 9 face restoration models. (c) Multi-dimensional subjective preference evaluation in " + }, + { + "bbox": [ + 55, + 384, + 555, + 426 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 384, + 555, + 426 + ], + "type": "text", + "content": "-Bench. (d) 3D scatter plots of MOS distributions across FaceQ three subsets." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 202, + 212, + 407, + 380 + ], + "blocks": [ + { + "bbox": [ + 202, + 212, + 407, + 380 + ], + "lines": [ + { + "bbox": [ + 202, + 212, + 407, + 380 + ], + "spans": [ + { + "bbox": [ + 202, + 212, + 407, + 380 + ], + "type": "image", + "image_path": "50a31b7cb83ff8691ad34b7636a1481484922af6f46b1659ee084656b72f3aeb.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 407, + 212, + 553, + 382 + ], + "blocks": [ + { + "bbox": [ + 407, + 212, + 553, + 382 + ], + "lines": [ + { + "bbox": [ + 407, + 212, + 553, + 382 + ], + "spans": [ + { + "bbox": [ + 407, + 212, + 553, + 382 + ], + "type": "image", + "image_path": "cf85e2bc4cae21c3c7775c8426f5183047677a9fc1bbce0840f052dde717a00a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 151, + 435, + 200, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 435, + 200, + 448 + ], + "spans": [ + { + "bbox": [ + 151, + 435, + 200, + 448 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 54, + 462, + 297, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 462, + 297, + 689 + ], + "spans": [ + { + "bbox": [ + 54, + 462, + 297, + 689 + ], + "type": "text", + "content": "Recent artificial intelligence (AI) generative models have demonstrated remarkable capabilities in image production, and have been widely applied to face image generation, customization, and restoration. However, many AI-generated faces (AIGFs) still suffer from issues such as unique distortions, unrealistic details, and unexpected identity shifts, underscoring the need for a comprehensive quality evaluation method for AIGFs. To this end, we introduce FaceQ, the first comprehensive AI-generated Face image database with fine-grained Quality annotations aligned with human preferences, which consists of 12K images and 491K ratings across multiple dimensions. Using the FaceQ database, we establish F-Bench, a benchmark for comparing and evaluating face generation, customization, and restoration models, highlighting strengths and weaknesses across various prompts and evaluation dimensions. Additionally, we assess the performance of existing image quality assessment (IQA) methods on FaceQ, and further propose a large multimodal model (LMM) based Face quality Evaluator (F" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 436, + 556, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 556, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 556, + 496 + ], + "type": "text", + "content": "Eval) to accurately assess the multi-dimensional quality of generated faces in a one-for-all manner. Extensive experimental results demonstrate the state-of-the-art performance of our F-Eval. The project page is: https://media-sjtu.github.io/F-Bench/." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 523, + 394, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 523, + 394, + 536 + ], + "spans": [ + { + "bbox": [ + 314, + 523, + 394, + 536 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 544, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 544, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 312, + 544, + 555, + 676 + ], + "type": "text", + "content": "Generative models, including VAE [47], GAN [26], and diffusion models [33, 74, 80], have made remarkable progress in recent years, and have been extended to the fields of text-to-image generation [3], text-driven image editing [44], image restoration [99], etc, leading to high-quality artificial intelligence generated images (AIGIs). As an important aspect of AIGIs, many studies have also investigated face generation [100], face customization [57, 108], and face restoration [62, 87], producing creative, high-quality face images. Despite these advances, many AI-generated faces (AIGFs) still suffer from perceptual deficiencies." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "Many natural face image quality assessment databases have been established in the literature [48, 51, 59, 86]. However, natural images are mainly degraded by distor" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 220, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 220, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 220, + 713 + ], + "type": "text", + "content": "* Equal contribution. † Corresponding author." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10982" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 92, + 553, + 156 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 555, + 92 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 555, + 92 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 555, + 92 + ], + "type": "text", + "content": "Table 1. Comparison between FaceQ and existing AIGC quality assessment databases (top) as well as natural face quality assessment databases (bottom). FaceQ is the first large-scale AI-generated, customized and restored face quality assessment database." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 92, + 553, + 156 + ], + "lines": [ + { + "bbox": [ + 55, + 92, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 92, + 553, + 156 + ], + "type": "table", + "html": "
DatabaseDomainSourceImagesScoresDimensions
AGIQA-3K (TCSVT2024) [51]GeneralAI-generated2,982MOSPerception, Alignment
AIGCIQA (CVPR2024) [86]GeneralAI-generated2,400MOSQuality, Authenticity, Correspondence
RichHF-18K (CVPR2024) [59]GeneralAI-generated17,760MOSPlausibility, Aesthetics, Text-image Alignment, Overall
PIQ23 (CVPR2023) [13]FaceReal5,116PairOverall, Details, Exposure
FaceQFaceAI-generated, customized and restored12,255MOSQuality, Authenticity, ID Fidelity, Correspondence
", + "image_path": "242cd0ee9f6c552fbdf553649873ffa2f11f98908744e1f2698cbb8ec4ab3d82.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 158, + 294, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 158, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 54, + 158, + 294, + 326 + ], + "type": "text", + "content": "tions such as noise, blur, compression [23, 66, 110], etc., while AIGIs generally suffer from unique distortions including unrealistic structures and text-image inconsistencies [59, 86], etc. Recently, many studies have investigated human preferences for AIGIs [39, 50, 59, 86, 103]. However, these general AIGI quality assessment works may not be well aligned with human preferences on AIGFs, since the perceptual quality of AIGFs can be significantly affected by small deviations in facial attributes,(e.g., eye size, skin texture, identity shifts), while the number of AIGFs in these AIGI databases is limited. Moreover, current AIGI quality assessment works mainly focus on the image generation, while human preferences for face customization have rarely been considered, where identity fidelity is significant." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 327, + 296, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 296, + 568 + ], + "type": "text", + "content": "The above limitation highlights the necessity of establishing a comprehensive quality evaluator for AIGFs. To this end, we propose FaceQ, the first AI-generated Face image Quality assessment database with three popular tasks. Specifically, FaceQ contains " + }, + { + "bbox": [ + 56, + 327, + 296, + 568 + ], + "type": "inline_equation", + "content": "12\\mathrm{K}+" + }, + { + "bbox": [ + 56, + 327, + 296, + 568 + ], + "type": "text", + "content": " images obtained from 29 generative models across three tasks: 1) face generation, 2) face customization, and 3) face restoration, as shown in Fig.1. Based on the generated images, we collect " + }, + { + "bbox": [ + 56, + 327, + 296, + 568 + ], + "type": "inline_equation", + "content": "491\\mathrm{K}+" + }, + { + "bbox": [ + 56, + 327, + 296, + 568 + ], + "type": "text", + "content": " quality ratings from multiple dimensions including quality, authenticity, identity (ID) fidelity, and text-image correspondence, respectively, finally forming " + }, + { + "bbox": [ + 56, + 327, + 296, + 568 + ], + "type": "inline_equation", + "content": "32\\mathrm{K}+" + }, + { + "bbox": [ + 56, + 327, + 296, + 568 + ], + "type": "text", + "content": " mean opinion scores (MOSs). Using FaceQ, we establish a human preference benchmark, i.e., F-Bench, for face generation, customization, and restoration models, and thoroughly analyze the generation performance across different evaluation perspectives, models, and categories. In addition, we also compare the correlation of existing image quality assessment (IQA), face quality assessment (FQA), AIGI quality assessment, and preference evaluation metrics with human preferences." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": "Furthermore, we present F-Eval, the first generated Face image quality Evaluation method designed to perform comprehensive multi-dimensional quality evaluation for generated, customized, and restorted faces in a one-for-all manner. Specifically, F-Eval leverages large multimodal models (LMMs) [11] and adopts instruction tuning [63] for preference-related vision-language adaption. We also apply low-rank adaptation (LoRA) [35] techniques for more efficient feature refining. Extensive experimental results demonstrate that F-Eval outperforms existing quality assessment methods in terms of multiple dimensions relevant to human preference for all three tasks. The main contribu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 158, + 494, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 158, + 494, + 171 + ], + "spans": [ + { + "bbox": [ + 313, + 158, + 494, + 171 + ], + "type": "text", + "content": "tions of our work are highlighted as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 175, + 553, + 342 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 313, + 175, + 553, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 175, + 553, + 233 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 553, + 233 + ], + "type": "text", + "content": "- We construct FaceQ, the first generated face quality assessment database consisting " + }, + { + "bbox": [ + 313, + 175, + 553, + 233 + ], + "type": "inline_equation", + "content": "491\\mathrm{K}+" + }, + { + "bbox": [ + 313, + 175, + 553, + 233 + ], + "type": "text", + "content": " human preference ratings from quality, authenticity, ID fidelity, correspondence dimensions, on " + }, + { + "bbox": [ + 313, + 175, + 553, + 233 + ], + "type": "inline_equation", + "content": "12\\mathrm{K}+" + }, + { + "bbox": [ + 313, + 175, + 553, + 233 + ], + "type": "text", + "content": " generative face images across three tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 235, + 553, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 235, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 235, + 553, + 293 + ], + "type": "text", + "content": "- We establish F-Bench, a comprehensive human preference benchmark for 29 face generation, customization, and restoration methods. It also extensively evaluates the alignment of 26 existing quality assessment methods with human preferences." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 294, + 553, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 553, + 342 + ], + "type": "text", + "content": "- We present F-Eval, a LMM-based quality assessment model to predict multi-dimensional human preference on generated faces. Extensive experiments demonstrate the state-of-art QA performance of the proposed F-Eval." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 357, + 401, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 401, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 401, + 369 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 378, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 555, + 653 + ], + "type": "text", + "content": "Face Generation & Customization & Restoration. (1) For face generation models, earlier works such as GANs [26, 43, 85] have demonstrated significant influences. Recently, diffusion models [73, 74, 77] have achieved rapid developments in text-to-image generation. These models can be categorized into pixel-space diffusion models [1, 72, 73, 76], and latent space diffusion models [17, 58, 71, 74]. (2) For face customization, this is a recently defined task that aims to create images of a specific person with the input of an identity reference image and prompts [98]. It can be divided into two categories including conventional test-time optimization [24, 49, 75] and tuning-free customization methods utilizing pre-trained diffusion models [20, 30, 57, 91, 101, 108]. (3) For face restoration, which aims to restore high-quality face images from their low-quality counterparts. Some traditional restoration methods utilize geometric priors [16, 19, 46] or reference priors [54, 55] while in recent years, exploring generative priors such as GAN priors [15, 28, 94, 107, 113] and diffusion priors [36, 37, 62, 84, 87, 97, 109] has become a prevalent trend. Despite these improvements, the generated, customized, and restored images may suffer from various quality issues, highlighting the necessity of evaluators." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "Face Image Quality Assessment Database. Existing face image quality assessment (FIQA) can be divided into Biometric FIQA [12, 31, 60, 69] and Generic FIQA [13, 18, 82, 92]. BFIQA methods are most developed to evaluate the biometric utility for face recognition. GFIQA meth" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10983" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 242, + 203 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 242, + 203 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 242, + 203 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 242, + 203 + ], + "type": "image", + "image_path": "0ca6f16b7c42b7cffbe0d97807f01593ad18168ac065d62cf16b2ebdca062ecf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 204, + 555, + 237 + ], + "lines": [ + { + "bbox": [ + 55, + 204, + 555, + 237 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 555, + 237 + ], + "type": "text", + "content": "Figure 2. Rating comparisons of eight dimensions. Each column presents a pair of intuitive examples of each dimension, with red indicating the better rating and blue indicating the worse one. From left to right, the subsets are face generation, face customization, and face restoration subsets. The last row displays the corresponding prompts, reference image-prompt pairs, and the GT-LQ image pairs." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 244, + 70, + 425, + 203 + ], + "blocks": [ + { + "bbox": [ + 244, + 70, + 425, + 203 + ], + "lines": [ + { + "bbox": [ + 244, + 70, + 425, + 203 + ], + "spans": [ + { + "bbox": [ + 244, + 70, + 425, + 203 + ], + "type": "image", + "image_path": "15cf956e90dd11acd6ae04bce2981f7b38875f0c25c21bc20e03bb64418aa45d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 428, + 70, + 554, + 203 + ], + "blocks": [ + { + "bbox": [ + 428, + 70, + 554, + 203 + ], + "lines": [ + { + "bbox": [ + 428, + 70, + 554, + 203 + ], + "spans": [ + { + "bbox": [ + 428, + 70, + 554, + 203 + ], + "type": "image", + "image_path": "959bcd90fe7fba8303ac06245c48e7d7a9e96871c01f3bfb428c7c084619b239.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 240, + 294, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 240, + 294, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 294, + 276 + ], + "type": "text", + "content": "ods [82] focus on perceptual degradation in face restoration task. Both of these databases consist of natural face images, which has distinctly different nature with generated faces." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 277, + 295, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 295, + 361 + ], + "type": "text", + "content": "AIGC Quality Assessment Database. In recent years, plenty of researchers have raised interest in assessing AI-generated images(AIGIs) quality assessment [23, 40, 48, 51, 86, 104] by building various databases as shown in Tab 1. All of these existing database contain various categories of general AIGI images, leaving a notable gap in datasets dedicated to AI-generated face images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 370, + 223, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 370, + 223, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 223, + 384 + ], + "type": "text", + "content": "3. FaceQ Database Construction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 391, + 296, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 391, + 296, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 391, + 296, + 462 + ], + "type": "text", + "content": "In this section, we introduce the proposed generative face quality assessment database FaceQ. The database includes 12,255 images and 491,130 human opinion ratings. Our FaceQ demonstrates the merits of comprehensive generated visual content and multi-dimensional labeling, which can significantly facilitate the advancement of AIGC research." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 469, + 151, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 469, + 151, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 469, + 151, + 480 + ], + "type": "text", + "content": "3.1. Data Collection" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 487, + 208, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 208, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 208, + 498 + ], + "type": "text", + "content": "3.1.1. FaceQ-Gen Subset Collection" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 502, + 295, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 295, + 561 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 295, + 561 + ], + "type": "text", + "content": "Prompt Sources. The prompts in FaceQ-Gen subset are sourced from MS-COCO [61] [57] and GPT-4o [10]. These face-centric prompts can be categorized into nine types: General, Accessories, Clothes, Action, View, Expression, Background, Style, and Facial Attributes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 562, + 295, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 562, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 562, + 295, + 658 + ], + "type": "text", + "content": "Models Collection. We collect 14 representative face generation models, including Deep Floyd [1], Stable Diffusion V1.5 [74], Stable Diffusion V2.1 [74], SD-XL [71], Realistic Vision [8], Stable Cascade [70], ProtoVision V6.6 [7], SD3 [9], Playground V2.5 [52], Kolors [6], PixArt-alpha [17], Hunyuan [58], and Flux-dev [3]. Negative prompts and safe sensors are enabled. In total, FaceQ-Gen includes 4,032 generated face images." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 662, + 206, + 674 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 662, + 206, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 662, + 206, + 674 + ], + "type": "text", + "content": "3.1.2. FaceQ-Cus Subset Collection" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Identity and Prompt Sources. Face customization relies on a reference identity image and a guiding prompt. The reference images in FaceQ-Cus are partially self-collected," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 240, + 555, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 240, + 555, + 265 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 555, + 265 + ], + "type": "text", + "content": "with the rest sourced from [57]. The prompts are selected and filtered from the FaceQ-Gen prompt list." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 266, + 555, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 555, + 338 + ], + "type": "text", + "content": "Models Collection. We collect six representative face customization methods, including ELITE [98], FastComposer [101], IP-Adapter-FaceID [4], InstantID [91], IP-Adapter-Face-ID-Plus [5], and PhotoMaker V2 [57]. Here, we use single reference image as input. In total, FaceQ-Cus includes 4,200 generated face images." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 346, + 463, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 463, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 463, + 357 + ], + "type": "text", + "content": "3.1.3. FaceQ-Res Subset Collection" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 361, + 554, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 554, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 554, + 468 + ], + "type": "text", + "content": "Low Quality Images Source. (1) Synthetic: We construct two degraded pipelines following previous works [15, 93, 95, 97] to simulate real-world degradations on both indomain data from VFHQ [90], CelebRef [56] and out-of-domain self-collected high-quality Asian face images (2) Real-world: Real-world low-quality images are sourced from various public datasets, including AgeAB [68], CelebChild [28], Wider [105], LFW [38] and MegaFace [45], ensuring the degradation diversity." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 470, + 554, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 470, + 554, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 554, + 578 + ], + "type": "text", + "content": "Models Collection and Face Image Restoration. We evaluate both diffusion-prior-based methods such as DiffBIR [62], DifFace [109], StableSR [87], and GAN or Transformer-based methods GFPGAN [94], GPEN [107], CodeFormer [113], VQFR [28] and Face-SPARNet [15]. Each method is benchmarked on both the " + }, + { + "bbox": [ + 313, + 470, + 554, + 578 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 470, + 554, + 578 + ], + "type": "text", + "content": "-Res synthetic dataset and the real-world dataset. In total, FaceQ-Res include 4023 restored images, with 2007 images in the real-world case and 2016 images in the synthetic case." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 587, + 512, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 587, + 512, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 587, + 512, + 601 + ], + "type": "text", + "content": "3.2. Human Preference Dimension Design" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 605, + 554, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 666 + ], + "type": "text", + "content": "The evaluation dimensions of generated faces need specific design to adequately reflect human preference, given the complexity of perceptual factors involved. To the end, we decompose \"quality\" into four core dimensions: Quality, Authenticity, ID Fidelity, and Correspondence." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 666, + 553, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 553, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 553, + 700 + ], + "type": "text", + "content": "Quality. This dimension evaluates the overall perceptual quality of the image, considering factors such as color accuracy, blur, noise, and artifacts. It is used in all tasks." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "type": "text", + "content": "Authenticity. This dimension evaluates how closely the" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10984" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 83, + 219, + 137 + ], + "blocks": [ + { + "bbox": [ + 58, + 83, + 219, + 137 + ], + "lines": [ + { + "bbox": [ + 58, + 83, + 219, + 137 + ], + "spans": [ + { + "bbox": [ + 58, + 83, + 219, + 137 + ], + "type": "image", + "image_path": "906cf53f1707bfcabc0ffea135230ee9f39b5e4307d42b8189ee2f896f3f2245.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 137, + 198, + 145 + ], + "lines": [ + { + "bbox": [ + 102, + 137, + 198, + 145 + ], + "spans": [ + { + "bbox": [ + 102, + 137, + 198, + 145 + ], + "type": "text", + "content": "(a) Face generation MOS distribution" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 214, + 83, + 385, + 138 + ], + "blocks": [ + { + "bbox": [ + 214, + 83, + 385, + 138 + ], + "lines": [ + { + "bbox": [ + 214, + 83, + 385, + 138 + ], + "spans": [ + { + "bbox": [ + 214, + 83, + 385, + 138 + ], + "type": "image", + "image_path": "f75d4c146079fe3a3a587bb40ab780de22b24212ea513086b7e2ba41df47cde3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 254, + 138, + 359, + 145 + ], + "lines": [ + { + "bbox": [ + 254, + 138, + 359, + 145 + ], + "spans": [ + { + "bbox": [ + 254, + 138, + 359, + 145 + ], + "type": "text", + "content": "(b) Face customization MOS distribution" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 385, + 83, + 553, + 138 + ], + "blocks": [ + { + "bbox": [ + 385, + 83, + 553, + 138 + ], + "lines": [ + { + "bbox": [ + 385, + 83, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 385, + 83, + 553, + 138 + ], + "type": "image", + "image_path": "e6fc96cb02e178976b89dc1fedc2b07e1fc527d99f6a9e0dbe615ac2b0e0470d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 424, + 138, + 521, + 145 + ], + "lines": [ + { + "bbox": [ + 424, + 138, + 521, + 145 + ], + "spans": [ + { + "bbox": [ + 424, + 138, + 521, + 145 + ], + "type": "text", + "content": "(c) Face restoration MOS distribution" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 59, + 148, + 154, + 182 + ], + "blocks": [ + { + "bbox": [ + 59, + 148, + 154, + 182 + ], + "lines": [ + { + "bbox": [ + 59, + 148, + 154, + 182 + ], + "spans": [ + { + "bbox": [ + 59, + 148, + 154, + 182 + ], + "type": "image", + "image_path": "3c50297297cdc84bc438869a8f6c868263c2dfd6249d458e8a5464371dc5356b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 182, + 129, + 190 + ], + "lines": [ + { + "bbox": [ + 85, + 182, + 129, + 190 + ], + "spans": [ + { + "bbox": [ + 85, + 182, + 129, + 190 + ], + "type": "text", + "content": "(d) Hunyuan-DiT" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 191, + 553, + 213 + ], + "lines": [ + { + "bbox": [ + 55, + 191, + 553, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 191, + 553, + 213 + ], + "type": "text", + "content": "Figure 3. Mean Opinion Score (MOS) distribution histograms and kernel density curves. (a)-(c) MOS distributions for three subsets. (d)-(h) Model-wise MOS distributions. MOS Distributions of other methods are provided in the supplementary material." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 154, + 148, + 253, + 182 + ], + "blocks": [ + { + "bbox": [ + 154, + 148, + 253, + 182 + ], + "lines": [ + { + "bbox": [ + 154, + 148, + 253, + 182 + ], + "spans": [ + { + "bbox": [ + 154, + 148, + 253, + 182 + ], + "type": "image", + "image_path": "5c5a432204711046296b8eb590eb28e6cfa3397c63fa6bd83fcdec4b96b1c1c7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 195, + 182, + 215, + 190 + ], + "lines": [ + { + "bbox": [ + 195, + 182, + 215, + 190 + ], + "spans": [ + { + "bbox": [ + 195, + 182, + 215, + 190 + ], + "type": "text", + "content": "(e) SD3" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 255, + 148, + 353, + 182 + ], + "blocks": [ + { + "bbox": [ + 255, + 148, + 353, + 182 + ], + "lines": [ + { + "bbox": [ + 255, + 148, + 353, + 182 + ], + "spans": [ + { + "bbox": [ + 255, + 148, + 353, + 182 + ], + "type": "image", + "image_path": "79d7ab75f175a10401291e5733ca09c3a9a45d4049bcae52a7f8d3669e2ca469.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 293, + 182, + 325, + 189 + ], + "lines": [ + { + "bbox": [ + 293, + 182, + 325, + 189 + ], + "spans": [ + { + "bbox": [ + 293, + 182, + 325, + 189 + ], + "type": "text", + "content": "(f) InstantID" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 353, + 148, + 451, + 182 + ], + "blocks": [ + { + "bbox": [ + 353, + 148, + 451, + 182 + ], + "lines": [ + { + "bbox": [ + 353, + 148, + 451, + 182 + ], + "spans": [ + { + "bbox": [ + 353, + 148, + 451, + 182 + ], + "type": "image", + "image_path": "96e20e2700d895b48c3bae974748669a32f3565913b33186b0a6b646c5450d6c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 373, + 182, + 430, + 190 + ], + "lines": [ + { + "bbox": [ + 373, + 182, + 430, + 190 + ], + "spans": [ + { + "bbox": [ + 373, + 182, + 430, + 190 + ], + "type": "text", + "content": "(g) VQFR (real world)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 451, + 148, + 550, + 182 + ], + "blocks": [ + { + "bbox": [ + 451, + 148, + 550, + 182 + ], + "lines": [ + { + "bbox": [ + 451, + 148, + 550, + 182 + ], + "spans": [ + { + "bbox": [ + 451, + 148, + 550, + 182 + ], + "type": "image", + "image_path": "5549685e6ba630acb1c517d5728a32ae0ca6ceee01efa7ff46dfa9400665b798.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 469, + 182, + 529, + 190 + ], + "lines": [ + { + "bbox": [ + 469, + 182, + 529, + 190 + ], + "spans": [ + { + "bbox": [ + 469, + 182, + 529, + 190 + ], + "type": "text", + "content": "(h) StableSR (synthetic)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 224, + 295, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 224, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 295, + 258 + ], + "type": "text", + "content": "generated image resembles a natural, real-life photograph, with particular emphasis on realistic skin texture, facial details, and wrinkles. It is used in face generation task." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 259, + 295, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 259, + 295, + 295 + ], + "spans": [ + { + "bbox": [ + 55, + 259, + 295, + 295 + ], + "type": "text", + "content": "ID Fidelity. This dimension evaluates how well the identity of the reference image is preserved in the customized or restored image in face customization and restoration tasks." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 296, + 295, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 296, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 296, + 295, + 319 + ], + "type": "text", + "content": "Correspondence. This dimension evaluates text-to-image alignment in face generation and customization tasks." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 328, + 185, + 341 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 328, + 185, + 341 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 185, + 341 + ], + "type": "text", + "content": "3.3. Subjective Experiment" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 346, + 296, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 296, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 296, + 477 + ], + "type": "text", + "content": "Main Assessment. To ensure a comprehensive, fair, and reliable subjective assessment, 180 participants with normal or corrected-to-normal eyesight are recruited for the experiment. They first take expert-led tutorials before the main assessment. In the main assessment, participants are asked to rate the image by dragging a sliding window from 0 to 5 with a two-digit decimal from different dimensions. All images are displayed on 27-inch 4K Dell monitors, randomly presented, at their original resolution under standard lighting conditions. The experiment follows the ITU-R BT.500-14 guidelines [78] for subjective evaluations." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 479, + 296, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 296, + 551 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 296, + 551 + ], + "type": "text", + "content": "Subjective Score Processing. To eliminate the inter-annotator variability, outlier detection based on Kurtosis is conducted with a rejection rate of " + }, + { + "bbox": [ + 55, + 479, + 296, + 551 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 55, + 479, + 296, + 551 + ], + "type": "text", + "content": " [78]. As a result, each image is rated by at least 15 subjects. These valid ratings are converted and scaled into Z-scores ranging from [0, 100] with the following formulas:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 84, + 559, + 295, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 559, + 295, + 585 + ], + "spans": [ + { + "bbox": [ + 84, + 559, + 295, + 585 + ], + "type": "interline_equation", + "content": "z _ {i j} = \\frac {r _ {i j} - \\mu_ {i}}{\\sigma_ {i}}, \\quad z _ {i j} ^ {\\prime} = \\frac {1 0 0 (z _ {i j} + 3)}{6}, \\tag {1}", + "image_path": "ca2eb5d5d6ddc22231373225f7dd3b0863c7575944ffc0ee3e4c4d71e87bc666.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 61, + 606, + 295, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 606, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 61, + 606, + 295, + 644 + ], + "type": "interline_equation", + "content": "\\mu_ {i} = \\frac {1}{M _ {i}} \\sum_ {j = 1} ^ {M _ {i}} r _ {i j}, \\sigma_ {i} = \\sqrt {\\frac {1}{M _ {i} - 1} \\sum_ {j = 1} ^ {M _ {i}} (r _ {i j} - \\mu_ {i}) ^ {2}}, \\tag {2}", + "image_path": "b0d5862f27b73027a77be5ccfa2e65b34e50c68b3676b4d8eda6291fa37ffc58.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "r_{ij}" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " represents the original ratings provided by the " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "-th evaluator for the " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "-th image, " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "M_i" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " denotes the total number of images that were assessed by the " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "-th participant in this group. Finally, the mean opinion score (MOS) of the image " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "MOS_j" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": ") is computed by averaging the rescaled" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 314, + 224, + 395, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 224, + 395, + 235 + ], + "spans": [ + { + "bbox": [ + 314, + 224, + 395, + 235 + ], + "type": "text", + "content": "z-scores as follows:" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 380, + 243, + 554, + 276 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 243, + 554, + 276 + ], + "spans": [ + { + "bbox": [ + 380, + 243, + 554, + 276 + ], + "type": "interline_equation", + "content": "M O S _ {j} = \\frac {1}{N _ {j}} \\sum_ {i = 1} ^ {N _ {j}} z _ {i j} ^ {\\prime} \\tag {3}", + "image_path": "e64f83eaa2af2f63fc05c93d55bf52f1481034a1d63ceba831744443c2a884ed.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 284, + 553, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 284, + 553, + 310 + ], + "spans": [ + { + "bbox": [ + 313, + 284, + 553, + 310 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 284, + 553, + 310 + ], + "type": "inline_equation", + "content": "N_{j}" + }, + { + "bbox": [ + 313, + 284, + 553, + 310 + ], + "type": "text", + "content": " denotes the number of valid subjects for image " + }, + { + "bbox": [ + 313, + 284, + 553, + 310 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 284, + 553, + 310 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 284, + 553, + 310 + ], + "type": "inline_equation", + "content": "z_{ij}^{\\prime}" + }, + { + "bbox": [ + 313, + 284, + 553, + 310 + ], + "type": "text", + "content": " denotes rescaled z-scores." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 319, + 553, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 553, + 345 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 553, + 345 + ], + "type": "text", + "content": "4. F-Bench: Benchmarking Current Generative Face Models" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 353, + 554, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 353, + 554, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 353, + 554, + 425 + ], + "type": "text", + "content": "In this section, we establish a human preference benchmark F-Bench for today's state-of-the-art face generation, customization, and restoration methods, from quality, authenticity, ID fidelity, and correspondence dimensions. This benchmark analyzes the strengths and weaknesses, providing new insights and potential directions for future work." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 432, + 432, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 432, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 432, + 445 + ], + "type": "text", + "content": "4.1. Perspective Analysis" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 449, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 653 + ], + "type": "text", + "content": "Fig.3 demonstrate the MOS distribution across eight dimensions in the three subsets, highlighting distinct variations between dimensions. (1) In the face generation task, the Correspondence score is the highest among the three dimensions, indicating that existing generative models perform well in aligning with prompts. The distribution of Authenticity is skewed towards lower scores, suggesting that current face generation models still fall short in generating photorealistic face images that satisfy human preferences. (2) In face customization, models perform the worst in the Quality dimension overall, demonstrating the difficulty of injecting identity information without compromising image quality. Identity Fidelity shows a bimodal distribution, reflecting the dichotomous nature of human judgment regarding whether the generated face accurately represents the same identity. (3) In face restoration, models perform better in Quality than in Identity Fidelity." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 660, + 451, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 660, + 451, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 660, + 451, + 673 + ], + "type": "text", + "content": "4.2. Model-wise Comparison" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": "Face Generation. Fig.4a presents the average MOS scores across three dimensions: Quality, Authenticity, and Correspondence. Correspondence scores show a steady increase" + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10985" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 99, + 70, + 237, + 155 + ], + "blocks": [ + { + "bbox": [ + 99, + 70, + 237, + 155 + ], + "lines": [ + { + "bbox": [ + 99, + 70, + 237, + 155 + ], + "spans": [ + { + "bbox": [ + 99, + 70, + 237, + 155 + ], + "type": "image", + "image_path": "303eb0332611bac5a2f83daa16a9a9abd49abc0fce780e81bf45e786228ec912.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 159, + 555, + 182 + ], + "lines": [ + { + "bbox": [ + 55, + 159, + 555, + 182 + ], + "spans": [ + { + "bbox": [ + 55, + 159, + 555, + 182 + ], + "type": "text", + "content": "Figure 4. Average MOS score comparison across all models and dimensions. (a) Face generation. (b) Face customization. (c) Face restoration. The models are arranged in a clockwise order by release date." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 250, + 70, + 400, + 156 + ], + "blocks": [ + { + "bbox": [ + 250, + 70, + 400, + 156 + ], + "lines": [ + { + "bbox": [ + 250, + 70, + 400, + 156 + ], + "spans": [ + { + "bbox": [ + 250, + 70, + 400, + 156 + ], + "type": "image", + "image_path": "f59aec7a75dddcb25645866c238883c65a864408f00f39ecea66cc85f857374e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 406, + 70, + 547, + 156 + ], + "blocks": [ + { + "bbox": [ + 406, + 70, + 547, + 156 + ], + "lines": [ + { + "bbox": [ + 406, + 70, + 547, + 156 + ], + "spans": [ + { + "bbox": [ + 406, + 70, + 547, + 156 + ], + "type": "image", + "image_path": "23104c2b45121fa81dc3d3fde8bf89c52a2708806106f4cf5126a2ed3de561b8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 56, + 185, + 553, + 268 + ], + "blocks": [ + { + "bbox": [ + 56, + 185, + 553, + 268 + ], + "lines": [ + { + "bbox": [ + 56, + 185, + 553, + 268 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 553, + 268 + ], + "type": "image", + "image_path": "58c75820987283522713a3ba1b67a33e0791394c6a0c2832d19922b94416984a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 269, + 542, + 281 + ], + "lines": [ + { + "bbox": [ + 66, + 269, + 542, + 281 + ], + "spans": [ + { + "bbox": [ + 66, + 269, + 542, + 281 + ], + "type": "text", + "content": "Figure 5. Violin plots of quality, authenticity and correspondence scores in FaceQ-Gen subset across nine prompt categories." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 286, + 295, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 286, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 54, + 286, + 295, + 464 + ], + "type": "text", + "content": "over time, indicating consistent improvement in text comprehension. Quality scores generally rise with the release year, although Hunyuan [58] and Dreamlike [2] exhibit deviations. Hunyuan [58] oversmoothes the facial details and performs worse compared to its contemporaries, while Dreamlike [2] performs the opposite. Authenticity scores exhibit the greatest variation among models, suggesting that current face generation models place insufficient emphasis on authenticity. Except for Realistic Vision [8], SD3 [9], Flux-dev [3], and Dreamlike [2], most models lag in authenticity relative to quality, highlighting a common phenomenon of high-quality but less realistic outputs in generative models. Fig.3d,e give some examples of MOS distributions of different methods, revealing a reasonable fine-grained performance difference." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 465, + 295, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 465, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 55, + 465, + 295, + 644 + ], + "type": "text", + "content": "Face Customization. Fig.4b is the performance radar chart for face customization models. IP-Adapter-FaceID-Plus[5] achieves the best performance across all dimensions, followed by PhotoMaker [57]. This highlights their great robustness in identity encoders. ELITE [98] scores the lowest in all dimensions. IP-Adapter-FaceID [4] produces high-quality images but lacks accuracy in Identity Fidelity. This may be because the base model [71] introduces many AI artifacts, which affect both Identity Fidelity and Quality. InstantID [91] and FastComposer [101] score well in Identity Fidelity but underperform significantly in Quality and Correspondence, indicating strong identity preservation but limited text-driven editing capabilities. Fig.3h illustrate the model-wise MOS distributions, revealing similar trends to those observed in the radar chart." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 645, + 295, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 645, + 295, + 704 + ], + "spans": [ + { + "bbox": [ + 55, + 645, + 295, + 704 + ], + "type": "text", + "content": "Face Restoration. Fig.4c displays the average performance comparison of nine face restoration models. DR2 [97] exhibits the weakest performance in both Quality and Identity fidelity. This may be due to the degradation removal process of DR2 [97], which results in the loss of fine de" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 286, + 555, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 286, + 555, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 286, + 555, + 406 + ], + "type": "text", + "content": "tails and excessive smoothing. In contrast, StableSR [89] and CodeFormer [113] achieve the best overall performance among the nine face restoration models, successfully preserving both intricate details and accurate identity representations. When comparing the performance on real-world and synthetic data, most methods struggle to achieve satisfactory results on complex real-world degradations, except for SAPRNet [15] and DiffBIR [62]. StableSR [89] and GFPGAN [94] show strong robustness, performing consistently well on both real-world and synthetic data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 411, + 446, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 446, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 446, + 423 + ], + "type": "text", + "content": "4.3. Class-wise Comparison" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "content": "For the face generation task, we analyze the Mean Opinion Score (MOS) distributions of Quality, Authenticity, and Correspondence across nine prompt categories, as shown in Fig.5. The violin plots highlight distinct performance patterns across these categories. The Action, Expression, and Facial Attributes categories exhibit lower MOS values in all three dimensions, with notable variability. In the Action category, limited visibility of the face due to body movements affects the generation of fine facial details, resulting in reduced scores, especially in Quality and Correspondence. The Expression category poses challenges in maintaining spatial relationships and physical constraints, leading to distortion and artifacts in expressions with exaggerated facial movements, such as laughing or crying, thereby impacting both Quality and Authenticity. For Facial Attributes, generating images with precise attributes (e.g., specific eye color, hair style) is challenging, particularly when multiple attributes are specified, which lowers the scores. In contrast, the General and Background categories achieve the highest MOS scores across all dimensions, as simpler or less specific prompts are easier for the model to interpret accurately. The Clothes and Accessories categories also perform well, indicating that the model handles clothes and accessory prompts effectively. Overall, prompts with specific or" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "10986" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 209, + 242 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 209, + 242 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 209, + 242 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 209, + 242 + ], + "type": "image", + "image_path": "ea49b49f9d03c0009d83adb59ba0355c21572a7476d5ebc10ade03953940f09e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 243, + 555, + 299 + ], + "lines": [ + { + "bbox": [ + 54, + 243, + 555, + 299 + ], + "spans": [ + { + "bbox": [ + 54, + 243, + 555, + 299 + ], + "type": "text", + "content": "Figure 6. The overall framework of F-Eval. F-Eval can evaluate quality, authenticity, correspondence, and identity fidelity in a one-for-all framework. It can process both single and paired images, along with prompts, to produce quality scores. It consists of three encoders, including a vision encoder, a face encoder, and a text tokenizer to process multi-modal inputs. These features are projected into the same space by trained projectors. A pre-trained large language model is utilized to fuse the features while fine-tuned with four LoRA experts. Specific LoRA will be activated by dimension ID, which is classified by a trainable router." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 219, + 70, + 553, + 242 + ], + "blocks": [ + { + "bbox": [ + 219, + 70, + 553, + 242 + ], + "lines": [ + { + "bbox": [ + 219, + 70, + 553, + 242 + ], + "spans": [ + { + "bbox": [ + 219, + 70, + 553, + 242 + ], + "type": "image", + "image_path": "eabc3a39c23181b6f185bfa475841063d2fa90b82c807841ba30c66e7609deb8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 301, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 301, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 301, + 296, + 361 + ], + "type": "text", + "content": "complex attribute requirements tend to result in lower MOS scores, while general or simple prompts yield consistently high scores across dimensions. This can be attributed to the absence of high-quality fine-grained text-face pair datasets during the training of generative models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 370, + 149, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 370, + 149, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 149, + 384 + ], + "type": "text", + "content": "5. F-Eval Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 391, + 296, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 391, + 296, + 439 + ], + "spans": [ + { + "bbox": [ + 55, + 391, + 296, + 439 + ], + "type": "text", + "content": "In this section, we introduce our one-for-all generated face quality assessment method, F-Eval, towards predicting quality scores from multiple dimensions for three tasks using one model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 445, + 176, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 176, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 176, + 456 + ], + "type": "text", + "content": "5.1. Overall Architecture" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 462, + 296, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 296, + 606 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 296, + 606 + ], + "type": "text", + "content": "The overall architecture of F-Eval is depicted in Fig. 6. As shown on the left, F-Eval can assess four dimensions: quality, authenticity, text-image correspondence, and identity fidelity, with multimodal inputs, i.e.text prompts, image and optional reference image, to output quality-related answers. On the right, the input image(s) can originate from all three tasks. F-Eval employs a vision encoder and a face encoder to extract visual features, with projectors projecting both into a unified language space. User prompts are encoded into textual embeddings through a tokenizer. These multimodal features are concatenated and fed into a pre-trained large language model for the final output." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 612, + 143, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 143, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 143, + 625 + ], + "type": "text", + "content": "5.2. Model Design" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "Vision Encoding. For input image(s), F-Eval adopt a vision encoder to extract global features and a face encoder captures face features related to facial structure and identity. Specifically, the vision encoder, Qwen-VL-2.5-ViT [11], is tuned on FaceQ to adjust different dimensions. Face encoder is Arcface [22], which is frozen. Two two-layer multi-layer perception projectors then map the features." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 301, + 555, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 555, + 387 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 555, + 387 + ], + "type": "text", + "content": "Feature Fusion via the LLM. For input prompts, it is tokenized into text tokens and then concatenated with visual tokens. Here, Qwen-VL2.5-7B [11] is employed as the pretrained LLM backbone to combine visual tokens and text tokens to perform multi-modal learning.. Finally, the output features of the LLM are decoded with a text decoder and then projected to the prompt-corresponded quality space." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 393, + 468, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 468, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 468, + 407 + ], + "type": "text", + "content": "5.3. Multi-Expert LoRA Tuning" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 411, + 555, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 555, + 554 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 555, + 554 + ], + "type": "text", + "content": "Instruction Tuning. Previous IQA methods are often designed for specific dimensions [106, 112], addressing only one aspect of the problem or requiring retraining of model weights for different dimensions. Leveraging the generalization ability of LLMs, instruction tuning allows for the integration of different dimensions within a unified framework. Therefore, we construct diverse instruction data for the four facial quality evaluation dimensions in the FaceQ dataset and fine-tune F-Eval by adjusting the visual encoder, projector, and LLM parameters to adapt it to the specific task. As a result, F-Eval is capable of evaluating facial quality using unified weights." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 554, + 556, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 554, + 556, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 554, + 556, + 687 + ], + "type": "text", + "content": "Mixture of LoRA Experts (MoLE). Tuning a holistic quality model for different dimensions using a single LoRA expert introduces inherent interference between dimensions. Hence, F-Eval integrates four LoRA experts for four dimensions when tuning vision encoder and LLM, constituting MoLE modules. A router is trained to classify input prompts into the four dimensions and output the corresponding dimension ID " + }, + { + "bbox": [ + 313, + 554, + 556, + 687 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 554, + 556, + 687 + ], + "type": "text", + "content": ". Specifically, for a frozen linear layer " + }, + { + "bbox": [ + 313, + 554, + 556, + 687 + ], + "type": "inline_equation", + "content": "h = Wx" + }, + { + "bbox": [ + 313, + 554, + 556, + 687 + ], + "type": "text", + "content": " with input " + }, + { + "bbox": [ + 313, + 554, + 556, + 687 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^{d_i}" + }, + { + "bbox": [ + 313, + 554, + 556, + 687 + ], + "type": "text", + "content": " and weight matrix " + }, + { + "bbox": [ + 313, + 554, + 556, + 687 + ], + "type": "inline_equation", + "content": "W \\in \\mathbb{R}^{d_o \\times d_i}" + }, + { + "bbox": [ + 313, + 554, + 556, + 687 + ], + "type": "text", + "content": ", a LoRA expert is trained to learn a low-rank decomposed update:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 358, + 695, + 555, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 695, + 555, + 717 + ], + "spans": [ + { + "bbox": [ + 358, + 695, + 555, + 717 + ], + "type": "interline_equation", + "content": "h = W x + \\Delta W x = W x + \\frac {\\alpha}{r} B A x, \\tag {4}", + "image_path": "915e177f62b899dee612f1c85984e5d690bc21aabd8fb062b3df8995745039ac.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10987" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 96, + 555, + 344 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 2. Performance comparison on FaceQ-Gen and FaceQ-Cus subsets. " + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "inline_equation", + "content": "\\spadesuit" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "inline_equation", + "content": "\\clubsuit" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "inline_equation", + "content": "\\diamond" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "inline_equation", + "content": "\\heartsuit" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": " denote traditional IQA models, face IQA models, classical deep learning-based IQA models, and AIGC IQA models respectively. * indicates fine-tuned results." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 96, + 555, + 344 + ], + "lines": [ + { + "bbox": [ + 57, + 96, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 555, + 344 + ], + "type": "table", + "html": "
TaskFace GenerationFace Customization
DimensionQualityAuthenticityCorrespondenceQualityID FidelityCorrespondence
ModelSRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑
◆NIQE [67]0.17260.11420.25530.08690.05710.13460.07520.05070.10980.13650.08990.14360.14900.09910.13100.11860.07840.1245
◆ILNIQE [111]0.19810.13190.31590.06750.04450.15050.08890.05910.16650.20320.13270.24790.21920.14520.27780.18780.12510.2275
◆HOSA [102]0.25370.17080.29720.11030.07280.14460.1220.08140.14950.10190.06880.05970.07530.05060.05520.10360.06960.0748
◆BPRI-PSS [64]0.23620.15950.24130.17000.11420.17720.16730.11210.16470.02380.01660.03140.14250.09720.17580.11000.07290.1067
◆FISBLIM [27]0.09010.05950.04330.00280.00030.01750.10000.06650.01960.16140.10240.16900.02560.01730.01260.08520.05590.0854
◆ArcFace[22]N/AN/AN/AN/AN/AN/AN/AN/AN/A0.30990.20940.36050.50620.34390.55720.18370.12430.2244
◆SER-FIQ[83]0.15520.10340.15610.04700.03140.05750.11300.0750.1110-0.1241-0.080-0.1241-0.0109-0.00780.0042-0.1680-0.1121-0.1911
◆DSL-FIQA[18]0.56060.39270.59450.31560.20960.39600.37060.25140.43550.36150.24500.36510.24290.16520.25190.10400.07030.0974
◆IFQA[41]0.23130.15490.32610.15970.10620.20970.15970.10770.22600.24010.16140.22200.15030.09730.17390.37310.25460.3941
◆CNNIAQ*[42]0.42190.28950.35100.30370.20540.29730.28240.19310.25950.62440.43760.66200.65110.44410.63810.37650.25450.3264
◇VGG16*[79]0.58460.41410.60060.50100.35080.51810.43490.29940.47230.72170.53240.78570.67350.47170.66360.56350.39320.5589
◇VGG19*[79]0.57280.40430.56440.46810.32480.44790.41710.28730.47540.79520.60310.80960.75440.54170.75550.65850.47420.6465
◇ResNet18*[29]0.61500.44170.66600.46000.31600.46040.48290.33600.46280.79660.60130.80060.74730.53350.76560.63080.44880.6191
◇ResNet34*[29]0.60920.43540.65370.55640.39490.54010.44100.30650.52070.81410.62110.81510.79690.58780.81570.57080.40000.5810
◇HyperIAQ*[81]0.63330.45450.65910.61060.44060.60420.42350.29120.42400.85240.66310.82050.84190.64630.85880.66280.47570.6454
◇TRES*[25]0.77660.58670.80470.64580.47030.65790.57080.40380.62700.88830.70830.89510.83300.64090.86580.73930.54970.7400
◆MANIAQ*[106]0.78710.59510.81500.73080.54360.77940.62780.44430.70660.89700.71850.89520.85900.66540.87960.78490.59410.7822
◇PromptIAQ*[14]0.78230.59070.82450.72320.53270.75580.65140.46980.70650.90710.73150.88210.86930.67920.88760.83480.63420.8281
♥CLIPScore[32]0.09670.06420.13150.06400.04080.08170.23960.15910.27330.31260.21060.31850.18160.12180.18110.70080.50980.6986
♥BLIPScore[53]0.13060.08720.19290.09350.06130.15690.20510.13640.29630.30730.20630.31360.12130.07960.12660.67070.47970.6123
♥ImageReward [103]0.38490.26310.43580.26060.17360.31810.51550.35420.58710.32320.21660.31960.19050.12700.18390.77140.56730.7740
♥MINTIAQ[88]0.70020.51600.80470.62170.44500.68260.54910.38910.65840.53780.37750.55740.40880.28160.40410.28860.19320.2959
♥MINTIAQ*[88]0.83120.64740.89740.81770.63060.85110.79080.59910.86670.89040.70920.88160.85240.66970.87550.83910.64130.8318
F-Eval (Ours)0.84860.66700.90850.83120.65850.85780.84710.66370.91060.94620.79610.94610.91880.76400.93220.94600.79590.9457
", + "image_path": "2404cf1bdcbb2950b5e72e8a7e17a7c42e44040aba12ba64fcc397b03b4e2d71.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 347, + 295, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 347, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 295, + 407 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 347, + 295, + 407 + ], + "type": "inline_equation", + "content": "A\\in \\mathbb{R}^{r\\times d_i}" + }, + { + "bbox": [ + 55, + 347, + 295, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 347, + 295, + 407 + ], + "type": "inline_equation", + "content": "B\\in \\mathbb{R}^{d_o\\times r}" + }, + { + "bbox": [ + 55, + 347, + 295, + 407 + ], + "type": "text", + "content": " are the low-rank matrices, " + }, + { + "bbox": [ + 55, + 347, + 295, + 407 + ], + "type": "inline_equation", + "content": "r\\ll \\min (d_o,d_i)" + }, + { + "bbox": [ + 55, + 347, + 295, + 407 + ], + "type": "text", + "content": " is the rank, and " + }, + { + "bbox": [ + 55, + 347, + 295, + 407 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 347, + 295, + 407 + ], + "type": "text", + "content": " is scaling factor. For prompt of certain dimension, the corresponding expert is activated based on the dimension ID " + }, + { + "bbox": [ + 55, + 347, + 295, + 407 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 347, + 295, + 407 + ], + "type": "text", + "content": ", which can be formulated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 404, + 295, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 404, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 404, + 295, + 426 + ], + "type": "interline_equation", + "content": "E _ {k} (x) = \\frac {\\alpha_ {k}}{r _ {k}} B _ {k} A _ {k} x, \\tag {5}", + "image_path": "7294e54019f1a4d86d6a60ca2ae4ac3cd4a6446edff83539a1912634ed62ded2.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 430, + 208, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 208, + 442 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 208, + 442 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 430, + 208, + 442 + ], + "type": "inline_equation", + "content": "E_{k}(\\cdot)" + }, + { + "bbox": [ + 55, + 430, + 208, + 442 + ], + "type": "text", + "content": " is the k-th LoRA expert." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 452, + 137, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 452, + 137, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 452, + 137, + 466 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 471, + 188, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 471, + 188, + 484 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 188, + 484 + ], + "type": "text", + "content": "6.1. Implementation Details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 488, + 296, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 488, + 296, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 488, + 296, + 597 + ], + "type": "text", + "content": "The LLM used in F-Eval is Qwen-2.5VL [11], with an input token channel dimension of 3584. The vision encoder, face encoder, and LLM are all frozen during training. The ranks of LoRA experts are 16,32,16,16, respectively. The LoRA alphas are set equal to LoRA ranks. LoRA dropout is set as 0.05. The model is trained on 2 NVIDIA GeForce RTX 3090 GPUs with flash-attn [21] enabled. We use AdamW Optimizer to train the network for 10 epochs with a batch size of 4 and learning rate " + }, + { + "bbox": [ + 55, + 488, + 296, + 597 + ], + "type": "inline_equation", + "content": "1e^{-4}" + }, + { + "bbox": [ + 55, + 488, + 296, + 597 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 600, + 254, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 254, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 254, + 613 + ], + "type": "text", + "content": "6.2. Comparison with existing QA Models" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": "We conduct comprehensive experiments on the proposed FaceQ database to evaluate the effectiveness of both current quality assessment methods and our proposed methods, as shown in Tab.2, Tab.3, and Tab.4. The evaluation metrics are Spearman rank-order correlation coefficient (SRCC), Kendall rank order correlation coefficient (KRCC), and Pearson linear correlation coefficient (PLCC), which assess the correlation between predicted scores by the QA" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 348, + 555, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 555, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 555, + 552 + ], + "type": "text", + "content": "methods and the MOS scores provided by humans. Higher metrics indicated better alignment with human preferences. The comparison QA models falls into four categories: 1) Traditional IQA Models. NR-IQA models include NIQE [67], ILNIQE [111], HOSA [102], BPRI-PSS [64], FISBLIM [27]. FR-IQA models include LPIPS [112], SSIM [96] and PSNR [34]. 2) Classical Deep learning-based IQA Models. This category includes CNNIQA [42], VGG16 [79], VGG19 [79], ResNet18 [29], ResNet34 [29], Hyper-IQA [81], TReS [25], MANIQA [106], PromptIQA[14]. All models are retrained on FaceQ with an 80:20 training/testing split. 3) AIGC IQA Models. This category includes pre-trained vision-language models, including CLIP-Score [32], BLIPScore [53], zero-shot metric ImageReward [103], and the AIGC-specific QA model MINTIQA [88]. 4) Face IQA Models. This category includes SER-FIQ [83], DSL-FIQA [18], ArcFace [22], IFQA[41]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 562, + 440, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 562, + 440, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 562, + 440, + 575 + ], + "type": "text", + "content": "6.3. Performance Analysis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 554, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 629 + ], + "type": "text", + "content": "Traditional IQA Models. Traditional IQA methods, including both NR-IQA and FR-IQA models, generally perform poorly across all tasks and dimensions, with low SRCC, KRCC, and PLCC scores." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 630, + 554, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 701 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 701 + ], + "type": "text", + "content": "Classical Deep Learning-based IQA Models. After being trained on our dataset, deep learning-based IQA models, such as MANIQA [106] and HyperIQA [81], significantly outperform traditional models in terms of quality and authenticity. However, these methods face challenges in text understanding, leading to unsatisfactory correlations." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 701, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 554, + 714 + ], + "type": "text", + "content": "AIGC IQA Models. CLIPScore [32] BLIPScore [53], and" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10988" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 116, + 294, + 313 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 295, + 114 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 295, + 114 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 295, + 114 + ], + "type": "text", + "content": "Table 3. Performance comparison on FaceQ-Res synthetic subset. " + }, + { + "bbox": [ + 55, + 71, + 295, + 114 + ], + "type": "inline_equation", + "content": "\\spadesuit" + }, + { + "bbox": [ + 55, + 71, + 295, + 114 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 71, + 295, + 114 + ], + "type": "inline_equation", + "content": "\\heartsuit" + }, + { + "bbox": [ + 55, + 71, + 295, + 114 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 71, + 295, + 114 + ], + "type": "inline_equation", + "content": "\\clubsuit" + }, + { + "bbox": [ + 55, + 71, + 295, + 114 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 71, + 295, + 114 + ], + "type": "inline_equation", + "content": "\\diamond" + }, + { + "bbox": [ + 55, + 71, + 295, + 114 + ], + "type": "text", + "content": " denote traditional NR-IQA models, FR-IQA models, face IQA models, and classical deep learning-based IQA models respectively. * indicates fine-tuned results." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 116, + 294, + 313 + ], + "lines": [ + { + "bbox": [ + 56, + 116, + 294, + 313 + ], + "spans": [ + { + "bbox": [ + 56, + 116, + 294, + 313 + ], + "type": "table", + "html": "
Dimension ModelQualityID Fidelity
SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑
▲BMPRI[65]0.21610.14280.18620.11130.07390.1069
▲NIQE[67]0.17980.12130.16420.21930.14800.2118
▲ILNIQE[111]0.25050.17270.31800.26960.18570.3321
▲HOSA[102]0.37580.25220.37480.34180.34550.2301
▲BPII-LSSn [64]0.13980.09570.09990.12440.08370.0973
▲BPII [64]0.16370.10650.14800.01950.01210.0281
▲FISBLIM [27]0.15890.10890.20380.10540.07090.1715
\\( \\diamondsuit \\)LPIPS[112]0.39090.26580.46430.43720.30180.4870
\\( \\diamondsuit \\)SSIM[96]0.10170.06750.04880.09330.06010.0504
\\( \\diamondsuit \\)PSNR [34]0.01980.01130.05360.09860.06620.1212
\\( \\diamondsuit \\)ArcFace[22]0.34660.23270.37740.57280.40300.5762
\\( \\diamondsuit \\)SER-FIQ [25]0.38130.25840.37460.39510.26410.3707
\\( \\diamondsuit \\)DSL-FIQA [18]0.63870.45080.65280.48820.33600.5379
\\( \\diamondsuit \\)IFQA [41]0.22730.15270.33480.32140.21850.3874
\\( \\diamondsuit \\)CNNIA* [42]0.46610.31970.38360.46780.32160.4288
\\( \\diamondsuit \\)ResNet18* [29]0.72540.52950.69820.66620.48020.6527
\\( \\diamondsuit \\)ResNet34* [29]0.66930.49160.65500.59050.41700.6084
\\( \\diamondsuit \\)VGG16* [79]0.74870.55130.71010.61770.44130.6381
\\( \\diamondsuit \\)VGG19* [79]0.69940.51260.67220.60380.42780.5999
\\( \\diamondsuit \\)HyperIA* [81]0.82690.63230.82280.76150.55850.7471
\\( \\diamondsuit \\)MANIA* [106]0.82870.64030.87780.77040.56600.7619
\\( \\diamondsuit \\)TReS* [25]0.86560.68300.86220.79510.58790.7829
\\( \\diamondsuit \\)PromptIA* [14]0.85190.66610.87330.82200.62750.8254
F-Eval (Ours)0.86920.68550.90090.85070.67310.8726
", + "image_path": "65b474d478b7053cea0a3ed0d9efa13383f5095e8a8cf4b7d99eff41fd4c767b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 361, + 294, + 536 + ], + "blocks": [ + { + "bbox": [ + 55, + 316, + 295, + 360 + ], + "lines": [ + { + "bbox": [ + 55, + 316, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 295, + 360 + ], + "type": "text", + "content": "Table 4. Performance comparison on FaceQ-Res real-world subset. " + }, + { + "bbox": [ + 55, + 316, + 295, + 360 + ], + "type": "inline_equation", + "content": "\\spadesuit" + }, + { + "bbox": [ + 55, + 316, + 295, + 360 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 316, + 295, + 360 + ], + "type": "inline_equation", + "content": "\\clubsuit" + }, + { + "bbox": [ + 55, + 316, + 295, + 360 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 316, + 295, + 360 + ], + "type": "inline_equation", + "content": "\\diamond" + }, + { + "bbox": [ + 55, + 316, + 295, + 360 + ], + "type": "text", + "content": " denote traditional IQA models, face IQA models, and classical deep learning-based IQA models respectively. * indicates fine-tuned results." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 361, + 294, + 536 + ], + "lines": [ + { + "bbox": [ + 56, + 361, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 294, + 536 + ], + "type": "table", + "html": "
Dimension ModelQualityID Fidelity
SRCC↑KRCC↑PLCC↑SRCC↑KRCC↑PLCC↑
▲BMPRII[65]0.15820.10930.27990.01490.01080.1386
▲NIQE[67]0.17830.11860.19620.12450.08100.1701
▲ILNIQE[111]0.22300.16100.22910.19640.13620.2340
▲HOSA[102]0.39270.42770.26540.33280.22340.3813
▲BPRI-LSSn [64]0.27150.18400.30220.20590.13800.2475
▲BPRI [64]0.19200.13130.22000.08460.05700.1136
▲FISBLIM [27]0.25120.17050.27250.23410.15810.2699
▲ArcFace[22]0.23710.15650.41830.49560.34700.6330
▲SER-FIQ [25]-0.0482-0.02990.10230.04210.02990.1584
▲DSL-FIQA [18]0.65090.46410.64940.48400.33620.5118
▲IFQA [41]0.02510.02100.12210.08930.06030.2086
◇CNNIQA* [42]0.36500.25060.47050.32520.21970.3112
◇VGG16* [79]0.49960.34390.42690.41170.28590.4622
◇VGG19* [79]0.58710.41810.51330.30540.20440.3594
◇ResNet18* [29]0.59770.42210.57220.55540.39090.6161
◇ResNet34* [29]0.62590.44750.61680.50360.35390.6136
◇HyperiQA* [81]0.79260.59650.79260.77830.58450.8163
◇MANIQA* [106]0.83560.64820.84670.75970.56550.8081
◇TReS* [25]0.82870.63910.85010.72870.54410.8009
◇PromptIQA* [14]0.81150.61860.86450.73980.55260.8297
F-Eval(Ours)0.84480.65770.87050.79570.60570.8366
", + "image_path": "23070b0e5dae106f1be8f9cb016e40e7ca63b3bdba0a1cb43a109df1d31d079d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 545, + 295, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 545, + 295, + 642 + ], + "spans": [ + { + "bbox": [ + 54, + 545, + 295, + 642 + ], + "type": "text", + "content": "ImageReward [103] exhibit weak correlations with human preference, likely due to their focus on high-level semantics rather than low-level details crucial for image quality assessment. MINTIQA [88] achieves best performance among current AIGC IQA methods, but its performance is still unsatisfactory, particularly on the customization task. Fine-tuning on our dataset leads to significant improvements, highlighting the great potential of FaceQ." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 642, + 295, + 714 + ], + "type": "text", + "content": "Face IQA Models. ArcFace similarity [22] demonstrates its ability to extract identity-related features. However, it lacks perceptual sensitivity for image quality, leading to poor performance in quality assessment. DSL-FIQA [18], as the latest method, achieves the best performance in predicting quality scores among face IQA methods, but it is" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "type": "text", + "content": "still far from satisfactory. SER-FIQ fails to align with human perception across most dimensions and tasks. Experiments confirm that these natural face evaluation algorithms struggle to assess the complex characteristics of AIGFs, highlighting the need for an effective AIGF evaluator." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 133, + 555, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 555, + 239 + ], + "type": "text", + "content": "F-Eval (Ours). Our method achieves optimal performance across all tasks and dimensions, with the most significant gain in the customization task, as existing methods rarely focus human preferences in face customization. Among the four dimensions, the improvements in correspondence and ID fidelity stand out, thanks to dual-stream vision encoding and the powerful LLM backbone for text understanding and modality fusion. The MoLE technique allows the model to learn distinct features for each dimension." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 243, + 408, + 256 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 408, + 256 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 408, + 256 + ], + "type": "text", + "content": "6.4. Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 261, + 553, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 261, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 553, + 285 + ], + "type": "text", + "content": "We conducted the ablation study to validate the utility of the core components of F-Eval in Tab. 5." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 285, + 554, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 554, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 554, + 357 + ], + "type": "text", + "content": "Effectiveness of Vision Encoding. 1) The effectiveness of the face encoder is demonstrated by comparing the second and last rows. Without the face encoder, a significant performance drop is observed, indicating that the extracted features help the LLM interpret face images. 2) Removing the projector leads to a dramatic performance drop." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 357, + 554, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 554, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 554, + 452 + ], + "type": "text", + "content": "Effectiveness of MoLE. We conducted three sets of LoRA adaptation experiments: fine-tuning a single LoRA for all dimensions and tasks, fine-tuning three LoRA experts for three tasks, and fine-tuning four LoRA experts for four dimensions. The results show that fine-tuning multiple dimensions with a single LoRA leads to performance deterioration, which further confirms the independence of the characteristics across different dimensions." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 316, + 478, + 553, + 515 + ], + "blocks": [ + { + "bbox": [ + 313, + 454, + 553, + 477 + ], + "lines": [ + { + "bbox": [ + 313, + 454, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 553, + 477 + ], + "type": "text", + "content": "Table 5. Ablation study of the proposed F-Eval method. Here only present SRCC due to space limitations." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 478, + 553, + 515 + ], + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 515 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 515 + ], + "type": "table", + "html": "
Feature & StrategyFace GenerationFace CustomizationFace Restoration synFace Restoration rw
VisionProjectorFaceSingle LoRATask LoRADim. LoRAQualityAuthen.Corres.QualityID.Corres.QualityID.Corres.QualityID.QualityID.
✓✓✓0.8080.7500.8460.9240.8430.9270.7770.7760.7760.7070.748
✓✓✓0.8150.7520.8350.9240.8550.9210.8120.7830.7160.603
✓✓✓✓✓0.7930.7760.8190.9410.8850.9420.8380.8230.8020.671
✓✓✓✓0.8230.7800.8380.9420.8730.9420.8240.8260.7540.638
✓✓✓✓0.8490.8310.8470.9460.9190.9460.8690.8510.8450.796
", + "image_path": "9ebdecafbdb12c7efb3beb5e68bbe7d6c72c20c390f185fb5b9be00ebc00d9e6.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 526, + 388, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 388, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 388, + 538 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": "In this work, we conduct a pioneering generative face image quality assessment study. Specifically, we introduce FaceQ, the first generated face quality assessment database comprising 12,255 face images and 491,130 human preference ratings. Then, we establish F-Bench, a benchmark for evaluating the capabilities of existing face generation, customization, and restoration models, while also reveals the limitations of current Image Quality Assessment (IQA) and Face Quality Assessment (FQA) methods. Furthermore, we propose an evaluation model F-Eval to accurately predict human preferences on generated face images, which demonstrates state-of-the-art performance. We believe this work will contribute to advancing generative models and help overcome current limitations in AIGC community." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10989" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 198 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 198 + ], + "type": "text", + "content": "This work is supported by the National Natural Science Foundation of China under Grants 62271308, 62401365, 62225112, 62132006, U24A20220, STCSM under Grants 24ZR1432000, 24511106902, 24511106900, 22511105700, 22DZ2229005, 111 plan under Grant Number BP0719010, China Postdoctoral Science Foundation under Grants BX20250411, 2025M773473, and State Key Laboratory of UHD Video and Audio Production and Presentation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 209, + 115, + 221 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 115, + 221 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 115, + 221 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 229, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 66, + 229, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 229, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 66, + 229, + 294, + 251 + ], + "type": "text", + "content": "[1] Deep Floyd. https://github.com/deep-floyd/IF, Accessed: 2024-10-03. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 252, + 296, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 252, + 296, + 283 + ], + "spans": [ + { + "bbox": [ + 66, + 252, + 296, + 283 + ], + "type": "text", + "content": "[2] Dreamlike V2. https://huggingface.co/dreamlike-art/dreamlike-photoreal-2.0, Accessed: 2024-10-03.5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 286, + 294, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 286, + 294, + 317 + ], + "spans": [ + { + "bbox": [ + 66, + 286, + 294, + 317 + ], + "type": "text", + "content": "[3] Flux-dev. https://huggingface.co/black-forest-labs/FLUX.1-dev, Accessed: 2024-10-03.1,3,5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 320, + 293, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 320, + 293, + 342 + ], + "spans": [ + { + "bbox": [ + 66, + 320, + 293, + 342 + ], + "type": "text", + "content": "[4] IP-Adapter-FaceID-SDXL. https://huggingface.co/h94/IP-Adapter, Accessed: 2024-10-03. 3, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 343, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 343, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 66, + 343, + 294, + 374 + ], + "type": "text", + "content": "[5] IP-Adapter-FaceID-Plus. https://huggingface.co/h94/IP-Adapter-FaceID, Accessed: 2024-10-03.3,5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 377, + 294, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 377, + 294, + 398 + ], + "spans": [ + { + "bbox": [ + 66, + 377, + 294, + 398 + ], + "type": "text", + "content": "[6] Kcolors. https://huggingface.co/Kwai-Kcolors/Kcolors, Accessed: 2024-10-03. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 400, + 294, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 400, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 66, + 400, + 294, + 431 + ], + "type": "text", + "content": "[7] ProtoVision V6.6. https://huggingface.co/stablediffusionapi/protovision-xl-v6.6, Accessed: 2024-10-03. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 66, + 434, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 434, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 66, + 434, + 294, + 464 + ], + "type": "text", + "content": "[8] Realistic Vision V5.1. https://huggingface.co/SG161222/Realistic_Vision_V5.1_noVAE, Accessed: 2024-10-03. 3, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 66, + 468, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 468, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 66, + 468, + 294, + 498 + ], + "type": "text", + "content": "[9] SD3. https://huggingface.co/stabilityai/stable-diffusion-3-medium, Accessed: 2024-10-03. 3,5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 501, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 501, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 62, + 501, + 294, + 521 + ], + "type": "text", + "content": "[10] GPT-4o. https://chatgpt.com/, Accessed: 2024-10-10.3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 524, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 524, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 62, + 524, + 294, + 567 + ], + "type": "text", + "content": "[11] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. 2, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 569, + 296, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 569, + 296, + 623 + ], + "spans": [ + { + "bbox": [ + 62, + 569, + 296, + 623 + ], + "type": "text", + "content": "[12] Fadi Boutros, Meiling Fang, Marcel Klemt, Biying Fu, and Naser Damer. Cr-fiqa: face image quality assessment by learning sample relative classifiability. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5836-5845, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 62, + 624, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 624, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 62, + 624, + 294, + 678 + ], + "type": "text", + "content": "[13] Nicolas Chahine, Stefania Calarasanu, Davide Garcia-Civiero, Theo Cayla, Sira Ferradans, and Jean Ponce. An image quality assessment dataset for portraits. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9968-9978, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 62, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 681, + 294, + 713 + ], + "type": "text", + "content": "[14] Chen et al. Promptia: Boosting the performance and generalization for no-reference image quality assessment via prompts. In ECCV, 2024. 7, 8" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 320, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 553, + 116 + ], + "type": "text", + "content": "[15] Chaofeng Chen, Dihong Gong, Hao Wang, Zhifeng Li, and Kwan-Yee K Wong. Learning spatial attention for face super-resolution. IEEE Transactions on Image Processing (TIP), 30:1219-1231, 2020. 2, 3, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 119, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 119, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 321, + 119, + 553, + 183 + ], + "type": "text", + "content": "[16] Chaofeng Chen, Xiaoming Li, Lingbo Yang, Xianhui Lin, Lei Zhang, and Kwan-Yee K Wong. Progressive semantic-aware style transformation for blind face restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11896-11905, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 186, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 186, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 321, + 186, + 553, + 240 + ], + "type": "text", + "content": "[17] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arxiv:2310.00426, 2023. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 243, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 243, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 321, + 243, + 553, + 307 + ], + "type": "text", + "content": "[18] Wei-Ting Chen, Gurunandan Krishnan, Qiang Gao, SyYen Kuo, Sizhou Ma, and Jian Wang. Dsl-fqa: Assessing facial image quality via dual-set degradation learning and landmark-guided transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2931-2941, 2024. 2, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 309, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 309, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 321, + 309, + 553, + 363 + ], + "type": "text", + "content": "[19] Yu Chen, Ying Tai, Xiaoming Liu, Chunhua Shen, and Jian Yang. Fsrnet: End-to-end learning face super-resolution with facial priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2492-2501, 2018. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 365, + 553, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 365, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 321, + 365, + 553, + 418 + ], + "type": "text", + "content": "[20] Siying Cui, Jiankang Deng, Jia Guo, Xiang An, Yongle Zhao, Xinyu Wei, and Ziyong Feng. Idapter: Learning mixed features for tuning-free personalization of text-to-image models. arXiv preprint arXiv:2403.13535, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 422, + 553, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 422, + 553, + 463 + ], + "spans": [ + { + "bbox": [ + 321, + 422, + 553, + 463 + ], + "type": "text", + "content": "[21] Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning. In Proceedings of the International Conference on Learning Representations (ICLR). 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 467, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 467, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 321, + 467, + 553, + 521 + ], + "type": "text", + "content": "[22] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4690-4699, 2019. 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 321, + 523, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 523, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 321, + 523, + 553, + 567 + ], + "type": "text", + "content": "[23] Huiyu Duan, Qiang Hu, Jiarui Wang, Liu Yang, Zitong Xu, Lu Liu, Xiongkuo Min, Chunlei Cai, Tianxiao Ye, Xiaoyun Zhang, et al. Fine-grained user generated content video quality assessment. In m, 2025. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 321, + 568, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 568, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 321, + 568, + 553, + 622 + ], + "type": "text", + "content": "[24] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 321, + 624, + 555, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 624, + 555, + 678 + ], + "spans": [ + { + "bbox": [ + 321, + 624, + 555, + 678 + ], + "type": "text", + "content": "[25] S. Alireza Golestaneh, Saba Dadsetan, and Kris M. Kitani. No-reference image quality assessment via transformers, relative ranking, and self-consistency. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1220-1230, 2022. 7, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 321, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 321, + 681, + 553, + 713 + ], + "type": "text", + "content": "[26] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Pro" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10990" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 81, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 81, + 73, + 294, + 95 + ], + "type": "text", + "content": "ceedings of the Advances in Neural Information Processing Systems, 2014. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 96, + 295, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 96, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 61, + 96, + 295, + 151 + ], + "type": "text", + "content": "[27] Ke Gu, Guangtao Zhai, Min Liu, Xiaokang Yang, Wenjun Zhang, Xianghui Sun, Wanhong Chen, and Ying Zuo. Fisblim: A five-step blind metric for quality assessment of multiply distorted images. In Proceedings of the SiPS, pages 241-246, 2013. 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 152, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 152, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 61, + 152, + 294, + 196 + ], + "type": "text", + "content": "[28] Yuchao Gu, Xintao Wang, Liangbin Xie, Chao Dong, Gen Li, Ying Shan, and Ming-Ming Cheng. Vqfr: Blind face restoration with vector-quantized dictionary and parallel decoder. arXiv preprint arXiv:2205.06803, 2022. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 198, + 295, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 198, + 295, + 241 + ], + "spans": [ + { + "bbox": [ + 62, + 198, + 295, + 241 + ], + "type": "text", + "content": "[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 243, + 294, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 243, + 294, + 286 + ], + "spans": [ + { + "bbox": [ + 62, + 243, + 294, + 286 + ], + "type": "text", + "content": "[30] Xuanhua He, Quande Liu, Shengju Qian, Xin Wang, Tao Hu, Ke Cao, Keyu Yan, Man Zhou, and Jie Zhang. Id-animator: Zero-shot identity-preserving human video generation. arXiv preprint arXiv:2404.15275, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 288, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 288, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 62, + 288, + 294, + 342 + ], + "type": "text", + "content": "[31] Javier Hernandez-Ortega, Javier Galbally, Julian Fierrez, Rudolf Haraksim, and Laurent Beslay. Faceqnet: Quality assessment for face recognition based on deep learning. In 2019 International Conference on Biometrics (ICB), pages 1–8. IEEE, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 344, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 344, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 62, + 344, + 294, + 387 + ], + "type": "text", + "content": "[32] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 388, + 294, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 388, + 294, + 432 + ], + "spans": [ + { + "bbox": [ + 62, + 388, + 294, + 432 + ], + "type": "text", + "content": "[33] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 33: 6840-6851, 2020. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 434, + 294, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 434, + 294, + 477 + ], + "spans": [ + { + "bbox": [ + 62, + 434, + 294, + 477 + ], + "type": "text", + "content": "[34] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssim. In Proceedings of the IEEE International Conference on Pattern Recognition (ICPR), pages 2366-2369. IEEE, 2010. 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 479, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 479, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 62, + 479, + 294, + 533 + ], + "type": "text", + "content": "[35] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 535, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 535, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 62, + 535, + 294, + 590 + ], + "type": "text", + "content": "[36] Qiang Hu, Qihan He, Houqiang Zhong, Guo Lu, Xiaoyun Zhang, Guangtao Zhai, and Yanfeng Wang. Varfvv: View-adaptive real-time interactive free-view video streaming with edge computing. IEEE Journal on Selected Areas in Communications, 2025. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 591, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 591, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 62, + 591, + 294, + 657 + ], + "type": "text", + "content": "[37] Qiang Hu, Zihan Zheng, Houqiang Zhong, Sihua Fu, Li Song, Xiaoyun Zhang, Guangtao Zhai, and Yanfeng Wang. 4dgc: Rate-aware 4d gaussian compression for efficient streamable free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 875-885, 2025. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 658, + 294, + 712 + ], + "type": "text", + "content": "[38] Gary B. Huang, Manu Ramesh, Tamara Berg, and Erik Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. Technical Report 07-49, University of Massachusetts, Amherst, 2007. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 321, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 321, + 73, + 553, + 127 + ], + "type": "text", + "content": "[39] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 36:78723-78747, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 129, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 129, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 320, + 129, + 553, + 183 + ], + "type": "text", + "content": "[40] Haoyun Jiang, Fei Huang, Qiang Hu, Minmin Sun, Shuai Xiao, Yong Li, Junyang Lin, Jiangchao Yao, et al. Catekv: On sequential consistency for long-context llm inference acceleration. In Proceedings of International Conference on Machine Learning (ICML), 2025. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 184, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 184, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 320, + 184, + 553, + 237 + ], + "type": "text", + "content": "[41] Byungho Jo, Donghyeon Cho, In Kyu Park, and Sungeun Hong. Ifqa: Interpretable face quality assessment. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 3444-3453, 2023. 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 239, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 239, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 320, + 239, + 553, + 282 + ], + "type": "text", + "content": "[42] Le Kang, Peng Ye, Yi Li, and David Doermann. Convolutional neural networks for no-reference image quality assessment. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014. 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 283, + 553, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 283, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 320, + 283, + 553, + 336 + ], + "type": "text", + "content": "[43] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 4401-4410, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 338, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 338, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 320, + 338, + 553, + 403 + ], + "type": "text", + "content": "[44] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6007-6017, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 404, + 555, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 404, + 555, + 459 + ], + "spans": [ + { + "bbox": [ + 320, + 404, + 555, + 459 + ], + "type": "text", + "content": "[45] Ira Kemelmacher-Shlizerman, Steven M Seitz, Daniel Miller, and Evan Brossard. The megaface benchmark: 1 million faces for recognition at scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4873-4882, 2016. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 460, + 553, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 460, + 553, + 492 + ], + "spans": [ + { + "bbox": [ + 320, + 460, + 553, + 492 + ], + "type": "text", + "content": "[46] Deokyun Kim, Minseon Kim, Gihyun Kwon, and Dae-Shik Kim. Progressive face super-resolution via attention to facial landmark. arXiv preprint arXiv:1908.08239, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 493, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 493, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 320, + 493, + 553, + 514 + ], + "type": "text", + "content": "[47] Diederik P Kingma, Max Welling, et al. Auto-encoding variational bayes. 2013. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 515, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 515, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 320, + 515, + 553, + 569 + ], + "type": "text", + "content": "[48] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Matiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. Proceedings of Advances in neural information processing systems (Neurips), 36:36652-36663, 2023. 1, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 571, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 571, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 320, + 571, + 553, + 624 + ], + "type": "text", + "content": "[49] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1931-1941, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 625, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 625, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 320, + 625, + 553, + 657 + ], + "type": "text", + "content": "[50] Li et al. Agiqa-3k: An open database for ai-generated image quality assessment. arXiv preprint arXiv:2306.04717, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 320, + 658, + 553, + 712 + ], + "type": "text", + "content": "[51] Chunyi Li, Zicheng Zhang, Haoning Wu, Wei Sun, Xiongkuo Min, Xiaohong Liu, Guangtao Zhai, and Weisi Lin. Agiaq-3k: An open database for ai-generated image quality assessment. arXiv preprint arXiv: 2306.04717, 2023. 1, 2, 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "10991" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "type": "text", + "content": "[52] Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Linmiao Xu, and Suhail Doshi. Playground v2.5: Three insights towards enhancing aesthetic quality in text-to-image generation, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 118, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 118, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 61, + 118, + 294, + 173 + ], + "type": "text", + "content": "[53] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proceedings of International Conference on Machine learning (ICML), pages 12888-12900, 2022. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 175, + 295, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 175, + 295, + 218 + ], + "spans": [ + { + "bbox": [ + 62, + 175, + 295, + 218 + ], + "type": "text", + "content": "[54] Xiaoming Li, Ming Liu, Yuting Ye, Wangmeng Zuo, Liang Lin, and Ruigang Yang. Learning warped guidance for blind face restoration. In The European Conference on Computer Vision (ECCV), 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 220, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 220, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 62, + 220, + 294, + 285 + ], + "type": "text", + "content": "[55] Xiaoming Li, Wenyu Li, Dongwei Ren, Hongzhi Zhang, Meng Wang, and Wangmeng Zuo. Enhanced blind face restoration with multi-exemplar images and adaptive spatial feature fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2706-2715, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 287, + 294, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 287, + 294, + 340 + ], + "spans": [ + { + "bbox": [ + 62, + 287, + 294, + 340 + ], + "type": "text", + "content": "[56] Xiaoming Li, Shiguang Zhang, Shangchen Zhou, Lei Zhang, and Wangmeng Zuo. Learning dual memory dictionaries for blind face restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 343, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 343, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 62, + 343, + 294, + 387 + ], + "type": "text", + "content": "[57] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, Ming-Ming Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. arXiv preprint arXiv:2312.04461, 2023. 1, 2, 3, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 388, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 388, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 62, + 388, + 294, + 453 + ], + "type": "text", + "content": "[58] Zhimin Li, Jianwei Zhang, Qin Lin, Jiangfeng Xiong, Yanxin Long, Xinchi Deng, Yingfang Zhang, Xingchao Liu, Minbin Huang, Zedong Xiao, et al. Hunyuan-dit: A powerful multi-resolution diffusion transformer with fine-grained chinese understanding. arXiv preprint arXiv:2405.08748, 2024. 2, 3, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 456, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 456, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 62, + 456, + 294, + 521 + ], + "type": "text", + "content": "[59] Youwei Liang, Junfeng He, Gang Li, Peizhao Li, Arseniy Klimovskiy, Nicholas Carolan, Jiao Sun, Jordi Pont-Tuset, Sarah Young, Feng Yang, et al. Rich human feedback for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19401-19411, 2024. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 523, + 294, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 523, + 294, + 576 + ], + "spans": [ + { + "bbox": [ + 62, + 523, + 294, + 576 + ], + "type": "text", + "content": "[60] Zhang Lijun, Shao Xiaohu, Yang Fei, Deng Pingling, Zhou Xiangdong, and Shi Yu. Multi-branch face quality assessment for face recognition. In 2019 IEEE 19th International Conference on Communication Technology (ICCT), pages 1659–1664. IEEE, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 579, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 579, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 62, + 579, + 294, + 634 + ], + "type": "text", + "content": "[61] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of European conference on computer vision (ECCV), pages 740-755. Springer, 2014. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 635, + 294, + 688 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 635, + 294, + 688 + ], + "spans": [ + { + "bbox": [ + 62, + 635, + 294, + 688 + ], + "type": "text", + "content": "[62] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Ben Fei, Bo Dai, Wanli Ouyang, Yu Qiao, and Chao Dong. Diff-bir: Towards blind image restoration with generative diffusion prior. arXiv preprint arXiv:2308.15070, 2023. 1, 2, 3, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 692, + 295, + 713 + ], + "type": "text", + "content": "[63] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Proceedings of" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 321, + 74, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 339, + 74, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 74, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 339, + 74, + 553, + 95 + ], + "type": "text", + "content": "the Advances in Neural Information Processing Systems (NeurIPS), pages 34892-34916, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 321, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 321, + 96, + 553, + 139 + ], + "type": "text", + "content": "[64] Xiongkuo Min, Ke Gu, Guangtao Zhai, Jing Liu, Xiaokang Yang, and Chang Wen Chen. Blind quality assessment based on pseudo-reference image. IEEE Transactions on Multimedia (TMM), 20(8):2049-2062, 2018. 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 140, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 140, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 321, + 140, + 553, + 182 + ], + "type": "text", + "content": "[65] Xiongkuo Min, Guangtao Zhai, Ke Gu, Yutao Liu, and Xiaokang Yang. Blind image quality estimation via distortion aggravation. IEEE Transactions on Broadcasting, 64(2): 508-517, 2018. 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 184, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 184, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 321, + 184, + 553, + 227 + ], + "type": "text", + "content": "[66] Xiongkuo Min, Huiyu Duan, Wei Sun, Yucheng Zhu, and Guangtao Zhai. Perceptual video quality assessment: A survey. Science China Information Sciences, 67(11): 211301, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 228, + 553, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 228, + 553, + 260 + ], + "spans": [ + { + "bbox": [ + 321, + 228, + 553, + 260 + ], + "type": "text", + "content": "[67] Anish Mittal, Rajiv Soundararajan, and Alan C. Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal Processing Letters, 20(3):209-212, 2013. 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 262, + 553, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 262, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 321, + 262, + 553, + 326 + ], + "type": "text", + "content": "[68] Stylianos Moschoglou, Athanasios Papaioannou, Christos Sagonas, Jiankang Deng, Irene Kotsia, and Stefanos Zafeiriou. Agedb: the first manually collected, in-the-wild age database. In proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 51-59, 2017. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 327, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 327, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 321, + 327, + 553, + 392 + ], + "type": "text", + "content": "[69] Fu-Zhao Ou, Xingyu Chen, Ruixin Zhang, Yuge Huang, Shaoxin Li, Jilin Li, Yong Li, Liujuan Cao, and Yuan-Gen Wang. Sdd-fiqa: Unsupervised face image quality assessment with similarity distribution distance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7670–7679, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 394, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 394, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 321, + 394, + 553, + 437 + ], + "type": "text", + "content": "[70] Pablo Pernias, Dominic Rampas, Mats L Richter, Christopher J Pal, and Marc Aubreville. Würstchen: An efficient architecture for large-scale text-to-image diffusion models. arXiv preprint arXiv:2306.00637, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 438, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 438, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 321, + 438, + 553, + 491 + ], + "type": "text", + "content": "[71] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 3, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 493, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 493, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 321, + 493, + 553, + 525 + ], + "type": "text", + "content": "[72] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 526, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 526, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 321, + 526, + 553, + 568 + ], + "type": "text", + "content": "[73] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 571, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 571, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 321, + 571, + 553, + 624 + ], + "type": "text", + "content": "[74] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 625, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 625, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 321, + 625, + 553, + 689 + ], + "type": "text", + "content": "[75] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22500-22510, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 321, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 321, + 692, + 553, + 713 + ], + "type": "text", + "content": "[76] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "10992" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 72, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 81, + 72, + 296, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 72, + 296, + 127 + ], + "spans": [ + { + "bbox": [ + 81, + 72, + 296, + 127 + ], + "type": "text", + "content": "Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 128, + 296, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 128, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 61, + 128, + 296, + 205 + ], + "type": "text", + "content": "[77] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 35:36479-36494, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 205, + 295, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 205, + 295, + 237 + ], + "spans": [ + { + "bbox": [ + 61, + 205, + 295, + 237 + ], + "type": "text", + "content": "[78] B Series. Methodology for the subjective assessment of the quality of television pictures. Recommendation ITU-R BT, 500(13), 2012. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 239, + 295, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 295, + 271 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 295, + 271 + ], + "type": "text", + "content": "[79] Karen Simonyan and Andrew Zisserman. Very Deep Convolutional Networks for Large-Scale Image Recognition. arXiv e-prints, art. arXiv:1409.1556, 2014. 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 272, + 295, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 272, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 272, + 295, + 293 + ], + "type": "text", + "content": "[80] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv:2010.02502, 2020. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 295, + 295, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 295, + 295, + 348 + ], + "spans": [ + { + "bbox": [ + 62, + 295, + 295, + 348 + ], + "type": "text", + "content": "[81] Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3664-3673, 2020. 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 350, + 295, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 350, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 62, + 350, + 295, + 403 + ], + "type": "text", + "content": "[82] Shaolin Su, Hanhe Lin, Vlad Hosu, Oliver Wiedemann, Jinqiu Sun, Yu Zhu, Hantao Liu, Yanning Zhang, and Dietmar Saupe. Going the extra mile in face image quality assessment: A novel database and model. IEEE Transactions on Multimedia, 2023. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 404, + 295, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 404, + 295, + 470 + ], + "spans": [ + { + "bbox": [ + 62, + 404, + 295, + 470 + ], + "type": "text", + "content": "[83] Philipp Terhorst, Jan Niklas Kolf, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. Ser-fiq: Unsupervised estimation of face image quality based on stochastic embedding robustness. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5651-5660, 2020. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 471, + 295, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 471, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 62, + 471, + 295, + 525 + ], + "type": "text", + "content": "[84] Yuan Tian, Kaiyuan Ji, Rongzhao Zhang, Yankai Jiang, Chunyi Li, Xiaosong Wang, and Guangtao Zhai. Towards all-in-one medical image re-identification. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 30774–30786, 2025. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 526, + 295, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 526, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 62, + 526, + 295, + 569 + ], + "type": "text", + "content": "[85] Yuri Viazovetskyi, Vladimir Ivashkin, and Evgeny Kashin. Stylegan2 distillation for feed-forward image manipulation. Proceedings of the European conference on computer vision (ECCV), pages 170-186, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 571, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 571, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 62, + 571, + 295, + 624 + ], + "type": "text", + "content": "[86] Jiarui Wang, Huiyu Duan, Jing Liu, Shi Chen, Xiongkuo Min, and Guangtao Zhai. Aigciqa2023: A large-scale image quality assessment database for ai generated images: from the perspectives of quality, authenticity and correspondence. 2023. 1, 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 625, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 625, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 62, + 625, + 295, + 668 + ], + "type": "text", + "content": "[87] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. arXiv preprint arXiv:2305.07015, 2023. 1, 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 295, + 712 + ], + "type": "text", + "content": "[88] Jiarui Wang, Huiyu Duan, Guangtao Zhai, and Xiongkuo Min. Understanding and evaluating human preferences for ai generated images with instruction tuning. arXiv preprint arXiv:2405.07346, 2024. 7, 8" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 320, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 553, + 116 + ], + "type": "text", + "content": "[89] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin C.K. Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. 2024. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 118, + 553, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 118, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 320, + 118, + 553, + 151 + ], + "type": "text", + "content": "[90] Liangbin Xie. Xintao Wang, Honglun Zhang, Chao Dong, and Ying Shan. Vfhq: A high-quality dataset and benchmark for video face super-resolution, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 320, + 152, + 553, + 195 + ], + "type": "text", + "content": "[91] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 2, 3, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 198, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 198, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 320, + 198, + 553, + 252 + ], + "type": "text", + "content": "[92] Tao Wang, Kaihao Zhang, Xuanxi Chen, Wenhan Luo, Jiankang Deng, Tong Lu, Xiaochun Cao, Wei Liu, Hongdong Li, and Stefanos Zafeiriou. A survey of deep face restoration: Denoise, super-resolution, deblur, artifact removal. arXiv preprint arXiv:2211.02831, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 254, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 254, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 320, + 254, + 553, + 297 + ], + "type": "text", + "content": "[93] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In International Conference on Computer Vision Workshops (ICCVW). 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 298, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 298, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 320, + 298, + 553, + 342 + ], + "type": "text", + "content": "[94] Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. Towards real-world blind face restoration with generative facial prior. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 3, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 344, + 553, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 344, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 320, + 344, + 553, + 388 + ], + "type": "text", + "content": "[95] Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. Towards real-world blind face restoration with generative facial prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 389, + 553, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 389, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 320, + 389, + 553, + 432 + ], + "type": "text", + "content": "[96] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing (TIP), 13(4):600-612, 2004. 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 434, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 434, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 320, + 434, + 553, + 499 + ], + "type": "text", + "content": "[97] Zhixin Wang, Ziying Zhang, Xiaoyun Zhang, Huangjie Zheng, Mingyuan Zhou, Ya Zhang, and Yanfeng Wang. Dr2: Diffusion-based robust degradation remover for blind face restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1704-1713, 2023. 2, 3, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 501, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 501, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 320, + 501, + 553, + 567 + ], + "type": "text", + "content": "[98] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15943-15953, 2023. 2, 3, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 568, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 568, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 320, + 568, + 553, + 623 + ], + "type": "text", + "content": "[99] Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xinglong Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. Diffir: Efficient diffusion model for image restoration. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13095-13105, 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 624, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 624, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 315, + 624, + 553, + 678 + ], + "type": "text", + "content": "[100] Weihao Xia, Yujiu Yang, Jing-Hao Xue, and Baoyuan Wu. Tedigan: Text-guided diverse face image generation and manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2256-2265, 2021. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[101] Guangxuan Xiao, Tianwei Yin, William T Freeman, Frédo Durand, and Song Han. Fastcomposer: Tuning-free multisubject image generation with localized attention. Inter-" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10993" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 710 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 81, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 81, + 72, + 296, + 95 + ], + "type": "text", + "content": "national Journal of Computer Vision (IJCV), 133(3):1175-1194, 2025. 2, 3, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 149 + ], + "type": "text", + "content": "[102] Jingtao Xu, Peng Ye, Qiaohong Li, Haiqing Du, Yong Liu, and David S. Doermann. Blind image quality assessment based on high order statistics aggregation. IEEE Transactions on Image Processing (TIP), 25:4444-4457, 2016. 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 295, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 295, + 217 + ], + "type": "text", + "content": "[103] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 36, 2024. 2, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 295, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 295, + 284 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 295, + 284 + ], + "type": "text", + "content": "[104] Liu Yang, Huiyu Duan, Long Teng, Yucheng Zhu, Xiaohong Liu, Menghan Hu, Xiongkuo Min, Guangtao Zhai, and Patrick Le Callet. Aigcoiaq2024: Perceptual quality assessment of ai generated omnidirectional images. In 2024 IEEE International Conference on Image Processing (ICIP), pages 1239-1245. IEEE, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 285, + 296, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 285, + 296, + 329 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 296, + 329 + ], + "type": "text", + "content": "[105] Shuo Yang, Ping Luo, Chen-Change Loy, and Xiaou Tang. Wider face: A face detection benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5525–5533, 2016. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 295, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 295, + 384 + ], + "type": "text", + "content": "[106] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. MANIQA: Multi-dimension Attention Network for No-Reference Image Quality Assessment. arXiv e-prints, art. arXiv:2204.08958, 2022. 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 386, + 295, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 386, + 295, + 440 + ], + "spans": [ + { + "bbox": [ + 56, + 386, + 295, + 440 + ], + "type": "text", + "content": "[107] Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. Gan prior embedded network for blind face restoration in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 672-681, 2021. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 441, + 295, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 441, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 56, + 441, + 295, + 485 + ], + "type": "text", + "content": "[108] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 487, + 295, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 487, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 56, + 487, + 295, + 518 + ], + "type": "text", + "content": "[109] Zongsheng Yue and Chen Change Loy. Difface: Blind face restoration with diffused error contraction. arXiv preprint arXiv:2212.06512, 2022. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 520, + 295, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 520, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 56, + 520, + 295, + 552 + ], + "type": "text", + "content": "[110] Guangtao Zhai and Xiongkuo Min. Perceptual image quality assessment: a survey. Science China Information Sciences, 63:1-52, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 554, + 295, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 554, + 295, + 597 + ], + "spans": [ + { + "bbox": [ + 56, + 554, + 295, + 597 + ], + "type": "text", + "content": "[111] Lin Zhang, Lei Zhang, and Alan C. Bovik. A feature-enriched completely blind image quality evaluator. IEEE Transactions on Image Processing (TIP), 24(8):2579-2591, 2015. 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 599, + 295, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 599, + 295, + 654 + ], + "spans": [ + { + "bbox": [ + 56, + 599, + 295, + 654 + ], + "type": "text", + "content": "[112] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 655, + 295, + 710 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 655, + 295, + 710 + ], + "spans": [ + { + "bbox": [ + 56, + 655, + 295, + 710 + ], + "type": "text", + "content": "[113] Shangchen Zhou, Kelvin Chan, Chongyi Li, and Chen Change Loy. Towards robust blind face restoration with codebook lookup transformer. Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 35:30599-30611, 2022. 2, 3, 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "10994" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/FALCON_ Resolving Visual Redundancy and Fragmentation in High-resolution Multimodal Large Language Models via Visual Registers/3c258128-6a38-46a9-bf2a-aabe422e5840_content_list.json b/2025/FALCON_ Resolving Visual Redundancy and Fragmentation in High-resolution Multimodal Large Language Models via Visual Registers/3c258128-6a38-46a9-bf2a-aabe422e5840_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fd00b83f33d5d04035b26b9282891ef797ce58f9 --- /dev/null +++ b/2025/FALCON_ Resolving Visual Redundancy and Fragmentation in High-resolution Multimodal Large Language Models via Visual Registers/3c258128-6a38-46a9-bf2a-aabe422e5840_content_list.json @@ -0,0 +1,1423 @@ +[ + { + "type": "text", + "text": "FALCON: Resolving Visual Redundancy and Fragmentation in High-resolution Multimodal Large Language Models via Visual Registers", + "text_level": 1, + "bbox": [ + 98, + 128, + 898, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Renshan Zhang $^{1}$ , Rui Shao $^{1\\dagger}$ , Gongwei Chen $^{1}$ , Miao Zhang $^{1}$ , Kaiwen Zhou $^{2}$ , Weili Guan $^{1}$ , Liqiang Nie $^{1\\dagger}$ $^{1}$ Harbin Institute of Technology, Shenzhen $^{2}$ Huawei Noah's Ark Lab", + "bbox": [ + 220, + 208, + 776, + 262 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zhangrenshan@stu.hit.edu.cn shaorui@hit.edu.cn", + "bbox": [ + 264, + 265, + 730, + 279 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "chengongwei@hit.edu.cn nieliqiang@gmail.com", + "bbox": [ + 284, + 282, + 722, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/JiuTian-VL/FALCON", + "bbox": [ + 316, + 300, + 674, + 314 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 349, + 326, + 366 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The incorporation of high-resolution visual input equips multimodal large language models (MLLMs) with enhanced visual perception capabilities for real-world tasks. However, most existing high-resolution MLLMs rely on a cropping-based approach to process images, which leads to fragmented visual encoding and a sharp increase in redundant tokens. To tackle these issues, we propose the FALCON model. FALCON introduces a novel visual register technique to simultaneously: 1) Eliminate redundant tokens at the stage of visual encoding. To directly address the visual redundancy present in the output of vision encoder, we propose a Register-based Representation Compacting (ReCompact) mechanism. This mechanism introduces a set of learnable visual registers designed to adaptively aggregate essential information while discarding redundancy. It enables the encoder to produce a more compact visual representation with a minimal number of output tokens, thus eliminating the need for an additional compression module. 2) Ensure continuity in visual encoding. To address the potential encoding errors caused by fragmented visual inputs, we develop a Register Interactive Attention (ReAtten) module. This module facilitates effective and efficient information exchange across sub-images by enabling interactions between visual registers. It ensures the continuity of visual semantics throughout the encoding. We conduct comprehensive experiments with FALCON on high-resolution benchmarks across a wide range of scenarios. FALCON demonstrates superior performance with a remarkable 9-fold reduction in visual tokens.", + "bbox": [ + 88, + 381, + 485, + 819 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 821, + 223, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, multimodal large language models (MLLMs) [2, 9, 30, 32, 59, 60] have made significant advancements in", + "bbox": [ + 89, + 847, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d729d5f337aba80dea70000a3b80a55cd315708e8941f483cec9704ffe13e941.jpg", + "image_caption": [ + "Figure 1. Comparison between existing high-resolution MLLMs and FALCON. FALCON eliminates visual redundancy and ensures the continuity of high-resolution encoding through the use of ReCompact and ReAtten mechanisms." + ], + "image_footnote": [], + "bbox": [ + 517, + 349, + 905, + 583 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "performing various vision-language tasks [45, 46, 61], facilitating more natural human-machine interactions. Early MLLMs [9, 32, 59] typically processed images at a fixed low resolution, such as $224 \\times 224$ . This limited their capabilities in many real-world applications that require high-resolution input. To equip MLLMs with high-resolution visual perception, recent efforts [15, 17, 33, 58, 67] have introduced a cropping-based high-resolution processing method. Specifically, As shown in Fig. 1(b), this approach crops high-resolution images into multiple non-overlapping sub-images that match the pre-trained resolution of the encoder. Each sub-image is processed independently by the vision encoder. Then the resulting visual features are concatenated and fed into the large language model (LLM).", + "bbox": [ + 509, + 656, + 906, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although effective, cropping-based methods still face issues of: 1) Visual Redundancy. The number of visual to-", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding authors", + "bbox": [ + 109, + 886, + 238, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23530", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "kens grows dramatically with increasing resolution. While this enriches the visual details, it also leads to an increase in redundant tokens. These redundant tokens significantly increase the computational burden on LLM and interfere with the processing of visual information [65]. To address these issues, as illustrated in Fig. 1(b), existing methods typically rely on additional compression techniques. Representative approaches include combining learnable queries with cross-attention [1, 28, 60] or simply employing pooling layers [48, 55]. However, these methods either demand substantial training resources or fail to fully address redundancy. 2) Visual Fragmentation. Cropping an image into multiple non-overlapping sub-images may disrupt visual semantics and potentially lead to encoding errors. As shown in Fig. 1(a), in an OCR task, the word \"pineapple\" may be incorrectly segmented into \"pine\" and \"apple\", causing the LLM to misinterpret the word. Similarly, fragmentation can also hinder the identification of objects. For example, the \"rubin vase\" pattern illustrated in Fig. 1(a) depicts both a vase and two faces simultaneously. However, the fragmentation could lead to the vase being unrecognizable.", + "bbox": [ + 89, + 90, + 480, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these issues, as shown in Fig. 1(c), we propose the FALCON model. FALCON introduces a novel visual register technique to simultaneously: 1) Eliminate redundant tokens during visual encoding. To eliminate redundancy directly at the stage of visual encoding, we propose a novel Register-based Representation Compacting (ReCompact) mechanism. This mechanism introduces a set of learnable visual registers as additional inputs to the vision encoder. The registers are designed to adaptively aggregate essential information from image tokens and discard redundant content. Then the outputs corresponding to the visual registers are considered as a non-redundant visual representation, serving as the compact input for the LLM instead of a large number of image tokens. This mechanism enables the vision encoder to convey comprehensive visual information in a compact representation using minimal tokens, thus removing the need for an additional compression module. 2) Ensure continuity in visual encoding. To avoid the potential encoding errors caused by fragmented visual inputs, we develop a novel Register Interactive Attention (ReAtten) module. This module leverages the compact and comprehensive representation produced by visual registers to facilitate information exchange across sub-images. It gathers visual registers from different sub-images and employs a specialized cross-ViT attention module to enable interactions among them. Such interaction enables thorough information exchange across the entire image, thus ensuring visual continuity throughout the encoding process. Additionally, the compactness of the visual registers further enhances the efficiency of these interactions. Consequently, this module provides LLMs with a reliable foundation to accurately interpret visual content and reduce hallucinations.", + "bbox": [ + 91, + 417, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Note that the widely used Abstractor-like modules [28, 33, 58, 60], including QFormer [9], ReSampler [1], follow a similar approach by using query tokens to extract information from image tokens after visual encoding. The key difference is that the visual register mechanism directly adopts the pretrained ViT to aggregate information into a set of specified register tokens. This design is based on the observation that a well-pretrained ViT can naturally gather global information into some specific tokens through self-attention [10]. It offers the visual register mechanism several strengths in terms of: 1) Training Efficiency. The substantial number of randomly initialized parameters in Abstractor-like modules requires extensive data for pretraining. For instance, QFormer [22], Abstractor [60], and Resampler [1] use large corpora of 129M, 400M, and 1.4B samples, respectively. In contrast, by leveraging well-pretrained ViT parameters, the visual register mechanism requires only 3.5M samples for pretraining. The ablation results in Fig. 4 further demonstrate this advantages. 2) Addressing Visual Fragmentation. By leveraging the compact and comprehensive representation produced by visual registers, the proposed ReAtten module can effectively facilitate information exchange across sub-images. This is an important function that is notably lacking in Abstractor-like modules. Combined with ReCompact, it enables the visual register mechanism to simultaneously address the issues of visual redundancy and fragmentation. The ablation results in Tab. 5 validate the effectiveness of ReAtten.", + "bbox": [ + 511, + 90, + 903, + 512 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We summarize our main contributions as follows:", + "bbox": [ + 532, + 515, + 861, + 527 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To address the issue of visual redundancy directly at the encoding phase, we introduce the Register-based Representation Compacting (ReCompact) mechanism. This mechanism enables vision encoder to capture rich visual information using only a minimal number of tokens, effectively eliminating redundancy and obviating the need for additional compression modules.", + "- To address the issue of visual fragmentation caused by image cropping, we propose the register interactive attention (ReAtten) module. This module ensures the semantic continuity of high-resolution visual encoding in an efficient and effective manner.", + "- We introduce FALCON, a model that leverages the proposed visual register technique. Comprehensive evaluations on high-resolution visual understanding benchmarks highlight the superiority of FALCON. Notably, FALCON achieves exceptional results with a remarkable reduction in visual tokens by $9 \\times$ ." + ], + "bbox": [ + 513, + 531, + 903, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 820, + 653, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Multimodal Large Language Models", + "text_level": 1, + "bbox": [ + 513, + 847, + 831, + 863 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Research on MLLMs [1, 2, 30, 47, 59, 60] has made significant strides in recent years. By incorporating image in", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23531", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2f27ac45349ad0f7f21ad2a796584b3c4076140f766605856e347b5f07931124.jpg", + "image_caption": [ + "Figure 2. Overall framework of the proposed FALCON. The model introduces a set of learnable visual registers, which are fed into the ViT along with the image tokens from each sub-image. During the visual encoding, the ReCompact mechanism enables information from the image tokens to be aggregated into the visual registers, while the proposed ReAtten facilitates information exchange among the visual registers across different sub-images. Ultimately, these visual registers provide a compact visual representation as input to the LLM." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 903, + 256 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "puts into pretrained LLMs [8, 51, 52], MLLMs demonstrate strong capabilities across various vision-language tasks. Typically, these models use a pretrained vision encoder [43, 62] to transform input images into a set of visual tokens. These tokens are then projected into the input space of the LLM through a projection module (e.g., MLP [30, 32], QFormer [9] and Abstractor [60]) and processed alongside text inputs. Building on the progress of MLLMs, recent studies have shifted from designing specialized methods for certain tasks [44-46] to employing MLLMs as generalist agents, capable of handling diverse practical tasks such as document analysis [15, 57, 64], online video assistant [24], retrieval-augmented generation [40], game agents [27, 29], robot manipulation [18, 21, 39], and GUI agent [3, 5, 54]. Despite these advances, MLLMs still struggle with the fixed low resolution of pretrained image encoders, which can limit their effectiveness in handling high-resolution images.", + "bbox": [ + 91, + 330, + 483, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. High-resolution MLLMs", + "text_level": 1, + "bbox": [ + 89, + 603, + 316, + 618 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the limitation of MLLMs in handling only low-resolution visual inputs, early studies attempted to either directly incorporate vision encoders that were pre-trained on higher-resolution images [36, 38], or to improve the resolution of pre-trained encoders through additional training [1, 60]. While achieving some progress, they still remain restricted by the fixed resolutions. To accommodate the varying resolutions and aspect ratios, cropping-based techniques have been widely adopted in recent works [15, 17, 33, 58, 67]. However, these approaches still face challenges related to visual redundancy and fragmentation.", + "bbox": [ + 89, + 626, + 482, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To mitigate the issue of redundancy, existing models typically require an additional compression module. For example, Abstractor-like modules utilizing learnable queries combined with cross-attention are widely adopted [1, 28, 33, 58, 60]. These modules typically require pretraining on massive amounts of data, which diminishes their practical efficacy. In contrast, some approaches use pooling", + "bbox": [ + 89, + 795, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "layers [48, 55] or convolutional layers [15, 36]. Though resource-efficient, they often fail to address the issue of redundancy. In this work, we present an initial exploration to directly eliminate redundancy during the visual encoding phase, thereby removing the need for additional compression. To address the issue of fragmentation, several concurrent studies also make preliminary investigation. TextMonkey [33] uses Shifted Window Attention [35] to exchange information between sub-images. But the exchange is still limited to local regions. MiniMonkey [16] proposed a Complementary Image Pyramid strategy to generate complementary cropping schemes. However, this strategy inevitably encodes the same visual content multiple times, further exacerbating visual redundancy. In this work, we propose the ReAtten module to facilitate continuous visual encoding through the interaction among visual registers.", + "bbox": [ + 511, + 330, + 906, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. FALCON", + "text_level": 1, + "bbox": [ + 511, + 588, + 619, + 603 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall framework of FALCON is illustrated in Fig. 2. The method begins with shape-adaptive cropping [58], where the image is cropped into multiple non-overlapping sub-images, each matching the resolution of the pre-trained vision encoder. Additionally, a resized version of the full image is used to supplement global layout information. To directly eliminate redundancy during visual encoding, we propose Register-based Representation Compacting (Re-Compact) mechanism. This mechanism introduces a set of learnable visual registers, which are paired with image tokens from each sub-image and fed into the vision encoder to capture rich visual information. To ensure the continuity of visual semantics throughout the encoding, a novel Register Interactive Attention (ReAtten) module is integrated into the Vision Transformer (ViT) [11] to facilitate information exchange between sub-images via the visual registers. Finally, the compact visual representations produced by the visual registers are processed through a simple MLP module before being fed into the LLM for further analysis.", + "bbox": [ + 511, + 612, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23532", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Mitigating Visual Redundancy via ReCompact", + "text_level": 1, + "bbox": [ + 89, + 90, + 483, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Existing MLLMs typically use a ViT-based vision encoder, such as CLIP [43]. For an input image of size $H \\times W$ , it is patched into multiple $P \\times P$ patches. Each patch is treated as an individual image token, resulting in $N = \\frac{H}{P} \\times \\frac{W}{P}$ tokens, denoted as $I = \\{i_1, \\dots, i_N\\}$ . In most current MLLMs, output features corresponding to all image tokens are usually preserved in their entirety. However, when high-resolution images are included, the number of image tokens increases dramatically, placing a substantial burden on the computational resources of LLMs. Consequently, most high-resolution MLLMs necessitate an additional compression module, which can be resource-inefficient. Previous research has shown that most image tokens contain only localized information from their surrounding areas [10]. It suggests that a significant portion of these tokens may be redundant, particularly those originating from background regions. This observation motivates us to directly eliminate redundancy in the output of the vision encoder, thereby reducing the number of visual tokens and avoiding the need for an additional compression module.", + "bbox": [ + 89, + 112, + 483, + 414 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To achieve this, we propose a novel Register-based Representation Compacting (ReCompact) mechanism. This mechanism employs a set of visual registers to aggregate essential information from image tokens, resulting in a more compact visual representation. Specifically, as shown in Fig. 2, we introduce a set of learnable visual registers, denoted by $R = \\{r_1,\\dots ,r_M\\}$ , where $M$ represents the number of visual registers, and $M\\ll N$ . Given a single input image, the image tokens are concatenated with visual registers to create a new input for ViT, represented as $V = \\mathrm{Concat}(I,R) = \\{i_1,\\dots ,i_N,r_1,\\dots ,r_M\\}$ . During visual encoding, the self-attention mechanism in ViT enables the visual register to adaptively aggregate important information from image tokens while discarding redundant content. Denote $X_{k,l}\\in \\mathcal{R}^{(N + M)\\times D}$ as the input hidden states of the $k$ -th sub-image at the $l$ -th layer, where $D$ is the dimensionality. The self-attention is formulated as:", + "bbox": [ + 89, + 415, + 483, + 674 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {X} _ {k, l} = \\operatorname {S e l f - A t t e n} \\left(X _ {k, l}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {X _ {k , l} X _ {k , l} ^ {T}}{\\sqrt {D _ {\\text {k e y}}}}\\right) X _ {k, l} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 684, + 483, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Notably, we do not apply any attention masks for specific constraints. Our experiments show that pretraining on a modest amount of data is sufficient to effectively adapt the pretrained ViT to the proposed ReCompact, enabling it to aggregate essential information within the visual registers. After visual encoding, The output features corresponding to the visual registers, denoted as $F = \\{f_{1},\\dots ,f_{M}\\}$ , are then used as the visual representation of the image and fed into the LLM for various vision-language tasks.", + "bbox": [ + 89, + 733, + 482, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To further adapt the ReCompact mechanism for cropping-based high-resolution image perception, we share", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the same set of visual registers as input for different sub-images. The input for the $k$ -th sub-image is formulated as $V_{k} = \\mathrm{Concat}(I_{k},R)$ , where $k = 1,2,\\dots ,N_{c}$ and $N_{c}$ denotes the total number of sub-images. After encoding, the outputs of the visual registers for all sub-images are concatenated, yielding $F_{hr} = \\mathrm{Concat}(F_1,F_2,\\dots ,F_{N_c})$ which serves as the input to the LLM. The ReCompact mechanism enables the vision encoder to represent rich visual information using only a minimal number of $M$ tokens per sub-image, which significantly reduces the visual redundancy and eliminates the need for additional compression.", + "bbox": [ + 511, + 90, + 906, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Ensuring Visual Continuity via ReAtten", + "text_level": 1, + "bbox": [ + 511, + 266, + 856, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When using the cropping-based high-resolution image encoding method, each sub-image is independently encoded by the vision encoder. This process can disrupt semantic coherence and result in the misrepresentation of visual concepts. To preserve the continuity of visual semantics, it is crucial to enable information exchange between sub-images during encoding. A naive approach would be to directly concatenate the image features from all sub-images and apply self-attention. However, the quadratic computational complexity makes this approach impractical. In Sec. 3.1, we introduced the ReCompact mechanism, where the visual registers effectively aggregate rich visual information while discarding redundancy during visual encoding, allowing them to serve as compact yet comprehensive image representations. Consequently, interactions between visual registers can naturally facilitate sufficient information exchange across different sub-images in an efficient manner.", + "bbox": [ + 511, + 287, + 906, + 544 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To enable information exchange between visual registers across different sub-images, we introduce a novel Register Interactive Attention (ReAtten) module. As shown in Fig. 2, the ReAtten module is integrated into each of the $L$ layers of the ViT. After the self-attention process in Eq. (1), we acquire $\\hat{X}_{k,l} = \\{\\hat{X}_{k,l}^{I},\\hat{X}_{k,l}^{R}\\}$ , where $\\hat{X}^I\\in \\mathcal{R}^{N\\times D}$ and $\\hat{X}^R\\in \\mathcal{R}^{M\\times D}$ represent the hidden states corresponding to image tokens and visual registers, respectively. The states of the visual registers from different sub-images are concatenated to form $\\hat{X}_l^R = \\mathrm{Concat}\\left(\\{\\hat{X}_{k,l}^R\\}_{k = 1}^{N_c}\\right)\\in$ $\\mathcal{R}^{M\\cdot N_c\\times D}$ . Then these concatenated registers are processed via ReAtten to enable information exchange across subimages, which can be formulated as:", + "bbox": [ + 511, + 545, + 908, + 752 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\overline {{X}} _ {l} ^ {R} = \\mathrm {R e A t t e n} (\\hat {X} _ {l} ^ {R}) = \\hat {X} _ {l} ^ {R} + \\mathrm {C r o s s - V i T - A t t e n} (\\hat {X} _ {l} ^ {R}) (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 762, + 906, + 785 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where Cross-ViT-Atten is implemented as a self-attention layer. Once the information exchange is completed, the registers and image tokens from a same sub-image are concatenated and then passed into the feedforward network (FFN) for transformation, yielding the output of the $l$ -th layer:", + "bbox": [ + 511, + 792, + 906, + 869 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nX _ {k, l + 1} = \\operatorname {F F N} \\left(\\operatorname {C o n c a t} \\left(\\hat {X} _ {k, l} ^ {I}, \\bar {X} _ {k, l} ^ {R}\\right)\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 878, + 906, + 905 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23533", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7fea22fdc3cc3b493e312b32f4a98ebdb9f316bfc9d08bbff01a1afa4a45a549.jpg", + "image_caption": [ + "Figure 3. The progressive training pipeline. **Stage 0:** Align the frozen vision and language backbones by training only the MLP and registers. **Stage 1:** Tune all parameters with caption and OCR data on low-resolution for coarse alignment. **Stage 2:** Integrate high-resolution input along with fine-grained captions and OCR to achieve refined alignment. **Stage 3:** Froze ViT and introduce ReAtten module for high-resolution instruction tuning." + ], + "image_footnote": [], + "bbox": [ + 109, + 87, + 460, + 339 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By integrating ReAtten into each layer of the ViT model, FALCON enables effective information exchange between cropped sub-images through visual registers, thereby ensuring semantic continuity of the high-resolution image throughout the visual encoding process.", + "bbox": [ + 89, + 452, + 483, + 529 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Learning Expressive Visual Representation via Progressive Training Pipeline", + "text_level": 1, + "bbox": [ + 89, + 541, + 482, + 573 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To empower the ReCompact mechanism with the capability to capture rich visual patterns and to enable smoother initialization of the ReAtten module under high-resolution conditions, we develop a specialized progressive training pipeline. As illustrated in Fig. 3, this pipeline progressively transitions from low to high resolution across four stages.", + "bbox": [ + 89, + 580, + 482, + 671 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Stage 0: Static Alignment on Low-resolution. Since both the visual registers and MLP projector are randomly initialized, in stage 0, we aim to achieve a initial alignment between the pretrained vision encoder and the LLM. To facilitate this, we begin by using low-resolution images as inputs, which are directly resized to match the resolution of the vision encoder. We then statically pretrain the model by keeping both the vision encoder and LLM frozen, tuning only the visual registers and MLP using image caption data.", + "bbox": [ + 89, + 672, + 482, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Stage 1: Coarse Alignment on Low-resolution. To better adapt the pre-trained vision encoder to our newly introduced ReCompact mechanism, in stage 1, we unfreeze all model parameters to enable a more thorough tuning. To facilitate the visual registers to progressively learn capturing visual information from coarse to fine granularity, we con", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tinue to use low-resolution inputs and employ high-quality, extended captions rich in visual content, along with full-text OCR data for training. This design allows the registers to acquire holistic scene understanding ability.", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Stage 2: Refined Alignment on High-resolution. To further enhance the capacity of the visual register to capture fine-grained visual details, in stage 2, we transition the model to process high-resolution image for refined alignment. Additionally, we employ datasets that includes a series of fine-grained captions and OCR tasks, such as regional captions and text location. This facilitates the visual registers to capture comprehensive fine-grained details.", + "bbox": [ + 511, + 152, + 903, + 271 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Stage 3: Instruction Tuning on High-resolution. To adapt the well-aligned visual register for a wide range of real-world tasks, in stage 3, we perform instruction tuning with a mixture of datasets covering various tasks, such as VQA, document understanding, and science reasoning. Additionally, the proposed ReAtten module is integrated to address visual fragmentation caused by cropping-based high-resolution image encoding. Since the vision encoder with ReCompact is well-aligned, the vision encoder is kept frozen and other parameters including ReAtten are trainable during instruction tuning. Notably, to facilitate a smoother initialization for the newly introduced ReAtten modules, we initialize parameters of each ReAtten module with the ViT self-attention parameters at the corresponding depth. This design enables ReAtten to effectively leverage pre-trained parameters, avoiding the need for extensive fine-tuning.", + "bbox": [ + 511, + 272, + 906, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 527, + 645, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We implement FALCON with SigLIP-L/16-384px [62] as the vision encoder and Llama-3.1-8B-Instruct [12] as the LLM backbone. A two-layer MLP module with GeLU [13] activation is served as the vision-language projection layer. The number of visual registers is set to 64, achieving remarkable visual token compression rates of $9 \\times$ . The maximum number of sub-images is set to 16.", + "bbox": [ + 511, + 551, + 905, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Results", + "text_level": 1, + "bbox": [ + 511, + 667, + 714, + 683 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate FALCON using the recently proposed MMERealWorld [66] benchmark, a comprehensive and largescale benchmark specifically designed to assess the high-resolution understanding capabilities of MLLMs. MMERealWorld comprises 13,366 high-resolution images with an average resolution of $2,000 \\times 1,500$ pixels, spanning five domains: OCR in the wild, Remote Sensing, Diagrams and Tables, Monitoring, and Autonomous Driving. This benchmark is well-suited for evaluating the practical high-resolution understanding capabilities of MLLMs.", + "bbox": [ + 511, + 688, + 903, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Tab. 1, FALCON significantly outperform other advanced MLLMs in both perception and reasoning tasks. FALCON achieves state-of-the-art performance with only 64 tokens per sub-image, resulting in a remarkable vi", + "bbox": [ + 511, + 839, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23534", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Evaluation on MME-RealWorld Benchmark. \"OCR\", \"RS\", \"DT\", \"MO\", and \"AD\" stand for OCR in the Wild, Remote Sensing, Diagram and Table, Monitoring, and Autonomous Driving, respectively. \"Avg\" denotes the weighted average accuracy across all samples, while \"Avg-C\" denotes the weighted average accuracy across different domains. \"Sizev\" indicates the image resolution supported by the model, where $\\times n$ signifies a cropping-based high-resolution processing approach with up to $n$ sub-images. \"Tokenc\" denotes the number of visual tokens after compression. † For models utilizing hybrid vision encoders, \"Sizev\" represents the setting of encoder with the highest resolution. We gray out the models that are either closed-source or have at least twice the number of visual tokens compared to ours. Among these, the lighter gray denotes scores below ours. The underline indicates that the second performance is achieved. Notably, FALCON achieves state-of-the-art performance with just 64 tokens per sub-image, surpassing most models that have large amount of visual tokens and many advanced closed-source commercial models.", + "bbox": [ + 89, + 88, + 906, + 215 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/16635d14cdbcc4f4fe268638884ac70a1a1f19331fea4bda213dfe27f740dd5a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TypeMethod\\( Size^v \\)\\( Token^c \\)PerceptionReasoning
Task SplitOCRRSDTMOADAvgAvg-COCRDTMOADAvgAvg-C
Visual Tokens\\(\\nabla 1024\\)MiniGPT-v2 [4]\\( 448^2 \\)25639.023.320.419.326.026.925.630.020.416.923.723.022.7
Qwen-VL-Chat [1]\\( 448^2 \\)25632.415.115.622.115.120.820.128.613.616.524.622.020.8
LLaVA1.5-13B [30]\\( 336^2 \\)57644.123.320.220.526.128.426.830.220.827.524.825.525.8
ShareGPT4V-13B [6]\\( 336^2 \\)57644.623.120.219.326.128.426.626.020.827.324.624.624.7
TextMonkey [33]\\( 448^2 \\times 4 \\)256 × 437.311.75.916.114.318.217.130.42.24.420.016.014.3
Mini-Gemini-7B-HD† [26]\\( 336^2 \\times 4 \\)57642.031.322.334.224.831.130.935.424.625.923.326.127.3
DeepSeek-VL† [36]\\( 1024^2 \\)57649.625.523.427.033.433.131.845.223.816.727.328.028.3
MiniCPM-V 2.5 [56]\\( 448^2 \\times 9 \\)96 × 966.827.752.838.734.247.444.044.031.837.031.034.536.0
Monkey [28]\\( 448^2 \\times 4 \\)256 × 454.625.032.528.029.736.334.027.220.827.333.028.827.1
FALCON\\( 384^2 \\times 16 \\)64 × 1666.449.647.240.436.550.348.055.439.244.434.540.743.4
Visual Tokens\\(\\nabla 2048\\)LLaVA-Next-Qwen72B [31]\\( 336^2 \\times 4 \\)576 × 437.129.127.729.418.029.028.317.234.227.329.727.927.1
LLaVA-Next-liama3 [31]\\( 336^2 \\times 4 \\)576 × 447.925.426.619.518.730.127.655.223.421.130.732.132.6
mPLUG-DocOwl 1.5 [15]\\( 448^2 \\times 9 \\)256 × 951.223.729.325.028.333.731.542.619.820.526.026.927.2
CogVLM2-liama3-Chat [14]\\( 1344^2 \\)230470.028.847.533.730.245.842.054.032.841.231.237.339.8
IXC2.5 [63]\\( 560^2 \\times 24 \\)400 × 2469.336.163.939.533.652.548.553.441.017.730.033.935.5
LLaVA-OneVision [20]\\( 384^2 \\times 9 \\)729 × 978.753.560.740.345.859.655.861.840.040.834.141.244.2
Closed-sourceGPT-4o-mini--62.56.744.226.524.237.132.847.039.825.826.832.534.9
Gemini-1.5-pro [50]--67.614.039.931.126.639.635.952.733.228.319.229.233.4
GPT-4o--77.728.946.733.922.446.441.961.444.836.526.437.642.3
Claude 3.5 Sonnet--72.525.767.432.240.852.947.761.961.241.831.944.149.2
", + "bbox": [ + 94, + 224, + 901, + 565 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "sual compression rate of 9 times. In addition, FALCON also outperforms most models that have twice as many visual tokens. Specifically, FALCON outperforms CogVLM2-llama3-Chat [14] in average scores on both perception and reasoning task, which uses 2304 tokens per-image. It also surpasses IXC2.5 [63] in average scores on the reasoning tasks, which has a maximum token counts of 9600. Notably, FALCON achieves higher average scores in both perception and reasoning tasks compared to three advanced commercial models, GPT-4o, GPT-4o-mini and Gemini1.5-pro [50], with it notably outperforming all commercial models in the sub-tasks of remote sensing and monitoring. These results demonstrate that the ReCompact and ReAtten technique proposed by FALCON effectively eliminate visual redundancy and ensure visual continuity, allowing comprehensive and detailed visual information to be conveyed with a highly compact visual representation.", + "bbox": [ + 89, + 573, + 482, + 829 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To further assess FALCON's capabilities in high-resolution understanding, we conduct evaluations on more high-resolution benchmarks: V* [53], DocVQA [42], TextVQA [49] and ChartQA [41]. V* focuses on evaluating", + "bbox": [ + 89, + 839, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "models' capabilities for fine-grained attribute recognition and spatial reasoning. DocVQA challenges models to comprehend text-dense document images. ChartQA evaluates models' proficiency in interpreting various types of chart, while TextVQA assesses the ability to recognize small text in real-world scenes. As shown in Tab. 2, FALCON demonstrates superior performance across all these benchmarks.", + "bbox": [ + 511, + 573, + 903, + 678 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate the performance of FALCON across a broader range of domains, we further conduct experiments with diverse popular benchmarks. As shown in Tab. 4, POPE [25], ScienceQA [37], MMBench [34] and SEED-Bench [19] are selected to assess models' capabilities in hallucination, scientific reasoning, and comprehensive ability. FALCON achieves SOTA performance across all these benchmarks. The results demonstrate that FALCON exhibits superior capabilities across a wide range of tasks.", + "bbox": [ + 511, + 680, + 906, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 830, + 666, + 845 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We perform extensive ablation studies to validate the effectiveness of the proposed ReCompact and ReAtten.", + "bbox": [ + 511, + 853, + 903, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ablation on Training Data. To eliminate the influx", + "bbox": [ + 531, + 885, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23535", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d317d5be1748869dd71a6b23903c1097fcf424c80a45494cda3caca2b06ef27f.jpg", + "table_caption": [ + "Table 2. Evaluations on diverse high-resolution Benchmarks. \"V* ${}_{Avg}$ \" represents the weighted average score for the two subtasks in the ${V}^{ * }$ benchmark. † denotes OCR inputs are utilized." + ], + "table_footnote": [], + "table_body": "
ModelV*AvgDocVQATextVQAChartQA
Low-resolution MLLMs
Qwen-VL [1]-65.163.859.0
InstructBLIP [9]34.04.550.710.9
LLaVA-v1.5-7B [30]48.78.558.2†17.8
mPLUG-Owl2 [60]-17.958.2†22.8
High-resolution MLLMs
UReader [58]-65.457.659.3
Monkey [28]-66.565.859.0
LlvaF-FlexAttn [23]54.4-48.9-
S2-Wrapper-7B [48]55.5-61.0-
LlvaNxt-Vicuna7B [31]61.8-64.954.3
SliME-7B [67]--64.433.9
FALCON64.474.466.568.8
", + "bbox": [ + 99, + 142, + 472, + 339 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Ablation results on training data. We train FALCON with two-stage pipeline and exactly the same dataset as LLaVAv1.5 [30]. The results confirm the superiority of the proposed method when the influence of training data is excluded.", + "bbox": [ + 89, + 343, + 483, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/033cda0170d90fa03b901d164de4d812f479bd3f65f7a6d76335bf8e0d8b24d8.jpg", + "image_caption": [ + "Figure 4. Comparison of different compression methods. We report the sum of the \"Avg-C\" scores for the perception and reasoning tasks in MME-RealWorld. The superior performance demonstrates that ReCompact effectively eliminates redundancy and produces a more expressive and compact visual representation." + ], + "image_footnote": [], + "bbox": [ + 91, + 411, + 483, + 573 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ence of training data and demonstrate the advantages of the proposed method, we utilize exactly the same dataset as LLaVA-v1.5 [30] and employ a two-stage pipeline for training. As shown in Tab. 3, the proposed method exhibits significant performance advantages across multiple benchmarks when trained on the same dataset.", + "bbox": [ + 89, + 657, + 483, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of ReCompact mechanism. To verify the advantages of the proposed ReCompact mechanism in reducing visual redundancy and compressing visual representations during the visual encoding, we compare the Re-Compact with other commonly used compression methods. This includes pooling [48, 55], Pixel Shuffle [7] and Abstractor [60], where the Abstractor serves as a representative module that combines query tokens with cross-attention.", + "bbox": [ + 89, + 750, + 482, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Fig. 4, ReCompact demonstrates superior performance compared to other compression methods.", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1150b29a95b08510f57717b0145b9822dac0fba53380b848ac9d6d1e3113870e.jpg", + "table_caption": [ + "Table 4. Evaluation on benchmarks across diverse domains. The POPE, SQA, MMBench and SEED-Bench are used to assess hallucination, scientific reasoning, and overall abilities, respectively." + ], + "table_footnote": [], + "table_body": "
ModelSQAPOPEMMBSEED-Img
Low-Resolution MLLMs
Qwen-VL [1]67.1-41.662.3
InstructBLIP [9]63.178.938.458.8
LlaVA-v1.5-7B [30]66.885.962.865.8
mPLUG-Owl2 [60]68.785.863.564.5
High-Resolution MLLMs
Monkey [28]69.483.757.464.3
LlaVA-FlexAttn [23]-85.965.7-
S2-Wrapper-7B [48]--66.267.9
LLaVA-Next-Vicuna7B [31]70.186.566.569.6
SliME-7B [67]76.885.463.167.0
FALCON88.888.076.173.3
", + "bbox": [ + 521, + 142, + 895, + 327 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/de7f372c5cc25c3bf236cf9c0af8cca376b8630b98794d3b553b18055604e759.jpg", + "table_caption": [ + "Table 5. Comparison of different methods for maintaining visual continuity. The \"Baseline\" model refers to a condition where no specific methods are applied to preserve visual continuity. We report the \"Avg-C\" score for MME-RealWorld. The results show that the ReAtten module more effectively addresses the issue of visual fragmentation compared to other methods, thereby improving overall performance and reducing hallucinations." + ], + "table_footnote": [], + "table_body": "
MethodV*AVgMME-RealWorldPOPE
PerceptionReasoning
Baseline51.338.235.685.7
CIP [16]50.338.437.286.3
W-Atten [33]60.241.038.186.4
ReAtten61.342.139.087.3
", + "bbox": [ + 544, + 440, + 870, + 529 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "These results indicate that ReCompact effectively eliminates visual redundancy and produces a more expressive compact visual representation. In contrast, we argue that the pooling layer and Pixel Shuffle merely reduces the number of tokens by merging adjacent ones, failing to eliminate redundancy that interferes with the LLM's understanding of visual content. Additionally, although Abstractor uses more parameters, its performance is generally lower than that of pooling and ReCompact. This result suggests that even millions of samples may be insufficient for fully training Abstractor, which diminishes their practical efficacy.", + "bbox": [ + 511, + 536, + 906, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of ReAtten Module. To validate the advantage of the proposed ReAtten Module, we compare ReAtten with other contemporary methods, including the Shifted Window Attention (W-Attn) [35] adopted by TextMonkey [33] and the Complementary Image Pyramid (CIP) introduced by MiniMonkey [16].", + "bbox": [ + 511, + 702, + 905, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Tab. 5, the model incorporating ReAtten achieves significant improvements over the baseline across all benchmarks and demonstrates superiority compared to other methods. These results indicate that the ReAtten module effectively enhances the ability of the model in high-resolution perception and mitigates hallucinations by addressing the issue of visual fragmentation. In contrast, we", + "bbox": [ + 511, + 795, + 905, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23536", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/69a61bdb051446db242affaf1ca3d6af18de2936a64fc459257188dc0fac5037.jpg", + "image_caption": [ + "(a) Input Image" + ], + "image_footnote": [], + "bbox": [ + 99, + 109, + 333, + 233 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8ceb0250439bd1470cb952aa6838c3da9139588af5237a80a9f74e4373251c0c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 344, + 92, + 467, + 166 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bd66f1083bbf9ca89b53d071f4a99d6ba15b6ea2d41b3298cad141cc222eef11.jpg", + "image_caption": [ + "(b) Register-to-1" + ], + "image_footnote": [], + "bbox": [ + 344, + 167, + 467, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5059284c6bb109e87b39beafc78b52016928dfabae1f0d576551336ded1848a6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 467, + 93, + 586, + 166 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/95f15239a632e544a78a17d734546edff477e66bf722bd4973c44a87b7dcafed.jpg", + "image_caption": [ + "Image Attention" + ], + "image_footnote": [], + "bbox": [ + 467, + 167, + 586, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d67645d774671878f7e1abc1423fa5dc384083272d41db0887c544e8031ffb02.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 598, + 93, + 828, + 167 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1253deae450cc82ae011ce0b29a5cf61ef43fe4fbaff73bfe804d96497da7381.jpg", + "image_caption": [ + "(c) Ablation on ReAtten" + ], + "image_footnote": [], + "bbox": [ + 599, + 170, + 828, + 241 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "w/ ReAtten", + "bbox": [ + 834, + 123, + 890, + 150 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "w/o ReAtten", + "bbox": [ + 834, + 196, + 890, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/813b26bb016597ba79d2f958654b8d68b721c43981cec7fa7c93c4039fc1a5c1.jpg", + "image_caption": [ + "Figure 5. Visualization of the register-to-image attention map. We randomly select visual registers and extract the attention scores that each register assigns to image tokens. These scores are then visualized as heatmaps overlaid on the input image. The results in (b) show that each visual register focuses on specific parts of the image, capturing rich visual patterns. Meanwhile, registers pay minimal attention to background areas, effectively avoiding the inclusion of redundancy. We also conduct qualitative ablation studies of the ReAtten module. As shown in (c), in the ViT model without ReAtten, the attention patterns across different sub-images appear extremely fragmented. In contrast, the ViT model with ReAtten shows continuous attention patterns, indicating effective information interaction between sub-images.", + "Figure 6. Ablation on different visual register counts. Increasing the number of visual registers generally enhances accuracy across both benchmarks. However, the improvement rate noticeably tapers off when the register count exceeds 64. Based on this observation, to better balance the efficiency and performance, we select register counts of 64 for FALCON." + ], + "image_footnote": [], + "bbox": [ + 112, + 369, + 460, + 483 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "argue that the information exchange in W-Atten is limited to a local window, which hinders sufficient information exchange among all sub-images. Meanwhile, CIP introduces further redundancy by repeatedly encoding the same visual content, which can even lead to a decline in performance.", + "bbox": [ + 89, + 582, + 482, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation on the Number of Visual Registers. To investigate how visual register count affects model performance, we conduct experiments with visual register counts set to 36, 64, and 144. These counts correspond to visual token compression rates of $16 \\times$ , $9 \\times$ , and $4 \\times$ , respectively. As shown in Fig. 6, increasing the number of visual registers generally enhances model's high-resolution comprehension capabilities and reduces hallucinations. However, when the register count exceeds 64, the improvement gains start tapering off noticeably. Meanwhile, increasing visual registers demands more computational resources. Compared to the baseline model with 36 registers, models configured with 64 and 144 registers require 1.15 and 2.41 times more training time, respectively. Base on these observations, to achieve a better balance between efficiency and performance, we select 64 as the register counts for FALCON.", + "bbox": [ + 89, + 659, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Qualitative Analysis", + "text_level": 1, + "bbox": [ + 511, + 369, + 705, + 386 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To qualitatively analyze the proposed visual register technique, we visualize the register-to-image attention maps within ViT. Specifically, for a given visual register, we extract the attention scores it assigns to image tokens in each sub-image and concatenate them to construct a complete heatmap. As shown in Fig. 5(b), each visual register focuses on relatively distinct areas of the image, highlighting that it captures a diverse range of visual details. Moreover, the visual registers tend to avoid background areas, indicating their effectiveness in identifying and discarding visual redundancy. These properties enable visual registers to effectively produce a comprehensive and non-redundant visual representation. Additionally, to qualitatively evaluate the effectiveness of the ReAtten module, we visualize the attention maps within ViT models with and without ReAtten. As shown in Fig. 5(c), in the ViT model without ReAtten, the attention patterns across different sub-images appear highly fragmented. In contrast, the ViT model with ReAtten exhibits continuous attention patterns, indicating effective information exchange between sub-images.", + "bbox": [ + 511, + 391, + 906, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 708, + 633, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To address visual redundancy and fragmentation in high-resolution MLLMs, we propose FALCON. FALCON employs an innovative visual register technique that simultaneously addresses both challenges. This technique uses a Re-Compact mechanism to adaptively aggregate essential visual information through visual registers, creating a compact, non-redundant representation. Additionally, a ReAtten module is introduced to further enhances visual continuity by enabling information exchange among sub-images via visual registers. Extensive experiments demonstrate FALCON's superiority in high-resolution understanding.", + "bbox": [ + 511, + 734, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23537", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgement", + "text_level": 1, + "bbox": [ + 89, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This study is supported by National Natural Science Foundation of China (Grant No. 62306090, No. U24A20328), Natural Science Foundation of Guangdong Province of China (Grant No. 2024A1515010147), Shenzhen Science and Technology Program (KQTD20240729102207002), Natural Science Foundation of Beijing, China (Grant No. L254018), Jiangsu Science and Technology Major Program (Grant No. BG2024041) and China Postdoctoral Science Foundation (Grant No. 2024M764192).", + "bbox": [ + 89, + 114, + 483, + 251 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 267, + 187, + 282 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 1(2):3, 2023. 2, 3, 6, 7", + "[2] Gongwei Chen, Leyang Shen, Rui Shao, Xiang Deng, and Liqiang Nie. Lion: Empowering multimodal large language model with dual-level visual knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26540-26550, 2024. 1, 2", + "[3] Gongwei Chen, Xurui Zhou, Rui Shao, Yibo Lyu, Kaiwen Zhou, Wentao Li, Yinchuan Li, Zhongang Qi, and Liqiang Nie. Less is more: Empowering gui agent with context-aware simplification. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2025. 3", + "[4] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 6", + "[5] Jingxuan Chen, Derek Yuen, Bin Xie, Yuhao Yang, Gongwei Chen, Zhihao Wu, Li Yixing, Xurui Zhou, Weiwen Liu, Shuai Wang, Kaiwen Zhou, Rui Shao, Liqiang Nie, Yasheng Wang, Jianye HAO, Jun Wang, and Kun Shao. Spa-bench: A comprehensive benchmark for smartphone agent evaluation. In The Thirteenth International Conference on Learning Representations, 2025. 3", + "[6] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. In European Conference on Computer Vision, pages 370-387. Springer, 2024. 6", + "[7] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internv1: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24185–24198, 2024. 7", + "[8] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source" + ], + "bbox": [ + 99, + 292, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "chatbot impressing gpt-4 with $90\\%$ chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2(3):6, 2023. 3", + "[9] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1, 2, 3, 7", + "[10] Timothee Darcet, Maxime Aquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. In The Twelfth International Conference on Learning Representations, 2024. 2, 4", + "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 3", + "[12] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024.5", + "[13] Dan Hendrycks and Kevin Gimpel. Bridging nonlinearities and stochastic regularizers with gaussian error linear units. arXiv preprint arXiv:1606.08415, 2016. 5", + "[14] Wenyi Hong, Weihan Wang, Ming Ding, Wenmeng Yu, Qingsong Lv, Yan Wang, Yean Cheng, Shiyu Huang, Junhui Ji, Zhao Xue, et al. Cogvlm2: Visual language models for image and video understanding. arXiv preprint arXiv:2408.16500, 2024. 6", + "[15] Anwen Hu, Haiyang Xu, Jiabo Ye, Ming Yan, Liang Zhang, Bo Zhang, Ji Zhang, Qin Jin, Fei Huang, and Jingren Zhou. mplug-docowl 1.5: Unified structure learning forOCR-free document understanding. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 3096-3120, 2024. 1, 3, 6", + "[16] Mingxin Huang, Yuliang Liu, Dingkang Liang, Lianwen Jin, and Xiang Bai. Mini-monkey: Alleviating the semantic saw-tooth effect for lightweight MLLMs via complementary image pyramid. In The Thirteenth International Conference on Learning Representations, 2025. 3, 7", + "[17] Runhui Huang, Xinpeng Ding, Chunwei Wang, Jianhua Han, Yulong Liu, Hengshuang Zhao, Hang Xu, Lu Hou, Wei Zhang, and Xiaodan Liang. Hires-llava: Restoring fragmentation input in high-resolution large vision-language models. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 29814–29824, 2025. 1, 3", + "[18] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan P Foster, Pannag R Sanketi, Quan Vuong, et al. Openvla: An open-source vision-language-action model. In 8th Annual Conference on Robot Learning, 2024. 3", + "[19] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking mul" + ], + "bbox": [ + 516, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23538", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "timodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 6", + "[20] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. LLaVA-onevision: Easy visual task transfer. Transactions on Machine Learning Research, 2025. 6", + "[21] Hao Li, Qi Lv, Rui Shao, Xiang Deng, Yinchuan Li, Jianye HAO, and Liqiang Nie. STAR: Learning diverse robot skill abstractions through rotation-augmented vector quantization. In Forty-second International Conference on Machine Learning, 2025. 3", + "[22] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 2", + "[23] Junyan Li, Delin Chen, Tianle Cai, Peihao Chen, Yining Hong, Zhenfang Chen, Yikang Shen, and Chuang Gan. Flex-attention for efficient high-resolution vision-language models. In European Conference on Computer Vision, pages 286–302, 2024. 7", + "[24] Wei Li, Bing Hu, Rui Shao, Leyang Shen, and Liqiang Nie. Lion-fs: Fast & slow video-language thinker as online video assistant. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 3240-3251, 2025. 3", + "[25] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In The 2023 Conference on Empirical Methods in Natural Language Processing, 2023. 6", + "[26] Yanwei Li, Yuechen Zhang, Chengyao Wang, Zhisheng Zhong, Yixin Chen, Ruihang Chu, Shaoteng Liu, and Jiaya Jia. Mini-gemini: Mining the potential of multi-modality vision language models. arXiv preprint arXiv:2403.18814, 2024. 6", + "[27] Zaijing Li, Yuquan Xie, Rui Shao, Gongwei Chen, Dongmei Jiang, and Liqiang Nie. Optimus-1: Hybrid multimodal memory empowered agents excel in long-horizon tasks. In Advances in neural information processing systems, 2024. 3", + "[28] Zhang Li, Biao Yang, Qiang Liu, Zhiyin Ma, Shuo Zhang, Jingxu Yang, Yabo Sun, Yuliang Liu, and Xiang Bai. Monkey: Image resolution and text label are important things for large multi-modal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26763-26773, 2024. 2, 3, 6, 7", + "[29] Zaijing Li, Yuquan Xie, Rui Shao, Gongwei Chen, Dongmei Jiang, and Liqiang Nie. Optimus-2: Multimodal minercraft agent with goal-observation-action conditioned policy. In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2025. 3", + "[30] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 1, 2, 3, 6, 7", + "[31] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,ocr,and world knowledge,2024.6,7" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 3", + "[33] Yuliang Liu, Biao Yang, Qiang Liu, Zhang Li, Zhiyin Ma, Shuo Zhang, and Xiang Bai. Textmonkey: AnOCR-free large multimodal model for understanding document. arXiv preprint arXiv:2403.04473, 2024. 1, 2, 3, 6, 7", + "[34] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European Conference on Computer Vision, pages 216-233. Springer, 2025. 6", + "[35] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 3, 7", + "[36] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Yaofeng Sun, et al. Deepseek-vl: towards real-world vision-language understanding. arXiv preprint arXiv:2403.05525, 2024. 3, 6", + "[37] Pan Lu, Swaroop Mishra, Tony Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In The 36th Conference on Neural Information Processing Systems (NeurIPS), 2022. 6", + "[38] Gen Luo, Yiyi Zhou, Yuxin Zhang, Xiawu Zheng, Xiaoshuai Sun, and Rongrong Ji. Feast your eyes: Mixture-of-resolution adaptation for multimodal large language models. arXiv preprint arXiv:2403.03003, 2024. 3", + "[39] Qi Lv, Hao Li, Xiang Deng, Rui Shao, Yinchuan Li, Jianye Hao, Longxiang Gao, Michael Yu Wang, and Liqiang Nie. Spatial-temporal graph diffusion policy with kinematic modeling for bimanual robotic manipulation. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 17394-17404, 2025. 3", + "[40] Yibo Lyu, Rui Shao, Gongwei Chen, Yijie Zhu, Weili Guan, and Liqiang Nie. Puma: Layer-pruned language model for efficient unified multimodal retrieval with modality-adaptive learning. In Proceedings of the 33nd ACM International Conference on Multimedia, 2025. 3", + "[41] Ahmed Masry, Xuan Long Do, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. In Findings of the Association for Computational Linguistics: ACL 2022, pages 2263-2279, 2022. 6", + "[42] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 6", + "[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 4" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23539", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[44] Rui Shao, Xiangyuan Lan, Jiawei Li, and Pong C Yuen. Multi-adversarial discriminative deep domain generalization for face presentation attack detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10023-10031, 2019. 3", + "[45] Rui Shao, Tianxing Wu, and Ziwei Liu. Detecting and grounding multi-modal media manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6904-6913, 2023. 1", + "[46] Rui Shao, Tianxing Wu, Jianlong Wu, Liqiang Nie, and Zwei Liu. Detecting and grounding multi-modal media manipulation and beyond. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1, 3", + "[47] Leyang Shen, Gongwei Chen, Rui Shao, Weili Guan, and Liqiang Nie. Mone: Mixture of multimodal experts for generalist multimodal large language models. In Advances in neural information processing systems, 2024. 2", + "[48] Baifeng Shi, Ziyang Wu, Maolin Mao, Xin Wang, and Trevor Darrell. When do we not need larger vision models? In European Conference on Computer Vision, pages 444-462, 2024. 2, 3, 7", + "[49] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 6", + "[50] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 6", + "[51] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3", + "[52] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3", + "[53] Penghao Wu and Saining Xie. V*: Guided visual search as a core mechanism in multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13084-13094, 2024. 6", + "[54] Bin Xie, Rui Shao, Gongwei Chen, Kaiwen Zhou, Yinchuan Li, Jie Liu, Min Zhang, and Liqiang Nie. Gui-explorer: Autonomous exploration and mining of transition-aware knowledge for gui agent. In Annual Meeting of the Association for Computational Linguistics (ACL), 2025. 3", + "[55] Linli Yao, Lei Li, Shuhuai Ren, Lean Wang, Yuanxin Liu, Xu Sun, and Lu Hou. Deco: Decoupling token compression from semantic abstraction in multimodal large language models. arXiv preprint arXiv:2405.20985, 2024. 2, 3, 7", + "[56] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 6" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[57] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Yuhao Dan, Chenlin Zhao, Guohai Xu, Chenliang Li, Junfeng Tian, et al. mplug-docowl: Modularized multimodal large language model for document understanding. arXiv preprint arXiv:2307.02499, 2023. 3", + "[58] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Guohai Xu, Chenliang Li, Junfeng Tian, Qi Qian, Ji Zhang, et al. Ureader: UniversalOCR-free visually-situated language understanding with multimodal large language model. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 2841-2858, 2023. 1, 2, 3, 7", + "[59] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 1, 2", + "[60] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Anwen Hu, Haowei Liu, Qi Qian, Ji Zhang, and Fei Huang. mplug-owl2: Revolutionizing multi-modal large language model with modality collaboration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13040–13051, 2024. 1, 2, 3, 7", + "[61] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In European Conference on Computer Vision (ECCV), pages 69-85. Springer, 2016. 1", + "[62] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 3, 5", + "[63] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, et al. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output. arXiv preprint arXiv:2407.03320, 2024. 6", + "[64] Renshan Zhang, Yibo Lyu, Rui Shao, Gongwei Chen, Weili Guan, and Liqiang Nie. Token-level correlation-guided compression for efficient multimodal document understanding. arXiv preprint arXiv:2407.14439, 2024. 3", + "[65] Xiaofeng Zhang, Yihao Quan, Chen Shen, Xiaosong Yuan, Shaotian Yan, Liang Xie, Wenxiao Wang, Chaochen Gu, Hao Tang, and Jieping Ye. From redundancy to relevance: Enhancing explainability in multimodal large language models. Annual Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics, 2025. 2", + "[66] YiFan Zhang, Huanyu Zhang, Haochen Tian, Chaoyou Fu, Shuangqing Zhang, Junfei Wu, Feng Li, Kun Wang, Qing-song Wen, Zhang Zhang, Liang Wang, and Rong Jin. MMErealworld: Could your multimodal LLM challenge high-resolution real-world scenarios that are difficult for humans? In The Thirteenth International Conference on Learning Representations, 2025. 5", + "[67] Yi-Fan Zhang, Qingsong Wen, Chaoyou Fu, Xue Wang, Zhang Zhang, Liang Wang, and Rong Jin. Beyond llava-hd: Diving into high-resolution large multimodal models. arXiv preprint arXiv:2406.08487, 2024. 1, 3, 7" + ], + "bbox": [ + 516, + 90, + 903, + 893 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "23540", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/FALCON_ Resolving Visual Redundancy and Fragmentation in High-resolution Multimodal Large Language Models via Visual Registers/3c258128-6a38-46a9-bf2a-aabe422e5840_model.json b/2025/FALCON_ Resolving Visual Redundancy and Fragmentation in High-resolution Multimodal Large Language Models via Visual Registers/3c258128-6a38-46a9-bf2a-aabe422e5840_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0cdd300ee63543e9b9478253792eaecccebb32a5 --- /dev/null +++ b/2025/FALCON_ Resolving Visual Redundancy and Fragmentation in High-resolution Multimodal Large Language Models via Visual Registers/3c258128-6a38-46a9-bf2a-aabe422e5840_model.json @@ -0,0 +1,2213 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.099, + 0.13, + 0.9, + 0.177 + ], + "angle": 0, + "content": "FALCON: Resolving Visual Redundancy and Fragmentation in High-resolution Multimodal Large Language Models via Visual Registers" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.209, + 0.777, + 0.263 + ], + "angle": 0, + "content": "Renshan Zhang\\(^{1}\\), Rui Shao\\(^{1\\dagger}\\), Gongwei Chen\\(^{1}\\), Miao Zhang\\(^{1}\\), Kaiwen Zhou\\(^{2}\\), Weili Guan\\(^{1}\\), Liqiang Nie\\(^{1\\dagger}\\) \n\\(^{1}\\)Harbin Institute of Technology, Shenzhen \\(^{2}\\)Huawei Noah's Ark Lab" + }, + { + "type": "text", + "bbox": [ + 0.266, + 0.266, + 0.732, + 0.28 + ], + "angle": 0, + "content": "zhangrenshan@stu.hit.edu.cn shaorui@hit.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.285, + 0.283, + 0.723, + 0.298 + ], + "angle": 0, + "content": "chengongwei@hit.edu.cn nieliqiang@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.318, + 0.301, + 0.676, + 0.315 + ], + "angle": 0, + "content": "https://github.com/JiuTian-VL/FALCON" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.351, + 0.327, + 0.367 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.382, + 0.486, + 0.82 + ], + "angle": 0, + "content": "The incorporation of high-resolution visual input equips multimodal large language models (MLLMs) with enhanced visual perception capabilities for real-world tasks. However, most existing high-resolution MLLMs rely on a cropping-based approach to process images, which leads to fragmented visual encoding and a sharp increase in redundant tokens. To tackle these issues, we propose the FALCON model. FALCON introduces a novel visual register technique to simultaneously: 1) Eliminate redundant tokens at the stage of visual encoding. To directly address the visual redundancy present in the output of vision encoder, we propose a Register-based Representation Compacting (ReCompact) mechanism. This mechanism introduces a set of learnable visual registers designed to adaptively aggregate essential information while discarding redundancy. It enables the encoder to produce a more compact visual representation with a minimal number of output tokens, thus eliminating the need for an additional compression module. 2) Ensure continuity in visual encoding. To address the potential encoding errors caused by fragmented visual inputs, we develop a Register Interactive Attention (ReAtten) module. This module facilitates effective and efficient information exchange across sub-images by enabling interactions between visual registers. It ensures the continuity of visual semantics throughout the encoding. We conduct comprehensive experiments with FALCON on high-resolution benchmarks across a wide range of scenarios. FALCON demonstrates superior performance with a remarkable 9-fold reduction in visual tokens." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.823, + 0.224, + 0.839 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.848, + 0.483, + 0.879 + ], + "angle": 0, + "content": "Recently, multimodal large language models (MLLMs) [2, 9, 30, 32, 59, 60] have made significant advancements in" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.35, + 0.906, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.594, + 0.907, + 0.65 + ], + "angle": 0, + "content": "Figure 1. Comparison between existing high-resolution MLLMs and FALCON. FALCON eliminates visual redundancy and ensures the continuity of high-resolution encoding through the use of ReCompact and ReAtten mechanisms." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.657, + 0.907, + 0.868 + ], + "angle": 0, + "content": "performing various vision-language tasks [45, 46, 61], facilitating more natural human-machine interactions. Early MLLMs [9, 32, 59] typically processed images at a fixed low resolution, such as \\(224 \\times 224\\). This limited their capabilities in many real-world applications that require high-resolution input. To equip MLLMs with high-resolution visual perception, recent efforts [15, 17, 33, 58, 67] have introduced a cropping-based high-resolution processing method. Specifically, As shown in Fig. 1(b), this approach crops high-resolution images into multiple non-overlapping sub-images that match the pre-trained resolution of the encoder. Each sub-image is processed independently by the vision encoder. Then the resulting visual features are concatenated and fed into the large language model (LLM)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Although effective, cropping-based methods still face issues of: 1) Visual Redundancy. The number of visual to-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.24, + 0.9 + ], + "angle": 0, + "content": "†Corresponding authors" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23530" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.41 + ], + "angle": 0, + "content": "kens grows dramatically with increasing resolution. While this enriches the visual details, it also leads to an increase in redundant tokens. These redundant tokens significantly increase the computational burden on LLM and interfere with the processing of visual information [65]. To address these issues, as illustrated in Fig. 1(b), existing methods typically rely on additional compression techniques. Representative approaches include combining learnable queries with cross-attention [1, 28, 60] or simply employing pooling layers [48, 55]. However, these methods either demand substantial training resources or fail to fully address redundancy. 2) Visual Fragmentation. Cropping an image into multiple non-overlapping sub-images may disrupt visual semantics and potentially lead to encoding errors. As shown in Fig. 1(a), in an OCR task, the word \"pineapple\" may be incorrectly segmented into \"pine\" and \"apple\", causing the LLM to misinterpret the word. Similarly, fragmentation can also hinder the identification of objects. For example, the \"rubin vase\" pattern illustrated in Fig. 1(a) depicts both a vase and two faces simultaneously. However, the fragmentation could lead to the vase being unrecognizable." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.418, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To address these issues, as shown in Fig. 1(c), we propose the FALCON model. FALCON introduces a novel visual register technique to simultaneously: 1) Eliminate redundant tokens during visual encoding. To eliminate redundancy directly at the stage of visual encoding, we propose a novel Register-based Representation Compacting (ReCompact) mechanism. This mechanism introduces a set of learnable visual registers as additional inputs to the vision encoder. The registers are designed to adaptively aggregate essential information from image tokens and discard redundant content. Then the outputs corresponding to the visual registers are considered as a non-redundant visual representation, serving as the compact input for the LLM instead of a large number of image tokens. This mechanism enables the vision encoder to convey comprehensive visual information in a compact representation using minimal tokens, thus removing the need for an additional compression module. 2) Ensure continuity in visual encoding. To avoid the potential encoding errors caused by fragmented visual inputs, we develop a novel Register Interactive Attention (ReAtten) module. This module leverages the compact and comprehensive representation produced by visual registers to facilitate information exchange across sub-images. It gathers visual registers from different sub-images and employs a specialized cross-ViT attention module to enable interactions among them. Such interaction enables thorough information exchange across the entire image, thus ensuring visual continuity throughout the encoding process. Additionally, the compactness of the visual registers further enhances the efficiency of these interactions. Consequently, this module provides LLMs with a reliable foundation to accurately interpret visual content and reduce hallucinations." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.513 + ], + "angle": 0, + "content": "Note that the widely used Abstractor-like modules [28, 33, 58, 60], including QFormer [9], ReSampler [1], follow a similar approach by using query tokens to extract information from image tokens after visual encoding. The key difference is that the visual register mechanism directly adopts the pretrained ViT to aggregate information into a set of specified register tokens. This design is based on the observation that a well-pretrained ViT can naturally gather global information into some specific tokens through self-attention [10]. It offers the visual register mechanism several strengths in terms of: 1) Training Efficiency. The substantial number of randomly initialized parameters in Abstractor-like modules requires extensive data for pretraining. For instance, QFormer [22], Abstractor [60], and Resampler [1] use large corpora of 129M, 400M, and 1.4B samples, respectively. In contrast, by leveraging well-pretrained ViT parameters, the visual register mechanism requires only 3.5M samples for pretraining. The ablation results in Fig. 4 further demonstrate this advantages. 2) Addressing Visual Fragmentation. By leveraging the compact and comprehensive representation produced by visual registers, the proposed ReAtten module can effectively facilitate information exchange across sub-images. This is an important function that is notably lacking in Abstractor-like modules. Combined with ReCompact, it enables the visual register mechanism to simultaneously address the issues of visual redundancy and fragmentation. The ablation results in Tab. 5 validate the effectiveness of ReAtten." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.516, + 0.862, + 0.529 + ], + "angle": 0, + "content": "We summarize our main contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.532, + 0.905, + 0.637 + ], + "angle": 0, + "content": "- To address the issue of visual redundancy directly at the encoding phase, we introduce the Register-based Representation Compacting (ReCompact) mechanism. This mechanism enables vision encoder to capture rich visual information using only a minimal number of tokens, effectively eliminating redundancy and obviating the need for additional compression modules." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.638, + 0.905, + 0.711 + ], + "angle": 0, + "content": "- To address the issue of visual fragmentation caused by image cropping, we propose the register interactive attention (ReAtten) module. This module ensures the semantic continuity of high-resolution visual encoding in an efficient and effective manner." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.713, + 0.905, + 0.803 + ], + "angle": 0, + "content": "- We introduce FALCON, a model that leverages the proposed visual register technique. Comprehensive evaluations on high-resolution visual understanding benchmarks highlight the superiority of FALCON. Notably, FALCON achieves exceptional results with a remarkable reduction in visual tokens by \\(9 \\times\\)." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.532, + 0.905, + 0.803 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.821, + 0.655, + 0.836 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.848, + 0.833, + 0.864 + ], + "angle": 0, + "content": "2.1. Multimodal Large Language Models" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Research on MLLMs [1, 2, 30, 47, 59, 60] has made significant strides in recent years. By incorporating image in" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23531" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.088, + 0.905, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.267, + 0.908, + 0.325 + ], + "angle": 0, + "content": "Figure 2. Overall framework of the proposed FALCON. The model introduces a set of learnable visual registers, which are fed into the ViT along with the image tokens from each sub-image. During the visual encoding, the ReCompact mechanism enables information from the image tokens to be aggregated into the visual registers, while the proposed ReAtten facilitates information exchange among the visual registers across different sub-images. Ultimately, these visual registers provide a compact visual representation as input to the LLM." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.332, + 0.485, + 0.591 + ], + "angle": 0, + "content": "puts into pretrained LLMs [8, 51, 52], MLLMs demonstrate strong capabilities across various vision-language tasks. Typically, these models use a pretrained vision encoder [43, 62] to transform input images into a set of visual tokens. These tokens are then projected into the input space of the LLM through a projection module (e.g., MLP [30, 32], QFormer [9] and Abstractor [60]) and processed alongside text inputs. Building on the progress of MLLMs, recent studies have shifted from designing specialized methods for certain tasks [44-46] to employing MLLMs as generalist agents, capable of handling diverse practical tasks such as document analysis [15, 57, 64], online video assistant [24], retrieval-augmented generation [40], game agents [27, 29], robot manipulation [18, 21, 39], and GUI agent [3, 5, 54]. Despite these advances, MLLMs still struggle with the fixed low resolution of pretrained image encoders, which can limit their effectiveness in handling high-resolution images." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.604, + 0.318, + 0.619 + ], + "angle": 0, + "content": "2.2. High-resolution MLLMs" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.627, + 0.483, + 0.794 + ], + "angle": 0, + "content": "To address the limitation of MLLMs in handling only low-resolution visual inputs, early studies attempted to either directly incorporate vision encoders that were pre-trained on higher-resolution images [36, 38], or to improve the resolution of pre-trained encoders through additional training [1, 60]. While achieving some progress, they still remain restricted by the fixed resolutions. To accommodate the varying resolutions and aspect ratios, cropping-based techniques have been widely adopted in recent works [15, 17, 33, 58, 67]. However, these approaches still face challenges related to visual redundancy and fragmentation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.903 + ], + "angle": 0, + "content": "To mitigate the issue of redundancy, existing models typically require an additional compression module. For example, Abstractor-like modules utilizing learnable queries combined with cross-attention are widely adopted [1, 28, 33, 58, 60]. These modules typically require pretraining on massive amounts of data, which diminishes their practical efficacy. In contrast, some approaches use pooling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.332, + 0.908, + 0.575 + ], + "angle": 0, + "content": "layers [48, 55] or convolutional layers [15, 36]. Though resource-efficient, they often fail to address the issue of redundancy. In this work, we present an initial exploration to directly eliminate redundancy during the visual encoding phase, thereby removing the need for additional compression. To address the issue of fragmentation, several concurrent studies also make preliminary investigation. TextMonkey [33] uses Shifted Window Attention [35] to exchange information between sub-images. But the exchange is still limited to local regions. MiniMonkey [16] proposed a Complementary Image Pyramid strategy to generate complementary cropping schemes. However, this strategy inevitably encodes the same visual content multiple times, further exacerbating visual redundancy. In this work, we propose the ReAtten module to facilitate continuous visual encoding through the interaction among visual registers." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.589, + 0.62, + 0.604 + ], + "angle": 0, + "content": "3. FALCON" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.907, + 0.901 + ], + "angle": 0, + "content": "The overall framework of FALCON is illustrated in Fig. 2. The method begins with shape-adaptive cropping [58], where the image is cropped into multiple non-overlapping sub-images, each matching the resolution of the pre-trained vision encoder. Additionally, a resized version of the full image is used to supplement global layout information. To directly eliminate redundancy during visual encoding, we propose Register-based Representation Compacting (Re-Compact) mechanism. This mechanism introduces a set of learnable visual registers, which are paired with image tokens from each sub-image and fed into the vision encoder to capture rich visual information. To ensure the continuity of visual semantics throughout the encoding, a novel Register Interactive Attention (ReAtten) module is integrated into the Vision Transformer (ViT) [11] to facilitate information exchange between sub-images via the visual registers. Finally, the compact visual representations produced by the visual registers are processed through a simple MLP module before being fed into the LLM for further analysis." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23532" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.484, + 0.108 + ], + "angle": 0, + "content": "3.1. Mitigating Visual Redundancy via ReCompact" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.484, + 0.415 + ], + "angle": 0, + "content": "Existing MLLMs typically use a ViT-based vision encoder, such as CLIP [43]. For an input image of size \\( H \\times W \\), it is patched into multiple \\( P \\times P \\) patches. Each patch is treated as an individual image token, resulting in \\( N = \\frac{H}{P} \\times \\frac{W}{P} \\) tokens, denoted as \\( I = \\{i_1, \\dots, i_N\\} \\). In most current MLLMs, output features corresponding to all image tokens are usually preserved in their entirety. However, when high-resolution images are included, the number of image tokens increases dramatically, placing a substantial burden on the computational resources of LLMs. Consequently, most high-resolution MLLMs necessitate an additional compression module, which can be resource-inefficient. Previous research has shown that most image tokens contain only localized information from their surrounding areas [10]. It suggests that a significant portion of these tokens may be redundant, particularly those originating from background regions. This observation motivates us to directly eliminate redundancy in the output of the vision encoder, thereby reducing the number of visual tokens and avoiding the need for an additional compression module." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.416, + 0.484, + 0.675 + ], + "angle": 0, + "content": "To achieve this, we propose a novel Register-based Representation Compacting (ReCompact) mechanism. This mechanism employs a set of visual registers to aggregate essential information from image tokens, resulting in a more compact visual representation. Specifically, as shown in Fig. 2, we introduce a set of learnable visual registers, denoted by \\( R = \\{r_1,\\dots ,r_M\\} \\), where \\( M \\) represents the number of visual registers, and \\( M\\ll N \\). Given a single input image, the image tokens are concatenated with visual registers to create a new input for ViT, represented as \\( V = \\mathrm{Concat}(I,R) = \\{i_1,\\dots ,i_N,r_1,\\dots ,r_M\\} \\). During visual encoding, the self-attention mechanism in ViT enables the visual register to adaptively aggregate important information from image tokens while discarding redundant content. Denote \\( X_{k,l}\\in \\mathcal{R}^{(N + M)\\times D} \\) as the input hidden states of the \\( k \\)-th sub-image at the \\( l \\)-th layer, where \\( D \\) is the dimensionality. The self-attention is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.685, + 0.484, + 0.725 + ], + "angle": 0, + "content": "\\[\n\\hat {X} _ {k, l} = \\operatorname {S e l f - A t t e n} \\left(X _ {k, l}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {X _ {k , l} X _ {k , l} ^ {T}}{\\sqrt {D _ {\\text {k e y}}}}\\right) X _ {k, l} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.734, + 0.483, + 0.87 + ], + "angle": 0, + "content": "Notably, we do not apply any attention masks for specific constraints. Our experiments show that pretraining on a modest amount of data is sufficient to effectively adapt the pretrained ViT to the proposed ReCompact, enabling it to aggregate essential information within the visual registers. After visual encoding, The output features corresponding to the visual registers, denoted as \\( F = \\{f_{1},\\dots ,f_{M}\\} \\), are then used as the visual representation of the image and fed into the LLM for various vision-language tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "To further adapt the ReCompact mechanism for cropping-based high-resolution image perception, we share" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.259 + ], + "angle": 0, + "content": "the same set of visual registers as input for different sub-images. The input for the \\(k\\)-th sub-image is formulated as \\(V_{k} = \\mathrm{Concat}(I_{k},R)\\), where \\(k = 1,2,\\dots ,N_{c}\\) and \\(N_{c}\\) denotes the total number of sub-images. After encoding, the outputs of the visual registers for all sub-images are concatenated, yielding \\(F_{hr} = \\mathrm{Concat}(F_1,F_2,\\dots ,F_{N_c})\\) which serves as the input to the LLM. The ReCompact mechanism enables the vision encoder to represent rich visual information using only a minimal number of \\(M\\) tokens per sub-image, which significantly reduces the visual redundancy and eliminates the need for additional compression." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.267, + 0.857, + 0.283 + ], + "angle": 0, + "content": "3.2. Ensuring Visual Continuity via ReAtten" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.907, + 0.545 + ], + "angle": 0, + "content": "When using the cropping-based high-resolution image encoding method, each sub-image is independently encoded by the vision encoder. This process can disrupt semantic coherence and result in the misrepresentation of visual concepts. To preserve the continuity of visual semantics, it is crucial to enable information exchange between sub-images during encoding. A naive approach would be to directly concatenate the image features from all sub-images and apply self-attention. However, the quadratic computational complexity makes this approach impractical. In Sec. 3.1, we introduced the ReCompact mechanism, where the visual registers effectively aggregate rich visual information while discarding redundancy during visual encoding, allowing them to serve as compact yet comprehensive image representations. Consequently, interactions between visual registers can naturally facilitate sufficient information exchange across different sub-images in an efficient manner." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.546, + 0.909, + 0.753 + ], + "angle": 0, + "content": "To enable information exchange between visual registers across different sub-images, we introduce a novel Register Interactive Attention (ReAtten) module. As shown in Fig. 2, the ReAtten module is integrated into each of the \\(L\\) layers of the ViT. After the self-attention process in Eq. (1), we acquire \\(\\hat{X}_{k,l} = \\{\\hat{X}_{k,l}^{I},\\hat{X}_{k,l}^{R}\\}\\), where \\(\\hat{X}^I\\in \\mathcal{R}^{N\\times D}\\) and \\(\\hat{X}^R\\in \\mathcal{R}^{M\\times D}\\) represent the hidden states corresponding to image tokens and visual registers, respectively. The states of the visual registers from different sub-images are concatenated to form \\(\\hat{X}_l^R = \\mathrm{Concat}\\left(\\{\\hat{X}_{k,l}^R\\}_{k = 1}^{N_c}\\right)\\in\\) \\(\\mathcal{R}^{M\\cdot N_c\\times D}\\). Then these concatenated registers are processed via ReAtten to enable information exchange across subimages, which can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.763, + 0.907, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\overline {{X}} _ {l} ^ {R} = \\mathrm {R e A t t e n} (\\hat {X} _ {l} ^ {R}) = \\hat {X} _ {l} ^ {R} + \\mathrm {C r o s s - V i T - A t t e n} (\\hat {X} _ {l} ^ {R}) (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.793, + 0.907, + 0.87 + ], + "angle": 0, + "content": "where Cross-ViT-Atten is implemented as a self-attention layer. Once the information exchange is completed, the registers and image tokens from a same sub-image are concatenated and then passed into the feedforward network (FFN) for transformation, yielding the output of the \\( l \\)-th layer:" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.879, + 0.907, + 0.906 + ], + "angle": 0, + "content": "\\[\nX _ {k, l + 1} = \\operatorname {F F N} \\left(\\operatorname {C o n c a t} \\left(\\hat {X} _ {k, l} ^ {I}, \\bar {X} _ {k, l} ^ {R}\\right)\\right) \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "23533" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.088, + 0.462, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.351, + 0.483, + 0.448 + ], + "angle": 0, + "content": "Figure 3. The progressive training pipeline. **Stage 0:** Align the frozen vision and language backbones by training only the MLP and registers. **Stage 1:** Tune all parameters with caption and OCR data on low-resolution for coarse alignment. **Stage 2:** Integrate high-resolution input along with fine-grained captions and OCR to achieve refined alignment. **Stage 3:** Froze ViT and introduce ReAtten module for high-resolution instruction tuning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.453, + 0.484, + 0.53 + ], + "angle": 0, + "content": "By integrating ReAtten into each layer of the ViT model, FALCON enables effective information exchange between cropped sub-images through visual registers, thereby ensuring semantic continuity of the high-resolution image throughout the visual encoding process." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.542, + 0.483, + 0.574 + ], + "angle": 0, + "content": "3.3. Learning Expressive Visual Representation via Progressive Training Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.581, + 0.483, + 0.672 + ], + "angle": 0, + "content": "To empower the ReCompact mechanism with the capability to capture rich visual patterns and to enable smoother initialization of the ReAtten module under high-resolution conditions, we develop a specialized progressive training pipeline. As illustrated in Fig. 3, this pipeline progressively transitions from low to high resolution across four stages." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.809 + ], + "angle": 0, + "content": "Stage 0: Static Alignment on Low-resolution. Since both the visual registers and MLP projector are randomly initialized, in stage 0, we aim to achieve a initial alignment between the pretrained vision encoder and the LLM. To facilitate this, we begin by using low-resolution images as inputs, which are directly resized to match the resolution of the vision encoder. We then statically pretrain the model by keeping both the vision encoder and LLM frozen, tuning only the visual registers and MLP using image caption data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Stage 1: Coarse Alignment on Low-resolution. To better adapt the pre-trained vision encoder to our newly introduced ReCompact mechanism, in stage 1, we unfreeze all model parameters to enable a more thorough tuning. To facilitate the visual registers to progressively learn capturing visual information from coarse to fine granularity, we con" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "tinue to use low-resolution inputs and employ high-quality, extended captions rich in visual content, along with full-text OCR data for training. This design allows the registers to acquire holistic scene understanding ability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.153, + 0.905, + 0.272 + ], + "angle": 0, + "content": "Stage 2: Refined Alignment on High-resolution. To further enhance the capacity of the visual register to capture fine-grained visual details, in stage 2, we transition the model to process high-resolution image for refined alignment. Additionally, we employ datasets that includes a series of fine-grained captions and OCR tasks, such as regional captions and text location. This facilitates the visual registers to capture comprehensive fine-grained details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.273, + 0.907, + 0.515 + ], + "angle": 0, + "content": "Stage 3: Instruction Tuning on High-resolution. To adapt the well-aligned visual register for a wide range of real-world tasks, in stage 3, we perform instruction tuning with a mixture of datasets covering various tasks, such as VQA, document understanding, and science reasoning. Additionally, the proposed ReAtten module is integrated to address visual fragmentation caused by cropping-based high-resolution image encoding. Since the vision encoder with ReCompact is well-aligned, the vision encoder is kept frozen and other parameters including ReAtten are trainable during instruction tuning. Notably, to facilitate a smoother initialization for the newly introduced ReAtten modules, we initialize parameters of each ReAtten module with the ViT self-attention parameters at the corresponding depth. This design enables ReAtten to effectively leverage pre-trained parameters, avoiding the need for extensive fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.528, + 0.646, + 0.544 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.552, + 0.906, + 0.659 + ], + "angle": 0, + "content": "We implement FALCON with SigLIP-L/16-384px [62] as the vision encoder and Llama-3.1-8B-Instruct [12] as the LLM backbone. A two-layer MLP module with GeLU [13] activation is served as the vision-language projection layer. The number of visual registers is set to 64, achieving remarkable visual token compression rates of \\(9 \\times\\). The maximum number of sub-images is set to 16." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.668, + 0.715, + 0.684 + ], + "angle": 0, + "content": "4.1. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.905, + 0.84 + ], + "angle": 0, + "content": "We evaluate FALCON using the recently proposed MMERealWorld [66] benchmark, a comprehensive and largescale benchmark specifically designed to assess the high-resolution understanding capabilities of MLLMs. MMERealWorld comprises 13,366 high-resolution images with an average resolution of \\(2,000 \\times 1,500\\) pixels, spanning five domains: OCR in the wild, Remote Sensing, Diagrams and Tables, Monitoring, and Autonomous Driving. This benchmark is well-suited for evaluating the practical high-resolution understanding capabilities of MLLMs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.901 + ], + "angle": 0, + "content": "As shown in Tab. 1, FALCON significantly outperform other advanced MLLMs in both perception and reasoning tasks. FALCON achieves state-of-the-art performance with only 64 tokens per sub-image, resulting in a remarkable vi" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23534" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.216 + ], + "angle": 0, + "content": "Table 1. Evaluation on MME-RealWorld Benchmark. \"OCR\", \"RS\", \"DT\", \"MO\", and \"AD\" stand for OCR in the Wild, Remote Sensing, Diagram and Table, Monitoring, and Autonomous Driving, respectively. \"Avg\" denotes the weighted average accuracy across all samples, while \"Avg-C\" denotes the weighted average accuracy across different domains. \"Sizev\" indicates the image resolution supported by the model, where \\( \\times n \\) signifies a cropping-based high-resolution processing approach with up to \\( n \\) sub-images. \"Tokenc\" denotes the number of visual tokens after compression. † For models utilizing hybrid vision encoders, \"Sizev\" represents the setting of encoder with the highest resolution. We gray out the models that are either closed-source or have at least twice the number of visual tokens compared to ours. Among these, the lighter gray denotes scores below ours. The underline indicates that the second performance is achieved. Notably, FALCON achieves state-of-the-art performance with just 64 tokens per sub-image, surpassing most models that have large amount of visual tokens and many advanced closed-source commercial models." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.226, + 0.903, + 0.566 + ], + "angle": 0, + "content": "
TypeMethod\\( Size^v \\)\\( Token^c \\)PerceptionReasoning
Task SplitOCRRSDTMOADAvgAvg-COCRDTMOADAvgAvg-C
Visual Tokens\\(\\nabla 1024\\)MiniGPT-v2 [4]\\( 448^2 \\)25639.023.320.419.326.026.925.630.020.416.923.723.022.7
Qwen-VL-Chat [1]\\( 448^2 \\)25632.415.115.622.115.120.820.128.613.616.524.622.020.8
LLaVA1.5-13B [30]\\( 336^2 \\)57644.123.320.220.526.128.426.830.220.827.524.825.525.8
ShareGPT4V-13B [6]\\( 336^2 \\)57644.623.120.219.326.128.426.626.020.827.324.624.624.7
TextMonkey [33]\\( 448^2 \\times 4 \\)256 × 437.311.75.916.114.318.217.130.42.24.420.016.014.3
Mini-Gemini-7B-HD† [26]\\( 336^2 \\times 4 \\)57642.031.322.334.224.831.130.935.424.625.923.326.127.3
DeepSeek-VL† [36]\\( 1024^2 \\)57649.625.523.427.033.433.131.845.223.816.727.328.028.3
MiniCPM-V 2.5 [56]\\( 448^2 \\times 9 \\)96 × 966.827.752.838.734.247.444.044.031.837.031.034.536.0
Monkey [28]\\( 448^2 \\times 4 \\)256 × 454.625.032.528.029.736.334.027.220.827.333.028.827.1
FALCON\\( 384^2 \\times 16 \\)64 × 1666.449.647.240.436.550.348.055.439.244.434.540.743.4
Visual Tokens\\(\\nabla 2048\\)LLaVA-Next-Qwen72B [31]\\( 336^2 \\times 4 \\)576 × 437.129.127.729.418.029.028.317.234.227.329.727.927.1
LLaVA-Next-liama3 [31]\\( 336^2 \\times 4 \\)576 × 447.925.426.619.518.730.127.655.223.421.130.732.132.6
mPLUG-DocOwl 1.5 [15]\\( 448^2 \\times 9 \\)256 × 951.223.729.325.028.333.731.542.619.820.526.026.927.2
CogVLM2-liama3-Chat [14]\\( 1344^2 \\)230470.028.847.533.730.245.842.054.032.841.231.237.339.8
IXC2.5 [63]\\( 560^2 \\times 24 \\)400 × 2469.336.163.939.533.652.548.553.441.017.730.033.935.5
LLaVA-OneVision [20]\\( 384^2 \\times 9 \\)729 × 978.753.560.740.345.859.655.861.840.040.834.141.244.2
Closed-sourceGPT-4o-mini--62.56.744.226.524.237.132.847.039.825.826.832.534.9
Gemini-1.5-pro [50]--67.614.039.931.126.639.635.952.733.228.319.229.233.4
GPT-4o--77.728.946.733.922.446.441.961.444.836.526.437.642.3
Claude 3.5 Sonnet--72.525.767.432.240.852.947.761.961.241.831.944.149.2
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.574, + 0.483, + 0.83 + ], + "angle": 0, + "content": "sual compression rate of 9 times. In addition, FALCON also outperforms most models that have twice as many visual tokens. Specifically, FALCON outperforms CogVLM2-llama3-Chat [14] in average scores on both perception and reasoning task, which uses 2304 tokens per-image. It also surpasses IXC2.5 [63] in average scores on the reasoning tasks, which has a maximum token counts of 9600. Notably, FALCON achieves higher average scores in both perception and reasoning tasks compared to three advanced commercial models, GPT-4o, GPT-4o-mini and Gemini1.5-pro [50], with it notably outperforming all commercial models in the sub-tasks of remote sensing and monitoring. These results demonstrate that the ReCompact and ReAtten technique proposed by FALCON effectively eliminate visual redundancy and ensure visual continuity, allowing comprehensive and detailed visual information to be conveyed with a highly compact visual representation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.903 + ], + "angle": 0, + "content": "To further assess FALCON's capabilities in high-resolution understanding, we conduct evaluations on more high-resolution benchmarks: V* [53], DocVQA [42], TextVQA [49] and ChartQA [41]. V* focuses on evaluating" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.574, + 0.905, + 0.679 + ], + "angle": 0, + "content": "models' capabilities for fine-grained attribute recognition and spatial reasoning. DocVQA challenges models to comprehend text-dense document images. ChartQA evaluates models' proficiency in interpreting various types of chart, while TextVQA assesses the ability to recognize small text in real-world scenes. As shown in Tab. 2, FALCON demonstrates superior performance across all these benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.681, + 0.907, + 0.817 + ], + "angle": 0, + "content": "To evaluate the performance of FALCON across a broader range of domains, we further conduct experiments with diverse popular benchmarks. As shown in Tab. 4, POPE [25], ScienceQA [37], MMBench [34] and SEED-Bench [19] are selected to assess models' capabilities in hallucination, scientific reasoning, and comprehensive ability. FALCON achieves SOTA performance across all these benchmarks. The results demonstrate that FALCON exhibits superior capabilities across a wide range of tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.831, + 0.667, + 0.847 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.854, + 0.905, + 0.884 + ], + "angle": 0, + "content": "We perform extensive ablation studies to validate the effectiveness of the proposed ReCompact and ReAtten." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Ablation on Training Data. To eliminate the influx" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23535" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.089, + 0.486, + 0.133 + ], + "angle": 0, + "content": "Table 2. Evaluations on diverse high-resolution Benchmarks. \"V* \\( {}_{Avg} \\) \" represents the weighted average score for the two subtasks in the \\( {V}^{ * } \\) benchmark. † denotes OCR inputs are utilized." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.143, + 0.473, + 0.34 + ], + "angle": 0, + "content": "
ModelV*AvgDocVQATextVQAChartQA
Low-resolution MLLMs
Qwen-VL [1]-65.163.859.0
InstructBLIP [9]34.04.550.710.9
LLaVA-v1.5-7B [30]48.78.558.2†17.8
mPLUG-Owl2 [60]-17.958.2†22.8
High-resolution MLLMs
UReader [58]-65.457.659.3
Monkey [28]-66.565.859.0
LlvaF-FlexAttn [23]54.4-48.9-
S2-Wrapper-7B [48]55.5-61.0-
LlvaNxt-Vicuna7B [31]61.8-64.954.3
SliME-7B [67]--64.433.9
FALCON64.474.466.568.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.344, + 0.485, + 0.4 + ], + "angle": 0, + "content": "Table 3. Ablation results on training data. We train FALCON with two-stage pipeline and exactly the same dataset as LLaVAv1.5 [30]. The results confirm the superiority of the proposed method when the influence of training data is excluded." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.412, + 0.485, + 0.574 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.583, + 0.483, + 0.655 + ], + "angle": 0, + "content": "Figure 4. Comparison of different compression methods. We report the sum of the \"Avg-C\" scores for the perception and reasoning tasks in MME-RealWorld. The superior performance demonstrates that ReCompact effectively eliminates redundancy and produces a more expressive and compact visual representation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.484, + 0.748 + ], + "angle": 0, + "content": "ence of training data and demonstrate the advantages of the proposed method, we utilize exactly the same dataset as LLaVA-v1.5 [30] and employ a two-stage pipeline for training. As shown in Tab. 3, the proposed method exhibits significant performance advantages across multiple benchmarks when trained on the same dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.87 + ], + "angle": 0, + "content": "Effectiveness of ReCompact mechanism. To verify the advantages of the proposed ReCompact mechanism in reducing visual redundancy and compressing visual representations during the visual encoding, we compare the Re-Compact with other commonly used compression methods. This includes pooling [48, 55], Pixel Shuffle [7] and Abstractor [60], where the Abstractor serves as a representative module that combines query tokens with cross-attention." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "As shown in Fig. 4, ReCompact demonstrates superior performance compared to other compression methods." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.907, + 0.133 + ], + "angle": 0, + "content": "Table 4. Evaluation on benchmarks across diverse domains. The POPE, SQA, MMBench and SEED-Bench are used to assess hallucination, scientific reasoning, and overall abilities, respectively." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.143, + 0.897, + 0.328 + ], + "angle": 0, + "content": "
ModelSQAPOPEMMBSEED-Img
Low-Resolution MLLMs
Qwen-VL [1]67.1-41.662.3
InstructBLIP [9]63.178.938.458.8
LlaVA-v1.5-7B [30]66.885.962.865.8
mPLUG-Owl2 [60]68.785.863.564.5
High-Resolution MLLMs
Monkey [28]69.483.757.464.3
LlaVA-FlexAttn [23]-85.965.7-
S2-Wrapper-7B [48]--66.267.9
LLaVA-Next-Vicuna7B [31]70.186.566.569.6
SliME-7B [67]76.885.463.167.0
FALCON88.888.076.173.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.332, + 0.907, + 0.43 + ], + "angle": 0, + "content": "Table 5. Comparison of different methods for maintaining visual continuity. The \"Baseline\" model refers to a condition where no specific methods are applied to preserve visual continuity. We report the \"Avg-C\" score for MME-RealWorld. The results show that the ReAtten module more effectively addresses the issue of visual fragmentation compared to other methods, thereby improving overall performance and reducing hallucinations." + }, + { + "type": "table", + "bbox": [ + 0.545, + 0.441, + 0.872, + 0.53 + ], + "angle": 0, + "content": "
MethodV*AVgMME-RealWorldPOPE
PerceptionReasoning
Baseline51.338.235.685.7
CIP [16]50.338.437.286.3
W-Atten [33]60.241.038.186.4
ReAtten61.342.139.087.3
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.907, + 0.702 + ], + "angle": 0, + "content": "These results indicate that ReCompact effectively eliminates visual redundancy and produces a more expressive compact visual representation. In contrast, we argue that the pooling layer and Pixel Shuffle merely reduces the number of tokens by merging adjacent ones, failing to eliminate redundancy that interferes with the LLM's understanding of visual content. Additionally, although Abstractor uses more parameters, its performance is generally lower than that of pooling and ReCompact. This result suggests that even millions of samples may be insufficient for fully training Abstractor, which diminishes their practical efficacy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.703, + 0.906, + 0.795 + ], + "angle": 0, + "content": "Effectiveness of ReAtten Module. To validate the advantage of the proposed ReAtten Module, we compare ReAtten with other contemporary methods, including the Shifted Window Attention (W-Attn) [35] adopted by TextMonkey [33] and the Complementary Image Pyramid (CIP) introduced by MiniMonkey [16]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.903 + ], + "angle": 0, + "content": "As shown in Tab. 5, the model incorporating ReAtten achieves significant improvements over the baseline across all benchmarks and demonstrates superiority compared to other methods. These results indicate that the ReAtten module effectively enhances the ability of the model in high-resolution perception and mitigates hallucinations by addressing the issue of visual fragmentation. In contrast, we" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23536" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.111, + 0.334, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.155, + 0.251, + 0.276, + 0.267 + ], + "angle": 0, + "content": "(a) Input Image" + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.093, + 0.468, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.169, + 0.468, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.346, + 0.25, + 0.468, + 0.265 + ], + "angle": 0, + "content": "(b) Register-to-1" + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.094, + 0.588, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.169, + 0.588, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.468, + 0.25, + 0.588, + 0.265 + ], + "angle": 0, + "content": "Image Attention" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.094, + 0.829, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.6, + 0.171, + 0.829, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.625, + 0.25, + 0.806, + 0.265 + ], + "angle": 0, + "content": "(c) Ablation on ReAtten" + }, + { + "type": "text", + "bbox": [ + 0.835, + 0.124, + 0.892, + 0.151 + ], + "angle": 0, + "content": "w/ ReAtten" + }, + { + "type": "text", + "bbox": [ + 0.835, + 0.197, + 0.892, + 0.222 + ], + "angle": 0, + "content": "w/o ReAtten" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.281, + 0.907, + 0.366 + ], + "angle": 0, + "content": "Figure 5. Visualization of the register-to-image attention map. We randomly select visual registers and extract the attention scores that each register assigns to image tokens. These scores are then visualized as heatmaps overlaid on the input image. The results in (b) show that each visual register focuses on specific parts of the image, capturing rich visual patterns. Meanwhile, registers pay minimal attention to background areas, effectively avoiding the inclusion of redundancy. We also conduct qualitative ablation studies of the ReAtten module. As shown in (c), in the ViT model without ReAtten, the attention patterns across different sub-images appear extremely fragmented. In contrast, the ViT model with ReAtten shows continuous attention patterns, indicating effective information interaction between sub-images." + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.37, + 0.462, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.497, + 0.483, + 0.58 + ], + "angle": 0, + "content": "Figure 6. Ablation on different visual register counts. Increasing the number of visual registers generally enhances accuracy across both benchmarks. However, the improvement rate noticeably tapers off when the register count exceeds 64. Based on this observation, to better balance the efficiency and performance, we select register counts of 64 for FALCON." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.583, + 0.483, + 0.659 + ], + "angle": 0, + "content": "argue that the information exchange in W-Atten is limited to a local window, which hinders sufficient information exchange among all sub-images. Meanwhile, CIP introduces further redundancy by repeatedly encoding the same visual content, which can even lead to a decline in performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Ablation on the Number of Visual Registers. To investigate how visual register count affects model performance, we conduct experiments with visual register counts set to 36, 64, and 144. These counts correspond to visual token compression rates of \\(16 \\times\\), \\(9 \\times\\), and \\(4 \\times\\), respectively. As shown in Fig. 6, increasing the number of visual registers generally enhances model's high-resolution comprehension capabilities and reduces hallucinations. However, when the register count exceeds 64, the improvement gains start tapering off noticeably. Meanwhile, increasing visual registers demands more computational resources. Compared to the baseline model with 36 registers, models configured with 64 and 144 registers require 1.15 and 2.41 times more training time, respectively. Base on these observations, to achieve a better balance between efficiency and performance, we select 64 as the register counts for FALCON." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.37, + 0.706, + 0.387 + ], + "angle": 0, + "content": "4.3. Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.392, + 0.907, + 0.695 + ], + "angle": 0, + "content": "To qualitatively analyze the proposed visual register technique, we visualize the register-to-image attention maps within ViT. Specifically, for a given visual register, we extract the attention scores it assigns to image tokens in each sub-image and concatenate them to construct a complete heatmap. As shown in Fig. 5(b), each visual register focuses on relatively distinct areas of the image, highlighting that it captures a diverse range of visual details. Moreover, the visual registers tend to avoid background areas, indicating their effectiveness in identifying and discarding visual redundancy. These properties enable visual registers to effectively produce a comprehensive and non-redundant visual representation. Additionally, to qualitatively evaluate the effectiveness of the ReAtten module, we visualize the attention maps within ViT models with and without ReAtten. As shown in Fig. 5(c), in the ViT model without ReAtten, the attention patterns across different sub-images appear highly fragmented. In contrast, the ViT model with ReAtten exhibits continuous attention patterns, indicating effective information exchange between sub-images." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.709, + 0.634, + 0.725 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.903 + ], + "angle": 0, + "content": "To address visual redundancy and fragmentation in high-resolution MLLMs, we propose FALCON. FALCON employs an innovative visual register technique that simultaneously addresses both challenges. This technique uses a Re-Compact mechanism to adaptively aggregate essential visual information through visual registers, creating a compact, non-redundant representation. Additionally, a ReAtten module is introduced to further enhances visual continuity by enabling information exchange among sub-images via visual registers. Extensive experiments demonstrate FALCON's superiority in high-resolution understanding." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23537" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.271, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.484, + 0.252 + ], + "angle": 0, + "content": "This study is supported by National Natural Science Foundation of China (Grant No. 62306090, No. U24A20328), Natural Science Foundation of Guangdong Province of China (Grant No. 2024A1515010147), Shenzhen Science and Technology Program (KQTD20240729102207002), Natural Science Foundation of Beijing, China (Grant No. L254018), Jiangsu Science and Technology Major Program (Grant No. BG2024041) and China Postdoctoral Science Foundation (Grant No. 2024M764192)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.268, + 0.188, + 0.284 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.293, + 0.484, + 0.363 + ], + "angle": 0, + "content": "[1] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 1(2):3, 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.364, + 0.484, + 0.434 + ], + "angle": 0, + "content": "[2] Gongwei Chen, Leyang Shen, Rui Shao, Xiang Deng, and Liqiang Nie. Lion: Empowering multimodal large language model with dual-level visual knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26540-26550, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.435, + 0.484, + 0.517 + ], + "angle": 0, + "content": "[3] Gongwei Chen, Xurui Zhou, Rui Shao, Yibo Lyu, Kaiwen Zhou, Wentao Li, Yinchuan Li, Zhongang Qi, and Liqiang Nie. Less is more: Empowering gui agent with context-aware simplification. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.52, + 0.484, + 0.603 + ], + "angle": 0, + "content": "[4] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.605, + 0.484, + 0.702 + ], + "angle": 0, + "content": "[5] Jingxuan Chen, Derek Yuen, Bin Xie, Yuhao Yang, Gongwei Chen, Zhihao Wu, Li Yixing, Xurui Zhou, Weiwen Liu, Shuai Wang, Kaiwen Zhou, Rui Shao, Liqiang Nie, Yasheng Wang, Jianye HAO, Jun Wang, and Kun Shao. Spa-bench: A comprehensive benchmark for smartphone agent evaluation. In The Thirteenth International Conference on Learning Representations, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.704, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[6] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. In European Conference on Computer Vision, pages 370-387. Springer, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.775, + 0.484, + 0.858 + ], + "angle": 0, + "content": "[7] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internv1: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24185–24198, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.86, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[8] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.293, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.134 + ], + "angle": 0, + "content": "chatbot impressing gpt-4 with \\(90\\%\\) chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2(3):6, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.136, + 0.906, + 0.219 + ], + "angle": 0, + "content": "[9] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.906, + 0.275 + ], + "angle": 0, + "content": "[10] Timothee Darcet, Maxime Aquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. In The Twelfth International Conference on Learning Representations, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.906, + 0.374 + ], + "angle": 0, + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.906, + 0.445 + ], + "angle": 0, + "content": "[12] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.905, + 0.49 + ], + "angle": 0, + "content": "[13] Dan Hendrycks and Kevin Gimpel. Bridging nonlinearities and stochastic regularizers with gaussian error linear units. arXiv preprint arXiv:1606.08415, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.906, + 0.559 + ], + "angle": 0, + "content": "[14] Wenyi Hong, Weihan Wang, Ming Ding, Wenmeng Yu, Qingsong Lv, Yan Wang, Yean Cheng, Shiyu Huang, Junhui Ji, Zhao Xue, et al. Cogvlm2: Visual language models for image and video understanding. arXiv preprint arXiv:2408.16500, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.562, + 0.906, + 0.645 + ], + "angle": 0, + "content": "[15] Anwen Hu, Haiyang Xu, Jiabo Ye, Ming Yan, Liang Zhang, Bo Zhang, Ji Zhang, Qin Jin, Fei Huang, and Jingren Zhou. mplug-docowl 1.5: Unified structure learning forOCR-free document understanding. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 3096-3120, 2024. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.647, + 0.906, + 0.716 + ], + "angle": 0, + "content": "[16] Mingxin Huang, Yuliang Liu, Dingkang Liang, Lianwen Jin, and Xiang Bai. Mini-monkey: Alleviating the semantic saw-tooth effect for lightweight MLLMs via complementary image pyramid. In The Thirteenth International Conference on Learning Representations, 2025. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.718, + 0.906, + 0.801 + ], + "angle": 0, + "content": "[17] Runhui Huang, Xinpeng Ding, Chunwei Wang, Jianhua Han, Yulong Liu, Hengshuang Zhao, Hang Xu, Lu Hou, Wei Zhang, and Xiaodan Liang. Hires-llava: Restoring fragmentation input in high-resolution large vision-language models. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 29814–29824, 2025. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.802, + 0.906, + 0.872 + ], + "angle": 0, + "content": "[18] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan P Foster, Pannag R Sanketi, Quan Vuong, et al. Openvla: An open-source vision-language-action model. In 8th Annual Conference on Robot Learning, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.906, + 0.902 + ], + "angle": 0, + "content": "[19] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking mul" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23538" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "timodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.188 + ], + "angle": 0, + "content": "[20] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. LLaVA-onevision: Easy visual task transfer. Transactions on Machine Learning Research, 2025. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.484, + 0.259 + ], + "angle": 0, + "content": "[21] Hao Li, Qi Lv, Rui Shao, Xiang Deng, Yinchuan Li, Jianye HAO, and Liqiang Nie. STAR: Learning diverse robot skill abstractions through rotation-augmented vector quantization. In Forty-second International Conference on Machine Learning, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.484, + 0.328 + ], + "angle": 0, + "content": "[22] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.484, + 0.397 + ], + "angle": 0, + "content": "[23] Junyan Li, Delin Chen, Tianle Cai, Peihao Chen, Yining Hong, Zhenfang Chen, Yikang Shen, and Chuang Gan. Flex-attention for efficient high-resolution vision-language models. In European Conference on Computer Vision, pages 286–302, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.483, + 0.455 + ], + "angle": 0, + "content": "[24] Wei Li, Bing Hu, Rui Shao, Leyang Shen, and Liqiang Nie. Lion-fs: Fast & slow video-language thinker as online video assistant. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 3240-3251, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.457, + 0.483, + 0.511 + ], + "angle": 0, + "content": "[25] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In The 2023 Conference on Empirical Methods in Natural Language Processing, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.512, + 0.483, + 0.579 + ], + "angle": 0, + "content": "[26] Yanwei Li, Yuechen Zhang, Chengyao Wang, Zhisheng Zhong, Yixin Chen, Ruihang Chu, Shaoteng Liu, and Jiaya Jia. Mini-gemini: Mining the potential of multi-modality vision language models. arXiv preprint arXiv:2403.18814, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.483, + 0.636 + ], + "angle": 0, + "content": "[27] Zaijing Li, Yuquan Xie, Rui Shao, Gongwei Chen, Dongmei Jiang, and Liqiang Nie. Optimus-1: Hybrid multimodal memory empowered agents excel in long-horizon tasks. In Advances in neural information processing systems, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.638, + 0.483, + 0.719 + ], + "angle": 0, + "content": "[28] Zhang Li, Biao Yang, Qiang Liu, Zhiyin Ma, Shuo Zhang, Jingxu Yang, Yabo Sun, Yuliang Liu, and Xiang Bai. Monkey: Image resolution and text label are important things for large multi-modal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26763-26773, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.721, + 0.483, + 0.788 + ], + "angle": 0, + "content": "[29] Zaijing Li, Yuquan Xie, Rui Shao, Gongwei Chen, Dongmei Jiang, and Liqiang Nie. Optimus-2: Multimodal minercraft agent with goal-observation-action conditioned policy. In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[30] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[31] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,ocr,and world knowledge,2024.6,7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[32] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[33] Yuliang Liu, Biao Yang, Qiang Liu, Zhang Li, Zhiyin Ma, Shuo Zhang, and Xiang Bai. Textmonkey: AnOCR-free large multimodal model for understanding document. arXiv preprint arXiv:2403.04473, 2024. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[34] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European Conference on Computer Vision, pages 216-233. Springer, 2025. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[35] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.331, + 0.905, + 0.397 + ], + "angle": 0, + "content": "[36] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Yaofeng Sun, et al. Deepseek-vl: towards real-world vision-language understanding. arXiv preprint arXiv:2403.05525, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[37] Pan Lu, Swaroop Mishra, Tony Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In The 36th Conference on Neural Information Processing Systems (NeurIPS), 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[38] Gen Luo, Yiyi Zhou, Yuxin Zhang, Xiawu Zheng, Xiaoshuai Sun, and Rongrong Ji. Feast your eyes: Mixture-of-resolution adaptation for multimodal large language models. arXiv preprint arXiv:2403.03003, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.54, + 0.908, + 0.621 + ], + "angle": 0, + "content": "[39] Qi Lv, Hao Li, Xiang Deng, Rui Shao, Yinchuan Li, Jianye Hao, Longxiang Gao, Michael Yu Wang, and Liqiang Nie. Spatial-temporal graph diffusion policy with kinematic modeling for bimanual robotic manipulation. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 17394-17404, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[40] Yibo Lyu, Rui Shao, Gongwei Chen, Yijie Zhu, Weili Guan, and Liqiang Nie. Puma: Layer-pruned language model for efficient unified multimodal retrieval with modality-adaptive learning. In Proceedings of the 33nd ACM International Conference on Multimedia, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.693, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[41] Ahmed Masry, Xuan Long Do, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. In Findings of the Association for Computational Linguistics: ACL 2022, pages 2263-2279, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.763, + 0.907, + 0.816 + ], + "angle": 0, + "content": "[42] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23539" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[44] Rui Shao, Xiangyuan Lan, Jiawei Li, and Pong C Yuen. Multi-adversarial discriminative deep domain generalization for face presentation attack detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10023-10031, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[45] Rui Shao, Tianxing Wu, and Ziwei Liu. Detecting and grounding multi-modal media manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6904-6913, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.482, + 0.273 + ], + "angle": 0, + "content": "[46] Rui Shao, Tianxing Wu, Jianlong Wu, Liqiang Nie, and Zwei Liu. Detecting and grounding multi-modal media manipulation and beyond. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.273, + 0.482, + 0.33 + ], + "angle": 0, + "content": "[47] Leyang Shen, Gongwei Chen, Rui Shao, Weili Guan, and Liqiang Nie. Mone: Mixture of multimodal experts for generalist multimodal large language models. In Advances in neural information processing systems, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.384 + ], + "angle": 0, + "content": "[48] Baifeng Shi, Ziyang Wu, Maolin Mao, Xin Wang, and Trevor Darrell. When do we not need larger vision models? In European Conference on Computer Vision, pages 444-462, 2024. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.385, + 0.482, + 0.455 + ], + "angle": 0, + "content": "[49] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.482, + 0.523 + ], + "angle": 0, + "content": "[50] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.524, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[51] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[52] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.664, + 0.482, + 0.72 + ], + "angle": 0, + "content": "[53] Penghao Wu and Saining Xie. V*: Guided visual search as a core mechanism in multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13084-13094, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.482, + 0.789 + ], + "angle": 0, + "content": "[54] Bin Xie, Rui Shao, Gongwei Chen, Kaiwen Zhou, Yinchuan Li, Jie Liu, Min Zhang, and Liqiang Nie. Gui-explorer: Autonomous exploration and mining of transition-aware knowledge for gui agent. In Annual Meeting of the Association for Computational Linguistics (ACL), 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[55] Linli Yao, Lei Li, Shuhuai Ren, Lean Wang, Yuanxin Liu, Xu Sun, and Lu Hou. Deco: Decoupling token compression from semantic abstraction in multimodal large language models. arXiv preprint arXiv:2405.20985, 2024. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[56] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[57] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Yuhao Dan, Chenlin Zhao, Guohai Xu, Chenliang Li, Junfeng Tian, et al. mplug-docowl: Modularized multimodal large language model for document understanding. arXiv preprint arXiv:2307.02499, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[58] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Guohai Xu, Chenliang Li, Junfeng Tian, Qi Qian, Ji Zhang, et al. Ureader: UniversalOCR-free visually-situated language understanding with multimodal large language model. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 2841-2858, 2023. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.316 + ], + "angle": 0, + "content": "[59] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.905, + 0.401 + ], + "angle": 0, + "content": "[60] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Anwen Hu, Haowei Liu, Qi Qian, Ji Zhang, and Fei Huang. mplug-owl2: Revolutionizing multi-modal large language model with modality collaboration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13040–13051, 2024. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.402, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[61] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In European Conference on Computer Vision (ECCV), pages 69-85. Springer, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.459, + 0.905, + 0.515 + ], + "angle": 0, + "content": "[62] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.515, + 0.905, + 0.584 + ], + "angle": 0, + "content": "[63] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, et al. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output. arXiv preprint arXiv:2407.03320, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.585, + 0.905, + 0.641 + ], + "angle": 0, + "content": "[64] Renshan Zhang, Yibo Lyu, Rui Shao, Gongwei Chen, Weili Guan, and Liqiang Nie. Token-level correlation-guided compression for efficient multimodal document understanding. arXiv preprint arXiv:2407.14439, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.642, + 0.905, + 0.738 + ], + "angle": 0, + "content": "[65] Xiaofeng Zhang, Yihao Quan, Chen Shen, Xiaosong Yuan, Shaotian Yan, Liang Xie, Wenxiao Wang, Chaochen Gu, Hao Tang, and Jieping Ye. From redundancy to relevance: Enhancing explainability in multimodal large language models. Annual Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.739, + 0.905, + 0.837 + ], + "angle": 0, + "content": "[66] YiFan Zhang, Huanyu Zhang, Haochen Tian, Chaoyou Fu, Shuangqing Zhang, Junfei Wu, Feng Li, Kun Wang, Qing-song Wen, Zhang Zhang, Liang Wang, and Rong Jin. MMErealworld: Could your multimodal LLM challenge high-resolution real-world scenarios that are difficult for humans? In The Thirteenth International Conference on Learning Representations, 2025. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.838, + 0.905, + 0.895 + ], + "angle": 0, + "content": "[67] Yi-Fan Zhang, Qingsong Wen, Chaoyou Fu, Xue Wang, Zhang Zhang, Liang Wang, and Rong Jin. Beyond llava-hd: Diving into high-resolution large multimodal models. arXiv preprint arXiv:2406.08487, 2024. 1, 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.895 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23540" + } + ] +] \ No newline at end of file diff --git a/2025/FALCON_ Resolving Visual Redundancy and Fragmentation in High-resolution Multimodal Large Language Models via Visual Registers/3c258128-6a38-46a9-bf2a-aabe422e5840_origin.pdf b/2025/FALCON_ Resolving Visual Redundancy and Fragmentation in High-resolution Multimodal Large Language Models via Visual Registers/3c258128-6a38-46a9-bf2a-aabe422e5840_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3e385d798016bfb8d00444de11b4f36081f2537b --- /dev/null +++ b/2025/FALCON_ Resolving Visual Redundancy and Fragmentation in High-resolution Multimodal Large Language Models via Visual Registers/3c258128-6a38-46a9-bf2a-aabe422e5840_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:075df46486e01a567d03842cc69b9c6fa1c261d4b17210f41149259820e81c24 +size 1696542 diff --git a/2025/FALCON_ Resolving Visual Redundancy and Fragmentation in High-resolution Multimodal Large Language Models via Visual Registers/full.md b/2025/FALCON_ Resolving Visual Redundancy and Fragmentation in High-resolution Multimodal Large Language Models via Visual Registers/full.md new file mode 100644 index 0000000000000000000000000000000000000000..66d4a0a10b0471347fd02f9ebf3bb9acac3b0af4 --- /dev/null +++ b/2025/FALCON_ Resolving Visual Redundancy and Fragmentation in High-resolution Multimodal Large Language Models via Visual Registers/full.md @@ -0,0 +1,286 @@ +# FALCON: Resolving Visual Redundancy and Fragmentation in High-resolution Multimodal Large Language Models via Visual Registers + +Renshan Zhang $^{1}$ , Rui Shao $^{1\dagger}$ , Gongwei Chen $^{1}$ , Miao Zhang $^{1}$ , Kaiwen Zhou $^{2}$ , Weili Guan $^{1}$ , Liqiang Nie $^{1\dagger}$ $^{1}$ Harbin Institute of Technology, Shenzhen $^{2}$ Huawei Noah's Ark Lab + +zhangrenshan@stu.hit.edu.cn shaorui@hit.edu.cn + +chengongwei@hit.edu.cn nieliqiang@gmail.com + +https://github.com/JiuTian-VL/FALCON + +# Abstract + +The incorporation of high-resolution visual input equips multimodal large language models (MLLMs) with enhanced visual perception capabilities for real-world tasks. However, most existing high-resolution MLLMs rely on a cropping-based approach to process images, which leads to fragmented visual encoding and a sharp increase in redundant tokens. To tackle these issues, we propose the FALCON model. FALCON introduces a novel visual register technique to simultaneously: 1) Eliminate redundant tokens at the stage of visual encoding. To directly address the visual redundancy present in the output of vision encoder, we propose a Register-based Representation Compacting (ReCompact) mechanism. This mechanism introduces a set of learnable visual registers designed to adaptively aggregate essential information while discarding redundancy. It enables the encoder to produce a more compact visual representation with a minimal number of output tokens, thus eliminating the need for an additional compression module. 2) Ensure continuity in visual encoding. To address the potential encoding errors caused by fragmented visual inputs, we develop a Register Interactive Attention (ReAtten) module. This module facilitates effective and efficient information exchange across sub-images by enabling interactions between visual registers. It ensures the continuity of visual semantics throughout the encoding. We conduct comprehensive experiments with FALCON on high-resolution benchmarks across a wide range of scenarios. FALCON demonstrates superior performance with a remarkable 9-fold reduction in visual tokens. + +# 1. Introduction + +Recently, multimodal large language models (MLLMs) [2, 9, 30, 32, 59, 60] have made significant advancements in + +![](images/d729d5f337aba80dea70000a3b80a55cd315708e8941f483cec9704ffe13e941.jpg) +Figure 1. Comparison between existing high-resolution MLLMs and FALCON. FALCON eliminates visual redundancy and ensures the continuity of high-resolution encoding through the use of ReCompact and ReAtten mechanisms. + +performing various vision-language tasks [45, 46, 61], facilitating more natural human-machine interactions. Early MLLMs [9, 32, 59] typically processed images at a fixed low resolution, such as $224 \times 224$ . This limited their capabilities in many real-world applications that require high-resolution input. To equip MLLMs with high-resolution visual perception, recent efforts [15, 17, 33, 58, 67] have introduced a cropping-based high-resolution processing method. Specifically, As shown in Fig. 1(b), this approach crops high-resolution images into multiple non-overlapping sub-images that match the pre-trained resolution of the encoder. Each sub-image is processed independently by the vision encoder. Then the resulting visual features are concatenated and fed into the large language model (LLM). + +Although effective, cropping-based methods still face issues of: 1) Visual Redundancy. The number of visual to- + +kens grows dramatically with increasing resolution. While this enriches the visual details, it also leads to an increase in redundant tokens. These redundant tokens significantly increase the computational burden on LLM and interfere with the processing of visual information [65]. To address these issues, as illustrated in Fig. 1(b), existing methods typically rely on additional compression techniques. Representative approaches include combining learnable queries with cross-attention [1, 28, 60] or simply employing pooling layers [48, 55]. However, these methods either demand substantial training resources or fail to fully address redundancy. 2) Visual Fragmentation. Cropping an image into multiple non-overlapping sub-images may disrupt visual semantics and potentially lead to encoding errors. As shown in Fig. 1(a), in an OCR task, the word "pineapple" may be incorrectly segmented into "pine" and "apple", causing the LLM to misinterpret the word. Similarly, fragmentation can also hinder the identification of objects. For example, the "rubin vase" pattern illustrated in Fig. 1(a) depicts both a vase and two faces simultaneously. However, the fragmentation could lead to the vase being unrecognizable. + +To address these issues, as shown in Fig. 1(c), we propose the FALCON model. FALCON introduces a novel visual register technique to simultaneously: 1) Eliminate redundant tokens during visual encoding. To eliminate redundancy directly at the stage of visual encoding, we propose a novel Register-based Representation Compacting (ReCompact) mechanism. This mechanism introduces a set of learnable visual registers as additional inputs to the vision encoder. The registers are designed to adaptively aggregate essential information from image tokens and discard redundant content. Then the outputs corresponding to the visual registers are considered as a non-redundant visual representation, serving as the compact input for the LLM instead of a large number of image tokens. This mechanism enables the vision encoder to convey comprehensive visual information in a compact representation using minimal tokens, thus removing the need for an additional compression module. 2) Ensure continuity in visual encoding. To avoid the potential encoding errors caused by fragmented visual inputs, we develop a novel Register Interactive Attention (ReAtten) module. This module leverages the compact and comprehensive representation produced by visual registers to facilitate information exchange across sub-images. It gathers visual registers from different sub-images and employs a specialized cross-ViT attention module to enable interactions among them. Such interaction enables thorough information exchange across the entire image, thus ensuring visual continuity throughout the encoding process. Additionally, the compactness of the visual registers further enhances the efficiency of these interactions. Consequently, this module provides LLMs with a reliable foundation to accurately interpret visual content and reduce hallucinations. + +Note that the widely used Abstractor-like modules [28, 33, 58, 60], including QFormer [9], ReSampler [1], follow a similar approach by using query tokens to extract information from image tokens after visual encoding. The key difference is that the visual register mechanism directly adopts the pretrained ViT to aggregate information into a set of specified register tokens. This design is based on the observation that a well-pretrained ViT can naturally gather global information into some specific tokens through self-attention [10]. It offers the visual register mechanism several strengths in terms of: 1) Training Efficiency. The substantial number of randomly initialized parameters in Abstractor-like modules requires extensive data for pretraining. For instance, QFormer [22], Abstractor [60], and Resampler [1] use large corpora of 129M, 400M, and 1.4B samples, respectively. In contrast, by leveraging well-pretrained ViT parameters, the visual register mechanism requires only 3.5M samples for pretraining. The ablation results in Fig. 4 further demonstrate this advantages. 2) Addressing Visual Fragmentation. By leveraging the compact and comprehensive representation produced by visual registers, the proposed ReAtten module can effectively facilitate information exchange across sub-images. This is an important function that is notably lacking in Abstractor-like modules. Combined with ReCompact, it enables the visual register mechanism to simultaneously address the issues of visual redundancy and fragmentation. The ablation results in Tab. 5 validate the effectiveness of ReAtten. + +We summarize our main contributions as follows: + +- To address the issue of visual redundancy directly at the encoding phase, we introduce the Register-based Representation Compacting (ReCompact) mechanism. This mechanism enables vision encoder to capture rich visual information using only a minimal number of tokens, effectively eliminating redundancy and obviating the need for additional compression modules. +- To address the issue of visual fragmentation caused by image cropping, we propose the register interactive attention (ReAtten) module. This module ensures the semantic continuity of high-resolution visual encoding in an efficient and effective manner. +- We introduce FALCON, a model that leverages the proposed visual register technique. Comprehensive evaluations on high-resolution visual understanding benchmarks highlight the superiority of FALCON. Notably, FALCON achieves exceptional results with a remarkable reduction in visual tokens by $9 \times$ . + +# 2. Related Work + +# 2.1. Multimodal Large Language Models + +Research on MLLMs [1, 2, 30, 47, 59, 60] has made significant strides in recent years. By incorporating image in + +![](images/2f27ac45349ad0f7f21ad2a796584b3c4076140f766605856e347b5f07931124.jpg) +Figure 2. Overall framework of the proposed FALCON. The model introduces a set of learnable visual registers, which are fed into the ViT along with the image tokens from each sub-image. During the visual encoding, the ReCompact mechanism enables information from the image tokens to be aggregated into the visual registers, while the proposed ReAtten facilitates information exchange among the visual registers across different sub-images. Ultimately, these visual registers provide a compact visual representation as input to the LLM. + +puts into pretrained LLMs [8, 51, 52], MLLMs demonstrate strong capabilities across various vision-language tasks. Typically, these models use a pretrained vision encoder [43, 62] to transform input images into a set of visual tokens. These tokens are then projected into the input space of the LLM through a projection module (e.g., MLP [30, 32], QFormer [9] and Abstractor [60]) and processed alongside text inputs. Building on the progress of MLLMs, recent studies have shifted from designing specialized methods for certain tasks [44-46] to employing MLLMs as generalist agents, capable of handling diverse practical tasks such as document analysis [15, 57, 64], online video assistant [24], retrieval-augmented generation [40], game agents [27, 29], robot manipulation [18, 21, 39], and GUI agent [3, 5, 54]. Despite these advances, MLLMs still struggle with the fixed low resolution of pretrained image encoders, which can limit their effectiveness in handling high-resolution images. + +# 2.2. High-resolution MLLMs + +To address the limitation of MLLMs in handling only low-resolution visual inputs, early studies attempted to either directly incorporate vision encoders that were pre-trained on higher-resolution images [36, 38], or to improve the resolution of pre-trained encoders through additional training [1, 60]. While achieving some progress, they still remain restricted by the fixed resolutions. To accommodate the varying resolutions and aspect ratios, cropping-based techniques have been widely adopted in recent works [15, 17, 33, 58, 67]. However, these approaches still face challenges related to visual redundancy and fragmentation. + +To mitigate the issue of redundancy, existing models typically require an additional compression module. For example, Abstractor-like modules utilizing learnable queries combined with cross-attention are widely adopted [1, 28, 33, 58, 60]. These modules typically require pretraining on massive amounts of data, which diminishes their practical efficacy. In contrast, some approaches use pooling + +layers [48, 55] or convolutional layers [15, 36]. Though resource-efficient, they often fail to address the issue of redundancy. In this work, we present an initial exploration to directly eliminate redundancy during the visual encoding phase, thereby removing the need for additional compression. To address the issue of fragmentation, several concurrent studies also make preliminary investigation. TextMonkey [33] uses Shifted Window Attention [35] to exchange information between sub-images. But the exchange is still limited to local regions. MiniMonkey [16] proposed a Complementary Image Pyramid strategy to generate complementary cropping schemes. However, this strategy inevitably encodes the same visual content multiple times, further exacerbating visual redundancy. In this work, we propose the ReAtten module to facilitate continuous visual encoding through the interaction among visual registers. + +# 3. FALCON + +The overall framework of FALCON is illustrated in Fig. 2. The method begins with shape-adaptive cropping [58], where the image is cropped into multiple non-overlapping sub-images, each matching the resolution of the pre-trained vision encoder. Additionally, a resized version of the full image is used to supplement global layout information. To directly eliminate redundancy during visual encoding, we propose Register-based Representation Compacting (Re-Compact) mechanism. This mechanism introduces a set of learnable visual registers, which are paired with image tokens from each sub-image and fed into the vision encoder to capture rich visual information. To ensure the continuity of visual semantics throughout the encoding, a novel Register Interactive Attention (ReAtten) module is integrated into the Vision Transformer (ViT) [11] to facilitate information exchange between sub-images via the visual registers. Finally, the compact visual representations produced by the visual registers are processed through a simple MLP module before being fed into the LLM for further analysis. + +# 3.1. Mitigating Visual Redundancy via ReCompact + +Existing MLLMs typically use a ViT-based vision encoder, such as CLIP [43]. For an input image of size $H \times W$ , it is patched into multiple $P \times P$ patches. Each patch is treated as an individual image token, resulting in $N = \frac{H}{P} \times \frac{W}{P}$ tokens, denoted as $I = \{i_1, \dots, i_N\}$ . In most current MLLMs, output features corresponding to all image tokens are usually preserved in their entirety. However, when high-resolution images are included, the number of image tokens increases dramatically, placing a substantial burden on the computational resources of LLMs. Consequently, most high-resolution MLLMs necessitate an additional compression module, which can be resource-inefficient. Previous research has shown that most image tokens contain only localized information from their surrounding areas [10]. It suggests that a significant portion of these tokens may be redundant, particularly those originating from background regions. This observation motivates us to directly eliminate redundancy in the output of the vision encoder, thereby reducing the number of visual tokens and avoiding the need for an additional compression module. + +To achieve this, we propose a novel Register-based Representation Compacting (ReCompact) mechanism. This mechanism employs a set of visual registers to aggregate essential information from image tokens, resulting in a more compact visual representation. Specifically, as shown in Fig. 2, we introduce a set of learnable visual registers, denoted by $R = \{r_1,\dots ,r_M\}$ , where $M$ represents the number of visual registers, and $M\ll N$ . Given a single input image, the image tokens are concatenated with visual registers to create a new input for ViT, represented as $V = \mathrm{Concat}(I,R) = \{i_1,\dots ,i_N,r_1,\dots ,r_M\}$ . During visual encoding, the self-attention mechanism in ViT enables the visual register to adaptively aggregate important information from image tokens while discarding redundant content. Denote $X_{k,l}\in \mathcal{R}^{(N + M)\times D}$ as the input hidden states of the $k$ -th sub-image at the $l$ -th layer, where $D$ is the dimensionality. The self-attention is formulated as: + +$$ +\hat {X} _ {k, l} = \operatorname {S e l f - A t t e n} \left(X _ {k, l}\right) = \operatorname {S o f t m a x} \left(\frac {X _ {k , l} X _ {k , l} ^ {T}}{\sqrt {D _ {\text {k e y}}}}\right) X _ {k, l} \tag {1} +$$ + +Notably, we do not apply any attention masks for specific constraints. Our experiments show that pretraining on a modest amount of data is sufficient to effectively adapt the pretrained ViT to the proposed ReCompact, enabling it to aggregate essential information within the visual registers. After visual encoding, The output features corresponding to the visual registers, denoted as $F = \{f_{1},\dots ,f_{M}\}$ , are then used as the visual representation of the image and fed into the LLM for various vision-language tasks. + +To further adapt the ReCompact mechanism for cropping-based high-resolution image perception, we share + +the same set of visual registers as input for different sub-images. The input for the $k$ -th sub-image is formulated as $V_{k} = \mathrm{Concat}(I_{k},R)$ , where $k = 1,2,\dots ,N_{c}$ and $N_{c}$ denotes the total number of sub-images. After encoding, the outputs of the visual registers for all sub-images are concatenated, yielding $F_{hr} = \mathrm{Concat}(F_1,F_2,\dots ,F_{N_c})$ which serves as the input to the LLM. The ReCompact mechanism enables the vision encoder to represent rich visual information using only a minimal number of $M$ tokens per sub-image, which significantly reduces the visual redundancy and eliminates the need for additional compression. + +# 3.2. Ensuring Visual Continuity via ReAtten + +When using the cropping-based high-resolution image encoding method, each sub-image is independently encoded by the vision encoder. This process can disrupt semantic coherence and result in the misrepresentation of visual concepts. To preserve the continuity of visual semantics, it is crucial to enable information exchange between sub-images during encoding. A naive approach would be to directly concatenate the image features from all sub-images and apply self-attention. However, the quadratic computational complexity makes this approach impractical. In Sec. 3.1, we introduced the ReCompact mechanism, where the visual registers effectively aggregate rich visual information while discarding redundancy during visual encoding, allowing them to serve as compact yet comprehensive image representations. Consequently, interactions between visual registers can naturally facilitate sufficient information exchange across different sub-images in an efficient manner. + +To enable information exchange between visual registers across different sub-images, we introduce a novel Register Interactive Attention (ReAtten) module. As shown in Fig. 2, the ReAtten module is integrated into each of the $L$ layers of the ViT. After the self-attention process in Eq. (1), we acquire $\hat{X}_{k,l} = \{\hat{X}_{k,l}^{I},\hat{X}_{k,l}^{R}\}$ , where $\hat{X}^I\in \mathcal{R}^{N\times D}$ and $\hat{X}^R\in \mathcal{R}^{M\times D}$ represent the hidden states corresponding to image tokens and visual registers, respectively. The states of the visual registers from different sub-images are concatenated to form $\hat{X}_l^R = \mathrm{Concat}\left(\{\hat{X}_{k,l}^R\}_{k = 1}^{N_c}\right)\in$ $\mathcal{R}^{M\cdot N_c\times D}$ . Then these concatenated registers are processed via ReAtten to enable information exchange across subimages, which can be formulated as: + +$$ +\overline {{X}} _ {l} ^ {R} = \mathrm {R e A t t e n} (\hat {X} _ {l} ^ {R}) = \hat {X} _ {l} ^ {R} + \mathrm {C r o s s - V i T - A t t e n} (\hat {X} _ {l} ^ {R}) (2) +$$ + +where Cross-ViT-Atten is implemented as a self-attention layer. Once the information exchange is completed, the registers and image tokens from a same sub-image are concatenated and then passed into the feedforward network (FFN) for transformation, yielding the output of the $l$ -th layer: + +$$ +X _ {k, l + 1} = \operatorname {F F N} \left(\operatorname {C o n c a t} \left(\hat {X} _ {k, l} ^ {I}, \bar {X} _ {k, l} ^ {R}\right)\right) \tag {3} +$$ + +![](images/7fea22fdc3cc3b493e312b32f4a98ebdb9f316bfc9d08bbff01a1afa4a45a549.jpg) +Figure 3. The progressive training pipeline. **Stage 0:** Align the frozen vision and language backbones by training only the MLP and registers. **Stage 1:** Tune all parameters with caption and OCR data on low-resolution for coarse alignment. **Stage 2:** Integrate high-resolution input along with fine-grained captions and OCR to achieve refined alignment. **Stage 3:** Froze ViT and introduce ReAtten module for high-resolution instruction tuning. + +By integrating ReAtten into each layer of the ViT model, FALCON enables effective information exchange between cropped sub-images through visual registers, thereby ensuring semantic continuity of the high-resolution image throughout the visual encoding process. + +# 3.3. Learning Expressive Visual Representation via Progressive Training Pipeline + +To empower the ReCompact mechanism with the capability to capture rich visual patterns and to enable smoother initialization of the ReAtten module under high-resolution conditions, we develop a specialized progressive training pipeline. As illustrated in Fig. 3, this pipeline progressively transitions from low to high resolution across four stages. + +Stage 0: Static Alignment on Low-resolution. Since both the visual registers and MLP projector are randomly initialized, in stage 0, we aim to achieve a initial alignment between the pretrained vision encoder and the LLM. To facilitate this, we begin by using low-resolution images as inputs, which are directly resized to match the resolution of the vision encoder. We then statically pretrain the model by keeping both the vision encoder and LLM frozen, tuning only the visual registers and MLP using image caption data. + +Stage 1: Coarse Alignment on Low-resolution. To better adapt the pre-trained vision encoder to our newly introduced ReCompact mechanism, in stage 1, we unfreeze all model parameters to enable a more thorough tuning. To facilitate the visual registers to progressively learn capturing visual information from coarse to fine granularity, we con + +tinue to use low-resolution inputs and employ high-quality, extended captions rich in visual content, along with full-text OCR data for training. This design allows the registers to acquire holistic scene understanding ability. + +Stage 2: Refined Alignment on High-resolution. To further enhance the capacity of the visual register to capture fine-grained visual details, in stage 2, we transition the model to process high-resolution image for refined alignment. Additionally, we employ datasets that includes a series of fine-grained captions and OCR tasks, such as regional captions and text location. This facilitates the visual registers to capture comprehensive fine-grained details. + +Stage 3: Instruction Tuning on High-resolution. To adapt the well-aligned visual register for a wide range of real-world tasks, in stage 3, we perform instruction tuning with a mixture of datasets covering various tasks, such as VQA, document understanding, and science reasoning. Additionally, the proposed ReAtten module is integrated to address visual fragmentation caused by cropping-based high-resolution image encoding. Since the vision encoder with ReCompact is well-aligned, the vision encoder is kept frozen and other parameters including ReAtten are trainable during instruction tuning. Notably, to facilitate a smoother initialization for the newly introduced ReAtten modules, we initialize parameters of each ReAtten module with the ViT self-attention parameters at the corresponding depth. This design enables ReAtten to effectively leverage pre-trained parameters, avoiding the need for extensive fine-tuning. + +# 4. Experiments + +We implement FALCON with SigLIP-L/16-384px [62] as the vision encoder and Llama-3.1-8B-Instruct [12] as the LLM backbone. A two-layer MLP module with GeLU [13] activation is served as the vision-language projection layer. The number of visual registers is set to 64, achieving remarkable visual token compression rates of $9 \times$ . The maximum number of sub-images is set to 16. + +# 4.1. Experimental Results + +We evaluate FALCON using the recently proposed MMERealWorld [66] benchmark, a comprehensive and largescale benchmark specifically designed to assess the high-resolution understanding capabilities of MLLMs. MMERealWorld comprises 13,366 high-resolution images with an average resolution of $2,000 \times 1,500$ pixels, spanning five domains: OCR in the wild, Remote Sensing, Diagrams and Tables, Monitoring, and Autonomous Driving. This benchmark is well-suited for evaluating the practical high-resolution understanding capabilities of MLLMs. + +As shown in Tab. 1, FALCON significantly outperform other advanced MLLMs in both perception and reasoning tasks. FALCON achieves state-of-the-art performance with only 64 tokens per sub-image, resulting in a remarkable vi + +Table 1. Evaluation on MME-RealWorld Benchmark. "OCR", "RS", "DT", "MO", and "AD" stand for OCR in the Wild, Remote Sensing, Diagram and Table, Monitoring, and Autonomous Driving, respectively. "Avg" denotes the weighted average accuracy across all samples, while "Avg-C" denotes the weighted average accuracy across different domains. "Sizev" indicates the image resolution supported by the model, where $\times n$ signifies a cropping-based high-resolution processing approach with up to $n$ sub-images. "Tokenc" denotes the number of visual tokens after compression. † For models utilizing hybrid vision encoders, "Sizev" represents the setting of encoder with the highest resolution. We gray out the models that are either closed-source or have at least twice the number of visual tokens compared to ours. Among these, the lighter gray denotes scores below ours. The underline indicates that the second performance is achieved. Notably, FALCON achieves state-of-the-art performance with just 64 tokens per sub-image, surpassing most models that have large amount of visual tokens and many advanced closed-source commercial models. + +
TypeMethod\( Size^v \)\( Token^c \)PerceptionReasoning
Task SplitOCRRSDTMOADAvgAvg-COCRDTMOADAvgAvg-C
Visual Tokens\(\nabla 1024\)MiniGPT-v2 [4]\( 448^2 \)25639.023.320.419.326.026.925.630.020.416.923.723.022.7
Qwen-VL-Chat [1]\( 448^2 \)25632.415.115.622.115.120.820.128.613.616.524.622.020.8
LLaVA1.5-13B [30]\( 336^2 \)57644.123.320.220.526.128.426.830.220.827.524.825.525.8
ShareGPT4V-13B [6]\( 336^2 \)57644.623.120.219.326.128.426.626.020.827.324.624.624.7
TextMonkey [33]\( 448^2 \times 4 \)256 × 437.311.75.916.114.318.217.130.42.24.420.016.014.3
Mini-Gemini-7B-HD† [26]\( 336^2 \times 4 \)57642.031.322.334.224.831.130.935.424.625.923.326.127.3
DeepSeek-VL† [36]\( 1024^2 \)57649.625.523.427.033.433.131.845.223.816.727.328.028.3
MiniCPM-V 2.5 [56]\( 448^2 \times 9 \)96 × 966.827.752.838.734.247.444.044.031.837.031.034.536.0
Monkey [28]\( 448^2 \times 4 \)256 × 454.625.032.528.029.736.334.027.220.827.333.028.827.1
FALCON\( 384^2 \times 16 \)64 × 1666.449.647.240.436.550.348.055.439.244.434.540.743.4
Visual Tokens\(\nabla 2048\)LLaVA-Next-Qwen72B [31]\( 336^2 \times 4 \)576 × 437.129.127.729.418.029.028.317.234.227.329.727.927.1
LLaVA-Next-liama3 [31]\( 336^2 \times 4 \)576 × 447.925.426.619.518.730.127.655.223.421.130.732.132.6
mPLUG-DocOwl 1.5 [15]\( 448^2 \times 9 \)256 × 951.223.729.325.028.333.731.542.619.820.526.026.927.2
CogVLM2-liama3-Chat [14]\( 1344^2 \)230470.028.847.533.730.245.842.054.032.841.231.237.339.8
IXC2.5 [63]\( 560^2 \times 24 \)400 × 2469.336.163.939.533.652.548.553.441.017.730.033.935.5
LLaVA-OneVision [20]\( 384^2 \times 9 \)729 × 978.753.560.740.345.859.655.861.840.040.834.141.244.2
Closed-sourceGPT-4o-mini--62.56.744.226.524.237.132.847.039.825.826.832.534.9
Gemini-1.5-pro [50]--67.614.039.931.126.639.635.952.733.228.319.229.233.4
GPT-4o--77.728.946.733.922.446.441.961.444.836.526.437.642.3
Claude 3.5 Sonnet--72.525.767.432.240.852.947.761.961.241.831.944.149.2
+ +sual compression rate of 9 times. In addition, FALCON also outperforms most models that have twice as many visual tokens. Specifically, FALCON outperforms CogVLM2-llama3-Chat [14] in average scores on both perception and reasoning task, which uses 2304 tokens per-image. It also surpasses IXC2.5 [63] in average scores on the reasoning tasks, which has a maximum token counts of 9600. Notably, FALCON achieves higher average scores in both perception and reasoning tasks compared to three advanced commercial models, GPT-4o, GPT-4o-mini and Gemini1.5-pro [50], with it notably outperforming all commercial models in the sub-tasks of remote sensing and monitoring. These results demonstrate that the ReCompact and ReAtten technique proposed by FALCON effectively eliminate visual redundancy and ensure visual continuity, allowing comprehensive and detailed visual information to be conveyed with a highly compact visual representation. + +To further assess FALCON's capabilities in high-resolution understanding, we conduct evaluations on more high-resolution benchmarks: V* [53], DocVQA [42], TextVQA [49] and ChartQA [41]. V* focuses on evaluating + +models' capabilities for fine-grained attribute recognition and spatial reasoning. DocVQA challenges models to comprehend text-dense document images. ChartQA evaluates models' proficiency in interpreting various types of chart, while TextVQA assesses the ability to recognize small text in real-world scenes. As shown in Tab. 2, FALCON demonstrates superior performance across all these benchmarks. + +To evaluate the performance of FALCON across a broader range of domains, we further conduct experiments with diverse popular benchmarks. As shown in Tab. 4, POPE [25], ScienceQA [37], MMBench [34] and SEED-Bench [19] are selected to assess models' capabilities in hallucination, scientific reasoning, and comprehensive ability. FALCON achieves SOTA performance across all these benchmarks. The results demonstrate that FALCON exhibits superior capabilities across a wide range of tasks. + +# 4.2. Ablation Study + +We perform extensive ablation studies to validate the effectiveness of the proposed ReCompact and ReAtten. + +Ablation on Training Data. To eliminate the influx + +Table 2. Evaluations on diverse high-resolution Benchmarks. "V* ${}_{Avg}$ " represents the weighted average score for the two subtasks in the ${V}^{ * }$ benchmark. † denotes OCR inputs are utilized. + +
ModelV*AvgDocVQATextVQAChartQA
Low-resolution MLLMs
Qwen-VL [1]-65.163.859.0
InstructBLIP [9]34.04.550.710.9
LLaVA-v1.5-7B [30]48.78.558.2†17.8
mPLUG-Owl2 [60]-17.958.2†22.8
High-resolution MLLMs
UReader [58]-65.457.659.3
Monkey [28]-66.565.859.0
LlvaF-FlexAttn [23]54.4-48.9-
S2-Wrapper-7B [48]55.5-61.0-
LlvaNxt-Vicuna7B [31]61.8-64.954.3
SliME-7B [67]--64.433.9
FALCON64.474.466.568.8
+ +Table 3. Ablation results on training data. We train FALCON with two-stage pipeline and exactly the same dataset as LLaVAv1.5 [30]. The results confirm the superiority of the proposed method when the influence of training data is excluded. + +![](images/033cda0170d90fa03b901d164de4d812f479bd3f65f7a6d76335bf8e0d8b24d8.jpg) +Figure 4. Comparison of different compression methods. We report the sum of the "Avg-C" scores for the perception and reasoning tasks in MME-RealWorld. The superior performance demonstrates that ReCompact effectively eliminates redundancy and produces a more expressive and compact visual representation. + +ence of training data and demonstrate the advantages of the proposed method, we utilize exactly the same dataset as LLaVA-v1.5 [30] and employ a two-stage pipeline for training. As shown in Tab. 3, the proposed method exhibits significant performance advantages across multiple benchmarks when trained on the same dataset. + +Effectiveness of ReCompact mechanism. To verify the advantages of the proposed ReCompact mechanism in reducing visual redundancy and compressing visual representations during the visual encoding, we compare the Re-Compact with other commonly used compression methods. This includes pooling [48, 55], Pixel Shuffle [7] and Abstractor [60], where the Abstractor serves as a representative module that combines query tokens with cross-attention. + +As shown in Fig. 4, ReCompact demonstrates superior performance compared to other compression methods. + +Table 4. Evaluation on benchmarks across diverse domains. The POPE, SQA, MMBench and SEED-Bench are used to assess hallucination, scientific reasoning, and overall abilities, respectively. + +
ModelSQAPOPEMMBSEED-Img
Low-Resolution MLLMs
Qwen-VL [1]67.1-41.662.3
InstructBLIP [9]63.178.938.458.8
LlaVA-v1.5-7B [30]66.885.962.865.8
mPLUG-Owl2 [60]68.785.863.564.5
High-Resolution MLLMs
Monkey [28]69.483.757.464.3
LlaVA-FlexAttn [23]-85.965.7-
S2-Wrapper-7B [48]--66.267.9
LLaVA-Next-Vicuna7B [31]70.186.566.569.6
SliME-7B [67]76.885.463.167.0
FALCON88.888.076.173.3
+ +Table 5. Comparison of different methods for maintaining visual continuity. The "Baseline" model refers to a condition where no specific methods are applied to preserve visual continuity. We report the "Avg-C" score for MME-RealWorld. The results show that the ReAtten module more effectively addresses the issue of visual fragmentation compared to other methods, thereby improving overall performance and reducing hallucinations. + +
MethodV*AVgMME-RealWorldPOPE
PerceptionReasoning
Baseline51.338.235.685.7
CIP [16]50.338.437.286.3
W-Atten [33]60.241.038.186.4
ReAtten61.342.139.087.3
+ +These results indicate that ReCompact effectively eliminates visual redundancy and produces a more expressive compact visual representation. In contrast, we argue that the pooling layer and Pixel Shuffle merely reduces the number of tokens by merging adjacent ones, failing to eliminate redundancy that interferes with the LLM's understanding of visual content. Additionally, although Abstractor uses more parameters, its performance is generally lower than that of pooling and ReCompact. This result suggests that even millions of samples may be insufficient for fully training Abstractor, which diminishes their practical efficacy. + +Effectiveness of ReAtten Module. To validate the advantage of the proposed ReAtten Module, we compare ReAtten with other contemporary methods, including the Shifted Window Attention (W-Attn) [35] adopted by TextMonkey [33] and the Complementary Image Pyramid (CIP) introduced by MiniMonkey [16]. + +As shown in Tab. 5, the model incorporating ReAtten achieves significant improvements over the baseline across all benchmarks and demonstrates superiority compared to other methods. These results indicate that the ReAtten module effectively enhances the ability of the model in high-resolution perception and mitigates hallucinations by addressing the issue of visual fragmentation. In contrast, we + +![](images/69a61bdb051446db242affaf1ca3d6af18de2936a64fc459257188dc0fac5037.jpg) +(a) Input Image + +![](images/8ceb0250439bd1470cb952aa6838c3da9139588af5237a80a9f74e4373251c0c.jpg) + +![](images/bd66f1083bbf9ca89b53d071f4a99d6ba15b6ea2d41b3298cad141cc222eef11.jpg) +(b) Register-to-1 + +![](images/5059284c6bb109e87b39beafc78b52016928dfabae1f0d576551336ded1848a6.jpg) + +![](images/95f15239a632e544a78a17d734546edff477e66bf722bd4973c44a87b7dcafed.jpg) +Image Attention + +![](images/d67645d774671878f7e1abc1423fa5dc384083272d41db0887c544e8031ffb02.jpg) + +![](images/1253deae450cc82ae011ce0b29a5cf61ef43fe4fbaff73bfe804d96497da7381.jpg) +(c) Ablation on ReAtten + +w/ ReAtten + +w/o ReAtten + +![](images/813b26bb016597ba79d2f958654b8d68b721c43981cec7fa7c93c4039fc1a5c1.jpg) +Figure 5. Visualization of the register-to-image attention map. We randomly select visual registers and extract the attention scores that each register assigns to image tokens. These scores are then visualized as heatmaps overlaid on the input image. The results in (b) show that each visual register focuses on specific parts of the image, capturing rich visual patterns. Meanwhile, registers pay minimal attention to background areas, effectively avoiding the inclusion of redundancy. We also conduct qualitative ablation studies of the ReAtten module. As shown in (c), in the ViT model without ReAtten, the attention patterns across different sub-images appear extremely fragmented. In contrast, the ViT model with ReAtten shows continuous attention patterns, indicating effective information interaction between sub-images. +Figure 6. Ablation on different visual register counts. Increasing the number of visual registers generally enhances accuracy across both benchmarks. However, the improvement rate noticeably tapers off when the register count exceeds 64. Based on this observation, to better balance the efficiency and performance, we select register counts of 64 for FALCON. + +argue that the information exchange in W-Atten is limited to a local window, which hinders sufficient information exchange among all sub-images. Meanwhile, CIP introduces further redundancy by repeatedly encoding the same visual content, which can even lead to a decline in performance. + +Ablation on the Number of Visual Registers. To investigate how visual register count affects model performance, we conduct experiments with visual register counts set to 36, 64, and 144. These counts correspond to visual token compression rates of $16 \times$ , $9 \times$ , and $4 \times$ , respectively. As shown in Fig. 6, increasing the number of visual registers generally enhances model's high-resolution comprehension capabilities and reduces hallucinations. However, when the register count exceeds 64, the improvement gains start tapering off noticeably. Meanwhile, increasing visual registers demands more computational resources. Compared to the baseline model with 36 registers, models configured with 64 and 144 registers require 1.15 and 2.41 times more training time, respectively. Base on these observations, to achieve a better balance between efficiency and performance, we select 64 as the register counts for FALCON. + +# 4.3. Qualitative Analysis + +To qualitatively analyze the proposed visual register technique, we visualize the register-to-image attention maps within ViT. Specifically, for a given visual register, we extract the attention scores it assigns to image tokens in each sub-image and concatenate them to construct a complete heatmap. As shown in Fig. 5(b), each visual register focuses on relatively distinct areas of the image, highlighting that it captures a diverse range of visual details. Moreover, the visual registers tend to avoid background areas, indicating their effectiveness in identifying and discarding visual redundancy. These properties enable visual registers to effectively produce a comprehensive and non-redundant visual representation. Additionally, to qualitatively evaluate the effectiveness of the ReAtten module, we visualize the attention maps within ViT models with and without ReAtten. As shown in Fig. 5(c), in the ViT model without ReAtten, the attention patterns across different sub-images appear highly fragmented. In contrast, the ViT model with ReAtten exhibits continuous attention patterns, indicating effective information exchange between sub-images. + +# 5. Conclusion + +To address visual redundancy and fragmentation in high-resolution MLLMs, we propose FALCON. FALCON employs an innovative visual register technique that simultaneously addresses both challenges. This technique uses a Re-Compact mechanism to adaptively aggregate essential visual information through visual registers, creating a compact, non-redundant representation. Additionally, a ReAtten module is introduced to further enhances visual continuity by enabling information exchange among sub-images via visual registers. Extensive experiments demonstrate FALCON's superiority in high-resolution understanding. + +# 6. Acknowledgement + +This study is supported by National Natural Science Foundation of China (Grant No. 62306090, No. U24A20328), Natural Science Foundation of Guangdong Province of China (Grant No. 2024A1515010147), Shenzhen Science and Technology Program (KQTD20240729102207002), Natural Science Foundation of Beijing, China (Grant No. L254018), Jiangsu Science and Technology Major Program (Grant No. BG2024041) and China Postdoctoral Science Foundation (Grant No. 2024M764192). + +# References + +[1] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 1(2):3, 2023. 2, 3, 6, 7 +[2] Gongwei Chen, Leyang Shen, Rui Shao, Xiang Deng, and Liqiang Nie. Lion: Empowering multimodal large language model with dual-level visual knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26540-26550, 2024. 1, 2 +[3] Gongwei Chen, Xurui Zhou, Rui Shao, Yibo Lyu, Kaiwen Zhou, Wentao Li, Yinchuan Li, Zhongang Qi, and Liqiang Nie. Less is more: Empowering gui agent with context-aware simplification. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2025. 3 +[4] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 6 +[5] Jingxuan Chen, Derek Yuen, Bin Xie, Yuhao Yang, Gongwei Chen, Zhihao Wu, Li Yixing, Xurui Zhou, Weiwen Liu, Shuai Wang, Kaiwen Zhou, Rui Shao, Liqiang Nie, Yasheng Wang, Jianye HAO, Jun Wang, and Kun Shao. Spa-bench: A comprehensive benchmark for smartphone agent evaluation. In The Thirteenth International Conference on Learning Representations, 2025. 3 +[6] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. In European Conference on Computer Vision, pages 370-387. Springer, 2024. 6 +[7] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internv1: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24185–24198, 2024. 7 +[8] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source + +chatbot impressing gpt-4 with $90\%$ chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2(3):6, 2023. 3 +[9] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1, 2, 3, 7 +[10] Timothee Darcet, Maxime Aquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. In The Twelfth International Conference on Learning Representations, 2024. 2, 4 +[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 3 +[12] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024.5 +[13] Dan Hendrycks and Kevin Gimpel. Bridging nonlinearities and stochastic regularizers with gaussian error linear units. arXiv preprint arXiv:1606.08415, 2016. 5 +[14] Wenyi Hong, Weihan Wang, Ming Ding, Wenmeng Yu, Qingsong Lv, Yan Wang, Yean Cheng, Shiyu Huang, Junhui Ji, Zhao Xue, et al. Cogvlm2: Visual language models for image and video understanding. arXiv preprint arXiv:2408.16500, 2024. 6 +[15] Anwen Hu, Haiyang Xu, Jiabo Ye, Ming Yan, Liang Zhang, Bo Zhang, Ji Zhang, Qin Jin, Fei Huang, and Jingren Zhou. mplug-docowl 1.5: Unified structure learning forOCR-free document understanding. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 3096-3120, 2024. 1, 3, 6 +[16] Mingxin Huang, Yuliang Liu, Dingkang Liang, Lianwen Jin, and Xiang Bai. Mini-monkey: Alleviating the semantic saw-tooth effect for lightweight MLLMs via complementary image pyramid. In The Thirteenth International Conference on Learning Representations, 2025. 3, 7 +[17] Runhui Huang, Xinpeng Ding, Chunwei Wang, Jianhua Han, Yulong Liu, Hengshuang Zhao, Hang Xu, Lu Hou, Wei Zhang, and Xiaodan Liang. Hires-llava: Restoring fragmentation input in high-resolution large vision-language models. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 29814–29824, 2025. 1, 3 +[18] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan P Foster, Pannag R Sanketi, Quan Vuong, et al. Openvla: An open-source vision-language-action model. In 8th Annual Conference on Robot Learning, 2024. 3 +[19] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking mul + +timodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 6 +[20] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. LLaVA-onevision: Easy visual task transfer. Transactions on Machine Learning Research, 2025. 6 +[21] Hao Li, Qi Lv, Rui Shao, Xiang Deng, Yinchuan Li, Jianye HAO, and Liqiang Nie. STAR: Learning diverse robot skill abstractions through rotation-augmented vector quantization. In Forty-second International Conference on Machine Learning, 2025. 3 +[22] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 2 +[23] Junyan Li, Delin Chen, Tianle Cai, Peihao Chen, Yining Hong, Zhenfang Chen, Yikang Shen, and Chuang Gan. Flex-attention for efficient high-resolution vision-language models. In European Conference on Computer Vision, pages 286–302, 2024. 7 +[24] Wei Li, Bing Hu, Rui Shao, Leyang Shen, and Liqiang Nie. Lion-fs: Fast & slow video-language thinker as online video assistant. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 3240-3251, 2025. 3 +[25] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In The 2023 Conference on Empirical Methods in Natural Language Processing, 2023. 6 +[26] Yanwei Li, Yuechen Zhang, Chengyao Wang, Zhisheng Zhong, Yixin Chen, Ruihang Chu, Shaoteng Liu, and Jiaya Jia. Mini-gemini: Mining the potential of multi-modality vision language models. arXiv preprint arXiv:2403.18814, 2024. 6 +[27] Zaijing Li, Yuquan Xie, Rui Shao, Gongwei Chen, Dongmei Jiang, and Liqiang Nie. Optimus-1: Hybrid multimodal memory empowered agents excel in long-horizon tasks. In Advances in neural information processing systems, 2024. 3 +[28] Zhang Li, Biao Yang, Qiang Liu, Zhiyin Ma, Shuo Zhang, Jingxu Yang, Yabo Sun, Yuliang Liu, and Xiang Bai. Monkey: Image resolution and text label are important things for large multi-modal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26763-26773, 2024. 2, 3, 6, 7 +[29] Zaijing Li, Yuquan Xie, Rui Shao, Gongwei Chen, Dongmei Jiang, and Liqiang Nie. Optimus-2: Multimodal minercraft agent with goal-observation-action conditioned policy. In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2025. 3 +[30] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 1, 2, 3, 6, 7 +[31] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,ocr,and world knowledge,2024.6,7 + +[32] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 3 +[33] Yuliang Liu, Biao Yang, Qiang Liu, Zhang Li, Zhiyin Ma, Shuo Zhang, and Xiang Bai. Textmonkey: AnOCR-free large multimodal model for understanding document. arXiv preprint arXiv:2403.04473, 2024. 1, 2, 3, 6, 7 +[34] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European Conference on Computer Vision, pages 216-233. Springer, 2025. 6 +[35] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 3, 7 +[36] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Yaofeng Sun, et al. Deepseek-vl: towards real-world vision-language understanding. arXiv preprint arXiv:2403.05525, 2024. 3, 6 +[37] Pan Lu, Swaroop Mishra, Tony Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In The 36th Conference on Neural Information Processing Systems (NeurIPS), 2022. 6 +[38] Gen Luo, Yiyi Zhou, Yuxin Zhang, Xiawu Zheng, Xiaoshuai Sun, and Rongrong Ji. Feast your eyes: Mixture-of-resolution adaptation for multimodal large language models. arXiv preprint arXiv:2403.03003, 2024. 3 +[39] Qi Lv, Hao Li, Xiang Deng, Rui Shao, Yinchuan Li, Jianye Hao, Longxiang Gao, Michael Yu Wang, and Liqiang Nie. Spatial-temporal graph diffusion policy with kinematic modeling for bimanual robotic manipulation. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 17394-17404, 2025. 3 +[40] Yibo Lyu, Rui Shao, Gongwei Chen, Yijie Zhu, Weili Guan, and Liqiang Nie. Puma: Layer-pruned language model for efficient unified multimodal retrieval with modality-adaptive learning. In Proceedings of the 33nd ACM International Conference on Multimedia, 2025. 3 +[41] Ahmed Masry, Xuan Long Do, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. In Findings of the Association for Computational Linguistics: ACL 2022, pages 2263-2279, 2022. 6 +[42] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 6 +[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 4 + +[44] Rui Shao, Xiangyuan Lan, Jiawei Li, and Pong C Yuen. Multi-adversarial discriminative deep domain generalization for face presentation attack detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10023-10031, 2019. 3 +[45] Rui Shao, Tianxing Wu, and Ziwei Liu. Detecting and grounding multi-modal media manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6904-6913, 2023. 1 +[46] Rui Shao, Tianxing Wu, Jianlong Wu, Liqiang Nie, and Zwei Liu. Detecting and grounding multi-modal media manipulation and beyond. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1, 3 +[47] Leyang Shen, Gongwei Chen, Rui Shao, Weili Guan, and Liqiang Nie. Mone: Mixture of multimodal experts for generalist multimodal large language models. In Advances in neural information processing systems, 2024. 2 +[48] Baifeng Shi, Ziyang Wu, Maolin Mao, Xin Wang, and Trevor Darrell. When do we not need larger vision models? In European Conference on Computer Vision, pages 444-462, 2024. 2, 3, 7 +[49] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 6 +[50] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 6 +[51] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3 +[52] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3 +[53] Penghao Wu and Saining Xie. V*: Guided visual search as a core mechanism in multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13084-13094, 2024. 6 +[54] Bin Xie, Rui Shao, Gongwei Chen, Kaiwen Zhou, Yinchuan Li, Jie Liu, Min Zhang, and Liqiang Nie. Gui-explorer: Autonomous exploration and mining of transition-aware knowledge for gui agent. In Annual Meeting of the Association for Computational Linguistics (ACL), 2025. 3 +[55] Linli Yao, Lei Li, Shuhuai Ren, Lean Wang, Yuanxin Liu, Xu Sun, and Lu Hou. Deco: Decoupling token compression from semantic abstraction in multimodal large language models. arXiv preprint arXiv:2405.20985, 2024. 2, 3, 7 +[56] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 6 + +[57] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Yuhao Dan, Chenlin Zhao, Guohai Xu, Chenliang Li, Junfeng Tian, et al. mplug-docowl: Modularized multimodal large language model for document understanding. arXiv preprint arXiv:2307.02499, 2023. 3 +[58] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Guohai Xu, Chenliang Li, Junfeng Tian, Qi Qian, Ji Zhang, et al. Ureader: UniversalOCR-free visually-situated language understanding with multimodal large language model. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 2841-2858, 2023. 1, 2, 3, 7 +[59] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 1, 2 +[60] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Anwen Hu, Haowei Liu, Qi Qian, Ji Zhang, and Fei Huang. mplug-owl2: Revolutionizing multi-modal large language model with modality collaboration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13040–13051, 2024. 1, 2, 3, 7 +[61] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In European Conference on Computer Vision (ECCV), pages 69-85. Springer, 2016. 1 +[62] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 3, 5 +[63] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, et al. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output. arXiv preprint arXiv:2407.03320, 2024. 6 +[64] Renshan Zhang, Yibo Lyu, Rui Shao, Gongwei Chen, Weili Guan, and Liqiang Nie. Token-level correlation-guided compression for efficient multimodal document understanding. arXiv preprint arXiv:2407.14439, 2024. 3 +[65] Xiaofeng Zhang, Yihao Quan, Chen Shen, Xiaosong Yuan, Shaotian Yan, Liang Xie, Wenxiao Wang, Chaochen Gu, Hao Tang, and Jieping Ye. From redundancy to relevance: Enhancing explainability in multimodal large language models. Annual Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics, 2025. 2 +[66] YiFan Zhang, Huanyu Zhang, Haochen Tian, Chaoyou Fu, Shuangqing Zhang, Junfei Wu, Feng Li, Kun Wang, Qing-song Wen, Zhang Zhang, Liang Wang, and Rong Jin. MMErealworld: Could your multimodal LLM challenge high-resolution real-world scenarios that are difficult for humans? In The Thirteenth International Conference on Learning Representations, 2025. 5 +[67] Yi-Fan Zhang, Qingsong Wen, Chaoyou Fu, Xue Wang, Zhang Zhang, Liang Wang, and Rong Jin. Beyond llava-hd: Diving into high-resolution large multimodal models. arXiv preprint arXiv:2406.08487, 2024. 1, 3, 7 \ No newline at end of file diff --git a/2025/FALCON_ Resolving Visual Redundancy and Fragmentation in High-resolution Multimodal Large Language Models via Visual Registers/images.zip b/2025/FALCON_ Resolving Visual Redundancy and Fragmentation in High-resolution Multimodal Large Language Models via Visual Registers/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..11a363228b52591fe3043cd7912d64ac1aad6825 --- /dev/null +++ b/2025/FALCON_ Resolving Visual Redundancy and Fragmentation in High-resolution Multimodal Large Language Models via Visual Registers/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a8d875546674402cb0adc83d0ee36eaa52d295521332afa97f5fd62a23f2683f +size 681680 diff --git a/2025/FALCON_ Resolving Visual Redundancy and Fragmentation in High-resolution Multimodal Large Language Models via Visual Registers/layout.json b/2025/FALCON_ Resolving Visual Redundancy and Fragmentation in High-resolution Multimodal Large Language Models via Visual Registers/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0a2fac5d7ce0a79f06c8a6fcc103e08ee3af2229 --- /dev/null +++ b/2025/FALCON_ Resolving Visual Redundancy and Fragmentation in High-resolution Multimodal Large Language Models via Visual Registers/layout.json @@ -0,0 +1,7680 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 60, + 102, + 550, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 102, + 550, + 140 + ], + "spans": [ + { + "bbox": [ + 60, + 102, + 550, + 140 + ], + "type": "text", + "content": "FALCON: Resolving Visual Redundancy and Fragmentation in High-resolution Multimodal Large Language Models via Visual Registers" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 135, + 165, + 475, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 165, + 475, + 208 + ], + "spans": [ + { + "bbox": [ + 135, + 165, + 475, + 208 + ], + "type": "text", + "content": "Renshan Zhang" + }, + { + "bbox": [ + 135, + 165, + 475, + 208 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 135, + 165, + 475, + 208 + ], + "type": "text", + "content": ", Rui Shao" + }, + { + "bbox": [ + 135, + 165, + 475, + 208 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 135, + 165, + 475, + 208 + ], + "type": "text", + "content": ", Gongwei Chen" + }, + { + "bbox": [ + 135, + 165, + 475, + 208 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 135, + 165, + 475, + 208 + ], + "type": "text", + "content": ", Miao Zhang" + }, + { + "bbox": [ + 135, + 165, + 475, + 208 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 135, + 165, + 475, + 208 + ], + "type": "text", + "content": ", Kaiwen Zhou" + }, + { + "bbox": [ + 135, + 165, + 475, + 208 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 135, + 165, + 475, + 208 + ], + "type": "text", + "content": ", Weili Guan" + }, + { + "bbox": [ + 135, + 165, + 475, + 208 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 135, + 165, + 475, + 208 + ], + "type": "text", + "content": ", Liqiang Nie" + }, + { + "bbox": [ + 135, + 165, + 475, + 208 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 135, + 165, + 475, + 208 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 135, + 165, + 475, + 208 + ], + "type": "text", + "content": "Harbin Institute of Technology, Shenzhen " + }, + { + "bbox": [ + 135, + 165, + 475, + 208 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 135, + 165, + 475, + 208 + ], + "type": "text", + "content": "Huawei Noah's Ark Lab" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 162, + 210, + 447, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 210, + 447, + 221 + ], + "spans": [ + { + "bbox": [ + 162, + 210, + 447, + 221 + ], + "type": "text", + "content": "zhangrenshan@stu.hit.edu.cn shaorui@hit.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 174, + 224, + 442, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 224, + 442, + 236 + ], + "spans": [ + { + "bbox": [ + 174, + 224, + 442, + 236 + ], + "type": "text", + "content": "chengongwei@hit.edu.cn nieliqiang@gmail.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 194, + 238, + 413, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 238, + 413, + 249 + ], + "spans": [ + { + "bbox": [ + 194, + 238, + 413, + 249 + ], + "type": "text", + "content": "https://github.com/JiuTian-VL/FALCON" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 152, + 277, + 200, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 277, + 200, + 290 + ], + "spans": [ + { + "bbox": [ + 152, + 277, + 200, + 290 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 302, + 297, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 302, + 297, + 649 + ], + "spans": [ + { + "bbox": [ + 54, + 302, + 297, + 649 + ], + "type": "text", + "content": "The incorporation of high-resolution visual input equips multimodal large language models (MLLMs) with enhanced visual perception capabilities for real-world tasks. However, most existing high-resolution MLLMs rely on a cropping-based approach to process images, which leads to fragmented visual encoding and a sharp increase in redundant tokens. To tackle these issues, we propose the FALCON model. FALCON introduces a novel visual register technique to simultaneously: 1) Eliminate redundant tokens at the stage of visual encoding. To directly address the visual redundancy present in the output of vision encoder, we propose a Register-based Representation Compacting (ReCompact) mechanism. This mechanism introduces a set of learnable visual registers designed to adaptively aggregate essential information while discarding redundancy. It enables the encoder to produce a more compact visual representation with a minimal number of output tokens, thus eliminating the need for an additional compression module. 2) Ensure continuity in visual encoding. To address the potential encoding errors caused by fragmented visual inputs, we develop a Register Interactive Attention (ReAtten) module. This module facilitates effective and efficient information exchange across sub-images by enabling interactions between visual registers. It ensures the continuity of visual semantics throughout the encoding. We conduct comprehensive experiments with FALCON on high-resolution benchmarks across a wide range of scenarios. FALCON demonstrates superior performance with a remarkable 9-fold reduction in visual tokens." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 651, + 137, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 137, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 137, + 664 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 671, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 295, + 696 + ], + "type": "text", + "content": "Recently, multimodal large language models (MLLMs) [2, 9, 30, 32, 59, 60] have made significant advancements in" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 277, + 554, + 462 + ], + "blocks": [ + { + "bbox": [ + 317, + 277, + 554, + 462 + ], + "lines": [ + { + "bbox": [ + 317, + 277, + 554, + 462 + ], + "spans": [ + { + "bbox": [ + 317, + 277, + 554, + 462 + ], + "type": "image", + "image_path": "d729d5f337aba80dea70000a3b80a55cd315708e8941f483cec9704ffe13e941.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 470, + 555, + 514 + ], + "lines": [ + { + "bbox": [ + 313, + 470, + 555, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 555, + 514 + ], + "type": "text", + "content": "Figure 1. Comparison between existing high-resolution MLLMs and FALCON. FALCON eliminates visual redundancy and ensures the continuity of high-resolution encoding through the use of ReCompact and ReAtten mechanisms." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 520, + 555, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 520, + 555, + 687 + ], + "spans": [ + { + "bbox": [ + 312, + 520, + 555, + 687 + ], + "type": "text", + "content": "performing various vision-language tasks [45, 46, 61], facilitating more natural human-machine interactions. Early MLLMs [9, 32, 59] typically processed images at a fixed low resolution, such as " + }, + { + "bbox": [ + 312, + 520, + 555, + 687 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 312, + 520, + 555, + 687 + ], + "type": "text", + "content": ". This limited their capabilities in many real-world applications that require high-resolution input. To equip MLLMs with high-resolution visual perception, recent efforts [15, 17, 33, 58, 67] have introduced a cropping-based high-resolution processing method. Specifically, As shown in Fig. 1(b), this approach crops high-resolution images into multiple non-overlapping sub-images that match the pre-trained resolution of the encoder. Each sub-image is processed independently by the vision encoder. Then the resulting visual features are concatenated and fed into the large language model (LLM)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Although effective, cropping-based methods still face issues of: 1) Visual Redundancy. The number of visual to-" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 146, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 146, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 146, + 712 + ], + "type": "text", + "content": "†Corresponding authors" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23530" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 324 + ], + "type": "text", + "content": "kens grows dramatically with increasing resolution. While this enriches the visual details, it also leads to an increase in redundant tokens. These redundant tokens significantly increase the computational burden on LLM and interfere with the processing of visual information [65]. To address these issues, as illustrated in Fig. 1(b), existing methods typically rely on additional compression techniques. Representative approaches include combining learnable queries with cross-attention [1, 28, 60] or simply employing pooling layers [48, 55]. However, these methods either demand substantial training resources or fail to fully address redundancy. 2) Visual Fragmentation. Cropping an image into multiple non-overlapping sub-images may disrupt visual semantics and potentially lead to encoding errors. As shown in Fig. 1(a), in an OCR task, the word \"pineapple\" may be incorrectly segmented into \"pine\" and \"apple\", causing the LLM to misinterpret the word. Similarly, fragmentation can also hinder the identification of objects. For example, the \"rubin vase\" pattern illustrated in Fig. 1(a) depicts both a vase and two faces simultaneously. However, the fragmentation could lead to the vase being unrecognizable." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 331, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 295, + 713 + ], + "type": "text", + "content": "To address these issues, as shown in Fig. 1(c), we propose the FALCON model. FALCON introduces a novel visual register technique to simultaneously: 1) Eliminate redundant tokens during visual encoding. To eliminate redundancy directly at the stage of visual encoding, we propose a novel Register-based Representation Compacting (ReCompact) mechanism. This mechanism introduces a set of learnable visual registers as additional inputs to the vision encoder. The registers are designed to adaptively aggregate essential information from image tokens and discard redundant content. Then the outputs corresponding to the visual registers are considered as a non-redundant visual representation, serving as the compact input for the LLM instead of a large number of image tokens. This mechanism enables the vision encoder to convey comprehensive visual information in a compact representation using minimal tokens, thus removing the need for an additional compression module. 2) Ensure continuity in visual encoding. To avoid the potential encoding errors caused by fragmented visual inputs, we develop a novel Register Interactive Attention (ReAtten) module. This module leverages the compact and comprehensive representation produced by visual registers to facilitate information exchange across sub-images. It gathers visual registers from different sub-images and employs a specialized cross-ViT attention module to enable interactions among them. Such interaction enables thorough information exchange across the entire image, thus ensuring visual continuity throughout the encoding process. Additionally, the compactness of the visual registers further enhances the efficiency of these interactions. Consequently, this module provides LLMs with a reliable foundation to accurately interpret visual content and reduce hallucinations." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 313, + 72, + 553, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 406 + ], + "type": "text", + "content": "Note that the widely used Abstractor-like modules [28, 33, 58, 60], including QFormer [9], ReSampler [1], follow a similar approach by using query tokens to extract information from image tokens after visual encoding. The key difference is that the visual register mechanism directly adopts the pretrained ViT to aggregate information into a set of specified register tokens. This design is based on the observation that a well-pretrained ViT can naturally gather global information into some specific tokens through self-attention [10]. It offers the visual register mechanism several strengths in terms of: 1) Training Efficiency. The substantial number of randomly initialized parameters in Abstractor-like modules requires extensive data for pretraining. For instance, QFormer [22], Abstractor [60], and Resampler [1] use large corpora of 129M, 400M, and 1.4B samples, respectively. In contrast, by leveraging well-pretrained ViT parameters, the visual register mechanism requires only 3.5M samples for pretraining. The ablation results in Fig. 4 further demonstrate this advantages. 2) Addressing Visual Fragmentation. By leveraging the compact and comprehensive representation produced by visual registers, the proposed ReAtten module can effectively facilitate information exchange across sub-images. This is an important function that is notably lacking in Abstractor-like modules. Combined with ReCompact, it enables the visual register mechanism to simultaneously address the issues of visual redundancy and fragmentation. The ablation results in Tab. 5 validate the effectiveness of ReAtten." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 326, + 408, + 527, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 408, + 527, + 418 + ], + "spans": [ + { + "bbox": [ + 326, + 408, + 527, + 418 + ], + "type": "text", + "content": "We summarize our main contributions as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 421, + 553, + 635 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 314, + 421, + 553, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 421, + 553, + 504 + ], + "spans": [ + { + "bbox": [ + 314, + 421, + 553, + 504 + ], + "type": "text", + "content": "- To address the issue of visual redundancy directly at the encoding phase, we introduce the Register-based Representation Compacting (ReCompact) mechanism. This mechanism enables vision encoder to capture rich visual information using only a minimal number of tokens, effectively eliminating redundancy and obviating the need for additional compression modules." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 505, + 553, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 505, + 553, + 563 + ], + "spans": [ + { + "bbox": [ + 314, + 505, + 553, + 563 + ], + "type": "text", + "content": "- To address the issue of visual fragmentation caused by image cropping, we propose the register interactive attention (ReAtten) module. This module ensures the semantic continuity of high-resolution visual encoding in an efficient and effective manner." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 564, + 553, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 564, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 314, + 564, + 553, + 635 + ], + "type": "text", + "content": "- We introduce FALCON, a model that leverages the proposed visual register technique. Comprehensive evaluations on high-resolution visual understanding benchmarks highlight the superiority of FALCON. Notably, FALCON achieves exceptional results with a remarkable reduction in visual tokens by " + }, + { + "bbox": [ + 314, + 564, + 553, + 635 + ], + "type": "inline_equation", + "content": "9 \\times" + }, + { + "bbox": [ + 314, + 564, + 553, + 635 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 650, + 400, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 650, + 400, + 662 + ], + "spans": [ + { + "bbox": [ + 314, + 650, + 400, + 662 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 671, + 509, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 671, + 509, + 684 + ], + "spans": [ + { + "bbox": [ + 314, + 671, + 509, + 684 + ], + "type": "text", + "content": "2.1. Multimodal Large Language Models" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Research on MLLMs [1, 2, 30, 47, 59, 60] has made significant strides in recent years. By incorporating image in" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23531" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 553, + 203 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 553, + 203 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 553, + 203 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 553, + 203 + ], + "type": "image", + "image_path": "2f27ac45349ad0f7f21ad2a796584b3c4076140f766605856e347b5f07931124.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 211, + 555, + 257 + ], + "lines": [ + { + "bbox": [ + 54, + 211, + 555, + 257 + ], + "spans": [ + { + "bbox": [ + 54, + 211, + 555, + 257 + ], + "type": "text", + "content": "Figure 2. Overall framework of the proposed FALCON. The model introduces a set of learnable visual registers, which are fed into the ViT along with the image tokens from each sub-image. During the visual encoding, the ReCompact mechanism enables information from the image tokens to be aggregated into the visual registers, while the proposed ReAtten facilitates information exchange among the visual registers across different sub-images. Ultimately, these visual registers provide a compact visual representation as input to the LLM." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 262, + 296, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 262, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 56, + 262, + 296, + 468 + ], + "type": "text", + "content": "puts into pretrained LLMs [8, 51, 52], MLLMs demonstrate strong capabilities across various vision-language tasks. Typically, these models use a pretrained vision encoder [43, 62] to transform input images into a set of visual tokens. These tokens are then projected into the input space of the LLM through a projection module (e.g., MLP [30, 32], QFormer [9] and Abstractor [60]) and processed alongside text inputs. Building on the progress of MLLMs, recent studies have shifted from designing specialized methods for certain tasks [44-46] to employing MLLMs as generalist agents, capable of handling diverse practical tasks such as document analysis [15, 57, 64], online video assistant [24], retrieval-augmented generation [40], game agents [27, 29], robot manipulation [18, 21, 39], and GUI agent [3, 5, 54]. Despite these advances, MLLMs still struggle with the fixed low resolution of pretrained image encoders, which can limit their effectiveness in handling high-resolution images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 478, + 194, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 478, + 194, + 490 + ], + "spans": [ + { + "bbox": [ + 55, + 478, + 194, + 490 + ], + "type": "text", + "content": "2.2. High-resolution MLLMs" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 496, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 295, + 628 + ], + "type": "text", + "content": "To address the limitation of MLLMs in handling only low-resolution visual inputs, early studies attempted to either directly incorporate vision encoders that were pre-trained on higher-resolution images [36, 38], or to improve the resolution of pre-trained encoders through additional training [1, 60]. While achieving some progress, they still remain restricted by the fixed resolutions. To accommodate the varying resolutions and aspect ratios, cropping-based techniques have been widely adopted in recent works [15, 17, 33, 58, 67]. However, these approaches still face challenges related to visual redundancy and fragmentation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "text", + "content": "To mitigate the issue of redundancy, existing models typically require an additional compression module. For example, Abstractor-like modules utilizing learnable queries combined with cross-attention are widely adopted [1, 28, 33, 58, 60]. These modules typically require pretraining on massive amounts of data, which diminishes their practical efficacy. In contrast, some approaches use pooling" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 262, + 555, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 555, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 555, + 455 + ], + "type": "text", + "content": "layers [48, 55] or convolutional layers [15, 36]. Though resource-efficient, they often fail to address the issue of redundancy. In this work, we present an initial exploration to directly eliminate redundancy during the visual encoding phase, thereby removing the need for additional compression. To address the issue of fragmentation, several concurrent studies also make preliminary investigation. TextMonkey [33] uses Shifted Window Attention [35] to exchange information between sub-images. But the exchange is still limited to local regions. MiniMonkey [16] proposed a Complementary Image Pyramid strategy to generate complementary cropping schemes. However, this strategy inevitably encodes the same visual content multiple times, further exacerbating visual redundancy. In this work, we propose the ReAtten module to facilitate continuous visual encoding through the interaction among visual registers." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 466, + 379, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 466, + 379, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 466, + 379, + 478 + ], + "type": "text", + "content": "3. FALCON" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 485, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 555, + 713 + ], + "type": "text", + "content": "The overall framework of FALCON is illustrated in Fig. 2. The method begins with shape-adaptive cropping [58], where the image is cropped into multiple non-overlapping sub-images, each matching the resolution of the pre-trained vision encoder. Additionally, a resized version of the full image is used to supplement global layout information. To directly eliminate redundancy during visual encoding, we propose Register-based Representation Compacting (Re-Compact) mechanism. This mechanism introduces a set of learnable visual registers, which are paired with image tokens from each sub-image and fed into the vision encoder to capture rich visual information. To ensure the continuity of visual semantics throughout the encoding, a novel Register Interactive Attention (ReAtten) module is integrated into the Vision Transformer (ViT) [11] to facilitate information exchange between sub-images via the visual registers. Finally, the compact visual representations produced by the visual registers are processed through a simple MLP module before being fed into the LLM for further analysis." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23532" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 85 + ], + "type": "text", + "content": "3.1. Mitigating Visual Redundancy via ReCompact" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 328 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 328 + ], + "type": "text", + "content": "Existing MLLMs typically use a ViT-based vision encoder, such as CLIP [43]. For an input image of size " + }, + { + "bbox": [ + 55, + 89, + 296, + 328 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 55, + 89, + 296, + 328 + ], + "type": "text", + "content": ", it is patched into multiple " + }, + { + "bbox": [ + 55, + 89, + 296, + 328 + ], + "type": "inline_equation", + "content": "P \\times P" + }, + { + "bbox": [ + 55, + 89, + 296, + 328 + ], + "type": "text", + "content": " patches. Each patch is treated as an individual image token, resulting in " + }, + { + "bbox": [ + 55, + 89, + 296, + 328 + ], + "type": "inline_equation", + "content": "N = \\frac{H}{P} \\times \\frac{W}{P}" + }, + { + "bbox": [ + 55, + 89, + 296, + 328 + ], + "type": "text", + "content": " tokens, denoted as " + }, + { + "bbox": [ + 55, + 89, + 296, + 328 + ], + "type": "inline_equation", + "content": "I = \\{i_1, \\dots, i_N\\}" + }, + { + "bbox": [ + 55, + 89, + 296, + 328 + ], + "type": "text", + "content": ". In most current MLLMs, output features corresponding to all image tokens are usually preserved in their entirety. However, when high-resolution images are included, the number of image tokens increases dramatically, placing a substantial burden on the computational resources of LLMs. Consequently, most high-resolution MLLMs necessitate an additional compression module, which can be resource-inefficient. Previous research has shown that most image tokens contain only localized information from their surrounding areas [10]. It suggests that a significant portion of these tokens may be redundant, particularly those originating from background regions. This observation motivates us to directly eliminate redundancy in the output of the vision encoder, thereby reducing the number of visual tokens and avoiding the need for an additional compression module." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 329, + 296, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 329, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 329, + 296, + 534 + ], + "type": "text", + "content": "To achieve this, we propose a novel Register-based Representation Compacting (ReCompact) mechanism. This mechanism employs a set of visual registers to aggregate essential information from image tokens, resulting in a more compact visual representation. Specifically, as shown in Fig. 2, we introduce a set of learnable visual registers, denoted by " + }, + { + "bbox": [ + 55, + 329, + 296, + 534 + ], + "type": "inline_equation", + "content": "R = \\{r_1,\\dots ,r_M\\}" + }, + { + "bbox": [ + 55, + 329, + 296, + 534 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 329, + 296, + 534 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 329, + 296, + 534 + ], + "type": "text", + "content": " represents the number of visual registers, and " + }, + { + "bbox": [ + 55, + 329, + 296, + 534 + ], + "type": "inline_equation", + "content": "M\\ll N" + }, + { + "bbox": [ + 55, + 329, + 296, + 534 + ], + "type": "text", + "content": ". Given a single input image, the image tokens are concatenated with visual registers to create a new input for ViT, represented as " + }, + { + "bbox": [ + 55, + 329, + 296, + 534 + ], + "type": "inline_equation", + "content": "V = \\mathrm{Concat}(I,R) = \\{i_1,\\dots ,i_N,r_1,\\dots ,r_M\\}" + }, + { + "bbox": [ + 55, + 329, + 296, + 534 + ], + "type": "text", + "content": ". During visual encoding, the self-attention mechanism in ViT enables the visual register to adaptively aggregate important information from image tokens while discarding redundant content. Denote " + }, + { + "bbox": [ + 55, + 329, + 296, + 534 + ], + "type": "inline_equation", + "content": "X_{k,l}\\in \\mathcal{R}^{(N + M)\\times D}" + }, + { + "bbox": [ + 55, + 329, + 296, + 534 + ], + "type": "text", + "content": " as the input hidden states of the " + }, + { + "bbox": [ + 55, + 329, + 296, + 534 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 329, + 296, + 534 + ], + "type": "text", + "content": "-th sub-image at the " + }, + { + "bbox": [ + 55, + 329, + 296, + 534 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 329, + 296, + 534 + ], + "type": "text", + "content": "-th layer, where " + }, + { + "bbox": [ + 55, + 329, + 296, + 534 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 329, + 296, + 534 + ], + "type": "text", + "content": " is the dimensionality. The self-attention is formulated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 542, + 296, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 542, + 296, + 574 + ], + "spans": [ + { + "bbox": [ + 63, + 542, + 296, + 574 + ], + "type": "interline_equation", + "content": "\\hat {X} _ {k, l} = \\operatorname {S e l f - A t t e n} \\left(X _ {k, l}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {X _ {k , l} X _ {k , l} ^ {T}}{\\sqrt {D _ {\\text {k e y}}}}\\right) X _ {k, l} \\tag {1}", + "image_path": "adeee11300c00dfc417baf6feab0592ff0ff66a022ef4ea8047c2e7747def3ba.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 581, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 295, + 689 + ], + "type": "text", + "content": "Notably, we do not apply any attention masks for specific constraints. Our experiments show that pretraining on a modest amount of data is sufficient to effectively adapt the pretrained ViT to the proposed ReCompact, enabling it to aggregate essential information within the visual registers. After visual encoding, The output features corresponding to the visual registers, denoted as " + }, + { + "bbox": [ + 55, + 581, + 295, + 689 + ], + "type": "inline_equation", + "content": "F = \\{f_{1},\\dots ,f_{M}\\}" + }, + { + "bbox": [ + 55, + 581, + 295, + 689 + ], + "type": "text", + "content": ", are then used as the visual representation of the image and fed into the LLM for various vision-language tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "To further adapt the ReCompact mechanism for cropping-based high-resolution image perception, we share" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 205 + ], + "type": "text", + "content": "the same set of visual registers as input for different sub-images. The input for the " + }, + { + "bbox": [ + 313, + 72, + 555, + 205 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 72, + 555, + 205 + ], + "type": "text", + "content": "-th sub-image is formulated as " + }, + { + "bbox": [ + 313, + 72, + 555, + 205 + ], + "type": "inline_equation", + "content": "V_{k} = \\mathrm{Concat}(I_{k},R)" + }, + { + "bbox": [ + 313, + 72, + 555, + 205 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 72, + 555, + 205 + ], + "type": "inline_equation", + "content": "k = 1,2,\\dots ,N_{c}" + }, + { + "bbox": [ + 313, + 72, + 555, + 205 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 205 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 313, + 72, + 555, + 205 + ], + "type": "text", + "content": " denotes the total number of sub-images. After encoding, the outputs of the visual registers for all sub-images are concatenated, yielding " + }, + { + "bbox": [ + 313, + 72, + 555, + 205 + ], + "type": "inline_equation", + "content": "F_{hr} = \\mathrm{Concat}(F_1,F_2,\\dots ,F_{N_c})" + }, + { + "bbox": [ + 313, + 72, + 555, + 205 + ], + "type": "text", + "content": " which serves as the input to the LLM. The ReCompact mechanism enables the vision encoder to represent rich visual information using only a minimal number of " + }, + { + "bbox": [ + 313, + 72, + 555, + 205 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 72, + 555, + 205 + ], + "type": "text", + "content": " tokens per sub-image, which significantly reduces the visual redundancy and eliminates the need for additional compression." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 211, + 524, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 211, + 524, + 224 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 524, + 224 + ], + "type": "text", + "content": "3.2. Ensuring Visual Continuity via ReAtten" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 228, + 555, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 431 + ], + "type": "text", + "content": "When using the cropping-based high-resolution image encoding method, each sub-image is independently encoded by the vision encoder. This process can disrupt semantic coherence and result in the misrepresentation of visual concepts. To preserve the continuity of visual semantics, it is crucial to enable information exchange between sub-images during encoding. A naive approach would be to directly concatenate the image features from all sub-images and apply self-attention. However, the quadratic computational complexity makes this approach impractical. In Sec. 3.1, we introduced the ReCompact mechanism, where the visual registers effectively aggregate rich visual information while discarding redundancy during visual encoding, allowing them to serve as compact yet comprehensive image representations. Consequently, interactions between visual registers can naturally facilitate sufficient information exchange across different sub-images in an efficient manner." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 432, + 556, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 556, + 596 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 556, + 596 + ], + "type": "text", + "content": "To enable information exchange between visual registers across different sub-images, we introduce a novel Register Interactive Attention (ReAtten) module. As shown in Fig. 2, the ReAtten module is integrated into each of the " + }, + { + "bbox": [ + 313, + 432, + 556, + 596 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 432, + 556, + 596 + ], + "type": "text", + "content": " layers of the ViT. After the self-attention process in Eq. (1), we acquire " + }, + { + "bbox": [ + 313, + 432, + 556, + 596 + ], + "type": "inline_equation", + "content": "\\hat{X}_{k,l} = \\{\\hat{X}_{k,l}^{I},\\hat{X}_{k,l}^{R}\\}" + }, + { + "bbox": [ + 313, + 432, + 556, + 596 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 432, + 556, + 596 + ], + "type": "inline_equation", + "content": "\\hat{X}^I\\in \\mathcal{R}^{N\\times D}" + }, + { + "bbox": [ + 313, + 432, + 556, + 596 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 432, + 556, + 596 + ], + "type": "inline_equation", + "content": "\\hat{X}^R\\in \\mathcal{R}^{M\\times D}" + }, + { + "bbox": [ + 313, + 432, + 556, + 596 + ], + "type": "text", + "content": " represent the hidden states corresponding to image tokens and visual registers, respectively. The states of the visual registers from different sub-images are concatenated to form " + }, + { + "bbox": [ + 313, + 432, + 556, + 596 + ], + "type": "inline_equation", + "content": "\\hat{X}_l^R = \\mathrm{Concat}\\left(\\{\\hat{X}_{k,l}^R\\}_{k = 1}^{N_c}\\right)\\in" + }, + { + "bbox": [ + 313, + 432, + 556, + 596 + ], + "type": "inline_equation", + "content": "\\mathcal{R}^{M\\cdot N_c\\times D}" + }, + { + "bbox": [ + 313, + 432, + 556, + 596 + ], + "type": "text", + "content": ". Then these concatenated registers are processed via ReAtten to enable information exchange across subimages, which can be formulated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 318, + 604, + 555, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 604, + 555, + 622 + ], + "spans": [ + { + "bbox": [ + 318, + 604, + 555, + 622 + ], + "type": "interline_equation", + "content": "\\overline {{X}} _ {l} ^ {R} = \\mathrm {R e A t t e n} (\\hat {X} _ {l} ^ {R}) = \\hat {X} _ {l} ^ {R} + \\mathrm {C r o s s - V i T - A t t e n} (\\hat {X} _ {l} ^ {R}) (2)", + "image_path": "856a95cbca31ce19f642462b2a4e2b70ce69b92138e43d00e5723534bdc29fca.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 628, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 555, + 689 + ], + "type": "text", + "content": "where Cross-ViT-Atten is implemented as a self-attention layer. Once the information exchange is completed, the registers and image tokens from a same sub-image are concatenated and then passed into the feedforward network (FFN) for transformation, yielding the output of the " + }, + { + "bbox": [ + 313, + 628, + 555, + 689 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 628, + 555, + 689 + ], + "type": "text", + "content": "-th layer:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 357, + 696, + 555, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 696, + 555, + 717 + ], + "spans": [ + { + "bbox": [ + 357, + 696, + 555, + 717 + ], + "type": "interline_equation", + "content": "X _ {k, l + 1} = \\operatorname {F F N} \\left(\\operatorname {C o n c a t} \\left(\\hat {X} _ {k, l} ^ {I}, \\bar {X} _ {k, l} ^ {R}\\right)\\right) \\tag {3}", + "image_path": "df01b980e76f85c80e6f7ed2844b103ab68bd616235b3fe53d24e95f4ceae49a.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "23533" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 69, + 282, + 269 + ], + "blocks": [ + { + "bbox": [ + 67, + 69, + 282, + 269 + ], + "lines": [ + { + "bbox": [ + 67, + 69, + 282, + 269 + ], + "spans": [ + { + "bbox": [ + 67, + 69, + 282, + 269 + ], + "type": "image", + "image_path": "7fea22fdc3cc3b493e312b32f4a98ebdb9f316bfc9d08bbff01a1afa4a45a549.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 277, + 295, + 354 + ], + "lines": [ + { + "bbox": [ + 55, + 277, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 295, + 354 + ], + "type": "text", + "content": "Figure 3. The progressive training pipeline. **Stage 0:** Align the frozen vision and language backbones by training only the MLP and registers. **Stage 1:** Tune all parameters with caption and OCR data on low-resolution for coarse alignment. **Stage 2:** Integrate high-resolution input along with fine-grained captions and OCR to achieve refined alignment. **Stage 3:** Froze ViT and introduce ReAtten module for high-resolution instruction tuning." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 358, + 296, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 296, + 419 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 296, + 419 + ], + "type": "text", + "content": "By integrating ReAtten into each layer of the ViT model, FALCON enables effective information exchange between cropped sub-images through visual registers, thereby ensuring semantic continuity of the high-resolution image throughout the visual encoding process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 429, + 295, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 429, + 295, + 454 + ], + "spans": [ + { + "bbox": [ + 55, + 429, + 295, + 454 + ], + "type": "text", + "content": "3.3. Learning Expressive Visual Representation via Progressive Training Pipeline" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 460, + 295, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 532 + ], + "type": "text", + "content": "To empower the ReCompact mechanism with the capability to capture rich visual patterns and to enable smoother initialization of the ReAtten module under high-resolution conditions, we develop a specialized progressive training pipeline. As illustrated in Fig. 3, this pipeline progressively transitions from low to high resolution across four stages." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 533, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 640 + ], + "type": "text", + "content": "Stage 0: Static Alignment on Low-resolution. Since both the visual registers and MLP projector are randomly initialized, in stage 0, we aim to achieve a initial alignment between the pretrained vision encoder and the LLM. To facilitate this, we begin by using low-resolution images as inputs, which are directly resized to match the resolution of the vision encoder. We then statically pretrain the model by keeping both the vision encoder and LLM frozen, tuning only the visual registers and MLP using image caption data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "Stage 1: Coarse Alignment on Low-resolution. To better adapt the pre-trained vision encoder to our newly introduced ReCompact mechanism, in stage 1, we unfreeze all model parameters to enable a more thorough tuning. To facilitate the visual registers to progressively learn capturing visual information from coarse to fine granularity, we con" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "tinue to use low-resolution inputs and employ high-quality, extended captions rich in visual content, along with full-text OCR data for training. This design allows the registers to acquire holistic scene understanding ability." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 121, + 553, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 553, + 215 + ], + "type": "text", + "content": "Stage 2: Refined Alignment on High-resolution. To further enhance the capacity of the visual register to capture fine-grained visual details, in stage 2, we transition the model to process high-resolution image for refined alignment. Additionally, we employ datasets that includes a series of fine-grained captions and OCR tasks, such as regional captions and text location. This facilitates the visual registers to capture comprehensive fine-grained details." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 216, + 555, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 555, + 407 + ], + "type": "text", + "content": "Stage 3: Instruction Tuning on High-resolution. To adapt the well-aligned visual register for a wide range of real-world tasks, in stage 3, we perform instruction tuning with a mixture of datasets covering various tasks, such as VQA, document understanding, and science reasoning. Additionally, the proposed ReAtten module is integrated to address visual fragmentation caused by cropping-based high-resolution image encoding. Since the vision encoder with ReCompact is well-aligned, the vision encoder is kept frozen and other parameters including ReAtten are trainable during instruction tuning. Notably, to facilitate a smoother initialization for the newly introduced ReAtten modules, we initialize parameters of each ReAtten module with the ViT self-attention parameters at the corresponding depth. This design enables ReAtten to effectively leverage pre-trained parameters, avoiding the need for extensive fine-tuning." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 418, + 395, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 395, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 395, + 430 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 437, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 521 + ], + "type": "text", + "content": "We implement FALCON with SigLIP-L/16-384px [62] as the vision encoder and Llama-3.1-8B-Instruct [12] as the LLM backbone. A two-layer MLP module with GeLU [13] activation is served as the vision-language projection layer. The number of visual registers is set to 64, achieving remarkable visual token compression rates of " + }, + { + "bbox": [ + 313, + 437, + 554, + 521 + ], + "type": "inline_equation", + "content": "9 \\times" + }, + { + "bbox": [ + 313, + 437, + 554, + 521 + ], + "type": "text", + "content": ". The maximum number of sub-images is set to 16." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 529, + 437, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 437, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 437, + 541 + ], + "type": "text", + "content": "4.1. Experimental Results" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 545, + 553, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 553, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 553, + 665 + ], + "type": "text", + "content": "We evaluate FALCON using the recently proposed MMERealWorld [66] benchmark, a comprehensive and largescale benchmark specifically designed to assess the high-resolution understanding capabilities of MLLMs. MMERealWorld comprises 13,366 high-resolution images with an average resolution of " + }, + { + "bbox": [ + 313, + 545, + 553, + 665 + ], + "type": "inline_equation", + "content": "2,000 \\times 1,500" + }, + { + "bbox": [ + 313, + 545, + 553, + 665 + ], + "type": "text", + "content": " pixels, spanning five domains: OCR in the wild, Remote Sensing, Diagrams and Tables, Monitoring, and Autonomous Driving. This benchmark is well-suited for evaluating the practical high-resolution understanding capabilities of MLLMs." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": "As shown in Tab. 1, FALCON significantly outperform other advanced MLLMs in both perception and reasoning tasks. FALCON achieves state-of-the-art performance with only 64 tokens per sub-image, resulting in a remarkable vi" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23534" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 171 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 171 + ], + "type": "text", + "content": "Table 1. Evaluation on MME-RealWorld Benchmark. \"OCR\", \"RS\", \"DT\", \"MO\", and \"AD\" stand for OCR in the Wild, Remote Sensing, Diagram and Table, Monitoring, and Autonomous Driving, respectively. \"Avg\" denotes the weighted average accuracy across all samples, while \"Avg-C\" denotes the weighted average accuracy across different domains. \"Sizev\" indicates the image resolution supported by the model, where " + }, + { + "bbox": [ + 55, + 70, + 555, + 171 + ], + "type": "inline_equation", + "content": "\\times n" + }, + { + "bbox": [ + 55, + 70, + 555, + 171 + ], + "type": "text", + "content": " signifies a cropping-based high-resolution processing approach with up to " + }, + { + "bbox": [ + 55, + 70, + 555, + 171 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 70, + 555, + 171 + ], + "type": "text", + "content": " sub-images. \"Tokenc\" denotes the number of visual tokens after compression. † For models utilizing hybrid vision encoders, \"Sizev\" represents the setting of encoder with the highest resolution. We gray out the models that are either closed-source or have at least twice the number of visual tokens compared to ours. Among these, the lighter gray denotes scores below ours. The underline indicates that the second performance is achieved. Notably, FALCON achieves state-of-the-art performance with just 64 tokens per sub-image, surpassing most models that have large amount of visual tokens and many advanced closed-source commercial models." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 178, + 552, + 448 + ], + "blocks": [ + { + "bbox": [ + 58, + 178, + 552, + 448 + ], + "lines": [ + { + "bbox": [ + 58, + 178, + 552, + 448 + ], + "spans": [ + { + "bbox": [ + 58, + 178, + 552, + 448 + ], + "type": "table", + "html": "
TypeMethod\\( Size^v \\)\\( Token^c \\)PerceptionReasoning
Task SplitOCRRSDTMOADAvgAvg-COCRDTMOADAvgAvg-C
Visual Tokens\\(\\nabla 1024\\)MiniGPT-v2 [4]\\( 448^2 \\)25639.023.320.419.326.026.925.630.020.416.923.723.022.7
Qwen-VL-Chat [1]\\( 448^2 \\)25632.415.115.622.115.120.820.128.613.616.524.622.020.8
LLaVA1.5-13B [30]\\( 336^2 \\)57644.123.320.220.526.128.426.830.220.827.524.825.525.8
ShareGPT4V-13B [6]\\( 336^2 \\)57644.623.120.219.326.128.426.626.020.827.324.624.624.7
TextMonkey [33]\\( 448^2 \\times 4 \\)256 × 437.311.75.916.114.318.217.130.42.24.420.016.014.3
Mini-Gemini-7B-HD† [26]\\( 336^2 \\times 4 \\)57642.031.322.334.224.831.130.935.424.625.923.326.127.3
DeepSeek-VL† [36]\\( 1024^2 \\)57649.625.523.427.033.433.131.845.223.816.727.328.028.3
MiniCPM-V 2.5 [56]\\( 448^2 \\times 9 \\)96 × 966.827.752.838.734.247.444.044.031.837.031.034.536.0
Monkey [28]\\( 448^2 \\times 4 \\)256 × 454.625.032.528.029.736.334.027.220.827.333.028.827.1
FALCON\\( 384^2 \\times 16 \\)64 × 1666.449.647.240.436.550.348.055.439.244.434.540.743.4
Visual Tokens\\(\\nabla 2048\\)LLaVA-Next-Qwen72B [31]\\( 336^2 \\times 4 \\)576 × 437.129.127.729.418.029.028.317.234.227.329.727.927.1
LLaVA-Next-liama3 [31]\\( 336^2 \\times 4 \\)576 × 447.925.426.619.518.730.127.655.223.421.130.732.132.6
mPLUG-DocOwl 1.5 [15]\\( 448^2 \\times 9 \\)256 × 951.223.729.325.028.333.731.542.619.820.526.026.927.2
CogVLM2-liama3-Chat [14]\\( 1344^2 \\)230470.028.847.533.730.245.842.054.032.841.231.237.339.8
IXC2.5 [63]\\( 560^2 \\times 24 \\)400 × 2469.336.163.939.533.652.548.553.441.017.730.033.935.5
LLaVA-OneVision [20]\\( 384^2 \\times 9 \\)729 × 978.753.560.740.345.859.655.861.840.040.834.141.244.2
Closed-sourceGPT-4o-mini--62.56.744.226.524.237.132.847.039.825.826.832.534.9
Gemini-1.5-pro [50]--67.614.039.931.126.639.635.952.733.228.319.229.233.4
GPT-4o--77.728.946.733.922.446.441.961.444.836.526.437.642.3
Claude 3.5 Sonnet--72.525.767.432.240.852.947.761.961.241.831.944.149.2
", + "image_path": "16635d14cdbcc4f4fe268638884ac70a1a1f19331fea4bda213dfe27f740dd5a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 454, + 295, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 454, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 295, + 657 + ], + "type": "text", + "content": "sual compression rate of 9 times. In addition, FALCON also outperforms most models that have twice as many visual tokens. Specifically, FALCON outperforms CogVLM2-llama3-Chat [14] in average scores on both perception and reasoning task, which uses 2304 tokens per-image. It also surpasses IXC2.5 [63] in average scores on the reasoning tasks, which has a maximum token counts of 9600. Notably, FALCON achieves higher average scores in both perception and reasoning tasks compared to three advanced commercial models, GPT-4o, GPT-4o-mini and Gemini1.5-pro [50], with it notably outperforming all commercial models in the sub-tasks of remote sensing and monitoring. These results demonstrate that the ReCompact and ReAtten technique proposed by FALCON effectively eliminate visual redundancy and ensure visual continuity, allowing comprehensive and detailed visual information to be conveyed with a highly compact visual representation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "type": "text", + "content": "To further assess FALCON's capabilities in high-resolution understanding, we conduct evaluations on more high-resolution benchmarks: V* [53], DocVQA [42], TextVQA [49] and ChartQA [41]. V* focuses on evaluating" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 454, + 553, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 553, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 553, + 537 + ], + "type": "text", + "content": "models' capabilities for fine-grained attribute recognition and spatial reasoning. DocVQA challenges models to comprehend text-dense document images. ChartQA evaluates models' proficiency in interpreting various types of chart, while TextVQA assesses the ability to recognize small text in real-world scenes. As shown in Tab. 2, FALCON demonstrates superior performance across all these benchmarks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 539, + 555, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 539, + 555, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 555, + 647 + ], + "type": "text", + "content": "To evaluate the performance of FALCON across a broader range of domains, we further conduct experiments with diverse popular benchmarks. As shown in Tab. 4, POPE [25], ScienceQA [37], MMBench [34] and SEED-Bench [19] are selected to assess models' capabilities in hallucination, scientific reasoning, and comprehensive ability. FALCON achieves SOTA performance across all these benchmarks. The results demonstrate that FALCON exhibits superior capabilities across a wide range of tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 658, + 408, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 408, + 670 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 408, + 670 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 676, + 553, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 676, + 553, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 676, + 553, + 700 + ], + "type": "text", + "content": "We perform extensive ablation studies to validate the effectiveness of the proposed ReCompact and ReAtten." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 325, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 553, + 713 + ], + "type": "text", + "content": "Ablation on Training Data. To eliminate the influx" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23535" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 113, + 289, + 269 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "type": "text", + "content": "Table 2. Evaluations on diverse high-resolution Benchmarks. \"V* " + }, + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "type": "inline_equation", + "content": "{}_{Avg}" + }, + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "type": "text", + "content": " \" represents the weighted average score for the two subtasks in the " + }, + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "type": "inline_equation", + "content": "{V}^{ * }" + }, + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "type": "text", + "content": " benchmark. † denotes OCR inputs are utilized." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 113, + 289, + 269 + ], + "lines": [ + { + "bbox": [ + 61, + 113, + 289, + 269 + ], + "spans": [ + { + "bbox": [ + 61, + 113, + 289, + 269 + ], + "type": "table", + "html": "
ModelV*AvgDocVQATextVQAChartQA
Low-resolution MLLMs
Qwen-VL [1]-65.163.859.0
InstructBLIP [9]34.04.550.710.9
LLaVA-v1.5-7B [30]48.78.558.2†17.8
mPLUG-Owl2 [60]-17.958.2†22.8
High-resolution MLLMs
UReader [58]-65.457.659.3
Monkey [28]-66.565.859.0
LlvaF-FlexAttn [23]54.4-48.9-
S2-Wrapper-7B [48]55.5-61.0-
LlvaNxt-Vicuna7B [31]61.8-64.954.3
SliME-7B [67]--64.433.9
FALCON64.474.466.568.8
", + "image_path": "d317d5be1748869dd71a6b23903c1097fcf424c80a45494cda3caca2b06ef27f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 272, + 296, + 316 + ], + "lines": [ + { + "bbox": [ + 55, + 272, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 272, + 296, + 316 + ], + "type": "text", + "content": "Table 3. Ablation results on training data. We train FALCON with two-stage pipeline and exactly the same dataset as LLaVAv1.5 [30]. The results confirm the superiority of the proposed method when the influence of training data is excluded." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 56, + 326, + 296, + 454 + ], + "blocks": [ + { + "bbox": [ + 56, + 326, + 296, + 454 + ], + "lines": [ + { + "bbox": [ + 56, + 326, + 296, + 454 + ], + "spans": [ + { + "bbox": [ + 56, + 326, + 296, + 454 + ], + "type": "image", + "image_path": "033cda0170d90fa03b901d164de4d812f479bd3f65f7a6d76335bf8e0d8b24d8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 461, + 295, + 518 + ], + "lines": [ + { + "bbox": [ + 55, + 461, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 295, + 518 + ], + "type": "text", + "content": "Figure 4. Comparison of different compression methods. We report the sum of the \"Avg-C\" scores for the perception and reasoning tasks in MME-RealWorld. The superior performance demonstrates that ReCompact effectively eliminates redundancy and produces a more expressive and compact visual representation." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 521, + 296, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 296, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 296, + 592 + ], + "type": "text", + "content": "ence of training data and demonstrate the advantages of the proposed method, we utilize exactly the same dataset as LLaVA-v1.5 [30] and employ a two-stage pipeline for training. As shown in Tab. 3, the proposed method exhibits significant performance advantages across multiple benchmarks when trained on the same dataset." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "content": "Effectiveness of ReCompact mechanism. To verify the advantages of the proposed ReCompact mechanism in reducing visual redundancy and compressing visual representations during the visual encoding, we compare the Re-Compact with other commonly used compression methods. This includes pooling [48, 55], Pixel Shuffle [7] and Abstractor [60], where the Abstractor serves as a representative module that combines query tokens with cross-attention." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "As shown in Fig. 4, ReCompact demonstrates superior performance compared to other compression methods." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 319, + 113, + 548, + 259 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 105 + ], + "type": "text", + "content": "Table 4. Evaluation on benchmarks across diverse domains. The POPE, SQA, MMBench and SEED-Bench are used to assess hallucination, scientific reasoning, and overall abilities, respectively." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 113, + 548, + 259 + ], + "lines": [ + { + "bbox": [ + 319, + 113, + 548, + 259 + ], + "spans": [ + { + "bbox": [ + 319, + 113, + 548, + 259 + ], + "type": "table", + "html": "
ModelSQAPOPEMMBSEED-Img
Low-Resolution MLLMs
Qwen-VL [1]67.1-41.662.3
InstructBLIP [9]63.178.938.458.8
LlaVA-v1.5-7B [30]66.885.962.865.8
mPLUG-Owl2 [60]68.785.863.564.5
High-Resolution MLLMs
Monkey [28]69.483.757.464.3
LlaVA-FlexAttn [23]-85.965.7-
S2-Wrapper-7B [48]--66.267.9
LLaVA-Next-Vicuna7B [31]70.186.566.569.6
SliME-7B [67]76.885.463.167.0
FALCON88.888.076.173.3
", + "image_path": "1150b29a95b08510f57717b0145b9822dac0fba53380b848ac9d6d1e3113870e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 333, + 349, + 533, + 419 + ], + "blocks": [ + { + "bbox": [ + 313, + 262, + 555, + 340 + ], + "lines": [ + { + "bbox": [ + 313, + 262, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 555, + 340 + ], + "type": "text", + "content": "Table 5. Comparison of different methods for maintaining visual continuity. The \"Baseline\" model refers to a condition where no specific methods are applied to preserve visual continuity. We report the \"Avg-C\" score for MME-RealWorld. The results show that the ReAtten module more effectively addresses the issue of visual fragmentation compared to other methods, thereby improving overall performance and reducing hallucinations." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 333, + 349, + 533, + 419 + ], + "lines": [ + { + "bbox": [ + 333, + 349, + 533, + 419 + ], + "spans": [ + { + "bbox": [ + 333, + 349, + 533, + 419 + ], + "type": "table", + "html": "
MethodV*AVgMME-RealWorldPOPE
PerceptionReasoning
Baseline51.338.235.685.7
CIP [16]50.338.437.286.3
W-Atten [33]60.241.038.186.4
ReAtten61.342.139.087.3
", + "image_path": "de7f372c5cc25c3bf236cf9c0af8cca376b8630b98794d3b553b18055604e759.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 425, + 555, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 555, + 555 + ], + "type": "text", + "content": "These results indicate that ReCompact effectively eliminates visual redundancy and produces a more expressive compact visual representation. In contrast, we argue that the pooling layer and Pixel Shuffle merely reduces the number of tokens by merging adjacent ones, failing to eliminate redundancy that interferes with the LLM's understanding of visual content. Additionally, although Abstractor uses more parameters, its performance is generally lower than that of pooling and ReCompact. This result suggests that even millions of samples may be insufficient for fully training Abstractor, which diminishes their practical efficacy." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 556, + 554, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 554, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 554, + 629 + ], + "type": "text", + "content": "Effectiveness of ReAtten Module. To validate the advantage of the proposed ReAtten Module, we compare ReAtten with other contemporary methods, including the Shifted Window Attention (W-Attn) [35] adopted by TextMonkey [33] and the Complementary Image Pyramid (CIP) introduced by MiniMonkey [16]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 630, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 715 + ], + "type": "text", + "content": "As shown in Tab. 5, the model incorporating ReAtten achieves significant improvements over the baseline across all benchmarks and demonstrates superiority compared to other methods. These results indicate that the ReAtten module effectively enhances the ability of the model in high-resolution perception and mitigates hallucinations by addressing the issue of visual fragmentation. In contrast, we" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23536" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 87, + 204, + 185 + ], + "blocks": [ + { + "bbox": [ + 61, + 87, + 204, + 185 + ], + "lines": [ + { + "bbox": [ + 61, + 87, + 204, + 185 + ], + "spans": [ + { + "bbox": [ + 61, + 87, + 204, + 185 + ], + "type": "image", + "image_path": "69a61bdb051446db242affaf1ca3d6af18de2936a64fc459257188dc0fac5037.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 198, + 168, + 211 + ], + "lines": [ + { + "bbox": [ + 94, + 198, + 168, + 211 + ], + "spans": [ + { + "bbox": [ + 94, + 198, + 168, + 211 + ], + "type": "text", + "content": "(a) Input Image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 211, + 73, + 286, + 132 + ], + "blocks": [ + { + "bbox": [ + 211, + 73, + 286, + 132 + ], + "lines": [ + { + "bbox": [ + 211, + 73, + 286, + 132 + ], + "spans": [ + { + "bbox": [ + 211, + 73, + 286, + 132 + ], + "type": "image", + "image_path": "8ceb0250439bd1470cb952aa6838c3da9139588af5237a80a9f74e4373251c0c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 211, + 133, + 286, + 190 + ], + "blocks": [ + { + "bbox": [ + 211, + 133, + 286, + 190 + ], + "lines": [ + { + "bbox": [ + 211, + 133, + 286, + 190 + ], + "spans": [ + { + "bbox": [ + 211, + 133, + 286, + 190 + ], + "type": "image", + "image_path": "bd66f1083bbf9ca89b53d071f4a99d6ba15b6ea2d41b3298cad141cc222eef11.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 198, + 286, + 209 + ], + "lines": [ + { + "bbox": [ + 211, + 198, + 286, + 209 + ], + "spans": [ + { + "bbox": [ + 211, + 198, + 286, + 209 + ], + "type": "text", + "content": "(b) Register-to-1" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 286, + 74, + 359, + 132 + ], + "blocks": [ + { + "bbox": [ + 286, + 74, + 359, + 132 + ], + "lines": [ + { + "bbox": [ + 286, + 74, + 359, + 132 + ], + "spans": [ + { + "bbox": [ + 286, + 74, + 359, + 132 + ], + "type": "image", + "image_path": "5059284c6bb109e87b39beafc78b52016928dfabae1f0d576551336ded1848a6.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 286, + 133, + 359, + 190 + ], + "blocks": [ + { + "bbox": [ + 286, + 133, + 359, + 190 + ], + "lines": [ + { + "bbox": [ + 286, + 133, + 359, + 190 + ], + "spans": [ + { + "bbox": [ + 286, + 133, + 359, + 190 + ], + "type": "image", + "image_path": "95f15239a632e544a78a17d734546edff477e66bf722bd4973c44a87b7dcafed.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 286, + 198, + 359, + 209 + ], + "lines": [ + { + "bbox": [ + 286, + 198, + 359, + 209 + ], + "spans": [ + { + "bbox": [ + 286, + 198, + 359, + 209 + ], + "type": "text", + "content": "Image Attention" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 366, + 74, + 507, + 133 + ], + "blocks": [ + { + "bbox": [ + 366, + 74, + 507, + 133 + ], + "lines": [ + { + "bbox": [ + 366, + 74, + 507, + 133 + ], + "spans": [ + { + "bbox": [ + 366, + 74, + 507, + 133 + ], + "type": "image", + "image_path": "d67645d774671878f7e1abc1423fa5dc384083272d41db0887c544e8031ffb02.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 367, + 135, + 507, + 191 + ], + "blocks": [ + { + "bbox": [ + 367, + 135, + 507, + 191 + ], + "lines": [ + { + "bbox": [ + 367, + 135, + 507, + 191 + ], + "spans": [ + { + "bbox": [ + 367, + 135, + 507, + 191 + ], + "type": "image", + "image_path": "1253deae450cc82ae011ce0b29a5cf61ef43fe4fbaff73bfe804d96497da7381.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 382, + 198, + 493, + 209 + ], + "lines": [ + { + "bbox": [ + 382, + 198, + 493, + 209 + ], + "spans": [ + { + "bbox": [ + 382, + 198, + 493, + 209 + ], + "type": "text", + "content": "(c) Ablation on ReAtten" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 511, + 98, + 545, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 511, + 98, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 511, + 98, + 545, + 119 + ], + "type": "text", + "content": "w/ ReAtten" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 511, + 156, + 545, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 511, + 156, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 511, + 156, + 545, + 175 + ], + "type": "text", + "content": "w/o ReAtten" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 69, + 293, + 282, + 383 + ], + "blocks": [ + { + "bbox": [ + 54, + 222, + 555, + 289 + ], + "lines": [ + { + "bbox": [ + 54, + 222, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 54, + 222, + 555, + 289 + ], + "type": "text", + "content": "Figure 5. Visualization of the register-to-image attention map. We randomly select visual registers and extract the attention scores that each register assigns to image tokens. These scores are then visualized as heatmaps overlaid on the input image. The results in (b) show that each visual register focuses on specific parts of the image, capturing rich visual patterns. Meanwhile, registers pay minimal attention to background areas, effectively avoiding the inclusion of redundancy. We also conduct qualitative ablation studies of the ReAtten module. As shown in (c), in the ViT model without ReAtten, the attention patterns across different sub-images appear extremely fragmented. In contrast, the ViT model with ReAtten shows continuous attention patterns, indicating effective information interaction between sub-images." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 69, + 293, + 282, + 383 + ], + "lines": [ + { + "bbox": [ + 69, + 293, + 282, + 383 + ], + "spans": [ + { + "bbox": [ + 69, + 293, + 282, + 383 + ], + "type": "image", + "image_path": "813b26bb016597ba79d2f958654b8d68b721c43981cec7fa7c93c4039fc1a5c1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 393, + 295, + 459 + ], + "lines": [ + { + "bbox": [ + 55, + 393, + 295, + 459 + ], + "spans": [ + { + "bbox": [ + 55, + 393, + 295, + 459 + ], + "type": "text", + "content": "Figure 6. Ablation on different visual register counts. Increasing the number of visual registers generally enhances accuracy across both benchmarks. However, the improvement rate noticeably tapers off when the register count exceeds 64. Based on this observation, to better balance the efficiency and performance, we select register counts of 64 for FALCON." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 461, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 295, + 521 + ], + "type": "text", + "content": "argue that the information exchange in W-Atten is limited to a local window, which hinders sufficient information exchange among all sub-images. Meanwhile, CIP introduces further redundancy by repeatedly encoding the same visual content, which can even lead to a decline in performance." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": "Ablation on the Number of Visual Registers. To investigate how visual register count affects model performance, we conduct experiments with visual register counts set to 36, 64, and 144. These counts correspond to visual token compression rates of " + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "inline_equation", + "content": "16 \\times" + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "inline_equation", + "content": "9 \\times" + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": ", respectively. As shown in Fig. 6, increasing the number of visual registers generally enhances model's high-resolution comprehension capabilities and reduces hallucinations. However, when the register count exceeds 64, the improvement gains start tapering off noticeably. Meanwhile, increasing visual registers demands more computational resources. Compared to the baseline model with 36 registers, models configured with 64 and 144 registers require 1.15 and 2.41 times more training time, respectively. Base on these observations, to achieve a better balance between efficiency and performance, we select 64 as the register counts for FALCON." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 293, + 432, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 432, + 306 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 432, + 306 + ], + "type": "text", + "content": "4.3. Qualitative Analysis" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 310, + 555, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 310, + 555, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 310, + 555, + 550 + ], + "type": "text", + "content": "To qualitatively analyze the proposed visual register technique, we visualize the register-to-image attention maps within ViT. Specifically, for a given visual register, we extract the attention scores it assigns to image tokens in each sub-image and concatenate them to construct a complete heatmap. As shown in Fig. 5(b), each visual register focuses on relatively distinct areas of the image, highlighting that it captures a diverse range of visual details. Moreover, the visual registers tend to avoid background areas, indicating their effectiveness in identifying and discarding visual redundancy. These properties enable visual registers to effectively produce a comprehensive and non-redundant visual representation. Additionally, to qualitatively evaluate the effectiveness of the ReAtten module, we visualize the attention maps within ViT models with and without ReAtten. As shown in Fig. 5(c), in the ViT model without ReAtten, the attention patterns across different sub-images appear highly fragmented. In contrast, the ViT model with ReAtten exhibits continuous attention patterns, indicating effective information exchange between sub-images." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 561, + 388, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 561, + 388, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 561, + 388, + 574 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": "To address visual redundancy and fragmentation in high-resolution MLLMs, we propose FALCON. FALCON employs an innovative visual register technique that simultaneously addresses both challenges. This technique uses a Re-Compact mechanism to adaptively aggregate essential visual information through visual registers, creating a compact, non-redundant representation. Additionally, a ReAtten module is introduced to further enhances visual continuity by enabling information exchange among sub-images via visual registers. Extensive experiments demonstrate FALCON's superiority in high-resolution understanding." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23537" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "type": "text", + "content": "6. Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 199 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 199 + ], + "type": "text", + "content": "This study is supported by National Natural Science Foundation of China (Grant No. 62306090, No. U24A20328), Natural Science Foundation of Guangdong Province of China (Grant No. 2024A1515010147), Shenzhen Science and Technology Program (KQTD20240729102207002), Natural Science Foundation of Beijing, China (Grant No. L254018), Jiangsu Science and Technology Major Program (Grant No. BG2024041) and China Postdoctoral Science Foundation (Grant No. 2024M764192)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 212, + 115, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 212, + 115, + 224 + ], + "spans": [ + { + "bbox": [ + 56, + 212, + 115, + 224 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 232, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 61, + 232, + 296, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 232, + 296, + 287 + ], + "spans": [ + { + "bbox": [ + 61, + 232, + 296, + 287 + ], + "type": "text", + "content": "[1] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 1(2):3, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 288, + 296, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 288, + 296, + 343 + ], + "spans": [ + { + "bbox": [ + 61, + 288, + 296, + 343 + ], + "type": "text", + "content": "[2] Gongwei Chen, Leyang Shen, Rui Shao, Xiang Deng, and Liqiang Nie. Lion: Empowering multimodal large language model with dual-level visual knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26540-26550, 2024. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 344, + 296, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 344, + 296, + 409 + ], + "spans": [ + { + "bbox": [ + 62, + 344, + 296, + 409 + ], + "type": "text", + "content": "[3] Gongwei Chen, Xurui Zhou, Rui Shao, Yibo Lyu, Kaiwen Zhou, Wentao Li, Yinchuan Li, Zhongang Qi, and Liqiang Nie. Less is more: Empowering gui agent with context-aware simplification. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2025. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 411, + 296, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 411, + 296, + 477 + ], + "spans": [ + { + "bbox": [ + 62, + 411, + 296, + 477 + ], + "type": "text", + "content": "[4] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 479, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 479, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 479, + 296, + 555 + ], + "type": "text", + "content": "[5] Jingxuan Chen, Derek Yuen, Bin Xie, Yuhao Yang, Gongwei Chen, Zhihao Wu, Li Yixing, Xurui Zhou, Weiwen Liu, Shuai Wang, Kaiwen Zhou, Rui Shao, Liqiang Nie, Yasheng Wang, Jianye HAO, Jun Wang, and Kun Shao. Spa-bench: A comprehensive benchmark for smartphone agent evaluation. In The Thirteenth International Conference on Learning Representations, 2025. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 557, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 557, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 557, + 296, + 612 + ], + "type": "text", + "content": "[6] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. In European Conference on Computer Vision, pages 370-387. Springer, 2024. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 613, + 296, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 613, + 296, + 679 + ], + "spans": [ + { + "bbox": [ + 62, + 613, + 296, + 679 + ], + "type": "text", + "content": "[7] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internv1: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24185–24198, 2024. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 681, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 681, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 62, + 681, + 296, + 714 + ], + "type": "text", + "content": "[8] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "text", + "content": "chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "text", + "content": " chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2(3):6, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 320, + 107, + 554, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 107, + 554, + 173 + ], + "spans": [ + { + "bbox": [ + 320, + 107, + 554, + 173 + ], + "type": "text", + "content": "[9] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1, 2, 3, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 175, + 554, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 554, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 554, + 217 + ], + "type": "text", + "content": "[10] Timothee Darcet, Maxime Aquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. In The Twelfth International Conference on Learning Representations, 2024. 2, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 220, + 554, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 554, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 554, + 296 + ], + "type": "text", + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 298, + 554, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 554, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 554, + 352 + ], + "type": "text", + "content": "[12] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024.5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 354, + 553, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 388 + ], + "type": "text", + "content": "[13] Dan Hendrycks and Kevin Gimpel. Bridging nonlinearities and stochastic regularizers with gaussian error linear units. arXiv preprint arXiv:1606.08415, 2016. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 388, + 554, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 554, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 554, + 442 + ], + "type": "text", + "content": "[14] Wenyi Hong, Weihan Wang, Ming Ding, Wenmeng Yu, Qingsong Lv, Yan Wang, Yean Cheng, Shiyu Huang, Junhui Ji, Zhao Xue, et al. Cogvlm2: Visual language models for image and video understanding. arXiv preprint arXiv:2408.16500, 2024. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 445, + 554, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 554, + 510 + ], + "type": "text", + "content": "[15] Anwen Hu, Haiyang Xu, Jiabo Ye, Ming Yan, Liang Zhang, Bo Zhang, Ji Zhang, Qin Jin, Fei Huang, and Jingren Zhou. mplug-docowl 1.5: Unified structure learning forOCR-free document understanding. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 3096-3120, 2024. 1, 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 512, + 554, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 512, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 512, + 554, + 567 + ], + "type": "text", + "content": "[16] Mingxin Huang, Yuliang Liu, Dingkang Liang, Lianwen Jin, and Xiang Bai. Mini-monkey: Alleviating the semantic saw-tooth effect for lightweight MLLMs via complementary image pyramid. In The Thirteenth International Conference on Learning Representations, 2025. 3, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 568, + 554, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 568, + 554, + 634 + ], + "spans": [ + { + "bbox": [ + 317, + 568, + 554, + 634 + ], + "type": "text", + "content": "[17] Runhui Huang, Xinpeng Ding, Chunwei Wang, Jianhua Han, Yulong Liu, Hengshuang Zhao, Hang Xu, Lu Hou, Wei Zhang, and Xiaodan Liang. Hires-llava: Restoring fragmentation input in high-resolution large vision-language models. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 29814–29824, 2025. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 635, + 554, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 554, + 690 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 554, + 690 + ], + "type": "text", + "content": "[18] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan P Foster, Pannag R Sanketi, Quan Vuong, et al. Openvla: An open-source vision-language-action model. In 8th Annual Conference on Robot Learning, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 691, + 554, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 554, + 714 + ], + "type": "text", + "content": "[19] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking mul" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23538" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "text", + "content": "timodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 148 + ], + "type": "text", + "content": "[20] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. LLaVA-onevision: Easy visual task transfer. Transactions on Machine Learning Research, 2025. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 296, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 296, + 205 + ], + "type": "text", + "content": "[21] Hao Li, Qi Lv, Rui Shao, Xiang Deng, Yinchuan Li, Jianye HAO, and Liqiang Nie. STAR: Learning diverse robot skill abstractions through rotation-augmented vector quantization. In Forty-second International Conference on Machine Learning, 2025. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 296, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 296, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 296, + 259 + ], + "type": "text", + "content": "[22] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 296, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 296, + 314 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 296, + 314 + ], + "type": "text", + "content": "[23] Junyan Li, Delin Chen, Tianle Cai, Peihao Chen, Yining Hong, Zhenfang Chen, Yikang Shen, and Chuang Gan. Flex-attention for efficient high-resolution vision-language models. In European Conference on Computer Vision, pages 286–302, 2024. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 316, + 295, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 295, + 360 + ], + "type": "text", + "content": "[24] Wei Li, Bing Hu, Rui Shao, Leyang Shen, and Liqiang Nie. Lion-fs: Fast & slow video-language thinker as online video assistant. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 3240-3251, 2025. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 361, + 295, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 295, + 404 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 295, + 404 + ], + "type": "text", + "content": "[25] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In The 2023 Conference on Empirical Methods in Natural Language Processing, 2023. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 405, + 295, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 405, + 295, + 458 + ], + "spans": [ + { + "bbox": [ + 56, + 405, + 295, + 458 + ], + "type": "text", + "content": "[26] Yanwei Li, Yuechen Zhang, Chengyao Wang, Zhisheng Zhong, Yixin Chen, Ruihang Chu, Shaoteng Liu, and Jiaya Jia. Mini-gemini: Mining the potential of multi-modality vision language models. arXiv preprint arXiv:2403.18814, 2024. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 460, + 295, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 295, + 503 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 295, + 503 + ], + "type": "text", + "content": "[27] Zaijing Li, Yuquan Xie, Rui Shao, Gongwei Chen, Dongmei Jiang, and Liqiang Nie. Optimus-1: Hybrid multimodal memory empowered agents excel in long-horizon tasks. In Advances in neural information processing systems, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 505, + 295, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 505, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 56, + 505, + 295, + 569 + ], + "type": "text", + "content": "[28] Zhang Li, Biao Yang, Qiang Liu, Zhiyin Ma, Shuo Zhang, Jingxu Yang, Yabo Sun, Yuliang Liu, and Xiang Bai. Monkey: Image resolution and text label are important things for large multi-modal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26763-26773, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 571, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 571, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 571, + 295, + 624 + ], + "type": "text", + "content": "[29] Zaijing Li, Yuquan Xie, Rui Shao, Gongwei Chen, Dongmei Jiang, and Liqiang Nie. Optimus-2: Multimodal minercraft agent with goal-observation-action conditioned policy. In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2025. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 625, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 295, + 678 + ], + "type": "text", + "content": "[30] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "type": "text", + "content": "[31] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,ocr,and world knowledge,2024.6,7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[32] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 149 + ], + "type": "text", + "content": "[33] Yuliang Liu, Biao Yang, Qiang Liu, Zhang Li, Zhiyin Ma, Shuo Zhang, and Xiang Bai. Textmonkey: AnOCR-free large multimodal model for understanding document. arXiv preprint arXiv:2403.04473, 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 151, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 553, + 205 + ], + "type": "text", + "content": "[34] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European Conference on Computer Vision, pages 216-233. Springer, 2025. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 206, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 259 + ], + "type": "text", + "content": "[35] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 3, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 262, + 553, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 553, + 314 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 553, + 314 + ], + "type": "text", + "content": "[36] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Yaofeng Sun, et al. Deepseek-vl: towards real-world vision-language understanding. arXiv preprint arXiv:2403.05525, 2024. 3, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 316, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 381 + ], + "type": "text", + "content": "[37] Pan Lu, Swaroop Mishra, Tony Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In The 36th Conference on Neural Information Processing Systems (NeurIPS), 2022. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "type": "text", + "content": "[38] Gen Luo, Yiyi Zhou, Yuxin Zhang, Xiawu Zheng, Xiaoshuai Sun, and Rongrong Ji. Feast your eyes: Mixture-of-resolution adaptation for multimodal large language models. arXiv preprint arXiv:2403.03003, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 427, + 555, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 427, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 427, + 555, + 491 + ], + "type": "text", + "content": "[39] Qi Lv, Hao Li, Xiang Deng, Rui Shao, Yinchuan Li, Jianye Hao, Longxiang Gao, Michael Yu Wang, and Liqiang Nie. Spatial-temporal graph diffusion policy with kinematic modeling for bimanual robotic manipulation. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 17394-17404, 2025. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 493, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 553, + 547 + ], + "type": "text", + "content": "[40] Yibo Lyu, Rui Shao, Gongwei Chen, Yijie Zhu, Weili Guan, and Liqiang Nie. Puma: Layer-pruned language model for efficient unified multimodal retrieval with modality-adaptive learning. In Proceedings of the 33nd ACM International Conference on Multimedia, 2025. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 548, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 602 + ], + "type": "text", + "content": "[41] Ahmed Masry, Xuan Long Do, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. In Findings of the Association for Computational Linguistics: ACL 2022, pages 2263-2279, 2022. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 604, + 555, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 604, + 555, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 604, + 555, + 646 + ], + "type": "text", + "content": "[42] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "text", + "content": "[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 4" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23539" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "text", + "content": "[44] Rui Shao, Xiangyuan Lan, Jiawei Li, and Pong C Yuen. Multi-adversarial discriminative deep domain generalization for face presentation attack detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10023-10031, 2019. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 294, + 172 + ], + "type": "text", + "content": "[45] Rui Shao, Tianxing Wu, and Ziwei Liu. Detecting and grounding multi-modal media manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6904-6913, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 294, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 294, + 216 + ], + "type": "text", + "content": "[46] Rui Shao, Tianxing Wu, Jianlong Wu, Liqiang Nie, and Zwei Liu. Detecting and grounding multi-modal media manipulation and beyond. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 216, + 294, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 216, + 294, + 261 + ], + "spans": [ + { + "bbox": [ + 56, + 216, + 294, + 261 + ], + "type": "text", + "content": "[47] Leyang Shen, Gongwei Chen, Rui Shao, Weili Guan, and Liqiang Nie. Mone: Mixture of multimodal experts for generalist multimodal large language models. In Advances in neural information processing systems, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 304 + ], + "type": "text", + "content": "[48] Baifeng Shi, Ziyang Wu, Maolin Mao, Xin Wang, and Trevor Darrell. When do we not need larger vision models? In European Conference on Computer Vision, pages 444-462, 2024. 2, 3, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 304, + 294, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 294, + 360 + ], + "type": "text", + "content": "[49] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 360, + 294, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 294, + 414 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 294, + 414 + ], + "type": "text", + "content": "[50] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 415, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 470 + ], + "type": "text", + "content": "[51] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "type": "text", + "content": "[52] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 525, + 294, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 294, + 570 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 294, + 570 + ], + "type": "text", + "content": "[53] Penghao Wu and Saining Xie. V*: Guided visual search as a core mechanism in multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13084-13094, 2024. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 570, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 294, + 624 + ], + "type": "text", + "content": "[54] Bin Xie, Rui Shao, Gongwei Chen, Kaiwen Zhou, Yinchuan Li, Jie Liu, Min Zhang, and Liqiang Nie. Gui-explorer: Autonomous exploration and mining of transition-aware knowledge for gui agent. In Annual Meeting of the Association for Computational Linguistics (ACL), 2025. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 669 + ], + "type": "text", + "content": "[55] Linli Yao, Lei Li, Shuhuai Ren, Lean Wang, Yuanxin Liu, Xu Sun, and Lu Hou. Deco: Decoupling token compression from semantic abstraction in multimodal large language models. arXiv preprint arXiv:2405.20985, 2024. 2, 3, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "type": "text", + "content": "[56] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 708 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "text", + "content": "[57] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Yuhao Dan, Chenlin Zhao, Guohai Xu, Chenliang Li, Junfeng Tian, et al. mplug-docowl: Modularized multimodal large language model for document understanding. arXiv preprint arXiv:2307.02499, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 195 + ], + "type": "text", + "content": "[58] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Guohai Xu, Chenliang Li, Junfeng Tian, Qi Qian, Ji Zhang, et al. Ureader: UniversalOCR-free visually-situated language understanding with multimodal large language model. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 2841-2858, 2023. 1, 2, 3, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 195, + 553, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 250 + ], + "type": "text", + "content": "[59] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 251, + 553, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 317 + ], + "type": "text", + "content": "[60] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Anwen Hu, Haowei Liu, Qi Qian, Ji Zhang, and Fei Huang. mplug-owl2: Revolutionizing multi-modal large language model with modality collaboration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13040–13051, 2024. 1, 2, 3, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 318, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 318, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 318, + 553, + 363 + ], + "type": "text", + "content": "[61] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In European Conference on Computer Vision (ECCV), pages 69-85. Springer, 2016. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 363, + 553, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 363, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 316, + 363, + 553, + 407 + ], + "type": "text", + "content": "[62] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 3, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 407, + 553, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 407, + 553, + 462 + ], + "spans": [ + { + "bbox": [ + 316, + 407, + 553, + 462 + ], + "type": "text", + "content": "[63] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, et al. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output. arXiv preprint arXiv:2407.03320, 2024. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 463, + 553, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 463, + 553, + 507 + ], + "spans": [ + { + "bbox": [ + 316, + 463, + 553, + 507 + ], + "type": "text", + "content": "[64] Renshan Zhang, Yibo Lyu, Rui Shao, Gongwei Chen, Weili Guan, and Liqiang Nie. Token-level correlation-guided compression for efficient multimodal document understanding. arXiv preprint arXiv:2407.14439, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 508, + 553, + 584 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 508, + 553, + 584 + ], + "spans": [ + { + "bbox": [ + 316, + 508, + 553, + 584 + ], + "type": "text", + "content": "[65] Xiaofeng Zhang, Yihao Quan, Chen Shen, Xiaosong Yuan, Shaotian Yan, Liang Xie, Wenxiao Wang, Chaochen Gu, Hao Tang, and Jieping Ye. From redundancy to relevance: Enhancing explainability in multimodal large language models. Annual Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics, 2025. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 585, + 553, + 662 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 585, + 553, + 662 + ], + "spans": [ + { + "bbox": [ + 316, + 585, + 553, + 662 + ], + "type": "text", + "content": "[66] YiFan Zhang, Huanyu Zhang, Haochen Tian, Chaoyou Fu, Shuangqing Zhang, Junfei Wu, Feng Li, Kun Wang, Qing-song Wen, Zhang Zhang, Liang Wang, and Rong Jin. MMErealworld: Could your multimodal LLM challenge high-resolution real-world scenarios that are difficult for humans? In The Thirteenth International Conference on Learning Representations, 2025. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 663, + 553, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 663, + 553, + 708 + ], + "spans": [ + { + "bbox": [ + 316, + 663, + 553, + 708 + ], + "type": "text", + "content": "[67] Yi-Fan Zhang, Qingsong Wen, Chaoyou Fu, Xue Wang, Zhang Zhang, Liang Wang, and Rong Jin. Beyond llava-hd: Diving into high-resolution large multimodal models. arXiv preprint arXiv:2406.08487, 2024. 1, 3, 7" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23540" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/FA_ Forced Prompt Learning of Vision-Language Models for Out-of-Distribution Detection/8ae5a438-b943-4e71-b751-475d9f24c3b7_content_list.json b/2025/FA_ Forced Prompt Learning of Vision-Language Models for Out-of-Distribution Detection/8ae5a438-b943-4e71-b751-475d9f24c3b7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..56c69ba980ec5f83d070c90c6c1d43637fcf8c9f --- /dev/null +++ b/2025/FA_ Forced Prompt Learning of Vision-Language Models for Out-of-Distribution Detection/8ae5a438-b943-4e71-b751-475d9f24c3b7_content_list.json @@ -0,0 +1,1365 @@ +[ + { + "type": "text", + "text": "FA: Forced Prompt Learning of Vision-Language Models for Out-of-Distribution Detection", + "text_level": 1, + "bbox": [ + 192, + 130, + 805, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xinhua Lu $^{1,2,4}$ , Runhe Lai $^{1,2,4}$ , Yanqi Wu $^{1,2,4}$ , Kanghao Chen $^{3\\dagger}$ , Wei-Shi Zheng $^{1,2,4}$ , Ruixuan Wang $^{1,2,4\\dagger}$ $^{1}$ Sun Yat-sen University $^{2}$ Peng Cheng Laboratory \n $^{3}$ Hong Kong University of Science and Technology (Guangzhou) \n $^{4}$ Key Laboratory of Machine Intelligence and Advanced Computing", + "bbox": [ + 99, + 200, + 906, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{luxh55, lairh5, wuyq268}@mail2.sysu.edu.cn, kchen879@connect.hkust-gz.edu.cn, wszheng@ieee.org, wangruix5@mail.sysu.edu.cn", + "bbox": [ + 158, + 277, + 844, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 344, + 326, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pre-trained vision-language models (VLMs) have advanced out-of-distribution (OOD) detection recently. However, existing CLIP-based methods often focus on learning OOD-related knowledge to improve OOD detection, showing limited generalization or reliance on external largescale auxiliary datasets. In this study, instead of delving into the intricate OOD-related knowledge, we propose an innovative CLIP-based framework based on Forced prompt leArning (FA), designed to make full use of the InDistribution (ID) knowledge and ultimately boost the effectiveness of OOD detection. Our key insight is to learn a prompt (i.e. forced prompt) that contains more diversified and richer descriptions of the ID classes beyond the textual semantics of class labels. Specifically, it promotes better discernment for ID images, by forcing more notable semantic similarity between ID images and the learnable forced prompt. Moreover, we introduce a forced coefficient, encouraging the forced prompt to learn more comprehensive and nuanced descriptions of the ID classes. In this way, FA is capable of achieving notable improvements in OOD detection, even when trained without any external auxiliary datasets, while maintaining an identical number of trainable parameters as CoOp. Extensive empirical evaluations confirm our method consistently outperforms current state-of-the-art methods. Code is available at https://github.com/0xFIFA/FA.", + "bbox": [ + 86, + 377, + 483, + 770 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 801, + 220, + 816 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "AI models often encounter Out-of-Distribution (OOD) samples [14, 25, 28, 38], which differ from the distribution of the training data, when deployed in real-world applications.", + "bbox": [ + 89, + 828, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d9ad327290586f4d05943d5523b636a0b2470f791793f567897944331faba407.jpg", + "image_caption": [ + "Figure 1. Comparison of CLIP-based OOD detection methods. Existing methods focus on OOD-related knowledge to learn a complex ID/OOD decision boundary as shown in the orange area in (b). In contrast, our FA aims to fully exploit ID knowledge by forcing the model to learn richer descriptions of the ID classes beyond the textual semantics of class labels. These richer descriptions of the ID classes compared to the reference text allow better discernment of ID/OOD samples (orange vs. blue area in (a))." + ], + "image_footnote": [], + "bbox": [ + 519, + 344, + 898, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Detecting OOD data is vital for the reliability of AI systems. This problem becomes harder in fields like intelligent diagnosis, where only a small amount of labeled data is available, motivating the problem of few-shot OOD detection.", + "bbox": [ + 511, + 686, + 905, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To tackle the OOD detection problem, previous single-modal post-hoc methods [14, 16, 24, 27, 41] achieve notable success based on a model pre-trained on the ID dataset. However, they come with limitations, such as requiring substantial computational resources and annotation costs for training, which hinder the performance of few-shot OOD detection. With the development of large pre-trained vision-language models (VLMs) like CLIP [37], few-shot OOD detection has achieved remarkable performance. Most of the CLIP-based OOD detection methods [30, 32] focus on", + "bbox": [ + 511, + 750, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 114, + 886, + 243, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1152", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "leveraging the powerful generalization capability of CLIP to improve OOD detection.", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent CLIP-based OOD detection methods [1, 11, 23, 33, 34, 46, 48] aim to learn OOD-related knowledge to improve OOD detection performance. For example, recent CLIP-based works [11, 46] improve OOD detection by relying on training on external large-scale auxiliary datasets, which entails significant resource overhead. To eliminate dependence on external datasets, existing methods [1, 33, 48] learn OOD-related knowledge from exposed outliers based only on ID training data, such as background regions in ID training images. However, these OOD features extracted from specific regions are difficult to match with infinite OOD data encountered in practice. Moreover, some methods [23, 34] aim to learn negative prompts semantically opposite to the ID class labels, while these limited negative prompts are often insufficient to capture the distinctions between the diverse and numerous OOD data and the ID data. Overall, improving OOD detection by learning OOD-related knowledge shows inherent limitations or requires significant computational resources as well as labor costs. Inspired by the method [44] for open-set recognition, which demonstrates that enhancing closed-set accuracy can typically improve open-set recognition capabilities, we are motivated to explore ways to improve the identification of ID images, thereby enhancing OOD detection performance.", + "bbox": [ + 91, + 125, + 483, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, rather than focusing on the intricate OOD-related knowledge, we propose a novel CLIP-based framework based on Forced prompt leArning (FA), which aims to fully exploit the ID knowledge and ultimately improve OOD detection. Our key insight is to learn a prompt that contains richer knowledge beyond the textual semantics of class labels. It facilitates advanced discernment for ID images, by forcing higher semantic similarity between ID images and the learnable prompt (i.e. forced prompt), which provides more diversified and richer descriptions of the ID classes. In this way, FA can achieve significant improvements in OOD detection performance, even when trained without any additional external data, while maintaining the same number of learnable parameters as CoOp [54].", + "bbox": [ + 89, + 505, + 483, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, we introduce a novel forced prompt along with the original prompt, both of which are initialized identically. FA preserves the generalization capability of the VLMs (i.e. CLIP) by freezing the original prompt and optimizing a trainable copy (i.e. forced prompt). In particular, the forced prompt treats the original prompt as a reference, while forcing the text features associated with the forced prompt to be more salient compared to those associated with the original prompt. In this way, it effectively improves the OOD performance even without sacrificing the ID classification capability of the model. Additionally, we introduce a forced coefficient, encouraging the forced", + "bbox": [ + 89, + 719, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "prompt to learn more comprehensive connotations. Experimentally, FA achieves superior few-shot OOD detection performance across diverse OOD benchmarks. Compared to current SOTA methods [48] that focus on learning OOD-related knowledge, our method significantly reduces the average FPR95 score from $31.62\\%$ to $27.81\\%$ and improves the average AUROC from $92.01\\%$ to $93.26\\%$ even in 1-shot OOD detection on ImageNet-1k. The main contributions are summarized as follows:", + "bbox": [ + 511, + 90, + 903, + 226 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a simple yet effective framework, which fully exploits the ID knowledge to improve OOD detection without focusing on intricate OOD-related knowledge.", + "- We present a Forced prompt leArning (FA) strategy to exploit richer knowledge beyond the textual semantics of class labels.", + "- We evaluate our method on diverse OOD benchmarks, showing that our model consistently outperforms current state-of-the-art methods." + ], + "bbox": [ + 511, + 231, + 903, + 371 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 511, + 388, + 651, + 404 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prompt Learning. In recent years, pre-trained vision-language models (VLMs) such as CLIP [37] have demonstrated powerful few-shot learning capabilities in both visual and textual domains. However, the design of the prompt greatly affects the performance of VLMs for downstream tasks, which may require manually crafting numerous prompts. Inspired by prompt learning studies in natural language processing (NLP) [26], CoOp [54] uses a set of learnable context vectors to transfer CLIP for specific downstream tasks, becoming a pioneering method for subsequent studies [4, 33, 53]. Currently, prompt learning is widely used in OOD detection mainly to represent various OOD-related knowledge. In this work, we learn the forced prompt, which contains richer descriptions of the ID classes to improve OOD detection.", + "bbox": [ + 511, + 415, + 903, + 641 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Out-of-Distribution Detection. Conventional methods explore OOD detection for single-modal models. One line of studies designs score functions to differentiate ID and OOD data based on output from the pre-trained models, such as logit output [14, 16, 24, 27] or outputs from the penultimate layer [22, 41, 42, 50]. These methods are called post-hoc methods. Another line of studies [9, 10, 18, 31, 39, 43, 55] adopts various training strategies to learn a reliable decision boundary between ID and OOD data.", + "bbox": [ + 511, + 643, + 905, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, the development of OOD detectors based on VLMs, especially CLIP, has received much attention as VLMs have demonstrated their remarkable generalization capability in both visual and textual domains. Some studies leverage real outlier information from external auxiliary OOD datasets [11, 46] or extensive corpora [5, 20] to promote OOD detection. However, this is impractical in real-world scenarios, where outliers are infinite and agnostic.", + "bbox": [ + 511, + 779, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1153", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d6c07d7693e54a2e5e7e9bf056c4470e042bca80776d2e49bb596b306ede69d4.jpg", + "image_caption": [ + "Figure 2. Overview of the proposed FA framework. Our framework includes the learnable forced prompt and the frozen original prompt, both of which are initialized identically by a manual template \"a photo of a [class-c]\". The learnable forced prompt treats the frozen original prompt as a reference, forcing its text features to become more salient compared to those generated by the original prompt. The richer and more diversified description of ID classes learned by FA can ultimately improve OOD detection." + ], + "image_footnote": [], + "bbox": [ + 174, + 93, + 815, + 334 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To eliminate the dependence on external data, recent methods [1, 23, 32, 33, 48] were developed under the assumption that only ID data are available. As a representative zero-shot method that only uses available ID labels, MCM [30] employs softmax scaling to align visual features with ID text concepts for OOD detection. GL-MCM [32] further introduces a local maximum concept matching(L-MCM) score to improve the separability of the local information. Compared with zero-shot methods, which may undergo a domain gap with ID downstream data, prompt learning methods achieve better OOD detection performance with access to few-shot ID samples. For prompt learning OOD detection methods, LoCoOp [33] and SCT [48] keep the textual embeddings of ID classes away from ID-irrelevant local region embeddings in the multi-modal embedding space by their proposed entropy maximization strategy, which ensures that OOD embeddings can be dissimilar to any textual embedding of ID classes. To detect challenging OOD samples to improve OOD detection, ID-like [1] explores the vicinity of ID samples to construct OOD samples correlated to the ID and refines the additional \"ID-like\" text embeddings to fine-grained differences. In addition, NegPrompt [23] and LSN [34] introduce the negative connotations of ID categories which can be represented by additional negative prompts, enabling more accurate detection of OOD samples. Overall, various OOD-related knowledge is commonly adopted to address the model's overconfidence in current research. In this work, we focus on leveraging ID knowledge to improve OOD detection, instead of exploring complex OOD-related knowledge.", + "bbox": [ + 91, + 409, + 483, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 407, + 651, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prompt Learning with CLIP. CLIP contains an image encoder $f(\\cdot)$ and a text encoder $g(\\cdot)$ , designed to extract features from images and text descriptions respectively. Generally, for an ID dataset which contains $C$ categories, the hand-crafted prompt $\\hat{\\mathbf{u}}_c$ \"a photo of a [class-c]\" is designed to match images, where class-c represents the class name. Formally, these prompts are individually encoded in $\\hat{\\mathbf{t}}_c = g(\\hat{\\mathbf{u}}_c) \\in \\mathbb{R}^{d \\times 1}$ , $c = 1, \\dots, C$ . Given an image $\\mathbf{x}$ , it can be encoded in $\\mathbf{z} = f(\\mathbf{x}) \\in \\mathbb{R}^{d \\times 1}$ . Because image features and corresponding text features are aligned on the multimodal embedding space, the cosine similarity $\\cos(\\mathbf{z}, \\hat{\\mathbf{t}}_c)$ of $\\mathbf{z}$ and $\\hat{\\mathbf{t}}_c$ can represent the matching degree between the image $\\mathbf{x}$ and text description $\\hat{\\mathbf{u}}_c$ . Therefore, if the text feature $\\mathbf{t}_c$ of class $c$ has the highest similarity to $\\mathbf{z}$ , the image $\\mathbf{x}$ will be considered to be class $c$ .", + "bbox": [ + 511, + 431, + 906, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To effectively transfer CLIP to downstream image recognition tasks, CoOp [54] sets a portion of tokens in the text prompt as continuous learnable parameters. Concretely, CoOp initializes the text prompts as $\\mathbf{u}_c = [\\mathbf{v}_1,\\dots ,\\mathbf{v}_L,\\mathbf{w}_c],c\\in \\{1,\\dots ,C\\}$ , where $L$ is the length of the prompt's tokens, $\\mathbf{v}_i$ ( $i\\in \\{1,\\ldots ,L\\}$ ) is the learnable vector with the same dimension as the word embedding and $\\mathbf{w}_c$ is the word embedding for the class name of class $c$ . The text prompt $\\mathbf{u}_c$ is encoded into features $\\mathbf{t}_c = g(\\mathbf{u}_c)$ . Together with the image feature $\\mathbf{z}$ , the probability for the image $\\mathbf{x}$ to be classified into class $c$ is defined as", + "bbox": [ + 511, + 659, + 908, + 825 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np (y = c \\mid \\mathbf {x}) = \\frac {e ^ {\\cos (\\mathbf {z} , \\mathbf {t} _ {c}) / \\tau}}{\\sum_ {j = 1} ^ {C} e ^ {\\cos (\\mathbf {z} , \\mathbf {t} _ {j}) / \\tau}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 835, + 906, + 875 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\tau$ is a fixed temperature scaling hyper-parameter.", + "bbox": [ + 513, + 886, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1154", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overall, the prompt's tokens can be trained to align with training data by minimizing the cross-entropy loss with the prediction probability from Eq. (1).", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "OOD Detection. In the OOD detection task, the model is expected to determine whether a test image belongs to one of the learned ID classes. Therefore, OOD detection can be seen as a binary classification problem to distinguish ID images from OOD images in a test set $\\mathcal{D}_{test}$ as follows", + "bbox": [ + 89, + 136, + 480, + 212 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD (\\mathbf {x}) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} S (\\mathbf {x}) \\geq \\mu \\\\ 0, & \\text {i f} S (\\mathbf {x}) < \\mu \\end{array} , \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 223, + 480, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{x} \\in \\mathcal{D}_{test}$ , 1 and 0 respectively indicate that $\\mathbf{x}$ is classified as the ID class and the OOD class by the OOD detector $D(\\cdot)$ , $S(\\cdot)$ is certain score function, and $\\mu$ is a predefined threshold constant.", + "bbox": [ + 89, + 268, + 482, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 89, + 344, + 225, + 361 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Overview", + "text_level": 1, + "bbox": [ + 89, + 369, + 200, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this study, we propose a novel CLIP-based framework based on Forced prompt leArning (FA), designed to fully exploit the ID knowledge from few-shot ID samples by forcing the prompt to learn richer knowledge beyond the textual semantics of class labels. As shown in Fig. 2, we introduce a forced prompt along with the original prompt, both of which are initialized with the same semantics. By freezing the original prompt and optimizing a trainable copy, ID images will show higher semantic similarity to the forced prompt compared to the original prompt. Moreover, to encourage descriptions of ID classes learned by the forced prompt to become more detailed and comprehensive, we also introduce a forced coefficient. During testing, the model is capable of effectively distinguishing between ID data and OOD data, benefiting from the diversified descriptions of the ID classes provided by the forced prompt.", + "bbox": [ + 89, + 391, + 482, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Forced prompt learning", + "text_level": 1, + "bbox": [ + 89, + 643, + 310, + 659 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Merely the semantic information conveyed by class names is insufficient to comprehensively encompass the discriminative information of each class. In order to explore richer semantic descriptions of the ID classes beyond the textual semantics of class labels, a novel forced prompt along with the original prompt is proposed. To implement the above objective, we introduce a simple yet effective training strategy using the forced cross-entropy (FCE) loss based on the forced prompt as follows", + "bbox": [ + 89, + 665, + 482, + 801 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {F C E} = \\mathbb {E} _ {\\left(\\mathbf {x}, y _ {c}\\right) \\sim \\mathcal {D} _ {t r a i n} ^ {I D}} \\left[ - \\log \\frac {e ^ {s _ {c} ^ {f} / \\tau}}{\\sum_ {j = 1} ^ {C} e ^ {s _ {j} ^ {f} / \\tau} + \\sum_ {j = 1} ^ {C} e ^ {s _ {j} ^ {o} / \\tau}} \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 811, + 480, + 867 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the ID training dataset $\\mathcal{D}_{train}^{ID}$ consists of ID image-label pairs $(\\mathbf{x},y_{c})$ , $s_j^f = \\cos (\\mathbf{z},\\mathbf{t}_j^f)$ and $s_j^o = \\cos (\\mathbf{z},\\mathbf{t}_j^o)$", + "bbox": [ + 89, + 867, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "represent the similarity between the image feature and the prompt feature corresponding to the forced prompt and the original prompt, respectively.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on this loss function, the ID image features will be forced to show higher cosine similarity to the learnable forced prompt compared to the original text prompt. This is because the cosine similarity between the ID image features and the original text prompt is already high before any further prompt learning. Consequently, the forced prompt is forced to uncover richer discriminative information of ID classes, so that the text features of the forced prompt achieve a higher cosine similarity with the image features compared to those of the original prompt. Benefiting from this information specific to the ID classes, the cosine similarity between the image features of an OOD image and the text features of both prompts may exhibit less distinction compared to those of the ID data. Therefore, even when trained without any reliance on external auxiliary datasets, the detector will have a notable capability to distinguish ID data from OOD data.", + "bbox": [ + 511, + 140, + 906, + 397 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, this naive FCE loss may not work effectively because of its inherent limitations. Concretely, employing random initialization for both prompts fails to fully leverage the semantic information provided by CLIP's prior knowledge, resulting in limited performance. To solve this limitation, we propose to initialize the forced prompt using the manual template (i.e. \"a photo of a [class-c]\") identified to the original prompt. In this way, both prompts with manual initialization will have clear semantic information compared to random initialization, thereby improving the model's generalization capability. Specifically, the class labels are utilized in both prompts by concatenating with the prompt embedding to provide foundational semantic information related to the category names, based on CLIP's prior knowledge. In particular, for the forced prompt, we adopt the shared learnable vector across all classes rather than the independent learnable vector for each class. This choice is inspired by CoOp [54], which demonstrates that using an independent learnable vector mostly underperforms the shared learnable vector in challenging low-data scenarios since the former has more parameters and requires more data for training. Formally, we utilize the embeddings of the manual template to initialize the prompt's embeddings of both prompts, formulated as $\\mathbf{u}_c = [\\mathbf{v}_1,\\dots ,\\mathbf{v}_L,\\mathbf{w}_c], c\\in$ $\\{1,\\dots ,C\\}$ , which will be fed to the text encoder to obtain prompt feature $\\mathbf{t}_c$ (see Section 3). Here, $L$ is the length of the token (e.g., for the \"a photo of a\", $L = 4$ ). Then we freeze the original prompt and the class label part of both the prompts to preserve the generalization capability of CLIP, while only making the forced prompt learnable. Note that our method effectively improves OOD detection performance while maintaining consistency with CoOp in the number of learnable parameters without leveraging ad", + "bbox": [ + 511, + 401, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1155", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ditional learnable prompts, unlike existing work [1, 23, 34].", + "bbox": [ + 89, + 90, + 480, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although the above prompt design achieves moderate performance, relying solely on a single original prompt as a reference is somewhat insufficient to develop a comprehensive capability, particularly with the limitations of ID data (e.g., few-shot OOD). To encourage the forced prompt to capture more comprehensive and nuanced descriptions of the ID classes, we introduce a forced coefficient $K$ ( $K \\geq 0, K \\in \\mathbb{N}$ ), which indicates the intensity with which the model is compelled to learn from the data. Formally, we first derive the original text features $\\{\\mathbf{t}_1^o, \\dots, \\mathbf{t}_C^o\\}$ from the original prompt and the image features $\\mathbf{z}$ from the corresponding image. Then, we compute the cosine similarity between the original text features and the image features for $K$ iterations, which will be used in the subsequent computation of the cross-entropy loss function. Notably, when $K = 0$ , no original prompt is used, and the model operates equivalently to CoOp. Based on the forced coefficient, we refine Eq. (3) as follows", + "bbox": [ + 89, + 107, + 482, + 378 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {F C E - K} = \\mathbb {E} _ {\\left(\\mathbf {x}, y _ {c}\\right) \\sim \\mathcal {D} _ {t r a i n} ^ {I D}} \\left[ - \\log \\frac {e ^ {s _ {c} ^ {f} / \\tau}}{\\sum_ {j = 1} ^ {C} e ^ {s _ {j} ^ {f} / \\tau} + K \\sum_ {j = 1} ^ {C} e ^ {s _ {j} ^ {o} / \\tau}} \\right]. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 391, + 483, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on this function, the cosine similarity between the ID image features and the text features of the forced prompt must become more salient as the coefficient $K$ increases.", + "bbox": [ + 89, + 446, + 482, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Model inference", + "text_level": 1, + "bbox": [ + 89, + 502, + 251, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During model inference, for the downstream classification task, we adopt the same strategy as CLIP, depending solely on the forced prompt [1]. For the OOD detection task, our method can be flexibly combined with different score functions, such as the MCM [30] score and the GL-MCM [32] score. The MCM score is defined as the maximum similarity between the global image features $\\mathbf{z}^g$ (i.e. $\\mathbf{z}$ ) and all text features $\\mathbf{t}_c^a$ (i.e. the concatenation of text features $\\mathbf{t}_c^f$ and $\\mathbf{t}_c^o$ ) after applying softmax with temperature $\\tau_0$ , i.e.,", + "bbox": [ + 89, + 525, + 483, + 661 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS _ {\\mathrm {M C M}} (\\mathbf {x}) = \\max _ {c} \\frac {e ^ {\\cos \\left(\\mathbf {z} ^ {g} , \\mathbf {t} _ {c} ^ {a}\\right) / \\tau_ {0}}}{\\sum_ {j = 1} ^ {C} e ^ {\\cos \\left(\\mathbf {z} ^ {g} , \\mathbf {t} _ {j} ^ {f}\\right) / \\tau_ {0}} + K e ^ {\\cos \\left(\\mathbf {z} ^ {g} , \\mathbf {t} _ {j} ^ {o}\\right) / \\tau_ {0}}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 672, + 482, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "while the GL-MCM score simultaneously considers both global and local features, which can be expressed as", + "bbox": [ + 89, + 727, + 482, + 757 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS _ {\\mathrm {G L - M C M}} (\\mathbf {x}) = S _ {\\mathrm {M C M}} (\\mathbf {x}) + S _ {\\mathrm {L - M C M}} (\\mathbf {x}), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 770, + 482, + 787 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS _ {\\mathrm {L - M C M}} (\\mathbf {x}) = \\max _ {i, c} \\frac {e ^ {\\cos \\left(\\mathbf {z} _ {i} ^ {l} , \\mathbf {t} _ {c} ^ {a}\\right) / \\tau_ {0}}}{\\sum_ {j = 1} ^ {C} e ^ {\\cos \\left(\\mathbf {z} _ {i} ^ {l} , \\mathbf {t} _ {j} ^ {f}\\right) / \\tau_ {0}} + K e ^ {\\cos \\left(\\mathbf {z} _ {i} ^ {l} , \\mathbf {t} _ {j} ^ {o}\\right) / \\tau_ {0}}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 800, + 480, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{z}_i^l (i\\in \\{1,\\dots ,N\\})$ represents $N$ extracted local features generated by CLIP's image encoder [32, 33]. We set $\\tau_0 = 1$ during model inference.", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 513, + 90, + 645, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experimental details", + "text_level": 1, + "bbox": [ + 513, + 114, + 709, + 131 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We use a popular benchmark for conventional OOD detection [19, 30, 32, 33], where ImageNet-1k [7] serves as the ID dataset, and the OOD datasets are iNaturalist [17], SUN [47], Places [52], and Texture [6]. Moreover, inspired by existing studies [1, 2], we use datasets like OpenImage-O [45], NINCO [2], and ImageNet-O [15], which are cleaner and more realistic, to simulate the more challenging OOD detection. Besides, we also utilize other widely adopted datasets for few-shot settings as ID datasets [49, 51, 54], which include StanfordCars [21], UCF101 [40], Caltech101 [12], Flowers102 [35], EuroSAT [13], FGVCAircraft [29], OxfordPets [36], and Food101 [3], considering factors such as image resolution and the number of classes.", + "bbox": [ + 511, + 138, + 906, + 348 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Setup. Following previous studies [33], we use the ViTB/16 [8] as the backbone model. For our model, when using ImageNet-1k as the ID dataset, the epochs are respectively set to 30 and 50 for the 1-shot and 16-shot settings, while for other ID datasets, we set the epoch to 200 following CoOp [54]. The value of forced coefficient $K$ is uniformly set to 3, which will be further discussed in the sensitivity study. Other hyperparameters are as follows: learning rate $= 2\\mathrm{e} - 3$ , batch size $= 160$ , SGD (momentum $= 0.9$ , weight decay $= 5\\mathrm{e} - 4$ ) as the optimizer with a cosine scheduler, $\\tau = 1$ in Eq. (3) and Eq. (4). All experiments on our model can be conducted on a single Nvidia A30 GPU. The average experiment results (including our reproductions) over four runs are reported for comparison.", + "bbox": [ + 511, + 349, + 906, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparison Methods. To validate the effectiveness of our FA fairly, we mainly compare it with CLIP-based OOD detection methods that do not use real outliers (e.g. OOD labels). For previous post-hoc methods, including MSP [14], ODIN [24], Energy [27], ReAct [41], and MaxLogit [16], we adapt these methods with the CLIP image encoder as the CLIP-based post-hoc methods. For zero-shot methods, we select MCM [30], GL-MCM [32] and CLIPN [46] as baselines. For prompt learning methods, we adopt CoOp [54], LoCoOp [33], IDLike [1], and SCT [48] as baselines.", + "bbox": [ + 511, + 563, + 906, + 714 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. For evaluation, we adopt the following metrics: (1) The False Positive Rate at $95\\%$ True Positive Rate for in-distribution samples (FPR95); (2) The Area Under the Receiver Operating Characteristic Curve (AUROC); (3) In-distribution data classification Top-1 accuracy (ID ACC).", + "bbox": [ + 511, + 714, + 903, + 790 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Main results", + "text_level": 1, + "bbox": [ + 513, + 801, + 645, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Conventional OOD Detection. Table 1 summarizes our comparison results on the ImageNet-1k benchmarks, which show that our FA achieves state-of-the-art OOD detection performance among other CLIP-based methods under different few-shot settings (More details under the 4-shot set", + "bbox": [ + 511, + 824, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1156", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/df04df2d128cd259127b9519eb43a98675da97ff4d11497a8e58d74958be8209.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodiNaturalistSUNPlacesTexturesAverage
FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑
Zero-shot methods
MCM31.9594.1637.2292.5542.9890.1058.3585.8342.6390.66
GL-MCM15.0996.7229.0893.4137.0790.3758.9483.1135.0490.90
CLIPN19.1796.1726.4394.0232.2692.6241.2390.1230.2193.19
CLIP-based post-hoc methods
\\( MSP^† \\)74.5777.7476.9573.9779.7272.1873.6674.8476.2274.68
\\( ODIN^† \\)98.9357.7388.7278.4287.8076.8885.4771.4990.2371.13
\\( Energy^† \\)64.9887.1846.4291.1757.4087.3350.3988.2254.8088.48
\\( ReAct^† \\)65.5786.8746.1791.0456.8587.4249.8888.1354.6288.37
\\( MaxLogit^† \\)60.8888.0344.8391.1655.5487.4548.7288.6352.4988.82
Prompt learning based methods
\\( CoOp_{MCM} \\)41.14±9.3991.47±2.1239.06±4.1791.68±0.8345.38±4.8889.16±1.2851.37±3.0987.82±1.1244.24±1.3990.03±0.32
\\( CoOp_{GL} \\)23.30±6.3794.59±1.5132.08±3.8792.21±0.9539.22±4.5589.59±1.5855.78±3.4683.26±1.3837.59±0.3489.91±0.31
\\( LoCoOp_{MCM} \\)36.64±4.8792.78±0.9331.86±3.7993.50±0.9238.81±3.3790.70±0.9748.11±2.3189.43±0.6838.85±2.6791.59±0.58
\\( LoCoOp_{GL} \\)21.97±2.9195.39±0.5924.95±2.3794.42±0.5934.14±2.6691.14±0.6949.04±2.7787.73±0.9632.53±2.1992.17±0.47
IDLike17.73±1.9196.68±0.3148.17±1.3989.53±0.5850.43±6.5888.27±2.2729.12±7.6493.25±2.1636.36±3.8691.93±1.18
\\( LSN^† \\)59.28±7.0287.20±3.1540.15±0.8291.47±0.1446.11±1.8688.74±0.5760.34±0.1483.92±0.4251.47±1.5387.84±0.58
\\( NegPrompt^† \\)65.03±8.6984.56±2.5244.39±1.6689.63±0.6651.31±6.2186.55±2.1987.60±1.6163.76±3.0262.08±3.7181.13±1.78
\\( SCT_{MCM} \\)41.93±12.1791.77±2.4030.39±2.1793.76±0.3938.73±2.5490.78±0.3446.78±3.5188.96±1.1639.46±3.3991.32±0.90
\\( SCT_{GL} \\)20.57±10.295.63±1.8924.56±3.0394.39±0.5333.27±2.9691.27±0.5748.12±2.9786.76±0.8931.62±3.1992.01±0.77
\\( FA_{MCM}(Ours) \\)25.50±2.7294.72±0.2736.24±3.1592.40±0.9535.38±2.7591.99±0.6828.34±0.8793.95±0.1931.37±1.0793.25±0.32
\\( FA_{GL}(Ours) \\)14.12±1.3296.76±0.1029.99±1.5792.95±0.6632.48±1.4891.83±0.4934.66±1.2191.50±0.3627.81±0.4493.26±0.27
Prompt learning based methods
\\( CoOp_{MCM} \\)30.26±1.9893.43±0.8134.69±0.4392.59±0.1441.91±0.7190.11±0.2344.68±2.1189.95±0.4837.89±0.7191.52±0.29
\\( CoOp_{GL} \\)15.96±1.6796.11±0.5527.26±1.9993.29±0.4635.36±2.0890.58±0.6448.63±2.1186.11±0.5931.81±1.2791.51±0.39
\\( LoCoOp_{MCM} \\)27.35±3.1994.12±0.8030.93±1.1993.75±0.2638.26±1.5391.12±0.2741.36±2.5690.99±0.5334.47±0.7392.49±0.14
\\( LoCoOp_{GL} \\)18.46±1.3895.85±0.6322.43±0.9695.15±0.2231.53±1.5292.15±0.2243.35±3.1289.38±0.7828.94±1.2993.13±0.17
IDLike19.23±10.596.70±1.5854.15±2.8887.64±1.1956.63±0.0785.86±0.4434.69±6.4191.90±2.3141.18±1.7390.53±0.01
\\( LSN^† \\)36.17±4.8192.66±1.1634.27±0.4493.53±0.2041.47±0.8590.52±0.3746.43±0.6089.38±0.2439.58±0.7391.53±0.09
\\( NegPrompt^† \\)37.79±0.1190.49±0.0132.11±3.7792.25±1.0035.52±0.4191.16±0.0343.93±9.0988.38±3.3137.34±1.4190.57±0.59
\\( SCT_{MCM} \\)29.41±2.1993.76±0.5527.28±2.8094.22±0.4036.35±2.1391.16±0.3342.25±1.8990.52±0.4833.82±1.7892.42±0.35
\\( SCT_{GL} \\)15.19±2.1696.71±0.4420.00±0.6195.57±0.1129.71±0.8592.37±0.1244.17±0.8288.59±0.3927.27±0.4493.31±0.17
\\( FA_{MCM}(Ours) \\)25.79±1.4894.29±0.3533.54±1.1793.02±0.1933.77±1.6492.64±0.4223.17±1.3195.14±0.2629.07±1.1193.77±0.19
\\( FA_{GL}(Ours) \\)14.49±1.2796.48±0.2927.65±1.0893.46±0.1831.09±1.3892.44±0.3429.50±0.8992.93±0.1825.68±0.5893.82±0.11
", + "bbox": [ + 89, + 87, + 906, + 575 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/455dd344dd79e267e82084b1808e48db370c8956b7fe646176768df80503b981.jpg", + "table_caption": [ + "Table 1. Comparison results on ImageNet-1k OOD benchmarks. All results using the same backbone ViT-B/16. The results marked with $\\dagger$ are taken from [48]; the others are our reproductions. The prompt learning based methods are run under four trials, reporting the mean and standard deviation of the performance. The subscripts ${}_{\\mathrm{{MCM}}}$ and ${}_{\\mathrm{{GL}}}$ indicate the use of the MCM score and the GL-MCM score. The best and second-best results are indicated in bold and underline. $\\uparrow$ indicates larger values are better; $\\downarrow$ indicates smaller values are better. All values are percentages." + ], + "table_footnote": [], + "table_body": "
Method1-shot4-shot16-shot
CoOp67.44±0.5069.71±0.0770.99±0.14
LoCoOp67.40±0.6469.55±0.1071.53±0.17
IDLike68.17±0.5768.91±0.1469.46±0.02
SCT68.63±0.1369.93±0.2271.78±0.05
FA(Ours)68.67±0.3969.96±0.0471.02±0.09
", + "bbox": [ + 117, + 670, + 455, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparison results in ID Top-1 accuracy on ImageNet-1k under different few-shot settings. Other notations are the same as Tab. 1.", + "bbox": [ + 89, + 784, + 482, + 825 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ting are provided in the Appendix A). Specifically, in the 16-shot setting, FA with GL-MCM score $\\left(\\mathrm{FA}_{\\mathrm{GL}}\\right)$ outperforms the best baseline $\\mathrm{SCT}_{\\mathrm{GL}}$ (average FPR95 of $25.68\\%$ vs. $27.27\\%$ , and average AUROC of $93.82\\%$ vs. $93.31\\%$ ).", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Additionally, in the 1-shot setting, $\\mathrm{FA}_{\\mathrm{GL}}$ surpasses $\\mathrm{SCT}_{\\mathrm{GL}}$ by a large margin, showing improvements of at least $1.25\\%$ and $3.81\\%$ on average AUROC and FPR95. More importantly, even the average score of $\\mathrm{FA}_{\\mathrm{MCM}}$ has exceeded $\\mathrm{SCT}_{\\mathrm{GL}}$ in the 1-shot setting. Moreover, our FA effectively improves the OOD performance without sacrificing the ID classification capability of the model, as shown in Tab. 2.", + "bbox": [ + 511, + 672, + 906, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In particular, we can also observe that our method performs not well on the SUN dataset, as shown in Tab. 1. This is related to the presence of a significant number of samples belonging to ID category in these datasets, as demonstrated by recent studies [1]. In other words, the SUN dataset in the popular OOD benchmark requires more thorough cleaning to be effectively used as an OOD dataset.", + "bbox": [ + 511, + 779, + 908, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Challenging OOD Detection. Recent studies [1, 2] have", + "bbox": [ + 511, + 883, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1157", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/686b772dca172fa9cee16de40529ad055d9d0de180ebf534e139582902000e4e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodOpenImage-ONINCOImageNet-OAverage
FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑
CoOpMCM37.95±0.5991.99±0.2278.00±0.2373.00±0.5970.14±0.8681.34±0.8362.03±0.3882.11±0.48
CoOpGL32.23±1.0092.22±0.2572.04±0.8175.35±0.8867.35±1.1978.44±0.9457.21±0.9681.99±0.58
LoCoOpMCM37.03±0.9492.24±0.2177.96±0.6372.78±0.7970.96±0.7581.25±0.3961.98±0.5582.09±0.35
LoCoOpGL33.87±1.2392.53±0.0775.16±0.9272.97±0.3968.06±1.9281.19±0.2559.03±1.2782.23±0.17
IDLike54.43±2.3587.89±0.4578.93±1.7969.32±0.4284.08±1.4567.45±0.8672.48±0.8974.89±0.28
SCTMCM37.28±1.1692.04±0.3378.51±0.8671.09±1.1971.45±0.5481.29±0.1562.41±0.7381.48±0.52
SCTGL32.24±0.3492.75±0.1774.15±0.5373.14±0.9468.35±0.6280.84±0.4558.25±0.3982.24±0.44
FAmCM(Ours)37.93±1.6591.84±0.4570.44±1.2477.77±0.9463.84±1.3682.71±0.3057.40±1.3484.08±0.47
FAGL(Ours)32.10±0.8992.39±0.2865.61±1.3279.01±0.7163.13±0.9880.39±0.2153.61±0.8983.93±0.31
", + "bbox": [ + 91, + 88, + 906, + 237 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c89bb2e0ccfc86f226a8bd042677dedbc56597f4d2e5952389cc4f3c7082576e.jpg", + "table_caption": [ + "Table 3. Challenging OOD detection results on cleaner OOD datasets in the 16-shot setting. We use the same notation as Tab. 1." + ], + "table_footnote": [], + "table_body": "
ID DatasetMethodiNaturalistSUNPlacesTexturesAverage
FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑
UCF101CoOpMCM14.40±6.6197.41±1.3637.09±3.9493.14±0.4840.07±2.7791.89±0.6533.24±4.2193.19±1.2231.21±3.5893.91±0.45
CoOpGL9.65±6.3598.18±0.9828.54±5.0794.51±0.9232.94±1.7493.28±0.3235.96±2.6591.83±1.4226.77±3.2294.45±0.42
LoCoOpMCM2.38±0.9899.21±0.2219.73±4.2496.11±0.7822.57±3.0194.96±0.7715.89±3.1296.12±0.6715.14±1.9196.60±0.33
LoCoOpGL1.13±0.2299.65±0.0917.93±4.0996.51±0.8922.34±4.5895.03±1.0618.17±3.0195.25±0.7114.89±2.4996.61±0.58
IDLike66.63±27.689.75±5.5287.88±6.5678.04±6.1484.47±12.176.04±8.5349.07±11.291.19±2.7472.01±14.083.76±5.64
SCTMCM1.51±0.5199.44±0.1024.17±5.5495.21±0.8826.62±4.4193.75±0.8115.34±3.8495.98±0.6416.91±3.2096.09±0.48
SCTGL0.79±0.2799.75±0.0418.69±2.4195.86±0.5623.77±2.4793.98±0.4717.44±2.6294.52±0.5415.17±1.5696.03±0.25
FAmCM(Ours)0.12±0.1199.94±0.013.06±1.3399.34±0.234.68±1.0099.08±0.262.68±0.5199.31±0.432.63±0.5899.42±0.22
FAGL(Ours)0.13±0.0299.94±0.024.98±1.0299.01±0.227.46±0.8698.44±0.275.67±0.8398.75±0.194.56±0.4799.03±0.13
EuroSATCoOpMCM99.14±0.7331.95±9.1798.29±0.4944.36±4.1997.28±0.8547.51±4.6386.69±4.3769.17±3.1595.35±1.0348.25±3.09
CoOpGL92.45±4.2657.59±10.490.97±3.4961.89±4.8388.69±3.8564.00±4.9262.37±1.9383.58±2.3183.62±1.1866.77±2.83
LoCoOpMCM96.51±3.3353.49±8.0693.52±5.6855.59±9.1992.88±4.0557.89±5.1976.45±7.4878.59±2.6689.84±4.2961.39±3.90
LoCoOpGL86.13±12.272.68±7.9787.81±6.3768.58±6.4085.93±4.1569.90±3.4054.89±4.6787.00±1.0178.69±4.6574.54±3.59
IDLike98.02±2.1373.32±5.8395.74±2.8465.19±6.8694.79±3.6769.44±5.0663.34±9.1086.68±2.9887.97±4.2173.66±5.04
SCTMCM99.79±0.2232.90±6.7398.36±0.8444.30±3.4197.77±0.8648.08±1.6483.98±6.9273.25±3.5894.97±2.1949.63±2.39
SCTGL96.57±5.4158.21±5.3192.41±4.5561.54±3.4790.49±5.0263.33±2.7155.29±14.785.48±3.6883.69±7.0267.14±2.20
FAmCM(Ours)85.24±6.8878.62±3.1836.24±10.892.13±2.4428.17±10.294.05±1.6810.36±3.9797.71±0.6939.99±7.7990.63±1.91
FAGL(Ours)77.82±7.2886.88±2.1556.10±8.0388.43±2.2649.87±7.6989.07±1.498.99±2.5197.74±0.3648.19±6.2390.53±1.42
", + "bbox": [ + 91, + 271, + 906, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Representative OOD detection results with various ID datasets in the 16-shot setting. We use the same notation as Tab. 1.", + "bbox": [ + 112, + 551, + 880, + 566 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/076c5ec2f75ee8cd53e39c8b2940dcccfd6dbd0bae3b8b4fc0d38aacd046cb8c.jpg", + "image_caption": [ + "Figure 3. A portion of OOD detection average performance using other ID datasets under the 16-shot setting. We report the mean and standard deviation of the AUROC for our FA (blue) and CoOp (orange) using the MCM score. More details can be found in the Appendix B." + ], + "image_footnote": [], + "bbox": [ + 109, + 593, + 457, + 719 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "shown that some of the used OOD datasets contain more or less images of objects that belong to ID classes in ImageNet-1k [7]. Hence, probing the true performance of OOD detectors for ImageNet-1k demands OOD datasets that are both challenging and genuinely OOD. As shown in Tab. 3, to fully substantiate the effectiveness of our", + "bbox": [ + 89, + 809, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "FA, we leverage 3 additional cleaner and more challenging OOD datasets, including OpenImage-O [45], NINCO [2], and ImageNet-O [15], inspired by [2]. Notably, both our $\\mathrm{FA}_{\\mathrm{MCM}}$ and $\\mathrm{FA}_{\\mathrm{GL}}$ consistently outperform current SOTA methods in terms of average FPR95 and AUROC under different few-shot settings (More details under the 1-shot and 4-shot settings are provided in the Appendix A).", + "bbox": [ + 511, + 593, + 906, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Various ID Datasets. Considering factors like dataset variety, image resolution, and the number of classes [30], we also leverage 8 additional widely used datasets for few-shot setting as ID datasets for comparison [30, 34, 54]. Overall, despite our model is simple, it achieves superior average OOD performance on these ID datasets compared to other methods. We selected representative results using UCF101 [40] and EuroSAT [13] as ID datasets in the 16-shot setting, as shown in Tab. 4. Since using the other datasets (i.e. Food101 [3], StandfordCars [21], Caltech101 [12], FGVCAircraft [29], Flowers102 [35], and OxfordPets [36]) as ID datasets for OOD detection presents relatively lower difficulty as shown in Fig. 3, we leave more", + "bbox": [ + 511, + 704, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1158", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3179384cf6c5f32757ff2ec23c1096e85edf7724d0bee176cc0262fa0bb2070b.jpg", + "image_caption": [ + "(a) Conventional OOD benchmark" + ], + "image_footnote": [], + "bbox": [ + 112, + 88, + 281, + 190 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/473ae12adfd35b414a654564acef41ba335a1dae659dd6e9bbf391ca6c4e51ab.jpg", + "image_caption": [ + "(b)Challenging OOD benchmark", + "Figure 4. Sensitivity study of hyperparameter $K$ . We report the average FPR95 (Blue) and AUROC (Orange) of our $\\mathrm{FA}_{\\mathrm{GL}}$ on different OOD benchmarks using ImageNet-1k as the ID dataset under 16-shot setting. Dashed lines represent the performance of the best baseline (i.e. $\\mathrm{SCT}_{\\mathrm{GL}}$ )." + ], + "image_footnote": [], + "bbox": [ + 290, + 88, + 460, + 190 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "experimental details in Appendix B.", + "bbox": [ + 89, + 287, + 331, + 303 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In particular, for the ID dataset EuroSAT, $\\mathrm{FA}_{\\mathrm{MCM}}$ surpasses the best-competing method $\\mathrm{LoCoOp_{GL}}$ by $16.09\\%$ in average AUROC and $38.7\\%$ in average FPR95, underscoring its effectiveness and versatility even when the ID dataset is of low resolution (64x64 pixels).", + "bbox": [ + 89, + 303, + 483, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation study", + "text_level": 1, + "bbox": [ + 89, + 390, + 240, + 406 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Influence of the $\\mathcal{L}_{FCE-K}$ . As shown in Tab. 5, we trained the model $\\mathrm{FA}_{CE}$ and $\\mathrm{FA}_{FCE-K}$ using the standard cross-entropy loss $\\mathcal{L}_{CE}$ and our $\\mathcal{L}_{FCE-K}$ , respectively. Our $\\mathrm{FA}_{FCE-K}$ significantly outperforms $\\mathrm{FA}_{CE}$ under the same inference setting, which demonstrate the effectiveness of our $\\mathcal{L}_{FCE-K}$ .", + "bbox": [ + 89, + 412, + 483, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Initialization of the forced and original prompt. As shown in Tab. 6, we perform experiments comparing different initialization scenarios for both prompts. Manual initialization uses the embeddings of \"a photo of a [class-c]\" to initialize the prompt's embeddings, while the random initialization setting is based on CoOp [54]. The results show that employing consistent manual initialization for both prompts enables the forced prompt to effectively learn more specific descriptions of the ID classes.", + "bbox": [ + 89, + 503, + 483, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Influence of the shared learnable vector. We also explore the influence of the shared learnable vector and the independent learnable vector for the forced prompts. As shown in Tab. 7, we can observe that using the shared learnable vector performs well in the few-shot setting since the independent learnable vector for each class has more parameters and requires more training data [54].", + "bbox": [ + 89, + 640, + 483, + 747 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Sensitivity study", + "text_level": 1, + "bbox": [ + 89, + 757, + 253, + 773 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Fig. 4, we explore the influence of the hyperparameter $K$ . Overall, with $K$ ranging from 1 to 6, the results indicate that our proposed framework is insensitive to the choice of $K$ . Concretely, our method is stable and outperforms the best baseline $\\mathrm{SCT}_{\\mathrm{GL}}$ in average AUROC and FPR95 on both conventional and challenging OOD detection benchmarks. Specially, as $K$ increases, the average AUROC shows an upward trend and then tends to be sta", + "bbox": [ + 89, + 779, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ble (see Fig. 4 (a)), showing a bottleneck in capturing more comprehensive semantic representations of the ID classes.", + "bbox": [ + 511, + 90, + 903, + 122 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0d2ab2d3dc31e25d8dcba5b51277f58ee42c764870cb61853ec9ae9292522cbb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMCMGL-MCM
FPR95↓AUROC↑FPR95↓AUROC↑
FAce41.43±3.1191.01±0.6734.29±1.9490.99±0.56
FAFCE-K29.07±1.1193.77±0.1925.68±0.5893.82±0.11
", + "bbox": [ + 521, + 136, + 898, + 204 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3a5bbe0446b33e23b45d99fc3a7d3742084c89f429d48ecaf48249aafdb4719f.jpg", + "table_caption": [ + "Table 5. Influence of the $\\mathcal{L}_{FCE-K}$ . MCM and GL-MCM refer to the score functions used during model inference." + ], + "table_footnote": [], + "table_body": "
Forced-M-InitOriginal-M-InitFA MCMFA GL
FPR95↓AUROC↑FPR95↓AUROC↑
--40.71±4.0091.28±0.8234.88±2.9292.12±0.76
✓-40.81±2.2591.25±0.6436.03±2.8491.62±1.01
-✓31.35±1.3493.35±0.3426.39±1.0293.64±0.19
✓✓29.07±1.1193.77±0.1925.68±0.5893.82±0.11
", + "bbox": [ + 521, + 257, + 898, + 354 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c2c3d01904873d81b6eba1a93aff2071de38602fb85df39b9cc6d5be0849582c.jpg", + "table_caption": [ + "Table 6. Ablation study of different initialization scenarios for the forced and original prompts. \"M-Init\" represents manual initialization. $\\checkmark$ represents using manual initialization; - represents using random initialization." + ], + "table_footnote": [], + "table_body": "
Shared- VectorM-InitFA MCMFA GL
FPR95↓AUROC↑FPR95↓AUROC↑
--44.02±2.2490.17±0.4246.29±1.3388.24±0.41
✓-40.71±4.0091.28±0.8234.88±2.9292.12±0.76
-✓34.69±0.6292.57±0.1229.27±0.5592.82±0.80
✓✓29.07±1.1193.77±0.1925.68±0.5893.82±0.11
", + "bbox": [ + 521, + 435, + 898, + 534 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Ablation study of the shared learnable vector. $\\sqrt{}$ - for \"Shared-vector\" represents using the shared / independent learnable vector for the forced prompt; $\\sqrt{}$ - for \"M-Init\" represents using manual / random initialization for the forced prompt.", + "bbox": [ + 511, + 540, + 905, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 617, + 633, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this study, we propose a novel CLIP-based framework for OOD detection based on Forced prompt leArning (FA), which focuses on fully exploiting the ID knowledge to effectively improve OOD detection without exploring complex OOD-related knowledge. We introduce a learnable forced prompt in addition to the frozen original prompt, both of which are using the same manual initialization. The forced prompt treats the original prompt as a reference, forcing itself to learn more diversified semantic descriptions of the ID classes rather than being limited to the textual semantics of class labels. We also introduce a forced coefficient to facilitate the forced prompt in learning more nuanced descriptions of the ID classes. Comprehensive experimental evaluations demonstrate that our method consistently surpasses current state-of-the-art methods on diverse benchmarks. We expect that our study can bring new insight on VLMs-based OOD detection methods and inspire more future research.", + "bbox": [ + 511, + 643, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1159", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported in part by the National Natural Science Foundation of China (grant No. 62071502), the Major Key Project of PCL (grant No. PCL2023A09), and Guangdong Excellent Youth Team Program (grant No. 2023B1515040025).", + "bbox": [ + 89, + 114, + 483, + 190 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 203, + 187, + 218 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yichen Bai, Zongbo Han, Bing Cao, Xiaoheng Jiang, Qinghua Hu, and Changqing Zhang. Id-like prompt learning for few-shot out-of-distribution detection. In CVPR, pages 17480-17489, 2024. 2, 3, 5, 6", + "[2] Julian Bitterwolf, Maximilian Müller, and Matthias Hein. In or out? fixing imagenet out-of-distribution detection evaluation. In ICML, pages 2471-2506, 2023. 5, 6, 7", + "[3] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 - mining discriminative components with random forests. In ECCV, pages 446-461, 2014. 5, 7, 11", + "[4] Guangyi Chen, Weiran Yao, Xiangchen Song, Xinyue Li, Yongming Rao, and Kun Zhang. PLOT: prompt learning with optimal transport for vision-language models. In ICLR, 2023. 2", + "[5] Mengyuan Chen, Junyu Gao, and Changsheng Xu. Conjugated semantic pool improves OOD detection with pretrained vision-language models. In NeurIPS, 2024. 2", + "[6] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In CVPR, pages 3606-3613, 2014. 5", + "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255, 2009. 5, 7", + "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 5", + "[9] Xuefeng Du, Zhaoning Wang, Mu Cai, and Yixuan Li. VOS: learning what you don't know by virtual outlier synthesis. In ICLR, 2022. 2", + "[10] Xuefeng Du, Yiyou Sun, Jerry Zhu, and Yixuan Li. Dream the impossible: Outlier imagination with diffusion models. NeurIPS, 36:60878-60901, 2023. 2", + "[11] Sepideh Esmaeilpour, Bing Liu, Eric Robertson, and Lei Shu. Zero-shot out-of-distribution detection based on the pre-trained model clip. In AAAI, pages 6568-6576, 2022. 2", + "[12] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In CVPRW, page 178, 2004. 5, 7, 11", + "[13] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 5, 7" + ], + "bbox": [ + 93, + 227, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In ICLR, 2017. 1, 2, 5", + "[15] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In CVPR, pages 15262-15271, 2021. 5, 7", + "[16] Dan Hendrycks, Steven Basart, Mantas Mazeika, Andy Zou, Joseph Kwon, Mohammadreza Mostajabi, Jacob Steinhardt, and Dawn Song. Scaling out-of-distribution detection for real-world settings. In ICML, pages 8759-8773, 2022. 1, 2, 5", + "[17] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alexander Shepard, Hartwig Adam, Pietro Perona, and Serge J. Belongie. The inaturalist species classification and detection dataset. In CVPR, pages 8769-8778, 2018. 5", + "[18] Yen-Chang Hsu, Yilin Shen, Hongxia Jin, and Zsolt Kira. Generalized ODIN: detecting out-of-distribution image without learning from out-of-distribution data. In CVPR, pages 10948-10957, 2020. 2", + "[19] Rui Huang and Yixuan Li. MOS: towards scaling out-of-distribution detection for large semantic space. In CVPR, pages 8710-8719, 2021. 5", + "[20] Xue Jiang, Feng Liu, Zhen Fang, Hong Chen, Tongliang Liu, Feng Zheng, and Bo Han. Negative label guided OOD detection with pretrained vision-language models. In ICLR, 2024. 2", + "[21] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCVW, pages 554–561, 2013. 5, 7, 11", + "[22] Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. In NeurIPS, pages 7167-7177, 2018. 2", + "[23] Tianqi Li, Guansong Pang, Xiao Bai, Wenjun Miao, and Jin Zheng. Learning transferable negative prompts for out-of-distribution detection. In CVPR, pages 17584-17594, 2024. 2, 3, 5", + "[24] Shiyu Liang, Yixuan Li, and R. Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. In ICLR, 2018. 1, 2, 5", + "[25] Jiashuo Liu, Zheyan Shen, Yue He, Xingxuan Zhang, Renzhe Xu, Han Yu, and Peng Cui. Towards out-of-distribution generalization: A survey. arXiv preprint arXiv:2108.13624, 2021. 1", + "[26] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM computing surveys, 55(9): 1-35, 2023. 2", + "[27] Weitang Liu, Xiaoyun Wang, John D. Owens, and Yixuan Li. Energy-based out-of-distribution detection. In NeurIPS, 2020. 1, 2, 5", + "[28] Shuo Lu, Yingsheng Wang, Lijun Sheng, Aihua Zheng, Lingxiao He, and Jian Liang. Recent advances in ood detection: Problems and approaches. arXiv preprint arXiv:2409.11884, 2024. 1" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1160", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 5, 7, 11", + "[30] Yifei Ming, Ziyang Cai, Jiuming Gu, Yiyou Sun, Wei Li, and Yixuan Li. Delving into out-of-distribution detection with vision-language representations. In NeurIPS, 2022. 1, 3, 5, 7", + "[31] Yifei Ming, Yiyou Sun, Ousmane Dia, and Yixuan Li. How to exploit hyperspherical embeddings for out-of-distribution detection? In ICLR, 2023. 2", + "[32] Atsuyuki Miyai, Qing Yu, Go Irie, and Kiyoharu Aizawa. Zero-shot in-distribution detection in multi-object settings using vision-language foundation models. arXiv preprint arXiv:2304.04521, 2023. 1, 3, 5", + "[33] Atsuyuki Miyai, Qing Yu, Go Irie, and Kiyoharu Aizawa. Locoop: Few-shot out-of-distribution detection via prompt learning. In NeurIPS, 2023. 2, 3, 5", + "[34] Jun Nie, Yonggang Zhang, Zhen Fang, Tongliang Liu, Bo Han, and Xinmei Tian. Out-of-distribution detection with negative prompts. In ICLR, 2024. 2, 3, 5, 7", + "[35] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729, 2008. 5, 7, 11", + "[36] Omkar M. Parkhi, Andreae Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In CVPR, pages 3498-3505, 2012. 5, 7, 11", + "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 1, 2", + "[38] Mohammadreza Salehi, Hossein Mirzaei, Dan Hendrycks, Yixuan Li, Mohammad Hossein Rohban, and Mohammad Sabokrou. A unified survey on anomaly, novelty, open-set, and out-of-distribution detection: Solutions and future challenges. arXiv preprint arXiv:2110.14051, 2021. 1", + "[39] Vikash Sehwag, Mung Chiang, and Prateek Mittal. SSD: A unified framework for self-supervised outlier detection. In ICLR, 2021. 2", + "[40] K Soomro. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012.5,7", + "[41] Yiyou Sun, Chuan Guo, and Yixuan Li. React: Out-of-distribution detection with rectified activations. In NeurIPS, pages 144-157, 2021. 1, 2, 5", + "[42] Yiyou Sun, Yifei Ming, Xiaojin Zhu, and Yixuan Li. Out-of-distribution detection with deep nearest neighbors. In ICML, pages 20827-20840, 2022. 2", + "[43] Leitian Tao, Xuefeng Du, Jerry Zhu, and Yixuan Li. Nonparametric outlier synthesis. In ICLR, 2023. 2", + "[44] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Open-set recognition: A good closed-set classifier is all you need. In ICLR, 2022. 2" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[45] Haoqi Wang, Zhizhong Li, Litong Feng, and Wayne Zhang. Vim: Out-of-distribution with virtual-logit matching. In CVPR, pages 4911-4920, 2022. 5, 7", + "[46] Hualiang Wang, Yi Li, Huifeng Yao, and Xiaomeng Li. CLIPN for zero-shot OOD detection: Teaching CLIP to say no. In ICCV, pages 1802-1812, 2023. 2, 5", + "[47] Jianxiong Xiao, James Hays, Krista A. Ehinger, Aude Oliva, and Antonio Torralba. SUN database: Large-scale scene recognition from abbey to zoo. In CVPR, pages 3485-3492, 2010. 5", + "[48] Geng Yu, Jianing Zhu, Jiangchao Yao, and Bo Han. Self-calibrated tuning of vision-language models for out-of-distribution detection. In NeurIPS, 2024. 2, 3, 5, 6", + "[49] Tao Yu, Zhihe Lu, Xin Jin, Zhibo Chen, and Xinchao Wang. Task residual for tuning vision-language models. In CVPR, pages 10899-10909, 2023. 5", + "[50] Yeonguk Yu, Sungho Shin, Seongju Lee, Changyun Jun, and Kyoobin Lee. Block selection method for using feature norm in out-of-distribution detection. In CVPR, pages 15701-15711, 2023. 2", + "[51] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kun-chang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of CLIP for few-shot classification. In ECCV, pages 493-510, 2022. 5", + "[52] Bolei Zhou, Ágata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE TPAMI, 40(6):1452-1464, 2018. 5", + "[53] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In CVPR, pages 16795-16804, 2022. 2", + "[54] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 130(9):2337-2348, 2022. 2, 3, 4, 5, 7, 8", + "[55] Yibo Zhou. Rethinking reconstruction autoencoder-based out-of-distribution detection. In CVPR, pages 7379-7387, 2022. 2" + ], + "bbox": [ + 514, + 90, + 903, + 614 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1161", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/FA_ Forced Prompt Learning of Vision-Language Models for Out-of-Distribution Detection/8ae5a438-b943-4e71-b751-475d9f24c3b7_model.json b/2025/FA_ Forced Prompt Learning of Vision-Language Models for Out-of-Distribution Detection/8ae5a438-b943-4e71-b751-475d9f24c3b7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..535baf8dc437b9aa206e2dec9381aa46fe2aed76 --- /dev/null +++ b/2025/FA_ Forced Prompt Learning of Vision-Language Models for Out-of-Distribution Detection/8ae5a438-b943-4e71-b751-475d9f24c3b7_model.json @@ -0,0 +1,1991 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.194, + 0.131, + 0.806, + 0.175 + ], + "angle": 0, + "content": "FA: Forced Prompt Learning of Vision-Language Models for Out-of-Distribution Detection" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.202, + 0.907, + 0.277 + ], + "angle": 0, + "content": "Xinhua Lu\\(^{1,2,4}\\), Runhe Lai\\(^{1,2,4}\\), Yanqi Wu\\(^{1,2,4}\\), Kanghao Chen\\(^{3\\dagger}\\), Wei-Shi Zheng\\(^{1,2,4}\\), Ruixuan Wang\\(^{1,2,4\\dagger}\\) \n\\(^{1}\\)Sun Yat-sen University \\(^{2}\\)Peng Cheng Laboratory \n\\(^{3}\\)Hong Kong University of Science and Technology (Guangzhou) \n\\(^{4}\\)Key Laboratory of Machine Intelligence and Advanced Computing" + }, + { + "type": "text", + "bbox": [ + 0.159, + 0.278, + 0.846, + 0.31 + ], + "angle": 0, + "content": "{luxh55, lairh5, wuyq268}@mail2.sysu.edu.cn, kchen879@connect.hkust-gz.edu.cn, wszheng@ieee.org, wangruix5@mail.sysu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.345, + 0.327, + 0.361 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.378, + 0.485, + 0.771 + ], + "angle": 0, + "content": "Pre-trained vision-language models (VLMs) have advanced out-of-distribution (OOD) detection recently. However, existing CLIP-based methods often focus on learning OOD-related knowledge to improve OOD detection, showing limited generalization or reliance on external largescale auxiliary datasets. In this study, instead of delving into the intricate OOD-related knowledge, we propose an innovative CLIP-based framework based on Forced prompt leArning (FA), designed to make full use of the InDistribution (ID) knowledge and ultimately boost the effectiveness of OOD detection. Our key insight is to learn a prompt (i.e. forced prompt) that contains more diversified and richer descriptions of the ID classes beyond the textual semantics of class labels. Specifically, it promotes better discernment for ID images, by forcing more notable semantic similarity between ID images and the learnable forced prompt. Moreover, we introduce a forced coefficient, encouraging the forced prompt to learn more comprehensive and nuanced descriptions of the ID classes. In this way, FA is capable of achieving notable improvements in OOD detection, even when trained without any external auxiliary datasets, while maintaining an identical number of trainable parameters as CoOp. Extensive empirical evaluations confirm our method consistently outperforms current state-of-the-art methods. Code is available at https://github.com/0xFIFA/FA." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.803, + 0.222, + 0.818 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.829, + 0.484, + 0.875 + ], + "angle": 0, + "content": "AI models often encounter Out-of-Distribution (OOD) samples [14, 25, 28, 38], which differ from the distribution of the training data, when deployed in real-world applications." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.345, + 0.9, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.563, + 0.907, + 0.675 + ], + "angle": 0, + "content": "Figure 1. Comparison of CLIP-based OOD detection methods. Existing methods focus on OOD-related knowledge to learn a complex ID/OOD decision boundary as shown in the orange area in (b). In contrast, our FA aims to fully exploit ID knowledge by forcing the model to learn richer descriptions of the ID classes beyond the textual semantics of class labels. These richer descriptions of the ID classes compared to the reference text allow better discernment of ID/OOD samples (orange vs. blue area in (a))." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.688, + 0.906, + 0.749 + ], + "angle": 0, + "content": "Detecting OOD data is vital for the reliability of AI systems. This problem becomes harder in fields like intelligent diagnosis, where only a small amount of labeled data is available, motivating the problem of few-shot OOD detection." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.902 + ], + "angle": 0, + "content": "To tackle the OOD detection problem, previous single-modal post-hoc methods [14, 16, 24, 27, 41] achieve notable success based on a model pre-trained on the ID dataset. However, they come with limitations, such as requiring substantial computational resources and annotation costs for training, which hinder the performance of few-shot OOD detection. With the development of large pre-trained vision-language models (VLMs) like CLIP [37], few-shot OOD detection has achieved remarkable performance. Most of the CLIP-based OOD detection methods [30, 32] focus on" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.887, + 0.245, + 0.901 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1152" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "leveraging the powerful generalization capability of CLIP to improve OOD detection." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.125, + 0.484, + 0.502 + ], + "angle": 0, + "content": "Recent CLIP-based OOD detection methods [1, 11, 23, 33, 34, 46, 48] aim to learn OOD-related knowledge to improve OOD detection performance. For example, recent CLIP-based works [11, 46] improve OOD detection by relying on training on external large-scale auxiliary datasets, which entails significant resource overhead. To eliminate dependence on external datasets, existing methods [1, 33, 48] learn OOD-related knowledge from exposed outliers based only on ID training data, such as background regions in ID training images. However, these OOD features extracted from specific regions are difficult to match with infinite OOD data encountered in practice. Moreover, some methods [23, 34] aim to learn negative prompts semantically opposite to the ID class labels, while these limited negative prompts are often insufficient to capture the distinctions between the diverse and numerous OOD data and the ID data. Overall, improving OOD detection by learning OOD-related knowledge shows inherent limitations or requires significant computational resources as well as labor costs. Inspired by the method [44] for open-set recognition, which demonstrates that enhancing closed-set accuracy can typically improve open-set recognition capabilities, we are motivated to explore ways to improve the identification of ID images, thereby enhancing OOD detection performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.506, + 0.484, + 0.718 + ], + "angle": 0, + "content": "In this paper, rather than focusing on the intricate OOD-related knowledge, we propose a novel CLIP-based framework based on Forced prompt leArning (FA), which aims to fully exploit the ID knowledge and ultimately improve OOD detection. Our key insight is to learn a prompt that contains richer knowledge beyond the textual semantics of class labels. It facilitates advanced discernment for ID images, by forcing higher semantic similarity between ID images and the learnable prompt (i.e. forced prompt), which provides more diversified and richer descriptions of the ID classes. In this way, FA can achieve significant improvements in OOD detection performance, even when trained without any additional external data, while maintaining the same number of learnable parameters as CoOp [54]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Specifically, we introduce a novel forced prompt along with the original prompt, both of which are initialized identically. FA preserves the generalization capability of the VLMs (i.e. CLIP) by freezing the original prompt and optimizing a trainable copy (i.e. forced prompt). In particular, the forced prompt treats the original prompt as a reference, while forcing the text features associated with the forced prompt to be more salient compared to those associated with the original prompt. In this way, it effectively improves the OOD performance even without sacrificing the ID classification capability of the model. Additionally, we introduce a forced coefficient, encouraging the forced" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.227 + ], + "angle": 0, + "content": "prompt to learn more comprehensive connotations. Experimentally, FA achieves superior few-shot OOD detection performance across diverse OOD benchmarks. Compared to current SOTA methods [48] that focus on learning OOD-related knowledge, our method significantly reduces the average FPR95 score from \\(31.62\\%\\) to \\(27.81\\%\\) and improves the average AUROC from \\(92.01\\%\\) to \\(93.26\\%\\) even in 1-shot OOD detection on ImageNet-1k. The main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.232, + 0.905, + 0.277 + ], + "angle": 0, + "content": "- We propose a simple yet effective framework, which fully exploits the ID knowledge to improve OOD detection without focusing on intricate OOD-related knowledge." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.28, + 0.905, + 0.324 + ], + "angle": 0, + "content": "- We present a Forced prompt leArning (FA) strategy to exploit richer knowledge beyond the textual semantics of class labels." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.328, + 0.905, + 0.372 + ], + "angle": 0, + "content": "- We evaluate our method on diverse OOD benchmarks, showing that our model consistently outperforms current state-of-the-art methods." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.232, + 0.905, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.389, + 0.652, + 0.405 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.416, + 0.905, + 0.642 + ], + "angle": 0, + "content": "Prompt Learning. In recent years, pre-trained vision-language models (VLMs) such as CLIP [37] have demonstrated powerful few-shot learning capabilities in both visual and textual domains. However, the design of the prompt greatly affects the performance of VLMs for downstream tasks, which may require manually crafting numerous prompts. Inspired by prompt learning studies in natural language processing (NLP) [26], CoOp [54] uses a set of learnable context vectors to transfer CLIP for specific downstream tasks, becoming a pioneering method for subsequent studies [4, 33, 53]. Currently, prompt learning is widely used in OOD detection mainly to represent various OOD-related knowledge. In this work, we learn the forced prompt, which contains richer descriptions of the ID classes to improve OOD detection." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.906, + 0.779 + ], + "angle": 0, + "content": "Out-of-Distribution Detection. Conventional methods explore OOD detection for single-modal models. One line of studies designs score functions to differentiate ID and OOD data based on output from the pre-trained models, such as logit output [14, 16, 24, 27] or outputs from the penultimate layer [22, 41, 42, 50]. These methods are called post-hoc methods. Another line of studies [9, 10, 18, 31, 39, 43, 55] adopts various training strategies to learn a reliable decision boundary between ID and OOD data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Recently, the development of OOD detectors based on VLMs, especially CLIP, has received much attention as VLMs have demonstrated their remarkable generalization capability in both visual and textual domains. Some studies leverage real outlier information from external auxiliary OOD datasets [11, 46] or extensive corpora [5, 20] to promote OOD detection. However, this is impractical in real-world scenarios, where outliers are infinite and agnostic." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1153" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.176, + 0.094, + 0.816, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.342, + 0.908, + 0.398 + ], + "angle": 0, + "content": "Figure 2. Overview of the proposed FA framework. Our framework includes the learnable forced prompt and the frozen original prompt, both of which are initialized identically by a manual template \"a photo of a [class-c]\". The learnable forced prompt treats the frozen original prompt as a reference, forcing its text features to become more salient compared to those generated by the original prompt. The richer and more diversified description of ID classes learned by FA can ultimately improve OOD detection." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.41, + 0.485, + 0.863 + ], + "angle": 0, + "content": "To eliminate the dependence on external data, recent methods [1, 23, 32, 33, 48] were developed under the assumption that only ID data are available. As a representative zero-shot method that only uses available ID labels, MCM [30] employs softmax scaling to align visual features with ID text concepts for OOD detection. GL-MCM [32] further introduces a local maximum concept matching(L-MCM) score to improve the separability of the local information. Compared with zero-shot methods, which may undergo a domain gap with ID downstream data, prompt learning methods achieve better OOD detection performance with access to few-shot ID samples. For prompt learning OOD detection methods, LoCoOp [33] and SCT [48] keep the textual embeddings of ID classes away from ID-irrelevant local region embeddings in the multi-modal embedding space by their proposed entropy maximization strategy, which ensures that OOD embeddings can be dissimilar to any textual embedding of ID classes. To detect challenging OOD samples to improve OOD detection, ID-like [1] explores the vicinity of ID samples to construct OOD samples correlated to the ID and refines the additional \"ID-like\" text embeddings to fine-grained differences. In addition, NegPrompt [23] and LSN [34] introduce the negative connotations of ID categories which can be represented by additional negative prompts, enabling more accurate detection of OOD samples. Overall, various OOD-related knowledge is commonly adopted to address the model's overconfidence in current research. In this work, we focus on leveraging ID knowledge to improve OOD detection, instead of exploring complex OOD-related knowledge." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.408, + 0.652, + 0.423 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.433, + 0.907, + 0.659 + ], + "angle": 0, + "content": "Prompt Learning with CLIP. CLIP contains an image encoder \\( f(\\cdot) \\) and a text encoder \\( g(\\cdot) \\), designed to extract features from images and text descriptions respectively. Generally, for an ID dataset which contains \\( C \\) categories, the hand-crafted prompt \\( \\hat{\\mathbf{u}}_c \\) \"a photo of a [class-c]\" is designed to match images, where class-c represents the class name. Formally, these prompts are individually encoded in \\( \\hat{\\mathbf{t}}_c = g(\\hat{\\mathbf{u}}_c) \\in \\mathbb{R}^{d \\times 1} \\), \\( c = 1, \\dots, C \\). Given an image \\( \\mathbf{x} \\), it can be encoded in \\( \\mathbf{z} = f(\\mathbf{x}) \\in \\mathbb{R}^{d \\times 1} \\). Because image features and corresponding text features are aligned on the multimodal embedding space, the cosine similarity \\( \\cos(\\mathbf{z}, \\hat{\\mathbf{t}}_c) \\) of \\( \\mathbf{z} \\) and \\( \\hat{\\mathbf{t}}_c \\) can represent the matching degree between the image \\( \\mathbf{x} \\) and text description \\( \\hat{\\mathbf{u}}_c \\). Therefore, if the text feature \\( \\mathbf{t}_c \\) of class \\( c \\) has the highest similarity to \\( \\mathbf{z} \\), the image \\( \\mathbf{x} \\) will be considered to be class \\( c \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.909, + 0.827 + ], + "angle": 0, + "content": "To effectively transfer CLIP to downstream image recognition tasks, CoOp [54] sets a portion of tokens in the text prompt as continuous learnable parameters. Concretely, CoOp initializes the text prompts as \\(\\mathbf{u}_c = [\\mathbf{v}_1,\\dots ,\\mathbf{v}_L,\\mathbf{w}_c],c\\in \\{1,\\dots ,C\\}\\), where \\(L\\) is the length of the prompt's tokens, \\(\\mathbf{v}_i\\) (\\(i\\in \\{1,\\ldots ,L\\}\\)) is the learnable vector with the same dimension as the word embedding and \\(\\mathbf{w}_c\\) is the word embedding for the class name of class \\(c\\). The text prompt \\(\\mathbf{u}_c\\) is encoded into features \\(\\mathbf{t}_c = g(\\mathbf{u}_c)\\). Together with the image feature \\(\\mathbf{z}\\), the probability for the image \\(\\mathbf{x}\\) to be classified into class \\(c\\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.836, + 0.907, + 0.876 + ], + "angle": 0, + "content": "\\[\np (y = c \\mid \\mathbf {x}) = \\frac {e ^ {\\cos (\\mathbf {z} , \\mathbf {t} _ {c}) / \\tau}}{\\sum_ {j = 1} ^ {C} e ^ {\\cos (\\mathbf {z} , \\mathbf {t} _ {j}) / \\tau}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.887, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\(\\tau\\) is a fixed temperature scaling hyper-parameter." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1154" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "Overall, the prompt's tokens can be trained to align with training data by minimizing the cross-entropy loss with the prediction probability from Eq. (1)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.137, + 0.482, + 0.213 + ], + "angle": 0, + "content": "OOD Detection. In the OOD detection task, the model is expected to determine whether a test image belongs to one of the learned ID classes. Therefore, OOD detection can be seen as a binary classification problem to distinguish ID images from OOD images in a test set \\(\\mathcal{D}_{test}\\) as follows" + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.224, + 0.482, + 0.26 + ], + "angle": 0, + "content": "\\[\nD (\\mathbf {x}) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} S (\\mathbf {x}) \\geq \\mu \\\\ 0, & \\text {i f} S (\\mathbf {x}) < \\mu \\end{array} , \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.27, + 0.483, + 0.331 + ], + "angle": 0, + "content": "where \\(\\mathbf{x} \\in \\mathcal{D}_{test}\\), 1 and 0 respectively indicate that \\(\\mathbf{x}\\) is classified as the ID class and the OOD class by the OOD detector \\(D(\\cdot)\\), \\(S(\\cdot)\\) is certain score function, and \\(\\mu\\) is a predefined threshold constant." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.345, + 0.226, + 0.362 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.37, + 0.201, + 0.384 + ], + "angle": 0, + "content": "4.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.392, + 0.483, + 0.634 + ], + "angle": 0, + "content": "In this study, we propose a novel CLIP-based framework based on Forced prompt leArning (FA), designed to fully exploit the ID knowledge from few-shot ID samples by forcing the prompt to learn richer knowledge beyond the textual semantics of class labels. As shown in Fig. 2, we introduce a forced prompt along with the original prompt, both of which are initialized with the same semantics. By freezing the original prompt and optimizing a trainable copy, ID images will show higher semantic similarity to the forced prompt compared to the original prompt. Moreover, to encourage descriptions of ID classes learned by the forced prompt to become more detailed and comprehensive, we also introduce a forced coefficient. During testing, the model is capable of effectively distinguishing between ID data and OOD data, benefiting from the diversified descriptions of the ID classes provided by the forced prompt." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.644, + 0.312, + 0.66 + ], + "angle": 0, + "content": "4.2. Forced prompt learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.666, + 0.483, + 0.803 + ], + "angle": 0, + "content": "Merely the semantic information conveyed by class names is insufficient to comprehensively encompass the discriminative information of each class. In order to explore richer semantic descriptions of the ID classes beyond the textual semantics of class labels, a novel forced prompt along with the original prompt is proposed. To implement the above objective, we introduce a simple yet effective training strategy using the forced cross-entropy (FCE) loss based on the forced prompt as follows" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.812, + 0.482, + 0.868 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {F C E} = \\mathbb {E} _ {\\left(\\mathbf {x}, y _ {c}\\right) \\sim \\mathcal {D} _ {t r a i n} ^ {I D}} \\left[ - \\log \\frac {e ^ {s _ {c} ^ {f} / \\tau}}{\\sum_ {j = 1} ^ {C} e ^ {s _ {j} ^ {f} / \\tau} + \\sum_ {j = 1} ^ {C} e ^ {s _ {j} ^ {o} / \\tau}} \\right], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.868, + 0.483, + 0.904 + ], + "angle": 0, + "content": "where the ID training dataset \\(\\mathcal{D}_{train}^{ID}\\) consists of ID image-label pairs \\((\\mathbf{x},y_{c})\\), \\(s_j^f = \\cos (\\mathbf{z},\\mathbf{t}_j^f)\\) and \\(s_j^o = \\cos (\\mathbf{z},\\mathbf{t}_j^o)\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "represent the similarity between the image feature and the prompt feature corresponding to the forced prompt and the original prompt, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.141, + 0.907, + 0.398 + ], + "angle": 0, + "content": "Based on this loss function, the ID image features will be forced to show higher cosine similarity to the learnable forced prompt compared to the original text prompt. This is because the cosine similarity between the ID image features and the original text prompt is already high before any further prompt learning. Consequently, the forced prompt is forced to uncover richer discriminative information of ID classes, so that the text features of the forced prompt achieve a higher cosine similarity with the image features compared to those of the original prompt. Benefiting from this information specific to the ID classes, the cosine similarity between the image features of an OOD image and the text features of both prompts may exhibit less distinction compared to those of the ID data. Therefore, even when trained without any reliance on external auxiliary datasets, the detector will have a notable capability to distinguish ID data from OOD data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.402, + 0.908, + 0.902 + ], + "angle": 0, + "content": "However, this naive FCE loss may not work effectively because of its inherent limitations. Concretely, employing random initialization for both prompts fails to fully leverage the semantic information provided by CLIP's prior knowledge, resulting in limited performance. To solve this limitation, we propose to initialize the forced prompt using the manual template (i.e. \"a photo of a [class-c]\") identified to the original prompt. In this way, both prompts with manual initialization will have clear semantic information compared to random initialization, thereby improving the model's generalization capability. Specifically, the class labels are utilized in both prompts by concatenating with the prompt embedding to provide foundational semantic information related to the category names, based on CLIP's prior knowledge. In particular, for the forced prompt, we adopt the shared learnable vector across all classes rather than the independent learnable vector for each class. This choice is inspired by CoOp [54], which demonstrates that using an independent learnable vector mostly underperforms the shared learnable vector in challenging low-data scenarios since the former has more parameters and requires more data for training. Formally, we utilize the embeddings of the manual template to initialize the prompt's embeddings of both prompts, formulated as \\(\\mathbf{u}_c = [\\mathbf{v}_1,\\dots ,\\mathbf{v}_L,\\mathbf{w}_c], c\\in\\) \\(\\{1,\\dots ,C\\}\\), which will be fed to the text encoder to obtain prompt feature \\(\\mathbf{t}_c\\) (see Section 3). Here, \\(L\\) is the length of the token (e.g., for the \"a photo of a\", \\(L = 4\\)). Then we freeze the original prompt and the class label part of both the prompts to preserve the generalization capability of CLIP, while only making the forced prompt learnable. Note that our method effectively improves OOD detection performance while maintaining consistency with CoOp in the number of learnable parameters without leveraging ad" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1155" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.107 + ], + "angle": 0, + "content": "ditional learnable prompts, unlike existing work [1, 23, 34]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.108, + 0.483, + 0.38 + ], + "angle": 0, + "content": "Although the above prompt design achieves moderate performance, relying solely on a single original prompt as a reference is somewhat insufficient to develop a comprehensive capability, particularly with the limitations of ID data (e.g., few-shot OOD). To encourage the forced prompt to capture more comprehensive and nuanced descriptions of the ID classes, we introduce a forced coefficient \\( K \\) (\\( K \\geq 0, K \\in \\mathbb{N} \\)), which indicates the intensity with which the model is compelled to learn from the data. Formally, we first derive the original text features \\( \\{\\mathbf{t}_1^o, \\dots, \\mathbf{t}_C^o\\} \\) from the original prompt and the image features \\( \\mathbf{z} \\) from the corresponding image. Then, we compute the cosine similarity between the original text features and the image features for \\( K \\) iterations, which will be used in the subsequent computation of the cross-entropy loss function. Notably, when \\( K = 0 \\), no original prompt is used, and the model operates equivalently to CoOp. Based on the forced coefficient, we refine Eq. (3) as follows" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.392, + 0.485, + 0.446 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {F C E - K} = \\mathbb {E} _ {\\left(\\mathbf {x}, y _ {c}\\right) \\sim \\mathcal {D} _ {t r a i n} ^ {I D}} \\left[ - \\log \\frac {e ^ {s _ {c} ^ {f} / \\tau}}{\\sum_ {j = 1} ^ {C} e ^ {s _ {j} ^ {f} / \\tau} + K \\sum_ {j = 1} ^ {C} e ^ {s _ {j} ^ {o} / \\tau}} \\right]. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.447, + 0.483, + 0.492 + ], + "angle": 0, + "content": "Based on this function, the cosine similarity between the ID image features and the text features of the forced prompt must become more salient as the coefficient \\( K \\) increases." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.503, + 0.253, + 0.518 + ], + "angle": 0, + "content": "4.3. Model inference" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.526, + 0.484, + 0.662 + ], + "angle": 0, + "content": "During model inference, for the downstream classification task, we adopt the same strategy as CLIP, depending solely on the forced prompt [1]. For the OOD detection task, our method can be flexibly combined with different score functions, such as the MCM [30] score and the GL-MCM [32] score. The MCM score is defined as the maximum similarity between the global image features \\(\\mathbf{z}^g\\) (i.e. \\(\\mathbf{z}\\)) and all text features \\(\\mathbf{t}_c^a\\) (i.e. the concatenation of text features \\(\\mathbf{t}_c^f\\) and \\(\\mathbf{t}_c^o\\)) after applying softmax with temperature \\(\\tau_0\\), i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.673, + 0.483, + 0.727 + ], + "angle": 0, + "content": "\\[\nS _ {\\mathrm {M C M}} (\\mathbf {x}) = \\max _ {c} \\frac {e ^ {\\cos \\left(\\mathbf {z} ^ {g} , \\mathbf {t} _ {c} ^ {a}\\right) / \\tau_ {0}}}{\\sum_ {j = 1} ^ {C} e ^ {\\cos \\left(\\mathbf {z} ^ {g} , \\mathbf {t} _ {j} ^ {f}\\right) / \\tau_ {0}} + K e ^ {\\cos \\left(\\mathbf {z} ^ {g} , \\mathbf {t} _ {j} ^ {o}\\right) / \\tau_ {0}}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.728, + 0.483, + 0.758 + ], + "angle": 0, + "content": "while the GL-MCM score simultaneously considers both global and local features, which can be expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.771, + 0.483, + 0.789 + ], + "angle": 0, + "content": "\\[\nS _ {\\mathrm {G L - M C M}} (\\mathbf {x}) = S _ {\\mathrm {M C M}} (\\mathbf {x}) + S _ {\\mathrm {L - M C M}} (\\mathbf {x}), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.801, + 0.482, + 0.855 + ], + "angle": 0, + "content": "\\[\nS _ {\\mathrm {L - M C M}} (\\mathbf {x}) = \\max _ {i, c} \\frac {e ^ {\\cos \\left(\\mathbf {z} _ {i} ^ {l} , \\mathbf {t} _ {c} ^ {a}\\right) / \\tau_ {0}}}{\\sum_ {j = 1} ^ {C} e ^ {\\cos \\left(\\mathbf {z} _ {i} ^ {l} , \\mathbf {t} _ {j} ^ {f}\\right) / \\tau_ {0}} + K e ^ {\\cos \\left(\\mathbf {z} _ {i} ^ {l} , \\mathbf {t} _ {j} ^ {o}\\right) / \\tau_ {0}}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathbf{z}_i^l (i\\in \\{1,\\dots ,N\\})\\) represents \\(N\\) extracted local features generated by CLIP's image encoder [32, 33]. We set \\(\\tau_0 = 1\\) during model inference." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.646, + 0.107 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.115, + 0.71, + 0.132 + ], + "angle": 0, + "content": "5.1. Experimental details" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.139, + 0.907, + 0.349 + ], + "angle": 0, + "content": "Datasets. We use a popular benchmark for conventional OOD detection [19, 30, 32, 33], where ImageNet-1k [7] serves as the ID dataset, and the OOD datasets are iNaturalist [17], SUN [47], Places [52], and Texture [6]. Moreover, inspired by existing studies [1, 2], we use datasets like OpenImage-O [45], NINCO [2], and ImageNet-O [15], which are cleaner and more realistic, to simulate the more challenging OOD detection. Besides, we also utilize other widely adopted datasets for few-shot settings as ID datasets [49, 51, 54], which include StanfordCars [21], UCF101 [40], Caltech101 [12], Flowers102 [35], EuroSAT [13], FGVCAircraft [29], OxfordPets [36], and Food101 [3], considering factors such as image resolution and the number of classes." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.351, + 0.908, + 0.563 + ], + "angle": 0, + "content": "Setup. Following previous studies [33], we use the ViTB/16 [8] as the backbone model. For our model, when using ImageNet-1k as the ID dataset, the epochs are respectively set to 30 and 50 for the 1-shot and 16-shot settings, while for other ID datasets, we set the epoch to 200 following CoOp [54]. The value of forced coefficient \\( K \\) is uniformly set to 3, which will be further discussed in the sensitivity study. Other hyperparameters are as follows: learning rate \\( = 2\\mathrm{e} - 3 \\), batch size \\( = 160 \\), SGD (momentum \\( = 0.9 \\), weight decay \\( = 5\\mathrm{e} - 4 \\)) as the optimizer with a cosine scheduler, \\( \\tau = 1 \\) in Eq. (3) and Eq. (4). All experiments on our model can be conducted on a single Nvidia A30 GPU. The average experiment results (including our reproductions) over four runs are reported for comparison." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.564, + 0.907, + 0.715 + ], + "angle": 0, + "content": "Comparison Methods. To validate the effectiveness of our FA fairly, we mainly compare it with CLIP-based OOD detection methods that do not use real outliers (e.g. OOD labels). For previous post-hoc methods, including MSP [14], ODIN [24], Energy [27], ReAct [41], and MaxLogit [16], we adapt these methods with the CLIP image encoder as the CLIP-based post-hoc methods. For zero-shot methods, we select MCM [30], GL-MCM [32] and CLIPN [46] as baselines. For prompt learning methods, we adopt CoOp [54], LoCoOp [33], IDLike [1], and SCT [48] as baselines." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.715, + 0.905, + 0.791 + ], + "angle": 0, + "content": "Metrics. For evaluation, we adopt the following metrics: (1) The False Positive Rate at \\(95\\%\\) True Positive Rate for in-distribution samples (FPR95); (2) The Area Under the Receiver Operating Characteristic Curve (AUROC); (3) In-distribution data classification Top-1 accuracy (ID ACC)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.802, + 0.647, + 0.817 + ], + "angle": 0, + "content": "5.2. Main results" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.825, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Conventional OOD Detection. Table 1 summarizes our comparison results on the ImageNet-1k benchmarks, which show that our FA achieves state-of-the-art OOD detection performance among other CLIP-based methods under different few-shot settings (More details under the 4-shot set" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1156" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.091, + 0.088, + 0.907, + 0.577 + ], + "angle": 0, + "content": "
MethodiNaturalistSUNPlacesTexturesAverage
FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑
Zero-shot methods
MCM31.9594.1637.2292.5542.9890.1058.3585.8342.6390.66
GL-MCM15.0996.7229.0893.4137.0790.3758.9483.1135.0490.90
CLIPN19.1796.1726.4394.0232.2692.6241.2390.1230.2193.19
CLIP-based post-hoc methods
\\( MSP^† \\)74.5777.7476.9573.9779.7272.1873.6674.8476.2274.68
\\( ODIN^† \\)98.9357.7388.7278.4287.8076.8885.4771.4990.2371.13
\\( Energy^† \\)64.9887.1846.4291.1757.4087.3350.3988.2254.8088.48
\\( ReAct^† \\)65.5786.8746.1791.0456.8587.4249.8888.1354.6288.37
\\( MaxLogit^† \\)60.8888.0344.8391.1655.5487.4548.7288.6352.4988.82
Prompt learning based methods
\\( CoOp_{MCM} \\)41.14±9.3991.47±2.1239.06±4.1791.68±0.8345.38±4.8889.16±1.2851.37±3.0987.82±1.1244.24±1.3990.03±0.32
\\( CoOp_{GL} \\)23.30±6.3794.59±1.5132.08±3.8792.21±0.9539.22±4.5589.59±1.5855.78±3.4683.26±1.3837.59±0.3489.91±0.31
\\( LoCoOp_{MCM} \\)36.64±4.8792.78±0.9331.86±3.7993.50±0.9238.81±3.3790.70±0.9748.11±2.3189.43±0.6838.85±2.6791.59±0.58
\\( LoCoOp_{GL} \\)21.97±2.9195.39±0.5924.95±2.3794.42±0.5934.14±2.6691.14±0.6949.04±2.7787.73±0.9632.53±2.1992.17±0.47
IDLike17.73±1.9196.68±0.3148.17±1.3989.53±0.5850.43±6.5888.27±2.2729.12±7.6493.25±2.1636.36±3.8691.93±1.18
\\( LSN^† \\)59.28±7.0287.20±3.1540.15±0.8291.47±0.1446.11±1.8688.74±0.5760.34±0.1483.92±0.4251.47±1.5387.84±0.58
\\( NegPrompt^† \\)65.03±8.6984.56±2.5244.39±1.6689.63±0.6651.31±6.2186.55±2.1987.60±1.6163.76±3.0262.08±3.7181.13±1.78
\\( SCT_{MCM} \\)41.93±12.1791.77±2.4030.39±2.1793.76±0.3938.73±2.5490.78±0.3446.78±3.5188.96±1.1639.46±3.3991.32±0.90
\\( SCT_{GL} \\)20.57±10.295.63±1.8924.56±3.0394.39±0.5333.27±2.9691.27±0.5748.12±2.9786.76±0.8931.62±3.1992.01±0.77
\\( FA_{MCM}(Ours) \\)25.50±2.7294.72±0.2736.24±3.1592.40±0.9535.38±2.7591.99±0.6828.34±0.8793.95±0.1931.37±1.0793.25±0.32
\\( FA_{GL}(Ours) \\)14.12±1.3296.76±0.1029.99±1.5792.95±0.6632.48±1.4891.83±0.4934.66±1.2191.50±0.3627.81±0.4493.26±0.27
Prompt learning based methods
\\( CoOp_{MCM} \\)30.26±1.9893.43±0.8134.69±0.4392.59±0.1441.91±0.7190.11±0.2344.68±2.1189.95±0.4837.89±0.7191.52±0.29
\\( CoOp_{GL} \\)15.96±1.6796.11±0.5527.26±1.9993.29±0.4635.36±2.0890.58±0.6448.63±2.1186.11±0.5931.81±1.2791.51±0.39
\\( LoCoOp_{MCM} \\)27.35±3.1994.12±0.8030.93±1.1993.75±0.2638.26±1.5391.12±0.2741.36±2.5690.99±0.5334.47±0.7392.49±0.14
\\( LoCoOp_{GL} \\)18.46±1.3895.85±0.6322.43±0.9695.15±0.2231.53±1.5292.15±0.2243.35±3.1289.38±0.7828.94±1.2993.13±0.17
IDLike19.23±10.596.70±1.5854.15±2.8887.64±1.1956.63±0.0785.86±0.4434.69±6.4191.90±2.3141.18±1.7390.53±0.01
\\( LSN^† \\)36.17±4.8192.66±1.1634.27±0.4493.53±0.2041.47±0.8590.52±0.3746.43±0.6089.38±0.2439.58±0.7391.53±0.09
\\( NegPrompt^† \\)37.79±0.1190.49±0.0132.11±3.7792.25±1.0035.52±0.4191.16±0.0343.93±9.0988.38±3.3137.34±1.4190.57±0.59
\\( SCT_{MCM} \\)29.41±2.1993.76±0.5527.28±2.8094.22±0.4036.35±2.1391.16±0.3342.25±1.8990.52±0.4833.82±1.7892.42±0.35
\\( SCT_{GL} \\)15.19±2.1696.71±0.4420.00±0.6195.57±0.1129.71±0.8592.37±0.1244.17±0.8288.59±0.3927.27±0.4493.31±0.17
\\( FA_{MCM}(Ours) \\)25.79±1.4894.29±0.3533.54±1.1793.02±0.1933.77±1.6492.64±0.4223.17±1.3195.14±0.2629.07±1.1193.77±0.19
\\( FA_{GL}(Ours) \\)14.49±1.2796.48±0.2927.65±1.0893.46±0.1831.09±1.3892.44±0.3429.50±0.8992.93±0.1825.68±0.5893.82±0.11
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.582, + 0.907, + 0.653 + ], + "angle": 0, + "content": "Table 1. Comparison results on ImageNet-1k OOD benchmarks. All results using the same backbone ViT-B/16. The results marked with \\( \\dagger \\) are taken from [48]; the others are our reproductions. The prompt learning based methods are run under four trials, reporting the mean and standard deviation of the performance. The subscripts \\( {}_{\\mathrm{{MCM}}} \\) and \\( {}_{\\mathrm{{GL}}} \\) indicate the use of the MCM score and the GL-MCM score. The best and second-best results are indicated in bold and underline. \\( \\uparrow \\) indicates larger values are better; \\( \\downarrow \\) indicates smaller values are better. All values are percentages." + }, + { + "type": "table", + "bbox": [ + 0.119, + 0.671, + 0.457, + 0.78 + ], + "angle": 0, + "content": "
Method1-shot4-shot16-shot
CoOp67.44±0.5069.71±0.0770.99±0.14
LoCoOp67.40±0.6469.55±0.1071.53±0.17
IDLike68.17±0.5768.91±0.1469.46±0.02
SCT68.63±0.1369.93±0.2271.78±0.05
FA(Ours)68.67±0.3969.96±0.0471.02±0.09
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.785, + 0.483, + 0.826 + ], + "angle": 0, + "content": "Table 2. Comparison results in ID Top-1 accuracy on ImageNet-1k under different few-shot settings. Other notations are the same as Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "ting are provided in the Appendix A). Specifically, in the 16-shot setting, FA with GL-MCM score \\(\\left(\\mathrm{FA}_{\\mathrm{GL}}\\right)\\) outperforms the best baseline \\(\\mathrm{SCT}_{\\mathrm{GL}}\\) (average FPR95 of \\(25.68\\%\\) vs. \\(27.27\\%\\), and average AUROC of \\(93.82\\%\\) vs. \\(93.31\\%\\))." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.907, + 0.78 + ], + "angle": 0, + "content": "Additionally, in the 1-shot setting, \\(\\mathrm{FA}_{\\mathrm{GL}}\\) surpasses \\(\\mathrm{SCT}_{\\mathrm{GL}}\\) by a large margin, showing improvements of at least \\(1.25\\%\\) and \\(3.81\\%\\) on average AUROC and FPR95. More importantly, even the average score of \\(\\mathrm{FA}_{\\mathrm{MCM}}\\) has exceeded \\(\\mathrm{SCT}_{\\mathrm{GL}}\\) in the 1-shot setting. Moreover, our FA effectively improves the OOD performance without sacrificing the ID classification capability of the model, as shown in Tab. 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.884 + ], + "angle": 0, + "content": "In particular, we can also observe that our method performs not well on the SUN dataset, as shown in Tab. 1. This is related to the presence of a significant number of samples belonging to ID category in these datasets, as demonstrated by recent studies [1]. In other words, the SUN dataset in the popular OOD benchmark requires more thorough cleaning to be effectively used as an OOD dataset." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.885, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Challenging OOD Detection. Recent studies [1, 2] have" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1157" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.907, + 0.238 + ], + "angle": 0, + "content": "
MethodOpenImage-ONINCOImageNet-OAverage
FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑
CoOpMCM37.95±0.5991.99±0.2278.00±0.2373.00±0.5970.14±0.8681.34±0.8362.03±0.3882.11±0.48
CoOpGL32.23±1.0092.22±0.2572.04±0.8175.35±0.8867.35±1.1978.44±0.9457.21±0.9681.99±0.58
LoCoOpMCM37.03±0.9492.24±0.2177.96±0.6372.78±0.7970.96±0.7581.25±0.3961.98±0.5582.09±0.35
LoCoOpGL33.87±1.2392.53±0.0775.16±0.9272.97±0.3968.06±1.9281.19±0.2559.03±1.2782.23±0.17
IDLike54.43±2.3587.89±0.4578.93±1.7969.32±0.4284.08±1.4567.45±0.8672.48±0.8974.89±0.28
SCTMCM37.28±1.1692.04±0.3378.51±0.8671.09±1.1971.45±0.5481.29±0.1562.41±0.7381.48±0.52
SCTGL32.24±0.3492.75±0.1774.15±0.5373.14±0.9468.35±0.6280.84±0.4558.25±0.3982.24±0.44
FAmCM(Ours)37.93±1.6591.84±0.4570.44±1.2477.77±0.9463.84±1.3682.71±0.3057.40±1.3484.08±0.47
FAGL(Ours)32.10±0.8992.39±0.2865.61±1.3279.01±0.7163.13±0.9880.39±0.2153.61±0.8983.93±0.31
" + }, + { + "type": "table_caption", + "bbox": [ + 0.12, + 0.243, + 0.877, + 0.258 + ], + "angle": 0, + "content": "Table 3. Challenging OOD detection results on cleaner OOD datasets in the 16-shot setting. We use the same notation as Tab. 1." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.272, + 0.907, + 0.548 + ], + "angle": 0, + "content": "
ID DatasetMethodiNaturalistSUNPlacesTexturesAverage
FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑
UCF101CoOpMCM14.40±6.6197.41±1.3637.09±3.9493.14±0.4840.07±2.7791.89±0.6533.24±4.2193.19±1.2231.21±3.5893.91±0.45
CoOpGL9.65±6.3598.18±0.9828.54±5.0794.51±0.9232.94±1.7493.28±0.3235.96±2.6591.83±1.4226.77±3.2294.45±0.42
LoCoOpMCM2.38±0.9899.21±0.2219.73±4.2496.11±0.7822.57±3.0194.96±0.7715.89±3.1296.12±0.6715.14±1.9196.60±0.33
LoCoOpGL1.13±0.2299.65±0.0917.93±4.0996.51±0.8922.34±4.5895.03±1.0618.17±3.0195.25±0.7114.89±2.4996.61±0.58
IDLike66.63±27.689.75±5.5287.88±6.5678.04±6.1484.47±12.176.04±8.5349.07±11.291.19±2.7472.01±14.083.76±5.64
SCTMCM1.51±0.5199.44±0.1024.17±5.5495.21±0.8826.62±4.4193.75±0.8115.34±3.8495.98±0.6416.91±3.2096.09±0.48
SCTGL0.79±0.2799.75±0.0418.69±2.4195.86±0.5623.77±2.4793.98±0.4717.44±2.6294.52±0.5415.17±1.5696.03±0.25
FAmCM(Ours)0.12±0.1199.94±0.013.06±1.3399.34±0.234.68±1.0099.08±0.262.68±0.5199.31±0.432.63±0.5899.42±0.22
FAGL(Ours)0.13±0.0299.94±0.024.98±1.0299.01±0.227.46±0.8698.44±0.275.67±0.8398.75±0.194.56±0.4799.03±0.13
EuroSATCoOpMCM99.14±0.7331.95±9.1798.29±0.4944.36±4.1997.28±0.8547.51±4.6386.69±4.3769.17±3.1595.35±1.0348.25±3.09
CoOpGL92.45±4.2657.59±10.490.97±3.4961.89±4.8388.69±3.8564.00±4.9262.37±1.9383.58±2.3183.62±1.1866.77±2.83
LoCoOpMCM96.51±3.3353.49±8.0693.52±5.6855.59±9.1992.88±4.0557.89±5.1976.45±7.4878.59±2.6689.84±4.2961.39±3.90
LoCoOpGL86.13±12.272.68±7.9787.81±6.3768.58±6.4085.93±4.1569.90±3.4054.89±4.6787.00±1.0178.69±4.6574.54±3.59
IDLike98.02±2.1373.32±5.8395.74±2.8465.19±6.8694.79±3.6769.44±5.0663.34±9.1086.68±2.9887.97±4.2173.66±5.04
SCTMCM99.79±0.2232.90±6.7398.36±0.8444.30±3.4197.77±0.8648.08±1.6483.98±6.9273.25±3.5894.97±2.1949.63±2.39
SCTGL96.57±5.4158.21±5.3192.41±4.5561.54±3.4790.49±5.0263.33±2.7155.29±14.785.48±3.6883.69±7.0267.14±2.20
FAmCM(Ours)85.24±6.8878.62±3.1836.24±10.892.13±2.4428.17±10.294.05±1.6810.36±3.9797.71±0.6939.99±7.7990.63±1.91
FAGL(Ours)77.82±7.2886.88±2.1556.10±8.0388.43±2.2649.87±7.6989.07±1.498.99±2.5197.74±0.3648.19±6.2390.53±1.42
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.553, + 0.882, + 0.568 + ], + "angle": 0, + "content": "Table 4. Representative OOD detection results with various ID datasets in the 16-shot setting. We use the same notation as Tab. 1." + }, + { + "type": "image", + "bbox": [ + 0.11, + 0.594, + 0.459, + 0.72 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.723, + 0.483, + 0.793 + ], + "angle": 0, + "content": "Figure 3. A portion of OOD detection average performance using other ID datasets under the 16-shot setting. We report the mean and standard deviation of the AUROC for our FA (blue) and CoOp (orange) using the MCM score. More details can be found in the Appendix B." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.901 + ], + "angle": 0, + "content": "shown that some of the used OOD datasets contain more or less images of objects that belong to ID classes in ImageNet-1k [7]. Hence, probing the true performance of OOD detectors for ImageNet-1k demands OOD datasets that are both challenging and genuinely OOD. As shown in Tab. 3, to fully substantiate the effectiveness of our" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.594, + 0.907, + 0.701 + ], + "angle": 0, + "content": "FA, we leverage 3 additional cleaner and more challenging OOD datasets, including OpenImage-O [45], NINCO [2], and ImageNet-O [15], inspired by [2]. Notably, both our \\(\\mathrm{FA}_{\\mathrm{MCM}}\\) and \\(\\mathrm{FA}_{\\mathrm{GL}}\\) consistently outperform current SOTA methods in terms of average FPR95 and AUROC under different few-shot settings (More details under the 1-shot and 4-shot settings are provided in the Appendix A)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Various ID Datasets. Considering factors like dataset variety, image resolution, and the number of classes [30], we also leverage 8 additional widely used datasets for few-shot setting as ID datasets for comparison [30, 34, 54]. Overall, despite our model is simple, it achieves superior average OOD performance on these ID datasets compared to other methods. We selected representative results using UCF101 [40] and EuroSAT [13] as ID datasets in the 16-shot setting, as shown in Tab. 4. Since using the other datasets (i.e. Food101 [3], StandfordCars [21], Caltech101 [12], FGVCAircraft [29], Flowers102 [35], and OxfordPets [36]) as ID datasets for OOD detection presents relatively lower difficulty as shown in Fig. 3, we leave more" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1158" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.089, + 0.282, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.193, + 0.276, + 0.202 + ], + "angle": 0, + "content": "(a) Conventional OOD benchmark" + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.089, + 0.462, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.193, + 0.453, + 0.202 + ], + "angle": 0, + "content": "(b)Challenging OOD benchmark" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.21, + 0.483, + 0.281 + ], + "angle": 0, + "content": "Figure 4. Sensitivity study of hyperparameter \\( K \\). We report the average FPR95 (Blue) and AUROC (Orange) of our \\( \\mathrm{FA}_{\\mathrm{GL}} \\) on different OOD benchmarks using ImageNet-1k as the ID dataset under 16-shot setting. Dashed lines represent the performance of the best baseline (i.e. \\( \\mathrm{SCT}_{\\mathrm{GL}} \\))." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.289, + 0.332, + 0.304 + ], + "angle": 0, + "content": "experimental details in Appendix B." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.304, + 0.484, + 0.38 + ], + "angle": 0, + "content": "In particular, for the ID dataset EuroSAT, \\(\\mathrm{FA}_{\\mathrm{MCM}}\\) surpasses the best-competing method \\(\\mathrm{LoCoOp_{GL}}\\) by \\(16.09\\%\\) in average AUROC and \\(38.7\\%\\) in average FPR95, underscoring its effectiveness and versatility even when the ID dataset is of low resolution (64x64 pixels)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.391, + 0.241, + 0.407 + ], + "angle": 0, + "content": "5.3. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.413, + 0.484, + 0.505 + ], + "angle": 0, + "content": "Influence of the \\(\\mathcal{L}_{FCE-K}\\). As shown in Tab. 5, we trained the model \\(\\mathrm{FA}_{CE}\\) and \\(\\mathrm{FA}_{FCE-K}\\) using the standard cross-entropy loss \\(\\mathcal{L}_{CE}\\) and our \\(\\mathcal{L}_{FCE-K}\\), respectively. Our \\(\\mathrm{FA}_{FCE-K}\\) significantly outperforms \\(\\mathrm{FA}_{CE}\\) under the same inference setting, which demonstrate the effectiveness of our \\(\\mathcal{L}_{FCE-K}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.505, + 0.484, + 0.64 + ], + "angle": 0, + "content": "Initialization of the forced and original prompt. As shown in Tab. 6, we perform experiments comparing different initialization scenarios for both prompts. Manual initialization uses the embeddings of \"a photo of a [class-c]\" to initialize the prompt's embeddings, while the random initialization setting is based on CoOp [54]. The results show that employing consistent manual initialization for both prompts enables the forced prompt to effectively learn more specific descriptions of the ID classes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.641, + 0.484, + 0.748 + ], + "angle": 0, + "content": "Influence of the shared learnable vector. We also explore the influence of the shared learnable vector and the independent learnable vector for the forced prompts. As shown in Tab. 7, we can observe that using the shared learnable vector performs well in the few-shot setting since the independent learnable vector for each class has more parameters and requires more training data [54]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.758, + 0.254, + 0.774 + ], + "angle": 0, + "content": "5.4. Sensitivity study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "As shown in Fig. 4, we explore the influence of the hyperparameter \\(K\\). Overall, with \\(K\\) ranging from 1 to 6, the results indicate that our proposed framework is insensitive to the choice of \\(K\\). Concretely, our method is stable and outperforms the best baseline \\(\\mathrm{SCT}_{\\mathrm{GL}}\\) in average AUROC and FPR95 on both conventional and challenging OOD detection benchmarks. Specially, as \\(K\\) increases, the average AUROC shows an upward trend and then tends to be sta" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.123 + ], + "angle": 0, + "content": "ble (see Fig. 4 (a)), showing a bottleneck in capturing more comprehensive semantic representations of the ID classes." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.137, + 0.899, + 0.205 + ], + "angle": 0, + "content": "
MethodMCMGL-MCM
FPR95↓AUROC↑FPR95↓AUROC↑
FAce41.43±3.1191.01±0.6734.29±1.9490.99±0.56
FAFCE-K29.07±1.1193.77±0.1925.68±0.5893.82±0.11
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.211, + 0.905, + 0.24 + ], + "angle": 0, + "content": "Table 5. Influence of the \\(\\mathcal{L}_{FCE-K}\\). MCM and GL-MCM refer to the score functions used during model inference." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.258, + 0.899, + 0.356 + ], + "angle": 0, + "content": "
Forced-M-InitOriginal-M-InitFA MCMFA GL
FPR95↓AUROC↑FPR95↓AUROC↑
--40.71±4.0091.28±0.8234.88±2.9292.12±0.76
✓-40.81±2.2591.25±0.6436.03±2.8491.62±1.01
-✓31.35±1.3493.35±0.3426.39±1.0293.64±0.19
✓✓29.07±1.1193.77±0.1925.68±0.5893.82±0.11
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.362, + 0.905, + 0.418 + ], + "angle": 0, + "content": "Table 6. Ablation study of different initialization scenarios for the forced and original prompts. \"M-Init\" represents manual initialization. \\(\\checkmark\\) represents using manual initialization; - represents using random initialization." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.436, + 0.899, + 0.535 + ], + "angle": 0, + "content": "
Shared- VectorM-InitFA MCMFA GL
FPR95↓AUROC↑FPR95↓AUROC↑
--44.02±2.2490.17±0.4246.29±1.3388.24±0.41
✓-40.71±4.0091.28±0.8234.88±2.9292.12±0.76
-✓34.69±0.6292.57±0.1229.27±0.5592.82±0.80
✓✓29.07±1.1193.77±0.1925.68±0.5893.82±0.11
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.541, + 0.906, + 0.598 + ], + "angle": 0, + "content": "Table 7. Ablation study of the shared learnable vector. \\(\\sqrt{}\\) - for \"Shared-vector\" represents using the shared / independent learnable vector for the forced prompt; \\(\\sqrt{}\\) - for \"M-Init\" represents using manual / random initialization for the forced prompt." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.618, + 0.634, + 0.634 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this study, we propose a novel CLIP-based framework for OOD detection based on Forced prompt leArning (FA), which focuses on fully exploiting the ID knowledge to effectively improve OOD detection without exploring complex OOD-related knowledge. We introduce a learnable forced prompt in addition to the frozen original prompt, both of which are using the same manual initialization. The forced prompt treats the original prompt as a reference, forcing itself to learn more diversified semantic descriptions of the ID classes rather than being limited to the textual semantics of class labels. We also introduce a forced coefficient to facilitate the forced prompt in learning more nuanced descriptions of the ID classes. Comprehensive experimental evaluations demonstrate that our method consistently surpasses current state-of-the-art methods on diverse benchmarks. We expect that our study can bring new insight on VLMs-based OOD detection methods and inspire more future research." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1159" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.27, + 0.108 + ], + "angle": 0, + "content": "7. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.484, + 0.191 + ], + "angle": 0, + "content": "This work is supported in part by the National Natural Science Foundation of China (grant No. 62071502), the Major Key Project of PCL (grant No. PCL2023A09), and Guangdong Excellent Youth Team Program (grant No. 2023B1515040025)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.204, + 0.188, + 0.219 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.228, + 0.484, + 0.284 + ], + "angle": 0, + "content": "[1] Yichen Bai, Zongbo Han, Bing Cao, Xiaoheng Jiang, Qinghua Hu, and Changqing Zhang. Id-like prompt learning for few-shot out-of-distribution detection. In CVPR, pages 17480-17489, 2024. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.285, + 0.484, + 0.327 + ], + "angle": 0, + "content": "[2] Julian Bitterwolf, Maximilian Müller, and Matthias Hein. In or out? fixing imagenet out-of-distribution detection evaluation. In ICML, pages 2471-2506, 2023. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.328, + 0.483, + 0.369 + ], + "angle": 0, + "content": "[3] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 - mining discriminative components with random forests. In ECCV, pages 446-461, 2014. 5, 7, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.37, + 0.482, + 0.423 + ], + "angle": 0, + "content": "[4] Guangyi Chen, Weiran Yao, Xiangchen Song, Xinyue Li, Yongming Rao, and Kun Zhang. PLOT: prompt learning with optimal transport for vision-language models. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.425, + 0.483, + 0.467 + ], + "angle": 0, + "content": "[5] Mengyuan Chen, Junyu Gao, and Changsheng Xu. Conjugated semantic pool improves OOD detection with pretrained vision-language models. In NeurIPS, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.468, + 0.483, + 0.509 + ], + "angle": 0, + "content": "[6] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In CVPR, pages 3606-3613, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.51, + 0.483, + 0.551 + ], + "angle": 0, + "content": "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255, 2009. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.552, + 0.483, + 0.634 + ], + "angle": 0, + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.635, + 0.483, + 0.676 + ], + "angle": 0, + "content": "[9] Xuefeng Du, Zhaoning Wang, Mu Cai, and Yixuan Li. VOS: learning what you don't know by virtual outlier synthesis. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.678, + 0.483, + 0.718 + ], + "angle": 0, + "content": "[10] Xuefeng Du, Yiyou Sun, Jerry Zhu, and Yixuan Li. Dream the impossible: Outlier imagination with diffusion models. NeurIPS, 36:60878-60901, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.72, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[11] Sepideh Esmaeilpour, Bing Liu, Eric Robertson, and Lei Shu. Zero-shot out-of-distribution detection based on the pre-trained model clip. In AAAI, pages 6568-6576, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.483, + 0.831 + ], + "angle": 0, + "content": "[12] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In CVPRW, page 178, 2004. 5, 7, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 5, 7" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.228, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[14] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In ICLR, 2017. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[15] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In CVPR, pages 15262-15271, 2021. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.246 + ], + "angle": 0, + "content": "[16] Dan Hendrycks, Steven Basart, Mantas Mazeika, Andy Zou, Joseph Kwon, Mohammadreza Mostajabi, Jacob Steinhardt, and Dawn Song. Scaling out-of-distribution detection for real-world settings. In ICML, pages 8759-8773, 2022. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[17] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alexander Shepard, Hartwig Adam, Pietro Perona, and Serge J. Belongie. The inaturalist species classification and detection dataset. In CVPR, pages 8769-8778, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[18] Yen-Chang Hsu, Yilin Shen, Hongxia Jin, and Zsolt Kira. Generalized ODIN: detecting out-of-distribution image without learning from out-of-distribution data. In CVPR, pages 10948-10957, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.905, + 0.418 + ], + "angle": 0, + "content": "[19] Rui Huang and Yixuan Li. MOS: towards scaling out-of-distribution detection for large semantic space. In CVPR, pages 8710-8719, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[20] Xue Jiang, Feng Liu, Zhen Fang, Hong Chen, Tongliang Liu, Feng Zheng, and Bo Han. Negative label guided OOD detection with pretrained vision-language models. In ICLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.518 + ], + "angle": 0, + "content": "[21] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCVW, pages 554–561, 2013. 5, 7, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[22] Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. In NeurIPS, pages 7167-7177, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.576, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[23] Tianqi Li, Guansong Pang, Xiao Bai, Wenjun Miao, and Jin Zheng. Learning transferable negative prompts for out-of-distribution detection. In CVPR, pages 17584-17594, 2024. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.905, + 0.674 + ], + "angle": 0, + "content": "[24] Shiyu Liang, Yixuan Li, and R. Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. In ICLR, 2018. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.676, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[25] Jiashuo Liu, Zheyan Shen, Yue He, Xingxuan Zhang, Renzhe Xu, Han Yu, and Peng Cui. Towards out-of-distribution generalization: A survey. arXiv preprint arXiv:2108.13624, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[26] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM computing surveys, 55(9): 1-35, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[27] Weitang Liu, Xiaoyun Wang, John D. Owens, and Yixuan Li. Energy-based out-of-distribution detection. In NeurIPS, 2020. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[28] Shuo Lu, Yingsheng Wang, Lijun Sheng, Aihua Zheng, Lingxiao He, and Jian Liang. Recent advances in ood detection: Problems and approaches. arXiv preprint arXiv:2409.11884, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1160" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[29] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 5, 7, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.482, + 0.207 + ], + "angle": 0, + "content": "[30] Yifei Ming, Ziyang Cai, Jiuming Gu, Yiyou Sun, Wei Li, and Yixuan Li. Delving into out-of-distribution detection with vision-language representations. In NeurIPS, 2022. 1, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[31] Yifei Ming, Yiyou Sun, Ousmane Dia, and Yixuan Li. How to exploit hyperspherical embeddings for out-of-distribution detection? In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.251, + 0.482, + 0.307 + ], + "angle": 0, + "content": "[32] Atsuyuki Miyai, Qing Yu, Go Irie, and Kiyoharu Aizawa. Zero-shot in-distribution detection in multi-object settings using vision-language foundation models. arXiv preprint arXiv:2304.04521, 2023. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.309, + 0.482, + 0.352 + ], + "angle": 0, + "content": "[33] Atsuyuki Miyai, Qing Yu, Go Irie, and Kiyoharu Aizawa. Locoop: Few-shot out-of-distribution detection via prompt learning. In NeurIPS, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.353, + 0.482, + 0.395 + ], + "angle": 0, + "content": "[34] Jun Nie, Yonggang Zhang, Zhen Fang, Tongliang Liu, Bo Han, and Xinmei Tian. Out-of-distribution detection with negative prompts. In ICLR, 2024. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.397, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[35] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729, 2008. 5, 7, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.482, + 0.495 + ], + "angle": 0, + "content": "[36] Omkar M. Parkhi, Andreae Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In CVPR, pages 3498-3505, 2012. 5, 7, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.498, + 0.482, + 0.581 + ], + "angle": 0, + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.583, + 0.482, + 0.652 + ], + "angle": 0, + "content": "[38] Mohammadreza Salehi, Hossein Mirzaei, Dan Hendrycks, Yixuan Li, Mohammad Hossein Rohban, and Mohammad Sabokrou. A unified survey on anomaly, novelty, open-set, and out-of-distribution detection: Solutions and future challenges. arXiv preprint arXiv:2110.14051, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.655, + 0.482, + 0.695 + ], + "angle": 0, + "content": "[39] Vikash Sehwag, Mung Chiang, and Prateek Mittal. SSD: A unified framework for self-supervised outlier detection. In ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.698, + 0.482, + 0.739 + ], + "angle": 0, + "content": "[40] K Soomro. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012.5,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.742, + 0.482, + 0.784 + ], + "angle": 0, + "content": "[41] Yiyou Sun, Chuan Guo, and Yixuan Li. React: Out-of-distribution detection with rectified activations. In NeurIPS, pages 144-157, 2021. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.786, + 0.482, + 0.828 + ], + "angle": 0, + "content": "[42] Yiyou Sun, Yifei Ming, Xiaojin Zhu, and Yixuan Li. Out-of-distribution detection with deep nearest neighbors. In ICML, pages 20827-20840, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.83, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[43] Leitian Tao, Xuefeng Du, Jerry Zhu, and Yixuan Li. Nonparametric outlier synthesis. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[44] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Open-set recognition: A good closed-set classifier is all you need. In ICLR, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[45] Haoqi Wang, Zhizhong Li, Litong Feng, and Wayne Zhang. Vim: Out-of-distribution with virtual-logit matching. In CVPR, pages 4911-4920, 2022. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.136, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[46] Hualiang Wang, Yi Li, Huifeng Yao, and Xiaomeng Li. CLIPN for zero-shot OOD detection: Teaching CLIP to say no. In ICCV, pages 1802-1812, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.179, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[47] Jianxiong Xiao, James Hays, Krista A. Ehinger, Aude Oliva, and Antonio Torralba. SUN database: Large-scale scene recognition from abbey to zoo. In CVPR, pages 3485-3492, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.235, + 0.905, + 0.276 + ], + "angle": 0, + "content": "[48] Geng Yu, Jianing Zhu, Jiangchao Yao, and Bo Han. Self-calibrated tuning of vision-language models for out-of-distribution detection. In NeurIPS, 2024. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.278, + 0.905, + 0.32 + ], + "angle": 0, + "content": "[49] Tao Yu, Zhihe Lu, Xin Jin, Zhibo Chen, and Xinchao Wang. Task residual for tuning vision-language models. In CVPR, pages 10899-10909, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.321, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[50] Yeonguk Yu, Sungho Shin, Seongju Lee, Changyun Jun, and Kyoobin Lee. Block selection method for using feature norm in out-of-distribution detection. In CVPR, pages 15701-15711, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.377, + 0.905, + 0.433 + ], + "angle": 0, + "content": "[51] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kun-chang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of CLIP for few-shot classification. In ECCV, pages 493-510, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.434, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[52] Bolei Zhou, Ágata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE TPAMI, 40(6):1452-1464, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.49, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[53] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In CVPR, pages 16795-16804, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.533, + 0.905, + 0.575 + ], + "angle": 0, + "content": "[54] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 130(9):2337-2348, 2022. 2, 3, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.576, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[55] Yibo Zhou. Rethinking reconstruction autoencoder-based out-of-distribution detection. In CVPR, pages 7379-7387, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.616 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.956 + ], + "angle": 0, + "content": "1161" + } + ] +] \ No newline at end of file diff --git a/2025/FA_ Forced Prompt Learning of Vision-Language Models for Out-of-Distribution Detection/8ae5a438-b943-4e71-b751-475d9f24c3b7_origin.pdf b/2025/FA_ Forced Prompt Learning of Vision-Language Models for Out-of-Distribution Detection/8ae5a438-b943-4e71-b751-475d9f24c3b7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..811287a97da5f3209660d522743cb2e69cd746ea --- /dev/null +++ b/2025/FA_ Forced Prompt Learning of Vision-Language Models for Out-of-Distribution Detection/8ae5a438-b943-4e71-b751-475d9f24c3b7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:18e38caaeb6173bd843f51c3a5ab7517257dac1fedb30793b3fa99e78ce9ffff +size 507530 diff --git a/2025/FA_ Forced Prompt Learning of Vision-Language Models for Out-of-Distribution Detection/full.md b/2025/FA_ Forced Prompt Learning of Vision-Language Models for Out-of-Distribution Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..dbe0b2a5ccc07689b98c664e87dcb241f58afc57 --- /dev/null +++ b/2025/FA_ Forced Prompt Learning of Vision-Language Models for Out-of-Distribution Detection/full.md @@ -0,0 +1,278 @@ +# FA: Forced Prompt Learning of Vision-Language Models for Out-of-Distribution Detection + +Xinhua Lu $^{1,2,4}$ , Runhe Lai $^{1,2,4}$ , Yanqi Wu $^{1,2,4}$ , Kanghao Chen $^{3\dagger}$ , Wei-Shi Zheng $^{1,2,4}$ , Ruixuan Wang $^{1,2,4\dagger}$ $^{1}$ Sun Yat-sen University $^{2}$ Peng Cheng Laboratory + $^{3}$ Hong Kong University of Science and Technology (Guangzhou) + $^{4}$ Key Laboratory of Machine Intelligence and Advanced Computing + +{luxh55, lairh5, wuyq268}@mail2.sysu.edu.cn, kchen879@connect.hkust-gz.edu.cn, wszheng@ieee.org, wangruix5@mail.sysu.edu.cn + +# Abstract + +Pre-trained vision-language models (VLMs) have advanced out-of-distribution (OOD) detection recently. However, existing CLIP-based methods often focus on learning OOD-related knowledge to improve OOD detection, showing limited generalization or reliance on external largescale auxiliary datasets. In this study, instead of delving into the intricate OOD-related knowledge, we propose an innovative CLIP-based framework based on Forced prompt leArning (FA), designed to make full use of the InDistribution (ID) knowledge and ultimately boost the effectiveness of OOD detection. Our key insight is to learn a prompt (i.e. forced prompt) that contains more diversified and richer descriptions of the ID classes beyond the textual semantics of class labels. Specifically, it promotes better discernment for ID images, by forcing more notable semantic similarity between ID images and the learnable forced prompt. Moreover, we introduce a forced coefficient, encouraging the forced prompt to learn more comprehensive and nuanced descriptions of the ID classes. In this way, FA is capable of achieving notable improvements in OOD detection, even when trained without any external auxiliary datasets, while maintaining an identical number of trainable parameters as CoOp. Extensive empirical evaluations confirm our method consistently outperforms current state-of-the-art methods. Code is available at https://github.com/0xFIFA/FA. + +# 1. Introduction + +AI models often encounter Out-of-Distribution (OOD) samples [14, 25, 28, 38], which differ from the distribution of the training data, when deployed in real-world applications. + +![](images/d9ad327290586f4d05943d5523b636a0b2470f791793f567897944331faba407.jpg) +Figure 1. Comparison of CLIP-based OOD detection methods. Existing methods focus on OOD-related knowledge to learn a complex ID/OOD decision boundary as shown in the orange area in (b). In contrast, our FA aims to fully exploit ID knowledge by forcing the model to learn richer descriptions of the ID classes beyond the textual semantics of class labels. These richer descriptions of the ID classes compared to the reference text allow better discernment of ID/OOD samples (orange vs. blue area in (a)). + +Detecting OOD data is vital for the reliability of AI systems. This problem becomes harder in fields like intelligent diagnosis, where only a small amount of labeled data is available, motivating the problem of few-shot OOD detection. + +To tackle the OOD detection problem, previous single-modal post-hoc methods [14, 16, 24, 27, 41] achieve notable success based on a model pre-trained on the ID dataset. However, they come with limitations, such as requiring substantial computational resources and annotation costs for training, which hinder the performance of few-shot OOD detection. With the development of large pre-trained vision-language models (VLMs) like CLIP [37], few-shot OOD detection has achieved remarkable performance. Most of the CLIP-based OOD detection methods [30, 32] focus on + +leveraging the powerful generalization capability of CLIP to improve OOD detection. + +Recent CLIP-based OOD detection methods [1, 11, 23, 33, 34, 46, 48] aim to learn OOD-related knowledge to improve OOD detection performance. For example, recent CLIP-based works [11, 46] improve OOD detection by relying on training on external large-scale auxiliary datasets, which entails significant resource overhead. To eliminate dependence on external datasets, existing methods [1, 33, 48] learn OOD-related knowledge from exposed outliers based only on ID training data, such as background regions in ID training images. However, these OOD features extracted from specific regions are difficult to match with infinite OOD data encountered in practice. Moreover, some methods [23, 34] aim to learn negative prompts semantically opposite to the ID class labels, while these limited negative prompts are often insufficient to capture the distinctions between the diverse and numerous OOD data and the ID data. Overall, improving OOD detection by learning OOD-related knowledge shows inherent limitations or requires significant computational resources as well as labor costs. Inspired by the method [44] for open-set recognition, which demonstrates that enhancing closed-set accuracy can typically improve open-set recognition capabilities, we are motivated to explore ways to improve the identification of ID images, thereby enhancing OOD detection performance. + +In this paper, rather than focusing on the intricate OOD-related knowledge, we propose a novel CLIP-based framework based on Forced prompt leArning (FA), which aims to fully exploit the ID knowledge and ultimately improve OOD detection. Our key insight is to learn a prompt that contains richer knowledge beyond the textual semantics of class labels. It facilitates advanced discernment for ID images, by forcing higher semantic similarity between ID images and the learnable prompt (i.e. forced prompt), which provides more diversified and richer descriptions of the ID classes. In this way, FA can achieve significant improvements in OOD detection performance, even when trained without any additional external data, while maintaining the same number of learnable parameters as CoOp [54]. + +Specifically, we introduce a novel forced prompt along with the original prompt, both of which are initialized identically. FA preserves the generalization capability of the VLMs (i.e. CLIP) by freezing the original prompt and optimizing a trainable copy (i.e. forced prompt). In particular, the forced prompt treats the original prompt as a reference, while forcing the text features associated with the forced prompt to be more salient compared to those associated with the original prompt. In this way, it effectively improves the OOD performance even without sacrificing the ID classification capability of the model. Additionally, we introduce a forced coefficient, encouraging the forced + +prompt to learn more comprehensive connotations. Experimentally, FA achieves superior few-shot OOD detection performance across diverse OOD benchmarks. Compared to current SOTA methods [48] that focus on learning OOD-related knowledge, our method significantly reduces the average FPR95 score from $31.62\%$ to $27.81\%$ and improves the average AUROC from $92.01\%$ to $93.26\%$ even in 1-shot OOD detection on ImageNet-1k. The main contributions are summarized as follows: + +- We propose a simple yet effective framework, which fully exploits the ID knowledge to improve OOD detection without focusing on intricate OOD-related knowledge. +- We present a Forced prompt leArning (FA) strategy to exploit richer knowledge beyond the textual semantics of class labels. +- We evaluate our method on diverse OOD benchmarks, showing that our model consistently outperforms current state-of-the-art methods. + +# 2. Related work + +Prompt Learning. In recent years, pre-trained vision-language models (VLMs) such as CLIP [37] have demonstrated powerful few-shot learning capabilities in both visual and textual domains. However, the design of the prompt greatly affects the performance of VLMs for downstream tasks, which may require manually crafting numerous prompts. Inspired by prompt learning studies in natural language processing (NLP) [26], CoOp [54] uses a set of learnable context vectors to transfer CLIP for specific downstream tasks, becoming a pioneering method for subsequent studies [4, 33, 53]. Currently, prompt learning is widely used in OOD detection mainly to represent various OOD-related knowledge. In this work, we learn the forced prompt, which contains richer descriptions of the ID classes to improve OOD detection. + +Out-of-Distribution Detection. Conventional methods explore OOD detection for single-modal models. One line of studies designs score functions to differentiate ID and OOD data based on output from the pre-trained models, such as logit output [14, 16, 24, 27] or outputs from the penultimate layer [22, 41, 42, 50]. These methods are called post-hoc methods. Another line of studies [9, 10, 18, 31, 39, 43, 55] adopts various training strategies to learn a reliable decision boundary between ID and OOD data. + +Recently, the development of OOD detectors based on VLMs, especially CLIP, has received much attention as VLMs have demonstrated their remarkable generalization capability in both visual and textual domains. Some studies leverage real outlier information from external auxiliary OOD datasets [11, 46] or extensive corpora [5, 20] to promote OOD detection. However, this is impractical in real-world scenarios, where outliers are infinite and agnostic. + +![](images/d6c07d7693e54a2e5e7e9bf056c4470e042bca80776d2e49bb596b306ede69d4.jpg) +Figure 2. Overview of the proposed FA framework. Our framework includes the learnable forced prompt and the frozen original prompt, both of which are initialized identically by a manual template "a photo of a [class-c]". The learnable forced prompt treats the frozen original prompt as a reference, forcing its text features to become more salient compared to those generated by the original prompt. The richer and more diversified description of ID classes learned by FA can ultimately improve OOD detection. + +To eliminate the dependence on external data, recent methods [1, 23, 32, 33, 48] were developed under the assumption that only ID data are available. As a representative zero-shot method that only uses available ID labels, MCM [30] employs softmax scaling to align visual features with ID text concepts for OOD detection. GL-MCM [32] further introduces a local maximum concept matching(L-MCM) score to improve the separability of the local information. Compared with zero-shot methods, which may undergo a domain gap with ID downstream data, prompt learning methods achieve better OOD detection performance with access to few-shot ID samples. For prompt learning OOD detection methods, LoCoOp [33] and SCT [48] keep the textual embeddings of ID classes away from ID-irrelevant local region embeddings in the multi-modal embedding space by their proposed entropy maximization strategy, which ensures that OOD embeddings can be dissimilar to any textual embedding of ID classes. To detect challenging OOD samples to improve OOD detection, ID-like [1] explores the vicinity of ID samples to construct OOD samples correlated to the ID and refines the additional "ID-like" text embeddings to fine-grained differences. In addition, NegPrompt [23] and LSN [34] introduce the negative connotations of ID categories which can be represented by additional negative prompts, enabling more accurate detection of OOD samples. Overall, various OOD-related knowledge is commonly adopted to address the model's overconfidence in current research. In this work, we focus on leveraging ID knowledge to improve OOD detection, instead of exploring complex OOD-related knowledge. + +# 3. Preliminaries + +Prompt Learning with CLIP. CLIP contains an image encoder $f(\cdot)$ and a text encoder $g(\cdot)$ , designed to extract features from images and text descriptions respectively. Generally, for an ID dataset which contains $C$ categories, the hand-crafted prompt $\hat{\mathbf{u}}_c$ "a photo of a [class-c]" is designed to match images, where class-c represents the class name. Formally, these prompts are individually encoded in $\hat{\mathbf{t}}_c = g(\hat{\mathbf{u}}_c) \in \mathbb{R}^{d \times 1}$ , $c = 1, \dots, C$ . Given an image $\mathbf{x}$ , it can be encoded in $\mathbf{z} = f(\mathbf{x}) \in \mathbb{R}^{d \times 1}$ . Because image features and corresponding text features are aligned on the multimodal embedding space, the cosine similarity $\cos(\mathbf{z}, \hat{\mathbf{t}}_c)$ of $\mathbf{z}$ and $\hat{\mathbf{t}}_c$ can represent the matching degree between the image $\mathbf{x}$ and text description $\hat{\mathbf{u}}_c$ . Therefore, if the text feature $\mathbf{t}_c$ of class $c$ has the highest similarity to $\mathbf{z}$ , the image $\mathbf{x}$ will be considered to be class $c$ . + +To effectively transfer CLIP to downstream image recognition tasks, CoOp [54] sets a portion of tokens in the text prompt as continuous learnable parameters. Concretely, CoOp initializes the text prompts as $\mathbf{u}_c = [\mathbf{v}_1,\dots ,\mathbf{v}_L,\mathbf{w}_c],c\in \{1,\dots ,C\}$ , where $L$ is the length of the prompt's tokens, $\mathbf{v}_i$ ( $i\in \{1,\ldots ,L\}$ ) is the learnable vector with the same dimension as the word embedding and $\mathbf{w}_c$ is the word embedding for the class name of class $c$ . The text prompt $\mathbf{u}_c$ is encoded into features $\mathbf{t}_c = g(\mathbf{u}_c)$ . Together with the image feature $\mathbf{z}$ , the probability for the image $\mathbf{x}$ to be classified into class $c$ is defined as + +$$ +p (y = c \mid \mathbf {x}) = \frac {e ^ {\cos (\mathbf {z} , \mathbf {t} _ {c}) / \tau}}{\sum_ {j = 1} ^ {C} e ^ {\cos (\mathbf {z} , \mathbf {t} _ {j}) / \tau}}, \tag {1} +$$ + +where $\tau$ is a fixed temperature scaling hyper-parameter. + +Overall, the prompt's tokens can be trained to align with training data by minimizing the cross-entropy loss with the prediction probability from Eq. (1). + +OOD Detection. In the OOD detection task, the model is expected to determine whether a test image belongs to one of the learned ID classes. Therefore, OOD detection can be seen as a binary classification problem to distinguish ID images from OOD images in a test set $\mathcal{D}_{test}$ as follows + +$$ +D (\mathbf {x}) = \left\{ \begin{array}{l l} 1, & \text {i f} S (\mathbf {x}) \geq \mu \\ 0, & \text {i f} S (\mathbf {x}) < \mu \end{array} , \right. \tag {2} +$$ + +where $\mathbf{x} \in \mathcal{D}_{test}$ , 1 and 0 respectively indicate that $\mathbf{x}$ is classified as the ID class and the OOD class by the OOD detector $D(\cdot)$ , $S(\cdot)$ is certain score function, and $\mu$ is a predefined threshold constant. + +# 4. Methodology + +# 4.1. Overview + +In this study, we propose a novel CLIP-based framework based on Forced prompt leArning (FA), designed to fully exploit the ID knowledge from few-shot ID samples by forcing the prompt to learn richer knowledge beyond the textual semantics of class labels. As shown in Fig. 2, we introduce a forced prompt along with the original prompt, both of which are initialized with the same semantics. By freezing the original prompt and optimizing a trainable copy, ID images will show higher semantic similarity to the forced prompt compared to the original prompt. Moreover, to encourage descriptions of ID classes learned by the forced prompt to become more detailed and comprehensive, we also introduce a forced coefficient. During testing, the model is capable of effectively distinguishing between ID data and OOD data, benefiting from the diversified descriptions of the ID classes provided by the forced prompt. + +# 4.2. Forced prompt learning + +Merely the semantic information conveyed by class names is insufficient to comprehensively encompass the discriminative information of each class. In order to explore richer semantic descriptions of the ID classes beyond the textual semantics of class labels, a novel forced prompt along with the original prompt is proposed. To implement the above objective, we introduce a simple yet effective training strategy using the forced cross-entropy (FCE) loss based on the forced prompt as follows + +$$ +\mathcal {L} _ {F C E} = \mathbb {E} _ {\left(\mathbf {x}, y _ {c}\right) \sim \mathcal {D} _ {t r a i n} ^ {I D}} \left[ - \log \frac {e ^ {s _ {c} ^ {f} / \tau}}{\sum_ {j = 1} ^ {C} e ^ {s _ {j} ^ {f} / \tau} + \sum_ {j = 1} ^ {C} e ^ {s _ {j} ^ {o} / \tau}} \right], \tag {3} +$$ + +where the ID training dataset $\mathcal{D}_{train}^{ID}$ consists of ID image-label pairs $(\mathbf{x},y_{c})$ , $s_j^f = \cos (\mathbf{z},\mathbf{t}_j^f)$ and $s_j^o = \cos (\mathbf{z},\mathbf{t}_j^o)$ + +represent the similarity between the image feature and the prompt feature corresponding to the forced prompt and the original prompt, respectively. + +Based on this loss function, the ID image features will be forced to show higher cosine similarity to the learnable forced prompt compared to the original text prompt. This is because the cosine similarity between the ID image features and the original text prompt is already high before any further prompt learning. Consequently, the forced prompt is forced to uncover richer discriminative information of ID classes, so that the text features of the forced prompt achieve a higher cosine similarity with the image features compared to those of the original prompt. Benefiting from this information specific to the ID classes, the cosine similarity between the image features of an OOD image and the text features of both prompts may exhibit less distinction compared to those of the ID data. Therefore, even when trained without any reliance on external auxiliary datasets, the detector will have a notable capability to distinguish ID data from OOD data. + +However, this naive FCE loss may not work effectively because of its inherent limitations. Concretely, employing random initialization for both prompts fails to fully leverage the semantic information provided by CLIP's prior knowledge, resulting in limited performance. To solve this limitation, we propose to initialize the forced prompt using the manual template (i.e. "a photo of a [class-c]") identified to the original prompt. In this way, both prompts with manual initialization will have clear semantic information compared to random initialization, thereby improving the model's generalization capability. Specifically, the class labels are utilized in both prompts by concatenating with the prompt embedding to provide foundational semantic information related to the category names, based on CLIP's prior knowledge. In particular, for the forced prompt, we adopt the shared learnable vector across all classes rather than the independent learnable vector for each class. This choice is inspired by CoOp [54], which demonstrates that using an independent learnable vector mostly underperforms the shared learnable vector in challenging low-data scenarios since the former has more parameters and requires more data for training. Formally, we utilize the embeddings of the manual template to initialize the prompt's embeddings of both prompts, formulated as $\mathbf{u}_c = [\mathbf{v}_1,\dots ,\mathbf{v}_L,\mathbf{w}_c], c\in$ $\{1,\dots ,C\}$ , which will be fed to the text encoder to obtain prompt feature $\mathbf{t}_c$ (see Section 3). Here, $L$ is the length of the token (e.g., for the "a photo of a", $L = 4$ ). Then we freeze the original prompt and the class label part of both the prompts to preserve the generalization capability of CLIP, while only making the forced prompt learnable. Note that our method effectively improves OOD detection performance while maintaining consistency with CoOp in the number of learnable parameters without leveraging ad + +ditional learnable prompts, unlike existing work [1, 23, 34]. + +Although the above prompt design achieves moderate performance, relying solely on a single original prompt as a reference is somewhat insufficient to develop a comprehensive capability, particularly with the limitations of ID data (e.g., few-shot OOD). To encourage the forced prompt to capture more comprehensive and nuanced descriptions of the ID classes, we introduce a forced coefficient $K$ ( $K \geq 0, K \in \mathbb{N}$ ), which indicates the intensity with which the model is compelled to learn from the data. Formally, we first derive the original text features $\{\mathbf{t}_1^o, \dots, \mathbf{t}_C^o\}$ from the original prompt and the image features $\mathbf{z}$ from the corresponding image. Then, we compute the cosine similarity between the original text features and the image features for $K$ iterations, which will be used in the subsequent computation of the cross-entropy loss function. Notably, when $K = 0$ , no original prompt is used, and the model operates equivalently to CoOp. Based on the forced coefficient, we refine Eq. (3) as follows + +$$ +\mathcal {L} _ {F C E - K} = \mathbb {E} _ {\left(\mathbf {x}, y _ {c}\right) \sim \mathcal {D} _ {t r a i n} ^ {I D}} \left[ - \log \frac {e ^ {s _ {c} ^ {f} / \tau}}{\sum_ {j = 1} ^ {C} e ^ {s _ {j} ^ {f} / \tau} + K \sum_ {j = 1} ^ {C} e ^ {s _ {j} ^ {o} / \tau}} \right]. \tag {4} +$$ + +Based on this function, the cosine similarity between the ID image features and the text features of the forced prompt must become more salient as the coefficient $K$ increases. + +# 4.3. Model inference + +During model inference, for the downstream classification task, we adopt the same strategy as CLIP, depending solely on the forced prompt [1]. For the OOD detection task, our method can be flexibly combined with different score functions, such as the MCM [30] score and the GL-MCM [32] score. The MCM score is defined as the maximum similarity between the global image features $\mathbf{z}^g$ (i.e. $\mathbf{z}$ ) and all text features $\mathbf{t}_c^a$ (i.e. the concatenation of text features $\mathbf{t}_c^f$ and $\mathbf{t}_c^o$ ) after applying softmax with temperature $\tau_0$ , i.e., + +$$ +S _ {\mathrm {M C M}} (\mathbf {x}) = \max _ {c} \frac {e ^ {\cos \left(\mathbf {z} ^ {g} , \mathbf {t} _ {c} ^ {a}\right) / \tau_ {0}}}{\sum_ {j = 1} ^ {C} e ^ {\cos \left(\mathbf {z} ^ {g} , \mathbf {t} _ {j} ^ {f}\right) / \tau_ {0}} + K e ^ {\cos \left(\mathbf {z} ^ {g} , \mathbf {t} _ {j} ^ {o}\right) / \tau_ {0}}}, \tag {5} +$$ + +while the GL-MCM score simultaneously considers both global and local features, which can be expressed as + +$$ +S _ {\mathrm {G L - M C M}} (\mathbf {x}) = S _ {\mathrm {M C M}} (\mathbf {x}) + S _ {\mathrm {L - M C M}} (\mathbf {x}), \tag {6} +$$ + +$$ +S _ {\mathrm {L - M C M}} (\mathbf {x}) = \max _ {i, c} \frac {e ^ {\cos \left(\mathbf {z} _ {i} ^ {l} , \mathbf {t} _ {c} ^ {a}\right) / \tau_ {0}}}{\sum_ {j = 1} ^ {C} e ^ {\cos \left(\mathbf {z} _ {i} ^ {l} , \mathbf {t} _ {j} ^ {f}\right) / \tau_ {0}} + K e ^ {\cos \left(\mathbf {z} _ {i} ^ {l} , \mathbf {t} _ {j} ^ {o}\right) / \tau_ {0}}}, \tag {7} +$$ + +where $\mathbf{z}_i^l (i\in \{1,\dots ,N\})$ represents $N$ extracted local features generated by CLIP's image encoder [32, 33]. We set $\tau_0 = 1$ during model inference. + +# 5. Experiments + +# 5.1. Experimental details + +Datasets. We use a popular benchmark for conventional OOD detection [19, 30, 32, 33], where ImageNet-1k [7] serves as the ID dataset, and the OOD datasets are iNaturalist [17], SUN [47], Places [52], and Texture [6]. Moreover, inspired by existing studies [1, 2], we use datasets like OpenImage-O [45], NINCO [2], and ImageNet-O [15], which are cleaner and more realistic, to simulate the more challenging OOD detection. Besides, we also utilize other widely adopted datasets for few-shot settings as ID datasets [49, 51, 54], which include StanfordCars [21], UCF101 [40], Caltech101 [12], Flowers102 [35], EuroSAT [13], FGVCAircraft [29], OxfordPets [36], and Food101 [3], considering factors such as image resolution and the number of classes. + +Setup. Following previous studies [33], we use the ViTB/16 [8] as the backbone model. For our model, when using ImageNet-1k as the ID dataset, the epochs are respectively set to 30 and 50 for the 1-shot and 16-shot settings, while for other ID datasets, we set the epoch to 200 following CoOp [54]. The value of forced coefficient $K$ is uniformly set to 3, which will be further discussed in the sensitivity study. Other hyperparameters are as follows: learning rate $= 2\mathrm{e} - 3$ , batch size $= 160$ , SGD (momentum $= 0.9$ , weight decay $= 5\mathrm{e} - 4$ ) as the optimizer with a cosine scheduler, $\tau = 1$ in Eq. (3) and Eq. (4). All experiments on our model can be conducted on a single Nvidia A30 GPU. The average experiment results (including our reproductions) over four runs are reported for comparison. + +Comparison Methods. To validate the effectiveness of our FA fairly, we mainly compare it with CLIP-based OOD detection methods that do not use real outliers (e.g. OOD labels). For previous post-hoc methods, including MSP [14], ODIN [24], Energy [27], ReAct [41], and MaxLogit [16], we adapt these methods with the CLIP image encoder as the CLIP-based post-hoc methods. For zero-shot methods, we select MCM [30], GL-MCM [32] and CLIPN [46] as baselines. For prompt learning methods, we adopt CoOp [54], LoCoOp [33], IDLike [1], and SCT [48] as baselines. + +Metrics. For evaluation, we adopt the following metrics: (1) The False Positive Rate at $95\%$ True Positive Rate for in-distribution samples (FPR95); (2) The Area Under the Receiver Operating Characteristic Curve (AUROC); (3) In-distribution data classification Top-1 accuracy (ID ACC). + +# 5.2. Main results + +Conventional OOD Detection. Table 1 summarizes our comparison results on the ImageNet-1k benchmarks, which show that our FA achieves state-of-the-art OOD detection performance among other CLIP-based methods under different few-shot settings (More details under the 4-shot set + +
MethodiNaturalistSUNPlacesTexturesAverage
FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑
Zero-shot methods
MCM31.9594.1637.2292.5542.9890.1058.3585.8342.6390.66
GL-MCM15.0996.7229.0893.4137.0790.3758.9483.1135.0490.90
CLIPN19.1796.1726.4394.0232.2692.6241.2390.1230.2193.19
CLIP-based post-hoc methods
\( MSP^† \)74.5777.7476.9573.9779.7272.1873.6674.8476.2274.68
\( ODIN^† \)98.9357.7388.7278.4287.8076.8885.4771.4990.2371.13
\( Energy^† \)64.9887.1846.4291.1757.4087.3350.3988.2254.8088.48
\( ReAct^† \)65.5786.8746.1791.0456.8587.4249.8888.1354.6288.37
\( MaxLogit^† \)60.8888.0344.8391.1655.5487.4548.7288.6352.4988.82
Prompt learning based methods
\( CoOp_{MCM} \)41.14±9.3991.47±2.1239.06±4.1791.68±0.8345.38±4.8889.16±1.2851.37±3.0987.82±1.1244.24±1.3990.03±0.32
\( CoOp_{GL} \)23.30±6.3794.59±1.5132.08±3.8792.21±0.9539.22±4.5589.59±1.5855.78±3.4683.26±1.3837.59±0.3489.91±0.31
\( LoCoOp_{MCM} \)36.64±4.8792.78±0.9331.86±3.7993.50±0.9238.81±3.3790.70±0.9748.11±2.3189.43±0.6838.85±2.6791.59±0.58
\( LoCoOp_{GL} \)21.97±2.9195.39±0.5924.95±2.3794.42±0.5934.14±2.6691.14±0.6949.04±2.7787.73±0.9632.53±2.1992.17±0.47
IDLike17.73±1.9196.68±0.3148.17±1.3989.53±0.5850.43±6.5888.27±2.2729.12±7.6493.25±2.1636.36±3.8691.93±1.18
\( LSN^† \)59.28±7.0287.20±3.1540.15±0.8291.47±0.1446.11±1.8688.74±0.5760.34±0.1483.92±0.4251.47±1.5387.84±0.58
\( NegPrompt^† \)65.03±8.6984.56±2.5244.39±1.6689.63±0.6651.31±6.2186.55±2.1987.60±1.6163.76±3.0262.08±3.7181.13±1.78
\( SCT_{MCM} \)41.93±12.1791.77±2.4030.39±2.1793.76±0.3938.73±2.5490.78±0.3446.78±3.5188.96±1.1639.46±3.3991.32±0.90
\( SCT_{GL} \)20.57±10.295.63±1.8924.56±3.0394.39±0.5333.27±2.9691.27±0.5748.12±2.9786.76±0.8931.62±3.1992.01±0.77
\( FA_{MCM}(Ours) \)25.50±2.7294.72±0.2736.24±3.1592.40±0.9535.38±2.7591.99±0.6828.34±0.8793.95±0.1931.37±1.0793.25±0.32
\( FA_{GL}(Ours) \)14.12±1.3296.76±0.1029.99±1.5792.95±0.6632.48±1.4891.83±0.4934.66±1.2191.50±0.3627.81±0.4493.26±0.27
Prompt learning based methods
\( CoOp_{MCM} \)30.26±1.9893.43±0.8134.69±0.4392.59±0.1441.91±0.7190.11±0.2344.68±2.1189.95±0.4837.89±0.7191.52±0.29
\( CoOp_{GL} \)15.96±1.6796.11±0.5527.26±1.9993.29±0.4635.36±2.0890.58±0.6448.63±2.1186.11±0.5931.81±1.2791.51±0.39
\( LoCoOp_{MCM} \)27.35±3.1994.12±0.8030.93±1.1993.75±0.2638.26±1.5391.12±0.2741.36±2.5690.99±0.5334.47±0.7392.49±0.14
\( LoCoOp_{GL} \)18.46±1.3895.85±0.6322.43±0.9695.15±0.2231.53±1.5292.15±0.2243.35±3.1289.38±0.7828.94±1.2993.13±0.17
IDLike19.23±10.596.70±1.5854.15±2.8887.64±1.1956.63±0.0785.86±0.4434.69±6.4191.90±2.3141.18±1.7390.53±0.01
\( LSN^† \)36.17±4.8192.66±1.1634.27±0.4493.53±0.2041.47±0.8590.52±0.3746.43±0.6089.38±0.2439.58±0.7391.53±0.09
\( NegPrompt^† \)37.79±0.1190.49±0.0132.11±3.7792.25±1.0035.52±0.4191.16±0.0343.93±9.0988.38±3.3137.34±1.4190.57±0.59
\( SCT_{MCM} \)29.41±2.1993.76±0.5527.28±2.8094.22±0.4036.35±2.1391.16±0.3342.25±1.8990.52±0.4833.82±1.7892.42±0.35
\( SCT_{GL} \)15.19±2.1696.71±0.4420.00±0.6195.57±0.1129.71±0.8592.37±0.1244.17±0.8288.59±0.3927.27±0.4493.31±0.17
\( FA_{MCM}(Ours) \)25.79±1.4894.29±0.3533.54±1.1793.02±0.1933.77±1.6492.64±0.4223.17±1.3195.14±0.2629.07±1.1193.77±0.19
\( FA_{GL}(Ours) \)14.49±1.2796.48±0.2927.65±1.0893.46±0.1831.09±1.3892.44±0.3429.50±0.8992.93±0.1825.68±0.5893.82±0.11
+ +Table 1. Comparison results on ImageNet-1k OOD benchmarks. All results using the same backbone ViT-B/16. The results marked with $\dagger$ are taken from [48]; the others are our reproductions. The prompt learning based methods are run under four trials, reporting the mean and standard deviation of the performance. The subscripts ${}_{\mathrm{{MCM}}}$ and ${}_{\mathrm{{GL}}}$ indicate the use of the MCM score and the GL-MCM score. The best and second-best results are indicated in bold and underline. $\uparrow$ indicates larger values are better; $\downarrow$ indicates smaller values are better. All values are percentages. + +
Method1-shot4-shot16-shot
CoOp67.44±0.5069.71±0.0770.99±0.14
LoCoOp67.40±0.6469.55±0.1071.53±0.17
IDLike68.17±0.5768.91±0.1469.46±0.02
SCT68.63±0.1369.93±0.2271.78±0.05
FA(Ours)68.67±0.3969.96±0.0471.02±0.09
+ +Table 2. Comparison results in ID Top-1 accuracy on ImageNet-1k under different few-shot settings. Other notations are the same as Tab. 1. + +ting are provided in the Appendix A). Specifically, in the 16-shot setting, FA with GL-MCM score $\left(\mathrm{FA}_{\mathrm{GL}}\right)$ outperforms the best baseline $\mathrm{SCT}_{\mathrm{GL}}$ (average FPR95 of $25.68\%$ vs. $27.27\%$ , and average AUROC of $93.82\%$ vs. $93.31\%$ ). + +Additionally, in the 1-shot setting, $\mathrm{FA}_{\mathrm{GL}}$ surpasses $\mathrm{SCT}_{\mathrm{GL}}$ by a large margin, showing improvements of at least $1.25\%$ and $3.81\%$ on average AUROC and FPR95. More importantly, even the average score of $\mathrm{FA}_{\mathrm{MCM}}$ has exceeded $\mathrm{SCT}_{\mathrm{GL}}$ in the 1-shot setting. Moreover, our FA effectively improves the OOD performance without sacrificing the ID classification capability of the model, as shown in Tab. 2. + +In particular, we can also observe that our method performs not well on the SUN dataset, as shown in Tab. 1. This is related to the presence of a significant number of samples belonging to ID category in these datasets, as demonstrated by recent studies [1]. In other words, the SUN dataset in the popular OOD benchmark requires more thorough cleaning to be effectively used as an OOD dataset. + +Challenging OOD Detection. Recent studies [1, 2] have + +
MethodOpenImage-ONINCOImageNet-OAverage
FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑
CoOpMCM37.95±0.5991.99±0.2278.00±0.2373.00±0.5970.14±0.8681.34±0.8362.03±0.3882.11±0.48
CoOpGL32.23±1.0092.22±0.2572.04±0.8175.35±0.8867.35±1.1978.44±0.9457.21±0.9681.99±0.58
LoCoOpMCM37.03±0.9492.24±0.2177.96±0.6372.78±0.7970.96±0.7581.25±0.3961.98±0.5582.09±0.35
LoCoOpGL33.87±1.2392.53±0.0775.16±0.9272.97±0.3968.06±1.9281.19±0.2559.03±1.2782.23±0.17
IDLike54.43±2.3587.89±0.4578.93±1.7969.32±0.4284.08±1.4567.45±0.8672.48±0.8974.89±0.28
SCTMCM37.28±1.1692.04±0.3378.51±0.8671.09±1.1971.45±0.5481.29±0.1562.41±0.7381.48±0.52
SCTGL32.24±0.3492.75±0.1774.15±0.5373.14±0.9468.35±0.6280.84±0.4558.25±0.3982.24±0.44
FAmCM(Ours)37.93±1.6591.84±0.4570.44±1.2477.77±0.9463.84±1.3682.71±0.3057.40±1.3484.08±0.47
FAGL(Ours)32.10±0.8992.39±0.2865.61±1.3279.01±0.7163.13±0.9880.39±0.2153.61±0.8983.93±0.31
+ +Table 3. Challenging OOD detection results on cleaner OOD datasets in the 16-shot setting. We use the same notation as Tab. 1. + +
ID DatasetMethodiNaturalistSUNPlacesTexturesAverage
FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑
UCF101CoOpMCM14.40±6.6197.41±1.3637.09±3.9493.14±0.4840.07±2.7791.89±0.6533.24±4.2193.19±1.2231.21±3.5893.91±0.45
CoOpGL9.65±6.3598.18±0.9828.54±5.0794.51±0.9232.94±1.7493.28±0.3235.96±2.6591.83±1.4226.77±3.2294.45±0.42
LoCoOpMCM2.38±0.9899.21±0.2219.73±4.2496.11±0.7822.57±3.0194.96±0.7715.89±3.1296.12±0.6715.14±1.9196.60±0.33
LoCoOpGL1.13±0.2299.65±0.0917.93±4.0996.51±0.8922.34±4.5895.03±1.0618.17±3.0195.25±0.7114.89±2.4996.61±0.58
IDLike66.63±27.689.75±5.5287.88±6.5678.04±6.1484.47±12.176.04±8.5349.07±11.291.19±2.7472.01±14.083.76±5.64
SCTMCM1.51±0.5199.44±0.1024.17±5.5495.21±0.8826.62±4.4193.75±0.8115.34±3.8495.98±0.6416.91±3.2096.09±0.48
SCTGL0.79±0.2799.75±0.0418.69±2.4195.86±0.5623.77±2.4793.98±0.4717.44±2.6294.52±0.5415.17±1.5696.03±0.25
FAmCM(Ours)0.12±0.1199.94±0.013.06±1.3399.34±0.234.68±1.0099.08±0.262.68±0.5199.31±0.432.63±0.5899.42±0.22
FAGL(Ours)0.13±0.0299.94±0.024.98±1.0299.01±0.227.46±0.8698.44±0.275.67±0.8398.75±0.194.56±0.4799.03±0.13
EuroSATCoOpMCM99.14±0.7331.95±9.1798.29±0.4944.36±4.1997.28±0.8547.51±4.6386.69±4.3769.17±3.1595.35±1.0348.25±3.09
CoOpGL92.45±4.2657.59±10.490.97±3.4961.89±4.8388.69±3.8564.00±4.9262.37±1.9383.58±2.3183.62±1.1866.77±2.83
LoCoOpMCM96.51±3.3353.49±8.0693.52±5.6855.59±9.1992.88±4.0557.89±5.1976.45±7.4878.59±2.6689.84±4.2961.39±3.90
LoCoOpGL86.13±12.272.68±7.9787.81±6.3768.58±6.4085.93±4.1569.90±3.4054.89±4.6787.00±1.0178.69±4.6574.54±3.59
IDLike98.02±2.1373.32±5.8395.74±2.8465.19±6.8694.79±3.6769.44±5.0663.34±9.1086.68±2.9887.97±4.2173.66±5.04
SCTMCM99.79±0.2232.90±6.7398.36±0.8444.30±3.4197.77±0.8648.08±1.6483.98±6.9273.25±3.5894.97±2.1949.63±2.39
SCTGL96.57±5.4158.21±5.3192.41±4.5561.54±3.4790.49±5.0263.33±2.7155.29±14.785.48±3.6883.69±7.0267.14±2.20
FAmCM(Ours)85.24±6.8878.62±3.1836.24±10.892.13±2.4428.17±10.294.05±1.6810.36±3.9797.71±0.6939.99±7.7990.63±1.91
FAGL(Ours)77.82±7.2886.88±2.1556.10±8.0388.43±2.2649.87±7.6989.07±1.498.99±2.5197.74±0.3648.19±6.2390.53±1.42
+ +Table 4. Representative OOD detection results with various ID datasets in the 16-shot setting. We use the same notation as Tab. 1. + +![](images/076c5ec2f75ee8cd53e39c8b2940dcccfd6dbd0bae3b8b4fc0d38aacd046cb8c.jpg) +Figure 3. A portion of OOD detection average performance using other ID datasets under the 16-shot setting. We report the mean and standard deviation of the AUROC for our FA (blue) and CoOp (orange) using the MCM score. More details can be found in the Appendix B. + +shown that some of the used OOD datasets contain more or less images of objects that belong to ID classes in ImageNet-1k [7]. Hence, probing the true performance of OOD detectors for ImageNet-1k demands OOD datasets that are both challenging and genuinely OOD. As shown in Tab. 3, to fully substantiate the effectiveness of our + +FA, we leverage 3 additional cleaner and more challenging OOD datasets, including OpenImage-O [45], NINCO [2], and ImageNet-O [15], inspired by [2]. Notably, both our $\mathrm{FA}_{\mathrm{MCM}}$ and $\mathrm{FA}_{\mathrm{GL}}$ consistently outperform current SOTA methods in terms of average FPR95 and AUROC under different few-shot settings (More details under the 1-shot and 4-shot settings are provided in the Appendix A). + +Various ID Datasets. Considering factors like dataset variety, image resolution, and the number of classes [30], we also leverage 8 additional widely used datasets for few-shot setting as ID datasets for comparison [30, 34, 54]. Overall, despite our model is simple, it achieves superior average OOD performance on these ID datasets compared to other methods. We selected representative results using UCF101 [40] and EuroSAT [13] as ID datasets in the 16-shot setting, as shown in Tab. 4. Since using the other datasets (i.e. Food101 [3], StandfordCars [21], Caltech101 [12], FGVCAircraft [29], Flowers102 [35], and OxfordPets [36]) as ID datasets for OOD detection presents relatively lower difficulty as shown in Fig. 3, we leave more + +![](images/3179384cf6c5f32757ff2ec23c1096e85edf7724d0bee176cc0262fa0bb2070b.jpg) +(a) Conventional OOD benchmark + +![](images/473ae12adfd35b414a654564acef41ba335a1dae659dd6e9bbf391ca6c4e51ab.jpg) +(b)Challenging OOD benchmark +Figure 4. Sensitivity study of hyperparameter $K$ . We report the average FPR95 (Blue) and AUROC (Orange) of our $\mathrm{FA}_{\mathrm{GL}}$ on different OOD benchmarks using ImageNet-1k as the ID dataset under 16-shot setting. Dashed lines represent the performance of the best baseline (i.e. $\mathrm{SCT}_{\mathrm{GL}}$ ). + +experimental details in Appendix B. + +In particular, for the ID dataset EuroSAT, $\mathrm{FA}_{\mathrm{MCM}}$ surpasses the best-competing method $\mathrm{LoCoOp_{GL}}$ by $16.09\%$ in average AUROC and $38.7\%$ in average FPR95, underscoring its effectiveness and versatility even when the ID dataset is of low resolution (64x64 pixels). + +# 5.3. Ablation study + +Influence of the $\mathcal{L}_{FCE-K}$ . As shown in Tab. 5, we trained the model $\mathrm{FA}_{CE}$ and $\mathrm{FA}_{FCE-K}$ using the standard cross-entropy loss $\mathcal{L}_{CE}$ and our $\mathcal{L}_{FCE-K}$ , respectively. Our $\mathrm{FA}_{FCE-K}$ significantly outperforms $\mathrm{FA}_{CE}$ under the same inference setting, which demonstrate the effectiveness of our $\mathcal{L}_{FCE-K}$ . + +Initialization of the forced and original prompt. As shown in Tab. 6, we perform experiments comparing different initialization scenarios for both prompts. Manual initialization uses the embeddings of "a photo of a [class-c]" to initialize the prompt's embeddings, while the random initialization setting is based on CoOp [54]. The results show that employing consistent manual initialization for both prompts enables the forced prompt to effectively learn more specific descriptions of the ID classes. + +Influence of the shared learnable vector. We also explore the influence of the shared learnable vector and the independent learnable vector for the forced prompts. As shown in Tab. 7, we can observe that using the shared learnable vector performs well in the few-shot setting since the independent learnable vector for each class has more parameters and requires more training data [54]. + +# 5.4. Sensitivity study + +As shown in Fig. 4, we explore the influence of the hyperparameter $K$ . Overall, with $K$ ranging from 1 to 6, the results indicate that our proposed framework is insensitive to the choice of $K$ . Concretely, our method is stable and outperforms the best baseline $\mathrm{SCT}_{\mathrm{GL}}$ in average AUROC and FPR95 on both conventional and challenging OOD detection benchmarks. Specially, as $K$ increases, the average AUROC shows an upward trend and then tends to be sta + +ble (see Fig. 4 (a)), showing a bottleneck in capturing more comprehensive semantic representations of the ID classes. + +
MethodMCMGL-MCM
FPR95↓AUROC↑FPR95↓AUROC↑
FAce41.43±3.1191.01±0.6734.29±1.9490.99±0.56
FAFCE-K29.07±1.1193.77±0.1925.68±0.5893.82±0.11
+ +Table 5. Influence of the $\mathcal{L}_{FCE-K}$ . MCM and GL-MCM refer to the score functions used during model inference. + +
Forced-M-InitOriginal-M-InitFA MCMFA GL
FPR95↓AUROC↑FPR95↓AUROC↑
--40.71±4.0091.28±0.8234.88±2.9292.12±0.76
✓-40.81±2.2591.25±0.6436.03±2.8491.62±1.01
-✓31.35±1.3493.35±0.3426.39±1.0293.64±0.19
✓✓29.07±1.1193.77±0.1925.68±0.5893.82±0.11
+ +Table 6. Ablation study of different initialization scenarios for the forced and original prompts. "M-Init" represents manual initialization. $\checkmark$ represents using manual initialization; - represents using random initialization. + +
Shared- VectorM-InitFA MCMFA GL
FPR95↓AUROC↑FPR95↓AUROC↑
--44.02±2.2490.17±0.4246.29±1.3388.24±0.41
✓-40.71±4.0091.28±0.8234.88±2.9292.12±0.76
-✓34.69±0.6292.57±0.1229.27±0.5592.82±0.80
✓✓29.07±1.1193.77±0.1925.68±0.5893.82±0.11
+ +Table 7. Ablation study of the shared learnable vector. $\sqrt{}$ - for "Shared-vector" represents using the shared / independent learnable vector for the forced prompt; $\sqrt{}$ - for "M-Init" represents using manual / random initialization for the forced prompt. + +# 6. Conclusion + +In this study, we propose a novel CLIP-based framework for OOD detection based on Forced prompt leArning (FA), which focuses on fully exploiting the ID knowledge to effectively improve OOD detection without exploring complex OOD-related knowledge. We introduce a learnable forced prompt in addition to the frozen original prompt, both of which are using the same manual initialization. The forced prompt treats the original prompt as a reference, forcing itself to learn more diversified semantic descriptions of the ID classes rather than being limited to the textual semantics of class labels. We also introduce a forced coefficient to facilitate the forced prompt in learning more nuanced descriptions of the ID classes. Comprehensive experimental evaluations demonstrate that our method consistently surpasses current state-of-the-art methods on diverse benchmarks. We expect that our study can bring new insight on VLMs-based OOD detection methods and inspire more future research. + +# 7. Acknowledgments + +This work is supported in part by the National Natural Science Foundation of China (grant No. 62071502), the Major Key Project of PCL (grant No. PCL2023A09), and Guangdong Excellent Youth Team Program (grant No. 2023B1515040025). + +# References + +[1] Yichen Bai, Zongbo Han, Bing Cao, Xiaoheng Jiang, Qinghua Hu, and Changqing Zhang. Id-like prompt learning for few-shot out-of-distribution detection. In CVPR, pages 17480-17489, 2024. 2, 3, 5, 6 +[2] Julian Bitterwolf, Maximilian Müller, and Matthias Hein. In or out? fixing imagenet out-of-distribution detection evaluation. In ICML, pages 2471-2506, 2023. 5, 6, 7 +[3] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 - mining discriminative components with random forests. In ECCV, pages 446-461, 2014. 5, 7, 11 +[4] Guangyi Chen, Weiran Yao, Xiangchen Song, Xinyue Li, Yongming Rao, and Kun Zhang. PLOT: prompt learning with optimal transport for vision-language models. In ICLR, 2023. 2 +[5] Mengyuan Chen, Junyu Gao, and Changsheng Xu. Conjugated semantic pool improves OOD detection with pretrained vision-language models. In NeurIPS, 2024. 2 +[6] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In CVPR, pages 3606-3613, 2014. 5 +[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255, 2009. 5, 7 +[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 5 +[9] Xuefeng Du, Zhaoning Wang, Mu Cai, and Yixuan Li. VOS: learning what you don't know by virtual outlier synthesis. In ICLR, 2022. 2 +[10] Xuefeng Du, Yiyou Sun, Jerry Zhu, and Yixuan Li. Dream the impossible: Outlier imagination with diffusion models. NeurIPS, 36:60878-60901, 2023. 2 +[11] Sepideh Esmaeilpour, Bing Liu, Eric Robertson, and Lei Shu. Zero-shot out-of-distribution detection based on the pre-trained model clip. In AAAI, pages 6568-6576, 2022. 2 +[12] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In CVPRW, page 178, 2004. 5, 7, 11 +[13] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 5, 7 + +[14] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In ICLR, 2017. 1, 2, 5 +[15] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In CVPR, pages 15262-15271, 2021. 5, 7 +[16] Dan Hendrycks, Steven Basart, Mantas Mazeika, Andy Zou, Joseph Kwon, Mohammadreza Mostajabi, Jacob Steinhardt, and Dawn Song. Scaling out-of-distribution detection for real-world settings. In ICML, pages 8759-8773, 2022. 1, 2, 5 +[17] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alexander Shepard, Hartwig Adam, Pietro Perona, and Serge J. Belongie. The inaturalist species classification and detection dataset. In CVPR, pages 8769-8778, 2018. 5 +[18] Yen-Chang Hsu, Yilin Shen, Hongxia Jin, and Zsolt Kira. Generalized ODIN: detecting out-of-distribution image without learning from out-of-distribution data. In CVPR, pages 10948-10957, 2020. 2 +[19] Rui Huang and Yixuan Li. MOS: towards scaling out-of-distribution detection for large semantic space. In CVPR, pages 8710-8719, 2021. 5 +[20] Xue Jiang, Feng Liu, Zhen Fang, Hong Chen, Tongliang Liu, Feng Zheng, and Bo Han. Negative label guided OOD detection with pretrained vision-language models. In ICLR, 2024. 2 +[21] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCVW, pages 554–561, 2013. 5, 7, 11 +[22] Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. In NeurIPS, pages 7167-7177, 2018. 2 +[23] Tianqi Li, Guansong Pang, Xiao Bai, Wenjun Miao, and Jin Zheng. Learning transferable negative prompts for out-of-distribution detection. In CVPR, pages 17584-17594, 2024. 2, 3, 5 +[24] Shiyu Liang, Yixuan Li, and R. Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. In ICLR, 2018. 1, 2, 5 +[25] Jiashuo Liu, Zheyan Shen, Yue He, Xingxuan Zhang, Renzhe Xu, Han Yu, and Peng Cui. Towards out-of-distribution generalization: A survey. arXiv preprint arXiv:2108.13624, 2021. 1 +[26] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM computing surveys, 55(9): 1-35, 2023. 2 +[27] Weitang Liu, Xiaoyun Wang, John D. Owens, and Yixuan Li. Energy-based out-of-distribution detection. In NeurIPS, 2020. 1, 2, 5 +[28] Shuo Lu, Yingsheng Wang, Lijun Sheng, Aihua Zheng, Lingxiao He, and Jian Liang. Recent advances in ood detection: Problems and approaches. arXiv preprint arXiv:2409.11884, 2024. 1 + +[29] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 5, 7, 11 +[30] Yifei Ming, Ziyang Cai, Jiuming Gu, Yiyou Sun, Wei Li, and Yixuan Li. Delving into out-of-distribution detection with vision-language representations. In NeurIPS, 2022. 1, 3, 5, 7 +[31] Yifei Ming, Yiyou Sun, Ousmane Dia, and Yixuan Li. How to exploit hyperspherical embeddings for out-of-distribution detection? In ICLR, 2023. 2 +[32] Atsuyuki Miyai, Qing Yu, Go Irie, and Kiyoharu Aizawa. Zero-shot in-distribution detection in multi-object settings using vision-language foundation models. arXiv preprint arXiv:2304.04521, 2023. 1, 3, 5 +[33] Atsuyuki Miyai, Qing Yu, Go Irie, and Kiyoharu Aizawa. Locoop: Few-shot out-of-distribution detection via prompt learning. In NeurIPS, 2023. 2, 3, 5 +[34] Jun Nie, Yonggang Zhang, Zhen Fang, Tongliang Liu, Bo Han, and Xinmei Tian. Out-of-distribution detection with negative prompts. In ICLR, 2024. 2, 3, 5, 7 +[35] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729, 2008. 5, 7, 11 +[36] Omkar M. Parkhi, Andreae Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In CVPR, pages 3498-3505, 2012. 5, 7, 11 +[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 1, 2 +[38] Mohammadreza Salehi, Hossein Mirzaei, Dan Hendrycks, Yixuan Li, Mohammad Hossein Rohban, and Mohammad Sabokrou. A unified survey on anomaly, novelty, open-set, and out-of-distribution detection: Solutions and future challenges. arXiv preprint arXiv:2110.14051, 2021. 1 +[39] Vikash Sehwag, Mung Chiang, and Prateek Mittal. SSD: A unified framework for self-supervised outlier detection. In ICLR, 2021. 2 +[40] K Soomro. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012.5,7 +[41] Yiyou Sun, Chuan Guo, and Yixuan Li. React: Out-of-distribution detection with rectified activations. In NeurIPS, pages 144-157, 2021. 1, 2, 5 +[42] Yiyou Sun, Yifei Ming, Xiaojin Zhu, and Yixuan Li. Out-of-distribution detection with deep nearest neighbors. In ICML, pages 20827-20840, 2022. 2 +[43] Leitian Tao, Xuefeng Du, Jerry Zhu, and Yixuan Li. Nonparametric outlier synthesis. In ICLR, 2023. 2 +[44] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Open-set recognition: A good closed-set classifier is all you need. In ICLR, 2022. 2 + +[45] Haoqi Wang, Zhizhong Li, Litong Feng, and Wayne Zhang. Vim: Out-of-distribution with virtual-logit matching. In CVPR, pages 4911-4920, 2022. 5, 7 +[46] Hualiang Wang, Yi Li, Huifeng Yao, and Xiaomeng Li. CLIPN for zero-shot OOD detection: Teaching CLIP to say no. In ICCV, pages 1802-1812, 2023. 2, 5 +[47] Jianxiong Xiao, James Hays, Krista A. Ehinger, Aude Oliva, and Antonio Torralba. SUN database: Large-scale scene recognition from abbey to zoo. In CVPR, pages 3485-3492, 2010. 5 +[48] Geng Yu, Jianing Zhu, Jiangchao Yao, and Bo Han. Self-calibrated tuning of vision-language models for out-of-distribution detection. In NeurIPS, 2024. 2, 3, 5, 6 +[49] Tao Yu, Zhihe Lu, Xin Jin, Zhibo Chen, and Xinchao Wang. Task residual for tuning vision-language models. In CVPR, pages 10899-10909, 2023. 5 +[50] Yeonguk Yu, Sungho Shin, Seongju Lee, Changyun Jun, and Kyoobin Lee. Block selection method for using feature norm in out-of-distribution detection. In CVPR, pages 15701-15711, 2023. 2 +[51] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kun-chang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of CLIP for few-shot classification. In ECCV, pages 493-510, 2022. 5 +[52] Bolei Zhou, Ágata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE TPAMI, 40(6):1452-1464, 2018. 5 +[53] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In CVPR, pages 16795-16804, 2022. 2 +[54] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 130(9):2337-2348, 2022. 2, 3, 4, 5, 7, 8 +[55] Yibo Zhou. Rethinking reconstruction autoencoder-based out-of-distribution detection. In CVPR, pages 7379-7387, 2022. 2 \ No newline at end of file diff --git a/2025/FA_ Forced Prompt Learning of Vision-Language Models for Out-of-Distribution Detection/images.zip b/2025/FA_ Forced Prompt Learning of Vision-Language Models for Out-of-Distribution Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..608ea3e5cc463c7ff84bbbc13ff70a35817877db --- /dev/null +++ b/2025/FA_ Forced Prompt Learning of Vision-Language Models for Out-of-Distribution Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da8df17e489d4ce68a281c2181b907e1f7378687f0c50f80edf7b91e094f83a6 +size 822440 diff --git a/2025/FA_ Forced Prompt Learning of Vision-Language Models for Out-of-Distribution Detection/layout.json b/2025/FA_ Forced Prompt Learning of Vision-Language Models for Out-of-Distribution Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e33c89e219c9a69e9dfeb84dacf641a1ef5e5b9e --- /dev/null +++ b/2025/FA_ Forced Prompt Learning of Vision-Language Models for Out-of-Distribution Detection/layout.json @@ -0,0 +1,8556 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 118, + 103, + 493, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 103, + 493, + 138 + ], + "spans": [ + { + "bbox": [ + 118, + 103, + 493, + 138 + ], + "type": "text", + "content": "FA: Forced Prompt Learning of Vision-Language Models for Out-of-Distribution Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 159, + 555, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 159, + 555, + 219 + ], + "spans": [ + { + "bbox": [ + 61, + 159, + 555, + 219 + ], + "type": "text", + "content": "Xinhua Lu" + }, + { + "bbox": [ + 61, + 159, + 555, + 219 + ], + "type": "inline_equation", + "content": "^{1,2,4}" + }, + { + "bbox": [ + 61, + 159, + 555, + 219 + ], + "type": "text", + "content": ", Runhe Lai" + }, + { + "bbox": [ + 61, + 159, + 555, + 219 + ], + "type": "inline_equation", + "content": "^{1,2,4}" + }, + { + "bbox": [ + 61, + 159, + 555, + 219 + ], + "type": "text", + "content": ", Yanqi Wu" + }, + { + "bbox": [ + 61, + 159, + 555, + 219 + ], + "type": "inline_equation", + "content": "^{1,2,4}" + }, + { + "bbox": [ + 61, + 159, + 555, + 219 + ], + "type": "text", + "content": ", Kanghao Chen" + }, + { + "bbox": [ + 61, + 159, + 555, + 219 + ], + "type": "inline_equation", + "content": "^{3\\dagger}" + }, + { + "bbox": [ + 61, + 159, + 555, + 219 + ], + "type": "text", + "content": ", Wei-Shi Zheng" + }, + { + "bbox": [ + 61, + 159, + 555, + 219 + ], + "type": "inline_equation", + "content": "^{1,2,4}" + }, + { + "bbox": [ + 61, + 159, + 555, + 219 + ], + "type": "text", + "content": ", Ruixuan Wang" + }, + { + "bbox": [ + 61, + 159, + 555, + 219 + ], + "type": "inline_equation", + "content": "^{1,2,4\\dagger}" + }, + { + "bbox": [ + 61, + 159, + 555, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 159, + 555, + 219 + ], + "type": "text", + "content": "Sun Yat-sen University " + }, + { + "bbox": [ + 61, + 159, + 555, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 159, + 555, + 219 + ], + "type": "text", + "content": "Peng Cheng Laboratory \n" + }, + { + "bbox": [ + 61, + 159, + 555, + 219 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 159, + 555, + 219 + ], + "type": "text", + "content": "Hong Kong University of Science and Technology (Guangzhou) \n" + }, + { + "bbox": [ + 61, + 159, + 555, + 219 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 61, + 159, + 555, + 219 + ], + "type": "text", + "content": "Key Laboratory of Machine Intelligence and Advanced Computing" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 97, + 220, + 517, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 220, + 517, + 245 + ], + "spans": [ + { + "bbox": [ + 97, + 220, + 517, + 245 + ], + "type": "text", + "content": "{luxh55, lairh5, wuyq268}@mail2.sysu.edu.cn, kchen879@connect.hkust-gz.edu.cn, wszheng@ieee.org, wangruix5@mail.sysu.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "spans": [ + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 299, + 296, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 299, + 296, + 610 + ], + "spans": [ + { + "bbox": [ + 53, + 299, + 296, + 610 + ], + "type": "text", + "content": "Pre-trained vision-language models (VLMs) have advanced out-of-distribution (OOD) detection recently. However, existing CLIP-based methods often focus on learning OOD-related knowledge to improve OOD detection, showing limited generalization or reliance on external largescale auxiliary datasets. In this study, instead of delving into the intricate OOD-related knowledge, we propose an innovative CLIP-based framework based on Forced prompt leArning (FA), designed to make full use of the InDistribution (ID) knowledge and ultimately boost the effectiveness of OOD detection. Our key insight is to learn a prompt (i.e. forced prompt) that contains more diversified and richer descriptions of the ID classes beyond the textual semantics of class labels. Specifically, it promotes better discernment for ID images, by forcing more notable semantic similarity between ID images and the learnable forced prompt. Moreover, we introduce a forced coefficient, encouraging the forced prompt to learn more comprehensive and nuanced descriptions of the ID classes. In this way, FA is capable of achieving notable improvements in OOD detection, even when trained without any external auxiliary datasets, while maintaining an identical number of trainable parameters as CoOp. Extensive empirical evaluations confirm our method consistently outperforms current state-of-the-art methods. Code is available at https://github.com/0xFIFA/FA." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 635, + 135, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 135, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 135, + 647 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 656, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 656, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 656, + 296, + 693 + ], + "type": "text", + "content": "AI models often encounter Out-of-Distribution (OOD) samples [14, 25, 28, 38], which differ from the distribution of the training data, when deployed in real-world applications." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 318, + 273, + 550, + 437 + ], + "blocks": [ + { + "bbox": [ + 318, + 273, + 550, + 437 + ], + "lines": [ + { + "bbox": [ + 318, + 273, + 550, + 437 + ], + "spans": [ + { + "bbox": [ + 318, + 273, + 550, + 437 + ], + "type": "image", + "image_path": "d9ad327290586f4d05943d5523b636a0b2470f791793f567897944331faba407.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 445, + 555, + 534 + ], + "lines": [ + { + "bbox": [ + 313, + 445, + 555, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 555, + 534 + ], + "type": "text", + "content": "Figure 1. Comparison of CLIP-based OOD detection methods. Existing methods focus on OOD-related knowledge to learn a complex ID/OOD decision boundary as shown in the orange area in (b). In contrast, our FA aims to fully exploit ID knowledge by forcing the model to learn richer descriptions of the ID classes beyond the textual semantics of class labels. These richer descriptions of the ID classes compared to the reference text allow better discernment of ID/OOD samples (orange vs. blue area in (a))." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 544, + 554, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 554, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 554, + 593 + ], + "type": "text", + "content": "Detecting OOD data is vital for the reliability of AI systems. This problem becomes harder in fields like intelligent diagnosis, where only a small amount of labeled data is available, motivating the problem of few-shot OOD detection." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "content": "To tackle the OOD detection problem, previous single-modal post-hoc methods [14, 16, 24, 27, 41] achieve notable success based on a model pre-trained on the ID dataset. However, they come with limitations, such as requiring substantial computational resources and annotation costs for training, which hinder the performance of few-shot OOD detection. With the development of large pre-trained vision-language models (VLMs) like CLIP [37], few-shot OOD detection has achieved remarkable performance. Most of the CLIP-based OOD detection methods [30, 32] focus on" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 149, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 149, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 149, + 713 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1152" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "leveraging the powerful generalization capability of CLIP to improve OOD detection." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 99, + 296, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 99, + 296, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 99, + 296, + 397 + ], + "type": "text", + "content": "Recent CLIP-based OOD detection methods [1, 11, 23, 33, 34, 46, 48] aim to learn OOD-related knowledge to improve OOD detection performance. For example, recent CLIP-based works [11, 46] improve OOD detection by relying on training on external large-scale auxiliary datasets, which entails significant resource overhead. To eliminate dependence on external datasets, existing methods [1, 33, 48] learn OOD-related knowledge from exposed outliers based only on ID training data, such as background regions in ID training images. However, these OOD features extracted from specific regions are difficult to match with infinite OOD data encountered in practice. Moreover, some methods [23, 34] aim to learn negative prompts semantically opposite to the ID class labels, while these limited negative prompts are often insufficient to capture the distinctions between the diverse and numerous OOD data and the ID data. Overall, improving OOD detection by learning OOD-related knowledge shows inherent limitations or requires significant computational resources as well as labor costs. Inspired by the method [44] for open-set recognition, which demonstrates that enhancing closed-set accuracy can typically improve open-set recognition capabilities, we are motivated to explore ways to improve the identification of ID images, thereby enhancing OOD detection performance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 400, + 296, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 400, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 296, + 568 + ], + "type": "text", + "content": "In this paper, rather than focusing on the intricate OOD-related knowledge, we propose a novel CLIP-based framework based on Forced prompt leArning (FA), which aims to fully exploit the ID knowledge and ultimately improve OOD detection. Our key insight is to learn a prompt that contains richer knowledge beyond the textual semantics of class labels. It facilitates advanced discernment for ID images, by forcing higher semantic similarity between ID images and the learnable prompt (i.e. forced prompt), which provides more diversified and richer descriptions of the ID classes. In this way, FA can achieve significant improvements in OOD detection performance, even when trained without any additional external data, while maintaining the same number of learnable parameters as CoOp [54]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": "Specifically, we introduce a novel forced prompt along with the original prompt, both of which are initialized identically. FA preserves the generalization capability of the VLMs (i.e. CLIP) by freezing the original prompt and optimizing a trainable copy (i.e. forced prompt). In particular, the forced prompt treats the original prompt as a reference, while forcing the text features associated with the forced prompt to be more salient compared to those associated with the original prompt. In this way, it effectively improves the OOD performance even without sacrificing the ID classification capability of the model. Additionally, we introduce a forced coefficient, encouraging the forced" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": "prompt to learn more comprehensive connotations. Experimentally, FA achieves superior few-shot OOD detection performance across diverse OOD benchmarks. Compared to current SOTA methods [48] that focus on learning OOD-related knowledge, our method significantly reduces the average FPR95 score from " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "31.62\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "27.81\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": " and improves the average AUROC from " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "92.01\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "93.26\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": " even in 1-shot OOD detection on ImageNet-1k. The main contributions are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 183, + 553, + 294 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 313, + 183, + 553, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 183, + 553, + 219 + ], + "spans": [ + { + "bbox": [ + 313, + 183, + 553, + 219 + ], + "type": "text", + "content": "- We propose a simple yet effective framework, which fully exploits the ID knowledge to improve OOD detection without focusing on intricate OOD-related knowledge." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 221, + 553, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 221, + 553, + 256 + ], + "spans": [ + { + "bbox": [ + 313, + 221, + 553, + 256 + ], + "type": "text", + "content": "- We present a Forced prompt leArning (FA) strategy to exploit richer knowledge beyond the textual semantics of class labels." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 259, + 553, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 259, + 553, + 294 + ], + "spans": [ + { + "bbox": [ + 313, + 259, + 553, + 294 + ], + "type": "text", + "content": "- We evaluate our method on diverse OOD benchmarks, showing that our model consistently outperforms current state-of-the-art methods." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 308, + 399, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 399, + 320 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 399, + 320 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 329, + 553, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 553, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 553, + 508 + ], + "type": "text", + "content": "Prompt Learning. In recent years, pre-trained vision-language models (VLMs) such as CLIP [37] have demonstrated powerful few-shot learning capabilities in both visual and textual domains. However, the design of the prompt greatly affects the performance of VLMs for downstream tasks, which may require manually crafting numerous prompts. Inspired by prompt learning studies in natural language processing (NLP) [26], CoOp [54] uses a set of learnable context vectors to transfer CLIP for specific downstream tasks, becoming a pioneering method for subsequent studies [4, 33, 53]. Currently, prompt learning is widely used in OOD detection mainly to represent various OOD-related knowledge. In this work, we learn the forced prompt, which contains richer descriptions of the ID classes to improve OOD detection." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 510, + 554, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 554, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 554, + 616 + ], + "type": "text", + "content": "Out-of-Distribution Detection. Conventional methods explore OOD detection for single-modal models. One line of studies designs score functions to differentiate ID and OOD data based on output from the pre-trained models, such as logit output [14, 16, 24, 27] or outputs from the penultimate layer [22, 41, 42, 50]. These methods are called post-hoc methods. Another line of studies [9, 10, 18, 31, 39, 43, 55] adopts various training strategies to learn a reliable decision boundary between ID and OOD data." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": "Recently, the development of OOD detectors based on VLMs, especially CLIP, has received much attention as VLMs have demonstrated their remarkable generalization capability in both visual and textual domains. Some studies leverage real outlier information from external auxiliary OOD datasets [11, 46] or extensive corpora [5, 20] to promote OOD detection. However, this is impractical in real-world scenarios, where outliers are infinite and agnostic." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1153" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 74, + 499, + 265 + ], + "blocks": [ + { + "bbox": [ + 107, + 74, + 499, + 265 + ], + "lines": [ + { + "bbox": [ + 107, + 74, + 499, + 265 + ], + "spans": [ + { + "bbox": [ + 107, + 74, + 499, + 265 + ], + "type": "image", + "image_path": "d6c07d7693e54a2e5e7e9bf056c4470e042bca80776d2e49bb596b306ede69d4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 270, + 555, + 315 + ], + "lines": [ + { + "bbox": [ + 54, + 270, + 555, + 315 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 555, + 315 + ], + "type": "text", + "content": "Figure 2. Overview of the proposed FA framework. Our framework includes the learnable forced prompt and the frozen original prompt, both of which are initialized identically by a manual template \"a photo of a [class-c]\". The learnable forced prompt treats the frozen original prompt as a reference, forcing its text features to become more salient compared to those generated by the original prompt. The richer and more diversified description of ID classes learned by FA can ultimately improve OOD detection." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 324, + 296, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 324, + 296, + 683 + ], + "spans": [ + { + "bbox": [ + 56, + 324, + 296, + 683 + ], + "type": "text", + "content": "To eliminate the dependence on external data, recent methods [1, 23, 32, 33, 48] were developed under the assumption that only ID data are available. As a representative zero-shot method that only uses available ID labels, MCM [30] employs softmax scaling to align visual features with ID text concepts for OOD detection. GL-MCM [32] further introduces a local maximum concept matching(L-MCM) score to improve the separability of the local information. Compared with zero-shot methods, which may undergo a domain gap with ID downstream data, prompt learning methods achieve better OOD detection performance with access to few-shot ID samples. For prompt learning OOD detection methods, LoCoOp [33] and SCT [48] keep the textual embeddings of ID classes away from ID-irrelevant local region embeddings in the multi-modal embedding space by their proposed entropy maximization strategy, which ensures that OOD embeddings can be dissimilar to any textual embedding of ID classes. To detect challenging OOD samples to improve OOD detection, ID-like [1] explores the vicinity of ID samples to construct OOD samples correlated to the ID and refines the additional \"ID-like\" text embeddings to fine-grained differences. In addition, NegPrompt [23] and LSN [34] introduce the negative connotations of ID categories which can be represented by additional negative prompts, enabling more accurate detection of OOD samples. Overall, various OOD-related knowledge is commonly adopted to address the model's overconfidence in current research. In this work, we focus on leveraging ID knowledge to improve OOD detection, instead of exploring complex OOD-related knowledge." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 323, + 399, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 399, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 399, + 335 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 342, + 555, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 521 + ], + "type": "text", + "content": "Prompt Learning with CLIP. CLIP contains an image encoder " + }, + { + "bbox": [ + 313, + 342, + 555, + 521 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 313, + 342, + 555, + 521 + ], + "type": "text", + "content": " and a text encoder " + }, + { + "bbox": [ + 313, + 342, + 555, + 521 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 313, + 342, + 555, + 521 + ], + "type": "text", + "content": ", designed to extract features from images and text descriptions respectively. Generally, for an ID dataset which contains " + }, + { + "bbox": [ + 313, + 342, + 555, + 521 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 342, + 555, + 521 + ], + "type": "text", + "content": " categories, the hand-crafted prompt " + }, + { + "bbox": [ + 313, + 342, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{u}}_c" + }, + { + "bbox": [ + 313, + 342, + 555, + 521 + ], + "type": "text", + "content": " \"a photo of a [class-c]\" is designed to match images, where class-c represents the class name. Formally, these prompts are individually encoded in " + }, + { + "bbox": [ + 313, + 342, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{t}}_c = g(\\hat{\\mathbf{u}}_c) \\in \\mathbb{R}^{d \\times 1}" + }, + { + "bbox": [ + 313, + 342, + 555, + 521 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 342, + 555, + 521 + ], + "type": "inline_equation", + "content": "c = 1, \\dots, C" + }, + { + "bbox": [ + 313, + 342, + 555, + 521 + ], + "type": "text", + "content": ". Given an image " + }, + { + "bbox": [ + 313, + 342, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 342, + 555, + 521 + ], + "type": "text", + "content": ", it can be encoded in " + }, + { + "bbox": [ + 313, + 342, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{z} = f(\\mathbf{x}) \\in \\mathbb{R}^{d \\times 1}" + }, + { + "bbox": [ + 313, + 342, + 555, + 521 + ], + "type": "text", + "content": ". Because image features and corresponding text features are aligned on the multimodal embedding space, the cosine similarity " + }, + { + "bbox": [ + 313, + 342, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\cos(\\mathbf{z}, \\hat{\\mathbf{t}}_c)" + }, + { + "bbox": [ + 313, + 342, + 555, + 521 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 342, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 313, + 342, + 555, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 342, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{t}}_c" + }, + { + "bbox": [ + 313, + 342, + 555, + 521 + ], + "type": "text", + "content": " can represent the matching degree between the image " + }, + { + "bbox": [ + 313, + 342, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 342, + 555, + 521 + ], + "type": "text", + "content": " and text description " + }, + { + "bbox": [ + 313, + 342, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{u}}_c" + }, + { + "bbox": [ + 313, + 342, + 555, + 521 + ], + "type": "text", + "content": ". Therefore, if the text feature " + }, + { + "bbox": [ + 313, + 342, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_c" + }, + { + "bbox": [ + 313, + 342, + 555, + 521 + ], + "type": "text", + "content": " of class " + }, + { + "bbox": [ + 313, + 342, + 555, + 521 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 342, + 555, + 521 + ], + "type": "text", + "content": " has the highest similarity to " + }, + { + "bbox": [ + 313, + 342, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 313, + 342, + 555, + 521 + ], + "type": "text", + "content": ", the image " + }, + { + "bbox": [ + 313, + 342, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 342, + 555, + 521 + ], + "type": "text", + "content": " will be considered to be class " + }, + { + "bbox": [ + 313, + 342, + 555, + 521 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 342, + 555, + 521 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "type": "text", + "content": "To effectively transfer CLIP to downstream image recognition tasks, CoOp [54] sets a portion of tokens in the text prompt as continuous learnable parameters. Concretely, CoOp initializes the text prompts as " + }, + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_c = [\\mathbf{v}_1,\\dots ,\\mathbf{v}_L,\\mathbf{w}_c],c\\in \\{1,\\dots ,C\\}" + }, + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "type": "text", + "content": " is the length of the prompt's tokens, " + }, + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i" + }, + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "type": "inline_equation", + "content": "i\\in \\{1,\\ldots ,L\\}" + }, + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "type": "text", + "content": ") is the learnable vector with the same dimension as the word embedding and " + }, + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_c" + }, + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "type": "text", + "content": " is the word embedding for the class name of class " + }, + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "type": "text", + "content": ". The text prompt " + }, + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_c" + }, + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "type": "text", + "content": " is encoded into features " + }, + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_c = g(\\mathbf{u}_c)" + }, + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "type": "text", + "content": ". Together with the image feature " + }, + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "type": "text", + "content": ", the probability for the image " + }, + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "type": "text", + "content": " to be classified into class " + }, + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 362, + 662, + 555, + 693 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 662, + 555, + 693 + ], + "spans": [ + { + "bbox": [ + 362, + 662, + 555, + 693 + ], + "type": "interline_equation", + "content": "p (y = c \\mid \\mathbf {x}) = \\frac {e ^ {\\cos (\\mathbf {z} , \\mathbf {t} _ {c}) / \\tau}}{\\sum_ {j = 1} ^ {C} e ^ {\\cos (\\mathbf {z} , \\mathbf {t} _ {j}) / \\tau}}, \\tag {1}", + "image_path": "435ddaae65bf56b276bec652d2081707b3fbb7f88b7020a4d0f5d95b1123c630.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 702, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 702, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 314, + 702, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 314, + 702, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 314, + 702, + 555, + 714 + ], + "type": "text", + "content": " is a fixed temperature scaling hyper-parameter." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1154" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "Overall, the prompt's tokens can be trained to align with training data by minimizing the cross-entropy loss with the prediction probability from Eq. (1)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 108, + 294, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 108, + 294, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 108, + 294, + 168 + ], + "type": "text", + "content": "OOD Detection. In the OOD detection task, the model is expected to determine whether a test image belongs to one of the learned ID classes. Therefore, OOD detection can be seen as a binary classification problem to distinguish ID images from OOD images in a test set " + }, + { + "bbox": [ + 55, + 108, + 294, + 168 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{test}" + }, + { + "bbox": [ + 55, + 108, + 294, + 168 + ], + "type": "text", + "content": " as follows" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 122, + 177, + 294, + 205 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 177, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 122, + 177, + 294, + 205 + ], + "type": "interline_equation", + "content": "D (\\mathbf {x}) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} S (\\mathbf {x}) \\geq \\mu \\\\ 0, & \\text {i f} S (\\mathbf {x}) < \\mu \\end{array} , \\right. \\tag {2}", + "image_path": "f3bd5bb3cf11246a8581c76a717b459b7bbeda6c6e6b4b9e53632f45b27f6c2d.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 213, + 295, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 213, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 295, + 262 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 213, + 295, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathcal{D}_{test}" + }, + { + "bbox": [ + 55, + 213, + 295, + 262 + ], + "type": "text", + "content": ", 1 and 0 respectively indicate that " + }, + { + "bbox": [ + 55, + 213, + 295, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 213, + 295, + 262 + ], + "type": "text", + "content": " is classified as the ID class and the OOD class by the OOD detector " + }, + { + "bbox": [ + 55, + 213, + 295, + 262 + ], + "type": "inline_equation", + "content": "D(\\cdot)" + }, + { + "bbox": [ + 55, + 213, + 295, + 262 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 213, + 295, + 262 + ], + "type": "inline_equation", + "content": "S(\\cdot)" + }, + { + "bbox": [ + 55, + 213, + 295, + 262 + ], + "type": "text", + "content": " is certain score function, and " + }, + { + "bbox": [ + 55, + 213, + 295, + 262 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 213, + 295, + 262 + ], + "type": "text", + "content": " is a predefined threshold constant." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 273, + 138, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 273, + 138, + 286 + ], + "spans": [ + { + "bbox": [ + 55, + 273, + 138, + 286 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 293, + 123, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 123, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 123, + 304 + ], + "type": "text", + "content": "4.1. Overview" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 310, + 295, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 310, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 310, + 295, + 502 + ], + "type": "text", + "content": "In this study, we propose a novel CLIP-based framework based on Forced prompt leArning (FA), designed to fully exploit the ID knowledge from few-shot ID samples by forcing the prompt to learn richer knowledge beyond the textual semantics of class labels. As shown in Fig. 2, we introduce a forced prompt along with the original prompt, both of which are initialized with the same semantics. By freezing the original prompt and optimizing a trainable copy, ID images will show higher semantic similarity to the forced prompt compared to the original prompt. Moreover, to encourage descriptions of ID classes learned by the forced prompt to become more detailed and comprehensive, we also introduce a forced coefficient. During testing, the model is capable of effectively distinguishing between ID data and OOD data, benefiting from the diversified descriptions of the ID classes provided by the forced prompt." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 510, + 190, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 190, + 522 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 190, + 522 + ], + "type": "text", + "content": "4.2. Forced prompt learning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 527, + 295, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 527, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 527, + 295, + 635 + ], + "type": "text", + "content": "Merely the semantic information conveyed by class names is insufficient to comprehensively encompass the discriminative information of each class. In order to explore richer semantic descriptions of the ID classes beyond the textual semantics of class labels, a novel forced prompt along with the original prompt is proposed. To implement the above objective, we introduce a simple yet effective training strategy using the forced cross-entropy (FCE) loss based on the forced prompt as follows" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 59, + 643, + 294, + 687 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 643, + 294, + 687 + ], + "spans": [ + { + "bbox": [ + 59, + 643, + 294, + 687 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {F C E} = \\mathbb {E} _ {\\left(\\mathbf {x}, y _ {c}\\right) \\sim \\mathcal {D} _ {t r a i n} ^ {I D}} \\left[ - \\log \\frac {e ^ {s _ {c} ^ {f} / \\tau}}{\\sum_ {j = 1} ^ {C} e ^ {s _ {j} ^ {f} / \\tau} + \\sum_ {j = 1} ^ {C} e ^ {s _ {j} ^ {o} / \\tau}} \\right], \\tag {3}", + "image_path": "4c4c763eb6b63a023428c948a3b598dc0125d4cb675114d27028ce68ef4d772d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 687, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 687, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 687, + 295, + 715 + ], + "type": "text", + "content": "where the ID training dataset " + }, + { + "bbox": [ + 55, + 687, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{train}^{ID}" + }, + { + "bbox": [ + 55, + 687, + 295, + 715 + ], + "type": "text", + "content": " consists of ID image-label pairs " + }, + { + "bbox": [ + 55, + 687, + 295, + 715 + ], + "type": "inline_equation", + "content": "(\\mathbf{x},y_{c})" + }, + { + "bbox": [ + 55, + 687, + 295, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 687, + 295, + 715 + ], + "type": "inline_equation", + "content": "s_j^f = \\cos (\\mathbf{z},\\mathbf{t}_j^f)" + }, + { + "bbox": [ + 55, + 687, + 295, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 687, + 295, + 715 + ], + "type": "inline_equation", + "content": "s_j^o = \\cos (\\mathbf{z},\\mathbf{t}_j^o)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "represent the similarity between the image feature and the prompt feature corresponding to the forced prompt and the original prompt, respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 111, + 555, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 111, + 555, + 315 + ], + "spans": [ + { + "bbox": [ + 313, + 111, + 555, + 315 + ], + "type": "text", + "content": "Based on this loss function, the ID image features will be forced to show higher cosine similarity to the learnable forced prompt compared to the original text prompt. This is because the cosine similarity between the ID image features and the original text prompt is already high before any further prompt learning. Consequently, the forced prompt is forced to uncover richer discriminative information of ID classes, so that the text features of the forced prompt achieve a higher cosine similarity with the image features compared to those of the original prompt. Benefiting from this information specific to the ID classes, the cosine similarity between the image features of an OOD image and the text features of both prompts may exhibit less distinction compared to those of the ID data. Therefore, even when trained without any reliance on external auxiliary datasets, the detector will have a notable capability to distinguish ID data from OOD data." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 318, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 555, + 714 + ], + "type": "text", + "content": "However, this naive FCE loss may not work effectively because of its inherent limitations. Concretely, employing random initialization for both prompts fails to fully leverage the semantic information provided by CLIP's prior knowledge, resulting in limited performance. To solve this limitation, we propose to initialize the forced prompt using the manual template (i.e. \"a photo of a [class-c]\") identified to the original prompt. In this way, both prompts with manual initialization will have clear semantic information compared to random initialization, thereby improving the model's generalization capability. Specifically, the class labels are utilized in both prompts by concatenating with the prompt embedding to provide foundational semantic information related to the category names, based on CLIP's prior knowledge. In particular, for the forced prompt, we adopt the shared learnable vector across all classes rather than the independent learnable vector for each class. This choice is inspired by CoOp [54], which demonstrates that using an independent learnable vector mostly underperforms the shared learnable vector in challenging low-data scenarios since the former has more parameters and requires more data for training. Formally, we utilize the embeddings of the manual template to initialize the prompt's embeddings of both prompts, formulated as " + }, + { + "bbox": [ + 313, + 318, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_c = [\\mathbf{v}_1,\\dots ,\\mathbf{v}_L,\\mathbf{w}_c], c\\in" + }, + { + "bbox": [ + 313, + 318, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\{1,\\dots ,C\\}" + }, + { + "bbox": [ + 313, + 318, + 555, + 714 + ], + "type": "text", + "content": ", which will be fed to the text encoder to obtain prompt feature " + }, + { + "bbox": [ + 313, + 318, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_c" + }, + { + "bbox": [ + 313, + 318, + 555, + 714 + ], + "type": "text", + "content": " (see Section 3). Here, " + }, + { + "bbox": [ + 313, + 318, + 555, + 714 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 318, + 555, + 714 + ], + "type": "text", + "content": " is the length of the token (e.g., for the \"a photo of a\", " + }, + { + "bbox": [ + 313, + 318, + 555, + 714 + ], + "type": "inline_equation", + "content": "L = 4" + }, + { + "bbox": [ + 313, + 318, + 555, + 714 + ], + "type": "text", + "content": "). Then we freeze the original prompt and the class label part of both the prompts to preserve the generalization capability of CLIP, while only making the forced prompt learnable. Note that our method effectively improves OOD detection performance while maintaining consistency with CoOp in the number of learnable parameters without leveraging ad" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1155" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 84 + ], + "type": "text", + "content": "ditional learnable prompts, unlike existing work [1, 23, 34]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 85, + 295, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 85, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 85, + 295, + 300 + ], + "type": "text", + "content": "Although the above prompt design achieves moderate performance, relying solely on a single original prompt as a reference is somewhat insufficient to develop a comprehensive capability, particularly with the limitations of ID data (e.g., few-shot OOD). To encourage the forced prompt to capture more comprehensive and nuanced descriptions of the ID classes, we introduce a forced coefficient " + }, + { + "bbox": [ + 55, + 85, + 295, + 300 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 85, + 295, + 300 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 85, + 295, + 300 + ], + "type": "inline_equation", + "content": "K \\geq 0, K \\in \\mathbb{N}" + }, + { + "bbox": [ + 55, + 85, + 295, + 300 + ], + "type": "text", + "content": "), which indicates the intensity with which the model is compelled to learn from the data. Formally, we first derive the original text features " + }, + { + "bbox": [ + 55, + 85, + 295, + 300 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{t}_1^o, \\dots, \\mathbf{t}_C^o\\}" + }, + { + "bbox": [ + 55, + 85, + 295, + 300 + ], + "type": "text", + "content": " from the original prompt and the image features " + }, + { + "bbox": [ + 55, + 85, + 295, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 55, + 85, + 295, + 300 + ], + "type": "text", + "content": " from the corresponding image. Then, we compute the cosine similarity between the original text features and the image features for " + }, + { + "bbox": [ + 55, + 85, + 295, + 300 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 85, + 295, + 300 + ], + "type": "text", + "content": " iterations, which will be used in the subsequent computation of the cross-entropy loss function. Notably, when " + }, + { + "bbox": [ + 55, + 85, + 295, + 300 + ], + "type": "inline_equation", + "content": "K = 0" + }, + { + "bbox": [ + 55, + 85, + 295, + 300 + ], + "type": "text", + "content": ", no original prompt is used, and the model operates equivalently to CoOp. Based on the forced coefficient, we refine Eq. (3) as follows" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 310, + 296, + 353 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 310, + 296, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 310, + 296, + 353 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {F C E - K} = \\mathbb {E} _ {\\left(\\mathbf {x}, y _ {c}\\right) \\sim \\mathcal {D} _ {t r a i n} ^ {I D}} \\left[ - \\log \\frac {e ^ {s _ {c} ^ {f} / \\tau}}{\\sum_ {j = 1} ^ {C} e ^ {s _ {j} ^ {f} / \\tau} + K \\sum_ {j = 1} ^ {C} e ^ {s _ {j} ^ {o} / \\tau}} \\right]. \\tag {4}", + "image_path": "4e27628c12a7725afbb5819949df1aed15ddec91543a6cce321ab912c72de60a.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 354, + 295, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 295, + 389 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 295, + 389 + ], + "type": "text", + "content": "Based on this function, the cosine similarity between the ID image features and the text features of the forced prompt must become more salient as the coefficient " + }, + { + "bbox": [ + 55, + 354, + 295, + 389 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 354, + 295, + 389 + ], + "type": "text", + "content": " increases." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 398, + 154, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 398, + 154, + 410 + ], + "spans": [ + { + "bbox": [ + 55, + 398, + 154, + 410 + ], + "type": "text", + "content": "4.3. Model inference" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 416, + 296, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 416, + 296, + 524 + ], + "spans": [ + { + "bbox": [ + 55, + 416, + 296, + 524 + ], + "type": "text", + "content": "During model inference, for the downstream classification task, we adopt the same strategy as CLIP, depending solely on the forced prompt [1]. For the OOD detection task, our method can be flexibly combined with different score functions, such as the MCM [30] score and the GL-MCM [32] score. The MCM score is defined as the maximum similarity between the global image features " + }, + { + "bbox": [ + 55, + 416, + 296, + 524 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^g" + }, + { + "bbox": [ + 55, + 416, + 296, + 524 + ], + "type": "text", + "content": " (i.e. " + }, + { + "bbox": [ + 55, + 416, + 296, + 524 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 55, + 416, + 296, + 524 + ], + "type": "text", + "content": ") and all text features " + }, + { + "bbox": [ + 55, + 416, + 296, + 524 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_c^a" + }, + { + "bbox": [ + 55, + 416, + 296, + 524 + ], + "type": "text", + "content": " (i.e. the concatenation of text features " + }, + { + "bbox": [ + 55, + 416, + 296, + 524 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_c^f" + }, + { + "bbox": [ + 55, + 416, + 296, + 524 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 416, + 296, + 524 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_c^o" + }, + { + "bbox": [ + 55, + 416, + 296, + 524 + ], + "type": "text", + "content": ") after applying softmax with temperature " + }, + { + "bbox": [ + 55, + 416, + 296, + 524 + ], + "type": "inline_equation", + "content": "\\tau_0" + }, + { + "bbox": [ + 55, + 416, + 296, + 524 + ], + "type": "text", + "content": ", i.e.," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 533, + 295, + 575 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 533, + 295, + 575 + ], + "spans": [ + { + "bbox": [ + 58, + 533, + 295, + 575 + ], + "type": "interline_equation", + "content": "S _ {\\mathrm {M C M}} (\\mathbf {x}) = \\max _ {c} \\frac {e ^ {\\cos \\left(\\mathbf {z} ^ {g} , \\mathbf {t} _ {c} ^ {a}\\right) / \\tau_ {0}}}{\\sum_ {j = 1} ^ {C} e ^ {\\cos \\left(\\mathbf {z} ^ {g} , \\mathbf {t} _ {j} ^ {f}\\right) / \\tau_ {0}} + K e ^ {\\cos \\left(\\mathbf {z} ^ {g} , \\mathbf {t} _ {j} ^ {o}\\right) / \\tau_ {0}}}, \\tag {5}", + "image_path": "6002875049349633c1b1f5cce0d27da175badfb629f5cabe1a9239330e8d6100.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 576, + 295, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 295, + 600 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 295, + 600 + ], + "type": "text", + "content": "while the GL-MCM score simultaneously considers both global and local features, which can be expressed as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 610, + 295, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 610, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 104, + 610, + 295, + 624 + ], + "type": "interline_equation", + "content": "S _ {\\mathrm {G L - M C M}} (\\mathbf {x}) = S _ {\\mathrm {M C M}} (\\mathbf {x}) + S _ {\\mathrm {L - M C M}} (\\mathbf {x}), \\tag {6}", + "image_path": "ee88d6c81f64e1d580eca4b4102e83a64a7fc6aed855b3c407d1b1f61c3eacd3.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 59, + 634, + 294, + 677 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 634, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 59, + 634, + 294, + 677 + ], + "type": "interline_equation", + "content": "S _ {\\mathrm {L - M C M}} (\\mathbf {x}) = \\max _ {i, c} \\frac {e ^ {\\cos \\left(\\mathbf {z} _ {i} ^ {l} , \\mathbf {t} _ {c} ^ {a}\\right) / \\tau_ {0}}}{\\sum_ {j = 1} ^ {C} e ^ {\\cos \\left(\\mathbf {z} _ {i} ^ {l} , \\mathbf {t} _ {j} ^ {f}\\right) / \\tau_ {0}} + K e ^ {\\cos \\left(\\mathbf {z} _ {i} ^ {l} , \\mathbf {t} _ {j} ^ {o}\\right) / \\tau_ {0}}}, \\tag {7}", + "image_path": "5cb592dd1ae00b71004d8646c8642f1b0472c7f7b028db53cf26bcb79aad167a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i^l (i\\in \\{1,\\dots ,N\\})" + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": " represents " + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": " extracted local features generated by CLIP's image encoder [32, 33]. We set " + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\tau_0 = 1" + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": " during model inference." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 72, + 395, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 395, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 395, + 84 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 91, + 434, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 91, + 434, + 104 + ], + "spans": [ + { + "bbox": [ + 314, + 91, + 434, + 104 + ], + "type": "text", + "content": "5.1. Experimental details" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 110, + 555, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 110, + 555, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 110, + 555, + 276 + ], + "type": "text", + "content": "Datasets. We use a popular benchmark for conventional OOD detection [19, 30, 32, 33], where ImageNet-1k [7] serves as the ID dataset, and the OOD datasets are iNaturalist [17], SUN [47], Places [52], and Texture [6]. Moreover, inspired by existing studies [1, 2], we use datasets like OpenImage-O [45], NINCO [2], and ImageNet-O [15], which are cleaner and more realistic, to simulate the more challenging OOD detection. Besides, we also utilize other widely adopted datasets for few-shot settings as ID datasets [49, 51, 54], which include StanfordCars [21], UCF101 [40], Caltech101 [12], Flowers102 [35], EuroSAT [13], FGVCAircraft [29], OxfordPets [36], and Food101 [3], considering factors such as image resolution and the number of classes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 277, + 555, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 555, + 445 + ], + "type": "text", + "content": "Setup. Following previous studies [33], we use the ViTB/16 [8] as the backbone model. For our model, when using ImageNet-1k as the ID dataset, the epochs are respectively set to 30 and 50 for the 1-shot and 16-shot settings, while for other ID datasets, we set the epoch to 200 following CoOp [54]. The value of forced coefficient " + }, + { + "bbox": [ + 313, + 277, + 555, + 445 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 277, + 555, + 445 + ], + "type": "text", + "content": " is uniformly set to 3, which will be further discussed in the sensitivity study. Other hyperparameters are as follows: learning rate " + }, + { + "bbox": [ + 313, + 277, + 555, + 445 + ], + "type": "inline_equation", + "content": "= 2\\mathrm{e} - 3" + }, + { + "bbox": [ + 313, + 277, + 555, + 445 + ], + "type": "text", + "content": ", batch size " + }, + { + "bbox": [ + 313, + 277, + 555, + 445 + ], + "type": "inline_equation", + "content": "= 160" + }, + { + "bbox": [ + 313, + 277, + 555, + 445 + ], + "type": "text", + "content": ", SGD (momentum " + }, + { + "bbox": [ + 313, + 277, + 555, + 445 + ], + "type": "inline_equation", + "content": "= 0.9" + }, + { + "bbox": [ + 313, + 277, + 555, + 445 + ], + "type": "text", + "content": ", weight decay " + }, + { + "bbox": [ + 313, + 277, + 555, + 445 + ], + "type": "inline_equation", + "content": "= 5\\mathrm{e} - 4" + }, + { + "bbox": [ + 313, + 277, + 555, + 445 + ], + "type": "text", + "content": ") as the optimizer with a cosine scheduler, " + }, + { + "bbox": [ + 313, + 277, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\tau = 1" + }, + { + "bbox": [ + 313, + 277, + 555, + 445 + ], + "type": "text", + "content": " in Eq. (3) and Eq. (4). All experiments on our model can be conducted on a single Nvidia A30 GPU. The average experiment results (including our reproductions) over four runs are reported for comparison." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 446, + 555, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 446, + 555, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 446, + 555, + 566 + ], + "type": "text", + "content": "Comparison Methods. To validate the effectiveness of our FA fairly, we mainly compare it with CLIP-based OOD detection methods that do not use real outliers (e.g. OOD labels). For previous post-hoc methods, including MSP [14], ODIN [24], Energy [27], ReAct [41], and MaxLogit [16], we adapt these methods with the CLIP image encoder as the CLIP-based post-hoc methods. For zero-shot methods, we select MCM [30], GL-MCM [32] and CLIPN [46] as baselines. For prompt learning methods, we adopt CoOp [54], LoCoOp [33], IDLike [1], and SCT [48] as baselines." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 566, + 553, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 566, + 553, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 566, + 553, + 626 + ], + "type": "text", + "content": "Metrics. For evaluation, we adopt the following metrics: (1) The False Positive Rate at " + }, + { + "bbox": [ + 313, + 566, + 553, + 626 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 313, + 566, + 553, + 626 + ], + "type": "text", + "content": " True Positive Rate for in-distribution samples (FPR95); (2) The Area Under the Receiver Operating Characteristic Curve (AUROC); (3) In-distribution data classification Top-1 accuracy (ID ACC)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 635, + 395, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 635, + 395, + 647 + ], + "spans": [ + { + "bbox": [ + 314, + 635, + 395, + 647 + ], + "type": "text", + "content": "5.2. Main results" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "content": "Conventional OOD Detection. Table 1 summarizes our comparison results on the ImageNet-1k benchmarks, which show that our FA achieves state-of-the-art OOD detection performance among other CLIP-based methods under different few-shot settings (More details under the 4-shot set" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1156" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 69, + 555, + 456 + ], + "blocks": [ + { + "bbox": [ + 55, + 69, + 555, + 456 + ], + "lines": [ + { + "bbox": [ + 55, + 69, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 69, + 555, + 456 + ], + "type": "table", + "html": "
MethodiNaturalistSUNPlacesTexturesAverage
FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑
Zero-shot methods
MCM31.9594.1637.2292.5542.9890.1058.3585.8342.6390.66
GL-MCM15.0996.7229.0893.4137.0790.3758.9483.1135.0490.90
CLIPN19.1796.1726.4394.0232.2692.6241.2390.1230.2193.19
CLIP-based post-hoc methods
\\( MSP^† \\)74.5777.7476.9573.9779.7272.1873.6674.8476.2274.68
\\( ODIN^† \\)98.9357.7388.7278.4287.8076.8885.4771.4990.2371.13
\\( Energy^† \\)64.9887.1846.4291.1757.4087.3350.3988.2254.8088.48
\\( ReAct^† \\)65.5786.8746.1791.0456.8587.4249.8888.1354.6288.37
\\( MaxLogit^† \\)60.8888.0344.8391.1655.5487.4548.7288.6352.4988.82
Prompt learning based methods
\\( CoOp_{MCM} \\)41.14±9.3991.47±2.1239.06±4.1791.68±0.8345.38±4.8889.16±1.2851.37±3.0987.82±1.1244.24±1.3990.03±0.32
\\( CoOp_{GL} \\)23.30±6.3794.59±1.5132.08±3.8792.21±0.9539.22±4.5589.59±1.5855.78±3.4683.26±1.3837.59±0.3489.91±0.31
\\( LoCoOp_{MCM} \\)36.64±4.8792.78±0.9331.86±3.7993.50±0.9238.81±3.3790.70±0.9748.11±2.3189.43±0.6838.85±2.6791.59±0.58
\\( LoCoOp_{GL} \\)21.97±2.9195.39±0.5924.95±2.3794.42±0.5934.14±2.6691.14±0.6949.04±2.7787.73±0.9632.53±2.1992.17±0.47
IDLike17.73±1.9196.68±0.3148.17±1.3989.53±0.5850.43±6.5888.27±2.2729.12±7.6493.25±2.1636.36±3.8691.93±1.18
\\( LSN^† \\)59.28±7.0287.20±3.1540.15±0.8291.47±0.1446.11±1.8688.74±0.5760.34±0.1483.92±0.4251.47±1.5387.84±0.58
\\( NegPrompt^† \\)65.03±8.6984.56±2.5244.39±1.6689.63±0.6651.31±6.2186.55±2.1987.60±1.6163.76±3.0262.08±3.7181.13±1.78
\\( SCT_{MCM} \\)41.93±12.1791.77±2.4030.39±2.1793.76±0.3938.73±2.5490.78±0.3446.78±3.5188.96±1.1639.46±3.3991.32±0.90
\\( SCT_{GL} \\)20.57±10.295.63±1.8924.56±3.0394.39±0.5333.27±2.9691.27±0.5748.12±2.9786.76±0.8931.62±3.1992.01±0.77
\\( FA_{MCM}(Ours) \\)25.50±2.7294.72±0.2736.24±3.1592.40±0.9535.38±2.7591.99±0.6828.34±0.8793.95±0.1931.37±1.0793.25±0.32
\\( FA_{GL}(Ours) \\)14.12±1.3296.76±0.1029.99±1.5792.95±0.6632.48±1.4891.83±0.4934.66±1.2191.50±0.3627.81±0.4493.26±0.27
Prompt learning based methods
\\( CoOp_{MCM} \\)30.26±1.9893.43±0.8134.69±0.4392.59±0.1441.91±0.7190.11±0.2344.68±2.1189.95±0.4837.89±0.7191.52±0.29
\\( CoOp_{GL} \\)15.96±1.6796.11±0.5527.26±1.9993.29±0.4635.36±2.0890.58±0.6448.63±2.1186.11±0.5931.81±1.2791.51±0.39
\\( LoCoOp_{MCM} \\)27.35±3.1994.12±0.8030.93±1.1993.75±0.2638.26±1.5391.12±0.2741.36±2.5690.99±0.5334.47±0.7392.49±0.14
\\( LoCoOp_{GL} \\)18.46±1.3895.85±0.6322.43±0.9695.15±0.2231.53±1.5292.15±0.2243.35±3.1289.38±0.7828.94±1.2993.13±0.17
IDLike19.23±10.596.70±1.5854.15±2.8887.64±1.1956.63±0.0785.86±0.4434.69±6.4191.90±2.3141.18±1.7390.53±0.01
\\( LSN^† \\)36.17±4.8192.66±1.1634.27±0.4493.53±0.2041.47±0.8590.52±0.3746.43±0.6089.38±0.2439.58±0.7391.53±0.09
\\( NegPrompt^† \\)37.79±0.1190.49±0.0132.11±3.7792.25±1.0035.52±0.4191.16±0.0343.93±9.0988.38±3.3137.34±1.4190.57±0.59
\\( SCT_{MCM} \\)29.41±2.1993.76±0.5527.28±2.8094.22±0.4036.35±2.1391.16±0.3342.25±1.8990.52±0.4833.82±1.7892.42±0.35
\\( SCT_{GL} \\)15.19±2.1696.71±0.4420.00±0.6195.57±0.1129.71±0.8592.37±0.1244.17±0.8288.59±0.3927.27±0.4493.31±0.17
\\( FA_{MCM}(Ours) \\)25.79±1.4894.29±0.3533.54±1.1793.02±0.1933.77±1.6492.64±0.4223.17±1.3195.14±0.2629.07±1.1193.77±0.19
\\( FA_{GL}(Ours) \\)14.49±1.2796.48±0.2927.65±1.0893.46±0.1831.09±1.3892.44±0.3429.50±0.8992.93±0.1825.68±0.5893.82±0.11
", + "image_path": "df04df2d128cd259127b9519eb43a98675da97ff4d11497a8e58d74958be8209.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 72, + 531, + 279, + 617 + ], + "blocks": [ + { + "bbox": [ + 55, + 460, + 555, + 517 + ], + "lines": [ + { + "bbox": [ + 55, + 460, + 555, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 555, + 517 + ], + "type": "text", + "content": "Table 1. Comparison results on ImageNet-1k OOD benchmarks. All results using the same backbone ViT-B/16. The results marked with " + }, + { + "bbox": [ + 55, + 460, + 555, + 517 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 460, + 555, + 517 + ], + "type": "text", + "content": " are taken from [48]; the others are our reproductions. The prompt learning based methods are run under four trials, reporting the mean and standard deviation of the performance. The subscripts " + }, + { + "bbox": [ + 55, + 460, + 555, + 517 + ], + "type": "inline_equation", + "content": "{}_{\\mathrm{{MCM}}}" + }, + { + "bbox": [ + 55, + 460, + 555, + 517 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 460, + 555, + 517 + ], + "type": "inline_equation", + "content": "{}_{\\mathrm{{GL}}}" + }, + { + "bbox": [ + 55, + 460, + 555, + 517 + ], + "type": "text", + "content": " indicate the use of the MCM score and the GL-MCM score. The best and second-best results are indicated in bold and underline. " + }, + { + "bbox": [ + 55, + 460, + 555, + 517 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 55, + 460, + 555, + 517 + ], + "type": "text", + "content": " indicates larger values are better; " + }, + { + "bbox": [ + 55, + 460, + 555, + 517 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 55, + 460, + 555, + 517 + ], + "type": "text", + "content": " indicates smaller values are better. All values are percentages." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 72, + 531, + 279, + 617 + ], + "lines": [ + { + "bbox": [ + 72, + 531, + 279, + 617 + ], + "spans": [ + { + "bbox": [ + 72, + 531, + 279, + 617 + ], + "type": "table", + "html": "
Method1-shot4-shot16-shot
CoOp67.44±0.5069.71±0.0770.99±0.14
LoCoOp67.40±0.6469.55±0.1071.53±0.17
IDLike68.17±0.5768.91±0.1469.46±0.02
SCT68.63±0.1369.93±0.2271.78±0.05
FA(Ours)68.67±0.3969.96±0.0471.02±0.09
", + "image_path": "455dd344dd79e267e82084b1808e48db370c8956b7fe646176768df80503b981.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 621, + 295, + 654 + ], + "lines": [ + { + "bbox": [ + 55, + 621, + 295, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 621, + 295, + 654 + ], + "type": "text", + "content": "Table 2. Comparison results in ID Top-1 accuracy on ImageNet-1k under different few-shot settings. Other notations are the same as Tab. 1." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "ting are provided in the Appendix A). Specifically, in the 16-shot setting, FA with GL-MCM score " + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\left(\\mathrm{FA}_{\\mathrm{GL}}\\right)" + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": " outperforms the best baseline " + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{SCT}_{\\mathrm{GL}}" + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": " (average FPR95 of " + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "inline_equation", + "content": "25.68\\%" + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "inline_equation", + "content": "27.27\\%" + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": ", and average AUROC of " + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "inline_equation", + "content": "93.82\\%" + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "inline_equation", + "content": "93.31\\%" + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 533, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 617 + ], + "type": "text", + "content": "Additionally, in the 1-shot setting, " + }, + { + "bbox": [ + 313, + 533, + 555, + 617 + ], + "type": "inline_equation", + "content": "\\mathrm{FA}_{\\mathrm{GL}}" + }, + { + "bbox": [ + 313, + 533, + 555, + 617 + ], + "type": "text", + "content": " surpasses " + }, + { + "bbox": [ + 313, + 533, + 555, + 617 + ], + "type": "inline_equation", + "content": "\\mathrm{SCT}_{\\mathrm{GL}}" + }, + { + "bbox": [ + 313, + 533, + 555, + 617 + ], + "type": "text", + "content": " by a large margin, showing improvements of at least " + }, + { + "bbox": [ + 313, + 533, + 555, + 617 + ], + "type": "inline_equation", + "content": "1.25\\%" + }, + { + "bbox": [ + 313, + 533, + 555, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 533, + 555, + 617 + ], + "type": "inline_equation", + "content": "3.81\\%" + }, + { + "bbox": [ + 313, + 533, + 555, + 617 + ], + "type": "text", + "content": " on average AUROC and FPR95. More importantly, even the average score of " + }, + { + "bbox": [ + 313, + 533, + 555, + 617 + ], + "type": "inline_equation", + "content": "\\mathrm{FA}_{\\mathrm{MCM}}" + }, + { + "bbox": [ + 313, + 533, + 555, + 617 + ], + "type": "text", + "content": " has exceeded " + }, + { + "bbox": [ + 313, + 533, + 555, + 617 + ], + "type": "inline_equation", + "content": "\\mathrm{SCT}_{\\mathrm{GL}}" + }, + { + "bbox": [ + 313, + 533, + 555, + 617 + ], + "type": "text", + "content": " in the 1-shot setting. Moreover, our FA effectively improves the OOD performance without sacrificing the ID classification capability of the model, as shown in Tab. 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 617, + 556, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 700 + ], + "type": "text", + "content": "In particular, we can also observe that our method performs not well on the SUN dataset, as shown in Tab. 1. This is related to the presence of a significant number of samples belonging to ID category in these datasets, as demonstrated by recent studies [1]. In other words, the SUN dataset in the popular OOD benchmark requires more thorough cleaning to be effectively used as an OOD dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 700, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 700, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 700, + 555, + 714 + ], + "type": "text", + "content": "Challenging OOD Detection. Recent studies [1, 2] have" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1157" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 555, + 188 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 555, + 188 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 555, + 188 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 555, + 188 + ], + "type": "table", + "html": "
MethodOpenImage-ONINCOImageNet-OAverage
FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑
CoOpMCM37.95±0.5991.99±0.2278.00±0.2373.00±0.5970.14±0.8681.34±0.8362.03±0.3882.11±0.48
CoOpGL32.23±1.0092.22±0.2572.04±0.8175.35±0.8867.35±1.1978.44±0.9457.21±0.9681.99±0.58
LoCoOpMCM37.03±0.9492.24±0.2177.96±0.6372.78±0.7970.96±0.7581.25±0.3961.98±0.5582.09±0.35
LoCoOpGL33.87±1.2392.53±0.0775.16±0.9272.97±0.3968.06±1.9281.19±0.2559.03±1.2782.23±0.17
IDLike54.43±2.3587.89±0.4578.93±1.7969.32±0.4284.08±1.4567.45±0.8672.48±0.8974.89±0.28
SCTMCM37.28±1.1692.04±0.3378.51±0.8671.09±1.1971.45±0.5481.29±0.1562.41±0.7381.48±0.52
SCTGL32.24±0.3492.75±0.1774.15±0.5373.14±0.9468.35±0.6280.84±0.4558.25±0.3982.24±0.44
FAmCM(Ours)37.93±1.6591.84±0.4570.44±1.2477.77±0.9463.84±1.3682.71±0.3057.40±1.3484.08±0.47
FAGL(Ours)32.10±0.8992.39±0.2865.61±1.3279.01±0.7163.13±0.9880.39±0.2153.61±0.8983.93±0.31
", + "image_path": "686b772dca172fa9cee16de40529ad055d9d0de180ebf534e139582902000e4e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 215, + 555, + 434 + ], + "blocks": [ + { + "bbox": [ + 73, + 192, + 536, + 204 + ], + "lines": [ + { + "bbox": [ + 73, + 192, + 536, + 204 + ], + "spans": [ + { + "bbox": [ + 73, + 192, + 536, + 204 + ], + "type": "text", + "content": "Table 3. Challenging OOD detection results on cleaner OOD datasets in the 16-shot setting. We use the same notation as Tab. 1." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 215, + 555, + 434 + ], + "lines": [ + { + "bbox": [ + 56, + 215, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 56, + 215, + 555, + 434 + ], + "type": "table", + "html": "
ID DatasetMethodiNaturalistSUNPlacesTexturesAverage
FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑
UCF101CoOpMCM14.40±6.6197.41±1.3637.09±3.9493.14±0.4840.07±2.7791.89±0.6533.24±4.2193.19±1.2231.21±3.5893.91±0.45
CoOpGL9.65±6.3598.18±0.9828.54±5.0794.51±0.9232.94±1.7493.28±0.3235.96±2.6591.83±1.4226.77±3.2294.45±0.42
LoCoOpMCM2.38±0.9899.21±0.2219.73±4.2496.11±0.7822.57±3.0194.96±0.7715.89±3.1296.12±0.6715.14±1.9196.60±0.33
LoCoOpGL1.13±0.2299.65±0.0917.93±4.0996.51±0.8922.34±4.5895.03±1.0618.17±3.0195.25±0.7114.89±2.4996.61±0.58
IDLike66.63±27.689.75±5.5287.88±6.5678.04±6.1484.47±12.176.04±8.5349.07±11.291.19±2.7472.01±14.083.76±5.64
SCTMCM1.51±0.5199.44±0.1024.17±5.5495.21±0.8826.62±4.4193.75±0.8115.34±3.8495.98±0.6416.91±3.2096.09±0.48
SCTGL0.79±0.2799.75±0.0418.69±2.4195.86±0.5623.77±2.4793.98±0.4717.44±2.6294.52±0.5415.17±1.5696.03±0.25
FAmCM(Ours)0.12±0.1199.94±0.013.06±1.3399.34±0.234.68±1.0099.08±0.262.68±0.5199.31±0.432.63±0.5899.42±0.22
FAGL(Ours)0.13±0.0299.94±0.024.98±1.0299.01±0.227.46±0.8698.44±0.275.67±0.8398.75±0.194.56±0.4799.03±0.13
EuroSATCoOpMCM99.14±0.7331.95±9.1798.29±0.4944.36±4.1997.28±0.8547.51±4.6386.69±4.3769.17±3.1595.35±1.0348.25±3.09
CoOpGL92.45±4.2657.59±10.490.97±3.4961.89±4.8388.69±3.8564.00±4.9262.37±1.9383.58±2.3183.62±1.1866.77±2.83
LoCoOpMCM96.51±3.3353.49±8.0693.52±5.6855.59±9.1992.88±4.0557.89±5.1976.45±7.4878.59±2.6689.84±4.2961.39±3.90
LoCoOpGL86.13±12.272.68±7.9787.81±6.3768.58±6.4085.93±4.1569.90±3.4054.89±4.6787.00±1.0178.69±4.6574.54±3.59
IDLike98.02±2.1373.32±5.8395.74±2.8465.19±6.8694.79±3.6769.44±5.0663.34±9.1086.68±2.9887.97±4.2173.66±5.04
SCTMCM99.79±0.2232.90±6.7398.36±0.8444.30±3.4197.77±0.8648.08±1.6483.98±6.9273.25±3.5894.97±2.1949.63±2.39
SCTGL96.57±5.4158.21±5.3192.41±4.5561.54±3.4790.49±5.0263.33±2.7155.29±14.785.48±3.6883.69±7.0267.14±2.20
FAmCM(Ours)85.24±6.8878.62±3.1836.24±10.892.13±2.4428.17±10.294.05±1.6810.36±3.9797.71±0.6939.99±7.7990.63±1.91
FAGL(Ours)77.82±7.2886.88±2.1556.10±8.0388.43±2.2649.87±7.6989.07±1.498.99±2.5197.74±0.3648.19±6.2390.53±1.42
", + "image_path": "c89bb2e0ccfc86f226a8bd042677dedbc56597f4d2e5952389cc4f3c7082576e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 437, + 539, + 449 + ], + "lines": [ + { + "bbox": [ + 69, + 437, + 539, + 449 + ], + "spans": [ + { + "bbox": [ + 69, + 437, + 539, + 449 + ], + "type": "text", + "content": "Table 4. Representative OOD detection results with various ID datasets in the 16-shot setting. We use the same notation as Tab. 1." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 67, + 470, + 280, + 570 + ], + "blocks": [ + { + "bbox": [ + 67, + 470, + 280, + 570 + ], + "lines": [ + { + "bbox": [ + 67, + 470, + 280, + 570 + ], + "spans": [ + { + "bbox": [ + 67, + 470, + 280, + 570 + ], + "type": "image", + "image_path": "076c5ec2f75ee8cd53e39c8b2940dcccfd6dbd0bae3b8b4fc0d38aacd046cb8c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 572, + 295, + 628 + ], + "lines": [ + { + "bbox": [ + 55, + 572, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 572, + 295, + 628 + ], + "type": "text", + "content": "Figure 3. A portion of OOD detection average performance using other ID datasets under the 16-shot setting. We report the mean and standard deviation of the AUROC for our FA (blue) and CoOp (orange) using the MCM score. More details can be found in the Appendix B." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": "shown that some of the used OOD datasets contain more or less images of objects that belong to ID classes in ImageNet-1k [7]. Hence, probing the true performance of OOD detectors for ImageNet-1k demands OOD datasets that are both challenging and genuinely OOD. As shown in Tab. 3, to fully substantiate the effectiveness of our" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 470, + 555, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 470, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 555, + 555 + ], + "type": "text", + "content": "FA, we leverage 3 additional cleaner and more challenging OOD datasets, including OpenImage-O [45], NINCO [2], and ImageNet-O [15], inspired by [2]. Notably, both our " + }, + { + "bbox": [ + 313, + 470, + 555, + 555 + ], + "type": "inline_equation", + "content": "\\mathrm{FA}_{\\mathrm{MCM}}" + }, + { + "bbox": [ + 313, + 470, + 555, + 555 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 470, + 555, + 555 + ], + "type": "inline_equation", + "content": "\\mathrm{FA}_{\\mathrm{GL}}" + }, + { + "bbox": [ + 313, + 470, + 555, + 555 + ], + "type": "text", + "content": " consistently outperform current SOTA methods in terms of average FPR95 and AUROC under different few-shot settings (More details under the 1-shot and 4-shot settings are provided in the Appendix A)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": "Various ID Datasets. Considering factors like dataset variety, image resolution, and the number of classes [30], we also leverage 8 additional widely used datasets for few-shot setting as ID datasets for comparison [30, 34, 54]. Overall, despite our model is simple, it achieves superior average OOD performance on these ID datasets compared to other methods. We selected representative results using UCF101 [40] and EuroSAT [13] as ID datasets in the 16-shot setting, as shown in Tab. 4. Since using the other datasets (i.e. Food101 [3], StandfordCars [21], Caltech101 [12], FGVCAircraft [29], Flowers102 [35], and OxfordPets [36]) as ID datasets for OOD detection presents relatively lower difficulty as shown in Fig. 3, we leave more" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1158" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 70, + 172, + 151 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 172, + 151 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 172, + 151 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 172, + 151 + ], + "type": "image", + "image_path": "3179384cf6c5f32757ff2ec23c1096e85edf7724d0bee176cc0262fa0bb2070b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 152, + 168, + 159 + ], + "lines": [ + { + "bbox": [ + 72, + 152, + 168, + 159 + ], + "spans": [ + { + "bbox": [ + 72, + 152, + 168, + 159 + ], + "type": "text", + "content": "(a) Conventional OOD benchmark" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 178, + 70, + 282, + 151 + ], + "blocks": [ + { + "bbox": [ + 178, + 70, + 282, + 151 + ], + "lines": [ + { + "bbox": [ + 178, + 70, + 282, + 151 + ], + "spans": [ + { + "bbox": [ + 178, + 70, + 282, + 151 + ], + "type": "image", + "image_path": "473ae12adfd35b414a654564acef41ba335a1dae659dd6e9bbf391ca6c4e51ab.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 152, + 277, + 159 + ], + "lines": [ + { + "bbox": [ + 184, + 152, + 277, + 159 + ], + "spans": [ + { + "bbox": [ + 184, + 152, + 277, + 159 + ], + "type": "text", + "content": "(b)Challenging OOD benchmark" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 166, + 295, + 222 + ], + "lines": [ + { + "bbox": [ + 55, + 166, + 295, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 166, + 295, + 222 + ], + "type": "text", + "content": "Figure 4. Sensitivity study of hyperparameter " + }, + { + "bbox": [ + 55, + 166, + 295, + 222 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 166, + 295, + 222 + ], + "type": "text", + "content": ". We report the average FPR95 (Blue) and AUROC (Orange) of our " + }, + { + "bbox": [ + 55, + 166, + 295, + 222 + ], + "type": "inline_equation", + "content": "\\mathrm{FA}_{\\mathrm{GL}}" + }, + { + "bbox": [ + 55, + 166, + 295, + 222 + ], + "type": "text", + "content": " on different OOD benchmarks using ImageNet-1k as the ID dataset under 16-shot setting. Dashed lines represent the performance of the best baseline (i.e. " + }, + { + "bbox": [ + 55, + 166, + 295, + 222 + ], + "type": "inline_equation", + "content": "\\mathrm{SCT}_{\\mathrm{GL}}" + }, + { + "bbox": [ + 55, + 166, + 295, + 222 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 228, + 203, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 203, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 203, + 240 + ], + "type": "text", + "content": "experimental details in Appendix B." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 240, + 296, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 240, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 296, + 300 + ], + "type": "text", + "content": "In particular, for the ID dataset EuroSAT, " + }, + { + "bbox": [ + 55, + 240, + 296, + 300 + ], + "type": "inline_equation", + "content": "\\mathrm{FA}_{\\mathrm{MCM}}" + }, + { + "bbox": [ + 55, + 240, + 296, + 300 + ], + "type": "text", + "content": " surpasses the best-competing method " + }, + { + "bbox": [ + 55, + 240, + 296, + 300 + ], + "type": "inline_equation", + "content": "\\mathrm{LoCoOp_{GL}}" + }, + { + "bbox": [ + 55, + 240, + 296, + 300 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 55, + 240, + 296, + 300 + ], + "type": "inline_equation", + "content": "16.09\\%" + }, + { + "bbox": [ + 55, + 240, + 296, + 300 + ], + "type": "text", + "content": " in average AUROC and " + }, + { + "bbox": [ + 55, + 240, + 296, + 300 + ], + "type": "inline_equation", + "content": "38.7\\%" + }, + { + "bbox": [ + 55, + 240, + 296, + 300 + ], + "type": "text", + "content": " in average FPR95, underscoring its effectiveness and versatility even when the ID dataset is of low resolution (64x64 pixels)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 309, + 147, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 309, + 147, + 322 + ], + "spans": [ + { + "bbox": [ + 55, + 309, + 147, + 322 + ], + "type": "text", + "content": "5.3. Ablation study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 327, + 296, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 399 + ], + "type": "text", + "content": "Influence of the " + }, + { + "bbox": [ + 55, + 327, + 296, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{FCE-K}" + }, + { + "bbox": [ + 55, + 327, + 296, + 399 + ], + "type": "text", + "content": ". As shown in Tab. 5, we trained the model " + }, + { + "bbox": [ + 55, + 327, + 296, + 399 + ], + "type": "inline_equation", + "content": "\\mathrm{FA}_{CE}" + }, + { + "bbox": [ + 55, + 327, + 296, + 399 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 327, + 296, + 399 + ], + "type": "inline_equation", + "content": "\\mathrm{FA}_{FCE-K}" + }, + { + "bbox": [ + 55, + 327, + 296, + 399 + ], + "type": "text", + "content": " using the standard cross-entropy loss " + }, + { + "bbox": [ + 55, + 327, + 296, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CE}" + }, + { + "bbox": [ + 55, + 327, + 296, + 399 + ], + "type": "text", + "content": " and our " + }, + { + "bbox": [ + 55, + 327, + 296, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{FCE-K}" + }, + { + "bbox": [ + 55, + 327, + 296, + 399 + ], + "type": "text", + "content": ", respectively. Our " + }, + { + "bbox": [ + 55, + 327, + 296, + 399 + ], + "type": "inline_equation", + "content": "\\mathrm{FA}_{FCE-K}" + }, + { + "bbox": [ + 55, + 327, + 296, + 399 + ], + "type": "text", + "content": " significantly outperforms " + }, + { + "bbox": [ + 55, + 327, + 296, + 399 + ], + "type": "inline_equation", + "content": "\\mathrm{FA}_{CE}" + }, + { + "bbox": [ + 55, + 327, + 296, + 399 + ], + "type": "text", + "content": " under the same inference setting, which demonstrate the effectiveness of our " + }, + { + "bbox": [ + 55, + 327, + 296, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{FCE-K}" + }, + { + "bbox": [ + 55, + 327, + 296, + 399 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 399, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 296, + 506 + ], + "type": "text", + "content": "Initialization of the forced and original prompt. As shown in Tab. 6, we perform experiments comparing different initialization scenarios for both prompts. Manual initialization uses the embeddings of \"a photo of a [class-c]\" to initialize the prompt's embeddings, while the random initialization setting is based on CoOp [54]. The results show that employing consistent manual initialization for both prompts enables the forced prompt to effectively learn more specific descriptions of the ID classes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 507, + 296, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 507, + 296, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 507, + 296, + 592 + ], + "type": "text", + "content": "Influence of the shared learnable vector. We also explore the influence of the shared learnable vector and the independent learnable vector for the forced prompts. As shown in Tab. 7, we can observe that using the shared learnable vector performs well in the few-shot setting since the independent learnable vector for each class has more parameters and requires more training data [54]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 600, + 155, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 155, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 155, + 613 + ], + "type": "text", + "content": "5.4. Sensitivity study" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": "As shown in Fig. 4, we explore the influence of the hyperparameter " + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": ". Overall, with " + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": " ranging from 1 to 6, the results indicate that our proposed framework is insensitive to the choice of " + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": ". Concretely, our method is stable and outperforms the best baseline " + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{SCT}_{\\mathrm{GL}}" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": " in average AUROC and FPR95 on both conventional and challenging OOD detection benchmarks. Specially, as " + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": " increases, the average AUROC shows an upward trend and then tends to be sta" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "type": "text", + "content": "ble (see Fig. 4 (a)), showing a bottleneck in capturing more comprehensive semantic representations of the ID classes." + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 319, + 108, + 550, + 162 + ], + "blocks": [ + { + "bbox": [ + 319, + 108, + 550, + 162 + ], + "lines": [ + { + "bbox": [ + 319, + 108, + 550, + 162 + ], + "spans": [ + { + "bbox": [ + 319, + 108, + 550, + 162 + ], + "type": "table", + "html": "
MethodMCMGL-MCM
FPR95↓AUROC↑FPR95↓AUROC↑
FAce41.43±3.1191.01±0.6734.29±1.9490.99±0.56
FAFCE-K29.07±1.1193.77±0.1925.68±0.5893.82±0.11
", + "image_path": "0d2ab2d3dc31e25d8dcba5b51277f58ee42c764870cb61853ec9ae9292522cbb.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 319, + 204, + 550, + 281 + ], + "blocks": [ + { + "bbox": [ + 313, + 167, + 553, + 190 + ], + "lines": [ + { + "bbox": [ + 313, + 167, + 553, + 190 + ], + "spans": [ + { + "bbox": [ + 313, + 167, + 553, + 190 + ], + "type": "text", + "content": "Table 5. Influence of the " + }, + { + "bbox": [ + 313, + 167, + 553, + 190 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{FCE-K}" + }, + { + "bbox": [ + 313, + 167, + 553, + 190 + ], + "type": "text", + "content": ". MCM and GL-MCM refer to the score functions used during model inference." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 204, + 550, + 281 + ], + "lines": [ + { + "bbox": [ + 319, + 204, + 550, + 281 + ], + "spans": [ + { + "bbox": [ + 319, + 204, + 550, + 281 + ], + "type": "table", + "html": "
Forced-M-InitOriginal-M-InitFA MCMFA GL
FPR95↓AUROC↑FPR95↓AUROC↑
--40.71±4.0091.28±0.8234.88±2.9292.12±0.76
✓-40.81±2.2591.25±0.6436.03±2.8491.62±1.01
-✓31.35±1.3493.35±0.3426.39±1.0293.64±0.19
✓✓29.07±1.1193.77±0.1925.68±0.5893.82±0.11
", + "image_path": "3a5bbe0446b33e23b45d99fc3a7d3742084c89f429d48ecaf48249aafdb4719f.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 319, + 345, + 550, + 423 + ], + "blocks": [ + { + "bbox": [ + 313, + 286, + 553, + 331 + ], + "lines": [ + { + "bbox": [ + 313, + 286, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 286, + 553, + 331 + ], + "type": "text", + "content": "Table 6. Ablation study of different initialization scenarios for the forced and original prompts. \"M-Init\" represents manual initialization. " + }, + { + "bbox": [ + 313, + 286, + 553, + 331 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 313, + 286, + 553, + 331 + ], + "type": "text", + "content": " represents using manual initialization; - represents using random initialization." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 345, + 550, + 423 + ], + "lines": [ + { + "bbox": [ + 319, + 345, + 550, + 423 + ], + "spans": [ + { + "bbox": [ + 319, + 345, + 550, + 423 + ], + "type": "table", + "html": "
Shared- VectorM-InitFA MCMFA GL
FPR95↓AUROC↑FPR95↓AUROC↑
--44.02±2.2490.17±0.4246.29±1.3388.24±0.41
✓-40.71±4.0091.28±0.8234.88±2.9292.12±0.76
-✓34.69±0.6292.57±0.1229.27±0.5592.82±0.80
✓✓29.07±1.1193.77±0.1925.68±0.5893.82±0.11
", + "image_path": "c2c3d01904873d81b6eba1a93aff2071de38602fb85df39b9cc6d5be0849582c.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 428, + 554, + 473 + ], + "lines": [ + { + "bbox": [ + 313, + 428, + 554, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 428, + 554, + 473 + ], + "type": "text", + "content": "Table 7. Ablation study of the shared learnable vector. " + }, + { + "bbox": [ + 313, + 428, + 554, + 473 + ], + "type": "inline_equation", + "content": "\\sqrt{}" + }, + { + "bbox": [ + 313, + 428, + 554, + 473 + ], + "type": "text", + "content": " - for \"Shared-vector\" represents using the shared / independent learnable vector for the forced prompt; " + }, + { + "bbox": [ + 313, + 428, + 554, + 473 + ], + "type": "inline_equation", + "content": "\\sqrt{}" + }, + { + "bbox": [ + 313, + 428, + 554, + 473 + ], + "type": "text", + "content": " - for \"M-Init\" represents using manual / random initialization for the forced prompt." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 314, + 489, + 388, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 489, + 388, + 502 + ], + "spans": [ + { + "bbox": [ + 314, + 489, + 388, + 502 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": "In this study, we propose a novel CLIP-based framework for OOD detection based on Forced prompt leArning (FA), which focuses on fully exploiting the ID knowledge to effectively improve OOD detection without exploring complex OOD-related knowledge. We introduce a learnable forced prompt in addition to the frozen original prompt, both of which are using the same manual initialization. The forced prompt treats the original prompt as a reference, forcing itself to learn more diversified semantic descriptions of the ID classes rather than being limited to the textual semantics of class labels. We also introduce a forced coefficient to facilitate the forced prompt in learning more nuanced descriptions of the ID classes. Comprehensive experimental evaluations demonstrate that our method consistently surpasses current state-of-the-art methods on diverse benchmarks. We expect that our study can bring new insight on VLMs-based OOD detection methods and inspire more future research." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1159" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "text", + "content": "7. Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 151 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 151 + ], + "type": "text", + "content": "This work is supported in part by the National Natural Science Foundation of China (grant No. 62071502), the Major Key Project of PCL (grant No. PCL2023A09), and Guangdong Excellent Youth Team Program (grant No. 2023B1515040025)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 161, + 115, + 173 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 161, + 115, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 161, + 115, + 173 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 180, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 180, + 296, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 180, + 296, + 224 + ], + "spans": [ + { + "bbox": [ + 61, + 180, + 296, + 224 + ], + "type": "text", + "content": "[1] Yichen Bai, Zongbo Han, Bing Cao, Xiaoheng Jiang, Qinghua Hu, and Changqing Zhang. Id-like prompt learning for few-shot out-of-distribution detection. In CVPR, pages 17480-17489, 2024. 2, 3, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 225, + 296, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 225, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 61, + 225, + 296, + 258 + ], + "type": "text", + "content": "[2] Julian Bitterwolf, Maximilian Müller, and Matthias Hein. In or out? fixing imagenet out-of-distribution detection evaluation. In ICML, pages 2471-2506, 2023. 5, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 259, + 295, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 259, + 295, + 292 + ], + "spans": [ + { + "bbox": [ + 61, + 259, + 295, + 292 + ], + "type": "text", + "content": "[3] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 - mining discriminative components with random forests. In ECCV, pages 446-461, 2014. 5, 7, 11" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 293, + 294, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 293, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 62, + 293, + 294, + 335 + ], + "type": "text", + "content": "[4] Guangyi Chen, Weiran Yao, Xiangchen Song, Xinyue Li, Yongming Rao, and Kun Zhang. PLOT: prompt learning with optimal transport for vision-language models. In ICLR, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 336, + 295, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 336, + 295, + 369 + ], + "spans": [ + { + "bbox": [ + 62, + 336, + 295, + 369 + ], + "type": "text", + "content": "[5] Mengyuan Chen, Junyu Gao, and Changsheng Xu. Conjugated semantic pool improves OOD detection with pretrained vision-language models. In NeurIPS, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 370, + 295, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 370, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 62, + 370, + 295, + 403 + ], + "type": "text", + "content": "[6] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In CVPR, pages 3606-3613, 2014. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 403, + 295, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 295, + 436 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 295, + 436 + ], + "type": "text", + "content": "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255, 2009. 5, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 437, + 295, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 437, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 62, + 437, + 295, + 502 + ], + "type": "text", + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 502, + 295, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 295, + 535 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 295, + 535 + ], + "type": "text", + "content": "[9] Xuefeng Du, Zhaoning Wang, Mu Cai, and Yixuan Li. VOS: learning what you don't know by virtual outlier synthesis. In ICLR, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 536, + 295, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 536, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 57, + 536, + 295, + 568 + ], + "type": "text", + "content": "[10] Xuefeng Du, Yiyou Sun, Jerry Zhu, and Yixuan Li. Dream the impossible: Outlier imagination with diffusion models. NeurIPS, 36:60878-60901, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 570, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 570, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 57, + 570, + 295, + 612 + ], + "type": "text", + "content": "[11] Sepideh Esmaeilpour, Bing Liu, Eric Robertson, and Lei Shu. Zero-shot out-of-distribution detection based on the pre-trained model clip. In AAAI, pages 6568-6576, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 614, + 295, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 295, + 658 + ], + "type": "text", + "content": "[12] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In CVPRW, page 178, 2004. 5, 7, 11" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "text", + "content": "[13] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 5, 7" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[14] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In ICLR, 2017. 1, 2, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 107, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 140 + ], + "type": "text", + "content": "[15] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In CVPR, pages 15262-15271, 2021. 5, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 141, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 194 + ], + "type": "text", + "content": "[16] Dan Hendrycks, Steven Basart, Mantas Mazeika, Andy Zou, Joseph Kwon, Mohammadreza Mostajabi, Jacob Steinhardt, and Dawn Song. Scaling out-of-distribution detection for real-world settings. In ICML, pages 8759-8773, 2022. 1, 2, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 197, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 251 + ], + "type": "text", + "content": "[17] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alexander Shepard, Hartwig Adam, Pietro Perona, and Serge J. Belongie. The inaturalist species classification and detection dataset. In CVPR, pages 8769-8778, 2018. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "type": "text", + "content": "[18] Yen-Chang Hsu, Yilin Shen, Hongxia Jin, and Zsolt Kira. Generalized ODIN: detecting out-of-distribution image without learning from out-of-distribution data. In CVPR, pages 10948-10957, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 298, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 553, + 331 + ], + "type": "text", + "content": "[19] Rui Huang and Yixuan Li. MOS: towards scaling out-of-distribution detection for large semantic space. In CVPR, pages 8710-8719, 2021. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 332, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 374 + ], + "type": "text", + "content": "[20] Xue Jiang, Feng Liu, Zhen Fang, Hong Chen, Tongliang Liu, Feng Zheng, and Bo Han. Negative label guided OOD detection with pretrained vision-language models. In ICLR, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 376, + 553, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 410 + ], + "type": "text", + "content": "[21] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCVW, pages 554–561, 2013. 5, 7, 11" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 411, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 453 + ], + "type": "text", + "content": "[22] Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. In NeurIPS, pages 7167-7177, 2018. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 456, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 553, + 498 + ], + "type": "text", + "content": "[23] Tianqi Li, Guansong Pang, Xiao Bai, Wenjun Miao, and Jin Zheng. Learning transferable negative prompts for out-of-distribution detection. In CVPR, pages 17584-17594, 2024. 2, 3, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 500, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 553, + 533 + ], + "type": "text", + "content": "[24] Shiyu Liang, Yixuan Li, and R. Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. In ICLR, 2018. 1, 2, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 535, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 535, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 535, + 553, + 578 + ], + "type": "text", + "content": "[25] Jiashuo Liu, Zheyan Shen, Yue He, Xingxuan Zhang, Renzhe Xu, Han Yu, and Peng Cui. Towards out-of-distribution generalization: A survey. arXiv preprint arXiv:2108.13624, 2021. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 579, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 553, + 633 + ], + "type": "text", + "content": "[26] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM computing surveys, 55(9): 1-35, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 635, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 668 + ], + "type": "text", + "content": "[27] Weitang Liu, Xiaoyun Wang, John D. Owens, and Yixuan Li. Energy-based out-of-distribution detection. In NeurIPS, 2020. 1, 2, 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[28] Shuo Lu, Yingsheng Wang, Lijun Sheng, Aihua Zheng, Lingxiao He, and Jian Liang. Recent advances in ood detection: Problems and approaches. arXiv preprint arXiv:2409.11884, 2024. 1" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1160" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[29] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 5, 7, 11" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 163 + ], + "type": "text", + "content": "[30] Yifei Ming, Ziyang Cai, Jiuming Gu, Yiyou Sun, Wei Li, and Yixuan Li. Delving into out-of-distribution detection with vision-language representations. In NeurIPS, 2022. 1, 3, 5, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 164, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 294, + 196 + ], + "type": "text", + "content": "[31] Yifei Ming, Yiyou Sun, Ousmane Dia, and Yixuan Li. How to exploit hyperspherical embeddings for out-of-distribution detection? In ICLR, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 243 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 243 + ], + "type": "text", + "content": "[32] Atsuyuki Miyai, Qing Yu, Go Irie, and Kiyoharu Aizawa. Zero-shot in-distribution detection in multi-object settings using vision-language foundation models. arXiv preprint arXiv:2304.04521, 2023. 1, 3, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 244, + 294, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 244, + 294, + 278 + ], + "spans": [ + { + "bbox": [ + 56, + 244, + 294, + 278 + ], + "type": "text", + "content": "[33] Atsuyuki Miyai, Qing Yu, Go Irie, and Kiyoharu Aizawa. Locoop: Few-shot out-of-distribution detection via prompt learning. In NeurIPS, 2023. 2, 3, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 279, + 294, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 279, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 56, + 279, + 294, + 312 + ], + "type": "text", + "content": "[34] Jun Nie, Yonggang Zhang, Zhen Fang, Tongliang Liu, Bo Han, and Xinmei Tian. Out-of-distribution detection with negative prompts. In ICLR, 2024. 2, 3, 5, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 314, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 314, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 314, + 294, + 358 + ], + "type": "text", + "content": "[35] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729, 2008. 5, 7, 11" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 360, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 294, + 392 + ], + "type": "text", + "content": "[36] Omkar M. Parkhi, Andreae Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In CVPR, pages 3498-3505, 2012. 5, 7, 11" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 394, + 294, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 394, + 294, + 460 + ], + "spans": [ + { + "bbox": [ + 56, + 394, + 294, + 460 + ], + "type": "text", + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 461, + 294, + 516 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 461, + 294, + 516 + ], + "spans": [ + { + "bbox": [ + 56, + 461, + 294, + 516 + ], + "type": "text", + "content": "[38] Mohammadreza Salehi, Hossein Mirzaei, Dan Hendrycks, Yixuan Li, Mohammad Hossein Rohban, and Mohammad Sabokrou. A unified survey on anomaly, novelty, open-set, and out-of-distribution detection: Solutions and future challenges. arXiv preprint arXiv:2110.14051, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 518, + 294, + 550 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 518, + 294, + 550 + ], + "spans": [ + { + "bbox": [ + 56, + 518, + 294, + 550 + ], + "type": "text", + "content": "[39] Vikash Sehwag, Mung Chiang, and Prateek Mittal. SSD: A unified framework for self-supervised outlier detection. In ICLR, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 552, + 294, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 552, + 294, + 585 + ], + "spans": [ + { + "bbox": [ + 56, + 552, + 294, + 585 + ], + "type": "text", + "content": "[40] K Soomro. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012.5,7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 587, + 294, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 587, + 294, + 620 + ], + "spans": [ + { + "bbox": [ + 56, + 587, + 294, + 620 + ], + "type": "text", + "content": "[41] Yiyou Sun, Chuan Guo, and Yixuan Li. React: Out-of-distribution detection with rectified activations. In NeurIPS, pages 144-157, 2021. 1, 2, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 622, + 294, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 622, + 294, + 655 + ], + "spans": [ + { + "bbox": [ + 56, + 622, + 294, + 655 + ], + "type": "text", + "content": "[42] Yiyou Sun, Yifei Ming, Xiaojin Zhu, and Yixuan Li. Out-of-distribution detection with deep nearest neighbors. In ICML, pages 20827-20840, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 657, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 657, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 657, + 294, + 678 + ], + "type": "text", + "content": "[43] Leitian Tao, Xuefeng Du, Jerry Zhu, and Yixuan Li. Nonparametric outlier synthesis. In ICLR, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "text", + "content": "[44] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Open-set recognition: A good closed-set classifier is all you need. In ICLR, 2022. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 487 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 106 + ], + "type": "text", + "content": "[45] Haoqi Wang, Zhizhong Li, Litong Feng, and Wayne Zhang. Vim: Out-of-distribution with virtual-logit matching. In CVPR, pages 4911-4920, 2022. 5, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 107, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 107, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 315, + 107, + 553, + 140 + ], + "type": "text", + "content": "[46] Hualiang Wang, Yi Li, Huifeng Yao, and Xiaomeng Li. CLIPN for zero-shot OOD detection: Teaching CLIP to say no. In ICCV, pages 1802-1812, 2023. 2, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 141, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 141, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 315, + 141, + 553, + 184 + ], + "type": "text", + "content": "[47] Jianxiong Xiao, James Hays, Krista A. Ehinger, Aude Oliva, and Antonio Torralba. SUN database: Large-scale scene recognition from abbey to zoo. In CVPR, pages 3485-3492, 2010. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 186, + 553, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 186, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 315, + 186, + 553, + 218 + ], + "type": "text", + "content": "[48] Geng Yu, Jianing Zhu, Jiangchao Yao, and Bo Han. Self-calibrated tuning of vision-language models for out-of-distribution detection. In NeurIPS, 2024. 2, 3, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 220, + 553, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 220, + 553, + 253 + ], + "spans": [ + { + "bbox": [ + 315, + 220, + 553, + 253 + ], + "type": "text", + "content": "[49] Tao Yu, Zhihe Lu, Xin Jin, Zhibo Chen, and Xinchao Wang. Task residual for tuning vision-language models. In CVPR, pages 10899-10909, 2023. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 254, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 254, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 315, + 254, + 553, + 297 + ], + "type": "text", + "content": "[50] Yeonguk Yu, Sungho Shin, Seongju Lee, Changyun Jun, and Kyoobin Lee. Block selection method for using feature norm in out-of-distribution detection. In CVPR, pages 15701-15711, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 298, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 298, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 315, + 298, + 553, + 342 + ], + "type": "text", + "content": "[51] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kun-chang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of CLIP for few-shot classification. In ECCV, pages 493-510, 2022. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 343, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 343, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 315, + 343, + 553, + 387 + ], + "type": "text", + "content": "[52] Bolei Zhou, Ágata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE TPAMI, 40(6):1452-1464, 2018. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 388, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 388, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 315, + 388, + 553, + 421 + ], + "type": "text", + "content": "[53] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In CVPR, pages 16795-16804, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 422, + 553, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 422, + 553, + 455 + ], + "spans": [ + { + "bbox": [ + 315, + 422, + 553, + 455 + ], + "type": "text", + "content": "[54] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 130(9):2337-2348, 2022. 2, 3, 4, 5, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 456, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 456, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 315, + 456, + 553, + 487 + ], + "type": "text", + "content": "[55] Yibo Zhou. Rethinking reconstruction autoencoder-based out-of-distribution detection. In CVPR, pages 7379-7387, 2022. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1161" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/FB-Diff_ Fourier Basis-guided Diffusion for Temporal Interpolation of 4D Medical Imaging/c8ad7765-483e-400a-90bd-72090b0597d0_content_list.json b/2025/FB-Diff_ Fourier Basis-guided Diffusion for Temporal Interpolation of 4D Medical Imaging/c8ad7765-483e-400a-90bd-72090b0597d0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6324301be6b59c6fdbf2edb3605694cbe61f7a19 --- /dev/null +++ b/2025/FB-Diff_ Fourier Basis-guided Diffusion for Temporal Interpolation of 4D Medical Imaging/c8ad7765-483e-400a-90bd-72090b0597d0_content_list.json @@ -0,0 +1,1354 @@ +[ + { + "type": "text", + "text": "FB-Diff: Fourier Basis-guided Diffusion for Temporal Interpolation of 4D Medical Imaging", + "text_level": 1, + "bbox": [ + 129, + 128, + 869, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xin You $^{1,2,3}$ Runze Yang $^{1}$ Chuyan Zhang $^{1}$ Zhongliang Jiang $^{2}$ Jie Yang $^{1,*}$ Nassir Navab $^{2,*}$", + "bbox": [ + 99, + 200, + 898, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Institute of Medical Robotics, Shanghai Jiao Tong University, Shanghai, China", + "bbox": [ + 184, + 238, + 818, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Computer Aided Medical Procedures, Technical University of Munich, Munich, Germany", + "bbox": [ + 142, + 257, + 859, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Munich Center for Machine Learning, Munich, Germany", + "bbox": [ + 271, + 275, + 730, + 294 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "xin.you@tum.de, nassir.navab@tum.de", + "bbox": [ + 361, + 296, + 635, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 345, + 326, + 361 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The temporal interpolation task for 4D medical imaging, plays a crucial role in clinical practice of respiratory motion modeling. Following the simplified linear-motion hypothesis, existing approaches adopt optical flow-based models to interpolate intermediate frames. However, realistic respiratory motions should be nonlinear and quasi-periodic with specific frequencies. Intuited by this property, we resolve the temporal interpolation task from the frequency perspective, and propose a Fourier Basis-guided Diffusion model, termed FB-Diff. Specifically, due to the regular motion discipline of respiration, physiological motion priors are introduced to describe general characteristics of temporal data distributions. Then a Fourier motion operator is elaborately devised to extract Fourier bases by incorporating physiological motion priors and case-specific spectral information in the feature space of Variational Autoencoder. Well-learned Fourier bases can better simulate respiratory motions with motion patterns of specific frequencies. Conditioned on starting and ending frames, the diffusion model further leverages well-learned Fourier bases via the basis interaction operator, which promotes the temporal interpolation task in a generative manner. Extensive results demonstrate that FB-Diff achieves state-of-the-art (SOTA) perceptual performance with better temporal consistency while maintaining promising reconstruction metrics. Codes are available here.", + "bbox": [ + 91, + 378, + 483, + 770 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 799, + 220, + 814 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4D medical images, which depict 3D volumes with temporal variations, are essential in clinical practice for capturing dynamic changes and monitoring the progression of diseases over time [12, 25, 49, 52]. A significant application is the temporal modeling for breathing-induced motions of", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cardiac or pulmonary anatomical structures. That task aims to enrich motion visualizations by generating intermediate frames under the condition of starting and ending frames, which is consistent with the concept of Video Frame Interpolation (VFI) in natural scenarios.", + "bbox": [ + 511, + 347, + 906, + 422 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "VFI has been thoroughly investigated through systematic studies in recent years [10, 24, 31, 37, 39, 41, 58, 61, 62]. However, due to unique properties and constraints of medical imaging, it remains a challenge to directly implement these frameworks into 4D medical video interpolation with maintaining equally promising performance. Specifically, the quality of ground truth intermediate frames in medical imaging is often compromised due to factors such as imaging noise, unstable breathing, causing unexpected artifacts [17, 25, 43, 49]. Thus, the lower image quality compared to videos in the natural domain, will degrade the interpolation performance of VFI models. Besides, potential patients' movement might induce an additional disturbance to the temporal modeling [5, 38].", + "bbox": [ + 511, + 426, + 908, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Furthermore, there exists a discrepancy on the type of motions between natural and medical scenes. The former focuses on the object movement and illumination intensity change. While in the medical domain, motions can be depicted as subtle anatomical variance with quasi-periodic motion discipline [51, 56]. Thus, a specific motion prior of anatomies is effective to boost the temporal interpolation, especially for the medical data with limited resources from a small pool of individuals.", + "bbox": [ + 511, + 641, + 908, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "On account of these challenges, we present the following question: \"How can we devise a specialist VFI model tailored for 4D temporal medical volumes?\" Existing methods [2, 13, 29, 30, 51] on medical VFI face a major shortcoming: the linear motion hypothesis in a breathing period is required for the regression of intermediate frames as revealed in Fig.1(a). However, respiratory motions should be a nonlinear oscillatory process that cannot be simply mod-", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28010", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/49b55a0a073f7468310942ce3ca259cec8925fa39bd51feff78d4aec58660ba8.jpg", + "image_caption": [ + "(b) Nonlinear & quasi-periodic motion $\\Longrightarrow$ Our model", + "Figure 1. Temporal Interpolation for respiratory motions. (a) Simplified linear-motion hypothesis: interpolating intermediate frames by regressing optical flows with linear variations. (b) Nonlinear & quasi-periodic motions: leveraging diffusion models conditioned on Fourier motion bases, which represent various motion patterns indicating nonlinear and locally inconsistent anatomical variations." + ], + "image_footnote": [], + "bbox": [ + 140, + 88, + 861, + 358 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "eled with a linear equation [32, 51]. Also, these methods fail to recognize that respiratory motions in heart or lung, are indeed a quasi-periodic motion process with specific frequencies by extending the timeline. Intuited by this, we resolve the respiratory motion simulation from a frequency perspective. In Fig.1(b), each medical volume in the temporal domain can be viewed as a specific value in the periodic curve. That periodic curve can be decomposed as a group of sine and cosine bases with multiple frequencies [46]. Comparatively, there exists a group of Fourier bases, which are capable of reconstructing original temporal sequences. Here Fourier bases with different frequencies refer to various motion patterns, which indicate nonlinear and locally anisotropic variations of anatomical structures. In our work, Fourier bases are acquired by the Fast Fourier Transform (FFT) on the temporal domain.", + "bbox": [ + 88, + 458, + 482, + 699 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, 4D data distributions of the respiratory process depict a regular motion discipline, with temporal variations from inhaling to exhaling, or from diastole to systole. Motivated by that observation, physiology motion priors are introduced to encode generic characteristics of the respiratory process. Then a Fourier Motion Operator $O_{FM}$ is elaborately devised to yield enhanced motion bases by incorporating physiology motion priors and case-specific spectral information in the feature space. Specifically, A Variational Autoencoder (VAE) is introduced to map the video space into the feature space. Physiology motion priors are defined as learnable frequency embeddings, which can be updated by pretraining a Variational Autoen", + "bbox": [ + 89, + 704, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "coder (VAE) on given temporal data. And case-specific frequency information can be extracted by FFT in the temporal domain. Well-learned physiology motion priors can enrich Fourier bases with generic properties of respiratory motions. Then enhanced Fourier motion bases can better simulate the quasi-periodic respiratory process with motion patterns of specific frequencies. On the grounds that diffusion models show promising performance in generating realistic videos [15, 16, 35, 48], a Markov chain process is learned to transform Gaussian noise into target videos by the proposed Fourier Bases-guided Diffusion, termed FB-Diff. Since only starting and ending frames are available during inference, they serve as conditional prompts for intermediate frame interpolation. Meanwhile, we devise the basis interaction operator to inject enhanced Fourier bases into the conditional diffusion model. Then, FB-Diff can well address the temporal interpolation task in a generative manner.", + "bbox": [ + 511, + 458, + 906, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our proposed method has achieved state-of-the-art (SOTA) interpolation performance compared with other methods for 4D medical imaging, particularly on the mid-most frames. For quantitative results, more comprehensive evaluation metrics are adopted in this work, including traditional reconstruction metrics and feature-level perceptual metrics. FB-Diff not only presents promising reconstruction metrics but also superior perceptual results. Qualitative visualizations reveal better temporal consistency by FB-Diff compared with other linear-motion-based methods. Our main contributions are summarized as follows:", + "bbox": [ + 511, + 734, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28011", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ea0242ad2ded6b4b4995d4b5240260ecbefd64127560851e8856bf2d9f92bc00.jpg", + "image_caption": [ + "Figure 2. The pipeline of FB-Diff. (a) The Fourier motion operator is proposed to extract Fourier bases in a feature space of VAE. Bases consist of learnable generic motion priors and case-specific frequency information. (b) Fourier bases are first adapted to the masked video domain under the guidance of well-learned motion priors and pretrained VAE weights. Domain-adapted bases serve as the conditional input to boost the video synthesis by diffusion models." + ], + "image_footnote": [], + "bbox": [ + 127, + 88, + 869, + 496 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a diffusion model termed FB-Diff for the temporal interpolation of 4D medical imaging. FB-Diff leverages Fourier motion bases, which indicate motion patterns with specific frequencies to resolve the respiratory modeling with nonlinear and quasi-periodic motions.", + "- We introduce the Fourier motion operator $O_{FM}$ to extract Fourier bases, containing physiology motion priors and case-specific frequency representations.", + "- We devise the basis interaction operator to inject Fourier bases into the diffusion model. Fourier bases can boost the video synthesis for temporal interpolation.", + "- Extensive results demonstrate that FB-Diff reveals SOTA perceptual performance with better temporal consistency while maintaining promising reconstruction metrics." + ], + "bbox": [ + 89, + 589, + 482, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 814, + 232, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video Frame Interpolation. Popular VFI techniques are mainly classified into motion-free and motion-based approaches, depending on whether optical flows are incorporated or not. Motion-based: These approaches warp", + "bbox": [ + 89, + 839, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "input frames forward or backward based on optical flows acquired by either off-the-shelf networks [11, 23, 44, 45, 50] or custom-designed flow estimators [20, 21, 27, 31, 33, 58]. Warped intermediate frames are then refined by delicately designed networks [3, 26] to enhance visual quality. Recently, Transformer-based motion estimators are proposed to raise the upper limit of VFI performance [36, 41, 58] Motion-Free: Motion-free methods rely on implicit spatiotemporal modeling [8, 9, 19, 28] to generate the intermediate frame. However, due to a limited quality of 4D medical videos and quasi-periodic motion priors, existing VFI frameworks are not feasible. That is why we aim to propose a specialist VFI model for 4D medical sequences.", + "bbox": [ + 511, + 589, + 906, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4D Medical VFI. Existing methods on this topic mainly leverage volumetric motions for 4D temporal interpolation. Specifically, SVIN [13] estimates forward and backward deformation fields, which will yield precise intermediate frames by linearly combining bidirectional information. MPVF [51] resolves various magnitudes of motions by proposing a multi-pyramid voxel flows model that takes", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28012", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "multi-scale voxel flows into account. DDM [29] can learn spatial deformation information between the source and target volumes and provide a latent code for generating intermediate frames along a geodesic path. UVI-Net [30] utilizes the flow calculation model with the time-domain cycle-consistency constraint and linear motion hypothesis, to realize motion modeling in an unsupervised style. However, the linear hypothesis for motions will induce temporal inconsistency and spatial distortion.", + "bbox": [ + 89, + 90, + 483, + 226 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diffusion-based Video Interpolation. Denoising diffusion probabilistic models [15, 16, 35] reveal promising performance in generating realistic images or videos. Some methods maximize the potential of diffusion models to resolve video interpolation tasks. Essentially, these approaches implicitly learn the temporal motions from the starting frame to the ending frame [1, 7, 24]. LDMVFI [10] proposes a latent diffusion model, which approaches the VFI problem from a generative perspective by formulating it as a conditional synthesis issue. However, these diffusion models do not leverage effective conditional guidance, which is crucial for medical VFI with intrinsic motion priors. Thus, we introduce Fourier bases to enhance the conditional synthesis.", + "bbox": [ + 89, + 229, + 483, + 428 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 439, + 225, + 457 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Fourier Motion Bases", + "text_level": 1, + "bbox": [ + 89, + 464, + 294, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since respiratory motions for heart or lung exhibit a quasi-periodic characteristic, we intend to resolve this task from a frequency perspective. Specifically, the respiration can be viewed as an oscillatory process with certain frequencies. Thus, we draw an analogy between 2D time-varying curves and 4D temporal medical volumes. Curves can be decomposed into a group of harmonic oscillators with various frequencies [46]. Similarly, temporal volumes can be restored via a stack of extracted Fourier bases, which are embodied with various motion patterns of anatomical deformation. The whole process can be formulated as:", + "bbox": [ + 89, + 486, + 483, + 652 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {B} _ {k} = \\sum_ {n = 0} ^ {N - 1} V _ {n} e ^ {- i 2 \\pi k \\frac {n}{N}}, k = 0, 1, \\dots , N - 1, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 652, + 482, + 693 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV _ {n} = \\frac {1}{T} \\sum_ {k = 0} ^ {N - 1} \\mathcal {B} _ {k} e ^ {i 2 \\pi n \\frac {k}{N}}, n = 0, 1, \\dots , N - 1, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 696, + 482, + 736 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{B}_k$ means the $k_{th}$ Fourier basis in the frequency domain, $V_{n}$ refers to the $n_{th}$ medical volume, and $N$ is the frame number. These Fourier bases describe locally anisotropic anatomical variations with different motion intensities, promoting the performance of the temporal interpolation model $\\mathcal{I}$ . Thus, the goal of our work is to deduce 4D medical videos $V$ conditioned on two prompting frames $V_{0}, V_{N-1}$ and Fourier motion bases $\\mathcal{B}$ :", + "bbox": [ + 89, + 746, + 483, + 867 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {I} ^ {*} = \\underset {\\mathcal {I}} {\\arg \\min } \\| \\boldsymbol {V} - \\mathcal {I} \\left(V _ {0}, V _ {N - 1}; \\mathcal {B}\\right) \\| _ {\\mathcal {D}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 878, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3a5fb39ec16e2282dd96e4173cb949fd02dec76d57657897bdfbad4585e8955a.jpg", + "image_caption": [ + "Figure 3. The details of Fourier motion operator. Fourier bases are decomposed as physiology motion priors and case-specific Fourier features, corresponding to phase and magnitude components." + ], + "image_footnote": [], + "bbox": [ + 527, + 88, + 893, + 263 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathcal{I}^*$ refers to the optimal solution in the interpolation model set, and $\\mathcal{D}$ represents the distance-based normed space. We systematically organize the following sections along Fourier bases. Section 3.2 introduces the design of the Fourier Motion Operator, which constructs Fourier bases with physiology motion priors and case-specific frequency information. Section 3.3 adapts Fourier bases to the to-be-interpolated video domain with well-learned physiology motion priors. Section 3.4 provides details of the diffusion model conditioned on Fourier bases.", + "bbox": [ + 511, + 325, + 906, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Fourier Motion Operator", + "text_level": 1, + "bbox": [ + 513, + 486, + 746, + 502 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To explicitly extract the Fourier bases with various motion patterns for the respiration process, we introduce the Fourier Operator $O_{\\mathcal{F}}$ as a neural operator. As stated in Eq.(2), $O_{\\mathcal{F}}$ will operate in the temporal domain, not the spatial domain. Since the common Fourier operator $O_{\\mathcal{F}}$ only mines case-specific frequency representations, general physiology characteristics inside respiratory data are ignored. Inspired by that notion, we devise the Fourier motion operator $O_{FM}$ as a neural operator in the feature space. To map the image space into feature space, we introduce a Variational Autoencoder (VAE) as revealed in Fig.2(a). Due to no built-in 4D convolutional operator, we replace it with 3D spatial convolution and 1D temporal convolution as [34, 59] did. The spatial layer in Fig.2(a) consists of spatial convolution, GELU, and group normalization operations.", + "bbox": [ + 511, + 508, + 906, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Fourier motion operator $O_{FM}$ is devised to acquire spectral characteristics with both physiology motion priors and case-specific representations in the feature space. Firstly, physiology motion priors are introduced as learnable frequency embeddings, which are updated by backward gradients of restoring given temporal data via VAE. Well-trained frequency embeddings $\\theta$ can represent the generic discipline of the regular respiration process [54], with $\\theta_{k}$ referring to the $k_{th}$ element $(k = 0,1,\\dots,N - 1)$ . Then Fourier bases $\\mathcal{B}$ are made up of physiology motion priors $\\theta$ and case-specific features $F$ , which arise from the output of", + "bbox": [ + 511, + 734, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28013", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "spatial and temporal layers. As depicted by Fig.3, $\\mathcal{B}$ with a complex type is yielded in the following style:", + "bbox": [ + 89, + 90, + 485, + 122 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {B} = \\underbrace {\\mathcal {M} (\\left\\| O _ {\\mathcal {F}} (F) \\right\\| _ {2})} _ {\\text {C a s e - S p e c i f i c}} \\cdot \\underbrace {e ^ {i U (\\tau (\\theta))}} _ {\\text {G e n e r a l}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 130, + 482, + 172 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\tau$ means the channel transform, $U$ refers to the upsampling which unifies the resolution of $\\theta$ and $\\pmb{F}$ . These operations transform general motion priors into the phase part of Fourier bases, which determines the motion variation standing for the whole dataset distribution. Then $\\| \\cdot \\| _2$ means the $L_{2}$ norm, $\\mathcal{M}$ is the convolutional mapping operation. They will output motion bases' magnitude information, which selectively highlights regions in spectral maps relevant to case-specific information.", + "bbox": [ + 89, + 181, + 483, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Meanwhile, feature maps $F$ are further enriched with more fine-grained frequency details. The specific mechanism can be described with Eq.(5):", + "bbox": [ + 89, + 318, + 483, + 364 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {F}} = \\boldsymbol {F} + \\sigma \\mathcal {O} _ {\\mathcal {F}} ^ {- 1} (\\mathcal {B}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 373, + 482, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "here $O_{\\mathcal{F}}^{-1}$ and $\\sigma$ refer to the inverse Fourier transform and Sigmoid operators, which will boost deep features $\\hat{F}$ with enriched details, including shape and texture information.", + "bbox": [ + 89, + 402, + 483, + 452 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Frame Inpainting-based Basis Synthesis", + "text_level": 1, + "bbox": [ + 89, + 462, + 434, + 478 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Due to the fact that only starting and ending frames can be acquired during inference, a significant step is to adapt Fourier bases to the masked video domain. Motivated by the fact that strong correlations exist between Fourier motion bases and inpainted temporal sequences, we plan to jointly optimize these two variables. To accomplish that goal, we repurpose well-calibrated physiology motion priors from VAE, which prove to be effective in providing the motion guidance for intermediate frame inpainting. Also, pretrained weights of VAE encode implicit temporal dependencies between frames, enhancing the reconstruction of masked video sequences via the finetuned VAE noted as $\\mathcal{G}$ . A mathematical formulation is defined as:", + "bbox": [ + 89, + 484, + 483, + 679 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {V}, \\hat {\\mathcal {B}} = \\mathcal {G} (\\hat {\\boldsymbol {\\theta}}, V _ {\\boldsymbol {M}}) = \\mathcal {G} (\\hat {\\boldsymbol {\\theta}}, V _ {0},.., \\mathbf {0},.., V _ {N - 1}), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 689, + 482, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\hat{\\theta}$ and $V_{M}$ correspond to well-trained general motions and zero-masked videos, $\\hat{V}$ and $\\hat{B}$ refer to inpainted videos and domain-adapted Fourier bases.", + "bbox": [ + 89, + 718, + 482, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The interaction between temporal video representations and $\\hat{\\mathcal{B}}$ is implemented via the Fourier motion operator as shown in Fig.2(b). For the joint optimization process, video sequences will contribute temporal variations, enhancing the synthesis of adapted Fourier bases. Conversely, Fourier bases for zero-masked inputs inherently generate a spectral map that emphasizes regions with motion in the temporal axis. This activated map aids in temporal frame inpainting by exerting attention on regions with different magnitudes.", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e46e1aa2636d3a69e2489f0e0819f6dad5bd9bb3792fe805fd5471e932837caa.jpg", + "image_caption": [ + "Figure 4. The architecture of basis interaction operator. Fourier motion bases selectively enhance video representations by measuring the motion intensity of different regions. Then those enhanced temporal representations can enrich denoised embeddings from the diffusion model." + ], + "image_footnote": [], + "bbox": [ + 535, + 90, + 885, + 231 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Basis-guided 4D Conditional Diffusion", + "text_level": 1, + "bbox": [ + 511, + 320, + 844, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "VAE in Fig.2(a) is not only devised for universal motion priors, but also for the latent diffusion model (LDM), which shows better compatibility for extra conditional inputs. Besides, LDM synthesizes latent embeddings occupying less GPU memory compared to 4D temporal volumes. And the VAE decoder is leveraged to reconstruct interpolated video sequences with synthesized latent embeddings.", + "bbox": [ + 511, + 340, + 905, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the specific diffusion process, the whole video sequence $V$ is first encoded as a latent embedding $z_0 = \\mathcal{G}_e(V)$ . Then the latent code $z_0$ is perturbed as:", + "bbox": [ + 511, + 448, + 905, + 494 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} z _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , \\epsilon \\sim \\mathcal {N} (0, 1), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 498, + 903, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\overline{\\alpha}_t = \\prod_{i=1}^t (1 - \\beta_t)$ with $\\beta_t$ is the noise coefficient at time step $t$ , and $t$ is uniformly sampled from the timestep index set $[1,2,\\dots,T]$ . This process can be regarded as a Markov chain, which incrementally adds Gaussian noise to the latent code $z_0$ . The denoising model $d$ receives $z_t$ as input and is optimized under conditional guidance to learn the latent space distribution with the objective function:", + "bbox": [ + 511, + 522, + 905, + 628 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\epsilon} = E _ {z _ {t}, \\epsilon \\sim \\mathcal {N} (0, 1)} \\| \\epsilon - \\epsilon_ {d} \\left(z _ {t}, t; \\hat {\\mathcal {B}}, \\mathcal {G} _ {e} \\left(\\boldsymbol {V} _ {\\boldsymbol {M}}\\right), \\hat {\\boldsymbol {V}}\\right) \\| ^ {2}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 633, + 903, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "here $\\epsilon_d$ is the regressed noise, adapted Fourier bases $\\hat{\\mathcal{B}}$ and inpainted coarse videos $\\hat{V}$ are incorporated as the conditional guidance. Moreover, feature maps of starting and ending frames by the VAE encoder $\\mathcal{G}_e$ serve as a strong prompt for the 4D video synthesis. Then during the training process of FB-Diff, the total loss $\\mathcal{L}$ consists of the denoising loss $\\mathcal{L}_{\\epsilon}$ and the $L_{2}$ -norm reconstruction loss $\\mathcal{L}_r(\\hat{V},V)$ .", + "bbox": [ + 511, + 657, + 905, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To inject Fourier motion bases into LDM and harness coarsely inpainted results, we design the Basis Interaction Operator (BIO) as shown in Fig.4. Specifically, inpainted videos are mapped into video feature representations via spatial & temporal layers and the downsampling operation. Fourier motion bases with various motion patterns, selectively enhance video representations by measuring the motion intensity of different regions. Those enhanced temporal representations can enrich denoised embeddings from", + "bbox": [ + 511, + 765, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28014", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/cad301eea28cf45b15abfeb328188c97382b503a0137fcc1c8a0ba6185357e99.jpg", + "table_caption": [ + "Table 1. Baseline comparison with flow-based and diffusion-based models. (Bold: the best, Underlined: the second best. Δ values represent the relative difference between our model and UVI-Net, and red and blue numbers stand for better and worse results)." + ], + "table_footnote": [], + "table_body": "
ModelACDC Cardiac4D Lung
PSNR (dB) ↑LPIPS ↓FID ↓FVD ↓PSNR (dB) ↑LPIPS ↓FID ↓FVD ↓
SVIN [13]31.43±0.4211.563±0.20626.393.630.49±0.3042.650±0.24538.5125.6
Voxelmorph [2]30.77±0.5021.969±0.19729.4102.129.90±0.3732.815±0.26043.2149.0
Transmorph [6]29.81±0.5322.136±0.25531.5109.529.04±0.4032.946±0.29140.5143.5
MPVF [51]31.53±0.4101.594±0.20124.891.330.76±0.2652.418±0.24639.4120.3
UVI-Net [30]32.16±0.4021.662±0.24523.094.231.57±0.3112.211±0.21637.8121.7
IFRNet [31]30.58±0.4081.752±0.24927.897.530.05±0.2802.316±0.27339.7124.0
LDMVFI [10]27.11±0.4602.943±0.41028.899.226.31±0.4533.659±0.34142.1142.7
LDDM [7]24.53±0.4812.634±0.30833.4105.725.19±0.2732.914±0.28748.7146.3
Conditional Diff [16]26.59±0.5452.460±0.35730.195.725.95±0.3913.294±0.37545.0133.0
DDM [29]29.79±0.5042.089±0.35227.4110.329.67±0.4202.705±0.33040.8165.5
FB-Diff30.95±0.3861.452±0.20521.588.930.18±0.2872.066±0.21537.7113.6
Δ value1.21 ↓0.210 ↓1.5 ↓5.3 ↓1.39 ↓0.145 ↓0.1 ↓8.1 ↓
", + "bbox": [ + 135, + 127, + 867, + 333 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the diffusion model, promoting a promising synthesis of highlighted regions. Thus, BIO can boost well-interpolated videos with spatial and temporal consistency.", + "bbox": [ + 89, + 369, + 485, + 417 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 429, + 225, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 89, + 454, + 299, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. To evaluate the interpolation performance of our framework, we conduct experiments on public ACDC cardiac [4] and 4D-Lung datasets [22]. For ACDC with the MRI modality, volume sequences between the end-diastolic and end-systolic phases are extracted as valid 4D data. Of all the 150 4D sequences, cases with identity 1-100, 101-120, and 121-150 serve as the training, validation, and testing sets. Besides, all MRI volumes are cropped to $128 \\times 128 \\times 32$ , and the frame number ranges from 6 to 16 [53, 57]. For 4D-Lung CT images, the end-inspiratory and end-expiratory scans are set as the initial and final images. We collect 125 4D videos from [22], with the resolution equal to $6 \\times 128 \\times 128 \\times 128$ . They are split as $80 / 15 / 30$ cases for training, validation, and inference. We uniform the frame number as 16 with zero paddings.", + "bbox": [ + 88, + 477, + 482, + 704 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "**Evaluations.** We select quantitative evaluation metrics including Peak Signal-to-Noise Ratio (PSNR) [18], Learned Perceptual Image Patch Similarity (LPIPS) [55, 60], Fréchet Inception Distance (FID) [14], and Fréchet Video Distance (FVD) [47]. Compared to PSNR, which is the traditional pixel-wise reconstruction metric, LPIPS and FVD evaluate perceptual quality from the perspective of human vision [10, 24, 61], which is more feasible and favored in clinical practice. Specifically, LPIPS reveals important perceptual similarity like human vision. FID and FVD evaluate the dataset distribution bias between realistic and synthesized data, including spatial and temporal consistency.", + "bbox": [ + 88, + 704, + 482, + 885 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Details. The VAE maps the image space into", + "bbox": [ + 89, + 885, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the downsampled latent space with a ratio of $1/8$ . And the proposed conditional diffusion model is implemented with a lightweight UNet. All models are trained using the AdamW optimizer with the linear warm-up strategy. The batch size, diffusion step $T$ , and training epoch are equal to 2, 1000, and 500. Experiments are implemented based on Pytorch and 2 NVIDIA RTX 4090 GPUs. More details about the network and training settings are described in the Appendix.", + "bbox": [ + 511, + 369, + 906, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparisons with SOTA baselines", + "text_level": 1, + "bbox": [ + 511, + 500, + 815, + 516 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Comparison. We comprehensively conduct experiments on ACDC cardiac and 4D Lung datasets. Quantitative results are illustrated in Tab.1. Our model achieves the best performance across the benchmark on perceptual evaluation metrics from the perspective of human vision [10, 61], which are clinically favored compared with the pixel-centric reconstruction metric [7]. Specifically, FB-Diff outperforms UVI-Net [30] with $0.210 \\downarrow$ LPIPS and $5.3 \\downarrow$ FVD values on ACDC. Also, our model maintains quite satisfactory PSNR values of $30.95\\mathrm{dB}$ and $30.18\\mathrm{dB}$ on ACDC and 4D Lung. Although diffusion models are not sensitive to pixel-wise intensities [15], the conditional guidance of Fourier bases and coarsely inpainted videos is effective to boost the interpolation performance of diffusion models. Tab.1 reveals that FB-Diff significantly surpasses existing diffusion models on all metrics.", + "bbox": [ + 511, + 522, + 906, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Furthermore, the interpolation process becomes most challenging for the midmost frames since they are positioned farthest from the two reference frames [24]. And our model can better resolve this issue due to stronger temporal representations based on quasi-periodic motion modeling. Thus, a quantitative evaluation of central frames is necessary. As shown in Tab.2, FB-Diff demonstrates greater performance advantages, with $4.3\\downarrow$ FID and $8.2\\downarrow$ FVD values compared with UVI-Net. Additionally, FB-Diff acquires a competi", + "bbox": [ + 511, + 765, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28015", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8da4c198ac582ec7f985250dd8520efe4944f7261aca5a0b3ecc632bc930d0c8.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison of the midmost frame interpolation between different models. VM: Voxelmorph." + ], + "image_footnote": [], + "bbox": [ + 137, + 89, + 867, + 349 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/da4ed1f522b999eaef26e04949fe27039f74ff39048bbe97f6a0fa1e5ba4dbe0.jpg", + "table_caption": [ + "Table 2. Interpolation performance of the midmost frames on ACDC cardiac. The choice of midmost frames: if the frame number $N$ is odd, three intermediate frames are selected, otherwise only two frames are chosen." + ], + "table_footnote": [], + "table_body": "
ModelPSNR (dB) ↑LPIPS ↓FID ↓FVD ↓
SVIN [13]26.01±0.4672.673±0.36138.9142.8
DDM [29]24.37±0.5112.756±0.33041.5145.1
UVI-Net [30]26.25±0.4102.542±0.37437.0137.1
FB-Diff26.18±0.4132.287±0.30532.7128.9
Δ value0.07 ↓0.255 ↓4.3 ↓8.2 ↓
", + "bbox": [ + 104, + 462, + 473, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tive PSNR value, which demonstrates the effectiveness of respiration-related modeling via Fourier bases.", + "bbox": [ + 89, + 564, + 483, + 594 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Comparison. Qualitative results on the interpolated midmost frame are depicted in Fig.5. SVIN [13], Voxelmorph [2], and DDM [29] leverage the single-direction flow estimation to interpolate intermediate frames with linear motions. Compared with that, UVI-Net [30] constructs the interpolation model with time-domain cycle-consistency, and achieves better reconstruction performance as expected. However, these methods essentially overlook the nonlinear and quasi-periodic motion discipline between temporal sequences, resulting in unsatisfactory interpolated results. Instead, FB-Diff exhibits advantageous predictions for midmost frames under the guidance of various motion patterns, indicating locally anisotropic motions with different intensities. Besides, other diffusion models do not work well due to the lack of fine-grained conditional guidance.", + "bbox": [ + 89, + 595, + 483, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 832, + 243, + 848 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Key Components. An ablation study is conducted to investigate the significance of various components in FB-Diff as shown in Tab.3. (a) Physiology Motion Priors $\\theta$ . By", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f006ddcb334f88c13fa38ae31482e173fdb0836eb48bf0dd0ce0e5815b0219d4.jpg", + "table_caption": [ + "Table 3. Ablation study on key components. $\\theta$ : physiology motion priors, $\\hat{B}$ : Fourier motion bases, $\\hat{V}$ : coarsely inpainted videos." + ], + "table_footnote": [], + "table_body": "
SettingsACDC Cardiac4D Lung
θˆBˆVPSNR (dB) ↑LPIPS ↓PSNR (dB) ↑LPIPS ↓
X✓✓29.761.71029.532.413
XX✓28.172.10528.402.971
✓✓X30.291.65529.712.246
✓✓✓30.951.45230.182.066
", + "bbox": [ + 527, + 434, + 895, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "removing $\\theta$ from the Fourier motion operator, interpolation performance decreases significantly due to the loss of general spectral characteristics from respiratory data. Therefore, Fourier bases lack the precise phase information. (b) Fourier bases $\\hat{\\mathcal{B}}$ . Without $\\hat{\\mathcal{B}}$ as conditional guidance, the conditional diffusion model fails to perceive explicit motion patterns with different frequencies, which provide the intensity of anisotropic motions. As a result, the generative model cannot well learn the regular temporal variations, with 2.78 dB $\\downarrow$ and 1.78 dB $\\downarrow$ PSNR values on ACDC and 4D Lung. (c) Inpainted Videos $\\hat{V}$ . Here the coarsely reconstructed sequences serve as the output by the finedtuned VAE as revealed in Fig.2. And the joint optimization process boosts the spatial and temporal representations of $\\hat{V}$ . Consequently, discarding this component will reduce fine-grained details of the diffusion model's output.", + "bbox": [ + 511, + 536, + 906, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Hyper-settings of Fourier Bases. (a) The number of Fourier Bases $N_{Fb}$ . We first analyze the quantitative effects of the frequency number on the interpolation performance. As revealed in Tab.4, reducing the frequency number will cause degraded performance of the temporal simulation. Specifically, 1.30 dB lower PSNR and 0.535 higher LPIPS values are induced by decreasing $N_{Fb}$ from 16 to 8. That phenomenon indicates that some elements in the group of", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28016", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c70cbbc9d0193a624785ed0adda338f619441ab48549cd94deb117f6ac5a06fa.jpg", + "image_caption": [ + "Figure 6. Visualizations on the frequency spectral of Fourier bases. $V_{0}$ , $V_{N - 1}$ : starting and ending frames. $f_{1} - f_{8}$ refer to different motion patterns, which highlight different regions under different frequency. There also appear locally anisotropic motion intensities inside a specific motion pattern." + ], + "image_footnote": [], + "bbox": [ + 112, + 88, + 885, + 178 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e1109913f67f4e7e5b943dde274bf8c0476fa60b1a7c67b26c2d1d3363d2b6ea.jpg", + "table_caption": [ + "Table 4. Ablation study on the hyper-settings of Fourier bases. $N_{Fb}$ : the frequency number of Fourier bases." + ], + "table_footnote": [], + "table_body": "
SettingsACDC Cardiac4D Lung
PSNR (dB) ↑LPIPS ↓PSNR (dB) ↑LPIPS ↓
\\(N_{Fb}\\)=16 (Ours)30.951.45230.182.066
\\(N_{Fb}\\)=1230.761.64029.972.570
\\(N_{Fb}\\)=829.651.98729.502.745
w/o Real part29.181.87229.562.453
w/o Imaginary part29.881.68129.672.401
", + "bbox": [ + 93, + 292, + 483, + 386 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "bases with the frequency ranging from $f_{8}$ to $f_{15}$ are consistent with specific frequencies of the respiratory process. (b) Real & Imaginary Parts. $\\hat{\\mathcal{B}}$ bears real and imaginary parts after combining generic priors and case-specific spectral representations, which perform as the phase and magnitude information. Eliminating either the real or imaginary part will convert the complex data type into the real one, resulting in the loss of phase information and influencing the temporal interpolation.", + "bbox": [ + 89, + 402, + 483, + 540 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Visualization Analysis", + "text_level": 1, + "bbox": [ + 89, + 546, + 299, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization of Fourier bases. To further validate the effectiveness and interpretability of our proposed Fourier bases, we visualize extracted motion bases from the most shadow layer. As revealed in Fig.6, $f_{0} - f_{7}$ refers to the magnitude of the first eight motion bases in ascending order of frequency. Specifically, $f_{0}$ shows a low-frequency motion pattern highlighting no region, which is strongly related to static backgrounds. $f_{1} - f_{3}$ exhibit a large motion intensity in the highlighted right and left ventricle regions. Those regions are indeed dynamic foregrounds requiring special attention. $f_{4}$ depicts an activation in the myocardial region, which contracts inward along the circumferential direction. $f_{5} - f_{7}$ highlight regions with vast deformations in the temporal domain. Particularly, $f_{6}$ illustrates a large motion on boundaries of anatomical structures. In summary, these motion patterns reveal locally anisotropic motion intensities, and are harmonized with specific frequencies of the respiratory process, thus promoting the temporal interpolation.", + "bbox": [ + 89, + 568, + 482, + 840 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization of temporal anatomical variations. To demonstrate the effectiveness of FB-Diff on temporal simulation, we visualize temporal variation maps for a perceptual comparison. Fig.7 provides interpolated intermediate", + "bbox": [ + 89, + 840, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bfd652572d8a4fb20e6f20ddca808de4f098967ae86b12ade2703c4b76e21236.jpg", + "image_caption": [ + "Figure 7. Temporal variation comparison between FB-Diff and existing methods with the linear motion hypothesis." + ], + "image_footnote": [], + "bbox": [ + 531, + 253, + 890, + 501 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "frames at different timepoints. According to the referenced video, the anatomical motion in the temporal domain is nonlinear, which cannot be precisely simulated by existing methods with the linear motion hypothesis. In contrast, our predictions show better temporal consistency and continuity, particularly for frames at $t = 0.6$ and 0.8.", + "bbox": [ + 511, + 547, + 906, + 640 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 660, + 633, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our proposed FB-Diff resolves the temporal interpolation of 4D medical imaging from a frequency perspective, and can model the respiration-related process with nonlinear and quasi-periodic motions. A Fourier motion operator is devised to extract Fourier bases by incorporating physiological motion priors and case-specific spectral information in the feature space. And the diffusion model further leverages Fourier bases and two prompting frames to achieve the precise interpolation. Extensive results show that FB-Diff achieves SOTA perceptual performance with better temporal consistency while maintaining promising reconstruction metrics. Ultimately, it requires more conditional guidance such as electrocardiogram signals to simulate the highly unstable breathing, which is promising in future work.", + "bbox": [ + 511, + 688, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28017", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. We thank the anonymous reviewers for their thorough and constructive feedbacks. This research was supported in part by the project of the National Natural Science Foundation of China (No. 62376153), in part by the institute of Munich Center of Machine Learning.", + "bbox": [ + 89, + 90, + 485, + 167 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 181, + 187, + 196 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Juyoung Bae, Elizabeth Tong, and Hao Chen. Conditional diffusion model for versatile temporal inpainting in 4d cerebral ct perfusion imaging. In MICCAI, pages 67-77. Springer, 2024. 4", + "[2] Guha Balakrishnan, Amy Zhao, Mert R Sabuncu, John Guttag, and Adrian V Dalca. Voxelmorph: a learning framework for deformable medical image registration. IEEE transactions on medical imaging, 38(8):1788-1800, 2019. 1, 6, 7", + "[3] Wenbo Bao, Wei-Sheng Lai, Xiaoyun Zhang, Zhiyong Gao, and Ming-Hsuan Yang. Memc-net: Motion estimation and motion compensation driven neural network for video interpolation and enhancement. IEEE transactions on pattern analysis and machine intelligence, 43(3):933-948, 2019. 3", + "[4] Olivier Bernard, Alain Lalande, Clement Zotti, Frederick Cervenansky, Xin Yang, Pheng-Ann Heng, Irem Cetin, Karim Lekadir, Oscar Camara, Miguel Angel Gonzalez Ballester, et al. Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: is the problem solved? IEEE transactions on medical imaging, 37 (11):2514–2525, 2018. 6", + "[5] Rhydian Caines, Naomi K Sisson, and Carl G Rowbottom. 4dct and vmat for lung patients with irregular breathing. Journal of Applied Clinical Medical Physics, 23(1):e13453, 2022. 1", + "[6] Junyu Chen, Eric C Frey, Yufan He, William P Segars, Ye Li, and Yong Du. Transmorph: Transformer for unsupervised medical image registration. Medical image analysis, 82:102615, 2022. 6", + "[7] Tingxiu Chen, Yilei Shi, Zixuan Zheng, Bingcong Yan, Jingliang Hu, Xiao Xiang Zhu, and Lichao Mou. Ultrasound image-to-video synthesis via latent dynamic diffusion models. In MICCAI, pages 764-774. Springer, 2024. 4, 6", + "[8] Myungsub Choi, Heewon Kim, Bohyung Han, Ning Xu, and Kyoung Mu Lee. Channel attention is all you need for video frame interpolation. In AAAI, pages 10663-10671, 2020. 3", + "[9] Myungsub Choi, Suyoung Lee, Heewon Kim, and Kyoung Mu Lee. Motion-aware dynamic architecture for efficient frame interpolation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13839-13848, 2021. 3", + "[10] Duolikun Danier, Fan Zhang, and David Bull. Ldmvfi: Video frame interpolation with latent diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1472-1480, 2024. 1, 4, 6", + "[11] Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick Van Der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In ICCV, pages 2758-2766, 2015. 3" + ], + "bbox": [ + 93, + 205, + 480, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Jan Ehrhardt, Cristian Lorenz, et al. 4D modeling and estimation of respiratory motion for radiation therapy. Springer, 2013. 1", + "[13] Yuyu Guo, Lei Bi, Euijoon Ahn, Dagan Feng, Qian Wang, and Jinman Kim. A spatiotemporal volumetric interpolation network for 4d dynamic medical image. In CVPR, pages 4726-4735, 2020. 1, 3, 6, 7", + "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 30, 2017. 6", + "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 4, 6", + "[16] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. NeurIPS, 35:8633-8646, 2022. 2, 4, 6", + "[17] Kan N Hor, Rolf Baumann, Gianni Pedrizzetti, Gianni Tonti, William M Gottliebson, Michael Taylor, D Woodrow Benson, and Wojciech Mazur. Magnetic resonance derived myocardial strain assessment using feature tracking. Journal of visualized experiments: JoVE, (48):2356, 2011. 1", + "[18] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssim. In 2010 20th international conference on pattern recognition, pages 2366-2369. IEEE, 2010. 6", + "[19] Mengshun Hu, Kui Jiang, Liang Liao, Jing Xiao, Junjun Jiang, and Zheng Wang. Spatial-temporal space hand-in-hand: Spatial-temporal video super-resolution via cycle-projected mutual learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3574-3583, 2022. 3", + "[20] Mengshun Hu, Kui Jiang, Zhihang Zhong, Zheng Wang, and Yinqiang Zheng. Iq-vfi: Implicit quadratic motion estimation for video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6410-6419, 2024. 3", + "[21] Zhewei Huang, Tianyuan Zhang, Wen Heng, Boxin Shi, and Shuchang Zhou. Real-time intermediate flow estimation for video frame interpolation. In European Conference on Computer Vision, pages 624-642. Springer, 2022. 3", + "[22] Geoffrey D Hugo, Elisabeth Weiss, William C Sleeman, Salim Balik, Paul J Keall, Jun Lu, and Jeffrey F Williamson. Data from 4d lung imaging of nsclc patients. (No Title), 2016. 6", + "[23] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In CVPR, pages 2462-2470, 2017. 3", + "[24] Siddhant Jain, Daniel Watson, Eric Tabellion, Ben Poole, Janne Kontkanen, et al. Video interpolation with diffusion models. In CVPR, pages 7341-7351, 2024. 1, 4, 6", + "[25] Mi-Young Jeung, Philippe Germain, et al. Myocardial tagging with mr imaging: overview of normal and pathologic findings. Radiographics, 32(5):1381-1398, 2012. 1", + "[26] Huaizu Jiang, Deqing Sun, Varun Jampani, Ming-Hsuan Yang, Erik Learned-Miller, and Jan Kautz. Super slomo: High quality estimation of multiple intermediate frames for" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28018", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "video interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9000-9008, 2018. 3", + "[27] Xin Jin, Longhai Wu, Jie Chen, Youxin Chen, Jayoon Koo, and Cheul-hee Hahm. A unified pyramid recurrent network for video frame interpolation. In CVPR, pages 1578-1587, 2023. 3", + "[28] Tarun Kalluri, Deepak Pathak, Manmohan Chandraker, and Du Tran. Flavr: Flow-agnostic video representations for fast frame interpolation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2071–2082, 2023. 3", + "[29] Boah Kim and Jong Chul Ye. Diffusion deformable model for 4d temporal medical image generation. In MICCAI, pages 539-548. Springer, 2022. 1, 4, 6, 7", + "[30] JungEun Kim, Hangyul Yoon, Geondo Park, Kyungsu Kim, and Eunho Yang. Data-efficient unsupervised interpolation without any intermediate frame for 4d medical images. In CVPR, pages 11353-11364, 2024. 1, 4, 6, 7", + "[31] Lingtong Kong, Boyuan Jiang, Donghao Luo, Wenqing Chu, Xiaoming Huang, Ying Tai, Chengjie Wang, and Jie Yang. Ifrnet: Intermediate feature refine network for efficient frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1969-1978, 2022. 1, 3, 6", + "[32] Xia Li, Runzhao Yang, Xiangtai Li, Antony Lomax, Ye Zhang, and Joachim Buhmann. Cpt-interp: Continuous spatial and temporal motion modeling for 4d medical image interpolation. arXiv preprint arXiv:2405.15385, 2024. 2", + "[33] Zhen Li, Zuo-Liang Zhu, Ling-Hao Han, Qibin Hou, Chun-Le Guo, and Ming-Ming Cheng. Amt: All-pairs multi-field transforms for efficient frame interpolation. In CVPR, pages 9801-9810, 2023. 3", + "[34] Hanwen Liang, Yuyang Yin, Dejia Xu, Hanxue Liang, Zhangyang Wang, Konstantinos N Plataniotis, Yao Zhao, and Yunchao Wei. Diffusion4d: Fast spatial-temporal consistent 4d generation via video diffusion models. arXiv preprint arXiv:2405.16645, 2024. 4", + "[35] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv preprint arXiv:2402.17177, 2024. 2, 4", + "[36] Liying Lu, Ruizheng Wu, Huaijia Lin, Jiangbo Lu, and Jiaya Jia. Video frame interpolation with transformer. In CVPR, pages 3532-3542, 2022. 3", + "[37] Simone Meyer, Abdelaziz Djerouah, Brian McWilliams, Alexander Sorkine-Hornung, Markus Gross, and Christopher Schroers. Phasenet for video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 498-507, 2018. 1", + "[38] Kotaro Mizuno and Masahiro Muto. Preoperative evaluation of pleural adhesion in patients with lung tumors using four-dimensional computed tomography performed during natural breathing. *Medicine*, 100(47):e27800, 2021. 1", + "[39] Simon Niklaus and Feng Liu. Softmax splatting for video frame interpolation. In Proceedings of the IEEE/CVF con" + ], + "bbox": [ + 91, + 90, + 483, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ference on computer vision and pattern recognition, pages 5437-5446, 2020. 1", + "[40] David Ouyang, Bryan He, Amirata Ghorbani, Neal Yuan, Joseph Ebinger, Curtis P Langlotz, Paul A Heidenreich, Robert A Harrington, David H Liang, Euan A Ashley, et al. Video-based ai for beat-to-beat assessment of cardiac function. Nature, 580(7802):252-256, 2020. 1", + "[41] Junheum Park, Jintae Kim, and Chang-Su Kim. Biformer: Learning bilateral motion estimation via bilateral transformer for 4k video frame interpolation. In CVPR, pages 1568-1577, 2023. 1, 3", + "[42] Saikat Roy, Gregor Koehler, Constantin Ulrich, Michael Baumgartner, Jens Petersen, Fabian Isensee, Paul F Jaeger, and Klaus H Maier-Hein. Mednext: transformer-driven scaling of convnets for medical image segmentation. In MICCAI, pages 405–415. Springer, 2023. 1", + "[43] Otto A Smiseth, Hans Torp, Anders Opdahl, Kristina H Haugaa, and Stig Urheim. Myocardial strain imaging: how useful is it in clinical decision making? European heart journal, 37 (15):1196-1207, 2016. 1", + "[44] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In CVPR, pages 8934-8943, 2018. 3", + "[45] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, pages 402-419. Springer, 2020. 3", + "[46] Georgi P Tolstov. Fourier series. Courier Corporation, 2012. 2, 4", + "[47] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 6", + "[48] Bingchao Wang, Zhiwei Ning, Jianyu Ding, Xuanang Gao, Yin Li, Dongsheng Jiang, Jie Yang, and Wei Liu. Fix-clip: Dual-branch hierarchical contrastive learning via synthetic captions for better understanding of long text. arXiv preprint arXiv:2507.10095, 2025. 2", + "[49] Lu Wang, Shelly Hayes, Kamen Paskalev, Lihui Jin, Mark K Buyyounouski, Charlie C-M Ma, and Steve Feigenberg. Dosimetric comparison of stereotactic body radiotherapy using 4d ct and multiphase ct images for treatment planning of lung cancer: evaluation of the impact on daily dose coverage. *Radiotherapy and Oncology*, 91(3):314–324, 2009. 1", + "[50] Yihan Wang, Lahav Lipson, and Jia Deng. Sea-raft: Simple, efficient, accurate raft for optical flow. In ECCV, pages 36-54. Springer, 2024. 3", + "[51] Tzu-Ti Wei, Chin Kuo, Yu-Chee Tseng, and Jen-Jee Chen. Mpvf: 4d medical image inpainting by multi-pyramid voxel flows. IEEE Journal of Biomedical and Health Informatics, 2023. 1, 2, 3, 6", + "[52] Xin You, Ming Ding, Minghui Zhang, Yangqian Wu, Yi Yu, Yun Gu, and Jie Yang. Semantic difference guidance for the uncertain boundary segmentation of ct left atrial appendage. In MICCAI, pages 121-131. Springer, 2023. 1", + "[53] Xin You, Junjun He, Jie Yang, and Yun Gu. Learning with explicit shape priors for medical image segmentation. IEEE Transactions on Medical Imaging, 2024. 6" + ], + "bbox": [ + 516, + 92, + 903, + 902 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28019", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[54] Xin You, Yixin Lou, Minghui Zhang, Jie Yang, and Yun Gu. Slord: Structural low-rank descriptors for shape consistency in vertebrae segmentation. IEEE Journal of Biomedical and Health Informatics, 2025. 4", + "[55] Xin You, Minghui Zhang, Hanxiao Zhang, Jie Yang, and Nassir Navab. Temporal differential fields for 4d motion modeling via image-to-video synthesis. arXiv preprint arXiv:2505.17333, 2025. 6", + "[56] Weihao Yu, Yuanhao Cai, Ruyi Zha, Zhiwen Fan, Chenxin Li, and Yixuan Yuan. X $^2$ -gaussian: 4d radiative gaussian splatting for continuous-time tomographic reconstruction. arXiv preprint arXiv:2503.21779, 2025. 1", + "[57] Chuyan Zhang, Hao Zheng, Xin You, Yefeng Zheng, and Yun Gu. Pass: test-time prompting to adapt styles and semantic shapes in medical image segmentation. IEEE Transactions on Medical Imaging, 2024. 6", + "[58] Guozhen Zhang, Yuhan Zhu, Haonan Wang, Youxin Chen, Gangshan Wu, and Limin Wang. Extracting motion and appearance via inter-frame attention for efficient video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5682-5692, 2023. 1, 3", + "[59] Haiyu Zhang, Xinyuan Chen, Yaohui Wang, Xihui Liu, Yunhong Wang, and Yu Qiao. 4diffusion: Multi-view video diffusion model for 4d generation. NeurIPS, 37:15272-15295, 2025. 4", + "[60] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 6", + "[61] Zhihang Zhong, Gurunandan Krishnan, Xiao Sun, Yu Qiao, Sizhuo Ma, and Jian Wang. Clearer frames, anytime: Resolving velocity ambiguity in video frame interpolation. In ECCV, pages 346-363. Springer, 2024. 1, 6", + "[62] Kun Zhou, Wenbo Li, Xiaoguang Han, and Jiangbo Lu. Exploring motion ambiguity and alignment for high-quality video frame interpolation. In CVPR, pages 22169-22179, 2023. 1" + ], + "bbox": [ + 91, + 92, + 480, + 626 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "28020", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/FB-Diff_ Fourier Basis-guided Diffusion for Temporal Interpolation of 4D Medical Imaging/c8ad7765-483e-400a-90bd-72090b0597d0_model.json b/2025/FB-Diff_ Fourier Basis-guided Diffusion for Temporal Interpolation of 4D Medical Imaging/c8ad7765-483e-400a-90bd-72090b0597d0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b915f852ce9d37a5d9ca728b3b2e138b071b4d10 --- /dev/null +++ b/2025/FB-Diff_ Fourier Basis-guided Diffusion for Temporal Interpolation of 4D Medical Imaging/c8ad7765-483e-400a-90bd-72090b0597d0_model.json @@ -0,0 +1,2103 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.13, + 0.13, + 0.87, + 0.176 + ], + "angle": 0, + "content": "FB-Diff: Fourier Basis-guided Diffusion for Temporal Interpolation of 4D Medical Imaging" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.202, + 0.9, + 0.222 + ], + "angle": 0, + "content": "Xin You\\(^{1,2,3}\\) Runze Yang\\(^{1}\\) Chuyan Zhang\\(^{1}\\) Zhongliang Jiang\\(^{2}\\) Jie Yang\\(^{1,*}\\) Nassir Navab\\(^{2,*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.185, + 0.239, + 0.819, + 0.258 + ], + "angle": 0, + "content": "\\(^{1}\\) Institute of Medical Robotics, Shanghai Jiao Tong University, Shanghai, China" + }, + { + "type": "text", + "bbox": [ + 0.143, + 0.258, + 0.86, + 0.277 + ], + "angle": 0, + "content": "\\(^{2}\\) Computer Aided Medical Procedures, Technical University of Munich, Munich, Germany" + }, + { + "type": "text", + "bbox": [ + 0.272, + 0.276, + 0.732, + 0.295 + ], + "angle": 0, + "content": "3 Munich Center for Machine Learning, Munich, Germany" + }, + { + "type": "text", + "bbox": [ + 0.362, + 0.297, + 0.637, + 0.31 + ], + "angle": 0, + "content": "xin.you@tum.de, nassir.navab@tum.de" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.347, + 0.327, + 0.362 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.379, + 0.485, + 0.771 + ], + "angle": 0, + "content": "The temporal interpolation task for 4D medical imaging, plays a crucial role in clinical practice of respiratory motion modeling. Following the simplified linear-motion hypothesis, existing approaches adopt optical flow-based models to interpolate intermediate frames. However, realistic respiratory motions should be nonlinear and quasi-periodic with specific frequencies. Intuited by this property, we resolve the temporal interpolation task from the frequency perspective, and propose a Fourier Basis-guided Diffusion model, termed FB-Diff. Specifically, due to the regular motion discipline of respiration, physiological motion priors are introduced to describe general characteristics of temporal data distributions. Then a Fourier motion operator is elaborately devised to extract Fourier bases by incorporating physiological motion priors and case-specific spectral information in the feature space of Variational Autoencoder. Well-learned Fourier bases can better simulate respiratory motions with motion patterns of specific frequencies. Conditioned on starting and ending frames, the diffusion model further leverages well-learned Fourier bases via the basis interaction operator, which promotes the temporal interpolation task in a generative manner. Extensive results demonstrate that FB-Diff achieves state-of-the-art (SOTA) perceptual performance with better temporal consistency while maintaining promising reconstruction metrics. Codes are available here." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.8, + 0.222, + 0.815 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "4D medical images, which depict 3D volumes with temporal variations, are essential in clinical practice for capturing dynamic changes and monitoring the progression of diseases over time [12, 25, 49, 52]. A significant application is the temporal modeling for breathing-induced motions of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.348, + 0.907, + 0.424 + ], + "angle": 0, + "content": "cardiac or pulmonary anatomical structures. That task aims to enrich motion visualizations by generating intermediate frames under the condition of starting and ending frames, which is consistent with the concept of Video Frame Interpolation (VFI) in natural scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.427, + 0.909, + 0.639 + ], + "angle": 0, + "content": "VFI has been thoroughly investigated through systematic studies in recent years [10, 24, 31, 37, 39, 41, 58, 61, 62]. However, due to unique properties and constraints of medical imaging, it remains a challenge to directly implement these frameworks into 4D medical video interpolation with maintaining equally promising performance. Specifically, the quality of ground truth intermediate frames in medical imaging is often compromised due to factors such as imaging noise, unstable breathing, causing unexpected artifacts [17, 25, 43, 49]. Thus, the lower image quality compared to videos in the natural domain, will degrade the interpolation performance of VFI models. Besides, potential patients' movement might induce an additional disturbance to the temporal modeling [5, 38]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.909, + 0.778 + ], + "angle": 0, + "content": "Furthermore, there exists a discrepancy on the type of motions between natural and medical scenes. The former focuses on the object movement and illumination intensity change. While in the medical domain, motions can be depicted as subtle anatomical variance with quasi-periodic motion discipline [51, 56]. Thus, a specific motion prior of anatomies is effective to boost the temporal interpolation, especially for the medical data with limited resources from a small pool of individuals." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "On account of these challenges, we present the following question: \"How can we devise a specialist VFI model tailored for 4D temporal medical volumes?\" Existing methods [2, 13, 29, 30, 51] on medical VFI face a major shortcoming: the linear motion hypothesis in a breathing period is required for the regression of intermediate frames as revealed in Fig.1(a). However, respiratory motions should be a nonlinear oscillatory process that cannot be simply mod-" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28010" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.141, + 0.089, + 0.862, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.334, + 0.363, + 0.684, + 0.379 + ], + "angle": 0, + "content": "(b) Nonlinear & quasi-periodic motion \\(\\Longrightarrow\\) Our model" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.39, + 0.908, + 0.434 + ], + "angle": 0, + "content": "Figure 1. Temporal Interpolation for respiratory motions. (a) Simplified linear-motion hypothesis: interpolating intermediate frames by regressing optical flows with linear variations. (b) Nonlinear & quasi-periodic motions: leveraging diffusion models conditioned on Fourier motion bases, which represent various motion patterns indicating nonlinear and locally inconsistent anatomical variations." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.459, + 0.483, + 0.7 + ], + "angle": 0, + "content": "eled with a linear equation [32, 51]. Also, these methods fail to recognize that respiratory motions in heart or lung, are indeed a quasi-periodic motion process with specific frequencies by extending the timeline. Intuited by this, we resolve the respiratory motion simulation from a frequency perspective. In Fig.1(b), each medical volume in the temporal domain can be viewed as a specific value in the periodic curve. That periodic curve can be decomposed as a group of sine and cosine bases with multiple frequencies [46]. Comparatively, there exists a group of Fourier bases, which are capable of reconstructing original temporal sequences. Here Fourier bases with different frequencies refer to various motion patterns, which indicate nonlinear and locally anisotropic variations of anatomical structures. In our work, Fourier bases are acquired by the Fast Fourier Transform (FFT) on the temporal domain." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In this work, 4D data distributions of the respiratory process depict a regular motion discipline, with temporal variations from inhaling to exhaling, or from diastole to systole. Motivated by that observation, physiology motion priors are introduced to encode generic characteristics of the respiratory process. Then a Fourier Motion Operator \\(O_{FM}\\) is elaborately devised to yield enhanced motion bases by incorporating physiology motion priors and case-specific spectral information in the feature space. Specifically, A Variational Autoencoder (VAE) is introduced to map the video space into the feature space. Physiology motion priors are defined as learnable frequency embeddings, which can be updated by pretraining a Variational Autoen" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.459, + 0.907, + 0.732 + ], + "angle": 0, + "content": "coder (VAE) on given temporal data. And case-specific frequency information can be extracted by FFT in the temporal domain. Well-learned physiology motion priors can enrich Fourier bases with generic properties of respiratory motions. Then enhanced Fourier motion bases can better simulate the quasi-periodic respiratory process with motion patterns of specific frequencies. On the grounds that diffusion models show promising performance in generating realistic videos [15, 16, 35, 48], a Markov chain process is learned to transform Gaussian noise into target videos by the proposed Fourier Bases-guided Diffusion, termed FB-Diff. Since only starting and ending frames are available during inference, they serve as conditional prompts for intermediate frame interpolation. Meanwhile, we devise the basis interaction operator to inject enhanced Fourier bases into the conditional diffusion model. Then, FB-Diff can well address the temporal interpolation task in a generative manner." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Our proposed method has achieved state-of-the-art (SOTA) interpolation performance compared with other methods for 4D medical imaging, particularly on the mid-most frames. For quantitative results, more comprehensive evaluation metrics are adopted in this work, including traditional reconstruction metrics and feature-level perceptual metrics. FB-Diff not only presents promising reconstruction metrics but also superior perceptual results. Qualitative visualizations reveal better temporal consistency by FB-Diff compared with other linear-motion-based methods. Our main contributions are summarized as follows:" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28011" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.129, + 0.089, + 0.87, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.507, + 0.907, + 0.568 + ], + "angle": 0, + "content": "Figure 2. The pipeline of FB-Diff. (a) The Fourier motion operator is proposed to extract Fourier bases in a feature space of VAE. Bases consist of learnable generic motion priors and case-specific frequency information. (b) Fourier bases are first adapted to the masked video domain under the guidance of well-learned motion priors and pretrained VAE weights. Domain-adapted bases serve as the conditional input to boost the video synthesis by diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.59, + 0.483, + 0.666 + ], + "angle": 0, + "content": "- We propose a diffusion model termed FB-Diff for the temporal interpolation of 4D medical imaging. FB-Diff leverages Fourier motion bases, which indicate motion patterns with specific frequencies to resolve the respiratory modeling with nonlinear and quasi-periodic motions." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.666, + 0.483, + 0.711 + ], + "angle": 0, + "content": "- We introduce the Fourier motion operator \\( O_{FM} \\) to extract Fourier bases, containing physiology motion priors and case-specific frequency representations." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.711, + 0.483, + 0.756 + ], + "angle": 0, + "content": "- We devise the basis interaction operator to inject Fourier bases into the diffusion model. Fourier bases can boost the video synthesis for temporal interpolation." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.756, + 0.483, + 0.801 + ], + "angle": 0, + "content": "- Extensive results demonstrate that FB-Diff reveals SOTA perceptual performance with better temporal consistency while maintaining promising reconstruction metrics." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.59, + 0.483, + 0.801 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.815, + 0.233, + 0.831 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Video Frame Interpolation. Popular VFI techniques are mainly classified into motion-free and motion-based approaches, depending on whether optical flows are incorporated or not. Motion-based: These approaches warp" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.59, + 0.907, + 0.788 + ], + "angle": 0, + "content": "input frames forward or backward based on optical flows acquired by either off-the-shelf networks [11, 23, 44, 45, 50] or custom-designed flow estimators [20, 21, 27, 31, 33, 58]. Warped intermediate frames are then refined by delicately designed networks [3, 26] to enhance visual quality. Recently, Transformer-based motion estimators are proposed to raise the upper limit of VFI performance [36, 41, 58] Motion-Free: Motion-free methods rely on implicit spatiotemporal modeling [8, 9, 19, 28] to generate the intermediate frame. However, due to a limited quality of 4D medical videos and quasi-periodic motion priors, existing VFI frameworks are not feasible. That is why we aim to propose a specialist VFI model for 4D medical sequences." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "4D Medical VFI. Existing methods on this topic mainly leverage volumetric motions for 4D temporal interpolation. Specifically, SVIN [13] estimates forward and backward deformation fields, which will yield precise intermediate frames by linearly combining bidirectional information. MPVF [51] resolves various magnitudes of motions by proposing a multi-pyramid voxel flows model that takes" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28012" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.227 + ], + "angle": 0, + "content": "multi-scale voxel flows into account. DDM [29] can learn spatial deformation information between the source and target volumes and provide a latent code for generating intermediate frames along a geodesic path. UVI-Net [30] utilizes the flow calculation model with the time-domain cycle-consistency constraint and linear motion hypothesis, to realize motion modeling in an unsupervised style. However, the linear hypothesis for motions will induce temporal inconsistency and spatial distortion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.231, + 0.485, + 0.429 + ], + "angle": 0, + "content": "Diffusion-based Video Interpolation. Denoising diffusion probabilistic models [15, 16, 35] reveal promising performance in generating realistic images or videos. Some methods maximize the potential of diffusion models to resolve video interpolation tasks. Essentially, these approaches implicitly learn the temporal motions from the starting frame to the ending frame [1, 7, 24]. LDMVFI [10] proposes a latent diffusion model, which approaches the VFI problem from a generative perspective by formulating it as a conditional synthesis issue. However, these diffusion models do not leverage effective conditional guidance, which is crucial for medical VFI with intrinsic motion priors. Thus, we introduce Fourier bases to enhance the conditional synthesis." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.44, + 0.226, + 0.458 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.465, + 0.295, + 0.48 + ], + "angle": 0, + "content": "3.1. Fourier Motion Bases" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.487, + 0.484, + 0.653 + ], + "angle": 0, + "content": "Since respiratory motions for heart or lung exhibit a quasi-periodic characteristic, we intend to resolve this task from a frequency perspective. Specifically, the respiration can be viewed as an oscillatory process with certain frequencies. Thus, we draw an analogy between 2D time-varying curves and 4D temporal medical volumes. Curves can be decomposed into a group of harmonic oscillators with various frequencies [46]. Similarly, temporal volumes can be restored via a stack of extracted Fourier bases, which are embodied with various motion patterns of anatomical deformation. The whole process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.654, + 0.483, + 0.694 + ], + "angle": 0, + "content": "\\[\n\\mathcal {B} _ {k} = \\sum_ {n = 0} ^ {N - 1} V _ {n} e ^ {- i 2 \\pi k \\frac {n}{N}}, k = 0, 1, \\dots , N - 1, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.126, + 0.697, + 0.483, + 0.737 + ], + "angle": 0, + "content": "\\[\nV _ {n} = \\frac {1}{T} \\sum_ {k = 0} ^ {N - 1} \\mathcal {B} _ {k} e ^ {i 2 \\pi n \\frac {k}{N}}, n = 0, 1, \\dots , N - 1, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.747, + 0.484, + 0.868 + ], + "angle": 0, + "content": "where \\(\\mathcal{B}_k\\) means the \\(k_{th}\\) Fourier basis in the frequency domain, \\(V_{n}\\) refers to the \\(n_{th}\\) medical volume, and \\(N\\) is the frame number. These Fourier bases describe locally anisotropic anatomical variations with different motion intensities, promoting the performance of the temporal interpolation model \\(\\mathcal{I}\\). Thus, the goal of our work is to deduce 4D medical videos \\(V\\) conditioned on two prompting frames \\(V_{0}, V_{N-1}\\) and Fourier motion bases \\(\\mathcal{B}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.879, + 0.483, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {I} ^ {*} = \\underset {\\mathcal {I}} {\\arg \\min } \\| \\boldsymbol {V} - \\mathcal {I} \\left(V _ {0}, V _ {N - 1}; \\mathcal {B}\\right) \\| _ {\\mathcal {D}}, \\tag {3}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.089, + 0.895, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.276, + 0.907, + 0.318 + ], + "angle": 0, + "content": "Figure 3. The details of Fourier motion operator. Fourier bases are decomposed as physiology motion priors and case-specific Fourier features, corresponding to phase and magnitude components." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.326, + 0.907, + 0.476 + ], + "angle": 0, + "content": "\\(\\mathcal{I}^*\\) refers to the optimal solution in the interpolation model set, and \\(\\mathcal{D}\\) represents the distance-based normed space. We systematically organize the following sections along Fourier bases. Section 3.2 introduces the design of the Fourier Motion Operator, which constructs Fourier bases with physiology motion priors and case-specific frequency information. Section 3.3 adapts Fourier bases to the to-be-interpolated video domain with well-learned physiology motion priors. Section 3.4 provides details of the diffusion model conditioned on Fourier bases." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.487, + 0.747, + 0.503 + ], + "angle": 0, + "content": "3.2. Fourier Motion Operator" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.907, + 0.735 + ], + "angle": 0, + "content": "To explicitly extract the Fourier bases with various motion patterns for the respiration process, we introduce the Fourier Operator \\(O_{\\mathcal{F}}\\) as a neural operator. As stated in Eq.(2), \\(O_{\\mathcal{F}}\\) will operate in the temporal domain, not the spatial domain. Since the common Fourier operator \\(O_{\\mathcal{F}}\\) only mines case-specific frequency representations, general physiology characteristics inside respiratory data are ignored. Inspired by that notion, we devise the Fourier motion operator \\(O_{FM}\\) as a neural operator in the feature space. To map the image space into feature space, we introduce a Variational Autoencoder (VAE) as revealed in Fig.2(a). Due to no built-in 4D convolutional operator, we replace it with 3D spatial convolution and 1D temporal convolution as [34, 59] did. The spatial layer in Fig.2(a) consists of spatial convolution, GELU, and group normalization operations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.736, + 0.909, + 0.902 + ], + "angle": 0, + "content": "The Fourier motion operator \\(O_{FM}\\) is devised to acquire spectral characteristics with both physiology motion priors and case-specific representations in the feature space. Firstly, physiology motion priors are introduced as learnable frequency embeddings, which are updated by backward gradients of restoring given temporal data via VAE. Well-trained frequency embeddings \\(\\theta\\) can represent the generic discipline of the regular respiration process [54], with \\(\\theta_{k}\\) referring to the \\(k_{th}\\) element \\((k = 0,1,\\dots,N - 1)\\). Then Fourier bases \\(\\mathcal{B}\\) are made up of physiology motion priors \\(\\theta\\) and case-specific features \\(F\\), which arise from the output of" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28013" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.123 + ], + "angle": 0, + "content": "spatial and temporal layers. As depicted by Fig.3, \\(\\mathcal{B}\\) with a complex type is yielded in the following style:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.131, + 0.483, + 0.173 + ], + "angle": 0, + "content": "\\[\n\\mathcal {B} = \\underbrace {\\mathcal {M} (\\left\\| O _ {\\mathcal {F}} (F) \\right\\| _ {2})} _ {\\text {C a s e - S p e c i f i c}} \\cdot \\underbrace {e ^ {i U (\\tau (\\theta))}} _ {\\text {G e n e r a l}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.183, + 0.485, + 0.318 + ], + "angle": 0, + "content": "\\(\\tau\\) means the channel transform, \\(U\\) refers to the upsampling which unifies the resolution of \\(\\theta\\) and \\(\\pmb{F}\\). These operations transform general motion priors into the phase part of Fourier bases, which determines the motion variation standing for the whole dataset distribution. Then \\(\\| \\cdot \\| _2\\) means the \\(L_{2}\\) norm, \\(\\mathcal{M}\\) is the convolutional mapping operation. They will output motion bases' magnitude information, which selectively highlights regions in spectral maps relevant to case-specific information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.319, + 0.485, + 0.365 + ], + "angle": 0, + "content": "Meanwhile, feature maps \\(F\\) are further enriched with more fine-grained frequency details. The specific mechanism can be described with Eq.(5):" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.374, + 0.483, + 0.395 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {F}} = \\boldsymbol {F} + \\sigma \\mathcal {O} _ {\\mathcal {F}} ^ {- 1} (\\mathcal {B}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.404, + 0.484, + 0.453 + ], + "angle": 0, + "content": "here \\(O_{\\mathcal{F}}^{-1}\\) and \\(\\sigma\\) refer to the inverse Fourier transform and Sigmoid operators, which will boost deep features \\(\\hat{F}\\) with enriched details, including shape and texture information." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.463, + 0.435, + 0.479 + ], + "angle": 0, + "content": "3.3. Frame Inpainting-based Basis Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.485, + 0.485, + 0.68 + ], + "angle": 0, + "content": "Due to the fact that only starting and ending frames can be acquired during inference, a significant step is to adapt Fourier bases to the masked video domain. Motivated by the fact that strong correlations exist between Fourier motion bases and inpainted temporal sequences, we plan to jointly optimize these two variables. To accomplish that goal, we repurpose well-calibrated physiology motion priors from VAE, which prove to be effective in providing the motion guidance for intermediate frame inpainting. Also, pretrained weights of VAE encode implicit temporal dependencies between frames, enhancing the reconstruction of masked video sequences via the finetuned VAE noted as \\(\\mathcal{G}\\). A mathematical formulation is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.69, + 0.483, + 0.711 + ], + "angle": 0, + "content": "\\[\n\\hat {V}, \\hat {\\mathcal {B}} = \\mathcal {G} (\\hat {\\boldsymbol {\\theta}}, V _ {\\boldsymbol {M}}) = \\mathcal {G} (\\hat {\\boldsymbol {\\theta}}, V _ {0},.., \\mathbf {0},.., V _ {N - 1}), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.483, + 0.764 + ], + "angle": 0, + "content": "\\(\\hat{\\theta}\\) and \\(V_{M}\\) correspond to well-trained general motions and zero-masked videos, \\(\\hat{V}\\) and \\(\\hat{B}\\) refer to inpainted videos and domain-adapted Fourier bases." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "The interaction between temporal video representations and \\(\\hat{\\mathcal{B}}\\) is implemented via the Fourier motion operator as shown in Fig.2(b). For the joint optimization process, video sequences will contribute temporal variations, enhancing the synthesis of adapted Fourier bases. Conversely, Fourier bases for zero-masked inputs inherently generate a spectral map that emphasizes regions with motion in the temporal axis. This activated map aids in temporal frame inpainting by exerting attention on regions with different magnitudes." + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.091, + 0.887, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.244, + 0.906, + 0.313 + ], + "angle": 0, + "content": "Figure 4. The architecture of basis interaction operator. Fourier motion bases selectively enhance video representations by measuring the motion intensity of different regions. Then those enhanced temporal representations can enrich denoised embeddings from the diffusion model." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.321, + 0.845, + 0.337 + ], + "angle": 0, + "content": "3.4. Basis-guided 4D Conditional Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.342, + 0.906, + 0.448 + ], + "angle": 0, + "content": "VAE in Fig.2(a) is not only devised for universal motion priors, but also for the latent diffusion model (LDM), which shows better compatibility for extra conditional inputs. Besides, LDM synthesizes latent embeddings occupying less GPU memory compared to 4D temporal volumes. And the VAE decoder is leveraged to reconstruct interpolated video sequences with synthesized latent embeddings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.449, + 0.906, + 0.495 + ], + "angle": 0, + "content": "For the specific diffusion process, the whole video sequence \\( V \\) is first encoded as a latent embedding \\( z_0 = \\mathcal{G}_e(V) \\). Then the latent code \\( z_0 \\) is perturbed as:" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.499, + 0.905, + 0.518 + ], + "angle": 0, + "content": "\\[\nz _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} z _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , \\epsilon \\sim \\mathcal {N} (0, 1), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.523, + 0.906, + 0.63 + ], + "angle": 0, + "content": "where \\(\\overline{\\alpha}_t = \\prod_{i=1}^t (1 - \\beta_t)\\) with \\(\\beta_t\\) is the noise coefficient at time step \\(t\\), and \\(t\\) is uniformly sampled from the timestep index set \\([1,2,\\dots,T]\\). This process can be regarded as a Markov chain, which incrementally adds Gaussian noise to the latent code \\(z_0\\). The denoising model \\(d\\) receives \\(z_t\\) as input and is optimized under conditional guidance to learn the latent space distribution with the objective function:" + }, + { + "type": "equation", + "bbox": [ + 0.539, + 0.634, + 0.905, + 0.654 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\epsilon} = E _ {z _ {t}, \\epsilon \\sim \\mathcal {N} (0, 1)} \\| \\epsilon - \\epsilon_ {d} \\left(z _ {t}, t; \\hat {\\mathcal {B}}, \\mathcal {G} _ {e} \\left(\\boldsymbol {V} _ {\\boldsymbol {M}}\\right), \\hat {\\boldsymbol {V}}\\right) \\| ^ {2}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.906, + 0.765 + ], + "angle": 0, + "content": "here \\(\\epsilon_d\\) is the regressed noise, adapted Fourier bases \\(\\hat{\\mathcal{B}}\\) and inpainted coarse videos \\(\\hat{V}\\) are incorporated as the conditional guidance. Moreover, feature maps of starting and ending frames by the VAE encoder \\(\\mathcal{G}_e\\) serve as a strong prompt for the 4D video synthesis. Then during the training process of FB-Diff, the total loss \\(\\mathcal{L}\\) consists of the denoising loss \\(\\mathcal{L}_{\\epsilon}\\) and the \\(L_{2}\\)-norm reconstruction loss \\(\\mathcal{L}_r(\\hat{V},V)\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.907, + 0.902 + ], + "angle": 0, + "content": "To inject Fourier motion bases into LDM and harness coarsely inpainted results, we design the Basis Interaction Operator (BIO) as shown in Fig.4. Specifically, inpainted videos are mapped into video feature representations via spatial & temporal layers and the downsampling operation. Fourier motion bases with various motion patterns, selectively enhance video representations by measuring the motion intensity of different regions. Those enhanced temporal representations can enrich denoised embeddings from" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28014" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 1. Baseline comparison with flow-based and diffusion-based models. (Bold: the best, Underlined: the second best. Δ values represent the relative difference between our model and UVI-Net, and red and blue numbers stand for better and worse results)." + }, + { + "type": "table", + "bbox": [ + 0.137, + 0.128, + 0.868, + 0.334 + ], + "angle": 0, + "content": "
ModelACDC Cardiac4D Lung
PSNR (dB) ↑LPIPS ↓FID ↓FVD ↓PSNR (dB) ↑LPIPS ↓FID ↓FVD ↓
SVIN [13]31.43±0.4211.563±0.20626.393.630.49±0.3042.650±0.24538.5125.6
Voxelmorph [2]30.77±0.5021.969±0.19729.4102.129.90±0.3732.815±0.26043.2149.0
Transmorph [6]29.81±0.5322.136±0.25531.5109.529.04±0.4032.946±0.29140.5143.5
MPVF [51]31.53±0.4101.594±0.20124.891.330.76±0.2652.418±0.24639.4120.3
UVI-Net [30]32.16±0.4021.662±0.24523.094.231.57±0.3112.211±0.21637.8121.7
IFRNet [31]30.58±0.4081.752±0.24927.897.530.05±0.2802.316±0.27339.7124.0
LDMVFI [10]27.11±0.4602.943±0.41028.899.226.31±0.4533.659±0.34142.1142.7
LDDM [7]24.53±0.4812.634±0.30833.4105.725.19±0.2732.914±0.28748.7146.3
Conditional Diff [16]26.59±0.5452.460±0.35730.195.725.95±0.3913.294±0.37545.0133.0
DDM [29]29.79±0.5042.089±0.35227.4110.329.67±0.4202.705±0.33040.8165.5
FB-Diff30.95±0.3861.452±0.20521.588.930.18±0.2872.066±0.21537.7113.6
Δ value1.21 ↓0.210 ↓1.5 ↓5.3 ↓1.39 ↓0.145 ↓0.1 ↓8.1 ↓
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.37, + 0.486, + 0.418 + ], + "angle": 0, + "content": "the diffusion model, promoting a promising synthesis of highlighted regions. Thus, BIO can boost well-interpolated videos with spatial and temporal consistency." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.43, + 0.226, + 0.447 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.455, + 0.3, + 0.472 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.478, + 0.483, + 0.705 + ], + "angle": 0, + "content": "Dataset. To evaluate the interpolation performance of our framework, we conduct experiments on public ACDC cardiac [4] and 4D-Lung datasets [22]. For ACDC with the MRI modality, volume sequences between the end-diastolic and end-systolic phases are extracted as valid 4D data. Of all the 150 4D sequences, cases with identity 1-100, 101-120, and 121-150 serve as the training, validation, and testing sets. Besides, all MRI volumes are cropped to \\(128 \\times 128 \\times 32\\), and the frame number ranges from 6 to 16 [53, 57]. For 4D-Lung CT images, the end-inspiratory and end-expiratory scans are set as the initial and final images. We collect 125 4D videos from [22], with the resolution equal to \\(6 \\times 128 \\times 128 \\times 128\\). They are split as \\(80 / 15 / 30\\) cases for training, validation, and inference. We uniform the frame number as 16 with zero paddings." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.705, + 0.483, + 0.886 + ], + "angle": 0, + "content": "**Evaluations.** We select quantitative evaluation metrics including Peak Signal-to-Noise Ratio (PSNR) [18], Learned Perceptual Image Patch Similarity (LPIPS) [55, 60], Fréchet Inception Distance (FID) [14], and Fréchet Video Distance (FVD) [47]. Compared to PSNR, which is the traditional pixel-wise reconstruction metric, LPIPS and FVD evaluate perceptual quality from the perspective of human vision [10, 24, 61], which is more feasible and favored in clinical practice. Specifically, LPIPS reveals important perceptual similarity like human vision. FID and FVD evaluate the dataset distribution bias between realistic and synthesized data, including spatial and temporal consistency." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Training Details. The VAE maps the image space into" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.37, + 0.907, + 0.492 + ], + "angle": 0, + "content": "the downsampled latent space with a ratio of \\(1/8\\). And the proposed conditional diffusion model is implemented with a lightweight UNet. All models are trained using the AdamW optimizer with the linear warm-up strategy. The batch size, diffusion step \\(T\\), and training epoch are equal to 2, 1000, and 500. Experiments are implemented based on Pytorch and 2 NVIDIA RTX 4090 GPUs. More details about the network and training settings are described in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.501, + 0.816, + 0.517 + ], + "angle": 0, + "content": "4.2. Comparisons with SOTA baselines" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.523, + 0.907, + 0.765 + ], + "angle": 0, + "content": "Quantitative Comparison. We comprehensively conduct experiments on ACDC cardiac and 4D Lung datasets. Quantitative results are illustrated in Tab.1. Our model achieves the best performance across the benchmark on perceptual evaluation metrics from the perspective of human vision [10, 61], which are clinically favored compared with the pixel-centric reconstruction metric [7]. Specifically, FB-Diff outperforms UVI-Net [30] with \\(0.210 \\downarrow\\) LPIPS and \\(5.3 \\downarrow\\) FVD values on ACDC. Also, our model maintains quite satisfactory PSNR values of \\(30.95\\mathrm{dB}\\) and \\(30.18\\mathrm{dB}\\) on ACDC and 4D Lung. Although diffusion models are not sensitive to pixel-wise intensities [15], the conditional guidance of Fourier bases and coarsely inpainted videos is effective to boost the interpolation performance of diffusion models. Tab.1 reveals that FB-Diff significantly surpasses existing diffusion models on all metrics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Furthermore, the interpolation process becomes most challenging for the midmost frames since they are positioned farthest from the two reference frames [24]. And our model can better resolve this issue due to stronger temporal representations based on quasi-periodic motion modeling. Thus, a quantitative evaluation of central frames is necessary. As shown in Tab.2, FB-Diff demonstrates greater performance advantages, with \\(4.3\\downarrow\\) FID and \\(8.2\\downarrow\\) FVD values compared with UVI-Net. Additionally, FB-Diff acquires a competi" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28015" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.138, + 0.09, + 0.868, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.159, + 0.358, + 0.837, + 0.373 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison of the midmost frame interpolation between different models. VM: Voxelmorph." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.397, + 0.484, + 0.455 + ], + "angle": 0, + "content": "Table 2. Interpolation performance of the midmost frames on ACDC cardiac. The choice of midmost frames: if the frame number \\( N \\) is odd, three intermediate frames are selected, otherwise only two frames are chosen." + }, + { + "type": "table", + "bbox": [ + 0.106, + 0.463, + 0.475, + 0.548 + ], + "angle": 0, + "content": "
ModelPSNR (dB) ↑LPIPS ↓FID ↓FVD ↓
SVIN [13]26.01±0.4672.673±0.36138.9142.8
DDM [29]24.37±0.5112.756±0.33041.5145.1
UVI-Net [30]26.25±0.4102.542±0.37437.0137.1
FB-Diff26.18±0.4132.287±0.30532.7128.9
Δ value0.07 ↓0.255 ↓4.3 ↓8.2 ↓
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.565, + 0.484, + 0.595 + ], + "angle": 0, + "content": "tive PSNR value, which demonstrates the effectiveness of respiration-related modeling via Fourier bases." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.596, + 0.484, + 0.822 + ], + "angle": 0, + "content": "Qualitative Comparison. Qualitative results on the interpolated midmost frame are depicted in Fig.5. SVIN [13], Voxelmorph [2], and DDM [29] leverage the single-direction flow estimation to interpolate intermediate frames with linear motions. Compared with that, UVI-Net [30] constructs the interpolation model with time-domain cycle-consistency, and achieves better reconstruction performance as expected. However, these methods essentially overlook the nonlinear and quasi-periodic motion discipline between temporal sequences, resulting in unsatisfactory interpolated results. Instead, FB-Diff exhibits advantageous predictions for midmost frames under the guidance of various motion patterns, indicating locally anisotropic motions with different intensities. Besides, other diffusion models do not work well due to the lack of fine-grained conditional guidance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.245, + 0.849 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Key Components. An ablation study is conducted to investigate the significance of various components in FB-Diff as shown in Tab.3. (a) Physiology Motion Priors \\(\\theta\\). By" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.397, + 0.907, + 0.426 + ], + "angle": 0, + "content": "Table 3. Ablation study on key components. \\(\\theta\\): physiology motion priors, \\(\\hat{B}\\): Fourier motion bases, \\(\\hat{V}\\): coarsely inpainted videos." + }, + { + "type": "table", + "bbox": [ + 0.528, + 0.435, + 0.897, + 0.519 + ], + "angle": 0, + "content": "
SettingsACDC Cardiac4D Lung
θˆBˆVPSNR (dB) ↑LPIPS ↓PSNR (dB) ↑LPIPS ↓
X✓✓29.761.71029.532.413
XX✓28.172.10528.402.971
✓✓X30.291.65529.712.246
✓✓✓30.951.45230.182.066
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.907, + 0.779 + ], + "angle": 0, + "content": "removing \\(\\theta\\) from the Fourier motion operator, interpolation performance decreases significantly due to the loss of general spectral characteristics from respiratory data. Therefore, Fourier bases lack the precise phase information. (b) Fourier bases \\(\\hat{\\mathcal{B}}\\). Without \\(\\hat{\\mathcal{B}}\\) as conditional guidance, the conditional diffusion model fails to perceive explicit motion patterns with different frequencies, which provide the intensity of anisotropic motions. As a result, the generative model cannot well learn the regular temporal variations, with 2.78 dB \\(\\downarrow\\) and 1.78 dB \\(\\downarrow\\) PSNR values on ACDC and 4D Lung. (c) Inpainted Videos \\(\\hat{V}\\). Here the coarsely reconstructed sequences serve as the output by the finedtuned VAE as revealed in Fig.2. And the joint optimization process boosts the spatial and temporal representations of \\(\\hat{V}\\). Consequently, discarding this component will reduce fine-grained details of the diffusion model's output." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Hyper-settings of Fourier Bases. (a) The number of Fourier Bases \\( N_{Fb} \\). We first analyze the quantitative effects of the frequency number on the interpolation performance. As revealed in Tab.4, reducing the frequency number will cause degraded performance of the temporal simulation. Specifically, 1.30 dB lower PSNR and 0.535 higher LPIPS values are induced by decreasing \\( N_{Fb} \\) from 16 to 8. That phenomenon indicates that some elements in the group of" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28016" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.089, + 0.887, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.188, + 0.908, + 0.232 + ], + "angle": 0, + "content": "Figure 6. Visualizations on the frequency spectral of Fourier bases. \\( V_{0} \\), \\( V_{N - 1} \\): starting and ending frames. \\( f_{1} - f_{8} \\) refer to different motion patterns, which highlight different regions under different frequency. There also appear locally anisotropic motion intensities inside a specific motion pattern." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.255, + 0.483, + 0.283 + ], + "angle": 0, + "content": "Table 4. Ablation study on the hyper-settings of Fourier bases. \\(N_{Fb}\\): the frequency number of Fourier bases." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.294, + 0.485, + 0.387 + ], + "angle": 0, + "content": "
SettingsACDC Cardiac4D Lung
PSNR (dB) ↑LPIPS ↓PSNR (dB) ↑LPIPS ↓
\\(N_{Fb}\\)=16 (Ours)30.951.45230.182.066
\\(N_{Fb}\\)=1230.761.64029.972.570
\\(N_{Fb}\\)=829.651.98729.502.745
w/o Real part29.181.87229.562.453
w/o Imaginary part29.881.68129.672.401
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.404, + 0.484, + 0.541 + ], + "angle": 0, + "content": "bases with the frequency ranging from \\( f_{8} \\) to \\( f_{15} \\) are consistent with specific frequencies of the respiratory process. (b) Real & Imaginary Parts. \\( \\hat{\\mathcal{B}} \\) bears real and imaginary parts after combining generic priors and case-specific spectral representations, which perform as the phase and magnitude information. Eliminating either the real or imaginary part will convert the complex data type into the real one, resulting in the loss of phase information and influencing the temporal interpolation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.547, + 0.3, + 0.563 + ], + "angle": 0, + "content": "4.4. Visualization Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.483, + 0.841 + ], + "angle": 0, + "content": "Visualization of Fourier bases. To further validate the effectiveness and interpretability of our proposed Fourier bases, we visualize extracted motion bases from the most shadow layer. As revealed in Fig.6, \\( f_{0} - f_{7} \\) refers to the magnitude of the first eight motion bases in ascending order of frequency. Specifically, \\( f_{0} \\) shows a low-frequency motion pattern highlighting no region, which is strongly related to static backgrounds. \\( f_{1} - f_{3} \\) exhibit a large motion intensity in the highlighted right and left ventricle regions. Those regions are indeed dynamic foregrounds requiring special attention. \\( f_{4} \\) depicts an activation in the myocardial region, which contracts inward along the circumferential direction. \\( f_{5} - f_{7} \\) highlight regions with vast deformations in the temporal domain. Particularly, \\( f_{6} \\) illustrates a large motion on boundaries of anatomical structures. In summary, these motion patterns reveal locally anisotropic motion intensities, and are harmonized with specific frequencies of the respiratory process, thus promoting the temporal interpolation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Visualization of temporal anatomical variations. To demonstrate the effectiveness of FB-Diff on temporal simulation, we visualize temporal variation maps for a perceptual comparison. Fig.7 provides interpolated intermediate" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.254, + 0.891, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.513, + 0.907, + 0.542 + ], + "angle": 0, + "content": "Figure 7. Temporal variation comparison between FB-Diff and existing methods with the linear motion hypothesis." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.549, + 0.907, + 0.641 + ], + "angle": 0, + "content": "frames at different timepoints. According to the referenced video, the anatomical motion in the temporal domain is nonlinear, which cannot be precisely simulated by existing methods with the linear motion hypothesis. In contrast, our predictions show better temporal consistency and continuity, particularly for frames at \\( t = 0.6 \\) and 0.8." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.661, + 0.634, + 0.679 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Our proposed FB-Diff resolves the temporal interpolation of 4D medical imaging from a frequency perspective, and can model the respiration-related process with nonlinear and quasi-periodic motions. A Fourier motion operator is devised to extract Fourier bases by incorporating physiological motion priors and case-specific spectral information in the feature space. And the diffusion model further leverages Fourier bases and two prompting frames to achieve the precise interpolation. Extensive results show that FB-Diff achieves SOTA perceptual performance with better temporal consistency while maintaining promising reconstruction metrics. Ultimately, it requires more conditional guidance such as electrocardiogram signals to simulate the highly unstable breathing, which is promising in future work." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28017" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.486, + 0.168 + ], + "angle": 0, + "content": "Acknowledgement. We thank the anonymous reviewers for their thorough and constructive feedbacks. This research was supported in part by the project of the National Natural Science Foundation of China (No. 62376153), in part by the institute of Munich Center of Machine Learning." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.182, + 0.188, + 0.198 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.207, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[1] Juyoung Bae, Elizabeth Tong, and Hao Chen. Conditional diffusion model for versatile temporal inpainting in 4d cerebral ct perfusion imaging. In MICCAI, pages 67-77. Springer, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.264, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[2] Guha Balakrishnan, Amy Zhao, Mert R Sabuncu, John Guttag, and Adrian V Dalca. Voxelmorph: a learning framework for deformable medical image registration. IEEE transactions on medical imaging, 38(8):1788-1800, 2019. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.321, + 0.482, + 0.39 + ], + "angle": 0, + "content": "[3] Wenbo Bao, Wei-Sheng Lai, Xiaoyun Zhang, Zhiyong Gao, and Ming-Hsuan Yang. Memc-net: Motion estimation and motion compensation driven neural network for video interpolation and enhancement. IEEE transactions on pattern analysis and machine intelligence, 43(3):933-948, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.392, + 0.482, + 0.488 + ], + "angle": 0, + "content": "[4] Olivier Bernard, Alain Lalande, Clement Zotti, Frederick Cervenansky, Xin Yang, Pheng-Ann Heng, Irem Cetin, Karim Lekadir, Oscar Camara, Miguel Angel Gonzalez Ballester, et al. Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: is the problem solved? IEEE transactions on medical imaging, 37 (11):2514–2525, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.49, + 0.482, + 0.544 + ], + "angle": 0, + "content": "[5] Rhydian Caines, Naomi K Sisson, and Carl G Rowbottom. 4dct and vmat for lung patients with irregular breathing. Journal of Applied Clinical Medical Physics, 23(1):e13453, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.547, + 0.482, + 0.601 + ], + "angle": 0, + "content": "[6] Junyu Chen, Eric C Frey, Yufan He, William P Segars, Ye Li, and Yong Du. Transmorph: Transformer for unsupervised medical image registration. Medical image analysis, 82:102615, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.604, + 0.482, + 0.66 + ], + "angle": 0, + "content": "[7] Tingxiu Chen, Yilei Shi, Zixuan Zheng, Bingcong Yan, Jingliang Hu, Xiao Xiang Zhu, and Lichao Mou. Ultrasound image-to-video synthesis via latent dynamic diffusion models. In MICCAI, pages 764-774. Springer, 2024. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.661, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[8] Myungsub Choi, Heewon Kim, Bohyung Han, Ning Xu, and Kyoung Mu Lee. Channel attention is all you need for video frame interpolation. In AAAI, pages 10663-10671, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.704, + 0.482, + 0.771 + ], + "angle": 0, + "content": "[9] Myungsub Choi, Suyoung Lee, Heewon Kim, and Kyoung Mu Lee. Motion-aware dynamic architecture for efficient frame interpolation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13839-13848, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[10] Duolikun Danier, Fan Zhang, and David Bull. Ldmvfi: Video frame interpolation with latent diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1472-1480, 2024. 1, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[11] Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick Van Der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In ICCV, pages 2758-2766, 2015. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.207, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "[12] Jan Ehrhardt, Cristian Lorenz, et al. 4D modeling and estimation of respiratory motion for radiation therapy. Springer, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[13] Yuyu Guo, Lei Bi, Euijoon Ahn, Dagan Feng, Qian Wang, and Jinman Kim. A spatiotemporal volumetric interpolation network for 4d dynamic medical image. In CVPR, pages 4726-4735, 2020. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.193, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.293, + 0.905, + 0.333 + ], + "angle": 0, + "content": "[16] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. NeurIPS, 35:8633-8646, 2022. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.335, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[17] Kan N Hor, Rolf Baumann, Gianni Pedrizzetti, Gianni Tonti, William M Gottliebson, Michael Taylor, D Woodrow Benson, and Wojciech Mazur. Magnetic resonance derived myocardial strain assessment using feature tracking. Journal of visualized experiments: JoVE, (48):2356, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.407, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[18] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssim. In 2010 20th international conference on pattern recognition, pages 2366-2369. IEEE, 2010. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.449, + 0.905, + 0.53 + ], + "angle": 0, + "content": "[19] Mengshun Hu, Kui Jiang, Liang Liao, Jing Xiao, Junjun Jiang, and Zheng Wang. Spatial-temporal space hand-in-hand: Spatial-temporal video super-resolution via cycle-projected mutual learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3574-3583, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.533, + 0.907, + 0.601 + ], + "angle": 0, + "content": "[20] Mengshun Hu, Kui Jiang, Zhihang Zhong, Zheng Wang, and Yinqiang Zheng. Iq-vfi: Implicit quadratic motion estimation for video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6410-6419, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.604, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[21] Zhewei Huang, Tianyuan Zhang, Wen Heng, Boxin Shi, and Shuchang Zhou. Real-time intermediate flow estimation for video frame interpolation. In European Conference on Computer Vision, pages 624-642. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.661, + 0.905, + 0.713 + ], + "angle": 0, + "content": "[22] Geoffrey D Hugo, Elisabeth Weiss, William C Sleeman, Salim Balik, Paul J Keall, Jun Lu, and Jeffrey F Williamson. Data from 4d lung imaging of nsclc patients. (No Title), 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.717, + 0.905, + 0.772 + ], + "angle": 0, + "content": "[23] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In CVPR, pages 2462-2470, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.774, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[24] Siddhant Jain, Daniel Watson, Eric Tabellion, Ben Poole, Janne Kontkanen, et al. Video interpolation with diffusion models. In CVPR, pages 7341-7351, 2024. 1, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.817, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[25] Mi-Young Jeung, Philippe Germain, et al. Myocardial tagging with mr imaging: overview of normal and pathologic findings. Radiographics, 32(5):1381-1398, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Huaizu Jiang, Deqing Sun, Varun Jampani, Ming-Hsuan Yang, Erik Learned-Miller, and Jan Kautz. Super slomo: High quality estimation of multiple intermediate frames for" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28018" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.123, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "video interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9000-9008, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.19 + ], + "angle": 0, + "content": "[27] Xin Jin, Longhai Wu, Jie Chen, Youxin Chen, Jayoon Koo, and Cheul-hee Hahm. A unified pyramid recurrent network for video frame interpolation. In CVPR, pages 1578-1587, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.484, + 0.261 + ], + "angle": 0, + "content": "[28] Tarun Kalluri, Deepak Pathak, Manmohan Chandraker, and Du Tran. Flavr: Flow-agnostic video representations for fast frame interpolation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2071–2082, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.483, + 0.305 + ], + "angle": 0, + "content": "[29] Boah Kim and Jong Chul Ye. Diffusion deformable model for 4d temporal medical image generation. In MICCAI, pages 539-548. Springer, 2022. 1, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.483, + 0.362 + ], + "angle": 0, + "content": "[30] JungEun Kim, Hangyul Yoon, Geondo Park, Kyungsu Kim, and Eunho Yang. Data-efficient unsupervised interpolation without any intermediate frame for 4d medical images. In CVPR, pages 11353-11364, 2024. 1, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.363, + 0.483, + 0.446 + ], + "angle": 0, + "content": "[31] Lingtong Kong, Boyuan Jiang, Donghao Luo, Wenqing Chu, Xiaoming Huang, Ying Tai, Chengjie Wang, and Jie Yang. Ifrnet: Intermediate feature refine network for efficient frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1969-1978, 2022. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.483, + 0.504 + ], + "angle": 0, + "content": "[32] Xia Li, Runzhao Yang, Xiangtai Li, Antony Lomax, Ye Zhang, and Joachim Buhmann. Cpt-interp: Continuous spatial and temporal motion modeling for 4d medical image interpolation. arXiv preprint arXiv:2405.15385, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.483, + 0.56 + ], + "angle": 0, + "content": "[33] Zhen Li, Zuo-Liang Zhu, Ling-Hao Han, Qibin Hou, Chun-Le Guo, and Ming-Ming Cheng. Amt: All-pairs multi-field transforms for efficient frame interpolation. In CVPR, pages 9801-9810, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.561, + 0.483, + 0.63 + ], + "angle": 0, + "content": "[34] Hanwen Liang, Yuyang Yin, Dejia Xu, Hanxue Liang, Zhangyang Wang, Konstantinos N Plataniotis, Yao Zhao, and Yunchao Wei. Diffusion4d: Fast spatial-temporal consistent 4d generation via video diffusion models. arXiv preprint arXiv:2405.16645, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.483, + 0.702 + ], + "angle": 0, + "content": "[35] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv preprint arXiv:2402.17177, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.703, + 0.483, + 0.745 + ], + "angle": 0, + "content": "[36] Liying Lu, Ruizheng Wu, Huaijia Lin, Jiangbo Lu, and Jiaya Jia. Video frame interpolation with transformer. In CVPR, pages 3532-3542, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[37] Simone Meyer, Abdelaziz Djerouah, Brian McWilliams, Alexander Sorkine-Hornung, Markus Gross, and Christopher Schroers. Phasenet for video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 498-507, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[38] Kotaro Mizuno and Masahiro Muto. Preoperative evaluation of pleural adhesion in patients with lung tumors using four-dimensional computed tomography performed during natural breathing. *Medicine*, 100(47):e27800, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.483, + 0.903 + ], + "angle": 0, + "content": "[39] Simon Niklaus and Feng Liu. Softmax splatting for video frame interpolation. In Proceedings of the IEEE/CVF con" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "ference on computer vision and pattern recognition, pages 5437-5446, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[40] David Ouyang, Bryan He, Amirata Ghorbani, Neal Yuan, Joseph Ebinger, Curtis P Langlotz, Paul A Heidenreich, Robert A Harrington, David H Liang, Euan A Ashley, et al. Video-based ai for beat-to-beat assessment of cardiac function. Nature, 580(7802):252-256, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[41] Junheum Park, Jintae Kim, and Chang-Su Kim. Biformer: Learning bilateral motion estimation via bilateral transformer for 4k video frame interpolation. In CVPR, pages 1568-1577, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[42] Saikat Roy, Gregor Koehler, Constantin Ulrich, Michael Baumgartner, Jens Petersen, Fabian Isensee, Paul F Jaeger, and Klaus H Maier-Hein. Mednext: transformer-driven scaling of convnets for medical image segmentation. In MICCAI, pages 405–415. Springer, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[43] Otto A Smiseth, Hans Torp, Anders Opdahl, Kristina H Haugaa, and Stig Urheim. Myocardial strain imaging: how useful is it in clinical decision making? European heart journal, 37 (15):1196-1207, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.905, + 0.419 + ], + "angle": 0, + "content": "[44] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In CVPR, pages 8934-8943, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.462 + ], + "angle": 0, + "content": "[45] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, pages 402-419. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.463, + 0.905, + 0.49 + ], + "angle": 0, + "content": "[46] Georgi P Tolstov. Fourier series. Courier Corporation, 2012. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.905, + 0.547 + ], + "angle": 0, + "content": "[47] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.905, + 0.617 + ], + "angle": 0, + "content": "[48] Bingchao Wang, Zhiwei Ning, Jianyu Ding, Xuanang Gao, Yin Li, Dongsheng Jiang, Jie Yang, and Wei Liu. Fix-clip: Dual-branch hierarchical contrastive learning via synthetic captions for better understanding of long text. arXiv preprint arXiv:2507.10095, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.618, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[49] Lu Wang, Shelly Hayes, Kamen Paskalev, Lihui Jin, Mark K Buyyounouski, Charlie C-M Ma, and Steve Feigenberg. Dosimetric comparison of stereotactic body radiotherapy using 4d ct and multiphase ct images for treatment planning of lung cancer: evaluation of the impact on daily dose coverage. *Radiotherapy and Oncology*, 91(3):314–324, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.703, + 0.905, + 0.745 + ], + "angle": 0, + "content": "[50] Yihan Wang, Lahav Lipson, and Jia Deng. Sea-raft: Simple, efficient, accurate raft for optical flow. In ECCV, pages 36-54. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[51] Tzu-Ti Wei, Chin Kuo, Yu-Chee Tseng, and Jen-Jee Chen. Mpvf: 4d medical image inpainting by multi-pyramid voxel flows. IEEE Journal of Biomedical and Health Informatics, 2023. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.859 + ], + "angle": 0, + "content": "[52] Xin You, Ming Ding, Minghui Zhang, Yangqian Wu, Yi Yu, Yun Gu, and Jie Yang. Semantic difference guidance for the uncertain boundary segmentation of ct left atrial appendage. In MICCAI, pages 121-131. Springer, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.903 + ], + "angle": 0, + "content": "[53] Xin You, Junjun He, Jie Yang, and Yun Gu. Learning with explicit shape priors for medical image segmentation. IEEE Transactions on Medical Imaging, 2024. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28019" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[54] Xin You, Yixin Lou, Minghui Zhang, Jie Yang, and Yun Gu. Slord: Structural low-rank descriptors for shape consistency in vertebrae segmentation. IEEE Journal of Biomedical and Health Informatics, 2025. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[55] Xin You, Minghui Zhang, Hanxiao Zhang, Jie Yang, and Nassir Navab. Temporal differential fields for 4d motion modeling via image-to-video synthesis. arXiv preprint arXiv:2505.17333, 2025. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[56] Weihao Yu, Yuanhao Cai, Ruyi Zha, Zhiwen Fan, Chenxin Li, and Yixuan Yuan. X\\(^2\\)-gaussian: 4d radiative gaussian splatting for continuous-time tomographic reconstruction. arXiv preprint arXiv:2503.21779, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[57] Chuyan Zhang, Hao Zheng, Xin You, Yefeng Zheng, and Yun Gu. Pass: test-time prompting to adapt styles and semantic shapes in medical image segmentation. IEEE Transactions on Medical Imaging, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.401 + ], + "angle": 0, + "content": "[58] Guozhen Zhang, Yuhan Zhu, Haonan Wang, Youxin Chen, Gangshan Wu, and Limin Wang. Extracting motion and appearance via inter-frame attention for efficient video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5682-5692, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.403, + 0.482, + 0.457 + ], + "angle": 0, + "content": "[59] Haiyu Zhang, Xinyuan Chen, Yaohui Wang, Xihui Liu, Yunhong Wang, and Yu Qiao. 4diffusion: Multi-view video diffusion model for 4d generation. NeurIPS, 37:15272-15295, 2025. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.46, + 0.482, + 0.514 + ], + "angle": 0, + "content": "[60] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.517, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[61] Zhihang Zhong, Gurunandan Krishnan, Xiao Sun, Yu Qiao, Sizhuo Ma, and Jian Wang. Clearer frames, anytime: Resolving velocity ambiguity in video frame interpolation. In ECCV, pages 346-363. Springer, 2024. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.574, + 0.482, + 0.627 + ], + "angle": 0, + "content": "[62] Kun Zhou, Wenbo Li, Xiaoguang Han, and Jiangbo Lu. Exploring motion ambiguity and alignment for high-quality video frame interpolation. In CVPR, pages 22169-22179, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.627 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28020" + } + ] +] \ No newline at end of file diff --git a/2025/FB-Diff_ Fourier Basis-guided Diffusion for Temporal Interpolation of 4D Medical Imaging/c8ad7765-483e-400a-90bd-72090b0597d0_origin.pdf b/2025/FB-Diff_ Fourier Basis-guided Diffusion for Temporal Interpolation of 4D Medical Imaging/c8ad7765-483e-400a-90bd-72090b0597d0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e3ef78af4d3225474693827ae7c67fde9c69e850 --- /dev/null +++ b/2025/FB-Diff_ Fourier Basis-guided Diffusion for Temporal Interpolation of 4D Medical Imaging/c8ad7765-483e-400a-90bd-72090b0597d0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:653558f2b24b5d0c568508bdf144a2826f57693e44313fcade3f65fe9d99a214 +size 20055998 diff --git a/2025/FB-Diff_ Fourier Basis-guided Diffusion for Temporal Interpolation of 4D Medical Imaging/full.md b/2025/FB-Diff_ Fourier Basis-guided Diffusion for Temporal Interpolation of 4D Medical Imaging/full.md new file mode 100644 index 0000000000000000000000000000000000000000..745aa82b90b89574540d18b55a0869d7f4be0a20 --- /dev/null +++ b/2025/FB-Diff_ Fourier Basis-guided Diffusion for Temporal Interpolation of 4D Medical Imaging/full.md @@ -0,0 +1,285 @@ +# FB-Diff: Fourier Basis-guided Diffusion for Temporal Interpolation of 4D Medical Imaging + +Xin You $^{1,2,3}$ Runze Yang $^{1}$ Chuyan Zhang $^{1}$ Zhongliang Jiang $^{2}$ Jie Yang $^{1,*}$ Nassir Navab $^{2,*}$ + +$^{1}$ Institute of Medical Robotics, Shanghai Jiao Tong University, Shanghai, China + +$^{2}$ Computer Aided Medical Procedures, Technical University of Munich, Munich, Germany + +3 Munich Center for Machine Learning, Munich, Germany + +xin.you@tum.de, nassir.navab@tum.de + +# Abstract + +The temporal interpolation task for 4D medical imaging, plays a crucial role in clinical practice of respiratory motion modeling. Following the simplified linear-motion hypothesis, existing approaches adopt optical flow-based models to interpolate intermediate frames. However, realistic respiratory motions should be nonlinear and quasi-periodic with specific frequencies. Intuited by this property, we resolve the temporal interpolation task from the frequency perspective, and propose a Fourier Basis-guided Diffusion model, termed FB-Diff. Specifically, due to the regular motion discipline of respiration, physiological motion priors are introduced to describe general characteristics of temporal data distributions. Then a Fourier motion operator is elaborately devised to extract Fourier bases by incorporating physiological motion priors and case-specific spectral information in the feature space of Variational Autoencoder. Well-learned Fourier bases can better simulate respiratory motions with motion patterns of specific frequencies. Conditioned on starting and ending frames, the diffusion model further leverages well-learned Fourier bases via the basis interaction operator, which promotes the temporal interpolation task in a generative manner. Extensive results demonstrate that FB-Diff achieves state-of-the-art (SOTA) perceptual performance with better temporal consistency while maintaining promising reconstruction metrics. Codes are available here. + +# 1. Introduction + +4D medical images, which depict 3D volumes with temporal variations, are essential in clinical practice for capturing dynamic changes and monitoring the progression of diseases over time [12, 25, 49, 52]. A significant application is the temporal modeling for breathing-induced motions of + +cardiac or pulmonary anatomical structures. That task aims to enrich motion visualizations by generating intermediate frames under the condition of starting and ending frames, which is consistent with the concept of Video Frame Interpolation (VFI) in natural scenarios. + +VFI has been thoroughly investigated through systematic studies in recent years [10, 24, 31, 37, 39, 41, 58, 61, 62]. However, due to unique properties and constraints of medical imaging, it remains a challenge to directly implement these frameworks into 4D medical video interpolation with maintaining equally promising performance. Specifically, the quality of ground truth intermediate frames in medical imaging is often compromised due to factors such as imaging noise, unstable breathing, causing unexpected artifacts [17, 25, 43, 49]. Thus, the lower image quality compared to videos in the natural domain, will degrade the interpolation performance of VFI models. Besides, potential patients' movement might induce an additional disturbance to the temporal modeling [5, 38]. + +Furthermore, there exists a discrepancy on the type of motions between natural and medical scenes. The former focuses on the object movement and illumination intensity change. While in the medical domain, motions can be depicted as subtle anatomical variance with quasi-periodic motion discipline [51, 56]. Thus, a specific motion prior of anatomies is effective to boost the temporal interpolation, especially for the medical data with limited resources from a small pool of individuals. + +On account of these challenges, we present the following question: "How can we devise a specialist VFI model tailored for 4D temporal medical volumes?" Existing methods [2, 13, 29, 30, 51] on medical VFI face a major shortcoming: the linear motion hypothesis in a breathing period is required for the regression of intermediate frames as revealed in Fig.1(a). However, respiratory motions should be a nonlinear oscillatory process that cannot be simply mod- + +![](images/49b55a0a073f7468310942ce3ca259cec8925fa39bd51feff78d4aec58660ba8.jpg) +(b) Nonlinear & quasi-periodic motion $\Longrightarrow$ Our model +Figure 1. Temporal Interpolation for respiratory motions. (a) Simplified linear-motion hypothesis: interpolating intermediate frames by regressing optical flows with linear variations. (b) Nonlinear & quasi-periodic motions: leveraging diffusion models conditioned on Fourier motion bases, which represent various motion patterns indicating nonlinear and locally inconsistent anatomical variations. + +eled with a linear equation [32, 51]. Also, these methods fail to recognize that respiratory motions in heart or lung, are indeed a quasi-periodic motion process with specific frequencies by extending the timeline. Intuited by this, we resolve the respiratory motion simulation from a frequency perspective. In Fig.1(b), each medical volume in the temporal domain can be viewed as a specific value in the periodic curve. That periodic curve can be decomposed as a group of sine and cosine bases with multiple frequencies [46]. Comparatively, there exists a group of Fourier bases, which are capable of reconstructing original temporal sequences. Here Fourier bases with different frequencies refer to various motion patterns, which indicate nonlinear and locally anisotropic variations of anatomical structures. In our work, Fourier bases are acquired by the Fast Fourier Transform (FFT) on the temporal domain. + +In this work, 4D data distributions of the respiratory process depict a regular motion discipline, with temporal variations from inhaling to exhaling, or from diastole to systole. Motivated by that observation, physiology motion priors are introduced to encode generic characteristics of the respiratory process. Then a Fourier Motion Operator $O_{FM}$ is elaborately devised to yield enhanced motion bases by incorporating physiology motion priors and case-specific spectral information in the feature space. Specifically, A Variational Autoencoder (VAE) is introduced to map the video space into the feature space. Physiology motion priors are defined as learnable frequency embeddings, which can be updated by pretraining a Variational Autoen + +coder (VAE) on given temporal data. And case-specific frequency information can be extracted by FFT in the temporal domain. Well-learned physiology motion priors can enrich Fourier bases with generic properties of respiratory motions. Then enhanced Fourier motion bases can better simulate the quasi-periodic respiratory process with motion patterns of specific frequencies. On the grounds that diffusion models show promising performance in generating realistic videos [15, 16, 35, 48], a Markov chain process is learned to transform Gaussian noise into target videos by the proposed Fourier Bases-guided Diffusion, termed FB-Diff. Since only starting and ending frames are available during inference, they serve as conditional prompts for intermediate frame interpolation. Meanwhile, we devise the basis interaction operator to inject enhanced Fourier bases into the conditional diffusion model. Then, FB-Diff can well address the temporal interpolation task in a generative manner. + +Our proposed method has achieved state-of-the-art (SOTA) interpolation performance compared with other methods for 4D medical imaging, particularly on the mid-most frames. For quantitative results, more comprehensive evaluation metrics are adopted in this work, including traditional reconstruction metrics and feature-level perceptual metrics. FB-Diff not only presents promising reconstruction metrics but also superior perceptual results. Qualitative visualizations reveal better temporal consistency by FB-Diff compared with other linear-motion-based methods. Our main contributions are summarized as follows: + +![](images/ea0242ad2ded6b4b4995d4b5240260ecbefd64127560851e8856bf2d9f92bc00.jpg) +Figure 2. The pipeline of FB-Diff. (a) The Fourier motion operator is proposed to extract Fourier bases in a feature space of VAE. Bases consist of learnable generic motion priors and case-specific frequency information. (b) Fourier bases are first adapted to the masked video domain under the guidance of well-learned motion priors and pretrained VAE weights. Domain-adapted bases serve as the conditional input to boost the video synthesis by diffusion models. + +- We propose a diffusion model termed FB-Diff for the temporal interpolation of 4D medical imaging. FB-Diff leverages Fourier motion bases, which indicate motion patterns with specific frequencies to resolve the respiratory modeling with nonlinear and quasi-periodic motions. +- We introduce the Fourier motion operator $O_{FM}$ to extract Fourier bases, containing physiology motion priors and case-specific frequency representations. +- We devise the basis interaction operator to inject Fourier bases into the diffusion model. Fourier bases can boost the video synthesis for temporal interpolation. +- Extensive results demonstrate that FB-Diff reveals SOTA perceptual performance with better temporal consistency while maintaining promising reconstruction metrics. + +# 2. Related Work + +Video Frame Interpolation. Popular VFI techniques are mainly classified into motion-free and motion-based approaches, depending on whether optical flows are incorporated or not. Motion-based: These approaches warp + +input frames forward or backward based on optical flows acquired by either off-the-shelf networks [11, 23, 44, 45, 50] or custom-designed flow estimators [20, 21, 27, 31, 33, 58]. Warped intermediate frames are then refined by delicately designed networks [3, 26] to enhance visual quality. Recently, Transformer-based motion estimators are proposed to raise the upper limit of VFI performance [36, 41, 58] Motion-Free: Motion-free methods rely on implicit spatiotemporal modeling [8, 9, 19, 28] to generate the intermediate frame. However, due to a limited quality of 4D medical videos and quasi-periodic motion priors, existing VFI frameworks are not feasible. That is why we aim to propose a specialist VFI model for 4D medical sequences. + +4D Medical VFI. Existing methods on this topic mainly leverage volumetric motions for 4D temporal interpolation. Specifically, SVIN [13] estimates forward and backward deformation fields, which will yield precise intermediate frames by linearly combining bidirectional information. MPVF [51] resolves various magnitudes of motions by proposing a multi-pyramid voxel flows model that takes + +multi-scale voxel flows into account. DDM [29] can learn spatial deformation information between the source and target volumes and provide a latent code for generating intermediate frames along a geodesic path. UVI-Net [30] utilizes the flow calculation model with the time-domain cycle-consistency constraint and linear motion hypothesis, to realize motion modeling in an unsupervised style. However, the linear hypothesis for motions will induce temporal inconsistency and spatial distortion. + +Diffusion-based Video Interpolation. Denoising diffusion probabilistic models [15, 16, 35] reveal promising performance in generating realistic images or videos. Some methods maximize the potential of diffusion models to resolve video interpolation tasks. Essentially, these approaches implicitly learn the temporal motions from the starting frame to the ending frame [1, 7, 24]. LDMVFI [10] proposes a latent diffusion model, which approaches the VFI problem from a generative perspective by formulating it as a conditional synthesis issue. However, these diffusion models do not leverage effective conditional guidance, which is crucial for medical VFI with intrinsic motion priors. Thus, we introduce Fourier bases to enhance the conditional synthesis. + +# 3. Methodology + +# 3.1. Fourier Motion Bases + +Since respiratory motions for heart or lung exhibit a quasi-periodic characteristic, we intend to resolve this task from a frequency perspective. Specifically, the respiration can be viewed as an oscillatory process with certain frequencies. Thus, we draw an analogy between 2D time-varying curves and 4D temporal medical volumes. Curves can be decomposed into a group of harmonic oscillators with various frequencies [46]. Similarly, temporal volumes can be restored via a stack of extracted Fourier bases, which are embodied with various motion patterns of anatomical deformation. The whole process can be formulated as: + +$$ +\mathcal {B} _ {k} = \sum_ {n = 0} ^ {N - 1} V _ {n} e ^ {- i 2 \pi k \frac {n}{N}}, k = 0, 1, \dots , N - 1, \tag {1} +$$ + +$$ +V _ {n} = \frac {1}{T} \sum_ {k = 0} ^ {N - 1} \mathcal {B} _ {k} e ^ {i 2 \pi n \frac {k}{N}}, n = 0, 1, \dots , N - 1, \tag {2} +$$ + +where $\mathcal{B}_k$ means the $k_{th}$ Fourier basis in the frequency domain, $V_{n}$ refers to the $n_{th}$ medical volume, and $N$ is the frame number. These Fourier bases describe locally anisotropic anatomical variations with different motion intensities, promoting the performance of the temporal interpolation model $\mathcal{I}$ . Thus, the goal of our work is to deduce 4D medical videos $V$ conditioned on two prompting frames $V_{0}, V_{N-1}$ and Fourier motion bases $\mathcal{B}$ : + +$$ +\mathcal {I} ^ {*} = \underset {\mathcal {I}} {\arg \min } \| \boldsymbol {V} - \mathcal {I} \left(V _ {0}, V _ {N - 1}; \mathcal {B}\right) \| _ {\mathcal {D}}, \tag {3} +$$ + +![](images/3a5fb39ec16e2282dd96e4173cb949fd02dec76d57657897bdfbad4585e8955a.jpg) +Figure 3. The details of Fourier motion operator. Fourier bases are decomposed as physiology motion priors and case-specific Fourier features, corresponding to phase and magnitude components. + +$\mathcal{I}^*$ refers to the optimal solution in the interpolation model set, and $\mathcal{D}$ represents the distance-based normed space. We systematically organize the following sections along Fourier bases. Section 3.2 introduces the design of the Fourier Motion Operator, which constructs Fourier bases with physiology motion priors and case-specific frequency information. Section 3.3 adapts Fourier bases to the to-be-interpolated video domain with well-learned physiology motion priors. Section 3.4 provides details of the diffusion model conditioned on Fourier bases. + +# 3.2. Fourier Motion Operator + +To explicitly extract the Fourier bases with various motion patterns for the respiration process, we introduce the Fourier Operator $O_{\mathcal{F}}$ as a neural operator. As stated in Eq.(2), $O_{\mathcal{F}}$ will operate in the temporal domain, not the spatial domain. Since the common Fourier operator $O_{\mathcal{F}}$ only mines case-specific frequency representations, general physiology characteristics inside respiratory data are ignored. Inspired by that notion, we devise the Fourier motion operator $O_{FM}$ as a neural operator in the feature space. To map the image space into feature space, we introduce a Variational Autoencoder (VAE) as revealed in Fig.2(a). Due to no built-in 4D convolutional operator, we replace it with 3D spatial convolution and 1D temporal convolution as [34, 59] did. The spatial layer in Fig.2(a) consists of spatial convolution, GELU, and group normalization operations. + +The Fourier motion operator $O_{FM}$ is devised to acquire spectral characteristics with both physiology motion priors and case-specific representations in the feature space. Firstly, physiology motion priors are introduced as learnable frequency embeddings, which are updated by backward gradients of restoring given temporal data via VAE. Well-trained frequency embeddings $\theta$ can represent the generic discipline of the regular respiration process [54], with $\theta_{k}$ referring to the $k_{th}$ element $(k = 0,1,\dots,N - 1)$ . Then Fourier bases $\mathcal{B}$ are made up of physiology motion priors $\theta$ and case-specific features $F$ , which arise from the output of + +spatial and temporal layers. As depicted by Fig.3, $\mathcal{B}$ with a complex type is yielded in the following style: + +$$ +\mathcal {B} = \underbrace {\mathcal {M} (\left\| O _ {\mathcal {F}} (F) \right\| _ {2})} _ {\text {C a s e - S p e c i f i c}} \cdot \underbrace {e ^ {i U (\tau (\theta))}} _ {\text {G e n e r a l}}, \tag {4} +$$ + +$\tau$ means the channel transform, $U$ refers to the upsampling which unifies the resolution of $\theta$ and $\pmb{F}$ . These operations transform general motion priors into the phase part of Fourier bases, which determines the motion variation standing for the whole dataset distribution. Then $\| \cdot \| _2$ means the $L_{2}$ norm, $\mathcal{M}$ is the convolutional mapping operation. They will output motion bases' magnitude information, which selectively highlights regions in spectral maps relevant to case-specific information. + +Meanwhile, feature maps $F$ are further enriched with more fine-grained frequency details. The specific mechanism can be described with Eq.(5): + +$$ +\hat {\boldsymbol {F}} = \boldsymbol {F} + \sigma \mathcal {O} _ {\mathcal {F}} ^ {- 1} (\mathcal {B}), \tag {5} +$$ + +here $O_{\mathcal{F}}^{-1}$ and $\sigma$ refer to the inverse Fourier transform and Sigmoid operators, which will boost deep features $\hat{F}$ with enriched details, including shape and texture information. + +# 3.3. Frame Inpainting-based Basis Synthesis + +Due to the fact that only starting and ending frames can be acquired during inference, a significant step is to adapt Fourier bases to the masked video domain. Motivated by the fact that strong correlations exist between Fourier motion bases and inpainted temporal sequences, we plan to jointly optimize these two variables. To accomplish that goal, we repurpose well-calibrated physiology motion priors from VAE, which prove to be effective in providing the motion guidance for intermediate frame inpainting. Also, pretrained weights of VAE encode implicit temporal dependencies between frames, enhancing the reconstruction of masked video sequences via the finetuned VAE noted as $\mathcal{G}$ . A mathematical formulation is defined as: + +$$ +\hat {V}, \hat {\mathcal {B}} = \mathcal {G} (\hat {\boldsymbol {\theta}}, V _ {\boldsymbol {M}}) = \mathcal {G} (\hat {\boldsymbol {\theta}}, V _ {0},.., \mathbf {0},.., V _ {N - 1}), \tag {6} +$$ + +$\hat{\theta}$ and $V_{M}$ correspond to well-trained general motions and zero-masked videos, $\hat{V}$ and $\hat{B}$ refer to inpainted videos and domain-adapted Fourier bases. + +The interaction between temporal video representations and $\hat{\mathcal{B}}$ is implemented via the Fourier motion operator as shown in Fig.2(b). For the joint optimization process, video sequences will contribute temporal variations, enhancing the synthesis of adapted Fourier bases. Conversely, Fourier bases for zero-masked inputs inherently generate a spectral map that emphasizes regions with motion in the temporal axis. This activated map aids in temporal frame inpainting by exerting attention on regions with different magnitudes. + +![](images/e46e1aa2636d3a69e2489f0e0819f6dad5bd9bb3792fe805fd5471e932837caa.jpg) +Figure 4. The architecture of basis interaction operator. Fourier motion bases selectively enhance video representations by measuring the motion intensity of different regions. Then those enhanced temporal representations can enrich denoised embeddings from the diffusion model. + +# 3.4. Basis-guided 4D Conditional Diffusion + +VAE in Fig.2(a) is not only devised for universal motion priors, but also for the latent diffusion model (LDM), which shows better compatibility for extra conditional inputs. Besides, LDM synthesizes latent embeddings occupying less GPU memory compared to 4D temporal volumes. And the VAE decoder is leveraged to reconstruct interpolated video sequences with synthesized latent embeddings. + +For the specific diffusion process, the whole video sequence $V$ is first encoded as a latent embedding $z_0 = \mathcal{G}_e(V)$ . Then the latent code $z_0$ is perturbed as: + +$$ +z _ {t} = \sqrt {\bar {\alpha} _ {t}} z _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon , \epsilon \sim \mathcal {N} (0, 1), \tag {7} +$$ + +where $\overline{\alpha}_t = \prod_{i=1}^t (1 - \beta_t)$ with $\beta_t$ is the noise coefficient at time step $t$ , and $t$ is uniformly sampled from the timestep index set $[1,2,\dots,T]$ . This process can be regarded as a Markov chain, which incrementally adds Gaussian noise to the latent code $z_0$ . The denoising model $d$ receives $z_t$ as input and is optimized under conditional guidance to learn the latent space distribution with the objective function: + +$$ +\mathcal {L} _ {\epsilon} = E _ {z _ {t}, \epsilon \sim \mathcal {N} (0, 1)} \| \epsilon - \epsilon_ {d} \left(z _ {t}, t; \hat {\mathcal {B}}, \mathcal {G} _ {e} \left(\boldsymbol {V} _ {\boldsymbol {M}}\right), \hat {\boldsymbol {V}}\right) \| ^ {2}. \tag {8} +$$ + +here $\epsilon_d$ is the regressed noise, adapted Fourier bases $\hat{\mathcal{B}}$ and inpainted coarse videos $\hat{V}$ are incorporated as the conditional guidance. Moreover, feature maps of starting and ending frames by the VAE encoder $\mathcal{G}_e$ serve as a strong prompt for the 4D video synthesis. Then during the training process of FB-Diff, the total loss $\mathcal{L}$ consists of the denoising loss $\mathcal{L}_{\epsilon}$ and the $L_{2}$ -norm reconstruction loss $\mathcal{L}_r(\hat{V},V)$ . + +To inject Fourier motion bases into LDM and harness coarsely inpainted results, we design the Basis Interaction Operator (BIO) as shown in Fig.4. Specifically, inpainted videos are mapped into video feature representations via spatial & temporal layers and the downsampling operation. Fourier motion bases with various motion patterns, selectively enhance video representations by measuring the motion intensity of different regions. Those enhanced temporal representations can enrich denoised embeddings from + +Table 1. Baseline comparison with flow-based and diffusion-based models. (Bold: the best, Underlined: the second best. Δ values represent the relative difference between our model and UVI-Net, and red and blue numbers stand for better and worse results). + +
ModelACDC Cardiac4D Lung
PSNR (dB) ↑LPIPS ↓FID ↓FVD ↓PSNR (dB) ↑LPIPS ↓FID ↓FVD ↓
SVIN [13]31.43±0.4211.563±0.20626.393.630.49±0.3042.650±0.24538.5125.6
Voxelmorph [2]30.77±0.5021.969±0.19729.4102.129.90±0.3732.815±0.26043.2149.0
Transmorph [6]29.81±0.5322.136±0.25531.5109.529.04±0.4032.946±0.29140.5143.5
MPVF [51]31.53±0.4101.594±0.20124.891.330.76±0.2652.418±0.24639.4120.3
UVI-Net [30]32.16±0.4021.662±0.24523.094.231.57±0.3112.211±0.21637.8121.7
IFRNet [31]30.58±0.4081.752±0.24927.897.530.05±0.2802.316±0.27339.7124.0
LDMVFI [10]27.11±0.4602.943±0.41028.899.226.31±0.4533.659±0.34142.1142.7
LDDM [7]24.53±0.4812.634±0.30833.4105.725.19±0.2732.914±0.28748.7146.3
Conditional Diff [16]26.59±0.5452.460±0.35730.195.725.95±0.3913.294±0.37545.0133.0
DDM [29]29.79±0.5042.089±0.35227.4110.329.67±0.4202.705±0.33040.8165.5
FB-Diff30.95±0.3861.452±0.20521.588.930.18±0.2872.066±0.21537.7113.6
Δ value1.21 ↓0.210 ↓1.5 ↓5.3 ↓1.39 ↓0.145 ↓0.1 ↓8.1 ↓
+ +the diffusion model, promoting a promising synthesis of highlighted regions. Thus, BIO can boost well-interpolated videos with spatial and temporal consistency. + +# 4. Experiments + +# 4.1. Experimental Settings + +Dataset. To evaluate the interpolation performance of our framework, we conduct experiments on public ACDC cardiac [4] and 4D-Lung datasets [22]. For ACDC with the MRI modality, volume sequences between the end-diastolic and end-systolic phases are extracted as valid 4D data. Of all the 150 4D sequences, cases with identity 1-100, 101-120, and 121-150 serve as the training, validation, and testing sets. Besides, all MRI volumes are cropped to $128 \times 128 \times 32$ , and the frame number ranges from 6 to 16 [53, 57]. For 4D-Lung CT images, the end-inspiratory and end-expiratory scans are set as the initial and final images. We collect 125 4D videos from [22], with the resolution equal to $6 \times 128 \times 128 \times 128$ . They are split as $80 / 15 / 30$ cases for training, validation, and inference. We uniform the frame number as 16 with zero paddings. + +**Evaluations.** We select quantitative evaluation metrics including Peak Signal-to-Noise Ratio (PSNR) [18], Learned Perceptual Image Patch Similarity (LPIPS) [55, 60], Fréchet Inception Distance (FID) [14], and Fréchet Video Distance (FVD) [47]. Compared to PSNR, which is the traditional pixel-wise reconstruction metric, LPIPS and FVD evaluate perceptual quality from the perspective of human vision [10, 24, 61], which is more feasible and favored in clinical practice. Specifically, LPIPS reveals important perceptual similarity like human vision. FID and FVD evaluate the dataset distribution bias between realistic and synthesized data, including spatial and temporal consistency. + +Training Details. The VAE maps the image space into + +the downsampled latent space with a ratio of $1/8$ . And the proposed conditional diffusion model is implemented with a lightweight UNet. All models are trained using the AdamW optimizer with the linear warm-up strategy. The batch size, diffusion step $T$ , and training epoch are equal to 2, 1000, and 500. Experiments are implemented based on Pytorch and 2 NVIDIA RTX 4090 GPUs. More details about the network and training settings are described in the Appendix. + +# 4.2. Comparisons with SOTA baselines + +Quantitative Comparison. We comprehensively conduct experiments on ACDC cardiac and 4D Lung datasets. Quantitative results are illustrated in Tab.1. Our model achieves the best performance across the benchmark on perceptual evaluation metrics from the perspective of human vision [10, 61], which are clinically favored compared with the pixel-centric reconstruction metric [7]. Specifically, FB-Diff outperforms UVI-Net [30] with $0.210 \downarrow$ LPIPS and $5.3 \downarrow$ FVD values on ACDC. Also, our model maintains quite satisfactory PSNR values of $30.95\mathrm{dB}$ and $30.18\mathrm{dB}$ on ACDC and 4D Lung. Although diffusion models are not sensitive to pixel-wise intensities [15], the conditional guidance of Fourier bases and coarsely inpainted videos is effective to boost the interpolation performance of diffusion models. Tab.1 reveals that FB-Diff significantly surpasses existing diffusion models on all metrics. + +Furthermore, the interpolation process becomes most challenging for the midmost frames since they are positioned farthest from the two reference frames [24]. And our model can better resolve this issue due to stronger temporal representations based on quasi-periodic motion modeling. Thus, a quantitative evaluation of central frames is necessary. As shown in Tab.2, FB-Diff demonstrates greater performance advantages, with $4.3\downarrow$ FID and $8.2\downarrow$ FVD values compared with UVI-Net. Additionally, FB-Diff acquires a competi + +![](images/8da4c198ac582ec7f985250dd8520efe4944f7261aca5a0b3ecc632bc930d0c8.jpg) +Figure 5. Qualitative comparison of the midmost frame interpolation between different models. VM: Voxelmorph. + +Table 2. Interpolation performance of the midmost frames on ACDC cardiac. The choice of midmost frames: if the frame number $N$ is odd, three intermediate frames are selected, otherwise only two frames are chosen. + +
ModelPSNR (dB) ↑LPIPS ↓FID ↓FVD ↓
SVIN [13]26.01±0.4672.673±0.36138.9142.8
DDM [29]24.37±0.5112.756±0.33041.5145.1
UVI-Net [30]26.25±0.4102.542±0.37437.0137.1
FB-Diff26.18±0.4132.287±0.30532.7128.9
Δ value0.07 ↓0.255 ↓4.3 ↓8.2 ↓
+ +tive PSNR value, which demonstrates the effectiveness of respiration-related modeling via Fourier bases. + +Qualitative Comparison. Qualitative results on the interpolated midmost frame are depicted in Fig.5. SVIN [13], Voxelmorph [2], and DDM [29] leverage the single-direction flow estimation to interpolate intermediate frames with linear motions. Compared with that, UVI-Net [30] constructs the interpolation model with time-domain cycle-consistency, and achieves better reconstruction performance as expected. However, these methods essentially overlook the nonlinear and quasi-periodic motion discipline between temporal sequences, resulting in unsatisfactory interpolated results. Instead, FB-Diff exhibits advantageous predictions for midmost frames under the guidance of various motion patterns, indicating locally anisotropic motions with different intensities. Besides, other diffusion models do not work well due to the lack of fine-grained conditional guidance. + +# 4.3. Ablation Study + +Key Components. An ablation study is conducted to investigate the significance of various components in FB-Diff as shown in Tab.3. (a) Physiology Motion Priors $\theta$ . By + +Table 3. Ablation study on key components. $\theta$ : physiology motion priors, $\hat{B}$ : Fourier motion bases, $\hat{V}$ : coarsely inpainted videos. + +
SettingsACDC Cardiac4D Lung
θˆBˆVPSNR (dB) ↑LPIPS ↓PSNR (dB) ↑LPIPS ↓
X✓✓29.761.71029.532.413
XX✓28.172.10528.402.971
✓✓X30.291.65529.712.246
✓✓✓30.951.45230.182.066
+ +removing $\theta$ from the Fourier motion operator, interpolation performance decreases significantly due to the loss of general spectral characteristics from respiratory data. Therefore, Fourier bases lack the precise phase information. (b) Fourier bases $\hat{\mathcal{B}}$ . Without $\hat{\mathcal{B}}$ as conditional guidance, the conditional diffusion model fails to perceive explicit motion patterns with different frequencies, which provide the intensity of anisotropic motions. As a result, the generative model cannot well learn the regular temporal variations, with 2.78 dB $\downarrow$ and 1.78 dB $\downarrow$ PSNR values on ACDC and 4D Lung. (c) Inpainted Videos $\hat{V}$ . Here the coarsely reconstructed sequences serve as the output by the finedtuned VAE as revealed in Fig.2. And the joint optimization process boosts the spatial and temporal representations of $\hat{V}$ . Consequently, discarding this component will reduce fine-grained details of the diffusion model's output. + +Hyper-settings of Fourier Bases. (a) The number of Fourier Bases $N_{Fb}$ . We first analyze the quantitative effects of the frequency number on the interpolation performance. As revealed in Tab.4, reducing the frequency number will cause degraded performance of the temporal simulation. Specifically, 1.30 dB lower PSNR and 0.535 higher LPIPS values are induced by decreasing $N_{Fb}$ from 16 to 8. That phenomenon indicates that some elements in the group of + +![](images/c70cbbc9d0193a624785ed0adda338f619441ab48549cd94deb117f6ac5a06fa.jpg) +Figure 6. Visualizations on the frequency spectral of Fourier bases. $V_{0}$ , $V_{N - 1}$ : starting and ending frames. $f_{1} - f_{8}$ refer to different motion patterns, which highlight different regions under different frequency. There also appear locally anisotropic motion intensities inside a specific motion pattern. + +Table 4. Ablation study on the hyper-settings of Fourier bases. $N_{Fb}$ : the frequency number of Fourier bases. + +
SettingsACDC Cardiac4D Lung
PSNR (dB) ↑LPIPS ↓PSNR (dB) ↑LPIPS ↓
\(N_{Fb}\)=16 (Ours)30.951.45230.182.066
\(N_{Fb}\)=1230.761.64029.972.570
\(N_{Fb}\)=829.651.98729.502.745
w/o Real part29.181.87229.562.453
w/o Imaginary part29.881.68129.672.401
+ +bases with the frequency ranging from $f_{8}$ to $f_{15}$ are consistent with specific frequencies of the respiratory process. (b) Real & Imaginary Parts. $\hat{\mathcal{B}}$ bears real and imaginary parts after combining generic priors and case-specific spectral representations, which perform as the phase and magnitude information. Eliminating either the real or imaginary part will convert the complex data type into the real one, resulting in the loss of phase information and influencing the temporal interpolation. + +# 4.4. Visualization Analysis + +Visualization of Fourier bases. To further validate the effectiveness and interpretability of our proposed Fourier bases, we visualize extracted motion bases from the most shadow layer. As revealed in Fig.6, $f_{0} - f_{7}$ refers to the magnitude of the first eight motion bases in ascending order of frequency. Specifically, $f_{0}$ shows a low-frequency motion pattern highlighting no region, which is strongly related to static backgrounds. $f_{1} - f_{3}$ exhibit a large motion intensity in the highlighted right and left ventricle regions. Those regions are indeed dynamic foregrounds requiring special attention. $f_{4}$ depicts an activation in the myocardial region, which contracts inward along the circumferential direction. $f_{5} - f_{7}$ highlight regions with vast deformations in the temporal domain. Particularly, $f_{6}$ illustrates a large motion on boundaries of anatomical structures. In summary, these motion patterns reveal locally anisotropic motion intensities, and are harmonized with specific frequencies of the respiratory process, thus promoting the temporal interpolation. + +Visualization of temporal anatomical variations. To demonstrate the effectiveness of FB-Diff on temporal simulation, we visualize temporal variation maps for a perceptual comparison. Fig.7 provides interpolated intermediate + +![](images/bfd652572d8a4fb20e6f20ddca808de4f098967ae86b12ade2703c4b76e21236.jpg) +Figure 7. Temporal variation comparison between FB-Diff and existing methods with the linear motion hypothesis. + +frames at different timepoints. According to the referenced video, the anatomical motion in the temporal domain is nonlinear, which cannot be precisely simulated by existing methods with the linear motion hypothesis. In contrast, our predictions show better temporal consistency and continuity, particularly for frames at $t = 0.6$ and 0.8. + +# 5. Conclusion + +Our proposed FB-Diff resolves the temporal interpolation of 4D medical imaging from a frequency perspective, and can model the respiration-related process with nonlinear and quasi-periodic motions. A Fourier motion operator is devised to extract Fourier bases by incorporating physiological motion priors and case-specific spectral information in the feature space. And the diffusion model further leverages Fourier bases and two prompting frames to achieve the precise interpolation. Extensive results show that FB-Diff achieves SOTA perceptual performance with better temporal consistency while maintaining promising reconstruction metrics. Ultimately, it requires more conditional guidance such as electrocardiogram signals to simulate the highly unstable breathing, which is promising in future work. + +Acknowledgement. We thank the anonymous reviewers for their thorough and constructive feedbacks. This research was supported in part by the project of the National Natural Science Foundation of China (No. 62376153), in part by the institute of Munich Center of Machine Learning. + +# References + +[1] Juyoung Bae, Elizabeth Tong, and Hao Chen. Conditional diffusion model for versatile temporal inpainting in 4d cerebral ct perfusion imaging. In MICCAI, pages 67-77. Springer, 2024. 4 +[2] Guha Balakrishnan, Amy Zhao, Mert R Sabuncu, John Guttag, and Adrian V Dalca. Voxelmorph: a learning framework for deformable medical image registration. IEEE transactions on medical imaging, 38(8):1788-1800, 2019. 1, 6, 7 +[3] Wenbo Bao, Wei-Sheng Lai, Xiaoyun Zhang, Zhiyong Gao, and Ming-Hsuan Yang. Memc-net: Motion estimation and motion compensation driven neural network for video interpolation and enhancement. IEEE transactions on pattern analysis and machine intelligence, 43(3):933-948, 2019. 3 +[4] Olivier Bernard, Alain Lalande, Clement Zotti, Frederick Cervenansky, Xin Yang, Pheng-Ann Heng, Irem Cetin, Karim Lekadir, Oscar Camara, Miguel Angel Gonzalez Ballester, et al. Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: is the problem solved? IEEE transactions on medical imaging, 37 (11):2514–2525, 2018. 6 +[5] Rhydian Caines, Naomi K Sisson, and Carl G Rowbottom. 4dct and vmat for lung patients with irregular breathing. Journal of Applied Clinical Medical Physics, 23(1):e13453, 2022. 1 +[6] Junyu Chen, Eric C Frey, Yufan He, William P Segars, Ye Li, and Yong Du. Transmorph: Transformer for unsupervised medical image registration. Medical image analysis, 82:102615, 2022. 6 +[7] Tingxiu Chen, Yilei Shi, Zixuan Zheng, Bingcong Yan, Jingliang Hu, Xiao Xiang Zhu, and Lichao Mou. Ultrasound image-to-video synthesis via latent dynamic diffusion models. In MICCAI, pages 764-774. Springer, 2024. 4, 6 +[8] Myungsub Choi, Heewon Kim, Bohyung Han, Ning Xu, and Kyoung Mu Lee. Channel attention is all you need for video frame interpolation. In AAAI, pages 10663-10671, 2020. 3 +[9] Myungsub Choi, Suyoung Lee, Heewon Kim, and Kyoung Mu Lee. Motion-aware dynamic architecture for efficient frame interpolation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13839-13848, 2021. 3 +[10] Duolikun Danier, Fan Zhang, and David Bull. Ldmvfi: Video frame interpolation with latent diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1472-1480, 2024. 1, 4, 6 +[11] Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick Van Der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In ICCV, pages 2758-2766, 2015. 3 + +[12] Jan Ehrhardt, Cristian Lorenz, et al. 4D modeling and estimation of respiratory motion for radiation therapy. Springer, 2013. 1 +[13] Yuyu Guo, Lei Bi, Euijoon Ahn, Dagan Feng, Qian Wang, and Jinman Kim. A spatiotemporal volumetric interpolation network for 4d dynamic medical image. In CVPR, pages 4726-4735, 2020. 1, 3, 6, 7 +[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 30, 2017. 6 +[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 4, 6 +[16] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. NeurIPS, 35:8633-8646, 2022. 2, 4, 6 +[17] Kan N Hor, Rolf Baumann, Gianni Pedrizzetti, Gianni Tonti, William M Gottliebson, Michael Taylor, D Woodrow Benson, and Wojciech Mazur. Magnetic resonance derived myocardial strain assessment using feature tracking. Journal of visualized experiments: JoVE, (48):2356, 2011. 1 +[18] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssim. In 2010 20th international conference on pattern recognition, pages 2366-2369. IEEE, 2010. 6 +[19] Mengshun Hu, Kui Jiang, Liang Liao, Jing Xiao, Junjun Jiang, and Zheng Wang. Spatial-temporal space hand-in-hand: Spatial-temporal video super-resolution via cycle-projected mutual learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3574-3583, 2022. 3 +[20] Mengshun Hu, Kui Jiang, Zhihang Zhong, Zheng Wang, and Yinqiang Zheng. Iq-vfi: Implicit quadratic motion estimation for video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6410-6419, 2024. 3 +[21] Zhewei Huang, Tianyuan Zhang, Wen Heng, Boxin Shi, and Shuchang Zhou. Real-time intermediate flow estimation for video frame interpolation. In European Conference on Computer Vision, pages 624-642. Springer, 2022. 3 +[22] Geoffrey D Hugo, Elisabeth Weiss, William C Sleeman, Salim Balik, Paul J Keall, Jun Lu, and Jeffrey F Williamson. Data from 4d lung imaging of nsclc patients. (No Title), 2016. 6 +[23] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In CVPR, pages 2462-2470, 2017. 3 +[24] Siddhant Jain, Daniel Watson, Eric Tabellion, Ben Poole, Janne Kontkanen, et al. Video interpolation with diffusion models. In CVPR, pages 7341-7351, 2024. 1, 4, 6 +[25] Mi-Young Jeung, Philippe Germain, et al. Myocardial tagging with mr imaging: overview of normal and pathologic findings. Radiographics, 32(5):1381-1398, 2012. 1 +[26] Huaizu Jiang, Deqing Sun, Varun Jampani, Ming-Hsuan Yang, Erik Learned-Miller, and Jan Kautz. Super slomo: High quality estimation of multiple intermediate frames for + +video interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9000-9008, 2018. 3 +[27] Xin Jin, Longhai Wu, Jie Chen, Youxin Chen, Jayoon Koo, and Cheul-hee Hahm. A unified pyramid recurrent network for video frame interpolation. In CVPR, pages 1578-1587, 2023. 3 +[28] Tarun Kalluri, Deepak Pathak, Manmohan Chandraker, and Du Tran. Flavr: Flow-agnostic video representations for fast frame interpolation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2071–2082, 2023. 3 +[29] Boah Kim and Jong Chul Ye. Diffusion deformable model for 4d temporal medical image generation. In MICCAI, pages 539-548. Springer, 2022. 1, 4, 6, 7 +[30] JungEun Kim, Hangyul Yoon, Geondo Park, Kyungsu Kim, and Eunho Yang. Data-efficient unsupervised interpolation without any intermediate frame for 4d medical images. In CVPR, pages 11353-11364, 2024. 1, 4, 6, 7 +[31] Lingtong Kong, Boyuan Jiang, Donghao Luo, Wenqing Chu, Xiaoming Huang, Ying Tai, Chengjie Wang, and Jie Yang. Ifrnet: Intermediate feature refine network for efficient frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1969-1978, 2022. 1, 3, 6 +[32] Xia Li, Runzhao Yang, Xiangtai Li, Antony Lomax, Ye Zhang, and Joachim Buhmann. Cpt-interp: Continuous spatial and temporal motion modeling for 4d medical image interpolation. arXiv preprint arXiv:2405.15385, 2024. 2 +[33] Zhen Li, Zuo-Liang Zhu, Ling-Hao Han, Qibin Hou, Chun-Le Guo, and Ming-Ming Cheng. Amt: All-pairs multi-field transforms for efficient frame interpolation. In CVPR, pages 9801-9810, 2023. 3 +[34] Hanwen Liang, Yuyang Yin, Dejia Xu, Hanxue Liang, Zhangyang Wang, Konstantinos N Plataniotis, Yao Zhao, and Yunchao Wei. Diffusion4d: Fast spatial-temporal consistent 4d generation via video diffusion models. arXiv preprint arXiv:2405.16645, 2024. 4 +[35] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv preprint arXiv:2402.17177, 2024. 2, 4 +[36] Liying Lu, Ruizheng Wu, Huaijia Lin, Jiangbo Lu, and Jiaya Jia. Video frame interpolation with transformer. In CVPR, pages 3532-3542, 2022. 3 +[37] Simone Meyer, Abdelaziz Djerouah, Brian McWilliams, Alexander Sorkine-Hornung, Markus Gross, and Christopher Schroers. Phasenet for video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 498-507, 2018. 1 +[38] Kotaro Mizuno and Masahiro Muto. Preoperative evaluation of pleural adhesion in patients with lung tumors using four-dimensional computed tomography performed during natural breathing. *Medicine*, 100(47):e27800, 2021. 1 +[39] Simon Niklaus and Feng Liu. Softmax splatting for video frame interpolation. In Proceedings of the IEEE/CVF con + +ference on computer vision and pattern recognition, pages 5437-5446, 2020. 1 +[40] David Ouyang, Bryan He, Amirata Ghorbani, Neal Yuan, Joseph Ebinger, Curtis P Langlotz, Paul A Heidenreich, Robert A Harrington, David H Liang, Euan A Ashley, et al. Video-based ai for beat-to-beat assessment of cardiac function. Nature, 580(7802):252-256, 2020. 1 +[41] Junheum Park, Jintae Kim, and Chang-Su Kim. Biformer: Learning bilateral motion estimation via bilateral transformer for 4k video frame interpolation. In CVPR, pages 1568-1577, 2023. 1, 3 +[42] Saikat Roy, Gregor Koehler, Constantin Ulrich, Michael Baumgartner, Jens Petersen, Fabian Isensee, Paul F Jaeger, and Klaus H Maier-Hein. Mednext: transformer-driven scaling of convnets for medical image segmentation. In MICCAI, pages 405–415. Springer, 2023. 1 +[43] Otto A Smiseth, Hans Torp, Anders Opdahl, Kristina H Haugaa, and Stig Urheim. Myocardial strain imaging: how useful is it in clinical decision making? European heart journal, 37 (15):1196-1207, 2016. 1 +[44] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In CVPR, pages 8934-8943, 2018. 3 +[45] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, pages 402-419. Springer, 2020. 3 +[46] Georgi P Tolstov. Fourier series. Courier Corporation, 2012. 2, 4 +[47] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 6 +[48] Bingchao Wang, Zhiwei Ning, Jianyu Ding, Xuanang Gao, Yin Li, Dongsheng Jiang, Jie Yang, and Wei Liu. Fix-clip: Dual-branch hierarchical contrastive learning via synthetic captions for better understanding of long text. arXiv preprint arXiv:2507.10095, 2025. 2 +[49] Lu Wang, Shelly Hayes, Kamen Paskalev, Lihui Jin, Mark K Buyyounouski, Charlie C-M Ma, and Steve Feigenberg. Dosimetric comparison of stereotactic body radiotherapy using 4d ct and multiphase ct images for treatment planning of lung cancer: evaluation of the impact on daily dose coverage. *Radiotherapy and Oncology*, 91(3):314–324, 2009. 1 +[50] Yihan Wang, Lahav Lipson, and Jia Deng. Sea-raft: Simple, efficient, accurate raft for optical flow. In ECCV, pages 36-54. Springer, 2024. 3 +[51] Tzu-Ti Wei, Chin Kuo, Yu-Chee Tseng, and Jen-Jee Chen. Mpvf: 4d medical image inpainting by multi-pyramid voxel flows. IEEE Journal of Biomedical and Health Informatics, 2023. 1, 2, 3, 6 +[52] Xin You, Ming Ding, Minghui Zhang, Yangqian Wu, Yi Yu, Yun Gu, and Jie Yang. Semantic difference guidance for the uncertain boundary segmentation of ct left atrial appendage. In MICCAI, pages 121-131. Springer, 2023. 1 +[53] Xin You, Junjun He, Jie Yang, and Yun Gu. Learning with explicit shape priors for medical image segmentation. IEEE Transactions on Medical Imaging, 2024. 6 + +[54] Xin You, Yixin Lou, Minghui Zhang, Jie Yang, and Yun Gu. Slord: Structural low-rank descriptors for shape consistency in vertebrae segmentation. IEEE Journal of Biomedical and Health Informatics, 2025. 4 +[55] Xin You, Minghui Zhang, Hanxiao Zhang, Jie Yang, and Nassir Navab. Temporal differential fields for 4d motion modeling via image-to-video synthesis. arXiv preprint arXiv:2505.17333, 2025. 6 +[56] Weihao Yu, Yuanhao Cai, Ruyi Zha, Zhiwen Fan, Chenxin Li, and Yixuan Yuan. X $^2$ -gaussian: 4d radiative gaussian splatting for continuous-time tomographic reconstruction. arXiv preprint arXiv:2503.21779, 2025. 1 +[57] Chuyan Zhang, Hao Zheng, Xin You, Yefeng Zheng, and Yun Gu. Pass: test-time prompting to adapt styles and semantic shapes in medical image segmentation. IEEE Transactions on Medical Imaging, 2024. 6 +[58] Guozhen Zhang, Yuhan Zhu, Haonan Wang, Youxin Chen, Gangshan Wu, and Limin Wang. Extracting motion and appearance via inter-frame attention for efficient video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5682-5692, 2023. 1, 3 +[59] Haiyu Zhang, Xinyuan Chen, Yaohui Wang, Xihui Liu, Yunhong Wang, and Yu Qiao. 4diffusion: Multi-view video diffusion model for 4d generation. NeurIPS, 37:15272-15295, 2025. 4 +[60] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 6 +[61] Zhihang Zhong, Gurunandan Krishnan, Xiao Sun, Yu Qiao, Sizhuo Ma, and Jian Wang. Clearer frames, anytime: Resolving velocity ambiguity in video frame interpolation. In ECCV, pages 346-363. Springer, 2024. 1, 6 +[62] Kun Zhou, Wenbo Li, Xiaoguang Han, and Jiangbo Lu. Exploring motion ambiguity and alignment for high-quality video frame interpolation. In CVPR, pages 22169-22179, 2023. 1 \ No newline at end of file diff --git a/2025/FB-Diff_ Fourier Basis-guided Diffusion for Temporal Interpolation of 4D Medical Imaging/images.zip b/2025/FB-Diff_ Fourier Basis-guided Diffusion for Temporal Interpolation of 4D Medical Imaging/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2d141fa907aaf5ee36371490142df65b870a799b --- /dev/null +++ b/2025/FB-Diff_ Fourier Basis-guided Diffusion for Temporal Interpolation of 4D Medical Imaging/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9faa601e3014562f4e3b38fc14b87899f5ab2811fd80e55c8e30e35f915caa05 +size 740286 diff --git a/2025/FB-Diff_ Fourier Basis-guided Diffusion for Temporal Interpolation of 4D Medical Imaging/layout.json b/2025/FB-Diff_ Fourier Basis-guided Diffusion for Temporal Interpolation of 4D Medical Imaging/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a90c5eaf2499627516f3c0a84480c3261d3ac37b --- /dev/null +++ b/2025/FB-Diff_ Fourier Basis-guided Diffusion for Temporal Interpolation of 4D Medical Imaging/layout.json @@ -0,0 +1,8554 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 79, + 102, + 532, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 102, + 532, + 139 + ], + "spans": [ + { + "bbox": [ + 79, + 102, + 532, + 139 + ], + "type": "text", + "content": "FB-Diff: Fourier Basis-guided Diffusion for Temporal Interpolation of 4D Medical Imaging" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 159, + 550, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 159, + 550, + 175 + ], + "spans": [ + { + "bbox": [ + 61, + 159, + 550, + 175 + ], + "type": "text", + "content": "Xin You" + }, + { + "bbox": [ + 61, + 159, + 550, + 175 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 61, + 159, + 550, + 175 + ], + "type": "text", + "content": " Runze Yang" + }, + { + "bbox": [ + 61, + 159, + 550, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 159, + 550, + 175 + ], + "type": "text", + "content": " Chuyan Zhang" + }, + { + "bbox": [ + 61, + 159, + 550, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 159, + 550, + 175 + ], + "type": "text", + "content": " Zhongliang Jiang" + }, + { + "bbox": [ + 61, + 159, + 550, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 159, + 550, + 175 + ], + "type": "text", + "content": " Jie Yang" + }, + { + "bbox": [ + 61, + 159, + 550, + 175 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 61, + 159, + 550, + 175 + ], + "type": "text", + "content": " Nassir Navab" + }, + { + "bbox": [ + 61, + 159, + 550, + 175 + ], + "type": "inline_equation", + "content": "^{2,*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 113, + 189, + 501, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 189, + 501, + 204 + ], + "spans": [ + { + "bbox": [ + 113, + 189, + 501, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 189, + 501, + 204 + ], + "type": "text", + "content": " Institute of Medical Robotics, Shanghai Jiao Tong University, Shanghai, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 87, + 204, + 526, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 204, + 526, + 219 + ], + "spans": [ + { + "bbox": [ + 87, + 204, + 526, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 87, + 204, + 526, + 219 + ], + "type": "text", + "content": " Computer Aided Medical Procedures, Technical University of Munich, Munich, Germany" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 166, + 218, + 447, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 218, + 447, + 233 + ], + "spans": [ + { + "bbox": [ + 166, + 218, + 447, + 233 + ], + "type": "text", + "content": "3 Munich Center for Machine Learning, Munich, Germany" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 221, + 235, + 389, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 235, + 389, + 245 + ], + "spans": [ + { + "bbox": [ + 221, + 235, + 389, + 245 + ], + "type": "text", + "content": "xin.you@tum.de, nassir.navab@tum.de" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 152, + 274, + 200, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 274, + 200, + 286 + ], + "spans": [ + { + "bbox": [ + 152, + 274, + 200, + 286 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 300, + 296, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 300, + 296, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 300, + 296, + 610 + ], + "type": "text", + "content": "The temporal interpolation task for 4D medical imaging, plays a crucial role in clinical practice of respiratory motion modeling. Following the simplified linear-motion hypothesis, existing approaches adopt optical flow-based models to interpolate intermediate frames. However, realistic respiratory motions should be nonlinear and quasi-periodic with specific frequencies. Intuited by this property, we resolve the temporal interpolation task from the frequency perspective, and propose a Fourier Basis-guided Diffusion model, termed FB-Diff. Specifically, due to the regular motion discipline of respiration, physiological motion priors are introduced to describe general characteristics of temporal data distributions. Then a Fourier motion operator is elaborately devised to extract Fourier bases by incorporating physiological motion priors and case-specific spectral information in the feature space of Variational Autoencoder. Well-learned Fourier bases can better simulate respiratory motions with motion patterns of specific frequencies. Conditioned on starting and ending frames, the diffusion model further leverages well-learned Fourier bases via the basis interaction operator, which promotes the temporal interpolation task in a generative manner. Extensive results demonstrate that FB-Diff achieves state-of-the-art (SOTA) perceptual performance with better temporal consistency while maintaining promising reconstruction metrics. Codes are available here." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 633, + 135, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 633, + 135, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 633, + 135, + 645 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "4D medical images, which depict 3D volumes with temporal variations, are essential in clinical practice for capturing dynamic changes and monitoring the progression of diseases over time [12, 25, 49, 52]. A significant application is the temporal modeling for breathing-induced motions of" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 275, + 555, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 555, + 335 + ], + "type": "text", + "content": "cardiac or pulmonary anatomical structures. That task aims to enrich motion visualizations by generating intermediate frames under the condition of starting and ending frames, which is consistent with the concept of Video Frame Interpolation (VFI) in natural scenarios." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 338, + 556, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 556, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 556, + 506 + ], + "type": "text", + "content": "VFI has been thoroughly investigated through systematic studies in recent years [10, 24, 31, 37, 39, 41, 58, 61, 62]. However, due to unique properties and constraints of medical imaging, it remains a challenge to directly implement these frameworks into 4D medical video interpolation with maintaining equally promising performance. Specifically, the quality of ground truth intermediate frames in medical imaging is often compromised due to factors such as imaging noise, unstable breathing, causing unexpected artifacts [17, 25, 43, 49]. Thus, the lower image quality compared to videos in the natural domain, will degrade the interpolation performance of VFI models. Besides, potential patients' movement might induce an additional disturbance to the temporal modeling [5, 38]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 508, + 556, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 556, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 556, + 616 + ], + "type": "text", + "content": "Furthermore, there exists a discrepancy on the type of motions between natural and medical scenes. The former focuses on the object movement and illumination intensity change. While in the medical domain, motions can be depicted as subtle anatomical variance with quasi-periodic motion discipline [51, 56]. Thus, a specific motion prior of anatomies is effective to boost the temporal interpolation, especially for the medical data with limited resources from a small pool of individuals." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "On account of these challenges, we present the following question: \"How can we devise a specialist VFI model tailored for 4D temporal medical volumes?\" Existing methods [2, 13, 29, 30, 51] on medical VFI face a major shortcoming: the linear motion hypothesis in a breathing period is required for the regression of intermediate frames as revealed in Fig.1(a). However, respiratory motions should be a nonlinear oscillatory process that cannot be simply mod-" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28010" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 70, + 527, + 284 + ], + "blocks": [ + { + "bbox": [ + 86, + 70, + 527, + 284 + ], + "lines": [ + { + "bbox": [ + 86, + 70, + 527, + 284 + ], + "spans": [ + { + "bbox": [ + 86, + 70, + 527, + 284 + ], + "type": "image", + "image_path": "49b55a0a073f7468310942ce3ca259cec8925fa39bd51feff78d4aec58660ba8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 204, + 287, + 418, + 300 + ], + "lines": [ + { + "bbox": [ + 204, + 287, + 418, + 300 + ], + "spans": [ + { + "bbox": [ + 204, + 287, + 418, + 300 + ], + "type": "text", + "content": "(b) Nonlinear & quasi-periodic motion " + }, + { + "bbox": [ + 204, + 287, + 418, + 300 + ], + "type": "inline_equation", + "content": "\\Longrightarrow" + }, + { + "bbox": [ + 204, + 287, + 418, + 300 + ], + "type": "text", + "content": " Our model" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 308, + 555, + 343 + ], + "lines": [ + { + "bbox": [ + 54, + 308, + 555, + 343 + ], + "spans": [ + { + "bbox": [ + 54, + 308, + 555, + 343 + ], + "type": "text", + "content": "Figure 1. Temporal Interpolation for respiratory motions. (a) Simplified linear-motion hypothesis: interpolating intermediate frames by regressing optical flows with linear variations. (b) Nonlinear & quasi-periodic motions: leveraging diffusion models conditioned on Fourier motion bases, which represent various motion patterns indicating nonlinear and locally inconsistent anatomical variations." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 363, + 295, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 363, + 295, + 554 + ], + "spans": [ + { + "bbox": [ + 54, + 363, + 295, + 554 + ], + "type": "text", + "content": "eled with a linear equation [32, 51]. Also, these methods fail to recognize that respiratory motions in heart or lung, are indeed a quasi-periodic motion process with specific frequencies by extending the timeline. Intuited by this, we resolve the respiratory motion simulation from a frequency perspective. In Fig.1(b), each medical volume in the temporal domain can be viewed as a specific value in the periodic curve. That periodic curve can be decomposed as a group of sine and cosine bases with multiple frequencies [46]. Comparatively, there exists a group of Fourier bases, which are capable of reconstructing original temporal sequences. Here Fourier bases with different frequencies refer to various motion patterns, which indicate nonlinear and locally anisotropic variations of anatomical structures. In our work, Fourier bases are acquired by the Fast Fourier Transform (FFT) on the temporal domain." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": "In this work, 4D data distributions of the respiratory process depict a regular motion discipline, with temporal variations from inhaling to exhaling, or from diastole to systole. Motivated by that observation, physiology motion priors are introduced to encode generic characteristics of the respiratory process. Then a Fourier Motion Operator " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "O_{FM}" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": " is elaborately devised to yield enhanced motion bases by incorporating physiology motion priors and case-specific spectral information in the feature space. Specifically, A Variational Autoencoder (VAE) is introduced to map the video space into the feature space. Physiology motion priors are defined as learnable frequency embeddings, which can be updated by pretraining a Variational Autoen" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 363, + 555, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 555, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 555, + 579 + ], + "type": "text", + "content": "coder (VAE) on given temporal data. And case-specific frequency information can be extracted by FFT in the temporal domain. Well-learned physiology motion priors can enrich Fourier bases with generic properties of respiratory motions. Then enhanced Fourier motion bases can better simulate the quasi-periodic respiratory process with motion patterns of specific frequencies. On the grounds that diffusion models show promising performance in generating realistic videos [15, 16, 35, 48], a Markov chain process is learned to transform Gaussian noise into target videos by the proposed Fourier Bases-guided Diffusion, termed FB-Diff. Since only starting and ending frames are available during inference, they serve as conditional prompts for intermediate frame interpolation. Meanwhile, we devise the basis interaction operator to inject enhanced Fourier bases into the conditional diffusion model. Then, FB-Diff can well address the temporal interpolation task in a generative manner." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "content": "Our proposed method has achieved state-of-the-art (SOTA) interpolation performance compared with other methods for 4D medical imaging, particularly on the mid-most frames. For quantitative results, more comprehensive evaluation metrics are adopted in this work, including traditional reconstruction metrics and feature-level perceptual metrics. FB-Diff not only presents promising reconstruction metrics but also superior perceptual results. Qualitative visualizations reveal better temporal consistency by FB-Diff compared with other linear-motion-based methods. Our main contributions are summarized as follows:" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28011" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 78, + 70, + 532, + 393 + ], + "blocks": [ + { + "bbox": [ + 78, + 70, + 532, + 393 + ], + "lines": [ + { + "bbox": [ + 78, + 70, + 532, + 393 + ], + "spans": [ + { + "bbox": [ + 78, + 70, + 532, + 393 + ], + "type": "image", + "image_path": "ea0242ad2ded6b4b4995d4b5240260ecbefd64127560851e8856bf2d9f92bc00.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 401, + 555, + 449 + ], + "lines": [ + { + "bbox": [ + 55, + 401, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 555, + 449 + ], + "type": "text", + "content": "Figure 2. The pipeline of FB-Diff. (a) The Fourier motion operator is proposed to extract Fourier bases in a feature space of VAE. Bases consist of learnable generic motion priors and case-specific frequency information. (b) Fourier bases are first adapted to the masked video domain under the guidance of well-learned motion priors and pretrained VAE weights. Domain-adapted bases serve as the conditional input to boost the video synthesis by diffusion models." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 467, + 295, + 634 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 467, + 295, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 467, + 295, + 527 + ], + "spans": [ + { + "bbox": [ + 55, + 467, + 295, + 527 + ], + "type": "text", + "content": "- We propose a diffusion model termed FB-Diff for the temporal interpolation of 4D medical imaging. FB-Diff leverages Fourier motion bases, which indicate motion patterns with specific frequencies to resolve the respiratory modeling with nonlinear and quasi-periodic motions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 527, + 295, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 527, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 527, + 295, + 563 + ], + "type": "text", + "content": "- We introduce the Fourier motion operator " + }, + { + "bbox": [ + 55, + 527, + 295, + 563 + ], + "type": "inline_equation", + "content": "O_{FM}" + }, + { + "bbox": [ + 55, + 527, + 295, + 563 + ], + "type": "text", + "content": " to extract Fourier bases, containing physiology motion priors and case-specific frequency representations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 563, + 295, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 295, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 295, + 598 + ], + "type": "text", + "content": "- We devise the basis interaction operator to inject Fourier bases into the diffusion model. Fourier bases can boost the video synthesis for temporal interpolation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 598, + 295, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 598, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 598, + 295, + 634 + ], + "type": "text", + "content": "- Extensive results demonstrate that FB-Diff reveals SOTA perceptual performance with better temporal consistency while maintaining promising reconstruction metrics." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 645, + 142, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 645, + 142, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 645, + 142, + 658 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "type": "text", + "content": "Video Frame Interpolation. Popular VFI techniques are mainly classified into motion-free and motion-based approaches, depending on whether optical flows are incorporated or not. Motion-based: These approaches warp" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 467, + 555, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 555, + 624 + ], + "type": "text", + "content": "input frames forward or backward based on optical flows acquired by either off-the-shelf networks [11, 23, 44, 45, 50] or custom-designed flow estimators [20, 21, 27, 31, 33, 58]. Warped intermediate frames are then refined by delicately designed networks [3, 26] to enhance visual quality. Recently, Transformer-based motion estimators are proposed to raise the upper limit of VFI performance [36, 41, 58] Motion-Free: Motion-free methods rely on implicit spatiotemporal modeling [8, 9, 19, 28] to generate the intermediate frame. However, due to a limited quality of 4D medical videos and quasi-periodic motion priors, existing VFI frameworks are not feasible. That is why we aim to propose a specialist VFI model for 4D medical sequences." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "4D Medical VFI. Existing methods on this topic mainly leverage volumetric motions for 4D temporal interpolation. Specifically, SVIN [13] estimates forward and backward deformation fields, which will yield precise intermediate frames by linearly combining bidirectional information. MPVF [51] resolves various magnitudes of motions by proposing a multi-pyramid voxel flows model that takes" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28012" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "type": "text", + "content": "multi-scale voxel flows into account. DDM [29] can learn spatial deformation information between the source and target volumes and provide a latent code for generating intermediate frames along a geodesic path. UVI-Net [30] utilizes the flow calculation model with the time-domain cycle-consistency constraint and linear motion hypothesis, to realize motion modeling in an unsupervised style. However, the linear hypothesis for motions will induce temporal inconsistency and spatial distortion." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 182, + 296, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 182, + 296, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 296, + 339 + ], + "type": "text", + "content": "Diffusion-based Video Interpolation. Denoising diffusion probabilistic models [15, 16, 35] reveal promising performance in generating realistic images or videos. Some methods maximize the potential of diffusion models to resolve video interpolation tasks. Essentially, these approaches implicitly learn the temporal motions from the starting frame to the ending frame [1, 7, 24]. LDMVFI [10] proposes a latent diffusion model, which approaches the VFI problem from a generative perspective by formulating it as a conditional synthesis issue. However, these diffusion models do not leverage effective conditional guidance, which is crucial for medical VFI with intrinsic motion priors. Thus, we introduce Fourier bases to enhance the conditional synthesis." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 348, + 138, + 362 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 138, + 362 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 138, + 362 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 368, + 180, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 368, + 180, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 368, + 180, + 380 + ], + "type": "text", + "content": "3.1. Fourier Motion Bases" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 385, + 296, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 385, + 296, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 385, + 296, + 517 + ], + "type": "text", + "content": "Since respiratory motions for heart or lung exhibit a quasi-periodic characteristic, we intend to resolve this task from a frequency perspective. Specifically, the respiration can be viewed as an oscillatory process with certain frequencies. Thus, we draw an analogy between 2D time-varying curves and 4D temporal medical volumes. Curves can be decomposed into a group of harmonic oscillators with various frequencies [46]. Similarly, temporal volumes can be restored via a stack of extracted Fourier bases, which are embodied with various motion patterns of anatomical deformation. The whole process can be formulated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 80, + 517, + 295, + 549 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 517, + 295, + 549 + ], + "spans": [ + { + "bbox": [ + 80, + 517, + 295, + 549 + ], + "type": "interline_equation", + "content": "\\mathcal {B} _ {k} = \\sum_ {n = 0} ^ {N - 1} V _ {n} e ^ {- i 2 \\pi k \\frac {n}{N}}, k = 0, 1, \\dots , N - 1, \\tag {1}", + "image_path": "ee4b60718d3368be731cfcefe0eca32565f0394f7b43b54d18dbc27fc3301e33.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 77, + 552, + 295, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 552, + 295, + 583 + ], + "spans": [ + { + "bbox": [ + 77, + 552, + 295, + 583 + ], + "type": "interline_equation", + "content": "V _ {n} = \\frac {1}{T} \\sum_ {k = 0} ^ {N - 1} \\mathcal {B} _ {k} e ^ {i 2 \\pi n \\frac {k}{N}}, n = 0, 1, \\dots , N - 1, \\tag {2}", + "image_path": "d3d0677121b442fb97797247315ae2de63522731356a1480a56e9813b8021ee8.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 591, + 296, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 591, + 296, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 591, + 296, + 687 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 591, + 296, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_k" + }, + { + "bbox": [ + 55, + 591, + 296, + 687 + ], + "type": "text", + "content": " means the " + }, + { + "bbox": [ + 55, + 591, + 296, + 687 + ], + "type": "inline_equation", + "content": "k_{th}" + }, + { + "bbox": [ + 55, + 591, + 296, + 687 + ], + "type": "text", + "content": " Fourier basis in the frequency domain, " + }, + { + "bbox": [ + 55, + 591, + 296, + 687 + ], + "type": "inline_equation", + "content": "V_{n}" + }, + { + "bbox": [ + 55, + 591, + 296, + 687 + ], + "type": "text", + "content": " refers to the " + }, + { + "bbox": [ + 55, + 591, + 296, + 687 + ], + "type": "inline_equation", + "content": "n_{th}" + }, + { + "bbox": [ + 55, + 591, + 296, + 687 + ], + "type": "text", + "content": " medical volume, and " + }, + { + "bbox": [ + 55, + 591, + 296, + 687 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 591, + 296, + 687 + ], + "type": "text", + "content": " is the frame number. These Fourier bases describe locally anisotropic anatomical variations with different motion intensities, promoting the performance of the temporal interpolation model " + }, + { + "bbox": [ + 55, + 591, + 296, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 55, + 591, + 296, + 687 + ], + "type": "text", + "content": ". Thus, the goal of our work is to deduce 4D medical videos " + }, + { + "bbox": [ + 55, + 591, + 296, + 687 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 591, + 296, + 687 + ], + "type": "text", + "content": " conditioned on two prompting frames " + }, + { + "bbox": [ + 55, + 591, + 296, + 687 + ], + "type": "inline_equation", + "content": "V_{0}, V_{N-1}" + }, + { + "bbox": [ + 55, + 591, + 296, + 687 + ], + "type": "text", + "content": " and Fourier motion bases " + }, + { + "bbox": [ + 55, + 591, + 296, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 55, + 591, + 296, + 687 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 91, + 696, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 696, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 91, + 696, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {I} ^ {*} = \\underset {\\mathcal {I}} {\\arg \\min } \\| \\boldsymbol {V} - \\mathcal {I} \\left(V _ {0}, V _ {N - 1}; \\mathcal {B}\\right) \\| _ {\\mathcal {D}}, \\tag {3}", + "image_path": "f3695dee50facbd0d20eb1e0a9bea38b7c2f1915d5644b80b7978cad003b9e35.jpg" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 323, + 70, + 547, + 209 + ], + "blocks": [ + { + "bbox": [ + 323, + 70, + 547, + 209 + ], + "lines": [ + { + "bbox": [ + 323, + 70, + 547, + 209 + ], + "spans": [ + { + "bbox": [ + 323, + 70, + 547, + 209 + ], + "type": "image", + "image_path": "3a5fb39ec16e2282dd96e4173cb949fd02dec76d57657897bdfbad4585e8955a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 218, + 555, + 251 + ], + "lines": [ + { + "bbox": [ + 313, + 218, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 218, + 555, + 251 + ], + "type": "text", + "content": "Figure 3. The details of Fourier motion operator. Fourier bases are decomposed as physiology motion priors and case-specific Fourier features, corresponding to phase and magnitude components." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 258, + 555, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^*" + }, + { + "bbox": [ + 313, + 258, + 555, + 376 + ], + "type": "text", + "content": " refers to the optimal solution in the interpolation model set, and " + }, + { + "bbox": [ + 313, + 258, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 258, + 555, + 376 + ], + "type": "text", + "content": " represents the distance-based normed space. We systematically organize the following sections along Fourier bases. Section 3.2 introduces the design of the Fourier Motion Operator, which constructs Fourier bases with physiology motion priors and case-specific frequency information. Section 3.3 adapts Fourier bases to the to-be-interpolated video domain with well-learned physiology motion priors. Section 3.4 provides details of the diffusion model conditioned on Fourier bases." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 385, + 457, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 385, + 457, + 398 + ], + "spans": [ + { + "bbox": [ + 314, + 385, + 457, + 398 + ], + "type": "text", + "content": "3.2. Fourier Motion Operator" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 403, + 555, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 555, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 555, + 582 + ], + "type": "text", + "content": "To explicitly extract the Fourier bases with various motion patterns for the respiration process, we introduce the Fourier Operator " + }, + { + "bbox": [ + 313, + 403, + 555, + 582 + ], + "type": "inline_equation", + "content": "O_{\\mathcal{F}}" + }, + { + "bbox": [ + 313, + 403, + 555, + 582 + ], + "type": "text", + "content": " as a neural operator. As stated in Eq.(2), " + }, + { + "bbox": [ + 313, + 403, + 555, + 582 + ], + "type": "inline_equation", + "content": "O_{\\mathcal{F}}" + }, + { + "bbox": [ + 313, + 403, + 555, + 582 + ], + "type": "text", + "content": " will operate in the temporal domain, not the spatial domain. Since the common Fourier operator " + }, + { + "bbox": [ + 313, + 403, + 555, + 582 + ], + "type": "inline_equation", + "content": "O_{\\mathcal{F}}" + }, + { + "bbox": [ + 313, + 403, + 555, + 582 + ], + "type": "text", + "content": " only mines case-specific frequency representations, general physiology characteristics inside respiratory data are ignored. Inspired by that notion, we devise the Fourier motion operator " + }, + { + "bbox": [ + 313, + 403, + 555, + 582 + ], + "type": "inline_equation", + "content": "O_{FM}" + }, + { + "bbox": [ + 313, + 403, + 555, + 582 + ], + "type": "text", + "content": " as a neural operator in the feature space. To map the image space into feature space, we introduce a Variational Autoencoder (VAE) as revealed in Fig.2(a). Due to no built-in 4D convolutional operator, we replace it with 3D spatial convolution and 1D temporal convolution as [34, 59] did. The spatial layer in Fig.2(a) consists of spatial convolution, GELU, and group normalization operations." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "content": "The Fourier motion operator " + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "inline_equation", + "content": "O_{FM}" + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "content": " is devised to acquire spectral characteristics with both physiology motion priors and case-specific representations in the feature space. Firstly, physiology motion priors are introduced as learnable frequency embeddings, which are updated by backward gradients of restoring given temporal data via VAE. Well-trained frequency embeddings " + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "content": " can represent the generic discipline of the regular respiration process [54], with " + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\theta_{k}" + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "content": " referring to the " + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "inline_equation", + "content": "k_{th}" + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "content": " element " + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "inline_equation", + "content": "(k = 0,1,\\dots,N - 1)" + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "content": ". Then Fourier bases " + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "content": " are made up of physiology motion priors " + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "content": " and case-specific features " + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "content": ", which arise from the output of" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28013" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "content": "spatial and temporal layers. As depicted by Fig.3, " + }, + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "content": " with a complex type is yielded in the following style:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 103, + 295, + 137 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 103, + 295, + 137 + ], + "spans": [ + { + "bbox": [ + 106, + 103, + 295, + 137 + ], + "type": "interline_equation", + "content": "\\mathcal {B} = \\underbrace {\\mathcal {M} (\\left\\| O _ {\\mathcal {F}} (F) \\right\\| _ {2})} _ {\\text {C a s e - S p e c i f i c}} \\cdot \\underbrace {e ^ {i U (\\tau (\\theta))}} _ {\\text {G e n e r a l}}, \\tag {4}", + "image_path": "992857e209f258d53e36524b8b4414162ba43779c75113f435079cb6abe56948.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 144, + 296, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 296, + 251 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 144, + 296, + 251 + ], + "type": "text", + "content": " means the channel transform, " + }, + { + "bbox": [ + 55, + 144, + 296, + 251 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 55, + 144, + 296, + 251 + ], + "type": "text", + "content": " refers to the upsampling which unifies the resolution of " + }, + { + "bbox": [ + 55, + 144, + 296, + 251 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 144, + 296, + 251 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 144, + 296, + 251 + ], + "type": "inline_equation", + "content": "\\pmb{F}" + }, + { + "bbox": [ + 55, + 144, + 296, + 251 + ], + "type": "text", + "content": ". These operations transform general motion priors into the phase part of Fourier bases, which determines the motion variation standing for the whole dataset distribution. Then " + }, + { + "bbox": [ + 55, + 144, + 296, + 251 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\| _2" + }, + { + "bbox": [ + 55, + 144, + 296, + 251 + ], + "type": "text", + "content": " means the " + }, + { + "bbox": [ + 55, + 144, + 296, + 251 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 55, + 144, + 296, + 251 + ], + "type": "text", + "content": " norm, " + }, + { + "bbox": [ + 55, + 144, + 296, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 144, + 296, + 251 + ], + "type": "text", + "content": " is the convolutional mapping operation. They will output motion bases' magnitude information, which selectively highlights regions in spectral maps relevant to case-specific information." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 252, + 296, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 252, + 296, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 296, + 289 + ], + "type": "text", + "content": "Meanwhile, feature maps " + }, + { + "bbox": [ + 55, + 252, + 296, + 289 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 252, + 296, + 289 + ], + "type": "text", + "content": " are further enriched with more fine-grained frequency details. The specific mechanism can be described with Eq.(5):" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 296, + 295, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 296, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 133, + 296, + 295, + 312 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {F}} = \\boldsymbol {F} + \\sigma \\mathcal {O} _ {\\mathcal {F}} ^ {- 1} (\\mathcal {B}), \\tag {5}", + "image_path": "1d008d0f24d1d0619b625a04ce15682c7deb7e48b38601ab26d1cfe3bb3b6300.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 319, + 296, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 296, + 358 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 296, + 358 + ], + "type": "text", + "content": "here " + }, + { + "bbox": [ + 55, + 319, + 296, + 358 + ], + "type": "inline_equation", + "content": "O_{\\mathcal{F}}^{-1}" + }, + { + "bbox": [ + 55, + 319, + 296, + 358 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 319, + 296, + 358 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 319, + 296, + 358 + ], + "type": "text", + "content": " refer to the inverse Fourier transform and Sigmoid operators, which will boost deep features " + }, + { + "bbox": [ + 55, + 319, + 296, + 358 + ], + "type": "inline_equation", + "content": "\\hat{F}" + }, + { + "bbox": [ + 55, + 319, + 296, + 358 + ], + "type": "text", + "content": " with enriched details, including shape and texture information." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 366, + 266, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 266, + 379 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 266, + 379 + ], + "type": "text", + "content": "3.3. Frame Inpainting-based Basis Synthesis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 384, + 296, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 296, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 296, + 538 + ], + "type": "text", + "content": "Due to the fact that only starting and ending frames can be acquired during inference, a significant step is to adapt Fourier bases to the masked video domain. Motivated by the fact that strong correlations exist between Fourier motion bases and inpainted temporal sequences, we plan to jointly optimize these two variables. To accomplish that goal, we repurpose well-calibrated physiology motion priors from VAE, which prove to be effective in providing the motion guidance for intermediate frame inpainting. Also, pretrained weights of VAE encode implicit temporal dependencies between frames, enhancing the reconstruction of masked video sequences via the finetuned VAE noted as " + }, + { + "bbox": [ + 55, + 384, + 296, + 538 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 55, + 384, + 296, + 538 + ], + "type": "text", + "content": ". A mathematical formulation is defined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 85, + 546, + 295, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 546, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 85, + 546, + 295, + 563 + ], + "type": "interline_equation", + "content": "\\hat {V}, \\hat {\\mathcal {B}} = \\mathcal {G} (\\hat {\\boldsymbol {\\theta}}, V _ {\\boldsymbol {M}}) = \\mathcal {G} (\\hat {\\boldsymbol {\\theta}}, V _ {0},.., \\mathbf {0},.., V _ {N - 1}), \\tag {6}", + "image_path": "f65868c56995ac5f8c82022878e4813ba97ec3d6e8e5260422d6316617938379.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 569, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}" + }, + { + "bbox": [ + 55, + 569, + 295, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 569, + 295, + 605 + ], + "type": "inline_equation", + "content": "V_{M}" + }, + { + "bbox": [ + 55, + 569, + 295, + 605 + ], + "type": "text", + "content": " correspond to well-trained general motions and zero-masked videos, " + }, + { + "bbox": [ + 55, + 569, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\hat{V}" + }, + { + "bbox": [ + 55, + 569, + 295, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 569, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\hat{B}" + }, + { + "bbox": [ + 55, + 569, + 295, + 605 + ], + "type": "text", + "content": " refer to inpainted videos and domain-adapted Fourier bases." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "The interaction between temporal video representations and " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{B}}" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " is implemented via the Fourier motion operator as shown in Fig.2(b). For the joint optimization process, video sequences will contribute temporal variations, enhancing the synthesis of adapted Fourier bases. Conversely, Fourier bases for zero-masked inputs inherently generate a spectral map that emphasizes regions with motion in the temporal axis. This activated map aids in temporal frame inpainting by exerting attention on regions with different magnitudes." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 328, + 72, + 542, + 183 + ], + "blocks": [ + { + "bbox": [ + 328, + 72, + 542, + 183 + ], + "lines": [ + { + "bbox": [ + 328, + 72, + 542, + 183 + ], + "spans": [ + { + "bbox": [ + 328, + 72, + 542, + 183 + ], + "type": "image", + "image_path": "e46e1aa2636d3a69e2489f0e0819f6dad5bd9bb3792fe805fd5471e932837caa.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 193, + 554, + 247 + ], + "lines": [ + { + "bbox": [ + 313, + 193, + 554, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 193, + 554, + 247 + ], + "type": "text", + "content": "Figure 4. The architecture of basis interaction operator. Fourier motion bases selectively enhance video representations by measuring the motion intensity of different regions. Then those enhanced temporal representations can enrich denoised embeddings from the diffusion model." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 254, + 517, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 254, + 517, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 254, + 517, + 266 + ], + "type": "text", + "content": "3.4. Basis-guided 4D Conditional Diffusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 270, + 554, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 554, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 554, + 354 + ], + "type": "text", + "content": "VAE in Fig.2(a) is not only devised for universal motion priors, but also for the latent diffusion model (LDM), which shows better compatibility for extra conditional inputs. Besides, LDM synthesizes latent embeddings occupying less GPU memory compared to 4D temporal volumes. And the VAE decoder is leveraged to reconstruct interpolated video sequences with synthesized latent embeddings." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 355, + 554, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 355, + 554, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 355, + 554, + 392 + ], + "type": "text", + "content": "For the specific diffusion process, the whole video sequence " + }, + { + "bbox": [ + 313, + 355, + 554, + 392 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 355, + 554, + 392 + ], + "type": "text", + "content": " is first encoded as a latent embedding " + }, + { + "bbox": [ + 313, + 355, + 554, + 392 + ], + "type": "inline_equation", + "content": "z_0 = \\mathcal{G}_e(V)" + }, + { + "bbox": [ + 313, + 355, + 554, + 392 + ], + "type": "text", + "content": ". Then the latent code " + }, + { + "bbox": [ + 313, + 355, + 554, + 392 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 313, + 355, + 554, + 392 + ], + "type": "text", + "content": " is perturbed as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 354, + 395, + 553, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 395, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 354, + 395, + 553, + 410 + ], + "type": "interline_equation", + "content": "z _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} z _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , \\epsilon \\sim \\mathcal {N} (0, 1), \\tag {7}", + "image_path": "ac70d33aa37240cd5d360162c8f73aa85e33eb4bc899f8c51abfd1e85072f39e.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 414, + 554, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 554, + 498 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 414, + 554, + 498 + ], + "type": "inline_equation", + "content": "\\overline{\\alpha}_t = \\prod_{i=1}^t (1 - \\beta_t)" + }, + { + "bbox": [ + 313, + 414, + 554, + 498 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 414, + 554, + 498 + ], + "type": "inline_equation", + "content": "\\beta_t" + }, + { + "bbox": [ + 313, + 414, + 554, + 498 + ], + "type": "text", + "content": " is the noise coefficient at time step " + }, + { + "bbox": [ + 313, + 414, + 554, + 498 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 414, + 554, + 498 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 414, + 554, + 498 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 414, + 554, + 498 + ], + "type": "text", + "content": " is uniformly sampled from the timestep index set " + }, + { + "bbox": [ + 313, + 414, + 554, + 498 + ], + "type": "inline_equation", + "content": "[1,2,\\dots,T]" + }, + { + "bbox": [ + 313, + 414, + 554, + 498 + ], + "type": "text", + "content": ". This process can be regarded as a Markov chain, which incrementally adds Gaussian noise to the latent code " + }, + { + "bbox": [ + 313, + 414, + 554, + 498 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 313, + 414, + 554, + 498 + ], + "type": "text", + "content": ". The denoising model " + }, + { + "bbox": [ + 313, + 414, + 554, + 498 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 414, + 554, + 498 + ], + "type": "text", + "content": " receives " + }, + { + "bbox": [ + 313, + 414, + 554, + 498 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 313, + 414, + 554, + 498 + ], + "type": "text", + "content": " as input and is optimized under conditional guidance to learn the latent space distribution with the objective function:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 329, + 502, + 553, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 502, + 553, + 517 + ], + "spans": [ + { + "bbox": [ + 329, + 502, + 553, + 517 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\epsilon} = E _ {z _ {t}, \\epsilon \\sim \\mathcal {N} (0, 1)} \\| \\epsilon - \\epsilon_ {d} \\left(z _ {t}, t; \\hat {\\mathcal {B}}, \\mathcal {G} _ {e} \\left(\\boldsymbol {V} _ {\\boldsymbol {M}}\\right), \\hat {\\boldsymbol {V}}\\right) \\| ^ {2}. \\tag {8}", + "image_path": "c908fd410fffbc44c6673593a01e531d45abf01053e9b6c698f52719c828acc9.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 521, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 554, + 605 + ], + "type": "text", + "content": "here " + }, + { + "bbox": [ + 313, + 521, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\epsilon_d" + }, + { + "bbox": [ + 313, + 521, + 554, + 605 + ], + "type": "text", + "content": " is the regressed noise, adapted Fourier bases " + }, + { + "bbox": [ + 313, + 521, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{B}}" + }, + { + "bbox": [ + 313, + 521, + 554, + 605 + ], + "type": "text", + "content": " and inpainted coarse videos " + }, + { + "bbox": [ + 313, + 521, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\hat{V}" + }, + { + "bbox": [ + 313, + 521, + 554, + 605 + ], + "type": "text", + "content": " are incorporated as the conditional guidance. Moreover, feature maps of starting and ending frames by the VAE encoder " + }, + { + "bbox": [ + 313, + 521, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_e" + }, + { + "bbox": [ + 313, + 521, + 554, + 605 + ], + "type": "text", + "content": " serve as a strong prompt for the 4D video synthesis. Then during the training process of FB-Diff, the total loss " + }, + { + "bbox": [ + 313, + 521, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 313, + 521, + 554, + 605 + ], + "type": "text", + "content": " consists of the denoising loss " + }, + { + "bbox": [ + 313, + 521, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\epsilon}" + }, + { + "bbox": [ + 313, + 521, + 554, + 605 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 313, + 521, + 554, + 605 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 313, + 521, + 554, + 605 + ], + "type": "text", + "content": "-norm reconstruction loss " + }, + { + "bbox": [ + 313, + 521, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_r(\\hat{V},V)" + }, + { + "bbox": [ + 313, + 521, + 554, + 605 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 606, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 555, + 714 + ], + "type": "text", + "content": "To inject Fourier motion bases into LDM and harness coarsely inpainted results, we design the Basis Interaction Operator (BIO) as shown in Fig.4. Specifically, inpainted videos are mapped into video feature representations via spatial & temporal layers and the downsampling operation. Fourier motion bases with various motion patterns, selectively enhance video representations by measuring the motion intensity of different regions. Those enhanced temporal representations can enrich denoised embeddings from" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28014" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 83, + 101, + 531, + 264 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Baseline comparison with flow-based and diffusion-based models. (Bold: the best, Underlined: the second best. Δ values represent the relative difference between our model and UVI-Net, and red and blue numbers stand for better and worse results)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 83, + 101, + 531, + 264 + ], + "lines": [ + { + "bbox": [ + 83, + 101, + 531, + 264 + ], + "spans": [ + { + "bbox": [ + 83, + 101, + 531, + 264 + ], + "type": "table", + "html": "
ModelACDC Cardiac4D Lung
PSNR (dB) ↑LPIPS ↓FID ↓FVD ↓PSNR (dB) ↑LPIPS ↓FID ↓FVD ↓
SVIN [13]31.43±0.4211.563±0.20626.393.630.49±0.3042.650±0.24538.5125.6
Voxelmorph [2]30.77±0.5021.969±0.19729.4102.129.90±0.3732.815±0.26043.2149.0
Transmorph [6]29.81±0.5322.136±0.25531.5109.529.04±0.4032.946±0.29140.5143.5
MPVF [51]31.53±0.4101.594±0.20124.891.330.76±0.2652.418±0.24639.4120.3
UVI-Net [30]32.16±0.4021.662±0.24523.094.231.57±0.3112.211±0.21637.8121.7
IFRNet [31]30.58±0.4081.752±0.24927.897.530.05±0.2802.316±0.27339.7124.0
LDMVFI [10]27.11±0.4602.943±0.41028.899.226.31±0.4533.659±0.34142.1142.7
LDDM [7]24.53±0.4812.634±0.30833.4105.725.19±0.2732.914±0.28748.7146.3
Conditional Diff [16]26.59±0.5452.460±0.35730.195.725.95±0.3913.294±0.37545.0133.0
DDM [29]29.79±0.5042.089±0.35227.4110.329.67±0.4202.705±0.33040.8165.5
FB-Diff30.95±0.3861.452±0.20521.588.930.18±0.2872.066±0.21537.7113.6
Δ value1.21 ↓0.210 ↓1.5 ↓5.3 ↓1.39 ↓0.145 ↓0.1 ↓8.1 ↓
", + "image_path": "cad301eea28cf45b15abfeb328188c97382b503a0137fcc1c8a0ba6185357e99.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 293, + 297, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 297, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 297, + 331 + ], + "type": "text", + "content": "the diffusion model, promoting a promising synthesis of highlighted regions. Thus, BIO can boost well-interpolated videos with spatial and temporal consistency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 340, + 138, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 138, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 138, + 354 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 360, + 183, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 183, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 183, + 373 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 378, + 295, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 378, + 295, + 558 + ], + "spans": [ + { + "bbox": [ + 54, + 378, + 295, + 558 + ], + "type": "text", + "content": "Dataset. To evaluate the interpolation performance of our framework, we conduct experiments on public ACDC cardiac [4] and 4D-Lung datasets [22]. For ACDC with the MRI modality, volume sequences between the end-diastolic and end-systolic phases are extracted as valid 4D data. Of all the 150 4D sequences, cases with identity 1-100, 101-120, and 121-150 serve as the training, validation, and testing sets. Besides, all MRI volumes are cropped to " + }, + { + "bbox": [ + 54, + 378, + 295, + 558 + ], + "type": "inline_equation", + "content": "128 \\times 128 \\times 32" + }, + { + "bbox": [ + 54, + 378, + 295, + 558 + ], + "type": "text", + "content": ", and the frame number ranges from 6 to 16 [53, 57]. For 4D-Lung CT images, the end-inspiratory and end-expiratory scans are set as the initial and final images. We collect 125 4D videos from [22], with the resolution equal to " + }, + { + "bbox": [ + 54, + 378, + 295, + 558 + ], + "type": "inline_equation", + "content": "6 \\times 128 \\times 128 \\times 128" + }, + { + "bbox": [ + 54, + 378, + 295, + 558 + ], + "type": "text", + "content": ". They are split as " + }, + { + "bbox": [ + 54, + 378, + 295, + 558 + ], + "type": "inline_equation", + "content": "80 / 15 / 30" + }, + { + "bbox": [ + 54, + 378, + 295, + 558 + ], + "type": "text", + "content": " cases for training, validation, and inference. We uniform the frame number as 16 with zero paddings." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 558, + 295, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 558, + 295, + 701 + ], + "spans": [ + { + "bbox": [ + 54, + 558, + 295, + 701 + ], + "type": "text", + "content": "**Evaluations.** We select quantitative evaluation metrics including Peak Signal-to-Noise Ratio (PSNR) [18], Learned Perceptual Image Patch Similarity (LPIPS) [55, 60], Fréchet Inception Distance (FID) [14], and Fréchet Video Distance (FVD) [47]. Compared to PSNR, which is the traditional pixel-wise reconstruction metric, LPIPS and FVD evaluate perceptual quality from the perspective of human vision [10, 24, 61], which is more feasible and favored in clinical practice. Specifically, LPIPS reveals important perceptual similarity like human vision. FID and FVD evaluate the dataset distribution bias between realistic and synthesized data, including spatial and temporal consistency." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "content": "Training Details. The VAE maps the image space into" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 293, + 555, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 555, + 389 + ], + "type": "text", + "content": "the downsampled latent space with a ratio of " + }, + { + "bbox": [ + 313, + 293, + 555, + 389 + ], + "type": "inline_equation", + "content": "1/8" + }, + { + "bbox": [ + 313, + 293, + 555, + 389 + ], + "type": "text", + "content": ". And the proposed conditional diffusion model is implemented with a lightweight UNet. All models are trained using the AdamW optimizer with the linear warm-up strategy. The batch size, diffusion step " + }, + { + "bbox": [ + 313, + 293, + 555, + 389 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 293, + 555, + 389 + ], + "type": "text", + "content": ", and training epoch are equal to 2, 1000, and 500. Experiments are implemented based on Pytorch and 2 NVIDIA RTX 4090 GPUs. More details about the network and training settings are described in the Appendix." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 396, + 499, + 409 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 499, + 409 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 499, + 409 + ], + "type": "text", + "content": "4.2. Comparisons with SOTA baselines" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 414, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 555, + 605 + ], + "type": "text", + "content": "Quantitative Comparison. We comprehensively conduct experiments on ACDC cardiac and 4D Lung datasets. Quantitative results are illustrated in Tab.1. Our model achieves the best performance across the benchmark on perceptual evaluation metrics from the perspective of human vision [10, 61], which are clinically favored compared with the pixel-centric reconstruction metric [7]. Specifically, FB-Diff outperforms UVI-Net [30] with " + }, + { + "bbox": [ + 313, + 414, + 555, + 605 + ], + "type": "inline_equation", + "content": "0.210 \\downarrow" + }, + { + "bbox": [ + 313, + 414, + 555, + 605 + ], + "type": "text", + "content": " LPIPS and " + }, + { + "bbox": [ + 313, + 414, + 555, + 605 + ], + "type": "inline_equation", + "content": "5.3 \\downarrow" + }, + { + "bbox": [ + 313, + 414, + 555, + 605 + ], + "type": "text", + "content": " FVD values on ACDC. Also, our model maintains quite satisfactory PSNR values of " + }, + { + "bbox": [ + 313, + 414, + 555, + 605 + ], + "type": "inline_equation", + "content": "30.95\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 414, + 555, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 414, + 555, + 605 + ], + "type": "inline_equation", + "content": "30.18\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 414, + 555, + 605 + ], + "type": "text", + "content": " on ACDC and 4D Lung. Although diffusion models are not sensitive to pixel-wise intensities [15], the conditional guidance of Fourier bases and coarsely inpainted videos is effective to boost the interpolation performance of diffusion models. Tab.1 reveals that FB-Diff significantly surpasses existing diffusion models on all metrics." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": "Furthermore, the interpolation process becomes most challenging for the midmost frames since they are positioned farthest from the two reference frames [24]. And our model can better resolve this issue due to stronger temporal representations based on quasi-periodic motion modeling. Thus, a quantitative evaluation of central frames is necessary. As shown in Tab.2, FB-Diff demonstrates greater performance advantages, with " + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "inline_equation", + "content": "4.3\\downarrow" + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": " FID and " + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "inline_equation", + "content": "8.2\\downarrow" + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": " FVD values compared with UVI-Net. Additionally, FB-Diff acquires a competi" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28015" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 84, + 71, + 531, + 277 + ], + "blocks": [ + { + "bbox": [ + 84, + 71, + 531, + 277 + ], + "lines": [ + { + "bbox": [ + 84, + 71, + 531, + 277 + ], + "spans": [ + { + "bbox": [ + 84, + 71, + 531, + 277 + ], + "type": "image", + "image_path": "8da4c198ac582ec7f985250dd8520efe4944f7261aca5a0b3ecc632bc930d0c8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 97, + 283, + 512, + 295 + ], + "lines": [ + { + "bbox": [ + 97, + 283, + 512, + 295 + ], + "spans": [ + { + "bbox": [ + 97, + 283, + 512, + 295 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison of the midmost frame interpolation between different models. VM: Voxelmorph." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 64, + 366, + 290, + 434 + ], + "blocks": [ + { + "bbox": [ + 55, + 314, + 296, + 360 + ], + "lines": [ + { + "bbox": [ + 55, + 314, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 314, + 296, + 360 + ], + "type": "text", + "content": "Table 2. Interpolation performance of the midmost frames on ACDC cardiac. The choice of midmost frames: if the frame number " + }, + { + "bbox": [ + 55, + 314, + 296, + 360 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 314, + 296, + 360 + ], + "type": "text", + "content": " is odd, three intermediate frames are selected, otherwise only two frames are chosen." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 64, + 366, + 290, + 434 + ], + "lines": [ + { + "bbox": [ + 64, + 366, + 290, + 434 + ], + "spans": [ + { + "bbox": [ + 64, + 366, + 290, + 434 + ], + "type": "table", + "html": "
ModelPSNR (dB) ↑LPIPS ↓FID ↓FVD ↓
SVIN [13]26.01±0.4672.673±0.36138.9142.8
DDM [29]24.37±0.5112.756±0.33041.5145.1
UVI-Net [30]26.25±0.4102.542±0.37437.0137.1
FB-Diff26.18±0.4132.287±0.30532.7128.9
Δ value0.07 ↓0.255 ↓4.3 ↓8.2 ↓
", + "image_path": "da4ed1f522b999eaef26e04949fe27039f74ff39048bbe97f6a0fa1e5ba4dbe0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 447, + 296, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 296, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 296, + 471 + ], + "type": "text", + "content": "tive PSNR value, which demonstrates the effectiveness of respiration-related modeling via Fourier bases." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 472, + 296, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 296, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 296, + 651 + ], + "type": "text", + "content": "Qualitative Comparison. Qualitative results on the interpolated midmost frame are depicted in Fig.5. SVIN [13], Voxelmorph [2], and DDM [29] leverage the single-direction flow estimation to interpolate intermediate frames with linear motions. Compared with that, UVI-Net [30] constructs the interpolation model with time-domain cycle-consistency, and achieves better reconstruction performance as expected. However, these methods essentially overlook the nonlinear and quasi-periodic motion discipline between temporal sequences, resulting in unsatisfactory interpolated results. Instead, FB-Diff exhibits advantageous predictions for midmost frames under the guidance of various motion patterns, indicating locally anisotropic motions with different intensities. Besides, other diffusion models do not work well due to the lack of fine-grained conditional guidance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 659, + 149, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 149, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 149, + 672 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Key Components. An ablation study is conducted to investigate the significance of various components in FB-Diff as shown in Tab.3. (a) Physiology Motion Priors " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": ". By" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 323, + 344, + 548, + 411 + ], + "blocks": [ + { + "bbox": [ + 313, + 314, + 555, + 337 + ], + "lines": [ + { + "bbox": [ + 313, + 314, + 555, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 555, + 337 + ], + "type": "text", + "content": "Table 3. Ablation study on key components. " + }, + { + "bbox": [ + 313, + 314, + 555, + 337 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 314, + 555, + 337 + ], + "type": "text", + "content": ": physiology motion priors, " + }, + { + "bbox": [ + 313, + 314, + 555, + 337 + ], + "type": "inline_equation", + "content": "\\hat{B}" + }, + { + "bbox": [ + 313, + 314, + 555, + 337 + ], + "type": "text", + "content": ": Fourier motion bases, " + }, + { + "bbox": [ + 313, + 314, + 555, + 337 + ], + "type": "inline_equation", + "content": "\\hat{V}" + }, + { + "bbox": [ + 313, + 314, + 555, + 337 + ], + "type": "text", + "content": ": coarsely inpainted videos." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 344, + 548, + 411 + ], + "lines": [ + { + "bbox": [ + 323, + 344, + 548, + 411 + ], + "spans": [ + { + "bbox": [ + 323, + 344, + 548, + 411 + ], + "type": "table", + "html": "
SettingsACDC Cardiac4D Lung
θˆBˆVPSNR (dB) ↑LPIPS ↓PSNR (dB) ↑LPIPS ↓
X✓✓29.761.71029.532.413
XX✓28.172.10528.402.971
✓✓X30.291.65529.712.246
✓✓✓30.951.45230.182.066
", + "image_path": "f006ddcb334f88c13fa38ae31482e173fdb0836eb48bf0dd0ce0e5815b0219d4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 425, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 555, + 616 + ], + "type": "text", + "content": "removing " + }, + { + "bbox": [ + 313, + 425, + 555, + 616 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 425, + 555, + 616 + ], + "type": "text", + "content": " from the Fourier motion operator, interpolation performance decreases significantly due to the loss of general spectral characteristics from respiratory data. Therefore, Fourier bases lack the precise phase information. (b) Fourier bases " + }, + { + "bbox": [ + 313, + 425, + 555, + 616 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{B}}" + }, + { + "bbox": [ + 313, + 425, + 555, + 616 + ], + "type": "text", + "content": ". Without " + }, + { + "bbox": [ + 313, + 425, + 555, + 616 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{B}}" + }, + { + "bbox": [ + 313, + 425, + 555, + 616 + ], + "type": "text", + "content": " as conditional guidance, the conditional diffusion model fails to perceive explicit motion patterns with different frequencies, which provide the intensity of anisotropic motions. As a result, the generative model cannot well learn the regular temporal variations, with 2.78 dB " + }, + { + "bbox": [ + 313, + 425, + 555, + 616 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 313, + 425, + 555, + 616 + ], + "type": "text", + "content": " and 1.78 dB " + }, + { + "bbox": [ + 313, + 425, + 555, + 616 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 313, + 425, + 555, + 616 + ], + "type": "text", + "content": " PSNR values on ACDC and 4D Lung. (c) Inpainted Videos " + }, + { + "bbox": [ + 313, + 425, + 555, + 616 + ], + "type": "inline_equation", + "content": "\\hat{V}" + }, + { + "bbox": [ + 313, + 425, + 555, + 616 + ], + "type": "text", + "content": ". Here the coarsely reconstructed sequences serve as the output by the finedtuned VAE as revealed in Fig.2. And the joint optimization process boosts the spatial and temporal representations of " + }, + { + "bbox": [ + 313, + 425, + 555, + 616 + ], + "type": "inline_equation", + "content": "\\hat{V}" + }, + { + "bbox": [ + 313, + 425, + 555, + 616 + ], + "type": "text", + "content": ". Consequently, discarding this component will reduce fine-grained details of the diffusion model's output." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "Hyper-settings of Fourier Bases. (a) The number of Fourier Bases " + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "inline_equation", + "content": "N_{Fb}" + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": ". We first analyze the quantitative effects of the frequency number on the interpolation performance. As revealed in Tab.4, reducing the frequency number will cause degraded performance of the temporal simulation. Specifically, 1.30 dB lower PSNR and 0.535 higher LPIPS values are induced by decreasing " + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "inline_equation", + "content": "N_{Fb}" + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": " from 16 to 8. That phenomenon indicates that some elements in the group of" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28016" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 70, + 542, + 141 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 542, + 141 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 542, + 141 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 542, + 141 + ], + "type": "image", + "image_path": "c70cbbc9d0193a624785ed0adda338f619441ab48549cd94deb117f6ac5a06fa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 148, + 555, + 183 + ], + "lines": [ + { + "bbox": [ + 54, + 148, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 54, + 148, + 555, + 183 + ], + "type": "text", + "content": "Figure 6. Visualizations on the frequency spectral of Fourier bases. " + }, + { + "bbox": [ + 54, + 148, + 555, + 183 + ], + "type": "inline_equation", + "content": "V_{0}" + }, + { + "bbox": [ + 54, + 148, + 555, + 183 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 148, + 555, + 183 + ], + "type": "inline_equation", + "content": "V_{N - 1}" + }, + { + "bbox": [ + 54, + 148, + 555, + 183 + ], + "type": "text", + "content": ": starting and ending frames. " + }, + { + "bbox": [ + 54, + 148, + 555, + 183 + ], + "type": "inline_equation", + "content": "f_{1} - f_{8}" + }, + { + "bbox": [ + 54, + 148, + 555, + 183 + ], + "type": "text", + "content": " refer to different motion patterns, which highlight different regions under different frequency. There also appear locally anisotropic motion intensities inside a specific motion pattern." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 232, + 296, + 306 + ], + "blocks": [ + { + "bbox": [ + 55, + 201, + 295, + 224 + ], + "lines": [ + { + "bbox": [ + 55, + 201, + 295, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 295, + 224 + ], + "type": "text", + "content": "Table 4. Ablation study on the hyper-settings of Fourier bases. " + }, + { + "bbox": [ + 55, + 201, + 295, + 224 + ], + "type": "inline_equation", + "content": "N_{Fb}" + }, + { + "bbox": [ + 55, + 201, + 295, + 224 + ], + "type": "text", + "content": ": the frequency number of Fourier bases." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 232, + 296, + 306 + ], + "lines": [ + { + "bbox": [ + 57, + 232, + 296, + 306 + ], + "spans": [ + { + "bbox": [ + 57, + 232, + 296, + 306 + ], + "type": "table", + "html": "
SettingsACDC Cardiac4D Lung
PSNR (dB) ↑LPIPS ↓PSNR (dB) ↑LPIPS ↓
\\(N_{Fb}\\)=16 (Ours)30.951.45230.182.066
\\(N_{Fb}\\)=1230.761.64029.972.570
\\(N_{Fb}\\)=829.651.98729.502.745
w/o Real part29.181.87229.562.453
w/o Imaginary part29.881.68129.672.401
", + "image_path": "e1109913f67f4e7e5b943dde274bf8c0476fa60b1a7c67b26c2d1d3363d2b6ea.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 319, + 296, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 296, + 428 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 296, + 428 + ], + "type": "text", + "content": "bases with the frequency ranging from " + }, + { + "bbox": [ + 55, + 319, + 296, + 428 + ], + "type": "inline_equation", + "content": "f_{8}" + }, + { + "bbox": [ + 55, + 319, + 296, + 428 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 319, + 296, + 428 + ], + "type": "inline_equation", + "content": "f_{15}" + }, + { + "bbox": [ + 55, + 319, + 296, + 428 + ], + "type": "text", + "content": " are consistent with specific frequencies of the respiratory process. (b) Real & Imaginary Parts. " + }, + { + "bbox": [ + 55, + 319, + 296, + 428 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{B}}" + }, + { + "bbox": [ + 55, + 319, + 296, + 428 + ], + "type": "text", + "content": " bears real and imaginary parts after combining generic priors and case-specific spectral representations, which perform as the phase and magnitude information. Eliminating either the real or imaginary part will convert the complex data type into the real one, resulting in the loss of phase information and influencing the temporal interpolation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 433, + 183, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 183, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 183, + 445 + ], + "type": "text", + "content": "4.4. Visualization Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 450, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 666 + ], + "type": "text", + "content": "Visualization of Fourier bases. To further validate the effectiveness and interpretability of our proposed Fourier bases, we visualize extracted motion bases from the most shadow layer. As revealed in Fig.6, " + }, + { + "bbox": [ + 55, + 450, + 295, + 666 + ], + "type": "inline_equation", + "content": "f_{0} - f_{7}" + }, + { + "bbox": [ + 55, + 450, + 295, + 666 + ], + "type": "text", + "content": " refers to the magnitude of the first eight motion bases in ascending order of frequency. Specifically, " + }, + { + "bbox": [ + 55, + 450, + 295, + 666 + ], + "type": "inline_equation", + "content": "f_{0}" + }, + { + "bbox": [ + 55, + 450, + 295, + 666 + ], + "type": "text", + "content": " shows a low-frequency motion pattern highlighting no region, which is strongly related to static backgrounds. " + }, + { + "bbox": [ + 55, + 450, + 295, + 666 + ], + "type": "inline_equation", + "content": "f_{1} - f_{3}" + }, + { + "bbox": [ + 55, + 450, + 295, + 666 + ], + "type": "text", + "content": " exhibit a large motion intensity in the highlighted right and left ventricle regions. Those regions are indeed dynamic foregrounds requiring special attention. " + }, + { + "bbox": [ + 55, + 450, + 295, + 666 + ], + "type": "inline_equation", + "content": "f_{4}" + }, + { + "bbox": [ + 55, + 450, + 295, + 666 + ], + "type": "text", + "content": " depicts an activation in the myocardial region, which contracts inward along the circumferential direction. " + }, + { + "bbox": [ + 55, + 450, + 295, + 666 + ], + "type": "inline_equation", + "content": "f_{5} - f_{7}" + }, + { + "bbox": [ + 55, + 450, + 295, + 666 + ], + "type": "text", + "content": " highlight regions with vast deformations in the temporal domain. Particularly, " + }, + { + "bbox": [ + 55, + 450, + 295, + 666 + ], + "type": "inline_equation", + "content": "f_{6}" + }, + { + "bbox": [ + 55, + 450, + 295, + 666 + ], + "type": "text", + "content": " illustrates a large motion on boundaries of anatomical structures. In summary, these motion patterns reveal locally anisotropic motion intensities, and are harmonized with specific frequencies of the respiratory process, thus promoting the temporal interpolation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": "Visualization of temporal anatomical variations. To demonstrate the effectiveness of FB-Diff on temporal simulation, we visualize temporal variation maps for a perceptual comparison. Fig.7 provides interpolated intermediate" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 325, + 201, + 545, + 397 + ], + "blocks": [ + { + "bbox": [ + 325, + 201, + 545, + 397 + ], + "lines": [ + { + "bbox": [ + 325, + 201, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 325, + 201, + 545, + 397 + ], + "type": "image", + "image_path": "bfd652572d8a4fb20e6f20ddca808de4f098967ae86b12ade2703c4b76e21236.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 406, + 555, + 429 + ], + "lines": [ + { + "bbox": [ + 313, + 406, + 555, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 555, + 429 + ], + "type": "text", + "content": "Figure 7. Temporal variation comparison between FB-Diff and existing methods with the linear motion hypothesis." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 434, + 555, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 555, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 555, + 507 + ], + "type": "text", + "content": "frames at different timepoints. According to the referenced video, the anatomical motion in the temporal domain is nonlinear, which cannot be precisely simulated by existing methods with the linear motion hypothesis. In contrast, our predictions show better temporal consistency and continuity, particularly for frames at " + }, + { + "bbox": [ + 313, + 434, + 555, + 507 + ], + "type": "inline_equation", + "content": "t = 0.6" + }, + { + "bbox": [ + 313, + 434, + 555, + 507 + ], + "type": "text", + "content": " and 0.8." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 523, + 388, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 523, + 388, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 523, + 388, + 537 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 545, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 715 + ], + "type": "text", + "content": "Our proposed FB-Diff resolves the temporal interpolation of 4D medical imaging from a frequency perspective, and can model the respiration-related process with nonlinear and quasi-periodic motions. A Fourier motion operator is devised to extract Fourier bases by incorporating physiological motion priors and case-specific spectral information in the feature space. And the diffusion model further leverages Fourier bases and two prompting frames to achieve the precise interpolation. Extensive results show that FB-Diff achieves SOTA perceptual performance with better temporal consistency while maintaining promising reconstruction metrics. Ultimately, it requires more conditional guidance such as electrocardiogram signals to simulate the highly unstable breathing, which is promising in future work." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28017" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 133 + ], + "type": "text", + "content": "Acknowledgement. We thank the anonymous reviewers for their thorough and constructive feedbacks. This research was supported in part by the project of the National Natural Science Foundation of China (No. 62376153), in part by the institute of Munich Center of Machine Learning." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 144, + 115, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 144, + 115, + 156 + ], + "spans": [ + { + "bbox": [ + 56, + 144, + 115, + 156 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 163, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 163, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 163, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 61, + 163, + 294, + 206 + ], + "type": "text", + "content": "[1] Juyoung Bae, Elizabeth Tong, and Hao Chen. Conditional diffusion model for versatile temporal inpainting in 4d cerebral ct perfusion imaging. In MICCAI, pages 67-77. Springer, 2024. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 209, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 209, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 61, + 209, + 294, + 252 + ], + "type": "text", + "content": "[2] Guha Balakrishnan, Amy Zhao, Mert R Sabuncu, John Guttag, and Adrian V Dalca. Voxelmorph: a learning framework for deformable medical image registration. IEEE transactions on medical imaging, 38(8):1788-1800, 2019. 1, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 254, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 254, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 62, + 254, + 294, + 308 + ], + "type": "text", + "content": "[3] Wenbo Bao, Wei-Sheng Lai, Xiaoyun Zhang, Zhiyong Gao, and Ming-Hsuan Yang. Memc-net: Motion estimation and motion compensation driven neural network for video interpolation and enhancement. IEEE transactions on pattern analysis and machine intelligence, 43(3):933-948, 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 310, + 294, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 310, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 62, + 310, + 294, + 386 + ], + "type": "text", + "content": "[4] Olivier Bernard, Alain Lalande, Clement Zotti, Frederick Cervenansky, Xin Yang, Pheng-Ann Heng, Irem Cetin, Karim Lekadir, Oscar Camara, Miguel Angel Gonzalez Ballester, et al. Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: is the problem solved? IEEE transactions on medical imaging, 37 (11):2514–2525, 2018. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 388, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 388, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 388, + 294, + 430 + ], + "type": "text", + "content": "[5] Rhydian Caines, Naomi K Sisson, and Carl G Rowbottom. 4dct and vmat for lung patients with irregular breathing. Journal of Applied Clinical Medical Physics, 23(1):e13453, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 433, + 294, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 433, + 294, + 475 + ], + "spans": [ + { + "bbox": [ + 62, + 433, + 294, + 475 + ], + "type": "text", + "content": "[6] Junyu Chen, Eric C Frey, Yufan He, William P Segars, Ye Li, and Yong Du. Transmorph: Transformer for unsupervised medical image registration. Medical image analysis, 82:102615, 2022. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 478, + 294, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 478, + 294, + 522 + ], + "spans": [ + { + "bbox": [ + 62, + 478, + 294, + 522 + ], + "type": "text", + "content": "[7] Tingxiu Chen, Yilei Shi, Zixuan Zheng, Bingcong Yan, Jingliang Hu, Xiao Xiang Zhu, and Lichao Mou. Ultrasound image-to-video synthesis via latent dynamic diffusion models. In MICCAI, pages 764-774. Springer, 2024. 4, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 523, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 523, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 523, + 294, + 555 + ], + "type": "text", + "content": "[8] Myungsub Choi, Heewon Kim, Bohyung Han, Ning Xu, and Kyoung Mu Lee. Channel attention is all you need for video frame interpolation. In AAAI, pages 10663-10671, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 557, + 294, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 557, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 62, + 557, + 294, + 610 + ], + "type": "text", + "content": "[9] Myungsub Choi, Suyoung Lee, Heewon Kim, and Kyoung Mu Lee. Motion-aware dynamic architecture for efficient frame interpolation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13839-13848, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 613, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 294, + 656 + ], + "type": "text", + "content": "[10] Duolikun Danier, Fan Zhang, and David Bull. Ldmvfi: Video frame interpolation with latent diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1472-1480, 2024. 1, 4, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 294, + 713 + ], + "type": "text", + "content": "[11] Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick Van Der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In ICCV, pages 2758-2766, 2015. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "type": "text", + "content": "[12] Jan Ehrhardt, Cristian Lorenz, et al. 4D modeling and estimation of respiratory motion for radiation therapy. Springer, 2013. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 150 + ], + "type": "text", + "content": "[13] Yuyu Guo, Lei Bi, Euijoon Ahn, Dagan Feng, Qian Wang, and Jinman Kim. A spatiotemporal volumetric interpolation network for 4d dynamic medical image. In CVPR, pages 4726-4735, 2020. 1, 3, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 553, + 195 + ], + "type": "text", + "content": "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 30, 2017. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 198, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 230 + ], + "type": "text", + "content": "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 4, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 232, + 553, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 232, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 553, + 263 + ], + "type": "text", + "content": "[16] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. NeurIPS, 35:8633-8646, 2022. 2, 4, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 265, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 265, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 317, + 265, + 553, + 319 + ], + "type": "text", + "content": "[17] Kan N Hor, Rolf Baumann, Gianni Pedrizzetti, Gianni Tonti, William M Gottliebson, Michael Taylor, D Woodrow Benson, and Wojciech Mazur. Magnetic resonance derived myocardial strain assessment using feature tracking. Journal of visualized experiments: JoVE, (48):2356, 2011. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 322, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 322, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 317, + 322, + 553, + 354 + ], + "type": "text", + "content": "[18] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssim. In 2010 20th international conference on pattern recognition, pages 2366-2369. IEEE, 2010. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 355, + 553, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 355, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 317, + 355, + 553, + 419 + ], + "type": "text", + "content": "[19] Mengshun Hu, Kui Jiang, Liang Liao, Jing Xiao, Junjun Jiang, and Zheng Wang. Spatial-temporal space hand-in-hand: Spatial-temporal video super-resolution via cycle-projected mutual learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3574-3583, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 422, + 555, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 555, + 475 + ], + "type": "text", + "content": "[20] Mengshun Hu, Kui Jiang, Zhihang Zhong, Zheng Wang, and Yinqiang Zheng. Iq-vfi: Implicit quadratic motion estimation for video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6410-6419, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 478, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 478, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 317, + 478, + 553, + 521 + ], + "type": "text", + "content": "[21] Zhewei Huang, Tianyuan Zhang, Wen Heng, Boxin Shi, and Shuchang Zhou. Real-time intermediate flow estimation for video frame interpolation. In European Conference on Computer Vision, pages 624-642. Springer, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 523, + 553, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 523, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 317, + 523, + 553, + 564 + ], + "type": "text", + "content": "[22] Geoffrey D Hugo, Elisabeth Weiss, William C Sleeman, Salim Balik, Paul J Keall, Jun Lu, and Jeffrey F Williamson. Data from 4d lung imaging of nsclc patients. (No Title), 2016. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 567, + 553, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 567, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 317, + 567, + 553, + 611 + ], + "type": "text", + "content": "[23] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In CVPR, pages 2462-2470, 2017. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 613, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 553, + 645 + ], + "type": "text", + "content": "[24] Siddhant Jain, Daniel Watson, Eric Tabellion, Ben Poole, Janne Kontkanen, et al. Video interpolation with diffusion models. In CVPR, pages 7341-7351, 2024. 1, 4, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 647, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 679 + ], + "type": "text", + "content": "[25] Mi-Young Jeung, Philippe Germain, et al. Myocardial tagging with mr imaging: overview of normal and pathologic findings. Radiographics, 32(5):1381-1398, 2012. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "type": "text", + "content": "[26] Huaizu Jiang, Deqing Sun, Varun Jampani, Ming-Hsuan Yang, Erik Learned-Miller, and Jan Kautz. Super slomo: High quality estimation of multiple intermediate frames for" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28018" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 715 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 75, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 296, + 106 + ], + "type": "text", + "content": "video interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9000-9008, 2018. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 150 + ], + "type": "text", + "content": "[27] Xin Jin, Longhai Wu, Jie Chen, Youxin Chen, Jayoon Koo, and Cheul-hee Hahm. A unified pyramid recurrent network for video frame interpolation. In CVPR, pages 1578-1587, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 296, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 296, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 296, + 206 + ], + "type": "text", + "content": "[28] Tarun Kalluri, Deepak Pathak, Manmohan Chandraker, and Du Tran. Flavr: Flow-agnostic video representations for fast frame interpolation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2071–2082, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 295, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 295, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 295, + 241 + ], + "type": "text", + "content": "[29] Boah Kim and Jong Chul Ye. Diffusion deformable model for 4d temporal medical image generation. In MICCAI, pages 539-548. Springer, 2022. 1, 4, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 243, + 295, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 295, + 286 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 295, + 286 + ], + "type": "text", + "content": "[30] JungEun Kim, Hangyul Yoon, Geondo Park, Kyungsu Kim, and Eunho Yang. Data-efficient unsupervised interpolation without any intermediate frame for 4d medical images. In CVPR, pages 11353-11364, 2024. 1, 4, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 287, + 295, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 287, + 295, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 287, + 295, + 353 + ], + "type": "text", + "content": "[31] Lingtong Kong, Boyuan Jiang, Donghao Luo, Wenqing Chu, Xiaoming Huang, Ying Tai, Chengjie Wang, and Jie Yang. Ifrnet: Intermediate feature refine network for efficient frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1969-1978, 2022. 1, 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 295, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 295, + 399 + ], + "type": "text", + "content": "[32] Xia Li, Runzhao Yang, Xiangtai Li, Antony Lomax, Ye Zhang, and Joachim Buhmann. Cpt-interp: Continuous spatial and temporal motion modeling for 4d medical image interpolation. arXiv preprint arXiv:2405.15385, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 295, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 295, + 443 + ], + "type": "text", + "content": "[33] Zhen Li, Zuo-Liang Zhu, Ling-Hao Han, Qibin Hou, Chun-Le Guo, and Ming-Ming Cheng. Amt: All-pairs multi-field transforms for efficient frame interpolation. In CVPR, pages 9801-9810, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 444, + 295, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 444, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 444, + 295, + 498 + ], + "type": "text", + "content": "[34] Hanwen Liang, Yuyang Yin, Dejia Xu, Hanxue Liang, Zhangyang Wang, Konstantinos N Plataniotis, Yao Zhao, and Yunchao Wei. Diffusion4d: Fast spatial-temporal consistent 4d generation via video diffusion models. arXiv preprint arXiv:2405.16645, 2024. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 500, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 295, + 555 + ], + "type": "text", + "content": "[35] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv preprint arXiv:2402.17177, 2024. 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 556, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 556, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 556, + 295, + 590 + ], + "type": "text", + "content": "[36] Liying Lu, Ruizheng Wu, Huaijia Lin, Jiangbo Lu, and Jiaya Jia. Video frame interpolation with transformer. In CVPR, pages 3532-3542, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 590, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 295, + 646 + ], + "type": "text", + "content": "[37] Simone Meyer, Abdelaziz Djerouah, Brian McWilliams, Alexander Sorkine-Hornung, Markus Gross, and Christopher Schroers. Phasenet for video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 498-507, 2018. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 295, + 690 + ], + "type": "text", + "content": "[38] Kotaro Mizuno and Masahiro Muto. Preoperative evaluation of pleural adhesion in patients with lung tumors using four-dimensional computed tomography performed during natural breathing. *Medicine*, 100(47):e27800, 2021. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 295, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 295, + 715 + ], + "type": "text", + "content": "[39] Simon Niklaus and Feng Liu. Softmax splatting for video frame interpolation. In Proceedings of the IEEE/CVF con" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 715 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "ference on computer vision and pattern recognition, pages 5437-5446, 2020. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "text", + "content": "[40] David Ouyang, Bryan He, Amirata Ghorbani, Neal Yuan, Joseph Ebinger, Curtis P Langlotz, Paul A Heidenreich, Robert A Harrington, David H Liang, Euan A Ashley, et al. Video-based ai for beat-to-beat assessment of cardiac function. Nature, 580(7802):252-256, 2020. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "text", + "content": "[41] Junheum Park, Jintae Kim, and Chang-Su Kim. Biformer: Learning bilateral motion estimation via bilateral transformer for 4k video frame interpolation. In CVPR, pages 1568-1577, 2023. 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 197, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 252 + ], + "type": "text", + "content": "[42] Saikat Roy, Gregor Koehler, Constantin Ulrich, Michael Baumgartner, Jens Petersen, Fabian Isensee, Paul F Jaeger, and Klaus H Maier-Hein. Mednext: transformer-driven scaling of convnets for medical image segmentation. In MICCAI, pages 405–415. Springer, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "type": "text", + "content": "[43] Otto A Smiseth, Hans Torp, Anders Opdahl, Kristina H Haugaa, and Stig Urheim. Myocardial strain imaging: how useful is it in clinical decision making? European heart journal, 37 (15):1196-1207, 2016. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 297, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 331 + ], + "type": "text", + "content": "[44] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In CVPR, pages 8934-8943, 2018. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 332, + 553, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 365 + ], + "type": "text", + "content": "[45] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, pages 402-419. Springer, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 366, + 553, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 366, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 316, + 366, + 553, + 388 + ], + "type": "text", + "content": "[46] Georgi P Tolstov. Fourier series. Courier Corporation, 2012. 2, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 388, + 553, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 433 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 433 + ], + "type": "text", + "content": "[47] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 434, + 553, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 488 + ], + "type": "text", + "content": "[48] Bingchao Wang, Zhiwei Ning, Jianyu Ding, Xuanang Gao, Yin Li, Dongsheng Jiang, Jie Yang, and Wei Liu. Fix-clip: Dual-branch hierarchical contrastive learning via synthetic captions for better understanding of long text. arXiv preprint arXiv:2507.10095, 2025. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 489, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 489, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 489, + 553, + 555 + ], + "type": "text", + "content": "[49] Lu Wang, Shelly Hayes, Kamen Paskalev, Lihui Jin, Mark K Buyyounouski, Charlie C-M Ma, and Steve Feigenberg. Dosimetric comparison of stereotactic body radiotherapy using 4d ct and multiphase ct images for treatment planning of lung cancer: evaluation of the impact on daily dose coverage. *Radiotherapy and Oncology*, 91(3):314–324, 2009. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 556, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 556, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 556, + 553, + 590 + ], + "type": "text", + "content": "[50] Yihan Wang, Lahav Lipson, and Jia Deng. Sea-raft: Simple, efficient, accurate raft for optical flow. In ECCV, pages 36-54. Springer, 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "type": "text", + "content": "[51] Tzu-Ti Wei, Chin Kuo, Yu-Chee Tseng, and Jen-Jee Chen. Mpvf: 4d medical image inpainting by multi-pyramid voxel flows. IEEE Journal of Biomedical and Health Informatics, 2023. 1, 2, 3, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 635, + 553, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 680 + ], + "type": "text", + "content": "[52] Xin You, Ming Ding, Minghui Zhang, Yangqian Wu, Yi Yu, Yun Gu, and Jie Yang. Semantic difference guidance for the uncertain boundary segmentation of ct left atrial appendage. In MICCAI, pages 121-131. Springer, 2023. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 681, + 553, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 715 + ], + "type": "text", + "content": "[53] Xin You, Junjun He, Jie Yang, and Yun Gu. Learning with explicit shape priors for medical image segmentation. IEEE Transactions on Medical Imaging, 2024. 6" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28019" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 496 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[54] Xin You, Yixin Lou, Minghui Zhang, Jie Yang, and Yun Gu. Slord: Structural low-rank descriptors for shape consistency in vertebrae segmentation. IEEE Journal of Biomedical and Health Informatics, 2025. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "text", + "content": "[55] Xin You, Minghui Zhang, Hanxiao Zhang, Jie Yang, and Nassir Navab. Temporal differential fields for 4d motion modeling via image-to-video synthesis. arXiv preprint arXiv:2505.17333, 2025. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "text", + "content": "[56] Weihao Yu, Yuanhao Cai, Ruyi Zha, Zhiwen Fan, Chenxin Li, and Yixuan Yuan. X" + }, + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "text", + "content": "-gaussian: 4d radiative gaussian splatting for continuous-time tomographic reconstruction. arXiv preprint arXiv:2503.21779, 2025. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "type": "text", + "content": "[57] Chuyan Zhang, Hao Zheng, Xin You, Yefeng Zheng, and Yun Gu. Pass: test-time prompting to adapt styles and semantic shapes in medical image segmentation. IEEE Transactions on Medical Imaging, 2024. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 294, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 317 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 317 + ], + "type": "text", + "content": "[58] Guozhen Zhang, Yuhan Zhu, Haonan Wang, Youxin Chen, Gangshan Wu, and Limin Wang. Extracting motion and appearance via inter-frame attention for efficient video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5682-5692, 2023. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 319, + 294, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 294, + 361 + ], + "type": "text", + "content": "[59] Haiyu Zhang, Xinyuan Chen, Yaohui Wang, Xihui Liu, Yunhong Wang, and Yu Qiao. 4diffusion: Multi-view video diffusion model for 4d generation. NeurIPS, 37:15272-15295, 2025. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 364, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 364, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 364, + 294, + 407 + ], + "type": "text", + "content": "[60] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 409, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 409, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 409, + 294, + 453 + ], + "type": "text", + "content": "[61] Zhihang Zhong, Gurunandan Krishnan, Xiao Sun, Yu Qiao, Sizhuo Ma, and Jian Wang. Clearer frames, anytime: Resolving velocity ambiguity in video frame interpolation. In ECCV, pages 346-363. Springer, 2024. 1, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 454, + 294, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 454, + 294, + 496 + ], + "spans": [ + { + "bbox": [ + 56, + 454, + 294, + 496 + ], + "type": "text", + "content": "[62] Kun Zhou, Wenbo Li, Xiaoguang Han, and Jiangbo Lu. Exploring motion ambiguity and alignment for high-quality video frame interpolation. In CVPR, pages 22169-22179, 2023. 1" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28020" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/FDPT_ Federated Discrete Prompt Tuning for Black-Box Visual-Language Models/5d5a1b2f-9957-4fb4-b461-7bf3e1ebccfb_content_list.json b/2025/FDPT_ Federated Discrete Prompt Tuning for Black-Box Visual-Language Models/5d5a1b2f-9957-4fb4-b461-7bf3e1ebccfb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..43109ce513b7a44d57ef7f678ecbd38aef8a95f4 --- /dev/null +++ b/2025/FDPT_ Federated Discrete Prompt Tuning for Black-Box Visual-Language Models/5d5a1b2f-9957-4fb4-b461-7bf3e1ebccfb_content_list.json @@ -0,0 +1,1400 @@ +[ + { + "type": "text", + "text": "FDPT: Federated Discrete Prompt Tuning for Black-Box Visual-Language Models", + "text_level": 1, + "bbox": [ + 125, + 128, + 872, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiaqi Wu $^{1,8,10\\dagger}$ , Simin Chen $^{2\\dagger}$ , Jing Tang $^{3}$ , Yuzhe YANG $^{4}$ , Yiming Chen $^{5}$ , Lixu Wang $^{6}$ , Song Lin $^{7}$ , Zehua Wang $^{8,9}$ , Wei Chen $^{9*}$ , and Zijian Tian $^{10}$", + "bbox": [ + 109, + 202, + 885, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Tsinghua University, $^{2}$ University of Texas at Dallas, $^{3}$ Huazhong University of Science and Technology,", + "bbox": [ + 163, + 241, + 808, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4The Chinese University of Hong Kong, 5National University of Singapore, 6Nanyang Technological University,", + "bbox": [ + 163, + 257, + 831, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "7Beijing Baidu Netcom Science Technology Co., Ltd., 8University of British Columbia,", + "bbox": [ + 236, + 275, + 759, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{9}$ China University of Mining and Technology, $^{10}$ China University of Mining and Technology (Beijing)", + "bbox": [ + 163, + 292, + 802, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "wjq11346@student.ubc.ca, sxc180080@utdallas.edu, focusers@163.com, yuzheyang@link.cuhk.edu.cn,", + "bbox": [ + 150, + 311, + 841, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yiming.chen@u.nus.edu, wanglixu4334@gmail.com, linsong04@baidu.com, zwang@ece.ubc.ca,", + "bbox": [ + 183, + 329, + 810, + 344 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "chenwdavior@163.com，Tianzj0726@126.com", + "bbox": [ + 349, + 347, + 640, + 361 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 397, + 326, + 412 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-Language Models (VLMs) have driven significant advancements in multimodal AI. Fine-tuning these models with user data enhances adaptability but poses privacy risks. While federated learning (FL) mitigates user data privacy concerns, it fails to protect model properties. Existing methods relying on black-box VLM APIs often require access to prediction logits, making them vulnerable to inversion attacks. Additionally, optimizing tuning complexity and data transmission efficiency in federated VLM scenarios remains a challenge. To address these challenges, we propose FDPT—the first federated discrete prompt tuning method utilizing black-box VLMs. During client optimization stage, FDPT employs an agent-driven framework leveraging large language models (LLMs) with enhanced reasoning capacities to systematically optimize discrete prompt representations, and also utilizes feedback mechanisms and chain of thought to enhance prediction accuracy. Importantly, it performs optimization by relying not on the predicted logic vectors output by LLMs but on textual results, avoiding reverse attack risks. During global aggregation stage, We mimic human electoral activities by employing evolutionary computation methods underpinned by semantic similarity computation to implement enhanced zero-order optimization for acquiring representative global tokens, thereby achieving knowledge aggregation. FDPT significantly outperforms nine state-of-the-art methods in image classification and visual question-answering, reducing communication overhead while generating highly transferable optimized prompts. Additionally, it exhibits im", + "bbox": [ + 88, + 428, + 485, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "proved robustness to data heterogeneity.", + "bbox": [ + 511, + 398, + 779, + 414 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 446, + 645, + 463 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "General-purpose Vision-Language Models (VLMs) [45] integrate vast visual and linguistic information, advancing various vision tasks [36]. To better adapt to user needs, these models can undergo fine-tuning or prompt-tuning [27], but such modifications introduce privacy risks [13, 18, 34, 38]. Federated Learning (FL) [22] addresses these concerns by decentralizing data storage, reducing single-point breaches, and exchanging model weights or embeddings instead of raw data. Various FL-based approaches [19, 29, 39] have been explored for tuning large models, effectively preserving user data privacy.", + "bbox": [ + 509, + 474, + 906, + 641 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the success of FL, protecting the privacy of VLMs themselves remains a critical yet largely unexplored challenge. Existing FL approaches use zero-order optimization to estimate gradients for prompt tuning on black-box VLMs [31, 41]. For example, FedBPT [31] employs prompt exchange and gradient-free optimization, reducing communication and memory costs by avoiding direct parameter access and backpropagation on client devices. However, several limitations remain. First, these methods remain vulnerable to security risks as they rely on prediction logits, exposing them to inversion attacks that can infer model parameters [7, 47]. Second, zero-order optimization techniques [31, 41] often suffer from limited accuracy and adaptability in prompt fine-tuning. Finally, soft prompt methods [15, 46] lack interpretability and incur high communication costs in federated settings.", + "bbox": [ + 511, + 642, + 908, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address these challenges, we propose FDPT, the first", + "bbox": [ + 532, + 885, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Equal contribution", + "bbox": [ + 107, + 875, + 217, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author", + "bbox": [ + 111, + 887, + 233, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2461", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "federated discrete prompt optimization method comprising two stages: client-side optimization and global aggregation. In the client-side optimization phase, we introduce AI Agent-based Client Prompt Optimization, which refines token-level prompts using large language models (LLMs). We design a feedback mechanism and Chain-of-Thought measures to guide the agent in selecting optimal discrete tokens, overcoming the limitations of existing zero-order optimization methods in prompt tuning. Unlike traditional approaches that rely on prediction logits, our method uses only black-box VLM text outputs, reducing the risk of inversion attacks and enhancing model privacy protection. Afterwards, in the global aggregation phase, to reduce communication overhead in federated settings, we use discrete tokens for knowledge exchange instead of soft prompts. Inspired by human voting mechanisms, we propose a Self-Attention-Guided Evolutionary Computation method to identify representative tokens for constructing global prompts. Unlike the coarse sampling in traditional zero-order optimization, our approach clusters tokens in semantic space and then applies similarity calculations to refine the most accurate results, enhancing tuning efficiency.", + "bbox": [ + 89, + 90, + 480, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Detailed experimental results demonstrate the advantages of our method. In Visual Question Answering (VQA) tasks, under the GPT-4 black-box setting, our method outperforms the classic black-box prompt optimization method FedBPT [31], achieving $11.1\\%$ and $7.4\\%$ improvements on the Art and Chemistry categories of the MMMU dataset [43], respectively. In traditional computer vision (CV) classification tasks, using the Qwen2-VL-72B model [3], our approach achieves $71.8\\%$ accuracy on CIFAR-100, surpassing FedBPT by $5.3\\%$ , demonstrating its superior performance. Additionally, FDPT significantly reduces communication costs without compromising optimization performance and enhances interpretability through discrete prompts. The contributions of this paper are as follows:", + "bbox": [ + 89, + 449, + 482, + 660 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose FDPT, a first federated black-box discrete prompt optimization framework that aggregates knowledge from multiple parties to optimize prompts while ensuring the privacy of both users and easily overlooked VLMs.", + "- We introduce an AI Agent-based Client Prompt Optimization for token-level prompts and a Self-attention-guided Evolutionary Computation method for selecting global tokens, offering better predictive performance than traditional zero-order optimization.", + "- Extensive experiments demonstrate that our method achieves superior performance across various text-vision tasks, exhibits low communication overhead, and is resilient to data heterogeneity." + ], + "bbox": [ + 89, + 686, + 478, + 893 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Background & Related Work", + "text_level": 1, + "bbox": [ + 513, + 89, + 782, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual Language Models. Visual-Language Models (VLMs) [5, 45] are AI systems designed to process and understand both visual and textual information, enabling them to perform tasks such as Visual Question Answering (VQA) [43] and traditional visual classification tasks [20]. Traditional VLMs, like Qwen-VL [3], effectively integrate visual features with textual data, excelling in both VQA and image classification tasks. Flamingo [1] leverages few-shot learning capabilities, allowing it to quickly adapt to new tasks and perform well across both VQA and classification. Gemini [32] focuses on reasoning and multimodal understanding, achieving strong results in complex VQA tasks while also performing well in image and text classification. LLaVA [25] benefits from large-scale pretraining, efficiently handling intricate visuals and long-form text, with exceptional performance in both VQA and classification tasks. Newer models like InternVL2 [8], Qwen2-VL [35], and GPT-4o [16] build on these foundations, enhancing multimodal reasoning, improving the fusion of visual and textual data, and offering broader adaptability across both VQA and classification tasks.", + "bbox": [ + 511, + 114, + 903, + 431 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Federated Learning. Federated Learning (FL) [22] is a decentralized approach allowing multiple clients to collaboratively train a shared model without sharing raw data, thus preserving privacy and defending against attacks [6, 24]. Recent works have further improved FL's robustness in handling data heterogeneity [26, 33], label noise [17], and personalization needs [11]. With the rise of large language models (LLMs), Federated Learning (FL) has gained new applications, effectively addressing the challenges of knowledge exchange and privacy protection between independent LLMs. FedPrompt [46] combines FL with prompt tuning, utilizing split aggregation to significantly reduce communication costs. FedBPT [31] presents a federated black-box prompt tuning framework that avoids direct access to model parameters, instead optimizing prompts using gradient-free methods. FedDTPT [40] realizes discrete prompt tuning for Large Language Models (LLMs). However, these methods still face issues such as security risks in large models and high communication overhead. We propose a Federated Discrete Prompt Tuning method, which, compared to existing approaches, performs zero-order optimization at the token level, mitigating the risk of adversarial users leveraging logits generated by the large model for backdoor attacks. Additionally, by utilizing discrete tokens for knowledge exchange, it significantly reduces communication overhead.", + "bbox": [ + 511, + 452, + 903, + 844 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prompt Tuning. Prompt tuning [49] is an emerging technique for adapting large pre-trained models to specific tasks", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2462", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/750758aadf814a2906dd59c770d1ad113ac648049b233a009a000c5306cbf8d5.jpg", + "image_caption": [ + "Figure 1. Overview of FDPT. The proposed framework consists of the client-side local Agent-based Client Prompt Optimization and the server-side Semantic-Similarity-Guided Evolutionary Computation for global selection. The detailed process is described in Sec. 3.2." + ], + "image_footnote": [], + "bbox": [ + 171, + 88, + 823, + 321 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with minimal additional training. In VLMs, prompt tuning [14] bridges visual and textual modalities, enabling tasks like image captioning [42, 44]. The Prompt tuning can be represented as finding an optimal prompt $\\mathbf{p}^*$ to maximize the model's performance on the dataset $\\mathcal{D} = \\{(\\mathbf{x},\\mathbf{y})\\}$ :", + "bbox": [ + 89, + 386, + 482, + 463 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} ^ {*} = \\arg \\max f _ {\\mathrm {V L M}} (\\mathbf {p}; \\mathcal {D}) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 484, + 482, + 501 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $f_{\\mathrm{VLM}}(\\mathbf{p};\\mathcal{D})$ denotes the performance of the VLM using prompt $\\mathbf{p}$ over the dataset $\\mathcal{D}$ . Continual Prompt Tuning (soft prompt tuning) involves fine-tuning a small set of trainable parameters (soft prompts) while keeping the core model fixed. However, it suffers from poor interpretability, and in Federated Learning (FL), it incurs high communication overhead between clients. Discrete Prompt Tuning (DPT), in contrast, optimizes specific discrete tokens from a fixed vocabulary. A notable approach, RLPrompt, uses reinforcement learning (RL) to select the best tokens based on rewards. However, it has limited optimization efficiency due to the computationally intensive nature of RL. Regarding above limitations, the proposed Federated Discrete Prompt Tuning method aims to combine the strengths of discrete prompt tuning with the constraints and needs of Federated Learning. By leveraging an AI agent (based on large language models) with robust reasoning capabilities, the method can generate more accurate and efficient local discrete tokens. Furthermore, by incorporating semantic similarity calculations into an evolutionary computing framework, the method can generate more precise global tokens that are shared across clients. Additionally, this reduces communication costs between clients, as discrete tokens are much smaller in parameters compared to soft prompts.", + "bbox": [ + 91, + 523, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 383, + 604, + 400 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 513, + 409, + 717, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we consider a distributed scenario [22], where $N$ clients each possess a private dataset $\\mathcal{D}_k = \\{(\\mathcal{X}_k,\\mathcal{Y}_k)\\}$ consisting of $n_k$ samples $\\{(x_i^k,y_i^k)\\}_{i\\in [n_k]}$ . Each client can only access its own private data and does not wish to share it with others. This scenario is realistic, as hospitals and other organizations prioritize protecting sensitive data. Moreover, such distributed settings have been widely studied in existing research.", + "bbox": [ + 511, + 431, + 906, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Then, our federated discrete prompt optimization can be formulated as an optimization problem that aims to find the optimal prompt instruction $p^*$ that satisfies the following equation:", + "bbox": [ + 511, + 551, + 906, + 612 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np ^ {*} = \\arg \\min _ {p} \\sum_ {k \\in [ K ]} \\frac {n ^ {k}}{\\sum_ {k \\in [ K ]} n ^ {k}} \\sum_ {i \\in [ n ^ {k} ]} \\mathcal {L} \\left(\\mathrm {V L M} \\left(p; x _ {i} ^ {k}\\right), y _ {i} ^ {k}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 619, + 903, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{L}(\\cdot)$ is the objective function, which could be the error rate, and $\\mathrm{VLM}(\\cdot)$ represents the Vision-Language Model (VLM). The above equation can be interpreted as learning a globally optimal natural language prompt by minimizing the weighted average loss across all clients, where each client's contribution is proportional to its dataset size.", + "bbox": [ + 511, + 674, + 906, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Notice that if clients were permitted to share their private data and had white-box access to the VLM, the equation above could be easily solved by approximating the objective function with a differentiable function, such as cross-entropy loss. In this case, standard backpropagation could be applied to the global dataset to search for the optimal prompt instruction. However, due to commercial concerns, modern VLMs are typically deployed as API ser", + "bbox": [ + 511, + 779, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2463", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "vices, which are black-boxes to the clients. Additionally, each client is unwilling to share their data with others due to privacy concerns. Consequently, we must impose the following constraints:", + "bbox": [ + 89, + 90, + 483, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Blackbox: $\\frac{\\partial\\mathcal{L}(\\mathrm{VLM}(p;x),y)}{\\partial p} = \\emptyset$", + "bbox": [ + 91, + 159, + 357, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Privacy: $\\forall k,j\\in [K],k\\neq j,$ $\\mathrm{Client}_k$ cannot access $\\mathcal{D}_j$ (3)", + "bbox": [ + 91, + 194, + 482, + 226 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The first constraint indicates that the gradient is not feasible, while the second constraint specifies that each client can only access its own data.", + "bbox": [ + 89, + 234, + 483, + 281 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Design Overview", + "text_level": 1, + "bbox": [ + 89, + 289, + 256, + 305 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The algorithm overview is illustrated in Figure 1. Consistent with the traditional federated learning paradigm [22], each client communicates solely with the server using the prompt information, while keeping their data local. This ensures compliance with the data privacy constraints outlined in Eq.(3). Specifically, at the beginning, the server distributes the Global Prompt to each client. Then our approach follows these steps during each global communication round to search an optimal prompt: (1) Client Optimization: Each client utilizes an Agent-based Client Prompt Optimization method for personalized token-level prompt optimization. This method is preferred over backpropagation and ensures adherence to the black-box constraints in Eq.(3), which is more realistic in the context of VLM as a service; (2) Server Aggregation: Each client uploads its locally optimized prompt to the server, where knowledge aggregation is performed using the Semantic-similarity-guided Evolutionary Computation method. The subsequent sections provide a detailed explanation of each step in our approach.", + "bbox": [ + 89, + 311, + 483, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Client Optimization", + "text_level": 1, + "bbox": [ + 89, + 622, + 281, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Upon receiving the global discrete prompt $p_0$ from the server in the current round, each client iteratively updates it using an agent-based approach, as shown in Alg. 1. Initially, the agent queries the LLM to randomly replace some tokens in $p_0$ , generating a set of prompt mutants $p_1, p_2, \\ldots, p_M$ (Line 1). Each prompt is then combined with the client's local dataset and fed into the VLM to measure the accuracy, which serves as the prompt mutant's score. FDPT maintains a priority queue to manage these mutants, prioritizing those with higher accuracy (Lines 2-6). FDPT iteratively selects the highest-priority prompt from the queue as the seed, asks the LLM to mutate the important tokens of the seed, and updates the priority queue with the newly generated prompt mutants (Lines 7-13). Specifically, token importance is computed as follows: (1), FDPT compares each prompt mutant $p_i$ with the seed prompt $p_0$ token by token, constructing a difference matrix $M$ , where $M_i^j = 0$ if the", + "bbox": [ + 89, + 643, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Agent-Based Prompt Optimization" + ], + "code_body": "Input: Initial discrete prompt $p_0$ , Large Language Model (LLM), Vision-Language Model (VLM), client local dataset $\\mathcal{D}$ , number of mutants $M$ \nOutput: Optimized prompt $\\hat{p}$ \n1: Initialization: $\\{p_1, p_2, \\ldots, p_M\\} =$ Mutator( $LLM, p_0$ ) $\\triangleright$ Mutate $p_0$ by randomly replacing tokens \n2: $\\mathcal{Q} =$ Priority Queue $\\triangleright$ Initialize an empty queue \n3: for each prompt $p_i$ in $\\{p_0, p_1, \\ldots, p_M\\}$ do \n4: $Acc_i =$ Evaluate(VLM, $p_i$ , $\\mathcal{D}$ ) $\\triangleright$ Evaluation \n5: Push $(\\mathcal{Q}, p_i, Acc_i)$ $\\triangleright$ Update the priority queue \n6: end for \n7: while stopping criterion not met do \n8: $p_{seed} =$ Pop $(\\mathcal{Q}) \\triangleright$ Get the highest-priority prompt \n9: Position = Importance Estimator $\\triangleright$ Select important tokens \n10: $p_{new} =$ Mutator( $LLM, p_{seed}$ , Position) \n11: $Acc_{new} =$ Evaluate(VLM, $p_{new}$ , $\\mathcal{D}$ ) $\\triangleright$ Evaluation \n12: Push $(\\mathcal{Q}, p_{new}, Acc_{new})$ $\\triangleright$ Update the queue \n13: end while \n14: Return highest-scoring prompt from $\\mathcal{Q}$", + "bbox": [ + 513, + 109, + 906, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$j$ -th token of $p_i$ and $p_0$ are different, and $M_i^j = 1$ otherwise. (2), FDPT computes the accuracy gains $V$ , where $V_i = \\mathrm{Acc}_i - \\mathrm{Acc}_0$ . (3) The importance score of each token position is then given by $S = M^{-1}V$ . This score estimation equation reflects the degree to which a token's presence or absence affects accuracy, weighted by its position in the prompt. Finally, the top $K$ token positions are selected based on the importance score $S$ .", + "bbox": [ + 511, + 458, + 906, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Server Aggregation", + "text_level": 1, + "bbox": [ + 513, + 589, + 699, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After local client optimization, each client uploads its optimized discrete prompt to the server for knowledge exchange. The server then aggregates these local optimal prompts for the next round of optimization. Specifically, the server aggregation includes the following steps (as shown in Algorithm 2):", + "bbox": [ + 511, + 611, + 905, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Token-level Semantic Representation. In total client uploaded prompts, e.g. the initial population $\\mathcal{P}_{\\mathrm{init}}$ , for each token in a prompt, we use a pre-trained text encoder of the Roberta to get its token-level semantic representation, denoted as", + "bbox": [ + 511, + 702, + 905, + 777 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} _ {i} = [ e _ {i} ^ {1}, e _ {i} ^ {2}, \\dots , e _ {i} ^ {K} ] e _ {i} ^ {j} = f _ {e m b e d} (t k _ {i} ^ {j}) \\qquad (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 787, + 906, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{embed}(\\cdot)$ is the pre-trained encoder, and $tk_i^j$ is the $j^{th}$ token of the $i^{th}$ prompt.", + "bbox": [ + 511, + 819, + 905, + 851 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reputation Test. We first evaluate the reputation of each token, defined as the semantic-similarity weight (denoted", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2464", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\alpha_{g,h})$ that each token assigns to all other tokens. This is computed using a cosine-similarity mechanism [12] based on embeddings:", + "bbox": [ + 89, + 90, + 485, + 137 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {g} = \\sum_ {h} \\alpha_ {g, h} = \\sum_ {h} \\operatorname {s o f t m a x} \\left(\\frac {\\mathbf {e} _ {g} \\cdot \\mathbf {e} _ {h}}{\\sqrt {d}}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 147, + 482, + 184 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{e}_g$ and $\\mathbf{e}_h$ represent the embedding of a token $g$ and $h$ , respectively, and $d$ is the dimensionality of the embeddings. A higher semantic-similarity weight $\\alpha_{g,h}$ signifies a stronger influence over other tokens. We define the total semantic-similarity weight of token $g$ as $\\alpha_{g}$ , which represents the token's semantic importance among all other tokens, referred to as its Reputation.", + "bbox": [ + 89, + 195, + 483, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Aggregation. Next, tokens are clustered based on their semantic similarity by grouping embeddings that are close to each other in the high-dimensional semantic space:", + "bbox": [ + 89, + 320, + 483, + 367 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} = \\left\\{\\mathcal {C} _ {1}, \\mathcal {C} _ {2}, \\dots , \\mathcal {C} _ {Z} \\right\\}, \\mathcal {C} _ {z} = \\left\\{\\mathbf {e} _ {h} \\mid \\| \\mathbf {e} _ {h} - \\mathbf {e} _ {g} \\| < \\epsilon , \\forall h \\in \\mathcal {P} _ {\\text {i n i t}} \\right\\}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 377, + 483, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\epsilon$ is a pre-defined distance threshold. $C_z$ denotes a cluster, or Community. Embeddings with equivalent syntactic meanings are grouped into a cluster, from which a single representative is subsequently selected.", + "bbox": [ + 89, + 407, + 483, + 469 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Election. Finally, within each cluster (Community), the representative token is determined by semantic-similarity weights (reputation). Tokens with higher weights have stronger semantic dependencies on other tokens and are more likely to be selected:", + "bbox": [ + 89, + 487, + 483, + 563 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {r} _ {z} = \\underset {\\mathbf {e} _ {n} \\in \\mathcal {C} _ {z}} {\\arg \\max } \\alpha_ {n}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 577, + 482, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{r}_z$ denotes the representative token in Community $\\mathcal{C}_z$ . Importantly, communities containing only a single token are discarded, as these tokens contain only client-specific information, following the principle of \"majority rule\" [30] in Voting. The selected tokens are then arranged in their original order (using, for example, a bubble sort [2]) to form a representative prompt, or the \"Representatives\", denoted as $\\mathcal{P}_{\\mathrm{new}} = \\{\\mathbf{r}_1,\\mathbf{r}_2,\\dots ,\\mathbf{r}_Z\\}$ .", + "bbox": [ + 89, + 609, + 483, + 731 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Evaluation", + "text_level": 1, + "bbox": [ + 89, + 744, + 207, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Evaluation Setup", + "text_level": 1, + "bbox": [ + 89, + 768, + 259, + 786 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Backbone API. In our experiments, we selected four backbone models: Qwen2-VL-72B [35], InternVL2-Llama3-76B, TeleMM [23], and GPT4o.", + "bbox": [ + 89, + 790, + 483, + 837 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset and Evaluation. We conduct experiments on two tasks: traditional image classification and Visual Question Answer (VQA). For image classification, we use the", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 2 Semantic-similarity-Guided Evolutionary Computation for Discrete Prompt Optimization" + ], + "code_body": "Input: I clients with discrete prompt tokens $\\{\\mathcal{P}_{\\mathrm{client}_i}\\}_{i = 1}^I$ \nOutput: Global prompt tokens $\\mathcal{P}_{\\mathrm{new}}$ \n1: Initialization Initial population $\\mathcal{P}_{\\mathrm{init}}$ \n2: for each token $tk$ in a prompt $\\in \\mathcal{P}_{\\mathrm{init}}$ do \n3: Achieve mapping according to Eq.4 \n4: end for \n5: for each token $\\mathbf{t}_g$ do $\\triangleright$ Reputation Test \n6: for each token $\\mathbf{t}_h$ do \n7: Semantic-similarity weight according to Eq.5 \n8: end for \n9: Get total semantic-similarity weight $\\alpha_{g} = \\sum_{h}\\alpha_{g,h}$ \n10: end for \n11: Cluster tokens according to Eq.6 Aggregation \n12: for each Community $C_z\\in \\mathcal{C}$ do Election \n13: Select representative token according to Eq.7 \n14: if $C_z$ contains only one token then \n15: Discard Community $C_z$ \n16: end if \n17: end for \n18: Form Representatives $\\mathcal{P}_{\\mathrm{new}} = \\{\\mathbf{r}_1,\\mathbf{r}_2,\\dots ,\\mathbf{r}_Z\\}$ from selected representative tokens", + "bbox": [ + 511, + 125, + 906, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MNIST [9], CIFAR-10, CIFAR-100, and TinyImageNet (TinyImg) [20] datasets. For VQA, we utilize the Art, Chemistry, and Finance sections of the MMMU dataset [43], along with the RealWorldQA (RWQA) dataset. The task objective is to leverage optimized prompts for accurate image classification on the image datasets, and for VQA, to assist in parsing textual content within the datasets to enable accurate classification or decision-making.", + "bbox": [ + 511, + 473, + 906, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use Accuracy as the evaluation metric for the image classification task. To maintain consistency across tasks, we also constrain predictions in the VQA task to the range of available labels and evaluate performance using Accuracy.", + "bbox": [ + 511, + 595, + 906, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparison Baselines. We compare our method against eight state-of-the-art (SOTA) approaches: FedPrompt [46], FedPepTAO [4], FedTPG [29], HPT [37], CoOp [48], BBT [10], FedBPT [31], BPTVLM [41], and a manual prompting method (Manual Prompt). Notably, HPT [37], CoOp [48], BBT[10], and BPTVLM [41] are not originally designed for federated learning. To adapt these methods to a federated setting, we apply a weighted averaging aggregation approach, following the implementations of FedPrompt [46]and FedBPT [31]. The results are reported using federated versions of these models: FedAvg-HPT, FedAvg-CoOp, FedAvg-BBT, and Fed-BPTVLM. Additionally, we categorize these methods into White-box and Black-box approaches based on whether they require access to model parameters. The White-box methods include FedPrompt [46],", + "bbox": [ + 511, + 674, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2465", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/9aaa57d4b4beb88de6977a74ffcc74e8af97332a32ca82a86645ed5c709770d7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BackboneMethodCV ClassificationVQAAverage
MNISTCifar-10Cifar-100TinyImgArtChemistryFinanceRWQA
Qwen2-VL-72BWhite BoxFedPrompt90.479.573.560.974.663.370.580.4
FedPepTAO86.778.372.959.569.158.868.577.9
FedTPG84.276.871.359.272.457.562.675.4
FedAvg-HPT86.978.470.258.775.753.565.978.3
FedAvg-CoOp85.778.171.459.270.261.863.376.5
Black BoxManual70.762.253.547.354.745.851.858.3
FedAvg-BBT78.972.362.751.763.553.960.369.7
FedBPT80.072.064.253.168.050.659.067.8
Fed-BPTVLM80.773.162.350.965.253.363.370.8
Our method90.578.571.857.874.060.867.278.5
InternVL2-Llama3-76BWhite BoxFedPrompt88.181.374.460.270.258.862.976.1
FedPepTAO87.377.871.258.565.255.363.075.4
FedTPG86.279.371.757.468.054.260.773.8
FedAvg-HPT85.178.772.358.469.153.062.074.3
FedAvg-CoOp85.779.171.658.171.856.263.971.7
Black BoxManual71.863.751.845.452.546.353.861.3
FedAvg-BBT80.071.165.252.760.251.953.870.5
FedBPT80.772.766.652.361.350.658.467.1
Fed-BPTVLM79.571.467.253.064.153.056.769.8
Our method89.679.071.557.370.755.963.675.5
TeleMMBlack BoxManual69.465.553.345.853.045.052.557.9
FedAvg-BBT80.970.663.149.156.948.856.464.1
FedBPT80.071.361.950.960.851.557.466.3
Fed-BPTVLM79.171.062.749.759.152.357.768.5
Our method88.478.969.357.464.655.761.675.0
GPT4oBlack BoxManual74.468.259.152.861.958.661.671.9
FedAvg-BBT81.076.767.656.164.657.563.669.2
FedBPT83.278.368.456.669.159.368.569.9
Fed-BPTVLM81.779.067.957.367.457.965.271.6
Our method92.585.373.962.776.866.772.878.5
", + "bbox": [ + 106, + 87, + 890, + 587 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Performance Comparison on CV Classification and VQA Tasks. The accuracy results are measured in percentages (\\%). Bold values indicate the overall best results, and underlined values indicate the best results among Black Box methods.", + "bbox": [ + 89, + 595, + 906, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "FedPepTAO [4], FedTPG [29], FedAvg-HPT, and FedAvg-CoOp, while the Black-box methods consist of Manual, FedAvg-BBT, FedBPT [31], and Fed-BPTVLM. Detailed descriptions of all baselines are provided in Appendix A.", + "bbox": [ + 89, + 651, + 482, + 712 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation. The federated learning setup for our experiments is as follows: we use a default of 10 clients, with a client participation rate of $100\\%$ in each training round. Additionally, we adopt a $k$ -shot mechanism for prompt tuning, following the approach in work [31], with the default value of $k$ set to 50. For image classification datasets, we use the original training and test sets. For each Visual QA dataset, we randomly select 50 samples as the training set (from the validation set for MMMU and the test set for RealWorldQA), with the remaining samples used for testing. In terms of baseline implementations, we adapt traditional", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "prompt optimization methods (HPT, CoOp, BBT, BBT-VLM) to the federated learning framework by employing the classical Federated Averaging (FedAvg) approach. For black-box backbone model configurations, we extend soft prompt tuning methodologies, such as FedBPT, by mapping the soft tokens to discrete tokens, through the RoBERTa encoder's tokenization mechanism, before feeding them into API-based models. In addition, all experiments are conducted on two NVIDIA GeForce RTX A800 GPUs. Detailed hyperparameter settings are provided in Appendix B.", + "bbox": [ + 511, + 651, + 906, + 803 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Effectiveness Results", + "text_level": 1, + "bbox": [ + 511, + 816, + 709, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first evaluate the effectiveness of our method on traditional CV classification tasks and VQA tasks. Experimental results show that, compared to other methods, our method demonstrates significant performance improvements in both", + "bbox": [ + 511, + 839, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2466", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tasks, particularly excelling in black-box environments, with only a small gap from white-box algorithms.", + "bbox": [ + 89, + 90, + 480, + 122 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Traditional CV Classification Tasks. In traditional CV classification tasks, our method achieves outstanding performance across all model configurations. For example, in the GPT4o black-box setting, our method achieves improvements of $11.2\\%$ , $7.0\\%$ , $5.5\\%$ , and $6.1\\%$ over the closest-performing FedBPT [31] method on the MNIST, Cifar-10, Cifar-100, and TinyImg datasets, respectively. In the TeleMM black-box setting, our method also performs exceptionally well, achieving $88.4\\%$ and $78.9\\%$ accuracy on MNIST and Cifar-10, respectively, which is an improvement of $9.3\\%$ and $8.3\\%$ over FedAvg-BBT. Additionally, for the Qwen2-VL-72B model, our method achieves accuracies of $78.5\\%$ and $71.8\\%$ on the Cifar-10 and Cifar-100 datasets, respectively, representing improvements of $7.1\\%$ and $5.3\\%$ over FedBPT [31], further demonstrating its advantage in traditional CV tasks.", + "bbox": [ + 89, + 141, + 482, + 383 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "VQA Tasks. Our method also shows strong performance in VQA tasks. In the GPT4o black-box setting, our method achieves improvements of $11.1\\%$ , $7.4\\%$ , $4.3\\%$ , and $8.6\\%$ over the FedBPT [31] method on the Art, Chemistry, Finance, and RWQA datasets, respectively. In TeleMM setting, our method's performance remains very stable, achieving improvements of $7.7\\%$ and $10.9\\%$ over FedAvg-BBT on the Art and RWQA datasets, respectively.", + "bbox": [ + 89, + 402, + 482, + 525 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Prompt Transferability Evaluation", + "text_level": 1, + "bbox": [ + 89, + 534, + 393, + 550 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the prompt transferability experiment, we evaluate the performance of prompts optimized on one backbone model when applied to another, as shown in Table 2. The results indicate that prompts exhibit a degree of transferability, maintaining high robustness across models with minimal accuracy loss, as referenced in Table 1. Additionally, compared to the Qwen2-VL transfer results, the InternVL2 transfer showed better accuracy, further validating the high applicability and robustness of prompts between certain models.", + "bbox": [ + 89, + 556, + 482, + 708 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/329c6aa54a41fb62a1627c40a1023d44ac408383efbb017c4bba9319fb55c93f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelCifar-10Cifar-100TinyImg
Qwen2-VL→TeleMM8070.558.1
InternVL2→GPT4o87.872.363.5
", + "bbox": [ + 94, + 717, + 478, + 765 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Communication Overhead Experiment", + "text_level": 1, + "bbox": [ + 89, + 832, + 426, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the communication overhead experiment, our method shows a significant advantage in parameter efficiency, as illustrated in Table 3. Specifically, white-box methods like", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "FedPrompt [46] and FedTPG [29] and most black-box methods such as FedAvg-BBT and Fed-BPTVLM require hundreds of thousands of parameters, whereas our method requires only 150. This substantial reduction effectively minimizes latency and resource consumption associated with communication. This advantage primarily arises from a key difference in algorithm design: while other methods use soft prompts for knowledge exchange between clients, our method employs discrete prompt tokens. Soft prompts typically require a large number of continuous parameters, whereas discrete prompt tokens enable knowledge transfer with only a small set of parameters, significantly reducing communication overhead.", + "bbox": [ + 511, + 90, + 903, + 287 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/aa4c6eea7331c9d07ab9724ccd8d4dca825b99624935b534b17ae42878f44315.jpg", + "table_caption": [ + "Table 2. Transferability Experiment Results. The accuracy results are measured in percentages $(\\%)$ ." + ], + "table_footnote": [], + "table_body": "
TypeMethodParameters
White BoxFedPrompt614k
FedPepTAO1796k
FedTPG614k
FedAvg-HPT614k
FedAvg-CoOp614k
Black BoxFedAvg-BBT614k
FedBPT40.8k
Fed-BPTVLM614k
Our method150
", + "bbox": [ + 517, + 297, + 901, + 448 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Communication Overhead Experiment results. The measurements are in Tensors within the PyTorch Framework. Bold values indicate the best results, and underlined values indicate the second-best results.", + "bbox": [ + 511, + 457, + 906, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 545, + 678, + 560 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the improvements ablation study based on Qwen2-VL72B API, our method consistently achieves the best performance across all metrics, further validating the effectiveness of its key components, as shown in Table 4. Compared to cases where specific modules are removed or replaced, our method demonstrates significant improvements on the Cifar-10, Cifar-100, and TinyImg datasets. Server-side experiments. Our method improves accuracy on the Cifar-10 and Cifar-100 datasets by $2.0\\%$ and $1.2\\%$ , respectively, compared to the setting without the clustering strategy. Furthermore, compared to the fixed clustering approach, our method achieves accuracy gains of $0.7\\%$ on Cifar-10 and $2.5\\%$ on Cifar-100, underscoring the importance of a dynamic clustering strategy. Client-side experiments. Removing the G-F mechanism results in a significant accuracy drop across all datasets. Furthermore, our method achieves a $1.0\\%$ accuracy improvement on Cifar-100 and $1.6\\%$ on TinyImg compared to the configuration without the E-EB strategy. Compared to the configuration without the PPR-CoT strategy, our method improves accuracy by $3.2\\%$ on Cifar-10 and $2.5\\%$ on Cifar-100, underscoring the pivotal role of the PPR-CoT strategy in enhancing model perfor", + "bbox": [ + 511, + 568, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2467", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/014723ac8c30ea6bf4eb63108b72b1e7123449f98af58dfb737d99169f31fe74.jpg", + "table_caption": [ + "mance." + ], + "table_footnote": [], + "table_body": "
SiteMethodCifar-10Cifar-100TinyImg
Serverw Embedding-fedavg71.962.554.2
w Random search65.357.353.9
w/o Clustering Strategy76.570.655.9
w Fixed Clusters77.869.356.3
Clientw/o G-F Mechanism71.865.553.5
w/o E-EB Strategy76.970.856.2
w/o PPR-CoT75.369.356.7
Our method78.571.857.8
", + "bbox": [ + 94, + 118, + 480, + 262 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. The Improvements Ablation Studies Results. The accuracy is measured in percentages (\\%). Performance decreases with the removal or modification of any key component, underscoring the effectiveness of the improvements. \"w Embedding-fedavg\" denotes the use of a weighted averaging aggregation method, as in FedBPT [31] and FedAvg-BBT. \"w Random search\" refers to the removal of weight comparison, using random search instead to select representative tokens. \"w/o Clustering Strategy\" indicates the removal of the clustering step. \"w Fixed Clusters\" implies a fixed number of clusters, which affects the count of representative tokens.", + "bbox": [ + 89, + 271, + 483, + 424 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Initial Prompt Impact. We evaluate the impact of the Initial Prompt on performance using GPT-4o as the backbone black-box VLM for classification tasks across datasets, comparing it with three typical black-box algorithms. Figure 2 shows that different prompt configurations significantly affect model performance, with our algorithm consistently achieving the best outcomes. Specifically, detailed prompts enhance model accuracy by providing richer information, which supports more effective learning and adaptation. In contrast, shorter prompts reduce accuracy, indicating that limited information constrains performance. Thus, prompt design is critical in classification tasks, and our algorithm exhibits strong robustness and performance across various prompt configurations. We present all Initial Prompts in Appendix C.", + "bbox": [ + 88, + 473, + 486, + 700 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bdf692c71c5accc0fa2d0073f9c851bbd72af7066710b1a942b3d1ad6d0a7bcb.jpg", + "image_caption": [ + "Figure 2. Initial Prompt Impact. Prompt-1 represents a \"shorter initial prompt,\" Prompt-2 a \"standard initial prompt,\" and Prompt-3 a \"detailed initial prompt.\"" + ], + "image_footnote": [], + "bbox": [ + 122, + 729, + 444, + 840 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Heterogeneous Data Robustness Experiment", + "text_level": 1, + "bbox": [ + 511, + 90, + 890, + 107 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the heterogeneous data robustness experiment, we use the Qwen2-VL-72B API as the backbone VLM model to evaluate the performance of different algorithms on heterogeneous data in the TinyImagenet dataset, as shown in Figure 3. Data is distributed across 10 clients using a Dirichlet Distribution [21] with varying hyperparameters dir, where a smaller Dirichlet parameter (e.g., dir=0.05) indicates greater data heterogeneity. Details of these heterogeneous data distributions are provided in Appendix D. The results demonstrate that all methods show some robustness to data heterogeneity, aligning with expectations [28, 31] for model stability on heterogeneous data. Additionally, as heterogeneity increases, our method exhibits superior performance, with minimal accuracy reduction, outperforming other methods under all heterogeneity conditions.", + "bbox": [ + 511, + 114, + 906, + 340 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/da5f1c208403cb7a251f64a0194070670d99b45fd8172029e77c6a9e485be995.jpg", + "image_caption": [ + "Figure 3. Heterogeneous Data Robustness Experiment results on the TinyImageNet dataset. The parameter \"dir\" is the Dirichlet Distribution hyperparameter; smaller values of dir indicate higher data heterogeneity across the 10 clients." + ], + "image_footnote": [], + "bbox": [ + 552, + 354, + 864, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 556, + 633, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose FDPT, the first federated black-box discrete prompt tuning framework based on visual language models. FDPT uses an Agent-based Client Prompt Optimization strategy to achieve token-level prompt optimization on the client side without traditional prediction logits, thereby preventing inversion attacks. On the server side, a Semantic-Similarity-Guided Evolutionary Computation framework is implemented to execute more precise zeroth-order optimization, which enables the obtention of representative global tokens. Moreover, discrete token exchange between clients offers high security and low communication overhead. Moreover, discrete tokens exhibit enhanced interpretability compared to soft prompts, which facilitates human experts in designing corresponding instructions. Our framework exhibits: (1) statistically significant accuracy enhancements in conventional visual recognition and cross-modal reasoning tasks; (2) optimized communication efficiency through adaptive token selection mechanisms; (3) cross-domain transferability of learned prompt representations; and (4) strengthened resilience to data distribution shifts, particularly in heterogeneous learning environments.", + "bbox": [ + 511, + 583, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2468", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This research was supported by the National Natural Science Foundation of China under Grant 52074305, 52274160, 51874300 and the National Natural Science Foundation of China-Shanxi Provincial People's Government Coal-based Low Carbon Joint Fund U1510115, the Academic Research Projects of Beijing Union University (No.ZK90202106) and R&D Program of Beijing Municipal Education Commission (KM202211417005).", + "bbox": [ + 89, + 90, + 485, + 227 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 239, + 187, + 257 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 2", + "[2] Owen Astrachan. Bubble sort: an archaeological algorithmic analysis. ACM Sigcse Bulletin, 35(1):1-5, 2003. 5", + "[3] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 2", + "[4] Tianshi Che, Ji Liu, Yang Zhou, Jiaxiang Ren, Jiwen Zhou, Victor S Sheng, Huaiyu Dai, and Dejing Dou. Federated learning of large language models with parameter-efficient prompt tuning and adaptive optimization. arXiv preprint arXiv:2310.15080, 2023. 5, 6", + "[5] Simin Chen, Zihe Song, Mirazul Haque, Cong Liu, and Wei Yang. Nicsslowdown: Evaluating the efficiency robustness of neural image caption generation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15365-15374, 2022. 2", + "[6] Simin Chen, Hanlin Chen, Mirazul Haque, Cong Liu, and Wei Yang. The dark side of dynamic routing neural networks: Towards efficiency backdoor injection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24585-24594, 2023. 2", + "[7] Yiyi Chen, Heather Lent, and Johannes Bjerva. Text embedding inversion security for multilingual language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7808-7827, 2024. 1", + "[8] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024. 2", + "[9] Li Deng. The mnist database of handwritten digit images for machine learning research [best of the web]. IEEE signal processing magazine, 29(6):141-142, 2012. 5", + "[10] Shizhe Diao, Zhichao Huang, Ruijia Xu, Xuechun Li, Yong Lin, Xiao Zhou, and Tong Zhang. Black-box prompt learning for pre-trained language models. Trans. Mach. Learn. Res., 2023, 2023. 5" + ], + "bbox": [ + 94, + 265, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Xiuwen Fang and Mang Ye. Robust federated learning with noisy and heterogeneous clients. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10072-10081, 2022. 2", + "[12] Cong Gao, Wenfeng Li, Lijun He, and Lingchong Zhong. A distance and cosine similarity-based fitness evaluation mechanism for large-scale many-objective optimization. Engineering Applications of Artificial Intelligence, 133:108127, 2024. 5", + "[13] Aditya Golatkar, Alessandro Achille, Yu-Xiang Wang, Aaron Roth, Michael Kearns, and Stefano Soatto. Mixed differential privacy in computer vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8376-8386, 2022. 1", + "[14] Xiang Gu, Shuchao Pang, Anan Du, Yifei Wang, Jixiang Miao, and Jorge Díez. Dynamic multimodal prompt tuning: Boost few-shot learning with vlm-guided point cloud models. In ECAI, pages 761-768. IOS Press, 2024. 3", + "[15] Tao Guo, Song Guo, Junxiao Wang, Xueyang Tang, and Wenchao Xu. Promptfl: Let federated participants cooperatively learn prompts instead of models-federated learning in age of foundation model. IEEE Transactions on Mobile Computing, 2023. 1", + "[16] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli-hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 2", + "[17] Xinyuan Ji, Zhaowei Zhu, Wei Xi, Olga Gadyatskaya, Zi long Song, Yong Cai, and Yang Liu. Fedfixer: Mitigating heterogeneous label noise in federated learning. In AAAI, pages 12830-12838. AAAI Press, 2024. 2", + "[18] Ruochen Jiao, Shaoyuan Xie, Justin Yue, Takami Sato, Lixu Wang, Yixuan Wang, Qi Alfred Chen, and Qi Zhu. Exploring backdoor attacks against large language model-based decision making. arXiv preprint arXiv:2405.20774, 2024. 1", + "[19] Weirui Kuang, Bingchen Qian, Zitao Li, Daoyuan Chen, Dawei Gao, Xuchen Pan, Yuexiang Xie, Yaliang Li, Bolin Ding, and Jingren Zhou. Federatedscope-llm: A comprehensive package for fine-tuning large language models in federated learning. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 5260–5271, 2024. 1", + "[20] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3, 2015. 2, 5", + "[21] Qinbin Li, Yiqun Diao, Quan Chen, and Bingsheng He. Federated learning on non-iid data silos: An experimental study. In 2022 IEEE 38th international conference on data engineering (ICDE), pages 965-978. IEEE, 2022. 8", + "[22] Tian Li, Anit Kumar Sahu, Ameet Talwalkar, and Virginia Smith. Federated learning: Challenges, methods, and future directions. IEEE signal processing magazine, 37(3):50-60, 2020. 1, 2, 3, 4", + "[23] Xiang Li, Yiqun Yao, Xin Jiang, Xuezhi Fang, Chao Wang, Xinzhang Liu, Zihan Wang, Yu Zhao, Xin Wang, Yuyao Huang, Shuangyong Song, Yongxiang Li, Zheng Zhang, Bo Zhao, Aixin Sun, Yequan Wang, Zhongjiang He, Zhongyuan Wang, Xuelong Li, and Tiejun Huang. Tele-flm technical report, 2024. 5" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2469", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Zexin Li, Bangjie Yin, Taiping Yao, Junfeng Guo, Shouhong Ding, Simin Chen, and Cong Liu. Sibling-attack: Rethinking transferable adversarial attacks against face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24626–24637, 2023. 2", + "[25] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 2", + "[26] Mulei Ma, Chenyu Gong, Liekang Zeng, Yang Yang, and Liantao Wu. Flocoff: Data heterogeneity resilient federated learning with communication-efficient edge offloading. IEEE Journal on Selected Areas in Communications, 42(11): 3262-3277, 2024. 2", + "[27] Yifei Ming and Yixuan Li. How does fine-tuning impact out-of-distribution detection for vision-language models? International Journal of Computer Vision, 132(2):596-609, 2024. 1", + "[28] John Nguyen, Jianyu Wang, Kshitiz Malik, Maziar Sanjabi, and Michael Rabbat. Where to begin? on the impact of pre-training and initialization in federated learning. arXiv preprint arXiv:2206.15387, 2022. 8", + "[29] Chen Qiu, Xingyu Li, Chaithanya Kumar Mummadi, Madan Ravi Ganesh, Zhenzhen Li, Lu Peng, and Wan-Yi Lin. Federated text-driven prompt generation for vision-language models. In The Twelfth International Conference on Learning Representations, 2024. 1, 5, 6, 7", + "[30] Mathias Risse. Arguing for majority rule*. Journal of Political Philosophy, 12(1), 2004. 5", + "[31] Jingwei Sun, Ziyue Xu, Hongxu Yin, Dong Yang, Daguang Xu, Yiran Chen, and Holger R Roth. Fedbpt: Efficient federated black-box prompt tuning for large language models. arXiv preprint arXiv:2310.01467, 2023. 1, 2, 5, 6, 7, 8", + "[32] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 2", + "[33] Lixu Wang, Shichao Xu, Xiao Wang, and Qi Zhu. Addressing class imbalance in federated learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10165-10173, 2021. 2", + "[34] Lixu Wang, Shichao Xu, Ruiqi Xu, Xiao Wang, and Qi Zhu. Non-transferable learning: A new approach for model ownership verification and applicability authorization. In International Conference on Learning Representation (ICLR-2022), 2022. 1", + "[35] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 2, 5", + "[36] Tao Wang, Yushu Zhang, Shuren Qi, Ruoyu Zhao, Xia Zhihua, and Jian Weng. Security and privacy on generative data in aigc: A survey. ACM Computing Surveys, 2023. 1", + "[37] Yubin Wang, Xinyang Jiang, De Cheng, Dongsheng Li, and Cairong Zhao. Learning hierarchical prompt with structured linguistic knowledge for vision-language models. In AAAI, pages 5749-5757. AAAI Press, 2024. 5" + ], + "bbox": [ + 91, + 90, + 480, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Yubo Wang, Chaohu Liu, Yanqiu Qu, Haoyu Cao, Deqiang Jiang, and Linli Xu. Break the visual perception: Adversarial attacks targeting encoded visual tokens of large vision-language models. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 1072-1081, 2024. 1", + "[39] Feijie Wu, Zitao Li, Yaliang Li, Bolin Ding, and Jing Gao. Fedbiot: Llm local fine-tuning in federated learning without full model. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 3345-3355, 2024. 1", + "[40] Jiaqi Wu, Simin Chen, Yuzhe Yang, Yijiang Li, Shiyue Hou, Rui Jing, Zehua Wang, Wei Chen, and Zijian Tian. Feddtpt: federated discrete and transferable prompt tuning for black-box large language models. arXiv preprint arXiv:2411.00985, 2024. 2", + "[41] Lang Yu, Qin Chen, Jiaju Lin, and Liang He. Black-box prompt tuning for vision-language model as a service. In IJCAI, pages 1686-1694, 2023. 1, 5", + "[42] Bowen Yuan, Sisi You, and Bing-Kun Bao. Self-pt: Adaptive self-prompt tuning for low-resource visual question answering. In ACM Multimedia, pages 5089-5098. ACM, 2023. 3", + "[43] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9556-9567, 2024. 2, 5", + "[44] Jingjing Zhang, Shancheng Fang, Zhendong Mao, Zhiwei Zhang, and Yongdong Zhang. Fine-tuning with multi-modal entity prompts for news image captioning. In ACM Multimedia, pages 4365-4373. ACM, 2022. 3", + "[45] Jingyi Zhang, Jiaxing Huang, Sheng Jin, and Shijian Lu. Vision-language models for vision tasks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1, 2", + "[46] Haodong Zhao, Wei Du, Fangqi Li, Peixuan Li, and Gongshen Liu. Fedprompt: Communication-efficient and privacy-preserving prompt tuning in federated learning. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 1, 2, 5, 7", + "[47] Xuejun Zhao, Wencan Zhang, Xiaokui Xiao, and Brian Lim. Exploiting explanations for model inversion attacks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 682-692, 2021. 1", + "[48] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 5", + "[49] Yu Zhu, Kang Li, Lequan Yu, and Pheng-Ann Heng. Memory-efficient prompt tuning for incremental histopathology classification. In AAAI, pages 7802-7810. AAAI Press, 2024. 2" + ], + "bbox": [ + 516, + 90, + 903, + 878 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2470", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/FDPT_ Federated Discrete Prompt Tuning for Black-Box Visual-Language Models/5d5a1b2f-9957-4fb4-b461-7bf3e1ebccfb_model.json b/2025/FDPT_ Federated Discrete Prompt Tuning for Black-Box Visual-Language Models/5d5a1b2f-9957-4fb4-b461-7bf3e1ebccfb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c8c94d1ccd38a47928f6c91dbf2858ed9ef569cd --- /dev/null +++ b/2025/FDPT_ Federated Discrete Prompt Tuning for Black-Box Visual-Language Models/5d5a1b2f-9957-4fb4-b461-7bf3e1ebccfb_model.json @@ -0,0 +1,1947 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.126, + 0.13, + 0.873, + 0.175 + ], + "angle": 0, + "content": "FDPT: Federated Discrete Prompt Tuning for Black-Box Visual-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.203, + 0.886, + 0.24 + ], + "angle": 0, + "content": "Jiaqi Wu\\(^{1,8,10\\dagger}\\), Simin Chen\\(^{2\\dagger}\\), Jing Tang\\(^{3}\\), Yuzhe YANG\\(^{4}\\), Yiming Chen\\(^{5}\\), Lixu Wang\\(^{6}\\), Song Lin\\(^{7}\\), Zehua Wang\\(^{8,9}\\), Wei Chen\\(^{9*}\\), and Zijian Tian\\(^{10}\\)" + }, + { + "type": "text", + "bbox": [ + 0.164, + 0.242, + 0.809, + 0.258 + ], + "angle": 0, + "content": "\\(^{1}\\)Tsinghua University, \\(^{2}\\)University of Texas at Dallas, \\(^{3}\\)Huazhong University of Science and Technology," + }, + { + "type": "text", + "bbox": [ + 0.164, + 0.258, + 0.833, + 0.275 + ], + "angle": 0, + "content": "4The Chinese University of Hong Kong, 5National University of Singapore, 6Nanyang Technological University," + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.276, + 0.76, + 0.292 + ], + "angle": 0, + "content": "7Beijing Baidu Netcom Science Technology Co., Ltd., 8University of British Columbia," + }, + { + "type": "text", + "bbox": [ + 0.164, + 0.294, + 0.803, + 0.31 + ], + "angle": 0, + "content": "\\(^{9}\\)China University of Mining and Technology, \\(^{10}\\)China University of Mining and Technology (Beijing)" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.312, + 0.843, + 0.327 + ], + "angle": 0, + "content": "wjq11346@student.ubc.ca, sxc180080@utdallas.edu, focusers@163.com, yuzheyang@link.cuhk.edu.cn," + }, + { + "type": "text", + "bbox": [ + 0.184, + 0.33, + 0.812, + 0.345 + ], + "angle": 0, + "content": "yiming.chen@u.nus.edu, wanglixu4334@gmail.com, linsong04@baidu.com, zwang@ece.ubc.ca," + }, + { + "type": "text", + "bbox": [ + 0.351, + 0.348, + 0.642, + 0.362 + ], + "angle": 0, + "content": "chenwdavior@163.com，Tianzj0726@126.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.398, + 0.327, + 0.413 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.429, + 0.486, + 0.868 + ], + "angle": 0, + "content": "Vision-Language Models (VLMs) have driven significant advancements in multimodal AI. Fine-tuning these models with user data enhances adaptability but poses privacy risks. While federated learning (FL) mitigates user data privacy concerns, it fails to protect model properties. Existing methods relying on black-box VLM APIs often require access to prediction logits, making them vulnerable to inversion attacks. Additionally, optimizing tuning complexity and data transmission efficiency in federated VLM scenarios remains a challenge. To address these challenges, we propose FDPT—the first federated discrete prompt tuning method utilizing black-box VLMs. During client optimization stage, FDPT employs an agent-driven framework leveraging large language models (LLMs) with enhanced reasoning capacities to systematically optimize discrete prompt representations, and also utilizes feedback mechanisms and chain of thought to enhance prediction accuracy. Importantly, it performs optimization by relying not on the predicted logic vectors output by LLMs but on textual results, avoiding reverse attack risks. During global aggregation stage, We mimic human electoral activities by employing evolutionary computation methods underpinned by semantic similarity computation to implement enhanced zero-order optimization for acquiring representative global tokens, thereby achieving knowledge aggregation. FDPT significantly outperforms nine state-of-the-art methods in image classification and visual question-answering, reducing communication overhead while generating highly transferable optimized prompts. Additionally, it exhibits im" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.4, + 0.78, + 0.415 + ], + "angle": 0, + "content": "proved robustness to data heterogeneity." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.448, + 0.646, + 0.464 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.475, + 0.907, + 0.642 + ], + "angle": 0, + "content": "General-purpose Vision-Language Models (VLMs) [45] integrate vast visual and linguistic information, advancing various vision tasks [36]. To better adapt to user needs, these models can undergo fine-tuning or prompt-tuning [27], but such modifications introduce privacy risks [13, 18, 34, 38]. Federated Learning (FL) [22] addresses these concerns by decentralizing data storage, reducing single-point breaches, and exchanging model weights or embeddings instead of raw data. Various FL-based approaches [19, 29, 39] have been explored for tuning large models, effectively preserving user data privacy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.909, + 0.884 + ], + "angle": 0, + "content": "Despite the success of FL, protecting the privacy of VLMs themselves remains a critical yet largely unexplored challenge. Existing FL approaches use zero-order optimization to estimate gradients for prompt tuning on black-box VLMs [31, 41]. For example, FedBPT [31] employs prompt exchange and gradient-free optimization, reducing communication and memory costs by avoiding direct parameter access and backpropagation on client devices. However, several limitations remain. First, these methods remain vulnerable to security risks as they rely on prediction logits, exposing them to inversion attacks that can infer model parameters [7, 47]. Second, zero-order optimization techniques [31, 41] often suffer from limited accuracy and adaptability in prompt fine-tuning. Finally, soft prompt methods [15, 46] lack interpretability and incur high communication costs in federated settings." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "To address these challenges, we propose FDPT, the first" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.218, + 0.888 + ], + "angle": 0, + "content": "\\(\\dagger\\) Equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.112, + 0.888, + 0.235, + 0.9 + ], + "angle": 0, + "content": "* Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2461" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.424 + ], + "angle": 0, + "content": "federated discrete prompt optimization method comprising two stages: client-side optimization and global aggregation. In the client-side optimization phase, we introduce AI Agent-based Client Prompt Optimization, which refines token-level prompts using large language models (LLMs). We design a feedback mechanism and Chain-of-Thought measures to guide the agent in selecting optimal discrete tokens, overcoming the limitations of existing zero-order optimization methods in prompt tuning. Unlike traditional approaches that rely on prediction logits, our method uses only black-box VLM text outputs, reducing the risk of inversion attacks and enhancing model privacy protection. Afterwards, in the global aggregation phase, to reduce communication overhead in federated settings, we use discrete tokens for knowledge exchange instead of soft prompts. Inspired by human voting mechanisms, we propose a Self-Attention-Guided Evolutionary Computation method to identify representative tokens for constructing global prompts. Unlike the coarse sampling in traditional zero-order optimization, our approach clusters tokens in semantic space and then applies similarity calculations to refine the most accurate results, enhancing tuning efficiency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.45, + 0.483, + 0.661 + ], + "angle": 0, + "content": "Detailed experimental results demonstrate the advantages of our method. In Visual Question Answering (VQA) tasks, under the GPT-4 black-box setting, our method outperforms the classic black-box prompt optimization method FedBPT [31], achieving \\(11.1\\%\\) and \\(7.4\\%\\) improvements on the Art and Chemistry categories of the MMMU dataset [43], respectively. In traditional computer vision (CV) classification tasks, using the Qwen2-VL-72B model [3], our approach achieves \\(71.8\\%\\) accuracy on CIFAR-100, surpassing FedBPT by \\(5.3\\%\\), demonstrating its superior performance. Additionally, FDPT significantly reduces communication costs without compromising optimization performance and enhances interpretability through discrete prompts. The contributions of this paper are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.687, + 0.48, + 0.759 + ], + "angle": 0, + "content": "- We propose FDPT, a first federated black-box discrete prompt optimization framework that aggregates knowledge from multiple parties to optimize prompts while ensuring the privacy of both users and easily overlooked VLMs." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.763, + 0.48, + 0.834 + ], + "angle": 0, + "content": "- We introduce an AI Agent-based Client Prompt Optimization for token-level prompts and a Self-attention-guided Evolutionary Computation method for selecting global tokens, offering better predictive performance than traditional zero-order optimization." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.838, + 0.48, + 0.895 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate that our method achieves superior performance across various text-vision tasks, exhibits low communication overhead, and is resilient to data heterogeneity." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.687, + 0.48, + 0.895 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.783, + 0.107 + ], + "angle": 0, + "content": "2. Background & Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.115, + 0.905, + 0.433 + ], + "angle": 0, + "content": "Visual Language Models. Visual-Language Models (VLMs) [5, 45] are AI systems designed to process and understand both visual and textual information, enabling them to perform tasks such as Visual Question Answering (VQA) [43] and traditional visual classification tasks [20]. Traditional VLMs, like Qwen-VL [3], effectively integrate visual features with textual data, excelling in both VQA and image classification tasks. Flamingo [1] leverages few-shot learning capabilities, allowing it to quickly adapt to new tasks and perform well across both VQA and classification. Gemini [32] focuses on reasoning and multimodal understanding, achieving strong results in complex VQA tasks while also performing well in image and text classification. LLaVA [25] benefits from large-scale pretraining, efficiently handling intricate visuals and long-form text, with exceptional performance in both VQA and classification tasks. Newer models like InternVL2 [8], Qwen2-VL [35], and GPT-4o [16] build on these foundations, enhancing multimodal reasoning, improving the fusion of visual and textual data, and offering broader adaptability across both VQA and classification tasks." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.453, + 0.905, + 0.845 + ], + "angle": 0, + "content": "Federated Learning. Federated Learning (FL) [22] is a decentralized approach allowing multiple clients to collaboratively train a shared model without sharing raw data, thus preserving privacy and defending against attacks [6, 24]. Recent works have further improved FL's robustness in handling data heterogeneity [26, 33], label noise [17], and personalization needs [11]. With the rise of large language models (LLMs), Federated Learning (FL) has gained new applications, effectively addressing the challenges of knowledge exchange and privacy protection between independent LLMs. FedPrompt [46] combines FL with prompt tuning, utilizing split aggregation to significantly reduce communication costs. FedBPT [31] presents a federated black-box prompt tuning framework that avoids direct access to model parameters, instead optimizing prompts using gradient-free methods. FedDTPT [40] realizes discrete prompt tuning for Large Language Models (LLMs). However, these methods still face issues such as security risks in large models and high communication overhead. We propose a Federated Discrete Prompt Tuning method, which, compared to existing approaches, performs zero-order optimization at the token level, mitigating the risk of adversarial users leveraging logits generated by the large model for backdoor attacks. Additionally, by utilizing discrete tokens for knowledge exchange, it significantly reduces communication overhead." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Prompt Tuning. Prompt tuning [49] is an emerging technique for adapting large pre-trained models to specific tasks" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2462" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.172, + 0.089, + 0.825, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.332, + 0.908, + 0.361 + ], + "angle": 0, + "content": "Figure 1. Overview of FDPT. The proposed framework consists of the client-side local Agent-based Client Prompt Optimization and the server-side Semantic-Similarity-Guided Evolutionary Computation for global selection. The detailed process is described in Sec. 3.2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.387, + 0.483, + 0.464 + ], + "angle": 0, + "content": "with minimal additional training. In VLMs, prompt tuning [14] bridges visual and textual modalities, enabling tasks like image captioning [42, 44]. The Prompt tuning can be represented as finding an optimal prompt \\(\\mathbf{p}^*\\) to maximize the model's performance on the dataset \\(\\mathcal{D} = \\{(\\mathbf{x},\\mathbf{y})\\}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.196, + 0.485, + 0.483, + 0.502 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} ^ {*} = \\arg \\max f _ {\\mathrm {V L M}} (\\mathbf {p}; \\mathcal {D}) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.524, + 0.485, + 0.902 + ], + "angle": 0, + "content": "where \\( f_{\\mathrm{VLM}}(\\mathbf{p};\\mathcal{D}) \\) denotes the performance of the VLM using prompt \\( \\mathbf{p} \\) over the dataset \\( \\mathcal{D} \\). Continual Prompt Tuning (soft prompt tuning) involves fine-tuning a small set of trainable parameters (soft prompts) while keeping the core model fixed. However, it suffers from poor interpretability, and in Federated Learning (FL), it incurs high communication overhead between clients. Discrete Prompt Tuning (DPT), in contrast, optimizes specific discrete tokens from a fixed vocabulary. A notable approach, RLPrompt, uses reinforcement learning (RL) to select the best tokens based on rewards. However, it has limited optimization efficiency due to the computationally intensive nature of RL. Regarding above limitations, the proposed Federated Discrete Prompt Tuning method aims to combine the strengths of discrete prompt tuning with the constraints and needs of Federated Learning. By leveraging an AI agent (based on large language models) with robust reasoning capabilities, the method can generate more accurate and efficient local discrete tokens. Furthermore, by incorporating semantic similarity calculations into an evolutionary computing framework, the method can generate more precise global tokens that are shared across clients. Additionally, this reduces communication costs between clients, as discrete tokens are much smaller in parameters compared to soft prompts." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.385, + 0.605, + 0.401 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.41, + 0.718, + 0.424 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.432, + 0.907, + 0.552 + ], + "angle": 0, + "content": "In this work, we consider a distributed scenario [22], where \\(N\\) clients each possess a private dataset \\(\\mathcal{D}_k = \\{(\\mathcal{X}_k,\\mathcal{Y}_k)\\}\\) consisting of \\(n_k\\) samples \\(\\{(x_i^k,y_i^k)\\}_{i\\in [n_k]}\\). Each client can only access its own private data and does not wish to share it with others. This scenario is realistic, as hospitals and other organizations prioritize protecting sensitive data. Moreover, such distributed settings have been widely studied in existing research." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.907, + 0.613 + ], + "angle": 0, + "content": "Then, our federated discrete prompt optimization can be formulated as an optimization problem that aims to find the optimal prompt instruction \\( p^* \\) that satisfies the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.621, + 0.905, + 0.674 + ], + "angle": 0, + "content": "\\[\np ^ {*} = \\arg \\min _ {p} \\sum_ {k \\in [ K ]} \\frac {n ^ {k}}{\\sum_ {k \\in [ K ]} n ^ {k}} \\sum_ {i \\in [ n ^ {k} ]} \\mathcal {L} \\left(\\mathrm {V L M} \\left(p; x _ {i} ^ {k}\\right), y _ {i} ^ {k}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.779 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}(\\cdot)\\) is the objective function, which could be the error rate, and \\(\\mathrm{VLM}(\\cdot)\\) represents the Vision-Language Model (VLM). The above equation can be interpreted as learning a globally optimal natural language prompt by minimizing the weighted average loss across all clients, where each client's contribution is proportional to its dataset size." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Notice that if clients were permitted to share their private data and had white-box access to the VLM, the equation above could be easily solved by approximating the objective function with a differentiable function, such as cross-entropy loss. In this case, standard backpropagation could be applied to the global dataset to search for the optimal prompt instruction. However, due to commercial concerns, modern VLMs are typically deployed as API ser" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2463" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.152 + ], + "angle": 0, + "content": "vices, which are black-boxes to the clients. Additionally, each client is unwilling to share their data with others due to privacy concerns. Consequently, we must impose the following constraints:" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.16, + 0.358, + 0.193 + ], + "angle": 0, + "content": "Blackbox: \\(\\frac{\\partial\\mathcal{L}(\\mathrm{VLM}(p;x),y)}{\\partial p} = \\emptyset\\)" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.195, + 0.483, + 0.227 + ], + "angle": 0, + "content": "Privacy: \\(\\forall k,j\\in [K],k\\neq j,\\) \\(\\mathrm{Client}_k\\) cannot access \\(\\mathcal{D}_j\\) (3)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.236, + 0.484, + 0.282 + ], + "angle": 0, + "content": "The first constraint indicates that the gradient is not feasible, while the second constraint specifies that each client can only access its own data." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.29, + 0.258, + 0.306 + ], + "angle": 0, + "content": "3.2. Design Overview" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.312, + 0.485, + 0.614 + ], + "angle": 0, + "content": "The algorithm overview is illustrated in Figure 1. Consistent with the traditional federated learning paradigm [22], each client communicates solely with the server using the prompt information, while keeping their data local. This ensures compliance with the data privacy constraints outlined in Eq.(3). Specifically, at the beginning, the server distributes the Global Prompt to each client. Then our approach follows these steps during each global communication round to search an optimal prompt: (1) Client Optimization: Each client utilizes an Agent-based Client Prompt Optimization method for personalized token-level prompt optimization. This method is preferred over backpropagation and ensures adherence to the black-box constraints in Eq.(3), which is more realistic in the context of VLM as a service; (2) Server Aggregation: Each client uploads its locally optimized prompt to the server, where knowledge aggregation is performed using the Semantic-similarity-guided Evolutionary Computation method. The subsequent sections provide a detailed explanation of each step in our approach." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.623, + 0.282, + 0.639 + ], + "angle": 0, + "content": "3.3. Client Optimization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Upon receiving the global discrete prompt \\( p_0 \\) from the server in the current round, each client iteratively updates it using an agent-based approach, as shown in Alg. 1. Initially, the agent queries the LLM to randomly replace some tokens in \\( p_0 \\), generating a set of prompt mutants \\( p_1, p_2, \\ldots, p_M \\) (Line 1). Each prompt is then combined with the client's local dataset and fed into the VLM to measure the accuracy, which serves as the prompt mutant's score. FDPT maintains a priority queue to manage these mutants, prioritizing those with higher accuracy (Lines 2-6). FDPT iteratively selects the highest-priority prompt from the queue as the seed, asks the LLM to mutate the important tokens of the seed, and updates the priority queue with the newly generated prompt mutants (Lines 7-13). Specifically, token importance is computed as follows: (1), FDPT compares each prompt mutant \\( p_i \\) with the seed prompt \\( p_0 \\) token by token, constructing a difference matrix \\( M \\), where \\( M_i^j = 0 \\) if the" + }, + { + "type": "code_caption", + "bbox": [ + 0.515, + 0.091, + 0.836, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Agent-Based Prompt Optimization" + }, + { + "type": "algorithm", + "bbox": [ + 0.514, + 0.11, + 0.907, + 0.433 + ], + "angle": 0, + "content": "Input: Initial discrete prompt \\(p_0\\), Large Language Model (LLM), Vision-Language Model (VLM), client local dataset \\(\\mathcal{D}\\), number of mutants \\(M\\) \nOutput: Optimized prompt \\(\\hat{p}\\) \n1: Initialization: \\(\\{p_1, p_2, \\ldots, p_M\\} =\\) Mutator( \\(LLM, p_0\\) ) \\(\\triangleright\\) Mutate \\(p_0\\) by randomly replacing tokens \n2: \\(\\mathcal{Q} =\\) Priority Queue \\(\\triangleright\\) Initialize an empty queue \n3: for each prompt \\(p_i\\) in \\(\\{p_0, p_1, \\ldots, p_M\\}\\) do \n4: \\(Acc_i =\\) Evaluate(VLM, \\(p_i\\), \\(\\mathcal{D}\\)) \\(\\triangleright\\) Evaluation \n5: Push \\((\\mathcal{Q}, p_i, Acc_i)\\) \\(\\triangleright\\) Update the priority queue \n6: end for \n7: while stopping criterion not met do \n8: \\(p_{seed} =\\) Pop \\((\\mathcal{Q}) \\triangleright\\) Get the highest-priority prompt \n9: Position = Importance Estimator \\(\\triangleright\\) Select important tokens \n10: \\(p_{new} =\\) Mutator( \\(LLM, p_{seed}\\), Position) \n11: \\(Acc_{new} =\\) Evaluate(VLM, \\(p_{new}\\), \\(\\mathcal{D}\\)) \\(\\triangleright\\) Evaluation \n12: Push \\((\\mathcal{Q}, p_{new}, Acc_{new})\\) \\(\\triangleright\\) Update the queue \n13: end while \n14: Return highest-scoring prompt from \\(\\mathcal{Q}\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.459, + 0.907, + 0.581 + ], + "angle": 0, + "content": "\\(j\\)-th token of \\(p_i\\) and \\(p_0\\) are different, and \\(M_i^j = 1\\) otherwise. (2), FDPT computes the accuracy gains \\(V\\), where \\(V_i = \\mathrm{Acc}_i - \\mathrm{Acc}_0\\). (3) The importance score of each token position is then given by \\(S = M^{-1}V\\). This score estimation equation reflects the degree to which a token's presence or absence affects accuracy, weighted by its position in the prompt. Finally, the top \\(K\\) token positions are selected based on the importance score \\(S\\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.59, + 0.7, + 0.606 + ], + "angle": 0, + "content": "3.4. Server Aggregation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.612, + 0.906, + 0.702 + ], + "angle": 0, + "content": "After local client optimization, each client uploads its optimized discrete prompt to the server for knowledge exchange. The server then aggregates these local optimal prompts for the next round of optimization. Specifically, the server aggregation includes the following steps (as shown in Algorithm 2):" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.703, + 0.906, + 0.778 + ], + "angle": 0, + "content": "Token-level Semantic Representation. In total client uploaded prompts, e.g. the initial population \\(\\mathcal{P}_{\\mathrm{init}}\\), for each token in a prompt, we use a pre-trained text encoder of the Roberta to get its token-level semantic representation, denoted as" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.789, + 0.907, + 0.809 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} _ {i} = [ e _ {i} ^ {1}, e _ {i} ^ {2}, \\dots , e _ {i} ^ {K} ] e _ {i} ^ {j} = f _ {e m b e d} (t k _ {i} ^ {j}) \\qquad (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.82, + 0.906, + 0.852 + ], + "angle": 0, + "content": "where \\( f_{embed}(\\cdot) \\) is the pre-trained encoder, and \\( tk_i^j \\) is the \\( j^{th} \\) token of the \\( i^{th} \\) prompt." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Reputation Test. We first evaluate the reputation of each token, defined as the semantic-similarity weight (denoted" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2464" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.138 + ], + "angle": 0, + "content": "\\(\\alpha_{g,h})\\) that each token assigns to all other tokens. This is computed using a cosine-similarity mechanism [12] based on embeddings:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.148, + 0.483, + 0.185 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {g} = \\sum_ {h} \\alpha_ {g, h} = \\sum_ {h} \\operatorname {s o f t m a x} \\left(\\frac {\\mathbf {e} _ {g} \\cdot \\mathbf {e} _ {h}}{\\sqrt {d}}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.196, + 0.485, + 0.304 + ], + "angle": 0, + "content": "where \\(\\mathbf{e}_g\\) and \\(\\mathbf{e}_h\\) represent the embedding of a token \\(g\\) and \\(h\\), respectively, and \\(d\\) is the dimensionality of the embeddings. A higher semantic-similarity weight \\(\\alpha_{g,h}\\) signifies a stronger influence over other tokens. We define the total semantic-similarity weight of token \\(g\\) as \\(\\alpha_{g}\\), which represents the token's semantic importance among all other tokens, referred to as its Reputation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.321, + 0.485, + 0.368 + ], + "angle": 0, + "content": "Aggregation. Next, tokens are clustered based on their semantic similarity by grouping embeddings that are close to each other in the high-dimensional semantic space:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.378, + 0.484, + 0.408 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} = \\left\\{\\mathcal {C} _ {1}, \\mathcal {C} _ {2}, \\dots , \\mathcal {C} _ {Z} \\right\\}, \\mathcal {C} _ {z} = \\left\\{\\mathbf {e} _ {h} \\mid \\| \\mathbf {e} _ {h} - \\mathbf {e} _ {g} \\| < \\epsilon , \\forall h \\in \\mathcal {P} _ {\\text {i n i t}} \\right\\}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.409, + 0.485, + 0.47 + ], + "angle": 0, + "content": "where \\(\\epsilon\\) is a pre-defined distance threshold. \\(C_z\\) denotes a cluster, or Community. Embeddings with equivalent syntactic meanings are grouped into a cluster, from which a single representative is subsequently selected." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.488, + 0.484, + 0.564 + ], + "angle": 0, + "content": "Election. Finally, within each cluster (Community), the representative token is determined by semantic-similarity weights (reputation). Tokens with higher weights have stronger semantic dependencies on other tokens and are more likely to be selected:" + }, + { + "type": "equation", + "bbox": [ + 0.22, + 0.578, + 0.483, + 0.6 + ], + "angle": 0, + "content": "\\[\n\\mathbf {r} _ {z} = \\underset {\\mathbf {e} _ {n} \\in \\mathcal {C} _ {z}} {\\arg \\max } \\alpha_ {n}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.61, + 0.484, + 0.732 + ], + "angle": 0, + "content": "where \\(\\mathbf{r}_z\\) denotes the representative token in Community \\(\\mathcal{C}_z\\). Importantly, communities containing only a single token are discarded, as these tokens contain only client-specific information, following the principle of \"majority rule\" [30] in Voting. The selected tokens are then arranged in their original order (using, for example, a bubble sort [2]) to form a representative prompt, or the \"Representatives\", denoted as \\(\\mathcal{P}_{\\mathrm{new}} = \\{\\mathbf{r}_1,\\mathbf{r}_2,\\dots ,\\mathbf{r}_Z\\}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.745, + 0.208, + 0.76 + ], + "angle": 0, + "content": "4. Evaluation" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.769, + 0.26, + 0.787 + ], + "angle": 0, + "content": "4.1. Evaluation Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.791, + 0.484, + 0.838 + ], + "angle": 0, + "content": "Backbone API. In our experiments, we selected four backbone models: Qwen2-VL-72B [35], InternVL2-Llama3-76B, TeleMM [23], and GPT4o." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Dataset and Evaluation. We conduct experiments on two tasks: traditional image classification and Visual Question Answer (VQA). For image classification, we use the" + }, + { + "type": "code_caption", + "bbox": [ + 0.515, + 0.091, + 0.907, + 0.122 + ], + "angle": 0, + "content": "Algorithm 2 Semantic-similarity-Guided Evolutionary Computation for Discrete Prompt Optimization" + }, + { + "type": "algorithm", + "bbox": [ + 0.513, + 0.125, + 0.907, + 0.451 + ], + "angle": 0, + "content": "Input: I clients with discrete prompt tokens \\(\\{\\mathcal{P}_{\\mathrm{client}_i}\\}_{i = 1}^I\\) \nOutput: Global prompt tokens \\(\\mathcal{P}_{\\mathrm{new}}\\) \n1: Initialization Initial population \\(\\mathcal{P}_{\\mathrm{init}}\\) \n2: for each token \\(tk\\) in a prompt \\(\\in \\mathcal{P}_{\\mathrm{init}}\\) do \n3: Achieve mapping according to Eq.4 \n4: end for \n5: for each token \\(\\mathbf{t}_g\\) do \\(\\triangleright\\) Reputation Test \n6: for each token \\(\\mathbf{t}_h\\) do \n7: Semantic-similarity weight according to Eq.5 \n8: end for \n9: Get total semantic-similarity weight \\(\\alpha_{g} = \\sum_{h}\\alpha_{g,h}\\) \n10: end for \n11: Cluster tokens according to Eq.6 Aggregation \n12: for each Community \\(C_z\\in \\mathcal{C}\\) do Election \n13: Select representative token according to Eq.7 \n14: if \\(C_z\\) contains only one token then \n15: Discard Community \\(C_z\\) \n16: end if \n17: end for \n18: Form Representatives \\(\\mathcal{P}_{\\mathrm{new}} = \\{\\mathbf{r}_1,\\mathbf{r}_2,\\dots ,\\mathbf{r}_Z\\}\\) from selected representative tokens" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.474, + 0.907, + 0.596 + ], + "angle": 0, + "content": "MNIST [9], CIFAR-10, CIFAR-100, and TinyImageNet (TinyImg) [20] datasets. For VQA, we utilize the Art, Chemistry, and Finance sections of the MMMU dataset [43], along with the RealWorldQA (RWQA) dataset. The task objective is to leverage optimized prompts for accurate image classification on the image datasets, and for VQA, to assist in parsing textual content within the datasets to enable accurate classification or decision-making." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.597, + 0.907, + 0.658 + ], + "angle": 0, + "content": "We use Accuracy as the evaluation metric for the image classification task. To maintain consistency across tasks, we also constrain predictions in the VQA task to the range of available labels and evaluate performance using Accuracy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Comparison Baselines. We compare our method against eight state-of-the-art (SOTA) approaches: FedPrompt [46], FedPepTAO [4], FedTPG [29], HPT [37], CoOp [48], BBT [10], FedBPT [31], BPTVLM [41], and a manual prompting method (Manual Prompt). Notably, HPT [37], CoOp [48], BBT[10], and BPTVLM [41] are not originally designed for federated learning. To adapt these methods to a federated setting, we apply a weighted averaging aggregation approach, following the implementations of FedPrompt [46]and FedBPT [31]. The results are reported using federated versions of these models: FedAvg-HPT, FedAvg-CoOp, FedAvg-BBT, and Fed-BPTVLM. Additionally, we categorize these methods into White-box and Black-box approaches based on whether they require access to model parameters. The White-box methods include FedPrompt [46]," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2465" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.107, + 0.088, + 0.892, + 0.588 + ], + "angle": 0, + "content": "
BackboneMethodCV ClassificationVQAAverage
MNISTCifar-10Cifar-100TinyImgArtChemistryFinanceRWQA
Qwen2-VL-72BWhite BoxFedPrompt90.479.573.560.974.663.370.580.4
FedPepTAO86.778.372.959.569.158.868.577.9
FedTPG84.276.871.359.272.457.562.675.4
FedAvg-HPT86.978.470.258.775.753.565.978.3
FedAvg-CoOp85.778.171.459.270.261.863.376.5
Black BoxManual70.762.253.547.354.745.851.858.3
FedAvg-BBT78.972.362.751.763.553.960.369.7
FedBPT80.072.064.253.168.050.659.067.8
Fed-BPTVLM80.773.162.350.965.253.363.370.8
Our method90.578.571.857.874.060.867.278.5
InternVL2-Llama3-76BWhite BoxFedPrompt88.181.374.460.270.258.862.976.1
FedPepTAO87.377.871.258.565.255.363.075.4
FedTPG86.279.371.757.468.054.260.773.8
FedAvg-HPT85.178.772.358.469.153.062.074.3
FedAvg-CoOp85.779.171.658.171.856.263.971.7
Black BoxManual71.863.751.845.452.546.353.861.3
FedAvg-BBT80.071.165.252.760.251.953.870.5
FedBPT80.772.766.652.361.350.658.467.1
Fed-BPTVLM79.571.467.253.064.153.056.769.8
Our method89.679.071.557.370.755.963.675.5
TeleMMBlack BoxManual69.465.553.345.853.045.052.557.9
FedAvg-BBT80.970.663.149.156.948.856.464.1
FedBPT80.071.361.950.960.851.557.466.3
Fed-BPTVLM79.171.062.749.759.152.357.768.5
Our method88.478.969.357.464.655.761.675.0
GPT4oBlack BoxManual74.468.259.152.861.958.661.671.9
FedAvg-BBT81.076.767.656.164.657.563.669.2
FedBPT83.278.368.456.669.159.368.569.9
Fed-BPTVLM81.779.067.957.367.457.965.271.6
Our method92.585.373.962.776.866.772.878.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.597, + 0.908, + 0.626 + ], + "angle": 0, + "content": "Table 1. Performance Comparison on CV Classification and VQA Tasks. The accuracy results are measured in percentages (\\%). Bold values indicate the overall best results, and underlined values indicate the best results among Black Box methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.652, + 0.483, + 0.713 + ], + "angle": 0, + "content": "FedPepTAO [4], FedTPG [29], FedAvg-HPT, and FedAvg-CoOp, while the Black-box methods consist of Manual, FedAvg-BBT, FedBPT [31], and Fed-BPTVLM. Detailed descriptions of all baselines are provided in Appendix A." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Implementation. The federated learning setup for our experiments is as follows: we use a default of 10 clients, with a client participation rate of \\(100\\%\\) in each training round. Additionally, we adopt a \\(k\\)-shot mechanism for prompt tuning, following the approach in work [31], with the default value of \\(k\\) set to 50. For image classification datasets, we use the original training and test sets. For each Visual QA dataset, we randomly select 50 samples as the training set (from the validation set for MMMU and the test set for RealWorldQA), with the remaining samples used for testing. In terms of baseline implementations, we adapt traditional" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.652, + 0.907, + 0.804 + ], + "angle": 0, + "content": "prompt optimization methods (HPT, CoOp, BBT, BBT-VLM) to the federated learning framework by employing the classical Federated Averaging (FedAvg) approach. For black-box backbone model configurations, we extend soft prompt tuning methodologies, such as FedBPT, by mapping the soft tokens to discrete tokens, through the RoBERTa encoder's tokenization mechanism, before feeding them into API-based models. In addition, all experiments are conducted on two NVIDIA GeForce RTX A800 GPUs. Detailed hyperparameter settings are provided in Appendix B." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.817, + 0.71, + 0.832 + ], + "angle": 0, + "content": "4.2. Effectiveness Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We first evaluate the effectiveness of our method on traditional CV classification tasks and VQA tasks. Experimental results show that, compared to other methods, our method demonstrates significant performance improvements in both" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2466" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.482, + 0.123 + ], + "angle": 0, + "content": "tasks, particularly excelling in black-box environments, with only a small gap from white-box algorithms." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.142, + 0.483, + 0.385 + ], + "angle": 0, + "content": "Traditional CV Classification Tasks. In traditional CV classification tasks, our method achieves outstanding performance across all model configurations. For example, in the GPT4o black-box setting, our method achieves improvements of \\(11.2\\%\\), \\(7.0\\%\\), \\(5.5\\%\\), and \\(6.1\\%\\) over the closest-performing FedBPT [31] method on the MNIST, Cifar-10, Cifar-100, and TinyImg datasets, respectively. In the TeleMM black-box setting, our method also performs exceptionally well, achieving \\(88.4\\%\\) and \\(78.9\\%\\) accuracy on MNIST and Cifar-10, respectively, which is an improvement of \\(9.3\\%\\) and \\(8.3\\%\\) over FedAvg-BBT. Additionally, for the Qwen2-VL-72B model, our method achieves accuracies of \\(78.5\\%\\) and \\(71.8\\%\\) on the Cifar-10 and Cifar-100 datasets, respectively, representing improvements of \\(7.1\\%\\) and \\(5.3\\%\\) over FedBPT [31], further demonstrating its advantage in traditional CV tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.404, + 0.483, + 0.526 + ], + "angle": 0, + "content": "VQA Tasks. Our method also shows strong performance in VQA tasks. In the GPT4o black-box setting, our method achieves improvements of \\(11.1\\%\\), \\(7.4\\%\\), \\(4.3\\%\\), and \\(8.6\\%\\) over the FedBPT [31] method on the Art, Chemistry, Finance, and RWQA datasets, respectively. In TeleMM setting, our method's performance remains very stable, achieving improvements of \\(7.7\\%\\) and \\(10.9\\%\\) over FedAvg-BBT on the Art and RWQA datasets, respectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.535, + 0.395, + 0.551 + ], + "angle": 0, + "content": "4.3. Prompt Transferability Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.558, + 0.483, + 0.709 + ], + "angle": 0, + "content": "In the prompt transferability experiment, we evaluate the performance of prompts optimized on one backbone model when applied to another, as shown in Table 2. The results indicate that prompts exhibit a degree of transferability, maintaining high robustness across models with minimal accuracy loss, as referenced in Table 1. Additionally, compared to the Qwen2-VL transfer results, the InternVL2 transfer showed better accuracy, further validating the high applicability and robustness of prompts between certain models." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.718, + 0.48, + 0.766 + ], + "angle": 0, + "content": "
ModelCifar-10Cifar-100TinyImg
Qwen2-VL→TeleMM8070.558.1
InternVL2→GPT4o87.872.363.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.775, + 0.483, + 0.804 + ], + "angle": 0, + "content": "Table 2. Transferability Experiment Results. The accuracy results are measured in percentages \\((\\%)\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.427, + 0.851 + ], + "angle": 0, + "content": "4.4. Communication Overhead Experiment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In the communication overhead experiment, our method shows a significant advantage in parameter efficiency, as illustrated in Table 3. Specifically, white-box methods like" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.288 + ], + "angle": 0, + "content": "FedPrompt [46] and FedTPG [29] and most black-box methods such as FedAvg-BBT and Fed-BPTVLM require hundreds of thousands of parameters, whereas our method requires only 150. This substantial reduction effectively minimizes latency and resource consumption associated with communication. This advantage primarily arises from a key difference in algorithm design: while other methods use soft prompts for knowledge exchange between clients, our method employs discrete prompt tokens. Soft prompts typically require a large number of continuous parameters, whereas discrete prompt tokens enable knowledge transfer with only a small set of parameters, significantly reducing communication overhead." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.299, + 0.903, + 0.449 + ], + "angle": 0, + "content": "
TypeMethodParameters
White BoxFedPrompt614k
FedPepTAO1796k
FedTPG614k
FedAvg-HPT614k
FedAvg-CoOp614k
Black BoxFedAvg-BBT614k
FedBPT40.8k
Fed-BPTVLM614k
Our method150
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.458, + 0.907, + 0.514 + ], + "angle": 0, + "content": "Table 3. Communication Overhead Experiment results. The measurements are in Tensors within the PyTorch Framework. Bold values indicate the best results, and underlined values indicate the second-best results." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.546, + 0.679, + 0.561 + ], + "angle": 0, + "content": "4.5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In the improvements ablation study based on Qwen2-VL72B API, our method consistently achieves the best performance across all metrics, further validating the effectiveness of its key components, as shown in Table 4. Compared to cases where specific modules are removed or replaced, our method demonstrates significant improvements on the Cifar-10, Cifar-100, and TinyImg datasets. Server-side experiments. Our method improves accuracy on the Cifar-10 and Cifar-100 datasets by \\(2.0\\%\\) and \\(1.2\\%\\), respectively, compared to the setting without the clustering strategy. Furthermore, compared to the fixed clustering approach, our method achieves accuracy gains of \\(0.7\\%\\) on Cifar-10 and \\(2.5\\%\\) on Cifar-100, underscoring the importance of a dynamic clustering strategy. Client-side experiments. Removing the G-F mechanism results in a significant accuracy drop across all datasets. Furthermore, our method achieves a \\(1.0\\%\\) accuracy improvement on Cifar-100 and \\(1.6\\%\\) on TinyImg compared to the configuration without the E-EB strategy. Compared to the configuration without the PPR-CoT strategy, our method improves accuracy by \\(3.2\\%\\) on Cifar-10 and \\(2.5\\%\\) on Cifar-100, underscoring the pivotal role of the PPR-CoT strategy in enhancing model perfor" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2467" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.093, + 0.143, + 0.105 + ], + "angle": 0, + "content": "mance." + }, + { + "type": "table", + "bbox": [ + 0.096, + 0.119, + 0.481, + 0.263 + ], + "angle": 0, + "content": "
SiteMethodCifar-10Cifar-100TinyImg
Serverw Embedding-fedavg71.962.554.2
w Random search65.357.353.9
w/o Clustering Strategy76.570.655.9
w Fixed Clusters77.869.356.3
Clientw/o G-F Mechanism71.865.553.5
w/o E-EB Strategy76.970.856.2
w/o PPR-CoT75.369.356.7
Our method78.571.857.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.272, + 0.484, + 0.425 + ], + "angle": 0, + "content": "Table 4. The Improvements Ablation Studies Results. The accuracy is measured in percentages (\\%). Performance decreases with the removal or modification of any key component, underscoring the effectiveness of the improvements. \"w Embedding-fedavg\" denotes the use of a weighted averaging aggregation method, as in FedBPT [31] and FedAvg-BBT. \"w Random search\" refers to the removal of weight comparison, using random search instead to select representative tokens. \"w/o Clustering Strategy\" indicates the removal of the clustering step. \"w Fixed Clusters\" implies a fixed number of clusters, which affects the count of representative tokens." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.474, + 0.487, + 0.702 + ], + "angle": 0, + "content": "Initial Prompt Impact. We evaluate the impact of the Initial Prompt on performance using GPT-4o as the backbone black-box VLM for classification tasks across datasets, comparing it with three typical black-box algorithms. Figure 2 shows that different prompt configurations significantly affect model performance, with our algorithm consistently achieving the best outcomes. Specifically, detailed prompts enhance model accuracy by providing richer information, which supports more effective learning and adaptation. In contrast, shorter prompts reduce accuracy, indicating that limited information constrains performance. Thus, prompt design is critical in classification tasks, and our algorithm exhibits strong robustness and performance across various prompt configurations. We present all Initial Prompts in Appendix C." + }, + { + "type": "image", + "bbox": [ + 0.123, + 0.73, + 0.446, + 0.842 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.855, + 0.483, + 0.898 + ], + "angle": 0, + "content": "Figure 2. Initial Prompt Impact. Prompt-1 represents a \"shorter initial prompt,\" Prompt-2 a \"standard initial prompt,\" and Prompt-3 a \"detailed initial prompt.\"" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.091, + 0.892, + 0.108 + ], + "angle": 0, + "content": "4.6. Heterogeneous Data Robustness Experiment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.115, + 0.908, + 0.342 + ], + "angle": 0, + "content": "In the heterogeneous data robustness experiment, we use the Qwen2-VL-72B API as the backbone VLM model to evaluate the performance of different algorithms on heterogeneous data in the TinyImagenet dataset, as shown in Figure 3. Data is distributed across 10 clients using a Dirichlet Distribution [21] with varying hyperparameters dir, where a smaller Dirichlet parameter (e.g., dir=0.05) indicates greater data heterogeneity. Details of these heterogeneous data distributions are provided in Appendix D. The results demonstrate that all methods show some robustness to data heterogeneity, aligning with expectations [28, 31] for model stability on heterogeneous data. Additionally, as heterogeneity increases, our method exhibits superior performance, with minimal accuracy reduction, outperforming other methods under all heterogeneity conditions." + }, + { + "type": "image", + "bbox": [ + 0.553, + 0.355, + 0.866, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.466, + 0.909, + 0.522 + ], + "angle": 0, + "content": "Figure 3. Heterogeneous Data Robustness Experiment results on the TinyImageNet dataset. The parameter \"dir\" is the Dirichlet Distribution hyperparameter; smaller values of dir indicate higher data heterogeneity across the 10 clients." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.557, + 0.634, + 0.573 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We propose FDPT, the first federated black-box discrete prompt tuning framework based on visual language models. FDPT uses an Agent-based Client Prompt Optimization strategy to achieve token-level prompt optimization on the client side without traditional prediction logits, thereby preventing inversion attacks. On the server side, a Semantic-Similarity-Guided Evolutionary Computation framework is implemented to execute more precise zeroth-order optimization, which enables the obtention of representative global tokens. Moreover, discrete token exchange between clients offers high security and low communication overhead. Moreover, discrete tokens exhibit enhanced interpretability compared to soft prompts, which facilitates human experts in designing corresponding instructions. Our framework exhibits: (1) statistically significant accuracy enhancements in conventional visual recognition and cross-modal reasoning tasks; (2) optimized communication efficiency through adaptive token selection mechanisms; (3) cross-domain transferability of learned prompt representations; and (4) strengthened resilience to data distribution shifts, particularly in heterogeneous learning environments." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2468" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.228 + ], + "angle": 0, + "content": "Acknowledgements. This research was supported by the National Natural Science Foundation of China under Grant 52074305, 52274160, 51874300 and the National Natural Science Foundation of China-Shanxi Provincial People's Government Coal-based Low Carbon Joint Fund U1510115, the Academic Research Projects of Beijing Union University (No.ZK90202106) and R&D Program of Beijing Municipal Education Commission (KM202211417005)." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.241, + 0.188, + 0.258 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.266, + 0.482, + 0.348 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.352, + 0.483, + 0.379 + ], + "angle": 0, + "content": "[2] Owen Astrachan. Bubble sort: an archaeological algorithmic analysis. ACM Sigcse Bulletin, 35(1):1-5, 2003. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.381, + 0.483, + 0.434 + ], + "angle": 0, + "content": "[3] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.437, + 0.483, + 0.505 + ], + "angle": 0, + "content": "[4] Tianshi Che, Ji Liu, Yang Zhou, Jiaxiang Ren, Jiwen Zhou, Victor S Sheng, Huaiyu Dai, and Dejing Dou. Federated learning of large language models with parameter-efficient prompt tuning and adaptive optimization. arXiv preprint arXiv:2310.15080, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.508, + 0.483, + 0.577 + ], + "angle": 0, + "content": "[5] Simin Chen, Zihe Song, Mirazul Haque, Cong Liu, and Wei Yang. Nicsslowdown: Evaluating the efficiency robustness of neural image caption generation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15365-15374, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.579, + 0.483, + 0.647 + ], + "angle": 0, + "content": "[6] Simin Chen, Hanlin Chen, Mirazul Haque, Cong Liu, and Wei Yang. The dark side of dynamic routing neural networks: Towards efficiency backdoor injection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24585-24594, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.649, + 0.483, + 0.717 + ], + "angle": 0, + "content": "[7] Yiyi Chen, Heather Lent, and Johannes Bjerva. Text embedding inversion security for multilingual language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7808-7827, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.72, + 0.483, + 0.801 + ], + "angle": 0, + "content": "[8] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.804, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[9] Li Deng. The mnist database of handwritten digit images for machine learning research [best of the web]. IEEE signal processing magazine, 29(6):141-142, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.095, + 0.847, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[10] Shizhe Diao, Zhichao Huang, Ruijia Xu, Xuechun Li, Yong Lin, Xiao Zhou, and Tong Zhang. Black-box prompt learning for pre-trained language models. Trans. Mach. Learn. Res., 2023, 2023. 5" + }, + { + "type": "list", + "bbox": [ + 0.095, + 0.266, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.147 + ], + "angle": 0, + "content": "[11] Xiuwen Fang and Mang Ye. Robust federated learning with noisy and heterogeneous clients. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10072-10081, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.907, + 0.216 + ], + "angle": 0, + "content": "[12] Cong Gao, Wenfeng Li, Lijun He, and Lingchong Zhong. A distance and cosine similarity-based fitness evaluation mechanism for large-scale many-objective optimization. Engineering Applications of Artificial Intelligence, 133:108127, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.908, + 0.288 + ], + "angle": 0, + "content": "[13] Aditya Golatkar, Alessandro Achille, Yu-Xiang Wang, Aaron Roth, Michael Kearns, and Stefano Soatto. Mixed differential privacy in computer vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8376-8386, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.289, + 0.906, + 0.343 + ], + "angle": 0, + "content": "[14] Xiang Gu, Shuchao Pang, Anan Du, Yifei Wang, Jixiang Miao, and Jorge Díez. Dynamic multimodal prompt tuning: Boost few-shot learning with vlm-guided point cloud models. In ECAI, pages 761-768. IOS Press, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.345, + 0.906, + 0.413 + ], + "angle": 0, + "content": "[15] Tao Guo, Song Guo, Junxiao Wang, Xueyang Tang, and Wenchao Xu. Promptfl: Let federated participants cooperatively learn prompts instead of models-federated learning in age of foundation model. IEEE Transactions on Mobile Computing, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.415, + 0.906, + 0.468 + ], + "angle": 0, + "content": "[16] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli-hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.906, + 0.525 + ], + "angle": 0, + "content": "[17] Xinyuan Ji, Zhaowei Zhu, Wei Xi, Olga Gadyatskaya, Zi long Song, Yong Cai, and Yang Liu. Fedfixer: Mitigating heterogeneous label noise in federated learning. In AAAI, pages 12830-12838. AAAI Press, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.906, + 0.58 + ], + "angle": 0, + "content": "[18] Ruochen Jiao, Shaoyuan Xie, Justin Yue, Takami Sato, Lixu Wang, Yixuan Wang, Qi Alfred Chen, and Qi Zhu. Exploring backdoor attacks against large language model-based decision making. arXiv preprint arXiv:2405.20774, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.582, + 0.906, + 0.677 + ], + "angle": 0, + "content": "[19] Weirui Kuang, Bingchen Qian, Zitao Li, Daoyuan Chen, Dawei Gao, Xuchen Pan, Yuexiang Xie, Yaliang Li, Bolin Ding, and Jingren Zhou. Federatedscope-llm: A comprehensive package for fine-tuning large language models in federated learning. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 5260–5271, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[20] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3, 2015. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.707, + 0.906, + 0.762 + ], + "angle": 0, + "content": "[21] Qinbin Li, Yiqun Diao, Quan Chen, and Bingsheng He. Federated learning on non-iid data silos: An experimental study. In 2022 IEEE 38th international conference on data engineering (ICDE), pages 965-978. IEEE, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.763, + 0.906, + 0.816 + ], + "angle": 0, + "content": "[22] Tian Li, Anit Kumar Sahu, Ameet Talwalkar, and Virginia Smith. Federated learning: Challenges, methods, and future directions. IEEE signal processing magazine, 37(3):50-60, 2020. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[23] Xiang Li, Yiqun Yao, Xin Jiang, Xuezhi Fang, Chao Wang, Xinzhang Liu, Zihan Wang, Yu Zhao, Xin Wang, Yuyao Huang, Shuangyong Song, Yongxiang Li, Zheng Zhang, Bo Zhao, Aixin Sun, Yequan Wang, Zhongjiang He, Zhongyuan Wang, Xuelong Li, and Tiejun Huang. Tele-flm technical report, 2024. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2469" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.162 + ], + "angle": 0, + "content": "[24] Zexin Li, Bangjie Yin, Taiping Yao, Junfeng Guo, Shouhong Ding, Simin Chen, and Cong Liu. Sibling-attack: Rethinking transferable adversarial attacks against face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24626–24637, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.192 + ], + "angle": 0, + "content": "[25] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[26] Mulei Ma, Chenyu Gong, Liekang Zeng, Yang Yang, and Liantao Wu. Flocoff: Data heterogeneity resilient federated learning with communication-efficient edge offloading. IEEE Journal on Selected Areas in Communications, 42(11): 3262-3277, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[27] Yifei Ming and Yixuan Li. How does fine-tuning impact out-of-distribution detection for vision-language models? International Journal of Computer Vision, 132(2):596-609, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.319, + 0.482, + 0.376 + ], + "angle": 0, + "content": "[28] John Nguyen, Jianyu Wang, Kshitiz Malik, Maziar Sanjabi, and Michael Rabbat. Where to begin? on the impact of pre-training and initialization in federated learning. arXiv preprint arXiv:2206.15387, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.482, + 0.446 + ], + "angle": 0, + "content": "[29] Chen Qiu, Xingyu Li, Chaithanya Kumar Mummadi, Madan Ravi Ganesh, Zhenzhen Li, Lu Peng, and Wan-Yi Lin. Federated text-driven prompt generation for vision-language models. In The Twelfth International Conference on Learning Representations, 2024. 1, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.482, + 0.476 + ], + "angle": 0, + "content": "[30] Mathias Risse. Arguing for majority rule*. Journal of Political Philosophy, 12(1), 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.482, + 0.533 + ], + "angle": 0, + "content": "[31] Jingwei Sun, Ziyue Xu, Hongxu Yin, Dong Yang, Daguang Xu, Yiran Chen, and Holger R Roth. Fedbpt: Efficient federated black-box prompt tuning for large language models. arXiv preprint arXiv:2310.01467, 2023. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.534, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[32] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.659 + ], + "angle": 0, + "content": "[33] Lixu Wang, Shichao Xu, Xiao Wang, and Qi Zhu. Addressing class imbalance in federated learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10165-10173, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[34] Lixu Wang, Shichao Xu, Ruiqi Xu, Xiao Wang, and Qi Zhu. Non-transferable learning: A new approach for model ownership verification and applicability authorization. In International Conference on Learning Representation (ICLR-2022), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[35] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[36] Tao Wang, Yushu Zhang, Shuren Qi, Ruoyu Zhao, Xia Zhihua, and Jian Weng. Security and privacy on generative data in aigc: A survey. ACM Computing Surveys, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[37] Yubin Wang, Xinyang Jiang, De Cheng, Dongsheng Li, and Cairong Zhao. Learning hierarchical prompt with structured linguistic knowledge for vision-language models. In AAAI, pages 5749-5757. AAAI Press, 2024. 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[38] Yubo Wang, Chaohu Liu, Yanqiu Qu, Haoyu Cao, Deqiang Jiang, and Linli Xu. Break the visual perception: Adversarial attacks targeting encoded visual tokens of large vision-language models. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 1072-1081, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.177, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[39] Feijie Wu, Zitao Li, Yaliang Li, Bolin Ding, and Jing Gao. Fedbiot: Llm local fine-tuning in federated learning without full model. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 3345-3355, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.316 + ], + "angle": 0, + "content": "[40] Jiaqi Wu, Simin Chen, Yuzhe Yang, Yijiang Li, Shiyue Hou, Rui Jing, Zehua Wang, Wei Chen, and Zijian Tian. Feddtpt: federated discrete and transferable prompt tuning for black-box large language models. arXiv preprint arXiv:2411.00985, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.318, + 0.905, + 0.359 + ], + "angle": 0, + "content": "[41] Lang Yu, Qin Chen, Jiaju Lin, and Liang He. Black-box prompt tuning for vision-language model as a service. In IJCAI, pages 1686-1694, 2023. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.36, + 0.905, + 0.415 + ], + "angle": 0, + "content": "[42] Bowen Yuan, Sisi You, and Bing-Kun Bao. Self-pt: Adaptive self-prompt tuning for low-resource visual question answering. In ACM Multimedia, pages 5089-5098. ACM, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.417, + 0.905, + 0.514 + ], + "angle": 0, + "content": "[43] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9556-9567, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.515, + 0.905, + 0.571 + ], + "angle": 0, + "content": "[44] Jingjing Zhang, Shancheng Fang, Zhendong Mao, Zhiwei Zhang, and Yongdong Zhang. Fine-tuning with multi-modal entity prompts for news image captioning. In ACM Multimedia, pages 4365-4373. ACM, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.572, + 0.905, + 0.627 + ], + "angle": 0, + "content": "[45] Jingyi Zhang, Jiaxing Huang, Sheng Jin, and Shijian Lu. Vision-language models for vision tasks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.629, + 0.905, + 0.711 + ], + "angle": 0, + "content": "[46] Haodong Zhao, Wei Du, Fangqi Li, Peixuan Li, and Gongshen Liu. Fedprompt: Communication-efficient and privacy-preserving prompt tuning in federated learning. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.712, + 0.905, + 0.769 + ], + "angle": 0, + "content": "[47] Xuejun Zhao, Wencan Zhang, Xiaokui Xiao, and Brian Lim. Exploiting explanations for model inversion attacks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 682-692, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.77, + 0.905, + 0.825 + ], + "angle": 0, + "content": "[48] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.826, + 0.905, + 0.88 + ], + "angle": 0, + "content": "[49] Yu Zhu, Kang Li, Lequan Yu, and Pheng-Ann Heng. Memory-efficient prompt tuning for incremental histopathology classification. In AAAI, pages 7802-7810. AAAI Press, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.88 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "2470" + } + ] +] \ No newline at end of file diff --git a/2025/FDPT_ Federated Discrete Prompt Tuning for Black-Box Visual-Language Models/5d5a1b2f-9957-4fb4-b461-7bf3e1ebccfb_origin.pdf b/2025/FDPT_ Federated Discrete Prompt Tuning for Black-Box Visual-Language Models/5d5a1b2f-9957-4fb4-b461-7bf3e1ebccfb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e5271696da42bbcc488f26c8991847c799f2723b --- /dev/null +++ b/2025/FDPT_ Federated Discrete Prompt Tuning for Black-Box Visual-Language Models/5d5a1b2f-9957-4fb4-b461-7bf3e1ebccfb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7e77759311c5ec215b41c0829835d98dea3446c5140161c8bcddf7f8195473ab +size 1010182 diff --git a/2025/FDPT_ Federated Discrete Prompt Tuning for Black-Box Visual-Language Models/full.md b/2025/FDPT_ Federated Discrete Prompt Tuning for Black-Box Visual-Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..722c0671d3678cf59f7cc27be636c52d384ef30b --- /dev/null +++ b/2025/FDPT_ Federated Discrete Prompt Tuning for Black-Box Visual-Language Models/full.md @@ -0,0 +1,305 @@ +# FDPT: Federated Discrete Prompt Tuning for Black-Box Visual-Language Models + +Jiaqi Wu $^{1,8,10\dagger}$ , Simin Chen $^{2\dagger}$ , Jing Tang $^{3}$ , Yuzhe YANG $^{4}$ , Yiming Chen $^{5}$ , Lixu Wang $^{6}$ , Song Lin $^{7}$ , Zehua Wang $^{8,9}$ , Wei Chen $^{9*}$ , and Zijian Tian $^{10}$ + +$^{1}$ Tsinghua University, $^{2}$ University of Texas at Dallas, $^{3}$ Huazhong University of Science and Technology, + +4The Chinese University of Hong Kong, 5National University of Singapore, 6Nanyang Technological University, + +7Beijing Baidu Netcom Science Technology Co., Ltd., 8University of British Columbia, + +$^{9}$ China University of Mining and Technology, $^{10}$ China University of Mining and Technology (Beijing) + +wjq11346@student.ubc.ca, sxc180080@utdallas.edu, focusers@163.com, yuzheyang@link.cuhk.edu.cn, + +yiming.chen@u.nus.edu, wanglixu4334@gmail.com, linsong04@baidu.com, zwang@ece.ubc.ca, + +chenwdavior@163.com，Tianzj0726@126.com + +# Abstract + +Vision-Language Models (VLMs) have driven significant advancements in multimodal AI. Fine-tuning these models with user data enhances adaptability but poses privacy risks. While federated learning (FL) mitigates user data privacy concerns, it fails to protect model properties. Existing methods relying on black-box VLM APIs often require access to prediction logits, making them vulnerable to inversion attacks. Additionally, optimizing tuning complexity and data transmission efficiency in federated VLM scenarios remains a challenge. To address these challenges, we propose FDPT—the first federated discrete prompt tuning method utilizing black-box VLMs. During client optimization stage, FDPT employs an agent-driven framework leveraging large language models (LLMs) with enhanced reasoning capacities to systematically optimize discrete prompt representations, and also utilizes feedback mechanisms and chain of thought to enhance prediction accuracy. Importantly, it performs optimization by relying not on the predicted logic vectors output by LLMs but on textual results, avoiding reverse attack risks. During global aggregation stage, We mimic human electoral activities by employing evolutionary computation methods underpinned by semantic similarity computation to implement enhanced zero-order optimization for acquiring representative global tokens, thereby achieving knowledge aggregation. FDPT significantly outperforms nine state-of-the-art methods in image classification and visual question-answering, reducing communication overhead while generating highly transferable optimized prompts. Additionally, it exhibits im + +proved robustness to data heterogeneity. + +# 1. Introduction + +General-purpose Vision-Language Models (VLMs) [45] integrate vast visual and linguistic information, advancing various vision tasks [36]. To better adapt to user needs, these models can undergo fine-tuning or prompt-tuning [27], but such modifications introduce privacy risks [13, 18, 34, 38]. Federated Learning (FL) [22] addresses these concerns by decentralizing data storage, reducing single-point breaches, and exchanging model weights or embeddings instead of raw data. Various FL-based approaches [19, 29, 39] have been explored for tuning large models, effectively preserving user data privacy. + +Despite the success of FL, protecting the privacy of VLMs themselves remains a critical yet largely unexplored challenge. Existing FL approaches use zero-order optimization to estimate gradients for prompt tuning on black-box VLMs [31, 41]. For example, FedBPT [31] employs prompt exchange and gradient-free optimization, reducing communication and memory costs by avoiding direct parameter access and backpropagation on client devices. However, several limitations remain. First, these methods remain vulnerable to security risks as they rely on prediction logits, exposing them to inversion attacks that can infer model parameters [7, 47]. Second, zero-order optimization techniques [31, 41] often suffer from limited accuracy and adaptability in prompt fine-tuning. Finally, soft prompt methods [15, 46] lack interpretability and incur high communication costs in federated settings. + +To address these challenges, we propose FDPT, the first + +federated discrete prompt optimization method comprising two stages: client-side optimization and global aggregation. In the client-side optimization phase, we introduce AI Agent-based Client Prompt Optimization, which refines token-level prompts using large language models (LLMs). We design a feedback mechanism and Chain-of-Thought measures to guide the agent in selecting optimal discrete tokens, overcoming the limitations of existing zero-order optimization methods in prompt tuning. Unlike traditional approaches that rely on prediction logits, our method uses only black-box VLM text outputs, reducing the risk of inversion attacks and enhancing model privacy protection. Afterwards, in the global aggregation phase, to reduce communication overhead in federated settings, we use discrete tokens for knowledge exchange instead of soft prompts. Inspired by human voting mechanisms, we propose a Self-Attention-Guided Evolutionary Computation method to identify representative tokens for constructing global prompts. Unlike the coarse sampling in traditional zero-order optimization, our approach clusters tokens in semantic space and then applies similarity calculations to refine the most accurate results, enhancing tuning efficiency. + +Detailed experimental results demonstrate the advantages of our method. In Visual Question Answering (VQA) tasks, under the GPT-4 black-box setting, our method outperforms the classic black-box prompt optimization method FedBPT [31], achieving $11.1\%$ and $7.4\%$ improvements on the Art and Chemistry categories of the MMMU dataset [43], respectively. In traditional computer vision (CV) classification tasks, using the Qwen2-VL-72B model [3], our approach achieves $71.8\%$ accuracy on CIFAR-100, surpassing FedBPT by $5.3\%$ , demonstrating its superior performance. Additionally, FDPT significantly reduces communication costs without compromising optimization performance and enhances interpretability through discrete prompts. The contributions of this paper are as follows: + +- We propose FDPT, a first federated black-box discrete prompt optimization framework that aggregates knowledge from multiple parties to optimize prompts while ensuring the privacy of both users and easily overlooked VLMs. +- We introduce an AI Agent-based Client Prompt Optimization for token-level prompts and a Self-attention-guided Evolutionary Computation method for selecting global tokens, offering better predictive performance than traditional zero-order optimization. +- Extensive experiments demonstrate that our method achieves superior performance across various text-vision tasks, exhibits low communication overhead, and is resilient to data heterogeneity. + +# 2. Background & Related Work + +Visual Language Models. Visual-Language Models (VLMs) [5, 45] are AI systems designed to process and understand both visual and textual information, enabling them to perform tasks such as Visual Question Answering (VQA) [43] and traditional visual classification tasks [20]. Traditional VLMs, like Qwen-VL [3], effectively integrate visual features with textual data, excelling in both VQA and image classification tasks. Flamingo [1] leverages few-shot learning capabilities, allowing it to quickly adapt to new tasks and perform well across both VQA and classification. Gemini [32] focuses on reasoning and multimodal understanding, achieving strong results in complex VQA tasks while also performing well in image and text classification. LLaVA [25] benefits from large-scale pretraining, efficiently handling intricate visuals and long-form text, with exceptional performance in both VQA and classification tasks. Newer models like InternVL2 [8], Qwen2-VL [35], and GPT-4o [16] build on these foundations, enhancing multimodal reasoning, improving the fusion of visual and textual data, and offering broader adaptability across both VQA and classification tasks. + +Federated Learning. Federated Learning (FL) [22] is a decentralized approach allowing multiple clients to collaboratively train a shared model without sharing raw data, thus preserving privacy and defending against attacks [6, 24]. Recent works have further improved FL's robustness in handling data heterogeneity [26, 33], label noise [17], and personalization needs [11]. With the rise of large language models (LLMs), Federated Learning (FL) has gained new applications, effectively addressing the challenges of knowledge exchange and privacy protection between independent LLMs. FedPrompt [46] combines FL with prompt tuning, utilizing split aggregation to significantly reduce communication costs. FedBPT [31] presents a federated black-box prompt tuning framework that avoids direct access to model parameters, instead optimizing prompts using gradient-free methods. FedDTPT [40] realizes discrete prompt tuning for Large Language Models (LLMs). However, these methods still face issues such as security risks in large models and high communication overhead. We propose a Federated Discrete Prompt Tuning method, which, compared to existing approaches, performs zero-order optimization at the token level, mitigating the risk of adversarial users leveraging logits generated by the large model for backdoor attacks. Additionally, by utilizing discrete tokens for knowledge exchange, it significantly reduces communication overhead. + +Prompt Tuning. Prompt tuning [49] is an emerging technique for adapting large pre-trained models to specific tasks + +![](images/750758aadf814a2906dd59c770d1ad113ac648049b233a009a000c5306cbf8d5.jpg) +Figure 1. Overview of FDPT. The proposed framework consists of the client-side local Agent-based Client Prompt Optimization and the server-side Semantic-Similarity-Guided Evolutionary Computation for global selection. The detailed process is described in Sec. 3.2. + +with minimal additional training. In VLMs, prompt tuning [14] bridges visual and textual modalities, enabling tasks like image captioning [42, 44]. The Prompt tuning can be represented as finding an optimal prompt $\mathbf{p}^*$ to maximize the model's performance on the dataset $\mathcal{D} = \{(\mathbf{x},\mathbf{y})\}$ : + +$$ +\mathbf {p} ^ {*} = \arg \max f _ {\mathrm {V L M}} (\mathbf {p}; \mathcal {D}) \tag {1} +$$ + +where $f_{\mathrm{VLM}}(\mathbf{p};\mathcal{D})$ denotes the performance of the VLM using prompt $\mathbf{p}$ over the dataset $\mathcal{D}$ . Continual Prompt Tuning (soft prompt tuning) involves fine-tuning a small set of trainable parameters (soft prompts) while keeping the core model fixed. However, it suffers from poor interpretability, and in Federated Learning (FL), it incurs high communication overhead between clients. Discrete Prompt Tuning (DPT), in contrast, optimizes specific discrete tokens from a fixed vocabulary. A notable approach, RLPrompt, uses reinforcement learning (RL) to select the best tokens based on rewards. However, it has limited optimization efficiency due to the computationally intensive nature of RL. Regarding above limitations, the proposed Federated Discrete Prompt Tuning method aims to combine the strengths of discrete prompt tuning with the constraints and needs of Federated Learning. By leveraging an AI agent (based on large language models) with robust reasoning capabilities, the method can generate more accurate and efficient local discrete tokens. Furthermore, by incorporating semantic similarity calculations into an evolutionary computing framework, the method can generate more precise global tokens that are shared across clients. Additionally, this reduces communication costs between clients, as discrete tokens are much smaller in parameters compared to soft prompts. + +# 3. Method + +# 3.1. Problem Formulation + +In this work, we consider a distributed scenario [22], where $N$ clients each possess a private dataset $\mathcal{D}_k = \{(\mathcal{X}_k,\mathcal{Y}_k)\}$ consisting of $n_k$ samples $\{(x_i^k,y_i^k)\}_{i\in [n_k]}$ . Each client can only access its own private data and does not wish to share it with others. This scenario is realistic, as hospitals and other organizations prioritize protecting sensitive data. Moreover, such distributed settings have been widely studied in existing research. + +Then, our federated discrete prompt optimization can be formulated as an optimization problem that aims to find the optimal prompt instruction $p^*$ that satisfies the following equation: + +$$ +p ^ {*} = \arg \min _ {p} \sum_ {k \in [ K ]} \frac {n ^ {k}}{\sum_ {k \in [ K ]} n ^ {k}} \sum_ {i \in [ n ^ {k} ]} \mathcal {L} \left(\mathrm {V L M} \left(p; x _ {i} ^ {k}\right), y _ {i} ^ {k}\right) \tag {2} +$$ + +where $\mathcal{L}(\cdot)$ is the objective function, which could be the error rate, and $\mathrm{VLM}(\cdot)$ represents the Vision-Language Model (VLM). The above equation can be interpreted as learning a globally optimal natural language prompt by minimizing the weighted average loss across all clients, where each client's contribution is proportional to its dataset size. + +Notice that if clients were permitted to share their private data and had white-box access to the VLM, the equation above could be easily solved by approximating the objective function with a differentiable function, such as cross-entropy loss. In this case, standard backpropagation could be applied to the global dataset to search for the optimal prompt instruction. However, due to commercial concerns, modern VLMs are typically deployed as API ser + +vices, which are black-boxes to the clients. Additionally, each client is unwilling to share their data with others due to privacy concerns. Consequently, we must impose the following constraints: + +Blackbox: $\frac{\partial\mathcal{L}(\mathrm{VLM}(p;x),y)}{\partial p} = \emptyset$ + +Privacy: $\forall k,j\in [K],k\neq j,$ $\mathrm{Client}_k$ cannot access $\mathcal{D}_j$ (3) + +The first constraint indicates that the gradient is not feasible, while the second constraint specifies that each client can only access its own data. + +# 3.2. Design Overview + +The algorithm overview is illustrated in Figure 1. Consistent with the traditional federated learning paradigm [22], each client communicates solely with the server using the prompt information, while keeping their data local. This ensures compliance with the data privacy constraints outlined in Eq.(3). Specifically, at the beginning, the server distributes the Global Prompt to each client. Then our approach follows these steps during each global communication round to search an optimal prompt: (1) Client Optimization: Each client utilizes an Agent-based Client Prompt Optimization method for personalized token-level prompt optimization. This method is preferred over backpropagation and ensures adherence to the black-box constraints in Eq.(3), which is more realistic in the context of VLM as a service; (2) Server Aggregation: Each client uploads its locally optimized prompt to the server, where knowledge aggregation is performed using the Semantic-similarity-guided Evolutionary Computation method. The subsequent sections provide a detailed explanation of each step in our approach. + +# 3.3. Client Optimization + +Upon receiving the global discrete prompt $p_0$ from the server in the current round, each client iteratively updates it using an agent-based approach, as shown in Alg. 1. Initially, the agent queries the LLM to randomly replace some tokens in $p_0$ , generating a set of prompt mutants $p_1, p_2, \ldots, p_M$ (Line 1). Each prompt is then combined with the client's local dataset and fed into the VLM to measure the accuracy, which serves as the prompt mutant's score. FDPT maintains a priority queue to manage these mutants, prioritizing those with higher accuracy (Lines 2-6). FDPT iteratively selects the highest-priority prompt from the queue as the seed, asks the LLM to mutate the important tokens of the seed, and updates the priority queue with the newly generated prompt mutants (Lines 7-13). Specifically, token importance is computed as follows: (1), FDPT compares each prompt mutant $p_i$ with the seed prompt $p_0$ token by token, constructing a difference matrix $M$ , where $M_i^j = 0$ if the + +Algorithm 1 Agent-Based Prompt Optimization +Input: Initial discrete prompt $p_0$ , Large Language Model (LLM), Vision-Language Model (VLM), client local dataset $\mathcal{D}$ , number of mutants $M$ +Output: Optimized prompt $\hat{p}$ +1: Initialization: $\{p_1, p_2, \ldots, p_M\} =$ Mutator( $LLM, p_0$ ) $\triangleright$ Mutate $p_0$ by randomly replacing tokens +2: $\mathcal{Q} =$ Priority Queue $\triangleright$ Initialize an empty queue +3: for each prompt $p_i$ in $\{p_0, p_1, \ldots, p_M\}$ do +4: $Acc_i =$ Evaluate(VLM, $p_i$ , $\mathcal{D}$ ) $\triangleright$ Evaluation +5: Push $(\mathcal{Q}, p_i, Acc_i)$ $\triangleright$ Update the priority queue +6: end for +7: while stopping criterion not met do +8: $p_{seed} =$ Pop $(\mathcal{Q}) \triangleright$ Get the highest-priority prompt +9: Position = Importance Estimator $\triangleright$ Select important tokens +10: $p_{new} =$ Mutator( $LLM, p_{seed}$ , Position) +11: $Acc_{new} =$ Evaluate(VLM, $p_{new}$ , $\mathcal{D}$ ) $\triangleright$ Evaluation +12: Push $(\mathcal{Q}, p_{new}, Acc_{new})$ $\triangleright$ Update the queue +13: end while +14: Return highest-scoring prompt from $\mathcal{Q}$ + +$j$ -th token of $p_i$ and $p_0$ are different, and $M_i^j = 1$ otherwise. (2), FDPT computes the accuracy gains $V$ , where $V_i = \mathrm{Acc}_i - \mathrm{Acc}_0$ . (3) The importance score of each token position is then given by $S = M^{-1}V$ . This score estimation equation reflects the degree to which a token's presence or absence affects accuracy, weighted by its position in the prompt. Finally, the top $K$ token positions are selected based on the importance score $S$ . + +# 3.4. Server Aggregation + +After local client optimization, each client uploads its optimized discrete prompt to the server for knowledge exchange. The server then aggregates these local optimal prompts for the next round of optimization. Specifically, the server aggregation includes the following steps (as shown in Algorithm 2): + +Token-level Semantic Representation. In total client uploaded prompts, e.g. the initial population $\mathcal{P}_{\mathrm{init}}$ , for each token in a prompt, we use a pre-trained text encoder of the Roberta to get its token-level semantic representation, denoted as + +$$ +\mathcal {E} _ {i} = [ e _ {i} ^ {1}, e _ {i} ^ {2}, \dots , e _ {i} ^ {K} ] e _ {i} ^ {j} = f _ {e m b e d} (t k _ {i} ^ {j}) \qquad (4) +$$ + +where $f_{embed}(\cdot)$ is the pre-trained encoder, and $tk_i^j$ is the $j^{th}$ token of the $i^{th}$ prompt. + +Reputation Test. We first evaluate the reputation of each token, defined as the semantic-similarity weight (denoted + +$\alpha_{g,h})$ that each token assigns to all other tokens. This is computed using a cosine-similarity mechanism [12] based on embeddings: + +$$ +\alpha_ {g} = \sum_ {h} \alpha_ {g, h} = \sum_ {h} \operatorname {s o f t m a x} \left(\frac {\mathbf {e} _ {g} \cdot \mathbf {e} _ {h}}{\sqrt {d}}\right), \tag {5} +$$ + +where $\mathbf{e}_g$ and $\mathbf{e}_h$ represent the embedding of a token $g$ and $h$ , respectively, and $d$ is the dimensionality of the embeddings. A higher semantic-similarity weight $\alpha_{g,h}$ signifies a stronger influence over other tokens. We define the total semantic-similarity weight of token $g$ as $\alpha_{g}$ , which represents the token's semantic importance among all other tokens, referred to as its Reputation. + +Aggregation. Next, tokens are clustered based on their semantic similarity by grouping embeddings that are close to each other in the high-dimensional semantic space: + +$$ +\mathcal {C} = \left\{\mathcal {C} _ {1}, \mathcal {C} _ {2}, \dots , \mathcal {C} _ {Z} \right\}, \mathcal {C} _ {z} = \left\{\mathbf {e} _ {h} \mid \| \mathbf {e} _ {h} - \mathbf {e} _ {g} \| < \epsilon , \forall h \in \mathcal {P} _ {\text {i n i t}} \right\}, \tag {6} +$$ + +where $\epsilon$ is a pre-defined distance threshold. $C_z$ denotes a cluster, or Community. Embeddings with equivalent syntactic meanings are grouped into a cluster, from which a single representative is subsequently selected. + +Election. Finally, within each cluster (Community), the representative token is determined by semantic-similarity weights (reputation). Tokens with higher weights have stronger semantic dependencies on other tokens and are more likely to be selected: + +$$ +\mathbf {r} _ {z} = \underset {\mathbf {e} _ {n} \in \mathcal {C} _ {z}} {\arg \max } \alpha_ {n}, \tag {7} +$$ + +where $\mathbf{r}_z$ denotes the representative token in Community $\mathcal{C}_z$ . Importantly, communities containing only a single token are discarded, as these tokens contain only client-specific information, following the principle of "majority rule" [30] in Voting. The selected tokens are then arranged in their original order (using, for example, a bubble sort [2]) to form a representative prompt, or the "Representatives", denoted as $\mathcal{P}_{\mathrm{new}} = \{\mathbf{r}_1,\mathbf{r}_2,\dots ,\mathbf{r}_Z\}$ . + +# 4. Evaluation + +# 4.1. Evaluation Setup + +Backbone API. In our experiments, we selected four backbone models: Qwen2-VL-72B [35], InternVL2-Llama3-76B, TeleMM [23], and GPT4o. + +Dataset and Evaluation. We conduct experiments on two tasks: traditional image classification and Visual Question Answer (VQA). For image classification, we use the + +Algorithm 2 Semantic-similarity-Guided Evolutionary Computation for Discrete Prompt Optimization +Input: I clients with discrete prompt tokens $\{\mathcal{P}_{\mathrm{client}_i}\}_{i = 1}^I$ +Output: Global prompt tokens $\mathcal{P}_{\mathrm{new}}$ +1: Initialization Initial population $\mathcal{P}_{\mathrm{init}}$ +2: for each token $tk$ in a prompt $\in \mathcal{P}_{\mathrm{init}}$ do +3: Achieve mapping according to Eq.4 +4: end for +5: for each token $\mathbf{t}_g$ do $\triangleright$ Reputation Test +6: for each token $\mathbf{t}_h$ do +7: Semantic-similarity weight according to Eq.5 +8: end for +9: Get total semantic-similarity weight $\alpha_{g} = \sum_{h}\alpha_{g,h}$ +10: end for +11: Cluster tokens according to Eq.6 Aggregation +12: for each Community $C_z\in \mathcal{C}$ do Election +13: Select representative token according to Eq.7 +14: if $C_z$ contains only one token then +15: Discard Community $C_z$ +16: end if +17: end for +18: Form Representatives $\mathcal{P}_{\mathrm{new}} = \{\mathbf{r}_1,\mathbf{r}_2,\dots ,\mathbf{r}_Z\}$ from selected representative tokens + +MNIST [9], CIFAR-10, CIFAR-100, and TinyImageNet (TinyImg) [20] datasets. For VQA, we utilize the Art, Chemistry, and Finance sections of the MMMU dataset [43], along with the RealWorldQA (RWQA) dataset. The task objective is to leverage optimized prompts for accurate image classification on the image datasets, and for VQA, to assist in parsing textual content within the datasets to enable accurate classification or decision-making. + +We use Accuracy as the evaluation metric for the image classification task. To maintain consistency across tasks, we also constrain predictions in the VQA task to the range of available labels and evaluate performance using Accuracy. + +Comparison Baselines. We compare our method against eight state-of-the-art (SOTA) approaches: FedPrompt [46], FedPepTAO [4], FedTPG [29], HPT [37], CoOp [48], BBT [10], FedBPT [31], BPTVLM [41], and a manual prompting method (Manual Prompt). Notably, HPT [37], CoOp [48], BBT[10], and BPTVLM [41] are not originally designed for federated learning. To adapt these methods to a federated setting, we apply a weighted averaging aggregation approach, following the implementations of FedPrompt [46]and FedBPT [31]. The results are reported using federated versions of these models: FedAvg-HPT, FedAvg-CoOp, FedAvg-BBT, and Fed-BPTVLM. Additionally, we categorize these methods into White-box and Black-box approaches based on whether they require access to model parameters. The White-box methods include FedPrompt [46], + +
BackboneMethodCV ClassificationVQAAverage
MNISTCifar-10Cifar-100TinyImgArtChemistryFinanceRWQA
Qwen2-VL-72BWhite BoxFedPrompt90.479.573.560.974.663.370.580.4
FedPepTAO86.778.372.959.569.158.868.577.9
FedTPG84.276.871.359.272.457.562.675.4
FedAvg-HPT86.978.470.258.775.753.565.978.3
FedAvg-CoOp85.778.171.459.270.261.863.376.5
Black BoxManual70.762.253.547.354.745.851.858.3
FedAvg-BBT78.972.362.751.763.553.960.369.7
FedBPT80.072.064.253.168.050.659.067.8
Fed-BPTVLM80.773.162.350.965.253.363.370.8
Our method90.578.571.857.874.060.867.278.5
InternVL2-Llama3-76BWhite BoxFedPrompt88.181.374.460.270.258.862.976.1
FedPepTAO87.377.871.258.565.255.363.075.4
FedTPG86.279.371.757.468.054.260.773.8
FedAvg-HPT85.178.772.358.469.153.062.074.3
FedAvg-CoOp85.779.171.658.171.856.263.971.7
Black BoxManual71.863.751.845.452.546.353.861.3
FedAvg-BBT80.071.165.252.760.251.953.870.5
FedBPT80.772.766.652.361.350.658.467.1
Fed-BPTVLM79.571.467.253.064.153.056.769.8
Our method89.679.071.557.370.755.963.675.5
TeleMMBlack BoxManual69.465.553.345.853.045.052.557.9
FedAvg-BBT80.970.663.149.156.948.856.464.1
FedBPT80.071.361.950.960.851.557.466.3
Fed-BPTVLM79.171.062.749.759.152.357.768.5
Our method88.478.969.357.464.655.761.675.0
GPT4oBlack BoxManual74.468.259.152.861.958.661.671.9
FedAvg-BBT81.076.767.656.164.657.563.669.2
FedBPT83.278.368.456.669.159.368.569.9
Fed-BPTVLM81.779.067.957.367.457.965.271.6
Our method92.585.373.962.776.866.772.878.5
+ +Table 1. Performance Comparison on CV Classification and VQA Tasks. The accuracy results are measured in percentages (\%). Bold values indicate the overall best results, and underlined values indicate the best results among Black Box methods. + +FedPepTAO [4], FedTPG [29], FedAvg-HPT, and FedAvg-CoOp, while the Black-box methods consist of Manual, FedAvg-BBT, FedBPT [31], and Fed-BPTVLM. Detailed descriptions of all baselines are provided in Appendix A. + +Implementation. The federated learning setup for our experiments is as follows: we use a default of 10 clients, with a client participation rate of $100\%$ in each training round. Additionally, we adopt a $k$ -shot mechanism for prompt tuning, following the approach in work [31], with the default value of $k$ set to 50. For image classification datasets, we use the original training and test sets. For each Visual QA dataset, we randomly select 50 samples as the training set (from the validation set for MMMU and the test set for RealWorldQA), with the remaining samples used for testing. In terms of baseline implementations, we adapt traditional + +prompt optimization methods (HPT, CoOp, BBT, BBT-VLM) to the federated learning framework by employing the classical Federated Averaging (FedAvg) approach. For black-box backbone model configurations, we extend soft prompt tuning methodologies, such as FedBPT, by mapping the soft tokens to discrete tokens, through the RoBERTa encoder's tokenization mechanism, before feeding them into API-based models. In addition, all experiments are conducted on two NVIDIA GeForce RTX A800 GPUs. Detailed hyperparameter settings are provided in Appendix B. + +# 4.2. Effectiveness Results + +We first evaluate the effectiveness of our method on traditional CV classification tasks and VQA tasks. Experimental results show that, compared to other methods, our method demonstrates significant performance improvements in both + +tasks, particularly excelling in black-box environments, with only a small gap from white-box algorithms. + +Traditional CV Classification Tasks. In traditional CV classification tasks, our method achieves outstanding performance across all model configurations. For example, in the GPT4o black-box setting, our method achieves improvements of $11.2\%$ , $7.0\%$ , $5.5\%$ , and $6.1\%$ over the closest-performing FedBPT [31] method on the MNIST, Cifar-10, Cifar-100, and TinyImg datasets, respectively. In the TeleMM black-box setting, our method also performs exceptionally well, achieving $88.4\%$ and $78.9\%$ accuracy on MNIST and Cifar-10, respectively, which is an improvement of $9.3\%$ and $8.3\%$ over FedAvg-BBT. Additionally, for the Qwen2-VL-72B model, our method achieves accuracies of $78.5\%$ and $71.8\%$ on the Cifar-10 and Cifar-100 datasets, respectively, representing improvements of $7.1\%$ and $5.3\%$ over FedBPT [31], further demonstrating its advantage in traditional CV tasks. + +VQA Tasks. Our method also shows strong performance in VQA tasks. In the GPT4o black-box setting, our method achieves improvements of $11.1\%$ , $7.4\%$ , $4.3\%$ , and $8.6\%$ over the FedBPT [31] method on the Art, Chemistry, Finance, and RWQA datasets, respectively. In TeleMM setting, our method's performance remains very stable, achieving improvements of $7.7\%$ and $10.9\%$ over FedAvg-BBT on the Art and RWQA datasets, respectively. + +# 4.3. Prompt Transferability Evaluation + +In the prompt transferability experiment, we evaluate the performance of prompts optimized on one backbone model when applied to another, as shown in Table 2. The results indicate that prompts exhibit a degree of transferability, maintaining high robustness across models with minimal accuracy loss, as referenced in Table 1. Additionally, compared to the Qwen2-VL transfer results, the InternVL2 transfer showed better accuracy, further validating the high applicability and robustness of prompts between certain models. + +
ModelCifar-10Cifar-100TinyImg
Qwen2-VL→TeleMM8070.558.1
InternVL2→GPT4o87.872.363.5
+ +# 4.4. Communication Overhead Experiment + +In the communication overhead experiment, our method shows a significant advantage in parameter efficiency, as illustrated in Table 3. Specifically, white-box methods like + +FedPrompt [46] and FedTPG [29] and most black-box methods such as FedAvg-BBT and Fed-BPTVLM require hundreds of thousands of parameters, whereas our method requires only 150. This substantial reduction effectively minimizes latency and resource consumption associated with communication. This advantage primarily arises from a key difference in algorithm design: while other methods use soft prompts for knowledge exchange between clients, our method employs discrete prompt tokens. Soft prompts typically require a large number of continuous parameters, whereas discrete prompt tokens enable knowledge transfer with only a small set of parameters, significantly reducing communication overhead. + +Table 2. Transferability Experiment Results. The accuracy results are measured in percentages $(\%)$ . + +
TypeMethodParameters
White BoxFedPrompt614k
FedPepTAO1796k
FedTPG614k
FedAvg-HPT614k
FedAvg-CoOp614k
Black BoxFedAvg-BBT614k
FedBPT40.8k
Fed-BPTVLM614k
Our method150
+ +Table 3. Communication Overhead Experiment results. The measurements are in Tensors within the PyTorch Framework. Bold values indicate the best results, and underlined values indicate the second-best results. + +# 4.5. Ablation Studies + +In the improvements ablation study based on Qwen2-VL72B API, our method consistently achieves the best performance across all metrics, further validating the effectiveness of its key components, as shown in Table 4. Compared to cases where specific modules are removed or replaced, our method demonstrates significant improvements on the Cifar-10, Cifar-100, and TinyImg datasets. Server-side experiments. Our method improves accuracy on the Cifar-10 and Cifar-100 datasets by $2.0\%$ and $1.2\%$ , respectively, compared to the setting without the clustering strategy. Furthermore, compared to the fixed clustering approach, our method achieves accuracy gains of $0.7\%$ on Cifar-10 and $2.5\%$ on Cifar-100, underscoring the importance of a dynamic clustering strategy. Client-side experiments. Removing the G-F mechanism results in a significant accuracy drop across all datasets. Furthermore, our method achieves a $1.0\%$ accuracy improvement on Cifar-100 and $1.6\%$ on TinyImg compared to the configuration without the E-EB strategy. Compared to the configuration without the PPR-CoT strategy, our method improves accuracy by $3.2\%$ on Cifar-10 and $2.5\%$ on Cifar-100, underscoring the pivotal role of the PPR-CoT strategy in enhancing model perfor + +mance. + +
SiteMethodCifar-10Cifar-100TinyImg
Serverw Embedding-fedavg71.962.554.2
w Random search65.357.353.9
w/o Clustering Strategy76.570.655.9
w Fixed Clusters77.869.356.3
Clientw/o G-F Mechanism71.865.553.5
w/o E-EB Strategy76.970.856.2
w/o PPR-CoT75.369.356.7
Our method78.571.857.8
+ +Table 4. The Improvements Ablation Studies Results. The accuracy is measured in percentages (\%). Performance decreases with the removal or modification of any key component, underscoring the effectiveness of the improvements. "w Embedding-fedavg" denotes the use of a weighted averaging aggregation method, as in FedBPT [31] and FedAvg-BBT. "w Random search" refers to the removal of weight comparison, using random search instead to select representative tokens. "w/o Clustering Strategy" indicates the removal of the clustering step. "w Fixed Clusters" implies a fixed number of clusters, which affects the count of representative tokens. + +Initial Prompt Impact. We evaluate the impact of the Initial Prompt on performance using GPT-4o as the backbone black-box VLM for classification tasks across datasets, comparing it with three typical black-box algorithms. Figure 2 shows that different prompt configurations significantly affect model performance, with our algorithm consistently achieving the best outcomes. Specifically, detailed prompts enhance model accuracy by providing richer information, which supports more effective learning and adaptation. In contrast, shorter prompts reduce accuracy, indicating that limited information constrains performance. Thus, prompt design is critical in classification tasks, and our algorithm exhibits strong robustness and performance across various prompt configurations. We present all Initial Prompts in Appendix C. + +![](images/bdf692c71c5accc0fa2d0073f9c851bbd72af7066710b1a942b3d1ad6d0a7bcb.jpg) +Figure 2. Initial Prompt Impact. Prompt-1 represents a "shorter initial prompt," Prompt-2 a "standard initial prompt," and Prompt-3 a "detailed initial prompt." + +# 4.6. Heterogeneous Data Robustness Experiment + +In the heterogeneous data robustness experiment, we use the Qwen2-VL-72B API as the backbone VLM model to evaluate the performance of different algorithms on heterogeneous data in the TinyImagenet dataset, as shown in Figure 3. Data is distributed across 10 clients using a Dirichlet Distribution [21] with varying hyperparameters dir, where a smaller Dirichlet parameter (e.g., dir=0.05) indicates greater data heterogeneity. Details of these heterogeneous data distributions are provided in Appendix D. The results demonstrate that all methods show some robustness to data heterogeneity, aligning with expectations [28, 31] for model stability on heterogeneous data. Additionally, as heterogeneity increases, our method exhibits superior performance, with minimal accuracy reduction, outperforming other methods under all heterogeneity conditions. + +![](images/da5f1c208403cb7a251f64a0194070670d99b45fd8172029e77c6a9e485be995.jpg) +Figure 3. Heterogeneous Data Robustness Experiment results on the TinyImageNet dataset. The parameter "dir" is the Dirichlet Distribution hyperparameter; smaller values of dir indicate higher data heterogeneity across the 10 clients. + +# 5. Conclusion + +We propose FDPT, the first federated black-box discrete prompt tuning framework based on visual language models. FDPT uses an Agent-based Client Prompt Optimization strategy to achieve token-level prompt optimization on the client side without traditional prediction logits, thereby preventing inversion attacks. On the server side, a Semantic-Similarity-Guided Evolutionary Computation framework is implemented to execute more precise zeroth-order optimization, which enables the obtention of representative global tokens. Moreover, discrete token exchange between clients offers high security and low communication overhead. Moreover, discrete tokens exhibit enhanced interpretability compared to soft prompts, which facilitates human experts in designing corresponding instructions. Our framework exhibits: (1) statistically significant accuracy enhancements in conventional visual recognition and cross-modal reasoning tasks; (2) optimized communication efficiency through adaptive token selection mechanisms; (3) cross-domain transferability of learned prompt representations; and (4) strengthened resilience to data distribution shifts, particularly in heterogeneous learning environments. + +Acknowledgements. This research was supported by the National Natural Science Foundation of China under Grant 52074305, 52274160, 51874300 and the National Natural Science Foundation of China-Shanxi Provincial People's Government Coal-based Low Carbon Joint Fund U1510115, the Academic Research Projects of Beijing Union University (No.ZK90202106) and R&D Program of Beijing Municipal Education Commission (KM202211417005). + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 2 +[2] Owen Astrachan. Bubble sort: an archaeological algorithmic analysis. ACM Sigcse Bulletin, 35(1):1-5, 2003. 5 +[3] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 2 +[4] Tianshi Che, Ji Liu, Yang Zhou, Jiaxiang Ren, Jiwen Zhou, Victor S Sheng, Huaiyu Dai, and Dejing Dou. Federated learning of large language models with parameter-efficient prompt tuning and adaptive optimization. arXiv preprint arXiv:2310.15080, 2023. 5, 6 +[5] Simin Chen, Zihe Song, Mirazul Haque, Cong Liu, and Wei Yang. Nicsslowdown: Evaluating the efficiency robustness of neural image caption generation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15365-15374, 2022. 2 +[6] Simin Chen, Hanlin Chen, Mirazul Haque, Cong Liu, and Wei Yang. The dark side of dynamic routing neural networks: Towards efficiency backdoor injection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24585-24594, 2023. 2 +[7] Yiyi Chen, Heather Lent, and Johannes Bjerva. Text embedding inversion security for multilingual language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7808-7827, 2024. 1 +[8] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024. 2 +[9] Li Deng. The mnist database of handwritten digit images for machine learning research [best of the web]. IEEE signal processing magazine, 29(6):141-142, 2012. 5 +[10] Shizhe Diao, Zhichao Huang, Ruijia Xu, Xuechun Li, Yong Lin, Xiao Zhou, and Tong Zhang. Black-box prompt learning for pre-trained language models. Trans. Mach. Learn. Res., 2023, 2023. 5 + +[11] Xiuwen Fang and Mang Ye. Robust federated learning with noisy and heterogeneous clients. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10072-10081, 2022. 2 +[12] Cong Gao, Wenfeng Li, Lijun He, and Lingchong Zhong. A distance and cosine similarity-based fitness evaluation mechanism for large-scale many-objective optimization. Engineering Applications of Artificial Intelligence, 133:108127, 2024. 5 +[13] Aditya Golatkar, Alessandro Achille, Yu-Xiang Wang, Aaron Roth, Michael Kearns, and Stefano Soatto. Mixed differential privacy in computer vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8376-8386, 2022. 1 +[14] Xiang Gu, Shuchao Pang, Anan Du, Yifei Wang, Jixiang Miao, and Jorge Díez. Dynamic multimodal prompt tuning: Boost few-shot learning with vlm-guided point cloud models. In ECAI, pages 761-768. IOS Press, 2024. 3 +[15] Tao Guo, Song Guo, Junxiao Wang, Xueyang Tang, and Wenchao Xu. Promptfl: Let federated participants cooperatively learn prompts instead of models-federated learning in age of foundation model. IEEE Transactions on Mobile Computing, 2023. 1 +[16] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli-hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 2 +[17] Xinyuan Ji, Zhaowei Zhu, Wei Xi, Olga Gadyatskaya, Zi long Song, Yong Cai, and Yang Liu. Fedfixer: Mitigating heterogeneous label noise in federated learning. In AAAI, pages 12830-12838. AAAI Press, 2024. 2 +[18] Ruochen Jiao, Shaoyuan Xie, Justin Yue, Takami Sato, Lixu Wang, Yixuan Wang, Qi Alfred Chen, and Qi Zhu. Exploring backdoor attacks against large language model-based decision making. arXiv preprint arXiv:2405.20774, 2024. 1 +[19] Weirui Kuang, Bingchen Qian, Zitao Li, Daoyuan Chen, Dawei Gao, Xuchen Pan, Yuexiang Xie, Yaliang Li, Bolin Ding, and Jingren Zhou. Federatedscope-llm: A comprehensive package for fine-tuning large language models in federated learning. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 5260–5271, 2024. 1 +[20] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3, 2015. 2, 5 +[21] Qinbin Li, Yiqun Diao, Quan Chen, and Bingsheng He. Federated learning on non-iid data silos: An experimental study. In 2022 IEEE 38th international conference on data engineering (ICDE), pages 965-978. IEEE, 2022. 8 +[22] Tian Li, Anit Kumar Sahu, Ameet Talwalkar, and Virginia Smith. Federated learning: Challenges, methods, and future directions. IEEE signal processing magazine, 37(3):50-60, 2020. 1, 2, 3, 4 +[23] Xiang Li, Yiqun Yao, Xin Jiang, Xuezhi Fang, Chao Wang, Xinzhang Liu, Zihan Wang, Yu Zhao, Xin Wang, Yuyao Huang, Shuangyong Song, Yongxiang Li, Zheng Zhang, Bo Zhao, Aixin Sun, Yequan Wang, Zhongjiang He, Zhongyuan Wang, Xuelong Li, and Tiejun Huang. Tele-flm technical report, 2024. 5 + +[24] Zexin Li, Bangjie Yin, Taiping Yao, Junfeng Guo, Shouhong Ding, Simin Chen, and Cong Liu. Sibling-attack: Rethinking transferable adversarial attacks against face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24626–24637, 2023. 2 +[25] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 2 +[26] Mulei Ma, Chenyu Gong, Liekang Zeng, Yang Yang, and Liantao Wu. Flocoff: Data heterogeneity resilient federated learning with communication-efficient edge offloading. IEEE Journal on Selected Areas in Communications, 42(11): 3262-3277, 2024. 2 +[27] Yifei Ming and Yixuan Li. How does fine-tuning impact out-of-distribution detection for vision-language models? International Journal of Computer Vision, 132(2):596-609, 2024. 1 +[28] John Nguyen, Jianyu Wang, Kshitiz Malik, Maziar Sanjabi, and Michael Rabbat. Where to begin? on the impact of pre-training and initialization in federated learning. arXiv preprint arXiv:2206.15387, 2022. 8 +[29] Chen Qiu, Xingyu Li, Chaithanya Kumar Mummadi, Madan Ravi Ganesh, Zhenzhen Li, Lu Peng, and Wan-Yi Lin. Federated text-driven prompt generation for vision-language models. In The Twelfth International Conference on Learning Representations, 2024. 1, 5, 6, 7 +[30] Mathias Risse. Arguing for majority rule*. Journal of Political Philosophy, 12(1), 2004. 5 +[31] Jingwei Sun, Ziyue Xu, Hongxu Yin, Dong Yang, Daguang Xu, Yiran Chen, and Holger R Roth. Fedbpt: Efficient federated black-box prompt tuning for large language models. arXiv preprint arXiv:2310.01467, 2023. 1, 2, 5, 6, 7, 8 +[32] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 2 +[33] Lixu Wang, Shichao Xu, Xiao Wang, and Qi Zhu. Addressing class imbalance in federated learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10165-10173, 2021. 2 +[34] Lixu Wang, Shichao Xu, Ruiqi Xu, Xiao Wang, and Qi Zhu. Non-transferable learning: A new approach for model ownership verification and applicability authorization. In International Conference on Learning Representation (ICLR-2022), 2022. 1 +[35] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 2, 5 +[36] Tao Wang, Yushu Zhang, Shuren Qi, Ruoyu Zhao, Xia Zhihua, and Jian Weng. Security and privacy on generative data in aigc: A survey. ACM Computing Surveys, 2023. 1 +[37] Yubin Wang, Xinyang Jiang, De Cheng, Dongsheng Li, and Cairong Zhao. Learning hierarchical prompt with structured linguistic knowledge for vision-language models. In AAAI, pages 5749-5757. AAAI Press, 2024. 5 + +[38] Yubo Wang, Chaohu Liu, Yanqiu Qu, Haoyu Cao, Deqiang Jiang, and Linli Xu. Break the visual perception: Adversarial attacks targeting encoded visual tokens of large vision-language models. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 1072-1081, 2024. 1 +[39] Feijie Wu, Zitao Li, Yaliang Li, Bolin Ding, and Jing Gao. Fedbiot: Llm local fine-tuning in federated learning without full model. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 3345-3355, 2024. 1 +[40] Jiaqi Wu, Simin Chen, Yuzhe Yang, Yijiang Li, Shiyue Hou, Rui Jing, Zehua Wang, Wei Chen, and Zijian Tian. Feddtpt: federated discrete and transferable prompt tuning for black-box large language models. arXiv preprint arXiv:2411.00985, 2024. 2 +[41] Lang Yu, Qin Chen, Jiaju Lin, and Liang He. Black-box prompt tuning for vision-language model as a service. In IJCAI, pages 1686-1694, 2023. 1, 5 +[42] Bowen Yuan, Sisi You, and Bing-Kun Bao. Self-pt: Adaptive self-prompt tuning for low-resource visual question answering. In ACM Multimedia, pages 5089-5098. ACM, 2023. 3 +[43] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9556-9567, 2024. 2, 5 +[44] Jingjing Zhang, Shancheng Fang, Zhendong Mao, Zhiwei Zhang, and Yongdong Zhang. Fine-tuning with multi-modal entity prompts for news image captioning. In ACM Multimedia, pages 4365-4373. ACM, 2022. 3 +[45] Jingyi Zhang, Jiaxing Huang, Sheng Jin, and Shijian Lu. Vision-language models for vision tasks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1, 2 +[46] Haodong Zhao, Wei Du, Fangqi Li, Peixuan Li, and Gongshen Liu. Fedprompt: Communication-efficient and privacy-preserving prompt tuning in federated learning. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 1, 2, 5, 7 +[47] Xuejun Zhao, Wencan Zhang, Xiaokui Xiao, and Brian Lim. Exploiting explanations for model inversion attacks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 682-692, 2021. 1 +[48] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 5 +[49] Yu Zhu, Kang Li, Lequan Yu, and Pheng-Ann Heng. Memory-efficient prompt tuning for incremental histopathology classification. In AAAI, pages 7802-7810. AAAI Press, 2024. 2 \ No newline at end of file diff --git a/2025/FDPT_ Federated Discrete Prompt Tuning for Black-Box Visual-Language Models/images.zip b/2025/FDPT_ Federated Discrete Prompt Tuning for Black-Box Visual-Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..932ae7c16ab915a3c07e8547cb44a9e0e1fcfed3 --- /dev/null +++ b/2025/FDPT_ Federated Discrete Prompt Tuning for Black-Box Visual-Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8990294b7dd8868717761cf6465792a4389da01c15b456e9b5f97f6f69648548 +size 498192 diff --git a/2025/FDPT_ Federated Discrete Prompt Tuning for Black-Box Visual-Language Models/layout.json b/2025/FDPT_ Federated Discrete Prompt Tuning for Black-Box Visual-Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cd1e93bdef5550347c9008b2d0f55fba781159bc --- /dev/null +++ b/2025/FDPT_ Federated Discrete Prompt Tuning for Black-Box Visual-Language Models/layout.json @@ -0,0 +1,8686 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 77, + 102, + 534, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 102, + 534, + 138 + ], + "spans": [ + { + "bbox": [ + 77, + 102, + 534, + 138 + ], + "type": "text", + "content": "FDPT: Federated Discrete Prompt Tuning for Black-Box Visual-Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 160, + 542, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 160, + 542, + 190 + ], + "spans": [ + { + "bbox": [ + 67, + 160, + 542, + 190 + ], + "type": "text", + "content": "Jiaqi Wu" + }, + { + "bbox": [ + 67, + 160, + 542, + 190 + ], + "type": "inline_equation", + "content": "^{1,8,10\\dagger}" + }, + { + "bbox": [ + 67, + 160, + 542, + 190 + ], + "type": "text", + "content": ", Simin Chen" + }, + { + "bbox": [ + 67, + 160, + 542, + 190 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 67, + 160, + 542, + 190 + ], + "type": "text", + "content": ", Jing Tang" + }, + { + "bbox": [ + 67, + 160, + 542, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 67, + 160, + 542, + 190 + ], + "type": "text", + "content": ", Yuzhe YANG" + }, + { + "bbox": [ + 67, + 160, + 542, + 190 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 67, + 160, + 542, + 190 + ], + "type": "text", + "content": ", Yiming Chen" + }, + { + "bbox": [ + 67, + 160, + 542, + 190 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 67, + 160, + 542, + 190 + ], + "type": "text", + "content": ", Lixu Wang" + }, + { + "bbox": [ + 67, + 160, + 542, + 190 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 67, + 160, + 542, + 190 + ], + "type": "text", + "content": ", Song Lin" + }, + { + "bbox": [ + 67, + 160, + 542, + 190 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 67, + 160, + 542, + 190 + ], + "type": "text", + "content": ", Zehua Wang" + }, + { + "bbox": [ + 67, + 160, + 542, + 190 + ], + "type": "inline_equation", + "content": "^{8,9}" + }, + { + "bbox": [ + 67, + 160, + 542, + 190 + ], + "type": "text", + "content": ", Wei Chen" + }, + { + "bbox": [ + 67, + 160, + 542, + 190 + ], + "type": "inline_equation", + "content": "^{9*}" + }, + { + "bbox": [ + 67, + 160, + 542, + 190 + ], + "type": "text", + "content": ", and Zijian Tian" + }, + { + "bbox": [ + 67, + 160, + 542, + 190 + ], + "type": "inline_equation", + "content": "^{10}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 100, + 191, + 495, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 191, + 495, + 204 + ], + "spans": [ + { + "bbox": [ + 100, + 191, + 495, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 100, + 191, + 495, + 204 + ], + "type": "text", + "content": "Tsinghua University, " + }, + { + "bbox": [ + 100, + 191, + 495, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 100, + 191, + 495, + 204 + ], + "type": "text", + "content": "University of Texas at Dallas, " + }, + { + "bbox": [ + 100, + 191, + 495, + 204 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 100, + 191, + 495, + 204 + ], + "type": "text", + "content": "Huazhong University of Science and Technology," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 100, + 204, + 509, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 204, + 509, + 217 + ], + "spans": [ + { + "bbox": [ + 100, + 204, + 509, + 217 + ], + "type": "text", + "content": "4The Chinese University of Hong Kong, 5National University of Singapore, 6Nanyang Technological University," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 145, + 218, + 465, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 218, + 465, + 231 + ], + "spans": [ + { + "bbox": [ + 145, + 218, + 465, + 231 + ], + "type": "text", + "content": "7Beijing Baidu Netcom Science Technology Co., Ltd., 8University of British Columbia," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 100, + 232, + 491, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 232, + 491, + 245 + ], + "spans": [ + { + "bbox": [ + 100, + 232, + 491, + 245 + ], + "type": "inline_equation", + "content": "^{9}" + }, + { + "bbox": [ + 100, + 232, + 491, + 245 + ], + "type": "text", + "content": "China University of Mining and Technology, " + }, + { + "bbox": [ + 100, + 232, + 491, + 245 + ], + "type": "inline_equation", + "content": "^{10}" + }, + { + "bbox": [ + 100, + 232, + 491, + 245 + ], + "type": "text", + "content": "China University of Mining and Technology (Beijing)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 92, + 247, + 515, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 247, + 515, + 258 + ], + "spans": [ + { + "bbox": [ + 92, + 247, + 515, + 258 + ], + "type": "text", + "content": "wjq11346@student.ubc.ca, sxc180080@utdallas.edu, focusers@163.com, yuzheyang@link.cuhk.edu.cn," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 112, + 261, + 496, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 261, + 496, + 273 + ], + "spans": [ + { + "bbox": [ + 112, + 261, + 496, + 273 + ], + "type": "text", + "content": "yiming.chen@u.nus.edu, wanglixu4334@gmail.com, linsong04@baidu.com, zwang@ece.ubc.ca," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 214, + 275, + 392, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 275, + 392, + 286 + ], + "spans": [ + { + "bbox": [ + 214, + 275, + 392, + 286 + ], + "type": "text", + "content": "chenwdavior@163.com，Tianzj0726@126.com" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 151, + 315, + 200, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 315, + 200, + 327 + ], + "spans": [ + { + "bbox": [ + 151, + 315, + 200, + 327 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 54, + 339, + 297, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 339, + 297, + 687 + ], + "spans": [ + { + "bbox": [ + 54, + 339, + 297, + 687 + ], + "type": "text", + "content": "Vision-Language Models (VLMs) have driven significant advancements in multimodal AI. Fine-tuning these models with user data enhances adaptability but poses privacy risks. While federated learning (FL) mitigates user data privacy concerns, it fails to protect model properties. Existing methods relying on black-box VLM APIs often require access to prediction logits, making them vulnerable to inversion attacks. Additionally, optimizing tuning complexity and data transmission efficiency in federated VLM scenarios remains a challenge. To address these challenges, we propose FDPT—the first federated discrete prompt tuning method utilizing black-box VLMs. During client optimization stage, FDPT employs an agent-driven framework leveraging large language models (LLMs) with enhanced reasoning capacities to systematically optimize discrete prompt representations, and also utilizes feedback mechanisms and chain of thought to enhance prediction accuracy. Importantly, it performs optimization by relying not on the predicted logic vectors output by LLMs but on textual results, avoiding reverse attack risks. During global aggregation stage, We mimic human electoral activities by employing evolutionary computation methods underpinned by semantic similarity computation to implement enhanced zero-order optimization for acquiring representative global tokens, thereby achieving knowledge aggregation. FDPT significantly outperforms nine state-of-the-art methods in image classification and visual question-answering, reducing communication overhead while generating highly transferable optimized prompts. Additionally, it exhibits im" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 316, + 477, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 477, + 328 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 477, + 328 + ], + "type": "text", + "content": "proved robustness to data heterogeneity." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 354, + 395, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 354, + 395, + 367 + ], + "spans": [ + { + "bbox": [ + 314, + 354, + 395, + 367 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 376, + 555, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 376, + 555, + 508 + ], + "spans": [ + { + "bbox": [ + 312, + 376, + 555, + 508 + ], + "type": "text", + "content": "General-purpose Vision-Language Models (VLMs) [45] integrate vast visual and linguistic information, advancing various vision tasks [36]. To better adapt to user needs, these models can undergo fine-tuning or prompt-tuning [27], but such modifications introduce privacy risks [13, 18, 34, 38]. Federated Learning (FL) [22] addresses these concerns by decentralizing data storage, reducing single-point breaches, and exchanging model weights or embeddings instead of raw data. Various FL-based approaches [19, 29, 39] have been explored for tuning large models, effectively preserving user data privacy." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 509, + 556, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 556, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 556, + 700 + ], + "type": "text", + "content": "Despite the success of FL, protecting the privacy of VLMs themselves remains a critical yet largely unexplored challenge. Existing FL approaches use zero-order optimization to estimate gradients for prompt tuning on black-box VLMs [31, 41]. For example, FedBPT [31] employs prompt exchange and gradient-free optimization, reducing communication and memory costs by avoiding direct parameter access and backpropagation on client devices. However, several limitations remain. First, these methods remain vulnerable to security risks as they rely on prediction logits, exposing them to inversion attacks that can infer model parameters [7, 47]. Second, zero-order optimization techniques [31, 41] often suffer from limited accuracy and adaptability in prompt fine-tuning. Finally, soft prompt methods [15, 46] lack interpretability and incur high communication costs in federated settings." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "content": "To address these challenges, we propose FDPT, the first" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 133, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 133, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 133, + 703 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 66, + 693, + 133, + 703 + ], + "type": "text", + "content": " Equal contribution" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 68, + 703, + 143, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 703, + 143, + 712 + ], + "spans": [ + { + "bbox": [ + 68, + 703, + 143, + 712 + ], + "type": "text", + "content": "* Corresponding author" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2461" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 335 + ], + "type": "text", + "content": "federated discrete prompt optimization method comprising two stages: client-side optimization and global aggregation. In the client-side optimization phase, we introduce AI Agent-based Client Prompt Optimization, which refines token-level prompts using large language models (LLMs). We design a feedback mechanism and Chain-of-Thought measures to guide the agent in selecting optimal discrete tokens, overcoming the limitations of existing zero-order optimization methods in prompt tuning. Unlike traditional approaches that rely on prediction logits, our method uses only black-box VLM text outputs, reducing the risk of inversion attacks and enhancing model privacy protection. Afterwards, in the global aggregation phase, to reduce communication overhead in federated settings, we use discrete tokens for knowledge exchange instead of soft prompts. Inspired by human voting mechanisms, we propose a Self-Attention-Guided Evolutionary Computation method to identify representative tokens for constructing global prompts. Unlike the coarse sampling in traditional zero-order optimization, our approach clusters tokens in semantic space and then applies similarity calculations to refine the most accurate results, enhancing tuning efficiency." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 356, + 295, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 356, + 295, + 523 + ], + "spans": [ + { + "bbox": [ + 55, + 356, + 295, + 523 + ], + "type": "text", + "content": "Detailed experimental results demonstrate the advantages of our method. In Visual Question Answering (VQA) tasks, under the GPT-4 black-box setting, our method outperforms the classic black-box prompt optimization method FedBPT [31], achieving " + }, + { + "bbox": [ + 55, + 356, + 295, + 523 + ], + "type": "inline_equation", + "content": "11.1\\%" + }, + { + "bbox": [ + 55, + 356, + 295, + 523 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 356, + 295, + 523 + ], + "type": "inline_equation", + "content": "7.4\\%" + }, + { + "bbox": [ + 55, + 356, + 295, + 523 + ], + "type": "text", + "content": " improvements on the Art and Chemistry categories of the MMMU dataset [43], respectively. In traditional computer vision (CV) classification tasks, using the Qwen2-VL-72B model [3], our approach achieves " + }, + { + "bbox": [ + 55, + 356, + 295, + 523 + ], + "type": "inline_equation", + "content": "71.8\\%" + }, + { + "bbox": [ + 55, + 356, + 295, + 523 + ], + "type": "text", + "content": " accuracy on CIFAR-100, surpassing FedBPT by " + }, + { + "bbox": [ + 55, + 356, + 295, + 523 + ], + "type": "inline_equation", + "content": "5.3\\%" + }, + { + "bbox": [ + 55, + 356, + 295, + 523 + ], + "type": "text", + "content": ", demonstrating its superior performance. Additionally, FDPT significantly reduces communication costs without compromising optimization performance and enhances interpretability through discrete prompts. The contributions of this paper are as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 544, + 293, + 708 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 544, + 293, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 293, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 293, + 601 + ], + "type": "text", + "content": "- We propose FDPT, a first federated black-box discrete prompt optimization framework that aggregates knowledge from multiple parties to optimize prompts while ensuring the privacy of both users and easily overlooked VLMs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 604, + 293, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 293, + 660 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 293, + 660 + ], + "type": "text", + "content": "- We introduce an AI Agent-based Client Prompt Optimization for token-level prompts and a Self-attention-guided Evolutionary Computation method for selecting global tokens, offering better predictive performance than traditional zero-order optimization." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 663, + 293, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 663, + 293, + 708 + ], + "spans": [ + { + "bbox": [ + 55, + 663, + 293, + 708 + ], + "type": "text", + "content": "- Extensive experiments demonstrate that our method achieves superior performance across various text-vision tasks, exhibits low communication overhead, and is resilient to data heterogeneity." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 71, + 479, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 479, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 479, + 84 + ], + "type": "text", + "content": "2. Background & Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 91, + 553, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 553, + 342 + ], + "type": "text", + "content": "Visual Language Models. Visual-Language Models (VLMs) [5, 45] are AI systems designed to process and understand both visual and textual information, enabling them to perform tasks such as Visual Question Answering (VQA) [43] and traditional visual classification tasks [20]. Traditional VLMs, like Qwen-VL [3], effectively integrate visual features with textual data, excelling in both VQA and image classification tasks. Flamingo [1] leverages few-shot learning capabilities, allowing it to quickly adapt to new tasks and perform well across both VQA and classification. Gemini [32] focuses on reasoning and multimodal understanding, achieving strong results in complex VQA tasks while also performing well in image and text classification. LLaVA [25] benefits from large-scale pretraining, efficiently handling intricate visuals and long-form text, with exceptional performance in both VQA and classification tasks. Newer models like InternVL2 [8], Qwen2-VL [35], and GPT-4o [16] build on these foundations, enhancing multimodal reasoning, improving the fusion of visual and textual data, and offering broader adaptability across both VQA and classification tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 358, + 553, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 553, + 669 + ], + "type": "text", + "content": "Federated Learning. Federated Learning (FL) [22] is a decentralized approach allowing multiple clients to collaboratively train a shared model without sharing raw data, thus preserving privacy and defending against attacks [6, 24]. Recent works have further improved FL's robustness in handling data heterogeneity [26, 33], label noise [17], and personalization needs [11]. With the rise of large language models (LLMs), Federated Learning (FL) has gained new applications, effectively addressing the challenges of knowledge exchange and privacy protection between independent LLMs. FedPrompt [46] combines FL with prompt tuning, utilizing split aggregation to significantly reduce communication costs. FedBPT [31] presents a federated black-box prompt tuning framework that avoids direct access to model parameters, instead optimizing prompts using gradient-free methods. FedDTPT [40] realizes discrete prompt tuning for Large Language Models (LLMs). However, these methods still face issues such as security risks in large models and high communication overhead. We propose a Federated Discrete Prompt Tuning method, which, compared to existing approaches, performs zero-order optimization at the token level, mitigating the risk of adversarial users leveraging logits generated by the large model for backdoor attacks. Additionally, by utilizing discrete tokens for knowledge exchange, it significantly reduces communication overhead." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Prompt Tuning. Prompt tuning [49] is an emerging technique for adapting large pre-trained models to specific tasks" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2462" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 105, + 70, + 504, + 255 + ], + "blocks": [ + { + "bbox": [ + 105, + 70, + 504, + 255 + ], + "lines": [ + { + "bbox": [ + 105, + 70, + 504, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 70, + 504, + 255 + ], + "type": "image", + "image_path": "750758aadf814a2906dd59c770d1ad113ac648049b233a009a000c5306cbf8d5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 262, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 55, + 262, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 555, + 285 + ], + "type": "text", + "content": "Figure 1. Overview of FDPT. The proposed framework consists of the client-side local Agent-based Client Prompt Optimization and the server-side Semantic-Similarity-Guided Evolutionary Computation for global selection. The detailed process is described in Sec. 3.2." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 306, + 295, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 306, + 295, + 367 + ], + "spans": [ + { + "bbox": [ + 55, + 306, + 295, + 367 + ], + "type": "text", + "content": "with minimal additional training. In VLMs, prompt tuning [14] bridges visual and textual modalities, enabling tasks like image captioning [42, 44]. The Prompt tuning can be represented as finding an optimal prompt " + }, + { + "bbox": [ + 55, + 306, + 295, + 367 + ], + "type": "inline_equation", + "content": "\\mathbf{p}^*" + }, + { + "bbox": [ + 55, + 306, + 295, + 367 + ], + "type": "text", + "content": " to maximize the model's performance on the dataset " + }, + { + "bbox": [ + 55, + 306, + 295, + 367 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(\\mathbf{x},\\mathbf{y})\\}" + }, + { + "bbox": [ + 55, + 306, + 295, + 367 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 119, + 384, + 295, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 384, + 295, + 397 + ], + "spans": [ + { + "bbox": [ + 119, + 384, + 295, + 397 + ], + "type": "interline_equation", + "content": "\\mathbf {p} ^ {*} = \\arg \\max f _ {\\mathrm {V L M}} (\\mathbf {p}; \\mathcal {D}) \\tag {1}", + "image_path": "21d11fe42047988db36b834c31ba2862a5d66c63524b46fb80e7c20110aa9c15.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 415, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 56, + 415, + 296, + 714 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{VLM}}(\\mathbf{p};\\mathcal{D})" + }, + { + "bbox": [ + 56, + 415, + 296, + 714 + ], + "type": "text", + "content": " denotes the performance of the VLM using prompt " + }, + { + "bbox": [ + 56, + 415, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 56, + 415, + 296, + 714 + ], + "type": "text", + "content": " over the dataset " + }, + { + "bbox": [ + 56, + 415, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 56, + 415, + 296, + 714 + ], + "type": "text", + "content": ". Continual Prompt Tuning (soft prompt tuning) involves fine-tuning a small set of trainable parameters (soft prompts) while keeping the core model fixed. However, it suffers from poor interpretability, and in Federated Learning (FL), it incurs high communication overhead between clients. Discrete Prompt Tuning (DPT), in contrast, optimizes specific discrete tokens from a fixed vocabulary. A notable approach, RLPrompt, uses reinforcement learning (RL) to select the best tokens based on rewards. However, it has limited optimization efficiency due to the computationally intensive nature of RL. Regarding above limitations, the proposed Federated Discrete Prompt Tuning method aims to combine the strengths of discrete prompt tuning with the constraints and needs of Federated Learning. By leveraging an AI agent (based on large language models) with robust reasoning capabilities, the method can generate more accurate and efficient local discrete tokens. Furthermore, by incorporating semantic similarity calculations into an evolutionary computing framework, the method can generate more precise global tokens that are shared across clients. Additionally, this reduces communication costs between clients, as discrete tokens are much smaller in parameters compared to soft prompts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 304, + 370, + 317 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 304, + 370, + 317 + ], + "spans": [ + { + "bbox": [ + 314, + 304, + 370, + 317 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 324, + 439, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 324, + 439, + 335 + ], + "spans": [ + { + "bbox": [ + 314, + 324, + 439, + 335 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 342, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 437 + ], + "type": "text", + "content": "In this work, we consider a distributed scenario [22], where " + }, + { + "bbox": [ + 313, + 342, + 555, + 437 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 342, + 555, + 437 + ], + "type": "text", + "content": " clients each possess a private dataset " + }, + { + "bbox": [ + 313, + 342, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_k = \\{(\\mathcal{X}_k,\\mathcal{Y}_k)\\}" + }, + { + "bbox": [ + 313, + 342, + 555, + 437 + ], + "type": "text", + "content": " consisting of " + }, + { + "bbox": [ + 313, + 342, + 555, + 437 + ], + "type": "inline_equation", + "content": "n_k" + }, + { + "bbox": [ + 313, + 342, + 555, + 437 + ], + "type": "text", + "content": " samples " + }, + { + "bbox": [ + 313, + 342, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\{(x_i^k,y_i^k)\\}_{i\\in [n_k]}" + }, + { + "bbox": [ + 313, + 342, + 555, + 437 + ], + "type": "text", + "content": ". Each client can only access its own private data and does not wish to share it with others. This scenario is realistic, as hospitals and other organizations prioritize protecting sensitive data. Moreover, such distributed settings have been widely studied in existing research." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 437, + 555, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 485 + ], + "type": "text", + "content": "Then, our federated discrete prompt optimization can be formulated as an optimization problem that aims to find the optimal prompt instruction " + }, + { + "bbox": [ + 313, + 437, + 555, + 485 + ], + "type": "inline_equation", + "content": "p^*" + }, + { + "bbox": [ + 313, + 437, + 555, + 485 + ], + "type": "text", + "content": " that satisfies the following equation:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 491, + 553, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 553, + 533 + ], + "type": "interline_equation", + "content": "p ^ {*} = \\arg \\min _ {p} \\sum_ {k \\in [ K ]} \\frac {n ^ {k}}{\\sum_ {k \\in [ K ]} n ^ {k}} \\sum_ {i \\in [ n ^ {k} ]} \\mathcal {L} \\left(\\mathrm {V L M} \\left(p; x _ {i} ^ {k}\\right), y _ {i} ^ {k}\\right) \\tag {2}", + "image_path": "4cafb757712a3dbdbfe89ecdac98d85e09efc04971e7acfc30236733cdff4dab.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 534, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 534, + 555, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\cdot)" + }, + { + "bbox": [ + 313, + 534, + 555, + 616 + ], + "type": "text", + "content": " is the objective function, which could be the error rate, and " + }, + { + "bbox": [ + 313, + 534, + 555, + 616 + ], + "type": "inline_equation", + "content": "\\mathrm{VLM}(\\cdot)" + }, + { + "bbox": [ + 313, + 534, + 555, + 616 + ], + "type": "text", + "content": " represents the Vision-Language Model (VLM). The above equation can be interpreted as learning a globally optimal natural language prompt by minimizing the weighted average loss across all clients, where each client's contribution is proportional to its dataset size." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": "Notice that if clients were permitted to share their private data and had white-box access to the VLM, the equation above could be easily solved by approximating the objective function with a differentiable function, such as cross-entropy loss. In this case, standard backpropagation could be applied to the global dataset to search for the optimal prompt instruction. However, due to commercial concerns, modern VLMs are typically deployed as API ser" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2463" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": "vices, which are black-boxes to the clients. Additionally, each client is unwilling to share their data with others due to privacy concerns. Consequently, we must impose the following constraints:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 126, + 219, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 126, + 219, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 126, + 219, + 152 + ], + "type": "text", + "content": "Blackbox: " + }, + { + "bbox": [ + 56, + 126, + 219, + 152 + ], + "type": "inline_equation", + "content": "\\frac{\\partial\\mathcal{L}(\\mathrm{VLM}(p;x),y)}{\\partial p} = \\emptyset" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 154, + 295, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 154, + 295, + 179 + ], + "spans": [ + { + "bbox": [ + 56, + 154, + 295, + 179 + ], + "type": "text", + "content": "Privacy: " + }, + { + "bbox": [ + 56, + 154, + 295, + 179 + ], + "type": "inline_equation", + "content": "\\forall k,j\\in [K],k\\neq j," + }, + { + "bbox": [ + 56, + 154, + 295, + 179 + ], + "type": "inline_equation", + "content": "\\mathrm{Client}_k" + }, + { + "bbox": [ + 56, + 154, + 295, + 179 + ], + "type": "text", + "content": " cannot access " + }, + { + "bbox": [ + 56, + 154, + 295, + 179 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_j" + }, + { + "bbox": [ + 56, + 154, + 295, + 179 + ], + "type": "text", + "content": " (3)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 186, + 296, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 186, + 296, + 223 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 296, + 223 + ], + "type": "text", + "content": "The first constraint indicates that the gradient is not feasible, while the second constraint specifies that each client can only access its own data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 229, + 157, + 242 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 229, + 157, + 242 + ], + "spans": [ + { + "bbox": [ + 55, + 229, + 157, + 242 + ], + "type": "text", + "content": "3.2. Design Overview" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 247, + 296, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 296, + 486 + ], + "type": "text", + "content": "The algorithm overview is illustrated in Figure 1. Consistent with the traditional federated learning paradigm [22], each client communicates solely with the server using the prompt information, while keeping their data local. This ensures compliance with the data privacy constraints outlined in Eq.(3). Specifically, at the beginning, the server distributes the Global Prompt to each client. Then our approach follows these steps during each global communication round to search an optimal prompt: (1) Client Optimization: Each client utilizes an Agent-based Client Prompt Optimization method for personalized token-level prompt optimization. This method is preferred over backpropagation and ensures adherence to the black-box constraints in Eq.(3), which is more realistic in the context of VLM as a service; (2) Server Aggregation: Each client uploads its locally optimized prompt to the server, where knowledge aggregation is performed using the Semantic-similarity-guided Evolutionary Computation method. The subsequent sections provide a detailed explanation of each step in our approach." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 493, + 172, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 493, + 172, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 493, + 172, + 506 + ], + "type": "text", + "content": "3.3. Client Optimization" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": "Upon receiving the global discrete prompt " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": " from the server in the current round, each client iteratively updates it using an agent-based approach, as shown in Alg. 1. Initially, the agent queries the LLM to randomly replace some tokens in " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": ", generating a set of prompt mutants " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "p_1, p_2, \\ldots, p_M" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": " (Line 1). Each prompt is then combined with the client's local dataset and fed into the VLM to measure the accuracy, which serves as the prompt mutant's score. FDPT maintains a priority queue to manage these mutants, prioritizing those with higher accuracy (Lines 2-6). FDPT iteratively selects the highest-priority prompt from the queue as the seed, asks the LLM to mutate the important tokens of the seed, and updates the priority queue with the newly generated prompt mutants (Lines 7-13). Specifically, token importance is computed as follows: (1), FDPT compares each prompt mutant " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": " with the seed prompt " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": " token by token, constructing a difference matrix " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "M_i^j = 0" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": " if the" + } + ] + } + ], + "index": 7 + }, + { + "type": "code", + "bbox": [ + 314, + 87, + 555, + 342 + ], + "blocks": [ + { + "bbox": [ + 315, + 72, + 511, + 84 + ], + "lines": [ + { + "bbox": [ + 315, + 72, + 511, + 84 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 511, + 84 + ], + "type": "text", + "content": "Algorithm 1 Agent-Based Prompt Optimization" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "lines": [ + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "text", + "content": "Input: Initial discrete prompt " + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "text", + "content": ", Large Language Model (LLM), Vision-Language Model (VLM), client local dataset " + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "text", + "content": ", number of mutants " + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "text", + "content": " \nOutput: Optimized prompt " + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "inline_equation", + "content": "\\hat{p}" + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "text", + "content": " \n1: Initialization: " + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "inline_equation", + "content": "\\{p_1, p_2, \\ldots, p_M\\} =" + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "text", + "content": " Mutator( " + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "inline_equation", + "content": "LLM, p_0" + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "text", + "content": " ) " + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "text", + "content": " Mutate " + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "text", + "content": " by randomly replacing tokens \n2: " + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{Q} =" + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "text", + "content": " Priority Queue " + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "text", + "content": " Initialize an empty queue \n3: for each prompt " + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "inline_equation", + "content": "\\{p_0, p_1, \\ldots, p_M\\}" + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "text", + "content": " do \n4: " + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "inline_equation", + "content": "Acc_i =" + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "text", + "content": " Evaluate(VLM, " + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "text", + "content": ") " + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "text", + "content": " Evaluation \n5: Push " + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "inline_equation", + "content": "(\\mathcal{Q}, p_i, Acc_i)" + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "text", + "content": " Update the priority queue \n6: end for \n7: while stopping criterion not met do \n8: " + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "inline_equation", + "content": "p_{seed} =" + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "text", + "content": " Pop " + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "inline_equation", + "content": "(\\mathcal{Q}) \\triangleright" + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "text", + "content": " Get the highest-priority prompt \n9: Position = Importance Estimator " + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "text", + "content": " Select important tokens \n10: " + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "inline_equation", + "content": "p_{new} =" + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "text", + "content": " Mutator( " + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "inline_equation", + "content": "LLM, p_{seed}" + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "text", + "content": ", Position) \n11: " + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "inline_equation", + "content": "Acc_{new} =" + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "text", + "content": " Evaluate(VLM, " + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "inline_equation", + "content": "p_{new}" + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "text", + "content": ") " + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "text", + "content": " Evaluation \n12: Push " + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "inline_equation", + "content": "(\\mathcal{Q}, p_{new}, Acc_{new})" + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "text", + "content": " Update the queue \n13: end while \n14: Return highest-scoring prompt from " + }, + { + "bbox": [ + 314, + 87, + 555, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "code_body" + } + ], + "index": 9, + "sub_type": "algorithm" + }, + { + "bbox": [ + 313, + 363, + 555, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 555, + 460 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 363, + 555, + 460 + ], + "type": "text", + "content": "-th token of " + }, + { + "bbox": [ + 313, + 363, + 555, + 460 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 313, + 363, + 555, + 460 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 363, + 555, + 460 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 313, + 363, + 555, + 460 + ], + "type": "text", + "content": " are different, and " + }, + { + "bbox": [ + 313, + 363, + 555, + 460 + ], + "type": "inline_equation", + "content": "M_i^j = 1" + }, + { + "bbox": [ + 313, + 363, + 555, + 460 + ], + "type": "text", + "content": " otherwise. (2), FDPT computes the accuracy gains " + }, + { + "bbox": [ + 313, + 363, + 555, + 460 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 363, + 555, + 460 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 363, + 555, + 460 + ], + "type": "inline_equation", + "content": "V_i = \\mathrm{Acc}_i - \\mathrm{Acc}_0" + }, + { + "bbox": [ + 313, + 363, + 555, + 460 + ], + "type": "text", + "content": ". (3) The importance score of each token position is then given by " + }, + { + "bbox": [ + 313, + 363, + 555, + 460 + ], + "type": "inline_equation", + "content": "S = M^{-1}V" + }, + { + "bbox": [ + 313, + 363, + 555, + 460 + ], + "type": "text", + "content": ". This score estimation equation reflects the degree to which a token's presence or absence affects accuracy, weighted by its position in the prompt. Finally, the top " + }, + { + "bbox": [ + 313, + 363, + 555, + 460 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 363, + 555, + 460 + ], + "type": "text", + "content": " token positions are selected based on the importance score " + }, + { + "bbox": [ + 313, + 363, + 555, + 460 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 363, + 555, + 460 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 467, + 428, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 467, + 428, + 479 + ], + "spans": [ + { + "bbox": [ + 314, + 467, + 428, + 479 + ], + "type": "text", + "content": "3.4. Server Aggregation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 484, + 554, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 554, + 555 + ], + "type": "text", + "content": "After local client optimization, each client uploads its optimized discrete prompt to the server for knowledge exchange. The server then aggregates these local optimal prompts for the next round of optimization. Specifically, the server aggregation includes the following steps (as shown in Algorithm 2):" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 556, + 554, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 554, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 554, + 616 + ], + "type": "text", + "content": "Token-level Semantic Representation. In total client uploaded prompts, e.g. the initial population " + }, + { + "bbox": [ + 313, + 556, + 554, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{init}}" + }, + { + "bbox": [ + 313, + 556, + 554, + 616 + ], + "type": "text", + "content": ", for each token in a prompt, we use a pre-trained text encoder of the Roberta to get its token-level semantic representation, denoted as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 348, + 624, + 555, + 640 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 624, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 348, + 624, + 555, + 640 + ], + "type": "interline_equation", + "content": "\\mathcal {E} _ {i} = [ e _ {i} ^ {1}, e _ {i} ^ {2}, \\dots , e _ {i} ^ {K} ] e _ {i} ^ {j} = f _ {e m b e d} (t k _ {i} ^ {j}) \\qquad (4)", + "image_path": "4ad5db8d67f3dd13288dd7e5bf18293cd82c924f29d6dc359de4c8e8aab2650e.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 649, + 554, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 649, + 554, + 674 + ], + "spans": [ + { + "bbox": [ + 313, + 649, + 554, + 674 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 649, + 554, + 674 + ], + "type": "inline_equation", + "content": "f_{embed}(\\cdot)" + }, + { + "bbox": [ + 313, + 649, + 554, + 674 + ], + "type": "text", + "content": " is the pre-trained encoder, and " + }, + { + "bbox": [ + 313, + 649, + 554, + 674 + ], + "type": "inline_equation", + "content": "tk_i^j" + }, + { + "bbox": [ + 313, + 649, + 554, + 674 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 313, + 649, + 554, + 674 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 313, + 649, + 554, + 674 + ], + "type": "text", + "content": " token of the " + }, + { + "bbox": [ + 313, + 649, + 554, + 674 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 313, + 649, + 554, + 674 + ], + "type": "text", + "content": " prompt." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Reputation Test. We first evaluate the reputation of each token, defined as the semantic-similarity weight (denoted" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2464" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "inline_equation", + "content": "\\alpha_{g,h})" + }, + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "content": " that each token assigns to all other tokens. This is computed using a cosine-similarity mechanism [12] based on embeddings:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 88, + 117, + 295, + 146 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 117, + 295, + 146 + ], + "spans": [ + { + "bbox": [ + 88, + 117, + 295, + 146 + ], + "type": "interline_equation", + "content": "\\alpha_ {g} = \\sum_ {h} \\alpha_ {g, h} = \\sum_ {h} \\operatorname {s o f t m a x} \\left(\\frac {\\mathbf {e} _ {g} \\cdot \\mathbf {e} _ {h}}{\\sqrt {d}}\\right), \\tag {5}", + "image_path": "3ac684fa79436530a81b287c04219c4836218ec8c30209b4215a08873374f337.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 155, + 296, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 155, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 155, + 296, + 240 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 155, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_g" + }, + { + "bbox": [ + 55, + 155, + 296, + 240 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 155, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_h" + }, + { + "bbox": [ + 55, + 155, + 296, + 240 + ], + "type": "text", + "content": " represent the embedding of a token " + }, + { + "bbox": [ + 55, + 155, + 296, + 240 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 155, + 296, + 240 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 155, + 296, + 240 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 155, + 296, + 240 + ], + "type": "text", + "content": ", respectively, and " + }, + { + "bbox": [ + 55, + 155, + 296, + 240 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 155, + 296, + 240 + ], + "type": "text", + "content": " is the dimensionality of the embeddings. A higher semantic-similarity weight " + }, + { + "bbox": [ + 55, + 155, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\alpha_{g,h}" + }, + { + "bbox": [ + 55, + 155, + 296, + 240 + ], + "type": "text", + "content": " signifies a stronger influence over other tokens. We define the total semantic-similarity weight of token " + }, + { + "bbox": [ + 55, + 155, + 296, + 240 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 155, + 296, + 240 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 155, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\alpha_{g}" + }, + { + "bbox": [ + 55, + 155, + 296, + 240 + ], + "type": "text", + "content": ", which represents the token's semantic importance among all other tokens, referred to as its Reputation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 254, + 296, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 254, + 296, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 254, + 296, + 291 + ], + "type": "text", + "content": "Aggregation. Next, tokens are clustered based on their semantic similarity by grouping embeddings that are close to each other in the high-dimensional semantic space:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 299, + 296, + 323 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 299, + 296, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 299, + 296, + 323 + ], + "type": "interline_equation", + "content": "\\mathcal {C} = \\left\\{\\mathcal {C} _ {1}, \\mathcal {C} _ {2}, \\dots , \\mathcal {C} _ {Z} \\right\\}, \\mathcal {C} _ {z} = \\left\\{\\mathbf {e} _ {h} \\mid \\| \\mathbf {e} _ {h} - \\mathbf {e} _ {g} \\| < \\epsilon , \\forall h \\in \\mathcal {P} _ {\\text {i n i t}} \\right\\}, \\tag {6}", + "image_path": "060a4ddb909e26e40bc2688d6f7c3eddba8419696a2a5fa17a36bbb740255f58.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 323, + 296, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 296, + 372 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 296, + 372 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 323, + 296, + 372 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 323, + 296, + 372 + ], + "type": "text", + "content": " is a pre-defined distance threshold. " + }, + { + "bbox": [ + 55, + 323, + 296, + 372 + ], + "type": "inline_equation", + "content": "C_z" + }, + { + "bbox": [ + 55, + 323, + 296, + 372 + ], + "type": "text", + "content": " denotes a cluster, or Community. Embeddings with equivalent syntactic meanings are grouped into a cluster, from which a single representative is subsequently selected." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 386, + 296, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 386, + 296, + 446 + ], + "spans": [ + { + "bbox": [ + 55, + 386, + 296, + 446 + ], + "type": "text", + "content": "Election. Finally, within each cluster (Community), the representative token is determined by semantic-similarity weights (reputation). Tokens with higher weights have stronger semantic dependencies on other tokens and are more likely to be selected:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 457, + 295, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 457, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 134, + 457, + 295, + 475 + ], + "type": "interline_equation", + "content": "\\mathbf {r} _ {z} = \\underset {\\mathbf {e} _ {n} \\in \\mathcal {C} _ {z}} {\\arg \\max } \\alpha_ {n}, \\tag {7}", + "image_path": "68175fc0023d2939e9a0b0385ad7b922839c2401000bb359cf2fe3b0fb8c3dd6.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 483, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 296, + 579 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 483, + 296, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_z" + }, + { + "bbox": [ + 55, + 483, + 296, + 579 + ], + "type": "text", + "content": " denotes the representative token in Community " + }, + { + "bbox": [ + 55, + 483, + 296, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_z" + }, + { + "bbox": [ + 55, + 483, + 296, + 579 + ], + "type": "text", + "content": ". Importantly, communities containing only a single token are discarded, as these tokens contain only client-specific information, following the principle of \"majority rule\" [30] in Voting. The selected tokens are then arranged in their original order (using, for example, a bubble sort [2]) to form a representative prompt, or the \"Representatives\", denoted as " + }, + { + "bbox": [ + 55, + 483, + 296, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{new}} = \\{\\mathbf{r}_1,\\mathbf{r}_2,\\dots ,\\mathbf{r}_Z\\}" + }, + { + "bbox": [ + 55, + 483, + 296, + 579 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 590, + 127, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 127, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 127, + 601 + ], + "type": "text", + "content": "4. Evaluation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 609, + 159, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 159, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 159, + 623 + ], + "type": "text", + "content": "4.1. Evaluation Setup" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 626, + 296, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 626, + 296, + 663 + ], + "spans": [ + { + "bbox": [ + 55, + 626, + 296, + 663 + ], + "type": "text", + "content": "Backbone API. In our experiments, we selected four backbone models: Qwen2-VL-72B [35], InternVL2-Llama3-76B, TeleMM [23], and GPT4o." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Dataset and Evaluation. We conduct experiments on two tasks: traditional image classification and Visual Question Answer (VQA). For image classification, we use the" + } + ] + } + ], + "index": 12 + }, + { + "type": "code", + "bbox": [ + 313, + 99, + 555, + 357 + ], + "blocks": [ + { + "bbox": [ + 315, + 72, + 555, + 96 + ], + "lines": [ + { + "bbox": [ + 315, + 72, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 555, + 96 + ], + "type": "text", + "content": "Algorithm 2 Semantic-similarity-Guided Evolutionary Computation for Discrete Prompt Optimization" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 313, + 99, + 555, + 357 + ], + "lines": [ + { + "bbox": [ + 313, + 99, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 99, + 555, + 357 + ], + "type": "text", + "content": "Input: I clients with discrete prompt tokens " + }, + { + "bbox": [ + 313, + 99, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{P}_{\\mathrm{client}_i}\\}_{i = 1}^I" + }, + { + "bbox": [ + 313, + 99, + 555, + 357 + ], + "type": "text", + "content": " \nOutput: Global prompt tokens " + }, + { + "bbox": [ + 313, + 99, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{new}}" + }, + { + "bbox": [ + 313, + 99, + 555, + 357 + ], + "type": "text", + "content": " \n1: Initialization Initial population " + }, + { + "bbox": [ + 313, + 99, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{init}}" + }, + { + "bbox": [ + 313, + 99, + 555, + 357 + ], + "type": "text", + "content": " \n2: for each token " + }, + { + "bbox": [ + 313, + 99, + 555, + 357 + ], + "type": "inline_equation", + "content": "tk" + }, + { + "bbox": [ + 313, + 99, + 555, + 357 + ], + "type": "text", + "content": " in a prompt " + }, + { + "bbox": [ + 313, + 99, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\in \\mathcal{P}_{\\mathrm{init}}" + }, + { + "bbox": [ + 313, + 99, + 555, + 357 + ], + "type": "text", + "content": " do \n3: Achieve mapping according to Eq.4 \n4: end for \n5: for each token " + }, + { + "bbox": [ + 313, + 99, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_g" + }, + { + "bbox": [ + 313, + 99, + 555, + 357 + ], + "type": "text", + "content": " do " + }, + { + "bbox": [ + 313, + 99, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 313, + 99, + 555, + 357 + ], + "type": "text", + "content": " Reputation Test \n6: for each token " + }, + { + "bbox": [ + 313, + 99, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_h" + }, + { + "bbox": [ + 313, + 99, + 555, + 357 + ], + "type": "text", + "content": " do \n7: Semantic-similarity weight according to Eq.5 \n8: end for \n9: Get total semantic-similarity weight " + }, + { + "bbox": [ + 313, + 99, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\alpha_{g} = \\sum_{h}\\alpha_{g,h}" + }, + { + "bbox": [ + 313, + 99, + 555, + 357 + ], + "type": "text", + "content": " \n10: end for \n11: Cluster tokens according to Eq.6 Aggregation \n12: for each Community " + }, + { + "bbox": [ + 313, + 99, + 555, + 357 + ], + "type": "inline_equation", + "content": "C_z\\in \\mathcal{C}" + }, + { + "bbox": [ + 313, + 99, + 555, + 357 + ], + "type": "text", + "content": " do Election \n13: Select representative token according to Eq.7 \n14: if " + }, + { + "bbox": [ + 313, + 99, + 555, + 357 + ], + "type": "inline_equation", + "content": "C_z" + }, + { + "bbox": [ + 313, + 99, + 555, + 357 + ], + "type": "text", + "content": " contains only one token then \n15: Discard Community " + }, + { + "bbox": [ + 313, + 99, + 555, + 357 + ], + "type": "inline_equation", + "content": "C_z" + }, + { + "bbox": [ + 313, + 99, + 555, + 357 + ], + "type": "text", + "content": " \n16: end if \n17: end for \n18: Form Representatives " + }, + { + "bbox": [ + 313, + 99, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{new}} = \\{\\mathbf{r}_1,\\mathbf{r}_2,\\dots ,\\mathbf{r}_Z\\}" + }, + { + "bbox": [ + 313, + 99, + 555, + 357 + ], + "type": "text", + "content": " from selected representative tokens" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "code_body" + } + ], + "index": 14, + "sub_type": "algorithm" + }, + { + "bbox": [ + 313, + 375, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 375, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 555, + 472 + ], + "type": "text", + "content": "MNIST [9], CIFAR-10, CIFAR-100, and TinyImageNet (TinyImg) [20] datasets. For VQA, we utilize the Art, Chemistry, and Finance sections of the MMMU dataset [43], along with the RealWorldQA (RWQA) dataset. The task objective is to leverage optimized prompts for accurate image classification on the image datasets, and for VQA, to assist in parsing textual content within the datasets to enable accurate classification or decision-making." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 472, + 555, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 555, + 521 + ], + "type": "text", + "content": "We use Accuracy as the evaluation metric for the image classification task. To maintain consistency across tasks, we also constrain predictions in the VQA task to the range of available labels and evaluate performance using Accuracy." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "content": "Comparison Baselines. We compare our method against eight state-of-the-art (SOTA) approaches: FedPrompt [46], FedPepTAO [4], FedTPG [29], HPT [37], CoOp [48], BBT [10], FedBPT [31], BPTVLM [41], and a manual prompting method (Manual Prompt). Notably, HPT [37], CoOp [48], BBT[10], and BPTVLM [41] are not originally designed for federated learning. To adapt these methods to a federated setting, we apply a weighted averaging aggregation approach, following the implementations of FedPrompt [46]and FedBPT [31]. The results are reported using federated versions of these models: FedAvg-HPT, FedAvg-CoOp, FedAvg-BBT, and Fed-BPTVLM. Additionally, we categorize these methods into White-box and Black-box approaches based on whether they require access to model parameters. The White-box methods include FedPrompt [46]," + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2465" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 69, + 545, + 465 + ], + "blocks": [ + { + "bbox": [ + 65, + 69, + 545, + 465 + ], + "lines": [ + { + "bbox": [ + 65, + 69, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 65, + 69, + 545, + 465 + ], + "type": "table", + "html": "
BackboneMethodCV ClassificationVQAAverage
MNISTCifar-10Cifar-100TinyImgArtChemistryFinanceRWQA
Qwen2-VL-72BWhite BoxFedPrompt90.479.573.560.974.663.370.580.4
FedPepTAO86.778.372.959.569.158.868.577.9
FedTPG84.276.871.359.272.457.562.675.4
FedAvg-HPT86.978.470.258.775.753.565.978.3
FedAvg-CoOp85.778.171.459.270.261.863.376.5
Black BoxManual70.762.253.547.354.745.851.858.3
FedAvg-BBT78.972.362.751.763.553.960.369.7
FedBPT80.072.064.253.168.050.659.067.8
Fed-BPTVLM80.773.162.350.965.253.363.370.8
Our method90.578.571.857.874.060.867.278.5
InternVL2-Llama3-76BWhite BoxFedPrompt88.181.374.460.270.258.862.976.1
FedPepTAO87.377.871.258.565.255.363.075.4
FedTPG86.279.371.757.468.054.260.773.8
FedAvg-HPT85.178.772.358.469.153.062.074.3
FedAvg-CoOp85.779.171.658.171.856.263.971.7
Black BoxManual71.863.751.845.452.546.353.861.3
FedAvg-BBT80.071.165.252.760.251.953.870.5
FedBPT80.772.766.652.361.350.658.467.1
Fed-BPTVLM79.571.467.253.064.153.056.769.8
Our method89.679.071.557.370.755.963.675.5
TeleMMBlack BoxManual69.465.553.345.853.045.052.557.9
FedAvg-BBT80.970.663.149.156.948.856.464.1
FedBPT80.071.361.950.960.851.557.466.3
Fed-BPTVLM79.171.062.749.759.152.357.768.5
Our method88.478.969.357.464.655.761.675.0
GPT4oBlack BoxManual74.468.259.152.861.958.661.671.9
FedAvg-BBT81.076.767.656.164.657.563.669.2
FedBPT83.278.368.456.669.159.368.569.9
Fed-BPTVLM81.779.067.957.367.457.965.271.6
Our method92.585.373.962.776.866.772.878.5
", + "image_path": "9aaa57d4b4beb88de6977a74ffcc74e8af97332a32ca82a86645ed5c709770d7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 472, + 555, + 495 + ], + "lines": [ + { + "bbox": [ + 55, + 472, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 555, + 495 + ], + "type": "text", + "content": "Table 1. Performance Comparison on CV Classification and VQA Tasks. The accuracy results are measured in percentages (\\%). Bold values indicate the overall best results, and underlined values indicate the best results among Black Box methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 516, + 295, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 295, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 295, + 564 + ], + "type": "text", + "content": "FedPepTAO [4], FedTPG [29], FedAvg-HPT, and FedAvg-CoOp, while the Black-box methods consist of Manual, FedAvg-BBT, FedBPT [31], and Fed-BPTVLM. Detailed descriptions of all baselines are provided in Appendix A." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "Implementation. The federated learning setup for our experiments is as follows: we use a default of 10 clients, with a client participation rate of " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": " in each training round. Additionally, we adopt a " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "-shot mechanism for prompt tuning, following the approach in work [31], with the default value of " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": " set to 50. For image classification datasets, we use the original training and test sets. For each Visual QA dataset, we randomly select 50 samples as the training set (from the validation set for MMMU and the test set for RealWorldQA), with the remaining samples used for testing. In terms of baseline implementations, we adapt traditional" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 516, + 555, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 516, + 555, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 516, + 555, + 636 + ], + "type": "text", + "content": "prompt optimization methods (HPT, CoOp, BBT, BBT-VLM) to the federated learning framework by employing the classical Federated Averaging (FedAvg) approach. For black-box backbone model configurations, we extend soft prompt tuning methodologies, such as FedBPT, by mapping the soft tokens to discrete tokens, through the RoBERTa encoder's tokenization mechanism, before feeding them into API-based models. In addition, all experiments are conducted on two NVIDIA GeForce RTX A800 GPUs. Detailed hyperparameter settings are provided in Appendix B." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 647, + 434, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 434, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 434, + 658 + ], + "type": "text", + "content": "4.2. Effectiveness Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": "We first evaluate the effectiveness of our method on traditional CV classification tasks and VQA tasks. Experimental results show that, compared to other methods, our method demonstrates significant performance improvements in both" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2466" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 97 + ], + "type": "text", + "content": "tasks, particularly excelling in black-box environments, with only a small gap from white-box algorithms." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 112, + 295, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 112, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 112, + 295, + 304 + ], + "type": "text", + "content": "Traditional CV Classification Tasks. In traditional CV classification tasks, our method achieves outstanding performance across all model configurations. For example, in the GPT4o black-box setting, our method achieves improvements of " + }, + { + "bbox": [ + 55, + 112, + 295, + 304 + ], + "type": "inline_equation", + "content": "11.2\\%" + }, + { + "bbox": [ + 55, + 112, + 295, + 304 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 112, + 295, + 304 + ], + "type": "inline_equation", + "content": "7.0\\%" + }, + { + "bbox": [ + 55, + 112, + 295, + 304 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 112, + 295, + 304 + ], + "type": "inline_equation", + "content": "5.5\\%" + }, + { + "bbox": [ + 55, + 112, + 295, + 304 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 112, + 295, + 304 + ], + "type": "inline_equation", + "content": "6.1\\%" + }, + { + "bbox": [ + 55, + 112, + 295, + 304 + ], + "type": "text", + "content": " over the closest-performing FedBPT [31] method on the MNIST, Cifar-10, Cifar-100, and TinyImg datasets, respectively. In the TeleMM black-box setting, our method also performs exceptionally well, achieving " + }, + { + "bbox": [ + 55, + 112, + 295, + 304 + ], + "type": "inline_equation", + "content": "88.4\\%" + }, + { + "bbox": [ + 55, + 112, + 295, + 304 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 112, + 295, + 304 + ], + "type": "inline_equation", + "content": "78.9\\%" + }, + { + "bbox": [ + 55, + 112, + 295, + 304 + ], + "type": "text", + "content": " accuracy on MNIST and Cifar-10, respectively, which is an improvement of " + }, + { + "bbox": [ + 55, + 112, + 295, + 304 + ], + "type": "inline_equation", + "content": "9.3\\%" + }, + { + "bbox": [ + 55, + 112, + 295, + 304 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 112, + 295, + 304 + ], + "type": "inline_equation", + "content": "8.3\\%" + }, + { + "bbox": [ + 55, + 112, + 295, + 304 + ], + "type": "text", + "content": " over FedAvg-BBT. Additionally, for the Qwen2-VL-72B model, our method achieves accuracies of " + }, + { + "bbox": [ + 55, + 112, + 295, + 304 + ], + "type": "inline_equation", + "content": "78.5\\%" + }, + { + "bbox": [ + 55, + 112, + 295, + 304 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 112, + 295, + 304 + ], + "type": "inline_equation", + "content": "71.8\\%" + }, + { + "bbox": [ + 55, + 112, + 295, + 304 + ], + "type": "text", + "content": " on the Cifar-10 and Cifar-100 datasets, respectively, representing improvements of " + }, + { + "bbox": [ + 55, + 112, + 295, + 304 + ], + "type": "inline_equation", + "content": "7.1\\%" + }, + { + "bbox": [ + 55, + 112, + 295, + 304 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 112, + 295, + 304 + ], + "type": "inline_equation", + "content": "5.3\\%" + }, + { + "bbox": [ + 55, + 112, + 295, + 304 + ], + "type": "text", + "content": " over FedBPT [31], further demonstrating its advantage in traditional CV tasks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 319, + 295, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 295, + 416 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 295, + 416 + ], + "type": "text", + "content": "VQA Tasks. Our method also shows strong performance in VQA tasks. In the GPT4o black-box setting, our method achieves improvements of " + }, + { + "bbox": [ + 55, + 319, + 295, + 416 + ], + "type": "inline_equation", + "content": "11.1\\%" + }, + { + "bbox": [ + 55, + 319, + 295, + 416 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 319, + 295, + 416 + ], + "type": "inline_equation", + "content": "7.4\\%" + }, + { + "bbox": [ + 55, + 319, + 295, + 416 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 319, + 295, + 416 + ], + "type": "inline_equation", + "content": "4.3\\%" + }, + { + "bbox": [ + 55, + 319, + 295, + 416 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 319, + 295, + 416 + ], + "type": "inline_equation", + "content": "8.6\\%" + }, + { + "bbox": [ + 55, + 319, + 295, + 416 + ], + "type": "text", + "content": " over the FedBPT [31] method on the Art, Chemistry, Finance, and RWQA datasets, respectively. In TeleMM setting, our method's performance remains very stable, achieving improvements of " + }, + { + "bbox": [ + 55, + 319, + 295, + 416 + ], + "type": "inline_equation", + "content": "7.7\\%" + }, + { + "bbox": [ + 55, + 319, + 295, + 416 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 319, + 295, + 416 + ], + "type": "inline_equation", + "content": "10.9\\%" + }, + { + "bbox": [ + 55, + 319, + 295, + 416 + ], + "type": "text", + "content": " over FedAvg-BBT on the Art and RWQA datasets, respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 423, + 241, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 241, + 436 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 241, + 436 + ], + "type": "text", + "content": "4.3. Prompt Transferability Evaluation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 441, + 295, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 295, + 561 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 295, + 561 + ], + "type": "text", + "content": "In the prompt transferability experiment, we evaluate the performance of prompts optimized on one backbone model when applied to another, as shown in Table 2. The results indicate that prompts exhibit a degree of transferability, maintaining high robustness across models with minimal accuracy loss, as referenced in Table 1. Additionally, compared to the Qwen2-VL transfer results, the InternVL2 transfer showed better accuracy, further validating the high applicability and robustness of prompts between certain models." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 58, + 568, + 293, + 606 + ], + "blocks": [ + { + "bbox": [ + 58, + 568, + 293, + 606 + ], + "lines": [ + { + "bbox": [ + 58, + 568, + 293, + 606 + ], + "spans": [ + { + "bbox": [ + 58, + 568, + 293, + 606 + ], + "type": "table", + "html": "
ModelCifar-10Cifar-100TinyImg
Qwen2-VL→TeleMM8070.558.1
InternVL2→GPT4o87.872.363.5
", + "image_path": "329c6aa54a41fb62a1627c40a1023d44ac408383efbb017c4bba9319fb55c93f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 659, + 261, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 261, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 261, + 673 + ], + "type": "text", + "content": "4.4. Communication Overhead Experiment" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "In the communication overhead experiment, our method shows a significant advantage in parameter efficiency, as illustrated in Table 3. Specifically, white-box methods like" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": "FedPrompt [46] and FedTPG [29] and most black-box methods such as FedAvg-BBT and Fed-BPTVLM require hundreds of thousands of parameters, whereas our method requires only 150. This substantial reduction effectively minimizes latency and resource consumption associated with communication. This advantage primarily arises from a key difference in algorithm design: while other methods use soft prompts for knowledge exchange between clients, our method employs discrete prompt tokens. Soft prompts typically require a large number of continuous parameters, whereas discrete prompt tokens enable knowledge transfer with only a small set of parameters, significantly reducing communication overhead." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 317, + 236, + 552, + 355 + ], + "blocks": [ + { + "bbox": [ + 55, + 613, + 295, + 636 + ], + "lines": [ + { + "bbox": [ + 55, + 613, + 295, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 613, + 295, + 636 + ], + "type": "text", + "content": "Table 2. Transferability Experiment Results. The accuracy results are measured in percentages " + }, + { + "bbox": [ + 55, + 613, + 295, + 636 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 55, + 613, + 295, + 636 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 236, + 552, + 355 + ], + "lines": [ + { + "bbox": [ + 317, + 236, + 552, + 355 + ], + "spans": [ + { + "bbox": [ + 317, + 236, + 552, + 355 + ], + "type": "table", + "html": "
TypeMethodParameters
White BoxFedPrompt614k
FedPepTAO1796k
FedTPG614k
FedAvg-HPT614k
FedAvg-CoOp614k
Black BoxFedAvg-BBT614k
FedBPT40.8k
Fed-BPTVLM614k
Our method150
", + "image_path": "aa4c6eea7331c9d07ab9724ccd8d4dca825b99624935b534b17ae42878f44315.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 362, + 555, + 407 + ], + "lines": [ + { + "bbox": [ + 313, + 362, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 362, + 555, + 407 + ], + "type": "text", + "content": "Table 3. Communication Overhead Experiment results. The measurements are in Tensors within the PyTorch Framework. Bold values indicate the best results, and underlined values indicate the second-best results." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 432, + 415, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 415, + 444 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 415, + 444 + ], + "type": "text", + "content": "4.5. Ablation Studies" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "content": "In the improvements ablation study based on Qwen2-VL72B API, our method consistently achieves the best performance across all metrics, further validating the effectiveness of its key components, as shown in Table 4. Compared to cases where specific modules are removed or replaced, our method demonstrates significant improvements on the Cifar-10, Cifar-100, and TinyImg datasets. Server-side experiments. Our method improves accuracy on the Cifar-10 and Cifar-100 datasets by " + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "inline_equation", + "content": "2.0\\%" + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "content": ", respectively, compared to the setting without the clustering strategy. Furthermore, compared to the fixed clustering approach, our method achieves accuracy gains of " + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "content": " on Cifar-10 and " + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "content": " on Cifar-100, underscoring the importance of a dynamic clustering strategy. Client-side experiments. Removing the G-F mechanism results in a significant accuracy drop across all datasets. Furthermore, our method achieves a " + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "inline_equation", + "content": "1.0\\%" + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "content": " accuracy improvement on Cifar-100 and " + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "content": " on TinyImg compared to the configuration without the E-EB strategy. Compared to the configuration without the PPR-CoT strategy, our method improves accuracy by " + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "inline_equation", + "content": "3.2\\%" + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "content": " on Cifar-10 and " + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "content": " on Cifar-100, underscoring the pivotal role of the PPR-CoT strategy in enhancing model perfor" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2467" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 94, + 294, + 208 + ], + "blocks": [ + { + "bbox": [ + 56, + 73, + 87, + 83 + ], + "lines": [ + { + "bbox": [ + 56, + 73, + 87, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 87, + 83 + ], + "type": "text", + "content": "mance." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 94, + 294, + 208 + ], + "lines": [ + { + "bbox": [ + 58, + 94, + 294, + 208 + ], + "spans": [ + { + "bbox": [ + 58, + 94, + 294, + 208 + ], + "type": "table", + "html": "
SiteMethodCifar-10Cifar-100TinyImg
Serverw Embedding-fedavg71.962.554.2
w Random search65.357.353.9
w/o Clustering Strategy76.570.655.9
w Fixed Clusters77.869.356.3
Clientw/o G-F Mechanism71.865.553.5
w/o E-EB Strategy76.970.856.2
w/o PPR-CoT75.369.356.7
Our method78.571.857.8
", + "image_path": "014723ac8c30ea6bf4eb63108b72b1e7123449f98af58dfb737d99169f31fe74.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 215, + 296, + 336 + ], + "lines": [ + { + "bbox": [ + 55, + 215, + 296, + 336 + ], + "spans": [ + { + "bbox": [ + 55, + 215, + 296, + 336 + ], + "type": "text", + "content": "Table 4. The Improvements Ablation Studies Results. The accuracy is measured in percentages (\\%). Performance decreases with the removal or modification of any key component, underscoring the effectiveness of the improvements. \"w Embedding-fedavg\" denotes the use of a weighted averaging aggregation method, as in FedBPT [31] and FedAvg-BBT. \"w Random search\" refers to the removal of weight comparison, using random search instead to select representative tokens. \"w/o Clustering Strategy\" indicates the removal of the clustering step. \"w Fixed Clusters\" implies a fixed number of clusters, which affects the count of representative tokens." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 375, + 298, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 375, + 298, + 555 + ], + "spans": [ + { + "bbox": [ + 54, + 375, + 298, + 555 + ], + "type": "text", + "content": "Initial Prompt Impact. We evaluate the impact of the Initial Prompt on performance using GPT-4o as the backbone black-box VLM for classification tasks across datasets, comparing it with three typical black-box algorithms. Figure 2 shows that different prompt configurations significantly affect model performance, with our algorithm consistently achieving the best outcomes. Specifically, detailed prompts enhance model accuracy by providing richer information, which supports more effective learning and adaptation. In contrast, shorter prompts reduce accuracy, indicating that limited information constrains performance. Thus, prompt design is critical in classification tasks, and our algorithm exhibits strong robustness and performance across various prompt configurations. We present all Initial Prompts in Appendix C." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 75, + 578, + 272, + 666 + ], + "blocks": [ + { + "bbox": [ + 75, + 578, + 272, + 666 + ], + "lines": [ + { + "bbox": [ + 75, + 578, + 272, + 666 + ], + "spans": [ + { + "bbox": [ + 75, + 578, + 272, + 666 + ], + "type": "image", + "image_path": "bdf692c71c5accc0fa2d0073f9c851bbd72af7066710b1a942b3d1ad6d0a7bcb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 677, + 295, + 711 + ], + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 711 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 711 + ], + "type": "text", + "content": "Figure 2. Initial Prompt Impact. Prompt-1 represents a \"shorter initial prompt,\" Prompt-2 a \"standard initial prompt,\" and Prompt-3 a \"detailed initial prompt.\"" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 545, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 545, + 85 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 545, + 85 + ], + "type": "text", + "content": "4.6. Heterogeneous Data Robustness Experiment" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 91, + 555, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 270 + ], + "type": "text", + "content": "In the heterogeneous data robustness experiment, we use the Qwen2-VL-72B API as the backbone VLM model to evaluate the performance of different algorithms on heterogeneous data in the TinyImagenet dataset, as shown in Figure 3. Data is distributed across 10 clients using a Dirichlet Distribution [21] with varying hyperparameters dir, where a smaller Dirichlet parameter (e.g., dir=0.05) indicates greater data heterogeneity. Details of these heterogeneous data distributions are provided in Appendix D. The results demonstrate that all methods show some robustness to data heterogeneity, aligning with expectations [28, 31] for model stability on heterogeneous data. Additionally, as heterogeneity increases, our method exhibits superior performance, with minimal accuracy reduction, outperforming other methods under all heterogeneity conditions." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 338, + 281, + 529, + 361 + ], + "blocks": [ + { + "bbox": [ + 338, + 281, + 529, + 361 + ], + "lines": [ + { + "bbox": [ + 338, + 281, + 529, + 361 + ], + "spans": [ + { + "bbox": [ + 338, + 281, + 529, + 361 + ], + "type": "image", + "image_path": "da5f1c208403cb7a251f64a0194070670d99b45fd8172029e77c6a9e485be995.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 369, + 556, + 413 + ], + "lines": [ + { + "bbox": [ + 313, + 369, + 556, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 556, + 413 + ], + "type": "text", + "content": "Figure 3. Heterogeneous Data Robustness Experiment results on the TinyImageNet dataset. The parameter \"dir\" is the Dirichlet Distribution hyperparameter; smaller values of dir indicate higher data heterogeneity across the 10 clients." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 441, + 388, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 388, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 388, + 453 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 462, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 556, + 715 + ], + "type": "text", + "content": "We propose FDPT, the first federated black-box discrete prompt tuning framework based on visual language models. FDPT uses an Agent-based Client Prompt Optimization strategy to achieve token-level prompt optimization on the client side without traditional prediction logits, thereby preventing inversion attacks. On the server side, a Semantic-Similarity-Guided Evolutionary Computation framework is implemented to execute more precise zeroth-order optimization, which enables the obtention of representative global tokens. Moreover, discrete token exchange between clients offers high security and low communication overhead. Moreover, discrete tokens exhibit enhanced interpretability compared to soft prompts, which facilitates human experts in designing corresponding instructions. Our framework exhibits: (1) statistically significant accuracy enhancements in conventional visual recognition and cross-modal reasoning tasks; (2) optimized communication efficiency through adaptive token selection mechanisms; (3) cross-domain transferability of learned prompt representations; and (4) strengthened resilience to data distribution shifts, particularly in heterogeneous learning environments." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2468" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 180 + ], + "type": "text", + "content": "Acknowledgements. This research was supported by the National Natural Science Foundation of China under Grant 52074305, 52274160, 51874300 and the National Natural Science Foundation of China-Shanxi Provincial People's Government Coal-based Low Carbon Joint Fund U1510115, the Academic Research Projects of Beijing Union University (No.ZK90202106) and R&D Program of Beijing Municipal Education Commission (KM202211417005)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 190, + 115, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 190, + 115, + 204 + ], + "spans": [ + { + "bbox": [ + 56, + 190, + 115, + 204 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 210, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 210, + 294, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 210, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 61, + 210, + 294, + 275 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 278, + 295, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 278, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 62, + 278, + 295, + 300 + ], + "type": "text", + "content": "[2] Owen Astrachan. Bubble sort: an archaeological algorithmic analysis. ACM Sigcse Bulletin, 35(1):1-5, 2003. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 301, + 295, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 301, + 295, + 343 + ], + "spans": [ + { + "bbox": [ + 62, + 301, + 295, + 343 + ], + "type": "text", + "content": "[3] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 346, + 295, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 346, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 62, + 346, + 295, + 399 + ], + "type": "text", + "content": "[4] Tianshi Che, Ji Liu, Yang Zhou, Jiaxiang Ren, Jiwen Zhou, Victor S Sheng, Huaiyu Dai, and Dejing Dou. Federated learning of large language models with parameter-efficient prompt tuning and adaptive optimization. arXiv preprint arXiv:2310.15080, 2023. 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 402, + 295, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 402, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 402, + 295, + 456 + ], + "type": "text", + "content": "[5] Simin Chen, Zihe Song, Mirazul Haque, Cong Liu, and Wei Yang. Nicsslowdown: Evaluating the efficiency robustness of neural image caption generation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15365-15374, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 458, + 295, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 458, + 295, + 512 + ], + "spans": [ + { + "bbox": [ + 62, + 458, + 295, + 512 + ], + "type": "text", + "content": "[6] Simin Chen, Hanlin Chen, Mirazul Haque, Cong Liu, and Wei Yang. The dark side of dynamic routing neural networks: Towards efficiency backdoor injection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24585-24594, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 514, + 295, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 295, + 567 + ], + "type": "text", + "content": "[7] Yiyi Chen, Heather Lent, and Johannes Bjerva. Text embedding inversion security for multilingual language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7808-7827, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 570, + 295, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 570, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 62, + 570, + 295, + 634 + ], + "type": "text", + "content": "[8] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 636, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 636, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 62, + 636, + 295, + 668 + ], + "type": "text", + "content": "[9] Li Deng. The mnist database of handwritten digit images for machine learning research [best of the web]. IEEE signal processing magazine, 29(6):141-142, 2012. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 58, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 670, + 294, + 712 + ], + "type": "text", + "content": "[10] Shizhe Diao, Zhichao Huang, Ruijia Xu, Xuechun Li, Yong Lin, Xiao Zhou, and Tong Zhang. Black-box prompt learning for pre-trained language models. Trans. Mach. Learn. Res., 2023, 2023. 5" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "text", + "content": "[11] Xiuwen Fang and Mang Ye. Robust federated learning with noisy and heterogeneous clients. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10072-10081, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 118, + 555, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 555, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 555, + 171 + ], + "type": "text", + "content": "[12] Cong Gao, Wenfeng Li, Lijun He, and Lingchong Zhong. A distance and cosine similarity-based fitness evaluation mechanism for large-scale many-objective optimization. Engineering Applications of Artificial Intelligence, 133:108127, 2024. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 173, + 555, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 555, + 228 + ], + "type": "text", + "content": "[13] Aditya Golatkar, Alessandro Achille, Yu-Xiang Wang, Aaron Roth, Michael Kearns, and Stefano Soatto. Mixed differential privacy in computer vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8376-8386, 2022. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 228, + 554, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 228, + 554, + 271 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 554, + 271 + ], + "type": "text", + "content": "[14] Xiang Gu, Shuchao Pang, Anan Du, Yifei Wang, Jixiang Miao, and Jorge Díez. Dynamic multimodal prompt tuning: Boost few-shot learning with vlm-guided point cloud models. In ECAI, pages 761-768. IOS Press, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 273, + 554, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 273, + 554, + 327 + ], + "spans": [ + { + "bbox": [ + 317, + 273, + 554, + 327 + ], + "type": "text", + "content": "[15] Tao Guo, Song Guo, Junxiao Wang, Xueyang Tang, and Wenchao Xu. Promptfl: Let federated participants cooperatively learn prompts instead of models-federated learning in age of foundation model. IEEE Transactions on Mobile Computing, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 328, + 554, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 328, + 554, + 370 + ], + "spans": [ + { + "bbox": [ + 317, + 328, + 554, + 370 + ], + "type": "text", + "content": "[16] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli-hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 372, + 554, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 554, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 554, + 415 + ], + "type": "text", + "content": "[17] Xinyuan Ji, Zhaowei Zhu, Wei Xi, Olga Gadyatskaya, Zi long Song, Yong Cai, and Yang Liu. Fedfixer: Mitigating heterogeneous label noise in federated learning. In AAAI, pages 12830-12838. AAAI Press, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 416, + 554, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 554, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 554, + 459 + ], + "type": "text", + "content": "[18] Ruochen Jiao, Shaoyuan Xie, Justin Yue, Takami Sato, Lixu Wang, Yixuan Wang, Qi Alfred Chen, and Qi Zhu. Exploring backdoor attacks against large language model-based decision making. arXiv preprint arXiv:2405.20774, 2024. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 460, + 554, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 554, + 536 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 554, + 536 + ], + "type": "text", + "content": "[19] Weirui Kuang, Bingchen Qian, Zitao Li, Daoyuan Chen, Dawei Gao, Xuchen Pan, Yuexiang Xie, Yaliang Li, Bolin Ding, and Jingren Zhou. Federatedscope-llm: A comprehensive package for fine-tuning large language models in federated learning. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 5260–5271, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 537, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 558 + ], + "type": "text", + "content": "[20] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3, 2015. 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 559, + 554, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 554, + 603 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 554, + 603 + ], + "type": "text", + "content": "[21] Qinbin Li, Yiqun Diao, Quan Chen, and Bingsheng He. Federated learning on non-iid data silos: An experimental study. In 2022 IEEE 38th international conference on data engineering (ICDE), pages 965-978. IEEE, 2022. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 604, + 554, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 604, + 554, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 604, + 554, + 646 + ], + "type": "text", + "content": "[22] Tian Li, Anit Kumar Sahu, Ameet Talwalkar, and Virginia Smith. Federated learning: Challenges, methods, and future directions. IEEE signal processing magazine, 37(3):50-60, 2020. 1, 2, 3, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 554, + 712 + ], + "type": "text", + "content": "[23] Xiang Li, Yiqun Yao, Xin Jiang, Xuezhi Fang, Chao Wang, Xinzhang Liu, Zihan Wang, Yu Zhao, Xin Wang, Yuyao Huang, Shuangyong Song, Yongxiang Li, Zheng Zhang, Bo Zhao, Aixin Sun, Yequan Wang, Zhongjiang He, Zhongyuan Wang, Xuelong Li, and Tiejun Huang. Tele-flm technical report, 2024. 5" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2469" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 128 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 128 + ], + "type": "text", + "content": "[24] Zexin Li, Bangjie Yin, Taiping Yao, Junfeng Guo, Shouhong Ding, Simin Chen, and Cong Liu. Sibling-attack: Rethinking transferable adversarial attacks against face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24626–24637, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 152 + ], + "type": "text", + "content": "[25] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 206 + ], + "type": "text", + "content": "[26] Mulei Ma, Chenyu Gong, Liekang Zeng, Yang Yang, and Liantao Wu. Flocoff: Data heterogeneity resilient federated learning with communication-efficient edge offloading. IEEE Journal on Selected Areas in Communications, 42(11): 3262-3277, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "type": "text", + "content": "[27] Yifei Ming and Yixuan Li. How does fine-tuning impact out-of-distribution detection for vision-language models? International Journal of Computer Vision, 132(2):596-609, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 252, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 252, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 252, + 294, + 297 + ], + "type": "text", + "content": "[28] John Nguyen, Jianyu Wang, Kshitiz Malik, Maziar Sanjabi, and Michael Rabbat. Where to begin? on the impact of pre-training and initialization in federated learning. arXiv preprint arXiv:2206.15387, 2022. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 298, + 294, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 294, + 353 + ], + "type": "text", + "content": "[29] Chen Qiu, Xingyu Li, Chaithanya Kumar Mummadi, Madan Ravi Ganesh, Zhenzhen Li, Lu Peng, and Wan-Yi Lin. Federated text-driven prompt generation for vision-language models. In The Twelfth International Conference on Learning Representations, 2024. 1, 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 376 + ], + "type": "text", + "content": "[30] Mathias Risse. Arguing for majority rule*. Journal of Political Philosophy, 12(1), 2004. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 377, + 294, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 294, + 422 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 294, + 422 + ], + "type": "text", + "content": "[31] Jingwei Sun, Ziyue Xu, Hongxu Yin, Dong Yang, Daguang Xu, Yiran Chen, and Holger R Roth. Fedbpt: Efficient federated black-box prompt tuning for large language models. arXiv preprint arXiv:2310.01467, 2023. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 422, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 476 + ], + "type": "text", + "content": "[32] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 478, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 521 + ], + "type": "text", + "content": "[33] Lixu Wang, Shichao Xu, Xiao Wang, and Qi Zhu. Addressing class imbalance in federated learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10165-10173, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "type": "text", + "content": "[34] Lixu Wang, Shichao Xu, Ruiqi Xu, Xiao Wang, and Qi Zhu. Non-transferable learning: A new approach for model ownership verification and applicability authorization. In International Conference on Learning Representation (ICLR-2022), 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "type": "text", + "content": "[35] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 2, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 635, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 669 + ], + "type": "text", + "content": "[36] Tao Wang, Yushu Zhang, Shuren Qi, Ruoyu Zhao, Xia Zhihua, and Jian Weng. Security and privacy on generative data in aigc: A survey. ACM Computing Surveys, 2023. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 714 + ], + "type": "text", + "content": "[37] Yubin Wang, Xinyang Jiang, De Cheng, Dongsheng Li, and Cairong Zhao. Learning hierarchical prompt with structured linguistic knowledge for vision-language models. In AAAI, pages 5749-5757. AAAI Press, 2024. 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 696 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 139 + ], + "type": "text", + "content": "[38] Yubo Wang, Chaohu Liu, Yanqiu Qu, Haoyu Cao, Deqiang Jiang, and Linli Xu. Break the visual perception: Adversarial attacks targeting encoded visual tokens of large vision-language models. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 1072-1081, 2024. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "type": "text", + "content": "[39] Feijie Wu, Zitao Li, Yaliang Li, Bolin Ding, and Jing Gao. Fedbiot: Llm local fine-tuning in federated learning without full model. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 3345-3355, 2024. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 195, + 553, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 250 + ], + "type": "text", + "content": "[40] Jiaqi Wu, Simin Chen, Yuzhe Yang, Yijiang Li, Shiyue Hou, Rui Jing, Zehua Wang, Wei Chen, and Zijian Tian. Feddtpt: federated discrete and transferable prompt tuning for black-box large language models. arXiv preprint arXiv:2411.00985, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 251, + 553, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 284 + ], + "type": "text", + "content": "[41] Lang Yu, Qin Chen, Jiaju Lin, and Liang He. Black-box prompt tuning for vision-language model as a service. In IJCAI, pages 1686-1694, 2023. 1, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 285, + 553, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 285, + 553, + 328 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 553, + 328 + ], + "type": "text", + "content": "[42] Bowen Yuan, Sisi You, and Bing-Kun Bao. Self-pt: Adaptive self-prompt tuning for low-resource visual question answering. In ACM Multimedia, pages 5089-5098. ACM, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 330, + 553, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 330, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 316, + 330, + 553, + 407 + ], + "type": "text", + "content": "[43] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9556-9567, 2024. 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 407, + 553, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 407, + 553, + 452 + ], + "spans": [ + { + "bbox": [ + 316, + 407, + 553, + 452 + ], + "type": "text", + "content": "[44] Jingjing Zhang, Shancheng Fang, Zhendong Mao, Zhiwei Zhang, and Yongdong Zhang. Fine-tuning with multi-modal entity prompts for news image captioning. In ACM Multimedia, pages 4365-4373. ACM, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 453, + 553, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 453, + 553, + 496 + ], + "spans": [ + { + "bbox": [ + 316, + 453, + 553, + 496 + ], + "type": "text", + "content": "[45] Jingyi Zhang, Jiaxing Huang, Sheng Jin, and Shijian Lu. Vision-language models for vision tasks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 498, + 553, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 553, + 563 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 553, + 563 + ], + "type": "text", + "content": "[46] Haodong Zhao, Wei Du, Fangqi Li, Peixuan Li, and Gongshen Liu. Fedprompt: Communication-efficient and privacy-preserving prompt tuning in federated learning. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 1, 2, 5, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 563, + 553, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 563, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 316, + 563, + 553, + 609 + ], + "type": "text", + "content": "[47] Xuejun Zhao, Wencan Zhang, Xiaokui Xiao, and Brian Lim. Exploiting explanations for model inversion attacks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 682-692, 2021. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 609, + 553, + 653 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 609, + 553, + 653 + ], + "spans": [ + { + "bbox": [ + 316, + 609, + 553, + 653 + ], + "type": "text", + "content": "[48] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 654, + 553, + 696 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 654, + 553, + 696 + ], + "spans": [ + { + "bbox": [ + 316, + 654, + 553, + 696 + ], + "type": "text", + "content": "[49] Yu Zhu, Kang Li, Lequan Yu, and Pheng-Ann Heng. Memory-efficient prompt tuning for incremental histopathology classification. In AAAI, pages 7802-7810. AAAI Press, 2024. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2470" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/FE-CLIP_ Frequency Enhanced CLIP Model for Zero-Shot Anomaly Detection and Segmentation/e20fdad0-0c02-434f-a4f5-0e729b38aab0_content_list.json b/2025/FE-CLIP_ Frequency Enhanced CLIP Model for Zero-Shot Anomaly Detection and Segmentation/e20fdad0-0c02-434f-a4f5-0e729b38aab0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9d55cc493edce48bdad08fd7cf2047b5e1e58a9b --- /dev/null +++ b/2025/FE-CLIP_ Frequency Enhanced CLIP Model for Zero-Shot Anomaly Detection and Segmentation/e20fdad0-0c02-434f-a4f5-0e729b38aab0_content_list.json @@ -0,0 +1,1608 @@ +[ + { + "type": "text", + "text": "FE-CLIP: Frequency Enhanced CLIP Model for Zero-Shot Anomaly Detection and Segmentation", + "text_level": 1, + "bbox": [ + 233, + 130, + 759, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tao Gong $^{1,2,3}$ , Qi Chu $^{1,2,3*}$ , Bin Liu $^{1,2,3}$ , Wei Zhou $^{4}$ , Nenghai Yu $^{1,2,3}$", + "bbox": [ + 212, + 202, + 782, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1School of Cyber Science and Technology, University of Science and Technology of China", + "bbox": [ + 142, + 220, + 854, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2Anhui Province Key Laboratory of Digital Security", + "bbox": [ + 294, + 238, + 704, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3the CCCD Key Lab of Ministry of Culture and Tourism", + "bbox": [ + 276, + 256, + 720, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ Ling Yang Industrial Internet Co., Ltd.", + "bbox": [ + 341, + 273, + 653, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{tgong, qchu, flowice, ynh} @ustc.edu.cn, weizhou8@iflytek.com", + "bbox": [ + 238, + 292, + 751, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 344, + 326, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zero-shot anomaly detection (ZSAD) requires detection models trained using auxiliary data to detect anomalies without any training sample in a target dataset. It is challenging since the models need to generalize to anomalies across different domains. Recently, CLIP-based anomaly detection methods, such as WinCLIP and AnomalyCLIP, have demonstrated superior performance in the ZSAD task, due to the strong zero-shot recognition of the CLIP model. However, they overlook the utilization of frequency information of images. In this paper, we find that frequency information could benefit the ZSAD task, since some properties of the anomaly area, such as appearance defects, can also be reflected based on its frequency information. To this end, We propose Frequency Enhanced CLIP (FE-CLIP), taking advantage of two different but complementary frequency-aware clues, (1) Frequency-aware Feature Extraction adapter, and (2) Local Frequency Statistics adapter, in the visual encoder of CLIP, to deeply mine frequency information for the ZSAD task. We apply DCT as the frequency-domain transformation. Through comprehensive experiments, we show that the proposed FE-CLIP has good generalization across different domains and achieves superior zero-shot performance of detecting and segmenting anomalies in 10 datasets of highly diverse class semantics from various defect inspections and medical domains. Besides, FE-CLIP also achieves superior performance under the few-normal-shot anomaly detection settings.", + "bbox": [ + 91, + 377, + 482, + 785 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 816, + 220, + 832 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anomaly Detection aims to predict an image or a pixel as normal or anomalous, and has been widely applied in", + "bbox": [ + 89, + 843, + 482, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "various applications, such as industrial defect inspection [4, 5, 7, 16, 28, 35, 37, 47, 51] and medical image analysis [11, 27, 31, 42, 43]. Existing AD approaches typically assume that training examples in a target application domain are available for learning the detection models. However, this assumption may not hold in various scenarios, such as i) when accessing training data violates data privacy policies, or ii) when the target domain does not have relevant training data. Therefore, Zero-shot anomaly detection (ZSAD) is an emerging task for AD in such scenarios, to which the AD methods mentioned above are not viable, as it requires detection models to detect anomalies without any training sample in a target dataset. ZSAD is challenging since the models need to generalize to anomalies across different domains where the appearance of foreground objects, abnormal regions, and background features can vary significantly.", + "bbox": [ + 511, + 345, + 906, + 589 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, large pre-trained vision-language models (VLMs) have demonstrated strong zero-shot recognition ability in various vision tasks. Particularly, being pretrained using millions/billions of image-text pairs, CLIP [33] has been applied to empower various downstream tasks with its strong generalization capability. WinCLIP [17] is the first work to apply the CLIP model in the ZSAD line. It uses a large number of hand-crafted text prompts and extraction of window/patch/image-level features aligned with text to finish the ZSAD task. However, WinCLIP involves multiple forward passes of image patches for anomaly segmentation, which leads to inefficiency. Besides, WinCLIP only performs the testing of the CLIP model in the ZSAD without any fine-tuning using the anomaly detection task data, which also limits its performance in the ZSAD task. Subsequently, several works [8, 22, 32, 52] attempt to adapt the CLIP into the ZSAD task by fintuning. Specifically, AnomalyCLIP [52] learns object-agnostic text prompts that capture generic normality and abnormality in an image regardless of its foreground objects. ClipSAM [22] introduces", + "bbox": [ + 511, + 598, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author", + "bbox": [ + 109, + 887, + 232, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21220", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1c26cd97864a68e083005162dfc3547c7765e45f23d3c57b6e61b7a48aa42a10.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 90, + 189, + 162 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9d558e64152f6531e8db98707af6e0a4f628d3d575aee100ae81214939f2f93e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 194, + 90, + 284, + 162 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/481e6f3295a2142aeb98cd2466fb958c8eb1f96e00ab27f6e11da1e0a5400fcb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 90, + 377, + 162 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e0912137d64943e50914be6d383485a7da6123f6874c46937667ee4e5b732b41.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 380, + 90, + 470, + 162 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/624acac42e6ce9861fb5a31243faba7f9dca38d1893b808927d4a803f61fb3d6.jpg", + "image_caption": [ + "(a) normal" + ], + "image_footnote": [], + "bbox": [ + 101, + 162, + 189, + 233 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1d0eda1e2bd73fd2b08f254c32a05142fe642f32d8b33e44f39c0fe41c77365b.jpg", + "image_caption": [ + "(b) hole" + ], + "image_footnote": [], + "bbox": [ + 194, + 162, + 284, + 233 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/43a1f479838ef21d8c8f7fbcfdcdee60401235314c45aa66c30cbcd9d46cca12.jpg", + "image_caption": [ + "(c) crack", + "Figure 1. The local frequency statistics of images. The second row represents the local frequency statistics of non-overlapped image regions. There are four columns, standing for the normal object, object with a hole, object with a crack, and object with print, respectively. Please zoom in and pay attention to the red region (i.e. the anomaly region of the input image) for a better view." + ], + "image_footnote": [], + "bbox": [ + 287, + 162, + 377, + 233 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c6e47206040e53189505e927df62368744ab12c30d1fdec37db0117a4990442e.jpg", + "image_caption": [ + "(d) print" + ], + "image_footnote": [], + "bbox": [ + 380, + 162, + 470, + 233 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "a unified multiscale cross-modal interaction module for interacting language with visual features at multiple stages of the CLIP visual encoder to reason anomaly positions. VCP-CLIP [32] propose the Pre-VCP module and PostVCP module to perform visual context prompting to activate CLIP's anomalous semantic perception ability. AdaCLIP [8] incorporates static and dynamic learnable prompts to adapt CLIP for ZSAD. However, all of the methods mentioned above overlook the utilization of frequency information of images.", + "bbox": [ + 88, + 369, + 482, + 521 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we find that frequency information could benefit the ZSAD task. Some properties of the anomaly area, such as appearance defects, can also be reflected based on its frequency information, since high-frequency components represent fine details (e.g. soft borders) that often exist in the anomaly area. For example, as shown in Figure 1, the frequency statistics in the red region are different from the frequency statistics in the surrounding regions of the input abnormal image and the corresponding region of the normal image. These specific frequency patterns could be helpful for anomaly detection.", + "bbox": [ + 88, + 522, + 482, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we introduce a novel approach, namely the Frequency Enhanced CLIP (FE-CLIP), to inject frequency information into CLIP for accurate ZSAD across different domains. FE-CLIP takes advantage of two different but complementary frequency-aware clues, (1) Frequency-aware Feature Extraction (FFE) adapter, and (2) Local Frequency Statistics (LFS) adapter, in the visual encoder of CLIP, to deeply mine frequency information. Specifically, the FFE adapter uses DCT [1] to transform the image features into the frequency domain in order to mine the frequency information, then uses inverse DCT to transform back to the spatial domain. The FFE adapter describes the frequency-aware patterns in the spatial domain, but does not explicitly render the frequency information directly in", + "bbox": [ + 89, + 689, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the neural networks. Therefore, we use LFS to render the frequency information directly in the frequency domain. In each densely but regularly sampled local spatial patch, the statistics are gathered by counting the mean frequency responses. The averaged frequency statistics are a multichannel spatial map, where the number of channels is identical to the number of channels in the spatial domain. The local frequency statistics also follow the spatial layouts as the features of the CLIP visual encoder, thus also enjoying effective representation learning powered by the CLIP visual encoder. Meanwhile, since the frequency-aware feature extraction and local frequency statistics are complementary to each other but both of them share inherently similar frequency-aware semantics, thus they can be progressively fused during the feature learning process.", + "bbox": [ + 511, + 90, + 903, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Through comprehensive experiments, we show that the proposed FE-CLIP has good generalization across different domains and achieves superior zero-shot performance of detecting and segmenting anomalies in 10 datasets of highly diverse class semantics from various defect inspections and medical domains. Our contributions in this paper are summarized as follows:", + "bbox": [ + 511, + 319, + 903, + 422 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We find that frequency information could benefit the ZSAD task, and propose a novel Frequency Enhanced CLIP (FE-CLIP) method to inject frequency information into CLIP for accurate ZSAD across different domains.", + "2. We propose two different but complementary frequency-aware adapters, Frequency-aware Feature Extraction (FFE) adapter, and Local Frequency Statistics (LFS) adapter, in the visual encoder of CLIP, to deeply mine the frequency patterns of anomaly region.", + "3. Extensive experiments show that the proposed FE-CLIP has good generalization across different domains and achieves superior zero-shot performance of detecting and segmenting anomalies in 10 datasets of highly diverse class semantics from various defect inspections and medical domains. Besides, the proposed FE-CLIP also achieves superior performance under the few-normal-shot anomaly detection settings." + ], + "bbox": [ + 514, + 425, + 906, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 700, + 653, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Vision-language modeling", + "text_level": 1, + "bbox": [ + 511, + 726, + 748, + 742 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Among recent successes of large pre-trained vision-language models (VLM) [2, 20, 33], CLIP [33] is the first to perform pre-training on web-scale image-text data, showing unprecedented generality: e.g., its language-driven zero-shot inference, improved both effective robustness [40] and perceptual alignment [12]. Many following VLM works explored large-scale pre-training in different aspects, e.g., scaling up data [20], efficient designs [21], multi-tasks [29], etc. To democratize large-scale VLM for the usages in different domains, a billion-scale data LAION5B [39], a", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21221", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/aa8fb52490b307caa182158c166aa25e1db21b01264cdc8cc19722b5bec154fb.jpg", + "image_caption": [ + "Figure 2. The framework of the proposed FE-CLIP. The text encoder and visual encoder are from the CLIP model, and the FFE and LFS are the proposed adapters. We only show the class token embedding after the last block of the CLIP visual encoder for convenience. Actually, the CLIP visual encoder can output the class token after every block." + ], + "image_footnote": [], + "bbox": [ + 158, + 98, + 846, + 359 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "code base of OpenCLIP with pre-trained models are open-sourced. Other works presented CLIP's promise in zero/few-shot transfer to downstream tasks beyond classification [13, 36, 41, 46]. Good prompt engineering and tuning [33, 50] can non-trivially benefit generalization performances. Moreover, some other works [34, 48, 49] leverage the pre-trained CLIP for language-guided detection and segmentation with promising performances. In this paper, we mainly focus on adapting CLIP into the ZSAD task with superior performance by injecting the frequency information into the visual encoder of CLIP.", + "bbox": [ + 88, + 445, + 483, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Zero-shot Anomaly detection", + "text_level": 1, + "bbox": [ + 89, + 621, + 352, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ZSAD relies on the model's strong transferability to handle unseen anomalies [3]. A very recent approach WinCLIP [17] presents a seminal work that leverages CLIP [33] for zero-shot anomaly classification and segmentation. WinCLIP proposes a compositional ensemble on state words and prompt templates and efficient extraction and aggregation of window/patch/image-level features aligned with text to finish the ZSAD task. To tackle this inefficiency, APRILGAN [9] introduces a fine-tuning strategy by adding extra learnable linear projection layers in multiple stages of the CLIP visual encoder to enhance the modeling of local visual semantics. AnomalyCLIP [52] learns object-agnostic text prompts that capture generic normality and abnormality in an image regardless of its foreground objects. ClipSAM [22] introduces a unified multiscale cross-modal interaction module for interacting language with visual features at multiple stages of the CLIP visual encoder to reason anomaly", + "bbox": [ + 89, + 643, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "positions. Myriad [25] and AnomalyGPT [14] utilize large language models for zero-shot/few-shot anomaly detection. They employ an image decoder to provide fine-grained semantics and design a prompt learner to fine-tune the large language models using prompt embeddings. However, all of the methods mentioned above overlook the utilization of frequency information of images. In this paper, we find that frequency information could benefit the ZSAD task, and propose the FE-CLIP method which utilizes two different but complementary adapters, the FFE adapter and the LFS adapter, into the visual encoder of CLIP, to deeply mine the frequency information for ZSAD task.", + "bbox": [ + 511, + 445, + 908, + 627 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 641, + 607, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 511, + 666, + 643, + 683 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CLIP [33] is a large-scale pretraining method offering a joint vision-language representation. It consists of a text encoder $T(\\cdot)$ and a visual encoder $F(\\cdot)$ . Both encoders are mainstream multi-block networks such as ViT [10]. Using text prompts is a typical way to acquire the embeddings of different classes for zero-shot classification. The text prompts usually combine a text prompt template $\\Omega$ with the category name $c$ , where the text prompt template $\\Omega$ usually is defined as A photo of a [cls] and [cls] represents the target category name. The text prompts are passed through $T(\\cdot)$ to obtain its corresponding textual embedding $g_{c} \\in \\mathbb{R}^{D}$ where $D$ denotes the channel dimension. Then, an input image is passed through $F(\\cdot)$ to get the visual representations, where the class token $f \\in \\mathbb{R}^{D}$ is the", + "bbox": [ + 511, + 688, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21222", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "global visual embedding of the input image, and patch tokens $f^{m} \\in \\mathbb{R}^{H \\times W \\times D}$ are treated as local visual embeddings of the input image. $H$ and $W$ denote the height and width of the patch tokens, respectively. CLIP can perform zero-shot recognition by measuring the similarity between textual and visual embeddings. Specifically, given a target class set $\\mathcal{C}$ and an image, CLIP predicts the probability of the image belonging to the $c$ category as follows:", + "bbox": [ + 89, + 90, + 483, + 212 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np (y = c) = P \\left(g _ {c}, f\\right) = \\frac {\\exp \\left(< g _ {c} , f > / \\tau\\right)}{\\sum_ {c \\in \\mathcal {C}} \\exp \\left(< g _ {c} , f >\\right) / \\tau} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 222, + 482, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tau$ denotes the temperature and the operator $< ,\\cdot ,>$ represents the computation of cosine similarity. The computation can be extended from global visual embeddings to local visual embeddings to derive the corresponding segmentation maps $M_{c}\\in \\mathbb{R}^{H\\times W}$ , if each entry $(i,j)$ is computed as $P\\left(g_c,f^{m(i,j)}\\right)$", + "bbox": [ + 89, + 267, + 483, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Frequency Enhanced CLIP", + "text_level": 1, + "bbox": [ + 89, + 366, + 338, + 382 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this paper, we find that frequency information could benefit the ZSAD task, and propose a novel Frequency Enhanced CLIP (FE-CLIP) method to inject frequency information into CLIP [33] for accurate ZSAD across different domains. We inject the frequency information by the features of the CLIP visual encoder rather than the frequency of the input image, since the CLIP visual encoder is trained to take RGB image as input. Typically, the CLIP visual encoder comprises a series of block layers. From the bottom to the top of the layers, the visual encoder gradually learns the visual patterns at different levels of abstraction. As shown in Figure 2, we add Frequency-aware Feature Extraction (FFE) adapter and Local Frequency Statistics (LFS) adapter across multiple blocks levels in the visual encoder $F(\\cdot)$ of CLIP, while keeping its original backbone unchanged, thus enabling frequency information injection at multiple feature levels. To be specific, assuming the visual encoder consists of $N$ blocks ( $N = 4$ ), the features (i.e. the patch tokens) after the $n$ -th block are denoted as $f_{n}^{m} \\in \\mathbb{R}^{H \\times W \\times D}$ , and the class token after the $n$ -th block is denoted as $f_{n} \\in \\mathbb{R}^{D}$ , where $n = 1, 2,.., N$ . Firstly, the features $f_{n}^{m}$ are passed through an FFE adapter and an LFS adapter to get the features $f_{n,ffe}^{m}$ and $f_{n,lfs}^{m}$ , respectively, which contain the frequency information. We can get the frequency-aware feature $\\hat{f}_{n}^{m}$ as follows:", + "bbox": [ + 89, + 388, + 483, + 767 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {f} _ {n} ^ {m} = \\lambda (f _ {n, f f e} ^ {m} + f _ {n, l f s} ^ {m}) + (1 - \\lambda) f _ {n} ^ {m} \\qquad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 777, + 482, + 797 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with $\\hat{f}_n^m$ serving as the input for the next block. The class token $f_n$ will also be injected with frequency information after the next block due to the self-attention mechanism [44] in the blocks. We set $\\lambda = 0.1$ to preserve the original knowledge of the CLIP visual encoder. Then, the normal text prompt and abnormal text prompt are passed through", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ada198cb4923503e658aeebf6adf459bb11b39ce42a82b39c983defa58a9c7c5.jpg", + "image_caption": [ + "Figure 3. The proposed Frequency-aware Feature Extraction (FFE) adapter." + ], + "image_footnote": [], + "bbox": [ + 524, + 88, + 897, + 157 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "text encoder $T(\\cdot)$ to get corresponding textual embedding $g_{r}$ and $g_{a}$ , respectively. $g_{r}$ and $g_{a}$ constitute the target class set $\\mathcal{C}$ . We simply use A photo of a normal object and A photo of a damaged object for the normal text prompt and abnormal text prompt, respectively, rather than the complex compositional ensemble on state words and prompt templates used in WinCLIP [17], since we find that the simple text prompt also works well. Finally, we compute the abnormal segmentation map $M_{a,n} \\in \\mathbb{R}^{H \\times W}$ based on the equation 1, where each entry $(i,j)$ is computed as $P\\left(g_{a}, fc(\\hat{f}_{n}^{m(i,j)})\\right)$ . The $fc$ means that we use a single learnable fc to align the dimension of visual features $\\hat{f}_{n}^{m}$ to the dimension of text features. The abnormal score $S_{a,n}$ is also computed based on the equation 1 as $P\\left(g_{a}, fc_{clip}(f_{n})\\right)$ . The $fc_{clip}$ denotes the frozen project layer of the visual encoder in CLIP to align the dimension of class token $f_{n}$ and the dimension of text features. The final prediction of the abnormal score $S_{a}$ and abnormal segmentation map $M_{a}$ are computed by averaging $S_{a,n}$ and $M_{a,n}$ across $N$ blocks, respectively, as follows:", + "bbox": [ + 511, + 220, + 906, + 531 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {a} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} S _ {a, n}, \\quad M _ {a} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} M _ {a, n} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 541, + 906, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Frequency-aware Feature Extraction Adapter", + "text_level": 1, + "bbox": [ + 511, + 590, + 901, + 607 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The specific frequency patterns could be beneficial for anomaly detection. To mine the frequency information, we propose the Frequency-aware Feature Extraction (FFE) adapter in the visual encoder of CLIP. Specifically, given the patch tokens of the $n$ -th block $f_{n}^{m}\\in \\mathbb{R}^{H\\times W\\times D}$ , we first split the $f_{n}^{m}$ into non-overlapped windows with total $(H / P)\\times (W / P)$ windows and each window containing $P\\times P$ patch tokens ( $P$ is set to 3 by default). As shown in Figure 3, we apply DCT [1] transformation to each window to extract the DCT features $f_{n,ffe,1}^{m}$ which represents the frequency information. Then, we pass $f_{n,ffe,1}^{m}$ through one linear layer followed by one GELU( $\\cdot$ ) activation layer to further extract the frequency features $f_{n,ffe,2}^{m}$ . Finally, we apply the inverse DCT transformation to $f_{n,ffe,2}^{m}$ to get the frequency-aware features $f_{n,ffe}^{m}$ in the spatial domain. The FFE adapter describes the frequency-aware patterns in the spatial domain, but does not explicitly render the frequency information directly in the neural networks. In the next sub-section, we propose an adapter that renders the fre", + "bbox": [ + 511, + 613, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21223", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5e04a0e617d3013bf98a724038f9b30a721e7183e7309f2327e0716c29a434c0.jpg", + "image_caption": [ + "Figure 4. The proposed Local Frequency Statistics (LFS) adapter." + ], + "image_footnote": [], + "bbox": [ + 101, + 89, + 473, + 165 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "quency information directly in the neural networks.", + "bbox": [ + 89, + 231, + 429, + 247 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Local Frequency Statistics Adapter", + "text_level": 1, + "bbox": [ + 89, + 253, + 397, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The aforementioned FFE adapter has provided frequency-aware representation, but it has to represent the frequency-aware patterns back into the spatial domain, thus failing to directly utilize the frequency patterns in the frequency domain. Therefore, we propose the Local Frequency Statistics (LFS) adapter in the visual encoder of CLIP to render the frequency patterns directly in the frequency domain. As shown in Figure 4, given the patch tokens of the $n$ -th block $f_{n}^{m}\\in \\mathbb{R}^{H\\times W\\times D}$ , we first apply a Sliding Window DCT (SW DCT) on $f_{n}^{m}$ (i.e., taking DCT [1] densely on sliding windows of the features) to extract the localized frequency responses $f_{n,lfs,1}^{m}$ . The size of sliding windows is $Q\\times Q$ ( $Q$ is set to 3 by default). Therefore, there are $Q\\times Q$ group frequency responses in $f_{n,lfs,1}^{m}$ , where the first one represents the low-frequency responses and the last one represents the high-frequency responses. Then we count the mean of $f_{n,lfs,1}^{m}$ across $Q\\times Q$ groups to get the mean frequency responses $f_{n,lfs,2}^{m}$ . The $f_{n,lfs,2}^{m}$ is a multi-channel spatial map that shares the same layout as $f_{n}^{m}$ . The LFS adapter outputs $f_{n,lfs}^{m}$ after $f_{n,lfs,2}^{m}$ being passed through a convolutional layer followed by one $GELU(\\cdot)$ activation layer. The LFS adapter provides a localized aperture to detect detailed abnormal frequency distributions.", + "bbox": [ + 91, + 277, + 483, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Loss Functions", + "text_level": 1, + "bbox": [ + 89, + 632, + 243, + 647 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let $S_{gt}$ denote the ground-truth image label where 1 represents the abnormal image and 0 represents the normal image. The loss for the anomaly classification is defined as follows:", + "bbox": [ + 89, + 654, + 482, + 715 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {c l s} = - \\frac {1}{N} \\sum_ {N} ^ {n = 1} \\left[ S _ {g t} \\log \\left(S _ {a, n}\\right) + \\left(1 - S _ {g t}\\right) \\log \\left(1 - S _ {a, n}\\right) \\right] \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 723, + 480, + 775 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\log$ denotes the log operation. Similarly, let $M_{gt}$ denote the ground-truth pixel label of the input image, where 1 means abnormal pixel region and 0 means normal pixel region. The loss for the anomaly segmentation is defined as follows:", + "bbox": [ + 89, + 776, + 482, + 844 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {m a s k}} = \\frac {1}{N} \\sum_ {N} ^ {n = 1} [ \\operatorname {F o c a l} (U p (M _ {a, n}), M _ {g t}) + D i c e (U p (M _ {a, n}), M _ {g t}) ] \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 851, + 506, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $Focal(\\cdot, \\cdot)$ and $Dice(\\cdot, \\cdot)$ denote the focal loss [26] and dice loss [23], respectively. The operator $Up(\\cdot)$ represents the unsampling operation. During training, both the visual encoder and the text encoder of CLIP [33] are Frozen. Only the FFE adapters and LFS adapters are optimized by the loss function $L_{total} = L_{cls} + L_{mask}$ .", + "bbox": [ + 511, + 90, + 905, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 196, + 645, + 213 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 511, + 222, + 799, + 237 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To verify the effectiveness of the proposed method FE-CLIP, we conduct comprehensive experiments across 10 real-world anomaly detection datasets, covering various industrial inspection scenarios and medical imaging domains. In industrial inspection, we use MVtec AD [4], VisA [54], MPDD [18], BTAD [30], DAGM [45], and DTD-Synthetic [3]. In medical imaging, we use colon polyp detection datasets CVC-ClinicDB [6] and Kvasir [19], brain tumor detection datasets BrainMRI [38] and Br35H [15].", + "bbox": [ + 511, + 244, + 906, + 380 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As for the evaluation metrics, following previous works [17, 52], we use Area Under the Receiver Operating Characteristic curve (AUROC). Additionally, Average Precision (AP) for Zero-Shot Anomaly Detection (ZSAD) and PRO [5] for Zeor-Shot Anomaly Segmentation (ZSAS) are also used to provide a more in-depth analysis of the results.", + "bbox": [ + 511, + 381, + 905, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 482, + 730, + 498 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use the publicly available CLIP model (VITL/14@336px) as our backbone. Model parameters of CLIP are all frozen, and only the parameters of FFE adapters and LFS adapters are learnable. Following AnomalyCLIP [52], we fine-tune FE-CLIP using the test data of MVtec AD and evaluate the ZSAD performance on other datasets. As for MVtec AD, we fine-tune FE-CLIP on the test data of VisA. The proposed FE-CLIP is trained by 9 epochs with Adam optimizer. The learning rate is set to 5e-4 and the total batch size is 16. All experiments are conducted in PyTorch-1.13.0 with four NVIDIA RTX 3090 24GB GPUs. All experiment results are averaged under five runs. We report dataset-level results, which are averaged across their respective sub-datasets.", + "bbox": [ + 511, + 503, + 906, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Main Results", + "text_level": 1, + "bbox": [ + 511, + 726, + 651, + 741 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ZSAD performance on diverse anomaly scenarios. Table 1 shows the ZSAD results of FE-CLIP with other competing methods over six industrial defect datasets and two medical domain datasets. As shown in Table 1, FE-CLIP can achieve higher AUROC and AP for the ZSAD task on most industrial inspection datasets and medical datasets, such as MVTec AD, DTD-Synthetic, and BrainMRI, when compared with other competing methods, such as WinCLIP [17], AnomalyCLIP [52] and AdaCLIP [8]. This shows that (1) the frequency information can benefit the ZSAD task", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21224", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/88e54f6673c8da6d807a01d5faebda72997adb2739eb01cc3f9d34d01e05d850.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsIndustrial DefectsMedical Anomalies
MVTec ADVisAMPDDBTADDAGMDTD-SyntheticBrainMRIBr35H
CLIP [33]74.1 / 87.666.4 / 71.554.3 / 65.434.5 / 52.579.6 / 59.071.6 / 85.773.9 / 81.778.4 / 78.8
CoOp [50]88.8 / 94.862.8 / 68.155.1 / 64.266.8 / 77.487.5 / 74.6- / -61.3 / 44.986.0 / 87.5
WinCLIP [17]91.8 / 96.578.1 / 81.263.6 / 69.968.2 / 70.991.8 / 79.593.2 / 92.686.6 / 91.580.5 / 82.2
APRIL-GAN [9]86.1 / 93.578.0 / 81.478.0 / 81.473.6 / 68.694.4 / 83.886.4 / 95.089.3 / 90.993.1 / 92.9
AnomalyCLIP [52]91.5 / 96.282.1 / 85.477.0 / 82.088.3 / 87.397.5 / 92.393.5 / 97.090.3 / 92.294.6 / 94.7
AdaCLIP [8]89.2 / -85.8 / -76.0 / -88.6 / -99.1 / -95.5 / -94.8 / -97.7 / -
FE-CLIP91.9 / 96.584.6 / 86.678.0 / 82.690.3 / 90.097.5 / 92.398.3 / 99.494.8 / 93.896.8 / 93.8
", + "bbox": [ + 99, + 88, + 903, + 224 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c57f4830242ac8bbbdc5be1a89ca7c3e40937918fd9fef0a46665eef220eb416.jpg", + "table_caption": [ + "Table 1. AUROC / AP results for Zero-Shot Anomaly Detection (ZSAD) task on eight real-world anomaly detection datasets. Best results are highlighted in red." + ], + "table_footnote": [], + "table_body": "
MethodsIndustrial DefectsMedical Anomalies
MVTec ADVisAMPDDBTADDAGMDTD-SyntheticCVC-ClinicDBKvasir
CLIP [33]38.4 / 11.346.6 / 14.862.1, 33.030.6 / 4.430.6 / 4.433.9 / 12.547.5 / 18.944.6 / 17.7
CoOp [50]33.3 / 6.724.2 / 3.815.4 / 2.315.4 / 2.317.5 / 2.1- / -17.5 / 2.144.1 / 3.5
WinCLIP [17]85.1 / 64.679.6 / 56.879.6 / 56.872.7 / 27.387.6 / 65.783.9 / 57.851.2 / 13.869.7 / 24.5
APRIL-GAN [9]87.6 / 84.094.2 / 86.894.1 / 83.260.8 / 25.082.4 / 66.295.3 / 86.980.5 / 60.775.0 / 36.2
AnomalyCLIP [52]91.1 / 81.495.5 / 87.096.5 / 88.794.2 / 74.895.6 / 91.097.9 / 92.382.9 / 67.878.9 / 45.6
AdaCLIP [8]88.7 / -95.5 / -96.1 / -92.1 / -91.5 / -97.9 / -84.4 / -- / -
VCP-CLIP [32]92.0 / 87.395.7 / 90.7- / -94.1 / 74.699.4 / 98.3- / -- / -- / -
FE-CLIP92.6 / 88.395.9 / 92.897.0 / 90.595.6 / 80.498.5 / 96.699.0 / 97.484.5 / 69.279.8 / 60.6
", + "bbox": [ + 98, + 276, + 903, + 424 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. AUROC / PRO results for Zero-Shot Anomaly Segmentation (ZSAS) task on eight real-world anomaly detection datasets. Best results are highlighted in red. Note that the image-level medical AD datasets do not contain segmentation ground truth, so the pixel-level medical AD datasets are different from the image-level datasets.", + "bbox": [ + 89, + 431, + 906, + 476 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "due to the superior performance of FE-CLIP, and (2) the frequency information helps FE-CLIP achieve good generalization across datasets from different domains.", + "bbox": [ + 89, + 502, + 482, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ZSAS performance on diverse anomaly scenarios. We show the ZSAS results of FE-CLIP with other competing methods over six industrial defect datasets and two medical domain datasets in Table 2. Table 2 shows that FE-CLIP achieves higher AUROC and PRO for the ZSAS task on almost all of the industrial inspection datasets and medical datasets, except for the DAGM dataset, when compared with other competing methods, such as WinCLIP [17], AnomalyCLIP [52], AdaCLIP [8] and VCP-CLIP [32]. This further demonstrates that (1) the frequency information can benefit the ZSAS task and (2) the frequency information helps FE-CLIP achieve good generalization across datasets from different domains.", + "bbox": [ + 88, + 550, + 482, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Discussion. We also note that the FE-CLIP can defeat almost all of the competing methods on the ZSAS task, while FE-CLIP only performs better than other methods over most of the datasets on ZSAD task. The reason may be that the frequency information can be directly injected into the patch tokens (i.e. the feature map used for the ZSAS task) through the FFE adapter and LFS adapter, while the frequency information can only be implicitly involved in the global class token (i.e. the feature vector used for the ZSAD task) from the patch tokens based on the self", + "bbox": [ + 88, + 750, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "attention mechanism.", + "bbox": [ + 511, + 503, + 658, + 515 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Overall, the proposed FE-CLIP achieves superior zero-shot performance of detecting and segmenting anomalies in datasets of highly diverse class semantics from various defect inspection and medical imaging domains.", + "bbox": [ + 511, + 518, + 906, + 579 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 590, + 666, + 607 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The effectiveness of FFE and LFS We conduct ablation experiments on MVTec AD dataset and VisA dataset to validate the effectiveness of the proposed FFE adapter and LFS adapter. The results are shown in Table 3. The \"CLIP $[33] + \\text{Conv Adapter}\"$ means we add $N$ conv adapters, where each of them consists of one convolutional layer followed by one $GELU(\\cdot)$ activation layer to keep the same learnable parameters with the LFS adapter into the visual encoder of CLIP. It can be concluded that both the FFE adapter and LFS adapter can achieve better performance than the conv adapter in MVTec AD dataset and VisA dataset across the ZSAD task and ZSAS task. This also shows that the frequency information can benefit the ZSAD task and ZSAS task. When combined the FFE adapter with the LFS adapter, the FE-CLIP achieves higher performance than the FFE adapter and LFS adapter. This demonstrates that the FFE adapter and LFS adapter can mine different but complementary frequency information to further boost the performance of model on the ZSAD and ZSAS tasks.", + "bbox": [ + 511, + 613, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21225", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e3bab8a5ceac6a321f616e8e408d0e1a32fed1ffdfa9a58e13fc88d7235db46a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsMVTec ADVisA
Pixel-levelImage-levelPixel-levelImage-level
CLIP[33] + Conv Adapter89.0 / 85.187.8 / 94.594.8 / 89.882.3 / 83.5
CLIP[33] + FFE Adapter90.6 / 86.290.0 / 95.495.1 / 91.383.3 / 85.0
CLIP[33] + LFS Adapter91.5 / 87.190.6 / 95.795.5 / 91.883.9 / 85.7
FE-CLIP92.6 / 88.391.9 / 96.595.9 / 92.884.6 / 86.6
", + "bbox": [ + 215, + 88, + 782, + 198 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9555c9f026c27fb73638766794429890f5e41a7a55540e2c1819321e780ef512.jpg", + "table_caption": [ + "Table 3. Module ablation on the MVTec AD and the VisA datasets. The metric of Pixel-level is AUROC / PRO for ZSAS task, and the metric of Image-level is AUROC / AP for ZSAD task. Best results are highlighted in red." + ], + "table_footnote": [], + "table_body": "
MethodsMVTec ADVisA
Pixel-levelImage-levelPixel-levelImage-level
FFT92.8 / 88.392.0 / 96.295.5 / 93.084.1 / 86.8
DCT92.6 / 88.391.9 / 96.595.9 / 92.884.6 / 86.6
", + "bbox": [ + 91, + 258, + 488, + 335 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Compared with using FFT to extract the frequency information in the FFE and LFS modules. The metric of Pixel-level is AUROC / PRO for ZSAS task, and the metric of Image-level is AUROC / AP for ZSAD task. Best results are highlighted in red.", + "bbox": [ + 89, + 343, + 483, + 401 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d6f13747d83cf192722772b460591f1b66576a4774d0cb3f468528e3c50c073d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 426, + 169, + 487 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/19181edfeb38d426337f67abc575606f9e563ded07ed5ace7016579f93df73eb.jpg", + "image_caption": [ + "(a) Input Image" + ], + "image_footnote": [], + "bbox": [ + 96, + 488, + 169, + 545 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/919a719e30bba0e8afc8a81c79f7db0966c64af1ac9771609e1bf42911a92bc0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 173, + 426, + 248, + 487 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5131f2a5216916bc8b7df84d3705250873a75f7cc4c87f10966a1a98a7f44d48.jpg", + "image_caption": [ + "(b) $f_{fe}^{m}$ from FFE Adapter" + ], + "image_footnote": [], + "bbox": [ + 173, + 488, + 248, + 545 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2d4704cb6283321213688f49e1fcf5b6c8a811acb9be816fdb5fea86e934c725.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 251, + 426, + 325, + 488 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/61cd967df442bba1adb5db52b48daa51e95ab22d3049bf2ad4684b7194503f66.jpg", + "image_caption": [ + "(c) $f_{N,lf_s}^m$ from LFS Adapter" + ], + "image_footnote": [], + "bbox": [ + 251, + 488, + 325, + 545 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1aa7f5f88c38dc5ab12d1f85613f49cf78fae46ee08df725bb534ffe4d2c00ce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 326, + 426, + 401, + 487 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e7f85ba6cab1ad3ddce06249e23fbd3b3b9a739c90497e16a7968c548736d774.jpg", + "image_caption": [ + "(d) $\\hat{f}_N^m$ from equation (2)", + "Figure 5. The visualization of the learned frequency features. Please zoom in for a better view." + ], + "image_footnote": [], + "bbox": [ + 326, + 488, + 401, + 545 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8dacbcc1c5ca34bb5c9e58ea87bc5dfde841ac5e0ddc160659a58225a04e3901.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 401, + 426, + 478, + 487 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d9d43740d3f9c1df4bb794417b67fcfcae4ceb31aa44f2b19cc30e03997fc358.jpg", + "image_caption": [ + "(e) Final prediction" + ], + "image_footnote": [], + "bbox": [ + 401, + 488, + 478, + 545 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Would FFT work instead of DCT We attempt to use FFT to replace the DCT in the FFE and LFS modules to extract the frequency information. However, there are real values and imaginary values after FFT. To deal with this, the real values and imaginary values are separately processed and then combined to transform back to the spatial domain in the FFE module. As for the LFS module, the averaged local frequency statistics are computed from the real values and imaginary values, respectively, to get the averaged real local frequency statistics and the averaged imaginary local frequency statistics, and then they are summed to get the final averaged local frequency statistics. The results are shown in Table 4. This shows that using FFT achieves similar results with DCT, which means that the key is to inject the frequency information rather than the way of frequency transformation. We choose DCT since DCT is more convenient to extract frequency.", + "bbox": [ + 89, + 643, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Analysis", + "text_level": 1, + "bbox": [ + 511, + 261, + 614, + 277 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Time cost. We report the inference time of APRIL-GAN, AnomalyCLIP, VCP-CLIP, AdaCLIP, and FE-CLIP with PyTorch-1.13.0 on one NVIDIA RTX 3090 24GB GPU. The inference time is 105, 120, 128, 160, and $125\\mathrm{ms}$ , respectively. FE-CLIP achieves a better trade-off between performance and speed.", + "bbox": [ + 511, + 285, + 905, + 376 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visualization of learned frequency features. To show why frequency information is effective for the ZSAD task, we visualize the frequency-aware features from the last block of the visual encoder of CLIP as example, where the FFE and LFS adapters are added. As shown in Figure 5, both the features $f_{N,ffe}^{m}$ from FFE adapter and the features $f_{N,lfs}^{m}$ from LFS adapter highlight the abnormal region of images. By injecting the features $f_{N,ffe}^{m}$ and $f_{N,lfs}^{m}$ into the visual encoder of CLIP, the obtained features $\\hat{f}_N^m$ can successfully localize the main abnormal region, in favor of the final prediction of the model.", + "bbox": [ + 511, + 378, + 906, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "When frequency information can be beneficial. Some properties of anomaly area, such as appearance defects, can be reflected based on its frequency information, since high-frequency components represent fine details (e.g. soft borders) that often exist in anomaly area. In these cases, These specific frequency patterns could be beneficial for anomaly detection. For example, as shown in Figure 1, the frequency statistics in the red region are different from the frequency statistics in the surrounding regions of the input abnormal image and the corresponding region of the normal image.", + "bbox": [ + 511, + 549, + 905, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Let's consider images of normal dumplings and images of squashed dumplings, where squashed dumplings are anomalies. In this case, if the squashed dumplings reveal the meat filling, the frequency of squashed dumplings could still stand out and frequency information could be helpful to detect anomalies. However, if the dumplings are squashed to be more like a meat pie without revealing the meat filling, the frequency of this squashed dumpling might not stand out and frequency information might not work well.", + "bbox": [ + 511, + 702, + 905, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We validate the FE-CLIP across different domains with 10 datasets of highly diverse class semantics from various industrial inspections and medical domains, and the results of FE-CLIP on the datasets of medical domains are trained", + "bbox": [ + 511, + 839, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21226", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f2dc48cc22edaa802f805f5bb14bad245619d22856f888d906f56cd06a4950e7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SetupMethodsMVTec ADVisA
Pixel-levelImage-levelPixel-levelImage-level
1-shotPatchCore [37]92.0 / 79.783.4 / 92.295.4 / 80.579.9 / 82.8
WinCLIP [17]95.2 / 87.193.1 / 96.596.4 / 85.183.8 / 85.1
AnomalyGPT [14]95.3 / -94.1 / -96.2 / -87.4 / -
PromptAD [24]95.9 / -94.6 / -96.7 / -86.9 / -
FE-CLIP96.1 / 90.995.6 / 98.197.9 / 93.988.7 / 90.5
2-shotPatchCore [37]93.3 / 82.386.3 / 93.896.1 / 82.681.6 / 84.8
WinCLIP [17]96.0 / 88.494.4 / 97.096.8 / 86.284.6 / 85.8
AnomalyGPT [14]95.6 / -95.5 / -96.4 / -88.6 / -
PromptAD [24]96.2 / -95.7 / -97.1 / -88.3 / -
InCTRL [53]- / -94.0 / 96.9- / -85.8 / 87.7
FE-CLIP96.5 / 91.596.2 / 98.398.2 / 94.390.3 / 91.6
4-shotPatchCore [37]94.3 / 84.388.8 / 94.596.8 / 84.985.3 / 87.5
WinCLIP [17]96.2 / 89.095.2 / 97.397.2 / 87.687.3 / 88.8
AnomalyGPT [14]96.2 / -96.3 / -96.7 / -90.6 / -
PromptAD [24]96.5 / -96.6 / -97.4 / -89.1 / -
InCTRL [53]- / -94.5 / 97.2- / -87.7 / 90.2
FE-CLIP96.7 / 91.796.6 / 98.498.4 / 95.090.7 / 91.9
", + "bbox": [ + 205, + 87, + 792, + 409 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Compared with other competing methods under few-normal-shot anomaly detection setting. The metric of Pixel-level is AUROC / PRO for the few-normal-shot anomaly segmentation task, and the metric of Image-level is AUROC / AP for the few-normal-shot anomaly detection task. Best results are highlighted in red.", + "bbox": [ + 88, + 419, + 906, + 463 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "by the industrial inspection MVTec AD dataset, as shown in Table 1 and Table 2. The superior performance of FE-CLIP demonstrates that frequency information could be beneficial for many cases of anomaly detection, and the proposed FE-CLIP has good generalization across different domains.", + "bbox": [ + 88, + 479, + 486, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. FE-CLIP with Few Normal Shots", + "text_level": 1, + "bbox": [ + 89, + 571, + 383, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The performance of FE-CLIP can also be further boosted when equipped with few-shot normal images, just like WinCLIP [17] and AnomalyGPT [14]. We adopt the same way as AnomalyGPT to compute the anomaly segmentation map with few-shot normal images. In short words, we use the same visual encoder of FE-CLIP to extract intermediate patch-level features from normal images and store them in memory banks, then calculate the cosine similarity distance between each patch token of input image and its most similar counterpart in the memory banks to obtain the anomaly segmentation map $S_{a,few\\_shot}$ . Finally, we add the anomaly segmentation map $S_{a}$ from the FE-CLIP model with $S_{a,few\\_shot}$ to get the final anomaly segmentation map, and add the anomaly score $M_{a}$ from the FE-CLIP model with the max value of $S_{a,few\\_shot}$ to get the final anomaly score. Please refer to [14] for more technical details.", + "bbox": [ + 91, + 595, + 482, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We compare the FE-CLIP with other competing methods, such as AnomalyGPT [14], PromptAD [24], and InCTRL [53], under the few-shot anomaly detection setting. As shown in Table 5, the FE-CLIP outperforms all other com", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "peting methods across different metrics, including AUROC and PRO metrics for the anomaly detection task, and AUROC and AP metrics for the anomaly segmentation task, on both datasets. The experiments under few-normal-shot setting further demonstrate the superiority of the FE-CLIP.", + "bbox": [ + 511, + 479, + 906, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 579, + 633, + 594 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we find that frequency information can benefit the ZSAD task. To this end, we propose FE-CLIP, which adds the proposed FFE and LFS adapters into the visual encoder of CLIP, to deeply mine the frequency patterns for the visual features. Extensive experiments on 10 datasets of highly diverse class semantics from various defect inspections and medical domains demonstrate that (1) the frequency information can benefit the ZSAD task, and (2) the frequency information can help the model have good generalization across datasets from different domains. We also conduct experiments under the few-normal-shot anomaly detection setting to further validate the superiority of the proposed FE-CLIP. We hope the community of anomaly detection will be aware of the effectiveness of frequency information from our work and explore more ways to utilize the frequency information for anomaly detection in the future.", + "bbox": [ + 511, + 607, + 906, + 849 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 511, + 854, + 681, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the Anhui Provincial Science and Technology Major Project (No. 2023z020006).", + "bbox": [ + 511, + 872, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21227", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Nasir Ahmed, T. Natarajan, and Kamisetty R Rao. Discrete cosine transform. IEEE transactions on Computers, 100(1): 90-93, 1974. 2, 4, 5", + "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 2", + "[3] Toshimichi Aota, Lloyd Teh Tzer Tong, and Takayuki Okatani. Zero-shot versus many-shot: Unsupervised texture anomaly detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5564-5572, 2023. 3, 5", + "[4] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad-a comprehensive real-world dataset for unsupervised anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9592-9600, 2019. 1, 5", + "[5] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Uninformed students: Student-teacher anomaly detection with discriminative latent embeddings. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4183–4192, 2020. 1, 5", + "[6] Jorge Bernal, F Javier Sánchez, Gloria Fernández-Esparrach, Debora Gil, Cristina Rodríguez, and Fernando Vilarino. Wm-dova maps for accurate polyp highlighting in colonoscopy: Validation vs. saliency maps from physicians. Computerized medical imaging and graphics, 43:99-111, 2015. 5", + "[7] Tri Cao, Jiawen Zhu, and Guansong Pang. Anomaly detection under distribution shift. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6511-6523, 2023. 1", + "[8] Yunkang Cao, Jiangning Zhang, Luca Frittoli, Yuqi Cheng, Weiming Shen, and Giacomo Boracchi. Adaclip: Adapting clip with hybrid learnable prompts for zero-shot anomaly detection. In European Conference on Computer Vision, 2024. 1, 2, 5, 6", + "[9] Xuhai Chen, Yue Han, and Jiangning Zhang. A zero-/few-shot anomaly classification and segmentation method for cvpr 2023 vand workshop challenge tracks 1&2: 1st place on zero-shot ad and 4th place on few-shot ad. arXiv preprint arXiv:2305.17382, 2023. 3, 6", + "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.3", + "[11] Tharindu Fernando, Harshala Gammulle, Simon Denman, Sridha Sridharan, and Clinton Fookes. Deep learning for medical anomaly detection-a survey. ACM Computing Surveys (CSUR), 54(7):1-37, 2021. 1", + "[12] Gabriel Goh, Nick Cammarata, Chelsea Voss, Shan Carter, Michael Petrov, Ludwig Schubert, Alec Radford, and Chris" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Olah. Multimodal neurons in artificial neural networks. Distill, 6(3):e30, 2021. 2", + "[13] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021. 3", + "[14] Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Ming Tang, and Jinqiao Wang. Anomalygpt: Detecting industrial anomalies using large vision-language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1932-1940, 2024. 3, 8", + "[15] A. Hamada. Br35h: Brain tumor detection 2020. In https://www.kaggle.com/datasets/ahmedhamada0/brain-tumor-detection, 2020. 5", + "[16] Chaoqin Huang, Haoyan Guan, Aofan Jiang, Ya Zhang, Michael Spratling, and Yan-Feng Wang. Registration based few-shot anomaly detection. In European Conference on Computer Vision, pages 303-319. Springer, 2022. 1", + "[17] Jongheon Jeong, Yang Zou, Taewan Kim, Dongqing Zhang, Avinash Ravichandran, and Onkar Dabeer. Winclip: Zero/few-shot anomaly classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19606-19616, 2023. 1, 3, 4, 5, 6, 8", + "[18] Stepan Jezek, Martin Jonak, Radim Burget, Pavel Dvorak, and Milos Skotak. Deep learning-based defect detection of metal parts: evaluating current methods in complex conditions. In 2021 13th International congress on ultra modern telecommunications and control systems and workshops (ICUMT), pages 66-71. IEEE, 2021. 5", + "[19] Debesh Jha, Pia H Smedsrud, Michael A Riegler, Pål Halvorsen, Thomas De Lange, Dag Johansen, and Håvard D Johansen. Kvasir-seg: A segmented polyp dataset. In MultiMedia Modeling: 26th International Conference, MMM 2020, Daejeon, South Korea, January 5–8, 2020, Proceedings, Part II 26, pages 451–462. Springer, 2020. 5", + "[20] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 2", + "[21] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 2", + "[22] Shengze Li, Jianjian Cao, Peng Ye, Yuhan Ding, Chongjun Tu, and Tao Chen. Clipsam: Clip and sam collaboration for zero-shot anomaly segmentation. arXiv preprint arXiv:2401.12665, 2024. 1, 3", + "[23] Xiaoya Li, Xiaofei Sun, Yuxian Meng, Junjun Liang, Fei Wu, and Jiwei Li. Dice loss for data-imbalanced nlp tasks. arXiv preprint arXiv:1911.02855, 2019. 5", + "[24] Xiaofan Li, Zhizhong Zhang, Xin Tan, Chengwei Chen, Yanyun Qu, Yuan Xie, and Lizhuang Ma. Promptad: Learning prompts with only normal samples for few-shot anomaly" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21228", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024. 8", + "[25] Yuanze Li, Haolin Wang, Shihao Yuan, Ming Liu, Debin Zhao, Yiwen Guo, Chen Xu, Guangming Shi, and Wangmeng Zuo. Myriad: Large multimodal model by applying vision experts for industrial anomaly detection. arXiv preprint arXiv:2310.19070, 2023. 3", + "[26] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5", + "[27] Jie Liu, Yixiao Zhang, Jie-Neng Chen, Junfei Xiao, Yongyi Lu, Bennett A Landman, Yixuan Yuan, Alan Yuille, Yucheng Tang, and Zongwei Zhou. Clip-driven universal model for organ segmentation and tumor detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21152-21164, 2023. 1", + "[28] Philipp Liznerski, Lukas Ruff, Robert A Vandermeulen, Billy Joe Franks, Marius Kloft, and Klaus-Robert Müller. Explainable deep one-class classification. arXiv preprint arXiv:2007.01760, 2020. 1", + "[29] Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, and Aniruddha Kembhavi. Unified-io: A unified model for vision, language, and multi-modal tasks. In The Eleventh International Conference on Learning Representations, 2022. 2", + "[30] Pankaj Mishra, Riccardo Verk, Daniele Fornasier, Claudio Piciarelli, and Gian Luca Foresti. Vt-adj: A vision transformer network for image anomaly detection and localization. In 2021 IEEE 30th International Symposium on Industrial Electronics (ISIE), pages 01-06. IEEE, 2021. 5", + "[31] Ziyuan Qin, Huahui Yi, Qicheng Lao, and Kang Li. Medical image understanding with pretrained vision language models: A comprehensive study. arXiv preprint arXiv:2209.15517, 2022. 1", + "[32] Zhen Qu, Xian Tao, Mukesh Prasad, Fei Shen, Zhengtao Zhang, Xinyi Gong, and Guiguang Ding. Vcp-clip: A visual context prompting model for zero-shot anomaly segmentation. In European Conference on Computer Vision, 2024. 1, 2, 6", + "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 1, 2, 3, 4, 5, 6, 7", + "[34] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18082-18091, 2022. 3", + "[35] Tal Reiss and Yedid Hoshen. Mean-shifted contrastive loss for anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2155-2162, 2023. 1", + "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image" + ], + "bbox": [ + 91, + 92, + 482, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3", + "[37] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14318-14328, 2022. 1, 8", + "[38] Mohammadreza Salehi, Niousha Sadjadi, Soroosh Baselizadeh, Mohammad H Rohban, and Hamid R Rabiee. Multiresolution knowledge distillation for anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14902-14912, 2021. 5", + "[39] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2", + "[40] Rohan Taori, Achal Dave, Vaishaal Shankar, Nicholas Carlini, Benjamin Recht, and Ludwig Schmidt. Measuring robustness to natural distribution shifts in image classification. Advances in Neural Information Processing Systems, 33:18583-18599, 2020. 2", + "[41] Yoad Tewel, Yoav Shalev, Idan Schwartz, and Lior Wolf. Zerocap: Zero-shot image-to-text generation for visual-semantic arithmetic. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17918-17928, 2022. 3", + "[42] Yu Tian, Guansong Pang, Fengbei Liu, Yuanhong Chen, Seon Ho Shin, Johan W Verjans, Rajvinder Singh, and Gustavo Carneiro. Constrained contrastive distribution learning for unsupervised anomaly detection and localisation in medical images. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27-October 1, 2021, Proceedings, Part V 24, pages 128-140. Springer, 2021. 1", + "[43] Yu Tian, Fengbei Liu, Guansong Pang, Yuanhong Chen, Yuyuan Liu, Johan W Verjans, Rajvinder Singh, and Gustavo Carneiro. Self-supervised pseudo multi-class pre-training for unsupervised anomaly detection and segmentation in medical images. Medical image analysis, 90:102930, 2023. 1", + "[44] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4", + "[45] Matthias Wieler and Tobias Hahn. Weakly supervised learning for industrial optical inspection. In DAGM symposium in, page 11, 2007. 5", + "[46] Mengde Xu, Zheng Zhang, Fangyun Wei, Yutong Lin, Yue Cao, Han Hu, and Xiang Bai. A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In European Conference on Computer Vision, pages 736-753. Springer, 2022. 3", + "[47] Zhiyuan You, Lei Cui, Yujun Shen, Kai Yang, Xin Lu, Yu Zheng, and Xinyi Le. A unified model for multi-class" + ], + "bbox": [ + 516, + 92, + 906, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21229", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "anomaly detection. Advances in Neural Information Processing Systems, 35:4571-4584, 2022. 1", + "[48] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16793-16803, 2022. 3", + "[49] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In European Conference on Computer Vision, pages 696-712. Springer, 2022. 3", + "[50] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 3, 6", + "[51] Qihang Zhou, Shibo He, Haoyu Liu, Tao Chen, and Jiming Chen. Pull & push: Leveraging differential knowledge distillation for efficient unsupervised anomaly detection and localization. IEEE Transactions on Circuits and Systems for Video Technology, 2022. 1", + "[52] Qihang Zhou, Guansong Pang, Yu Tian, Shibo He, and Jiming Chen. Anomalyclip: Object-agnostic prompt learning for zero-shot anomaly detection. The thirteenth International Conference on Learning Representations, 2024. 1, 3, 5, 6", + "[53] Jiawen Zhu and Guansong Pang. Toward generalist anomaly detection via in-context residual learning with few-shot sample prompts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024. 8", + "[54] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. In European Conference on Computer Vision, pages 392-408. Springer, 2022. 5" + ], + "bbox": [ + 91, + 90, + 482, + 556 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "21230", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/FE-CLIP_ Frequency Enhanced CLIP Model for Zero-Shot Anomaly Detection and Segmentation/e20fdad0-0c02-434f-a4f5-0e729b38aab0_model.json b/2025/FE-CLIP_ Frequency Enhanced CLIP Model for Zero-Shot Anomaly Detection and Segmentation/e20fdad0-0c02-434f-a4f5-0e729b38aab0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8d4433b1a445c584473ca11f6493b3f9ffb22628 --- /dev/null +++ b/2025/FE-CLIP_ Frequency Enhanced CLIP Model for Zero-Shot Anomaly Detection and Segmentation/e20fdad0-0c02-434f-a4f5-0e729b38aab0_model.json @@ -0,0 +1,2290 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.131, + 0.761, + 0.175 + ], + "angle": 0, + "content": "FE-CLIP: Frequency Enhanced CLIP Model for Zero-Shot Anomaly Detection and Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.213, + 0.203, + 0.783, + 0.221 + ], + "angle": 0, + "content": "Tao Gong \\(^{1,2,3}\\), Qi Chu \\(^{1,2,3*}\\), Bin Liu \\(^{1,2,3}\\), Wei Zhou \\(^{4}\\), Nenghai Yu \\(^{1,2,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.143, + 0.222, + 0.856, + 0.24 + ], + "angle": 0, + "content": "1School of Cyber Science and Technology, University of Science and Technology of China" + }, + { + "type": "text", + "bbox": [ + 0.295, + 0.239, + 0.705, + 0.257 + ], + "angle": 0, + "content": "2Anhui Province Key Laboratory of Digital Security" + }, + { + "type": "text", + "bbox": [ + 0.277, + 0.257, + 0.721, + 0.275 + ], + "angle": 0, + "content": "3the CCCD Key Lab of Ministry of Culture and Tourism" + }, + { + "type": "text", + "bbox": [ + 0.342, + 0.275, + 0.654, + 0.292 + ], + "angle": 0, + "content": "\\(^{4}\\)Ling Yang Industrial Internet Co., Ltd." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.293, + 0.753, + 0.31 + ], + "angle": 0, + "content": "{tgong, qchu, flowice, ynh} @ustc.edu.cn, weizhou8@iflytek.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.345, + 0.327, + 0.361 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.378, + 0.483, + 0.786 + ], + "angle": 0, + "content": "Zero-shot anomaly detection (ZSAD) requires detection models trained using auxiliary data to detect anomalies without any training sample in a target dataset. It is challenging since the models need to generalize to anomalies across different domains. Recently, CLIP-based anomaly detection methods, such as WinCLIP and AnomalyCLIP, have demonstrated superior performance in the ZSAD task, due to the strong zero-shot recognition of the CLIP model. However, they overlook the utilization of frequency information of images. In this paper, we find that frequency information could benefit the ZSAD task, since some properties of the anomaly area, such as appearance defects, can also be reflected based on its frequency information. To this end, We propose Frequency Enhanced CLIP (FE-CLIP), taking advantage of two different but complementary frequency-aware clues, (1) Frequency-aware Feature Extraction adapter, and (2) Local Frequency Statistics adapter, in the visual encoder of CLIP, to deeply mine frequency information for the ZSAD task. We apply DCT as the frequency-domain transformation. Through comprehensive experiments, we show that the proposed FE-CLIP has good generalization across different domains and achieves superior zero-shot performance of detecting and segmenting anomalies in 10 datasets of highly diverse class semantics from various defect inspections and medical domains. Besides, FE-CLIP also achieves superior performance under the few-normal-shot anomaly detection settings." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.817, + 0.222, + 0.833 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.844, + 0.483, + 0.875 + ], + "angle": 0, + "content": "Anomaly Detection aims to predict an image or a pixel as normal or anomalous, and has been widely applied in" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.346, + 0.908, + 0.59 + ], + "angle": 0, + "content": "various applications, such as industrial defect inspection [4, 5, 7, 16, 28, 35, 37, 47, 51] and medical image analysis [11, 27, 31, 42, 43]. Existing AD approaches typically assume that training examples in a target application domain are available for learning the detection models. However, this assumption may not hold in various scenarios, such as i) when accessing training data violates data privacy policies, or ii) when the target domain does not have relevant training data. Therefore, Zero-shot anomaly detection (ZSAD) is an emerging task for AD in such scenarios, to which the AD methods mentioned above are not viable, as it requires detection models to detect anomalies without any training sample in a target dataset. ZSAD is challenging since the models need to generalize to anomalies across different domains where the appearance of foreground objects, abnormal regions, and background features can vary significantly." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Recently, large pre-trained vision-language models (VLMs) have demonstrated strong zero-shot recognition ability in various vision tasks. Particularly, being pretrained using millions/billions of image-text pairs, CLIP [33] has been applied to empower various downstream tasks with its strong generalization capability. WinCLIP [17] is the first work to apply the CLIP model in the ZSAD line. It uses a large number of hand-crafted text prompts and extraction of window/patch/image-level features aligned with text to finish the ZSAD task. However, WinCLIP involves multiple forward passes of image patches for anomaly segmentation, which leads to inefficiency. Besides, WinCLIP only performs the testing of the CLIP model in the ZSAD without any fine-tuning using the anomaly detection task data, which also limits its performance in the ZSAD task. Subsequently, several works [8, 22, 32, 52] attempt to adapt the CLIP into the ZSAD task by fintuning. Specifically, AnomalyCLIP [52] learns object-agnostic text prompts that capture generic normality and abnormality in an image regardless of its foreground objects. ClipSAM [22] introduces" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.233, + 0.901 + ], + "angle": 0, + "content": "*Corresponding Author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21220" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.091, + 0.191, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.196, + 0.092, + 0.285, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.092, + 0.379, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.092, + 0.472, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.164, + 0.191, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.236, + 0.171, + 0.246 + ], + "angle": 0, + "content": "(a) normal" + }, + { + "type": "image", + "bbox": [ + 0.196, + 0.164, + 0.285, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.222, + 0.236, + 0.259, + 0.245 + ], + "angle": 0, + "content": "(b) hole" + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.164, + 0.379, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.315, + 0.236, + 0.354, + 0.246 + ], + "angle": 0, + "content": "(c) crack" + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.164, + 0.472, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.408, + 0.236, + 0.446, + 0.246 + ], + "angle": 0, + "content": "(d) print" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.258, + 0.483, + 0.342 + ], + "angle": 0, + "content": "Figure 1. The local frequency statistics of images. The second row represents the local frequency statistics of non-overlapped image regions. There are four columns, standing for the normal object, object with a hole, object with a crack, and object with print, respectively. Please zoom in and pay attention to the red region (i.e. the anomaly region of the input image) for a better view." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.371, + 0.483, + 0.522 + ], + "angle": 0, + "content": "a unified multiscale cross-modal interaction module for interacting language with visual features at multiple stages of the CLIP visual encoder to reason anomaly positions. VCP-CLIP [32] propose the Pre-VCP module and PostVCP module to perform visual context prompting to activate CLIP's anomalous semantic perception ability. AdaCLIP [8] incorporates static and dynamic learnable prompts to adapt CLIP for ZSAD. However, all of the methods mentioned above overlook the utilization of frequency information of images." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.523, + 0.483, + 0.689 + ], + "angle": 0, + "content": "In this paper, we find that frequency information could benefit the ZSAD task. Some properties of the anomaly area, such as appearance defects, can also be reflected based on its frequency information, since high-frequency components represent fine details (e.g. soft borders) that often exist in the anomaly area. For example, as shown in Figure 1, the frequency statistics in the red region are different from the frequency statistics in the surrounding regions of the input abnormal image and the corresponding region of the normal image. These specific frequency patterns could be helpful for anomaly detection." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To this end, we introduce a novel approach, namely the Frequency Enhanced CLIP (FE-CLIP), to inject frequency information into CLIP for accurate ZSAD across different domains. FE-CLIP takes advantage of two different but complementary frequency-aware clues, (1) Frequency-aware Feature Extraction (FFE) adapter, and (2) Local Frequency Statistics (LFS) adapter, in the visual encoder of CLIP, to deeply mine frequency information. Specifically, the FFE adapter uses DCT [1] to transform the image features into the frequency domain in order to mine the frequency information, then uses inverse DCT to transform back to the spatial domain. The FFE adapter describes the frequency-aware patterns in the spatial domain, but does not explicitly render the frequency information directly in" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.319 + ], + "angle": 0, + "content": "the neural networks. Therefore, we use LFS to render the frequency information directly in the frequency domain. In each densely but regularly sampled local spatial patch, the statistics are gathered by counting the mean frequency responses. The averaged frequency statistics are a multichannel spatial map, where the number of channels is identical to the number of channels in the spatial domain. The local frequency statistics also follow the spatial layouts as the features of the CLIP visual encoder, thus also enjoying effective representation learning powered by the CLIP visual encoder. Meanwhile, since the frequency-aware feature extraction and local frequency statistics are complementary to each other but both of them share inherently similar frequency-aware semantics, thus they can be progressively fused during the feature learning process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.32, + 0.905, + 0.424 + ], + "angle": 0, + "content": "Through comprehensive experiments, we show that the proposed FE-CLIP has good generalization across different domains and achieves superior zero-shot performance of detecting and segmenting anomalies in 10 datasets of highly diverse class semantics from various defect inspections and medical domains. Our contributions in this paper are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.426, + 0.905, + 0.486 + ], + "angle": 0, + "content": "1. We find that frequency information could benefit the ZSAD task, and propose a novel Frequency Enhanced CLIP (FE-CLIP) method to inject frequency information into CLIP for accurate ZSAD across different domains." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.487, + 0.905, + 0.563 + ], + "angle": 0, + "content": "2. We propose two different but complementary frequency-aware adapters, Frequency-aware Feature Extraction (FFE) adapter, and Local Frequency Statistics (LFS) adapter, in the visual encoder of CLIP, to deeply mine the frequency patterns of anomaly region." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.563, + 0.907, + 0.683 + ], + "angle": 0, + "content": "3. Extensive experiments show that the proposed FE-CLIP has good generalization across different domains and achieves superior zero-shot performance of detecting and segmenting anomalies in 10 datasets of highly diverse class semantics from various defect inspections and medical domains. Besides, the proposed FE-CLIP also achieves superior performance under the few-normal-shot anomaly detection settings." + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.426, + 0.907, + 0.683 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.701, + 0.655, + 0.716 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.727, + 0.749, + 0.743 + ], + "angle": 0, + "content": "2.1. Vision-language modeling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Among recent successes of large pre-trained vision-language models (VLM) [2, 20, 33], CLIP [33] is the first to perform pre-training on web-scale image-text data, showing unprecedented generality: e.g., its language-driven zero-shot inference, improved both effective robustness [40] and perceptual alignment [12]. Many following VLM works explored large-scale pre-training in different aspects, e.g., scaling up data [20], efficient designs [21], multi-tasks [29], etc. To democratize large-scale VLM for the usages in different domains, a billion-scale data LAION5B [39], a" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21221" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.159, + 0.099, + 0.848, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.377, + 0.908, + 0.421 + ], + "angle": 0, + "content": "Figure 2. The framework of the proposed FE-CLIP. The text encoder and visual encoder are from the CLIP model, and the FFE and LFS are the proposed adapters. We only show the class token embedding after the last block of the CLIP visual encoder for convenience. Actually, the CLIP visual encoder can output the class token after every block." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.446, + 0.484, + 0.611 + ], + "angle": 0, + "content": "code base of OpenCLIP with pre-trained models are open-sourced. Other works presented CLIP's promise in zero/few-shot transfer to downstream tasks beyond classification [13, 36, 41, 46]. Good prompt engineering and tuning [33, 50] can non-trivially benefit generalization performances. Moreover, some other works [34, 48, 49] leverage the pre-trained CLIP for language-guided detection and segmentation with promising performances. In this paper, we mainly focus on adapting CLIP into the ZSAD task with superior performance by injecting the frequency information into the visual encoder of CLIP." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.622, + 0.353, + 0.639 + ], + "angle": 0, + "content": "2.2. Zero-shot Anomaly detection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.484, + 0.903 + ], + "angle": 0, + "content": "ZSAD relies on the model's strong transferability to handle unseen anomalies [3]. A very recent approach WinCLIP [17] presents a seminal work that leverages CLIP [33] for zero-shot anomaly classification and segmentation. WinCLIP proposes a compositional ensemble on state words and prompt templates and efficient extraction and aggregation of window/patch/image-level features aligned with text to finish the ZSAD task. To tackle this inefficiency, APRILGAN [9] introduces a fine-tuning strategy by adding extra learnable linear projection layers in multiple stages of the CLIP visual encoder to enhance the modeling of local visual semantics. AnomalyCLIP [52] learns object-agnostic text prompts that capture generic normality and abnormality in an image regardless of its foreground objects. ClipSAM [22] introduces a unified multiscale cross-modal interaction module for interacting language with visual features at multiple stages of the CLIP visual encoder to reason anomaly" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.446, + 0.909, + 0.628 + ], + "angle": 0, + "content": "positions. Myriad [25] and AnomalyGPT [14] utilize large language models for zero-shot/few-shot anomaly detection. They employ an image decoder to provide fine-grained semantics and design a prompt learner to fine-tune the large language models using prompt embeddings. However, all of the methods mentioned above overlook the utilization of frequency information of images. In this paper, we find that frequency information could benefit the ZSAD task, and propose the FE-CLIP method which utilizes two different but complementary adapters, the FFE adapter and the LFS adapter, into the visual encoder of CLIP, to deeply mine the frequency information for ZSAD task." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.642, + 0.608, + 0.658 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.667, + 0.644, + 0.684 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.903 + ], + "angle": 0, + "content": "CLIP [33] is a large-scale pretraining method offering a joint vision-language representation. It consists of a text encoder \\( T(\\cdot) \\) and a visual encoder \\( F(\\cdot) \\). Both encoders are mainstream multi-block networks such as ViT [10]. Using text prompts is a typical way to acquire the embeddings of different classes for zero-shot classification. The text prompts usually combine a text prompt template \\( \\Omega \\) with the category name \\( c \\), where the text prompt template \\( \\Omega \\) usually is defined as A photo of a [cls] and [cls] represents the target category name. The text prompts are passed through \\( T(\\cdot) \\) to obtain its corresponding textual embedding \\( g_{c} \\in \\mathbb{R}^{D} \\) where \\( D \\) denotes the channel dimension. Then, an input image is passed through \\( F(\\cdot) \\) to get the visual representations, where the class token \\( f \\in \\mathbb{R}^{D} \\) is the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21222" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.213 + ], + "angle": 0, + "content": "global visual embedding of the input image, and patch tokens \\( f^{m} \\in \\mathbb{R}^{H \\times W \\times D} \\) are treated as local visual embeddings of the input image. \\( H \\) and \\( W \\) denote the height and width of the patch tokens, respectively. CLIP can perform zero-shot recognition by measuring the similarity between textual and visual embeddings. Specifically, given a target class set \\( \\mathcal{C} \\) and an image, CLIP predicts the probability of the image belonging to the \\( c \\) category as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.223, + 0.483, + 0.259 + ], + "angle": 0, + "content": "\\[\np (y = c) = P \\left(g _ {c}, f\\right) = \\frac {\\exp \\left(< g _ {c} , f > / \\tau\\right)}{\\sum_ {c \\in \\mathcal {C}} \\exp \\left(< g _ {c} , f >\\right) / \\tau} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.268, + 0.484, + 0.36 + ], + "angle": 0, + "content": "where \\(\\tau\\) denotes the temperature and the operator \\(< ,\\cdot ,>\\) represents the computation of cosine similarity. The computation can be extended from global visual embeddings to local visual embeddings to derive the corresponding segmentation maps \\(M_{c}\\in \\mathbb{R}^{H\\times W}\\) , if each entry \\((i,j)\\) is computed as \\(P\\left(g_c,f^{m(i,j)}\\right)\\)" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.367, + 0.339, + 0.383 + ], + "angle": 0, + "content": "3.2. Frequency Enhanced CLIP" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.389, + 0.484, + 0.768 + ], + "angle": 0, + "content": "In this paper, we find that frequency information could benefit the ZSAD task, and propose a novel Frequency Enhanced CLIP (FE-CLIP) method to inject frequency information into CLIP [33] for accurate ZSAD across different domains. We inject the frequency information by the features of the CLIP visual encoder rather than the frequency of the input image, since the CLIP visual encoder is trained to take RGB image as input. Typically, the CLIP visual encoder comprises a series of block layers. From the bottom to the top of the layers, the visual encoder gradually learns the visual patterns at different levels of abstraction. As shown in Figure 2, we add Frequency-aware Feature Extraction (FFE) adapter and Local Frequency Statistics (LFS) adapter across multiple blocks levels in the visual encoder \\( F(\\cdot) \\) of CLIP, while keeping its original backbone unchanged, thus enabling frequency information injection at multiple feature levels. To be specific, assuming the visual encoder consists of \\( N \\) blocks (\\( N = 4 \\)), the features (i.e. the patch tokens) after the \\( n \\)-th block are denoted as \\( f_{n}^{m} \\in \\mathbb{R}^{H \\times W \\times D} \\), and the class token after the \\( n \\)-th block is denoted as \\( f_{n} \\in \\mathbb{R}^{D} \\), where \\( n = 1, 2,.., N \\). Firstly, the features \\( f_{n}^{m} \\) are passed through an FFE adapter and an LFS adapter to get the features \\( f_{n,ffe}^{m} \\) and \\( f_{n,lfs}^{m} \\), respectively, which contain the frequency information. We can get the frequency-aware feature \\( \\hat{f}_{n}^{m} \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.778, + 0.483, + 0.799 + ], + "angle": 0, + "content": "\\[\n\\hat {f} _ {n} ^ {m} = \\lambda (f _ {n, f f e} ^ {m} + f _ {n, l f s} ^ {m}) + (1 - \\lambda) f _ {n} ^ {m} \\qquad (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "with \\(\\hat{f}_n^m\\) serving as the input for the next block. The class token \\(f_n\\) will also be injected with frequency information after the next block due to the self-attention mechanism [44] in the blocks. We set \\(\\lambda = 0.1\\) to preserve the original knowledge of the CLIP visual encoder. Then, the normal text prompt and abnormal text prompt are passed through" + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.089, + 0.898, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.169, + 0.905, + 0.197 + ], + "angle": 0, + "content": "Figure 3. The proposed Frequency-aware Feature Extraction (FFE) adapter." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.222, + 0.907, + 0.532 + ], + "angle": 0, + "content": "text encoder \\( T(\\cdot) \\) to get corresponding textual embedding \\( g_{r} \\) and \\( g_{a} \\), respectively. \\( g_{r} \\) and \\( g_{a} \\) constitute the target class set \\( \\mathcal{C} \\). We simply use A photo of a normal object and A photo of a damaged object for the normal text prompt and abnormal text prompt, respectively, rather than the complex compositional ensemble on state words and prompt templates used in WinCLIP [17], since we find that the simple text prompt also works well. Finally, we compute the abnormal segmentation map \\( M_{a,n} \\in \\mathbb{R}^{H \\times W} \\) based on the equation 1, where each entry \\( (i,j) \\) is computed as \\( P\\left(g_{a}, fc(\\hat{f}_{n}^{m(i,j)})\\right) \\). The \\( fc \\) means that we use a single learnable fc to align the dimension of visual features \\( \\hat{f}_{n}^{m} \\) to the dimension of text features. The abnormal score \\( S_{a,n} \\) is also computed based on the equation 1 as \\( P\\left(g_{a}, fc_{clip}(f_{n})\\right) \\). The \\( fc_{clip} \\) denotes the frozen project layer of the visual encoder in CLIP to align the dimension of class token \\( f_{n} \\) and the dimension of text features. The final prediction of the abnormal score \\( S_{a} \\) and abnormal segmentation map \\( M_{a} \\) are computed by averaging \\( S_{a,n} \\) and \\( M_{a,n} \\) across \\( N \\) blocks, respectively, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.542, + 0.907, + 0.584 + ], + "angle": 0, + "content": "\\[\nS _ {a} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} S _ {a, n}, \\quad M _ {a} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} M _ {a, n} \\tag {3}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.592, + 0.902, + 0.608 + ], + "angle": 0, + "content": "3.3. Frequency-aware Feature Extraction Adapter" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.901 + ], + "angle": 0, + "content": "The specific frequency patterns could be beneficial for anomaly detection. To mine the frequency information, we propose the Frequency-aware Feature Extraction (FFE) adapter in the visual encoder of CLIP. Specifically, given the patch tokens of the \\(n\\)-th block \\(f_{n}^{m}\\in \\mathbb{R}^{H\\times W\\times D}\\), we first split the \\(f_{n}^{m}\\) into non-overlapped windows with total \\((H / P)\\times (W / P)\\) windows and each window containing \\(P\\times P\\) patch tokens (\\(P\\) is set to 3 by default). As shown in Figure 3, we apply DCT [1] transformation to each window to extract the DCT features \\(f_{n,ffe,1}^{m}\\) which represents the frequency information. Then, we pass \\(f_{n,ffe,1}^{m}\\) through one linear layer followed by one GELU(\\(\\cdot\\)) activation layer to further extract the frequency features \\(f_{n,ffe,2}^{m}\\). Finally, we apply the inverse DCT transformation to \\(f_{n,ffe,2}^{m}\\) to get the frequency-aware features \\(f_{n,ffe}^{m}\\) in the spatial domain. The FFE adapter describes the frequency-aware patterns in the spatial domain, but does not explicitly render the frequency information directly in the neural networks. In the next sub-section, we propose an adapter that renders the fre" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21223" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.09, + 0.475, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.177, + 0.483, + 0.207 + ], + "angle": 0, + "content": "Figure 4. The proposed Local Frequency Statistics (LFS) adapter." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.232, + 0.431, + 0.248 + ], + "angle": 0, + "content": "quency information directly in the neural networks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.255, + 0.398, + 0.273 + ], + "angle": 0, + "content": "3.4. Local Frequency Statistics Adapter" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.278, + 0.485, + 0.626 + ], + "angle": 0, + "content": "The aforementioned FFE adapter has provided frequency-aware representation, but it has to represent the frequency-aware patterns back into the spatial domain, thus failing to directly utilize the frequency patterns in the frequency domain. Therefore, we propose the Local Frequency Statistics (LFS) adapter in the visual encoder of CLIP to render the frequency patterns directly in the frequency domain. As shown in Figure 4, given the patch tokens of the \\(n\\)-th block \\(f_{n}^{m}\\in \\mathbb{R}^{H\\times W\\times D}\\), we first apply a Sliding Window DCT (SW DCT) on \\(f_{n}^{m}\\) (i.e., taking DCT [1] densely on sliding windows of the features) to extract the localized frequency responses \\(f_{n,lfs,1}^{m}\\). The size of sliding windows is \\(Q\\times Q\\) (\\(Q\\) is set to 3 by default). Therefore, there are \\(Q\\times Q\\) group frequency responses in \\(f_{n,lfs,1}^{m}\\), where the first one represents the low-frequency responses and the last one represents the high-frequency responses. Then we count the mean of \\(f_{n,lfs,1}^{m}\\) across \\(Q\\times Q\\) groups to get the mean frequency responses \\(f_{n,lfs,2}^{m}\\). The \\(f_{n,lfs,2}^{m}\\) is a multi-channel spatial map that shares the same layout as \\(f_{n}^{m}\\). The LFS adapter outputs \\(f_{n,lfs}^{m}\\) after \\(f_{n,lfs,2}^{m}\\) being passed through a convolutional layer followed by one \\(GELU(\\cdot)\\) activation layer. The LFS adapter provides a localized aperture to detect detailed abnormal frequency distributions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.633, + 0.245, + 0.648 + ], + "angle": 0, + "content": "3.5. Loss Functions" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.655, + 0.483, + 0.716 + ], + "angle": 0, + "content": "Let \\( S_{gt} \\) denote the ground-truth image label where 1 represents the abnormal image and 0 represents the normal image. The loss for the anomaly classification is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.724, + 0.482, + 0.776 + ], + "angle": 0, + "content": "\\[\nL _ {c l s} = - \\frac {1}{N} \\sum_ {N} ^ {n = 1} \\left[ S _ {g t} \\log \\left(S _ {a, n}\\right) + \\left(1 - S _ {g t}\\right) \\log \\left(1 - S _ {a, n}\\right) \\right] \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.777, + 0.483, + 0.845 + ], + "angle": 0, + "content": "where \\( \\log \\) denotes the log operation. Similarly, let \\( M_{gt} \\) denote the ground-truth pixel label of the input image, where 1 means abnormal pixel region and 0 means normal pixel region. The loss for the anomaly segmentation is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.852, + 0.507, + 0.902 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {m a s k}} = \\frac {1}{N} \\sum_ {N} ^ {n = 1} [ \\operatorname {F o c a l} (U p (M _ {a, n}), M _ {g t}) + D i c e (U p (M _ {a, n}), M _ {g t}) ] \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.183 + ], + "angle": 0, + "content": "where \\( Focal(\\cdot, \\cdot) \\) and \\( Dice(\\cdot, \\cdot) \\) denote the focal loss [26] and dice loss [23], respectively. The operator \\( Up(\\cdot) \\) represents the unsampling operation. During training, both the visual encoder and the text encoder of CLIP [33] are Frozen. Only the FFE adapters and LFS adapters are optimized by the loss function \\( L_{total} = L_{cls} + L_{mask} \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.197, + 0.646, + 0.214 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.223, + 0.8, + 0.238 + ], + "angle": 0, + "content": "4.1. Datasets and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.245, + 0.907, + 0.381 + ], + "angle": 0, + "content": "To verify the effectiveness of the proposed method FE-CLIP, we conduct comprehensive experiments across 10 real-world anomaly detection datasets, covering various industrial inspection scenarios and medical imaging domains. In industrial inspection, we use MVtec AD [4], VisA [54], MPDD [18], BTAD [30], DAGM [45], and DTD-Synthetic [3]. In medical imaging, we use colon polyp detection datasets CVC-ClinicDB [6] and Kvasir [19], brain tumor detection datasets BrainMRI [38] and Br35H [15]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.382, + 0.906, + 0.472 + ], + "angle": 0, + "content": "As for the evaluation metrics, following previous works [17, 52], we use Area Under the Receiver Operating Characteristic curve (AUROC). Additionally, Average Precision (AP) for Zero-Shot Anomaly Detection (ZSAD) and PRO [5] for Zeor-Shot Anomaly Segmentation (ZSAS) are also used to provide a more in-depth analysis of the results." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.483, + 0.731, + 0.499 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.505, + 0.907, + 0.716 + ], + "angle": 0, + "content": "We use the publicly available CLIP model (VITL/14@336px) as our backbone. Model parameters of CLIP are all frozen, and only the parameters of FFE adapters and LFS adapters are learnable. Following AnomalyCLIP [52], we fine-tune FE-CLIP using the test data of MVtec AD and evaluate the ZSAD performance on other datasets. As for MVtec AD, we fine-tune FE-CLIP on the test data of VisA. The proposed FE-CLIP is trained by 9 epochs with Adam optimizer. The learning rate is set to 5e-4 and the total batch size is 16. All experiments are conducted in PyTorch-1.13.0 with four NVIDIA RTX 3090 24GB GPUs. All experiment results are averaged under five runs. We report dataset-level results, which are averaged across their respective sub-datasets." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.727, + 0.652, + 0.742 + ], + "angle": 0, + "content": "4.3. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "ZSAD performance on diverse anomaly scenarios. Table 1 shows the ZSAD results of FE-CLIP with other competing methods over six industrial defect datasets and two medical domain datasets. As shown in Table 1, FE-CLIP can achieve higher AUROC and AP for the ZSAD task on most industrial inspection datasets and medical datasets, such as MVTec AD, DTD-Synthetic, and BrainMRI, when compared with other competing methods, such as WinCLIP [17], AnomalyCLIP [52] and AdaCLIP [8]. This shows that (1) the frequency information can benefit the ZSAD task" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21224" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.089, + 0.905, + 0.226 + ], + "angle": 0, + "content": "
MethodsIndustrial DefectsMedical Anomalies
MVTec ADVisAMPDDBTADDAGMDTD-SyntheticBrainMRIBr35H
CLIP [33]74.1 / 87.666.4 / 71.554.3 / 65.434.5 / 52.579.6 / 59.071.6 / 85.773.9 / 81.778.4 / 78.8
CoOp [50]88.8 / 94.862.8 / 68.155.1 / 64.266.8 / 77.487.5 / 74.6- / -61.3 / 44.986.0 / 87.5
WinCLIP [17]91.8 / 96.578.1 / 81.263.6 / 69.968.2 / 70.991.8 / 79.593.2 / 92.686.6 / 91.580.5 / 82.2
APRIL-GAN [9]86.1 / 93.578.0 / 81.478.0 / 81.473.6 / 68.694.4 / 83.886.4 / 95.089.3 / 90.993.1 / 92.9
AnomalyCLIP [52]91.5 / 96.282.1 / 85.477.0 / 82.088.3 / 87.397.5 / 92.393.5 / 97.090.3 / 92.294.6 / 94.7
AdaCLIP [8]89.2 / -85.8 / -76.0 / -88.6 / -99.1 / -95.5 / -94.8 / -97.7 / -
FE-CLIP91.9 / 96.584.6 / 86.678.0 / 82.690.3 / 90.097.5 / 92.398.3 / 99.494.8 / 93.896.8 / 93.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.235, + 0.907, + 0.266 + ], + "angle": 0, + "content": "Table 1. AUROC / AP results for Zero-Shot Anomaly Detection (ZSAD) task on eight real-world anomaly detection datasets. Best results are highlighted in red." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.277, + 0.905, + 0.425 + ], + "angle": 0, + "content": "
MethodsIndustrial DefectsMedical Anomalies
MVTec ADVisAMPDDBTADDAGMDTD-SyntheticCVC-ClinicDBKvasir
CLIP [33]38.4 / 11.346.6 / 14.862.1, 33.030.6 / 4.430.6 / 4.433.9 / 12.547.5 / 18.944.6 / 17.7
CoOp [50]33.3 / 6.724.2 / 3.815.4 / 2.315.4 / 2.317.5 / 2.1- / -17.5 / 2.144.1 / 3.5
WinCLIP [17]85.1 / 64.679.6 / 56.879.6 / 56.872.7 / 27.387.6 / 65.783.9 / 57.851.2 / 13.869.7 / 24.5
APRIL-GAN [9]87.6 / 84.094.2 / 86.894.1 / 83.260.8 / 25.082.4 / 66.295.3 / 86.980.5 / 60.775.0 / 36.2
AnomalyCLIP [52]91.1 / 81.495.5 / 87.096.5 / 88.794.2 / 74.895.6 / 91.097.9 / 92.382.9 / 67.878.9 / 45.6
AdaCLIP [8]88.7 / -95.5 / -96.1 / -92.1 / -91.5 / -97.9 / -84.4 / -- / -
VCP-CLIP [32]92.0 / 87.395.7 / 90.7- / -94.1 / 74.699.4 / 98.3- / -- / -- / -
FE-CLIP92.6 / 88.395.9 / 92.897.0 / 90.595.6 / 80.498.5 / 96.699.0 / 97.484.5 / 69.279.8 / 60.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.433, + 0.908, + 0.477 + ], + "angle": 0, + "content": "Table 2. AUROC / PRO results for Zero-Shot Anomaly Segmentation (ZSAS) task on eight real-world anomaly detection datasets. Best results are highlighted in red. Note that the image-level medical AD datasets do not contain segmentation ground truth, so the pixel-level medical AD datasets are different from the image-level datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.503, + 0.483, + 0.548 + ], + "angle": 0, + "content": "due to the superior performance of FE-CLIP, and (2) the frequency information helps FE-CLIP achieve good generalization across datasets from different domains." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.551, + 0.483, + 0.746 + ], + "angle": 0, + "content": "ZSAS performance on diverse anomaly scenarios. We show the ZSAS results of FE-CLIP with other competing methods over six industrial defect datasets and two medical domain datasets in Table 2. Table 2 shows that FE-CLIP achieves higher AUROC and PRO for the ZSAS task on almost all of the industrial inspection datasets and medical datasets, except for the DAGM dataset, when compared with other competing methods, such as WinCLIP [17], AnomalyCLIP [52], AdaCLIP [8] and VCP-CLIP [32]. This further demonstrates that (1) the frequency information can benefit the ZSAS task and (2) the frequency information helps FE-CLIP achieve good generalization across datasets from different domains." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Discussion. We also note that the FE-CLIP can defeat almost all of the competing methods on the ZSAS task, while FE-CLIP only performs better than other methods over most of the datasets on ZSAD task. The reason may be that the frequency information can be directly injected into the patch tokens (i.e. the feature map used for the ZSAS task) through the FFE adapter and LFS adapter, while the frequency information can only be implicitly involved in the global class token (i.e. the feature vector used for the ZSAD task) from the patch tokens based on the self" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.504, + 0.659, + 0.516 + ], + "angle": 0, + "content": "attention mechanism." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.519, + 0.907, + 0.58 + ], + "angle": 0, + "content": "Overall, the proposed FE-CLIP achieves superior zero-shot performance of detecting and segmenting anomalies in datasets of highly diverse class semantics from various defect inspection and medical imaging domains." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.591, + 0.667, + 0.608 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.903 + ], + "angle": 0, + "content": "The effectiveness of FFE and LFS We conduct ablation experiments on MVTec AD dataset and VisA dataset to validate the effectiveness of the proposed FFE adapter and LFS adapter. The results are shown in Table 3. The \"CLIP \\([33] + \\text{Conv Adapter}\"\\) means we add \\(N\\) conv adapters, where each of them consists of one convolutional layer followed by one \\(GELU(\\cdot)\\) activation layer to keep the same learnable parameters with the LFS adapter into the visual encoder of CLIP. It can be concluded that both the FFE adapter and LFS adapter can achieve better performance than the conv adapter in MVTec AD dataset and VisA dataset across the ZSAD task and ZSAS task. This also shows that the frequency information can benefit the ZSAD task and ZSAS task. When combined the FFE adapter with the LFS adapter, the FE-CLIP achieves higher performance than the FFE adapter and LFS adapter. This demonstrates that the FFE adapter and LFS adapter can mine different but complementary frequency information to further boost the performance of model on the ZSAD and ZSAS tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21225" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.217, + 0.089, + 0.784, + 0.199 + ], + "angle": 0, + "content": "
MethodsMVTec ADVisA
Pixel-levelImage-levelPixel-levelImage-level
CLIP[33] + Conv Adapter89.0 / 85.187.8 / 94.594.8 / 89.882.3 / 83.5
CLIP[33] + FFE Adapter90.6 / 86.290.0 / 95.495.1 / 91.383.3 / 85.0
CLIP[33] + LFS Adapter91.5 / 87.190.6 / 95.795.5 / 91.883.9 / 85.7
FE-CLIP92.6 / 88.391.9 / 96.595.9 / 92.884.6 / 86.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.207, + 0.908, + 0.237 + ], + "angle": 0, + "content": "Table 3. Module ablation on the MVTec AD and the VisA datasets. The metric of Pixel-level is AUROC / PRO for ZSAS task, and the metric of Image-level is AUROC / AP for ZSAD task. Best results are highlighted in red." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.26, + 0.489, + 0.336 + ], + "angle": 0, + "content": "
MethodsMVTec ADVisA
Pixel-levelImage-levelPixel-levelImage-level
FFT92.8 / 88.392.0 / 96.295.5 / 93.084.1 / 86.8
DCT92.6 / 88.391.9 / 96.595.9 / 92.884.6 / 86.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.344, + 0.484, + 0.402 + ], + "angle": 0, + "content": "Table 4. Compared with using FFT to extract the frequency information in the FFE and LFS modules. The metric of Pixel-level is AUROC / PRO for ZSAS task, and the metric of Image-level is AUROC / AP for ZSAD task. Best results are highlighted in red." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.428, + 0.171, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.489, + 0.171, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.103, + 0.547, + 0.166, + 0.556 + ], + "angle": 0, + "content": "(a) Input Image" + }, + { + "type": "image", + "bbox": [ + 0.174, + 0.428, + 0.249, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.174, + 0.489, + 0.249, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.182, + 0.547, + 0.241, + 0.565 + ], + "angle": 0, + "content": "(b) \\(f_{fe}^{m}\\) from FFE Adapter" + }, + { + "type": "image", + "bbox": [ + 0.252, + 0.428, + 0.326, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.252, + 0.489, + 0.326, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.26, + 0.547, + 0.316, + 0.565 + ], + "angle": 0, + "content": "(c) \\(f_{N,lf_s}^m\\) from LFS Adapter" + }, + { + "type": "image", + "bbox": [ + 0.328, + 0.428, + 0.402, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.328, + 0.489, + 0.402, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.34, + 0.547, + 0.389, + 0.565 + ], + "angle": 0, + "content": "(d) \\(\\hat{f}_N^m\\) from equation (2)" + }, + { + "type": "image", + "bbox": [ + 0.403, + 0.428, + 0.48, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.403, + 0.489, + 0.48, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.422, + 0.547, + 0.461, + 0.564 + ], + "angle": 0, + "content": "(e) Final prediction" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.576, + 0.483, + 0.604 + ], + "angle": 0, + "content": "Figure 5. The visualization of the learned frequency features. Please zoom in for a better view." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Would FFT work instead of DCT We attempt to use FFT to replace the DCT in the FFE and LFS modules to extract the frequency information. However, there are real values and imaginary values after FFT. To deal with this, the real values and imaginary values are separately processed and then combined to transform back to the spatial domain in the FFE module. As for the LFS module, the averaged local frequency statistics are computed from the real values and imaginary values, respectively, to get the averaged real local frequency statistics and the averaged imaginary local frequency statistics, and then they are summed to get the final averaged local frequency statistics. The results are shown in Table 4. This shows that using FFT achieves similar results with DCT, which means that the key is to inject the frequency information rather than the way of frequency transformation. We choose DCT since DCT is more convenient to extract frequency." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.262, + 0.616, + 0.278 + ], + "angle": 0, + "content": "4.5. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.286, + 0.906, + 0.377 + ], + "angle": 0, + "content": "Time cost. We report the inference time of APRIL-GAN, AnomalyCLIP, VCP-CLIP, AdaCLIP, and FE-CLIP with PyTorch-1.13.0 on one NVIDIA RTX 3090 24GB GPU. The inference time is 105, 120, 128, 160, and \\(125\\mathrm{ms}\\), respectively. FE-CLIP achieves a better trade-off between performance and speed." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.379, + 0.907, + 0.548 + ], + "angle": 0, + "content": "Visualization of learned frequency features. To show why frequency information is effective for the ZSAD task, we visualize the frequency-aware features from the last block of the visual encoder of CLIP as example, where the FFE and LFS adapters are added. As shown in Figure 5, both the features \\( f_{N,ffe}^{m} \\) from FFE adapter and the features \\( f_{N,lfs}^{m} \\) from LFS adapter highlight the abnormal region of images. By injecting the features \\( f_{N,ffe}^{m} \\) and \\( f_{N,lfs}^{m} \\) into the visual encoder of CLIP, the obtained features \\( \\hat{f}_N^m \\) can successfully localize the main abnormal region, in favor of the final prediction of the model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.55, + 0.906, + 0.702 + ], + "angle": 0, + "content": "When frequency information can be beneficial. Some properties of anomaly area, such as appearance defects, can be reflected based on its frequency information, since high-frequency components represent fine details (e.g. soft borders) that often exist in anomaly area. In these cases, These specific frequency patterns could be beneficial for anomaly detection. For example, as shown in Figure 1, the frequency statistics in the red region are different from the frequency statistics in the surrounding regions of the input abnormal image and the corresponding region of the normal image." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.703, + 0.906, + 0.839 + ], + "angle": 0, + "content": "Let's consider images of normal dumplings and images of squashed dumplings, where squashed dumplings are anomalies. In this case, if the squashed dumplings reveal the meat filling, the frequency of squashed dumplings could still stand out and frequency information could be helpful to detect anomalies. However, if the dumplings are squashed to be more like a meat pie without revealing the meat filling, the frequency of this squashed dumpling might not stand out and frequency information might not work well." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.901 + ], + "angle": 0, + "content": "We validate the FE-CLIP across different domains with 10 datasets of highly diverse class semantics from various industrial inspections and medical domains, and the results of FE-CLIP on the datasets of medical domains are trained" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21226" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.207, + 0.088, + 0.794, + 0.41 + ], + "angle": 0, + "content": "
SetupMethodsMVTec ADVisA
Pixel-levelImage-levelPixel-levelImage-level
1-shotPatchCore [37]92.0 / 79.783.4 / 92.295.4 / 80.579.9 / 82.8
WinCLIP [17]95.2 / 87.193.1 / 96.596.4 / 85.183.8 / 85.1
AnomalyGPT [14]95.3 / -94.1 / -96.2 / -87.4 / -
PromptAD [24]95.9 / -94.6 / -96.7 / -86.9 / -
FE-CLIP96.1 / 90.995.6 / 98.197.9 / 93.988.7 / 90.5
2-shotPatchCore [37]93.3 / 82.386.3 / 93.896.1 / 82.681.6 / 84.8
WinCLIP [17]96.0 / 88.494.4 / 97.096.8 / 86.284.6 / 85.8
AnomalyGPT [14]95.6 / -95.5 / -96.4 / -88.6 / -
PromptAD [24]96.2 / -95.7 / -97.1 / -88.3 / -
InCTRL [53]- / -94.0 / 96.9- / -85.8 / 87.7
FE-CLIP96.5 / 91.596.2 / 98.398.2 / 94.390.3 / 91.6
4-shotPatchCore [37]94.3 / 84.388.8 / 94.596.8 / 84.985.3 / 87.5
WinCLIP [17]96.2 / 89.095.2 / 97.397.2 / 87.687.3 / 88.8
AnomalyGPT [14]96.2 / -96.3 / -96.7 / -90.6 / -
PromptAD [24]96.5 / -96.6 / -97.4 / -89.1 / -
InCTRL [53]- / -94.5 / 97.2- / -87.7 / 90.2
FE-CLIP96.7 / 91.796.6 / 98.498.4 / 95.090.7 / 91.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.42, + 0.908, + 0.464 + ], + "angle": 0, + "content": "Table 5. Compared with other competing methods under few-normal-shot anomaly detection setting. The metric of Pixel-level is AUROC / PRO for the few-normal-shot anomaly segmentation task, and the metric of Image-level is AUROC / AP for the few-normal-shot anomaly detection task. Best results are highlighted in red." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.481, + 0.487, + 0.558 + ], + "angle": 0, + "content": "by the industrial inspection MVTec AD dataset, as shown in Table 1 and Table 2. The superior performance of FE-CLIP demonstrates that frequency information could be beneficial for many cases of anomaly detection, and the proposed FE-CLIP has good generalization across different domains." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.573, + 0.384, + 0.589 + ], + "angle": 0, + "content": "4.6. FE-CLIP with Few Normal Shots" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.597, + 0.483, + 0.839 + ], + "angle": 0, + "content": "The performance of FE-CLIP can also be further boosted when equipped with few-shot normal images, just like WinCLIP [17] and AnomalyGPT [14]. We adopt the same way as AnomalyGPT to compute the anomaly segmentation map with few-shot normal images. In short words, we use the same visual encoder of FE-CLIP to extract intermediate patch-level features from normal images and store them in memory banks, then calculate the cosine similarity distance between each patch token of input image and its most similar counterpart in the memory banks to obtain the anomaly segmentation map \\( S_{a,few\\_shot} \\). Finally, we add the anomaly segmentation map \\( S_{a} \\) from the FE-CLIP model with \\( S_{a,few\\_shot} \\) to get the final anomaly segmentation map, and add the anomaly score \\( M_{a} \\) from the FE-CLIP model with the max value of \\( S_{a,few\\_shot} \\) to get the final anomaly score. Please refer to [14] for more technical details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We compare the FE-CLIP with other competing methods, such as AnomalyGPT [14], PromptAD [24], and InCTRL [53], under the few-shot anomaly detection setting. As shown in Table 5, the FE-CLIP outperforms all other com" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.481, + 0.907, + 0.558 + ], + "angle": 0, + "content": "peting methods across different metrics, including AUROC and PRO metrics for the anomaly detection task, and AUROC and AP metrics for the anomaly segmentation task, on both datasets. The experiments under few-normal-shot setting further demonstrate the superiority of the FE-CLIP." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.58, + 0.634, + 0.595 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.608, + 0.907, + 0.851 + ], + "angle": 0, + "content": "In this paper, we find that frequency information can benefit the ZSAD task. To this end, we propose FE-CLIP, which adds the proposed FFE and LFS adapters into the visual encoder of CLIP, to deeply mine the frequency patterns for the visual features. Extensive experiments on 10 datasets of highly diverse class semantics from various defect inspections and medical domains demonstrate that (1) the frequency information can benefit the ZSAD task, and (2) the frequency information can help the model have good generalization across datasets from different domains. We also conduct experiments under the few-normal-shot anomaly detection setting to further validate the superiority of the proposed FE-CLIP. We hope the community of anomaly detection will be aware of the effectiveness of frequency information from our work and explore more ways to utilize the frequency information for anomaly detection in the future." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.855, + 0.682, + 0.871 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.873, + 0.907, + 0.903 + ], + "angle": 0, + "content": "This work was supported by the Anhui Provincial Science and Technology Major Project (No. 2023z020006)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "21227" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.156 + ], + "angle": 0, + "content": "[1] Nasir Ahmed, T. Natarajan, and Kamisetty R Rao. Discrete cosine transform. IEEE transactions on Computers, 100(1): 90-93, 1974. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.158, + 0.484, + 0.24 + ], + "angle": 0, + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.242, + 0.483, + 0.311 + ], + "angle": 0, + "content": "[3] Toshimichi Aota, Lloyd Teh Tzer Tong, and Takayuki Okatani. Zero-shot versus many-shot: Unsupervised texture anomaly detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5564-5572, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.312, + 0.483, + 0.382 + ], + "angle": 0, + "content": "[4] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad-a comprehensive real-world dataset for unsupervised anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9592-9600, 2019. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.382, + 0.483, + 0.452 + ], + "angle": 0, + "content": "[5] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Uninformed students: Student-teacher anomaly detection with discriminative latent embeddings. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4183–4192, 2020. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.453, + 0.483, + 0.535 + ], + "angle": 0, + "content": "[6] Jorge Bernal, F Javier Sánchez, Gloria Fernández-Esparrach, Debora Gil, Cristina Rodríguez, and Fernando Vilarino. Wm-dova maps for accurate polyp highlighting in colonoscopy: Validation vs. saliency maps from physicians. Computerized medical imaging and graphics, 43:99-111, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.536, + 0.483, + 0.592 + ], + "angle": 0, + "content": "[7] Tri Cao, Jiawen Zhu, and Guansong Pang. Anomaly detection under distribution shift. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6511-6523, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.593, + 0.483, + 0.661 + ], + "angle": 0, + "content": "[8] Yunkang Cao, Jiangning Zhang, Luca Frittoli, Yuqi Cheng, Weiming Shen, and Giacomo Boracchi. Adaclip: Adapting clip with hybrid learnable prompts for zero-shot anomaly detection. In European Conference on Computer Vision, 2024. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.663, + 0.483, + 0.732 + ], + "angle": 0, + "content": "[9] Xuhai Chen, Yue Han, and Jiangning Zhang. A zero-/few-shot anomaly classification and segmentation method for cvpr 2023 vand workshop challenge tracks 1&2: 1st place on zero-shot ad and 4th place on few-shot ad. arXiv preprint arXiv:2305.17382, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.733, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.873 + ], + "angle": 0, + "content": "[11] Tharindu Fernando, Harshala Gammulle, Simon Denman, Sridha Sridharan, and Clinton Fookes. Deep learning for medical anomaly detection-a survey. ACM Computing Surveys (CSUR), 54(7):1-37, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Gabriel Goh, Nick Cammarata, Chelsea Voss, Shan Carter, Michael Petrov, Ludwig Schubert, Alec Radford, and Chris" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Olah. Multimodal neurons in artificial neural networks. Distill, 6(3):e30, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[13] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[14] Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Ming Tang, and Jinqiao Wang. Anomalygpt: Detecting industrial anomalies using large vision-language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1932-1940, 2024. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.904, + 0.29 + ], + "angle": 0, + "content": "[15] A. Hamada. Br35h: Brain tumor detection 2020. In https://www.kaggle.com/datasets/ahmedhamada0/brain-tumor-detection, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.293, + 0.905, + 0.348 + ], + "angle": 0, + "content": "[16] Chaoqin Huang, Haoyan Guan, Aofan Jiang, Ya Zhang, Michael Spratling, and Yan-Feng Wang. Registration based few-shot anomaly detection. In European Conference on Computer Vision, pages 303-319. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.35, + 0.905, + 0.433 + ], + "angle": 0, + "content": "[17] Jongheon Jeong, Yang Zou, Taewan Kim, Dongqing Zhang, Avinash Ravichandran, and Onkar Dabeer. Winclip: Zero/few-shot anomaly classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19606-19616, 2023. 1, 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.435, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[18] Stepan Jezek, Martin Jonak, Radim Burget, Pavel Dvorak, and Milos Skotak. Deep learning-based defect detection of metal parts: evaluating current methods in complex conditions. In 2021 13th International congress on ultra modern telecommunications and control systems and workshops (ICUMT), pages 66-71. IEEE, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[19] Debesh Jha, Pia H Smedsrud, Michael A Riegler, Pål Halvorsen, Thomas De Lange, Dag Johansen, and Håvard D Johansen. Kvasir-seg: A segmented polyp dataset. In MultiMedia Modeling: 26th International Conference, MMM 2020, Daejeon, South Korea, January 5–8, 2020, Proceedings, Part II 26, pages 451–462. Springer, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.686 + ], + "angle": 0, + "content": "[20] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.688, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[21] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.759, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[22] Shengze Li, Jianjian Cao, Peng Ye, Yuhan Ding, Chongjun Tu, and Tao Chen. Clipsam: Clip and sam collaboration for zero-shot anomaly segmentation. arXiv preprint arXiv:2401.12665, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[23] Xiaoya Li, Xiaofei Sun, Yuxian Meng, Junjun Liang, Fei Wu, and Jiwei Li. Dice loss for data-imbalanced nlp tasks. arXiv preprint arXiv:1911.02855, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] Xiaofan Li, Zhizhong Zhang, Xin Tan, Chengwei Chen, Yanyun Qu, Yuan Xie, and Lizhuang Ma. Promptad: Learning prompts with only normal samples for few-shot anomaly" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21228" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.483, + 0.191 + ], + "angle": 0, + "content": "[25] Yuanze Li, Haolin Wang, Shihao Yuan, Ming Liu, Debin Zhao, Yiwen Guo, Chen Xu, Guangming Shi, and Wangmeng Zuo. Myriad: Large multimodal model by applying vision experts for industrial anomaly detection. arXiv preprint arXiv:2310.19070, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.483, + 0.248 + ], + "angle": 0, + "content": "[26] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.483, + 0.332 + ], + "angle": 0, + "content": "[27] Jie Liu, Yixiao Zhang, Jie-Neng Chen, Junfei Xiao, Yongyi Lu, Bennett A Landman, Yixuan Yuan, Alan Yuille, Yucheng Tang, and Zongwei Zhou. Clip-driven universal model for organ segmentation and tumor detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21152-21164, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.483, + 0.389 + ], + "angle": 0, + "content": "[28] Philipp Liznerski, Lukas Ruff, Robert A Vandermeulen, Billy Joe Franks, Marius Kloft, and Klaus-Robert Müller. Explainable deep one-class classification. arXiv preprint arXiv:2007.01760, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.483, + 0.459 + ], + "angle": 0, + "content": "[29] Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, and Aniruddha Kembhavi. Unified-io: A unified model for vision, language, and multi-modal tasks. In The Eleventh International Conference on Learning Representations, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.483, + 0.532 + ], + "angle": 0, + "content": "[30] Pankaj Mishra, Riccardo Verk, Daniele Fornasier, Claudio Piciarelli, and Gian Luca Foresti. Vt-adj: A vision transformer network for image anomaly detection and localization. In 2021 IEEE 30th International Symposium on Industrial Electronics (ISIE), pages 01-06. IEEE, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.483, + 0.588 + ], + "angle": 0, + "content": "[31] Ziyuan Qin, Huahui Yi, Qicheng Lao, and Kang Li. Medical image understanding with pretrained vision language models: A comprehensive study. arXiv preprint arXiv:2209.15517, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.483, + 0.658 + ], + "angle": 0, + "content": "[32] Zhen Qu, Xian Tao, Mukesh Prasad, Fei Shen, Zhengtao Zhang, Xinyi Gong, and Guiguang Ding. Vcp-clip: A visual context prompting model for zero-shot anomaly segmentation. In European Conference on Computer Vision, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.483, + 0.743 + ], + "angle": 0, + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.745, + 0.483, + 0.828 + ], + "angle": 0, + "content": "[34] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18082-18091, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.83, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[35] Tal Reiss and Yedid Hoshen. Mean-shifted contrastive loss for anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2155-2162, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.908, + 0.135 + ], + "angle": 0, + "content": "synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.908, + 0.206 + ], + "angle": 0, + "content": "[37] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14318-14328, 2022. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.208, + 0.906, + 0.289 + ], + "angle": 0, + "content": "[38] Mohammadreza Salehi, Niousha Sadjadi, Soroosh Baselizadeh, Mohammad H Rohban, and Hamid R Rabiee. Multiresolution knowledge distillation for anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14902-14912, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.906, + 0.375 + ], + "angle": 0, + "content": "[39] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.906, + 0.445 + ], + "angle": 0, + "content": "[40] Rohan Taori, Achal Dave, Vaishaal Shankar, Nicholas Carlini, Benjamin Recht, and Ludwig Schmidt. Measuring robustness to natural distribution shifts in image classification. Advances in Neural Information Processing Systems, 33:18583-18599, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.906, + 0.516 + ], + "angle": 0, + "content": "[41] Yoad Tewel, Yoav Shalev, Idan Schwartz, and Lior Wolf. Zerocap: Zero-shot image-to-text generation for visual-semantic arithmetic. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17918-17928, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.518, + 0.906, + 0.63 + ], + "angle": 0, + "content": "[42] Yu Tian, Guansong Pang, Fengbei Liu, Yuanhong Chen, Seon Ho Shin, Johan W Verjans, Rajvinder Singh, and Gustavo Carneiro. Constrained contrastive distribution learning for unsupervised anomaly detection and localisation in medical images. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27-October 1, 2021, Proceedings, Part V 24, pages 128-140. Springer, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.631, + 0.906, + 0.7 + ], + "angle": 0, + "content": "[43] Yu Tian, Fengbei Liu, Guansong Pang, Yuanhong Chen, Yuyuan Liu, Johan W Verjans, Rajvinder Singh, and Gustavo Carneiro. Self-supervised pseudo multi-class pre-training for unsupervised anomaly detection and segmentation in medical images. Medical image analysis, 90:102930, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.702, + 0.906, + 0.758 + ], + "angle": 0, + "content": "[44] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.759, + 0.906, + 0.801 + ], + "angle": 0, + "content": "[45] Matthias Wieler and Tobias Hahn. Weakly supervised learning for industrial optical inspection. In DAGM symposium in, page 11, 2007. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.906, + 0.872 + ], + "angle": 0, + "content": "[46] Mengde Xu, Zheng Zhang, Fangyun Wei, Yutong Lin, Yue Cao, Han Hu, and Xiang Bai. A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In European Conference on Computer Vision, pages 736-753. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.906, + 0.902 + ], + "angle": 0, + "content": "[47] Zhiyuan You, Lei Cui, Yujun Shen, Kai Yang, Xin Lu, Yu Zheng, and Xinyi Le. A unified model for multi-class" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21229" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "anomaly detection. Advances in Neural Information Processing Systems, 35:4571-4584, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[48] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16793-16803, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[49] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In European Conference on Computer Vision, pages 696-712. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.303 + ], + "angle": 0, + "content": "[50] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.482, + 0.374 + ], + "angle": 0, + "content": "[51] Qihang Zhou, Shibo He, Haoyu Liu, Tao Chen, and Jiming Chen. Pull & push: Leveraging differential knowledge distillation for efficient unsupervised anomaly detection and localization. IEEE Transactions on Circuits and Systems for Video Technology, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[52] Qihang Zhou, Guansong Pang, Yu Tian, Shibo He, and Jiming Chen. Anomalyclip: Object-agnostic prompt learning for zero-shot anomaly detection. The thirteenth International Conference on Learning Representations, 2024. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.482, + 0.488 + ], + "angle": 0, + "content": "[53] Jiawen Zhu and Guansong Pang. Toward generalist anomaly detection via in-context residual learning with few-shot sample prompts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.482, + 0.558 + ], + "angle": 0, + "content": "[54] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. In European Conference on Computer Vision, pages 392-408. Springer, 2022. 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.558 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "21230" + } + ] +] \ No newline at end of file diff --git a/2025/FE-CLIP_ Frequency Enhanced CLIP Model for Zero-Shot Anomaly Detection and Segmentation/e20fdad0-0c02-434f-a4f5-0e729b38aab0_origin.pdf b/2025/FE-CLIP_ Frequency Enhanced CLIP Model for Zero-Shot Anomaly Detection and Segmentation/e20fdad0-0c02-434f-a4f5-0e729b38aab0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..711d46319baba0d522a78337c7ec1e6c7441d7e8 --- /dev/null +++ b/2025/FE-CLIP_ Frequency Enhanced CLIP Model for Zero-Shot Anomaly Detection and Segmentation/e20fdad0-0c02-434f-a4f5-0e729b38aab0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:935cd259c07bd745260e1b7d9bfc25470a9bd91e8fad2571f51e5ed497887923 +size 458215 diff --git a/2025/FE-CLIP_ Frequency Enhanced CLIP Model for Zero-Shot Anomaly Detection and Segmentation/full.md b/2025/FE-CLIP_ Frequency Enhanced CLIP Model for Zero-Shot Anomaly Detection and Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..674d3cc467c127e52150eca08dfea28c31b7fc46 --- /dev/null +++ b/2025/FE-CLIP_ Frequency Enhanced CLIP Model for Zero-Shot Anomaly Detection and Segmentation/full.md @@ -0,0 +1,312 @@ +# FE-CLIP: Frequency Enhanced CLIP Model for Zero-Shot Anomaly Detection and Segmentation + +Tao Gong $^{1,2,3}$ , Qi Chu $^{1,2,3*}$ , Bin Liu $^{1,2,3}$ , Wei Zhou $^{4}$ , Nenghai Yu $^{1,2,3}$ + +1School of Cyber Science and Technology, University of Science and Technology of China + +2Anhui Province Key Laboratory of Digital Security + +3the CCCD Key Lab of Ministry of Culture and Tourism + +$^{4}$ Ling Yang Industrial Internet Co., Ltd. + +{tgong, qchu, flowice, ynh} @ustc.edu.cn, weizhou8@iflytek.com + +# Abstract + +Zero-shot anomaly detection (ZSAD) requires detection models trained using auxiliary data to detect anomalies without any training sample in a target dataset. It is challenging since the models need to generalize to anomalies across different domains. Recently, CLIP-based anomaly detection methods, such as WinCLIP and AnomalyCLIP, have demonstrated superior performance in the ZSAD task, due to the strong zero-shot recognition of the CLIP model. However, they overlook the utilization of frequency information of images. In this paper, we find that frequency information could benefit the ZSAD task, since some properties of the anomaly area, such as appearance defects, can also be reflected based on its frequency information. To this end, We propose Frequency Enhanced CLIP (FE-CLIP), taking advantage of two different but complementary frequency-aware clues, (1) Frequency-aware Feature Extraction adapter, and (2) Local Frequency Statistics adapter, in the visual encoder of CLIP, to deeply mine frequency information for the ZSAD task. We apply DCT as the frequency-domain transformation. Through comprehensive experiments, we show that the proposed FE-CLIP has good generalization across different domains and achieves superior zero-shot performance of detecting and segmenting anomalies in 10 datasets of highly diverse class semantics from various defect inspections and medical domains. Besides, FE-CLIP also achieves superior performance under the few-normal-shot anomaly detection settings. + +# 1. Introduction + +Anomaly Detection aims to predict an image or a pixel as normal or anomalous, and has been widely applied in + +various applications, such as industrial defect inspection [4, 5, 7, 16, 28, 35, 37, 47, 51] and medical image analysis [11, 27, 31, 42, 43]. Existing AD approaches typically assume that training examples in a target application domain are available for learning the detection models. However, this assumption may not hold in various scenarios, such as i) when accessing training data violates data privacy policies, or ii) when the target domain does not have relevant training data. Therefore, Zero-shot anomaly detection (ZSAD) is an emerging task for AD in such scenarios, to which the AD methods mentioned above are not viable, as it requires detection models to detect anomalies without any training sample in a target dataset. ZSAD is challenging since the models need to generalize to anomalies across different domains where the appearance of foreground objects, abnormal regions, and background features can vary significantly. + +Recently, large pre-trained vision-language models (VLMs) have demonstrated strong zero-shot recognition ability in various vision tasks. Particularly, being pretrained using millions/billions of image-text pairs, CLIP [33] has been applied to empower various downstream tasks with its strong generalization capability. WinCLIP [17] is the first work to apply the CLIP model in the ZSAD line. It uses a large number of hand-crafted text prompts and extraction of window/patch/image-level features aligned with text to finish the ZSAD task. However, WinCLIP involves multiple forward passes of image patches for anomaly segmentation, which leads to inefficiency. Besides, WinCLIP only performs the testing of the CLIP model in the ZSAD without any fine-tuning using the anomaly detection task data, which also limits its performance in the ZSAD task. Subsequently, several works [8, 22, 32, 52] attempt to adapt the CLIP into the ZSAD task by fintuning. Specifically, AnomalyCLIP [52] learns object-agnostic text prompts that capture generic normality and abnormality in an image regardless of its foreground objects. ClipSAM [22] introduces + +![](images/1c26cd97864a68e083005162dfc3547c7765e45f23d3c57b6e61b7a48aa42a10.jpg) + +![](images/9d558e64152f6531e8db98707af6e0a4f628d3d575aee100ae81214939f2f93e.jpg) + +![](images/481e6f3295a2142aeb98cd2466fb958c8eb1f96e00ab27f6e11da1e0a5400fcb.jpg) + +![](images/e0912137d64943e50914be6d383485a7da6123f6874c46937667ee4e5b732b41.jpg) + +![](images/624acac42e6ce9861fb5a31243faba7f9dca38d1893b808927d4a803f61fb3d6.jpg) +(a) normal + +![](images/1d0eda1e2bd73fd2b08f254c32a05142fe642f32d8b33e44f39c0fe41c77365b.jpg) +(b) hole + +![](images/43a1f479838ef21d8c8f7fbcfdcdee60401235314c45aa66c30cbcd9d46cca12.jpg) +(c) crack +Figure 1. The local frequency statistics of images. The second row represents the local frequency statistics of non-overlapped image regions. There are four columns, standing for the normal object, object with a hole, object with a crack, and object with print, respectively. Please zoom in and pay attention to the red region (i.e. the anomaly region of the input image) for a better view. + +![](images/c6e47206040e53189505e927df62368744ab12c30d1fdec37db0117a4990442e.jpg) +(d) print + +a unified multiscale cross-modal interaction module for interacting language with visual features at multiple stages of the CLIP visual encoder to reason anomaly positions. VCP-CLIP [32] propose the Pre-VCP module and PostVCP module to perform visual context prompting to activate CLIP's anomalous semantic perception ability. AdaCLIP [8] incorporates static and dynamic learnable prompts to adapt CLIP for ZSAD. However, all of the methods mentioned above overlook the utilization of frequency information of images. + +In this paper, we find that frequency information could benefit the ZSAD task. Some properties of the anomaly area, such as appearance defects, can also be reflected based on its frequency information, since high-frequency components represent fine details (e.g. soft borders) that often exist in the anomaly area. For example, as shown in Figure 1, the frequency statistics in the red region are different from the frequency statistics in the surrounding regions of the input abnormal image and the corresponding region of the normal image. These specific frequency patterns could be helpful for anomaly detection. + +To this end, we introduce a novel approach, namely the Frequency Enhanced CLIP (FE-CLIP), to inject frequency information into CLIP for accurate ZSAD across different domains. FE-CLIP takes advantage of two different but complementary frequency-aware clues, (1) Frequency-aware Feature Extraction (FFE) adapter, and (2) Local Frequency Statistics (LFS) adapter, in the visual encoder of CLIP, to deeply mine frequency information. Specifically, the FFE adapter uses DCT [1] to transform the image features into the frequency domain in order to mine the frequency information, then uses inverse DCT to transform back to the spatial domain. The FFE adapter describes the frequency-aware patterns in the spatial domain, but does not explicitly render the frequency information directly in + +the neural networks. Therefore, we use LFS to render the frequency information directly in the frequency domain. In each densely but regularly sampled local spatial patch, the statistics are gathered by counting the mean frequency responses. The averaged frequency statistics are a multichannel spatial map, where the number of channels is identical to the number of channels in the spatial domain. The local frequency statistics also follow the spatial layouts as the features of the CLIP visual encoder, thus also enjoying effective representation learning powered by the CLIP visual encoder. Meanwhile, since the frequency-aware feature extraction and local frequency statistics are complementary to each other but both of them share inherently similar frequency-aware semantics, thus they can be progressively fused during the feature learning process. + +Through comprehensive experiments, we show that the proposed FE-CLIP has good generalization across different domains and achieves superior zero-shot performance of detecting and segmenting anomalies in 10 datasets of highly diverse class semantics from various defect inspections and medical domains. Our contributions in this paper are summarized as follows: + +1. We find that frequency information could benefit the ZSAD task, and propose a novel Frequency Enhanced CLIP (FE-CLIP) method to inject frequency information into CLIP for accurate ZSAD across different domains. +2. We propose two different but complementary frequency-aware adapters, Frequency-aware Feature Extraction (FFE) adapter, and Local Frequency Statistics (LFS) adapter, in the visual encoder of CLIP, to deeply mine the frequency patterns of anomaly region. +3. Extensive experiments show that the proposed FE-CLIP has good generalization across different domains and achieves superior zero-shot performance of detecting and segmenting anomalies in 10 datasets of highly diverse class semantics from various defect inspections and medical domains. Besides, the proposed FE-CLIP also achieves superior performance under the few-normal-shot anomaly detection settings. + +# 2. Related Work + +# 2.1. Vision-language modeling + +Among recent successes of large pre-trained vision-language models (VLM) [2, 20, 33], CLIP [33] is the first to perform pre-training on web-scale image-text data, showing unprecedented generality: e.g., its language-driven zero-shot inference, improved both effective robustness [40] and perceptual alignment [12]. Many following VLM works explored large-scale pre-training in different aspects, e.g., scaling up data [20], efficient designs [21], multi-tasks [29], etc. To democratize large-scale VLM for the usages in different domains, a billion-scale data LAION5B [39], a + +![](images/aa8fb52490b307caa182158c166aa25e1db21b01264cdc8cc19722b5bec154fb.jpg) +Figure 2. The framework of the proposed FE-CLIP. The text encoder and visual encoder are from the CLIP model, and the FFE and LFS are the proposed adapters. We only show the class token embedding after the last block of the CLIP visual encoder for convenience. Actually, the CLIP visual encoder can output the class token after every block. + +code base of OpenCLIP with pre-trained models are open-sourced. Other works presented CLIP's promise in zero/few-shot transfer to downstream tasks beyond classification [13, 36, 41, 46]. Good prompt engineering and tuning [33, 50] can non-trivially benefit generalization performances. Moreover, some other works [34, 48, 49] leverage the pre-trained CLIP for language-guided detection and segmentation with promising performances. In this paper, we mainly focus on adapting CLIP into the ZSAD task with superior performance by injecting the frequency information into the visual encoder of CLIP. + +# 2.2. Zero-shot Anomaly detection + +ZSAD relies on the model's strong transferability to handle unseen anomalies [3]. A very recent approach WinCLIP [17] presents a seminal work that leverages CLIP [33] for zero-shot anomaly classification and segmentation. WinCLIP proposes a compositional ensemble on state words and prompt templates and efficient extraction and aggregation of window/patch/image-level features aligned with text to finish the ZSAD task. To tackle this inefficiency, APRILGAN [9] introduces a fine-tuning strategy by adding extra learnable linear projection layers in multiple stages of the CLIP visual encoder to enhance the modeling of local visual semantics. AnomalyCLIP [52] learns object-agnostic text prompts that capture generic normality and abnormality in an image regardless of its foreground objects. ClipSAM [22] introduces a unified multiscale cross-modal interaction module for interacting language with visual features at multiple stages of the CLIP visual encoder to reason anomaly + +positions. Myriad [25] and AnomalyGPT [14] utilize large language models for zero-shot/few-shot anomaly detection. They employ an image decoder to provide fine-grained semantics and design a prompt learner to fine-tune the large language models using prompt embeddings. However, all of the methods mentioned above overlook the utilization of frequency information of images. In this paper, we find that frequency information could benefit the ZSAD task, and propose the FE-CLIP method which utilizes two different but complementary adapters, the FFE adapter and the LFS adapter, into the visual encoder of CLIP, to deeply mine the frequency information for ZSAD task. + +# 3. Method + +# 3.1. Preliminary + +CLIP [33] is a large-scale pretraining method offering a joint vision-language representation. It consists of a text encoder $T(\cdot)$ and a visual encoder $F(\cdot)$ . Both encoders are mainstream multi-block networks such as ViT [10]. Using text prompts is a typical way to acquire the embeddings of different classes for zero-shot classification. The text prompts usually combine a text prompt template $\Omega$ with the category name $c$ , where the text prompt template $\Omega$ usually is defined as A photo of a [cls] and [cls] represents the target category name. The text prompts are passed through $T(\cdot)$ to obtain its corresponding textual embedding $g_{c} \in \mathbb{R}^{D}$ where $D$ denotes the channel dimension. Then, an input image is passed through $F(\cdot)$ to get the visual representations, where the class token $f \in \mathbb{R}^{D}$ is the + +global visual embedding of the input image, and patch tokens $f^{m} \in \mathbb{R}^{H \times W \times D}$ are treated as local visual embeddings of the input image. $H$ and $W$ denote the height and width of the patch tokens, respectively. CLIP can perform zero-shot recognition by measuring the similarity between textual and visual embeddings. Specifically, given a target class set $\mathcal{C}$ and an image, CLIP predicts the probability of the image belonging to the $c$ category as follows: + +$$ +p (y = c) = P \left(g _ {c}, f\right) = \frac {\exp \left(< g _ {c} , f > / \tau\right)}{\sum_ {c \in \mathcal {C}} \exp \left(< g _ {c} , f >\right) / \tau} \tag {1} +$$ + +where $\tau$ denotes the temperature and the operator $< ,\cdot ,>$ represents the computation of cosine similarity. The computation can be extended from global visual embeddings to local visual embeddings to derive the corresponding segmentation maps $M_{c}\in \mathbb{R}^{H\times W}$ , if each entry $(i,j)$ is computed as $P\left(g_c,f^{m(i,j)}\right)$ + +# 3.2. Frequency Enhanced CLIP + +In this paper, we find that frequency information could benefit the ZSAD task, and propose a novel Frequency Enhanced CLIP (FE-CLIP) method to inject frequency information into CLIP [33] for accurate ZSAD across different domains. We inject the frequency information by the features of the CLIP visual encoder rather than the frequency of the input image, since the CLIP visual encoder is trained to take RGB image as input. Typically, the CLIP visual encoder comprises a series of block layers. From the bottom to the top of the layers, the visual encoder gradually learns the visual patterns at different levels of abstraction. As shown in Figure 2, we add Frequency-aware Feature Extraction (FFE) adapter and Local Frequency Statistics (LFS) adapter across multiple blocks levels in the visual encoder $F(\cdot)$ of CLIP, while keeping its original backbone unchanged, thus enabling frequency information injection at multiple feature levels. To be specific, assuming the visual encoder consists of $N$ blocks ( $N = 4$ ), the features (i.e. the patch tokens) after the $n$ -th block are denoted as $f_{n}^{m} \in \mathbb{R}^{H \times W \times D}$ , and the class token after the $n$ -th block is denoted as $f_{n} \in \mathbb{R}^{D}$ , where $n = 1, 2,.., N$ . Firstly, the features $f_{n}^{m}$ are passed through an FFE adapter and an LFS adapter to get the features $f_{n,ffe}^{m}$ and $f_{n,lfs}^{m}$ , respectively, which contain the frequency information. We can get the frequency-aware feature $\hat{f}_{n}^{m}$ as follows: + +$$ +\hat {f} _ {n} ^ {m} = \lambda (f _ {n, f f e} ^ {m} + f _ {n, l f s} ^ {m}) + (1 - \lambda) f _ {n} ^ {m} \qquad (2) +$$ + +with $\hat{f}_n^m$ serving as the input for the next block. The class token $f_n$ will also be injected with frequency information after the next block due to the self-attention mechanism [44] in the blocks. We set $\lambda = 0.1$ to preserve the original knowledge of the CLIP visual encoder. Then, the normal text prompt and abnormal text prompt are passed through + +![](images/ada198cb4923503e658aeebf6adf459bb11b39ce42a82b39c983defa58a9c7c5.jpg) +Figure 3. The proposed Frequency-aware Feature Extraction (FFE) adapter. + +text encoder $T(\cdot)$ to get corresponding textual embedding $g_{r}$ and $g_{a}$ , respectively. $g_{r}$ and $g_{a}$ constitute the target class set $\mathcal{C}$ . We simply use A photo of a normal object and A photo of a damaged object for the normal text prompt and abnormal text prompt, respectively, rather than the complex compositional ensemble on state words and prompt templates used in WinCLIP [17], since we find that the simple text prompt also works well. Finally, we compute the abnormal segmentation map $M_{a,n} \in \mathbb{R}^{H \times W}$ based on the equation 1, where each entry $(i,j)$ is computed as $P\left(g_{a}, fc(\hat{f}_{n}^{m(i,j)})\right)$ . The $fc$ means that we use a single learnable fc to align the dimension of visual features $\hat{f}_{n}^{m}$ to the dimension of text features. The abnormal score $S_{a,n}$ is also computed based on the equation 1 as $P\left(g_{a}, fc_{clip}(f_{n})\right)$ . The $fc_{clip}$ denotes the frozen project layer of the visual encoder in CLIP to align the dimension of class token $f_{n}$ and the dimension of text features. The final prediction of the abnormal score $S_{a}$ and abnormal segmentation map $M_{a}$ are computed by averaging $S_{a,n}$ and $M_{a,n}$ across $N$ blocks, respectively, as follows: + +$$ +S _ {a} = \frac {1}{N} \sum_ {n = 1} ^ {N} S _ {a, n}, \quad M _ {a} = \frac {1}{N} \sum_ {n = 1} ^ {N} M _ {a, n} \tag {3} +$$ + +# 3.3. Frequency-aware Feature Extraction Adapter + +The specific frequency patterns could be beneficial for anomaly detection. To mine the frequency information, we propose the Frequency-aware Feature Extraction (FFE) adapter in the visual encoder of CLIP. Specifically, given the patch tokens of the $n$ -th block $f_{n}^{m}\in \mathbb{R}^{H\times W\times D}$ , we first split the $f_{n}^{m}$ into non-overlapped windows with total $(H / P)\times (W / P)$ windows and each window containing $P\times P$ patch tokens ( $P$ is set to 3 by default). As shown in Figure 3, we apply DCT [1] transformation to each window to extract the DCT features $f_{n,ffe,1}^{m}$ which represents the frequency information. Then, we pass $f_{n,ffe,1}^{m}$ through one linear layer followed by one GELU( $\cdot$ ) activation layer to further extract the frequency features $f_{n,ffe,2}^{m}$ . Finally, we apply the inverse DCT transformation to $f_{n,ffe,2}^{m}$ to get the frequency-aware features $f_{n,ffe}^{m}$ in the spatial domain. The FFE adapter describes the frequency-aware patterns in the spatial domain, but does not explicitly render the frequency information directly in the neural networks. In the next sub-section, we propose an adapter that renders the fre + +![](images/5e04a0e617d3013bf98a724038f9b30a721e7183e7309f2327e0716c29a434c0.jpg) +Figure 4. The proposed Local Frequency Statistics (LFS) adapter. + +quency information directly in the neural networks. + +# 3.4. Local Frequency Statistics Adapter + +The aforementioned FFE adapter has provided frequency-aware representation, but it has to represent the frequency-aware patterns back into the spatial domain, thus failing to directly utilize the frequency patterns in the frequency domain. Therefore, we propose the Local Frequency Statistics (LFS) adapter in the visual encoder of CLIP to render the frequency patterns directly in the frequency domain. As shown in Figure 4, given the patch tokens of the $n$ -th block $f_{n}^{m}\in \mathbb{R}^{H\times W\times D}$ , we first apply a Sliding Window DCT (SW DCT) on $f_{n}^{m}$ (i.e., taking DCT [1] densely on sliding windows of the features) to extract the localized frequency responses $f_{n,lfs,1}^{m}$ . The size of sliding windows is $Q\times Q$ ( $Q$ is set to 3 by default). Therefore, there are $Q\times Q$ group frequency responses in $f_{n,lfs,1}^{m}$ , where the first one represents the low-frequency responses and the last one represents the high-frequency responses. Then we count the mean of $f_{n,lfs,1}^{m}$ across $Q\times Q$ groups to get the mean frequency responses $f_{n,lfs,2}^{m}$ . The $f_{n,lfs,2}^{m}$ is a multi-channel spatial map that shares the same layout as $f_{n}^{m}$ . The LFS adapter outputs $f_{n,lfs}^{m}$ after $f_{n,lfs,2}^{m}$ being passed through a convolutional layer followed by one $GELU(\cdot)$ activation layer. The LFS adapter provides a localized aperture to detect detailed abnormal frequency distributions. + +# 3.5. Loss Functions + +Let $S_{gt}$ denote the ground-truth image label where 1 represents the abnormal image and 0 represents the normal image. The loss for the anomaly classification is defined as follows: + +$$ +L _ {c l s} = - \frac {1}{N} \sum_ {N} ^ {n = 1} \left[ S _ {g t} \log \left(S _ {a, n}\right) + \left(1 - S _ {g t}\right) \log \left(1 - S _ {a, n}\right) \right] \tag {4} +$$ + +where $\log$ denotes the log operation. Similarly, let $M_{gt}$ denote the ground-truth pixel label of the input image, where 1 means abnormal pixel region and 0 means normal pixel region. The loss for the anomaly segmentation is defined as follows: + +$$ +L _ {\text {m a s k}} = \frac {1}{N} \sum_ {N} ^ {n = 1} [ \operatorname {F o c a l} (U p (M _ {a, n}), M _ {g t}) + D i c e (U p (M _ {a, n}), M _ {g t}) ] \tag {5} +$$ + +where $Focal(\cdot, \cdot)$ and $Dice(\cdot, \cdot)$ denote the focal loss [26] and dice loss [23], respectively. The operator $Up(\cdot)$ represents the unsampling operation. During training, both the visual encoder and the text encoder of CLIP [33] are Frozen. Only the FFE adapters and LFS adapters are optimized by the loss function $L_{total} = L_{cls} + L_{mask}$ . + +# 4. Experiments + +# 4.1. Datasets and Evaluation Metrics + +To verify the effectiveness of the proposed method FE-CLIP, we conduct comprehensive experiments across 10 real-world anomaly detection datasets, covering various industrial inspection scenarios and medical imaging domains. In industrial inspection, we use MVtec AD [4], VisA [54], MPDD [18], BTAD [30], DAGM [45], and DTD-Synthetic [3]. In medical imaging, we use colon polyp detection datasets CVC-ClinicDB [6] and Kvasir [19], brain tumor detection datasets BrainMRI [38] and Br35H [15]. + +As for the evaluation metrics, following previous works [17, 52], we use Area Under the Receiver Operating Characteristic curve (AUROC). Additionally, Average Precision (AP) for Zero-Shot Anomaly Detection (ZSAD) and PRO [5] for Zeor-Shot Anomaly Segmentation (ZSAS) are also used to provide a more in-depth analysis of the results. + +# 4.2. Implementation Details + +We use the publicly available CLIP model (VITL/14@336px) as our backbone. Model parameters of CLIP are all frozen, and only the parameters of FFE adapters and LFS adapters are learnable. Following AnomalyCLIP [52], we fine-tune FE-CLIP using the test data of MVtec AD and evaluate the ZSAD performance on other datasets. As for MVtec AD, we fine-tune FE-CLIP on the test data of VisA. The proposed FE-CLIP is trained by 9 epochs with Adam optimizer. The learning rate is set to 5e-4 and the total batch size is 16. All experiments are conducted in PyTorch-1.13.0 with four NVIDIA RTX 3090 24GB GPUs. All experiment results are averaged under five runs. We report dataset-level results, which are averaged across their respective sub-datasets. + +# 4.3. Main Results + +ZSAD performance on diverse anomaly scenarios. Table 1 shows the ZSAD results of FE-CLIP with other competing methods over six industrial defect datasets and two medical domain datasets. As shown in Table 1, FE-CLIP can achieve higher AUROC and AP for the ZSAD task on most industrial inspection datasets and medical datasets, such as MVTec AD, DTD-Synthetic, and BrainMRI, when compared with other competing methods, such as WinCLIP [17], AnomalyCLIP [52] and AdaCLIP [8]. This shows that (1) the frequency information can benefit the ZSAD task + +
MethodsIndustrial DefectsMedical Anomalies
MVTec ADVisAMPDDBTADDAGMDTD-SyntheticBrainMRIBr35H
CLIP [33]74.1 / 87.666.4 / 71.554.3 / 65.434.5 / 52.579.6 / 59.071.6 / 85.773.9 / 81.778.4 / 78.8
CoOp [50]88.8 / 94.862.8 / 68.155.1 / 64.266.8 / 77.487.5 / 74.6- / -61.3 / 44.986.0 / 87.5
WinCLIP [17]91.8 / 96.578.1 / 81.263.6 / 69.968.2 / 70.991.8 / 79.593.2 / 92.686.6 / 91.580.5 / 82.2
APRIL-GAN [9]86.1 / 93.578.0 / 81.478.0 / 81.473.6 / 68.694.4 / 83.886.4 / 95.089.3 / 90.993.1 / 92.9
AnomalyCLIP [52]91.5 / 96.282.1 / 85.477.0 / 82.088.3 / 87.397.5 / 92.393.5 / 97.090.3 / 92.294.6 / 94.7
AdaCLIP [8]89.2 / -85.8 / -76.0 / -88.6 / -99.1 / -95.5 / -94.8 / -97.7 / -
FE-CLIP91.9 / 96.584.6 / 86.678.0 / 82.690.3 / 90.097.5 / 92.398.3 / 99.494.8 / 93.896.8 / 93.8
+ +Table 1. AUROC / AP results for Zero-Shot Anomaly Detection (ZSAD) task on eight real-world anomaly detection datasets. Best results are highlighted in red. + +
MethodsIndustrial DefectsMedical Anomalies
MVTec ADVisAMPDDBTADDAGMDTD-SyntheticCVC-ClinicDBKvasir
CLIP [33]38.4 / 11.346.6 / 14.862.1, 33.030.6 / 4.430.6 / 4.433.9 / 12.547.5 / 18.944.6 / 17.7
CoOp [50]33.3 / 6.724.2 / 3.815.4 / 2.315.4 / 2.317.5 / 2.1- / -17.5 / 2.144.1 / 3.5
WinCLIP [17]85.1 / 64.679.6 / 56.879.6 / 56.872.7 / 27.387.6 / 65.783.9 / 57.851.2 / 13.869.7 / 24.5
APRIL-GAN [9]87.6 / 84.094.2 / 86.894.1 / 83.260.8 / 25.082.4 / 66.295.3 / 86.980.5 / 60.775.0 / 36.2
AnomalyCLIP [52]91.1 / 81.495.5 / 87.096.5 / 88.794.2 / 74.895.6 / 91.097.9 / 92.382.9 / 67.878.9 / 45.6
AdaCLIP [8]88.7 / -95.5 / -96.1 / -92.1 / -91.5 / -97.9 / -84.4 / -- / -
VCP-CLIP [32]92.0 / 87.395.7 / 90.7- / -94.1 / 74.699.4 / 98.3- / -- / -- / -
FE-CLIP92.6 / 88.395.9 / 92.897.0 / 90.595.6 / 80.498.5 / 96.699.0 / 97.484.5 / 69.279.8 / 60.6
+ +Table 2. AUROC / PRO results for Zero-Shot Anomaly Segmentation (ZSAS) task on eight real-world anomaly detection datasets. Best results are highlighted in red. Note that the image-level medical AD datasets do not contain segmentation ground truth, so the pixel-level medical AD datasets are different from the image-level datasets. + +due to the superior performance of FE-CLIP, and (2) the frequency information helps FE-CLIP achieve good generalization across datasets from different domains. + +ZSAS performance on diverse anomaly scenarios. We show the ZSAS results of FE-CLIP with other competing methods over six industrial defect datasets and two medical domain datasets in Table 2. Table 2 shows that FE-CLIP achieves higher AUROC and PRO for the ZSAS task on almost all of the industrial inspection datasets and medical datasets, except for the DAGM dataset, when compared with other competing methods, such as WinCLIP [17], AnomalyCLIP [52], AdaCLIP [8] and VCP-CLIP [32]. This further demonstrates that (1) the frequency information can benefit the ZSAS task and (2) the frequency information helps FE-CLIP achieve good generalization across datasets from different domains. + +Discussion. We also note that the FE-CLIP can defeat almost all of the competing methods on the ZSAS task, while FE-CLIP only performs better than other methods over most of the datasets on ZSAD task. The reason may be that the frequency information can be directly injected into the patch tokens (i.e. the feature map used for the ZSAS task) through the FFE adapter and LFS adapter, while the frequency information can only be implicitly involved in the global class token (i.e. the feature vector used for the ZSAD task) from the patch tokens based on the self + +attention mechanism. + +Overall, the proposed FE-CLIP achieves superior zero-shot performance of detecting and segmenting anomalies in datasets of highly diverse class semantics from various defect inspection and medical imaging domains. + +# 4.4. Ablation Study + +The effectiveness of FFE and LFS We conduct ablation experiments on MVTec AD dataset and VisA dataset to validate the effectiveness of the proposed FFE adapter and LFS adapter. The results are shown in Table 3. The "CLIP $[33] + \text{Conv Adapter}"$ means we add $N$ conv adapters, where each of them consists of one convolutional layer followed by one $GELU(\cdot)$ activation layer to keep the same learnable parameters with the LFS adapter into the visual encoder of CLIP. It can be concluded that both the FFE adapter and LFS adapter can achieve better performance than the conv adapter in MVTec AD dataset and VisA dataset across the ZSAD task and ZSAS task. This also shows that the frequency information can benefit the ZSAD task and ZSAS task. When combined the FFE adapter with the LFS adapter, the FE-CLIP achieves higher performance than the FFE adapter and LFS adapter. This demonstrates that the FFE adapter and LFS adapter can mine different but complementary frequency information to further boost the performance of model on the ZSAD and ZSAS tasks. + +
MethodsMVTec ADVisA
Pixel-levelImage-levelPixel-levelImage-level
CLIP[33] + Conv Adapter89.0 / 85.187.8 / 94.594.8 / 89.882.3 / 83.5
CLIP[33] + FFE Adapter90.6 / 86.290.0 / 95.495.1 / 91.383.3 / 85.0
CLIP[33] + LFS Adapter91.5 / 87.190.6 / 95.795.5 / 91.883.9 / 85.7
FE-CLIP92.6 / 88.391.9 / 96.595.9 / 92.884.6 / 86.6
+ +Table 3. Module ablation on the MVTec AD and the VisA datasets. The metric of Pixel-level is AUROC / PRO for ZSAS task, and the metric of Image-level is AUROC / AP for ZSAD task. Best results are highlighted in red. + +
MethodsMVTec ADVisA
Pixel-levelImage-levelPixel-levelImage-level
FFT92.8 / 88.392.0 / 96.295.5 / 93.084.1 / 86.8
DCT92.6 / 88.391.9 / 96.595.9 / 92.884.6 / 86.6
+ +Table 4. Compared with using FFT to extract the frequency information in the FFE and LFS modules. The metric of Pixel-level is AUROC / PRO for ZSAS task, and the metric of Image-level is AUROC / AP for ZSAD task. Best results are highlighted in red. + +![](images/d6f13747d83cf192722772b460591f1b66576a4774d0cb3f468528e3c50c073d.jpg) + +![](images/19181edfeb38d426337f67abc575606f9e563ded07ed5ace7016579f93df73eb.jpg) +(a) Input Image + +![](images/919a719e30bba0e8afc8a81c79f7db0966c64af1ac9771609e1bf42911a92bc0.jpg) + +![](images/5131f2a5216916bc8b7df84d3705250873a75f7cc4c87f10966a1a98a7f44d48.jpg) +(b) $f_{fe}^{m}$ from FFE Adapter + +![](images/2d4704cb6283321213688f49e1fcf5b6c8a811acb9be816fdb5fea86e934c725.jpg) + +![](images/61cd967df442bba1adb5db52b48daa51e95ab22d3049bf2ad4684b7194503f66.jpg) +(c) $f_{N,lf_s}^m$ from LFS Adapter + +![](images/1aa7f5f88c38dc5ab12d1f85613f49cf78fae46ee08df725bb534ffe4d2c00ce.jpg) + +![](images/e7f85ba6cab1ad3ddce06249e23fbd3b3b9a739c90497e16a7968c548736d774.jpg) +(d) $\hat{f}_N^m$ from equation (2) +Figure 5. The visualization of the learned frequency features. Please zoom in for a better view. + +![](images/8dacbcc1c5ca34bb5c9e58ea87bc5dfde841ac5e0ddc160659a58225a04e3901.jpg) + +![](images/d9d43740d3f9c1df4bb794417b67fcfcae4ceb31aa44f2b19cc30e03997fc358.jpg) +(e) Final prediction + +Would FFT work instead of DCT We attempt to use FFT to replace the DCT in the FFE and LFS modules to extract the frequency information. However, there are real values and imaginary values after FFT. To deal with this, the real values and imaginary values are separately processed and then combined to transform back to the spatial domain in the FFE module. As for the LFS module, the averaged local frequency statistics are computed from the real values and imaginary values, respectively, to get the averaged real local frequency statistics and the averaged imaginary local frequency statistics, and then they are summed to get the final averaged local frequency statistics. The results are shown in Table 4. This shows that using FFT achieves similar results with DCT, which means that the key is to inject the frequency information rather than the way of frequency transformation. We choose DCT since DCT is more convenient to extract frequency. + +# 4.5. Analysis + +Time cost. We report the inference time of APRIL-GAN, AnomalyCLIP, VCP-CLIP, AdaCLIP, and FE-CLIP with PyTorch-1.13.0 on one NVIDIA RTX 3090 24GB GPU. The inference time is 105, 120, 128, 160, and $125\mathrm{ms}$ , respectively. FE-CLIP achieves a better trade-off between performance and speed. + +Visualization of learned frequency features. To show why frequency information is effective for the ZSAD task, we visualize the frequency-aware features from the last block of the visual encoder of CLIP as example, where the FFE and LFS adapters are added. As shown in Figure 5, both the features $f_{N,ffe}^{m}$ from FFE adapter and the features $f_{N,lfs}^{m}$ from LFS adapter highlight the abnormal region of images. By injecting the features $f_{N,ffe}^{m}$ and $f_{N,lfs}^{m}$ into the visual encoder of CLIP, the obtained features $\hat{f}_N^m$ can successfully localize the main abnormal region, in favor of the final prediction of the model. + +When frequency information can be beneficial. Some properties of anomaly area, such as appearance defects, can be reflected based on its frequency information, since high-frequency components represent fine details (e.g. soft borders) that often exist in anomaly area. In these cases, These specific frequency patterns could be beneficial for anomaly detection. For example, as shown in Figure 1, the frequency statistics in the red region are different from the frequency statistics in the surrounding regions of the input abnormal image and the corresponding region of the normal image. + +Let's consider images of normal dumplings and images of squashed dumplings, where squashed dumplings are anomalies. In this case, if the squashed dumplings reveal the meat filling, the frequency of squashed dumplings could still stand out and frequency information could be helpful to detect anomalies. However, if the dumplings are squashed to be more like a meat pie without revealing the meat filling, the frequency of this squashed dumpling might not stand out and frequency information might not work well. + +We validate the FE-CLIP across different domains with 10 datasets of highly diverse class semantics from various industrial inspections and medical domains, and the results of FE-CLIP on the datasets of medical domains are trained + +
SetupMethodsMVTec ADVisA
Pixel-levelImage-levelPixel-levelImage-level
1-shotPatchCore [37]92.0 / 79.783.4 / 92.295.4 / 80.579.9 / 82.8
WinCLIP [17]95.2 / 87.193.1 / 96.596.4 / 85.183.8 / 85.1
AnomalyGPT [14]95.3 / -94.1 / -96.2 / -87.4 / -
PromptAD [24]95.9 / -94.6 / -96.7 / -86.9 / -
FE-CLIP96.1 / 90.995.6 / 98.197.9 / 93.988.7 / 90.5
2-shotPatchCore [37]93.3 / 82.386.3 / 93.896.1 / 82.681.6 / 84.8
WinCLIP [17]96.0 / 88.494.4 / 97.096.8 / 86.284.6 / 85.8
AnomalyGPT [14]95.6 / -95.5 / -96.4 / -88.6 / -
PromptAD [24]96.2 / -95.7 / -97.1 / -88.3 / -
InCTRL [53]- / -94.0 / 96.9- / -85.8 / 87.7
FE-CLIP96.5 / 91.596.2 / 98.398.2 / 94.390.3 / 91.6
4-shotPatchCore [37]94.3 / 84.388.8 / 94.596.8 / 84.985.3 / 87.5
WinCLIP [17]96.2 / 89.095.2 / 97.397.2 / 87.687.3 / 88.8
AnomalyGPT [14]96.2 / -96.3 / -96.7 / -90.6 / -
PromptAD [24]96.5 / -96.6 / -97.4 / -89.1 / -
InCTRL [53]- / -94.5 / 97.2- / -87.7 / 90.2
FE-CLIP96.7 / 91.796.6 / 98.498.4 / 95.090.7 / 91.9
+ +Table 5. Compared with other competing methods under few-normal-shot anomaly detection setting. The metric of Pixel-level is AUROC / PRO for the few-normal-shot anomaly segmentation task, and the metric of Image-level is AUROC / AP for the few-normal-shot anomaly detection task. Best results are highlighted in red. + +by the industrial inspection MVTec AD dataset, as shown in Table 1 and Table 2. The superior performance of FE-CLIP demonstrates that frequency information could be beneficial for many cases of anomaly detection, and the proposed FE-CLIP has good generalization across different domains. + +# 4.6. FE-CLIP with Few Normal Shots + +The performance of FE-CLIP can also be further boosted when equipped with few-shot normal images, just like WinCLIP [17] and AnomalyGPT [14]. We adopt the same way as AnomalyGPT to compute the anomaly segmentation map with few-shot normal images. In short words, we use the same visual encoder of FE-CLIP to extract intermediate patch-level features from normal images and store them in memory banks, then calculate the cosine similarity distance between each patch token of input image and its most similar counterpart in the memory banks to obtain the anomaly segmentation map $S_{a,few\_shot}$ . Finally, we add the anomaly segmentation map $S_{a}$ from the FE-CLIP model with $S_{a,few\_shot}$ to get the final anomaly segmentation map, and add the anomaly score $M_{a}$ from the FE-CLIP model with the max value of $S_{a,few\_shot}$ to get the final anomaly score. Please refer to [14] for more technical details. + +We compare the FE-CLIP with other competing methods, such as AnomalyGPT [14], PromptAD [24], and InCTRL [53], under the few-shot anomaly detection setting. As shown in Table 5, the FE-CLIP outperforms all other com + +peting methods across different metrics, including AUROC and PRO metrics for the anomaly detection task, and AUROC and AP metrics for the anomaly segmentation task, on both datasets. The experiments under few-normal-shot setting further demonstrate the superiority of the FE-CLIP. + +# 5. Conclusion + +In this paper, we find that frequency information can benefit the ZSAD task. To this end, we propose FE-CLIP, which adds the proposed FFE and LFS adapters into the visual encoder of CLIP, to deeply mine the frequency patterns for the visual features. Extensive experiments on 10 datasets of highly diverse class semantics from various defect inspections and medical domains demonstrate that (1) the frequency information can benefit the ZSAD task, and (2) the frequency information can help the model have good generalization across datasets from different domains. We also conduct experiments under the few-normal-shot anomaly detection setting to further validate the superiority of the proposed FE-CLIP. We hope the community of anomaly detection will be aware of the effectiveness of frequency information from our work and explore more ways to utilize the frequency information for anomaly detection in the future. + +# Acknowledgements + +This work was supported by the Anhui Provincial Science and Technology Major Project (No. 2023z020006). + +# References + +[1] Nasir Ahmed, T. Natarajan, and Kamisetty R Rao. Discrete cosine transform. IEEE transactions on Computers, 100(1): 90-93, 1974. 2, 4, 5 +[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 2 +[3] Toshimichi Aota, Lloyd Teh Tzer Tong, and Takayuki Okatani. Zero-shot versus many-shot: Unsupervised texture anomaly detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5564-5572, 2023. 3, 5 +[4] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad-a comprehensive real-world dataset for unsupervised anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9592-9600, 2019. 1, 5 +[5] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Uninformed students: Student-teacher anomaly detection with discriminative latent embeddings. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4183–4192, 2020. 1, 5 +[6] Jorge Bernal, F Javier Sánchez, Gloria Fernández-Esparrach, Debora Gil, Cristina Rodríguez, and Fernando Vilarino. Wm-dova maps for accurate polyp highlighting in colonoscopy: Validation vs. saliency maps from physicians. Computerized medical imaging and graphics, 43:99-111, 2015. 5 +[7] Tri Cao, Jiawen Zhu, and Guansong Pang. Anomaly detection under distribution shift. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6511-6523, 2023. 1 +[8] Yunkang Cao, Jiangning Zhang, Luca Frittoli, Yuqi Cheng, Weiming Shen, and Giacomo Boracchi. Adaclip: Adapting clip with hybrid learnable prompts for zero-shot anomaly detection. In European Conference on Computer Vision, 2024. 1, 2, 5, 6 +[9] Xuhai Chen, Yue Han, and Jiangning Zhang. A zero-/few-shot anomaly classification and segmentation method for cvpr 2023 vand workshop challenge tracks 1&2: 1st place on zero-shot ad and 4th place on few-shot ad. arXiv preprint arXiv:2305.17382, 2023. 3, 6 +[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.3 +[11] Tharindu Fernando, Harshala Gammulle, Simon Denman, Sridha Sridharan, and Clinton Fookes. Deep learning for medical anomaly detection-a survey. ACM Computing Surveys (CSUR), 54(7):1-37, 2021. 1 +[12] Gabriel Goh, Nick Cammarata, Chelsea Voss, Shan Carter, Michael Petrov, Ludwig Schubert, Alec Radford, and Chris + +Olah. Multimodal neurons in artificial neural networks. Distill, 6(3):e30, 2021. 2 +[13] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021. 3 +[14] Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Ming Tang, and Jinqiao Wang. Anomalygpt: Detecting industrial anomalies using large vision-language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1932-1940, 2024. 3, 8 +[15] A. Hamada. Br35h: Brain tumor detection 2020. In https://www.kaggle.com/datasets/ahmedhamada0/brain-tumor-detection, 2020. 5 +[16] Chaoqin Huang, Haoyan Guan, Aofan Jiang, Ya Zhang, Michael Spratling, and Yan-Feng Wang. Registration based few-shot anomaly detection. In European Conference on Computer Vision, pages 303-319. Springer, 2022. 1 +[17] Jongheon Jeong, Yang Zou, Taewan Kim, Dongqing Zhang, Avinash Ravichandran, and Onkar Dabeer. Winclip: Zero/few-shot anomaly classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19606-19616, 2023. 1, 3, 4, 5, 6, 8 +[18] Stepan Jezek, Martin Jonak, Radim Burget, Pavel Dvorak, and Milos Skotak. Deep learning-based defect detection of metal parts: evaluating current methods in complex conditions. In 2021 13th International congress on ultra modern telecommunications and control systems and workshops (ICUMT), pages 66-71. IEEE, 2021. 5 +[19] Debesh Jha, Pia H Smedsrud, Michael A Riegler, Pål Halvorsen, Thomas De Lange, Dag Johansen, and Håvard D Johansen. Kvasir-seg: A segmented polyp dataset. In MultiMedia Modeling: 26th International Conference, MMM 2020, Daejeon, South Korea, January 5–8, 2020, Proceedings, Part II 26, pages 451–462. Springer, 2020. 5 +[20] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 2 +[21] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 2 +[22] Shengze Li, Jianjian Cao, Peng Ye, Yuhan Ding, Chongjun Tu, and Tao Chen. Clipsam: Clip and sam collaboration for zero-shot anomaly segmentation. arXiv preprint arXiv:2401.12665, 2024. 1, 3 +[23] Xiaoya Li, Xiaofei Sun, Yuxian Meng, Junjun Liang, Fei Wu, and Jiwei Li. Dice loss for data-imbalanced nlp tasks. arXiv preprint arXiv:1911.02855, 2019. 5 +[24] Xiaofan Li, Zhizhong Zhang, Xin Tan, Chengwei Chen, Yanyun Qu, Yuan Xie, and Lizhuang Ma. Promptad: Learning prompts with only normal samples for few-shot anomaly + +detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024. 8 +[25] Yuanze Li, Haolin Wang, Shihao Yuan, Ming Liu, Debin Zhao, Yiwen Guo, Chen Xu, Guangming Shi, and Wangmeng Zuo. Myriad: Large multimodal model by applying vision experts for industrial anomaly detection. arXiv preprint arXiv:2310.19070, 2023. 3 +[26] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5 +[27] Jie Liu, Yixiao Zhang, Jie-Neng Chen, Junfei Xiao, Yongyi Lu, Bennett A Landman, Yixuan Yuan, Alan Yuille, Yucheng Tang, and Zongwei Zhou. Clip-driven universal model for organ segmentation and tumor detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21152-21164, 2023. 1 +[28] Philipp Liznerski, Lukas Ruff, Robert A Vandermeulen, Billy Joe Franks, Marius Kloft, and Klaus-Robert Müller. Explainable deep one-class classification. arXiv preprint arXiv:2007.01760, 2020. 1 +[29] Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, and Aniruddha Kembhavi. Unified-io: A unified model for vision, language, and multi-modal tasks. In The Eleventh International Conference on Learning Representations, 2022. 2 +[30] Pankaj Mishra, Riccardo Verk, Daniele Fornasier, Claudio Piciarelli, and Gian Luca Foresti. Vt-adj: A vision transformer network for image anomaly detection and localization. In 2021 IEEE 30th International Symposium on Industrial Electronics (ISIE), pages 01-06. IEEE, 2021. 5 +[31] Ziyuan Qin, Huahui Yi, Qicheng Lao, and Kang Li. Medical image understanding with pretrained vision language models: A comprehensive study. arXiv preprint arXiv:2209.15517, 2022. 1 +[32] Zhen Qu, Xian Tao, Mukesh Prasad, Fei Shen, Zhengtao Zhang, Xinyi Gong, and Guiguang Ding. Vcp-clip: A visual context prompting model for zero-shot anomaly segmentation. In European Conference on Computer Vision, 2024. 1, 2, 6 +[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 1, 2, 3, 4, 5, 6, 7 +[34] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18082-18091, 2022. 3 +[35] Tal Reiss and Yedid Hoshen. Mean-shifted contrastive loss for anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2155-2162, 2023. 1 +[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image + +synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3 +[37] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14318-14328, 2022. 1, 8 +[38] Mohammadreza Salehi, Niousha Sadjadi, Soroosh Baselizadeh, Mohammad H Rohban, and Hamid R Rabiee. Multiresolution knowledge distillation for anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14902-14912, 2021. 5 +[39] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2 +[40] Rohan Taori, Achal Dave, Vaishaal Shankar, Nicholas Carlini, Benjamin Recht, and Ludwig Schmidt. Measuring robustness to natural distribution shifts in image classification. Advances in Neural Information Processing Systems, 33:18583-18599, 2020. 2 +[41] Yoad Tewel, Yoav Shalev, Idan Schwartz, and Lior Wolf. Zerocap: Zero-shot image-to-text generation for visual-semantic arithmetic. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17918-17928, 2022. 3 +[42] Yu Tian, Guansong Pang, Fengbei Liu, Yuanhong Chen, Seon Ho Shin, Johan W Verjans, Rajvinder Singh, and Gustavo Carneiro. Constrained contrastive distribution learning for unsupervised anomaly detection and localisation in medical images. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27-October 1, 2021, Proceedings, Part V 24, pages 128-140. Springer, 2021. 1 +[43] Yu Tian, Fengbei Liu, Guansong Pang, Yuanhong Chen, Yuyuan Liu, Johan W Verjans, Rajvinder Singh, and Gustavo Carneiro. Self-supervised pseudo multi-class pre-training for unsupervised anomaly detection and segmentation in medical images. Medical image analysis, 90:102930, 2023. 1 +[44] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4 +[45] Matthias Wieler and Tobias Hahn. Weakly supervised learning for industrial optical inspection. In DAGM symposium in, page 11, 2007. 5 +[46] Mengde Xu, Zheng Zhang, Fangyun Wei, Yutong Lin, Yue Cao, Han Hu, and Xiang Bai. A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In European Conference on Computer Vision, pages 736-753. Springer, 2022. 3 +[47] Zhiyuan You, Lei Cui, Yujun Shen, Kai Yang, Xin Lu, Yu Zheng, and Xinyi Le. A unified model for multi-class + +anomaly detection. Advances in Neural Information Processing Systems, 35:4571-4584, 2022. 1 +[48] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16793-16803, 2022. 3 +[49] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In European Conference on Computer Vision, pages 696-712. Springer, 2022. 3 +[50] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 3, 6 +[51] Qihang Zhou, Shibo He, Haoyu Liu, Tao Chen, and Jiming Chen. Pull & push: Leveraging differential knowledge distillation for efficient unsupervised anomaly detection and localization. IEEE Transactions on Circuits and Systems for Video Technology, 2022. 1 +[52] Qihang Zhou, Guansong Pang, Yu Tian, Shibo He, and Jiming Chen. Anomalyclip: Object-agnostic prompt learning for zero-shot anomaly detection. The thirteenth International Conference on Learning Representations, 2024. 1, 3, 5, 6 +[53] Jiawen Zhu and Guansong Pang. Toward generalist anomaly detection via in-context residual learning with few-shot sample prompts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024. 8 +[54] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. In European Conference on Computer Vision, pages 392-408. Springer, 2022. 5 \ No newline at end of file diff --git a/2025/FE-CLIP_ Frequency Enhanced CLIP Model for Zero-Shot Anomaly Detection and Segmentation/images.zip b/2025/FE-CLIP_ Frequency Enhanced CLIP Model for Zero-Shot Anomaly Detection and Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a80e06015b94002b85b6c137aab6303f6b2ea0d0 --- /dev/null +++ b/2025/FE-CLIP_ Frequency Enhanced CLIP Model for Zero-Shot Anomaly Detection and Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:39549efcd17267f76f7a6cbfff9e05a46ef3588783f651c8f33d3d8a277dd204 +size 482533 diff --git a/2025/FE-CLIP_ Frequency Enhanced CLIP Model for Zero-Shot Anomaly Detection and Segmentation/layout.json b/2025/FE-CLIP_ Frequency Enhanced CLIP Model for Zero-Shot Anomaly Detection and Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4953fc0210ba79ba9db8166bdf44042006ce9fc9 --- /dev/null +++ b/2025/FE-CLIP_ Frequency Enhanced CLIP Model for Zero-Shot Anomaly Detection and Segmentation/layout.json @@ -0,0 +1,9363 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 143, + 103, + 465, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 103, + 465, + 138 + ], + "spans": [ + { + "bbox": [ + 143, + 103, + 465, + 138 + ], + "type": "text", + "content": "FE-CLIP: Frequency Enhanced CLIP Model for Zero-Shot Anomaly Detection and Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 160, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 479, + 175 + ], + "type": "text", + "content": "Tao Gong " + }, + { + "bbox": [ + 130, + 160, + 479, + 175 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 130, + 160, + 479, + 175 + ], + "type": "text", + "content": ", Qi Chu " + }, + { + "bbox": [ + 130, + 160, + 479, + 175 + ], + "type": "inline_equation", + "content": "^{1,2,3*}" + }, + { + "bbox": [ + 130, + 160, + 479, + 175 + ], + "type": "text", + "content": ", Bin Liu " + }, + { + "bbox": [ + 130, + 160, + 479, + 175 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 130, + 160, + 479, + 175 + ], + "type": "text", + "content": ", Wei Zhou " + }, + { + "bbox": [ + 130, + 160, + 479, + 175 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 130, + 160, + 479, + 175 + ], + "type": "text", + "content": ", Nenghai Yu " + }, + { + "bbox": [ + 130, + 160, + 479, + 175 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 87, + 175, + 523, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 175, + 523, + 190 + ], + "spans": [ + { + "bbox": [ + 87, + 175, + 523, + 190 + ], + "type": "text", + "content": "1School of Cyber Science and Technology, University of Science and Technology of China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 180, + 189, + 431, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 189, + 431, + 203 + ], + "spans": [ + { + "bbox": [ + 180, + 189, + 431, + 203 + ], + "type": "text", + "content": "2Anhui Province Key Laboratory of Digital Security" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 169, + 203, + 441, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 203, + 441, + 217 + ], + "spans": [ + { + "bbox": [ + 169, + 203, + 441, + 217 + ], + "type": "text", + "content": "3the CCCD Key Lab of Ministry of Culture and Tourism" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 209, + 217, + 400, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 217, + 400, + 231 + ], + "spans": [ + { + "bbox": [ + 209, + 217, + 400, + 231 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 209, + 217, + 400, + 231 + ], + "type": "text", + "content": "Ling Yang Industrial Internet Co., Ltd." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 146, + 232, + 460, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 232, + 460, + 245 + ], + "spans": [ + { + "bbox": [ + 146, + 232, + 460, + 245 + ], + "type": "text", + "content": "{tgong, qchu, flowice, ynh} @ustc.edu.cn, weizhou8@iflytek.com" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "spans": [ + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 299, + 295, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 295, + 622 + ], + "type": "text", + "content": "Zero-shot anomaly detection (ZSAD) requires detection models trained using auxiliary data to detect anomalies without any training sample in a target dataset. It is challenging since the models need to generalize to anomalies across different domains. Recently, CLIP-based anomaly detection methods, such as WinCLIP and AnomalyCLIP, have demonstrated superior performance in the ZSAD task, due to the strong zero-shot recognition of the CLIP model. However, they overlook the utilization of frequency information of images. In this paper, we find that frequency information could benefit the ZSAD task, since some properties of the anomaly area, such as appearance defects, can also be reflected based on its frequency information. To this end, We propose Frequency Enhanced CLIP (FE-CLIP), taking advantage of two different but complementary frequency-aware clues, (1) Frequency-aware Feature Extraction adapter, and (2) Local Frequency Statistics adapter, in the visual encoder of CLIP, to deeply mine frequency information for the ZSAD task. We apply DCT as the frequency-domain transformation. Through comprehensive experiments, we show that the proposed FE-CLIP has good generalization across different domains and achieves superior zero-shot performance of detecting and segmenting anomalies in 10 datasets of highly diverse class semantics from various defect inspections and medical domains. Besides, FE-CLIP also achieves superior performance under the few-normal-shot anomaly detection settings." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 135, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 135, + 659 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 135, + 659 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 668, + 295, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 668, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 668, + 295, + 693 + ], + "type": "text", + "content": "Anomaly Detection aims to predict an image or a pixel as normal or anomalous, and has been widely applied in" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 274, + 555, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 555, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 555, + 467 + ], + "type": "text", + "content": "various applications, such as industrial defect inspection [4, 5, 7, 16, 28, 35, 37, 47, 51] and medical image analysis [11, 27, 31, 42, 43]. Existing AD approaches typically assume that training examples in a target application domain are available for learning the detection models. However, this assumption may not hold in various scenarios, such as i) when accessing training data violates data privacy policies, or ii) when the target domain does not have relevant training data. Therefore, Zero-shot anomaly detection (ZSAD) is an emerging task for AD in such scenarios, to which the AD methods mentioned above are not viable, as it requires detection models to detect anomalies without any training sample in a target dataset. ZSAD is challenging since the models need to generalize to anomalies across different domains where the appearance of foreground objects, abnormal regions, and background features can vary significantly." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": "Recently, large pre-trained vision-language models (VLMs) have demonstrated strong zero-shot recognition ability in various vision tasks. Particularly, being pretrained using millions/billions of image-text pairs, CLIP [33] has been applied to empower various downstream tasks with its strong generalization capability. WinCLIP [17] is the first work to apply the CLIP model in the ZSAD line. It uses a large number of hand-crafted text prompts and extraction of window/patch/image-level features aligned with text to finish the ZSAD task. However, WinCLIP involves multiple forward passes of image patches for anomaly segmentation, which leads to inefficiency. Besides, WinCLIP only performs the testing of the CLIP model in the ZSAD without any fine-tuning using the anomaly detection task data, which also limits its performance in the ZSAD task. Subsequently, several works [8, 22, 32, 52] attempt to adapt the CLIP into the ZSAD task by fintuning. Specifically, AnomalyCLIP [52] learns object-agnostic text prompts that capture generic normality and abnormality in an image regardless of its foreground objects. ClipSAM [22] introduces" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 142, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 142, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 142, + 713 + ], + "type": "text", + "content": "*Corresponding Author" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21220" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 72, + 116, + 129 + ], + "blocks": [ + { + "bbox": [ + 62, + 72, + 116, + 129 + ], + "lines": [ + { + "bbox": [ + 62, + 72, + 116, + 129 + ], + "spans": [ + { + "bbox": [ + 62, + 72, + 116, + 129 + ], + "type": "image", + "image_path": "1c26cd97864a68e083005162dfc3547c7765e45f23d3c57b6e61b7a48aa42a10.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 119, + 72, + 174, + 129 + ], + "blocks": [ + { + "bbox": [ + 119, + 72, + 174, + 129 + ], + "lines": [ + { + "bbox": [ + 119, + 72, + 174, + 129 + ], + "spans": [ + { + "bbox": [ + 119, + 72, + 174, + 129 + ], + "type": "image", + "image_path": "9d558e64152f6531e8db98707af6e0a4f628d3d575aee100ae81214939f2f93e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 176, + 72, + 231, + 129 + ], + "blocks": [ + { + "bbox": [ + 176, + 72, + 231, + 129 + ], + "lines": [ + { + "bbox": [ + 176, + 72, + 231, + 129 + ], + "spans": [ + { + "bbox": [ + 176, + 72, + 231, + 129 + ], + "type": "image", + "image_path": "481e6f3295a2142aeb98cd2466fb958c8eb1f96e00ab27f6e11da1e0a5400fcb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 233, + 72, + 288, + 129 + ], + "blocks": [ + { + "bbox": [ + 233, + 72, + 288, + 129 + ], + "lines": [ + { + "bbox": [ + 233, + 72, + 288, + 129 + ], + "spans": [ + { + "bbox": [ + 233, + 72, + 288, + 129 + ], + "type": "image", + "image_path": "e0912137d64943e50914be6d383485a7da6123f6874c46937667ee4e5b732b41.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 62, + 129, + 116, + 185 + ], + "blocks": [ + { + "bbox": [ + 62, + 129, + 116, + 185 + ], + "lines": [ + { + "bbox": [ + 62, + 129, + 116, + 185 + ], + "spans": [ + { + "bbox": [ + 62, + 129, + 116, + 185 + ], + "type": "image", + "image_path": "624acac42e6ce9861fb5a31243faba7f9dca38d1893b808927d4a803f61fb3d6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 186, + 104, + 194 + ], + "lines": [ + { + "bbox": [ + 75, + 186, + 104, + 194 + ], + "spans": [ + { + "bbox": [ + 75, + 186, + 104, + 194 + ], + "type": "text", + "content": "(a) normal" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 119, + 129, + 174, + 185 + ], + "blocks": [ + { + "bbox": [ + 119, + 129, + 174, + 185 + ], + "lines": [ + { + "bbox": [ + 119, + 129, + 174, + 185 + ], + "spans": [ + { + "bbox": [ + 119, + 129, + 174, + 185 + ], + "type": "image", + "image_path": "1d0eda1e2bd73fd2b08f254c32a05142fe642f32d8b33e44f39c0fe41c77365b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 186, + 158, + 194 + ], + "lines": [ + { + "bbox": [ + 135, + 186, + 158, + 194 + ], + "spans": [ + { + "bbox": [ + 135, + 186, + 158, + 194 + ], + "type": "text", + "content": "(b) hole" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 176, + 129, + 231, + 185 + ], + "blocks": [ + { + "bbox": [ + 176, + 129, + 231, + 185 + ], + "lines": [ + { + "bbox": [ + 176, + 129, + 231, + 185 + ], + "spans": [ + { + "bbox": [ + 176, + 129, + 231, + 185 + ], + "type": "image", + "image_path": "43a1f479838ef21d8c8f7fbcfdcdee60401235314c45aa66c30cbcd9d46cca12.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 186, + 216, + 194 + ], + "lines": [ + { + "bbox": [ + 192, + 186, + 216, + 194 + ], + "spans": [ + { + "bbox": [ + 192, + 186, + 216, + 194 + ], + "type": "text", + "content": "(c) crack" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 204, + 295, + 270 + ], + "lines": [ + { + "bbox": [ + 55, + 204, + 295, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 295, + 270 + ], + "type": "text", + "content": "Figure 1. The local frequency statistics of images. The second row represents the local frequency statistics of non-overlapped image regions. There are four columns, standing for the normal object, object with a hole, object with a crack, and object with print, respectively. Please zoom in and pay attention to the red region (i.e. the anomaly region of the input image) for a better view." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 233, + 129, + 288, + 185 + ], + "blocks": [ + { + "bbox": [ + 233, + 129, + 288, + 185 + ], + "lines": [ + { + "bbox": [ + 233, + 129, + 288, + 185 + ], + "spans": [ + { + "bbox": [ + 233, + 129, + 288, + 185 + ], + "type": "image", + "image_path": "c6e47206040e53189505e927df62368744ab12c30d1fdec37db0117a4990442e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 186, + 272, + 194 + ], + "lines": [ + { + "bbox": [ + 249, + 186, + 272, + 194 + ], + "spans": [ + { + "bbox": [ + 249, + 186, + 272, + 194 + ], + "type": "text", + "content": "(d) print" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 54, + 293, + 295, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 293, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 54, + 293, + 295, + 413 + ], + "type": "text", + "content": "a unified multiscale cross-modal interaction module for interacting language with visual features at multiple stages of the CLIP visual encoder to reason anomaly positions. VCP-CLIP [32] propose the Pre-VCP module and PostVCP module to perform visual context prompting to activate CLIP's anomalous semantic perception ability. AdaCLIP [8] incorporates static and dynamic learnable prompts to adapt CLIP for ZSAD. However, all of the methods mentioned above overlook the utilization of frequency information of images." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 54, + 414, + 295, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 414, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 54, + 414, + 295, + 545 + ], + "type": "text", + "content": "In this paper, we find that frequency information could benefit the ZSAD task. Some properties of the anomaly area, such as appearance defects, can also be reflected based on its frequency information, since high-frequency components represent fine details (e.g. soft borders) that often exist in the anomaly area. For example, as shown in Figure 1, the frequency statistics in the red region are different from the frequency statistics in the surrounding regions of the input abnormal image and the corresponding region of the normal image. These specific frequency patterns could be helpful for anomaly detection." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 546, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 714 + ], + "type": "text", + "content": "To this end, we introduce a novel approach, namely the Frequency Enhanced CLIP (FE-CLIP), to inject frequency information into CLIP for accurate ZSAD across different domains. FE-CLIP takes advantage of two different but complementary frequency-aware clues, (1) Frequency-aware Feature Extraction (FFE) adapter, and (2) Local Frequency Statistics (LFS) adapter, in the visual encoder of CLIP, to deeply mine frequency information. Specifically, the FFE adapter uses DCT [1] to transform the image features into the frequency domain in order to mine the frequency information, then uses inverse DCT to transform back to the spatial domain. The FFE adapter describes the frequency-aware patterns in the spatial domain, but does not explicitly render the frequency information directly in" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 72, + 553, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 252 + ], + "type": "text", + "content": "the neural networks. Therefore, we use LFS to render the frequency information directly in the frequency domain. In each densely but regularly sampled local spatial patch, the statistics are gathered by counting the mean frequency responses. The averaged frequency statistics are a multichannel spatial map, where the number of channels is identical to the number of channels in the spatial domain. The local frequency statistics also follow the spatial layouts as the features of the CLIP visual encoder, thus also enjoying effective representation learning powered by the CLIP visual encoder. Meanwhile, since the frequency-aware feature extraction and local frequency statistics are complementary to each other but both of them share inherently similar frequency-aware semantics, thus they can be progressively fused during the feature learning process." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 253, + 553, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 553, + 335 + ], + "type": "text", + "content": "Through comprehensive experiments, we show that the proposed FE-CLIP has good generalization across different domains and achieves superior zero-shot performance of detecting and segmenting anomalies in 10 datasets of highly diverse class semantics from various defect inspections and medical domains. Our contributions in this paper are summarized as follows:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 337, + 555, + 540 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 315, + 337, + 553, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 337, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 315, + 337, + 553, + 384 + ], + "type": "text", + "content": "1. We find that frequency information could benefit the ZSAD task, and propose a novel Frequency Enhanced CLIP (FE-CLIP) method to inject frequency information into CLIP for accurate ZSAD across different domains." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 385, + 553, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 385, + 553, + 445 + ], + "spans": [ + { + "bbox": [ + 315, + 385, + 553, + 445 + ], + "type": "text", + "content": "2. We propose two different but complementary frequency-aware adapters, Frequency-aware Feature Extraction (FFE) adapter, and Local Frequency Statistics (LFS) adapter, in the visual encoder of CLIP, to deeply mine the frequency patterns of anomaly region." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 445, + 555, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 445, + 555, + 540 + ], + "spans": [ + { + "bbox": [ + 315, + 445, + 555, + 540 + ], + "type": "text", + "content": "3. Extensive experiments show that the proposed FE-CLIP has good generalization across different domains and achieves superior zero-shot performance of detecting and segmenting anomalies in 10 datasets of highly diverse class semantics from various defect inspections and medical domains. Besides, the proposed FE-CLIP also achieves superior performance under the few-normal-shot anomaly detection settings." + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 555, + 400, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 555, + 400, + 567 + ], + "spans": [ + { + "bbox": [ + 314, + 555, + 400, + 567 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 575, + 458, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 575, + 458, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 575, + 458, + 588 + ], + "type": "text", + "content": "2.1. Vision-language modeling" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "Among recent successes of large pre-trained vision-language models (VLM) [2, 20, 33], CLIP [33] is the first to perform pre-training on web-scale image-text data, showing unprecedented generality: e.g., its language-driven zero-shot inference, improved both effective robustness [40] and perceptual alignment [12]. Many following VLM works explored large-scale pre-training in different aspects, e.g., scaling up data [20], efficient designs [21], multi-tasks [29], etc. To democratize large-scale VLM for the usages in different domains, a billion-scale data LAION5B [39], a" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21221" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 97, + 78, + 518, + 285 + ], + "blocks": [ + { + "bbox": [ + 97, + 78, + 518, + 285 + ], + "lines": [ + { + "bbox": [ + 97, + 78, + 518, + 285 + ], + "spans": [ + { + "bbox": [ + 97, + 78, + 518, + 285 + ], + "type": "image", + "image_path": "aa8fb52490b307caa182158c166aa25e1db21b01264cdc8cc19722b5bec154fb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 298, + 555, + 333 + ], + "lines": [ + { + "bbox": [ + 54, + 298, + 555, + 333 + ], + "spans": [ + { + "bbox": [ + 54, + 298, + 555, + 333 + ], + "type": "text", + "content": "Figure 2. The framework of the proposed FE-CLIP. The text encoder and visual encoder are from the CLIP model, and the FFE and LFS are the proposed adapters. We only show the class token embedding after the last block of the CLIP visual encoder for convenience. Actually, the CLIP visual encoder can output the class token after every block." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 353, + 296, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 353, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 54, + 353, + 296, + 483 + ], + "type": "text", + "content": "code base of OpenCLIP with pre-trained models are open-sourced. Other works presented CLIP's promise in zero/few-shot transfer to downstream tasks beyond classification [13, 36, 41, 46]. Good prompt engineering and tuning [33, 50] can non-trivially benefit generalization performances. Moreover, some other works [34, 48, 49] leverage the pre-trained CLIP for language-guided detection and segmentation with promising performances. In this paper, we mainly focus on adapting CLIP into the ZSAD task with superior performance by injecting the frequency information into the visual encoder of CLIP." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 492, + 216, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 492, + 216, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 492, + 216, + 506 + ], + "type": "text", + "content": "2.2. Zero-shot Anomaly detection" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 510, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 715 + ], + "type": "text", + "content": "ZSAD relies on the model's strong transferability to handle unseen anomalies [3]. A very recent approach WinCLIP [17] presents a seminal work that leverages CLIP [33] for zero-shot anomaly classification and segmentation. WinCLIP proposes a compositional ensemble on state words and prompt templates and efficient extraction and aggregation of window/patch/image-level features aligned with text to finish the ZSAD task. To tackle this inefficiency, APRILGAN [9] introduces a fine-tuning strategy by adding extra learnable linear projection layers in multiple stages of the CLIP visual encoder to enhance the modeling of local visual semantics. AnomalyCLIP [52] learns object-agnostic text prompts that capture generic normality and abnormality in an image regardless of its foreground objects. ClipSAM [22] introduces a unified multiscale cross-modal interaction module for interacting language with visual features at multiple stages of the CLIP visual encoder to reason anomaly" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 353, + 556, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 353, + 556, + 497 + ], + "spans": [ + { + "bbox": [ + 313, + 353, + 556, + 497 + ], + "type": "text", + "content": "positions. Myriad [25] and AnomalyGPT [14] utilize large language models for zero-shot/few-shot anomaly detection. They employ an image decoder to provide fine-grained semantics and design a prompt learner to fine-tune the large language models using prompt embeddings. However, all of the methods mentioned above overlook the utilization of frequency information of images. In this paper, we find that frequency information could benefit the ZSAD task, and propose the FE-CLIP method which utilizes two different but complementary adapters, the FFE adapter and the LFS adapter, into the visual encoder of CLIP, to deeply mine the frequency information for ZSAD task." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 508, + 372, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 372, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 372, + 521 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 528, + 394, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 394, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 394, + 541 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": "CLIP [33] is a large-scale pretraining method offering a joint vision-language representation. It consists of a text encoder " + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "inline_equation", + "content": "T(\\cdot)" + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": " and a visual encoder " + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "inline_equation", + "content": "F(\\cdot)" + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": ". Both encoders are mainstream multi-block networks such as ViT [10]. Using text prompts is a typical way to acquire the embeddings of different classes for zero-shot classification. The text prompts usually combine a text prompt template " + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": " with the category name " + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": ", where the text prompt template " + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": " usually is defined as A photo of a [cls] and [cls] represents the target category name. The text prompts are passed through " + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "inline_equation", + "content": "T(\\cdot)" + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": " to obtain its corresponding textual embedding " + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "inline_equation", + "content": "g_{c} \\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": " denotes the channel dimension. Then, an input image is passed through " + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "inline_equation", + "content": "F(\\cdot)" + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": " to get the visual representations, where the class token " + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "inline_equation", + "content": "f \\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": " is the" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21222" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": "global visual embedding of the input image, and patch tokens " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "f^{m} \\in \\mathbb{R}^{H \\times W \\times D}" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": " are treated as local visual embeddings of the input image. " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": " denote the height and width of the patch tokens, respectively. CLIP can perform zero-shot recognition by measuring the similarity between textual and visual embeddings. Specifically, given a target class set " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": " and an image, CLIP predicts the probability of the image belonging to the " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": " category as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 176, + 295, + 205 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 176, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 61, + 176, + 295, + 205 + ], + "type": "interline_equation", + "content": "p (y = c) = P \\left(g _ {c}, f\\right) = \\frac {\\exp \\left(< g _ {c} , f > / \\tau\\right)}{\\sum_ {c \\in \\mathcal {C}} \\exp \\left(< g _ {c} , f >\\right) / \\tau} \\tag {1}", + "image_path": "fc079798e769533d7517adc1b9c589d41a05b241357c496e0cc66fd6994b9a17.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 212, + 296, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 212, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 212, + 296, + 285 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 212, + 296, + 285 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 212, + 296, + 285 + ], + "type": "text", + "content": " denotes the temperature and the operator " + }, + { + "bbox": [ + 55, + 212, + 296, + 285 + ], + "type": "inline_equation", + "content": "< ,\\cdot ,>" + }, + { + "bbox": [ + 55, + 212, + 296, + 285 + ], + "type": "text", + "content": " represents the computation of cosine similarity. The computation can be extended from global visual embeddings to local visual embeddings to derive the corresponding segmentation maps " + }, + { + "bbox": [ + 55, + 212, + 296, + 285 + ], + "type": "inline_equation", + "content": "M_{c}\\in \\mathbb{R}^{H\\times W}" + }, + { + "bbox": [ + 55, + 212, + 296, + 285 + ], + "type": "text", + "content": " , if each entry " + }, + { + "bbox": [ + 55, + 212, + 296, + 285 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 55, + 212, + 296, + 285 + ], + "type": "text", + "content": " is computed as " + }, + { + "bbox": [ + 55, + 212, + 296, + 285 + ], + "type": "inline_equation", + "content": "P\\left(g_c,f^{m(i,j)}\\right)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 290, + 207, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 290, + 207, + 303 + ], + "spans": [ + { + "bbox": [ + 55, + 290, + 207, + 303 + ], + "type": "text", + "content": "3.2. Frequency Enhanced CLIP" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 308, + 296, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 308, + 296, + 608 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 296, + 608 + ], + "type": "text", + "content": "In this paper, we find that frequency information could benefit the ZSAD task, and propose a novel Frequency Enhanced CLIP (FE-CLIP) method to inject frequency information into CLIP [33] for accurate ZSAD across different domains. We inject the frequency information by the features of the CLIP visual encoder rather than the frequency of the input image, since the CLIP visual encoder is trained to take RGB image as input. Typically, the CLIP visual encoder comprises a series of block layers. From the bottom to the top of the layers, the visual encoder gradually learns the visual patterns at different levels of abstraction. As shown in Figure 2, we add Frequency-aware Feature Extraction (FFE) adapter and Local Frequency Statistics (LFS) adapter across multiple blocks levels in the visual encoder " + }, + { + "bbox": [ + 55, + 308, + 296, + 608 + ], + "type": "inline_equation", + "content": "F(\\cdot)" + }, + { + "bbox": [ + 55, + 308, + 296, + 608 + ], + "type": "text", + "content": " of CLIP, while keeping its original backbone unchanged, thus enabling frequency information injection at multiple feature levels. To be specific, assuming the visual encoder consists of " + }, + { + "bbox": [ + 55, + 308, + 296, + 608 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 308, + 296, + 608 + ], + "type": "text", + "content": " blocks (" + }, + { + "bbox": [ + 55, + 308, + 296, + 608 + ], + "type": "inline_equation", + "content": "N = 4" + }, + { + "bbox": [ + 55, + 308, + 296, + 608 + ], + "type": "text", + "content": "), the features (i.e. the patch tokens) after the " + }, + { + "bbox": [ + 55, + 308, + 296, + 608 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 308, + 296, + 608 + ], + "type": "text", + "content": "-th block are denoted as " + }, + { + "bbox": [ + 55, + 308, + 296, + 608 + ], + "type": "inline_equation", + "content": "f_{n}^{m} \\in \\mathbb{R}^{H \\times W \\times D}" + }, + { + "bbox": [ + 55, + 308, + 296, + 608 + ], + "type": "text", + "content": ", and the class token after the " + }, + { + "bbox": [ + 55, + 308, + 296, + 608 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 308, + 296, + 608 + ], + "type": "text", + "content": "-th block is denoted as " + }, + { + "bbox": [ + 55, + 308, + 296, + 608 + ], + "type": "inline_equation", + "content": "f_{n} \\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 55, + 308, + 296, + 608 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 308, + 296, + 608 + ], + "type": "inline_equation", + "content": "n = 1, 2,.., N" + }, + { + "bbox": [ + 55, + 308, + 296, + 608 + ], + "type": "text", + "content": ". Firstly, the features " + }, + { + "bbox": [ + 55, + 308, + 296, + 608 + ], + "type": "inline_equation", + "content": "f_{n}^{m}" + }, + { + "bbox": [ + 55, + 308, + 296, + 608 + ], + "type": "text", + "content": " are passed through an FFE adapter and an LFS adapter to get the features " + }, + { + "bbox": [ + 55, + 308, + 296, + 608 + ], + "type": "inline_equation", + "content": "f_{n,ffe}^{m}" + }, + { + "bbox": [ + 55, + 308, + 296, + 608 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 308, + 296, + 608 + ], + "type": "inline_equation", + "content": "f_{n,lfs}^{m}" + }, + { + "bbox": [ + 55, + 308, + 296, + 608 + ], + "type": "text", + "content": ", respectively, which contain the frequency information. We can get the frequency-aware feature " + }, + { + "bbox": [ + 55, + 308, + 296, + 608 + ], + "type": "inline_equation", + "content": "\\hat{f}_{n}^{m}" + }, + { + "bbox": [ + 55, + 308, + 296, + 608 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 94, + 616, + 295, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 616, + 295, + 632 + ], + "spans": [ + { + "bbox": [ + 94, + 616, + 295, + 632 + ], + "type": "interline_equation", + "content": "\\hat {f} _ {n} ^ {m} = \\lambda (f _ {n, f f e} ^ {m} + f _ {n, l f s} ^ {m}) + (1 - \\lambda) f _ {n} ^ {m} \\qquad (2)", + "image_path": "15ae6cb6b7cfb1315205a975bc9a57ff3b70551e2e191505c6f55a3fc0558cf7.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{f}_n^m" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " serving as the input for the next block. The class token " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "f_n" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " will also be injected with frequency information after the next block due to the self-attention mechanism [44] in the blocks. We set " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda = 0.1" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " to preserve the original knowledge of the CLIP visual encoder. Then, the normal text prompt and abnormal text prompt are passed through" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 321, + 70, + 549, + 125 + ], + "blocks": [ + { + "bbox": [ + 321, + 70, + 549, + 125 + ], + "lines": [ + { + "bbox": [ + 321, + 70, + 549, + 125 + ], + "spans": [ + { + "bbox": [ + 321, + 70, + 549, + 125 + ], + "type": "image", + "image_path": "ada198cb4923503e658aeebf6adf459bb11b39ce42a82b39c983defa58a9c7c5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 133, + 553, + 156 + ], + "lines": [ + { + "bbox": [ + 313, + 133, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 553, + 156 + ], + "type": "text", + "content": "Figure 3. The proposed Frequency-aware Feature Extraction (FFE) adapter." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 175, + 555, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 175, + 555, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 555, + 421 + ], + "type": "text", + "content": "text encoder " + }, + { + "bbox": [ + 313, + 175, + 555, + 421 + ], + "type": "inline_equation", + "content": "T(\\cdot)" + }, + { + "bbox": [ + 313, + 175, + 555, + 421 + ], + "type": "text", + "content": " to get corresponding textual embedding " + }, + { + "bbox": [ + 313, + 175, + 555, + 421 + ], + "type": "inline_equation", + "content": "g_{r}" + }, + { + "bbox": [ + 313, + 175, + 555, + 421 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 175, + 555, + 421 + ], + "type": "inline_equation", + "content": "g_{a}" + }, + { + "bbox": [ + 313, + 175, + 555, + 421 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 313, + 175, + 555, + 421 + ], + "type": "inline_equation", + "content": "g_{r}" + }, + { + "bbox": [ + 313, + 175, + 555, + 421 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 175, + 555, + 421 + ], + "type": "inline_equation", + "content": "g_{a}" + }, + { + "bbox": [ + 313, + 175, + 555, + 421 + ], + "type": "text", + "content": " constitute the target class set " + }, + { + "bbox": [ + 313, + 175, + 555, + 421 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 175, + 555, + 421 + ], + "type": "text", + "content": ". We simply use A photo of a normal object and A photo of a damaged object for the normal text prompt and abnormal text prompt, respectively, rather than the complex compositional ensemble on state words and prompt templates used in WinCLIP [17], since we find that the simple text prompt also works well. Finally, we compute the abnormal segmentation map " + }, + { + "bbox": [ + 313, + 175, + 555, + 421 + ], + "type": "inline_equation", + "content": "M_{a,n} \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 313, + 175, + 555, + 421 + ], + "type": "text", + "content": " based on the equation 1, where each entry " + }, + { + "bbox": [ + 313, + 175, + 555, + 421 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 313, + 175, + 555, + 421 + ], + "type": "text", + "content": " is computed as " + }, + { + "bbox": [ + 313, + 175, + 555, + 421 + ], + "type": "inline_equation", + "content": "P\\left(g_{a}, fc(\\hat{f}_{n}^{m(i,j)})\\right)" + }, + { + "bbox": [ + 313, + 175, + 555, + 421 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 313, + 175, + 555, + 421 + ], + "type": "inline_equation", + "content": "fc" + }, + { + "bbox": [ + 313, + 175, + 555, + 421 + ], + "type": "text", + "content": " means that we use a single learnable fc to align the dimension of visual features " + }, + { + "bbox": [ + 313, + 175, + 555, + 421 + ], + "type": "inline_equation", + "content": "\\hat{f}_{n}^{m}" + }, + { + "bbox": [ + 313, + 175, + 555, + 421 + ], + "type": "text", + "content": " to the dimension of text features. The abnormal score " + }, + { + "bbox": [ + 313, + 175, + 555, + 421 + ], + "type": "inline_equation", + "content": "S_{a,n}" + }, + { + "bbox": [ + 313, + 175, + 555, + 421 + ], + "type": "text", + "content": " is also computed based on the equation 1 as " + }, + { + "bbox": [ + 313, + 175, + 555, + 421 + ], + "type": "inline_equation", + "content": "P\\left(g_{a}, fc_{clip}(f_{n})\\right)" + }, + { + "bbox": [ + 313, + 175, + 555, + 421 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 313, + 175, + 555, + 421 + ], + "type": "inline_equation", + "content": "fc_{clip}" + }, + { + "bbox": [ + 313, + 175, + 555, + 421 + ], + "type": "text", + "content": " denotes the frozen project layer of the visual encoder in CLIP to align the dimension of class token " + }, + { + "bbox": [ + 313, + 175, + 555, + 421 + ], + "type": "inline_equation", + "content": "f_{n}" + }, + { + "bbox": [ + 313, + 175, + 555, + 421 + ], + "type": "text", + "content": " and the dimension of text features. The final prediction of the abnormal score " + }, + { + "bbox": [ + 313, + 175, + 555, + 421 + ], + "type": "inline_equation", + "content": "S_{a}" + }, + { + "bbox": [ + 313, + 175, + 555, + 421 + ], + "type": "text", + "content": " and abnormal segmentation map " + }, + { + "bbox": [ + 313, + 175, + 555, + 421 + ], + "type": "inline_equation", + "content": "M_{a}" + }, + { + "bbox": [ + 313, + 175, + 555, + 421 + ], + "type": "text", + "content": " are computed by averaging " + }, + { + "bbox": [ + 313, + 175, + 555, + 421 + ], + "type": "inline_equation", + "content": "S_{a,n}" + }, + { + "bbox": [ + 313, + 175, + 555, + 421 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 175, + 555, + 421 + ], + "type": "inline_equation", + "content": "M_{a,n}" + }, + { + "bbox": [ + 313, + 175, + 555, + 421 + ], + "type": "text", + "content": " across " + }, + { + "bbox": [ + 313, + 175, + 555, + 421 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 175, + 555, + 421 + ], + "type": "text", + "content": " blocks, respectively, as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 349, + 429, + 555, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 429, + 555, + 462 + ], + "spans": [ + { + "bbox": [ + 349, + 429, + 555, + 462 + ], + "type": "interline_equation", + "content": "S _ {a} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} S _ {a, n}, \\quad M _ {a} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} M _ {a, n} \\tag {3}", + "image_path": "81f082cc03ed5ae474f07aa35fcbf3322bfb0239ea9d7efed746a3adf6e09f52.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 468, + 552, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 552, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 552, + 481 + ], + "type": "text", + "content": "3.3. Frequency-aware Feature Extraction Adapter" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": "The specific frequency patterns could be beneficial for anomaly detection. To mine the frequency information, we propose the Frequency-aware Feature Extraction (FFE) adapter in the visual encoder of CLIP. Specifically, given the patch tokens of the " + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": "-th block " + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "inline_equation", + "content": "f_{n}^{m}\\in \\mathbb{R}^{H\\times W\\times D}" + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": ", we first split the " + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "inline_equation", + "content": "f_{n}^{m}" + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": " into non-overlapped windows with total " + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "inline_equation", + "content": "(H / P)\\times (W / P)" + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": " windows and each window containing " + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "inline_equation", + "content": "P\\times P" + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": " patch tokens (" + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": " is set to 3 by default). As shown in Figure 3, we apply DCT [1] transformation to each window to extract the DCT features " + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "inline_equation", + "content": "f_{n,ffe,1}^{m}" + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": " which represents the frequency information. Then, we pass " + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "inline_equation", + "content": "f_{n,ffe,1}^{m}" + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": " through one linear layer followed by one GELU(" + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": ") activation layer to further extract the frequency features " + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "inline_equation", + "content": "f_{n,ffe,2}^{m}" + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": ". Finally, we apply the inverse DCT transformation to " + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "inline_equation", + "content": "f_{n,ffe,2}^{m}" + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": " to get the frequency-aware features " + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "inline_equation", + "content": "f_{n,ffe}^{m}" + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": " in the spatial domain. The FFE adapter describes the frequency-aware patterns in the spatial domain, but does not explicitly render the frequency information directly in the neural networks. In the next sub-section, we propose an adapter that renders the fre" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21223" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 71, + 290, + 131 + ], + "blocks": [ + { + "bbox": [ + 62, + 71, + 290, + 131 + ], + "lines": [ + { + "bbox": [ + 62, + 71, + 290, + 131 + ], + "spans": [ + { + "bbox": [ + 62, + 71, + 290, + 131 + ], + "type": "image", + "image_path": "5e04a0e617d3013bf98a724038f9b30a721e7183e7309f2327e0716c29a434c0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 140, + 295, + 163 + ], + "lines": [ + { + "bbox": [ + 55, + 140, + 295, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 295, + 163 + ], + "type": "text", + "content": "Figure 4. The proposed Local Frequency Statistics (LFS) adapter." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 183, + 263, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 183, + 263, + 196 + ], + "spans": [ + { + "bbox": [ + 55, + 183, + 263, + 196 + ], + "type": "text", + "content": "quency information directly in the neural networks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 201, + 243, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 201, + 243, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 243, + 216 + ], + "type": "text", + "content": "3.4. Local Frequency Statistics Adapter" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 296, + 495 + ], + "type": "text", + "content": "The aforementioned FFE adapter has provided frequency-aware representation, but it has to represent the frequency-aware patterns back into the spatial domain, thus failing to directly utilize the frequency patterns in the frequency domain. Therefore, we propose the Local Frequency Statistics (LFS) adapter in the visual encoder of CLIP to render the frequency patterns directly in the frequency domain. As shown in Figure 4, given the patch tokens of the " + }, + { + "bbox": [ + 56, + 220, + 296, + 495 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 56, + 220, + 296, + 495 + ], + "type": "text", + "content": "-th block " + }, + { + "bbox": [ + 56, + 220, + 296, + 495 + ], + "type": "inline_equation", + "content": "f_{n}^{m}\\in \\mathbb{R}^{H\\times W\\times D}" + }, + { + "bbox": [ + 56, + 220, + 296, + 495 + ], + "type": "text", + "content": ", we first apply a Sliding Window DCT (SW DCT) on " + }, + { + "bbox": [ + 56, + 220, + 296, + 495 + ], + "type": "inline_equation", + "content": "f_{n}^{m}" + }, + { + "bbox": [ + 56, + 220, + 296, + 495 + ], + "type": "text", + "content": " (i.e., taking DCT [1] densely on sliding windows of the features) to extract the localized frequency responses " + }, + { + "bbox": [ + 56, + 220, + 296, + 495 + ], + "type": "inline_equation", + "content": "f_{n,lfs,1}^{m}" + }, + { + "bbox": [ + 56, + 220, + 296, + 495 + ], + "type": "text", + "content": ". The size of sliding windows is " + }, + { + "bbox": [ + 56, + 220, + 296, + 495 + ], + "type": "inline_equation", + "content": "Q\\times Q" + }, + { + "bbox": [ + 56, + 220, + 296, + 495 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 56, + 220, + 296, + 495 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 56, + 220, + 296, + 495 + ], + "type": "text", + "content": " is set to 3 by default). Therefore, there are " + }, + { + "bbox": [ + 56, + 220, + 296, + 495 + ], + "type": "inline_equation", + "content": "Q\\times Q" + }, + { + "bbox": [ + 56, + 220, + 296, + 495 + ], + "type": "text", + "content": " group frequency responses in " + }, + { + "bbox": [ + 56, + 220, + 296, + 495 + ], + "type": "inline_equation", + "content": "f_{n,lfs,1}^{m}" + }, + { + "bbox": [ + 56, + 220, + 296, + 495 + ], + "type": "text", + "content": ", where the first one represents the low-frequency responses and the last one represents the high-frequency responses. Then we count the mean of " + }, + { + "bbox": [ + 56, + 220, + 296, + 495 + ], + "type": "inline_equation", + "content": "f_{n,lfs,1}^{m}" + }, + { + "bbox": [ + 56, + 220, + 296, + 495 + ], + "type": "text", + "content": " across " + }, + { + "bbox": [ + 56, + 220, + 296, + 495 + ], + "type": "inline_equation", + "content": "Q\\times Q" + }, + { + "bbox": [ + 56, + 220, + 296, + 495 + ], + "type": "text", + "content": " groups to get the mean frequency responses " + }, + { + "bbox": [ + 56, + 220, + 296, + 495 + ], + "type": "inline_equation", + "content": "f_{n,lfs,2}^{m}" + }, + { + "bbox": [ + 56, + 220, + 296, + 495 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 56, + 220, + 296, + 495 + ], + "type": "inline_equation", + "content": "f_{n,lfs,2}^{m}" + }, + { + "bbox": [ + 56, + 220, + 296, + 495 + ], + "type": "text", + "content": " is a multi-channel spatial map that shares the same layout as " + }, + { + "bbox": [ + 56, + 220, + 296, + 495 + ], + "type": "inline_equation", + "content": "f_{n}^{m}" + }, + { + "bbox": [ + 56, + 220, + 296, + 495 + ], + "type": "text", + "content": ". The LFS adapter outputs " + }, + { + "bbox": [ + 56, + 220, + 296, + 495 + ], + "type": "inline_equation", + "content": "f_{n,lfs}^{m}" + }, + { + "bbox": [ + 56, + 220, + 296, + 495 + ], + "type": "text", + "content": " after " + }, + { + "bbox": [ + 56, + 220, + 296, + 495 + ], + "type": "inline_equation", + "content": "f_{n,lfs,2}^{m}" + }, + { + "bbox": [ + 56, + 220, + 296, + 495 + ], + "type": "text", + "content": " being passed through a convolutional layer followed by one " + }, + { + "bbox": [ + 56, + 220, + 296, + 495 + ], + "type": "inline_equation", + "content": "GELU(\\cdot)" + }, + { + "bbox": [ + 56, + 220, + 296, + 495 + ], + "type": "text", + "content": " activation layer. The LFS adapter provides a localized aperture to detect detailed abnormal frequency distributions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 501, + 149, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 501, + 149, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 501, + 149, + 513 + ], + "type": "text", + "content": "3.5. Loss Functions" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 518, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 295, + 567 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 55, + 518, + 295, + 567 + ], + "type": "inline_equation", + "content": "S_{gt}" + }, + { + "bbox": [ + 55, + 518, + 295, + 567 + ], + "type": "text", + "content": " denote the ground-truth image label where 1 represents the abnormal image and 0 represents the normal image. The loss for the anomaly classification is defined as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 573, + 294, + 614 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 573, + 294, + 614 + ], + "spans": [ + { + "bbox": [ + 58, + 573, + 294, + 614 + ], + "type": "interline_equation", + "content": "L _ {c l s} = - \\frac {1}{N} \\sum_ {N} ^ {n = 1} \\left[ S _ {g t} \\log \\left(S _ {a, n}\\right) + \\left(1 - S _ {g t}\\right) \\log \\left(1 - S _ {a, n}\\right) \\right] \\tag {4}", + "image_path": "72a2eddc42236519d61b47c7dac391612580efff2fcf6f9cd7a230108c53df89.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 615, + 295, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 615, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 55, + 615, + 295, + 669 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 615, + 295, + 669 + ], + "type": "inline_equation", + "content": "\\log" + }, + { + "bbox": [ + 55, + 615, + 295, + 669 + ], + "type": "text", + "content": " denotes the log operation. Similarly, let " + }, + { + "bbox": [ + 55, + 615, + 295, + 669 + ], + "type": "inline_equation", + "content": "M_{gt}" + }, + { + "bbox": [ + 55, + 615, + 295, + 669 + ], + "type": "text", + "content": " denote the ground-truth pixel label of the input image, where 1 means abnormal pixel region and 0 means normal pixel region. The loss for the anomaly segmentation is defined as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 674, + 310, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 674, + 310, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 674, + 310, + 714 + ], + "type": "interline_equation", + "content": "L _ {\\text {m a s k}} = \\frac {1}{N} \\sum_ {N} ^ {n = 1} [ \\operatorname {F o c a l} (U p (M _ {a, n}), M _ {g t}) + D i c e (U p (M _ {a, n}), M _ {g t}) ] \\tag {5}", + "image_path": "17048860b9efbcf70c36c8d4d819271586551c8ccaa78902e41fd0b2bcbf7481.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "Focal(\\cdot, \\cdot)" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "Dice(\\cdot, \\cdot)" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " denote the focal loss [26] and dice loss [23], respectively. The operator " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "Up(\\cdot)" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " represents the unsampling operation. During training, both the visual encoder and the text encoder of CLIP [33] are Frozen. Only the FFE adapters and LFS adapters are optimized by the loss function " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "L_{total} = L_{cls} + L_{mask}" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 156, + 395, + 169 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 395, + 169 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 395, + 169 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 176, + 489, + 188 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 176, + 489, + 188 + ], + "spans": [ + { + "bbox": [ + 313, + 176, + 489, + 188 + ], + "type": "text", + "content": "4.1. Datasets and Evaluation Metrics" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 194, + 555, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 555, + 301 + ], + "type": "text", + "content": "To verify the effectiveness of the proposed method FE-CLIP, we conduct comprehensive experiments across 10 real-world anomaly detection datasets, covering various industrial inspection scenarios and medical imaging domains. In industrial inspection, we use MVtec AD [4], VisA [54], MPDD [18], BTAD [30], DAGM [45], and DTD-Synthetic [3]. In medical imaging, we use colon polyp detection datasets CVC-ClinicDB [6] and Kvasir [19], brain tumor detection datasets BrainMRI [38] and Br35H [15]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 302, + 554, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 302, + 554, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 302, + 554, + 373 + ], + "type": "text", + "content": "As for the evaluation metrics, following previous works [17, 52], we use Area Under the Receiver Operating Characteristic curve (AUROC). Additionally, Average Precision (AP) for Zero-Shot Anomaly Detection (ZSAD) and PRO [5] for Zeor-Shot Anomaly Segmentation (ZSAS) are also used to provide a more in-depth analysis of the results." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 382, + 447, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 382, + 447, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 447, + 395 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 399, + 555, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 555, + 567 + ], + "type": "text", + "content": "We use the publicly available CLIP model (VITL/14@336px) as our backbone. Model parameters of CLIP are all frozen, and only the parameters of FFE adapters and LFS adapters are learnable. Following AnomalyCLIP [52], we fine-tune FE-CLIP using the test data of MVtec AD and evaluate the ZSAD performance on other datasets. As for MVtec AD, we fine-tune FE-CLIP on the test data of VisA. The proposed FE-CLIP is trained by 9 epochs with Adam optimizer. The learning rate is set to 5e-4 and the total batch size is 16. All experiments are conducted in PyTorch-1.13.0 with four NVIDIA RTX 3090 24GB GPUs. All experiment results are averaged under five runs. We report dataset-level results, which are averaged across their respective sub-datasets." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 575, + 399, + 587 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 575, + 399, + 587 + ], + "spans": [ + { + "bbox": [ + 313, + 575, + 399, + 587 + ], + "type": "text", + "content": "4.3. Main Results" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "ZSAD performance on diverse anomaly scenarios. Table 1 shows the ZSAD results of FE-CLIP with other competing methods over six industrial defect datasets and two medical domain datasets. As shown in Table 1, FE-CLIP can achieve higher AUROC and AP for the ZSAD task on most industrial inspection datasets and medical datasets, such as MVTec AD, DTD-Synthetic, and BrainMRI, when compared with other competing methods, such as WinCLIP [17], AnomalyCLIP [52] and AdaCLIP [8]. This shows that (1) the frequency information can benefit the ZSAD task" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21224" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 553, + 178 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 553, + 178 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 553, + 178 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 553, + 178 + ], + "type": "table", + "html": "
MethodsIndustrial DefectsMedical Anomalies
MVTec ADVisAMPDDBTADDAGMDTD-SyntheticBrainMRIBr35H
CLIP [33]74.1 / 87.666.4 / 71.554.3 / 65.434.5 / 52.579.6 / 59.071.6 / 85.773.9 / 81.778.4 / 78.8
CoOp [50]88.8 / 94.862.8 / 68.155.1 / 64.266.8 / 77.487.5 / 74.6- / -61.3 / 44.986.0 / 87.5
WinCLIP [17]91.8 / 96.578.1 / 81.263.6 / 69.968.2 / 70.991.8 / 79.593.2 / 92.686.6 / 91.580.5 / 82.2
APRIL-GAN [9]86.1 / 93.578.0 / 81.478.0 / 81.473.6 / 68.694.4 / 83.886.4 / 95.089.3 / 90.993.1 / 92.9
AnomalyCLIP [52]91.5 / 96.282.1 / 85.477.0 / 82.088.3 / 87.397.5 / 92.393.5 / 97.090.3 / 92.294.6 / 94.7
AdaCLIP [8]89.2 / -85.8 / -76.0 / -88.6 / -99.1 / -95.5 / -94.8 / -97.7 / -
FE-CLIP91.9 / 96.584.6 / 86.678.0 / 82.690.3 / 90.097.5 / 92.398.3 / 99.494.8 / 93.896.8 / 93.8
", + "image_path": "88e54f6673c8da6d807a01d5faebda72997adb2739eb01cc3f9d34d01e05d850.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 60, + 219, + 553, + 336 + ], + "blocks": [ + { + "bbox": [ + 55, + 186, + 555, + 210 + ], + "lines": [ + { + "bbox": [ + 55, + 186, + 555, + 210 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 555, + 210 + ], + "type": "text", + "content": "Table 1. AUROC / AP results for Zero-Shot Anomaly Detection (ZSAD) task on eight real-world anomaly detection datasets. Best results are highlighted in red." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 219, + 553, + 336 + ], + "lines": [ + { + "bbox": [ + 60, + 219, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 60, + 219, + 553, + 336 + ], + "type": "table", + "html": "
MethodsIndustrial DefectsMedical Anomalies
MVTec ADVisAMPDDBTADDAGMDTD-SyntheticCVC-ClinicDBKvasir
CLIP [33]38.4 / 11.346.6 / 14.862.1, 33.030.6 / 4.430.6 / 4.433.9 / 12.547.5 / 18.944.6 / 17.7
CoOp [50]33.3 / 6.724.2 / 3.815.4 / 2.315.4 / 2.317.5 / 2.1- / -17.5 / 2.144.1 / 3.5
WinCLIP [17]85.1 / 64.679.6 / 56.879.6 / 56.872.7 / 27.387.6 / 65.783.9 / 57.851.2 / 13.869.7 / 24.5
APRIL-GAN [9]87.6 / 84.094.2 / 86.894.1 / 83.260.8 / 25.082.4 / 66.295.3 / 86.980.5 / 60.775.0 / 36.2
AnomalyCLIP [52]91.1 / 81.495.5 / 87.096.5 / 88.794.2 / 74.895.6 / 91.097.9 / 92.382.9 / 67.878.9 / 45.6
AdaCLIP [8]88.7 / -95.5 / -96.1 / -92.1 / -91.5 / -97.9 / -84.4 / -- / -
VCP-CLIP [32]92.0 / 87.395.7 / 90.7- / -94.1 / 74.699.4 / 98.3- / -- / -- / -
FE-CLIP92.6 / 88.395.9 / 92.897.0 / 90.595.6 / 80.498.5 / 96.699.0 / 97.484.5 / 69.279.8 / 60.6
", + "image_path": "c57f4830242ac8bbbdc5be1a89ca7c3e40937918fd9fef0a46665eef220eb416.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 342, + 555, + 377 + ], + "lines": [ + { + "bbox": [ + 55, + 342, + 555, + 377 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 555, + 377 + ], + "type": "text", + "content": "Table 2. AUROC / PRO results for Zero-Shot Anomaly Segmentation (ZSAS) task on eight real-world anomaly detection datasets. Best results are highlighted in red. Note that the image-level medical AD datasets do not contain segmentation ground truth, so the pixel-level medical AD datasets are different from the image-level datasets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 398, + 295, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 398, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 398, + 295, + 434 + ], + "type": "text", + "content": "due to the superior performance of FE-CLIP, and (2) the frequency information helps FE-CLIP achieve good generalization across datasets from different domains." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 436, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 436, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 54, + 436, + 295, + 590 + ], + "type": "text", + "content": "ZSAS performance on diverse anomaly scenarios. We show the ZSAS results of FE-CLIP with other competing methods over six industrial defect datasets and two medical domain datasets in Table 2. Table 2 shows that FE-CLIP achieves higher AUROC and PRO for the ZSAS task on almost all of the industrial inspection datasets and medical datasets, except for the DAGM dataset, when compared with other competing methods, such as WinCLIP [17], AnomalyCLIP [52], AdaCLIP [8] and VCP-CLIP [32]. This further demonstrates that (1) the frequency information can benefit the ZSAS task and (2) the frequency information helps FE-CLIP achieve good generalization across datasets from different domains." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": "Discussion. We also note that the FE-CLIP can defeat almost all of the competing methods on the ZSAS task, while FE-CLIP only performs better than other methods over most of the datasets on ZSAD task. The reason may be that the frequency information can be directly injected into the patch tokens (i.e. the feature map used for the ZSAS task) through the FFE adapter and LFS adapter, while the frequency information can only be implicitly involved in the global class token (i.e. the feature vector used for the ZSAD task) from the patch tokens based on the self" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 399, + 403, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 403, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 403, + 408 + ], + "type": "text", + "content": "attention mechanism." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 411, + 555, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 555, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 555, + 459 + ], + "type": "text", + "content": "Overall, the proposed FE-CLIP achieves superior zero-shot performance of detecting and segmenting anomalies in datasets of highly diverse class semantics from various defect inspection and medical imaging domains." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 468, + 408, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 408, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 408, + 481 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 486, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 715 + ], + "type": "text", + "content": "The effectiveness of FFE and LFS We conduct ablation experiments on MVTec AD dataset and VisA dataset to validate the effectiveness of the proposed FFE adapter and LFS adapter. The results are shown in Table 3. The \"CLIP " + }, + { + "bbox": [ + 313, + 486, + 555, + 715 + ], + "type": "inline_equation", + "content": "[33] + \\text{Conv Adapter}\"" + }, + { + "bbox": [ + 313, + 486, + 555, + 715 + ], + "type": "text", + "content": " means we add " + }, + { + "bbox": [ + 313, + 486, + 555, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 486, + 555, + 715 + ], + "type": "text", + "content": " conv adapters, where each of them consists of one convolutional layer followed by one " + }, + { + "bbox": [ + 313, + 486, + 555, + 715 + ], + "type": "inline_equation", + "content": "GELU(\\cdot)" + }, + { + "bbox": [ + 313, + 486, + 555, + 715 + ], + "type": "text", + "content": " activation layer to keep the same learnable parameters with the LFS adapter into the visual encoder of CLIP. It can be concluded that both the FFE adapter and LFS adapter can achieve better performance than the conv adapter in MVTec AD dataset and VisA dataset across the ZSAD task and ZSAS task. This also shows that the frequency information can benefit the ZSAD task and ZSAS task. When combined the FFE adapter with the LFS adapter, the FE-CLIP achieves higher performance than the FFE adapter and LFS adapter. This demonstrates that the FFE adapter and LFS adapter can mine different but complementary frequency information to further boost the performance of model on the ZSAD and ZSAS tasks." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21225" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 70, + 479, + 157 + ], + "blocks": [ + { + "bbox": [ + 132, + 70, + 479, + 157 + ], + "lines": [ + { + "bbox": [ + 132, + 70, + 479, + 157 + ], + "spans": [ + { + "bbox": [ + 132, + 70, + 479, + 157 + ], + "type": "table", + "html": "
MethodsMVTec ADVisA
Pixel-levelImage-levelPixel-levelImage-level
CLIP[33] + Conv Adapter89.0 / 85.187.8 / 94.594.8 / 89.882.3 / 83.5
CLIP[33] + FFE Adapter90.6 / 86.290.0 / 95.495.1 / 91.383.3 / 85.0
CLIP[33] + LFS Adapter91.5 / 87.190.6 / 95.795.5 / 91.883.9 / 85.7
FE-CLIP92.6 / 88.391.9 / 96.595.9 / 92.884.6 / 86.6
", + "image_path": "e3bab8a5ceac6a321f616e8e408d0e1a32fed1ffdfa9a58e13fc88d7235db46a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 205, + 299, + 266 + ], + "blocks": [ + { + "bbox": [ + 55, + 163, + 555, + 187 + ], + "lines": [ + { + "bbox": [ + 55, + 163, + 555, + 187 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 555, + 187 + ], + "type": "text", + "content": "Table 3. Module ablation on the MVTec AD and the VisA datasets. The metric of Pixel-level is AUROC / PRO for ZSAS task, and the metric of Image-level is AUROC / AP for ZSAD task. Best results are highlighted in red." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 205, + 299, + 266 + ], + "lines": [ + { + "bbox": [ + 56, + 205, + 299, + 266 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 299, + 266 + ], + "type": "table", + "html": "
MethodsMVTec ADVisA
Pixel-levelImage-levelPixel-levelImage-level
FFT92.8 / 88.392.0 / 96.295.5 / 93.084.1 / 86.8
DCT92.6 / 88.391.9 / 96.595.9 / 92.884.6 / 86.6
", + "image_path": "9555c9f026c27fb73638766794429890f5e41a7a55540e2c1819321e780ef512.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 272, + 296, + 318 + ], + "lines": [ + { + "bbox": [ + 55, + 272, + 296, + 318 + ], + "spans": [ + { + "bbox": [ + 55, + 272, + 296, + 318 + ], + "type": "text", + "content": "Table 4. Compared with using FFT to extract the frequency information in the FFE and LFS modules. The metric of Pixel-level is AUROC / PRO for ZSAS task, and the metric of Image-level is AUROC / AP for ZSAD task. Best results are highlighted in red." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 59, + 338, + 104, + 386 + ], + "blocks": [ + { + "bbox": [ + 59, + 338, + 104, + 386 + ], + "lines": [ + { + "bbox": [ + 59, + 338, + 104, + 386 + ], + "spans": [ + { + "bbox": [ + 59, + 338, + 104, + 386 + ], + "type": "image", + "image_path": "d6f13747d83cf192722772b460591f1b66576a4774d0cb3f468528e3c50c073d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 59, + 387, + 104, + 432 + ], + "blocks": [ + { + "bbox": [ + 59, + 387, + 104, + 432 + ], + "lines": [ + { + "bbox": [ + 59, + 387, + 104, + 432 + ], + "spans": [ + { + "bbox": [ + 59, + 387, + 104, + 432 + ], + "type": "image", + "image_path": "19181edfeb38d426337f67abc575606f9e563ded07ed5ace7016579f93df73eb.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 433, + 101, + 440 + ], + "lines": [ + { + "bbox": [ + 63, + 433, + 101, + 440 + ], + "spans": [ + { + "bbox": [ + 63, + 433, + 101, + 440 + ], + "type": "text", + "content": "(a) Input Image" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 106, + 338, + 152, + 386 + ], + "blocks": [ + { + "bbox": [ + 106, + 338, + 152, + 386 + ], + "lines": [ + { + "bbox": [ + 106, + 338, + 152, + 386 + ], + "spans": [ + { + "bbox": [ + 106, + 338, + 152, + 386 + ], + "type": "image", + "image_path": "919a719e30bba0e8afc8a81c79f7db0966c64af1ac9771609e1bf42911a92bc0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 106, + 387, + 152, + 432 + ], + "blocks": [ + { + "bbox": [ + 106, + 387, + 152, + 432 + ], + "lines": [ + { + "bbox": [ + 106, + 387, + 152, + 432 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 152, + 432 + ], + "type": "image", + "image_path": "5131f2a5216916bc8b7df84d3705250873a75f7cc4c87f10966a1a98a7f44d48.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 111, + 433, + 147, + 447 + ], + "lines": [ + { + "bbox": [ + 111, + 433, + 147, + 447 + ], + "spans": [ + { + "bbox": [ + 111, + 433, + 147, + 447 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 111, + 433, + 147, + 447 + ], + "type": "inline_equation", + "content": "f_{fe}^{m}" + }, + { + "bbox": [ + 111, + 433, + 147, + 447 + ], + "type": "text", + "content": " from FFE Adapter" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 154, + 338, + 199, + 387 + ], + "blocks": [ + { + "bbox": [ + 154, + 338, + 199, + 387 + ], + "lines": [ + { + "bbox": [ + 154, + 338, + 199, + 387 + ], + "spans": [ + { + "bbox": [ + 154, + 338, + 199, + 387 + ], + "type": "image", + "image_path": "2d4704cb6283321213688f49e1fcf5b6c8a811acb9be816fdb5fea86e934c725.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 154, + 387, + 199, + 432 + ], + "blocks": [ + { + "bbox": [ + 154, + 387, + 199, + 432 + ], + "lines": [ + { + "bbox": [ + 154, + 387, + 199, + 432 + ], + "spans": [ + { + "bbox": [ + 154, + 387, + 199, + 432 + ], + "type": "image", + "image_path": "61cd967df442bba1adb5db52b48daa51e95ab22d3049bf2ad4684b7194503f66.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 433, + 193, + 447 + ], + "lines": [ + { + "bbox": [ + 159, + 433, + 193, + 447 + ], + "spans": [ + { + "bbox": [ + 159, + 433, + 193, + 447 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 159, + 433, + 193, + 447 + ], + "type": "inline_equation", + "content": "f_{N,lf_s}^m" + }, + { + "bbox": [ + 159, + 433, + 193, + 447 + ], + "type": "text", + "content": " from LFS Adapter" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 200, + 338, + 246, + 386 + ], + "blocks": [ + { + "bbox": [ + 200, + 338, + 246, + 386 + ], + "lines": [ + { + "bbox": [ + 200, + 338, + 246, + 386 + ], + "spans": [ + { + "bbox": [ + 200, + 338, + 246, + 386 + ], + "type": "image", + "image_path": "1aa7f5f88c38dc5ab12d1f85613f49cf78fae46ee08df725bb534ffe4d2c00ce.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 200, + 387, + 246, + 432 + ], + "blocks": [ + { + "bbox": [ + 200, + 387, + 246, + 432 + ], + "lines": [ + { + "bbox": [ + 200, + 387, + 246, + 432 + ], + "spans": [ + { + "bbox": [ + 200, + 387, + 246, + 432 + ], + "type": "image", + "image_path": "e7f85ba6cab1ad3ddce06249e23fbd3b3b9a739c90497e16a7968c548736d774.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 208, + 433, + 238, + 447 + ], + "lines": [ + { + "bbox": [ + 208, + 433, + 238, + 447 + ], + "spans": [ + { + "bbox": [ + 208, + 433, + 238, + 447 + ], + "type": "text", + "content": "(d) " + }, + { + "bbox": [ + 208, + 433, + 238, + 447 + ], + "type": "inline_equation", + "content": "\\hat{f}_N^m" + }, + { + "bbox": [ + 208, + 433, + 238, + 447 + ], + "type": "text", + "content": " from equation (2)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 456, + 295, + 478 + ], + "lines": [ + { + "bbox": [ + 55, + 456, + 295, + 478 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 295, + 478 + ], + "type": "text", + "content": "Figure 5. The visualization of the learned frequency features. Please zoom in for a better view." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 246, + 338, + 293, + 386 + ], + "blocks": [ + { + "bbox": [ + 246, + 338, + 293, + 386 + ], + "lines": [ + { + "bbox": [ + 246, + 338, + 293, + 386 + ], + "spans": [ + { + "bbox": [ + 246, + 338, + 293, + 386 + ], + "type": "image", + "image_path": "8dacbcc1c5ca34bb5c9e58ea87bc5dfde841ac5e0ddc160659a58225a04e3901.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 246, + 387, + 293, + 432 + ], + "blocks": [ + { + "bbox": [ + 246, + 387, + 293, + 432 + ], + "lines": [ + { + "bbox": [ + 246, + 387, + 293, + 432 + ], + "spans": [ + { + "bbox": [ + 246, + 387, + 293, + 432 + ], + "type": "image", + "image_path": "d9d43740d3f9c1df4bb794417b67fcfcae4ceb31aa44f2b19cc30e03997fc358.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 258, + 433, + 282, + 446 + ], + "lines": [ + { + "bbox": [ + 258, + 433, + 282, + 446 + ], + "spans": [ + { + "bbox": [ + 258, + 433, + 282, + 446 + ], + "type": "text", + "content": "(e) Final prediction" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": "Would FFT work instead of DCT We attempt to use FFT to replace the DCT in the FFE and LFS modules to extract the frequency information. However, there are real values and imaginary values after FFT. To deal with this, the real values and imaginary values are separately processed and then combined to transform back to the spatial domain in the FFE module. As for the LFS module, the averaged local frequency statistics are computed from the real values and imaginary values, respectively, to get the averaged real local frequency statistics and the averaged imaginary local frequency statistics, and then they are summed to get the final averaged local frequency statistics. The results are shown in Table 4. This shows that using FFT achieves similar results with DCT, which means that the key is to inject the frequency information rather than the way of frequency transformation. We choose DCT since DCT is more convenient to extract frequency." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 207, + 376, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 207, + 376, + 220 + ], + "spans": [ + { + "bbox": [ + 313, + 207, + 376, + 220 + ], + "type": "text", + "content": "4.5. Analysis" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 226, + 554, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 226, + 554, + 298 + ], + "spans": [ + { + "bbox": [ + 313, + 226, + 554, + 298 + ], + "type": "text", + "content": "Time cost. We report the inference time of APRIL-GAN, AnomalyCLIP, VCP-CLIP, AdaCLIP, and FE-CLIP with PyTorch-1.13.0 on one NVIDIA RTX 3090 24GB GPU. The inference time is 105, 120, 128, 160, and " + }, + { + "bbox": [ + 313, + 226, + 554, + 298 + ], + "type": "inline_equation", + "content": "125\\mathrm{ms}" + }, + { + "bbox": [ + 313, + 226, + 554, + 298 + ], + "type": "text", + "content": ", respectively. FE-CLIP achieves a better trade-off between performance and speed." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 300, + 555, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 555, + 434 + ], + "type": "text", + "content": "Visualization of learned frequency features. To show why frequency information is effective for the ZSAD task, we visualize the frequency-aware features from the last block of the visual encoder of CLIP as example, where the FFE and LFS adapters are added. As shown in Figure 5, both the features " + }, + { + "bbox": [ + 313, + 300, + 555, + 434 + ], + "type": "inline_equation", + "content": "f_{N,ffe}^{m}" + }, + { + "bbox": [ + 313, + 300, + 555, + 434 + ], + "type": "text", + "content": " from FFE adapter and the features " + }, + { + "bbox": [ + 313, + 300, + 555, + 434 + ], + "type": "inline_equation", + "content": "f_{N,lfs}^{m}" + }, + { + "bbox": [ + 313, + 300, + 555, + 434 + ], + "type": "text", + "content": " from LFS adapter highlight the abnormal region of images. By injecting the features " + }, + { + "bbox": [ + 313, + 300, + 555, + 434 + ], + "type": "inline_equation", + "content": "f_{N,ffe}^{m}" + }, + { + "bbox": [ + 313, + 300, + 555, + 434 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 300, + 555, + 434 + ], + "type": "inline_equation", + "content": "f_{N,lfs}^{m}" + }, + { + "bbox": [ + 313, + 300, + 555, + 434 + ], + "type": "text", + "content": " into the visual encoder of CLIP, the obtained features " + }, + { + "bbox": [ + 313, + 300, + 555, + 434 + ], + "type": "inline_equation", + "content": "\\hat{f}_N^m" + }, + { + "bbox": [ + 313, + 300, + 555, + 434 + ], + "type": "text", + "content": " can successfully localize the main abnormal region, in favor of the final prediction of the model." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 435, + 554, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 554, + 555 + ], + "type": "text", + "content": "When frequency information can be beneficial. Some properties of anomaly area, such as appearance defects, can be reflected based on its frequency information, since high-frequency components represent fine details (e.g. soft borders) that often exist in anomaly area. In these cases, These specific frequency patterns could be beneficial for anomaly detection. For example, as shown in Figure 1, the frequency statistics in the red region are different from the frequency statistics in the surrounding regions of the input abnormal image and the corresponding region of the normal image." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 556, + 554, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 554, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 554, + 664 + ], + "type": "text", + "content": "Let's consider images of normal dumplings and images of squashed dumplings, where squashed dumplings are anomalies. In this case, if the squashed dumplings reveal the meat filling, the frequency of squashed dumplings could still stand out and frequency information could be helpful to detect anomalies. However, if the dumplings are squashed to be more like a meat pie without revealing the meat filling, the frequency of this squashed dumpling might not stand out and frequency information might not work well." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "content": "We validate the FE-CLIP across different domains with 10 datasets of highly diverse class semantics from various industrial inspections and medical domains, and the results of FE-CLIP on the datasets of medical domains are trained" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21226" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 126, + 69, + 485, + 324 + ], + "blocks": [ + { + "bbox": [ + 126, + 69, + 485, + 324 + ], + "lines": [ + { + "bbox": [ + 126, + 69, + 485, + 324 + ], + "spans": [ + { + "bbox": [ + 126, + 69, + 485, + 324 + ], + "type": "table", + "html": "
SetupMethodsMVTec ADVisA
Pixel-levelImage-levelPixel-levelImage-level
1-shotPatchCore [37]92.0 / 79.783.4 / 92.295.4 / 80.579.9 / 82.8
WinCLIP [17]95.2 / 87.193.1 / 96.596.4 / 85.183.8 / 85.1
AnomalyGPT [14]95.3 / -94.1 / -96.2 / -87.4 / -
PromptAD [24]95.9 / -94.6 / -96.7 / -86.9 / -
FE-CLIP96.1 / 90.995.6 / 98.197.9 / 93.988.7 / 90.5
2-shotPatchCore [37]93.3 / 82.386.3 / 93.896.1 / 82.681.6 / 84.8
WinCLIP [17]96.0 / 88.494.4 / 97.096.8 / 86.284.6 / 85.8
AnomalyGPT [14]95.6 / -95.5 / -96.4 / -88.6 / -
PromptAD [24]96.2 / -95.7 / -97.1 / -88.3 / -
InCTRL [53]- / -94.0 / 96.9- / -85.8 / 87.7
FE-CLIP96.5 / 91.596.2 / 98.398.2 / 94.390.3 / 91.6
4-shotPatchCore [37]94.3 / 84.388.8 / 94.596.8 / 84.985.3 / 87.5
WinCLIP [17]96.2 / 89.095.2 / 97.397.2 / 87.687.3 / 88.8
AnomalyGPT [14]96.2 / -96.3 / -96.7 / -90.6 / -
PromptAD [24]96.5 / -96.6 / -97.4 / -89.1 / -
InCTRL [53]- / -94.5 / 97.2- / -87.7 / 90.2
FE-CLIP96.7 / 91.796.6 / 98.498.4 / 95.090.7 / 91.9
", + "image_path": "f2dc48cc22edaa802f805f5bb14bad245619d22856f888d906f56cd06a4950e7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 332, + 555, + 367 + ], + "lines": [ + { + "bbox": [ + 54, + 332, + 555, + 367 + ], + "spans": [ + { + "bbox": [ + 54, + 332, + 555, + 367 + ], + "type": "text", + "content": "Table 5. Compared with other competing methods under few-normal-shot anomaly detection setting. The metric of Pixel-level is AUROC / PRO for the few-normal-shot anomaly segmentation task, and the metric of Image-level is AUROC / AP for the few-normal-shot anomaly detection task. Best results are highlighted in red." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 380, + 298, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 380, + 298, + 441 + ], + "spans": [ + { + "bbox": [ + 54, + 380, + 298, + 441 + ], + "type": "text", + "content": "by the industrial inspection MVTec AD dataset, as shown in Table 1 and Table 2. The superior performance of FE-CLIP demonstrates that frequency information could be beneficial for many cases of anomaly detection, and the proposed FE-CLIP has good generalization across different domains." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 453, + 235, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 235, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 235, + 466 + ], + "type": "text", + "content": "4.6. FE-CLIP with Few Normal Shots" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 472, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 472, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 56, + 472, + 295, + 664 + ], + "type": "text", + "content": "The performance of FE-CLIP can also be further boosted when equipped with few-shot normal images, just like WinCLIP [17] and AnomalyGPT [14]. We adopt the same way as AnomalyGPT to compute the anomaly segmentation map with few-shot normal images. In short words, we use the same visual encoder of FE-CLIP to extract intermediate patch-level features from normal images and store them in memory banks, then calculate the cosine similarity distance between each patch token of input image and its most similar counterpart in the memory banks to obtain the anomaly segmentation map " + }, + { + "bbox": [ + 56, + 472, + 295, + 664 + ], + "type": "inline_equation", + "content": "S_{a,few\\_shot}" + }, + { + "bbox": [ + 56, + 472, + 295, + 664 + ], + "type": "text", + "content": ". Finally, we add the anomaly segmentation map " + }, + { + "bbox": [ + 56, + 472, + 295, + 664 + ], + "type": "inline_equation", + "content": "S_{a}" + }, + { + "bbox": [ + 56, + 472, + 295, + 664 + ], + "type": "text", + "content": " from the FE-CLIP model with " + }, + { + "bbox": [ + 56, + 472, + 295, + 664 + ], + "type": "inline_equation", + "content": "S_{a,few\\_shot}" + }, + { + "bbox": [ + 56, + 472, + 295, + 664 + ], + "type": "text", + "content": " to get the final anomaly segmentation map, and add the anomaly score " + }, + { + "bbox": [ + 56, + 472, + 295, + 664 + ], + "type": "inline_equation", + "content": "M_{a}" + }, + { + "bbox": [ + 56, + 472, + 295, + 664 + ], + "type": "text", + "content": " from the FE-CLIP model with the max value of " + }, + { + "bbox": [ + 56, + 472, + 295, + 664 + ], + "type": "inline_equation", + "content": "S_{a,few\\_shot}" + }, + { + "bbox": [ + 56, + 472, + 295, + 664 + ], + "type": "text", + "content": " to get the final anomaly score. Please refer to [14] for more technical details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "We compare the FE-CLIP with other competing methods, such as AnomalyGPT [14], PromptAD [24], and InCTRL [53], under the few-shot anomaly detection setting. As shown in Table 5, the FE-CLIP outperforms all other com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 380, + 555, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 555, + 441 + ], + "type": "text", + "content": "peting methods across different metrics, including AUROC and PRO metrics for the anomaly detection task, and AUROC and AP metrics for the anomaly segmentation task, on both datasets. The experiments under few-normal-shot setting further demonstrate the superiority of the FE-CLIP." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 459, + 388, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 388, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 388, + 471 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 481, + 555, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 481, + 555, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 481, + 555, + 673 + ], + "type": "text", + "content": "In this paper, we find that frequency information can benefit the ZSAD task. To this end, we propose FE-CLIP, which adds the proposed FFE and LFS adapters into the visual encoder of CLIP, to deeply mine the frequency patterns for the visual features. Extensive experiments on 10 datasets of highly diverse class semantics from various defect inspections and medical domains demonstrate that (1) the frequency information can benefit the ZSAD task, and (2) the frequency information can help the model have good generalization across datasets from different domains. We also conduct experiments under the few-normal-shot anomaly detection setting to further validate the superiority of the proposed FE-CLIP. We hope the community of anomaly detection will be aware of the effectiveness of frequency information from our work and explore more ways to utilize the frequency information for anomaly detection in the future." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 677, + 417, + 689 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 417, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 417, + 689 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 691, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 691, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 691, + 555, + 715 + ], + "type": "text", + "content": "This work was supported by the Anhui Provincial Science and Technology Major Project (No. 2023z020006)." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "21227" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "text", + "content": "[1] Nasir Ahmed, T. Natarajan, and Kamisetty R Rao. Discrete cosine transform. IEEE transactions on Computers, 100(1): 90-93, 1974. 2, 4, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 296, + 190 + ], + "type": "text", + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 191, + 295, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 191, + 295, + 246 + ], + "spans": [ + { + "bbox": [ + 62, + 191, + 295, + 246 + ], + "type": "text", + "content": "[3] Toshimichi Aota, Lloyd Teh Tzer Tong, and Takayuki Okatani. Zero-shot versus many-shot: Unsupervised texture anomaly detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5564-5572, 2023. 3, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 247, + 295, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 295, + 302 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 295, + 302 + ], + "type": "text", + "content": "[4] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad-a comprehensive real-world dataset for unsupervised anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9592-9600, 2019. 1, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 302, + 295, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 302, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 62, + 302, + 295, + 357 + ], + "type": "text", + "content": "[5] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Uninformed students: Student-teacher anomaly detection with discriminative latent embeddings. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4183–4192, 2020. 1, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 358, + 295, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 358, + 295, + 423 + ], + "spans": [ + { + "bbox": [ + 62, + 358, + 295, + 423 + ], + "type": "text", + "content": "[6] Jorge Bernal, F Javier Sánchez, Gloria Fernández-Esparrach, Debora Gil, Cristina Rodríguez, and Fernando Vilarino. Wm-dova maps for accurate polyp highlighting in colonoscopy: Validation vs. saliency maps from physicians. Computerized medical imaging and graphics, 43:99-111, 2015. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 424, + 295, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 424, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 62, + 424, + 295, + 468 + ], + "type": "text", + "content": "[7] Tri Cao, Jiawen Zhu, and Guansong Pang. Anomaly detection under distribution shift. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6511-6523, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 469, + 295, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 469, + 295, + 523 + ], + "spans": [ + { + "bbox": [ + 62, + 469, + 295, + 523 + ], + "type": "text", + "content": "[8] Yunkang Cao, Jiangning Zhang, Luca Frittoli, Yuqi Cheng, Weiming Shen, and Giacomo Boracchi. Adaclip: Adapting clip with hybrid learnable prompts for zero-shot anomaly detection. In European Conference on Computer Vision, 2024. 1, 2, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 525, + 295, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 525, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 62, + 525, + 295, + 579 + ], + "type": "text", + "content": "[9] Xuhai Chen, Yue Han, and Jiangning Zhang. A zero-/few-shot anomaly classification and segmentation method for cvpr 2023 vand workshop challenge tracks 1&2: 1st place on zero-shot ad and 4th place on few-shot ad. arXiv preprint arXiv:2305.17382, 2023. 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 580, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 580, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 580, + 295, + 645 + ], + "type": "text", + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 647, + 295, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 691 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 691 + ], + "type": "text", + "content": "[11] Tharindu Fernando, Harshala Gammulle, Simon Denman, Sridha Sridharan, and Clinton Fookes. Deep learning for medical anomaly detection-a survey. ACM Computing Surveys (CSUR), 54(7):1-37, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[12] Gabriel Goh, Nick Cammarata, Chelsea Voss, Shan Carter, Michael Petrov, Ludwig Schubert, Alec Radford, and Chris" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Olah. Multimodal neurons in artificial neural networks. Distill, 6(3):e30, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "text", + "content": "[13] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 141, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 196 + ], + "type": "text", + "content": "[14] Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Ming Tang, and Jinqiao Wang. Anomalygpt: Detecting industrial anomalies using large vision-language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1932-1940, 2024. 3, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 198, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 229 + ], + "type": "text", + "content": "[15] A. Hamada. Br35h: Brain tumor detection 2020. In https://www.kaggle.com/datasets/ahmedhamada0/brain-tumor-detection, 2020. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 232, + 553, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 232, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 553, + 275 + ], + "type": "text", + "content": "[16] Chaoqin Huang, Haoyan Guan, Aofan Jiang, Ya Zhang, Michael Spratling, and Yan-Feng Wang. Registration based few-shot anomaly detection. In European Conference on Computer Vision, pages 303-319. Springer, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 277, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 277, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 277, + 553, + 342 + ], + "type": "text", + "content": "[17] Jongheon Jeong, Yang Zou, Taewan Kim, Dongqing Zhang, Avinash Ravichandran, and Onkar Dabeer. Winclip: Zero/few-shot anomaly classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19606-19616, 2023. 1, 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 344, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 344, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 344, + 553, + 409 + ], + "type": "text", + "content": "[18] Stepan Jezek, Martin Jonak, Radim Burget, Pavel Dvorak, and Milos Skotak. Deep learning-based defect detection of metal parts: evaluating current methods in complex conditions. In 2021 13th International congress on ultra modern telecommunications and control systems and workshops (ICUMT), pages 66-71. IEEE, 2021. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 411, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 477 + ], + "type": "text", + "content": "[19] Debesh Jha, Pia H Smedsrud, Michael A Riegler, Pål Halvorsen, Thomas De Lange, Dag Johansen, and Håvard D Johansen. Kvasir-seg: A segmented polyp dataset. In MultiMedia Modeling: 26th International Conference, MMM 2020, Daejeon, South Korea, January 5–8, 2020, Proceedings, Part II 26, pages 451–462. Springer, 2020. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 478, + 553, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 543 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 543 + ], + "type": "text", + "content": "[20] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 544, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 553, + 600 + ], + "type": "text", + "content": "[21] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "type": "text", + "content": "[22] Shengze Li, Jianjian Cao, Peng Ye, Yuhan Ding, Chongjun Tu, and Tao Chen. Clipsam: Clip and sam collaboration for zero-shot anomaly segmentation. arXiv preprint arXiv:2401.12665, 2024. 1, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 679 + ], + "type": "text", + "content": "[23] Xiaoya Li, Xiaofei Sun, Yuxian Meng, Junjun Liang, Fei Wu, and Jiwei Li. Dice loss for data-imbalanced nlp tasks. arXiv preprint arXiv:1911.02855, 2019. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[24] Xiaofan Li, Zhizhong Zhang, Xin Tan, Chengwei Chen, Yanyun Qu, Yuan Xie, and Lizhuang Ma. Promptad: Learning prompts with only normal samples for few-shot anomaly" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21228" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "text", + "content": "detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 295, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 295, + 151 + ], + "type": "text", + "content": "[25] Yuanze Li, Haolin Wang, Shihao Yuan, Ming Liu, Debin Zhao, Yiwen Guo, Chen Xu, Guangming Shi, and Wangmeng Zuo. Myriad: Large multimodal model by applying vision experts for industrial anomaly detection. arXiv preprint arXiv:2310.19070, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 295, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 295, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 295, + 196 + ], + "type": "text", + "content": "[26] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 295, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 295, + 262 + ], + "type": "text", + "content": "[27] Jie Liu, Yixiao Zhang, Jie-Neng Chen, Junfei Xiao, Yongyi Lu, Bennett A Landman, Yixuan Yuan, Alan Yuille, Yucheng Tang, and Zongwei Zhou. Clip-driven universal model for organ segmentation and tumor detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21152-21164, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 295, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 295, + 308 + ], + "type": "text", + "content": "[28] Philipp Liznerski, Lukas Ruff, Robert A Vandermeulen, Billy Joe Franks, Marius Kloft, and Klaus-Robert Müller. Explainable deep one-class classification. arXiv preprint arXiv:2007.01760, 2020. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 309, + 295, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 295, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 295, + 363 + ], + "type": "text", + "content": "[29] Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, and Aniruddha Kembhavi. Unified-io: A unified model for vision, language, and multi-modal tasks. In The Eleventh International Conference on Learning Representations, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 295, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 295, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 295, + 421 + ], + "type": "text", + "content": "[30] Pankaj Mishra, Riccardo Verk, Daniele Fornasier, Claudio Piciarelli, and Gian Luca Foresti. Vt-adj: A vision transformer network for image anomaly detection and localization. In 2021 IEEE 30th International Symposium on Industrial Electronics (ISIE), pages 01-06. IEEE, 2021. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 295, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 295, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 295, + 465 + ], + "type": "text", + "content": "[31] Ziyuan Qin, Huahui Yi, Qicheng Lao, and Kang Li. Medical image understanding with pretrained vision language models: A comprehensive study. arXiv preprint arXiv:2209.15517, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 467, + 295, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 295, + 521 + ], + "type": "text", + "content": "[32] Zhen Qu, Xian Tao, Mukesh Prasad, Fei Shen, Zhengtao Zhang, Xinyi Gong, and Guiguang Ding. Vcp-clip: A visual context prompting model for zero-shot anomaly segmentation. In European Conference on Computer Vision, 2024. 1, 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 523, + 295, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 295, + 588 + ], + "type": "text", + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 590, + 295, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 295, + 655 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 295, + 655 + ], + "type": "text", + "content": "[34] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18082-18091, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 657, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 657, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 657, + 295, + 690 + ], + "type": "text", + "content": "[35] Tal Reiss and Yedid Hoshen. Mean-shifted contrastive loss for anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2155-2162, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 692, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 295, + 714 + ], + "type": "text", + "content": "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "text", + "content": "synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 108, + 555, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 555, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 555, + 163 + ], + "type": "text", + "content": "[37] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14318-14328, 2022. 1, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 164, + 554, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 164, + 554, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 164, + 554, + 228 + ], + "type": "text", + "content": "[38] Mohammadreza Salehi, Niousha Sadjadi, Soroosh Baselizadeh, Mohammad H Rohban, and Hamid R Rabiee. Multiresolution knowledge distillation for anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14902-14912, 2021. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 230, + 554, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 554, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 554, + 297 + ], + "type": "text", + "content": "[39] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 298, + 554, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 554, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 554, + 352 + ], + "type": "text", + "content": "[40] Rohan Taori, Achal Dave, Vaishaal Shankar, Nicholas Carlini, Benjamin Recht, and Ludwig Schmidt. Measuring robustness to natural distribution shifts in image classification. Advances in Neural Information Processing Systems, 33:18583-18599, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 354, + 554, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 554, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 554, + 408 + ], + "type": "text", + "content": "[41] Yoad Tewel, Yoav Shalev, Idan Schwartz, and Lior Wolf. Zerocap: Zero-shot image-to-text generation for visual-semantic arithmetic. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17918-17928, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 410, + 554, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 410, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 410, + 554, + 498 + ], + "type": "text", + "content": "[42] Yu Tian, Guansong Pang, Fengbei Liu, Yuanhong Chen, Seon Ho Shin, Johan W Verjans, Rajvinder Singh, and Gustavo Carneiro. Constrained contrastive distribution learning for unsupervised anomaly detection and localisation in medical images. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27-October 1, 2021, Proceedings, Part V 24, pages 128-140. Springer, 2021. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 499, + 554, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 499, + 554, + 554 + ], + "spans": [ + { + "bbox": [ + 316, + 499, + 554, + 554 + ], + "type": "text", + "content": "[43] Yu Tian, Fengbei Liu, Guansong Pang, Yuanhong Chen, Yuyuan Liu, Johan W Verjans, Rajvinder Singh, and Gustavo Carneiro. Self-supervised pseudo multi-class pre-training for unsupervised anomaly detection and segmentation in medical images. Medical image analysis, 90:102930, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 555, + 554, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 554, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 554, + 600 + ], + "type": "text", + "content": "[44] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 601, + 554, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 554, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 554, + 634 + ], + "type": "text", + "content": "[45] Matthias Wieler and Tobias Hahn. Weakly supervised learning for industrial optical inspection. In DAGM symposium in, page 11, 2007. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 635, + 554, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 554, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 554, + 690 + ], + "type": "text", + "content": "[46] Mengde Xu, Zheng Zhang, Fangyun Wei, Yutong Lin, Yue Cao, Han Hu, and Xiang Bai. A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In European Conference on Computer Vision, pages 736-753. Springer, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 691, + 554, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 554, + 714 + ], + "type": "text", + "content": "[47] Zhiyuan You, Lei Cui, Yujun Shen, Kai Yang, Xin Lu, Yu Zheng, and Xinyi Le. A unified model for multi-class" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "21229" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 441 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "anomaly detection. Advances in Neural Information Processing Systems, 35:4571-4584, 2022. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "type": "text", + "content": "[48] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16793-16803, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 196 + ], + "type": "text", + "content": "[49] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In European Conference on Computer Vision, pages 696-712. Springer, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 239 + ], + "type": "text", + "content": "[50] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 294, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 296 + ], + "type": "text", + "content": "[51] Qihang Zhou, Shibo He, Haoyu Liu, Tao Chen, and Jiming Chen. Pull & push: Leveraging differential knowledge distillation for efficient unsupervised anomaly detection and localization. IEEE Transactions on Circuits and Systems for Video Technology, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 342 + ], + "type": "text", + "content": "[52] Qihang Zhou, Guansong Pang, Yu Tian, Shibo He, and Jiming Chen. Anomalyclip: Object-agnostic prompt learning for zero-shot anomaly detection. The thirteenth International Conference on Learning Representations, 2024. 1, 3, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 343, + 294, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 294, + 386 + ], + "type": "text", + "content": "[53] Jiawen Zhu and Guansong Pang. Toward generalist anomaly detection via in-context residual learning with few-shot sample prompts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 441 + ], + "type": "text", + "content": "[54] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. In European Conference on Computer Vision, pages 392-408. Springer, 2022. 5" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21230" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/FED-PsyAU_ Privacy-Preserving Micro-Expression Recognition via Psychological AU Coordination and Dynamic Facial Motion Modeling/e27ad4a5-3106-4216-bcea-a95709cd5f65_content_list.json b/2025/FED-PsyAU_ Privacy-Preserving Micro-Expression Recognition via Psychological AU Coordination and Dynamic Facial Motion Modeling/e27ad4a5-3106-4216-bcea-a95709cd5f65_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d805f87b3ab8c9f6ec46a462307cfdb8be001004 --- /dev/null +++ b/2025/FED-PsyAU_ Privacy-Preserving Micro-Expression Recognition via Psychological AU Coordination and Dynamic Facial Motion Modeling/e27ad4a5-3106-4216-bcea-a95709cd5f65_content_list.json @@ -0,0 +1,1830 @@ +[ + { + "type": "text", + "text": "FED-PsyAU: Privacy-Preserving Micro-Expression Recognition via Psychological AU Coordination and Dynamic Facial Motion Modeling", + "text_level": 1, + "bbox": [ + 145, + 130, + 851, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jingting Li $^{1,2}$ , Yu Qian $^{1,3}$ , Lin Zhao $^{1,2}$ , Su-Jing Wang $^{*1,2}$", + "bbox": [ + 267, + 202, + 730, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ State Key Laboratory of Cognitive Science and Mental Health, Institute of Psychology, Chinese Academy of Sciences, Beijing, 100101, China", + "bbox": [ + 153, + 220, + 844, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Department of Psychology, University of the Chinese Academy of Sciences, Beijing, 100049, China", + "bbox": [ + 102, + 256, + 897, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ School of Computer, Jiangsu University of Science and Technology, Zhenjiang, 212100, China", + "bbox": [ + 122, + 273, + 875, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{\\dagger}$ Equal contribution, \\*Corresponding author (wangsujing@psych.ac.cn)", + "bbox": [ + 202, + 294, + 787, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 344, + 326, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Micro-expressions (MEs) are brief, low-intensity, often localized facial expressions. They could reveal genuine emotions individuals may attempt to conceal, valuable in contexts like criminal interrogation and psychological counseling. However, ME recognition (MER) faces challenges, such as small sample sizes and subtle features, which hinder efficient modeling. Additionally, real-world applications encounter ME data privacy issues, leaving the task of enhancing recognition across settings under privacy constraints largely unexplored. To address these issues, we propose a FED-PsyAU research framework. We begin with a psychological study on the coordination of upper and lower facial action units (AUs) to provide structured prior knowledge of facial muscle dynamics. We then develop a DPK-GAT network that combines these psychological priors with statistical AU patterns, enabling hierarchical learning of facial motion features from regional to global levels, effectively enhancing MER performance. Additionally, our federated learning (FL) framework advances MER capabilities across multiple clients without data sharing, preserving privacy and alleviating the limited-sample issue for each client. Extensive experiments on commonly-used ME databases demonstrate the effectiveness of our approach. Our implementation is publicly available at https://github.com/MELABIPCAS/FED-PsyAU.git.", + "bbox": [ + 89, + 376, + 482, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 782, + 220, + 799 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Micro-expressions (MEs) [9], resulting from conflicts between voluntary and involuntary expressions, reveal true emotions, making them valuable in non-contact emotion detection applications like criminal interrogation and psychological counseling. MEs are characterized by their brief duration, low intensity, localized occurrence, and occasional", + "bbox": [ + 89, + 809, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "asymmetry [10]. These traits not only restrict large-scale data collection and annotation but also make it challenging for deep learning networks to effectively model their motion characteristics. Some methods focus on regions of interest (ROIs) [4, 21, 31, 49], while others utilize attention mechanisms [3, 14, 33] and transformer-based architectures [24, 50, 57] to capture salient features. Meanwhile, MEs are often complex manifestations involving the combination of multiple AUs. In contrast, individual action units (AUs) correspond to specific, anatomically-based facial muscle movements. Focusing on AUs allows us to model more fundamental and consistent facial appearance changes. Certain approaches use AU labels to construct adjacency matrices for graph convolutional networks (GCNs) [24, 39]. Yet, to the best of our knowledge, these methods often apply AU annotations directly without fully exploiting the anatomical structure and coordinated AU dynamics, limiting the network's understanding on ME.", + "bbox": [ + 511, + 345, + 903, + 618 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address these challenges, we leverage the theory that distinct neural pathways control upper and lower facial movements, resulting in two different motion patterns. Based on this, we conduct a psychological study to investigate upper/lower facial AU coordination, providing crucial prior knowledge on structural motion relationships underpinning MEs. Using both psychological priors and data-driven AU patterns, we design a Graph Attention Network (GAT) [47] based on dynamic prior knowledge, applied separately to the upper and lower regions. A final global GAT captures the topological relationships of muscle movements across the face. The resulting feature, combined with full-face optical flow (OF), is input into a subsequent dual-stream network, effectively integrating structural and dynamic information to enhance ME recognition (MER).", + "bbox": [ + 511, + 622, + 903, + 849 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Besides, as mentioned earlier, although MEs have valuable applications in many privacy-sensitive contexts, the sensitivity of biometric data makes it impractical to aggre", + "bbox": [ + 511, + 854, + 903, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14453", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "gate ME data across clients for training. Meanwhile, the sample size in a single client is often insufficient to train a robust model. Yet, the challenge of enhancing MER performance across diverse clients while preserving privacy remains largely unexplored. To address this, we adopt a federated learning (FL) framework, improving performance by merging and updating model parameters across clients without exchanging local data. This approach preserves data privacy and supports the MER practical deployment.", + "bbox": [ + 89, + 90, + 480, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Overall, our work leverages psychological and data-driven priors to enhance the network's learning of structured ME motion patterns, particularly in limited-sample conditions. Additionally, we introduce FL to advance MER in real-world scenarios. The contributions of this paper are:", + "bbox": [ + 89, + 227, + 480, + 305 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We conduct a psychological experiment to derive AU coordination between the upper and lower facial regions, providing structural priors for ME and facial expression analysis in computer vision.", + "- We design a local-to-global GAT network that integrates psychological priors and data-driven AU patterns in the upper and lower facial regions, improving the network's ability to learn both local and global structural facial dynamics. This feature, combined with global OF feature in a dual-stream network, enhances MER performance.", + "- We employ FL in MER to improve data privacy and overcome small sample size challenges by aggregating features from distributed data-protected sources." + ], + "bbox": [ + 91, + 306, + 482, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 518, + 241, + 535 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Micro-expression Recognition", + "text_level": 1, + "bbox": [ + 89, + 545, + 356, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In recent years, MER has gained increasing attention from computer vision researchers. Traditional machine learning methods primarily rely on handcrafted feature extraction. For instance, Local Binary Pattern (LBP) [43], effectively utilizes local texture features, has inspired various improvements [18, 49]. Furthermore, OF [26] has been explored for capturing object motion. Liu et al. [38] developed the Main Directional Mean OF (MDMO) method by aligning faces in the OF domain, while Liong et al. [34] added optical strain features for MER. However, traditional approaches depend heavily on manually crafted features, limiting both accuracy and generalizability. Meanwhile, deep learning methods, with robust feature extraction capabilities, can automatically learn multi-level representations to capture subtle ME changes. The availability of several spontaneous ME databases, such as CASME II [51], SAMM [6], CAS(ME) $^{3}$ [27], DFME [59] and others [1, 30, 32, 44, 54], has accelerated deep learning advancements in MER. Inspired by Liong et al.'s work [35] on sequence redundancy, many studies [4, 13, 50] now use keyframes (onset and apex) as input, reducing model parameters while achieving excellent results.", + "bbox": [ + 89, + 568, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite advancements in preprocessing and feature extraction, deep learning alone has yet to achieve high MER accuracy. As MEs stem from physiological muscle activity, incorporating prior knowledge offers meaningful context. FACS [10], which defines expressions through AUs, provides a basic theoretical framework for MER. Recent studies explore AU-ME relationships to enhance MER performance. Lei et al. [24] applied GCN [21] to model AU co-occurrences, especially around the eyebrow and mouth regions. Xie et al. [53] introduced an AU-assisted GCN, with modules for AU feature extraction and ME image generation. Wang et al. [48] used transformers to learn dynamic AU adjacency, boosting model generalization. These networks typically rely on AU labels from databases, without fully integrating structured facial muscle movements, AU interrelations, or their links to expressions. In this work, we refine structured ME feature extraction using a local-to-global GAT network, which integrates priors on facial AU coordination from psychological experiments with data-driven movement patterns.", + "bbox": [ + 511, + 90, + 903, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Federated Learning", + "text_level": 1, + "bbox": [ + 513, + 406, + 702, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To assess MER across databases, ME Grand Challenge (MEGC) 2018 [56] introduced the Holdout-database Evaluation and Composite Database Evaluation tasks, utilizing the CASME II and SAMM databases. MEGC 2019 [45] expanded the CDE task to include the SMIC databases. While research on MEs predominantly utilizes these public datasets, existing methodologies do not address the privacy concerns that arise in practical applications.", + "bbox": [ + 511, + 429, + 903, + 550 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "FL [41] could addresses ME data scarcity and privacy concerns through collaborative training. It includes two main phases: local training and global aggregation. Many existing FL works for expressions, while insightful, are often not open-source or use highly customized frameworks (e.g., non-random client data allocation [12], specialized parameter aggregation [11]) making standard methods more suitable for implementation and comparisons in our study. For instance, McMahan et al. [41] proposed FedAvg, a basic aggregation method that uses weighted averaging based on each client's data volume. Li et al. [28] introduced FedProx by adding a regularization term to the client loss function, aligning updates more closely with the global model. However, the one-global model-fits-all approach ignores client data diversity. To solve this, our approach extends FedProx by distributing a personalized global model to each client.", + "bbox": [ + 511, + 551, + 905, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Proposed Psychological Study on AU Incoordination", + "text_level": 1, + "bbox": [ + 511, + 810, + 903, + 844 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This study employed a behavioral experiment to investigate individuals' cognitive responses to AU incoordination in the upper and lower facial regions.", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14454", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d0be976c39a1a2b1350002c8024a300f5603459bf08e9cad7c5424168a69b569.jpg", + "image_caption": [ + "(a) AU $4 + 9$" + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 161, + 155 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6a6758a36e511f40688f624788965fb5f708c901579da499564d7275e17f1144.jpg", + "image_caption": [ + "(b) AU 6+12" + ], + "image_footnote": [], + "bbox": [ + 200, + 88, + 267, + 155 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/75bb2e5693552c09e2c1a618dc3ae130c5010e3798a28cb5539630fc5fb61851.jpg", + "image_caption": [ + "(c) AU 4+12", + "Figure 1. Examples of AU composite images. (a) and (b) are coordinated composites. (c) and (d) are uncoordinated composites." + ], + "image_footnote": [], + "bbox": [ + 305, + 88, + 372, + 155 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c488cf12372e861cf51dcda4ad2396759b6b73ada71b35e9bbaf1e1974f014a9.jpg", + "image_caption": [ + "(d) AU 6+15" + ], + "image_footnote": [], + "bbox": [ + 411, + 88, + 478, + 155 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The study included 30 college participants (20 female, Mean= 22.22 years), a sample size considered adequate for statistical reliability under the Central Limit Theorem [23]. All procedures followed the Declaration of Helsinki and were approved by the institutional review board of the Institute of Psychology, Chinese Academy of Sciences.", + "bbox": [ + 89, + 238, + 482, + 328 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This behavioral experiment employed a within-subjects design. The independent variable was the type of AU composite image (coordinated vs. uncoordinated AU composites), as shown in Fig. 1. The face images were segmented anatomically. The dividing line was consistently present in all stimuli, serving as a control variable and would not influence the experiment's outcomes. To control for individual facial differences in emotion recognition, the original images used for composites were sourced from the same person [10]. Seventy-two composite images were generated from upper and lower facial AUs across six basic emotions, with each emotion represented by six coordinated and six uncoordinated expressions.", + "bbox": [ + 89, + 329, + 482, + 525 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The dependent variables are accuracy rates and scores. Specifically, participants made two judgments for each composite image: whether it was coordinated or uncoordinated, and the degree of coordination or incoordination.", + "bbox": [ + 89, + 526, + 482, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Statistical analysis from this study showed that participants were significantly more accurate in recognizing coordinated AU composite images than uncoordinated ones (Please see Supplementary Materials (Suppl.) for more experimental details and statistical results). The study further highlighted specific AU combinations, such as AU4 and AU12, that participants quickly identified as uncoordinated, whereas combinations like AU6 and AU12 were rapidly recognized as coordinated. This suggests that these uncoordinated AU pairings may inherently convey emotional incoordination. Besides, the coordination patterns between specific AUs reveal systematic facial action structures. These patterns help the network distinguish both coordinated and uncoordinated facial movements, thus enhancing its ability to identify ME features more effectively.", + "bbox": [ + 89, + 587, + 482, + 814 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Our Proposed FED-PsyAU Framework", + "text_level": 1, + "bbox": [ + 89, + 829, + 444, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose the FED-PsyAU framework (Fig. 2), which embeds a psychology-driven MER model within a federated aggregation module. The model features a hierarchi", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "cal structure that learns from local ROIs to AU groups and finally to global facial regions, capturing dynamic topology while minimizing redundancy. It leverages a multiscale InceptionNet to extract and fuse global OF with structural facial features. The federated module enables privacy-preserving collaborative training to boost MER performance on each client.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Localized ROI Modeling", + "text_level": 1, + "bbox": [ + 511, + 217, + 741, + 233 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since MEs are manifested by muscle movements in key facial regions (e.g., eyes, mouth, nose) [10], we refine the standard 68 landmarks from Dlib [20] to precisely capture these changes. As shown in Fig. 3, we discard 10 less informative outer contour landmarks and add 7 new midpoints to better track movements in the cheek and eyebrow regions, resulting in a final set of 65 keypoints.", + "bbox": [ + 511, + 242, + 903, + 348 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Then, onset-apex OF inputs $(128 \\times 128$ pixels) are cropped into 65 ROIs, each measuring $5 \\times 5$ pixels, centered on selected facial landmarks $\\mathbf{p}_k = (x_k, y_k)$ , where $k \\in 1, 2, \\ldots, 65$ , denotes the landmark index. To enhance robustness to apex annotation, we also compute OF using a small window of frames around the apex. The ROI features input to the Local ROI Feature Extractor (LFE), denoted as $\\Theta_k$ , contain the horizontal and vertical components of OF and optical strain. The LFE module consists of three convolutional layers, each followed by batch normalization and ReLU activation. All kernel sizes were $3 \\times 3$ . The output channels in the first two layers increase from 32 to 64, while the final layer outputs 3 channels. This design helps the LFE capture complex local patterns, with the output features referred to as $\\Phi_k = \\mathrm{LFE}(\\Theta_k)$ .", + "bbox": [ + 511, + 352, + 906, + 579 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by Vision Transformer [8], we use a Spatial Structure Encoder (SSE), i.e., Transformer Encoder with three attention heads to capture diverse dependencies among ROIs, as MEs result from coordinated muscle movements across multiple areas. Precisely, we convert $\\Phi_k$ , into a token. To retain the positional information of the facial landmarks, a positional embedding $E_{k}$ is added based on the landmark index to each token, yielding $X_{k} = \\Phi_{k} + E_{k}$ . We add the output of the three head attention module to the input feature $X_{k}$ , and then normalize the result using LayerNorm. Finally, after passing through a feed-forward neural network and applying LayerNorm, the final output representation is obtained as $Z_{k}\\in \\mathbb{R}^{3\\times 5\\times 5}$ .", + "bbox": [ + 511, + 583, + 906, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.2. Relationship Modeling among AU Features", + "text_level": 1, + "bbox": [ + 511, + 799, + 880, + 815 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our proposed Relationship Modeling among AU Features (AFR) module consists of two components: an AU Feature Extractor (AFE) to derive features from AU-specific facial regions, and a GAT-structure to model the topological relationships between these AUs.", + "bbox": [ + 511, + 824, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14455", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/32a7893b63bbeb533929b4b91016e08a888aebebc0656d7b79a2840f881c5d6b.jpg", + "image_caption": [ + "Figure 2. The pipeline of our proposed FED-PsyAU framework, including MER network (black block) and the Federated Aggregation (FA) module (orange block). In MER network, the LRM module consists of the local ROI feature extractor (LFE) and the spatial structure encoder (SSE), which preserves spatial structure information while efficiently extracting ROI features; the AFR module includes the AU feature extractor (AFE) and the dynamic prior knowledge-based graph attention network (DPK-GAT) for effective AU feature extraction and capturing AU dynamic relationships, the DSI module use a dual-stream structure to capture multi-scale facial muscle topology and motion information for MER. Then, the FA module integrates the MER network into the FL framework." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 903, + 277 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6fff2cada31386a5b011df13bd87f2c5f07231243518331d7ca049c484b4cd7c.jpg", + "image_caption": [ + "Figure 3. ROI grouping strategy for AU. OF features from each ROI are processed through LFE and SSE to yield high-dimensional representations. ROIs are then grouped by AU movement patterns to create inputs for AFEs, ultimately producing AU node features for the GAT network." + ], + "image_footnote": [], + "bbox": [ + 101, + 392, + 473, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2.1. AU Feature Extractor", + "text_level": 1, + "bbox": [ + 89, + 695, + 290, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "MEs involve dynamic, coordinated interactions across multiple regions [10] and are characterized by subtle movements that differ from typical macro-expressions. Therefore, as shown in Fig. 3, when grouping ROIs for the AFE module, the AU selection and regionalization are based on: 1) AU frequency in MEs; 2) Grouping by primary/secondary muscle activation areas [7]. Notably, not all AUs employed in Section 3 are applied to MER. We specifically retain only the 12 AUs exhibiting the highest relevance to ME dynamics, as listed in Tab. 1. Detailed occurrences are provided in the Suppl.", + "bbox": [ + 88, + 715, + 482, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The first step we construct in AFE is Group Squeeze and", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/4434210dc95c750de1434423c9b2d3c7ecffecf56f8da8459c219c8ad5f1d05b.jpg", + "table_caption": [ + "Table 1. Selected AUs and Corresponding Regions (Primary and Secondary/ Coordinate). $g$ , $N_g$ and $N_{g\\_AU}$ represent ROI Group index, corresponding ROI number and AU amount." + ], + "table_footnote": [], + "table_body": "
gAUPrimarySecondary/ CoordinateNgNg-AU
11,2,4EyebrowsEyes233
25,6,7EyesEyebrows, Cheeks, Mouth483
39,10Nose, CheeksMouth, Chin382
412,14,15,17MouthEyes, Eyebrows, Nose524
", + "bbox": [ + 516, + 446, + 906, + 515 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Excitation (GSE). It computes the attention for each group of AUs (ROI group), which makes the model pay more attention to the critical ROIs. Specifically, for each ROI group input $R_{g} = \\{Z_{1}, Z_{2}, \\ldots, Z_{N_{g}}\\}$ , $g \\in \\{1, 2, 3, 4\\}$ , the GSE module performs global average pooling on each ROI and computes channel importance through a fully connected (FC) layer. The original features are then preserved through residual concatenation:", + "bbox": [ + 511, + 537, + 906, + 657 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\mathrm {G S E}} \\left(R _ {g}\\right) = \\left\\{W _ {k} ^ {\\text {a t t}} \\odot Z _ {k} + Z _ {k} \\mid Z _ {k} \\in R _ {g} \\right\\} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 667, + 906, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $W_{k}^{att}$ is the attention weight for each group of ROI, $\\odot$ denotes element-wise multiplication along the channel dimension of $Z_{k}$ . Next, $f_{\\mathrm{GSE}}(R_g)$ undergoes further feature extraction, reducing from $N_{g}$ composite features to $N_{g\\_AU}$ representative AU features, yielding the output feature $F_{g\\_AU}$ . Precisely, the process comprises two convolutional layers: a $3 \\times 3$ layer that doubles the input dimensions and a $1 \\times 1$ layer that restores them to the original size. Both layers are followed by Batch Normalization and ReLU activation for improved stability and representation.", + "bbox": [ + 511, + 693, + 908, + 845 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2.2. AU Relationship Modeling", + "text_level": 1, + "bbox": [ + 511, + 851, + 746, + 867 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prior Knowledge: To model AU dynamic relationships, we integrate two forms of prior knowledge: an \"intrinsic\" prior", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14456", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "from psychological principles and an \"extrinsic\" prior from data-driven patterns. The intrinsic prior, derived from our psychological findings on AU coordination (Sec. 3), provides structural guidance by defining the initial GAT adjacency matrix $A$ . The extrinsic prior, based on statistical AU co-occurrence patterns in the data, offers data-driven refinement that theoretical models may overlook. This extrinsic prior forms a prior attention matrix $D$ to guide weight allocation among AU nodes. By fusing this foundational psychological structure with real-world dynamics, our model achieves superior adaptability and generalization (see Suppl. for matrix details).", + "bbox": [ + 88, + 90, + 480, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Facial Segmentation: As presented in Sec. 3, the upper and lower facial regions, controlled by different neural pathways, exhibit distinct muscle movement patterns. AUs and their associated primary muscles in the upper and lower face can exhibit coordination (potential co-occurrence) or lack of coordination (infrequent simultaneous appearance). Leveraging these intra- and inter-regional interactions, we model AU relationships for the upper face, lower face, and whole face. Consistent with the experimental setup in Sec. 3, we segment the face into upper and lower parts based on the orbicularis oculi muscle. Specifically, following the primary occurrence regions of 12 listed AUs, the AUs in the first two groups from Tab. 1 are classified as upper face AUs, while those in the last two groups are classified as lower face AUs.", + "bbox": [ + 89, + 273, + 482, + 484 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {A U} ^ {U p p e r} = F _ {1 \\_ A U} \\cup F _ {2 \\_ A U}, F _ {A U} ^ {L o w e r} = F _ {3 \\_ A U} \\cup F _ {4 \\_ A U} (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 496, + 482, + 516 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We input the AU features from the upper and lower face: $F_{AU}^{Upper}$ , $F_{AU}^{Lower}$ into our proposed Dynamic Prior Knowledge GATs (DPK-GATs) in parallel (Details are presented in the next part of this subsection). These networks use psychology prior knowledge to build AU adjacency relationships and data-driven priors to guide attention learning, producing outputs $h^{Upper}$ and $h^{Lower}$ , which are combined into $h^{Local}$ for the next stage.", + "bbox": [ + 89, + 530, + 482, + 650 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the whole-face DPK-GAT, we construct the network similarly but set adjacency values within the same region to zero, ignoring intra-regional AU relationships. By learning cross-regional topological structures and statistical patterns, we obtain global AU relationship features, denoted as $h^{Global}$ , as illustrated in Fig. 4.", + "bbox": [ + 89, + 652, + 482, + 743 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "DPK-GAT: We model dynamic AU node relationships combing GAT and the prior knowledge. In particular, for the input feature belonging to $\\{F_{AU}^{Upper}, F_{AU}^{Lower}, h^{Local}\\}$ , a linear mapping projects it to a higher-dimensional space as $f$ . The similarity score $e_{ij}$ between AU nodes $i$ and $j$ , is calculated using the attention weight vector $\\vec{a}$ and concatenated node features $f_{i}$ and $f_{j}$ :", + "bbox": [ + 89, + 743, + 482, + 851 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ne _ {i j} = \\left\\{ \\begin{array}{l l} \\text {L e a k y R e L U} \\left(\\tilde {a} ^ {T} [ f _ {i} | | f _ {j} ]\\right), & \\text {i f} A _ {i j} > 0 \\\\ - \\infty , & \\text {i f} A _ {i j} = 0 \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 863, + 482, + 904 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3934cbe5af16a98d999e3491078e08adc38f6f6edc979040b58195a36422df20.jpg", + "image_caption": [ + "Figure 4. DPK-GAT from facial segmentations to global face. First, intra-regional AU node relationships are learned within the upper and lower facial regions. Second, cross-regional relationships are modeled to integrate interactions between these regions." + ], + "image_footnote": [], + "bbox": [ + 516, + 88, + 903, + 195 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, only neighboring nodes defined by the psychology-derived adjacency matrix $A$ are considered.", + "bbox": [ + 511, + 287, + 903, + 318 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final attention $\\alpha_{ij}$ is calculated by combining the data-driven prior attention $D_{ij}$ and the self-attention, obtained by the softmax operation applied to $e_{ij}$ :", + "bbox": [ + 511, + 319, + 903, + 364 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {i j} = (1 - \\beta) \\cdot \\operatorname {s o f t m a x} \\left(e _ {i j}\\right) + \\beta \\cdot D _ {i j} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 377, + 903, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where, $\\beta$ is the dynamic prior attention weight, updated based on the training rounds. Along with a residual concatenation, the final node representation $h_i$ is aggregated as:", + "bbox": [ + 511, + 404, + 903, + 463 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nh _ {i} = \\operatorname {E L U} \\left(\\sum_ {j \\in \\mathcal {N} (i)} \\alpha_ {i j} \\cdot f _ {j}\\right) + f _ {i} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 464, + 903, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{N}(i)$ denotes the neighboring nodes of $i$ . In our implementation, we employ three attention heads and a two-layer structure, enabling the extraction of complex AU dependencies.", + "bbox": [ + 511, + 489, + 903, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Dual-Stream-InceptionNet", + "text_level": 1, + "bbox": [ + 511, + 559, + 754, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Muscle movements vary widely across facial regions, making multi-scale information essential for the MER task. InceptionNet's core concept is leveraging parallel convolutional kernels to extract multi-scale features, enhancing the network's ability to capture fine details and overall structure [46]. We build on this by using a dual-stream InceptionNet as the ME feature extractor. One stream processes the AU relationship features $h^{Global}$ , while the other handles global OF features $\\Theta$ . This dual-stream design allows the model to capture the topological relationships of facial muscles through AUs and holistic motion via OF. The output tensors from both streams are flattened, concatenated, and fed into a FC layer for ME prediction.", + "bbox": [ + 511, + 580, + 905, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Federated Aggregation", + "text_level": 1, + "bbox": [ + 511, + 787, + 725, + 804 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Traditional centralized learning consolidates all data on a single server, enhancing model accuracy and generalization but neglecting privacy concerns in MER. FL could address this by training models locally and sharing only parameters, protecting data privacy and retaining each client's unique data characteristics.", + "bbox": [ + 511, + 809, + 905, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14457", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "FedProx [29] is a FL approach designed to address data heterogeneity among clients. It incorporates proximal terms during model aggregation to balance global and client-local learning, improving both performance and convergence speed. We extend FedProx by altering the server's processing and distribution strategy during aggregation, which we call Personalized FedProx (P-FedProx). Specifically, P-FedProx initializes each client's model $W_{i}^{t}$ in round $t$ by combining its previous model $\\mathring{W}_{i}^{t-1}$ with models from other clients $\\mathring{W}_{j}^{t-1}$ .", + "bbox": [ + 89, + 90, + 483, + 238 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nW _ {i} ^ {t} = \\theta \\cdot \\mathring {W} _ {i} ^ {t - 1} + \\sum_ {\\substack {j = 1 \\\\ j \\neq i}} ^ {n} \\omega_ {j} \\cdot \\mathring {W} _ {j} ^ {t - 1}, \\text {where} \\sum_ {\\substack {j = 1 \\\\ j \\neq i}} ^ {n} \\omega_ {j} = 1 - \\theta \\tag{6}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 246, + 483, + 272 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where weight $\\omega_{j}$ is proportional to $j$ th client's data size. For $i$ th client, $\\theta$ is set to 0.9 (optimal experimental result, see more detail in the Suppl.)", + "bbox": [ + 89, + 277, + 483, + 325 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.5. Loss Function", + "text_level": 1, + "bbox": [ + 89, + 332, + 235, + 347 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "AU Prediction Loss: To enhance MER, our network models the topological relationships and statistical patterns of AUs, making AU prediction an auxiliary task. The prediction is utilized in two stages: one enhances the AU feature extraction performance of the AFE module, and the other improves the GAT network's ability to learn AU topological structures. AU recognition is formulated as a multi-label classification problem, with the Binary Cross-Entropy Loss for each AU label defined as:", + "bbox": [ + 89, + 353, + 483, + 479 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {A}} = \\frac {\\sum_ {j = 1} ^ {M} \\left[ y _ {j} \\cdot \\log \\left(\\sigma \\left(\\hat {y} _ {j}\\right)\\right) + \\left(1 - y _ {j}\\right) \\cdot \\log \\left(1 - \\sigma \\left(\\hat {y} _ {j}\\right)\\right) \\right]}{- M} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 484, + 483, + 516 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $M$ is the total number of AUs, $y_{j}$ , $\\sigma (\\hat{y}_j)$ and $\\hat{y}_j$ represent respectively the ground truth label (0 or 1), the logit output and prediction after applying the sigmoid function for the $j$ -th AU. The loss function is consistent for both stages, i.e., $L_{\\mathrm{A}}^{\\mathrm{AFE}}$ and $L_{\\mathrm{A}}^{\\mathrm{GAT}}$ .", + "bbox": [ + 89, + 523, + 482, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MER Loss: The Cross-Entropy Loss function is utilized for emotion classification. Thus, our overall MER loss is composed as follows:", + "bbox": [ + 89, + 599, + 483, + 642 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {M E R}} = \\alpha_ {1} \\cdot \\left(- \\sum_ {i = 1} ^ {N} y _ {i} \\cdot \\log (\\hat {y} _ {i})\\right) + \\alpha_ {2} \\cdot L _ {\\mathrm {A}} ^ {\\mathrm {A F E}} + \\alpha_ {3} \\cdot L _ {\\mathrm {A}} ^ {\\mathrm {G A T}} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 648, + 482, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where, $N$ represents the number of classes, $y_{i}$ and $\\hat{y}_i$ are the one-hot encoded ground truth label and the predicted probability for class $i$ , $\\alpha_{1}$ , $\\alpha_{2}$ and $\\alpha_{3}$ are hyperparameters, set to 0.2, 0.8, and 0.8 to prioritize MER loss while balancing AU recognition (See parameter analysis in Suppl.).", + "bbox": [ + 89, + 674, + 482, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Federated Loss: During federated training, the local client's loss function includes the MER loss $L_{\\mathrm{MER}}$ and a proximal term:", + "bbox": [ + 89, + 750, + 483, + 796 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {F L}} = L _ {\\mathrm {M E R}} + \\frac {\\alpha_ {4}}{2} \\cdot \\left\\| W - W _ {t} \\right\\| ^ {2} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 804, + 482, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $W_{t}$ denotes the global model in the server's $t$ -th round of distribution, and $W$ represents the current model of the client. $\\alpha_{4}$ is the hyperparameter for the weight of the proximal term.", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 513, + 89, + 645, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Configuration", + "text_level": 1, + "bbox": [ + 513, + 122, + 658, + 138 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Databases: We use two recently published large scale and challenging ME databases: CAS(ME) $^3$ and DFME. CAS(ME) $^3$ -Part A offers 860 samples from 100 participants. DFME includes three subsets with (participants, samples): A (72,1,118), B (92,969), C (492,5,439). For Apex-first-frame scenario, we compute the OF between the apex and mid-frame towards the offset.", + "bbox": [ + 511, + 152, + 905, + 258 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pre-training: To enhance the model's ability to perceive facial motion patterns, we first pre-trained the AU module (LRM+AFR) on the DISFA [40] and CK [19] datasets. The effect of this pre-training is detailed in the Suppl.", + "bbox": [ + 511, + 266, + 905, + 327 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MER Experiment Settings: The evaluation follows standard protocols on two datasets. On $\\mathrm{CAS(ME)^3}$ , we perform 3-class classification (positive, negative, surprised) [45] with Leave-One-Subject-Out (LOSO) validation. On DFME, we conduct 7-class classification (Happiness, Surprise, Disgust, Sadness, Anger, Fear, and Contempt) using scale-adapted 10-fold cross-validation [59]. To address class imbalance, performance is measured by the UF1 and UAR. Experiments were implemented in PyTorch 1.13.0 and run on five NVIDIA GeForce RTX 4090 GPUs.", + "bbox": [ + 511, + 334, + 906, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Federated Experiment Settings: The DFME and $\\mathrm{CAS(ME)^3}$ datasets are randomly split into 5 and 2 local clients, respectively, with each client assigned an equal number of subjects. However, variations in microexpression (ME) counts per subject result in differing ME data distributions and quantities across clients, mirroring real-world data heterogeneity (See detailed data distributions in the Suppl.). Traditional LOSO and k-fold cross-validation can leak training samples in a federated setting, so we employ a random partition strategy with $70\\%$ of samples for training and $30\\%$ for testing, averaging results over 10 tests to minimize randomness. Furthermore, we assume an ideal FL scenario with no client dropouts and equal local updates per client. In FedAvg [41] and FedProx [28], the server assigns an identical global model to all clients each round; differently, FedProx adds a proximal term to reduce model bias in local training. Our P-FedProx enhances this by assigning tailored initial models, with $90\\%$ ( $\\theta$ ) weight on each client's own model and $10\\%$ on others'. Additionally, we compare against recent personalized FL methods: FedRep [5] uses shared representations with personalized heads, ELLP [61] employs parameter decoupling with Fisher Information weighting, and FedAS [55] applies parameter alignment and client synchronization. Unlike these approaches requiring architectural changes or complex procedures, P-FedProx achieves personalization through simple weighted initialization.", + "bbox": [ + 511, + 493, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14458", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/502801362957141a992ddc3c6071f9230c83567e960670e5400bcdd202050902.jpg", + "table_caption": [ + "Table 2. SOTA methods comparison on DFME and CAS(ME) ${}^{3}$" + ], + "table_footnote": [], + "table_body": "
CategoryMethodYearDFMECAS(ME)3
UF1UARUF1UAR
3D-CNNI3D [2]20170.29230.3058--
R3D [17]20180.21640.2313--
Hand-craftedLBP-TOP [49]20140.23360.2653--
MDMO [37]20150.24890.2939--
Deep LearningOFF-ApexNet [15]20190.23860.2806--
AlexNet [22]2012--0.25700.2634
STSTNet [36]20190.27140.31080.37950.3792
RCN-A [52]20200.27510.31230.39280.3893
MERSiam [58]20210.31840.35320.31840.3532
FGRL [24]20210.07360.14290.33330.2636
FR [24, 60]20220.35590.38140.34930.3413
MMNet [25]20220.26490.27760.37060.3646
BDCNN [4]20220.29750.33140.50500.5164
Micro-BERT [42]2023--0.56040.6125
HTNet [50]20240.32430.33680.57670.5415
HSTA [16]2024--0.5930.618
Ours-0.38530.39780.62210.6226
", + "bbox": [ + 94, + 113, + 482, + 333 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Comparison to State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 89, + 359, + 434, + 376 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 2 list the SOTA comparisons on DFME and CAS(ME) $^3$ . The confusion matrices are shown in the Suppl.", + "bbox": [ + 89, + 381, + 482, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Regarding DFME, classifying ME samples into seven categories poses significant challenges, resulting in relatively low UF1 and UAR scores across all methods. However, our proposed method outperforms all others in every metric. Specifically, Our method outperforms I3D by $9.30\\%$ in UF1 and $9.20\\%$ in UAR, and surpasses LBP-TOP by $15.17\\%$ in UF1 and $13.25\\%$ in UAR, highlighting its effectiveness in extracting critical and discriminative features from ME clips. Additionally, compared to recent deep learning methods like FR, our approach yields $2.94\\%$ and $1.64\\%$ higher UF1 and UAR, confirming the benefits of incorporating facial topology and structured muscle motion for enhanced performance.", + "bbox": [ + 88, + 412, + 482, + 608 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Regarding CAS(ME) $^3$ , the primary challenge lies in its sample complexity. Compared to baseline methods (STST-Net, RCN-A, FR), our proposed method achieves over approximately $20\\%$ higher UF1 and UAR scores, marking a significant advancement. Our method surpasses three recent SOTA approaches in MER, including HTNet, which employs a hierarchical Transformer for local feature learning via self-attention. Specifically, our approach outperforms HTNet by $4.54\\%$ in UF1 and $8.11\\%$ in UAR, leveraging psychological insights and statistical patterns to capture deeper facial muscle dynamics. Additionally, it exceeds Micro-BERT and HSTA by $6.17\\%$ and $2.91\\%$ in UF1, and $1.01\\%$ and $0.46\\%$ in UAR, respectively. By mining facial movement patterns to better understand topology, our model achieves improved metrics, indicating enhanced accuracy in classifying all emotion categories.", + "bbox": [ + 89, + 609, + 482, + 852 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 862, + 243, + 878 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 3 listed the ablation study results on DFME.", + "bbox": [ + 89, + 885, + 413, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2e3b25314fa1a633b5772fe92f6dcaf267d5d667bcbc6b3699ad73f96961a2fd.jpg", + "table_caption": [ + "Table 3. Ablation study. L, G, S and D represent Local (upper and lower face), Global, Single stream and Dual Stream, respectively." + ], + "table_footnote": [], + "table_body": "
SettingLRMAFRInceptionNetUF1↑UAR↑
LFESSELGSD
I✓0.32490.3404
II✓✓0.33620.3517
III✓✓✓0.34800.3624
IV✓✓✓✓0.36290.3723
V✓✓✓✓0.36040.3658
VI✓✓✓✓✓0.38530.3978
SettingWithout AU GroupWith AU GroupUF1↑UAR↑
VII✓0.36830.3793
VIII✓0.38530.3978
SettingPsychological priorData-driven PriorUF1↑UAR↑
IX0.36150.3701
X✓0.37230.3832
XI✓0.37050.3767
XII✓✓0.38530.3978
", + "bbox": [ + 516, + 127, + 903, + 348 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of AU Feature: Setting I, which uses only full face OF as input, serves as our baseline for evaluating the role of AU features in MER. In Setting II, integrating AU features into the Dual Stream InceptionNet leads to a $1.13\\%$ improvement in both UF1 and UAR. This improvement demonstrates that incorporating AU features significantly enhances the model's ability to recognize MEs, and underscores the value of fusing multiple feature types to strengthen model performance.", + "bbox": [ + 511, + 377, + 906, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of ROI Relationship Modeling: The results in Setting III show improvements of $1.18\\%$ in UF1 and $1.07\\%$ in UAR compared to Setting II, indicating that the model effectively captures complex dependencies between ROIs and extracts more discriminative feature representations.", + "bbox": [ + 511, + 516, + 906, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of AU grouping: We compared AU feature extraction from the global face with our AU grouping strategy (Setting VII vs. VIII), the performance improvement (UF1: $1.70\\%$ , UAR: $1.85\\%$ ), confirming that AU grouping enables the network to learn more representative AU features.", + "bbox": [ + 511, + 594, + 905, + 669 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of Modeling the Topological Structure of ME Muscle Movements: Compared to Setting III, Setting IV, including local muscle movement topology modeling, improves UF1 and UAR by $1.49\\%$ and $0.99\\%$ , respectively. Setting V, which adds global muscle movement topology modeling, shows gains of $1.24\\%$ in UF1 and $0.34\\%$ in UAR over Setting III. Combining both local and global topology modeling yields the best performance on DFME. These results demonstrate that incorporating topological modeling of muscle movements enhances MER.", + "bbox": [ + 511, + 671, + 905, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of Prior Knowledge: Tab. 3 presents the ablation experiments to evaluate the impact of using prior knowledge for constructing facial muscle topology. In particular, compared to Setting IX, Setting X, which uses only psychological knowledge, improves UF1 and UAR by $1.08\\%$ and", + "bbox": [ + 511, + 824, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14459", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7d0359ab475a41676321769bfdef60e3582e90475e1d72ee6269ed907c5d0965.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 112, + 128, + 222, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a86bcd8e0a2ef30abe66e39830767a2927e9ac24819ae8d569c5be5d4acfe90c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 225, + 130, + 333, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ed619347e1503f8d05331087100313350ff40d9aa809b553fab8d6c994e8edc0.jpg", + "image_caption": [ + "(a) UF1 comparison across different clients." + ], + "image_footnote": [], + "bbox": [ + 334, + 130, + 442, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4dca360b1ef279a4b74913e2b235d6d2581047bd809eaacd2dfa43f486177ef5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 130, + 553, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4d64020bb307952c01cb4203c06cbd5b28c204536d2809b574d8cc0743ca013d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 130, + 663, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7ce7a4851e79c5bfbe1887a3d439063c725adae50997589753a075c3fe04ae49.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 663, + 130, + 772, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d60e9d362a4ca3a84d4fb9c4de393ad2b5ae925c765090503f0a69e38f92bcc4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 772, + 130, + 883, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ca2225a778ac85bbd9856be72ef91346a4870f018b736395c5f8e1ce97c1f509.jpg", + "image_caption": [ + "Figure 5. Compare the performance of different FL frameworks on different clients: UF1 and convergence. See UAR comparison in Suppl." + ], + "image_footnote": [], + "bbox": [ + 112, + 247, + 225, + 345 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9758b0502f57eb674eb29a4c4cbace84da198af57d5284b0fb0e007d7bf93c9b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 225, + 247, + 334, + 345 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d6e5db3f15daf1da9eeb5b183ec476c91b6fe6759cc2c50eeadc434b1fbd0e13.jpg", + "image_caption": [ + "(b) Convergence comparison across different clients." + ], + "image_footnote": [], + "bbox": [ + 334, + 247, + 444, + 345 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/164efcb98172282f2bc845b5954f618bf38f57aef2f9040df21f9014b03d4a75.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 247, + 553, + 345 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0278f3eb28b06d6cbecc906342c457ab2c577fc37247aa6ffcea9c6160bebf87.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 247, + 663, + 345 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/56f21f9e921cc95ebcf6812103c58b7add37417461d1b73fabeea964d84ffd19.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 663, + 247, + 772, + 345 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5d110c145f2ae6a3be10e74af597573fbc31866014983bc5e3e2f04bef4c399e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 772, + 247, + 883, + 345 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$1.31\\%$ , respectively. Setting XI, incorporating only data-driven knowledge, shows gains of $0.9\\%$ in UF1 and $0.66\\%$ in UAR. When both psychological and data-driven knowledge are combined (Setting XII), the improvements reach $2.38\\%$ in UF1 and $2.77\\%$ in UAR, proving the effectiveness of the prior knowledge. Constructing a prior adjacency matrix allows the model to capture complex facial muscle topology, while data-driven knowledge directs attention to uncover intrinsic movements patterns of facial muscles.", + "bbox": [ + 89, + 414, + 483, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Federated Experiment", + "text_level": 1, + "bbox": [ + 89, + 561, + 302, + 577 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Since FL aims to enhance MER performance on individual clients in scenarios with limited sample sizes and data privacy issue, and given that DFME and $\\mathrm{CAS}(\\mathrm{ME})^3$ have large sample size, we partition the datasets as mentioned in Section 5.1 and treat the subdivisions as separate clients.", + "bbox": [ + 89, + 583, + 483, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While our focus is on application rather than fundamental FL innovation, our experiments confirm the effectiveness of our P-FedProx method, which is tailored for ME data distributions. Benchmarked against several open-source SOTA methods, our approach excels at mitigating small-sample challenges while preserving data privacy. Specifically (Fig. 5), the results show P-FedProx outperforms both the single-client baseline and traditional FL methods (FedAvg, FedProx), with particularly significant gains on clients with more challenging data distributions (Clients 6-7). Compared to recent personalized FL methods, our approach surpasses FedRep and ELLP and achieves performance competitive with the state-of-the-art FedAS, yet with a significantly simpler implementation. These consistent gains demonstrate that our weighted model initialization effectively addresses the data heterogeneity that challenges", + "bbox": [ + 89, + 659, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "traditional FL, thereby mitigating ME sample scarcity while ensuring data privacy.", + "bbox": [ + 511, + 415, + 903, + 445 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Meantime, we note that all FL methods achieved good performance on the local clients from DFME, which may be due to the fact that these clients have a similar amount of data and slight differences in data distribution, presenting fewer challenges for federated learning.", + "bbox": [ + 511, + 446, + 905, + 521 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Besides, we analyze the training loss progression of different federated algorithms over communication rounds. We found that P-FedProx demonstrates the smoothest convergence curves with minimal oscillations, particularly evident in Clients 1-5. On the other hand, P-FedProx converges significantly faster than other FL methods. This rapid early convergence suggests that personalized global model assignment provides clients with better initialization states.", + "bbox": [ + 511, + 523, + 905, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 662, + 633, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "MER is challenged by limited sample sizes, subtle features, and privacy concerns in real-world applications. To address these issues, we conduct a psychological study to provide structured insights into facial muscle dynamics. We then develop a FED-PsyAU framework that integrates these findings with data-driven patterns for hierarchical learning of dynamic features, enhancing MER performance. Additionally, the FL framework improves MER across clients while preserving data privacy through iterative model updates. The comprehensive experiments validate our approach's effectiveness. In the future, we will integrate additional AUs to build a more comprehensive ME topology information and further advance MER in real-world applications through FL and semi-supervised approaches.", + "bbox": [ + 511, + 688, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "FedAvg(2017) PersonalizedFedProx ELLP(2024) Baseline FedProx(2020) FedRep(2021) FedAS(2024)", + "bbox": [ + 321, + 92, + 674, + 109 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14460", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 263, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This research was partially funded by 1) the National Natural Science Foundation of China (62476269, 62276252); 2) the Youth Innovation Promotion Association CAS.", + "bbox": [ + 89, + 114, + 485, + 160 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 172, + 187, + 188 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Xianye Ben, Yi Ren, Junping Zhang, Su-Jing Wang, Kidiyo Kpalma, Weixiao Meng, and Yong-Jin Liu. Video-Based Facial Micro-Expression Analysis: A Survey of Datasets, Features and Algorithms. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):5826-5846, 2022. 2", + "[2] Joao Carreira and Andrew Zisserman. Quo Vadis, Action Recognition? A New Model and the Kinetics Dataset. In proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 7", + "[3] Boyu Chen, Zhihao Zhang, Nian Liu, Yang Tan, Xinyu Liu, and Tong Chen. Spatiotemporal Convolutional Neural Network with Convolutional Block Attention Module for Micro-Expression Recognition. Information, 11(8):380, 2020. 1", + "[4] Bin Chen, Kun-Hong Liu, Yong Xu, Qing-Qiang Wu, and Jun-Feng Yao. Block Division Convolutional Network With Implicit Deep Features Augmentation for Micro-Expression Recognition. IEEE Transactions on Multimedia, 25:1345–1358, 2022. 1, 2, 7", + "[5] Liam Collins, Hamed Hassani, Aryan Mokhtari, and Sanjay Shakkottai. Exploiting shared representations for personalized federated learning. In International conference on machine learning, pages 2089-2099. PMLR, 2021. 6", + "[6] Adrian K. Davison, Cliff Lansley, Nicholas Costen, Kevin Tan, and Moi Hoon Yap. SAMM: A Spontaneous MicroFacial Movement Dataset. IEEE Transactions on Affective Computing, 9(1):116-129, 2018. 2", + "[7] Zizhao Dong, Gang Wang, Shaoyuan Lu, Jingting Li, Wenjing Yan, and Su-Jing Wang. Spontaneous Facial Expressions and Micro-Expressions Coding: From brain to Face. Frontiers in Psychology, 12:784834, 2022. 4", + "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale, 2021. 3", + "[9] Paul Ekman and Wallace V Friesen. Nonverbal Leakage and Clues to Deception. Psychiatry, 32(1):88-106, 1969. 1", + "[10] Paul Ekman and Wallace V Friesen. Facial Action Coding System. Environmental Psychology & Nonverbal Behavior, 1978. 1, 2, 3, 4", + "[11] Ali et al. Privacy preserving personalization for video facial expression recognition using federated learning. IMCI'22, . 2", + "[12] Fan et al. Feeling without sharing: A federated video emotion recognition framework via privacy-agnostic hybrid aggregation. ACM MM'23, . 2", + "[13] Xinqi Fan, Xueli Chen, Mingjie Jiang, Ali Raza Shahid, and Hong Yan. SelfME: Self-Supervised Motion Learning for Micro-Expression Recognition. In Proceedings of" + ], + "bbox": [ + 93, + 198, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13834-13843, 2023. 2", + "[14] Viswanatha Reddy Gajjala, Sai Prasanna Teja Reddy, Snehasis Mukherjee, and Shiv Ram Dubey. MERANet: facial micro-expression recognition using 3D residual attention network. In Proceedings of the Twelfth Indian Conference on Computer Vision, Graphics and Image Processing, New York, NY, USA, 2021. Association for Computing Machinery. 1", + "[15] Yee Siang Gan, Sze-Teng Liong, Wei-Chuen Yau, Yen-Chang Huang, and Lit-Ken Tan. Off-ApexNet on microexpression recognition system. Signal Processing: Image Communication, 74:129-139, 2019. 7", + "[16] Haihong Hao, Shuo Wang, Huixia Ben, Yanbin Hao, Yansong Wang, and Weiwei Wang. Hierarchical Space-Time Attention for Micro-Expression recognition. arXiv preprint arXiv:2405.03202, 2024. 7", + "[17] Kensho Hara, Hirokatsu Kataoka, and Yutaka Satoh. Can Spatiotemporal 3D CNNs Retrace the History of 2D CNNs and Imagenet? In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 6546-6555, 2018. 7", + "[18] Xiaohua Huang, Su-Jing Wang, Guoying Zhao, and Matti Piteikainen. Facial Micro-Expression Recognition Using Spatiotemporal Local Binary Pattern with Integral Projection. In 2015 IEEE International Conference on Computer Vision Workshop (ICCVW), pages 1–9. IEEE, 2015. 2", + "[19] T. Kanade, J.F. Cohn, and Yingli Tian. Comprehensive database for facial expression analysis. In Proceedings Fourth IEEE International Conference on Automatic Face and Gesture Recognition (Cat. No. PR00580), pages 46-53, 2000. 6", + "[20] Davis E King. Dlib-ml: A Machine Learning Toolkit. The Journal of Machine Learning Research, 10:1755-1758, 2009. 3", + "[21] Thomas N Kipf and Max Welling. Semi-Supervised Classification with Graph Convolutional Networks. arXiv preprint arXiv:1609.02907, 2016. 1, 2", + "[22] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 7", + "[23] Sang Gyu Kwak and Jong Hae Kim. Central limit theorem: the cornerstone of modern statistics. Korean journal of anesthesiaology, 70(2):144, 2017. 3", + "[24] Ling Lei, Tong Chen, Shigang Li, and Jianfeng Li. Microexpression Recognition Based on Facial Graph Representation Learning and Facial Action Unit Fusion. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1571-1580. IEEE, 2021. 1, 2, 7", + "[25] Hanting Li, Mingzhe Sui, Zhaoqing Zhu, and Feng Zhao. Mmnet: Muscle motion-guided network for microexpression recognition. arXiv preprint arXiv:2201.05297, 2022. 7", + "[26] Jingting Li, Moi Hoon Yap, Wen-Huang Cheng, John See, Xiaopeng Hong, Xiaobai Li, and Su-Jing Wang. FME'21:" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14461", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1st Workshop on Facial Micro-Expression: Advanced Techniques for Facial Expressions Generation and Spotting. In Proceedings of the 29th ACM International Conference on Multimedia, pages 5700-5701. ACM, 2021. 2", + "[27] Jingting Li, Zizhao Dong, Shaoyuan Lu, Su-Jing Wang, Wen-Jing Yan, Yinhuan Ma, Ye Liu, Changbing Huang, and Xiaolan Fu. CAS(ME) $^3$ : A Third Generation Facial Spontaneous Micro-Expression Database With Depth Information and High Ecological Validity. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):2782-2800, 2022. 2", + "[28] Tian Li, Anit Kumar Sahu, Maziar Sanjabi, Manzil Zaheer, Ameet Talwalkar, and Virginia Smith. On the Convergence of Federated Optimization in Heterogeneous Networks. arXiv preprint arXiv:1812.06127, 2018. 2, 6", + "[29] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks, 2020. 6", + "[30] Xiaobai Li, Tomas Pfister, Xiaohua Huang, Guoying Zhao, and Matti Pietikainen. A Spontaneous Micro-expression Database: Inducement, collection and baseline. In 2013 10th IEEE International Conference and Workshops on Automatic Face and Gesture Recognition (FG), pages 1-6, 2013. 2", + "[31] Xiaobai Li, Xiaopeng Hong, Antti Moilanen, Xiaohua Huang, Tomas Pfister, Guoying Zhao, and Matti Pietikainen. Towards Reading Hidden Emotions: A Comparative Study of Spontaneous Micro-Expression Spotting and Recognition Methods. IEEE Transactions on Affective Computing, 9(4): 563–577, 2018. 1", + "[32] Xiaobai Li, Shiyang Cheng, Yante Li, Muzammil Behzad, Jie Shen, Stefanos Zafeiriou, Maja Pantic, and Guoying Zhao. 4DME: A Spontaneous 4D Micro-Expression Dataset With Multimodalities. IEEE Transactions on Affective Computing, 14(4):3031-3047, 2023. 2", + "[33] Yante Li, Xiaohua Huang, and Guoying Zhao. Microexpression action unit detection with spatial and channel attention. Neurocomputing, 436:221-231, 2021. 1", + "[34] Sze-Teng Liong, John See, Raphael C.-W. Phan, Anh Cat Le Ngo, Yee-Hui Oh, and KokSheik Wong. Subtle Expression Recognition Using Optical Strain Weighted Features. In Computer Vision - ACCV 2014 Workshops, pages 644-657. Springer International Publishing, 2015. Series Title: Lecture Notes in Computer Science. 2", + "[35] Sze-Teng Liong, John See, KokSheik Wong, and Raphael C.-W. Phan. Less is More: Micro-expression Recognition from Video using Apex Frame. Signal Processing: Image Communication, 62:82-92, 2018. 2", + "[36] Sze-Teng Liong, Y. S. Gan, John See, Huai-Qian Khor, and Yen-Chang Huang. Shallow Triple Stream Three-dimensional CNN (STSTNet) for Micro-expression Recognition. In 2019 14th IEEE International Conference on Automatic Face & Gesture Recognition (FG 2019), pages 1-5, 2019. 7", + "[37] Yong-Jin Liu, Jin-Kai Zhang, Wen-Jing Yan, Su-Jing Wang, Guoying Zhao, and Xiaolan Fu. A Main Directional Mean Optical Flow Feature for Spontaneous Micro-Expression Recognition. IEEE Transactions on Affective Computing, 7 (4):299–310, 2015. 7" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Yong-Jin Liu, Jin-Kai Zhang, Wen-Jing Yan, Su-Jing Wang, Guoying Zhao, and Xiaolan Fu. A main directional mean optical flow feature for spontaneous micro-expression recognition. IEEE Transactions on Affective Computing, 7(4):299-310, 2016. 2", + "[39] Ling Lo, Hong-Xia Xie, Hong-Han Shuai, and Wen-Huang Cheng. MER-GCN: Micro-Expression Recognition Based on Relation Modeling with Graph Convolutional Networks. In 2020 IEEE conference on multimedia information processing and retrieval (MIPR), pages 79–84. IEEE, 2020. 1", + "[40] S Mohammad Mavadati, Mohammad H Mahoor, Kevin Bartlett, Philip Trinh, and Jeffrey F Cohn. Disfa: A spontaneous facial action intensity database. IEEE Transactions on Affective Computing, 4(2):151-160, 2013. 6", + "[41] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-Efficient Learning of Deep Networks from Decentralized Data. In Proceedings of the 20th International Conference on Artificial Intelligence and Statistics, pages 1273-1282. PMLR, 2017. ISSN: 2640-3498. 2, 6", + "[42] Xuan-Bac Nguyen, Chi Nhan Duong, Xin Li, Susan Gauch, Han-Seok Seo, and Khoa Luu. Micron-BERT: BERT-Based Facial Micro-Expression Recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1482–1492, 2023. 7", + "[43] T. Ojala, M. Pietikainen, and T. Maenpaa. Multiresolution gray-scale and rotation invariant texture classification with local binary patterns. IEEE Transactions on Pattern Analysis and Machine Intelligence, 24(7):971-987, 2002. 2", + "[44] Fangbing Qu, Su-Jing Wang, Wen-Jing Yan, He Li, Shuhang Wu, and Xiaolan Fu. CAS(ME) $^2$ : A Database for Spontaneous Macro-Expression and Micro-Expression Spotting and Recognition. IEEE Transactions on Affective Computing, 9(4):424-436, 2018. 2", + "[45] John See, Moi Hoon Yap, Jingting Li, Xiaopeng Hong, and Su-Jing Wang. MEGC 2019 – The Second Facial Micro-Expressions Grand Challenge. In 2019 14th IEEE International Conference on Automatic Face & Gesture Recognition (FG 2019), pages 1–5, 2019. 2, 6", + "[46] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE Conference on Computer Cision and Pattern Recognition, pages 1-9, 2015. 5", + "[47] Petar Velicković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph Attention Networks. arXiv preprint arXiv:1710.10903, 2017. 1", + "[48] Lei Wang, Pinyi Huang, Wangyang Cai, and Xiyao Liu. Micro-expression recognition by fusing action unit detection and spatio-temporal features. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 5595-5599. IEEE, 2024. 2", + "[49] Su-Jing Wang, Wen-Jing Yan, Xiaobai Li, Guoying Zhao, Chun-Guang Zhou, Xiaolan Fu, Minghao Yang, and Jianhua Tao. Micro-expression recognition using color spaces." + ], + "bbox": [ + 516, + 90, + 905, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14462", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE Transactions on Image Processing, 24(12):6034-6047, 2015. 1, 2, 7", + "[50] Zhifeng Wang, Kaihao Zhang, Wenhan Luo, and Ramesh Sankaranarayana. Htnet for micro-expression recognition. Neurocomputing, 602:128196, 2024. 1, 2, 7", + "[51] Wen-Jing Yan, Qi Wu, Yong-Jin Liu, Su-Jing Wang, and Xiaolan Fu. CASME database: A dataset of spontaneous micro-expressions collected from neutralized faces. In 2013 10th IEEE International Conference and Workshops on Automatic Face and Gesture Recognition (FG), pages 1-7. IEEE, 2013. 2", + "[52] Zhaoqiang Xia, Wei Peng, Huai-Qian Khor, Xiaoyi Feng, and Guoying Zhao. Revealing the invisible with model and data shrinking for composite-database micro-expression recognition. IEEE Transactions on Image Processing, 29: 8590-8605, 2020. 7", + "[53] Hong-Xia Xie, Ling Lo, Hong-Han Shuai, and Wen-Huang Cheng. AU-assisted graph attention convolutional network for micro-expression recognition. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2871-2880. ACM, 2020. 2", + "[54] Wen-Jing Yan, Xiaobai Li, Su-Jing Wang, Guoying Zhao, Yong-Jin Liu, Yu-Hsin Chen, and Xiaolan Fu. CASME II: An improved spontaneous micro-expression database and the baseline evaluation. PLoS ONE, 9(1):e86041, 2014-01-27. 2", + "[55] Xiyuan Yang, Wenke Huang, and Mang Ye. Fedas: Bridging inconsistency in personalized federated learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11986-11995, 2024. 6", + "[56] Moi Hoon Yap, John See, Xiaopeng Hong, and Su-Jing Wang. Facial Micro-Expressions Grand Challenge 2018 Summary. In 2018 13th IEEE International Conference on Automatic Face & Gesture Recognition (FG 2018), pages 675–678, 2018. 2", + "[57] Zhijun Zhai, Jianhui Zhao, Chengjiang Long, Wenju Xu, Shuangjiang He, and Huijuan Zhao. Feature representation learning with adaptive displacement generation and transformer fusion for micro-expression recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22086-22095, 2023. 1", + "[58] Sirui Zhao, Hanqing Tao, Yangsong Zhang, Tong Xu, Kun Zhang, Zhongkai Hao, and Enhong Chen. A two-stage 3D CNN based learning method for spontaneous microexpression recognition. Neurocomputing, 448:276-289, 2021. 7", + "[59] Sirui Zhao, Huaying Tang, Xinglong Mao, Shifeng Liu, Yiming Zhang, Hao Wang, Tong Xu, and Enhong Chen. DFME: A New Benchmark for Dynamic Facial MicroExpression Recognition. IEEE Transactions on Affective Computing, 15(3):1371-1386, 2024. 2, 6", + "[60] Ling Zhou, Qirong Mao, Xiaohua Huang, Feifei Zhang, and Zhihong Zhang. Feature Refinement: An expression-specific feature learning and fusion method for micro-expression recognition. Pattern Recognition, 122:108275, 2022. 7", + "[61] Xu Zhou, Jie Li, Gongjin Lan, Rongrong Ni, Angelo Cangelosi, Jiaxin Wang, and Xiaofeng Liu. Efficient lower" + ], + "bbox": [ + 91, + 90, + 482, + 902 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "layers parameter decoupling personalized federated learning method of facial expression recognition for home care robots. Information Fusion, 106:102261, 2024. 6", + "bbox": [ + 545, + 90, + 903, + 133 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "14463", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/FED-PsyAU_ Privacy-Preserving Micro-Expression Recognition via Psychological AU Coordination and Dynamic Facial Motion Modeling/e27ad4a5-3106-4216-bcea-a95709cd5f65_model.json b/2025/FED-PsyAU_ Privacy-Preserving Micro-Expression Recognition via Psychological AU Coordination and Dynamic Facial Motion Modeling/e27ad4a5-3106-4216-bcea-a95709cd5f65_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6ad239d3622e92c03f3bedaffd13aaaf74439fd4 --- /dev/null +++ b/2025/FED-PsyAU_ Privacy-Preserving Micro-Expression Recognition via Psychological AU Coordination and Dynamic Facial Motion Modeling/e27ad4a5-3106-4216-bcea-a95709cd5f65_model.json @@ -0,0 +1,2554 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.147, + 0.131, + 0.852, + 0.177 + ], + "angle": 0, + "content": "FED-PsyAU: Privacy-Preserving Micro-Expression Recognition via Psychological AU Coordination and Dynamic Facial Motion Modeling" + }, + { + "type": "text", + "bbox": [ + 0.268, + 0.203, + 0.731, + 0.222 + ], + "angle": 0, + "content": "Jingting Li\\(^{1,2}\\), Yu Qian\\(^{1,3}\\), Lin Zhao\\(^{1,2}\\), Su-Jing Wang\\(^{*1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.155, + 0.222, + 0.845, + 0.257 + ], + "angle": 0, + "content": "\\(^{1}\\)State Key Laboratory of Cognitive Science and Mental Health, Institute of Psychology, Chinese Academy of Sciences, Beijing, 100101, China" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.257, + 0.898, + 0.275 + ], + "angle": 0, + "content": "\\(^{2}\\)Department of Psychology, University of the Chinese Academy of Sciences, Beijing, 100049, China" + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.275, + 0.876, + 0.293 + ], + "angle": 0, + "content": "\\(^{3}\\)School of Computer, Jiangsu University of Science and Technology, Zhenjiang, 212100, China" + }, + { + "type": "text", + "bbox": [ + 0.204, + 0.295, + 0.788, + 0.31 + ], + "angle": 0, + "content": "\\(^{\\dagger}\\)Equal contribution, \\*Corresponding author (wangsujing@psych.ac.cn)" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.345, + 0.327, + 0.361 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.377, + 0.483, + 0.755 + ], + "angle": 0, + "content": "Micro-expressions (MEs) are brief, low-intensity, often localized facial expressions. They could reveal genuine emotions individuals may attempt to conceal, valuable in contexts like criminal interrogation and psychological counseling. However, ME recognition (MER) faces challenges, such as small sample sizes and subtle features, which hinder efficient modeling. Additionally, real-world applications encounter ME data privacy issues, leaving the task of enhancing recognition across settings under privacy constraints largely unexplored. To address these issues, we propose a FED-PsyAU research framework. We begin with a psychological study on the coordination of upper and lower facial action units (AUs) to provide structured prior knowledge of facial muscle dynamics. We then develop a DPK-GAT network that combines these psychological priors with statistical AU patterns, enabling hierarchical learning of facial motion features from regional to global levels, effectively enhancing MER performance. Additionally, our federated learning (FL) framework advances MER capabilities across multiple clients without data sharing, preserving privacy and alleviating the limited-sample issue for each client. Extensive experiments on commonly-used ME databases demonstrate the effectiveness of our approach. Our implementation is publicly available at https://github.com/MELABIPCAS/FED-PsyAU.git." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.784, + 0.222, + 0.8 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Micro-expressions (MEs) [9], resulting from conflicts between voluntary and involuntary expressions, reveal true emotions, making them valuable in non-contact emotion detection applications like criminal interrogation and psychological counseling. MEs are characterized by their brief duration, low intensity, localized occurrence, and occasional" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.346, + 0.905, + 0.619 + ], + "angle": 0, + "content": "asymmetry [10]. These traits not only restrict large-scale data collection and annotation but also make it challenging for deep learning networks to effectively model their motion characteristics. Some methods focus on regions of interest (ROIs) [4, 21, 31, 49], while others utilize attention mechanisms [3, 14, 33] and transformer-based architectures [24, 50, 57] to capture salient features. Meanwhile, MEs are often complex manifestations involving the combination of multiple AUs. In contrast, individual action units (AUs) correspond to specific, anatomically-based facial muscle movements. Focusing on AUs allows us to model more fundamental and consistent facial appearance changes. Certain approaches use AU labels to construct adjacency matrices for graph convolutional networks (GCNs) [24, 39]. Yet, to the best of our knowledge, these methods often apply AU annotations directly without fully exploiting the anatomical structure and coordinated AU dynamics, limiting the network's understanding on ME." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.623, + 0.905, + 0.851 + ], + "angle": 0, + "content": "To address these challenges, we leverage the theory that distinct neural pathways control upper and lower facial movements, resulting in two different motion patterns. Based on this, we conduct a psychological study to investigate upper/lower facial AU coordination, providing crucial prior knowledge on structural motion relationships underpinning MEs. Using both psychological priors and data-driven AU patterns, we design a Graph Attention Network (GAT) [47] based on dynamic prior knowledge, applied separately to the upper and lower regions. A final global GAT captures the topological relationships of muscle movements across the face. The resulting feature, combined with full-face optical flow (OF), is input into a subsequent dual-stream network, effectively integrating structural and dynamic information to enhance ME recognition (MER)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.903 + ], + "angle": 0, + "content": "Besides, as mentioned earlier, although MEs have valuable applications in many privacy-sensitive contexts, the sensitivity of biometric data makes it impractical to aggre" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14453" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.227 + ], + "angle": 0, + "content": "gate ME data across clients for training. Meanwhile, the sample size in a single client is often insufficient to train a robust model. Yet, the challenge of enhancing MER performance across diverse clients while preserving privacy remains largely unexplored. To address this, we adopt a federated learning (FL) framework, improving performance by merging and updating model parameters across clients without exchanging local data. This approach preserves data privacy and supports the MER practical deployment." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.228, + 0.482, + 0.306 + ], + "angle": 0, + "content": "Overall, our work leverages psychological and data-driven priors to enhance the network's learning of structured ME motion patterns, particularly in limited-sample conditions. Additionally, we introduce FL to advance MER in real-world scenarios. The contributions of this paper are:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.308, + 0.483, + 0.367 + ], + "angle": 0, + "content": "- We conduct a psychological experiment to derive AU coordination between the upper and lower facial regions, providing structural priors for ME and facial expression analysis in computer vision." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.369, + 0.482, + 0.458 + ], + "angle": 0, + "content": "- We design a local-to-global GAT network that integrates psychological priors and data-driven AU patterns in the upper and lower facial regions, improving the network's ability to learn both local and global structural facial dynamics. This feature, combined with global OF feature in a dual-stream network, enhances MER performance." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.459, + 0.482, + 0.504 + ], + "angle": 0, + "content": "- We employ FL in MER to improve data privacy and overcome small sample size challenges by aggregating features from distributed data-protected sources." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.308, + 0.483, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.52, + 0.242, + 0.536 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.546, + 0.357, + 0.563 + ], + "angle": 0, + "content": "2.1. Micro-expression Recognition" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In recent years, MER has gained increasing attention from computer vision researchers. Traditional machine learning methods primarily rely on handcrafted feature extraction. For instance, Local Binary Pattern (LBP) [43], effectively utilizes local texture features, has inspired various improvements [18, 49]. Furthermore, OF [26] has been explored for capturing object motion. Liu et al. [38] developed the Main Directional Mean OF (MDMO) method by aligning faces in the OF domain, while Liong et al. [34] added optical strain features for MER. However, traditional approaches depend heavily on manually crafted features, limiting both accuracy and generalizability. Meanwhile, deep learning methods, with robust feature extraction capabilities, can automatically learn multi-level representations to capture subtle ME changes. The availability of several spontaneous ME databases, such as CASME II [51], SAMM [6], CAS(ME)\\(^{3}\\) [27], DFME [59] and others [1, 30, 32, 44, 54], has accelerated deep learning advancements in MER. Inspired by Liong et al.'s work [35] on sequence redundancy, many studies [4, 13, 50] now use keyframes (onset and apex) as input, reducing model parameters while achieving excellent results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.394 + ], + "angle": 0, + "content": "Despite advancements in preprocessing and feature extraction, deep learning alone has yet to achieve high MER accuracy. As MEs stem from physiological muscle activity, incorporating prior knowledge offers meaningful context. FACS [10], which defines expressions through AUs, provides a basic theoretical framework for MER. Recent studies explore AU-ME relationships to enhance MER performance. Lei et al. [24] applied GCN [21] to model AU co-occurrences, especially around the eyebrow and mouth regions. Xie et al. [53] introduced an AU-assisted GCN, with modules for AU feature extraction and ME image generation. Wang et al. [48] used transformers to learn dynamic AU adjacency, boosting model generalization. These networks typically rely on AU labels from databases, without fully integrating structured facial muscle movements, AU interrelations, or their links to expressions. In this work, we refine structured ME feature extraction using a local-to-global GAT network, which integrates priors on facial AU coordination from psychological experiments with data-driven movement patterns." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.407, + 0.704, + 0.423 + ], + "angle": 0, + "content": "2.2. Federated Learning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.43, + 0.905, + 0.551 + ], + "angle": 0, + "content": "To assess MER across databases, ME Grand Challenge (MEGC) 2018 [56] introduced the Holdout-database Evaluation and Composite Database Evaluation tasks, utilizing the CASME II and SAMM databases. MEGC 2019 [45] expanded the CDE task to include the SMIC databases. While research on MEs predominantly utilizes these public datasets, existing methodologies do not address the privacy concerns that arise in practical applications." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.906, + 0.794 + ], + "angle": 0, + "content": "FL [41] could addresses ME data scarcity and privacy concerns through collaborative training. It includes two main phases: local training and global aggregation. Many existing FL works for expressions, while insightful, are often not open-source or use highly customized frameworks (e.g., non-random client data allocation [12], specialized parameter aggregation [11]) making standard methods more suitable for implementation and comparisons in our study. For instance, McMahan et al. [41] proposed FedAvg, a basic aggregation method that uses weighted averaging based on each client's data volume. Li et al. [28] introduced FedProx by adding a regularization term to the client loss function, aligning updates more closely with the global model. However, the one-global model-fits-all approach ignores client data diversity. To solve this, our approach extends FedProx by distributing a personalized global model to each client." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.811, + 0.905, + 0.845 + ], + "angle": 0, + "content": "3. Proposed Psychological Study on AU Incoordination" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "This study employed a behavioral experiment to investigate individuals' cognitive responses to AU incoordination in the upper and lower facial regions." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14454" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.163, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.099, + 0.158, + 0.16, + 0.169 + ], + "angle": 0, + "content": "(a) AU \\(4 + 9\\)" + }, + { + "type": "image", + "bbox": [ + 0.201, + 0.089, + 0.268, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.202, + 0.158, + 0.269, + 0.169 + ], + "angle": 0, + "content": "(b) AU 6+12" + }, + { + "type": "image", + "bbox": [ + 0.307, + 0.089, + 0.373, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.307, + 0.158, + 0.375, + 0.169 + ], + "angle": 0, + "content": "(c) AU 4+12" + }, + { + "type": "image", + "bbox": [ + 0.413, + 0.089, + 0.479, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.413, + 0.158, + 0.481, + 0.169 + ], + "angle": 0, + "content": "(d) AU 6+15" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.182, + 0.483, + 0.211 + ], + "angle": 0, + "content": "Figure 1. Examples of AU composite images. (a) and (b) are coordinated composites. (c) and (d) are uncoordinated composites." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.239, + 0.483, + 0.329 + ], + "angle": 0, + "content": "The study included 30 college participants (20 female, Mean= 22.22 years), a sample size considered adequate for statistical reliability under the Central Limit Theorem [23]. All procedures followed the Declaration of Helsinki and were approved by the institutional review board of the Institute of Psychology, Chinese Academy of Sciences." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.33, + 0.483, + 0.526 + ], + "angle": 0, + "content": "This behavioral experiment employed a within-subjects design. The independent variable was the type of AU composite image (coordinated vs. uncoordinated AU composites), as shown in Fig. 1. The face images were segmented anatomically. The dividing line was consistently present in all stimuli, serving as a control variable and would not influence the experiment's outcomes. To control for individual facial differences in emotion recognition, the original images used for composites were sourced from the same person [10]. Seventy-two composite images were generated from upper and lower facial AUs across six basic emotions, with each emotion represented by six coordinated and six uncoordinated expressions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.527, + 0.483, + 0.587 + ], + "angle": 0, + "content": "The dependent variables are accuracy rates and scores. Specifically, participants made two judgments for each composite image: whether it was coordinated or uncoordinated, and the degree of coordination or incoordination." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.588, + 0.483, + 0.815 + ], + "angle": 0, + "content": "Statistical analysis from this study showed that participants were significantly more accurate in recognizing coordinated AU composite images than uncoordinated ones (Please see Supplementary Materials (Suppl.) for more experimental details and statistical results). The study further highlighted specific AU combinations, such as AU4 and AU12, that participants quickly identified as uncoordinated, whereas combinations like AU6 and AU12 were rapidly recognized as coordinated. This suggests that these uncoordinated AU pairings may inherently convey emotional incoordination. Besides, the coordination patterns between specific AUs reveal systematic facial action structures. These patterns help the network distinguish both coordinated and uncoordinated facial movements, thus enhancing its ability to identify ME features more effectively." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.83, + 0.445, + 0.848 + ], + "angle": 0, + "content": "4. Our Proposed FED-PsyAU Framework" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We propose the FED-PsyAU framework (Fig. 2), which embeds a psychology-driven MER model within a federated aggregation module. The model features a hierarchi" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "cal structure that learns from local ROIs to AU groups and finally to global facial regions, capturing dynamic topology while minimizing redundancy. It leverages a multiscale InceptionNet to extract and fuse global OF with structural facial features. The federated module enables privacy-preserving collaborative training to boost MER performance on each client." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.218, + 0.742, + 0.234 + ], + "angle": 0, + "content": "4.1. Localized ROI Modeling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.243, + 0.905, + 0.349 + ], + "angle": 0, + "content": "Since MEs are manifested by muscle movements in key facial regions (e.g., eyes, mouth, nose) [10], we refine the standard 68 landmarks from Dlib [20] to precisely capture these changes. As shown in Fig. 3, we discard 10 less informative outer contour landmarks and add 7 new midpoints to better track movements in the cheek and eyebrow regions, resulting in a final set of 65 keypoints." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.353, + 0.907, + 0.58 + ], + "angle": 0, + "content": "Then, onset-apex OF inputs \\( (128 \\times 128 \\) pixels) are cropped into 65 ROIs, each measuring \\( 5 \\times 5 \\) pixels, centered on selected facial landmarks \\( \\mathbf{p}_k = (x_k, y_k) \\), where \\( k \\in 1, 2, \\ldots, 65 \\), denotes the landmark index. To enhance robustness to apex annotation, we also compute OF using a small window of frames around the apex. The ROI features input to the Local ROI Feature Extractor (LFE), denoted as \\( \\Theta_k \\), contain the horizontal and vertical components of OF and optical strain. The LFE module consists of three convolutional layers, each followed by batch normalization and ReLU activation. All kernel sizes were \\( 3 \\times 3 \\). The output channels in the first two layers increase from 32 to 64, while the final layer outputs 3 channels. This design helps the LFE capture complex local patterns, with the output features referred to as \\( \\Phi_k = \\mathrm{LFE}(\\Theta_k) \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.907, + 0.78 + ], + "angle": 0, + "content": "Inspired by Vision Transformer [8], we use a Spatial Structure Encoder (SSE), i.e., Transformer Encoder with three attention heads to capture diverse dependencies among ROIs, as MEs result from coordinated muscle movements across multiple areas. Precisely, we convert \\(\\Phi_k\\), into a token. To retain the positional information of the facial landmarks, a positional embedding \\(E_{k}\\) is added based on the landmark index to each token, yielding \\(X_{k} = \\Phi_{k} + E_{k}\\). We add the output of the three head attention module to the input feature \\(X_{k}\\), and then normalize the result using LayerNorm. Finally, after passing through a feed-forward neural network and applying LayerNorm, the final output representation is obtained as \\(Z_{k}\\in \\mathbb{R}^{3\\times 5\\times 5}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.8, + 0.881, + 0.816 + ], + "angle": 0, + "content": "4.2. Relationship Modeling among AU Features" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Our proposed Relationship Modeling among AU Features (AFR) module consists of two components: an AU Feature Extractor (AFE) to derive features from AU-specific facial regions, and a GAT-structure to model the topological relationships between these AUs." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14455" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.905, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.286, + 0.907, + 0.373 + ], + "angle": 0, + "content": "Figure 2. The pipeline of our proposed FED-PsyAU framework, including MER network (black block) and the Federated Aggregation (FA) module (orange block). In MER network, the LRM module consists of the local ROI feature extractor (LFE) and the spatial structure encoder (SSE), which preserves spatial structure information while efficiently extracting ROI features; the AFR module includes the AU feature extractor (AFE) and the dynamic prior knowledge-based graph attention network (DPK-GAT) for effective AU feature extraction and capturing AU dynamic relationships, the DSI module use a dual-stream structure to capture multi-scale facial muscle topology and motion information for MER. Then, the FA module integrates the MER network into the FL framework." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.393, + 0.475, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.594, + 0.487, + 0.665 + ], + "angle": 0, + "content": "Figure 3. ROI grouping strategy for AU. OF features from each ROI are processed through LFE and SSE to yield high-dimensional representations. ROIs are then grouped by AU movement patterns to create inputs for AFEs, ultimately producing AU node features for the GAT network." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.696, + 0.291, + 0.71 + ], + "angle": 0, + "content": "4.2.1. AU Feature Extractor" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.717, + 0.483, + 0.884 + ], + "angle": 0, + "content": "MEs involve dynamic, coordinated interactions across multiple regions [10] and are characterized by subtle movements that differ from typical macro-expressions. Therefore, as shown in Fig. 3, when grouping ROIs for the AFE module, the AU selection and regionalization are based on: 1) AU frequency in MEs; 2) Grouping by primary/secondary muscle activation areas [7]. Notably, not all AUs employed in Section 3 are applied to MER. We specifically retain only the 12 AUs exhibiting the highest relevance to ME dynamics, as listed in Tab. 1. Detailed occurrences are provided in the Suppl." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The first step we construct in AFE is Group Squeeze and" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.395, + 0.907, + 0.438 + ], + "angle": 0, + "content": "Table 1. Selected AUs and Corresponding Regions (Primary and Secondary/ Coordinate). \\( g \\), \\( N_g \\) and \\( N_{g\\_AU} \\) represent ROI Group index, corresponding ROI number and AU amount." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.447, + 0.908, + 0.516 + ], + "angle": 0, + "content": "
gAUPrimarySecondary/ CoordinateNgNg-AU
11,2,4EyebrowsEyes233
25,6,7EyesEyebrows, Cheeks, Mouth483
39,10Nose, CheeksMouth, Chin382
412,14,15,17MouthEyes, Eyebrows, Nose524
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.907, + 0.659 + ], + "angle": 0, + "content": "Excitation (GSE). It computes the attention for each group of AUs (ROI group), which makes the model pay more attention to the critical ROIs. Specifically, for each ROI group input \\( R_{g} = \\{Z_{1}, Z_{2}, \\ldots, Z_{N_{g}}\\} \\), \\( g \\in \\{1, 2, 3, 4\\} \\), the GSE module performs global average pooling on each ROI and computes channel importance through a fully connected (FC) layer. The original features are then preserved through residual concatenation:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.668, + 0.907, + 0.687 + ], + "angle": 0, + "content": "\\[\nf _ {\\mathrm {G S E}} \\left(R _ {g}\\right) = \\left\\{W _ {k} ^ {\\text {a t t}} \\odot Z _ {k} + Z _ {k} \\mid Z _ {k} \\in R _ {g} \\right\\} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.694, + 0.909, + 0.846 + ], + "angle": 0, + "content": "where \\( W_{k}^{att} \\) is the attention weight for each group of ROI, \\( \\odot \\) denotes element-wise multiplication along the channel dimension of \\( Z_{k} \\). Next, \\( f_{\\mathrm{GSE}}(R_g) \\) undergoes further feature extraction, reducing from \\( N_{g} \\) composite features to \\( N_{g\\_AU} \\) representative AU features, yielding the output feature \\( F_{g\\_AU} \\). Precisely, the process comprises two convolutional layers: a \\( 3 \\times 3 \\) layer that doubles the input dimensions and a \\( 1 \\times 1 \\) layer that restores them to the original size. Both layers are followed by Batch Normalization and ReLU activation for improved stability and representation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.852, + 0.747, + 0.868 + ], + "angle": 0, + "content": "4.2.2. AU Relationship Modeling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "Prior Knowledge: To model AU dynamic relationships, we integrate two forms of prior knowledge: an \"intrinsic\" prior" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14456" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.091, + 0.482, + 0.273 + ], + "angle": 0, + "content": "from psychological principles and an \"extrinsic\" prior from data-driven patterns. The intrinsic prior, derived from our psychological findings on AU coordination (Sec. 3), provides structural guidance by defining the initial GAT adjacency matrix \\( A \\). The extrinsic prior, based on statistical AU co-occurrence patterns in the data, offers data-driven refinement that theoretical models may overlook. This extrinsic prior forms a prior attention matrix \\( D \\) to guide weight allocation among AU nodes. By fusing this foundational psychological structure with real-world dynamics, our model achieves superior adaptability and generalization (see Suppl. for matrix details)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.274, + 0.483, + 0.485 + ], + "angle": 0, + "content": "Facial Segmentation: As presented in Sec. 3, the upper and lower facial regions, controlled by different neural pathways, exhibit distinct muscle movement patterns. AUs and their associated primary muscles in the upper and lower face can exhibit coordination (potential co-occurrence) or lack of coordination (infrequent simultaneous appearance). Leveraging these intra- and inter-regional interactions, we model AU relationships for the upper face, lower face, and whole face. Consistent with the experimental setup in Sec. 3, we segment the face into upper and lower parts based on the orbicularis oculi muscle. Specifically, following the primary occurrence regions of 12 listed AUs, the AUs in the first two groups from Tab. 1 are classified as upper face AUs, while those in the last two groups are classified as lower face AUs." + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.497, + 0.483, + 0.517 + ], + "angle": 0, + "content": "\\[\nF _ {A U} ^ {U p p e r} = F _ {1 \\_ A U} \\cup F _ {2 \\_ A U}, F _ {A U} ^ {L o w e r} = F _ {3 \\_ A U} \\cup F _ {4 \\_ A U} (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.531, + 0.483, + 0.651 + ], + "angle": 0, + "content": "We input the AU features from the upper and lower face: \\( F_{AU}^{Upper} \\), \\( F_{AU}^{Lower} \\) into our proposed Dynamic Prior Knowledge GATs (DPK-GATs) in parallel (Details are presented in the next part of this subsection). These networks use psychology prior knowledge to build AU adjacency relationships and data-driven priors to guide attention learning, producing outputs \\( h^{Upper} \\) and \\( h^{Lower} \\), which are combined into \\( h^{Local} \\) for the next stage." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.653, + 0.483, + 0.744 + ], + "angle": 0, + "content": "In the whole-face DPK-GAT, we construct the network similarly but set adjacency values within the same region to zero, ignoring intra-regional AU relationships. By learning cross-regional topological structures and statistical patterns, we obtain global AU relationship features, denoted as \\( h^{Global} \\), as illustrated in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.744, + 0.483, + 0.852 + ], + "angle": 0, + "content": "DPK-GAT: We model dynamic AU node relationships combing GAT and the prior knowledge. In particular, for the input feature belonging to \\(\\{F_{AU}^{Upper}, F_{AU}^{Lower}, h^{Local}\\}\\), a linear mapping projects it to a higher-dimensional space as \\(f\\). The similarity score \\(e_{ij}\\) between AU nodes \\(i\\) and \\(j\\), is calculated using the attention weight vector \\(\\vec{a}\\) and concatenated node features \\(f_{i}\\) and \\(f_{j}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.128, + 0.864, + 0.483, + 0.905 + ], + "angle": 0, + "content": "\\[\ne _ {i j} = \\left\\{ \\begin{array}{l l} \\text {L e a k y R e L U} \\left(\\tilde {a} ^ {T} [ f _ {i} | | f _ {j} ]\\right), & \\text {i f} A _ {i j} > 0 \\\\ - \\infty , & \\text {i f} A _ {i j} = 0 \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.905, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.207, + 0.905, + 0.263 + ], + "angle": 0, + "content": "Figure 4. DPK-GAT from facial segmentations to global face. First, intra-regional AU node relationships are learned within the upper and lower facial regions. Second, cross-regional relationships are modeled to integrate interactions between these regions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.289, + 0.905, + 0.319 + ], + "angle": 0, + "content": "Here, only neighboring nodes defined by the psychology-derived adjacency matrix \\(A\\) are considered." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.32, + 0.905, + 0.366 + ], + "angle": 0, + "content": "The final attention \\(\\alpha_{ij}\\) is calculated by combining the data-driven prior attention \\(D_{ij}\\) and the self-attention, obtained by the softmax operation applied to \\(e_{ij}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.378, + 0.905, + 0.395 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {i j} = (1 - \\beta) \\cdot \\operatorname {s o f t m a x} \\left(e _ {i j}\\right) + \\beta \\cdot D _ {i j} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.405, + 0.905, + 0.464 + ], + "angle": 0, + "content": "where, \\(\\beta\\) is the dynamic prior attention weight, updated based on the training rounds. Along with a residual concatenation, the final node representation \\(h_i\\) is aggregated as:" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.465, + 0.905, + 0.485 + ], + "angle": 0, + "content": "\\[\nh _ {i} = \\operatorname {E L U} \\left(\\sum_ {j \\in \\mathcal {N} (i)} \\alpha_ {i j} \\cdot f _ {j}\\right) + f _ {i} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.49, + 0.905, + 0.552 + ], + "angle": 0, + "content": "where \\(\\mathcal{N}(i)\\) denotes the neighboring nodes of \\(i\\). In our implementation, we employ three attention heads and a two-layer structure, enabling the extraction of complex AU dependencies." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.56, + 0.755, + 0.577 + ], + "angle": 0, + "content": "4.3. Dual-Stream-InceptionNet" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.582, + 0.906, + 0.78 + ], + "angle": 0, + "content": "Muscle movements vary widely across facial regions, making multi-scale information essential for the MER task. InceptionNet's core concept is leveraging parallel convolutional kernels to extract multi-scale features, enhancing the network's ability to capture fine details and overall structure [46]. We build on this by using a dual-stream InceptionNet as the ME feature extractor. One stream processes the AU relationship features \\( h^{Global} \\), while the other handles global OF features \\( \\Theta \\). This dual-stream design allows the model to capture the topological relationships of facial muscles through AUs and holistic motion via OF. The output tensors from both streams are flattened, concatenated, and fed into a FC layer for ME prediction." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.789, + 0.727, + 0.805 + ], + "angle": 0, + "content": "4.4. Federated Aggregation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.906, + 0.9 + ], + "angle": 0, + "content": "Traditional centralized learning consolidates all data on a single server, enhancing model accuracy and generalization but neglecting privacy concerns in MER. FL could address this by training models locally and sharing only parameters, protecting data privacy and retaining each client's unique data characteristics." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14457" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.239 + ], + "angle": 0, + "content": "FedProx [29] is a FL approach designed to address data heterogeneity among clients. It incorporates proximal terms during model aggregation to balance global and client-local learning, improving both performance and convergence speed. We extend FedProx by altering the server's processing and distribution strategy during aggregation, which we call Personalized FedProx (P-FedProx). Specifically, P-FedProx initializes each client's model \\( W_{i}^{t} \\) in round \\( t \\) by combining its previous model \\( \\mathring{W}_{i}^{t-1} \\) with models from other clients \\( \\mathring{W}_{j}^{t-1} \\)." + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.247, + 0.484, + 0.273 + ], + "angle": 0, + "content": "\\[\nW _ {i} ^ {t} = \\theta \\cdot \\mathring {W} _ {i} ^ {t - 1} + \\sum_ {\\substack {j = 1 \\\\ j \\neq i}} ^ {n} \\omega_ {j} \\cdot \\mathring {W} _ {j} ^ {t - 1}, \\text {where} \\sum_ {\\substack {j = 1 \\\\ j \\neq i}} ^ {n} \\omega_ {j} = 1 - \\theta \\tag{6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.279, + 0.484, + 0.326 + ], + "angle": 0, + "content": "where weight \\(\\omega_{j}\\) is proportional to \\(j\\)th client's data size. For \\(i\\)th client, \\(\\theta\\) is set to 0.9 (optimal experimental result, see more detail in the Suppl.)" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.333, + 0.236, + 0.348 + ], + "angle": 0, + "content": "4.5. Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.354, + 0.484, + 0.48 + ], + "angle": 0, + "content": "AU Prediction Loss: To enhance MER, our network models the topological relationships and statistical patterns of AUs, making AU prediction an auxiliary task. The prediction is utilized in two stages: one enhances the AU feature extraction performance of the AFE module, and the other improves the GAT network's ability to learn AU topological structures. AU recognition is formulated as a multi-label classification problem, with the Binary Cross-Entropy Loss for each AU label defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.485, + 0.484, + 0.517 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {A}} = \\frac {\\sum_ {j = 1} ^ {M} \\left[ y _ {j} \\cdot \\log \\left(\\sigma \\left(\\hat {y} _ {j}\\right)\\right) + \\left(1 - y _ {j}\\right) \\cdot \\log \\left(1 - \\sigma \\left(\\hat {y} _ {j}\\right)\\right) \\right]}{- M} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.525, + 0.483, + 0.601 + ], + "angle": 0, + "content": "where \\(M\\) is the total number of AUs, \\(y_{j}\\), \\(\\sigma (\\hat{y}_j)\\) and \\(\\hat{y}_j\\) represent respectively the ground truth label (0 or 1), the logit output and prediction after applying the sigmoid function for the \\(j\\)-th AU. The loss function is consistent for both stages, i.e., \\(L_{\\mathrm{A}}^{\\mathrm{AFE}}\\) and \\(L_{\\mathrm{A}}^{\\mathrm{GAT}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.601, + 0.484, + 0.643 + ], + "angle": 0, + "content": "MER Loss: The Cross-Entropy Loss function is utilized for emotion classification. Thus, our overall MER loss is composed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.649, + 0.483, + 0.668 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {M E R}} = \\alpha_ {1} \\cdot \\left(- \\sum_ {i = 1} ^ {N} y _ {i} \\cdot \\log (\\hat {y} _ {i})\\right) + \\alpha_ {2} \\cdot L _ {\\mathrm {A}} ^ {\\mathrm {A F E}} + \\alpha_ {3} \\cdot L _ {\\mathrm {A}} ^ {\\mathrm {G A T}} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.751 + ], + "angle": 0, + "content": "where, \\(N\\) represents the number of classes, \\(y_{i}\\) and \\(\\hat{y}_i\\) are the one-hot encoded ground truth label and the predicted probability for class \\(i\\), \\(\\alpha_{1}\\), \\(\\alpha_{2}\\) and \\(\\alpha_{3}\\) are hyperparameters, set to 0.2, 0.8, and 0.8 to prioritize MER loss while balancing AU recognition (See parameter analysis in Suppl.)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.751, + 0.484, + 0.797 + ], + "angle": 0, + "content": "Federated Loss: During federated training, the local client's loss function includes the MER loss \\( L_{\\mathrm{MER}} \\) and a proximal term:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.805, + 0.483, + 0.833 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {F L}} = L _ {\\mathrm {M E R}} + \\frac {\\alpha_ {4}}{2} \\cdot \\left\\| W - W _ {t} \\right\\| ^ {2} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\( W_{t} \\) denotes the global model in the server's \\( t \\)-th round of distribution, and \\( W \\) represents the current model of the client. \\( \\alpha_{4} \\) is the hyperparameter for the weight of the proximal term." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.646, + 0.108 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.123, + 0.659, + 0.14 + ], + "angle": 0, + "content": "5.1. Configuration" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.153, + 0.906, + 0.259 + ], + "angle": 0, + "content": "Databases: We use two recently published large scale and challenging ME databases: CAS(ME)\\(^3\\) and DFME. CAS(ME)\\(^3\\)-Part A offers 860 samples from 100 participants. DFME includes three subsets with (participants, samples): A (72,1,118), B (92,969), C (492,5,439). For Apex-first-frame scenario, we compute the OF between the apex and mid-frame towards the offset." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.267, + 0.906, + 0.328 + ], + "angle": 0, + "content": "Pre-training: To enhance the model's ability to perceive facial motion patterns, we first pre-trained the AU module (LRM+AFR) on the DISFA [40] and CK [19] datasets. The effect of this pre-training is detailed in the Suppl." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.335, + 0.907, + 0.485 + ], + "angle": 0, + "content": "MER Experiment Settings: The evaluation follows standard protocols on two datasets. On \\(\\mathrm{CAS(ME)^3}\\), we perform 3-class classification (positive, negative, surprised) [45] with Leave-One-Subject-Out (LOSO) validation. On DFME, we conduct 7-class classification (Happiness, Surprise, Disgust, Sadness, Anger, Fear, and Contempt) using scale-adapted 10-fold cross-validation [59]. To address class imbalance, performance is measured by the UF1 and UAR. Experiments were implemented in PyTorch 1.13.0 and run on five NVIDIA GeForce RTX 4090 GPUs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.494, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Federated Experiment Settings: The DFME and \\(\\mathrm{CAS(ME)^3}\\) datasets are randomly split into 5 and 2 local clients, respectively, with each client assigned an equal number of subjects. However, variations in microexpression (ME) counts per subject result in differing ME data distributions and quantities across clients, mirroring real-world data heterogeneity (See detailed data distributions in the Suppl.). Traditional LOSO and k-fold cross-validation can leak training samples in a federated setting, so we employ a random partition strategy with \\(70\\%\\) of samples for training and \\(30\\%\\) for testing, averaging results over 10 tests to minimize randomness. Furthermore, we assume an ideal FL scenario with no client dropouts and equal local updates per client. In FedAvg [41] and FedProx [28], the server assigns an identical global model to all clients each round; differently, FedProx adds a proximal term to reduce model bias in local training. Our P-FedProx enhances this by assigning tailored initial models, with \\(90\\%\\) (\\(\\theta\\)) weight on each client's own model and \\(10\\%\\) on others'. Additionally, we compare against recent personalized FL methods: FedRep [5] uses shared representations with personalized heads, ELLP [61] employs parameter decoupling with Fisher Information weighting, and FedAS [55] applies parameter alignment and client synchronization. Unlike these approaches requiring architectural changes or complex procedures, P-FedProx achieves personalization through simple weighted initialization." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14458" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.1, + 0.089, + 0.474, + 0.104 + ], + "angle": 0, + "content": "Table 2. SOTA methods comparison on DFME and CAS(ME) \\( {}^{3} \\)" + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.114, + 0.483, + 0.334 + ], + "angle": 0, + "content": "
CategoryMethodYearDFMECAS(ME)3
UF1UARUF1UAR
3D-CNNI3D [2]20170.29230.3058--
R3D [17]20180.21640.2313--
Hand-craftedLBP-TOP [49]20140.23360.2653--
MDMO [37]20150.24890.2939--
Deep LearningOFF-ApexNet [15]20190.23860.2806--
AlexNet [22]2012--0.25700.2634
STSTNet [36]20190.27140.31080.37950.3792
RCN-A [52]20200.27510.31230.39280.3893
MERSiam [58]20210.31840.35320.31840.3532
FGRL [24]20210.07360.14290.33330.2636
FR [24, 60]20220.35590.38140.34930.3413
MMNet [25]20220.26490.27760.37060.3646
BDCNN [4]20220.29750.33140.50500.5164
Micro-BERT [42]2023--0.56040.6125
HTNet [50]20240.32430.33680.57670.5415
HSTA [16]2024--0.5930.618
Ours-0.38530.39780.62210.6226
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.36, + 0.436, + 0.377 + ], + "angle": 0, + "content": "5.2. Comparison to State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.382, + 0.483, + 0.413 + ], + "angle": 0, + "content": "Tab. 2 list the SOTA comparisons on DFME and CAS(ME)\\(^3\\). The confusion matrices are shown in the Suppl." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.414, + 0.483, + 0.609 + ], + "angle": 0, + "content": "Regarding DFME, classifying ME samples into seven categories poses significant challenges, resulting in relatively low UF1 and UAR scores across all methods. However, our proposed method outperforms all others in every metric. Specifically, Our method outperforms I3D by \\(9.30\\%\\) in UF1 and \\(9.20\\%\\) in UAR, and surpasses LBP-TOP by \\(15.17\\%\\) in UF1 and \\(13.25\\%\\) in UAR, highlighting its effectiveness in extracting critical and discriminative features from ME clips. Additionally, compared to recent deep learning methods like FR, our approach yields \\(2.94\\%\\) and \\(1.64\\%\\) higher UF1 and UAR, confirming the benefits of incorporating facial topology and structured muscle motion for enhanced performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.611, + 0.483, + 0.853 + ], + "angle": 0, + "content": "Regarding CAS(ME)\\(^3\\), the primary challenge lies in its sample complexity. Compared to baseline methods (STST-Net, RCN-A, FR), our proposed method achieves over approximately \\(20\\%\\) higher UF1 and UAR scores, marking a significant advancement. Our method surpasses three recent SOTA approaches in MER, including HTNet, which employs a hierarchical Transformer for local feature learning via self-attention. Specifically, our approach outperforms HTNet by \\(4.54\\%\\) in UF1 and \\(8.11\\%\\) in UAR, leveraging psychological insights and statistical patterns to capture deeper facial muscle dynamics. Additionally, it exceeds Micro-BERT and HSTA by \\(6.17\\%\\) and \\(2.91\\%\\) in UF1, and \\(1.01\\%\\) and \\(0.46\\%\\) in UAR, respectively. By mining facial movement patterns to better understand topology, our model achieves improved metrics, indicating enhanced accuracy in classifying all emotion categories." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.863, + 0.245, + 0.88 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.415, + 0.901 + ], + "angle": 0, + "content": "Tab. 3 listed the ablation study results on DFME." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.907, + 0.118 + ], + "angle": 0, + "content": "Table 3. Ablation study. L, G, S and D represent Local (upper and lower face), Global, Single stream and Dual Stream, respectively." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.128, + 0.905, + 0.349 + ], + "angle": 0, + "content": "
SettingLRMAFRInceptionNetUF1↑UAR↑
LFESSELGSD
I✓0.32490.3404
II✓✓0.33620.3517
III✓✓✓0.34800.3624
IV✓✓✓✓0.36290.3723
V✓✓✓✓0.36040.3658
VI✓✓✓✓✓0.38530.3978
SettingWithout AU GroupWith AU GroupUF1↑UAR↑
VII✓0.36830.3793
VIII✓0.38530.3978
SettingPsychological priorData-driven PriorUF1↑UAR↑
IX0.36150.3701
X✓0.37230.3832
XI✓0.37050.3767
XII✓✓0.38530.3978
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.378, + 0.907, + 0.514 + ], + "angle": 0, + "content": "Impact of AU Feature: Setting I, which uses only full face OF as input, serves as our baseline for evaluating the role of AU features in MER. In Setting II, integrating AU features into the Dual Stream InceptionNet leads to a \\(1.13\\%\\) improvement in both UF1 and UAR. This improvement demonstrates that incorporating AU features significantly enhances the model's ability to recognize MEs, and underscores the value of fusing multiple feature types to strengthen model performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.517, + 0.907, + 0.592 + ], + "angle": 0, + "content": "Impact of ROI Relationship Modeling: The results in Setting III show improvements of \\(1.18\\%\\) in UF1 and \\(1.07\\%\\) in UAR compared to Setting II, indicating that the model effectively captures complex dependencies between ROIs and extracts more discriminative feature representations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.595, + 0.906, + 0.67 + ], + "angle": 0, + "content": "Impact of AU grouping: We compared AU feature extraction from the global face with our AU grouping strategy (Setting VII vs. VIII), the performance improvement (UF1: \\(1.70\\%\\), UAR: \\(1.85\\%\\)), confirming that AU grouping enables the network to learn more representative AU features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.672, + 0.906, + 0.822 + ], + "angle": 0, + "content": "Impact of Modeling the Topological Structure of ME Muscle Movements: Compared to Setting III, Setting IV, including local muscle movement topology modeling, improves UF1 and UAR by \\(1.49\\%\\) and \\(0.99\\%\\), respectively. Setting V, which adds global muscle movement topology modeling, shows gains of \\(1.24\\%\\) in UF1 and \\(0.34\\%\\) in UAR over Setting III. Combining both local and global topology modeling yields the best performance on DFME. These results demonstrate that incorporating topological modeling of muscle movements enhances MER." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Impact of Prior Knowledge: Tab. 3 presents the ablation experiments to evaluate the impact of using prior knowledge for constructing facial muscle topology. In particular, compared to Setting IX, Setting X, which uses only psychological knowledge, improves UF1 and UAR by \\(1.08\\%\\) and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14459" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.322, + 0.093, + 0.676, + 0.111 + ], + "angle": 0, + "content": "FedAvg(2017) PersonalizedFedProx ELLP(2024) Baseline FedProx(2020) FedRep(2021) FedAS(2024)" + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.129, + 0.223, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.131, + 0.334, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.131, + 0.444, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.131, + 0.554, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.131, + 0.664, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.131, + 0.773, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.774, + 0.131, + 0.885, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.383, + 0.234, + 0.615, + 0.246 + ], + "angle": 0, + "content": "(a) UF1 comparison across different clients." + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.248, + 0.227, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.248, + 0.336, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.248, + 0.445, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.248, + 0.555, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.248, + 0.664, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.664, + 0.248, + 0.773, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.773, + 0.248, + 0.884, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.36, + 0.35, + 0.638, + 0.362 + ], + "angle": 0, + "content": "(b) Convergence comparison across different clients." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.374, + 0.905, + 0.39 + ], + "angle": 0, + "content": "Figure 5. Compare the performance of different FL frameworks on different clients: UF1 and convergence. See UAR comparison in Suppl." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.415, + 0.484, + 0.552 + ], + "angle": 0, + "content": "\\(1.31\\%\\), respectively. Setting XI, incorporating only data-driven knowledge, shows gains of \\(0.9\\%\\) in UF1 and \\(0.66\\%\\) in UAR. When both psychological and data-driven knowledge are combined (Setting XII), the improvements reach \\(2.38\\%\\) in UF1 and \\(2.77\\%\\) in UAR, proving the effectiveness of the prior knowledge. Constructing a prior adjacency matrix allows the model to capture complex facial muscle topology, while data-driven knowledge directs attention to uncover intrinsic movements patterns of facial muscles." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.562, + 0.303, + 0.578 + ], + "angle": 0, + "content": "5.4. Federated Experiment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.484, + 0.659 + ], + "angle": 0, + "content": "Since FL aims to enhance MER performance on individual clients in scenarios with limited sample sizes and data privacy issue, and given that DFME and \\(\\mathrm{CAS}(\\mathrm{ME})^3\\) have large sample size, we partition the datasets as mentioned in Section 5.1 and treat the subdivisions as separate clients." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.484, + 0.902 + ], + "angle": 0, + "content": "While our focus is on application rather than fundamental FL innovation, our experiments confirm the effectiveness of our P-FedProx method, which is tailored for ME data distributions. Benchmarked against several open-source SOTA methods, our approach excels at mitigating small-sample challenges while preserving data privacy. Specifically (Fig. 5), the results show P-FedProx outperforms both the single-client baseline and traditional FL methods (FedAvg, FedProx), with particularly significant gains on clients with more challenging data distributions (Clients 6-7). Compared to recent personalized FL methods, our approach surpasses FedRep and ELLP and achieves performance competitive with the state-of-the-art FedAS, yet with a significantly simpler implementation. These consistent gains demonstrate that our weighted model initialization effectively addresses the data heterogeneity that challenges" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.416, + 0.905, + 0.446 + ], + "angle": 0, + "content": "traditional FL, thereby mitigating ME sample scarcity while ensuring data privacy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.447, + 0.906, + 0.522 + ], + "angle": 0, + "content": "Meantime, we note that all FL methods achieved good performance on the local clients from DFME, which may be due to the fact that these clients have a similar amount of data and slight differences in data distribution, presenting fewer challenges for federated learning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.906, + 0.645 + ], + "angle": 0, + "content": "Besides, we analyze the training loss progression of different federated algorithms over communication rounds. We found that P-FedProx demonstrates the smoothest convergence curves with minimal oscillations, particularly evident in Clients 1-5. On the other hand, P-FedProx converges significantly faster than other FL methods. This rapid early convergence suggests that personalized global model assignment provides clients with better initialization states." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.663, + 0.634, + 0.679 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.901 + ], + "angle": 0, + "content": "MER is challenged by limited sample sizes, subtle features, and privacy concerns in real-world applications. To address these issues, we conduct a psychological study to provide structured insights into facial muscle dynamics. We then develop a FED-PsyAU framework that integrates these findings with data-driven patterns for hierarchical learning of dynamic features, enhancing MER performance. Additionally, the FL framework improves MER across clients while preserving data privacy through iterative model updates. The comprehensive experiments validate our approach's effectiveness. In the future, we will integrate additional AUs to build a more comprehensive ME topology information and further advance MER in real-world applications through FL and semi-supervised approaches." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14460" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.264, + 0.108 + ], + "angle": 0, + "content": "7. Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.486, + 0.161 + ], + "angle": 0, + "content": "This research was partially funded by 1) the National Natural Science Foundation of China (62476269, 62276252); 2) the Youth Innovation Promotion Association CAS." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.173, + 0.188, + 0.189 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.199, + 0.484, + 0.268 + ], + "angle": 0, + "content": "[1] Xianye Ben, Yi Ren, Junping Zhang, Su-Jing Wang, Kidiyo Kpalma, Weixiao Meng, and Yong-Jin Liu. Video-Based Facial Micro-Expression Analysis: A Survey of Datasets, Features and Algorithms. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):5826-5846, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.269, + 0.484, + 0.324 + ], + "angle": 0, + "content": "[2] Joao Carreira and Andrew Zisserman. Quo Vadis, Action Recognition? A New Model and the Kinetics Dataset. In proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.326, + 0.484, + 0.38 + ], + "angle": 0, + "content": "[3] Boyu Chen, Zhihao Zhang, Nian Liu, Yang Tan, Xinyu Liu, and Tong Chen. Spatiotemporal Convolutional Neural Network with Convolutional Block Attention Module for Micro-Expression Recognition. Information, 11(8):380, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.382, + 0.484, + 0.45 + ], + "angle": 0, + "content": "[4] Bin Chen, Kun-Hong Liu, Yong Xu, Qing-Qiang Wu, and Jun-Feng Yao. Block Division Convolutional Network With Implicit Deep Features Augmentation for Micro-Expression Recognition. IEEE Transactions on Multimedia, 25:1345–1358, 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.451, + 0.484, + 0.507 + ], + "angle": 0, + "content": "[5] Liam Collins, Hamed Hassani, Aryan Mokhtari, and Sanjay Shakkottai. Exploiting shared representations for personalized federated learning. In International conference on machine learning, pages 2089-2099. PMLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.508, + 0.484, + 0.563 + ], + "angle": 0, + "content": "[6] Adrian K. Davison, Cliff Lansley, Nicholas Costen, Kevin Tan, and Moi Hoon Yap. SAMM: A Spontaneous MicroFacial Movement Dataset. IEEE Transactions on Affective Computing, 9(1):116-129, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.484, + 0.619 + ], + "angle": 0, + "content": "[7] Zizhao Dong, Gang Wang, Shaoyuan Lu, Jingting Li, Wenjing Yan, and Su-Jing Wang. Spontaneous Facial Expressions and Micro-Expressions Coding: From brain to Face. Frontiers in Psychology, 12:784834, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.621, + 0.484, + 0.702 + ], + "angle": 0, + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.704, + 0.484, + 0.731 + ], + "angle": 0, + "content": "[9] Paul Ekman and Wallace V Friesen. Nonverbal Leakage and Clues to Deception. Psychiatry, 32(1):88-106, 1969. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.733, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[10] Paul Ekman and Wallace V Friesen. Facial Action Coding System. Environmental Psychology & Nonverbal Behavior, 1978. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.483, + 0.814 + ], + "angle": 0, + "content": "[11] Ali et al. Privacy preserving personalization for video facial expression recognition using federated learning. IMCI'22, . 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.859 + ], + "angle": 0, + "content": "[12] Fan et al. Feeling without sharing: A federated video emotion recognition framework via privacy-agnostic hybrid aggregation. ACM MM'23, . 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[13] Xinqi Fan, Xueli Chen, Mingjie Jiang, Ali Raza Shahid, and Hong Yan. SelfME: Self-Supervised Motion Learning for Micro-Expression Recognition. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.199, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13834-13843, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[14] Viswanatha Reddy Gajjala, Sai Prasanna Teja Reddy, Snehasis Mukherjee, and Shiv Ram Dubey. MERANet: facial micro-expression recognition using 3D residual attention network. In Proceedings of the Twelfth Indian Conference on Computer Vision, Graphics and Image Processing, New York, NY, USA, 2021. Association for Computing Machinery. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[15] Yee Siang Gan, Sze-Teng Liong, Wei-Chuen Yau, Yen-Chang Huang, and Lit-Ken Tan. Off-ApexNet on microexpression recognition system. Signal Processing: Image Communication, 74:129-139, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[16] Haihong Hao, Shuo Wang, Huixia Ben, Yanbin Hao, Yansong Wang, and Weiwei Wang. Hierarchical Space-Time Attention for Micro-Expression recognition. arXiv preprint arXiv:2405.03202, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.402 + ], + "angle": 0, + "content": "[17] Kensho Hara, Hirokatsu Kataoka, and Yutaka Satoh. Can Spatiotemporal 3D CNNs Retrace the History of 2D CNNs and Imagenet? In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 6546-6555, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.404, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[18] Xiaohua Huang, Su-Jing Wang, Guoying Zhao, and Matti Piteikainen. Facial Micro-Expression Recognition Using Spatiotemporal Local Binary Pattern with Integral Projection. In 2015 IEEE International Conference on Computer Vision Workshop (ICCVW), pages 1–9. IEEE, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.543 + ], + "angle": 0, + "content": "[19] T. Kanade, J.F. Cohn, and Yingli Tian. Comprehensive database for facial expression analysis. In Proceedings Fourth IEEE International Conference on Automatic Face and Gesture Recognition (Cat. No. PR00580), pages 46-53, 2000. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.546, + 0.905, + 0.586 + ], + "angle": 0, + "content": "[20] Davis E King. Dlib-ml: A Machine Learning Toolkit. The Journal of Machine Learning Research, 10:1755-1758, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.589, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[21] Thomas N Kipf and Max Welling. Semi-Supervised Classification with Graph Convolutional Networks. arXiv preprint arXiv:1609.02907, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.905, + 0.686 + ], + "angle": 0, + "content": "[22] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[23] Sang Gyu Kwak and Jong Hae Kim. Central limit theorem: the cornerstone of modern statistics. Korean journal of anesthesiaology, 70(2):144, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.733, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[24] Ling Lei, Tong Chen, Shigang Li, and Jianfeng Li. Microexpression Recognition Based on Facial Graph Representation Learning and Facial Action Unit Fusion. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1571-1580. IEEE, 2021. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[25] Hanting Li, Mingzhe Sui, Zhaoqing Zhu, and Feng Zhao. Mmnet: Muscle motion-guided network for microexpression recognition. arXiv preprint arXiv:2201.05297, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Jingting Li, Moi Hoon Yap, Wen-Huang Cheng, John See, Xiaopeng Hong, Xiaobai Li, and Su-Jing Wang. FME'21:" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14461" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "1st Workshop on Facial Micro-Expression: Advanced Techniques for Facial Expressions Generation and Spotting. In Proceedings of the 29th ACM International Conference on Multimedia, pages 5700-5701. ACM, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.244 + ], + "angle": 0, + "content": "[27] Jingting Li, Zizhao Dong, Shaoyuan Lu, Su-Jing Wang, Wen-Jing Yan, Yinhuan Ma, Ye Liu, Changbing Huang, and Xiaolan Fu. CAS(ME)\\(^3\\): A Third Generation Facial Spontaneous Micro-Expression Database With Depth Information and High Ecological Validity. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):2782-2800, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.246, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[28] Tian Li, Anit Kumar Sahu, Maziar Sanjabi, Manzil Zaheer, Ameet Talwalkar, and Virginia Smith. On the Convergence of Federated Optimization in Heterogeneous Networks. arXiv preprint arXiv:1812.06127, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.482, + 0.344 + ], + "angle": 0, + "content": "[29] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.413 + ], + "angle": 0, + "content": "[30] Xiaobai Li, Tomas Pfister, Xiaohua Huang, Guoying Zhao, and Matti Pietikainen. A Spontaneous Micro-expression Database: Inducement, collection and baseline. In 2013 10th IEEE International Conference and Workshops on Automatic Face and Gesture Recognition (FG), pages 1-6, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[31] Xiaobai Li, Xiaopeng Hong, Antti Moilanen, Xiaohua Huang, Tomas Pfister, Guoying Zhao, and Matti Pietikainen. Towards Reading Hidden Emotions: A Comparative Study of Spontaneous Micro-Expression Spotting and Recognition Methods. IEEE Transactions on Affective Computing, 9(4): 563–577, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.482, + 0.567 + ], + "angle": 0, + "content": "[32] Xiaobai Li, Shiyang Cheng, Yante Li, Muzammil Behzad, Jie Shen, Stefanos Zafeiriou, Maja Pantic, and Guoying Zhao. 4DME: A Spontaneous 4D Micro-Expression Dataset With Multimodalities. IEEE Transactions on Affective Computing, 14(4):3031-3047, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.482, + 0.609 + ], + "angle": 0, + "content": "[33] Yante Li, Xiaohua Huang, and Guoying Zhao. Microexpression action unit detection with spatial and channel attention. Neurocomputing, 436:221-231, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.609, + 0.482, + 0.692 + ], + "angle": 0, + "content": "[34] Sze-Teng Liong, John See, Raphael C.-W. Phan, Anh Cat Le Ngo, Yee-Hui Oh, and KokSheik Wong. Subtle Expression Recognition Using Optical Strain Weighted Features. In Computer Vision - ACCV 2014 Workshops, pages 644-657. Springer International Publishing, 2015. Series Title: Lecture Notes in Computer Science. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.748 + ], + "angle": 0, + "content": "[35] Sze-Teng Liong, John See, KokSheik Wong, and Raphael C.-W. Phan. Less is More: Micro-expression Recognition from Video using Apex Frame. Signal Processing: Image Communication, 62:82-92, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[36] Sze-Teng Liong, Y. S. Gan, John See, Huai-Qian Khor, and Yen-Chang Huang. Shallow Triple Stream Three-dimensional CNN (STSTNet) for Micro-expression Recognition. In 2019 14th IEEE International Conference on Automatic Face & Gesture Recognition (FG 2019), pages 1-5, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[37] Yong-Jin Liu, Jin-Kai Zhang, Wen-Jing Yan, Su-Jing Wang, Guoying Zhao, and Xiaolan Fu. A Main Directional Mean Optical Flow Feature for Spontaneous Micro-Expression Recognition. IEEE Transactions on Affective Computing, 7 (4):299–310, 2015. 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.162 + ], + "angle": 0, + "content": "[38] Yong-Jin Liu, Jin-Kai Zhang, Wen-Jing Yan, Su-Jing Wang, Guoying Zhao, and Xiaolan Fu. A main directional mean optical flow feature for spontaneous micro-expression recognition. IEEE Transactions on Affective Computing, 7(4):299-310, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.906, + 0.234 + ], + "angle": 0, + "content": "[39] Ling Lo, Hong-Xia Xie, Hong-Han Shuai, and Wen-Huang Cheng. MER-GCN: Micro-Expression Recognition Based on Relation Modeling with Graph Convolutional Networks. In 2020 IEEE conference on multimedia information processing and retrieval (MIPR), pages 79–84. IEEE, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.906, + 0.29 + ], + "angle": 0, + "content": "[40] S Mohammad Mavadati, Mohammad H Mahoor, Kevin Bartlett, Philip Trinh, and Jeffrey F Cohn. Disfa: A spontaneous facial action intensity database. IEEE Transactions on Affective Computing, 4(2):151-160, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.906, + 0.375 + ], + "angle": 0, + "content": "[41] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-Efficient Learning of Deep Networks from Decentralized Data. In Proceedings of the 20th International Conference on Artificial Intelligence and Statistics, pages 1273-1282. PMLR, 2017. ISSN: 2640-3498. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.906, + 0.447 + ], + "angle": 0, + "content": "[42] Xuan-Bac Nguyen, Chi Nhan Duong, Xin Li, Susan Gauch, Han-Seok Seo, and Khoa Luu. Micron-BERT: BERT-Based Facial Micro-Expression Recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1482–1492, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.906, + 0.503 + ], + "angle": 0, + "content": "[43] T. Ojala, M. Pietikainen, and T. Maenpaa. Multiresolution gray-scale and rotation invariant texture classification with local binary patterns. IEEE Transactions on Pattern Analysis and Machine Intelligence, 24(7):971-987, 2002. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.906, + 0.574 + ], + "angle": 0, + "content": "[44] Fangbing Qu, Su-Jing Wang, Wen-Jing Yan, He Li, Shuhang Wu, and Xiaolan Fu. CAS(ME)\\(^2\\): A Database for Spontaneous Macro-Expression and Micro-Expression Spotting and Recognition. IEEE Transactions on Affective Computing, 9(4):424-436, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.906, + 0.646 + ], + "angle": 0, + "content": "[45] John See, Moi Hoon Yap, Jingting Li, Xiaopeng Hong, and Su-Jing Wang. MEGC 2019 – The Second Facial Micro-Expressions Grand Challenge. In 2019 14th IEEE International Conference on Automatic Face & Gesture Recognition (FG 2019), pages 1–5, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.906, + 0.729 + ], + "angle": 0, + "content": "[46] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE Conference on Computer Cision and Pattern Recognition, pages 1-9, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.73, + 0.906, + 0.786 + ], + "angle": 0, + "content": "[47] Petar Velicković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph Attention Networks. arXiv preprint arXiv:1710.10903, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.906, + 0.859 + ], + "angle": 0, + "content": "[48] Lei Wang, Pinyi Huang, Wangyang Cai, and Xiyao Liu. Micro-expression recognition by fusing action unit detection and spatio-temporal features. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 5595-5599. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.906, + 0.902 + ], + "angle": 0, + "content": "[49] Su-Jing Wang, Wen-Jing Yan, Xiaobai Li, Guoying Zhao, Chun-Guang Zhou, Xiaolan Fu, Minghao Yang, and Jianhua Tao. Micro-expression recognition using color spaces." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14462" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "IEEE Transactions on Image Processing, 24(12):6034-6047, 2015. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.164 + ], + "angle": 0, + "content": "[50] Zhifeng Wang, Kaihao Zhang, Wenhan Luo, and Ramesh Sankaranarayana. Htnet for micro-expression recognition. Neurocomputing, 602:128196, 2024. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.483, + 0.248 + ], + "angle": 0, + "content": "[51] Wen-Jing Yan, Qi Wu, Yong-Jin Liu, Su-Jing Wang, and Xiaolan Fu. CASME database: A dataset of spontaneous micro-expressions collected from neutralized faces. In 2013 10th IEEE International Conference and Workshops on Automatic Face and Gesture Recognition (FG), pages 1-7. IEEE, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.483, + 0.319 + ], + "angle": 0, + "content": "[52] Zhaoqiang Xia, Wei Peng, Huai-Qian Khor, Xiaoyi Feng, and Guoying Zhao. Revealing the invisible with model and data shrinking for composite-database micro-expression recognition. IEEE Transactions on Image Processing, 29: 8590-8605, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.483, + 0.39 + ], + "angle": 0, + "content": "[53] Hong-Xia Xie, Ling Lo, Hong-Han Shuai, and Wen-Huang Cheng. AU-assisted graph attention convolutional network for micro-expression recognition. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2871-2880. ACM, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.483, + 0.459 + ], + "angle": 0, + "content": "[54] Wen-Jing Yan, Xiaobai Li, Su-Jing Wang, Guoying Zhao, Yong-Jin Liu, Yu-Hsin Chen, and Xiaolan Fu. CASME II: An improved spontaneous micro-expression database and the baseline evaluation. PLoS ONE, 9(1):e86041, 2014-01-27. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.483, + 0.518 + ], + "angle": 0, + "content": "[55] Xiyuan Yang, Wenke Huang, and Mang Ye. Fedas: Bridging inconsistency in personalized federated learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11986-11995, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.52, + 0.483, + 0.588 + ], + "angle": 0, + "content": "[56] Moi Hoon Yap, John See, Xiaopeng Hong, and Su-Jing Wang. Facial Micro-Expressions Grand Challenge 2018 Summary. In 2018 13th IEEE International Conference on Automatic Face & Gesture Recognition (FG 2018), pages 675–678, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.591, + 0.483, + 0.673 + ], + "angle": 0, + "content": "[57] Zhijun Zhai, Jianhui Zhao, Chengjiang Long, Wenju Xu, Shuangjiang He, and Huijuan Zhao. Feature representation learning with adaptive displacement generation and transformer fusion for micro-expression recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22086-22095, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.483, + 0.743 + ], + "angle": 0, + "content": "[58] Sirui Zhao, Hanqing Tao, Yangsong Zhang, Tong Xu, Kun Zhang, Zhongkai Hao, and Enhong Chen. A two-stage 3D CNN based learning method for spontaneous microexpression recognition. Neurocomputing, 448:276-289, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.483, + 0.814 + ], + "angle": 0, + "content": "[59] Sirui Zhao, Huaying Tang, Xinglong Mao, Shifeng Liu, Yiming Zhang, Hao Wang, Tong Xu, and Enhong Chen. DFME: A New Benchmark for Dynamic Facial MicroExpression Recognition. IEEE Transactions on Affective Computing, 15(3):1371-1386, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.483, + 0.873 + ], + "angle": 0, + "content": "[60] Ling Zhou, Qirong Mao, Xiaohua Huang, Feifei Zhang, and Zhihong Zhang. Feature Refinement: An expression-specific feature learning and fusion method for micro-expression recognition. Pattern Recognition, 122:108275, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.483, + 0.903 + ], + "angle": 0, + "content": "[61] Xu Zhou, Jie Li, Gongjin Lan, Rongrong Ni, Angelo Cangelosi, Jiaxin Wang, and Xiaofeng Liu. Efficient lower" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.547, + 0.092, + 0.905, + 0.135 + ], + "angle": 0, + "content": "layers parameter decoupling personalized federated learning method of facial expression recognition for home care robots. Information Fusion, 106:102261, 2024. 6" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14463" + } + ] +] \ No newline at end of file diff --git a/2025/FED-PsyAU_ Privacy-Preserving Micro-Expression Recognition via Psychological AU Coordination and Dynamic Facial Motion Modeling/e27ad4a5-3106-4216-bcea-a95709cd5f65_origin.pdf b/2025/FED-PsyAU_ Privacy-Preserving Micro-Expression Recognition via Psychological AU Coordination and Dynamic Facial Motion Modeling/e27ad4a5-3106-4216-bcea-a95709cd5f65_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5ae6bc64a1f511fb72234f20481bd54670c8aac5 --- /dev/null +++ b/2025/FED-PsyAU_ Privacy-Preserving Micro-Expression Recognition via Psychological AU Coordination and Dynamic Facial Motion Modeling/e27ad4a5-3106-4216-bcea-a95709cd5f65_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d538f27be030ef506a73a0045904e6f321eb940d7b1b4d6658e427b1f7a94666 +size 3450668 diff --git a/2025/FED-PsyAU_ Privacy-Preserving Micro-Expression Recognition via Psychological AU Coordination and Dynamic Facial Motion Modeling/full.md b/2025/FED-PsyAU_ Privacy-Preserving Micro-Expression Recognition via Psychological AU Coordination and Dynamic Facial Motion Modeling/full.md new file mode 100644 index 0000000000000000000000000000000000000000..bf785541c0252bfee01947b045531e030007dd7b --- /dev/null +++ b/2025/FED-PsyAU_ Privacy-Preserving Micro-Expression Recognition via Psychological AU Coordination and Dynamic Facial Motion Modeling/full.md @@ -0,0 +1,365 @@ +# FED-PsyAU: Privacy-Preserving Micro-Expression Recognition via Psychological AU Coordination and Dynamic Facial Motion Modeling + +Jingting Li $^{1,2}$ , Yu Qian $^{1,3}$ , Lin Zhao $^{1,2}$ , Su-Jing Wang $^{*1,2}$ + +$^{1}$ State Key Laboratory of Cognitive Science and Mental Health, Institute of Psychology, Chinese Academy of Sciences, Beijing, 100101, China + +$^{2}$ Department of Psychology, University of the Chinese Academy of Sciences, Beijing, 100049, China + +$^{3}$ School of Computer, Jiangsu University of Science and Technology, Zhenjiang, 212100, China + +$^{\dagger}$ Equal contribution, \*Corresponding author (wangsujing@psych.ac.cn) + +# Abstract + +Micro-expressions (MEs) are brief, low-intensity, often localized facial expressions. They could reveal genuine emotions individuals may attempt to conceal, valuable in contexts like criminal interrogation and psychological counseling. However, ME recognition (MER) faces challenges, such as small sample sizes and subtle features, which hinder efficient modeling. Additionally, real-world applications encounter ME data privacy issues, leaving the task of enhancing recognition across settings under privacy constraints largely unexplored. To address these issues, we propose a FED-PsyAU research framework. We begin with a psychological study on the coordination of upper and lower facial action units (AUs) to provide structured prior knowledge of facial muscle dynamics. We then develop a DPK-GAT network that combines these psychological priors with statistical AU patterns, enabling hierarchical learning of facial motion features from regional to global levels, effectively enhancing MER performance. Additionally, our federated learning (FL) framework advances MER capabilities across multiple clients without data sharing, preserving privacy and alleviating the limited-sample issue for each client. Extensive experiments on commonly-used ME databases demonstrate the effectiveness of our approach. Our implementation is publicly available at https://github.com/MELABIPCAS/FED-PsyAU.git. + +# 1. Introduction + +Micro-expressions (MEs) [9], resulting from conflicts between voluntary and involuntary expressions, reveal true emotions, making them valuable in non-contact emotion detection applications like criminal interrogation and psychological counseling. MEs are characterized by their brief duration, low intensity, localized occurrence, and occasional + +asymmetry [10]. These traits not only restrict large-scale data collection and annotation but also make it challenging for deep learning networks to effectively model their motion characteristics. Some methods focus on regions of interest (ROIs) [4, 21, 31, 49], while others utilize attention mechanisms [3, 14, 33] and transformer-based architectures [24, 50, 57] to capture salient features. Meanwhile, MEs are often complex manifestations involving the combination of multiple AUs. In contrast, individual action units (AUs) correspond to specific, anatomically-based facial muscle movements. Focusing on AUs allows us to model more fundamental and consistent facial appearance changes. Certain approaches use AU labels to construct adjacency matrices for graph convolutional networks (GCNs) [24, 39]. Yet, to the best of our knowledge, these methods often apply AU annotations directly without fully exploiting the anatomical structure and coordinated AU dynamics, limiting the network's understanding on ME. + +To address these challenges, we leverage the theory that distinct neural pathways control upper and lower facial movements, resulting in two different motion patterns. Based on this, we conduct a psychological study to investigate upper/lower facial AU coordination, providing crucial prior knowledge on structural motion relationships underpinning MEs. Using both psychological priors and data-driven AU patterns, we design a Graph Attention Network (GAT) [47] based on dynamic prior knowledge, applied separately to the upper and lower regions. A final global GAT captures the topological relationships of muscle movements across the face. The resulting feature, combined with full-face optical flow (OF), is input into a subsequent dual-stream network, effectively integrating structural and dynamic information to enhance ME recognition (MER). + +Besides, as mentioned earlier, although MEs have valuable applications in many privacy-sensitive contexts, the sensitivity of biometric data makes it impractical to aggre + +gate ME data across clients for training. Meanwhile, the sample size in a single client is often insufficient to train a robust model. Yet, the challenge of enhancing MER performance across diverse clients while preserving privacy remains largely unexplored. To address this, we adopt a federated learning (FL) framework, improving performance by merging and updating model parameters across clients without exchanging local data. This approach preserves data privacy and supports the MER practical deployment. + +Overall, our work leverages psychological and data-driven priors to enhance the network's learning of structured ME motion patterns, particularly in limited-sample conditions. Additionally, we introduce FL to advance MER in real-world scenarios. The contributions of this paper are: + +- We conduct a psychological experiment to derive AU coordination between the upper and lower facial regions, providing structural priors for ME and facial expression analysis in computer vision. +- We design a local-to-global GAT network that integrates psychological priors and data-driven AU patterns in the upper and lower facial regions, improving the network's ability to learn both local and global structural facial dynamics. This feature, combined with global OF feature in a dual-stream network, enhances MER performance. +- We employ FL in MER to improve data privacy and overcome small sample size challenges by aggregating features from distributed data-protected sources. + +# 2. Related Works + +# 2.1. Micro-expression Recognition + +In recent years, MER has gained increasing attention from computer vision researchers. Traditional machine learning methods primarily rely on handcrafted feature extraction. For instance, Local Binary Pattern (LBP) [43], effectively utilizes local texture features, has inspired various improvements [18, 49]. Furthermore, OF [26] has been explored for capturing object motion. Liu et al. [38] developed the Main Directional Mean OF (MDMO) method by aligning faces in the OF domain, while Liong et al. [34] added optical strain features for MER. However, traditional approaches depend heavily on manually crafted features, limiting both accuracy and generalizability. Meanwhile, deep learning methods, with robust feature extraction capabilities, can automatically learn multi-level representations to capture subtle ME changes. The availability of several spontaneous ME databases, such as CASME II [51], SAMM [6], CAS(ME) $^{3}$ [27], DFME [59] and others [1, 30, 32, 44, 54], has accelerated deep learning advancements in MER. Inspired by Liong et al.'s work [35] on sequence redundancy, many studies [4, 13, 50] now use keyframes (onset and apex) as input, reducing model parameters while achieving excellent results. + +Despite advancements in preprocessing and feature extraction, deep learning alone has yet to achieve high MER accuracy. As MEs stem from physiological muscle activity, incorporating prior knowledge offers meaningful context. FACS [10], which defines expressions through AUs, provides a basic theoretical framework for MER. Recent studies explore AU-ME relationships to enhance MER performance. Lei et al. [24] applied GCN [21] to model AU co-occurrences, especially around the eyebrow and mouth regions. Xie et al. [53] introduced an AU-assisted GCN, with modules for AU feature extraction and ME image generation. Wang et al. [48] used transformers to learn dynamic AU adjacency, boosting model generalization. These networks typically rely on AU labels from databases, without fully integrating structured facial muscle movements, AU interrelations, or their links to expressions. In this work, we refine structured ME feature extraction using a local-to-global GAT network, which integrates priors on facial AU coordination from psychological experiments with data-driven movement patterns. + +# 2.2. Federated Learning + +To assess MER across databases, ME Grand Challenge (MEGC) 2018 [56] introduced the Holdout-database Evaluation and Composite Database Evaluation tasks, utilizing the CASME II and SAMM databases. MEGC 2019 [45] expanded the CDE task to include the SMIC databases. While research on MEs predominantly utilizes these public datasets, existing methodologies do not address the privacy concerns that arise in practical applications. + +FL [41] could addresses ME data scarcity and privacy concerns through collaborative training. It includes two main phases: local training and global aggregation. Many existing FL works for expressions, while insightful, are often not open-source or use highly customized frameworks (e.g., non-random client data allocation [12], specialized parameter aggregation [11]) making standard methods more suitable for implementation and comparisons in our study. For instance, McMahan et al. [41] proposed FedAvg, a basic aggregation method that uses weighted averaging based on each client's data volume. Li et al. [28] introduced FedProx by adding a regularization term to the client loss function, aligning updates more closely with the global model. However, the one-global model-fits-all approach ignores client data diversity. To solve this, our approach extends FedProx by distributing a personalized global model to each client. + +# 3. Proposed Psychological Study on AU Incoordination + +This study employed a behavioral experiment to investigate individuals' cognitive responses to AU incoordination in the upper and lower facial regions. + +![](images/d0be976c39a1a2b1350002c8024a300f5603459bf08e9cad7c5424168a69b569.jpg) +(a) AU $4 + 9$ + +![](images/6a6758a36e511f40688f624788965fb5f708c901579da499564d7275e17f1144.jpg) +(b) AU 6+12 + +![](images/75bb2e5693552c09e2c1a618dc3ae130c5010e3798a28cb5539630fc5fb61851.jpg) +(c) AU 4+12 +Figure 1. Examples of AU composite images. (a) and (b) are coordinated composites. (c) and (d) are uncoordinated composites. + +![](images/c488cf12372e861cf51dcda4ad2396759b6b73ada71b35e9bbaf1e1974f014a9.jpg) +(d) AU 6+15 + +The study included 30 college participants (20 female, Mean= 22.22 years), a sample size considered adequate for statistical reliability under the Central Limit Theorem [23]. All procedures followed the Declaration of Helsinki and were approved by the institutional review board of the Institute of Psychology, Chinese Academy of Sciences. + +This behavioral experiment employed a within-subjects design. The independent variable was the type of AU composite image (coordinated vs. uncoordinated AU composites), as shown in Fig. 1. The face images were segmented anatomically. The dividing line was consistently present in all stimuli, serving as a control variable and would not influence the experiment's outcomes. To control for individual facial differences in emotion recognition, the original images used for composites were sourced from the same person [10]. Seventy-two composite images were generated from upper and lower facial AUs across six basic emotions, with each emotion represented by six coordinated and six uncoordinated expressions. + +The dependent variables are accuracy rates and scores. Specifically, participants made two judgments for each composite image: whether it was coordinated or uncoordinated, and the degree of coordination or incoordination. + +Statistical analysis from this study showed that participants were significantly more accurate in recognizing coordinated AU composite images than uncoordinated ones (Please see Supplementary Materials (Suppl.) for more experimental details and statistical results). The study further highlighted specific AU combinations, such as AU4 and AU12, that participants quickly identified as uncoordinated, whereas combinations like AU6 and AU12 were rapidly recognized as coordinated. This suggests that these uncoordinated AU pairings may inherently convey emotional incoordination. Besides, the coordination patterns between specific AUs reveal systematic facial action structures. These patterns help the network distinguish both coordinated and uncoordinated facial movements, thus enhancing its ability to identify ME features more effectively. + +# 4. Our Proposed FED-PsyAU Framework + +We propose the FED-PsyAU framework (Fig. 2), which embeds a psychology-driven MER model within a federated aggregation module. The model features a hierarchi + +cal structure that learns from local ROIs to AU groups and finally to global facial regions, capturing dynamic topology while minimizing redundancy. It leverages a multiscale InceptionNet to extract and fuse global OF with structural facial features. The federated module enables privacy-preserving collaborative training to boost MER performance on each client. + +# 4.1. Localized ROI Modeling + +Since MEs are manifested by muscle movements in key facial regions (e.g., eyes, mouth, nose) [10], we refine the standard 68 landmarks from Dlib [20] to precisely capture these changes. As shown in Fig. 3, we discard 10 less informative outer contour landmarks and add 7 new midpoints to better track movements in the cheek and eyebrow regions, resulting in a final set of 65 keypoints. + +Then, onset-apex OF inputs $(128 \times 128$ pixels) are cropped into 65 ROIs, each measuring $5 \times 5$ pixels, centered on selected facial landmarks $\mathbf{p}_k = (x_k, y_k)$ , where $k \in 1, 2, \ldots, 65$ , denotes the landmark index. To enhance robustness to apex annotation, we also compute OF using a small window of frames around the apex. The ROI features input to the Local ROI Feature Extractor (LFE), denoted as $\Theta_k$ , contain the horizontal and vertical components of OF and optical strain. The LFE module consists of three convolutional layers, each followed by batch normalization and ReLU activation. All kernel sizes were $3 \times 3$ . The output channels in the first two layers increase from 32 to 64, while the final layer outputs 3 channels. This design helps the LFE capture complex local patterns, with the output features referred to as $\Phi_k = \mathrm{LFE}(\Theta_k)$ . + +Inspired by Vision Transformer [8], we use a Spatial Structure Encoder (SSE), i.e., Transformer Encoder with three attention heads to capture diverse dependencies among ROIs, as MEs result from coordinated muscle movements across multiple areas. Precisely, we convert $\Phi_k$ , into a token. To retain the positional information of the facial landmarks, a positional embedding $E_{k}$ is added based on the landmark index to each token, yielding $X_{k} = \Phi_{k} + E_{k}$ . We add the output of the three head attention module to the input feature $X_{k}$ , and then normalize the result using LayerNorm. Finally, after passing through a feed-forward neural network and applying LayerNorm, the final output representation is obtained as $Z_{k}\in \mathbb{R}^{3\times 5\times 5}$ . + +# 4.2. Relationship Modeling among AU Features + +Our proposed Relationship Modeling among AU Features (AFR) module consists of two components: an AU Feature Extractor (AFE) to derive features from AU-specific facial regions, and a GAT-structure to model the topological relationships between these AUs. + +![](images/32a7893b63bbeb533929b4b91016e08a888aebebc0656d7b79a2840f881c5d6b.jpg) +Figure 2. The pipeline of our proposed FED-PsyAU framework, including MER network (black block) and the Federated Aggregation (FA) module (orange block). In MER network, the LRM module consists of the local ROI feature extractor (LFE) and the spatial structure encoder (SSE), which preserves spatial structure information while efficiently extracting ROI features; the AFR module includes the AU feature extractor (AFE) and the dynamic prior knowledge-based graph attention network (DPK-GAT) for effective AU feature extraction and capturing AU dynamic relationships, the DSI module use a dual-stream structure to capture multi-scale facial muscle topology and motion information for MER. Then, the FA module integrates the MER network into the FL framework. + +![](images/6fff2cada31386a5b011df13bd87f2c5f07231243518331d7ca049c484b4cd7c.jpg) +Figure 3. ROI grouping strategy for AU. OF features from each ROI are processed through LFE and SSE to yield high-dimensional representations. ROIs are then grouped by AU movement patterns to create inputs for AFEs, ultimately producing AU node features for the GAT network. + +# 4.2.1. AU Feature Extractor + +MEs involve dynamic, coordinated interactions across multiple regions [10] and are characterized by subtle movements that differ from typical macro-expressions. Therefore, as shown in Fig. 3, when grouping ROIs for the AFE module, the AU selection and regionalization are based on: 1) AU frequency in MEs; 2) Grouping by primary/secondary muscle activation areas [7]. Notably, not all AUs employed in Section 3 are applied to MER. We specifically retain only the 12 AUs exhibiting the highest relevance to ME dynamics, as listed in Tab. 1. Detailed occurrences are provided in the Suppl. + +The first step we construct in AFE is Group Squeeze and + +Table 1. Selected AUs and Corresponding Regions (Primary and Secondary/ Coordinate). $g$ , $N_g$ and $N_{g\_AU}$ represent ROI Group index, corresponding ROI number and AU amount. + +
gAUPrimarySecondary/ CoordinateNgNg-AU
11,2,4EyebrowsEyes233
25,6,7EyesEyebrows, Cheeks, Mouth483
39,10Nose, CheeksMouth, Chin382
412,14,15,17MouthEyes, Eyebrows, Nose524
+ +Excitation (GSE). It computes the attention for each group of AUs (ROI group), which makes the model pay more attention to the critical ROIs. Specifically, for each ROI group input $R_{g} = \{Z_{1}, Z_{2}, \ldots, Z_{N_{g}}\}$ , $g \in \{1, 2, 3, 4\}$ , the GSE module performs global average pooling on each ROI and computes channel importance through a fully connected (FC) layer. The original features are then preserved through residual concatenation: + +$$ +f _ {\mathrm {G S E}} \left(R _ {g}\right) = \left\{W _ {k} ^ {\text {a t t}} \odot Z _ {k} + Z _ {k} \mid Z _ {k} \in R _ {g} \right\} \tag {1} +$$ + +where $W_{k}^{att}$ is the attention weight for each group of ROI, $\odot$ denotes element-wise multiplication along the channel dimension of $Z_{k}$ . Next, $f_{\mathrm{GSE}}(R_g)$ undergoes further feature extraction, reducing from $N_{g}$ composite features to $N_{g\_AU}$ representative AU features, yielding the output feature $F_{g\_AU}$ . Precisely, the process comprises two convolutional layers: a $3 \times 3$ layer that doubles the input dimensions and a $1 \times 1$ layer that restores them to the original size. Both layers are followed by Batch Normalization and ReLU activation for improved stability and representation. + +# 4.2.2. AU Relationship Modeling + +Prior Knowledge: To model AU dynamic relationships, we integrate two forms of prior knowledge: an "intrinsic" prior + +from psychological principles and an "extrinsic" prior from data-driven patterns. The intrinsic prior, derived from our psychological findings on AU coordination (Sec. 3), provides structural guidance by defining the initial GAT adjacency matrix $A$ . The extrinsic prior, based on statistical AU co-occurrence patterns in the data, offers data-driven refinement that theoretical models may overlook. This extrinsic prior forms a prior attention matrix $D$ to guide weight allocation among AU nodes. By fusing this foundational psychological structure with real-world dynamics, our model achieves superior adaptability and generalization (see Suppl. for matrix details). + +Facial Segmentation: As presented in Sec. 3, the upper and lower facial regions, controlled by different neural pathways, exhibit distinct muscle movement patterns. AUs and their associated primary muscles in the upper and lower face can exhibit coordination (potential co-occurrence) or lack of coordination (infrequent simultaneous appearance). Leveraging these intra- and inter-regional interactions, we model AU relationships for the upper face, lower face, and whole face. Consistent with the experimental setup in Sec. 3, we segment the face into upper and lower parts based on the orbicularis oculi muscle. Specifically, following the primary occurrence regions of 12 listed AUs, the AUs in the first two groups from Tab. 1 are classified as upper face AUs, while those in the last two groups are classified as lower face AUs. + +$$ +F _ {A U} ^ {U p p e r} = F _ {1 \_ A U} \cup F _ {2 \_ A U}, F _ {A U} ^ {L o w e r} = F _ {3 \_ A U} \cup F _ {4 \_ A U} (2) +$$ + +We input the AU features from the upper and lower face: $F_{AU}^{Upper}$ , $F_{AU}^{Lower}$ into our proposed Dynamic Prior Knowledge GATs (DPK-GATs) in parallel (Details are presented in the next part of this subsection). These networks use psychology prior knowledge to build AU adjacency relationships and data-driven priors to guide attention learning, producing outputs $h^{Upper}$ and $h^{Lower}$ , which are combined into $h^{Local}$ for the next stage. + +In the whole-face DPK-GAT, we construct the network similarly but set adjacency values within the same region to zero, ignoring intra-regional AU relationships. By learning cross-regional topological structures and statistical patterns, we obtain global AU relationship features, denoted as $h^{Global}$ , as illustrated in Fig. 4. + +DPK-GAT: We model dynamic AU node relationships combing GAT and the prior knowledge. In particular, for the input feature belonging to $\{F_{AU}^{Upper}, F_{AU}^{Lower}, h^{Local}\}$ , a linear mapping projects it to a higher-dimensional space as $f$ . The similarity score $e_{ij}$ between AU nodes $i$ and $j$ , is calculated using the attention weight vector $\vec{a}$ and concatenated node features $f_{i}$ and $f_{j}$ : + +$$ +e _ {i j} = \left\{ \begin{array}{l l} \text {L e a k y R e L U} \left(\tilde {a} ^ {T} [ f _ {i} | | f _ {j} ]\right), & \text {i f} A _ {i j} > 0 \\ - \infty , & \text {i f} A _ {i j} = 0 \end{array} \right. \tag {3} +$$ + +![](images/3934cbe5af16a98d999e3491078e08adc38f6f6edc979040b58195a36422df20.jpg) +Figure 4. DPK-GAT from facial segmentations to global face. First, intra-regional AU node relationships are learned within the upper and lower facial regions. Second, cross-regional relationships are modeled to integrate interactions between these regions. + +Here, only neighboring nodes defined by the psychology-derived adjacency matrix $A$ are considered. + +The final attention $\alpha_{ij}$ is calculated by combining the data-driven prior attention $D_{ij}$ and the self-attention, obtained by the softmax operation applied to $e_{ij}$ : + +$$ +\alpha_ {i j} = (1 - \beta) \cdot \operatorname {s o f t m a x} \left(e _ {i j}\right) + \beta \cdot D _ {i j} \tag {4} +$$ + +where, $\beta$ is the dynamic prior attention weight, updated based on the training rounds. Along with a residual concatenation, the final node representation $h_i$ is aggregated as: + +$$ +h _ {i} = \operatorname {E L U} \left(\sum_ {j \in \mathcal {N} (i)} \alpha_ {i j} \cdot f _ {j}\right) + f _ {i} \tag {5} +$$ + +where $\mathcal{N}(i)$ denotes the neighboring nodes of $i$ . In our implementation, we employ three attention heads and a two-layer structure, enabling the extraction of complex AU dependencies. + +# 4.3. Dual-Stream-InceptionNet + +Muscle movements vary widely across facial regions, making multi-scale information essential for the MER task. InceptionNet's core concept is leveraging parallel convolutional kernels to extract multi-scale features, enhancing the network's ability to capture fine details and overall structure [46]. We build on this by using a dual-stream InceptionNet as the ME feature extractor. One stream processes the AU relationship features $h^{Global}$ , while the other handles global OF features $\Theta$ . This dual-stream design allows the model to capture the topological relationships of facial muscles through AUs and holistic motion via OF. The output tensors from both streams are flattened, concatenated, and fed into a FC layer for ME prediction. + +# 4.4. Federated Aggregation + +Traditional centralized learning consolidates all data on a single server, enhancing model accuracy and generalization but neglecting privacy concerns in MER. FL could address this by training models locally and sharing only parameters, protecting data privacy and retaining each client's unique data characteristics. + +FedProx [29] is a FL approach designed to address data heterogeneity among clients. It incorporates proximal terms during model aggregation to balance global and client-local learning, improving both performance and convergence speed. We extend FedProx by altering the server's processing and distribution strategy during aggregation, which we call Personalized FedProx (P-FedProx). Specifically, P-FedProx initializes each client's model $W_{i}^{t}$ in round $t$ by combining its previous model $\mathring{W}_{i}^{t-1}$ with models from other clients $\mathring{W}_{j}^{t-1}$ . + +$$ +W _ {i} ^ {t} = \theta \cdot \mathring {W} _ {i} ^ {t - 1} + \sum_ {\substack {j = 1 \\ j \neq i}} ^ {n} \omega_ {j} \cdot \mathring {W} _ {j} ^ {t - 1}, \text {where} \sum_ {\substack {j = 1 \\ j \neq i}} ^ {n} \omega_ {j} = 1 - \theta \tag{6} +$$ + +where weight $\omega_{j}$ is proportional to $j$ th client's data size. For $i$ th client, $\theta$ is set to 0.9 (optimal experimental result, see more detail in the Suppl.) + +# 4.5. Loss Function + +AU Prediction Loss: To enhance MER, our network models the topological relationships and statistical patterns of AUs, making AU prediction an auxiliary task. The prediction is utilized in two stages: one enhances the AU feature extraction performance of the AFE module, and the other improves the GAT network's ability to learn AU topological structures. AU recognition is formulated as a multi-label classification problem, with the Binary Cross-Entropy Loss for each AU label defined as: + +$$ +L _ {\mathrm {A}} = \frac {\sum_ {j = 1} ^ {M} \left[ y _ {j} \cdot \log \left(\sigma \left(\hat {y} _ {j}\right)\right) + \left(1 - y _ {j}\right) \cdot \log \left(1 - \sigma \left(\hat {y} _ {j}\right)\right) \right]}{- M} \tag {7} +$$ + +where $M$ is the total number of AUs, $y_{j}$ , $\sigma (\hat{y}_j)$ and $\hat{y}_j$ represent respectively the ground truth label (0 or 1), the logit output and prediction after applying the sigmoid function for the $j$ -th AU. The loss function is consistent for both stages, i.e., $L_{\mathrm{A}}^{\mathrm{AFE}}$ and $L_{\mathrm{A}}^{\mathrm{GAT}}$ . + +MER Loss: The Cross-Entropy Loss function is utilized for emotion classification. Thus, our overall MER loss is composed as follows: + +$$ +L _ {\text {M E R}} = \alpha_ {1} \cdot \left(- \sum_ {i = 1} ^ {N} y _ {i} \cdot \log (\hat {y} _ {i})\right) + \alpha_ {2} \cdot L _ {\mathrm {A}} ^ {\mathrm {A F E}} + \alpha_ {3} \cdot L _ {\mathrm {A}} ^ {\mathrm {G A T}} \tag {8} +$$ + +where, $N$ represents the number of classes, $y_{i}$ and $\hat{y}_i$ are the one-hot encoded ground truth label and the predicted probability for class $i$ , $\alpha_{1}$ , $\alpha_{2}$ and $\alpha_{3}$ are hyperparameters, set to 0.2, 0.8, and 0.8 to prioritize MER loss while balancing AU recognition (See parameter analysis in Suppl.). + +Federated Loss: During federated training, the local client's loss function includes the MER loss $L_{\mathrm{MER}}$ and a proximal term: + +$$ +L _ {\mathrm {F L}} = L _ {\mathrm {M E R}} + \frac {\alpha_ {4}}{2} \cdot \left\| W - W _ {t} \right\| ^ {2} \tag {9} +$$ + +where $W_{t}$ denotes the global model in the server's $t$ -th round of distribution, and $W$ represents the current model of the client. $\alpha_{4}$ is the hyperparameter for the weight of the proximal term. + +# 5. Experiments + +# 5.1. Configuration + +Databases: We use two recently published large scale and challenging ME databases: CAS(ME) $^3$ and DFME. CAS(ME) $^3$ -Part A offers 860 samples from 100 participants. DFME includes three subsets with (participants, samples): A (72,1,118), B (92,969), C (492,5,439). For Apex-first-frame scenario, we compute the OF between the apex and mid-frame towards the offset. + +Pre-training: To enhance the model's ability to perceive facial motion patterns, we first pre-trained the AU module (LRM+AFR) on the DISFA [40] and CK [19] datasets. The effect of this pre-training is detailed in the Suppl. + +MER Experiment Settings: The evaluation follows standard protocols on two datasets. On $\mathrm{CAS(ME)^3}$ , we perform 3-class classification (positive, negative, surprised) [45] with Leave-One-Subject-Out (LOSO) validation. On DFME, we conduct 7-class classification (Happiness, Surprise, Disgust, Sadness, Anger, Fear, and Contempt) using scale-adapted 10-fold cross-validation [59]. To address class imbalance, performance is measured by the UF1 and UAR. Experiments were implemented in PyTorch 1.13.0 and run on five NVIDIA GeForce RTX 4090 GPUs. + +Federated Experiment Settings: The DFME and $\mathrm{CAS(ME)^3}$ datasets are randomly split into 5 and 2 local clients, respectively, with each client assigned an equal number of subjects. However, variations in microexpression (ME) counts per subject result in differing ME data distributions and quantities across clients, mirroring real-world data heterogeneity (See detailed data distributions in the Suppl.). Traditional LOSO and k-fold cross-validation can leak training samples in a federated setting, so we employ a random partition strategy with $70\%$ of samples for training and $30\%$ for testing, averaging results over 10 tests to minimize randomness. Furthermore, we assume an ideal FL scenario with no client dropouts and equal local updates per client. In FedAvg [41] and FedProx [28], the server assigns an identical global model to all clients each round; differently, FedProx adds a proximal term to reduce model bias in local training. Our P-FedProx enhances this by assigning tailored initial models, with $90\%$ ( $\theta$ ) weight on each client's own model and $10\%$ on others'. Additionally, we compare against recent personalized FL methods: FedRep [5] uses shared representations with personalized heads, ELLP [61] employs parameter decoupling with Fisher Information weighting, and FedAS [55] applies parameter alignment and client synchronization. Unlike these approaches requiring architectural changes or complex procedures, P-FedProx achieves personalization through simple weighted initialization. + +Table 2. SOTA methods comparison on DFME and CAS(ME) ${}^{3}$ + +
CategoryMethodYearDFMECAS(ME)3
UF1UARUF1UAR
3D-CNNI3D [2]20170.29230.3058--
R3D [17]20180.21640.2313--
Hand-craftedLBP-TOP [49]20140.23360.2653--
MDMO [37]20150.24890.2939--
Deep LearningOFF-ApexNet [15]20190.23860.2806--
AlexNet [22]2012--0.25700.2634
STSTNet [36]20190.27140.31080.37950.3792
RCN-A [52]20200.27510.31230.39280.3893
MERSiam [58]20210.31840.35320.31840.3532
FGRL [24]20210.07360.14290.33330.2636
FR [24, 60]20220.35590.38140.34930.3413
MMNet [25]20220.26490.27760.37060.3646
BDCNN [4]20220.29750.33140.50500.5164
Micro-BERT [42]2023--0.56040.6125
HTNet [50]20240.32430.33680.57670.5415
HSTA [16]2024--0.5930.618
Ours-0.38530.39780.62210.6226
+ +# 5.2. Comparison to State-of-the-art Methods + +Tab. 2 list the SOTA comparisons on DFME and CAS(ME) $^3$ . The confusion matrices are shown in the Suppl. + +Regarding DFME, classifying ME samples into seven categories poses significant challenges, resulting in relatively low UF1 and UAR scores across all methods. However, our proposed method outperforms all others in every metric. Specifically, Our method outperforms I3D by $9.30\%$ in UF1 and $9.20\%$ in UAR, and surpasses LBP-TOP by $15.17\%$ in UF1 and $13.25\%$ in UAR, highlighting its effectiveness in extracting critical and discriminative features from ME clips. Additionally, compared to recent deep learning methods like FR, our approach yields $2.94\%$ and $1.64\%$ higher UF1 and UAR, confirming the benefits of incorporating facial topology and structured muscle motion for enhanced performance. + +Regarding CAS(ME) $^3$ , the primary challenge lies in its sample complexity. Compared to baseline methods (STST-Net, RCN-A, FR), our proposed method achieves over approximately $20\%$ higher UF1 and UAR scores, marking a significant advancement. Our method surpasses three recent SOTA approaches in MER, including HTNet, which employs a hierarchical Transformer for local feature learning via self-attention. Specifically, our approach outperforms HTNet by $4.54\%$ in UF1 and $8.11\%$ in UAR, leveraging psychological insights and statistical patterns to capture deeper facial muscle dynamics. Additionally, it exceeds Micro-BERT and HSTA by $6.17\%$ and $2.91\%$ in UF1, and $1.01\%$ and $0.46\%$ in UAR, respectively. By mining facial movement patterns to better understand topology, our model achieves improved metrics, indicating enhanced accuracy in classifying all emotion categories. + +# 5.3. Ablation Study + +Tab. 3 listed the ablation study results on DFME. + +Table 3. Ablation study. L, G, S and D represent Local (upper and lower face), Global, Single stream and Dual Stream, respectively. + +
SettingLRMAFRInceptionNetUF1↑UAR↑
LFESSELGSD
I✓0.32490.3404
II✓✓0.33620.3517
III✓✓✓0.34800.3624
IV✓✓✓✓0.36290.3723
V✓✓✓✓0.36040.3658
VI✓✓✓✓✓0.38530.3978
SettingWithout AU GroupWith AU GroupUF1↑UAR↑
VII✓0.36830.3793
VIII✓0.38530.3978
SettingPsychological priorData-driven PriorUF1↑UAR↑
IX0.36150.3701
X✓0.37230.3832
XI✓0.37050.3767
XII✓✓0.38530.3978
+ +Impact of AU Feature: Setting I, which uses only full face OF as input, serves as our baseline for evaluating the role of AU features in MER. In Setting II, integrating AU features into the Dual Stream InceptionNet leads to a $1.13\%$ improvement in both UF1 and UAR. This improvement demonstrates that incorporating AU features significantly enhances the model's ability to recognize MEs, and underscores the value of fusing multiple feature types to strengthen model performance. + +Impact of ROI Relationship Modeling: The results in Setting III show improvements of $1.18\%$ in UF1 and $1.07\%$ in UAR compared to Setting II, indicating that the model effectively captures complex dependencies between ROIs and extracts more discriminative feature representations. + +Impact of AU grouping: We compared AU feature extraction from the global face with our AU grouping strategy (Setting VII vs. VIII), the performance improvement (UF1: $1.70\%$ , UAR: $1.85\%$ ), confirming that AU grouping enables the network to learn more representative AU features. + +Impact of Modeling the Topological Structure of ME Muscle Movements: Compared to Setting III, Setting IV, including local muscle movement topology modeling, improves UF1 and UAR by $1.49\%$ and $0.99\%$ , respectively. Setting V, which adds global muscle movement topology modeling, shows gains of $1.24\%$ in UF1 and $0.34\%$ in UAR over Setting III. Combining both local and global topology modeling yields the best performance on DFME. These results demonstrate that incorporating topological modeling of muscle movements enhances MER. + +Impact of Prior Knowledge: Tab. 3 presents the ablation experiments to evaluate the impact of using prior knowledge for constructing facial muscle topology. In particular, compared to Setting IX, Setting X, which uses only psychological knowledge, improves UF1 and UAR by $1.08\%$ and + +![](images/7d0359ab475a41676321769bfdef60e3582e90475e1d72ee6269ed907c5d0965.jpg) + +![](images/a86bcd8e0a2ef30abe66e39830767a2927e9ac24819ae8d569c5be5d4acfe90c.jpg) + +![](images/ed619347e1503f8d05331087100313350ff40d9aa809b553fab8d6c994e8edc0.jpg) +(a) UF1 comparison across different clients. + +![](images/4dca360b1ef279a4b74913e2b235d6d2581047bd809eaacd2dfa43f486177ef5.jpg) + +![](images/4d64020bb307952c01cb4203c06cbd5b28c204536d2809b574d8cc0743ca013d.jpg) + +![](images/7ce7a4851e79c5bfbe1887a3d439063c725adae50997589753a075c3fe04ae49.jpg) + +![](images/d60e9d362a4ca3a84d4fb9c4de393ad2b5ae925c765090503f0a69e38f92bcc4.jpg) + +![](images/ca2225a778ac85bbd9856be72ef91346a4870f018b736395c5f8e1ce97c1f509.jpg) +Figure 5. Compare the performance of different FL frameworks on different clients: UF1 and convergence. See UAR comparison in Suppl. + +![](images/9758b0502f57eb674eb29a4c4cbace84da198af57d5284b0fb0e007d7bf93c9b.jpg) + +![](images/d6e5db3f15daf1da9eeb5b183ec476c91b6fe6759cc2c50eeadc434b1fbd0e13.jpg) +(b) Convergence comparison across different clients. + +![](images/164efcb98172282f2bc845b5954f618bf38f57aef2f9040df21f9014b03d4a75.jpg) + +![](images/0278f3eb28b06d6cbecc906342c457ab2c577fc37247aa6ffcea9c6160bebf87.jpg) + +![](images/56f21f9e921cc95ebcf6812103c58b7add37417461d1b73fabeea964d84ffd19.jpg) + +![](images/5d110c145f2ae6a3be10e74af597573fbc31866014983bc5e3e2f04bef4c399e.jpg) + +$1.31\%$ , respectively. Setting XI, incorporating only data-driven knowledge, shows gains of $0.9\%$ in UF1 and $0.66\%$ in UAR. When both psychological and data-driven knowledge are combined (Setting XII), the improvements reach $2.38\%$ in UF1 and $2.77\%$ in UAR, proving the effectiveness of the prior knowledge. Constructing a prior adjacency matrix allows the model to capture complex facial muscle topology, while data-driven knowledge directs attention to uncover intrinsic movements patterns of facial muscles. + +# 5.4. Federated Experiment + +Since FL aims to enhance MER performance on individual clients in scenarios with limited sample sizes and data privacy issue, and given that DFME and $\mathrm{CAS}(\mathrm{ME})^3$ have large sample size, we partition the datasets as mentioned in Section 5.1 and treat the subdivisions as separate clients. + +While our focus is on application rather than fundamental FL innovation, our experiments confirm the effectiveness of our P-FedProx method, which is tailored for ME data distributions. Benchmarked against several open-source SOTA methods, our approach excels at mitigating small-sample challenges while preserving data privacy. Specifically (Fig. 5), the results show P-FedProx outperforms both the single-client baseline and traditional FL methods (FedAvg, FedProx), with particularly significant gains on clients with more challenging data distributions (Clients 6-7). Compared to recent personalized FL methods, our approach surpasses FedRep and ELLP and achieves performance competitive with the state-of-the-art FedAS, yet with a significantly simpler implementation. These consistent gains demonstrate that our weighted model initialization effectively addresses the data heterogeneity that challenges + +traditional FL, thereby mitigating ME sample scarcity while ensuring data privacy. + +Meantime, we note that all FL methods achieved good performance on the local clients from DFME, which may be due to the fact that these clients have a similar amount of data and slight differences in data distribution, presenting fewer challenges for federated learning. + +Besides, we analyze the training loss progression of different federated algorithms over communication rounds. We found that P-FedProx demonstrates the smoothest convergence curves with minimal oscillations, particularly evident in Clients 1-5. On the other hand, P-FedProx converges significantly faster than other FL methods. This rapid early convergence suggests that personalized global model assignment provides clients with better initialization states. + +# 6. Conclusion + +MER is challenged by limited sample sizes, subtle features, and privacy concerns in real-world applications. To address these issues, we conduct a psychological study to provide structured insights into facial muscle dynamics. We then develop a FED-PsyAU framework that integrates these findings with data-driven patterns for hierarchical learning of dynamic features, enhancing MER performance. Additionally, the FL framework improves MER across clients while preserving data privacy through iterative model updates. The comprehensive experiments validate our approach's effectiveness. In the future, we will integrate additional AUs to build a more comprehensive ME topology information and further advance MER in real-world applications through FL and semi-supervised approaches. + +# 7. Acknowledgment + +This research was partially funded by 1) the National Natural Science Foundation of China (62476269, 62276252); 2) the Youth Innovation Promotion Association CAS. + +# References + +[1] Xianye Ben, Yi Ren, Junping Zhang, Su-Jing Wang, Kidiyo Kpalma, Weixiao Meng, and Yong-Jin Liu. Video-Based Facial Micro-Expression Analysis: A Survey of Datasets, Features and Algorithms. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):5826-5846, 2022. 2 +[2] Joao Carreira and Andrew Zisserman. Quo Vadis, Action Recognition? A New Model and the Kinetics Dataset. In proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 7 +[3] Boyu Chen, Zhihao Zhang, Nian Liu, Yang Tan, Xinyu Liu, and Tong Chen. Spatiotemporal Convolutional Neural Network with Convolutional Block Attention Module for Micro-Expression Recognition. Information, 11(8):380, 2020. 1 +[4] Bin Chen, Kun-Hong Liu, Yong Xu, Qing-Qiang Wu, and Jun-Feng Yao. Block Division Convolutional Network With Implicit Deep Features Augmentation for Micro-Expression Recognition. IEEE Transactions on Multimedia, 25:1345–1358, 2022. 1, 2, 7 +[5] Liam Collins, Hamed Hassani, Aryan Mokhtari, and Sanjay Shakkottai. Exploiting shared representations for personalized federated learning. In International conference on machine learning, pages 2089-2099. PMLR, 2021. 6 +[6] Adrian K. Davison, Cliff Lansley, Nicholas Costen, Kevin Tan, and Moi Hoon Yap. SAMM: A Spontaneous MicroFacial Movement Dataset. IEEE Transactions on Affective Computing, 9(1):116-129, 2018. 2 +[7] Zizhao Dong, Gang Wang, Shaoyuan Lu, Jingting Li, Wenjing Yan, and Su-Jing Wang. Spontaneous Facial Expressions and Micro-Expressions Coding: From brain to Face. Frontiers in Psychology, 12:784834, 2022. 4 +[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale, 2021. 3 +[9] Paul Ekman and Wallace V Friesen. Nonverbal Leakage and Clues to Deception. Psychiatry, 32(1):88-106, 1969. 1 +[10] Paul Ekman and Wallace V Friesen. Facial Action Coding System. Environmental Psychology & Nonverbal Behavior, 1978. 1, 2, 3, 4 +[11] Ali et al. Privacy preserving personalization for video facial expression recognition using federated learning. IMCI'22, . 2 +[12] Fan et al. Feeling without sharing: A federated video emotion recognition framework via privacy-agnostic hybrid aggregation. ACM MM'23, . 2 +[13] Xinqi Fan, Xueli Chen, Mingjie Jiang, Ali Raza Shahid, and Hong Yan. SelfME: Self-Supervised Motion Learning for Micro-Expression Recognition. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13834-13843, 2023. 2 +[14] Viswanatha Reddy Gajjala, Sai Prasanna Teja Reddy, Snehasis Mukherjee, and Shiv Ram Dubey. MERANet: facial micro-expression recognition using 3D residual attention network. In Proceedings of the Twelfth Indian Conference on Computer Vision, Graphics and Image Processing, New York, NY, USA, 2021. Association for Computing Machinery. 1 +[15] Yee Siang Gan, Sze-Teng Liong, Wei-Chuen Yau, Yen-Chang Huang, and Lit-Ken Tan. Off-ApexNet on microexpression recognition system. Signal Processing: Image Communication, 74:129-139, 2019. 7 +[16] Haihong Hao, Shuo Wang, Huixia Ben, Yanbin Hao, Yansong Wang, and Weiwei Wang. Hierarchical Space-Time Attention for Micro-Expression recognition. arXiv preprint arXiv:2405.03202, 2024. 7 +[17] Kensho Hara, Hirokatsu Kataoka, and Yutaka Satoh. Can Spatiotemporal 3D CNNs Retrace the History of 2D CNNs and Imagenet? In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 6546-6555, 2018. 7 +[18] Xiaohua Huang, Su-Jing Wang, Guoying Zhao, and Matti Piteikainen. Facial Micro-Expression Recognition Using Spatiotemporal Local Binary Pattern with Integral Projection. In 2015 IEEE International Conference on Computer Vision Workshop (ICCVW), pages 1–9. IEEE, 2015. 2 +[19] T. Kanade, J.F. Cohn, and Yingli Tian. Comprehensive database for facial expression analysis. In Proceedings Fourth IEEE International Conference on Automatic Face and Gesture Recognition (Cat. No. PR00580), pages 46-53, 2000. 6 +[20] Davis E King. Dlib-ml: A Machine Learning Toolkit. The Journal of Machine Learning Research, 10:1755-1758, 2009. 3 +[21] Thomas N Kipf and Max Welling. Semi-Supervised Classification with Graph Convolutional Networks. arXiv preprint arXiv:1609.02907, 2016. 1, 2 +[22] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 7 +[23] Sang Gyu Kwak and Jong Hae Kim. Central limit theorem: the cornerstone of modern statistics. Korean journal of anesthesiaology, 70(2):144, 2017. 3 +[24] Ling Lei, Tong Chen, Shigang Li, and Jianfeng Li. Microexpression Recognition Based on Facial Graph Representation Learning and Facial Action Unit Fusion. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1571-1580. IEEE, 2021. 1, 2, 7 +[25] Hanting Li, Mingzhe Sui, Zhaoqing Zhu, and Feng Zhao. Mmnet: Muscle motion-guided network for microexpression recognition. arXiv preprint arXiv:2201.05297, 2022. 7 +[26] Jingting Li, Moi Hoon Yap, Wen-Huang Cheng, John See, Xiaopeng Hong, Xiaobai Li, and Su-Jing Wang. FME'21: + +1st Workshop on Facial Micro-Expression: Advanced Techniques for Facial Expressions Generation and Spotting. In Proceedings of the 29th ACM International Conference on Multimedia, pages 5700-5701. ACM, 2021. 2 +[27] Jingting Li, Zizhao Dong, Shaoyuan Lu, Su-Jing Wang, Wen-Jing Yan, Yinhuan Ma, Ye Liu, Changbing Huang, and Xiaolan Fu. CAS(ME) $^3$ : A Third Generation Facial Spontaneous Micro-Expression Database With Depth Information and High Ecological Validity. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):2782-2800, 2022. 2 +[28] Tian Li, Anit Kumar Sahu, Maziar Sanjabi, Manzil Zaheer, Ameet Talwalkar, and Virginia Smith. On the Convergence of Federated Optimization in Heterogeneous Networks. arXiv preprint arXiv:1812.06127, 2018. 2, 6 +[29] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks, 2020. 6 +[30] Xiaobai Li, Tomas Pfister, Xiaohua Huang, Guoying Zhao, and Matti Pietikainen. A Spontaneous Micro-expression Database: Inducement, collection and baseline. In 2013 10th IEEE International Conference and Workshops on Automatic Face and Gesture Recognition (FG), pages 1-6, 2013. 2 +[31] Xiaobai Li, Xiaopeng Hong, Antti Moilanen, Xiaohua Huang, Tomas Pfister, Guoying Zhao, and Matti Pietikainen. Towards Reading Hidden Emotions: A Comparative Study of Spontaneous Micro-Expression Spotting and Recognition Methods. IEEE Transactions on Affective Computing, 9(4): 563–577, 2018. 1 +[32] Xiaobai Li, Shiyang Cheng, Yante Li, Muzammil Behzad, Jie Shen, Stefanos Zafeiriou, Maja Pantic, and Guoying Zhao. 4DME: A Spontaneous 4D Micro-Expression Dataset With Multimodalities. IEEE Transactions on Affective Computing, 14(4):3031-3047, 2023. 2 +[33] Yante Li, Xiaohua Huang, and Guoying Zhao. Microexpression action unit detection with spatial and channel attention. Neurocomputing, 436:221-231, 2021. 1 +[34] Sze-Teng Liong, John See, Raphael C.-W. Phan, Anh Cat Le Ngo, Yee-Hui Oh, and KokSheik Wong. Subtle Expression Recognition Using Optical Strain Weighted Features. In Computer Vision - ACCV 2014 Workshops, pages 644-657. Springer International Publishing, 2015. Series Title: Lecture Notes in Computer Science. 2 +[35] Sze-Teng Liong, John See, KokSheik Wong, and Raphael C.-W. Phan. Less is More: Micro-expression Recognition from Video using Apex Frame. Signal Processing: Image Communication, 62:82-92, 2018. 2 +[36] Sze-Teng Liong, Y. S. Gan, John See, Huai-Qian Khor, and Yen-Chang Huang. Shallow Triple Stream Three-dimensional CNN (STSTNet) for Micro-expression Recognition. In 2019 14th IEEE International Conference on Automatic Face & Gesture Recognition (FG 2019), pages 1-5, 2019. 7 +[37] Yong-Jin Liu, Jin-Kai Zhang, Wen-Jing Yan, Su-Jing Wang, Guoying Zhao, and Xiaolan Fu. A Main Directional Mean Optical Flow Feature for Spontaneous Micro-Expression Recognition. IEEE Transactions on Affective Computing, 7 (4):299–310, 2015. 7 + +[38] Yong-Jin Liu, Jin-Kai Zhang, Wen-Jing Yan, Su-Jing Wang, Guoying Zhao, and Xiaolan Fu. A main directional mean optical flow feature for spontaneous micro-expression recognition. IEEE Transactions on Affective Computing, 7(4):299-310, 2016. 2 +[39] Ling Lo, Hong-Xia Xie, Hong-Han Shuai, and Wen-Huang Cheng. MER-GCN: Micro-Expression Recognition Based on Relation Modeling with Graph Convolutional Networks. In 2020 IEEE conference on multimedia information processing and retrieval (MIPR), pages 79–84. IEEE, 2020. 1 +[40] S Mohammad Mavadati, Mohammad H Mahoor, Kevin Bartlett, Philip Trinh, and Jeffrey F Cohn. Disfa: A spontaneous facial action intensity database. IEEE Transactions on Affective Computing, 4(2):151-160, 2013. 6 +[41] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-Efficient Learning of Deep Networks from Decentralized Data. In Proceedings of the 20th International Conference on Artificial Intelligence and Statistics, pages 1273-1282. PMLR, 2017. ISSN: 2640-3498. 2, 6 +[42] Xuan-Bac Nguyen, Chi Nhan Duong, Xin Li, Susan Gauch, Han-Seok Seo, and Khoa Luu. Micron-BERT: BERT-Based Facial Micro-Expression Recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1482–1492, 2023. 7 +[43] T. Ojala, M. Pietikainen, and T. Maenpaa. Multiresolution gray-scale and rotation invariant texture classification with local binary patterns. IEEE Transactions on Pattern Analysis and Machine Intelligence, 24(7):971-987, 2002. 2 +[44] Fangbing Qu, Su-Jing Wang, Wen-Jing Yan, He Li, Shuhang Wu, and Xiaolan Fu. CAS(ME) $^2$ : A Database for Spontaneous Macro-Expression and Micro-Expression Spotting and Recognition. IEEE Transactions on Affective Computing, 9(4):424-436, 2018. 2 +[45] John See, Moi Hoon Yap, Jingting Li, Xiaopeng Hong, and Su-Jing Wang. MEGC 2019 – The Second Facial Micro-Expressions Grand Challenge. In 2019 14th IEEE International Conference on Automatic Face & Gesture Recognition (FG 2019), pages 1–5, 2019. 2, 6 +[46] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE Conference on Computer Cision and Pattern Recognition, pages 1-9, 2015. 5 +[47] Petar Velicković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph Attention Networks. arXiv preprint arXiv:1710.10903, 2017. 1 +[48] Lei Wang, Pinyi Huang, Wangyang Cai, and Xiyao Liu. Micro-expression recognition by fusing action unit detection and spatio-temporal features. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 5595-5599. IEEE, 2024. 2 +[49] Su-Jing Wang, Wen-Jing Yan, Xiaobai Li, Guoying Zhao, Chun-Guang Zhou, Xiaolan Fu, Minghao Yang, and Jianhua Tao. Micro-expression recognition using color spaces. + +IEEE Transactions on Image Processing, 24(12):6034-6047, 2015. 1, 2, 7 +[50] Zhifeng Wang, Kaihao Zhang, Wenhan Luo, and Ramesh Sankaranarayana. Htnet for micro-expression recognition. Neurocomputing, 602:128196, 2024. 1, 2, 7 +[51] Wen-Jing Yan, Qi Wu, Yong-Jin Liu, Su-Jing Wang, and Xiaolan Fu. CASME database: A dataset of spontaneous micro-expressions collected from neutralized faces. In 2013 10th IEEE International Conference and Workshops on Automatic Face and Gesture Recognition (FG), pages 1-7. IEEE, 2013. 2 +[52] Zhaoqiang Xia, Wei Peng, Huai-Qian Khor, Xiaoyi Feng, and Guoying Zhao. Revealing the invisible with model and data shrinking for composite-database micro-expression recognition. IEEE Transactions on Image Processing, 29: 8590-8605, 2020. 7 +[53] Hong-Xia Xie, Ling Lo, Hong-Han Shuai, and Wen-Huang Cheng. AU-assisted graph attention convolutional network for micro-expression recognition. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2871-2880. ACM, 2020. 2 +[54] Wen-Jing Yan, Xiaobai Li, Su-Jing Wang, Guoying Zhao, Yong-Jin Liu, Yu-Hsin Chen, and Xiaolan Fu. CASME II: An improved spontaneous micro-expression database and the baseline evaluation. PLoS ONE, 9(1):e86041, 2014-01-27. 2 +[55] Xiyuan Yang, Wenke Huang, and Mang Ye. Fedas: Bridging inconsistency in personalized federated learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11986-11995, 2024. 6 +[56] Moi Hoon Yap, John See, Xiaopeng Hong, and Su-Jing Wang. Facial Micro-Expressions Grand Challenge 2018 Summary. In 2018 13th IEEE International Conference on Automatic Face & Gesture Recognition (FG 2018), pages 675–678, 2018. 2 +[57] Zhijun Zhai, Jianhui Zhao, Chengjiang Long, Wenju Xu, Shuangjiang He, and Huijuan Zhao. Feature representation learning with adaptive displacement generation and transformer fusion for micro-expression recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22086-22095, 2023. 1 +[58] Sirui Zhao, Hanqing Tao, Yangsong Zhang, Tong Xu, Kun Zhang, Zhongkai Hao, and Enhong Chen. A two-stage 3D CNN based learning method for spontaneous microexpression recognition. Neurocomputing, 448:276-289, 2021. 7 +[59] Sirui Zhao, Huaying Tang, Xinglong Mao, Shifeng Liu, Yiming Zhang, Hao Wang, Tong Xu, and Enhong Chen. DFME: A New Benchmark for Dynamic Facial MicroExpression Recognition. IEEE Transactions on Affective Computing, 15(3):1371-1386, 2024. 2, 6 +[60] Ling Zhou, Qirong Mao, Xiaohua Huang, Feifei Zhang, and Zhihong Zhang. Feature Refinement: An expression-specific feature learning and fusion method for micro-expression recognition. Pattern Recognition, 122:108275, 2022. 7 +[61] Xu Zhou, Jie Li, Gongjin Lan, Rongrong Ni, Angelo Cangelosi, Jiaxin Wang, and Xiaofeng Liu. Efficient lower + +layers parameter decoupling personalized federated learning method of facial expression recognition for home care robots. Information Fusion, 106:102261, 2024. 6 \ No newline at end of file diff --git a/2025/FED-PsyAU_ Privacy-Preserving Micro-Expression Recognition via Psychological AU Coordination and Dynamic Facial Motion Modeling/images.zip b/2025/FED-PsyAU_ Privacy-Preserving Micro-Expression Recognition via Psychological AU Coordination and Dynamic Facial Motion Modeling/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..afbb0767bac45bc6a95642b03abfb25e83128a0b --- /dev/null +++ b/2025/FED-PsyAU_ Privacy-Preserving Micro-Expression Recognition via Psychological AU Coordination and Dynamic Facial Motion Modeling/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:09665861e9e0267082912531ad30b614ddbebe298e2232f10e1685a5dd8c77d1 +size 446997 diff --git a/2025/FED-PsyAU_ Privacy-Preserving Micro-Expression Recognition via Psychological AU Coordination and Dynamic Facial Motion Modeling/layout.json b/2025/FED-PsyAU_ Privacy-Preserving Micro-Expression Recognition via Psychological AU Coordination and Dynamic Facial Motion Modeling/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..18daafef99ed45d51be9101776d91276fc75eac4 --- /dev/null +++ b/2025/FED-PsyAU_ Privacy-Preserving Micro-Expression Recognition via Psychological AU Coordination and Dynamic Facial Motion Modeling/layout.json @@ -0,0 +1,10463 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 89, + 103, + 521, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 103, + 521, + 140 + ], + "spans": [ + { + "bbox": [ + 89, + 103, + 521, + 140 + ], + "type": "text", + "content": "FED-PsyAU: Privacy-Preserving Micro-Expression Recognition via Psychological AU Coordination and Dynamic Facial Motion Modeling" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 164, + 160, + 447, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 160, + 447, + 175 + ], + "spans": [ + { + "bbox": [ + 164, + 160, + 447, + 175 + ], + "type": "text", + "content": "Jingting Li" + }, + { + "bbox": [ + 164, + 160, + 447, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 164, + 160, + 447, + 175 + ], + "type": "text", + "content": ", Yu Qian" + }, + { + "bbox": [ + 164, + 160, + 447, + 175 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 164, + 160, + 447, + 175 + ], + "type": "text", + "content": ", Lin Zhao" + }, + { + "bbox": [ + 164, + 160, + 447, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 164, + 160, + 447, + 175 + ], + "type": "text", + "content": ", Su-Jing Wang" + }, + { + "bbox": [ + 164, + 160, + 447, + 175 + ], + "type": "inline_equation", + "content": "^{*1,2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 94, + 175, + 517, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 175, + 517, + 203 + ], + "spans": [ + { + "bbox": [ + 94, + 175, + 517, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 94, + 175, + 517, + 203 + ], + "type": "text", + "content": "State Key Laboratory of Cognitive Science and Mental Health, Institute of Psychology, Chinese Academy of Sciences, Beijing, 100101, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 63, + 203, + 549, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 203, + 549, + 217 + ], + "spans": [ + { + "bbox": [ + 63, + 203, + 549, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 63, + 203, + 549, + 217 + ], + "type": "text", + "content": "Department of Psychology, University of the Chinese Academy of Sciences, Beijing, 100049, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 75, + 217, + 536, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 217, + 536, + 232 + ], + "spans": [ + { + "bbox": [ + 75, + 217, + 536, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 75, + 217, + 536, + 232 + ], + "type": "text", + "content": "School of Computer, Jiangsu University of Science and Technology, Zhenjiang, 212100, China" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 124, + 233, + 482, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 233, + 482, + 245 + ], + "spans": [ + { + "bbox": [ + 124, + 233, + 482, + 245 + ], + "type": "inline_equation", + "content": "^{\\dagger}" + }, + { + "bbox": [ + 124, + 233, + 482, + 245 + ], + "type": "text", + "content": "Equal contribution, \\*Corresponding author (wangsujing@psych.ac.cn)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "spans": [ + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 298, + 295, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 298, + 295, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 298, + 295, + 597 + ], + "type": "text", + "content": "Micro-expressions (MEs) are brief, low-intensity, often localized facial expressions. They could reveal genuine emotions individuals may attempt to conceal, valuable in contexts like criminal interrogation and psychological counseling. However, ME recognition (MER) faces challenges, such as small sample sizes and subtle features, which hinder efficient modeling. Additionally, real-world applications encounter ME data privacy issues, leaving the task of enhancing recognition across settings under privacy constraints largely unexplored. To address these issues, we propose a FED-PsyAU research framework. We begin with a psychological study on the coordination of upper and lower facial action units (AUs) to provide structured prior knowledge of facial muscle dynamics. We then develop a DPK-GAT network that combines these psychological priors with statistical AU patterns, enabling hierarchical learning of facial motion features from regional to global levels, effectively enhancing MER performance. Additionally, our federated learning (FL) framework advances MER capabilities across multiple clients without data sharing, preserving privacy and alleviating the limited-sample issue for each client. Extensive experiments on commonly-used ME databases demonstrate the effectiveness of our approach. Our implementation is publicly available at https://github.com/MELABIPCAS/FED-PsyAU.git." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 620, + 135, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 620, + 135, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 620, + 135, + 633 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 641, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 713 + ], + "type": "text", + "content": "Micro-expressions (MEs) [9], resulting from conflicts between voluntary and involuntary expressions, reveal true emotions, making them valuable in non-contact emotion detection applications like criminal interrogation and psychological counseling. MEs are characterized by their brief duration, low intensity, localized occurrence, and occasional" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 274, + 553, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 553, + 490 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 553, + 490 + ], + "type": "text", + "content": "asymmetry [10]. These traits not only restrict large-scale data collection and annotation but also make it challenging for deep learning networks to effectively model their motion characteristics. Some methods focus on regions of interest (ROIs) [4, 21, 31, 49], while others utilize attention mechanisms [3, 14, 33] and transformer-based architectures [24, 50, 57] to capture salient features. Meanwhile, MEs are often complex manifestations involving the combination of multiple AUs. In contrast, individual action units (AUs) correspond to specific, anatomically-based facial muscle movements. Focusing on AUs allows us to model more fundamental and consistent facial appearance changes. Certain approaches use AU labels to construct adjacency matrices for graph convolutional networks (GCNs) [24, 39]. Yet, to the best of our knowledge, these methods often apply AU annotations directly without fully exploiting the anatomical structure and coordinated AU dynamics, limiting the network's understanding on ME." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 493, + 553, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 493, + 553, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 493, + 553, + 673 + ], + "type": "text", + "content": "To address these challenges, we leverage the theory that distinct neural pathways control upper and lower facial movements, resulting in two different motion patterns. Based on this, we conduct a psychological study to investigate upper/lower facial AU coordination, providing crucial prior knowledge on structural motion relationships underpinning MEs. Using both psychological priors and data-driven AU patterns, we design a Graph Attention Network (GAT) [47] based on dynamic prior knowledge, applied separately to the upper and lower regions. A final global GAT captures the topological relationships of muscle movements across the face. The resulting feature, combined with full-face optical flow (OF), is input into a subsequent dual-stream network, effectively integrating structural and dynamic information to enhance ME recognition (MER)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 715 + ], + "type": "text", + "content": "Besides, as mentioned earlier, although MEs have valuable applications in many privacy-sensitive contexts, the sensitivity of biometric data makes it impractical to aggre" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14453" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": "gate ME data across clients for training. Meanwhile, the sample size in a single client is often insufficient to train a robust model. Yet, the challenge of enhancing MER performance across diverse clients while preserving privacy remains largely unexplored. To address this, we adopt a federated learning (FL) framework, improving performance by merging and updating model parameters across clients without exchanging local data. This approach preserves data privacy and supports the MER practical deployment." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 180, + 294, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 180, + 294, + 242 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 294, + 242 + ], + "type": "text", + "content": "Overall, our work leverages psychological and data-driven priors to enhance the network's learning of structured ME motion patterns, particularly in limited-sample conditions. Additionally, we introduce FL to advance MER in real-world scenarios. The contributions of this paper are:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 243, + 295, + 399 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 56, + 243, + 295, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 295, + 290 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 295, + 290 + ], + "type": "text", + "content": "- We conduct a psychological experiment to derive AU coordination between the upper and lower facial regions, providing structural priors for ME and facial expression analysis in computer vision." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 292, + 294, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 292, + 294, + 362 + ], + "spans": [ + { + "bbox": [ + 56, + 292, + 294, + 362 + ], + "type": "text", + "content": "- We design a local-to-global GAT network that integrates psychological priors and data-driven AU patterns in the upper and lower facial regions, improving the network's ability to learn both local and global structural facial dynamics. This feature, combined with global OF feature in a dual-stream network, enhances MER performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 363, + 294, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 363, + 294, + 399 + ], + "spans": [ + { + "bbox": [ + 56, + 363, + 294, + 399 + ], + "type": "text", + "content": "- We employ FL in MER to improve data privacy and overcome small sample size challenges by aggregating features from distributed data-protected sources." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 411, + 148, + 424 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 148, + 424 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 148, + 424 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 432, + 218, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 432, + 218, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 432, + 218, + 445 + ], + "type": "text", + "content": "2.1. Micro-expression Recognition" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "type": "text", + "content": "In recent years, MER has gained increasing attention from computer vision researchers. Traditional machine learning methods primarily rely on handcrafted feature extraction. For instance, Local Binary Pattern (LBP) [43], effectively utilizes local texture features, has inspired various improvements [18, 49]. Furthermore, OF [26] has been explored for capturing object motion. Liu et al. [38] developed the Main Directional Mean OF (MDMO) method by aligning faces in the OF domain, while Liong et al. [34] added optical strain features for MER. However, traditional approaches depend heavily on manually crafted features, limiting both accuracy and generalizability. Meanwhile, deep learning methods, with robust feature extraction capabilities, can automatically learn multi-level representations to capture subtle ME changes. The availability of several spontaneous ME databases, such as CASME II [51], SAMM [6], CAS(ME)" + }, + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "type": "text", + "content": " [27], DFME [59] and others [1, 30, 32, 44, 54], has accelerated deep learning advancements in MER. Inspired by Liong et al.'s work [35] on sequence redundancy, many studies [4, 13, 50] now use keyframes (onset and apex) as input, reducing model parameters while achieving excellent results." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "text", + "content": "Despite advancements in preprocessing and feature extraction, deep learning alone has yet to achieve high MER accuracy. As MEs stem from physiological muscle activity, incorporating prior knowledge offers meaningful context. FACS [10], which defines expressions through AUs, provides a basic theoretical framework for MER. Recent studies explore AU-ME relationships to enhance MER performance. Lei et al. [24] applied GCN [21] to model AU co-occurrences, especially around the eyebrow and mouth regions. Xie et al. [53] introduced an AU-assisted GCN, with modules for AU feature extraction and ME image generation. Wang et al. [48] used transformers to learn dynamic AU adjacency, boosting model generalization. These networks typically rely on AU labels from databases, without fully integrating structured facial muscle movements, AU interrelations, or their links to expressions. In this work, we refine structured ME feature extraction using a local-to-global GAT network, which integrates priors on facial AU coordination from psychological experiments with data-driven movement patterns." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 322, + 430, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 322, + 430, + 335 + ], + "spans": [ + { + "bbox": [ + 314, + 322, + 430, + 335 + ], + "type": "text", + "content": "2.2. Federated Learning" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 340, + 553, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 553, + 436 + ], + "type": "text", + "content": "To assess MER across databases, ME Grand Challenge (MEGC) 2018 [56] introduced the Holdout-database Evaluation and Composite Database Evaluation tasks, utilizing the CASME II and SAMM databases. MEGC 2019 [45] expanded the CDE task to include the SMIC databases. While research on MEs predominantly utilizes these public datasets, existing methodologies do not address the privacy concerns that arise in practical applications." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 437, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 628 + ], + "type": "text", + "content": "FL [41] could addresses ME data scarcity and privacy concerns through collaborative training. It includes two main phases: local training and global aggregation. Many existing FL works for expressions, while insightful, are often not open-source or use highly customized frameworks (e.g., non-random client data allocation [12], specialized parameter aggregation [11]) making standard methods more suitable for implementation and comparisons in our study. For instance, McMahan et al. [41] proposed FedAvg, a basic aggregation method that uses weighted averaging based on each client's data volume. Li et al. [28] introduced FedProx by adding a regularization term to the client loss function, aligning updates more closely with the global model. However, the one-global model-fits-all approach ignores client data diversity. To solve this, our approach extends FedProx by distributing a personalized global model to each client." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 642, + 553, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 669 + ], + "type": "text", + "content": "3. Proposed Psychological Study on AU Incoordination" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "This study employed a behavioral experiment to investigate individuals' cognitive responses to AU incoordination in the upper and lower facial regions." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14454" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 99, + 123 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 99, + 123 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 99, + 123 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 99, + 123 + ], + "type": "image", + "image_path": "d0be976c39a1a2b1350002c8024a300f5603459bf08e9cad7c5424168a69b569.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 60, + 125, + 97, + 133 + ], + "lines": [ + { + "bbox": [ + 60, + 125, + 97, + 133 + ], + "spans": [ + { + "bbox": [ + 60, + 125, + 97, + 133 + ], + "type": "text", + "content": "(a) AU " + }, + { + "bbox": [ + 60, + 125, + 97, + 133 + ], + "type": "inline_equation", + "content": "4 + 9" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 123, + 70, + 164, + 123 + ], + "blocks": [ + { + "bbox": [ + 123, + 70, + 164, + 123 + ], + "lines": [ + { + "bbox": [ + 123, + 70, + 164, + 123 + ], + "spans": [ + { + "bbox": [ + 123, + 70, + 164, + 123 + ], + "type": "image", + "image_path": "6a6758a36e511f40688f624788965fb5f708c901579da499564d7275e17f1144.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 123, + 125, + 164, + 133 + ], + "lines": [ + { + "bbox": [ + 123, + 125, + 164, + 133 + ], + "spans": [ + { + "bbox": [ + 123, + 125, + 164, + 133 + ], + "type": "text", + "content": "(b) AU 6+12" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 187, + 70, + 228, + 123 + ], + "blocks": [ + { + "bbox": [ + 187, + 70, + 228, + 123 + ], + "lines": [ + { + "bbox": [ + 187, + 70, + 228, + 123 + ], + "spans": [ + { + "bbox": [ + 187, + 70, + 228, + 123 + ], + "type": "image", + "image_path": "75bb2e5693552c09e2c1a618dc3ae130c5010e3798a28cb5539630fc5fb61851.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 187, + 125, + 229, + 133 + ], + "lines": [ + { + "bbox": [ + 187, + 125, + 229, + 133 + ], + "spans": [ + { + "bbox": [ + 187, + 125, + 229, + 133 + ], + "type": "text", + "content": "(c) AU 4+12" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 144, + 295, + 167 + ], + "lines": [ + { + "bbox": [ + 55, + 144, + 295, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 295, + 167 + ], + "type": "text", + "content": "Figure 1. Examples of AU composite images. (a) and (b) are coordinated composites. (c) and (d) are uncoordinated composites." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 252, + 70, + 293, + 123 + ], + "blocks": [ + { + "bbox": [ + 252, + 70, + 293, + 123 + ], + "lines": [ + { + "bbox": [ + 252, + 70, + 293, + 123 + ], + "spans": [ + { + "bbox": [ + 252, + 70, + 293, + 123 + ], + "type": "image", + "image_path": "c488cf12372e861cf51dcda4ad2396759b6b73ada71b35e9bbaf1e1974f014a9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 252, + 125, + 294, + 133 + ], + "lines": [ + { + "bbox": [ + 252, + 125, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 252, + 125, + 294, + 133 + ], + "type": "text", + "content": "(d) AU 6+15" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 189, + 295, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 189, + 295, + 260 + ], + "spans": [ + { + "bbox": [ + 55, + 189, + 295, + 260 + ], + "type": "text", + "content": "The study included 30 college participants (20 female, Mean= 22.22 years), a sample size considered adequate for statistical reliability under the Central Limit Theorem [23]. All procedures followed the Declaration of Helsinki and were approved by the institutional review board of the Institute of Psychology, Chinese Academy of Sciences." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 261, + 295, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 261, + 295, + 416 + ], + "spans": [ + { + "bbox": [ + 55, + 261, + 295, + 416 + ], + "type": "text", + "content": "This behavioral experiment employed a within-subjects design. The independent variable was the type of AU composite image (coordinated vs. uncoordinated AU composites), as shown in Fig. 1. The face images were segmented anatomically. The dividing line was consistently present in all stimuli, serving as a control variable and would not influence the experiment's outcomes. To control for individual facial differences in emotion recognition, the original images used for composites were sourced from the same person [10]. Seventy-two composite images were generated from upper and lower facial AUs across six basic emotions, with each emotion represented by six coordinated and six uncoordinated expressions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 417, + 295, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 417, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 417, + 295, + 464 + ], + "type": "text", + "content": "The dependent variables are accuracy rates and scores. Specifically, participants made two judgments for each composite image: whether it was coordinated or uncoordinated, and the degree of coordination or incoordination." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 465, + 295, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 465, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 55, + 465, + 295, + 645 + ], + "type": "text", + "content": "Statistical analysis from this study showed that participants were significantly more accurate in recognizing coordinated AU composite images than uncoordinated ones (Please see Supplementary Materials (Suppl.) for more experimental details and statistical results). The study further highlighted specific AU combinations, such as AU4 and AU12, that participants quickly identified as uncoordinated, whereas combinations like AU6 and AU12 were rapidly recognized as coordinated. This suggests that these uncoordinated AU pairings may inherently convey emotional incoordination. Besides, the coordination patterns between specific AUs reveal systematic facial action structures. These patterns help the network distinguish both coordinated and uncoordinated facial movements, thus enhancing its ability to identify ME features more effectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 657, + 272, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 272, + 671 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 272, + 671 + ], + "type": "text", + "content": "4. Our Proposed FED-PsyAU Framework" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "We propose the FED-PsyAU framework (Fig. 2), which embeds a psychology-driven MER model within a federated aggregation module. The model features a hierarchi" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "cal structure that learns from local ROIs to AU groups and finally to global facial regions, capturing dynamic topology while minimizing redundancy. It leverages a multiscale InceptionNet to extract and fuse global OF with structural facial features. The federated module enables privacy-preserving collaborative training to boost MER performance on each client." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 172, + 454, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 172, + 454, + 185 + ], + "spans": [ + { + "bbox": [ + 313, + 172, + 454, + 185 + ], + "type": "text", + "content": "4.1. Localized ROI Modeling" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 192, + 553, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 553, + 276 + ], + "type": "text", + "content": "Since MEs are manifested by muscle movements in key facial regions (e.g., eyes, mouth, nose) [10], we refine the standard 68 landmarks from Dlib [20] to precisely capture these changes. As shown in Fig. 3, we discard 10 less informative outer contour landmarks and add 7 new midpoints to better track movements in the cheek and eyebrow regions, resulting in a final set of 65 keypoints." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 279, + 555, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 555, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 555, + 459 + ], + "type": "text", + "content": "Then, onset-apex OF inputs " + }, + { + "bbox": [ + 313, + 279, + 555, + 459 + ], + "type": "inline_equation", + "content": "(128 \\times 128" + }, + { + "bbox": [ + 313, + 279, + 555, + 459 + ], + "type": "text", + "content": " pixels) are cropped into 65 ROIs, each measuring " + }, + { + "bbox": [ + 313, + 279, + 555, + 459 + ], + "type": "inline_equation", + "content": "5 \\times 5" + }, + { + "bbox": [ + 313, + 279, + 555, + 459 + ], + "type": "text", + "content": " pixels, centered on selected facial landmarks " + }, + { + "bbox": [ + 313, + 279, + 555, + 459 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_k = (x_k, y_k)" + }, + { + "bbox": [ + 313, + 279, + 555, + 459 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 279, + 555, + 459 + ], + "type": "inline_equation", + "content": "k \\in 1, 2, \\ldots, 65" + }, + { + "bbox": [ + 313, + 279, + 555, + 459 + ], + "type": "text", + "content": ", denotes the landmark index. To enhance robustness to apex annotation, we also compute OF using a small window of frames around the apex. The ROI features input to the Local ROI Feature Extractor (LFE), denoted as " + }, + { + "bbox": [ + 313, + 279, + 555, + 459 + ], + "type": "inline_equation", + "content": "\\Theta_k" + }, + { + "bbox": [ + 313, + 279, + 555, + 459 + ], + "type": "text", + "content": ", contain the horizontal and vertical components of OF and optical strain. The LFE module consists of three convolutional layers, each followed by batch normalization and ReLU activation. All kernel sizes were " + }, + { + "bbox": [ + 313, + 279, + 555, + 459 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 313, + 279, + 555, + 459 + ], + "type": "text", + "content": ". The output channels in the first two layers increase from 32 to 64, while the final layer outputs 3 channels. This design helps the LFE capture complex local patterns, with the output features referred to as " + }, + { + "bbox": [ + 313, + 279, + 555, + 459 + ], + "type": "inline_equation", + "content": "\\Phi_k = \\mathrm{LFE}(\\Theta_k)" + }, + { + "bbox": [ + 313, + 279, + 555, + 459 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 462, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 617 + ], + "type": "text", + "content": "Inspired by Vision Transformer [8], we use a Spatial Structure Encoder (SSE), i.e., Transformer Encoder with three attention heads to capture diverse dependencies among ROIs, as MEs result from coordinated muscle movements across multiple areas. Precisely, we convert " + }, + { + "bbox": [ + 313, + 462, + 555, + 617 + ], + "type": "inline_equation", + "content": "\\Phi_k" + }, + { + "bbox": [ + 313, + 462, + 555, + 617 + ], + "type": "text", + "content": ", into a token. To retain the positional information of the facial landmarks, a positional embedding " + }, + { + "bbox": [ + 313, + 462, + 555, + 617 + ], + "type": "inline_equation", + "content": "E_{k}" + }, + { + "bbox": [ + 313, + 462, + 555, + 617 + ], + "type": "text", + "content": " is added based on the landmark index to each token, yielding " + }, + { + "bbox": [ + 313, + 462, + 555, + 617 + ], + "type": "inline_equation", + "content": "X_{k} = \\Phi_{k} + E_{k}" + }, + { + "bbox": [ + 313, + 462, + 555, + 617 + ], + "type": "text", + "content": ". We add the output of the three head attention module to the input feature " + }, + { + "bbox": [ + 313, + 462, + 555, + 617 + ], + "type": "inline_equation", + "content": "X_{k}" + }, + { + "bbox": [ + 313, + 462, + 555, + 617 + ], + "type": "text", + "content": ", and then normalize the result using LayerNorm. Finally, after passing through a feed-forward neural network and applying LayerNorm, the final output representation is obtained as " + }, + { + "bbox": [ + 313, + 462, + 555, + 617 + ], + "type": "inline_equation", + "content": "Z_{k}\\in \\mathbb{R}^{3\\times 5\\times 5}" + }, + { + "bbox": [ + 313, + 462, + 555, + 617 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 633, + 539, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 633, + 539, + 646 + ], + "spans": [ + { + "bbox": [ + 313, + 633, + 539, + 646 + ], + "type": "text", + "content": "4.2. Relationship Modeling among AU Features" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "type": "text", + "content": "Our proposed Relationship Modeling among AU Features (AFR) module consists of two components: an AU Feature Extractor (AFE) to derive features from AU-specific facial regions, and a GAT-structure to model the topological relationships between these AUs." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14455" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 553, + 220 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 553, + 220 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 553, + 220 + ], + "type": "image", + "image_path": "32a7893b63bbeb533929b4b91016e08a888aebebc0656d7b79a2840f881c5d6b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 226, + 555, + 295 + ], + "lines": [ + { + "bbox": [ + 54, + 226, + 555, + 295 + ], + "spans": [ + { + "bbox": [ + 54, + 226, + 555, + 295 + ], + "type": "text", + "content": "Figure 2. The pipeline of our proposed FED-PsyAU framework, including MER network (black block) and the Federated Aggregation (FA) module (orange block). In MER network, the LRM module consists of the local ROI feature extractor (LFE) and the spatial structure encoder (SSE), which preserves spatial structure information while efficiently extracting ROI features; the AFR module includes the AU feature extractor (AFE) and the dynamic prior knowledge-based graph attention network (DPK-GAT) for effective AU feature extraction and capturing AU dynamic relationships, the DSI module use a dual-stream structure to capture multi-scale facial muscle topology and motion information for MER. Then, the FA module integrates the MER network into the FL framework." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 62, + 311, + 290, + 462 + ], + "blocks": [ + { + "bbox": [ + 62, + 311, + 290, + 462 + ], + "lines": [ + { + "bbox": [ + 62, + 311, + 290, + 462 + ], + "spans": [ + { + "bbox": [ + 62, + 311, + 290, + 462 + ], + "type": "image", + "image_path": "6fff2cada31386a5b011df13bd87f2c5f07231243518331d7ca049c484b4cd7c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 470, + 298, + 526 + ], + "lines": [ + { + "bbox": [ + 54, + 470, + 298, + 526 + ], + "spans": [ + { + "bbox": [ + 54, + 470, + 298, + 526 + ], + "type": "text", + "content": "Figure 3. ROI grouping strategy for AU. OF features from each ROI are processed through LFE and SSE to yield high-dimensional representations. ROIs are then grouped by AU movement patterns to create inputs for AFEs, ultimately producing AU node features for the GAT network." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 551, + 178, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 178, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 178, + 562 + ], + "type": "text", + "content": "4.2.1. AU Feature Extractor" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 567, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 567, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 54, + 567, + 295, + 700 + ], + "type": "text", + "content": "MEs involve dynamic, coordinated interactions across multiple regions [10] and are characterized by subtle movements that differ from typical macro-expressions. Therefore, as shown in Fig. 3, when grouping ROIs for the AFE module, the AU selection and regionalization are based on: 1) AU frequency in MEs; 2) Grouping by primary/secondary muscle activation areas [7]. Notably, not all AUs employed in Section 3 are applied to MER. We specifically retain only the 12 AUs exhibiting the highest relevance to ME dynamics, as listed in Tab. 1. Detailed occurrences are provided in the Suppl." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "The first step we construct in AFE is Group Squeeze and" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 316, + 354, + 555, + 408 + ], + "blocks": [ + { + "bbox": [ + 313, + 312, + 555, + 346 + ], + "lines": [ + { + "bbox": [ + 313, + 312, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 555, + 346 + ], + "type": "text", + "content": "Table 1. Selected AUs and Corresponding Regions (Primary and Secondary/ Coordinate). " + }, + { + "bbox": [ + 313, + 312, + 555, + 346 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 312, + 555, + 346 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 312, + 555, + 346 + ], + "type": "inline_equation", + "content": "N_g" + }, + { + "bbox": [ + 313, + 312, + 555, + 346 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 312, + 555, + 346 + ], + "type": "inline_equation", + "content": "N_{g\\_AU}" + }, + { + "bbox": [ + 313, + 312, + 555, + 346 + ], + "type": "text", + "content": " represent ROI Group index, corresponding ROI number and AU amount." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 354, + 555, + 408 + ], + "lines": [ + { + "bbox": [ + 316, + 354, + 555, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 555, + 408 + ], + "type": "table", + "html": "
gAUPrimarySecondary/ CoordinateNgNg-AU
11,2,4EyebrowsEyes233
25,6,7EyesEyebrows, Cheeks, Mouth483
39,10Nose, CheeksMouth, Chin382
412,14,15,17MouthEyes, Eyebrows, Nose524
", + "image_path": "4434210dc95c750de1434423c9b2d3c7ecffecf56f8da8459c219c8ad5f1d05b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "text", + "content": "Excitation (GSE). It computes the attention for each group of AUs (ROI group), which makes the model pay more attention to the critical ROIs. Specifically, for each ROI group input " + }, + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "inline_equation", + "content": "R_{g} = \\{Z_{1}, Z_{2}, \\ldots, Z_{N_{g}}\\}" + }, + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "inline_equation", + "content": "g \\in \\{1, 2, 3, 4\\}" + }, + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "text", + "content": ", the GSE module performs global average pooling on each ROI and computes channel importance through a fully connected (FC) layer. The original features are then preserved through residual concatenation:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 344, + 529, + 555, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 529, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 344, + 529, + 555, + 544 + ], + "type": "interline_equation", + "content": "f _ {\\mathrm {G S E}} \\left(R _ {g}\\right) = \\left\\{W _ {k} ^ {\\text {a t t}} \\odot Z _ {k} + Z _ {k} \\mid Z _ {k} \\in R _ {g} \\right\\} \\tag {1}", + "image_path": "12545b8036c8da9a903a8baae783c70030efa87f462bd0e251d200a264b02a49.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 549, + 556, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 549, + 556, + 670 + ], + "spans": [ + { + "bbox": [ + 313, + 549, + 556, + 670 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 549, + 556, + 670 + ], + "type": "inline_equation", + "content": "W_{k}^{att}" + }, + { + "bbox": [ + 313, + 549, + 556, + 670 + ], + "type": "text", + "content": " is the attention weight for each group of ROI, " + }, + { + "bbox": [ + 313, + 549, + 556, + 670 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 313, + 549, + 556, + 670 + ], + "type": "text", + "content": " denotes element-wise multiplication along the channel dimension of " + }, + { + "bbox": [ + 313, + 549, + 556, + 670 + ], + "type": "inline_equation", + "content": "Z_{k}" + }, + { + "bbox": [ + 313, + 549, + 556, + 670 + ], + "type": "text", + "content": ". Next, " + }, + { + "bbox": [ + 313, + 549, + 556, + 670 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{GSE}}(R_g)" + }, + { + "bbox": [ + 313, + 549, + 556, + 670 + ], + "type": "text", + "content": " undergoes further feature extraction, reducing from " + }, + { + "bbox": [ + 313, + 549, + 556, + 670 + ], + "type": "inline_equation", + "content": "N_{g}" + }, + { + "bbox": [ + 313, + 549, + 556, + 670 + ], + "type": "text", + "content": " composite features to " + }, + { + "bbox": [ + 313, + 549, + 556, + 670 + ], + "type": "inline_equation", + "content": "N_{g\\_AU}" + }, + { + "bbox": [ + 313, + 549, + 556, + 670 + ], + "type": "text", + "content": " representative AU features, yielding the output feature " + }, + { + "bbox": [ + 313, + 549, + 556, + 670 + ], + "type": "inline_equation", + "content": "F_{g\\_AU}" + }, + { + "bbox": [ + 313, + 549, + 556, + 670 + ], + "type": "text", + "content": ". Precisely, the process comprises two convolutional layers: a " + }, + { + "bbox": [ + 313, + 549, + 556, + 670 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 313, + 549, + 556, + 670 + ], + "type": "text", + "content": " layer that doubles the input dimensions and a " + }, + { + "bbox": [ + 313, + 549, + 556, + 670 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 313, + 549, + 556, + 670 + ], + "type": "text", + "content": " layer that restores them to the original size. Both layers are followed by Batch Normalization and ReLU activation for improved stability and representation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 674, + 457, + 687 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 674, + 457, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 674, + 457, + 687 + ], + "type": "text", + "content": "4.2.2. AU Relationship Modeling" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Prior Knowledge: To model AU dynamic relationships, we integrate two forms of prior knowledge: an \"intrinsic\" prior" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14456" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 216 + ], + "type": "text", + "content": "from psychological principles and an \"extrinsic\" prior from data-driven patterns. The intrinsic prior, derived from our psychological findings on AU coordination (Sec. 3), provides structural guidance by defining the initial GAT adjacency matrix " + }, + { + "bbox": [ + 54, + 72, + 294, + 216 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 54, + 72, + 294, + 216 + ], + "type": "text", + "content": ". The extrinsic prior, based on statistical AU co-occurrence patterns in the data, offers data-driven refinement that theoretical models may overlook. This extrinsic prior forms a prior attention matrix " + }, + { + "bbox": [ + 54, + 72, + 294, + 216 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 54, + 72, + 294, + 216 + ], + "type": "text", + "content": " to guide weight allocation among AU nodes. By fusing this foundational psychological structure with real-world dynamics, our model achieves superior adaptability and generalization (see Suppl. for matrix details)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 217, + 295, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 295, + 384 + ], + "type": "text", + "content": "Facial Segmentation: As presented in Sec. 3, the upper and lower facial regions, controlled by different neural pathways, exhibit distinct muscle movement patterns. AUs and their associated primary muscles in the upper and lower face can exhibit coordination (potential co-occurrence) or lack of coordination (infrequent simultaneous appearance). Leveraging these intra- and inter-regional interactions, we model AU relationships for the upper face, lower face, and whole face. Consistent with the experimental setup in Sec. 3, we segment the face into upper and lower parts based on the orbicularis oculi muscle. Specifically, following the primary occurrence regions of 12 listed AUs, the AUs in the first two groups from Tab. 1 are classified as upper face AUs, while those in the last two groups are classified as lower face AUs." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 393, + 295, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 393, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 61, + 393, + 295, + 409 + ], + "type": "interline_equation", + "content": "F _ {A U} ^ {U p p e r} = F _ {1 \\_ A U} \\cup F _ {2 \\_ A U}, F _ {A U} ^ {L o w e r} = F _ {3 \\_ A U} \\cup F _ {4 \\_ A U} (2)", + "image_path": "92f8b5a58bffd6a9464f94d867d95d0590995669269776922f740efb0780d8af.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 420, + 295, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 420, + 295, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 420, + 295, + 515 + ], + "type": "text", + "content": "We input the AU features from the upper and lower face: " + }, + { + "bbox": [ + 55, + 420, + 295, + 515 + ], + "type": "inline_equation", + "content": "F_{AU}^{Upper}" + }, + { + "bbox": [ + 55, + 420, + 295, + 515 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 420, + 295, + 515 + ], + "type": "inline_equation", + "content": "F_{AU}^{Lower}" + }, + { + "bbox": [ + 55, + 420, + 295, + 515 + ], + "type": "text", + "content": " into our proposed Dynamic Prior Knowledge GATs (DPK-GATs) in parallel (Details are presented in the next part of this subsection). These networks use psychology prior knowledge to build AU adjacency relationships and data-driven priors to guide attention learning, producing outputs " + }, + { + "bbox": [ + 55, + 420, + 295, + 515 + ], + "type": "inline_equation", + "content": "h^{Upper}" + }, + { + "bbox": [ + 55, + 420, + 295, + 515 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 420, + 295, + 515 + ], + "type": "inline_equation", + "content": "h^{Lower}" + }, + { + "bbox": [ + 55, + 420, + 295, + 515 + ], + "type": "text", + "content": ", which are combined into " + }, + { + "bbox": [ + 55, + 420, + 295, + 515 + ], + "type": "inline_equation", + "content": "h^{Local}" + }, + { + "bbox": [ + 55, + 420, + 295, + 515 + ], + "type": "text", + "content": " for the next stage." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 517, + 295, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 295, + 589 + ], + "type": "text", + "content": "In the whole-face DPK-GAT, we construct the network similarly but set adjacency values within the same region to zero, ignoring intra-regional AU relationships. By learning cross-regional topological structures and statistical patterns, we obtain global AU relationship features, denoted as " + }, + { + "bbox": [ + 55, + 517, + 295, + 589 + ], + "type": "inline_equation", + "content": "h^{Global}" + }, + { + "bbox": [ + 55, + 517, + 295, + 589 + ], + "type": "text", + "content": ", as illustrated in Fig. 4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 589, + 295, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 589, + 295, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 589, + 295, + 674 + ], + "type": "text", + "content": "DPK-GAT: We model dynamic AU node relationships combing GAT and the prior knowledge. In particular, for the input feature belonging to " + }, + { + "bbox": [ + 55, + 589, + 295, + 674 + ], + "type": "inline_equation", + "content": "\\{F_{AU}^{Upper}, F_{AU}^{Lower}, h^{Local}\\}" + }, + { + "bbox": [ + 55, + 589, + 295, + 674 + ], + "type": "text", + "content": ", a linear mapping projects it to a higher-dimensional space as " + }, + { + "bbox": [ + 55, + 589, + 295, + 674 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 589, + 295, + 674 + ], + "type": "text", + "content": ". The similarity score " + }, + { + "bbox": [ + 55, + 589, + 295, + 674 + ], + "type": "inline_equation", + "content": "e_{ij}" + }, + { + "bbox": [ + 55, + 589, + 295, + 674 + ], + "type": "text", + "content": " between AU nodes " + }, + { + "bbox": [ + 55, + 589, + 295, + 674 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 589, + 295, + 674 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 589, + 295, + 674 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 589, + 295, + 674 + ], + "type": "text", + "content": ", is calculated using the attention weight vector " + }, + { + "bbox": [ + 55, + 589, + 295, + 674 + ], + "type": "inline_equation", + "content": "\\vec{a}" + }, + { + "bbox": [ + 55, + 589, + 295, + 674 + ], + "type": "text", + "content": " and concatenated node features " + }, + { + "bbox": [ + 55, + 589, + 295, + 674 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 55, + 589, + 295, + 674 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 589, + 295, + 674 + ], + "type": "inline_equation", + "content": "f_{j}" + }, + { + "bbox": [ + 55, + 589, + 295, + 674 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 78, + 684, + 295, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 684, + 295, + 716 + ], + "spans": [ + { + "bbox": [ + 78, + 684, + 295, + 716 + ], + "type": "interline_equation", + "content": "e _ {i j} = \\left\\{ \\begin{array}{l l} \\text {L e a k y R e L U} \\left(\\tilde {a} ^ {T} [ f _ {i} | | f _ {j} ]\\right), & \\text {i f} A _ {i j} > 0 \\\\ - \\infty , & \\text {i f} A _ {i j} = 0 \\end{array} \\right. \\tag {3}", + "image_path": "24351a4b78bd8a78ae10c753202f1f1891244815106b14b7edd7be4ff312f791.jpg" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 553, + 155 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 553, + 155 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 553, + 155 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 553, + 155 + ], + "type": "image", + "image_path": "3934cbe5af16a98d999e3491078e08adc38f6f6edc979040b58195a36422df20.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 163, + 553, + 208 + ], + "lines": [ + { + "bbox": [ + 313, + 163, + 553, + 208 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 553, + 208 + ], + "type": "text", + "content": "Figure 4. DPK-GAT from facial segmentations to global face. First, intra-regional AU node relationships are learned within the upper and lower facial regions. Second, cross-regional relationships are modeled to integrate interactions between these regions." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 228, + 553, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 553, + 252 + ], + "type": "text", + "content": "Here, only neighboring nodes defined by the psychology-derived adjacency matrix " + }, + { + "bbox": [ + 313, + 228, + 553, + 252 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 228, + 553, + 252 + ], + "type": "text", + "content": " are considered." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 253, + 553, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 553, + 289 + ], + "type": "text", + "content": "The final attention " + }, + { + "bbox": [ + 313, + 253, + 553, + 289 + ], + "type": "inline_equation", + "content": "\\alpha_{ij}" + }, + { + "bbox": [ + 313, + 253, + 553, + 289 + ], + "type": "text", + "content": " is calculated by combining the data-driven prior attention " + }, + { + "bbox": [ + 313, + 253, + 553, + 289 + ], + "type": "inline_equation", + "content": "D_{ij}" + }, + { + "bbox": [ + 313, + 253, + 553, + 289 + ], + "type": "text", + "content": " and the self-attention, obtained by the softmax operation applied to " + }, + { + "bbox": [ + 313, + 253, + 553, + 289 + ], + "type": "inline_equation", + "content": "e_{ij}" + }, + { + "bbox": [ + 313, + 253, + 553, + 289 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 353, + 299, + 553, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 299, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 353, + 299, + 553, + 312 + ], + "type": "interline_equation", + "content": "\\alpha_ {i j} = (1 - \\beta) \\cdot \\operatorname {s o f t m a x} \\left(e _ {i j}\\right) + \\beta \\cdot D _ {i j} \\tag {4}", + "image_path": "ef5b362860b5dc2123af35c3baa1dbbc1edf0b8a0e1e0482dad2d3400b9f1185.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 320, + 553, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 320, + 553, + 367 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 553, + 367 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 313, + 320, + 553, + 367 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 320, + 553, + 367 + ], + "type": "text", + "content": " is the dynamic prior attention weight, updated based on the training rounds. Along with a residual concatenation, the final node representation " + }, + { + "bbox": [ + 313, + 320, + 553, + 367 + ], + "type": "inline_equation", + "content": "h_i" + }, + { + "bbox": [ + 313, + 320, + 553, + 367 + ], + "type": "text", + "content": " is aggregated as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 363, + 368, + 553, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 368, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 363, + 368, + 553, + 384 + ], + "type": "interline_equation", + "content": "h _ {i} = \\operatorname {E L U} \\left(\\sum_ {j \\in \\mathcal {N} (i)} \\alpha_ {i j} \\cdot f _ {j}\\right) + f _ {i} \\tag {5}", + "image_path": "975b78a07fe646ccff43c5cea81e31c666d5f00db150a70b195f9dd5602d65ca.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 388, + 553, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 553, + 437 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 388, + 553, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(i)" + }, + { + "bbox": [ + 313, + 388, + 553, + 437 + ], + "type": "text", + "content": " denotes the neighboring nodes of " + }, + { + "bbox": [ + 313, + 388, + 553, + 437 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 388, + 553, + 437 + ], + "type": "text", + "content": ". In our implementation, we employ three attention heads and a two-layer structure, enabling the extraction of complex AU dependencies." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 443, + 462, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 443, + 462, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 462, + 456 + ], + "type": "text", + "content": "4.3. Dual-Stream-InceptionNet" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 460, + 554, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 554, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 554, + 617 + ], + "type": "text", + "content": "Muscle movements vary widely across facial regions, making multi-scale information essential for the MER task. InceptionNet's core concept is leveraging parallel convolutional kernels to extract multi-scale features, enhancing the network's ability to capture fine details and overall structure [46]. We build on this by using a dual-stream InceptionNet as the ME feature extractor. One stream processes the AU relationship features " + }, + { + "bbox": [ + 313, + 460, + 554, + 617 + ], + "type": "inline_equation", + "content": "h^{Global}" + }, + { + "bbox": [ + 313, + 460, + 554, + 617 + ], + "type": "text", + "content": ", while the other handles global OF features " + }, + { + "bbox": [ + 313, + 460, + 554, + 617 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 313, + 460, + 554, + 617 + ], + "type": "text", + "content": ". This dual-stream design allows the model to capture the topological relationships of facial muscles through AUs and holistic motion via OF. The output tensors from both streams are flattened, concatenated, and fed into a FC layer for ME prediction." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 624, + 444, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 444, + 637 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 444, + 637 + ], + "type": "text", + "content": "4.4. Federated Aggregation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 641, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 554, + 712 + ], + "type": "text", + "content": "Traditional centralized learning consolidates all data on a single server, enhancing model accuracy and generalization but neglecting privacy concerns in MER. FL could address this by training models locally and sharing only parameters, protecting data privacy and retaining each client's unique data characteristics." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14457" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 189 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 189 + ], + "type": "text", + "content": "FedProx [29] is a FL approach designed to address data heterogeneity among clients. It incorporates proximal terms during model aggregation to balance global and client-local learning, improving both performance and convergence speed. We extend FedProx by altering the server's processing and distribution strategy during aggregation, which we call Personalized FedProx (P-FedProx). Specifically, P-FedProx initializes each client's model " + }, + { + "bbox": [ + 55, + 72, + 296, + 189 + ], + "type": "inline_equation", + "content": "W_{i}^{t}" + }, + { + "bbox": [ + 55, + 72, + 296, + 189 + ], + "type": "text", + "content": " in round " + }, + { + "bbox": [ + 55, + 72, + 296, + 189 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 72, + 296, + 189 + ], + "type": "text", + "content": " by combining its previous model " + }, + { + "bbox": [ + 55, + 72, + 296, + 189 + ], + "type": "inline_equation", + "content": "\\mathring{W}_{i}^{t-1}" + }, + { + "bbox": [ + 55, + 72, + 296, + 189 + ], + "type": "text", + "content": " with models from other clients " + }, + { + "bbox": [ + 55, + 72, + 296, + 189 + ], + "type": "inline_equation", + "content": "\\mathring{W}_{j}^{t-1}" + }, + { + "bbox": [ + 55, + 72, + 296, + 189 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 195, + 296, + 216 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 195, + 296, + 216 + ], + "spans": [ + { + "bbox": [ + 61, + 195, + 296, + 216 + ], + "type": "interline_equation", + "content": "W _ {i} ^ {t} = \\theta \\cdot \\mathring {W} _ {i} ^ {t - 1} + \\sum_ {\\substack {j = 1 \\\\ j \\neq i}} ^ {n} \\omega_ {j} \\cdot \\mathring {W} _ {j} ^ {t - 1}, \\text {where} \\sum_ {\\substack {j = 1 \\\\ j \\neq i}} ^ {n} \\omega_ {j} = 1 - \\theta \\tag{6}", + "image_path": "cc61a8223def414af67e8564e63f05109af1bcab36fb81eb96c9f130a6595739.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 220, + 296, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 220, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 296, + 258 + ], + "type": "text", + "content": "where weight " + }, + { + "bbox": [ + 55, + 220, + 296, + 258 + ], + "type": "inline_equation", + "content": "\\omega_{j}" + }, + { + "bbox": [ + 55, + 220, + 296, + 258 + ], + "type": "text", + "content": " is proportional to " + }, + { + "bbox": [ + 55, + 220, + 296, + 258 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 220, + 296, + 258 + ], + "type": "text", + "content": "th client's data size. For " + }, + { + "bbox": [ + 55, + 220, + 296, + 258 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 220, + 296, + 258 + ], + "type": "text", + "content": "th client, " + }, + { + "bbox": [ + 55, + 220, + 296, + 258 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 220, + 296, + 258 + ], + "type": "text", + "content": " is set to 0.9 (optimal experimental result, see more detail in the Suppl.)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 263, + 144, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 263, + 144, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 263, + 144, + 275 + ], + "type": "text", + "content": "4.5. Loss Function" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 280, + 296, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 280, + 296, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 280, + 296, + 380 + ], + "type": "text", + "content": "AU Prediction Loss: To enhance MER, our network models the topological relationships and statistical patterns of AUs, making AU prediction an auxiliary task. The prediction is utilized in two stages: one enhances the AU feature extraction performance of the AFE module, and the other improves the GAT network's ability to learn AU topological structures. AU recognition is formulated as a multi-label classification problem, with the Binary Cross-Entropy Loss for each AU label defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 384, + 296, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 384, + 296, + 409 + ], + "spans": [ + { + "bbox": [ + 61, + 384, + 296, + 409 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {A}} = \\frac {\\sum_ {j = 1} ^ {M} \\left[ y _ {j} \\cdot \\log \\left(\\sigma \\left(\\hat {y} _ {j}\\right)\\right) + \\left(1 - y _ {j}\\right) \\cdot \\log \\left(1 - \\sigma \\left(\\hat {y} _ {j}\\right)\\right) \\right]}{- M} \\tag {7}", + "image_path": "411051c51721d711ea78a63a6f1a47c2bb61f1cabb0bb3c1e2622d380f02e700.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "text", + "content": " is the total number of AUs, " + }, + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "inline_equation", + "content": "y_{j}" + }, + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "inline_equation", + "content": "\\sigma (\\hat{y}_j)" + }, + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "inline_equation", + "content": "\\hat{y}_j" + }, + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "text", + "content": " represent respectively the ground truth label (0 or 1), the logit output and prediction after applying the sigmoid function for the " + }, + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "text", + "content": "-th AU. The loss function is consistent for both stages, i.e., " + }, + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{A}}^{\\mathrm{AFE}}" + }, + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{A}}^{\\mathrm{GAT}}" + }, + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 475, + 296, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 296, + 509 + ], + "type": "text", + "content": "MER Loss: The Cross-Entropy Loss function is utilized for emotion classification. Thus, our overall MER loss is composed as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 514, + 295, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 514, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 61, + 514, + 295, + 529 + ], + "type": "interline_equation", + "content": "L _ {\\text {M E R}} = \\alpha_ {1} \\cdot \\left(- \\sum_ {i = 1} ^ {N} y _ {i} \\cdot \\log (\\hat {y} _ {i})\\right) + \\alpha_ {2} \\cdot L _ {\\mathrm {A}} ^ {\\mathrm {A F E}} + \\alpha_ {3} \\cdot L _ {\\mathrm {A}} ^ {\\mathrm {G A T}} \\tag {8}", + "image_path": "23339518f469feeeeec4a8c337814d76663afbaca86042447a995e4a08e7453e.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 534, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 594 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 55, + 534, + 295, + 594 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 534, + 295, + 594 + ], + "type": "text", + "content": " represents the number of classes, " + }, + { + "bbox": [ + 55, + 534, + 295, + 594 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 55, + 534, + 295, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 534, + 295, + 594 + ], + "type": "inline_equation", + "content": "\\hat{y}_i" + }, + { + "bbox": [ + 55, + 534, + 295, + 594 + ], + "type": "text", + "content": " are the one-hot encoded ground truth label and the predicted probability for class " + }, + { + "bbox": [ + 55, + 534, + 295, + 594 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 534, + 295, + 594 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 534, + 295, + 594 + ], + "type": "inline_equation", + "content": "\\alpha_{1}" + }, + { + "bbox": [ + 55, + 534, + 295, + 594 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 534, + 295, + 594 + ], + "type": "inline_equation", + "content": "\\alpha_{2}" + }, + { + "bbox": [ + 55, + 534, + 295, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 534, + 295, + 594 + ], + "type": "inline_equation", + "content": "\\alpha_{3}" + }, + { + "bbox": [ + 55, + 534, + 295, + 594 + ], + "type": "text", + "content": " are hyperparameters, set to 0.2, 0.8, and 0.8 to prioritize MER loss while balancing AU recognition (See parameter analysis in Suppl.)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 594, + 296, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 631 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 631 + ], + "type": "text", + "content": "Federated Loss: During federated training, the local client's loss function includes the MER loss " + }, + { + "bbox": [ + 55, + 594, + 296, + 631 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{MER}}" + }, + { + "bbox": [ + 55, + 594, + 296, + 631 + ], + "type": "text", + "content": " and a proximal term:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 637, + 295, + 659 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 637, + 295, + 659 + ], + "spans": [ + { + "bbox": [ + 106, + 637, + 295, + 659 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {F L}} = L _ {\\mathrm {M E R}} + \\frac {\\alpha_ {4}}{2} \\cdot \\left\\| W - W _ {t} \\right\\| ^ {2} \\tag {9}", + "image_path": "5f4693c0ee05d54b13eb6fd5213b55b3bf438a71eff06387d9d078f7d36e6b74.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "W_{t}" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " denotes the global model in the server's " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "-th round of distribution, and " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " represents the current model of the client. " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\alpha_{4}" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " is the hyperparameter for the weight of the proximal term." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 71, + 395, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 395, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 395, + 85 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 97, + 403, + 110 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 97, + 403, + 110 + ], + "spans": [ + { + "bbox": [ + 314, + 97, + 403, + 110 + ], + "type": "text", + "content": "5.1. Configuration" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 121, + 554, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 554, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 554, + 205 + ], + "type": "text", + "content": "Databases: We use two recently published large scale and challenging ME databases: CAS(ME)" + }, + { + "bbox": [ + 313, + 121, + 554, + 205 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 313, + 121, + 554, + 205 + ], + "type": "text", + "content": " and DFME. CAS(ME)" + }, + { + "bbox": [ + 313, + 121, + 554, + 205 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 313, + 121, + 554, + 205 + ], + "type": "text", + "content": "-Part A offers 860 samples from 100 participants. DFME includes three subsets with (participants, samples): A (72,1,118), B (92,969), C (492,5,439). For Apex-first-frame scenario, we compute the OF between the apex and mid-frame towards the offset." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 211, + 554, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 211, + 554, + 259 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 554, + 259 + ], + "type": "text", + "content": "Pre-training: To enhance the model's ability to perceive facial motion patterns, we first pre-trained the AU module (LRM+AFR) on the DISFA [40] and CK [19] datasets. The effect of this pre-training is detailed in the Suppl." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 265, + 555, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 265, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 555, + 384 + ], + "type": "text", + "content": "MER Experiment Settings: The evaluation follows standard protocols on two datasets. On " + }, + { + "bbox": [ + 313, + 265, + 555, + 384 + ], + "type": "inline_equation", + "content": "\\mathrm{CAS(ME)^3}" + }, + { + "bbox": [ + 313, + 265, + 555, + 384 + ], + "type": "text", + "content": ", we perform 3-class classification (positive, negative, surprised) [45] with Leave-One-Subject-Out (LOSO) validation. On DFME, we conduct 7-class classification (Happiness, Surprise, Disgust, Sadness, Anger, Fear, and Contempt) using scale-adapted 10-fold cross-validation [59]. To address class imbalance, performance is measured by the UF1 and UAR. Experiments were implemented in PyTorch 1.13.0 and run on five NVIDIA GeForce RTX 4090 GPUs." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "text", + "content": "Federated Experiment Settings: The DFME and " + }, + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{CAS(ME)^3}" + }, + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "text", + "content": " datasets are randomly split into 5 and 2 local clients, respectively, with each client assigned an equal number of subjects. However, variations in microexpression (ME) counts per subject result in differing ME data distributions and quantities across clients, mirroring real-world data heterogeneity (See detailed data distributions in the Suppl.). Traditional LOSO and k-fold cross-validation can leak training samples in a federated setting, so we employ a random partition strategy with " + }, + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "text", + "content": " of samples for training and " + }, + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "text", + "content": " for testing, averaging results over 10 tests to minimize randomness. Furthermore, we assume an ideal FL scenario with no client dropouts and equal local updates per client. In FedAvg [41] and FedProx [28], the server assigns an identical global model to all clients each round; differently, FedProx adds a proximal term to reduce model bias in local training. Our P-FedProx enhances this by assigning tailored initial models, with " + }, + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "text", + "content": ") weight on each client's own model and " + }, + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "text", + "content": " on others'. Additionally, we compare against recent personalized FL methods: FedRep [5] uses shared representations with personalized heads, ELLP [61] employs parameter decoupling with Fisher Information weighting, and FedAS [55] applies parameter alignment and client synchronization. Unlike these approaches requiring architectural changes or complex procedures, P-FedProx achieves personalization through simple weighted initialization." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14458" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 90, + 295, + 264 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 290, + 82 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 290, + 82 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 290, + 82 + ], + "type": "text", + "content": "Table 2. SOTA methods comparison on DFME and CAS(ME) " + }, + { + "bbox": [ + 61, + 70, + 290, + 82 + ], + "type": "inline_equation", + "content": "{}^{3}" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 90, + 295, + 264 + ], + "lines": [ + { + "bbox": [ + 58, + 90, + 295, + 264 + ], + "spans": [ + { + "bbox": [ + 58, + 90, + 295, + 264 + ], + "type": "table", + "html": "
CategoryMethodYearDFMECAS(ME)3
UF1UARUF1UAR
3D-CNNI3D [2]20170.29230.3058--
R3D [17]20180.21640.2313--
Hand-craftedLBP-TOP [49]20140.23360.2653--
MDMO [37]20150.24890.2939--
Deep LearningOFF-ApexNet [15]20190.23860.2806--
AlexNet [22]2012--0.25700.2634
STSTNet [36]20190.27140.31080.37950.3792
RCN-A [52]20200.27510.31230.39280.3893
MERSiam [58]20210.31840.35320.31840.3532
FGRL [24]20210.07360.14290.33330.2636
FR [24, 60]20220.35590.38140.34930.3413
MMNet [25]20220.26490.27760.37060.3646
BDCNN [4]20220.29750.33140.50500.5164
Micro-BERT [42]2023--0.56040.6125
HTNet [50]20240.32430.33680.57670.5415
HSTA [16]2024--0.5930.618
Ours-0.38530.39780.62210.6226
", + "image_path": "502801362957141a992ddc3c6071f9230c83567e960670e5400bcdd202050902.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 285, + 266, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 266, + 298 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 266, + 298 + ], + "type": "text", + "content": "5.2. Comparison to State-of-the-art Methods" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 302, + 295, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 302, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 302, + 295, + 327 + ], + "type": "text", + "content": "Tab. 2 list the SOTA comparisons on DFME and CAS(ME)" + }, + { + "bbox": [ + 55, + 302, + 295, + 327 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 55, + 302, + 295, + 327 + ], + "type": "text", + "content": ". The confusion matrices are shown in the Suppl." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 327, + 295, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 327, + 295, + 482 + ], + "spans": [ + { + "bbox": [ + 54, + 327, + 295, + 482 + ], + "type": "text", + "content": "Regarding DFME, classifying ME samples into seven categories poses significant challenges, resulting in relatively low UF1 and UAR scores across all methods. However, our proposed method outperforms all others in every metric. Specifically, Our method outperforms I3D by " + }, + { + "bbox": [ + 54, + 327, + 295, + 482 + ], + "type": "inline_equation", + "content": "9.30\\%" + }, + { + "bbox": [ + 54, + 327, + 295, + 482 + ], + "type": "text", + "content": " in UF1 and " + }, + { + "bbox": [ + 54, + 327, + 295, + 482 + ], + "type": "inline_equation", + "content": "9.20\\%" + }, + { + "bbox": [ + 54, + 327, + 295, + 482 + ], + "type": "text", + "content": " in UAR, and surpasses LBP-TOP by " + }, + { + "bbox": [ + 54, + 327, + 295, + 482 + ], + "type": "inline_equation", + "content": "15.17\\%" + }, + { + "bbox": [ + 54, + 327, + 295, + 482 + ], + "type": "text", + "content": " in UF1 and " + }, + { + "bbox": [ + 54, + 327, + 295, + 482 + ], + "type": "inline_equation", + "content": "13.25\\%" + }, + { + "bbox": [ + 54, + 327, + 295, + 482 + ], + "type": "text", + "content": " in UAR, highlighting its effectiveness in extracting critical and discriminative features from ME clips. Additionally, compared to recent deep learning methods like FR, our approach yields " + }, + { + "bbox": [ + 54, + 327, + 295, + 482 + ], + "type": "inline_equation", + "content": "2.94\\%" + }, + { + "bbox": [ + 54, + 327, + 295, + 482 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 327, + 295, + 482 + ], + "type": "inline_equation", + "content": "1.64\\%" + }, + { + "bbox": [ + 54, + 327, + 295, + 482 + ], + "type": "text", + "content": " higher UF1 and UAR, confirming the benefits of incorporating facial topology and structured muscle motion for enhanced performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 483, + 295, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 295, + 675 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 295, + 675 + ], + "type": "text", + "content": "Regarding CAS(ME)" + }, + { + "bbox": [ + 55, + 483, + 295, + 675 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 55, + 483, + 295, + 675 + ], + "type": "text", + "content": ", the primary challenge lies in its sample complexity. Compared to baseline methods (STST-Net, RCN-A, FR), our proposed method achieves over approximately " + }, + { + "bbox": [ + 55, + 483, + 295, + 675 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 55, + 483, + 295, + 675 + ], + "type": "text", + "content": " higher UF1 and UAR scores, marking a significant advancement. Our method surpasses three recent SOTA approaches in MER, including HTNet, which employs a hierarchical Transformer for local feature learning via self-attention. Specifically, our approach outperforms HTNet by " + }, + { + "bbox": [ + 55, + 483, + 295, + 675 + ], + "type": "inline_equation", + "content": "4.54\\%" + }, + { + "bbox": [ + 55, + 483, + 295, + 675 + ], + "type": "text", + "content": " in UF1 and " + }, + { + "bbox": [ + 55, + 483, + 295, + 675 + ], + "type": "inline_equation", + "content": "8.11\\%" + }, + { + "bbox": [ + 55, + 483, + 295, + 675 + ], + "type": "text", + "content": " in UAR, leveraging psychological insights and statistical patterns to capture deeper facial muscle dynamics. Additionally, it exceeds Micro-BERT and HSTA by " + }, + { + "bbox": [ + 55, + 483, + 295, + 675 + ], + "type": "inline_equation", + "content": "6.17\\%" + }, + { + "bbox": [ + 55, + 483, + 295, + 675 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 483, + 295, + 675 + ], + "type": "inline_equation", + "content": "2.91\\%" + }, + { + "bbox": [ + 55, + 483, + 295, + 675 + ], + "type": "text", + "content": " in UF1, and " + }, + { + "bbox": [ + 55, + 483, + 295, + 675 + ], + "type": "inline_equation", + "content": "1.01\\%" + }, + { + "bbox": [ + 55, + 483, + 295, + 675 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 483, + 295, + 675 + ], + "type": "inline_equation", + "content": "0.46\\%" + }, + { + "bbox": [ + 55, + 483, + 295, + 675 + ], + "type": "text", + "content": " in UAR, respectively. By mining facial movement patterns to better understand topology, our model achieves improved metrics, indicating enhanced accuracy in classifying all emotion categories." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 683, + 149, + 696 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 149, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 149, + 696 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 701, + 253, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 253, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 253, + 713 + ], + "type": "text", + "content": "Tab. 3 listed the ablation study results on DFME." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 316, + 101, + 553, + 276 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 555, + 93 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 93 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 93 + ], + "type": "text", + "content": "Table 3. Ablation study. L, G, S and D represent Local (upper and lower face), Global, Single stream and Dual Stream, respectively." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 101, + 553, + 276 + ], + "lines": [ + { + "bbox": [ + 316, + 101, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 316, + 101, + 553, + 276 + ], + "type": "table", + "html": "
SettingLRMAFRInceptionNetUF1↑UAR↑
LFESSELGSD
I✓0.32490.3404
II✓✓0.33620.3517
III✓✓✓0.34800.3624
IV✓✓✓✓0.36290.3723
V✓✓✓✓0.36040.3658
VI✓✓✓✓✓0.38530.3978
SettingWithout AU GroupWith AU GroupUF1↑UAR↑
VII✓0.36830.3793
VIII✓0.38530.3978
SettingPsychological priorData-driven PriorUF1↑UAR↑
IX0.36150.3701
X✓0.37230.3832
XI✓0.37050.3767
XII✓✓0.38530.3978
", + "image_path": "2e3b25314fa1a633b5772fe92f6dcaf267d5d667bcbc6b3699ad73f96961a2fd.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 299, + 555, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 299, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 299, + 555, + 407 + ], + "type": "text", + "content": "Impact of AU Feature: Setting I, which uses only full face OF as input, serves as our baseline for evaluating the role of AU features in MER. In Setting II, integrating AU features into the Dual Stream InceptionNet leads to a " + }, + { + "bbox": [ + 313, + 299, + 555, + 407 + ], + "type": "inline_equation", + "content": "1.13\\%" + }, + { + "bbox": [ + 313, + 299, + 555, + 407 + ], + "type": "text", + "content": " improvement in both UF1 and UAR. This improvement demonstrates that incorporating AU features significantly enhances the model's ability to recognize MEs, and underscores the value of fusing multiple feature types to strengthen model performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 409, + 555, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 409, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 555, + 468 + ], + "type": "text", + "content": "Impact of ROI Relationship Modeling: The results in Setting III show improvements of " + }, + { + "bbox": [ + 313, + 409, + 555, + 468 + ], + "type": "inline_equation", + "content": "1.18\\%" + }, + { + "bbox": [ + 313, + 409, + 555, + 468 + ], + "type": "text", + "content": " in UF1 and " + }, + { + "bbox": [ + 313, + 409, + 555, + 468 + ], + "type": "inline_equation", + "content": "1.07\\%" + }, + { + "bbox": [ + 313, + 409, + 555, + 468 + ], + "type": "text", + "content": " in UAR compared to Setting II, indicating that the model effectively captures complex dependencies between ROIs and extracts more discriminative feature representations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 471, + 554, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 554, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 554, + 530 + ], + "type": "text", + "content": "Impact of AU grouping: We compared AU feature extraction from the global face with our AU grouping strategy (Setting VII vs. VIII), the performance improvement (UF1: " + }, + { + "bbox": [ + 313, + 471, + 554, + 530 + ], + "type": "inline_equation", + "content": "1.70\\%" + }, + { + "bbox": [ + 313, + 471, + 554, + 530 + ], + "type": "text", + "content": ", UAR: " + }, + { + "bbox": [ + 313, + 471, + 554, + 530 + ], + "type": "inline_equation", + "content": "1.85\\%" + }, + { + "bbox": [ + 313, + 471, + 554, + 530 + ], + "type": "text", + "content": "), confirming that AU grouping enables the network to learn more representative AU features." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 532, + 554, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 554, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 554, + 651 + ], + "type": "text", + "content": "Impact of Modeling the Topological Structure of ME Muscle Movements: Compared to Setting III, Setting IV, including local muscle movement topology modeling, improves UF1 and UAR by " + }, + { + "bbox": [ + 313, + 532, + 554, + 651 + ], + "type": "inline_equation", + "content": "1.49\\%" + }, + { + "bbox": [ + 313, + 532, + 554, + 651 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 532, + 554, + 651 + ], + "type": "inline_equation", + "content": "0.99\\%" + }, + { + "bbox": [ + 313, + 532, + 554, + 651 + ], + "type": "text", + "content": ", respectively. Setting V, which adds global muscle movement topology modeling, shows gains of " + }, + { + "bbox": [ + 313, + 532, + 554, + 651 + ], + "type": "inline_equation", + "content": "1.24\\%" + }, + { + "bbox": [ + 313, + 532, + 554, + 651 + ], + "type": "text", + "content": " in UF1 and " + }, + { + "bbox": [ + 313, + 532, + 554, + 651 + ], + "type": "inline_equation", + "content": "0.34\\%" + }, + { + "bbox": [ + 313, + 532, + 554, + 651 + ], + "type": "text", + "content": " in UAR over Setting III. Combining both local and global topology modeling yields the best performance on DFME. These results demonstrate that incorporating topological modeling of muscle movements enhances MER." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": "Impact of Prior Knowledge: Tab. 3 presents the ablation experiments to evaluate the impact of using prior knowledge for constructing facial muscle topology. In particular, compared to Setting IX, Setting X, which uses only psychological knowledge, improves UF1 and UAR by " + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "inline_equation", + "content": "1.08\\%" + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14459" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 102, + 136, + 182 + ], + "blocks": [ + { + "bbox": [ + 69, + 102, + 136, + 182 + ], + "lines": [ + { + "bbox": [ + 69, + 102, + 136, + 182 + ], + "spans": [ + { + "bbox": [ + 69, + 102, + 136, + 182 + ], + "type": "image", + "image_path": "7d0359ab475a41676321769bfdef60e3582e90475e1d72ee6269ed907c5d0965.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 138, + 103, + 204, + 182 + ], + "blocks": [ + { + "bbox": [ + 138, + 103, + 204, + 182 + ], + "lines": [ + { + "bbox": [ + 138, + 103, + 204, + 182 + ], + "spans": [ + { + "bbox": [ + 138, + 103, + 204, + 182 + ], + "type": "image", + "image_path": "a86bcd8e0a2ef30abe66e39830767a2927e9ac24819ae8d569c5be5d4acfe90c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 205, + 103, + 271, + 182 + ], + "blocks": [ + { + "bbox": [ + 205, + 103, + 271, + 182 + ], + "lines": [ + { + "bbox": [ + 205, + 103, + 271, + 182 + ], + "spans": [ + { + "bbox": [ + 205, + 103, + 271, + 182 + ], + "type": "image", + "image_path": "ed619347e1503f8d05331087100313350ff40d9aa809b553fab8d6c994e8edc0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 234, + 185, + 376, + 194 + ], + "lines": [ + { + "bbox": [ + 234, + 185, + 376, + 194 + ], + "spans": [ + { + "bbox": [ + 234, + 185, + 376, + 194 + ], + "type": "text", + "content": "(a) UF1 comparison across different clients." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 272, + 103, + 339, + 182 + ], + "blocks": [ + { + "bbox": [ + 272, + 103, + 339, + 182 + ], + "lines": [ + { + "bbox": [ + 272, + 103, + 339, + 182 + ], + "spans": [ + { + "bbox": [ + 272, + 103, + 339, + 182 + ], + "type": "image", + "image_path": "4dca360b1ef279a4b74913e2b235d6d2581047bd809eaacd2dfa43f486177ef5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 340, + 103, + 406, + 182 + ], + "blocks": [ + { + "bbox": [ + 340, + 103, + 406, + 182 + ], + "lines": [ + { + "bbox": [ + 340, + 103, + 406, + 182 + ], + "spans": [ + { + "bbox": [ + 340, + 103, + 406, + 182 + ], + "type": "image", + "image_path": "4d64020bb307952c01cb4203c06cbd5b28c204536d2809b574d8cc0743ca013d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 406, + 103, + 473, + 182 + ], + "blocks": [ + { + "bbox": [ + 406, + 103, + 473, + 182 + ], + "lines": [ + { + "bbox": [ + 406, + 103, + 473, + 182 + ], + "spans": [ + { + "bbox": [ + 406, + 103, + 473, + 182 + ], + "type": "image", + "image_path": "7ce7a4851e79c5bfbe1887a3d439063c725adae50997589753a075c3fe04ae49.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 473, + 103, + 541, + 182 + ], + "blocks": [ + { + "bbox": [ + 473, + 103, + 541, + 182 + ], + "lines": [ + { + "bbox": [ + 473, + 103, + 541, + 182 + ], + "spans": [ + { + "bbox": [ + 473, + 103, + 541, + 182 + ], + "type": "image", + "image_path": "d60e9d362a4ca3a84d4fb9c4de393ad2b5ae925c765090503f0a69e38f92bcc4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 69, + 196, + 138, + 274 + ], + "blocks": [ + { + "bbox": [ + 69, + 196, + 138, + 274 + ], + "lines": [ + { + "bbox": [ + 69, + 196, + 138, + 274 + ], + "spans": [ + { + "bbox": [ + 69, + 196, + 138, + 274 + ], + "type": "image", + "image_path": "ca2225a778ac85bbd9856be72ef91346a4870f018b736395c5f8e1ce97c1f509.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 296, + 553, + 308 + ], + "lines": [ + { + "bbox": [ + 55, + 296, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 296, + 553, + 308 + ], + "type": "text", + "content": "Figure 5. Compare the performance of different FL frameworks on different clients: UF1 and convergence. See UAR comparison in Suppl." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 138, + 196, + 205, + 274 + ], + "blocks": [ + { + "bbox": [ + 138, + 196, + 205, + 274 + ], + "lines": [ + { + "bbox": [ + 138, + 196, + 205, + 274 + ], + "spans": [ + { + "bbox": [ + 138, + 196, + 205, + 274 + ], + "type": "image", + "image_path": "9758b0502f57eb674eb29a4c4cbace84da198af57d5284b0fb0e007d7bf93c9b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 205, + 196, + 272, + 274 + ], + "blocks": [ + { + "bbox": [ + 205, + 196, + 272, + 274 + ], + "lines": [ + { + "bbox": [ + 205, + 196, + 272, + 274 + ], + "spans": [ + { + "bbox": [ + 205, + 196, + 272, + 274 + ], + "type": "image", + "image_path": "d6e5db3f15daf1da9eeb5b183ec476c91b6fe6759cc2c50eeadc434b1fbd0e13.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 277, + 390, + 286 + ], + "lines": [ + { + "bbox": [ + 220, + 277, + 390, + 286 + ], + "spans": [ + { + "bbox": [ + 220, + 277, + 390, + 286 + ], + "type": "text", + "content": "(b) Convergence comparison across different clients." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 272, + 196, + 339, + 274 + ], + "blocks": [ + { + "bbox": [ + 272, + 196, + 339, + 274 + ], + "lines": [ + { + "bbox": [ + 272, + 196, + 339, + 274 + ], + "spans": [ + { + "bbox": [ + 272, + 196, + 339, + 274 + ], + "type": "image", + "image_path": "164efcb98172282f2bc845b5954f618bf38f57aef2f9040df21f9014b03d4a75.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 340, + 196, + 406, + 274 + ], + "blocks": [ + { + "bbox": [ + 340, + 196, + 406, + 274 + ], + "lines": [ + { + "bbox": [ + 340, + 196, + 406, + 274 + ], + "spans": [ + { + "bbox": [ + 340, + 196, + 406, + 274 + ], + "type": "image", + "image_path": "0278f3eb28b06d6cbecc906342c457ab2c577fc37247aa6ffcea9c6160bebf87.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 406, + 196, + 473, + 274 + ], + "blocks": [ + { + "bbox": [ + 406, + 196, + 473, + 274 + ], + "lines": [ + { + "bbox": [ + 406, + 196, + 473, + 274 + ], + "spans": [ + { + "bbox": [ + 406, + 196, + 473, + 274 + ], + "type": "image", + "image_path": "56f21f9e921cc95ebcf6812103c58b7add37417461d1b73fabeea964d84ffd19.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 473, + 196, + 541, + 274 + ], + "blocks": [ + { + "bbox": [ + 473, + 196, + 541, + 274 + ], + "lines": [ + { + "bbox": [ + 473, + 196, + 541, + 274 + ], + "spans": [ + { + "bbox": [ + 473, + 196, + 541, + 274 + ], + "type": "image", + "image_path": "5d110c145f2ae6a3be10e74af597573fbc31866014983bc5e3e2f04bef4c399e.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 328, + 296, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 328, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 296, + 437 + ], + "type": "inline_equation", + "content": "1.31\\%" + }, + { + "bbox": [ + 55, + 328, + 296, + 437 + ], + "type": "text", + "content": ", respectively. Setting XI, incorporating only data-driven knowledge, shows gains of " + }, + { + "bbox": [ + 55, + 328, + 296, + 437 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 55, + 328, + 296, + 437 + ], + "type": "text", + "content": " in UF1 and " + }, + { + "bbox": [ + 55, + 328, + 296, + 437 + ], + "type": "inline_equation", + "content": "0.66\\%" + }, + { + "bbox": [ + 55, + 328, + 296, + 437 + ], + "type": "text", + "content": " in UAR. When both psychological and data-driven knowledge are combined (Setting XII), the improvements reach " + }, + { + "bbox": [ + 55, + 328, + 296, + 437 + ], + "type": "inline_equation", + "content": "2.38\\%" + }, + { + "bbox": [ + 55, + 328, + 296, + 437 + ], + "type": "text", + "content": " in UF1 and " + }, + { + "bbox": [ + 55, + 328, + 296, + 437 + ], + "type": "inline_equation", + "content": "2.77\\%" + }, + { + "bbox": [ + 55, + 328, + 296, + 437 + ], + "type": "text", + "content": " in UAR, proving the effectiveness of the prior knowledge. Constructing a prior adjacency matrix allows the model to capture complex facial muscle topology, while data-driven knowledge directs attention to uncover intrinsic movements patterns of facial muscles." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 445, + 185, + 457 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 185, + 457 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 185, + 457 + ], + "type": "text", + "content": "5.4. Federated Experiment" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 462, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 296, + 521 + ], + "type": "text", + "content": "Since FL aims to enhance MER performance on individual clients in scenarios with limited sample sizes and data privacy issue, and given that DFME and " + }, + { + "bbox": [ + 55, + 462, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\mathrm{CAS}(\\mathrm{ME})^3" + }, + { + "bbox": [ + 55, + 462, + 296, + 521 + ], + "type": "text", + "content": " have large sample size, we partition the datasets as mentioned in Section 5.1 and treat the subdivisions as separate clients." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "content": "While our focus is on application rather than fundamental FL innovation, our experiments confirm the effectiveness of our P-FedProx method, which is tailored for ME data distributions. Benchmarked against several open-source SOTA methods, our approach excels at mitigating small-sample challenges while preserving data privacy. Specifically (Fig. 5), the results show P-FedProx outperforms both the single-client baseline and traditional FL methods (FedAvg, FedProx), with particularly significant gains on clients with more challenging data distributions (Clients 6-7). Compared to recent personalized FL methods, our approach surpasses FedRep and ELLP and achieves performance competitive with the state-of-the-art FedAS, yet with a significantly simpler implementation. These consistent gains demonstrate that our weighted model initialization effectively addresses the data heterogeneity that challenges" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 329, + 553, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 553, + 353 + ], + "type": "text", + "content": "traditional FL, thereby mitigating ME sample scarcity while ensuring data privacy." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 354, + 554, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 554, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 554, + 413 + ], + "type": "text", + "content": "Meantime, we note that all FL methods achieved good performance on the local clients from DFME, which may be due to the fact that these clients have a similar amount of data and slight differences in data distribution, presenting fewer challenges for federated learning." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 415, + 554, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 554, + 510 + ], + "type": "text", + "content": "Besides, we analyze the training loss progression of different federated algorithms over communication rounds. We found that P-FedProx demonstrates the smoothest convergence curves with minimal oscillations, particularly evident in Clients 1-5. On the other hand, P-FedProx converges significantly faster than other FL methods. This rapid early convergence suggests that personalized global model assignment provides clients with better initialization states." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 314, + 525, + 388, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 525, + 388, + 537 + ], + "spans": [ + { + "bbox": [ + 314, + 525, + 388, + 537 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": "MER is challenged by limited sample sizes, subtle features, and privacy concerns in real-world applications. To address these issues, we conduct a psychological study to provide structured insights into facial muscle dynamics. We then develop a FED-PsyAU framework that integrates these findings with data-driven patterns for hierarchical learning of dynamic features, enhancing MER performance. Additionally, the FL framework improves MER across clients while preserving data privacy through iterative model updates. The comprehensive experiments validate our approach's effectiveness. In the future, we will integrate additional AUs to build a more comprehensive ME topology information and further advance MER in real-world applications through FL and semi-supervised approaches." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 73, + 413, + 87 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 73, + 413, + 87 + ], + "spans": [ + { + "bbox": [ + 197, + 73, + 413, + 87 + ], + "type": "text", + "content": "FedAvg(2017) PersonalizedFedProx ELLP(2024) Baseline FedProx(2020) FedRep(2021) FedAS(2024)" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14460" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 161, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 161, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 161, + 85 + ], + "type": "text", + "content": "7. Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 127 + ], + "type": "text", + "content": "This research was partially funded by 1) the National Natural Science Foundation of China (62476269, 62276252); 2) the Youth Innovation Promotion Association CAS." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 157, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 157, + 296, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 157, + 296, + 212 + ], + "spans": [ + { + "bbox": [ + 61, + 157, + 296, + 212 + ], + "type": "text", + "content": "[1] Xianye Ben, Yi Ren, Junping Zhang, Su-Jing Wang, Kidiyo Kpalma, Weixiao Meng, and Yong-Jin Liu. Video-Based Facial Micro-Expression Analysis: A Survey of Datasets, Features and Algorithms. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):5826-5846, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 213, + 296, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 213, + 296, + 256 + ], + "spans": [ + { + "bbox": [ + 61, + 213, + 296, + 256 + ], + "type": "text", + "content": "[2] Joao Carreira and Andrew Zisserman. Quo Vadis, Action Recognition? A New Model and the Kinetics Dataset. In proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 258, + 296, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 296, + 300 + ], + "type": "text", + "content": "[3] Boyu Chen, Zhihao Zhang, Nian Liu, Yang Tan, Xinyu Liu, and Tong Chen. Spatiotemporal Convolutional Neural Network with Convolutional Block Attention Module for Micro-Expression Recognition. Information, 11(8):380, 2020. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 302, + 296, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 302, + 296, + 356 + ], + "spans": [ + { + "bbox": [ + 62, + 302, + 296, + 356 + ], + "type": "text", + "content": "[4] Bin Chen, Kun-Hong Liu, Yong Xu, Qing-Qiang Wu, and Jun-Feng Yao. Block Division Convolutional Network With Implicit Deep Features Augmentation for Micro-Expression Recognition. IEEE Transactions on Multimedia, 25:1345–1358, 2022. 1, 2, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 357, + 296, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 357, + 296, + 401 + ], + "spans": [ + { + "bbox": [ + 62, + 357, + 296, + 401 + ], + "type": "text", + "content": "[5] Liam Collins, Hamed Hassani, Aryan Mokhtari, and Sanjay Shakkottai. Exploiting shared representations for personalized federated learning. In International conference on machine learning, pages 2089-2099. PMLR, 2021. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 402, + 296, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 402, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 62, + 402, + 296, + 445 + ], + "type": "text", + "content": "[6] Adrian K. Davison, Cliff Lansley, Nicholas Costen, Kevin Tan, and Moi Hoon Yap. SAMM: A Spontaneous MicroFacial Movement Dataset. IEEE Transactions on Affective Computing, 9(1):116-129, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 447, + 296, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 296, + 490 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 296, + 490 + ], + "type": "text", + "content": "[7] Zizhao Dong, Gang Wang, Shaoyuan Lu, Jingting Li, Wenjing Yan, and Su-Jing Wang. Spontaneous Facial Expressions and Micro-Expressions Coding: From brain to Face. Frontiers in Psychology, 12:784834, 2022. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 491, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 296, + 555 + ], + "type": "text", + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 557, + 296, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 557, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 557, + 296, + 578 + ], + "type": "text", + "content": "[9] Paul Ekman and Wallace V Friesen. Nonverbal Leakage and Clues to Deception. Psychiatry, 32(1):88-106, 1969. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 580, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 580, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 57, + 580, + 295, + 612 + ], + "type": "text", + "content": "[10] Paul Ekman and Wallace V Friesen. Facial Action Coding System. Environmental Psychology & Nonverbal Behavior, 1978. 1, 2, 3, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 613, + 295, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 295, + 644 + ], + "type": "text", + "content": "[11] Ali et al. Privacy preserving personalization for video facial expression recognition using federated learning. IMCI'22, . 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 647, + 295, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 680 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 680 + ], + "type": "text", + "content": "[12] Fan et al. Feeling without sharing: A federated video emotion recognition framework via privacy-agnostic hybrid aggregation. ACM MM'23, . 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "type": "text", + "content": "[13] Xinqi Fan, Xueli Chen, Mingjie Jiang, Ali Raza Shahid, and Hong Yan. SelfME: Self-Supervised Motion Learning for Micro-Expression Recognition. In Proceedings of" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13834-13843, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 96, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 173 + ], + "type": "text", + "content": "[14] Viswanatha Reddy Gajjala, Sai Prasanna Teja Reddy, Snehasis Mukherjee, and Shiv Ram Dubey. MERANet: facial micro-expression recognition using 3D residual attention network. In Proceedings of the Twelfth Indian Conference on Computer Vision, Graphics and Image Processing, New York, NY, USA, 2021. Association for Computing Machinery. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "type": "text", + "content": "[15] Yee Siang Gan, Sze-Teng Liong, Wei-Chuen Yau, Yen-Chang Huang, and Lit-Ken Tan. Off-ApexNet on microexpression recognition system. Signal Processing: Image Communication, 74:129-139, 2019. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "type": "text", + "content": "[16] Haihong Hao, Shuo Wang, Huixia Ben, Yanbin Hao, Yansong Wang, and Weiwei Wang. Hierarchical Space-Time Attention for Micro-Expression recognition. arXiv preprint arXiv:2405.03202, 2024. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 264, + 553, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 318 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 318 + ], + "type": "text", + "content": "[17] Kensho Hara, Hirokatsu Kataoka, and Yutaka Satoh. Can Spatiotemporal 3D CNNs Retrace the History of 2D CNNs and Imagenet? In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 6546-6555, 2018. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 319, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 374 + ], + "type": "text", + "content": "[18] Xiaohua Huang, Su-Jing Wang, Guoying Zhao, and Matti Piteikainen. Facial Micro-Expression Recognition Using Spatiotemporal Local Binary Pattern with Integral Projection. In 2015 IEEE International Conference on Computer Vision Workshop (ICCVW), pages 1–9. IEEE, 2015. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 376, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 430 + ], + "type": "text", + "content": "[19] T. Kanade, J.F. Cohn, and Yingli Tian. Comprehensive database for facial expression analysis. In Proceedings Fourth IEEE International Conference on Automatic Face and Gesture Recognition (Cat. No. PR00580), pages 46-53, 2000. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 432, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 432, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 432, + 553, + 464 + ], + "type": "text", + "content": "[20] Davis E King. Dlib-ml: A Machine Learning Toolkit. The Journal of Machine Learning Research, 10:1755-1758, 2009. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 466, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 466, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 466, + 553, + 498 + ], + "type": "text", + "content": "[21] Thomas N Kipf and Max Welling. Semi-Supervised Classification with Graph Convolutional Networks. arXiv preprint arXiv:1609.02907, 2016. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 500, + 553, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 553, + 543 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 553, + 543 + ], + "type": "text", + "content": "[22] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 545, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 578 + ], + "type": "text", + "content": "[23] Sang Gyu Kwak and Jong Hae Kim. Central limit theorem: the cornerstone of modern statistics. Korean journal of anesthesiaology, 70(2):144, 2017. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 580, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 580, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 580, + 553, + 644 + ], + "type": "text", + "content": "[24] Ling Lei, Tong Chen, Shigang Li, and Jianfeng Li. Microexpression Recognition Based on Facial Graph Representation Learning and Facial Action Unit Fusion. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1571-1580. IEEE, 2021. 1, 2, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[25] Hanting Li, Mingzhe Sui, Zhaoqing Zhu, and Feng Zhao. Mmnet: Muscle motion-guided network for microexpression recognition. arXiv preprint arXiv:2201.05297, 2022. 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[26] Jingting Li, Moi Hoon Yap, Wen-Huang Cheng, John See, Xiaopeng Hong, Xiaobai Li, and Su-Jing Wang. FME'21:" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14461" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 116 + ], + "type": "text", + "content": "1st Workshop on Facial Micro-Expression: Advanced Techniques for Facial Expressions Generation and Spotting. In Proceedings of the 29th ACM International Conference on Multimedia, pages 5700-5701. ACM, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 193 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 193 + ], + "type": "text", + "content": "[27] Jingting Li, Zizhao Dong, Shaoyuan Lu, Su-Jing Wang, Wen-Jing Yan, Yinhuan Ma, Ye Liu, Changbing Huang, and Xiaolan Fu. CAS(ME)" + }, + { + "bbox": [ + 56, + 118, + 295, + 193 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 56, + 118, + 295, + 193 + ], + "type": "text", + "content": ": A Third Generation Facial Spontaneous Micro-Expression Database With Depth Information and High Ecological Validity. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):2782-2800, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 194, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 294, + 239 + ], + "type": "text", + "content": "[28] Tian Li, Anit Kumar Sahu, Maziar Sanjabi, Manzil Zaheer, Ameet Talwalkar, and Virginia Smith. On the Convergence of Federated Optimization in Heterogeneous Networks. arXiv preprint arXiv:1812.06127, 2018. 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 239, + 294, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 272 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 272 + ], + "type": "text", + "content": "[29] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks, 2020. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 272, + 294, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 327 + ], + "type": "text", + "content": "[30] Xiaobai Li, Tomas Pfister, Xiaohua Huang, Guoying Zhao, and Matti Pietikainen. A Spontaneous Micro-expression Database: Inducement, collection and baseline. In 2013 10th IEEE International Conference and Workshops on Automatic Face and Gesture Recognition (FG), pages 1-6, 2013. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 392 + ], + "type": "text", + "content": "[31] Xiaobai Li, Xiaopeng Hong, Antti Moilanen, Xiaohua Huang, Tomas Pfister, Guoying Zhao, and Matti Pietikainen. Towards Reading Hidden Emotions: A Comparative Study of Spontaneous Micro-Expression Spotting and Recognition Methods. IEEE Transactions on Affective Computing, 9(4): 563–577, 2018. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 393, + 294, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 294, + 449 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 294, + 449 + ], + "type": "text", + "content": "[32] Xiaobai Li, Shiyang Cheng, Yante Li, Muzammil Behzad, Jie Shen, Stefanos Zafeiriou, Maja Pantic, and Guoying Zhao. 4DME: A Spontaneous 4D Micro-Expression Dataset With Multimodalities. IEEE Transactions on Affective Computing, 14(4):3031-3047, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 449, + 294, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 294, + 482 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 294, + 482 + ], + "type": "text", + "content": "[33] Yante Li, Xiaohua Huang, and Guoying Zhao. Microexpression action unit detection with spatial and channel attention. Neurocomputing, 436:221-231, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 482, + 294, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 482, + 294, + 548 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 294, + 548 + ], + "type": "text", + "content": "[34] Sze-Teng Liong, John See, Raphael C.-W. Phan, Anh Cat Le Ngo, Yee-Hui Oh, and KokSheik Wong. Subtle Expression Recognition Using Optical Strain Weighted Features. In Computer Vision - ACCV 2014 Workshops, pages 644-657. Springer International Publishing, 2015. Series Title: Lecture Notes in Computer Science. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 548, + 294, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 592 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 592 + ], + "type": "text", + "content": "[35] Sze-Teng Liong, John See, KokSheik Wong, and Raphael C.-W. Phan. Less is More: Micro-expression Recognition from Video using Apex Frame. Signal Processing: Image Communication, 62:82-92, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 592, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 657 + ], + "type": "text", + "content": "[36] Sze-Teng Liong, Y. S. Gan, John See, Huai-Qian Khor, and Yen-Chang Huang. Shallow Triple Stream Three-dimensional CNN (STSTNet) for Micro-expression Recognition. In 2019 14th IEEE International Conference on Automatic Face & Gesture Recognition (FG 2019), pages 1-5, 2019. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[37] Yong-Jin Liu, Jin-Kai Zhang, Wen-Jing Yan, Su-Jing Wang, Guoying Zhao, and Xiaolan Fu. A Main Directional Mean Optical Flow Feature for Spontaneous Micro-Expression Recognition. IEEE Transactions on Affective Computing, 7 (4):299–310, 2015. 7" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 714 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 128 + ], + "type": "text", + "content": "[38] Yong-Jin Liu, Jin-Kai Zhang, Wen-Jing Yan, Su-Jing Wang, Guoying Zhao, and Xiaolan Fu. A main directional mean optical flow feature for spontaneous micro-expression recognition. IEEE Transactions on Affective Computing, 7(4):299-310, 2016. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 129, + 554, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 554, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 554, + 185 + ], + "type": "text", + "content": "[39] Ling Lo, Hong-Xia Xie, Hong-Han Shuai, and Wen-Huang Cheng. MER-GCN: Micro-Expression Recognition Based on Relation Modeling with Graph Convolutional Networks. In 2020 IEEE conference on multimedia information processing and retrieval (MIPR), pages 79–84. IEEE, 2020. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 185, + 554, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 554, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 554, + 229 + ], + "type": "text", + "content": "[40] S Mohammad Mavadati, Mohammad H Mahoor, Kevin Bartlett, Philip Trinh, and Jeffrey F Cohn. Disfa: A spontaneous facial action intensity database. IEEE Transactions on Affective Computing, 4(2):151-160, 2013. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 230, + 554, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 554, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 554, + 297 + ], + "type": "text", + "content": "[41] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-Efficient Learning of Deep Networks from Decentralized Data. In Proceedings of the 20th International Conference on Artificial Intelligence and Statistics, pages 1273-1282. PMLR, 2017. ISSN: 2640-3498. 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 297, + 554, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 554, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 554, + 354 + ], + "type": "text", + "content": "[42] Xuan-Bac Nguyen, Chi Nhan Duong, Xin Li, Susan Gauch, Han-Seok Seo, and Khoa Luu. Micron-BERT: BERT-Based Facial Micro-Expression Recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1482–1492, 2023. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 354, + 554, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 554, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 554, + 398 + ], + "type": "text", + "content": "[43] T. Ojala, M. Pietikainen, and T. Maenpaa. Multiresolution gray-scale and rotation invariant texture classification with local binary patterns. IEEE Transactions on Pattern Analysis and Machine Intelligence, 24(7):971-987, 2002. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 399, + 554, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 554, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 554, + 454 + ], + "type": "text", + "content": "[44] Fangbing Qu, Su-Jing Wang, Wen-Jing Yan, He Li, Shuhang Wu, and Xiaolan Fu. CAS(ME)" + }, + { + "bbox": [ + 316, + 399, + 554, + 454 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 316, + 399, + 554, + 454 + ], + "type": "text", + "content": ": A Database for Spontaneous Macro-Expression and Micro-Expression Spotting and Recognition. IEEE Transactions on Affective Computing, 9(4):424-436, 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 456, + 554, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 554, + 511 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 554, + 511 + ], + "type": "text", + "content": "[45] John See, Moi Hoon Yap, Jingting Li, Xiaopeng Hong, and Su-Jing Wang. MEGC 2019 – The Second Facial Micro-Expressions Grand Challenge. In 2019 14th IEEE International Conference on Automatic Face & Gesture Recognition (FG 2019), pages 1–5, 2019. 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 512, + 554, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 554, + 577 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 554, + 577 + ], + "type": "text", + "content": "[46] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE Conference on Computer Cision and Pattern Recognition, pages 1-9, 2015. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 578, + 554, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 554, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 554, + 622 + ], + "type": "text", + "content": "[47] Petar Velicković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph Attention Networks. arXiv preprint arXiv:1710.10903, 2017. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 624, + 554, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 554, + 680 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 554, + 680 + ], + "type": "text", + "content": "[48] Lei Wang, Pinyi Huang, Wangyang Cai, and Xiyao Liu. Micro-expression recognition by fusing action unit detection and spatio-temporal features. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 5595-5599. IEEE, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 681, + 554, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 554, + 714 + ], + "type": "text", + "content": "[49] Su-Jing Wang, Wen-Jing Yan, Xiaobai Li, Guoying Zhao, Chun-Guang Zhou, Xiaolan Fu, Minghao Yang, and Jianhua Tao. Micro-expression recognition using color spaces." + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14462" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 715 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "IEEE Transactions on Image Processing, 24(12):6034-6047, 2015. 1, 2, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "type": "text", + "content": "[50] Zhifeng Wang, Kaihao Zhang, Wenhan Luo, and Ramesh Sankaranarayana. Htnet for micro-expression recognition. Neurocomputing, 602:128196, 2024. 1, 2, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 130, + 295, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 295, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 295, + 196 + ], + "type": "text", + "content": "[51] Wen-Jing Yan, Qi Wu, Yong-Jin Liu, Su-Jing Wang, and Xiaolan Fu. CASME database: A dataset of spontaneous micro-expressions collected from neutralized faces. In 2013 10th IEEE International Conference and Workshops on Automatic Face and Gesture Recognition (FG), pages 1-7. IEEE, 2013. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 295, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 295, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 295, + 252 + ], + "type": "text", + "content": "[52] Zhaoqiang Xia, Wei Peng, Huai-Qian Khor, Xiaoyi Feng, and Guoying Zhao. Revealing the invisible with model and data shrinking for composite-database micro-expression recognition. IEEE Transactions on Image Processing, 29: 8590-8605, 2020. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 295, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 295, + 308 + ], + "type": "text", + "content": "[53] Hong-Xia Xie, Ling Lo, Hong-Han Shuai, and Wen-Huang Cheng. AU-assisted graph attention convolutional network for micro-expression recognition. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2871-2880. ACM, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 310, + 295, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 295, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 295, + 363 + ], + "type": "text", + "content": "[54] Wen-Jing Yan, Xiaobai Li, Su-Jing Wang, Guoying Zhao, Yong-Jin Liu, Yu-Hsin Chen, and Xiaolan Fu. CASME II: An improved spontaneous micro-expression database and the baseline evaluation. PLoS ONE, 9(1):e86041, 2014-01-27. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 295, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 295, + 410 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 295, + 410 + ], + "type": "text", + "content": "[55] Xiyuan Yang, Wenke Huang, and Mang Ye. Fedas: Bridging inconsistency in personalized federated learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11986-11995, 2024. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 295, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 295, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 295, + 465 + ], + "type": "text", + "content": "[56] Moi Hoon Yap, John See, Xiaopeng Hong, and Su-Jing Wang. Facial Micro-Expressions Grand Challenge 2018 Summary. In 2018 13th IEEE International Conference on Automatic Face & Gesture Recognition (FG 2018), pages 675–678, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 468, + 295, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 295, + 533 + ], + "type": "text", + "content": "[57] Zhijun Zhai, Jianhui Zhao, Chengjiang Long, Wenju Xu, Shuangjiang He, and Huijuan Zhao. Feature representation learning with adaptive displacement generation and transformer fusion for micro-expression recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22086-22095, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 534, + 295, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 295, + 588 + ], + "type": "text", + "content": "[58] Sirui Zhao, Hanqing Tao, Yangsong Zhang, Tong Xu, Kun Zhang, Zhongkai Hao, and Enhong Chen. A two-stage 3D CNN based learning method for spontaneous microexpression recognition. Neurocomputing, 448:276-289, 2021. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 590, + 295, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 295, + 644 + ], + "type": "text", + "content": "[59] Sirui Zhao, Huaying Tang, Xinglong Mao, Shifeng Liu, Yiming Zhang, Hao Wang, Tong Xu, and Enhong Chen. DFME: A New Benchmark for Dynamic Facial MicroExpression Recognition. IEEE Transactions on Affective Computing, 15(3):1371-1386, 2024. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 295, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 295, + 691 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 295, + 691 + ], + "type": "text", + "content": "[60] Ling Zhou, Qirong Mao, Xiaohua Huang, Feifei Zhang, and Zhihong Zhang. Feature Refinement: An expression-specific feature learning and fusion method for micro-expression recognition. Pattern Recognition, 122:108275, 2022. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 692, + 295, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 295, + 715 + ], + "type": "text", + "content": "[61] Xu Zhou, Jie Li, Gongjin Lan, Rongrong Ni, Angelo Cangelosi, Jiaxin Wang, and Xiaofeng Liu. Efficient lower" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 334, + 72, + 553, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 334, + 72, + 553, + 106 + ], + "type": "text", + "content": "layers parameter decoupling personalized federated learning method of facial expression recognition for home care robots. Information Fusion, 106:102261, 2024. 6" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14463" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/FEVER-OOD_ Free Energy Vulnerability Elimination for Robust Out-of-Distribution Detection/0c16965c-d6fa-4547-a6ba-ab34512ce472_content_list.json b/2025/FEVER-OOD_ Free Energy Vulnerability Elimination for Robust Out-of-Distribution Detection/0c16965c-d6fa-4547-a6ba-ab34512ce472_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a674287341fc31e0bd42e78a236c51962bc22189 --- /dev/null +++ b/2025/FEVER-OOD_ Free Energy Vulnerability Elimination for Robust Out-of-Distribution Detection/0c16965c-d6fa-4547-a6ba-ab34512ce472_content_list.json @@ -0,0 +1,1554 @@ +[ + { + "type": "text", + "text": "FEVER-OOD: Free Energy Vulnerability Elimination for Robust Out-of-Distribution Detection", + "text_level": 1, + "bbox": [ + 169, + 130, + 828, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Brian K.S. Isaac-Medina1 Maurizio Che2 Yona Faline A. Gaus1 Samet Akcay3 Toby P. Breckon1 \n1Durham University, Durham, UK 2University of Vienna, Vienna, Austria 3Intel", + "bbox": [ + 99, + 184, + 939, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{brian.k.isaac-medina, yona.f.binti-abd-gaus, toby.breckon}@durham.ac.uk mauricio.adrian.che.moguel@univie.ac.at samet.akcay@intel.com", + "bbox": [ + 200, + 223, + 839, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Modern machine learning models, that excel on computer vision tasks such as classification and object detection, are often overconfident in their predictions for Out-of-Distribution (OOD) examples, resulting in unpredictable behaviour for open-set environments. Recent works have demonstrated that the free energy score is an effective measure of uncertainty for OOD detection given its close relationship to the data distribution. However, despite that free energy-based methods representing a significant empirical advance in OOD detection, our theoretical analysis reveals previously unexplored and inherent vulnerabilities within the free energy score formulation such that in-distribution and OOD instances can have distinct feature representations yet identical free energy scores. This phenomenon occurs when the vector representing the feature space difference between the in-distribution and OOD sample belongs to the null space of the last layer of a neural-based classifier. To mitigate these issues, we explore lower-dimensional feature spaces to reduce the null space footprint and introduce novel regularisation to maximize the least singular value of the final linear layer, hence enhancing intersample free energy separation. We refer to these techniques as Free Energy Vulnerability Elimination for Robust Out-of-Distribution Detection (FEVER-OOD). Extensive experimentation shows that FEVER-OOD improves OOD detection in CIFAR-10 with $28.22\\%$ FPR95 and 94.78 AUROC vs $33.66\\%$ FPR95 and 92.15 AUROC of the baseline model, CIFAR-100 with $42.77\\%$ FPR95 and 89.98 AUROC vs $50.96\\%$ FPR95 and 88.06 AUROC of the baseline model, and achieves state of the art OOD detection in Imagenet-100, with average OOD FPR95 of $36.50\\%$ and an AUROC of 92.74 when used with the Dream-OOD model, compared with a $39.33\\%$ and 91.84 AUROC without FEVER-OOD.", + "bbox": [ + 88, + 321, + 485, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 830, + 220, + 845 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Out-of-distribution (OOD) detection (c.f. anomaly/outlier detection), aims to identify abnormal samples that deviate significantly from a given data distribution. Whilst", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/848d7e663e2933ff92e896c599e56815165ef8b69813df492dfc4ac6e02b7a86.jpg", + "image_caption": [ + "Free Energy-based OOD (VOS)" + ], + "image_footnote": [], + "bbox": [ + 516, + 300, + 709, + 414 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b79fe209468046edba6d570b17932480933ce0e0db970c3334cea190f38ac4df.jpg", + "image_caption": [ + "FEVER-OOD (VOS)", + "Figure 1. FEVER-OOD improves baseline Free Energy-based OOD detection methods, where baseline methods predict OOD classes as in-distribution (left, blue), while FEVER-OOD classifies them as OOD (right, green)." + ], + "image_footnote": [], + "bbox": [ + 718, + 300, + 903, + 414 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "OOD is essential in developing deployable machine learning systems, to ensure valid inference only on data drawn from same distribution used in training [24, 36, 44], real-world deployments often involve scenarios where models encounter unseen OOD classes that could otherwise undermine decision stability and hence wider system robustness.", + "bbox": [ + 511, + 487, + 906, + 579 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Methods such as deep ensembles [23], ODIN [27], Mahalanobis distance-based detection [25], and generalized ODIN [16] all utilize softmax() to differentiate indistribution from OOD. Whilst intuitive, such approaches often struggle as deep neural networks commonly yield overly high softmax() scores, even for inputs far from the training data distribution [31]. Given OOD and indistribution samples should have significant spatial separation in any given feature space, Liu et al. [29] developed an alternative scoring function to effectively measure the alignment of a given sample to an a priori learned distribution. Denoted as the free energy scoring function, this formulation demonstrates superior OOD performance across both classification [7-9] and detection tasks [6, 7, 18, 21].", + "bbox": [ + 511, + 580, + 908, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite free energy-based methods representing a significant empirical advance in OOD detection, our theoretical analysis reveals previously unexplored and inherent vulnerabilities within the free energy score formulation that could impact their practical reliability. These vulnerabilities reveal that in-distribution and OOD instances can have distinct feature representations yet produce identical free en", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4529", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ergy scores. This phenomenon occurs in the feature space when the vector direction representing the difference between an in-distribution and an OOD sample lies within the null space of the last layer of neural-based classifier. Such regions, effectively \"blind spots\" in terms of OOD, lead to OOD and in-distribution generating identical or near-identical free energy score values.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To mitigate these issues, we explore lower-dimensional feature spaces to effectively reduce the size of the null space. In addition, we also introduce a novel regulariser to maximize the least singular value of the final linear layer, hence enhancing inter-sample free energy separation to improve OOD robustness. We refer to these techniques as Free energy Vulnerability Elimination for Robust Out-of-Distribution Detection (FEVER-OOD).", + "bbox": [ + 89, + 198, + 480, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our key contributions in this paper are:", + "bbox": [ + 89, + 321, + 431, + 335 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The identification of inherent free energy score vulnerabilities, where under null-space conditions, in-distribution and OOD instances will yield similar free energy scores despite differing feature representations. Denoted as Null Space Vulnerabilities (NSV) such conditions readily occur when the feature dimension exceeds the number of classes. Additionally, such free energy similarity can be influenced by the least singular vector (LSV) of the last linear layer of a neural-based classifier, leading to what we denote as Least Singular Value Vulnerabilities (LSVV). Both vulnerabilities negatively impact in-distribution to OOD discrimination.", + "- Novel formulations that mitigate the NSV and LSVV impact (Fig. 1). To address NSV, we introduce an additional layer to reduce the null space whilst for LSVV, we develop an LSV Regulariser that maximizes the least singular value, guaranteeing detectable energy variations from small feature-space differences and hence enhancing inter-sample energy scores discrimination. Furthermore, we explore a Condition Number Regulariser to promote balanced energy changes for displacements in all feature-space direction hence ensuring improved uniformity in energy score variations.", + "- Comprehensive experiments and ablation studies using established benchmark datasets over both object classification and detection tasks, that demonstrate the effectiveness of FEVER-OOD in improving OOD detection performance. Specifically, Dream-OOD [8] with FEVER-OOD trained on Imagenet-100 [35] as in-distribution decreases the FPR95 from $39.33\\%$ to $36.50\\%$ and increases the average AUROC from 91.84 to 92.74, achieving state-of-the-art performance for OOD detection." + ], + "bbox": [ + 89, + 337, + 480, + 820 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 832, + 232, + 848 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Identifying OOD samples is a crucial yet inherently challenging problem, as models are not exposed to, and hence cannot reliably differentiate, OOD samples.", + "bbox": [ + 89, + 859, + 480, + 905 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "OOD Detection: to overcome this challenge, researchers have begun incorporating additional data during training to establish a more conservative and safe decision boundary against OOD inputs. Some methods [1, 19] have used unlabeled data to regularise model training while still focusing on classifying labeled in-distribution data, while others [2] leverage human assistance to strategically label examples within a novel maximum disambiguation region. Despite its potential, leveraging unlabeled data is non-trivial due to the heterogeneous mix of in-distribution and OOD instances. The absence of a clean OOD set presents significant challenges in designing effective OOD learning algorithms. A different branch of works incorporate post-hoc methods, focusing on designing OOD scoring mechanisms at test time. Hendrycks and Gimpel [15] introduced the maximum predicted softmax() probability as an in-distribution score function, establishing an initial baseline for scoring OOD detection. Similarly, ODIN [27] aims to widen the gap between values in the softmax() score vectors by temperature scaling and gradient-based input perturbation, enhancing the separation between in-distribution and OOD instances. Lee et al. [25] use the Mahalanobis distance between the average softmax() vectors of images and incorporate intermediate features to improve detection performance and robustness. Nonetheless, the misalignment between the softmax() score and the true data density [29, 32] ultimately makes these techniques a suboptimal solution to OOD detection. Closer to our work, Cook et al. [4] identify null space adversarial attacks, proposing that instances with large magnitude of their projection onto the span of the null space of a neural network classifier would likely be an outlier. Similarly, Wang et al. [41] propose that since the features lie in low-dimensional manifolds, outliers are likely to deviate from the principal subspace, i.e. the span of the eigenvectors or the largest eigenvalues, of the in-distribution data. Therefore, OOD instances are detected with respect to the magnitude of the component of the perpendicular subspace to the principal subspace.", + "bbox": [ + 511, + 90, + 906, + 666 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Energy-based OOD Detection: to address these limitations, Liu et al. [29] propose to use a free energy-based in-distribution score function instead of the softmax() score to identify OOD instances. This approach regularises the model to produce lower energy values for in-distribution and higher energy values for auxiliary outliers, creating a significant energy gap that enhances the model ability to detect OOD samples. Energy-based OOD modeling has enabled novel outlier synthesis paradigms. For instance, Du et al. [7] propose virtual outlier synthesis (VOS) by sampling outliers from low-likelihood in-distribution regions. VOS has shown to be effective for classification and object-level OOD detection. Further outlier synthesis techniques based on VOS have been proposed, including using normalizing flows [21], sampling based on the in-distribution boundary", + "bbox": [ + 511, + 674, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4530", + "bbox": [ + 480, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "instances [39] or extending VOS when class labels are unavailable [18]. In addition, VOS techniques arensuitable for object detection [7, 21]. Whilst VOS and subsequent works synthesize outliers in the feature space, Dream-OOD [8] uses Stable Diffusion [34] to generate outliers in the pixel-space, learning better discriminative free-score values for OOD instances and providing visual examples of outliers. Energy-based methods have shown great performance, becoming state-of-the-art in several benchmarks (e.g., Dream-OOD [8] for ImageNet-100 in-distribution), hence we focus in the properties of these methods. Notwithstanding that the core idea of the energy score is that OOD samples must have a significantly different free energy than in-distribution sample, we examine some inherent vulnerabilities in the energy formulation based on the null space and least singular value of the last linear layer of neural-based classifiers.", + "bbox": [ + 89, + 90, + 483, + 332 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries: Free Energy Score", + "text_level": 1, + "bbox": [ + 89, + 338, + 395, + 356 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The goal of OOD detection is to detect outliers, $\\mathbf{v}$ , that have a low probability in the true (unknown) in-distribution density $p(\\mathbf{x})$ . However, the computation of $p(\\mathbf{x})$ is usually unfeasible or computationally untractable, so we instead resort to additional information. Now consider that we also have a set of in-distribution categories $\\{1,\\dots ,K\\}$ such that if we know the joint probability distribution, $p(\\mathbf{x},k)$ , for $k\\in \\{1,\\ldots ,K\\}$ , we could marginalize $p(\\mathbf{x})$ . Nonetheless, since the actual joint probabilities $p(\\mathbf{x},k)$ might be also untractable, we could instead use non-negative arbitrary potential functions $\\phi (\\mathbf{x},k)$ that measure the affinity of the states of $\\mathbf{x}$ and $k$ [12]. In this sense, we can construct the unnormalized probability:", + "bbox": [ + 89, + 363, + 483, + 561 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {p} (\\mathbf {x}) = \\sum_ {k = 1} ^ {K} \\phi (\\mathbf {x}, k) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 569, + 482, + 611 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "that gives an idea of the likelihood $\\mathbf{x}$ . Furthermore, energy-based models enforce the non-negative constraint of $\\phi$ by:", + "bbox": [ + 89, + 619, + 482, + 650 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\phi (\\mathbf {x}, k) = \\exp (- E (\\mathbf {x}, k)), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 659, + 482, + 676 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $E$ is an unbounded arbitrary function called the energy. By substituting Eq. (2) into Eq. (1) and taking the negative $\\log^1$ , we get a value known as the free energy:", + "bbox": [ + 89, + 685, + 483, + 733 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} (\\mathbf {x}) = - \\log \\tilde {p} (\\mathbf {x}) = - \\log \\sum_ {k = 1} ^ {K} \\exp (- E (\\mathbf {x}, k)), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 741, + 482, + 781 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and more recently used as a concept in OOD detection [7, 8, 18, 21, 29]. To get the actual density $p(\\mathbf{x})$ , we must compute the normalizing partition function of Eq. (1); since we are interested in distinguishing between in-distribution and OOD, $\\mathcal{F}(\\mathbf{x})$ thus serves well for this purpose.", + "bbox": [ + 89, + 790, + 483, + 866 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/bdd060b5b3baf52a609a84f74c4399fcafd9d4e86650030073fdec620e75a556.jpg", + "image_caption": [ + "Figure 2. Vulnerabilities of Free Energy-based OOD detection. Directions in the Null Space do not change the energy, while the LSV direction has minimal change, compared to other directions." + ], + "image_footnote": [], + "bbox": [ + 514, + 87, + 903, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In a general machine learning context, a classifier $f(\\mathbf{x};\\theta):\\mathbb{R}^d\\to \\mathbb{R}^K$ with learned parameters $\\theta$ assigns categorical unnormalized probabilities $f_{k}(\\mathbf{x};\\theta)$ , called logits, for a given $\\mathbf{x}\\in \\mathbb{R}^d$ . While the logits are usually normalized using the softmax() function in order to obtain a probability distribution, they can also be used as the energy $E$ in Eq. (2). This realization enables the use of the free energy (Eq. (3)) as a measure of uncertainty for a classifier:", + "bbox": [ + 511, + 287, + 906, + 409 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} (\\mathbf {x}) = - \\log \\sum_ {k = 1} ^ {K} \\exp \\left(f _ {k} (\\mathbf {x}; \\theta)\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 420, + 903, + 462 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Liu et al. [29] demonstrate that Eq. (4) is superior for effective OOD detection when compared to use of softmax() whilst Du et al. [7] explicitly use the free energy score (Eq. (4)) to perform binary classification between indistribution and OOD by training with synthetic outliers.", + "bbox": [ + 511, + 472, + 905, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Vulnerabilities of the Free Energy Score", + "text_level": 1, + "bbox": [ + 511, + 561, + 872, + 579 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although the free energy (Eq. (4)) is an effective uncertainty measure for OOD detection [7, 29], here we examine some inherent vulnerabilities when used in conjunction with a neural network based classifier. Firstly, Sec. 4.1 investigates OOD instances that have similar energies as in-distribution samples, whilst Sec. 4.2 discusses minimal changes in free energy, and finally Sec. 4.3 outlines our proposed approach to minimize the impact of these vulnerabilities.", + "bbox": [ + 511, + 587, + 905, + 708 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Null Space Vulnerability", + "text_level": 1, + "bbox": [ + 511, + 717, + 738, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Consider a neural network-based classifier $f = (g \\circ h)(\\mathbf{x})$ where $h(\\mathbf{x}): \\mathbb{R}^d \\to \\mathbb{R}^{d'}$ encodes the input into a $d'$ -dimensional feature vector and $g: \\mathbb{R}^{d'} \\to \\mathbb{R}^K$ is the last layer of the neural network that transforms the features into unnormalized class probabilities (i.e., the logits). If $g$ is a linear layer (as for most classifiers [5, 14, 43]), then:", + "bbox": [ + 511, + 739, + 905, + 830 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf (\\mathbf {x}) = \\mathrm {W} _ {c l s} ^ {\\top} h (\\mathbf {x}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 642, + 840, + 903, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathrm{W}_{cls} \\in \\mathbb{R}^{d' \\times K}$ are learnable weights. Hence, the logits $f_{k}(\\mathbf{x};\\theta)$ correspond to the $k$ -th index of $f(\\mathbf{x})$ .", + "bbox": [ + 511, + 869, + 903, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "$^{1}$ N.B. the signs of the free energy and the argument of the exponent of $\\phi$ are arbitrarily negative for historical reasons.", + "bbox": [ + 89, + 875, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4531", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The core idea of energy-based OOD is that outliers $\\mathbf{v}$ have a significantly different free energy, $\\mathcal{F}(\\mathbf{v})$ , from in-distribution samples, $\\mathcal{F}(\\mathbf{x})$ . Since OOD instances should be far from in-distribution samples in the feature space [7, 39], the expected distance from the inlier features $h(\\mathbf{x})$ and outlier features $h(\\mathbf{v})$ should be greater than a given boundary distance, $d_b$ , as follows:", + "bbox": [ + 89, + 90, + 483, + 196 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} \\left[ \\| h (\\mathbf {x}) - h (\\mathbf {v}) \\| \\right] \\geq d _ {b}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 205, + 483, + 224 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for a norm $\\|\\cdot\\|$ . In other words, the inliers remain bounded to a region in the feature space while outliers are unbounded. Therefore, the expected distance between inliers and outliers is greater than a bounding distance $d_b$ . Subsequently, we can rewrite Eq. (6) as:", + "bbox": [ + 89, + 232, + 483, + 309 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nh (\\mathbf {v}) = h (\\mathbf {x}) + \\boldsymbol {\\delta}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 222, + 318, + 482, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\delta$ is a vector with $\\mathbb{E}[\\|\\delta\\|] \\geq d_b$ . By substituting Eq. (7) into Eq. (5) we obtain the following:", + "bbox": [ + 89, + 344, + 483, + 377 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf (\\mathbf {v}) = \\mathrm {W} _ {c l s} ^ {\\top} (h (\\mathbf {x}) + \\boldsymbol {\\delta}) = f (\\mathbf {x}) + \\mathrm {W} _ {c l s} ^ {\\top} \\boldsymbol {\\delta}. \\qquad (8)\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 383, + 483, + 404 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However within this formulation, if $\\mathrm{W}_{cls}$ has more rows than columns, which occurs when the feature dimension is larger than the number of classes, the size of the null space of $\\mathrm{W}_{cls}^{\\top}$ , called the nullity, is nullity $(\\mathrm{W}_{cls}^{\\top}) = d' - \\operatorname{rank}(\\mathrm{W}_{cls})$ due to the rank-nullity theorem. Mathematically, this means that there exists a linear subspace $\\mathrm{Null}(\\mathrm{W}_{cls}^{\\top}) \\subset \\mathbb{R}^{d'}$ such that if $\\delta \\in \\mathrm{Null}(\\mathrm{W}_{cls}^{\\top})$ then $\\mathrm{W}_{cls}^{\\top} \\delta = 0$ . Within our OOD context, if an outlier instance, $\\mathbf{v}$ , has a feature representation, $h(\\mathbf{v})$ , such that its difference with respect to the inlier feature representation, $h(\\mathbf{x})$ , of any inlier instance, $\\mathbf{x}$ , is in the null space of the last layer of the classifier, then their free energy scores are the same:", + "bbox": [ + 89, + 411, + 483, + 606 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nh (\\mathbf {v}) - h (\\mathbf {x}) \\in \\operatorname {N u l l} \\left(\\mathrm {W} _ {c l s} ^ {\\top}\\right) \\Rightarrow \\mathcal {F} (\\mathbf {x}) = \\mathcal {F} (\\mathbf {v}). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 616, + 482, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Practically, Eq. (9) is telling us that it is possible for in-distribution and OOD instances to have different feature representations (satisfying the condition of Eq. (6)) and yet have the same free energy score, as depicted in Fig. 2. This fundamentally challenges the notion that free energy is an effective mechanism for OOD detection as any approaches reliant on Eq. (3), such as [7, 8, 18, 21, 29], contain inherent \"detection blind-spots\" as per Eq. (9). We denote Eq. (9) as the Null Space Vulnerabilities (NSV) of classifier $f$ with size nullity $(\\mathrm{W}_{cls}^{\\top}) = d' - \\mathrm{rank}(\\mathrm{W}_{cls})$ . An approach to minimize this vulnerability is outlined in Sec. 4.3.", + "bbox": [ + 89, + 643, + 483, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As an additional remark, NSV are inherent in energy-based methods; while some post hoc methods use the projection of features onto the null space as a discriminator between in-distribution and OOD [4, 41], energy-based methods use the classification logits for detecting outliers, hence the null space projection does not change the OOD score.", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Minimal Change of Free Energy", + "text_level": 1, + "bbox": [ + 511, + 90, + 799, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We now turn our attention to the case where $\\delta$ is not in the null space of $\\mathbf{W}_{cls}^{\\top}$ . Furthermore, we can focus on $\\delta$ orthogonal to $\\mathrm{Null}(\\mathbf{W}_{cls}^{\\top})$ , since every vector in the domain of $\\mathbf{W}_{cls}^{\\top}$ can be decomposed into one component in $\\mathrm{Null}(\\mathbf{W}_{cls}^{\\top})$ and one component in $\\mathrm{Null}(\\mathbf{W}_{cls}^{\\top})^{\\perp}$ , and the effect of the component in $\\mathrm{Null}(\\mathbf{W}_{cls}^{\\top})$ will be dealt with using the null space reduction method (Sec. 4.3).", + "bbox": [ + 511, + 112, + 906, + 218 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to create more discriminative OOD detectors based on the free energy score, we investigate cases where the difference in OOD and in-distribution free energy is minimal. We can state our problem formulation to find the case that minimizes the free energy difference:", + "bbox": [ + 511, + 219, + 906, + 295 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {\\substack {\\boldsymbol {\\delta}: \\| \\boldsymbol {\\delta} \\| \\geq d _ {b}, \\\\ \\boldsymbol {\\delta} \\in \\operatorname {Null} \\left(\\mathrm {W} _ {c l s} ^ {\\top}\\right) ^ {\\perp}}} | \\mathcal {F} (\\mathbf {v}) - \\mathcal {F} (\\mathbf {x}) | \\\\ = \\min _ {\\substack {\\boldsymbol {\\delta}: \\| \\boldsymbol {\\delta} \\| \\geq d _ {b}, \\\\ \\boldsymbol {\\delta} \\in \\operatorname {Null} \\left(\\mathrm {W} _ {c l s} ^ {\\top}\\right) ^ {\\perp}}} \\left| \\log \\frac {\\sum_ {k} ^ {K} \\exp \\left(f _ {k} (h (\\mathbf {x}))\\right)}{\\sum_ {k} ^ {K} \\exp \\left(f _ {k} (h (\\mathbf {x}) + \\boldsymbol {\\delta})\\right)} \\right|. \\end{array} \\tag{10}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 305, + 905, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Whilst in Eq. (10) we set $\\| \\delta \\| \\geq d_b$ instead of $\\mathbb{E}\\left[\\| \\delta \\| \\right]\\geq$ $d_{b}$ , in reality we are just interested in the case of non-zero $\\delta$ since otherwise the solution would be trivial. Whilst the solution to Eq. (10) may be challenging to obtain, we note the difference in free-energies tends to 0 if the value inside the $\\log ()$ function tends to 1. This is equivalent to making $f_{k}(h(\\mathbf{x}))$ and $f_{k}(h(\\mathbf{x}) + \\delta)$ similar for all $k$ . Therefore, considering that $f_{k}$ is in the form of Eq. (5), our problem consists on finding the following proxy minimum:", + "bbox": [ + 511, + 404, + 906, + 542 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{r l} \\underset {\\delta : \\| \\boldsymbol {\\delta} \\| \\geq d _ {b},} {\\min } & \\| \\mathrm {W} _ {c l s} ^ {\\top} (h (\\mathbf {x}) + \\boldsymbol {\\delta}) - \\mathrm {W} _ {c l s} ^ {\\top} h (\\mathbf {x}) \\| \\\\ & = \\underset {\\substack {\\boldsymbol {\\delta}: \\| \\boldsymbol {\\delta} \\| \\geq d _ {b}, \\\\ \\boldsymbol {\\delta} \\in \\operatorname {Null} \\left(\\mathrm {W} _ {c l s} ^ {\\top}\\right) ^ {\\perp}}} {\\min } \\| \\mathrm {W} _ {c l s} ^ {\\top} \\boldsymbol {\\delta} \\|. \\end{array} \\tag{11}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 551, + 905, + 631 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the $L^2$ -norm in Eq. (11), we thus obtain:", + "bbox": [ + 511, + 643, + 810, + 659 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\substack {\\boldsymbol {\\delta}: \\| \\boldsymbol {\\delta} \\| _ {2} \\geq d _ {b}, \\\\ \\boldsymbol {\\delta} \\in \\operatorname {Null} \\left(\\mathrm {W} _ {c l s} ^ {\\top}\\right) ^ {\\perp}}} \\| \\mathrm {W} _ {c l s} ^ {\\top} \\boldsymbol {\\delta} \\| _ {2} = d _ {b} \\sigma_ {m i n} \\left(\\mathrm {W} _ {c l s}\\right), \\tag{12}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 669, + 906, + 707 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\sigma_{\\min}(\\mathrm{W}_{cls})$ is the least singular value of $\\mathrm{W}_{cls}$ (see further mathematical formulation details in Supp. Mat.).", + "bbox": [ + 511, + 719, + 905, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Subsequently, Eq. (12) holds for the case where $\\delta$ is in the direction corresponding to the least singular value, such that the change in the free energy will be smaller than any other direction for $\\delta$ (and considering $\\delta \\in \\mathrm{Null}(\\mathrm{W}_{cls}^{\\top})^{\\perp}$ ). We name this result (from Eq. (12)) the Least Singular Value Vulnerabilities (LSVV) of $f$ . In our OOD context, it means that there could exist some outliers far from the inlier distribution, in terms of Eq. (6), with similar free energy scores because they may fall in the direction of the least singular value of $\\mathrm{W}_{cls}$ . Again, this fundamentally challenges", + "bbox": [ + 511, + 750, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4532", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the notion that free energy is an effective mechanism for OOD as free energy has been shown to be non-linear with respect to a given feature representation, $h(\\cdot)$ .", + "bbox": [ + 89, + 90, + 482, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This analysis is itself inspired by the seminal adversarial attacks analysis of Simon-Gabriel et al. [37], where they instead maximize the change in the classifier response for small input perturbations. Whilst here we are modelling the opposite problem, (i.e., the minimum change in the classifier for perturbations larger than a non-zero value), we could consider this analysis as an effective adversarial attack mechanism for the OOD domain.", + "bbox": [ + 89, + 137, + 483, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Null Space Reduction & Regularisation", + "text_level": 1, + "bbox": [ + 89, + 267, + 429, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we introduce our Free Energy Vulnerability Elimination for Robust OOD Detection (FEVER-OOD) framework, where we explore readily available methods to minimize the NSV (Sec. 4.1) and LSVV (Sec. 4.2) impact potential within free energy score based OOD approaches.", + "bbox": [ + 89, + 289, + 483, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Null Space Reduction (NSR). For NSV, we propose to reduce the null space by adding an extra linear layer $g' : \\mathbb{R}^{d'} \\to \\mathbb{R}^r$ , where we choose $r < d'$ to reduce the size of the NSV (Fig. 3). We denote this method as $r$ -null space reduction ( $r$ -NSR). Therefore, the new classifier becomes $f' = g \\circ g' \\circ h$ such that the feature representation of a sample is in $\\mathbb{R}^r$ and is given instead by $(g' \\circ h)(\\mathbf{x})$ . Effectively, within an OOD context, we are promoting better discrimination between OOD and in-distribution instances.", + "bbox": [ + 89, + 366, + 483, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Least Singular Value Regulariser (LSVR). To minimize the impact of LSVV, we maximise $\\sigma_{min}(\\mathrm{W}_{cls})$ by adding an LSVR to the loss function $\\mathcal{L}$ of a baseline method:", + "bbox": [ + 89, + 501, + 483, + 546 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {L S V} := \\mathcal {L} + \\lambda_ {L S V} \\sigma_ {\\min } ^ {- 1} \\left(\\mathrm {W} _ {c l s}\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 558, + 482, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{LSV}$ is a scalar hyperparameter that controls the contribution of the LSV regulariser. It is additionally noted that alternatively maximizing the LSV might also increase the rest of the singular values, resulting in more significant changes of energy in all directions. In the context of OOD detection, LSVR increases the discriminative capacity for OOD detection of baseline methods.", + "bbox": [ + 89, + 587, + 483, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Condition Number (CN) Regulariser. Whilst the LSV regulariser will increase the change in energy for $\\delta$ in the LSV direction, there is no constraint with regard to the distribution over all other directions. For this reason, we explore a Condition Number Regulariser (CNR):", + "bbox": [ + 89, + 693, + 483, + 768 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C N} := \\mathcal {L} + \\lambda_ {C N} \\kappa \\left(\\mathrm {W} _ {\\text {c l s}}\\right), \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 781, + 482, + 797 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\kappa (\\mathrm{W}_{cls})$ is the condition number of the WCLS matrix, i.e., the ratio of the greatest and least singular values. Intuitively, this will encourage equalized energy changes for displacements in all directions where we note that Eq. (14) similarly assumes use of the $L^2$ -norm as per Eq. (12). A different approach to the CNR is the orthogonal constraint", + "bbox": [ + 89, + 809, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9ce8f5922b4e4211f3267ae349a5be11a44052d8378b0f5c650be308d1177588.jpg", + "image_caption": [ + "Figure 3. Null space reduction. We add an extra layer $g'$ to create $r$ -dimensional features, reducing the size of the null space." + ], + "image_footnote": [], + "bbox": [ + 537, + 88, + 885, + 299 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\| W_{cls}W_{cls}^{\\top} - I\\|$ . However, this method adds the additional constraint that the singular values must be close to one, which might be a stronger prior than the CNR, hence we opt for the latter. The gradients for the singular values in Eqs. (13) and (14) are detailed by Lewis and Sendov [26] (Proposition 6.2) and further discussed in the Supp. Mat.", + "bbox": [ + 511, + 340, + 905, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experimental Details", + "text_level": 1, + "bbox": [ + 513, + 443, + 715, + 459 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We study the impact of FEVER-OOD on free energy baseline methods (Sec. 5.1) across different standard datasets (Sec. 5.2). Implementation details are given in Sec. 5.3.", + "bbox": [ + 511, + 469, + 905, + 515 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Baseline Methods and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 511, + 529, + 870, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Three free energy baseline methods for image OOD detection that use the free energy score are tested: VOS [7], FFS [21] and Dream-OOD [8]. These baseline methods learn a small MLP $\\psi : \\mathbb{R} \\to (0,1)$ that uses the free energy in Eq. (4) to predict an uncertainty score. Since OOD data is unavailable during training, these methods use synthetic outliers to train $\\psi$ . While VOS and FFS synthesize outliers in the feature space $h(\\mathbf{x})$ , Dream-OOD generates outliers in the pixel space using Stable Diffusion [34]. Additionally, VOS and FFS are also implemented as object detectors for object-level OOD detection, synthesizing outliers before the last layer of the classification head of a Faster RCNN [33].", + "bbox": [ + 511, + 551, + 905, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We follow standard evaluation metrics, reporting the false positive rate for OOD detection at $95\\%$ true positive rate of in-distribution detection (FPR95) and the area under the receiver operating characteristic curve (AUROC). We also assess the effect of our proposed techniques into the main in-distribution task (accuracy for classification and MS-COCO [28] mean average precision (mAP) for object detection). FPR95 and AUROC metrics are reported using a single intersection-over-union threshold for object detection, corresponding to the one that maximises the F-1 score, as described by Harakeh and Waslander [13].", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4533", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ad272b348d5f821114dc6e272341dc498cbe59becddba24f00ec480376c0be72.jpg", + "table_caption": [ + "Table 1. CIFAR-10 Results (ID Acc. = in-distribution accuracy; Null Space Reduction (NSR) methods = our approach)." + ], + "table_footnote": [], + "table_body": "
MethodFEVER-OODOOD DatasetsID Acc
TexturesSHVNPlaces365LSUNiSUNAvg
r-NSRRegularizerFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROC
VOS--50.05±5.7086.79±1.2839.15±10.5291.73±2.8840.88±1.7589.54±0.698.22±1.0298.36±0.2030.01±4.4494.31±0.8433.66±2.2192.15±0.5994.83±0.16
-λLSV=1.055.58±6.3586.71±2.3543.00±3.1692.09±0.9438.78±2.5191.14±0.567.36±1.5998.46±0.3221.23±6.9596.23±1.1733.19±3.2092.92±0.8594.66±0.18
-λCN=0.150.80±3.0685.39±0.7930.78±2.9594.04±0.5841.60±1.4088.77±0.748.50±0.8298.30±0.1030.91±4.5493.41±1.6832.52±1.2591.98±0.4694.69±0.16
96λLSV=1.040.86±4.8392.12±1.0541.92±16.1193.20±2.6135.69±2.2992.61±0.574.95±1.4998.93±0.2417.68±5.2697.04±0.8528.22±4.1894.78±0.8594.74±0.17
64λLSV=1.044.41±5.3991.26±1.2934.68±10.9394.39±1.7235.85±2.5692.42±0.535.68±0.5798.75±0.1324.03±5.9196.02±0.9528.93±3.8694.57±0.6694.75±0.13
32λLSV=0.00146.53±3.7789.67±1.0927.84±9.4795.29±1.0837.33±1.6391.77±0.395.65±0.7798.82±0.1124.09±6.7795.68±1.3528.29±3.3394.25±0.5594.68±0.07
10-53.20±3.9088.97±0.7335.62±9.6694.41±1.2845.73±6.2989.73±1.6911.72±3.3997.93±0.4641.80±14.3592.51±3.0337.61±4.0092.71±0.9791.95±2.32
FFS--52.86±4.4983.47±1.6738.67±11.2189.74±5.1944.65±1.2987.47±0.776.59±0.9298.67±0.1731.34±1.8893.24±0.9734.82±2.0390.52±0.9894.69±0.15
-λLSV=0.00150.74±3.9884.93±0.7532.25±12.1692.97±3.6042.99±2.3088.30±1.105.76±1.2398.82±0.2728.16±6.4594.18±1.9831.98±3.1091.84±1.2894.73±0.12
-λCN=1.050.08±3.8684.92±1.8530.91±6.1992.77±2.3345.69±2.5987.16±0.756.22±0.6598.73±0.1126.60±3.6494.80±0.6431.90±2.5191.67±0.9494.85±0.20
96λLSV=1.048.11±5.1589.84±1.5140.65±15.4793.32±3.1636.37±2.2892.29±0.605.22±1.3598.89±0.1824.84±10.0895.82±1.6831.04±5.0594.03±1.0694.71±0.08
64λLSV=1.051.70±5.6788.81±1.0933.02±14.0994.60±2.3039.05±3.7391.53±1.235.56±1.9598.84±0.3026.63±7.5195.65±1.1231.19±4.3293.89±0.6994.68±0.08
32-46.75±3.4989.61±1.4031.77±8.7394.44±1.3541.51±3.3890.72±1.024.93±1.1199.04±0.1628.92±11.2394.49±2.9430.78±3.7093.66±0.9594.79±0.18
10-54.72±5.8984.77±3.2345.70±12.9591.49±2.2741.52±3.2789.51±1.318.21±3.5998.33±0.6728.58±7.7594.30±1.8235.75±5.2691.68±1.3894.54±0.17
", + "bbox": [ + 94, + 107, + 906, + 282 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Datasets", + "text_level": 1, + "bbox": [ + 89, + 292, + 192, + 306 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We assess FEVER-OOD with standard datasets for both classification and object detection. For image-level OOD detection (i.e., image classification), we train VOS and FFS using the CIFAR-10 and CIFAR-100 [20] datasets as in-distribution while testing OOD detection in five non-overlapping datasets: Textures [3], SVHN [30], Places365 [47], LSUN [45] and iSUN [11]. We also test Dream-OOD with the CIFAR-100 dataset as in-distribution data with the same OOD datasets. Additionally, we also train Dream-OOD with the Imagenet-100 dataset, a 100 classes partition of Imagenet [35], as in-distribution while testing OOD detection in four datasets: Textures [3], Places365 [47], iNaturalist [40] and SUN [42]. We use the PASCAL VOC [10] dataset as in-distribution for object-level OOD detection using VOS and FFS and test for OOD detection on the MS-COCO dataset [28] removing images containing objects with overlapping in-distribution categories [7].", + "bbox": [ + 88, + 315, + 485, + 574 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Implementation details", + "text_level": 1, + "bbox": [ + 89, + 583, + 305, + 598 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We follow the established architectural choices of prior works. For image-level OOD detection with VOS and FFS, we use a WideResNet-40 [46] with a 128-dimensional feature space. We choose different NSR values ranging from the feature dim to the number of classes, gradually decreasing the nullity. For CIFAR-10, we evaluate the $\\{96,64,32,10\\}$ -NSR, while for CIFAR-100 we assess $\\{114,100\\}$ -NSR. With respect to Dream-OOD models, we use a ResNet-34 [14] with 512-dimensional feature space, evaluating $\\{256,128,100\\}$ -NSR for both CIFAR-100 and Imagenet-100. Similarly, we reduce the feature space of the classification head to 768, 512 and 265 dimensions for VOS and FFS object detection experiments, having an original 1,024-dimensional object-wise feature space (as per Faster RCNN [33]). With regards to the LSVR and CNR, we test with various loss weight values in all the architectures (1, 0.1, 0.01, and 0.001), to evaluate their impact on model performance (except for Imagenet-100, where we use $\\lambda_{LSV} = 0.001$ and $\\lambda_{CN} = 0.01$ based on CIFAR-100 experiments). We train all the models using the original", + "bbox": [ + 91, + 599, + 485, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "corresponding baseline settings. Details on the hyperparameters and the training regime for each model are given in the Supp. Mat. To obtain reliable comparisons we use deterministic learning via a fixed random seed for all experiments. We run all experiments 5 times and report mean and standard deviation for all metrics.", + "bbox": [ + 511, + 294, + 906, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Results", + "text_level": 1, + "bbox": [ + 511, + 395, + 601, + 410 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We discuss the performance of FEVER-OOD against different free energy based OOD detection models in classification and detection. Tabs. 1 to 3 compare the use of FEVER-OOD with the corresponding baseline methods. We report different NSR as well as the best results obtained using $\\lambda_{LSV}$ (Eq. (13)) or $\\lambda_{CN}$ (Eq. (14)), and their corresponding value (we show both cases when no NSR is applied). Extensive ablations are available in the Supp. Mat.", + "bbox": [ + 511, + 415, + 906, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CIFAR-10. Tab. 1 shows the results of the VOS and FFS architectures with CIFAR-10 as in-distribution. In general, NSR and LSVR improve OOD detection for both VOS and FFS models. The best performance using the VOS method is achieved using a 96-NSR with LSVR, using a $\\lambda_{LSV} = 1.0$ , where the average AUROC increases to 94.78 and the average FPR95 reduces to $28.22\\%$ , compared to an AUROC of 92.15 and $33.66\\%$ of the baseline VOS model without affecting the in-distribution accuracy. Similarly, the best FFS setting is achieved with a 96-NSR and $\\lambda_{LSV} = 1.0$ , resulting in 94.03 AUROC and $31.04\\%$ FPR95, compared to the baseline FFS model, with an AUROC of 90.52 and FPR95 of $34.82\\%$ , although the best FPR95 is achieved with 32-NSR and no singular value regularization, with an FPR95 of 30.78. In general, it is seen that a large NSR (i.e., 10-NSR) may decrease the AUROC and increase the FPR95. Since VOS and FFS are techniques that synthesize outliers in the feature space, a large reduction of the feature dimension might not allow for proper in-distribution understanding, impacting in the OOD training of these techniques. Additionally, it is also observed that LSVR yields better results than CNR, and a value for $\\lambda_{LSV} = 1.0$ is adequate for none to moderate NSR, while smaller values for LSVR (or none) are better for large NSR. While LSVR increases the differ", + "bbox": [ + 511, + 537, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4534", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d0107b85f7a5a0e33af91b81e5a80311a473c9b260a0ca7d6911b533d953983e.jpg", + "table_caption": [ + "Table 2. CIFAR-100 Results (ID Acc. = in-distribution accuracy; Null Space Reduction (NSR) methods = our approach)." + ], + "table_footnote": [], + "table_body": "
MethodFEVER-OODOOD DatasetsID Acc
r-NSRRegularizerTexturesSHVNPlaces365LSUNiSUNAvg
FPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROC
VOS--81.24±2.1176.01±0.8076.09±6.5183.18±3.5980.53±1.4676.25±0.7837.86±4.0693.21±0.6879.34±1.6076.43±1.4471.01±1.2381.02±0.4076.04±0.22
-λLSV=0.182.87±1.1176.49±0.9473.35±2.9684.57±1.5380.36±1.1676.89±0.7940.05±3.7492.83±0.4674.28±4.0380.25±3.3470.18±0.8782.21±0.5676.01±0.16
-λCN=0.0182.51±2.6476.04±1.9173.67±3.1085.31±1.2480.18±1.0876.59±0.2341.08±3.2792.75±0.6380.83±4.3773.79±5.5171.65±0.9680.90±1.1176.20±0.11
114λLSV=0.0180.02±2.7578.73±1.8383.45±2.5982.07±1.3678.41±0.9677.71±0.9328.90±3.6195.11±0.5872.07±9.0780.55±4.7068.57±2.2882.83±0.9175.72±0.16
100-80.47±2.3278.21±1.1479.10±7.4283.39±3.9179.19±1.0477.23±0.5028.99±2.5995.15±0.3975.51±6.6879.63±3.7068.65±2.2982.72±0.5775.54±0.19
FFS--82.87±1.3975.54±0.9376.32±6.1784.83±2.3681.14±1.0976.27±0.7436.27±4.9193.54±1.0382.83±2.9874.29±3.0871.89±1.6480.89±1.1176.04±0.11
-λLSV=0.0180.95±1.3476.53±0.7083.12±5.5981.61±2.9280.19±0.8476.58±0.4932.93±4.0594.20±0.5179.05±5.1677.28±3.2371.25±2.1181.24±0.7276.27±0.28
-λCN=0.00180.27±2.7676.78±1.5378.57±7.3282.57±2.9580.40±1.0976.40±0.5332.91±4.9094.06±1.0280.71±7.9074.49±5.5770.57±2.4180.86±0.8876.05±0.16
114λLSV=0.00179.26±2.6778.16±1.5074.28±7.1385.37±3.2079.45±0.8577.25±0.6124.10±1.4295.94±0.1773.61±5.8280.61±2.4966.14±1.9183.47±0.9875.45±0.37
100-77.69±2.9778.62±1.0277.84±9.6283.40±3.5280.41±0.7276.58±0.3521.91±0.9596.25±0.1179.60±6.0675.79±5.1767.49±2.0382.13±0.8575.48±0.28
Dream-OOD--62.20±1.0283.84±0.3873.05±1.9284.56±0.2177.95±1.9779.43±0.1739.90±2.0192.87±0.441.70±0.1199.58±0.0450.96±1.4488.06±0.6075.61±0.19
-λLSV=0.0158.45±2.2786.04±0.4168.75±1.8487.65±0.3077.45±2.1178.59±0.2215.45±2.5797.24±0.091.55±0.0399.63±0.4544.33±3.4889.83±0.1275.87±0.23
-λCN=0.00157.40±1.8886.28±0.3277.75±1.1185.13±0.4178.60±1.2378.73±0.1527.2±1.7795.01±0.421.55±0.0899.57±0.0748.5±1.5488.94±0.4376.32±0.14
256-60.00±3.2185.42±0.7967.50±2.4785.84±0.6675.90±1.1079.57±0.4319.85±1.0596.74±0.671.00±0.0299.78±0.0644.82±2.7289.47±0.24277.01±0.12
128λLSV=0.156.45±2.3987.73±1.0167.45±1.1287.53±0.4478.45±2.4677.24±0.6129.60±1.7394.81±0.401.95±0.1099.5±0.1046.78±2.7289.36±0.8376.21±0.35
100-57.55±2.4886.8±0.4454.45±3.0288.69±0.5975.80±1.2178.88±0.6924.45±1.5795.86±0.521.60±0.0799.65±0.2142.77±2.1089.98±0.1576.41±0.32
", + "bbox": [ + 109, + 104, + 880, + 268 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/68fa8866a7e13226a215d32b057d741526ad5890366a3a42010a74173933b055.jpg", + "table_caption": [ + "Table 3. ImageNet-100 Results (ID Acc. = in-distribution accuracy; Null Space Reduction (NSR) methods = our approach)." + ], + "table_footnote": [], + "table_body": "
MethodFEVER-OODOOD DatasetsID Acc
r-NSRRegularizeriNaturalistPlaces365SUNTexturesAvg
FPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROC
Dream-OOD--23.98±2.0895.94±0.2641.75±1.7292.48±0.1240.85±1.1392.76±0.0950.73±0.8586.21±0.2939.33±1.0891.84±0.1287.76±0.19
-λLSV=0.0159.91±3.0990.89±0.4866.26±2.1788.12±0.4774.26±1.5284.53±0.3360.56±1.2183.91±0.4265.25±1.5086.87±0.2587.85±0.11
-λLSV=0.00159.71±3.5390.97±0.7065.01±1.9088.39±0.4573.93±1.5884.69±0.4960.51±1.4684.01±0.5164.79±1.6987.01±0.4587.77±0.13
256-23.84±1.7395.80±0.3044.35±1.7192.14±0.3042.92±1.9892.46±0.3344.38±1.8188.60±0.5438.87±1.2992.25±0.2387.57±0.24
128λLSV=0.0123.46±2.5495.93±0.3643.37±1.4792.19±0.2242.42±1.6092.44±0.2941.86±1.0989.48±0.2737.78±1.4592.51±0.2287.60±0.11
100λLSV=0.0122.24±0.8196.16±0.1441.08±3.0392.59±0.4440.39±2.8492.91±0.4042.31±0.9789.30±0.3736.50±1.6792.74±0.2187.42±0.15
", + "bbox": [ + 109, + 292, + 880, + 383 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b5a0447239bfd21dc15f5b456962b4ed61103b23b89df1f7fdc9f6fc4365220a.jpg", + "image_caption": [ + "(a) Dataset: CIFAR-10", + "Figure 4. Singular Values Spectrum." + ], + "image_footnote": [], + "bbox": [ + 96, + 388, + 279, + 494 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c3e37d80961ce74aefb7bf09e7d5aaeea22dc52d68dccd6d83b173b223b81729.jpg", + "image_caption": [ + "(b) Dataset: CIFAR-100" + ], + "image_footnote": [], + "bbox": [ + 287, + 388, + 473, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ence in energy of outliers in or near the least singular vector direction, the CN regulariser makes equalized free energy changes for all directions. Since the latter effect might disable directions with smaller energy changes than others, it might also reduce the free energy change in the direction of the largest singular value. Fig. 4(a) shows the singular value spectrum of the baseline VOS model against the LSVR and CNR versions for the CIFAR-10 dataset. It is seen that the while both regularizers homogenize the singular values, the CNR also reduces the most significant singular value. Finally, our experiments show that VOS achieves better results than FFS in CIFAR-10 OOD detection.", + "bbox": [ + 89, + 522, + 483, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "CIFAR-100. Tab. 2 shows the results for VOS, FFS and Dream-OOD trained on the CIFAR-100 dataset as in-distribution. The best models for VOS and FFS are achieved with 114-NSR and LSVR, with a $68.57\\%$ FPR95 $(-2.44\\%)$ VOS) and 82.83 $(+1.81\\mathrm{VOS})$ AUROC for VOS-114-NSR and $66.14\\%$ FPR95 $(-5.75\\%)$ FFS) and 83.47 $(+2.58\\mathrm{FFS})$ AUROC for FFS-114-NSR. However, despite these improvements, the performance of virtual outlier synthesis methods is usually poor for several in-distribution classes [44], hence the enhancement of these architectures is still challenging. Regarding Dream-OOD, it is noted that the best results are obtained by Dream-OOD-100-NSR, having an FPR95 of $42.77\\%$ $(-8.19\\%)$ Dream-OOD) and 89.98", + "bbox": [ + 89, + 703, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "AUROC (+1.9 Dream-OOD). Finally, for the three energy techniques (VOS, FFS and Dream-OOD), we observed that the model does not converge when using 100-NSR and the CNR (see Supp. Mat.), suggesting that the conditioning number penalization does not allow the model to learn the in-distribution task. Fig. 4(b) shows the singular value spectrum for the LSVR and CNR versions of the VOS model. Differently from the CIFAR-10 experiments, it is seen that LSVR keeps the most significant singular values similar to the baseline but homogenizes the least significant singular values, while the CNR keeps all singular values at a low level, hence explaining why it fails for this dataset.", + "bbox": [ + 511, + 390, + 906, + 569 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Imagenet-100. We show the results of Dream-OOD with Imagenet-100 as in-distribution in Tab. 3. Similar as with CIFAR-100, the best results are obtained with NSR100, effectively eliminating the null space of the last layer (since it becomes square and is not singular). Contrary to CIFAR-100, LSVR has a negative impact when used alone. Considering that we use the same pre-training scheme as Du et al. [8], adding LSVR and CNR on an already pretrained last linear layer might affect the already formed feature space. Hence, these regularisers benefit untrained NSR modules. The best model, Dream-OOD-100-NSR with LSVR, achieves an average $36.50\\%$ FPR95 and 92.74 AUROC, compared to the baseline $39.33\\%$ FPR95 and 91.84 AUROC, becoming the state-of-the-art for Imagenet-100.", + "bbox": [ + 511, + 569, + 908, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "PASCAL VOC. We test FEVER-OOD techniques in object detection task, where the goal is to detect object-level OOD instances. Tab. 4 show the result for the baseline VOS and FFS, taking PASCAL VOC [10] as in-distribution dataset. For both tasks, we evaluate on two OOD datasets that contain subset of images from: MS-COCO [28] and OpenImages [22], following the original works [7, 21]. The best-performing model for VOS is VOS-768-NSR with CNR,", + "bbox": [ + 511, + 779, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4535", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/38dbf40693c2a1f54c600dfde7a1f99607541a80bb3b382b4e01f49999ed70f7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 102, + 88, + 269, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/faca7797e080a595d5658a989be5df461a6290e0819605a9804d8973355d6141.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 269, + 90, + 413, + 175 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ced249d0cb05ba7cd7dae5c191dd056ace4e257e8a829eee15d6f0bf62667384.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 90, + 722, + 175 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dc34406b1cea827db91f32e8f55f3b196923b6a6b91f5fa79bf82d32af4fa7b6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 722, + 90, + 883, + 175 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0131f23c020b950610ac0e882aeeaf1edd8e56e0d088678eb9446f5fe4cd56f6.jpg", + "image_caption": [ + "Figure 5. MS-COCO objects detected on OOD images by VOS baseline [7] (first row) and Fever-OOD (second row). Blue: OOD objects detected and mis-classified as being in-distribution. Green: the same OOD objects correctly detected as OOD by FEVER-OOD (ours)." + ], + "image_footnote": [], + "bbox": [ + 101, + 176, + 267, + 261 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e2d1f8f1cedb55b6fb5d8445d54255cf5535fafaca5776d32bedee50b1fd4b21.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 267, + 176, + 413, + 261 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fde8a35f769bbd6f5f32b25bc9dda203d2d11d2aa027c8ec6d8b373e4c0c396a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 176, + 722, + 261 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/12c8724f5737873132ceb2e02d3f4dab10f687b8d2e26550db93809d26955762.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 722, + 176, + 888, + 262 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/99ec6a99027725d4b081080c4d77dad883791b868892b91ecf0d7d6f85eb605c.jpg", + "table_caption": [ + "Table 4. VOS results with PASCAL VOC as in-distribution (Null Space Reduction (NSR) methods = our approach)." + ], + "table_footnote": [], + "table_body": "
MethodFEVER-OODOOD DatasetsmAP
r-NSRλLSVλCNMS-COCOOpenImages
FPR95AUROCFPR95AUROC
VOS---50.2987.7753.0986.580.489
-0.001-49.9688.1052.2685.800.489
--153.4286.6652.7385.770.490
768--48.9487.9452.3486.330.489
7680.01-50.0887.7851.7186.040.493
768-0.0147.4788.1549.3686.320.491
512--49.8488.4255.3385.330.486
5120.001-49.2688.3954.3285.820.489
512-0.0149.5987.4854.8185.480.490
256--47.8888.4952.4186.390.486
2560.01-48.5188.3650.5286.120.491
256-0.00149.188.2751.4786.080.489
FFS---50.7787.1853.7885.340.487
-0.01-53.0986.7553.4785.860.490
--0.0150.3687.6551.1985.980.485
768--51.5087.5953.3385.840.488
7680.01-47.4288.3652.4686.400.489
768-0.0151.1286.8255.4484.840.490
512--46.9388.8253.8986.280.491
5120.01-47.4288.3652.4686.400.488
512-0.0151.4287.0554.9884.550.490
256--51.1888.7453.6786.560.487
2560.01-47.9388.0447.9384.950.489
256-0.0152.9486.6252.8884.450.489
", + "bbox": [ + 94, + 337, + 480, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "achieving average FPR95 scores of $47.47\\%$ and $49.36\\%$ on MS-COCO and OpenImages, respectively, an improvement of nearly $4\\%$ over the VOS baseline. For FFS, we achieve a minimum FPR95 of $46.93\\%$ with FFS-512-NSR model and $47.93\\%$ for the FFS-256-NSR model with the LSVR for both OOD datasets. Additionally, our approach generally delivers higher AUROC scores compared to the FFS baseline across both OOD datasets. Qualitative results are illustrated in Fig. 5, where we visualize the prediction on MS-COCO OOD images, using the baseline VOS (top) and the best Fever-OOD model, i.e., VOS-768-NSR, with the PASCAL-VOC as the in-distribution dataset, showing that out framework outperforms baseline VOS in in detecting outlier objects with a significantly lower false positive rate.", + "bbox": [ + 88, + 619, + 482, + 833 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Computational overhead. The computational overhead introduced by our method is minimal since we only add an extra linear layer. For instance, the ResNet-34 used in Dream-OOD takes $4.42 \\pm 1.15$ ms/img and has 21.5M", + "bbox": [ + 88, + 839, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "params. Adding 256, 128, 100-NSR layers change the inference time to $3.91 \\pm 0.62$ , $4.44 \\pm 1.17$ , $4.13 \\pm 1.02$ ms/img, with an extra 105k, 27k, 10k params, which is negligible compared with the total size of ResNet-34.", + "bbox": [ + 511, + 304, + 905, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 369, + 633, + 383 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we present FEVER-OOD, a robust approach for addressing critical vulnerabilities found in existing energy-based OOD detection. Our method targets the Null Space Vulnerabilities (NSV), where in-distribution and OOD instances can produce similar free energy scores despite differing feature representations, and Least Singular Value Vulnerabilities (LSVV), where energy score similarities are influenced by the least singular vector of the classifier last linear layer. These vulnerabilities are especially problematic in high-dimensional feature spaces where the feature dimension exceeds the number of classes, impacting the effectiveness of OOD detection. To mitigate these issues, FEVER-OOD incorporates strategies such as reducing the null space through lower-dimensional feature spaces and introducing a Least Singular Value Regularizer to maximize the least singular value, enhancing the distinction between energy scores of different samples. Additionally, we propose a Condition Number Regularizer to promote uniform energy changes for feature-space displacements, fostering consistent energy variation and improving detection robustness. Our comprehensive evaluation on benchmark datasets, including applications to the Dream-OOD framework with ImageNet-100 as the ID dataset, demonstrates the significant effectiveness of FEVER-OOD. Specifically, our approach achieves a $10.13\\%$ reduction in the average false positive rate and a $1.6\\%$ increase in AUROC, setting a new state-of-the-art in OOD detection. Furthermore, our results highlight the generalization capability of FEVER-OOD across different detection architectures, showcasing its potential beyond conventional classification settings. Future extensions of FEVER-OOD could include applications to varied domains and integration with other OOD detection frameworks to further validate its versatility and performance in diverse machine learning contexts. This work provides valuable insights into enhancing OOD robustness and sets a foundation for more resilient free energy-based detection methods.", + "bbox": [ + 511, + 387, + 906, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4536", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 104, + 187, + 119 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Haoyue Bai, Gregory Canal, Xuefeng Du, Jeongyeol Kwon, Robert D Nowak, and Yixuan Li. Feed two birds with one scone: Exploiting wild data for both out-of-distribution generalization and detection. In International Conference on Machine Learning, pages 1454-1471. PMLR, 2023. 2", + "[2] Haoyue Bai, Jifan Zhang, and Robert Nowak. Aha: Human-assisted out-of-distribution generalization and detection. arXiv preprint arXiv:2410.08000, 2024. 2", + "[3] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3606-3613, 2014. 6", + "[4] Matthew Cook, Alina Zare, and Paul Gader. Outlier detection through null space analysis of neural networks. ICML Workshops, 2020. 2, 4", + "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 3", + "[6] Xuefeng Du, Xin Wang, Gabriel Gozum, and Yixuan Li. Unknown-aware object detection: Learning what you don't know from videos in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13678-13688, 2022. 1", + "[7] Xuefeng Du, Zhaoning Wang, Mu Cai, and Yixuan Li. Vos: Learning what you don't know by virtual outlier synthesis. In International Conference on Learning Representations, 2022. 1, 2, 3, 4, 5, 6, 7, 8, 9, 10", + "[8] Xuefeng Du, Yiyou Sun, Xiaojin Zhu, and Yixuan Li. Dream the impossible: Outlier imagination with diffusion models. In Advances in Neural Information Processing Systems, 2023. 2, 3, 4, 5, 7", + "[9] Xuefeng Du, Zhen Fang, Ilias Diakonikolas, and Yixuan Li. How does unlabeled data provably help out-of-distribution detection? arXiv preprint arXiv:2402.03502, 2024. 1", + "[10] Mark Everingham, Luc Van Gool, Christopher K. I. Williams, John M. Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International Journal of Computer Vision, 88(2):303-338, 2010. 6, 7, 2, 9, 10", + "[11] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 580-587, 2014. 6", + "[12] Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep Learning. MIT Press, 2016. http://www.deeplearningbook.org.3", + "[13] Ali Harakeh and Steven L. Waslander. Estimating and evaluating regression predictive uncertainty in deep object detectors. In International Conference on Learning Representations, 2021. 5" + ], + "bbox": [ + 93, + 128, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 3, 6, 2", + "[15] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In International Conference on Learning Representations, 2022. 2, 8", + "[16] Yen-Chang Hsu, Yilin Shen, Hongxia Jin, and Zsolt Kira. Generalized odin: Detecting out-of-distribution image without learning from out-of-distribution data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10951-10960, 2020. 1", + "[17] Rui Huang, Andrew Geng, and Yixuan Li. On the importance of gradients for detecting distributional shifts in the wild. Advances in Neural Information Processing Systems, 34:677-689, 2021. 8", + "[18] B.K.S. Isaac-Medina, Y.F.A. Gaus, N. Bhowmik, and T.P. Breckon. Towards open-world object-based anomaly detection via self-supervised outlier synthesis. In Proc. European Conference on Computer Vision. Springer, 2024. 1, 3, 4", + "[19] Julian Katz-Samuels, Julia B Nakhleh, Robert Nowak, and Yixuan Li. Training ood detectors in their natural habitats. In International Conference on Machine Learning, pages 10848-10865. PMLR, 2022. 2", + "[20] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009. 6, 2", + "[21] Nishant Kumar, Siniša Šegvic, Abouzar Eslami, and Stefan Gumhold. Normalizing flow based feature synthesis for outlier-aware object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5156-5165, 2023. 1, 2, 3, 4, 5, 7, 6, 9, 10", + "[22] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision, 128(7):1956-1981, 2020. 7, 6, 9", + "[23] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. Advances in Neural Information Processing Systems, 30, 2017. 1", + "[24] Hao Lang, Yinhe Zheng, Yixuan Li, Jian SUN, Fei Huang, and Yongbin Li. A survey on out-of-distribution detection in NLP. Transactions on Machine Learning Research, 2024. 1", + "[25] Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. Advances in Neural Information Processing Systems, 31, 2018. 1, 2", + "[26] Adrian S Lewis and Hristo S Sendov. Nonsmooth analysis of singular values. part i: Theory. Set-Valued Analysis, 13: 213–241, 2005. 5, 1", + "[27] Shiyu Liang, Yixuan Li, and R Srikant. Enhancing the reliability of out-of-distribution image detection in neural net" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4537", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "works. In 6th International Conference on Learning Representations, 2018, 2018. 1, 2, 8", + "[28] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5, 6, 7, 10", + "[29] Weitang Liu, Xiaoyun Wang, John Owens, and Yixuan Li. Energy-based out-of-distribution detection. Advances in Neural Information Processing Systems, 33:21464-21475, 2020. 1, 2, 3, 4", + "[30] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading digits in natural images with unsupervised feature learning. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning 2011, 2011. 6", + "[31] Anh Nguyen, Jason Yosinski, and Jeff Clune. Deep neural networks are easily fooled: High confidence predictions for unrecognizable images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 427-436, 2015. 1", + "[32] Bo Peng, Yadan Luo, Yonggang Zhang, Yixuan Li, and Zhen Fang. Conjnorm: Tractable density estimation for out-of-distribution detection. In The Twelfth International Conference on Learning Representations, 2024. 2", + "[33] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: towards real-time object detection with region proposal networks. In Proceedings of the 28th International Conference on Neural Information Processing Systems - Volume 1, page 91-99, Cambridge, MA, USA, 2015. MIT Press. 5, 6", + "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 3, 5", + "[35] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 115 (3):211-252, 2015. 2, 6", + "[36] Mohammadreza Salehi, Hossein Mirzaei, Dan Hendrycks, Yixuan Li, Mohammad Hossein Rohban, and Mohammad Sabokrou. A unified survey on anomaly, novelty, open-set, and out of-distribution detection: Solutions and future challenges. Transactions on Machine Learning Research, 2022. 1", + "[37] Carl-Johann Simon-Gabriel, Yann Ollivier, Leon Bottou, Bernhard Schölkopf, and David Lopez-Paz. First-order adversarial vulnerability of neural networks and input dimension. In Proceedings of the 36th International Conference on Machine Learning, pages 5809-5817. PMLR, 2019. 5", + "[38] Yiyou Sun, Yifei Ming, Xiaojin Zhu, and Yixuan Li. Out-of-distribution detection with deep nearest neighbors. In International Conference on Machine Learning, pages 20827-20840. PMLR, 2022. 8" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Leitian Tao, Xuefeng Du, Jerry Zhu, and Yixuan Li. Nonparametric outlier synthesis. In The Eleventh International Conference on Learning Representations, 2023. 3, 4, 8", + "[40] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8769-8778, 2018. 6", + "[41] Haoqi Wang, Zhizhong Li, Litong Feng, and Wayne Zhang. Vim: Out-of-distribution with virtual-logit matching. In CVPR, pages 4921-4930, 2022. 2, 4, 8", + "[42] Jianxiong Xiao, James Hays, Krista A. Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 3485-3492, 2010. 6", + "[43] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10687-10698, 2020. 3", + "[44] Jingkang Yang, Kaiyang Zhou, Yixuan Li, and Ziwei Liu. Generalized out-of-distribution detection: A survey. International Journal of Computer Vision, pages 1-28, 2024. 1, 7", + "[45] Fisher Yu, Yinda Zhang, Shuran Song, Ari Seff, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 6", + "[46] Erwan Zerhouni, David Lanyi, Matheus Viana, and Maria Gabrani. Wide residual networks for mitosis detection. In 2017 IEEE 14th International Symposium on Biomedical Imaging (ISBI 2017), pages 924-928. IEEE, 2017. 6", + "[47] Bolei Zhou, Agata Lapedriza, Antonio Torralba, and Aude Oliva. Places: An image database for deep scene understanding. Journal of Vision, 17(10):296-296, 2017. 6" + ], + "bbox": [ + 516, + 90, + 905, + 614 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4538", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/FEVER-OOD_ Free Energy Vulnerability Elimination for Robust Out-of-Distribution Detection/0c16965c-d6fa-4547-a6ba-ab34512ce472_model.json b/2025/FEVER-OOD_ Free Energy Vulnerability Elimination for Robust Out-of-Distribution Detection/0c16965c-d6fa-4547-a6ba-ab34512ce472_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2a354792d93d0505c0c1ef2fc5b1d534f31d89ea --- /dev/null +++ b/2025/FEVER-OOD_ Free Energy Vulnerability Elimination for Robust Out-of-Distribution Detection/0c16965c-d6fa-4547-a6ba-ab34512ce472_model.json @@ -0,0 +1,2123 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.171, + 0.131, + 0.83, + 0.175 + ], + "angle": 0, + "content": "FEVER-OOD: Free Energy Vulnerability Elimination for Robust Out-of-Distribution Detection" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.185, + 0.94, + 0.222 + ], + "angle": 0, + "content": "Brian K.S. Isaac-Medina1 Maurizio Che2 Yona Faline A. Gaus1 Samet Akcay3 Toby P. Breckon1 \n1Durham University, Durham, UK 2University of Vienna, Vienna, Austria 3Intel" + }, + { + "type": "text", + "bbox": [ + 0.202, + 0.224, + 0.841, + 0.257 + ], + "angle": 0, + "content": "{brian.k.isaac-medina, yona.f.binti-abd-gaus, toby.breckon}@durham.ac.uk mauricio.adrian.che.moguel@univie.ac.at samet.akcay@intel.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.323, + 0.486, + 0.822 + ], + "angle": 0, + "content": "Modern machine learning models, that excel on computer vision tasks such as classification and object detection, are often overconfident in their predictions for Out-of-Distribution (OOD) examples, resulting in unpredictable behaviour for open-set environments. Recent works have demonstrated that the free energy score is an effective measure of uncertainty for OOD detection given its close relationship to the data distribution. However, despite that free energy-based methods representing a significant empirical advance in OOD detection, our theoretical analysis reveals previously unexplored and inherent vulnerabilities within the free energy score formulation such that in-distribution and OOD instances can have distinct feature representations yet identical free energy scores. This phenomenon occurs when the vector representing the feature space difference between the in-distribution and OOD sample belongs to the null space of the last layer of a neural-based classifier. To mitigate these issues, we explore lower-dimensional feature spaces to reduce the null space footprint and introduce novel regularisation to maximize the least singular value of the final linear layer, hence enhancing intersample free energy separation. We refer to these techniques as Free Energy Vulnerability Elimination for Robust Out-of-Distribution Detection (FEVER-OOD). Extensive experimentation shows that FEVER-OOD improves OOD detection in CIFAR-10 with \\(28.22\\%\\) FPR95 and 94.78 AUROC vs \\(33.66\\%\\) FPR95 and 92.15 AUROC of the baseline model, CIFAR-100 with \\(42.77\\%\\) FPR95 and 89.98 AUROC vs \\(50.96\\%\\) FPR95 and 88.06 AUROC of the baseline model, and achieves state of the art OOD detection in Imagenet-100, with average OOD FPR95 of \\(36.50\\%\\) and an AUROC of 92.74 when used with the Dream-OOD model, compared with a \\(39.33\\%\\) and 91.84 AUROC without FEVER-OOD." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.831, + 0.222, + 0.846 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.485, + 0.901 + ], + "angle": 0, + "content": "Out-of-distribution (OOD) detection (c.f. anomaly/outlier detection), aims to identify abnormal samples that deviate significantly from a given data distribution. Whilst" + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.289, + 0.7, + 0.3 + ], + "angle": 0, + "content": "Free Energy-based OOD (VOS)" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.301, + 0.71, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.75, + 0.289, + 0.867, + 0.3 + ], + "angle": 0, + "content": "FEVER-OOD (VOS)" + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.301, + 0.905, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.426, + 0.907, + 0.482 + ], + "angle": 0, + "content": "Figure 1. FEVER-OOD improves baseline Free Energy-based OOD detection methods, where baseline methods predict OOD classes as in-distribution (left, blue), while FEVER-OOD classifies them as OOD (right, green)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.488, + 0.907, + 0.58 + ], + "angle": 0, + "content": "OOD is essential in developing deployable machine learning systems, to ensure valid inference only on data drawn from same distribution used in training [24, 36, 44], real-world deployments often involve scenarios where models encounter unseen OOD classes that could otherwise undermine decision stability and hence wider system robustness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.909, + 0.793 + ], + "angle": 0, + "content": "Methods such as deep ensembles [23], ODIN [27], Mahalanobis distance-based detection [25], and generalized ODIN [16] all utilize softmax() to differentiate indistribution from OOD. Whilst intuitive, such approaches often struggle as deep neural networks commonly yield overly high softmax() scores, even for inputs far from the training data distribution [31]. Given OOD and indistribution samples should have significant spatial separation in any given feature space, Liu et al. [29] developed an alternative scoring function to effectively measure the alignment of a given sample to an a priori learned distribution. Denoted as the free energy scoring function, this formulation demonstrates superior OOD performance across both classification [7-9] and detection tasks [6, 7, 18, 21]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Despite free energy-based methods representing a significant empirical advance in OOD detection, our theoretical analysis reveals previously unexplored and inherent vulnerabilities within the free energy score formulation that could impact their practical reliability. These vulnerabilities reveal that in-distribution and OOD instances can have distinct feature representations yet produce identical free en" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4529" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "ergy scores. This phenomenon occurs in the feature space when the vector direction representing the difference between an in-distribution and an OOD sample lies within the null space of the last layer of neural-based classifier. Such regions, effectively \"blind spots\" in terms of OOD, lead to OOD and in-distribution generating identical or near-identical free energy score values." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.199, + 0.482, + 0.319 + ], + "angle": 0, + "content": "To mitigate these issues, we explore lower-dimensional feature spaces to effectively reduce the size of the null space. In addition, we also introduce a novel regulariser to maximize the least singular value of the final linear layer, hence enhancing inter-sample free energy separation to improve OOD robustness. We refer to these techniques as Free energy Vulnerability Elimination for Robust Out-of-Distribution Detection (FEVER-OOD)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.322, + 0.433, + 0.337 + ], + "angle": 0, + "content": "In summary, our key contributions in this paper are:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.338, + 0.481, + 0.518 + ], + "angle": 0, + "content": "- The identification of inherent free energy score vulnerabilities, where under null-space conditions, in-distribution and OOD instances will yield similar free energy scores despite differing feature representations. Denoted as Null Space Vulnerabilities (NSV) such conditions readily occur when the feature dimension exceeds the number of classes. Additionally, such free energy similarity can be influenced by the least singular vector (LSV) of the last linear layer of a neural-based classifier, leading to what we denote as Least Singular Value Vulnerabilities (LSVV). Both vulnerabilities negatively impact in-distribution to OOD discrimination." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.52, + 0.481, + 0.685 + ], + "angle": 0, + "content": "- Novel formulations that mitigate the NSV and LSVV impact (Fig. 1). To address NSV, we introduce an additional layer to reduce the null space whilst for LSVV, we develop an LSV Regulariser that maximizes the least singular value, guaranteeing detectable energy variations from small feature-space differences and hence enhancing inter-sample energy scores discrimination. Furthermore, we explore a Condition Number Regulariser to promote balanced energy changes for displacements in all feature-space direction hence ensuring improved uniformity in energy score variations." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.686, + 0.481, + 0.821 + ], + "angle": 0, + "content": "- Comprehensive experiments and ablation studies using established benchmark datasets over both object classification and detection tasks, that demonstrate the effectiveness of FEVER-OOD in improving OOD detection performance. Specifically, Dream-OOD [8] with FEVER-OOD trained on Imagenet-100 [35] as in-distribution decreases the FPR95 from \\(39.33\\%\\) to \\(36.50\\%\\) and increases the average AUROC from 91.84 to 92.74, achieving state-of-the-art performance for OOD detection." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.338, + 0.481, + 0.821 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.233, + 0.849 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.86, + 0.482, + 0.906 + ], + "angle": 0, + "content": "Identifying OOD samples is a crucial yet inherently challenging problem, as models are not exposed to, and hence cannot reliably differentiate, OOD samples." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.667 + ], + "angle": 0, + "content": "OOD Detection: to overcome this challenge, researchers have begun incorporating additional data during training to establish a more conservative and safe decision boundary against OOD inputs. Some methods [1, 19] have used unlabeled data to regularise model training while still focusing on classifying labeled in-distribution data, while others [2] leverage human assistance to strategically label examples within a novel maximum disambiguation region. Despite its potential, leveraging unlabeled data is non-trivial due to the heterogeneous mix of in-distribution and OOD instances. The absence of a clean OOD set presents significant challenges in designing effective OOD learning algorithms. A different branch of works incorporate post-hoc methods, focusing on designing OOD scoring mechanisms at test time. Hendrycks and Gimpel [15] introduced the maximum predicted softmax() probability as an in-distribution score function, establishing an initial baseline for scoring OOD detection. Similarly, ODIN [27] aims to widen the gap between values in the softmax() score vectors by temperature scaling and gradient-based input perturbation, enhancing the separation between in-distribution and OOD instances. Lee et al. [25] use the Mahalanobis distance between the average softmax() vectors of images and incorporate intermediate features to improve detection performance and robustness. Nonetheless, the misalignment between the softmax() score and the true data density [29, 32] ultimately makes these techniques a suboptimal solution to OOD detection. Closer to our work, Cook et al. [4] identify null space adversarial attacks, proposing that instances with large magnitude of their projection onto the span of the null space of a neural network classifier would likely be an outlier. Similarly, Wang et al. [41] propose that since the features lie in low-dimensional manifolds, outliers are likely to deviate from the principal subspace, i.e. the span of the eigenvectors or the largest eigenvalues, of the in-distribution data. Therefore, OOD instances are detected with respect to the magnitude of the component of the perpendicular subspace to the principal subspace." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Energy-based OOD Detection: to address these limitations, Liu et al. [29] propose to use a free energy-based in-distribution score function instead of the softmax() score to identify OOD instances. This approach regularises the model to produce lower energy values for in-distribution and higher energy values for auxiliary outliers, creating a significant energy gap that enhances the model ability to detect OOD samples. Energy-based OOD modeling has enabled novel outlier synthesis paradigms. For instance, Du et al. [7] propose virtual outlier synthesis (VOS) by sampling outliers from low-likelihood in-distribution regions. VOS has shown to be effective for classification and object-level OOD detection. Further outlier synthesis techniques based on VOS have been proposed, including using normalizing flows [21], sampling based on the in-distribution boundary" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4530" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.333 + ], + "angle": 0, + "content": "instances [39] or extending VOS when class labels are unavailable [18]. In addition, VOS techniques arensuitable for object detection [7, 21]. Whilst VOS and subsequent works synthesize outliers in the feature space, Dream-OOD [8] uses Stable Diffusion [34] to generate outliers in the pixel-space, learning better discriminative free-score values for OOD instances and providing visual examples of outliers. Energy-based methods have shown great performance, becoming state-of-the-art in several benchmarks (e.g., Dream-OOD [8] for ImageNet-100 in-distribution), hence we focus in the properties of these methods. Notwithstanding that the core idea of the energy score is that OOD samples must have a significantly different free energy than in-distribution sample, we examine some inherent vulnerabilities in the energy formulation based on the null space and least singular value of the last linear layer of neural-based classifiers." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.339, + 0.396, + 0.357 + ], + "angle": 0, + "content": "3. Preliminaries: Free Energy Score" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.364, + 0.484, + 0.562 + ], + "angle": 0, + "content": "The goal of OOD detection is to detect outliers, \\(\\mathbf{v}\\), that have a low probability in the true (unknown) in-distribution density \\(p(\\mathbf{x})\\). However, the computation of \\(p(\\mathbf{x})\\) is usually unfeasible or computationally untractable, so we instead resort to additional information. Now consider that we also have a set of in-distribution categories \\(\\{1,\\dots ,K\\}\\) such that if we know the joint probability distribution, \\(p(\\mathbf{x},k)\\), for \\(k\\in \\{1,\\ldots ,K\\}\\), we could marginalize \\(p(\\mathbf{x})\\). Nonetheless, since the actual joint probabilities \\(p(\\mathbf{x},k)\\) might be also untractable, we could instead use non-negative arbitrary potential functions \\(\\phi (\\mathbf{x},k)\\) that measure the affinity of the states of \\(\\mathbf{x}\\) and \\(k\\) [12]. In this sense, we can construct the unnormalized probability:" + }, + { + "type": "equation", + "bbox": [ + 0.22, + 0.57, + 0.483, + 0.612 + ], + "angle": 0, + "content": "\\[\n\\tilde {p} (\\mathbf {x}) = \\sum_ {k = 1} ^ {K} \\phi (\\mathbf {x}, k) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.62, + 0.483, + 0.651 + ], + "angle": 0, + "content": "that gives an idea of the likelihood \\(\\mathbf{x}\\). Furthermore, energy-based models enforce the non-negative constraint of \\(\\phi\\) by:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.66, + 0.483, + 0.677 + ], + "angle": 0, + "content": "\\[\n\\phi (\\mathbf {x}, k) = \\exp (- E (\\mathbf {x}, k)), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.686, + 0.484, + 0.734 + ], + "angle": 0, + "content": "where \\( E \\) is an unbounded arbitrary function called the energy. By substituting Eq. (2) into Eq. (1) and taking the negative \\( \\log^1 \\), we get a value known as the free energy:" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.742, + 0.483, + 0.782 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} (\\mathbf {x}) = - \\log \\tilde {p} (\\mathbf {x}) = - \\log \\sum_ {k = 1} ^ {K} \\exp (- E (\\mathbf {x}, k)), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.791, + 0.484, + 0.867 + ], + "angle": 0, + "content": "and more recently used as a concept in OOD detection [7, 8, 18, 21, 29]. To get the actual density \\( p(\\mathbf{x}) \\), we must compute the normalizing partition function of Eq. (1); since we are interested in distinguishing between in-distribution and OOD, \\( \\mathcal{F}(\\mathbf{x}) \\) thus serves well for this purpose." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.088, + 0.905, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.238, + 0.905, + 0.281 + ], + "angle": 0, + "content": "Figure 2. Vulnerabilities of Free Energy-based OOD detection. Directions in the Null Space do not change the energy, while the LSV direction has minimal change, compared to other directions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.289, + 0.907, + 0.41 + ], + "angle": 0, + "content": "In a general machine learning context, a classifier \\( f(\\mathbf{x};\\theta):\\mathbb{R}^d\\to \\mathbb{R}^K \\) with learned parameters \\( \\theta \\) assigns categorical unnormalized probabilities \\( f_{k}(\\mathbf{x};\\theta) \\), called logits, for a given \\( \\mathbf{x}\\in \\mathbb{R}^d \\). While the logits are usually normalized using the softmax() function in order to obtain a probability distribution, they can also be used as the energy \\( E \\) in Eq. (2). This realization enables the use of the free energy (Eq. (3)) as a measure of uncertainty for a classifier:" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.421, + 0.905, + 0.463 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} (\\mathbf {x}) = - \\log \\sum_ {k = 1} ^ {K} \\exp \\left(f _ {k} (\\mathbf {x}; \\theta)\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.473, + 0.906, + 0.549 + ], + "angle": 0, + "content": "Liu et al. [29] demonstrate that Eq. (4) is superior for effective OOD detection when compared to use of softmax() whilst Du et al. [7] explicitly use the free energy score (Eq. (4)) to perform binary classification between indistribution and OOD by training with synthetic outliers." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.562, + 0.874, + 0.58 + ], + "angle": 0, + "content": "4. Vulnerabilities of the Free Energy Score" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.588, + 0.906, + 0.709 + ], + "angle": 0, + "content": "Although the free energy (Eq. (4)) is an effective uncertainty measure for OOD detection [7, 29], here we examine some inherent vulnerabilities when used in conjunction with a neural network based classifier. Firstly, Sec. 4.1 investigates OOD instances that have similar energies as in-distribution samples, whilst Sec. 4.2 discusses minimal changes in free energy, and finally Sec. 4.3 outlines our proposed approach to minimize the impact of these vulnerabilities." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.718, + 0.74, + 0.735 + ], + "angle": 0, + "content": "4.1. Null Space Vulnerability" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.74, + 0.906, + 0.831 + ], + "angle": 0, + "content": "Consider a neural network-based classifier \\( f = (g \\circ h)(\\mathbf{x}) \\) where \\( h(\\mathbf{x}): \\mathbb{R}^d \\to \\mathbb{R}^{d'} \\) encodes the input into a \\( d' \\)-dimensional feature vector and \\( g: \\mathbb{R}^{d'} \\to \\mathbb{R}^K \\) is the last layer of the neural network that transforms the features into unnormalized class probabilities (i.e., the logits). If \\( g \\) is a linear layer (as for most classifiers [5, 14, 43]), then:" + }, + { + "type": "equation", + "bbox": [ + 0.643, + 0.841, + 0.905, + 0.86 + ], + "angle": 0, + "content": "\\[\nf (\\mathbf {x}) = \\mathrm {W} _ {c l s} ^ {\\top} h (\\mathbf {x}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.87, + 0.905, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathrm{W}_{cls} \\in \\mathbb{R}^{d' \\times K}\\) are learnable weights. Hence, the logits \\(f_{k}(\\mathbf{x};\\theta)\\) correspond to the \\(k\\)-th index of \\(f(\\mathbf{x})\\)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.484, + 0.901 + ], + "angle": 0, + "content": "\\(^{1}\\)N.B. the signs of the free energy and the argument of the exponent of \\(\\phi\\) are arbitrarily negative for historical reasons." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4531" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.198 + ], + "angle": 0, + "content": "The core idea of energy-based OOD is that outliers \\(\\mathbf{v}\\) have a significantly different free energy, \\(\\mathcal{F}(\\mathbf{v})\\), from in-distribution samples, \\(\\mathcal{F}(\\mathbf{x})\\). Since OOD instances should be far from in-distribution samples in the feature space [7, 39], the expected distance from the inlier features \\(h(\\mathbf{x})\\) and outlier features \\(h(\\mathbf{v})\\) should be greater than a given boundary distance, \\(d_b\\), as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.207, + 0.484, + 0.226 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} \\left[ \\| h (\\mathbf {x}) - h (\\mathbf {v}) \\| \\right] \\geq d _ {b}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.233, + 0.484, + 0.31 + ], + "angle": 0, + "content": "for a norm \\( \\|\\cdot\\| \\). In other words, the inliers remain bounded to a region in the feature space while outliers are unbounded. Therefore, the expected distance between inliers and outliers is greater than a bounding distance \\( d_b \\). Subsequently, we can rewrite Eq. (6) as:" + }, + { + "type": "equation", + "bbox": [ + 0.223, + 0.319, + 0.483, + 0.337 + ], + "angle": 0, + "content": "\\[\nh (\\mathbf {v}) = h (\\mathbf {x}) + \\boldsymbol {\\delta}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.345, + 0.484, + 0.378 + ], + "angle": 0, + "content": "where \\(\\delta\\) is a vector with \\(\\mathbb{E}[\\|\\delta\\|] \\geq d_b\\). By substituting Eq. (7) into Eq. (5) we obtain the following:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.384, + 0.484, + 0.405 + ], + "angle": 0, + "content": "\\[\nf (\\mathbf {v}) = \\mathrm {W} _ {c l s} ^ {\\top} (h (\\mathbf {x}) + \\boldsymbol {\\delta}) = f (\\mathbf {x}) + \\mathrm {W} _ {c l s} ^ {\\top} \\boldsymbol {\\delta}. \\qquad (8)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.412, + 0.484, + 0.607 + ], + "angle": 0, + "content": "However within this formulation, if \\(\\mathrm{W}_{cls}\\) has more rows than columns, which occurs when the feature dimension is larger than the number of classes, the size of the null space of \\(\\mathrm{W}_{cls}^{\\top}\\), called the nullity, is nullity \\((\\mathrm{W}_{cls}^{\\top}) = d' - \\operatorname{rank}(\\mathrm{W}_{cls})\\) due to the rank-nullity theorem. Mathematically, this means that there exists a linear subspace \\(\\mathrm{Null}(\\mathrm{W}_{cls}^{\\top}) \\subset \\mathbb{R}^{d'}\\) such that if \\(\\delta \\in \\mathrm{Null}(\\mathrm{W}_{cls}^{\\top})\\) then \\(\\mathrm{W}_{cls}^{\\top} \\delta = 0\\). Within our OOD context, if an outlier instance, \\(\\mathbf{v}\\), has a feature representation, \\(h(\\mathbf{v})\\), such that its difference with respect to the inlier feature representation, \\(h(\\mathbf{x})\\), of any inlier instance, \\(\\mathbf{x}\\), is in the null space of the last layer of the classifier, then their free energy scores are the same:" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.617, + 0.483, + 0.637 + ], + "angle": 0, + "content": "\\[\nh (\\mathbf {v}) - h (\\mathbf {x}) \\in \\operatorname {N u l l} \\left(\\mathrm {W} _ {c l s} ^ {\\top}\\right) \\Rightarrow \\mathcal {F} (\\mathbf {x}) = \\mathcal {F} (\\mathbf {v}). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.484, + 0.81 + ], + "angle": 0, + "content": "Practically, Eq. (9) is telling us that it is possible for in-distribution and OOD instances to have different feature representations (satisfying the condition of Eq. (6)) and yet have the same free energy score, as depicted in Fig. 2. This fundamentally challenges the notion that free energy is an effective mechanism for OOD detection as any approaches reliant on Eq. (3), such as [7, 8, 18, 21, 29], contain inherent \"detection blind-spots\" as per Eq. (9). We denote Eq. (9) as the Null Space Vulnerabilities (NSV) of classifier \\( f \\) with size nullity \\( (\\mathrm{W}_{cls}^{\\top}) = d' - \\mathrm{rank}(\\mathrm{W}_{cls}) \\). An approach to minimize this vulnerability is outlined in Sec. 4.3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "As an additional remark, NSV are inherent in energy-based methods; while some post hoc methods use the projection of features onto the null space as a discriminator between in-distribution and OOD [4, 41], energy-based methods use the classification logits for detecting outliers, hence the null space projection does not change the OOD score." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.091, + 0.8, + 0.108 + ], + "angle": 0, + "content": "4.2. Minimal Change of Free Energy" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.113, + 0.907, + 0.219 + ], + "angle": 0, + "content": "We now turn our attention to the case where \\(\\delta\\) is not in the null space of \\(\\mathbf{W}_{cls}^{\\top}\\). Furthermore, we can focus on \\(\\delta\\) orthogonal to \\(\\mathrm{Null}(\\mathbf{W}_{cls}^{\\top})\\), since every vector in the domain of \\(\\mathbf{W}_{cls}^{\\top}\\) can be decomposed into one component in \\(\\mathrm{Null}(\\mathbf{W}_{cls}^{\\top})\\) and one component in \\(\\mathrm{Null}(\\mathbf{W}_{cls}^{\\top})^{\\perp}\\), and the effect of the component in \\(\\mathrm{Null}(\\mathbf{W}_{cls}^{\\top})\\) will be dealt with using the null space reduction method (Sec. 4.3)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.22, + 0.907, + 0.296 + ], + "angle": 0, + "content": "In order to create more discriminative OOD detectors based on the free energy score, we investigate cases where the difference in OOD and in-distribution free energy is minimal. We can state our problem formulation to find the case that minimizes the free energy difference:" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.306, + 0.906, + 0.395 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {\\substack {\\boldsymbol {\\delta}: \\| \\boldsymbol {\\delta} \\| \\geq d _ {b}, \\\\ \\boldsymbol {\\delta} \\in \\operatorname {Null} \\left(\\mathrm {W} _ {c l s} ^ {\\top}\\right) ^ {\\perp}}} | \\mathcal {F} (\\mathbf {v}) - \\mathcal {F} (\\mathbf {x}) | \\\\ = \\min _ {\\substack {\\boldsymbol {\\delta}: \\| \\boldsymbol {\\delta} \\| \\geq d _ {b}, \\\\ \\boldsymbol {\\delta} \\in \\operatorname {Null} \\left(\\mathrm {W} _ {c l s} ^ {\\top}\\right) ^ {\\perp}}} \\left| \\log \\frac {\\sum_ {k} ^ {K} \\exp \\left(f _ {k} (h (\\mathbf {x}))\\right)}{\\sum_ {k} ^ {K} \\exp \\left(f _ {k} (h (\\mathbf {x}) + \\boldsymbol {\\delta})\\right)} \\right|. \\end{array} \\tag{10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.405, + 0.907, + 0.543 + ], + "angle": 0, + "content": "Whilst in Eq. (10) we set \\(\\| \\delta \\| \\geq d_b\\) instead of \\(\\mathbb{E}\\left[\\| \\delta \\| \\right]\\geq\\) \\(d_{b}\\), in reality we are just interested in the case of non-zero \\(\\delta\\) since otherwise the solution would be trivial. Whilst the solution to Eq. (10) may be challenging to obtain, we note the difference in free-energies tends to 0 if the value inside the \\(\\log ()\\) function tends to 1. This is equivalent to making \\(f_{k}(h(\\mathbf{x}))\\) and \\(f_{k}(h(\\mathbf{x}) + \\delta)\\) similar for all \\(k\\). Therefore, considering that \\(f_{k}\\) is in the form of Eq. (5), our problem consists on finding the following proxy minimum:" + }, + { + "type": "equation", + "bbox": [ + 0.546, + 0.552, + 0.906, + 0.632 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{r l} \\underset {\\delta : \\| \\boldsymbol {\\delta} \\| \\geq d _ {b},} {\\min } & \\| \\mathrm {W} _ {c l s} ^ {\\top} (h (\\mathbf {x}) + \\boldsymbol {\\delta}) - \\mathrm {W} _ {c l s} ^ {\\top} h (\\mathbf {x}) \\| \\\\ & = \\underset {\\substack {\\boldsymbol {\\delta}: \\| \\boldsymbol {\\delta} \\| \\geq d _ {b}, \\\\ \\boldsymbol {\\delta} \\in \\operatorname {Null} \\left(\\mathrm {W} _ {c l s} ^ {\\top}\\right) ^ {\\perp}}} {\\min } \\| \\mathrm {W} _ {c l s} ^ {\\top} \\boldsymbol {\\delta} \\|. \\end{array} \\tag{11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.644, + 0.812, + 0.66 + ], + "angle": 0, + "content": "For the \\(L^2\\)-norm in Eq. (11), we thus obtain:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.67, + 0.907, + 0.708 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\substack {\\boldsymbol {\\delta}: \\| \\boldsymbol {\\delta} \\| _ {2} \\geq d _ {b}, \\\\ \\boldsymbol {\\delta} \\in \\operatorname {Null} \\left(\\mathrm {W} _ {c l s} ^ {\\top}\\right) ^ {\\perp}}} \\| \\mathrm {W} _ {c l s} ^ {\\top} \\boldsymbol {\\delta} \\| _ {2} = d _ {b} \\sigma_ {m i n} \\left(\\mathrm {W} _ {c l s}\\right), \\tag{12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.75 + ], + "angle": 0, + "content": "where \\(\\sigma_{\\min}(\\mathrm{W}_{cls})\\) is the least singular value of \\(\\mathrm{W}_{cls}\\) (see further mathematical formulation details in Supp. Mat.)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Subsequently, Eq. (12) holds for the case where \\(\\delta\\) is in the direction corresponding to the least singular value, such that the change in the free energy will be smaller than any other direction for \\(\\delta\\) (and considering \\(\\delta \\in \\mathrm{Null}(\\mathrm{W}_{cls}^{\\top})^{\\perp}\\)). We name this result (from Eq. (12)) the Least Singular Value Vulnerabilities (LSVV) of \\(f\\). In our OOD context, it means that there could exist some outliers far from the inlier distribution, in terms of Eq. (6), with similar free energy scores because they may fall in the direction of the least singular value of \\(\\mathrm{W}_{cls}\\). Again, this fundamentally challenges" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4532" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.137 + ], + "angle": 0, + "content": "the notion that free energy is an effective mechanism for OOD as free energy has been shown to be non-linear with respect to a given feature representation, \\( h(\\cdot) \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.138, + 0.484, + 0.258 + ], + "angle": 0, + "content": "This analysis is itself inspired by the seminal adversarial attacks analysis of Simon-Gabriel et al. [37], where they instead maximize the change in the classifier response for small input perturbations. Whilst here we are modelling the opposite problem, (i.e., the minimum change in the classifier for perturbations larger than a non-zero value), we could consider this analysis as an effective adversarial attack mechanism for the OOD domain." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.268, + 0.43, + 0.284 + ], + "angle": 0, + "content": "4.3. Null Space Reduction & Regularisation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.29, + 0.484, + 0.365 + ], + "angle": 0, + "content": "In this section, we introduce our Free Energy Vulnerability Elimination for Robust OOD Detection (FEVER-OOD) framework, where we explore readily available methods to minimize the NSV (Sec. 4.1) and LSVV (Sec. 4.2) impact potential within free energy score based OOD approaches." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.367, + 0.484, + 0.501 + ], + "angle": 0, + "content": "Null Space Reduction (NSR). For NSV, we propose to reduce the null space by adding an extra linear layer \\( g' : \\mathbb{R}^{d'} \\to \\mathbb{R}^r \\), where we choose \\( r < d' \\) to reduce the size of the NSV (Fig. 3). We denote this method as \\( r \\)-null space reduction (\\( r \\)-NSR). Therefore, the new classifier becomes \\( f' = g \\circ g' \\circ h \\) such that the feature representation of a sample is in \\( \\mathbb{R}^r \\) and is given instead by \\( (g' \\circ h)(\\mathbf{x}) \\). Effectively, within an OOD context, we are promoting better discrimination between OOD and in-distribution instances." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.502, + 0.484, + 0.547 + ], + "angle": 0, + "content": "Least Singular Value Regulariser (LSVR). To minimize the impact of LSVV, we maximise \\(\\sigma_{min}(\\mathrm{W}_{cls})\\) by adding an LSVR to the loss function \\(\\mathcal{L}\\) of a baseline method:" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.559, + 0.483, + 0.577 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {L S V} := \\mathcal {L} + \\lambda_ {L S V} \\sigma_ {\\min } ^ {- 1} \\left(\\mathrm {W} _ {c l s}\\right), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.588, + 0.484, + 0.693 + ], + "angle": 0, + "content": "where \\(\\lambda_{LSV}\\) is a scalar hyperparameter that controls the contribution of the LSV regulariser. It is additionally noted that alternatively maximizing the LSV might also increase the rest of the singular values, resulting in more significant changes of energy in all directions. In the context of OOD detection, LSVR increases the discriminative capacity for OOD detection of baseline methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.694, + 0.484, + 0.77 + ], + "angle": 0, + "content": "Condition Number (CN) Regulariser. Whilst the LSV regulariser will increase the change in energy for \\(\\delta\\) in the LSV direction, there is no constraint with regard to the distribution over all other directions. For this reason, we explore a Condition Number Regulariser (CNR):" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.782, + 0.483, + 0.799 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C N} := \\mathcal {L} + \\lambda_ {C N} \\kappa \\left(\\mathrm {W} _ {\\text {c l s}}\\right), \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.901 + ], + "angle": 0, + "content": "where \\(\\kappa (\\mathrm{W}_{cls})\\) is the condition number of the WCLS matrix, i.e., the ratio of the greatest and least singular values. Intuitively, this will encourage equalized energy changes for displacements in all directions where we note that Eq. (14) similarly assumes use of the \\(L^2\\)-norm as per Eq. (12). A different approach to the CNR is the orthogonal constraint" + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.089, + 0.887, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.303, + 0.906, + 0.332 + ], + "angle": 0, + "content": "Figure 3. Null space reduction. We add an extra layer \\( g' \\) to create \\( r \\)-dimensional features, reducing the size of the null space." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.342, + 0.906, + 0.435 + ], + "angle": 0, + "content": "\\(\\| W_{cls}W_{cls}^{\\top} - I\\|\\). However, this method adds the additional constraint that the singular values must be close to one, which might be a stronger prior than the CNR, hence we opt for the latter. The gradients for the singular values in Eqs. (13) and (14) are detailed by Lewis and Sendov [26] (Proposition 6.2) and further discussed in the Supp. Mat." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.444, + 0.716, + 0.46 + ], + "angle": 0, + "content": "5. Experimental Details" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.47, + 0.906, + 0.516 + ], + "angle": 0, + "content": "We study the impact of FEVER-OOD on free energy baseline methods (Sec. 5.1) across different standard datasets (Sec. 5.2). Implementation details are given in Sec. 5.3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.53, + 0.871, + 0.544 + ], + "angle": 0, + "content": "5.1. Baseline Methods and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.552, + 0.906, + 0.734 + ], + "angle": 0, + "content": "Three free energy baseline methods for image OOD detection that use the free energy score are tested: VOS [7], FFS [21] and Dream-OOD [8]. These baseline methods learn a small MLP \\(\\psi : \\mathbb{R} \\to (0,1)\\) that uses the free energy in Eq. (4) to predict an uncertainty score. Since OOD data is unavailable during training, these methods use synthetic outliers to train \\(\\psi\\). While VOS and FFS synthesize outliers in the feature space \\(h(\\mathbf{x})\\), Dream-OOD generates outliers in the pixel space using Stable Diffusion [34]. Additionally, VOS and FFS are also implemented as object detectors for object-level OOD detection, synthesizing outliers before the last layer of the classification head of a Faster RCNN [33]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We follow standard evaluation metrics, reporting the false positive rate for OOD detection at \\(95\\%\\) true positive rate of in-distribution detection (FPR95) and the area under the receiver operating characteristic curve (AUROC). We also assess the effect of our proposed techniques into the main in-distribution task (accuracy for classification and MS-COCO [28] mean average precision (mAP) for object detection). FPR95 and AUROC metrics are reported using a single intersection-over-union threshold for object detection, corresponding to the one that maximises the F-1 score, as described by Harakeh and Waslander [13]." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4533" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.141, + 0.089, + 0.856, + 0.104 + ], + "angle": 0, + "content": "Table 1. CIFAR-10 Results (ID Acc. = in-distribution accuracy; Null Space Reduction (NSR) methods = our approach)." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.108, + 0.907, + 0.283 + ], + "angle": 0, + "content": "
MethodFEVER-OODOOD DatasetsID Acc
TexturesSHVNPlaces365LSUNiSUNAvg
r-NSRRegularizerFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROC
VOS--50.05±5.7086.79±1.2839.15±10.5291.73±2.8840.88±1.7589.54±0.698.22±1.0298.36±0.2030.01±4.4494.31±0.8433.66±2.2192.15±0.5994.83±0.16
-λLSV=1.055.58±6.3586.71±2.3543.00±3.1692.09±0.9438.78±2.5191.14±0.567.36±1.5998.46±0.3221.23±6.9596.23±1.1733.19±3.2092.92±0.8594.66±0.18
-λCN=0.150.80±3.0685.39±0.7930.78±2.9594.04±0.5841.60±1.4088.77±0.748.50±0.8298.30±0.1030.91±4.5493.41±1.6832.52±1.2591.98±0.4694.69±0.16
96λLSV=1.040.86±4.8392.12±1.0541.92±16.1193.20±2.6135.69±2.2992.61±0.574.95±1.4998.93±0.2417.68±5.2697.04±0.8528.22±4.1894.78±0.8594.74±0.17
64λLSV=1.044.41±5.3991.26±1.2934.68±10.9394.39±1.7235.85±2.5692.42±0.535.68±0.5798.75±0.1324.03±5.9196.02±0.9528.93±3.8694.57±0.6694.75±0.13
32λLSV=0.00146.53±3.7789.67±1.0927.84±9.4795.29±1.0837.33±1.6391.77±0.395.65±0.7798.82±0.1124.09±6.7795.68±1.3528.29±3.3394.25±0.5594.68±0.07
10-53.20±3.9088.97±0.7335.62±9.6694.41±1.2845.73±6.2989.73±1.6911.72±3.3997.93±0.4641.80±14.3592.51±3.0337.61±4.0092.71±0.9791.95±2.32
FFS--52.86±4.4983.47±1.6738.67±11.2189.74±5.1944.65±1.2987.47±0.776.59±0.9298.67±0.1731.34±1.8893.24±0.9734.82±2.0390.52±0.9894.69±0.15
-λLSV=0.00150.74±3.9884.93±0.7532.25±12.1692.97±3.6042.99±2.3088.30±1.105.76±1.2398.82±0.2728.16±6.4594.18±1.9831.98±3.1091.84±1.2894.73±0.12
-λCN=1.050.08±3.8684.92±1.8530.91±6.1992.77±2.3345.69±2.5987.16±0.756.22±0.6598.73±0.1126.60±3.6494.80±0.6431.90±2.5191.67±0.9494.85±0.20
96λLSV=1.048.11±5.1589.84±1.5140.65±15.4793.32±3.1636.37±2.2892.29±0.605.22±1.3598.89±0.1824.84±10.0895.82±1.6831.04±5.0594.03±1.0694.71±0.08
64λLSV=1.051.70±5.6788.81±1.0933.02±14.0994.60±2.3039.05±3.7391.53±1.235.56±1.9598.84±0.3026.63±7.5195.65±1.1231.19±4.3293.89±0.6994.68±0.08
32-46.75±3.4989.61±1.4031.77±8.7394.44±1.3541.51±3.3890.72±1.024.93±1.1199.04±0.1628.92±11.2394.49±2.9430.78±3.7093.66±0.9594.79±0.18
10-54.72±5.8984.77±3.2345.70±12.9591.49±2.2741.52±3.2789.51±1.318.21±3.5998.33±0.6728.58±7.7594.30±1.8235.75±5.2691.68±1.3894.54±0.17
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.294, + 0.194, + 0.308 + ], + "angle": 0, + "content": "5.2. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.316, + 0.486, + 0.575 + ], + "angle": 0, + "content": "We assess FEVER-OOD with standard datasets for both classification and object detection. For image-level OOD detection (i.e., image classification), we train VOS and FFS using the CIFAR-10 and CIFAR-100 [20] datasets as in-distribution while testing OOD detection in five non-overlapping datasets: Textures [3], SVHN [30], Places365 [47], LSUN [45] and iSUN [11]. We also test Dream-OOD with the CIFAR-100 dataset as in-distribution data with the same OOD datasets. Additionally, we also train Dream-OOD with the Imagenet-100 dataset, a 100 classes partition of Imagenet [35], as in-distribution while testing OOD detection in four datasets: Textures [3], Places365 [47], iNaturalist [40] and SUN [42]. We use the PASCAL VOC [10] dataset as in-distribution for object-level OOD detection using VOS and FFS and test for OOD detection on the MS-COCO dataset [28] removing images containing objects with overlapping in-distribution categories [7]." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.584, + 0.307, + 0.599 + ], + "angle": 0, + "content": "5.3. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.6, + 0.486, + 0.902 + ], + "angle": 0, + "content": "We follow the established architectural choices of prior works. For image-level OOD detection with VOS and FFS, we use a WideResNet-40 [46] with a 128-dimensional feature space. We choose different NSR values ranging from the feature dim to the number of classes, gradually decreasing the nullity. For CIFAR-10, we evaluate the \\(\\{96,64,32,10\\}\\)-NSR, while for CIFAR-100 we assess \\(\\{114,100\\}\\)-NSR. With respect to Dream-OOD models, we use a ResNet-34 [14] with 512-dimensional feature space, evaluating \\(\\{256,128,100\\}\\)-NSR for both CIFAR-100 and Imagenet-100. Similarly, we reduce the feature space of the classification head to 768, 512 and 265 dimensions for VOS and FFS object detection experiments, having an original 1,024-dimensional object-wise feature space (as per Faster RCNN [33]). With regards to the LSVR and CNR, we test with various loss weight values in all the architectures (1, 0.1, 0.01, and 0.001), to evaluate their impact on model performance (except for Imagenet-100, where we use \\(\\lambda_{LSV} = 0.001\\) and \\(\\lambda_{CN} = 0.01\\) based on CIFAR-100 experiments). We train all the models using the original" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.295, + 0.907, + 0.385 + ], + "angle": 0, + "content": "corresponding baseline settings. Details on the hyperparameters and the training regime for each model are given in the Supp. Mat. To obtain reliable comparisons we use deterministic learning via a fixed random seed for all experiments. We run all experiments 5 times and report mean and standard deviation for all metrics." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.396, + 0.602, + 0.411 + ], + "angle": 0, + "content": "6. Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.416, + 0.907, + 0.538 + ], + "angle": 0, + "content": "We discuss the performance of FEVER-OOD against different free energy based OOD detection models in classification and detection. Tabs. 1 to 3 compare the use of FEVER-OOD with the corresponding baseline methods. We report different NSR as well as the best results obtained using \\(\\lambda_{LSV}\\) (Eq. (13)) or \\(\\lambda_{CN}\\) (Eq. (14)), and their corresponding value (we show both cases when no NSR is applied). Extensive ablations are available in the Supp. Mat." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.909, + 0.902 + ], + "angle": 0, + "content": "CIFAR-10. Tab. 1 shows the results of the VOS and FFS architectures with CIFAR-10 as in-distribution. In general, NSR and LSVR improve OOD detection for both VOS and FFS models. The best performance using the VOS method is achieved using a 96-NSR with LSVR, using a \\(\\lambda_{LSV} = 1.0\\), where the average AUROC increases to 94.78 and the average FPR95 reduces to \\(28.22\\%\\), compared to an AUROC of 92.15 and \\(33.66\\%\\) of the baseline VOS model without affecting the in-distribution accuracy. Similarly, the best FFS setting is achieved with a 96-NSR and \\(\\lambda_{LSV} = 1.0\\), resulting in 94.03 AUROC and \\(31.04\\%\\) FPR95, compared to the baseline FFS model, with an AUROC of 90.52 and FPR95 of \\(34.82\\%\\), although the best FPR95 is achieved with 32-NSR and no singular value regularization, with an FPR95 of 30.78. In general, it is seen that a large NSR (i.e., 10-NSR) may decrease the AUROC and increase the FPR95. Since VOS and FFS are techniques that synthesize outliers in the feature space, a large reduction of the feature dimension might not allow for proper in-distribution understanding, impacting in the OOD training of these techniques. Additionally, it is also observed that LSVR yields better results than CNR, and a value for \\(\\lambda_{LSV} = 1.0\\) is adequate for none to moderate NSR, while smaller values for LSVR (or none) are better for large NSR. While LSVR increases the differ" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4534" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.137, + 0.09, + 0.858, + 0.103 + ], + "angle": 0, + "content": "Table 2. CIFAR-100 Results (ID Acc. = in-distribution accuracy; Null Space Reduction (NSR) methods = our approach)." + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.105, + 0.882, + 0.269 + ], + "angle": 0, + "content": "
MethodFEVER-OODOOD DatasetsID Acc
r-NSRRegularizerTexturesSHVNPlaces365LSUNiSUNAvg
FPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROC
VOS--81.24±2.1176.01±0.8076.09±6.5183.18±3.5980.53±1.4676.25±0.7837.86±4.0693.21±0.6879.34±1.6076.43±1.4471.01±1.2381.02±0.4076.04±0.22
-λLSV=0.182.87±1.1176.49±0.9473.35±2.9684.57±1.5380.36±1.1676.89±0.7940.05±3.7492.83±0.4674.28±4.0380.25±3.3470.18±0.8782.21±0.5676.01±0.16
-λCN=0.0182.51±2.6476.04±1.9173.67±3.1085.31±1.2480.18±1.0876.59±0.2341.08±3.2792.75±0.6380.83±4.3773.79±5.5171.65±0.9680.90±1.1176.20±0.11
114λLSV=0.0180.02±2.7578.73±1.8383.45±2.5982.07±1.3678.41±0.9677.71±0.9328.90±3.6195.11±0.5872.07±9.0780.55±4.7068.57±2.2882.83±0.9175.72±0.16
100-80.47±2.3278.21±1.1479.10±7.4283.39±3.9179.19±1.0477.23±0.5028.99±2.5995.15±0.3975.51±6.6879.63±3.7068.65±2.2982.72±0.5775.54±0.19
FFS--82.87±1.3975.54±0.9376.32±6.1784.83±2.3681.14±1.0976.27±0.7436.27±4.9193.54±1.0382.83±2.9874.29±3.0871.89±1.6480.89±1.1176.04±0.11
-λLSV=0.0180.95±1.3476.53±0.7083.12±5.5981.61±2.9280.19±0.8476.58±0.4932.93±4.0594.20±0.5179.05±5.1677.28±3.2371.25±2.1181.24±0.7276.27±0.28
-λCN=0.00180.27±2.7676.78±1.5378.57±7.3282.57±2.9580.40±1.0976.40±0.5332.91±4.9094.06±1.0280.71±7.9074.49±5.5770.57±2.4180.86±0.8876.05±0.16
114λLSV=0.00179.26±2.6778.16±1.5074.28±7.1385.37±3.2079.45±0.8577.25±0.6124.10±1.4295.94±0.1773.61±5.8280.61±2.4966.14±1.9183.47±0.9875.45±0.37
100-77.69±2.9778.62±1.0277.84±9.6283.40±3.5280.41±0.7276.58±0.3521.91±0.9596.25±0.1179.60±6.0675.79±5.1767.49±2.0382.13±0.8575.48±0.28
Dream-OOD--62.20±1.0283.84±0.3873.05±1.9284.56±0.2177.95±1.9779.43±0.1739.90±2.0192.87±0.441.70±0.1199.58±0.0450.96±1.4488.06±0.6075.61±0.19
-λLSV=0.0158.45±2.2786.04±0.4168.75±1.8487.65±0.3077.45±2.1178.59±0.2215.45±2.5797.24±0.091.55±0.0399.63±0.4544.33±3.4889.83±0.1275.87±0.23
-λCN=0.00157.40±1.8886.28±0.3277.75±1.1185.13±0.4178.60±1.2378.73±0.1527.2±1.7795.01±0.421.55±0.0899.57±0.0748.5±1.5488.94±0.4376.32±0.14
256-60.00±3.2185.42±0.7967.50±2.4785.84±0.6675.90±1.1079.57±0.4319.85±1.0596.74±0.671.00±0.0299.78±0.0644.82±2.7289.47±0.24277.01±0.12
128λLSV=0.156.45±2.3987.73±1.0167.45±1.1287.53±0.4478.45±2.4677.24±0.6129.60±1.7394.81±0.401.95±0.1099.5±0.1046.78±2.7289.36±0.8376.21±0.35
100-57.55±2.4886.8±0.4454.45±3.0288.69±0.5975.80±1.2178.88±0.6924.45±1.5795.86±0.521.60±0.0799.65±0.2142.77±2.1089.98±0.1576.41±0.32
" + }, + { + "type": "table_caption", + "bbox": [ + 0.13, + 0.278, + 0.867, + 0.292 + ], + "angle": 0, + "content": "Table 3. ImageNet-100 Results (ID Acc. = in-distribution accuracy; Null Space Reduction (NSR) methods = our approach)." + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.293, + 0.882, + 0.384 + ], + "angle": 0, + "content": "
MethodFEVER-OODOOD DatasetsID Acc
r-NSRRegularizeriNaturalistPlaces365SUNTexturesAvg
FPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROC
Dream-OOD--23.98±2.0895.94±0.2641.75±1.7292.48±0.1240.85±1.1392.76±0.0950.73±0.8586.21±0.2939.33±1.0891.84±0.1287.76±0.19
-λLSV=0.0159.91±3.0990.89±0.4866.26±2.1788.12±0.4774.26±1.5284.53±0.3360.56±1.2183.91±0.4265.25±1.5086.87±0.2587.85±0.11
-λLSV=0.00159.71±3.5390.97±0.7065.01±1.9088.39±0.4573.93±1.5884.69±0.4960.51±1.4684.01±0.5164.79±1.6987.01±0.4587.77±0.13
256-23.84±1.7395.80±0.3044.35±1.7192.14±0.3042.92±1.9892.46±0.3344.38±1.8188.60±0.5438.87±1.2992.25±0.2387.57±0.24
128λLSV=0.0123.46±2.5495.93±0.3643.37±1.4792.19±0.2242.42±1.6092.44±0.2941.86±1.0989.48±0.2737.78±1.4592.51±0.2287.60±0.11
100λLSV=0.0122.24±0.8196.16±0.1441.08±3.0392.59±0.4440.39±2.8492.91±0.4042.31±0.9789.30±0.3736.50±1.6792.74±0.2187.42±0.15
" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.39, + 0.28, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.496, + 0.249, + 0.506 + ], + "angle": 0, + "content": "(a) Dataset: CIFAR-10" + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.39, + 0.475, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.316, + 0.496, + 0.445, + 0.506 + ], + "angle": 0, + "content": "(b) Dataset: CIFAR-100" + }, + { + "type": "image_caption", + "bbox": [ + 0.178, + 0.507, + 0.396, + 0.519 + ], + "angle": 0, + "content": "Figure 4. Singular Values Spectrum." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.523, + 0.484, + 0.702 + ], + "angle": 0, + "content": "ence in energy of outliers in or near the least singular vector direction, the CN regulariser makes equalized free energy changes for all directions. Since the latter effect might disable directions with smaller energy changes than others, it might also reduce the free energy change in the direction of the largest singular value. Fig. 4(a) shows the singular value spectrum of the baseline VOS model against the LSVR and CNR versions for the CIFAR-10 dataset. It is seen that the while both regularizers homogenize the singular values, the CNR also reduces the most significant singular value. Finally, our experiments show that VOS achieves better results than FFS in CIFAR-10 OOD detection." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.484, + 0.901 + ], + "angle": 0, + "content": "CIFAR-100. Tab. 2 shows the results for VOS, FFS and Dream-OOD trained on the CIFAR-100 dataset as in-distribution. The best models for VOS and FFS are achieved with 114-NSR and LSVR, with a \\(68.57\\%\\) FPR95 \\((-2.44\\%)\\) VOS) and 82.83 \\((+1.81\\mathrm{VOS})\\) AUROC for VOS-114-NSR and \\(66.14\\%\\) FPR95 \\((-5.75\\%)\\) FFS) and 83.47 \\((+2.58\\mathrm{FFS})\\) AUROC for FFS-114-NSR. However, despite these improvements, the performance of virtual outlier synthesis methods is usually poor for several in-distribution classes [44], hence the enhancement of these architectures is still challenging. Regarding Dream-OOD, it is noted that the best results are obtained by Dream-OOD-100-NSR, having an FPR95 of \\(42.77\\%\\) \\((-8.19\\%)\\) Dream-OOD) and 89.98" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.391, + 0.907, + 0.57 + ], + "angle": 0, + "content": "AUROC (+1.9 Dream-OOD). Finally, for the three energy techniques (VOS, FFS and Dream-OOD), we observed that the model does not converge when using 100-NSR and the CNR (see Supp. Mat.), suggesting that the conditioning number penalization does not allow the model to learn the in-distribution task. Fig. 4(b) shows the singular value spectrum for the LSVR and CNR versions of the VOS model. Differently from the CIFAR-10 experiments, it is seen that LSVR keeps the most significant singular values similar to the baseline but homogenizes the least significant singular values, while the CNR keeps all singular values at a low level, hence explaining why it fails for this dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.57, + 0.909, + 0.78 + ], + "angle": 0, + "content": "Imagenet-100. We show the results of Dream-OOD with Imagenet-100 as in-distribution in Tab. 3. Similar as with CIFAR-100, the best results are obtained with NSR100, effectively eliminating the null space of the last layer (since it becomes square and is not singular). Contrary to CIFAR-100, LSVR has a negative impact when used alone. Considering that we use the same pre-training scheme as Du et al. [8], adding LSVR and CNR on an already pretrained last linear layer might affect the already formed feature space. Hence, these regularisers benefit untrained NSR modules. The best model, Dream-OOD-100-NSR with LSVR, achieves an average \\(36.50\\%\\) FPR95 and 92.74 AUROC, compared to the baseline \\(39.33\\%\\) FPR95 and 91.84 AUROC, becoming the state-of-the-art for Imagenet-100." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "PASCAL VOC. We test FEVER-OOD techniques in object detection task, where the goal is to detect object-level OOD instances. Tab. 4 show the result for the baseline VOS and FFS, taking PASCAL VOC [10] as in-distribution dataset. For both tasks, we evaluate on two OOD datasets that contain subset of images from: MS-COCO [28] and OpenImages [22], following the original works [7, 21]. The best-performing model for VOS is VOS-768-NSR with CNR," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4535" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.089, + 0.27, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.27, + 0.091, + 0.414, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.415, + 0.091, + 0.723, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.723, + 0.092, + 0.885, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.178, + 0.268, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.268, + 0.178, + 0.414, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.415, + 0.178, + 0.723, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.723, + 0.178, + 0.89, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.268, + 0.905, + 0.296 + ], + "angle": 0, + "content": "Figure 5. MS-COCO objects detected on OOD images by VOS baseline [7] (first row) and Fever-OOD (second row). Blue: OOD objects detected and mis-classified as being in-distribution. Green: the same OOD objects correctly detected as OOD by FEVER-OOD (ours)." + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.303, + 0.445, + 0.331 + ], + "angle": 0, + "content": "Table 4. VOS results with PASCAL VOC as in-distribution (Null Space Reduction (NSR) methods = our approach)." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.338, + 0.482, + 0.611 + ], + "angle": 0, + "content": "
MethodFEVER-OODOOD DatasetsmAP
r-NSRλLSVλCNMS-COCOOpenImages
FPR95AUROCFPR95AUROC
VOS---50.2987.7753.0986.580.489
-0.001-49.9688.1052.2685.800.489
--153.4286.6652.7385.770.490
768--48.9487.9452.3486.330.489
7680.01-50.0887.7851.7186.040.493
768-0.0147.4788.1549.3686.320.491
512--49.8488.4255.3385.330.486
5120.001-49.2688.3954.3285.820.489
512-0.0149.5987.4854.8185.480.490
256--47.8888.4952.4186.390.486
2560.01-48.5188.3650.5286.120.491
256-0.00149.188.2751.4786.080.489
FFS---50.7787.1853.7885.340.487
-0.01-53.0986.7553.4785.860.490
--0.0150.3687.6551.1985.980.485
768--51.5087.5953.3385.840.488
7680.01-47.4288.3652.4686.400.489
768-0.0151.1286.8255.4484.840.490
512--46.9388.8253.8986.280.491
5120.01-47.4288.3652.4686.400.488
512-0.0151.4287.0554.9884.550.490
256--51.1888.7453.6786.560.487
2560.01-47.9388.0447.9384.950.489
256-0.0152.9486.6252.8884.450.489
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.621, + 0.483, + 0.834 + ], + "angle": 0, + "content": "achieving average FPR95 scores of \\(47.47\\%\\) and \\(49.36\\%\\) on MS-COCO and OpenImages, respectively, an improvement of nearly \\(4\\%\\) over the VOS baseline. For FFS, we achieve a minimum FPR95 of \\(46.93\\%\\) with FFS-512-NSR model and \\(47.93\\%\\) for the FFS-256-NSR model with the LSVR for both OOD datasets. Additionally, our approach generally delivers higher AUROC scores compared to the FFS baseline across both OOD datasets. Qualitative results are illustrated in Fig. 5, where we visualize the prediction on MS-COCO OOD images, using the baseline VOS (top) and the best Fever-OOD model, i.e., VOS-768-NSR, with the PASCAL-VOC as the in-distribution dataset, showing that out framework outperforms baseline VOS in in detecting outlier objects with a significantly lower false positive rate." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Computational overhead. The computational overhead introduced by our method is minimal since we only add an extra linear layer. For instance, the ResNet-34 used in Dream-OOD takes \\( 4.42 \\pm 1.15 \\) ms/img and has 21.5M" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.305, + 0.906, + 0.365 + ], + "angle": 0, + "content": "params. Adding 256, 128, 100-NSR layers change the inference time to \\(3.91 \\pm 0.62\\), \\(4.44 \\pm 1.17\\), \\(4.13 \\pm 1.02\\) ms/img, with an extra 105k, 27k, 10k params, which is negligible compared with the total size of ResNet-34." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.371, + 0.634, + 0.385 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.388, + 0.908, + 0.902 + ], + "angle": 0, + "content": "In this work, we present FEVER-OOD, a robust approach for addressing critical vulnerabilities found in existing energy-based OOD detection. Our method targets the Null Space Vulnerabilities (NSV), where in-distribution and OOD instances can produce similar free energy scores despite differing feature representations, and Least Singular Value Vulnerabilities (LSVV), where energy score similarities are influenced by the least singular vector of the classifier last linear layer. These vulnerabilities are especially problematic in high-dimensional feature spaces where the feature dimension exceeds the number of classes, impacting the effectiveness of OOD detection. To mitigate these issues, FEVER-OOD incorporates strategies such as reducing the null space through lower-dimensional feature spaces and introducing a Least Singular Value Regularizer to maximize the least singular value, enhancing the distinction between energy scores of different samples. Additionally, we propose a Condition Number Regularizer to promote uniform energy changes for feature-space displacements, fostering consistent energy variation and improving detection robustness. Our comprehensive evaluation on benchmark datasets, including applications to the Dream-OOD framework with ImageNet-100 as the ID dataset, demonstrates the significant effectiveness of FEVER-OOD. Specifically, our approach achieves a \\(10.13\\%\\) reduction in the average false positive rate and a \\(1.6\\%\\) increase in AUROC, setting a new state-of-the-art in OOD detection. Furthermore, our results highlight the generalization capability of FEVER-OOD across different detection architectures, showcasing its potential beyond conventional classification settings. Future extensions of FEVER-OOD could include applications to varied domains and integration with other OOD detection frameworks to further validate its versatility and performance in diverse machine learning contexts. This work provides valuable insights into enhancing OOD robustness and sets a foundation for more resilient free energy-based detection methods." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4536" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.105, + 0.188, + 0.12 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.13, + 0.483, + 0.198 + ], + "angle": 0, + "content": "[1] Haoyue Bai, Gregory Canal, Xuefeng Du, Jeongyeol Kwon, Robert D Nowak, and Yixuan Li. Feed two birds with one scone: Exploiting wild data for both out-of-distribution generalization and detection. In International Conference on Machine Learning, pages 1454-1471. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.201, + 0.483, + 0.241 + ], + "angle": 0, + "content": "[2] Haoyue Bai, Jifan Zhang, and Robert Nowak. Aha: Human-assisted out-of-distribution generalization and detection. arXiv preprint arXiv:2410.08000, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.243, + 0.482, + 0.31 + ], + "angle": 0, + "content": "[3] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3606-3613, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.313, + 0.483, + 0.353 + ], + "angle": 0, + "content": "[4] Matthew Cook, Alina Zare, and Paul Gader. Outlier detection through null space analysis of neural networks. ICML Workshops, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.355, + 0.483, + 0.45 + ], + "angle": 0, + "content": "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.453, + 0.483, + 0.521 + ], + "angle": 0, + "content": "[6] Xuefeng Du, Xin Wang, Gabriel Gozum, and Yixuan Li. Unknown-aware object detection: Learning what you don't know from videos in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13678-13688, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.523, + 0.483, + 0.577 + ], + "angle": 0, + "content": "[7] Xuefeng Du, Zhaoning Wang, Mu Cai, and Yixuan Li. Vos: Learning what you don't know by virtual outlier synthesis. In International Conference on Learning Representations, 2022. 1, 2, 3, 4, 5, 6, 7, 8, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.58, + 0.483, + 0.632 + ], + "angle": 0, + "content": "[8] Xuefeng Du, Yiyou Sun, Xiaojin Zhu, and Yixuan Li. Dream the impossible: Outlier imagination with diffusion models. In Advances in Neural Information Processing Systems, 2023. 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.635, + 0.483, + 0.676 + ], + "angle": 0, + "content": "[9] Xuefeng Du, Zhen Fang, Ilias Diakonikolas, and Yixuan Li. How does unlabeled data provably help out-of-distribution detection? arXiv preprint arXiv:2402.03502, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.677, + 0.483, + 0.732 + ], + "angle": 0, + "content": "[10] Mark Everingham, Luc Van Gool, Christopher K. I. Williams, John M. Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International Journal of Computer Vision, 88(2):303-338, 2010. 6, 7, 2, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.734, + 0.483, + 0.801 + ], + "angle": 0, + "content": "[11] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 580-587, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[12] Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep Learning. MIT Press, 2016. http://www.deeplearningbook.org.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[13] Ali Harakeh and Steven L. Waslander. Estimating and evaluating regression predictive uncertainty in deep object detectors. In International Conference on Learning Representations, 2021. 5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.13, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 3, 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[15] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In International Conference on Learning Representations, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.907, + 0.275 + ], + "angle": 0, + "content": "[16] Yen-Chang Hsu, Yilin Shen, Hongxia Jin, and Zsolt Kira. Generalized odin: Detecting out-of-distribution image without learning from out-of-distribution data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10951-10960, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[17] Rui Huang, Andrew Geng, and Yixuan Li. On the importance of gradients for detecting distributional shifts in the wild. Advances in Neural Information Processing Systems, 34:677-689, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[18] B.K.S. Isaac-Medina, Y.F.A. Gaus, N. Bhowmik, and T.P. Breckon. Towards open-world object-based anomaly detection via self-supervised outlier synthesis. In Proc. European Conference on Computer Vision. Springer, 2024. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[19] Julian Katz-Samuels, Julia B Nakhleh, Robert Nowak, and Yixuan Li. Training ood detectors in their natural habitats. In International Conference on Machine Learning, pages 10848-10865. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[20] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009. 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.491, + 0.907, + 0.572 + ], + "angle": 0, + "content": "[21] Nishant Kumar, Siniša Šegvic, Abouzar Eslami, and Stefan Gumhold. Normalizing flow based feature synthesis for outlier-aware object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5156-5165, 2023. 1, 2, 3, 4, 5, 7, 6, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.576, + 0.905, + 0.671 + ], + "angle": 0, + "content": "[22] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision, 128(7):1956-1981, 2020. 7, 6, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.674, + 0.905, + 0.729 + ], + "angle": 0, + "content": "[23] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. Advances in Neural Information Processing Systems, 30, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.731, + 0.905, + 0.772 + ], + "angle": 0, + "content": "[24] Hao Lang, Yinhe Zheng, Yixuan Li, Jian SUN, Fei Huang, and Yongbin Li. A survey on out-of-distribution detection in NLP. Transactions on Machine Learning Research, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.774, + 0.905, + 0.828 + ], + "angle": 0, + "content": "[25] Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. Advances in Neural Information Processing Systems, 31, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[26] Adrian S Lewis and Hristo S Sendov. Nonsmooth analysis of singular values. part i: Theory. Set-Valued Analysis, 13: 213–241, 2005. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Shiyu Liang, Yixuan Li, and R Srikant. Enhancing the reliability of out-of-distribution image detection in neural net" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4537" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "works. In 6th International Conference on Learning Representations, 2018, 2018. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[28] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5, 6, 7, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[29] Weitang Liu, Xiaoyun Wang, John Owens, and Yixuan Li. Energy-based out-of-distribution detection. Advances in Neural Information Processing Systems, 33:21464-21475, 2020. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[30] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading digits in natural images with unsupervised feature learning. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning 2011, 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.398 + ], + "angle": 0, + "content": "[31] Anh Nguyen, Jason Yosinski, and Jeff Clune. Deep neural networks are easily fooled: High confidence predictions for unrecognizable images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 427-436, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.455 + ], + "angle": 0, + "content": "[32] Bo Peng, Yadan Luo, Yonggang Zhang, Yixuan Li, and Zhen Fang. Conjnorm: Tractable density estimation for out-of-distribution detection. In The Twelfth International Conference on Learning Representations, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.457, + 0.482, + 0.537 + ], + "angle": 0, + "content": "[33] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: towards real-time object detection with region proposal networks. In Proceedings of the 28th International Conference on Neural Information Processing Systems - Volume 1, page 91-99, Cambridge, MA, USA, 2015. MIT Press. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.483, + 0.609 + ], + "angle": 0, + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.61, + 0.482, + 0.691 + ], + "angle": 0, + "content": "[35] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 115 (3):211-252, 2015. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.693, + 0.482, + 0.774 + ], + "angle": 0, + "content": "[36] Mohammadreza Salehi, Hossein Mirzaei, Dan Hendrycks, Yixuan Li, Mohammad Hossein Rohban, and Mohammad Sabokrou. A unified survey on anomaly, novelty, open-set, and out of-distribution detection: Solutions and future challenges. Transactions on Machine Learning Research, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[37] Carl-Johann Simon-Gabriel, Yann Ollivier, Leon Bottou, Bernhard Schölkopf, and David Lopez-Paz. First-order adversarial vulnerability of neural networks and input dimension. In Proceedings of the 36th International Conference on Machine Learning, pages 5809-5817. PMLR, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[38] Yiyou Sun, Yifei Ming, Xiaojin Zhu, and Yixuan Li. Out-of-distribution detection with deep nearest neighbors. In International Conference on Machine Learning, pages 20827-20840. PMLR, 2022. 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.135 + ], + "angle": 0, + "content": "[39] Leitian Tao, Xuefeng Du, Jerry Zhu, and Yixuan Li. Nonparametric outlier synthesis. In The Eleventh International Conference on Learning Representations, 2023. 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.906, + 0.218 + ], + "angle": 0, + "content": "[40] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8769-8778, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[41] Haoqi Wang, Zhizhong Li, Litong Feng, and Wayne Zhang. Vim: Out-of-distribution with virtual-logit matching. In CVPR, pages 4921-4930, 2022. 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[42] Jianxiong Xiao, James Hays, Krista A. Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 3485-3492, 2010. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.333, + 0.905, + 0.401 + ], + "angle": 0, + "content": "[43] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10687-10698, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.403, + 0.905, + 0.457 + ], + "angle": 0, + "content": "[44] Jingkang Yang, Kaiyang Zhou, Yixuan Li, and Ziwei Liu. Generalized out-of-distribution detection: A survey. International Journal of Computer Vision, pages 1-28, 2024. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.46, + 0.905, + 0.515 + ], + "angle": 0, + "content": "[45] Fisher Yu, Yinda Zhang, Shuran Song, Ari Seff, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.517, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[46] Erwan Zerhouni, David Lanyi, Matheus Viana, and Maria Gabrani. Wide residual networks for mitosis detection. In 2017 IEEE 14th International Symposium on Biomedical Imaging (ISBI 2017), pages 924-928. IEEE, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.574, + 0.905, + 0.615 + ], + "angle": 0, + "content": "[47] Bolei Zhou, Agata Lapedriza, Antonio Torralba, and Aude Oliva. Places: An image database for deep scene understanding. Journal of Vision, 17(10):296-296, 2017. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.615 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4538" + } + ] +] \ No newline at end of file diff --git a/2025/FEVER-OOD_ Free Energy Vulnerability Elimination for Robust Out-of-Distribution Detection/0c16965c-d6fa-4547-a6ba-ab34512ce472_origin.pdf b/2025/FEVER-OOD_ Free Energy Vulnerability Elimination for Robust Out-of-Distribution Detection/0c16965c-d6fa-4547-a6ba-ab34512ce472_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..546c5eb19ee36d892ee6fd3032e622c187fdc239 --- /dev/null +++ b/2025/FEVER-OOD_ Free Energy Vulnerability Elimination for Robust Out-of-Distribution Detection/0c16965c-d6fa-4547-a6ba-ab34512ce472_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a822b077f368b7c48330cafce9530eda4e1b480ee5acec541c6cca5370879728 +size 6553072 diff --git a/2025/FEVER-OOD_ Free Energy Vulnerability Elimination for Robust Out-of-Distribution Detection/full.md b/2025/FEVER-OOD_ Free Energy Vulnerability Elimination for Robust Out-of-Distribution Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e721073a0deddf63c7e882ba4c124c6459dfa31f --- /dev/null +++ b/2025/FEVER-OOD_ Free Energy Vulnerability Elimination for Robust Out-of-Distribution Detection/full.md @@ -0,0 +1,319 @@ +# FEVER-OOD: Free Energy Vulnerability Elimination for Robust Out-of-Distribution Detection + +Brian K.S. Isaac-Medina1 Maurizio Che2 Yona Faline A. Gaus1 Samet Akcay3 Toby P. Breckon1 +1Durham University, Durham, UK 2University of Vienna, Vienna, Austria 3Intel + +{brian.k.isaac-medina, yona.f.binti-abd-gaus, toby.breckon}@durham.ac.uk mauricio.adrian.che.moguel@univie.ac.at samet.akcay@intel.com + +# Abstract + +Modern machine learning models, that excel on computer vision tasks such as classification and object detection, are often overconfident in their predictions for Out-of-Distribution (OOD) examples, resulting in unpredictable behaviour for open-set environments. Recent works have demonstrated that the free energy score is an effective measure of uncertainty for OOD detection given its close relationship to the data distribution. However, despite that free energy-based methods representing a significant empirical advance in OOD detection, our theoretical analysis reveals previously unexplored and inherent vulnerabilities within the free energy score formulation such that in-distribution and OOD instances can have distinct feature representations yet identical free energy scores. This phenomenon occurs when the vector representing the feature space difference between the in-distribution and OOD sample belongs to the null space of the last layer of a neural-based classifier. To mitigate these issues, we explore lower-dimensional feature spaces to reduce the null space footprint and introduce novel regularisation to maximize the least singular value of the final linear layer, hence enhancing intersample free energy separation. We refer to these techniques as Free Energy Vulnerability Elimination for Robust Out-of-Distribution Detection (FEVER-OOD). Extensive experimentation shows that FEVER-OOD improves OOD detection in CIFAR-10 with $28.22\%$ FPR95 and 94.78 AUROC vs $33.66\%$ FPR95 and 92.15 AUROC of the baseline model, CIFAR-100 with $42.77\%$ FPR95 and 89.98 AUROC vs $50.96\%$ FPR95 and 88.06 AUROC of the baseline model, and achieves state of the art OOD detection in Imagenet-100, with average OOD FPR95 of $36.50\%$ and an AUROC of 92.74 when used with the Dream-OOD model, compared with a $39.33\%$ and 91.84 AUROC without FEVER-OOD. + +# 1. Introduction + +Out-of-distribution (OOD) detection (c.f. anomaly/outlier detection), aims to identify abnormal samples that deviate significantly from a given data distribution. Whilst + +![](images/848d7e663e2933ff92e896c599e56815165ef8b69813df492dfc4ac6e02b7a86.jpg) +Free Energy-based OOD (VOS) + +![](images/b79fe209468046edba6d570b17932480933ce0e0db970c3334cea190f38ac4df.jpg) +FEVER-OOD (VOS) +Figure 1. FEVER-OOD improves baseline Free Energy-based OOD detection methods, where baseline methods predict OOD classes as in-distribution (left, blue), while FEVER-OOD classifies them as OOD (right, green). + +OOD is essential in developing deployable machine learning systems, to ensure valid inference only on data drawn from same distribution used in training [24, 36, 44], real-world deployments often involve scenarios where models encounter unseen OOD classes that could otherwise undermine decision stability and hence wider system robustness. + +Methods such as deep ensembles [23], ODIN [27], Mahalanobis distance-based detection [25], and generalized ODIN [16] all utilize softmax() to differentiate indistribution from OOD. Whilst intuitive, such approaches often struggle as deep neural networks commonly yield overly high softmax() scores, even for inputs far from the training data distribution [31]. Given OOD and indistribution samples should have significant spatial separation in any given feature space, Liu et al. [29] developed an alternative scoring function to effectively measure the alignment of a given sample to an a priori learned distribution. Denoted as the free energy scoring function, this formulation demonstrates superior OOD performance across both classification [7-9] and detection tasks [6, 7, 18, 21]. + +Despite free energy-based methods representing a significant empirical advance in OOD detection, our theoretical analysis reveals previously unexplored and inherent vulnerabilities within the free energy score formulation that could impact their practical reliability. These vulnerabilities reveal that in-distribution and OOD instances can have distinct feature representations yet produce identical free en + +ergy scores. This phenomenon occurs in the feature space when the vector direction representing the difference between an in-distribution and an OOD sample lies within the null space of the last layer of neural-based classifier. Such regions, effectively "blind spots" in terms of OOD, lead to OOD and in-distribution generating identical or near-identical free energy score values. + +To mitigate these issues, we explore lower-dimensional feature spaces to effectively reduce the size of the null space. In addition, we also introduce a novel regulariser to maximize the least singular value of the final linear layer, hence enhancing inter-sample free energy separation to improve OOD robustness. We refer to these techniques as Free energy Vulnerability Elimination for Robust Out-of-Distribution Detection (FEVER-OOD). + +In summary, our key contributions in this paper are: + +- The identification of inherent free energy score vulnerabilities, where under null-space conditions, in-distribution and OOD instances will yield similar free energy scores despite differing feature representations. Denoted as Null Space Vulnerabilities (NSV) such conditions readily occur when the feature dimension exceeds the number of classes. Additionally, such free energy similarity can be influenced by the least singular vector (LSV) of the last linear layer of a neural-based classifier, leading to what we denote as Least Singular Value Vulnerabilities (LSVV). Both vulnerabilities negatively impact in-distribution to OOD discrimination. +- Novel formulations that mitigate the NSV and LSVV impact (Fig. 1). To address NSV, we introduce an additional layer to reduce the null space whilst for LSVV, we develop an LSV Regulariser that maximizes the least singular value, guaranteeing detectable energy variations from small feature-space differences and hence enhancing inter-sample energy scores discrimination. Furthermore, we explore a Condition Number Regulariser to promote balanced energy changes for displacements in all feature-space direction hence ensuring improved uniformity in energy score variations. +- Comprehensive experiments and ablation studies using established benchmark datasets over both object classification and detection tasks, that demonstrate the effectiveness of FEVER-OOD in improving OOD detection performance. Specifically, Dream-OOD [8] with FEVER-OOD trained on Imagenet-100 [35] as in-distribution decreases the FPR95 from $39.33\%$ to $36.50\%$ and increases the average AUROC from 91.84 to 92.74, achieving state-of-the-art performance for OOD detection. + +# 2. Related Work + +Identifying OOD samples is a crucial yet inherently challenging problem, as models are not exposed to, and hence cannot reliably differentiate, OOD samples. + +OOD Detection: to overcome this challenge, researchers have begun incorporating additional data during training to establish a more conservative and safe decision boundary against OOD inputs. Some methods [1, 19] have used unlabeled data to regularise model training while still focusing on classifying labeled in-distribution data, while others [2] leverage human assistance to strategically label examples within a novel maximum disambiguation region. Despite its potential, leveraging unlabeled data is non-trivial due to the heterogeneous mix of in-distribution and OOD instances. The absence of a clean OOD set presents significant challenges in designing effective OOD learning algorithms. A different branch of works incorporate post-hoc methods, focusing on designing OOD scoring mechanisms at test time. Hendrycks and Gimpel [15] introduced the maximum predicted softmax() probability as an in-distribution score function, establishing an initial baseline for scoring OOD detection. Similarly, ODIN [27] aims to widen the gap between values in the softmax() score vectors by temperature scaling and gradient-based input perturbation, enhancing the separation between in-distribution and OOD instances. Lee et al. [25] use the Mahalanobis distance between the average softmax() vectors of images and incorporate intermediate features to improve detection performance and robustness. Nonetheless, the misalignment between the softmax() score and the true data density [29, 32] ultimately makes these techniques a suboptimal solution to OOD detection. Closer to our work, Cook et al. [4] identify null space adversarial attacks, proposing that instances with large magnitude of their projection onto the span of the null space of a neural network classifier would likely be an outlier. Similarly, Wang et al. [41] propose that since the features lie in low-dimensional manifolds, outliers are likely to deviate from the principal subspace, i.e. the span of the eigenvectors or the largest eigenvalues, of the in-distribution data. Therefore, OOD instances are detected with respect to the magnitude of the component of the perpendicular subspace to the principal subspace. + +Energy-based OOD Detection: to address these limitations, Liu et al. [29] propose to use a free energy-based in-distribution score function instead of the softmax() score to identify OOD instances. This approach regularises the model to produce lower energy values for in-distribution and higher energy values for auxiliary outliers, creating a significant energy gap that enhances the model ability to detect OOD samples. Energy-based OOD modeling has enabled novel outlier synthesis paradigms. For instance, Du et al. [7] propose virtual outlier synthesis (VOS) by sampling outliers from low-likelihood in-distribution regions. VOS has shown to be effective for classification and object-level OOD detection. Further outlier synthesis techniques based on VOS have been proposed, including using normalizing flows [21], sampling based on the in-distribution boundary + +instances [39] or extending VOS when class labels are unavailable [18]. In addition, VOS techniques arensuitable for object detection [7, 21]. Whilst VOS and subsequent works synthesize outliers in the feature space, Dream-OOD [8] uses Stable Diffusion [34] to generate outliers in the pixel-space, learning better discriminative free-score values for OOD instances and providing visual examples of outliers. Energy-based methods have shown great performance, becoming state-of-the-art in several benchmarks (e.g., Dream-OOD [8] for ImageNet-100 in-distribution), hence we focus in the properties of these methods. Notwithstanding that the core idea of the energy score is that OOD samples must have a significantly different free energy than in-distribution sample, we examine some inherent vulnerabilities in the energy formulation based on the null space and least singular value of the last linear layer of neural-based classifiers. + +# 3. Preliminaries: Free Energy Score + +The goal of OOD detection is to detect outliers, $\mathbf{v}$ , that have a low probability in the true (unknown) in-distribution density $p(\mathbf{x})$ . However, the computation of $p(\mathbf{x})$ is usually unfeasible or computationally untractable, so we instead resort to additional information. Now consider that we also have a set of in-distribution categories $\{1,\dots ,K\}$ such that if we know the joint probability distribution, $p(\mathbf{x},k)$ , for $k\in \{1,\ldots ,K\}$ , we could marginalize $p(\mathbf{x})$ . Nonetheless, since the actual joint probabilities $p(\mathbf{x},k)$ might be also untractable, we could instead use non-negative arbitrary potential functions $\phi (\mathbf{x},k)$ that measure the affinity of the states of $\mathbf{x}$ and $k$ [12]. In this sense, we can construct the unnormalized probability: + +$$ +\tilde {p} (\mathbf {x}) = \sum_ {k = 1} ^ {K} \phi (\mathbf {x}, k) \tag {1} +$$ + +that gives an idea of the likelihood $\mathbf{x}$ . Furthermore, energy-based models enforce the non-negative constraint of $\phi$ by: + +$$ +\phi (\mathbf {x}, k) = \exp (- E (\mathbf {x}, k)), \tag {2} +$$ + +where $E$ is an unbounded arbitrary function called the energy. By substituting Eq. (2) into Eq. (1) and taking the negative $\log^1$ , we get a value known as the free energy: + +$$ +\mathcal {F} (\mathbf {x}) = - \log \tilde {p} (\mathbf {x}) = - \log \sum_ {k = 1} ^ {K} \exp (- E (\mathbf {x}, k)), \tag {3} +$$ + +and more recently used as a concept in OOD detection [7, 8, 18, 21, 29]. To get the actual density $p(\mathbf{x})$ , we must compute the normalizing partition function of Eq. (1); since we are interested in distinguishing between in-distribution and OOD, $\mathcal{F}(\mathbf{x})$ thus serves well for this purpose. + +![](images/bdd060b5b3baf52a609a84f74c4399fcafd9d4e86650030073fdec620e75a556.jpg) +Figure 2. Vulnerabilities of Free Energy-based OOD detection. Directions in the Null Space do not change the energy, while the LSV direction has minimal change, compared to other directions. + +In a general machine learning context, a classifier $f(\mathbf{x};\theta):\mathbb{R}^d\to \mathbb{R}^K$ with learned parameters $\theta$ assigns categorical unnormalized probabilities $f_{k}(\mathbf{x};\theta)$ , called logits, for a given $\mathbf{x}\in \mathbb{R}^d$ . While the logits are usually normalized using the softmax() function in order to obtain a probability distribution, they can also be used as the energy $E$ in Eq. (2). This realization enables the use of the free energy (Eq. (3)) as a measure of uncertainty for a classifier: + +$$ +\mathcal {F} (\mathbf {x}) = - \log \sum_ {k = 1} ^ {K} \exp \left(f _ {k} (\mathbf {x}; \theta)\right). \tag {4} +$$ + +Liu et al. [29] demonstrate that Eq. (4) is superior for effective OOD detection when compared to use of softmax() whilst Du et al. [7] explicitly use the free energy score (Eq. (4)) to perform binary classification between indistribution and OOD by training with synthetic outliers. + +# 4. Vulnerabilities of the Free Energy Score + +Although the free energy (Eq. (4)) is an effective uncertainty measure for OOD detection [7, 29], here we examine some inherent vulnerabilities when used in conjunction with a neural network based classifier. Firstly, Sec. 4.1 investigates OOD instances that have similar energies as in-distribution samples, whilst Sec. 4.2 discusses minimal changes in free energy, and finally Sec. 4.3 outlines our proposed approach to minimize the impact of these vulnerabilities. + +# 4.1. Null Space Vulnerability + +Consider a neural network-based classifier $f = (g \circ h)(\mathbf{x})$ where $h(\mathbf{x}): \mathbb{R}^d \to \mathbb{R}^{d'}$ encodes the input into a $d'$ -dimensional feature vector and $g: \mathbb{R}^{d'} \to \mathbb{R}^K$ is the last layer of the neural network that transforms the features into unnormalized class probabilities (i.e., the logits). If $g$ is a linear layer (as for most classifiers [5, 14, 43]), then: + +$$ +f (\mathbf {x}) = \mathrm {W} _ {c l s} ^ {\top} h (\mathbf {x}), \tag {5} +$$ + +where $\mathrm{W}_{cls} \in \mathbb{R}^{d' \times K}$ are learnable weights. Hence, the logits $f_{k}(\mathbf{x};\theta)$ correspond to the $k$ -th index of $f(\mathbf{x})$ . + +The core idea of energy-based OOD is that outliers $\mathbf{v}$ have a significantly different free energy, $\mathcal{F}(\mathbf{v})$ , from in-distribution samples, $\mathcal{F}(\mathbf{x})$ . Since OOD instances should be far from in-distribution samples in the feature space [7, 39], the expected distance from the inlier features $h(\mathbf{x})$ and outlier features $h(\mathbf{v})$ should be greater than a given boundary distance, $d_b$ , as follows: + +$$ +\mathbb {E} \left[ \| h (\mathbf {x}) - h (\mathbf {v}) \| \right] \geq d _ {b}, \tag {6} +$$ + +for a norm $\|\cdot\|$ . In other words, the inliers remain bounded to a region in the feature space while outliers are unbounded. Therefore, the expected distance between inliers and outliers is greater than a bounding distance $d_b$ . Subsequently, we can rewrite Eq. (6) as: + +$$ +h (\mathbf {v}) = h (\mathbf {x}) + \boldsymbol {\delta}, \tag {7} +$$ + +where $\delta$ is a vector with $\mathbb{E}[\|\delta\|] \geq d_b$ . By substituting Eq. (7) into Eq. (5) we obtain the following: + +$$ +f (\mathbf {v}) = \mathrm {W} _ {c l s} ^ {\top} (h (\mathbf {x}) + \boldsymbol {\delta}) = f (\mathbf {x}) + \mathrm {W} _ {c l s} ^ {\top} \boldsymbol {\delta}. \qquad (8) +$$ + +However within this formulation, if $\mathrm{W}_{cls}$ has more rows than columns, which occurs when the feature dimension is larger than the number of classes, the size of the null space of $\mathrm{W}_{cls}^{\top}$ , called the nullity, is nullity $(\mathrm{W}_{cls}^{\top}) = d' - \operatorname{rank}(\mathrm{W}_{cls})$ due to the rank-nullity theorem. Mathematically, this means that there exists a linear subspace $\mathrm{Null}(\mathrm{W}_{cls}^{\top}) \subset \mathbb{R}^{d'}$ such that if $\delta \in \mathrm{Null}(\mathrm{W}_{cls}^{\top})$ then $\mathrm{W}_{cls}^{\top} \delta = 0$ . Within our OOD context, if an outlier instance, $\mathbf{v}$ , has a feature representation, $h(\mathbf{v})$ , such that its difference with respect to the inlier feature representation, $h(\mathbf{x})$ , of any inlier instance, $\mathbf{x}$ , is in the null space of the last layer of the classifier, then their free energy scores are the same: + +$$ +h (\mathbf {v}) - h (\mathbf {x}) \in \operatorname {N u l l} \left(\mathrm {W} _ {c l s} ^ {\top}\right) \Rightarrow \mathcal {F} (\mathbf {x}) = \mathcal {F} (\mathbf {v}). \tag {9} +$$ + +Practically, Eq. (9) is telling us that it is possible for in-distribution and OOD instances to have different feature representations (satisfying the condition of Eq. (6)) and yet have the same free energy score, as depicted in Fig. 2. This fundamentally challenges the notion that free energy is an effective mechanism for OOD detection as any approaches reliant on Eq. (3), such as [7, 8, 18, 21, 29], contain inherent "detection blind-spots" as per Eq. (9). We denote Eq. (9) as the Null Space Vulnerabilities (NSV) of classifier $f$ with size nullity $(\mathrm{W}_{cls}^{\top}) = d' - \mathrm{rank}(\mathrm{W}_{cls})$ . An approach to minimize this vulnerability is outlined in Sec. 4.3. + +As an additional remark, NSV are inherent in energy-based methods; while some post hoc methods use the projection of features onto the null space as a discriminator between in-distribution and OOD [4, 41], energy-based methods use the classification logits for detecting outliers, hence the null space projection does not change the OOD score. + +# 4.2. Minimal Change of Free Energy + +We now turn our attention to the case where $\delta$ is not in the null space of $\mathbf{W}_{cls}^{\top}$ . Furthermore, we can focus on $\delta$ orthogonal to $\mathrm{Null}(\mathbf{W}_{cls}^{\top})$ , since every vector in the domain of $\mathbf{W}_{cls}^{\top}$ can be decomposed into one component in $\mathrm{Null}(\mathbf{W}_{cls}^{\top})$ and one component in $\mathrm{Null}(\mathbf{W}_{cls}^{\top})^{\perp}$ , and the effect of the component in $\mathrm{Null}(\mathbf{W}_{cls}^{\top})$ will be dealt with using the null space reduction method (Sec. 4.3). + +In order to create more discriminative OOD detectors based on the free energy score, we investigate cases where the difference in OOD and in-distribution free energy is minimal. We can state our problem formulation to find the case that minimizes the free energy difference: + +$$ +\begin{array}{l} \min _ {\substack {\boldsymbol {\delta}: \| \boldsymbol {\delta} \| \geq d _ {b}, \\ \boldsymbol {\delta} \in \operatorname {Null} \left(\mathrm {W} _ {c l s} ^ {\top}\right) ^ {\perp}}} | \mathcal {F} (\mathbf {v}) - \mathcal {F} (\mathbf {x}) | \\ = \min _ {\substack {\boldsymbol {\delta}: \| \boldsymbol {\delta} \| \geq d _ {b}, \\ \boldsymbol {\delta} \in \operatorname {Null} \left(\mathrm {W} _ {c l s} ^ {\top}\right) ^ {\perp}}} \left| \log \frac {\sum_ {k} ^ {K} \exp \left(f _ {k} (h (\mathbf {x}))\right)}{\sum_ {k} ^ {K} \exp \left(f _ {k} (h (\mathbf {x}) + \boldsymbol {\delta})\right)} \right|. \end{array} \tag{10} +$$ + +Whilst in Eq. (10) we set $\| \delta \| \geq d_b$ instead of $\mathbb{E}\left[\| \delta \| \right]\geq$ $d_{b}$ , in reality we are just interested in the case of non-zero $\delta$ since otherwise the solution would be trivial. Whilst the solution to Eq. (10) may be challenging to obtain, we note the difference in free-energies tends to 0 if the value inside the $\log ()$ function tends to 1. This is equivalent to making $f_{k}(h(\mathbf{x}))$ and $f_{k}(h(\mathbf{x}) + \delta)$ similar for all $k$ . Therefore, considering that $f_{k}$ is in the form of Eq. (5), our problem consists on finding the following proxy minimum: + +$$ +\begin{array}{r l} \underset {\delta : \| \boldsymbol {\delta} \| \geq d _ {b},} {\min } & \| \mathrm {W} _ {c l s} ^ {\top} (h (\mathbf {x}) + \boldsymbol {\delta}) - \mathrm {W} _ {c l s} ^ {\top} h (\mathbf {x}) \| \\ & = \underset {\substack {\boldsymbol {\delta}: \| \boldsymbol {\delta} \| \geq d _ {b}, \\ \boldsymbol {\delta} \in \operatorname {Null} \left(\mathrm {W} _ {c l s} ^ {\top}\right) ^ {\perp}}} {\min } \| \mathrm {W} _ {c l s} ^ {\top} \boldsymbol {\delta} \|. \end{array} \tag{11} +$$ + +For the $L^2$ -norm in Eq. (11), we thus obtain: + +$$ +\min _ {\substack {\boldsymbol {\delta}: \| \boldsymbol {\delta} \| _ {2} \geq d _ {b}, \\ \boldsymbol {\delta} \in \operatorname {Null} \left(\mathrm {W} _ {c l s} ^ {\top}\right) ^ {\perp}}} \| \mathrm {W} _ {c l s} ^ {\top} \boldsymbol {\delta} \| _ {2} = d _ {b} \sigma_ {m i n} \left(\mathrm {W} _ {c l s}\right), \tag{12} +$$ + +where $\sigma_{\min}(\mathrm{W}_{cls})$ is the least singular value of $\mathrm{W}_{cls}$ (see further mathematical formulation details in Supp. Mat.). + +Subsequently, Eq. (12) holds for the case where $\delta$ is in the direction corresponding to the least singular value, such that the change in the free energy will be smaller than any other direction for $\delta$ (and considering $\delta \in \mathrm{Null}(\mathrm{W}_{cls}^{\top})^{\perp}$ ). We name this result (from Eq. (12)) the Least Singular Value Vulnerabilities (LSVV) of $f$ . In our OOD context, it means that there could exist some outliers far from the inlier distribution, in terms of Eq. (6), with similar free energy scores because they may fall in the direction of the least singular value of $\mathrm{W}_{cls}$ . Again, this fundamentally challenges + +the notion that free energy is an effective mechanism for OOD as free energy has been shown to be non-linear with respect to a given feature representation, $h(\cdot)$ . + +This analysis is itself inspired by the seminal adversarial attacks analysis of Simon-Gabriel et al. [37], where they instead maximize the change in the classifier response for small input perturbations. Whilst here we are modelling the opposite problem, (i.e., the minimum change in the classifier for perturbations larger than a non-zero value), we could consider this analysis as an effective adversarial attack mechanism for the OOD domain. + +# 4.3. Null Space Reduction & Regularisation + +In this section, we introduce our Free Energy Vulnerability Elimination for Robust OOD Detection (FEVER-OOD) framework, where we explore readily available methods to minimize the NSV (Sec. 4.1) and LSVV (Sec. 4.2) impact potential within free energy score based OOD approaches. + +Null Space Reduction (NSR). For NSV, we propose to reduce the null space by adding an extra linear layer $g' : \mathbb{R}^{d'} \to \mathbb{R}^r$ , where we choose $r < d'$ to reduce the size of the NSV (Fig. 3). We denote this method as $r$ -null space reduction ( $r$ -NSR). Therefore, the new classifier becomes $f' = g \circ g' \circ h$ such that the feature representation of a sample is in $\mathbb{R}^r$ and is given instead by $(g' \circ h)(\mathbf{x})$ . Effectively, within an OOD context, we are promoting better discrimination between OOD and in-distribution instances. + +Least Singular Value Regulariser (LSVR). To minimize the impact of LSVV, we maximise $\sigma_{min}(\mathrm{W}_{cls})$ by adding an LSVR to the loss function $\mathcal{L}$ of a baseline method: + +$$ +\mathcal {L} _ {L S V} := \mathcal {L} + \lambda_ {L S V} \sigma_ {\min } ^ {- 1} \left(\mathrm {W} _ {c l s}\right), \tag {13} +$$ + +where $\lambda_{LSV}$ is a scalar hyperparameter that controls the contribution of the LSV regulariser. It is additionally noted that alternatively maximizing the LSV might also increase the rest of the singular values, resulting in more significant changes of energy in all directions. In the context of OOD detection, LSVR increases the discriminative capacity for OOD detection of baseline methods. + +Condition Number (CN) Regulariser. Whilst the LSV regulariser will increase the change in energy for $\delta$ in the LSV direction, there is no constraint with regard to the distribution over all other directions. For this reason, we explore a Condition Number Regulariser (CNR): + +$$ +\mathcal {L} _ {C N} := \mathcal {L} + \lambda_ {C N} \kappa \left(\mathrm {W} _ {\text {c l s}}\right), \tag {14} +$$ + +where $\kappa (\mathrm{W}_{cls})$ is the condition number of the WCLS matrix, i.e., the ratio of the greatest and least singular values. Intuitively, this will encourage equalized energy changes for displacements in all directions where we note that Eq. (14) similarly assumes use of the $L^2$ -norm as per Eq. (12). A different approach to the CNR is the orthogonal constraint + +![](images/9ce8f5922b4e4211f3267ae349a5be11a44052d8378b0f5c650be308d1177588.jpg) +Figure 3. Null space reduction. We add an extra layer $g'$ to create $r$ -dimensional features, reducing the size of the null space. + +$\| W_{cls}W_{cls}^{\top} - I\|$ . However, this method adds the additional constraint that the singular values must be close to one, which might be a stronger prior than the CNR, hence we opt for the latter. The gradients for the singular values in Eqs. (13) and (14) are detailed by Lewis and Sendov [26] (Proposition 6.2) and further discussed in the Supp. Mat. + +# 5. Experimental Details + +We study the impact of FEVER-OOD on free energy baseline methods (Sec. 5.1) across different standard datasets (Sec. 5.2). Implementation details are given in Sec. 5.3. + +# 5.1. Baseline Methods and Evaluation Metrics + +Three free energy baseline methods for image OOD detection that use the free energy score are tested: VOS [7], FFS [21] and Dream-OOD [8]. These baseline methods learn a small MLP $\psi : \mathbb{R} \to (0,1)$ that uses the free energy in Eq. (4) to predict an uncertainty score. Since OOD data is unavailable during training, these methods use synthetic outliers to train $\psi$ . While VOS and FFS synthesize outliers in the feature space $h(\mathbf{x})$ , Dream-OOD generates outliers in the pixel space using Stable Diffusion [34]. Additionally, VOS and FFS are also implemented as object detectors for object-level OOD detection, synthesizing outliers before the last layer of the classification head of a Faster RCNN [33]. + +We follow standard evaluation metrics, reporting the false positive rate for OOD detection at $95\%$ true positive rate of in-distribution detection (FPR95) and the area under the receiver operating characteristic curve (AUROC). We also assess the effect of our proposed techniques into the main in-distribution task (accuracy for classification and MS-COCO [28] mean average precision (mAP) for object detection). FPR95 and AUROC metrics are reported using a single intersection-over-union threshold for object detection, corresponding to the one that maximises the F-1 score, as described by Harakeh and Waslander [13]. + +Table 1. CIFAR-10 Results (ID Acc. = in-distribution accuracy; Null Space Reduction (NSR) methods = our approach). + +
MethodFEVER-OODOOD DatasetsID Acc
TexturesSHVNPlaces365LSUNiSUNAvg
r-NSRRegularizerFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROC
VOS--50.05±5.7086.79±1.2839.15±10.5291.73±2.8840.88±1.7589.54±0.698.22±1.0298.36±0.2030.01±4.4494.31±0.8433.66±2.2192.15±0.5994.83±0.16
-λLSV=1.055.58±6.3586.71±2.3543.00±3.1692.09±0.9438.78±2.5191.14±0.567.36±1.5998.46±0.3221.23±6.9596.23±1.1733.19±3.2092.92±0.8594.66±0.18
-λCN=0.150.80±3.0685.39±0.7930.78±2.9594.04±0.5841.60±1.4088.77±0.748.50±0.8298.30±0.1030.91±4.5493.41±1.6832.52±1.2591.98±0.4694.69±0.16
96λLSV=1.040.86±4.8392.12±1.0541.92±16.1193.20±2.6135.69±2.2992.61±0.574.95±1.4998.93±0.2417.68±5.2697.04±0.8528.22±4.1894.78±0.8594.74±0.17
64λLSV=1.044.41±5.3991.26±1.2934.68±10.9394.39±1.7235.85±2.5692.42±0.535.68±0.5798.75±0.1324.03±5.9196.02±0.9528.93±3.8694.57±0.6694.75±0.13
32λLSV=0.00146.53±3.7789.67±1.0927.84±9.4795.29±1.0837.33±1.6391.77±0.395.65±0.7798.82±0.1124.09±6.7795.68±1.3528.29±3.3394.25±0.5594.68±0.07
10-53.20±3.9088.97±0.7335.62±9.6694.41±1.2845.73±6.2989.73±1.6911.72±3.3997.93±0.4641.80±14.3592.51±3.0337.61±4.0092.71±0.9791.95±2.32
FFS--52.86±4.4983.47±1.6738.67±11.2189.74±5.1944.65±1.2987.47±0.776.59±0.9298.67±0.1731.34±1.8893.24±0.9734.82±2.0390.52±0.9894.69±0.15
-λLSV=0.00150.74±3.9884.93±0.7532.25±12.1692.97±3.6042.99±2.3088.30±1.105.76±1.2398.82±0.2728.16±6.4594.18±1.9831.98±3.1091.84±1.2894.73±0.12
-λCN=1.050.08±3.8684.92±1.8530.91±6.1992.77±2.3345.69±2.5987.16±0.756.22±0.6598.73±0.1126.60±3.6494.80±0.6431.90±2.5191.67±0.9494.85±0.20
96λLSV=1.048.11±5.1589.84±1.5140.65±15.4793.32±3.1636.37±2.2892.29±0.605.22±1.3598.89±0.1824.84±10.0895.82±1.6831.04±5.0594.03±1.0694.71±0.08
64λLSV=1.051.70±5.6788.81±1.0933.02±14.0994.60±2.3039.05±3.7391.53±1.235.56±1.9598.84±0.3026.63±7.5195.65±1.1231.19±4.3293.89±0.6994.68±0.08
32-46.75±3.4989.61±1.4031.77±8.7394.44±1.3541.51±3.3890.72±1.024.93±1.1199.04±0.1628.92±11.2394.49±2.9430.78±3.7093.66±0.9594.79±0.18
10-54.72±5.8984.77±3.2345.70±12.9591.49±2.2741.52±3.2789.51±1.318.21±3.5998.33±0.6728.58±7.7594.30±1.8235.75±5.2691.68±1.3894.54±0.17
+ +# 5.2. Datasets + +We assess FEVER-OOD with standard datasets for both classification and object detection. For image-level OOD detection (i.e., image classification), we train VOS and FFS using the CIFAR-10 and CIFAR-100 [20] datasets as in-distribution while testing OOD detection in five non-overlapping datasets: Textures [3], SVHN [30], Places365 [47], LSUN [45] and iSUN [11]. We also test Dream-OOD with the CIFAR-100 dataset as in-distribution data with the same OOD datasets. Additionally, we also train Dream-OOD with the Imagenet-100 dataset, a 100 classes partition of Imagenet [35], as in-distribution while testing OOD detection in four datasets: Textures [3], Places365 [47], iNaturalist [40] and SUN [42]. We use the PASCAL VOC [10] dataset as in-distribution for object-level OOD detection using VOS and FFS and test for OOD detection on the MS-COCO dataset [28] removing images containing objects with overlapping in-distribution categories [7]. + +# 5.3. Implementation details + +We follow the established architectural choices of prior works. For image-level OOD detection with VOS and FFS, we use a WideResNet-40 [46] with a 128-dimensional feature space. We choose different NSR values ranging from the feature dim to the number of classes, gradually decreasing the nullity. For CIFAR-10, we evaluate the $\{96,64,32,10\}$ -NSR, while for CIFAR-100 we assess $\{114,100\}$ -NSR. With respect to Dream-OOD models, we use a ResNet-34 [14] with 512-dimensional feature space, evaluating $\{256,128,100\}$ -NSR for both CIFAR-100 and Imagenet-100. Similarly, we reduce the feature space of the classification head to 768, 512 and 265 dimensions for VOS and FFS object detection experiments, having an original 1,024-dimensional object-wise feature space (as per Faster RCNN [33]). With regards to the LSVR and CNR, we test with various loss weight values in all the architectures (1, 0.1, 0.01, and 0.001), to evaluate their impact on model performance (except for Imagenet-100, where we use $\lambda_{LSV} = 0.001$ and $\lambda_{CN} = 0.01$ based on CIFAR-100 experiments). We train all the models using the original + +corresponding baseline settings. Details on the hyperparameters and the training regime for each model are given in the Supp. Mat. To obtain reliable comparisons we use deterministic learning via a fixed random seed for all experiments. We run all experiments 5 times and report mean and standard deviation for all metrics. + +# 6. Results + +We discuss the performance of FEVER-OOD against different free energy based OOD detection models in classification and detection. Tabs. 1 to 3 compare the use of FEVER-OOD with the corresponding baseline methods. We report different NSR as well as the best results obtained using $\lambda_{LSV}$ (Eq. (13)) or $\lambda_{CN}$ (Eq. (14)), and their corresponding value (we show both cases when no NSR is applied). Extensive ablations are available in the Supp. Mat. + +CIFAR-10. Tab. 1 shows the results of the VOS and FFS architectures with CIFAR-10 as in-distribution. In general, NSR and LSVR improve OOD detection for both VOS and FFS models. The best performance using the VOS method is achieved using a 96-NSR with LSVR, using a $\lambda_{LSV} = 1.0$ , where the average AUROC increases to 94.78 and the average FPR95 reduces to $28.22\%$ , compared to an AUROC of 92.15 and $33.66\%$ of the baseline VOS model without affecting the in-distribution accuracy. Similarly, the best FFS setting is achieved with a 96-NSR and $\lambda_{LSV} = 1.0$ , resulting in 94.03 AUROC and $31.04\%$ FPR95, compared to the baseline FFS model, with an AUROC of 90.52 and FPR95 of $34.82\%$ , although the best FPR95 is achieved with 32-NSR and no singular value regularization, with an FPR95 of 30.78. In general, it is seen that a large NSR (i.e., 10-NSR) may decrease the AUROC and increase the FPR95. Since VOS and FFS are techniques that synthesize outliers in the feature space, a large reduction of the feature dimension might not allow for proper in-distribution understanding, impacting in the OOD training of these techniques. Additionally, it is also observed that LSVR yields better results than CNR, and a value for $\lambda_{LSV} = 1.0$ is adequate for none to moderate NSR, while smaller values for LSVR (or none) are better for large NSR. While LSVR increases the differ + +Table 2. CIFAR-100 Results (ID Acc. = in-distribution accuracy; Null Space Reduction (NSR) methods = our approach). + +
MethodFEVER-OODOOD DatasetsID Acc
r-NSRRegularizerTexturesSHVNPlaces365LSUNiSUNAvg
FPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROC
VOS--81.24±2.1176.01±0.8076.09±6.5183.18±3.5980.53±1.4676.25±0.7837.86±4.0693.21±0.6879.34±1.6076.43±1.4471.01±1.2381.02±0.4076.04±0.22
-λLSV=0.182.87±1.1176.49±0.9473.35±2.9684.57±1.5380.36±1.1676.89±0.7940.05±3.7492.83±0.4674.28±4.0380.25±3.3470.18±0.8782.21±0.5676.01±0.16
-λCN=0.0182.51±2.6476.04±1.9173.67±3.1085.31±1.2480.18±1.0876.59±0.2341.08±3.2792.75±0.6380.83±4.3773.79±5.5171.65±0.9680.90±1.1176.20±0.11
114λLSV=0.0180.02±2.7578.73±1.8383.45±2.5982.07±1.3678.41±0.9677.71±0.9328.90±3.6195.11±0.5872.07±9.0780.55±4.7068.57±2.2882.83±0.9175.72±0.16
100-80.47±2.3278.21±1.1479.10±7.4283.39±3.9179.19±1.0477.23±0.5028.99±2.5995.15±0.3975.51±6.6879.63±3.7068.65±2.2982.72±0.5775.54±0.19
FFS--82.87±1.3975.54±0.9376.32±6.1784.83±2.3681.14±1.0976.27±0.7436.27±4.9193.54±1.0382.83±2.9874.29±3.0871.89±1.6480.89±1.1176.04±0.11
-λLSV=0.0180.95±1.3476.53±0.7083.12±5.5981.61±2.9280.19±0.8476.58±0.4932.93±4.0594.20±0.5179.05±5.1677.28±3.2371.25±2.1181.24±0.7276.27±0.28
-λCN=0.00180.27±2.7676.78±1.5378.57±7.3282.57±2.9580.40±1.0976.40±0.5332.91±4.9094.06±1.0280.71±7.9074.49±5.5770.57±2.4180.86±0.8876.05±0.16
114λLSV=0.00179.26±2.6778.16±1.5074.28±7.1385.37±3.2079.45±0.8577.25±0.6124.10±1.4295.94±0.1773.61±5.8280.61±2.4966.14±1.9183.47±0.9875.45±0.37
100-77.69±2.9778.62±1.0277.84±9.6283.40±3.5280.41±0.7276.58±0.3521.91±0.9596.25±0.1179.60±6.0675.79±5.1767.49±2.0382.13±0.8575.48±0.28
Dream-OOD--62.20±1.0283.84±0.3873.05±1.9284.56±0.2177.95±1.9779.43±0.1739.90±2.0192.87±0.441.70±0.1199.58±0.0450.96±1.4488.06±0.6075.61±0.19
-λLSV=0.0158.45±2.2786.04±0.4168.75±1.8487.65±0.3077.45±2.1178.59±0.2215.45±2.5797.24±0.091.55±0.0399.63±0.4544.33±3.4889.83±0.1275.87±0.23
-λCN=0.00157.40±1.8886.28±0.3277.75±1.1185.13±0.4178.60±1.2378.73±0.1527.2±1.7795.01±0.421.55±0.0899.57±0.0748.5±1.5488.94±0.4376.32±0.14
256-60.00±3.2185.42±0.7967.50±2.4785.84±0.6675.90±1.1079.57±0.4319.85±1.0596.74±0.671.00±0.0299.78±0.0644.82±2.7289.47±0.24277.01±0.12
128λLSV=0.156.45±2.3987.73±1.0167.45±1.1287.53±0.4478.45±2.4677.24±0.6129.60±1.7394.81±0.401.95±0.1099.5±0.1046.78±2.7289.36±0.8376.21±0.35
100-57.55±2.4886.8±0.4454.45±3.0288.69±0.5975.80±1.2178.88±0.6924.45±1.5795.86±0.521.60±0.0799.65±0.2142.77±2.1089.98±0.1576.41±0.32
+ +Table 3. ImageNet-100 Results (ID Acc. = in-distribution accuracy; Null Space Reduction (NSR) methods = our approach). + +
MethodFEVER-OODOOD DatasetsID Acc
r-NSRRegularizeriNaturalistPlaces365SUNTexturesAvg
FPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROC
Dream-OOD--23.98±2.0895.94±0.2641.75±1.7292.48±0.1240.85±1.1392.76±0.0950.73±0.8586.21±0.2939.33±1.0891.84±0.1287.76±0.19
-λLSV=0.0159.91±3.0990.89±0.4866.26±2.1788.12±0.4774.26±1.5284.53±0.3360.56±1.2183.91±0.4265.25±1.5086.87±0.2587.85±0.11
-λLSV=0.00159.71±3.5390.97±0.7065.01±1.9088.39±0.4573.93±1.5884.69±0.4960.51±1.4684.01±0.5164.79±1.6987.01±0.4587.77±0.13
256-23.84±1.7395.80±0.3044.35±1.7192.14±0.3042.92±1.9892.46±0.3344.38±1.8188.60±0.5438.87±1.2992.25±0.2387.57±0.24
128λLSV=0.0123.46±2.5495.93±0.3643.37±1.4792.19±0.2242.42±1.6092.44±0.2941.86±1.0989.48±0.2737.78±1.4592.51±0.2287.60±0.11
100λLSV=0.0122.24±0.8196.16±0.1441.08±3.0392.59±0.4440.39±2.8492.91±0.4042.31±0.9789.30±0.3736.50±1.6792.74±0.2187.42±0.15
+ +![](images/b5a0447239bfd21dc15f5b456962b4ed61103b23b89df1f7fdc9f6fc4365220a.jpg) +(a) Dataset: CIFAR-10 +Figure 4. Singular Values Spectrum. + +![](images/c3e37d80961ce74aefb7bf09e7d5aaeea22dc52d68dccd6d83b173b223b81729.jpg) +(b) Dataset: CIFAR-100 + +ence in energy of outliers in or near the least singular vector direction, the CN regulariser makes equalized free energy changes for all directions. Since the latter effect might disable directions with smaller energy changes than others, it might also reduce the free energy change in the direction of the largest singular value. Fig. 4(a) shows the singular value spectrum of the baseline VOS model against the LSVR and CNR versions for the CIFAR-10 dataset. It is seen that the while both regularizers homogenize the singular values, the CNR also reduces the most significant singular value. Finally, our experiments show that VOS achieves better results than FFS in CIFAR-10 OOD detection. + +CIFAR-100. Tab. 2 shows the results for VOS, FFS and Dream-OOD trained on the CIFAR-100 dataset as in-distribution. The best models for VOS and FFS are achieved with 114-NSR and LSVR, with a $68.57\%$ FPR95 $(-2.44\%)$ VOS) and 82.83 $(+1.81\mathrm{VOS})$ AUROC for VOS-114-NSR and $66.14\%$ FPR95 $(-5.75\%)$ FFS) and 83.47 $(+2.58\mathrm{FFS})$ AUROC for FFS-114-NSR. However, despite these improvements, the performance of virtual outlier synthesis methods is usually poor for several in-distribution classes [44], hence the enhancement of these architectures is still challenging. Regarding Dream-OOD, it is noted that the best results are obtained by Dream-OOD-100-NSR, having an FPR95 of $42.77\%$ $(-8.19\%)$ Dream-OOD) and 89.98 + +AUROC (+1.9 Dream-OOD). Finally, for the three energy techniques (VOS, FFS and Dream-OOD), we observed that the model does not converge when using 100-NSR and the CNR (see Supp. Mat.), suggesting that the conditioning number penalization does not allow the model to learn the in-distribution task. Fig. 4(b) shows the singular value spectrum for the LSVR and CNR versions of the VOS model. Differently from the CIFAR-10 experiments, it is seen that LSVR keeps the most significant singular values similar to the baseline but homogenizes the least significant singular values, while the CNR keeps all singular values at a low level, hence explaining why it fails for this dataset. + +Imagenet-100. We show the results of Dream-OOD with Imagenet-100 as in-distribution in Tab. 3. Similar as with CIFAR-100, the best results are obtained with NSR100, effectively eliminating the null space of the last layer (since it becomes square and is not singular). Contrary to CIFAR-100, LSVR has a negative impact when used alone. Considering that we use the same pre-training scheme as Du et al. [8], adding LSVR and CNR on an already pretrained last linear layer might affect the already formed feature space. Hence, these regularisers benefit untrained NSR modules. The best model, Dream-OOD-100-NSR with LSVR, achieves an average $36.50\%$ FPR95 and 92.74 AUROC, compared to the baseline $39.33\%$ FPR95 and 91.84 AUROC, becoming the state-of-the-art for Imagenet-100. + +PASCAL VOC. We test FEVER-OOD techniques in object detection task, where the goal is to detect object-level OOD instances. Tab. 4 show the result for the baseline VOS and FFS, taking PASCAL VOC [10] as in-distribution dataset. For both tasks, we evaluate on two OOD datasets that contain subset of images from: MS-COCO [28] and OpenImages [22], following the original works [7, 21]. The best-performing model for VOS is VOS-768-NSR with CNR, + +![](images/38dbf40693c2a1f54c600dfde7a1f99607541a80bb3b382b4e01f49999ed70f7.jpg) + +![](images/faca7797e080a595d5658a989be5df461a6290e0819605a9804d8973355d6141.jpg) + +![](images/ced249d0cb05ba7cd7dae5c191dd056ace4e257e8a829eee15d6f0bf62667384.jpg) + +![](images/dc34406b1cea827db91f32e8f55f3b196923b6a6b91f5fa79bf82d32af4fa7b6.jpg) + +![](images/0131f23c020b950610ac0e882aeeaf1edd8e56e0d088678eb9446f5fe4cd56f6.jpg) +Figure 5. MS-COCO objects detected on OOD images by VOS baseline [7] (first row) and Fever-OOD (second row). Blue: OOD objects detected and mis-classified as being in-distribution. Green: the same OOD objects correctly detected as OOD by FEVER-OOD (ours). + +![](images/e2d1f8f1cedb55b6fb5d8445d54255cf5535fafaca5776d32bedee50b1fd4b21.jpg) + +![](images/fde8a35f769bbd6f5f32b25bc9dda203d2d11d2aa027c8ec6d8b373e4c0c396a.jpg) + +![](images/12c8724f5737873132ceb2e02d3f4dab10f687b8d2e26550db93809d26955762.jpg) + +Table 4. VOS results with PASCAL VOC as in-distribution (Null Space Reduction (NSR) methods = our approach). + +
MethodFEVER-OODOOD DatasetsmAP
r-NSRλLSVλCNMS-COCOOpenImages
FPR95AUROCFPR95AUROC
VOS---50.2987.7753.0986.580.489
-0.001-49.9688.1052.2685.800.489
--153.4286.6652.7385.770.490
768--48.9487.9452.3486.330.489
7680.01-50.0887.7851.7186.040.493
768-0.0147.4788.1549.3686.320.491
512--49.8488.4255.3385.330.486
5120.001-49.2688.3954.3285.820.489
512-0.0149.5987.4854.8185.480.490
256--47.8888.4952.4186.390.486
2560.01-48.5188.3650.5286.120.491
256-0.00149.188.2751.4786.080.489
FFS---50.7787.1853.7885.340.487
-0.01-53.0986.7553.4785.860.490
--0.0150.3687.6551.1985.980.485
768--51.5087.5953.3385.840.488
7680.01-47.4288.3652.4686.400.489
768-0.0151.1286.8255.4484.840.490
512--46.9388.8253.8986.280.491
5120.01-47.4288.3652.4686.400.488
512-0.0151.4287.0554.9884.550.490
256--51.1888.7453.6786.560.487
2560.01-47.9388.0447.9384.950.489
256-0.0152.9486.6252.8884.450.489
+ +achieving average FPR95 scores of $47.47\%$ and $49.36\%$ on MS-COCO and OpenImages, respectively, an improvement of nearly $4\%$ over the VOS baseline. For FFS, we achieve a minimum FPR95 of $46.93\%$ with FFS-512-NSR model and $47.93\%$ for the FFS-256-NSR model with the LSVR for both OOD datasets. Additionally, our approach generally delivers higher AUROC scores compared to the FFS baseline across both OOD datasets. Qualitative results are illustrated in Fig. 5, where we visualize the prediction on MS-COCO OOD images, using the baseline VOS (top) and the best Fever-OOD model, i.e., VOS-768-NSR, with the PASCAL-VOC as the in-distribution dataset, showing that out framework outperforms baseline VOS in in detecting outlier objects with a significantly lower false positive rate. + +Computational overhead. The computational overhead introduced by our method is minimal since we only add an extra linear layer. For instance, the ResNet-34 used in Dream-OOD takes $4.42 \pm 1.15$ ms/img and has 21.5M + +params. Adding 256, 128, 100-NSR layers change the inference time to $3.91 \pm 0.62$ , $4.44 \pm 1.17$ , $4.13 \pm 1.02$ ms/img, with an extra 105k, 27k, 10k params, which is negligible compared with the total size of ResNet-34. + +# 7. Conclusion + +In this work, we present FEVER-OOD, a robust approach for addressing critical vulnerabilities found in existing energy-based OOD detection. Our method targets the Null Space Vulnerabilities (NSV), where in-distribution and OOD instances can produce similar free energy scores despite differing feature representations, and Least Singular Value Vulnerabilities (LSVV), where energy score similarities are influenced by the least singular vector of the classifier last linear layer. These vulnerabilities are especially problematic in high-dimensional feature spaces where the feature dimension exceeds the number of classes, impacting the effectiveness of OOD detection. To mitigate these issues, FEVER-OOD incorporates strategies such as reducing the null space through lower-dimensional feature spaces and introducing a Least Singular Value Regularizer to maximize the least singular value, enhancing the distinction between energy scores of different samples. Additionally, we propose a Condition Number Regularizer to promote uniform energy changes for feature-space displacements, fostering consistent energy variation and improving detection robustness. Our comprehensive evaluation on benchmark datasets, including applications to the Dream-OOD framework with ImageNet-100 as the ID dataset, demonstrates the significant effectiveness of FEVER-OOD. Specifically, our approach achieves a $10.13\%$ reduction in the average false positive rate and a $1.6\%$ increase in AUROC, setting a new state-of-the-art in OOD detection. Furthermore, our results highlight the generalization capability of FEVER-OOD across different detection architectures, showcasing its potential beyond conventional classification settings. Future extensions of FEVER-OOD could include applications to varied domains and integration with other OOD detection frameworks to further validate its versatility and performance in diverse machine learning contexts. This work provides valuable insights into enhancing OOD robustness and sets a foundation for more resilient free energy-based detection methods. + +# References + +[1] Haoyue Bai, Gregory Canal, Xuefeng Du, Jeongyeol Kwon, Robert D Nowak, and Yixuan Li. Feed two birds with one scone: Exploiting wild data for both out-of-distribution generalization and detection. In International Conference on Machine Learning, pages 1454-1471. PMLR, 2023. 2 +[2] Haoyue Bai, Jifan Zhang, and Robert Nowak. Aha: Human-assisted out-of-distribution generalization and detection. arXiv preprint arXiv:2410.08000, 2024. 2 +[3] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3606-3613, 2014. 6 +[4] Matthew Cook, Alina Zare, and Paul Gader. Outlier detection through null space analysis of neural networks. ICML Workshops, 2020. 2, 4 +[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 3 +[6] Xuefeng Du, Xin Wang, Gabriel Gozum, and Yixuan Li. Unknown-aware object detection: Learning what you don't know from videos in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13678-13688, 2022. 1 +[7] Xuefeng Du, Zhaoning Wang, Mu Cai, and Yixuan Li. Vos: Learning what you don't know by virtual outlier synthesis. In International Conference on Learning Representations, 2022. 1, 2, 3, 4, 5, 6, 7, 8, 9, 10 +[8] Xuefeng Du, Yiyou Sun, Xiaojin Zhu, and Yixuan Li. Dream the impossible: Outlier imagination with diffusion models. In Advances in Neural Information Processing Systems, 2023. 2, 3, 4, 5, 7 +[9] Xuefeng Du, Zhen Fang, Ilias Diakonikolas, and Yixuan Li. How does unlabeled data provably help out-of-distribution detection? arXiv preprint arXiv:2402.03502, 2024. 1 +[10] Mark Everingham, Luc Van Gool, Christopher K. I. Williams, John M. Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International Journal of Computer Vision, 88(2):303-338, 2010. 6, 7, 2, 9, 10 +[11] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 580-587, 2014. 6 +[12] Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep Learning. MIT Press, 2016. http://www.deeplearningbook.org.3 +[13] Ali Harakeh and Steven L. Waslander. Estimating and evaluating regression predictive uncertainty in deep object detectors. In International Conference on Learning Representations, 2021. 5 + +[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 3, 6, 2 +[15] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In International Conference on Learning Representations, 2022. 2, 8 +[16] Yen-Chang Hsu, Yilin Shen, Hongxia Jin, and Zsolt Kira. Generalized odin: Detecting out-of-distribution image without learning from out-of-distribution data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10951-10960, 2020. 1 +[17] Rui Huang, Andrew Geng, and Yixuan Li. On the importance of gradients for detecting distributional shifts in the wild. Advances in Neural Information Processing Systems, 34:677-689, 2021. 8 +[18] B.K.S. Isaac-Medina, Y.F.A. Gaus, N. Bhowmik, and T.P. Breckon. Towards open-world object-based anomaly detection via self-supervised outlier synthesis. In Proc. European Conference on Computer Vision. Springer, 2024. 1, 3, 4 +[19] Julian Katz-Samuels, Julia B Nakhleh, Robert Nowak, and Yixuan Li. Training ood detectors in their natural habitats. In International Conference on Machine Learning, pages 10848-10865. PMLR, 2022. 2 +[20] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009. 6, 2 +[21] Nishant Kumar, Siniša Šegvic, Abouzar Eslami, and Stefan Gumhold. Normalizing flow based feature synthesis for outlier-aware object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5156-5165, 2023. 1, 2, 3, 4, 5, 7, 6, 9, 10 +[22] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision, 128(7):1956-1981, 2020. 7, 6, 9 +[23] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. Advances in Neural Information Processing Systems, 30, 2017. 1 +[24] Hao Lang, Yinhe Zheng, Yixuan Li, Jian SUN, Fei Huang, and Yongbin Li. A survey on out-of-distribution detection in NLP. Transactions on Machine Learning Research, 2024. 1 +[25] Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. Advances in Neural Information Processing Systems, 31, 2018. 1, 2 +[26] Adrian S Lewis and Hristo S Sendov. Nonsmooth analysis of singular values. part i: Theory. Set-Valued Analysis, 13: 213–241, 2005. 5, 1 +[27] Shiyu Liang, Yixuan Li, and R Srikant. Enhancing the reliability of out-of-distribution image detection in neural net + +works. In 6th International Conference on Learning Representations, 2018, 2018. 1, 2, 8 +[28] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5, 6, 7, 10 +[29] Weitang Liu, Xiaoyun Wang, John Owens, and Yixuan Li. Energy-based out-of-distribution detection. Advances in Neural Information Processing Systems, 33:21464-21475, 2020. 1, 2, 3, 4 +[30] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading digits in natural images with unsupervised feature learning. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning 2011, 2011. 6 +[31] Anh Nguyen, Jason Yosinski, and Jeff Clune. Deep neural networks are easily fooled: High confidence predictions for unrecognizable images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 427-436, 2015. 1 +[32] Bo Peng, Yadan Luo, Yonggang Zhang, Yixuan Li, and Zhen Fang. Conjnorm: Tractable density estimation for out-of-distribution detection. In The Twelfth International Conference on Learning Representations, 2024. 2 +[33] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: towards real-time object detection with region proposal networks. In Proceedings of the 28th International Conference on Neural Information Processing Systems - Volume 1, page 91-99, Cambridge, MA, USA, 2015. MIT Press. 5, 6 +[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 3, 5 +[35] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 115 (3):211-252, 2015. 2, 6 +[36] Mohammadreza Salehi, Hossein Mirzaei, Dan Hendrycks, Yixuan Li, Mohammad Hossein Rohban, and Mohammad Sabokrou. A unified survey on anomaly, novelty, open-set, and out of-distribution detection: Solutions and future challenges. Transactions on Machine Learning Research, 2022. 1 +[37] Carl-Johann Simon-Gabriel, Yann Ollivier, Leon Bottou, Bernhard Schölkopf, and David Lopez-Paz. First-order adversarial vulnerability of neural networks and input dimension. In Proceedings of the 36th International Conference on Machine Learning, pages 5809-5817. PMLR, 2019. 5 +[38] Yiyou Sun, Yifei Ming, Xiaojin Zhu, and Yixuan Li. Out-of-distribution detection with deep nearest neighbors. In International Conference on Machine Learning, pages 20827-20840. PMLR, 2022. 8 + +[39] Leitian Tao, Xuefeng Du, Jerry Zhu, and Yixuan Li. Nonparametric outlier synthesis. In The Eleventh International Conference on Learning Representations, 2023. 3, 4, 8 +[40] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8769-8778, 2018. 6 +[41] Haoqi Wang, Zhizhong Li, Litong Feng, and Wayne Zhang. Vim: Out-of-distribution with virtual-logit matching. In CVPR, pages 4921-4930, 2022. 2, 4, 8 +[42] Jianxiong Xiao, James Hays, Krista A. Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 3485-3492, 2010. 6 +[43] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10687-10698, 2020. 3 +[44] Jingkang Yang, Kaiyang Zhou, Yixuan Li, and Ziwei Liu. Generalized out-of-distribution detection: A survey. International Journal of Computer Vision, pages 1-28, 2024. 1, 7 +[45] Fisher Yu, Yinda Zhang, Shuran Song, Ari Seff, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 6 +[46] Erwan Zerhouni, David Lanyi, Matheus Viana, and Maria Gabrani. Wide residual networks for mitosis detection. In 2017 IEEE 14th International Symposium on Biomedical Imaging (ISBI 2017), pages 924-928. IEEE, 2017. 6 +[47] Bolei Zhou, Agata Lapedriza, Antonio Torralba, and Aude Oliva. Places: An image database for deep scene understanding. Journal of Vision, 17(10):296-296, 2017. 6 \ No newline at end of file diff --git a/2025/FEVER-OOD_ Free Energy Vulnerability Elimination for Robust Out-of-Distribution Detection/images.zip b/2025/FEVER-OOD_ Free Energy Vulnerability Elimination for Robust Out-of-Distribution Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b4eee5146e1c4d9db99c0967106f528a9e777959 --- /dev/null +++ b/2025/FEVER-OOD_ Free Energy Vulnerability Elimination for Robust Out-of-Distribution Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:61131ac228dc93e622069ba5103a16caf47187cad4777ab16ead3bdce37a9f88 +size 702075 diff --git a/2025/FEVER-OOD_ Free Energy Vulnerability Elimination for Robust Out-of-Distribution Detection/layout.json b/2025/FEVER-OOD_ Free Energy Vulnerability Elimination for Robust Out-of-Distribution Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b3b9a5e59c7c3ef612cb48e178fc36b82931a3fa --- /dev/null +++ b/2025/FEVER-OOD_ Free Energy Vulnerability Elimination for Robust Out-of-Distribution Detection/layout.json @@ -0,0 +1,9398 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 104, + 103, + 507, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 103, + 507, + 138 + ], + "spans": [ + { + "bbox": [ + 104, + 103, + 507, + 138 + ], + "type": "text", + "content": "FEVER-OOD: Free Energy Vulnerability Elimination for Robust Out-of-Distribution Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 146, + 575, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 146, + 575, + 175 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 575, + 175 + ], + "type": "text", + "content": "Brian K.S. Isaac-Medina1 Maurizio Che2 Yona Faline A. Gaus1 Samet Akcay3 Toby P. Breckon1 \n1Durham University, Durham, UK 2University of Vienna, Vienna, Austria 3Intel" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 123, + 177, + 514, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 177, + 514, + 203 + ], + "spans": [ + { + "bbox": [ + 123, + 177, + 514, + 203 + ], + "type": "text", + "content": "{brian.k.isaac-medina, yona.f.binti-abd-gaus, toby.breckon}@durham.ac.uk mauricio.adrian.che.moguel@univie.ac.at samet.akcay@intel.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 255, + 297, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 255, + 297, + 651 + ], + "spans": [ + { + "bbox": [ + 54, + 255, + 297, + 651 + ], + "type": "text", + "content": "Modern machine learning models, that excel on computer vision tasks such as classification and object detection, are often overconfident in their predictions for Out-of-Distribution (OOD) examples, resulting in unpredictable behaviour for open-set environments. Recent works have demonstrated that the free energy score is an effective measure of uncertainty for OOD detection given its close relationship to the data distribution. However, despite that free energy-based methods representing a significant empirical advance in OOD detection, our theoretical analysis reveals previously unexplored and inherent vulnerabilities within the free energy score formulation such that in-distribution and OOD instances can have distinct feature representations yet identical free energy scores. This phenomenon occurs when the vector representing the feature space difference between the in-distribution and OOD sample belongs to the null space of the last layer of a neural-based classifier. To mitigate these issues, we explore lower-dimensional feature spaces to reduce the null space footprint and introduce novel regularisation to maximize the least singular value of the final linear layer, hence enhancing intersample free energy separation. We refer to these techniques as Free Energy Vulnerability Elimination for Robust Out-of-Distribution Detection (FEVER-OOD). Extensive experimentation shows that FEVER-OOD improves OOD detection in CIFAR-10 with " + }, + { + "bbox": [ + 54, + 255, + 297, + 651 + ], + "type": "inline_equation", + "content": "28.22\\%" + }, + { + "bbox": [ + 54, + 255, + 297, + 651 + ], + "type": "text", + "content": " FPR95 and 94.78 AUROC vs " + }, + { + "bbox": [ + 54, + 255, + 297, + 651 + ], + "type": "inline_equation", + "content": "33.66\\%" + }, + { + "bbox": [ + 54, + 255, + 297, + 651 + ], + "type": "text", + "content": " FPR95 and 92.15 AUROC of the baseline model, CIFAR-100 with " + }, + { + "bbox": [ + 54, + 255, + 297, + 651 + ], + "type": "inline_equation", + "content": "42.77\\%" + }, + { + "bbox": [ + 54, + 255, + 297, + 651 + ], + "type": "text", + "content": " FPR95 and 89.98 AUROC vs " + }, + { + "bbox": [ + 54, + 255, + 297, + 651 + ], + "type": "inline_equation", + "content": "50.96\\%" + }, + { + "bbox": [ + 54, + 255, + 297, + 651 + ], + "type": "text", + "content": " FPR95 and 88.06 AUROC of the baseline model, and achieves state of the art OOD detection in Imagenet-100, with average OOD FPR95 of " + }, + { + "bbox": [ + 54, + 255, + 297, + 651 + ], + "type": "inline_equation", + "content": "36.50\\%" + }, + { + "bbox": [ + 54, + 255, + 297, + 651 + ], + "type": "text", + "content": " and an AUROC of 92.74 when used with the Dream-OOD model, compared with a " + }, + { + "bbox": [ + 54, + 255, + 297, + 651 + ], + "type": "inline_equation", + "content": "39.33\\%" + }, + { + "bbox": [ + 54, + 255, + 297, + 651 + ], + "type": "text", + "content": " and 91.84 AUROC without FEVER-OOD." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 658, + 135, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 135, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 135, + 670 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Out-of-distribution (OOD) detection (c.f. anomaly/outlier detection), aims to identify abnormal samples that deviate significantly from a given data distribution. Whilst" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 238, + 434, + 328 + ], + "blocks": [ + { + "bbox": [ + 321, + 228, + 428, + 237 + ], + "lines": [ + { + "bbox": [ + 321, + 228, + 428, + 237 + ], + "spans": [ + { + "bbox": [ + 321, + 228, + 428, + 237 + ], + "type": "text", + "content": "Free Energy-based OOD (VOS)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 238, + 434, + 328 + ], + "lines": [ + { + "bbox": [ + 316, + 238, + 434, + 328 + ], + "spans": [ + { + "bbox": [ + 316, + 238, + 434, + 328 + ], + "type": "image", + "image_path": "848d7e663e2933ff92e896c599e56815165ef8b69813df492dfc4ac6e02b7a86.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 440, + 238, + 553, + 328 + ], + "blocks": [ + { + "bbox": [ + 459, + 228, + 530, + 237 + ], + "lines": [ + { + "bbox": [ + 459, + 228, + 530, + 237 + ], + "spans": [ + { + "bbox": [ + 459, + 228, + 530, + 237 + ], + "type": "text", + "content": "FEVER-OOD (VOS)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 440, + 238, + 553, + 328 + ], + "lines": [ + { + "bbox": [ + 440, + 238, + 553, + 328 + ], + "spans": [ + { + "bbox": [ + 440, + 238, + 553, + 328 + ], + "type": "image", + "image_path": "b79fe209468046edba6d570b17932480933ce0e0db970c3334cea190f38ac4df.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 337, + 555, + 381 + ], + "lines": [ + { + "bbox": [ + 313, + 337, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 555, + 381 + ], + "type": "text", + "content": "Figure 1. FEVER-OOD improves baseline Free Energy-based OOD detection methods, where baseline methods predict OOD classes as in-distribution (left, blue), while FEVER-OOD classifies them as OOD (right, green)." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 386, + 555, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 386, + 555, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 555, + 459 + ], + "type": "text", + "content": "OOD is essential in developing deployable machine learning systems, to ensure valid inference only on data drawn from same distribution used in training [24, 36, 44], real-world deployments often involve scenarios where models encounter unseen OOD classes that could otherwise undermine decision stability and hence wider system robustness." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 460, + 556, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 556, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 556, + 628 + ], + "type": "text", + "content": "Methods such as deep ensembles [23], ODIN [27], Mahalanobis distance-based detection [25], and generalized ODIN [16] all utilize softmax() to differentiate indistribution from OOD. Whilst intuitive, such approaches often struggle as deep neural networks commonly yield overly high softmax() scores, even for inputs far from the training data distribution [31]. Given OOD and indistribution samples should have significant spatial separation in any given feature space, Liu et al. [29] developed an alternative scoring function to effectively measure the alignment of a given sample to an a priori learned distribution. Denoted as the free energy scoring function, this formulation demonstrates superior OOD performance across both classification [7-9] and detection tasks [6, 7, 18, 21]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "Despite free energy-based methods representing a significant empirical advance in OOD detection, our theoretical analysis reveals previously unexplored and inherent vulnerabilities within the free energy score formulation that could impact their practical reliability. These vulnerabilities reveal that in-distribution and OOD instances can have distinct feature representations yet produce identical free en" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4529" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "ergy scores. This phenomenon occurs in the feature space when the vector direction representing the difference between an in-distribution and an OOD sample lies within the null space of the last layer of neural-based classifier. Such regions, effectively \"blind spots\" in terms of OOD, lead to OOD and in-distribution generating identical or near-identical free energy score values." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 157, + 294, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 157, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 55, + 157, + 294, + 252 + ], + "type": "text", + "content": "To mitigate these issues, we explore lower-dimensional feature spaces to effectively reduce the size of the null space. In addition, we also introduce a novel regulariser to maximize the least singular value of the final linear layer, hence enhancing inter-sample free energy separation to improve OOD robustness. We refer to these techniques as Free energy Vulnerability Elimination for Robust Out-of-Distribution Detection (FEVER-OOD)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 255, + 264, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 264, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 264, + 266 + ], + "type": "text", + "content": "In summary, our key contributions in this paper are:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 267, + 294, + 650 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 267, + 294, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 267, + 294, + 410 + ], + "spans": [ + { + "bbox": [ + 55, + 267, + 294, + 410 + ], + "type": "text", + "content": "- The identification of inherent free energy score vulnerabilities, where under null-space conditions, in-distribution and OOD instances will yield similar free energy scores despite differing feature representations. Denoted as Null Space Vulnerabilities (NSV) such conditions readily occur when the feature dimension exceeds the number of classes. Additionally, such free energy similarity can be influenced by the least singular vector (LSV) of the last linear layer of a neural-based classifier, leading to what we denote as Least Singular Value Vulnerabilities (LSVV). Both vulnerabilities negatively impact in-distribution to OOD discrimination." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 411, + 294, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 294, + 542 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 294, + 542 + ], + "type": "text", + "content": "- Novel formulations that mitigate the NSV and LSVV impact (Fig. 1). To address NSV, we introduce an additional layer to reduce the null space whilst for LSVV, we develop an LSV Regulariser that maximizes the least singular value, guaranteeing detectable energy variations from small feature-space differences and hence enhancing inter-sample energy scores discrimination. Furthermore, we explore a Condition Number Regulariser to promote balanced energy changes for displacements in all feature-space direction hence ensuring improved uniformity in energy score variations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 543, + 294, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 543, + 294, + 650 + ], + "spans": [ + { + "bbox": [ + 55, + 543, + 294, + 650 + ], + "type": "text", + "content": "- Comprehensive experiments and ablation studies using established benchmark datasets over both object classification and detection tasks, that demonstrate the effectiveness of FEVER-OOD in improving OOD detection performance. Specifically, Dream-OOD [8] with FEVER-OOD trained on Imagenet-100 [35] as in-distribution decreases the FPR95 from " + }, + { + "bbox": [ + 55, + 543, + 294, + 650 + ], + "type": "inline_equation", + "content": "39.33\\%" + }, + { + "bbox": [ + 55, + 543, + 294, + 650 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 543, + 294, + 650 + ], + "type": "inline_equation", + "content": "36.50\\%" + }, + { + "bbox": [ + 55, + 543, + 294, + 650 + ], + "type": "text", + "content": " and increases the average AUROC from 91.84 to 92.74, achieving state-of-the-art performance for OOD detection." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 659, + 142, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 142, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 142, + 672 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 681, + 294, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 681, + 294, + 717 + ], + "spans": [ + { + "bbox": [ + 55, + 681, + 294, + 717 + ], + "type": "text", + "content": "Identifying OOD samples is a crucial yet inherently challenging problem, as models are not exposed to, and hence cannot reliably differentiate, OOD samples." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 528 + ], + "type": "text", + "content": "OOD Detection: to overcome this challenge, researchers have begun incorporating additional data during training to establish a more conservative and safe decision boundary against OOD inputs. Some methods [1, 19] have used unlabeled data to regularise model training while still focusing on classifying labeled in-distribution data, while others [2] leverage human assistance to strategically label examples within a novel maximum disambiguation region. Despite its potential, leveraging unlabeled data is non-trivial due to the heterogeneous mix of in-distribution and OOD instances. The absence of a clean OOD set presents significant challenges in designing effective OOD learning algorithms. A different branch of works incorporate post-hoc methods, focusing on designing OOD scoring mechanisms at test time. Hendrycks and Gimpel [15] introduced the maximum predicted softmax() probability as an in-distribution score function, establishing an initial baseline for scoring OOD detection. Similarly, ODIN [27] aims to widen the gap between values in the softmax() score vectors by temperature scaling and gradient-based input perturbation, enhancing the separation between in-distribution and OOD instances. Lee et al. [25] use the Mahalanobis distance between the average softmax() vectors of images and incorporate intermediate features to improve detection performance and robustness. Nonetheless, the misalignment between the softmax() score and the true data density [29, 32] ultimately makes these techniques a suboptimal solution to OOD detection. Closer to our work, Cook et al. [4] identify null space adversarial attacks, proposing that instances with large magnitude of their projection onto the span of the null space of a neural network classifier would likely be an outlier. Similarly, Wang et al. [41] propose that since the features lie in low-dimensional manifolds, outliers are likely to deviate from the principal subspace, i.e. the span of the eigenvectors or the largest eigenvalues, of the in-distribution data. Therefore, OOD instances are detected with respect to the magnitude of the component of the perpendicular subspace to the principal subspace." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "content": "Energy-based OOD Detection: to address these limitations, Liu et al. [29] propose to use a free energy-based in-distribution score function instead of the softmax() score to identify OOD instances. This approach regularises the model to produce lower energy values for in-distribution and higher energy values for auxiliary outliers, creating a significant energy gap that enhances the model ability to detect OOD samples. Energy-based OOD modeling has enabled novel outlier synthesis paradigms. For instance, Du et al. [7] propose virtual outlier synthesis (VOS) by sampling outliers from low-likelihood in-distribution regions. VOS has shown to be effective for classification and object-level OOD detection. Further outlier synthesis techniques based on VOS have been proposed, including using normalizing flows [21], sampling based on the in-distribution boundary" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "type": "text", + "content": "4530" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 263 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 263 + ], + "type": "text", + "content": "instances [39] or extending VOS when class labels are unavailable [18]. In addition, VOS techniques arensuitable for object detection [7, 21]. Whilst VOS and subsequent works synthesize outliers in the feature space, Dream-OOD [8] uses Stable Diffusion [34] to generate outliers in the pixel-space, learning better discriminative free-score values for OOD instances and providing visual examples of outliers. Energy-based methods have shown great performance, becoming state-of-the-art in several benchmarks (e.g., Dream-OOD [8] for ImageNet-100 in-distribution), hence we focus in the properties of these methods. Notwithstanding that the core idea of the energy score is that OOD samples must have a significantly different free energy than in-distribution sample, we examine some inherent vulnerabilities in the energy formulation based on the null space and least singular value of the last linear layer of neural-based classifiers." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 268, + 242, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 268, + 242, + 282 + ], + "spans": [ + { + "bbox": [ + 55, + 268, + 242, + 282 + ], + "type": "text", + "content": "3. Preliminaries: Free Energy Score" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 288, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 296, + 445 + ], + "type": "text", + "content": "The goal of OOD detection is to detect outliers, " + }, + { + "bbox": [ + 55, + 288, + 296, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 55, + 288, + 296, + 445 + ], + "type": "text", + "content": ", that have a low probability in the true (unknown) in-distribution density " + }, + { + "bbox": [ + 55, + 288, + 296, + 445 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 288, + 296, + 445 + ], + "type": "text", + "content": ". However, the computation of " + }, + { + "bbox": [ + 55, + 288, + 296, + 445 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 288, + 296, + 445 + ], + "type": "text", + "content": " is usually unfeasible or computationally untractable, so we instead resort to additional information. Now consider that we also have a set of in-distribution categories " + }, + { + "bbox": [ + 55, + 288, + 296, + 445 + ], + "type": "inline_equation", + "content": "\\{1,\\dots ,K\\}" + }, + { + "bbox": [ + 55, + 288, + 296, + 445 + ], + "type": "text", + "content": " such that if we know the joint probability distribution, " + }, + { + "bbox": [ + 55, + 288, + 296, + 445 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x},k)" + }, + { + "bbox": [ + 55, + 288, + 296, + 445 + ], + "type": "text", + "content": ", for " + }, + { + "bbox": [ + 55, + 288, + 296, + 445 + ], + "type": "inline_equation", + "content": "k\\in \\{1,\\ldots ,K\\}" + }, + { + "bbox": [ + 55, + 288, + 296, + 445 + ], + "type": "text", + "content": ", we could marginalize " + }, + { + "bbox": [ + 55, + 288, + 296, + 445 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 288, + 296, + 445 + ], + "type": "text", + "content": ". Nonetheless, since the actual joint probabilities " + }, + { + "bbox": [ + 55, + 288, + 296, + 445 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x},k)" + }, + { + "bbox": [ + 55, + 288, + 296, + 445 + ], + "type": "text", + "content": " might be also untractable, we could instead use non-negative arbitrary potential functions " + }, + { + "bbox": [ + 55, + 288, + 296, + 445 + ], + "type": "inline_equation", + "content": "\\phi (\\mathbf{x},k)" + }, + { + "bbox": [ + 55, + 288, + 296, + 445 + ], + "type": "text", + "content": " that measure the affinity of the states of " + }, + { + "bbox": [ + 55, + 288, + 296, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 288, + 296, + 445 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 288, + 296, + 445 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 288, + 296, + 445 + ], + "type": "text", + "content": " [12]. In this sense, we can construct the unnormalized probability:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 451, + 295, + 484 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 451, + 295, + 484 + ], + "spans": [ + { + "bbox": [ + 134, + 451, + 295, + 484 + ], + "type": "interline_equation", + "content": "\\tilde {p} (\\mathbf {x}) = \\sum_ {k = 1} ^ {K} \\phi (\\mathbf {x}, k) \\tag {1}", + "image_path": "0107b0badcd6f1e7de8c6a251471368beb1f90206a26dd2526d31823c58ed45a.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 491, + 295, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 295, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 295, + 515 + ], + "type": "text", + "content": "that gives an idea of the likelihood " + }, + { + "bbox": [ + 55, + 491, + 295, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 491, + 295, + 515 + ], + "type": "text", + "content": ". Furthermore, energy-based models enforce the non-negative constraint of " + }, + { + "bbox": [ + 55, + 491, + 295, + 515 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 491, + 295, + 515 + ], + "type": "text", + "content": " by:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 117, + 522, + 295, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 522, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 117, + 522, + 295, + 536 + ], + "type": "interline_equation", + "content": "\\phi (\\mathbf {x}, k) = \\exp (- E (\\mathbf {x}, k)), \\tag {2}", + "image_path": "31b6f3e75c8e632be867bead9accc871229f5e556ed6df8a61b8c58d271c551d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 543, + 296, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 543, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 543, + 296, + 581 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 543, + 296, + 581 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 55, + 543, + 296, + 581 + ], + "type": "text", + "content": " is an unbounded arbitrary function called the energy. By substituting Eq. (2) into Eq. (1) and taking the negative " + }, + { + "bbox": [ + 55, + 543, + 296, + 581 + ], + "type": "inline_equation", + "content": "\\log^1" + }, + { + "bbox": [ + 55, + 543, + 296, + 581 + ], + "type": "text", + "content": ", we get a value known as the free energy:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 587, + 295, + 619 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 587, + 295, + 619 + ], + "spans": [ + { + "bbox": [ + 69, + 587, + 295, + 619 + ], + "type": "interline_equation", + "content": "\\mathcal {F} (\\mathbf {x}) = - \\log \\tilde {p} (\\mathbf {x}) = - \\log \\sum_ {k = 1} ^ {K} \\exp (- E (\\mathbf {x}, k)), \\tag {3}", + "image_path": "66c9ab4c61127ee9b4dffe23ebecfe5be78fb9be2367885b63e2dd44873bc064.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 626, + 296, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 626, + 296, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 626, + 296, + 686 + ], + "type": "text", + "content": "and more recently used as a concept in OOD detection [7, 8, 18, 21, 29]. To get the actual density " + }, + { + "bbox": [ + 55, + 626, + 296, + 686 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 626, + 296, + 686 + ], + "type": "text", + "content": ", we must compute the normalizing partition function of Eq. (1); since we are interested in distinguishing between in-distribution and OOD, " + }, + { + "bbox": [ + 55, + 626, + 296, + 686 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 626, + 296, + 686 + ], + "type": "text", + "content": " thus serves well for this purpose." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 315, + 69, + 553, + 180 + ], + "blocks": [ + { + "bbox": [ + 315, + 69, + 553, + 180 + ], + "lines": [ + { + "bbox": [ + 315, + 69, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 315, + 69, + 553, + 180 + ], + "type": "image", + "image_path": "bdd060b5b3baf52a609a84f74c4399fcafd9d4e86650030073fdec620e75a556.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 188, + 553, + 222 + ], + "lines": [ + { + "bbox": [ + 313, + 188, + 553, + 222 + ], + "spans": [ + { + "bbox": [ + 313, + 188, + 553, + 222 + ], + "type": "text", + "content": "Figure 2. Vulnerabilities of Free Energy-based OOD detection. Directions in the Null Space do not change the energy, while the LSV direction has minimal change, compared to other directions." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 228, + 555, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 324 + ], + "type": "text", + "content": "In a general machine learning context, a classifier " + }, + { + "bbox": [ + 313, + 228, + 555, + 324 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x};\\theta):\\mathbb{R}^d\\to \\mathbb{R}^K" + }, + { + "bbox": [ + 313, + 228, + 555, + 324 + ], + "type": "text", + "content": " with learned parameters " + }, + { + "bbox": [ + 313, + 228, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 228, + 555, + 324 + ], + "type": "text", + "content": " assigns categorical unnormalized probabilities " + }, + { + "bbox": [ + 313, + 228, + 555, + 324 + ], + "type": "inline_equation", + "content": "f_{k}(\\mathbf{x};\\theta)" + }, + { + "bbox": [ + 313, + 228, + 555, + 324 + ], + "type": "text", + "content": ", called logits, for a given " + }, + { + "bbox": [ + 313, + 228, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\mathbf{x}\\in \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 228, + 555, + 324 + ], + "type": "text", + "content": ". While the logits are usually normalized using the softmax() function in order to obtain a probability distribution, they can also be used as the energy " + }, + { + "bbox": [ + 313, + 228, + 555, + 324 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 313, + 228, + 555, + 324 + ], + "type": "text", + "content": " in Eq. (2). This realization enables the use of the free energy (Eq. (3)) as a measure of uncertainty for a classifier:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 364, + 333, + 553, + 366 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 333, + 553, + 366 + ], + "spans": [ + { + "bbox": [ + 364, + 333, + 553, + 366 + ], + "type": "interline_equation", + "content": "\\mathcal {F} (\\mathbf {x}) = - \\log \\sum_ {k = 1} ^ {K} \\exp \\left(f _ {k} (\\mathbf {x}; \\theta)\\right). \\tag {4}", + "image_path": "765358f88030e327baee55c55e880c524a6b0e06752b9770e41186d3f59567e5.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 374, + 554, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 554, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 554, + 434 + ], + "type": "text", + "content": "Liu et al. [29] demonstrate that Eq. (4) is superior for effective OOD detection when compared to use of softmax() whilst Du et al. [7] explicitly use the free energy score (Eq. (4)) to perform binary classification between indistribution and OOD by training with synthetic outliers." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 445, + 534, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 534, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 534, + 459 + ], + "type": "text", + "content": "4. Vulnerabilities of the Free Energy Score" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 465, + 554, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 465, + 554, + 561 + ], + "spans": [ + { + "bbox": [ + 313, + 465, + 554, + 561 + ], + "type": "text", + "content": "Although the free energy (Eq. (4)) is an effective uncertainty measure for OOD detection [7, 29], here we examine some inherent vulnerabilities when used in conjunction with a neural network based classifier. Firstly, Sec. 4.1 investigates OOD instances that have similar energies as in-distribution samples, whilst Sec. 4.2 discusses minimal changes in free energy, and finally Sec. 4.3 outlines our proposed approach to minimize the impact of these vulnerabilities." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 568, + 452, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 452, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 452, + 582 + ], + "type": "text", + "content": "4.1. Null Space Vulnerability" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 586, + 554, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 554, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 554, + 658 + ], + "type": "text", + "content": "Consider a neural network-based classifier " + }, + { + "bbox": [ + 313, + 586, + 554, + 658 + ], + "type": "inline_equation", + "content": "f = (g \\circ h)(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 586, + 554, + 658 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 586, + 554, + 658 + ], + "type": "inline_equation", + "content": "h(\\mathbf{x}): \\mathbb{R}^d \\to \\mathbb{R}^{d'}" + }, + { + "bbox": [ + 313, + 586, + 554, + 658 + ], + "type": "text", + "content": " encodes the input into a " + }, + { + "bbox": [ + 313, + 586, + 554, + 658 + ], + "type": "inline_equation", + "content": "d'" + }, + { + "bbox": [ + 313, + 586, + 554, + 658 + ], + "type": "text", + "content": "-dimensional feature vector and " + }, + { + "bbox": [ + 313, + 586, + 554, + 658 + ], + "type": "inline_equation", + "content": "g: \\mathbb{R}^{d'} \\to \\mathbb{R}^K" + }, + { + "bbox": [ + 313, + 586, + 554, + 658 + ], + "type": "text", + "content": " is the last layer of the neural network that transforms the features into unnormalized class probabilities (i.e., the logits). If " + }, + { + "bbox": [ + 313, + 586, + 554, + 658 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 586, + 554, + 658 + ], + "type": "text", + "content": " is a linear layer (as for most classifiers [5, 14, 43]), then:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 393, + 666, + 553, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 666, + 553, + 681 + ], + "spans": [ + { + "bbox": [ + 393, + 666, + 553, + 681 + ], + "type": "interline_equation", + "content": "f (\\mathbf {x}) = \\mathrm {W} _ {c l s} ^ {\\top} h (\\mathbf {x}), \\tag {5}", + "image_path": "1b57237e927f3b9d9b5dd2914d871f1f5a0aa40652379c2a407981852a7ed6de.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{W}_{cls} \\in \\mathbb{R}^{d' \\times K}" + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": " are learnable weights. Hence, the logits " + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "inline_equation", + "content": "f_{k}(\\mathbf{x};\\theta)" + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": " correspond to the " + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "-th index of " + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 296, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 296, + 713 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 55, + 693, + 296, + 713 + ], + "type": "text", + "content": "N.B. the signs of the free energy and the argument of the exponent of " + }, + { + "bbox": [ + 55, + 693, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 693, + 296, + 713 + ], + "type": "text", + "content": " are arbitrarily negative for historical reasons." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "4531" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": "The core idea of energy-based OOD is that outliers " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " have a significantly different free energy, " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\mathbf{v})" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": ", from in-distribution samples, " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": ". Since OOD instances should be far from in-distribution samples in the feature space [7, 39], the expected distance from the inlier features " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "h(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " and outlier features " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "h(\\mathbf{v})" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " should be greater than a given boundary distance, " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "d_b" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": ", as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 122, + 163, + 296, + 178 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 163, + 296, + 178 + ], + "spans": [ + { + "bbox": [ + 122, + 163, + 296, + 178 + ], + "type": "interline_equation", + "content": "\\mathbb {E} \\left[ \\| h (\\mathbf {x}) - h (\\mathbf {v}) \\| \\right] \\geq d _ {b}, \\tag {6}", + "image_path": "e06ccf197eda4483910d3313a65b828a864afaf66042748293a7cb1611cc62a1.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 184, + 296, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 184, + 296, + 245 + ], + "spans": [ + { + "bbox": [ + 55, + 184, + 296, + 245 + ], + "type": "text", + "content": "for a norm " + }, + { + "bbox": [ + 55, + 184, + 296, + 245 + ], + "type": "inline_equation", + "content": "\\|\\cdot\\|" + }, + { + "bbox": [ + 55, + 184, + 296, + 245 + ], + "type": "text", + "content": ". In other words, the inliers remain bounded to a region in the feature space while outliers are unbounded. Therefore, the expected distance between inliers and outliers is greater than a bounding distance " + }, + { + "bbox": [ + 55, + 184, + 296, + 245 + ], + "type": "inline_equation", + "content": "d_b" + }, + { + "bbox": [ + 55, + 184, + 296, + 245 + ], + "type": "text", + "content": ". Subsequently, we can rewrite Eq. (6) as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 136, + 252, + 295, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 252, + 295, + 266 + ], + "spans": [ + { + "bbox": [ + 136, + 252, + 295, + 266 + ], + "type": "interline_equation", + "content": "h (\\mathbf {v}) = h (\\mathbf {x}) + \\boldsymbol {\\delta}, \\tag {7}", + "image_path": "b021e30274d412415125683c6e83885ecc760d6d7494bfc49da837a4c6dd4d73.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 273, + 296, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 273, + 296, + 299 + ], + "spans": [ + { + "bbox": [ + 55, + 273, + 296, + 299 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 273, + 296, + 299 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 273, + 296, + 299 + ], + "type": "text", + "content": " is a vector with " + }, + { + "bbox": [ + 55, + 273, + 296, + 299 + ], + "type": "inline_equation", + "content": "\\mathbb{E}[\\|\\delta\\|] \\geq d_b" + }, + { + "bbox": [ + 55, + 273, + 296, + 299 + ], + "type": "text", + "content": ". By substituting Eq. (7) into Eq. (5) we obtain the following:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 85, + 304, + 296, + 320 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 304, + 296, + 320 + ], + "spans": [ + { + "bbox": [ + 85, + 304, + 296, + 320 + ], + "type": "interline_equation", + "content": "f (\\mathbf {v}) = \\mathrm {W} _ {c l s} ^ {\\top} (h (\\mathbf {x}) + \\boldsymbol {\\delta}) = f (\\mathbf {x}) + \\mathrm {W} _ {c l s} ^ {\\top} \\boldsymbol {\\delta}. \\qquad (8)", + "image_path": "7b5c2a22ccd7a5ac41f3bbafafdb58247cc23cbb213d343644da6fcb2cdf0340.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 326, + 296, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 326, + 296, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 296, + 480 + ], + "type": "text", + "content": "However within this formulation, if " + }, + { + "bbox": [ + 55, + 326, + 296, + 480 + ], + "type": "inline_equation", + "content": "\\mathrm{W}_{cls}" + }, + { + "bbox": [ + 55, + 326, + 296, + 480 + ], + "type": "text", + "content": " has more rows than columns, which occurs when the feature dimension is larger than the number of classes, the size of the null space of " + }, + { + "bbox": [ + 55, + 326, + 296, + 480 + ], + "type": "inline_equation", + "content": "\\mathrm{W}_{cls}^{\\top}" + }, + { + "bbox": [ + 55, + 326, + 296, + 480 + ], + "type": "text", + "content": ", called the nullity, is nullity " + }, + { + "bbox": [ + 55, + 326, + 296, + 480 + ], + "type": "inline_equation", + "content": "(\\mathrm{W}_{cls}^{\\top}) = d' - \\operatorname{rank}(\\mathrm{W}_{cls})" + }, + { + "bbox": [ + 55, + 326, + 296, + 480 + ], + "type": "text", + "content": " due to the rank-nullity theorem. Mathematically, this means that there exists a linear subspace " + }, + { + "bbox": [ + 55, + 326, + 296, + 480 + ], + "type": "inline_equation", + "content": "\\mathrm{Null}(\\mathrm{W}_{cls}^{\\top}) \\subset \\mathbb{R}^{d'}" + }, + { + "bbox": [ + 55, + 326, + 296, + 480 + ], + "type": "text", + "content": " such that if " + }, + { + "bbox": [ + 55, + 326, + 296, + 480 + ], + "type": "inline_equation", + "content": "\\delta \\in \\mathrm{Null}(\\mathrm{W}_{cls}^{\\top})" + }, + { + "bbox": [ + 55, + 326, + 296, + 480 + ], + "type": "text", + "content": " then " + }, + { + "bbox": [ + 55, + 326, + 296, + 480 + ], + "type": "inline_equation", + "content": "\\mathrm{W}_{cls}^{\\top} \\delta = 0" + }, + { + "bbox": [ + 55, + 326, + 296, + 480 + ], + "type": "text", + "content": ". Within our OOD context, if an outlier instance, " + }, + { + "bbox": [ + 55, + 326, + 296, + 480 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 55, + 326, + 296, + 480 + ], + "type": "text", + "content": ", has a feature representation, " + }, + { + "bbox": [ + 55, + 326, + 296, + 480 + ], + "type": "inline_equation", + "content": "h(\\mathbf{v})" + }, + { + "bbox": [ + 55, + 326, + 296, + 480 + ], + "type": "text", + "content": ", such that its difference with respect to the inlier feature representation, " + }, + { + "bbox": [ + 55, + 326, + 296, + 480 + ], + "type": "inline_equation", + "content": "h(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 326, + 296, + 480 + ], + "type": "text", + "content": ", of any inlier instance, " + }, + { + "bbox": [ + 55, + 326, + 296, + 480 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 326, + 296, + 480 + ], + "type": "text", + "content": ", is in the null space of the last layer of the classifier, then their free energy scores are the same:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 79, + 488, + 295, + 504 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 488, + 295, + 504 + ], + "spans": [ + { + "bbox": [ + 79, + 488, + 295, + 504 + ], + "type": "interline_equation", + "content": "h (\\mathbf {v}) - h (\\mathbf {x}) \\in \\operatorname {N u l l} \\left(\\mathrm {W} _ {c l s} ^ {\\top}\\right) \\Rightarrow \\mathcal {F} (\\mathbf {x}) = \\mathcal {F} (\\mathbf {v}). \\tag {9}", + "image_path": "75c6c1bb546416f07d4d591b039fc146d12b8e0afbf543e65e8eb3a102eebe5c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 510, + 296, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 641 + ], + "type": "text", + "content": "Practically, Eq. (9) is telling us that it is possible for in-distribution and OOD instances to have different feature representations (satisfying the condition of Eq. (6)) and yet have the same free energy score, as depicted in Fig. 2. This fundamentally challenges the notion that free energy is an effective mechanism for OOD detection as any approaches reliant on Eq. (3), such as [7, 8, 18, 21, 29], contain inherent \"detection blind-spots\" as per Eq. (9). We denote Eq. (9) as the Null Space Vulnerabilities (NSV) of classifier " + }, + { + "bbox": [ + 55, + 510, + 296, + 641 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 510, + 296, + 641 + ], + "type": "text", + "content": " with size nullity " + }, + { + "bbox": [ + 55, + 510, + 296, + 641 + ], + "type": "inline_equation", + "content": "(\\mathrm{W}_{cls}^{\\top}) = d' - \\mathrm{rank}(\\mathrm{W}_{cls})" + }, + { + "bbox": [ + 55, + 510, + 296, + 641 + ], + "type": "text", + "content": ". An approach to minimize this vulnerability is outlined in Sec. 4.3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "As an additional remark, NSV are inherent in energy-based methods; while some post hoc methods use the projection of features onto the null space as a discriminator between in-distribution and OOD [4, 41], energy-based methods use the classification logits for detecting outliers, hence the null space projection does not change the OOD score." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 489, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 489, + 85 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 489, + 85 + ], + "type": "text", + "content": "4.2. Minimal Change of Free Energy" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 89, + 555, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 555, + 173 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 555, + 173 + ], + "type": "text", + "content": "We now turn our attention to the case where " + }, + { + "bbox": [ + 313, + 89, + 555, + 173 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 89, + 555, + 173 + ], + "type": "text", + "content": " is not in the null space of " + }, + { + "bbox": [ + 313, + 89, + 555, + 173 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{cls}^{\\top}" + }, + { + "bbox": [ + 313, + 89, + 555, + 173 + ], + "type": "text", + "content": ". Furthermore, we can focus on " + }, + { + "bbox": [ + 313, + 89, + 555, + 173 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 89, + 555, + 173 + ], + "type": "text", + "content": " orthogonal to " + }, + { + "bbox": [ + 313, + 89, + 555, + 173 + ], + "type": "inline_equation", + "content": "\\mathrm{Null}(\\mathbf{W}_{cls}^{\\top})" + }, + { + "bbox": [ + 313, + 89, + 555, + 173 + ], + "type": "text", + "content": ", since every vector in the domain of " + }, + { + "bbox": [ + 313, + 89, + 555, + 173 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{cls}^{\\top}" + }, + { + "bbox": [ + 313, + 89, + 555, + 173 + ], + "type": "text", + "content": " can be decomposed into one component in " + }, + { + "bbox": [ + 313, + 89, + 555, + 173 + ], + "type": "inline_equation", + "content": "\\mathrm{Null}(\\mathbf{W}_{cls}^{\\top})" + }, + { + "bbox": [ + 313, + 89, + 555, + 173 + ], + "type": "text", + "content": " and one component in " + }, + { + "bbox": [ + 313, + 89, + 555, + 173 + ], + "type": "inline_equation", + "content": "\\mathrm{Null}(\\mathbf{W}_{cls}^{\\top})^{\\perp}" + }, + { + "bbox": [ + 313, + 89, + 555, + 173 + ], + "type": "text", + "content": ", and the effect of the component in " + }, + { + "bbox": [ + 313, + 89, + 555, + 173 + ], + "type": "inline_equation", + "content": "\\mathrm{Null}(\\mathbf{W}_{cls}^{\\top})" + }, + { + "bbox": [ + 313, + 89, + 555, + 173 + ], + "type": "text", + "content": " will be dealt with using the null space reduction method (Sec. 4.3)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 174, + 555, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 174, + 555, + 234 + ], + "spans": [ + { + "bbox": [ + 313, + 174, + 555, + 234 + ], + "type": "text", + "content": "In order to create more discriminative OOD detectors based on the free energy score, we investigate cases where the difference in OOD and in-distribution free energy is minimal. We can state our problem formulation to find the case that minimizes the free energy difference:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 328, + 242, + 554, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 242, + 554, + 312 + ], + "spans": [ + { + "bbox": [ + 328, + 242, + 554, + 312 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {\\substack {\\boldsymbol {\\delta}: \\| \\boldsymbol {\\delta} \\| \\geq d _ {b}, \\\\ \\boldsymbol {\\delta} \\in \\operatorname {Null} \\left(\\mathrm {W} _ {c l s} ^ {\\top}\\right) ^ {\\perp}}} | \\mathcal {F} (\\mathbf {v}) - \\mathcal {F} (\\mathbf {x}) | \\\\ = \\min _ {\\substack {\\boldsymbol {\\delta}: \\| \\boldsymbol {\\delta} \\| \\geq d _ {b}, \\\\ \\boldsymbol {\\delta} \\in \\operatorname {Null} \\left(\\mathrm {W} _ {c l s} ^ {\\top}\\right) ^ {\\perp}}} \\left| \\log \\frac {\\sum_ {k} ^ {K} \\exp \\left(f _ {k} (h (\\mathbf {x}))\\right)}{\\sum_ {k} ^ {K} \\exp \\left(f _ {k} (h (\\mathbf {x}) + \\boldsymbol {\\delta})\\right)} \\right|. \\end{array} \\tag{10}", + "image_path": "452a87f32824f18dc1b1d18f5153fcf8e62d96537e29a1c895e71c1f99be92a1.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 320, + 555, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 320, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 555, + 430 + ], + "type": "text", + "content": "Whilst in Eq. (10) we set " + }, + { + "bbox": [ + 313, + 320, + 555, + 430 + ], + "type": "inline_equation", + "content": "\\| \\delta \\| \\geq d_b" + }, + { + "bbox": [ + 313, + 320, + 555, + 430 + ], + "type": "text", + "content": " instead of " + }, + { + "bbox": [ + 313, + 320, + 555, + 430 + ], + "type": "inline_equation", + "content": "\\mathbb{E}\\left[\\| \\delta \\| \\right]\\geq" + }, + { + "bbox": [ + 313, + 320, + 555, + 430 + ], + "type": "inline_equation", + "content": "d_{b}" + }, + { + "bbox": [ + 313, + 320, + 555, + 430 + ], + "type": "text", + "content": ", in reality we are just interested in the case of non-zero " + }, + { + "bbox": [ + 313, + 320, + 555, + 430 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 320, + 555, + 430 + ], + "type": "text", + "content": " since otherwise the solution would be trivial. Whilst the solution to Eq. (10) may be challenging to obtain, we note the difference in free-energies tends to 0 if the value inside the " + }, + { + "bbox": [ + 313, + 320, + 555, + 430 + ], + "type": "inline_equation", + "content": "\\log ()" + }, + { + "bbox": [ + 313, + 320, + 555, + 430 + ], + "type": "text", + "content": " function tends to 1. This is equivalent to making " + }, + { + "bbox": [ + 313, + 320, + 555, + 430 + ], + "type": "inline_equation", + "content": "f_{k}(h(\\mathbf{x}))" + }, + { + "bbox": [ + 313, + 320, + 555, + 430 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 320, + 555, + 430 + ], + "type": "inline_equation", + "content": "f_{k}(h(\\mathbf{x}) + \\delta)" + }, + { + "bbox": [ + 313, + 320, + 555, + 430 + ], + "type": "text", + "content": " similar for all " + }, + { + "bbox": [ + 313, + 320, + 555, + 430 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 320, + 555, + 430 + ], + "type": "text", + "content": ". Therefore, considering that " + }, + { + "bbox": [ + 313, + 320, + 555, + 430 + ], + "type": "inline_equation", + "content": "f_{k}" + }, + { + "bbox": [ + 313, + 320, + 555, + 430 + ], + "type": "text", + "content": " is in the form of Eq. (5), our problem consists on finding the following proxy minimum:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 334, + 437, + 554, + 500 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 437, + 554, + 500 + ], + "spans": [ + { + "bbox": [ + 334, + 437, + 554, + 500 + ], + "type": "interline_equation", + "content": "\\begin{array}{r l} \\underset {\\delta : \\| \\boldsymbol {\\delta} \\| \\geq d _ {b},} {\\min } & \\| \\mathrm {W} _ {c l s} ^ {\\top} (h (\\mathbf {x}) + \\boldsymbol {\\delta}) - \\mathrm {W} _ {c l s} ^ {\\top} h (\\mathbf {x}) \\| \\\\ & = \\underset {\\substack {\\boldsymbol {\\delta}: \\| \\boldsymbol {\\delta} \\| \\geq d _ {b}, \\\\ \\boldsymbol {\\delta} \\in \\operatorname {Null} \\left(\\mathrm {W} _ {c l s} ^ {\\top}\\right) ^ {\\perp}}} {\\min } \\| \\mathrm {W} _ {c l s} ^ {\\top} \\boldsymbol {\\delta} \\|. \\end{array} \\tag{11}", + "image_path": "cdfc9fa733a738beb015b90d123e490deeae50a22dae39489772c509daec3f6a.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 510, + 496, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 496, + 522 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 496, + 522 + ], + "type": "text", + "content": "For the " + }, + { + "bbox": [ + 313, + 510, + 496, + 522 + ], + "type": "inline_equation", + "content": "L^2" + }, + { + "bbox": [ + 313, + 510, + 496, + 522 + ], + "type": "text", + "content": "-norm in Eq. (11), we thus obtain:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 348, + 530, + 555, + 560 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 530, + 555, + 560 + ], + "spans": [ + { + "bbox": [ + 348, + 530, + 555, + 560 + ], + "type": "interline_equation", + "content": "\\min _ {\\substack {\\boldsymbol {\\delta}: \\| \\boldsymbol {\\delta} \\| _ {2} \\geq d _ {b}, \\\\ \\boldsymbol {\\delta} \\in \\operatorname {Null} \\left(\\mathrm {W} _ {c l s} ^ {\\top}\\right) ^ {\\perp}}} \\| \\mathrm {W} _ {c l s} ^ {\\top} \\boldsymbol {\\delta} \\| _ {2} = d _ {b} \\sigma_ {m i n} \\left(\\mathrm {W} _ {c l s}\\right), \\tag{12}", + "image_path": "a122dfaf675872dcc61b63d222a437297f9b1f61c673b5d07c395660dd1273c7.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 570, + 554, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 594 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 570, + 554, + 594 + ], + "type": "inline_equation", + "content": "\\sigma_{\\min}(\\mathrm{W}_{cls})" + }, + { + "bbox": [ + 313, + 570, + 554, + 594 + ], + "type": "text", + "content": " is the least singular value of " + }, + { + "bbox": [ + 313, + 570, + 554, + 594 + ], + "type": "inline_equation", + "content": "\\mathrm{W}_{cls}" + }, + { + "bbox": [ + 313, + 570, + 554, + 594 + ], + "type": "text", + "content": " (see further mathematical formulation details in Supp. Mat.)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": "Subsequently, Eq. (12) holds for the case where " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": " is in the direction corresponding to the least singular value, such that the change in the free energy will be smaller than any other direction for " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": " (and considering " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\delta \\in \\mathrm{Null}(\\mathrm{W}_{cls}^{\\top})^{\\perp}" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": "). We name this result (from Eq. (12)) the Least Singular Value Vulnerabilities (LSVV) of " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": ". In our OOD context, it means that there could exist some outliers far from the inlier distribution, in terms of Eq. (6), with similar free energy scores because they may fall in the direction of the least singular value of " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{W}_{cls}" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": ". Again, this fundamentally challenges" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4532" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": "the notion that free energy is an effective mechanism for OOD as free energy has been shown to be non-linear with respect to a given feature representation, " + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "inline_equation", + "content": "h(\\cdot)" + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 109, + 296, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 109, + 296, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 109, + 296, + 204 + ], + "type": "text", + "content": "This analysis is itself inspired by the seminal adversarial attacks analysis of Simon-Gabriel et al. [37], where they instead maximize the change in the classifier response for small input perturbations. Whilst here we are modelling the opposite problem, (i.e., the minimum change in the classifier for perturbations larger than a non-zero value), we could consider this analysis as an effective adversarial attack mechanism for the OOD domain." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 212, + 263, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 212, + 263, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 212, + 263, + 224 + ], + "type": "text", + "content": "4.3. Null Space Reduction & Regularisation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 229, + 296, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 229, + 296, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 229, + 296, + 289 + ], + "type": "text", + "content": "In this section, we introduce our Free Energy Vulnerability Elimination for Robust OOD Detection (FEVER-OOD) framework, where we explore readily available methods to minimize the NSV (Sec. 4.1) and LSVV (Sec. 4.2) impact potential within free energy score based OOD approaches." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 290, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 290, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 290, + 296, + 396 + ], + "type": "text", + "content": "Null Space Reduction (NSR). For NSV, we propose to reduce the null space by adding an extra linear layer " + }, + { + "bbox": [ + 55, + 290, + 296, + 396 + ], + "type": "inline_equation", + "content": "g' : \\mathbb{R}^{d'} \\to \\mathbb{R}^r" + }, + { + "bbox": [ + 55, + 290, + 296, + 396 + ], + "type": "text", + "content": ", where we choose " + }, + { + "bbox": [ + 55, + 290, + 296, + 396 + ], + "type": "inline_equation", + "content": "r < d'" + }, + { + "bbox": [ + 55, + 290, + 296, + 396 + ], + "type": "text", + "content": " to reduce the size of the NSV (Fig. 3). We denote this method as " + }, + { + "bbox": [ + 55, + 290, + 296, + 396 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 290, + 296, + 396 + ], + "type": "text", + "content": "-null space reduction (" + }, + { + "bbox": [ + 55, + 290, + 296, + 396 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 290, + 296, + 396 + ], + "type": "text", + "content": "-NSR). Therefore, the new classifier becomes " + }, + { + "bbox": [ + 55, + 290, + 296, + 396 + ], + "type": "inline_equation", + "content": "f' = g \\circ g' \\circ h" + }, + { + "bbox": [ + 55, + 290, + 296, + 396 + ], + "type": "text", + "content": " such that the feature representation of a sample is in " + }, + { + "bbox": [ + 55, + 290, + 296, + 396 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^r" + }, + { + "bbox": [ + 55, + 290, + 296, + 396 + ], + "type": "text", + "content": " and is given instead by " + }, + { + "bbox": [ + 55, + 290, + 296, + 396 + ], + "type": "inline_equation", + "content": "(g' \\circ h)(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 290, + 296, + 396 + ], + "type": "text", + "content": ". Effectively, within an OOD context, we are promoting better discrimination between OOD and in-distribution instances." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 397, + 296, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 397, + 296, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 397, + 296, + 433 + ], + "type": "text", + "content": "Least Singular Value Regulariser (LSVR). To minimize the impact of LSVV, we maximise " + }, + { + "bbox": [ + 55, + 397, + 296, + 433 + ], + "type": "inline_equation", + "content": "\\sigma_{min}(\\mathrm{W}_{cls})" + }, + { + "bbox": [ + 55, + 397, + 296, + 433 + ], + "type": "text", + "content": " by adding an LSVR to the loss function " + }, + { + "bbox": [ + 55, + 397, + 296, + 433 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 55, + 397, + 296, + 433 + ], + "type": "text", + "content": " of a baseline method:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 442, + 295, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 442, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 295, + 456 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {L S V} := \\mathcal {L} + \\lambda_ {L S V} \\sigma_ {\\min } ^ {- 1} \\left(\\mathrm {W} _ {c l s}\\right), \\tag {13}", + "image_path": "429ee4be4598a47503084fad0021706cc7be88c2d7bb2d65c20f3654bae794de.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 465, + 296, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 465, + 296, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 465, + 296, + 548 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 465, + 296, + 548 + ], + "type": "inline_equation", + "content": "\\lambda_{LSV}" + }, + { + "bbox": [ + 55, + 465, + 296, + 548 + ], + "type": "text", + "content": " is a scalar hyperparameter that controls the contribution of the LSV regulariser. It is additionally noted that alternatively maximizing the LSV might also increase the rest of the singular values, resulting in more significant changes of energy in all directions. In the context of OOD detection, LSVR increases the discriminative capacity for OOD detection of baseline methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 549, + 296, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 549, + 296, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 549, + 296, + 609 + ], + "type": "text", + "content": "Condition Number (CN) Regulariser. Whilst the LSV regulariser will increase the change in energy for " + }, + { + "bbox": [ + 55, + 549, + 296, + 609 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 549, + 296, + 609 + ], + "type": "text", + "content": " in the LSV direction, there is no constraint with regard to the distribution over all other directions. For this reason, we explore a Condition Number Regulariser (CNR):" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 117, + 619, + 295, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 619, + 295, + 632 + ], + "spans": [ + { + "bbox": [ + 117, + 619, + 295, + 632 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C N} := \\mathcal {L} + \\lambda_ {C N} \\kappa \\left(\\mathrm {W} _ {\\text {c l s}}\\right), \\tag {14}", + "image_path": "0c8b3fcec226547c672963bb02c537c54fdb0568116e6335df3291f1e8d422f9.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\kappa (\\mathrm{W}_{cls})" + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": " is the condition number of the WCLS matrix, i.e., the ratio of the greatest and least singular values. Intuitively, this will encourage equalized energy changes for displacements in all directions where we note that Eq. (14) similarly assumes use of the " + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "inline_equation", + "content": "L^2" + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": "-norm as per Eq. (12). A different approach to the CNR is the orthogonal constraint" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 329, + 70, + 542, + 237 + ], + "blocks": [ + { + "bbox": [ + 329, + 70, + 542, + 237 + ], + "lines": [ + { + "bbox": [ + 329, + 70, + 542, + 237 + ], + "spans": [ + { + "bbox": [ + 329, + 70, + 542, + 237 + ], + "type": "image", + "image_path": "9ce8f5922b4e4211f3267ae349a5be11a44052d8378b0f5c650be308d1177588.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 239, + 554, + 262 + ], + "lines": [ + { + "bbox": [ + 313, + 239, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 554, + 262 + ], + "type": "text", + "content": "Figure 3. Null space reduction. We add an extra layer " + }, + { + "bbox": [ + 313, + 239, + 554, + 262 + ], + "type": "inline_equation", + "content": "g'" + }, + { + "bbox": [ + 313, + 239, + 554, + 262 + ], + "type": "text", + "content": " to create " + }, + { + "bbox": [ + 313, + 239, + 554, + 262 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 239, + 554, + 262 + ], + "type": "text", + "content": "-dimensional features, reducing the size of the null space." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 270, + 554, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 554, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 554, + 344 + ], + "type": "inline_equation", + "content": "\\| W_{cls}W_{cls}^{\\top} - I\\|" + }, + { + "bbox": [ + 313, + 270, + 554, + 344 + ], + "type": "text", + "content": ". However, this method adds the additional constraint that the singular values must be close to one, which might be a stronger prior than the CNR, hence we opt for the latter. The gradients for the singular values in Eqs. (13) and (14) are detailed by Lewis and Sendov [26] (Proposition 6.2) and further discussed in the Supp. Mat." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 351, + 438, + 364 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 351, + 438, + 364 + ], + "spans": [ + { + "bbox": [ + 314, + 351, + 438, + 364 + ], + "type": "text", + "content": "5. Experimental Details" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 372, + 554, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 554, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 554, + 408 + ], + "type": "text", + "content": "We study the impact of FEVER-OOD on free energy baseline methods (Sec. 5.1) across different standard datasets (Sec. 5.2). Implementation details are given in Sec. 5.3." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 419, + 533, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 533, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 533, + 430 + ], + "type": "text", + "content": "5.1. Baseline Methods and Evaluation Metrics" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 437, + 554, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 581 + ], + "type": "text", + "content": "Three free energy baseline methods for image OOD detection that use the free energy score are tested: VOS [7], FFS [21] and Dream-OOD [8]. These baseline methods learn a small MLP " + }, + { + "bbox": [ + 313, + 437, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\psi : \\mathbb{R} \\to (0,1)" + }, + { + "bbox": [ + 313, + 437, + 554, + 581 + ], + "type": "text", + "content": " that uses the free energy in Eq. (4) to predict an uncertainty score. Since OOD data is unavailable during training, these methods use synthetic outliers to train " + }, + { + "bbox": [ + 313, + 437, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 313, + 437, + 554, + 581 + ], + "type": "text", + "content": ". While VOS and FFS synthesize outliers in the feature space " + }, + { + "bbox": [ + 313, + 437, + 554, + 581 + ], + "type": "inline_equation", + "content": "h(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 437, + 554, + 581 + ], + "type": "text", + "content": ", Dream-OOD generates outliers in the pixel space using Stable Diffusion [34]. Additionally, VOS and FFS are also implemented as object detectors for object-level OOD detection, synthesizing outliers before the last layer of the classification head of a Faster RCNN [33]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "We follow standard evaluation metrics, reporting the false positive rate for OOD detection at " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " true positive rate of in-distribution detection (FPR95) and the area under the receiver operating characteristic curve (AUROC). We also assess the effect of our proposed techniques into the main in-distribution task (accuracy for classification and MS-COCO [28] mean average precision (mAP) for object detection). FPR95 and AUROC metrics are reported using a single intersection-over-union threshold for object detection, corresponding to the one that maximises the F-1 score, as described by Harakeh and Waslander [13]." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4533" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 85, + 555, + 224 + ], + "blocks": [ + { + "bbox": [ + 86, + 70, + 523, + 82 + ], + "lines": [ + { + "bbox": [ + 86, + 70, + 523, + 82 + ], + "spans": [ + { + "bbox": [ + 86, + 70, + 523, + 82 + ], + "type": "text", + "content": "Table 1. CIFAR-10 Results (ID Acc. = in-distribution accuracy; Null Space Reduction (NSR) methods = our approach)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 85, + 555, + 224 + ], + "lines": [ + { + "bbox": [ + 58, + 85, + 555, + 224 + ], + "spans": [ + { + "bbox": [ + 58, + 85, + 555, + 224 + ], + "type": "table", + "html": "
MethodFEVER-OODOOD DatasetsID Acc
TexturesSHVNPlaces365LSUNiSUNAvg
r-NSRRegularizerFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROC
VOS--50.05±5.7086.79±1.2839.15±10.5291.73±2.8840.88±1.7589.54±0.698.22±1.0298.36±0.2030.01±4.4494.31±0.8433.66±2.2192.15±0.5994.83±0.16
-λLSV=1.055.58±6.3586.71±2.3543.00±3.1692.09±0.9438.78±2.5191.14±0.567.36±1.5998.46±0.3221.23±6.9596.23±1.1733.19±3.2092.92±0.8594.66±0.18
-λCN=0.150.80±3.0685.39±0.7930.78±2.9594.04±0.5841.60±1.4088.77±0.748.50±0.8298.30±0.1030.91±4.5493.41±1.6832.52±1.2591.98±0.4694.69±0.16
96λLSV=1.040.86±4.8392.12±1.0541.92±16.1193.20±2.6135.69±2.2992.61±0.574.95±1.4998.93±0.2417.68±5.2697.04±0.8528.22±4.1894.78±0.8594.74±0.17
64λLSV=1.044.41±5.3991.26±1.2934.68±10.9394.39±1.7235.85±2.5692.42±0.535.68±0.5798.75±0.1324.03±5.9196.02±0.9528.93±3.8694.57±0.6694.75±0.13
32λLSV=0.00146.53±3.7789.67±1.0927.84±9.4795.29±1.0837.33±1.6391.77±0.395.65±0.7798.82±0.1124.09±6.7795.68±1.3528.29±3.3394.25±0.5594.68±0.07
10-53.20±3.9088.97±0.7335.62±9.6694.41±1.2845.73±6.2989.73±1.6911.72±3.3997.93±0.4641.80±14.3592.51±3.0337.61±4.0092.71±0.9791.95±2.32
FFS--52.86±4.4983.47±1.6738.67±11.2189.74±5.1944.65±1.2987.47±0.776.59±0.9298.67±0.1731.34±1.8893.24±0.9734.82±2.0390.52±0.9894.69±0.15
-λLSV=0.00150.74±3.9884.93±0.7532.25±12.1692.97±3.6042.99±2.3088.30±1.105.76±1.2398.82±0.2728.16±6.4594.18±1.9831.98±3.1091.84±1.2894.73±0.12
-λCN=1.050.08±3.8684.92±1.8530.91±6.1992.77±2.3345.69±2.5987.16±0.756.22±0.6598.73±0.1126.60±3.6494.80±0.6431.90±2.5191.67±0.9494.85±0.20
96λLSV=1.048.11±5.1589.84±1.5140.65±15.4793.32±3.1636.37±2.2892.29±0.605.22±1.3598.89±0.1824.84±10.0895.82±1.6831.04±5.0594.03±1.0694.71±0.08
64λLSV=1.051.70±5.6788.81±1.0933.02±14.0994.60±2.3039.05±3.7391.53±1.235.56±1.9598.84±0.3026.63±7.5195.65±1.1231.19±4.3293.89±0.6994.68±0.08
32-46.75±3.4989.61±1.4031.77±8.7394.44±1.3541.51±3.3890.72±1.024.93±1.1199.04±0.1628.92±11.2394.49±2.9430.78±3.7093.66±0.9594.79±0.18
10-54.72±5.8984.77±3.2345.70±12.9591.49±2.2741.52±3.2789.51±1.318.21±3.5998.33±0.6728.58±7.7594.30±1.8235.75±5.2691.68±1.3894.54±0.17
", + "image_path": "ad272b348d5f821114dc6e272341dc498cbe59becddba24f00ec480376c0be72.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 232, + 118, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 232, + 118, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 118, + 243 + ], + "type": "text", + "content": "5.2. Datasets" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 250, + 297, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 250, + 297, + 455 + ], + "spans": [ + { + "bbox": [ + 54, + 250, + 297, + 455 + ], + "type": "text", + "content": "We assess FEVER-OOD with standard datasets for both classification and object detection. For image-level OOD detection (i.e., image classification), we train VOS and FFS using the CIFAR-10 and CIFAR-100 [20] datasets as in-distribution while testing OOD detection in five non-overlapping datasets: Textures [3], SVHN [30], Places365 [47], LSUN [45] and iSUN [11]. We also test Dream-OOD with the CIFAR-100 dataset as in-distribution data with the same OOD datasets. Additionally, we also train Dream-OOD with the Imagenet-100 dataset, a 100 classes partition of Imagenet [35], as in-distribution while testing OOD detection in four datasets: Textures [3], Places365 [47], iNaturalist [40] and SUN [42]. We use the PASCAL VOC [10] dataset as in-distribution for object-level OOD detection using VOS and FFS and test for OOD detection on the MS-COCO dataset [28] removing images containing objects with overlapping in-distribution categories [7]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 462, + 187, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 187, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 187, + 474 + ], + "type": "text", + "content": "5.3. Implementation details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 475, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 475, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 475, + 297, + 714 + ], + "type": "text", + "content": "We follow the established architectural choices of prior works. For image-level OOD detection with VOS and FFS, we use a WideResNet-40 [46] with a 128-dimensional feature space. We choose different NSR values ranging from the feature dim to the number of classes, gradually decreasing the nullity. For CIFAR-10, we evaluate the " + }, + { + "bbox": [ + 56, + 475, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\{96,64,32,10\\}" + }, + { + "bbox": [ + 56, + 475, + 297, + 714 + ], + "type": "text", + "content": "-NSR, while for CIFAR-100 we assess " + }, + { + "bbox": [ + 56, + 475, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\{114,100\\}" + }, + { + "bbox": [ + 56, + 475, + 297, + 714 + ], + "type": "text", + "content": "-NSR. With respect to Dream-OOD models, we use a ResNet-34 [14] with 512-dimensional feature space, evaluating " + }, + { + "bbox": [ + 56, + 475, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\{256,128,100\\}" + }, + { + "bbox": [ + 56, + 475, + 297, + 714 + ], + "type": "text", + "content": "-NSR for both CIFAR-100 and Imagenet-100. Similarly, we reduce the feature space of the classification head to 768, 512 and 265 dimensions for VOS and FFS object detection experiments, having an original 1,024-dimensional object-wise feature space (as per Faster RCNN [33]). With regards to the LSVR and CNR, we test with various loss weight values in all the architectures (1, 0.1, 0.01, and 0.001), to evaluate their impact on model performance (except for Imagenet-100, where we use " + }, + { + "bbox": [ + 56, + 475, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{LSV} = 0.001" + }, + { + "bbox": [ + 56, + 475, + 297, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 475, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{CN} = 0.01" + }, + { + "bbox": [ + 56, + 475, + 297, + 714 + ], + "type": "text", + "content": " based on CIFAR-100 experiments). We train all the models using the original" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 233, + 555, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 233, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 233, + 555, + 304 + ], + "type": "text", + "content": "corresponding baseline settings. Details on the hyperparameters and the training regime for each model are given in the Supp. Mat. To obtain reliable comparisons we use deterministic learning via a fixed random seed for all experiments. We run all experiments 5 times and report mean and standard deviation for all metrics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 313, + 368, + 325 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 313, + 368, + 325 + ], + "spans": [ + { + "bbox": [ + 313, + 313, + 368, + 325 + ], + "type": "text", + "content": "6. Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 329, + 555, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 555, + 426 + ], + "type": "text", + "content": "We discuss the performance of FEVER-OOD against different free energy based OOD detection models in classification and detection. Tabs. 1 to 3 compare the use of FEVER-OOD with the corresponding baseline methods. We report different NSR as well as the best results obtained using " + }, + { + "bbox": [ + 313, + 329, + 555, + 426 + ], + "type": "inline_equation", + "content": "\\lambda_{LSV}" + }, + { + "bbox": [ + 313, + 329, + 555, + 426 + ], + "type": "text", + "content": " (Eq. (13)) or " + }, + { + "bbox": [ + 313, + 329, + 555, + 426 + ], + "type": "inline_equation", + "content": "\\lambda_{CN}" + }, + { + "bbox": [ + 313, + 329, + 555, + 426 + ], + "type": "text", + "content": " (Eq. (14)), and their corresponding value (we show both cases when no NSR is applied). Extensive ablations are available in the Supp. Mat." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 426, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 556, + 714 + ], + "type": "text", + "content": "CIFAR-10. Tab. 1 shows the results of the VOS and FFS architectures with CIFAR-10 as in-distribution. In general, NSR and LSVR improve OOD detection for both VOS and FFS models. The best performance using the VOS method is achieved using a 96-NSR with LSVR, using a " + }, + { + "bbox": [ + 313, + 426, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{LSV} = 1.0" + }, + { + "bbox": [ + 313, + 426, + 556, + 714 + ], + "type": "text", + "content": ", where the average AUROC increases to 94.78 and the average FPR95 reduces to " + }, + { + "bbox": [ + 313, + 426, + 556, + 714 + ], + "type": "inline_equation", + "content": "28.22\\%" + }, + { + "bbox": [ + 313, + 426, + 556, + 714 + ], + "type": "text", + "content": ", compared to an AUROC of 92.15 and " + }, + { + "bbox": [ + 313, + 426, + 556, + 714 + ], + "type": "inline_equation", + "content": "33.66\\%" + }, + { + "bbox": [ + 313, + 426, + 556, + 714 + ], + "type": "text", + "content": " of the baseline VOS model without affecting the in-distribution accuracy. Similarly, the best FFS setting is achieved with a 96-NSR and " + }, + { + "bbox": [ + 313, + 426, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{LSV} = 1.0" + }, + { + "bbox": [ + 313, + 426, + 556, + 714 + ], + "type": "text", + "content": ", resulting in 94.03 AUROC and " + }, + { + "bbox": [ + 313, + 426, + 556, + 714 + ], + "type": "inline_equation", + "content": "31.04\\%" + }, + { + "bbox": [ + 313, + 426, + 556, + 714 + ], + "type": "text", + "content": " FPR95, compared to the baseline FFS model, with an AUROC of 90.52 and FPR95 of " + }, + { + "bbox": [ + 313, + 426, + 556, + 714 + ], + "type": "inline_equation", + "content": "34.82\\%" + }, + { + "bbox": [ + 313, + 426, + 556, + 714 + ], + "type": "text", + "content": ", although the best FPR95 is achieved with 32-NSR and no singular value regularization, with an FPR95 of 30.78. In general, it is seen that a large NSR (i.e., 10-NSR) may decrease the AUROC and increase the FPR95. Since VOS and FFS are techniques that synthesize outliers in the feature space, a large reduction of the feature dimension might not allow for proper in-distribution understanding, impacting in the OOD training of these techniques. Additionally, it is also observed that LSVR yields better results than CNR, and a value for " + }, + { + "bbox": [ + 313, + 426, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{LSV} = 1.0" + }, + { + "bbox": [ + 313, + 426, + 556, + 714 + ], + "type": "text", + "content": " is adequate for none to moderate NSR, while smaller values for LSVR (or none) are better for large NSR. While LSVR increases the differ" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4534" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 83, + 539, + 213 + ], + "blocks": [ + { + "bbox": [ + 83, + 71, + 525, + 81 + ], + "lines": [ + { + "bbox": [ + 83, + 71, + 525, + 81 + ], + "spans": [ + { + "bbox": [ + 83, + 71, + 525, + 81 + ], + "type": "text", + "content": "Table 2. CIFAR-100 Results (ID Acc. = in-distribution accuracy; Null Space Reduction (NSR) methods = our approach)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 83, + 539, + 213 + ], + "lines": [ + { + "bbox": [ + 67, + 83, + 539, + 213 + ], + "spans": [ + { + "bbox": [ + 67, + 83, + 539, + 213 + ], + "type": "table", + "html": "
MethodFEVER-OODOOD DatasetsID Acc
r-NSRRegularizerTexturesSHVNPlaces365LSUNiSUNAvg
FPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROC
VOS--81.24±2.1176.01±0.8076.09±6.5183.18±3.5980.53±1.4676.25±0.7837.86±4.0693.21±0.6879.34±1.6076.43±1.4471.01±1.2381.02±0.4076.04±0.22
-λLSV=0.182.87±1.1176.49±0.9473.35±2.9684.57±1.5380.36±1.1676.89±0.7940.05±3.7492.83±0.4674.28±4.0380.25±3.3470.18±0.8782.21±0.5676.01±0.16
-λCN=0.0182.51±2.6476.04±1.9173.67±3.1085.31±1.2480.18±1.0876.59±0.2341.08±3.2792.75±0.6380.83±4.3773.79±5.5171.65±0.9680.90±1.1176.20±0.11
114λLSV=0.0180.02±2.7578.73±1.8383.45±2.5982.07±1.3678.41±0.9677.71±0.9328.90±3.6195.11±0.5872.07±9.0780.55±4.7068.57±2.2882.83±0.9175.72±0.16
100-80.47±2.3278.21±1.1479.10±7.4283.39±3.9179.19±1.0477.23±0.5028.99±2.5995.15±0.3975.51±6.6879.63±3.7068.65±2.2982.72±0.5775.54±0.19
FFS--82.87±1.3975.54±0.9376.32±6.1784.83±2.3681.14±1.0976.27±0.7436.27±4.9193.54±1.0382.83±2.9874.29±3.0871.89±1.6480.89±1.1176.04±0.11
-λLSV=0.0180.95±1.3476.53±0.7083.12±5.5981.61±2.9280.19±0.8476.58±0.4932.93±4.0594.20±0.5179.05±5.1677.28±3.2371.25±2.1181.24±0.7276.27±0.28
-λCN=0.00180.27±2.7676.78±1.5378.57±7.3282.57±2.9580.40±1.0976.40±0.5332.91±4.9094.06±1.0280.71±7.9074.49±5.5770.57±2.4180.86±0.8876.05±0.16
114λLSV=0.00179.26±2.6778.16±1.5074.28±7.1385.37±3.2079.45±0.8577.25±0.6124.10±1.4295.94±0.1773.61±5.8280.61±2.4966.14±1.9183.47±0.9875.45±0.37
100-77.69±2.9778.62±1.0277.84±9.6283.40±3.5280.41±0.7276.58±0.3521.91±0.9596.25±0.1179.60±6.0675.79±5.1767.49±2.0382.13±0.8575.48±0.28
Dream-OOD--62.20±1.0283.84±0.3873.05±1.9284.56±0.2177.95±1.9779.43±0.1739.90±2.0192.87±0.441.70±0.1199.58±0.0450.96±1.4488.06±0.6075.61±0.19
-λLSV=0.0158.45±2.2786.04±0.4168.75±1.8487.65±0.3077.45±2.1178.59±0.2215.45±2.5797.24±0.091.55±0.0399.63±0.4544.33±3.4889.83±0.1275.87±0.23
-λCN=0.00157.40±1.8886.28±0.3277.75±1.1185.13±0.4178.60±1.2378.73±0.1527.2±1.7795.01±0.421.55±0.0899.57±0.0748.5±1.5488.94±0.4376.32±0.14
256-60.00±3.2185.42±0.7967.50±2.4785.84±0.6675.90±1.1079.57±0.4319.85±1.0596.74±0.671.00±0.0299.78±0.0644.82±2.7289.47±0.24277.01±0.12
128λLSV=0.156.45±2.3987.73±1.0167.45±1.1287.53±0.4478.45±2.4677.24±0.6129.60±1.7394.81±0.401.95±0.1099.5±0.1046.78±2.7289.36±0.8376.21±0.35
100-57.55±2.4886.8±0.4454.45±3.0288.69±0.5975.80±1.2178.88±0.6924.45±1.5795.86±0.521.60±0.0799.65±0.2142.77±2.1089.98±0.1576.41±0.32
", + "image_path": "d0107b85f7a5a0e33af91b81e5a80311a473c9b260a0ca7d6911b533d953983e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 67, + 232, + 539, + 304 + ], + "blocks": [ + { + "bbox": [ + 79, + 220, + 530, + 231 + ], + "lines": [ + { + "bbox": [ + 79, + 220, + 530, + 231 + ], + "spans": [ + { + "bbox": [ + 79, + 220, + 530, + 231 + ], + "type": "text", + "content": "Table 3. ImageNet-100 Results (ID Acc. = in-distribution accuracy; Null Space Reduction (NSR) methods = our approach)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 232, + 539, + 304 + ], + "lines": [ + { + "bbox": [ + 67, + 232, + 539, + 304 + ], + "spans": [ + { + "bbox": [ + 67, + 232, + 539, + 304 + ], + "type": "table", + "html": "
MethodFEVER-OODOOD DatasetsID Acc
r-NSRRegularizeriNaturalistPlaces365SUNTexturesAvg
FPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROCFPR95AUROC
Dream-OOD--23.98±2.0895.94±0.2641.75±1.7292.48±0.1240.85±1.1392.76±0.0950.73±0.8586.21±0.2939.33±1.0891.84±0.1287.76±0.19
-λLSV=0.0159.91±3.0990.89±0.4866.26±2.1788.12±0.4774.26±1.5284.53±0.3360.56±1.2183.91±0.4265.25±1.5086.87±0.2587.85±0.11
-λLSV=0.00159.71±3.5390.97±0.7065.01±1.9088.39±0.4573.93±1.5884.69±0.4960.51±1.4684.01±0.5164.79±1.6987.01±0.4587.77±0.13
256-23.84±1.7395.80±0.3044.35±1.7192.14±0.3042.92±1.9892.46±0.3344.38±1.8188.60±0.5438.87±1.2992.25±0.2387.57±0.24
128λLSV=0.0123.46±2.5495.93±0.3643.37±1.4792.19±0.2242.42±1.6092.44±0.2941.86±1.0989.48±0.2737.78±1.4592.51±0.2287.60±0.11
100λLSV=0.0122.24±0.8196.16±0.1441.08±3.0392.59±0.4440.39±2.8492.91±0.4042.31±0.9789.30±0.3736.50±1.6792.74±0.2187.42±0.15
", + "image_path": "68fa8866a7e13226a215d32b057d741526ad5890366a3a42010a74173933b055.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 59, + 308, + 171, + 392 + ], + "blocks": [ + { + "bbox": [ + 59, + 308, + 171, + 392 + ], + "lines": [ + { + "bbox": [ + 59, + 308, + 171, + 392 + ], + "spans": [ + { + "bbox": [ + 59, + 308, + 171, + 392 + ], + "type": "image", + "image_path": "b5a0447239bfd21dc15f5b456962b4ed61103b23b89df1f7fdc9f6fc4365220a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 392, + 152, + 400 + ], + "lines": [ + { + "bbox": [ + 78, + 392, + 152, + 400 + ], + "spans": [ + { + "bbox": [ + 78, + 392, + 152, + 400 + ], + "type": "text", + "content": "(a) Dataset: CIFAR-10" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 108, + 401, + 242, + 411 + ], + "lines": [ + { + "bbox": [ + 108, + 401, + 242, + 411 + ], + "spans": [ + { + "bbox": [ + 108, + 401, + 242, + 411 + ], + "type": "text", + "content": "Figure 4. Singular Values Spectrum." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 176, + 308, + 290, + 392 + ], + "blocks": [ + { + "bbox": [ + 176, + 308, + 290, + 392 + ], + "lines": [ + { + "bbox": [ + 176, + 308, + 290, + 392 + ], + "spans": [ + { + "bbox": [ + 176, + 308, + 290, + 392 + ], + "type": "image", + "image_path": "c3e37d80961ce74aefb7bf09e7d5aaeea22dc52d68dccd6d83b173b223b81729.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 193, + 392, + 272, + 400 + ], + "lines": [ + { + "bbox": [ + 193, + 392, + 272, + 400 + ], + "spans": [ + { + "bbox": [ + 193, + 392, + 272, + 400 + ], + "type": "text", + "content": "(b) Dataset: CIFAR-100" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 414, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 414, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 414, + 296, + 555 + ], + "type": "text", + "content": "ence in energy of outliers in or near the least singular vector direction, the CN regulariser makes equalized free energy changes for all directions. Since the latter effect might disable directions with smaller energy changes than others, it might also reduce the free energy change in the direction of the largest singular value. Fig. 4(a) shows the singular value spectrum of the baseline VOS model against the LSVR and CNR versions for the CIFAR-10 dataset. It is seen that the while both regularizers homogenize the singular values, the CNR also reduces the most significant singular value. Finally, our experiments show that VOS achieves better results than FFS in CIFAR-10 OOD detection." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": "CIFAR-100. Tab. 2 shows the results for VOS, FFS and Dream-OOD trained on the CIFAR-100 dataset as in-distribution. The best models for VOS and FFS are achieved with 114-NSR and LSVR, with a " + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "inline_equation", + "content": "68.57\\%" + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": " FPR95 " + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "inline_equation", + "content": "(-2.44\\%)" + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": " VOS) and 82.83 " + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "inline_equation", + "content": "(+1.81\\mathrm{VOS})" + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": " AUROC for VOS-114-NSR and " + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "inline_equation", + "content": "66.14\\%" + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": " FPR95 " + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "inline_equation", + "content": "(-5.75\\%)" + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": " FFS) and 83.47 " + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "inline_equation", + "content": "(+2.58\\mathrm{FFS})" + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": " AUROC for FFS-114-NSR. However, despite these improvements, the performance of virtual outlier synthesis methods is usually poor for several in-distribution classes [44], hence the enhancement of these architectures is still challenging. Regarding Dream-OOD, it is noted that the best results are obtained by Dream-OOD-100-NSR, having an FPR95 of " + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "inline_equation", + "content": "42.77\\%" + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "inline_equation", + "content": "(-8.19\\%)" + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": " Dream-OOD) and 89.98" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 309, + 555, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 309, + 555, + 451 + ], + "spans": [ + { + "bbox": [ + 313, + 309, + 555, + 451 + ], + "type": "text", + "content": "AUROC (+1.9 Dream-OOD). Finally, for the three energy techniques (VOS, FFS and Dream-OOD), we observed that the model does not converge when using 100-NSR and the CNR (see Supp. Mat.), suggesting that the conditioning number penalization does not allow the model to learn the in-distribution task. Fig. 4(b) shows the singular value spectrum for the LSVR and CNR versions of the VOS model. Differently from the CIFAR-10 experiments, it is seen that LSVR keeps the most significant singular values similar to the baseline but homogenizes the least significant singular values, while the CNR keeps all singular values at a low level, hence explaining why it fails for this dataset." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 451, + 556, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 556, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 556, + 617 + ], + "type": "text", + "content": "Imagenet-100. We show the results of Dream-OOD with Imagenet-100 as in-distribution in Tab. 3. Similar as with CIFAR-100, the best results are obtained with NSR100, effectively eliminating the null space of the last layer (since it becomes square and is not singular). Contrary to CIFAR-100, LSVR has a negative impact when used alone. Considering that we use the same pre-training scheme as Du et al. [8], adding LSVR and CNR on an already pretrained last linear layer might affect the already formed feature space. Hence, these regularisers benefit untrained NSR modules. The best model, Dream-OOD-100-NSR with LSVR, achieves an average " + }, + { + "bbox": [ + 313, + 451, + 556, + 617 + ], + "type": "inline_equation", + "content": "36.50\\%" + }, + { + "bbox": [ + 313, + 451, + 556, + 617 + ], + "type": "text", + "content": " FPR95 and 92.74 AUROC, compared to the baseline " + }, + { + "bbox": [ + 313, + 451, + 556, + 617 + ], + "type": "inline_equation", + "content": "39.33\\%" + }, + { + "bbox": [ + 313, + 451, + 556, + 617 + ], + "type": "text", + "content": " FPR95 and 91.84 AUROC, becoming the state-of-the-art for Imagenet-100." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": "PASCAL VOC. We test FEVER-OOD techniques in object detection task, where the goal is to detect object-level OOD instances. Tab. 4 show the result for the baseline VOS and FFS, taking PASCAL VOC [10] as in-distribution dataset. For both tasks, we evaluate on two OOD datasets that contain subset of images from: MS-COCO [28] and OpenImages [22], following the original works [7, 21]. The best-performing model for VOS is VOS-768-NSR with CNR," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4535" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 70, + 165, + 140 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 165, + 140 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 165, + 140 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 165, + 140 + ], + "type": "image", + "image_path": "38dbf40693c2a1f54c600dfde7a1f99607541a80bb3b382b4e01f49999ed70f7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 165, + 72, + 253, + 139 + ], + "blocks": [ + { + "bbox": [ + 165, + 72, + 253, + 139 + ], + "lines": [ + { + "bbox": [ + 165, + 72, + 253, + 139 + ], + "spans": [ + { + "bbox": [ + 165, + 72, + 253, + 139 + ], + "type": "image", + "image_path": "faca7797e080a595d5658a989be5df461a6290e0819605a9804d8973355d6141.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 253, + 72, + 442, + 139 + ], + "blocks": [ + { + "bbox": [ + 253, + 72, + 442, + 139 + ], + "lines": [ + { + "bbox": [ + 253, + 72, + 442, + 139 + ], + "spans": [ + { + "bbox": [ + 253, + 72, + 442, + 139 + ], + "type": "image", + "image_path": "ced249d0cb05ba7cd7dae5c191dd056ace4e257e8a829eee15d6f0bf62667384.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 442, + 72, + 541, + 139 + ], + "blocks": [ + { + "bbox": [ + 442, + 72, + 541, + 139 + ], + "lines": [ + { + "bbox": [ + 442, + 72, + 541, + 139 + ], + "spans": [ + { + "bbox": [ + 442, + 72, + 541, + 139 + ], + "type": "image", + "image_path": "dc34406b1cea827db91f32e8f55f3b196923b6a6b91f5fa79bf82d32af4fa7b6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 62, + 140, + 164, + 207 + ], + "blocks": [ + { + "bbox": [ + 62, + 140, + 164, + 207 + ], + "lines": [ + { + "bbox": [ + 62, + 140, + 164, + 207 + ], + "spans": [ + { + "bbox": [ + 62, + 140, + 164, + 207 + ], + "type": "image", + "image_path": "0131f23c020b950610ac0e882aeeaf1edd8e56e0d088678eb9446f5fe4cd56f6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 212, + 553, + 234 + ], + "lines": [ + { + "bbox": [ + 55, + 212, + 553, + 234 + ], + "spans": [ + { + "bbox": [ + 55, + 212, + 553, + 234 + ], + "type": "text", + "content": "Figure 5. MS-COCO objects detected on OOD images by VOS baseline [7] (first row) and Fever-OOD (second row). Blue: OOD objects detected and mis-classified as being in-distribution. Green: the same OOD objects correctly detected as OOD by FEVER-OOD (ours)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 164, + 140, + 253, + 207 + ], + "blocks": [ + { + "bbox": [ + 164, + 140, + 253, + 207 + ], + "lines": [ + { + "bbox": [ + 164, + 140, + 253, + 207 + ], + "spans": [ + { + "bbox": [ + 164, + 140, + 253, + 207 + ], + "type": "image", + "image_path": "e2d1f8f1cedb55b6fb5d8445d54255cf5535fafaca5776d32bedee50b1fd4b21.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 253, + 140, + 442, + 207 + ], + "blocks": [ + { + "bbox": [ + 253, + 140, + 442, + 207 + ], + "lines": [ + { + "bbox": [ + 253, + 140, + 442, + 207 + ], + "spans": [ + { + "bbox": [ + 253, + 140, + 442, + 207 + ], + "type": "image", + "image_path": "fde8a35f769bbd6f5f32b25bc9dda203d2d11d2aa027c8ec6d8b373e4c0c396a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 442, + 140, + 544, + 208 + ], + "blocks": [ + { + "bbox": [ + 442, + 140, + 544, + 208 + ], + "lines": [ + { + "bbox": [ + 442, + 140, + 544, + 208 + ], + "spans": [ + { + "bbox": [ + 442, + 140, + 544, + 208 + ], + "type": "image", + "image_path": "12c8724f5737873132ceb2e02d3f4dab10f687b8d2e26550db93809d26955762.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 58, + 267, + 294, + 483 + ], + "blocks": [ + { + "bbox": [ + 55, + 239, + 272, + 262 + ], + "lines": [ + { + "bbox": [ + 55, + 239, + 272, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 272, + 262 + ], + "type": "text", + "content": "Table 4. VOS results with PASCAL VOC as in-distribution (Null Space Reduction (NSR) methods = our approach)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 267, + 294, + 483 + ], + "lines": [ + { + "bbox": [ + 58, + 267, + 294, + 483 + ], + "spans": [ + { + "bbox": [ + 58, + 267, + 294, + 483 + ], + "type": "table", + "html": "
MethodFEVER-OODOOD DatasetsmAP
r-NSRλLSVλCNMS-COCOOpenImages
FPR95AUROCFPR95AUROC
VOS---50.2987.7753.0986.580.489
-0.001-49.9688.1052.2685.800.489
--153.4286.6652.7385.770.490
768--48.9487.9452.3486.330.489
7680.01-50.0887.7851.7186.040.493
768-0.0147.4788.1549.3686.320.491
512--49.8488.4255.3385.330.486
5120.001-49.2688.3954.3285.820.489
512-0.0149.5987.4854.8185.480.490
256--47.8888.4952.4186.390.486
2560.01-48.5188.3650.5286.120.491
256-0.00149.188.2751.4786.080.489
FFS---50.7787.1853.7885.340.487
-0.01-53.0986.7553.4785.860.490
--0.0150.3687.6551.1985.980.485
768--51.5087.5953.3385.840.488
7680.01-47.4288.3652.4686.400.489
768-0.0151.1286.8255.4484.840.490
512--46.9388.8253.8986.280.491
5120.01-47.4288.3652.4686.400.488
512-0.0151.4287.0554.9884.550.490
256--51.1888.7453.6786.560.487
2560.01-47.9388.0447.9384.950.489
256-0.0152.9486.6252.8884.450.489
", + "image_path": "99ec6a99027725d4b081080c4d77dad883791b868892b91ecf0d7d6f85eb605c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 54, + 491, + 295, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 491, + 295, + 660 + ], + "spans": [ + { + "bbox": [ + 54, + 491, + 295, + 660 + ], + "type": "text", + "content": "achieving average FPR95 scores of " + }, + { + "bbox": [ + 54, + 491, + 295, + 660 + ], + "type": "inline_equation", + "content": "47.47\\%" + }, + { + "bbox": [ + 54, + 491, + 295, + 660 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 491, + 295, + 660 + ], + "type": "inline_equation", + "content": "49.36\\%" + }, + { + "bbox": [ + 54, + 491, + 295, + 660 + ], + "type": "text", + "content": " on MS-COCO and OpenImages, respectively, an improvement of nearly " + }, + { + "bbox": [ + 54, + 491, + 295, + 660 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 54, + 491, + 295, + 660 + ], + "type": "text", + "content": " over the VOS baseline. For FFS, we achieve a minimum FPR95 of " + }, + { + "bbox": [ + 54, + 491, + 295, + 660 + ], + "type": "inline_equation", + "content": "46.93\\%" + }, + { + "bbox": [ + 54, + 491, + 295, + 660 + ], + "type": "text", + "content": " with FFS-512-NSR model and " + }, + { + "bbox": [ + 54, + 491, + 295, + 660 + ], + "type": "inline_equation", + "content": "47.93\\%" + }, + { + "bbox": [ + 54, + 491, + 295, + 660 + ], + "type": "text", + "content": " for the FFS-256-NSR model with the LSVR for both OOD datasets. Additionally, our approach generally delivers higher AUROC scores compared to the FFS baseline across both OOD datasets. Qualitative results are illustrated in Fig. 5, where we visualize the prediction on MS-COCO OOD images, using the baseline VOS (top) and the best Fever-OOD model, i.e., VOS-768-NSR, with the PASCAL-VOC as the in-distribution dataset, showing that out framework outperforms baseline VOS in in detecting outlier objects with a significantly lower false positive rate." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 54, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 665, + 296, + 714 + ], + "type": "text", + "content": "Computational overhead. The computational overhead introduced by our method is minimal since we only add an extra linear layer. For instance, the ResNet-34 used in Dream-OOD takes " + }, + { + "bbox": [ + 54, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "4.42 \\pm 1.15" + }, + { + "bbox": [ + 54, + 665, + 296, + 714 + ], + "type": "text", + "content": " ms/img and has 21.5M" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 241, + 554, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 241, + 554, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 241, + 554, + 289 + ], + "type": "text", + "content": "params. Adding 256, 128, 100-NSR layers change the inference time to " + }, + { + "bbox": [ + 313, + 241, + 554, + 289 + ], + "type": "inline_equation", + "content": "3.91 \\pm 0.62" + }, + { + "bbox": [ + 313, + 241, + 554, + 289 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 241, + 554, + 289 + ], + "type": "inline_equation", + "content": "4.44 \\pm 1.17" + }, + { + "bbox": [ + 313, + 241, + 554, + 289 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 241, + 554, + 289 + ], + "type": "inline_equation", + "content": "4.13 \\pm 1.02" + }, + { + "bbox": [ + 313, + 241, + 554, + 289 + ], + "type": "text", + "content": " ms/img, with an extra 105k, 27k, 10k params, which is negligible compared with the total size of ResNet-34." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 293, + 388, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 388, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 388, + 304 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 307, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 555, + 714 + ], + "type": "text", + "content": "In this work, we present FEVER-OOD, a robust approach for addressing critical vulnerabilities found in existing energy-based OOD detection. Our method targets the Null Space Vulnerabilities (NSV), where in-distribution and OOD instances can produce similar free energy scores despite differing feature representations, and Least Singular Value Vulnerabilities (LSVV), where energy score similarities are influenced by the least singular vector of the classifier last linear layer. These vulnerabilities are especially problematic in high-dimensional feature spaces where the feature dimension exceeds the number of classes, impacting the effectiveness of OOD detection. To mitigate these issues, FEVER-OOD incorporates strategies such as reducing the null space through lower-dimensional feature spaces and introducing a Least Singular Value Regularizer to maximize the least singular value, enhancing the distinction between energy scores of different samples. Additionally, we propose a Condition Number Regularizer to promote uniform energy changes for feature-space displacements, fostering consistent energy variation and improving detection robustness. Our comprehensive evaluation on benchmark datasets, including applications to the Dream-OOD framework with ImageNet-100 as the ID dataset, demonstrates the significant effectiveness of FEVER-OOD. Specifically, our approach achieves a " + }, + { + "bbox": [ + 313, + 307, + 555, + 714 + ], + "type": "inline_equation", + "content": "10.13\\%" + }, + { + "bbox": [ + 313, + 307, + 555, + 714 + ], + "type": "text", + "content": " reduction in the average false positive rate and a " + }, + { + "bbox": [ + 313, + 307, + 555, + 714 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 313, + 307, + 555, + 714 + ], + "type": "text", + "content": " increase in AUROC, setting a new state-of-the-art in OOD detection. Furthermore, our results highlight the generalization capability of FEVER-OOD across different detection architectures, showcasing its potential beyond conventional classification settings. Future extensions of FEVER-OOD could include applications to varied domains and integration with other OOD detection frameworks to further validate its versatility and performance in diverse machine learning contexts. This work provides valuable insights into enhancing OOD robustness and sets a foundation for more resilient free energy-based detection methods." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4536" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 83, + 115, + 95 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 83, + 115, + 95 + ], + "spans": [ + { + "bbox": [ + 56, + 83, + 115, + 95 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 102, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 102, + 295, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 102, + 295, + 156 + ], + "spans": [ + { + "bbox": [ + 61, + 102, + 295, + 156 + ], + "type": "text", + "content": "[1] Haoyue Bai, Gregory Canal, Xuefeng Du, Jeongyeol Kwon, Robert D Nowak, and Yixuan Li. Feed two birds with one scone: Exploiting wild data for both out-of-distribution generalization and detection. In International Conference on Machine Learning, pages 1454-1471. PMLR, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 159, + 295, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 159, + 295, + 190 + ], + "spans": [ + { + "bbox": [ + 62, + 159, + 295, + 190 + ], + "type": "text", + "content": "[2] Haoyue Bai, Jifan Zhang, and Robert Nowak. Aha: Human-assisted out-of-distribution generalization and detection. arXiv preprint arXiv:2410.08000, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 192, + 294, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 192, + 294, + 245 + ], + "spans": [ + { + "bbox": [ + 62, + 192, + 294, + 245 + ], + "type": "text", + "content": "[3] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3606-3613, 2014. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 247, + 295, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 295, + 279 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 295, + 279 + ], + "type": "text", + "content": "[4] Matthew Cook, Alina Zare, and Paul Gader. Outlier detection through null space analysis of neural networks. ICML Workshops, 2020. 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 281, + 295, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 281, + 295, + 356 + ], + "spans": [ + { + "bbox": [ + 62, + 281, + 295, + 356 + ], + "type": "text", + "content": "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 358, + 295, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 358, + 295, + 412 + ], + "spans": [ + { + "bbox": [ + 62, + 358, + 295, + 412 + ], + "type": "text", + "content": "[6] Xuefeng Du, Xin Wang, Gabriel Gozum, and Yixuan Li. Unknown-aware object detection: Learning what you don't know from videos in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13678-13688, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 414, + 295, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 414, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 414, + 295, + 456 + ], + "type": "text", + "content": "[7] Xuefeng Du, Zhaoning Wang, Mu Cai, and Yixuan Li. Vos: Learning what you don't know by virtual outlier synthesis. In International Conference on Learning Representations, 2022. 1, 2, 3, 4, 5, 6, 7, 8, 9, 10" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 459, + 295, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 459, + 295, + 500 + ], + "spans": [ + { + "bbox": [ + 62, + 459, + 295, + 500 + ], + "type": "text", + "content": "[8] Xuefeng Du, Yiyou Sun, Xiaojin Zhu, and Yixuan Li. Dream the impossible: Outlier imagination with diffusion models. In Advances in Neural Information Processing Systems, 2023. 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 502, + 295, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 295, + 535 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 295, + 535 + ], + "type": "text", + "content": "[9] Xuefeng Du, Zhen Fang, Ilias Diakonikolas, and Yixuan Li. How does unlabeled data provably help out-of-distribution detection? arXiv preprint arXiv:2402.03502, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 536, + 295, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 536, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 57, + 536, + 295, + 579 + ], + "type": "text", + "content": "[10] Mark Everingham, Luc Van Gool, Christopher K. I. Williams, John M. Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International Journal of Computer Vision, 88(2):303-338, 2010. 6, 7, 2, 9, 10" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 581, + 295, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 581, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 57, + 581, + 295, + 634 + ], + "type": "text", + "content": "[11] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 580-587, 2014. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 636, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 668 + ], + "type": "text", + "content": "[12] Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep Learning. MIT Press, 2016. http://www.deeplearningbook.org.3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "text", + "content": "[13] Ali Harakeh and Steven L. Waslander. Estimating and evaluating regression predictive uncertainty in deep object detectors. In International Conference on Learning Representations, 2021. 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "text", + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 3, 6, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 119, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 161 + ], + "type": "text", + "content": "[15] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In International Conference on Learning Representations, 2022. 2, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 163, + 555, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 555, + 217 + ], + "type": "text", + "content": "[16] Yen-Chang Hsu, Yilin Shen, Hongxia Jin, and Zsolt Kira. Generalized odin: Detecting out-of-distribution image without learning from out-of-distribution data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10951-10960, 2020. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "type": "text", + "content": "[17] Rui Huang, Andrew Geng, and Yixuan Li. On the importance of gradients for detecting distributional shifts in the wild. Advances in Neural Information Processing Systems, 34:677-689, 2021. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "text", + "content": "[18] B.K.S. Isaac-Medina, Y.F.A. Gaus, N. Bhowmik, and T.P. Breckon. Towards open-world object-based anomaly detection via self-supervised outlier synthesis. In Proc. European Conference on Computer Vision. Springer, 2024. 1, 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 309, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 352 + ], + "type": "text", + "content": "[19] Julian Katz-Samuels, Julia B Nakhleh, Robert Nowak, and Yixuan Li. Training ood detectors in their natural habitats. In International Conference on Machine Learning, pages 10848-10865. PMLR, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 354, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 386 + ], + "type": "text", + "content": "[20] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009. 6, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 388, + 555, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 388, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 317, + 388, + 555, + 453 + ], + "type": "text", + "content": "[21] Nishant Kumar, Siniša Šegvic, Abouzar Eslami, and Stefan Gumhold. Normalizing flow based feature synthesis for outlier-aware object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5156-5165, 2023. 1, 2, 3, 4, 5, 7, 6, 9, 10" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 456, + 553, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 553, + 531 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 553, + 531 + ], + "type": "text", + "content": "[22] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision, 128(7):1956-1981, 2020. 7, 6, 9" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 533, + 553, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 533, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 317, + 533, + 553, + 577 + ], + "type": "text", + "content": "[23] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. Advances in Neural Information Processing Systems, 30, 2017. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 578, + 553, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 578, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 317, + 578, + 553, + 611 + ], + "type": "text", + "content": "[24] Hao Lang, Yinhe Zheng, Yixuan Li, Jian SUN, Fei Huang, and Yongbin Li. A survey on out-of-distribution detection in NLP. Transactions on Machine Learning Research, 2024. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 613, + 553, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 553, + 655 + ], + "type": "text", + "content": "[25] Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. Advances in Neural Information Processing Systems, 31, 2018. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "type": "text", + "content": "[26] Adrian S Lewis and Hristo S Sendov. Nonsmooth analysis of singular values. part i: Theory. Set-Valued Analysis, 13: 213–241, 2005. 5, 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[27] Shiyu Liang, Yixuan Li, and R Srikant. Enhancing the reliability of out-of-distribution image detection in neural net" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4537" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "works. In 6th International Conference on Learning Representations, 2018, 2018. 1, 2, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "type": "text", + "content": "[28] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5, 6, 7, 10" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 294, + 205 + ], + "type": "text", + "content": "[29] Weitang Liu, Xiaoyun Wang, John Owens, and Yixuan Li. Energy-based out-of-distribution detection. Advances in Neural Information Processing Systems, 33:21464-21475, 2020. 1, 2, 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 294, + 259 + ], + "type": "text", + "content": "[30] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading digits in natural images with unsupervised feature learning. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning 2011, 2011. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 294, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 315 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 315 + ], + "type": "text", + "content": "[31] Anh Nguyen, Jason Yosinski, and Jeff Clune. Deep neural networks are easily fooled: High confidence predictions for unrecognizable images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 427-436, 2015. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 316, + 294, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 360 + ], + "type": "text", + "content": "[32] Bo Peng, Yadan Luo, Yonggang Zhang, Yixuan Li, and Zhen Fang. Conjnorm: Tractable density estimation for out-of-distribution detection. In The Twelfth International Conference on Learning Representations, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 361, + 294, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 294, + 425 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 294, + 425 + ], + "type": "text", + "content": "[33] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: towards real-time object detection with region proposal networks. In Proceedings of the 28th International Conference on Neural Information Processing Systems - Volume 1, page 91-99, Cambridge, MA, USA, 2015. MIT Press. 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 426, + 295, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 295, + 482 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 295, + 482 + ], + "type": "text", + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 3, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 483, + 294, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 483, + 294, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 483, + 294, + 547 + ], + "type": "text", + "content": "[35] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 115 (3):211-252, 2015. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 548, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 613 + ], + "type": "text", + "content": "[36] Mohammadreza Salehi, Hossein Mirzaei, Dan Hendrycks, Yixuan Li, Mohammad Hossein Rohban, and Mohammad Sabokrou. A unified survey on anomaly, novelty, open-set, and out of-distribution detection: Solutions and future challenges. Transactions on Machine Learning Research, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 614, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 669 + ], + "type": "text", + "content": "[37] Carl-Johann Simon-Gabriel, Yann Ollivier, Leon Bottou, Bernhard Schölkopf, and David Lopez-Paz. First-order adversarial vulnerability of neural networks and input dimension. In Proceedings of the 36th International Conference on Machine Learning, pages 5809-5817. PMLR, 2019. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[38] Yiyou Sun, Yifei Ming, Xiaojin Zhu, and Yixuan Li. Out-of-distribution detection with deep nearest neighbors. In International Conference on Machine Learning, pages 20827-20840. PMLR, 2022. 8" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 487 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 106 + ], + "type": "text", + "content": "[39] Leitian Tao, Xuefeng Du, Jerry Zhu, and Yixuan Li. Nonparametric outlier synthesis. In The Eleventh International Conference on Learning Representations, 2023. 3, 4, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 107, + 554, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 554, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 554, + 172 + ], + "type": "text", + "content": "[40] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8769-8778, 2018. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 174, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 553, + 207 + ], + "type": "text", + "content": "[41] Haoqi Wang, Zhizhong Li, Litong Feng, and Wayne Zhang. Vim: Out-of-distribution with virtual-logit matching. In CVPR, pages 4921-4930, 2022. 2, 4, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "type": "text", + "content": "[42] Jianxiong Xiao, James Hays, Krista A. Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 3485-3492, 2010. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 263, + 553, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 263, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 316, + 263, + 553, + 317 + ], + "type": "text", + "content": "[43] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10687-10698, 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 319, + 553, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 361 + ], + "type": "text", + "content": "[44] Jingkang Yang, Kaiyang Zhou, Yixuan Li, and Ziwei Liu. Generalized out-of-distribution detection: A survey. International Journal of Computer Vision, pages 1-28, 2024. 1, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 364, + 553, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 364, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 316, + 364, + 553, + 407 + ], + "type": "text", + "content": "[45] Fisher Yu, Yinda Zhang, Shuran Song, Ari Seff, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 409, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 409, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 409, + 553, + 453 + ], + "type": "text", + "content": "[46] Erwan Zerhouni, David Lanyi, Matheus Viana, and Maria Gabrani. Wide residual networks for mitosis detection. In 2017 IEEE 14th International Symposium on Biomedical Imaging (ISBI 2017), pages 924-928. IEEE, 2017. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 454, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 454, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 454, + 553, + 487 + ], + "type": "text", + "content": "[47] Bolei Zhou, Agata Lapedriza, Antonio Torralba, and Aude Oliva. Places: An image database for deep scene understanding. Journal of Vision, 17(10):296-296, 2017. 6" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4538" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/FG-OrIU_ Towards Better Forgetting via Feature-Gradient Orthogonality for Incremental Unlearning/4a750062-438a-4c8c-b4db-9a6576c3de7a_content_list.json b/2025/FG-OrIU_ Towards Better Forgetting via Feature-Gradient Orthogonality for Incremental Unlearning/4a750062-438a-4c8c-b4db-9a6576c3de7a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c705d053ca952d8d85006ad77062765794a9c80e --- /dev/null +++ b/2025/FG-OrIU_ Towards Better Forgetting via Feature-Gradient Orthogonality for Incremental Unlearning/4a750062-438a-4c8c-b4db-9a6576c3de7a_content_list.json @@ -0,0 +1,1656 @@ +[ + { + "type": "text", + "text": "FG-OrIU: Towards Better Forgetting via Feature-Gradient Orthogonality for Incremental Unlearning", + "text_level": 1, + "bbox": [ + 109, + 128, + 888, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qian Feng $^{1}$ , JiaHang Tu $^{1}$ , Mintong Kang $^{2}$ , Hanbin Zhao $^{1*}$ , Chao Zhang $^{1}$ , Hui Qian $^{1}$", + "bbox": [ + 148, + 202, + 846, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1College of Computer Science and Technology, Zhejiang University \n2Department of Computer Science University of Illinois at Urbana-Champaign", + "bbox": [ + 184, + 222, + 805, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Incremental unlearning (IU) is critical for pre-trained models to comply with sequential data deletion requests, yet existing methods primarily suppress parameters or confuse knowledge without explicit constraints on both feature and gradient level, resulting in superficial forgetting where residual information remains recoverable. This incomplete forgetting risks security breaches and disrupts retention balance, especially in IU scenarios. We propose FG-OrIU (Feature-Gradient Orthogonality for Incremental Unlearning), the first framework unifying orthogonal constraints on both features and gradients level to achieve deep forgetting, where the forgetting effect is irreversible. FG-OrIU decomposes feature spaces via Singular Value Decomposition (SVD), separating forgetting and remaining class features into distinct subspaces. It then enforces dual constraints: feature orthogonal projection on both forgetting and remaining classes, while gradient orthogonal projection prevents the reintroduction of forgotten knowledge and disruption to remaining classes during updates. Additionally, dynamic subspace adaptation merges newly forgetting subspaces and contracts remaining subspaces, ensuring a stable balance between removal and retention across sequential unlearning tasks. Extensive experiments demonstrate the effectiveness of our method.", + "bbox": [ + 88, + 323, + 485, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 715, + 220, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, pre-training with fine-tuning [22, 51, 57] has dominated model development. However, large-scale datasets often contain harmful or private content, necessitating mechanisms to erase specific data traces. Machine Unlearning (MU) [4, 5, 7, 12, 16, 20, 24, 34, 41, 42, 50] addresses this by enabling pre-trained models (PTMs) to comply with deletion mandates like GDPR (General data protection regulation)'s \"right to be forgotten\" [36, 37]. While existing MU methods focus on static unlearning, for", + "bbox": [ + 89, + 739, + 483, + 877 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e8d792e49fac155c7c124f2cda35afc426466289ac4a664bdba1ab3bf5de5de0.jpg", + "image_caption": [ + "Figure 1. Here, we present three examples from forgetting classes, with the reconstructed images generated using Deep Image Prior (DIP) [46], where the last block's features from the unlearned model serve as the training target." + ], + "image_footnote": [], + "bbox": [ + 514, + 290, + 905, + 450 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "getting a single or small set of classes in one task, real-world scenarios require Incremental Unlearning (IU) [8-10, 17, 19, 25, 28, 31, 35, 48, 60, 61]. Regulatory shifts (e.g., evolving OpenAI safety policies $^{1}$ , EU AI Act $^{2}$ ) and dynamic content restrictions necessitate sequential unlearning of multiple classes/concepts without costly retraining. IU need not only erase targets in a single step but also preserve retained performance and ensure irreversible forgetting across tasks, making it critical for sustainable model governance in evolving ethical and legal landscapes.", + "bbox": [ + 511, + 512, + 906, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "IU methods broadly address two scenarios: (1) concept erasure in generative models, where sensitive concepts (e.g., nudity, violence) are iteratively removed from diffusion models [11, 45], and (2) instant- or class-level unlearning in vision classifiers, where PTMs forget specific classes (e.g., face identity, medical data) while retaining unrelated ones [26, 34]. In this work, we focus on the latter, specifically targeting class-level IU in vision PTMs. Our goal is threefold: (a) to ensure that the model completely forgets the specified classes, (b) to robustly preserve the performance of remaining classes, and (c) to maintain the effectiveness of previous forgetting across sequential unlearning procedures.", + "bbox": [ + 511, + 664, + 908, + 861 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author: Hanbin Zhao .", + "bbox": [ + 107, + 886, + 464, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1 https://openai.com/safety/", + "bbox": [ + 531, + 875, + 679, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{2}$ https://artificialintelligenceact.eu/the-act/", + "bbox": [ + 531, + 887, + 756, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1957", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing methods largely fall into two categories: parameter suppression, which selectively weakens target class-related parameters [15, 16, 52], and knowledge confusion, which distorts predictions via noisy labels or bad teacher distillation [12, 18, 29]. However, our empirical findings reveal that these methods lead to superficial forgetting, as the unlearned model still retains semantic information related to the forgetting classes, making the erased knowledge partially residual and recoverable. Specifically, reconstructing features from the last layer of the unlearned model using Deep Image Prior (DIP) [46] produces blurred yet recognizable images (as illustrated in Figure 1). Furthermore, fixing the backbone of the unlearned model and retraining only the classification head [10, 61], which recovers a portion of the original accuracy (as illustrated in Figure 3). The blurred image and the partial recovery of performance both indicate that existing methods merely degrade the features of forgetting classes rather than completely eliminating them. The degraded features still contain residual class-specific information, which leads to potential recovery of the forgetting effect. Besides, these degraded features not only enable recovery but also risk entanglement with the remaining class features [34]. Such entanglement destabilizes decision boundaries, degrading performance on the remaining classes, particularly in IU scenarios where sequential deletion requests amplify error propagation [17, 48]. To address these limitations and achieve deep forgetting, we argue that unlearning should explicitly constrain both feature representations and gradient updates. Since model updates rely on both forward computations (feature extraction) and backward optimization (gradient updates), it is critical to simultaneously suppress the features of forgetting classes and prevent their reintroduction during training updates. In Table 1, we summarize the differences between our methods and prior works.", + "bbox": [ + 91, + 90, + 480, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To enforce explicit constraints on both feature representations and gradient updates, we draw inspiration from orthogonal mechanisms in related domains. For feature-level constraints, recent work in text-to-image models [45, 55] demonstrates that orthogonalizing sensitive token embeddings effectively removes targeted concepts from diffusion outputs, suggesting geometric isolation as a viable strategy for feature suppression. For gradient-level constraints, inspired by [56], which reduces the mutual influence of different tasks in multi-task learning, and [40, 62], which mitigates the forgetting of old tasks in continual learning, we leverage the insight that updating in directions orthogonal to old/other task subspaces helps alleviate catastrophic forgetting. While these orthogonal principles show promise in adjacent fields like generative model editing and multitask/continual learning, their applicability to MU, particularly IU in vision PTMs, remains underexplored.", + "bbox": [ + 93, + 625, + 480, + 878 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose FG-OrIU (Feature-Gradient", + "bbox": [ + 109, + 883, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/f61b2e852e03ed485ce2a9396f3fc5ab8e7a540f3c48420f0e9de3360d3d7dad.jpg", + "table_caption": [ + "Table 1. Position, Complete, and Recover refers to, where the forgetting occurs, whether the forgetting is thorough, and whether the forgetting effect can be restored, respectively." + ], + "table_footnote": [], + "table_body": "
PositionCompleteRecover
superficial forgetting
Prior worksOnly Final OutputNoPartial
deep forgetting
OursFeature&GradientYesNo
", + "bbox": [ + 517, + 133, + 901, + 209 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Orthogonality for Incremental Unlearning), a unified framework for IU in vision PTMs that enforces explicit constraints at both feature and gradient levels through three stages: feature subspace decomposition, dual orthogonal projection, and dynamic subspace adaptation. In feature subspace decomposition, we decompose the feature space of each layer into two subspaces via Singular Value Decomposition (SVD): the forgetting subspace $S_f$ , which captures the discriminative features of the classes to be removed, and the remaining subspace $S_r$ , which encodes the features of the remaining classes. In dual orthogonal projection, we freeze the backbone of PTM and introduce lightweight LoRA modules for efficient updates. Then, features of samples from forgetting classes are orthogonalized against $S_f$ to eliminate correlations, while remaining classes features are aligned with $S_r$ to minimize interference through minimizing the projection residual. Before backward optimization, gradients are first projected onto $S_f$ to maximally perturb parameters critical for forgetting classes, then modified along the orthogonal complement of $S_r$ to prevent contamination of retained features. In dynamic subspace adaptation, FOrIU dynamically expands $S_f$ by merging new forgetting class subspaces and contracts $S_r$ by removing overlapping features, ensuring cumulative forgetting without degrading retained performance.", + "bbox": [ + 511, + 214, + 903, + 592 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are threefold:", + "text_level": 1, + "bbox": [ + 532, + 593, + 800, + 606 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We reveal that existing IU methods achieve superficial forgetting rather than deep forgetting, as their partial feature suppression leaves residual information recoverable.", + "- We propose the first framework unifying forgetting on both feature and gradient level via dual orthogonal projection, ensuring irreversible forgetting while preserving remaining class robustness.", + "- Extensive experiments demonstrate superior performance in MU and IU scenarios across multiple benchmarks." + ], + "bbox": [ + 513, + 609, + 903, + 746 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 753, + 653, + 768 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Machine Unlearning", + "text_level": 1, + "bbox": [ + 511, + 772, + 707, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Machine unlearning (MU) aims to remove the influence of specific data from pre-trained models (PTMs) without retraining from scratch [49, 58]. Existing MU methods are broadly categorized into exact and approximate unlearning. Exact MU [5, 32] ensures statistical indistinguishability from models retrained on remaining data but is computationally prohibitive for large PTMs. Approximate MU re", + "bbox": [ + 511, + 794, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1958", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "taxes this guarantee and is more practical. Current methods fall into two paradigms: (1) concept erasure in generative models [11, 45]; and (2) class-/sample-wise unlearning in classification models, which modifies pre-trained weights to forget target classes while preserving others [26, 34]. However, most studies focus on static unlearning, assuming a single-step forgetting task. Real-world applications, such as compliance with iterative AI safety policies (e.g., OpenAI's content guidelines) or evolving regulations like the EU AI Act, require incremental unlearning (IU) to process sequential deletion requests efficiently. Directly applying static methods to IU often leads to cumulative errors, model collapse, and incomplete forgetting [17, 48], underscoring the need for dedicated frameworks to ensure compliance in dynamic regulatory environments.", + "bbox": [ + 89, + 90, + 483, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Incremental Unlearning", + "text_level": 1, + "bbox": [ + 89, + 319, + 313, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing MU methods, like parameter suppression [15, 16, 52] and knowledge confusion methods [10, 12, 17, 18, 29, 35, 61], lead to superficial forgetting due to the absence of direct constraints on the feature representation. In particular, the degraded features trap them in a dilemma between forgetting and remaining, which becomes even more severe when directly applied to IU. Meanwhile, some continual learning methods [6, 27], originally designed to retain old knowledge while dynamically learning new knowledge, also face inherent limitations in IU [35]. Previous work focusing on IU has primarily relied on toy models, e.g., linear models [13], or imposed heavy constraints, such as adding a class-specific synthetic signal in the pre-training stage [43], which is not feasible for PTMs. Moreover, while some focus on dynamically performing both continual learning and selective forgetting within learned knowledge [8, 28, 31], our approach differs in that we emphasize achieving deep forgetting when handling sequential unlearning tasks. That is, rather than alternating between learning and forgetting, the pre-trained model should completely forget unwanted knowledge while efficiently retaining the rest.", + "bbox": [ + 89, + 339, + 483, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Problem Statement", + "text_level": 1, + "bbox": [ + 89, + 659, + 277, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following [60, 61], we first formally define Machine Unlearning (MU), i.e., static unlearning scenario. We then extend this definition to Incremental Unlearning (IU), i.e., dynamic unlearning scenario, where deletion requests arrive sequentially.", + "bbox": [ + 89, + 680, + 482, + 755 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\mathcal{M}$ be a pre-trained model trained on the dataset $\\mathbf{D}$ . We denote the mapping relationship of the model as $f_{\\mathcal{M}}(\\mathcal{X}_{\\mathbf{D}})\\rightarrow \\mathcal{Y}_{\\mathbf{D}}$ , where $\\mathcal{X}_{\\mathbf{D}}$ and $\\mathcal{Y}_{\\mathbf{D}}$ represent the input set and output set, respectively. For MU, we partition $\\mathbf{D}$ into $\\mathbf{D}_{\\mathbf{f}}$ (forgetting classes) and $\\mathbf{D}_{\\mathbf{r}}$ (remaining classes), under the constraint $|\\mathbf{D}_{\\mathbf{r}}| + |\\mathbf{D}_{\\mathbf{f}}|\\ll |\\mathbf{D}|$ . The forgetting algorithm $\\mathcal{F}$ produces $\\mathcal{M}' = \\mathcal{F}(\\mathcal{M},\\mathbf{D}_{\\mathbf{f}},\\mathbf{D}_{\\mathbf{r}})$ , altering the following mappings:", + "bbox": [ + 89, + 756, + 483, + 878 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\mathcal {M} ^ {\\prime}} \\left(\\mathcal {X} _ {\\mathbf {D} _ {\\mathrm {f}}}\\right) \\not \\longrightarrow \\mathcal {Y} _ {\\mathbf {D} _ {\\mathrm {f}}}, f _ {\\mathcal {M} ^ {\\prime}} \\left(\\mathcal {X} _ {\\mathbf {D} _ {\\mathrm {r}}}\\right) \\longrightarrow \\mathcal {Y} _ {\\mathbf {D} _ {\\mathrm {r}}}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 883, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $\\nrightarrow$ means the mapping relationship no longer holds.", + "bbox": [ + 511, + 90, + 903, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For IU, let there are a sequence of unlearning task, $\\{\\mathcal{T}_1,\\mathcal{T}_2,\\dots ,\\mathcal{T}_T\\}$ and $T$ is the number of unlearning tasks. And $\\{\\mathbf{D}_{f,t},\\mathbf{D}_{r,t}\\}$ is the forgetting classes and remaining classes belonging to $\\mathcal{T}_t$ where $\\mathbf{D}_{r,t} = \\mathbf{D} - \\cup_{i = 1}^{t}\\mathbf{D}_{f,i}$ . Starting from $\\mathcal{M}_0 = \\mathcal{M}$ , the forgetting algorithm $\\mathcal{F}$ iteratively processes till $t$ -th unlearning task $\\mathcal{T}_t$ to generate $\\mathcal{M}_t = \\mathcal{F}(\\mathcal{M}_{t - 1},\\mathbf{D}_{f,t},\\mathbf{D}_{r,t})$ which holds the followings:", + "bbox": [ + 511, + 107, + 903, + 213 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\mathcal {M} _ {t}} \\left(\\mathcal {X} _ {\\mathbf {D} _ {f, i}}\\right) \\not \\longrightarrow \\mathcal {Y} _ {\\mathbf {D} _ {f, i}}, f _ {\\mathcal {M} _ {t}} \\left(\\mathcal {X} _ {\\mathbf {D} _ {r, j}}\\right) \\longrightarrow \\mathcal {Y} _ {\\mathbf {D} _ {r, j}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 224, + 903, + 243 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Methods", + "text_level": 1, + "bbox": [ + 511, + 258, + 612, + 273 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce FG-OrIU, a comprehensive framework for Incremental Unlearning (IU) in pre-trained models (PTM). By imposing orthogonal constraints on both feature representations and gradient updates, FG-OrIU ensures thorough removal of unwanted knowledge while preserving performance on remaining classes. An overview of the framework is shown in Figure 2.", + "bbox": [ + 511, + 277, + 906, + 382 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Feature Subspace Decomposition", + "text_level": 1, + "bbox": [ + 511, + 386, + 803, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "PTMs acquire implicit knowledge from pre-training data, benefiting downstream tasks during inference. Unlearning aims to selectively remove the influence of forgetting classes from the PTM [10]. Unlike explicit knowledge in traditional Knowledge Bases (KBs) [38, 44], which can be directly edited, the knowledge in PTMs is implicit and intertwined across all parameters [10, 34, 64]. Effective unlearning requires precisely identifying feature spaces for forgetting and remaining. To achieve this, we decompose features into two subspaces: one capturing forgetting-class contributions and the other preserving remaining-class knowledge. Specifically, we perform singular value decomposition (SVD) on the feature matrix of forgetting classes and use the span of its dominant singular vectors to define the forgetting subspace. Similarly, we apply the same procedure to the remaining classes to construct its subspace. We then detail the process as follows.", + "bbox": [ + 511, + 402, + 905, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a PTM, $\\mathcal{M}$ , we construct a representation matrix $\\pmb{R}_{f,i} = \\left[\\pmb{x}_{1,i},\\dots,\\pmb{x}_{n_{f_i},i}\\right]\\in \\mathbb{R}^{n_{f_i}\\times d}$ at each layer $l$ where columns correspond to representations of $n_{f_i}$ samples from the remaining classes $\\mathbf{D}_{\\mathbf{f},\\mathbf{i}}$ , and $d$ is the output dimension3. SVD is applied to $\\pmb{R}_{f,i} = \\pmb{U}_{f,i}\\pmb{\\Sigma}_{f,i}(\\pmb{V}_{f,i})^T$ , followed by its $k$ -rank approximation $(\\pmb{R}_{f,i})_k$ satisfying $||(\\pmb{R}_{f,i})_k||_F^2\\geq \\epsilon ||\\pmb{R}_{f,i}||_F^2$ for a threshold $\\epsilon$ . The subspace $\\mathcal{S}_{f,i} = \\mathrm{span}\\{\\pmb{B}_{f,i}\\}$ is defining using the first $k$ vectors $\\pmb{B}_{f,i} = \\{\\pmb{u}_{1,i},\\dots,\\pmb{u}_{k,i}\\}$ from $\\pmb{U}_{f,i}$ , capturing directions with the highest singular values. Similarly, the feature matrix $\\pmb{R}_{r,i}$ for remaining classes $\\mathbf{D}_{\\mathbf{r},\\mathbf{i}}$ defines the subspace $\\mathcal{S}_{r,i} = \\mathrm{span}\\{\\pmb{B}_{r,i}\\}$ . The two sets of layer-wised feature subspace, $\\{\\mathcal{S}_{f,i}^1,\\dots,\\mathcal{S}_{f,i}^l\\}$ and $\\{\\mathcal{S}_{r,i}^1,\\dots,\\mathcal{S}_{r,i}^l\\}$ , represent the core part for forgetting and remaining classes,", + "bbox": [ + 511, + 657, + 906, + 876 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "3Here, we omit the notation $l$ for simplicity.", + "bbox": [ + 529, + 886, + 764, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1959", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/60c69cb922bd064d6bed1158b52380dd84e70f8643f9fe22e271c4f25a6466dc.jpg", + "image_caption": [ + "Figure 2. Overview of FG-OrIU: (a) For each unlearning task, we decompose the layer-wised- forgetting and remaining subspace separately (Sec. 4.1). (b) LoRA is inserted at each layer, with constraints applied at both the feature and gradient levels for forgetting and remaining (Sec. 4.2). (c) For new unlearning tasks, the forgetting subspace expands while the remaining contracts (Sec. 4.3)." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 898, + 296 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "respectively, isolating features to be erased while preserving those to retain. By explicitly decomposing feature subspaces across all layers, we strengthen supervision signals for parameter updates and mitigate the risk of gradient obfuscation [2] in deep networks, where gradient inaccuracies caused by layer-wise nonlinear transformations could otherwise weaken the unlearning effect. This multi-layer constraint ensures consistent suppression of forgetting class features throughout the model hierarchy.", + "bbox": [ + 88, + 345, + 483, + 482 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Dual Orthogonal Projection", + "text_level": 1, + "bbox": [ + 89, + 482, + 341, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prior to unlearning task $\\mathcal{T}_i$ , we freeze both the backbone and the classification head of the PTM and insert lightweight LoRA modules [23] at each layer. We then denote the model with LoRA inserted as $\\mathcal{M}_{\\theta}$ , where $\\theta$ represents the trainable parameters in LoRA. After unlearning, the updates in LoRA, $\\theta^{*}$ , will be merged into the initial weight of the backbone, $\\mathcal{M}_i \\xrightarrow[\\text{merge}]{\\theta^*} \\mathcal{M}_{i+1}$ .", + "bbox": [ + 89, + 503, + 483, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Building on the feature subspaces $S_f$ and $S_r$ defined in Section 4.1, we propose a dual orthogonal projection mechanism that operates through both forward feature alignment and backward gradient correction.", + "bbox": [ + 89, + 619, + 483, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2.1. Feature Orthogonal Projection", + "text_level": 1, + "bbox": [ + 89, + 685, + 352, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by [45, 55], for forgetting classes, achieving thorough removal at the feature level requires that the features generated by the unlearned model, $\\mathcal{M}_{\\theta}$ , deviate as much as possible from the feature subspace $S_{f}$ . Meanwhile, we aim to ensure that the unlearned model's feature subspace for remaining classes aligns with $S_{r}$ . The orthogonality constraint between $S_{f}$ and the unlearned features serves as a critical mechanism for facilitating class-specific forgetting: under a fixed classification head, features orthogonal to the original subspace $S_{f}$ induce maximal perturbations in class logits, as the inner product between orthogonal vectors diminishes to zero. This forces the unlearned model to suppress discriminative patterns associated with forgetting", + "bbox": [ + 89, + 704, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "classes in its intermediate representations. To do so, we propose the following two regularization constraints. For forgetting classes, let $\\boldsymbol{x}_f$ denote a batch of forgetting samples from $\\mathbf{D}_{\\mathbf{f},\\mathrm{i}}$ , and $h_f^l$ denote the $l$ -th layer features gained with the unlearned model $\\mathcal{M}_{\\theta}$ . For complete forgetting, we enforce:", + "bbox": [ + 511, + 345, + 906, + 436 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {h} _ {f} ^ {l} \\perp \\mathcal {S} _ {f, i} ^ {l} \\quad \\Rightarrow \\quad \\min | | \\mathbf {P} _ {f} ^ {l} \\boldsymbol {h} _ {f} ^ {l} | | _ {F} ^ {2} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 460, + 906, + 483 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{P}_f^l = \\mathbf{B}_{f,i}^l (\\mathbf{B}_{f,i}^l)^\\top$ is the projection matrix for the forgetting subspace $S_{f,i}^{l}$ and $\\| \\cdot \\| _F^2$ is the F-norm. Eq.(3) ensures that the features of $\\mathbf{D}_{\\mathbf{f},\\mathbf{i}}$ gained from the unlearned model $\\mathcal{M}_{\\theta}$ lie outside the original feature subspace gained from $\\mathcal{M}$ , erasing the unwanted information. Thus, we define the flop-f loss:", + "bbox": [ + 511, + 493, + 905, + 589 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {f o p - f} = \\sum_ {l = 1} ^ {m} \\left\\| \\mathbf {P} _ {f} ^ {l} \\boldsymbol {h} _ {f} ^ {l} \\right\\| _ {F} ^ {2}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 619, + 599, + 905, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $m$ denotes the total number of layers in the PTM. For remaining classes, we follow previous works and apply the standard cross-entropy loss by constraining the model's final output, as follows:", + "bbox": [ + 511, + 648, + 905, + 709 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c e} = \\mathcal {L} \\left(\\mathcal {X} _ {\\mathbf {D} _ {r, 1}}, \\mathcal {Y} _ {\\mathbf {D} _ {r, 1}}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 720, + 903, + 738 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}$ denotes the cross-entropy loss function. In addition, we also propose applying explicit constraint at the feature level for the remaining classes to better preserve the desired information. Specifically, $\\pmb{x}_r$ denote a batch of remaining samples from $\\mathbf{D}_{\\mathbf{r},\\mathbf{i}}$ , and $h_r^l$ denote the $l$ -th layer features gained with the unlearned model $\\mathcal{M}_{\\theta}$ . For better preservation, we maintain feature fidelity through:", + "bbox": [ + 511, + 748, + 906, + 854 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min \\left\\| \\boldsymbol {h} _ {r} ^ {l} - \\mathbf {P} _ {r} ^ {l} \\boldsymbol {h} _ {r} ^ {l} \\right\\| _ {F} ^ {2} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 878, + 903, + 897 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1960", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{P}_r^l = \\mathbf{B}_{r,i}^l (\\mathbf{B}_{r,i}^l)^\\top$ is the projection matrix for the remaining feature subspace $S_{r,i}^{l}$ .", + "bbox": [ + 89, + 88, + 482, + 125 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Eq.(6) ensures that the features of $\\mathbf{D}_{\\mathbf{r},\\mathbf{i}}$ gained from the unlearned model $\\mathcal{M}_{\\theta}$ , align with the original feature subspace gained from $\\mathcal{M}$ , maintaining the better preservation. Therefore, We draw the following $f_{op - r}$ loss:", + "bbox": [ + 89, + 125, + 482, + 185 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {f o p - r} = \\sum_ {l = 1} ^ {m} \\left\\| \\boldsymbol {h} _ {r} ^ {l} - \\mathbf {P} _ {r} ^ {l} \\boldsymbol {h} _ {r} ^ {l} \\right\\| _ {F} ^ {2}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 191, + 482, + 231 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, the total loss is:", + "bbox": [ + 109, + 237, + 272, + 252 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {c e} + \\lambda_ {1} \\mathcal {L} _ {f o p - f} + \\lambda_ {2} \\mathcal {L} _ {f o p - r}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 260, + 482, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{1},\\lambda_{2}$ are the coefficients during training.", + "bbox": [ + 89, + 284, + 413, + 300 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.2. Gradient Orthogonal Projection", + "text_level": 1, + "bbox": [ + 89, + 305, + 362, + 320 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To prevent parameter updates from reintroducing forgotten knowledge while preserving retained features, we design a gradient correction mechanism inspired by gradient projection techniques in multi-task learning [56] and continual learning [40, 62]. Given the raw gradient $g_{l}$ at layer $l$ , we apply the following modifications:", + "bbox": [ + 89, + 324, + 483, + 414 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\hat {g} _ {l} = g _ {l} - g _ {l} ^ {\\prime} - g _ {l} ^ {\\prime \\prime} \\tag {9} \\\\ = \\boldsymbol {g} _ {l} - \\underbrace {\\boldsymbol {g} _ {l} (\\boldsymbol {I} - \\mathbf {B} _ {f , i} ^ {l} (\\mathbf {B} _ {f , i} ^ {l}) ^ {\\top})} _ {(t e r m 1: g o p - f)} - \\underbrace {\\boldsymbol {g} _ {l} \\mathbf {B} _ {r , i} ^ {l} (\\mathbf {B} _ {r , i} ^ {l}) ^ {\\top}} _ {(t e r m 2: g o p - r)}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 421, + 482, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Term (gop-f), $g_{l}^{\\prime}$ : Removing the components of $g_{l}$ that are orthogonal to $S_{f,i}^{l}$ not only amplifies updates focusing on directions most critical for encoding forgetting class features, but also blocks parameter updates that could reintroduce removed knowledge.", + "bbox": [ + 89, + 491, + 483, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Term (gop-r), $g_{l}^{\\prime \\prime}$ : We aim to ensure that forgetting specific classes does not affect the remaining classes. Formally, for the unlearned model $f_{\\mathcal{M}_{\\theta}}$ and $x_{r}$ from the remaining classes $\\mathbf{D}_{\\mathbf{r},\\mathbf{i}}$ , the following equation holds: $f_{\\mathcal{M}_{\\theta}}(x_{r}) = f_{\\mathcal{M}}(x_{r})$ . Specifically, we consider the input and output of layer $l$ as follows:", + "bbox": [ + 89, + 566, + 483, + 657 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {W} _ {l} \\boldsymbol {x} _ {r} + \\Delta \\theta_ {l} \\boldsymbol {x} _ {r} = \\boldsymbol {W} _ {l} \\boldsymbol {x} _ {r} \\\\ \\Delta \\theta_ {l} \\boldsymbol {x} _ {\\boldsymbol {r}} = \\mathbf {0} \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 666, + 480, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Delta \\theta_{l}$ represents the change in the LoRA parameters inserted at layer $l$ , which corresponds to its gradient. $W_{l}$ is the parameter at layer $l$ from the pre-trained model, and these parameters remain frozen during unlearning.", + "bbox": [ + 89, + 708, + 482, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Eq.(10) indicates that if the update direction of the parameters, i.e., the gradient of the trainable LoRA parameters, lies in the complementary space of $S_{r,i}^{l}$ , i.e., $\\pmb{g}_{l} \\perp S_{r,i}^{l}$ , then $\\Delta \\theta \\pmb{x}_r = \\eta \\pmb{g}_l \\pmb{x}_r = \\mathbf{0}$ (where $\\eta$ is the learning rate), ensuring that the unlearned model's performance on the remaining classes is not affected. Therefore, we further refine the gradient by removing its components along $S_{r,i}^{l}$ .", + "bbox": [ + 89, + 768, + 482, + 875 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, modified gradient $\\hat{g}_l$ serves the real gradient for parameter updates at the layer $l$ .", + "bbox": [ + 89, + 875, + 482, + 906 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Dynamic Subspace Adaptation", + "text_level": 1, + "bbox": [ + 511, + 90, + 787, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When encountering a new unlearning task $\\mathcal{T}_{i + 1}$ , our method only needs to update the two decomposed feature subspaces. Since the forgetting classes $\\mathbf{D_{f,i + 1}}$ in $\\mathcal{T}_{i + 1}$ are a subset of the remaining classes from the previous task $\\mathcal{T}_i$ , i.e., $\\mathbf{D_{f,i + 1}}\\subseteq \\mathbf{D_{r,i}}$ , the feature subspaces can be efficiently adjusted. For simplicity, we omit the notion $l$ of layers.", + "bbox": [ + 511, + 106, + 903, + 195 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first construct a representation matrix $\\mathbf{R}_{f,i+1}$ using only the forgetting classes $\\mathbf{D}_{\\mathbf{f},i+1}$ . Before applying SVD and $k$ -rank approximation, we remove common bases already present in $S_{f,i}$ to ensure newly added bases are unique and orthogonal to existing ones:", + "bbox": [ + 511, + 196, + 903, + 271 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\hat {\\boldsymbol {R}} _ {f, i + 1} = \\boldsymbol {R} _ {f, i + 1} - \\boldsymbol {B} _ {f, i} \\left(\\boldsymbol {B} _ {f, i}\\right) ^ {T} \\left(\\boldsymbol {R} _ {f, i + 1}\\right) (11) \\\\ = \\boldsymbol {R} _ {f, i + 1} - \\boldsymbol {R} _ {f, i + 1} ^ {\\text {p r o j}}. (12) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 276, + 903, + 320 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, we perform SVD on $\\hat{R}_{f,i+1} = \\hat{U}_{f,i+1}\\hat{\\Sigma}_{f,i+1}(\\hat{V}_{f,i+1})^T$ and select $h$ new orthogonal from $\\hat{U}_{f,i+1}$ . Then, the feature subspace of forgetting classes is expanded as $\\mathcal{S}_{f,i+1} = \\text{span}\\{B_{f,i+1}\\}$ , where $\\mathbf{B}_{f,i+1} = [B_{f,i};\\mathbf{u}_1,\\dots,\\mathbf{u}_h]$ with $h$ new bases from $\\hat{U}_{f,i+1}$ . And the feature subspace $\\mathcal{S}_{r,i+1}$ of remaining classes is contracted to remove the overlapping features accordingly.", + "bbox": [ + 511, + 328, + 905, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 513, + 460, + 638, + 476 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 513, + 478, + 702, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets, Benchmarks and Pre-trained Models.", + "text_level": 1, + "bbox": [ + 511, + 500, + 846, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following [60], we evaluate FG-OrIU on Classification and Recognition tasks. For classification, we use ImageNet100 [39], CUB-200 [47], and Omnibenchmark [59], with a ViT-B/16 [14] pre-trained in PyTorch [33]. For face recognition, we use CASIA-Face100 [60], a subset of CASIA-WebFace [54], and MS-Celeb-100, sampled from [21]. The pre-trained model is the Face Transformer from [63]. We focus on class-unlearning scenarios with two benchmarks: (a) MU, varying numbers of classes are removed; (b) IU, 20 classes are removed per task over 4 tasks. All setting is in the form of Y-X, which means experiments start from a PTM (Y classes originally) and forget X classes. More details about implementations are in Appendix ??", + "bbox": [ + 511, + 513, + 905, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. Following [60], we evaluate performance using average accuracy (Acc) for both forgetting and remaining classes. The accuracy for forgetting classes should approach zero, while the accuracy for remaining classes should remain close to the original model's performance. Additionally, we adopt the H-Mean [43] metric to assess overall performance after learning task $\\mathcal{T}_t$ , computed as:", + "bbox": [ + 511, + 710, + 903, + 808 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nH - M e a n ^ {(t)} = \\frac {2 A c c _ {r} ^ {(t)} \\cdot D r o p ^ {(t)}}{A c c _ {r} ^ {(t)} + D r o p ^ {(t)}}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 824, + 903, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here $Acc_{r}^{(t)}$ represents accuracy on the retained dataset after task $\\mathcal{T}_t$ , while $Drop^{(t)} = Acc_{f}^{(t - 1)} - Acc_{f}^{(t)}$ mea", + "bbox": [ + 511, + 866, + 903, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1961", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/bf039ee8be30e903e540a9871cbbb7dd65d97a2eb85ba8f91aaf3d962c5015c8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsTunable Ratio ↓100-5100-10100-50100-90
H ↑Accr ↑Accf ↓H ↑Accr ↑Accf ↓H ↑Accr ↑Accf ↓H ↑Accr ↑Accf ↓
Pre-train--73.8570.88-73.8172.74-72.4574.88-72.3273.86
L2*99.73%67.1366.963.5867.7464.871.8664.3956.610.2452.4540.680.04
EWC*99.73%69.6568.690.2469.1665.920.0062.4853.610.0044.4131.750.00
MAS*99.73%69.7369.290.7269.3866.410.1262.6753.880.0046.7934.240.00
LwF99.73%67.9568.550.0070.0867.620.0063.8155.590.0061.2552.320.00
DER99.73%69.7068.550.0070.0867.620.0063.8155.590.0057.0346.440.00
DER++99.73%69.7068.560.0070.5868.550.0064.6156.820.0061.4052.540.00
FDR99.73%70.3169.740.0070.0367.510.0065.4058.040.0053.8542.370.00
Retrain100.00%16.499.330.0018.0210.280.0019.7111.350.0046.4733.900.00
SCRUB99.73%67.7864.940.0068.2664.310.0065.8258.710.0016.119.040.00
SCRUB-S99.73%70.2969.950.2471.1969.810.1257.4951.209.3417.539.940.00
LIRF*50.66%25.5667.6755.1326.3565.8356.2647.1358.9535.6254.4944.293.06
GS-LoRA1.28%71.0271.160.0071.7670.810.0071.2968.050.0273.7173.560.00
GS-LoRA++1.28%71.1271.360.0072.0471.350.0071.5668.520.0072.8571.860.00
FG-OrIU1.28%71.7872.710.0072.5272.310.0074.1873.510.0075.4977.150.00
", + "bbox": [ + 94, + 80, + 906, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Static Machine Unlearning results for face recognition task on CASIA-Face100. Accr (%) and Accf (%) are the accuracies of remaining and forgetting classes. * denotes the original methods with a rehearsal buffer. Retrain represents retraining the model using replay data and the training epoch is the same as other methods to ensure a fair comparison. Pre-train denotes the results before forgetting.", + "bbox": [ + 89, + 343, + 906, + 387 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4aa4a42371b07dd43463877fa058fb00049a3065512c6495781bb452334269bc.jpg", + "image_caption": [ + "Figure 3. Accuracy on forgetting classes when recovering on CASIA-Face100. The line (Pre-train) is the result before unlearning. The line (FG-OrIU) is Our method." + ], + "image_footnote": [], + "bbox": [ + 96, + 396, + 475, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "sures the performance drop on forgetting classes before and after training. After unlearning task $\\mathcal{T}_t$ , we evaluate performance on all previously forgetting classes from tasks $\\mathcal{T}_1, \\mathcal{T}_2, \\dots, \\mathcal{T}_{t-1}$ .", + "bbox": [ + 89, + 627, + 482, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. Following [60], we compare FG-OrIU against (a) continual learning (CL) methods, including L2 regularization, EWC [27], MAS [1], LwF [30], DER [6], and FDR [3], as well as (b) machine unlearning (MU) methods such as BAD-T [12], LIRF [53], SCRUB/SCRUB-S [29], GS-LoRA/GS-LoRA++ [60, 61], and retraining. To ensure forgetting occurs in the backbone, we freeze the classification head for all methods. In the retraining setup, we train a randomly initialized model using replay data, ensuring the number of training epochs matches that of other methods.", + "bbox": [ + 88, + 686, + 482, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Results and Comparison", + "text_level": 1, + "bbox": [ + 89, + 840, + 316, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2 presents the results for static Machine Unlearning task. FG-OrIU exhibits significantly degraded performance on forgetting classes while maintaining the perfor", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "mance of remaining classes. For example, in the settings where 50 and 90 classes are removed per task, FG-OrIU achieves improvements of $2.62\\%$ and $2.64\\%$ in $H$ , and $4.99\\%$ and $5.29\\%$ in $ACC_{r}$ compared to GS-LoRA. Notably, FG-OrIU keeps the pre-trained model frozen and only trains the LoRA modules, enabling effective and efficient forgetting. Table 3 and 4 show the results for Incremental Unlearning task. FG-OrIU consistently outperforms all methods. Specifically, CL methods struggle to achieve effective forgetting on ImageNet100, as seen in DER, which retains high $ACC_{f}$ of $10.40\\%$ , $10.90\\%$ , $8.80\\%$ , and $3.40\\%$ across four tasks, respectively. On the other hand, MU methods fail to maintain their forgetting effect over sequential unlearning tasks. For example, LIRF and SCRUB-S show an increase in $ACC_{o}$ on CASIA-Face100, which becomes more pronounced as the sequence of unlearning tasks grows. GS-LoRA is specifically designed for IU. Compared to it, our method achieves better forgetting while also demonstrating superior performance in preserving the remaining classes, i.e., substantial improvements in $ACC_{r}$ , particularly in long-sequence unlearning scenarios, achieving gains of $5.15\\%$ and $4.6\\%$ in the fourth unlearning task on two different datasets, respectively.", + "bbox": [ + 511, + 390, + 906, + 738 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 747, + 666, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our FG-OrIU consists of four components, i.e., feature orthogonal projection for forgetting, feature orthogonal projection for remaining, gradient orthogonal projection for forgetting, and gradient orthogonal projection for remaining, denoted as $fop - f$ , $fop - r$ , $gop - f$ , and $gop - r$ . We independently remove each of these components to demonstrate their individual contributions. We observe that when the regularization term or gradient modification related to remaining is removed, the performance on the remaining", + "bbox": [ + 509, + 763, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1962", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6005b36a826d2e2f50f4420788a4e0c576199b40c4a650a4e929c901f3ca41cd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsTask 1(100-20)Task 2(80-20)Task 3(60-20)Task 4(40-20)
H↑Accr↑Accf↓H↑Accr↑Accf↓Acco↓H↑Accr↑Accf↓Acco↓H↑Accr↑Accf↓Acco↓
Pre-train-74.3174.65-74.5073.80--74.8073.91--74.4775.11-
L2*66.9162.132.1666.6661.741.379.3666.4261.371.5411.8366.9561.020.978.00
EWC*67.7161.950.0067.7162.550.000.0067.0961.430.000.1068.0262.140.000.23
MAS*67.5261.630.0068.1263.250.000.0068.1563.230.000.0367.7061.610.000.00
LwF69.4365.040.2169.9466.470.000.2170.3467.110.000.0070.8967.120.000.04
DER70.7567.240.0068.8864.580.000.0068.9564.620.000.0069.4164.510.000.00
DER++70.0165.900.0068.9964.770.000.0069.9666.420.000.0069.8565.280.000.00
FDR68.2962.920.0067.3962.010.000.0069.1664.990.000.0072.5170.080.000.00
Retrain17.299.770.0031.1219.720.000.0041.1728.801.720.0052.4441.774.660.09
BAD-T68.9063.970.0069.5665.950.210.0070.2166.870.000.3472.7170.460.000.04
LIRF*30.5964.4654.6034.0562.0350.3443.5044.5662.5339.2936.5840.3662.6245.3427.96
SCRUB70.3966.590.0070.5567.850.320.0071.0168.330.000.4473.3671.680.000.05
SCRUB-S72.4170.340.0570.0671.535.1615.2273.3873.090.2415.8175.3376.240.686.47
GS-LoRA74.4074.160.0073.5973.370.000.0574.3674.880.060.0073.7672.450.001.93
GS-LoRA++73.9774.160.0073.4873.160.000.1074.2074.510.000.0075.2375.360.000.00
FG-OrIU74.4574.250.0074.6475.500.000.0075.4977.140.000.0077.7180.510.000.00
", + "bbox": [ + 93, + 88, + 906, + 323 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/481a8bb287438625803ad9178f1036d49e1f67c9e5fa94e78b3d1d7d86d090b8.jpg", + "table_caption": [ + "Table 3. Incremental Unlearning results for face recognition task on CASIA-Face100. $Acc_{o}(\\%)$ is the accuracy of old tasks, i.e., the accuracy on all previously forgetting classes in task $\\{\\mathcal{T}_1,\\dots ,\\mathcal{T}_{t - 1}\\}$ . With four tasks, each forgetting 20 classes, desired forgetting should consider both $Acc_{r}$ and $Acc_{f}$ . We bold the best and underline the second-best $H$ results." + ], + "table_footnote": [], + "table_body": "
MethodsTask 1(100-20)Task 2(80-20)Task 3(60-20)Task 4(40-20)
H↑Accr↑Accf↓H↑Accr↑Accf↓Acco↓H↑Accr↑Accf↓Acco↓H↑Accr↑Accf↓Acco↓
Pre-train-89.9387.40-90.0389.60--89.6590.80--92.8086.50-
L2*79.3988.3315.3083.5988.3710.3048.1084.0188.2010.6037.1582.7289.909.9020.97
EWC*82.6978.550.1085.7182.570.500.2087.3384.900.902.9086.9389.502.008.27
MAS*80.4474.500.0081.5174.770.000.6082.1875.050.000.6081.9377.900.101.77
LwF85.8285.150.9087.2685.400.402.4087.3584.150.002.3586.0585.700.101.07
DER82.8889.7310.4084.2290.5710.9050.8086.3691.208.8042.0088.5294.703.4024.90
DER++83.4589.739.4084.1390.6311.1050.1087.5091.306.8040.4089.2494.802.2020.80
FDR29.5117.780.5036.2122.700.100.0040.7926.300.000.0047.9133.502.401.27
Retrain48.2033.280.0056.2841.301.300.0061.1447.555.200.2067.2360.6011.000.53
BAD-T74.7465.290.0078.0969.210.000.1076.6566.320.000.0076.2568.170.000.00
LIRF*52.7264.2842.7161.0569.8335.3632.1560.9660.3229.1830.9158.5168.4635.4121.04
SCRUB75.8767.030.0076.2366.330.000.0075.4664.550.000.0077.5670.300.000.00
SCRUB-S83.4479.830.0082.5976.600.000.1079.8171.200.000.0078.2371.400.000.07
GS-LoRA86.1385.000.1087.3385.270.101.1087.8485.150.100.3586.3086.300.200.07
GS-LoRA++86.6385.980.1087.2285.230.300.3087.4884.400.000.6087.0087.500.000.07
FG-OrIU87.0186.730.1088.4587.430.100.1088.9087.350.300.0089.1592.100.100.00
", + "bbox": [ + 93, + 368, + 906, + 601 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/053cd906e44a256255dcfb553dbd3257dc76fd678f2569479e6eeea7accb2147.jpg", + "table_caption": [ + "Table 4. Incremental Unlearning results for classification task on ImageNet100." + ], + "table_footnote": [], + "table_body": "
VariantsTask 1(100-20)Task 2(80-20)
Accr↑Accf↓Accr↑Accf↓Acco↓
Pre-train74.3074.6074.5073.80-
FG-OrIU74.250.0075.500.000.00
w/o fop-f73.010.2076.240.421.28
w/oGOP-f73.920.1076.584.370.82
w/o fop-r73.920.0574.980.210.15
w/oGOP-r71.450.0073.960.000.00
", + "bbox": [ + 94, + 617, + 480, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "classes significantly declines. Meanwhile, when the regularization term or gradient modification related to forgetting is removed, the performance on the forgetting classes shows a slight improvement. Thus, each component provides clear benefits in achieving a better trade-off between forgetting and remaining while preserving the old forgetting effect.", + "bbox": [ + 89, + 763, + 482, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Discussion", + "text_level": 1, + "bbox": [ + 89, + 854, + 209, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "IU is designed to ensure strict privacy protection. This raises several key questions: (a) Does the unlearned model", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a1530471620ff3d069593db5b8450fb6e28a2da15ee1cd7eb6cde12403b88b2c.jpg", + "table_caption": [ + "Table 5. Ablation studies on CASIA-Face100." + ], + "table_footnote": [], + "table_body": "
ExpsMetircPre-trainDERFDRSCRUBGS-LoRAOurs
Exp1SSIM0.940.600.340.750.390.01
PSNR31.3221.2216.8923.4017.608.37
Exp2SSIM0.950.490.660.630.380.02
PSNR31.1519.1722.1521.1217.328.72
Exp3SSIM0.930.530.490.600.410.01
PSNR31.1020.0319.8322.0115.928.55
", + "bbox": [ + 516, + 627, + 903, + 727 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. SSIM $\\uparrow$ and PSNR(db) $\\uparrow$ scores for the reconstructed images using DIP. Here, Pre-train serves as the upper bound.", + "bbox": [ + 511, + 731, + 906, + 761 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "still retain any private information? (b) Can the forgetting effect be reversed? and (c) How does unlearning impact the feature subspace? To address these questions, we design the following experiment.", + "bbox": [ + 511, + 777, + 906, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4.1. Semantic Residual Analysis", + "text_level": 1, + "bbox": [ + 511, + 843, + 751, + 856 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We first investigate the semantic information contained in features after unlearning. We utilize Deep Image Prior (DIP) [46] to reconstruct an image from the features ob", + "bbox": [ + 511, + 856, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1963", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ff79610a9d205a4eee421c6e9612ec41dc8c28e591552eb48da22e0236ca7245.jpg", + "image_caption": [ + "(a) Pre-train" + ], + "image_footnote": [], + "bbox": [ + 98, + 89, + 336, + 273 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6f2b569fb9a2a2172283ea05658bc1fa68c98ff58a4b73c01c589025f16ea78e.jpg", + "image_caption": [ + "(b) Ours", + "Figure 4. TSNE visualization with 5 forgetting classes and 10 remaining classes, using different model as feature extractor: (a) the pretrained model, (b) the unlearned model through FG-OrIU, and (c) through GS-LoRA." + ], + "image_footnote": [], + "bbox": [ + 359, + 90, + 602, + 272 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3c353ae144c0369e8c264f890b96d1fcd642199b14b1c315ab8990ef3f35ca47.jpg", + "image_caption": [ + "(c) GS-LoRA" + ], + "image_footnote": [], + "bbox": [ + 625, + 90, + 905, + 272 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tained from the last layer of the unlearned model. Figure 1 presents the visualization of the reconstructed image, which allows us to infer the semantic information encoded in features. More results are provided in the Appendix ??", + "bbox": [ + 89, + 318, + 483, + 380 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Additionally, we measure the SSIM and PSNR scores for the images generated by DIP from the unlearned model using different IU methods as a quantitative results. As shown in Table 6, the SSIM values of the reconstructed images from the compared methods range between 0.35 and 0.75, with the highest being 0.75 for SCRUB on $exp1$ , while the PSNR values range between 16 and 24, with the highest being 23.40 for SCRUB on $exp1$ . This indicates that the reconstructed image achieves a high similarity with the original image, suggesting that the information in features has been partially preserved. In contrast, our method yields an SSIM of only 0.01 and a PSNR of 8.55, indicating a significant difference between the reconstructed and original images. This suggests that the semantic information at the feature level has been substantially obfuscated.", + "bbox": [ + 88, + 387, + 483, + 613 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4.2. Forgetting Recoverable Analysis", + "text_level": 1, + "bbox": [ + 89, + 628, + 362, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Given an unlearned model obtained through different IU methods, we freeze its backbone and retrain a new classification head using the entire training dataset, which includes both forgetting and remaining classes. As illustrated in Figure 3, MLP* represents a trivial unlearning method where the classification head's parameters for the forgetting classes are simply set to zero. Consequently, the underlying features remain intact, enabling effortless recovery of the forgotten knowledge. Additionally, existing IU methods primarily impose constraints on the final output rather than directly modifying feature representations. As a result, while features may be weakened to some extent, their core structure remains largely unchanged. Thus, these methods allow the forgotten knowledge to be recovered to $15\\% -70\\%$ of its original performance, which leads to superficial forgetting. In contrast, our method achieves deep forgetting, making recovery infeasible.", + "bbox": [ + 89, + 647, + 483, + 905 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4.3. Feature Subspace Analysis", + "text_level": 1, + "bbox": [ + 511, + 319, + 745, + 334 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To analyze the impact of direct constraints on features and gradients for forgetting and remaining classes, we conduct an experiment. We remove 20 classes while retaining 80, then extract features from 5 forgetting and 10 remaining classes and visualize them using t-SNE.", + "bbox": [ + 511, + 338, + 905, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A key observation is that GS-LoRA causes some forgetting class samples to move closer to or even overlap with certain remaining class samples (e.g., F21, $\\mathrm{F}74 \\rightarrow \\mathrm{R}92$ in c). This suggests that merely applying constraints at the final output, resulting in degraded features from the forgetting classes entangling with those of the remaining classes. Consequently, this degradation negatively impacts the performance of the remaining classes. In contrast, our method enforces forgetting at both feature and gradient level, ensuring that features from the forgetting classes do not shift toward any remaining class but instead become linearly inseparable within the forgetting set. Notably, an unexpected yet beneficial side effect emerges: in the original PTM, certain remaining class samples (e.g., R59) were slightly entangled with forgetting classes (e.g., F90 and F46). However, after applying our method, these classes become completely disentangled. This observation aligns with the results in Table 3, where the accuracy of the remaining classes improves after unlearning (e.g., $75.36 \\rightarrow 80.51$ ), demonstrating that our method not only enhances forgetting effectiveness but also benefits the generalization of the remaining classes.", + "bbox": [ + 511, + 414, + 906, + 731 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 746, + 633, + 760 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose FG-OrIU for incremental unlearning in vision PTMs. By enforcing orthogonal constraints on both features and gradients, FG-OrIU achieves irreversible forgetting while maintaining performance on the remaining classes. Its dynamic subspace adaptation ensures robust handling of sequential deletion requests. Extensive experiments confirm its superiority. While FG-OrIU is designed for class-level unlearning, extending it to sample-level settings remains a promising direction for future work.", + "bbox": [ + 511, + 763, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1964", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by National Natural Science Foundation of China under Grant 62402430, 62206248, 62476238, Zhejiang Provincial Natural Science Foundation of China under Grant LQN25F020008, and Aeronautical Science Foundation of China 20240048076001. Qian Feng would like to extend his gratitude to Jiahang Tu and Hanbin Zhao from Zhejiang University for their discussions regarding the methods and experiments. Additionally, he is thankful to Mintong Kang from UIUC for his dedicated suggestions on improving the overall writing of the paper.", + "bbox": [ + 89, + 114, + 485, + 282 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 292, + 187, + 308 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In Proceedings of the European conference on computer vision (ECCV), pages 139-154, 2018. 6", + "[2] Anish Athalye, Nicholas Carlini, and David Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In International conference on machine learning, pages 274-283. PMLR, 2018. 4", + "[3] Ari S Benjamin, David Rolnick, and Konrad Kording. Measuring and regularizing networks in function space. arXiv preprint arXiv:1805.08289, 2018. 6", + "[4] Jacopo Bonato, Marco Cotogni, and Luigi Sabetta. Is retain set all you need in machine unlearning? restoring performance of unlearned models with out-of-distribution images. In European Conference on Computer Vision, pages 1-19. Springer, 2024. 1", + "[5] Lucas Bourtoule, Varun Chandrasekaran, Christopher A Choquette-Choo, Hengrui Jia, Adelin Travers, Baiwu Zhang, David Lie, and Nicolas Papernot. Machine unlearning. In 2021 IEEE symposium on security and privacy (SP), pages 141-159. IEEE, 2021. 1, 2", + "[6] Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark experience for general continual learning: a strong, simple baseline. Advances in neural information processing systems, 33:15920-15930, 2020. 3, 6", + "[7] Yinzhi Cao and Junfeng Yang. Towards making systems forget with machine unlearning. In 2015 IEEE symposium on security and privacy, pages 463-480. IEEE, 2015. 1", + "[8] Romit Chatterjee, Vikram Chandawat, Ayush Tarun, Ankur Mali, and Murari Mandal. A unified framework for continual learning and machine unlearning. arXiv preprint arXiv:2408.11374, 2024. 1, 3", + "[9] Jiaao Chen and Diyi Yang. Unlearn what you want to forget: Efficient unlearning for llms. arXiv preprint arXiv:2310.20150, 2023.", + "[10] Xinwen Cheng, Zhehao Huang, Wenxin Zhou, Zhengbao He, Ruikai Yang, Yingwen Wu, and Xiaolin Huang. Remaining-data-free machine unlearning by suppressing sample contribution. arXiv preprint arXiv:2402.15109, 2024. 1, 2, 3" + ], + "bbox": [ + 93, + 316, + 485, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Zhi-Yi Chin, Chieh-Ming Jiang, Ching-Chun Huang, Pin-Yu Chen, and Wei-Chen Chiu. Prompting4debugging: Red-teaming text-to-image diffusion models by finding problematic prompts. arXiv preprint arXiv:2309.06135, 2023. 1, 3", + "[12] Vikram S Chandawat, Ayush K Tarun, Murari Mandal, and Mohan Kankanhalli. Can bad teaching induce forgetting? unlearning in deep networks using an incompetent teacher. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7210-7217, 2023. 1, 2, 3, 6", + "[13] Meng Ding, Jinhui Xu, and Kaiyi Ji. Why fine-tuning struggles with forgetting in machine unlearning? theoretical insights and a remedial approach. arXiv preprint arXiv:2410.03833, 2024. 3", + "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.5", + "[15] Chongyu Fan, Jiancheng Liu, Yihua Zhang, Eric Wong, Dennis Wei, and Sijia Liu. Salun: Empowering machine unlearning via gradient-based weight saliency in both image classification and generation. arXiv preprint arXiv:2310.12508, 2023. 2, 3", + "[16] Jack Foster, Stefan Schoepf, and Alexandra Brintrup. Fast machine unlearning without retraining through selective synaptic dampening. In Proceedings of the AAAI conference on artificial intelligence, pages 12043-12051, 2024. 1, 2, 3", + "[17] Chongyang Gao, Lixu Wang, Chenkai Weng, Xiao Wang, and Qi Zhu. Practical unlearning for large language models. arXiv preprint arXiv:2407.10223, 2024. 1, 2, 3", + "[18] Laura Graves, Vineel Nagisetty, and Vijay Ganesh. Amnesiac machine learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11516-11524, 2021. 2, 3", + "[19] Keltin Grimes, Collin Abidi, Cole Frank, and Shannon Gallagher. Gone but not forgotten: Improved benchmarks for machine unlearning. arXiv preprint arXiv:2405.19211, 2024. 1", + "[20] Chuan Guo, Tom Goldstein, Awni Hannun, and Laurens Van Der Maaten. Certified data removal from machine learning models. arXiv preprint arXiv:1911.03030, 2019. 1", + "[21] Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. Ms-celeb-1m: A dataset and benchmark for large-scale face recognition. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 87-102. Springer, 2016. 5", + "[22] Zeyu Han, Chao Gao, Jinyang Liu, Jeff Zhang, and Sai Qian Zhang. Parameter-efficient fine-tuning for large models: A comprehensive survey. arXiv preprint arXiv:2403.14608, 2024. 1", + "[23] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 4", + "[24] Zhehao Huang, Xinwen Cheng, JingHao Zheng, Haoran Wang, Zhengbao He, Tao Li, and Xiaolin Huang. Unified" + ], + "bbox": [ + 516, + 92, + 906, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1965", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "gradient-based machine unlearning with remain geometry enhancement. arXiv preprint arXiv:2409.19732, 2024. 1", + "[25] Joel Jang, Dongkeun Yoon, Sohee Yang, Sungmin Cha, Moontae Lee, Lajanugen Logeswaran, and Minjoon Seo. Knowledge unlearning for mitigating privacy risks in language models. arXiv preprint arXiv:2210.01504, 2022. 1", + "[26] Jinghan Jia, Jiancheng Liu, Parikshit Ram, Yuguang Yao, Gaowen Liu, Yang Liu, Pranay Sharma, and Sijia Liu. Model sparsity can simplify machine unlearning. Advances in Neural Information Processing Systems, 36:51584-51605, 2023. 1, 3", + "[27] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 3, 6", + "[28] Vishnuprasadh Kumaravelu, Sayanta Adhikari, and PK Srijith. Uncle: An unlearning framework for continual learning. 1, 3", + "[29] Meghdad Kurmanji, Peter Triantafillou, Jamie Hayes, and Eleni Triantafillou. Towards unbounded machine unlearning. Advances in neural information processing systems, 36: 1957-1987, 2023. 2, 3, 6", + "[30] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 6", + "[31] Bo Liu, Qiang Liu, and Peter Stone. Continual learning and private unlearning. In Conference on Lifelong Learning Agents, pages 243-254. PMLR, 2022. 1, 3", + "[32] Gaoyang Liu, Xiaogiang Ma, Yang Yang, Chen Wang, and Jiangchuan Liu. Federaser: Enabling efficient client-level data removal from federated learning models. In 2021 IEEE/ACM 29th International Symposium on Quality of Service (IWQOS), pages 1-10. IEEE, 2021. 2", + "[33] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5", + "[34] Laiqiao Qin, Tianqing Zhu, Linlin Wang, and Wanlei Zhou. Machine unlearning on pre-trained models by residual feature alignment using lora. arXiv preprint arXiv:2411.08443, 2024. 1, 2, 3", + "[35] Sadia Qureshi, Thanveer Shaik, Xiaohui Tao, Haoran Xie, Lin Li, Jianming Yong, and Xiaohua Jia. Exploring incremental unlearning: Techniques, challenges, and future directions. arXiv preprint arXiv:2502.16708, 2025. 1, 3", + "[36] Protection Regulation. General data protection regulation. Intouch, 25:1-5, 2018. 1", + "[37] Jeffrey Rosen. The right to be forgotten. *Stan. L. Rev. Online*, 64:88, 2011. 1", + "[38] Dan Roth and Wen-tau Yih. Probabilistic reasoning for entity & relation recognition. In COLING 2002: The 19th International Conference on Computational Linguistics, 2002. 3" + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 5", + "[40] Gobinda Saha, Isha Garg, and Kaushik Roy. Gradient projection memory for continual learning. arXiv preprint arXiv:2103.09762, 2021. 2, 5", + "[41] Thanveer Shaik, Xiaohui Tao, Haoran Xie, Lin Li, Xiaofeng Zhu, and Qing Li. Exploring the landscape of machine unlearning: A comprehensive survey and taxonomy. IEEE Transactions on Neural Networks and Learning Systems, 2024. 1", + "[42] Yingdan Shi and Ren Wang. Redefining machine unlearning: A conformal prediction-motivated approach. arXiv preprint arXiv:2501.19403, 2025. 1", + "[43] Takashi Shibata, Go Irie, Daiki Ikami, and Yu Mitsuzumi. Learning with selective forgetting. In *IJCAI*, page 6, 2021. 3, 5", + "[44] Mihai Surdeanu and Heng Ji. Overview of the english slot filling track at the tac2014 knowledge base population evaluation. In Proc. Text Analysis Conference (TAC2014), 2014. 3", + "[45] Jiahang Tu, Qian Feng, Chufan Chen, Jiahua Dong, Hanbin Zhao, Chao Zhang, and Hui Qian. Ce-sdwv: Effective and efficient concept erasure for text-to-image diffusion models via a semantic-driven word vocabulary. arXiv preprint arXiv:2501.15562, 2025. 1, 2, 3, 4", + "[46] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9446-9454, 2018. 1, 2, 7", + "[47] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5", + "[48] Cheng-Long Wang, Qi Li, Zihang Xiang, and Di Wang. Has approximate machine unlearning been evaluated properly? from auditing to side effects. arXiv e-prints, pages arXiv-2403, 2024. 1, 2, 3", + "[49] Heng Xu, Tianqing Zhu, Lefeng Zhang, Wanlei Zhou, and Philip S Yu. Machine unlearning: A survey. ACM Computing Surveys, 56(1):1-36, 2023. 2", + "[50] Jie Xu, Zihan Wu, Cong Wang, and Xiaohua Jia. Machine unlearning: Solutions and challenges. IEEE Transactions on Emerging Topics in Computational Intelligence, 2024. 1", + "[51] Lingling Xu, Haoran Xie, Si-Zhao Joe Qin, Xiaohui Tao, and Fu Lee Wang. Parameter-efficient fine-tuning methods for pretrained language models: A critical review and assessment. arXiv preprint arXiv:2312.12148, 2023. 1", + "[52] Haonan Yan, Xiaoguang Li, Ziyao Guo, Hui Li, Fenghua Li, and Xiaodong Lin. Arcane: An efficient architecture for exact machine unlearning. In *IJCAI*, page 19, 2022. 2, 3", + "[53] Jingwen Ye, Yifang Fu, Jie Song, Xingyi Yang, Songhua Liu, Xin Jin, Mingli Song, and Xinchao Wang. Learning with recoverable forgetting. In ECCV, 2022. 6", + "[54] Dong Yi, Zhen Lei, Shengcai Liao, and Stan Z Li. Learning face representation from scratch. arXiv preprint arXiv:1411.7923, 2014. 5" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1966", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[55] Jaehong Yoon, Shoubin Yu, Vaidehi Patil, Huaxiu Yao, and Mohit Bansal. Safree: Training-free and adaptive guard for safe text-to-image and video generation. arXiv preprint arXiv:2410.12761, 2024. 2, 4", + "[56] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. Advances in Neural Information Processing Systems, 33:5824-5836, 2020. 2, 5", + "[57] Dan Zhang, Tao Feng, Lilong Xue, Yuandong Wang, Yuxiao Dong, and Jie Tang. Parameter-efficient fine-tuning for foundation models. arXiv preprint arXiv:2501.13787, 2025. 1", + "[58] Haibo Zhang, Toru Nakamura, Takamasa Isohara, and Kouichi Sakurai. A review on machine unlearning. SN Computer Science, 4(4):337, 2023. 2", + "[59] Yuanhan Zhang, Zhenfei Yin, Jing Shao, and Ziwei Liu. Benchmarking omni-vision representation through the lens of visual realms. In European Conference on Computer Vision, pages 594-611. Springer, 2022. 5", + "[60] Hongbo Zhao, Bolin Ni, Junsong Fan, Yuxi Wang, Yuntao Chen, Gaofeng Meng, and Zhaoxiang Zhang. Continual forgetting for pre-trained vision models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28631-28642, 2024. 1, 3, 5, 6", + "[61] Hongbo Zhao, Fei Zhu, Bolin Ni, Feng Zhu, Gaofeng Meng, and Zhaoxiang Zhang. Practical continual forgetting for pretrained vision models. arXiv preprint arXiv:2501.09705, 2025. 1, 2, 3, 6", + "[62] Zhen Zhao, Zhizhong Zhang, Xin Tan, Jun Liu, Yanyun Qu, Yuan Xie, and Lizhuang Ma. Rethinking gradient projection continual learning: Stability/plasticity feature space decoupling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3718-3727, 2023. 2, 5", + "[63] Yaoyao Zhong and Weihong Deng. Face transformer for recognition. arXiv preprint arXiv:2103.14803, 2021. 5", + "[64] Chen Zhu, Ankit Singh Rawat, Manzil Zaheer, Srinadh Bhojanapalli, Daliang Li, Felix Yu, and Sanjiv Kumar. Modifying memories in transformer models. arXiv preprint arXiv:2012.00363, 2020. 3" + ], + "bbox": [ + 91, + 92, + 480, + 656 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1967", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/FG-OrIU_ Towards Better Forgetting via Feature-Gradient Orthogonality for Incremental Unlearning/4a750062-438a-4c8c-b4db-9a6576c3de7a_model.json b/2025/FG-OrIU_ Towards Better Forgetting via Feature-Gradient Orthogonality for Incremental Unlearning/4a750062-438a-4c8c-b4db-9a6576c3de7a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..80f305e72a93884bc00922bb8846a4c25aed802e --- /dev/null +++ b/2025/FG-OrIU_ Towards Better Forgetting via Feature-Gradient Orthogonality for Incremental Unlearning/4a750062-438a-4c8c-b4db-9a6576c3de7a_model.json @@ -0,0 +1,2389 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.13, + 0.89, + 0.177 + ], + "angle": 0, + "content": "FG-OrIU: Towards Better Forgetting via Feature-Gradient Orthogonality for Incremental Unlearning" + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.203, + 0.847, + 0.223 + ], + "angle": 0, + "content": "Qian Feng\\(^{1}\\), JiaHang Tu\\(^{1}\\), Mintong Kang\\(^{2}\\), Hanbin Zhao\\(^{1*}\\), Chao Zhang\\(^{1}\\), Hui Qian\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.186, + 0.223, + 0.807, + 0.258 + ], + "angle": 0, + "content": "1College of Computer Science and Technology, Zhejiang University \n2Department of Computer Science University of Illinois at Urbana-Champaign" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.324, + 0.486, + 0.687 + ], + "angle": 0, + "content": "Incremental unlearning (IU) is critical for pre-trained models to comply with sequential data deletion requests, yet existing methods primarily suppress parameters or confuse knowledge without explicit constraints on both feature and gradient level, resulting in superficial forgetting where residual information remains recoverable. This incomplete forgetting risks security breaches and disrupts retention balance, especially in IU scenarios. We propose FG-OrIU (Feature-Gradient Orthogonality for Incremental Unlearning), the first framework unifying orthogonal constraints on both features and gradients level to achieve deep forgetting, where the forgetting effect is irreversible. FG-OrIU decomposes feature spaces via Singular Value Decomposition (SVD), separating forgetting and remaining class features into distinct subspaces. It then enforces dual constraints: feature orthogonal projection on both forgetting and remaining classes, while gradient orthogonal projection prevents the reintroduction of forgotten knowledge and disruption to remaining classes during updates. Additionally, dynamic subspace adaptation merges newly forgetting subspaces and contracts remaining subspaces, ensuring a stable balance between removal and retention across sequential unlearning tasks. Extensive experiments demonstrate the effectiveness of our method." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.716, + 0.222, + 0.731 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.741, + 0.484, + 0.878 + ], + "angle": 0, + "content": "In recent years, pre-training with fine-tuning [22, 51, 57] has dominated model development. However, large-scale datasets often contain harmful or private content, necessitating mechanisms to erase specific data traces. Machine Unlearning (MU) [4, 5, 7, 12, 16, 20, 24, 34, 41, 42, 50] addresses this by enabling pre-trained models (PTMs) to comply with deletion mandates like GDPR (General data protection regulation)'s \"right to be forgotten\" [36, 37]. While existing MU methods focus on static unlearning, for" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.291, + 0.906, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.453, + 0.907, + 0.509 + ], + "angle": 0, + "content": "Figure 1. Here, we present three examples from forgetting classes, with the reconstructed images generated using Deep Image Prior (DIP) [46], where the last block's features from the unlearned model serve as the training target." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.513, + 0.907, + 0.664 + ], + "angle": 0, + "content": "getting a single or small set of classes in one task, real-world scenarios require Incremental Unlearning (IU) [8-10, 17, 19, 25, 28, 31, 35, 48, 60, 61]. Regulatory shifts (e.g., evolving OpenAI safety policies\\(^{1}\\), EU AI Act\\(^{2}\\)) and dynamic content restrictions necessitate sequential unlearning of multiple classes/concepts without costly retraining. IU need not only erase targets in a single step but also preserve retained performance and ensure irreversible forgetting across tasks, making it critical for sustainable model governance in evolving ethical and legal landscapes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.665, + 0.909, + 0.862 + ], + "angle": 0, + "content": "IU methods broadly address two scenarios: (1) concept erasure in generative models, where sensitive concepts (e.g., nudity, violence) are iteratively removed from diffusion models [11, 45], and (2) instant- or class-level unlearning in vision classifiers, where PTMs forget specific classes (e.g., face identity, medical data) while retaining unrelated ones [26, 34]. In this work, we focus on the latter, specifically targeting class-level IU in vision PTMs. Our goal is threefold: (a) to ensure that the model completely forgets the specified classes, (b) to robustly preserve the performance of remaining classes, and (c) to maintain the effectiveness of previous forgetting across sequential unlearning procedures." + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.465, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author: Hanbin Zhao ." + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.875, + 0.68, + 0.888 + ], + "angle": 0, + "content": "1 https://openai.com/safety/" + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.888, + 0.757, + 0.901 + ], + "angle": 0, + "content": "\\(^{2}\\)https://artificialintelligenceact.eu/the-act/" + }, + { + "type": "list", + "bbox": [ + 0.532, + 0.875, + 0.757, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1957" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.092, + 0.092, + 0.482, + 0.62 + ], + "angle": 0, + "content": "Existing methods largely fall into two categories: parameter suppression, which selectively weakens target class-related parameters [15, 16, 52], and knowledge confusion, which distorts predictions via noisy labels or bad teacher distillation [12, 18, 29]. However, our empirical findings reveal that these methods lead to superficial forgetting, as the unlearned model still retains semantic information related to the forgetting classes, making the erased knowledge partially residual and recoverable. Specifically, reconstructing features from the last layer of the unlearned model using Deep Image Prior (DIP) [46] produces blurred yet recognizable images (as illustrated in Figure 1). Furthermore, fixing the backbone of the unlearned model and retraining only the classification head [10, 61], which recovers a portion of the original accuracy (as illustrated in Figure 3). The blurred image and the partial recovery of performance both indicate that existing methods merely degrade the features of forgetting classes rather than completely eliminating them. The degraded features still contain residual class-specific information, which leads to potential recovery of the forgetting effect. Besides, these degraded features not only enable recovery but also risk entanglement with the remaining class features [34]. Such entanglement destabilizes decision boundaries, degrading performance on the remaining classes, particularly in IU scenarios where sequential deletion requests amplify error propagation [17, 48]. To address these limitations and achieve deep forgetting, we argue that unlearning should explicitly constrain both feature representations and gradient updates. Since model updates rely on both forward computations (feature extraction) and backward optimization (gradient updates), it is critical to simultaneously suppress the features of forgetting classes and prevent their reintroduction during training updates. In Table 1, we summarize the differences between our methods and prior works." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.625, + 0.482, + 0.88 + ], + "angle": 0, + "content": "To enforce explicit constraints on both feature representations and gradient updates, we draw inspiration from orthogonal mechanisms in related domains. For feature-level constraints, recent work in text-to-image models [45, 55] demonstrates that orthogonalizing sensitive token embeddings effectively removes targeted concepts from diffusion outputs, suggesting geometric isolation as a viable strategy for feature suppression. For gradient-level constraints, inspired by [56], which reduces the mutual influence of different tasks in multi-task learning, and [40, 62], which mitigates the forgetting of old tasks in continual learning, we leverage the insight that updating in directions orthogonal to old/other task subspaces helps alleviate catastrophic forgetting. While these orthogonal principles show promise in adjacent fields like generative model editing and multitask/continual learning, their applicability to MU, particularly IU in vision PTMs, remains underexplored." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.885, + 0.482, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose FG-OrIU (Feature-Gradient" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.905, + 0.131 + ], + "angle": 0, + "content": "Table 1. Position, Complete, and Recover refers to, where the forgetting occurs, whether the forgetting is thorough, and whether the forgetting effect can be restored, respectively." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.135, + 0.903, + 0.21 + ], + "angle": 0, + "content": "
PositionCompleteRecover
superficial forgetting
Prior worksOnly Final OutputNoPartial
deep forgetting
OursFeature&GradientYesNo
" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.215, + 0.905, + 0.593 + ], + "angle": 0, + "content": "Orthogonality for Incremental Unlearning), a unified framework for IU in vision PTMs that enforces explicit constraints at both feature and gradient levels through three stages: feature subspace decomposition, dual orthogonal projection, and dynamic subspace adaptation. In feature subspace decomposition, we decompose the feature space of each layer into two subspaces via Singular Value Decomposition (SVD): the forgetting subspace \\( S_f \\), which captures the discriminative features of the classes to be removed, and the remaining subspace \\( S_r \\), which encodes the features of the remaining classes. In dual orthogonal projection, we freeze the backbone of PTM and introduce lightweight LoRA modules for efficient updates. Then, features of samples from forgetting classes are orthogonalized against \\( S_f \\) to eliminate correlations, while remaining classes features are aligned with \\( S_r \\) to minimize interference through minimizing the projection residual. Before backward optimization, gradients are first projected onto \\( S_f \\) to maximally perturb parameters critical for forgetting classes, then modified along the orthogonal complement of \\( S_r \\) to prevent contamination of retained features. In dynamic subspace adaptation, FOrIU dynamically expands \\( S_f \\) by merging new forgetting class subspaces and contracts \\( S_r \\) by removing overlapping features, ensuring cumulative forgetting without degrading retained performance." + }, + { + "type": "title", + "bbox": [ + 0.534, + 0.594, + 0.802, + 0.607 + ], + "angle": 0, + "content": "Our main contributions are threefold:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.611, + 0.905, + 0.656 + ], + "angle": 0, + "content": "- We reveal that existing IU methods achieve superficial forgetting rather than deep forgetting, as their partial feature suppression leaves residual information recoverable." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.657, + 0.905, + 0.717 + ], + "angle": 0, + "content": "- We propose the first framework unifying forgetting on both feature and gradient level via dual orthogonal projection, ensuring irreversible forgetting while preserving remaining class robustness." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.718, + 0.905, + 0.747 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate superior performance in MU and IU scenarios across multiple benchmarks." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.611, + 0.905, + 0.747 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.755, + 0.655, + 0.77 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.773, + 0.709, + 0.789 + ], + "angle": 0, + "content": "2.1. Machine Unlearning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Machine unlearning (MU) aims to remove the influence of specific data from pre-trained models (PTMs) without retraining from scratch [49, 58]. Existing MU methods are broadly categorized into exact and approximate unlearning. Exact MU [5, 32] ensures statistical indistinguishability from models retrained on remaining data but is computationally prohibitive for large PTMs. Approximate MU re" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1958" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.318 + ], + "angle": 0, + "content": "taxes this guarantee and is more practical. Current methods fall into two paradigms: (1) concept erasure in generative models [11, 45]; and (2) class-/sample-wise unlearning in classification models, which modifies pre-trained weights to forget target classes while preserving others [26, 34]. However, most studies focus on static unlearning, assuming a single-step forgetting task. Real-world applications, such as compliance with iterative AI safety policies (e.g., OpenAI's content guidelines) or evolving regulations like the EU AI Act, require incremental unlearning (IU) to process sequential deletion requests efficiently. Directly applying static methods to IU often leads to cumulative errors, model collapse, and incomplete forgetting [17, 48], underscoring the need for dedicated frameworks to ensure compliance in dynamic regulatory environments." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.32, + 0.314, + 0.336 + ], + "angle": 0, + "content": "2.2. Incremental Unlearning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.34, + 0.485, + 0.658 + ], + "angle": 0, + "content": "Existing MU methods, like parameter suppression [15, 16, 52] and knowledge confusion methods [10, 12, 17, 18, 29, 35, 61], lead to superficial forgetting due to the absence of direct constraints on the feature representation. In particular, the degraded features trap them in a dilemma between forgetting and remaining, which becomes even more severe when directly applied to IU. Meanwhile, some continual learning methods [6, 27], originally designed to retain old knowledge while dynamically learning new knowledge, also face inherent limitations in IU [35]. Previous work focusing on IU has primarily relied on toy models, e.g., linear models [13], or imposed heavy constraints, such as adding a class-specific synthetic signal in the pre-training stage [43], which is not feasible for PTMs. Moreover, while some focus on dynamically performing both continual learning and selective forgetting within learned knowledge [8, 28, 31], our approach differs in that we emphasize achieving deep forgetting when handling sequential unlearning tasks. That is, rather than alternating between learning and forgetting, the pre-trained model should completely forget unwanted knowledge while efficiently retaining the rest." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.66, + 0.279, + 0.675 + ], + "angle": 0, + "content": "3. Problem Statement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.681, + 0.483, + 0.756 + ], + "angle": 0, + "content": "Following [60, 61], we first formally define Machine Unlearning (MU), i.e., static unlearning scenario. We then extend this definition to Incremental Unlearning (IU), i.e., dynamic unlearning scenario, where deletion requests arrive sequentially." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.757, + 0.484, + 0.879 + ], + "angle": 0, + "content": "Let \\(\\mathcal{M}\\) be a pre-trained model trained on the dataset \\(\\mathbf{D}\\). We denote the mapping relationship of the model as \\(f_{\\mathcal{M}}(\\mathcal{X}_{\\mathbf{D}})\\rightarrow \\mathcal{Y}_{\\mathbf{D}}\\), where \\(\\mathcal{X}_{\\mathbf{D}}\\) and \\(\\mathcal{Y}_{\\mathbf{D}}\\) represent the input set and output set, respectively. For MU, we partition \\(\\mathbf{D}\\) into \\(\\mathbf{D}_{\\mathbf{f}}\\) (forgetting classes) and \\(\\mathbf{D}_{\\mathbf{r}}\\) (remaining classes), under the constraint \\(|\\mathbf{D}_{\\mathbf{r}}| + |\\mathbf{D}_{\\mathbf{f}}|\\ll |\\mathbf{D}|\\). The forgetting algorithm \\(\\mathcal{F}\\) produces \\(\\mathcal{M}' = \\mathcal{F}(\\mathcal{M},\\mathbf{D}_{\\mathbf{f}},\\mathbf{D}_{\\mathbf{r}})\\), altering the following mappings:" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.885, + 0.483, + 0.904 + ], + "angle": 0, + "content": "\\[\nf _ {\\mathcal {M} ^ {\\prime}} \\left(\\mathcal {X} _ {\\mathbf {D} _ {\\mathrm {f}}}\\right) \\not \\longrightarrow \\mathcal {Y} _ {\\mathbf {D} _ {\\mathrm {f}}}, f _ {\\mathcal {M} ^ {\\prime}} \\left(\\mathcal {X} _ {\\mathbf {D} _ {\\mathrm {r}}}\\right) \\longrightarrow \\mathcal {Y} _ {\\mathbf {D} _ {\\mathrm {r}}}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.108 + ], + "angle": 0, + "content": "Here, \\(\\nrightarrow\\) means the mapping relationship no longer holds." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.108, + 0.905, + 0.214 + ], + "angle": 0, + "content": "For IU, let there are a sequence of unlearning task, \\(\\{\\mathcal{T}_1,\\mathcal{T}_2,\\dots ,\\mathcal{T}_T\\}\\) and \\(T\\) is the number of unlearning tasks. And \\(\\{\\mathbf{D}_{f,t},\\mathbf{D}_{r,t}\\}\\) is the forgetting classes and remaining classes belonging to \\(\\mathcal{T}_t\\) where \\(\\mathbf{D}_{r,t} = \\mathbf{D} - \\cup_{i = 1}^{t}\\mathbf{D}_{f,i}\\). Starting from \\(\\mathcal{M}_0 = \\mathcal{M}\\), the forgetting algorithm \\(\\mathcal{F}\\) iteratively processes till \\(t\\)-th unlearning task \\(\\mathcal{T}_t\\) to generate \\(\\mathcal{M}_t = \\mathcal{F}(\\mathcal{M}_{t - 1},\\mathbf{D}_{f,t},\\mathbf{D}_{r,t})\\) which holds the followings:" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.226, + 0.905, + 0.244 + ], + "angle": 0, + "content": "\\[\nf _ {\\mathcal {M} _ {t}} \\left(\\mathcal {X} _ {\\mathbf {D} _ {f, i}}\\right) \\not \\longrightarrow \\mathcal {Y} _ {\\mathbf {D} _ {f, i}}, f _ {\\mathcal {M} _ {t}} \\left(\\mathcal {X} _ {\\mathbf {D} _ {r, j}}\\right) \\longrightarrow \\mathcal {Y} _ {\\mathbf {D} _ {r, j}}, \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.259, + 0.613, + 0.274 + ], + "angle": 0, + "content": "4. Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.278, + 0.907, + 0.383 + ], + "angle": 0, + "content": "In this section, we introduce FG-OrIU, a comprehensive framework for Incremental Unlearning (IU) in pre-trained models (PTM). By imposing orthogonal constraints on both feature representations and gradient updates, FG-OrIU ensures thorough removal of unwanted knowledge while preserving performance on remaining classes. An overview of the framework is shown in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.387, + 0.805, + 0.403 + ], + "angle": 0, + "content": "4.1. Feature Subspace Decomposition" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.403, + 0.906, + 0.658 + ], + "angle": 0, + "content": "PTMs acquire implicit knowledge from pre-training data, benefiting downstream tasks during inference. Unlearning aims to selectively remove the influence of forgetting classes from the PTM [10]. Unlike explicit knowledge in traditional Knowledge Bases (KBs) [38, 44], which can be directly edited, the knowledge in PTMs is implicit and intertwined across all parameters [10, 34, 64]. Effective unlearning requires precisely identifying feature spaces for forgetting and remaining. To achieve this, we decompose features into two subspaces: one capturing forgetting-class contributions and the other preserving remaining-class knowledge. Specifically, we perform singular value decomposition (SVD) on the feature matrix of forgetting classes and use the span of its dominant singular vectors to define the forgetting subspace. Similarly, we apply the same procedure to the remaining classes to construct its subspace. We then detail the process as follows." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.877 + ], + "angle": 0, + "content": "Given a PTM, \\(\\mathcal{M}\\), we construct a representation matrix \\(\\pmb{R}_{f,i} = \\left[\\pmb{x}_{1,i},\\dots,\\pmb{x}_{n_{f_i},i}\\right]\\in \\mathbb{R}^{n_{f_i}\\times d}\\) at each layer \\(l\\) where columns correspond to representations of \\(n_{f_i}\\) samples from the remaining classes \\(\\mathbf{D}_{\\mathbf{f},\\mathbf{i}}\\), and \\(d\\) is the output dimension3. SVD is applied to \\(\\pmb{R}_{f,i} = \\pmb{U}_{f,i}\\pmb{\\Sigma}_{f,i}(\\pmb{V}_{f,i})^T\\), followed by its \\(k\\)-rank approximation \\((\\pmb{R}_{f,i})_k\\) satisfying \\(||(\\pmb{R}_{f,i})_k||_F^2\\geq \\epsilon ||\\pmb{R}_{f,i}||_F^2\\) for a threshold \\(\\epsilon\\). The subspace \\(\\mathcal{S}_{f,i} = \\mathrm{span}\\{\\pmb{B}_{f,i}\\}\\) is defining using the first \\(k\\) vectors \\(\\pmb{B}_{f,i} = \\{\\pmb{u}_{1,i},\\dots,\\pmb{u}_{k,i}\\}\\) from \\(\\pmb{U}_{f,i}\\), capturing directions with the highest singular values. Similarly, the feature matrix \\(\\pmb{R}_{r,i}\\) for remaining classes \\(\\mathbf{D}_{\\mathbf{r},\\mathbf{i}}\\) defines the subspace \\(\\mathcal{S}_{r,i} = \\mathrm{span}\\{\\pmb{B}_{r,i}\\}\\). The two sets of layer-wised feature subspace, \\(\\{\\mathcal{S}_{f,i}^1,\\dots,\\mathcal{S}_{f,i}^l\\}\\) and \\(\\{\\mathcal{S}_{r,i}^1,\\dots,\\mathcal{S}_{r,i}^l\\}\\), represent the core part for forgetting and remaining classes," + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.887, + 0.766, + 0.901 + ], + "angle": 0, + "content": "3Here, we omit the notation \\(l\\) for simplicity." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1959" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.089, + 0.9, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.3, + 0.907, + 0.343 + ], + "angle": 0, + "content": "Figure 2. Overview of FG-OrIU: (a) For each unlearning task, we decompose the layer-wised- forgetting and remaining subspace separately (Sec. 4.1). (b) LoRA is inserted at each layer, with constraints applied at both the feature and gradient levels for forgetting and remaining (Sec. 4.2). (c) For new unlearning tasks, the forgetting subspace expands while the remaining contracts (Sec. 4.3)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.347, + 0.484, + 0.483 + ], + "angle": 0, + "content": "respectively, isolating features to be erased while preserving those to retain. By explicitly decomposing feature subspaces across all layers, we strengthen supervision signals for parameter updates and mitigate the risk of gradient obfuscation [2] in deep networks, where gradient inaccuracies caused by layer-wise nonlinear transformations could otherwise weaken the unlearning effect. This multi-layer constraint ensures consistent suppression of forgetting class features throughout the model hierarchy." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.483, + 0.342, + 0.5 + ], + "angle": 0, + "content": "4.2. Dual Orthogonal Projection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.505, + 0.484, + 0.621 + ], + "angle": 0, + "content": "Prior to unlearning task \\(\\mathcal{T}_i\\), we freeze both the backbone and the classification head of the PTM and insert lightweight LoRA modules [23] at each layer. We then denote the model with LoRA inserted as \\(\\mathcal{M}_{\\theta}\\), where \\(\\theta\\) represents the trainable parameters in LoRA. After unlearning, the updates in LoRA, \\(\\theta^{*}\\), will be merged into the initial weight of the backbone, \\(\\mathcal{M}_i \\xrightarrow[\\text{merge}]{\\theta^*} \\mathcal{M}_{i+1}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.621, + 0.484, + 0.68 + ], + "angle": 0, + "content": "Building on the feature subspaces \\( S_f \\) and \\( S_r \\) defined in Section 4.1, we propose a dual orthogonal projection mechanism that operates through both forward feature alignment and backward gradient correction." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.686, + 0.353, + 0.701 + ], + "angle": 0, + "content": "4.2.1. Feature Orthogonal Projection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Inspired by [45, 55], for forgetting classes, achieving thorough removal at the feature level requires that the features generated by the unlearned model, \\(\\mathcal{M}_{\\theta}\\), deviate as much as possible from the feature subspace \\(S_{f}\\). Meanwhile, we aim to ensure that the unlearned model's feature subspace for remaining classes aligns with \\(S_{r}\\). The orthogonality constraint between \\(S_{f}\\) and the unlearned features serves as a critical mechanism for facilitating class-specific forgetting: under a fixed classification head, features orthogonal to the original subspace \\(S_{f}\\) induce maximal perturbations in class logits, as the inner product between orthogonal vectors diminishes to zero. This forces the unlearned model to suppress discriminative patterns associated with forgetting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.347, + 0.907, + 0.437 + ], + "angle": 0, + "content": "classes in its intermediate representations. To do so, we propose the following two regularization constraints. For forgetting classes, let \\( \\boldsymbol{x}_f \\) denote a batch of forgetting samples from \\( \\mathbf{D}_{\\mathbf{f},\\mathrm{i}} \\), and \\( h_f^l \\) denote the \\( l \\)-th layer features gained with the unlearned model \\( \\mathcal{M}_{\\theta} \\). For complete forgetting, we enforce:" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.462, + 0.907, + 0.484 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {h} _ {f} ^ {l} \\perp \\mathcal {S} _ {f, i} ^ {l} \\quad \\Rightarrow \\quad \\min | | \\mathbf {P} _ {f} ^ {l} \\boldsymbol {h} _ {f} ^ {l} | | _ {F} ^ {2} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.494, + 0.906, + 0.59 + ], + "angle": 0, + "content": "where \\(\\mathbf{P}_f^l = \\mathbf{B}_{f,i}^l (\\mathbf{B}_{f,i}^l)^\\top\\) is the projection matrix for the forgetting subspace \\(S_{f,i}^{l}\\) and \\(\\| \\cdot \\| _F^2\\) is the F-norm. Eq.(3) ensures that the features of \\(\\mathbf{D}_{\\mathbf{f},\\mathbf{i}}\\) gained from the unlearned model \\(\\mathcal{M}_{\\theta}\\) lie outside the original feature subspace gained from \\(\\mathcal{M}\\), erasing the unwanted information. Thus, we define the flop-f loss:" + }, + { + "type": "equation", + "bbox": [ + 0.62, + 0.6, + 0.906, + 0.639 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {f o p - f} = \\sum_ {l = 1} ^ {m} \\left\\| \\mathbf {P} _ {f} ^ {l} \\boldsymbol {h} _ {f} ^ {l} \\right\\| _ {F} ^ {2}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.649, + 0.906, + 0.71 + ], + "angle": 0, + "content": "where \\( m \\) denotes the total number of layers in the PTM. For remaining classes, we follow previous works and apply the standard cross-entropy loss by constraining the model's final output, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.721, + 0.905, + 0.739 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c e} = \\mathcal {L} \\left(\\mathcal {X} _ {\\mathbf {D} _ {r, 1}}, \\mathcal {Y} _ {\\mathbf {D} _ {r, 1}}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.749, + 0.907, + 0.855 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}\\) denotes the cross-entropy loss function. In addition, we also propose applying explicit constraint at the feature level for the remaining classes to better preserve the desired information. Specifically, \\(\\pmb{x}_r\\) denote a batch of remaining samples from \\(\\mathbf{D}_{\\mathbf{r},\\mathbf{i}}\\), and \\(h_r^l\\) denote the \\(l\\)-th layer features gained with the unlearned model \\(\\mathcal{M}_{\\theta}\\). For better preservation, we maintain feature fidelity through:" + }, + { + "type": "equation", + "bbox": [ + 0.64, + 0.879, + 0.905, + 0.898 + ], + "angle": 0, + "content": "\\[\n\\min \\left\\| \\boldsymbol {h} _ {r} ^ {l} - \\mathbf {P} _ {r} ^ {l} \\boldsymbol {h} _ {r} ^ {l} \\right\\| _ {F} ^ {2} \\tag {6}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1960" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.089, + 0.483, + 0.125 + ], + "angle": 0, + "content": "where \\(\\mathbf{P}_r^l = \\mathbf{B}_{r,i}^l (\\mathbf{B}_{r,i}^l)^\\top\\) is the projection matrix for the remaining feature subspace \\(S_{r,i}^{l}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.125, + 0.483, + 0.186 + ], + "angle": 0, + "content": "Eq.(6) ensures that the features of \\(\\mathbf{D}_{\\mathbf{r},\\mathbf{i}}\\) gained from the unlearned model \\(\\mathcal{M}_{\\theta}\\), align with the original feature subspace gained from \\(\\mathcal{M}\\), maintaining the better preservation. Therefore, We draw the following \\(f_{op - r}\\) loss:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.192, + 0.483, + 0.232 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {f o p - r} = \\sum_ {l = 1} ^ {m} \\left\\| \\boldsymbol {h} _ {r} ^ {l} - \\mathbf {P} _ {r} ^ {l} \\boldsymbol {h} _ {r} ^ {l} \\right\\| _ {F} ^ {2}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.238, + 0.274, + 0.253 + ], + "angle": 0, + "content": "Finally, the total loss is:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.261, + 0.483, + 0.278 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {c e} + \\lambda_ {1} \\mathcal {L} _ {f o p - f} + \\lambda_ {2} \\mathcal {L} _ {f o p - r}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.285, + 0.414, + 0.301 + ], + "angle": 0, + "content": "where \\(\\lambda_{1},\\lambda_{2}\\) are the coefficients during training." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.306, + 0.363, + 0.321 + ], + "angle": 0, + "content": "4.2.2. Gradient Orthogonal Projection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.325, + 0.484, + 0.415 + ], + "angle": 0, + "content": "To prevent parameter updates from reintroducing forgotten knowledge while preserving retained features, we design a gradient correction mechanism inspired by gradient projection techniques in multi-task learning [56] and continual learning [40, 62]. Given the raw gradient \\( g_{l} \\) at layer \\( l \\), we apply the following modifications:" + }, + { + "type": "equation", + "bbox": [ + 0.127, + 0.422, + 0.483, + 0.485 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\hat {g} _ {l} = g _ {l} - g _ {l} ^ {\\prime} - g _ {l} ^ {\\prime \\prime} \\tag {9} \\\\ = \\boldsymbol {g} _ {l} - \\underbrace {\\boldsymbol {g} _ {l} (\\boldsymbol {I} - \\mathbf {B} _ {f , i} ^ {l} (\\mathbf {B} _ {f , i} ^ {l}) ^ {\\top})} _ {(t e r m 1: g o p - f)} - \\underbrace {\\boldsymbol {g} _ {l} \\mathbf {B} _ {r , i} ^ {l} (\\mathbf {B} _ {r , i} ^ {l}) ^ {\\top}} _ {(t e r m 2: g o p - r)}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.484, + 0.568 + ], + "angle": 0, + "content": "Term (gop-f), \\( g_{l}^{\\prime} \\): Removing the components of \\( g_{l} \\) that are orthogonal to \\( S_{f,i}^{l} \\) not only amplifies updates focusing on directions most critical for encoding forgetting class features, but also blocks parameter updates that could reintroduce removed knowledge." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.484, + 0.659 + ], + "angle": 0, + "content": "Term (gop-r), \\( g_{l}^{\\prime \\prime} \\): We aim to ensure that forgetting specific classes does not affect the remaining classes. Formally, for the unlearned model \\( f_{\\mathcal{M}_{\\theta}} \\) and \\( x_{r} \\) from the remaining classes \\( \\mathbf{D}_{\\mathbf{r},\\mathbf{i}} \\), the following equation holds: \\( f_{\\mathcal{M}_{\\theta}}(x_{r}) = f_{\\mathcal{M}}(x_{r}) \\). Specifically, we consider the input and output of layer \\( l \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.667, + 0.482, + 0.702 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {W} _ {l} \\boldsymbol {x} _ {r} + \\Delta \\theta_ {l} \\boldsymbol {x} _ {r} = \\boldsymbol {W} _ {l} \\boldsymbol {x} _ {r} \\\\ \\Delta \\theta_ {l} \\boldsymbol {x} _ {\\boldsymbol {r}} = \\mathbf {0} \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.709, + 0.483, + 0.77 + ], + "angle": 0, + "content": "where \\(\\Delta \\theta_{l}\\) represents the change in the LoRA parameters inserted at layer \\(l\\), which corresponds to its gradient. \\(W_{l}\\) is the parameter at layer \\(l\\) from the pre-trained model, and these parameters remain frozen during unlearning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.77, + 0.483, + 0.876 + ], + "angle": 0, + "content": "Eq.(10) indicates that if the update direction of the parameters, i.e., the gradient of the trainable LoRA parameters, lies in the complementary space of \\( S_{r,i}^{l} \\), i.e., \\( \\pmb{g}_{l} \\perp S_{r,i}^{l} \\), then \\( \\Delta \\theta \\pmb{x}_r = \\eta \\pmb{g}_l \\pmb{x}_r = \\mathbf{0} \\) (where \\( \\eta \\) is the learning rate), ensuring that the unlearned model's performance on the remaining classes is not affected. Therefore, we further refine the gradient by removing its components along \\( S_{r,i}^{l} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.876, + 0.483, + 0.907 + ], + "angle": 0, + "content": "Finally, modified gradient \\(\\hat{g}_l\\) serves the real gradient for parameter updates at the layer \\(l\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.091, + 0.789, + 0.107 + ], + "angle": 0, + "content": "4.3. Dynamic Subspace Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.107, + 0.905, + 0.196 + ], + "angle": 0, + "content": "When encountering a new unlearning task \\(\\mathcal{T}_{i + 1}\\), our method only needs to update the two decomposed feature subspaces. Since the forgetting classes \\(\\mathbf{D_{f,i + 1}}\\) in \\(\\mathcal{T}_{i + 1}\\) are a subset of the remaining classes from the previous task \\(\\mathcal{T}_i\\), i.e., \\(\\mathbf{D_{f,i + 1}}\\subseteq \\mathbf{D_{r,i}}\\), the feature subspaces can be efficiently adjusted. For simplicity, we omit the notion \\(l\\) of layers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.197, + 0.905, + 0.272 + ], + "angle": 0, + "content": "We first construct a representation matrix \\( \\mathbf{R}_{f,i+1} \\) using only the forgetting classes \\( \\mathbf{D}_{\\mathbf{f},i+1} \\). Before applying SVD and \\( k \\)-rank approximation, we remove common bases already present in \\( S_{f,i} \\) to ensure newly added bases are unique and orthogonal to existing ones:" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.277, + 0.905, + 0.321 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\hat {\\boldsymbol {R}} _ {f, i + 1} = \\boldsymbol {R} _ {f, i + 1} - \\boldsymbol {B} _ {f, i} \\left(\\boldsymbol {B} _ {f, i}\\right) ^ {T} \\left(\\boldsymbol {R} _ {f, i + 1}\\right) (11) \\\\ = \\boldsymbol {R} _ {f, i + 1} - \\boldsymbol {R} _ {f, i + 1} ^ {\\text {p r o j}}. (12) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.329, + 0.906, + 0.456 + ], + "angle": 0, + "content": "Next, we perform SVD on \\(\\hat{R}_{f,i+1} = \\hat{U}_{f,i+1}\\hat{\\Sigma}_{f,i+1}(\\hat{V}_{f,i+1})^T\\) and select \\(h\\) new orthogonal from \\(\\hat{U}_{f,i+1}\\). Then, the feature subspace of forgetting classes is expanded as \\(\\mathcal{S}_{f,i+1} = \\text{span}\\{B_{f,i+1}\\}\\), where \\(\\mathbf{B}_{f,i+1} = [B_{f,i};\\mathbf{u}_1,\\dots,\\mathbf{u}_h]\\) with \\(h\\) new bases from \\(\\hat{U}_{f,i+1}\\). And the feature subspace \\(\\mathcal{S}_{r,i+1}\\) of remaining classes is contracted to remove the overlapping features accordingly." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.461, + 0.64, + 0.477 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.479, + 0.704, + 0.495 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.5, + 0.848, + 0.514 + ], + "angle": 0, + "content": "Datasets, Benchmarks and Pre-trained Models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.515, + 0.906, + 0.711 + ], + "angle": 0, + "content": "Following [60], we evaluate FG-OrIU on Classification and Recognition tasks. For classification, we use ImageNet100 [39], CUB-200 [47], and Omnibenchmark [59], with a ViT-B/16 [14] pre-trained in PyTorch [33]. For face recognition, we use CASIA-Face100 [60], a subset of CASIA-WebFace [54], and MS-Celeb-100, sampled from [21]. The pre-trained model is the Face Transformer from [63]. We focus on class-unlearning scenarios with two benchmarks: (a) MU, varying numbers of classes are removed; (b) IU, 20 classes are removed per task over 4 tasks. All setting is in the form of Y-X, which means experiments start from a PTM (Y classes originally) and forget X classes. More details about implementations are in Appendix ??" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.711, + 0.905, + 0.809 + ], + "angle": 0, + "content": "Metrics. Following [60], we evaluate performance using average accuracy (Acc) for both forgetting and remaining classes. The accuracy for forgetting classes should approach zero, while the accuracy for remaining classes should remain close to the original model's performance. Additionally, we adopt the H-Mean [43] metric to assess overall performance after learning task \\(\\mathcal{T}_t\\), computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.825, + 0.905, + 0.861 + ], + "angle": 0, + "content": "\\[\nH - M e a n ^ {(t)} = \\frac {2 A c c _ {r} ^ {(t)} \\cdot D r o p ^ {(t)}}{A c c _ {r} ^ {(t)} + D r o p ^ {(t)}}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.867, + 0.905, + 0.904 + ], + "angle": 0, + "content": "Here \\( Acc_{r}^{(t)} \\) represents accuracy on the retained dataset after task \\( \\mathcal{T}_t \\), while \\( Drop^{(t)} = Acc_{f}^{(t - 1)} - Acc_{f}^{(t)} \\) mea" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1961" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.082, + 0.908, + 0.337 + ], + "angle": 0, + "content": "
MethodsTunable Ratio ↓100-5100-10100-50100-90
H ↑Accr ↑Accf ↓H ↑Accr ↑Accf ↓H ↑Accr ↑Accf ↓H ↑Accr ↑Accf ↓
Pre-train--73.8570.88-73.8172.74-72.4574.88-72.3273.86
L2*99.73%67.1366.963.5867.7464.871.8664.3956.610.2452.4540.680.04
EWC*99.73%69.6568.690.2469.1665.920.0062.4853.610.0044.4131.750.00
MAS*99.73%69.7369.290.7269.3866.410.1262.6753.880.0046.7934.240.00
LwF99.73%67.9568.550.0070.0867.620.0063.8155.590.0061.2552.320.00
DER99.73%69.7068.550.0070.0867.620.0063.8155.590.0057.0346.440.00
DER++99.73%69.7068.560.0070.5868.550.0064.6156.820.0061.4052.540.00
FDR99.73%70.3169.740.0070.0367.510.0065.4058.040.0053.8542.370.00
Retrain100.00%16.499.330.0018.0210.280.0019.7111.350.0046.4733.900.00
SCRUB99.73%67.7864.940.0068.2664.310.0065.8258.710.0016.119.040.00
SCRUB-S99.73%70.2969.950.2471.1969.810.1257.4951.209.3417.539.940.00
LIRF*50.66%25.5667.6755.1326.3565.8356.2647.1358.9535.6254.4944.293.06
GS-LoRA1.28%71.0271.160.0071.7670.810.0071.2968.050.0273.7173.560.00
GS-LoRA++1.28%71.1271.360.0072.0471.350.0071.5668.520.0072.8571.860.00
FG-OrIU1.28%71.7872.710.0072.5272.310.0074.1873.510.0075.4977.150.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.344, + 0.908, + 0.388 + ], + "angle": 0, + "content": "Table 2. Static Machine Unlearning results for face recognition task on CASIA-Face100. Accr (%) and Accf (%) are the accuracies of remaining and forgetting classes. * denotes the original methods with a rehearsal buffer. Retrain represents retraining the model using replay data and the training epoch is the same as other methods to ensure a fair comparison. Pre-train denotes the results before forgetting." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.397, + 0.476, + 0.581 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.582, + 0.483, + 0.625 + ], + "angle": 0, + "content": "Figure 3. Accuracy on forgetting classes when recovering on CASIA-Face100. The line (Pre-train) is the result before unlearning. The line (FG-OrIU) is Our method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.483, + 0.688 + ], + "angle": 0, + "content": "sures the performance drop on forgetting classes before and after training. After unlearning task \\(\\mathcal{T}_t\\), we evaluate performance on all previously forgetting classes from tasks \\(\\mathcal{T}_1, \\mathcal{T}_2, \\dots, \\mathcal{T}_{t-1}\\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.688, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Baselines. Following [60], we compare FG-OrIU against (a) continual learning (CL) methods, including L2 regularization, EWC [27], MAS [1], LwF [30], DER [6], and FDR [3], as well as (b) machine unlearning (MU) methods such as BAD-T [12], LIRF [53], SCRUB/SCRUB-S [29], GS-LoRA/GS-LoRA++ [60, 61], and retraining. To ensure forgetting occurs in the backbone, we freeze the classification head for all methods. In the retraining setup, we train a randomly initialized model using replay data, ensuring the number of training epochs matches that of other methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.841, + 0.317, + 0.856 + ], + "angle": 0, + "content": "5.2. Results and Comparison" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Table 2 presents the results for static Machine Unlearning task. FG-OrIU exhibits significantly degraded performance on forgetting classes while maintaining the perfor" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.391, + 0.907, + 0.739 + ], + "angle": 0, + "content": "mance of remaining classes. For example, in the settings where 50 and 90 classes are removed per task, FG-OrIU achieves improvements of \\(2.62\\%\\) and \\(2.64\\%\\) in \\(H\\), and \\(4.99\\%\\) and \\(5.29\\%\\) in \\(ACC_{r}\\) compared to GS-LoRA. Notably, FG-OrIU keeps the pre-trained model frozen and only trains the LoRA modules, enabling effective and efficient forgetting. Table 3 and 4 show the results for Incremental Unlearning task. FG-OrIU consistently outperforms all methods. Specifically, CL methods struggle to achieve effective forgetting on ImageNet100, as seen in DER, which retains high \\(ACC_{f}\\) of \\(10.40\\%\\), \\(10.90\\%\\), \\(8.80\\%\\), and \\(3.40\\%\\) across four tasks, respectively. On the other hand, MU methods fail to maintain their forgetting effect over sequential unlearning tasks. For example, LIRF and SCRUB-S show an increase in \\(ACC_{o}\\) on CASIA-Face100, which becomes more pronounced as the sequence of unlearning tasks grows. GS-LoRA is specifically designed for IU. Compared to it, our method achieves better forgetting while also demonstrating superior performance in preserving the remaining classes, i.e., substantial improvements in \\(ACC_{r}\\), particularly in long-sequence unlearning scenarios, achieving gains of \\(5.15\\%\\) and \\(4.6\\%\\) in the fourth unlearning task on two different datasets, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.748, + 0.667, + 0.764 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Our FG-OrIU consists of four components, i.e., feature orthogonal projection for forgetting, feature orthogonal projection for remaining, gradient orthogonal projection for forgetting, and gradient orthogonal projection for remaining, denoted as \\( fop - f \\), \\( fop - r \\), \\( gop - f \\), and \\( gop - r \\). We independently remove each of these components to demonstrate their individual contributions. We observe that when the regularization term or gradient modification related to remaining is removed, the performance on the remaining" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1962" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.907, + 0.324 + ], + "angle": 0, + "content": "
MethodsTask 1(100-20)Task 2(80-20)Task 3(60-20)Task 4(40-20)
H↑Accr↑Accf↓H↑Accr↑Accf↓Acco↓H↑Accr↑Accf↓Acco↓H↑Accr↑Accf↓Acco↓
Pre-train-74.3174.65-74.5073.80--74.8073.91--74.4775.11-
L2*66.9162.132.1666.6661.741.379.3666.4261.371.5411.8366.9561.020.978.00
EWC*67.7161.950.0067.7162.550.000.0067.0961.430.000.1068.0262.140.000.23
MAS*67.5261.630.0068.1263.250.000.0068.1563.230.000.0367.7061.610.000.00
LwF69.4365.040.2169.9466.470.000.2170.3467.110.000.0070.8967.120.000.04
DER70.7567.240.0068.8864.580.000.0068.9564.620.000.0069.4164.510.000.00
DER++70.0165.900.0068.9964.770.000.0069.9666.420.000.0069.8565.280.000.00
FDR68.2962.920.0067.3962.010.000.0069.1664.990.000.0072.5170.080.000.00
Retrain17.299.770.0031.1219.720.000.0041.1728.801.720.0052.4441.774.660.09
BAD-T68.9063.970.0069.5665.950.210.0070.2166.870.000.3472.7170.460.000.04
LIRF*30.5964.4654.6034.0562.0350.3443.5044.5662.5339.2936.5840.3662.6245.3427.96
SCRUB70.3966.590.0070.5567.850.320.0071.0168.330.000.4473.3671.680.000.05
SCRUB-S72.4170.340.0570.0671.535.1615.2273.3873.090.2415.8175.3376.240.686.47
GS-LoRA74.4074.160.0073.5973.370.000.0574.3674.880.060.0073.7672.450.001.93
GS-LoRA++73.9774.160.0073.4873.160.000.1074.2074.510.000.0075.2375.360.000.00
FG-OrIU74.4574.250.0074.6475.500.000.0075.4977.140.000.0077.7180.510.000.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.326, + 0.907, + 0.368 + ], + "angle": 0, + "content": "Table 3. Incremental Unlearning results for face recognition task on CASIA-Face100. \\( Acc_{o}(\\%) \\) is the accuracy of old tasks, i.e., the accuracy on all previously forgetting classes in task \\( \\{\\mathcal{T}_1,\\dots ,\\mathcal{T}_{t - 1}\\} \\). With four tasks, each forgetting 20 classes, desired forgetting should consider both \\( Acc_{r} \\) and \\( Acc_{f} \\). We bold the best and underline the second-best \\( H \\) results." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.369, + 0.907, + 0.602 + ], + "angle": 0, + "content": "
MethodsTask 1(100-20)Task 2(80-20)Task 3(60-20)Task 4(40-20)
H↑Accr↑Accf↓H↑Accr↑Accf↓Acco↓H↑Accr↑Accf↓Acco↓H↑Accr↑Accf↓Acco↓
Pre-train-89.9387.40-90.0389.60--89.6590.80--92.8086.50-
L2*79.3988.3315.3083.5988.3710.3048.1084.0188.2010.6037.1582.7289.909.9020.97
EWC*82.6978.550.1085.7182.570.500.2087.3384.900.902.9086.9389.502.008.27
MAS*80.4474.500.0081.5174.770.000.6082.1875.050.000.6081.9377.900.101.77
LwF85.8285.150.9087.2685.400.402.4087.3584.150.002.3586.0585.700.101.07
DER82.8889.7310.4084.2290.5710.9050.8086.3691.208.8042.0088.5294.703.4024.90
DER++83.4589.739.4084.1390.6311.1050.1087.5091.306.8040.4089.2494.802.2020.80
FDR29.5117.780.5036.2122.700.100.0040.7926.300.000.0047.9133.502.401.27
Retrain48.2033.280.0056.2841.301.300.0061.1447.555.200.2067.2360.6011.000.53
BAD-T74.7465.290.0078.0969.210.000.1076.6566.320.000.0076.2568.170.000.00
LIRF*52.7264.2842.7161.0569.8335.3632.1560.9660.3229.1830.9158.5168.4635.4121.04
SCRUB75.8767.030.0076.2366.330.000.0075.4664.550.000.0077.5670.300.000.00
SCRUB-S83.4479.830.0082.5976.600.000.1079.8171.200.000.0078.2371.400.000.07
GS-LoRA86.1385.000.1087.3385.270.101.1087.8485.150.100.3586.3086.300.200.07
GS-LoRA++86.6385.980.1087.2285.230.300.3087.4884.400.000.6087.0087.500.000.07
FG-OrIU87.0186.730.1088.4587.430.100.1088.9087.350.300.0089.1592.100.100.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.262, + 0.604, + 0.733, + 0.618 + ], + "angle": 0, + "content": "Table 4. Incremental Unlearning results for classification task on ImageNet100." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.618, + 0.481, + 0.743 + ], + "angle": 0, + "content": "
VariantsTask 1(100-20)Task 2(80-20)
Accr↑Accf↓Accr↑Accf↓Acco↓
Pre-train74.3074.6074.5073.80-
FG-OrIU74.250.0075.500.000.00
w/o fop-f73.010.2076.240.421.28
w/oGOP-f73.920.1076.584.370.82
w/o fop-r73.920.0574.980.210.15
w/oGOP-r71.450.0073.960.000.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.149, + 0.747, + 0.425, + 0.761 + ], + "angle": 0, + "content": "Table 5. Ablation studies on CASIA-Face100." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.856 + ], + "angle": 0, + "content": "classes significantly declines. Meanwhile, when the regularization term or gradient modification related to forgetting is removed, the performance on the forgetting classes shows a slight improvement. Thus, each component provides clear benefits in achieving a better trade-off between forgetting and remaining while preserving the old forgetting effect." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.856, + 0.21, + 0.87 + ], + "angle": 0, + "content": "5.4. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "IU is designed to ensure strict privacy protection. This raises several key questions: (a) Does the unlearned model" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.628, + 0.905, + 0.728 + ], + "angle": 0, + "content": "
ExpsMetircPre-trainDERFDRSCRUBGS-LoRAOurs
Exp1SSIM0.940.600.340.750.390.01
PSNR31.3221.2216.8923.4017.608.37
Exp2SSIM0.950.490.660.630.380.02
PSNR31.1519.1722.1521.1217.328.72
Exp3SSIM0.930.530.490.600.410.01
PSNR31.1020.0319.8322.0115.928.55
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.732, + 0.907, + 0.762 + ], + "angle": 0, + "content": "Table 6. SSIM \\(\\uparrow\\) and PSNR(db) \\(\\uparrow\\) scores for the reconstructed images using DIP. Here, Pre-train serves as the upper bound." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.779, + 0.907, + 0.84 + ], + "angle": 0, + "content": "still retain any private information? (b) Can the forgetting effect be reversed? and (c) How does unlearning impact the feature subspace? To address these questions, we design the following experiment." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.844, + 0.753, + 0.857 + ], + "angle": 0, + "content": "5.4.1. Semantic Residual Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.857, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We first investigate the semantic information contained in features after unlearning. We utilize Deep Image Prior (DIP) [46] to reconstruct an image from the features ob" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1963" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.09, + 0.338, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.184, + 0.275, + 0.245, + 0.284 + ], + "angle": 0, + "content": "(a) Pre-train" + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.091, + 0.603, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.459, + 0.275, + 0.503, + 0.284 + ], + "angle": 0, + "content": "(b) Ours" + }, + { + "type": "image", + "bbox": [ + 0.626, + 0.091, + 0.906, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.714, + 0.275, + 0.781, + 0.284 + ], + "angle": 0, + "content": "(c) GS-LoRA" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.286, + 0.908, + 0.315 + ], + "angle": 0, + "content": "Figure 4. TSNE visualization with 5 forgetting classes and 10 remaining classes, using different model as feature extractor: (a) the pretrained model, (b) the unlearned model through FG-OrIU, and (c) through GS-LoRA." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.319, + 0.484, + 0.381 + ], + "angle": 0, + "content": "tained from the last layer of the unlearned model. Figure 1 presents the visualization of the reconstructed image, which allows us to infer the semantic information encoded in features. More results are provided in the Appendix ??" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.388, + 0.484, + 0.614 + ], + "angle": 0, + "content": "Additionally, we measure the SSIM and PSNR scores for the images generated by DIP from the unlearned model using different IU methods as a quantitative results. As shown in Table 6, the SSIM values of the reconstructed images from the compared methods range between 0.35 and 0.75, with the highest being 0.75 for SCRUB on \\(exp1\\), while the PSNR values range between 16 and 24, with the highest being 23.40 for SCRUB on \\(exp1\\). This indicates that the reconstructed image achieves a high similarity with the original image, suggesting that the information in features has been partially preserved. In contrast, our method yields an SSIM of only 0.01 and a PSNR of 8.55, indicating a significant difference between the reconstructed and original images. This suggests that the semantic information at the feature level has been substantially obfuscated." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.629, + 0.363, + 0.644 + ], + "angle": 0, + "content": "5.4.2. Forgetting Recoverable Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.648, + 0.484, + 0.906 + ], + "angle": 0, + "content": "Given an unlearned model obtained through different IU methods, we freeze its backbone and retrain a new classification head using the entire training dataset, which includes both forgetting and remaining classes. As illustrated in Figure 3, MLP* represents a trivial unlearning method where the classification head's parameters for the forgetting classes are simply set to zero. Consequently, the underlying features remain intact, enabling effortless recovery of the forgotten knowledge. Additionally, existing IU methods primarily impose constraints on the final output rather than directly modifying feature representations. As a result, while features may be weakened to some extent, their core structure remains largely unchanged. Thus, these methods allow the forgotten knowledge to be recovered to \\(15\\% -70\\%\\) of its original performance, which leads to superficial forgetting. In contrast, our method achieves deep forgetting, making recovery infeasible." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.32, + 0.746, + 0.335 + ], + "angle": 0, + "content": "5.4.3. Feature Subspace Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.339, + 0.906, + 0.414 + ], + "angle": 0, + "content": "To analyze the impact of direct constraints on features and gradients for forgetting and remaining classes, we conduct an experiment. We remove 20 classes while retaining 80, then extract features from 5 forgetting and 10 remaining classes and visualize them using t-SNE." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.415, + 0.907, + 0.732 + ], + "angle": 0, + "content": "A key observation is that GS-LoRA causes some forgetting class samples to move closer to or even overlap with certain remaining class samples (e.g., F21, \\( \\mathrm{F}74 \\rightarrow \\mathrm{R}92 \\) in c). This suggests that merely applying constraints at the final output, resulting in degraded features from the forgetting classes entangling with those of the remaining classes. Consequently, this degradation negatively impacts the performance of the remaining classes. In contrast, our method enforces forgetting at both feature and gradient level, ensuring that features from the forgetting classes do not shift toward any remaining class but instead become linearly inseparable within the forgetting set. Notably, an unexpected yet beneficial side effect emerges: in the original PTM, certain remaining class samples (e.g., R59) were slightly entangled with forgetting classes (e.g., F90 and F46). However, after applying our method, these classes become completely disentangled. This observation aligns with the results in Table 3, where the accuracy of the remaining classes improves after unlearning (e.g., \\( 75.36 \\rightarrow 80.51 \\)), demonstrating that our method not only enhances forgetting effectiveness but also benefits the generalization of the remaining classes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.747, + 0.634, + 0.761 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.901 + ], + "angle": 0, + "content": "In this work, we propose FG-OrIU for incremental unlearning in vision PTMs. By enforcing orthogonal constraints on both features and gradients, FG-OrIU achieves irreversible forgetting while maintaining performance on the remaining classes. Its dynamic subspace adaptation ensures robust handling of sequential deletion requests. Extensive experiments confirm its superiority. While FG-OrIU is designed for class-level unlearning, extending it to sample-level settings remains a promising direction for future work." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1964" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.283 + ], + "angle": 0, + "content": "This work was supported in part by National Natural Science Foundation of China under Grant 62402430, 62206248, 62476238, Zhejiang Provincial Natural Science Foundation of China under Grant LQN25F020008, and Aeronautical Science Foundation of China 20240048076001. Qian Feng would like to extend his gratitude to Jiahang Tu and Hanbin Zhao from Zhejiang University for their discussions regarding the methods and experiments. Additionally, he is thankful to Mintong Kang from UIUC for his dedicated suggestions on improving the overall writing of the paper." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.293, + 0.188, + 0.309 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.318, + 0.486, + 0.387 + ], + "angle": 0, + "content": "[1] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In Proceedings of the European conference on computer vision (ECCV), pages 139-154, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.388, + 0.484, + 0.444 + ], + "angle": 0, + "content": "[2] Anish Athalye, Nicholas Carlini, and David Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In International conference on machine learning, pages 274-283. PMLR, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.444, + 0.484, + 0.485 + ], + "angle": 0, + "content": "[3] Ari S Benjamin, David Rolnick, and Konrad Kording. Measuring and regularizing networks in function space. arXiv preprint arXiv:1805.08289, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.485, + 0.484, + 0.554 + ], + "angle": 0, + "content": "[4] Jacopo Bonato, Marco Cotogni, and Luigi Sabetta. Is retain set all you need in machine unlearning? restoring performance of unlearned models with out-of-distribution images. In European Conference on Computer Vision, pages 1-19. Springer, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.555, + 0.484, + 0.622 + ], + "angle": 0, + "content": "[5] Lucas Bourtoule, Varun Chandrasekaran, Christopher A Choquette-Choo, Hengrui Jia, Adelin Travers, Baiwu Zhang, David Lie, and Nicolas Papernot. Machine unlearning. In 2021 IEEE symposium on security and privacy (SP), pages 141-159. IEEE, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.623, + 0.484, + 0.692 + ], + "angle": 0, + "content": "[6] Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark experience for general continual learning: a strong, simple baseline. Advances in neural information processing systems, 33:15920-15930, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.693, + 0.484, + 0.735 + ], + "angle": 0, + "content": "[7] Yinzhi Cao and Junfeng Yang. Towards making systems forget with machine unlearning. In 2015 IEEE symposium on security and privacy, pages 463-480. IEEE, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.735, + 0.484, + 0.79 + ], + "angle": 0, + "content": "[8] Romit Chatterjee, Vikram Chandawat, Ayush Tarun, Ankur Mali, and Murari Mandal. A unified framework for continual learning and machine unlearning. arXiv preprint arXiv:2408.11374, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.79, + 0.484, + 0.831 + ], + "angle": 0, + "content": "[9] Jiaao Chen and Diyi Yang. Unlearn what you want to forget: Efficient unlearning for llms. arXiv preprint arXiv:2310.20150, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[10] Xinwen Cheng, Zhehao Huang, Wenxin Zhou, Zhengbao He, Ruikai Yang, Yingwen Wu, and Xiaolin Huang. Remaining-data-free machine unlearning by suppressing sample contribution. arXiv preprint arXiv:2402.15109, 2024. 1, 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.318, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.148 + ], + "angle": 0, + "content": "[11] Zhi-Yi Chin, Chieh-Ming Jiang, Ching-Chun Huang, Pin-Yu Chen, and Wei-Chen Chiu. Prompting4debugging: Red-teaming text-to-image diffusion models by finding problematic prompts. arXiv preprint arXiv:2309.06135, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.906, + 0.218 + ], + "angle": 0, + "content": "[12] Vikram S Chandawat, Ayush K Tarun, Murari Mandal, and Mohan Kankanhalli. Can bad teaching induce forgetting? unlearning in deep networks using an incompetent teacher. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7210-7217, 2023. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.906, + 0.273 + ], + "angle": 0, + "content": "[13] Meng Ding, Jinhui Xu, and Kaiyi Ji. Why fine-tuning struggles with forgetting in machine unlearning? theoretical insights and a remedial approach. arXiv preprint arXiv:2410.03833, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.906, + 0.356 + ], + "angle": 0, + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.906, + 0.425 + ], + "angle": 0, + "content": "[15] Chongyu Fan, Jiancheng Liu, Yihua Zhang, Eric Wong, Dennis Wei, and Sijia Liu. Salun: Empowering machine unlearning via gradient-based weight saliency in both image classification and generation. arXiv preprint arXiv:2310.12508, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.427, + 0.906, + 0.482 + ], + "angle": 0, + "content": "[16] Jack Foster, Stefan Schoepf, and Alexandra Brintrup. Fast machine unlearning without retraining through selective synaptic dampening. In Proceedings of the AAAI conference on artificial intelligence, pages 12043-12051, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.906, + 0.525 + ], + "angle": 0, + "content": "[17] Chongyang Gao, Lixu Wang, Chenkai Weng, Xiao Wang, and Qi Zhu. Practical unlearning for large language models. arXiv preprint arXiv:2407.10223, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.906, + 0.579 + ], + "angle": 0, + "content": "[18] Laura Graves, Vineel Nagisetty, and Vijay Ganesh. Amnesiac machine learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11516-11524, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.906, + 0.635 + ], + "angle": 0, + "content": "[19] Keltin Grimes, Collin Abidi, Cole Frank, and Shannon Gallagher. Gone but not forgotten: Improved benchmarks for machine unlearning. arXiv preprint arXiv:2405.19211, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.636, + 0.906, + 0.678 + ], + "angle": 0, + "content": "[20] Chuan Guo, Tom Goldstein, Awni Hannun, and Laurens Van Der Maaten. Certified data removal from machine learning models. arXiv preprint arXiv:1911.03030, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.679, + 0.906, + 0.761 + ], + "angle": 0, + "content": "[21] Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. Ms-celeb-1m: A dataset and benchmark for large-scale face recognition. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 87-102. Springer, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.906, + 0.815 + ], + "angle": 0, + "content": "[22] Zeyu Han, Chao Gao, Jinyang Liu, Jeff Zhang, and Sai Qian Zhang. Parameter-efficient fine-tuning for large models: A comprehensive survey. arXiv preprint arXiv:2403.14608, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.906, + 0.872 + ], + "angle": 0, + "content": "[23] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.906, + 0.903 + ], + "angle": 0, + "content": "[24] Zhehao Huang, Xinwen Cheng, JingHao Zheng, Haoran Wang, Zhengbao He, Tao Li, and Xiaolin Huang. Unified" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1965" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "gradient-based machine unlearning with remain geometry enhancement. arXiv preprint arXiv:2409.19732, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.123, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[25] Joel Jang, Dongkeun Yoon, Sohee Yang, Sungmin Cha, Moontae Lee, Lajanugen Logeswaran, and Minjoon Seo. Knowledge unlearning for mitigating privacy risks in language models. arXiv preprint arXiv:2210.01504, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.181, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[26] Jinghan Jia, Jiancheng Liu, Parikshit Ram, Yuguang Yao, Gaowen Liu, Yang Liu, Pranay Sharma, and Sijia Liu. Model sparsity can simplify machine unlearning. Advances in Neural Information Processing Systems, 36:51584-51605, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.252, + 0.482, + 0.333 + ], + "angle": 0, + "content": "[27] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.337, + 0.482, + 0.377 + ], + "angle": 0, + "content": "[28] Vishnuprasadh Kumaravelu, Sayanta Adhikari, and PK Srijith. Uncle: An unlearning framework for continual learning. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.381, + 0.482, + 0.436 + ], + "angle": 0, + "content": "[29] Meghdad Kurmanji, Peter Triantafillou, Jamie Hayes, and Eleni Triantafillou. Towards unbounded machine unlearning. Advances in neural information processing systems, 36: 1957-1987, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.439, + 0.482, + 0.48 + ], + "angle": 0, + "content": "[30] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.483, + 0.482, + 0.524 + ], + "angle": 0, + "content": "[31] Bo Liu, Qiang Liu, and Peter Stone. Continual learning and private unlearning. In Conference on Lifelong Learning Agents, pages 243-254. PMLR, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.527, + 0.482, + 0.595 + ], + "angle": 0, + "content": "[32] Gaoyang Liu, Xiaogiang Ma, Yang Yang, Chen Wang, and Jiangchuan Liu. Federaser: Enabling efficient client-level data removal from federated learning models. In 2021 IEEE/ACM 29th International Symposium on Quality of Service (IWQOS), pages 1-10. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.599, + 0.482, + 0.679 + ], + "angle": 0, + "content": "[33] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.684, + 0.482, + 0.737 + ], + "angle": 0, + "content": "[34] Laiqiao Qin, Tianqing Zhu, Linlin Wang, and Wanlei Zhou. Machine unlearning on pre-trained models by residual feature alignment using lora. arXiv preprint arXiv:2411.08443, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.742, + 0.482, + 0.796 + ], + "angle": 0, + "content": "[35] Sadia Qureshi, Thanveer Shaik, Xiaohui Tao, Haoran Xie, Lin Li, Jianming Yong, and Xiaohua Jia. Exploring incremental unlearning: Techniques, challenges, and future directions. arXiv preprint arXiv:2502.16708, 2025. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.8, + 0.482, + 0.826 + ], + "angle": 0, + "content": "[36] Protection Regulation. General data protection regulation. Intouch, 25:1-5, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.829, + 0.482, + 0.855 + ], + "angle": 0, + "content": "[37] Jeffrey Rosen. The right to be forgotten. *Stan. L. Rev. Online*, 64:88, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[38] Dan Roth and Wen-tau Yih. Probabilistic reasoning for entity & relation recognition. In COLING 2002: The 19th International Conference on Computational Linguistics, 2002. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.161 + ], + "angle": 0, + "content": "[39] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.906, + 0.203 + ], + "angle": 0, + "content": "[40] Gobinda Saha, Isha Garg, and Kaushik Roy. Gradient projection memory for continual learning. arXiv preprint arXiv:2103.09762, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.906, + 0.272 + ], + "angle": 0, + "content": "[41] Thanveer Shaik, Xiaohui Tao, Haoran Xie, Lin Li, Xiaofeng Zhu, and Qing Li. Exploring the landscape of machine unlearning: A comprehensive survey and taxonomy. IEEE Transactions on Neural Networks and Learning Systems, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.314 + ], + "angle": 0, + "content": "[42] Yingdan Shi and Ren Wang. Redefining machine unlearning: A conformal prediction-motivated approach. arXiv preprint arXiv:2501.19403, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.905, + 0.355 + ], + "angle": 0, + "content": "[43] Takashi Shibata, Go Irie, Daiki Ikami, and Yu Mitsuzumi. Learning with selective forgetting. In *IJCAI*, page 6, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.357, + 0.905, + 0.41 + ], + "angle": 0, + "content": "[44] Mihai Surdeanu and Heng Ji. Overview of the english slot filling track at the tac2014 knowledge base population evaluation. In Proc. Text Analysis Conference (TAC2014), 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.413, + 0.905, + 0.481 + ], + "angle": 0, + "content": "[45] Jiahang Tu, Qian Feng, Chufan Chen, Jiahua Dong, Hanbin Zhao, Chao Zhang, and Hui Qian. Ce-sdwv: Effective and efficient concept erasure for text-to-image diffusion models via a semantic-driven word vocabulary. arXiv preprint arXiv:2501.15562, 2025. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.483, + 0.905, + 0.536 + ], + "angle": 0, + "content": "[46] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9446-9454, 2018. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.578 + ], + "angle": 0, + "content": "[47] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.905, + 0.634 + ], + "angle": 0, + "content": "[48] Cheng-Long Wang, Qi Li, Zihang Xiang, and Di Wang. Has approximate machine unlearning been evaluated properly? from auditing to side effects. arXiv e-prints, pages arXiv-2403, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.905, + 0.677 + ], + "angle": 0, + "content": "[49] Heng Xu, Tianqing Zhu, Lefeng Zhang, Wanlei Zhou, and Philip S Yu. Machine unlearning: A survey. ACM Computing Surveys, 56(1):1-36, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[50] Jie Xu, Zihan Wu, Cong Wang, and Xiaohua Jia. Machine unlearning: Solutions and challenges. IEEE Transactions on Emerging Topics in Computational Intelligence, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.721, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[51] Lingling Xu, Haoran Xie, Si-Zhao Joe Qin, Xiaohui Tao, and Fu Lee Wang. Parameter-efficient fine-tuning methods for pretrained language models: A critical review and assessment. arXiv preprint arXiv:2312.12148, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.777, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[52] Haonan Yan, Xiaoguang Li, Ziyao Guo, Hui Li, Fenghua Li, and Xiaodong Lin. Arcane: An efficient architecture for exact machine unlearning. In *IJCAI*, page 19, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.859 + ], + "angle": 0, + "content": "[53] Jingwen Ye, Yifang Fu, Jie Song, Xingyi Yang, Songhua Liu, Xin Jin, Mingli Song, and Xinchao Wang. Learning with recoverable forgetting. In ECCV, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[54] Dong Yi, Zhen Lei, Shengcai Liao, and Stan Z Li. Learning face representation from scratch. arXiv preprint arXiv:1411.7923, 2014. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1966" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[55] Jaehong Yoon, Shoubin Yu, Vaidehi Patil, Huaxiu Yao, and Mohit Bansal. Safree: Training-free and adaptive guard for safe text-to-image and video generation. arXiv preprint arXiv:2410.12761, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[56] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. Advances in Neural Information Processing Systems, 33:5824-5836, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[57] Dan Zhang, Tao Feng, Lilong Xue, Yuandong Wang, Yuxiao Dong, and Jie Tang. Parameter-efficient fine-tuning for foundation models. arXiv preprint arXiv:2501.13787, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.304 + ], + "angle": 0, + "content": "[58] Haibo Zhang, Toru Nakamura, Takamasa Isohara, and Kouichi Sakurai. A review on machine unlearning. SN Computer Science, 4(4):337, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.482, + 0.361 + ], + "angle": 0, + "content": "[59] Yuanhan Zhang, Zhenfei Yin, Jing Shao, and Ziwei Liu. Benchmarking omni-vision representation through the lens of visual realms. In European Conference on Computer Vision, pages 594-611. Springer, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[60] Hongbo Zhao, Bolin Ni, Junsong Fan, Yuxi Wang, Yuntao Chen, Gaofeng Meng, and Zhaoxiang Zhang. Continual forgetting for pre-trained vision models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28631-28642, 2024. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.433, + 0.482, + 0.487 + ], + "angle": 0, + "content": "[61] Hongbo Zhao, Fei Zhu, Bolin Ni, Feng Zhu, Gaofeng Meng, and Zhaoxiang Zhang. Practical continual forgetting for pretrained vision models. arXiv preprint arXiv:2501.09705, 2025. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.489, + 0.482, + 0.571 + ], + "angle": 0, + "content": "[62] Zhen Zhao, Zhizhong Zhang, Xin Tan, Jun Liu, Yanyun Qu, Yuan Xie, and Lizhuang Ma. Rethinking gradient projection continual learning: Stability/plasticity feature space decoupling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3718-3727, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.573, + 0.482, + 0.601 + ], + "angle": 0, + "content": "[63] Yaoyao Zhong and Weihong Deng. Face transformer for recognition. arXiv preprint arXiv:2103.14803, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.602, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[64] Chen Zhu, Ankit Singh Rawat, Manzil Zaheer, Srinadh Bhojanapalli, Daliang Li, Felix Yu, and Sanjiv Kumar. Modifying memories in transformer models. arXiv preprint arXiv:2012.00363, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.657 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1967" + } + ] +] \ No newline at end of file diff --git a/2025/FG-OrIU_ Towards Better Forgetting via Feature-Gradient Orthogonality for Incremental Unlearning/4a750062-438a-4c8c-b4db-9a6576c3de7a_origin.pdf b/2025/FG-OrIU_ Towards Better Forgetting via Feature-Gradient Orthogonality for Incremental Unlearning/4a750062-438a-4c8c-b4db-9a6576c3de7a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d2a9e68d8f1fd8bc401f2d9e6432aa76bea976bc --- /dev/null +++ b/2025/FG-OrIU_ Towards Better Forgetting via Feature-Gradient Orthogonality for Incremental Unlearning/4a750062-438a-4c8c-b4db-9a6576c3de7a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5a083b55093ccc7215b852d819ddd021f03ab0f86b00323b27a01301d64d9c84 +size 9575532 diff --git a/2025/FG-OrIU_ Towards Better Forgetting via Feature-Gradient Orthogonality for Incremental Unlearning/full.md b/2025/FG-OrIU_ Towards Better Forgetting via Feature-Gradient Orthogonality for Incremental Unlearning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3261cc420cc21a7cc4e5bd56ab387bdb3c3f6bb1 --- /dev/null +++ b/2025/FG-OrIU_ Towards Better Forgetting via Feature-Gradient Orthogonality for Incremental Unlearning/full.md @@ -0,0 +1,337 @@ +# FG-OrIU: Towards Better Forgetting via Feature-Gradient Orthogonality for Incremental Unlearning + +Qian Feng $^{1}$ , JiaHang Tu $^{1}$ , Mintong Kang $^{2}$ , Hanbin Zhao $^{1*}$ , Chao Zhang $^{1}$ , Hui Qian $^{1}$ + +1College of Computer Science and Technology, Zhejiang University +2Department of Computer Science University of Illinois at Urbana-Champaign + +# Abstract + +Incremental unlearning (IU) is critical for pre-trained models to comply with sequential data deletion requests, yet existing methods primarily suppress parameters or confuse knowledge without explicit constraints on both feature and gradient level, resulting in superficial forgetting where residual information remains recoverable. This incomplete forgetting risks security breaches and disrupts retention balance, especially in IU scenarios. We propose FG-OrIU (Feature-Gradient Orthogonality for Incremental Unlearning), the first framework unifying orthogonal constraints on both features and gradients level to achieve deep forgetting, where the forgetting effect is irreversible. FG-OrIU decomposes feature spaces via Singular Value Decomposition (SVD), separating forgetting and remaining class features into distinct subspaces. It then enforces dual constraints: feature orthogonal projection on both forgetting and remaining classes, while gradient orthogonal projection prevents the reintroduction of forgotten knowledge and disruption to remaining classes during updates. Additionally, dynamic subspace adaptation merges newly forgetting subspaces and contracts remaining subspaces, ensuring a stable balance between removal and retention across sequential unlearning tasks. Extensive experiments demonstrate the effectiveness of our method. + +# 1. Introduction + +In recent years, pre-training with fine-tuning [22, 51, 57] has dominated model development. However, large-scale datasets often contain harmful or private content, necessitating mechanisms to erase specific data traces. Machine Unlearning (MU) [4, 5, 7, 12, 16, 20, 24, 34, 41, 42, 50] addresses this by enabling pre-trained models (PTMs) to comply with deletion mandates like GDPR (General data protection regulation)'s "right to be forgotten" [36, 37]. While existing MU methods focus on static unlearning, for + +![](images/e8d792e49fac155c7c124f2cda35afc426466289ac4a664bdba1ab3bf5de5de0.jpg) +Figure 1. Here, we present three examples from forgetting classes, with the reconstructed images generated using Deep Image Prior (DIP) [46], where the last block's features from the unlearned model serve as the training target. + +getting a single or small set of classes in one task, real-world scenarios require Incremental Unlearning (IU) [8-10, 17, 19, 25, 28, 31, 35, 48, 60, 61]. Regulatory shifts (e.g., evolving OpenAI safety policies $^{1}$ , EU AI Act $^{2}$ ) and dynamic content restrictions necessitate sequential unlearning of multiple classes/concepts without costly retraining. IU need not only erase targets in a single step but also preserve retained performance and ensure irreversible forgetting across tasks, making it critical for sustainable model governance in evolving ethical and legal landscapes. + +IU methods broadly address two scenarios: (1) concept erasure in generative models, where sensitive concepts (e.g., nudity, violence) are iteratively removed from diffusion models [11, 45], and (2) instant- or class-level unlearning in vision classifiers, where PTMs forget specific classes (e.g., face identity, medical data) while retaining unrelated ones [26, 34]. In this work, we focus on the latter, specifically targeting class-level IU in vision PTMs. Our goal is threefold: (a) to ensure that the model completely forgets the specified classes, (b) to robustly preserve the performance of remaining classes, and (c) to maintain the effectiveness of previous forgetting across sequential unlearning procedures. + +Existing methods largely fall into two categories: parameter suppression, which selectively weakens target class-related parameters [15, 16, 52], and knowledge confusion, which distorts predictions via noisy labels or bad teacher distillation [12, 18, 29]. However, our empirical findings reveal that these methods lead to superficial forgetting, as the unlearned model still retains semantic information related to the forgetting classes, making the erased knowledge partially residual and recoverable. Specifically, reconstructing features from the last layer of the unlearned model using Deep Image Prior (DIP) [46] produces blurred yet recognizable images (as illustrated in Figure 1). Furthermore, fixing the backbone of the unlearned model and retraining only the classification head [10, 61], which recovers a portion of the original accuracy (as illustrated in Figure 3). The blurred image and the partial recovery of performance both indicate that existing methods merely degrade the features of forgetting classes rather than completely eliminating them. The degraded features still contain residual class-specific information, which leads to potential recovery of the forgetting effect. Besides, these degraded features not only enable recovery but also risk entanglement with the remaining class features [34]. Such entanglement destabilizes decision boundaries, degrading performance on the remaining classes, particularly in IU scenarios where sequential deletion requests amplify error propagation [17, 48]. To address these limitations and achieve deep forgetting, we argue that unlearning should explicitly constrain both feature representations and gradient updates. Since model updates rely on both forward computations (feature extraction) and backward optimization (gradient updates), it is critical to simultaneously suppress the features of forgetting classes and prevent their reintroduction during training updates. In Table 1, we summarize the differences between our methods and prior works. + +To enforce explicit constraints on both feature representations and gradient updates, we draw inspiration from orthogonal mechanisms in related domains. For feature-level constraints, recent work in text-to-image models [45, 55] demonstrates that orthogonalizing sensitive token embeddings effectively removes targeted concepts from diffusion outputs, suggesting geometric isolation as a viable strategy for feature suppression. For gradient-level constraints, inspired by [56], which reduces the mutual influence of different tasks in multi-task learning, and [40, 62], which mitigates the forgetting of old tasks in continual learning, we leverage the insight that updating in directions orthogonal to old/other task subspaces helps alleviate catastrophic forgetting. While these orthogonal principles show promise in adjacent fields like generative model editing and multitask/continual learning, their applicability to MU, particularly IU in vision PTMs, remains underexplored. + +In this paper, we propose FG-OrIU (Feature-Gradient + +Table 1. Position, Complete, and Recover refers to, where the forgetting occurs, whether the forgetting is thorough, and whether the forgetting effect can be restored, respectively. + +
PositionCompleteRecover
superficial forgetting
Prior worksOnly Final OutputNoPartial
deep forgetting
OursFeature&GradientYesNo
+ +Orthogonality for Incremental Unlearning), a unified framework for IU in vision PTMs that enforces explicit constraints at both feature and gradient levels through three stages: feature subspace decomposition, dual orthogonal projection, and dynamic subspace adaptation. In feature subspace decomposition, we decompose the feature space of each layer into two subspaces via Singular Value Decomposition (SVD): the forgetting subspace $S_f$ , which captures the discriminative features of the classes to be removed, and the remaining subspace $S_r$ , which encodes the features of the remaining classes. In dual orthogonal projection, we freeze the backbone of PTM and introduce lightweight LoRA modules for efficient updates. Then, features of samples from forgetting classes are orthogonalized against $S_f$ to eliminate correlations, while remaining classes features are aligned with $S_r$ to minimize interference through minimizing the projection residual. Before backward optimization, gradients are first projected onto $S_f$ to maximally perturb parameters critical for forgetting classes, then modified along the orthogonal complement of $S_r$ to prevent contamination of retained features. In dynamic subspace adaptation, FOrIU dynamically expands $S_f$ by merging new forgetting class subspaces and contracts $S_r$ by removing overlapping features, ensuring cumulative forgetting without degrading retained performance. + +# Our main contributions are threefold: + +- We reveal that existing IU methods achieve superficial forgetting rather than deep forgetting, as their partial feature suppression leaves residual information recoverable. +- We propose the first framework unifying forgetting on both feature and gradient level via dual orthogonal projection, ensuring irreversible forgetting while preserving remaining class robustness. +- Extensive experiments demonstrate superior performance in MU and IU scenarios across multiple benchmarks. + +# 2. Related Work + +# 2.1. Machine Unlearning + +Machine unlearning (MU) aims to remove the influence of specific data from pre-trained models (PTMs) without retraining from scratch [49, 58]. Existing MU methods are broadly categorized into exact and approximate unlearning. Exact MU [5, 32] ensures statistical indistinguishability from models retrained on remaining data but is computationally prohibitive for large PTMs. Approximate MU re + +taxes this guarantee and is more practical. Current methods fall into two paradigms: (1) concept erasure in generative models [11, 45]; and (2) class-/sample-wise unlearning in classification models, which modifies pre-trained weights to forget target classes while preserving others [26, 34]. However, most studies focus on static unlearning, assuming a single-step forgetting task. Real-world applications, such as compliance with iterative AI safety policies (e.g., OpenAI's content guidelines) or evolving regulations like the EU AI Act, require incremental unlearning (IU) to process sequential deletion requests efficiently. Directly applying static methods to IU often leads to cumulative errors, model collapse, and incomplete forgetting [17, 48], underscoring the need for dedicated frameworks to ensure compliance in dynamic regulatory environments. + +# 2.2. Incremental Unlearning + +Existing MU methods, like parameter suppression [15, 16, 52] and knowledge confusion methods [10, 12, 17, 18, 29, 35, 61], lead to superficial forgetting due to the absence of direct constraints on the feature representation. In particular, the degraded features trap them in a dilemma between forgetting and remaining, which becomes even more severe when directly applied to IU. Meanwhile, some continual learning methods [6, 27], originally designed to retain old knowledge while dynamically learning new knowledge, also face inherent limitations in IU [35]. Previous work focusing on IU has primarily relied on toy models, e.g., linear models [13], or imposed heavy constraints, such as adding a class-specific synthetic signal in the pre-training stage [43], which is not feasible for PTMs. Moreover, while some focus on dynamically performing both continual learning and selective forgetting within learned knowledge [8, 28, 31], our approach differs in that we emphasize achieving deep forgetting when handling sequential unlearning tasks. That is, rather than alternating between learning and forgetting, the pre-trained model should completely forget unwanted knowledge while efficiently retaining the rest. + +# 3. Problem Statement + +Following [60, 61], we first formally define Machine Unlearning (MU), i.e., static unlearning scenario. We then extend this definition to Incremental Unlearning (IU), i.e., dynamic unlearning scenario, where deletion requests arrive sequentially. + +Let $\mathcal{M}$ be a pre-trained model trained on the dataset $\mathbf{D}$ . We denote the mapping relationship of the model as $f_{\mathcal{M}}(\mathcal{X}_{\mathbf{D}})\rightarrow \mathcal{Y}_{\mathbf{D}}$ , where $\mathcal{X}_{\mathbf{D}}$ and $\mathcal{Y}_{\mathbf{D}}$ represent the input set and output set, respectively. For MU, we partition $\mathbf{D}$ into $\mathbf{D}_{\mathbf{f}}$ (forgetting classes) and $\mathbf{D}_{\mathbf{r}}$ (remaining classes), under the constraint $|\mathbf{D}_{\mathbf{r}}| + |\mathbf{D}_{\mathbf{f}}|\ll |\mathbf{D}|$ . The forgetting algorithm $\mathcal{F}$ produces $\mathcal{M}' = \mathcal{F}(\mathcal{M},\mathbf{D}_{\mathbf{f}},\mathbf{D}_{\mathbf{r}})$ , altering the following mappings: + +$$ +f _ {\mathcal {M} ^ {\prime}} \left(\mathcal {X} _ {\mathbf {D} _ {\mathrm {f}}}\right) \not \longrightarrow \mathcal {Y} _ {\mathbf {D} _ {\mathrm {f}}}, f _ {\mathcal {M} ^ {\prime}} \left(\mathcal {X} _ {\mathbf {D} _ {\mathrm {r}}}\right) \longrightarrow \mathcal {Y} _ {\mathbf {D} _ {\mathrm {r}}}. \tag {1} +$$ + +Here, $\nrightarrow$ means the mapping relationship no longer holds. + +For IU, let there are a sequence of unlearning task, $\{\mathcal{T}_1,\mathcal{T}_2,\dots ,\mathcal{T}_T\}$ and $T$ is the number of unlearning tasks. And $\{\mathbf{D}_{f,t},\mathbf{D}_{r,t}\}$ is the forgetting classes and remaining classes belonging to $\mathcal{T}_t$ where $\mathbf{D}_{r,t} = \mathbf{D} - \cup_{i = 1}^{t}\mathbf{D}_{f,i}$ . Starting from $\mathcal{M}_0 = \mathcal{M}$ , the forgetting algorithm $\mathcal{F}$ iteratively processes till $t$ -th unlearning task $\mathcal{T}_t$ to generate $\mathcal{M}_t = \mathcal{F}(\mathcal{M}_{t - 1},\mathbf{D}_{f,t},\mathbf{D}_{r,t})$ which holds the followings: + +$$ +f _ {\mathcal {M} _ {t}} \left(\mathcal {X} _ {\mathbf {D} _ {f, i}}\right) \not \longrightarrow \mathcal {Y} _ {\mathbf {D} _ {f, i}}, f _ {\mathcal {M} _ {t}} \left(\mathcal {X} _ {\mathbf {D} _ {r, j}}\right) \longrightarrow \mathcal {Y} _ {\mathbf {D} _ {r, j}}, \tag {2} +$$ + +# 4. Methods + +In this section, we introduce FG-OrIU, a comprehensive framework for Incremental Unlearning (IU) in pre-trained models (PTM). By imposing orthogonal constraints on both feature representations and gradient updates, FG-OrIU ensures thorough removal of unwanted knowledge while preserving performance on remaining classes. An overview of the framework is shown in Figure 2. + +# 4.1. Feature Subspace Decomposition + +PTMs acquire implicit knowledge from pre-training data, benefiting downstream tasks during inference. Unlearning aims to selectively remove the influence of forgetting classes from the PTM [10]. Unlike explicit knowledge in traditional Knowledge Bases (KBs) [38, 44], which can be directly edited, the knowledge in PTMs is implicit and intertwined across all parameters [10, 34, 64]. Effective unlearning requires precisely identifying feature spaces for forgetting and remaining. To achieve this, we decompose features into two subspaces: one capturing forgetting-class contributions and the other preserving remaining-class knowledge. Specifically, we perform singular value decomposition (SVD) on the feature matrix of forgetting classes and use the span of its dominant singular vectors to define the forgetting subspace. Similarly, we apply the same procedure to the remaining classes to construct its subspace. We then detail the process as follows. + +Given a PTM, $\mathcal{M}$ , we construct a representation matrix $\pmb{R}_{f,i} = \left[\pmb{x}_{1,i},\dots,\pmb{x}_{n_{f_i},i}\right]\in \mathbb{R}^{n_{f_i}\times d}$ at each layer $l$ where columns correspond to representations of $n_{f_i}$ samples from the remaining classes $\mathbf{D}_{\mathbf{f},\mathbf{i}}$ , and $d$ is the output dimension3. SVD is applied to $\pmb{R}_{f,i} = \pmb{U}_{f,i}\pmb{\Sigma}_{f,i}(\pmb{V}_{f,i})^T$ , followed by its $k$ -rank approximation $(\pmb{R}_{f,i})_k$ satisfying $||(\pmb{R}_{f,i})_k||_F^2\geq \epsilon ||\pmb{R}_{f,i}||_F^2$ for a threshold $\epsilon$ . The subspace $\mathcal{S}_{f,i} = \mathrm{span}\{\pmb{B}_{f,i}\}$ is defining using the first $k$ vectors $\pmb{B}_{f,i} = \{\pmb{u}_{1,i},\dots,\pmb{u}_{k,i}\}$ from $\pmb{U}_{f,i}$ , capturing directions with the highest singular values. Similarly, the feature matrix $\pmb{R}_{r,i}$ for remaining classes $\mathbf{D}_{\mathbf{r},\mathbf{i}}$ defines the subspace $\mathcal{S}_{r,i} = \mathrm{span}\{\pmb{B}_{r,i}\}$ . The two sets of layer-wised feature subspace, $\{\mathcal{S}_{f,i}^1,\dots,\mathcal{S}_{f,i}^l\}$ and $\{\mathcal{S}_{r,i}^1,\dots,\mathcal{S}_{r,i}^l\}$ , represent the core part for forgetting and remaining classes, + +![](images/60c69cb922bd064d6bed1158b52380dd84e70f8643f9fe22e271c4f25a6466dc.jpg) +Figure 2. Overview of FG-OrIU: (a) For each unlearning task, we decompose the layer-wised- forgetting and remaining subspace separately (Sec. 4.1). (b) LoRA is inserted at each layer, with constraints applied at both the feature and gradient levels for forgetting and remaining (Sec. 4.2). (c) For new unlearning tasks, the forgetting subspace expands while the remaining contracts (Sec. 4.3). + +respectively, isolating features to be erased while preserving those to retain. By explicitly decomposing feature subspaces across all layers, we strengthen supervision signals for parameter updates and mitigate the risk of gradient obfuscation [2] in deep networks, where gradient inaccuracies caused by layer-wise nonlinear transformations could otherwise weaken the unlearning effect. This multi-layer constraint ensures consistent suppression of forgetting class features throughout the model hierarchy. + +# 4.2. Dual Orthogonal Projection + +Prior to unlearning task $\mathcal{T}_i$ , we freeze both the backbone and the classification head of the PTM and insert lightweight LoRA modules [23] at each layer. We then denote the model with LoRA inserted as $\mathcal{M}_{\theta}$ , where $\theta$ represents the trainable parameters in LoRA. After unlearning, the updates in LoRA, $\theta^{*}$ , will be merged into the initial weight of the backbone, $\mathcal{M}_i \xrightarrow[\text{merge}]{\theta^*} \mathcal{M}_{i+1}$ . + +Building on the feature subspaces $S_f$ and $S_r$ defined in Section 4.1, we propose a dual orthogonal projection mechanism that operates through both forward feature alignment and backward gradient correction. + +# 4.2.1. Feature Orthogonal Projection + +Inspired by [45, 55], for forgetting classes, achieving thorough removal at the feature level requires that the features generated by the unlearned model, $\mathcal{M}_{\theta}$ , deviate as much as possible from the feature subspace $S_{f}$ . Meanwhile, we aim to ensure that the unlearned model's feature subspace for remaining classes aligns with $S_{r}$ . The orthogonality constraint between $S_{f}$ and the unlearned features serves as a critical mechanism for facilitating class-specific forgetting: under a fixed classification head, features orthogonal to the original subspace $S_{f}$ induce maximal perturbations in class logits, as the inner product between orthogonal vectors diminishes to zero. This forces the unlearned model to suppress discriminative patterns associated with forgetting + +classes in its intermediate representations. To do so, we propose the following two regularization constraints. For forgetting classes, let $\boldsymbol{x}_f$ denote a batch of forgetting samples from $\mathbf{D}_{\mathbf{f},\mathrm{i}}$ , and $h_f^l$ denote the $l$ -th layer features gained with the unlearned model $\mathcal{M}_{\theta}$ . For complete forgetting, we enforce: + +$$ +\boldsymbol {h} _ {f} ^ {l} \perp \mathcal {S} _ {f, i} ^ {l} \quad \Rightarrow \quad \min | | \mathbf {P} _ {f} ^ {l} \boldsymbol {h} _ {f} ^ {l} | | _ {F} ^ {2} \tag {3} +$$ + +where $\mathbf{P}_f^l = \mathbf{B}_{f,i}^l (\mathbf{B}_{f,i}^l)^\top$ is the projection matrix for the forgetting subspace $S_{f,i}^{l}$ and $\| \cdot \| _F^2$ is the F-norm. Eq.(3) ensures that the features of $\mathbf{D}_{\mathbf{f},\mathbf{i}}$ gained from the unlearned model $\mathcal{M}_{\theta}$ lie outside the original feature subspace gained from $\mathcal{M}$ , erasing the unwanted information. Thus, we define the flop-f loss: + +$$ +\mathcal {L} _ {f o p - f} = \sum_ {l = 1} ^ {m} \left\| \mathbf {P} _ {f} ^ {l} \boldsymbol {h} _ {f} ^ {l} \right\| _ {F} ^ {2}. \tag {4} +$$ + +where $m$ denotes the total number of layers in the PTM. For remaining classes, we follow previous works and apply the standard cross-entropy loss by constraining the model's final output, as follows: + +$$ +\mathcal {L} _ {c e} = \mathcal {L} \left(\mathcal {X} _ {\mathbf {D} _ {r, 1}}, \mathcal {Y} _ {\mathbf {D} _ {r, 1}}\right), \tag {5} +$$ + +where $\mathcal{L}$ denotes the cross-entropy loss function. In addition, we also propose applying explicit constraint at the feature level for the remaining classes to better preserve the desired information. Specifically, $\pmb{x}_r$ denote a batch of remaining samples from $\mathbf{D}_{\mathbf{r},\mathbf{i}}$ , and $h_r^l$ denote the $l$ -th layer features gained with the unlearned model $\mathcal{M}_{\theta}$ . For better preservation, we maintain feature fidelity through: + +$$ +\min \left\| \boldsymbol {h} _ {r} ^ {l} - \mathbf {P} _ {r} ^ {l} \boldsymbol {h} _ {r} ^ {l} \right\| _ {F} ^ {2} \tag {6} +$$ + +where $\mathbf{P}_r^l = \mathbf{B}_{r,i}^l (\mathbf{B}_{r,i}^l)^\top$ is the projection matrix for the remaining feature subspace $S_{r,i}^{l}$ . + +Eq.(6) ensures that the features of $\mathbf{D}_{\mathbf{r},\mathbf{i}}$ gained from the unlearned model $\mathcal{M}_{\theta}$ , align with the original feature subspace gained from $\mathcal{M}$ , maintaining the better preservation. Therefore, We draw the following $f_{op - r}$ loss: + +$$ +\mathcal {L} _ {f o p - r} = \sum_ {l = 1} ^ {m} \left\| \boldsymbol {h} _ {r} ^ {l} - \mathbf {P} _ {r} ^ {l} \boldsymbol {h} _ {r} ^ {l} \right\| _ {F} ^ {2}. \tag {7} +$$ + +Finally, the total loss is: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {c e} + \lambda_ {1} \mathcal {L} _ {f o p - f} + \lambda_ {2} \mathcal {L} _ {f o p - r}, \tag {8} +$$ + +where $\lambda_{1},\lambda_{2}$ are the coefficients during training. + +# 4.2.2. Gradient Orthogonal Projection + +To prevent parameter updates from reintroducing forgotten knowledge while preserving retained features, we design a gradient correction mechanism inspired by gradient projection techniques in multi-task learning [56] and continual learning [40, 62]. Given the raw gradient $g_{l}$ at layer $l$ , we apply the following modifications: + +$$ +\begin{array}{l} \hat {g} _ {l} = g _ {l} - g _ {l} ^ {\prime} - g _ {l} ^ {\prime \prime} \tag {9} \\ = \boldsymbol {g} _ {l} - \underbrace {\boldsymbol {g} _ {l} (\boldsymbol {I} - \mathbf {B} _ {f , i} ^ {l} (\mathbf {B} _ {f , i} ^ {l}) ^ {\top})} _ {(t e r m 1: g o p - f)} - \underbrace {\boldsymbol {g} _ {l} \mathbf {B} _ {r , i} ^ {l} (\mathbf {B} _ {r , i} ^ {l}) ^ {\top}} _ {(t e r m 2: g o p - r)}, \\ \end{array} +$$ + +Term (gop-f), $g_{l}^{\prime}$ : Removing the components of $g_{l}$ that are orthogonal to $S_{f,i}^{l}$ not only amplifies updates focusing on directions most critical for encoding forgetting class features, but also blocks parameter updates that could reintroduce removed knowledge. + +Term (gop-r), $g_{l}^{\prime \prime}$ : We aim to ensure that forgetting specific classes does not affect the remaining classes. Formally, for the unlearned model $f_{\mathcal{M}_{\theta}}$ and $x_{r}$ from the remaining classes $\mathbf{D}_{\mathbf{r},\mathbf{i}}$ , the following equation holds: $f_{\mathcal{M}_{\theta}}(x_{r}) = f_{\mathcal{M}}(x_{r})$ . Specifically, we consider the input and output of layer $l$ as follows: + +$$ +\begin{array}{l} \boldsymbol {W} _ {l} \boldsymbol {x} _ {r} + \Delta \theta_ {l} \boldsymbol {x} _ {r} = \boldsymbol {W} _ {l} \boldsymbol {x} _ {r} \\ \Delta \theta_ {l} \boldsymbol {x} _ {\boldsymbol {r}} = \mathbf {0} \tag {10} \\ \end{array} +$$ + +where $\Delta \theta_{l}$ represents the change in the LoRA parameters inserted at layer $l$ , which corresponds to its gradient. $W_{l}$ is the parameter at layer $l$ from the pre-trained model, and these parameters remain frozen during unlearning. + +Eq.(10) indicates that if the update direction of the parameters, i.e., the gradient of the trainable LoRA parameters, lies in the complementary space of $S_{r,i}^{l}$ , i.e., $\pmb{g}_{l} \perp S_{r,i}^{l}$ , then $\Delta \theta \pmb{x}_r = \eta \pmb{g}_l \pmb{x}_r = \mathbf{0}$ (where $\eta$ is the learning rate), ensuring that the unlearned model's performance on the remaining classes is not affected. Therefore, we further refine the gradient by removing its components along $S_{r,i}^{l}$ . + +Finally, modified gradient $\hat{g}_l$ serves the real gradient for parameter updates at the layer $l$ . + +# 4.3. Dynamic Subspace Adaptation + +When encountering a new unlearning task $\mathcal{T}_{i + 1}$ , our method only needs to update the two decomposed feature subspaces. Since the forgetting classes $\mathbf{D_{f,i + 1}}$ in $\mathcal{T}_{i + 1}$ are a subset of the remaining classes from the previous task $\mathcal{T}_i$ , i.e., $\mathbf{D_{f,i + 1}}\subseteq \mathbf{D_{r,i}}$ , the feature subspaces can be efficiently adjusted. For simplicity, we omit the notion $l$ of layers. + +We first construct a representation matrix $\mathbf{R}_{f,i+1}$ using only the forgetting classes $\mathbf{D}_{\mathbf{f},i+1}$ . Before applying SVD and $k$ -rank approximation, we remove common bases already present in $S_{f,i}$ to ensure newly added bases are unique and orthogonal to existing ones: + +$$ +\begin{array}{l} \hat {\boldsymbol {R}} _ {f, i + 1} = \boldsymbol {R} _ {f, i + 1} - \boldsymbol {B} _ {f, i} \left(\boldsymbol {B} _ {f, i}\right) ^ {T} \left(\boldsymbol {R} _ {f, i + 1}\right) (11) \\ = \boldsymbol {R} _ {f, i + 1} - \boldsymbol {R} _ {f, i + 1} ^ {\text {p r o j}}. (12) \\ \end{array} +$$ + +Next, we perform SVD on $\hat{R}_{f,i+1} = \hat{U}_{f,i+1}\hat{\Sigma}_{f,i+1}(\hat{V}_{f,i+1})^T$ and select $h$ new orthogonal from $\hat{U}_{f,i+1}$ . Then, the feature subspace of forgetting classes is expanded as $\mathcal{S}_{f,i+1} = \text{span}\{B_{f,i+1}\}$ , where $\mathbf{B}_{f,i+1} = [B_{f,i};\mathbf{u}_1,\dots,\mathbf{u}_h]$ with $h$ new bases from $\hat{U}_{f,i+1}$ . And the feature subspace $\mathcal{S}_{r,i+1}$ of remaining classes is contracted to remove the overlapping features accordingly. + +# 5. Experiment + +# 5.1. Experimental Setup + +# Datasets, Benchmarks and Pre-trained Models. + +Following [60], we evaluate FG-OrIU on Classification and Recognition tasks. For classification, we use ImageNet100 [39], CUB-200 [47], and Omnibenchmark [59], with a ViT-B/16 [14] pre-trained in PyTorch [33]. For face recognition, we use CASIA-Face100 [60], a subset of CASIA-WebFace [54], and MS-Celeb-100, sampled from [21]. The pre-trained model is the Face Transformer from [63]. We focus on class-unlearning scenarios with two benchmarks: (a) MU, varying numbers of classes are removed; (b) IU, 20 classes are removed per task over 4 tasks. All setting is in the form of Y-X, which means experiments start from a PTM (Y classes originally) and forget X classes. More details about implementations are in Appendix ?? + +Metrics. Following [60], we evaluate performance using average accuracy (Acc) for both forgetting and remaining classes. The accuracy for forgetting classes should approach zero, while the accuracy for remaining classes should remain close to the original model's performance. Additionally, we adopt the H-Mean [43] metric to assess overall performance after learning task $\mathcal{T}_t$ , computed as: + +$$ +H - M e a n ^ {(t)} = \frac {2 A c c _ {r} ^ {(t)} \cdot D r o p ^ {(t)}}{A c c _ {r} ^ {(t)} + D r o p ^ {(t)}}. \tag {13} +$$ + +Here $Acc_{r}^{(t)}$ represents accuracy on the retained dataset after task $\mathcal{T}_t$ , while $Drop^{(t)} = Acc_{f}^{(t - 1)} - Acc_{f}^{(t)}$ mea + +
MethodsTunable Ratio ↓100-5100-10100-50100-90
H ↑Accr ↑Accf ↓H ↑Accr ↑Accf ↓H ↑Accr ↑Accf ↓H ↑Accr ↑Accf ↓
Pre-train--73.8570.88-73.8172.74-72.4574.88-72.3273.86
L2*99.73%67.1366.963.5867.7464.871.8664.3956.610.2452.4540.680.04
EWC*99.73%69.6568.690.2469.1665.920.0062.4853.610.0044.4131.750.00
MAS*99.73%69.7369.290.7269.3866.410.1262.6753.880.0046.7934.240.00
LwF99.73%67.9568.550.0070.0867.620.0063.8155.590.0061.2552.320.00
DER99.73%69.7068.550.0070.0867.620.0063.8155.590.0057.0346.440.00
DER++99.73%69.7068.560.0070.5868.550.0064.6156.820.0061.4052.540.00
FDR99.73%70.3169.740.0070.0367.510.0065.4058.040.0053.8542.370.00
Retrain100.00%16.499.330.0018.0210.280.0019.7111.350.0046.4733.900.00
SCRUB99.73%67.7864.940.0068.2664.310.0065.8258.710.0016.119.040.00
SCRUB-S99.73%70.2969.950.2471.1969.810.1257.4951.209.3417.539.940.00
LIRF*50.66%25.5667.6755.1326.3565.8356.2647.1358.9535.6254.4944.293.06
GS-LoRA1.28%71.0271.160.0071.7670.810.0071.2968.050.0273.7173.560.00
GS-LoRA++1.28%71.1271.360.0072.0471.350.0071.5668.520.0072.8571.860.00
FG-OrIU1.28%71.7872.710.0072.5272.310.0074.1873.510.0075.4977.150.00
+ +Table 2. Static Machine Unlearning results for face recognition task on CASIA-Face100. Accr (%) and Accf (%) are the accuracies of remaining and forgetting classes. * denotes the original methods with a rehearsal buffer. Retrain represents retraining the model using replay data and the training epoch is the same as other methods to ensure a fair comparison. Pre-train denotes the results before forgetting. + +![](images/4aa4a42371b07dd43463877fa058fb00049a3065512c6495781bb452334269bc.jpg) +Figure 3. Accuracy on forgetting classes when recovering on CASIA-Face100. The line (Pre-train) is the result before unlearning. The line (FG-OrIU) is Our method. + +sures the performance drop on forgetting classes before and after training. After unlearning task $\mathcal{T}_t$ , we evaluate performance on all previously forgetting classes from tasks $\mathcal{T}_1, \mathcal{T}_2, \dots, \mathcal{T}_{t-1}$ . + +Baselines. Following [60], we compare FG-OrIU against (a) continual learning (CL) methods, including L2 regularization, EWC [27], MAS [1], LwF [30], DER [6], and FDR [3], as well as (b) machine unlearning (MU) methods such as BAD-T [12], LIRF [53], SCRUB/SCRUB-S [29], GS-LoRA/GS-LoRA++ [60, 61], and retraining. To ensure forgetting occurs in the backbone, we freeze the classification head for all methods. In the retraining setup, we train a randomly initialized model using replay data, ensuring the number of training epochs matches that of other methods. + +# 5.2. Results and Comparison + +Table 2 presents the results for static Machine Unlearning task. FG-OrIU exhibits significantly degraded performance on forgetting classes while maintaining the perfor + +mance of remaining classes. For example, in the settings where 50 and 90 classes are removed per task, FG-OrIU achieves improvements of $2.62\%$ and $2.64\%$ in $H$ , and $4.99\%$ and $5.29\%$ in $ACC_{r}$ compared to GS-LoRA. Notably, FG-OrIU keeps the pre-trained model frozen and only trains the LoRA modules, enabling effective and efficient forgetting. Table 3 and 4 show the results for Incremental Unlearning task. FG-OrIU consistently outperforms all methods. Specifically, CL methods struggle to achieve effective forgetting on ImageNet100, as seen in DER, which retains high $ACC_{f}$ of $10.40\%$ , $10.90\%$ , $8.80\%$ , and $3.40\%$ across four tasks, respectively. On the other hand, MU methods fail to maintain their forgetting effect over sequential unlearning tasks. For example, LIRF and SCRUB-S show an increase in $ACC_{o}$ on CASIA-Face100, which becomes more pronounced as the sequence of unlearning tasks grows. GS-LoRA is specifically designed for IU. Compared to it, our method achieves better forgetting while also demonstrating superior performance in preserving the remaining classes, i.e., substantial improvements in $ACC_{r}$ , particularly in long-sequence unlearning scenarios, achieving gains of $5.15\%$ and $4.6\%$ in the fourth unlearning task on two different datasets, respectively. + +# 5.3. Ablation Study + +Our FG-OrIU consists of four components, i.e., feature orthogonal projection for forgetting, feature orthogonal projection for remaining, gradient orthogonal projection for forgetting, and gradient orthogonal projection for remaining, denoted as $fop - f$ , $fop - r$ , $gop - f$ , and $gop - r$ . We independently remove each of these components to demonstrate their individual contributions. We observe that when the regularization term or gradient modification related to remaining is removed, the performance on the remaining + +
MethodsTask 1(100-20)Task 2(80-20)Task 3(60-20)Task 4(40-20)
H↑Accr↑Accf↓H↑Accr↑Accf↓Acco↓H↑Accr↑Accf↓Acco↓H↑Accr↑Accf↓Acco↓
Pre-train-74.3174.65-74.5073.80--74.8073.91--74.4775.11-
L2*66.9162.132.1666.6661.741.379.3666.4261.371.5411.8366.9561.020.978.00
EWC*67.7161.950.0067.7162.550.000.0067.0961.430.000.1068.0262.140.000.23
MAS*67.5261.630.0068.1263.250.000.0068.1563.230.000.0367.7061.610.000.00
LwF69.4365.040.2169.9466.470.000.2170.3467.110.000.0070.8967.120.000.04
DER70.7567.240.0068.8864.580.000.0068.9564.620.000.0069.4164.510.000.00
DER++70.0165.900.0068.9964.770.000.0069.9666.420.000.0069.8565.280.000.00
FDR68.2962.920.0067.3962.010.000.0069.1664.990.000.0072.5170.080.000.00
Retrain17.299.770.0031.1219.720.000.0041.1728.801.720.0052.4441.774.660.09
BAD-T68.9063.970.0069.5665.950.210.0070.2166.870.000.3472.7170.460.000.04
LIRF*30.5964.4654.6034.0562.0350.3443.5044.5662.5339.2936.5840.3662.6245.3427.96
SCRUB70.3966.590.0070.5567.850.320.0071.0168.330.000.4473.3671.680.000.05
SCRUB-S72.4170.340.0570.0671.535.1615.2273.3873.090.2415.8175.3376.240.686.47
GS-LoRA74.4074.160.0073.5973.370.000.0574.3674.880.060.0073.7672.450.001.93
GS-LoRA++73.9774.160.0073.4873.160.000.1074.2074.510.000.0075.2375.360.000.00
FG-OrIU74.4574.250.0074.6475.500.000.0075.4977.140.000.0077.7180.510.000.00
+ +Table 3. Incremental Unlearning results for face recognition task on CASIA-Face100. $Acc_{o}(\%)$ is the accuracy of old tasks, i.e., the accuracy on all previously forgetting classes in task $\{\mathcal{T}_1,\dots ,\mathcal{T}_{t - 1}\}$ . With four tasks, each forgetting 20 classes, desired forgetting should consider both $Acc_{r}$ and $Acc_{f}$ . We bold the best and underline the second-best $H$ results. + +
MethodsTask 1(100-20)Task 2(80-20)Task 3(60-20)Task 4(40-20)
H↑Accr↑Accf↓H↑Accr↑Accf↓Acco↓H↑Accr↑Accf↓Acco↓H↑Accr↑Accf↓Acco↓
Pre-train-89.9387.40-90.0389.60--89.6590.80--92.8086.50-
L2*79.3988.3315.3083.5988.3710.3048.1084.0188.2010.6037.1582.7289.909.9020.97
EWC*82.6978.550.1085.7182.570.500.2087.3384.900.902.9086.9389.502.008.27
MAS*80.4474.500.0081.5174.770.000.6082.1875.050.000.6081.9377.900.101.77
LwF85.8285.150.9087.2685.400.402.4087.3584.150.002.3586.0585.700.101.07
DER82.8889.7310.4084.2290.5710.9050.8086.3691.208.8042.0088.5294.703.4024.90
DER++83.4589.739.4084.1390.6311.1050.1087.5091.306.8040.4089.2494.802.2020.80
FDR29.5117.780.5036.2122.700.100.0040.7926.300.000.0047.9133.502.401.27
Retrain48.2033.280.0056.2841.301.300.0061.1447.555.200.2067.2360.6011.000.53
BAD-T74.7465.290.0078.0969.210.000.1076.6566.320.000.0076.2568.170.000.00
LIRF*52.7264.2842.7161.0569.8335.3632.1560.9660.3229.1830.9158.5168.4635.4121.04
SCRUB75.8767.030.0076.2366.330.000.0075.4664.550.000.0077.5670.300.000.00
SCRUB-S83.4479.830.0082.5976.600.000.1079.8171.200.000.0078.2371.400.000.07
GS-LoRA86.1385.000.1087.3385.270.101.1087.8485.150.100.3586.3086.300.200.07
GS-LoRA++86.6385.980.1087.2285.230.300.3087.4884.400.000.6087.0087.500.000.07
FG-OrIU87.0186.730.1088.4587.430.100.1088.9087.350.300.0089.1592.100.100.00
+ +Table 4. Incremental Unlearning results for classification task on ImageNet100. + +
VariantsTask 1(100-20)Task 2(80-20)
Accr↑Accf↓Accr↑Accf↓Acco↓
Pre-train74.3074.6074.5073.80-
FG-OrIU74.250.0075.500.000.00
w/o fop-f73.010.2076.240.421.28
w/oGOP-f73.920.1076.584.370.82
w/o fop-r73.920.0574.980.210.15
w/oGOP-r71.450.0073.960.000.00
+ +classes significantly declines. Meanwhile, when the regularization term or gradient modification related to forgetting is removed, the performance on the forgetting classes shows a slight improvement. Thus, each component provides clear benefits in achieving a better trade-off between forgetting and remaining while preserving the old forgetting effect. + +# 5.4. Discussion + +IU is designed to ensure strict privacy protection. This raises several key questions: (a) Does the unlearned model + +Table 5. Ablation studies on CASIA-Face100. + +
ExpsMetircPre-trainDERFDRSCRUBGS-LoRAOurs
Exp1SSIM0.940.600.340.750.390.01
PSNR31.3221.2216.8923.4017.608.37
Exp2SSIM0.950.490.660.630.380.02
PSNR31.1519.1722.1521.1217.328.72
Exp3SSIM0.930.530.490.600.410.01
PSNR31.1020.0319.8322.0115.928.55
+ +Table 6. SSIM $\uparrow$ and PSNR(db) $\uparrow$ scores for the reconstructed images using DIP. Here, Pre-train serves as the upper bound. + +still retain any private information? (b) Can the forgetting effect be reversed? and (c) How does unlearning impact the feature subspace? To address these questions, we design the following experiment. + +# 5.4.1. Semantic Residual Analysis + +We first investigate the semantic information contained in features after unlearning. We utilize Deep Image Prior (DIP) [46] to reconstruct an image from the features ob + +![](images/ff79610a9d205a4eee421c6e9612ec41dc8c28e591552eb48da22e0236ca7245.jpg) +(a) Pre-train + +![](images/6f2b569fb9a2a2172283ea05658bc1fa68c98ff58a4b73c01c589025f16ea78e.jpg) +(b) Ours +Figure 4. TSNE visualization with 5 forgetting classes and 10 remaining classes, using different model as feature extractor: (a) the pretrained model, (b) the unlearned model through FG-OrIU, and (c) through GS-LoRA. + +![](images/3c353ae144c0369e8c264f890b96d1fcd642199b14b1c315ab8990ef3f35ca47.jpg) +(c) GS-LoRA + +tained from the last layer of the unlearned model. Figure 1 presents the visualization of the reconstructed image, which allows us to infer the semantic information encoded in features. More results are provided in the Appendix ?? + +Additionally, we measure the SSIM and PSNR scores for the images generated by DIP from the unlearned model using different IU methods as a quantitative results. As shown in Table 6, the SSIM values of the reconstructed images from the compared methods range between 0.35 and 0.75, with the highest being 0.75 for SCRUB on $exp1$ , while the PSNR values range between 16 and 24, with the highest being 23.40 for SCRUB on $exp1$ . This indicates that the reconstructed image achieves a high similarity with the original image, suggesting that the information in features has been partially preserved. In contrast, our method yields an SSIM of only 0.01 and a PSNR of 8.55, indicating a significant difference between the reconstructed and original images. This suggests that the semantic information at the feature level has been substantially obfuscated. + +# 5.4.2. Forgetting Recoverable Analysis + +Given an unlearned model obtained through different IU methods, we freeze its backbone and retrain a new classification head using the entire training dataset, which includes both forgetting and remaining classes. As illustrated in Figure 3, MLP* represents a trivial unlearning method where the classification head's parameters for the forgetting classes are simply set to zero. Consequently, the underlying features remain intact, enabling effortless recovery of the forgotten knowledge. Additionally, existing IU methods primarily impose constraints on the final output rather than directly modifying feature representations. As a result, while features may be weakened to some extent, their core structure remains largely unchanged. Thus, these methods allow the forgotten knowledge to be recovered to $15\% -70\%$ of its original performance, which leads to superficial forgetting. In contrast, our method achieves deep forgetting, making recovery infeasible. + +# 5.4.3. Feature Subspace Analysis + +To analyze the impact of direct constraints on features and gradients for forgetting and remaining classes, we conduct an experiment. We remove 20 classes while retaining 80, then extract features from 5 forgetting and 10 remaining classes and visualize them using t-SNE. + +A key observation is that GS-LoRA causes some forgetting class samples to move closer to or even overlap with certain remaining class samples (e.g., F21, $\mathrm{F}74 \rightarrow \mathrm{R}92$ in c). This suggests that merely applying constraints at the final output, resulting in degraded features from the forgetting classes entangling with those of the remaining classes. Consequently, this degradation negatively impacts the performance of the remaining classes. In contrast, our method enforces forgetting at both feature and gradient level, ensuring that features from the forgetting classes do not shift toward any remaining class but instead become linearly inseparable within the forgetting set. Notably, an unexpected yet beneficial side effect emerges: in the original PTM, certain remaining class samples (e.g., R59) were slightly entangled with forgetting classes (e.g., F90 and F46). However, after applying our method, these classes become completely disentangled. This observation aligns with the results in Table 3, where the accuracy of the remaining classes improves after unlearning (e.g., $75.36 \rightarrow 80.51$ ), demonstrating that our method not only enhances forgetting effectiveness but also benefits the generalization of the remaining classes. + +# 6. Conclusion + +In this work, we propose FG-OrIU for incremental unlearning in vision PTMs. By enforcing orthogonal constraints on both features and gradients, FG-OrIU achieves irreversible forgetting while maintaining performance on the remaining classes. Its dynamic subspace adaptation ensures robust handling of sequential deletion requests. Extensive experiments confirm its superiority. While FG-OrIU is designed for class-level unlearning, extending it to sample-level settings remains a promising direction for future work. + +# Acknowledgments + +This work was supported in part by National Natural Science Foundation of China under Grant 62402430, 62206248, 62476238, Zhejiang Provincial Natural Science Foundation of China under Grant LQN25F020008, and Aeronautical Science Foundation of China 20240048076001. Qian Feng would like to extend his gratitude to Jiahang Tu and Hanbin Zhao from Zhejiang University for their discussions regarding the methods and experiments. Additionally, he is thankful to Mintong Kang from UIUC for his dedicated suggestions on improving the overall writing of the paper. + +# References + +[1] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In Proceedings of the European conference on computer vision (ECCV), pages 139-154, 2018. 6 +[2] Anish Athalye, Nicholas Carlini, and David Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In International conference on machine learning, pages 274-283. PMLR, 2018. 4 +[3] Ari S Benjamin, David Rolnick, and Konrad Kording. Measuring and regularizing networks in function space. arXiv preprint arXiv:1805.08289, 2018. 6 +[4] Jacopo Bonato, Marco Cotogni, and Luigi Sabetta. Is retain set all you need in machine unlearning? restoring performance of unlearned models with out-of-distribution images. In European Conference on Computer Vision, pages 1-19. Springer, 2024. 1 +[5] Lucas Bourtoule, Varun Chandrasekaran, Christopher A Choquette-Choo, Hengrui Jia, Adelin Travers, Baiwu Zhang, David Lie, and Nicolas Papernot. Machine unlearning. In 2021 IEEE symposium on security and privacy (SP), pages 141-159. IEEE, 2021. 1, 2 +[6] Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark experience for general continual learning: a strong, simple baseline. Advances in neural information processing systems, 33:15920-15930, 2020. 3, 6 +[7] Yinzhi Cao and Junfeng Yang. Towards making systems forget with machine unlearning. In 2015 IEEE symposium on security and privacy, pages 463-480. IEEE, 2015. 1 +[8] Romit Chatterjee, Vikram Chandawat, Ayush Tarun, Ankur Mali, and Murari Mandal. A unified framework for continual learning and machine unlearning. arXiv preprint arXiv:2408.11374, 2024. 1, 3 +[9] Jiaao Chen and Diyi Yang. Unlearn what you want to forget: Efficient unlearning for llms. arXiv preprint arXiv:2310.20150, 2023. +[10] Xinwen Cheng, Zhehao Huang, Wenxin Zhou, Zhengbao He, Ruikai Yang, Yingwen Wu, and Xiaolin Huang. Remaining-data-free machine unlearning by suppressing sample contribution. arXiv preprint arXiv:2402.15109, 2024. 1, 2, 3 + +[11] Zhi-Yi Chin, Chieh-Ming Jiang, Ching-Chun Huang, Pin-Yu Chen, and Wei-Chen Chiu. Prompting4debugging: Red-teaming text-to-image diffusion models by finding problematic prompts. arXiv preprint arXiv:2309.06135, 2023. 1, 3 +[12] Vikram S Chandawat, Ayush K Tarun, Murari Mandal, and Mohan Kankanhalli. Can bad teaching induce forgetting? unlearning in deep networks using an incompetent teacher. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7210-7217, 2023. 1, 2, 3, 6 +[13] Meng Ding, Jinhui Xu, and Kaiyi Ji. Why fine-tuning struggles with forgetting in machine unlearning? theoretical insights and a remedial approach. arXiv preprint arXiv:2410.03833, 2024. 3 +[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.5 +[15] Chongyu Fan, Jiancheng Liu, Yihua Zhang, Eric Wong, Dennis Wei, and Sijia Liu. Salun: Empowering machine unlearning via gradient-based weight saliency in both image classification and generation. arXiv preprint arXiv:2310.12508, 2023. 2, 3 +[16] Jack Foster, Stefan Schoepf, and Alexandra Brintrup. Fast machine unlearning without retraining through selective synaptic dampening. In Proceedings of the AAAI conference on artificial intelligence, pages 12043-12051, 2024. 1, 2, 3 +[17] Chongyang Gao, Lixu Wang, Chenkai Weng, Xiao Wang, and Qi Zhu. Practical unlearning for large language models. arXiv preprint arXiv:2407.10223, 2024. 1, 2, 3 +[18] Laura Graves, Vineel Nagisetty, and Vijay Ganesh. Amnesiac machine learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11516-11524, 2021. 2, 3 +[19] Keltin Grimes, Collin Abidi, Cole Frank, and Shannon Gallagher. Gone but not forgotten: Improved benchmarks for machine unlearning. arXiv preprint arXiv:2405.19211, 2024. 1 +[20] Chuan Guo, Tom Goldstein, Awni Hannun, and Laurens Van Der Maaten. Certified data removal from machine learning models. arXiv preprint arXiv:1911.03030, 2019. 1 +[21] Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. Ms-celeb-1m: A dataset and benchmark for large-scale face recognition. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 87-102. Springer, 2016. 5 +[22] Zeyu Han, Chao Gao, Jinyang Liu, Jeff Zhang, and Sai Qian Zhang. Parameter-efficient fine-tuning for large models: A comprehensive survey. arXiv preprint arXiv:2403.14608, 2024. 1 +[23] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 4 +[24] Zhehao Huang, Xinwen Cheng, JingHao Zheng, Haoran Wang, Zhengbao He, Tao Li, and Xiaolin Huang. Unified + +gradient-based machine unlearning with remain geometry enhancement. arXiv preprint arXiv:2409.19732, 2024. 1 +[25] Joel Jang, Dongkeun Yoon, Sohee Yang, Sungmin Cha, Moontae Lee, Lajanugen Logeswaran, and Minjoon Seo. Knowledge unlearning for mitigating privacy risks in language models. arXiv preprint arXiv:2210.01504, 2022. 1 +[26] Jinghan Jia, Jiancheng Liu, Parikshit Ram, Yuguang Yao, Gaowen Liu, Yang Liu, Pranay Sharma, and Sijia Liu. Model sparsity can simplify machine unlearning. Advances in Neural Information Processing Systems, 36:51584-51605, 2023. 1, 3 +[27] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 3, 6 +[28] Vishnuprasadh Kumaravelu, Sayanta Adhikari, and PK Srijith. Uncle: An unlearning framework for continual learning. 1, 3 +[29] Meghdad Kurmanji, Peter Triantafillou, Jamie Hayes, and Eleni Triantafillou. Towards unbounded machine unlearning. Advances in neural information processing systems, 36: 1957-1987, 2023. 2, 3, 6 +[30] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 6 +[31] Bo Liu, Qiang Liu, and Peter Stone. Continual learning and private unlearning. In Conference on Lifelong Learning Agents, pages 243-254. PMLR, 2022. 1, 3 +[32] Gaoyang Liu, Xiaogiang Ma, Yang Yang, Chen Wang, and Jiangchuan Liu. Federaser: Enabling efficient client-level data removal from federated learning models. In 2021 IEEE/ACM 29th International Symposium on Quality of Service (IWQOS), pages 1-10. IEEE, 2021. 2 +[33] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5 +[34] Laiqiao Qin, Tianqing Zhu, Linlin Wang, and Wanlei Zhou. Machine unlearning on pre-trained models by residual feature alignment using lora. arXiv preprint arXiv:2411.08443, 2024. 1, 2, 3 +[35] Sadia Qureshi, Thanveer Shaik, Xiaohui Tao, Haoran Xie, Lin Li, Jianming Yong, and Xiaohua Jia. Exploring incremental unlearning: Techniques, challenges, and future directions. arXiv preprint arXiv:2502.16708, 2025. 1, 3 +[36] Protection Regulation. General data protection regulation. Intouch, 25:1-5, 2018. 1 +[37] Jeffrey Rosen. The right to be forgotten. *Stan. L. Rev. Online*, 64:88, 2011. 1 +[38] Dan Roth and Wen-tau Yih. Probabilistic reasoning for entity & relation recognition. In COLING 2002: The 19th International Conference on Computational Linguistics, 2002. 3 + +[39] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 5 +[40] Gobinda Saha, Isha Garg, and Kaushik Roy. Gradient projection memory for continual learning. arXiv preprint arXiv:2103.09762, 2021. 2, 5 +[41] Thanveer Shaik, Xiaohui Tao, Haoran Xie, Lin Li, Xiaofeng Zhu, and Qing Li. Exploring the landscape of machine unlearning: A comprehensive survey and taxonomy. IEEE Transactions on Neural Networks and Learning Systems, 2024. 1 +[42] Yingdan Shi and Ren Wang. Redefining machine unlearning: A conformal prediction-motivated approach. arXiv preprint arXiv:2501.19403, 2025. 1 +[43] Takashi Shibata, Go Irie, Daiki Ikami, and Yu Mitsuzumi. Learning with selective forgetting. In *IJCAI*, page 6, 2021. 3, 5 +[44] Mihai Surdeanu and Heng Ji. Overview of the english slot filling track at the tac2014 knowledge base population evaluation. In Proc. Text Analysis Conference (TAC2014), 2014. 3 +[45] Jiahang Tu, Qian Feng, Chufan Chen, Jiahua Dong, Hanbin Zhao, Chao Zhang, and Hui Qian. Ce-sdwv: Effective and efficient concept erasure for text-to-image diffusion models via a semantic-driven word vocabulary. arXiv preprint arXiv:2501.15562, 2025. 1, 2, 3, 4 +[46] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9446-9454, 2018. 1, 2, 7 +[47] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5 +[48] Cheng-Long Wang, Qi Li, Zihang Xiang, and Di Wang. Has approximate machine unlearning been evaluated properly? from auditing to side effects. arXiv e-prints, pages arXiv-2403, 2024. 1, 2, 3 +[49] Heng Xu, Tianqing Zhu, Lefeng Zhang, Wanlei Zhou, and Philip S Yu. Machine unlearning: A survey. ACM Computing Surveys, 56(1):1-36, 2023. 2 +[50] Jie Xu, Zihan Wu, Cong Wang, and Xiaohua Jia. Machine unlearning: Solutions and challenges. IEEE Transactions on Emerging Topics in Computational Intelligence, 2024. 1 +[51] Lingling Xu, Haoran Xie, Si-Zhao Joe Qin, Xiaohui Tao, and Fu Lee Wang. Parameter-efficient fine-tuning methods for pretrained language models: A critical review and assessment. arXiv preprint arXiv:2312.12148, 2023. 1 +[52] Haonan Yan, Xiaoguang Li, Ziyao Guo, Hui Li, Fenghua Li, and Xiaodong Lin. Arcane: An efficient architecture for exact machine unlearning. In *IJCAI*, page 19, 2022. 2, 3 +[53] Jingwen Ye, Yifang Fu, Jie Song, Xingyi Yang, Songhua Liu, Xin Jin, Mingli Song, and Xinchao Wang. Learning with recoverable forgetting. In ECCV, 2022. 6 +[54] Dong Yi, Zhen Lei, Shengcai Liao, and Stan Z Li. Learning face representation from scratch. arXiv preprint arXiv:1411.7923, 2014. 5 + +[55] Jaehong Yoon, Shoubin Yu, Vaidehi Patil, Huaxiu Yao, and Mohit Bansal. Safree: Training-free and adaptive guard for safe text-to-image and video generation. arXiv preprint arXiv:2410.12761, 2024. 2, 4 +[56] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. Advances in Neural Information Processing Systems, 33:5824-5836, 2020. 2, 5 +[57] Dan Zhang, Tao Feng, Lilong Xue, Yuandong Wang, Yuxiao Dong, and Jie Tang. Parameter-efficient fine-tuning for foundation models. arXiv preprint arXiv:2501.13787, 2025. 1 +[58] Haibo Zhang, Toru Nakamura, Takamasa Isohara, and Kouichi Sakurai. A review on machine unlearning. SN Computer Science, 4(4):337, 2023. 2 +[59] Yuanhan Zhang, Zhenfei Yin, Jing Shao, and Ziwei Liu. Benchmarking omni-vision representation through the lens of visual realms. In European Conference on Computer Vision, pages 594-611. Springer, 2022. 5 +[60] Hongbo Zhao, Bolin Ni, Junsong Fan, Yuxi Wang, Yuntao Chen, Gaofeng Meng, and Zhaoxiang Zhang. Continual forgetting for pre-trained vision models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28631-28642, 2024. 1, 3, 5, 6 +[61] Hongbo Zhao, Fei Zhu, Bolin Ni, Feng Zhu, Gaofeng Meng, and Zhaoxiang Zhang. Practical continual forgetting for pretrained vision models. arXiv preprint arXiv:2501.09705, 2025. 1, 2, 3, 6 +[62] Zhen Zhao, Zhizhong Zhang, Xin Tan, Jun Liu, Yanyun Qu, Yuan Xie, and Lizhuang Ma. Rethinking gradient projection continual learning: Stability/plasticity feature space decoupling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3718-3727, 2023. 2, 5 +[63] Yaoyao Zhong and Weihong Deng. Face transformer for recognition. arXiv preprint arXiv:2103.14803, 2021. 5 +[64] Chen Zhu, Ankit Singh Rawat, Manzil Zaheer, Srinadh Bhojanapalli, Daliang Li, Felix Yu, and Sanjiv Kumar. Modifying memories in transformer models. arXiv preprint arXiv:2012.00363, 2020. 3 \ No newline at end of file diff --git a/2025/FG-OrIU_ Towards Better Forgetting via Feature-Gradient Orthogonality for Incremental Unlearning/images.zip b/2025/FG-OrIU_ Towards Better Forgetting via Feature-Gradient Orthogonality for Incremental Unlearning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a3f09769f4e738ce541a4f1cafc89c61ba726e2a --- /dev/null +++ b/2025/FG-OrIU_ Towards Better Forgetting via Feature-Gradient Orthogonality for Incremental Unlearning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:abcaa2b3c95d2e078c8e4a9935c6d41881c44ed4eb9d8676175f0e444c0f5844 +size 749195 diff --git a/2025/FG-OrIU_ Towards Better Forgetting via Feature-Gradient Orthogonality for Incremental Unlearning/layout.json b/2025/FG-OrIU_ Towards Better Forgetting via Feature-Gradient Orthogonality for Incremental Unlearning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5201b760d7bc4b512c752c818281d9aaba07bf2b --- /dev/null +++ b/2025/FG-OrIU_ Towards Better Forgetting via Feature-Gradient Orthogonality for Incremental Unlearning/layout.json @@ -0,0 +1,10697 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 102, + 544, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 102, + 544, + 140 + ], + "spans": [ + { + "bbox": [ + 67, + 102, + 544, + 140 + ], + "type": "text", + "content": "FG-OrIU: Towards Better Forgetting via Feature-Gradient Orthogonality for Incremental Unlearning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 160, + 518, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 160, + 518, + 176 + ], + "spans": [ + { + "bbox": [ + 91, + 160, + 518, + 176 + ], + "type": "text", + "content": "Qian Feng" + }, + { + "bbox": [ + 91, + 160, + 518, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 160, + 518, + 176 + ], + "type": "text", + "content": ", JiaHang Tu" + }, + { + "bbox": [ + 91, + 160, + 518, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 160, + 518, + 176 + ], + "type": "text", + "content": ", Mintong Kang" + }, + { + "bbox": [ + 91, + 160, + 518, + 176 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 91, + 160, + 518, + 176 + ], + "type": "text", + "content": ", Hanbin Zhao" + }, + { + "bbox": [ + 91, + 160, + 518, + 176 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 91, + 160, + 518, + 176 + ], + "type": "text", + "content": ", Chao Zhang" + }, + { + "bbox": [ + 91, + 160, + 518, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 160, + 518, + 176 + ], + "type": "text", + "content": ", Hui Qian" + }, + { + "bbox": [ + 91, + 160, + 518, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 113, + 176, + 493, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 176, + 493, + 204 + ], + "spans": [ + { + "bbox": [ + 113, + 176, + 493, + 204 + ], + "type": "text", + "content": "1College of Computer Science and Technology, Zhejiang University \n2Department of Computer Science University of Illinois at Urbana-Champaign" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 256, + 297, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 256, + 297, + 544 + ], + "spans": [ + { + "bbox": [ + 54, + 256, + 297, + 544 + ], + "type": "text", + "content": "Incremental unlearning (IU) is critical for pre-trained models to comply with sequential data deletion requests, yet existing methods primarily suppress parameters or confuse knowledge without explicit constraints on both feature and gradient level, resulting in superficial forgetting where residual information remains recoverable. This incomplete forgetting risks security breaches and disrupts retention balance, especially in IU scenarios. We propose FG-OrIU (Feature-Gradient Orthogonality for Incremental Unlearning), the first framework unifying orthogonal constraints on both features and gradients level to achieve deep forgetting, where the forgetting effect is irreversible. FG-OrIU decomposes feature spaces via Singular Value Decomposition (SVD), separating forgetting and remaining class features into distinct subspaces. It then enforces dual constraints: feature orthogonal projection on both forgetting and remaining classes, while gradient orthogonal projection prevents the reintroduction of forgotten knowledge and disruption to remaining classes during updates. Additionally, dynamic subspace adaptation merges newly forgetting subspaces and contracts remaining subspaces, ensuring a stable balance between removal and retention across sequential unlearning tasks. Extensive experiments demonstrate the effectiveness of our method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 567, + 135, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 135, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 135, + 578 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 586, + 296, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 296, + 695 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 296, + 695 + ], + "type": "text", + "content": "In recent years, pre-training with fine-tuning [22, 51, 57] has dominated model development. However, large-scale datasets often contain harmful or private content, necessitating mechanisms to erase specific data traces. Machine Unlearning (MU) [4, 5, 7, 12, 16, 20, 24, 34, 41, 42, 50] addresses this by enabling pre-trained models (PTMs) to comply with deletion mandates like GDPR (General data protection regulation)'s \"right to be forgotten\" [36, 37]. While existing MU methods focus on static unlearning, for" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 315, + 230, + 554, + 357 + ], + "blocks": [ + { + "bbox": [ + 315, + 230, + 554, + 357 + ], + "lines": [ + { + "bbox": [ + 315, + 230, + 554, + 357 + ], + "spans": [ + { + "bbox": [ + 315, + 230, + 554, + 357 + ], + "type": "image", + "image_path": "e8d792e49fac155c7c124f2cda35afc426466289ac4a664bdba1ab3bf5de5de0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 358, + 555, + 403 + ], + "lines": [ + { + "bbox": [ + 313, + 358, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 555, + 403 + ], + "type": "text", + "content": "Figure 1. Here, we present three examples from forgetting classes, with the reconstructed images generated using Deep Image Prior (DIP) [46], where the last block's features from the unlearned model serve as the training target." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 406, + 555, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 555, + 525 + ], + "type": "text", + "content": "getting a single or small set of classes in one task, real-world scenarios require Incremental Unlearning (IU) [8-10, 17, 19, 25, 28, 31, 35, 48, 60, 61]. Regulatory shifts (e.g., evolving OpenAI safety policies" + }, + { + "bbox": [ + 313, + 406, + 555, + 525 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 313, + 406, + 555, + 525 + ], + "type": "text", + "content": ", EU AI Act" + }, + { + "bbox": [ + 313, + 406, + 555, + 525 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 313, + 406, + 555, + 525 + ], + "type": "text", + "content": ") and dynamic content restrictions necessitate sequential unlearning of multiple classes/concepts without costly retraining. IU need not only erase targets in a single step but also preserve retained performance and ensure irreversible forgetting across tasks, making it critical for sustainable model governance in evolving ethical and legal landscapes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 526, + 556, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 556, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 556, + 682 + ], + "type": "text", + "content": "IU methods broadly address two scenarios: (1) concept erasure in generative models, where sensitive concepts (e.g., nudity, violence) are iteratively removed from diffusion models [11, 45], and (2) instant- or class-level unlearning in vision classifiers, where PTMs forget specific classes (e.g., face identity, medical data) while retaining unrelated ones [26, 34]. In this work, we focus on the latter, specifically targeting class-level IU in vision PTMs. Our goal is threefold: (a) to ensure that the model completely forgets the specified classes, (b) to robustly preserve the performance of remaining classes, and (c) to maintain the effectiveness of previous forgetting across sequential unlearning procedures." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 284, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 284, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 284, + 713 + ], + "type": "text", + "content": "*Corresponding author: Hanbin Zhao ." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 325, + 693, + 416, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 693, + 416, + 703 + ], + "spans": [ + { + "bbox": [ + 325, + 693, + 416, + 703 + ], + "type": "text", + "content": "1 https://openai.com/safety/" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 325, + 703, + 463, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 703, + 463, + 713 + ], + "spans": [ + { + "bbox": [ + 325, + 703, + 463, + 713 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 325, + 703, + 463, + 713 + ], + "type": "text", + "content": "https://artificialintelligenceact.eu/the-act/" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1957" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 491 + ], + "type": "text", + "content": "Existing methods largely fall into two categories: parameter suppression, which selectively weakens target class-related parameters [15, 16, 52], and knowledge confusion, which distorts predictions via noisy labels or bad teacher distillation [12, 18, 29]. However, our empirical findings reveal that these methods lead to superficial forgetting, as the unlearned model still retains semantic information related to the forgetting classes, making the erased knowledge partially residual and recoverable. Specifically, reconstructing features from the last layer of the unlearned model using Deep Image Prior (DIP) [46] produces blurred yet recognizable images (as illustrated in Figure 1). Furthermore, fixing the backbone of the unlearned model and retraining only the classification head [10, 61], which recovers a portion of the original accuracy (as illustrated in Figure 3). The blurred image and the partial recovery of performance both indicate that existing methods merely degrade the features of forgetting classes rather than completely eliminating them. The degraded features still contain residual class-specific information, which leads to potential recovery of the forgetting effect. Besides, these degraded features not only enable recovery but also risk entanglement with the remaining class features [34]. Such entanglement destabilizes decision boundaries, degrading performance on the remaining classes, particularly in IU scenarios where sequential deletion requests amplify error propagation [17, 48]. To address these limitations and achieve deep forgetting, we argue that unlearning should explicitly constrain both feature representations and gradient updates. Since model updates rely on both forward computations (feature extraction) and backward optimization (gradient updates), it is critical to simultaneously suppress the features of forgetting classes and prevent their reintroduction during training updates. In Table 1, we summarize the differences between our methods and prior works." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 495, + 294, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 495, + 294, + 696 + ], + "spans": [ + { + "bbox": [ + 57, + 495, + 294, + 696 + ], + "type": "text", + "content": "To enforce explicit constraints on both feature representations and gradient updates, we draw inspiration from orthogonal mechanisms in related domains. For feature-level constraints, recent work in text-to-image models [45, 55] demonstrates that orthogonalizing sensitive token embeddings effectively removes targeted concepts from diffusion outputs, suggesting geometric isolation as a viable strategy for feature suppression. For gradient-level constraints, inspired by [56], which reduces the mutual influence of different tasks in multi-task learning, and [40, 62], which mitigates the forgetting of old tasks in continual learning, we leverage the insight that updating in directions orthogonal to old/other task subspaces helps alleviate catastrophic forgetting. While these orthogonal principles show promise in adjacent fields like generative model editing and multitask/continual learning, their applicability to MU, particularly IU in vision PTMs, remains underexplored." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 700, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 700, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 700, + 294, + 713 + ], + "type": "text", + "content": "In this paper, we propose FG-OrIU (Feature-Gradient" + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 317, + 106, + 552, + 166 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 553, + 103 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 553, + 103 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 553, + 103 + ], + "type": "text", + "content": "Table 1. Position, Complete, and Recover refers to, where the forgetting occurs, whether the forgetting is thorough, and whether the forgetting effect can be restored, respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 106, + 552, + 166 + ], + "lines": [ + { + "bbox": [ + 317, + 106, + 552, + 166 + ], + "spans": [ + { + "bbox": [ + 317, + 106, + 552, + 166 + ], + "type": "table", + "html": "
PositionCompleteRecover
superficial forgetting
Prior worksOnly Final OutputNoPartial
deep forgetting
OursFeature&GradientYesNo
", + "image_path": "f61b2e852e03ed485ce2a9396f3fc5ab8e7a540f3c48420f0e9de3360d3d7dad.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 170, + 553, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 170, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 170, + 553, + 469 + ], + "type": "text", + "content": "Orthogonality for Incremental Unlearning), a unified framework for IU in vision PTMs that enforces explicit constraints at both feature and gradient levels through three stages: feature subspace decomposition, dual orthogonal projection, and dynamic subspace adaptation. In feature subspace decomposition, we decompose the feature space of each layer into two subspaces via Singular Value Decomposition (SVD): the forgetting subspace " + }, + { + "bbox": [ + 313, + 170, + 553, + 469 + ], + "type": "inline_equation", + "content": "S_f" + }, + { + "bbox": [ + 313, + 170, + 553, + 469 + ], + "type": "text", + "content": ", which captures the discriminative features of the classes to be removed, and the remaining subspace " + }, + { + "bbox": [ + 313, + 170, + 553, + 469 + ], + "type": "inline_equation", + "content": "S_r" + }, + { + "bbox": [ + 313, + 170, + 553, + 469 + ], + "type": "text", + "content": ", which encodes the features of the remaining classes. In dual orthogonal projection, we freeze the backbone of PTM and introduce lightweight LoRA modules for efficient updates. Then, features of samples from forgetting classes are orthogonalized against " + }, + { + "bbox": [ + 313, + 170, + 553, + 469 + ], + "type": "inline_equation", + "content": "S_f" + }, + { + "bbox": [ + 313, + 170, + 553, + 469 + ], + "type": "text", + "content": " to eliminate correlations, while remaining classes features are aligned with " + }, + { + "bbox": [ + 313, + 170, + 553, + 469 + ], + "type": "inline_equation", + "content": "S_r" + }, + { + "bbox": [ + 313, + 170, + 553, + 469 + ], + "type": "text", + "content": " to minimize interference through minimizing the projection residual. Before backward optimization, gradients are first projected onto " + }, + { + "bbox": [ + 313, + 170, + 553, + 469 + ], + "type": "inline_equation", + "content": "S_f" + }, + { + "bbox": [ + 313, + 170, + 553, + 469 + ], + "type": "text", + "content": " to maximally perturb parameters critical for forgetting classes, then modified along the orthogonal complement of " + }, + { + "bbox": [ + 313, + 170, + 553, + 469 + ], + "type": "inline_equation", + "content": "S_r" + }, + { + "bbox": [ + 313, + 170, + 553, + 469 + ], + "type": "text", + "content": " to prevent contamination of retained features. In dynamic subspace adaptation, FOrIU dynamically expands " + }, + { + "bbox": [ + 313, + 170, + 553, + 469 + ], + "type": "inline_equation", + "content": "S_f" + }, + { + "bbox": [ + 313, + 170, + 553, + 469 + ], + "type": "text", + "content": " by merging new forgetting class subspaces and contracts " + }, + { + "bbox": [ + 313, + 170, + 553, + 469 + ], + "type": "inline_equation", + "content": "S_r" + }, + { + "bbox": [ + 313, + 170, + 553, + 469 + ], + "type": "text", + "content": " by removing overlapping features, ensuring cumulative forgetting without degrading retained performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 326, + 470, + 490, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 470, + 490, + 480 + ], + "spans": [ + { + "bbox": [ + 326, + 470, + 490, + 480 + ], + "type": "text", + "content": "Our main contributions are threefold:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 483, + 553, + 591 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 483, + 553, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 483, + 553, + 519 + ], + "spans": [ + { + "bbox": [ + 314, + 483, + 553, + 519 + ], + "type": "text", + "content": "- We reveal that existing IU methods achieve superficial forgetting rather than deep forgetting, as their partial feature suppression leaves residual information recoverable." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 520, + 553, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 520, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 314, + 520, + 553, + 567 + ], + "type": "text", + "content": "- We propose the first framework unifying forgetting on both feature and gradient level via dual orthogonal projection, ensuring irreversible forgetting while preserving remaining class robustness." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 568, + 553, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 568, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 314, + 568, + 553, + 591 + ], + "type": "text", + "content": "- Extensive experiments demonstrate superior performance in MU and IU scenarios across multiple benchmarks." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 597, + 400, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 400, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 400, + 609 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 612, + 433, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 433, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 433, + 624 + ], + "type": "text", + "content": "2.1. Machine Unlearning" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "content": "Machine unlearning (MU) aims to remove the influence of specific data from pre-trained models (PTMs) without retraining from scratch [49, 58]. Existing MU methods are broadly categorized into exact and approximate unlearning. Exact MU [5, 32] ensures statistical indistinguishability from models retrained on remaining data but is computationally prohibitive for large PTMs. Approximate MU re" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1958" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 251 + ], + "type": "text", + "content": "taxes this guarantee and is more practical. Current methods fall into two paradigms: (1) concept erasure in generative models [11, 45]; and (2) class-/sample-wise unlearning in classification models, which modifies pre-trained weights to forget target classes while preserving others [26, 34]. However, most studies focus on static unlearning, assuming a single-step forgetting task. Real-world applications, such as compliance with iterative AI safety policies (e.g., OpenAI's content guidelines) or evolving regulations like the EU AI Act, require incremental unlearning (IU) to process sequential deletion requests efficiently. Directly applying static methods to IU often leads to cumulative errors, model collapse, and incomplete forgetting [17, 48], underscoring the need for dedicated frameworks to ensure compliance in dynamic regulatory environments." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 253, + 192, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 253, + 192, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 192, + 266 + ], + "type": "text", + "content": "2.2. Incremental Unlearning" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 269, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 269, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 269, + 296, + 521 + ], + "type": "text", + "content": "Existing MU methods, like parameter suppression [15, 16, 52] and knowledge confusion methods [10, 12, 17, 18, 29, 35, 61], lead to superficial forgetting due to the absence of direct constraints on the feature representation. In particular, the degraded features trap them in a dilemma between forgetting and remaining, which becomes even more severe when directly applied to IU. Meanwhile, some continual learning methods [6, 27], originally designed to retain old knowledge while dynamically learning new knowledge, also face inherent limitations in IU [35]. Previous work focusing on IU has primarily relied on toy models, e.g., linear models [13], or imposed heavy constraints, such as adding a class-specific synthetic signal in the pre-training stage [43], which is not feasible for PTMs. Moreover, while some focus on dynamically performing both continual learning and selective forgetting within learned knowledge [8, 28, 31], our approach differs in that we emphasize achieving deep forgetting when handling sequential unlearning tasks. That is, rather than alternating between learning and forgetting, the pre-trained model should completely forget unwanted knowledge while efficiently retaining the rest." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 522, + 170, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 170, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 170, + 534 + ], + "type": "text", + "content": "3. Problem Statement" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 539, + 295, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 295, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 295, + 598 + ], + "type": "text", + "content": "Following [60, 61], we first formally define Machine Unlearning (MU), i.e., static unlearning scenario. We then extend this definition to Incremental Unlearning (IU), i.e., dynamic unlearning scenario, where deletion requests arrive sequentially." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 599, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 296, + 696 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 55, + 599, + 296, + 696 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 599, + 296, + 696 + ], + "type": "text", + "content": " be a pre-trained model trained on the dataset " + }, + { + "bbox": [ + 55, + 599, + 296, + 696 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 55, + 599, + 296, + 696 + ], + "type": "text", + "content": ". We denote the mapping relationship of the model as " + }, + { + "bbox": [ + 55, + 599, + 296, + 696 + ], + "type": "inline_equation", + "content": "f_{\\mathcal{M}}(\\mathcal{X}_{\\mathbf{D}})\\rightarrow \\mathcal{Y}_{\\mathbf{D}}" + }, + { + "bbox": [ + 55, + 599, + 296, + 696 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 599, + 296, + 696 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{\\mathbf{D}}" + }, + { + "bbox": [ + 55, + 599, + 296, + 696 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 599, + 296, + 696 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}_{\\mathbf{D}}" + }, + { + "bbox": [ + 55, + 599, + 296, + 696 + ], + "type": "text", + "content": " represent the input set and output set, respectively. For MU, we partition " + }, + { + "bbox": [ + 55, + 599, + 296, + 696 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 55, + 599, + 296, + 696 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 55, + 599, + 296, + 696 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathbf{f}}" + }, + { + "bbox": [ + 55, + 599, + 296, + 696 + ], + "type": "text", + "content": " (forgetting classes) and " + }, + { + "bbox": [ + 55, + 599, + 296, + 696 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathbf{r}}" + }, + { + "bbox": [ + 55, + 599, + 296, + 696 + ], + "type": "text", + "content": " (remaining classes), under the constraint " + }, + { + "bbox": [ + 55, + 599, + 296, + 696 + ], + "type": "inline_equation", + "content": "|\\mathbf{D}_{\\mathbf{r}}| + |\\mathbf{D}_{\\mathbf{f}}|\\ll |\\mathbf{D}|" + }, + { + "bbox": [ + 55, + 599, + 296, + 696 + ], + "type": "text", + "content": ". The forgetting algorithm " + }, + { + "bbox": [ + 55, + 599, + 296, + 696 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 599, + 296, + 696 + ], + "type": "text", + "content": " produces " + }, + { + "bbox": [ + 55, + 599, + 296, + 696 + ], + "type": "inline_equation", + "content": "\\mathcal{M}' = \\mathcal{F}(\\mathcal{M},\\mathbf{D}_{\\mathbf{f}},\\mathbf{D}_{\\mathbf{r}})" + }, + { + "bbox": [ + 55, + 599, + 296, + 696 + ], + "type": "text", + "content": ", altering the following mappings:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 87, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 87, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "f _ {\\mathcal {M} ^ {\\prime}} \\left(\\mathcal {X} _ {\\mathbf {D} _ {\\mathrm {f}}}\\right) \\not \\longrightarrow \\mathcal {Y} _ {\\mathbf {D} _ {\\mathrm {f}}}, f _ {\\mathcal {M} ^ {\\prime}} \\left(\\mathcal {X} _ {\\mathbf {D} _ {\\mathrm {r}}}\\right) \\longrightarrow \\mathcal {Y} _ {\\mathbf {D} _ {\\mathrm {r}}}. \\tag {1}", + "image_path": "253a1f37af928e594bc16591d260d4011ecf1ec5e44dda49c09f2d6cb148b6c2.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 85 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 85 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 72, + 553, + 85 + ], + "type": "inline_equation", + "content": "\\nrightarrow" + }, + { + "bbox": [ + 313, + 72, + 553, + 85 + ], + "type": "text", + "content": " means the mapping relationship no longer holds." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 85, + 553, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 85, + 553, + 169 + ], + "spans": [ + { + "bbox": [ + 313, + 85, + 553, + 169 + ], + "type": "text", + "content": "For IU, let there are a sequence of unlearning task, " + }, + { + "bbox": [ + 313, + 85, + 553, + 169 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{T}_1,\\mathcal{T}_2,\\dots ,\\mathcal{T}_T\\}" + }, + { + "bbox": [ + 313, + 85, + 553, + 169 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 85, + 553, + 169 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 85, + 553, + 169 + ], + "type": "text", + "content": " is the number of unlearning tasks. And " + }, + { + "bbox": [ + 313, + 85, + 553, + 169 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{D}_{f,t},\\mathbf{D}_{r,t}\\}" + }, + { + "bbox": [ + 313, + 85, + 553, + 169 + ], + "type": "text", + "content": " is the forgetting classes and remaining classes belonging to " + }, + { + "bbox": [ + 313, + 85, + 553, + 169 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_t" + }, + { + "bbox": [ + 313, + 85, + 553, + 169 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 85, + 553, + 169 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{r,t} = \\mathbf{D} - \\cup_{i = 1}^{t}\\mathbf{D}_{f,i}" + }, + { + "bbox": [ + 313, + 85, + 553, + 169 + ], + "type": "text", + "content": ". Starting from " + }, + { + "bbox": [ + 313, + 85, + 553, + 169 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_0 = \\mathcal{M}" + }, + { + "bbox": [ + 313, + 85, + 553, + 169 + ], + "type": "text", + "content": ", the forgetting algorithm " + }, + { + "bbox": [ + 313, + 85, + 553, + 169 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 85, + 553, + 169 + ], + "type": "text", + "content": " iteratively processes till " + }, + { + "bbox": [ + 313, + 85, + 553, + 169 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 85, + 553, + 169 + ], + "type": "text", + "content": "-th unlearning task " + }, + { + "bbox": [ + 313, + 85, + 553, + 169 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_t" + }, + { + "bbox": [ + 313, + 85, + 553, + 169 + ], + "type": "text", + "content": " to generate " + }, + { + "bbox": [ + 313, + 85, + 553, + 169 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_t = \\mathcal{F}(\\mathcal{M}_{t - 1},\\mathbf{D}_{f,t},\\mathbf{D}_{r,t})" + }, + { + "bbox": [ + 313, + 85, + 553, + 169 + ], + "type": "text", + "content": " which holds the followings:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 328, + 178, + 553, + 193 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 178, + 553, + 193 + ], + "spans": [ + { + "bbox": [ + 328, + 178, + 553, + 193 + ], + "type": "interline_equation", + "content": "f _ {\\mathcal {M} _ {t}} \\left(\\mathcal {X} _ {\\mathbf {D} _ {f, i}}\\right) \\not \\longrightarrow \\mathcal {Y} _ {\\mathbf {D} _ {f, i}}, f _ {\\mathcal {M} _ {t}} \\left(\\mathcal {X} _ {\\mathbf {D} _ {r, j}}\\right) \\longrightarrow \\mathcal {Y} _ {\\mathbf {D} _ {r, j}}, \\tag {2}", + "image_path": "52b02be30cbe3fcd779e9132505fe0c2b180ee0cd2b49ed070c8e36f1bfa2af6.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 205, + 375, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 375, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 375, + 217 + ], + "type": "text", + "content": "4. Methods" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 220, + 555, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 555, + 303 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 555, + 303 + ], + "type": "text", + "content": "In this section, we introduce FG-OrIU, a comprehensive framework for Incremental Unlearning (IU) in pre-trained models (PTM). By imposing orthogonal constraints on both feature representations and gradient updates, FG-OrIU ensures thorough removal of unwanted knowledge while preserving performance on remaining classes. An overview of the framework is shown in Figure 2." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 306, + 492, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 492, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 492, + 319 + ], + "type": "text", + "content": "4.1. Feature Subspace Decomposition" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 319, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 554, + 521 + ], + "type": "text", + "content": "PTMs acquire implicit knowledge from pre-training data, benefiting downstream tasks during inference. Unlearning aims to selectively remove the influence of forgetting classes from the PTM [10]. Unlike explicit knowledge in traditional Knowledge Bases (KBs) [38, 44], which can be directly edited, the knowledge in PTMs is implicit and intertwined across all parameters [10, 34, 64]. Effective unlearning requires precisely identifying feature spaces for forgetting and remaining. To achieve this, we decompose features into two subspaces: one capturing forgetting-class contributions and the other preserving remaining-class knowledge. Specifically, we perform singular value decomposition (SVD) on the feature matrix of forgetting classes and use the span of its dominant singular vectors to define the forgetting subspace. Similarly, we apply the same procedure to the remaining classes to construct its subspace. We then detail the process as follows." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 521, + 555, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 694 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 694 + ], + "type": "text", + "content": "Given a PTM, " + }, + { + "bbox": [ + 313, + 521, + 555, + 694 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 521, + 555, + 694 + ], + "type": "text", + "content": ", we construct a representation matrix " + }, + { + "bbox": [ + 313, + 521, + 555, + 694 + ], + "type": "inline_equation", + "content": "\\pmb{R}_{f,i} = \\left[\\pmb{x}_{1,i},\\dots,\\pmb{x}_{n_{f_i},i}\\right]\\in \\mathbb{R}^{n_{f_i}\\times d}" + }, + { + "bbox": [ + 313, + 521, + 555, + 694 + ], + "type": "text", + "content": " at each layer " + }, + { + "bbox": [ + 313, + 521, + 555, + 694 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 521, + 555, + 694 + ], + "type": "text", + "content": " where columns correspond to representations of " + }, + { + "bbox": [ + 313, + 521, + 555, + 694 + ], + "type": "inline_equation", + "content": "n_{f_i}" + }, + { + "bbox": [ + 313, + 521, + 555, + 694 + ], + "type": "text", + "content": " samples from the remaining classes " + }, + { + "bbox": [ + 313, + 521, + 555, + 694 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathbf{f},\\mathbf{i}}" + }, + { + "bbox": [ + 313, + 521, + 555, + 694 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 521, + 555, + 694 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 521, + 555, + 694 + ], + "type": "text", + "content": " is the output dimension3. SVD is applied to " + }, + { + "bbox": [ + 313, + 521, + 555, + 694 + ], + "type": "inline_equation", + "content": "\\pmb{R}_{f,i} = \\pmb{U}_{f,i}\\pmb{\\Sigma}_{f,i}(\\pmb{V}_{f,i})^T" + }, + { + "bbox": [ + 313, + 521, + 555, + 694 + ], + "type": "text", + "content": ", followed by its " + }, + { + "bbox": [ + 313, + 521, + 555, + 694 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 521, + 555, + 694 + ], + "type": "text", + "content": "-rank approximation " + }, + { + "bbox": [ + 313, + 521, + 555, + 694 + ], + "type": "inline_equation", + "content": "(\\pmb{R}_{f,i})_k" + }, + { + "bbox": [ + 313, + 521, + 555, + 694 + ], + "type": "text", + "content": " satisfying " + }, + { + "bbox": [ + 313, + 521, + 555, + 694 + ], + "type": "inline_equation", + "content": "||(\\pmb{R}_{f,i})_k||_F^2\\geq \\epsilon ||\\pmb{R}_{f,i}||_F^2" + }, + { + "bbox": [ + 313, + 521, + 555, + 694 + ], + "type": "text", + "content": " for a threshold " + }, + { + "bbox": [ + 313, + 521, + 555, + 694 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 521, + 555, + 694 + ], + "type": "text", + "content": ". The subspace " + }, + { + "bbox": [ + 313, + 521, + 555, + 694 + ], + "type": "inline_equation", + "content": "\\mathcal{S}_{f,i} = \\mathrm{span}\\{\\pmb{B}_{f,i}\\}" + }, + { + "bbox": [ + 313, + 521, + 555, + 694 + ], + "type": "text", + "content": " is defining using the first " + }, + { + "bbox": [ + 313, + 521, + 555, + 694 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 521, + 555, + 694 + ], + "type": "text", + "content": " vectors " + }, + { + "bbox": [ + 313, + 521, + 555, + 694 + ], + "type": "inline_equation", + "content": "\\pmb{B}_{f,i} = \\{\\pmb{u}_{1,i},\\dots,\\pmb{u}_{k,i}\\}" + }, + { + "bbox": [ + 313, + 521, + 555, + 694 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 521, + 555, + 694 + ], + "type": "inline_equation", + "content": "\\pmb{U}_{f,i}" + }, + { + "bbox": [ + 313, + 521, + 555, + 694 + ], + "type": "text", + "content": ", capturing directions with the highest singular values. Similarly, the feature matrix " + }, + { + "bbox": [ + 313, + 521, + 555, + 694 + ], + "type": "inline_equation", + "content": "\\pmb{R}_{r,i}" + }, + { + "bbox": [ + 313, + 521, + 555, + 694 + ], + "type": "text", + "content": " for remaining classes " + }, + { + "bbox": [ + 313, + 521, + 555, + 694 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathbf{r},\\mathbf{i}}" + }, + { + "bbox": [ + 313, + 521, + 555, + 694 + ], + "type": "text", + "content": " defines the subspace " + }, + { + "bbox": [ + 313, + 521, + 555, + 694 + ], + "type": "inline_equation", + "content": "\\mathcal{S}_{r,i} = \\mathrm{span}\\{\\pmb{B}_{r,i}\\}" + }, + { + "bbox": [ + 313, + 521, + 555, + 694 + ], + "type": "text", + "content": ". The two sets of layer-wised feature subspace, " + }, + { + "bbox": [ + 313, + 521, + 555, + 694 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{S}_{f,i}^1,\\dots,\\mathcal{S}_{f,i}^l\\}" + }, + { + "bbox": [ + 313, + 521, + 555, + 694 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 521, + 555, + 694 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{S}_{r,i}^1,\\dots,\\mathcal{S}_{r,i}^l\\}" + }, + { + "bbox": [ + 313, + 521, + 555, + 694 + ], + "type": "text", + "content": ", represent the core part for forgetting and remaining classes," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 324, + 702, + 468, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 702, + 468, + 713 + ], + "spans": [ + { + "bbox": [ + 324, + 702, + 468, + 713 + ], + "type": "text", + "content": "3Here, we omit the notation " + }, + { + "bbox": [ + 324, + 702, + 468, + 713 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 324, + 702, + 468, + 713 + ], + "type": "text", + "content": " for simplicity." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1959" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 550, + 235 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 550, + 235 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 550, + 235 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 550, + 235 + ], + "type": "image", + "image_path": "60c69cb922bd064d6bed1158b52380dd84e70f8643f9fe22e271c4f25a6466dc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 237, + 555, + 271 + ], + "lines": [ + { + "bbox": [ + 55, + 237, + 555, + 271 + ], + "spans": [ + { + "bbox": [ + 55, + 237, + 555, + 271 + ], + "type": "text", + "content": "Figure 2. Overview of FG-OrIU: (a) For each unlearning task, we decompose the layer-wised- forgetting and remaining subspace separately (Sec. 4.1). (b) LoRA is inserted at each layer, with constraints applied at both the feature and gradient levels for forgetting and remaining (Sec. 4.2). (c) For new unlearning tasks, the forgetting subspace expands while the remaining contracts (Sec. 4.3)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 274, + 296, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 274, + 296, + 382 + ], + "spans": [ + { + "bbox": [ + 54, + 274, + 296, + 382 + ], + "type": "text", + "content": "respectively, isolating features to be erased while preserving those to retain. By explicitly decomposing feature subspaces across all layers, we strengthen supervision signals for parameter updates and mitigate the risk of gradient obfuscation [2] in deep networks, where gradient inaccuracies caused by layer-wise nonlinear transformations could otherwise weaken the unlearning effect. This multi-layer constraint ensures consistent suppression of forgetting class features throughout the model hierarchy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 382, + 209, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 382, + 209, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 382, + 209, + 396 + ], + "type": "text", + "content": "4.2. Dual Orthogonal Projection" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 399, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 296, + 491 + ], + "type": "text", + "content": "Prior to unlearning task " + }, + { + "bbox": [ + 55, + 399, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i" + }, + { + "bbox": [ + 55, + 399, + 296, + 491 + ], + "type": "text", + "content": ", we freeze both the backbone and the classification head of the PTM and insert lightweight LoRA modules [23] at each layer. We then denote the model with LoRA inserted as " + }, + { + "bbox": [ + 55, + 399, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\theta}" + }, + { + "bbox": [ + 55, + 399, + 296, + 491 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 399, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 399, + 296, + 491 + ], + "type": "text", + "content": " represents the trainable parameters in LoRA. After unlearning, the updates in LoRA, " + }, + { + "bbox": [ + 55, + 399, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\theta^{*}" + }, + { + "bbox": [ + 55, + 399, + 296, + 491 + ], + "type": "text", + "content": ", will be merged into the initial weight of the backbone, " + }, + { + "bbox": [ + 55, + 399, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_i \\xrightarrow[\\text{merge}]{\\theta^*} \\mathcal{M}_{i+1}" + }, + { + "bbox": [ + 55, + 399, + 296, + 491 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 491, + 296, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 296, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 296, + 538 + ], + "type": "text", + "content": "Building on the feature subspaces " + }, + { + "bbox": [ + 55, + 491, + 296, + 538 + ], + "type": "inline_equation", + "content": "S_f" + }, + { + "bbox": [ + 55, + 491, + 296, + 538 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 491, + 296, + 538 + ], + "type": "inline_equation", + "content": "S_r" + }, + { + "bbox": [ + 55, + 491, + 296, + 538 + ], + "type": "text", + "content": " defined in Section 4.1, we propose a dual orthogonal projection mechanism that operates through both forward feature alignment and backward gradient correction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 543, + 216, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 543, + 216, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 543, + 216, + 555 + ], + "type": "text", + "content": "4.2.1. Feature Orthogonal Projection" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": "Inspired by [45, 55], for forgetting classes, achieving thorough removal at the feature level requires that the features generated by the unlearned model, " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\theta}" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": ", deviate as much as possible from the feature subspace " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "S_{f}" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": ". Meanwhile, we aim to ensure that the unlearned model's feature subspace for remaining classes aligns with " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "S_{r}" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": ". The orthogonality constraint between " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "S_{f}" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": " and the unlearned features serves as a critical mechanism for facilitating class-specific forgetting: under a fixed classification head, features orthogonal to the original subspace " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "S_{f}" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": " induce maximal perturbations in class logits, as the inner product between orthogonal vectors diminishes to zero. This forces the unlearned model to suppress discriminative patterns associated with forgetting" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 274, + 555, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 555, + 346 + ], + "type": "text", + "content": "classes in its intermediate representations. To do so, we propose the following two regularization constraints. For forgetting classes, let " + }, + { + "bbox": [ + 313, + 274, + 555, + 346 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_f" + }, + { + "bbox": [ + 313, + 274, + 555, + 346 + ], + "type": "text", + "content": " denote a batch of forgetting samples from " + }, + { + "bbox": [ + 313, + 274, + 555, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathbf{f},\\mathrm{i}}" + }, + { + "bbox": [ + 313, + 274, + 555, + 346 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 274, + 555, + 346 + ], + "type": "inline_equation", + "content": "h_f^l" + }, + { + "bbox": [ + 313, + 274, + 555, + 346 + ], + "type": "text", + "content": " denote the " + }, + { + "bbox": [ + 313, + 274, + 555, + 346 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 274, + 555, + 346 + ], + "type": "text", + "content": "-th layer features gained with the unlearned model " + }, + { + "bbox": [ + 313, + 274, + 555, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\theta}" + }, + { + "bbox": [ + 313, + 274, + 555, + 346 + ], + "type": "text", + "content": ". For complete forgetting, we enforce:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 363, + 365, + 555, + 383 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 365, + 555, + 383 + ], + "spans": [ + { + "bbox": [ + 363, + 365, + 555, + 383 + ], + "type": "interline_equation", + "content": "\\boldsymbol {h} _ {f} ^ {l} \\perp \\mathcal {S} _ {f, i} ^ {l} \\quad \\Rightarrow \\quad \\min | | \\mathbf {P} _ {f} ^ {l} \\boldsymbol {h} _ {f} ^ {l} | | _ {F} ^ {2} \\tag {3}", + "image_path": "8aa6b7bd4d246e4bc62ac959a4a43157fbcad0d4f3a3c4043ed93f70eb276a49.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 391, + 554, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 391, + 554, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 554, + 467 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 391, + 554, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_f^l = \\mathbf{B}_{f,i}^l (\\mathbf{B}_{f,i}^l)^\\top" + }, + { + "bbox": [ + 313, + 391, + 554, + 467 + ], + "type": "text", + "content": " is the projection matrix for the forgetting subspace " + }, + { + "bbox": [ + 313, + 391, + 554, + 467 + ], + "type": "inline_equation", + "content": "S_{f,i}^{l}" + }, + { + "bbox": [ + 313, + 391, + 554, + 467 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 391, + 554, + 467 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\| _F^2" + }, + { + "bbox": [ + 313, + 391, + 554, + 467 + ], + "type": "text", + "content": " is the F-norm. Eq.(3) ensures that the features of " + }, + { + "bbox": [ + 313, + 391, + 554, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathbf{f},\\mathbf{i}}" + }, + { + "bbox": [ + 313, + 391, + 554, + 467 + ], + "type": "text", + "content": " gained from the unlearned model " + }, + { + "bbox": [ + 313, + 391, + 554, + 467 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\theta}" + }, + { + "bbox": [ + 313, + 391, + 554, + 467 + ], + "type": "text", + "content": " lie outside the original feature subspace gained from " + }, + { + "bbox": [ + 313, + 391, + 554, + 467 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 391, + 554, + 467 + ], + "type": "text", + "content": ", erasing the unwanted information. Thus, we define the flop-f loss:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 379, + 475, + 554, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 379, + 475, + 554, + 506 + ], + "spans": [ + { + "bbox": [ + 379, + 475, + 554, + 506 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {f o p - f} = \\sum_ {l = 1} ^ {m} \\left\\| \\mathbf {P} _ {f} ^ {l} \\boldsymbol {h} _ {f} ^ {l} \\right\\| _ {F} ^ {2}. \\tag {4}", + "image_path": "78da4eb36f94c040039270e90e96117ef6b1f3c8b24a37072313a02975ab6ea9.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 514, + 554, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 554, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 554, + 562 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 514, + 554, + 562 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 514, + 554, + 562 + ], + "type": "text", + "content": " denotes the total number of layers in the PTM. For remaining classes, we follow previous works and apply the standard cross-entropy loss by constraining the model's final output, as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 383, + 571, + 553, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 571, + 553, + 585 + ], + "spans": [ + { + "bbox": [ + 383, + 571, + 553, + 585 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c e} = \\mathcal {L} \\left(\\mathcal {X} _ {\\mathbf {D} _ {r, 1}}, \\mathcal {Y} _ {\\mathbf {D} _ {r, 1}}\\right), \\tag {5}", + "image_path": "ecaabc8594d7a7ccd886b9a622a6576f3e44e0f108c939e16e922c9fa954bc8a.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 593, + 555, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 593, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 593, + 555, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 593, + 555, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 313, + 593, + 555, + 677 + ], + "type": "text", + "content": " denotes the cross-entropy loss function. In addition, we also propose applying explicit constraint at the feature level for the remaining classes to better preserve the desired information. Specifically, " + }, + { + "bbox": [ + 313, + 593, + 555, + 677 + ], + "type": "inline_equation", + "content": "\\pmb{x}_r" + }, + { + "bbox": [ + 313, + 593, + 555, + 677 + ], + "type": "text", + "content": " denote a batch of remaining samples from " + }, + { + "bbox": [ + 313, + 593, + 555, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathbf{r},\\mathbf{i}}" + }, + { + "bbox": [ + 313, + 593, + 555, + 677 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 593, + 555, + 677 + ], + "type": "inline_equation", + "content": "h_r^l" + }, + { + "bbox": [ + 313, + 593, + 555, + 677 + ], + "type": "text", + "content": " denote the " + }, + { + "bbox": [ + 313, + 593, + 555, + 677 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 593, + 555, + 677 + ], + "type": "text", + "content": "-th layer features gained with the unlearned model " + }, + { + "bbox": [ + 313, + 593, + 555, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\theta}" + }, + { + "bbox": [ + 313, + 593, + 555, + 677 + ], + "type": "text", + "content": ". For better preservation, we maintain feature fidelity through:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 391, + 696, + 553, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 696, + 553, + 711 + ], + "spans": [ + { + "bbox": [ + 391, + 696, + 553, + 711 + ], + "type": "interline_equation", + "content": "\\min \\left\\| \\boldsymbol {h} _ {r} ^ {l} - \\mathbf {P} _ {r} ^ {l} \\boldsymbol {h} _ {r} ^ {l} \\right\\| _ {F} ^ {2} \\tag {6}", + "image_path": "6ca0a73dd27971bab7e7e6f54325148773553c5ef2db0de5f3bad67154dbcfa1.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1960" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 70, + 295, + 99 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 70, + 295, + 99 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 295, + 99 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 70, + 295, + 99 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_r^l = \\mathbf{B}_{r,i}^l (\\mathbf{B}_{r,i}^l)^\\top" + }, + { + "bbox": [ + 55, + 70, + 295, + 99 + ], + "type": "text", + "content": " is the projection matrix for the remaining feature subspace " + }, + { + "bbox": [ + 55, + 70, + 295, + 99 + ], + "type": "inline_equation", + "content": "S_{r,i}^{l}" + }, + { + "bbox": [ + 55, + 70, + 295, + 99 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 99, + 295, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 99, + 295, + 147 + ], + "spans": [ + { + "bbox": [ + 55, + 99, + 295, + 147 + ], + "type": "text", + "content": "Eq.(6) ensures that the features of " + }, + { + "bbox": [ + 55, + 99, + 295, + 147 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathbf{r},\\mathbf{i}}" + }, + { + "bbox": [ + 55, + 99, + 295, + 147 + ], + "type": "text", + "content": " gained from the unlearned model " + }, + { + "bbox": [ + 55, + 99, + 295, + 147 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\theta}" + }, + { + "bbox": [ + 55, + 99, + 295, + 147 + ], + "type": "text", + "content": ", align with the original feature subspace gained from " + }, + { + "bbox": [ + 55, + 99, + 295, + 147 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 99, + 295, + 147 + ], + "type": "text", + "content": ", maintaining the better preservation. Therefore, We draw the following " + }, + { + "bbox": [ + 55, + 99, + 295, + 147 + ], + "type": "inline_equation", + "content": "f_{op - r}" + }, + { + "bbox": [ + 55, + 99, + 295, + 147 + ], + "type": "text", + "content": " loss:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 110, + 152, + 295, + 183 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 152, + 295, + 183 + ], + "spans": [ + { + "bbox": [ + 110, + 152, + 295, + 183 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {f o p - r} = \\sum_ {l = 1} ^ {m} \\left\\| \\boldsymbol {h} _ {r} ^ {l} - \\mathbf {P} _ {r} ^ {l} \\boldsymbol {h} _ {r} ^ {l} \\right\\| _ {F} ^ {2}. \\tag {7}", + "image_path": "bf931798708c0be39aabf0019dc3b56b7db443277cda40565570af00b1701ed2.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 188, + 167, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 188, + 167, + 200 + ], + "spans": [ + { + "bbox": [ + 67, + 188, + 167, + 200 + ], + "type": "text", + "content": "Finally, the total loss is:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 93, + 206, + 295, + 220 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 206, + 295, + 220 + ], + "spans": [ + { + "bbox": [ + 93, + 206, + 295, + 220 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {c e} + \\lambda_ {1} \\mathcal {L} _ {f o p - f} + \\lambda_ {2} \\mathcal {L} _ {f o p - r}, \\tag {8}", + "image_path": "40f1c00883697dd78d0ba4e2e7932f526568099e803864431fab1fb09033a895.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 225, + 253, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 225, + 253, + 238 + ], + "spans": [ + { + "bbox": [ + 55, + 225, + 253, + 238 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 225, + 253, + 238 + ], + "type": "inline_equation", + "content": "\\lambda_{1},\\lambda_{2}" + }, + { + "bbox": [ + 55, + 225, + 253, + 238 + ], + "type": "text", + "content": " are the coefficients during training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 242, + 222, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 242, + 222, + 254 + ], + "spans": [ + { + "bbox": [ + 55, + 242, + 222, + 254 + ], + "type": "text", + "content": "4.2.2. Gradient Orthogonal Projection" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 257, + 296, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 257, + 296, + 328 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 296, + 328 + ], + "type": "text", + "content": "To prevent parameter updates from reintroducing forgotten knowledge while preserving retained features, we design a gradient correction mechanism inspired by gradient projection techniques in multi-task learning [56] and continual learning [40, 62]. Given the raw gradient " + }, + { + "bbox": [ + 55, + 257, + 296, + 328 + ], + "type": "inline_equation", + "content": "g_{l}" + }, + { + "bbox": [ + 55, + 257, + 296, + 328 + ], + "type": "text", + "content": " at layer " + }, + { + "bbox": [ + 55, + 257, + 296, + 328 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 257, + 296, + 328 + ], + "type": "text", + "content": ", we apply the following modifications:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 77, + 334, + 295, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 334, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 77, + 334, + 295, + 384 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\hat {g} _ {l} = g _ {l} - g _ {l} ^ {\\prime} - g _ {l} ^ {\\prime \\prime} \\tag {9} \\\\ = \\boldsymbol {g} _ {l} - \\underbrace {\\boldsymbol {g} _ {l} (\\boldsymbol {I} - \\mathbf {B} _ {f , i} ^ {l} (\\mathbf {B} _ {f , i} ^ {l}) ^ {\\top})} _ {(t e r m 1: g o p - f)} - \\underbrace {\\boldsymbol {g} _ {l} \\mathbf {B} _ {r , i} ^ {l} (\\mathbf {B} _ {r , i} ^ {l}) ^ {\\top}} _ {(t e r m 2: g o p - r)}, \\\\ \\end{array}", + "image_path": "7f633089fcc3ef2b57bc37a9393f4b662dfad2a3454e2f661f65d59d6e7f8d80.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 389, + 296, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 296, + 449 + ], + "type": "text", + "content": "Term (gop-f), " + }, + { + "bbox": [ + 55, + 389, + 296, + 449 + ], + "type": "inline_equation", + "content": "g_{l}^{\\prime}" + }, + { + "bbox": [ + 55, + 389, + 296, + 449 + ], + "type": "text", + "content": ": Removing the components of " + }, + { + "bbox": [ + 55, + 389, + 296, + 449 + ], + "type": "inline_equation", + "content": "g_{l}" + }, + { + "bbox": [ + 55, + 389, + 296, + 449 + ], + "type": "text", + "content": " that are orthogonal to " + }, + { + "bbox": [ + 55, + 389, + 296, + 449 + ], + "type": "inline_equation", + "content": "S_{f,i}^{l}" + }, + { + "bbox": [ + 55, + 389, + 296, + 449 + ], + "type": "text", + "content": " not only amplifies updates focusing on directions most critical for encoding forgetting class features, but also blocks parameter updates that could reintroduce removed knowledge." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 449, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 296, + 521 + ], + "type": "text", + "content": "Term (gop-r), " + }, + { + "bbox": [ + 55, + 449, + 296, + 521 + ], + "type": "inline_equation", + "content": "g_{l}^{\\prime \\prime}" + }, + { + "bbox": [ + 55, + 449, + 296, + 521 + ], + "type": "text", + "content": ": We aim to ensure that forgetting specific classes does not affect the remaining classes. Formally, for the unlearned model " + }, + { + "bbox": [ + 55, + 449, + 296, + 521 + ], + "type": "inline_equation", + "content": "f_{\\mathcal{M}_{\\theta}}" + }, + { + "bbox": [ + 55, + 449, + 296, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 449, + 296, + 521 + ], + "type": "inline_equation", + "content": "x_{r}" + }, + { + "bbox": [ + 55, + 449, + 296, + 521 + ], + "type": "text", + "content": " from the remaining classes " + }, + { + "bbox": [ + 55, + 449, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathbf{r},\\mathbf{i}}" + }, + { + "bbox": [ + 55, + 449, + 296, + 521 + ], + "type": "text", + "content": ", the following equation holds: " + }, + { + "bbox": [ + 55, + 449, + 296, + 521 + ], + "type": "inline_equation", + "content": "f_{\\mathcal{M}_{\\theta}}(x_{r}) = f_{\\mathcal{M}}(x_{r})" + }, + { + "bbox": [ + 55, + 449, + 296, + 521 + ], + "type": "text", + "content": ". Specifically, we consider the input and output of layer " + }, + { + "bbox": [ + 55, + 449, + 296, + 521 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 449, + 296, + 521 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 122, + 528, + 294, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 528, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 122, + 528, + 294, + 555 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {W} _ {l} \\boldsymbol {x} _ {r} + \\Delta \\theta_ {l} \\boldsymbol {x} _ {r} = \\boldsymbol {W} _ {l} \\boldsymbol {x} _ {r} \\\\ \\Delta \\theta_ {l} \\boldsymbol {x} _ {\\boldsymbol {r}} = \\mathbf {0} \\tag {10} \\\\ \\end{array}", + "image_path": "270844bb70280528ccea52beec8bf38bac85419e529b6b14ec7ce0508142abd9.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 561, + 295, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 561, + 295, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 561, + 295, + 609 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 561, + 295, + 609 + ], + "type": "inline_equation", + "content": "\\Delta \\theta_{l}" + }, + { + "bbox": [ + 55, + 561, + 295, + 609 + ], + "type": "text", + "content": " represents the change in the LoRA parameters inserted at layer " + }, + { + "bbox": [ + 55, + 561, + 295, + 609 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 561, + 295, + 609 + ], + "type": "text", + "content": ", which corresponds to its gradient. " + }, + { + "bbox": [ + 55, + 561, + 295, + 609 + ], + "type": "inline_equation", + "content": "W_{l}" + }, + { + "bbox": [ + 55, + 561, + 295, + 609 + ], + "type": "text", + "content": " is the parameter at layer " + }, + { + "bbox": [ + 55, + 561, + 295, + 609 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 561, + 295, + 609 + ], + "type": "text", + "content": " from the pre-trained model, and these parameters remain frozen during unlearning." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 609, + 295, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 295, + 693 + ], + "type": "text", + "content": "Eq.(10) indicates that if the update direction of the parameters, i.e., the gradient of the trainable LoRA parameters, lies in the complementary space of " + }, + { + "bbox": [ + 55, + 609, + 295, + 693 + ], + "type": "inline_equation", + "content": "S_{r,i}^{l}" + }, + { + "bbox": [ + 55, + 609, + 295, + 693 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 55, + 609, + 295, + 693 + ], + "type": "inline_equation", + "content": "\\pmb{g}_{l} \\perp S_{r,i}^{l}" + }, + { + "bbox": [ + 55, + 609, + 295, + 693 + ], + "type": "text", + "content": ", then " + }, + { + "bbox": [ + 55, + 609, + 295, + 693 + ], + "type": "inline_equation", + "content": "\\Delta \\theta \\pmb{x}_r = \\eta \\pmb{g}_l \\pmb{x}_r = \\mathbf{0}" + }, + { + "bbox": [ + 55, + 609, + 295, + 693 + ], + "type": "text", + "content": " (where " + }, + { + "bbox": [ + 55, + 609, + 295, + 693 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 609, + 295, + 693 + ], + "type": "text", + "content": " is the learning rate), ensuring that the unlearned model's performance on the remaining classes is not affected. Therefore, we further refine the gradient by removing its components along " + }, + { + "bbox": [ + 55, + 609, + 295, + 693 + ], + "type": "inline_equation", + "content": "S_{r,i}^{l}" + }, + { + "bbox": [ + 55, + 609, + 295, + 693 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 693, + 295, + 718 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 718 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 718 + ], + "type": "text", + "content": "Finally, modified gradient " + }, + { + "bbox": [ + 55, + 693, + 295, + 718 + ], + "type": "inline_equation", + "content": "\\hat{g}_l" + }, + { + "bbox": [ + 55, + 693, + 295, + 718 + ], + "type": "text", + "content": " serves the real gradient for parameter updates at the layer " + }, + { + "bbox": [ + 55, + 693, + 295, + 718 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 693, + 295, + 718 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 72, + 482, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 482, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 482, + 84 + ], + "type": "text", + "content": "4.3. Dynamic Subspace Adaptation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 84, + 553, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 84, + 553, + 155 + ], + "spans": [ + { + "bbox": [ + 313, + 84, + 553, + 155 + ], + "type": "text", + "content": "When encountering a new unlearning task " + }, + { + "bbox": [ + 313, + 84, + 553, + 155 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{i + 1}" + }, + { + "bbox": [ + 313, + 84, + 553, + 155 + ], + "type": "text", + "content": ", our method only needs to update the two decomposed feature subspaces. Since the forgetting classes " + }, + { + "bbox": [ + 313, + 84, + 553, + 155 + ], + "type": "inline_equation", + "content": "\\mathbf{D_{f,i + 1}}" + }, + { + "bbox": [ + 313, + 84, + 553, + 155 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 84, + 553, + 155 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{i + 1}" + }, + { + "bbox": [ + 313, + 84, + 553, + 155 + ], + "type": "text", + "content": " are a subset of the remaining classes from the previous task " + }, + { + "bbox": [ + 313, + 84, + 553, + 155 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i" + }, + { + "bbox": [ + 313, + 84, + 553, + 155 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 313, + 84, + 553, + 155 + ], + "type": "inline_equation", + "content": "\\mathbf{D_{f,i + 1}}\\subseteq \\mathbf{D_{r,i}}" + }, + { + "bbox": [ + 313, + 84, + 553, + 155 + ], + "type": "text", + "content": ", the feature subspaces can be efficiently adjusted. For simplicity, we omit the notion " + }, + { + "bbox": [ + 313, + 84, + 553, + 155 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 84, + 553, + 155 + ], + "type": "text", + "content": " of layers." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 156, + 553, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 553, + 215 + ], + "type": "text", + "content": "We first construct a representation matrix " + }, + { + "bbox": [ + 313, + 156, + 553, + 215 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{f,i+1}" + }, + { + "bbox": [ + 313, + 156, + 553, + 215 + ], + "type": "text", + "content": " using only the forgetting classes " + }, + { + "bbox": [ + 313, + 156, + 553, + 215 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathbf{f},i+1}" + }, + { + "bbox": [ + 313, + 156, + 553, + 215 + ], + "type": "text", + "content": ". Before applying SVD and " + }, + { + "bbox": [ + 313, + 156, + 553, + 215 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 156, + 553, + 215 + ], + "type": "text", + "content": "-rank approximation, we remove common bases already present in " + }, + { + "bbox": [ + 313, + 156, + 553, + 215 + ], + "type": "inline_equation", + "content": "S_{f,i}" + }, + { + "bbox": [ + 313, + 156, + 553, + 215 + ], + "type": "text", + "content": " to ensure newly added bases are unique and orthogonal to existing ones:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 337, + 219, + 553, + 254 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 219, + 553, + 254 + ], + "spans": [ + { + "bbox": [ + 337, + 219, + 553, + 254 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\hat {\\boldsymbol {R}} _ {f, i + 1} = \\boldsymbol {R} _ {f, i + 1} - \\boldsymbol {B} _ {f, i} \\left(\\boldsymbol {B} _ {f, i}\\right) ^ {T} \\left(\\boldsymbol {R} _ {f, i + 1}\\right) (11) \\\\ = \\boldsymbol {R} _ {f, i + 1} - \\boldsymbol {R} _ {f, i + 1} ^ {\\text {p r o j}}. (12) \\\\ \\end{array}", + "image_path": "62fdac1cc286b69659f1d58077e44e37ea6654f804b22d4aa461027c548a8ecb.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 260, + 554, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 260, + 554, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 554, + 361 + ], + "type": "text", + "content": "Next, we perform SVD on " + }, + { + "bbox": [ + 313, + 260, + 554, + 361 + ], + "type": "inline_equation", + "content": "\\hat{R}_{f,i+1} = \\hat{U}_{f,i+1}\\hat{\\Sigma}_{f,i+1}(\\hat{V}_{f,i+1})^T" + }, + { + "bbox": [ + 313, + 260, + 554, + 361 + ], + "type": "text", + "content": " and select " + }, + { + "bbox": [ + 313, + 260, + 554, + 361 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 260, + 554, + 361 + ], + "type": "text", + "content": " new orthogonal from " + }, + { + "bbox": [ + 313, + 260, + 554, + 361 + ], + "type": "inline_equation", + "content": "\\hat{U}_{f,i+1}" + }, + { + "bbox": [ + 313, + 260, + 554, + 361 + ], + "type": "text", + "content": ". Then, the feature subspace of forgetting classes is expanded as " + }, + { + "bbox": [ + 313, + 260, + 554, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{S}_{f,i+1} = \\text{span}\\{B_{f,i+1}\\}" + }, + { + "bbox": [ + 313, + 260, + 554, + 361 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 260, + 554, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{f,i+1} = [B_{f,i};\\mathbf{u}_1,\\dots,\\mathbf{u}_h]" + }, + { + "bbox": [ + 313, + 260, + 554, + 361 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 260, + 554, + 361 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 260, + 554, + 361 + ], + "type": "text", + "content": " new bases from " + }, + { + "bbox": [ + 313, + 260, + 554, + 361 + ], + "type": "inline_equation", + "content": "\\hat{U}_{f,i+1}" + }, + { + "bbox": [ + 313, + 260, + 554, + 361 + ], + "type": "text", + "content": ". And the feature subspace " + }, + { + "bbox": [ + 313, + 260, + 554, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{S}_{r,i+1}" + }, + { + "bbox": [ + 313, + 260, + 554, + 361 + ], + "type": "text", + "content": " of remaining classes is contracted to remove the overlapping features accordingly." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 365, + 391, + 377 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 365, + 391, + 377 + ], + "spans": [ + { + "bbox": [ + 314, + 365, + 391, + 377 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 314, + 379, + 430, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 379, + 430, + 392 + ], + "spans": [ + { + "bbox": [ + 314, + 379, + 430, + 392 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 396, + 518, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 518, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 518, + 407 + ], + "type": "text", + "content": "Datasets, Benchmarks and Pre-trained Models." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 407, + 554, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 554, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 554, + 563 + ], + "type": "text", + "content": "Following [60], we evaluate FG-OrIU on Classification and Recognition tasks. For classification, we use ImageNet100 [39], CUB-200 [47], and Omnibenchmark [59], with a ViT-B/16 [14] pre-trained in PyTorch [33]. For face recognition, we use CASIA-Face100 [60], a subset of CASIA-WebFace [54], and MS-Celeb-100, sampled from [21]. The pre-trained model is the Face Transformer from [63]. We focus on class-unlearning scenarios with two benchmarks: (a) MU, varying numbers of classes are removed; (b) IU, 20 classes are removed per task over 4 tasks. All setting is in the form of Y-X, which means experiments start from a PTM (Y classes originally) and forget X classes. More details about implementations are in Appendix ??" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 563, + 553, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 553, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 553, + 640 + ], + "type": "text", + "content": "Metrics. Following [60], we evaluate performance using average accuracy (Acc) for both forgetting and remaining classes. The accuracy for forgetting classes should approach zero, while the accuracy for remaining classes should remain close to the original model's performance. Additionally, we adopt the H-Mean [43] metric to assess overall performance after learning task " + }, + { + "bbox": [ + 313, + 563, + 553, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_t" + }, + { + "bbox": [ + 313, + 563, + 553, + 640 + ], + "type": "text", + "content": ", computed as:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 368, + 653, + 553, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 653, + 553, + 681 + ], + "spans": [ + { + "bbox": [ + 368, + 653, + 553, + 681 + ], + "type": "interline_equation", + "content": "H - M e a n ^ {(t)} = \\frac {2 A c c _ {r} ^ {(t)} \\cdot D r o p ^ {(t)}}{A c c _ {r} ^ {(t)} + D r o p ^ {(t)}}. \\tag {13}", + "image_path": "348e3a86c538d62e1baf608a67460b4a4ab6baba7abe6ffbbb3a672a3e51bf7a.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 686, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 686, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 686, + 553, + 715 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 313, + 686, + 553, + 715 + ], + "type": "inline_equation", + "content": "Acc_{r}^{(t)}" + }, + { + "bbox": [ + 313, + 686, + 553, + 715 + ], + "type": "text", + "content": " represents accuracy on the retained dataset after task " + }, + { + "bbox": [ + 313, + 686, + 553, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_t" + }, + { + "bbox": [ + 313, + 686, + 553, + 715 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 313, + 686, + 553, + 715 + ], + "type": "inline_equation", + "content": "Drop^{(t)} = Acc_{f}^{(t - 1)} - Acc_{f}^{(t)}" + }, + { + "bbox": [ + 313, + 686, + 553, + 715 + ], + "type": "text", + "content": " mea" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1961" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 64, + 555, + 266 + ], + "blocks": [ + { + "bbox": [ + 58, + 64, + 555, + 266 + ], + "lines": [ + { + "bbox": [ + 58, + 64, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 58, + 64, + 555, + 266 + ], + "type": "table", + "html": "
MethodsTunable Ratio ↓100-5100-10100-50100-90
H ↑Accr ↑Accf ↓H ↑Accr ↑Accf ↓H ↑Accr ↑Accf ↓H ↑Accr ↑Accf ↓
Pre-train--73.8570.88-73.8172.74-72.4574.88-72.3273.86
L2*99.73%67.1366.963.5867.7464.871.8664.3956.610.2452.4540.680.04
EWC*99.73%69.6568.690.2469.1665.920.0062.4853.610.0044.4131.750.00
MAS*99.73%69.7369.290.7269.3866.410.1262.6753.880.0046.7934.240.00
LwF99.73%67.9568.550.0070.0867.620.0063.8155.590.0061.2552.320.00
DER99.73%69.7068.550.0070.0867.620.0063.8155.590.0057.0346.440.00
DER++99.73%69.7068.560.0070.5868.550.0064.6156.820.0061.4052.540.00
FDR99.73%70.3169.740.0070.0367.510.0065.4058.040.0053.8542.370.00
Retrain100.00%16.499.330.0018.0210.280.0019.7111.350.0046.4733.900.00
SCRUB99.73%67.7864.940.0068.2664.310.0065.8258.710.0016.119.040.00
SCRUB-S99.73%70.2969.950.2471.1969.810.1257.4951.209.3417.539.940.00
LIRF*50.66%25.5667.6755.1326.3565.8356.2647.1358.9535.6254.4944.293.06
GS-LoRA1.28%71.0271.160.0071.7670.810.0071.2968.050.0273.7173.560.00
GS-LoRA++1.28%71.1271.360.0072.0471.350.0071.5668.520.0072.8571.860.00
FG-OrIU1.28%71.7872.710.0072.5272.310.0074.1873.510.0075.4977.150.00
", + "image_path": "bf039ee8be30e903e540a9871cbbb7dd65d97a2eb85ba8f91aaf3d962c5015c8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 272, + 555, + 307 + ], + "lines": [ + { + "bbox": [ + 55, + 272, + 555, + 307 + ], + "spans": [ + { + "bbox": [ + 55, + 272, + 555, + 307 + ], + "type": "text", + "content": "Table 2. Static Machine Unlearning results for face recognition task on CASIA-Face100. Accr (%) and Accf (%) are the accuracies of remaining and forgetting classes. * denotes the original methods with a rehearsal buffer. Retrain represents retraining the model using replay data and the training epoch is the same as other methods to ensure a fair comparison. Pre-train denotes the results before forgetting." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 59, + 314, + 291, + 460 + ], + "blocks": [ + { + "bbox": [ + 59, + 314, + 291, + 460 + ], + "lines": [ + { + "bbox": [ + 59, + 314, + 291, + 460 + ], + "spans": [ + { + "bbox": [ + 59, + 314, + 291, + 460 + ], + "type": "image", + "image_path": "4aa4a42371b07dd43463877fa058fb00049a3065512c6495781bb452334269bc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 460, + 295, + 495 + ], + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 495 + ], + "type": "text", + "content": "Figure 3. Accuracy on forgetting classes when recovering on CASIA-Face100. The line (Pre-train) is the result before unlearning. The line (FG-OrIU) is Our method." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 497, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 295, + 544 + ], + "type": "text", + "content": "sures the performance drop on forgetting classes before and after training. After unlearning task " + }, + { + "bbox": [ + 55, + 497, + 295, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_t" + }, + { + "bbox": [ + 55, + 497, + 295, + 544 + ], + "type": "text", + "content": ", we evaluate performance on all previously forgetting classes from tasks " + }, + { + "bbox": [ + 55, + 497, + 295, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_1, \\mathcal{T}_2, \\dots, \\mathcal{T}_{t-1}" + }, + { + "bbox": [ + 55, + 497, + 295, + 544 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 544, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 544, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 54, + 544, + 295, + 664 + ], + "type": "text", + "content": "Baselines. Following [60], we compare FG-OrIU against (a) continual learning (CL) methods, including L2 regularization, EWC [27], MAS [1], LwF [30], DER [6], and FDR [3], as well as (b) machine unlearning (MU) methods such as BAD-T [12], LIRF [53], SCRUB/SCRUB-S [29], GS-LoRA/GS-LoRA++ [60, 61], and retraining. To ensure forgetting occurs in the backbone, we freeze the classification head for all methods. In the retraining setup, we train a randomly initialized model using replay data, ensuring the number of training epochs matches that of other methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 666, + 194, + 677 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 194, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 194, + 677 + ], + "type": "text", + "content": "5.2. Results and Comparison" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "Table 2 presents the results for static Machine Unlearning task. FG-OrIU exhibits significantly degraded performance on forgetting classes while maintaining the perfor" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 309, + 555, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 309, + 555, + 585 + ], + "spans": [ + { + "bbox": [ + 313, + 309, + 555, + 585 + ], + "type": "text", + "content": "mance of remaining classes. For example, in the settings where 50 and 90 classes are removed per task, FG-OrIU achieves improvements of " + }, + { + "bbox": [ + 313, + 309, + 555, + 585 + ], + "type": "inline_equation", + "content": "2.62\\%" + }, + { + "bbox": [ + 313, + 309, + 555, + 585 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 309, + 555, + 585 + ], + "type": "inline_equation", + "content": "2.64\\%" + }, + { + "bbox": [ + 313, + 309, + 555, + 585 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 309, + 555, + 585 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 309, + 555, + 585 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 309, + 555, + 585 + ], + "type": "inline_equation", + "content": "4.99\\%" + }, + { + "bbox": [ + 313, + 309, + 555, + 585 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 309, + 555, + 585 + ], + "type": "inline_equation", + "content": "5.29\\%" + }, + { + "bbox": [ + 313, + 309, + 555, + 585 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 309, + 555, + 585 + ], + "type": "inline_equation", + "content": "ACC_{r}" + }, + { + "bbox": [ + 313, + 309, + 555, + 585 + ], + "type": "text", + "content": " compared to GS-LoRA. Notably, FG-OrIU keeps the pre-trained model frozen and only trains the LoRA modules, enabling effective and efficient forgetting. Table 3 and 4 show the results for Incremental Unlearning task. FG-OrIU consistently outperforms all methods. Specifically, CL methods struggle to achieve effective forgetting on ImageNet100, as seen in DER, which retains high " + }, + { + "bbox": [ + 313, + 309, + 555, + 585 + ], + "type": "inline_equation", + "content": "ACC_{f}" + }, + { + "bbox": [ + 313, + 309, + 555, + 585 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 309, + 555, + 585 + ], + "type": "inline_equation", + "content": "10.40\\%" + }, + { + "bbox": [ + 313, + 309, + 555, + 585 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 309, + 555, + 585 + ], + "type": "inline_equation", + "content": "10.90\\%" + }, + { + "bbox": [ + 313, + 309, + 555, + 585 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 309, + 555, + 585 + ], + "type": "inline_equation", + "content": "8.80\\%" + }, + { + "bbox": [ + 313, + 309, + 555, + 585 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 309, + 555, + 585 + ], + "type": "inline_equation", + "content": "3.40\\%" + }, + { + "bbox": [ + 313, + 309, + 555, + 585 + ], + "type": "text", + "content": " across four tasks, respectively. On the other hand, MU methods fail to maintain their forgetting effect over sequential unlearning tasks. For example, LIRF and SCRUB-S show an increase in " + }, + { + "bbox": [ + 313, + 309, + 555, + 585 + ], + "type": "inline_equation", + "content": "ACC_{o}" + }, + { + "bbox": [ + 313, + 309, + 555, + 585 + ], + "type": "text", + "content": " on CASIA-Face100, which becomes more pronounced as the sequence of unlearning tasks grows. GS-LoRA is specifically designed for IU. Compared to it, our method achieves better forgetting while also demonstrating superior performance in preserving the remaining classes, i.e., substantial improvements in " + }, + { + "bbox": [ + 313, + 309, + 555, + 585 + ], + "type": "inline_equation", + "content": "ACC_{r}" + }, + { + "bbox": [ + 313, + 309, + 555, + 585 + ], + "type": "text", + "content": ", particularly in long-sequence unlearning scenarios, achieving gains of " + }, + { + "bbox": [ + 313, + 309, + 555, + 585 + ], + "type": "inline_equation", + "content": "5.15\\%" + }, + { + "bbox": [ + 313, + 309, + 555, + 585 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 309, + 555, + 585 + ], + "type": "inline_equation", + "content": "4.6\\%" + }, + { + "bbox": [ + 313, + 309, + 555, + 585 + ], + "type": "text", + "content": " in the fourth unlearning task on two different datasets, respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 592, + 408, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 592, + 408, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 592, + 408, + 605 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 605, + 555, + 715 + ], + "type": "text", + "content": "Our FG-OrIU consists of four components, i.e., feature orthogonal projection for forgetting, feature orthogonal projection for remaining, gradient orthogonal projection for forgetting, and gradient orthogonal projection for remaining, denoted as " + }, + { + "bbox": [ + 312, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "fop - f" + }, + { + "bbox": [ + 312, + 605, + 555, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 312, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "fop - r" + }, + { + "bbox": [ + 312, + 605, + 555, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 312, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "gop - f" + }, + { + "bbox": [ + 312, + 605, + 555, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 312, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "gop - r" + }, + { + "bbox": [ + 312, + 605, + 555, + 715 + ], + "type": "text", + "content": ". We independently remove each of these components to demonstrate their individual contributions. We observe that when the regularization term or gradient modification related to remaining is removed, the performance on the remaining" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1962" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 555, + 256 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 555, + 256 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 555, + 256 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 555, + 256 + ], + "type": "table", + "html": "
MethodsTask 1(100-20)Task 2(80-20)Task 3(60-20)Task 4(40-20)
H↑Accr↑Accf↓H↑Accr↑Accf↓Acco↓H↑Accr↑Accf↓Acco↓H↑Accr↑Accf↓Acco↓
Pre-train-74.3174.65-74.5073.80--74.8073.91--74.4775.11-
L2*66.9162.132.1666.6661.741.379.3666.4261.371.5411.8366.9561.020.978.00
EWC*67.7161.950.0067.7162.550.000.0067.0961.430.000.1068.0262.140.000.23
MAS*67.5261.630.0068.1263.250.000.0068.1563.230.000.0367.7061.610.000.00
LwF69.4365.040.2169.9466.470.000.2170.3467.110.000.0070.8967.120.000.04
DER70.7567.240.0068.8864.580.000.0068.9564.620.000.0069.4164.510.000.00
DER++70.0165.900.0068.9964.770.000.0069.9666.420.000.0069.8565.280.000.00
FDR68.2962.920.0067.3962.010.000.0069.1664.990.000.0072.5170.080.000.00
Retrain17.299.770.0031.1219.720.000.0041.1728.801.720.0052.4441.774.660.09
BAD-T68.9063.970.0069.5665.950.210.0070.2166.870.000.3472.7170.460.000.04
LIRF*30.5964.4654.6034.0562.0350.3443.5044.5662.5339.2936.5840.3662.6245.3427.96
SCRUB70.3966.590.0070.5567.850.320.0071.0168.330.000.4473.3671.680.000.05
SCRUB-S72.4170.340.0570.0671.535.1615.2273.3873.090.2415.8175.3376.240.686.47
GS-LoRA74.4074.160.0073.5973.370.000.0574.3674.880.060.0073.7672.450.001.93
GS-LoRA++73.9774.160.0073.4873.160.000.1074.2074.510.000.0075.2375.360.000.00
FG-OrIU74.4574.250.0074.6475.500.000.0075.4977.140.000.0077.7180.510.000.00
", + "image_path": "6005b36a826d2e2f50f4420788a4e0c576199b40c4a650a4e929c901f3ca41cd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 292, + 555, + 476 + ], + "blocks": [ + { + "bbox": [ + 55, + 258, + 555, + 291 + ], + "lines": [ + { + "bbox": [ + 55, + 258, + 555, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 555, + 291 + ], + "type": "text", + "content": "Table 3. Incremental Unlearning results for face recognition task on CASIA-Face100. " + }, + { + "bbox": [ + 55, + 258, + 555, + 291 + ], + "type": "inline_equation", + "content": "Acc_{o}(\\%)" + }, + { + "bbox": [ + 55, + 258, + 555, + 291 + ], + "type": "text", + "content": " is the accuracy of old tasks, i.e., the accuracy on all previously forgetting classes in task " + }, + { + "bbox": [ + 55, + 258, + 555, + 291 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{T}_1,\\dots ,\\mathcal{T}_{t - 1}\\}" + }, + { + "bbox": [ + 55, + 258, + 555, + 291 + ], + "type": "text", + "content": ". With four tasks, each forgetting 20 classes, desired forgetting should consider both " + }, + { + "bbox": [ + 55, + 258, + 555, + 291 + ], + "type": "inline_equation", + "content": "Acc_{r}" + }, + { + "bbox": [ + 55, + 258, + 555, + 291 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 258, + 555, + 291 + ], + "type": "inline_equation", + "content": "Acc_{f}" + }, + { + "bbox": [ + 55, + 258, + 555, + 291 + ], + "type": "text", + "content": ". We bold the best and underline the second-best " + }, + { + "bbox": [ + 55, + 258, + 555, + 291 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 258, + 555, + 291 + ], + "type": "text", + "content": " results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 292, + 555, + 476 + ], + "lines": [ + { + "bbox": [ + 57, + 292, + 555, + 476 + ], + "spans": [ + { + "bbox": [ + 57, + 292, + 555, + 476 + ], + "type": "table", + "html": "
MethodsTask 1(100-20)Task 2(80-20)Task 3(60-20)Task 4(40-20)
H↑Accr↑Accf↓H↑Accr↑Accf↓Acco↓H↑Accr↑Accf↓Acco↓H↑Accr↑Accf↓Acco↓
Pre-train-89.9387.40-90.0389.60--89.6590.80--92.8086.50-
L2*79.3988.3315.3083.5988.3710.3048.1084.0188.2010.6037.1582.7289.909.9020.97
EWC*82.6978.550.1085.7182.570.500.2087.3384.900.902.9086.9389.502.008.27
MAS*80.4474.500.0081.5174.770.000.6082.1875.050.000.6081.9377.900.101.77
LwF85.8285.150.9087.2685.400.402.4087.3584.150.002.3586.0585.700.101.07
DER82.8889.7310.4084.2290.5710.9050.8086.3691.208.8042.0088.5294.703.4024.90
DER++83.4589.739.4084.1390.6311.1050.1087.5091.306.8040.4089.2494.802.2020.80
FDR29.5117.780.5036.2122.700.100.0040.7926.300.000.0047.9133.502.401.27
Retrain48.2033.280.0056.2841.301.300.0061.1447.555.200.2067.2360.6011.000.53
BAD-T74.7465.290.0078.0969.210.000.1076.6566.320.000.0076.2568.170.000.00
LIRF*52.7264.2842.7161.0569.8335.3632.1560.9660.3229.1830.9158.5168.4635.4121.04
SCRUB75.8767.030.0076.2366.330.000.0075.4664.550.000.0077.5670.300.000.00
SCRUB-S83.4479.830.0082.5976.600.000.1079.8171.200.000.0078.2371.400.000.07
GS-LoRA86.1385.000.1087.3385.270.101.1087.8485.150.100.3586.3086.300.200.07
GS-LoRA++86.6385.980.1087.2285.230.300.3087.4884.400.000.6087.0087.500.000.07
FG-OrIU87.0186.730.1088.4587.430.100.1088.9087.350.300.0089.1592.100.100.00
", + "image_path": "481a8bb287438625803ad9178f1036d49e1f67c9e5fa94e78b3d1d7d86d090b8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 489, + 294, + 588 + ], + "blocks": [ + { + "bbox": [ + 160, + 478, + 448, + 489 + ], + "lines": [ + { + "bbox": [ + 160, + 478, + 448, + 489 + ], + "spans": [ + { + "bbox": [ + 160, + 478, + 448, + 489 + ], + "type": "text", + "content": "Table 4. Incremental Unlearning results for classification task on ImageNet100." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 489, + 294, + 588 + ], + "lines": [ + { + "bbox": [ + 58, + 489, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 58, + 489, + 294, + 588 + ], + "type": "table", + "html": "
VariantsTask 1(100-20)Task 2(80-20)
Accr↑Accf↓Accr↑Accf↓Acco↓
Pre-train74.3074.6074.5073.80-
FG-OrIU74.250.0075.500.000.00
w/o fop-f73.010.2076.240.421.28
w/oGOP-f73.920.1076.584.370.82
w/o fop-r73.920.0574.980.210.15
w/oGOP-r71.450.0073.960.000.00
", + "image_path": "053cd906e44a256255dcfb553dbd3257dc76fd678f2569479e6eeea7accb2147.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "text", + "content": "classes significantly declines. Meanwhile, when the regularization term or gradient modification related to forgetting is removed, the performance on the forgetting classes shows a slight improvement. Thus, each component provides clear benefits in achieving a better trade-off between forgetting and remaining while preserving the old forgetting effect." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 128, + 689 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 128, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 128, + 689 + ], + "type": "text", + "content": "5.4. Discussion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "IU is designed to ensure strict privacy protection. This raises several key questions: (a) Does the unlearned model" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 316, + 497, + 553, + 576 + ], + "blocks": [ + { + "bbox": [ + 91, + 591, + 260, + 602 + ], + "lines": [ + { + "bbox": [ + 91, + 591, + 260, + 602 + ], + "spans": [ + { + "bbox": [ + 91, + 591, + 260, + 602 + ], + "type": "text", + "content": "Table 5. Ablation studies on CASIA-Face100." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 497, + 553, + 576 + ], + "lines": [ + { + "bbox": [ + 316, + 497, + 553, + 576 + ], + "spans": [ + { + "bbox": [ + 316, + 497, + 553, + 576 + ], + "type": "table", + "html": "
ExpsMetircPre-trainDERFDRSCRUBGS-LoRAOurs
Exp1SSIM0.940.600.340.750.390.01
PSNR31.3221.2216.8923.4017.608.37
Exp2SSIM0.950.490.660.630.380.02
PSNR31.1519.1722.1521.1217.328.72
Exp3SSIM0.930.530.490.600.410.01
PSNR31.1020.0319.8322.0115.928.55
", + "image_path": "a1530471620ff3d069593db5b8450fb6e28a2da15ee1cd7eb6cde12403b88b2c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 579, + 555, + 603 + ], + "lines": [ + { + "bbox": [ + 313, + 579, + 555, + 603 + ], + "spans": [ + { + "bbox": [ + 313, + 579, + 555, + 603 + ], + "type": "text", + "content": "Table 6. SSIM " + }, + { + "bbox": [ + 313, + 579, + 555, + 603 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 313, + 579, + 555, + 603 + ], + "type": "text", + "content": " and PSNR(db) " + }, + { + "bbox": [ + 313, + 579, + 555, + 603 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 313, + 579, + 555, + 603 + ], + "type": "text", + "content": " scores for the reconstructed images using DIP. Here, Pre-train serves as the upper bound." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 616, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 555, + 665 + ], + "type": "text", + "content": "still retain any private information? (b) Can the forgetting effect be reversed? and (c) How does unlearning impact the feature subspace? To address these questions, we design the following experiment." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 668, + 460, + 678 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 668, + 460, + 678 + ], + "spans": [ + { + "bbox": [ + 313, + 668, + 460, + 678 + ], + "type": "text", + "content": "5.4.1. Semantic Residual Analysis" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 678, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 678, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 678, + 554, + 713 + ], + "type": "text", + "content": "We first investigate the semantic information contained in features after unlearning. We utilize Deep Image Prior (DIP) [46] to reconstruct an image from the features ob" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1963" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 71, + 206, + 217 + ], + "blocks": [ + { + "bbox": [ + 60, + 71, + 206, + 217 + ], + "lines": [ + { + "bbox": [ + 60, + 71, + 206, + 217 + ], + "spans": [ + { + "bbox": [ + 60, + 71, + 206, + 217 + ], + "type": "image", + "image_path": "ff79610a9d205a4eee421c6e9612ec41dc8c28e591552eb48da22e0236ca7245.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 112, + 217, + 149, + 224 + ], + "lines": [ + { + "bbox": [ + 112, + 217, + 149, + 224 + ], + "spans": [ + { + "bbox": [ + 112, + 217, + 149, + 224 + ], + "type": "text", + "content": "(a) Pre-train" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 220, + 72, + 369, + 216 + ], + "blocks": [ + { + "bbox": [ + 220, + 72, + 369, + 216 + ], + "lines": [ + { + "bbox": [ + 220, + 72, + 369, + 216 + ], + "spans": [ + { + "bbox": [ + 220, + 72, + 369, + 216 + ], + "type": "image", + "image_path": "6f2b569fb9a2a2172283ea05658bc1fa68c98ff58a4b73c01c589025f16ea78e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 280, + 217, + 307, + 224 + ], + "lines": [ + { + "bbox": [ + 280, + 217, + 307, + 224 + ], + "spans": [ + { + "bbox": [ + 280, + 217, + 307, + 224 + ], + "type": "text", + "content": "(b) Ours" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 226, + 555, + 249 + ], + "lines": [ + { + "bbox": [ + 55, + 226, + 555, + 249 + ], + "spans": [ + { + "bbox": [ + 55, + 226, + 555, + 249 + ], + "type": "text", + "content": "Figure 4. TSNE visualization with 5 forgetting classes and 10 remaining classes, using different model as feature extractor: (a) the pretrained model, (b) the unlearned model through FG-OrIU, and (c) through GS-LoRA." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 383, + 72, + 554, + 216 + ], + "blocks": [ + { + "bbox": [ + 383, + 72, + 554, + 216 + ], + "lines": [ + { + "bbox": [ + 383, + 72, + 554, + 216 + ], + "spans": [ + { + "bbox": [ + 383, + 72, + 554, + 216 + ], + "type": "image", + "image_path": "3c353ae144c0369e8c264f890b96d1fcd642199b14b1c315ab8990ef3f35ca47.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 436, + 217, + 477, + 224 + ], + "lines": [ + { + "bbox": [ + 436, + 217, + 477, + 224 + ], + "spans": [ + { + "bbox": [ + 436, + 217, + 477, + 224 + ], + "type": "text", + "content": "(c) GS-LoRA" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 252, + 296, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 252, + 296, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 296, + 301 + ], + "type": "text", + "content": "tained from the last layer of the unlearned model. Figure 1 presents the visualization of the reconstructed image, which allows us to infer the semantic information encoded in features. More results are provided in the Appendix ??" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 307, + 296, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 307, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 54, + 307, + 296, + 486 + ], + "type": "text", + "content": "Additionally, we measure the SSIM and PSNR scores for the images generated by DIP from the unlearned model using different IU methods as a quantitative results. As shown in Table 6, the SSIM values of the reconstructed images from the compared methods range between 0.35 and 0.75, with the highest being 0.75 for SCRUB on " + }, + { + "bbox": [ + 54, + 307, + 296, + 486 + ], + "type": "inline_equation", + "content": "exp1" + }, + { + "bbox": [ + 54, + 307, + 296, + 486 + ], + "type": "text", + "content": ", while the PSNR values range between 16 and 24, with the highest being 23.40 for SCRUB on " + }, + { + "bbox": [ + 54, + 307, + 296, + 486 + ], + "type": "inline_equation", + "content": "exp1" + }, + { + "bbox": [ + 54, + 307, + 296, + 486 + ], + "type": "text", + "content": ". This indicates that the reconstructed image achieves a high similarity with the original image, suggesting that the information in features has been partially preserved. In contrast, our method yields an SSIM of only 0.01 and a PSNR of 8.55, indicating a significant difference between the reconstructed and original images. This suggests that the semantic information at the feature level has been substantially obfuscated." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 498, + 222, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 222, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 222, + 510 + ], + "type": "text", + "content": "5.4.2. Forgetting Recoverable Analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 513, + 296, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 513, + 296, + 717 + ], + "spans": [ + { + "bbox": [ + 55, + 513, + 296, + 717 + ], + "type": "text", + "content": "Given an unlearned model obtained through different IU methods, we freeze its backbone and retrain a new classification head using the entire training dataset, which includes both forgetting and remaining classes. As illustrated in Figure 3, MLP* represents a trivial unlearning method where the classification head's parameters for the forgetting classes are simply set to zero. Consequently, the underlying features remain intact, enabling effortless recovery of the forgotten knowledge. Additionally, existing IU methods primarily impose constraints on the final output rather than directly modifying feature representations. As a result, while features may be weakened to some extent, their core structure remains largely unchanged. Thus, these methods allow the forgotten knowledge to be recovered to " + }, + { + "bbox": [ + 55, + 513, + 296, + 717 + ], + "type": "inline_equation", + "content": "15\\% -70\\%" + }, + { + "bbox": [ + 55, + 513, + 296, + 717 + ], + "type": "text", + "content": " of its original performance, which leads to superficial forgetting. In contrast, our method achieves deep forgetting, making recovery infeasible." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 253, + 456, + 265 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 456, + 265 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 456, + 265 + ], + "type": "text", + "content": "5.4.3. Feature Subspace Analysis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 268, + 554, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 268, + 554, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 268, + 554, + 327 + ], + "type": "text", + "content": "To analyze the impact of direct constraints on features and gradients for forgetting and remaining classes, we conduct an experiment. We remove 20 classes while retaining 80, then extract features from 5 forgetting and 10 remaining classes and visualize them using t-SNE." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 328, + 555, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 328, + 555, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 328, + 555, + 579 + ], + "type": "text", + "content": "A key observation is that GS-LoRA causes some forgetting class samples to move closer to or even overlap with certain remaining class samples (e.g., F21, " + }, + { + "bbox": [ + 313, + 328, + 555, + 579 + ], + "type": "inline_equation", + "content": "\\mathrm{F}74 \\rightarrow \\mathrm{R}92" + }, + { + "bbox": [ + 313, + 328, + 555, + 579 + ], + "type": "text", + "content": " in c). This suggests that merely applying constraints at the final output, resulting in degraded features from the forgetting classes entangling with those of the remaining classes. Consequently, this degradation negatively impacts the performance of the remaining classes. In contrast, our method enforces forgetting at both feature and gradient level, ensuring that features from the forgetting classes do not shift toward any remaining class but instead become linearly inseparable within the forgetting set. Notably, an unexpected yet beneficial side effect emerges: in the original PTM, certain remaining class samples (e.g., R59) were slightly entangled with forgetting classes (e.g., F90 and F46). However, after applying our method, these classes become completely disentangled. This observation aligns with the results in Table 3, where the accuracy of the remaining classes improves after unlearning (e.g., " + }, + { + "bbox": [ + 313, + 328, + 555, + 579 + ], + "type": "inline_equation", + "content": "75.36 \\rightarrow 80.51" + }, + { + "bbox": [ + 313, + 328, + 555, + 579 + ], + "type": "text", + "content": "), demonstrating that our method not only enhances forgetting effectiveness but also benefits the generalization of the remaining classes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 591, + 388, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 591, + 388, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 591, + 388, + 602 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": "In this work, we propose FG-OrIU for incremental unlearning in vision PTMs. By enforcing orthogonal constraints on both features and gradients, FG-OrIU achieves irreversible forgetting while maintaining performance on the remaining classes. Its dynamic subspace adaptation ensures robust handling of sequential deletion requests. Extensive experiments confirm its superiority. While FG-OrIU is designed for class-level unlearning, extending it to sample-level settings remains a promising direction for future work." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1964" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 224 + ], + "type": "text", + "content": "This work was supported in part by National Natural Science Foundation of China under Grant 62402430, 62206248, 62476238, Zhejiang Provincial Natural Science Foundation of China under Grant LQN25F020008, and Aeronautical Science Foundation of China 20240048076001. Qian Feng would like to extend his gratitude to Jiahang Tu and Hanbin Zhao from Zhejiang University for their discussions regarding the methods and experiments. Additionally, he is thankful to Mintong Kang from UIUC for his dedicated suggestions on improving the overall writing of the paper." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 232, + 115, + 244 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 115, + 244 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 115, + 244 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 251, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 251, + 297, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 251, + 297, + 306 + ], + "spans": [ + { + "bbox": [ + 61, + 251, + 297, + 306 + ], + "type": "text", + "content": "[1] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In Proceedings of the European conference on computer vision (ECCV), pages 139-154, 2018. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 307, + 296, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 307, + 296, + 351 + ], + "spans": [ + { + "bbox": [ + 61, + 307, + 296, + 351 + ], + "type": "text", + "content": "[2] Anish Athalye, Nicholas Carlini, and David Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In International conference on machine learning, pages 274-283. PMLR, 2018. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 351, + 296, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 351, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 351, + 296, + 384 + ], + "type": "text", + "content": "[3] Ari S Benjamin, David Rolnick, and Konrad Kording. Measuring and regularizing networks in function space. arXiv preprint arXiv:1805.08289, 2018. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 384, + 296, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 384, + 296, + 438 + ], + "spans": [ + { + "bbox": [ + 62, + 384, + 296, + 438 + ], + "type": "text", + "content": "[4] Jacopo Bonato, Marco Cotogni, and Luigi Sabetta. Is retain set all you need in machine unlearning? restoring performance of unlearned models with out-of-distribution images. In European Conference on Computer Vision, pages 1-19. Springer, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 439, + 296, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 439, + 296, + 492 + ], + "spans": [ + { + "bbox": [ + 62, + 439, + 296, + 492 + ], + "type": "text", + "content": "[5] Lucas Bourtoule, Varun Chandrasekaran, Christopher A Choquette-Choo, Hengrui Jia, Adelin Travers, Baiwu Zhang, David Lie, and Nicolas Papernot. Machine unlearning. In 2021 IEEE symposium on security and privacy (SP), pages 141-159. IEEE, 2021. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 493, + 296, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 493, + 296, + 548 + ], + "spans": [ + { + "bbox": [ + 62, + 493, + 296, + 548 + ], + "type": "text", + "content": "[6] Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark experience for general continual learning: a strong, simple baseline. Advances in neural information processing systems, 33:15920-15930, 2020. 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 548, + 296, + 582 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 548, + 296, + 582 + ], + "spans": [ + { + "bbox": [ + 62, + 548, + 296, + 582 + ], + "type": "text", + "content": "[7] Yinzhi Cao and Junfeng Yang. Towards making systems forget with machine unlearning. In 2015 IEEE symposium on security and privacy, pages 463-480. IEEE, 2015. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 582, + 296, + 625 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 582, + 296, + 625 + ], + "spans": [ + { + "bbox": [ + 62, + 582, + 296, + 625 + ], + "type": "text", + "content": "[8] Romit Chatterjee, Vikram Chandawat, Ayush Tarun, Ankur Mali, and Murari Mandal. A unified framework for continual learning and machine unlearning. arXiv preprint arXiv:2408.11374, 2024. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 625, + 296, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 625, + 296, + 658 + ], + "spans": [ + { + "bbox": [ + 62, + 625, + 296, + 658 + ], + "type": "text", + "content": "[9] Jiaao Chen and Diyi Yang. Unlearn what you want to forget: Efficient unlearning for llms. arXiv preprint arXiv:2310.20150, 2023." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 658, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 712 + ], + "type": "text", + "content": "[10] Xinwen Cheng, Zhehao Huang, Wenxin Zhou, Zhengbao He, Ruikai Yang, Yingwen Wu, and Xiaolin Huang. Remaining-data-free machine unlearning by suppressing sample contribution. arXiv preprint arXiv:2402.15109, 2024. 1, 2, 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 715 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 117 + ], + "type": "text", + "content": "[11] Zhi-Yi Chin, Chieh-Ming Jiang, Ching-Chun Huang, Pin-Yu Chen, and Wei-Chen Chiu. Prompting4debugging: Red-teaming text-to-image diffusion models by finding problematic prompts. arXiv preprint arXiv:2309.06135, 2023. 1, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 118, + 554, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 172 + ], + "type": "text", + "content": "[12] Vikram S Chandawat, Ayush K Tarun, Murari Mandal, and Mohan Kankanhalli. Can bad teaching induce forgetting? unlearning in deep networks using an incompetent teacher. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7210-7217, 2023. 1, 2, 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 173, + 554, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 554, + 216 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 554, + 216 + ], + "type": "text", + "content": "[13] Meng Ding, Jinhui Xu, and Kaiyi Ji. Why fine-tuning struggles with forgetting in machine unlearning? theoretical insights and a remedial approach. arXiv preprint arXiv:2410.03833, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 217, + 554, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 554, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 554, + 281 + ], + "type": "text", + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 283, + 554, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 554, + 336 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 554, + 336 + ], + "type": "text", + "content": "[15] Chongyu Fan, Jiancheng Liu, Yihua Zhang, Eric Wong, Dennis Wei, and Sijia Liu. Salun: Empowering machine unlearning via gradient-based weight saliency in both image classification and generation. arXiv preprint arXiv:2310.12508, 2023. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 338, + 554, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 554, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 554, + 381 + ], + "type": "text", + "content": "[16] Jack Foster, Stefan Schoepf, and Alexandra Brintrup. Fast machine unlearning without retraining through selective synaptic dampening. In Proceedings of the AAAI conference on artificial intelligence, pages 12043-12051, 2024. 1, 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 383, + 554, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 554, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 554, + 415 + ], + "type": "text", + "content": "[17] Chongyang Gao, Lixu Wang, Chenkai Weng, Xiao Wang, and Qi Zhu. Practical unlearning for large language models. arXiv preprint arXiv:2407.10223, 2024. 1, 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 415, + 554, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 554, + 458 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 554, + 458 + ], + "type": "text", + "content": "[18] Laura Graves, Vineel Nagisetty, and Vijay Ganesh. Amnesiac machine learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11516-11524, 2021. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 460, + 554, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 554, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 554, + 502 + ], + "type": "text", + "content": "[19] Keltin Grimes, Collin Abidi, Cole Frank, and Shannon Gallagher. Gone but not forgotten: Improved benchmarks for machine unlearning. arXiv preprint arXiv:2405.19211, 2024. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 503, + 554, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 503, + 554, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 503, + 554, + 536 + ], + "type": "text", + "content": "[20] Chuan Guo, Tom Goldstein, Awni Hannun, and Laurens Van Der Maaten. Certified data removal from machine learning models. arXiv preprint arXiv:1911.03030, 2019. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 537, + 554, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 537, + 554, + 602 + ], + "spans": [ + { + "bbox": [ + 317, + 537, + 554, + 602 + ], + "type": "text", + "content": "[21] Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. Ms-celeb-1m: A dataset and benchmark for large-scale face recognition. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 87-102. Springer, 2016. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 603, + 554, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 554, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 554, + 645 + ], + "type": "text", + "content": "[22] Zeyu Han, Chao Gao, Jinyang Liu, Jeff Zhang, and Sai Qian Zhang. Parameter-efficient fine-tuning for large models: A comprehensive survey. arXiv preprint arXiv:2403.14608, 2024. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 647, + 554, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 554, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 554, + 690 + ], + "type": "text", + "content": "[23] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 554, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 554, + 715 + ], + "type": "text", + "content": "[24] Zhehao Huang, Xinwen Cheng, JingHao Zheng, Haoran Wang, Zhengbao He, Tao Li, and Xiaolin Huang. Unified" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1965" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "text", + "content": "gradient-based machine unlearning with remain geometry enhancement. arXiv preprint arXiv:2409.19732, 2024. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 97, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 97, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 97, + 295, + 140 + ], + "type": "text", + "content": "[25] Joel Jang, Dongkeun Yoon, Sohee Yang, Sungmin Cha, Moontae Lee, Lajanugen Logeswaran, and Minjoon Seo. Knowledge unlearning for mitigating privacy risks in language models. arXiv preprint arXiv:2210.01504, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 143, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 143, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 143, + 294, + 196 + ], + "type": "text", + "content": "[26] Jinghan Jia, Jiancheng Liu, Parikshit Ram, Yuguang Yao, Gaowen Liu, Yang Liu, Pranay Sharma, and Sijia Liu. Model sparsity can simplify machine unlearning. Advances in Neural Information Processing Systems, 36:51584-51605, 2023. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 199, + 294, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 199, + 294, + 263 + ], + "spans": [ + { + "bbox": [ + 56, + 199, + 294, + 263 + ], + "type": "text", + "content": "[27] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 266, + 294, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 266, + 294, + 298 + ], + "spans": [ + { + "bbox": [ + 56, + 266, + 294, + 298 + ], + "type": "text", + "content": "[28] Vishnuprasadh Kumaravelu, Sayanta Adhikari, and PK Srijith. Uncle: An unlearning framework for continual learning. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 301, + 294, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 301, + 294, + 345 + ], + "spans": [ + { + "bbox": [ + 56, + 301, + 294, + 345 + ], + "type": "text", + "content": "[29] Meghdad Kurmanji, Peter Triantafillou, Jamie Hayes, and Eleni Triantafillou. Towards unbounded machine unlearning. Advances in neural information processing systems, 36: 1957-1987, 2023. 2, 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 347, + 294, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 347, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 56, + 347, + 294, + 380 + ], + "type": "text", + "content": "[30] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 382, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 382, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 382, + 294, + 415 + ], + "type": "text", + "content": "[31] Bo Liu, Qiang Liu, and Peter Stone. Continual learning and private unlearning. In Conference on Lifelong Learning Agents, pages 243-254. PMLR, 2022. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 417, + 294, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 417, + 294, + 471 + ], + "spans": [ + { + "bbox": [ + 56, + 417, + 294, + 471 + ], + "type": "text", + "content": "[32] Gaoyang Liu, Xiaogiang Ma, Yang Yang, Chen Wang, and Jiangchuan Liu. Federaser: Enabling efficient client-level data removal from federated learning models. In 2021 IEEE/ACM 29th International Symposium on Quality of Service (IWQOS), pages 1-10. IEEE, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 474, + 294, + 537 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 474, + 294, + 537 + ], + "spans": [ + { + "bbox": [ + 56, + 474, + 294, + 537 + ], + "type": "text", + "content": "[33] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 541, + 294, + 583 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 541, + 294, + 583 + ], + "spans": [ + { + "bbox": [ + 56, + 541, + 294, + 583 + ], + "type": "text", + "content": "[34] Laiqiao Qin, Tianqing Zhu, Linlin Wang, and Wanlei Zhou. Machine unlearning on pre-trained models by residual feature alignment using lora. arXiv preprint arXiv:2411.08443, 2024. 1, 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 587, + 294, + 630 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 587, + 294, + 630 + ], + "spans": [ + { + "bbox": [ + 56, + 587, + 294, + 630 + ], + "type": "text", + "content": "[35] Sadia Qureshi, Thanveer Shaik, Xiaohui Tao, Haoran Xie, Lin Li, Jianming Yong, and Xiaohua Jia. Exploring incremental unlearning: Techniques, challenges, and future directions. arXiv preprint arXiv:2502.16708, 2025. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 633, + 294, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 633, + 294, + 654 + ], + "spans": [ + { + "bbox": [ + 56, + 633, + 294, + 654 + ], + "type": "text", + "content": "[36] Protection Regulation. General data protection regulation. Intouch, 25:1-5, 2018. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 656, + 294, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 656, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 656, + 294, + 677 + ], + "type": "text", + "content": "[37] Jeffrey Rosen. The right to be forgotten. *Stan. L. Rev. Online*, 64:88, 2011. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "type": "text", + "content": "[38] Dan Roth and Wen-tau Yih. Probabilistic reasoning for entity & relation recognition. In COLING 2002: The 19th International Conference on Computational Linguistics, 2002. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 127 + ], + "type": "text", + "content": "[39] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 129, + 554, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 554, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 554, + 160 + ], + "type": "text", + "content": "[40] Gobinda Saha, Isha Garg, and Kaushik Roy. Gradient projection memory for continual learning. arXiv preprint arXiv:2103.09762, 2021. 2, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 162, + 554, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 554, + 215 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 554, + 215 + ], + "type": "text", + "content": "[41] Thanveer Shaik, Xiaohui Tao, Haoran Xie, Lin Li, Xiaofeng Zhu, and Qing Li. Exploring the landscape of machine unlearning: A comprehensive survey and taxonomy. IEEE Transactions on Neural Networks and Learning Systems, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 217, + 553, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 248 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 248 + ], + "type": "text", + "content": "[42] Yingdan Shi and Ren Wang. Redefining machine unlearning: A conformal prediction-motivated approach. arXiv preprint arXiv:2501.19403, 2025. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 250, + 553, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 553, + 281 + ], + "type": "text", + "content": "[43] Takashi Shibata, Go Irie, Daiki Ikami, and Yu Mitsuzumi. Learning with selective forgetting. In *IJCAI*, page 6, 2021. 3, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 282, + 553, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 282, + 553, + 324 + ], + "spans": [ + { + "bbox": [ + 316, + 282, + 553, + 324 + ], + "type": "text", + "content": "[44] Mihai Surdeanu and Heng Ji. Overview of the english slot filling track at the tac2014 knowledge base population evaluation. In Proc. Text Analysis Conference (TAC2014), 2014. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 327, + 553, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 380 + ], + "type": "text", + "content": "[45] Jiahang Tu, Qian Feng, Chufan Chen, Jiahua Dong, Hanbin Zhao, Chao Zhang, and Hui Qian. Ce-sdwv: Effective and efficient concept erasure for text-to-image diffusion models via a semantic-driven word vocabulary. arXiv preprint arXiv:2501.15562, 2025. 1, 2, 3, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 382, + 553, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 382, + 553, + 424 + ], + "spans": [ + { + "bbox": [ + 316, + 382, + 553, + 424 + ], + "type": "text", + "content": "[46] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9446-9454, 2018. 1, 2, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 426, + 553, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 457 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 457 + ], + "type": "text", + "content": "[47] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 460, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 502 + ], + "type": "text", + "content": "[48] Cheng-Long Wang, Qi Li, Zihang Xiang, and Di Wang. Has approximate machine unlearning been evaluated properly? from auditing to side effects. arXiv e-prints, pages arXiv-2403, 2024. 1, 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 504, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 553, + 536 + ], + "type": "text", + "content": "[49] Heng Xu, Tianqing Zhu, Lefeng Zhang, Wanlei Zhou, and Philip S Yu. Machine unlearning: A survey. ACM Computing Surveys, 56(1):1-36, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 537, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 569 + ], + "type": "text", + "content": "[50] Jie Xu, Zihan Wu, Cong Wang, and Xiaohua Jia. Machine unlearning: Solutions and challenges. IEEE Transactions on Emerging Topics in Computational Intelligence, 2024. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 571, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 571, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 571, + 553, + 613 + ], + "type": "text", + "content": "[51] Lingling Xu, Haoran Xie, Si-Zhao Joe Qin, Xiaohui Tao, and Fu Lee Wang. Parameter-efficient fine-tuning methods for pretrained language models: A critical review and assessment. arXiv preprint arXiv:2312.12148, 2023. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 615, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 615, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 615, + 553, + 646 + ], + "type": "text", + "content": "[52] Haonan Yan, Xiaoguang Li, Ziyao Guo, Hui Li, Fenghua Li, and Xiaodong Lin. Arcane: An efficient architecture for exact machine unlearning. In *IJCAI*, page 19, 2022. 2, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 647, + 553, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 680 + ], + "type": "text", + "content": "[53] Jingwen Ye, Yifang Fu, Jie Song, Xingyi Yang, Songhua Liu, Xin Jin, Mingli Song, and Xinchao Wang. Learning with recoverable forgetting. In ECCV, 2022. 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "type": "text", + "content": "[54] Dong Yi, Zhen Lei, Shengcai Liao, and Stan Z Li. Learning face representation from scratch. arXiv preprint arXiv:1411.7923, 2014. 5" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1966" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 520 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[55] Jaehong Yoon, Shoubin Yu, Vaidehi Patil, Huaxiu Yao, and Mohit Bansal. Safree: Training-free and adaptive guard for safe text-to-image and video generation. arXiv preprint arXiv:2410.12761, 2024. 2, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "text", + "content": "[56] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. Advances in Neural Information Processing Systems, 33:5824-5836, 2020. 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "text", + "content": "[57] Dan Zhang, Tao Feng, Lilong Xue, Yuandong Wang, Yuxiao Dong, and Jie Tang. Parameter-efficient fine-tuning for foundation models. arXiv preprint arXiv:2501.13787, 2025. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 240 + ], + "type": "text", + "content": "[58] Haibo Zhang, Toru Nakamura, Takamasa Isohara, and Kouichi Sakurai. A review on machine unlearning. SN Computer Science, 4(4):337, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 285 + ], + "type": "text", + "content": "[59] Yuanhan Zhang, Zhenfei Yin, Jing Shao, and Ziwei Liu. Benchmarking omni-vision representation through the lens of visual realms. In European Conference on Computer Vision, pages 594-611. Springer, 2022. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 286, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 342 + ], + "type": "text", + "content": "[60] Hongbo Zhao, Bolin Ni, Junsong Fan, Yuxi Wang, Yuntao Chen, Gaofeng Meng, and Zhaoxiang Zhang. Continual forgetting for pre-trained vision models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28631-28642, 2024. 1, 3, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 342, + 294, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 294, + 385 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 294, + 385 + ], + "type": "text", + "content": "[61] Hongbo Zhao, Fei Zhu, Bolin Ni, Feng Zhu, Gaofeng Meng, and Zhaoxiang Zhang. Practical continual forgetting for pretrained vision models. arXiv preprint arXiv:2501.09705, 2025. 1, 2, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 387, + 294, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 387, + 294, + 452 + ], + "spans": [ + { + "bbox": [ + 56, + 387, + 294, + 452 + ], + "type": "text", + "content": "[62] Zhen Zhao, Zhizhong Zhang, Xin Tan, Jun Liu, Yanyun Qu, Yuan Xie, and Lizhuang Ma. Rethinking gradient projection continual learning: Stability/plasticity feature space decoupling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3718-3727, 2023. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 453, + 294, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 453, + 294, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 453, + 294, + 475 + ], + "type": "text", + "content": "[63] Yaoyao Zhong and Weihong Deng. Face transformer for recognition. arXiv preprint arXiv:2103.14803, 2021. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 476, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 476, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 476, + 294, + 520 + ], + "type": "text", + "content": "[64] Chen Zhu, Ankit Singh Rawat, Manzil Zaheer, Srinadh Bhojanapalli, Daliang Li, Felix Yu, and Sanjiv Kumar. Modifying memories in transformer models. arXiv preprint arXiv:2012.00363, 2020. 3" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1967" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/FICGen_ Frequency-Inspired Contextual Disentanglement for Layout-driven Degraded Image Generation/24509018-dca5-4420-ad1a-dd11a5974d9d_content_list.json b/2025/FICGen_ Frequency-Inspired Contextual Disentanglement for Layout-driven Degraded Image Generation/24509018-dca5-4420-ad1a-dd11a5974d9d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f4675c641ef004bfd80bbb712262baaa6b3791ef --- /dev/null +++ b/2025/FICGen_ Frequency-Inspired Contextual Disentanglement for Layout-driven Degraded Image Generation/24509018-dca5-4420-ad1a-dd11a5974d9d_content_list.json @@ -0,0 +1,1664 @@ +[ + { + "type": "text", + "text": "FICGen: Frequency-Inspired Contextual Disentanglement for Layout-driven Degraded Image Generation", + "text_level": 1, + "bbox": [ + 186, + 130, + 813, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wenzhuang Wang $^{1,2}$ , Yifan Zhao $^{1*}$ , Mingcan Ma $^{2}$ , Ming Liu $^{2}$ , Zhonglin Jiang $^{2}$ , Yong Chen $^{2}$ , Jia Li $^{1*}$ $^{1}$ State Key Laboratory of Virtual Reality Technology and Systems, SCSE&QRI, Beihang University \n $^{2}$ Geely Automobile Research Institute (Ningbo) Co., Ltd", + "bbox": [ + 104, + 202, + 895, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{wz_wang, zhaoyf, jiali}@buaa.edu.cn", + "bbox": [ + 346, + 258, + 643, + 273 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e52df08b170e168c189b4efce8c392106b33dcbb6739c1715062db63cf4229e3.jpg", + "image_caption": [ + "(a) Contextual Illusion Dilemma" + ], + "image_footnote": [], + "bbox": [ + 106, + 304, + 705, + 424 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bae07c4cca4983817de4840ae34b4044aae24659988177a16250b6f238aa11ec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 306, + 874, + 431 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3f5b4734eab6bf4189c78fd9299c564b5beb98d241710036527c98722faee8b8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 107, + 436, + 705, + 503 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7f4d482c918dea9fcd464ff72fabf735832fb6b5680f8bbd1eb14eb100419a48.jpg", + "image_caption": [ + "(b) Robust Generation for Out-of-Distribution Layouts", + "(c) Alignment & Trainability" + ], + "image_footnote": [], + "bbox": [ + 107, + 503, + 705, + 570 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c40921c12eebe00483cafdbeb616afcdc8a55eef5f6faa8157f34168d6938478.jpg", + "image_caption": [ + "Figure 1. Visualizations of degraded image generation by our FICGen. FICGen achieves superior quality compared to existing L2I methods and demonstrates strong robustness to OOD layouts. Moreover, it shows excellent trainability for boosting object detectors, i.e., Faster R-CNN (R50) [29]. The alignment and trainability are evaluated on the ExDARK [18] and DIOR-H [14], respectively." + ], + "image_footnote": [], + "bbox": [ + 712, + 436, + 875, + 561 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 655, + 326, + 670 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Layout-to-image (L2I) generation has exhibited promising results in natural domains, but suffers from limited generative fidelity and weak alignment with user-provided layouts when applied to degraded scenes (i.e., low-light, underwater). We primarily attribute these limitations to the \"contextual illusion dilemma\" in degraded conditions, where foreground instances are overwhelmed by context-dominant frequency distributions. Motivated by this, our paper proposes a new Frequency-Inspired Contextual Disentangle-ment Generative (FICGen) paradigm, which seeks to transfer frequency knowledge of degraded images into the latent diffusion space, thereby facilitating the rendering of", + "bbox": [ + 86, + 689, + 485, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "degraded instances and their surroundings via contextual frequency-aware guidance. To be specific, FICGen consists of two major steps. Firstly, we introduce a learnable dual-query mechanism, each paired with a dedicated frequency resampler, to extract contextual frequency prototypes from pre-collected degraded exemplars in the training set. Secondly, a visual-frequency enhanced attention is employed to inject frequency prototypes into the degraded generation process. To alleviate the contextual illusion and attribute leakage, an instance coherence map is developed to regulate latent-space disentanglement between individual instances and their surroundings, coupled with an adaptive spatial-frequency aggregation module to reconstruct spatial-frequency mixed degraded representations. Extensive experiments on 5 benchmarks involving a variety of degraded scenarios—from severe low-light to mild blur—demonstrate", + "bbox": [ + 511, + 656, + 908, + 898 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Correspondence should be addressed to Yifan Zhao and Jia Li.", + "bbox": [ + 107, + 888, + 444, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19097", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "that FICGen consistently surpasses existing L2I methods in terms of generative fidelity, alignment and downstream auxiliary trainability.", + "bbox": [ + 89, + 90, + 485, + 137 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 165, + 222, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual perception [21, 33] depends on a large amount of annotated datasets, particularly in degraded conditions such as low-light [18], and adverse weather [13]. Unfortunately, acquiring adequate datasets for these extreme conditions remains a major obstacle due to the high cost of manual collection and annotation. For instance, the number of images in ExDARK (7,363) [18] is only about 1/20th that of COCO 2017 (118,287) [16], highlighting the severe data scarcity. In light of this, alternative strategies are emerging that shift the focus from a \"model-centric\" to a \"data-centric\" perspective, where advanced generative models are used to synthesize realistic images for bolstering downstream tasks [40].", + "bbox": [ + 88, + 190, + 485, + 372 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Witnessing the phenomenal capabilities of text-to-image (T2I) diffusion models [27, 32], numerous advances have been explored in conditioning image generation on visual labels, thereby enabling fine-grained controllability beyond what textual prompts alone can achieve. Such labels are often represented by masks [20, 45], depths [51], and bounding boxes (bboxes) [8, 15, 46, 53, 54], thereby facilitating visual generation via layout-to-image (L2I) synthesis. Among them, bboxes are the most intuitive for guiding the location of visual entities during L2I generation.", + "bbox": [ + 88, + 372, + 483, + 523 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite the success of L2I methods [15, 53, 54] in natural scenarios, they face challenges when adapting to practical degraded conditions. In Fig. 1 (a), aerial objects (i.e., vehicles) are often small in size and visually resemble surrounding structures (i.e., bridges), while underwater species (i.e., fish) frequently blend with other nearby creatures (i.e., corals), resulting in contextual entanglement. Such cases, difficult even for the human eye to distinguish and predominantly occurring in degraded scenarios, lead to generative illusions regarding object quantity, location, and interaction—termed the “contextual illusion dilemma” in our work.", + "bbox": [ + 88, + 523, + 483, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generally, as shown in Fig. 2, natural image exhibits a relatively balanced distribution of high-frequency (HF) and low-frequency (LF) components, making the distinction between instances and contexts pronounced. In contrast, HF details of instances in degraded contexts are attenuated, while LF contexts dominate the overall frequency distribution. Motivated by the frequency properties, we propose a new frequency-inspired contextual disentanglement paradigm to tackle the dilemma. Specifically, we first extract HF and LF prototypes from degraded contexts and associate them with each object and its surroundings. Second, a learnable dual-query mechanism equipped with frequency perceiver resamplers is introduced to simultaneously capture contextual frequency knowledge (i.e., boundaries and colours). Third, we inject", + "bbox": [ + 88, + 688, + 485, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/36cac4420640b7fe574c6332faec8ca5a94fd87d8eadf4e7f65fb4aafa82e5c9.jpg", + "image_caption": [ + "Figure 2. Frequency spectrum comparison between natural and degraded images." + ], + "image_footnote": [], + "bbox": [ + 516, + 89, + 903, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the knowledge into the latent diffusion space via a visual-frequency enhanced attention, enabling better perception of frequency attributes during degraded generation. Meanwhile, an instance coherence map is designed to disentangle degraded instances from their surroundings, mitigating contextual entanglement and attribute leakage. Finally, a spatial-frequency aggregation module is developed to adaptively recombine them to obtain refined degraded representations. With the above efforts, FICGen offers a unified approach for degraded image synthesis from a frequency-inspired perspective. We evaluate FICGen on 5 degraded benchmarks, covering low-light (ExDARK [18]), underwater (RUOD [6]), remote sensing (DIOR [14]), adverse weather (DAWN [13]), and blur (VOC 2012 [5]) conditions. Extensive experiments demonstrate that FICGen achieves superior image quality while maintaining strong alignment with geometric layouts. In Fig.1(b), FICGen can perform robust OOD generation on unseen layouts. Fig.1(c) shows that FICGen achieves better alignment than the real test set across all classes except \"boat\", effectively narrowing the real-synthetic domain gap. Furthermore, FICGen consistently improves downstream tasks by $\\sim 2.0$ mAP and substantially boosts detection accuracy for certain classes, i.e., $\\sim 6.0$ (32.2 vs. 38.1) on \"airport\", underscoring its strong trainability.", + "bbox": [ + 511, + 233, + 908, + 595 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are threefold: (1) We propose FICGen, the first attempt to address the contextual illusion dilemma in degraded image generation. (2) We introduce a learnable dual-query mechanism to capture contextual frequency knowledge, which is injected into the latent diffusion space via a visual-frequency enhanced attention module, and further complemented by an instance coherence map and a spatial-frequency aggregation module to enable latent-space disentanglement and consolidation. (3) Extensive experiments on 5 degraded datasets show that our FICGen outperforms SOTA L2I methods in both quantitative and qualitative evaluations. Notably, FICGen can be seamlessly integrated as a plug-and-play component to enhance other diffusion-based generative models.", + "bbox": [ + 511, + 597, + 910, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 821, + 653, + 838 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Text-Driven Image Synthesis", + "text_level": 1, + "bbox": [ + 511, + 847, + 769, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-driven image synthesis aims to create visual content conditioned on diverse textual descriptions. Early endeav-", + "bbox": [ + 511, + 869, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19098", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ors focus on leveraging conditional GANs [28, 44, 49, 50] to generate realistic images. Recent advances in diffusion [11, 23, 24, 27, 30, 32, 34] and auto-regressive models [7, 26, 48] have revolutionized image synthesis by enabling stable training and producing high-fidelity images. Typically, DALL-E [26] exemplifies an auto-regressive work that treats textual and visual tokens as a unified data sequence, demonstrating remarkable zero-shot capability. Among diffusion-based efforts, latent diffusion model (LDM) [30] efficiently implements a denoising Markov chain in a low-dimensional latent space, striking a balance between generative quality and computational cost. Nevertheless, these text-prompted methods struggle with fine-grained control over object attributes (i.e., position).", + "bbox": [ + 89, + 90, + 485, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Layout-Driven Image Synthesis", + "text_level": 1, + "bbox": [ + 89, + 311, + 372, + 328 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Layout-driven image synthesis [2, 15, 22, 46, 53] encodes geometric layouts as positional tokens to enable fine-grained controllability. However, these methods often suffer from attribute leakage, where certain layouts mistakenly attend to irrelevant regions. Multi-instance control [41, 42, 54] addresses this by partitioning layouts into disjoint components, allowing each layout to focus on its corresponding region. Nonetheless, these approaches are confined to natural image domains. AeroGen [36] pioneers layout-conditioned generation in the context of remote sensing imagery, but it still struggles with semantic ambiguity and limited layout controllability. CC-Diff [52] further improves contextual coherence by modeling foreground-background relations, but overlooks frequency-aware cues and remains vulnerable to contextual illusion in various degraded scenes. In contrast, we propose a frequency-inspired contextual disentanglement paradigm, which introduces contextual frequency guidance into the generation process, delivering more realistic degraded images across diverse circumstances.", + "bbox": [ + 89, + 333, + 485, + 621 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 632, + 184, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 657, + 232, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Latent Diffusion Model. Unlike vanilla DDPM [11], which generates images from Gaussian noise in pixel space, LDM [30] operates in a compressed latent space. Specifically, an input image $\\mathcal{I}$ is first encoded into a latent representation $z_{0}$ , which is then diffused into a noisy latent state via the forward process: $z_{t} = \\alpha_{t}z_{0} + \\sigma_{t}\\epsilon$ , where $\\alpha_{t}$ and $\\sigma_{t}$ are the fixed schedule, $\\epsilon \\sim \\mathcal{N}(0,1)$ is added standard Gaussian noise term and $t \\in [0,T]$ denotes the timestep. Then, a denoising U-Net $(\\epsilon_{\\theta})$ [31] is trained to predict the step-wise noise $\\epsilon$ , with the training objective $\\mathcal{L}_{LDM}$ defined in Eq.1.", + "bbox": [ + 89, + 679, + 485, + 830 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\theta} {\\operatorname {m i n}} \\mathcal {L} _ {L D M} = \\mathbb {E} _ {z _ {0}, \\epsilon \\sim \\mathcal {N} (0, 1), t, \\tau} [ | | \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, \\tau) | | _ {2} ^ {2} ], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 839, + 483, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\tau$ is the guidance embedding (i.e., text and bboxes).", + "bbox": [ + 89, + 869, + 477, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem Formulation. Given instance layouts comprising", + "bbox": [ + 89, + 885, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$N$ object bboxes: $\\mathcal{B} = (\\{l_i, b_i\\}_{i=1}^N)$ , where the location of $i$ -th object $b_i = [x_i^1, y_i^1, x_i^2, y_i^2]$ (i.e., top-left and bottom-right corners), and $l_i$ is its corresponding semantic category. $\\mathcal{V}$ denotes the global textual description of the degraded image. In this regard, our FICGen $(\\mathcal{G}(\\cdot))$ strives to generate a degraded image conditioned on both spatial layout and semantics: $\\mathcal{I} = \\mathcal{G}(\\mathcal{B}, \\epsilon, \\mathcal{V})$ .", + "bbox": [ + 511, + 90, + 908, + 198 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Bridge Queries with Frequency Prototypes", + "text_level": 1, + "bbox": [ + 511, + 205, + 877, + 222 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing L2I methods [2, 8, 15, 53, 54] primarily emphasize foreground layout conditions to render instances in natural domains, where fronts and contexts share a relatively uniform frequency distribution. Nonetheless, the naive strategy is inadequate for degraded domains with contextual illusion.", + "bbox": [ + 511, + 226, + 906, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Exploring degraded contexts reveals that frequency disentanglement, a longstanding yet insightful concept, effectively mines weakened HF instances from the dominant LF surroundings. To this end, we seek to extract contextual frequency knowledge and transfer it into the latent space, thus achieving more realistic degraded generation via frequency-aware guidance. To achieve this, we face two main issues: 1) how to represent the frequency knowledge for degraded contexts? and 2) how to transfer frequency knowledge into the latent diffusion space? The answers are detailed below.", + "bbox": [ + 511, + 303, + 908, + 453 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Frequency Prototypes Representation. For the former issue, our goal is to represent contextual frequency knowledge (i.e., textures) using frequency prototypes. To this end, we take two steps to manifest them, as depicted in Fig. 3(a). Taking the foreground exemplar as an example, we first build a class-specific dictionary by collecting same-class instances from the training set. A degraded instance is then sampled and assigned to the $i$ -th bbox based on its position. All selected instances form a foreground exemplar that is often overwhelmed in degraded scenarios. Second, stacked convolutional networks extract its intermediate feature maps $\\mathbf{X} \\in \\mathbb{R}^{H \\times W}$ , which are then transformed into the frequency domain $\\mathcal{F}(\\mathbf{X})$ via a Discrete Fourier Transform (DFT):", + "bbox": [ + 511, + 454, + 910, + 650 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {\\mathcal {F}} (u, v) = \\frac {1}{H \\times W} \\sum_ {h = 0} ^ {H - 1} \\sum_ {w = 0} ^ {W - 1} \\mathbf {X} (h, w) e ^ {- j 2 \\pi (u h + v w)}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 657, + 906, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{X}_{\\mathcal{F}}$ is the complex-valued DFT output, $|v|$ and $|u|$ denote the width and height of normalised frequencies. Next, to isolate HF components, a binary mask $\\mathbf{M}_{\\mathcal{F}}$ is applied:", + "bbox": [ + 511, + 713, + 906, + 758 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {\\mathcal {F}} (u, v) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} u \\notin \\left[ \\frac {H}{2} - \\gamma H, \\frac {H}{2} + \\gamma H \\right], \\\\ & v \\notin \\left[ \\frac {W}{2} - \\gamma W, \\frac {W}{2} + \\gamma W \\right], \\\\ 0, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 766, + 906, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\gamma$ controls the size of the HF region. The masked frequencies $\\mathbf{X}_{\\mathcal{F}}\\mathbf{M}_{\\mathcal{F}}$ are then enhanced using learnable channelwise weights $\\mathbf{W}_{\\mathcal{F}}$ , and projected back to the spatial domain:", + "bbox": [ + 511, + 830, + 908, + 876 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} ^ {\\uparrow} = \\mathcal {F} ^ {- 1} \\left(\\mathbf {X} _ {\\mathcal {F}} \\mathbf {M} _ {\\mathcal {F}}\\right) \\cdot \\mathbf {W} _ {\\mathcal {F}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 883, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19099", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fa0b8feb95be71860094d4ca4929b0b6f7d5b173ea8038d81185383cfb8a2272.jpg", + "image_caption": [ + "Figure 3. The overview pipeline of our FICGen, consisting of (a) Frequency Perceiver Resamplers, (b) Visual-Frequency Enhanced Attention, and (c) Adaptive Spatial-Frequency Aggregation." + ], + "image_footnote": [], + "bbox": [ + 94, + 85, + 903, + 334 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{F}^{-1}$ denotes the inverse DFT. The result $\\mathbf{X}^{\\dagger}$ represents the enhanced HF feature map. A similar process is applied to the complementary region using $1 - \\mathbf{M}_F$ , yielding the LF feature map $\\mathbf{X}^{\\downarrow}$ . Inspired by the prototypical learning, we adopt their mean pooled results as HF and LF prototypes, i.e., $\\mathbf{p}^{\\uparrow} = \\{p_i^\\uparrow\\}_{i=1}^N \\in \\mathbb{R}^{N \\times d}$ and $p^\\downarrow \\in \\mathbb{R}^d$ , where $\\mathbf{p}^{\\uparrow}$ corresponds to the HF representations of degraded instances, $p^\\downarrow$ reflects the LF characteristics of surrounding contexts, and $d$ denotes the feature dimension.", + "bbox": [ + 88, + 397, + 483, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Frequency Perceiver Resamplers. To address the latter issue, inspired by the Perceiver architecture [47], we adopt a learnable dual-query mechanism that employs separate frequency resamplers to bridge HF and LF prototypes, thereby capturing contextual frequency cues:", + "bbox": [ + 89, + 534, + 483, + 609 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {q} _ {i} ^ {\\uparrow} = \\operatorname {H F - R e s a m p l e r} \\left(\\mathbf {Q} ^ {\\uparrow}, \\phi_ {k 1} ^ {r} \\left(p _ {i} ^ {\\uparrow}\\right), \\phi_ {v 1} ^ {r} \\left(p _ {i} ^ {\\uparrow}\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 619, + 482, + 640 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {q} ^ {\\downarrow} = \\operatorname {L F - R e s a m p l e r} \\left(\\mathbf {Q} ^ {\\downarrow}, \\phi_ {k 1} ^ {g} \\left(p ^ {\\downarrow}\\right), \\phi_ {v 1} ^ {g} \\left(p ^ {\\downarrow}\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 641, + 483, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\phi_{*}^{r},\\phi_{*}^{g}\\in \\mathbb{R}^{d\\times d}$ are linear projection layers. The dual queries $\\mathbf{Q}^{\\uparrow}$ and $\\mathbf{Q}^{\\downarrow}$ are learned through their respective resamplers, each composed of multiple transformer blocks [38], enabling continuous interactions between the dual queries and frequency prototypes, yielding frequency-aware instance and context tokens, i.e., $\\mathbf{q}_i^\\uparrow$ and $\\mathbf{q}^{\\downarrow}$ .", + "bbox": [ + 89, + 671, + 483, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the above dedicated design, the synergy between frequency prototypes and the dual-query mechanism enables the simultaneous perception of essential frequency cues and contextual appearances, which are then incorporated into the degraded image generation process.", + "bbox": [ + 89, + 763, + 483, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Contextual Frequency Knowledge Transfer", + "text_level": 1, + "bbox": [ + 89, + 848, + 459, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Previous L2I efforts [2, 4, 15, 54] only focus on instance generation conditioned on layouts, while neglecting the sur", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "rounding context. This limitation is particularly detrimental in degraded scenarios involving contextual entanglement.", + "bbox": [ + 511, + 398, + 903, + 428 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Visual-Frequency Enhanced Attention. To transfer frequency properties into the degraded generation, we employ a visual-frequency enhanced attention module (Fig.3(c)) that injects frequency-aware tokens $\\mathbf{q}_i^\\uparrow$ and $\\mathbf{q}^{\\downarrow}$ into the latent diffusion space. Specifically, each bbox $b_{i}$ is first encoded using a Fourier embedding, followed by a multi-layer perceptron $(\\mathbf{E}_{box})$ to produce positional tokens. In parallel, its semantic class $l_{i}$ is encoded via a CLIP text encoder $(\\mathbf{E}_{clip})$ [25]. Notably, $\\mathbf{q}_i^\\uparrow$ is then integrated with them to provide frequency-aware cues for instance generation. As for the surrounding, $\\mathbf{q}^{\\downarrow}$ is fused with text tokens extracted from the global caption $\\mathcal{V}$ , thereby facilitating holistic contextual guidance.", + "bbox": [ + 511, + 429, + 906, + 625 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {R} _ {i} = \\left[ \\mathbf {q} _ {i} ^ {\\uparrow}; \\mathbf {E} _ {c l i p} \\left(l _ {i}\\right); \\mathbf {E} _ {b o x} (\\text {F o u r i e r} (b _ {i})) \\right], \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 637, + 906, + 656 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {G} = \\left[ \\mathbf {q} ^ {\\downarrow}; \\mathbf {E} _ {c l i p} (\\mathcal {Y}) \\right], \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 659, + 906, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $[;]$ represents concatenate operation.", + "bbox": [ + 511, + 689, + 797, + 704 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It is preferable for each layout to affect only its corresponding local region. To this end, an instance coherence map $\\hat{\\mathbf{M}}_i$ is utilized to isolate individual instances and enforce precise interactions between layout conditions and their associated latent features.", + "bbox": [ + 511, + 705, + 906, + 780 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {i} ^ {r} = \\operatorname {s o f t m a x} \\left(\\frac {Q _ {v} \\phi_ {k 2} ^ {r} (\\mathbf {R} _ {i})}{\\sqrt {d}}\\right) \\cdot \\phi_ {v 2} ^ {r} (\\mathbf {R} _ {i}) \\odot \\hat {\\mathbf {M}} _ {i}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 790, + 906, + 824 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} ^ {g} = \\operatorname {s o f t m a x} \\left(\\frac {Q _ {v} \\phi_ {k 2} ^ {g} (\\mathbf {G})}{\\sqrt {d}}\\right) \\cdot \\phi_ {v 2} ^ {g} (\\mathbf {G}) \\odot \\hat {\\mathbf {M}} ^ {g}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 825, + 906, + 859 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Q_{v}$ denotes the linear projection of latent features, $\\odot$ is element-wise multiplication, and $\\phi_{*}^{r},\\phi_{*}^{g}$ are linear layers.", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19100", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, $\\mathbf{f}_i^r$ indicates the latent feature of the $i$ -th instance, while $\\mathbf{f}^g$ corresponds to the contextual representation.", + "bbox": [ + 89, + 90, + 485, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Instance Coherence Map. To avert feature contamination and attribute leakage, we first construct an instance-level mask $\\hat{\\mathbf{M}}_i$ from each bbox layout, as in [42, 54], which enables precise generation of instance-specific latent features:", + "bbox": [ + 89, + 121, + 485, + 183 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {M}} _ {i} (x, y) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} [ x, y ] \\in b _ {i}, \\\\ 0, & \\text {i f} [ x, y ] \\notin b _ {i}, \\end{array} \\right. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 190, + 483, + 232 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $x$ and $y$ denote the mask coordinates. On the basis, we further separate foreground instances from their surroundings by constructing the complementary region $\\hat{\\mathbf{M}}^g$ , which encompasses all areas outside the instance-level masks, thus enabling contextual disentanglement:", + "bbox": [ + 89, + 239, + 483, + 315 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {M}} ^ {g} = 1 - \\sum_ {i = 1} ^ {N} \\hat {\\mathbf {M}} _ {i}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 323, + 483, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, we combine $\\mathbf{f}_i^r$ and $\\mathbf{f}^g$ to produce the unified latent degraded representations $\\mathbf{F} = [\\{\\mathbf{f}_i^r\\}_{i=1}^N, \\mathbf{f}^g]$ , which serve as the key for degraded image generation.", + "bbox": [ + 89, + 373, + 483, + 421 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Adaptive Spatial-Frequency Aggregation", + "text_level": 1, + "bbox": [ + 89, + 428, + 441, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After harvesting the degraded representations $\\mathbf{F}$ , we move forward to jointly consolidate and reconstruct them. In contrast to existing methods that either directly sum multiple instances [52] or combine them solely within the spatial domain [54], which often suffer from \"object omission and merging\" for overlapping instances, as illustrated in Fig. 3(c), we aim to adaptively aggregate the degraded instances and their surroundings in a spatial-frequency mixed manner. \"Spatial\" captures contextual relational dependencies, while \"frequency\" emphasizes fine-grained attributes across instances, i.e., boundary sharpness and textures, which can be formulated as follows:", + "bbox": [ + 88, + 449, + 485, + 630 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} ^ {s} = \\mathbf {S A M} \\left(\\left[ \\sum_ {i = 1} ^ {N} \\mathbf {f} _ {i} ^ {r}, \\mathbf {f} ^ {g} \\right]\\right), \\mathbf {F} ^ {f} = \\mathbf {F A M} \\left(\\left[ \\sum_ {i = 1} ^ {N} \\mathbf {f} _ {i} ^ {r}, \\mathbf {f} ^ {g} \\right]\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 638, + 483, + 679 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\delta^ {f i n a l} = \\sum_ {i = 1} ^ {N + 1} \\operatorname {s o f t m a x} \\left(\\delta^ {f u s i o n}\\right) \\left[ \\sum_ {i = 1} ^ {N} \\mathbf {f} _ {i} ^ {r}, \\mathbf {f} ^ {g} \\right], \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 688, + 483, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where SAM and FAM represent the spatial and frequency aggregation modules, respectively, consisting of standard self- and frequency-attention mechanisms. $\\delta^{fusion} = \\zeta * ([\\mathbf{F}^s,\\mathbf{F}^f])$ , where $\\zeta$ is a learnable single-layer depthwise convolution to enhance local perception. Finally, the refined degraded representation $\\delta^{final} \\in \\mathbb{R}^{h\\times w\\times c}$ is decoded into the target degraded image.", + "bbox": [ + 89, + 734, + 483, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Learning Procedure", + "text_level": 1, + "bbox": [ + 89, + 848, + 282, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During training, the parameters of the pretrained LDM remain fixed, with optimization focused solely on FICGen.", + "bbox": [ + 89, + 869, + 485, + 901 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c2364cd391b77b654d1c3d3f2e174c029435df0c31b9031d960db3e5fd4d818f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DetectorMethodFID ↓mAP ↑\\( AP_{50} \\) ↑\\( AP_{75} \\) ↑
DIOR-H [14]
Faster R-CNNOracle-33.455.635.0
Faster R-CNNMIGC [54]31.6421.838.417.5
Faster R-CNNCC-Diff [52]30.8823.642.421.4
Faster R-CNNFICGen (ours)31.2527.648.727.6
RUOD [6]
Faster R-CNNOracle-50.580.254.4
Faster R-CNNMIGC [54]26.5027.254.124.6
Faster R-CNNCC-Diff [52]25.2129.758.427.9
Faster R-CNNFICGen (ours)25.1037.068.636.5
blurred VOC 2012 [5]
Faster R-CNNOracle-31.556.532.4
Faster R-CNNMIGC [54]62.6634.765.833.1
Faster R-CNNCC-Diff [52]62.2036.767.636.3
Faster R-CNNFICGen (ours)58.0240.770.342.7
ExDARK [18]
Cascade R-CNNOracle-37.265.837.8
Cascade R-CNNMIGC [54]45.7632.463.529.5
Cascade R-CNNCC-Diff [52]44.2635.165.634.1
Cascade R-CNNFICGen (ours)42.4042.573.045.1
DAWN [13]
Cascade R-CNNOracle-27.346.426.3
Cascade R-CNNMIGC [54]70.1012.632.38.6
Cascade R-CNNCC-Diff [52]68.5615.733.914.8
Cascade R-CNNFICGen (ours)68.3122.644.321.5
", + "bbox": [ + 519, + 89, + 901, + 395 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparison of fidelity and alignment on various degraded conditions. The performance is evaluated by off-the-shelf detectors Faster R-CNN [29] and Cascade R-CNN [1] on synthetic test set. \"Oracle\" denotes the real test baseline. The best results are in red and the second best results are in underline.", + "bbox": [ + 511, + 405, + 906, + 476 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The learning objective integrates the original mean-square error loss and the contextual frequency guidance:", + "bbox": [ + 511, + 502, + 906, + 532 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta^ {\\prime}} \\mathcal {L} _ {F I C G e n} = \\mathbb {E} _ {z _ {0}, \\epsilon \\sim \\mathcal {N} (0, 1), t, \\mathcal {Y}} [ \\| \\epsilon - \\mathcal {G} _ {\\theta , \\theta^ {\\prime}} (z _ {t}, t, \\mathcal {Y}, \\mathcal {B}, \\mathcal {Q}) \\| _ {2} ^ {2} ], \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 542, + 929, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{Q} = [\\{\\mathbf{q}_i^\\uparrow\\}_{i=1}^N, \\mathbf{q}^\\downarrow]$ denotes the set of frequency-aware tokens. While $\\theta$ and $\\theta'$ denote the frozen and trainable parameters of the pretrained LDM and our FICGen, respectively.", + "bbox": [ + 511, + 580, + 906, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 511, + 642, + 638, + 659 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setups", + "text_level": 1, + "bbox": [ + 511, + 667, + 710, + 684 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. Our FICGen is evaluated on five degraded-context datasets: ExDARK [18], RUOD [6], DAWN [13], DIOR [14], and blurred VOC 2012 [5]. More details are provided in the Supplementary Material.", + "bbox": [ + 511, + 688, + 906, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. Our FICGen is built upon the pre-trained SDv1.5 checkpoint and deployed only on the mid-level (i.e., $8 \\times 8$ ) and lowest-resolution (i.e., $16 \\times 16$ ) decoder layers. For training, all images are resized to $512 \\times 512$ . We adopt the AdamW [19] optimizer with a learning rate of $1\\mathrm{e}^{-4}$ , $\\beta_{1} = 0.9$ , and $\\beta_{2} = 0.999$ , and fine-tune for 300 epochs on 8 A100 with a batch size of 320. For inference, we use the EulerDiscreteScheduler [12] with 50 sampling steps and a classifier-free guidance scale [10] of 7.5.", + "bbox": [ + 511, + 750, + 908, + 885 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. Following [2, 15], we comprehend", + "bbox": [ + 511, + 885, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19101", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/132e4ca7a2c5c573edbed77f99eab531d017588675f0b4d323799b3bbbd38482.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison of degraded images generated by different L2I methods across various scenarios. Zoom in for better details." + ], + "image_footnote": [], + "bbox": [ + 93, + 85, + 911, + 372 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a3a51c6fc63a2bf7b44533942ac9cb0b53f359dade392ba7b0c37bb21cea3d16.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFID ↓YOLO Score ↑
mAPAP50AP75
Oracle-39.257.143.0
LostGAN [35]57.1014.327.315.2
LayoutDiffusion [53]45.3120.037.419.3
ReCo [46]42.5621.140.723.1
MIGC [54]31.6424.541.825.6
GLIGEN [15]41.3125.844.427.8
CC-Diff [52]30.8826.444.228.5
AeroGen [36]38.5729.854.231.6
FICGen (ours)31.2531.249.934.6
", + "bbox": [ + 102, + 419, + 478, + 573 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "sively evaluate our FICGen from three aspects: fidelity, alignment, and trainability. Fidelity refers to the perceptual quality between synthetic and real images, measured by the Fréchet Inception Distance (FID) [9]. While alignment assesses spatial and semantic correspondences between layout conditions and generated instances, quantified via COCO-style average precision [16] with off-the-shelf detectors. Trainability examines the potential of synthetic images for auxiliary downstream training, assessed by mean Average Precision (mAP) under various training settings.", + "bbox": [ + 88, + 638, + 485, + 790 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 89, + 799, + 230, + 814 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.1. Quantitative Evaluations", + "text_level": 1, + "bbox": [ + 89, + 821, + 310, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fidelity and Alignment. Tab. 1 reports the quantitative comparison of fidelity and alignment via reporting the COCO-style AP, where pre-trained Faster R-CNN (R50) [29] and Cascade R-CNN (R50) [1] are employed to infer synthetic", + "bbox": [ + 89, + 839, + 485, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/09ea972b3f0f4b27d736a05ffb1d8c47d1a244c7bb47d87dafe4828ad6050061.jpg", + "table_caption": [ + "Table 2. Quantitative comparison on the DIOR-H test set, with detection performance evaluated using a pre-trained YOLOv8 [37]." + ], + "table_footnote": [], + "table_body": "
MethodExDARK [18]RUOD [6]
mAP↑AP50↑AP75↑mAP↑AP50↑AP75↑
Oracle38.367.839.858.485.564.8
MIGC [54]28.659.224.121.143.917.9
CC-Diff [52]31.361.828.829.757.828.0
FICGen (ours)38.568.539.537.167.136.7
", + "bbox": [ + 521, + 419, + 901, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Quantitative comparison on ExDARK and RUOD test set, evaluated by a pre-trained Deformable-DETR [55].", + "bbox": [ + 511, + 512, + 906, + 540 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "images, with the results evaluated against corresponding real annotations. Notably, our FICGen successfully outperforms the existing MIGC [54] and CC-Diff [52] in terms of both FID and AP metrics, achieving remarkable FID scores of 25.10 and 42.40 on the RUOD and ExDARK, respectively. Moreover, for remote sensing object detection, FICGen reaches $82\\%$ of the mAP relative to real data (i.e., 27.6 vs. 33.4), substantially outperforming MIGC (21.8) and CC-Diff (23.6) by a large margin. Intriguingly, FICGen even exceeds the real ExDARK test set in AP, highlighting its ability to generate degraded instances that precisely adhere to layouts while bridging the real-to-synthetic domain gap.", + "bbox": [ + 511, + 566, + 906, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To further verify the generative applicability across different detectors, we evaluate FICGen using YOLOv8 [37] and Deformable-DETR [55], as shown in Tabs. 2 and 3. FICGen surpasses AeroGen by 1.4 and 3.0 in mAP and $\\mathrm{AP}_{75}$ on the DIOR-H, and outperforms CC-Diff by 7.2 and 7.4 on the ExDARK and RUOD, respectively.", + "bbox": [ + 511, + 750, + 908, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "**Trainability.** Following the trainability protocol in [2], we use augmented annotations as layout inputs to generate synthetic training images and mix them with real ones, effectively doubling the training set. As shown in the last", + "bbox": [ + 511, + 839, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19102", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3664b6e34240d910051098c1baf49187167e7bf9b163af3a8f8e39767c9d6e7b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DetectorBackboneTraining SetObject Detection (AP) for Sampled Classes /%mAP\\( AP_{50} \\)\\( AP_{75} \\)
# Real# Syntheticbicyclebusmotorbikeboatcarcupbottledog
Train with Full Real Data
RetinaNetPVT-MR:5.1k-44.858.231.331.136.025.528.747.835.265.433.9
Faster R-CNNX101R:5.1k-49.060.835.630.139.530.934.149.338.669.339.6
Train with Pure Synthetic Data
RetinaNetPVT-M-MIGC (S: 5.1k)32.846.720.419.922.514.116.033.423.549.019.1
Faster R-CNNX101-MIGC (S: 5.1k)33.648.218.520.423.117.021.133.224.749.422.1
RetinaNetPVT-M-CC-Diff (S: 5.1k)34.345.521.620.222.015.416.634.924.149.920.6
Faster R-CNNX101-CC-Diff (S: 5.1k)38.047.521.519.124.318.321.833.825.552.321.7
RetinaNetPVT-M-FICGen (S: 5.1k)35.150.423.220.024.916.216.833.325.351.022.1
Faster R-CNNX101-FICGen (S: 5.1k)38.149.921.722.024.318.920.436.426.751.425.1
Train with Synthetic Data and Finetune on Real Data
RetinaNetPVT-MR:5.1kMIGC (S: 5.1k)47.362.134.732.637.427.629.649.037.167.536.9
Faster R-CNNX101R:5.1kMIGC (S: 5.1k)51.361.838.031.540.032.135.849.739.970.740.9
RetinaNetPVT-MR:5.1kCC-Diff (S: 5.1k)47.960.934.331.536.427.130.548.137.067.035.8
Faster R-CNNX101R:5.1kCC-Diff (S: 5.1k)50.262.238.731.140.632.035.449.439.870.440.7
RetinaNetPVT-MR:5.1kFICGen (S: 5.1k)47.562.235.733.737.827.630.947.637.667.736.8
Faster R-CNNX101R:5.1kFICGen(S: 5.1k)51.462.938.232.840.333.135.649.940.670.942.1
Train with Real & Synthetic Data
Faster R-CNNX101R:5.1kMIGC (S: 5.1k)51.763.036.231.739.731.835.249.639.569.739.8
Faster R-CNNX101R:5.1kCC-Diff (S: 5.1k)50.961.636.531.840.231.634.750.339.570.040.7
Faster R-CNNX101R:5.1kFICGen (S: 5.1k)50.561.838.631.640.432.636.549.140.170.541.9
", + "bbox": [ + 94, + 88, + 903, + 380 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/306c94d8f2f1053bde67494f9736e0b6c8f30ab31e27003cab907a28e5782771.jpg", + "table_caption": [ + "Table 4. Quantitative comparison of trainability on the ExDARK dataset, evaluated by RetinaNet [17] (PVT-M [39]) and Faster R-CNN [29] (X101 [43]) under three training settings." + ], + "table_footnote": [], + "table_body": "
MethodDIOR-H [14]DAWN [13]ExDARK [18]VOC 2012 [5]
airport ↑mAP ↑\\( AP_{50} \\) ↑\\( AP_{75} \\) ↑bus ↑mAP ↑\\( AP_{50} \\) ↑\\( AP_{75} \\) ↑motorbike ↑mAP ↑\\( AP_{50} \\) ↑\\( AP_{75} \\) ↑boat ↑mAP ↑\\( AP_{50} \\) ↑\\( AP_{75} \\) ↑
Oracle32.233.455.635.021.824.846.023.033.335.867.434.542.948.376.852.5
MIGC [54]36.934.156.136.121.624.943.324.634.437.167.836.642.549.477.254.2
CC-Diff [52]37.634.356.136.326.225.846.125.834.137.067.736.644.449.677.354.3
FICGen (ours)38.135.556.837.627.525.944.526.537.137.768.037.747.350.577.455.4
", + "bbox": [ + 94, + 433, + 903, + 506 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Quantitative comparison of trainability on the DIOR, DAWN and ExDARK and natural VOC 2012 datasets, with detection evaluated using the Faster R-CNN (R50) object detector.", + "bbox": [ + 89, + 517, + 906, + 545 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b0e26b309c3cc17a7a07893e74749b15db0dac1b8c80d21cb496ac653965b172.jpg", + "image_caption": [ + "Figure 5. Visualization results generated by our FICGen on unseen geometric layouts augmented via flipping and translation." + ], + "image_footnote": [], + "bbox": [ + 93, + 570, + 482, + 770 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "three rows of Tab. 4, we demonstrate for the first time that synthetic degraded images can significantly enhance the detection performance of downstream object detectors (i.e., +2.4 for bottle and +3.0 for motorbike). Moreover,", + "bbox": [ + 89, + 839, + 485, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "we observe that training with synthetic images followed by fine-tuning on real ones yields the best performance. Tab. 5 further provides a comparative analysis under various degraded conditions. Crucially, FICGen consistently achieves the highest gains, boosting AP by 5.7 for \"bus\" (DAWN) and 3.8 for \"motorbike\" (ExDARK), highlighting its efficacy in mitigating data scarcity in degraded scenarios.", + "bbox": [ + 511, + 571, + 906, + 676 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generalization. To evaluate the generalization of our FICGen, we conduct further experiments not only in severely degraded scenarios such as low-light and underwater conditions with horizontal layout control, but also in blurred and natural VOC 2012 [5], and oriented DIOR-R [3]. Extensive results demonstrate that synthetic images generated by FICGen exhibit superior quality and alignment. Further details are provided in the Supplementary Material.", + "bbox": [ + 511, + 678, + 908, + 799 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.2. Qualitative Evaluations", + "text_level": 1, + "bbox": [ + 511, + 806, + 723, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 4 presents a visual comparison of degraded images generated by various L2I methods. Our FICGen demonstrates superior performance in two key aspects: (a) Enhanced Layout Coherence. It resolves the dilemma of \"object omission and merging\", accurately aligning generated instances with", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19103", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/acffbba831add96592c62aec3840f73e866dc205ffd60a2392745ad6d05205be.jpg", + "image_caption": [ + "Figure 6. Performance convergence curves when generated degraded images are incorporated into downstream detection." + ], + "image_footnote": [], + "bbox": [ + 96, + 90, + 488, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e1c385a64c31b40b6b6a48e09da037633704c5910690148ca4fecd8708760b45.jpg", + "image_caption": [ + "Figure 7. Generalization results of FICGen on mildly blurred and natural scenes, as well as under oriented layout control." + ], + "image_footnote": [], + "bbox": [ + 94, + 257, + 488, + 592 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the provided layouts in terms of quantity, scale, and position (i.e., aerial tennis courts and rainy cars), and effectively handling small objects (i.e., aerial vehicles). (b) Strong Occlusion Handling. In heavily occluded scenarios (i.e., underwater corals and fish), FICGen reconstructs dense instances while maintaining spatial consistency, whereas MIGC and CC-Diff fail to enforce overlapped layout constraints. Furthermore, Fig. 5 shows that FICGen generalizes effectively to unseen layouts, including those extending beyond image boundaries. As illustrated in Fig. 6, the synthesized degraded images generated by FICGen enhance downstream detection performance by improving mAP and accelerating training convergence. Finally, Fig. 7 highlights the generalization capability of FICGen in blurred and natural scenes, as well as under oriented layout controls. Additional visualizations are provided in the Supplementary Material.", + "bbox": [ + 88, + 659, + 485, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/35cee3ec536b66b30238cc3ba224b8fdd6b62a79f669b641d55a94c8bd1e63e7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
FPRVFEAASFAmAP\\( AP_{50} \\)\\( AP_{75} \\)
XXX15.734.812.0
X✓X23.244.622.7
✓X✓18.536.313.9
✓✓X25.546.325.1
X✓✓25.646.925.3
✓✓✓27.648.727.6
", + "bbox": [ + 524, + 89, + 903, + 185 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Ablation study of FPR, VFEA, and ASFA modules, evaluated by the pre-trained Faster R-CNN (R50) on the DIOR test set.", + "bbox": [ + 511, + 194, + 906, + 222 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 256, + 678, + 271 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Frequency Perceiver Resamplers. The foundation of frequency prototype learning lies in the FPR module, which combines a dual-query mechanism with contextual frequency knowledge. As shown in the last two rows of Tab. 6, FPR significantly improves the alignment scores from 25.6/46.9/25.3 to 27.6/48.7/27.6, demonstrating the effectiveness of contextual frequency guidance.", + "bbox": [ + 511, + 281, + 906, + 387 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visual-Frequency Enhanced Attention. VFEA serves as a bridge to inject frequency knowledge (i.e., boundaries and textures) into the latent space while integrating the instance coherence map for contextual disentanglement. As shown in the second row of Tab. 6, even without FPR and ASFA, VFEA achieves competitive performance over the baseline, reaching 23.2/44.6/22.7 in AP.", + "bbox": [ + 511, + 391, + 908, + 496 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Adaptive Spatial-Frequency Aggregation. ASFA is designed to reconstruct refined degraded representations by integrating spatial and frequency perspectives. As evidenced by the results in the fourth and sixth rows of Tab. 6, incorporating ASFA improves the average precision (AP) from 25.5/46.3/25.1 to 27.6/48.7/27.6. This substantial gain underscores its critical role in consolidating contextual information and enhancing the fidelity of degraded image generation.", + "bbox": [ + 511, + 500, + 908, + 621 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 511, + 645, + 782, + 660 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce the task of degraded image generation and define the contextual illusion dilemma as its core challenge. To address this, we propose a frequency-inspired generative paradigm, FICGen. We first extract frequency prototypes from degraded contexts and employ a learnable dual-query mechanism to mine contextual frequency knowledge, which is then transferred into the latent diffusion space via a visual-frequency enhanced attention, coupled with an instance coherence map and a spatial-frequency aggregation module, to disentangle foreground instances from their surroundings and consolidate them into refined degraded representations. Extensive experiments on 5 degraded image datasets validate the effectiveness of FICGen. In future work, we plan to extend this frequency-inspired paradigm to 3D content generation for broader applications.", + "bbox": [ + 511, + 674, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19104", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is partially supported by grants from the National Natural Science Foundation of China (No. 62132002 and No. 62202010), Beijing Nova Program (No.20250484786), and the Fundamental Research Funds for the Central Universities.", + "bbox": [ + 89, + 114, + 485, + 189 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 202, + 187, + 218 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6154-6162, 2018. 5, 6", + "[2] Kai Chen, Enze Xie, Zhe Chen, Yibo Wang, Lanqing Hong, Zhenguo Li, and Dit-Yan Yeung. Geodiffusion: Text-prompted geometric control for object detection data generation. arXiv preprint arXiv:2306.04607, 2023. 3, 4, 5, 6", + "[3] Gong Cheng, Jiabao Wang, Ke Li, Xingxing Xie, Chunbo Lang, Yanqing Yao, and Junwei Han. Anchor-free oriented proposal generator for object detection. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2022. 7", + "[4] Jiaxin Cheng, Xiao Liang, Xingjian Shi, Tong He, Tianjun Xiao, and Mu Li. Layoutdiffuse: Adapting foundational diffusion models for layout-to-image generation. arXiv preprint arXiv:2302.08908, 2023. 4", + "[5] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 2, 5, 7", + "[6] Chenping Fu, Risheng Liu, Xin Fan, Puyang Chen, Hao Fu, Wanqi Yuan, Ming Zhu, and Zhongxuan Luo. Rethinking general underwater object detection: Datasets, challenges, and solutions. Neurocomputing, 517:243-256, 2023. 2, 5, 6", + "[7] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. In European Conference on Computer Vision, pages 89-106. Springer, 2022. 3", + "[8] Yuchao Gu, Yipin Zhou, Yunfan Ye, Yixin Nie, Licheng Yu, Pingchuan Ma, Kevin Qinghong Lin, and Mike Zheng Shou. Roictrl: Boosting instance control for visual generation. arXiv preprint arXiv:2411.17949, 2024. 2, 3", + "[9] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6", + "[10] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 5", + "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3", + "[12] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in neural information processing systems, 35:26565-26577, 2022. 5" + ], + "bbox": [ + 93, + 227, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Mourad A Kenk and Mahmoud Hassaballah. Dawn: vehicle detection in adverse weather nature dataset. arXiv preprint arXiv:2008.05402, 2020. 2, 5, 7", + "[14] Ke Li, Gang Wan, Gong Cheng, Liqui Meng, and Junwei Han. Object detection in optical remote sensing images: A survey and a new benchmark. ISPRS journal of photogrammetry and remote sensing, 159:296-307, 2020. 1, 2, 5, 7", + "[15] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22511-22521, 2023. 2, 3, 4, 5, 6", + "[16] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer vision-ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part v 13, pages 740-755. Springer, 2014. 2, 6", + "[17] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 7", + "[18] Yuen Peng Loh and Chee Seng Chan. Getting to know low-light images with the exclusively dark dataset. Computer Vision and Image Understanding, 178:30-42, 2019. 1, 2, 5, 6, 7", + "[19] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5", + "[20] Zhengyao Lv, Yuxiang Wei, Wangmeng Zuo, and KwanYee K Wong. Place: Adaptive layout-semantic fusion for semantic image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9264–9274, 2024. 2", + "[21] Mingcan Ma, Changqun Xia, Chenxi Xie, Xiaowu Chen, and Jia Li. Boosting broader receptive fields for salient object detection. IEEE Transactions on Image Processing, 32:1026-1038, 2023. 2", + "[22] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI conference on artificial intelligence, pages 4296–4304, 2024. 3", + "[23] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 3", + "[24] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 3", + "[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 4" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19105", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International conference on machine learning, pages 8821-8831. Pmlr, 2021. 3", + "[27] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2): 3, 2022. 2, 3", + "[28] Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In International conference on machine learning, pages 1060-1069. PMLR, 2016. 3", + "[29] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 1, 5, 6, 7", + "[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3", + "[31] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 3", + "[32] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2, 3", + "[33] Zhe Shan, Yang Liu, Lei Zhou, Cheng Yan, Heng Wang, and Xia Xie. Ros-sam: High-quality interactive segmentation for remote sensing moving object. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 3625-3635, 2025. 2", + "[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3", + "[35] Wei Sun and Tianfu Wu. Image synthesis from reconfigurable layout and style. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10531-10540, 2019. 6", + "[36] Datalog Tang, Xiangyong Cao, Xuan Wu, Jialin Li, Jing Yao, Xueru Bai, Dongsheng Jiang, Yin Li, and Deyu Meng. Aerogen: enhancing remote sensing object detection with diffusion-driven data generation. arXiv preprint arXiv:2411.15497, 2024. 3, 6", + "[37] Ultralytics. Ultralytics yolov8. https://github.com/ultralytics/ultralytics, 2023.6", + "[38] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 4", + "[39] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyra-" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "mid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 568-578, 2021. 7", + "[40] Wenzhuang Wang, Mingcan Ma, Yong Chen, Changqun Xia, Zhenbao Liang, and Jia Li. Freegen: Bridging visual-linguistic discrepancies towards diffusion-based pixel-level data synthesis. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7916–7924, 2025. 2", + "[41] Xudong Wang, Trevor Darrell, Sai Saketh Rambhatla, Rohit Girdhar, and Ishan Misra. Instancediffusion: Instance-level control for image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6232-6242, 2024. 3", + "[42] Yinwei Wu, Xianpan Zhou, Bing Ma, Xuefeng Su, Kai Ma, and Xinchao Wang. Ifadapter: Instance feature control for grounded text-to-image generation. arXiv preprint arXiv:2409.08240, 2024. 3, 5", + "[43] Saining Xie, Ross Girshick, Piotr Dólár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1492–1500, 2017. 7", + "[44] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Finegrained text to image generation with attentional generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1316-1324, 2018. 3", + "[45] Han Xue, Zhiwu Huang, Qianru Sun, Li Song, and Wenjun Zhang. Freestyle layout-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14256-14266, 2023. 2", + "[46] Zhengyuan Yang, Jianfeng Wang, Zhe Gan, Linjie Li, Kevin Lin, Chenfei Wu, Nan Duan, Zicheng Liu, Ce Liu, Michael Zeng, et al. Reco: Region-controlled text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14246-14255, 2023. 2, 3, 6", + "[47] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 4", + "[48] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 3", + "[49] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 5907-5915, 2017. 3", + "[50] Han Zhang, Jing Yu Koh, Jason Baldridge, Honglak Lee, and Yinfei Yang. Cross-modal contrastive learning for text-to-image generation. In Proceedings of the IEEE/CVF con" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19106", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ference on computer vision and pattern recognition, pages 833-842, 2021. 3", + "[51] Jingyang Zhang, Shiwei Li, Yuanxun Lu, Tian Fang, David McKinnon, Yanghai Tsin, Long Quan, and Yao Yao. Joint-net: Extending text-to-image diffusion for dense distribution modeling. arXiv preprint arXiv:2310.06347, 2023. 2", + "[52] Mu Zhang, Yunfan Liu, Yue Liu, Hongtian Yu, and Qixiang Ye. Cc-diff: Enhancing contextual coherence in remote sensing image synthesis. arXiv preprint arXiv:2412.08464, 2024. 3, 5, 6, 7", + "[53] Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22490-22499, 2023. 2, 3, 6", + "[54] Dewei Zhou, You Li, Fan Ma, Xiaoting Zhang, and Yi Yang. Mige: Multi-instance generation controller for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6818-6828, 2024. 2, 3, 4, 5, 6, 7", + "[55] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 6" + ], + "bbox": [ + 91, + 92, + 483, + 430 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "19107", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/FICGen_ Frequency-Inspired Contextual Disentanglement for Layout-driven Degraded Image Generation/24509018-dca5-4420-ad1a-dd11a5974d9d_model.json b/2025/FICGen_ Frequency-Inspired Contextual Disentanglement for Layout-driven Degraded Image Generation/24509018-dca5-4420-ad1a-dd11a5974d9d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..27fb05c4e5761d264f1daf0c4bab9bb64b0053a5 --- /dev/null +++ b/2025/FICGen_ Frequency-Inspired Contextual Disentanglement for Layout-driven Degraded Image Generation/24509018-dca5-4420-ad1a-dd11a5974d9d_model.json @@ -0,0 +1,2279 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.187, + 0.131, + 0.815, + 0.177 + ], + "angle": 0, + "content": "FICGen: Frequency-Inspired Contextual Disentanglement for Layout-driven Degraded Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.203, + 0.897, + 0.257 + ], + "angle": 0, + "content": "Wenzhuang Wang\\(^{1,2}\\), Yifan Zhao\\(^{1*}\\), Mingcan Ma\\(^{2}\\), Ming Liu\\(^{2}\\), Zhonglin Jiang\\(^{2}\\), Yong Chen\\(^{2}\\), Jia Li\\(^{1*}\\) \n\\(^{1}\\)State Key Laboratory of Virtual Reality Technology and Systems, SCSE&QRI, Beihang University \n\\(^{2}\\)Geely Automobile Research Institute (Ningbo) Co., Ltd" + }, + { + "type": "text", + "bbox": [ + 0.348, + 0.26, + 0.645, + 0.275 + ], + "angle": 0, + "content": "{wz_wang, zhaoyf, jiali}@buaa.edu.cn" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.305, + 0.707, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.309, + 0.425, + 0.493, + 0.435 + ], + "angle": 0, + "content": "(a) Contextual Illusion Dilemma" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.307, + 0.875, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.437, + 0.706, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.505, + 0.706, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.257, + 0.571, + 0.568, + 0.583 + ], + "angle": 0, + "content": "(b) Robust Generation for Out-of-Distribution Layouts" + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.437, + 0.877, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.713, + 0.57, + 0.877, + 0.584 + ], + "angle": 0, + "content": "(c) Alignment & Trainability" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.593, + 0.908, + 0.637 + ], + "angle": 0, + "content": "Figure 1. Visualizations of degraded image generation by our FICGen. FICGen achieves superior quality compared to existing L2I methods and demonstrates strong robustness to OOD layouts. Moreover, it shows excellent trainability for boosting object detectors, i.e., Faster R-CNN (R50) [29]. The alignment and trainability are evaluated on the ExDARK [18] and DIOR-H [14], respectively." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.656, + 0.327, + 0.671 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.69, + 0.486, + 0.871 + ], + "angle": 0, + "content": "Layout-to-image (L2I) generation has exhibited promising results in natural domains, but suffers from limited generative fidelity and weak alignment with user-provided layouts when applied to degraded scenes (i.e., low-light, underwater). We primarily attribute these limitations to the \"contextual illusion dilemma\" in degraded conditions, where foreground instances are overwhelmed by context-dominant frequency distributions. Motivated by this, our paper proposes a new Frequency-Inspired Contextual Disentangle-ment Generative (FICGen) paradigm, which seeks to transfer frequency knowledge of degraded images into the latent diffusion space, thereby facilitating the rendering of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.657, + 0.909, + 0.899 + ], + "angle": 0, + "content": "degraded instances and their surroundings via contextual frequency-aware guidance. To be specific, FICGen consists of two major steps. Firstly, we introduce a learnable dual-query mechanism, each paired with a dedicated frequency resampler, to extract contextual frequency prototypes from pre-collected degraded exemplars in the training set. Secondly, a visual-frequency enhanced attention is employed to inject frequency prototypes into the degraded generation process. To alleviate the contextual illusion and attribute leakage, an instance coherence map is developed to regulate latent-space disentanglement between individual instances and their surroundings, coupled with an adaptive spatial-frequency aggregation module to reconstruct spatial-frequency mixed degraded representations. Extensive experiments on 5 benchmarks involving a variety of degraded scenarios—from severe low-light to mild blur—demonstrate" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.889, + 0.446, + 0.902 + ], + "angle": 0, + "content": "*Correspondence should be addressed to Yifan Zhao and Jia Li." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19097" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.138 + ], + "angle": 0, + "content": "that FICGen consistently surpasses existing L2I methods in terms of generative fidelity, alignment and downstream auxiliary trainability." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.166, + 0.223, + 0.182 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.191, + 0.486, + 0.373 + ], + "angle": 0, + "content": "Visual perception [21, 33] depends on a large amount of annotated datasets, particularly in degraded conditions such as low-light [18], and adverse weather [13]. Unfortunately, acquiring adequate datasets for these extreme conditions remains a major obstacle due to the high cost of manual collection and annotation. For instance, the number of images in ExDARK (7,363) [18] is only about 1/20th that of COCO 2017 (118,287) [16], highlighting the severe data scarcity. In light of this, alternative strategies are emerging that shift the focus from a \"model-centric\" to a \"data-centric\" perspective, where advanced generative models are used to synthesize realistic images for bolstering downstream tasks [40]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.373, + 0.485, + 0.524 + ], + "angle": 0, + "content": "Witnessing the phenomenal capabilities of text-to-image (T2I) diffusion models [27, 32], numerous advances have been explored in conditioning image generation on visual labels, thereby enabling fine-grained controllability beyond what textual prompts alone can achieve. Such labels are often represented by masks [20, 45], depths [51], and bounding boxes (bboxes) [8, 15, 46, 53, 54], thereby facilitating visual generation via layout-to-image (L2I) synthesis. Among them, bboxes are the most intuitive for guiding the location of visual entities during L2I generation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.524, + 0.485, + 0.689 + ], + "angle": 0, + "content": "Despite the success of L2I methods [15, 53, 54] in natural scenarios, they face challenges when adapting to practical degraded conditions. In Fig. 1 (a), aerial objects (i.e., vehicles) are often small in size and visually resemble surrounding structures (i.e., bridges), while underwater species (i.e., fish) frequently blend with other nearby creatures (i.e., corals), resulting in contextual entanglement. Such cases, difficult even for the human eye to distinguish and predominantly occurring in degraded scenarios, lead to generative illusions regarding object quantity, location, and interaction—termed the “contextual illusion dilemma” in our work." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.689, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Generally, as shown in Fig. 2, natural image exhibits a relatively balanced distribution of high-frequency (HF) and low-frequency (LF) components, making the distinction between instances and contexts pronounced. In contrast, HF details of instances in degraded contexts are attenuated, while LF contexts dominate the overall frequency distribution. Motivated by the frequency properties, we propose a new frequency-inspired contextual disentanglement paradigm to tackle the dilemma. Specifically, we first extract HF and LF prototypes from degraded contexts and associate them with each object and its surroundings. Second, a learnable dual-query mechanism equipped with frequency perceiver resamplers is introduced to simultaneously capture contextual frequency knowledge (i.e., boundaries and colours). Third, we inject" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.09, + 0.905, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.179, + 0.907, + 0.207 + ], + "angle": 0, + "content": "Figure 2. Frequency spectrum comparison between natural and degraded images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.234, + 0.909, + 0.597 + ], + "angle": 0, + "content": "the knowledge into the latent diffusion space via a visual-frequency enhanced attention, enabling better perception of frequency attributes during degraded generation. Meanwhile, an instance coherence map is designed to disentangle degraded instances from their surroundings, mitigating contextual entanglement and attribute leakage. Finally, a spatial-frequency aggregation module is developed to adaptively recombine them to obtain refined degraded representations. With the above efforts, FICGen offers a unified approach for degraded image synthesis from a frequency-inspired perspective. We evaluate FICGen on 5 degraded benchmarks, covering low-light (ExDARK [18]), underwater (RUOD [6]), remote sensing (DIOR [14]), adverse weather (DAWN [13]), and blur (VOC 2012 [5]) conditions. Extensive experiments demonstrate that FICGen achieves superior image quality while maintaining strong alignment with geometric layouts. In Fig.1(b), FICGen can perform robust OOD generation on unseen layouts. Fig.1(c) shows that FICGen achieves better alignment than the real test set across all classes except \"boat\", effectively narrowing the real-synthetic domain gap. Furthermore, FICGen consistently improves downstream tasks by \\(\\sim 2.0\\) mAP and substantially boosts detection accuracy for certain classes, i.e., \\(\\sim 6.0\\) (32.2 vs. 38.1) on \"airport\", underscoring its strong trainability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.911, + 0.81 + ], + "angle": 0, + "content": "In summary, our contributions are threefold: (1) We propose FICGen, the first attempt to address the contextual illusion dilemma in degraded image generation. (2) We introduce a learnable dual-query mechanism to capture contextual frequency knowledge, which is injected into the latent diffusion space via a visual-frequency enhanced attention module, and further complemented by an instance coherence map and a spatial-frequency aggregation module to enable latent-space disentanglement and consolidation. (3) Extensive experiments on 5 degraded datasets show that our FICGen outperforms SOTA L2I methods in both quantitative and qualitative evaluations. Notably, FICGen can be seamlessly integrated as a plug-and-play component to enhance other diffusion-based generative models." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.823, + 0.655, + 0.839 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.771, + 0.866 + ], + "angle": 0, + "content": "2.1. Text-Driven Image Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.91, + 0.902 + ], + "angle": 0, + "content": "Text-driven image synthesis aims to create visual content conditioned on diverse textual descriptions. Early endeav-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19098" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.304 + ], + "angle": 0, + "content": "ors focus on leveraging conditional GANs [28, 44, 49, 50] to generate realistic images. Recent advances in diffusion [11, 23, 24, 27, 30, 32, 34] and auto-regressive models [7, 26, 48] have revolutionized image synthesis by enabling stable training and producing high-fidelity images. Typically, DALL-E [26] exemplifies an auto-regressive work that treats textual and visual tokens as a unified data sequence, demonstrating remarkable zero-shot capability. Among diffusion-based efforts, latent diffusion model (LDM) [30] efficiently implements a denoising Markov chain in a low-dimensional latent space, striking a balance between generative quality and computational cost. Nevertheless, these text-prompted methods struggle with fine-grained control over object attributes (i.e., position)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.312, + 0.373, + 0.329 + ], + "angle": 0, + "content": "2.2. Layout-Driven Image Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.334, + 0.486, + 0.622 + ], + "angle": 0, + "content": "Layout-driven image synthesis [2, 15, 22, 46, 53] encodes geometric layouts as positional tokens to enable fine-grained controllability. However, these methods often suffer from attribute leakage, where certain layouts mistakenly attend to irrelevant regions. Multi-instance control [41, 42, 54] addresses this by partitioning layouts into disjoint components, allowing each layout to focus on its corresponding region. Nonetheless, these approaches are confined to natural image domains. AeroGen [36] pioneers layout-conditioned generation in the context of remote sensing imagery, but it still struggles with semantic ambiguity and limited layout controllability. CC-Diff [52] further improves contextual coherence by modeling foreground-background relations, but overlooks frequency-aware cues and remains vulnerable to contextual illusion in various degraded scenes. In contrast, we propose a frequency-inspired contextual disentanglement paradigm, which introduces contextual frequency guidance into the generation process, delivering more realistic degraded images across diverse circumstances." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.633, + 0.185, + 0.649 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.658, + 0.233, + 0.673 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.68, + 0.486, + 0.832 + ], + "angle": 0, + "content": "Latent Diffusion Model. Unlike vanilla DDPM [11], which generates images from Gaussian noise in pixel space, LDM [30] operates in a compressed latent space. Specifically, an input image \\(\\mathcal{I}\\) is first encoded into a latent representation \\(z_{0}\\), which is then diffused into a noisy latent state via the forward process: \\(z_{t} = \\alpha_{t}z_{0} + \\sigma_{t}\\epsilon\\), where \\(\\alpha_{t}\\) and \\(\\sigma_{t}\\) are the fixed schedule, \\(\\epsilon \\sim \\mathcal{N}(0,1)\\) is added standard Gaussian noise term and \\(t \\in [0,T]\\) denotes the timestep. Then, a denoising U-Net \\((\\epsilon_{\\theta})\\) [31] is trained to predict the step-wise noise \\(\\epsilon\\), with the training objective \\(\\mathcal{L}_{LDM}\\) defined in Eq.1." + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.84, + 0.484, + 0.863 + ], + "angle": 0, + "content": "\\[\n\\underset {\\theta} {\\operatorname {m i n}} \\mathcal {L} _ {L D M} = \\mathbb {E} _ {z _ {0}, \\epsilon \\sim \\mathcal {N} (0, 1), t, \\tau} [ | | \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, \\tau) | | _ {2} ^ {2} ], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.478, + 0.885 + ], + "angle": 0, + "content": "where \\(\\tau\\) is the guidance embedding (i.e., text and bboxes)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Problem Formulation. Given instance layouts comprising" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.091, + 0.909, + 0.199 + ], + "angle": 0, + "content": "\\(N\\) object bboxes: \\(\\mathcal{B} = (\\{l_i, b_i\\}_{i=1}^N)\\), where the location of \\(i\\)-th object \\(b_i = [x_i^1, y_i^1, x_i^2, y_i^2]\\) (i.e., top-left and bottom-right corners), and \\(l_i\\) is its corresponding semantic category. \\(\\mathcal{V}\\) denotes the global textual description of the degraded image. In this regard, our FICGen \\((\\mathcal{G}(\\cdot))\\) strives to generate a degraded image conditioned on both spatial layout and semantics: \\(\\mathcal{I} = \\mathcal{G}(\\mathcal{B}, \\epsilon, \\mathcal{V})\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.206, + 0.879, + 0.223 + ], + "angle": 0, + "content": "3.2. Bridge Queries with Frequency Prototypes" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.227, + 0.908, + 0.303 + ], + "angle": 0, + "content": "Existing L2I methods [2, 8, 15, 53, 54] primarily emphasize foreground layout conditions to render instances in natural domains, where fronts and contexts share a relatively uniform frequency distribution. Nonetheless, the naive strategy is inadequate for degraded domains with contextual illusion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.304, + 0.909, + 0.454 + ], + "angle": 0, + "content": "Exploring degraded contexts reveals that frequency disentanglement, a longstanding yet insightful concept, effectively mines weakened HF instances from the dominant LF surroundings. To this end, we seek to extract contextual frequency knowledge and transfer it into the latent space, thus achieving more realistic degraded generation via frequency-aware guidance. To achieve this, we face two main issues: 1) how to represent the frequency knowledge for degraded contexts? and 2) how to transfer frequency knowledge into the latent diffusion space? The answers are detailed below." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.455, + 0.911, + 0.651 + ], + "angle": 0, + "content": "Frequency Prototypes Representation. For the former issue, our goal is to represent contextual frequency knowledge (i.e., textures) using frequency prototypes. To this end, we take two steps to manifest them, as depicted in Fig. 3(a). Taking the foreground exemplar as an example, we first build a class-specific dictionary by collecting same-class instances from the training set. A degraded instance is then sampled and assigned to the \\( i \\)-th bbox based on its position. All selected instances form a foreground exemplar that is often overwhelmed in degraded scenarios. Second, stacked convolutional networks extract its intermediate feature maps \\( \\mathbf{X} \\in \\mathbb{R}^{H \\times W} \\), which are then transformed into the frequency domain \\( \\mathcal{F}(\\mathbf{X}) \\) via a Discrete Fourier Transform (DFT):" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.659, + 0.907, + 0.713 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {\\mathcal {F}} (u, v) = \\frac {1}{H \\times W} \\sum_ {h = 0} ^ {H - 1} \\sum_ {w = 0} ^ {W - 1} \\mathbf {X} (h, w) e ^ {- j 2 \\pi (u h + v w)}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.714, + 0.908, + 0.76 + ], + "angle": 0, + "content": "where \\(\\mathbf{X}_{\\mathcal{F}}\\) is the complex-valued DFT output, \\(|v|\\) and \\(|u|\\) denote the width and height of normalised frequencies. Next, to isolate HF components, a binary mask \\(\\mathbf{M}_{\\mathcal{F}}\\) is applied:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.767, + 0.907, + 0.823 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {\\mathcal {F}} (u, v) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} u \\notin \\left[ \\frac {H}{2} - \\gamma H, \\frac {H}{2} + \\gamma H \\right], \\\\ & v \\notin \\left[ \\frac {W}{2} - \\gamma W, \\frac {W}{2} + \\gamma W \\right], \\\\ 0, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.831, + 0.91, + 0.877 + ], + "angle": 0, + "content": "where \\(\\gamma\\) controls the size of the HF region. The masked frequencies \\(\\mathbf{X}_{\\mathcal{F}}\\mathbf{M}_{\\mathcal{F}}\\) are then enhanced using learnable channelwise weights \\(\\mathbf{W}_{\\mathcal{F}}\\), and projected back to the spatial domain:" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.884, + 0.907, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} ^ {\\uparrow} = \\mathcal {F} ^ {- 1} \\left(\\mathbf {X} _ {\\mathcal {F}} \\mathbf {M} _ {\\mathcal {F}}\\right) \\cdot \\mathbf {W} _ {\\mathcal {F}}, \\tag {4}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19099" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.087, + 0.905, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.343, + 0.908, + 0.374 + ], + "angle": 0, + "content": "Figure 3. The overview pipeline of our FICGen, consisting of (a) Frequency Perceiver Resamplers, (b) Visual-Frequency Enhanced Attention, and (c) Adaptive Spatial-Frequency Aggregation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.398, + 0.484, + 0.534 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}^{-1}\\) denotes the inverse DFT. The result \\(\\mathbf{X}^{\\dagger}\\) represents the enhanced HF feature map. A similar process is applied to the complementary region using \\(1 - \\mathbf{M}_F\\), yielding the LF feature map \\(\\mathbf{X}^{\\downarrow}\\). Inspired by the prototypical learning, we adopt their mean pooled results as HF and LF prototypes, i.e., \\(\\mathbf{p}^{\\uparrow} = \\{p_i^\\uparrow\\}_{i=1}^N \\in \\mathbb{R}^{N \\times d}\\) and \\(p^\\downarrow \\in \\mathbb{R}^d\\), where \\(\\mathbf{p}^{\\uparrow}\\) corresponds to the HF representations of degraded instances, \\(p^\\downarrow\\) reflects the LF characteristics of surrounding contexts, and \\(d\\) denotes the feature dimension." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.535, + 0.484, + 0.61 + ], + "angle": 0, + "content": "Frequency Perceiver Resamplers. To address the latter issue, inspired by the Perceiver architecture [47], we adopt a learnable dual-query mechanism that employs separate frequency resamplers to bridge HF and LF prototypes, thereby capturing contextual frequency cues:" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.62, + 0.483, + 0.641 + ], + "angle": 0, + "content": "\\[\n\\mathbf {q} _ {i} ^ {\\uparrow} = \\operatorname {H F - R e s a m p l e r} \\left(\\mathbf {Q} ^ {\\uparrow}, \\phi_ {k 1} ^ {r} \\left(p _ {i} ^ {\\uparrow}\\right), \\phi_ {v 1} ^ {r} \\left(p _ {i} ^ {\\uparrow}\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.139, + 0.642, + 0.484, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\mathbf {q} ^ {\\downarrow} = \\operatorname {L F - R e s a m p l e r} \\left(\\mathbf {Q} ^ {\\downarrow}, \\phi_ {k 1} ^ {g} \\left(p ^ {\\downarrow}\\right), \\phi_ {v 1} ^ {g} \\left(p ^ {\\downarrow}\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.672, + 0.484, + 0.764 + ], + "angle": 0, + "content": "where \\(\\phi_{*}^{r},\\phi_{*}^{g}\\in \\mathbb{R}^{d\\times d}\\) are linear projection layers. The dual queries \\(\\mathbf{Q}^{\\uparrow}\\) and \\(\\mathbf{Q}^{\\downarrow}\\) are learned through their respective resamplers, each composed of multiple transformer blocks [38], enabling continuous interactions between the dual queries and frequency prototypes, yielding frequency-aware instance and context tokens, i.e., \\(\\mathbf{q}_i^\\uparrow\\) and \\(\\mathbf{q}^{\\downarrow}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.84 + ], + "angle": 0, + "content": "With the above dedicated design, the synergy between frequency prototypes and the dual-query mechanism enables the simultaneous perception of essential frequency cues and contextual appearances, which are then incorporated into the degraded image generation process." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.46, + 0.866 + ], + "angle": 0, + "content": "3.3. Contextual Frequency Knowledge Transfer" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Previous L2I efforts [2, 4, 15, 54] only focus on instance generation conditioned on layouts, while neglecting the sur" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.399, + 0.905, + 0.429 + ], + "angle": 0, + "content": "rounding context. This limitation is particularly detrimental in degraded scenarios involving contextual entanglement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.43, + 0.907, + 0.626 + ], + "angle": 0, + "content": "Visual-Frequency Enhanced Attention. To transfer frequency properties into the degraded generation, we employ a visual-frequency enhanced attention module (Fig.3(c)) that injects frequency-aware tokens \\(\\mathbf{q}_i^\\uparrow\\) and \\(\\mathbf{q}^{\\downarrow}\\) into the latent diffusion space. Specifically, each bbox \\(b_{i}\\) is first encoded using a Fourier embedding, followed by a multi-layer perceptron \\((\\mathbf{E}_{box})\\) to produce positional tokens. In parallel, its semantic class \\(l_{i}\\) is encoded via a CLIP text encoder \\((\\mathbf{E}_{clip})\\) [25]. Notably, \\(\\mathbf{q}_i^\\uparrow\\) is then integrated with them to provide frequency-aware cues for instance generation. As for the surrounding, \\(\\mathbf{q}^{\\downarrow}\\) is fused with text tokens extracted from the global caption \\(\\mathcal{V}\\), thereby facilitating holistic contextual guidance." + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.638, + 0.907, + 0.657 + ], + "angle": 0, + "content": "\\[\n\\mathbf {R} _ {i} = \\left[ \\mathbf {q} _ {i} ^ {\\uparrow}; \\mathbf {E} _ {c l i p} \\left(l _ {i}\\right); \\mathbf {E} _ {b o x} (\\text {F o u r i e r} (b _ {i})) \\right], \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.66, + 0.907, + 0.678 + ], + "angle": 0, + "content": "\\[\n\\mathbf {G} = \\left[ \\mathbf {q} ^ {\\downarrow}; \\mathbf {E} _ {c l i p} (\\mathcal {Y}) \\right], \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.69, + 0.799, + 0.705 + ], + "angle": 0, + "content": "where \\([;]\\) represents concatenate operation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.706, + 0.908, + 0.781 + ], + "angle": 0, + "content": "It is preferable for each layout to affect only its corresponding local region. To this end, an instance coherence map \\(\\hat{\\mathbf{M}}_i\\) is utilized to isolate individual instances and enforce precise interactions between layout conditions and their associated latent features." + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.791, + 0.907, + 0.825 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {i} ^ {r} = \\operatorname {s o f t m a x} \\left(\\frac {Q _ {v} \\phi_ {k 2} ^ {r} (\\mathbf {R} _ {i})}{\\sqrt {d}}\\right) \\cdot \\phi_ {v 2} ^ {r} (\\mathbf {R} _ {i}) \\odot \\hat {\\mathbf {M}} _ {i}, \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.826, + 0.907, + 0.86 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} ^ {g} = \\operatorname {s o f t m a x} \\left(\\frac {Q _ {v} \\phi_ {k 2} ^ {g} (\\mathbf {G})}{\\sqrt {d}}\\right) \\cdot \\phi_ {v 2} ^ {g} (\\mathbf {G}) \\odot \\hat {\\mathbf {M}} ^ {g}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "where \\(Q_{v}\\) denotes the linear projection of latent features, \\(\\odot\\) is element-wise multiplication, and \\(\\phi_{*}^{r},\\phi_{*}^{g}\\) are linear layers." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19100" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.121 + ], + "angle": 0, + "content": "Finally, \\(\\mathbf{f}_i^r\\) indicates the latent feature of the \\(i\\)-th instance, while \\(\\mathbf{f}^g\\) corresponds to the contextual representation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.486, + 0.184 + ], + "angle": 0, + "content": "Instance Coherence Map. To avert feature contamination and attribute leakage, we first construct an instance-level mask \\(\\hat{\\mathbf{M}}_i\\) from each bbox layout, as in [42, 54], which enables precise generation of instance-specific latent features:" + }, + { + "type": "equation", + "bbox": [ + 0.178, + 0.191, + 0.484, + 0.233 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {M}} _ {i} (x, y) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} [ x, y ] \\in b _ {i}, \\\\ 0, & \\text {i f} [ x, y ] \\notin b _ {i}, \\end{array} \\right. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.24, + 0.485, + 0.316 + ], + "angle": 0, + "content": "where \\( x \\) and \\( y \\) denote the mask coordinates. On the basis, we further separate foreground instances from their surroundings by constructing the complementary region \\( \\hat{\\mathbf{M}}^g \\), which encompasses all areas outside the instance-level masks, thus enabling contextual disentanglement:" + }, + { + "type": "equation", + "bbox": [ + 0.219, + 0.324, + 0.484, + 0.365 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {M}} ^ {g} = 1 - \\sum_ {i = 1} ^ {N} \\hat {\\mathbf {M}} _ {i}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.374, + 0.484, + 0.422 + ], + "angle": 0, + "content": "Then, we combine \\(\\mathbf{f}_i^r\\) and \\(\\mathbf{f}^g\\) to produce the unified latent degraded representations \\(\\mathbf{F} = [\\{\\mathbf{f}_i^r\\}_{i=1}^N, \\mathbf{f}^g]\\), which serve as the key for degraded image generation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.429, + 0.442, + 0.446 + ], + "angle": 0, + "content": "3.4. Adaptive Spatial-Frequency Aggregation" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.45, + 0.486, + 0.631 + ], + "angle": 0, + "content": "After harvesting the degraded representations \\(\\mathbf{F}\\), we move forward to jointly consolidate and reconstruct them. In contrast to existing methods that either directly sum multiple instances [52] or combine them solely within the spatial domain [54], which often suffer from \"object omission and merging\" for overlapping instances, as illustrated in Fig. 3(c), we aim to adaptively aggregate the degraded instances and their surroundings in a spatial-frequency mixed manner. \"Spatial\" captures contextual relational dependencies, while \"frequency\" emphasizes fine-grained attributes across instances, i.e., boundary sharpness and textures, which can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.639, + 0.484, + 0.68 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} ^ {s} = \\mathbf {S A M} \\left(\\left[ \\sum_ {i = 1} ^ {N} \\mathbf {f} _ {i} ^ {r}, \\mathbf {f} ^ {g} \\right]\\right), \\mathbf {F} ^ {f} = \\mathbf {F A M} \\left(\\left[ \\sum_ {i = 1} ^ {N} \\mathbf {f} _ {i} ^ {r}, \\mathbf {f} ^ {g} \\right]\\right), \\tag {13}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.689, + 0.484, + 0.731 + ], + "angle": 0, + "content": "\\[\n\\delta^ {f i n a l} = \\sum_ {i = 1} ^ {N + 1} \\operatorname {s o f t m a x} \\left(\\delta^ {f u s i o n}\\right) \\left[ \\sum_ {i = 1} ^ {N} \\mathbf {f} _ {i} ^ {r}, \\mathbf {f} ^ {g} \\right], \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.485, + 0.841 + ], + "angle": 0, + "content": "where SAM and FAM represent the spatial and frequency aggregation modules, respectively, consisting of standard self- and frequency-attention mechanisms. \\(\\delta^{fusion} = \\zeta * ([\\mathbf{F}^s,\\mathbf{F}^f])\\), where \\(\\zeta\\) is a learnable single-layer depthwise convolution to enhance local perception. Finally, the refined degraded representation \\(\\delta^{final} \\in \\mathbb{R}^{h\\times w\\times c}\\) is decoded into the target degraded image." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.283, + 0.866 + ], + "angle": 0, + "content": "3.5. Learning Procedure" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.486, + 0.902 + ], + "angle": 0, + "content": "During training, the parameters of the pretrained LDM remain fixed, with optimization focused solely on FICGen." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.09, + 0.902, + 0.396 + ], + "angle": 0, + "content": "
DetectorMethodFID ↓mAP ↑\\( AP_{50} \\) ↑\\( AP_{75} \\) ↑
DIOR-H [14]
Faster R-CNNOracle-33.455.635.0
Faster R-CNNMIGC [54]31.6421.838.417.5
Faster R-CNNCC-Diff [52]30.8823.642.421.4
Faster R-CNNFICGen (ours)31.2527.648.727.6
RUOD [6]
Faster R-CNNOracle-50.580.254.4
Faster R-CNNMIGC [54]26.5027.254.124.6
Faster R-CNNCC-Diff [52]25.2129.758.427.9
Faster R-CNNFICGen (ours)25.1037.068.636.5
blurred VOC 2012 [5]
Faster R-CNNOracle-31.556.532.4
Faster R-CNNMIGC [54]62.6634.765.833.1
Faster R-CNNCC-Diff [52]62.2036.767.636.3
Faster R-CNNFICGen (ours)58.0240.770.342.7
ExDARK [18]
Cascade R-CNNOracle-37.265.837.8
Cascade R-CNNMIGC [54]45.7632.463.529.5
Cascade R-CNNCC-Diff [52]44.2635.165.634.1
Cascade R-CNNFICGen (ours)42.4042.573.045.1
DAWN [13]
Cascade R-CNNOracle-27.346.426.3
Cascade R-CNNMIGC [54]70.1012.632.38.6
Cascade R-CNNCC-Diff [52]68.5615.733.914.8
Cascade R-CNNFICGen (ours)68.3122.644.321.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.406, + 0.908, + 0.477 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison of fidelity and alignment on various degraded conditions. The performance is evaluated by off-the-shelf detectors Faster R-CNN [29] and Cascade R-CNN [1] on synthetic test set. \"Oracle\" denotes the real test baseline. The best results are in red and the second best results are in underline." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.503, + 0.907, + 0.534 + ], + "angle": 0, + "content": "The learning objective integrates the original mean-square error loss and the contextual frequency guidance:" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.544, + 0.93, + 0.581 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta^ {\\prime}} \\mathcal {L} _ {F I C G e n} = \\mathbb {E} _ {z _ {0}, \\epsilon \\sim \\mathcal {N} (0, 1), t, \\mathcal {Y}} [ \\| \\epsilon - \\mathcal {G} _ {\\theta , \\theta^ {\\prime}} (z _ {t}, t, \\mathcal {Y}, \\mathcal {B}, \\mathcal {Q}) \\| _ {2} ^ {2} ], \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.582, + 0.908, + 0.63 + ], + "angle": 0, + "content": "where \\(\\mathcal{Q} = [\\{\\mathbf{q}_i^\\uparrow\\}_{i=1}^N, \\mathbf{q}^\\downarrow]\\) denotes the set of frequency-aware tokens. While \\(\\theta\\) and \\(\\theta'\\) denote the frozen and trainable parameters of the pretrained LDM and our FICGen, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.643, + 0.64, + 0.66 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.668, + 0.712, + 0.685 + ], + "angle": 0, + "content": "4.1. Experimental Setups" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.908, + 0.75 + ], + "angle": 0, + "content": "Datasets. Our FICGen is evaluated on five degraded-context datasets: ExDARK [18], RUOD [6], DAWN [13], DIOR [14], and blurred VOC 2012 [5]. More details are provided in the Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.886 + ], + "angle": 0, + "content": "Implementation Details. Our FICGen is built upon the pre-trained SDv1.5 checkpoint and deployed only on the mid-level (i.e., \\(8 \\times 8\\)) and lowest-resolution (i.e., \\(16 \\times 16\\)) decoder layers. For training, all images are resized to \\(512 \\times 512\\). We adopt the AdamW [19] optimizer with a learning rate of \\(1\\mathrm{e}^{-4}\\), \\(\\beta_{1} = 0.9\\), and \\(\\beta_{2} = 0.999\\), and fine-tune for 300 epochs on 8 A100 with a batch size of 320. For inference, we use the EulerDiscreteScheduler [12] with 50 sampling steps and a classifier-free guidance scale [10] of 7.5." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.908, + 0.902 + ], + "angle": 0, + "content": "Evaluation Metrics. Following [2, 15], we comprehend" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19101" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.086, + 0.912, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.382, + 0.908, + 0.397 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison of degraded images generated by different L2I methods across various scenarios. Zoom in for better details." + }, + { + "type": "table", + "bbox": [ + 0.103, + 0.42, + 0.48, + 0.574 + ], + "angle": 0, + "content": "
MethodFID ↓YOLO Score ↑
mAPAP50AP75
Oracle-39.257.143.0
LostGAN [35]57.1014.327.315.2
LayoutDiffusion [53]45.3120.037.419.3
ReCo [46]42.5621.140.723.1
MIGC [54]31.6424.541.825.6
GLIGEN [15]41.3125.844.427.8
CC-Diff [52]30.8826.444.228.5
AeroGen [36]38.5729.854.231.6
FICGen (ours)31.2531.249.934.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.584, + 0.483, + 0.613 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison on the DIOR-H test set, with detection performance evaluated using a pre-trained YOLOv8 [37]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.639, + 0.486, + 0.791 + ], + "angle": 0, + "content": "sively evaluate our FICGen from three aspects: fidelity, alignment, and trainability. Fidelity refers to the perceptual quality between synthetic and real images, measured by the Fréchet Inception Distance (FID) [9]. While alignment assesses spatial and semantic correspondences between layout conditions and generated instances, quantified via COCO-style average precision [16] with off-the-shelf detectors. Trainability examines the potential of synthetic images for auxiliary downstream training, assessed by mean Average Precision (mAP) under various training settings." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.8, + 0.231, + 0.815 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.822, + 0.312, + 0.836 + ], + "angle": 0, + "content": "4.2.1. Quantitative Evaluations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Fidelity and Alignment. Tab. 1 reports the quantitative comparison of fidelity and alignment via reporting the COCO-style AP, where pre-trained Faster R-CNN (R50) [29] and Cascade R-CNN (R50) [1] are employed to infer synthetic" + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.42, + 0.902, + 0.502 + ], + "angle": 0, + "content": "
MethodExDARK [18]RUOD [6]
mAP↑AP50↑AP75↑mAP↑AP50↑AP75↑
Oracle38.367.839.858.485.564.8
MIGC [54]28.659.224.121.143.917.9
CC-Diff [52]31.361.828.829.757.828.0
FICGen (ours)38.568.539.537.167.136.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.513, + 0.908, + 0.541 + ], + "angle": 0, + "content": "Table 3. Quantitative comparison on ExDARK and RUOD test set, evaluated by a pre-trained Deformable-DETR [55]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.907, + 0.749 + ], + "angle": 0, + "content": "images, with the results evaluated against corresponding real annotations. Notably, our FICGen successfully outperforms the existing MIGC [54] and CC-Diff [52] in terms of both FID and AP metrics, achieving remarkable FID scores of 25.10 and 42.40 on the RUOD and ExDARK, respectively. Moreover, for remote sensing object detection, FICGen reaches \\(82\\%\\) of the mAP relative to real data (i.e., 27.6 vs. 33.4), substantially outperforming MIGC (21.8) and CC-Diff (23.6) by a large margin. Intriguingly, FICGen even exceeds the real ExDARK test set in AP, highlighting its ability to generate degraded instances that precisely adhere to layouts while bridging the real-to-synthetic domain gap." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.84 + ], + "angle": 0, + "content": "To further verify the generative applicability across different detectors, we evaluate FICGen using YOLOv8 [37] and Deformable-DETR [55], as shown in Tabs. 2 and 3. FICGen surpasses AeroGen by 1.4 and 3.0 in mAP and \\(\\mathrm{AP}_{75}\\) on the DIOR-H, and outperforms CC-Diff by 7.2 and 7.4 on the ExDARK and RUOD, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.901 + ], + "angle": 0, + "content": "**Trainability.** Following the trainability protocol in [2], we use augmented annotations as layout inputs to generate synthetic training images and mix them with real ones, effectively doubling the training set. As shown in the last" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19102" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.904, + 0.381 + ], + "angle": 0, + "content": "
DetectorBackboneTraining SetObject Detection (AP) for Sampled Classes /%mAP\\( AP_{50} \\)\\( AP_{75} \\)
# Real# Syntheticbicyclebusmotorbikeboatcarcupbottledog
Train with Full Real Data
RetinaNetPVT-MR:5.1k-44.858.231.331.136.025.528.747.835.265.433.9
Faster R-CNNX101R:5.1k-49.060.835.630.139.530.934.149.338.669.339.6
Train with Pure Synthetic Data
RetinaNetPVT-M-MIGC (S: 5.1k)32.846.720.419.922.514.116.033.423.549.019.1
Faster R-CNNX101-MIGC (S: 5.1k)33.648.218.520.423.117.021.133.224.749.422.1
RetinaNetPVT-M-CC-Diff (S: 5.1k)34.345.521.620.222.015.416.634.924.149.920.6
Faster R-CNNX101-CC-Diff (S: 5.1k)38.047.521.519.124.318.321.833.825.552.321.7
RetinaNetPVT-M-FICGen (S: 5.1k)35.150.423.220.024.916.216.833.325.351.022.1
Faster R-CNNX101-FICGen (S: 5.1k)38.149.921.722.024.318.920.436.426.751.425.1
Train with Synthetic Data and Finetune on Real Data
RetinaNetPVT-MR:5.1kMIGC (S: 5.1k)47.362.134.732.637.427.629.649.037.167.536.9
Faster R-CNNX101R:5.1kMIGC (S: 5.1k)51.361.838.031.540.032.135.849.739.970.740.9
RetinaNetPVT-MR:5.1kCC-Diff (S: 5.1k)47.960.934.331.536.427.130.548.137.067.035.8
Faster R-CNNX101R:5.1kCC-Diff (S: 5.1k)50.262.238.731.140.632.035.449.439.870.440.7
RetinaNetPVT-MR:5.1kFICGen (S: 5.1k)47.562.235.733.737.827.630.947.637.667.736.8
Faster R-CNNX101R:5.1kFICGen(S: 5.1k)51.462.938.232.840.333.135.649.940.670.942.1
Train with Real & Synthetic Data
Faster R-CNNX101R:5.1kMIGC (S: 5.1k)51.763.036.231.739.731.835.249.639.569.739.8
Faster R-CNNX101R:5.1kCC-Diff (S: 5.1k)50.961.636.531.840.231.634.750.339.570.040.7
Faster R-CNNX101R:5.1kFICGen (S: 5.1k)50.561.838.631.640.432.636.549.140.170.541.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.392, + 0.907, + 0.421 + ], + "angle": 0, + "content": "Table 4. Quantitative comparison of trainability on the ExDARK dataset, evaluated by RetinaNet [17] (PVT-M [39]) and Faster R-CNN [29] (X101 [43]) under three training settings." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.434, + 0.905, + 0.507 + ], + "angle": 0, + "content": "
MethodDIOR-H [14]DAWN [13]ExDARK [18]VOC 2012 [5]
airport ↑mAP ↑\\( AP_{50} \\) ↑\\( AP_{75} \\) ↑bus ↑mAP ↑\\( AP_{50} \\) ↑\\( AP_{75} \\) ↑motorbike ↑mAP ↑\\( AP_{50} \\) ↑\\( AP_{75} \\) ↑boat ↑mAP ↑\\( AP_{50} \\) ↑\\( AP_{75} \\) ↑
Oracle32.233.455.635.021.824.846.023.033.335.867.434.542.948.376.852.5
MIGC [54]36.934.156.136.121.624.943.324.634.437.167.836.642.549.477.254.2
CC-Diff [52]37.634.356.136.326.225.846.125.834.137.067.736.644.449.677.354.3
FICGen (ours)38.135.556.837.627.525.944.526.537.137.768.037.747.350.577.455.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.518, + 0.907, + 0.546 + ], + "angle": 0, + "content": "Table 5. Quantitative comparison of trainability on the DIOR, DAWN and ExDARK and natural VOC 2012 datasets, with detection evaluated using the Faster R-CNN (R50) object detector." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.571, + 0.483, + 0.771 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.782, + 0.483, + 0.811 + ], + "angle": 0, + "content": "Figure 5. Visualization results generated by our FICGen on unseen geometric layouts augmented via flipping and translation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.486, + 0.903 + ], + "angle": 0, + "content": "three rows of Tab. 4, we demonstrate for the first time that synthetic degraded images can significantly enhance the detection performance of downstream object detectors (i.e., +2.4 for bottle and +3.0 for motorbike). Moreover," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.573, + 0.907, + 0.678 + ], + "angle": 0, + "content": "we observe that training with synthetic images followed by fine-tuning on real ones yields the best performance. Tab. 5 further provides a comparative analysis under various degraded conditions. Crucially, FICGen consistently achieves the highest gains, boosting AP by 5.7 for \"bus\" (DAWN) and 3.8 for \"motorbike\" (ExDARK), highlighting its efficacy in mitigating data scarcity in degraded scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.679, + 0.909, + 0.8 + ], + "angle": 0, + "content": "Generalization. To evaluate the generalization of our FICGen, we conduct further experiments not only in severely degraded scenarios such as low-light and underwater conditions with horizontal layout control, but also in blurred and natural VOC 2012 [5], and oriented DIOR-R [3]. Extensive results demonstrate that synthetic images generated by FICGen exhibit superior quality and alignment. Further details are provided in the Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.807, + 0.725, + 0.822 + ], + "angle": 0, + "content": "4.2.2. Qualitative Evaluations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Fig. 4 presents a visual comparison of degraded images generated by various L2I methods. Our FICGen demonstrates superior performance in two key aspects: (a) Enhanced Layout Coherence. It resolves the dilemma of \"object omission and merging\", accurately aligning generated instances with" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19103" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.092, + 0.49, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.214, + 0.486, + 0.243 + ], + "angle": 0, + "content": "Figure 6. Performance convergence curves when generated degraded images are incorporated into downstream detection." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.258, + 0.49, + 0.593 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.603, + 0.483, + 0.631 + ], + "angle": 0, + "content": "Figure 7. Generalization results of FICGen on mildly blurred and natural scenes, as well as under oriented layout control." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.66, + 0.486, + 0.903 + ], + "angle": 0, + "content": "the provided layouts in terms of quantity, scale, and position (i.e., aerial tennis courts and rainy cars), and effectively handling small objects (i.e., aerial vehicles). (b) Strong Occlusion Handling. In heavily occluded scenarios (i.e., underwater corals and fish), FICGen reconstructs dense instances while maintaining spatial consistency, whereas MIGC and CC-Diff fail to enforce overlapped layout constraints. Furthermore, Fig. 5 shows that FICGen generalizes effectively to unseen layouts, including those extending beyond image boundaries. As illustrated in Fig. 6, the synthesized degraded images generated by FICGen enhance downstream detection performance by improving mAP and accelerating training convergence. Finally, Fig. 7 highlights the generalization capability of FICGen in blurred and natural scenes, as well as under oriented layout controls. Additional visualizations are provided in the Supplementary Material." + }, + { + "type": "table", + "bbox": [ + 0.525, + 0.09, + 0.905, + 0.186 + ], + "angle": 0, + "content": "
FPRVFEAASFAmAP\\( AP_{50} \\)\\( AP_{75} \\)
XXX15.734.812.0
X✓X23.244.622.7
✓X✓18.536.313.9
✓✓X25.546.325.1
X✓✓25.646.925.3
✓✓✓27.648.727.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.195, + 0.908, + 0.223 + ], + "angle": 0, + "content": "Table 6. Ablation study of FPR, VFEA, and ASFA modules, evaluated by the pre-trained Faster R-CNN (R50) on the DIOR test set." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.257, + 0.679, + 0.272 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.282, + 0.908, + 0.388 + ], + "angle": 0, + "content": "Frequency Perceiver Resamplers. The foundation of frequency prototype learning lies in the FPR module, which combines a dual-query mechanism with contextual frequency knowledge. As shown in the last two rows of Tab. 6, FPR significantly improves the alignment scores from 25.6/46.9/25.3 to 27.6/48.7/27.6, demonstrating the effectiveness of contextual frequency guidance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.392, + 0.909, + 0.497 + ], + "angle": 0, + "content": "Visual-Frequency Enhanced Attention. VFEA serves as a bridge to inject frequency knowledge (i.e., boundaries and textures) into the latent space while integrating the instance coherence map for contextual disentanglement. As shown in the second row of Tab. 6, even without FPR and ASFA, VFEA achieves competitive performance over the baseline, reaching 23.2/44.6/22.7 in AP." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.501, + 0.909, + 0.622 + ], + "angle": 0, + "content": "Adaptive Spatial-Frequency Aggregation. ASFA is designed to reconstruct refined degraded representations by integrating spatial and frequency perspectives. As evidenced by the results in the fourth and sixth rows of Tab. 6, incorporating ASFA improves the average precision (AP) from 25.5/46.3/25.1 to 27.6/48.7/27.6. This substantial gain underscores its critical role in consolidating contextual information and enhancing the fidelity of degraded image generation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.646, + 0.783, + 0.661 + ], + "angle": 0, + "content": "5. Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduce the task of degraded image generation and define the contextual illusion dilemma as its core challenge. To address this, we propose a frequency-inspired generative paradigm, FICGen. We first extract frequency prototypes from degraded contexts and employ a learnable dual-query mechanism to mine contextual frequency knowledge, which is then transferred into the latent diffusion space via a visual-frequency enhanced attention, coupled with an instance coherence map and a spatial-frequency aggregation module, to disentangle foreground instances from their surroundings and consolidate them into refined degraded representations. Extensive experiments on 5 degraded image datasets validate the effectiveness of FICGen. In future work, we plan to extend this frequency-inspired paradigm to 3D content generation for broader applications." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19104" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.19 + ], + "angle": 0, + "content": "This work is partially supported by grants from the National Natural Science Foundation of China (No. 62132002 and No. 62202010), Beijing Nova Program (No.20250484786), and the Fundamental Research Funds for the Central Universities." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.203, + 0.188, + 0.219 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.228, + 0.484, + 0.284 + ], + "angle": 0, + "content": "[1] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6154-6162, 2018. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.285, + 0.484, + 0.354 + ], + "angle": 0, + "content": "[2] Kai Chen, Enze Xie, Zhe Chen, Yibo Wang, Lanqing Hong, Zhenguo Li, and Dit-Yan Yeung. Geodiffusion: Text-prompted geometric control for object detection data generation. arXiv preprint arXiv:2306.04607, 2023. 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.355, + 0.483, + 0.41 + ], + "angle": 0, + "content": "[3] Gong Cheng, Jiabao Wang, Ke Li, Xingxing Xie, Chunbo Lang, Yanqing Yao, and Junwei Han. Anchor-free oriented proposal generator for object detection. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.411, + 0.484, + 0.466 + ], + "angle": 0, + "content": "[4] Jiaxin Cheng, Xiao Liang, Xingjian Shi, Tong He, Tianjun Xiao, and Mu Li. Layoutdiffuse: Adapting foundational diffusion models for layout-to-image generation. arXiv preprint arXiv:2302.08908, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.468, + 0.484, + 0.522 + ], + "angle": 0, + "content": "[5] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.524, + 0.484, + 0.579 + ], + "angle": 0, + "content": "[6] Chenping Fu, Risheng Liu, Xin Fan, Puyang Chen, Hao Fu, Wanqi Yuan, Ming Zhu, and Zhongxuan Luo. Rethinking general underwater object detection: Datasets, challenges, and solutions. Neurocomputing, 517:243-256, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.58, + 0.484, + 0.647 + ], + "angle": 0, + "content": "[7] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. In European Conference on Computer Vision, pages 89-106. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.649, + 0.484, + 0.704 + ], + "angle": 0, + "content": "[8] Yuchao Gu, Yipin Zhou, Yunfan Ye, Yixin Nie, Licheng Yu, Pingchuan Ma, Kevin Qinghong Lin, and Mike Zheng Shou. Roictrl: Boosting instance control for visual generation. arXiv preprint arXiv:2411.17949, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.705, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[9] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.483, + 0.803 + ], + "angle": 0, + "content": "[10] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.484, + 0.845 + ], + "angle": 0, + "content": "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[12] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in neural information processing systems, 35:26565-26577, 2022. 5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.228, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.134 + ], + "angle": 0, + "content": "[13] Mourad A Kenk and Mahmoud Hassaballah. Dawn: vehicle detection in adverse weather nature dataset. arXiv preprint arXiv:2008.05402, 2020. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.907, + 0.19 + ], + "angle": 0, + "content": "[14] Ke Li, Gang Wan, Gong Cheng, Liqui Meng, and Junwei Han. Object detection in optical remote sensing images: A survey and a new benchmark. ISPRS journal of photogrammetry and remote sensing, 159:296-307, 2020. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.907, + 0.263 + ], + "angle": 0, + "content": "[15] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22511-22521, 2023. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.264, + 0.906, + 0.347 + ], + "angle": 0, + "content": "[16] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer vision-ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part v 13, pages 740-755. Springer, 2014. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.907, + 0.404 + ], + "angle": 0, + "content": "[17] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.907, + 0.46 + ], + "angle": 0, + "content": "[18] Yuen Peng Loh and Chee Seng Chan. Getting to know low-light images with the exclusively dark dataset. Computer Vision and Image Understanding, 178:30-42, 2019. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.906, + 0.49 + ], + "angle": 0, + "content": "[19] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.492, + 0.907, + 0.561 + ], + "angle": 0, + "content": "[20] Zhengyao Lv, Yuxiang Wei, Wangmeng Zuo, and KwanYee K Wong. Place: Adaptive layout-semantic fusion for semantic image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9264–9274, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.563, + 0.907, + 0.617 + ], + "angle": 0, + "content": "[21] Mingcan Ma, Changqun Xia, Chenxi Xie, Xiaowu Chen, and Jia Li. Boosting broader receptive fields for salient object detection. IEEE Transactions on Image Processing, 32:1026-1038, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.62, + 0.906, + 0.689 + ], + "angle": 0, + "content": "[22] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI conference on artificial intelligence, pages 4296–4304, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.691, + 0.906, + 0.758 + ], + "angle": 0, + "content": "[23] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.907, + 0.816 + ], + "angle": 0, + "content": "[24] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.907, + 0.9 + ], + "angle": 0, + "content": "[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19105" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.16 + ], + "angle": 0, + "content": "[26] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International conference on machine learning, pages 8821-8831. Pmlr, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.485, + 0.218 + ], + "angle": 0, + "content": "[27] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2): 3, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.484, + 0.276 + ], + "angle": 0, + "content": "[28] Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In International conference on machine learning, pages 1060-1069. PMLR, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.484, + 0.332 + ], + "angle": 0, + "content": "[29] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 1, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.484, + 0.403 + ], + "angle": 0, + "content": "[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.484, + 0.488 + ], + "angle": 0, + "content": "[31] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.484, + 0.572 + ], + "angle": 0, + "content": "[32] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.574, + 0.484, + 0.642 + ], + "angle": 0, + "content": "[33] Zhe Shan, Yang Liu, Lei Zhou, Cheng Yan, Heng Wang, and Xia Xie. Ros-sam: High-quality interactive segmentation for remote sensing moving object. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 3625-3635, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.645, + 0.484, + 0.684 + ], + "angle": 0, + "content": "[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.688, + 0.484, + 0.741 + ], + "angle": 0, + "content": "[35] Wei Sun and Tianfu Wu. Image synthesis from reconfigurable layout and style. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10531-10540, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.745, + 0.484, + 0.813 + ], + "angle": 0, + "content": "[36] Datalog Tang, Xiangyong Cao, Xuan Wu, Jialin Li, Jing Yao, Xueru Bai, Dongsheng Jiang, Yin Li, and Deyu Meng. Aerogen: enhancing remote sensing object detection with diffusion-driven data generation. arXiv preprint arXiv:2411.15497, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.816, + 0.484, + 0.842 + ], + "angle": 0, + "content": "[37] Ultralytics. Ultralytics yolov8. https://github.com/ultralytics/ultralytics, 2023.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[38] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[39] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyra-" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.908, + 0.147 + ], + "angle": 0, + "content": "mid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 568-578, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.907, + 0.219 + ], + "angle": 0, + "content": "[40] Wenzhuang Wang, Mingcan Ma, Yong Chen, Changqun Xia, Zhenbao Liang, and Jia Li. Freegen: Bridging visual-linguistic discrepancies towards diffusion-based pixel-level data synthesis. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7916–7924, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.907, + 0.29 + ], + "angle": 0, + "content": "[41] Xudong Wang, Trevor Darrell, Sai Saketh Rambhatla, Rohit Girdhar, and Ishan Misra. Instancediffusion: Instance-level control for image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6232-6242, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.907, + 0.346 + ], + "angle": 0, + "content": "[42] Yinwei Wu, Xianpan Zhou, Bing Ma, Xuefeng Su, Kai Ma, and Xinchao Wang. Ifadapter: Instance feature control for grounded text-to-image generation. arXiv preprint arXiv:2409.08240, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.907, + 0.417 + ], + "angle": 0, + "content": "[43] Saining Xie, Ross Girshick, Piotr Dólár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1492–1500, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.908, + 0.501 + ], + "angle": 0, + "content": "[44] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Finegrained text to image generation with attentional generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1316-1324, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.907, + 0.56 + ], + "angle": 0, + "content": "[45] Han Xue, Zhiwu Huang, Qianru Sun, Li Song, and Wenjun Zhang. Freestyle layout-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14256-14266, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.908, + 0.643 + ], + "angle": 0, + "content": "[46] Zhengyuan Yang, Jianfeng Wang, Zhe Gan, Linjie Li, Kevin Lin, Chenfei Wu, Nan Duan, Zicheng Liu, Ce Liu, Michael Zeng, et al. Reco: Region-controlled text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14246-14255, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.908, + 0.701 + ], + "angle": 0, + "content": "[47] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.908, + 0.772 + ], + "angle": 0, + "content": "[48] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.908, + 0.857 + ], + "angle": 0, + "content": "[49] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 5907-5915, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.908, + 0.901 + ], + "angle": 0, + "content": "[50] Han Zhang, Jing Yu Koh, Jason Baldridge, Honglak Lee, and Yinfei Yang. Cross-modal contrastive learning for text-to-image generation. In Proceedings of the IEEE/CVF con" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19106" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.123, + 0.093, + 0.482, + 0.119 + ], + "angle": 0, + "content": "ference on computer vision and pattern recognition, pages 833-842, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.177 + ], + "angle": 0, + "content": "[51] Jingyang Zhang, Shiwei Li, Yuanxun Lu, Tian Fang, David McKinnon, Yanghai Tsin, Long Quan, and Yao Yao. Joint-net: Extending text-to-image diffusion for dense distribution modeling. arXiv preprint arXiv:2310.06347, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.484, + 0.233 + ], + "angle": 0, + "content": "[52] Mu Zhang, Yunfan Liu, Yue Liu, Hongtian Yu, and Qixiang Ye. Cc-diff: Enhancing contextual coherence in remote sensing image synthesis. arXiv preprint arXiv:2412.08464, 2024. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.484, + 0.303 + ], + "angle": 0, + "content": "[53] Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22490-22499, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.484, + 0.373 + ], + "angle": 0, + "content": "[54] Dewei Zhou, You Li, Fan Ma, Xiaoting Zhang, and Yi Yang. Mige: Multi-instance generation controller for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6818-6828, 2024. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.484, + 0.431 + ], + "angle": 0, + "content": "[55] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "19107" + } + ] +] \ No newline at end of file diff --git a/2025/FICGen_ Frequency-Inspired Contextual Disentanglement for Layout-driven Degraded Image Generation/24509018-dca5-4420-ad1a-dd11a5974d9d_origin.pdf b/2025/FICGen_ Frequency-Inspired Contextual Disentanglement for Layout-driven Degraded Image Generation/24509018-dca5-4420-ad1a-dd11a5974d9d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..892dbefe4ea24e30391fed4cd632678d1c8b5f27 --- /dev/null +++ b/2025/FICGen_ Frequency-Inspired Contextual Disentanglement for Layout-driven Degraded Image Generation/24509018-dca5-4420-ad1a-dd11a5974d9d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:17463a68a087567c147f2ab3d3b08eb4077db181becfc57a6dd8197e32443a68 +size 21824563 diff --git a/2025/FICGen_ Frequency-Inspired Contextual Disentanglement for Layout-driven Degraded Image Generation/full.md b/2025/FICGen_ Frequency-Inspired Contextual Disentanglement for Layout-driven Degraded Image Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..da980c41a0433a1c2e34e86384f0dfb63f37466f --- /dev/null +++ b/2025/FICGen_ Frequency-Inspired Contextual Disentanglement for Layout-driven Degraded Image Generation/full.md @@ -0,0 +1,340 @@ +# FICGen: Frequency-Inspired Contextual Disentanglement for Layout-driven Degraded Image Generation + +Wenzhuang Wang $^{1,2}$ , Yifan Zhao $^{1*}$ , Mingcan Ma $^{2}$ , Ming Liu $^{2}$ , Zhonglin Jiang $^{2}$ , Yong Chen $^{2}$ , Jia Li $^{1*}$ $^{1}$ State Key Laboratory of Virtual Reality Technology and Systems, SCSE&QRI, Beihang University + $^{2}$ Geely Automobile Research Institute (Ningbo) Co., Ltd + +{wz_wang, zhaoyf, jiali}@buaa.edu.cn + +![](images/e52df08b170e168c189b4efce8c392106b33dcbb6739c1715062db63cf4229e3.jpg) +(a) Contextual Illusion Dilemma + +![](images/bae07c4cca4983817de4840ae34b4044aae24659988177a16250b6f238aa11ec.jpg) + +![](images/3f5b4734eab6bf4189c78fd9299c564b5beb98d241710036527c98722faee8b8.jpg) + +![](images/7f4d482c918dea9fcd464ff72fabf735832fb6b5680f8bbd1eb14eb100419a48.jpg) +(b) Robust Generation for Out-of-Distribution Layouts +(c) Alignment & Trainability + +![](images/c40921c12eebe00483cafdbeb616afcdc8a55eef5f6faa8157f34168d6938478.jpg) +Figure 1. Visualizations of degraded image generation by our FICGen. FICGen achieves superior quality compared to existing L2I methods and demonstrates strong robustness to OOD layouts. Moreover, it shows excellent trainability for boosting object detectors, i.e., Faster R-CNN (R50) [29]. The alignment and trainability are evaluated on the ExDARK [18] and DIOR-H [14], respectively. + +# Abstract + +Layout-to-image (L2I) generation has exhibited promising results in natural domains, but suffers from limited generative fidelity and weak alignment with user-provided layouts when applied to degraded scenes (i.e., low-light, underwater). We primarily attribute these limitations to the "contextual illusion dilemma" in degraded conditions, where foreground instances are overwhelmed by context-dominant frequency distributions. Motivated by this, our paper proposes a new Frequency-Inspired Contextual Disentangle-ment Generative (FICGen) paradigm, which seeks to transfer frequency knowledge of degraded images into the latent diffusion space, thereby facilitating the rendering of + +degraded instances and their surroundings via contextual frequency-aware guidance. To be specific, FICGen consists of two major steps. Firstly, we introduce a learnable dual-query mechanism, each paired with a dedicated frequency resampler, to extract contextual frequency prototypes from pre-collected degraded exemplars in the training set. Secondly, a visual-frequency enhanced attention is employed to inject frequency prototypes into the degraded generation process. To alleviate the contextual illusion and attribute leakage, an instance coherence map is developed to regulate latent-space disentanglement between individual instances and their surroundings, coupled with an adaptive spatial-frequency aggregation module to reconstruct spatial-frequency mixed degraded representations. Extensive experiments on 5 benchmarks involving a variety of degraded scenarios—from severe low-light to mild blur—demonstrate + +that FICGen consistently surpasses existing L2I methods in terms of generative fidelity, alignment and downstream auxiliary trainability. + +# 1. Introduction + +Visual perception [21, 33] depends on a large amount of annotated datasets, particularly in degraded conditions such as low-light [18], and adverse weather [13]. Unfortunately, acquiring adequate datasets for these extreme conditions remains a major obstacle due to the high cost of manual collection and annotation. For instance, the number of images in ExDARK (7,363) [18] is only about 1/20th that of COCO 2017 (118,287) [16], highlighting the severe data scarcity. In light of this, alternative strategies are emerging that shift the focus from a "model-centric" to a "data-centric" perspective, where advanced generative models are used to synthesize realistic images for bolstering downstream tasks [40]. + +Witnessing the phenomenal capabilities of text-to-image (T2I) diffusion models [27, 32], numerous advances have been explored in conditioning image generation on visual labels, thereby enabling fine-grained controllability beyond what textual prompts alone can achieve. Such labels are often represented by masks [20, 45], depths [51], and bounding boxes (bboxes) [8, 15, 46, 53, 54], thereby facilitating visual generation via layout-to-image (L2I) synthesis. Among them, bboxes are the most intuitive for guiding the location of visual entities during L2I generation. + +Despite the success of L2I methods [15, 53, 54] in natural scenarios, they face challenges when adapting to practical degraded conditions. In Fig. 1 (a), aerial objects (i.e., vehicles) are often small in size and visually resemble surrounding structures (i.e., bridges), while underwater species (i.e., fish) frequently blend with other nearby creatures (i.e., corals), resulting in contextual entanglement. Such cases, difficult even for the human eye to distinguish and predominantly occurring in degraded scenarios, lead to generative illusions regarding object quantity, location, and interaction—termed the “contextual illusion dilemma” in our work. + +Generally, as shown in Fig. 2, natural image exhibits a relatively balanced distribution of high-frequency (HF) and low-frequency (LF) components, making the distinction between instances and contexts pronounced. In contrast, HF details of instances in degraded contexts are attenuated, while LF contexts dominate the overall frequency distribution. Motivated by the frequency properties, we propose a new frequency-inspired contextual disentanglement paradigm to tackle the dilemma. Specifically, we first extract HF and LF prototypes from degraded contexts and associate them with each object and its surroundings. Second, a learnable dual-query mechanism equipped with frequency perceiver resamplers is introduced to simultaneously capture contextual frequency knowledge (i.e., boundaries and colours). Third, we inject + +![](images/36cac4420640b7fe574c6332faec8ca5a94fd87d8eadf4e7f65fb4aafa82e5c9.jpg) +Figure 2. Frequency spectrum comparison between natural and degraded images. + +the knowledge into the latent diffusion space via a visual-frequency enhanced attention, enabling better perception of frequency attributes during degraded generation. Meanwhile, an instance coherence map is designed to disentangle degraded instances from their surroundings, mitigating contextual entanglement and attribute leakage. Finally, a spatial-frequency aggregation module is developed to adaptively recombine them to obtain refined degraded representations. With the above efforts, FICGen offers a unified approach for degraded image synthesis from a frequency-inspired perspective. We evaluate FICGen on 5 degraded benchmarks, covering low-light (ExDARK [18]), underwater (RUOD [6]), remote sensing (DIOR [14]), adverse weather (DAWN [13]), and blur (VOC 2012 [5]) conditions. Extensive experiments demonstrate that FICGen achieves superior image quality while maintaining strong alignment with geometric layouts. In Fig.1(b), FICGen can perform robust OOD generation on unseen layouts. Fig.1(c) shows that FICGen achieves better alignment than the real test set across all classes except "boat", effectively narrowing the real-synthetic domain gap. Furthermore, FICGen consistently improves downstream tasks by $\sim 2.0$ mAP and substantially boosts detection accuracy for certain classes, i.e., $\sim 6.0$ (32.2 vs. 38.1) on "airport", underscoring its strong trainability. + +In summary, our contributions are threefold: (1) We propose FICGen, the first attempt to address the contextual illusion dilemma in degraded image generation. (2) We introduce a learnable dual-query mechanism to capture contextual frequency knowledge, which is injected into the latent diffusion space via a visual-frequency enhanced attention module, and further complemented by an instance coherence map and a spatial-frequency aggregation module to enable latent-space disentanglement and consolidation. (3) Extensive experiments on 5 degraded datasets show that our FICGen outperforms SOTA L2I methods in both quantitative and qualitative evaluations. Notably, FICGen can be seamlessly integrated as a plug-and-play component to enhance other diffusion-based generative models. + +# 2. Related Work + +# 2.1. Text-Driven Image Synthesis + +Text-driven image synthesis aims to create visual content conditioned on diverse textual descriptions. Early endeav- + +ors focus on leveraging conditional GANs [28, 44, 49, 50] to generate realistic images. Recent advances in diffusion [11, 23, 24, 27, 30, 32, 34] and auto-regressive models [7, 26, 48] have revolutionized image synthesis by enabling stable training and producing high-fidelity images. Typically, DALL-E [26] exemplifies an auto-regressive work that treats textual and visual tokens as a unified data sequence, demonstrating remarkable zero-shot capability. Among diffusion-based efforts, latent diffusion model (LDM) [30] efficiently implements a denoising Markov chain in a low-dimensional latent space, striking a balance between generative quality and computational cost. Nevertheless, these text-prompted methods struggle with fine-grained control over object attributes (i.e., position). + +# 2.2. Layout-Driven Image Synthesis + +Layout-driven image synthesis [2, 15, 22, 46, 53] encodes geometric layouts as positional tokens to enable fine-grained controllability. However, these methods often suffer from attribute leakage, where certain layouts mistakenly attend to irrelevant regions. Multi-instance control [41, 42, 54] addresses this by partitioning layouts into disjoint components, allowing each layout to focus on its corresponding region. Nonetheless, these approaches are confined to natural image domains. AeroGen [36] pioneers layout-conditioned generation in the context of remote sensing imagery, but it still struggles with semantic ambiguity and limited layout controllability. CC-Diff [52] further improves contextual coherence by modeling foreground-background relations, but overlooks frequency-aware cues and remains vulnerable to contextual illusion in various degraded scenes. In contrast, we propose a frequency-inspired contextual disentanglement paradigm, which introduces contextual frequency guidance into the generation process, delivering more realistic degraded images across diverse circumstances. + +# 3. Method + +# 3.1. Preliminaries + +Latent Diffusion Model. Unlike vanilla DDPM [11], which generates images from Gaussian noise in pixel space, LDM [30] operates in a compressed latent space. Specifically, an input image $\mathcal{I}$ is first encoded into a latent representation $z_{0}$ , which is then diffused into a noisy latent state via the forward process: $z_{t} = \alpha_{t}z_{0} + \sigma_{t}\epsilon$ , where $\alpha_{t}$ and $\sigma_{t}$ are the fixed schedule, $\epsilon \sim \mathcal{N}(0,1)$ is added standard Gaussian noise term and $t \in [0,T]$ denotes the timestep. Then, a denoising U-Net $(\epsilon_{\theta})$ [31] is trained to predict the step-wise noise $\epsilon$ , with the training objective $\mathcal{L}_{LDM}$ defined in Eq.1. + +$$ +\underset {\theta} {\operatorname {m i n}} \mathcal {L} _ {L D M} = \mathbb {E} _ {z _ {0}, \epsilon \sim \mathcal {N} (0, 1), t, \tau} [ | | \epsilon - \epsilon_ {\theta} (z _ {t}, t, \tau) | | _ {2} ^ {2} ], \tag {1} +$$ + +where $\tau$ is the guidance embedding (i.e., text and bboxes). + +Problem Formulation. Given instance layouts comprising + +$N$ object bboxes: $\mathcal{B} = (\{l_i, b_i\}_{i=1}^N)$ , where the location of $i$ -th object $b_i = [x_i^1, y_i^1, x_i^2, y_i^2]$ (i.e., top-left and bottom-right corners), and $l_i$ is its corresponding semantic category. $\mathcal{V}$ denotes the global textual description of the degraded image. In this regard, our FICGen $(\mathcal{G}(\cdot))$ strives to generate a degraded image conditioned on both spatial layout and semantics: $\mathcal{I} = \mathcal{G}(\mathcal{B}, \epsilon, \mathcal{V})$ . + +# 3.2. Bridge Queries with Frequency Prototypes + +Existing L2I methods [2, 8, 15, 53, 54] primarily emphasize foreground layout conditions to render instances in natural domains, where fronts and contexts share a relatively uniform frequency distribution. Nonetheless, the naive strategy is inadequate for degraded domains with contextual illusion. + +Exploring degraded contexts reveals that frequency disentanglement, a longstanding yet insightful concept, effectively mines weakened HF instances from the dominant LF surroundings. To this end, we seek to extract contextual frequency knowledge and transfer it into the latent space, thus achieving more realistic degraded generation via frequency-aware guidance. To achieve this, we face two main issues: 1) how to represent the frequency knowledge for degraded contexts? and 2) how to transfer frequency knowledge into the latent diffusion space? The answers are detailed below. + +Frequency Prototypes Representation. For the former issue, our goal is to represent contextual frequency knowledge (i.e., textures) using frequency prototypes. To this end, we take two steps to manifest them, as depicted in Fig. 3(a). Taking the foreground exemplar as an example, we first build a class-specific dictionary by collecting same-class instances from the training set. A degraded instance is then sampled and assigned to the $i$ -th bbox based on its position. All selected instances form a foreground exemplar that is often overwhelmed in degraded scenarios. Second, stacked convolutional networks extract its intermediate feature maps $\mathbf{X} \in \mathbb{R}^{H \times W}$ , which are then transformed into the frequency domain $\mathcal{F}(\mathbf{X})$ via a Discrete Fourier Transform (DFT): + +$$ +\mathbf {X} _ {\mathcal {F}} (u, v) = \frac {1}{H \times W} \sum_ {h = 0} ^ {H - 1} \sum_ {w = 0} ^ {W - 1} \mathbf {X} (h, w) e ^ {- j 2 \pi (u h + v w)}, \tag {2} +$$ + +where $\mathbf{X}_{\mathcal{F}}$ is the complex-valued DFT output, $|v|$ and $|u|$ denote the width and height of normalised frequencies. Next, to isolate HF components, a binary mask $\mathbf{M}_{\mathcal{F}}$ is applied: + +$$ +\mathbf {M} _ {\mathcal {F}} (u, v) = \left\{ \begin{array}{l l} 1, & \text {i f} u \notin \left[ \frac {H}{2} - \gamma H, \frac {H}{2} + \gamma H \right], \\ & v \notin \left[ \frac {W}{2} - \gamma W, \frac {W}{2} + \gamma W \right], \\ 0, & \text {o t h e r w i s e} \end{array} \right. \tag {3} +$$ + +where $\gamma$ controls the size of the HF region. The masked frequencies $\mathbf{X}_{\mathcal{F}}\mathbf{M}_{\mathcal{F}}$ are then enhanced using learnable channelwise weights $\mathbf{W}_{\mathcal{F}}$ , and projected back to the spatial domain: + +$$ +\mathbf {X} ^ {\uparrow} = \mathcal {F} ^ {- 1} \left(\mathbf {X} _ {\mathcal {F}} \mathbf {M} _ {\mathcal {F}}\right) \cdot \mathbf {W} _ {\mathcal {F}}, \tag {4} +$$ + +![](images/fa0b8feb95be71860094d4ca4929b0b6f7d5b173ea8038d81185383cfb8a2272.jpg) +Figure 3. The overview pipeline of our FICGen, consisting of (a) Frequency Perceiver Resamplers, (b) Visual-Frequency Enhanced Attention, and (c) Adaptive Spatial-Frequency Aggregation. + +where $\mathcal{F}^{-1}$ denotes the inverse DFT. The result $\mathbf{X}^{\dagger}$ represents the enhanced HF feature map. A similar process is applied to the complementary region using $1 - \mathbf{M}_F$ , yielding the LF feature map $\mathbf{X}^{\downarrow}$ . Inspired by the prototypical learning, we adopt their mean pooled results as HF and LF prototypes, i.e., $\mathbf{p}^{\uparrow} = \{p_i^\uparrow\}_{i=1}^N \in \mathbb{R}^{N \times d}$ and $p^\downarrow \in \mathbb{R}^d$ , where $\mathbf{p}^{\uparrow}$ corresponds to the HF representations of degraded instances, $p^\downarrow$ reflects the LF characteristics of surrounding contexts, and $d$ denotes the feature dimension. + +Frequency Perceiver Resamplers. To address the latter issue, inspired by the Perceiver architecture [47], we adopt a learnable dual-query mechanism that employs separate frequency resamplers to bridge HF and LF prototypes, thereby capturing contextual frequency cues: + +$$ +\mathbf {q} _ {i} ^ {\uparrow} = \operatorname {H F - R e s a m p l e r} \left(\mathbf {Q} ^ {\uparrow}, \phi_ {k 1} ^ {r} \left(p _ {i} ^ {\uparrow}\right), \phi_ {v 1} ^ {r} \left(p _ {i} ^ {\uparrow}\right)\right), \tag {5} +$$ + +$$ +\mathbf {q} ^ {\downarrow} = \operatorname {L F - R e s a m p l e r} \left(\mathbf {Q} ^ {\downarrow}, \phi_ {k 1} ^ {g} \left(p ^ {\downarrow}\right), \phi_ {v 1} ^ {g} \left(p ^ {\downarrow}\right)\right), \tag {6} +$$ + +where $\phi_{*}^{r},\phi_{*}^{g}\in \mathbb{R}^{d\times d}$ are linear projection layers. The dual queries $\mathbf{Q}^{\uparrow}$ and $\mathbf{Q}^{\downarrow}$ are learned through their respective resamplers, each composed of multiple transformer blocks [38], enabling continuous interactions between the dual queries and frequency prototypes, yielding frequency-aware instance and context tokens, i.e., $\mathbf{q}_i^\uparrow$ and $\mathbf{q}^{\downarrow}$ . + +With the above dedicated design, the synergy between frequency prototypes and the dual-query mechanism enables the simultaneous perception of essential frequency cues and contextual appearances, which are then incorporated into the degraded image generation process. + +# 3.3. Contextual Frequency Knowledge Transfer + +Previous L2I efforts [2, 4, 15, 54] only focus on instance generation conditioned on layouts, while neglecting the sur + +rounding context. This limitation is particularly detrimental in degraded scenarios involving contextual entanglement. + +Visual-Frequency Enhanced Attention. To transfer frequency properties into the degraded generation, we employ a visual-frequency enhanced attention module (Fig.3(c)) that injects frequency-aware tokens $\mathbf{q}_i^\uparrow$ and $\mathbf{q}^{\downarrow}$ into the latent diffusion space. Specifically, each bbox $b_{i}$ is first encoded using a Fourier embedding, followed by a multi-layer perceptron $(\mathbf{E}_{box})$ to produce positional tokens. In parallel, its semantic class $l_{i}$ is encoded via a CLIP text encoder $(\mathbf{E}_{clip})$ [25]. Notably, $\mathbf{q}_i^\uparrow$ is then integrated with them to provide frequency-aware cues for instance generation. As for the surrounding, $\mathbf{q}^{\downarrow}$ is fused with text tokens extracted from the global caption $\mathcal{V}$ , thereby facilitating holistic contextual guidance. + +$$ +\mathbf {R} _ {i} = \left[ \mathbf {q} _ {i} ^ {\uparrow}; \mathbf {E} _ {c l i p} \left(l _ {i}\right); \mathbf {E} _ {b o x} (\text {F o u r i e r} (b _ {i})) \right], \tag {7} +$$ + +$$ +\mathbf {G} = \left[ \mathbf {q} ^ {\downarrow}; \mathbf {E} _ {c l i p} (\mathcal {Y}) \right], \tag {8} +$$ + +where $[;]$ represents concatenate operation. + +It is preferable for each layout to affect only its corresponding local region. To this end, an instance coherence map $\hat{\mathbf{M}}_i$ is utilized to isolate individual instances and enforce precise interactions between layout conditions and their associated latent features. + +$$ +\mathbf {f} _ {i} ^ {r} = \operatorname {s o f t m a x} \left(\frac {Q _ {v} \phi_ {k 2} ^ {r} (\mathbf {R} _ {i})}{\sqrt {d}}\right) \cdot \phi_ {v 2} ^ {r} (\mathbf {R} _ {i}) \odot \hat {\mathbf {M}} _ {i}, \tag {9} +$$ + +$$ +\mathbf {f} ^ {g} = \operatorname {s o f t m a x} \left(\frac {Q _ {v} \phi_ {k 2} ^ {g} (\mathbf {G})}{\sqrt {d}}\right) \cdot \phi_ {v 2} ^ {g} (\mathbf {G}) \odot \hat {\mathbf {M}} ^ {g}, \tag {10} +$$ + +where $Q_{v}$ denotes the linear projection of latent features, $\odot$ is element-wise multiplication, and $\phi_{*}^{r},\phi_{*}^{g}$ are linear layers. + +Finally, $\mathbf{f}_i^r$ indicates the latent feature of the $i$ -th instance, while $\mathbf{f}^g$ corresponds to the contextual representation. + +Instance Coherence Map. To avert feature contamination and attribute leakage, we first construct an instance-level mask $\hat{\mathbf{M}}_i$ from each bbox layout, as in [42, 54], which enables precise generation of instance-specific latent features: + +$$ +\hat {\mathbf {M}} _ {i} (x, y) = \left\{ \begin{array}{l l} 1, & \text {i f} [ x, y ] \in b _ {i}, \\ 0, & \text {i f} [ x, y ] \notin b _ {i}, \end{array} \right. \tag {11} +$$ + +where $x$ and $y$ denote the mask coordinates. On the basis, we further separate foreground instances from their surroundings by constructing the complementary region $\hat{\mathbf{M}}^g$ , which encompasses all areas outside the instance-level masks, thus enabling contextual disentanglement: + +$$ +\hat {\mathbf {M}} ^ {g} = 1 - \sum_ {i = 1} ^ {N} \hat {\mathbf {M}} _ {i}. \tag {12} +$$ + +Then, we combine $\mathbf{f}_i^r$ and $\mathbf{f}^g$ to produce the unified latent degraded representations $\mathbf{F} = [\{\mathbf{f}_i^r\}_{i=1}^N, \mathbf{f}^g]$ , which serve as the key for degraded image generation. + +# 3.4. Adaptive Spatial-Frequency Aggregation + +After harvesting the degraded representations $\mathbf{F}$ , we move forward to jointly consolidate and reconstruct them. In contrast to existing methods that either directly sum multiple instances [52] or combine them solely within the spatial domain [54], which often suffer from "object omission and merging" for overlapping instances, as illustrated in Fig. 3(c), we aim to adaptively aggregate the degraded instances and their surroundings in a spatial-frequency mixed manner. "Spatial" captures contextual relational dependencies, while "frequency" emphasizes fine-grained attributes across instances, i.e., boundary sharpness and textures, which can be formulated as follows: + +$$ +\mathbf {F} ^ {s} = \mathbf {S A M} \left(\left[ \sum_ {i = 1} ^ {N} \mathbf {f} _ {i} ^ {r}, \mathbf {f} ^ {g} \right]\right), \mathbf {F} ^ {f} = \mathbf {F A M} \left(\left[ \sum_ {i = 1} ^ {N} \mathbf {f} _ {i} ^ {r}, \mathbf {f} ^ {g} \right]\right), \tag {13} +$$ + +$$ +\delta^ {f i n a l} = \sum_ {i = 1} ^ {N + 1} \operatorname {s o f t m a x} \left(\delta^ {f u s i o n}\right) \left[ \sum_ {i = 1} ^ {N} \mathbf {f} _ {i} ^ {r}, \mathbf {f} ^ {g} \right], \tag {14} +$$ + +where SAM and FAM represent the spatial and frequency aggregation modules, respectively, consisting of standard self- and frequency-attention mechanisms. $\delta^{fusion} = \zeta * ([\mathbf{F}^s,\mathbf{F}^f])$ , where $\zeta$ is a learnable single-layer depthwise convolution to enhance local perception. Finally, the refined degraded representation $\delta^{final} \in \mathbb{R}^{h\times w\times c}$ is decoded into the target degraded image. + +# 3.5. Learning Procedure + +During training, the parameters of the pretrained LDM remain fixed, with optimization focused solely on FICGen. + +
DetectorMethodFID ↓mAP ↑\( AP_{50} \) ↑\( AP_{75} \) ↑
DIOR-H [14]
Faster R-CNNOracle-33.455.635.0
Faster R-CNNMIGC [54]31.6421.838.417.5
Faster R-CNNCC-Diff [52]30.8823.642.421.4
Faster R-CNNFICGen (ours)31.2527.648.727.6
RUOD [6]
Faster R-CNNOracle-50.580.254.4
Faster R-CNNMIGC [54]26.5027.254.124.6
Faster R-CNNCC-Diff [52]25.2129.758.427.9
Faster R-CNNFICGen (ours)25.1037.068.636.5
blurred VOC 2012 [5]
Faster R-CNNOracle-31.556.532.4
Faster R-CNNMIGC [54]62.6634.765.833.1
Faster R-CNNCC-Diff [52]62.2036.767.636.3
Faster R-CNNFICGen (ours)58.0240.770.342.7
ExDARK [18]
Cascade R-CNNOracle-37.265.837.8
Cascade R-CNNMIGC [54]45.7632.463.529.5
Cascade R-CNNCC-Diff [52]44.2635.165.634.1
Cascade R-CNNFICGen (ours)42.4042.573.045.1
DAWN [13]
Cascade R-CNNOracle-27.346.426.3
Cascade R-CNNMIGC [54]70.1012.632.38.6
Cascade R-CNNCC-Diff [52]68.5615.733.914.8
Cascade R-CNNFICGen (ours)68.3122.644.321.5
+ +Table 1. Quantitative comparison of fidelity and alignment on various degraded conditions. The performance is evaluated by off-the-shelf detectors Faster R-CNN [29] and Cascade R-CNN [1] on synthetic test set. "Oracle" denotes the real test baseline. The best results are in red and the second best results are in underline. + +The learning objective integrates the original mean-square error loss and the contextual frequency guidance: + +$$ +\min _ {\theta^ {\prime}} \mathcal {L} _ {F I C G e n} = \mathbb {E} _ {z _ {0}, \epsilon \sim \mathcal {N} (0, 1), t, \mathcal {Y}} [ \| \epsilon - \mathcal {G} _ {\theta , \theta^ {\prime}} (z _ {t}, t, \mathcal {Y}, \mathcal {B}, \mathcal {Q}) \| _ {2} ^ {2} ], \tag {15} +$$ + +where $\mathcal{Q} = [\{\mathbf{q}_i^\uparrow\}_{i=1}^N, \mathbf{q}^\downarrow]$ denotes the set of frequency-aware tokens. While $\theta$ and $\theta'$ denote the frozen and trainable parameters of the pretrained LDM and our FICGen, respectively. + +# 4. Experiment + +# 4.1. Experimental Setups + +Datasets. Our FICGen is evaluated on five degraded-context datasets: ExDARK [18], RUOD [6], DAWN [13], DIOR [14], and blurred VOC 2012 [5]. More details are provided in the Supplementary Material. + +Implementation Details. Our FICGen is built upon the pre-trained SDv1.5 checkpoint and deployed only on the mid-level (i.e., $8 \times 8$ ) and lowest-resolution (i.e., $16 \times 16$ ) decoder layers. For training, all images are resized to $512 \times 512$ . We adopt the AdamW [19] optimizer with a learning rate of $1\mathrm{e}^{-4}$ , $\beta_{1} = 0.9$ , and $\beta_{2} = 0.999$ , and fine-tune for 300 epochs on 8 A100 with a batch size of 320. For inference, we use the EulerDiscreteScheduler [12] with 50 sampling steps and a classifier-free guidance scale [10] of 7.5. + +Evaluation Metrics. Following [2, 15], we comprehend + +![](images/132e4ca7a2c5c573edbed77f99eab531d017588675f0b4d323799b3bbbd38482.jpg) +Figure 4. Qualitative comparison of degraded images generated by different L2I methods across various scenarios. Zoom in for better details. + +
MethodFID ↓YOLO Score ↑
mAPAP50AP75
Oracle-39.257.143.0
LostGAN [35]57.1014.327.315.2
LayoutDiffusion [53]45.3120.037.419.3
ReCo [46]42.5621.140.723.1
MIGC [54]31.6424.541.825.6
GLIGEN [15]41.3125.844.427.8
CC-Diff [52]30.8826.444.228.5
AeroGen [36]38.5729.854.231.6
FICGen (ours)31.2531.249.934.6
+ +sively evaluate our FICGen from three aspects: fidelity, alignment, and trainability. Fidelity refers to the perceptual quality between synthetic and real images, measured by the Fréchet Inception Distance (FID) [9]. While alignment assesses spatial and semantic correspondences between layout conditions and generated instances, quantified via COCO-style average precision [16] with off-the-shelf detectors. Trainability examines the potential of synthetic images for auxiliary downstream training, assessed by mean Average Precision (mAP) under various training settings. + +# 4.2. Main Results + +# 4.2.1. Quantitative Evaluations + +Fidelity and Alignment. Tab. 1 reports the quantitative comparison of fidelity and alignment via reporting the COCO-style AP, where pre-trained Faster R-CNN (R50) [29] and Cascade R-CNN (R50) [1] are employed to infer synthetic + +Table 2. Quantitative comparison on the DIOR-H test set, with detection performance evaluated using a pre-trained YOLOv8 [37]. + +
MethodExDARK [18]RUOD [6]
mAP↑AP50↑AP75↑mAP↑AP50↑AP75↑
Oracle38.367.839.858.485.564.8
MIGC [54]28.659.224.121.143.917.9
CC-Diff [52]31.361.828.829.757.828.0
FICGen (ours)38.568.539.537.167.136.7
+ +Table 3. Quantitative comparison on ExDARK and RUOD test set, evaluated by a pre-trained Deformable-DETR [55]. + +images, with the results evaluated against corresponding real annotations. Notably, our FICGen successfully outperforms the existing MIGC [54] and CC-Diff [52] in terms of both FID and AP metrics, achieving remarkable FID scores of 25.10 and 42.40 on the RUOD and ExDARK, respectively. Moreover, for remote sensing object detection, FICGen reaches $82\%$ of the mAP relative to real data (i.e., 27.6 vs. 33.4), substantially outperforming MIGC (21.8) and CC-Diff (23.6) by a large margin. Intriguingly, FICGen even exceeds the real ExDARK test set in AP, highlighting its ability to generate degraded instances that precisely adhere to layouts while bridging the real-to-synthetic domain gap. + +To further verify the generative applicability across different detectors, we evaluate FICGen using YOLOv8 [37] and Deformable-DETR [55], as shown in Tabs. 2 and 3. FICGen surpasses AeroGen by 1.4 and 3.0 in mAP and $\mathrm{AP}_{75}$ on the DIOR-H, and outperforms CC-Diff by 7.2 and 7.4 on the ExDARK and RUOD, respectively. + +**Trainability.** Following the trainability protocol in [2], we use augmented annotations as layout inputs to generate synthetic training images and mix them with real ones, effectively doubling the training set. As shown in the last + +
DetectorBackboneTraining SetObject Detection (AP) for Sampled Classes /%mAP\( AP_{50} \)\( AP_{75} \)
# Real# Syntheticbicyclebusmotorbikeboatcarcupbottledog
Train with Full Real Data
RetinaNetPVT-MR:5.1k-44.858.231.331.136.025.528.747.835.265.433.9
Faster R-CNNX101R:5.1k-49.060.835.630.139.530.934.149.338.669.339.6
Train with Pure Synthetic Data
RetinaNetPVT-M-MIGC (S: 5.1k)32.846.720.419.922.514.116.033.423.549.019.1
Faster R-CNNX101-MIGC (S: 5.1k)33.648.218.520.423.117.021.133.224.749.422.1
RetinaNetPVT-M-CC-Diff (S: 5.1k)34.345.521.620.222.015.416.634.924.149.920.6
Faster R-CNNX101-CC-Diff (S: 5.1k)38.047.521.519.124.318.321.833.825.552.321.7
RetinaNetPVT-M-FICGen (S: 5.1k)35.150.423.220.024.916.216.833.325.351.022.1
Faster R-CNNX101-FICGen (S: 5.1k)38.149.921.722.024.318.920.436.426.751.425.1
Train with Synthetic Data and Finetune on Real Data
RetinaNetPVT-MR:5.1kMIGC (S: 5.1k)47.362.134.732.637.427.629.649.037.167.536.9
Faster R-CNNX101R:5.1kMIGC (S: 5.1k)51.361.838.031.540.032.135.849.739.970.740.9
RetinaNetPVT-MR:5.1kCC-Diff (S: 5.1k)47.960.934.331.536.427.130.548.137.067.035.8
Faster R-CNNX101R:5.1kCC-Diff (S: 5.1k)50.262.238.731.140.632.035.449.439.870.440.7
RetinaNetPVT-MR:5.1kFICGen (S: 5.1k)47.562.235.733.737.827.630.947.637.667.736.8
Faster R-CNNX101R:5.1kFICGen(S: 5.1k)51.462.938.232.840.333.135.649.940.670.942.1
Train with Real & Synthetic Data
Faster R-CNNX101R:5.1kMIGC (S: 5.1k)51.763.036.231.739.731.835.249.639.569.739.8
Faster R-CNNX101R:5.1kCC-Diff (S: 5.1k)50.961.636.531.840.231.634.750.339.570.040.7
Faster R-CNNX101R:5.1kFICGen (S: 5.1k)50.561.838.631.640.432.636.549.140.170.541.9
+ +Table 4. Quantitative comparison of trainability on the ExDARK dataset, evaluated by RetinaNet [17] (PVT-M [39]) and Faster R-CNN [29] (X101 [43]) under three training settings. + +
MethodDIOR-H [14]DAWN [13]ExDARK [18]VOC 2012 [5]
airport ↑mAP ↑\( AP_{50} \) ↑\( AP_{75} \) ↑bus ↑mAP ↑\( AP_{50} \) ↑\( AP_{75} \) ↑motorbike ↑mAP ↑\( AP_{50} \) ↑\( AP_{75} \) ↑boat ↑mAP ↑\( AP_{50} \) ↑\( AP_{75} \) ↑
Oracle32.233.455.635.021.824.846.023.033.335.867.434.542.948.376.852.5
MIGC [54]36.934.156.136.121.624.943.324.634.437.167.836.642.549.477.254.2
CC-Diff [52]37.634.356.136.326.225.846.125.834.137.067.736.644.449.677.354.3
FICGen (ours)38.135.556.837.627.525.944.526.537.137.768.037.747.350.577.455.4
+ +Table 5. Quantitative comparison of trainability on the DIOR, DAWN and ExDARK and natural VOC 2012 datasets, with detection evaluated using the Faster R-CNN (R50) object detector. + +![](images/b0e26b309c3cc17a7a07893e74749b15db0dac1b8c80d21cb496ac653965b172.jpg) +Figure 5. Visualization results generated by our FICGen on unseen geometric layouts augmented via flipping and translation. + +three rows of Tab. 4, we demonstrate for the first time that synthetic degraded images can significantly enhance the detection performance of downstream object detectors (i.e., +2.4 for bottle and +3.0 for motorbike). Moreover, + +we observe that training with synthetic images followed by fine-tuning on real ones yields the best performance. Tab. 5 further provides a comparative analysis under various degraded conditions. Crucially, FICGen consistently achieves the highest gains, boosting AP by 5.7 for "bus" (DAWN) and 3.8 for "motorbike" (ExDARK), highlighting its efficacy in mitigating data scarcity in degraded scenarios. + +Generalization. To evaluate the generalization of our FICGen, we conduct further experiments not only in severely degraded scenarios such as low-light and underwater conditions with horizontal layout control, but also in blurred and natural VOC 2012 [5], and oriented DIOR-R [3]. Extensive results demonstrate that synthetic images generated by FICGen exhibit superior quality and alignment. Further details are provided in the Supplementary Material. + +# 4.2.2. Qualitative Evaluations + +Fig. 4 presents a visual comparison of degraded images generated by various L2I methods. Our FICGen demonstrates superior performance in two key aspects: (a) Enhanced Layout Coherence. It resolves the dilemma of "object omission and merging", accurately aligning generated instances with + +![](images/acffbba831add96592c62aec3840f73e866dc205ffd60a2392745ad6d05205be.jpg) +Figure 6. Performance convergence curves when generated degraded images are incorporated into downstream detection. + +![](images/e1c385a64c31b40b6b6a48e09da037633704c5910690148ca4fecd8708760b45.jpg) +Figure 7. Generalization results of FICGen on mildly blurred and natural scenes, as well as under oriented layout control. + +the provided layouts in terms of quantity, scale, and position (i.e., aerial tennis courts and rainy cars), and effectively handling small objects (i.e., aerial vehicles). (b) Strong Occlusion Handling. In heavily occluded scenarios (i.e., underwater corals and fish), FICGen reconstructs dense instances while maintaining spatial consistency, whereas MIGC and CC-Diff fail to enforce overlapped layout constraints. Furthermore, Fig. 5 shows that FICGen generalizes effectively to unseen layouts, including those extending beyond image boundaries. As illustrated in Fig. 6, the synthesized degraded images generated by FICGen enhance downstream detection performance by improving mAP and accelerating training convergence. Finally, Fig. 7 highlights the generalization capability of FICGen in blurred and natural scenes, as well as under oriented layout controls. Additional visualizations are provided in the Supplementary Material. + +
FPRVFEAASFAmAP\( AP_{50} \)\( AP_{75} \)
XXX15.734.812.0
X✓X23.244.622.7
✓X✓18.536.313.9
✓✓X25.546.325.1
X✓✓25.646.925.3
✓✓✓27.648.727.6
+ +Table 6. Ablation study of FPR, VFEA, and ASFA modules, evaluated by the pre-trained Faster R-CNN (R50) on the DIOR test set. + +# 4.3. Ablation Studies + +Frequency Perceiver Resamplers. The foundation of frequency prototype learning lies in the FPR module, which combines a dual-query mechanism with contextual frequency knowledge. As shown in the last two rows of Tab. 6, FPR significantly improves the alignment scores from 25.6/46.9/25.3 to 27.6/48.7/27.6, demonstrating the effectiveness of contextual frequency guidance. + +Visual-Frequency Enhanced Attention. VFEA serves as a bridge to inject frequency knowledge (i.e., boundaries and textures) into the latent space while integrating the instance coherence map for contextual disentanglement. As shown in the second row of Tab. 6, even without FPR and ASFA, VFEA achieves competitive performance over the baseline, reaching 23.2/44.6/22.7 in AP. + +Adaptive Spatial-Frequency Aggregation. ASFA is designed to reconstruct refined degraded representations by integrating spatial and frequency perspectives. As evidenced by the results in the fourth and sixth rows of Tab. 6, incorporating ASFA improves the average precision (AP) from 25.5/46.3/25.1 to 27.6/48.7/27.6. This substantial gain underscores its critical role in consolidating contextual information and enhancing the fidelity of degraded image generation. + +# 5. Conclusion and Future Work + +In this paper, we introduce the task of degraded image generation and define the contextual illusion dilemma as its core challenge. To address this, we propose a frequency-inspired generative paradigm, FICGen. We first extract frequency prototypes from degraded contexts and employ a learnable dual-query mechanism to mine contextual frequency knowledge, which is then transferred into the latent diffusion space via a visual-frequency enhanced attention, coupled with an instance coherence map and a spatial-frequency aggregation module, to disentangle foreground instances from their surroundings and consolidate them into refined degraded representations. Extensive experiments on 5 degraded image datasets validate the effectiveness of FICGen. In future work, we plan to extend this frequency-inspired paradigm to 3D content generation for broader applications. + +# Acknowledgements + +This work is partially supported by grants from the National Natural Science Foundation of China (No. 62132002 and No. 62202010), Beijing Nova Program (No.20250484786), and the Fundamental Research Funds for the Central Universities. + +# References + +[1] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6154-6162, 2018. 5, 6 +[2] Kai Chen, Enze Xie, Zhe Chen, Yibo Wang, Lanqing Hong, Zhenguo Li, and Dit-Yan Yeung. Geodiffusion: Text-prompted geometric control for object detection data generation. arXiv preprint arXiv:2306.04607, 2023. 3, 4, 5, 6 +[3] Gong Cheng, Jiabao Wang, Ke Li, Xingxing Xie, Chunbo Lang, Yanqing Yao, and Junwei Han. Anchor-free oriented proposal generator for object detection. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2022. 7 +[4] Jiaxin Cheng, Xiao Liang, Xingjian Shi, Tong He, Tianjun Xiao, and Mu Li. Layoutdiffuse: Adapting foundational diffusion models for layout-to-image generation. arXiv preprint arXiv:2302.08908, 2023. 4 +[5] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 2, 5, 7 +[6] Chenping Fu, Risheng Liu, Xin Fan, Puyang Chen, Hao Fu, Wanqi Yuan, Ming Zhu, and Zhongxuan Luo. Rethinking general underwater object detection: Datasets, challenges, and solutions. Neurocomputing, 517:243-256, 2023. 2, 5, 6 +[7] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. In European Conference on Computer Vision, pages 89-106. Springer, 2022. 3 +[8] Yuchao Gu, Yipin Zhou, Yunfan Ye, Yixin Nie, Licheng Yu, Pingchuan Ma, Kevin Qinghong Lin, and Mike Zheng Shou. Roictrl: Boosting instance control for visual generation. arXiv preprint arXiv:2411.17949, 2024. 2, 3 +[9] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6 +[10] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 5 +[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3 +[12] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in neural information processing systems, 35:26565-26577, 2022. 5 + +[13] Mourad A Kenk and Mahmoud Hassaballah. Dawn: vehicle detection in adverse weather nature dataset. arXiv preprint arXiv:2008.05402, 2020. 2, 5, 7 +[14] Ke Li, Gang Wan, Gong Cheng, Liqui Meng, and Junwei Han. Object detection in optical remote sensing images: A survey and a new benchmark. ISPRS journal of photogrammetry and remote sensing, 159:296-307, 2020. 1, 2, 5, 7 +[15] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22511-22521, 2023. 2, 3, 4, 5, 6 +[16] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer vision-ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part v 13, pages 740-755. Springer, 2014. 2, 6 +[17] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 7 +[18] Yuen Peng Loh and Chee Seng Chan. Getting to know low-light images with the exclusively dark dataset. Computer Vision and Image Understanding, 178:30-42, 2019. 1, 2, 5, 6, 7 +[19] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5 +[20] Zhengyao Lv, Yuxiang Wei, Wangmeng Zuo, and KwanYee K Wong. Place: Adaptive layout-semantic fusion for semantic image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9264–9274, 2024. 2 +[21] Mingcan Ma, Changqun Xia, Chenxi Xie, Xiaowu Chen, and Jia Li. Boosting broader receptive fields for salient object detection. IEEE Transactions on Image Processing, 32:1026-1038, 2023. 2 +[22] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI conference on artificial intelligence, pages 4296–4304, 2024. 3 +[23] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 3 +[24] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 3 +[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 4 + +[26] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International conference on machine learning, pages 8821-8831. Pmlr, 2021. 3 +[27] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2): 3, 2022. 2, 3 +[28] Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In International conference on machine learning, pages 1060-1069. PMLR, 2016. 3 +[29] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 1, 5, 6, 7 +[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3 +[31] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 3 +[32] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2, 3 +[33] Zhe Shan, Yang Liu, Lei Zhou, Cheng Yan, Heng Wang, and Xia Xie. Ros-sam: High-quality interactive segmentation for remote sensing moving object. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 3625-3635, 2025. 2 +[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3 +[35] Wei Sun and Tianfu Wu. Image synthesis from reconfigurable layout and style. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10531-10540, 2019. 6 +[36] Datalog Tang, Xiangyong Cao, Xuan Wu, Jialin Li, Jing Yao, Xueru Bai, Dongsheng Jiang, Yin Li, and Deyu Meng. Aerogen: enhancing remote sensing object detection with diffusion-driven data generation. arXiv preprint arXiv:2411.15497, 2024. 3, 6 +[37] Ultralytics. Ultralytics yolov8. https://github.com/ultralytics/ultralytics, 2023.6 +[38] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 4 +[39] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyra- + +mid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 568-578, 2021. 7 +[40] Wenzhuang Wang, Mingcan Ma, Yong Chen, Changqun Xia, Zhenbao Liang, and Jia Li. Freegen: Bridging visual-linguistic discrepancies towards diffusion-based pixel-level data synthesis. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7916–7924, 2025. 2 +[41] Xudong Wang, Trevor Darrell, Sai Saketh Rambhatla, Rohit Girdhar, and Ishan Misra. Instancediffusion: Instance-level control for image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6232-6242, 2024. 3 +[42] Yinwei Wu, Xianpan Zhou, Bing Ma, Xuefeng Su, Kai Ma, and Xinchao Wang. Ifadapter: Instance feature control for grounded text-to-image generation. arXiv preprint arXiv:2409.08240, 2024. 3, 5 +[43] Saining Xie, Ross Girshick, Piotr Dólár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1492–1500, 2017. 7 +[44] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Finegrained text to image generation with attentional generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1316-1324, 2018. 3 +[45] Han Xue, Zhiwu Huang, Qianru Sun, Li Song, and Wenjun Zhang. Freestyle layout-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14256-14266, 2023. 2 +[46] Zhengyuan Yang, Jianfeng Wang, Zhe Gan, Linjie Li, Kevin Lin, Chenfei Wu, Nan Duan, Zicheng Liu, Ce Liu, Michael Zeng, et al. Reco: Region-controlled text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14246-14255, 2023. 2, 3, 6 +[47] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 4 +[48] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 3 +[49] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 5907-5915, 2017. 3 +[50] Han Zhang, Jing Yu Koh, Jason Baldridge, Honglak Lee, and Yinfei Yang. Cross-modal contrastive learning for text-to-image generation. In Proceedings of the IEEE/CVF con + +ference on computer vision and pattern recognition, pages 833-842, 2021. 3 +[51] Jingyang Zhang, Shiwei Li, Yuanxun Lu, Tian Fang, David McKinnon, Yanghai Tsin, Long Quan, and Yao Yao. Joint-net: Extending text-to-image diffusion for dense distribution modeling. arXiv preprint arXiv:2310.06347, 2023. 2 +[52] Mu Zhang, Yunfan Liu, Yue Liu, Hongtian Yu, and Qixiang Ye. Cc-diff: Enhancing contextual coherence in remote sensing image synthesis. arXiv preprint arXiv:2412.08464, 2024. 3, 5, 6, 7 +[53] Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22490-22499, 2023. 2, 3, 6 +[54] Dewei Zhou, You Li, Fan Ma, Xiaoting Zhang, and Yi Yang. Mige: Multi-instance generation controller for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6818-6828, 2024. 2, 3, 4, 5, 6, 7 +[55] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 6 \ No newline at end of file diff --git a/2025/FICGen_ Frequency-Inspired Contextual Disentanglement for Layout-driven Degraded Image Generation/images.zip b/2025/FICGen_ Frequency-Inspired Contextual Disentanglement for Layout-driven Degraded Image Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c83dce4fe569a6020ccfd1dac3b6339e3dfb72bc --- /dev/null +++ b/2025/FICGen_ Frequency-Inspired Contextual Disentanglement for Layout-driven Degraded Image Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13d9eb80ab00a58c75db1c74259611884bc76789cd8e7ec429c44a6ed3bc92b6 +size 1257945 diff --git a/2025/FICGen_ Frequency-Inspired Contextual Disentanglement for Layout-driven Degraded Image Generation/layout.json b/2025/FICGen_ Frequency-Inspired Contextual Disentanglement for Layout-driven Degraded Image Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7594e1c31a38a7b95d8cc69d0aea331282bf1170 --- /dev/null +++ b/2025/FICGen_ Frequency-Inspired Contextual Disentanglement for Layout-driven Degraded Image Generation/layout.json @@ -0,0 +1,8862 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 114, + 103, + 498, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 103, + 498, + 140 + ], + "spans": [ + { + "bbox": [ + 114, + 103, + 498, + 140 + ], + "type": "text", + "content": "FICGen: Frequency-Inspired Contextual Disentanglement for Layout-driven Degraded Image Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 160, + 548, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 160, + 548, + 203 + ], + "spans": [ + { + "bbox": [ + 64, + 160, + 548, + 203 + ], + "type": "text", + "content": "Wenzhuang Wang" + }, + { + "bbox": [ + 64, + 160, + 548, + 203 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 64, + 160, + 548, + 203 + ], + "type": "text", + "content": ", Yifan Zhao" + }, + { + "bbox": [ + 64, + 160, + 548, + 203 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 64, + 160, + 548, + 203 + ], + "type": "text", + "content": ", Mingcan Ma" + }, + { + "bbox": [ + 64, + 160, + 548, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 64, + 160, + 548, + 203 + ], + "type": "text", + "content": ", Ming Liu" + }, + { + "bbox": [ + 64, + 160, + 548, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 64, + 160, + 548, + 203 + ], + "type": "text", + "content": ", Zhonglin Jiang" + }, + { + "bbox": [ + 64, + 160, + 548, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 64, + 160, + 548, + 203 + ], + "type": "text", + "content": ", Yong Chen" + }, + { + "bbox": [ + 64, + 160, + 548, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 64, + 160, + 548, + 203 + ], + "type": "text", + "content": ", Jia Li" + }, + { + "bbox": [ + 64, + 160, + 548, + 203 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 64, + 160, + 548, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 160, + 548, + 203 + ], + "type": "text", + "content": "State Key Laboratory of Virtual Reality Technology and Systems, SCSE&QRI, Beihang University \n" + }, + { + "bbox": [ + 64, + 160, + 548, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 64, + 160, + 548, + 203 + ], + "type": "text", + "content": "Geely Automobile Research Institute (Ningbo) Co., Ltd" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 212, + 205, + 394, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 205, + 394, + 217 + ], + "spans": [ + { + "bbox": [ + 212, + 205, + 394, + 217 + ], + "type": "text", + "content": "{wz_wang, zhaoyf, jiali}@buaa.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 65, + 241, + 432, + 336 + ], + "blocks": [ + { + "bbox": [ + 65, + 241, + 432, + 336 + ], + "lines": [ + { + "bbox": [ + 65, + 241, + 432, + 336 + ], + "spans": [ + { + "bbox": [ + 65, + 241, + 432, + 336 + ], + "type": "image", + "image_path": "e52df08b170e168c189b4efce8c392106b33dcbb6739c1715062db63cf4229e3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 189, + 336, + 301, + 344 + ], + "lines": [ + { + "bbox": [ + 189, + 336, + 301, + 344 + ], + "spans": [ + { + "bbox": [ + 189, + 336, + 301, + 344 + ], + "type": "text", + "content": "(a) Contextual Illusion Dilemma" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 434, + 243, + 535, + 342 + ], + "blocks": [ + { + "bbox": [ + 434, + 243, + 535, + 342 + ], + "lines": [ + { + "bbox": [ + 434, + 243, + 535, + 342 + ], + "spans": [ + { + "bbox": [ + 434, + 243, + 535, + 342 + ], + "type": "image", + "image_path": "bae07c4cca4983817de4840ae34b4044aae24659988177a16250b6f238aa11ec.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 66, + 346, + 432, + 399 + ], + "blocks": [ + { + "bbox": [ + 66, + 346, + 432, + 399 + ], + "lines": [ + { + "bbox": [ + 66, + 346, + 432, + 399 + ], + "spans": [ + { + "bbox": [ + 66, + 346, + 432, + 399 + ], + "type": "image", + "image_path": "3f5b4734eab6bf4189c78fd9299c564b5beb98d241710036527c98722faee8b8.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 66, + 399, + 432, + 452 + ], + "blocks": [ + { + "bbox": [ + 66, + 399, + 432, + 452 + ], + "lines": [ + { + "bbox": [ + 66, + 399, + 432, + 452 + ], + "spans": [ + { + "bbox": [ + 66, + 399, + 432, + 452 + ], + "type": "image", + "image_path": "7f4d482c918dea9fcd464ff72fabf735832fb6b5680f8bbd1eb14eb100419a48.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 452, + 347, + 461 + ], + "lines": [ + { + "bbox": [ + 157, + 452, + 347, + 461 + ], + "spans": [ + { + "bbox": [ + 157, + 452, + 347, + 461 + ], + "type": "text", + "content": "(b) Robust Generation for Out-of-Distribution Layouts" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 436, + 451, + 536, + 462 + ], + "lines": [ + { + "bbox": [ + 436, + 451, + 536, + 462 + ], + "spans": [ + { + "bbox": [ + 436, + 451, + 536, + 462 + ], + "type": "text", + "content": "(c) Alignment & Trainability" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 436, + 346, + 536, + 445 + ], + "blocks": [ + { + "bbox": [ + 436, + 346, + 536, + 445 + ], + "lines": [ + { + "bbox": [ + 436, + 346, + 536, + 445 + ], + "spans": [ + { + "bbox": [ + 436, + 346, + 536, + 445 + ], + "type": "image", + "image_path": "c40921c12eebe00483cafdbeb616afcdc8a55eef5f6faa8157f34168d6938478.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 469, + 555, + 504 + ], + "lines": [ + { + "bbox": [ + 54, + 469, + 555, + 504 + ], + "spans": [ + { + "bbox": [ + 54, + 469, + 555, + 504 + ], + "type": "text", + "content": "Figure 1. Visualizations of degraded image generation by our FICGen. FICGen achieves superior quality compared to existing L2I methods and demonstrates strong robustness to OOD layouts. Moreover, it shows excellent trainability for boosting object detectors, i.e., Faster R-CNN (R50) [29]. The alignment and trainability are evaluated on the ExDARK [18] and DIOR-H [14], respectively." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 151, + 519, + 200, + 531 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 519, + 200, + 531 + ], + "spans": [ + { + "bbox": [ + 151, + 519, + 200, + 531 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 53, + 546, + 297, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 546, + 297, + 689 + ], + "spans": [ + { + "bbox": [ + 53, + 546, + 297, + 689 + ], + "type": "text", + "content": "Layout-to-image (L2I) generation has exhibited promising results in natural domains, but suffers from limited generative fidelity and weak alignment with user-provided layouts when applied to degraded scenes (i.e., low-light, underwater). We primarily attribute these limitations to the \"contextual illusion dilemma\" in degraded conditions, where foreground instances are overwhelmed by context-dominant frequency distributions. Motivated by this, our paper proposes a new Frequency-Inspired Contextual Disentangle-ment Generative (FICGen) paradigm, which seeks to transfer frequency knowledge of degraded images into the latent diffusion space, thereby facilitating the rendering of" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 520, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 556, + 712 + ], + "type": "text", + "content": "degraded instances and their surroundings via contextual frequency-aware guidance. To be specific, FICGen consists of two major steps. Firstly, we introduce a learnable dual-query mechanism, each paired with a dedicated frequency resampler, to extract contextual frequency prototypes from pre-collected degraded exemplars in the training set. Secondly, a visual-frequency enhanced attention is employed to inject frequency prototypes into the degraded generation process. To alleviate the contextual illusion and attribute leakage, an instance coherence map is developed to regulate latent-space disentanglement between individual instances and their surroundings, coupled with an adaptive spatial-frequency aggregation module to reconstruct spatial-frequency mixed degraded representations. Extensive experiments on 5 benchmarks involving a variety of degraded scenarios—from severe low-light to mild blur—demonstrate" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 704, + 272, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 704, + 272, + 714 + ], + "spans": [ + { + "bbox": [ + 66, + 704, + 272, + 714 + ], + "type": "text", + "content": "*Correspondence should be addressed to Yifan Zhao and Jia Li." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19097" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "content": "that FICGen consistently surpasses existing L2I methods in terms of generative fidelity, alignment and downstream auxiliary trainability." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 131, + 136, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 131, + 136, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 131, + 136, + 144 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 151, + 297, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 151, + 297, + 295 + ], + "spans": [ + { + "bbox": [ + 54, + 151, + 297, + 295 + ], + "type": "text", + "content": "Visual perception [21, 33] depends on a large amount of annotated datasets, particularly in degraded conditions such as low-light [18], and adverse weather [13]. Unfortunately, acquiring adequate datasets for these extreme conditions remains a major obstacle due to the high cost of manual collection and annotation. For instance, the number of images in ExDARK (7,363) [18] is only about 1/20th that of COCO 2017 (118,287) [16], highlighting the severe data scarcity. In light of this, alternative strategies are emerging that shift the focus from a \"model-centric\" to a \"data-centric\" perspective, where advanced generative models are used to synthesize realistic images for bolstering downstream tasks [40]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 295, + 296, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 295, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 54, + 295, + 296, + 415 + ], + "type": "text", + "content": "Witnessing the phenomenal capabilities of text-to-image (T2I) diffusion models [27, 32], numerous advances have been explored in conditioning image generation on visual labels, thereby enabling fine-grained controllability beyond what textual prompts alone can achieve. Such labels are often represented by masks [20, 45], depths [51], and bounding boxes (bboxes) [8, 15, 46, 53, 54], thereby facilitating visual generation via layout-to-image (L2I) synthesis. Among them, bboxes are the most intuitive for guiding the location of visual entities during L2I generation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 415, + 296, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 415, + 296, + 545 + ], + "spans": [ + { + "bbox": [ + 54, + 415, + 296, + 545 + ], + "type": "text", + "content": "Despite the success of L2I methods [15, 53, 54] in natural scenarios, they face challenges when adapting to practical degraded conditions. In Fig. 1 (a), aerial objects (i.e., vehicles) are often small in size and visually resemble surrounding structures (i.e., bridges), while underwater species (i.e., fish) frequently blend with other nearby creatures (i.e., corals), resulting in contextual entanglement. Such cases, difficult even for the human eye to distinguish and predominantly occurring in degraded scenarios, lead to generative illusions regarding object quantity, location, and interaction—termed the “contextual illusion dilemma” in our work." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 545, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 545, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 545, + 297, + 714 + ], + "type": "text", + "content": "Generally, as shown in Fig. 2, natural image exhibits a relatively balanced distribution of high-frequency (HF) and low-frequency (LF) components, making the distinction between instances and contexts pronounced. In contrast, HF details of instances in degraded contexts are attenuated, while LF contexts dominate the overall frequency distribution. Motivated by the frequency properties, we propose a new frequency-inspired contextual disentanglement paradigm to tackle the dilemma. Specifically, we first extract HF and LF prototypes from degraded contexts and associate them with each object and its surroundings. Second, a learnable dual-query mechanism equipped with frequency perceiver resamplers is introduced to simultaneously capture contextual frequency knowledge (i.e., boundaries and colours). Third, we inject" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 71, + 553, + 133 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 553, + 133 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 553, + 133 + ], + "type": "image", + "image_path": "36cac4420640b7fe574c6332faec8ca5a94fd87d8eadf4e7f65fb4aafa82e5c9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 141, + 555, + 163 + ], + "lines": [ + { + "bbox": [ + 313, + 141, + 555, + 163 + ], + "spans": [ + { + "bbox": [ + 313, + 141, + 555, + 163 + ], + "type": "text", + "content": "Figure 2. Frequency spectrum comparison between natural and degraded images." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 185, + 556, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 185, + 556, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 185, + 556, + 472 + ], + "type": "text", + "content": "the knowledge into the latent diffusion space via a visual-frequency enhanced attention, enabling better perception of frequency attributes during degraded generation. Meanwhile, an instance coherence map is designed to disentangle degraded instances from their surroundings, mitigating contextual entanglement and attribute leakage. Finally, a spatial-frequency aggregation module is developed to adaptively recombine them to obtain refined degraded representations. With the above efforts, FICGen offers a unified approach for degraded image synthesis from a frequency-inspired perspective. We evaluate FICGen on 5 degraded benchmarks, covering low-light (ExDARK [18]), underwater (RUOD [6]), remote sensing (DIOR [14]), adverse weather (DAWN [13]), and blur (VOC 2012 [5]) conditions. Extensive experiments demonstrate that FICGen achieves superior image quality while maintaining strong alignment with geometric layouts. In Fig.1(b), FICGen can perform robust OOD generation on unseen layouts. Fig.1(c) shows that FICGen achieves better alignment than the real test set across all classes except \"boat\", effectively narrowing the real-synthetic domain gap. Furthermore, FICGen consistently improves downstream tasks by " + }, + { + "bbox": [ + 313, + 185, + 556, + 472 + ], + "type": "inline_equation", + "content": "\\sim 2.0" + }, + { + "bbox": [ + 313, + 185, + 556, + 472 + ], + "type": "text", + "content": " mAP and substantially boosts detection accuracy for certain classes, i.e., " + }, + { + "bbox": [ + 313, + 185, + 556, + 472 + ], + "type": "inline_equation", + "content": "\\sim 6.0" + }, + { + "bbox": [ + 313, + 185, + 556, + 472 + ], + "type": "text", + "content": " (32.2 vs. 38.1) on \"airport\", underscoring its strong trainability." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 473, + 557, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 557, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 557, + 641 + ], + "type": "text", + "content": "In summary, our contributions are threefold: (1) We propose FICGen, the first attempt to address the contextual illusion dilemma in degraded image generation. (2) We introduce a learnable dual-query mechanism to capture contextual frequency knowledge, which is injected into the latent diffusion space via a visual-frequency enhanced attention module, and further complemented by an instance coherence map and a spatial-frequency aggregation module to enable latent-space disentanglement and consolidation. (3) Extensive experiments on 5 degraded datasets show that our FICGen outperforms SOTA L2I methods in both quantitative and qualitative evaluations. Notably, FICGen can be seamlessly integrated as a plug-and-play component to enhance other diffusion-based generative models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 651, + 400, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 651, + 400, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 651, + 400, + 664 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 671, + 471, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 471, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 471, + 685 + ], + "type": "text", + "content": "2.1. Text-Driven Image Synthesis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "content": "Text-driven image synthesis aims to create visual content conditioned on diverse textual descriptions. Early endeav-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19098" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 240 + ], + "type": "text", + "content": "ors focus on leveraging conditional GANs [28, 44, 49, 50] to generate realistic images. Recent advances in diffusion [11, 23, 24, 27, 30, 32, 34] and auto-regressive models [7, 26, 48] have revolutionized image synthesis by enabling stable training and producing high-fidelity images. Typically, DALL-E [26] exemplifies an auto-regressive work that treats textual and visual tokens as a unified data sequence, demonstrating remarkable zero-shot capability. Among diffusion-based efforts, latent diffusion model (LDM) [30] efficiently implements a denoising Markov chain in a low-dimensional latent space, striking a balance between generative quality and computational cost. Nevertheless, these text-prompted methods struggle with fine-grained control over object attributes (i.e., position)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 247, + 228, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 228, + 260 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 228, + 260 + ], + "type": "text", + "content": "2.2. Layout-Driven Image Synthesis" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 264, + 297, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 264, + 297, + 492 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 297, + 492 + ], + "type": "text", + "content": "Layout-driven image synthesis [2, 15, 22, 46, 53] encodes geometric layouts as positional tokens to enable fine-grained controllability. However, these methods often suffer from attribute leakage, where certain layouts mistakenly attend to irrelevant regions. Multi-instance control [41, 42, 54] addresses this by partitioning layouts into disjoint components, allowing each layout to focus on its corresponding region. Nonetheless, these approaches are confined to natural image domains. AeroGen [36] pioneers layout-conditioned generation in the context of remote sensing imagery, but it still struggles with semantic ambiguity and limited layout controllability. CC-Diff [52] further improves contextual coherence by modeling foreground-background relations, but overlooks frequency-aware cues and remains vulnerable to contextual illusion in various degraded scenes. In contrast, we propose a frequency-inspired contextual disentanglement paradigm, which introduces contextual frequency guidance into the generation process, delivering more realistic degraded images across diverse circumstances." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 501, + 113, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 501, + 113, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 501, + 113, + 514 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 521, + 142, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 142, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 142, + 533 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 538, + 297, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 538, + 297, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 538, + 297, + 658 + ], + "type": "text", + "content": "Latent Diffusion Model. Unlike vanilla DDPM [11], which generates images from Gaussian noise in pixel space, LDM [30] operates in a compressed latent space. Specifically, an input image " + }, + { + "bbox": [ + 55, + 538, + 297, + 658 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 55, + 538, + 297, + 658 + ], + "type": "text", + "content": " is first encoded into a latent representation " + }, + { + "bbox": [ + 55, + 538, + 297, + 658 + ], + "type": "inline_equation", + "content": "z_{0}" + }, + { + "bbox": [ + 55, + 538, + 297, + 658 + ], + "type": "text", + "content": ", which is then diffused into a noisy latent state via the forward process: " + }, + { + "bbox": [ + 55, + 538, + 297, + 658 + ], + "type": "inline_equation", + "content": "z_{t} = \\alpha_{t}z_{0} + \\sigma_{t}\\epsilon" + }, + { + "bbox": [ + 55, + 538, + 297, + 658 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 538, + 297, + 658 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 55, + 538, + 297, + 658 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 538, + 297, + 658 + ], + "type": "inline_equation", + "content": "\\sigma_{t}" + }, + { + "bbox": [ + 55, + 538, + 297, + 658 + ], + "type": "text", + "content": " are the fixed schedule, " + }, + { + "bbox": [ + 55, + 538, + 297, + 658 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,1)" + }, + { + "bbox": [ + 55, + 538, + 297, + 658 + ], + "type": "text", + "content": " is added standard Gaussian noise term and " + }, + { + "bbox": [ + 55, + 538, + 297, + 658 + ], + "type": "inline_equation", + "content": "t \\in [0,T]" + }, + { + "bbox": [ + 55, + 538, + 297, + 658 + ], + "type": "text", + "content": " denotes the timestep. Then, a denoising U-Net " + }, + { + "bbox": [ + 55, + 538, + 297, + 658 + ], + "type": "inline_equation", + "content": "(\\epsilon_{\\theta})" + }, + { + "bbox": [ + 55, + 538, + 297, + 658 + ], + "type": "text", + "content": " [31] is trained to predict the step-wise noise " + }, + { + "bbox": [ + 55, + 538, + 297, + 658 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 538, + 297, + 658 + ], + "type": "text", + "content": ", with the training objective " + }, + { + "bbox": [ + 55, + 538, + 297, + 658 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{LDM}" + }, + { + "bbox": [ + 55, + 538, + 297, + 658 + ], + "type": "text", + "content": " defined in Eq.1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 65, + 665, + 296, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 665, + 296, + 683 + ], + "spans": [ + { + "bbox": [ + 65, + 665, + 296, + 683 + ], + "type": "interline_equation", + "content": "\\underset {\\theta} {\\operatorname {m i n}} \\mathcal {L} _ {L D M} = \\mathbb {E} _ {z _ {0}, \\epsilon \\sim \\mathcal {N} (0, 1), t, \\tau} [ | | \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, \\tau) | | _ {2} ^ {2} ], \\tag {1}", + "image_path": "ccce50cee0734becf12a8ff8041af2cbb145da8b6da64b70baba46bad898960f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 292, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 292, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 292, + 700 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 292, + 700 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 689, + 292, + 700 + ], + "type": "text", + "content": " is the guidance embedding (i.e., text and bboxes)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "content": "Problem Formulation. Given instance layouts comprising" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 556, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 556, + 157 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 556, + 157 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 72, + 556, + 157 + ], + "type": "text", + "content": " object bboxes: " + }, + { + "bbox": [ + 313, + 72, + 556, + 157 + ], + "type": "inline_equation", + "content": "\\mathcal{B} = (\\{l_i, b_i\\}_{i=1}^N)" + }, + { + "bbox": [ + 313, + 72, + 556, + 157 + ], + "type": "text", + "content": ", where the location of " + }, + { + "bbox": [ + 313, + 72, + 556, + 157 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 72, + 556, + 157 + ], + "type": "text", + "content": "-th object " + }, + { + "bbox": [ + 313, + 72, + 556, + 157 + ], + "type": "inline_equation", + "content": "b_i = [x_i^1, y_i^1, x_i^2, y_i^2]" + }, + { + "bbox": [ + 313, + 72, + 556, + 157 + ], + "type": "text", + "content": " (i.e., top-left and bottom-right corners), and " + }, + { + "bbox": [ + 313, + 72, + 556, + 157 + ], + "type": "inline_equation", + "content": "l_i" + }, + { + "bbox": [ + 313, + 72, + 556, + 157 + ], + "type": "text", + "content": " is its corresponding semantic category. " + }, + { + "bbox": [ + 313, + 72, + 556, + 157 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 72, + 556, + 157 + ], + "type": "text", + "content": " denotes the global textual description of the degraded image. In this regard, our FICGen " + }, + { + "bbox": [ + 313, + 72, + 556, + 157 + ], + "type": "inline_equation", + "content": "(\\mathcal{G}(\\cdot))" + }, + { + "bbox": [ + 313, + 72, + 556, + 157 + ], + "type": "text", + "content": " strives to generate a degraded image conditioned on both spatial layout and semantics: " + }, + { + "bbox": [ + 313, + 72, + 556, + 157 + ], + "type": "inline_equation", + "content": "\\mathcal{I} = \\mathcal{G}(\\mathcal{B}, \\epsilon, \\mathcal{V})" + }, + { + "bbox": [ + 313, + 72, + 556, + 157 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 163, + 537, + 176 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 163, + 537, + 176 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 537, + 176 + ], + "type": "text", + "content": "3.2. Bridge Queries with Frequency Prototypes" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 179, + 555, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 179, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 179, + 555, + 239 + ], + "type": "text", + "content": "Existing L2I methods [2, 8, 15, 53, 54] primarily emphasize foreground layout conditions to render instances in natural domains, where fronts and contexts share a relatively uniform frequency distribution. Nonetheless, the naive strategy is inadequate for degraded domains with contextual illusion." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 240, + 556, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 240, + 556, + 359 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 556, + 359 + ], + "type": "text", + "content": "Exploring degraded contexts reveals that frequency disentanglement, a longstanding yet insightful concept, effectively mines weakened HF instances from the dominant LF surroundings. To this end, we seek to extract contextual frequency knowledge and transfer it into the latent space, thus achieving more realistic degraded generation via frequency-aware guidance. To achieve this, we face two main issues: 1) how to represent the frequency knowledge for degraded contexts? and 2) how to transfer frequency knowledge into the latent diffusion space? The answers are detailed below." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 360, + 557, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 557, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 557, + 515 + ], + "type": "text", + "content": "Frequency Prototypes Representation. For the former issue, our goal is to represent contextual frequency knowledge (i.e., textures) using frequency prototypes. To this end, we take two steps to manifest them, as depicted in Fig. 3(a). Taking the foreground exemplar as an example, we first build a class-specific dictionary by collecting same-class instances from the training set. A degraded instance is then sampled and assigned to the " + }, + { + "bbox": [ + 313, + 360, + 557, + 515 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 360, + 557, + 515 + ], + "type": "text", + "content": "-th bbox based on its position. All selected instances form a foreground exemplar that is often overwhelmed in degraded scenarios. Second, stacked convolutional networks extract its intermediate feature maps " + }, + { + "bbox": [ + 313, + 360, + 557, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 313, + 360, + 557, + 515 + ], + "type": "text", + "content": ", which are then transformed into the frequency domain " + }, + { + "bbox": [ + 313, + 360, + 557, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\mathbf{X})" + }, + { + "bbox": [ + 313, + 360, + 557, + 515 + ], + "type": "text", + "content": " via a Discrete Fourier Transform (DFT):" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 323, + 521, + 555, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 521, + 555, + 564 + ], + "spans": [ + { + "bbox": [ + 323, + 521, + 555, + 564 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {\\mathcal {F}} (u, v) = \\frac {1}{H \\times W} \\sum_ {h = 0} ^ {H - 1} \\sum_ {w = 0} ^ {W - 1} \\mathbf {X} (h, w) e ^ {- j 2 \\pi (u h + v w)}, \\tag {2}", + "image_path": "224b322ea0cdf8e3dc795a2b977fb20f0774717778ef34c411ca6748fccfdfe1.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 565, + 555, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 565, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 565, + 555, + 601 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 565, + 555, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathcal{F}}" + }, + { + "bbox": [ + 313, + 565, + 555, + 601 + ], + "type": "text", + "content": " is the complex-valued DFT output, " + }, + { + "bbox": [ + 313, + 565, + 555, + 601 + ], + "type": "inline_equation", + "content": "|v|" + }, + { + "bbox": [ + 313, + 565, + 555, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 565, + 555, + 601 + ], + "type": "inline_equation", + "content": "|u|" + }, + { + "bbox": [ + 313, + 565, + 555, + 601 + ], + "type": "text", + "content": " denote the width and height of normalised frequencies. Next, to isolate HF components, a binary mask " + }, + { + "bbox": [ + 313, + 565, + 555, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathcal{F}}" + }, + { + "bbox": [ + 313, + 565, + 555, + 601 + ], + "type": "text", + "content": " is applied:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 607, + 555, + 651 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 607, + 555, + 651 + ], + "spans": [ + { + "bbox": [ + 321, + 607, + 555, + 651 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {\\mathcal {F}} (u, v) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} u \\notin \\left[ \\frac {H}{2} - \\gamma H, \\frac {H}{2} + \\gamma H \\right], \\\\ & v \\notin \\left[ \\frac {W}{2} - \\gamma W, \\frac {W}{2} + \\gamma W \\right], \\\\ 0, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {3}", + "image_path": "414ae5d301c4990dbf7c475ccf7e0e15b9293adfd9b2ddb9f8e2eb1e050524af.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 658, + 556, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 556, + 694 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 556, + 694 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 658, + 556, + 694 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 658, + 556, + 694 + ], + "type": "text", + "content": " controls the size of the HF region. The masked frequencies " + }, + { + "bbox": [ + 313, + 658, + 556, + 694 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathcal{F}}\\mathbf{M}_{\\mathcal{F}}" + }, + { + "bbox": [ + 313, + 658, + 556, + 694 + ], + "type": "text", + "content": " are then enhanced using learnable channelwise weights " + }, + { + "bbox": [ + 313, + 658, + 556, + 694 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathcal{F}}" + }, + { + "bbox": [ + 313, + 658, + 556, + 694 + ], + "type": "text", + "content": ", and projected back to the spatial domain:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 376, + 700, + 555, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 700, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 376, + 700, + 555, + 715 + ], + "type": "interline_equation", + "content": "\\mathbf {X} ^ {\\uparrow} = \\mathcal {F} ^ {- 1} \\left(\\mathbf {X} _ {\\mathcal {F}} \\mathbf {M} _ {\\mathcal {F}}\\right) \\cdot \\mathbf {W} _ {\\mathcal {F}}, \\tag {4}", + "image_path": "076a0efabaa3be4e48e0450f26958ec988a08edd75224f3a164ceac804e4efe9.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19099" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 68, + 553, + 265 + ], + "blocks": [ + { + "bbox": [ + 58, + 68, + 553, + 265 + ], + "lines": [ + { + "bbox": [ + 58, + 68, + 553, + 265 + ], + "spans": [ + { + "bbox": [ + 58, + 68, + 553, + 265 + ], + "type": "image", + "image_path": "fa0b8feb95be71860094d4ca4929b0b6f7d5b173ea8038d81185383cfb8a2272.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 271, + 555, + 296 + ], + "lines": [ + { + "bbox": [ + 55, + 271, + 555, + 296 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 555, + 296 + ], + "type": "text", + "content": "Figure 3. The overview pipeline of our FICGen, consisting of (a) Frequency Perceiver Resamplers, (b) Visual-Frequency Enhanced Attention, and (c) Adaptive Spatial-Frequency Aggregation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 315, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 315, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 54, + 315, + 296, + 422 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 54, + 315, + 296, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{-1}" + }, + { + "bbox": [ + 54, + 315, + 296, + 422 + ], + "type": "text", + "content": " denotes the inverse DFT. The result " + }, + { + "bbox": [ + 54, + 315, + 296, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{\\dagger}" + }, + { + "bbox": [ + 54, + 315, + 296, + 422 + ], + "type": "text", + "content": " represents the enhanced HF feature map. A similar process is applied to the complementary region using " + }, + { + "bbox": [ + 54, + 315, + 296, + 422 + ], + "type": "inline_equation", + "content": "1 - \\mathbf{M}_F" + }, + { + "bbox": [ + 54, + 315, + 296, + 422 + ], + "type": "text", + "content": ", yielding the LF feature map " + }, + { + "bbox": [ + 54, + 315, + 296, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{\\downarrow}" + }, + { + "bbox": [ + 54, + 315, + 296, + 422 + ], + "type": "text", + "content": ". Inspired by the prototypical learning, we adopt their mean pooled results as HF and LF prototypes, i.e., " + }, + { + "bbox": [ + 54, + 315, + 296, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{p}^{\\uparrow} = \\{p_i^\\uparrow\\}_{i=1}^N \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 54, + 315, + 296, + 422 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 315, + 296, + 422 + ], + "type": "inline_equation", + "content": "p^\\downarrow \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 54, + 315, + 296, + 422 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 54, + 315, + 296, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{p}^{\\uparrow}" + }, + { + "bbox": [ + 54, + 315, + 296, + 422 + ], + "type": "text", + "content": " corresponds to the HF representations of degraded instances, " + }, + { + "bbox": [ + 54, + 315, + 296, + 422 + ], + "type": "inline_equation", + "content": "p^\\downarrow" + }, + { + "bbox": [ + 54, + 315, + 296, + 422 + ], + "type": "text", + "content": " reflects the LF characteristics of surrounding contexts, and " + }, + { + "bbox": [ + 54, + 315, + 296, + 422 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 54, + 315, + 296, + 422 + ], + "type": "text", + "content": " denotes the feature dimension." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 423, + 296, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 296, + 483 + ], + "type": "text", + "content": "Frequency Perceiver Resamplers. To address the latter issue, inspired by the Perceiver architecture [47], we adopt a learnable dual-query mechanism that employs separate frequency resamplers to bridge HF and LF prototypes, thereby capturing contextual frequency cues:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 84, + 491, + 295, + 507 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 491, + 295, + 507 + ], + "spans": [ + { + "bbox": [ + 84, + 491, + 295, + 507 + ], + "type": "interline_equation", + "content": "\\mathbf {q} _ {i} ^ {\\uparrow} = \\operatorname {H F - R e s a m p l e r} \\left(\\mathbf {Q} ^ {\\uparrow}, \\phi_ {k 1} ^ {r} \\left(p _ {i} ^ {\\uparrow}\\right), \\phi_ {v 1} ^ {r} \\left(p _ {i} ^ {\\uparrow}\\right)\\right), \\tag {5}", + "image_path": "8f20deb4af99cd08b1c4ede2b16339906775a7d22f19291b5d15ac1105ea0d86.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 85, + 508, + 296, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 508, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 85, + 508, + 296, + 525 + ], + "type": "interline_equation", + "content": "\\mathbf {q} ^ {\\downarrow} = \\operatorname {L F - R e s a m p l e r} \\left(\\mathbf {Q} ^ {\\downarrow}, \\phi_ {k 1} ^ {g} \\left(p ^ {\\downarrow}\\right), \\phi_ {v 1} ^ {g} \\left(p ^ {\\downarrow}\\right)\\right), \\tag {6}", + "image_path": "fca5579d2e9dbd1a99b1a50f3191ea58ef2c907922e061c01ac60e2e2eba2609.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 532, + 296, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 296, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 296, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 532, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\phi_{*}^{r},\\phi_{*}^{g}\\in \\mathbb{R}^{d\\times d}" + }, + { + "bbox": [ + 55, + 532, + 296, + 605 + ], + "type": "text", + "content": " are linear projection layers. The dual queries " + }, + { + "bbox": [ + 55, + 532, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^{\\uparrow}" + }, + { + "bbox": [ + 55, + 532, + 296, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 532, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^{\\downarrow}" + }, + { + "bbox": [ + 55, + 532, + 296, + 605 + ], + "type": "text", + "content": " are learned through their respective resamplers, each composed of multiple transformer blocks [38], enabling continuous interactions between the dual queries and frequency prototypes, yielding frequency-aware instance and context tokens, i.e., " + }, + { + "bbox": [ + 55, + 532, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_i^\\uparrow" + }, + { + "bbox": [ + 55, + 532, + 296, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 532, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{q}^{\\downarrow}" + }, + { + "bbox": [ + 55, + 532, + 296, + 605 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 296, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 665 + ], + "type": "text", + "content": "With the above dedicated design, the synergy between frequency prototypes and the dual-query mechanism enables the simultaneous perception of essential frequency cues and contextual appearances, which are then incorporated into the degraded image generation process." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 672, + 281, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 281, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 281, + 685 + ], + "type": "text", + "content": "3.3. Contextual Frequency Knowledge Transfer" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Previous L2I efforts [2, 4, 15, 54] only focus on instance generation conditioned on layouts, while neglecting the sur" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 316, + 553, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 553, + 339 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 553, + 339 + ], + "type": "text", + "content": "rounding context. This limitation is particularly detrimental in degraded scenarios involving contextual entanglement." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 340, + 555, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 555, + 495 + ], + "type": "text", + "content": "Visual-Frequency Enhanced Attention. To transfer frequency properties into the degraded generation, we employ a visual-frequency enhanced attention module (Fig.3(c)) that injects frequency-aware tokens " + }, + { + "bbox": [ + 313, + 340, + 555, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_i^\\uparrow" + }, + { + "bbox": [ + 313, + 340, + 555, + 495 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 340, + 555, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{q}^{\\downarrow}" + }, + { + "bbox": [ + 313, + 340, + 555, + 495 + ], + "type": "text", + "content": " into the latent diffusion space. Specifically, each bbox " + }, + { + "bbox": [ + 313, + 340, + 555, + 495 + ], + "type": "inline_equation", + "content": "b_{i}" + }, + { + "bbox": [ + 313, + 340, + 555, + 495 + ], + "type": "text", + "content": " is first encoded using a Fourier embedding, followed by a multi-layer perceptron " + }, + { + "bbox": [ + 313, + 340, + 555, + 495 + ], + "type": "inline_equation", + "content": "(\\mathbf{E}_{box})" + }, + { + "bbox": [ + 313, + 340, + 555, + 495 + ], + "type": "text", + "content": " to produce positional tokens. In parallel, its semantic class " + }, + { + "bbox": [ + 313, + 340, + 555, + 495 + ], + "type": "inline_equation", + "content": "l_{i}" + }, + { + "bbox": [ + 313, + 340, + 555, + 495 + ], + "type": "text", + "content": " is encoded via a CLIP text encoder " + }, + { + "bbox": [ + 313, + 340, + 555, + 495 + ], + "type": "inline_equation", + "content": "(\\mathbf{E}_{clip})" + }, + { + "bbox": [ + 313, + 340, + 555, + 495 + ], + "type": "text", + "content": " [25]. Notably, " + }, + { + "bbox": [ + 313, + 340, + 555, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_i^\\uparrow" + }, + { + "bbox": [ + 313, + 340, + 555, + 495 + ], + "type": "text", + "content": " is then integrated with them to provide frequency-aware cues for instance generation. As for the surrounding, " + }, + { + "bbox": [ + 313, + 340, + 555, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{q}^{\\downarrow}" + }, + { + "bbox": [ + 313, + 340, + 555, + 495 + ], + "type": "text", + "content": " is fused with text tokens extracted from the global caption " + }, + { + "bbox": [ + 313, + 340, + 555, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 340, + 555, + 495 + ], + "type": "text", + "content": ", thereby facilitating holistic contextual guidance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 355, + 505, + 555, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 505, + 555, + 520 + ], + "spans": [ + { + "bbox": [ + 355, + 505, + 555, + 520 + ], + "type": "interline_equation", + "content": "\\mathbf {R} _ {i} = \\left[ \\mathbf {q} _ {i} ^ {\\uparrow}; \\mathbf {E} _ {c l i p} \\left(l _ {i}\\right); \\mathbf {E} _ {b o x} (\\text {F o u r i e r} (b _ {i})) \\right], \\tag {7}", + "image_path": "21a93a40789c968f0a6223f52592599e6800b2345e576a8844847409e7cd61ad.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 358, + 522, + 555, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 522, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 358, + 522, + 555, + 536 + ], + "type": "interline_equation", + "content": "\\mathbf {G} = \\left[ \\mathbf {q} ^ {\\downarrow}; \\mathbf {E} _ {c l i p} (\\mathcal {Y}) \\right], \\tag {8}", + "image_path": "89aafda540ffa5ddbd1bafdf33e27eb66483d8054212ebf51b1c45e932f27176.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 546, + 488, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 488, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 488, + 558 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 546, + 488, + 558 + ], + "type": "inline_equation", + "content": "[;]" + }, + { + "bbox": [ + 313, + 546, + 488, + 558 + ], + "type": "text", + "content": " represents concatenate operation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 559, + 555, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 555, + 618 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 555, + 618 + ], + "type": "text", + "content": "It is preferable for each layout to affect only its corresponding local region. To this end, an instance coherence map " + }, + { + "bbox": [ + 313, + 559, + 555, + 618 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}_i" + }, + { + "bbox": [ + 313, + 559, + 555, + 618 + ], + "type": "text", + "content": " is utilized to isolate individual instances and enforce precise interactions between layout conditions and their associated latent features." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 343, + 626, + 555, + 653 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 626, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 343, + 626, + 555, + 653 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {i} ^ {r} = \\operatorname {s o f t m a x} \\left(\\frac {Q _ {v} \\phi_ {k 2} ^ {r} (\\mathbf {R} _ {i})}{\\sqrt {d}}\\right) \\cdot \\phi_ {v 2} ^ {r} (\\mathbf {R} _ {i}) \\odot \\hat {\\mathbf {M}} _ {i}, \\tag {9}", + "image_path": "8d41ff526ff580f6cecbf297a70edd4774da35a9a3950ba2c665fd8f6bc75cc6.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 343, + 654, + 555, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 654, + 555, + 681 + ], + "spans": [ + { + "bbox": [ + 343, + 654, + 555, + 681 + ], + "type": "interline_equation", + "content": "\\mathbf {f} ^ {g} = \\operatorname {s o f t m a x} \\left(\\frac {Q _ {v} \\phi_ {k 2} ^ {g} (\\mathbf {G})}{\\sqrt {d}}\\right) \\cdot \\phi_ {v 2} ^ {g} (\\mathbf {G}) \\odot \\hat {\\mathbf {M}} ^ {g}, \\tag {10}", + "image_path": "4ce95599af012f817e2fffd8426c1ab41e054587a0f637affc214b31e06dc6cc.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "Q_{v}" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": " denotes the linear projection of latent features, " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": " is element-wise multiplication, and " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\phi_{*}^{r},\\phi_{*}^{g}" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": " are linear layers." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19100" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 95 + ], + "type": "text", + "content": "Finally, " + }, + { + "bbox": [ + 55, + 72, + 297, + 95 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i^r" + }, + { + "bbox": [ + 55, + 72, + 297, + 95 + ], + "type": "text", + "content": " indicates the latent feature of the " + }, + { + "bbox": [ + 55, + 72, + 297, + 95 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 72, + 297, + 95 + ], + "type": "text", + "content": "-th instance, while " + }, + { + "bbox": [ + 55, + 72, + 297, + 95 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^g" + }, + { + "bbox": [ + 55, + 72, + 297, + 95 + ], + "type": "text", + "content": " corresponds to the contextual representation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 297, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 297, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 297, + 145 + ], + "type": "text", + "content": "Instance Coherence Map. To avert feature contamination and attribute leakage, we first construct an instance-level mask " + }, + { + "bbox": [ + 55, + 96, + 297, + 145 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}_i" + }, + { + "bbox": [ + 55, + 96, + 297, + 145 + ], + "type": "text", + "content": " from each bbox layout, as in [42, 54], which enables precise generation of instance-specific latent features:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 108, + 151, + 296, + 184 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 151, + 296, + 184 + ], + "spans": [ + { + "bbox": [ + 108, + 151, + 296, + 184 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {M}} _ {i} (x, y) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} [ x, y ] \\in b _ {i}, \\\\ 0, & \\text {i f} [ x, y ] \\notin b _ {i}, \\end{array} \\right. \\tag {11}", + "image_path": "71e769de3b584184af140339b293369061c4ad0c4000ec28a8c1867a5343a122.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 190, + 296, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 190, + 296, + 250 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 296, + 250 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 190, + 296, + 250 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 190, + 296, + 250 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 190, + 296, + 250 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 190, + 296, + 250 + ], + "type": "text", + "content": " denote the mask coordinates. On the basis, we further separate foreground instances from their surroundings by constructing the complementary region " + }, + { + "bbox": [ + 55, + 190, + 296, + 250 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}^g" + }, + { + "bbox": [ + 55, + 190, + 296, + 250 + ], + "type": "text", + "content": ", which encompasses all areas outside the instance-level masks, thus enabling contextual disentanglement:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 256, + 296, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 256, + 296, + 289 + ], + "spans": [ + { + "bbox": [ + 134, + 256, + 296, + 289 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {M}} ^ {g} = 1 - \\sum_ {i = 1} ^ {N} \\hat {\\mathbf {M}} _ {i}. \\tag {12}", + "image_path": "e40a681cea66ff489f5d8ad8b1fa4fa8b2a09baff51dbfcc9c0cac6473646142.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 296, + 296, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 296, + 296, + 334 + ], + "spans": [ + { + "bbox": [ + 55, + 296, + 296, + 334 + ], + "type": "text", + "content": "Then, we combine " + }, + { + "bbox": [ + 55, + 296, + 296, + 334 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i^r" + }, + { + "bbox": [ + 55, + 296, + 296, + 334 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 296, + 296, + 334 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^g" + }, + { + "bbox": [ + 55, + 296, + 296, + 334 + ], + "type": "text", + "content": " to produce the unified latent degraded representations " + }, + { + "bbox": [ + 55, + 296, + 296, + 334 + ], + "type": "inline_equation", + "content": "\\mathbf{F} = [\\{\\mathbf{f}_i^r\\}_{i=1}^N, \\mathbf{f}^g]" + }, + { + "bbox": [ + 55, + 296, + 296, + 334 + ], + "type": "text", + "content": ", which serve as the key for degraded image generation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 339, + 270, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 339, + 270, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 339, + 270, + 353 + ], + "type": "text", + "content": "3.4. Adaptive Spatial-Frequency Aggregation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 356, + 297, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 356, + 297, + 499 + ], + "spans": [ + { + "bbox": [ + 54, + 356, + 297, + 499 + ], + "type": "text", + "content": "After harvesting the degraded representations " + }, + { + "bbox": [ + 54, + 356, + 297, + 499 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 54, + 356, + 297, + 499 + ], + "type": "text", + "content": ", we move forward to jointly consolidate and reconstruct them. In contrast to existing methods that either directly sum multiple instances [52] or combine them solely within the spatial domain [54], which often suffer from \"object omission and merging\" for overlapping instances, as illustrated in Fig. 3(c), we aim to adaptively aggregate the degraded instances and their surroundings in a spatial-frequency mixed manner. \"Spatial\" captures contextual relational dependencies, while \"frequency\" emphasizes fine-grained attributes across instances, i.e., boundary sharpness and textures, which can be formulated as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 506, + 296, + 538 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 506, + 296, + 538 + ], + "spans": [ + { + "bbox": [ + 66, + 506, + 296, + 538 + ], + "type": "interline_equation", + "content": "\\mathbf {F} ^ {s} = \\mathbf {S A M} \\left(\\left[ \\sum_ {i = 1} ^ {N} \\mathbf {f} _ {i} ^ {r}, \\mathbf {f} ^ {g} \\right]\\right), \\mathbf {F} ^ {f} = \\mathbf {F A M} \\left(\\left[ \\sum_ {i = 1} ^ {N} \\mathbf {f} _ {i} ^ {r}, \\mathbf {f} ^ {g} \\right]\\right), \\tag {13}", + "image_path": "426bc184dd9a0cda7bc04bf42ea640478ebf7019eb8ab51d94ab45f8bca48521.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 88, + 545, + 296, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 545, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 88, + 545, + 296, + 578 + ], + "type": "interline_equation", + "content": "\\delta^ {f i n a l} = \\sum_ {i = 1} ^ {N + 1} \\operatorname {s o f t m a x} \\left(\\delta^ {f u s i o n}\\right) \\left[ \\sum_ {i = 1} ^ {N} \\mathbf {f} _ {i} ^ {r}, \\mathbf {f} ^ {g} \\right], \\tag {14}", + "image_path": "3f1d9fbbbf4379bc6f7d5aa25f5d5d8122001d69a9a06e6fbe4d23df858425ae.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 582, + 296, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 666 + ], + "type": "text", + "content": "where SAM and FAM represent the spatial and frequency aggregation modules, respectively, consisting of standard self- and frequency-attention mechanisms. " + }, + { + "bbox": [ + 55, + 582, + 296, + 666 + ], + "type": "inline_equation", + "content": "\\delta^{fusion} = \\zeta * ([\\mathbf{F}^s,\\mathbf{F}^f])" + }, + { + "bbox": [ + 55, + 582, + 296, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 582, + 296, + 666 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 55, + 582, + 296, + 666 + ], + "type": "text", + "content": " is a learnable single-layer depthwise convolution to enhance local perception. Finally, the refined degraded representation " + }, + { + "bbox": [ + 55, + 582, + 296, + 666 + ], + "type": "inline_equation", + "content": "\\delta^{final} \\in \\mathbb{R}^{h\\times w\\times c}" + }, + { + "bbox": [ + 55, + 582, + 296, + 666 + ], + "type": "text", + "content": " is decoded into the target degraded image." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 672, + 173, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 173, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 173, + 685 + ], + "type": "text", + "content": "3.5. Learning Procedure" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 689, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 297, + 714 + ], + "type": "text", + "content": "During training, the parameters of the pretrained LDM remain fixed, with optimization focused solely on FICGen." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 318, + 71, + 552, + 313 + ], + "blocks": [ + { + "bbox": [ + 318, + 71, + 552, + 313 + ], + "lines": [ + { + "bbox": [ + 318, + 71, + 552, + 313 + ], + "spans": [ + { + "bbox": [ + 318, + 71, + 552, + 313 + ], + "type": "table", + "html": "
DetectorMethodFID ↓mAP ↑\\( AP_{50} \\) ↑\\( AP_{75} \\) ↑
DIOR-H [14]
Faster R-CNNOracle-33.455.635.0
Faster R-CNNMIGC [54]31.6421.838.417.5
Faster R-CNNCC-Diff [52]30.8823.642.421.4
Faster R-CNNFICGen (ours)31.2527.648.727.6
RUOD [6]
Faster R-CNNOracle-50.580.254.4
Faster R-CNNMIGC [54]26.5027.254.124.6
Faster R-CNNCC-Diff [52]25.2129.758.427.9
Faster R-CNNFICGen (ours)25.1037.068.636.5
blurred VOC 2012 [5]
Faster R-CNNOracle-31.556.532.4
Faster R-CNNMIGC [54]62.6634.765.833.1
Faster R-CNNCC-Diff [52]62.2036.767.636.3
Faster R-CNNFICGen (ours)58.0240.770.342.7
ExDARK [18]
Cascade R-CNNOracle-37.265.837.8
Cascade R-CNNMIGC [54]45.7632.463.529.5
Cascade R-CNNCC-Diff [52]44.2635.165.634.1
Cascade R-CNNFICGen (ours)42.4042.573.045.1
DAWN [13]
Cascade R-CNNOracle-27.346.426.3
Cascade R-CNNMIGC [54]70.1012.632.38.6
Cascade R-CNNCC-Diff [52]68.5615.733.914.8
Cascade R-CNNFICGen (ours)68.3122.644.321.5
", + "image_path": "c2364cd391b77b654d1c3d3f2e174c029435df0c31b9031d960db3e5fd4d818f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 321, + 555, + 377 + ], + "lines": [ + { + "bbox": [ + 313, + 321, + 555, + 377 + ], + "spans": [ + { + "bbox": [ + 313, + 321, + 555, + 377 + ], + "type": "text", + "content": "Table 1. Quantitative comparison of fidelity and alignment on various degraded conditions. The performance is evaluated by off-the-shelf detectors Faster R-CNN [29] and Cascade R-CNN [1] on synthetic test set. \"Oracle\" denotes the real test baseline. The best results are in red and the second best results are in underline." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 398, + 555, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 398, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 555, + 422 + ], + "type": "text", + "content": "The learning objective integrates the original mean-square error loss and the contextual frequency guidance:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 430, + 569, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 569, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 569, + 460 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta^ {\\prime}} \\mathcal {L} _ {F I C G e n} = \\mathbb {E} _ {z _ {0}, \\epsilon \\sim \\mathcal {N} (0, 1), t, \\mathcal {Y}} [ \\| \\epsilon - \\mathcal {G} _ {\\theta , \\theta^ {\\prime}} (z _ {t}, t, \\mathcal {Y}, \\mathcal {B}, \\mathcal {Q}) \\| _ {2} ^ {2} ], \\tag {15}", + "image_path": "828ff8657b233751d482e86ce9508f8c3195e4101ed2b4af4f3caf2967c09616.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 460, + 555, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 498 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 460, + 555, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{Q} = [\\{\\mathbf{q}_i^\\uparrow\\}_{i=1}^N, \\mathbf{q}^\\downarrow]" + }, + { + "bbox": [ + 313, + 460, + 555, + 498 + ], + "type": "text", + "content": " denotes the set of frequency-aware tokens. While " + }, + { + "bbox": [ + 313, + 460, + 555, + 498 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 460, + 555, + 498 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 460, + 555, + 498 + ], + "type": "inline_equation", + "content": "\\theta'" + }, + { + "bbox": [ + 313, + 460, + 555, + 498 + ], + "type": "text", + "content": " denote the frozen and trainable parameters of the pretrained LDM and our FICGen, respectively." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 509, + 391, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 391, + 522 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 391, + 522 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 529, + 435, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 435, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 435, + 542 + ], + "type": "text", + "content": "4.1. Experimental Setups" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 545, + 555, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 594 + ], + "type": "text", + "content": "Datasets. Our FICGen is evaluated on five degraded-context datasets: ExDARK [18], RUOD [6], DAWN [13], DIOR [14], and blurred VOC 2012 [5]. More details are provided in the Supplementary Material." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 594, + 556, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 701 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 701 + ], + "type": "text", + "content": "Implementation Details. Our FICGen is built upon the pre-trained SDv1.5 checkpoint and deployed only on the mid-level (i.e., " + }, + { + "bbox": [ + 313, + 594, + 556, + 701 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 313, + 594, + 556, + 701 + ], + "type": "text", + "content": ") and lowest-resolution (i.e., " + }, + { + "bbox": [ + 313, + 594, + 556, + 701 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 313, + 594, + 556, + 701 + ], + "type": "text", + "content": ") decoder layers. For training, all images are resized to " + }, + { + "bbox": [ + 313, + 594, + 556, + 701 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 594, + 556, + 701 + ], + "type": "text", + "content": ". We adopt the AdamW [19] optimizer with a learning rate of " + }, + { + "bbox": [ + 313, + 594, + 556, + 701 + ], + "type": "inline_equation", + "content": "1\\mathrm{e}^{-4}" + }, + { + "bbox": [ + 313, + 594, + 556, + 701 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 594, + 556, + 701 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 313, + 594, + 556, + 701 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 594, + 556, + 701 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.999" + }, + { + "bbox": [ + 313, + 594, + 556, + 701 + ], + "type": "text", + "content": ", and fine-tune for 300 epochs on 8 A100 with a batch size of 320. For inference, we use the EulerDiscreteScheduler [12] with 50 sampling steps and a classifier-free guidance scale [10] of 7.5." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 555, + 714 + ], + "type": "text", + "content": "Evaluation Metrics. Following [2, 15], we comprehend" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19101" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 68, + 558, + 295 + ], + "blocks": [ + { + "bbox": [ + 57, + 68, + 558, + 295 + ], + "lines": [ + { + "bbox": [ + 57, + 68, + 558, + 295 + ], + "spans": [ + { + "bbox": [ + 57, + 68, + 558, + 295 + ], + "type": "image", + "image_path": "132e4ca7a2c5c573edbed77f99eab531d017588675f0b4d323799b3bbbd38482.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 302, + 555, + 314 + ], + "lines": [ + { + "bbox": [ + 55, + 302, + 555, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 302, + 555, + 314 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison of degraded images generated by different L2I methods across various scenarios. Zoom in for better details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 63, + 332, + 293, + 454 + ], + "blocks": [ + { + "bbox": [ + 63, + 332, + 293, + 454 + ], + "lines": [ + { + "bbox": [ + 63, + 332, + 293, + 454 + ], + "spans": [ + { + "bbox": [ + 63, + 332, + 293, + 454 + ], + "type": "table", + "html": "
MethodFID ↓YOLO Score ↑
mAPAP50AP75
Oracle-39.257.143.0
LostGAN [35]57.1014.327.315.2
LayoutDiffusion [53]45.3120.037.419.3
ReCo [46]42.5621.140.723.1
MIGC [54]31.6424.541.825.6
GLIGEN [15]41.3125.844.427.8
CC-Diff [52]30.8826.444.228.5
AeroGen [36]38.5729.854.231.6
FICGen (ours)31.2531.249.934.6
", + "image_path": "a3a51c6fc63a2bf7b44533942ac9cb0b53f359dade392ba7b0c37bb21cea3d16.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 506, + 297, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 506, + 297, + 626 + ], + "spans": [ + { + "bbox": [ + 54, + 506, + 297, + 626 + ], + "type": "text", + "content": "sively evaluate our FICGen from three aspects: fidelity, alignment, and trainability. Fidelity refers to the perceptual quality between synthetic and real images, measured by the Fréchet Inception Distance (FID) [9]. While alignment assesses spatial and semantic correspondences between layout conditions and generated instances, quantified via COCO-style average precision [16] with off-the-shelf detectors. Trainability examines the potential of synthetic images for auxiliary downstream training, assessed by mean Average Precision (mAP) under various training settings." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 633, + 141, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 633, + 141, + 645 + ], + "spans": [ + { + "bbox": [ + 55, + 633, + 141, + 645 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 651, + 190, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 190, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 190, + 662 + ], + "type": "text", + "content": "4.2.1. Quantitative Evaluations" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 297, + 714 + ], + "type": "text", + "content": "Fidelity and Alignment. Tab. 1 reports the quantitative comparison of fidelity and alignment via reporting the COCO-style AP, where pre-trained Faster R-CNN (R50) [29] and Cascade R-CNN (R50) [1] are employed to infer synthetic" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 319, + 332, + 552, + 397 + ], + "blocks": [ + { + "bbox": [ + 55, + 462, + 295, + 485 + ], + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 485 + ], + "type": "text", + "content": "Table 2. Quantitative comparison on the DIOR-H test set, with detection performance evaluated using a pre-trained YOLOv8 [37]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 332, + 552, + 397 + ], + "lines": [ + { + "bbox": [ + 319, + 332, + 552, + 397 + ], + "spans": [ + { + "bbox": [ + 319, + 332, + 552, + 397 + ], + "type": "table", + "html": "
MethodExDARK [18]RUOD [6]
mAP↑AP50↑AP75↑mAP↑AP50↑AP75↑
Oracle38.367.839.858.485.564.8
MIGC [54]28.659.224.121.143.917.9
CC-Diff [52]31.361.828.829.757.828.0
FICGen (ours)38.568.539.537.167.136.7
", + "image_path": "09ea972b3f0f4b27d736a05ffb1d8c47d1a244c7bb47d87dafe4828ad6050061.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 406, + 555, + 428 + ], + "lines": [ + { + "bbox": [ + 313, + 406, + 555, + 428 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 555, + 428 + ], + "type": "text", + "content": "Table 3. Quantitative comparison on ExDARK and RUOD test set, evaluated by a pre-trained Deformable-DETR [55]." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 449, + 555, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 593 + ], + "type": "text", + "content": "images, with the results evaluated against corresponding real annotations. Notably, our FICGen successfully outperforms the existing MIGC [54] and CC-Diff [52] in terms of both FID and AP metrics, achieving remarkable FID scores of 25.10 and 42.40 on the RUOD and ExDARK, respectively. Moreover, for remote sensing object detection, FICGen reaches " + }, + { + "bbox": [ + 313, + 449, + 555, + 593 + ], + "type": "inline_equation", + "content": "82\\%" + }, + { + "bbox": [ + 313, + 449, + 555, + 593 + ], + "type": "text", + "content": " of the mAP relative to real data (i.e., 27.6 vs. 33.4), substantially outperforming MIGC (21.8) and CC-Diff (23.6) by a large margin. Intriguingly, FICGen even exceeds the real ExDARK test set in AP, highlighting its ability to generate degraded instances that precisely adhere to layouts while bridging the real-to-synthetic domain gap." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 594, + 556, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 665 + ], + "type": "text", + "content": "To further verify the generative applicability across different detectors, we evaluate FICGen using YOLOv8 [37] and Deformable-DETR [55], as shown in Tabs. 2 and 3. FICGen surpasses AeroGen by 1.4 and 3.0 in mAP and " + }, + { + "bbox": [ + 313, + 594, + 556, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{75}" + }, + { + "bbox": [ + 313, + 594, + 556, + 665 + ], + "type": "text", + "content": " on the DIOR-H, and outperforms CC-Diff by 7.2 and 7.4 on the ExDARK and RUOD, respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "content": "**Trainability.** Following the trainability protocol in [2], we use augmented annotations as layout inputs to generate synthetic training images and mix them with real ones, effectively doubling the training set. As shown in the last" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19102" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 553, + 301 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 301 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 301 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 301 + ], + "type": "table", + "html": "
DetectorBackboneTraining SetObject Detection (AP) for Sampled Classes /%mAP\\( AP_{50} \\)\\( AP_{75} \\)
# Real# Syntheticbicyclebusmotorbikeboatcarcupbottledog
Train with Full Real Data
RetinaNetPVT-MR:5.1k-44.858.231.331.136.025.528.747.835.265.433.9
Faster R-CNNX101R:5.1k-49.060.835.630.139.530.934.149.338.669.339.6
Train with Pure Synthetic Data
RetinaNetPVT-M-MIGC (S: 5.1k)32.846.720.419.922.514.116.033.423.549.019.1
Faster R-CNNX101-MIGC (S: 5.1k)33.648.218.520.423.117.021.133.224.749.422.1
RetinaNetPVT-M-CC-Diff (S: 5.1k)34.345.521.620.222.015.416.634.924.149.920.6
Faster R-CNNX101-CC-Diff (S: 5.1k)38.047.521.519.124.318.321.833.825.552.321.7
RetinaNetPVT-M-FICGen (S: 5.1k)35.150.423.220.024.916.216.833.325.351.022.1
Faster R-CNNX101-FICGen (S: 5.1k)38.149.921.722.024.318.920.436.426.751.425.1
Train with Synthetic Data and Finetune on Real Data
RetinaNetPVT-MR:5.1kMIGC (S: 5.1k)47.362.134.732.637.427.629.649.037.167.536.9
Faster R-CNNX101R:5.1kMIGC (S: 5.1k)51.361.838.031.540.032.135.849.739.970.740.9
RetinaNetPVT-MR:5.1kCC-Diff (S: 5.1k)47.960.934.331.536.427.130.548.137.067.035.8
Faster R-CNNX101R:5.1kCC-Diff (S: 5.1k)50.262.238.731.140.632.035.449.439.870.440.7
RetinaNetPVT-MR:5.1kFICGen (S: 5.1k)47.562.235.733.737.827.630.947.637.667.736.8
Faster R-CNNX101R:5.1kFICGen(S: 5.1k)51.462.938.232.840.333.135.649.940.670.942.1
Train with Real & Synthetic Data
Faster R-CNNX101R:5.1kMIGC (S: 5.1k)51.763.036.231.739.731.835.249.639.569.739.8
Faster R-CNNX101R:5.1kCC-Diff (S: 5.1k)50.961.636.531.840.231.634.750.339.570.040.7
Faster R-CNNX101R:5.1kFICGen (S: 5.1k)50.561.838.631.640.432.636.549.140.170.541.9
", + "image_path": "3664b6e34240d910051098c1baf49187167e7bf9b163af3a8f8e39767c9d6e7b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 343, + 553, + 401 + ], + "blocks": [ + { + "bbox": [ + 55, + 310, + 555, + 333 + ], + "lines": [ + { + "bbox": [ + 55, + 310, + 555, + 333 + ], + "spans": [ + { + "bbox": [ + 55, + 310, + 555, + 333 + ], + "type": "text", + "content": "Table 4. Quantitative comparison of trainability on the ExDARK dataset, evaluated by RetinaNet [17] (PVT-M [39]) and Faster R-CNN [29] (X101 [43]) under three training settings." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 343, + 553, + 401 + ], + "lines": [ + { + "bbox": [ + 58, + 343, + 553, + 401 + ], + "spans": [ + { + "bbox": [ + 58, + 343, + 553, + 401 + ], + "type": "table", + "html": "
MethodDIOR-H [14]DAWN [13]ExDARK [18]VOC 2012 [5]
airport ↑mAP ↑\\( AP_{50} \\) ↑\\( AP_{75} \\) ↑bus ↑mAP ↑\\( AP_{50} \\) ↑\\( AP_{75} \\) ↑motorbike ↑mAP ↑\\( AP_{50} \\) ↑\\( AP_{75} \\) ↑boat ↑mAP ↑\\( AP_{50} \\) ↑\\( AP_{75} \\) ↑
Oracle32.233.455.635.021.824.846.023.033.335.867.434.542.948.376.852.5
MIGC [54]36.934.156.136.121.624.943.324.634.437.167.836.642.549.477.254.2
CC-Diff [52]37.634.356.136.326.225.846.125.834.137.067.736.644.449.677.354.3
FICGen (ours)38.135.556.837.627.525.944.526.537.137.768.037.747.350.577.455.4
", + "image_path": "306c94d8f2f1053bde67494f9736e0b6c8f30ab31e27003cab907a28e5782771.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 410, + 555, + 432 + ], + "lines": [ + { + "bbox": [ + 55, + 410, + 555, + 432 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 555, + 432 + ], + "type": "text", + "content": "Table 5. Quantitative comparison of trainability on the DIOR, DAWN and ExDARK and natural VOC 2012 datasets, with detection evaluated using the Faster R-CNN (R50) object detector." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 57, + 452, + 295, + 610 + ], + "blocks": [ + { + "bbox": [ + 57, + 452, + 295, + 610 + ], + "lines": [ + { + "bbox": [ + 57, + 452, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 57, + 452, + 295, + 610 + ], + "type": "image", + "image_path": "b0e26b309c3cc17a7a07893e74749b15db0dac1b8c80d21cb496ac653965b172.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 619, + 295, + 642 + ], + "lines": [ + { + "bbox": [ + 55, + 619, + 295, + 642 + ], + "spans": [ + { + "bbox": [ + 55, + 619, + 295, + 642 + ], + "type": "text", + "content": "Figure 5. Visualization results generated by our FICGen on unseen geometric layouts augmented via flipping and translation." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 297, + 715 + ], + "type": "text", + "content": "three rows of Tab. 4, we demonstrate for the first time that synthetic degraded images can significantly enhance the detection performance of downstream object detectors (i.e., +2.4 for bottle and +3.0 for motorbike). Moreover," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 453, + 555, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 555, + 536 + ], + "type": "text", + "content": "we observe that training with synthetic images followed by fine-tuning on real ones yields the best performance. Tab. 5 further provides a comparative analysis under various degraded conditions. Crucially, FICGen consistently achieves the highest gains, boosting AP by 5.7 for \"bus\" (DAWN) and 3.8 for \"motorbike\" (ExDARK), highlighting its efficacy in mitigating data scarcity in degraded scenarios." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 537, + 556, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 537, + 556, + 633 + ], + "spans": [ + { + "bbox": [ + 313, + 537, + 556, + 633 + ], + "type": "text", + "content": "Generalization. To evaluate the generalization of our FICGen, we conduct further experiments not only in severely degraded scenarios such as low-light and underwater conditions with horizontal layout control, but also in blurred and natural VOC 2012 [5], and oriented DIOR-R [3]. Extensive results demonstrate that synthetic images generated by FICGen exhibit superior quality and alignment. Further details are provided in the Supplementary Material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 639, + 443, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 443, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 443, + 651 + ], + "type": "text", + "content": "4.2.2. Qualitative Evaluations" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "Fig. 4 presents a visual comparison of degraded images generated by various L2I methods. Our FICGen demonstrates superior performance in two key aspects: (a) Enhanced Layout Coherence. It resolves the dilemma of \"object omission and merging\", accurately aligning generated instances with" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19103" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 299, + 159 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 299, + 159 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 299, + 159 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 299, + 159 + ], + "type": "image", + "image_path": "acffbba831add96592c62aec3840f73e866dc205ffd60a2392745ad6d05205be.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 169, + 297, + 192 + ], + "lines": [ + { + "bbox": [ + 55, + 169, + 297, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 169, + 297, + 192 + ], + "type": "text", + "content": "Figure 6. Performance convergence curves when generated degraded images are incorporated into downstream detection." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 204, + 299, + 469 + ], + "blocks": [ + { + "bbox": [ + 58, + 204, + 299, + 469 + ], + "lines": [ + { + "bbox": [ + 58, + 204, + 299, + 469 + ], + "spans": [ + { + "bbox": [ + 58, + 204, + 299, + 469 + ], + "type": "image", + "image_path": "e1c385a64c31b40b6b6a48e09da037633704c5910690148ca4fecd8708760b45.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 477, + 295, + 499 + ], + "lines": [ + { + "bbox": [ + 55, + 477, + 295, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 477, + 295, + 499 + ], + "type": "text", + "content": "Figure 7. Generalization results of FICGen on mildly blurred and natural scenes, as well as under oriented layout control." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 522, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 522, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 522, + 297, + 715 + ], + "type": "text", + "content": "the provided layouts in terms of quantity, scale, and position (i.e., aerial tennis courts and rainy cars), and effectively handling small objects (i.e., aerial vehicles). (b) Strong Occlusion Handling. In heavily occluded scenarios (i.e., underwater corals and fish), FICGen reconstructs dense instances while maintaining spatial consistency, whereas MIGC and CC-Diff fail to enforce overlapped layout constraints. Furthermore, Fig. 5 shows that FICGen generalizes effectively to unseen layouts, including those extending beyond image boundaries. As illustrated in Fig. 6, the synthesized degraded images generated by FICGen enhance downstream detection performance by improving mAP and accelerating training convergence. Finally, Fig. 7 highlights the generalization capability of FICGen in blurred and natural scenes, as well as under oriented layout controls. Additional visualizations are provided in the Supplementary Material." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 321, + 71, + 553, + 147 + ], + "blocks": [ + { + "bbox": [ + 321, + 71, + 553, + 147 + ], + "lines": [ + { + "bbox": [ + 321, + 71, + 553, + 147 + ], + "spans": [ + { + "bbox": [ + 321, + 71, + 553, + 147 + ], + "type": "table", + "html": "
FPRVFEAASFAmAP\\( AP_{50} \\)\\( AP_{75} \\)
XXX15.734.812.0
X✓X23.244.622.7
✓X✓18.536.313.9
✓✓X25.546.325.1
X✓✓25.646.925.3
✓✓✓27.648.727.6
", + "image_path": "35cee3ec536b66b30238cc3ba224b8fdd6b62a79f669b641d55a94c8bd1e63e7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 154, + 555, + 176 + ], + "lines": [ + { + "bbox": [ + 313, + 154, + 555, + 176 + ], + "spans": [ + { + "bbox": [ + 313, + 154, + 555, + 176 + ], + "type": "text", + "content": "Table 6. Ablation study of FPR, VFEA, and ASFA modules, evaluated by the pre-trained Faster R-CNN (R50) on the DIOR test set." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 203, + 415, + 215 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 203, + 415, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 203, + 415, + 215 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 223, + 555, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 223, + 555, + 307 + ], + "spans": [ + { + "bbox": [ + 313, + 223, + 555, + 307 + ], + "type": "text", + "content": "Frequency Perceiver Resamplers. The foundation of frequency prototype learning lies in the FPR module, which combines a dual-query mechanism with contextual frequency knowledge. As shown in the last two rows of Tab. 6, FPR significantly improves the alignment scores from 25.6/46.9/25.3 to 27.6/48.7/27.6, demonstrating the effectiveness of contextual frequency guidance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 310, + 556, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 310, + 556, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 310, + 556, + 393 + ], + "type": "text", + "content": "Visual-Frequency Enhanced Attention. VFEA serves as a bridge to inject frequency knowledge (i.e., boundaries and textures) into the latent space while integrating the instance coherence map for contextual disentanglement. As shown in the second row of Tab. 6, even without FPR and ASFA, VFEA achieves competitive performance over the baseline, reaching 23.2/44.6/22.7 in AP." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 396, + 556, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 556, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 556, + 492 + ], + "type": "text", + "content": "Adaptive Spatial-Frequency Aggregation. ASFA is designed to reconstruct refined degraded representations by integrating spatial and frequency perspectives. As evidenced by the results in the fourth and sixth rows of Tab. 6, incorporating ASFA improves the average precision (AP) from 25.5/46.3/25.1 to 27.6/48.7/27.6. This substantial gain underscores its critical role in consolidating contextual information and enhancing the fidelity of degraded image generation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 511, + 479, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 511, + 479, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 511, + 479, + 523 + ], + "type": "text", + "content": "5. Conclusion and Future Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "content": "In this paper, we introduce the task of degraded image generation and define the contextual illusion dilemma as its core challenge. To address this, we propose a frequency-inspired generative paradigm, FICGen. We first extract frequency prototypes from degraded contexts and employ a learnable dual-query mechanism to mine contextual frequency knowledge, which is then transferred into the latent diffusion space via a visual-frequency enhanced attention, coupled with an instance coherence map and a spatial-frequency aggregation module, to disentangle foreground instances from their surroundings and consolidate them into refined degraded representations. Extensive experiments on 5 degraded image datasets validate the effectiveness of FICGen. In future work, we plan to extend this frequency-inspired paradigm to 3D content generation for broader applications." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19104" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 150 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 150 + ], + "type": "text", + "content": "This work is partially supported by grants from the National Natural Science Foundation of China (No. 62132002 and No. 62202010), Beijing Nova Program (No.20250484786), and the Fundamental Research Funds for the Central Universities." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 160, + 115, + 173 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 160, + 115, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 160, + 115, + 173 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 180, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 180, + 296, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 180, + 296, + 224 + ], + "spans": [ + { + "bbox": [ + 61, + 180, + 296, + 224 + ], + "type": "text", + "content": "[1] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6154-6162, 2018. 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 225, + 296, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 225, + 296, + 280 + ], + "spans": [ + { + "bbox": [ + 61, + 225, + 296, + 280 + ], + "type": "text", + "content": "[2] Kai Chen, Enze Xie, Zhe Chen, Yibo Wang, Lanqing Hong, Zhenguo Li, and Dit-Yan Yeung. Geodiffusion: Text-prompted geometric control for object detection data generation. arXiv preprint arXiv:2306.04607, 2023. 3, 4, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 281, + 295, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 281, + 295, + 324 + ], + "spans": [ + { + "bbox": [ + 61, + 281, + 295, + 324 + ], + "type": "text", + "content": "[3] Gong Cheng, Jiabao Wang, Ke Li, Xingxing Xie, Chunbo Lang, Yanqing Yao, and Junwei Han. Anchor-free oriented proposal generator for object detection. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2022. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 325, + 296, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 325, + 296, + 369 + ], + "spans": [ + { + "bbox": [ + 62, + 325, + 296, + 369 + ], + "type": "text", + "content": "[4] Jiaxin Cheng, Xiao Liang, Xingjian Shi, Tong He, Tianjun Xiao, and Mu Li. Layoutdiffuse: Adapting foundational diffusion models for layout-to-image generation. arXiv preprint arXiv:2302.08908, 2023. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 370, + 296, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 370, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 370, + 296, + 413 + ], + "type": "text", + "content": "[5] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 2, 5, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 415, + 296, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 296, + 458 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 296, + 458 + ], + "type": "text", + "content": "[6] Chenping Fu, Risheng Liu, Xin Fan, Puyang Chen, Hao Fu, Wanqi Yuan, Ming Zhu, and Zhongxuan Luo. Rethinking general underwater object detection: Datasets, challenges, and solutions. Neurocomputing, 517:243-256, 2023. 2, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 459, + 296, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 459, + 296, + 512 + ], + "spans": [ + { + "bbox": [ + 62, + 459, + 296, + 512 + ], + "type": "text", + "content": "[7] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. In European Conference on Computer Vision, pages 89-106. Springer, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 514, + 296, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 296, + 557 + ], + "type": "text", + "content": "[8] Yuchao Gu, Yipin Zhou, Yunfan Ye, Yixin Nie, Licheng Yu, Pingchuan Ma, Kevin Qinghong Lin, and Mike Zheng Shou. Roictrl: Boosting instance control for visual generation. arXiv preprint arXiv:2411.17949, 2024. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 558, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 558, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 558, + 296, + 612 + ], + "type": "text", + "content": "[9] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 613, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 295, + 635 + ], + "type": "text", + "content": "[10] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 636, + 296, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 296, + 669 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 296, + 669 + ], + "type": "text", + "content": "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 712 + ], + "type": "text", + "content": "[12] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in neural information processing systems, 35:26565-26577, 2022. 5" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 106 + ], + "type": "text", + "content": "[13] Mourad A Kenk and Mahmoud Hassaballah. Dawn: vehicle detection in adverse weather nature dataset. arXiv preprint arXiv:2008.05402, 2020. 2, 5, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 108, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 555, + 150 + ], + "type": "text", + "content": "[14] Ke Li, Gang Wan, Gong Cheng, Liqui Meng, and Junwei Han. Object detection in optical remote sensing images: A survey and a new benchmark. ISPRS journal of photogrammetry and remote sensing, 159:296-307, 2020. 1, 2, 5, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 152, + 555, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 555, + 208 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 555, + 208 + ], + "type": "text", + "content": "[15] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22511-22521, 2023. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 209, + 554, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 554, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 554, + 274 + ], + "type": "text", + "content": "[16] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer vision-ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part v 13, pages 740-755. Springer, 2014. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 276, + 555, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 555, + 319 + ], + "type": "text", + "content": "[17] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 321, + 555, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 555, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 555, + 364 + ], + "type": "text", + "content": "[18] Yuen Peng Loh and Chee Seng Chan. Getting to know low-light images with the exclusively dark dataset. Computer Vision and Image Understanding, 178:30-42, 2019. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 365, + 554, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 554, + 388 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 554, + 388 + ], + "type": "text", + "content": "[19] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 389, + 555, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 389, + 555, + 444 + ], + "spans": [ + { + "bbox": [ + 317, + 389, + 555, + 444 + ], + "type": "text", + "content": "[20] Zhengyao Lv, Yuxiang Wei, Wangmeng Zuo, and KwanYee K Wong. Place: Adaptive layout-semantic fusion for semantic image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9264–9274, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 445, + 555, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 555, + 488 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 555, + 488 + ], + "type": "text", + "content": "[21] Mingcan Ma, Changqun Xia, Chenxi Xie, Xiaowu Chen, and Jia Li. Boosting broader receptive fields for salient object detection. IEEE Transactions on Image Processing, 32:1026-1038, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 491, + 554, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 491, + 554, + 545 + ], + "spans": [ + { + "bbox": [ + 317, + 491, + 554, + 545 + ], + "type": "text", + "content": "[22] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI conference on artificial intelligence, pages 4296–4304, 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 547, + 554, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 547, + 554, + 600 + ], + "spans": [ + { + "bbox": [ + 317, + 547, + 554, + 600 + ], + "type": "text", + "content": "[23] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 602, + 555, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 555, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 555, + 646 + ], + "type": "text", + "content": "[24] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 647, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 555, + 712 + ], + "type": "text", + "content": "[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 4" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19105" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 126 + ], + "type": "text", + "content": "[26] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International conference on machine learning, pages 8821-8831. Pmlr, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 296, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 296, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 296, + 172 + ], + "type": "text", + "content": "[27] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2): 3, 2022. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 296, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 296, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 296, + 218 + ], + "type": "text", + "content": "[28] Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In International conference on machine learning, pages 1060-1069. PMLR, 2016. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 296, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 296, + 262 + ], + "type": "text", + "content": "[29] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 1, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 296, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 296, + 319 + ], + "type": "text", + "content": "[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 320, + 296, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 296, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 296, + 386 + ], + "type": "text", + "content": "[31] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 296, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 296, + 453 + ], + "type": "text", + "content": "[32] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 454, + 296, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 454, + 296, + 508 + ], + "spans": [ + { + "bbox": [ + 56, + 454, + 296, + 508 + ], + "type": "text", + "content": "[33] Zhe Shan, Yang Liu, Lei Zhou, Cheng Yan, Heng Wang, and Xia Xie. Ros-sam: High-quality interactive segmentation for remote sensing moving object. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 3625-3635, 2025. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 510, + 296, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 510, + 296, + 541 + ], + "spans": [ + { + "bbox": [ + 56, + 510, + 296, + 541 + ], + "type": "text", + "content": "[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 544, + 296, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 296, + 586 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 296, + 586 + ], + "type": "text", + "content": "[35] Wei Sun and Tianfu Wu. Image synthesis from reconfigurable layout and style. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10531-10540, 2019. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 590, + 296, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 296, + 643 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 296, + 643 + ], + "type": "text", + "content": "[36] Datalog Tang, Xiangyong Cao, Xuan Wu, Jialin Li, Jing Yao, Xueru Bai, Dongsheng Jiang, Yin Li, and Deyu Meng. Aerogen: enhancing remote sensing object detection with diffusion-driven data generation. arXiv preprint arXiv:2411.15497, 2024. 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 646, + 296, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 646, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 56, + 646, + 296, + 666 + ], + "type": "text", + "content": "[37] Ultralytics. Ultralytics yolov8. https://github.com/ultralytics/ultralytics, 2023.6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 669, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 296, + 689 + ], + "type": "text", + "content": "[38] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 692, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 296, + 713 + ], + "type": "text", + "content": "[39] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyra-" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 116 + ], + "type": "text", + "content": "mid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 568-578, 2021. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 555, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 555, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 555, + 173 + ], + "type": "text", + "content": "[40] Wenzhuang Wang, Mingcan Ma, Yong Chen, Changqun Xia, Zhenbao Liang, and Jia Li. Freegen: Bridging visual-linguistic discrepancies towards diffusion-based pixel-level data synthesis. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7916–7924, 2025. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 175, + 555, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 555, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 555, + 229 + ], + "type": "text", + "content": "[41] Xudong Wang, Trevor Darrell, Sai Saketh Rambhatla, Rohit Girdhar, and Ishan Misra. Instancediffusion: Instance-level control for image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6232-6242, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 231, + 555, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 555, + 274 + ], + "type": "text", + "content": "[42] Yinwei Wu, Xianpan Zhou, Bing Ma, Xuefeng Su, Kai Ma, and Xinchao Wang. Ifadapter: Instance feature control for grounded text-to-image generation. arXiv preprint arXiv:2409.08240, 2024. 3, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 276, + 555, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 555, + 330 + ], + "type": "text", + "content": "[43] Saining Xie, Ross Girshick, Piotr Dólár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1492–1500, 2017. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 332, + 555, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 555, + 396 + ], + "type": "text", + "content": "[44] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Finegrained text to image generation with attentional generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1316-1324, 2018. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 399, + 555, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 555, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 555, + 443 + ], + "type": "text", + "content": "[45] Han Xue, Zhiwu Huang, Qianru Sun, Li Song, and Wenjun Zhang. Freestyle layout-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14256-14266, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 445, + 555, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 555, + 509 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 555, + 509 + ], + "type": "text", + "content": "[46] Zhengyuan Yang, Jianfeng Wang, Zhe Gan, Linjie Li, Kevin Lin, Chenfei Wu, Nan Duan, Zicheng Liu, Ce Liu, Michael Zeng, et al. Reco: Region-controlled text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14246-14255, 2023. 2, 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 512, + 555, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 555, + 555 + ], + "type": "text", + "content": "[47] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 557, + 555, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 555, + 611 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 555, + 611 + ], + "type": "text", + "content": "[48] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 613, + 555, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 555, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 555, + 678 + ], + "type": "text", + "content": "[49] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 5907-5915, 2017. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 681, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 555, + 713 + ], + "type": "text", + "content": "[50] Han Zhang, Jing Yu Koh, Jason Baldridge, Honglak Lee, and Yinfei Yang. Cross-modal contrastive learning for text-to-image generation. In Proceedings of the IEEE/CVF con" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19106" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 341 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 75, + 73, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 73, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 75, + 73, + 294, + 94 + ], + "type": "text", + "content": "ference on computer vision and pattern recognition, pages 833-842, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "text", + "content": "[51] Jingyang Zhang, Shiwei Li, Yuanxun Lu, Tian Fang, David McKinnon, Yanghai Tsin, Long Quan, and Yao Yao. Joint-net: Extending text-to-image diffusion for dense distribution modeling. arXiv preprint arXiv:2310.06347, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 296, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 296, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 296, + 184 + ], + "type": "text", + "content": "[52] Mu Zhang, Yunfan Liu, Yue Liu, Hongtian Yu, and Qixiang Ye. Cc-diff: Enhancing contextual coherence in remote sensing image synthesis. arXiv preprint arXiv:2412.08464, 2024. 3, 5, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 296, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 296, + 239 + ], + "type": "text", + "content": "[53] Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22490-22499, 2023. 2, 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 296, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 296, + 295 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 296, + 295 + ], + "type": "text", + "content": "[54] Dewei Zhou, You Li, Fan Ma, Xiaoting Zhang, and Yi Yang. Mige: Multi-instance generation controller for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6818-6828, 2024. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 296, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 296, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 296, + 341 + ], + "type": "text", + "content": "[55] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 6" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19107" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/FIND_ Few-Shot Anomaly Inspection with Normal-Only Multi-Modal Data/94dd2e05-5a4d-4859-bd44-9496d30c1756_content_list.json b/2025/FIND_ Few-Shot Anomaly Inspection with Normal-Only Multi-Modal Data/94dd2e05-5a4d-4859-bd44-9496d30c1756_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4014227926a7edb24c96e944e2611c22c6858a88 --- /dev/null +++ b/2025/FIND_ Few-Shot Anomaly Inspection with Normal-Only Multi-Modal Data/94dd2e05-5a4d-4859-bd44-9496d30c1756_content_list.json @@ -0,0 +1,1290 @@ +[ + { + "type": "text", + "text": "FIND: Few-Shot Anomaly Inspection with Normal-Only Multi-Modal Data", + "text_level": 1, + "bbox": [ + 119, + 130, + 875, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yiting Li $^{1}$ , Fayao Liu $^{1\\dagger}$ , Jingyi Liao $^{1}$ , Sichao Tian $^{2}$ , Chuan-Sheng Foo $^{1}$ , Xulei Yang $^{1\\dagger}$ $^{1}$ Institute for Infocomm Research, A*STAR, Singapore $^{2}$ Chinese Academy of Medical Sciences and Peking Union Medical College, China \n{li_yiting, liu_fayao, liao_jingyi, foo_chuan_sheng, yang_xulei} @i2r.a-star.edu.sg, {sctian} @implad.ac.cn", + "bbox": [ + 101, + 179, + 931, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multi-modal anomaly detection (MAD) improves industrial inspection by exploiting complementary 2D and 3D data. However, existing methods struggle in few-shot scenarios due to limited data and modality gaps. Current approaches either fuse multi-modal features or align cross-modal representations; however, they often suffer from high false-positive rates and fail to detect subtle defects, especially when training samples are scarce. To address these challenges, we propose the first few-shot MAD method FIND, a novel dual-student framework that integrates intramodal reverse distillation and cross-modal feature mapping. FIND employs modality-specific teachers and two collaborative students: an intra-modal student for fine-grained anomaly localization via reverse distillation, and a cross-modal student that captures inter-modal correspondences to detect cross-modal inconsistencies. Extensive experiments on MVTec-3D-AD and Eyecandies show that FIND significantly outperforms state-of-the-art methods in both full-shot and few-shot settings. Ablation studies validate the complementary roles of intra- and cross-modal distillation. Our work significantly advances MAD robustness in data-scarce industrial applications.", + "bbox": [ + 88, + 316, + 485, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 681, + 220, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multi-modal anomaly detection (MAD) has garnered increasing interest in industrial inspection [4, 17, 18, 20, 21, 26, 31, 34, 40], as relying on a single modality (e.g., RGB) may fail to capture crucial geometric cues under varying surface textures. By integrating heterogeneous data such as point clouds or depth maps, MAD methods can leverage multi-modal complementarity to enhance inspection quality. Since anomalous data are rare, most existing anomaly detection (AD) methods are trained only on normal data. Our work follows this protocol.", + "bbox": [ + 89, + 707, + 482, + 858 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While significant progress has been made in MAD, exist", + "bbox": [ + 109, + 859, + 482, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f0cbbd1f2cc09eb4ec229d1a47d39b1ce2c60b7a00d469d6a98d1fa8e28946f4.jpg", + "image_caption": [ + "Figure 1. Comparison with state-of-the-art methods on MVTec-3D-AD in terms of I-AUROC (left) and AUPRO@1% (right). FIND substantially outperforms prior methods (BTF [11], AST [24], M3DM [31], 3D-ADNAS [21] and CFM [5]) in both few-shot and full-shot settings." + ], + "image_footnote": [], + "bbox": [ + 517, + 286, + 712, + 412 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/79b2b557084b07c5b10597bbffcd215dc41f8a5d4e9b7b5ef38b7b4ef10a39c7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 714, + 286, + 901, + 412 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing approaches struggle to retain robustness in few-shot scenarios, where limited normal multi-modal training data is available—a critical limitation given the high cost of acquiring such data [8, 12, 14, 16, 27]. This is evidenced in Fig. 1, which compares various methods using different amounts of normal training samples, with notable performance drops observed. In this work, we propose FIND, a simple yet effective dual-student MAD framework focused on Few-Shot Anomaly Inspection with Normal-Only Multi-Modal Data. It effectively minimizes modality interference while exploiting multi-modal complementarity, significantly enhancing anomaly detection and localization.", + "bbox": [ + 509, + 522, + 906, + 703 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In single-modality anomaly detection, teacher-student (T-S) frameworks achieve strong results [6, 23, 28, 38]. A pretrained teacher extracts generic features, while a student trained on normal data mimics the teacher. Anomalies emerge as discrepancies between their features at inference. Reverse distillation [7, 28] refines this by reconstructing teacher features from compact one-class embeddings, enabling fine-grained defect localization.", + "bbox": [ + 509, + 704, + 908, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, extending the success of T-S frameworks to MAD remains a challenge. Existing methods either fuse modalities by concatenating RGB and depth at the input level [24] or across multiple feature levels [9], which risks modality interference. Specifically, MMRD [9] introduces", + "bbox": [ + 511, + 825, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding authors.", + "bbox": [ + 109, + 886, + 241, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23290", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a multi-modal reverse distillation paradigm that performs a weighted fusion of image and depth data across multiple feature levels, creating a multi-modal teacher target for a learnable student decoder tasked with restoring the teacher's multi-modal representations. In contrast, cross-modal Feature Mapping (CFM) [5], although not explicitly formulated within the T-S framework, uses domain-specific teachers for cross-modal distillation and performs late fusion on score maps instead of relying on feature fusion. The features of different modalities are aligned through an MLP-based projection network. CFM trains two student models to map features from one modality to another by mimicking the corresponding target modality teacher. By learning the expected correspondence between an object's color and shape, any deviation from this correspondence at test time indicates an anomaly. However, the inherent modality gap can lead to large reconstruction errors even for normal samples, resulting in high false positive rates and limiting the model's ability to distinguish genuine anomalies. These limitations result in high false positives and coarse localization (see Fig. 2 for examples), which are further exacerbated in few-shot scenarios, where limited normal samples hinder the effective alignment of the two modalities.", + "bbox": [ + 89, + 90, + 483, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the challenges stated above, we present FIND, which unifies intra-modal reverse distillation and cross-modal distillation within a single T-S framework. Our intuition is to minimize modality interference while exploiting the complementary strengths of multiple modalities. Specifically, each modality is paired with a modality-specific teacher, an intra-modal student, and a cross-modal student. The intra-modal student is trained via reverse distillation to model fine-grained, modality-specific normal appearances, while the cross-modal student is designed to capture inter-modal correspondences for robust anomaly detection. This decoupled design allows each student to focus on its specific task: the intra-modal student is highly sensitive to prominent modality-specific anomalies and provides finer-grained localization, while the cross-modal student detects anomalies that disrupt the learned normal correspondence between different modalities. This synergy enables precise localization and robustness to various types of anomalies, even with limited training data (see Fig. 2 for comparisons). To further mitigate the modality gap and thus reduce false positives in cross-modal distillation, we propose using surface normals as a geometric proxy, enabling the use of shared RGB and 3D encoders. We find this simple strategy to be highly effective in achieving accurate MAD. Our contributions are summarized as follows:", + "bbox": [ + 91, + 440, + 483, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We propose the first few-shot MAD method FIND, which synergistically integrates intra-modal reverse distillation and cross-modal distillation. By decoupling the learning objectives for intra- and cross-modal students, FIND significantly enhances MAD performance.", + "bbox": [ + 101, + 824, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ff2fcea15e646f283f76ecb9b49af151a96db198ed4a9fbdd816ba8cd45acf7b.jpg", + "image_caption": [ + "Figure 2. Compared to CFM under the 10-shot setting, our anomaly maps are sharper and show fewer false positives, focusing on anomalous patterns rather than minor texture variations." + ], + "image_footnote": [], + "bbox": [ + 524, + 88, + 905, + 305 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We show that surface normals are a highly effective geometric modality for MAD, effectively bridging the modality gap in cross-modal distillation, reducing false positives, and enhancing anomaly detection precision.", + "- FIND sets a new state-of-the-art in both multi-modal anomaly detection and localization tasks, outperforming existing methods in both full-shot and few-shot settings." + ], + "bbox": [ + 524, + 388, + 903, + 494 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related works", + "text_level": 1, + "bbox": [ + 513, + 511, + 658, + 526 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-modal Anomaly Detection With the development of 3D sensors, multi-modal anomaly detection has gained increasing attention [4, 9, 21, 36], as 3D information (e.g., point clouds) offers richer geometric cues and captures shape deformations that may be invisible in RGB domain. Recently, researchers have explored cross-modal alignment methods that fuse RGB and 3D data [24, 31] for more robust anomaly localization. For instance, CFM [5] proposes a novel framework that learns to map 2D and 3D features to each other, thereby pinpointing anomalies where the two modalities disagree; [29] introduce a local-to-global self-supervised adaptation mechanism to finetune pre-trained 2D/3D networks for more precise cross-modal alignment; [31] design a hybrid fusion scheme with multiple memory banks to achieve consistent performance gains over single-modality baselines. In parallel, memory-based 3D AD approaches [21, 31] have also shown strong potential by storing normal embeddings from both color and geometry in dedicated banks and measuring anomalous deviations at test time. For example, [4] propose a shape-guided dual-memory method that leverages signed distance fields (SDF) and RGB cues for complementary anomaly localization; [18] advocate an internal spatial modality perception module to distill intrinsic 3D structures into a memory", + "bbox": [ + 511, + 537, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23291", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "bank; [26] handle incomplete multi-modal data via cross-modal distillation and memory retrieval. Although existing methods have significantly advanced the state of the art in MAD, most still rely on either large-scale normal training sets or carefully aligned networks for each modality. In scenarios where only a few normal samples are available, the modality gap between 2D images and 3D point clouds can hinder reliable cross-modal feature alignment, thus limiting the models' ability to detect subtle anomalies.", + "bbox": [ + 89, + 90, + 480, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Few-shot Anomaly Detection In resource-constrained scenarios, collecting large amount of normal data can be both costly and impractical. Consequently, few-shot anomaly detection (FSAD) becomes a promising alternative. Current research on FSAD has primarily focused on RGB domains. AnoPLe [16] proposes bi-directional prompt learning with only normal samples, leveraging a pretrained vision-language model to detect anomalies from both textual and visual cues. FastRecon [8] proposes a lightweight feature-based regression with distribution regularization to obtain the optimal transformation from support to query features, it reconstructs a normal version of test sample's features from visually similar support features and detects any large deviations as anomalies. WinCLIP [14] explores a window-based feature fusion strategy to extract multi-scale features for zero-/few-shot anomaly classification and segmentation. FCOT [27] leverages a distribution-based Conditional Transport metric to minimize feature discrepancies between a few support images and query images, while updating a memory bank to precisely match anomalous regions. RegAD [12] uses image registration to align query images with support images in a category-agnostic manner. Despite remarkable progress, existing FSAD approaches still face limitations when extended to multi-modal scenarios due to the lack of capabilities for capturing the shape distortions that 3D data would reveal. This limitation motivates our work on multi-modal anomaly detection under few-shot settings, by capturing shape-induced defects that 2D-only methods may overlook.", + "bbox": [ + 91, + 231, + 483, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Cross-modal knowledge distillation Cross-modal knowledge distillation (KD) has attracted increasing interest for transferring knowledge between heterogeneous modalities (e.g. images, LiDAR, thermal, audio) [3, 15, 25, 30, 33]. A core challenge is the modality gap issue due to the discrepancies in data/feature distribution between modalities [13, 19, 22]. [19] demonstrates that modality gap in multimodal models such as CLIP is caused by a combination of model initialization and contrastive learning optimization. Recent works tackle this by aligning feature representations and customizing distillation strategies to each modality pair. For example, a LiDAR-based teacher can guide a camera-based student to improve monocular detection accuracy [10, 15, 32]. C2KD [13] propose a bidirectional distillation framework where teacher and student", + "bbox": [ + 89, + 674, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "learn from each other, coupled with on-the-fly sample filtering to discard misaligned targets. Cross-modal KD is applied in RGB-Thermal (RGB-T) tracking [37, 39], a two-stream RGB-T network can serve as a powerful teacher to a lightweight one-stream tracker; distilling both modality-specific and shared features helps the student network approach the teacher's performance, while greatly reducing computation. In audio-visual learning [35], visual models have been used to teach audio models to predict spatial or semantic information from sound, often by enforcing feature alignments to handle the modality gap.", + "bbox": [ + 511, + 90, + 903, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 271, + 604, + 286 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Framework Overview", + "text_level": 1, + "bbox": [ + 511, + 296, + 718, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "FIND consists of two single-teacher-dual-student branches designed to detect anomalies from multi-modal 2D and 3D inputs. Fig. 3 shows an overview of our method. Each branch is composed of a modality-specific teacher $(\\mathbf{T}_{2\\mathrm{D}}$ or $\\mathbf{T}_{3\\mathrm{D}})$ , a cross-modal distillation student $(\\mathbf{S}_{3D\\rightarrow 2D}$ or $\\mathbf{S}_{2D\\rightarrow 3D})$ , and an intra-modal reverse distillation student $(\\mathbf{S}_{2D\\rightarrow 2D}$ or $\\mathbf{S}_{3D\\rightarrow 3D})$ . Our motivation is to integrate cross-modal distillation with reverse distillation in a unified framework that minimizes modality interference while effectively exploiting their complementarity for anomaly detection. In essence, the Cross-Modal Student distills knowledge across the RGB and 3D streams, learning how normal features in one modality correspond to those in the other. The Intra-Modal Student, on the other hand, learns to reproduce the normal pattern of individual modality through reverse distillation. During training, both students receive supervision derived from a shared set of teacher features so that each modality learns a more comprehensive representation.", + "bbox": [ + 511, + 318, + 906, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Cross-modal Teacher-Student Learning", + "text_level": 1, + "bbox": [ + 511, + 614, + 856, + 631 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We aim to leverage the complementary knowledge between two different modalities, namely 2D and 3D, for anomaly detection through a cross-modal distillation procedure.", + "bbox": [ + 511, + 637, + 903, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Bridging the Modality Gap. Since features extracted by 2D backbones for RGB images and 3D point-based networks for point clouds follow different distributions, existing MAD methods face the modality gap issue due to the use of separate feature backbones for each modality. This discrepancy becomes even more significant in few-shot learning scenarios, where a student model with limited training data struggles to align with the teacher's normal feature distribution, leading to a high false positive rate in detection performance. To bridge the modality gap between the teacher and its cross-modal student, we propose extracting surface normal maps from raw point clouds to encode 3D geometry, enabling a shared backbone network used for both 2D and 3D modalities. Our key insight is that sur", + "bbox": [ + 511, + 689, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23292", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5342df6477af3055d6cba544d92ea4c5936d83068d902b863cb94e210ac8dbba.jpg", + "image_caption": [ + "Figure 3. FIND consists of two single-teacher-dual-student branches (left and right) designed to detect anomalies from multi-modal inputs, i.e., 2D RGB and 3D surface-normal. Each branch consists of a modality-specific teacher (center), a cross-modal distillation student (top), and an intra-modal reverse distillation student (bottom). The cross-modal student performs patch-wise mapping to the teacher modality for capturing inter-modality inconsistencies. In parallel, the intra-modal student conducts layer-wise feature inversion back to the original input resolution, enabling pixel-level localization of defects. By combining the results from the two pairs of intra-modal and cross-modal students, our approach robustly identifies various types of anomalies, even in few-shot scenarios." + ], + "image_footnote": [], + "bbox": [ + 98, + 85, + 486, + 333 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fdcec2dc2e27da84955a78ed877e7a2b9baf617194c3a98b20023ff009851cb5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 498, + 85, + 903, + 332 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "face normal maps capture fine-grained surface orientation details and preserve subtle textures more effectively than depth maps. This contrasts with previous work, which uses either raw point clouds or depth maps. Raw point clouds exacerbate the domain gap, while depth images tend to smooth out small shape variations and are less sensitive to small defects. These findings allow us to effectively bridge the modality gap and therefore significantly reduce false positives.", + "bbox": [ + 88, + 454, + 480, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Cross-modal Distillation. Let $\\mathbf{T}_{2D}$ and $\\mathbf{T}_{3D}$ denote the two pre-trained Vision Transformer (ViT) teachers for the 2D RGB images and 3D surface normals, respectively. Inspired by CFM [5], each teacher receives input from its native domain and outputs high-level features $\\mathbf{f}_{2D}^T$ or $\\mathbf{f}_{3D}^T$ during training. In parallel, the corresponding cross-modal distillation student, $\\mathbf{S}_{3D\\rightarrow 2D}$ or $\\mathbf{S}_{2D\\rightarrow 3D}$ , takes features from one modality as input and attempts to reproduce the output of the teacher network from the other modality:", + "bbox": [ + 88, + 597, + 482, + 733 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {2 D} ^ {S} = \\mathbf {S} _ {3 D \\rightarrow 2 D} \\big (\\mathbf {f} _ {3 D} ^ {T} \\big), \\quad \\mathbf {f} _ {3 D} ^ {S} = \\mathbf {S} _ {2 D \\rightarrow 3 D} \\big (\\mathbf {f} _ {2 D} ^ {T} \\big). \\qquad (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 739, + 482, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overall distillation loss is applied bidirectionally:", + "bbox": [ + 109, + 768, + 465, + 784 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c r o s s}} = \\left(1 - \\cos \\langle \\mathbf {f} _ {2 D} ^ {T}, \\mathbf {f} _ {2 D} ^ {S} \\rangle\\right) + \\left(1 - \\cos \\langle \\mathbf {f} _ {3 D} ^ {T}, \\mathbf {f} _ {3 D} ^ {S} \\rangle\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 806, + 482, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Intra-Modal Teacher-Student Learning", + "text_level": 1, + "bbox": [ + 89, + 848, + 431, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While the cross-modal student excels in identifying semantic relationships between modalities, it relies on high-level", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "feature representations with low spatial resolution, which may overlook subtle defects. To address this limitation, we introduce an intra-modal student for each modality, which performs intra-modal reconstruction on both feature level and image level. As a result, the intra-modal student specializes in capturing modality-specific fine-grained details for anomaly detection. In particular, the intra-modal student adopts a decoder-only architecture to perform multi-stage reverse distillation, progressively reconstructing back to the original image space for fine-grained anomaly localization.", + "bbox": [ + 511, + 454, + 906, + 607 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-Layer Reverse Distillation. To preserve both low-level textures and high-level semantics, we adopt a multilayer feature inversion strategy for the training of intramodal student. Let $\\{F_T^k\\}_{k=1}^K$ denote a set of $K$ feature maps selected from different layers of the teacher's encoder, with layer 1 representing the shallow representation and layer $K$ being the deepest representation. In addition, $k = 0$ denotes the original input image space. Our two intra-modal students, $\\mathbf{S}_{2D \\to 2D}$ , $\\mathbf{S}_{3D \\to 3D}$ , are designed to reverse the encoding process in a top-down manner. Specifically, we use $D^k$ to denote the decoding module in the $k$ -th layer of an intra-modal student, with each decoder block $D^k(\\cdot)$ reconstructing the next fine-grained feature map: $F_S^{k-1} = D^k(F_S^k)$ , where $F_S^k$ denotes the output of the $k$ -th level of the intra-modal student and $F_T^K$ serves as the student's starting point. The final output $F_S^0$ reaches the original image resolution, enabling fine-grained anomaly localization.", + "bbox": [ + 511, + 612, + 908, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We then enforce the consistency between the teacher and the student's predictions to ensure that normal appearance is", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23293", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "reconstructed across different hidden layers. In addition, we incorporate an image-level reconstruction term that compares the final output $F_{S}^{0}$ against the original input image $I$ , further preserving fidelity to subtle textures. Concretely, we calculate the intra-modal loss as:", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {i n t r a}} = \\sum_ {(S, T)} \\Big (\\sum_ {k = 1} ^ {K - 1} \\big (1 - \\cos \\langle F _ {S} ^ {k}, F _ {T} ^ {k} \\rangle \\big) + \\left\\| F _ {S} ^ {0} - I \\right\\| ^ {2} \\Big), \\\\ (S, T) \\in \\left\\{\\left(\\mathbf {S} _ {2 D \\rightarrow 2 D}, \\mathbf {T} _ {2 \\mathrm {D}}\\right), \\left(\\mathbf {S} _ {3 D \\rightarrow 3 D}, \\mathbf {T} _ {3 \\mathrm {D}}\\right)\\right\\}. \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 172, + 480, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This encourages the intra-modal students to capture both high-level semantics and fine-grained details of teacher's normality representation, even in few-shot settings.", + "bbox": [ + 89, + 250, + 482, + 296 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training and Inference", + "text_level": 1, + "bbox": [ + 89, + 309, + 305, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training. Our method jointly trains the cross-modal and intra-modal students using the respective objectives introduced in Sections 3.2 and 3.3. Specifically, we minimize the sum of the cross-modal loss in Eq. (2) and the intramodal loss in Eq. (3), yielding the final objective:", + "bbox": [ + 89, + 330, + 483, + 407 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {c r o s s}} + \\mathcal {L} _ {\\text {i n t r a}}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 416, + 480, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training is carried out on a few normal samples. The cross-modal student learns to reconstruct teacher features from the complementary modality, while the intra-modal student focuses on restoring its own modality in a reverse distillation manner. This combined strategy leverages both inter-modal consistency and fine-grained modality-specific cues for robust and fine-grained anomaly detection.", + "bbox": [ + 89, + 441, + 483, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Anomaly Scoring. At test time, each of the four students $\\mathbf{S}_{2D\\rightarrow 3D}$ , $\\mathbf{S}_{3D\\rightarrow 2D}$ , $\\mathbf{S}_{2D\\rightarrow 2D}$ , $\\mathbf{S}_{3D\\rightarrow 3D}$ produces an anomaly map by measuring the discrepancy between its reconstructed output and the corresponding teacher's feature. In particular, for the intra-modal students $(S_{2D\\rightarrow 2D}$ and $S_{3D\\rightarrow 3D})$ , the intra-modal anomaly maps $\\mathbf{A}_{2D\\rightarrow 2D}$ and $\\mathbf{A}_{3D\\rightarrow 3D}$ are calculated by aggregating the intra-modal teacher-student discrepancy across multiple layers. For each hidden layer, we additionally apply $\\mathrm{L}_2$ -normalize to both the teacher's feature and the student's reconstructed feature at every spatial position. For each layer, we calculate the Euclidean distance between the student's reconstructed output and the teacher's original output to obtain a pixel-level error map. Next, we apply Gaussian smoothing to the obtained error map for a more coherent anomaly response. We repeat this process for all $K - 1$ hidden layers and the final output layer, and then average all the error maps to obtain the two intra-modal maps $\\mathbf{A}_{2D\\rightarrow 2D}$ and $\\mathbf{A}_{3D\\rightarrow 3D}$ for both 2D and 3D modalities. The final intra-modal anomaly map $\\mathbf{A}_{\\mathrm{intra}}$ is computed by averaging the two intra-modal maps $\\mathbf{A}_{2D\\rightarrow 2D}$ and $\\mathbf{A}_{3D\\rightarrow 3D}$ .", + "bbox": [ + 89, + 551, + 483, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The cross-modal students $(S_{2D\\to 3D}$ and $S_{3D\\to 2D})$ operate similarly by measuring the discrepancy between the", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "student's reconstructed feature in one modality and the teacher's feature in the other modality. The resulting cross-modal anomaly maps from each student are denoted as $\\mathbf{A}_{2D\\to 3D}$ and $\\mathbf{A}_{3D\\to 2D}$ . Likewise, we average the two cross-modal maps to obtain a single cross-modal anomaly map $\\mathbf{A}_{\\mathrm{cross}}$ . The final overall anomaly map is obtained by the following:", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} _ {\\text {f i n a l}} = \\alpha \\mathbf {A} _ {\\text {c r o s s}} + (1 - \\alpha) \\mathbf {A} _ {\\text {i n t r a}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 205, + 903, + 224 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ adjusts the relative influence of cross-modal and intra-modal anomaly map. High values in $\\mathbf{A}_{\\mathrm{final}}$ indicate regions where the student network fails to mimic the teacher. By combining multi-modal alignment and reverse distillation reconstruction, our framework captures diverse anomaly signatures even in few-shot scenarios.", + "bbox": [ + 511, + 232, + 903, + 324 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 511, + 335, + 638, + 353 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 361, + 702, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets and Protocol. We evaluate on two public benchmarks: MVTec-3D-AD [1] and Eyecandies [2]. Following a standard few-shot normal-only protocol, we limit the available normal training images per class to $k \\in \\{5, 10, 50\\}$ , randomly sampling and averaging results over ten runs. We also evaluate under a full-shot (all normal samples) setting for completeness.", + "bbox": [ + 511, + 383, + 905, + 489 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We compare against: (1) BTF [11], a classical 3D feature-based method without deep learning; (2) AST [24], an asymmetric T-S approach trained on normal samples; (3) M3DM [31], a hybrid fusion model exploiting memory banks; (4) CFM [5], which aligns 2D-3D features without memory banks; and (5) 3D-ADNAS [21], which uses neural architecture search for multi-modal fusion.", + "bbox": [ + 511, + 494, + 905, + 601 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. We follow standard anomaly detection metrics: (a) I-AUROC for image-level anomaly classification, (b) P-AUROC (pixel-level AUROC) for localization, and (c) AUPRO (area under per-region overlap) at specific false positive rates (FPR), typically $1\\%$ and $30\\%$ . Higher values denote better performance.", + "bbox": [ + 511, + 606, + 905, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. The intra-modal student is implemented as a multi-stage decoder, where each stage consists of three sequential convolutional blocks with normalization layers and GeLU activation to ensure nonlinearity. In particular, the final block employs ConvTranspose2d layers to up-sample the feature map back to the original image resolution and progressively decreases the dimensionality of the output channel to the final 3 channels (RGB). We implement our cross-modal student network as an encoder-decoder style network. The encoder first encodes the input patches into a hidden representation with two convolutional blocks, while the decoder reconstructs them in the cross-modal feature space through another convolutional", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23294", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/cc972593557fb6dad684c6e4ea02a058c494a1e3331605ab7b4c09246d345f71.jpg", + "table_caption": [ + "Table 1. Comparison of anomaly detection and localization performance on MVTec-3D-AD and Eyecandies with varying numbers of training shots. $\\mathbf{{ROC}}{}_{I},{\\mathbf{{ROC}}}_{P}$ ,and $\\mathbf{{PRO}}$ denote I-AUROC, P-AUROC, and AUPRO respectively. Bold indicates the best performance." + ], + "table_footnote": [], + "table_body": "
k-ShotMethodMVTec-3D-ADEyecandies
\\( ROC_I \\)\\( ROC_P \\)PRO-30%PRO-1%\\( ROC_I \\)\\( ROC_P \\)PRO-30%PRO-1%
5BTF [11]0.6710.9800.9200.2880.6520.8150.7380.202
AST [24]0.6800.9500.9030.1580.6330.7410.6910.143
M3DM [31]0.8220.9840.9370.3300.7640.8710.8070.251
CFM [5]0.8110.9860.9490.3820.7950.8790.8010.260
3D-ADNAS [21]0.826---0.7750.875--
FIND (Ours)0.8990.9910.9630.4200.8400.9510.8620.324
10BTF [11]0.6950.9830.9280.3080.6850.8340.8060.217
AST [24]0.6890.9460.8350.1740.6710.7670.6240.161
M3DM [31]0.8450.9860.9430.3550.8240.8900.8120.266
CFM [5]0.8450.9870.9540.3980.8380.9030.8250.284
3D-ADNAS [21]0.848---0.8070.869--
FIND (Ours)0.9210.9930.9670.4340.8680.9590.8890.350
50BTF [11]0.8060.9890.9470.3560.7210.8560.8240.240
AST [24]0.7940.9740.9290.3350.7390.8620.7150.118
M3DM [31]0.9070.9890.9550.3870.8360.9330.8460.304
CFM [5]0.9060.9910.9650.4310.8520.9260.8510.315
3D-ADNAS [21]0.890---0.8680.912--
FIND (Ours)0.9520.9950.9730.4550.8970.9760.9070.365
ALLBTF [11]0.8650.9920.9590.3830.7400.8830.8450.261
AST [24]0.9370.9760.9440.3980.7800.9020.7440.149
M3DM [31]0.9450.9920.9640.3940.8820.9770.8870.331
CFM [5]0.9540.9930.9710.4550.8810.9740.8870.335
3D-ADNAS [21]0.951---0.9460.970--
FIND (Ours)0.9780.9950.9830.4560.9550.9830.9250.395
", + "bbox": [ + 173, + 128, + 823, + 478 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ac3325aabec7bc9b416b026048ce9ad337930e0b874126ab72e88a3ef7548fa9.jpg", + "table_caption": [ + "Table 2. Comparisons of per-category anomaly detection performance on MVTec-3D-AD." + ], + "table_footnote": [], + "table_body": "
MethodBagelCable GlandCarrotCookieDowelFoamPeachPotatoRopeTireMean
I-AUROC
BTF [11]0.9380.7650.9720.8880.9600.6640.9040.9290.9820.7260.865
AST [24]0.9830.8730.9760.9710.9320.8850.9740.9811.0000.7970.937
M3DM [31]0.9940.9090.9720.9760.9600.9420.9730.8990.9720.8500.945
CFM [5]0.9940.8880.9840.9930.9800.8880.9410.9430.9800.9530.954
3D-ADNAS [21]0.9971.0000.9710.9860.9660.9480.8970.8731.0000.8670.951
Shape Guided [4]0.9860.8940.9830.9910.9760.8570.9900.9650.9900.8690.947
LSFA [29]1.0000.9390.9820.9890.9610.9510.9830.9620.9890.9510.971
FIND (Ours)0.9840.9700.9850.9720.9820.9790.9740.9800.9880.9700.978
AUPRO@30%
BTF [11]0.9760.9690.9790.9730.9330.8880.8960.9120.9500.9710.959
AST [24]0.9700.9470.9810.9390.9130.9060.9790.9820.8890.9400.944
M3DM [31]0.9700.9710.9790.9500.9410.9320.9770.9710.9710.9750.964
CFM [5]0.9790.9720.9820.9450.9500.9680.9800.9430.9500.9810.971
Shape Guided [4]0.9810.9730.9820.9710.9620.9780.9810.9830.9740.9750.976
LSFA [29]0.9860.9740.9810.9460.9250.9410.9830.9830.9740.9830.968
FIND (Ours)0.9890.9770.9810.9830.9820.9860.9840.9800.9900.9780.983
", + "bbox": [ + 151, + 515, + 844, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "blocks, ensuring both preservation of the locality and improved alignment between the modalities.", + "bbox": [ + 89, + 787, + 482, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To ensure a fair comparison with previous work, we use the same ViT-B/8 DINO backbone as CFM [5], which is shared for both RGB and surface normal modalities. We treat surface normals as 3-channel RGB-style inputs and normalize them using ImageNet statistics before feeding", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "them into DINO. During training, the backbone weights remain fixed, and we optimize only the four student networks. We extract hidden features from teacher's 4, 6, 8-th layers for intra-modal reverse distillation, while the 10-th layer's feature is employed as the input for cross-modal student. We use the ADAM optimizer for training with a learning rate of $1 \\times 10^{-4}$ , and set $\\alpha = 0.5$ for all-shot tasks and", + "bbox": [ + 511, + 787, + 906, + 893 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23295", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/314955c28ff1a8386b969f81c76c131051abd5da2b3e22b3dbefefd4130e5d40.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 90, + 477, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7f6bf279ded68aa23f81b46a77a119b9a3007ebfa48616f8b39567b252c56576.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 257, + 475, + 393 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fac46808338bade6a8a5bf56d51c18da0b15e10cb1b8373eb55801e0c6ce9564.jpg", + "image_caption": [ + "Figure 4. Qualitative results on MVTec-3D-AD with various types of anomalies. From left to right: input RGB images, surface normals, GT anomaly maps, anomaly maps from 2D student set, anomaly maps from 3D student set, and our final predictions. Our final model consistently produces more accurate anomaly maps across different anomaly types." + ], + "image_footnote": [], + "bbox": [ + 101, + 405, + 475, + 542 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8dbab268d4a2f5c6efed4f14cb6a01bcef492b9245964b7b81a2159d9e92a61b.jpg", + "image_caption": [ + "Figure 5. Anomaly score distributions on bagel and cookie (MVTec-3D-AD, 10-shot). Our method shows better separation between normal and anomalous samples." + ], + "image_footnote": [], + "bbox": [ + 94, + 657, + 482, + 746 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$\\alpha = 0.8$ for few-shot tasks.", + "bbox": [ + 89, + 825, + 272, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For MVTec-3D-AD, each sample includes an RGB image and an unstructured point cloud. We first organize the point cloud into a 2D grid(XYZ map) corresponding to the rgb image size $H \\times W$ . We then compute per-pixel surface", + "bbox": [ + 89, + 839, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2388f5486de21d29ca1148893218b92a203a8215fa9a1804586e03e95745688f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CategoryIntra-studentCross-studentIntra + Cross
ImagePixelImagePixelImagePixel
Bagel96.399.798.899.599.599.6
Cable Gland74.497.679.098.382.498.3
Carrot97.499.899.499.997.399.8
Cookie86.498.491.897.292.597.8
Dowel75.298.183.899.382.699.2
Foam78.999.197.699.898.199.8
Peach97.099.895.799.999.499.9
Potato92.999.892.799.995.899.9
Rope97.699.896.699.898.899.8
Tire66.798.775.199.274.499.2
Average86.399.191.199.392.199.3
", + "bbox": [ + 517, + 89, + 903, + 318 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/937c3808e4f9468cb412eb3b99b88496b195f6f24d945e72b19428927a64cc2e.jpg", + "table_caption": [ + "Table 3. Ablation studies on intra-modal, cross-modal and dual students. We report the 10-shot categorical performance in image and pixel AUROC (\\%) on the MVTec-3D dataset." + ], + "table_footnote": [], + "table_body": "
CategoryDepthPoint CloudSurface Normal
ImagePixelImagePixelImagePixel
Bagel99.099.498.399.599.599.6
Cable Gland68.897.774.197.682.498.3
Carrot82.799.392.499.197.399.8
Cookie84.498.590.797.292.597.8
Dowel87.199.075.599.082.699.2
Foam81.689.892.499.598.199.8
Peach85.799.393.799.699.499.9
Potato68.499.184.398.795.899.9
Rope95.799.694.299.798.899.8
Tire66.897.673.699.374.499.2
Average82.097.986.898.992.199.3
", + "bbox": [ + 514, + 405, + 911, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablation studies on 3D modality input selection. We report the 10-shot categorical performance in image and pixel AUROC (\\%) on the MVTec-3D dataset.", + "bbox": [ + 511, + 643, + 906, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "normals by fitting a local plane to the neighbors of each point (or using the cross-product of adjacent triangle faces in a mesh representation). This produces a 3-channel normal vector map for the surface of the object. Following prior work [5, 31], we remove background points and fill the removed regions in the RGB image with a neutral value (black) to avoid false anomalies from the background. We finally resize both the RGB image and the normal map to $224 \\times 224$ pixels before feeding them to the model. For Eyecandies, we use normal maps with global lighting as the 3D modality.", + "bbox": [ + 511, + 734, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23296", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Primary Results", + "text_level": 1, + "bbox": [ + 89, + 90, + 253, + 107 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Few-shot Performance on MVtec-3D-AD. Table 1 shows that FIND exceeds all competitors: BTF, AST, M3DM, CFM and 3D-ADNAS, in all few-shot settings (5, 10, 50 shots) and evaluation metrics at both the image and pixel levels. For example, in the challenging 5-shot setting, our method achieves an I-AUROC of 0.899. With 10-shot training, the I-AUROC performance is increased to 0.921. Even with 50 shots, FIND continues to outperform strong memory-based methods (e.g., M3DM). These consistent gains highlight the effectiveness of FIND in limited data scenarios.", + "bbox": [ + 89, + 114, + 485, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Full-shot Performance on MVtec-3D-AD. Table 1 presents the anomaly detection performance in the full-shot setting at both the image and pixel levels, while Table 2 reports results for each category. FIND surpasses existing methods in both image-level and pixel-level detection, achieving the highest average I-AUROC $(97.8\\%)$ and AUPRO@ $30\\%$ $(98.3\\%)$ , demonstrating its robustness across various object classes. Notably, our approach excels in challenging categories (e.g., Cable Gland, Foam), indicating its ability to capture fine-grained anomalous patterns.", + "bbox": [ + 89, + 289, + 483, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results on Eyecandies. As shown in Table 1, FIND also sets a new state-of-the-art on Eyecandies under both few-shot and full-shot protocols, significantly outperforming existing methods. For example, at 5-shot, FIND achieves an I-AUROC score of 0.840, exceeding the other baselines significantly. This performance gain extends to pixel-level metrics (P-AUROC and AUPRO), demonstrating a superior localization of subtle anomalies. Furthermore, with 50 normal samples for training, our method even outperforms the full-shot trained M3DM and CFM in terms of AUROC and AUPRO. These results highlight the effectiveness of our approach in various shots of training data scenarios, reinforcing its robustness and generalizability.", + "bbox": [ + 89, + 449, + 483, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Qualitative Analysis", + "text_level": 1, + "bbox": [ + 89, + 662, + 284, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 4 shows that our method is robust to various types of anomaly (shape-only, color-only, combined) by effectively exploiting 2D-3D complementarity. The 2D student set $(\\mathbf{S}_{3D\\rightarrow 2D},\\mathbf{S}_{2D\\rightarrow 2D})$ captures subtle color anomalies often overlooked by geometric cues, while the 3D student set $(\\mathbf{S}_{2D\\rightarrow 3D},\\mathbf{S}_{3D\\rightarrow 3D})$ detects shape deformations that are imperceptible in RGB. Combining these maps improves localization accuracy and reduces false positives.", + "bbox": [ + 89, + 686, + 483, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 5 further compares the anomaly score distributions of normal and anomalous samples under the 10-shot setting for CFM and FIND on two MVTec-3D-AD classes (bagel and cookie). Our method yields more clearly separable distributions between normal and abnormal samples, contributing to improved detection accuracy.", + "bbox": [ + 89, + 810, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 513, + 90, + 679, + 106 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Intra-modal vs. Cross-modal vs. Dual Student. We evaluate the contributions of the intra-modal and cross-modal students, with results shown in Table 3. Intra-modal students capture modality-specific anomalies and guide the model to reproduce a normal version of the query sample at the pixel level, enhancing the detection of local defects. Removing them causes a notable drop in detection accuracy for categories such as cable gland, potato, and rope. However, intra-modal students operating within a single modality cannot leverage complementary information from the other modality, leaving those cross-modal anomalies undetected. In contrast, cross-modal students learn the expected normal correspondence between an object's color and shape appearance and detect anomalies by identifying inconsistencies between modalities. By combining both intra-modal and cross-modal students, we leverage their complementary strengths and achieve the best results.", + "bbox": [ + 511, + 112, + 906, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3D Modality Input Choices. In a 10-shot setting on MVTec-3D-AD, we compare raw point clouds (with a PointNet encoder), depth maps(from point clouds), and surface-normal maps for anomaly detection. The results are shown in Table 4. Surface normals achieve the highest I-AUROC, outperforming depth maps and raw point clouds. This indicates that normal maps preserve fine-grained geometric details that can highlight small defects. In contrast, depth maps lack this level of detail, resulting in lower accuracy. The point cloud baseline performs worst, likely due to a modality gap introduced by using a separate PointNet-based 3D feature extractor. Such gaps arise from misaligned representations when using a point cloud network versus image-based features. Our results show that surface normal maps provide the most effective shape modality for 3D anomaly detection in few-shot scenarios.", + "bbox": [ + 511, + 375, + 908, + 618 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 630, + 633, + 646 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we address the critical challenges of multimodal anomaly detection in few-shot scenarios by introducing FIND, a dual-student framework that integrates intramodal reverse distillation and cross-modal feature mapping. FIND overcomes data scarcity and modality gaps by exploiting collaborative learning between two students: one dedicated to fine-grained intra-modal anomaly localization through reverse distillation, and another focused on cross-modal inconsistency detection. Extensive experiments demonstrate that FIND achieves new state-of-the-art results in both few-shot and full-shot settings.", + "bbox": [ + 511, + 656, + 908, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 513, + 835, + 681, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported by the Agency for Science, Technology and Research (A*STAR) under its MTC Programmatic Funds (Grant No. M23L7b0021).", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23297", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Paul Bergmann, Xin Jin, David Sattlegger, and Carsten Steger. The mvtec 3d-ad dataset for unsupervised 3d anomaly detection and localization. In Proceedings of the 17th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications, VISIGRAPP, pages 202-213, 2022. 5", + "[2] Luca Bonfiglioli, Marco Toschi, Davide Silvestri, Nicola Fioraio, and Daniele De Gregorio. The eyecandies dataset for unsupervised multimodal anomaly detection and localization. In 16th Asian Conference on Computer Vision (ACCV), pages 459-475, 2022. 5", + "[3] Mengxi Chen and ... Enhanced multimodal representation learning with cross-modal KD. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11766-11775, 2023. 3", + "[4] Yu-Min Chu, Liu Chieh, Ting-I Hsieh, Hwann-Tzong Chen, and Tyng-Luh Liu. Shape-guided dual-memory learning for 3d anomaly detection. 2023. 1, 2, 6", + "[5] Alex Costanzino, Pierluigi Zama Ramirez, Giuseppe Lisanti, and Luigi Di Stefano. Multimodal industrial anomaly detection by crossmodal feature mapping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17234-17243, 2024. 1, 2, 4, 5, 6, 7", + "[6] Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one-class embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9737-9746, 2022. 1", + "[7] Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one-class embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9737-9746, 2022. 1", + "[8] Zheng Fang, Xiaoyang Wang, Haocheng Li, Jiejie Liu, Qiugui Hu, and Jimin Xiao. Fastrecon: Few-shot industrial anomaly detection via fast feature reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17481-17490, 2023. 1, 3", + "[9] Zhihao Gu, Jiangning Zhang, Liang Liu, Xu Chen, Jinlong Peng, Zhenye Gan, Guannan Jiang, Annan Shu, Yabiao Wang, and Lizhuang Ma. Rethinking reverse distillation for multi-modal anomaly detection. In Thirty-Eighth AAAI Conference on Artificial Intelligence, AAAI, pages 8445-8453, 2024. 1, 2", + "[10] Yu Hong, Hang Dai, and Yong Ding. Cross-modality knowledge distillation network for monocular 3d object detection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 87-104, 2022. 3", + "[11] Eliahu Horwitz and Yedid Hoshen. Back to the feature: classical 3d features are (almost) all you need for 3d anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2968-2977, 2023. 1, 5, 6", + "[12] Chaoqin Huang, Haoyan Guan, Aofan Jiang, Ya Zhang, Michael Spratling, and Yan-Feng Wang. Registration based few-shot anomaly detection. In European Conference on Computer Vision, pages 303-319. Springer, 2022. 1, 3" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Fushuo Huo, Wenchao Xu, Jingcai Guo, Haozhao Wang, and Song Guo. $C^2$ kd: Bridging the modality gap for cross-modal knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7151-7160, 2024. 3", + "[14] Jongheon Jeong, Yang Zou, Taewan Kim, Dongqing Zhang, Avinash Ravichandran, and Onkar Dabeer. Winclip: Zero/few-shot anomaly classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19606-19616, 2023. 1, 3", + "[15] Pilhyeon Lee and ... Decomposed cross-modal distillation for RGB-based temporal action detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages -, 2023. CVPR 2023. 3", + "[16] Yujin Lee, Seoyoon Jang, and Hyunsoo Yoon. Anople: Few-shot anomaly detection via bi-directional prompt learning with only normal samples. arXiv preprint arXiv:2408.13516, 2024. 1, 3", + "[17] Wenqiao Li, Xiaohao Xu, Yao Gu, Bozhong Zheng, Shenghua Gao, and Yingna Wu. Towards scalable 3d anomaly detection and localization: A benchmark via 3d anomaly synthesis and a self-supervised learning network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22207-22216, 2024. 1", + "[18] Hanzhe Liang, Guoyang Xie, Chengbin Hou, Bingshu Wang, Can Gao, and Jinbao Wang. Look inside for more: Internal spatial modality perception for 3d anomaly detection. arXiv preprint arXiv:2412.13461, 2024. 1, 2", + "[19] Victor Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Y Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. Advances in Neural Information Processing Systems, 35:17612-17625, 2022. 3", + "[20] Chieh Liu, Yu-Min Chu, Ting-I Hsieh, Hwann-Tzong Chen, and Tyng-Luh Liu. Learning diffusion models for multi-view anomaly detection. In European Conference on Computer Vision, pages 328-345. Springer, 2025. 1", + "[21] Kaifang Long, Guoyang Xie, Lianbo Ma, Jiaqi Liu, and Zhichao Lu. Revisiting multimodal fusion for 3d anomaly detection from an architectural perspective. In Thirty-Eighth AAAI Conference on Artificial Intelligence, AAAI, 2025. 1, 2, 5, 6", + "[22] Xiang Ma, Xuemei Li, Lexin Fang, and Caiming Zhang. Bridging the modality gap: Dimension information alignment and sparse spatial constraint for image-text matching. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 5074-5082, 2024. 3", + "[23] Nicolae-Cătălin Ristea, Neelu Madan, Radu Tudor Ionescu, Kamal Nasrollahi, Fahad Shahbaz Khan, Thomas B. Moeslund, and Mubarak Shah. Self-supervised predictive convolutional attentive block for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13576-13586, 2022. 1", + "[24] Marco Rudolph, Tom Wehrbein, Bodo Rosenhahn, and Bastian Wandt. Asymmetric student-teacher networks for industrial anomaly detection. In IEEE/CVF Winter Conference on" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23298", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Applications of Computer Vision, WACV, pages 2592-2602, 2023. 1, 2, 5, 6", + "[25] Pritam Sarkar and Ali Etemad. XKD: Cross-modal knowledge distillation with domain alignment for video representation learning. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2024. to appear. 3", + "[26] Wenbo Sui, Daniel Lichau, Josselin Lefèvre, and Harold Philippeau. Incomplete multimodal industrial anomaly detection via cross-modal distillation, 2024. 1, 3", + "[27] Long Tian, Hongyi Zhao, Ruiying Lu, Rongrong Wang, Yu-jie Wu, Liming Wang, Xiongpeng He, and Xiyang Liu. Foct: Few-shot industrial anomaly detection with foreground-aware online conditional transport. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6241-6249, 2024. 1, 3", + "[28] Dinh Tien Tran, Anh Tuan Nguyen, Nguyen Hoang Tran, Ta Duc Huy, Soan T. M. Duong, Chanh D. Tr. Nguyen, and Steven Q. H. Truong. Revisiting reverse distillation for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24511-24520, 2023. 1", + "[29] Yuanpeng Tu, Boshen Zhang, Liang Liu, Yuxi Li, Jiangning Zhang, Yabiao Wang, Chengjie Wang, and Cairong Zhao. Self-supervised feature adaptation for 3d industrial anomaly detection. In European Conference on Computer Vision, pages 75-91. Springer, 2025. 2, 6", + "[30] Huiqun Wang, Yiping Bao, Panwang Pan, Zeming Li, Xiao Liu, Ruijie Yang, and Di Huang. Multi-modal relation distillation for unified 3d representation learning. In European Conference on Computer Vision (ECCV), 2024. to appear. 3", + "[31] Yue Wang, Jinlong Peng, Jiangning Zhang, Ran Yi, Yabiao Wang, and Chengjie Wang. Multimodal industrial anomaly detection via hybrid fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8032-8041, 2023. 1, 2, 5, 6, 7", + "[32] Zeyu Wang, Dingwen Li, Chenxu Luo, Cihang Xie, and Xiaodong Yang. DistillBEV: Boosting multi-camera 3d object detection with cross-modal knowledge distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8637-8646, 2023. 3", + "[33] Zihui Xue, Zhengqi Gao, Sucheng Ren, and Hang Zhao. The modality focusing hypothesis: Towards understanding cross-modal knowledge distillation. In International Conference on Learning Representations (ICLR), 2023. Top $5\\%$ at ICLR 2023. 3", + "[34] Xu Yan, Heshen Zhan, Chaoda Zheng, Jiantao Gao, Ruimao Zhang, Shuguang Cui, and Zhen Li. Let images give you more: Point cloud cross-modal training for shape analysis. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 1", + "[35] Heeseung Yun, Joonil Na, and Gunhee Kim. Dense 2d-3d indoor prediction with sound via aligned cross-modal distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7829-7838, 2024. 3", + "[36] Vitjan Zavrtanik, Matej Kristan, and Danijel Skocaj. Cheating depth: Enhancing 3d surface anomaly detection via depth" + ], + "bbox": [ + 91, + 90, + 482, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "simulation. In IEEE/CVF Winter Conference on Applications of Computer Vision, WACV, pages 2153-2161. IEEE, 2024. 2", + "[37] Tianlu Zhang, Hongyuan Guo, Qiang Jiao, Qiang Zhang, and Jungong Han. Efficient rgb-t tracking via cross-modality distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5815-5824, 2023. 3", + "[38] Xuan Zhang, Shiyu Li, Xi Li, Ping Huang, Jiulong Shan, and Ting Chen. Destseg: Segmentation guided denoising student-teacher for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3914–3923, 2023. 1", + "[39] Shengchao Zhou, Weizhou Liu, Chen Hu, Shuchang Zhou, and Chao Ma. Unidistill: A universal cross-modality knowledge distillation framework for 3d object detection in bird's-eye view. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5116-5125, 2023. 3", + "[40] Zheyuan Zhou, Le Wang, Naiyu Fang, Zili Wang, Lemiao Qiu, and Shuyou Zhang. R3d-ad: Reconstruction via diffusion for 3d anomaly detection. In European Conference on Computer Vision, pages 91-107. Springer, 2025. 1" + ], + "bbox": [ + 516, + 92, + 906, + 415 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23299", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/FIND_ Few-Shot Anomaly Inspection with Normal-Only Multi-Modal Data/94dd2e05-5a4d-4859-bd44-9496d30c1756_model.json b/2025/FIND_ Few-Shot Anomaly Inspection with Normal-Only Multi-Modal Data/94dd2e05-5a4d-4859-bd44-9496d30c1756_model.json new file mode 100644 index 0000000000000000000000000000000000000000..53de5f6143bac12fb2bd8cda5ce57668fead6bda --- /dev/null +++ b/2025/FIND_ Few-Shot Anomaly Inspection with Normal-Only Multi-Modal Data/94dd2e05-5a4d-4859-bd44-9496d30c1756_model.json @@ -0,0 +1,1760 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.12, + 0.131, + 0.877, + 0.152 + ], + "angle": 0, + "content": "FIND: Few-Shot Anomaly Inspection with Normal-Only Multi-Modal Data" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.18, + 0.932, + 0.253 + ], + "angle": 0, + "content": "Yiting Li\\(^{1}\\), Fayao Liu\\(^{1\\dagger}\\), Jingyi Liao\\(^{1}\\), Sichao Tian\\(^{2}\\), Chuan-Sheng Foo\\(^{1}\\), Xulei Yang\\(^{1\\dagger}\\) \n\\(^{1}\\)Institute for Infocomm Research, A*STAR, Singapore \\(^{2}\\)Chinese Academy of Medical Sciences and Peking Union Medical College, China \n{li_yiting, liu_fayao, liao_jingyi, foo_chuan_sheng, yang_xulei} @i2r.a-star.edu.sg, {sctian} @implad.ac.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.318, + 0.486, + 0.651 + ], + "angle": 0, + "content": "Multi-modal anomaly detection (MAD) improves industrial inspection by exploiting complementary 2D and 3D data. However, existing methods struggle in few-shot scenarios due to limited data and modality gaps. Current approaches either fuse multi-modal features or align cross-modal representations; however, they often suffer from high false-positive rates and fail to detect subtle defects, especially when training samples are scarce. To address these challenges, we propose the first few-shot MAD method FIND, a novel dual-student framework that integrates intramodal reverse distillation and cross-modal feature mapping. FIND employs modality-specific teachers and two collaborative students: an intra-modal student for fine-grained anomaly localization via reverse distillation, and a cross-modal student that captures inter-modal correspondences to detect cross-modal inconsistencies. Extensive experiments on MVTec-3D-AD and Eyecandies show that FIND significantly outperforms state-of-the-art methods in both full-shot and few-shot settings. Ablation studies validate the complementary roles of intra- and cross-modal distillation. Our work significantly advances MAD robustness in data-scarce industrial applications." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.222, + 0.699 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.708, + 0.483, + 0.859 + ], + "angle": 0, + "content": "Multi-modal anomaly detection (MAD) has garnered increasing interest in industrial inspection [4, 17, 18, 20, 21, 26, 31, 34, 40], as relying on a single modality (e.g., RGB) may fail to capture crucial geometric cues under varying surface textures. By integrating heterogeneous data such as point clouds or depth maps, MAD methods can leverage multi-modal complementarity to enhance inspection quality. Since anomalous data are rare, most existing anomaly detection (AD) methods are trained only on normal data. Our work follows this protocol." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.86, + 0.483, + 0.875 + ], + "angle": 0, + "content": "While significant progress has been made in MAD, exist" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.287, + 0.713, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.287, + 0.902, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.426, + 0.907, + 0.497 + ], + "angle": 0, + "content": "Figure 1. Comparison with state-of-the-art methods on MVTec-3D-AD in terms of I-AUROC (left) and AUPRO@1% (right). FIND substantially outperforms prior methods (BTF [11], AST [24], M3DM [31], 3D-ADNAS [21] and CFM [5]) in both few-shot and full-shot settings." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.523, + 0.907, + 0.704 + ], + "angle": 0, + "content": "ing approaches struggle to retain robustness in few-shot scenarios, where limited normal multi-modal training data is available—a critical limitation given the high cost of acquiring such data [8, 12, 14, 16, 27]. This is evidenced in Fig. 1, which compares various methods using different amounts of normal training samples, with notable performance drops observed. In this work, we propose FIND, a simple yet effective dual-student MAD framework focused on Few-Shot Anomaly Inspection with Normal-Only Multi-Modal Data. It effectively minimizes modality interference while exploiting multi-modal complementarity, significantly enhancing anomaly detection and localization." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.705, + 0.909, + 0.825 + ], + "angle": 0, + "content": "In single-modality anomaly detection, teacher-student (T-S) frameworks achieve strong results [6, 23, 28, 38]. A pretrained teacher extracts generic features, while a student trained on normal data mimics the teacher. Anomalies emerge as discrepancies between their features at inference. Reverse distillation [7, 28] refines this by reconstructing teacher features from compact one-class embeddings, enabling fine-grained defect localization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.901 + ], + "angle": 0, + "content": "However, extending the success of T-S frameworks to MAD remains a challenge. Existing methods either fuse modalities by concatenating RGB and depth at the input level [24] or across multiple feature levels [9], which risks modality interference. Specifically, MMRD [9] introduces" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.242, + 0.9 + ], + "angle": 0, + "content": "†Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23290" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.438 + ], + "angle": 0, + "content": "a multi-modal reverse distillation paradigm that performs a weighted fusion of image and depth data across multiple feature levels, creating a multi-modal teacher target for a learnable student decoder tasked with restoring the teacher's multi-modal representations. In contrast, cross-modal Feature Mapping (CFM) [5], although not explicitly formulated within the T-S framework, uses domain-specific teachers for cross-modal distillation and performs late fusion on score maps instead of relying on feature fusion. The features of different modalities are aligned through an MLP-based projection network. CFM trains two student models to map features from one modality to another by mimicking the corresponding target modality teacher. By learning the expected correspondence between an object's color and shape, any deviation from this correspondence at test time indicates an anomaly. However, the inherent modality gap can lead to large reconstruction errors even for normal samples, resulting in high false positive rates and limiting the model's ability to distinguish genuine anomalies. These limitations result in high false positives and coarse localization (see Fig. 2 for examples), which are further exacerbated in few-shot scenarios, where limited normal samples hinder the effective alignment of the two modalities." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.441, + 0.485, + 0.818 + ], + "angle": 0, + "content": "To address the challenges stated above, we present FIND, which unifies intra-modal reverse distillation and cross-modal distillation within a single T-S framework. Our intuition is to minimize modality interference while exploiting the complementary strengths of multiple modalities. Specifically, each modality is paired with a modality-specific teacher, an intra-modal student, and a cross-modal student. The intra-modal student is trained via reverse distillation to model fine-grained, modality-specific normal appearances, while the cross-modal student is designed to capture inter-modal correspondences for robust anomaly detection. This decoupled design allows each student to focus on its specific task: the intra-modal student is highly sensitive to prominent modality-specific anomalies and provides finer-grained localization, while the cross-modal student detects anomalies that disrupt the learned normal correspondence between different modalities. This synergy enables precise localization and robustness to various types of anomalies, even with limited training data (see Fig. 2 for comparisons). To further mitigate the modality gap and thus reduce false positives in cross-modal distillation, we propose using surface normals as a geometric proxy, enabling the use of shared RGB and 3D encoders. We find this simple strategy to be highly effective in achieving accurate MAD. Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.825, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- We propose the first few-shot MAD method FIND, which synergistically integrates intra-modal reverse distillation and cross-modal distillation. By decoupling the learning objectives for intra- and cross-modal students, FIND significantly enhances MAD performance." + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.089, + 0.906, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.317, + 0.907, + 0.36 + ], + "angle": 0, + "content": "Figure 2. Compared to CFM under the 10-shot setting, our anomaly maps are sharper and show fewer false positives, focusing on anomalous patterns rather than minor texture variations." + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.389, + 0.905, + 0.448 + ], + "angle": 0, + "content": "- We show that surface normals are a highly effective geometric modality for MAD, effectively bridging the modality gap in cross-modal distillation, reducing false positives, and enhancing anomaly detection precision." + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.449, + 0.905, + 0.495 + ], + "angle": 0, + "content": "- FIND sets a new state-of-the-art in both multi-modal anomaly detection and localization tasks, outperforming existing methods in both full-shot and few-shot settings." + }, + { + "type": "list", + "bbox": [ + 0.525, + 0.389, + 0.905, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.512, + 0.659, + 0.527 + ], + "angle": 0, + "content": "2. Related works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Multi-modal Anomaly Detection With the development of 3D sensors, multi-modal anomaly detection has gained increasing attention [4, 9, 21, 36], as 3D information (e.g., point clouds) offers richer geometric cues and captures shape deformations that may be invisible in RGB domain. Recently, researchers have explored cross-modal alignment methods that fuse RGB and 3D data [24, 31] for more robust anomaly localization. For instance, CFM [5] proposes a novel framework that learns to map 2D and 3D features to each other, thereby pinpointing anomalies where the two modalities disagree; [29] introduce a local-to-global self-supervised adaptation mechanism to finetune pre-trained 2D/3D networks for more precise cross-modal alignment; [31] design a hybrid fusion scheme with multiple memory banks to achieve consistent performance gains over single-modality baselines. In parallel, memory-based 3D AD approaches [21, 31] have also shown strong potential by storing normal embeddings from both color and geometry in dedicated banks and measuring anomalous deviations at test time. For example, [4] propose a shape-guided dual-memory method that leverages signed distance fields (SDF) and RGB cues for complementary anomaly localization; [18] advocate an internal spatial modality perception module to distill intrinsic 3D structures into a memory" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23291" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.226 + ], + "angle": 0, + "content": "bank; [26] handle incomplete multi-modal data via cross-modal distillation and memory retrieval. Although existing methods have significantly advanced the state of the art in MAD, most still rely on either large-scale normal training sets or carefully aligned networks for each modality. In scenarios where only a few normal samples are available, the modality gap between 2D images and 3D point clouds can hinder reliable cross-modal feature alignment, thus limiting the models' ability to detect subtle anomalies." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.232, + 0.485, + 0.669 + ], + "angle": 0, + "content": "Few-shot Anomaly Detection In resource-constrained scenarios, collecting large amount of normal data can be both costly and impractical. Consequently, few-shot anomaly detection (FSAD) becomes a promising alternative. Current research on FSAD has primarily focused on RGB domains. AnoPLe [16] proposes bi-directional prompt learning with only normal samples, leveraging a pretrained vision-language model to detect anomalies from both textual and visual cues. FastRecon [8] proposes a lightweight feature-based regression with distribution regularization to obtain the optimal transformation from support to query features, it reconstructs a normal version of test sample's features from visually similar support features and detects any large deviations as anomalies. WinCLIP [14] explores a window-based feature fusion strategy to extract multi-scale features for zero-/few-shot anomaly classification and segmentation. FCOT [27] leverages a distribution-based Conditional Transport metric to minimize feature discrepancies between a few support images and query images, while updating a memory bank to precisely match anomalous regions. RegAD [12] uses image registration to align query images with support images in a category-agnostic manner. Despite remarkable progress, existing FSAD approaches still face limitations when extended to multi-modal scenarios due to the lack of capabilities for capturing the shape distortions that 3D data would reveal. This limitation motivates our work on multi-modal anomaly detection under few-shot settings, by capturing shape-induced defects that 2D-only methods may overlook." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Cross-modal knowledge distillation Cross-modal knowledge distillation (KD) has attracted increasing interest for transferring knowledge between heterogeneous modalities (e.g. images, LiDAR, thermal, audio) [3, 15, 25, 30, 33]. A core challenge is the modality gap issue due to the discrepancies in data/feature distribution between modalities [13, 19, 22]. [19] demonstrates that modality gap in multimodal models such as CLIP is caused by a combination of model initialization and contrastive learning optimization. Recent works tackle this by aligning feature representations and customizing distillation strategies to each modality pair. For example, a LiDAR-based teacher can guide a camera-based student to improve monocular detection accuracy [10, 15, 32]. C2KD [13] propose a bidirectional distillation framework where teacher and student" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.259 + ], + "angle": 0, + "content": "learn from each other, coupled with on-the-fly sample filtering to discard misaligned targets. Cross-modal KD is applied in RGB-Thermal (RGB-T) tracking [37, 39], a two-stream RGB-T network can serve as a powerful teacher to a lightweight one-stream tracker; distilling both modality-specific and shared features helps the student network approach the teacher's performance, while greatly reducing computation. In audio-visual learning [35], visual models have been used to teach audio models to predict spatial or semantic information from sound, often by enforcing feature alignments to handle the modality gap." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.272, + 0.605, + 0.287 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.297, + 0.719, + 0.312 + ], + "angle": 0, + "content": "3.1. Framework Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.319, + 0.908, + 0.605 + ], + "angle": 0, + "content": "FIND consists of two single-teacher-dual-student branches designed to detect anomalies from multi-modal 2D and 3D inputs. Fig. 3 shows an overview of our method. Each branch is composed of a modality-specific teacher \\((\\mathbf{T}_{2\\mathrm{D}}\\) or \\(\\mathbf{T}_{3\\mathrm{D}})\\), a cross-modal distillation student \\((\\mathbf{S}_{3D\\rightarrow 2D}\\) or \\(\\mathbf{S}_{2D\\rightarrow 3D})\\), and an intra-modal reverse distillation student \\((\\mathbf{S}_{2D\\rightarrow 2D}\\) or \\(\\mathbf{S}_{3D\\rightarrow 3D})\\). Our motivation is to integrate cross-modal distillation with reverse distillation in a unified framework that minimizes modality interference while effectively exploiting their complementarity for anomaly detection. In essence, the Cross-Modal Student distills knowledge across the RGB and 3D streams, learning how normal features in one modality correspond to those in the other. The Intra-Modal Student, on the other hand, learns to reproduce the normal pattern of individual modality through reverse distillation. During training, both students receive supervision derived from a shared set of teacher features so that each modality learns a more comprehensive representation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.616, + 0.857, + 0.632 + ], + "angle": 0, + "content": "3.2. Cross-modal Teacher-Student Learning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.638, + 0.905, + 0.683 + ], + "angle": 0, + "content": "We aim to leverage the complementary knowledge between two different modalities, namely 2D and 3D, for anomaly detection through a cross-modal distillation procedure." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Bridging the Modality Gap. Since features extracted by 2D backbones for RGB images and 3D point-based networks for point clouds follow different distributions, existing MAD methods face the modality gap issue due to the use of separate feature backbones for each modality. This discrepancy becomes even more significant in few-shot learning scenarios, where a student model with limited training data struggles to align with the teacher's normal feature distribution, leading to a high false positive rate in detection performance. To bridge the modality gap between the teacher and its cross-modal student, we propose extracting surface normal maps from raw point clouds to encode 3D geometry, enabling a shared backbone network used for both 2D and 3D modalities. Our key insight is that sur" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23292" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.086, + 0.488, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.087, + 0.904, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.345, + 0.907, + 0.43 + ], + "angle": 0, + "content": "Figure 3. FIND consists of two single-teacher-dual-student branches (left and right) designed to detect anomalies from multi-modal inputs, i.e., 2D RGB and 3D surface-normal. Each branch consists of a modality-specific teacher (center), a cross-modal distillation student (top), and an intra-modal reverse distillation student (bottom). The cross-modal student performs patch-wise mapping to the teacher modality for capturing inter-modality inconsistencies. In parallel, the intra-modal student conducts layer-wise feature inversion back to the original input resolution, enabling pixel-level localization of defects. By combining the results from the two pairs of intra-modal and cross-modal students, our approach robustly identifies various types of anomalies, even in few-shot scenarios." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.455, + 0.482, + 0.59 + ], + "angle": 0, + "content": "face normal maps capture fine-grained surface orientation details and preserve subtle textures more effectively than depth maps. This contrasts with previous work, which uses either raw point clouds or depth maps. Raw point clouds exacerbate the domain gap, while depth images tend to smooth out small shape variations and are less sensitive to small defects. These findings allow us to effectively bridge the modality gap and therefore significantly reduce false positives." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.598, + 0.483, + 0.734 + ], + "angle": 0, + "content": "Cross-modal Distillation. Let \\(\\mathbf{T}_{2D}\\) and \\(\\mathbf{T}_{3D}\\) denote the two pre-trained Vision Transformer (ViT) teachers for the 2D RGB images and 3D surface normals, respectively. Inspired by CFM [5], each teacher receives input from its native domain and outputs high-level features \\(\\mathbf{f}_{2D}^T\\) or \\(\\mathbf{f}_{3D}^T\\) during training. In parallel, the corresponding cross-modal distillation student, \\(\\mathbf{S}_{3D\\rightarrow 2D}\\) or \\(\\mathbf{S}_{2D\\rightarrow 3D}\\), takes features from one modality as input and attempts to reproduce the output of the teacher network from the other modality:" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.741, + 0.483, + 0.763 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {2 D} ^ {S} = \\mathbf {S} _ {3 D \\rightarrow 2 D} \\big (\\mathbf {f} _ {3 D} ^ {T} \\big), \\quad \\mathbf {f} _ {3 D} ^ {S} = \\mathbf {S} _ {2 D \\rightarrow 3 D} \\big (\\mathbf {f} _ {2 D} ^ {T} \\big). \\qquad (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.77, + 0.466, + 0.785 + ], + "angle": 0, + "content": "The overall distillation loss is applied bidirectionally:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.807, + 0.483, + 0.84 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c r o s s}} = \\left(1 - \\cos \\langle \\mathbf {f} _ {2 D} ^ {T}, \\mathbf {f} _ {2 D} ^ {S} \\rangle\\right) + \\left(1 - \\cos \\langle \\mathbf {f} _ {3 D} ^ {T}, \\mathbf {f} _ {3 D} ^ {S} \\rangle\\right). \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.433, + 0.866 + ], + "angle": 0, + "content": "3.3. Intra-Modal Teacher-Student Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "While the cross-modal student excels in identifying semantic relationships between modalities, it relies on high-level" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.455, + 0.907, + 0.608 + ], + "angle": 0, + "content": "feature representations with low spatial resolution, which may overlook subtle defects. To address this limitation, we introduce an intra-modal student for each modality, which performs intra-modal reconstruction on both feature level and image level. As a result, the intra-modal student specializes in capturing modality-specific fine-grained details for anomaly detection. In particular, the intra-modal student adopts a decoder-only architecture to perform multi-stage reverse distillation, progressively reconstructing back to the original image space for fine-grained anomaly localization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.909, + 0.87 + ], + "angle": 0, + "content": "Multi-Layer Reverse Distillation. To preserve both low-level textures and high-level semantics, we adopt a multilayer feature inversion strategy for the training of intramodal student. Let \\(\\{F_T^k\\}_{k=1}^K\\) denote a set of \\(K\\) feature maps selected from different layers of the teacher's encoder, with layer 1 representing the shallow representation and layer \\(K\\) being the deepest representation. In addition, \\(k = 0\\) denotes the original input image space. Our two intra-modal students, \\(\\mathbf{S}_{2D \\to 2D}\\), \\(\\mathbf{S}_{3D \\to 3D}\\), are designed to reverse the encoding process in a top-down manner. Specifically, we use \\(D^k\\) to denote the decoding module in the \\(k\\)-th layer of an intra-modal student, with each decoder block \\(D^k(\\cdot)\\) reconstructing the next fine-grained feature map: \\(F_S^{k-1} = D^k(F_S^k)\\), where \\(F_S^k\\) denotes the output of the \\(k\\)-th level of the intra-modal student and \\(F_T^K\\) serves as the student's starting point. The final output \\(F_S^0\\) reaches the original image resolution, enabling fine-grained anomaly localization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "We then enforce the consistency between the teacher and the student's predictions to ensure that normal appearance is" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23293" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "reconstructed across different hidden layers. In addition, we incorporate an image-level reconstruction term that compares the final output \\( F_{S}^{0} \\) against the original input image \\( I \\), further preserving fidelity to subtle textures. Concretely, we calculate the intra-modal loss as:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.173, + 0.482, + 0.25 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {i n t r a}} = \\sum_ {(S, T)} \\Big (\\sum_ {k = 1} ^ {K - 1} \\big (1 - \\cos \\langle F _ {S} ^ {k}, F _ {T} ^ {k} \\rangle \\big) + \\left\\| F _ {S} ^ {0} - I \\right\\| ^ {2} \\Big), \\\\ (S, T) \\in \\left\\{\\left(\\mathbf {S} _ {2 D \\rightarrow 2 D}, \\mathbf {T} _ {2 \\mathrm {D}}\\right), \\left(\\mathbf {S} _ {3 D \\rightarrow 3 D}, \\mathbf {T} _ {3 \\mathrm {D}}\\right)\\right\\}. \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.251, + 0.483, + 0.297 + ], + "angle": 0, + "content": "This encourages the intra-modal students to capture both high-level semantics and fine-grained details of teacher's normality representation, even in few-shot settings." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.31, + 0.306, + 0.327 + ], + "angle": 0, + "content": "3.4. Training and Inference" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.332, + 0.484, + 0.409 + ], + "angle": 0, + "content": "Training. Our method jointly trains the cross-modal and intra-modal students using the respective objectives introduced in Sections 3.2 and 3.3. Specifically, we minimize the sum of the cross-modal loss in Eq. (2) and the intramodal loss in Eq. (3), yielding the final objective:" + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.417, + 0.482, + 0.433 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {c r o s s}} + \\mathcal {L} _ {\\text {i n t r a}}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.442, + 0.484, + 0.548 + ], + "angle": 0, + "content": "Training is carried out on a few normal samples. The cross-modal student learns to reconstruct teacher features from the complementary modality, while the intra-modal student focuses on restoring its own modality in a reverse distillation manner. This combined strategy leverages both inter-modal consistency and fine-grained modality-specific cues for robust and fine-grained anomaly detection." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.484, + 0.87 + ], + "angle": 0, + "content": "Anomaly Scoring. At test time, each of the four students \\(\\mathbf{S}_{2D\\rightarrow 3D}\\), \\(\\mathbf{S}_{3D\\rightarrow 2D}\\), \\(\\mathbf{S}_{2D\\rightarrow 2D}\\), \\(\\mathbf{S}_{3D\\rightarrow 3D}\\) produces an anomaly map by measuring the discrepancy between its reconstructed output and the corresponding teacher's feature. In particular, for the intra-modal students \\((S_{2D\\rightarrow 2D}\\) and \\(S_{3D\\rightarrow 3D})\\), the intra-modal anomaly maps \\(\\mathbf{A}_{2D\\rightarrow 2D}\\) and \\(\\mathbf{A}_{3D\\rightarrow 3D}\\) are calculated by aggregating the intra-modal teacher-student discrepancy across multiple layers. For each hidden layer, we additionally apply \\(\\mathrm{L}_2\\) -normalize to both the teacher's feature and the student's reconstructed feature at every spatial position. For each layer, we calculate the Euclidean distance between the student's reconstructed output and the teacher's original output to obtain a pixel-level error map. Next, we apply Gaussian smoothing to the obtained error map for a more coherent anomaly response. We repeat this process for all \\(K - 1\\) hidden layers and the final output layer, and then average all the error maps to obtain the two intra-modal maps \\(\\mathbf{A}_{2D\\rightarrow 2D}\\) and \\(\\mathbf{A}_{3D\\rightarrow 3D}\\) for both 2D and 3D modalities. The final intra-modal anomaly map \\(\\mathbf{A}_{\\mathrm{intra}}\\) is computed by averaging the two intra-modal maps \\(\\mathbf{A}_{2D\\rightarrow 2D}\\) and \\(\\mathbf{A}_{3D\\rightarrow 3D}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The cross-modal students \\((S_{2D\\to 3D}\\) and \\(S_{3D\\to 2D})\\) operate similarly by measuring the discrepancy between the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "student's reconstructed feature in one modality and the teacher's feature in the other modality. The resulting cross-modal anomaly maps from each student are denoted as \\(\\mathbf{A}_{2D\\to 3D}\\) and \\(\\mathbf{A}_{3D\\to 2D}\\). Likewise, we average the two cross-modal maps to obtain a single cross-modal anomaly map \\(\\mathbf{A}_{\\mathrm{cross}}\\). The final overall anomaly map is obtained by the following:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.207, + 0.905, + 0.226 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} _ {\\text {f i n a l}} = \\alpha \\mathbf {A} _ {\\text {c r o s s}} + (1 - \\alpha) \\mathbf {A} _ {\\text {i n t r a}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.233, + 0.905, + 0.325 + ], + "angle": 0, + "content": "where \\(\\alpha\\) adjusts the relative influence of cross-modal and intra-modal anomaly map. High values in \\(\\mathbf{A}_{\\mathrm{final}}\\) indicate regions where the student network fails to mimic the teacher. By combining multi-modal alignment and reverse distillation reconstruction, our framework captures diverse anomaly signatures even in few-shot scenarios." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.337, + 0.64, + 0.354 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.362, + 0.704, + 0.379 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.384, + 0.906, + 0.49 + ], + "angle": 0, + "content": "Datasets and Protocol. We evaluate on two public benchmarks: MVTec-3D-AD [1] and Eyecandies [2]. Following a standard few-shot normal-only protocol, we limit the available normal training images per class to \\( k \\in \\{5, 10, 50\\} \\), randomly sampling and averaging results over ten runs. We also evaluate under a full-shot (all normal samples) setting for completeness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.496, + 0.906, + 0.602 + ], + "angle": 0, + "content": "Baselines. We compare against: (1) BTF [11], a classical 3D feature-based method without deep learning; (2) AST [24], an asymmetric T-S approach trained on normal samples; (3) M3DM [31], a hybrid fusion model exploiting memory banks; (4) CFM [5], which aligns 2D-3D features without memory banks; and (5) 3D-ADNAS [21], which uses neural architecture search for multi-modal fusion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.607, + 0.906, + 0.699 + ], + "angle": 0, + "content": "Evaluation Metrics. We follow standard anomaly detection metrics: (a) I-AUROC for image-level anomaly classification, (b) P-AUROC (pixel-level AUROC) for localization, and (c) AUPRO (area under per-region overlap) at specific false positive rates (FPR), typically \\(1\\%\\) and \\(30\\%\\). Higher values denote better performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. The intra-modal student is implemented as a multi-stage decoder, where each stage consists of three sequential convolutional blocks with normalization layers and GeLU activation to ensure nonlinearity. In particular, the final block employs ConvTranspose2d layers to up-sample the feature map back to the original image resolution and progressively decreases the dimensionality of the output channel to the final 3 channels (RGB). We implement our cross-modal student network as an encoder-decoder style network. The encoder first encodes the input patches into a hidden representation with two convolutional blocks, while the decoder reconstructs them in the cross-modal feature space through another convolutional" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23294" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 1. Comparison of anomaly detection and localization performance on MVTec-3D-AD and Eyecandies with varying numbers of training shots. \\( \\mathbf{{ROC}}{}_{I},{\\mathbf{{ROC}}}_{P} \\) ,and \\( \\mathbf{{PRO}} \\) denote I-AUROC, P-AUROC, and AUPRO respectively. Bold indicates the best performance." + }, + { + "type": "table", + "bbox": [ + 0.174, + 0.129, + 0.825, + 0.479 + ], + "angle": 0, + "content": "
k-ShotMethodMVTec-3D-ADEyecandies
\\( ROC_I \\)\\( ROC_P \\)PRO-30%PRO-1%\\( ROC_I \\)\\( ROC_P \\)PRO-30%PRO-1%
5BTF [11]0.6710.9800.9200.2880.6520.8150.7380.202
AST [24]0.6800.9500.9030.1580.6330.7410.6910.143
M3DM [31]0.8220.9840.9370.3300.7640.8710.8070.251
CFM [5]0.8110.9860.9490.3820.7950.8790.8010.260
3D-ADNAS [21]0.826---0.7750.875--
FIND (Ours)0.8990.9910.9630.4200.8400.9510.8620.324
10BTF [11]0.6950.9830.9280.3080.6850.8340.8060.217
AST [24]0.6890.9460.8350.1740.6710.7670.6240.161
M3DM [31]0.8450.9860.9430.3550.8240.8900.8120.266
CFM [5]0.8450.9870.9540.3980.8380.9030.8250.284
3D-ADNAS [21]0.848---0.8070.869--
FIND (Ours)0.9210.9930.9670.4340.8680.9590.8890.350
50BTF [11]0.8060.9890.9470.3560.7210.8560.8240.240
AST [24]0.7940.9740.9290.3350.7390.8620.7150.118
M3DM [31]0.9070.9890.9550.3870.8360.9330.8460.304
CFM [5]0.9060.9910.9650.4310.8520.9260.8510.315
3D-ADNAS [21]0.890---0.8680.912--
FIND (Ours)0.9520.9950.9730.4550.8970.9760.9070.365
ALLBTF [11]0.8650.9920.9590.3830.7400.8830.8450.261
AST [24]0.9370.9760.9440.3980.7800.9020.7440.149
M3DM [31]0.9450.9920.9640.3940.8820.9770.8870.331
CFM [5]0.9540.9930.9710.4550.8810.9740.8870.335
3D-ADNAS [21]0.951---0.9460.970--
FIND (Ours)0.9780.9950.9830.4560.9550.9830.9250.395
" + }, + { + "type": "table_caption", + "bbox": [ + 0.23, + 0.49, + 0.766, + 0.506 + ], + "angle": 0, + "content": "Table 2. Comparisons of per-category anomaly detection performance on MVTec-3D-AD." + }, + { + "type": "table", + "bbox": [ + 0.153, + 0.516, + 0.845, + 0.764 + ], + "angle": 0, + "content": "
MethodBagelCable GlandCarrotCookieDowelFoamPeachPotatoRopeTireMean
I-AUROC
BTF [11]0.9380.7650.9720.8880.9600.6640.9040.9290.9820.7260.865
AST [24]0.9830.8730.9760.9710.9320.8850.9740.9811.0000.7970.937
M3DM [31]0.9940.9090.9720.9760.9600.9420.9730.8990.9720.8500.945
CFM [5]0.9940.8880.9840.9930.9800.8880.9410.9430.9800.9530.954
3D-ADNAS [21]0.9971.0000.9710.9860.9660.9480.8970.8731.0000.8670.951
Shape Guided [4]0.9860.8940.9830.9910.9760.8570.9900.9650.9900.8690.947
LSFA [29]1.0000.9390.9820.9890.9610.9510.9830.9620.9890.9510.971
FIND (Ours)0.9840.9700.9850.9720.9820.9790.9740.9800.9880.9700.978
AUPRO@30%
BTF [11]0.9760.9690.9790.9730.9330.8880.8960.9120.9500.9710.959
AST [24]0.9700.9470.9810.9390.9130.9060.9790.9820.8890.9400.944
M3DM [31]0.9700.9710.9790.9500.9410.9320.9770.9710.9710.9750.964
CFM [5]0.9790.9720.9820.9450.9500.9680.9800.9430.9500.9810.971
Shape Guided [4]0.9810.9730.9820.9710.9620.9780.9810.9830.9740.9750.976
LSFA [29]0.9860.9740.9810.9460.9250.9410.9830.9830.9740.9830.968
FIND (Ours)0.9890.9770.9810.9830.9820.9860.9840.9800.9900.9780.983
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.788, + 0.483, + 0.818 + ], + "angle": 0, + "content": "blocks, ensuring both preservation of the locality and improved alignment between the modalities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "To ensure a fair comparison with previous work, we use the same ViT-B/8 DINO backbone as CFM [5], which is shared for both RGB and surface normal modalities. We treat surface normals as 3-channel RGB-style inputs and normalize them using ImageNet statistics before feeding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.788, + 0.907, + 0.895 + ], + "angle": 0, + "content": "them into DINO. During training, the backbone weights remain fixed, and we optimize only the four student networks. We extract hidden features from teacher's 4, 6, 8-th layers for intra-modal reverse distillation, while the 10-th layer's feature is employed as the input for cross-modal student. We use the ADAM optimizer for training with a learning rate of \\(1 \\times 10^{-4}\\), and set \\(\\alpha = 0.5\\) for all-shot tasks and" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23295" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.092, + 0.478, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.258, + 0.476, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.406, + 0.476, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.561, + 0.483, + 0.645 + ], + "angle": 0, + "content": "Figure 4. Qualitative results on MVTec-3D-AD with various types of anomalies. From left to right: input RGB images, surface normals, GT anomaly maps, anomaly maps from 2D student set, anomaly maps from 3D student set, and our final predictions. Our final model consistently produces more accurate anomaly maps across different anomaly types." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.658, + 0.483, + 0.747 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.757, + 0.483, + 0.8 + ], + "angle": 0, + "content": "Figure 5. Anomaly score distributions on bagel and cookie (MVTec-3D-AD, 10-shot). Our method shows better separation between normal and anomalous samples." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.826, + 0.274, + 0.839 + ], + "angle": 0, + "content": "\\(\\alpha = 0.8\\) for few-shot tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "For MVTec-3D-AD, each sample includes an RGB image and an unstructured point cloud. We first organize the point cloud into a 2D grid(XYZ map) corresponding to the rgb image size \\( H \\times W \\). We then compute per-pixel surface" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.09, + 0.905, + 0.319 + ], + "angle": 0, + "content": "
CategoryIntra-studentCross-studentIntra + Cross
ImagePixelImagePixelImagePixel
Bagel96.399.798.899.599.599.6
Cable Gland74.497.679.098.382.498.3
Carrot97.499.899.499.997.399.8
Cookie86.498.491.897.292.597.8
Dowel75.298.183.899.382.699.2
Foam78.999.197.699.898.199.8
Peach97.099.895.799.999.499.9
Potato92.999.892.799.995.899.9
Rope97.699.896.699.898.899.8
Tire66.798.775.199.274.499.2
Average86.399.191.199.392.199.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.329, + 0.907, + 0.371 + ], + "angle": 0, + "content": "Table 3. Ablation studies on intra-modal, cross-modal and dual students. We report the 10-shot categorical performance in image and pixel AUROC (\\%) on the MVTec-3D dataset." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.406, + 0.913, + 0.637 + ], + "angle": 0, + "content": "
CategoryDepthPoint CloudSurface Normal
ImagePixelImagePixelImagePixel
Bagel99.099.498.399.599.599.6
Cable Gland68.897.774.197.682.498.3
Carrot82.799.392.499.197.399.8
Cookie84.498.590.797.292.597.8
Dowel87.199.075.599.082.699.2
Foam81.689.892.499.598.199.8
Peach85.799.393.799.699.499.9
Potato68.499.184.398.795.899.9
Rope95.799.694.299.798.899.8
Tire66.897.673.699.374.499.2
Average82.097.986.898.992.199.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.645, + 0.907, + 0.687 + ], + "angle": 0, + "content": "Table 4. Ablation studies on 3D modality input selection. We report the 10-shot categorical performance in image and pixel AUROC (\\%) on the MVTec-3D dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.903 + ], + "angle": 0, + "content": "normals by fitting a local plane to the neighbors of each point (or using the cross-product of adjacent triangle faces in a mesh representation). This produces a 3-channel normal vector map for the surface of the object. Following prior work [5, 31], we remove background points and fill the removed regions in the RGB image with a neutral value (black) to avoid false anomalies from the background. We finally resize both the RGB image and the normal map to \\(224 \\times 224\\) pixels before feeding them to the model. For Eyecandies, we use normal maps with global lighting as the 3D modality." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23296" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.254, + 0.108 + ], + "angle": 0, + "content": "4.2. Primary Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.283 + ], + "angle": 0, + "content": "Few-shot Performance on MVtec-3D-AD. Table 1 shows that FIND exceeds all competitors: BTF, AST, M3DM, CFM and 3D-ADNAS, in all few-shot settings (5, 10, 50 shots) and evaluation metrics at both the image and pixel levels. For example, in the challenging 5-shot setting, our method achieves an I-AUROC of 0.899. With 10-shot training, the I-AUROC performance is increased to 0.921. Even with 50 shots, FIND continues to outperform strong memory-based methods (e.g., M3DM). These consistent gains highlight the effectiveness of FIND in limited data scenarios." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.29, + 0.485, + 0.442 + ], + "angle": 0, + "content": "Full-shot Performance on MVtec-3D-AD. Table 1 presents the anomaly detection performance in the full-shot setting at both the image and pixel levels, while Table 2 reports results for each category. FIND surpasses existing methods in both image-level and pixel-level detection, achieving the highest average I-AUROC \\((97.8\\%)\\) and AUPRO@ \\(30\\%\\) \\((98.3\\%)\\), demonstrating its robustness across various object classes. Notably, our approach excels in challenging categories (e.g., Cable Gland, Foam), indicating its ability to capture fine-grained anomalous patterns." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.45, + 0.485, + 0.648 + ], + "angle": 0, + "content": "Results on Eyecandies. As shown in Table 1, FIND also sets a new state-of-the-art on Eyecandies under both few-shot and full-shot protocols, significantly outperforming existing methods. For example, at 5-shot, FIND achieves an I-AUROC score of 0.840, exceeding the other baselines significantly. This performance gain extends to pixel-level metrics (P-AUROC and AUPRO), demonstrating a superior localization of subtle anomalies. Furthermore, with 50 normal samples for training, our method even outperforms the full-shot trained M3DM and CFM in terms of AUROC and AUPRO. These results highlight the effectiveness of our approach in various shots of training data scenarios, reinforcing its robustness and generalizability." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.663, + 0.285, + 0.68 + ], + "angle": 0, + "content": "4.3. Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.687, + 0.485, + 0.809 + ], + "angle": 0, + "content": "Fig. 4 shows that our method is robust to various types of anomaly (shape-only, color-only, combined) by effectively exploiting 2D-3D complementarity. The 2D student set \\((\\mathbf{S}_{3D\\rightarrow 2D},\\mathbf{S}_{2D\\rightarrow 2D})\\) captures subtle color anomalies often overlooked by geometric cues, while the 3D student set \\((\\mathbf{S}_{2D\\rightarrow 3D},\\mathbf{S}_{3D\\rightarrow 3D})\\) detects shape deformations that are imperceptible in RGB. Combining these maps improves localization accuracy and reduces false positives." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Fig. 5 further compares the anomaly score distributions of normal and anomalous samples under the 10-shot setting for CFM and FIND on two MVTec-3D-AD classes (bagel and cookie). Our method yields more clearly separable distributions between normal and abnormal samples, contributing to improved detection accuracy." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.68, + 0.107 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.113, + 0.907, + 0.371 + ], + "angle": 0, + "content": "Intra-modal vs. Cross-modal vs. Dual Student. We evaluate the contributions of the intra-modal and cross-modal students, with results shown in Table 3. Intra-modal students capture modality-specific anomalies and guide the model to reproduce a normal version of the query sample at the pixel level, enhancing the detection of local defects. Removing them causes a notable drop in detection accuracy for categories such as cable gland, potato, and rope. However, intra-modal students operating within a single modality cannot leverage complementary information from the other modality, leaving those cross-modal anomalies undetected. In contrast, cross-modal students learn the expected normal correspondence between an object's color and shape appearance and detect anomalies by identifying inconsistencies between modalities. By combining both intra-modal and cross-modal students, we leverage their complementary strengths and achieve the best results." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.376, + 0.909, + 0.619 + ], + "angle": 0, + "content": "3D Modality Input Choices. In a 10-shot setting on MVTec-3D-AD, we compare raw point clouds (with a PointNet encoder), depth maps(from point clouds), and surface-normal maps for anomaly detection. The results are shown in Table 4. Surface normals achieve the highest I-AUROC, outperforming depth maps and raw point clouds. This indicates that normal maps preserve fine-grained geometric details that can highlight small defects. In contrast, depth maps lack this level of detail, resulting in lower accuracy. The point cloud baseline performs worst, likely due to a modality gap introduced by using a separate PointNet-based 3D feature extractor. Such gaps arise from misaligned representations when using a point cloud network versus image-based features. Our results show that surface normal maps provide the most effective shape modality for 3D anomaly detection in few-shot scenarios." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.631, + 0.634, + 0.647 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.657, + 0.909, + 0.825 + ], + "angle": 0, + "content": "In this work, we address the critical challenges of multimodal anomaly detection in few-shot scenarios by introducing FIND, a dual-student framework that integrates intramodal reverse distillation and cross-modal feature mapping. FIND overcomes data scarcity and modality gaps by exploiting collaborative learning between two students: one dedicated to fine-grained intra-modal anomaly localization through reverse distillation, and another focused on cross-modal inconsistency detection. Extensive experiments demonstrate that FIND achieves new state-of-the-art results in both few-shot and full-shot settings." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.837, + 0.683, + 0.854 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.908, + 0.903 + ], + "angle": 0, + "content": "This work is supported by the Agency for Science, Technology and Research (A*STAR) under its MTC Programmatic Funds (Grant No. M23L7b0021)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23297" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.198 + ], + "angle": 0, + "content": "[1] Paul Bergmann, Xin Jin, David Sattlegger, and Carsten Steger. The mvtec 3d-ad dataset for unsupervised 3d anomaly detection and localization. In Proceedings of the 17th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications, VISIGRAPP, pages 202-213, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.201, + 0.484, + 0.27 + ], + "angle": 0, + "content": "[2] Luca Bonfiglioli, Marco Toschi, Davide Silvestri, Nicola Fioraio, and Daniele De Gregorio. The eyecandies dataset for unsupervised multimodal anomaly detection and localization. In 16th Asian Conference on Computer Vision (ACCV), pages 459-475, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.272, + 0.484, + 0.327 + ], + "angle": 0, + "content": "[3] Mengxi Chen and ... Enhanced multimodal representation learning with cross-modal KD. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11766-11775, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.33, + 0.483, + 0.371 + ], + "angle": 0, + "content": "[4] Yu-Min Chu, Liu Chieh, Ting-I Hsieh, Hwann-Tzong Chen, and Tyng-Luh Liu. Shape-guided dual-memory learning for 3d anomaly detection. 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.374, + 0.484, + 0.442 + ], + "angle": 0, + "content": "[5] Alex Costanzino, Pierluigi Zama Ramirez, Giuseppe Lisanti, and Luigi Di Stefano. Multimodal industrial anomaly detection by crossmodal feature mapping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17234-17243, 2024. 1, 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.445, + 0.484, + 0.5 + ], + "angle": 0, + "content": "[6] Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one-class embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9737-9746, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.503, + 0.484, + 0.558 + ], + "angle": 0, + "content": "[7] Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one-class embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9737-9746, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.561, + 0.483, + 0.629 + ], + "angle": 0, + "content": "[8] Zheng Fang, Xiaoyang Wang, Haocheng Li, Jiejie Liu, Qiugui Hu, and Jimin Xiao. Fastrecon: Few-shot industrial anomaly detection via fast feature reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17481-17490, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.632, + 0.483, + 0.713 + ], + "angle": 0, + "content": "[9] Zhihao Gu, Jiangning Zhang, Liang Liu, Xu Chen, Jinlong Peng, Zhenye Gan, Guannan Jiang, Annan Shu, Yabiao Wang, and Lizhuang Ma. Rethinking reverse distillation for multi-modal anomaly detection. In Thirty-Eighth AAAI Conference on Artificial Intelligence, AAAI, pages 8445-8453, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.717, + 0.483, + 0.771 + ], + "angle": 0, + "content": "[10] Yu Hong, Hang Dai, and Yong Ding. Cross-modality knowledge distillation network for monocular 3d object detection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 87-104, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.483, + 0.842 + ], + "angle": 0, + "content": "[11] Eliahu Horwitz and Yedid Hoshen. Back to the feature: classical 3d features are (almost) all you need for 3d anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2968-2977, 2023. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Chaoqin Huang, Haoyan Guan, Aofan Jiang, Ya Zhang, Michael Spratling, and Yan-Feng Wang. Registration based few-shot anomaly detection. In European Conference on Computer Vision, pages 303-319. Springer, 2022. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.162 + ], + "angle": 0, + "content": "[13] Fushuo Huo, Wenchao Xu, Jingcai Guo, Haozhao Wang, and Song Guo. \\(C^2\\)kd: Bridging the modality gap for cross-modal knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7151-7160, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.165, + 0.906, + 0.233 + ], + "angle": 0, + "content": "[14] Jongheon Jeong, Yang Zou, Taewan Kim, Dongqing Zhang, Avinash Ravichandran, and Onkar Dabeer. Winclip: Zero/few-shot anomaly classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19606-19616, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.235, + 0.908, + 0.29 + ], + "angle": 0, + "content": "[15] Pilhyeon Lee and ... Decomposed cross-modal distillation for RGB-based temporal action detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages -, 2023. CVPR 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.292, + 0.906, + 0.346 + ], + "angle": 0, + "content": "[16] Yujin Lee, Seoyoon Jang, and Hyunsoo Yoon. Anople: Few-shot anomaly detection via bi-directional prompt learning with only normal samples. arXiv preprint arXiv:2408.13516, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.349, + 0.906, + 0.445 + ], + "angle": 0, + "content": "[17] Wenqiao Li, Xiaohao Xu, Yao Gu, Bozhong Zheng, Shenghua Gao, and Yingna Wu. Towards scalable 3d anomaly detection and localization: A benchmark via 3d anomaly synthesis and a self-supervised learning network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22207-22216, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[18] Hanzhe Liang, Guoyang Xie, Chengbin Hou, Bingshu Wang, Can Gao, and Jinbao Wang. Look inside for more: Internal spatial modality perception for 3d anomaly detection. arXiv preprint arXiv:2412.13461, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.505, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[19] Victor Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Y Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. Advances in Neural Information Processing Systems, 35:17612-17625, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.576, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[20] Chieh Liu, Yu-Min Chu, Ting-I Hsieh, Hwann-Tzong Chen, and Tyng-Luh Liu. Learning diffusion models for multi-view anomaly detection. In European Conference on Computer Vision, pages 328-345. Springer, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.633, + 0.906, + 0.701 + ], + "angle": 0, + "content": "[21] Kaifang Long, Guoyang Xie, Lianbo Ma, Jiaqi Liu, and Zhichao Lu. Revisiting multimodal fusion for 3d anomaly detection from an architectural perspective. In Thirty-Eighth AAAI Conference on Artificial Intelligence, AAAI, 2025. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.704, + 0.906, + 0.773 + ], + "angle": 0, + "content": "[22] Xiang Ma, Xuemei Li, Lexin Fang, and Caiming Zhang. Bridging the modality gap: Dimension information alignment and sparse spatial constraint for image-text matching. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 5074-5082, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.775, + 0.906, + 0.858 + ], + "angle": 0, + "content": "[23] Nicolae-Cătălin Ristea, Neelu Madan, Radu Tudor Ionescu, Kamal Nasrollahi, Fahad Shahbaz Khan, Thomas B. Moeslund, and Mubarak Shah. Self-supervised predictive convolutional attentive block for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13576-13586, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] Marco Rudolph, Tom Wehrbein, Bodo Rosenhahn, and Bastian Wandt. Asymmetric student-teacher networks for industrial anomaly detection. In IEEE/CVF Winter Conference on" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23298" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "Applications of Computer Vision, WACV, pages 2592-2602, 2023. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[25] Pritam Sarkar and Ali Etemad. XKD: Cross-modal knowledge distillation with domain alignment for video representation learning. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2024. to appear. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.22 + ], + "angle": 0, + "content": "[26] Wenbo Sui, Daniel Lichau, Josselin Lefèvre, and Harold Philippeau. Incomplete multimodal industrial anomaly detection via cross-modal distillation, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.222, + 0.482, + 0.305 + ], + "angle": 0, + "content": "[27] Long Tian, Hongyi Zhao, Ruiying Lu, Rongrong Wang, Yu-jie Wu, Liming Wang, Xiongpeng He, and Xiyang Liu. Foct: Few-shot industrial anomaly detection with foreground-aware online conditional transport. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6241-6249, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.482, + 0.39 + ], + "angle": 0, + "content": "[28] Dinh Tien Tran, Anh Tuan Nguyen, Nguyen Hoang Tran, Ta Duc Huy, Soan T. M. Duong, Chanh D. Tr. Nguyen, and Steven Q. H. Truong. Revisiting reverse distillation for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24511-24520, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.482, + 0.461 + ], + "angle": 0, + "content": "[29] Yuanpeng Tu, Boshen Zhang, Liang Liu, Yuxi Li, Jiangning Zhang, Yabiao Wang, Chengjie Wang, and Cairong Zhao. Self-supervised feature adaptation for 3d industrial anomaly detection. In European Conference on Computer Vision, pages 75-91. Springer, 2025. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.463, + 0.482, + 0.518 + ], + "angle": 0, + "content": "[30] Huiqun Wang, Yiping Bao, Panwang Pan, Zeming Li, Xiao Liu, Ruijie Yang, and Di Huang. Multi-modal relation distillation for unified 3d representation learning. In European Conference on Computer Vision (ECCV), 2024. to appear. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.52, + 0.482, + 0.589 + ], + "angle": 0, + "content": "[31] Yue Wang, Jinlong Peng, Jiangning Zhang, Ran Yi, Yabiao Wang, and Chengjie Wang. Multimodal industrial anomaly detection via hybrid fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8032-8041, 2023. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.591, + 0.482, + 0.659 + ], + "angle": 0, + "content": "[32] Zeyu Wang, Dingwen Li, Chenxu Luo, Cihang Xie, and Xiaodong Yang. DistillBEV: Boosting multi-camera 3d object detection with cross-modal knowledge distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8637-8646, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.729 + ], + "angle": 0, + "content": "[33] Zihui Xue, Zhengqi Gao, Sucheng Ren, and Hang Zhao. The modality focusing hypothesis: Towards understanding cross-modal knowledge distillation. In International Conference on Learning Representations (ICLR), 2023. Top \\(5\\%\\) at ICLR 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[34] Xu Yan, Heshen Zhan, Chaoda Zheng, Jiantao Gao, Ruimao Zhang, Shuguang Cui, and Zhen Li. Let images give you more: Point cloud cross-modal training for shape analysis. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[35] Heeseung Yun, Joonil Na, and Gunhee Kim. Dense 2d-3d indoor prediction with sound via aligned cross-modal distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7829-7838, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[36] Vitjan Zavrtanik, Matej Kristan, and Danijel Skocaj. Cheating depth: Enhancing 3d surface anomaly detection via depth" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "simulation. In IEEE/CVF Winter Conference on Applications of Computer Vision, WACV, pages 2153-2161. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[37] Tianlu Zhang, Hongyuan Guo, Qiang Jiao, Qiang Zhang, and Jungong Han. Efficient rgb-t tracking via cross-modality distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5815-5824, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.907, + 0.275 + ], + "angle": 0, + "content": "[38] Xuan Zhang, Shiyu Li, Xi Li, Ping Huang, Jiulong Shan, and Ting Chen. Destseg: Segmentation guided denoising student-teacher for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3914–3923, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.358 + ], + "angle": 0, + "content": "[39] Shengchao Zhou, Weizhou Liu, Chen Hu, Shuchang Zhou, and Chao Ma. Unidistill: A universal cross-modality knowledge distillation framework for 3d object detection in bird's-eye view. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5116-5125, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.361, + 0.905, + 0.416 + ], + "angle": 0, + "content": "[40] Zheyuan Zhou, Le Wang, Naiyu Fang, Zili Wang, Lemiao Qiu, and Shuyou Zhang. R3d-ad: Reconstruction via diffusion for 3d anomaly detection. In European Conference on Computer Vision, pages 91-107. Springer, 2025. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23299" + } + ] +] \ No newline at end of file diff --git a/2025/FIND_ Few-Shot Anomaly Inspection with Normal-Only Multi-Modal Data/94dd2e05-5a4d-4859-bd44-9496d30c1756_origin.pdf b/2025/FIND_ Few-Shot Anomaly Inspection with Normal-Only Multi-Modal Data/94dd2e05-5a4d-4859-bd44-9496d30c1756_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..43aac26cff2a8716441c8e851dce7c28afdf0c8e --- /dev/null +++ b/2025/FIND_ Few-Shot Anomaly Inspection with Normal-Only Multi-Modal Data/94dd2e05-5a4d-4859-bd44-9496d30c1756_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eda0475d24ba7482568f3045d21dca386d481c84f1ecbc9c86ed5a4216a4744e +size 28956473 diff --git a/2025/FIND_ Few-Shot Anomaly Inspection with Normal-Only Multi-Modal Data/full.md b/2025/FIND_ Few-Shot Anomaly Inspection with Normal-Only Multi-Modal Data/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e8bb84bf129e948360d36ca3d0ff5c510aee5668 --- /dev/null +++ b/2025/FIND_ Few-Shot Anomaly Inspection with Normal-Only Multi-Modal Data/full.md @@ -0,0 +1,246 @@ +# FIND: Few-Shot Anomaly Inspection with Normal-Only Multi-Modal Data + +Yiting Li $^{1}$ , Fayao Liu $^{1\dagger}$ , Jingyi Liao $^{1}$ , Sichao Tian $^{2}$ , Chuan-Sheng Foo $^{1}$ , Xulei Yang $^{1\dagger}$ $^{1}$ Institute for Infocomm Research, A*STAR, Singapore $^{2}$ Chinese Academy of Medical Sciences and Peking Union Medical College, China +{li_yiting, liu_fayao, liao_jingyi, foo_chuan_sheng, yang_xulei} @i2r.a-star.edu.sg, {sctian} @implad.ac.cn + +# Abstract + +Multi-modal anomaly detection (MAD) improves industrial inspection by exploiting complementary 2D and 3D data. However, existing methods struggle in few-shot scenarios due to limited data and modality gaps. Current approaches either fuse multi-modal features or align cross-modal representations; however, they often suffer from high false-positive rates and fail to detect subtle defects, especially when training samples are scarce. To address these challenges, we propose the first few-shot MAD method FIND, a novel dual-student framework that integrates intramodal reverse distillation and cross-modal feature mapping. FIND employs modality-specific teachers and two collaborative students: an intra-modal student for fine-grained anomaly localization via reverse distillation, and a cross-modal student that captures inter-modal correspondences to detect cross-modal inconsistencies. Extensive experiments on MVTec-3D-AD and Eyecandies show that FIND significantly outperforms state-of-the-art methods in both full-shot and few-shot settings. Ablation studies validate the complementary roles of intra- and cross-modal distillation. Our work significantly advances MAD robustness in data-scarce industrial applications. + +# 1. Introduction + +Multi-modal anomaly detection (MAD) has garnered increasing interest in industrial inspection [4, 17, 18, 20, 21, 26, 31, 34, 40], as relying on a single modality (e.g., RGB) may fail to capture crucial geometric cues under varying surface textures. By integrating heterogeneous data such as point clouds or depth maps, MAD methods can leverage multi-modal complementarity to enhance inspection quality. Since anomalous data are rare, most existing anomaly detection (AD) methods are trained only on normal data. Our work follows this protocol. + +While significant progress has been made in MAD, exist + +![](images/f0cbbd1f2cc09eb4ec229d1a47d39b1ce2c60b7a00d469d6a98d1fa8e28946f4.jpg) +Figure 1. Comparison with state-of-the-art methods on MVTec-3D-AD in terms of I-AUROC (left) and AUPRO@1% (right). FIND substantially outperforms prior methods (BTF [11], AST [24], M3DM [31], 3D-ADNAS [21] and CFM [5]) in both few-shot and full-shot settings. + +![](images/79b2b557084b07c5b10597bbffcd215dc41f8a5d4e9b7b5ef38b7b4ef10a39c7.jpg) + +ing approaches struggle to retain robustness in few-shot scenarios, where limited normal multi-modal training data is available—a critical limitation given the high cost of acquiring such data [8, 12, 14, 16, 27]. This is evidenced in Fig. 1, which compares various methods using different amounts of normal training samples, with notable performance drops observed. In this work, we propose FIND, a simple yet effective dual-student MAD framework focused on Few-Shot Anomaly Inspection with Normal-Only Multi-Modal Data. It effectively minimizes modality interference while exploiting multi-modal complementarity, significantly enhancing anomaly detection and localization. + +In single-modality anomaly detection, teacher-student (T-S) frameworks achieve strong results [6, 23, 28, 38]. A pretrained teacher extracts generic features, while a student trained on normal data mimics the teacher. Anomalies emerge as discrepancies between their features at inference. Reverse distillation [7, 28] refines this by reconstructing teacher features from compact one-class embeddings, enabling fine-grained defect localization. + +However, extending the success of T-S frameworks to MAD remains a challenge. Existing methods either fuse modalities by concatenating RGB and depth at the input level [24] or across multiple feature levels [9], which risks modality interference. Specifically, MMRD [9] introduces + +a multi-modal reverse distillation paradigm that performs a weighted fusion of image and depth data across multiple feature levels, creating a multi-modal teacher target for a learnable student decoder tasked with restoring the teacher's multi-modal representations. In contrast, cross-modal Feature Mapping (CFM) [5], although not explicitly formulated within the T-S framework, uses domain-specific teachers for cross-modal distillation and performs late fusion on score maps instead of relying on feature fusion. The features of different modalities are aligned through an MLP-based projection network. CFM trains two student models to map features from one modality to another by mimicking the corresponding target modality teacher. By learning the expected correspondence between an object's color and shape, any deviation from this correspondence at test time indicates an anomaly. However, the inherent modality gap can lead to large reconstruction errors even for normal samples, resulting in high false positive rates and limiting the model's ability to distinguish genuine anomalies. These limitations result in high false positives and coarse localization (see Fig. 2 for examples), which are further exacerbated in few-shot scenarios, where limited normal samples hinder the effective alignment of the two modalities. + +To address the challenges stated above, we present FIND, which unifies intra-modal reverse distillation and cross-modal distillation within a single T-S framework. Our intuition is to minimize modality interference while exploiting the complementary strengths of multiple modalities. Specifically, each modality is paired with a modality-specific teacher, an intra-modal student, and a cross-modal student. The intra-modal student is trained via reverse distillation to model fine-grained, modality-specific normal appearances, while the cross-modal student is designed to capture inter-modal correspondences for robust anomaly detection. This decoupled design allows each student to focus on its specific task: the intra-modal student is highly sensitive to prominent modality-specific anomalies and provides finer-grained localization, while the cross-modal student detects anomalies that disrupt the learned normal correspondence between different modalities. This synergy enables precise localization and robustness to various types of anomalies, even with limited training data (see Fig. 2 for comparisons). To further mitigate the modality gap and thus reduce false positives in cross-modal distillation, we propose using surface normals as a geometric proxy, enabling the use of shared RGB and 3D encoders. We find this simple strategy to be highly effective in achieving accurate MAD. Our contributions are summarized as follows: + +- We propose the first few-shot MAD method FIND, which synergistically integrates intra-modal reverse distillation and cross-modal distillation. By decoupling the learning objectives for intra- and cross-modal students, FIND significantly enhances MAD performance. + +![](images/ff2fcea15e646f283f76ecb9b49af151a96db198ed4a9fbdd816ba8cd45acf7b.jpg) +Figure 2. Compared to CFM under the 10-shot setting, our anomaly maps are sharper and show fewer false positives, focusing on anomalous patterns rather than minor texture variations. + +- We show that surface normals are a highly effective geometric modality for MAD, effectively bridging the modality gap in cross-modal distillation, reducing false positives, and enhancing anomaly detection precision. +- FIND sets a new state-of-the-art in both multi-modal anomaly detection and localization tasks, outperforming existing methods in both full-shot and few-shot settings. + +# 2. Related works + +Multi-modal Anomaly Detection With the development of 3D sensors, multi-modal anomaly detection has gained increasing attention [4, 9, 21, 36], as 3D information (e.g., point clouds) offers richer geometric cues and captures shape deformations that may be invisible in RGB domain. Recently, researchers have explored cross-modal alignment methods that fuse RGB and 3D data [24, 31] for more robust anomaly localization. For instance, CFM [5] proposes a novel framework that learns to map 2D and 3D features to each other, thereby pinpointing anomalies where the two modalities disagree; [29] introduce a local-to-global self-supervised adaptation mechanism to finetune pre-trained 2D/3D networks for more precise cross-modal alignment; [31] design a hybrid fusion scheme with multiple memory banks to achieve consistent performance gains over single-modality baselines. In parallel, memory-based 3D AD approaches [21, 31] have also shown strong potential by storing normal embeddings from both color and geometry in dedicated banks and measuring anomalous deviations at test time. For example, [4] propose a shape-guided dual-memory method that leverages signed distance fields (SDF) and RGB cues for complementary anomaly localization; [18] advocate an internal spatial modality perception module to distill intrinsic 3D structures into a memory + +bank; [26] handle incomplete multi-modal data via cross-modal distillation and memory retrieval. Although existing methods have significantly advanced the state of the art in MAD, most still rely on either large-scale normal training sets or carefully aligned networks for each modality. In scenarios where only a few normal samples are available, the modality gap between 2D images and 3D point clouds can hinder reliable cross-modal feature alignment, thus limiting the models' ability to detect subtle anomalies. + +Few-shot Anomaly Detection In resource-constrained scenarios, collecting large amount of normal data can be both costly and impractical. Consequently, few-shot anomaly detection (FSAD) becomes a promising alternative. Current research on FSAD has primarily focused on RGB domains. AnoPLe [16] proposes bi-directional prompt learning with only normal samples, leveraging a pretrained vision-language model to detect anomalies from both textual and visual cues. FastRecon [8] proposes a lightweight feature-based regression with distribution regularization to obtain the optimal transformation from support to query features, it reconstructs a normal version of test sample's features from visually similar support features and detects any large deviations as anomalies. WinCLIP [14] explores a window-based feature fusion strategy to extract multi-scale features for zero-/few-shot anomaly classification and segmentation. FCOT [27] leverages a distribution-based Conditional Transport metric to minimize feature discrepancies between a few support images and query images, while updating a memory bank to precisely match anomalous regions. RegAD [12] uses image registration to align query images with support images in a category-agnostic manner. Despite remarkable progress, existing FSAD approaches still face limitations when extended to multi-modal scenarios due to the lack of capabilities for capturing the shape distortions that 3D data would reveal. This limitation motivates our work on multi-modal anomaly detection under few-shot settings, by capturing shape-induced defects that 2D-only methods may overlook. + +Cross-modal knowledge distillation Cross-modal knowledge distillation (KD) has attracted increasing interest for transferring knowledge between heterogeneous modalities (e.g. images, LiDAR, thermal, audio) [3, 15, 25, 30, 33]. A core challenge is the modality gap issue due to the discrepancies in data/feature distribution between modalities [13, 19, 22]. [19] demonstrates that modality gap in multimodal models such as CLIP is caused by a combination of model initialization and contrastive learning optimization. Recent works tackle this by aligning feature representations and customizing distillation strategies to each modality pair. For example, a LiDAR-based teacher can guide a camera-based student to improve monocular detection accuracy [10, 15, 32]. C2KD [13] propose a bidirectional distillation framework where teacher and student + +learn from each other, coupled with on-the-fly sample filtering to discard misaligned targets. Cross-modal KD is applied in RGB-Thermal (RGB-T) tracking [37, 39], a two-stream RGB-T network can serve as a powerful teacher to a lightweight one-stream tracker; distilling both modality-specific and shared features helps the student network approach the teacher's performance, while greatly reducing computation. In audio-visual learning [35], visual models have been used to teach audio models to predict spatial or semantic information from sound, often by enforcing feature alignments to handle the modality gap. + +# 3. Method + +# 3.1. Framework Overview + +FIND consists of two single-teacher-dual-student branches designed to detect anomalies from multi-modal 2D and 3D inputs. Fig. 3 shows an overview of our method. Each branch is composed of a modality-specific teacher $(\mathbf{T}_{2\mathrm{D}}$ or $\mathbf{T}_{3\mathrm{D}})$ , a cross-modal distillation student $(\mathbf{S}_{3D\rightarrow 2D}$ or $\mathbf{S}_{2D\rightarrow 3D})$ , and an intra-modal reverse distillation student $(\mathbf{S}_{2D\rightarrow 2D}$ or $\mathbf{S}_{3D\rightarrow 3D})$ . Our motivation is to integrate cross-modal distillation with reverse distillation in a unified framework that minimizes modality interference while effectively exploiting their complementarity for anomaly detection. In essence, the Cross-Modal Student distills knowledge across the RGB and 3D streams, learning how normal features in one modality correspond to those in the other. The Intra-Modal Student, on the other hand, learns to reproduce the normal pattern of individual modality through reverse distillation. During training, both students receive supervision derived from a shared set of teacher features so that each modality learns a more comprehensive representation. + +# 3.2. Cross-modal Teacher-Student Learning + +We aim to leverage the complementary knowledge between two different modalities, namely 2D and 3D, for anomaly detection through a cross-modal distillation procedure. + +Bridging the Modality Gap. Since features extracted by 2D backbones for RGB images and 3D point-based networks for point clouds follow different distributions, existing MAD methods face the modality gap issue due to the use of separate feature backbones for each modality. This discrepancy becomes even more significant in few-shot learning scenarios, where a student model with limited training data struggles to align with the teacher's normal feature distribution, leading to a high false positive rate in detection performance. To bridge the modality gap between the teacher and its cross-modal student, we propose extracting surface normal maps from raw point clouds to encode 3D geometry, enabling a shared backbone network used for both 2D and 3D modalities. Our key insight is that sur + +![](images/5342df6477af3055d6cba544d92ea4c5936d83068d902b863cb94e210ac8dbba.jpg) +Figure 3. FIND consists of two single-teacher-dual-student branches (left and right) designed to detect anomalies from multi-modal inputs, i.e., 2D RGB and 3D surface-normal. Each branch consists of a modality-specific teacher (center), a cross-modal distillation student (top), and an intra-modal reverse distillation student (bottom). The cross-modal student performs patch-wise mapping to the teacher modality for capturing inter-modality inconsistencies. In parallel, the intra-modal student conducts layer-wise feature inversion back to the original input resolution, enabling pixel-level localization of defects. By combining the results from the two pairs of intra-modal and cross-modal students, our approach robustly identifies various types of anomalies, even in few-shot scenarios. + +![](images/fdcec2dc2e27da84955a78ed877e7a2b9baf617194c3a98b20023ff009851cb5.jpg) + +face normal maps capture fine-grained surface orientation details and preserve subtle textures more effectively than depth maps. This contrasts with previous work, which uses either raw point clouds or depth maps. Raw point clouds exacerbate the domain gap, while depth images tend to smooth out small shape variations and are less sensitive to small defects. These findings allow us to effectively bridge the modality gap and therefore significantly reduce false positives. + +Cross-modal Distillation. Let $\mathbf{T}_{2D}$ and $\mathbf{T}_{3D}$ denote the two pre-trained Vision Transformer (ViT) teachers for the 2D RGB images and 3D surface normals, respectively. Inspired by CFM [5], each teacher receives input from its native domain and outputs high-level features $\mathbf{f}_{2D}^T$ or $\mathbf{f}_{3D}^T$ during training. In parallel, the corresponding cross-modal distillation student, $\mathbf{S}_{3D\rightarrow 2D}$ or $\mathbf{S}_{2D\rightarrow 3D}$ , takes features from one modality as input and attempts to reproduce the output of the teacher network from the other modality: + +$$ +\mathbf {f} _ {2 D} ^ {S} = \mathbf {S} _ {3 D \rightarrow 2 D} \big (\mathbf {f} _ {3 D} ^ {T} \big), \quad \mathbf {f} _ {3 D} ^ {S} = \mathbf {S} _ {2 D \rightarrow 3 D} \big (\mathbf {f} _ {2 D} ^ {T} \big). \qquad (1) +$$ + +The overall distillation loss is applied bidirectionally: + +$$ +\mathcal {L} _ {\text {c r o s s}} = \left(1 - \cos \langle \mathbf {f} _ {2 D} ^ {T}, \mathbf {f} _ {2 D} ^ {S} \rangle\right) + \left(1 - \cos \langle \mathbf {f} _ {3 D} ^ {T}, \mathbf {f} _ {3 D} ^ {S} \rangle\right). \tag {2} +$$ + +# 3.3. Intra-Modal Teacher-Student Learning + +While the cross-modal student excels in identifying semantic relationships between modalities, it relies on high-level + +feature representations with low spatial resolution, which may overlook subtle defects. To address this limitation, we introduce an intra-modal student for each modality, which performs intra-modal reconstruction on both feature level and image level. As a result, the intra-modal student specializes in capturing modality-specific fine-grained details for anomaly detection. In particular, the intra-modal student adopts a decoder-only architecture to perform multi-stage reverse distillation, progressively reconstructing back to the original image space for fine-grained anomaly localization. + +Multi-Layer Reverse Distillation. To preserve both low-level textures and high-level semantics, we adopt a multilayer feature inversion strategy for the training of intramodal student. Let $\{F_T^k\}_{k=1}^K$ denote a set of $K$ feature maps selected from different layers of the teacher's encoder, with layer 1 representing the shallow representation and layer $K$ being the deepest representation. In addition, $k = 0$ denotes the original input image space. Our two intra-modal students, $\mathbf{S}_{2D \to 2D}$ , $\mathbf{S}_{3D \to 3D}$ , are designed to reverse the encoding process in a top-down manner. Specifically, we use $D^k$ to denote the decoding module in the $k$ -th layer of an intra-modal student, with each decoder block $D^k(\cdot)$ reconstructing the next fine-grained feature map: $F_S^{k-1} = D^k(F_S^k)$ , where $F_S^k$ denotes the output of the $k$ -th level of the intra-modal student and $F_T^K$ serves as the student's starting point. The final output $F_S^0$ reaches the original image resolution, enabling fine-grained anomaly localization. + +We then enforce the consistency between the teacher and the student's predictions to ensure that normal appearance is + +reconstructed across different hidden layers. In addition, we incorporate an image-level reconstruction term that compares the final output $F_{S}^{0}$ against the original input image $I$ , further preserving fidelity to subtle textures. Concretely, we calculate the intra-modal loss as: + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {i n t r a}} = \sum_ {(S, T)} \Big (\sum_ {k = 1} ^ {K - 1} \big (1 - \cos \langle F _ {S} ^ {k}, F _ {T} ^ {k} \rangle \big) + \left\| F _ {S} ^ {0} - I \right\| ^ {2} \Big), \\ (S, T) \in \left\{\left(\mathbf {S} _ {2 D \rightarrow 2 D}, \mathbf {T} _ {2 \mathrm {D}}\right), \left(\mathbf {S} _ {3 D \rightarrow 3 D}, \mathbf {T} _ {3 \mathrm {D}}\right)\right\}. \tag {3} \\ \end{array} +$$ + +This encourages the intra-modal students to capture both high-level semantics and fine-grained details of teacher's normality representation, even in few-shot settings. + +# 3.4. Training and Inference + +Training. Our method jointly trains the cross-modal and intra-modal students using the respective objectives introduced in Sections 3.2 and 3.3. Specifically, we minimize the sum of the cross-modal loss in Eq. (2) and the intramodal loss in Eq. (3), yielding the final objective: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {c r o s s}} + \mathcal {L} _ {\text {i n t r a}}. \tag {4} +$$ + +Training is carried out on a few normal samples. The cross-modal student learns to reconstruct teacher features from the complementary modality, while the intra-modal student focuses on restoring its own modality in a reverse distillation manner. This combined strategy leverages both inter-modal consistency and fine-grained modality-specific cues for robust and fine-grained anomaly detection. + +Anomaly Scoring. At test time, each of the four students $\mathbf{S}_{2D\rightarrow 3D}$ , $\mathbf{S}_{3D\rightarrow 2D}$ , $\mathbf{S}_{2D\rightarrow 2D}$ , $\mathbf{S}_{3D\rightarrow 3D}$ produces an anomaly map by measuring the discrepancy between its reconstructed output and the corresponding teacher's feature. In particular, for the intra-modal students $(S_{2D\rightarrow 2D}$ and $S_{3D\rightarrow 3D})$ , the intra-modal anomaly maps $\mathbf{A}_{2D\rightarrow 2D}$ and $\mathbf{A}_{3D\rightarrow 3D}$ are calculated by aggregating the intra-modal teacher-student discrepancy across multiple layers. For each hidden layer, we additionally apply $\mathrm{L}_2$ -normalize to both the teacher's feature and the student's reconstructed feature at every spatial position. For each layer, we calculate the Euclidean distance between the student's reconstructed output and the teacher's original output to obtain a pixel-level error map. Next, we apply Gaussian smoothing to the obtained error map for a more coherent anomaly response. We repeat this process for all $K - 1$ hidden layers and the final output layer, and then average all the error maps to obtain the two intra-modal maps $\mathbf{A}_{2D\rightarrow 2D}$ and $\mathbf{A}_{3D\rightarrow 3D}$ for both 2D and 3D modalities. The final intra-modal anomaly map $\mathbf{A}_{\mathrm{intra}}$ is computed by averaging the two intra-modal maps $\mathbf{A}_{2D\rightarrow 2D}$ and $\mathbf{A}_{3D\rightarrow 3D}$ . + +The cross-modal students $(S_{2D\to 3D}$ and $S_{3D\to 2D})$ operate similarly by measuring the discrepancy between the + +student's reconstructed feature in one modality and the teacher's feature in the other modality. The resulting cross-modal anomaly maps from each student are denoted as $\mathbf{A}_{2D\to 3D}$ and $\mathbf{A}_{3D\to 2D}$ . Likewise, we average the two cross-modal maps to obtain a single cross-modal anomaly map $\mathbf{A}_{\mathrm{cross}}$ . The final overall anomaly map is obtained by the following: + +$$ +\mathbf {A} _ {\text {f i n a l}} = \alpha \mathbf {A} _ {\text {c r o s s}} + (1 - \alpha) \mathbf {A} _ {\text {i n t r a}}, \tag {5} +$$ + +where $\alpha$ adjusts the relative influence of cross-modal and intra-modal anomaly map. High values in $\mathbf{A}_{\mathrm{final}}$ indicate regions where the student network fails to mimic the teacher. By combining multi-modal alignment and reverse distillation reconstruction, our framework captures diverse anomaly signatures even in few-shot scenarios. + +# 4. Experiment + +# 4.1. Experimental Setup + +Datasets and Protocol. We evaluate on two public benchmarks: MVTec-3D-AD [1] and Eyecandies [2]. Following a standard few-shot normal-only protocol, we limit the available normal training images per class to $k \in \{5, 10, 50\}$ , randomly sampling and averaging results over ten runs. We also evaluate under a full-shot (all normal samples) setting for completeness. + +Baselines. We compare against: (1) BTF [11], a classical 3D feature-based method without deep learning; (2) AST [24], an asymmetric T-S approach trained on normal samples; (3) M3DM [31], a hybrid fusion model exploiting memory banks; (4) CFM [5], which aligns 2D-3D features without memory banks; and (5) 3D-ADNAS [21], which uses neural architecture search for multi-modal fusion. + +Evaluation Metrics. We follow standard anomaly detection metrics: (a) I-AUROC for image-level anomaly classification, (b) P-AUROC (pixel-level AUROC) for localization, and (c) AUPRO (area under per-region overlap) at specific false positive rates (FPR), typically $1\%$ and $30\%$ . Higher values denote better performance. + +Implementation Details. The intra-modal student is implemented as a multi-stage decoder, where each stage consists of three sequential convolutional blocks with normalization layers and GeLU activation to ensure nonlinearity. In particular, the final block employs ConvTranspose2d layers to up-sample the feature map back to the original image resolution and progressively decreases the dimensionality of the output channel to the final 3 channels (RGB). We implement our cross-modal student network as an encoder-decoder style network. The encoder first encodes the input patches into a hidden representation with two convolutional blocks, while the decoder reconstructs them in the cross-modal feature space through another convolutional + +Table 1. Comparison of anomaly detection and localization performance on MVTec-3D-AD and Eyecandies with varying numbers of training shots. $\mathbf{{ROC}}{}_{I},{\mathbf{{ROC}}}_{P}$ ,and $\mathbf{{PRO}}$ denote I-AUROC, P-AUROC, and AUPRO respectively. Bold indicates the best performance. + +
k-ShotMethodMVTec-3D-ADEyecandies
\( ROC_I \)\( ROC_P \)PRO-30%PRO-1%\( ROC_I \)\( ROC_P \)PRO-30%PRO-1%
5BTF [11]0.6710.9800.9200.2880.6520.8150.7380.202
AST [24]0.6800.9500.9030.1580.6330.7410.6910.143
M3DM [31]0.8220.9840.9370.3300.7640.8710.8070.251
CFM [5]0.8110.9860.9490.3820.7950.8790.8010.260
3D-ADNAS [21]0.826---0.7750.875--
FIND (Ours)0.8990.9910.9630.4200.8400.9510.8620.324
10BTF [11]0.6950.9830.9280.3080.6850.8340.8060.217
AST [24]0.6890.9460.8350.1740.6710.7670.6240.161
M3DM [31]0.8450.9860.9430.3550.8240.8900.8120.266
CFM [5]0.8450.9870.9540.3980.8380.9030.8250.284
3D-ADNAS [21]0.848---0.8070.869--
FIND (Ours)0.9210.9930.9670.4340.8680.9590.8890.350
50BTF [11]0.8060.9890.9470.3560.7210.8560.8240.240
AST [24]0.7940.9740.9290.3350.7390.8620.7150.118
M3DM [31]0.9070.9890.9550.3870.8360.9330.8460.304
CFM [5]0.9060.9910.9650.4310.8520.9260.8510.315
3D-ADNAS [21]0.890---0.8680.912--
FIND (Ours)0.9520.9950.9730.4550.8970.9760.9070.365
ALLBTF [11]0.8650.9920.9590.3830.7400.8830.8450.261
AST [24]0.9370.9760.9440.3980.7800.9020.7440.149
M3DM [31]0.9450.9920.9640.3940.8820.9770.8870.331
CFM [5]0.9540.9930.9710.4550.8810.9740.8870.335
3D-ADNAS [21]0.951---0.9460.970--
FIND (Ours)0.9780.9950.9830.4560.9550.9830.9250.395
+ +Table 2. Comparisons of per-category anomaly detection performance on MVTec-3D-AD. + +
MethodBagelCable GlandCarrotCookieDowelFoamPeachPotatoRopeTireMean
I-AUROC
BTF [11]0.9380.7650.9720.8880.9600.6640.9040.9290.9820.7260.865
AST [24]0.9830.8730.9760.9710.9320.8850.9740.9811.0000.7970.937
M3DM [31]0.9940.9090.9720.9760.9600.9420.9730.8990.9720.8500.945
CFM [5]0.9940.8880.9840.9930.9800.8880.9410.9430.9800.9530.954
3D-ADNAS [21]0.9971.0000.9710.9860.9660.9480.8970.8731.0000.8670.951
Shape Guided [4]0.9860.8940.9830.9910.9760.8570.9900.9650.9900.8690.947
LSFA [29]1.0000.9390.9820.9890.9610.9510.9830.9620.9890.9510.971
FIND (Ours)0.9840.9700.9850.9720.9820.9790.9740.9800.9880.9700.978
AUPRO@30%
BTF [11]0.9760.9690.9790.9730.9330.8880.8960.9120.9500.9710.959
AST [24]0.9700.9470.9810.9390.9130.9060.9790.9820.8890.9400.944
M3DM [31]0.9700.9710.9790.9500.9410.9320.9770.9710.9710.9750.964
CFM [5]0.9790.9720.9820.9450.9500.9680.9800.9430.9500.9810.971
Shape Guided [4]0.9810.9730.9820.9710.9620.9780.9810.9830.9740.9750.976
LSFA [29]0.9860.9740.9810.9460.9250.9410.9830.9830.9740.9830.968
FIND (Ours)0.9890.9770.9810.9830.9820.9860.9840.9800.9900.9780.983
+ +blocks, ensuring both preservation of the locality and improved alignment between the modalities. + +To ensure a fair comparison with previous work, we use the same ViT-B/8 DINO backbone as CFM [5], which is shared for both RGB and surface normal modalities. We treat surface normals as 3-channel RGB-style inputs and normalize them using ImageNet statistics before feeding + +them into DINO. During training, the backbone weights remain fixed, and we optimize only the four student networks. We extract hidden features from teacher's 4, 6, 8-th layers for intra-modal reverse distillation, while the 10-th layer's feature is employed as the input for cross-modal student. We use the ADAM optimizer for training with a learning rate of $1 \times 10^{-4}$ , and set $\alpha = 0.5$ for all-shot tasks and + +![](images/314955c28ff1a8386b969f81c76c131051abd5da2b3e22b3dbefefd4130e5d40.jpg) + +![](images/7f6bf279ded68aa23f81b46a77a119b9a3007ebfa48616f8b39567b252c56576.jpg) + +![](images/fac46808338bade6a8a5bf56d51c18da0b15e10cb1b8373eb55801e0c6ce9564.jpg) +Figure 4. Qualitative results on MVTec-3D-AD with various types of anomalies. From left to right: input RGB images, surface normals, GT anomaly maps, anomaly maps from 2D student set, anomaly maps from 3D student set, and our final predictions. Our final model consistently produces more accurate anomaly maps across different anomaly types. + +![](images/8dbab268d4a2f5c6efed4f14cb6a01bcef492b9245964b7b81a2159d9e92a61b.jpg) +Figure 5. Anomaly score distributions on bagel and cookie (MVTec-3D-AD, 10-shot). Our method shows better separation between normal and anomalous samples. + +$\alpha = 0.8$ for few-shot tasks. + +For MVTec-3D-AD, each sample includes an RGB image and an unstructured point cloud. We first organize the point cloud into a 2D grid(XYZ map) corresponding to the rgb image size $H \times W$ . We then compute per-pixel surface + +
CategoryIntra-studentCross-studentIntra + Cross
ImagePixelImagePixelImagePixel
Bagel96.399.798.899.599.599.6
Cable Gland74.497.679.098.382.498.3
Carrot97.499.899.499.997.399.8
Cookie86.498.491.897.292.597.8
Dowel75.298.183.899.382.699.2
Foam78.999.197.699.898.199.8
Peach97.099.895.799.999.499.9
Potato92.999.892.799.995.899.9
Rope97.699.896.699.898.899.8
Tire66.798.775.199.274.499.2
Average86.399.191.199.392.199.3
+ +Table 3. Ablation studies on intra-modal, cross-modal and dual students. We report the 10-shot categorical performance in image and pixel AUROC (\%) on the MVTec-3D dataset. + +
CategoryDepthPoint CloudSurface Normal
ImagePixelImagePixelImagePixel
Bagel99.099.498.399.599.599.6
Cable Gland68.897.774.197.682.498.3
Carrot82.799.392.499.197.399.8
Cookie84.498.590.797.292.597.8
Dowel87.199.075.599.082.699.2
Foam81.689.892.499.598.199.8
Peach85.799.393.799.699.499.9
Potato68.499.184.398.795.899.9
Rope95.799.694.299.798.899.8
Tire66.897.673.699.374.499.2
Average82.097.986.898.992.199.3
+ +Table 4. Ablation studies on 3D modality input selection. We report the 10-shot categorical performance in image and pixel AUROC (\%) on the MVTec-3D dataset. + +normals by fitting a local plane to the neighbors of each point (or using the cross-product of adjacent triangle faces in a mesh representation). This produces a 3-channel normal vector map for the surface of the object. Following prior work [5, 31], we remove background points and fill the removed regions in the RGB image with a neutral value (black) to avoid false anomalies from the background. We finally resize both the RGB image and the normal map to $224 \times 224$ pixels before feeding them to the model. For Eyecandies, we use normal maps with global lighting as the 3D modality. + +# 4.2. Primary Results + +Few-shot Performance on MVtec-3D-AD. Table 1 shows that FIND exceeds all competitors: BTF, AST, M3DM, CFM and 3D-ADNAS, in all few-shot settings (5, 10, 50 shots) and evaluation metrics at both the image and pixel levels. For example, in the challenging 5-shot setting, our method achieves an I-AUROC of 0.899. With 10-shot training, the I-AUROC performance is increased to 0.921. Even with 50 shots, FIND continues to outperform strong memory-based methods (e.g., M3DM). These consistent gains highlight the effectiveness of FIND in limited data scenarios. + +Full-shot Performance on MVtec-3D-AD. Table 1 presents the anomaly detection performance in the full-shot setting at both the image and pixel levels, while Table 2 reports results for each category. FIND surpasses existing methods in both image-level and pixel-level detection, achieving the highest average I-AUROC $(97.8\%)$ and AUPRO@ $30\%$ $(98.3\%)$ , demonstrating its robustness across various object classes. Notably, our approach excels in challenging categories (e.g., Cable Gland, Foam), indicating its ability to capture fine-grained anomalous patterns. + +Results on Eyecandies. As shown in Table 1, FIND also sets a new state-of-the-art on Eyecandies under both few-shot and full-shot protocols, significantly outperforming existing methods. For example, at 5-shot, FIND achieves an I-AUROC score of 0.840, exceeding the other baselines significantly. This performance gain extends to pixel-level metrics (P-AUROC and AUPRO), demonstrating a superior localization of subtle anomalies. Furthermore, with 50 normal samples for training, our method even outperforms the full-shot trained M3DM and CFM in terms of AUROC and AUPRO. These results highlight the effectiveness of our approach in various shots of training data scenarios, reinforcing its robustness and generalizability. + +# 4.3. Qualitative Analysis + +Fig. 4 shows that our method is robust to various types of anomaly (shape-only, color-only, combined) by effectively exploiting 2D-3D complementarity. The 2D student set $(\mathbf{S}_{3D\rightarrow 2D},\mathbf{S}_{2D\rightarrow 2D})$ captures subtle color anomalies often overlooked by geometric cues, while the 3D student set $(\mathbf{S}_{2D\rightarrow 3D},\mathbf{S}_{3D\rightarrow 3D})$ detects shape deformations that are imperceptible in RGB. Combining these maps improves localization accuracy and reduces false positives. + +Fig. 5 further compares the anomaly score distributions of normal and anomalous samples under the 10-shot setting for CFM and FIND on two MVTec-3D-AD classes (bagel and cookie). Our method yields more clearly separable distributions between normal and abnormal samples, contributing to improved detection accuracy. + +# 4.4. Ablation Studies + +Intra-modal vs. Cross-modal vs. Dual Student. We evaluate the contributions of the intra-modal and cross-modal students, with results shown in Table 3. Intra-modal students capture modality-specific anomalies and guide the model to reproduce a normal version of the query sample at the pixel level, enhancing the detection of local defects. Removing them causes a notable drop in detection accuracy for categories such as cable gland, potato, and rope. However, intra-modal students operating within a single modality cannot leverage complementary information from the other modality, leaving those cross-modal anomalies undetected. In contrast, cross-modal students learn the expected normal correspondence between an object's color and shape appearance and detect anomalies by identifying inconsistencies between modalities. By combining both intra-modal and cross-modal students, we leverage their complementary strengths and achieve the best results. + +3D Modality Input Choices. In a 10-shot setting on MVTec-3D-AD, we compare raw point clouds (with a PointNet encoder), depth maps(from point clouds), and surface-normal maps for anomaly detection. The results are shown in Table 4. Surface normals achieve the highest I-AUROC, outperforming depth maps and raw point clouds. This indicates that normal maps preserve fine-grained geometric details that can highlight small defects. In contrast, depth maps lack this level of detail, resulting in lower accuracy. The point cloud baseline performs worst, likely due to a modality gap introduced by using a separate PointNet-based 3D feature extractor. Such gaps arise from misaligned representations when using a point cloud network versus image-based features. Our results show that surface normal maps provide the most effective shape modality for 3D anomaly detection in few-shot scenarios. + +# 5. Conclusion + +In this work, we address the critical challenges of multimodal anomaly detection in few-shot scenarios by introducing FIND, a dual-student framework that integrates intramodal reverse distillation and cross-modal feature mapping. FIND overcomes data scarcity and modality gaps by exploiting collaborative learning between two students: one dedicated to fine-grained intra-modal anomaly localization through reverse distillation, and another focused on cross-modal inconsistency detection. Extensive experiments demonstrate that FIND achieves new state-of-the-art results in both few-shot and full-shot settings. + +# Acknowledgements + +This work is supported by the Agency for Science, Technology and Research (A*STAR) under its MTC Programmatic Funds (Grant No. M23L7b0021). + +# References + +[1] Paul Bergmann, Xin Jin, David Sattlegger, and Carsten Steger. The mvtec 3d-ad dataset for unsupervised 3d anomaly detection and localization. In Proceedings of the 17th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications, VISIGRAPP, pages 202-213, 2022. 5 +[2] Luca Bonfiglioli, Marco Toschi, Davide Silvestri, Nicola Fioraio, and Daniele De Gregorio. The eyecandies dataset for unsupervised multimodal anomaly detection and localization. In 16th Asian Conference on Computer Vision (ACCV), pages 459-475, 2022. 5 +[3] Mengxi Chen and ... Enhanced multimodal representation learning with cross-modal KD. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11766-11775, 2023. 3 +[4] Yu-Min Chu, Liu Chieh, Ting-I Hsieh, Hwann-Tzong Chen, and Tyng-Luh Liu. Shape-guided dual-memory learning for 3d anomaly detection. 2023. 1, 2, 6 +[5] Alex Costanzino, Pierluigi Zama Ramirez, Giuseppe Lisanti, and Luigi Di Stefano. Multimodal industrial anomaly detection by crossmodal feature mapping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17234-17243, 2024. 1, 2, 4, 5, 6, 7 +[6] Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one-class embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9737-9746, 2022. 1 +[7] Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one-class embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9737-9746, 2022. 1 +[8] Zheng Fang, Xiaoyang Wang, Haocheng Li, Jiejie Liu, Qiugui Hu, and Jimin Xiao. Fastrecon: Few-shot industrial anomaly detection via fast feature reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17481-17490, 2023. 1, 3 +[9] Zhihao Gu, Jiangning Zhang, Liang Liu, Xu Chen, Jinlong Peng, Zhenye Gan, Guannan Jiang, Annan Shu, Yabiao Wang, and Lizhuang Ma. Rethinking reverse distillation for multi-modal anomaly detection. In Thirty-Eighth AAAI Conference on Artificial Intelligence, AAAI, pages 8445-8453, 2024. 1, 2 +[10] Yu Hong, Hang Dai, and Yong Ding. Cross-modality knowledge distillation network for monocular 3d object detection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 87-104, 2022. 3 +[11] Eliahu Horwitz and Yedid Hoshen. Back to the feature: classical 3d features are (almost) all you need for 3d anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2968-2977, 2023. 1, 5, 6 +[12] Chaoqin Huang, Haoyan Guan, Aofan Jiang, Ya Zhang, Michael Spratling, and Yan-Feng Wang. Registration based few-shot anomaly detection. In European Conference on Computer Vision, pages 303-319. Springer, 2022. 1, 3 + +[13] Fushuo Huo, Wenchao Xu, Jingcai Guo, Haozhao Wang, and Song Guo. $C^2$ kd: Bridging the modality gap for cross-modal knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7151-7160, 2024. 3 +[14] Jongheon Jeong, Yang Zou, Taewan Kim, Dongqing Zhang, Avinash Ravichandran, and Onkar Dabeer. Winclip: Zero/few-shot anomaly classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19606-19616, 2023. 1, 3 +[15] Pilhyeon Lee and ... Decomposed cross-modal distillation for RGB-based temporal action detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages -, 2023. CVPR 2023. 3 +[16] Yujin Lee, Seoyoon Jang, and Hyunsoo Yoon. Anople: Few-shot anomaly detection via bi-directional prompt learning with only normal samples. arXiv preprint arXiv:2408.13516, 2024. 1, 3 +[17] Wenqiao Li, Xiaohao Xu, Yao Gu, Bozhong Zheng, Shenghua Gao, and Yingna Wu. Towards scalable 3d anomaly detection and localization: A benchmark via 3d anomaly synthesis and a self-supervised learning network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22207-22216, 2024. 1 +[18] Hanzhe Liang, Guoyang Xie, Chengbin Hou, Bingshu Wang, Can Gao, and Jinbao Wang. Look inside for more: Internal spatial modality perception for 3d anomaly detection. arXiv preprint arXiv:2412.13461, 2024. 1, 2 +[19] Victor Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Y Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. Advances in Neural Information Processing Systems, 35:17612-17625, 2022. 3 +[20] Chieh Liu, Yu-Min Chu, Ting-I Hsieh, Hwann-Tzong Chen, and Tyng-Luh Liu. Learning diffusion models for multi-view anomaly detection. In European Conference on Computer Vision, pages 328-345. Springer, 2025. 1 +[21] Kaifang Long, Guoyang Xie, Lianbo Ma, Jiaqi Liu, and Zhichao Lu. Revisiting multimodal fusion for 3d anomaly detection from an architectural perspective. In Thirty-Eighth AAAI Conference on Artificial Intelligence, AAAI, 2025. 1, 2, 5, 6 +[22] Xiang Ma, Xuemei Li, Lexin Fang, and Caiming Zhang. Bridging the modality gap: Dimension information alignment and sparse spatial constraint for image-text matching. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 5074-5082, 2024. 3 +[23] Nicolae-Cătălin Ristea, Neelu Madan, Radu Tudor Ionescu, Kamal Nasrollahi, Fahad Shahbaz Khan, Thomas B. Moeslund, and Mubarak Shah. Self-supervised predictive convolutional attentive block for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13576-13586, 2022. 1 +[24] Marco Rudolph, Tom Wehrbein, Bodo Rosenhahn, and Bastian Wandt. Asymmetric student-teacher networks for industrial anomaly detection. In IEEE/CVF Winter Conference on + +Applications of Computer Vision, WACV, pages 2592-2602, 2023. 1, 2, 5, 6 +[25] Pritam Sarkar and Ali Etemad. XKD: Cross-modal knowledge distillation with domain alignment for video representation learning. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2024. to appear. 3 +[26] Wenbo Sui, Daniel Lichau, Josselin Lefèvre, and Harold Philippeau. Incomplete multimodal industrial anomaly detection via cross-modal distillation, 2024. 1, 3 +[27] Long Tian, Hongyi Zhao, Ruiying Lu, Rongrong Wang, Yu-jie Wu, Liming Wang, Xiongpeng He, and Xiyang Liu. Foct: Few-shot industrial anomaly detection with foreground-aware online conditional transport. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6241-6249, 2024. 1, 3 +[28] Dinh Tien Tran, Anh Tuan Nguyen, Nguyen Hoang Tran, Ta Duc Huy, Soan T. M. Duong, Chanh D. Tr. Nguyen, and Steven Q. H. Truong. Revisiting reverse distillation for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24511-24520, 2023. 1 +[29] Yuanpeng Tu, Boshen Zhang, Liang Liu, Yuxi Li, Jiangning Zhang, Yabiao Wang, Chengjie Wang, and Cairong Zhao. Self-supervised feature adaptation for 3d industrial anomaly detection. In European Conference on Computer Vision, pages 75-91. Springer, 2025. 2, 6 +[30] Huiqun Wang, Yiping Bao, Panwang Pan, Zeming Li, Xiao Liu, Ruijie Yang, and Di Huang. Multi-modal relation distillation for unified 3d representation learning. In European Conference on Computer Vision (ECCV), 2024. to appear. 3 +[31] Yue Wang, Jinlong Peng, Jiangning Zhang, Ran Yi, Yabiao Wang, and Chengjie Wang. Multimodal industrial anomaly detection via hybrid fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8032-8041, 2023. 1, 2, 5, 6, 7 +[32] Zeyu Wang, Dingwen Li, Chenxu Luo, Cihang Xie, and Xiaodong Yang. DistillBEV: Boosting multi-camera 3d object detection with cross-modal knowledge distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8637-8646, 2023. 3 +[33] Zihui Xue, Zhengqi Gao, Sucheng Ren, and Hang Zhao. The modality focusing hypothesis: Towards understanding cross-modal knowledge distillation. In International Conference on Learning Representations (ICLR), 2023. Top $5\%$ at ICLR 2023. 3 +[34] Xu Yan, Heshen Zhan, Chaoda Zheng, Jiantao Gao, Ruimao Zhang, Shuguang Cui, and Zhen Li. Let images give you more: Point cloud cross-modal training for shape analysis. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 1 +[35] Heeseung Yun, Joonil Na, and Gunhee Kim. Dense 2d-3d indoor prediction with sound via aligned cross-modal distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7829-7838, 2024. 3 +[36] Vitjan Zavrtanik, Matej Kristan, and Danijel Skocaj. Cheating depth: Enhancing 3d surface anomaly detection via depth + +simulation. In IEEE/CVF Winter Conference on Applications of Computer Vision, WACV, pages 2153-2161. IEEE, 2024. 2 +[37] Tianlu Zhang, Hongyuan Guo, Qiang Jiao, Qiang Zhang, and Jungong Han. Efficient rgb-t tracking via cross-modality distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5815-5824, 2023. 3 +[38] Xuan Zhang, Shiyu Li, Xi Li, Ping Huang, Jiulong Shan, and Ting Chen. Destseg: Segmentation guided denoising student-teacher for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3914–3923, 2023. 1 +[39] Shengchao Zhou, Weizhou Liu, Chen Hu, Shuchang Zhou, and Chao Ma. Unidistill: A universal cross-modality knowledge distillation framework for 3d object detection in bird's-eye view. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5116-5125, 2023. 3 +[40] Zheyuan Zhou, Le Wang, Naiyu Fang, Zili Wang, Lemiao Qiu, and Shuyou Zhang. R3d-ad: Reconstruction via diffusion for 3d anomaly detection. In European Conference on Computer Vision, pages 91-107. Springer, 2025. 1 \ No newline at end of file diff --git a/2025/FIND_ Few-Shot Anomaly Inspection with Normal-Only Multi-Modal Data/images.zip b/2025/FIND_ Few-Shot Anomaly Inspection with Normal-Only Multi-Modal Data/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..389a88be7e4c7d9e39fa7d6c46687b3177279e44 --- /dev/null +++ b/2025/FIND_ Few-Shot Anomaly Inspection with Normal-Only Multi-Modal Data/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0b95d07d4af18ae10c7cdbf8ab03608bd96c11c459cba81209dd1cf79019b69f +size 622606 diff --git a/2025/FIND_ Few-Shot Anomaly Inspection with Normal-Only Multi-Modal Data/layout.json b/2025/FIND_ Few-Shot Anomaly Inspection with Normal-Only Multi-Modal Data/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b6704ef236fefd18921673fc90bdc63167ce68e9 --- /dev/null +++ b/2025/FIND_ Few-Shot Anomaly Inspection with Normal-Only Multi-Modal Data/layout.json @@ -0,0 +1,6870 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 73, + 103, + 536, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 103, + 536, + 120 + ], + "spans": [ + { + "bbox": [ + 73, + 103, + 536, + 120 + ], + "type": "text", + "content": "FIND: Few-Shot Anomaly Inspection with Normal-Only Multi-Modal Data" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 142, + 570, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 142, + 570, + 200 + ], + "spans": [ + { + "bbox": [ + 62, + 142, + 570, + 200 + ], + "type": "text", + "content": "Yiting Li" + }, + { + "bbox": [ + 62, + 142, + 570, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 142, + 570, + 200 + ], + "type": "text", + "content": ", Fayao Liu" + }, + { + "bbox": [ + 62, + 142, + 570, + 200 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 62, + 142, + 570, + 200 + ], + "type": "text", + "content": ", Jingyi Liao" + }, + { + "bbox": [ + 62, + 142, + 570, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 142, + 570, + 200 + ], + "type": "text", + "content": ", Sichao Tian" + }, + { + "bbox": [ + 62, + 142, + 570, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 62, + 142, + 570, + 200 + ], + "type": "text", + "content": ", Chuan-Sheng Foo" + }, + { + "bbox": [ + 62, + 142, + 570, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 142, + 570, + 200 + ], + "type": "text", + "content": ", Xulei Yang" + }, + { + "bbox": [ + 62, + 142, + 570, + 200 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 62, + 142, + 570, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 142, + 570, + 200 + ], + "type": "text", + "content": "Institute for Infocomm Research, A*STAR, Singapore " + }, + { + "bbox": [ + 62, + 142, + 570, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 62, + 142, + 570, + 200 + ], + "type": "text", + "content": "Chinese Academy of Medical Sciences and Peking Union Medical College, China \n{li_yiting, liu_fayao, liao_jingyi, foo_chuan_sheng, yang_xulei} @i2r.a-star.edu.sg, {sctian} @implad.ac.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 251, + 297, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 251, + 297, + 515 + ], + "spans": [ + { + "bbox": [ + 54, + 251, + 297, + 515 + ], + "type": "text", + "content": "Multi-modal anomaly detection (MAD) improves industrial inspection by exploiting complementary 2D and 3D data. However, existing methods struggle in few-shot scenarios due to limited data and modality gaps. Current approaches either fuse multi-modal features or align cross-modal representations; however, they often suffer from high false-positive rates and fail to detect subtle defects, especially when training samples are scarce. To address these challenges, we propose the first few-shot MAD method FIND, a novel dual-student framework that integrates intramodal reverse distillation and cross-modal feature mapping. FIND employs modality-specific teachers and two collaborative students: an intra-modal student for fine-grained anomaly localization via reverse distillation, and a cross-modal student that captures inter-modal correspondences to detect cross-modal inconsistencies. Extensive experiments on MVTec-3D-AD and Eyecandies show that FIND significantly outperforms state-of-the-art methods in both full-shot and few-shot settings. Ablation studies validate the complementary roles of intra- and cross-modal distillation. Our work significantly advances MAD robustness in data-scarce industrial applications." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 540, + 135, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 135, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 135, + 553 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 560, + 295, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 560, + 295, + 680 + ], + "spans": [ + { + "bbox": [ + 55, + 560, + 295, + 680 + ], + "type": "text", + "content": "Multi-modal anomaly detection (MAD) has garnered increasing interest in industrial inspection [4, 17, 18, 20, 21, 26, 31, 34, 40], as relying on a single modality (e.g., RGB) may fail to capture crucial geometric cues under varying surface textures. By integrating heterogeneous data such as point clouds or depth maps, MAD methods can leverage multi-modal complementarity to enhance inspection quality. Since anomalous data are rare, most existing anomaly detection (AD) methods are trained only on normal data. Our work follows this protocol." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 681, + 295, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 681, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 67, + 681, + 295, + 693 + ], + "type": "text", + "content": "While significant progress has been made in MAD, exist" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 227, + 436, + 327 + ], + "blocks": [ + { + "bbox": [ + 317, + 227, + 436, + 327 + ], + "lines": [ + { + "bbox": [ + 317, + 227, + 436, + 327 + ], + "spans": [ + { + "bbox": [ + 317, + 227, + 436, + 327 + ], + "type": "image", + "image_path": "f0cbbd1f2cc09eb4ec229d1a47d39b1ce2c60b7a00d469d6a98d1fa8e28946f4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 337, + 555, + 393 + ], + "lines": [ + { + "bbox": [ + 313, + 337, + 555, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 555, + 393 + ], + "type": "text", + "content": "Figure 1. Comparison with state-of-the-art methods on MVTec-3D-AD in terms of I-AUROC (left) and AUPRO@1% (right). FIND substantially outperforms prior methods (BTF [11], AST [24], M3DM [31], 3D-ADNAS [21] and CFM [5]) in both few-shot and full-shot settings." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 437, + 227, + 552, + 327 + ], + "blocks": [ + { + "bbox": [ + 437, + 227, + 552, + 327 + ], + "lines": [ + { + "bbox": [ + 437, + 227, + 552, + 327 + ], + "spans": [ + { + "bbox": [ + 437, + 227, + 552, + 327 + ], + "type": "image", + "image_path": "79b2b557084b07c5b10597bbffcd215dc41f8a5d4e9b7b5ef38b7b4ef10a39c7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 414, + 555, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 414, + 555, + 557 + ], + "spans": [ + { + "bbox": [ + 312, + 414, + 555, + 557 + ], + "type": "text", + "content": "ing approaches struggle to retain robustness in few-shot scenarios, where limited normal multi-modal training data is available—a critical limitation given the high cost of acquiring such data [8, 12, 14, 16, 27]. This is evidenced in Fig. 1, which compares various methods using different amounts of normal training samples, with notable performance drops observed. In this work, we propose FIND, a simple yet effective dual-student MAD framework focused on Few-Shot Anomaly Inspection with Normal-Only Multi-Modal Data. It effectively minimizes modality interference while exploiting multi-modal complementarity, significantly enhancing anomaly detection and localization." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 558, + 556, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 558, + 556, + 653 + ], + "spans": [ + { + "bbox": [ + 312, + 558, + 556, + 653 + ], + "type": "text", + "content": "In single-modality anomaly detection, teacher-student (T-S) frameworks achieve strong results [6, 23, 28, 38]. A pretrained teacher extracts generic features, while a student trained on normal data mimics the teacher. Anomalies emerge as discrepancies between their features at inference. Reverse distillation [7, 28] refines this by reconstructing teacher features from compact one-class embeddings, enabling fine-grained defect localization." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "content": "However, extending the success of T-S frameworks to MAD remains a challenge. Existing methods either fuse modalities by concatenating RGB and depth at the input level [24] or across multiple feature levels [9], which risks modality interference. Specifically, MMRD [9] introduces" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 148, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 148, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 148, + 712 + ], + "type": "text", + "content": "†Corresponding authors." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23290" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 346 + ], + "type": "text", + "content": "a multi-modal reverse distillation paradigm that performs a weighted fusion of image and depth data across multiple feature levels, creating a multi-modal teacher target for a learnable student decoder tasked with restoring the teacher's multi-modal representations. In contrast, cross-modal Feature Mapping (CFM) [5], although not explicitly formulated within the T-S framework, uses domain-specific teachers for cross-modal distillation and performs late fusion on score maps instead of relying on feature fusion. The features of different modalities are aligned through an MLP-based projection network. CFM trains two student models to map features from one modality to another by mimicking the corresponding target modality teacher. By learning the expected correspondence between an object's color and shape, any deviation from this correspondence at test time indicates an anomaly. However, the inherent modality gap can lead to large reconstruction errors even for normal samples, resulting in high false positive rates and limiting the model's ability to distinguish genuine anomalies. These limitations result in high false positives and coarse localization (see Fig. 2 for examples), which are further exacerbated in few-shot scenarios, where limited normal samples hinder the effective alignment of the two modalities." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 349, + 296, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 349, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 349, + 296, + 647 + ], + "type": "text", + "content": "To address the challenges stated above, we present FIND, which unifies intra-modal reverse distillation and cross-modal distillation within a single T-S framework. Our intuition is to minimize modality interference while exploiting the complementary strengths of multiple modalities. Specifically, each modality is paired with a modality-specific teacher, an intra-modal student, and a cross-modal student. The intra-modal student is trained via reverse distillation to model fine-grained, modality-specific normal appearances, while the cross-modal student is designed to capture inter-modal correspondences for robust anomaly detection. This decoupled design allows each student to focus on its specific task: the intra-modal student is highly sensitive to prominent modality-specific anomalies and provides finer-grained localization, while the cross-modal student detects anomalies that disrupt the learned normal correspondence between different modalities. This synergy enables precise localization and robustness to various types of anomalies, even with limited training data (see Fig. 2 for comparisons). To further mitigate the modality gap and thus reduce false positives in cross-modal distillation, we propose using surface normals as a geometric proxy, enabling the use of shared RGB and 3D encoders. We find this simple strategy to be highly effective in achieving accurate MAD. Our contributions are summarized as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 653, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 653, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 653, + 295, + 713 + ], + "type": "text", + "content": "- We propose the first few-shot MAD method FIND, which synergistically integrates intra-modal reverse distillation and cross-modal distillation. By decoupling the learning objectives for intra- and cross-modal students, FIND significantly enhances MAD performance." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 321, + 70, + 554, + 242 + ], + "blocks": [ + { + "bbox": [ + 321, + 70, + 554, + 242 + ], + "lines": [ + { + "bbox": [ + 321, + 70, + 554, + 242 + ], + "spans": [ + { + "bbox": [ + 321, + 70, + 554, + 242 + ], + "type": "image", + "image_path": "ff2fcea15e646f283f76ecb9b49af151a96db198ed4a9fbdd816ba8cd45acf7b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 251, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 313, + 251, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 555, + 285 + ], + "type": "text", + "content": "Figure 2. Compared to CFM under the 10-shot setting, our anomaly maps are sharper and show fewer false positives, focusing on anomalous patterns rather than minor texture variations." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 321, + 308, + 553, + 392 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 321, + 308, + 553, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 308, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 321, + 308, + 553, + 354 + ], + "type": "text", + "content": "- We show that surface normals are a highly effective geometric modality for MAD, effectively bridging the modality gap in cross-modal distillation, reducing false positives, and enhancing anomaly detection precision." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 321, + 355, + 553, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 355, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 321, + 355, + 553, + 392 + ], + "type": "text", + "content": "- FIND sets a new state-of-the-art in both multi-modal anomaly detection and localization tasks, outperforming existing methods in both full-shot and few-shot settings." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 405, + 403, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 405, + 403, + 417 + ], + "spans": [ + { + "bbox": [ + 314, + 405, + 403, + 417 + ], + "type": "text", + "content": "2. Related works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 426, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 715 + ], + "type": "text", + "content": "Multi-modal Anomaly Detection With the development of 3D sensors, multi-modal anomaly detection has gained increasing attention [4, 9, 21, 36], as 3D information (e.g., point clouds) offers richer geometric cues and captures shape deformations that may be invisible in RGB domain. Recently, researchers have explored cross-modal alignment methods that fuse RGB and 3D data [24, 31] for more robust anomaly localization. For instance, CFM [5] proposes a novel framework that learns to map 2D and 3D features to each other, thereby pinpointing anomalies where the two modalities disagree; [29] introduce a local-to-global self-supervised adaptation mechanism to finetune pre-trained 2D/3D networks for more precise cross-modal alignment; [31] design a hybrid fusion scheme with multiple memory banks to achieve consistent performance gains over single-modality baselines. In parallel, memory-based 3D AD approaches [21, 31] have also shown strong potential by storing normal embeddings from both color and geometry in dedicated banks and measuring anomalous deviations at test time. For example, [4] propose a shape-guided dual-memory method that leverages signed distance fields (SDF) and RGB cues for complementary anomaly localization; [18] advocate an internal spatial modality perception module to distill intrinsic 3D structures into a memory" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23291" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "type": "text", + "content": "bank; [26] handle incomplete multi-modal data via cross-modal distillation and memory retrieval. Although existing methods have significantly advanced the state of the art in MAD, most still rely on either large-scale normal training sets or carefully aligned networks for each modality. In scenarios where only a few normal samples are available, the modality gap between 2D images and 3D point clouds can hinder reliable cross-modal feature alignment, thus limiting the models' ability to detect subtle anomalies." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 183, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 183, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 183, + 296, + 529 + ], + "type": "text", + "content": "Few-shot Anomaly Detection In resource-constrained scenarios, collecting large amount of normal data can be both costly and impractical. Consequently, few-shot anomaly detection (FSAD) becomes a promising alternative. Current research on FSAD has primarily focused on RGB domains. AnoPLe [16] proposes bi-directional prompt learning with only normal samples, leveraging a pretrained vision-language model to detect anomalies from both textual and visual cues. FastRecon [8] proposes a lightweight feature-based regression with distribution regularization to obtain the optimal transformation from support to query features, it reconstructs a normal version of test sample's features from visually similar support features and detects any large deviations as anomalies. WinCLIP [14] explores a window-based feature fusion strategy to extract multi-scale features for zero-/few-shot anomaly classification and segmentation. FCOT [27] leverages a distribution-based Conditional Transport metric to minimize feature discrepancies between a few support images and query images, while updating a memory bank to precisely match anomalous regions. RegAD [12] uses image registration to align query images with support images in a category-agnostic manner. Despite remarkable progress, existing FSAD approaches still face limitations when extended to multi-modal scenarios due to the lack of capabilities for capturing the shape distortions that 3D data would reveal. This limitation motivates our work on multi-modal anomaly detection under few-shot settings, by capturing shape-induced defects that 2D-only methods may overlook." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": "Cross-modal knowledge distillation Cross-modal knowledge distillation (KD) has attracted increasing interest for transferring knowledge between heterogeneous modalities (e.g. images, LiDAR, thermal, audio) [3, 15, 25, 30, 33]. A core challenge is the modality gap issue due to the discrepancies in data/feature distribution between modalities [13, 19, 22]. [19] demonstrates that modality gap in multimodal models such as CLIP is caused by a combination of model initialization and contrastive learning optimization. Recent works tackle this by aligning feature representations and customizing distillation strategies to each modality pair. For example, a LiDAR-based teacher can guide a camera-based student to improve monocular detection accuracy [10, 15, 32]. C2KD [13] propose a bidirectional distillation framework where teacher and student" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "type": "text", + "content": "learn from each other, coupled with on-the-fly sample filtering to discard misaligned targets. Cross-modal KD is applied in RGB-Thermal (RGB-T) tracking [37, 39], a two-stream RGB-T network can serve as a powerful teacher to a lightweight one-stream tracker; distilling both modality-specific and shared features helps the student network approach the teacher's performance, while greatly reducing computation. In audio-visual learning [35], visual models have been used to teach audio models to predict spatial or semantic information from sound, often by enforcing feature alignments to handle the modality gap." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 215, + 370, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 215, + 370, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 215, + 370, + 227 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 235, + 440, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 235, + 440, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 235, + 440, + 247 + ], + "type": "text", + "content": "3.1. Framework Overview" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 252, + 555, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 252, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 252, + 555, + 479 + ], + "type": "text", + "content": "FIND consists of two single-teacher-dual-student branches designed to detect anomalies from multi-modal 2D and 3D inputs. Fig. 3 shows an overview of our method. Each branch is composed of a modality-specific teacher " + }, + { + "bbox": [ + 313, + 252, + 555, + 479 + ], + "type": "inline_equation", + "content": "(\\mathbf{T}_{2\\mathrm{D}}" + }, + { + "bbox": [ + 313, + 252, + 555, + 479 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 252, + 555, + 479 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{3\\mathrm{D}})" + }, + { + "bbox": [ + 313, + 252, + 555, + 479 + ], + "type": "text", + "content": ", a cross-modal distillation student " + }, + { + "bbox": [ + 313, + 252, + 555, + 479 + ], + "type": "inline_equation", + "content": "(\\mathbf{S}_{3D\\rightarrow 2D}" + }, + { + "bbox": [ + 313, + 252, + 555, + 479 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 252, + 555, + 479 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{2D\\rightarrow 3D})" + }, + { + "bbox": [ + 313, + 252, + 555, + 479 + ], + "type": "text", + "content": ", and an intra-modal reverse distillation student " + }, + { + "bbox": [ + 313, + 252, + 555, + 479 + ], + "type": "inline_equation", + "content": "(\\mathbf{S}_{2D\\rightarrow 2D}" + }, + { + "bbox": [ + 313, + 252, + 555, + 479 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 252, + 555, + 479 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{3D\\rightarrow 3D})" + }, + { + "bbox": [ + 313, + 252, + 555, + 479 + ], + "type": "text", + "content": ". Our motivation is to integrate cross-modal distillation with reverse distillation in a unified framework that minimizes modality interference while effectively exploiting their complementarity for anomaly detection. In essence, the Cross-Modal Student distills knowledge across the RGB and 3D streams, learning how normal features in one modality correspond to those in the other. The Intra-Modal Student, on the other hand, learns to reproduce the normal pattern of individual modality through reverse distillation. During training, both students receive supervision derived from a shared set of teacher features so that each modality learns a more comprehensive representation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 487, + 524, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 524, + 500 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 524, + 500 + ], + "type": "text", + "content": "3.2. Cross-modal Teacher-Student Learning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 505, + 553, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 505, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 505, + 553, + 540 + ], + "type": "text", + "content": "We aim to leverage the complementary knowledge between two different modalities, namely 2D and 3D, for anomaly detection through a cross-modal distillation procedure." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": "Bridging the Modality Gap. Since features extracted by 2D backbones for RGB images and 3D point-based networks for point clouds follow different distributions, existing MAD methods face the modality gap issue due to the use of separate feature backbones for each modality. This discrepancy becomes even more significant in few-shot learning scenarios, where a student model with limited training data struggles to align with the teacher's normal feature distribution, leading to a high false positive rate in detection performance. To bridge the modality gap between the teacher and its cross-modal student, we propose extracting surface normal maps from raw point clouds to encode 3D geometry, enabling a shared backbone network used for both 2D and 3D modalities. Our key insight is that sur" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23292" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 68, + 298, + 264 + ], + "blocks": [ + { + "bbox": [ + 60, + 68, + 298, + 264 + ], + "lines": [ + { + "bbox": [ + 60, + 68, + 298, + 264 + ], + "spans": [ + { + "bbox": [ + 60, + 68, + 298, + 264 + ], + "type": "image", + "image_path": "5342df6477af3055d6cba544d92ea4c5936d83068d902b863cb94e210ac8dbba.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 273, + 555, + 340 + ], + "lines": [ + { + "bbox": [ + 54, + 273, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 54, + 273, + 555, + 340 + ], + "type": "text", + "content": "Figure 3. FIND consists of two single-teacher-dual-student branches (left and right) designed to detect anomalies from multi-modal inputs, i.e., 2D RGB and 3D surface-normal. Each branch consists of a modality-specific teacher (center), a cross-modal distillation student (top), and an intra-modal reverse distillation student (bottom). The cross-modal student performs patch-wise mapping to the teacher modality for capturing inter-modality inconsistencies. In parallel, the intra-modal student conducts layer-wise feature inversion back to the original input resolution, enabling pixel-level localization of defects. By combining the results from the two pairs of intra-modal and cross-modal students, our approach robustly identifies various types of anomalies, even in few-shot scenarios." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 305, + 68, + 553, + 263 + ], + "blocks": [ + { + "bbox": [ + 305, + 68, + 553, + 263 + ], + "lines": [ + { + "bbox": [ + 305, + 68, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 305, + 68, + 553, + 263 + ], + "type": "image", + "image_path": "fdcec2dc2e27da84955a78ed877e7a2b9baf617194c3a98b20023ff009851cb5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 360, + 294, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 360, + 294, + 467 + ], + "spans": [ + { + "bbox": [ + 54, + 360, + 294, + 467 + ], + "type": "text", + "content": "face normal maps capture fine-grained surface orientation details and preserve subtle textures more effectively than depth maps. This contrasts with previous work, which uses either raw point clouds or depth maps. Raw point clouds exacerbate the domain gap, while depth images tend to smooth out small shape variations and are less sensitive to small defects. These findings allow us to effectively bridge the modality gap and therefore significantly reduce false positives." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 473, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 473, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 54, + 473, + 295, + 581 + ], + "type": "text", + "content": "Cross-modal Distillation. Let " + }, + { + "bbox": [ + 54, + 473, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{2D}" + }, + { + "bbox": [ + 54, + 473, + 295, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 473, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{3D}" + }, + { + "bbox": [ + 54, + 473, + 295, + 581 + ], + "type": "text", + "content": " denote the two pre-trained Vision Transformer (ViT) teachers for the 2D RGB images and 3D surface normals, respectively. Inspired by CFM [5], each teacher receives input from its native domain and outputs high-level features " + }, + { + "bbox": [ + 54, + 473, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{2D}^T" + }, + { + "bbox": [ + 54, + 473, + 295, + 581 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 54, + 473, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{3D}^T" + }, + { + "bbox": [ + 54, + 473, + 295, + 581 + ], + "type": "text", + "content": " during training. In parallel, the corresponding cross-modal distillation student, " + }, + { + "bbox": [ + 54, + 473, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{3D\\rightarrow 2D}" + }, + { + "bbox": [ + 54, + 473, + 295, + 581 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 54, + 473, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{2D\\rightarrow 3D}" + }, + { + "bbox": [ + 54, + 473, + 295, + 581 + ], + "type": "text", + "content": ", takes features from one modality as input and attempts to reproduce the output of the teacher network from the other modality:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 71, + 586, + 295, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 586, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 71, + 586, + 295, + 604 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {2 D} ^ {S} = \\mathbf {S} _ {3 D \\rightarrow 2 D} \\big (\\mathbf {f} _ {3 D} ^ {T} \\big), \\quad \\mathbf {f} _ {3 D} ^ {S} = \\mathbf {S} _ {2 D \\rightarrow 3 D} \\big (\\mathbf {f} _ {2 D} ^ {T} \\big). \\qquad (1)", + "image_path": "8d91abc6a794ca6d9a2a14332994e2b09bfa4f5888a873cd80f12291c9a57828.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 609, + 285, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 609, + 285, + 621 + ], + "spans": [ + { + "bbox": [ + 67, + 609, + 285, + 621 + ], + "type": "text", + "content": "The overall distillation loss is applied bidirectionally:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 639, + 295, + 665 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 639, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 61, + 639, + 295, + 665 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c r o s s}} = \\left(1 - \\cos \\langle \\mathbf {f} _ {2 D} ^ {T}, \\mathbf {f} _ {2 D} ^ {S} \\rangle\\right) + \\left(1 - \\cos \\langle \\mathbf {f} _ {3 D} ^ {T}, \\mathbf {f} _ {3 D} ^ {S} \\rangle\\right). \\tag {2}", + "image_path": "f17005dfef9b143e6b5acbde30f7a77c372103e127e4cf427651ca5d5cbb8529.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 672, + 264, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 264, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 264, + 685 + ], + "type": "text", + "content": "3.3. Intra-Modal Teacher-Student Learning" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "While the cross-modal student excels in identifying semantic relationships between modalities, it relies on high-level" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 360, + 555, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 555, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 555, + 481 + ], + "type": "text", + "content": "feature representations with low spatial resolution, which may overlook subtle defects. To address this limitation, we introduce an intra-modal student for each modality, which performs intra-modal reconstruction on both feature level and image level. As a result, the intra-modal student specializes in capturing modality-specific fine-grained details for anomaly detection. In particular, the intra-modal student adopts a decoder-only architecture to perform multi-stage reverse distillation, progressively reconstructing back to the original image space for fine-grained anomaly localization." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 485, + 556, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 556, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 556, + 689 + ], + "type": "text", + "content": "Multi-Layer Reverse Distillation. To preserve both low-level textures and high-level semantics, we adopt a multilayer feature inversion strategy for the training of intramodal student. Let " + }, + { + "bbox": [ + 313, + 485, + 556, + 689 + ], + "type": "inline_equation", + "content": "\\{F_T^k\\}_{k=1}^K" + }, + { + "bbox": [ + 313, + 485, + 556, + 689 + ], + "type": "text", + "content": " denote a set of " + }, + { + "bbox": [ + 313, + 485, + 556, + 689 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 485, + 556, + 689 + ], + "type": "text", + "content": " feature maps selected from different layers of the teacher's encoder, with layer 1 representing the shallow representation and layer " + }, + { + "bbox": [ + 313, + 485, + 556, + 689 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 485, + 556, + 689 + ], + "type": "text", + "content": " being the deepest representation. In addition, " + }, + { + "bbox": [ + 313, + 485, + 556, + 689 + ], + "type": "inline_equation", + "content": "k = 0" + }, + { + "bbox": [ + 313, + 485, + 556, + 689 + ], + "type": "text", + "content": " denotes the original input image space. Our two intra-modal students, " + }, + { + "bbox": [ + 313, + 485, + 556, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{2D \\to 2D}" + }, + { + "bbox": [ + 313, + 485, + 556, + 689 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 485, + 556, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{3D \\to 3D}" + }, + { + "bbox": [ + 313, + 485, + 556, + 689 + ], + "type": "text", + "content": ", are designed to reverse the encoding process in a top-down manner. Specifically, we use " + }, + { + "bbox": [ + 313, + 485, + 556, + 689 + ], + "type": "inline_equation", + "content": "D^k" + }, + { + "bbox": [ + 313, + 485, + 556, + 689 + ], + "type": "text", + "content": " to denote the decoding module in the " + }, + { + "bbox": [ + 313, + 485, + 556, + 689 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 485, + 556, + 689 + ], + "type": "text", + "content": "-th layer of an intra-modal student, with each decoder block " + }, + { + "bbox": [ + 313, + 485, + 556, + 689 + ], + "type": "inline_equation", + "content": "D^k(\\cdot)" + }, + { + "bbox": [ + 313, + 485, + 556, + 689 + ], + "type": "text", + "content": " reconstructing the next fine-grained feature map: " + }, + { + "bbox": [ + 313, + 485, + 556, + 689 + ], + "type": "inline_equation", + "content": "F_S^{k-1} = D^k(F_S^k)" + }, + { + "bbox": [ + 313, + 485, + 556, + 689 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 485, + 556, + 689 + ], + "type": "inline_equation", + "content": "F_S^k" + }, + { + "bbox": [ + 313, + 485, + 556, + 689 + ], + "type": "text", + "content": " denotes the output of the " + }, + { + "bbox": [ + 313, + 485, + 556, + 689 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 485, + 556, + 689 + ], + "type": "text", + "content": "-th level of the intra-modal student and " + }, + { + "bbox": [ + 313, + 485, + 556, + 689 + ], + "type": "inline_equation", + "content": "F_T^K" + }, + { + "bbox": [ + 313, + 485, + 556, + 689 + ], + "type": "text", + "content": " serves as the student's starting point. The final output " + }, + { + "bbox": [ + 313, + 485, + 556, + 689 + ], + "type": "inline_equation", + "content": "F_S^0" + }, + { + "bbox": [ + 313, + 485, + 556, + 689 + ], + "type": "text", + "content": " reaches the original image resolution, enabling fine-grained anomaly localization." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "We then enforce the consistency between the teacher and the student's predictions to ensure that normal appearance is" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23293" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "reconstructed across different hidden layers. In addition, we incorporate an image-level reconstruction term that compares the final output " + }, + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "inline_equation", + "content": "F_{S}^{0}" + }, + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": " against the original input image " + }, + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": ", further preserving fidelity to subtle textures. Concretely, we calculate the intra-modal loss as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 59, + 137, + 294, + 198 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 137, + 294, + 198 + ], + "spans": [ + { + "bbox": [ + 59, + 137, + 294, + 198 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {i n t r a}} = \\sum_ {(S, T)} \\Big (\\sum_ {k = 1} ^ {K - 1} \\big (1 - \\cos \\langle F _ {S} ^ {k}, F _ {T} ^ {k} \\rangle \\big) + \\left\\| F _ {S} ^ {0} - I \\right\\| ^ {2} \\Big), \\\\ (S, T) \\in \\left\\{\\left(\\mathbf {S} _ {2 D \\rightarrow 2 D}, \\mathbf {T} _ {2 \\mathrm {D}}\\right), \\left(\\mathbf {S} _ {3 D \\rightarrow 3 D}, \\mathbf {T} _ {3 \\mathrm {D}}\\right)\\right\\}. \\tag {3} \\\\ \\end{array}", + "image_path": "545ec837486ebe695ed91917a2c381e761993d85225f3c6cee5d9d7ee4d17ba8.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 198, + 295, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 198, + 295, + 235 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 295, + 235 + ], + "type": "text", + "content": "This encourages the intra-modal students to capture both high-level semantics and fine-grained details of teacher's normality representation, even in few-shot settings." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 245, + 187, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 245, + 187, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 245, + 187, + 258 + ], + "type": "text", + "content": "3.4. Training and Inference" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 262, + 296, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 296, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 296, + 323 + ], + "type": "text", + "content": "Training. Our method jointly trains the cross-modal and intra-modal students using the respective objectives introduced in Sections 3.2 and 3.3. Specifically, we minimize the sum of the cross-modal loss in Eq. (2) and the intramodal loss in Eq. (3), yielding the final objective:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 121, + 330, + 294, + 342 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 330, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 121, + 330, + 294, + 342 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {c r o s s}} + \\mathcal {L} _ {\\text {i n t r a}}. \\tag {4}", + "image_path": "64d2f02fda0dff173fb0bb16bdae2220696155157ffbfa4b04d9065a6af25ffa.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 350, + 296, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 296, + 434 + ], + "type": "text", + "content": "Training is carried out on a few normal samples. The cross-modal student learns to reconstruct teacher features from the complementary modality, while the intra-modal student focuses on restoring its own modality in a reverse distillation manner. This combined strategy leverages both inter-modal consistency and fine-grained modality-specific cues for robust and fine-grained anomaly detection." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 437, + 296, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 296, + 689 + ], + "type": "text", + "content": "Anomaly Scoring. At test time, each of the four students " + }, + { + "bbox": [ + 55, + 437, + 296, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{2D\\rightarrow 3D}" + }, + { + "bbox": [ + 55, + 437, + 296, + 689 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 437, + 296, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{3D\\rightarrow 2D}" + }, + { + "bbox": [ + 55, + 437, + 296, + 689 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 437, + 296, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{2D\\rightarrow 2D}" + }, + { + "bbox": [ + 55, + 437, + 296, + 689 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 437, + 296, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{3D\\rightarrow 3D}" + }, + { + "bbox": [ + 55, + 437, + 296, + 689 + ], + "type": "text", + "content": " produces an anomaly map by measuring the discrepancy between its reconstructed output and the corresponding teacher's feature. In particular, for the intra-modal students " + }, + { + "bbox": [ + 55, + 437, + 296, + 689 + ], + "type": "inline_equation", + "content": "(S_{2D\\rightarrow 2D}" + }, + { + "bbox": [ + 55, + 437, + 296, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 437, + 296, + 689 + ], + "type": "inline_equation", + "content": "S_{3D\\rightarrow 3D})" + }, + { + "bbox": [ + 55, + 437, + 296, + 689 + ], + "type": "text", + "content": ", the intra-modal anomaly maps " + }, + { + "bbox": [ + 55, + 437, + 296, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{2D\\rightarrow 2D}" + }, + { + "bbox": [ + 55, + 437, + 296, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 437, + 296, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{3D\\rightarrow 3D}" + }, + { + "bbox": [ + 55, + 437, + 296, + 689 + ], + "type": "text", + "content": " are calculated by aggregating the intra-modal teacher-student discrepancy across multiple layers. For each hidden layer, we additionally apply " + }, + { + "bbox": [ + 55, + 437, + 296, + 689 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_2" + }, + { + "bbox": [ + 55, + 437, + 296, + 689 + ], + "type": "text", + "content": " -normalize to both the teacher's feature and the student's reconstructed feature at every spatial position. For each layer, we calculate the Euclidean distance between the student's reconstructed output and the teacher's original output to obtain a pixel-level error map. Next, we apply Gaussian smoothing to the obtained error map for a more coherent anomaly response. We repeat this process for all " + }, + { + "bbox": [ + 55, + 437, + 296, + 689 + ], + "type": "inline_equation", + "content": "K - 1" + }, + { + "bbox": [ + 55, + 437, + 296, + 689 + ], + "type": "text", + "content": " hidden layers and the final output layer, and then average all the error maps to obtain the two intra-modal maps " + }, + { + "bbox": [ + 55, + 437, + 296, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{2D\\rightarrow 2D}" + }, + { + "bbox": [ + 55, + 437, + 296, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 437, + 296, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{3D\\rightarrow 3D}" + }, + { + "bbox": [ + 55, + 437, + 296, + 689 + ], + "type": "text", + "content": " for both 2D and 3D modalities. The final intra-modal anomaly map " + }, + { + "bbox": [ + 55, + 437, + 296, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{\\mathrm{intra}}" + }, + { + "bbox": [ + 55, + 437, + 296, + 689 + ], + "type": "text", + "content": " is computed by averaging the two intra-modal maps " + }, + { + "bbox": [ + 55, + 437, + 296, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{2D\\rightarrow 2D}" + }, + { + "bbox": [ + 55, + 437, + 296, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 437, + 296, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{3D\\rightarrow 3D}" + }, + { + "bbox": [ + 55, + 437, + 296, + 689 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "The cross-modal students " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "(S_{2D\\to 3D}" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "S_{3D\\to 2D})" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": " operate similarly by measuring the discrepancy between the" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "student's reconstructed feature in one modality and the teacher's feature in the other modality. The resulting cross-modal anomaly maps from each student are denoted as " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{2D\\to 3D}" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{3D\\to 2D}" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": ". Likewise, we average the two cross-modal maps to obtain a single cross-modal anomaly map " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{\\mathrm{cross}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": ". The final overall anomaly map is obtained by the following:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 353, + 163, + 553, + 178 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 163, + 553, + 178 + ], + "spans": [ + { + "bbox": [ + 353, + 163, + 553, + 178 + ], + "type": "interline_equation", + "content": "\\mathbf {A} _ {\\text {f i n a l}} = \\alpha \\mathbf {A} _ {\\text {c r o s s}} + (1 - \\alpha) \\mathbf {A} _ {\\text {i n t r a}}, \\tag {5}", + "image_path": "14e594756e0a3a2f3ce5982502241d495e133a6f81c992c54b1da6787f5c62c7.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 184, + 553, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 184, + 553, + 257 + ], + "spans": [ + { + "bbox": [ + 313, + 184, + 553, + 257 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 184, + 553, + 257 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 184, + 553, + 257 + ], + "type": "text", + "content": " adjusts the relative influence of cross-modal and intra-modal anomaly map. High values in " + }, + { + "bbox": [ + 313, + 184, + 553, + 257 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{\\mathrm{final}}" + }, + { + "bbox": [ + 313, + 184, + 553, + 257 + ], + "type": "text", + "content": " indicate regions where the student network fails to mimic the teacher. By combining multi-modal alignment and reverse distillation reconstruction, our framework captures diverse anomaly signatures even in few-shot scenarios." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 266, + 391, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 391, + 280 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 391, + 280 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 286, + 430, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 286, + 430, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 286, + 430, + 300 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 304, + 554, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 554, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 554, + 388 + ], + "type": "text", + "content": "Datasets and Protocol. We evaluate on two public benchmarks: MVTec-3D-AD [1] and Eyecandies [2]. Following a standard few-shot normal-only protocol, we limit the available normal training images per class to " + }, + { + "bbox": [ + 313, + 304, + 554, + 388 + ], + "type": "inline_equation", + "content": "k \\in \\{5, 10, 50\\}" + }, + { + "bbox": [ + 313, + 304, + 554, + 388 + ], + "type": "text", + "content": ", randomly sampling and averaging results over ten runs. We also evaluate under a full-shot (all normal samples) setting for completeness." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 392, + 554, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 554, + 476 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 554, + 476 + ], + "type": "text", + "content": "Baselines. We compare against: (1) BTF [11], a classical 3D feature-based method without deep learning; (2) AST [24], an asymmetric T-S approach trained on normal samples; (3) M3DM [31], a hybrid fusion model exploiting memory banks; (4) CFM [5], which aligns 2D-3D features without memory banks; and (5) 3D-ADNAS [21], which uses neural architecture search for multi-modal fusion." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 480, + 554, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 554, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 554, + 553 + ], + "type": "text", + "content": "Evaluation Metrics. We follow standard anomaly detection metrics: (a) I-AUROC for image-level anomaly classification, (b) P-AUROC (pixel-level AUROC) for localization, and (c) AUPRO (area under per-region overlap) at specific false positive rates (FPR), typically " + }, + { + "bbox": [ + 313, + 480, + 554, + 553 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 480, + 554, + 553 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 480, + 554, + 553 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 313, + 480, + 554, + 553 + ], + "type": "text", + "content": ". Higher values denote better performance." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "Implementation Details. The intra-modal student is implemented as a multi-stage decoder, where each stage consists of three sequential convolutional blocks with normalization layers and GeLU activation to ensure nonlinearity. In particular, the final block employs ConvTranspose2d layers to up-sample the feature map back to the original image resolution and progressively decreases the dimensionality of the output channel to the final 3 channels (RGB). We implement our cross-modal student network as an encoder-decoder style network. The encoder first encodes the input patches into a hidden representation with two convolutional blocks, while the decoder reconstructs them in the cross-modal feature space through another convolutional" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23294" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 106, + 102, + 504, + 379 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Comparison of anomaly detection and localization performance on MVTec-3D-AD and Eyecandies with varying numbers of training shots. " + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "inline_equation", + "content": "\\mathbf{{ROC}}{}_{I},{\\mathbf{{ROC}}}_{P}" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": " ,and " + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "inline_equation", + "content": "\\mathbf{{PRO}}" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": " denote I-AUROC, P-AUROC, and AUPRO respectively. Bold indicates the best performance." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 106, + 102, + 504, + 379 + ], + "lines": [ + { + "bbox": [ + 106, + 102, + 504, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 102, + 504, + 379 + ], + "type": "table", + "html": "
k-ShotMethodMVTec-3D-ADEyecandies
\\( ROC_I \\)\\( ROC_P \\)PRO-30%PRO-1%\\( ROC_I \\)\\( ROC_P \\)PRO-30%PRO-1%
5BTF [11]0.6710.9800.9200.2880.6520.8150.7380.202
AST [24]0.6800.9500.9030.1580.6330.7410.6910.143
M3DM [31]0.8220.9840.9370.3300.7640.8710.8070.251
CFM [5]0.8110.9860.9490.3820.7950.8790.8010.260
3D-ADNAS [21]0.826---0.7750.875--
FIND (Ours)0.8990.9910.9630.4200.8400.9510.8620.324
10BTF [11]0.6950.9830.9280.3080.6850.8340.8060.217
AST [24]0.6890.9460.8350.1740.6710.7670.6240.161
M3DM [31]0.8450.9860.9430.3550.8240.8900.8120.266
CFM [5]0.8450.9870.9540.3980.8380.9030.8250.284
3D-ADNAS [21]0.848---0.8070.869--
FIND (Ours)0.9210.9930.9670.4340.8680.9590.8890.350
50BTF [11]0.8060.9890.9470.3560.7210.8560.8240.240
AST [24]0.7940.9740.9290.3350.7390.8620.7150.118
M3DM [31]0.9070.9890.9550.3870.8360.9330.8460.304
CFM [5]0.9060.9910.9650.4310.8520.9260.8510.315
3D-ADNAS [21]0.890---0.8680.912--
FIND (Ours)0.9520.9950.9730.4550.8970.9760.9070.365
ALLBTF [11]0.8650.9920.9590.3830.7400.8830.8450.261
AST [24]0.9370.9760.9440.3980.7800.9020.7440.149
M3DM [31]0.9450.9920.9640.3940.8820.9770.8870.331
CFM [5]0.9540.9930.9710.4550.8810.9740.8870.335
3D-ADNAS [21]0.951---0.9460.970--
FIND (Ours)0.9780.9950.9830.4560.9550.9830.9250.395
", + "image_path": "cc972593557fb6dad684c6e4ea02a058c494a1e3331605ab7b4c09246d345f71.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 93, + 408, + 517, + 605 + ], + "blocks": [ + { + "bbox": [ + 140, + 388, + 468, + 400 + ], + "lines": [ + { + "bbox": [ + 140, + 388, + 468, + 400 + ], + "spans": [ + { + "bbox": [ + 140, + 388, + 468, + 400 + ], + "type": "text", + "content": "Table 2. Comparisons of per-category anomaly detection performance on MVTec-3D-AD." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 93, + 408, + 517, + 605 + ], + "lines": [ + { + "bbox": [ + 93, + 408, + 517, + 605 + ], + "spans": [ + { + "bbox": [ + 93, + 408, + 517, + 605 + ], + "type": "table", + "html": "
MethodBagelCable GlandCarrotCookieDowelFoamPeachPotatoRopeTireMean
I-AUROC
BTF [11]0.9380.7650.9720.8880.9600.6640.9040.9290.9820.7260.865
AST [24]0.9830.8730.9760.9710.9320.8850.9740.9811.0000.7970.937
M3DM [31]0.9940.9090.9720.9760.9600.9420.9730.8990.9720.8500.945
CFM [5]0.9940.8880.9840.9930.9800.8880.9410.9430.9800.9530.954
3D-ADNAS [21]0.9971.0000.9710.9860.9660.9480.8970.8731.0000.8670.951
Shape Guided [4]0.9860.8940.9830.9910.9760.8570.9900.9650.9900.8690.947
LSFA [29]1.0000.9390.9820.9890.9610.9510.9830.9620.9890.9510.971
FIND (Ours)0.9840.9700.9850.9720.9820.9790.9740.9800.9880.9700.978
AUPRO@30%
BTF [11]0.9760.9690.9790.9730.9330.8880.8960.9120.9500.9710.959
AST [24]0.9700.9470.9810.9390.9130.9060.9790.9820.8890.9400.944
M3DM [31]0.9700.9710.9790.9500.9410.9320.9770.9710.9710.9750.964
CFM [5]0.9790.9720.9820.9450.9500.9680.9800.9430.9500.9810.971
Shape Guided [4]0.9810.9730.9820.9710.9620.9780.9810.9830.9740.9750.976
LSFA [29]0.9860.9740.9810.9460.9250.9410.9830.9830.9740.9830.968
FIND (Ours)0.9890.9770.9810.9830.9820.9860.9840.9800.9900.9780.983
", + "image_path": "ac3325aabec7bc9b416b026048ce9ad337930e0b874126ab72e88a3ef7548fa9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 624, + 295, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 295, + 647 + ], + "type": "text", + "content": "blocks, ensuring both preservation of the locality and improved alignment between the modalities." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "To ensure a fair comparison with previous work, we use the same ViT-B/8 DINO backbone as CFM [5], which is shared for both RGB and surface normal modalities. We treat surface normals as 3-channel RGB-style inputs and normalize them using ImageNet statistics before feeding" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 624, + 555, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 555, + 708 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 555, + 708 + ], + "type": "text", + "content": "them into DINO. During training, the backbone weights remain fixed, and we optimize only the four student networks. We extract hidden features from teacher's 4, 6, 8-th layers for intra-modal reverse distillation, while the 10-th layer's feature is employed as the input for cross-modal student. We use the ADAM optimizer for training with a learning rate of " + }, + { + "bbox": [ + 313, + 624, + 555, + 708 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 313, + 624, + 555, + 708 + ], + "type": "text", + "content": ", and set " + }, + { + "bbox": [ + 313, + 624, + 555, + 708 + ], + "type": "inline_equation", + "content": "\\alpha = 0.5" + }, + { + "bbox": [ + 313, + 624, + 555, + 708 + ], + "type": "text", + "content": " for all-shot tasks and" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23295" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 72, + 292, + 194 + ], + "blocks": [ + { + "bbox": [ + 62, + 72, + 292, + 194 + ], + "lines": [ + { + "bbox": [ + 62, + 72, + 292, + 194 + ], + "spans": [ + { + "bbox": [ + 62, + 72, + 292, + 194 + ], + "type": "image", + "image_path": "314955c28ff1a8386b969f81c76c131051abd5da2b3e22b3dbefefd4130e5d40.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 62, + 204, + 291, + 312 + ], + "blocks": [ + { + "bbox": [ + 62, + 204, + 291, + 312 + ], + "lines": [ + { + "bbox": [ + 62, + 204, + 291, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 204, + 291, + 312 + ], + "type": "image", + "image_path": "7f6bf279ded68aa23f81b46a77a119b9a3007ebfa48616f8b39567b252c56576.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 62, + 321, + 291, + 430 + ], + "blocks": [ + { + "bbox": [ + 62, + 321, + 291, + 430 + ], + "lines": [ + { + "bbox": [ + 62, + 321, + 291, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 321, + 291, + 430 + ], + "type": "image", + "image_path": "fac46808338bade6a8a5bf56d51c18da0b15e10cb1b8373eb55801e0c6ce9564.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 444, + 295, + 510 + ], + "lines": [ + { + "bbox": [ + 55, + 444, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 444, + 295, + 510 + ], + "type": "text", + "content": "Figure 4. Qualitative results on MVTec-3D-AD with various types of anomalies. From left to right: input RGB images, surface normals, GT anomaly maps, anomaly maps from 2D student set, anomaly maps from 3D student set, and our final predictions. Our final model consistently produces more accurate anomaly maps across different anomaly types." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 58, + 521, + 295, + 591 + ], + "blocks": [ + { + "bbox": [ + 58, + 521, + 295, + 591 + ], + "lines": [ + { + "bbox": [ + 58, + 521, + 295, + 591 + ], + "spans": [ + { + "bbox": [ + 58, + 521, + 295, + 591 + ], + "type": "image", + "image_path": "8dbab268d4a2f5c6efed4f14cb6a01bcef492b9245964b7b81a2159d9e92a61b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 599, + 295, + 633 + ], + "lines": [ + { + "bbox": [ + 55, + 599, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 295, + 633 + ], + "type": "text", + "content": "Figure 5. Anomaly score distributions on bagel and cookie (MVTec-3D-AD, 10-shot). Our method shows better separation between normal and anomalous samples." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 654, + 167, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 167, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 167, + 664 + ], + "type": "inline_equation", + "content": "\\alpha = 0.8" + }, + { + "bbox": [ + 55, + 654, + 167, + 664 + ], + "type": "text", + "content": " for few-shot tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": "For MVTec-3D-AD, each sample includes an RGB image and an unstructured point cloud. We first organize the point cloud into a 2D grid(XYZ map) corresponding to the rgb image size " + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": ". We then compute per-pixel surface" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 71, + 553, + 252 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 553, + 252 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 553, + 252 + ], + "type": "table", + "html": "
CategoryIntra-studentCross-studentIntra + Cross
ImagePixelImagePixelImagePixel
Bagel96.399.798.899.599.599.6
Cable Gland74.497.679.098.382.498.3
Carrot97.499.899.499.997.399.8
Cookie86.498.491.897.292.597.8
Dowel75.298.183.899.382.699.2
Foam78.999.197.699.898.199.8
Peach97.099.895.799.999.499.9
Potato92.999.892.799.995.899.9
Rope97.699.896.699.898.899.8
Tire66.798.775.199.274.499.2
Average86.399.191.199.392.199.3
", + "image_path": "2388f5486de21d29ca1148893218b92a203a8215fa9a1804586e03e95745688f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 315, + 321, + 558, + 504 + ], + "blocks": [ + { + "bbox": [ + 313, + 260, + 555, + 293 + ], + "lines": [ + { + "bbox": [ + 313, + 260, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 555, + 293 + ], + "type": "text", + "content": "Table 3. Ablation studies on intra-modal, cross-modal and dual students. We report the 10-shot categorical performance in image and pixel AUROC (\\%) on the MVTec-3D dataset." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 321, + 558, + 504 + ], + "lines": [ + { + "bbox": [ + 315, + 321, + 558, + 504 + ], + "spans": [ + { + "bbox": [ + 315, + 321, + 558, + 504 + ], + "type": "table", + "html": "
CategoryDepthPoint CloudSurface Normal
ImagePixelImagePixelImagePixel
Bagel99.099.498.399.599.599.6
Cable Gland68.897.774.197.682.498.3
Carrot82.799.392.499.197.399.8
Cookie84.498.590.797.292.597.8
Dowel87.199.075.599.082.699.2
Foam81.689.892.499.598.199.8
Peach85.799.393.799.699.499.9
Potato68.499.184.398.795.899.9
Rope95.799.694.299.798.899.8
Tire66.897.673.699.374.499.2
Average82.097.986.898.992.199.3
", + "image_path": "937c3808e4f9468cb412eb3b99b88496b195f6f24d945e72b19428927a64cc2e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 510, + 555, + 544 + ], + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 544 + ], + "type": "text", + "content": "Table 4. Ablation studies on 3D modality input selection. We report the 10-shot categorical performance in image and pixel AUROC (\\%) on the MVTec-3D dataset." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": "normals by fitting a local plane to the neighbors of each point (or using the cross-product of adjacent triangle faces in a mesh representation). This produces a 3-channel normal vector map for the surface of the object. Following prior work [5, 31], we remove background points and fill the removed regions in the RGB image with a neutral value (black) to avoid false anomalies from the background. We finally resize both the RGB image and the normal map to " + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": " pixels before feeding them to the model. For Eyecandies, we use normal maps with global lighting as the 3D modality." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23296" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 155, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 155, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 155, + 85 + ], + "type": "text", + "content": "4.2. Primary Results" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 224 + ], + "type": "text", + "content": "Few-shot Performance on MVtec-3D-AD. Table 1 shows that FIND exceeds all competitors: BTF, AST, M3DM, CFM and 3D-ADNAS, in all few-shot settings (5, 10, 50 shots) and evaluation metrics at both the image and pixel levels. For example, in the challenging 5-shot setting, our method achieves an I-AUROC of 0.899. With 10-shot training, the I-AUROC performance is increased to 0.921. Even with 50 shots, FIND continues to outperform strong memory-based methods (e.g., M3DM). These consistent gains highlight the effectiveness of FIND in limited data scenarios." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 229, + 296, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 229, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 229, + 296, + 350 + ], + "type": "text", + "content": "Full-shot Performance on MVtec-3D-AD. Table 1 presents the anomaly detection performance in the full-shot setting at both the image and pixel levels, while Table 2 reports results for each category. FIND surpasses existing methods in both image-level and pixel-level detection, achieving the highest average I-AUROC " + }, + { + "bbox": [ + 55, + 229, + 296, + 350 + ], + "type": "inline_equation", + "content": "(97.8\\%)" + }, + { + "bbox": [ + 55, + 229, + 296, + 350 + ], + "type": "text", + "content": " and AUPRO@ " + }, + { + "bbox": [ + 55, + 229, + 296, + 350 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 55, + 229, + 296, + 350 + ], + "type": "inline_equation", + "content": "(98.3\\%)" + }, + { + "bbox": [ + 55, + 229, + 296, + 350 + ], + "type": "text", + "content": ", demonstrating its robustness across various object classes. Notably, our approach excels in challenging categories (e.g., Cable Gland, Foam), indicating its ability to capture fine-grained anomalous patterns." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 356, + 296, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 356, + 296, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 356, + 296, + 513 + ], + "type": "text", + "content": "Results on Eyecandies. As shown in Table 1, FIND also sets a new state-of-the-art on Eyecandies under both few-shot and full-shot protocols, significantly outperforming existing methods. For example, at 5-shot, FIND achieves an I-AUROC score of 0.840, exceeding the other baselines significantly. This performance gain extends to pixel-level metrics (P-AUROC and AUPRO), demonstrating a superior localization of subtle anomalies. Furthermore, with 50 normal samples for training, our method even outperforms the full-shot trained M3DM and CFM in terms of AUROC and AUPRO. These results highlight the effectiveness of our approach in various shots of training data scenarios, reinforcing its robustness and generalizability." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 525, + 174, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 174, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 174, + 538 + ], + "type": "text", + "content": "4.3. Qualitative Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 544, + 296, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 296, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 296, + 640 + ], + "type": "text", + "content": "Fig. 4 shows that our method is robust to various types of anomaly (shape-only, color-only, combined) by effectively exploiting 2D-3D complementarity. The 2D student set " + }, + { + "bbox": [ + 55, + 544, + 296, + 640 + ], + "type": "inline_equation", + "content": "(\\mathbf{S}_{3D\\rightarrow 2D},\\mathbf{S}_{2D\\rightarrow 2D})" + }, + { + "bbox": [ + 55, + 544, + 296, + 640 + ], + "type": "text", + "content": " captures subtle color anomalies often overlooked by geometric cues, while the 3D student set " + }, + { + "bbox": [ + 55, + 544, + 296, + 640 + ], + "type": "inline_equation", + "content": "(\\mathbf{S}_{2D\\rightarrow 3D},\\mathbf{S}_{3D\\rightarrow 3D})" + }, + { + "bbox": [ + 55, + 544, + 296, + 640 + ], + "type": "text", + "content": " detects shape deformations that are imperceptible in RGB. Combining these maps improves localization accuracy and reduces false positives." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": "Fig. 5 further compares the anomaly score distributions of normal and anomalous samples under the 10-shot setting for CFM and FIND on two MVTec-3D-AD classes (bagel and cookie). Our method yields more clearly separable distributions between normal and abnormal samples, contributing to improved detection accuracy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 72, + 416, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 416, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 416, + 84 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 89, + 555, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 555, + 293 + ], + "type": "text", + "content": "Intra-modal vs. Cross-modal vs. Dual Student. We evaluate the contributions of the intra-modal and cross-modal students, with results shown in Table 3. Intra-modal students capture modality-specific anomalies and guide the model to reproduce a normal version of the query sample at the pixel level, enhancing the detection of local defects. Removing them causes a notable drop in detection accuracy for categories such as cable gland, potato, and rope. However, intra-modal students operating within a single modality cannot leverage complementary information from the other modality, leaving those cross-modal anomalies undetected. In contrast, cross-modal students learn the expected normal correspondence between an object's color and shape appearance and detect anomalies by identifying inconsistencies between modalities. By combining both intra-modal and cross-modal students, we leverage their complementary strengths and achieve the best results." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 297, + 556, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 556, + 490 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 556, + 490 + ], + "type": "text", + "content": "3D Modality Input Choices. In a 10-shot setting on MVTec-3D-AD, we compare raw point clouds (with a PointNet encoder), depth maps(from point clouds), and surface-normal maps for anomaly detection. The results are shown in Table 4. Surface normals achieve the highest I-AUROC, outperforming depth maps and raw point clouds. This indicates that normal maps preserve fine-grained geometric details that can highlight small defects. In contrast, depth maps lack this level of detail, resulting in lower accuracy. The point cloud baseline performs worst, likely due to a modality gap introduced by using a separate PointNet-based 3D feature extractor. Such gaps arise from misaligned representations when using a point cloud network versus image-based features. Our results show that surface normal maps provide the most effective shape modality for 3D anomaly detection in few-shot scenarios." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 499, + 388, + 512 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 499, + 388, + 512 + ], + "spans": [ + { + "bbox": [ + 314, + 499, + 388, + 512 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 520, + 556, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 556, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 556, + 653 + ], + "type": "text", + "content": "In this work, we address the critical challenges of multimodal anomaly detection in few-shot scenarios by introducing FIND, a dual-student framework that integrates intramodal reverse distillation and cross-modal feature mapping. FIND overcomes data scarcity and modality gaps by exploiting collaborative learning between two students: one dedicated to fine-grained intra-modal anomaly localization through reverse distillation, and another focused on cross-modal inconsistency detection. Extensive experiments demonstrate that FIND achieves new state-of-the-art results in both few-shot and full-shot settings." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 662, + 417, + 676 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 662, + 417, + 676 + ], + "spans": [ + { + "bbox": [ + 314, + 662, + 417, + 676 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "This work is supported by the Agency for Science, Technology and Research (A*STAR) under its MTC Programmatic Funds (Grant No. M23L7b0021)." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23297" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "text", + "content": "[1] Paul Bergmann, Xin Jin, David Sattlegger, and Carsten Steger. The mvtec 3d-ad dataset for unsupervised 3d anomaly detection and localization. In Proceedings of the 17th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications, VISIGRAPP, pages 202-213, 2022. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 159, + 296, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 159, + 296, + 213 + ], + "spans": [ + { + "bbox": [ + 61, + 159, + 296, + 213 + ], + "type": "text", + "content": "[2] Luca Bonfiglioli, Marco Toschi, Davide Silvestri, Nicola Fioraio, and Daniele De Gregorio. The eyecandies dataset for unsupervised multimodal anomaly detection and localization. In 16th Asian Conference on Computer Vision (ACCV), pages 459-475, 2022. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 215, + 296, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 215, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 215, + 296, + 258 + ], + "type": "text", + "content": "[3] Mengxi Chen and ... Enhanced multimodal representation learning with cross-modal KD. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11766-11775, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 261, + 295, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 261, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 261, + 295, + 293 + ], + "type": "text", + "content": "[4] Yu-Min Chu, Liu Chieh, Ting-I Hsieh, Hwann-Tzong Chen, and Tyng-Luh Liu. Shape-guided dual-memory learning for 3d anomaly detection. 2023. 1, 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 296, + 296, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 296, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 296, + 296, + 350 + ], + "type": "text", + "content": "[5] Alex Costanzino, Pierluigi Zama Ramirez, Giuseppe Lisanti, and Luigi Di Stefano. Multimodal industrial anomaly detection by crossmodal feature mapping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17234-17243, 2024. 1, 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 352, + 296, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 352, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 352, + 296, + 396 + ], + "type": "text", + "content": "[6] Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one-class embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9737-9746, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 398, + 296, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 398, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 398, + 296, + 441 + ], + "type": "text", + "content": "[7] Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one-class embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9737-9746, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 444, + 295, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 444, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 62, + 444, + 295, + 498 + ], + "type": "text", + "content": "[8] Zheng Fang, Xiaoyang Wang, Haocheng Li, Jiejie Liu, Qiugui Hu, and Jimin Xiao. Fastrecon: Few-shot industrial anomaly detection via fast feature reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17481-17490, 2023. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 500, + 295, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 500, + 295, + 564 + ], + "spans": [ + { + "bbox": [ + 62, + 500, + 295, + 564 + ], + "type": "text", + "content": "[9] Zhihao Gu, Jiangning Zhang, Liang Liu, Xu Chen, Jinlong Peng, Zhenye Gan, Guannan Jiang, Annan Shu, Yabiao Wang, and Lizhuang Ma. Rethinking reverse distillation for multi-modal anomaly detection. In Thirty-Eighth AAAI Conference on Artificial Intelligence, AAAI, pages 8445-8453, 2024. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 567, + 295, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 567, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 57, + 567, + 295, + 610 + ], + "type": "text", + "content": "[10] Yu Hong, Hang Dai, and Yong Ding. Cross-modality knowledge distillation network for monocular 3d object detection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 87-104, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 613, + 295, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 295, + 666 + ], + "type": "text", + "content": "[11] Eliahu Horwitz and Yedid Hoshen. Back to the feature: classical 3d features are (almost) all you need for 3d anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2968-2977, 2023. 1, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[12] Chaoqin Huang, Haoyan Guan, Aofan Jiang, Ya Zhang, Michael Spratling, and Yan-Feng Wang. Registration based few-shot anomaly detection. In European Conference on Computer Vision, pages 303-319. Springer, 2022. 1, 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "type": "text", + "content": "[13] Fushuo Huo, Wenchao Xu, Jingcai Guo, Haozhao Wang, and Song Guo. " + }, + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "type": "inline_equation", + "content": "C^2" + }, + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "type": "text", + "content": "kd: Bridging the modality gap for cross-modal knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7151-7160, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 130, + 554, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 130, + 554, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 130, + 554, + 184 + ], + "type": "text", + "content": "[14] Jongheon Jeong, Yang Zou, Taewan Kim, Dongqing Zhang, Avinash Ravichandran, and Onkar Dabeer. Winclip: Zero/few-shot anomaly classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19606-19616, 2023. 1, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 186, + 555, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 555, + 229 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 555, + 229 + ], + "type": "text", + "content": "[15] Pilhyeon Lee and ... Decomposed cross-modal distillation for RGB-based temporal action detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages -, 2023. CVPR 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 231, + 554, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 231, + 554, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 231, + 554, + 274 + ], + "type": "text", + "content": "[16] Yujin Lee, Seoyoon Jang, and Hyunsoo Yoon. Anople: Few-shot anomaly detection via bi-directional prompt learning with only normal samples. arXiv preprint arXiv:2408.13516, 2024. 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 276, + 554, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 276, + 554, + 352 + ], + "spans": [ + { + "bbox": [ + 317, + 276, + 554, + 352 + ], + "type": "text", + "content": "[17] Wenqiao Li, Xiaohao Xu, Yao Gu, Bozhong Zheng, Shenghua Gao, and Yingna Wu. Towards scalable 3d anomaly detection and localization: A benchmark via 3d anomaly synthesis and a self-supervised learning network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22207-22216, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 354, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 398 + ], + "type": "text", + "content": "[18] Hanzhe Liang, Guoyang Xie, Chengbin Hou, Bingshu Wang, Can Gao, and Jinbao Wang. Look inside for more: Internal spatial modality perception for 3d anomaly detection. arXiv preprint arXiv:2412.13461, 2024. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 399, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 553, + 453 + ], + "type": "text", + "content": "[19] Victor Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Y Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. Advances in Neural Information Processing Systems, 35:17612-17625, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 456, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 553, + 499 + ], + "type": "text", + "content": "[20] Chieh Liu, Yu-Min Chu, Ting-I Hsieh, Hwann-Tzong Chen, and Tyng-Luh Liu. Learning diffusion models for multi-view anomaly detection. In European Conference on Computer Vision, pages 328-345. Springer, 2025. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 501, + 554, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 501, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 501, + 554, + 555 + ], + "type": "text", + "content": "[21] Kaifang Long, Guoyang Xie, Lianbo Ma, Jiaqi Liu, and Zhichao Lu. Revisiting multimodal fusion for 3d anomaly detection from an architectural perspective. In Thirty-Eighth AAAI Conference on Artificial Intelligence, AAAI, 2025. 1, 2, 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 557, + 554, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 557, + 554, + 612 + ], + "spans": [ + { + "bbox": [ + 317, + 557, + 554, + 612 + ], + "type": "text", + "content": "[22] Xiang Ma, Xuemei Li, Lexin Fang, and Caiming Zhang. Bridging the modality gap: Dimension information alignment and sparse spatial constraint for image-text matching. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 5074-5082, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 613, + 554, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 554, + 679 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 554, + 679 + ], + "type": "text", + "content": "[23] Nicolae-Cătălin Ristea, Neelu Madan, Radu Tudor Ionescu, Kamal Nasrollahi, Fahad Shahbaz Khan, Thomas B. Moeslund, and Mubarak Shah. Self-supervised predictive convolutional attentive block for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13576-13586, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[24] Marco Rudolph, Tom Wehrbein, Bodo Rosenhahn, and Bastian Wandt. Asymmetric student-teacher networks for industrial anomaly detection. In IEEE/CVF Winter Conference on" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23298" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "Applications of Computer Vision, WACV, pages 2592-2602, 2023. 1, 2, 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[25] Pritam Sarkar and Ali Etemad. XKD: Cross-modal knowledge distillation with domain alignment for video representation learning. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2024. to appear. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "type": "text", + "content": "[26] Wenbo Sui, Daniel Lichau, Josselin Lefèvre, and Harold Philippeau. Incomplete multimodal industrial anomaly detection via cross-modal distillation, 2024. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 294, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 241 + ], + "type": "text", + "content": "[27] Long Tian, Hongyi Zhao, Ruiying Lu, Rongrong Wang, Yu-jie Wu, Liming Wang, Xiongpeng He, and Xiyang Liu. Foct: Few-shot industrial anomaly detection with foreground-aware online conditional transport. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6241-6249, 2024. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 243, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 294, + 308 + ], + "type": "text", + "content": "[28] Dinh Tien Tran, Anh Tuan Nguyen, Nguyen Hoang Tran, Ta Duc Huy, Soan T. M. Duong, Chanh D. Tr. Nguyen, and Steven Q. H. Truong. Revisiting reverse distillation for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24511-24520, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 310, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 294, + 365 + ], + "type": "text", + "content": "[29] Yuanpeng Tu, Boshen Zhang, Liang Liu, Yuxi Li, Jiangning Zhang, Yabiao Wang, Chengjie Wang, and Cairong Zhao. Self-supervised feature adaptation for 3d industrial anomaly detection. In European Conference on Computer Vision, pages 75-91. Springer, 2025. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 366, + 294, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 366, + 294, + 410 + ], + "spans": [ + { + "bbox": [ + 56, + 366, + 294, + 410 + ], + "type": "text", + "content": "[30] Huiqun Wang, Yiping Bao, Panwang Pan, Zeming Li, Xiao Liu, Ruijie Yang, and Di Huang. Multi-modal relation distillation for unified 3d representation learning. In European Conference on Computer Vision (ECCV), 2024. to appear. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 294, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 466 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 466 + ], + "type": "text", + "content": "[31] Yue Wang, Jinlong Peng, Jiangning Zhang, Ran Yi, Yabiao Wang, and Chengjie Wang. Multimodal industrial anomaly detection via hybrid fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8032-8041, 2023. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 468, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 294, + 521 + ], + "type": "text", + "content": "[32] Zeyu Wang, Dingwen Li, Chenxu Luo, Cihang Xie, and Xiaodong Yang. DistillBEV: Boosting multi-camera 3d object detection with cross-modal knowledge distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8637-8646, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 523, + 294, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 577 + ], + "type": "text", + "content": "[33] Zihui Xue, Zhengqi Gao, Sucheng Ren, and Hang Zhao. The modality focusing hypothesis: Towards understanding cross-modal knowledge distillation. In International Conference on Learning Representations (ICLR), 2023. Top " + }, + { + "bbox": [ + 56, + 523, + 294, + 577 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 56, + 523, + 294, + 577 + ], + "type": "text", + "content": " at ICLR 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "type": "text", + "content": "[34] Xu Yan, Heshen Zhan, Chaoda Zheng, Jiantao Gao, Ruimao Zhang, Shuguang Cui, and Zhen Li. Let images give you more: Point cloud cross-modal training for shape analysis. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "text", + "content": "[35] Heeseung Yun, Joonil Na, and Gunhee Kim. Dense 2d-3d indoor prediction with sound via aligned cross-modal distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7829-7838, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "text", + "content": "[36] Vitjan Zavrtanik, Matej Kristan, and Danijel Skocaj. Cheating depth: Enhancing 3d surface anomaly detection via depth" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 329 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 105 + ], + "type": "text", + "content": "simulation. In IEEE/CVF Winter Conference on Applications of Computer Vision, WACV, pages 2153-2161. IEEE, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "type": "text", + "content": "[37] Tianlu Zhang, Hongyuan Guo, Qiang Jiao, Qiang Zhang, and Jungong Han. Efficient rgb-t tracking via cross-modality distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5815-5824, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 555, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 555, + 217 + ], + "type": "text", + "content": "[38] Xuan Zhang, Shiyu Li, Xi Li, Ping Huang, Jiulong Shan, and Ting Chen. Destseg: Segmentation guided denoising student-teacher for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3914–3923, 2023. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 219, + 553, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 283 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 283 + ], + "type": "text", + "content": "[39] Shengchao Zhou, Weizhou Liu, Chen Hu, Shuchang Zhou, and Chao Ma. Unidistill: A universal cross-modality knowledge distillation framework for 3d object detection in bird's-eye view. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5116-5125, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 285, + 553, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 285, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 553, + 329 + ], + "type": "text", + "content": "[40] Zheyuan Zhou, Le Wang, Naiyu Fang, Zili Wang, Lemiao Qiu, and Shuyou Zhang. R3d-ad: Reconstruction via diffusion for 3d anomaly detection. In European Conference on Computer Vision, pages 91-107. Springer, 2025. 1" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23299" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/FLOAT_ Generative Motion Latent Flow Matching for Audio-driven Talking Portrait/5eb040a4-3c63-4680-a42f-da4ce86274a7_content_list.json b/2025/FLOAT_ Generative Motion Latent Flow Matching for Audio-driven Talking Portrait/5eb040a4-3c63-4680-a42f-da4ce86274a7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5c31b33641c76eac6a7dbc911f34b37f145b8928 --- /dev/null +++ b/2025/FLOAT_ Generative Motion Latent Flow Matching for Audio-driven Talking Portrait/5eb040a4-3c63-4680-a42f-da4ce86274a7_content_list.json @@ -0,0 +1,1855 @@ +[ + { + "type": "text", + "text": "FLOAT: Generative Motion Latent Flow Matching for Audio-driven Talking Portrait", + "text_level": 1, + "bbox": [ + 114, + 128, + 883, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Taekyung $\\mathbf{K}\\mathrm{i}^{1\\star}$ Dongchan Min Gyeongsu Chae", + "bbox": [ + 267, + 202, + 723, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ KAIST $^{2}$ DeepBrain AI Inc.", + "bbox": [ + 390, + 222, + 607, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{taekyung.ki, alsehdcks95}@kaist.ac.kr gc@deepbrain.io", + "bbox": [ + 253, + 241, + 745, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://deepbrainai-research.github.io/float/", + "bbox": [ + 295, + 258, + 696, + 273 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/426ecc2dd395074757380a5c7cc5bc5fe8afac3b04755b79925b975e9e7bc694.jpg", + "image_caption": [ + "Figure 1. FLOAT can generate a talking portrait video from a single source image and audio where the talking motion is generated by the motion latent flow matching. It can enhance the emotion-related talking motion by leveraging speech-driven emotion labels, a natural way of emotion-aware motion control." + ], + "image_footnote": [], + "bbox": [ + 161, + 306, + 831, + 498 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 559, + 326, + 574 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the rapid advancement of diffusion-based generative models, portrait image animation has achieved remarkable results. However, it still faces challenges in temporally consistent video generation and fast sampling due to its iterative sampling nature. This paper presents FLOAT, an audio-driven talking portrait video generation method based on flow matching generative model. Instead of a pixel-based latent space, we take advantage of a learned orthogonal motion latent space, enabling efficient generation and editing of temporally consistent motion. To achieve this, we introduce a transformer-based vector field predictor with an effective frame-wise conditioning mechanism. Additionally, our method supports speech-driven emotion enhancement, enabling a natural incorporation of expressive motions. Extensive experiments demonstrate that our method outperforms state-of-the-art audio-driven talking portrait methods in terms of visual quality, motion fidelity, and efficiency.", + "bbox": [ + 86, + 590, + 485, + 849 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 559, + 643, + 574 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Animating a single image using a driving audio (i.e., audiodriven talking portrait generation) has gained significant attention in recent years for its great potential in avatar creation, video conferencing, virtual avatar chat, and user-friendly customer service. It aims to synthesize natural talking motion from audio signals, including accurate lip synchronization, rhythmical head movements, and fine-grained facial expressions. However, generating such motion solely from audio is extremely challenging due to its one-to-many correlation between audio and motion. In the earlier stage of this field, many works [8, 19, 28, 47, 51, 86] focus on generating accurate lip movements by relying on learned audio-lip alignment losses [9, 45].", + "bbox": [ + 511, + 589, + 906, + 787 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To comprehensively extend the range of motion, some works [45, 66, 84] incorporate probabilistic generative models, such as VAE [29] and normalizing flow [53], turning the motion generation into probabilistic sampling. However, these models still lack expressiveness in generated motion due to the limited capacity of these generative models.", + "bbox": [ + 511, + 790, + 908, + 881 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent talking portrait generation methods [7, 21, 26, 36,", + "bbox": [ + 531, + 885, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*This work was done during South Korea Mandatory Military Service at DeepBrain AI Inc.", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14699", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "44, 63, 68, 71, 76, 79], powered by diffusion-based generative models [23, 61], successfully mitigate this expressiveness issue. EMO [68] introduces a promising approach to this field [7, 26, 71, 76, 79] by employing a strong pretrained image diffusion model (i.e., StableDiffusion [54]) and lifting it into video generation [24]. However, it still faces challenges in generating temporally coherent videos and achieving sampling efficiency, requiring tens of minutes for a few seconds of video. Moreover, they heavily rely on auxiliary facial prior, such as bounding boxes [68, 79], 2D landmarks and skeletons [7, 26, 82], or 3D meshes [76], which significantly restricts the diversity and the fidelity of head movements due to their strong spatial bias.", + "bbox": [ + 89, + 90, + 480, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we present FLOAT, an audio-driven talking portrait video generation model based on flow matching generative model in a motion latent space. Flow matching [35, 37] has emerged as a promising alternative to diffusion models due to its fast and high-quality sampling. By modeling talking motion within a learned motion latent space [75], we can more efficiently sample temporally consistent motion latents. This is achieved by a simple yet effective transformer-based [70] vector field predictor, inspired by DiT [48]. Since our motion latent space has orthogonal structure, our method can manipulate head motion of the generated video using its basis. Furthermore, our method supports natural emotion-aware motion enhancement driven by speech. Our contributions are summarized as follows:", + "bbox": [ + 89, + 287, + 480, + 513 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present, FLOAT, flow matching based audio-driven talking portrait generation model using a learned orthogonal motion latent space, enabling to generate talking portrait videos with reduced sampling steps.", + "- We introduce a simple yet effective transformer-based flow vector field predictor for temporally consistent motion latent sampling, which also enables the speech-driven emotional controls.", + "- Extensive experiments demonstrate that FLOAT achieves state-of-the-art performance compared to both diffusion- and non-diffusion-based methods." + ], + "bbox": [ + 91, + 516, + 480, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 699, + 238, + 714 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Diffusion Models and Flow Matching", + "text_level": 1, + "bbox": [ + 89, + 724, + 413, + 741 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion Models Diffusion models or score-based generative models [13, 23, 46, 54, 60, 61] are generative models that gradually diffuse input signals into Gaussian noise and learn the denoising reverse process for the generative modeling. They have shown remarkable results in various generation tasks, such as unconditional image and video generation [4, 16, 48], text-to-image generation [52, 54, 55], text-to-video generation [4, 20], conditional image generation [24, 82], and 3D human generation [31, 64, 67].", + "bbox": [ + 89, + 748, + 480, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Accelerating Diffusion Models While diffusion models", + "bbox": [ + 89, + 885, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "demonstrate superior performance, their iterative sampling nature still bottlenecks the efficient generation compared to VAEs [29], normalizing flow [53], and GANs [18]. To overcome this limitation, several works have been developed to boost the sampling speed of the diffusion models. StableDiffusion (SD) [54] partially mitigates this problem by moving the diffusion process from the pixel space to the spatial latent space, establishing itself as a pivotal framework among diffusion models. Another line of research has developed the sampling solvers [40, 41] based on ordinary differential equations (ODEs). Meanwhile, model distillation [22] has been introduced to transfer the knowledge of the learned diffusion models into a student model, enabling one (or a few) steps of generation [27, 34, 38, 42, 62]. However, these approaches involve substantial effort to create a well-trained diffusion model and suffer from training instability.", + "bbox": [ + 511, + 90, + 903, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Flow Matching Flow matching [35, 37] stands out as an alternative to diffusion models for its high sampling speed and competitive sample quality compared to diffusion models [10, 17, 32, 35, 50]. It belongs to the family of flow-based generative models, which estimates a transformation (referred to as a flow) between a prior distribution (e.g., Gaussian) and a target distribution. Unlike the normalizing flow [14, 53] that directly estimates the noise-to-data transformation under specific architectural constraints (e.g., affine coupling), flow matching regresses the time-dependent vector field that generates this flow by solving its corresponding ODEs [6] with flexible architectures. One specific design of flow matching is an optimal transport (OT) based one, which transforms the data distribution along the straight path with constant velocity [35].", + "bbox": [ + 511, + 334, + 903, + 560 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our audio-driven talking portrait method employs flow matching to generate the natural talking motions. Thanks to the architectural flexibility of flow matching, we use transformer-encoder architecture [70] to estimate the generating vector field, allowing us to take the video temporal consistency into account.", + "bbox": [ + 511, + 561, + 903, + 651 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Audio-driven Portrait Animation", + "text_level": 1, + "bbox": [ + 511, + 666, + 803, + 680 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Audio-driven portrait animation is the task of generating a realistic talking portrait video using a single portrait image and driving audio [45, 73, 84, 87, 88]. Since audio-to-motion relation is basically a one-to-many problem, several works utilize additional facial prior for driving conditions, e.g., 2D facial landmarks [7, 21, 26, 71, 76, 88], 3D prior [8, 43, 44, 81, 84], or emotional labels [25, 65, 80]. In earlier stages, most works [8, 19, 28, 51] focused on generating accurate lip motion from audio by utilizing the lip-sync discriminator [9]. These approaches have advanced to generating audio-related head poses in a probabilistic way. For example, StyleTalker [45] uses normalizing flow [14, 53] to generate the head motion from audio, while SadTalker [84] uses audio-conditional variational inference [29] to learn", + "bbox": [ + 511, + 689, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14700", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the 3DMM coefficients [2], bridging the intermediate representations of a pre-trained portrait animator [74].", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Meanwhile, several works [25, 65, 72, 77] focus on an emotion-aware talking portrait generation. In particular, EAMM [25] considers an emotion as the complementary displacement of facial motion, and learns these displacement from an emotion label extracted from the image.", + "bbox": [ + 89, + 121, + 482, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent audio-driven talking portrait methods powered by diffusion models show remarkable results [7, 26, 36, 44, 68, 71, 76, 79, 80]. Specifically, EMO [68] and subsequent extensions [7, 71, 76, 79] utilize the pre-trained SD [54] as their backbone to leverage generative prior trained on the large-scale image datasets. They introduce additional modules, e.g., ReferenceNet [24] and Temporal Transformer [20], to preserve input identity and enhance the video temporal consistency, respectively. However, these modules introduces additional computational cost, requiring several minutes for a few seconds of video, and still suffer from video-level artifacts, such as noisy frames, and flickering.", + "bbox": [ + 89, + 196, + 483, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "VASA-1 [80] addresses the sampling time issue by sampling motion latents [15], producing lifelike talking portraits. Our method takes advantage of this approach. However, unlike [80], our motion latent space has a strong linear orthogonal structure represented by a computable basis, enabling to manipulate the generated motion at the test-time without external driving signals. Based on this orthogonality, we employ OT-based flow matching for motion latent sampling along a straight line with reduced sampling steps.", + "bbox": [ + 89, + 378, + 483, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries: (Conditional) Flow Matching", + "text_level": 1, + "bbox": [ + 89, + 525, + 483, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $x \\in \\mathbb{R}^d$ be a data, $t \\in [0,1]$ be the time, and $q$ be a unknown target distribution. We can define a flow as a time-dependent transformation $\\varphi_t : [0,1] \\times \\mathbb{R}^d \\to \\mathbb{R}^d$ that transforms a tractable prior distribution $p_0$ to the distribution $p_1 \\approx q$ . This flow $\\varphi_t$ further introduces a probability flow path $p_t : [0,1] \\times \\mathbb{R}^d \\to \\mathbb{R}_{>0}$ and a generating vector field $v_t : [0,1] \\times \\mathbb{R}^d \\to \\mathbb{R}^d$ where $p_t$ is defined by the push-forwarding", + "bbox": [ + 88, + 550, + 483, + 672 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {t} (x) = p _ {0} \\left(\\varphi_ {t} ^ {- 1} (x)\\right) \\det \\left| \\frac {\\partial \\varphi_ {t} ^ {- 1} (x)}{\\partial x} \\right|, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 678, + 482, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and $v_{t}$ generates $\\varphi_{t}$ by means of an ordinary differential equation (ODE) [6]:", + "bbox": [ + 89, + 720, + 482, + 750 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {d}{d t} \\varphi_ {t} (x) = v _ {t} \\left(\\varphi_ {t} (x)\\right) \\quad \\text {a n d} \\quad \\varphi_ {0} (x) = x. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 757, + 482, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Flow matching [35] aims to estimate the target generating vector field $u_{t}$ with a neural network parameterized by $\\theta$ :", + "bbox": [ + 89, + 792, + 483, + 823 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {F M}} (\\theta) := \\left\\| v _ {t} (x; \\theta) - u _ {t} (x) \\right\\| _ {2} ^ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 844, + 482, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $t \\sim \\mathcal{U}[0,1]$ and $x \\sim p_t(x)$ . However, the target generating vector field $u_t$ and the sample distribution $p_t$ are", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "intractable. To address this issue, [35] proposes a method for constructing a \"conditional\" probability path $p_t(\\cdot | x_1)$ as well as target \"conditional\" vector field $u_t(\\cdot | x_1)$ using a sample $x_1 \\sim q$ as a condition. And they prove that the following objective", + "bbox": [ + 511, + 90, + 905, + 167 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C F M}} (\\theta) := \\left\\| v _ {t} (x; \\theta) - u _ {t} (x \\mid x _ {1}) \\right\\| _ {2} ^ {2}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 174, + 903, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $t\\sim \\mathcal{U}[0,1]$ and $x\\sim p_t(x|x_1)$ , is equivalent to (3) with respect to the gradient $\\nabla_{\\theta}$ .", + "bbox": [ + 511, + 200, + 903, + 229 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "One natural way of constructing $u_{t}(\\cdot |x_{1})$ is a \"straight line\" that connects $x_0\\sim p_0$ and $x_{1}\\sim q$ , drawing an optimal transport (OT) path with constant velocity [35]. Specifically, a linear time interpolation between $x_0$ and $x_{1}$ gives us the flow $x_{t} = \\varphi_{t}(x) = (1 - t)x_{0} + tx_{1}$ , the conditional probability path $p_t(x|x_1)$ defined via the affine transformation $p_t(x|x_1) = \\mathcal{N}(x|tx_1,(1 - t)^2 I)$ , and the target generating vector field $u_{t}(x|x_{1}) = x_{1} - x_{0}$ . This specific choice turns the objective (4) into", + "bbox": [ + 511, + 229, + 905, + 366 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {O T}} (\\theta) := \\left\\| v _ {t} ((1 - t) x _ {0} + t x _ {1}; \\theta) - (x _ {1} - x _ {0}) \\right\\| _ {2} ^ {2}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 375, + 903, + 392 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $t\\sim \\mathcal{U}[0,1]$ $x_0\\sim p_0$ , and $x_{1}\\sim q$ , all of which are tractable.", + "bbox": [ + 511, + 400, + 903, + 429 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Classifier-free Vector Field [10] formulates a classifier-free vector field (CFV) technique for flow matching, which enables class-conditional sampling more controllable manner without any extra classifier trained on noisy trajectory. Formally, CFV compute the modified vector field $\\tilde{v}_t$ by", + "bbox": [ + 511, + 430, + 905, + 506 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {v} _ {t} \\left(x _ {t}, c; \\theta\\right) \\approx \\gamma v _ {t} \\left(x _ {t}, c; \\theta\\right) + (1 - \\gamma) v _ {t} \\left(x _ {t}, c = \\emptyset ; \\theta\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 513, + 903, + 531 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\gamma$ denotes the guidance scale. $v_{t}(x_{t},c = \\emptyset ;\\theta)$ is the predicted vector field without a driving condition $c$ . For more details, please refer to [10, 35].", + "bbox": [ + 511, + 539, + 905, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method: Flow Matching for Audio-driven Talking Portrait", + "text_level": 1, + "bbox": [ + 511, + 595, + 905, + 632 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We provide an overview of FLOAT in Fig. 2. Given source image $S \\in \\mathbb{R}^{3 \\times H \\times W}$ , and a driving audio signal $a^{1:L} \\in \\mathbb{R}^{L \\times d_a}$ of length $L$ , our method generates a video", + "bbox": [ + 511, + 640, + 905, + 685 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {D} ^ {1: L} = \\left(\\hat {D} ^ {l}\\right) _ {l = 1} ^ {L} \\in \\mathbb {R} ^ {L \\times 3 \\times H \\times W} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 691, + 903, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of $L$ frames, featuring audio-synchronized talking head motions, including both verbal and non-verbal motions. Our method consists of two phases. First, we pre-train a motion auto-encoder, which provides us with the expressive and smooth motion latent space for the talking portraits (Sec. 4.1). Next, we employ OT-based flow matching [35] to generate a sequence of motion latents with a transformer-based vector field predictor using the driving audio, which is decoded to the talking portrait videos (Sec. 4.2). We also incorporate speech-driven emotions as the driving conditions, achieving automatic emotion-aware talking portrait generation without any extra user input for emotion.", + "bbox": [ + 511, + 719, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14701", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a0dbc4d307dc1ce96a87d77c15d61dadc1ceb8953c76d8b4364fe5b1d5c18c47.jpg", + "image_caption": [ + "Figure 2. Overview of FLOAT. We encode the source image $S \\in \\mathbb{R}^{3 \\times H \\times W}$ into the latent with the explicit identity-motion decomposition $w_{s} = w_{s \\to r} + w_{r \\to s} \\in \\mathbb{R}^{d}$ . Given audio segments $a^{-L':L} \\in \\mathbb{R}^{(L' + L) \\times d_{a}}$ of the length $L' + L$ and the reference motion $w_{r \\to s} \\in \\mathbb{R}^{d}$ , and the speech-driven emotion label $w_{e} \\in \\mathbb{R}^{7}$ , a flow matching transformer estimates the generating vector field $v_{t}(\\varphi_{t}(x_{0}), \\mathbf{c}_{t}; \\theta) \\in \\mathbb{R}^{L \\times d}$ from noisy motion latents, which is used to solve corresponding ODE and generates the motion latents $w_{r \\to \\hat{D}^{1:L}}$ . Finally, the sequence of latents $w_{S \\to \\hat{D}^{1:L}} := (w_{S \\to r} + w_{r \\to \\hat{D}^{l}})_{l=1}^{L}$ are decoded into the video $\\hat{D}^{1:L} \\in \\mathbb{R}^{L \\times 3 \\times H \\times W}$ ." + ], + "image_footnote": [], + "bbox": [ + 101, + 89, + 893, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Motion Latent Auto-encoder", + "text_level": 1, + "bbox": [ + 89, + 342, + 346, + 356 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recent talking portrait methods utilize the VAE of StableDiffusion (SD) [54] due to its rich semantic pixel-based latent space. However, they often struggle to generate temporally consistent frames when lifted to video generating tasks [7, 24, 68, 79, 89]. Thus, our first goal for realistic talking portrait is to obtain good motion latent space, capturing both global (e.g., head motion) and fine-grained local (e.g., facial expressions, mouth and pupil movement) dynamics.", + "bbox": [ + 89, + 363, + 482, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instead of VAE of SD, we pre-train a motion latent autoencoder that encodes images into motion latents. This is done by training an auto-encoder to reconstruct a driving image from a source image sampled from the same video clip, enforcing the encoder to implicitly capture both temporally adjacent and distant motions. Inspired by LIA [75], we introduce a learned orthonormal basis that decomposes the motion along distinct orthogonal directions. Specifically, our motion auto-encoder encodes the source $S$ into the latent $w_{S} \\in \\mathbb{R}^{d}$ with following explicit decomposition:", + "bbox": [ + 89, + 484, + 483, + 636 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw _ {S} := w _ {S \\rightarrow r} + w _ {r \\rightarrow S}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 648, + 480, + 661 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $w_{S\\to r}\\in \\mathbb{R}^d$ is the identity latent and", + "bbox": [ + 89, + 671, + 380, + 686 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw _ {r \\rightarrow S} = \\sum_ {m = 1} ^ {M} \\lambda_ {m} (S) \\cdot \\mathbf {v} _ {m} \\in \\mathbb {R} ^ {d} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 178, + 696, + 480, + 736 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "is the motion latent with $\\lambda(S) := (\\lambda_m(S))_{m=1}^M \\in \\mathbb{R}^M$ being the source-dependent motion coefficients that span the learned source-agnostic motion basis $V := \\{\\mathbf{v}_m\\}_{m=1}^M \\subseteq \\mathbb{R}^d$ . In this space, $\\lambda_m(S)$ is the intensity of the motion direction $\\mathbf{v}_m$ .", + "bbox": [ + 89, + 747, + 482, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 6, our method enables motion editing of the sampled (generated) motion using only the basis $V$ and its orthogonality, as stated in Eq. (15).", + "bbox": [ + 89, + 825, + 482, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Improving Fidelity of Facial Components: $\\mathcal{L}_{\\mathrm{comp - lp}}$ The expressiveness of generated motions and the image fidelity", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5333d2792990caafb26491a4c5f2820a0f715920319784d7894970a3369fd475.jpg", + "image_caption": [ + "Figure 3. Efficacy of $\\mathcal{L}_{\\mathrm{comp - lp}}$ for fine-grained motion and fidelity." + ], + "image_footnote": [], + "bbox": [ + 517, + 339, + 898, + 467 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "are determined by the motion space and the motion autoencoder. However, as resolution increases, fine details in small facial regions (e.g., teeth, eyeballs) often get buried in large-scale dynamics. To address this issue, we propose a facial component perceptual loss $\\mathcal{L}_{\\mathrm{comp - lp}}$ using [59, 83] that significantly improves the image fidelity (e.g., teeth and eyes) as well as fine-grained motions (e.g., eyeball and eyebrows movements). As shown in Fig. 3, $\\mathcal{L}_{\\mathrm{comp - lp}}$ allows us to generate high-fidelity facial components and their fine-grained motions without relying on pre-trained foundation models, such as StableDiffusion [54].", + "bbox": [ + 511, + 505, + 906, + 671 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Flow Matching in Motion Latent Space", + "text_level": 1, + "bbox": [ + 511, + 681, + 849, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Armed with this linear orthogonal space, we employ OT-based flow matching [35, 37] for the motion sampling. Specifically, we predict a vector field $v_{t}(x_{t},\\mathbf{c}_{t};\\theta)\\in \\mathbb{R}^{L\\times d}$ where $x_{t}$ is the sample at flow time $t\\in [0,1]$ , and $\\mathbf{c}_t\\in \\mathbb{R}^{L\\times h}$ represents the driving conditions for $L$ consequent frames. This vector field generates the flow $\\varphi_t:[0,1]\\times \\mathbb{R}^{L\\times d}\\to \\mathbb{R}^{L\\times d}$ of $L$ frames by solving ODE (Eq. (2)). As illustrated in Fig. 4, we build our vector field predictor upon the transformer encoder [70] architecture. Specifically, we adopt DiT [48] architecture, but decouple frame-wise conditioning from time-axis attention mechanism, which enables us to model temporally consistent motion latents.", + "bbox": [ + 511, + 704, + 906, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In DiT [48], distinct semantic tokens are modulated by a", + "bbox": [ + 531, + 885, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14702", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/13dc8fc6b5b49496a5c35bf6618e63a51d637fbb5f4fc548a2fd89effb668ceb.jpg", + "image_caption": [ + "Figure 4. Frame-wise vector field predictor block at inference." + ], + "image_footnote": [], + "bbox": [ + 106, + 89, + 464, + 215 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "single diffusion time step embedding and class embedding through adaptive layer normalization (AdaLN). In contrast, our vector field predictor modulates each $l$ -th input latent with its corresponding $l$ -th condition and then combines their temporal relations through a masked self-attention layer that attends to $2 \\cdot T$ neighboring frames. Formally, for each $l$ -th frame, frame-wise AdaLN and frame-wise gating are computed by", + "bbox": [ + 89, + 244, + 483, + 367 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\gamma_ {i} ^ {l} \\times \\operatorname {L N} \\left(X _ {t} ^ {l}\\right) + \\beta_ {i} ^ {l} \\in \\mathbb {R} ^ {h} \\quad \\text {a n d} \\quad \\alpha_ {i} ^ {l} \\times X _ {t} ^ {l} \\in \\mathbb {R} ^ {h}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 375, + 480, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "respectively, where $i \\in \\{1,2\\}$ , $h$ is the hidden dimension, $\\mathrm{LN}(\\cdot)$ denotes layer norm [33], and $X_{t}^{l}$ is the $l$ -th input for each operation at flow time $t \\in [0,1]$ . The coefficients $\\alpha_{i}^{l}, \\beta_{i}^{l}, \\gamma_{i}^{l} \\in \\mathbb{R}^{h}$ are computed from the condition $\\mathbf{c}_{t}^{l} \\in \\mathbb{R}^{h}$ through a linear layer, ToScaleShift, as depicted in Fig. 4.", + "bbox": [ + 89, + 402, + 483, + 477 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Speech-driven Emotion Enhancement How can we make talking motions more expressive and natural? During talking, humans naturally reflect their emotions through their voices, and these emotions influence talking motions. For instance, a person who speaks sadly may be more likely to shake the head and avoid eye contact. This non-verbal motion derived from emotions crucially impacts the naturalness of a talking portrait.", + "bbox": [ + 89, + 478, + 483, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Existing works [25, 72, 80] use image-emotion paired data or image-driven emotion predictor [56] to generate the emotion-aware motion. In contrast, we incorporate speech-driven emotions, a more intuitive way of controlling emotion for audio-driven talking portrait. Specifically, we utilize a pre-trained speech emotion predictor [49] that produces softmax probabilities of seven distinct emotions: angry, disgust, fear, happy, neutral, sad, and surprise, which we then input into the vector field predictor.", + "bbox": [ + 89, + 598, + 482, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, as people do not always speak with a single, clear emotion, determining emotions solely from audio is often ambiguous [25]. Naive introduction of speech-driven emotion can make emotion-aware motion generation more challenging. To address this issue, we inject the emotions together with other driving conditions at training phase and modify them at inference phase.", + "bbox": [ + 89, + 734, + 482, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Driving Conditions We concatenate the audio representation $a^{1:L} \\in \\mathbb{R}^{L \\times d_a}$ of a pre-trained Wav2Vec2.0 [1], the speech emotion label $w_e \\in \\mathbb{R}^7$ , and the source motion latent $w_{r \\to S} \\in \\mathbb{R}^d$ . Next, we add the flow time step embedding", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\operatorname{Emb}(t) \\in \\mathbb{R}^h$ to these conditions, producing $\\mathbf{c}_t \\in \\mathbb{R}^{L \\times h}$ via a linear layer, ToCondition, as depicted in Fig. 2, where $\\operatorname{Emb}(t)$ is computed using the sinusoidal position embedding [70].", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training We train FLOAT by reconstructing a target vector field computed from driving frames using the corresponding audio segments and a source motion latent. We choose a pair of driving motions and corresponding audio $(w_{r\\rightarrow D^{1:L}},a^{1:L})$ , and construct the target vector field $u_{t}(x|w_{r\\rightarrow D^{1:L}}) = w_{r\\rightarrow D^{1:L}} - x_{0}\\in \\mathbb{R}^{L\\times d}$ with noisy input $\\varphi_t(x_0) = (1 - t)x_0 + tw_{r\\to D^{1:L}}(t\\sim \\mathcal{U}[0,1]$ and $x_0\\sim$ $\\mathcal{N}(0^{1:L},I)$", + "bbox": [ + 511, + 151, + 905, + 271 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For smooth transitions of sequences longer than the window length $L$ , we incorporate last $L'$ audio features and motion latents $w_{r\\rightarrow D^{-L':0}}$ from the preceding window as additional input.", + "bbox": [ + 511, + 272, + 905, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The flow matching objective $\\mathcal{L}_{\\mathrm{OT}}(\\theta)$ is defined by", + "bbox": [ + 532, + 332, + 862, + 348 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {O T}} (\\theta) = \\left\\| v _ {t} ^ {1: L} \\left(x _ {t}, \\mathbf {c} _ {t}; \\theta\\right) - u _ {t} \\left(x \\mid w _ {r \\rightarrow D ^ {1: L}}\\right)\\right\\|, \\tag {11} \\\\ + \\left\\| v _ {t} ^ {- L ^ {\\prime}: 0} \\left(x _ {t}, \\mathbf {c} _ {t}; \\theta\\right) - w _ {r \\rightarrow D ^ {- L ^ {\\prime}: 0}} \\right\\|, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 357, + 903, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $x_{t} \\coloneqq [w_{r \\to D^{-L' \\cdot 0}} | \\varphi_{t}(x_{0})] \\in \\mathbb{R}^{(-L' + L) \\times d}$ is the concatenated input, $\\mathbf{c}_{t} \\in \\mathbb{R}^{(-L' + L) \\times h}$ is the driving condition consisting of $[t, w_{r \\to S}, w_{e}, a^{1:L}, a^{-L':0}]$ . Note that $w_{e}$ and $w_{r \\to S}$ are shared across the $L' + L$ frames. We incorporate a velocity loss [67] to supervise temporal consistency:", + "bbox": [ + 511, + 409, + 905, + 489 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {v e l}} (\\theta) = \\left\\| \\Delta v _ {t} - \\Delta u _ {t} \\right\\|, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 500, + 903, + 516 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Delta v_{t}$ and $\\Delta u_{t}$ are the one-frame difference along the time-axis for the prediction $v_{t} \\in \\mathbb{R}^{(-L' + L) \\times d}$ and the target $[w_{r \\to D^{-L':0}} | u_{t}] \\in \\mathbb{R}^{(-L' + L) \\times d}$ , respectively.", + "bbox": [ + 511, + 526, + 905, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The total objective $\\mathcal{L}_{\\mathrm{total}}(\\theta)$ is", + "bbox": [ + 532, + 571, + 732, + 587 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} (\\theta) = \\lambda_ {\\mathrm {O T}} \\mathcal {L} _ {\\mathrm {O T}} (\\theta) + \\lambda_ {\\mathrm {v e l}} \\mathcal {L} _ {\\mathrm {v e l}} (\\theta), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 598, + 903, + 614 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{\\mathrm{OT}}$ and $\\lambda_{\\mathrm{vel}}$ are the balancing coefficients. During training, we apply dropout to $w_{r}$ , $w_{e}$ , and $a^{1:L}$ with a probability of 0.1 for CFV. Additionally, we apply dropout to the preceding audio and motion latents with a probability 0.5 for smooth transition in the initial window.", + "bbox": [ + 511, + 625, + 905, + 699 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inference During inference, we sample the generating vector field from noise $x_0$ , using the driving conditions $w_{r\\rightarrow S}$ , $w_e$ , and $a^{1:L}$ , as well as the $L'$ frames of preceding audio and generated motion latents.", + "bbox": [ + 511, + 700, + 905, + 760 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We extend the CFV [10] to an incremental CFV to separately adjust the audio and emotion, inspired by [3]:", + "bbox": [ + 511, + 761, + 905, + 792 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\tilde {v} _ {t} \\approx v _ {t} (x _ {0}, \\mathbf {c} _ {t} | _ {\\{a ^ {1: L}, w _ {e} \\}}) \\\\ + \\gamma_ {a} \\left[ v _ {t} \\left(x _ {0}, \\mathbf {c} _ {t} \\mid_ {w _ {e}}\\right) - v _ {t} \\left(x _ {0}, \\mathbf {c} _ {t} \\mid_ {\\{a ^ {1: L}, w _ {e} \\}} \\right] \\right. \\\\ + \\gamma_ {e} \\left[ v _ {t} \\left(x _ {0}, \\mathbf {c} _ {t}\\right) - v _ {t} \\left(x _ {0}, \\mathbf {c} _ {t} \\mid_ {w _ {e}}\\right) \\right], \\tag {14} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 801, + 903, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\gamma_{a}$ and $\\gamma_{e}$ are the guidance scales for audio and emotion, respectively. $\\mathbf{c}_t|_{\\{x,y\\}}$ denotes the driving condition", + "bbox": [ + 511, + 869, + 905, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14703", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/e5afec730d66757140b8157b78f57d5906c6a01e55b2d65608480adec298dc2b.jpg", + "table_caption": [ + "Table 1. Quantitative comparison results with state-of-the-art methods on HDTF [85] / RAVDESS [39]. The best result for each metric is in bold, and the second-best result is underlined." + ], + "table_footnote": [], + "table_body": "
MethodImage & Video GenerationLip Synchronization
FID ↓FVD ↓CSIM ↑E-FID ↓P-FID ↓LSE-D ↓LSE-C ↑
SadTalker† [84]71.952 / 119.430339.058 / 376.2940.644 / 0.6441.914 / 3.5001.456 / 2.0457.947 / 7.2737.305 / 4.748
EDTalk† [66]50.078 / 75.020211.284 / 304.9330.626 / 0.6761.579 / 3.4680.054 / 0.0908.123 / 7.6827.623 / 5.318
AniTalker† [36]39.512 / 70.430184.454 / 265.3410.643 / 0.7251.830 / 2.3300.092 / 0.1267.907 / 8.1767.288 / 4.555
Halo [79]25.363 / 57.648197.196 / 375.5570.869 / 0.8601.039 / 2.4920.037 / 0.0507.792 / 7.6137.582 / 4.795
EchoMimic [7]33.552 / 81.839296.757 / 320.2200.823 / 0.8051.234 / 3.2010.023 / 0.0478.903 / 8.1616.242 / 4.144
FLOAT (Ours)21.100 / 31.681162.052 / 166.3590.843 / 0.8101.229 / 1.3670.032 / 0.0317.290 / 6.9948.222 / 5.730
", + "bbox": [ + 99, + 121, + 898, + 237 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/daff56d026ccb964d277b4599dfc4f045c6db093de7e7c345df483f887a0ba5e.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison results with state-of-the-art methods on HDTF [85] / RAVDESS [39]. Please refer to supplementary videos. Note that we additionally provide a video comparison with EMO [68] and VASA-1 [80] using their video demonstration." + ], + "image_footnote": [], + "bbox": [ + 102, + 250, + 459, + 654 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3255bf62b78bce64af2a45ed15e8d0a73102a854ff37627a755a8f9541ebd61a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 462, + 369, + 550, + 449 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1c5bddc5d113e1988b2eff754be5d5038bf803625fe43d6a11361efc5373dec3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 470, + 450, + 535, + 518 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f4481d782c00a7ecbdb51e51ffd279fa5a42c4221caa8dc70e74f38ff5c8d111.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 247, + 883, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "without the condition $x$ and $y$ . We set $\\gamma_{a} = 2$ and $\\gamma_{e} = 1$ based on the ablation studies on $\\gamma_{a}$ and $\\gamma_{e}$ provided in supplementary materials.", + "bbox": [ + 89, + 710, + 482, + 755 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After sampling, ODE solver receives the estimated vector field to compute the motion latents through numerical integration. We empirically find that FLOAT can generate reasonable motion with around 10 number of function evaluations (NFE). Please refer to supplementary videos.", + "bbox": [ + 89, + 767, + 482, + 843 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Lastly, we add the source identity latent to the generated motion latents and decode them into video frames using the motion latent decoder.", + "bbox": [ + 89, + 854, + 483, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 513, + 709, + 645, + 726 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Dataset and Pre-processing", + "text_level": 1, + "bbox": [ + 511, + 738, + 759, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For training the motion latent auto-encoder, we use three open-source datasets: HDTF [85], RAVDESS [39], and VFHQ [78]. When training FLOAT, we exclude VFHQ because it does not support the synchronized audio. HDTF [85] is for high-definition talking face generation, containing videos of over 300 unique identities. RAVDESS [39] includes more than 2,400 emotion-intensive videos of 24 different identities. VFHQ [78] is designed for high-resolution video super-resolution and includes a large num", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14704", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ber of unique identities, which compensates the limited number of identities of the preceding datasets. Following the strategy of [58], we first convert each video to 25 FPS and resample the audio into $16\\mathrm{kHz}$ . Then, we crop and resize the facial region to $512^2$ resolution [5]. After the pre-processing, for HDTF, we use a total of 11.3 hours of 240 videos featuring 230 different identities for training, and videos of 78 different identities, each 15 seconds long, for test. For RAVDESS, we use videos of 22 identities for training, and videos of the remaining 2 identities for test, with each 3-4 seconds long and representing 14 emotional intensities. Note that the identities in the training and test are disjoint in both datasets.", + "bbox": [ + 89, + 90, + 483, + 287 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 301, + 307, + 318 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The motion latent dimension is set to $d = 512$ with $M = 20$ distinct orthogonal directions. For the vector predictor, we use 8 attention heads, a hidden dimension $h = 1024$ , and an attention window length $T = 2$ . Considering the length of the training video clips, we set $L = 50$ frames with preceding $L' = 10$ frames at once, encompassing 2.4 seconds of video. We employ the Adam optimizer [30] with a batch size of 8 and a learning rate of $10^{-5}$ . We use $L1$ distance for the norm $\\| \\cdot \\|$ in the training objective. We set the balancing coefficients to $\\lambda_{\\mathrm{OT}} = \\lambda_{\\mathrm{vel}} = 1$ . The entire training takes about 2 days for 2,000k steps on a single NVIDIA A100 GPU. We use Euler method [35] for the ODE solver.", + "bbox": [ + 89, + 325, + 483, + 507 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Evaluation", + "text_level": 1, + "bbox": [ + 89, + 523, + 210, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metrics and Baselines For evaluating the image and video generation quality, we measure Fréchet Inception Distance (FID) [57] and 16 frames Fréchet Video Distance (FVD) [69]. For facial identity, expression and head motion, we measure Cosine Similarity of identity embedding (CSIM) [11], Expression FID (E-FID) [68] and Pose FID (P-FID), respectively. Lastly, we measure Lip-Sync Error Distance and Confidence (LSE-D and LSE-C [51]) for audio-visual alignment.", + "bbox": [ + 89, + 547, + 483, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our method with state-of-the-art audio-driven talking portrait methods whose official implementations are publicly available. For non-diffusion methods, we compare with SadTalker [84] and EDTalk [66]. For diffusion methods, we compare with AniTalker [36], Hallo [79], and EchoMimic [7].", + "bbox": [ + 89, + 685, + 483, + 775 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison Results In Tab. 1 and Fig. 5, we show the quantitative and qualitative comparison results, respectively. FLOAT outperforms other methods on most of the metrics and visual quality in both datasets.", + "bbox": [ + 89, + 777, + 483, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Additionally, we provide video comparison results with EMO [68] and VASA-1 [80] in the supplementary materials, using their demonstration videos due to the infeasibility of direct implementation.", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eaef4948f10feb64b30b606b91b0571f4844c966563af3497ed62e22a9a64d1f.jpg", + "image_caption": [ + "Figure 6. Test-time pose editing using $\\lambda$ -control $(\\lambda_{15}(\\hat{D}) \\pm 10)$ ." + ], + "image_footnote": [], + "bbox": [ + 522, + 89, + 895, + 260 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Applications", + "text_level": 1, + "bbox": [ + 513, + 290, + 645, + 306 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Test-time Pose Editing via Orthonormal Basis $V$ Since FLOAT learns the underlying motion latent structure, it is natural to assume that for any sampled motion latent $w_{r\\rightarrow \\hat{D}}$ , there exist motion coefficients $\\{\\lambda_m(\\hat{D})\\}_{m = 1}^M$ satisfying the representation in Eq. (9): $w_{r\\rightarrow \\hat{D}} = \\sum_{m = 1}^{M}\\lambda_{m}(\\hat{D})\\cdot \\mathbf{v}_{m}$ .", + "bbox": [ + 511, + 311, + 903, + 391 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We can always compute these coefficients in closed form by taking inner products between the sampled motion $w_{r\\rightarrow \\hat{D}}$ and the learned orthonormal basis $V$ :", + "bbox": [ + 511, + 392, + 903, + 438 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\langle w _ {r \\rightarrow \\hat {D}}, \\mathbf {v} _ {k} \\right\\rangle = \\left\\langle \\sum_ {m = 1} ^ {M} \\lambda_ {m} (\\hat {D}) \\cdot \\mathbf {v} _ {m}, \\mathbf {v} _ {k} \\right\\rangle = \\lambda_ {k} (\\hat {D}), \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 445, + 903, + 486 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\langle \\mathbf{v}_m, \\mathbf{v}_k \\rangle = \\delta_{m,k}$ and $\\delta$ is Kronecker delta. At this point, we can edit the sampled motions by editing the corresponding coefficients (e.g., via linear operation) and combining them back into the motion latent. As shown in Fig. 6, it allows us to control head direction without interfering with other motions due to the orthogonality of the basis. We refer to this test-time editing technique as $\\lambda$ -control.", + "bbox": [ + 511, + 492, + 905, + 598 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Additional Driving Signals In Fig. 7 and Tab. 2, we experiment with additional driving conditions, head poses and image-driven emotion labels, to explore additional controllability in our method. We employ 3DMM head pose parameters $p \\in \\mathbb{R}^6$ [2] extracted by [12]. We concatenate a sequence of pose parameters $p^{1:L} \\in \\mathbb{R}^{L \\times 6}$ with the other driving conditions, and then map them to $c_t^{1:L} \\in \\mathbb{R}^{L \\times h}$ . We also experiment on image-driven emotion [56] for frame-wise emotion control rather than the long-term emotion enhancement. FLOAT can effectively accommodate these additional conditions, highlighting its flexibility across diverse control signals.", + "bbox": [ + 511, + 598, + 905, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Redirecting Speech-driven Emotion Since FLOAT learns diverse emotions in the emotion-intensive data distribution [39], the generated emotion-aware motion can be modified by redirecting the speech-driven emotion label toward a different emotion at inference time. As illustrated in Fig. 8, this technique is particularly beneficial for manual redirection when the emotion predicted from speech is complex or ambiguous.", + "bbox": [ + 511, + 779, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14705", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6ca3a17d0829e37887e40238950c7f077f253cdc06d11a9d01aa8966bbe9dd95.jpg", + "image_caption": [ + "Figure 7. Additional conditioning results of FLOAT. 3Dpose, $S2E$ , and $I2E$ denote 3D head pose parameters [12], speech-to-emotion [49], and image-to-emotion [56], respectively." + ], + "image_footnote": [], + "bbox": [ + 114, + 89, + 486, + 261 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/af02109081a547b53339d991fc0888c4352bbee856a7594687667110672402fa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 490, + 90, + 879, + 261 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/28f033061734716b9434f2325442c02abf8bcef3a6bb0e062ee3907bd5dd31d3.jpg", + "table_caption": [ + "Table 2. Quantitative results of FLOAT with additional conditions (HDTF [85] / RAVDESS [39]). $S2E$ , $I2E$ , and 3DPose denote speech-to-emotion [49], image-to-emotion [56], and 3DMM pose parameters [12], respectively." + ], + "table_footnote": [], + "table_body": "
ConfigurationsFID ↓FVD ↓E-FID ↓P-FID ↓LSE-D ↓
AFLOAT(Ours)21.100 / 31.681162.052 / 166.3591.229 / 1.3670.032 / 0.0317.290 / 6.994
BA + 3Dpose19.721 / 29.721126.663 / 112.8940.926 / 1.1520.012 / 0.0167.516 / 7.047
CA - S2E21.235 / 32.035155.032 / 166.8661.254 / 1.5020.031 / 0.0257.264 / 7.222
DA - S2E + I2E21/528 / 31.609158.577 / 162.3691.158 / 1.3050.034 / 0.0227.183 / 7.150
", + "bbox": [ + 93, + 364, + 478, + 411 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7e0ea533f52484b01bbefc6180adff4b2c616f773c8e0a9eb1c83bcb066f4fa5.jpg", + "image_caption": [ + "Figure 8. Redirecting the unclear emotion prediction to a desirable one-hot encoding, which can be further intensified by the CFV." + ], + "image_footnote": [], + "bbox": [ + 93, + 417, + 478, + 511 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 559, + 254, + 574 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation on Frame-wise AdaLN We compare frame-wise AdaLN (and gating) followed by masked self-attention to separate conditioning from attending, with a cross-attention that performs conditioning and attending simultaneously. As shown in Tab. 3, both approaches achieve competitive image and video quality, while frame-wise AdaLN provides better expression generation and lip synchronization. We observe that frame-wise AdaLN can achieve more diverse head motions than the cross-attention. Please refer to supplementary videos.", + "bbox": [ + 89, + 582, + 482, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation on Flow Matching We compare flow matching with two types of diffusion models: $\\epsilon$ -prediction (noise) and $x_0$ -prediction (signal) [52, 67]. In both cases, we adopt our vector predictor architecture as denoising networks. We adopt diffusion training settings of VASA-1 [80] (500 diffusion steps with a cosine noise scheduler [46] and 50 DDIM denoising steps) for the indirect comparison with [80]. Notably, diffusion and flow matching achieve competitive results on image quality while the latter achieves the better lip synchronization. In Fig. 9, we compare the forward pass efficiency by measuring frames per second (FPS) of each", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8d462d551dd1b2925762d737f5b543dda08baae00af25d2a0f9735f6e95d2541.jpg", + "table_caption": [ + "Table 3. Ablation studies of FLOAT on HDTF [85]. The best result for each metric is in bold, and the second-best result is underlined." + ], + "table_footnote": [], + "table_body": "
MethodFID ↓FVD ↓E-FID ↓LSE-D ↓#NFEs ↓
Ours (w. Cross-Attn.)21.873162.7021.4527.75710
Ours (w. Diff., c-pred.)21.190161.6661.2139.92250
Ours (w. Diff., x0-pred.)21.697162.8471.2789.04850
FLOAT (Ours)21.100162.0521.2297.29010
", + "bbox": [ + 524, + 340, + 890, + 398 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/41c53dddeebfba3786224a8801316f9ace4e5d7e82874f9a46e3782adc4c2a59.jpg", + "image_caption": [ + "Figure 9. Comparison of the forward pass efficiency. We compute FPS on a single NVIDIA V100 GPU." + ], + "image_footnote": [], + "bbox": [ + 527, + 402, + 890, + 488 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "model. Thanks to the compact motion latent representation and OT-based flow matching, FLOAT achieves the highest FPS, superior lip-sync performance, dynamic head motion, and the lowest NFEs.", + "bbox": [ + 511, + 537, + 906, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 625, + 633, + 640 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We proposed FLOAT, a flow matching based audio-driven talking portrait generation model leveraging a learned motion latent space. We introduced a transformer-based vector field predictor, enabling temporally consistent motion generation. Additionally, we incorporated speech-driven emotion labels into the motion sampling process to improve the naturalness of the audio-driven talking motions. FLOAT addresses current core limitations of diffusion-based talking portrait video generation methods by reducing the sampling time through flow matching while achieving the remarkable sample quality. Extensive experiments verified that FLOAT achieves state-of-the-art performance in terms of visual quality, motion fidelity, and efficiency.", + "bbox": [ + 511, + 654, + 906, + 851 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Discussion We leave further discussion considering limitations, future work, and ethical considerations in the supplementary materials.", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14706", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449-12460, 2020. 5", + "[2] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques, pages 187-194, 1999. 3, 7", + "[3] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18392-18402, 2023. 5", + "[4] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators, 2024. 2", + "[5] Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2d & 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In International Conference on Computer Vision, 2017. 7", + "[6] Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations. Advances in neural information processing systems, 31, 2018. 2, 3", + "[7] Zhiyuan Chen, Jiajiong Cao, Zhiquan Chen, Yuming Li, and Chenguang Ma. Echomimic: Lifelike audio-driven portrait animations through editable landmark conditions. arXiv preprint arXiv:2407.08136, 2024. 1, 2, 3, 4, 6, 7", + "[8] Kun Cheng, Xiaodong Cun, Yong Zhang, Menghan Xia, Fei Yin, Mingrui Zhu, Xuan Wang, Jue Wang, and Nannan Wang. Videoretalking: Audio-based lip synchronization for talking head video editing in the wild. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 1, 2", + "[9] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In Asian Conference on Computer Vision, pages 251-263, 2016. 1, 2", + "[10] Quan Dao, Hao Phung, Binh Nguyen, and Anh Tran. Flow matching in latent space. arXiv preprint arXiv:2307.08698, 2023. 2, 3, 5", + "[11] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4690-4699, 2019. 7", + "[12] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 7, 8", + "[13] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2", + "[14] Laurent Dinh, Jascha Sohl-Dickstein, and Samy Ben" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "gio. Density estimation using real nvp. arXiv preprint arXiv:1605.08803, 2016. 2", + "[15] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2663-2671, 2022. 3", + "[16] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 2", + "[17] Johannes S Fischer, Ming Gui, Pingchuan Ma, Nick Stracke, Stefan A Baumann, and Björn Ommer. Boosting latent diffusion with flow matching. arXiv preprint arXiv:2312.07360, 2023. 2", + "[18] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2", + "[19] Jiazhi Guan, Zhanwang Zhang, Hang Zhou, Tianshu Hu, Kaisiyuan Wang, Dongliang He, Haocheng Feng, Jingtuo Liu, Errui Ding, Ziwei Liu, et al. Stylesync: High-fidelity generalized and personalized lip sync in style-based generator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1505-1515, 2023. 1, 2", + "[20] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 2, 3", + "[21] Tianyu He, Junliang Guo, Runyi Yu, Yuchi Wang, Jialiang Zhu, Kaikai An, Leyi Li, Xu Tan, Chunyu Wang, Han Hu, et al. Gaia: Zero-shot talking avatar generation. arXiv preprint arXiv:2311.15230, 2023. 1, 2", + "[22] Geoffrey Hinton. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2", + "[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2", + "[24] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8153-8163, 2024. 2, 3, 4", + "[25] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Qianyi Wu, Wayne Wu, Feng Xu, and Xun Cao. Eamm: One-shot emotional talking face via audio-based emotion-aware motion model. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2, 3, 5", + "[26] Jianwen Jiang, Chao Liang, Jiaqi Yang, Gaojie Lin, Tianyun Zhong, and Yanbo Zheng. Loopy: Taming audio-driven portrait avatar with long-term motion dependency. arXiv preprint arXiv:2409.02634, 2024. 1, 2, 3", + "[27] Minguk Kang, Richard Zhang, Connelly Barnes, Sylvain Paris, Suha Kwak, Jaesik Park, Eli Shechtman, Jun-Yan Zhu," + ], + "bbox": [ + 517, + 93, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14707", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Taesung Park. Distilling diffusion models into conditional gans. arXiv preprint arXiv:2405.05967, 2024. 2", + "[28] Taekyung Ki and Dongchan Min. Stylelipsync: Style-based personalized lip-sync video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22841-22850, 2023. 1, 2", + "[29] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 1, 2", + "[30] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 7", + "[31] Tobias Kirschstein, Simon Giebenhain, and Matthias Nießner. Diffusionavatars: Deferred diffusion for high-fidelity 3d head avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5492, 2024. 2", + "[32] Matthew Le, Apoorv Vyas, Bowen Shi, Brian Karrer, Leda Sari, Rashel Moritz, Mary Williamson, Vimal Manohar, Yossi Adi, Jay Mahadeokar, et al. Voicebox: Text-guided multilingual universal speech generation at scale. Advances in neural information processing systems, 36, 2024. 2", + "[33] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. ArXiv e-prints, pages arXiv-1607, 2016. 5", + "[34] Senmao Li, Taihang Hu, Fahad Shahbaz Khan, Linxuan Li, Shiqi Yang, Yaxing Wang, Ming-Ming Cheng, and Jian Yang. Faster diffusion: Rethinking the role of unet encoder in diffusion models. arXiv preprint arXiv:2312.09608, 2023. 2", + "[35] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022. 2, 3, 4, 7", + "[36] Tao Liu, Feilong Chen, Shuai Fan, Chenpeng Du, Qi Chen, Xie Chen, and Kai Yu. Anitalker: Animate vivid and diverse talking faces through identity-decoupled facial motion encoding. arXiv preprint arXiv:2405.03121, 2024. 1, 3, 6, 7", + "[37] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 2, 4", + "[38] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, et al. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. In The Twelfth International Conference on Learning Representations, 2023. 2", + "[39] Steven R Livingstone and Frank A Russo. The ryerson audio-visual database of emotional speech and song (ravdess): A dynamic, multimodal set of facial and vocal expressions in north american english. PloS one, 13(5): e0196391, 2018. 6, 7, 8", + "[40] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 2", + "[41] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 2" + ], + "bbox": [ + 91, + 90, + 483, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 2", + "[43] Yifeng Ma, Suzhen Wang, Zhipeng Hu, Changjie Fan, Tangjie Lv, Yu Ding, Zhidong Deng, and Xin Yu. Styletalk: One-shot talking head generation with controllable speaking styles. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1896-1904, 2023. 2", + "[44] Yifeng Ma, Shiwei Zhang, Jiayu Wang, Xiang Wang, Yingya Zhang, and Zhidong Deng. Dreamtalk: When expressive talking head generation meets diffusion probabilistic models. arXiv preprint arXiv:2312.09767, 2023. 2, 3", + "[45] Dongchan Min, Minyoung Song, Eunji Ko, and Sung Ju Hwang. Stylistalker: One-shot style-based audio-driven talking head video generation. arXiv preprint arXiv:2208.10922, 2022. 1, 2", + "[46] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162-8171. PMLR, 2021. 2, 8", + "[47] Se Jin Park, Minsu Kim, Joanna Hong, Jeongsoo Choi, and Yong Man Ro. Synctalkface: Talking face generation with precise lip-syncing via audio-lip memory. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2062-2070, 2022. 1", + "[48] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 2, 4", + "[49] Leonardo Pepino, Pablo Riera, and Luciana Ferrer. Emotion recognition from speech using wav2vec 2.0 embeddings. arXiv preprint arXiv:2104.03502, 2021. 5, 8", + "[50] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models. arXiv preprint arXiv:2410.13720, 2024. 2", + "[51] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020. 1, 2, 7", + "[52] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2, 8", + "[53] Danilo Rezende and Shakir Mohamed. Variational inference with normalizing flows. In International conference on machine learning, pages 1530-1538. PMLR, 2015. 1, 2", + "[54] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 2, 3, 4", + "[55] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour," + ], + "bbox": [ + 516, + 92, + 906, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14708", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2", + "[56] Andrey V Savchenko. Hsemotion: High-speed emotion recognition library. Software Impacts, 14:100433, 2022. 5, 7, 8", + "[57] Maximilian Seitzer. pytorch-fid: FID Score for PyTorch. https://github.com/mseitzer/pytorch-fid, 2020. Version 0.3.0.7", + "[58] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. Advances in neural information processing systems, 32, 2019. 7", + "[59] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 4", + "[60] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2", + "[61] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2", + "[62] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arXiv preprint arXiv:2303.01469, 2023. 2", + "[63] Michal Stypulkowski, Konstantinos Vougioukas, Sen He, Maciej Zięba, Stavros Petridis, and Maja Pantic. Diffused heads: Diffusion models beat gans on talking-face generation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5091-5100, 2024. 2", + "[64] Zhiyao Sun, Tian Lv, Sheng Ye, Matthieu Lin, Jenny Sheng, Yu-Hui Wen, Minjing Yu, and Yong-jin Liu. Diffpositalk: Speech-driven stylistic 3d facial animation and head pose generation via diffusion models. ACM Transactions on Graphics (TOG), 43(4):1-9, 2024. 2", + "[65] Shuai Tan, Bin Ji, and Ye Pan. Emmn: Emotional motion memory network for audio-driven emotional talking face generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22146-22156, 2023. 2, 3", + "[66] Shuai Tan, Bin Ji, Mengxiao Bi, and Ye Pan. Edtalk: Efficient disentanglement for emotional talking head synthesis. In European Conference on Computer Vision, pages 398-416. Springer, 2025. 1, 6, 7", + "[67] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations, 2023. 2, 5, 8", + "[68] Linrui Tian, Qi Wang, Bang Zhang, and Liefeng Bo. Emo: Emote portrait alive-generating expressive portrait videos with audio2video diffusion model under weak conditions. arXiv preprint arXiv:2402.17485, 2024. 2, 3, 4, 6, 7", + "[69] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 7" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[70] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2, 4, 5", + "[71] Cong Wang, Kuan Tian, Jun Zhang, Yonghang Guan, Feng Luo, Fei Shen, Zhiwei Jiang, Qing Gu, Xiao Han, and Wei Yang. V-express: Conditional dropout for progressive training of portrait video generation. arXiv preprint arXiv:2406.02511, 2024. 2, 3", + "[72] Kaisiyuan Wang, Qianyi Wu, Linsen Song, Zhuoqian Yang, Wayne Wu, Chen Qian, Ran He, Yu Qiao, and Chen Change Loy. Mead: A large-scale audio-visual dataset for emotional talking-face generation. In European Conference on Computer Vision, pages 700–717. Springer, 2020. 3, 5", + "[73] Suzhen Wang, Lincheng Li, Yu Ding, Changjie Fan, and Xin Yu. Audio2head: Audio-driven one-shot talking-head generation with natural head motion. arXiv preprint arXiv:2107.09293, 2021. 2", + "[74] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10039-10049, 2021. 3", + "[75] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. Latent image animator: Learning to animate images via latent space navigation. arXiv preprint arXiv:2203.09043, 2022. 2, 4", + "[76] Huawei Wei, Zejun Yang, and Zhisheng Wang. Aniportrait: Audio-driven synthesis of photorealistic portrait animation. arXiv preprint arXiv:2403.17694, 2024. 2, 3", + "[77] Yibo Xia, Lizhen Wang, Xiang Deng, Xiaoyan Luo, and Yebin Liu. Gmtalker: Gaussian mixture based emotional talking video portraits. arXiv preprint arXiv:2312.07669, 2023. 3", + "[78] Liangbin Xie, Xintao Wang, Honglun Zhang, Chao Dong, and Ying Shan. Vfhq: A high-quality dataset and benchmark for video face super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 657-666, 2022. 6", + "[79] Mingwang Xu, Hui Li, Qingkun Su, Hanlin Shang, Liwei Zhang, Ce Liu, Jingdong Wang, Luc Van Gool, Yao Yao, and Siyu Zhu. Hallo: Hierarchical audio-driven visual synthesis for portrait image animation. arXiv preprint arXiv:2406.08801, 2024. 2, 3, 4, 6, 7", + "[80] Sicheng Xu, Guojun Chen, Yu-Xiao Guo, Jiaolong Yang, Chong Li, Zhenyu Zang, Yizhong Zhang, Xin Tong, and Baining Guo. Vasa-1: Lifelike audio-driven talking faces generated in real time. arXiv preprint arXiv:2404.10667, 2024. 2, 3, 5, 6, 7, 8", + "[81] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. In European conference on computer vision, pages 85-101. Springer, 2022. 2", + "[82] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "14709", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[83] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 4", + "[84] Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8652-8661, 2023. 1, 2, 6, 7", + "[85] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3661–3670, 2021. 6, 8", + "[86] Zhimeng Zhang, Zhipeng Hu, Wenjin Deng, Changjie Fan, Tangjie Lv, and Yu Ding. Dinet: Deformation inpainting network for realistic face visually dubbing on high resolution video. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3543-3551, 2023. 1", + "[87] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4176–4186, 2021. 2", + "[88] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. Makelttalk: speaker-aware talking-head animation. ACM Transactions On Graphics (TOG), 39(6):1-15, 2020. 2", + "[89] Shenhao Zhu, Junming Leo Chen, Zuozhuo Dai, Yinghui Xu, Xun Cao, Yao Yao, Hao Zhu, and Siyu Zhu. Champ: Controllable and consistent human image animation with 3d parametric guidance. arXiv preprint arXiv:2403.14781, 2024. 4" + ], + "bbox": [ + 91, + 90, + 485, + 609 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "14710", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/FLOAT_ Generative Motion Latent Flow Matching for Audio-driven Talking Portrait/5eb040a4-3c63-4680-a42f-da4ce86274a7_model.json b/2025/FLOAT_ Generative Motion Latent Flow Matching for Audio-driven Talking Portrait/5eb040a4-3c63-4680-a42f-da4ce86274a7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..83b8a8696d2d78482f9c208bf56a7e61e34b9ed2 --- /dev/null +++ b/2025/FLOAT_ Generative Motion Latent Flow Matching for Audio-driven Talking Portrait/5eb040a4-3c63-4680-a42f-da4ce86274a7_model.json @@ -0,0 +1,2853 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.13, + 0.885, + 0.175 + ], + "angle": 0, + "content": "FLOAT: Generative Motion Latent Flow Matching for Audio-driven Talking Portrait" + }, + { + "type": "text", + "bbox": [ + 0.269, + 0.203, + 0.725, + 0.222 + ], + "angle": 0, + "content": "Taekyung \\(\\mathbf{K}\\mathrm{i}^{1\\star}\\) Dongchan Min Gyeongsu Chae" + }, + { + "type": "text", + "bbox": [ + 0.391, + 0.223, + 0.608, + 0.24 + ], + "angle": 0, + "content": "\\(^{1}\\)KAIST \\(^{2}\\)DeepBrain AI Inc." + }, + { + "type": "text", + "bbox": [ + 0.254, + 0.242, + 0.746, + 0.257 + ], + "angle": 0, + "content": "{taekyung.ki, alsehdcks95}@kaist.ac.kr gc@deepbrain.io" + }, + { + "type": "text", + "bbox": [ + 0.297, + 0.26, + 0.697, + 0.274 + ], + "angle": 0, + "content": "https://deepbrainai-research.github.io/float/" + }, + { + "type": "image", + "bbox": [ + 0.162, + 0.308, + 0.832, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.505, + 0.908, + 0.547 + ], + "angle": 0, + "content": "Figure 1. FLOAT can generate a talking portrait video from a single source image and audio where the talking motion is generated by the motion latent flow matching. It can enhance the emotion-related talking motion by leveraging speech-driven emotion labels, a natural way of emotion-aware motion control." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.56, + 0.327, + 0.575 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.592, + 0.486, + 0.851 + ], + "angle": 0, + "content": "With the rapid advancement of diffusion-based generative models, portrait image animation has achieved remarkable results. However, it still faces challenges in temporally consistent video generation and fast sampling due to its iterative sampling nature. This paper presents FLOAT, an audio-driven talking portrait video generation method based on flow matching generative model. Instead of a pixel-based latent space, we take advantage of a learned orthogonal motion latent space, enabling efficient generation and editing of temporally consistent motion. To achieve this, we introduce a transformer-based vector field predictor with an effective frame-wise conditioning mechanism. Additionally, our method supports speech-driven emotion enhancement, enabling a natural incorporation of expressive motions. Extensive experiments demonstrate that our method outperforms state-of-the-art audio-driven talking portrait methods in terms of visual quality, motion fidelity, and efficiency." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.56, + 0.645, + 0.575 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.59, + 0.907, + 0.788 + ], + "angle": 0, + "content": "Animating a single image using a driving audio (i.e., audiodriven talking portrait generation) has gained significant attention in recent years for its great potential in avatar creation, video conferencing, virtual avatar chat, and user-friendly customer service. It aims to synthesize natural talking motion from audio signals, including accurate lip synchronization, rhythmical head movements, and fine-grained facial expressions. However, generating such motion solely from audio is extremely challenging due to its one-to-many correlation between audio and motion. In the earlier stage of this field, many works [8, 19, 28, 47, 51, 86] focus on generating accurate lip movements by relying on learned audio-lip alignment losses [9, 45]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.791, + 0.909, + 0.882 + ], + "angle": 0, + "content": "To comprehensively extend the range of motion, some works [45, 66, 84] incorporate probabilistic generative models, such as VAE [29] and normalizing flow [53], turning the motion generation into probabilistic sampling. However, these models still lack expressiveness in generated motion due to the limited capacity of these generative models." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Recent talking portrait generation methods [7, 21, 26, 36," + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "*This work was done during South Korea Mandatory Military Service at DeepBrain AI Inc." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14699" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.288 + ], + "angle": 0, + "content": "44, 63, 68, 71, 76, 79], powered by diffusion-based generative models [23, 61], successfully mitigate this expressiveness issue. EMO [68] introduces a promising approach to this field [7, 26, 71, 76, 79] by employing a strong pretrained image diffusion model (i.e., StableDiffusion [54]) and lifting it into video generation [24]. However, it still faces challenges in generating temporally coherent videos and achieving sampling efficiency, requiring tens of minutes for a few seconds of video. Moreover, they heavily rely on auxiliary facial prior, such as bounding boxes [68, 79], 2D landmarks and skeletons [7, 26, 82], or 3D meshes [76], which significantly restricts the diversity and the fidelity of head movements due to their strong spatial bias." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.289, + 0.482, + 0.515 + ], + "angle": 0, + "content": "In this paper, we present FLOAT, an audio-driven talking portrait video generation model based on flow matching generative model in a motion latent space. Flow matching [35, 37] has emerged as a promising alternative to diffusion models due to its fast and high-quality sampling. By modeling talking motion within a learned motion latent space [75], we can more efficiently sample temporally consistent motion latents. This is achieved by a simple yet effective transformer-based [70] vector field predictor, inspired by DiT [48]. Since our motion latent space has orthogonal structure, our method can manipulate head motion of the generated video using its basis. Furthermore, our method supports natural emotion-aware motion enhancement driven by speech. Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.517, + 0.481, + 0.576 + ], + "angle": 0, + "content": "- We present, FLOAT, flow matching based audio-driven talking portrait generation model using a learned orthogonal motion latent space, enabling to generate talking portrait videos with reduced sampling steps." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.578, + 0.48, + 0.635 + ], + "angle": 0, + "content": "- We introduce a simple yet effective transformer-based flow vector field predictor for temporally consistent motion latent sampling, which also enables the speech-driven emotional controls." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.638, + 0.48, + 0.68 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate that FLOAT achieves state-of-the-art performance compared to both diffusion- and non-diffusion-based methods." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.517, + 0.481, + 0.68 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.7, + 0.24, + 0.715 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.726, + 0.414, + 0.742 + ], + "angle": 0, + "content": "2.1. Diffusion Models and Flow Matching" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.749, + 0.482, + 0.884 + ], + "angle": 0, + "content": "Diffusion Models Diffusion models or score-based generative models [13, 23, 46, 54, 60, 61] are generative models that gradually diffuse input signals into Gaussian noise and learn the denoising reverse process for the generative modeling. They have shown remarkable results in various generation tasks, such as unconditional image and video generation [4, 16, 48], text-to-image generation [52, 54, 55], text-to-video generation [4, 20], conditional image generation [24, 82], and 3D human generation [31, 64, 67]." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Accelerating Diffusion Models While diffusion models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.333 + ], + "angle": 0, + "content": "demonstrate superior performance, their iterative sampling nature still bottlenecks the efficient generation compared to VAEs [29], normalizing flow [53], and GANs [18]. To overcome this limitation, several works have been developed to boost the sampling speed of the diffusion models. StableDiffusion (SD) [54] partially mitigates this problem by moving the diffusion process from the pixel space to the spatial latent space, establishing itself as a pivotal framework among diffusion models. Another line of research has developed the sampling solvers [40, 41] based on ordinary differential equations (ODEs). Meanwhile, model distillation [22] has been introduced to transfer the knowledge of the learned diffusion models into a student model, enabling one (or a few) steps of generation [27, 34, 38, 42, 62]. However, these approaches involve substantial effort to create a well-trained diffusion model and suffer from training instability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.335, + 0.905, + 0.561 + ], + "angle": 0, + "content": "Flow Matching Flow matching [35, 37] stands out as an alternative to diffusion models for its high sampling speed and competitive sample quality compared to diffusion models [10, 17, 32, 35, 50]. It belongs to the family of flow-based generative models, which estimates a transformation (referred to as a flow) between a prior distribution (e.g., Gaussian) and a target distribution. Unlike the normalizing flow [14, 53] that directly estimates the noise-to-data transformation under specific architectural constraints (e.g., affine coupling), flow matching regresses the time-dependent vector field that generates this flow by solving its corresponding ODEs [6] with flexible architectures. One specific design of flow matching is an optimal transport (OT) based one, which transforms the data distribution along the straight path with constant velocity [35]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.563, + 0.905, + 0.652 + ], + "angle": 0, + "content": "Our audio-driven talking portrait method employs flow matching to generate the natural talking motions. Thanks to the architectural flexibility of flow matching, we use transformer-encoder architecture [70] to estimate the generating vector field, allowing us to take the video temporal consistency into account." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.667, + 0.805, + 0.681 + ], + "angle": 0, + "content": "2.2. Audio-driven Portrait Animation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Audio-driven portrait animation is the task of generating a realistic talking portrait video using a single portrait image and driving audio [45, 73, 84, 87, 88]. Since audio-to-motion relation is basically a one-to-many problem, several works utilize additional facial prior for driving conditions, e.g., 2D facial landmarks [7, 21, 26, 71, 76, 88], 3D prior [8, 43, 44, 81, 84], or emotional labels [25, 65, 80]. In earlier stages, most works [8, 19, 28, 51] focused on generating accurate lip motion from audio by utilizing the lip-sync discriminator [9]. These approaches have advanced to generating audio-related head poses in a probabilistic way. For example, StyleTalker [45] uses normalizing flow [14, 53] to generate the head motion from audio, while SadTalker [84] uses audio-conditional variational inference [29] to learn" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14700" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "the 3DMM coefficients [2], bridging the intermediate representations of a pre-trained portrait animator [74]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.483, + 0.197 + ], + "angle": 0, + "content": "Meanwhile, several works [25, 65, 72, 77] focus on an emotion-aware talking portrait generation. In particular, EAMM [25] considers an emotion as the complementary displacement of facial motion, and learns these displacement from an emotion label extracted from the image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.198, + 0.484, + 0.379 + ], + "angle": 0, + "content": "Recent audio-driven talking portrait methods powered by diffusion models show remarkable results [7, 26, 36, 44, 68, 71, 76, 79, 80]. Specifically, EMO [68] and subsequent extensions [7, 71, 76, 79] utilize the pre-trained SD [54] as their backbone to leverage generative prior trained on the large-scale image datasets. They introduce additional modules, e.g., ReferenceNet [24] and Temporal Transformer [20], to preserve input identity and enhance the video temporal consistency, respectively. However, these modules introduces additional computational cost, requiring several minutes for a few seconds of video, and still suffer from video-level artifacts, such as noisy frames, and flickering." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.379, + 0.484, + 0.515 + ], + "angle": 0, + "content": "VASA-1 [80] addresses the sampling time issue by sampling motion latents [15], producing lifelike talking portraits. Our method takes advantage of this approach. However, unlike [80], our motion latent space has a strong linear orthogonal structure represented by a computable basis, enabling to manipulate the generated motion at the test-time without external driving signals. Based on this orthogonality, we employ OT-based flow matching for motion latent sampling along a straight line with reduced sampling steps." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.526, + 0.484, + 0.543 + ], + "angle": 0, + "content": "3. Preliminaries: (Conditional) Flow Matching" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.551, + 0.484, + 0.673 + ], + "angle": 0, + "content": "Let \\( x \\in \\mathbb{R}^d \\) be a data, \\( t \\in [0,1] \\) be the time, and \\( q \\) be a unknown target distribution. We can define a flow as a time-dependent transformation \\( \\varphi_t : [0,1] \\times \\mathbb{R}^d \\to \\mathbb{R}^d \\) that transforms a tractable prior distribution \\( p_0 \\) to the distribution \\( p_1 \\approx q \\). This flow \\( \\varphi_t \\) further introduces a probability flow path \\( p_t : [0,1] \\times \\mathbb{R}^d \\to \\mathbb{R}_{>0} \\) and a generating vector field \\( v_t : [0,1] \\times \\mathbb{R}^d \\to \\mathbb{R}^d \\) where \\( p_t \\) is defined by the push-forwarding" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.679, + 0.483, + 0.713 + ], + "angle": 0, + "content": "\\[\np _ {t} (x) = p _ {0} \\left(\\varphi_ {t} ^ {- 1} (x)\\right) \\det \\left| \\frac {\\partial \\varphi_ {t} ^ {- 1} (x)}{\\partial x} \\right|, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.721, + 0.483, + 0.751 + ], + "angle": 0, + "content": "and \\( v_{t} \\) generates \\( \\varphi_{t} \\) by means of an ordinary differential equation (ODE) [6]:" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.758, + 0.483, + 0.787 + ], + "angle": 0, + "content": "\\[\n\\frac {d}{d t} \\varphi_ {t} (x) = v _ {t} \\left(\\varphi_ {t} (x)\\right) \\quad \\text {a n d} \\quad \\varphi_ {0} (x) = x. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.793, + 0.484, + 0.824 + ], + "angle": 0, + "content": "Flow matching [35] aims to estimate the target generating vector field \\( u_{t} \\) with a neural network parameterized by \\( \\theta \\):" + }, + { + "type": "equation", + "bbox": [ + 0.178, + 0.845, + 0.483, + 0.864 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {F M}} (\\theta) := \\left\\| v _ {t} (x; \\theta) - u _ {t} (x) \\right\\| _ {2} ^ {2}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(t \\sim \\mathcal{U}[0,1]\\) and \\(x \\sim p_t(x)\\). However, the target generating vector field \\(u_t\\) and the sample distribution \\(p_t\\) are" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.168 + ], + "angle": 0, + "content": "intractable. To address this issue, [35] proposes a method for constructing a \"conditional\" probability path \\( p_t(\\cdot | x_1) \\) as well as target \"conditional\" vector field \\( u_t(\\cdot | x_1) \\) using a sample \\( x_1 \\sim q \\) as a condition. And they prove that the following objective" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.175, + 0.905, + 0.192 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C F M}} (\\theta) := \\left\\| v _ {t} (x; \\theta) - u _ {t} (x \\mid x _ {1}) \\right\\| _ {2} ^ {2}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.201, + 0.905, + 0.231 + ], + "angle": 0, + "content": "where \\(t\\sim \\mathcal{U}[0,1]\\) and \\(x\\sim p_t(x|x_1)\\), is equivalent to (3) with respect to the gradient \\(\\nabla_{\\theta}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.231, + 0.906, + 0.367 + ], + "angle": 0, + "content": "One natural way of constructing \\( u_{t}(\\cdot |x_{1}) \\) is a \"straight line\" that connects \\( x_0\\sim p_0 \\) and \\( x_{1}\\sim q \\), drawing an optimal transport (OT) path with constant velocity [35]. Specifically, a linear time interpolation between \\( x_0 \\) and \\( x_{1} \\) gives us the flow \\( x_{t} = \\varphi_{t}(x) = (1 - t)x_{0} + tx_{1} \\), the conditional probability path \\( p_t(x|x_1) \\) defined via the affine transformation \\( p_t(x|x_1) = \\mathcal{N}(x|tx_1,(1 - t)^2 I) \\), and the target generating vector field \\( u_{t}(x|x_{1}) = x_{1} - x_{0} \\). This specific choice turns the objective (4) into" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.375, + 0.905, + 0.393 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {O T}} (\\theta) := \\left\\| v _ {t} ((1 - t) x _ {0} + t x _ {1}; \\theta) - (x _ {1} - x _ {0}) \\right\\| _ {2} ^ {2}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.401, + 0.905, + 0.43 + ], + "angle": 0, + "content": "where \\(t\\sim \\mathcal{U}[0,1]\\) \\(x_0\\sim p_0\\) , and \\(x_{1}\\sim q\\) , all of which are tractable." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.431, + 0.906, + 0.507 + ], + "angle": 0, + "content": "Classifier-free Vector Field [10] formulates a classifier-free vector field (CFV) technique for flow matching, which enables class-conditional sampling more controllable manner without any extra classifier trained on noisy trajectory. Formally, CFV compute the modified vector field \\(\\tilde{v}_t\\) by" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.515, + 0.905, + 0.532 + ], + "angle": 0, + "content": "\\[\n\\tilde {v} _ {t} \\left(x _ {t}, c; \\theta\\right) \\approx \\gamma v _ {t} \\left(x _ {t}, c; \\theta\\right) + (1 - \\gamma) v _ {t} \\left(x _ {t}, c = \\emptyset ; \\theta\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.54, + 0.906, + 0.586 + ], + "angle": 0, + "content": "where \\(\\gamma\\) denotes the guidance scale. \\(v_{t}(x_{t},c = \\emptyset ;\\theta)\\) is the predicted vector field without a driving condition \\(c\\). For more details, please refer to [10, 35]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.597, + 0.906, + 0.633 + ], + "angle": 0, + "content": "4. Method: Flow Matching for Audio-driven Talking Portrait" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.641, + 0.906, + 0.686 + ], + "angle": 0, + "content": "We provide an overview of FLOAT in Fig. 2. Given source image \\( S \\in \\mathbb{R}^{3 \\times H \\times W} \\), and a driving audio signal \\( a^{1:L} \\in \\mathbb{R}^{L \\times d_a} \\) of length \\( L \\), our method generates a video" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.693, + 0.905, + 0.712 + ], + "angle": 0, + "content": "\\[\n\\hat {D} ^ {1: L} = \\left(\\hat {D} ^ {l}\\right) _ {l = 1} ^ {L} \\in \\mathbb {R} ^ {L \\times 3 \\times H \\times W} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.901 + ], + "angle": 0, + "content": "of \\(L\\) frames, featuring audio-synchronized talking head motions, including both verbal and non-verbal motions. Our method consists of two phases. First, we pre-train a motion auto-encoder, which provides us with the expressive and smooth motion latent space for the talking portraits (Sec. 4.1). Next, we employ OT-based flow matching [35] to generate a sequence of motion latents with a transformer-based vector field predictor using the driving audio, which is decoded to the talking portrait videos (Sec. 4.2). We also incorporate speech-driven emotions as the driving conditions, achieving automatic emotion-aware talking portrait generation without any extra user input for emotion." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14701" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.09, + 0.895, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.254, + 0.908, + 0.329 + ], + "angle": 0, + "content": "Figure 2. Overview of FLOAT. We encode the source image \\(S \\in \\mathbb{R}^{3 \\times H \\times W}\\) into the latent with the explicit identity-motion decomposition \\(w_{s} = w_{s \\to r} + w_{r \\to s} \\in \\mathbb{R}^{d}\\). Given audio segments \\(a^{-L':L} \\in \\mathbb{R}^{(L' + L) \\times d_{a}}\\) of the length \\(L' + L\\) and the reference motion \\(w_{r \\to s} \\in \\mathbb{R}^{d}\\), and the speech-driven emotion label \\(w_{e} \\in \\mathbb{R}^{7}\\), a flow matching transformer estimates the generating vector field \\(v_{t}(\\varphi_{t}(x_{0}), \\mathbf{c}_{t}; \\theta) \\in \\mathbb{R}^{L \\times d}\\) from noisy motion latents, which is used to solve corresponding ODE and generates the motion latents \\(w_{r \\to \\hat{D}^{1:L}}\\). Finally, the sequence of latents \\(w_{S \\to \\hat{D}^{1:L}} := (w_{S \\to r} + w_{r \\to \\hat{D}^{l}})_{l=1}^{L}\\) are decoded into the video \\(\\hat{D}^{1:L} \\in \\mathbb{R}^{L \\times 3 \\times H \\times W}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.343, + 0.347, + 0.357 + ], + "angle": 0, + "content": "4.1. Motion Latent Auto-encoder" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.364, + 0.483, + 0.485 + ], + "angle": 0, + "content": "Recent talking portrait methods utilize the VAE of StableDiffusion (SD) [54] due to its rich semantic pixel-based latent space. However, they often struggle to generate temporally consistent frames when lifted to video generating tasks [7, 24, 68, 79, 89]. Thus, our first goal for realistic talking portrait is to obtain good motion latent space, capturing both global (e.g., head motion) and fine-grained local (e.g., facial expressions, mouth and pupil movement) dynamics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.486, + 0.484, + 0.637 + ], + "angle": 0, + "content": "Instead of VAE of SD, we pre-train a motion latent autoencoder that encodes images into motion latents. This is done by training an auto-encoder to reconstruct a driving image from a source image sampled from the same video clip, enforcing the encoder to implicitly capture both temporally adjacent and distant motions. Inspired by LIA [75], we introduce a learned orthonormal basis that decomposes the motion along distinct orthogonal directions. Specifically, our motion auto-encoder encodes the source \\(S\\) into the latent \\(w_{S} \\in \\mathbb{R}^{d}\\) with following explicit decomposition:" + }, + { + "type": "equation", + "bbox": [ + 0.208, + 0.649, + 0.482, + 0.662 + ], + "angle": 0, + "content": "\\[\nw _ {S} := w _ {S \\rightarrow r} + w _ {r \\rightarrow S}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.672, + 0.381, + 0.687 + ], + "angle": 0, + "content": "where \\(w_{S\\to r}\\in \\mathbb{R}^d\\) is the identity latent and" + }, + { + "type": "equation", + "bbox": [ + 0.179, + 0.697, + 0.482, + 0.737 + ], + "angle": 0, + "content": "\\[\nw _ {r \\rightarrow S} = \\sum_ {m = 1} ^ {M} \\lambda_ {m} (S) \\cdot \\mathbf {v} _ {m} \\in \\mathbb {R} ^ {d} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.748, + 0.483, + 0.825 + ], + "angle": 0, + "content": "is the motion latent with \\(\\lambda(S) := (\\lambda_m(S))_{m=1}^M \\in \\mathbb{R}^M\\) being the source-dependent motion coefficients that span the learned source-agnostic motion basis \\(V := \\{\\mathbf{v}_m\\}_{m=1}^M \\subseteq \\mathbb{R}^d\\). In this space, \\(\\lambda_m(S)\\) is the intensity of the motion direction \\(\\mathbf{v}_m\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.871 + ], + "angle": 0, + "content": "As shown in Fig. 6, our method enables motion editing of the sampled (generated) motion using only the basis \\(V\\) and its orthogonality, as stated in Eq. (15)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Improving Fidelity of Facial Components: \\(\\mathcal{L}_{\\mathrm{comp - lp}}\\) The expressiveness of generated motions and the image fidelity" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.34, + 0.9, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.471, + 0.903, + 0.485 + ], + "angle": 0, + "content": "Figure 3. Efficacy of \\(\\mathcal{L}_{\\mathrm{comp - lp}}\\) for fine-grained motion and fidelity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.506, + 0.907, + 0.672 + ], + "angle": 0, + "content": "are determined by the motion space and the motion autoencoder. However, as resolution increases, fine details in small facial regions (e.g., teeth, eyeballs) often get buried in large-scale dynamics. To address this issue, we propose a facial component perceptual loss \\(\\mathcal{L}_{\\mathrm{comp - lp}}\\) using [59, 83] that significantly improves the image fidelity (e.g., teeth and eyes) as well as fine-grained motions (e.g., eyeball and eyebrows movements). As shown in Fig. 3, \\(\\mathcal{L}_{\\mathrm{comp - lp}}\\) allows us to generate high-fidelity facial components and their fine-grained motions without relying on pre-trained foundation models, such as StableDiffusion [54]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.682, + 0.851, + 0.699 + ], + "angle": 0, + "content": "4.2. Flow Matching in Motion Latent Space" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.885 + ], + "angle": 0, + "content": "Armed with this linear orthogonal space, we employ OT-based flow matching [35, 37] for the motion sampling. Specifically, we predict a vector field \\( v_{t}(x_{t},\\mathbf{c}_{t};\\theta)\\in \\mathbb{R}^{L\\times d} \\) where \\( x_{t} \\) is the sample at flow time \\( t\\in [0,1] \\), and \\( \\mathbf{c}_t\\in \\mathbb{R}^{L\\times h} \\) represents the driving conditions for \\( L \\) consequent frames. This vector field generates the flow \\( \\varphi_t:[0,1]\\times \\mathbb{R}^{L\\times d}\\to \\mathbb{R}^{L\\times d} \\) of \\( L \\) frames by solving ODE (Eq. (2)). As illustrated in Fig. 4, we build our vector field predictor upon the transformer encoder [70] architecture. Specifically, we adopt DiT [48] architecture, but decouple frame-wise conditioning from time-axis attention mechanism, which enables us to model temporally consistent motion latents." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "In DiT [48], distinct semantic tokens are modulated by a" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14702" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.09, + 0.465, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.219, + 0.472, + 0.233 + ], + "angle": 0, + "content": "Figure 4. Frame-wise vector field predictor block at inference." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.246, + 0.484, + 0.368 + ], + "angle": 0, + "content": "single diffusion time step embedding and class embedding through adaptive layer normalization (AdaLN). In contrast, our vector field predictor modulates each \\(l\\)-th input latent with its corresponding \\(l\\)-th condition and then combines their temporal relations through a masked self-attention layer that attends to \\(2 \\cdot T\\) neighboring frames. Formally, for each \\(l\\)-th frame, frame-wise AdaLN and frame-wise gating are computed by" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.375, + 0.482, + 0.395 + ], + "angle": 0, + "content": "\\[\n\\gamma_ {i} ^ {l} \\times \\operatorname {L N} \\left(X _ {t} ^ {l}\\right) + \\beta_ {i} ^ {l} \\in \\mathbb {R} ^ {h} \\quad \\text {a n d} \\quad \\alpha_ {i} ^ {l} \\times X _ {t} ^ {l} \\in \\mathbb {R} ^ {h}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.403, + 0.484, + 0.478 + ], + "angle": 0, + "content": "respectively, where \\(i \\in \\{1,2\\}\\), \\(h\\) is the hidden dimension, \\(\\mathrm{LN}(\\cdot)\\) denotes layer norm [33], and \\(X_{t}^{l}\\) is the \\(l\\)-th input for each operation at flow time \\(t \\in [0,1]\\). The coefficients \\(\\alpha_{i}^{l}, \\beta_{i}^{l}, \\gamma_{i}^{l} \\in \\mathbb{R}^{h}\\) are computed from the condition \\(\\mathbf{c}_{t}^{l} \\in \\mathbb{R}^{h}\\) through a linear layer, ToScaleShift, as depicted in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.479, + 0.484, + 0.599 + ], + "angle": 0, + "content": "Speech-driven Emotion Enhancement How can we make talking motions more expressive and natural? During talking, humans naturally reflect their emotions through their voices, and these emotions influence talking motions. For instance, a person who speaks sadly may be more likely to shake the head and avoid eye contact. This non-verbal motion derived from emotions crucially impacts the naturalness of a talking portrait." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.734 + ], + "angle": 0, + "content": "Existing works [25, 72, 80] use image-emotion paired data or image-driven emotion predictor [56] to generate the emotion-aware motion. In contrast, we incorporate speech-driven emotions, a more intuitive way of controlling emotion for audio-driven talking portrait. Specifically, we utilize a pre-trained speech emotion predictor [49] that produces softmax probabilities of seven distinct emotions: angry, disgust, fear, happy, neutral, sad, and surprise, which we then input into the vector field predictor." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.84 + ], + "angle": 0, + "content": "However, as people do not always speak with a single, clear emotion, determining emotions solely from audio is often ambiguous [25]. Naive introduction of speech-driven emotion can make emotion-aware motion generation more challenging. To address this issue, we inject the emotions together with other driving conditions at training phase and modify them at inference phase." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Driving Conditions We concatenate the audio representation \\( a^{1:L} \\in \\mathbb{R}^{L \\times d_a} \\) of a pre-trained Wav2Vec2.0 [1], the speech emotion label \\( w_e \\in \\mathbb{R}^7 \\), and the source motion latent \\( w_{r \\to S} \\in \\mathbb{R}^d \\). Next, we add the flow time step embedding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.091, + 0.905, + 0.152 + ], + "angle": 0, + "content": "\\(\\operatorname{Emb}(t) \\in \\mathbb{R}^h\\) to these conditions, producing \\(\\mathbf{c}_t \\in \\mathbb{R}^{L \\times h}\\) via a linear layer, ToCondition, as depicted in Fig. 2, where \\(\\operatorname{Emb}(t)\\) is computed using the sinusoidal position embedding [70]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.152, + 0.906, + 0.272 + ], + "angle": 0, + "content": "Training We train FLOAT by reconstructing a target vector field computed from driving frames using the corresponding audio segments and a source motion latent. We choose a pair of driving motions and corresponding audio \\((w_{r\\rightarrow D^{1:L}},a^{1:L})\\) , and construct the target vector field \\(u_{t}(x|w_{r\\rightarrow D^{1:L}}) = w_{r\\rightarrow D^{1:L}} - x_{0}\\in \\mathbb{R}^{L\\times d}\\) with noisy input \\(\\varphi_t(x_0) = (1 - t)x_0 + tw_{r\\to D^{1:L}}(t\\sim \\mathcal{U}[0,1]\\) and \\(x_0\\sim\\) \\(\\mathcal{N}(0^{1:L},I)\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.273, + 0.906, + 0.332 + ], + "angle": 0, + "content": "For smooth transitions of sequences longer than the window length \\( L \\), we incorporate last \\( L' \\) audio features and motion latents \\( w_{r\\rightarrow D^{-L':0}} \\) from the preceding window as additional input." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.333, + 0.864, + 0.349 + ], + "angle": 0, + "content": "The flow matching objective \\(\\mathcal{L}_{\\mathrm{OT}}(\\theta)\\) is defined by" + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.358, + 0.905, + 0.4 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {O T}} (\\theta) = \\left\\| v _ {t} ^ {1: L} \\left(x _ {t}, \\mathbf {c} _ {t}; \\theta\\right) - u _ {t} \\left(x \\mid w _ {r \\rightarrow D ^ {1: L}}\\right)\\right\\|, \\tag {11} \\\\ + \\left\\| v _ {t} ^ {- L ^ {\\prime}: 0} \\left(x _ {t}, \\mathbf {c} _ {t}; \\theta\\right) - w _ {r \\rightarrow D ^ {- L ^ {\\prime}: 0}} \\right\\|, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.906, + 0.49 + ], + "angle": 0, + "content": "where \\( x_{t} \\coloneqq [w_{r \\to D^{-L' \\cdot 0}} | \\varphi_{t}(x_{0})] \\in \\mathbb{R}^{(-L' + L) \\times d} \\) is the concatenated input, \\( \\mathbf{c}_{t} \\in \\mathbb{R}^{(-L' + L) \\times h} \\) is the driving condition consisting of \\( [t, w_{r \\to S}, w_{e}, a^{1:L}, a^{-L':0}] \\). Note that \\( w_{e} \\) and \\( w_{r \\to S} \\) are shared across the \\( L' + L \\) frames. We incorporate a velocity loss [67] to supervise temporal consistency:" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.5, + 0.905, + 0.517 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {v e l}} (\\theta) = \\left\\| \\Delta v _ {t} - \\Delta u _ {t} \\right\\|, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.527, + 0.906, + 0.573 + ], + "angle": 0, + "content": "where \\(\\Delta v_{t}\\) and \\(\\Delta u_{t}\\) are the one-frame difference along the time-axis for the prediction \\(v_{t} \\in \\mathbb{R}^{(-L' + L) \\times d}\\) and the target \\([w_{r \\to D^{-L':0}} | u_{t}] \\in \\mathbb{R}^{(-L' + L) \\times d}\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.573, + 0.733, + 0.588 + ], + "angle": 0, + "content": "The total objective \\(\\mathcal{L}_{\\mathrm{total}}(\\theta)\\) is" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.599, + 0.905, + 0.615 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} (\\theta) = \\lambda_ {\\mathrm {O T}} \\mathcal {L} _ {\\mathrm {O T}} (\\theta) + \\lambda_ {\\mathrm {v e l}} \\mathcal {L} _ {\\mathrm {v e l}} (\\theta), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.626, + 0.906, + 0.7 + ], + "angle": 0, + "content": "where \\(\\lambda_{\\mathrm{OT}}\\) and \\(\\lambda_{\\mathrm{vel}}\\) are the balancing coefficients. During training, we apply dropout to \\(w_{r}\\), \\(w_{e}\\), and \\(a^{1:L}\\) with a probability of 0.1 for CFV. Additionally, we apply dropout to the preceding audio and motion latents with a probability 0.5 for smooth transition in the initial window." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.701, + 0.906, + 0.761 + ], + "angle": 0, + "content": "Inference During inference, we sample the generating vector field from noise \\( x_0 \\), using the driving conditions \\( w_{r\\rightarrow S} \\), \\( w_e \\), and \\( a^{1:L} \\), as well as the \\( L' \\) frames of preceding audio and generated motion latents." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.762, + 0.906, + 0.793 + ], + "angle": 0, + "content": "We extend the CFV [10] to an incremental CFV to separately adjust the audio and emotion, inspired by [3]:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.803, + 0.905, + 0.861 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\tilde {v} _ {t} \\approx v _ {t} (x _ {0}, \\mathbf {c} _ {t} | _ {\\{a ^ {1: L}, w _ {e} \\}}) \\\\ + \\gamma_ {a} \\left[ v _ {t} \\left(x _ {0}, \\mathbf {c} _ {t} \\mid_ {w _ {e}}\\right) - v _ {t} \\left(x _ {0}, \\mathbf {c} _ {t} \\mid_ {\\{a ^ {1: L}, w _ {e} \\}} \\right] \\right. \\\\ + \\gamma_ {e} \\left[ v _ {t} \\left(x _ {0}, \\mathbf {c} _ {t}\\right) - v _ {t} \\left(x _ {0}, \\mathbf {c} _ {t} \\mid_ {w _ {e}}\\right) \\right], \\tag {14} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.903 + ], + "angle": 0, + "content": "where \\(\\gamma_{a}\\) and \\(\\gamma_{e}\\) are the guidance scales for audio and emotion, respectively. \\(\\mathbf{c}_t|_{\\{x,y\\}}\\) denotes the driving condition" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14703" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.09, + 0.908, + 0.118 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison results with state-of-the-art methods on HDTF [85] / RAVDESS [39]. The best result for each metric is in bold, and the second-best result is underlined." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.122, + 0.899, + 0.238 + ], + "angle": 0, + "content": "
MethodImage & Video GenerationLip Synchronization
FID ↓FVD ↓CSIM ↑E-FID ↓P-FID ↓LSE-D ↓LSE-C ↑
SadTalker† [84]71.952 / 119.430339.058 / 376.2940.644 / 0.6441.914 / 3.5001.456 / 2.0457.947 / 7.2737.305 / 4.748
EDTalk† [66]50.078 / 75.020211.284 / 304.9330.626 / 0.6761.579 / 3.4680.054 / 0.0908.123 / 7.6827.623 / 5.318
AniTalker† [36]39.512 / 70.430184.454 / 265.3410.643 / 0.7251.830 / 2.3300.092 / 0.1267.907 / 8.1767.288 / 4.555
Halo [79]25.363 / 57.648197.196 / 375.5570.869 / 0.8601.039 / 2.4920.037 / 0.0507.792 / 7.6137.582 / 4.795
EchoMimic [7]33.552 / 81.839296.757 / 320.2200.823 / 0.8051.234 / 3.2010.023 / 0.0478.903 / 8.1616.242 / 4.144
FLOAT (Ours)21.100 / 31.681162.052 / 166.3590.843 / 0.8101.229 / 1.3670.032 / 0.0317.290 / 6.9948.222 / 5.730
" + }, + { + "type": "image", + "bbox": [ + 0.103, + 0.25, + 0.46, + 0.655 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.463, + 0.371, + 0.551, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.471, + 0.452, + 0.537, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.248, + 0.885, + 0.655 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.658, + 0.908, + 0.687 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison results with state-of-the-art methods on HDTF [85] / RAVDESS [39]. Please refer to supplementary videos. Note that we additionally provide a video comparison with EMO [68] and VASA-1 [80] using their video demonstration." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.711, + 0.483, + 0.756 + ], + "angle": 0, + "content": "without the condition \\( x \\) and \\( y \\). We set \\( \\gamma_{a} = 2 \\) and \\( \\gamma_{e} = 1 \\) based on the ablation studies on \\( \\gamma_{a} \\) and \\( \\gamma_{e} \\) provided in supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.768, + 0.483, + 0.844 + ], + "angle": 0, + "content": "After sampling, ODE solver receives the estimated vector field to compute the motion latents through numerical integration. We empirically find that FLOAT can generate reasonable motion with around 10 number of function evaluations (NFE). Please refer to supplementary videos." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.9 + ], + "angle": 0, + "content": "Lastly, we add the source identity latent to the generated motion latents and decode them into video frames using the motion latent decoder." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.71, + 0.646, + 0.727 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.739, + 0.76, + 0.755 + ], + "angle": 0, + "content": "5.1. Dataset and Pre-processing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "For training the motion latent auto-encoder, we use three open-source datasets: HDTF [85], RAVDESS [39], and VFHQ [78]. When training FLOAT, we exclude VFHQ because it does not support the synchronized audio. HDTF [85] is for high-definition talking face generation, containing videos of over 300 unique identities. RAVDESS [39] includes more than 2,400 emotion-intensive videos of 24 different identities. VFHQ [78] is designed for high-resolution video super-resolution and includes a large num" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14704" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.288 + ], + "angle": 0, + "content": "ber of unique identities, which compensates the limited number of identities of the preceding datasets. Following the strategy of [58], we first convert each video to 25 FPS and resample the audio into \\(16\\mathrm{kHz}\\). Then, we crop and resize the facial region to \\(512^2\\) resolution [5]. After the pre-processing, for HDTF, we use a total of 11.3 hours of 240 videos featuring 230 different identities for training, and videos of 78 different identities, each 15 seconds long, for test. For RAVDESS, we use videos of 22 identities for training, and videos of the remaining 2 identities for test, with each 3-4 seconds long and representing 14 emotional intensities. Note that the identities in the training and test are disjoint in both datasets." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.303, + 0.308, + 0.319 + ], + "angle": 0, + "content": "5.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.327, + 0.484, + 0.508 + ], + "angle": 0, + "content": "The motion latent dimension is set to \\( d = 512 \\) with \\( M = 20 \\) distinct orthogonal directions. For the vector predictor, we use 8 attention heads, a hidden dimension \\( h = 1024 \\), and an attention window length \\( T = 2 \\). Considering the length of the training video clips, we set \\( L = 50 \\) frames with preceding \\( L' = 10 \\) frames at once, encompassing 2.4 seconds of video. We employ the Adam optimizer [30] with a batch size of 8 and a learning rate of \\( 10^{-5} \\). We use \\( L1 \\) distance for the norm \\( \\| \\cdot \\| \\) in the training objective. We set the balancing coefficients to \\( \\lambda_{\\mathrm{OT}} = \\lambda_{\\mathrm{vel}} = 1 \\). The entire training takes about 2 days for 2,000k steps on a single NVIDIA A100 GPU. We use Euler method [35] for the ODE solver." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.524, + 0.212, + 0.538 + ], + "angle": 0, + "content": "5.3. Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.548, + 0.484, + 0.683 + ], + "angle": 0, + "content": "Metrics and Baselines For evaluating the image and video generation quality, we measure Fréchet Inception Distance (FID) [57] and 16 frames Fréchet Video Distance (FVD) [69]. For facial identity, expression and head motion, we measure Cosine Similarity of identity embedding (CSIM) [11], Expression FID (E-FID) [68] and Pose FID (P-FID), respectively. Lastly, we measure Lip-Sync Error Distance and Confidence (LSE-D and LSE-C [51]) for audio-visual alignment." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.686, + 0.484, + 0.776 + ], + "angle": 0, + "content": "We compare our method with state-of-the-art audio-driven talking portrait methods whose official implementations are publicly available. For non-diffusion methods, we compare with SadTalker [84] and EDTalk [66]. For diffusion methods, we compare with AniTalker [36], Hallo [79], and EchoMimic [7]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.778, + 0.484, + 0.839 + ], + "angle": 0, + "content": "Comparison Results In Tab. 1 and Fig. 5, we show the quantitative and qualitative comparison results, respectively. FLOAT outperforms other methods on most of the metrics and visual quality in both datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Additionally, we provide video comparison results with EMO [68] and VASA-1 [80] in the supplementary materials, using their demonstration videos due to the infeasibility of direct implementation." + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.09, + 0.897, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.263, + 0.902, + 0.278 + ], + "angle": 0, + "content": "Figure 6. Test-time pose editing using \\(\\lambda\\)-control \\((\\lambda_{15}(\\hat{D}) \\pm 10)\\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.291, + 0.646, + 0.307 + ], + "angle": 0, + "content": "5.4. Applications" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.312, + 0.905, + 0.392 + ], + "angle": 0, + "content": "Test-time Pose Editing via Orthonormal Basis \\(V\\) Since FLOAT learns the underlying motion latent structure, it is natural to assume that for any sampled motion latent \\(w_{r\\rightarrow \\hat{D}}\\), there exist motion coefficients \\(\\{\\lambda_m(\\hat{D})\\}_{m = 1}^M\\) satisfying the representation in Eq. (9): \\(w_{r\\rightarrow \\hat{D}} = \\sum_{m = 1}^{M}\\lambda_{m}(\\hat{D})\\cdot \\mathbf{v}_{m}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.393, + 0.905, + 0.439 + ], + "angle": 0, + "content": "We can always compute these coefficients in closed form by taking inner products between the sampled motion \\( w_{r\\rightarrow \\hat{D}} \\) and the learned orthonormal basis \\( V \\):" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.446, + 0.905, + 0.487 + ], + "angle": 0, + "content": "\\[\n\\left\\langle w _ {r \\rightarrow \\hat {D}}, \\mathbf {v} _ {k} \\right\\rangle = \\left\\langle \\sum_ {m = 1} ^ {M} \\lambda_ {m} (\\hat {D}) \\cdot \\mathbf {v} _ {m}, \\mathbf {v} _ {k} \\right\\rangle = \\lambda_ {k} (\\hat {D}), \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.906, + 0.599 + ], + "angle": 0, + "content": "where \\(\\langle \\mathbf{v}_m, \\mathbf{v}_k \\rangle = \\delta_{m,k}\\) and \\(\\delta\\) is Kronecker delta. At this point, we can edit the sampled motions by editing the corresponding coefficients (e.g., via linear operation) and combining them back into the motion latent. As shown in Fig. 6, it allows us to control head direction without interfering with other motions due to the orthogonality of the basis. We refer to this test-time editing technique as \\(\\lambda\\)-control." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.906, + 0.78 + ], + "angle": 0, + "content": "Additional Driving Signals In Fig. 7 and Tab. 2, we experiment with additional driving conditions, head poses and image-driven emotion labels, to explore additional controllability in our method. We employ 3DMM head pose parameters \\( p \\in \\mathbb{R}^6 \\) [2] extracted by [12]. We concatenate a sequence of pose parameters \\( p^{1:L} \\in \\mathbb{R}^{L \\times 6} \\) with the other driving conditions, and then map them to \\( c_t^{1:L} \\in \\mathbb{R}^{L \\times h} \\). We also experiment on image-driven emotion [56] for frame-wise emotion control rather than the long-term emotion enhancement. FLOAT can effectively accommodate these additional conditions, highlighting its flexibility across diverse control signals." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Redirecting Speech-driven Emotion Since FLOAT learns diverse emotions in the emotion-intensive data distribution [39], the generated emotion-aware motion can be modified by redirecting the speech-driven emotion label toward a different emotion at inference time. As illustrated in Fig. 8, this technique is particularly beneficial for manual redirection when the emotion predicted from speech is complex or ambiguous." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14705" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.115, + 0.09, + 0.488, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.491, + 0.091, + 0.88, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.269, + 0.908, + 0.298 + ], + "angle": 0, + "content": "Figure 7. Additional conditioning results of FLOAT. 3Dpose, \\( S2E \\), and \\( I2E \\) denote 3D head pose parameters [12], speech-to-emotion [49], and image-to-emotion [56], respectively." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.305, + 0.484, + 0.362 + ], + "angle": 0, + "content": "Table 2. Quantitative results of FLOAT with additional conditions (HDTF [85] / RAVDESS [39]). \\( S2E \\), \\( I2E \\), and 3DPose denote speech-to-emotion [49], image-to-emotion [56], and 3DMM pose parameters [12], respectively." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.366, + 0.479, + 0.412 + ], + "angle": 0, + "content": "
ConfigurationsFID ↓FVD ↓E-FID ↓P-FID ↓LSE-D ↓
AFLOAT(Ours)21.100 / 31.681162.052 / 166.3591.229 / 1.3670.032 / 0.0317.290 / 6.994
BA + 3Dpose19.721 / 29.721126.663 / 112.8940.926 / 1.1520.012 / 0.0167.516 / 7.047
CA - S2E21.235 / 32.035155.032 / 166.8661.254 / 1.5020.031 / 0.0257.264 / 7.222
DA - S2E + I2E21/528 / 31.609158.577 / 162.3691.158 / 1.3050.034 / 0.0227.183 / 7.150
" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.419, + 0.48, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.518, + 0.484, + 0.547 + ], + "angle": 0, + "content": "Figure 8. Redirecting the unclear emotion prediction to a desirable one-hot encoding, which can be further intensified by the CFV." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.56, + 0.255, + 0.575 + ], + "angle": 0, + "content": "5.5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.583, + 0.483, + 0.734 + ], + "angle": 0, + "content": "Ablation on Frame-wise AdaLN We compare frame-wise AdaLN (and gating) followed by masked self-attention to separate conditioning from attending, with a cross-attention that performs conditioning and attending simultaneously. As shown in Tab. 3, both approaches achieve competitive image and video quality, while frame-wise AdaLN provides better expression generation and lip synchronization. We observe that frame-wise AdaLN can achieve more diverse head motions than the cross-attention. Please refer to supplementary videos." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Ablation on Flow Matching We compare flow matching with two types of diffusion models: \\(\\epsilon\\)-prediction (noise) and \\(x_0\\)-prediction (signal) [52, 67]. In both cases, we adopt our vector predictor architecture as denoising networks. We adopt diffusion training settings of VASA-1 [80] (500 diffusion steps with a cosine noise scheduler [46] and 50 DDIM denoising steps) for the indirect comparison with [80]. Notably, diffusion and flow matching achieve competitive results on image quality while the latter achieves the better lip synchronization. In Fig. 9, we compare the forward pass efficiency by measuring frames per second (FPS) of each" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.305, + 0.907, + 0.334 + ], + "angle": 0, + "content": "Table 3. Ablation studies of FLOAT on HDTF [85]. The best result for each metric is in bold, and the second-best result is underlined." + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.342, + 0.892, + 0.399 + ], + "angle": 0, + "content": "
MethodFID ↓FVD ↓E-FID ↓LSE-D ↓#NFEs ↓
Ours (w. Cross-Attn.)21.873162.7021.4527.75710
Ours (w. Diff., c-pred.)21.190161.6661.2139.92250
Ours (w. Diff., x0-pred.)21.697162.8471.2789.04850
FLOAT (Ours)21.100162.0521.2297.29010
" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.404, + 0.892, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.493, + 0.907, + 0.52 + ], + "angle": 0, + "content": "Figure 9. Comparison of the forward pass efficiency. We compute FPS on a single NVIDIA V100 GPU." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.907, + 0.598 + ], + "angle": 0, + "content": "model. Thanks to the compact motion latent representation and OT-based flow matching, FLOAT achieves the highest FPS, superior lip-sync performance, dynamic head motion, and the lowest NFEs." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.626, + 0.634, + 0.641 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.655, + 0.907, + 0.852 + ], + "angle": 0, + "content": "We proposed FLOAT, a flow matching based audio-driven talking portrait generation model leveraging a learned motion latent space. We introduced a transformer-based vector field predictor, enabling temporally consistent motion generation. Additionally, we incorporated speech-driven emotion labels into the motion sampling process to improve the naturalness of the audio-driven talking motions. FLOAT addresses current core limitations of diffusion-based talking portrait video generation methods by reducing the sampling time through flow matching while achieving the remarkable sample quality. Extensive experiments verified that FLOAT achieves state-of-the-art performance in terms of visual quality, motion fidelity, and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Discussion We leave further discussion considering limitations, future work, and ethical considerations in the supplementary materials." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14706" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449-12460, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.228 + ], + "angle": 0, + "content": "[2] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques, pages 187-194, 1999. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.231, + 0.483, + 0.298 + ], + "angle": 0, + "content": "[3] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18392-18402, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.301, + 0.483, + 0.356 + ], + "angle": 0, + "content": "[4] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.359, + 0.483, + 0.413 + ], + "angle": 0, + "content": "[5] Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2d & 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In International Conference on Computer Vision, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.416, + 0.483, + 0.47 + ], + "angle": 0, + "content": "[6] Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations. Advances in neural information processing systems, 31, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.473, + 0.483, + 0.528 + ], + "angle": 0, + "content": "[7] Zhiyuan Chen, Jiajiong Cao, Zhiquan Chen, Yuming Li, and Chenguang Ma. Echomimic: Lifelike audio-driven portrait animations through editable landmark conditions. arXiv preprint arXiv:2407.08136, 2024. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.531, + 0.483, + 0.599 + ], + "angle": 0, + "content": "[8] Kun Cheng, Xiaodong Cun, Yong Zhang, Menghan Xia, Fei Yin, Mingrui Zhu, Xuan Wang, Jue Wang, and Nannan Wang. Videoretalking: Audio-based lip synchronization for talking head video editing in the wild. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.602, + 0.483, + 0.642 + ], + "angle": 0, + "content": "[9] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In Asian Conference on Computer Vision, pages 251-263, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.645, + 0.483, + 0.685 + ], + "angle": 0, + "content": "[10] Quan Dao, Hao Phung, Binh Nguyen, and Anh Tran. Flow matching in latent space. arXiv preprint arXiv:2307.08698, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.688, + 0.483, + 0.757 + ], + "angle": 0, + "content": "[11] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4690-4699, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.76, + 0.483, + 0.841 + ], + "angle": 0, + "content": "[12] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.844, + 0.483, + 0.885 + ], + "angle": 0, + "content": "[13] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.888, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[14] Laurent Dinh, Jascha Sohl-Dickstein, and Samy Ben" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.094, + 0.905, + 0.119 + ], + "angle": 0, + "content": "gio. Density estimation using real nvp. arXiv preprint arXiv:1605.08803, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.905, + 0.191 + ], + "angle": 0, + "content": "[15] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2663-2671, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.193, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[16] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.278, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[17] Johannes S Fischer, Ming Gui, Pingchuan Ma, Nick Stracke, Stefan A Baumann, and Björn Ommer. Boosting latent diffusion with flow matching. arXiv preprint arXiv:2312.07360, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.334, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[18] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.391, + 0.905, + 0.487 + ], + "angle": 0, + "content": "[19] Jiazhi Guan, Zhanwang Zhang, Hang Zhou, Tianshu Hu, Kaisiyuan Wang, Dongliang He, Haocheng Feng, Jingtuo Liu, Errui Ding, Ziwei Liu, et al. Stylesync: High-fidelity generalized and personalized lip sync in style-based generator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1505-1515, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.49, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[20] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.561, + 0.905, + 0.615 + ], + "angle": 0, + "content": "[21] Tianyu He, Junliang Guo, Runyi Yu, Yuchi Wang, Jialiang Zhu, Kaikai An, Leyi Li, Xu Tan, Chunyu Wang, Han Hu, et al. Gaia: Zero-shot talking avatar generation. arXiv preprint arXiv:2311.15230, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.617, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[22] Geoffrey Hinton. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.647, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.689, + 0.907, + 0.744 + ], + "angle": 0, + "content": "[24] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8153-8163, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.747, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[25] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Qianyi Wu, Wayne Wu, Feng Xu, and Xun Cao. Eamm: One-shot emotional talking face via audio-based emotion-aware motion model. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.817, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[26] Jianwen Jiang, Chao Liang, Jiaqi Yang, Gaojie Lin, Tianyun Zhong, and Yanbo Zheng. Loopy: Taming audio-driven portrait avatar with long-term motion dependency. arXiv preprint arXiv:2409.02634, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Minguk Kang, Richard Zhang, Connelly Barnes, Sylvain Paris, Suha Kwak, Jaesik Park, Eli Shechtman, Jun-Yan Zhu," + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.094, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14707" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "and Taesung Park. Distilling diffusion models into conditional gans. arXiv preprint arXiv:2405.05967, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.123, + 0.484, + 0.178 + ], + "angle": 0, + "content": "[28] Taekyung Ki and Dongchan Min. Stylelipsync: Style-based personalized lip-sync video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22841-22850, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.181, + 0.482, + 0.209 + ], + "angle": 0, + "content": "[29] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.211, + 0.482, + 0.239 + ], + "angle": 0, + "content": "[30] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.241, + 0.483, + 0.31 + ], + "angle": 0, + "content": "[31] Tobias Kirschstein, Simon Giebenhain, and Matthias Nießner. Diffusionavatars: Deferred diffusion for high-fidelity 3d head avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5492, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.312, + 0.483, + 0.382 + ], + "angle": 0, + "content": "[32] Matthew Le, Apoorv Vyas, Bowen Shi, Brian Karrer, Leda Sari, Rashel Moritz, Mary Williamson, Vimal Manohar, Yossi Adi, Jay Mahadeokar, et al. Voicebox: Text-guided multilingual universal speech generation at scale. Advances in neural information processing systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.384, + 0.482, + 0.424 + ], + "angle": 0, + "content": "[33] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. ArXiv e-prints, pages arXiv-1607, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.483, + 0.495 + ], + "angle": 0, + "content": "[34] Senmao Li, Taihang Hu, Fahad Shahbaz Khan, Linxuan Li, Shiqi Yang, Yaxing Wang, Ming-Ming Cheng, and Jian Yang. Faster diffusion: Rethinking the role of unet encoder in diffusion models. arXiv preprint arXiv:2312.09608, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.499, + 0.482, + 0.541 + ], + "angle": 0, + "content": "[35] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022. 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.543, + 0.483, + 0.599 + ], + "angle": 0, + "content": "[36] Tao Liu, Feilong Chen, Shuai Fan, Chenpeng Du, Qi Chen, Xie Chen, and Kai Yu. Anitalker: Animate vivid and diverse talking faces through identity-decoupled facial motion encoding. arXiv preprint arXiv:2405.03121, 2024. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.601, + 0.482, + 0.642 + ], + "angle": 0, + "content": "[37] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.645, + 0.483, + 0.7 + ], + "angle": 0, + "content": "[38] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, et al. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. In The Twelfth International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.703, + 0.483, + 0.771 + ], + "angle": 0, + "content": "[39] Steven R Livingstone and Frank A Russo. The ryerson audio-visual database of emotional speech and song (ravdess): A dynamic, multimodal set of facial and vocal expressions in north american english. PloS one, 13(5): e0196391, 2018. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.774, + 0.483, + 0.842 + ], + "angle": 0, + "content": "[40] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[41] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[42] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.906, + 0.219 + ], + "angle": 0, + "content": "[43] Yifeng Ma, Suzhen Wang, Zhipeng Hu, Changjie Fan, Tangjie Lv, Yu Ding, Zhidong Deng, and Xin Yu. Styletalk: One-shot talking head generation with controllable speaking styles. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1896-1904, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[44] Yifeng Ma, Shiwei Zhang, Jiayu Wang, Xiang Wang, Yingya Zhang, and Zhidong Deng. Dreamtalk: When expressive talking head generation meets diffusion probabilistic models. arXiv preprint arXiv:2312.09767, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[45] Dongchan Min, Minyoung Song, Eunji Ko, and Sung Ju Hwang. Stylistalker: One-shot style-based audio-driven talking head video generation. arXiv preprint arXiv:2208.10922, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.906, + 0.388 + ], + "angle": 0, + "content": "[46] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162-8171. PMLR, 2021. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.907, + 0.459 + ], + "angle": 0, + "content": "[47] Se Jin Park, Minsu Kim, Joanna Hong, Jeongsoo Choi, and Yong Man Ro. Synctalkface: Talking face generation with precise lip-syncing via audio-lip memory. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2062-2070, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.906, + 0.516 + ], + "angle": 0, + "content": "[48] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[49] Leonardo Pepino, Pablo Riera, and Luciana Ferrer. Emotion recognition from speech using wav2vec 2.0 embeddings. arXiv preprint arXiv:2104.03502, 2021. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.562, + 0.906, + 0.629 + ], + "angle": 0, + "content": "[50] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models. arXiv preprint arXiv:2410.13720, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.906, + 0.7 + ], + "angle": 0, + "content": "[51] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.703, + 0.906, + 0.758 + ], + "angle": 0, + "content": "[52] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[53] Danilo Rezende and Shakir Mohamed. Variational inference with normalizing flows. In International conference on machine learning, pages 1530-1538. PMLR, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.907, + 0.871 + ], + "angle": 0, + "content": "[54] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[55] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "14708" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.148 + ], + "angle": 0, + "content": "Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.149, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[56] Andrey V Savchenko. Hsemotion: High-speed emotion recognition library. Software Impacts, 14:100433, 2022. 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.191, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[57] Maximilian Seitzer. pytorch-fid: FID Score for PyTorch. https://github.com/mseitzer/pytorch-fid, 2020. Version 0.3.0.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.483, + 0.287 + ], + "angle": 0, + "content": "[58] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. Advances in neural information processing systems, 32, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.288, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[59] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.33, + 0.482, + 0.371 + ], + "angle": 0, + "content": "[60] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.482, + 0.427 + ], + "angle": 0, + "content": "[61] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.428, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[62] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arXiv preprint arXiv:2303.01469, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.47, + 0.482, + 0.539 + ], + "angle": 0, + "content": "[63] Michal Stypulkowski, Konstantinos Vougioukas, Sen He, Maciej Zięba, Stavros Petridis, and Maja Pantic. Diffused heads: Diffusion models beat gans on talking-face generation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5091-5100, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.54, + 0.482, + 0.608 + ], + "angle": 0, + "content": "[64] Zhiyao Sun, Tian Lv, Sheng Ye, Matthieu Lin, Jenny Sheng, Yu-Hui Wen, Minjing Yu, and Yong-jin Liu. Diffpositalk: Speech-driven stylistic 3d facial animation and head pose generation via diffusion models. ACM Transactions on Graphics (TOG), 43(4):1-9, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.609, + 0.482, + 0.677 + ], + "angle": 0, + "content": "[65] Shuai Tan, Bin Ji, and Ye Pan. Emmn: Emotional motion memory network for audio-driven emotional talking face generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22146-22156, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.734 + ], + "angle": 0, + "content": "[66] Shuai Tan, Bin Ji, Mengxiao Bi, and Ye Pan. Edtalk: Efficient disentanglement for emotional talking head synthesis. In European Conference on Computer Vision, pages 398-416. Springer, 2025. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.482, + 0.789 + ], + "angle": 0, + "content": "[67] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations, 2023. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[68] Linrui Tian, Qi Wang, Bang Zhang, and Liefeng Bo. Emo: Emote portrait alive-generating expressive portrait videos with audio2video diffusion model under weak conditions. arXiv preprint arXiv:2402.17485, 2024. 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[69] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.121 + ], + "angle": 0, + "content": "[70] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.906, + 0.191 + ], + "angle": 0, + "content": "[71] Cong Wang, Kuan Tian, Jun Zhang, Yonghang Guan, Feng Luo, Fei Shen, Zhiwei Jiang, Qing Gu, Xiao Han, and Wei Yang. V-express: Conditional dropout for progressive training of portrait video generation. arXiv preprint arXiv:2406.02511, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.263 + ], + "angle": 0, + "content": "[72] Kaisiyuan Wang, Qianyi Wu, Linsen Song, Zhuoqian Yang, Wayne Wu, Chen Qian, Ran He, Yu Qiao, and Chen Change Loy. Mead: A large-scale audio-visual dataset for emotional talking-face generation. In European Conference on Computer Vision, pages 700–717. Springer, 2020. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[73] Suzhen Wang, Lincheng Li, Yu Ding, Changjie Fan, and Xin Yu. Audio2head: Audio-driven one-shot talking-head generation with natural head motion. arXiv preprint arXiv:2107.09293, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[74] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10039-10049, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[75] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. Latent image animator: Learning to animate images via latent space navigation. arXiv preprint arXiv:2203.09043, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.904, + 0.49 + ], + "angle": 0, + "content": "[76] Huawei Wei, Zejun Yang, and Zhisheng Wang. Aniportrait: Audio-driven synthesis of photorealistic portrait animation. arXiv preprint arXiv:2403.17694, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[77] Yibo Xia, Lizhen Wang, Xiang Deng, Xiaoyan Luo, and Yebin Liu. Gmtalker: Gaussian mixture based emotional talking video portraits. arXiv preprint arXiv:2312.07669, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.907, + 0.618 + ], + "angle": 0, + "content": "[78] Liangbin Xie, Xintao Wang, Honglun Zhang, Chao Dong, and Ying Shan. Vfhq: A high-quality dataset and benchmark for video face super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 657-666, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.62, + 0.906, + 0.688 + ], + "angle": 0, + "content": "[79] Mingwang Xu, Hui Li, Qingkun Su, Hanlin Shang, Liwei Zhang, Ce Liu, Jingdong Wang, Luc Van Gool, Yao Yao, and Siyu Zhu. Hallo: Hierarchical audio-driven visual synthesis for portrait image animation. arXiv preprint arXiv:2406.08801, 2024. 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.69, + 0.906, + 0.758 + ], + "angle": 0, + "content": "[80] Sicheng Xu, Guojun Chen, Yu-Xiao Guo, Jiaolong Yang, Chong Li, Zhenyu Zang, Yizhong Zhang, Xin Tong, and Baining Guo. Vasa-1: Lifelike audio-driven talking faces generated in real time. arXiv preprint arXiv:2404.10667, 2024. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.906, + 0.844 + ], + "angle": 0, + "content": "[81] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. In European conference on computer vision, pages 85-101. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[82] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14709" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[83] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.483, + 0.258 + ], + "angle": 0, + "content": "[84] Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8652-8661, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.262, + 0.483, + 0.33 + ], + "angle": 0, + "content": "[85] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3661–3670, 2021. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.332, + 0.483, + 0.401 + ], + "angle": 0, + "content": "[86] Zhimeng Zhang, Zhipeng Hu, Wenjin Deng, Changjie Fan, Tangjie Lv, and Yu Ding. Dinet: Deformation inpainting network for realistic face visually dubbing on high resolution video. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3543-3551, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.403, + 0.483, + 0.485 + ], + "angle": 0, + "content": "[87] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4176–4186, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.487, + 0.483, + 0.541 + ], + "angle": 0, + "content": "[88] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. Makelttalk: speaker-aware talking-head animation. ACM Transactions On Graphics (TOG), 39(6):1-15, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.543, + 0.483, + 0.611 + ], + "angle": 0, + "content": "[89] Shenhao Zhu, Junming Leo Chen, Zuozhuo Dai, Yinghui Xu, Xun Cao, Yao Yao, Hao Zhu, and Siyu Zhu. Champ: Controllable and consistent human image animation with 3d parametric guidance. arXiv preprint arXiv:2403.14781, 2024. 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.611 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14710" + } + ] +] \ No newline at end of file diff --git a/2025/FLOAT_ Generative Motion Latent Flow Matching for Audio-driven Talking Portrait/5eb040a4-3c63-4680-a42f-da4ce86274a7_origin.pdf b/2025/FLOAT_ Generative Motion Latent Flow Matching for Audio-driven Talking Portrait/5eb040a4-3c63-4680-a42f-da4ce86274a7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2f32b07eea1981a1616c6132873ff671c6fd2a70 --- /dev/null +++ b/2025/FLOAT_ Generative Motion Latent Flow Matching for Audio-driven Talking Portrait/5eb040a4-3c63-4680-a42f-da4ce86274a7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4bee473d0fba45cddc5e582b44a4d0a1d4aa6a6df4f325004a64bb0f4b2cb432 +size 5803476 diff --git a/2025/FLOAT_ Generative Motion Latent Flow Matching for Audio-driven Talking Portrait/full.md b/2025/FLOAT_ Generative Motion Latent Flow Matching for Audio-driven Talking Portrait/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d3f96d6788dbd8092fd2288906df85ee16dcef82 --- /dev/null +++ b/2025/FLOAT_ Generative Motion Latent Flow Matching for Audio-driven Talking Portrait/full.md @@ -0,0 +1,401 @@ +# FLOAT: Generative Motion Latent Flow Matching for Audio-driven Talking Portrait + +Taekyung $\mathbf{K}\mathrm{i}^{1\star}$ Dongchan Min Gyeongsu Chae + +$^{1}$ KAIST $^{2}$ DeepBrain AI Inc. + +{taekyung.ki, alsehdcks95}@kaist.ac.kr gc@deepbrain.io + +https://deepbrainai-research.github.io/float/ + +![](images/426ecc2dd395074757380a5c7cc5bc5fe8afac3b04755b79925b975e9e7bc694.jpg) +Figure 1. FLOAT can generate a talking portrait video from a single source image and audio where the talking motion is generated by the motion latent flow matching. It can enhance the emotion-related talking motion by leveraging speech-driven emotion labels, a natural way of emotion-aware motion control. + +# Abstract + +With the rapid advancement of diffusion-based generative models, portrait image animation has achieved remarkable results. However, it still faces challenges in temporally consistent video generation and fast sampling due to its iterative sampling nature. This paper presents FLOAT, an audio-driven talking portrait video generation method based on flow matching generative model. Instead of a pixel-based latent space, we take advantage of a learned orthogonal motion latent space, enabling efficient generation and editing of temporally consistent motion. To achieve this, we introduce a transformer-based vector field predictor with an effective frame-wise conditioning mechanism. Additionally, our method supports speech-driven emotion enhancement, enabling a natural incorporation of expressive motions. Extensive experiments demonstrate that our method outperforms state-of-the-art audio-driven talking portrait methods in terms of visual quality, motion fidelity, and efficiency. + +# 1. Introduction + +Animating a single image using a driving audio (i.e., audiodriven talking portrait generation) has gained significant attention in recent years for its great potential in avatar creation, video conferencing, virtual avatar chat, and user-friendly customer service. It aims to synthesize natural talking motion from audio signals, including accurate lip synchronization, rhythmical head movements, and fine-grained facial expressions. However, generating such motion solely from audio is extremely challenging due to its one-to-many correlation between audio and motion. In the earlier stage of this field, many works [8, 19, 28, 47, 51, 86] focus on generating accurate lip movements by relying on learned audio-lip alignment losses [9, 45]. + +To comprehensively extend the range of motion, some works [45, 66, 84] incorporate probabilistic generative models, such as VAE [29] and normalizing flow [53], turning the motion generation into probabilistic sampling. However, these models still lack expressiveness in generated motion due to the limited capacity of these generative models. + +Recent talking portrait generation methods [7, 21, 26, 36, + +44, 63, 68, 71, 76, 79], powered by diffusion-based generative models [23, 61], successfully mitigate this expressiveness issue. EMO [68] introduces a promising approach to this field [7, 26, 71, 76, 79] by employing a strong pretrained image diffusion model (i.e., StableDiffusion [54]) and lifting it into video generation [24]. However, it still faces challenges in generating temporally coherent videos and achieving sampling efficiency, requiring tens of minutes for a few seconds of video. Moreover, they heavily rely on auxiliary facial prior, such as bounding boxes [68, 79], 2D landmarks and skeletons [7, 26, 82], or 3D meshes [76], which significantly restricts the diversity and the fidelity of head movements due to their strong spatial bias. + +In this paper, we present FLOAT, an audio-driven talking portrait video generation model based on flow matching generative model in a motion latent space. Flow matching [35, 37] has emerged as a promising alternative to diffusion models due to its fast and high-quality sampling. By modeling talking motion within a learned motion latent space [75], we can more efficiently sample temporally consistent motion latents. This is achieved by a simple yet effective transformer-based [70] vector field predictor, inspired by DiT [48]. Since our motion latent space has orthogonal structure, our method can manipulate head motion of the generated video using its basis. Furthermore, our method supports natural emotion-aware motion enhancement driven by speech. Our contributions are summarized as follows: + +- We present, FLOAT, flow matching based audio-driven talking portrait generation model using a learned orthogonal motion latent space, enabling to generate talking portrait videos with reduced sampling steps. +- We introduce a simple yet effective transformer-based flow vector field predictor for temporally consistent motion latent sampling, which also enables the speech-driven emotional controls. +- Extensive experiments demonstrate that FLOAT achieves state-of-the-art performance compared to both diffusion- and non-diffusion-based methods. + +# 2. Related Works + +# 2.1. Diffusion Models and Flow Matching + +Diffusion Models Diffusion models or score-based generative models [13, 23, 46, 54, 60, 61] are generative models that gradually diffuse input signals into Gaussian noise and learn the denoising reverse process for the generative modeling. They have shown remarkable results in various generation tasks, such as unconditional image and video generation [4, 16, 48], text-to-image generation [52, 54, 55], text-to-video generation [4, 20], conditional image generation [24, 82], and 3D human generation [31, 64, 67]. + +Accelerating Diffusion Models While diffusion models + +demonstrate superior performance, their iterative sampling nature still bottlenecks the efficient generation compared to VAEs [29], normalizing flow [53], and GANs [18]. To overcome this limitation, several works have been developed to boost the sampling speed of the diffusion models. StableDiffusion (SD) [54] partially mitigates this problem by moving the diffusion process from the pixel space to the spatial latent space, establishing itself as a pivotal framework among diffusion models. Another line of research has developed the sampling solvers [40, 41] based on ordinary differential equations (ODEs). Meanwhile, model distillation [22] has been introduced to transfer the knowledge of the learned diffusion models into a student model, enabling one (or a few) steps of generation [27, 34, 38, 42, 62]. However, these approaches involve substantial effort to create a well-trained diffusion model and suffer from training instability. + +Flow Matching Flow matching [35, 37] stands out as an alternative to diffusion models for its high sampling speed and competitive sample quality compared to diffusion models [10, 17, 32, 35, 50]. It belongs to the family of flow-based generative models, which estimates a transformation (referred to as a flow) between a prior distribution (e.g., Gaussian) and a target distribution. Unlike the normalizing flow [14, 53] that directly estimates the noise-to-data transformation under specific architectural constraints (e.g., affine coupling), flow matching regresses the time-dependent vector field that generates this flow by solving its corresponding ODEs [6] with flexible architectures. One specific design of flow matching is an optimal transport (OT) based one, which transforms the data distribution along the straight path with constant velocity [35]. + +Our audio-driven talking portrait method employs flow matching to generate the natural talking motions. Thanks to the architectural flexibility of flow matching, we use transformer-encoder architecture [70] to estimate the generating vector field, allowing us to take the video temporal consistency into account. + +# 2.2. Audio-driven Portrait Animation + +Audio-driven portrait animation is the task of generating a realistic talking portrait video using a single portrait image and driving audio [45, 73, 84, 87, 88]. Since audio-to-motion relation is basically a one-to-many problem, several works utilize additional facial prior for driving conditions, e.g., 2D facial landmarks [7, 21, 26, 71, 76, 88], 3D prior [8, 43, 44, 81, 84], or emotional labels [25, 65, 80]. In earlier stages, most works [8, 19, 28, 51] focused on generating accurate lip motion from audio by utilizing the lip-sync discriminator [9]. These approaches have advanced to generating audio-related head poses in a probabilistic way. For example, StyleTalker [45] uses normalizing flow [14, 53] to generate the head motion from audio, while SadTalker [84] uses audio-conditional variational inference [29] to learn + +the 3DMM coefficients [2], bridging the intermediate representations of a pre-trained portrait animator [74]. + +Meanwhile, several works [25, 65, 72, 77] focus on an emotion-aware talking portrait generation. In particular, EAMM [25] considers an emotion as the complementary displacement of facial motion, and learns these displacement from an emotion label extracted from the image. + +Recent audio-driven talking portrait methods powered by diffusion models show remarkable results [7, 26, 36, 44, 68, 71, 76, 79, 80]. Specifically, EMO [68] and subsequent extensions [7, 71, 76, 79] utilize the pre-trained SD [54] as their backbone to leverage generative prior trained on the large-scale image datasets. They introduce additional modules, e.g., ReferenceNet [24] and Temporal Transformer [20], to preserve input identity and enhance the video temporal consistency, respectively. However, these modules introduces additional computational cost, requiring several minutes for a few seconds of video, and still suffer from video-level artifacts, such as noisy frames, and flickering. + +VASA-1 [80] addresses the sampling time issue by sampling motion latents [15], producing lifelike talking portraits. Our method takes advantage of this approach. However, unlike [80], our motion latent space has a strong linear orthogonal structure represented by a computable basis, enabling to manipulate the generated motion at the test-time without external driving signals. Based on this orthogonality, we employ OT-based flow matching for motion latent sampling along a straight line with reduced sampling steps. + +# 3. Preliminaries: (Conditional) Flow Matching + +Let $x \in \mathbb{R}^d$ be a data, $t \in [0,1]$ be the time, and $q$ be a unknown target distribution. We can define a flow as a time-dependent transformation $\varphi_t : [0,1] \times \mathbb{R}^d \to \mathbb{R}^d$ that transforms a tractable prior distribution $p_0$ to the distribution $p_1 \approx q$ . This flow $\varphi_t$ further introduces a probability flow path $p_t : [0,1] \times \mathbb{R}^d \to \mathbb{R}_{>0}$ and a generating vector field $v_t : [0,1] \times \mathbb{R}^d \to \mathbb{R}^d$ where $p_t$ is defined by the push-forwarding + +$$ +p _ {t} (x) = p _ {0} \left(\varphi_ {t} ^ {- 1} (x)\right) \det \left| \frac {\partial \varphi_ {t} ^ {- 1} (x)}{\partial x} \right|, \tag {1} +$$ + +and $v_{t}$ generates $\varphi_{t}$ by means of an ordinary differential equation (ODE) [6]: + +$$ +\frac {d}{d t} \varphi_ {t} (x) = v _ {t} \left(\varphi_ {t} (x)\right) \quad \text {a n d} \quad \varphi_ {0} (x) = x. \tag {2} +$$ + +Flow matching [35] aims to estimate the target generating vector field $u_{t}$ with a neural network parameterized by $\theta$ : + +$$ +\mathcal {L} _ {\mathrm {F M}} (\theta) := \left\| v _ {t} (x; \theta) - u _ {t} (x) \right\| _ {2} ^ {2}, \tag {3} +$$ + +where $t \sim \mathcal{U}[0,1]$ and $x \sim p_t(x)$ . However, the target generating vector field $u_t$ and the sample distribution $p_t$ are + +intractable. To address this issue, [35] proposes a method for constructing a "conditional" probability path $p_t(\cdot | x_1)$ as well as target "conditional" vector field $u_t(\cdot | x_1)$ using a sample $x_1 \sim q$ as a condition. And they prove that the following objective + +$$ +\mathcal {L} _ {\mathrm {C F M}} (\theta) := \left\| v _ {t} (x; \theta) - u _ {t} (x \mid x _ {1}) \right\| _ {2} ^ {2}, \tag {4} +$$ + +where $t\sim \mathcal{U}[0,1]$ and $x\sim p_t(x|x_1)$ , is equivalent to (3) with respect to the gradient $\nabla_{\theta}$ . + +One natural way of constructing $u_{t}(\cdot |x_{1})$ is a "straight line" that connects $x_0\sim p_0$ and $x_{1}\sim q$ , drawing an optimal transport (OT) path with constant velocity [35]. Specifically, a linear time interpolation between $x_0$ and $x_{1}$ gives us the flow $x_{t} = \varphi_{t}(x) = (1 - t)x_{0} + tx_{1}$ , the conditional probability path $p_t(x|x_1)$ defined via the affine transformation $p_t(x|x_1) = \mathcal{N}(x|tx_1,(1 - t)^2 I)$ , and the target generating vector field $u_{t}(x|x_{1}) = x_{1} - x_{0}$ . This specific choice turns the objective (4) into + +$$ +\mathcal {L} _ {\mathrm {O T}} (\theta) := \left\| v _ {t} ((1 - t) x _ {0} + t x _ {1}; \theta) - (x _ {1} - x _ {0}) \right\| _ {2} ^ {2}, \tag {5} +$$ + +where $t\sim \mathcal{U}[0,1]$ $x_0\sim p_0$ , and $x_{1}\sim q$ , all of which are tractable. + +Classifier-free Vector Field [10] formulates a classifier-free vector field (CFV) technique for flow matching, which enables class-conditional sampling more controllable manner without any extra classifier trained on noisy trajectory. Formally, CFV compute the modified vector field $\tilde{v}_t$ by + +$$ +\tilde {v} _ {t} \left(x _ {t}, c; \theta\right) \approx \gamma v _ {t} \left(x _ {t}, c; \theta\right) + (1 - \gamma) v _ {t} \left(x _ {t}, c = \emptyset ; \theta\right), \tag {6} +$$ + +where $\gamma$ denotes the guidance scale. $v_{t}(x_{t},c = \emptyset ;\theta)$ is the predicted vector field without a driving condition $c$ . For more details, please refer to [10, 35]. + +# 4. Method: Flow Matching for Audio-driven Talking Portrait + +We provide an overview of FLOAT in Fig. 2. Given source image $S \in \mathbb{R}^{3 \times H \times W}$ , and a driving audio signal $a^{1:L} \in \mathbb{R}^{L \times d_a}$ of length $L$ , our method generates a video + +$$ +\hat {D} ^ {1: L} = \left(\hat {D} ^ {l}\right) _ {l = 1} ^ {L} \in \mathbb {R} ^ {L \times 3 \times H \times W} \tag {7} +$$ + +of $L$ frames, featuring audio-synchronized talking head motions, including both verbal and non-verbal motions. Our method consists of two phases. First, we pre-train a motion auto-encoder, which provides us with the expressive and smooth motion latent space for the talking portraits (Sec. 4.1). Next, we employ OT-based flow matching [35] to generate a sequence of motion latents with a transformer-based vector field predictor using the driving audio, which is decoded to the talking portrait videos (Sec. 4.2). We also incorporate speech-driven emotions as the driving conditions, achieving automatic emotion-aware talking portrait generation without any extra user input for emotion. + +![](images/a0dbc4d307dc1ce96a87d77c15d61dadc1ceb8953c76d8b4364fe5b1d5c18c47.jpg) +Figure 2. Overview of FLOAT. We encode the source image $S \in \mathbb{R}^{3 \times H \times W}$ into the latent with the explicit identity-motion decomposition $w_{s} = w_{s \to r} + w_{r \to s} \in \mathbb{R}^{d}$ . Given audio segments $a^{-L':L} \in \mathbb{R}^{(L' + L) \times d_{a}}$ of the length $L' + L$ and the reference motion $w_{r \to s} \in \mathbb{R}^{d}$ , and the speech-driven emotion label $w_{e} \in \mathbb{R}^{7}$ , a flow matching transformer estimates the generating vector field $v_{t}(\varphi_{t}(x_{0}), \mathbf{c}_{t}; \theta) \in \mathbb{R}^{L \times d}$ from noisy motion latents, which is used to solve corresponding ODE and generates the motion latents $w_{r \to \hat{D}^{1:L}}$ . Finally, the sequence of latents $w_{S \to \hat{D}^{1:L}} := (w_{S \to r} + w_{r \to \hat{D}^{l}})_{l=1}^{L}$ are decoded into the video $\hat{D}^{1:L} \in \mathbb{R}^{L \times 3 \times H \times W}$ . + +# 4.1. Motion Latent Auto-encoder + +Recent talking portrait methods utilize the VAE of StableDiffusion (SD) [54] due to its rich semantic pixel-based latent space. However, they often struggle to generate temporally consistent frames when lifted to video generating tasks [7, 24, 68, 79, 89]. Thus, our first goal for realistic talking portrait is to obtain good motion latent space, capturing both global (e.g., head motion) and fine-grained local (e.g., facial expressions, mouth and pupil movement) dynamics. + +Instead of VAE of SD, we pre-train a motion latent autoencoder that encodes images into motion latents. This is done by training an auto-encoder to reconstruct a driving image from a source image sampled from the same video clip, enforcing the encoder to implicitly capture both temporally adjacent and distant motions. Inspired by LIA [75], we introduce a learned orthonormal basis that decomposes the motion along distinct orthogonal directions. Specifically, our motion auto-encoder encodes the source $S$ into the latent $w_{S} \in \mathbb{R}^{d}$ with following explicit decomposition: + +$$ +w _ {S} := w _ {S \rightarrow r} + w _ {r \rightarrow S}, \tag {8} +$$ + +where $w_{S\to r}\in \mathbb{R}^d$ is the identity latent and + +$$ +w _ {r \rightarrow S} = \sum_ {m = 1} ^ {M} \lambda_ {m} (S) \cdot \mathbf {v} _ {m} \in \mathbb {R} ^ {d} \tag {9} +$$ + +is the motion latent with $\lambda(S) := (\lambda_m(S))_{m=1}^M \in \mathbb{R}^M$ being the source-dependent motion coefficients that span the learned source-agnostic motion basis $V := \{\mathbf{v}_m\}_{m=1}^M \subseteq \mathbb{R}^d$ . In this space, $\lambda_m(S)$ is the intensity of the motion direction $\mathbf{v}_m$ . + +As shown in Fig. 6, our method enables motion editing of the sampled (generated) motion using only the basis $V$ and its orthogonality, as stated in Eq. (15). + +Improving Fidelity of Facial Components: $\mathcal{L}_{\mathrm{comp - lp}}$ The expressiveness of generated motions and the image fidelity + +![](images/5333d2792990caafb26491a4c5f2820a0f715920319784d7894970a3369fd475.jpg) +Figure 3. Efficacy of $\mathcal{L}_{\mathrm{comp - lp}}$ for fine-grained motion and fidelity. + +are determined by the motion space and the motion autoencoder. However, as resolution increases, fine details in small facial regions (e.g., teeth, eyeballs) often get buried in large-scale dynamics. To address this issue, we propose a facial component perceptual loss $\mathcal{L}_{\mathrm{comp - lp}}$ using [59, 83] that significantly improves the image fidelity (e.g., teeth and eyes) as well as fine-grained motions (e.g., eyeball and eyebrows movements). As shown in Fig. 3, $\mathcal{L}_{\mathrm{comp - lp}}$ allows us to generate high-fidelity facial components and their fine-grained motions without relying on pre-trained foundation models, such as StableDiffusion [54]. + +# 4.2. Flow Matching in Motion Latent Space + +Armed with this linear orthogonal space, we employ OT-based flow matching [35, 37] for the motion sampling. Specifically, we predict a vector field $v_{t}(x_{t},\mathbf{c}_{t};\theta)\in \mathbb{R}^{L\times d}$ where $x_{t}$ is the sample at flow time $t\in [0,1]$ , and $\mathbf{c}_t\in \mathbb{R}^{L\times h}$ represents the driving conditions for $L$ consequent frames. This vector field generates the flow $\varphi_t:[0,1]\times \mathbb{R}^{L\times d}\to \mathbb{R}^{L\times d}$ of $L$ frames by solving ODE (Eq. (2)). As illustrated in Fig. 4, we build our vector field predictor upon the transformer encoder [70] architecture. Specifically, we adopt DiT [48] architecture, but decouple frame-wise conditioning from time-axis attention mechanism, which enables us to model temporally consistent motion latents. + +In DiT [48], distinct semantic tokens are modulated by a + +![](images/13dc8fc6b5b49496a5c35bf6618e63a51d637fbb5f4fc548a2fd89effb668ceb.jpg) +Figure 4. Frame-wise vector field predictor block at inference. + +single diffusion time step embedding and class embedding through adaptive layer normalization (AdaLN). In contrast, our vector field predictor modulates each $l$ -th input latent with its corresponding $l$ -th condition and then combines their temporal relations through a masked self-attention layer that attends to $2 \cdot T$ neighboring frames. Formally, for each $l$ -th frame, frame-wise AdaLN and frame-wise gating are computed by + +$$ +\gamma_ {i} ^ {l} \times \operatorname {L N} \left(X _ {t} ^ {l}\right) + \beta_ {i} ^ {l} \in \mathbb {R} ^ {h} \quad \text {a n d} \quad \alpha_ {i} ^ {l} \times X _ {t} ^ {l} \in \mathbb {R} ^ {h}, \tag {10} +$$ + +respectively, where $i \in \{1,2\}$ , $h$ is the hidden dimension, $\mathrm{LN}(\cdot)$ denotes layer norm [33], and $X_{t}^{l}$ is the $l$ -th input for each operation at flow time $t \in [0,1]$ . The coefficients $\alpha_{i}^{l}, \beta_{i}^{l}, \gamma_{i}^{l} \in \mathbb{R}^{h}$ are computed from the condition $\mathbf{c}_{t}^{l} \in \mathbb{R}^{h}$ through a linear layer, ToScaleShift, as depicted in Fig. 4. + +Speech-driven Emotion Enhancement How can we make talking motions more expressive and natural? During talking, humans naturally reflect their emotions through their voices, and these emotions influence talking motions. For instance, a person who speaks sadly may be more likely to shake the head and avoid eye contact. This non-verbal motion derived from emotions crucially impacts the naturalness of a talking portrait. + +Existing works [25, 72, 80] use image-emotion paired data or image-driven emotion predictor [56] to generate the emotion-aware motion. In contrast, we incorporate speech-driven emotions, a more intuitive way of controlling emotion for audio-driven talking portrait. Specifically, we utilize a pre-trained speech emotion predictor [49] that produces softmax probabilities of seven distinct emotions: angry, disgust, fear, happy, neutral, sad, and surprise, which we then input into the vector field predictor. + +However, as people do not always speak with a single, clear emotion, determining emotions solely from audio is often ambiguous [25]. Naive introduction of speech-driven emotion can make emotion-aware motion generation more challenging. To address this issue, we inject the emotions together with other driving conditions at training phase and modify them at inference phase. + +Driving Conditions We concatenate the audio representation $a^{1:L} \in \mathbb{R}^{L \times d_a}$ of a pre-trained Wav2Vec2.0 [1], the speech emotion label $w_e \in \mathbb{R}^7$ , and the source motion latent $w_{r \to S} \in \mathbb{R}^d$ . Next, we add the flow time step embedding + +$\operatorname{Emb}(t) \in \mathbb{R}^h$ to these conditions, producing $\mathbf{c}_t \in \mathbb{R}^{L \times h}$ via a linear layer, ToCondition, as depicted in Fig. 2, where $\operatorname{Emb}(t)$ is computed using the sinusoidal position embedding [70]. + +Training We train FLOAT by reconstructing a target vector field computed from driving frames using the corresponding audio segments and a source motion latent. We choose a pair of driving motions and corresponding audio $(w_{r\rightarrow D^{1:L}},a^{1:L})$ , and construct the target vector field $u_{t}(x|w_{r\rightarrow D^{1:L}}) = w_{r\rightarrow D^{1:L}} - x_{0}\in \mathbb{R}^{L\times d}$ with noisy input $\varphi_t(x_0) = (1 - t)x_0 + tw_{r\to D^{1:L}}(t\sim \mathcal{U}[0,1]$ and $x_0\sim$ $\mathcal{N}(0^{1:L},I)$ + +For smooth transitions of sequences longer than the window length $L$ , we incorporate last $L'$ audio features and motion latents $w_{r\rightarrow D^{-L':0}}$ from the preceding window as additional input. + +The flow matching objective $\mathcal{L}_{\mathrm{OT}}(\theta)$ is defined by + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {O T}} (\theta) = \left\| v _ {t} ^ {1: L} \left(x _ {t}, \mathbf {c} _ {t}; \theta\right) - u _ {t} \left(x \mid w _ {r \rightarrow D ^ {1: L}}\right)\right\|, \tag {11} \\ + \left\| v _ {t} ^ {- L ^ {\prime}: 0} \left(x _ {t}, \mathbf {c} _ {t}; \theta\right) - w _ {r \rightarrow D ^ {- L ^ {\prime}: 0}} \right\|, \\ \end{array} +$$ + +where $x_{t} \coloneqq [w_{r \to D^{-L' \cdot 0}} | \varphi_{t}(x_{0})] \in \mathbb{R}^{(-L' + L) \times d}$ is the concatenated input, $\mathbf{c}_{t} \in \mathbb{R}^{(-L' + L) \times h}$ is the driving condition consisting of $[t, w_{r \to S}, w_{e}, a^{1:L}, a^{-L':0}]$ . Note that $w_{e}$ and $w_{r \to S}$ are shared across the $L' + L$ frames. We incorporate a velocity loss [67] to supervise temporal consistency: + +$$ +\mathcal {L} _ {\mathrm {v e l}} (\theta) = \left\| \Delta v _ {t} - \Delta u _ {t} \right\|, \tag {12} +$$ + +where $\Delta v_{t}$ and $\Delta u_{t}$ are the one-frame difference along the time-axis for the prediction $v_{t} \in \mathbb{R}^{(-L' + L) \times d}$ and the target $[w_{r \to D^{-L':0}} | u_{t}] \in \mathbb{R}^{(-L' + L) \times d}$ , respectively. + +The total objective $\mathcal{L}_{\mathrm{total}}(\theta)$ is + +$$ +\mathcal {L} _ {\text {t o t a l}} (\theta) = \lambda_ {\mathrm {O T}} \mathcal {L} _ {\mathrm {O T}} (\theta) + \lambda_ {\mathrm {v e l}} \mathcal {L} _ {\mathrm {v e l}} (\theta), \tag {13} +$$ + +where $\lambda_{\mathrm{OT}}$ and $\lambda_{\mathrm{vel}}$ are the balancing coefficients. During training, we apply dropout to $w_{r}$ , $w_{e}$ , and $a^{1:L}$ with a probability of 0.1 for CFV. Additionally, we apply dropout to the preceding audio and motion latents with a probability 0.5 for smooth transition in the initial window. + +Inference During inference, we sample the generating vector field from noise $x_0$ , using the driving conditions $w_{r\rightarrow S}$ , $w_e$ , and $a^{1:L}$ , as well as the $L'$ frames of preceding audio and generated motion latents. + +We extend the CFV [10] to an incremental CFV to separately adjust the audio and emotion, inspired by [3]: + +$$ +\begin{array}{l} \tilde {v} _ {t} \approx v _ {t} (x _ {0}, \mathbf {c} _ {t} | _ {\{a ^ {1: L}, w _ {e} \}}) \\ + \gamma_ {a} \left[ v _ {t} \left(x _ {0}, \mathbf {c} _ {t} \mid_ {w _ {e}}\right) - v _ {t} \left(x _ {0}, \mathbf {c} _ {t} \mid_ {\{a ^ {1: L}, w _ {e} \}} \right] \right. \\ + \gamma_ {e} \left[ v _ {t} \left(x _ {0}, \mathbf {c} _ {t}\right) - v _ {t} \left(x _ {0}, \mathbf {c} _ {t} \mid_ {w _ {e}}\right) \right], \tag {14} \\ \end{array} +$$ + +where $\gamma_{a}$ and $\gamma_{e}$ are the guidance scales for audio and emotion, respectively. $\mathbf{c}_t|_{\{x,y\}}$ denotes the driving condition + +Table 1. Quantitative comparison results with state-of-the-art methods on HDTF [85] / RAVDESS [39]. The best result for each metric is in bold, and the second-best result is underlined. + +
MethodImage & Video GenerationLip Synchronization
FID ↓FVD ↓CSIM ↑E-FID ↓P-FID ↓LSE-D ↓LSE-C ↑
SadTalker† [84]71.952 / 119.430339.058 / 376.2940.644 / 0.6441.914 / 3.5001.456 / 2.0457.947 / 7.2737.305 / 4.748
EDTalk† [66]50.078 / 75.020211.284 / 304.9330.626 / 0.6761.579 / 3.4680.054 / 0.0908.123 / 7.6827.623 / 5.318
AniTalker† [36]39.512 / 70.430184.454 / 265.3410.643 / 0.7251.830 / 2.3300.092 / 0.1267.907 / 8.1767.288 / 4.555
Halo [79]25.363 / 57.648197.196 / 375.5570.869 / 0.8601.039 / 2.4920.037 / 0.0507.792 / 7.6137.582 / 4.795
EchoMimic [7]33.552 / 81.839296.757 / 320.2200.823 / 0.8051.234 / 3.2010.023 / 0.0478.903 / 8.1616.242 / 4.144
FLOAT (Ours)21.100 / 31.681162.052 / 166.3590.843 / 0.8101.229 / 1.3670.032 / 0.0317.290 / 6.9948.222 / 5.730
+ +![](images/daff56d026ccb964d277b4599dfc4f045c6db093de7e7c345df483f887a0ba5e.jpg) +Figure 5. Qualitative comparison results with state-of-the-art methods on HDTF [85] / RAVDESS [39]. Please refer to supplementary videos. Note that we additionally provide a video comparison with EMO [68] and VASA-1 [80] using their video demonstration. + +![](images/3255bf62b78bce64af2a45ed15e8d0a73102a854ff37627a755a8f9541ebd61a.jpg) + +![](images/1c5bddc5d113e1988b2eff754be5d5038bf803625fe43d6a11361efc5373dec3.jpg) + +![](images/f4481d782c00a7ecbdb51e51ffd279fa5a42c4221caa8dc70e74f38ff5c8d111.jpg) + +without the condition $x$ and $y$ . We set $\gamma_{a} = 2$ and $\gamma_{e} = 1$ based on the ablation studies on $\gamma_{a}$ and $\gamma_{e}$ provided in supplementary materials. + +After sampling, ODE solver receives the estimated vector field to compute the motion latents through numerical integration. We empirically find that FLOAT can generate reasonable motion with around 10 number of function evaluations (NFE). Please refer to supplementary videos. + +Lastly, we add the source identity latent to the generated motion latents and decode them into video frames using the motion latent decoder. + +# 5. Experiments + +# 5.1. Dataset and Pre-processing + +For training the motion latent auto-encoder, we use three open-source datasets: HDTF [85], RAVDESS [39], and VFHQ [78]. When training FLOAT, we exclude VFHQ because it does not support the synchronized audio. HDTF [85] is for high-definition talking face generation, containing videos of over 300 unique identities. RAVDESS [39] includes more than 2,400 emotion-intensive videos of 24 different identities. VFHQ [78] is designed for high-resolution video super-resolution and includes a large num + +ber of unique identities, which compensates the limited number of identities of the preceding datasets. Following the strategy of [58], we first convert each video to 25 FPS and resample the audio into $16\mathrm{kHz}$ . Then, we crop and resize the facial region to $512^2$ resolution [5]. After the pre-processing, for HDTF, we use a total of 11.3 hours of 240 videos featuring 230 different identities for training, and videos of 78 different identities, each 15 seconds long, for test. For RAVDESS, we use videos of 22 identities for training, and videos of the remaining 2 identities for test, with each 3-4 seconds long and representing 14 emotional intensities. Note that the identities in the training and test are disjoint in both datasets. + +# 5.2. Implementation Details + +The motion latent dimension is set to $d = 512$ with $M = 20$ distinct orthogonal directions. For the vector predictor, we use 8 attention heads, a hidden dimension $h = 1024$ , and an attention window length $T = 2$ . Considering the length of the training video clips, we set $L = 50$ frames with preceding $L' = 10$ frames at once, encompassing 2.4 seconds of video. We employ the Adam optimizer [30] with a batch size of 8 and a learning rate of $10^{-5}$ . We use $L1$ distance for the norm $\| \cdot \|$ in the training objective. We set the balancing coefficients to $\lambda_{\mathrm{OT}} = \lambda_{\mathrm{vel}} = 1$ . The entire training takes about 2 days for 2,000k steps on a single NVIDIA A100 GPU. We use Euler method [35] for the ODE solver. + +# 5.3. Evaluation + +Metrics and Baselines For evaluating the image and video generation quality, we measure Fréchet Inception Distance (FID) [57] and 16 frames Fréchet Video Distance (FVD) [69]. For facial identity, expression and head motion, we measure Cosine Similarity of identity embedding (CSIM) [11], Expression FID (E-FID) [68] and Pose FID (P-FID), respectively. Lastly, we measure Lip-Sync Error Distance and Confidence (LSE-D and LSE-C [51]) for audio-visual alignment. + +We compare our method with state-of-the-art audio-driven talking portrait methods whose official implementations are publicly available. For non-diffusion methods, we compare with SadTalker [84] and EDTalk [66]. For diffusion methods, we compare with AniTalker [36], Hallo [79], and EchoMimic [7]. + +Comparison Results In Tab. 1 and Fig. 5, we show the quantitative and qualitative comparison results, respectively. FLOAT outperforms other methods on most of the metrics and visual quality in both datasets. + +Additionally, we provide video comparison results with EMO [68] and VASA-1 [80] in the supplementary materials, using their demonstration videos due to the infeasibility of direct implementation. + +![](images/eaef4948f10feb64b30b606b91b0571f4844c966563af3497ed62e22a9a64d1f.jpg) +Figure 6. Test-time pose editing using $\lambda$ -control $(\lambda_{15}(\hat{D}) \pm 10)$ . + +# 5.4. Applications + +Test-time Pose Editing via Orthonormal Basis $V$ Since FLOAT learns the underlying motion latent structure, it is natural to assume that for any sampled motion latent $w_{r\rightarrow \hat{D}}$ , there exist motion coefficients $\{\lambda_m(\hat{D})\}_{m = 1}^M$ satisfying the representation in Eq. (9): $w_{r\rightarrow \hat{D}} = \sum_{m = 1}^{M}\lambda_{m}(\hat{D})\cdot \mathbf{v}_{m}$ . + +We can always compute these coefficients in closed form by taking inner products between the sampled motion $w_{r\rightarrow \hat{D}}$ and the learned orthonormal basis $V$ : + +$$ +\left\langle w _ {r \rightarrow \hat {D}}, \mathbf {v} _ {k} \right\rangle = \left\langle \sum_ {m = 1} ^ {M} \lambda_ {m} (\hat {D}) \cdot \mathbf {v} _ {m}, \mathbf {v} _ {k} \right\rangle = \lambda_ {k} (\hat {D}), \tag {15} +$$ + +where $\langle \mathbf{v}_m, \mathbf{v}_k \rangle = \delta_{m,k}$ and $\delta$ is Kronecker delta. At this point, we can edit the sampled motions by editing the corresponding coefficients (e.g., via linear operation) and combining them back into the motion latent. As shown in Fig. 6, it allows us to control head direction without interfering with other motions due to the orthogonality of the basis. We refer to this test-time editing technique as $\lambda$ -control. + +Additional Driving Signals In Fig. 7 and Tab. 2, we experiment with additional driving conditions, head poses and image-driven emotion labels, to explore additional controllability in our method. We employ 3DMM head pose parameters $p \in \mathbb{R}^6$ [2] extracted by [12]. We concatenate a sequence of pose parameters $p^{1:L} \in \mathbb{R}^{L \times 6}$ with the other driving conditions, and then map them to $c_t^{1:L} \in \mathbb{R}^{L \times h}$ . We also experiment on image-driven emotion [56] for frame-wise emotion control rather than the long-term emotion enhancement. FLOAT can effectively accommodate these additional conditions, highlighting its flexibility across diverse control signals. + +Redirecting Speech-driven Emotion Since FLOAT learns diverse emotions in the emotion-intensive data distribution [39], the generated emotion-aware motion can be modified by redirecting the speech-driven emotion label toward a different emotion at inference time. As illustrated in Fig. 8, this technique is particularly beneficial for manual redirection when the emotion predicted from speech is complex or ambiguous. + +![](images/6ca3a17d0829e37887e40238950c7f077f253cdc06d11a9d01aa8966bbe9dd95.jpg) +Figure 7. Additional conditioning results of FLOAT. 3Dpose, $S2E$ , and $I2E$ denote 3D head pose parameters [12], speech-to-emotion [49], and image-to-emotion [56], respectively. + +![](images/af02109081a547b53339d991fc0888c4352bbee856a7594687667110672402fa.jpg) + +Table 2. Quantitative results of FLOAT with additional conditions (HDTF [85] / RAVDESS [39]). $S2E$ , $I2E$ , and 3DPose denote speech-to-emotion [49], image-to-emotion [56], and 3DMM pose parameters [12], respectively. + +
ConfigurationsFID ↓FVD ↓E-FID ↓P-FID ↓LSE-D ↓
AFLOAT(Ours)21.100 / 31.681162.052 / 166.3591.229 / 1.3670.032 / 0.0317.290 / 6.994
BA + 3Dpose19.721 / 29.721126.663 / 112.8940.926 / 1.1520.012 / 0.0167.516 / 7.047
CA - S2E21.235 / 32.035155.032 / 166.8661.254 / 1.5020.031 / 0.0257.264 / 7.222
DA - S2E + I2E21/528 / 31.609158.577 / 162.3691.158 / 1.3050.034 / 0.0227.183 / 7.150
+ +![](images/7e0ea533f52484b01bbefc6180adff4b2c616f773c8e0a9eb1c83bcb066f4fa5.jpg) +Figure 8. Redirecting the unclear emotion prediction to a desirable one-hot encoding, which can be further intensified by the CFV. + +# 5.5. Ablation Studies + +Ablation on Frame-wise AdaLN We compare frame-wise AdaLN (and gating) followed by masked self-attention to separate conditioning from attending, with a cross-attention that performs conditioning and attending simultaneously. As shown in Tab. 3, both approaches achieve competitive image and video quality, while frame-wise AdaLN provides better expression generation and lip synchronization. We observe that frame-wise AdaLN can achieve more diverse head motions than the cross-attention. Please refer to supplementary videos. + +Ablation on Flow Matching We compare flow matching with two types of diffusion models: $\epsilon$ -prediction (noise) and $x_0$ -prediction (signal) [52, 67]. In both cases, we adopt our vector predictor architecture as denoising networks. We adopt diffusion training settings of VASA-1 [80] (500 diffusion steps with a cosine noise scheduler [46] and 50 DDIM denoising steps) for the indirect comparison with [80]. Notably, diffusion and flow matching achieve competitive results on image quality while the latter achieves the better lip synchronization. In Fig. 9, we compare the forward pass efficiency by measuring frames per second (FPS) of each + +Table 3. Ablation studies of FLOAT on HDTF [85]. The best result for each metric is in bold, and the second-best result is underlined. + +
MethodFID ↓FVD ↓E-FID ↓LSE-D ↓#NFEs ↓
Ours (w. Cross-Attn.)21.873162.7021.4527.75710
Ours (w. Diff., c-pred.)21.190161.6661.2139.92250
Ours (w. Diff., x0-pred.)21.697162.8471.2789.04850
FLOAT (Ours)21.100162.0521.2297.29010
+ +![](images/41c53dddeebfba3786224a8801316f9ace4e5d7e82874f9a46e3782adc4c2a59.jpg) +Figure 9. Comparison of the forward pass efficiency. We compute FPS on a single NVIDIA V100 GPU. + +model. Thanks to the compact motion latent representation and OT-based flow matching, FLOAT achieves the highest FPS, superior lip-sync performance, dynamic head motion, and the lowest NFEs. + +# 6. Conclusion + +We proposed FLOAT, a flow matching based audio-driven talking portrait generation model leveraging a learned motion latent space. We introduced a transformer-based vector field predictor, enabling temporally consistent motion generation. Additionally, we incorporated speech-driven emotion labels into the motion sampling process to improve the naturalness of the audio-driven talking motions. FLOAT addresses current core limitations of diffusion-based talking portrait video generation methods by reducing the sampling time through flow matching while achieving the remarkable sample quality. Extensive experiments verified that FLOAT achieves state-of-the-art performance in terms of visual quality, motion fidelity, and efficiency. + +Discussion We leave further discussion considering limitations, future work, and ethical considerations in the supplementary materials. + +# References + +[1] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449-12460, 2020. 5 +[2] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques, pages 187-194, 1999. 3, 7 +[3] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18392-18402, 2023. 5 +[4] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators, 2024. 2 +[5] Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2d & 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In International Conference on Computer Vision, 2017. 7 +[6] Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations. Advances in neural information processing systems, 31, 2018. 2, 3 +[7] Zhiyuan Chen, Jiajiong Cao, Zhiquan Chen, Yuming Li, and Chenguang Ma. Echomimic: Lifelike audio-driven portrait animations through editable landmark conditions. arXiv preprint arXiv:2407.08136, 2024. 1, 2, 3, 4, 6, 7 +[8] Kun Cheng, Xiaodong Cun, Yong Zhang, Menghan Xia, Fei Yin, Mingrui Zhu, Xuan Wang, Jue Wang, and Nannan Wang. Videoretalking: Audio-based lip synchronization for talking head video editing in the wild. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 1, 2 +[9] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In Asian Conference on Computer Vision, pages 251-263, 2016. 1, 2 +[10] Quan Dao, Hao Phung, Binh Nguyen, and Anh Tran. Flow matching in latent space. arXiv preprint arXiv:2307.08698, 2023. 2, 3, 5 +[11] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4690-4699, 2019. 7 +[12] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 7, 8 +[13] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2 +[14] Laurent Dinh, Jascha Sohl-Dickstein, and Samy Ben + +gio. Density estimation using real nvp. arXiv preprint arXiv:1605.08803, 2016. 2 +[15] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2663-2671, 2022. 3 +[16] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 2 +[17] Johannes S Fischer, Ming Gui, Pingchuan Ma, Nick Stracke, Stefan A Baumann, and Björn Ommer. Boosting latent diffusion with flow matching. arXiv preprint arXiv:2312.07360, 2023. 2 +[18] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2 +[19] Jiazhi Guan, Zhanwang Zhang, Hang Zhou, Tianshu Hu, Kaisiyuan Wang, Dongliang He, Haocheng Feng, Jingtuo Liu, Errui Ding, Ziwei Liu, et al. Stylesync: High-fidelity generalized and personalized lip sync in style-based generator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1505-1515, 2023. 1, 2 +[20] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 2, 3 +[21] Tianyu He, Junliang Guo, Runyi Yu, Yuchi Wang, Jialiang Zhu, Kaikai An, Leyi Li, Xu Tan, Chunyu Wang, Han Hu, et al. Gaia: Zero-shot talking avatar generation. arXiv preprint arXiv:2311.15230, 2023. 1, 2 +[22] Geoffrey Hinton. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2 +[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2 +[24] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8153-8163, 2024. 2, 3, 4 +[25] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Qianyi Wu, Wayne Wu, Feng Xu, and Xun Cao. Eamm: One-shot emotional talking face via audio-based emotion-aware motion model. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2, 3, 5 +[26] Jianwen Jiang, Chao Liang, Jiaqi Yang, Gaojie Lin, Tianyun Zhong, and Yanbo Zheng. Loopy: Taming audio-driven portrait avatar with long-term motion dependency. arXiv preprint arXiv:2409.02634, 2024. 1, 2, 3 +[27] Minguk Kang, Richard Zhang, Connelly Barnes, Sylvain Paris, Suha Kwak, Jaesik Park, Eli Shechtman, Jun-Yan Zhu, + +and Taesung Park. Distilling diffusion models into conditional gans. arXiv preprint arXiv:2405.05967, 2024. 2 +[28] Taekyung Ki and Dongchan Min. Stylelipsync: Style-based personalized lip-sync video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22841-22850, 2023. 1, 2 +[29] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 1, 2 +[30] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 7 +[31] Tobias Kirschstein, Simon Giebenhain, and Matthias Nießner. Diffusionavatars: Deferred diffusion for high-fidelity 3d head avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5492, 2024. 2 +[32] Matthew Le, Apoorv Vyas, Bowen Shi, Brian Karrer, Leda Sari, Rashel Moritz, Mary Williamson, Vimal Manohar, Yossi Adi, Jay Mahadeokar, et al. Voicebox: Text-guided multilingual universal speech generation at scale. Advances in neural information processing systems, 36, 2024. 2 +[33] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. ArXiv e-prints, pages arXiv-1607, 2016. 5 +[34] Senmao Li, Taihang Hu, Fahad Shahbaz Khan, Linxuan Li, Shiqi Yang, Yaxing Wang, Ming-Ming Cheng, and Jian Yang. Faster diffusion: Rethinking the role of unet encoder in diffusion models. arXiv preprint arXiv:2312.09608, 2023. 2 +[35] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022. 2, 3, 4, 7 +[36] Tao Liu, Feilong Chen, Shuai Fan, Chenpeng Du, Qi Chen, Xie Chen, and Kai Yu. Anitalker: Animate vivid and diverse talking faces through identity-decoupled facial motion encoding. arXiv preprint arXiv:2405.03121, 2024. 1, 3, 6, 7 +[37] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 2, 4 +[38] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, et al. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. In The Twelfth International Conference on Learning Representations, 2023. 2 +[39] Steven R Livingstone and Frank A Russo. The ryerson audio-visual database of emotional speech and song (ravdess): A dynamic, multimodal set of facial and vocal expressions in north american english. PloS one, 13(5): e0196391, 2018. 6, 7, 8 +[40] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 2 +[41] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 2 + +[42] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 2 +[43] Yifeng Ma, Suzhen Wang, Zhipeng Hu, Changjie Fan, Tangjie Lv, Yu Ding, Zhidong Deng, and Xin Yu. Styletalk: One-shot talking head generation with controllable speaking styles. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1896-1904, 2023. 2 +[44] Yifeng Ma, Shiwei Zhang, Jiayu Wang, Xiang Wang, Yingya Zhang, and Zhidong Deng. Dreamtalk: When expressive talking head generation meets diffusion probabilistic models. arXiv preprint arXiv:2312.09767, 2023. 2, 3 +[45] Dongchan Min, Minyoung Song, Eunji Ko, and Sung Ju Hwang. Stylistalker: One-shot style-based audio-driven talking head video generation. arXiv preprint arXiv:2208.10922, 2022. 1, 2 +[46] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162-8171. PMLR, 2021. 2, 8 +[47] Se Jin Park, Minsu Kim, Joanna Hong, Jeongsoo Choi, and Yong Man Ro. Synctalkface: Talking face generation with precise lip-syncing via audio-lip memory. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2062-2070, 2022. 1 +[48] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 2, 4 +[49] Leonardo Pepino, Pablo Riera, and Luciana Ferrer. Emotion recognition from speech using wav2vec 2.0 embeddings. arXiv preprint arXiv:2104.03502, 2021. 5, 8 +[50] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models. arXiv preprint arXiv:2410.13720, 2024. 2 +[51] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020. 1, 2, 7 +[52] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2, 8 +[53] Danilo Rezende and Shakir Mohamed. Variational inference with normalizing flows. In International conference on machine learning, pages 1530-1538. PMLR, 2015. 1, 2 +[54] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 2, 3, 4 +[55] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, + +Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2 +[56] Andrey V Savchenko. Hsemotion: High-speed emotion recognition library. Software Impacts, 14:100433, 2022. 5, 7, 8 +[57] Maximilian Seitzer. pytorch-fid: FID Score for PyTorch. https://github.com/mseitzer/pytorch-fid, 2020. Version 0.3.0.7 +[58] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. Advances in neural information processing systems, 32, 2019. 7 +[59] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 4 +[60] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2 +[61] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2 +[62] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arXiv preprint arXiv:2303.01469, 2023. 2 +[63] Michal Stypulkowski, Konstantinos Vougioukas, Sen He, Maciej Zięba, Stavros Petridis, and Maja Pantic. Diffused heads: Diffusion models beat gans on talking-face generation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5091-5100, 2024. 2 +[64] Zhiyao Sun, Tian Lv, Sheng Ye, Matthieu Lin, Jenny Sheng, Yu-Hui Wen, Minjing Yu, and Yong-jin Liu. Diffpositalk: Speech-driven stylistic 3d facial animation and head pose generation via diffusion models. ACM Transactions on Graphics (TOG), 43(4):1-9, 2024. 2 +[65] Shuai Tan, Bin Ji, and Ye Pan. Emmn: Emotional motion memory network for audio-driven emotional talking face generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22146-22156, 2023. 2, 3 +[66] Shuai Tan, Bin Ji, Mengxiao Bi, and Ye Pan. Edtalk: Efficient disentanglement for emotional talking head synthesis. In European Conference on Computer Vision, pages 398-416. Springer, 2025. 1, 6, 7 +[67] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations, 2023. 2, 5, 8 +[68] Linrui Tian, Qi Wang, Bang Zhang, and Liefeng Bo. Emo: Emote portrait alive-generating expressive portrait videos with audio2video diffusion model under weak conditions. arXiv preprint arXiv:2402.17485, 2024. 2, 3, 4, 6, 7 +[69] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 7 + +[70] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2, 4, 5 +[71] Cong Wang, Kuan Tian, Jun Zhang, Yonghang Guan, Feng Luo, Fei Shen, Zhiwei Jiang, Qing Gu, Xiao Han, and Wei Yang. V-express: Conditional dropout for progressive training of portrait video generation. arXiv preprint arXiv:2406.02511, 2024. 2, 3 +[72] Kaisiyuan Wang, Qianyi Wu, Linsen Song, Zhuoqian Yang, Wayne Wu, Chen Qian, Ran He, Yu Qiao, and Chen Change Loy. Mead: A large-scale audio-visual dataset for emotional talking-face generation. In European Conference on Computer Vision, pages 700–717. Springer, 2020. 3, 5 +[73] Suzhen Wang, Lincheng Li, Yu Ding, Changjie Fan, and Xin Yu. Audio2head: Audio-driven one-shot talking-head generation with natural head motion. arXiv preprint arXiv:2107.09293, 2021. 2 +[74] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10039-10049, 2021. 3 +[75] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. Latent image animator: Learning to animate images via latent space navigation. arXiv preprint arXiv:2203.09043, 2022. 2, 4 +[76] Huawei Wei, Zejun Yang, and Zhisheng Wang. Aniportrait: Audio-driven synthesis of photorealistic portrait animation. arXiv preprint arXiv:2403.17694, 2024. 2, 3 +[77] Yibo Xia, Lizhen Wang, Xiang Deng, Xiaoyan Luo, and Yebin Liu. Gmtalker: Gaussian mixture based emotional talking video portraits. arXiv preprint arXiv:2312.07669, 2023. 3 +[78] Liangbin Xie, Xintao Wang, Honglun Zhang, Chao Dong, and Ying Shan. Vfhq: A high-quality dataset and benchmark for video face super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 657-666, 2022. 6 +[79] Mingwang Xu, Hui Li, Qingkun Su, Hanlin Shang, Liwei Zhang, Ce Liu, Jingdong Wang, Luc Van Gool, Yao Yao, and Siyu Zhu. Hallo: Hierarchical audio-driven visual synthesis for portrait image animation. arXiv preprint arXiv:2406.08801, 2024. 2, 3, 4, 6, 7 +[80] Sicheng Xu, Guojun Chen, Yu-Xiao Guo, Jiaolong Yang, Chong Li, Zhenyu Zang, Yizhong Zhang, Xin Tong, and Baining Guo. Vasa-1: Lifelike audio-driven talking faces generated in real time. arXiv preprint arXiv:2404.10667, 2024. 2, 3, 5, 6, 7, 8 +[81] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. In European conference on computer vision, pages 85-101. Springer, 2022. 2 +[82] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2 + +[83] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 4 +[84] Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8652-8661, 2023. 1, 2, 6, 7 +[85] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3661–3670, 2021. 6, 8 +[86] Zhimeng Zhang, Zhipeng Hu, Wenjin Deng, Changjie Fan, Tangjie Lv, and Yu Ding. Dinet: Deformation inpainting network for realistic face visually dubbing on high resolution video. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3543-3551, 2023. 1 +[87] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4176–4186, 2021. 2 +[88] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. Makelttalk: speaker-aware talking-head animation. ACM Transactions On Graphics (TOG), 39(6):1-15, 2020. 2 +[89] Shenhao Zhu, Junming Leo Chen, Zuozhuo Dai, Yinghui Xu, Xun Cao, Yao Yao, Hao Zhu, and Siyu Zhu. Champ: Controllable and consistent human image animation with 3d parametric guidance. arXiv preprint arXiv:2403.14781, 2024. 4 \ No newline at end of file diff --git a/2025/FLOAT_ Generative Motion Latent Flow Matching for Audio-driven Talking Portrait/images.zip b/2025/FLOAT_ Generative Motion Latent Flow Matching for Audio-driven Talking Portrait/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..22894b08ff8a1e7562b80936930be5f8ef576340 --- /dev/null +++ b/2025/FLOAT_ Generative Motion Latent Flow Matching for Audio-driven Talking Portrait/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5ddfae5469aecea37c7049addb6ec3d8184468251a66f58d459a1874653cc82e +size 743883 diff --git a/2025/FLOAT_ Generative Motion Latent Flow Matching for Audio-driven Talking Portrait/layout.json b/2025/FLOAT_ Generative Motion Latent Flow Matching for Audio-driven Talking Portrait/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..92679fd8b055e1dcc8e6747c6ea209e1513d76b0 --- /dev/null +++ b/2025/FLOAT_ Generative Motion Latent Flow Matching for Audio-driven Talking Portrait/layout.json @@ -0,0 +1,11927 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 70, + 102, + 541, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 102, + 541, + 138 + ], + "spans": [ + { + "bbox": [ + 70, + 102, + 541, + 138 + ], + "type": "text", + "content": "FLOAT: Generative Motion Latent Flow Matching for Audio-driven Talking Portrait" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 164, + 160, + 443, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 160, + 443, + 175 + ], + "spans": [ + { + "bbox": [ + 164, + 160, + 443, + 175 + ], + "type": "text", + "content": "Taekyung " + }, + { + "bbox": [ + 164, + 160, + 443, + 175 + ], + "type": "inline_equation", + "content": "\\mathbf{K}\\mathrm{i}^{1\\star}" + }, + { + "bbox": [ + 164, + 160, + 443, + 175 + ], + "type": "text", + "content": " Dongchan Min Gyeongsu Chae" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 239, + 176, + 372, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 176, + 372, + 190 + ], + "spans": [ + { + "bbox": [ + 239, + 176, + 372, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 239, + 176, + 372, + 190 + ], + "type": "text", + "content": "KAIST " + }, + { + "bbox": [ + 239, + 176, + 372, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 239, + 176, + 372, + 190 + ], + "type": "text", + "content": "DeepBrain AI Inc." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 155, + 191, + 456, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 191, + 456, + 203 + ], + "spans": [ + { + "bbox": [ + 155, + 191, + 456, + 203 + ], + "type": "text", + "content": "{taekyung.ki, alsehdcks95}@kaist.ac.kr gc@deepbrain.io" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 181, + 205, + 426, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 205, + 426, + 217 + ], + "spans": [ + { + "bbox": [ + 181, + 205, + 426, + 217 + ], + "type": "text", + "content": "https://deepbrainai-research.github.io/float/" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 99, + 243, + 509, + 395 + ], + "blocks": [ + { + "bbox": [ + 99, + 243, + 509, + 395 + ], + "lines": [ + { + "bbox": [ + 99, + 243, + 509, + 395 + ], + "spans": [ + { + "bbox": [ + 99, + 243, + 509, + 395 + ], + "type": "image", + "image_path": "426ecc2dd395074757380a5c7cc5bc5fe8afac3b04755b79925b975e9e7bc694.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 399, + 555, + 433 + ], + "lines": [ + { + "bbox": [ + 54, + 399, + 555, + 433 + ], + "spans": [ + { + "bbox": [ + 54, + 399, + 555, + 433 + ], + "type": "text", + "content": "Figure 1. FLOAT can generate a talking portrait video from a single source image and audio where the talking motion is generated by the motion latent flow matching. It can enhance the emotion-related talking motion by leveraging speech-driven emotion labels, a natural way of emotion-aware motion control." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 443, + 200, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 443, + 200, + 455 + ], + "spans": [ + { + "bbox": [ + 151, + 443, + 200, + 455 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 468, + 297, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 468, + 297, + 673 + ], + "spans": [ + { + "bbox": [ + 53, + 468, + 297, + 673 + ], + "type": "text", + "content": "With the rapid advancement of diffusion-based generative models, portrait image animation has achieved remarkable results. However, it still faces challenges in temporally consistent video generation and fast sampling due to its iterative sampling nature. This paper presents FLOAT, an audio-driven talking portrait video generation method based on flow matching generative model. Instead of a pixel-based latent space, we take advantage of a learned orthogonal motion latent space, enabling efficient generation and editing of temporally consistent motion. To achieve this, we introduce a transformer-based vector field predictor with an effective frame-wise conditioning mechanism. Additionally, our method supports speech-driven emotion enhancement, enabling a natural incorporation of expressive motions. Extensive experiments demonstrate that our method outperforms state-of-the-art audio-driven talking portrait methods in terms of visual quality, motion fidelity, and efficiency." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 443, + 394, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 443, + 394, + 455 + ], + "spans": [ + { + "bbox": [ + 315, + 443, + 394, + 455 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 467, + 555, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 555, + 624 + ], + "type": "text", + "content": "Animating a single image using a driving audio (i.e., audiodriven talking portrait generation) has gained significant attention in recent years for its great potential in avatar creation, video conferencing, virtual avatar chat, and user-friendly customer service. It aims to synthesize natural talking motion from audio signals, including accurate lip synchronization, rhythmical head movements, and fine-grained facial expressions. However, generating such motion solely from audio is extremely challenging due to its one-to-many correlation between audio and motion. In the earlier stage of this field, many works [8, 19, 28, 47, 51, 86] focus on generating accurate lip movements by relying on learned audio-lip alignment losses [9, 45]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 626, + 556, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 626, + 556, + 698 + ], + "spans": [ + { + "bbox": [ + 313, + 626, + 556, + 698 + ], + "type": "text", + "content": "To comprehensively extend the range of motion, some works [45, 66, 84] incorporate probabilistic generative models, such as VAE [29] and normalizing flow [53], turning the motion generation into probabilistic sampling. However, these models still lack expressiveness in generated motion due to the limited capacity of these generative models." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "type": "text", + "content": "Recent talking portrait generation methods [7, 21, 26, 36," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "*This work was done during South Korea Mandatory Military Service at DeepBrain AI Inc." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14699" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": "44, 63, 68, 71, 76, 79], powered by diffusion-based generative models [23, 61], successfully mitigate this expressiveness issue. EMO [68] introduces a promising approach to this field [7, 26, 71, 76, 79] by employing a strong pretrained image diffusion model (i.e., StableDiffusion [54]) and lifting it into video generation [24]. However, it still faces challenges in generating temporally coherent videos and achieving sampling efficiency, requiring tens of minutes for a few seconds of video. Moreover, they heavily rely on auxiliary facial prior, such as bounding boxes [68, 79], 2D landmarks and skeletons [7, 26, 82], or 3D meshes [76], which significantly restricts the diversity and the fidelity of head movements due to their strong spatial bias." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 228, + 294, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 294, + 407 + ], + "type": "text", + "content": "In this paper, we present FLOAT, an audio-driven talking portrait video generation model based on flow matching generative model in a motion latent space. Flow matching [35, 37] has emerged as a promising alternative to diffusion models due to its fast and high-quality sampling. By modeling talking motion within a learned motion latent space [75], we can more efficiently sample temporally consistent motion latents. This is achieved by a simple yet effective transformer-based [70] vector field predictor, inspired by DiT [48]. Since our motion latent space has orthogonal structure, our method can manipulate head motion of the generated video using its basis. Furthermore, our method supports natural emotion-aware motion enhancement driven by speech. Our contributions are summarized as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 409, + 294, + 538 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 56, + 409, + 294, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 409, + 294, + 456 + ], + "spans": [ + { + "bbox": [ + 56, + 409, + 294, + 456 + ], + "type": "text", + "content": "- We present, FLOAT, flow matching based audio-driven talking portrait generation model using a learned orthogonal motion latent space, enabling to generate talking portrait videos with reduced sampling steps." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 457, + 293, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 457, + 293, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 457, + 293, + 502 + ], + "type": "text", + "content": "- We introduce a simple yet effective transformer-based flow vector field predictor for temporally consistent motion latent sampling, which also enables the speech-driven emotional controls." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 505, + 293, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 505, + 293, + 538 + ], + "spans": [ + { + "bbox": [ + 56, + 505, + 293, + 538 + ], + "type": "text", + "content": "- Extensive experiments demonstrate that FLOAT achieves state-of-the-art performance compared to both diffusion- and non-diffusion-based methods." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 554, + 146, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 554, + 146, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 554, + 146, + 566 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 574, + 253, + 587 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 253, + 587 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 253, + 587 + ], + "type": "text", + "content": "2.1. Diffusion Models and Flow Matching" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 593, + 294, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 593, + 294, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 593, + 294, + 700 + ], + "type": "text", + "content": "Diffusion Models Diffusion models or score-based generative models [13, 23, 46, 54, 60, 61] are generative models that gradually diffuse input signals into Gaussian noise and learn the denoising reverse process for the generative modeling. They have shown remarkable results in various generation tasks, such as unconditional image and video generation [4, 16, 48], text-to-image generation [52, 54, 55], text-to-video generation [4, 20], conditional image generation [24, 82], and 3D human generation [31, 64, 67]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 701, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 294, + 713 + ], + "type": "text", + "content": "Accelerating Diffusion Models While diffusion models" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "type": "text", + "content": "demonstrate superior performance, their iterative sampling nature still bottlenecks the efficient generation compared to VAEs [29], normalizing flow [53], and GANs [18]. To overcome this limitation, several works have been developed to boost the sampling speed of the diffusion models. StableDiffusion (SD) [54] partially mitigates this problem by moving the diffusion process from the pixel space to the spatial latent space, establishing itself as a pivotal framework among diffusion models. Another line of research has developed the sampling solvers [40, 41] based on ordinary differential equations (ODEs). Meanwhile, model distillation [22] has been introduced to transfer the knowledge of the learned diffusion models into a student model, enabling one (or a few) steps of generation [27, 34, 38, 42, 62]. However, these approaches involve substantial effort to create a well-trained diffusion model and suffer from training instability." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 265, + 553, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 265, + 553, + 444 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 553, + 444 + ], + "type": "text", + "content": "Flow Matching Flow matching [35, 37] stands out as an alternative to diffusion models for its high sampling speed and competitive sample quality compared to diffusion models [10, 17, 32, 35, 50]. It belongs to the family of flow-based generative models, which estimates a transformation (referred to as a flow) between a prior distribution (e.g., Gaussian) and a target distribution. Unlike the normalizing flow [14, 53] that directly estimates the noise-to-data transformation under specific architectural constraints (e.g., affine coupling), flow matching regresses the time-dependent vector field that generates this flow by solving its corresponding ODEs [6] with flexible architectures. One specific design of flow matching is an optimal transport (OT) based one, which transforms the data distribution along the straight path with constant velocity [35]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 445, + 553, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 553, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 553, + 516 + ], + "type": "text", + "content": "Our audio-driven talking portrait method employs flow matching to generate the natural talking motions. Thanks to the architectural flexibility of flow matching, we use transformer-encoder architecture [70] to estimate the generating vector field, allowing us to take the video temporal consistency into account." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 528, + 492, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 492, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 492, + 539 + ], + "type": "text", + "content": "2.2. Audio-driven Portrait Animation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 546, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 553, + 713 + ], + "type": "text", + "content": "Audio-driven portrait animation is the task of generating a realistic talking portrait video using a single portrait image and driving audio [45, 73, 84, 87, 88]. Since audio-to-motion relation is basically a one-to-many problem, several works utilize additional facial prior for driving conditions, e.g., 2D facial landmarks [7, 21, 26, 71, 76, 88], 3D prior [8, 43, 44, 81, 84], or emotional labels [25, 65, 80]. In earlier stages, most works [8, 19, 28, 51] focused on generating accurate lip motion from audio by utilizing the lip-sync discriminator [9]. These approaches have advanced to generating audio-related head poses in a probabilistic way. For example, StyleTalker [45] uses normalizing flow [14, 53] to generate the head motion from audio, while SadTalker [84] uses audio-conditional variational inference [29] to learn" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14700" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "the 3DMM coefficients [2], bridging the intermediate representations of a pre-trained portrait animator [74]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 295, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 295, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 295, + 156 + ], + "type": "text", + "content": "Meanwhile, several works [25, 65, 72, 77] focus on an emotion-aware talking portrait generation. In particular, EAMM [25] considers an emotion as the complementary displacement of facial motion, and learns these displacement from an emotion label extracted from the image." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 156, + 296, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 156, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 296, + 300 + ], + "type": "text", + "content": "Recent audio-driven talking portrait methods powered by diffusion models show remarkable results [7, 26, 36, 44, 68, 71, 76, 79, 80]. Specifically, EMO [68] and subsequent extensions [7, 71, 76, 79] utilize the pre-trained SD [54] as their backbone to leverage generative prior trained on the large-scale image datasets. They introduce additional modules, e.g., ReferenceNet [24] and Temporal Transformer [20], to preserve input identity and enhance the video temporal consistency, respectively. However, these modules introduces additional computational cost, requiring several minutes for a few seconds of video, and still suffer from video-level artifacts, such as noisy frames, and flickering." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "text", + "content": "VASA-1 [80] addresses the sampling time issue by sampling motion latents [15], producing lifelike talking portraits. Our method takes advantage of this approach. However, unlike [80], our motion latent space has a strong linear orthogonal structure represented by a computable basis, enabling to manipulate the generated motion at the test-time without external driving signals. Based on this orthogonality, we employ OT-based flow matching for motion latent sampling along a straight line with reduced sampling steps." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 416, + 296, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 416, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 416, + 296, + 430 + ], + "type": "text", + "content": "3. Preliminaries: (Conditional) Flow Matching" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 436, + 296, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 436, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 54, + 436, + 296, + 533 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 54, + 436, + 296, + 533 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 54, + 436, + 296, + 533 + ], + "type": "text", + "content": " be a data, " + }, + { + "bbox": [ + 54, + 436, + 296, + 533 + ], + "type": "inline_equation", + "content": "t \\in [0,1]" + }, + { + "bbox": [ + 54, + 436, + 296, + 533 + ], + "type": "text", + "content": " be the time, and " + }, + { + "bbox": [ + 54, + 436, + 296, + 533 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 54, + 436, + 296, + 533 + ], + "type": "text", + "content": " be a unknown target distribution. We can define a flow as a time-dependent transformation " + }, + { + "bbox": [ + 54, + 436, + 296, + 533 + ], + "type": "inline_equation", + "content": "\\varphi_t : [0,1] \\times \\mathbb{R}^d \\to \\mathbb{R}^d" + }, + { + "bbox": [ + 54, + 436, + 296, + 533 + ], + "type": "text", + "content": " that transforms a tractable prior distribution " + }, + { + "bbox": [ + 54, + 436, + 296, + 533 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 54, + 436, + 296, + 533 + ], + "type": "text", + "content": " to the distribution " + }, + { + "bbox": [ + 54, + 436, + 296, + 533 + ], + "type": "inline_equation", + "content": "p_1 \\approx q" + }, + { + "bbox": [ + 54, + 436, + 296, + 533 + ], + "type": "text", + "content": ". This flow " + }, + { + "bbox": [ + 54, + 436, + 296, + 533 + ], + "type": "inline_equation", + "content": "\\varphi_t" + }, + { + "bbox": [ + 54, + 436, + 296, + 533 + ], + "type": "text", + "content": " further introduces a probability flow path " + }, + { + "bbox": [ + 54, + 436, + 296, + 533 + ], + "type": "inline_equation", + "content": "p_t : [0,1] \\times \\mathbb{R}^d \\to \\mathbb{R}_{>0}" + }, + { + "bbox": [ + 54, + 436, + 296, + 533 + ], + "type": "text", + "content": " and a generating vector field " + }, + { + "bbox": [ + 54, + 436, + 296, + 533 + ], + "type": "inline_equation", + "content": "v_t : [0,1] \\times \\mathbb{R}^d \\to \\mathbb{R}^d" + }, + { + "bbox": [ + 54, + 436, + 296, + 533 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 54, + 436, + 296, + 533 + ], + "type": "inline_equation", + "content": "p_t" + }, + { + "bbox": [ + 54, + 436, + 296, + 533 + ], + "type": "text", + "content": " is defined by the push-forwarding" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 97, + 537, + 295, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 537, + 295, + 564 + ], + "spans": [ + { + "bbox": [ + 97, + 537, + 295, + 564 + ], + "type": "interline_equation", + "content": "p _ {t} (x) = p _ {0} \\left(\\varphi_ {t} ^ {- 1} (x)\\right) \\det \\left| \\frac {\\partial \\varphi_ {t} ^ {- 1} (x)}{\\partial x} \\right|, \\tag {1}", + "image_path": "9c059b842235991bd02a17df5a1b2d8f9f4ae30504a7075dc46038daa9fb8900.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 571, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 571, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 295, + 594 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 55, + 571, + 295, + 594 + ], + "type": "inline_equation", + "content": "v_{t}" + }, + { + "bbox": [ + 55, + 571, + 295, + 594 + ], + "type": "text", + "content": " generates " + }, + { + "bbox": [ + 55, + 571, + 295, + 594 + ], + "type": "inline_equation", + "content": "\\varphi_{t}" + }, + { + "bbox": [ + 55, + 571, + 295, + 594 + ], + "type": "text", + "content": " by means of an ordinary differential equation (ODE) [6]:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 90, + 600, + 295, + 623 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 600, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 90, + 600, + 295, + 623 + ], + "type": "interline_equation", + "content": "\\frac {d}{d t} \\varphi_ {t} (x) = v _ {t} \\left(\\varphi_ {t} (x)\\right) \\quad \\text {a n d} \\quad \\varphi_ {0} (x) = x. \\tag {2}", + "image_path": "102ab355edee3256496a4b24226db60840cf779e76905a3d952a01e954b88422.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 628, + 296, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 296, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 296, + 652 + ], + "type": "text", + "content": "Flow matching [35] aims to estimate the target generating vector field " + }, + { + "bbox": [ + 55, + 628, + 296, + 652 + ], + "type": "inline_equation", + "content": "u_{t}" + }, + { + "bbox": [ + 55, + 628, + 296, + 652 + ], + "type": "text", + "content": " with a neural network parameterized by " + }, + { + "bbox": [ + 55, + 628, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 628, + 296, + 652 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 108, + 669, + 295, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 669, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 108, + 669, + 295, + 684 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {F M}} (\\theta) := \\left\\| v _ {t} (x; \\theta) - u _ {t} (x) \\right\\| _ {2} ^ {2}, \\tag {3}", + "image_path": "05d156bef76c4712ae91e04a434e6a8c1091e9cb574790428fdc68066bdf2f36.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "t \\sim \\mathcal{U}[0,1]" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "x \\sim p_t(x)" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": ". However, the target generating vector field " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "u_t" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " and the sample distribution " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "p_t" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " are" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "intractable. To address this issue, [35] proposes a method for constructing a \"conditional\" probability path " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "p_t(\\cdot | x_1)" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " as well as target \"conditional\" vector field " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "u_t(\\cdot | x_1)" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " using a sample " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "x_1 \\sim q" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " as a condition. And they prove that the following objective" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 359, + 138, + 553, + 152 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 138, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 359, + 138, + 553, + 152 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C F M}} (\\theta) := \\left\\| v _ {t} (x; \\theta) - u _ {t} (x \\mid x _ {1}) \\right\\| _ {2} ^ {2}, \\tag {4}", + "image_path": "607f9413d9d0f11bded532cc8b25b85dac9766ac8ffecde2f350001c264e9fd8.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 159, + 553, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 159, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 313, + 159, + 553, + 182 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 159, + 553, + 182 + ], + "type": "inline_equation", + "content": "t\\sim \\mathcal{U}[0,1]" + }, + { + "bbox": [ + 313, + 159, + 553, + 182 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 159, + 553, + 182 + ], + "type": "inline_equation", + "content": "x\\sim p_t(x|x_1)" + }, + { + "bbox": [ + 313, + 159, + 553, + 182 + ], + "type": "text", + "content": ", is equivalent to (3) with respect to the gradient " + }, + { + "bbox": [ + 313, + 159, + 553, + 182 + ], + "type": "inline_equation", + "content": "\\nabla_{\\theta}" + }, + { + "bbox": [ + 313, + 159, + 553, + 182 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 182, + 554, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 554, + 290 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 554, + 290 + ], + "type": "text", + "content": "One natural way of constructing " + }, + { + "bbox": [ + 313, + 182, + 554, + 290 + ], + "type": "inline_equation", + "content": "u_{t}(\\cdot |x_{1})" + }, + { + "bbox": [ + 313, + 182, + 554, + 290 + ], + "type": "text", + "content": " is a \"straight line\" that connects " + }, + { + "bbox": [ + 313, + 182, + 554, + 290 + ], + "type": "inline_equation", + "content": "x_0\\sim p_0" + }, + { + "bbox": [ + 313, + 182, + 554, + 290 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 182, + 554, + 290 + ], + "type": "inline_equation", + "content": "x_{1}\\sim q" + }, + { + "bbox": [ + 313, + 182, + 554, + 290 + ], + "type": "text", + "content": ", drawing an optimal transport (OT) path with constant velocity [35]. Specifically, a linear time interpolation between " + }, + { + "bbox": [ + 313, + 182, + 554, + 290 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 182, + 554, + 290 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 182, + 554, + 290 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 313, + 182, + 554, + 290 + ], + "type": "text", + "content": " gives us the flow " + }, + { + "bbox": [ + 313, + 182, + 554, + 290 + ], + "type": "inline_equation", + "content": "x_{t} = \\varphi_{t}(x) = (1 - t)x_{0} + tx_{1}" + }, + { + "bbox": [ + 313, + 182, + 554, + 290 + ], + "type": "text", + "content": ", the conditional probability path " + }, + { + "bbox": [ + 313, + 182, + 554, + 290 + ], + "type": "inline_equation", + "content": "p_t(x|x_1)" + }, + { + "bbox": [ + 313, + 182, + 554, + 290 + ], + "type": "text", + "content": " defined via the affine transformation " + }, + { + "bbox": [ + 313, + 182, + 554, + 290 + ], + "type": "inline_equation", + "content": "p_t(x|x_1) = \\mathcal{N}(x|tx_1,(1 - t)^2 I)" + }, + { + "bbox": [ + 313, + 182, + 554, + 290 + ], + "type": "text", + "content": ", and the target generating vector field " + }, + { + "bbox": [ + 313, + 182, + 554, + 290 + ], + "type": "inline_equation", + "content": "u_{t}(x|x_{1}) = x_{1} - x_{0}" + }, + { + "bbox": [ + 313, + 182, + 554, + 290 + ], + "type": "text", + "content": ". This specific choice turns the objective (4) into" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 324, + 297, + 553, + 311 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 297, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 324, + 297, + 553, + 311 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {O T}} (\\theta) := \\left\\| v _ {t} ((1 - t) x _ {0} + t x _ {1}; \\theta) - (x _ {1} - x _ {0}) \\right\\| _ {2} ^ {2}, \\tag {5}", + "image_path": "96401672f2a093864aff4416d6feaeed49a61000d930d3c187cc862b3da02039.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 317, + 553, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 317, + 553, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 317, + 553, + 340 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 317, + 553, + 340 + ], + "type": "inline_equation", + "content": "t\\sim \\mathcal{U}[0,1]" + }, + { + "bbox": [ + 313, + 317, + 553, + 340 + ], + "type": "inline_equation", + "content": "x_0\\sim p_0" + }, + { + "bbox": [ + 313, + 317, + 553, + 340 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 313, + 317, + 553, + 340 + ], + "type": "inline_equation", + "content": "x_{1}\\sim q" + }, + { + "bbox": [ + 313, + 317, + 553, + 340 + ], + "type": "text", + "content": " , all of which are tractable." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 341, + 554, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 554, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 554, + 401 + ], + "type": "text", + "content": "Classifier-free Vector Field [10] formulates a classifier-free vector field (CFV) technique for flow matching, which enables class-conditional sampling more controllable manner without any extra classifier trained on noisy trajectory. Formally, CFV compute the modified vector field " + }, + { + "bbox": [ + 313, + 341, + 554, + 401 + ], + "type": "inline_equation", + "content": "\\tilde{v}_t" + }, + { + "bbox": [ + 313, + 341, + 554, + 401 + ], + "type": "text", + "content": " by" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 319, + 407, + 553, + 421 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 407, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 319, + 407, + 553, + 421 + ], + "type": "interline_equation", + "content": "\\tilde {v} _ {t} \\left(x _ {t}, c; \\theta\\right) \\approx \\gamma v _ {t} \\left(x _ {t}, c; \\theta\\right) + (1 - \\gamma) v _ {t} \\left(x _ {t}, c = \\emptyset ; \\theta\\right), \\tag {6}", + "image_path": "f799526e0006d8509a9dac3309b46d1122575740a59d23f72a856b9b99604e47.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 427, + 554, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 427, + 554, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 427, + 554, + 464 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 427, + 554, + 464 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 427, + 554, + 464 + ], + "type": "text", + "content": " denotes the guidance scale. " + }, + { + "bbox": [ + 313, + 427, + 554, + 464 + ], + "type": "inline_equation", + "content": "v_{t}(x_{t},c = \\emptyset ;\\theta)" + }, + { + "bbox": [ + 313, + 427, + 554, + 464 + ], + "type": "text", + "content": " is the predicted vector field without a driving condition " + }, + { + "bbox": [ + 313, + 427, + 554, + 464 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 427, + 554, + 464 + ], + "type": "text", + "content": ". For more details, please refer to [10, 35]." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 472, + 554, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 554, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 554, + 501 + ], + "type": "text", + "content": "4. Method: Flow Matching for Audio-driven Talking Portrait" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 507, + 554, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 554, + 543 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 554, + 543 + ], + "type": "text", + "content": "We provide an overview of FLOAT in Fig. 2. Given source image " + }, + { + "bbox": [ + 313, + 507, + 554, + 543 + ], + "type": "inline_equation", + "content": "S \\in \\mathbb{R}^{3 \\times H \\times W}" + }, + { + "bbox": [ + 313, + 507, + 554, + 543 + ], + "type": "text", + "content": ", and a driving audio signal " + }, + { + "bbox": [ + 313, + 507, + 554, + 543 + ], + "type": "inline_equation", + "content": "a^{1:L} \\in \\mathbb{R}^{L \\times d_a}" + }, + { + "bbox": [ + 313, + 507, + 554, + 543 + ], + "type": "text", + "content": " of length " + }, + { + "bbox": [ + 313, + 507, + 554, + 543 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 507, + 554, + 543 + ], + "type": "text", + "content": ", our method generates a video" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 367, + 548, + 553, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 548, + 553, + 563 + ], + "spans": [ + { + "bbox": [ + 367, + 548, + 553, + 563 + ], + "type": "interline_equation", + "content": "\\hat {D} ^ {1: L} = \\left(\\hat {D} ^ {l}\\right) _ {l = 1} ^ {L} \\in \\mathbb {R} ^ {L \\times 3 \\times H \\times W} \\tag {7}", + "image_path": "27731320b789a8fa088f3778090bd0ddc80f76221d2cb40017bd07a5de7c1016.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "content": "of " + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "content": " frames, featuring audio-synchronized talking head motions, including both verbal and non-verbal motions. Our method consists of two phases. First, we pre-train a motion auto-encoder, which provides us with the expressive and smooth motion latent space for the talking portraits (Sec. 4.1). Next, we employ OT-based flow matching [35] to generate a sequence of motion latents with a transformer-based vector field predictor using the driving audio, which is decoded to the talking portrait videos (Sec. 4.2). We also incorporate speech-driven emotions as the driving conditions, achieving automatic emotion-aware talking portrait generation without any extra user input for emotion." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14701" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 71, + 547, + 194 + ], + "blocks": [ + { + "bbox": [ + 62, + 71, + 547, + 194 + ], + "lines": [ + { + "bbox": [ + 62, + 71, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 62, + 71, + 547, + 194 + ], + "type": "image", + "image_path": "a0dbc4d307dc1ce96a87d77c15d61dadc1ceb8953c76d8b4364fe5b1d5c18c47.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 201, + 555, + 260 + ], + "lines": [ + { + "bbox": [ + 55, + 201, + 555, + 260 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 555, + 260 + ], + "type": "text", + "content": "Figure 2. Overview of FLOAT. We encode the source image " + }, + { + "bbox": [ + 55, + 201, + 555, + 260 + ], + "type": "inline_equation", + "content": "S \\in \\mathbb{R}^{3 \\times H \\times W}" + }, + { + "bbox": [ + 55, + 201, + 555, + 260 + ], + "type": "text", + "content": " into the latent with the explicit identity-motion decomposition " + }, + { + "bbox": [ + 55, + 201, + 555, + 260 + ], + "type": "inline_equation", + "content": "w_{s} = w_{s \\to r} + w_{r \\to s} \\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 55, + 201, + 555, + 260 + ], + "type": "text", + "content": ". Given audio segments " + }, + { + "bbox": [ + 55, + 201, + 555, + 260 + ], + "type": "inline_equation", + "content": "a^{-L':L} \\in \\mathbb{R}^{(L' + L) \\times d_{a}}" + }, + { + "bbox": [ + 55, + 201, + 555, + 260 + ], + "type": "text", + "content": " of the length " + }, + { + "bbox": [ + 55, + 201, + 555, + 260 + ], + "type": "inline_equation", + "content": "L' + L" + }, + { + "bbox": [ + 55, + 201, + 555, + 260 + ], + "type": "text", + "content": " and the reference motion " + }, + { + "bbox": [ + 55, + 201, + 555, + 260 + ], + "type": "inline_equation", + "content": "w_{r \\to s} \\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 55, + 201, + 555, + 260 + ], + "type": "text", + "content": ", and the speech-driven emotion label " + }, + { + "bbox": [ + 55, + 201, + 555, + 260 + ], + "type": "inline_equation", + "content": "w_{e} \\in \\mathbb{R}^{7}" + }, + { + "bbox": [ + 55, + 201, + 555, + 260 + ], + "type": "text", + "content": ", a flow matching transformer estimates the generating vector field " + }, + { + "bbox": [ + 55, + 201, + 555, + 260 + ], + "type": "inline_equation", + "content": "v_{t}(\\varphi_{t}(x_{0}), \\mathbf{c}_{t}; \\theta) \\in \\mathbb{R}^{L \\times d}" + }, + { + "bbox": [ + 55, + 201, + 555, + 260 + ], + "type": "text", + "content": " from noisy motion latents, which is used to solve corresponding ODE and generates the motion latents " + }, + { + "bbox": [ + 55, + 201, + 555, + 260 + ], + "type": "inline_equation", + "content": "w_{r \\to \\hat{D}^{1:L}}" + }, + { + "bbox": [ + 55, + 201, + 555, + 260 + ], + "type": "text", + "content": ". Finally, the sequence of latents " + }, + { + "bbox": [ + 55, + 201, + 555, + 260 + ], + "type": "inline_equation", + "content": "w_{S \\to \\hat{D}^{1:L}} := (w_{S \\to r} + w_{r \\to \\hat{D}^{l}})_{l=1}^{L}" + }, + { + "bbox": [ + 55, + 201, + 555, + 260 + ], + "type": "text", + "content": " are decoded into the video " + }, + { + "bbox": [ + 55, + 201, + 555, + 260 + ], + "type": "inline_equation", + "content": "\\hat{D}^{1:L} \\in \\mathbb{R}^{L \\times 3 \\times H \\times W}" + }, + { + "bbox": [ + 55, + 201, + 555, + 260 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 271, + 212, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 271, + 212, + 282 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 212, + 282 + ], + "type": "text", + "content": "4.1. Motion Latent Auto-encoder" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 288, + 295, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 295, + 384 + ], + "type": "text", + "content": "Recent talking portrait methods utilize the VAE of StableDiffusion (SD) [54] due to its rich semantic pixel-based latent space. However, they often struggle to generate temporally consistent frames when lifted to video generating tasks [7, 24, 68, 79, 89]. Thus, our first goal for realistic talking portrait is to obtain good motion latent space, capturing both global (e.g., head motion) and fine-grained local (e.g., facial expressions, mouth and pupil movement) dynamics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 384, + 296, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 296, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 296, + 504 + ], + "type": "text", + "content": "Instead of VAE of SD, we pre-train a motion latent autoencoder that encodes images into motion latents. This is done by training an auto-encoder to reconstruct a driving image from a source image sampled from the same video clip, enforcing the encoder to implicitly capture both temporally adjacent and distant motions. Inspired by LIA [75], we introduce a learned orthonormal basis that decomposes the motion along distinct orthogonal directions. Specifically, our motion auto-encoder encodes the source " + }, + { + "bbox": [ + 55, + 384, + 296, + 504 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 384, + 296, + 504 + ], + "type": "text", + "content": " into the latent " + }, + { + "bbox": [ + 55, + 384, + 296, + 504 + ], + "type": "inline_equation", + "content": "w_{S} \\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 55, + 384, + 296, + 504 + ], + "type": "text", + "content": " with following explicit decomposition:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 514, + 294, + 524 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 514, + 294, + 524 + ], + "spans": [ + { + "bbox": [ + 127, + 514, + 294, + 524 + ], + "type": "interline_equation", + "content": "w _ {S} := w _ {S \\rightarrow r} + w _ {r \\rightarrow S}, \\tag {8}", + "image_path": "acd113dc3a461292cadad5127f839d6071d51356f720cc52752d64596cf375e4.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 532, + 233, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 233, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 233, + 544 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 532, + 233, + 544 + ], + "type": "inline_equation", + "content": "w_{S\\to r}\\in \\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 532, + 233, + 544 + ], + "type": "text", + "content": " is the identity latent and" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 109, + 552, + 294, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 552, + 294, + 583 + ], + "spans": [ + { + "bbox": [ + 109, + 552, + 294, + 583 + ], + "type": "interline_equation", + "content": "w _ {r \\rightarrow S} = \\sum_ {m = 1} ^ {M} \\lambda_ {m} (S) \\cdot \\mathbf {v} _ {m} \\in \\mathbb {R} ^ {d} \\tag {9}", + "image_path": "a49f7a5eee6d18f8a5c95c244caaf5dc0324c9f7277953cdab519acf8d141a1d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 592, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 592, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 295, + 653 + ], + "type": "text", + "content": "is the motion latent with " + }, + { + "bbox": [ + 55, + 592, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\lambda(S) := (\\lambda_m(S))_{m=1}^M \\in \\mathbb{R}^M" + }, + { + "bbox": [ + 55, + 592, + 295, + 653 + ], + "type": "text", + "content": " being the source-dependent motion coefficients that span the learned source-agnostic motion basis " + }, + { + "bbox": [ + 55, + 592, + 295, + 653 + ], + "type": "inline_equation", + "content": "V := \\{\\mathbf{v}_m\\}_{m=1}^M \\subseteq \\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 592, + 295, + 653 + ], + "type": "text", + "content": ". In this space, " + }, + { + "bbox": [ + 55, + 592, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\lambda_m(S)" + }, + { + "bbox": [ + 55, + 592, + 295, + 653 + ], + "type": "text", + "content": " is the intensity of the motion direction " + }, + { + "bbox": [ + 55, + 592, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_m" + }, + { + "bbox": [ + 55, + 592, + 295, + 653 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 654, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 689 + ], + "type": "text", + "content": "As shown in Fig. 6, our method enables motion editing of the sampled (generated) motion using only the basis " + }, + { + "bbox": [ + 55, + 654, + 295, + 689 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 654, + 295, + 689 + ], + "type": "text", + "content": " and its orthogonality, as stated in Eq. (15)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Improving Fidelity of Facial Components: " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{comp - lp}}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " The expressiveness of generated motions and the image fidelity" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 269, + 550, + 370 + ], + "blocks": [ + { + "bbox": [ + 317, + 269, + 550, + 370 + ], + "lines": [ + { + "bbox": [ + 317, + 269, + 550, + 370 + ], + "spans": [ + { + "bbox": [ + 317, + 269, + 550, + 370 + ], + "type": "image", + "image_path": "5333d2792990caafb26491a4c5f2820a0f715920319784d7894970a3369fd475.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 373, + 552, + 384 + ], + "lines": [ + { + "bbox": [ + 315, + 373, + 552, + 384 + ], + "spans": [ + { + "bbox": [ + 315, + 373, + 552, + 384 + ], + "type": "text", + "content": "Figure 3. Efficacy of " + }, + { + "bbox": [ + 315, + 373, + 552, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{comp - lp}}" + }, + { + "bbox": [ + 315, + 373, + 552, + 384 + ], + "type": "text", + "content": " for fine-grained motion and fidelity." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 400, + 555, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 555, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 555, + 532 + ], + "type": "text", + "content": "are determined by the motion space and the motion autoencoder. However, as resolution increases, fine details in small facial regions (e.g., teeth, eyeballs) often get buried in large-scale dynamics. To address this issue, we propose a facial component perceptual loss " + }, + { + "bbox": [ + 313, + 400, + 555, + 532 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{comp - lp}}" + }, + { + "bbox": [ + 313, + 400, + 555, + 532 + ], + "type": "text", + "content": " using [59, 83] that significantly improves the image fidelity (e.g., teeth and eyes) as well as fine-grained motions (e.g., eyeball and eyebrows movements). As shown in Fig. 3, " + }, + { + "bbox": [ + 313, + 400, + 555, + 532 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{comp - lp}}" + }, + { + "bbox": [ + 313, + 400, + 555, + 532 + ], + "type": "text", + "content": " allows us to generate high-fidelity facial components and their fine-grained motions without relying on pre-trained foundation models, such as StableDiffusion [54]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 540, + 520, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 520, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 520, + 553 + ], + "type": "text", + "content": "4.2. Flow Matching in Motion Latent Space" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 558, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 700 + ], + "type": "text", + "content": "Armed with this linear orthogonal space, we employ OT-based flow matching [35, 37] for the motion sampling. Specifically, we predict a vector field " + }, + { + "bbox": [ + 313, + 558, + 555, + 700 + ], + "type": "inline_equation", + "content": "v_{t}(x_{t},\\mathbf{c}_{t};\\theta)\\in \\mathbb{R}^{L\\times d}" + }, + { + "bbox": [ + 313, + 558, + 555, + 700 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 558, + 555, + 700 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 313, + 558, + 555, + 700 + ], + "type": "text", + "content": " is the sample at flow time " + }, + { + "bbox": [ + 313, + 558, + 555, + 700 + ], + "type": "inline_equation", + "content": "t\\in [0,1]" + }, + { + "bbox": [ + 313, + 558, + 555, + 700 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 558, + 555, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_t\\in \\mathbb{R}^{L\\times h}" + }, + { + "bbox": [ + 313, + 558, + 555, + 700 + ], + "type": "text", + "content": " represents the driving conditions for " + }, + { + "bbox": [ + 313, + 558, + 555, + 700 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 558, + 555, + 700 + ], + "type": "text", + "content": " consequent frames. This vector field generates the flow " + }, + { + "bbox": [ + 313, + 558, + 555, + 700 + ], + "type": "inline_equation", + "content": "\\varphi_t:[0,1]\\times \\mathbb{R}^{L\\times d}\\to \\mathbb{R}^{L\\times d}" + }, + { + "bbox": [ + 313, + 558, + 555, + 700 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 558, + 555, + 700 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 558, + 555, + 700 + ], + "type": "text", + "content": " frames by solving ODE (Eq. (2)). As illustrated in Fig. 4, we build our vector field predictor upon the transformer encoder [70] architecture. Specifically, we adopt DiT [48] architecture, but decouple frame-wise conditioning from time-axis attention mechanism, which enables us to model temporally consistent motion latents." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "type": "text", + "content": "In DiT [48], distinct semantic tokens are modulated by a" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14702" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 71, + 284, + 171 + ], + "blocks": [ + { + "bbox": [ + 65, + 71, + 284, + 171 + ], + "lines": [ + { + "bbox": [ + 65, + 71, + 284, + 171 + ], + "spans": [ + { + "bbox": [ + 65, + 71, + 284, + 171 + ], + "type": "image", + "image_path": "13dc8fc6b5b49496a5c35bf6618e63a51d637fbb5f4fc548a2fd89effb668ceb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 173, + 288, + 184 + ], + "lines": [ + { + "bbox": [ + 61, + 173, + 288, + 184 + ], + "spans": [ + { + "bbox": [ + 61, + 173, + 288, + 184 + ], + "type": "text", + "content": "Figure 4. Frame-wise vector field predictor block at inference." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 194, + 296, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 194, + 296, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 296, + 291 + ], + "type": "text", + "content": "single diffusion time step embedding and class embedding through adaptive layer normalization (AdaLN). In contrast, our vector field predictor modulates each " + }, + { + "bbox": [ + 55, + 194, + 296, + 291 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 194, + 296, + 291 + ], + "type": "text", + "content": "-th input latent with its corresponding " + }, + { + "bbox": [ + 55, + 194, + 296, + 291 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 194, + 296, + 291 + ], + "type": "text", + "content": "-th condition and then combines their temporal relations through a masked self-attention layer that attends to " + }, + { + "bbox": [ + 55, + 194, + 296, + 291 + ], + "type": "inline_equation", + "content": "2 \\cdot T" + }, + { + "bbox": [ + 55, + 194, + 296, + 291 + ], + "type": "text", + "content": " neighboring frames. Formally, for each " + }, + { + "bbox": [ + 55, + 194, + 296, + 291 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 194, + 296, + 291 + ], + "type": "text", + "content": "-th frame, frame-wise AdaLN and frame-wise gating are computed by" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 297, + 294, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 297, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 67, + 297, + 294, + 312 + ], + "type": "interline_equation", + "content": "\\gamma_ {i} ^ {l} \\times \\operatorname {L N} \\left(X _ {t} ^ {l}\\right) + \\beta_ {i} ^ {l} \\in \\mathbb {R} ^ {h} \\quad \\text {a n d} \\quad \\alpha_ {i} ^ {l} \\times X _ {t} ^ {l} \\in \\mathbb {R} ^ {h}, \\tag {10}", + "image_path": "e14898079ca51a523808791b42cc7c6bb6eae8de18404987e0834b692d317898.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 319, + 296, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 296, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 296, + 378 + ], + "type": "text", + "content": "respectively, where " + }, + { + "bbox": [ + 55, + 319, + 296, + 378 + ], + "type": "inline_equation", + "content": "i \\in \\{1,2\\}" + }, + { + "bbox": [ + 55, + 319, + 296, + 378 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 319, + 296, + 378 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 319, + 296, + 378 + ], + "type": "text", + "content": " is the hidden dimension, " + }, + { + "bbox": [ + 55, + 319, + 296, + 378 + ], + "type": "inline_equation", + "content": "\\mathrm{LN}(\\cdot)" + }, + { + "bbox": [ + 55, + 319, + 296, + 378 + ], + "type": "text", + "content": " denotes layer norm [33], and " + }, + { + "bbox": [ + 55, + 319, + 296, + 378 + ], + "type": "inline_equation", + "content": "X_{t}^{l}" + }, + { + "bbox": [ + 55, + 319, + 296, + 378 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 55, + 319, + 296, + 378 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 319, + 296, + 378 + ], + "type": "text", + "content": "-th input for each operation at flow time " + }, + { + "bbox": [ + 55, + 319, + 296, + 378 + ], + "type": "inline_equation", + "content": "t \\in [0,1]" + }, + { + "bbox": [ + 55, + 319, + 296, + 378 + ], + "type": "text", + "content": ". The coefficients " + }, + { + "bbox": [ + 55, + 319, + 296, + 378 + ], + "type": "inline_equation", + "content": "\\alpha_{i}^{l}, \\beta_{i}^{l}, \\gamma_{i}^{l} \\in \\mathbb{R}^{h}" + }, + { + "bbox": [ + 55, + 319, + 296, + 378 + ], + "type": "text", + "content": " are computed from the condition " + }, + { + "bbox": [ + 55, + 319, + 296, + 378 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{t}^{l} \\in \\mathbb{R}^{h}" + }, + { + "bbox": [ + 55, + 319, + 296, + 378 + ], + "type": "text", + "content": " through a linear layer, ToScaleShift, as depicted in Fig. 4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 379, + 296, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 379, + 296, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 379, + 296, + 474 + ], + "type": "text", + "content": "Speech-driven Emotion Enhancement How can we make talking motions more expressive and natural? During talking, humans naturally reflect their emotions through their voices, and these emotions influence talking motions. For instance, a person who speaks sadly may be more likely to shake the head and avoid eye contact. This non-verbal motion derived from emotions crucially impacts the naturalness of a talking portrait." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 474, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 581 + ], + "type": "text", + "content": "Existing works [25, 72, 80] use image-emotion paired data or image-driven emotion predictor [56] to generate the emotion-aware motion. In contrast, we incorporate speech-driven emotions, a more intuitive way of controlling emotion for audio-driven talking portrait. Specifically, we utilize a pre-trained speech emotion predictor [49] that produces softmax probabilities of seven distinct emotions: angry, disgust, fear, happy, neutral, sad, and surprise, which we then input into the vector field predictor." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 665 + ], + "type": "text", + "content": "However, as people do not always speak with a single, clear emotion, determining emotions solely from audio is often ambiguous [25]. Naive introduction of speech-driven emotion can make emotion-aware motion generation more challenging. To address this issue, we inject the emotions together with other driving conditions at training phase and modify them at inference phase." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Driving Conditions We concatenate the audio representation " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "a^{1:L} \\in \\mathbb{R}^{L \\times d_a}" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " of a pre-trained Wav2Vec2.0 [1], the speech emotion label " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "w_e \\in \\mathbb{R}^7" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ", and the source motion latent " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "w_{r \\to S} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ". Next, we add the flow time step embedding" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "\\operatorname{Emb}(t) \\in \\mathbb{R}^h" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": " to these conditions, producing " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_t \\in \\mathbb{R}^{L \\times h}" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": " via a linear layer, ToCondition, as depicted in Fig. 2, where " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "\\operatorname{Emb}(t)" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": " is computed using the sinusoidal position embedding [70]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 120, + 554, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 120, + 554, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 120, + 554, + 215 + ], + "type": "text", + "content": "Training We train FLOAT by reconstructing a target vector field computed from driving frames using the corresponding audio segments and a source motion latent. We choose a pair of driving motions and corresponding audio " + }, + { + "bbox": [ + 313, + 120, + 554, + 215 + ], + "type": "inline_equation", + "content": "(w_{r\\rightarrow D^{1:L}},a^{1:L})" + }, + { + "bbox": [ + 313, + 120, + 554, + 215 + ], + "type": "text", + "content": " , and construct the target vector field " + }, + { + "bbox": [ + 313, + 120, + 554, + 215 + ], + "type": "inline_equation", + "content": "u_{t}(x|w_{r\\rightarrow D^{1:L}}) = w_{r\\rightarrow D^{1:L}} - x_{0}\\in \\mathbb{R}^{L\\times d}" + }, + { + "bbox": [ + 313, + 120, + 554, + 215 + ], + "type": "text", + "content": " with noisy input " + }, + { + "bbox": [ + 313, + 120, + 554, + 215 + ], + "type": "inline_equation", + "content": "\\varphi_t(x_0) = (1 - t)x_0 + tw_{r\\to D^{1:L}}(t\\sim \\mathcal{U}[0,1]" + }, + { + "bbox": [ + 313, + 120, + 554, + 215 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 120, + 554, + 215 + ], + "type": "inline_equation", + "content": "x_0\\sim" + }, + { + "bbox": [ + 313, + 120, + 554, + 215 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0^{1:L},I)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 216, + 554, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 554, + 262 + ], + "type": "text", + "content": "For smooth transitions of sequences longer than the window length " + }, + { + "bbox": [ + 313, + 216, + 554, + 262 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 216, + 554, + 262 + ], + "type": "text", + "content": ", we incorporate last " + }, + { + "bbox": [ + 313, + 216, + 554, + 262 + ], + "type": "inline_equation", + "content": "L'" + }, + { + "bbox": [ + 313, + 216, + 554, + 262 + ], + "type": "text", + "content": " audio features and motion latents " + }, + { + "bbox": [ + 313, + 216, + 554, + 262 + ], + "type": "inline_equation", + "content": "w_{r\\rightarrow D^{-L':0}}" + }, + { + "bbox": [ + 313, + 216, + 554, + 262 + ], + "type": "text", + "content": " from the preceding window as additional input." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 326, + 263, + 528, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 263, + 528, + 276 + ], + "spans": [ + { + "bbox": [ + 326, + 263, + 528, + 276 + ], + "type": "text", + "content": "The flow matching objective " + }, + { + "bbox": [ + 326, + 263, + 528, + 276 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{OT}}(\\theta)" + }, + { + "bbox": [ + 326, + 263, + 528, + 276 + ], + "type": "text", + "content": " is defined by" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 331, + 283, + 553, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 283, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 331, + 283, + 553, + 316 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {O T}} (\\theta) = \\left\\| v _ {t} ^ {1: L} \\left(x _ {t}, \\mathbf {c} _ {t}; \\theta\\right) - u _ {t} \\left(x \\mid w _ {r \\rightarrow D ^ {1: L}}\\right)\\right\\|, \\tag {11} \\\\ + \\left\\| v _ {t} ^ {- L ^ {\\prime}: 0} \\left(x _ {t}, \\mathbf {c} _ {t}; \\theta\\right) - w _ {r \\rightarrow D ^ {- L ^ {\\prime}: 0}} \\right\\|, \\\\ \\end{array}", + "image_path": "e5c3c93e5c0b175bad72f974fb90a6d40ffbf25481e814a1f6d31a00e3358436.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 324, + 554, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 554, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 554, + 388 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 324, + 554, + 388 + ], + "type": "inline_equation", + "content": "x_{t} \\coloneqq [w_{r \\to D^{-L' \\cdot 0}} | \\varphi_{t}(x_{0})] \\in \\mathbb{R}^{(-L' + L) \\times d}" + }, + { + "bbox": [ + 313, + 324, + 554, + 388 + ], + "type": "text", + "content": " is the concatenated input, " + }, + { + "bbox": [ + 313, + 324, + 554, + 388 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{t} \\in \\mathbb{R}^{(-L' + L) \\times h}" + }, + { + "bbox": [ + 313, + 324, + 554, + 388 + ], + "type": "text", + "content": " is the driving condition consisting of " + }, + { + "bbox": [ + 313, + 324, + 554, + 388 + ], + "type": "inline_equation", + "content": "[t, w_{r \\to S}, w_{e}, a^{1:L}, a^{-L':0}]" + }, + { + "bbox": [ + 313, + 324, + 554, + 388 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 313, + 324, + 554, + 388 + ], + "type": "inline_equation", + "content": "w_{e}" + }, + { + "bbox": [ + 313, + 324, + 554, + 388 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 324, + 554, + 388 + ], + "type": "inline_equation", + "content": "w_{r \\to S}" + }, + { + "bbox": [ + 313, + 324, + 554, + 388 + ], + "type": "text", + "content": " are shared across the " + }, + { + "bbox": [ + 313, + 324, + 554, + 388 + ], + "type": "inline_equation", + "content": "L' + L" + }, + { + "bbox": [ + 313, + 324, + 554, + 388 + ], + "type": "text", + "content": " frames. We incorporate a velocity loss [67] to supervise temporal consistency:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 382, + 396, + 553, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 396, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 382, + 396, + 553, + 409 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {v e l}} (\\theta) = \\left\\| \\Delta v _ {t} - \\Delta u _ {t} \\right\\|, \\tag {12}", + "image_path": "1132bd7a68e7a149f5fbd4cecae1f864718f4d457b364c8d970c3fbf1b55967c.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 417, + 554, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 417, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 417, + 554, + 453 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 417, + 554, + 453 + ], + "type": "inline_equation", + "content": "\\Delta v_{t}" + }, + { + "bbox": [ + 313, + 417, + 554, + 453 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 417, + 554, + 453 + ], + "type": "inline_equation", + "content": "\\Delta u_{t}" + }, + { + "bbox": [ + 313, + 417, + 554, + 453 + ], + "type": "text", + "content": " are the one-frame difference along the time-axis for the prediction " + }, + { + "bbox": [ + 313, + 417, + 554, + 453 + ], + "type": "inline_equation", + "content": "v_{t} \\in \\mathbb{R}^{(-L' + L) \\times d}" + }, + { + "bbox": [ + 313, + 417, + 554, + 453 + ], + "type": "text", + "content": " and the target " + }, + { + "bbox": [ + 313, + 417, + 554, + 453 + ], + "type": "inline_equation", + "content": "[w_{r \\to D^{-L':0}} | u_{t}] \\in \\mathbb{R}^{(-L' + L) \\times d}" + }, + { + "bbox": [ + 313, + 417, + 554, + 453 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 326, + 453, + 448, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 453, + 448, + 465 + ], + "spans": [ + { + "bbox": [ + 326, + 453, + 448, + 465 + ], + "type": "text", + "content": "The total objective " + }, + { + "bbox": [ + 326, + 453, + 448, + 465 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{total}}(\\theta)" + }, + { + "bbox": [ + 326, + 453, + 448, + 465 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 358, + 474, + 553, + 487 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 474, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 358, + 474, + 553, + 487 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} (\\theta) = \\lambda_ {\\mathrm {O T}} \\mathcal {L} _ {\\mathrm {O T}} (\\theta) + \\lambda_ {\\mathrm {v e l}} \\mathcal {L} _ {\\mathrm {v e l}} (\\theta), \\tag {13}", + "image_path": "d34dbc1800bea4c7b4753bd241a90371a56ddbab4883109f7885cc1f1a423b6c.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 495, + 554, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 554, + 554 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 554, + 554 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 495, + 554, + 554 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{OT}}" + }, + { + "bbox": [ + 313, + 495, + 554, + 554 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 495, + 554, + 554 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{vel}}" + }, + { + "bbox": [ + 313, + 495, + 554, + 554 + ], + "type": "text", + "content": " are the balancing coefficients. During training, we apply dropout to " + }, + { + "bbox": [ + 313, + 495, + 554, + 554 + ], + "type": "inline_equation", + "content": "w_{r}" + }, + { + "bbox": [ + 313, + 495, + 554, + 554 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 495, + 554, + 554 + ], + "type": "inline_equation", + "content": "w_{e}" + }, + { + "bbox": [ + 313, + 495, + 554, + 554 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 495, + 554, + 554 + ], + "type": "inline_equation", + "content": "a^{1:L}" + }, + { + "bbox": [ + 313, + 495, + 554, + 554 + ], + "type": "text", + "content": " with a probability of 0.1 for CFV. Additionally, we apply dropout to the preceding audio and motion latents with a probability 0.5 for smooth transition in the initial window." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 555, + 554, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 554, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 554, + 602 + ], + "type": "text", + "content": "Inference During inference, we sample the generating vector field from noise " + }, + { + "bbox": [ + 313, + 555, + 554, + 602 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 555, + 554, + 602 + ], + "type": "text", + "content": ", using the driving conditions " + }, + { + "bbox": [ + 313, + 555, + 554, + 602 + ], + "type": "inline_equation", + "content": "w_{r\\rightarrow S}" + }, + { + "bbox": [ + 313, + 555, + 554, + 602 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 555, + 554, + 602 + ], + "type": "inline_equation", + "content": "w_e" + }, + { + "bbox": [ + 313, + 555, + 554, + 602 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 555, + 554, + 602 + ], + "type": "inline_equation", + "content": "a^{1:L}" + }, + { + "bbox": [ + 313, + 555, + 554, + 602 + ], + "type": "text", + "content": ", as well as the " + }, + { + "bbox": [ + 313, + 555, + 554, + 602 + ], + "type": "inline_equation", + "content": "L'" + }, + { + "bbox": [ + 313, + 555, + 554, + 602 + ], + "type": "text", + "content": " frames of preceding audio and generated motion latents." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 603, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 603, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 603, + 554, + 628 + ], + "type": "text", + "content": "We extend the CFV [10] to an incremental CFV to separately adjust the audio and emotion, inspired by [3]:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 343, + 635, + 553, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 635, + 553, + 681 + ], + "spans": [ + { + "bbox": [ + 343, + 635, + 553, + 681 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\tilde {v} _ {t} \\approx v _ {t} (x _ {0}, \\mathbf {c} _ {t} | _ {\\{a ^ {1: L}, w _ {e} \\}}) \\\\ + \\gamma_ {a} \\left[ v _ {t} \\left(x _ {0}, \\mathbf {c} _ {t} \\mid_ {w _ {e}}\\right) - v _ {t} \\left(x _ {0}, \\mathbf {c} _ {t} \\mid_ {\\{a ^ {1: L}, w _ {e} \\}} \\right] \\right. \\\\ + \\gamma_ {e} \\left[ v _ {t} \\left(x _ {0}, \\mathbf {c} _ {t}\\right) - v _ {t} \\left(x _ {0}, \\mathbf {c} _ {t} \\mid_ {w _ {e}}\\right) \\right], \\tag {14} \\\\ \\end{array}", + "image_path": "2135c7c784e391756fa40c7441d302bfc3e071e9fdea7c90db04c11e358201d3.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\gamma_{a}" + }, + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\gamma_{e}" + }, + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "text", + "content": " are the guidance scales for audio and emotion, respectively. " + }, + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_t|_{\\{x,y\\}}" + }, + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "text", + "content": " denotes the driving condition" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14703" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 96, + 550, + 188 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 555, + 93 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 555, + 93 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 555, + 93 + ], + "type": "text", + "content": "Table 1. Quantitative comparison results with state-of-the-art methods on HDTF [85] / RAVDESS [39]. The best result for each metric is in bold, and the second-best result is underlined." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 96, + 550, + 188 + ], + "lines": [ + { + "bbox": [ + 61, + 96, + 550, + 188 + ], + "spans": [ + { + "bbox": [ + 61, + 96, + 550, + 188 + ], + "type": "table", + "html": "
MethodImage & Video GenerationLip Synchronization
FID ↓FVD ↓CSIM ↑E-FID ↓P-FID ↓LSE-D ↓LSE-C ↑
SadTalker† [84]71.952 / 119.430339.058 / 376.2940.644 / 0.6441.914 / 3.5001.456 / 2.0457.947 / 7.2737.305 / 4.748
EDTalk† [66]50.078 / 75.020211.284 / 304.9330.626 / 0.6761.579 / 3.4680.054 / 0.0908.123 / 7.6827.623 / 5.318
AniTalker† [36]39.512 / 70.430184.454 / 265.3410.643 / 0.7251.830 / 2.3300.092 / 0.1267.907 / 8.1767.288 / 4.555
Halo [79]25.363 / 57.648197.196 / 375.5570.869 / 0.8601.039 / 2.4920.037 / 0.0507.792 / 7.6137.582 / 4.795
EchoMimic [7]33.552 / 81.839296.757 / 320.2200.823 / 0.8051.234 / 3.2010.023 / 0.0478.903 / 8.1616.242 / 4.144
FLOAT (Ours)21.100 / 31.681162.052 / 166.3590.843 / 0.8101.229 / 1.3670.032 / 0.0317.290 / 6.9948.222 / 5.730
", + "image_path": "e5afec730d66757140b8157b78f57d5906c6a01e55b2d65608480adec298dc2b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 63, + 198, + 281, + 518 + ], + "blocks": [ + { + "bbox": [ + 63, + 198, + 281, + 518 + ], + "lines": [ + { + "bbox": [ + 63, + 198, + 281, + 518 + ], + "spans": [ + { + "bbox": [ + 63, + 198, + 281, + 518 + ], + "type": "image", + "image_path": "daff56d026ccb964d277b4599dfc4f045c6db093de7e7c345df483f887a0ba5e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 521, + 555, + 544 + ], + "lines": [ + { + "bbox": [ + 55, + 521, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 555, + 544 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison results with state-of-the-art methods on HDTF [85] / RAVDESS [39]. Please refer to supplementary videos. Note that we additionally provide a video comparison with EMO [68] and VASA-1 [80] using their video demonstration." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 283, + 293, + 337, + 356 + ], + "blocks": [ + { + "bbox": [ + 283, + 293, + 337, + 356 + ], + "lines": [ + { + "bbox": [ + 283, + 293, + 337, + 356 + ], + "spans": [ + { + "bbox": [ + 283, + 293, + 337, + 356 + ], + "type": "image", + "image_path": "3255bf62b78bce64af2a45ed15e8d0a73102a854ff37627a755a8f9541ebd61a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 288, + 357, + 328, + 411 + ], + "blocks": [ + { + "bbox": [ + 288, + 357, + 328, + 411 + ], + "lines": [ + { + "bbox": [ + 288, + 357, + 328, + 411 + ], + "spans": [ + { + "bbox": [ + 288, + 357, + 328, + 411 + ], + "type": "image", + "image_path": "1c5bddc5d113e1988b2eff754be5d5038bf803625fe43d6a11361efc5373dec3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 342, + 196, + 541, + 518 + ], + "blocks": [ + { + "bbox": [ + 342, + 196, + 541, + 518 + ], + "lines": [ + { + "bbox": [ + 342, + 196, + 541, + 518 + ], + "spans": [ + { + "bbox": [ + 342, + 196, + 541, + 518 + ], + "type": "image", + "image_path": "f4481d782c00a7ecbdb51e51ffd279fa5a42c4221caa8dc70e74f38ff5c8d111.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 563, + 295, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 295, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 295, + 598 + ], + "type": "text", + "content": "without the condition " + }, + { + "bbox": [ + 55, + 563, + 295, + 598 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 563, + 295, + 598 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 563, + 295, + 598 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 563, + 295, + 598 + ], + "type": "text", + "content": ". We set " + }, + { + "bbox": [ + 55, + 563, + 295, + 598 + ], + "type": "inline_equation", + "content": "\\gamma_{a} = 2" + }, + { + "bbox": [ + 55, + 563, + 295, + 598 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 563, + 295, + 598 + ], + "type": "inline_equation", + "content": "\\gamma_{e} = 1" + }, + { + "bbox": [ + 55, + 563, + 295, + 598 + ], + "type": "text", + "content": " based on the ablation studies on " + }, + { + "bbox": [ + 55, + 563, + 295, + 598 + ], + "type": "inline_equation", + "content": "\\gamma_{a}" + }, + { + "bbox": [ + 55, + 563, + 295, + 598 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 563, + 295, + 598 + ], + "type": "inline_equation", + "content": "\\gamma_{e}" + }, + { + "bbox": [ + 55, + 563, + 295, + 598 + ], + "type": "text", + "content": " provided in supplementary materials." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 608, + 295, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 608, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 55, + 608, + 295, + 668 + ], + "type": "text", + "content": "After sampling, ODE solver receives the estimated vector field to compute the motion latents through numerical integration. We empirically find that FLOAT can generate reasonable motion with around 10 number of function evaluations (NFE). Please refer to supplementary videos." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 712 + ], + "type": "text", + "content": "Lastly, we add the source identity latent to the generated motion latents and decode them into video frames using the motion latent decoder." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 562, + 395, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 562, + 395, + 575 + ], + "spans": [ + { + "bbox": [ + 314, + 562, + 395, + 575 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 585, + 465, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 585, + 465, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 585, + 465, + 597 + ], + "type": "text", + "content": "5.1. Dataset and Pre-processing" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "For training the motion latent auto-encoder, we use three open-source datasets: HDTF [85], RAVDESS [39], and VFHQ [78]. When training FLOAT, we exclude VFHQ because it does not support the synchronized audio. HDTF [85] is for high-definition talking face generation, containing videos of over 300 unique identities. RAVDESS [39] includes more than 2,400 emotion-intensive videos of 24 different identities. VFHQ [78] is designed for high-resolution video super-resolution and includes a large num" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14704" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": "ber of unique identities, which compensates the limited number of identities of the preceding datasets. Following the strategy of [58], we first convert each video to 25 FPS and resample the audio into " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "16\\mathrm{kHz}" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": ". Then, we crop and resize the facial region to " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "512^2" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": " resolution [5]. After the pre-processing, for HDTF, we use a total of 11.3 hours of 240 videos featuring 230 different identities for training, and videos of 78 different identities, each 15 seconds long, for test. For RAVDESS, we use videos of 22 identities for training, and videos of the remaining 2 identities for test, with each 3-4 seconds long and representing 14 emotional intensities. Note that the identities in the training and test are disjoint in both datasets." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 239, + 188, + 252 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 188, + 252 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 188, + 252 + ], + "type": "text", + "content": "5.2. Implementation Details" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 258, + 296, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 296, + 402 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 296, + 402 + ], + "type": "text", + "content": "The motion latent dimension is set to " + }, + { + "bbox": [ + 55, + 258, + 296, + 402 + ], + "type": "inline_equation", + "content": "d = 512" + }, + { + "bbox": [ + 55, + 258, + 296, + 402 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 258, + 296, + 402 + ], + "type": "inline_equation", + "content": "M = 20" + }, + { + "bbox": [ + 55, + 258, + 296, + 402 + ], + "type": "text", + "content": " distinct orthogonal directions. For the vector predictor, we use 8 attention heads, a hidden dimension " + }, + { + "bbox": [ + 55, + 258, + 296, + 402 + ], + "type": "inline_equation", + "content": "h = 1024" + }, + { + "bbox": [ + 55, + 258, + 296, + 402 + ], + "type": "text", + "content": ", and an attention window length " + }, + { + "bbox": [ + 55, + 258, + 296, + 402 + ], + "type": "inline_equation", + "content": "T = 2" + }, + { + "bbox": [ + 55, + 258, + 296, + 402 + ], + "type": "text", + "content": ". Considering the length of the training video clips, we set " + }, + { + "bbox": [ + 55, + 258, + 296, + 402 + ], + "type": "inline_equation", + "content": "L = 50" + }, + { + "bbox": [ + 55, + 258, + 296, + 402 + ], + "type": "text", + "content": " frames with preceding " + }, + { + "bbox": [ + 55, + 258, + 296, + 402 + ], + "type": "inline_equation", + "content": "L' = 10" + }, + { + "bbox": [ + 55, + 258, + 296, + 402 + ], + "type": "text", + "content": " frames at once, encompassing 2.4 seconds of video. We employ the Adam optimizer [30] with a batch size of 8 and a learning rate of " + }, + { + "bbox": [ + 55, + 258, + 296, + 402 + ], + "type": "inline_equation", + "content": "10^{-5}" + }, + { + "bbox": [ + 55, + 258, + 296, + 402 + ], + "type": "text", + "content": ". We use " + }, + { + "bbox": [ + 55, + 258, + 296, + 402 + ], + "type": "inline_equation", + "content": "L1" + }, + { + "bbox": [ + 55, + 258, + 296, + 402 + ], + "type": "text", + "content": " distance for the norm " + }, + { + "bbox": [ + 55, + 258, + 296, + 402 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|" + }, + { + "bbox": [ + 55, + 258, + 296, + 402 + ], + "type": "text", + "content": " in the training objective. We set the balancing coefficients to " + }, + { + "bbox": [ + 55, + 258, + 296, + 402 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{OT}} = \\lambda_{\\mathrm{vel}} = 1" + }, + { + "bbox": [ + 55, + 258, + 296, + 402 + ], + "type": "text", + "content": ". The entire training takes about 2 days for 2,000k steps on a single NVIDIA A100 GPU. We use Euler method [35] for the ODE solver." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 415, + 129, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 129, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 129, + 426 + ], + "type": "text", + "content": "5.3. Evaluation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 434, + 296, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 296, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 296, + 540 + ], + "type": "text", + "content": "Metrics and Baselines For evaluating the image and video generation quality, we measure Fréchet Inception Distance (FID) [57] and 16 frames Fréchet Video Distance (FVD) [69]. For facial identity, expression and head motion, we measure Cosine Similarity of identity embedding (CSIM) [11], Expression FID (E-FID) [68] and Pose FID (P-FID), respectively. Lastly, we measure Lip-Sync Error Distance and Confidence (LSE-D and LSE-C [51]) for audio-visual alignment." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 543, + 296, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 543, + 296, + 614 + ], + "spans": [ + { + "bbox": [ + 55, + 543, + 296, + 614 + ], + "type": "text", + "content": "We compare our method with state-of-the-art audio-driven talking portrait methods whose official implementations are publicly available. For non-diffusion methods, we compare with SadTalker [84] and EDTalk [66]. For diffusion methods, we compare with AniTalker [36], Hallo [79], and EchoMimic [7]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 616, + 296, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 296, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 296, + 664 + ], + "type": "text", + "content": "Comparison Results In Tab. 1 and Fig. 5, we show the quantitative and qualitative comparison results, respectively. FLOAT outperforms other methods on most of the metrics and visual quality in both datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "Additionally, we provide video comparison results with EMO [68] and VASA-1 [80] in the supplementary materials, using their demonstration videos due to the infeasibility of direct implementation." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 320, + 71, + 548, + 206 + ], + "blocks": [ + { + "bbox": [ + 320, + 71, + 548, + 206 + ], + "lines": [ + { + "bbox": [ + 320, + 71, + 548, + 206 + ], + "spans": [ + { + "bbox": [ + 320, + 71, + 548, + 206 + ], + "type": "image", + "image_path": "eaef4948f10feb64b30b606b91b0571f4844c966563af3497ed62e22a9a64d1f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 208, + 552, + 220 + ], + "lines": [ + { + "bbox": [ + 316, + 208, + 552, + 220 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 552, + 220 + ], + "type": "text", + "content": "Figure 6. Test-time pose editing using " + }, + { + "bbox": [ + 316, + 208, + 552, + 220 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 316, + 208, + 552, + 220 + ], + "type": "text", + "content": "-control " + }, + { + "bbox": [ + 316, + 208, + 552, + 220 + ], + "type": "inline_equation", + "content": "(\\lambda_{15}(\\hat{D}) \\pm 10)" + }, + { + "bbox": [ + 316, + 208, + 552, + 220 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 230, + 395, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 230, + 395, + 243 + ], + "spans": [ + { + "bbox": [ + 314, + 230, + 395, + 243 + ], + "type": "text", + "content": "5.4. Applications" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 247, + 553, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 553, + 310 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 553, + 310 + ], + "type": "text", + "content": "Test-time Pose Editing via Orthonormal Basis " + }, + { + "bbox": [ + 313, + 247, + 553, + 310 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 247, + 553, + 310 + ], + "type": "text", + "content": " Since FLOAT learns the underlying motion latent structure, it is natural to assume that for any sampled motion latent " + }, + { + "bbox": [ + 313, + 247, + 553, + 310 + ], + "type": "inline_equation", + "content": "w_{r\\rightarrow \\hat{D}}" + }, + { + "bbox": [ + 313, + 247, + 553, + 310 + ], + "type": "text", + "content": ", there exist motion coefficients " + }, + { + "bbox": [ + 313, + 247, + 553, + 310 + ], + "type": "inline_equation", + "content": "\\{\\lambda_m(\\hat{D})\\}_{m = 1}^M" + }, + { + "bbox": [ + 313, + 247, + 553, + 310 + ], + "type": "text", + "content": " satisfying the representation in Eq. (9): " + }, + { + "bbox": [ + 313, + 247, + 553, + 310 + ], + "type": "inline_equation", + "content": "w_{r\\rightarrow \\hat{D}} = \\sum_{m = 1}^{M}\\lambda_{m}(\\hat{D})\\cdot \\mathbf{v}_{m}" + }, + { + "bbox": [ + 313, + 247, + 553, + 310 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 311, + 553, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 553, + 347 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 553, + 347 + ], + "type": "text", + "content": "We can always compute these coefficients in closed form by taking inner products between the sampled motion " + }, + { + "bbox": [ + 313, + 311, + 553, + 347 + ], + "type": "inline_equation", + "content": "w_{r\\rightarrow \\hat{D}}" + }, + { + "bbox": [ + 313, + 311, + 553, + 347 + ], + "type": "text", + "content": " and the learned orthonormal basis " + }, + { + "bbox": [ + 313, + 311, + 553, + 347 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 311, + 553, + 347 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 323, + 353, + 553, + 385 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 353, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 323, + 353, + 553, + 385 + ], + "type": "interline_equation", + "content": "\\left\\langle w _ {r \\rightarrow \\hat {D}}, \\mathbf {v} _ {k} \\right\\rangle = \\left\\langle \\sum_ {m = 1} ^ {M} \\lambda_ {m} (\\hat {D}) \\cdot \\mathbf {v} _ {m}, \\mathbf {v} _ {k} \\right\\rangle = \\lambda_ {k} (\\hat {D}), \\tag {15}", + "image_path": "349aa663bf27b33180f6b5c6f6208a7b5cac9ac03d208b5cfe8522ef589e2bda.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 390, + 554, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 554, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 554, + 474 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 390, + 554, + 474 + ], + "type": "inline_equation", + "content": "\\langle \\mathbf{v}_m, \\mathbf{v}_k \\rangle = \\delta_{m,k}" + }, + { + "bbox": [ + 313, + 390, + 554, + 474 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 390, + 554, + 474 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 390, + 554, + 474 + ], + "type": "text", + "content": " is Kronecker delta. At this point, we can edit the sampled motions by editing the corresponding coefficients (e.g., via linear operation) and combining them back into the motion latent. As shown in Fig. 6, it allows us to control head direction without interfering with other motions due to the orthogonality of the basis. We refer to this test-time editing technique as " + }, + { + "bbox": [ + 313, + 390, + 554, + 474 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 390, + 554, + 474 + ], + "type": "text", + "content": "-control." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 474, + 554, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 554, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 554, + 617 + ], + "type": "text", + "content": "Additional Driving Signals In Fig. 7 and Tab. 2, we experiment with additional driving conditions, head poses and image-driven emotion labels, to explore additional controllability in our method. We employ 3DMM head pose parameters " + }, + { + "bbox": [ + 313, + 474, + 554, + 617 + ], + "type": "inline_equation", + "content": "p \\in \\mathbb{R}^6" + }, + { + "bbox": [ + 313, + 474, + 554, + 617 + ], + "type": "text", + "content": " [2] extracted by [12]. We concatenate a sequence of pose parameters " + }, + { + "bbox": [ + 313, + 474, + 554, + 617 + ], + "type": "inline_equation", + "content": "p^{1:L} \\in \\mathbb{R}^{L \\times 6}" + }, + { + "bbox": [ + 313, + 474, + 554, + 617 + ], + "type": "text", + "content": " with the other driving conditions, and then map them to " + }, + { + "bbox": [ + 313, + 474, + 554, + 617 + ], + "type": "inline_equation", + "content": "c_t^{1:L} \\in \\mathbb{R}^{L \\times h}" + }, + { + "bbox": [ + 313, + 474, + 554, + 617 + ], + "type": "text", + "content": ". We also experiment on image-driven emotion [56] for frame-wise emotion control rather than the long-term emotion enhancement. FLOAT can effectively accommodate these additional conditions, highlighting its flexibility across diverse control signals." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": "Redirecting Speech-driven Emotion Since FLOAT learns diverse emotions in the emotion-intensive data distribution [39], the generated emotion-aware motion can be modified by redirecting the speech-driven emotion label toward a different emotion at inference time. As illustrated in Fig. 8, this technique is particularly beneficial for manual redirection when the emotion predicted from speech is complex or ambiguous." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14705" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 71, + 298, + 207 + ], + "blocks": [ + { + "bbox": [ + 70, + 71, + 298, + 207 + ], + "lines": [ + { + "bbox": [ + 70, + 71, + 298, + 207 + ], + "spans": [ + { + "bbox": [ + 70, + 71, + 298, + 207 + ], + "type": "image", + "image_path": "6ca3a17d0829e37887e40238950c7f077f253cdc06d11a9d01aa8966bbe9dd95.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 213, + 555, + 236 + ], + "lines": [ + { + "bbox": [ + 55, + 213, + 555, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 555, + 236 + ], + "type": "text", + "content": "Figure 7. Additional conditioning results of FLOAT. 3Dpose, " + }, + { + "bbox": [ + 55, + 213, + 555, + 236 + ], + "type": "inline_equation", + "content": "S2E" + }, + { + "bbox": [ + 55, + 213, + 555, + 236 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 213, + 555, + 236 + ], + "type": "inline_equation", + "content": "I2E" + }, + { + "bbox": [ + 55, + 213, + 555, + 236 + ], + "type": "text", + "content": " denote 3D head pose parameters [12], speech-to-emotion [49], and image-to-emotion [56], respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 300, + 72, + 538, + 207 + ], + "blocks": [ + { + "bbox": [ + 300, + 72, + 538, + 207 + ], + "lines": [ + { + "bbox": [ + 300, + 72, + 538, + 207 + ], + "spans": [ + { + "bbox": [ + 300, + 72, + 538, + 207 + ], + "type": "image", + "image_path": "af02109081a547b53339d991fc0888c4352bbee856a7594687667110672402fa.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 57, + 289, + 293, + 326 + ], + "blocks": [ + { + "bbox": [ + 55, + 241, + 296, + 286 + ], + "lines": [ + { + "bbox": [ + 55, + 241, + 296, + 286 + ], + "spans": [ + { + "bbox": [ + 55, + 241, + 296, + 286 + ], + "type": "text", + "content": "Table 2. Quantitative results of FLOAT with additional conditions (HDTF [85] / RAVDESS [39]). " + }, + { + "bbox": [ + 55, + 241, + 296, + 286 + ], + "type": "inline_equation", + "content": "S2E" + }, + { + "bbox": [ + 55, + 241, + 296, + 286 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 241, + 296, + 286 + ], + "type": "inline_equation", + "content": "I2E" + }, + { + "bbox": [ + 55, + 241, + 296, + 286 + ], + "type": "text", + "content": ", and 3DPose denote speech-to-emotion [49], image-to-emotion [56], and 3DMM pose parameters [12], respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 289, + 293, + 326 + ], + "lines": [ + { + "bbox": [ + 57, + 289, + 293, + 326 + ], + "spans": [ + { + "bbox": [ + 57, + 289, + 293, + 326 + ], + "type": "table", + "html": "
ConfigurationsFID ↓FVD ↓E-FID ↓P-FID ↓LSE-D ↓
AFLOAT(Ours)21.100 / 31.681162.052 / 166.3591.229 / 1.3670.032 / 0.0317.290 / 6.994
BA + 3Dpose19.721 / 29.721126.663 / 112.8940.926 / 1.1520.012 / 0.0167.516 / 7.047
CA - S2E21.235 / 32.035155.032 / 166.8661.254 / 1.5020.031 / 0.0257.264 / 7.222
DA - S2E + I2E21/528 / 31.609158.577 / 162.3691.158 / 1.3050.034 / 0.0227.183 / 7.150
", + "image_path": "28f033061734716b9434f2325442c02abf8bcef3a6bb0e062ee3907bd5dd31d3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 57, + 331, + 293, + 405 + ], + "blocks": [ + { + "bbox": [ + 57, + 331, + 293, + 405 + ], + "lines": [ + { + "bbox": [ + 57, + 331, + 293, + 405 + ], + "spans": [ + { + "bbox": [ + 57, + 331, + 293, + 405 + ], + "type": "image", + "image_path": "7e0ea533f52484b01bbefc6180adff4b2c616f773c8e0a9eb1c83bcb066f4fa5.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 410, + 296, + 433 + ], + "lines": [ + { + "bbox": [ + 55, + 410, + 296, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 296, + 433 + ], + "type": "text", + "content": "Figure 8. Redirecting the unclear emotion prediction to a desirable one-hot encoding, which can be further intensified by the CFV." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 443, + 156, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 443, + 156, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 156, + 455 + ], + "type": "text", + "content": "5.5. Ablation Studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 461, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 295, + 581 + ], + "type": "text", + "content": "Ablation on Frame-wise AdaLN We compare frame-wise AdaLN (and gating) followed by masked self-attention to separate conditioning from attending, with a cross-attention that performs conditioning and attending simultaneously. As shown in Tab. 3, both approaches achieve competitive image and video quality, while frame-wise AdaLN provides better expression generation and lip synchronization. We observe that frame-wise AdaLN can achieve more diverse head motions than the cross-attention. Please refer to supplementary videos." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "Ablation on Flow Matching We compare flow matching with two types of diffusion models: " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "-prediction (noise) and " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "-prediction (signal) [52, 67]. In both cases, we adopt our vector predictor architecture as denoising networks. We adopt diffusion training settings of VASA-1 [80] (500 diffusion steps with a cosine noise scheduler [46] and 50 DDIM denoising steps) for the indirect comparison with [80]. Notably, diffusion and flow matching achieve competitive results on image quality while the latter achieves the better lip synchronization. In Fig. 9, we compare the forward pass efficiency by measuring frames per second (FPS) of each" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 321, + 270, + 545, + 316 + ], + "blocks": [ + { + "bbox": [ + 313, + 241, + 555, + 264 + ], + "lines": [ + { + "bbox": [ + 313, + 241, + 555, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 241, + 555, + 264 + ], + "type": "text", + "content": "Table 3. Ablation studies of FLOAT on HDTF [85]. The best result for each metric is in bold, and the second-best result is underlined." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 270, + 545, + 316 + ], + "lines": [ + { + "bbox": [ + 321, + 270, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 321, + 270, + 545, + 316 + ], + "type": "table", + "html": "
MethodFID ↓FVD ↓E-FID ↓LSE-D ↓#NFEs ↓
Ours (w. Cross-Attn.)21.873162.7021.4527.75710
Ours (w. Diff., c-pred.)21.190161.6661.2139.92250
Ours (w. Diff., x0-pred.)21.697162.8471.2789.04850
FLOAT (Ours)21.100162.0521.2297.29010
", + "image_path": "8d462d551dd1b2925762d737f5b543dda08baae00af25d2a0f9735f6e95d2541.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 323, + 319, + 545, + 387 + ], + "blocks": [ + { + "bbox": [ + 323, + 319, + 545, + 387 + ], + "lines": [ + { + "bbox": [ + 323, + 319, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 323, + 319, + 545, + 387 + ], + "type": "image", + "image_path": "41c53dddeebfba3786224a8801316f9ace4e5d7e82874f9a46e3782adc4c2a59.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 390, + 555, + 411 + ], + "lines": [ + { + "bbox": [ + 313, + 390, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 555, + 411 + ], + "type": "text", + "content": "Figure 9. Comparison of the forward pass efficiency. We compute FPS on a single NVIDIA V100 GPU." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 426, + 555, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 473 + ], + "type": "text", + "content": "model. Thanks to the compact motion latent representation and OT-based flow matching, FLOAT achieves the highest FPS, superior lip-sync performance, dynamic head motion, and the lowest NFEs." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 495, + 388, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 495, + 388, + 507 + ], + "spans": [ + { + "bbox": [ + 314, + 495, + 388, + 507 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 518, + 555, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 555, + 674 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 555, + 674 + ], + "type": "text", + "content": "We proposed FLOAT, a flow matching based audio-driven talking portrait generation model leveraging a learned motion latent space. We introduced a transformer-based vector field predictor, enabling temporally consistent motion generation. Additionally, we incorporated speech-driven emotion labels into the motion sampling process to improve the naturalness of the audio-driven talking motions. FLOAT addresses current core limitations of diffusion-based talking portrait video generation methods by reducing the sampling time through flow matching while achieving the remarkable sample quality. Extensive experiments verified that FLOAT achieves state-of-the-art performance in terms of visual quality, motion fidelity, and efficiency." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "Discussion We leave further discussion considering limitations, future work, and ethical considerations in the supplementary materials." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14706" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449-12460, 2020. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 180 + ], + "type": "text", + "content": "[2] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques, pages 187-194, 1999. 3, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 182, + 295, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 295, + 236 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 295, + 236 + ], + "type": "text", + "content": "[3] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18392-18402, 2023. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 238, + 295, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 238, + 295, + 281 + ], + "spans": [ + { + "bbox": [ + 62, + 238, + 295, + 281 + ], + "type": "text", + "content": "[4] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 284, + 295, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 284, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 62, + 284, + 295, + 327 + ], + "type": "text", + "content": "[5] Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2d & 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In International Conference on Computer Vision, 2017. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 329, + 295, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 329, + 295, + 372 + ], + "spans": [ + { + "bbox": [ + 62, + 329, + 295, + 372 + ], + "type": "text", + "content": "[6] Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations. Advances in neural information processing systems, 31, 2018. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 374, + 295, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 374, + 295, + 418 + ], + "spans": [ + { + "bbox": [ + 62, + 374, + 295, + 418 + ], + "type": "text", + "content": "[7] Zhiyuan Chen, Jiajiong Cao, Zhiquan Chen, Yuming Li, and Chenguang Ma. Echomimic: Lifelike audio-driven portrait animations through editable landmark conditions. arXiv preprint arXiv:2407.08136, 2024. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 420, + 295, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 420, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 62, + 420, + 295, + 474 + ], + "type": "text", + "content": "[8] Kun Cheng, Xiaodong Cun, Yong Zhang, Menghan Xia, Fei Yin, Mingrui Zhu, Xuan Wang, Jue Wang, and Nannan Wang. Videoretalking: Audio-based lip synchronization for talking head video editing in the wild. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 476, + 295, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 476, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 62, + 476, + 295, + 508 + ], + "type": "text", + "content": "[9] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In Asian Conference on Computer Vision, pages 251-263, 2016. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 510, + 295, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 510, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 57, + 510, + 295, + 542 + ], + "type": "text", + "content": "[10] Quan Dao, Hao Phung, Binh Nguyen, and Anh Tran. Flow matching in latent space. arXiv preprint arXiv:2307.08698, 2023. 2, 3, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 544, + 295, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 544, + 295, + 599 + ], + "spans": [ + { + "bbox": [ + 57, + 544, + 295, + 599 + ], + "type": "text", + "content": "[11] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4690-4699, 2019. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 601, + 295, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 295, + 666 + ], + "type": "text", + "content": "[12] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 668, + 295, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 668, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 57, + 668, + 295, + 700 + ], + "type": "text", + "content": "[13] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 703, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 703, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 703, + 295, + 713 + ], + "type": "text", + "content": "[14] Laurent Dinh, Jascha Sohl-Dickstein, and Samy Ben" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 74, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 94 + ], + "type": "text", + "content": "gio. Density estimation using real nvp. arXiv preprint arXiv:1605.08803, 2016. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 97, + 553, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 553, + 151 + ], + "type": "text", + "content": "[15] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2663-2671, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 152, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 553, + 217 + ], + "type": "text", + "content": "[16] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 220, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 220, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 220, + 553, + 262 + ], + "type": "text", + "content": "[17] Johannes S Fischer, Ming Gui, Pingchuan Ma, Nick Stracke, Stefan A Baumann, and Björn Ommer. Boosting latent diffusion with flow matching. arXiv preprint arXiv:2312.07360, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 264, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 264, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 317, + 264, + 553, + 308 + ], + "type": "text", + "content": "[18] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 309, + 553, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 309, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 317, + 309, + 553, + 385 + ], + "type": "text", + "content": "[19] Jiazhi Guan, Zhanwang Zhang, Hang Zhou, Tianshu Hu, Kaisiyuan Wang, Dongliang He, Haocheng Feng, Jingtuo Liu, Errui Ding, Ziwei Liu, et al. Stylesync: High-fidelity generalized and personalized lip sync in style-based generator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1505-1515, 2023. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 388, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 388, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 317, + 388, + 553, + 442 + ], + "type": "text", + "content": "[20] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 444, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 444, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 317, + 444, + 553, + 487 + ], + "type": "text", + "content": "[21] Tianyu He, Junliang Guo, Runyi Yu, Yuchi Wang, Jialiang Zhu, Kaikai An, Leyi Li, Xu Tan, Chunyu Wang, Han Hu, et al. Gaia: Zero-shot talking avatar generation. arXiv preprint arXiv:2311.15230, 2023. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 488, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 488, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 488, + 553, + 510 + ], + "type": "text", + "content": "[22] Geoffrey Hinton. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 512, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 512, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 512, + 553, + 544 + ], + "type": "text", + "content": "[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 545, + 555, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 545, + 555, + 589 + ], + "spans": [ + { + "bbox": [ + 317, + 545, + 555, + 589 + ], + "type": "text", + "content": "[24] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8153-8163, 2024. 2, 3, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 591, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 591, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 317, + 591, + 553, + 644 + ], + "type": "text", + "content": "[25] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Qianyi Wu, Wayne Wu, Feng Xu, and Xun Cao. Eamm: One-shot emotional talking face via audio-based emotion-aware motion model. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2, 3, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "type": "text", + "content": "[26] Jianwen Jiang, Chao Liang, Jiaqi Yang, Gaojie Lin, Tianyun Zhong, and Yanbo Zheng. Loopy: Taming audio-driven portrait avatar with long-term motion dependency. arXiv preprint arXiv:2409.02634, 2024. 1, 2, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 692, + 553, + 713 + ], + "type": "text", + "content": "[27] Minguk Kang, Richard Zhang, Connelly Barnes, Sylvain Paris, Suha Kwak, Jaesik Park, Eli Shechtman, Jun-Yan Zhu," + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14707" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "text", + "content": "and Taesung Park. Distilling diffusion models into conditional gans. arXiv preprint arXiv:2405.05967, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 97, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 97, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 57, + 97, + 296, + 140 + ], + "type": "text", + "content": "[28] Taekyung Ki and Dongchan Min. Stylelipsync: Style-based personalized lip-sync video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22841-22850, 2023. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 143, + 294, + 165 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 143, + 294, + 165 + ], + "spans": [ + { + "bbox": [ + 57, + 143, + 294, + 165 + ], + "type": "text", + "content": "[29] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 167, + 294, + 189 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 167, + 294, + 189 + ], + "spans": [ + { + "bbox": [ + 57, + 167, + 294, + 189 + ], + "type": "text", + "content": "[30] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 190, + 295, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 190, + 295, + 245 + ], + "spans": [ + { + "bbox": [ + 56, + 190, + 295, + 245 + ], + "type": "text", + "content": "[31] Tobias Kirschstein, Simon Giebenhain, and Matthias Nießner. Diffusionavatars: Deferred diffusion for high-fidelity 3d head avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5492, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 247, + 295, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 247, + 295, + 302 + ], + "spans": [ + { + "bbox": [ + 56, + 247, + 295, + 302 + ], + "type": "text", + "content": "[32] Matthew Le, Apoorv Vyas, Bowen Shi, Brian Karrer, Leda Sari, Rashel Moritz, Mary Williamson, Vimal Manohar, Yossi Adi, Jay Mahadeokar, et al. Voicebox: Text-guided multilingual universal speech generation at scale. Advances in neural information processing systems, 36, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 304, + 294, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 304, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 57, + 304, + 294, + 335 + ], + "type": "text", + "content": "[33] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. ArXiv e-prints, pages arXiv-1607, 2016. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 338, + 295, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 295, + 392 + ], + "type": "text", + "content": "[34] Senmao Li, Taihang Hu, Fahad Shahbaz Khan, Linxuan Li, Shiqi Yang, Yaxing Wang, Ming-Ming Cheng, and Jian Yang. Faster diffusion: Rethinking the role of unet encoder in diffusion models. arXiv preprint arXiv:2312.09608, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 395, + 294, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 395, + 294, + 428 + ], + "spans": [ + { + "bbox": [ + 57, + 395, + 294, + 428 + ], + "type": "text", + "content": "[35] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022. 2, 3, 4, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 430, + 295, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 430, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 56, + 430, + 295, + 474 + ], + "type": "text", + "content": "[36] Tao Liu, Feilong Chen, Shuai Fan, Chenpeng Du, Qi Chen, Xie Chen, and Kai Yu. Anitalker: Animate vivid and diverse talking faces through identity-decoupled facial motion encoding. arXiv preprint arXiv:2405.03121, 2024. 1, 3, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 475, + 294, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 475, + 294, + 508 + ], + "spans": [ + { + "bbox": [ + 56, + 475, + 294, + 508 + ], + "type": "text", + "content": "[37] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 510, + 295, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 510, + 295, + 554 + ], + "spans": [ + { + "bbox": [ + 56, + 510, + 295, + 554 + ], + "type": "text", + "content": "[38] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, et al. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. In The Twelfth International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 556, + 295, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 556, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 556, + 295, + 610 + ], + "type": "text", + "content": "[39] Steven R Livingstone and Frank A Russo. The ryerson audio-visual database of emotional speech and song (ravdess): A dynamic, multimodal set of facial and vocal expressions in north american english. PloS one, 13(5): e0196391, 2018. 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 613, + 295, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 295, + 666 + ], + "type": "text", + "content": "[40] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "text", + "content": "[41] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[42] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 118, + 554, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 173 + ], + "type": "text", + "content": "[43] Yifeng Ma, Suzhen Wang, Zhipeng Hu, Changjie Fan, Tangjie Lv, Yu Ding, Zhidong Deng, and Xin Yu. Styletalk: One-shot talking head generation with controllable speaking styles. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1896-1904, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "type": "text", + "content": "[44] Yifeng Ma, Shiwei Zhang, Jiayu Wang, Xiang Wang, Yingya Zhang, and Zhidong Deng. Dreamtalk: When expressive talking head generation meets diffusion probabilistic models. arXiv preprint arXiv:2312.09767, 2023. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "type": "text", + "content": "[45] Dongchan Min, Minyoung Song, Eunji Ko, and Sung Ju Hwang. Stylistalker: One-shot style-based audio-driven talking head video generation. arXiv preprint arXiv:2208.10922, 2022. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 264, + 554, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 554, + 307 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 554, + 307 + ], + "type": "text", + "content": "[46] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162-8171. PMLR, 2021. 2, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 309, + 555, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 555, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 555, + 363 + ], + "type": "text", + "content": "[47] Se Jin Park, Minsu Kim, Joanna Hong, Jeongsoo Choi, and Yong Man Ro. Synctalkface: Talking face generation with precise lip-syncing via audio-lip memory. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2062-2070, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 365, + 554, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 554, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 554, + 408 + ], + "type": "text", + "content": "[48] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 2, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 411, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 443 + ], + "type": "text", + "content": "[49] Leonardo Pepino, Pablo Riera, and Luciana Ferrer. Emotion recognition from speech using wav2vec 2.0 embeddings. arXiv preprint arXiv:2104.03502, 2021. 5, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 445, + 554, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 554, + 498 + ], + "type": "text", + "content": "[50] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models. arXiv preprint arXiv:2410.13720, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 501, + 554, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 554, + 554 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 554, + 554 + ], + "type": "text", + "content": "[51] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020. 1, 2, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 556, + 554, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 556, + 554, + 600 + ], + "spans": [ + { + "bbox": [ + 317, + 556, + 554, + 600 + ], + "type": "text", + "content": "[52] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 601, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 634 + ], + "type": "text", + "content": "[53] Danilo Rezende and Shakir Mohamed. Variational inference with normalizing flows. In International conference on machine learning, pages 1530-1538. PMLR, 2015. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 635, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 555, + 689 + ], + "type": "text", + "content": "[54] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 2, 3, 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 692, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 714 + ], + "type": "text", + "content": "[55] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour," + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14708" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 117 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 117 + ], + "type": "text", + "content": "Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 118, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 118, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 57, + 118, + 295, + 149 + ], + "type": "text", + "content": "[56] Andrey V Savchenko. Hsemotion: High-speed emotion recognition library. Software Impacts, 14:100433, 2022. 5, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 151, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 151, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 57, + 151, + 294, + 183 + ], + "type": "text", + "content": "[57] Maximilian Seitzer. pytorch-fid: FID Score for PyTorch. https://github.com/mseitzer/pytorch-fid, 2020. Version 0.3.0.7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 184, + 295, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 295, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 295, + 227 + ], + "type": "text", + "content": "[58] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. Advances in neural information processing systems, 32, 2019. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 228, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 228, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 57, + 228, + 294, + 259 + ], + "type": "text", + "content": "[59] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 261, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 261, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 57, + 261, + 294, + 293 + ], + "type": "text", + "content": "[60] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 294, + 294, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 294, + 338 + ], + "type": "text", + "content": "[61] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 338, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 338, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 57, + 338, + 294, + 370 + ], + "type": "text", + "content": "[62] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arXiv preprint arXiv:2303.01469, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 372, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 372, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 57, + 372, + 294, + 426 + ], + "type": "text", + "content": "[63] Michal Stypulkowski, Konstantinos Vougioukas, Sen He, Maciej Zięba, Stavros Petridis, and Maja Pantic. Diffused heads: Diffusion models beat gans on talking-face generation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5091-5100, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 427, + 294, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 427, + 294, + 481 + ], + "spans": [ + { + "bbox": [ + 56, + 427, + 294, + 481 + ], + "type": "text", + "content": "[64] Zhiyao Sun, Tian Lv, Sheng Ye, Matthieu Lin, Jenny Sheng, Yu-Hui Wen, Minjing Yu, and Yong-jin Liu. Diffpositalk: Speech-driven stylistic 3d facial animation and head pose generation via diffusion models. ACM Transactions on Graphics (TOG), 43(4):1-9, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 482, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 482, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 294, + 536 + ], + "type": "text", + "content": "[65] Shuai Tan, Bin Ji, and Ye Pan. Emmn: Emotional motion memory network for audio-driven emotional talking face generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22146-22156, 2023. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 537, + 294, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 581 + ], + "type": "text", + "content": "[66] Shuai Tan, Bin Ji, Mengxiao Bi, and Ye Pan. Edtalk: Efficient disentanglement for emotional talking head synthesis. In European Conference on Computer Vision, pages 398-416. Springer, 2025. 1, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 582, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 294, + 624 + ], + "type": "text", + "content": "[67] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations, 2023. 2, 5, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 625, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 669 + ], + "type": "text", + "content": "[68] Linrui Tian, Qi Wang, Bang Zhang, and Liefeng Bo. Emo: Emote portrait alive-generating expressive portrait videos with audio2video diffusion model under weak conditions. arXiv preprint arXiv:2402.17485, 2024. 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "text", + "content": "[69] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 7" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 95 + ], + "type": "text", + "content": "[70] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2, 4, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 97, + 554, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 554, + 151 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 554, + 151 + ], + "type": "text", + "content": "[71] Cong Wang, Kuan Tian, Jun Zhang, Yonghang Guan, Feng Luo, Fei Shen, Zhiwei Jiang, Qing Gu, Xiao Han, and Wei Yang. V-express: Conditional dropout for progressive training of portrait video generation. arXiv preprint arXiv:2406.02511, 2024. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 152, + 553, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 208 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 208 + ], + "type": "text", + "content": "[72] Kaisiyuan Wang, Qianyi Wu, Linsen Song, Zhuoqian Yang, Wayne Wu, Chen Qian, Ran He, Yu Qiao, and Chen Change Loy. Mead: A large-scale audio-visual dataset for emotional talking-face generation. In European Conference on Computer Vision, pages 700–717. Springer, 2020. 3, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "type": "text", + "content": "[73] Suzhen Wang, Lincheng Li, Yu Ding, Changjie Fan, and Xin Yu. Audio2head: Audio-driven one-shot talking-head generation with natural head motion. arXiv preprint arXiv:2107.09293, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 254, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 308 + ], + "type": "text", + "content": "[74] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10039-10049, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 309, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 353 + ], + "type": "text", + "content": "[75] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. Latent image animator: Learning to animate images via latent space navigation. arXiv preprint arXiv:2203.09043, 2022. 2, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 355, + 553, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 553, + 388 + ], + "type": "text", + "content": "[76] Huawei Wei, Zejun Yang, and Zhisheng Wang. Aniportrait: Audio-driven synthesis of photorealistic portrait animation. arXiv preprint arXiv:2403.17694, 2024. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "type": "text", + "content": "[77] Yibo Xia, Lizhen Wang, Xiang Deng, Xiaoyan Luo, and Yebin Liu. Gmtalker: Gaussian mixture based emotional talking video portraits. arXiv preprint arXiv:2312.07669, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 434, + 555, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 555, + 489 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 555, + 489 + ], + "type": "text", + "content": "[78] Liangbin Xie, Xintao Wang, Honglun Zhang, Chao Dong, and Ying Shan. Vfhq: A high-quality dataset and benchmark for video face super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 657-666, 2022. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 491, + 554, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 491, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 491, + 554, + 544 + ], + "type": "text", + "content": "[79] Mingwang Xu, Hui Li, Qingkun Su, Hanlin Shang, Liwei Zhang, Ce Liu, Jingdong Wang, Luc Van Gool, Yao Yao, and Siyu Zhu. Hallo: Hierarchical audio-driven visual synthesis for portrait image animation. arXiv preprint arXiv:2406.08801, 2024. 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 546, + 554, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 546, + 554, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 546, + 554, + 600 + ], + "type": "text", + "content": "[80] Sicheng Xu, Guojun Chen, Yu-Xiao Guo, Jiaolong Yang, Chong Li, Zhenyu Zang, Yizhong Zhang, Xin Tong, and Baining Guo. Vasa-1: Lifelike audio-driven talking faces generated in real time. arXiv preprint arXiv:2404.10667, 2024. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 602, + 554, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 554, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 554, + 668 + ], + "type": "text", + "content": "[81] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. In European conference on computer vision, pages 85-101. Springer, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "text", + "content": "[82] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14709" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 483 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "text", + "content": "[83] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 295, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 295, + 204 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 295, + 204 + ], + "type": "text", + "content": "[84] Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8652-8661, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 207, + 295, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 207, + 295, + 261 + ], + "spans": [ + { + "bbox": [ + 56, + 207, + 295, + 261 + ], + "type": "text", + "content": "[85] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3661–3670, 2021. 6, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 262, + 295, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 262, + 295, + 317 + ], + "spans": [ + { + "bbox": [ + 56, + 262, + 295, + 317 + ], + "type": "text", + "content": "[86] Zhimeng Zhang, Zhipeng Hu, Wenjin Deng, Changjie Fan, Tangjie Lv, and Yu Ding. Dinet: Deformation inpainting network for realistic face visually dubbing on high resolution video. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3543-3551, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 319, + 295, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 295, + 384 + ], + "type": "text", + "content": "[87] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4176–4186, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 385, + 295, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 385, + 295, + 428 + ], + "spans": [ + { + "bbox": [ + 56, + 385, + 295, + 428 + ], + "type": "text", + "content": "[88] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. Makelttalk: speaker-aware talking-head animation. ACM Transactions On Graphics (TOG), 39(6):1-15, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 430, + 295, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 430, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 56, + 430, + 295, + 483 + ], + "type": "text", + "content": "[89] Shenhao Zhu, Junming Leo Chen, Zuozhuo Dai, Yinghui Xu, Xun Cao, Yao Yao, Hao Zhu, and Siyu Zhu. Champ: Controllable and consistent human image animation with 3d parametric guidance. arXiv preprint arXiv:2403.14781, 2024. 4" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14710" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/FLOSS_ Free Lunch in Open-vocabulary Semantic Segmentation/759adae1-3d57-4b38-8780-a8c8f1ca245e_content_list.json b/2025/FLOSS_ Free Lunch in Open-vocabulary Semantic Segmentation/759adae1-3d57-4b38-8780-a8c8f1ca245e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9cfe9a3434a6ba475f210bcd606259f3bef95d6c --- /dev/null +++ b/2025/FLOSS_ Free Lunch in Open-vocabulary Semantic Segmentation/759adae1-3d57-4b38-8780-a8c8f1ca245e_content_list.json @@ -0,0 +1,1616 @@ +[ + { + "type": "text", + "text": "FLOSS: Free Lunch in Open-vocabulary Semantic Segmentation", + "text_level": 1, + "bbox": [ + 169, + 130, + 826, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yasser Benigim $^{1*}$ Mohammad Fahes $^{1}$ Tuan-Hung $\\mathrm{Vu}^{1,2}$ Andrei Bursuc $^{1,2}$ Raoul de Charette $^{1}$ Inria $^{2}$ Valeo.ai", + "bbox": [ + 101, + 178, + 916, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 250, + 326, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we challenge the conventional practice in Open-Vocabulary Semantic Segmentation (OVSS) of using averaged class-wise text embeddings, which are typically obtained by encoding each class name with multiple templates (e.g., a photo of , a sketch of a ). We investigate the impact of templates for OVSS, and find that for each class, there exist single-template classifiers—which we refer to as class-experts—that significantly outperform the conventional averaged classifier. First, to identify these class-experts, we introduce a novel approach that estimates them without any labeled data or training. By leveraging the class-wise prediction entropy of single-template classifiers, we select those yielding the lowest entropy as the most reliable class-experts. Second, we combine the outputs of class-experts in a new fusion process. Our plug-and-play method, coined FLOSS, is orthogonal and complementary to existing OVSS methods, offering an improvement without the need for additional labels or training. Extensive experiments show that FLOSS consistently enhances state-of-the-art OVSS models, generalizes well across datasets with different distribution shifts, and delivers substantial improvements in low-data scenarios where only a few unlabeled images are available. Our code is available at https://github.com/yasserben/FLOSS.", + "bbox": [ + 86, + 282, + 483, + 659 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 686, + 220, + 702 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the past decade, advances in deep learning and the growing amount of training data have enabled the challenging task of semantic segmentation, which involves assigning semantic labels to each pixel in an image. Initially, this was limited to predefined categories [9, 10, 57, 69], but recent models rely on vision-language alignment [47] for open-vocabulary segmentation, where the semantic categories can be dynamically expressed at runtime in natural language.", + "bbox": [ + 89, + 712, + 482, + 848 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In Open-Vocabulary Semantic Segmentation (OVSS), the CLIP model [47] is commonly used as the backbone,", + "bbox": [ + 89, + 848, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3a640d3901aa085653a9588d6e8c542193f0e637180093783acb282b862cd372.jpg", + "image_caption": [ + "Figure 1. Using individual vs average template classifiers. Empirically, we observe that for each class there exist individual templates (colored dots) that lead to classifiers performing better than the popular CLIP classifier, which is built from the averaging of all 80 templates embeddings $(-)$ . The analysis was conducted on Cityscapes using the CLIP-DINOiser model." + ], + "image_footnote": [], + "bbox": [ + 513, + 250, + 906, + 407 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "which is either retrained [43], fine-tuned [12, 38], or frozen and adjusted to extend global image-language alignment to denser pixel-language alignment. In this work, we study CLIP's original classification paradigm and focus on methods that keep all CLIP parameters frozen [23, 63, 79]. Among these, MaskCLIP [79] leverages CLIP's embedding layer for segmentation, CLIP-DINOiser [66] enhances features via distillation from an external DINO [6] network, while NACLIP [23] modifies CLIP's self-attention to improve spatial consistency. A common aspect of prior OVSS methods is their reliance on CLIP's default classification process, where image or patch embeddings are compared with text embeddings representing semantic categories. To represent a class, a common practice is to use multiple templates to encode class names (e.g., \"a photo of a car\", \"a sketch of a car\", etc) and average their embeddings. Empirically, Radford et al. [47] engineered $M = 80$ templates to improve ImageNet zero-shot classification. For segmentation, OVSS methods [23, 66, 79] adopt by default the same templates to construct segmentation classifiers. While this \"template-averaging\" strategy works well overall, it may not always provide the best classifier for each semantic concept.", + "bbox": [ + 511, + 529, + 906, + 877 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1\"Template-averaging\" is an imprecise but convenient shorthand.", + "bbox": [ + 531, + 886, + 877, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done during an internship at Inria.", + "bbox": [ + 107, + 886, + 328, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21471", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our paper explores the impact of individual templates in OVSS – an aspect that has been overlooked in the literature. Fig. 1 shows an intriguing empirical observation: for each class, some individual templates outperform the averaged classifier that uses all templates. We call such templates class-experts, highlighting that they are trivial to identify using semantic ground truths.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, without labels, two challenges arise: 1) How to identify class-experts without ground truth? 2) How to fuse predictions from class-experts?", + "bbox": [ + 89, + 198, + 482, + 243 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose that entropy, an unsupervised metric, can help identify expert templates without labels. This leads to a new task—training- and label-free template selection for OVSS—which improves segmentation by selecting better text template subsets for a given dataset of unlabeled images. Our method shifts the attention to the text modality, unlike previous works that tweak the visual encoder, and also aims to improve the performance in an inductive setting, where the selected templates are applied to a different subset of images.", + "bbox": [ + 86, + 244, + 482, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We summarize our contributions as follows:", + "bbox": [ + 89, + 395, + 383, + 409 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We analyze the current template-averaging practice for OVSS, showing that better class-wise classifiers can be constructed using only a subset of all templates, and that these expert templates can be identified without labels.", + "- We propose the novel task of training- and label-free template selection for OVSS. Given an unlabeled set of images, this task aims to identify class-wise experts that improve OVSS performance on unseen counterparts of the dataset.", + "- We introduce FLOSS, which leverages class-wise per-pixel entropy to select expert templates and employs a simple and effective fusion scheme to combine individual expert predictions." + ], + "bbox": [ + 89, + 411, + 482, + 606 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With extensive experiments, we demonstrate that FLOSS consistently improves the state-of-the-art performance on OVSS benchmarks and exhibits generalization capabilities in the choice of experts. Furthermore, with only a few unlabeled images, our approach can select reasonably good expert templates, resulting in performance boost.", + "bbox": [ + 89, + 608, + 482, + 699 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 714, + 238, + 729 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision-Language Model (VLM) transferability. Since CLIP's release, its impressive zero-shot classification performance [47] and robustness [41, 58, 59] have made it a foundation for various adaptation and transfer learning tasks. Few-shot supervised fine-tuning, where only limited labeled data is available for downstream tasks like image classification, has gained significant attention. The goal is to leverage the vast knowledge of the VLM acquired through training on an immense image collection by steering it towards a task of interest via few labeled sam", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "plies. Parameter efficiency is central to this setting. To ensure it, prompt learning [8, 33, 80-82], lightweight MLP training [19], and linear probing [29, 55] have been proposed. Another line of research explores training-free few-shot adaptation [64, 77, 83], where a classifier is built using few labeled samples and ensembled with the zero-shot classifier. Additionally, unsupervised adaptation has been explored [26, 28, 56], alongside test-time adaptation, where prompts are learned per test image [54].", + "bbox": [ + 511, + 90, + 903, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While using VLMs for classification tasks aligns with their pre-training focus on global image recognition, applying them to dense prediction tasks like semantic segmentation is more challenging due to the gap between global image understanding and pixel-level prediction. We summarize related efforts in the following.", + "bbox": [ + 511, + 229, + 903, + 321 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Open-Vocabulary Semantic Segmentation (OVSS). OVSS uses VLMs like CLIP for text-based segmentation. However, CLIP's global pooling layers limit its ability to generate dense pixel-level, language-aligned features [31, 79]. To address this, several training-free methods densify CLIP without altering its parameters, preserving its image-language alignment. MaskCLIP [79] was the first to replace the attention pooling layer with a convolutional layer. Later works further densify CLIP by aggregating multiple image views [31, 66], integrating priors from other vision foundation models for better pooling [35, 66], extracting different information from the self-attention layers [3, 23, 37], or removing anomaly tokens [2]. Seg-TTO [14] leverages an LLM to generate discriminative attributes for rare classes in specialized domains (e.g., medical sciences, earth monitoring) that use technical terminology (e.g., mediastinum), operating in a test-time optimization framework while maintaining zero-shot capabilities. Other approaches fine-tune CLIP for segmentation tasks using weak supervision through image-level tags, captions [21, 48], or class-agnostic object masks [21, 49]. Others employ full supervision [12, 36, 38, 74] on densely annotated datasets like COCO Stuff [5], following strategies similar to early inductive zero-shot semantic segmentation methods [4, 67]. Among them, some fine-tune CLIP's image encoder [36, 38, 74], others combine this with additional backbones and segmentation heads [38], while some learn side networks while keeping CLIP frozen [74]. Many other approaches have been proposed in this active research area [24, 68, 72, 73, 76]. While effective, these methods require labeled data and computational resources, limiting their practicality in resource-constrained scenarios.", + "bbox": [ + 511, + 338, + 903, + 837 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our work focuses on training-free approaches, evaluating FLOSS on three key methods: MaskCLIP [79], CLIP-DINOiser [66] and NACLIP [23], while being applicable to any text-based segmentation framework.", + "bbox": [ + 511, + 839, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21472", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prompting in VLMs. CLIP's image-language interface has inspired the adoption of various prompt engineering practices from LLMs to better reveal the knowledge encapsulated in the vision encoder and boost performance. The seminal CLIP work [47] introduced dataset-specific templates (e.g., \"a photo of \") to bridge the gap between training captions and test prompts for zero-shot classification. Subsequent methods have improved prompts by crowdsourcing templates [1], adjusting class names [21, 60], or modifying templates for tasks like object detection [21, 22] and occupancy prediction [60]. While effective, manual prompt tuning is labor-intensive, leading to automated strategies like adding random words or characters to the original template [50], enriching class names with WordNet concepts [20] or synonyms [39], learning more informative class names [27], and using LLMs to generate more expressive descriptions starting from simple class names or retrieve them from a dataset [51] and further use them as prompts. Some approaches also learn prompts for task-specific fitting [45] or test-time adaptation [54], and others mine them with external LLMs [16].", + "bbox": [ + 89, + 90, + 483, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our approach differs by selecting expert templates for each class in an unsupervised manner, starting from predefined templates. While previous methods have focused on image classification, we apply this strategy to semantic segmentation using CLIP's original ImageNet templates, which have been used by default in prior works.", + "bbox": [ + 89, + 412, + 483, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 530, + 230, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "OVSS aims to produce a soft-segmentation map $\\hat{\\mathbf{S}}\\in [0,1]^{H\\times W\\times K}$ given an image $\\mathbf{l}\\in \\mathbb{R}^{H\\times W\\times 3}$ and a set of classes $\\{\\mathcal{C}_k\\}_{k\\in [1,K]}$ expressed in natural language. The final segmentation map $\\hat{\\mathbf{P}}$ is derived by applying arg max(·) on $\\hat{\\mathbf{S}}$ along the $K$ -dimension. Leveraging the aligned vision encoder $f(\\cdot)$ and text encoder $g(\\cdot)$ of CLIP, the standard practice consists in classifying an image patch by comparing its vision embedding to the text embedding of class-specific prompts constructed using predefined templates $\\{\\mathcal{T}_m\\}_{m\\in [1,M]}$ . An example of prompt is \"a bright photo of a car\", denoted", + "bbox": [ + 89, + 561, + 483, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\mathcal{P}_{m,k}$ for compactness. Instead of a single template, practitioners use \"template-averaging\" to increase robustness and improve overall performance by averaging the embeddings of prompts obtained using all available templates [40, 46, 47, 50], which is a form of ensembling on the text encoder side. This writes $\\bar{c}_k = \\frac{1}{M}\\sum_{m = 1}^{M}g(\\mathcal{P}_{m,k})$ where $\\bar{c}_k$ is the resulting representation of class $C_k$ . Subsequently, for each patch, the predicted class is the one that maximizes the cosine similarity between the visual and the", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "text embeddings:", + "bbox": [ + 513, + 90, + 630, + 106 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {k} = \\arg \\max _ {k} f (\\mathcal {V}) ^ {\\mathrm {T}} \\bar {c} _ {k}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 109, + 906, + 133 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{V}$ represents the visual patch. All embeddings are $\\ell_2$ -normalized. The text embeddings of $K$ classes define a classifier that we denote:", + "bbox": [ + 511, + 138, + 906, + 183 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {W} \\left(\\left\\{\\mathcal {T} _ {m} \\right\\} _ {m \\in [ 1, M ]}\\right) = \\left\\{\\bar {c} _ {k} \\right\\} _ {k \\in [ 1, K ]}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 190, + 906, + 208 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As in the above equation, the classifier can be represented as a function of the templates used to compute the average embeddings of all classes.", + "bbox": [ + 511, + 212, + 906, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Empirical observations on single templates. OVSS models [23, 66, 68, 79] typically rely on the $M = 80$ ImageNet templates of CLIP [47] to compute the average classifier of Eq. (2), which we refer to as $W_{\\mathrm{CLIP}}$ for brevity. While using $W_{\\mathrm{CLIP}}$ was shown to be robust [47, 79], the templates were originally hand-crafted for zero-shot classification on ImageNet [47], and their individual effect on class-wise performance remains unexplored for OVSS. Rather than $W_{\\mathrm{CLIP}}$ that uses all $M$ templates, one can construct $M$ distinct classifiers, each utilizing a single template $\\mathcal{T}_m$ . This writes:", + "bbox": [ + 511, + 262, + 906, + 426 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {W} \\left(\\mathcal {T} _ {m}\\right) = \\left\\{g \\left(\\mathcal {P} _ {m, k}\\right) \\right\\} _ {k \\in [ 1, K ]}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 434, + 903, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We design a systematic experiment to evaluate the performance of single-template classifiers $\\mathbf{W}(\\mathcal{T}_m)$ from Eq. (3). Fig. 1 shows the per-class performance on Cityscapes [13] of each of the 80 single-template classifiers (plotted as colored dots, e.g., ●), where colors encode the template identifier (i.e., m), along with the average $W_{\\mathrm{CLIP}}$ classifier (plotted as —). This leads to two key observations:", + "bbox": [ + 511, + 455, + 905, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "a) For each class, there exist single-template classifiers outperforming $W_{\\mathrm{CLIP}}$ on this class; we refer to the corresponding templates as class-experts. More precisely, the set of class-experts is defined as:", + "bbox": [ + 511, + 577, + 906, + 637 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} _ {k} = \\left\\{\\mathcal {T} _ {m} \\mid \\Omega_ {k} \\left(\\boldsymbol {W} \\left(\\mathcal {T} _ {m}\\right)\\right) > \\Omega_ {k} \\left(\\boldsymbol {W} _ {\\text {C L I P}}\\right), m \\in [ 1, M ] \\right\\}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 643, + 903, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\Omega_{k}(\\cdot)$ denotes the IoU performance on class $k$ .", + "bbox": [ + 511, + 674, + 867, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "b) A single-template classifier excelling on one class may exhibit suboptimal performance on others, implying that for $k, l \\in [1, K]$ the expert sets $\\mathcal{E}_k$ and $\\mathcal{E}_l$ may differ.", + "bbox": [ + 511, + 689, + 905, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following the above observations, class-experts are readily identified when ground truth labels are accessible, through direct performance comparison. We refer readers to Appendix A.3 for similar observations across datasets and models.", + "bbox": [ + 511, + 734, + 905, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, we are interested in the case where an unlabeled dataset is provided. Therefore, two questions stem from the previous empirical observations: (i) How to identify class-experts in an unsupervised, training-free manner? (ii) Given these identified class-experts, how to derive the final semantic predictions?", + "bbox": [ + 511, + 810, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21473", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8894a9fd171e353ebbc1bfe019331e61c5666bae754d5c2136c107ed1ca1a873.jpg", + "image_caption": [ + "Unsupervised identification of class-experts" + ], + "image_footnote": [], + "bbox": [ + 96, + 107, + 488, + 213 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8e8322105906b9164149c821e09e9c34d9f4406143d13ba5d567163264ce7c59.jpg", + "image_caption": [ + "Expert Fusion", + "Figure 2. Overview of FLOSS. Left: for each class, selected expert templates $(\\checkmark)$ construct the estimated set $\\hat{\\mathcal{E}}$ , which classifies the input image and yield the soft-segmentation map $\\mathbf{S}$ . In this example we set $N = 3$ for the $\\mathrm{Top - N}(\\dots)$ template indices corresponding to lowest entropy entries to select. Right: we explain the fusion strategy for two cases. For the pixel $x_{1}$ , the set of experts that predict their own class of expertise $\\mathcal{X}(x_1)$ contains 2 experts; consequently, the decision is taken by looking at the maximum softmax scores of those two, leading to the decision of classifying $x_{1}$ as \"person\" with the higher score. For the pixel $x_{2}$ , there is no expert predicting its own class of expertise, i.e., $\\mathcal{X}(x_2) = \\emptyset$ , and the decision is taken by the default classifier $\\mathbf{W}_{\\mathrm{CLIP}}$ ; the pixel $x_{2}$ thus takes the scores from $\\mathbf{S}_{\\mathrm{CLIP}}(x_2)$ ." + ], + "image_footnote": [], + "bbox": [ + 506, + 109, + 903, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 89, + 349, + 181, + 366 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given access to a dataset $\\mathcal{D}$ of unlabeled images, our goal is to improve the performance of OVSS using the aforementioned class-experts, without training or access to any labels. To do so, for each class $k\\in [1,K]$ our method first identifies a small set of class-experts among the $M$ predefined ImageNet templates of CLIP (Sec. 4.1) by leveraging unsupervised metrics. It then builds on a simple scheme to fuse class-experts' predictions into a single OVSS semantic output (Sec. 4.2).", + "bbox": [ + 89, + 377, + 483, + 512 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Unsupervised identification of class-experts", + "text_level": 1, + "bbox": [ + 89, + 523, + 457, + 540 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Referring to our definition of class-experts in Sec. 3, for each of the $K$ classes, our goal is to estimate a set of class-experts $\\hat{\\mathcal{E}}_k$ , among the $M$ pre-defined CLIP templates. Being in an unsupervised setting, we have no access to the true classifier performance (i.e., $\\Omega(\\cdot)$ in Eq. (4)) and therefore propose relying on unsupervised metrics which act as performance proxy [62, 70]. We leverage entropy as it is an established measure of uncertainty [18, 30], commonly adopted in deep learning literature [32, 34, 61]. The entropy value of a softmax prediction indicates a classifier's confidence, which empirically provides a reliable indicator of prediction quality. In our setting, entropy measures the confidence of single-template classifiers for each class, helping identify class-experts. We also study other unsupervised metrics in the ablations (Sec. 5.2).", + "bbox": [ + 89, + 546, + 482, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In detail, we estimate class-experts as the set of individual templates whose associated classifiers (cf., Eq. (3)) exhibit low entropy. Specifically, for each template $m$ and class $k$ , we compute the average entropy $\\mathcal{H}_{m,k}$ of the $C_{m,k}$ pixels predicted as class $k$ by $\\pmb{W}(\\mathcal{T}_m)$ in all images of $\\mathcal{D}$ :", + "bbox": [ + 89, + 773, + 482, + 849 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {H} _ {m, k} = - \\frac {1}{C _ {m , k}} \\sum_ {i = 1} ^ {C _ {m, k}} \\mathbf {q} _ {i} ^ {\\mathrm {T}} \\log (\\mathbf {q} _ {i}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 862, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{q}_i\\in \\Delta^{K - 1}$ is pixel-wise probability prediction obtained by applying softmax(·) over $K$ class-wise cosine similarities, and $\\Delta^{K - 1}$ is the $(K - 1)$ -simplex in $\\mathbb{R}^K$ . Subsequently, for class $k$ , the estimated set of experts is obtained as the $N$ templates yielding the lowest entropy:", + "bbox": [ + 511, + 351, + 906, + 429 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {E}} _ {k} = \\left\\{\\mathcal {T} _ {\\hat {m}} \\mid \\hat {m} \\in \\operatorname {T o p - N} (\\underset {m} {\\arg \\operatorname {s o r t}} \\mathcal {H} _ {m, k}) \\right\\}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 439, + 903, + 463 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{\\mathcal{E}}_k$ is expected to be an estimation of $\\mathcal{E}_k$ (Eq. (4)); the arg sort operator returns the sorted list of $M$ template indices so that $\\mathcal{H}_{m,k}$ are in ascending order. Finally, $\\mathrm{Top - N}(\\cdot)$ thus returns $\\mathbf{N}$ template indices corresponding to the lowest entropy entries. In practice, we use $N = 4$ and validate this choice through an ablation study in our experimental section. Following Eq. (2), the expert classifier of class $k$ , denoted $\\pmb {W}(\\hat{\\mathcal{E}}_k)$ , is constructed by averaging the embeddings obtained from the $N$ templates in $\\hat{\\mathcal{E}}_k$ for each of the $K$ class names.", + "bbox": [ + 511, + 476, + 906, + 631 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The above process leads to $K$ classifiers $\\{\\pmb{W}(\\hat{\\mathcal{E}}_k)\\}_{k\\in [1,K]}$ , all obtained in a fully unsupervised manner, and each is expected to excel on one specific class. However, it is important to note that although a classifier is expert of a single class, it outputs a full semantic map containing all $K$ classes. Therefore, the above formulation produces $K$ soft-segmentation maps, $\\{\\hat{\\mathbf{S}}_k\\}_{k\\in [1,K]}$ . To output a single OVSS map, we introduce a simple scheme to fuse all expert predictions.", + "bbox": [ + 511, + 632, + 906, + 771 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Fusion of expert predictions", + "text_level": 1, + "bbox": [ + 511, + 781, + 764, + 797 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To fuse all predicted soft-segmentation maps $\\{\\hat{\\mathbf{S}}_k\\}_{k\\in [1,K]}$ into a single output $\\hat{\\mathbf{S}}$ , we rely on the fact that each map $\\hat{\\mathbf{S}}_k$ excels at segmenting class $k$ . Therefore, for each pixel, the class assigned is the one having the highest probability among the expert classifiers which predicted their own class of expertise. Occasionally, when no expert predicts its own", + "bbox": [ + 511, + 804, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21474", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "class of expertise for a given pixel, we simply revert to using the $W_{\\mathrm{CLIP}}$ classifier. In practice, we observe that only $\\approx 2\\%$ of pixels fall under the latter case. Therefore, the vast majority of decisions are carried out by experts.", + "bbox": [ + 89, + 90, + 482, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Formally, for a given pixel $x$ , $\\hat{\\mathbf{S}}_k(x)$ is the $K$ -dimensional softmax probability of the expert $k$ at the position of $x$ . Let $\\mathcal{X}(x) = \\{k \\mid \\arg \\max_{i \\in [1, K]} \\hat{\\mathbf{S}}_k(x)[i] = k\\}$ be the index set of experts that predicted their own class of expertise for $x$ . The prediction of $x$ is determined as:", + "bbox": [ + 89, + 152, + 483, + 239 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {P}} (x) = \\left\\{ \\begin{array}{l l} k ^ {*} = \\underset {k \\in \\mathcal {X} (x)} {\\arg \\max } \\hat {\\mathbf {S}} _ {k} (x) [ k ], & \\text {i f} \\mathcal {X} (x) \\neq \\varnothing , \\\\ k ^ {\\dagger} = \\underset {k \\in [ 1, K ]} {\\arg \\max } \\mathbf {S} _ {\\operatorname {C L I P}} (x) [ k ], & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 250, + 480, + 318 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fig. 2 illustrates the unsupervised identification of class-experts as well as the fusion strategy. Overall, our method incurs computational overhead primarily from computing cosine similarities for the $K$ class-experts, which scales with the number of classes and becomes more significant when $K$ is very large ( $>100$ ). However, this overhead is constrained to the similarity computations since the vision encoder is shared across all experts, requiring only a single forward pass regardless of the number of classes. We provide computational details in Appendix A.2.", + "bbox": [ + 88, + 319, + 483, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 482, + 225, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Experimental setup. We evaluate our method on three state-of-the-art open-vocabulary semantic segmentation models: the seminal MaskCLIP [79] and two recent models CLIP-DINOiser [66] and NACLIP [23]. CLIP-DINOiser trains a convolution layer using only 1K unlabeled images from ImageNet [15] for fast inference on pixel-level segmentation tasks, effectively imitating the DINO priors for pooling guidance and removing the need for DINO encoder at runtime. MaskCLIP and NACLIP are training-free adaptation methods of CLIP.", + "bbox": [ + 89, + 508, + 483, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For a fair comparison, we report the version of NACLIP without refinement. We primarily select training-free models that keep the image encoder frozen and produce unaltered, dense image-language features. This enables us to assess the impact of the class-experts identification strategy by using the original CLIP pre-trained features and ImageNet templates. Unless otherwise stated, all models utilize CLIP ViT-B/16 backbone. CLIP-DINOiser and MaskCLIP are based on OpenCLIP [11], while NACLIP employs OpenAI's original CLIP [47]. For all models - CLIP-DINOiser, NACLIP, and MaskCLIP- we follow the authors' evaluation protocol. For CLIP-DINOiser and MaskCLIP, we resize images to 448 pixels on the shortest side with a sliding window of size 448 and stride 224. For NACLIP, images are resized to 336 pixels (except for Cityscapes which are resized to 560 pixels) with a 224 window and stride 112.", + "bbox": [ + 89, + 659, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f051b3f4d1281ea70252ffb49bf3af721a805db56259c88c8fbe96cc2aafbfcf.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCSVOC20PC59ADEStuffAvg
CLIP [47]6.749.111.23.25.715.2
GroupViT [71]11.179.723.49.215.327.7
CLIP-Surgery [37]31.4--12.921.9-
GEM [3]--32.615.7--
SCLIP [63]32.280.434.216.122.437.1
CLIP-DIY [65]11.679.719.89.913.326.9
TCL [7]23.177.530.314.919.633.1
ReCo [53]21.157.822.311.214.825.4
MaskCLIP [79]25.061.825.514.217.528.7
+ FLOSS25.861.826.214.917.829.3
NACLIP [23]35.579.735.217.423.338.2
+ FLOSS37.080.235.918.423.639.0
CLIP-DINOiser [66]31.380.935.920.024.638.5
+ FLOSS34.682.336.220.724.739.7
", + "bbox": [ + 514, + 88, + 905, + 295 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. OVSS across datasets and models. We report mIoU metric for eight OVSS baselines and three models either using the average templates (i.e., MaskCLIP, NACLIP, and CLIP-DINOiser) or with our method (i.e., +FLOSS). FLOSS consistently improves OVSS models across datasets of varying complexity, from urban scenes (Cityscapes, 19 classes) to general objects (COCO-Stuff, 171 classes). Bold highlights the best performance.", + "bbox": [ + 511, + 306, + 906, + 405 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets and Metric. We evaluate FLOSS across multiple semantic segmentation benchmarks: PASCAL CONTEXT 59 [42] (PC59), PASCAL VOC 20 [17] (VOC20), COCO-Stuff [5] (Stuff), ADE20K [78] (ADE), and Cityscapes [13] (CS). Note that these datasets exhibit different ontology and semantic granularity, ranging from 19 classes (CS) to 171 classes (Stuff). Additionally, we demonstrate the generalization capabilities of our class-experts identified on Cityscapes to other driving datasets such as BDD-100K [75] and MAPILLARY [44] as well as ACDC [52] under Night, Fog, Rain and Snow conditions. Importantly, our approach is training-free and uses no labels. Performance is measured using mean Intersection over Union (mIoU).", + "bbox": [ + 511, + 429, + 906, + 640 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. The only hyperparameter of FLOSS, being the number of class-experts to identify in Eq. (6), is set to $N = 4$ in all experiments. While our method is training-free, class-experts are estimated using the training set of each dataset, while performance is always reported on the corresponding validation set.", + "bbox": [ + 511, + 643, + 908, + 736 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Main results", + "text_level": 1, + "bbox": [ + 511, + 742, + 648, + 757 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Tab. 1 reports the mIoU of OVSS baselines across five datasets showing FLOSS plugged into three different models: MaskCLIP, NACLIP and CLIP-DINOiser. The performance gains vary with dataset complexity, showing smaller improvements on VOC20 $[+0.05, +0.5, +1.4]$ or PC59 $[+0.7, +0.7, +0.3]$ than on CS $[+0.8, +1.5, +3.3]$ . We attribute this to VOC20 and PC59 being closer to ImageNet which was leveraged in CLIP [47] to engineer the set of 80 templates used in all models. The benefit of our method", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21475", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f0227326c31fe3ebe4ddc064854bdd999b0c22b39b3ed846e1c041074c9c17c1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCSVOC20PC59ADEStuffAvg
MaskCLIP + FLOSS51%54%56%57%56%55%
NACLIP + FLOSS49%20%57%57%52%47%
CLIP-DINOiser + FLOSS62%41%57%45%44%50%
", + "bbox": [ + 93, + 88, + 482, + 145 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/905a10623f4db0ba0380402e2a1773150cc53ed1975652074d978e0ac7bacc7f.jpg", + "table_caption": [ + "Table 2. Quality of our estimated class-experts. We measure the quality of our estimated experts by evaluating their intersection with the true experts, cf. Eq. (8)." + ], + "table_footnote": [], + "table_body": "
Methodin domain CSout of domain
NightFogRainSnowBDDMAPAvg
MaskCLIP24.513.320.321.019.922.725.820.5
+ FLOSS25.813.921.722.321.322.926.821.5
NACLIP35.523.332.930.732.631.435.331.0
+ FLOSS37.023.632.531.033.632.336.531.6
CLIP-DINOiser31.312.724.127.927.231.335.226.4
+ FLOSS34.616.629.029.829.432.236.628.9
", + "bbox": [ + 91, + 210, + 483, + 321 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Generalization of experts across datasets. We evaluate the transferability of experts identified on the Cityscapes (CS) dataset to unseen datasets that share the same semantic classes and exhibit changes in lighting (Night), weather (Fog, Rain, Snow), or location and scenery (BDD, MAP). Results show a consistent improvement when using FLOSS, across almost all settings.", + "bbox": [ + 89, + 333, + 483, + 417 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "is evident on challenging benchmarks like ADE and Stuff, which have a large number of semantic categories making accurate segmentation particularly hard.", + "bbox": [ + 89, + 441, + 482, + 488 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quality of experts. We evaluate the quality $\\hat{\\rho}_k$ of our experts by measuring, for each class $k$ , the intersection of the estimated class-expert $\\hat{\\mathcal{E}}_k$ with the true set of experts $\\mathcal{E}_k$ :", + "bbox": [ + 89, + 491, + 483, + 537 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\rho} _ {k} = 1 0 0 \\times \\frac {\\left| \\hat {\\mathcal {E}} _ {k} \\cap \\mathcal {E} _ {k} \\right|}{\\left| \\hat {\\mathcal {E}} _ {k} \\right|} = 1 0 0 \\times \\frac {\\left| \\hat {\\mathcal {E}} _ {k} \\cap \\mathcal {E} _ {k} \\right|}{N}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 547, + 482, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Where $|.|$ denotes the cardinality of a set. Quality averaged over classes is reported in Tab. 2 and shows that regardless of the model used, we correctly estimate approximately half of the true experts. In some scenarios, such as PASCAL VOC 20 with NACLIP, even when only $20\\%$ of predicted experts are true experts, it is sufficient to bring some improvements ( $+0.5$ mIoU in Tab. 1). Per-class quality is reported in Appendix A.4, showing some variability.", + "bbox": [ + 89, + 594, + 483, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Generalization to unseen datasets. We now evaluate the generalization of experts to other unseen datasets sharing the same label set but exhibiting distribution shifts. In Tab. 3, we report performance of FLOSS when experts are identified on Cityscapes and used to evaluate performance on six unseen datasets which include four ACDC conditions (Night, Fog, Rain, Snow), BDD-100K, and MAPILLARY. Overall, we consistently improve the performance with notable gains of up to $+4.9$ mIoU for CLIP-DINOiser on ACDC Fog. This demonstrates that our class-wise experts are readily usable on other datasets sharing the same semantic classes, despite the presence of distribution shifts.", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a3551e18aa80179087cf1d94d21fdec9e5e993b0e2b3c7c1c3d0e3b297761cb0.jpg", + "image_caption": [ + "(a) Cityscapes", + "Figure 3. Effect of the number of images. We evaluate CLIP-DINOiser with FLOSS on urban scenes (a) and general object segmentation (b), when varying the number of images used to identify class-experts. Results show the average of 10 seeds, with standard deviation. Despite larger variation, we note the robustness of our method, even when having access to very few unlabeled images." + ], + "image_footnote": [], + "bbox": [ + 513, + 85, + 707, + 186 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/955bcc5f50b34a848238d5a50be45dadf422d01cd017d2dcab2dfaca88143656.jpg", + "image_caption": [ + "(b) PASCAL VOC 20" + ], + "image_footnote": [], + "bbox": [ + 710, + 88, + 903, + 186 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Low-data regimes. To investigate the data efficiency of FLOSS, we study its performance in lower data regimes. Fig. 3 reports the mIoU of CLIP-DINOiser on Cityscapes and PASCAL VOC 20, as the number of images sampled from their respective training sets (containing 2,975 and 1,464 images, respectively) varies. The figure shows the average and standard deviation over 10 seeds (i.e., randomly sampling 10 different image sets). On both datasets, the general trend shows that our method benefits from accessing more images, which stems from the better estimation of class-experts. Remarkably, our method performs well in a very low-data regime, outperforming the default CLIP-DINOiser with only 1 image on Cityscapes and around 25 images on PASCAL VOC 20. We conjecture that the latter requires more images due to its lower per-image class density compared to Cityscapes, which exhibits higher variability and density per image.", + "bbox": [ + 511, + 330, + 906, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Low-data regimes & domain shift. To further corroborate its practicality, we evaluate FLOSS under stricter conditions combining both low-data regimes and domain shift. Specifically, we use only a single CS image to identify experts that are tested on 6 shifted domains. Tab. 4 provides evidence on the effectiveness of FLOSS in this strict setting, leading to an average improvement of $1.9\\%$ across OOD datasets.", + "bbox": [ + 511, + 599, + 905, + 707 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Expert transferability. We investigate the cross-dataset transferability of class-experts by evaluating whether experts identified on one dataset can effectively improve performance on semantically related but distinct datasets. We employ COCO-Stuff (171 classes) to identify class-experts for CLIP-DINOiser and apply them to Cityscapes, PASCAL VOC 20, and PASCAL CONTEXT 59, which share 10, 16, and 37 common classes with COCO-Stuff, respectively. For non-overlapping classes, we use the default average classifier $W_{\\mathrm{CLIP}}$ . As reported in Tab. 5, FLOSS shows consistent improvements, confirming expert transferability across datasets.", + "bbox": [ + 511, + 719, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21476", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7a39889922c6d7faa44ce922202182fb10a2113e64d03e0d70441ec39c15b22f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Methodin domain CSout of domain
NightFogRainSnowBDDMAPAvg
CLIP-DINOiser31.312.724.127.927.231.335.226.4
+ FLOSS33.216.228.229.328.731.735.628.3
", + "bbox": [ + 91, + 88, + 482, + 143 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1fe61797bfb1597d562dfcc0bea8de282b6a715118bfd086f55369743a38c001.jpg", + "table_caption": [ + "Table 4. FLOSS with 1 image & domain shift. Experts are identified using a single Cityscapes (CS) image and evaluated on unseen datasets that share the same semantic classes and exhibit changes in lighting (Night), weather (Fog, Rain, Snow), or location and scenery (BDD, MAP). We observe a consistent boost." + ], + "table_footnote": [], + "table_body": "
Methodin-domain Stuffout of domain
CS commonallVOC20 commonallPC59 commonall
CLIP-DINOiser24.634.531.180.580.839.336.0
+ FLOSS24.735.832.282.382.440.036.2
", + "bbox": [ + 91, + 241, + 483, + 308 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 416, + 243, + 431 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fusion strategies for expert predictions. We investigate alternative strategies to fuse the expert predictions (cf., Sec. 4.2), reporting performance on Cityscapes in Tab. 6. Simply averaging all soft segmentation maps (\"Average-all\") irrespective of the templates class of expertise leads to suboptimal results across all models. When accounting for templates' expertise in fusion, conflicts arise often as several experts may predict their own class of expertise. We therefore explore three strategies to resolve conflicts: \"Default\" which falls back to $W_{\\mathrm{CLIP}}$ predictions, \"Average\" which averages the probability vectors of conflicting experts, and \"Highest\" which selects the most confident expert as described in Sec. 4.2. The latter consistently yields the best performance across all models, validating our fusion scheme.", + "bbox": [ + 88, + 438, + 482, + 664 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metrics for expert identification. We replace entropy in Eq. (6) with other unsupervised metrics to act as proxies of the template class-wise performance, reporting performance on Cityscapes using the CLIP-DINOiser model in Fig. 4a.", + "bbox": [ + 89, + 671, + 482, + 747 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(i) Avg. Probability computes the average probability assigned to class $k$ across all pixels predicted as class $k$ .", + "(ii) $MaNo$ [70] is adapted as it was originally designed for assessing general model performance in image classification, to evaluate class-wise performance in semantic segmentation.", + "(iii) \"ITI\" [25] is the Inter-class separability to Intra-class similarity ratio, and captures both the separability of the classes in the feature space as well as the ability to keep samples of the same class close together." + ], + "bbox": [ + 89, + 747, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6fea1e91f20c515277267c1165e8571ff5789ed7d0df6cdf457c1e538003c189.jpg", + "table_caption": [ + "Table 5. Cross-dataset expert transferability. Class-experts identified on COCO-Stuff are applied to CS, VOC20, and PC59. Results are reported for common classes shared with COCO-Stuff and all classes. FLOSS demonstrates consistent improvements across all target datasets." + ], + "table_footnote": [], + "table_body": "
FusionConflictMaskCLIP +FLOSSNACLIP +FLOSSCLIP-DINOiser +FLOSS
Average-all24.736.631.2
ExpertDefault24.836.632.0
ExpertAverage25.636.734.4
ExpertHighest25.837.034.6
", + "bbox": [ + 537, + 88, + 887, + 184 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Strategies for the fusion of expert predictions. Exploring various fusion strategies on Cityscapes shows the benefit of our strategy, which outperforms all others across models.", + "bbox": [ + 511, + 194, + 906, + 237 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/35f018e50f043a0a4c394456be68202bbc9065a31f331cacd0ad8061f09e7407.jpg", + "image_caption": [ + "(a) Expert metrics" + ], + "image_footnote": [], + "bbox": [ + 514, + 252, + 689, + 338 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5cd0243995006990d162def40ea3ee48a4b6dee6952de3490583bee4abf28408.jpg", + "image_caption": [ + "(b) Effect of Top-N" + ], + "image_footnote": [], + "bbox": [ + 694, + 252, + 905, + 358 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/104618af780225ba9347ec74934024773ff5ddc0e7c15be94356703eec0c6d69.jpg", + "image_caption": [ + "Figure 4. Ablation of experts. We ablate our choices of metrics to identify experts as well as the number of experts per class. (a) shows unsupervised metrics to identify the experts, showing that our choice of entropy yields the best performance, although 'Avg. Probability' performs on par. (b) reports performance of FLOSS when varying the number of $N$ experts. Results are reported on the Cityscapes dataset using the CLIP-DINOiser model.", + "Figure 5. Effect of class-expert quality. We report the performance of FLOSS in the oracle-based setting, where each class-wise set of 4 experts contains a ratio $(\\rho)$ of true experts. The experiment is conducted on Cityscapes using the CLIP-DINOiser model." + ], + "image_footnote": [], + "bbox": [ + 598, + 498, + 820, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "These metrics are detailed in Appendix A.5.", + "bbox": [ + 511, + 728, + 807, + 743 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Overall, our entropy measure (cf. Eq. (5)) is the most effective, although Avg. Probability performs on par.", + "bbox": [ + 511, + 744, + 903, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of class-expert identification quality. At the core of our method is the identification of class-experts. While we demonstrated there are templates which are class-expert, for each class there exist also under-performing templates (cf. Fig. 1) which, if selected, could harm the performance.", + "bbox": [ + 511, + 779, + 906, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To measure the impact of experts' quality on FLOSS, we devise an oracle-based experiment using $N = 4$ templates per-class, where we have access to ground truth la", + "bbox": [ + 511, + 854, + 905, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21477", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/840e7a385f70f7b5a999234090df76da968e3692fa5bf6995f02f0a260f4e4ee.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodViT-B/16ViT-L/14
CSPC59Avg.CSPC59Avg.
SCLIP [63]32.234.233.221.325.223.3
GEM [3]32.635.634.127.128.127.6
NACLIP [23]35.535.235.431.432.131.8
+ FLOSS37.035.936.532.432.432.4
", + "bbox": [ + 112, + 88, + 460, + 188 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Effect of backbone. Results (mIoU) show that NACLIP + FLOSS outperforms NACLIP across ViT-B/16 and ViT-L/14 CLIP backbones.", + "bbox": [ + 89, + 199, + 482, + 241 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "bels to identify the true class-experts. We randomly sample a fixed ratio $\\rho$ of templates among true experts (i.e., $\\mathcal{E}_k$ from Eq. (4)) and the remaining ones among non-experts (i.e., $\\{\\mathcal{T}_m\\} \\backslash \\mathcal{E}_k$ ). For example, a ratio of $\\rho = 75\\%$ indicates that 3 templates are experts and 1 is non-expert. The outcome of this experiment on CS using the CLIP-DINOiser model is reported in Fig. 5 and shows an expected performance boost when the ratio of true experts increases. It also highlights that $50\\%$ of true experts is sufficient for FLOSS to surpass the original CLIP-DINOiser. Interestingly, the performance of CLIP-DINOiser + FLOSS in Tab. 1 (34.6) slightly surpasses the performance at $\\rho = 100\\%$ shown in Fig. 5 (34.1). This however does not indicate that our estimated templates in Tab. 1 are all true experts - as reflected by Tab. 2 - but rather that our correctly estimated experts are on average better than random true experts.", + "bbox": [ + 88, + 282, + 482, + 525 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further measure the upper bound of CLIP-DINOiser + FLOSS, we implement an oracle version using only the best true class-expert (i.e., $\\{\\mathcal{T}_{\\tilde{m}} \\mid \\tilde{m} \\in \\mathrm{Top-N}(\\Omega_k(W(\\mathcal{T}_m)))\\}$ ). On CS, the oracle with $N \\in \\{1, 2, 3, 4\\}$ achieves respectively 37.1/37.0/37.3/37.5%, showing the potential for improvement assuming the identification of better experts.", + "bbox": [ + 89, + 531, + 483, + 623 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "FLOSS using the validation set. First, we investigate the effectiveness of FLOSS in a transductive setting. Using the validation set of CS for expert selection, CLIP-DINOiser + FLOSS achieves $34.4\\%$ mIoU on the same set, compared to $31.3\\%$ for CLIP-DINOiser. Second, using a portion of the validation set to select experts, i.e., 5/25/50/75 val images, we achieve 32.8/33.0/33.0/33.3% on the rest of the validation images, compared to $30.9\\%$ for CLIP-DINOiser alone.", + "bbox": [ + 89, + 640, + 482, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of different backbones. Tab. 7 demonstrates the effectiveness of our proposed FLOSS when integrated with NACLIP using the ViT-L/14 backbone. Our approach achieves consistent improvements, reaching $32.4\\%$ mIoU on both Cityscapes and PC59 datasets, leading to a 0.6 gain in average performance over NACLIP. These results validate the effectiveness of our method when using large-scale vision transformers.", + "bbox": [ + 89, + 780, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3be22c283219e3b72b7b87d86c2aa877d663ef3468c30965790df2fce340d449.jpg", + "image_caption": [ + "Figure 6. Performance of best-class experts. Performance gains (IoU) when using best class-experts versus standard template-averaging. Results are shown for CLIP-DINOiser, MaskCLIP and NACLIP. The bars represent the IoU difference between using best class-experts (selected with ground-truth labels) and the default template-averaging approach." + ], + "image_footnote": [], + "bbox": [ + 513, + 88, + 903, + 186 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Perspectives and Future Work", + "text_level": 1, + "bbox": [ + 511, + 306, + 792, + 323 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our experiments show that entropy is a strong estimator of class-experts. However, future research could develop even better proxies for prediction accuracy. Template-based methods appear especially promising, with significant potential for improving OVSS performance. As illustrated in Fig. 6, there is still a large gap to close—for example, selecting the best expert for the \"sky\" class could improve IoU by over 30 percentage points. These \"upper-bounds\" could be further increased by using more or augmented templates beyond the 80 currently used.", + "bbox": [ + 511, + 332, + 906, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 500, + 633, + 516 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a new task for improving OVSS models without having access to labels and without training, from a prompt template perspective. Our work is motivated by an intriguing observation: for each class, some templates excel in segmenting it. We refer to these templates as class-experts. Given a set of unlabeled images, our FLOSS uses the entropy of predictions as an unsupervised metric to identify these class-experts, thus not requiring any training or labels. Once selected, we propose a simple scheme for fusing the predictions of the class-experts, resulting in an improved overall OVSS performance in the inductive setting, even in the presence of distribution shifts. Additionally, we corroborate the effectiveness of FLOSS, which benefits from being plug-and-play, and further show that only few unlabeled images can be sufficient for expert selection. Furthermore, the oracle performance shows the potential of the class-experts, paving the way for future research.", + "bbox": [ + 511, + 527, + 906, + 784 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment. This research was partially funded by the French Agence Nationale de la Recherche (ANR) with the project SIGHT (ANR-20-CE23-0016) and supported by ELSA - European Lighthouse on Secure and Safe AI funded by the European Union under grant agreement No. 101070617. We sincerely thank Telecom Paris for providing the resources necessary to run our experiments and Nacereddine Laddaoui for his help with infrastructure. We are also grateful to Ivan Lopes for proofreading.", + "bbox": [ + 511, + 789, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21478", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Stephen H Bach, Victor Sanh, Zheng-Xin Yong, Albert Webson, Colin Raffel, Nihal V Nayak, Abheesht Sharma, Taewoon Kim, M Saiful Bari, Thibault Fevry, et al. Prompt-source: An integrated development environment and repository for natural language prompts. In ACL, 2022. 3", + "[2] Sule Bai, Yong Liu, Yifei Han, Haoji Zhang, and Yansong Tang. Self-calibrated clip for training-free open-vocabulary segmentation. arXiv preprint arXiv:2411.15869, 2024. 2", + "[3] Walid Bousselham, Felix Petersen, Vittorio Ferrari, and Hilde Kuehne. Grounding everything: Emerging localization properties in vision-language transformers. In CVPR, 2024. 2, 5, 8", + "[4] Maxime Bucher, Tuan-Hung Vu, Matthieu Cord, and Patrick Pérez. Zero-shot semantic segmentation. In NeurlPS, 2019. 2", + "[5] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In CVPR, 2018. 2, 5", + "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 1", + "[7] Junbum Cha, Jonghwan Mun, and Byungseok Roh. Learning to generate text-grounded mask for open-world semantic segmentation from only image-text pairs. In CVPR, 2023. 5", + "[8] Guangyi Chen, Weiran Yao, Xiangchen Song, Xinyue Li, Yongming Rao, and Kun Zhang. Plot: Prompt learning with optimal transport for vision-language models. In ICLR, 2023. 2", + "[9] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. TPAMI, 2017. 1", + "[10] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 1", + "[11] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, 2023. 5", + "[12] Seokju Cho, Heeseong Shin, Sunghwan Hong, Anurag Arnab, Paul Hongsuck Seo, and Seungryong Kim. Cat-seg: Cost aggregation for open-vocabulary semantic segmentation. In CVPR, 2024. 1, 2", + "[13] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 3, 5", + "[14] Ulindu De Silva, Didula Samaraweera, Sasini Wanigathunga, Kavindu Kariyawasam, Kanchana Ranasinghe, Muzammal Naseer, and Ranga Rodrigo. Test-time optimization for domain adaptive open vocabulary segmentation. arXiv preprint arXiv:2501.04696, 2025. 2" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 5", + "[16] Reza Esfandiarpoor, Cristina Menghini, and Stephen H Bach. If clip could talk: Understanding vision-language model representations through their preferred concept descriptions. In EMNLP, 2024. 3", + "[17] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results, 2012. 5", + "[18] Yarin Gal. Uncertainty in deep learning. PhD Thesis, University of Cambridge, 2016. 4", + "[19] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. IJCV, 2024. 2", + "[20] Yunhao Ge, Jie Ren, Andrew Gallagher, Yuxiao Wang, Ming-Hsuan Yang, Hartwig Adam, Laurent Itti, Balaji Lakshminarayanan, and Jiaping Zhao. Improving zero-shot generalization and robustness of multi-modal models. In CVPR, 2023. 3", + "[21] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In ECCV, 2022. 2, 3", + "[22] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In ICLR, 2022. 3", + "[23] Sina Hajimiri, Ismail Ben Ayed, and Jose Dolz. Pay attention to your neighbours: Training-free open-vocabulary semantic segmentation. In WACV, 2025. 1, 2, 3, 5, 8", + "[24] Cong Han, Yujie Zhong, Dengjie Li, Kai Han, and Lin Ma. Open-vocabulary semantic segmentation with decoupled one-pass network. In ICCV, 2023. 2", + "[25] Dapeng Hu, Jian Liang, Jun Hao Liew, Chuhui Xue, Song Bai, and Xinchao Wang. Mixed samples as probes for unsupervised model selection in domain adaptation. In NeurIPS, 2023. 7", + "[26] Xuefeng Hu, Ke Zhang, Lu Xia, Albert Chen, Jiajia Luo, Yuyin Sun, Ken Wang, Nan Qiao, Xiao Zeng, Min Sun, et al. Reclip: Refine contrastive language image pre-training with source free domain adaptation. In WACV, 2024. 2", + "[27] Haiwen Huang, Songyou Peng, Dan Zhang, and Andreas Geiger. Renovating names in open-vocabulary segmentation benchmarks. In NeurIPS, 2024. 3", + "[28] Tony Huang, Jack Chu, and Fangyun Wei. Unsupervised prompt learning for vision-language models. arXiv preprint arXiv:2204.03649, 2022. 2", + "[29] Yunshi Huang, Fereshteh Shakeri, Jose Dolz, Malik Boudiaf, Houda Bahig, and Ismail Ben Ayed. Lp++: A surprisingly strong linear probe for few-shot clip. In CVPR, 2024. 2", + "[30] Eyke Hüllermeier and Willem Waegeman. *Aleatoric and epistemic uncertainty in machine learning: An introduction to concepts and methods.* ML, 2021. 4", + "[31] Krishna Murthy Jatavallabhula, Alihusein Kuwajerwala, Qiao Gu, Mohd Omama, Tao Chen, Shuang Li, Ganesh Iyer, Soroush Saryazdi, Nikhil Keetha, Ayush Tewari, et al. ConceptFusion: Open-set multimodal 3d mapping. In RSS, 2023." + ], + "bbox": [ + 516, + 92, + 903, + 887 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21479", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? In NeurIPS, 2017. 4", + "[33] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In CVPR, 2023. 2", + "[34] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. In NeurIPS, 2017. 4", + "[35] Mengcheng Lan, Chaofeng Chen, Yiping Ke, Xinjiang Wang, Litong Feng, and Wayne Zhang. Proxyclip: Proxy attention improves clip for open-vocabulary segmentation. In ECCV, 2024. 2", + "[36] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. In ICLR, 2022. 2", + "[37] Yi Li, Hualiang Wang, Yiqun Duan, Jiheng Zhang, and Xiaomeng Li. A closer look at the explainability of contrastive language-image pre-training. PR, 2025. 2, 5", + "[38] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In CVPR, 2023. 1, 2", + "[39] Yuqi Lin, Minghao Chen, Wenxiao Wang, Boxi Wu, Ke Li, Binbin Lin, Haifeng Liu, and Xiaofei He. Clip is also an efficient segmenter: A text-driven approach for weakly supervised semantic segmentation. In CVPR, 2023. 3", + "[40] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. In ICLR, 2023. 3", + "[41] Matthias Minderer, Josip Djolonga, Rob Romijnders, Frances Hubis, Xiaohua Zhai, Neil Houlsby, Dustin Tran, and Mario Lucic. Revisiting the calibration of modern neural networks. In NeurIPS, 2021. 2", + "[42] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The role of context for object detection and semantic segmentation in the wild. In CVPR, 2014. 5", + "[43] Jishnu Mukhoti, Tsung-Yu Lin, Omid Poursaeed, Rui Wang, Ashish Shah, Philip HS Torr, and Ser-Nam Lim. Open vocabulary semantic segmentation with patch aligned contrastive learning. In CVPR, 2023. 1", + "[44] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In ICCV, 2017. 5", + "[45] Sarah Parisot, Yongxin Yang, and Steven McDonagh. Learning to name classes for vision and language models. In CVPR, 2023. 3", + "[46] Sarah Pratt, Ian Covert, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. In ICCV, 2023. 3", + "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3, 5", + "[48] Kanchana Ranasinghe, Brandon McKinzie, Sachin Ravi, Yinfei Yang, Alexander Toshev, and Jonathon Shlens. Per" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ceptual grouping in contrastive vision-language models. In ICCV, 2023. 2", + "[49] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In CVPR, 2022. 2", + "[50] Karsten Roth, Jae Myung Kim, A Koepke, Oriol Vinyals, Cordelia Schmid, and Zeynep Akata. Waffling around for performance: Visual classification with random words and broad concepts. In ICCV, 2023. 3", + "[51] Ohad Rubin, Jonathan Herzig, and Jonathan Berant. Learning to retrieve prompts for in-context learning. In *NAACL-HLT*, 2022. 3", + "[52] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding. In CVPR, 2021. 5", + "[53] Gyungin Shin, Weidi Xie, and Samuel Albanie. Reco: Retrieve and co-segment for zero-shot transfer. In NeurIPS, 2022. 5", + "[54] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. In NeurIPS, 2022. 2, 3", + "[55] Julio Silva-Rodriguez, Sina Hajimiri, Ismail Ben Ayed, and Jose Dolz. A closer look at the few-shot adaptation of large vision-language models. In CVPR, 2024. 2", + "[56] Vladan Stojnic, Yannis Kalantidis, and Giorgos Tolias. Label propagation for zero-shot classification with vision-language models. In CVPR, 2024. 2", + "[57] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In ICCV, 2021. 1", + "[58] Weijie Tu, Weijian Deng, and Tom Gedeon. A closer look at the robustness of contrastive language-image pre-training (clip). In NeurIPS, 2023. 2", + "[59] Weijie Tu, Weijian Deng, Dylan Campbell, Stephen Gould, and Tom Gedeon. An empirical study into what matters for calibrating vision-language models. In ICML, 2024. 2", + "[60] Antonin Vobecky, Oriane Simeoni, David Hurych, Spyridon Gidaris, Andrei Bursuc, Patrick Pérez, and Josef Sivic. Pop-3d: Open-vocabulary 3d occupancy prediction from images. In NeurIPS, 2023. 3", + "[61] Tuan-Hung Vu, Himalaya Jain, Maxime Bucher, Matthieu Cord, and Patrick Pérez. Advent: Adversarial entropy minimization for domain adaptation in semantic segmentation. In CVPR, 2019. 4", + "[62] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In ICLR, 2021. 4", + "[63] Feng Wang, Jieru Mei, and Alan Yuille. Sclip: Rethinking self-attention for dense vision-language inference. In ECCV, 2024. 1, 5, 8", + "[64] Zhengbo Wang, Jian Liang, Lijun Sheng, Ran He, Zilei Wang, and Tieniu Tan. A hard-to-beat baseline for training-free CLIP-based adaptation. In ICLR, 2024. 2", + "[65] Monika Wysoczańska, Michael Ramamonjisoa, Tomasz Trzeciński, and Oriane Simeoni. Clip-diy: Clip dense infer" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21480", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ence yields open-vocabulary semantic segmentation for-free. In WACV, 2024. 5", + "[66] Monika Wysoczanska, Oriane Simeoni, Michael Ramamonjisoa, Andrei Bursuc, Tomasz Trzciński, and Patrick Pérez. Clip-dinoiser: Teaching clip a few dino tricks for open-vocabulary semantic segmentation. In ECCV, 2024. 1, 2, 3, 5", + "[67] Yongqin Xian, Subhabrata Choudhury, Yang He, Bernt Schiele, and Zeynep Akata. Semantic projection network for zero-and few-label semantic segmentation. In CVPR, 2019. 2", + "[68] Bin Xie, Jiale Cao, Jin Xie, Fahad Shahbaz Khan, and Yanwei Pang. Sed: A simple encoder-decoder for open-vocabulary semantic segmentation. In CVPR, 2024. 2, 3", + "[69] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. In NeurIPS, 2021. 1", + "[70] Renchunzi Xie, Ambroise Odonnat, Vasilii Feofanov, Weijian Deng, Jianfeng Zhang, and Bo An. Mano: Exploiting matrix norm for unsupervised accuracy estimation under distribution shifts. In NeurIPS, 2024. 4, 7", + "[71] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. Groupvit: Semantic segmentation emerges from text supervision. In CVPR, 2022. 5", + "[72] Jilan Xu, Junlin Hou, Yuejie Zhang, Rui Feng, Yi Wang, Yu Qiao, and Weidi Xie. Learning open-vocabulary semantic segmentation models from natural language supervision. In CVPR, 2023. 2", + "[73] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In CVPR, 2023. 2", + "[74] Mengde Xu, Zheng Zhang, Fangyun Wei, Han Hu, and Xi-ang Bai. Side adapter network for open-vocabulary semantic segmentation. In CVPR, 2023. 2", + "[75] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darryll. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2020. 5", + "[76] Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang-Chieh Chen. Convolutions die hard: Open-vocabulary segmentation with single frozen convolutional clip. In NeurIPS, 2023. 2", + "[77] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. In ECCV, 2022. 2", + "[78] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. IJCV, 2019. 5", + "[79] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In ECCV, 2022. 1, 2, 3, 5", + "[80] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In CVPR, 2022. 2" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[81] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 2022.", + "[82] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. In ICCV, 2023. 2", + "[83] Xiangyang Zhu, Renrui Zhang, Bowei He, Aojun Zhou, Dong Wang, Bin Zhao, and Peng Gao. Not all features matter: Enhancing few-shot clip with adaptive prior refinement. In ICCV, 2023. 2" + ], + "bbox": [ + 516, + 90, + 903, + 232 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "21481", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/FLOSS_ Free Lunch in Open-vocabulary Semantic Segmentation/759adae1-3d57-4b38-8780-a8c8f1ca245e_model.json b/2025/FLOSS_ Free Lunch in Open-vocabulary Semantic Segmentation/759adae1-3d57-4b38-8780-a8c8f1ca245e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1dd8cff7af218554d522e0b575e76a8cd41e8632 --- /dev/null +++ b/2025/FLOSS_ Free Lunch in Open-vocabulary Semantic Segmentation/759adae1-3d57-4b38-8780-a8c8f1ca245e_model.json @@ -0,0 +1,2609 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.171, + 0.131, + 0.828, + 0.154 + ], + "angle": 0, + "content": "FLOSS: Free Lunch in Open-vocabulary Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.179, + 0.918, + 0.217 + ], + "angle": 0, + "content": "Yasser Benigim \\(^{1*}\\) Mohammad Fahes \\(^{1}\\) Tuan-Hung \\(\\mathrm{Vu}^{1,2}\\) Andrei Bursuc \\(^{1,2}\\) Raoul de Charette \\(^{1}\\) Inria \\(^{2}\\) Valeo.ai" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.251, + 0.328, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.284, + 0.485, + 0.66 + ], + "angle": 0, + "content": "In this paper, we challenge the conventional practice in Open-Vocabulary Semantic Segmentation (OVSS) of using averaged class-wise text embeddings, which are typically obtained by encoding each class name with multiple templates (e.g., a photo of , a sketch of a ). We investigate the impact of templates for OVSS, and find that for each class, there exist single-template classifiers—which we refer to as class-experts—that significantly outperform the conventional averaged classifier. First, to identify these class-experts, we introduce a novel approach that estimates them without any labeled data or training. By leveraging the class-wise prediction entropy of single-template classifiers, we select those yielding the lowest entropy as the most reliable class-experts. Second, we combine the outputs of class-experts in a new fusion process. Our plug-and-play method, coined FLOSS, is orthogonal and complementary to existing OVSS methods, offering an improvement without the need for additional labels or training. Extensive experiments show that FLOSS consistently enhances state-of-the-art OVSS models, generalizes well across datasets with different distribution shifts, and delivers substantial improvements in low-data scenarios where only a few unlabeled images are available. Our code is available at https://github.com/yasserben/FLOSS." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.688, + 0.222, + 0.703 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.713, + 0.483, + 0.849 + ], + "angle": 0, + "content": "In the past decade, advances in deep learning and the growing amount of training data have enabled the challenging task of semantic segmentation, which involves assigning semantic labels to each pixel in an image. Initially, this was limited to predefined categories [9, 10, 57, 69], but recent models rely on vision-language alignment [47] for open-vocabulary segmentation, where the semantic categories can be dynamically expressed at runtime in natural language." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.849, + 0.483, + 0.88 + ], + "angle": 0, + "content": "In Open-Vocabulary Semantic Segmentation (OVSS), the CLIP model [47] is commonly used as the backbone," + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.25, + 0.907, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.419, + 0.908, + 0.504 + ], + "angle": 0, + "content": "Figure 1. Using individual vs average template classifiers. Empirically, we observe that for each class there exist individual templates (colored dots) that lead to classifiers performing better than the popular CLIP classifier, which is built from the averaging of all 80 templates embeddings \\((-)\\). The analysis was conducted on Cityscapes using the CLIP-DINOiser model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.53, + 0.907, + 0.878 + ], + "angle": 0, + "content": "which is either retrained [43], fine-tuned [12, 38], or frozen and adjusted to extend global image-language alignment to denser pixel-language alignment. In this work, we study CLIP's original classification paradigm and focus on methods that keep all CLIP parameters frozen [23, 63, 79]. Among these, MaskCLIP [79] leverages CLIP's embedding layer for segmentation, CLIP-DINOiser [66] enhances features via distillation from an external DINO [6] network, while NACLIP [23] modifies CLIP's self-attention to improve spatial consistency. A common aspect of prior OVSS methods is their reliance on CLIP's default classification process, where image or patch embeddings are compared with text embeddings representing semantic categories. To represent a class, a common practice is to use multiple templates to encode class names (e.g., \"a photo of a car\", \"a sketch of a car\", etc) and average their embeddings. Empirically, Radford et al. [47] engineered \\( M = 80 \\) templates to improve ImageNet zero-shot classification. For segmentation, OVSS methods [23, 66, 79] adopt by default the same templates to construct segmentation classifiers. While this \"template-averaging\" strategy works well overall, it may not always provide the best classifier for each semantic concept." + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.887, + 0.878, + 0.902 + ], + "angle": 0, + "content": "1\"Template-averaging\" is an imprecise but convenient shorthand." + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.329, + 0.901 + ], + "angle": 0, + "content": "*Work done during an internship at Inria." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21471" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "Our paper explores the impact of individual templates in OVSS – an aspect that has been overlooked in the literature. Fig. 1 shows an intriguing empirical observation: for each class, some individual templates outperform the averaged classifier that uses all templates. We call such templates class-experts, highlighting that they are trivial to identify using semantic ground truths." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.199, + 0.483, + 0.244 + ], + "angle": 0, + "content": "However, without labels, two challenges arise: 1) How to identify class-experts without ground truth? 2) How to fuse predictions from class-experts?" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.245, + 0.483, + 0.395 + ], + "angle": 0, + "content": "We propose that entropy, an unsupervised metric, can help identify expert templates without labels. This leads to a new task—training- and label-free template selection for OVSS—which improves segmentation by selecting better text template subsets for a given dataset of unlabeled images. Our method shifts the attention to the text modality, unlike previous works that tweak the visual encoder, and also aims to improve the performance in an inductive setting, where the selected templates are applied to a different subset of images." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.396, + 0.384, + 0.41 + ], + "angle": 0, + "content": "We summarize our contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.412, + 0.483, + 0.471 + ], + "angle": 0, + "content": "- We analyze the current template-averaging practice for OVSS, showing that better class-wise classifiers can be constructed using only a subset of all templates, and that these expert templates can be identified without labels." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.473, + 0.483, + 0.546 + ], + "angle": 0, + "content": "- We propose the novel task of training- and label-free template selection for OVSS. Given an unlabeled set of images, this task aims to identify class-wise experts that improve OVSS performance on unseen counterparts of the dataset." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.548, + 0.483, + 0.607 + ], + "angle": 0, + "content": "- We introduce FLOSS, which leverages class-wise per-pixel entropy to select expert templates and employs a simple and effective fusion scheme to combine individual expert predictions." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.412, + 0.483, + 0.607 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.609, + 0.483, + 0.7 + ], + "angle": 0, + "content": "With extensive experiments, we demonstrate that FLOSS consistently improves the state-of-the-art performance on OVSS benchmarks and exhibits generalization capabilities in the choice of experts. Furthermore, with only a few unlabeled images, our approach can select reasonably good expert templates, resulting in performance boost." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.715, + 0.24, + 0.731 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Vision-Language Model (VLM) transferability. Since CLIP's release, its impressive zero-shot classification performance [47] and robustness [41, 58, 59] have made it a foundation for various adaptation and transfer learning tasks. Few-shot supervised fine-tuning, where only limited labeled data is available for downstream tasks like image classification, has gained significant attention. The goal is to leverage the vast knowledge of the VLM acquired through training on an immense image collection by steering it towards a task of interest via few labeled sam" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.228 + ], + "angle": 0, + "content": "plies. Parameter efficiency is central to this setting. To ensure it, prompt learning [8, 33, 80-82], lightweight MLP training [19], and linear probing [29, 55] have been proposed. Another line of research explores training-free few-shot adaptation [64, 77, 83], where a classifier is built using few labeled samples and ensembled with the zero-shot classifier. Additionally, unsupervised adaptation has been explored [26, 28, 56], alongside test-time adaptation, where prompts are learned per test image [54]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.23, + 0.905, + 0.322 + ], + "angle": 0, + "content": "While using VLMs for classification tasks aligns with their pre-training focus on global image recognition, applying them to dense prediction tasks like semantic segmentation is more challenging due to the gap between global image understanding and pixel-level prediction. We summarize related efforts in the following." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.339, + 0.905, + 0.838 + ], + "angle": 0, + "content": "Open-Vocabulary Semantic Segmentation (OVSS). OVSS uses VLMs like CLIP for text-based segmentation. However, CLIP's global pooling layers limit its ability to generate dense pixel-level, language-aligned features [31, 79]. To address this, several training-free methods densify CLIP without altering its parameters, preserving its image-language alignment. MaskCLIP [79] was the first to replace the attention pooling layer with a convolutional layer. Later works further densify CLIP by aggregating multiple image views [31, 66], integrating priors from other vision foundation models for better pooling [35, 66], extracting different information from the self-attention layers [3, 23, 37], or removing anomaly tokens [2]. Seg-TTO [14] leverages an LLM to generate discriminative attributes for rare classes in specialized domains (e.g., medical sciences, earth monitoring) that use technical terminology (e.g., mediastinum), operating in a test-time optimization framework while maintaining zero-shot capabilities. Other approaches fine-tune CLIP for segmentation tasks using weak supervision through image-level tags, captions [21, 48], or class-agnostic object masks [21, 49]. Others employ full supervision [12, 36, 38, 74] on densely annotated datasets like COCO Stuff [5], following strategies similar to early inductive zero-shot semantic segmentation methods [4, 67]. Among them, some fine-tune CLIP's image encoder [36, 38, 74], others combine this with additional backbones and segmentation heads [38], while some learn side networks while keeping CLIP frozen [74]. Many other approaches have been proposed in this active research area [24, 68, 72, 73, 76]. While effective, these methods require labeled data and computational resources, limiting their practicality in resource-constrained scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Our work focuses on training-free approaches, evaluating FLOSS on three key methods: MaskCLIP [79], CLIP-DINOiser [66] and NACLIP [23], while being applicable to any text-based segmentation framework." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21472" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.409 + ], + "angle": 0, + "content": "Prompting in VLMs. CLIP's image-language interface has inspired the adoption of various prompt engineering practices from LLMs to better reveal the knowledge encapsulated in the vision encoder and boost performance. The seminal CLIP work [47] introduced dataset-specific templates (e.g., \"a photo of \") to bridge the gap between training captions and test prompts for zero-shot classification. Subsequent methods have improved prompts by crowdsourcing templates [1], adjusting class names [21, 60], or modifying templates for tasks like object detection [21, 22] and occupancy prediction [60]. While effective, manual prompt tuning is labor-intensive, leading to automated strategies like adding random words or characters to the original template [50], enriching class names with WordNet concepts [20] or synonyms [39], learning more informative class names [27], and using LLMs to generate more expressive descriptions starting from simple class names or retrieve them from a dataset [51] and further use them as prompts. Some approaches also learn prompts for task-specific fitting [45] or test-time adaptation [54], and others mine them with external LLMs [16]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.413, + 0.484, + 0.504 + ], + "angle": 0, + "content": "Our approach differs by selecting expert templates for each class in an unsupervised manner, starting from predefined templates. While previous methods have focused on image classification, we apply this strategy to semantic segmentation using CLIP's original ImageNet templates, which have been used by default in prior works." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.531, + 0.231, + 0.547 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.562, + 0.485, + 0.751 + ], + "angle": 0, + "content": "OVSS aims to produce a soft-segmentation map \\(\\hat{\\mathbf{S}}\\in [0,1]^{H\\times W\\times K}\\) given an image \\(\\mathbf{l}\\in \\mathbb{R}^{H\\times W\\times 3}\\) and a set of classes \\(\\{\\mathcal{C}_k\\}_{k\\in [1,K]}\\) expressed in natural language. The final segmentation map \\(\\hat{\\mathbf{P}}\\) is derived by applying arg max(·) on \\(\\hat{\\mathbf{S}}\\) along the \\(K\\) -dimension. Leveraging the aligned vision encoder \\(f(\\cdot)\\) and text encoder \\(g(\\cdot)\\) of CLIP, the standard practice consists in classifying an image patch by comparing its vision embedding to the text embedding of class-specific prompts constructed using predefined templates \\(\\{\\mathcal{T}_m\\}_{m\\in [1,M]}\\). An example of prompt is \"a bright photo of a car\", denoted" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.751, + 0.484, + 0.901 + ], + "angle": 0, + "content": "\\(\\mathcal{P}_{m,k}\\) for compactness. Instead of a single template, practitioners use \"template-averaging\" to increase robustness and improve overall performance by averaging the embeddings of prompts obtained using all available templates [40, 46, 47, 50], which is a form of ensembling on the text encoder side. This writes \\(\\bar{c}_k = \\frac{1}{M}\\sum_{m = 1}^{M}g(\\mathcal{P}_{m,k})\\) where \\(\\bar{c}_k\\) is the resulting representation of class \\(C_k\\). Subsequently, for each patch, the predicted class is the one that maximizes the cosine similarity between the visual and the" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.632, + 0.107 + ], + "angle": 0, + "content": "text embeddings:" + }, + { + "type": "equation", + "bbox": [ + 0.628, + 0.111, + 0.907, + 0.134 + ], + "angle": 0, + "content": "\\[\n\\hat {k} = \\arg \\max _ {k} f (\\mathcal {V}) ^ {\\mathrm {T}} \\bar {c} _ {k}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.139, + 0.907, + 0.184 + ], + "angle": 0, + "content": "where \\(\\mathcal{V}\\) represents the visual patch. All embeddings are \\(\\ell_2\\)-normalized. The text embeddings of \\(K\\) classes define a classifier that we denote:" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.191, + 0.907, + 0.209 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {W} \\left(\\left\\{\\mathcal {T} _ {m} \\right\\} _ {m \\in [ 1, M ]}\\right) = \\left\\{\\bar {c} _ {k} \\right\\} _ {k \\in [ 1, K ]}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.213, + 0.907, + 0.258 + ], + "angle": 0, + "content": "As in the above equation, the classifier can be represented as a function of the templates used to compute the average embeddings of all classes." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.263, + 0.907, + 0.428 + ], + "angle": 0, + "content": "Empirical observations on single templates. OVSS models [23, 66, 68, 79] typically rely on the \\(M = 80\\) ImageNet templates of CLIP [47] to compute the average classifier of Eq. (2), which we refer to as \\(W_{\\mathrm{CLIP}}\\) for brevity. While using \\(W_{\\mathrm{CLIP}}\\) was shown to be robust [47, 79], the templates were originally hand-crafted for zero-shot classification on ImageNet [47], and their individual effect on class-wise performance remains unexplored for OVSS. Rather than \\(W_{\\mathrm{CLIP}}\\) that uses all \\(M\\) templates, one can construct \\(M\\) distinct classifiers, each utilizing a single template \\(\\mathcal{T}_m\\). This writes:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.435, + 0.905, + 0.453 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {W} \\left(\\mathcal {T} _ {m}\\right) = \\left\\{g \\left(\\mathcal {P} _ {m, k}\\right) \\right\\} _ {k \\in [ 1, K ]}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.457, + 0.906, + 0.576 + ], + "angle": 0, + "content": "We design a systematic experiment to evaluate the performance of single-template classifiers \\( \\mathbf{W}(\\mathcal{T}_m) \\) from Eq. (3). Fig. 1 shows the per-class performance on Cityscapes [13] of each of the 80 single-template classifiers (plotted as colored dots, e.g., ●), where colors encode the template identifier (i.e., m), along with the average \\( W_{\\mathrm{CLIP}} \\) classifier (plotted as —). This leads to two key observations:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.578, + 0.907, + 0.638 + ], + "angle": 0, + "content": "a) For each class, there exist single-template classifiers outperforming \\( W_{\\mathrm{CLIP}} \\) on this class; we refer to the corresponding templates as class-experts. More precisely, the set of class-experts is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.644, + 0.905, + 0.674 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} _ {k} = \\left\\{\\mathcal {T} _ {m} \\mid \\Omega_ {k} \\left(\\boldsymbol {W} \\left(\\mathcal {T} _ {m}\\right)\\right) > \\Omega_ {k} \\left(\\boldsymbol {W} _ {\\text {C L I P}}\\right), m \\in [ 1, M ] \\right\\}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.675, + 0.868, + 0.689 + ], + "angle": 0, + "content": "where \\(\\Omega_{k}(\\cdot)\\) denotes the IoU performance on class \\(k\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.69, + 0.906, + 0.735 + ], + "angle": 0, + "content": "b) A single-template classifier excelling on one class may exhibit suboptimal performance on others, implying that for \\( k, l \\in [1, K] \\) the expert sets \\( \\mathcal{E}_k \\) and \\( \\mathcal{E}_l \\) may differ." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.736, + 0.906, + 0.809 + ], + "angle": 0, + "content": "Following the above observations, class-experts are readily identified when ground truth labels are accessible, through direct performance comparison. We refer readers to Appendix A.3 for similar observations across datasets and models." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.811, + 0.907, + 0.902 + ], + "angle": 0, + "content": "However, we are interested in the case where an unlabeled dataset is provided. Therefore, two questions stem from the previous empirical observations: (i) How to identify class-experts in an unsupervised, training-free manner? (ii) Given these identified class-experts, how to derive the final semantic predictions?" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21473" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.176, + 0.092, + 0.408, + 0.105 + ], + "angle": 0, + "content": "Unsupervised identification of class-experts" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.108, + 0.49, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.667, + 0.093, + 0.745, + 0.105 + ], + "angle": 0, + "content": "Expert Fusion" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.111, + 0.905, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.242, + 0.907, + 0.327 + ], + "angle": 0, + "content": "Figure 2. Overview of FLOSS. Left: for each class, selected expert templates \\((\\checkmark)\\) construct the estimated set \\(\\hat{\\mathcal{E}}\\), which classifies the input image and yield the soft-segmentation map \\(\\mathbf{S}\\). In this example we set \\(N = 3\\) for the \\(\\mathrm{Top - N}(\\dots)\\) template indices corresponding to lowest entropy entries to select. Right: we explain the fusion strategy for two cases. For the pixel \\(x_{1}\\), the set of experts that predict their own class of expertise \\(\\mathcal{X}(x_1)\\) contains 2 experts; consequently, the decision is taken by looking at the maximum softmax scores of those two, leading to the decision of classifying \\(x_{1}\\) as \"person\" with the higher score. For the pixel \\(x_{2}\\), there is no expert predicting its own class of expertise, i.e., \\(\\mathcal{X}(x_2) = \\emptyset\\), and the decision is taken by the default classifier \\(\\mathbf{W}_{\\mathrm{CLIP}}\\); the pixel \\(x_{2}\\) thus takes the scores from \\(\\mathbf{S}_{\\mathrm{CLIP}}(x_2)\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.351, + 0.182, + 0.367 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.378, + 0.484, + 0.513 + ], + "angle": 0, + "content": "Given access to a dataset \\(\\mathcal{D}\\) of unlabeled images, our goal is to improve the performance of OVSS using the aforementioned class-experts, without training or access to any labels. To do so, for each class \\(k\\in [1,K]\\) our method first identifies a small set of class-experts among the \\(M\\) predefined ImageNet templates of CLIP (Sec. 4.1) by leveraging unsupervised metrics. It then builds on a simple scheme to fuse class-experts' predictions into a single OVSS semantic output (Sec. 4.2)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.525, + 0.458, + 0.541 + ], + "angle": 0, + "content": "4.1. Unsupervised identification of class-experts" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.547, + 0.483, + 0.773 + ], + "angle": 0, + "content": "Referring to our definition of class-experts in Sec. 3, for each of the \\(K\\) classes, our goal is to estimate a set of class-experts \\(\\hat{\\mathcal{E}}_k\\), among the \\(M\\) pre-defined CLIP templates. Being in an unsupervised setting, we have no access to the true classifier performance (i.e., \\(\\Omega(\\cdot)\\) in Eq. (4)) and therefore propose relying on unsupervised metrics which act as performance proxy [62, 70]. We leverage entropy as it is an established measure of uncertainty [18, 30], commonly adopted in deep learning literature [32, 34, 61]. The entropy value of a softmax prediction indicates a classifier's confidence, which empirically provides a reliable indicator of prediction quality. In our setting, entropy measures the confidence of single-template classifiers for each class, helping identify class-experts. We also study other unsupervised metrics in the ablations (Sec. 5.2)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.775, + 0.483, + 0.851 + ], + "angle": 0, + "content": "In detail, we estimate class-experts as the set of individual templates whose associated classifiers (cf., Eq. (3)) exhibit low entropy. Specifically, for each template \\(m\\) and class \\(k\\), we compute the average entropy \\(\\mathcal{H}_{m,k}\\) of the \\(C_{m,k}\\) pixels predicted as class \\(k\\) by \\(\\pmb{W}(\\mathcal{T}_m)\\) in all images of \\(\\mathcal{D}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.863, + 0.483, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {H} _ {m, k} = - \\frac {1}{C _ {m , k}} \\sum_ {i = 1} ^ {C _ {m, k}} \\mathbf {q} _ {i} ^ {\\mathrm {T}} \\log (\\mathbf {q} _ {i}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.352, + 0.907, + 0.43 + ], + "angle": 0, + "content": "where \\(\\mathbf{q}_i\\in \\Delta^{K - 1}\\) is pixel-wise probability prediction obtained by applying softmax(·) over \\(K\\) class-wise cosine similarities, and \\(\\Delta^{K - 1}\\) is the \\((K - 1)\\)-simplex in \\(\\mathbb{R}^K\\). Subsequently, for class \\(k\\), the estimated set of experts is obtained as the \\(N\\) templates yielding the lowest entropy:" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.44, + 0.905, + 0.464 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {E}} _ {k} = \\left\\{\\mathcal {T} _ {\\hat {m}} \\mid \\hat {m} \\in \\operatorname {T o p - N} (\\underset {m} {\\arg \\operatorname {s o r t}} \\mathcal {H} _ {m, k}) \\right\\}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.907, + 0.632 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathcal{E}}_k\\) is expected to be an estimation of \\(\\mathcal{E}_k\\) (Eq. (4)); the arg sort operator returns the sorted list of \\(M\\) template indices so that \\(\\mathcal{H}_{m,k}\\) are in ascending order. Finally, \\(\\mathrm{Top - N}(\\cdot)\\) thus returns \\(\\mathbf{N}\\) template indices corresponding to the lowest entropy entries. In practice, we use \\(N = 4\\) and validate this choice through an ablation study in our experimental section. Following Eq. (2), the expert classifier of class \\(k\\), denoted \\(\\pmb {W}(\\hat{\\mathcal{E}}_k)\\), is constructed by averaging the embeddings obtained from the \\(N\\) templates in \\(\\hat{\\mathcal{E}}_k\\) for each of the \\(K\\) class names." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.633, + 0.907, + 0.772 + ], + "angle": 0, + "content": "The above process leads to \\(K\\) classifiers \\(\\{\\pmb{W}(\\hat{\\mathcal{E}}_k)\\}_{k\\in [1,K]}\\), all obtained in a fully unsupervised manner, and each is expected to excel on one specific class. However, it is important to note that although a classifier is expert of a single class, it outputs a full semantic map containing all \\(K\\) classes. Therefore, the above formulation produces \\(K\\) soft-segmentation maps, \\(\\{\\hat{\\mathbf{S}}_k\\}_{k\\in [1,K]}\\). To output a single OVSS map, we introduce a simple scheme to fuse all expert predictions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.782, + 0.766, + 0.798 + ], + "angle": 0, + "content": "4.2. Fusion of expert predictions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.805, + 0.907, + 0.902 + ], + "angle": 0, + "content": "To fuse all predicted soft-segmentation maps \\(\\{\\hat{\\mathbf{S}}_k\\}_{k\\in [1,K]}\\) into a single output \\(\\hat{\\mathbf{S}}\\), we rely on the fact that each map \\(\\hat{\\mathbf{S}}_k\\) excels at segmenting class \\(k\\). Therefore, for each pixel, the class assigned is the one having the highest probability among the expert classifiers which predicted their own class of expertise. Occasionally, when no expert predicts its own" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21474" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.152 + ], + "angle": 0, + "content": "class of expertise for a given pixel, we simply revert to using the \\( W_{\\mathrm{CLIP}} \\) classifier. In practice, we observe that only \\( \\approx 2\\% \\) of pixels fall under the latter case. Therefore, the vast majority of decisions are carried out by experts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.153, + 0.484, + 0.241 + ], + "angle": 0, + "content": "Formally, for a given pixel \\(x\\), \\(\\hat{\\mathbf{S}}_k(x)\\) is the \\(K\\)-dimensional softmax probability of the expert \\(k\\) at the position of \\(x\\). Let \\(\\mathcal{X}(x) = \\{k \\mid \\arg \\max_{i \\in [1, K]} \\hat{\\mathbf{S}}_k(x)[i] = k\\}\\) be the index set of experts that predicted their own class of expertise for \\(x\\). The prediction of \\(x\\) is determined as:" + }, + { + "type": "equation", + "bbox": [ + 0.103, + 0.25, + 0.482, + 0.319 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {P}} (x) = \\left\\{ \\begin{array}{l l} k ^ {*} = \\underset {k \\in \\mathcal {X} (x)} {\\arg \\max } \\hat {\\mathbf {S}} _ {k} (x) [ k ], & \\text {i f} \\mathcal {X} (x) \\neq \\varnothing , \\\\ k ^ {\\dagger} = \\underset {k \\in [ 1, K ]} {\\arg \\max } \\mathbf {S} _ {\\operatorname {C L I P}} (x) [ k ], & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.32, + 0.485, + 0.471 + ], + "angle": 0, + "content": "Fig. 2 illustrates the unsupervised identification of class-experts as well as the fusion strategy. Overall, our method incurs computational overhead primarily from computing cosine similarities for the \\(K\\) class-experts, which scales with the number of classes and becomes more significant when \\(K\\) is very large (\\(>100\\)). However, this overhead is constrained to the similarity computations since the vision encoder is shared across all experts, requiring only a single forward pass regardless of the number of classes. We provide computational details in Appendix A.2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.483, + 0.226, + 0.5 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.509, + 0.484, + 0.658 + ], + "angle": 0, + "content": "Experimental setup. We evaluate our method on three state-of-the-art open-vocabulary semantic segmentation models: the seminal MaskCLIP [79] and two recent models CLIP-DINOiser [66] and NACLIP [23]. CLIP-DINOiser trains a convolution layer using only 1K unlabeled images from ImageNet [15] for fast inference on pixel-level segmentation tasks, effectively imitating the DINO priors for pooling guidance and removing the need for DINO encoder at runtime. MaskCLIP and NACLIP are training-free adaptation methods of CLIP." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.484, + 0.901 + ], + "angle": 0, + "content": "For a fair comparison, we report the version of NACLIP without refinement. We primarily select training-free models that keep the image encoder frozen and produce unaltered, dense image-language features. This enables us to assess the impact of the class-experts identification strategy by using the original CLIP pre-trained features and ImageNet templates. Unless otherwise stated, all models utilize CLIP ViT-B/16 backbone. CLIP-DINOiser and MaskCLIP are based on OpenCLIP [11], while NACLIP employs OpenAI's original CLIP [47]. For all models - CLIP-DINOiser, NACLIP, and MaskCLIP- we follow the authors' evaluation protocol. For CLIP-DINOiser and MaskCLIP, we resize images to 448 pixels on the shortest side with a sliding window of size 448 and stride 224. For NACLIP, images are resized to 336 pixels (except for Cityscapes which are resized to 560 pixels) with a 224 window and stride 112." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.089, + 0.906, + 0.296 + ], + "angle": 0, + "content": "
MethodCSVOC20PC59ADEStuffAvg
CLIP [47]6.749.111.23.25.715.2
GroupViT [71]11.179.723.49.215.327.7
CLIP-Surgery [37]31.4--12.921.9-
GEM [3]--32.615.7--
SCLIP [63]32.280.434.216.122.437.1
CLIP-DIY [65]11.679.719.89.913.326.9
TCL [7]23.177.530.314.919.633.1
ReCo [53]21.157.822.311.214.825.4
MaskCLIP [79]25.061.825.514.217.528.7
+ FLOSS25.861.826.214.917.829.3
NACLIP [23]35.579.735.217.423.338.2
+ FLOSS37.080.235.918.423.639.0
CLIP-DINOiser [66]31.380.935.920.024.638.5
+ FLOSS34.682.336.220.724.739.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.307, + 0.907, + 0.406 + ], + "angle": 0, + "content": "Table 1. OVSS across datasets and models. We report mIoU metric for eight OVSS baselines and three models either using the average templates (i.e., MaskCLIP, NACLIP, and CLIP-DINOiser) or with our method (i.e., +FLOSS). FLOSS consistently improves OVSS models across datasets of varying complexity, from urban scenes (Cityscapes, 19 classes) to general objects (COCO-Stuff, 171 classes). Bold highlights the best performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.43, + 0.907, + 0.641 + ], + "angle": 0, + "content": "Datasets and Metric. We evaluate FLOSS across multiple semantic segmentation benchmarks: PASCAL CONTEXT 59 [42] (PC59), PASCAL VOC 20 [17] (VOC20), COCO-Stuff [5] (Stuff), ADE20K [78] (ADE), and Cityscapes [13] (CS). Note that these datasets exhibit different ontology and semantic granularity, ranging from 19 classes (CS) to 171 classes (Stuff). Additionally, we demonstrate the generalization capabilities of our class-experts identified on Cityscapes to other driving datasets such as BDD-100K [75] and MAPILLARY [44] as well as ACDC [52] under Night, Fog, Rain and Snow conditions. Importantly, our approach is training-free and uses no labels. Performance is measured using mean Intersection over Union (mIoU)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.737 + ], + "angle": 0, + "content": "Implementation details. The only hyperparameter of FLOSS, being the number of class-experts to identify in Eq. (6), is set to \\( N = 4 \\) in all experiments. While our method is training-free, class-experts are estimated using the training set of each dataset, while performance is always reported on the corresponding validation set." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.649, + 0.758 + ], + "angle": 0, + "content": "5.1. Main results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Tab. 1 reports the mIoU of OVSS baselines across five datasets showing FLOSS plugged into three different models: MaskCLIP, NACLIP and CLIP-DINOiser. The performance gains vary with dataset complexity, showing smaller improvements on VOC20 \\([+0.05, +0.5, +1.4]\\) or PC59 \\([+0.7, +0.7, +0.3]\\) than on CS \\([+0.8, +1.5, +3.3]\\). We attribute this to VOC20 and PC59 being closer to ImageNet which was leveraged in CLIP [47] to engineer the set of 80 templates used in all models. The benefit of our method" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21475" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.483, + 0.146 + ], + "angle": 0, + "content": "
MethodCSVOC20PC59ADEStuffAvg
MaskCLIP + FLOSS51%54%56%57%56%55%
NACLIP + FLOSS49%20%57%57%52%47%
CLIP-DINOiser + FLOSS62%41%57%45%44%50%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.157, + 0.483, + 0.2 + ], + "angle": 0, + "content": "Table 2. Quality of our estimated class-experts. We measure the quality of our estimated experts by evaluating their intersection with the true experts, cf. Eq. (8)." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.212, + 0.485, + 0.323 + ], + "angle": 0, + "content": "
Methodin domain CSout of domain
NightFogRainSnowBDDMAPAvg
MaskCLIP24.513.320.321.019.922.725.820.5
+ FLOSS25.813.921.722.321.322.926.821.5
NACLIP35.523.332.930.732.631.435.331.0
+ FLOSS37.023.632.531.033.632.336.531.6
CLIP-DINOiser31.312.724.127.927.231.335.226.4
+ FLOSS34.616.629.029.829.432.236.628.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.334, + 0.484, + 0.418 + ], + "angle": 0, + "content": "Table 3. Generalization of experts across datasets. We evaluate the transferability of experts identified on the Cityscapes (CS) dataset to unseen datasets that share the same semantic classes and exhibit changes in lighting (Night), weather (Fog, Rain, Snow), or location and scenery (BDD, MAP). Results show a consistent improvement when using FLOSS, across almost all settings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.443, + 0.483, + 0.489 + ], + "angle": 0, + "content": "is evident on challenging benchmarks like ADE and Stuff, which have a large number of semantic categories making accurate segmentation particularly hard." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.484, + 0.539 + ], + "angle": 0, + "content": "Quality of experts. We evaluate the quality \\(\\hat{\\rho}_k\\) of our experts by measuring, for each class \\(k\\), the intersection of the estimated class-expert \\(\\hat{\\mathcal{E}}_k\\) with the true set of experts \\(\\mathcal{E}_k\\):" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.548, + 0.483, + 0.587 + ], + "angle": 0, + "content": "\\[\n\\hat {\\rho} _ {k} = 1 0 0 \\times \\frac {\\left| \\hat {\\mathcal {E}} _ {k} \\cap \\mathcal {E} _ {k} \\right|}{\\left| \\hat {\\mathcal {E}} _ {k} \\right|} = 1 0 0 \\times \\frac {\\left| \\hat {\\mathcal {E}} _ {k} \\cap \\mathcal {E} _ {k} \\right|}{N}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.595, + 0.484, + 0.716 + ], + "angle": 0, + "content": "Where \\( |.| \\) denotes the cardinality of a set. Quality averaged over classes is reported in Tab. 2 and shows that regardless of the model used, we correctly estimate approximately half of the true experts. In some scenarios, such as PASCAL VOC 20 with NACLIP, even when only \\( 20\\% \\) of predicted experts are true experts, it is sufficient to bring some improvements (\\( +0.5 \\) mIoU in Tab. 1). Per-class quality is reported in Appendix A.4, showing some variability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Generalization to unseen datasets. We now evaluate the generalization of experts to other unseen datasets sharing the same label set but exhibiting distribution shifts. In Tab. 3, we report performance of FLOSS when experts are identified on Cityscapes and used to evaluate performance on six unseen datasets which include four ACDC conditions (Night, Fog, Rain, Snow), BDD-100K, and MAPILLARY. Overall, we consistently improve the performance with notable gains of up to \\(+4.9\\) mIoU for CLIP-DINOiser on ACDC Fog. This demonstrates that our class-wise experts are readily usable on other datasets sharing the same semantic classes, despite the presence of distribution shifts." + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.087, + 0.709, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.574, + 0.188, + 0.652, + 0.201 + ], + "angle": 0, + "content": "(a) Cityscapes" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.089, + 0.905, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.751, + 0.188, + 0.868, + 0.2 + ], + "angle": 0, + "content": "(b) PASCAL VOC 20" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.213, + 0.908, + 0.298 + ], + "angle": 0, + "content": "Figure 3. Effect of the number of images. We evaluate CLIP-DINOiser with FLOSS on urban scenes (a) and general object segmentation (b), when varying the number of images used to identify class-experts. Results show the average of 10 seeds, with standard deviation. Despite larger variation, we note the robustness of our method, even when having access to very few unlabeled images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.332, + 0.907, + 0.59 + ], + "angle": 0, + "content": "Low-data regimes. To investigate the data efficiency of FLOSS, we study its performance in lower data regimes. Fig. 3 reports the mIoU of CLIP-DINOiser on Cityscapes and PASCAL VOC 20, as the number of images sampled from their respective training sets (containing 2,975 and 1,464 images, respectively) varies. The figure shows the average and standard deviation over 10 seeds (i.e., randomly sampling 10 different image sets). On both datasets, the general trend shows that our method benefits from accessing more images, which stems from the better estimation of class-experts. Remarkably, our method performs well in a very low-data regime, outperforming the default CLIP-DINOiser with only 1 image on Cityscapes and around 25 images on PASCAL VOC 20. We conjecture that the latter requires more images due to its lower per-image class density compared to Cityscapes, which exhibits higher variability and density per image." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.601, + 0.906, + 0.708 + ], + "angle": 0, + "content": "Low-data regimes & domain shift. To further corroborate its practicality, we evaluate FLOSS under stricter conditions combining both low-data regimes and domain shift. Specifically, we use only a single CS image to identify experts that are tested on 6 shifted domains. Tab. 4 provides evidence on the effectiveness of FLOSS in this strict setting, leading to an average improvement of \\(1.9\\%\\) across OOD datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Expert transferability. We investigate the cross-dataset transferability of class-experts by evaluating whether experts identified on one dataset can effectively improve performance on semantically related but distinct datasets. We employ COCO-Stuff (171 classes) to identify class-experts for CLIP-DINOiser and apply them to Cityscapes, PASCAL VOC 20, and PASCAL CONTEXT 59, which share 10, 16, and 37 common classes with COCO-Stuff, respectively. For non-overlapping classes, we use the default average classifier \\( W_{\\mathrm{CLIP}} \\). As reported in Tab. 5, FLOSS shows consistent improvements, confirming expert transferability across datasets." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21476" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.483, + 0.145 + ], + "angle": 0, + "content": "
Methodin domain CSout of domain
NightFogRainSnowBDDMAPAvg
CLIP-DINOiser31.312.724.127.927.231.335.226.4
+ FLOSS33.216.228.229.328.731.735.628.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.156, + 0.483, + 0.227 + ], + "angle": 0, + "content": "Table 4. FLOSS with 1 image & domain shift. Experts are identified using a single Cityscapes (CS) image and evaluated on unseen datasets that share the same semantic classes and exhibit changes in lighting (Night), weather (Fog, Rain, Snow), or location and scenery (BDD, MAP). We observe a consistent boost." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.242, + 0.485, + 0.309 + ], + "angle": 0, + "content": "
Methodin-domain Stuffout of domain
CS commonallVOC20 commonallPC59 commonall
CLIP-DINOiser24.634.531.180.580.839.336.0
+ FLOSS24.735.832.282.382.440.036.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.319, + 0.484, + 0.389 + ], + "angle": 0, + "content": "Table 5. Cross-dataset expert transferability. Class-experts identified on COCO-Stuff are applied to CS, VOC20, and PC59. Results are reported for common classes shared with COCO-Stuff and all classes. FLOSS demonstrates consistent improvements across all target datasets." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.417, + 0.245, + 0.433 + ], + "angle": 0, + "content": "5.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.439, + 0.483, + 0.665 + ], + "angle": 0, + "content": "Fusion strategies for expert predictions. We investigate alternative strategies to fuse the expert predictions (cf., Sec. 4.2), reporting performance on Cityscapes in Tab. 6. Simply averaging all soft segmentation maps (\"Average-all\") irrespective of the templates class of expertise leads to suboptimal results across all models. When accounting for templates' expertise in fusion, conflicts arise often as several experts may predict their own class of expertise. We therefore explore three strategies to resolve conflicts: \"Default\" which falls back to \\( W_{\\mathrm{CLIP}} \\) predictions, \"Average\" which averages the probability vectors of conflicting experts, and \"Highest\" which selects the most confident expert as described in Sec. 4.2. The latter consistently yields the best performance across all models, validating our fusion scheme." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.672, + 0.483, + 0.748 + ], + "angle": 0, + "content": "Metrics for expert identification. We replace entropy in Eq. (6) with other unsupervised metrics to act as proxies of the template class-wise performance, reporting performance on Cityscapes using the CLIP-DINOiser model in Fig. 4a." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.748, + 0.483, + 0.778 + ], + "angle": 0, + "content": "(i) Avg. Probability computes the average probability assigned to class \\( k \\) across all pixels predicted as class \\( k \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.779, + 0.483, + 0.839 + ], + "angle": 0, + "content": "(ii) \\( MaNo \\) [70] is adapted as it was originally designed for assessing general model performance in image classification, to evaluate class-wise performance in semantic segmentation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "(iii) \"ITI\" [25] is the Inter-class separability to Intra-class similarity ratio, and captures both the separability of the classes in the feature space as well as the ability to keep samples of the same class close together." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.748, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.539, + 0.089, + 0.888, + 0.185 + ], + "angle": 0, + "content": "
FusionConflictMaskCLIP +FLOSSNACLIP +FLOSSCLIP-DINOiser +FLOSS
Average-all24.736.631.2
ExpertDefault24.836.632.0
ExpertAverage25.636.734.4
ExpertHighest25.837.034.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.195, + 0.907, + 0.238 + ], + "angle": 0, + "content": "Table 6. Strategies for the fusion of expert predictions. Exploring various fusion strategies on Cityscapes shows the benefit of our strategy, which outperforms all others across models." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.253, + 0.691, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.554, + 0.342, + 0.652, + 0.354 + ], + "angle": 0, + "content": "(a) Expert metrics" + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.253, + 0.906, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.749, + 0.361, + 0.853, + 0.373 + ], + "angle": 0, + "content": "(b) Effect of Top-N" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.385, + 0.907, + 0.485 + ], + "angle": 0, + "content": "Figure 4. Ablation of experts. We ablate our choices of metrics to identify experts as well as the number of experts per class. (a) shows unsupervised metrics to identify the experts, showing that our choice of entropy yields the best performance, although 'Avg. Probability' performs on par. (b) reports performance of FLOSS when varying the number of \\( N \\) experts. Results are reported on the Cityscapes dataset using the CLIP-DINOiser model." + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.499, + 0.821, + 0.621 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.631, + 0.907, + 0.701 + ], + "angle": 0, + "content": "Figure 5. Effect of class-expert quality. We report the performance of FLOSS in the oracle-based setting, where each class-wise set of 4 experts contains a ratio \\((\\rho)\\) of true experts. The experiment is conducted on Cityscapes using the CLIP-DINOiser model." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.729, + 0.808, + 0.744 + ], + "angle": 0, + "content": "These metrics are detailed in Appendix A.5." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.745, + 0.905, + 0.775 + ], + "angle": 0, + "content": "Overall, our entropy measure (cf. Eq. (5)) is the most effective, although Avg. Probability performs on par." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.855 + ], + "angle": 0, + "content": "Impact of class-expert identification quality. At the core of our method is the identification of class-experts. While we demonstrated there are templates which are class-expert, for each class there exist also under-performing templates (cf. Fig. 1) which, if selected, could harm the performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.902 + ], + "angle": 0, + "content": "To measure the impact of experts' quality on FLOSS, we devise an oracle-based experiment using \\( N = 4 \\) templates per-class, where we have access to ground truth la" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21477" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.113, + 0.089, + 0.462, + 0.189 + ], + "angle": 0, + "content": "
MethodViT-B/16ViT-L/14
CSPC59Avg.CSPC59Avg.
SCLIP [63]32.234.233.221.325.223.3
GEM [3]32.635.634.127.128.127.6
NACLIP [23]35.535.235.431.432.131.8
+ FLOSS37.035.936.532.432.432.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.2, + 0.483, + 0.242 + ], + "angle": 0, + "content": "Table 7. Effect of backbone. Results (mIoU) show that NACLIP + FLOSS outperforms NACLIP across ViT-B/16 and ViT-L/14 CLIP backbones." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.283, + 0.483, + 0.526 + ], + "angle": 0, + "content": "bels to identify the true class-experts. We randomly sample a fixed ratio \\(\\rho\\) of templates among true experts (i.e., \\(\\mathcal{E}_k\\) from Eq. (4)) and the remaining ones among non-experts (i.e., \\(\\{\\mathcal{T}_m\\} \\backslash \\mathcal{E}_k\\)). For example, a ratio of \\(\\rho = 75\\%\\) indicates that 3 templates are experts and 1 is non-expert. The outcome of this experiment on CS using the CLIP-DINOiser model is reported in Fig. 5 and shows an expected performance boost when the ratio of true experts increases. It also highlights that \\(50\\%\\) of true experts is sufficient for FLOSS to surpass the original CLIP-DINOiser. Interestingly, the performance of CLIP-DINOiser + FLOSS in Tab. 1 (34.6) slightly surpasses the performance at \\(\\rho = 100\\%\\) shown in Fig. 5 (34.1). This however does not indicate that our estimated templates in Tab. 1 are all true experts - as reflected by Tab. 2 - but rather that our correctly estimated experts are on average better than random true experts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.532, + 0.484, + 0.624 + ], + "angle": 0, + "content": "To further measure the upper bound of CLIP-DINOiser + FLOSS, we implement an oracle version using only the best true class-expert (i.e., \\(\\{\\mathcal{T}_{\\tilde{m}} \\mid \\tilde{m} \\in \\mathrm{Top-N}(\\Omega_k(W(\\mathcal{T}_m)))\\}\\)). On CS, the oracle with \\(N \\in \\{1, 2, 3, 4\\}\\) achieves respectively 37.1/37.0/37.3/37.5%, showing the potential for improvement assuming the identification of better experts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.641, + 0.483, + 0.763 + ], + "angle": 0, + "content": "FLOSS using the validation set. First, we investigate the effectiveness of FLOSS in a transductive setting. Using the validation set of CS for expert selection, CLIP-DINOiser + FLOSS achieves \\(34.4\\%\\) mIoU on the same set, compared to \\(31.3\\%\\) for CLIP-DINOiser. Second, using a portion of the validation set to select experts, i.e., 5/25/50/75 val images, we achieve 32.8/33.0/33.0/33.3% on the rest of the validation images, compared to \\(30.9\\%\\) for CLIP-DINOiser alone." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Impact of different backbones. Tab. 7 demonstrates the effectiveness of our proposed FLOSS when integrated with NACLIP using the ViT-L/14 backbone. Our approach achieves consistent improvements, reaching \\(32.4\\%\\) mIoU on both Cityscapes and PC59 datasets, leading to a 0.6 gain in average performance over NACLIP. These results validate the effectiveness of our method when using large-scale vision transformers." + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.089, + 0.905, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.196, + 0.907, + 0.28 + ], + "angle": 0, + "content": "Figure 6. Performance of best-class experts. Performance gains (IoU) when using best class-experts versus standard template-averaging. Results are shown for CLIP-DINOiser, MaskCLIP and NACLIP. The bars represent the IoU difference between using best class-experts (selected with ground-truth labels) and the default template-averaging approach." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.307, + 0.794, + 0.324 + ], + "angle": 0, + "content": "6. Perspectives and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.333, + 0.907, + 0.485 + ], + "angle": 0, + "content": "Our experiments show that entropy is a strong estimator of class-experts. However, future research could develop even better proxies for prediction accuracy. Template-based methods appear especially promising, with significant potential for improving OVSS performance. As illustrated in Fig. 6, there is still a large gap to close—for example, selecting the best expert for the \"sky\" class could improve IoU by over 30 percentage points. These \"upper-bounds\" could be further increased by using more or augmented templates beyond the 80 currently used." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.501, + 0.634, + 0.517 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.528, + 0.907, + 0.785 + ], + "angle": 0, + "content": "We propose a new task for improving OVSS models without having access to labels and without training, from a prompt template perspective. Our work is motivated by an intriguing observation: for each class, some templates excel in segmenting it. We refer to these templates as class-experts. Given a set of unlabeled images, our FLOSS uses the entropy of predictions as an unsupervised metric to identify these class-experts, thus not requiring any training or labels. Once selected, we propose a simple scheme for fusing the predictions of the class-experts, resulting in an improved overall OVSS performance in the inductive setting, even in the presence of distribution shifts. Additionally, we corroborate the effectiveness of FLOSS, which benefits from being plug-and-play, and further show that only few unlabeled images can be sufficient for expert selection. Furthermore, the oracle performance shows the potential of the class-experts, paving the way for future research." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.79, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Acknowledgment. This research was partially funded by the French Agence Nationale de la Recherche (ANR) with the project SIGHT (ANR-20-CE23-0016) and supported by ELSA - European Lighthouse on Secure and Safe AI funded by the European Union under grant agreement No. 101070617. We sincerely thank Telecom Paris for providing the resources necessary to run our experiments and Nacereddine Laddaoui for his help with infrastructure. We are also grateful to Ivan Lopes for proofreading." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21478" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Stephen H Bach, Victor Sanh, Zheng-Xin Yong, Albert Webson, Colin Raffel, Nihal V Nayak, Abheesht Sharma, Taewoon Kim, M Saiful Bari, Thibault Fevry, et al. Prompt-source: An integrated development environment and repository for natural language prompts. In ACL, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.185, + 0.484, + 0.227 + ], + "angle": 0, + "content": "[2] Sule Bai, Yong Liu, Yifei Han, Haoji Zhang, and Yansong Tang. Self-calibrated clip for training-free open-vocabulary segmentation. arXiv preprint arXiv:2411.15869, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.228, + 0.483, + 0.282 + ], + "angle": 0, + "content": "[3] Walid Bousselham, Felix Petersen, Vittorio Ferrari, and Hilde Kuehne. Grounding everything: Emerging localization properties in vision-language transformers. In CVPR, 2024. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.284, + 0.482, + 0.324 + ], + "angle": 0, + "content": "[4] Maxime Bucher, Tuan-Hung Vu, Matthieu Cord, and Patrick Pérez. Zero-shot semantic segmentation. In NeurlPS, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.326, + 0.483, + 0.367 + ], + "angle": 0, + "content": "[5] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In CVPR, 2018. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.368, + 0.482, + 0.422 + ], + "angle": 0, + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.424, + 0.482, + 0.467 + ], + "angle": 0, + "content": "[7] Junbum Cha, Jonghwan Mun, and Byungseok Roh. Learning to generate text-grounded mask for open-world semantic segmentation from only image-text pairs. In CVPR, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.468, + 0.482, + 0.521 + ], + "angle": 0, + "content": "[8] Guangyi Chen, Weiran Yao, Xiangchen Song, Xinyue Li, Yongming Rao, and Kun Zhang. Plot: Prompt learning with optimal transport for vision-language models. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.523, + 0.482, + 0.578 + ], + "angle": 0, + "content": "[9] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. TPAMI, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.58, + 0.482, + 0.633 + ], + "angle": 0, + "content": "[10] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.635, + 0.482, + 0.703 + ], + "angle": 0, + "content": "[11] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.705, + 0.482, + 0.76 + ], + "angle": 0, + "content": "[12] Seokju Cho, Heeseong Shin, Sunghwan Hong, Anurag Arnab, Paul Hongsuck Seo, and Seungryong Kim. Cat-seg: Cost aggregation for open-vocabulary semantic segmentation. In CVPR, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[13] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.831, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[14] Ulindu De Silva, Didula Samaraweera, Sasini Wanigathunga, Kavindu Kariyawasam, Kanchana Ranasinghe, Muzammal Naseer, and Ranga Rodrigo. Test-time optimization for domain adaptive open vocabulary segmentation. arXiv preprint arXiv:2501.04696, 2025. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[16] Reza Esfandiarpoor, Cristina Menghini, and Stephen H Bach. If clip could talk: Understanding vision-language model representations through their preferred concept descriptions. In EMNLP, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.19, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[17] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.233, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[18] Yarin Gal. Uncertainty in deep learning. PhD Thesis, University of Cambridge, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.261, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[19] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. IJCV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.905, + 0.383 + ], + "angle": 0, + "content": "[20] Yunhao Ge, Jie Ren, Andrew Gallagher, Yuxiao Wang, Ming-Hsuan Yang, Hartwig Adam, Laurent Itti, Balaji Lakshminarayanan, and Jiaping Zhao. Improving zero-shot generalization and robustness of multi-modal models. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.385, + 0.905, + 0.426 + ], + "angle": 0, + "content": "[21] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In ECCV, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[22] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In ICLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.511 + ], + "angle": 0, + "content": "[23] Sina Hajimiri, Ismail Ben Ayed, and Jose Dolz. Pay attention to your neighbours: Training-free open-vocabulary semantic segmentation. In WACV, 2025. 1, 2, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.553 + ], + "angle": 0, + "content": "[24] Cong Han, Yujie Zhong, Dengjie Li, Kai Han, and Lin Ma. Open-vocabulary semantic segmentation with decoupled one-pass network. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[25] Dapeng Hu, Jian Liang, Jun Hao Liew, Chuhui Xue, Song Bai, and Xinchao Wang. Mixed samples as probes for unsupervised model selection in domain adaptation. In NeurIPS, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[26] Xuefeng Hu, Ke Zhang, Lu Xia, Albert Chen, Jiajia Luo, Yuyin Sun, Ken Wang, Nan Qiao, Xiao Zeng, Min Sun, et al. Reclip: Refine contrastive language image pre-training with source free domain adaptation. In WACV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[27] Haiwen Huang, Songyou Peng, Dan Zhang, and Andreas Geiger. Renovating names in open-vocabulary segmentation benchmarks. In NeurIPS, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.706, + 0.905, + 0.747 + ], + "angle": 0, + "content": "[28] Tony Huang, Jack Chu, and Fangyun Wei. Unsupervised prompt learning for vision-language models. arXiv preprint arXiv:2204.03649, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.748, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[29] Yunshi Huang, Fereshteh Shakeri, Jose Dolz, Malik Boudiaf, Houda Bahig, and Ismail Ben Ayed. Lp++: A surprisingly strong linear probe for few-shot clip. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[30] Eyke Hüllermeier and Willem Waegeman. *Aleatoric and epistemic uncertainty in machine learning: An introduction to concepts and methods.* ML, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.888 + ], + "angle": 0, + "content": "[31] Krishna Murthy Jatavallabhula, Alihusein Kuwajerwala, Qiao Gu, Mohd Omama, Tao Chen, Shuang Li, Ganesh Iyer, Soroush Saryazdi, Nikhil Keetha, Ayush Tewari, et al. ConceptFusion: Open-set multimodal 3d mapping. In RSS, 2023." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.888 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21479" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.133 + ], + "angle": 0, + "content": "[32] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? In NeurIPS, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.177 + ], + "angle": 0, + "content": "[33] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.22 + ], + "angle": 0, + "content": "[34] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. In NeurIPS, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.222, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[35] Mengcheng Lan, Chaofeng Chen, Yiping Ke, Xinjiang Wang, Litong Feng, and Wayne Zhang. Proxyclip: Proxy attention improves clip for open-vocabulary segmentation. In ECCV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[36] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.362 + ], + "angle": 0, + "content": "[37] Yi Li, Hualiang Wang, Yiqun Duan, Jiheng Zhang, and Xiaomeng Li. A closer look at the explainability of contrastive language-image pre-training. PR, 2025. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.364, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[38] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In CVPR, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.475 + ], + "angle": 0, + "content": "[39] Yuqi Lin, Minghao Chen, Wenxiao Wang, Boxi Wu, Ke Li, Binbin Lin, Haifeng Liu, and Xiaofei He. Clip is also an efficient segmenter: A text-driven approach for weakly supervised semantic segmentation. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.482, + 0.504 + ], + "angle": 0, + "content": "[40] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. In ICLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.506, + 0.482, + 0.56 + ], + "angle": 0, + "content": "[41] Matthias Minderer, Josip Djolonga, Rob Romijnders, Frances Hubis, Xiaohua Zhai, Neil Houlsby, Dustin Tran, and Mario Lucic. Revisiting the calibration of modern neural networks. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.616 + ], + "angle": 0, + "content": "[42] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The role of context for object detection and semantic segmentation in the wild. In CVPR, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[43] Jishnu Mukhoti, Tsung-Yu Lin, Omid Poursaeed, Rui Wang, Ashish Shah, Philip HS Torr, and Ser-Nam Lim. Open vocabulary semantic segmentation with patch aligned contrastive learning. In CVPR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.482, + 0.716 + ], + "angle": 0, + "content": "[44] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In ICCV, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[45] Sarah Parisot, Yongxin Yang, and Steven McDonagh. Learning to name classes for vision and language models. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[46] Sarah Pratt, Ian Covert, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[48] Kanchana Ranasinghe, Brandon McKinzie, Sachin Ravi, Yinfei Yang, Alexander Toshev, and Jonathon Shlens. Per" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "ceptual grouping in contrastive vision-language models. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[49] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[50] Karsten Roth, Jae Myung Kim, A Koepke, Oriol Vinyals, Cordelia Schmid, and Zeynep Akata. Waffling around for performance: Visual classification with random words and broad concepts. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[51] Ohad Rubin, Jonathan Herzig, and Jonathan Berant. Learning to retrieve prompts for in-context learning. In *NAACL-HLT*, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[52] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding. In CVPR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.359 + ], + "angle": 0, + "content": "[53] Gyungin Shin, Weidi Xie, and Samuel Albanie. Reco: Retrieve and co-segment for zero-shot transfer. In NeurIPS, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[54] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. In NeurIPS, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[55] Julio Silva-Rodriguez, Sina Hajimiri, Ismail Ben Ayed, and Jose Dolz. A closer look at the few-shot adaptation of large vision-language models. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[56] Vladan Stojnic, Yannis Kalantidis, and Giorgos Tolias. Label propagation for zero-shot classification with vision-language models. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.544 + ], + "angle": 0, + "content": "[57] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In ICCV, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[58] Weijie Tu, Weijian Deng, and Tom Gedeon. A closer look at the robustness of contrastive language-image pre-training (clip). In NeurIPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.591, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[59] Weijie Tu, Weijian Deng, Dylan Campbell, Stephen Gould, and Tom Gedeon. An empirical study into what matters for calibrating vision-language models. In ICML, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[60] Antonin Vobecky, Oriane Simeoni, David Hurych, Spyridon Gidaris, Andrei Bursuc, Patrick Pérez, and Josef Sivic. Pop-3d: Open-vocabulary 3d occupancy prediction from images. In NeurIPS, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[61] Tuan-Hung Vu, Himalaya Jain, Maxime Bucher, Matthieu Cord, and Patrick Pérez. Advent: Adversarial entropy minimization for domain adaptation in semantic segmentation. In CVPR, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[62] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In ICLR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.828 + ], + "angle": 0, + "content": "[63] Feng Wang, Jieru Mei, and Alan Yuille. Sclip: Rethinking self-attention for dense vision-language inference. In ECCV, 2024. 1, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[64] Zhengbo Wang, Jian Liang, Lijun Sheng, Ran He, Zilei Wang, and Tieniu Tan. A hard-to-beat baseline for training-free CLIP-based adaptation. In ICLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[65] Monika Wysoczańska, Michael Ramamonjisoa, Tomasz Trzeciński, and Oriane Simeoni. Clip-diy: Clip dense infer" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21480" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "ence yields open-vocabulary semantic segmentation for-free. In WACV, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.188 + ], + "angle": 0, + "content": "[66] Monika Wysoczanska, Oriane Simeoni, Michael Ramamonjisoa, Andrei Bursuc, Tomasz Trzciński, and Patrick Pérez. Clip-dinoiser: Teaching clip a few dino tricks for open-vocabulary semantic segmentation. In ECCV, 2024. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.244 + ], + "angle": 0, + "content": "[67] Yongqin Xian, Subhabrata Choudhury, Yang He, Bernt Schiele, and Zeynep Akata. Semantic projection network for zero-and few-label semantic segmentation. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.246, + 0.482, + 0.287 + ], + "angle": 0, + "content": "[68] Bin Xie, Jiale Cao, Jin Xie, Fahad Shahbaz Khan, and Yanwei Pang. Sed: A simple encoder-decoder for open-vocabulary semantic segmentation. In CVPR, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.482, + 0.342 + ], + "angle": 0, + "content": "[69] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. In NeurIPS, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.398 + ], + "angle": 0, + "content": "[70] Renchunzi Xie, Ambroise Odonnat, Vasilii Feofanov, Weijian Deng, Jianfeng Zhang, and Bo An. Mano: Exploiting matrix norm for unsupervised accuracy estimation under distribution shifts. In NeurIPS, 2024. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[71] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. Groupvit: Semantic segmentation emerges from text supervision. In CVPR, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.482, + 0.509 + ], + "angle": 0, + "content": "[72] Jilan Xu, Junlin Hou, Yuejie Zhang, Rui Feng, Yi Wang, Yu Qiao, and Weidi Xie. Learning open-vocabulary semantic segmentation models from natural language supervision. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.565 + ], + "angle": 0, + "content": "[73] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[74] Mengde Xu, Zheng Zhang, Fangyun Wei, Han Hu, and Xi-ang Bai. Side adapter network for open-vocabulary semantic segmentation. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.609, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[75] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darryll. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[76] Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang-Chieh Chen. Convolutions die hard: Open-vocabulary segmentation with single frozen convolutional clip. In NeurIPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.482, + 0.774 + ], + "angle": 0, + "content": "[77] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.828 + ], + "angle": 0, + "content": "[78] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. IJCV, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.482, + 0.859 + ], + "angle": 0, + "content": "[79] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In ECCV, 2022. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[80] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In CVPR, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.133 + ], + "angle": 0, + "content": "[81] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.175 + ], + "angle": 0, + "content": "[82] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[83] Xiangyang Zhu, Renrui Zhang, Bowei He, Aojun Zhou, Dong Wang, Bin Zhao, and Peng Gao. Not all features matter: Enhancing few-shot clip with adaptive prior refinement. In ICCV, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21481" + } + ] +] \ No newline at end of file diff --git a/2025/FLOSS_ Free Lunch in Open-vocabulary Semantic Segmentation/759adae1-3d57-4b38-8780-a8c8f1ca245e_origin.pdf b/2025/FLOSS_ Free Lunch in Open-vocabulary Semantic Segmentation/759adae1-3d57-4b38-8780-a8c8f1ca245e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c5ba0af7145d74c2ae545c34fcebf9c5c1ee1542 --- /dev/null +++ b/2025/FLOSS_ Free Lunch in Open-vocabulary Semantic Segmentation/759adae1-3d57-4b38-8780-a8c8f1ca245e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c16d42e28ff041605f88555233443aba44a3eba665a7ff484aea4b22c4711d3b +size 634373 diff --git a/2025/FLOSS_ Free Lunch in Open-vocabulary Semantic Segmentation/full.md b/2025/FLOSS_ Free Lunch in Open-vocabulary Semantic Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d3e0600fa51845973c4da7d67bd8f612d69bb93d --- /dev/null +++ b/2025/FLOSS_ Free Lunch in Open-vocabulary Semantic Segmentation/full.md @@ -0,0 +1,344 @@ +# FLOSS: Free Lunch in Open-vocabulary Semantic Segmentation + +Yasser Benigim $^{1*}$ Mohammad Fahes $^{1}$ Tuan-Hung $\mathrm{Vu}^{1,2}$ Andrei Bursuc $^{1,2}$ Raoul de Charette $^{1}$ Inria $^{2}$ Valeo.ai + +# Abstract + +In this paper, we challenge the conventional practice in Open-Vocabulary Semantic Segmentation (OVSS) of using averaged class-wise text embeddings, which are typically obtained by encoding each class name with multiple templates (e.g., a photo of , a sketch of a ). We investigate the impact of templates for OVSS, and find that for each class, there exist single-template classifiers—which we refer to as class-experts—that significantly outperform the conventional averaged classifier. First, to identify these class-experts, we introduce a novel approach that estimates them without any labeled data or training. By leveraging the class-wise prediction entropy of single-template classifiers, we select those yielding the lowest entropy as the most reliable class-experts. Second, we combine the outputs of class-experts in a new fusion process. Our plug-and-play method, coined FLOSS, is orthogonal and complementary to existing OVSS methods, offering an improvement without the need for additional labels or training. Extensive experiments show that FLOSS consistently enhances state-of-the-art OVSS models, generalizes well across datasets with different distribution shifts, and delivers substantial improvements in low-data scenarios where only a few unlabeled images are available. Our code is available at https://github.com/yasserben/FLOSS. + +# 1. Introduction + +In the past decade, advances in deep learning and the growing amount of training data have enabled the challenging task of semantic segmentation, which involves assigning semantic labels to each pixel in an image. Initially, this was limited to predefined categories [9, 10, 57, 69], but recent models rely on vision-language alignment [47] for open-vocabulary segmentation, where the semantic categories can be dynamically expressed at runtime in natural language. + +In Open-Vocabulary Semantic Segmentation (OVSS), the CLIP model [47] is commonly used as the backbone, + +![](images/3a640d3901aa085653a9588d6e8c542193f0e637180093783acb282b862cd372.jpg) +Figure 1. Using individual vs average template classifiers. Empirically, we observe that for each class there exist individual templates (colored dots) that lead to classifiers performing better than the popular CLIP classifier, which is built from the averaging of all 80 templates embeddings $(-)$ . The analysis was conducted on Cityscapes using the CLIP-DINOiser model. + +which is either retrained [43], fine-tuned [12, 38], or frozen and adjusted to extend global image-language alignment to denser pixel-language alignment. In this work, we study CLIP's original classification paradigm and focus on methods that keep all CLIP parameters frozen [23, 63, 79]. Among these, MaskCLIP [79] leverages CLIP's embedding layer for segmentation, CLIP-DINOiser [66] enhances features via distillation from an external DINO [6] network, while NACLIP [23] modifies CLIP's self-attention to improve spatial consistency. A common aspect of prior OVSS methods is their reliance on CLIP's default classification process, where image or patch embeddings are compared with text embeddings representing semantic categories. To represent a class, a common practice is to use multiple templates to encode class names (e.g., "a photo of a car", "a sketch of a car", etc) and average their embeddings. Empirically, Radford et al. [47] engineered $M = 80$ templates to improve ImageNet zero-shot classification. For segmentation, OVSS methods [23, 66, 79] adopt by default the same templates to construct segmentation classifiers. While this "template-averaging" strategy works well overall, it may not always provide the best classifier for each semantic concept. + +Our paper explores the impact of individual templates in OVSS – an aspect that has been overlooked in the literature. Fig. 1 shows an intriguing empirical observation: for each class, some individual templates outperform the averaged classifier that uses all templates. We call such templates class-experts, highlighting that they are trivial to identify using semantic ground truths. + +However, without labels, two challenges arise: 1) How to identify class-experts without ground truth? 2) How to fuse predictions from class-experts? + +We propose that entropy, an unsupervised metric, can help identify expert templates without labels. This leads to a new task—training- and label-free template selection for OVSS—which improves segmentation by selecting better text template subsets for a given dataset of unlabeled images. Our method shifts the attention to the text modality, unlike previous works that tweak the visual encoder, and also aims to improve the performance in an inductive setting, where the selected templates are applied to a different subset of images. + +We summarize our contributions as follows: + +- We analyze the current template-averaging practice for OVSS, showing that better class-wise classifiers can be constructed using only a subset of all templates, and that these expert templates can be identified without labels. +- We propose the novel task of training- and label-free template selection for OVSS. Given an unlabeled set of images, this task aims to identify class-wise experts that improve OVSS performance on unseen counterparts of the dataset. +- We introduce FLOSS, which leverages class-wise per-pixel entropy to select expert templates and employs a simple and effective fusion scheme to combine individual expert predictions. + +With extensive experiments, we demonstrate that FLOSS consistently improves the state-of-the-art performance on OVSS benchmarks and exhibits generalization capabilities in the choice of experts. Furthermore, with only a few unlabeled images, our approach can select reasonably good expert templates, resulting in performance boost. + +# 2. Related Works + +Vision-Language Model (VLM) transferability. Since CLIP's release, its impressive zero-shot classification performance [47] and robustness [41, 58, 59] have made it a foundation for various adaptation and transfer learning tasks. Few-shot supervised fine-tuning, where only limited labeled data is available for downstream tasks like image classification, has gained significant attention. The goal is to leverage the vast knowledge of the VLM acquired through training on an immense image collection by steering it towards a task of interest via few labeled sam + +plies. Parameter efficiency is central to this setting. To ensure it, prompt learning [8, 33, 80-82], lightweight MLP training [19], and linear probing [29, 55] have been proposed. Another line of research explores training-free few-shot adaptation [64, 77, 83], where a classifier is built using few labeled samples and ensembled with the zero-shot classifier. Additionally, unsupervised adaptation has been explored [26, 28, 56], alongside test-time adaptation, where prompts are learned per test image [54]. + +While using VLMs for classification tasks aligns with their pre-training focus on global image recognition, applying them to dense prediction tasks like semantic segmentation is more challenging due to the gap between global image understanding and pixel-level prediction. We summarize related efforts in the following. + +Open-Vocabulary Semantic Segmentation (OVSS). OVSS uses VLMs like CLIP for text-based segmentation. However, CLIP's global pooling layers limit its ability to generate dense pixel-level, language-aligned features [31, 79]. To address this, several training-free methods densify CLIP without altering its parameters, preserving its image-language alignment. MaskCLIP [79] was the first to replace the attention pooling layer with a convolutional layer. Later works further densify CLIP by aggregating multiple image views [31, 66], integrating priors from other vision foundation models for better pooling [35, 66], extracting different information from the self-attention layers [3, 23, 37], or removing anomaly tokens [2]. Seg-TTO [14] leverages an LLM to generate discriminative attributes for rare classes in specialized domains (e.g., medical sciences, earth monitoring) that use technical terminology (e.g., mediastinum), operating in a test-time optimization framework while maintaining zero-shot capabilities. Other approaches fine-tune CLIP for segmentation tasks using weak supervision through image-level tags, captions [21, 48], or class-agnostic object masks [21, 49]. Others employ full supervision [12, 36, 38, 74] on densely annotated datasets like COCO Stuff [5], following strategies similar to early inductive zero-shot semantic segmentation methods [4, 67]. Among them, some fine-tune CLIP's image encoder [36, 38, 74], others combine this with additional backbones and segmentation heads [38], while some learn side networks while keeping CLIP frozen [74]. Many other approaches have been proposed in this active research area [24, 68, 72, 73, 76]. While effective, these methods require labeled data and computational resources, limiting their practicality in resource-constrained scenarios. + +Our work focuses on training-free approaches, evaluating FLOSS on three key methods: MaskCLIP [79], CLIP-DINOiser [66] and NACLIP [23], while being applicable to any text-based segmentation framework. + +Prompting in VLMs. CLIP's image-language interface has inspired the adoption of various prompt engineering practices from LLMs to better reveal the knowledge encapsulated in the vision encoder and boost performance. The seminal CLIP work [47] introduced dataset-specific templates (e.g., "a photo of ") to bridge the gap between training captions and test prompts for zero-shot classification. Subsequent methods have improved prompts by crowdsourcing templates [1], adjusting class names [21, 60], or modifying templates for tasks like object detection [21, 22] and occupancy prediction [60]. While effective, manual prompt tuning is labor-intensive, leading to automated strategies like adding random words or characters to the original template [50], enriching class names with WordNet concepts [20] or synonyms [39], learning more informative class names [27], and using LLMs to generate more expressive descriptions starting from simple class names or retrieve them from a dataset [51] and further use them as prompts. Some approaches also learn prompts for task-specific fitting [45] or test-time adaptation [54], and others mine them with external LLMs [16]. + +Our approach differs by selecting expert templates for each class in an unsupervised manner, starting from predefined templates. While previous methods have focused on image classification, we apply this strategy to semantic segmentation using CLIP's original ImageNet templates, which have been used by default in prior works. + +# 3. Preliminaries + +OVSS aims to produce a soft-segmentation map $\hat{\mathbf{S}}\in [0,1]^{H\times W\times K}$ given an image $\mathbf{l}\in \mathbb{R}^{H\times W\times 3}$ and a set of classes $\{\mathcal{C}_k\}_{k\in [1,K]}$ expressed in natural language. The final segmentation map $\hat{\mathbf{P}}$ is derived by applying arg max(·) on $\hat{\mathbf{S}}$ along the $K$ -dimension. Leveraging the aligned vision encoder $f(\cdot)$ and text encoder $g(\cdot)$ of CLIP, the standard practice consists in classifying an image patch by comparing its vision embedding to the text embedding of class-specific prompts constructed using predefined templates $\{\mathcal{T}_m\}_{m\in [1,M]}$ . An example of prompt is "a bright photo of a car", denoted + +$\mathcal{P}_{m,k}$ for compactness. Instead of a single template, practitioners use "template-averaging" to increase robustness and improve overall performance by averaging the embeddings of prompts obtained using all available templates [40, 46, 47, 50], which is a form of ensembling on the text encoder side. This writes $\bar{c}_k = \frac{1}{M}\sum_{m = 1}^{M}g(\mathcal{P}_{m,k})$ where $\bar{c}_k$ is the resulting representation of class $C_k$ . Subsequently, for each patch, the predicted class is the one that maximizes the cosine similarity between the visual and the + +text embeddings: + +$$ +\hat {k} = \arg \max _ {k} f (\mathcal {V}) ^ {\mathrm {T}} \bar {c} _ {k}, \tag {1} +$$ + +where $\mathcal{V}$ represents the visual patch. All embeddings are $\ell_2$ -normalized. The text embeddings of $K$ classes define a classifier that we denote: + +$$ +\boldsymbol {W} \left(\left\{\mathcal {T} _ {m} \right\} _ {m \in [ 1, M ]}\right) = \left\{\bar {c} _ {k} \right\} _ {k \in [ 1, K ]}. \tag {2} +$$ + +As in the above equation, the classifier can be represented as a function of the templates used to compute the average embeddings of all classes. + +Empirical observations on single templates. OVSS models [23, 66, 68, 79] typically rely on the $M = 80$ ImageNet templates of CLIP [47] to compute the average classifier of Eq. (2), which we refer to as $W_{\mathrm{CLIP}}$ for brevity. While using $W_{\mathrm{CLIP}}$ was shown to be robust [47, 79], the templates were originally hand-crafted for zero-shot classification on ImageNet [47], and their individual effect on class-wise performance remains unexplored for OVSS. Rather than $W_{\mathrm{CLIP}}$ that uses all $M$ templates, one can construct $M$ distinct classifiers, each utilizing a single template $\mathcal{T}_m$ . This writes: + +$$ +\boldsymbol {W} \left(\mathcal {T} _ {m}\right) = \left\{g \left(\mathcal {P} _ {m, k}\right) \right\} _ {k \in [ 1, K ]}. \tag {3} +$$ + +We design a systematic experiment to evaluate the performance of single-template classifiers $\mathbf{W}(\mathcal{T}_m)$ from Eq. (3). Fig. 1 shows the per-class performance on Cityscapes [13] of each of the 80 single-template classifiers (plotted as colored dots, e.g., ●), where colors encode the template identifier (i.e., m), along with the average $W_{\mathrm{CLIP}}$ classifier (plotted as —). This leads to two key observations: + +a) For each class, there exist single-template classifiers outperforming $W_{\mathrm{CLIP}}$ on this class; we refer to the corresponding templates as class-experts. More precisely, the set of class-experts is defined as: + +$$ +\mathcal {E} _ {k} = \left\{\mathcal {T} _ {m} \mid \Omega_ {k} \left(\boldsymbol {W} \left(\mathcal {T} _ {m}\right)\right) > \Omega_ {k} \left(\boldsymbol {W} _ {\text {C L I P}}\right), m \in [ 1, M ] \right\}, \tag {4} +$$ + +where $\Omega_{k}(\cdot)$ denotes the IoU performance on class $k$ . + +b) A single-template classifier excelling on one class may exhibit suboptimal performance on others, implying that for $k, l \in [1, K]$ the expert sets $\mathcal{E}_k$ and $\mathcal{E}_l$ may differ. + +Following the above observations, class-experts are readily identified when ground truth labels are accessible, through direct performance comparison. We refer readers to Appendix A.3 for similar observations across datasets and models. + +However, we are interested in the case where an unlabeled dataset is provided. Therefore, two questions stem from the previous empirical observations: (i) How to identify class-experts in an unsupervised, training-free manner? (ii) Given these identified class-experts, how to derive the final semantic predictions? + +![](images/8894a9fd171e353ebbc1bfe019331e61c5666bae754d5c2136c107ed1ca1a873.jpg) +Unsupervised identification of class-experts + +![](images/8e8322105906b9164149c821e09e9c34d9f4406143d13ba5d567163264ce7c59.jpg) +Expert Fusion +Figure 2. Overview of FLOSS. Left: for each class, selected expert templates $(\checkmark)$ construct the estimated set $\hat{\mathcal{E}}$ , which classifies the input image and yield the soft-segmentation map $\mathbf{S}$ . In this example we set $N = 3$ for the $\mathrm{Top - N}(\dots)$ template indices corresponding to lowest entropy entries to select. Right: we explain the fusion strategy for two cases. For the pixel $x_{1}$ , the set of experts that predict their own class of expertise $\mathcal{X}(x_1)$ contains 2 experts; consequently, the decision is taken by looking at the maximum softmax scores of those two, leading to the decision of classifying $x_{1}$ as "person" with the higher score. For the pixel $x_{2}$ , there is no expert predicting its own class of expertise, i.e., $\mathcal{X}(x_2) = \emptyset$ , and the decision is taken by the default classifier $\mathbf{W}_{\mathrm{CLIP}}$ ; the pixel $x_{2}$ thus takes the scores from $\mathbf{S}_{\mathrm{CLIP}}(x_2)$ . + +# 4. Method + +Given access to a dataset $\mathcal{D}$ of unlabeled images, our goal is to improve the performance of OVSS using the aforementioned class-experts, without training or access to any labels. To do so, for each class $k\in [1,K]$ our method first identifies a small set of class-experts among the $M$ predefined ImageNet templates of CLIP (Sec. 4.1) by leveraging unsupervised metrics. It then builds on a simple scheme to fuse class-experts' predictions into a single OVSS semantic output (Sec. 4.2). + +# 4.1. Unsupervised identification of class-experts + +Referring to our definition of class-experts in Sec. 3, for each of the $K$ classes, our goal is to estimate a set of class-experts $\hat{\mathcal{E}}_k$ , among the $M$ pre-defined CLIP templates. Being in an unsupervised setting, we have no access to the true classifier performance (i.e., $\Omega(\cdot)$ in Eq. (4)) and therefore propose relying on unsupervised metrics which act as performance proxy [62, 70]. We leverage entropy as it is an established measure of uncertainty [18, 30], commonly adopted in deep learning literature [32, 34, 61]. The entropy value of a softmax prediction indicates a classifier's confidence, which empirically provides a reliable indicator of prediction quality. In our setting, entropy measures the confidence of single-template classifiers for each class, helping identify class-experts. We also study other unsupervised metrics in the ablations (Sec. 5.2). + +In detail, we estimate class-experts as the set of individual templates whose associated classifiers (cf., Eq. (3)) exhibit low entropy. Specifically, for each template $m$ and class $k$ , we compute the average entropy $\mathcal{H}_{m,k}$ of the $C_{m,k}$ pixels predicted as class $k$ by $\pmb{W}(\mathcal{T}_m)$ in all images of $\mathcal{D}$ : + +$$ +\mathcal {H} _ {m, k} = - \frac {1}{C _ {m , k}} \sum_ {i = 1} ^ {C _ {m, k}} \mathbf {q} _ {i} ^ {\mathrm {T}} \log (\mathbf {q} _ {i}), \tag {5} +$$ + +where $\mathbf{q}_i\in \Delta^{K - 1}$ is pixel-wise probability prediction obtained by applying softmax(·) over $K$ class-wise cosine similarities, and $\Delta^{K - 1}$ is the $(K - 1)$ -simplex in $\mathbb{R}^K$ . Subsequently, for class $k$ , the estimated set of experts is obtained as the $N$ templates yielding the lowest entropy: + +$$ +\hat {\mathcal {E}} _ {k} = \left\{\mathcal {T} _ {\hat {m}} \mid \hat {m} \in \operatorname {T o p - N} (\underset {m} {\arg \operatorname {s o r t}} \mathcal {H} _ {m, k}) \right\}, \tag {6} +$$ + +where $\hat{\mathcal{E}}_k$ is expected to be an estimation of $\mathcal{E}_k$ (Eq. (4)); the arg sort operator returns the sorted list of $M$ template indices so that $\mathcal{H}_{m,k}$ are in ascending order. Finally, $\mathrm{Top - N}(\cdot)$ thus returns $\mathbf{N}$ template indices corresponding to the lowest entropy entries. In practice, we use $N = 4$ and validate this choice through an ablation study in our experimental section. Following Eq. (2), the expert classifier of class $k$ , denoted $\pmb {W}(\hat{\mathcal{E}}_k)$ , is constructed by averaging the embeddings obtained from the $N$ templates in $\hat{\mathcal{E}}_k$ for each of the $K$ class names. + +The above process leads to $K$ classifiers $\{\pmb{W}(\hat{\mathcal{E}}_k)\}_{k\in [1,K]}$ , all obtained in a fully unsupervised manner, and each is expected to excel on one specific class. However, it is important to note that although a classifier is expert of a single class, it outputs a full semantic map containing all $K$ classes. Therefore, the above formulation produces $K$ soft-segmentation maps, $\{\hat{\mathbf{S}}_k\}_{k\in [1,K]}$ . To output a single OVSS map, we introduce a simple scheme to fuse all expert predictions. + +# 4.2. Fusion of expert predictions + +To fuse all predicted soft-segmentation maps $\{\hat{\mathbf{S}}_k\}_{k\in [1,K]}$ into a single output $\hat{\mathbf{S}}$ , we rely on the fact that each map $\hat{\mathbf{S}}_k$ excels at segmenting class $k$ . Therefore, for each pixel, the class assigned is the one having the highest probability among the expert classifiers which predicted their own class of expertise. Occasionally, when no expert predicts its own + +class of expertise for a given pixel, we simply revert to using the $W_{\mathrm{CLIP}}$ classifier. In practice, we observe that only $\approx 2\%$ of pixels fall under the latter case. Therefore, the vast majority of decisions are carried out by experts. + +Formally, for a given pixel $x$ , $\hat{\mathbf{S}}_k(x)$ is the $K$ -dimensional softmax probability of the expert $k$ at the position of $x$ . Let $\mathcal{X}(x) = \{k \mid \arg \max_{i \in [1, K]} \hat{\mathbf{S}}_k(x)[i] = k\}$ be the index set of experts that predicted their own class of expertise for $x$ . The prediction of $x$ is determined as: + +$$ +\hat {\mathbf {P}} (x) = \left\{ \begin{array}{l l} k ^ {*} = \underset {k \in \mathcal {X} (x)} {\arg \max } \hat {\mathbf {S}} _ {k} (x) [ k ], & \text {i f} \mathcal {X} (x) \neq \varnothing , \\ k ^ {\dagger} = \underset {k \in [ 1, K ]} {\arg \max } \mathbf {S} _ {\operatorname {C L I P}} (x) [ k ], & \text {o t h e r w i s e .} \end{array} \right. \tag {7} +$$ + +Fig. 2 illustrates the unsupervised identification of class-experts as well as the fusion strategy. Overall, our method incurs computational overhead primarily from computing cosine similarities for the $K$ class-experts, which scales with the number of classes and becomes more significant when $K$ is very large ( $>100$ ). However, this overhead is constrained to the similarity computations since the vision encoder is shared across all experts, requiring only a single forward pass regardless of the number of classes. We provide computational details in Appendix A.2. + +# 5. Experiments + +Experimental setup. We evaluate our method on three state-of-the-art open-vocabulary semantic segmentation models: the seminal MaskCLIP [79] and two recent models CLIP-DINOiser [66] and NACLIP [23]. CLIP-DINOiser trains a convolution layer using only 1K unlabeled images from ImageNet [15] for fast inference on pixel-level segmentation tasks, effectively imitating the DINO priors for pooling guidance and removing the need for DINO encoder at runtime. MaskCLIP and NACLIP are training-free adaptation methods of CLIP. + +For a fair comparison, we report the version of NACLIP without refinement. We primarily select training-free models that keep the image encoder frozen and produce unaltered, dense image-language features. This enables us to assess the impact of the class-experts identification strategy by using the original CLIP pre-trained features and ImageNet templates. Unless otherwise stated, all models utilize CLIP ViT-B/16 backbone. CLIP-DINOiser and MaskCLIP are based on OpenCLIP [11], while NACLIP employs OpenAI's original CLIP [47]. For all models - CLIP-DINOiser, NACLIP, and MaskCLIP- we follow the authors' evaluation protocol. For CLIP-DINOiser and MaskCLIP, we resize images to 448 pixels on the shortest side with a sliding window of size 448 and stride 224. For NACLIP, images are resized to 336 pixels (except for Cityscapes which are resized to 560 pixels) with a 224 window and stride 112. + +
MethodCSVOC20PC59ADEStuffAvg
CLIP [47]6.749.111.23.25.715.2
GroupViT [71]11.179.723.49.215.327.7
CLIP-Surgery [37]31.4--12.921.9-
GEM [3]--32.615.7--
SCLIP [63]32.280.434.216.122.437.1
CLIP-DIY [65]11.679.719.89.913.326.9
TCL [7]23.177.530.314.919.633.1
ReCo [53]21.157.822.311.214.825.4
MaskCLIP [79]25.061.825.514.217.528.7
+ FLOSS25.861.826.214.917.829.3
NACLIP [23]35.579.735.217.423.338.2
+ FLOSS37.080.235.918.423.639.0
CLIP-DINOiser [66]31.380.935.920.024.638.5
+ FLOSS34.682.336.220.724.739.7
+ +Table 1. OVSS across datasets and models. We report mIoU metric for eight OVSS baselines and three models either using the average templates (i.e., MaskCLIP, NACLIP, and CLIP-DINOiser) or with our method (i.e., +FLOSS). FLOSS consistently improves OVSS models across datasets of varying complexity, from urban scenes (Cityscapes, 19 classes) to general objects (COCO-Stuff, 171 classes). Bold highlights the best performance. + +Datasets and Metric. We evaluate FLOSS across multiple semantic segmentation benchmarks: PASCAL CONTEXT 59 [42] (PC59), PASCAL VOC 20 [17] (VOC20), COCO-Stuff [5] (Stuff), ADE20K [78] (ADE), and Cityscapes [13] (CS). Note that these datasets exhibit different ontology and semantic granularity, ranging from 19 classes (CS) to 171 classes (Stuff). Additionally, we demonstrate the generalization capabilities of our class-experts identified on Cityscapes to other driving datasets such as BDD-100K [75] and MAPILLARY [44] as well as ACDC [52] under Night, Fog, Rain and Snow conditions. Importantly, our approach is training-free and uses no labels. Performance is measured using mean Intersection over Union (mIoU). + +Implementation details. The only hyperparameter of FLOSS, being the number of class-experts to identify in Eq. (6), is set to $N = 4$ in all experiments. While our method is training-free, class-experts are estimated using the training set of each dataset, while performance is always reported on the corresponding validation set. + +# 5.1. Main results + +Tab. 1 reports the mIoU of OVSS baselines across five datasets showing FLOSS plugged into three different models: MaskCLIP, NACLIP and CLIP-DINOiser. The performance gains vary with dataset complexity, showing smaller improvements on VOC20 $[+0.05, +0.5, +1.4]$ or PC59 $[+0.7, +0.7, +0.3]$ than on CS $[+0.8, +1.5, +3.3]$ . We attribute this to VOC20 and PC59 being closer to ImageNet which was leveraged in CLIP [47] to engineer the set of 80 templates used in all models. The benefit of our method + +
MethodCSVOC20PC59ADEStuffAvg
MaskCLIP + FLOSS51%54%56%57%56%55%
NACLIP + FLOSS49%20%57%57%52%47%
CLIP-DINOiser + FLOSS62%41%57%45%44%50%
+ +Table 2. Quality of our estimated class-experts. We measure the quality of our estimated experts by evaluating their intersection with the true experts, cf. Eq. (8). + +
Methodin domain CSout of domain
NightFogRainSnowBDDMAPAvg
MaskCLIP24.513.320.321.019.922.725.820.5
+ FLOSS25.813.921.722.321.322.926.821.5
NACLIP35.523.332.930.732.631.435.331.0
+ FLOSS37.023.632.531.033.632.336.531.6
CLIP-DINOiser31.312.724.127.927.231.335.226.4
+ FLOSS34.616.629.029.829.432.236.628.9
+ +Table 3. Generalization of experts across datasets. We evaluate the transferability of experts identified on the Cityscapes (CS) dataset to unseen datasets that share the same semantic classes and exhibit changes in lighting (Night), weather (Fog, Rain, Snow), or location and scenery (BDD, MAP). Results show a consistent improvement when using FLOSS, across almost all settings. + +is evident on challenging benchmarks like ADE and Stuff, which have a large number of semantic categories making accurate segmentation particularly hard. + +Quality of experts. We evaluate the quality $\hat{\rho}_k$ of our experts by measuring, for each class $k$ , the intersection of the estimated class-expert $\hat{\mathcal{E}}_k$ with the true set of experts $\mathcal{E}_k$ : + +$$ +\hat {\rho} _ {k} = 1 0 0 \times \frac {\left| \hat {\mathcal {E}} _ {k} \cap \mathcal {E} _ {k} \right|}{\left| \hat {\mathcal {E}} _ {k} \right|} = 1 0 0 \times \frac {\left| \hat {\mathcal {E}} _ {k} \cap \mathcal {E} _ {k} \right|}{N}. \tag {8} +$$ + +Where $|.|$ denotes the cardinality of a set. Quality averaged over classes is reported in Tab. 2 and shows that regardless of the model used, we correctly estimate approximately half of the true experts. In some scenarios, such as PASCAL VOC 20 with NACLIP, even when only $20\%$ of predicted experts are true experts, it is sufficient to bring some improvements ( $+0.5$ mIoU in Tab. 1). Per-class quality is reported in Appendix A.4, showing some variability. + +Generalization to unseen datasets. We now evaluate the generalization of experts to other unseen datasets sharing the same label set but exhibiting distribution shifts. In Tab. 3, we report performance of FLOSS when experts are identified on Cityscapes and used to evaluate performance on six unseen datasets which include four ACDC conditions (Night, Fog, Rain, Snow), BDD-100K, and MAPILLARY. Overall, we consistently improve the performance with notable gains of up to $+4.9$ mIoU for CLIP-DINOiser on ACDC Fog. This demonstrates that our class-wise experts are readily usable on other datasets sharing the same semantic classes, despite the presence of distribution shifts. + +![](images/a3551e18aa80179087cf1d94d21fdec9e5e993b0e2b3c7c1c3d0e3b297761cb0.jpg) +(a) Cityscapes +Figure 3. Effect of the number of images. We evaluate CLIP-DINOiser with FLOSS on urban scenes (a) and general object segmentation (b), when varying the number of images used to identify class-experts. Results show the average of 10 seeds, with standard deviation. Despite larger variation, we note the robustness of our method, even when having access to very few unlabeled images. + +![](images/955bcc5f50b34a848238d5a50be45dadf422d01cd017d2dcab2dfaca88143656.jpg) +(b) PASCAL VOC 20 + +Low-data regimes. To investigate the data efficiency of FLOSS, we study its performance in lower data regimes. Fig. 3 reports the mIoU of CLIP-DINOiser on Cityscapes and PASCAL VOC 20, as the number of images sampled from their respective training sets (containing 2,975 and 1,464 images, respectively) varies. The figure shows the average and standard deviation over 10 seeds (i.e., randomly sampling 10 different image sets). On both datasets, the general trend shows that our method benefits from accessing more images, which stems from the better estimation of class-experts. Remarkably, our method performs well in a very low-data regime, outperforming the default CLIP-DINOiser with only 1 image on Cityscapes and around 25 images on PASCAL VOC 20. We conjecture that the latter requires more images due to its lower per-image class density compared to Cityscapes, which exhibits higher variability and density per image. + +Low-data regimes & domain shift. To further corroborate its practicality, we evaluate FLOSS under stricter conditions combining both low-data regimes and domain shift. Specifically, we use only a single CS image to identify experts that are tested on 6 shifted domains. Tab. 4 provides evidence on the effectiveness of FLOSS in this strict setting, leading to an average improvement of $1.9\%$ across OOD datasets. + +Expert transferability. We investigate the cross-dataset transferability of class-experts by evaluating whether experts identified on one dataset can effectively improve performance on semantically related but distinct datasets. We employ COCO-Stuff (171 classes) to identify class-experts for CLIP-DINOiser and apply them to Cityscapes, PASCAL VOC 20, and PASCAL CONTEXT 59, which share 10, 16, and 37 common classes with COCO-Stuff, respectively. For non-overlapping classes, we use the default average classifier $W_{\mathrm{CLIP}}$ . As reported in Tab. 5, FLOSS shows consistent improvements, confirming expert transferability across datasets. + +
Methodin domain CSout of domain
NightFogRainSnowBDDMAPAvg
CLIP-DINOiser31.312.724.127.927.231.335.226.4
+ FLOSS33.216.228.229.328.731.735.628.3
+ +Table 4. FLOSS with 1 image & domain shift. Experts are identified using a single Cityscapes (CS) image and evaluated on unseen datasets that share the same semantic classes and exhibit changes in lighting (Night), weather (Fog, Rain, Snow), or location and scenery (BDD, MAP). We observe a consistent boost. + +
Methodin-domain Stuffout of domain
CS commonallVOC20 commonallPC59 commonall
CLIP-DINOiser24.634.531.180.580.839.336.0
+ FLOSS24.735.832.282.382.440.036.2
+ +# 5.2. Ablation Study + +Fusion strategies for expert predictions. We investigate alternative strategies to fuse the expert predictions (cf., Sec. 4.2), reporting performance on Cityscapes in Tab. 6. Simply averaging all soft segmentation maps ("Average-all") irrespective of the templates class of expertise leads to suboptimal results across all models. When accounting for templates' expertise in fusion, conflicts arise often as several experts may predict their own class of expertise. We therefore explore three strategies to resolve conflicts: "Default" which falls back to $W_{\mathrm{CLIP}}$ predictions, "Average" which averages the probability vectors of conflicting experts, and "Highest" which selects the most confident expert as described in Sec. 4.2. The latter consistently yields the best performance across all models, validating our fusion scheme. + +Metrics for expert identification. We replace entropy in Eq. (6) with other unsupervised metrics to act as proxies of the template class-wise performance, reporting performance on Cityscapes using the CLIP-DINOiser model in Fig. 4a. + +(i) Avg. Probability computes the average probability assigned to class $k$ across all pixels predicted as class $k$ . +(ii) $MaNo$ [70] is adapted as it was originally designed for assessing general model performance in image classification, to evaluate class-wise performance in semantic segmentation. +(iii) "ITI" [25] is the Inter-class separability to Intra-class similarity ratio, and captures both the separability of the classes in the feature space as well as the ability to keep samples of the same class close together. + +Table 5. Cross-dataset expert transferability. Class-experts identified on COCO-Stuff are applied to CS, VOC20, and PC59. Results are reported for common classes shared with COCO-Stuff and all classes. FLOSS demonstrates consistent improvements across all target datasets. + +
FusionConflictMaskCLIP +FLOSSNACLIP +FLOSSCLIP-DINOiser +FLOSS
Average-all24.736.631.2
ExpertDefault24.836.632.0
ExpertAverage25.636.734.4
ExpertHighest25.837.034.6
+ +Table 6. Strategies for the fusion of expert predictions. Exploring various fusion strategies on Cityscapes shows the benefit of our strategy, which outperforms all others across models. + +![](images/35f018e50f043a0a4c394456be68202bbc9065a31f331cacd0ad8061f09e7407.jpg) +(a) Expert metrics + +![](images/5cd0243995006990d162def40ea3ee48a4b6dee6952de3490583bee4abf28408.jpg) +(b) Effect of Top-N + +![](images/104618af780225ba9347ec74934024773ff5ddc0e7c15be94356703eec0c6d69.jpg) +Figure 4. Ablation of experts. We ablate our choices of metrics to identify experts as well as the number of experts per class. (a) shows unsupervised metrics to identify the experts, showing that our choice of entropy yields the best performance, although 'Avg. Probability' performs on par. (b) reports performance of FLOSS when varying the number of $N$ experts. Results are reported on the Cityscapes dataset using the CLIP-DINOiser model. +Figure 5. Effect of class-expert quality. We report the performance of FLOSS in the oracle-based setting, where each class-wise set of 4 experts contains a ratio $(\rho)$ of true experts. The experiment is conducted on Cityscapes using the CLIP-DINOiser model. + +These metrics are detailed in Appendix A.5. + +Overall, our entropy measure (cf. Eq. (5)) is the most effective, although Avg. Probability performs on par. + +Impact of class-expert identification quality. At the core of our method is the identification of class-experts. While we demonstrated there are templates which are class-expert, for each class there exist also under-performing templates (cf. Fig. 1) which, if selected, could harm the performance. + +To measure the impact of experts' quality on FLOSS, we devise an oracle-based experiment using $N = 4$ templates per-class, where we have access to ground truth la + +
MethodViT-B/16ViT-L/14
CSPC59Avg.CSPC59Avg.
SCLIP [63]32.234.233.221.325.223.3
GEM [3]32.635.634.127.128.127.6
NACLIP [23]35.535.235.431.432.131.8
+ FLOSS37.035.936.532.432.432.4
+ +Table 7. Effect of backbone. Results (mIoU) show that NACLIP + FLOSS outperforms NACLIP across ViT-B/16 and ViT-L/14 CLIP backbones. + +bels to identify the true class-experts. We randomly sample a fixed ratio $\rho$ of templates among true experts (i.e., $\mathcal{E}_k$ from Eq. (4)) and the remaining ones among non-experts (i.e., $\{\mathcal{T}_m\} \backslash \mathcal{E}_k$ ). For example, a ratio of $\rho = 75\%$ indicates that 3 templates are experts and 1 is non-expert. The outcome of this experiment on CS using the CLIP-DINOiser model is reported in Fig. 5 and shows an expected performance boost when the ratio of true experts increases. It also highlights that $50\%$ of true experts is sufficient for FLOSS to surpass the original CLIP-DINOiser. Interestingly, the performance of CLIP-DINOiser + FLOSS in Tab. 1 (34.6) slightly surpasses the performance at $\rho = 100\%$ shown in Fig. 5 (34.1). This however does not indicate that our estimated templates in Tab. 1 are all true experts - as reflected by Tab. 2 - but rather that our correctly estimated experts are on average better than random true experts. + +To further measure the upper bound of CLIP-DINOiser + FLOSS, we implement an oracle version using only the best true class-expert (i.e., $\{\mathcal{T}_{\tilde{m}} \mid \tilde{m} \in \mathrm{Top-N}(\Omega_k(W(\mathcal{T}_m)))\}$ ). On CS, the oracle with $N \in \{1, 2, 3, 4\}$ achieves respectively 37.1/37.0/37.3/37.5%, showing the potential for improvement assuming the identification of better experts. + +FLOSS using the validation set. First, we investigate the effectiveness of FLOSS in a transductive setting. Using the validation set of CS for expert selection, CLIP-DINOiser + FLOSS achieves $34.4\%$ mIoU on the same set, compared to $31.3\%$ for CLIP-DINOiser. Second, using a portion of the validation set to select experts, i.e., 5/25/50/75 val images, we achieve 32.8/33.0/33.0/33.3% on the rest of the validation images, compared to $30.9\%$ for CLIP-DINOiser alone. + +Impact of different backbones. Tab. 7 demonstrates the effectiveness of our proposed FLOSS when integrated with NACLIP using the ViT-L/14 backbone. Our approach achieves consistent improvements, reaching $32.4\%$ mIoU on both Cityscapes and PC59 datasets, leading to a 0.6 gain in average performance over NACLIP. These results validate the effectiveness of our method when using large-scale vision transformers. + +![](images/3be22c283219e3b72b7b87d86c2aa877d663ef3468c30965790df2fce340d449.jpg) +Figure 6. Performance of best-class experts. Performance gains (IoU) when using best class-experts versus standard template-averaging. Results are shown for CLIP-DINOiser, MaskCLIP and NACLIP. The bars represent the IoU difference between using best class-experts (selected with ground-truth labels) and the default template-averaging approach. + +# 6. Perspectives and Future Work + +Our experiments show that entropy is a strong estimator of class-experts. However, future research could develop even better proxies for prediction accuracy. Template-based methods appear especially promising, with significant potential for improving OVSS performance. As illustrated in Fig. 6, there is still a large gap to close—for example, selecting the best expert for the "sky" class could improve IoU by over 30 percentage points. These "upper-bounds" could be further increased by using more or augmented templates beyond the 80 currently used. + +# 7. Conclusion + +We propose a new task for improving OVSS models without having access to labels and without training, from a prompt template perspective. Our work is motivated by an intriguing observation: for each class, some templates excel in segmenting it. We refer to these templates as class-experts. Given a set of unlabeled images, our FLOSS uses the entropy of predictions as an unsupervised metric to identify these class-experts, thus not requiring any training or labels. Once selected, we propose a simple scheme for fusing the predictions of the class-experts, resulting in an improved overall OVSS performance in the inductive setting, even in the presence of distribution shifts. Additionally, we corroborate the effectiveness of FLOSS, which benefits from being plug-and-play, and further show that only few unlabeled images can be sufficient for expert selection. Furthermore, the oracle performance shows the potential of the class-experts, paving the way for future research. + +Acknowledgment. This research was partially funded by the French Agence Nationale de la Recherche (ANR) with the project SIGHT (ANR-20-CE23-0016) and supported by ELSA - European Lighthouse on Secure and Safe AI funded by the European Union under grant agreement No. 101070617. We sincerely thank Telecom Paris for providing the resources necessary to run our experiments and Nacereddine Laddaoui for his help with infrastructure. We are also grateful to Ivan Lopes for proofreading. + +# References + +[1] Stephen H Bach, Victor Sanh, Zheng-Xin Yong, Albert Webson, Colin Raffel, Nihal V Nayak, Abheesht Sharma, Taewoon Kim, M Saiful Bari, Thibault Fevry, et al. Prompt-source: An integrated development environment and repository for natural language prompts. In ACL, 2022. 3 +[2] Sule Bai, Yong Liu, Yifei Han, Haoji Zhang, and Yansong Tang. Self-calibrated clip for training-free open-vocabulary segmentation. arXiv preprint arXiv:2411.15869, 2024. 2 +[3] Walid Bousselham, Felix Petersen, Vittorio Ferrari, and Hilde Kuehne. Grounding everything: Emerging localization properties in vision-language transformers. In CVPR, 2024. 2, 5, 8 +[4] Maxime Bucher, Tuan-Hung Vu, Matthieu Cord, and Patrick Pérez. Zero-shot semantic segmentation. In NeurlPS, 2019. 2 +[5] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In CVPR, 2018. 2, 5 +[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 1 +[7] Junbum Cha, Jonghwan Mun, and Byungseok Roh. Learning to generate text-grounded mask for open-world semantic segmentation from only image-text pairs. In CVPR, 2023. 5 +[8] Guangyi Chen, Weiran Yao, Xiangchen Song, Xinyue Li, Yongming Rao, and Kun Zhang. Plot: Prompt learning with optimal transport for vision-language models. In ICLR, 2023. 2 +[9] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. TPAMI, 2017. 1 +[10] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 1 +[11] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, 2023. 5 +[12] Seokju Cho, Heeseong Shin, Sunghwan Hong, Anurag Arnab, Paul Hongsuck Seo, and Seungryong Kim. Cat-seg: Cost aggregation for open-vocabulary semantic segmentation. In CVPR, 2024. 1, 2 +[13] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 3, 5 +[14] Ulindu De Silva, Didula Samaraweera, Sasini Wanigathunga, Kavindu Kariyawasam, Kanchana Ranasinghe, Muzammal Naseer, and Ranga Rodrigo. Test-time optimization for domain adaptive open vocabulary segmentation. arXiv preprint arXiv:2501.04696, 2025. 2 + +[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 5 +[16] Reza Esfandiarpoor, Cristina Menghini, and Stephen H Bach. If clip could talk: Understanding vision-language model representations through their preferred concept descriptions. In EMNLP, 2024. 3 +[17] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results, 2012. 5 +[18] Yarin Gal. Uncertainty in deep learning. PhD Thesis, University of Cambridge, 2016. 4 +[19] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. IJCV, 2024. 2 +[20] Yunhao Ge, Jie Ren, Andrew Gallagher, Yuxiao Wang, Ming-Hsuan Yang, Hartwig Adam, Laurent Itti, Balaji Lakshminarayanan, and Jiaping Zhao. Improving zero-shot generalization and robustness of multi-modal models. In CVPR, 2023. 3 +[21] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In ECCV, 2022. 2, 3 +[22] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In ICLR, 2022. 3 +[23] Sina Hajimiri, Ismail Ben Ayed, and Jose Dolz. Pay attention to your neighbours: Training-free open-vocabulary semantic segmentation. In WACV, 2025. 1, 2, 3, 5, 8 +[24] Cong Han, Yujie Zhong, Dengjie Li, Kai Han, and Lin Ma. Open-vocabulary semantic segmentation with decoupled one-pass network. In ICCV, 2023. 2 +[25] Dapeng Hu, Jian Liang, Jun Hao Liew, Chuhui Xue, Song Bai, and Xinchao Wang. Mixed samples as probes for unsupervised model selection in domain adaptation. In NeurIPS, 2023. 7 +[26] Xuefeng Hu, Ke Zhang, Lu Xia, Albert Chen, Jiajia Luo, Yuyin Sun, Ken Wang, Nan Qiao, Xiao Zeng, Min Sun, et al. Reclip: Refine contrastive language image pre-training with source free domain adaptation. In WACV, 2024. 2 +[27] Haiwen Huang, Songyou Peng, Dan Zhang, and Andreas Geiger. Renovating names in open-vocabulary segmentation benchmarks. In NeurIPS, 2024. 3 +[28] Tony Huang, Jack Chu, and Fangyun Wei. Unsupervised prompt learning for vision-language models. arXiv preprint arXiv:2204.03649, 2022. 2 +[29] Yunshi Huang, Fereshteh Shakeri, Jose Dolz, Malik Boudiaf, Houda Bahig, and Ismail Ben Ayed. Lp++: A surprisingly strong linear probe for few-shot clip. In CVPR, 2024. 2 +[30] Eyke Hüllermeier and Willem Waegeman. *Aleatoric and epistemic uncertainty in machine learning: An introduction to concepts and methods.* ML, 2021. 4 +[31] Krishna Murthy Jatavallabhula, Alihusein Kuwajerwala, Qiao Gu, Mohd Omama, Tao Chen, Shuang Li, Ganesh Iyer, Soroush Saryazdi, Nikhil Keetha, Ayush Tewari, et al. ConceptFusion: Open-set multimodal 3d mapping. In RSS, 2023. + +[32] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? In NeurIPS, 2017. 4 +[33] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In CVPR, 2023. 2 +[34] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. In NeurIPS, 2017. 4 +[35] Mengcheng Lan, Chaofeng Chen, Yiping Ke, Xinjiang Wang, Litong Feng, and Wayne Zhang. Proxyclip: Proxy attention improves clip for open-vocabulary segmentation. In ECCV, 2024. 2 +[36] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. In ICLR, 2022. 2 +[37] Yi Li, Hualiang Wang, Yiqun Duan, Jiheng Zhang, and Xiaomeng Li. A closer look at the explainability of contrastive language-image pre-training. PR, 2025. 2, 5 +[38] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In CVPR, 2023. 1, 2 +[39] Yuqi Lin, Minghao Chen, Wenxiao Wang, Boxi Wu, Ke Li, Binbin Lin, Haifeng Liu, and Xiaofei He. Clip is also an efficient segmenter: A text-driven approach for weakly supervised semantic segmentation. In CVPR, 2023. 3 +[40] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. In ICLR, 2023. 3 +[41] Matthias Minderer, Josip Djolonga, Rob Romijnders, Frances Hubis, Xiaohua Zhai, Neil Houlsby, Dustin Tran, and Mario Lucic. Revisiting the calibration of modern neural networks. In NeurIPS, 2021. 2 +[42] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The role of context for object detection and semantic segmentation in the wild. In CVPR, 2014. 5 +[43] Jishnu Mukhoti, Tsung-Yu Lin, Omid Poursaeed, Rui Wang, Ashish Shah, Philip HS Torr, and Ser-Nam Lim. Open vocabulary semantic segmentation with patch aligned contrastive learning. In CVPR, 2023. 1 +[44] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In ICCV, 2017. 5 +[45] Sarah Parisot, Yongxin Yang, and Steven McDonagh. Learning to name classes for vision and language models. In CVPR, 2023. 3 +[46] Sarah Pratt, Ian Covert, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. In ICCV, 2023. 3 +[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3, 5 +[48] Kanchana Ranasinghe, Brandon McKinzie, Sachin Ravi, Yinfei Yang, Alexander Toshev, and Jonathon Shlens. Per + +ceptual grouping in contrastive vision-language models. In ICCV, 2023. 2 +[49] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In CVPR, 2022. 2 +[50] Karsten Roth, Jae Myung Kim, A Koepke, Oriol Vinyals, Cordelia Schmid, and Zeynep Akata. Waffling around for performance: Visual classification with random words and broad concepts. In ICCV, 2023. 3 +[51] Ohad Rubin, Jonathan Herzig, and Jonathan Berant. Learning to retrieve prompts for in-context learning. In *NAACL-HLT*, 2022. 3 +[52] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding. In CVPR, 2021. 5 +[53] Gyungin Shin, Weidi Xie, and Samuel Albanie. Reco: Retrieve and co-segment for zero-shot transfer. In NeurIPS, 2022. 5 +[54] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. In NeurIPS, 2022. 2, 3 +[55] Julio Silva-Rodriguez, Sina Hajimiri, Ismail Ben Ayed, and Jose Dolz. A closer look at the few-shot adaptation of large vision-language models. In CVPR, 2024. 2 +[56] Vladan Stojnic, Yannis Kalantidis, and Giorgos Tolias. Label propagation for zero-shot classification with vision-language models. In CVPR, 2024. 2 +[57] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In ICCV, 2021. 1 +[58] Weijie Tu, Weijian Deng, and Tom Gedeon. A closer look at the robustness of contrastive language-image pre-training (clip). In NeurIPS, 2023. 2 +[59] Weijie Tu, Weijian Deng, Dylan Campbell, Stephen Gould, and Tom Gedeon. An empirical study into what matters for calibrating vision-language models. In ICML, 2024. 2 +[60] Antonin Vobecky, Oriane Simeoni, David Hurych, Spyridon Gidaris, Andrei Bursuc, Patrick Pérez, and Josef Sivic. Pop-3d: Open-vocabulary 3d occupancy prediction from images. In NeurIPS, 2023. 3 +[61] Tuan-Hung Vu, Himalaya Jain, Maxime Bucher, Matthieu Cord, and Patrick Pérez. Advent: Adversarial entropy minimization for domain adaptation in semantic segmentation. In CVPR, 2019. 4 +[62] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In ICLR, 2021. 4 +[63] Feng Wang, Jieru Mei, and Alan Yuille. Sclip: Rethinking self-attention for dense vision-language inference. In ECCV, 2024. 1, 5, 8 +[64] Zhengbo Wang, Jian Liang, Lijun Sheng, Ran He, Zilei Wang, and Tieniu Tan. A hard-to-beat baseline for training-free CLIP-based adaptation. In ICLR, 2024. 2 +[65] Monika Wysoczańska, Michael Ramamonjisoa, Tomasz Trzeciński, and Oriane Simeoni. Clip-diy: Clip dense infer + +ence yields open-vocabulary semantic segmentation for-free. In WACV, 2024. 5 +[66] Monika Wysoczanska, Oriane Simeoni, Michael Ramamonjisoa, Andrei Bursuc, Tomasz Trzciński, and Patrick Pérez. Clip-dinoiser: Teaching clip a few dino tricks for open-vocabulary semantic segmentation. In ECCV, 2024. 1, 2, 3, 5 +[67] Yongqin Xian, Subhabrata Choudhury, Yang He, Bernt Schiele, and Zeynep Akata. Semantic projection network for zero-and few-label semantic segmentation. In CVPR, 2019. 2 +[68] Bin Xie, Jiale Cao, Jin Xie, Fahad Shahbaz Khan, and Yanwei Pang. Sed: A simple encoder-decoder for open-vocabulary semantic segmentation. In CVPR, 2024. 2, 3 +[69] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. In NeurIPS, 2021. 1 +[70] Renchunzi Xie, Ambroise Odonnat, Vasilii Feofanov, Weijian Deng, Jianfeng Zhang, and Bo An. Mano: Exploiting matrix norm for unsupervised accuracy estimation under distribution shifts. In NeurIPS, 2024. 4, 7 +[71] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. Groupvit: Semantic segmentation emerges from text supervision. In CVPR, 2022. 5 +[72] Jilan Xu, Junlin Hou, Yuejie Zhang, Rui Feng, Yi Wang, Yu Qiao, and Weidi Xie. Learning open-vocabulary semantic segmentation models from natural language supervision. In CVPR, 2023. 2 +[73] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In CVPR, 2023. 2 +[74] Mengde Xu, Zheng Zhang, Fangyun Wei, Han Hu, and Xi-ang Bai. Side adapter network for open-vocabulary semantic segmentation. In CVPR, 2023. 2 +[75] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darryll. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2020. 5 +[76] Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang-Chieh Chen. Convolutions die hard: Open-vocabulary segmentation with single frozen convolutional clip. In NeurIPS, 2023. 2 +[77] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. In ECCV, 2022. 2 +[78] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. IJCV, 2019. 5 +[79] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In ECCV, 2022. 1, 2, 3, 5 +[80] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In CVPR, 2022. 2 + +[81] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 2022. +[82] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. In ICCV, 2023. 2 +[83] Xiangyang Zhu, Renrui Zhang, Bowei He, Aojun Zhou, Dong Wang, Bin Zhao, and Peng Gao. Not all features matter: Enhancing few-shot clip with adaptive prior refinement. In ICCV, 2023. 2 \ No newline at end of file diff --git a/2025/FLOSS_ Free Lunch in Open-vocabulary Semantic Segmentation/images.zip b/2025/FLOSS_ Free Lunch in Open-vocabulary Semantic Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..96448ef5688134d950c91a7ada6e1e10200abde4 --- /dev/null +++ b/2025/FLOSS_ Free Lunch in Open-vocabulary Semantic Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3cc69858f76f8f7eb5379f7e5b9aea5da7b0e6e2b7c8fefd0bad712ec01e58da +size 385359 diff --git a/2025/FLOSS_ Free Lunch in Open-vocabulary Semantic Segmentation/layout.json b/2025/FLOSS_ Free Lunch in Open-vocabulary Semantic Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..11fa8ae6a0d63c6ff44a7d4cccf9b80dbc3b9429 --- /dev/null +++ b/2025/FLOSS_ Free Lunch in Open-vocabulary Semantic Segmentation/layout.json @@ -0,0 +1,10727 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 104, + 103, + 506, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 103, + 506, + 121 + ], + "spans": [ + { + "bbox": [ + 104, + 103, + 506, + 121 + ], + "type": "text", + "content": "FLOSS: Free Lunch in Open-vocabulary Semantic Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 141, + 561, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 141, + 561, + 171 + ], + "spans": [ + { + "bbox": [ + 62, + 141, + 561, + 171 + ], + "type": "text", + "content": "Yasser Benigim " + }, + { + "bbox": [ + 62, + 141, + 561, + 171 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 62, + 141, + 561, + 171 + ], + "type": "text", + "content": " Mohammad Fahes " + }, + { + "bbox": [ + 62, + 141, + 561, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 141, + 561, + 171 + ], + "type": "text", + "content": " Tuan-Hung " + }, + { + "bbox": [ + 62, + 141, + 561, + 171 + ], + "type": "inline_equation", + "content": "\\mathrm{Vu}^{1,2}" + }, + { + "bbox": [ + 62, + 141, + 561, + 171 + ], + "type": "text", + "content": " Andrei Bursuc " + }, + { + "bbox": [ + 62, + 141, + 561, + 171 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 62, + 141, + 561, + 171 + ], + "type": "text", + "content": " Raoul de Charette " + }, + { + "bbox": [ + 62, + 141, + 561, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 141, + 561, + 171 + ], + "type": "text", + "content": " Inria " + }, + { + "bbox": [ + 62, + 141, + 561, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 62, + 141, + 561, + 171 + ], + "type": "text", + "content": " Valeo.ai" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 198, + 200, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 198, + 200, + 211 + ], + "spans": [ + { + "bbox": [ + 151, + 198, + 200, + 211 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 224, + 296, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 296, + 522 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 296, + 522 + ], + "type": "text", + "content": "In this paper, we challenge the conventional practice in Open-Vocabulary Semantic Segmentation (OVSS) of using averaged class-wise text embeddings, which are typically obtained by encoding each class name with multiple templates (e.g., a photo of , a sketch of a ). We investigate the impact of templates for OVSS, and find that for each class, there exist single-template classifiers—which we refer to as class-experts—that significantly outperform the conventional averaged classifier. First, to identify these class-experts, we introduce a novel approach that estimates them without any labeled data or training. By leveraging the class-wise prediction entropy of single-template classifiers, we select those yielding the lowest entropy as the most reliable class-experts. Second, we combine the outputs of class-experts in a new fusion process. Our plug-and-play method, coined FLOSS, is orthogonal and complementary to existing OVSS methods, offering an improvement without the need for additional labels or training. Extensive experiments show that FLOSS consistently enhances state-of-the-art OVSS models, generalizes well across datasets with different distribution shifts, and delivers substantial improvements in low-data scenarios where only a few unlabeled images are available. Our code is available at https://github.com/yasserben/FLOSS." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 544, + 135, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 135, + 556 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 135, + 556 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 564, + 295, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 295, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 295, + 672 + ], + "type": "text", + "content": "In the past decade, advances in deep learning and the growing amount of training data have enabled the challenging task of semantic segmentation, which involves assigning semantic labels to each pixel in an image. Initially, this was limited to predefined categories [9, 10, 57, 69], but recent models rely on vision-language alignment [47] for open-vocabulary segmentation, where the semantic categories can be dynamically expressed at runtime in natural language." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 672, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 295, + 696 + ], + "type": "text", + "content": "In Open-Vocabulary Semantic Segmentation (OVSS), the CLIP model [47] is commonly used as the backbone," + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 314, + 198, + 555, + 323 + ], + "blocks": [ + { + "bbox": [ + 314, + 198, + 555, + 323 + ], + "lines": [ + { + "bbox": [ + 314, + 198, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 314, + 198, + 555, + 323 + ], + "type": "image", + "image_path": "3a640d3901aa085653a9588d6e8c542193f0e637180093783acb282b862cd372.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 331, + 555, + 399 + ], + "lines": [ + { + "bbox": [ + 313, + 331, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 555, + 399 + ], + "type": "text", + "content": "Figure 1. Using individual vs average template classifiers. Empirically, we observe that for each class there exist individual templates (colored dots) that lead to classifiers performing better than the popular CLIP classifier, which is built from the averaging of all 80 templates embeddings " + }, + { + "bbox": [ + 313, + 331, + 555, + 399 + ], + "type": "inline_equation", + "content": "(-)" + }, + { + "bbox": [ + 313, + 331, + 555, + 399 + ], + "type": "text", + "content": ". The analysis was conducted on Cityscapes using the CLIP-DINOiser model." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 419, + 555, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 555, + 695 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 555, + 695 + ], + "type": "text", + "content": "which is either retrained [43], fine-tuned [12, 38], or frozen and adjusted to extend global image-language alignment to denser pixel-language alignment. In this work, we study CLIP's original classification paradigm and focus on methods that keep all CLIP parameters frozen [23, 63, 79]. Among these, MaskCLIP [79] leverages CLIP's embedding layer for segmentation, CLIP-DINOiser [66] enhances features via distillation from an external DINO [6] network, while NACLIP [23] modifies CLIP's self-attention to improve spatial consistency. A common aspect of prior OVSS methods is their reliance on CLIP's default classification process, where image or patch embeddings are compared with text embeddings representing semantic categories. To represent a class, a common practice is to use multiple templates to encode class names (e.g., \"a photo of a car\", \"a sketch of a car\", etc) and average their embeddings. Empirically, Radford et al. [47] engineered " + }, + { + "bbox": [ + 313, + 419, + 555, + 695 + ], + "type": "inline_equation", + "content": "M = 80" + }, + { + "bbox": [ + 313, + 419, + 555, + 695 + ], + "type": "text", + "content": " templates to improve ImageNet zero-shot classification. For segmentation, OVSS methods [23, 66, 79] adopt by default the same templates to construct segmentation classifiers. While this \"template-averaging\" strategy works well overall, it may not always provide the best classifier for each semantic concept." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 325, + 702, + 537, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 702, + 537, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 702, + 537, + 714 + ], + "type": "text", + "content": "1\"Template-averaging\" is an imprecise but convenient shorthand." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 66, + 702, + 201, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 201, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 201, + 713 + ], + "type": "text", + "content": "*Work done during an internship at Inria." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21471" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "Our paper explores the impact of individual templates in OVSS – an aspect that has been overlooked in the literature. Fig. 1 shows an intriguing empirical observation: for each class, some individual templates outperform the averaged classifier that uses all templates. We call such templates class-experts, highlighting that they are trivial to identify using semantic ground truths." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 157, + 295, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 157, + 295, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 157, + 295, + 193 + ], + "type": "text", + "content": "However, without labels, two challenges arise: 1) How to identify class-experts without ground truth? 2) How to fuse predictions from class-experts?" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 194, + 295, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 295, + 312 + ], + "type": "text", + "content": "We propose that entropy, an unsupervised metric, can help identify expert templates without labels. This leads to a new task—training- and label-free template selection for OVSS—which improves segmentation by selecting better text template subsets for a given dataset of unlabeled images. Our method shifts the attention to the text modality, unlike previous works that tweak the visual encoder, and also aims to improve the performance in an inductive setting, where the selected templates are applied to a different subset of images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 313, + 235, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 313, + 235, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 235, + 324 + ], + "type": "text", + "content": "We summarize our contributions as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 326, + 295, + 480 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 326, + 295, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 326, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 295, + 373 + ], + "type": "text", + "content": "- We analyze the current template-averaging practice for OVSS, showing that better class-wise classifiers can be constructed using only a subset of all templates, and that these expert templates can be identified without labels." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 374, + 295, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 374, + 295, + 432 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 295, + 432 + ], + "type": "text", + "content": "- We propose the novel task of training- and label-free template selection for OVSS. Given an unlabeled set of images, this task aims to identify class-wise experts that improve OVSS performance on unseen counterparts of the dataset." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 434, + 295, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 295, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 295, + 480 + ], + "type": "text", + "content": "- We introduce FLOSS, which leverages class-wise per-pixel entropy to select expert templates and employs a simple and effective fusion scheme to combine individual expert predictions." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 482, + 295, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 482, + 295, + 554 + ], + "spans": [ + { + "bbox": [ + 55, + 482, + 295, + 554 + ], + "type": "text", + "content": "With extensive experiments, we demonstrate that FLOSS consistently improves the state-of-the-art performance on OVSS benchmarks and exhibits generalization capabilities in the choice of experts. Furthermore, with only a few unlabeled images, our approach can select reasonably good expert templates, resulting in performance boost." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 566, + 146, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 566, + 146, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 566, + 146, + 578 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "Vision-Language Model (VLM) transferability. Since CLIP's release, its impressive zero-shot classification performance [47] and robustness [41, 58, 59] have made it a foundation for various adaptation and transfer learning tasks. Few-shot supervised fine-tuning, where only limited labeled data is available for downstream tasks like image classification, has gained significant attention. The goal is to leverage the vast knowledge of the VLM acquired through training on an immense image collection by steering it towards a task of interest via few labeled sam" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": "plies. Parameter efficiency is central to this setting. To ensure it, prompt learning [8, 33, 80-82], lightweight MLP training [19], and linear probing [29, 55] have been proposed. Another line of research explores training-free few-shot adaptation [64, 77, 83], where a classifier is built using few labeled samples and ensembled with the zero-shot classifier. Additionally, unsupervised adaptation has been explored [26, 28, 56], alongside test-time adaptation, where prompts are learned per test image [54]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 182, + 553, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 553, + 255 + ], + "type": "text", + "content": "While using VLMs for classification tasks aligns with their pre-training focus on global image recognition, applying them to dense prediction tasks like semantic segmentation is more challenging due to the gap between global image understanding and pixel-level prediction. We summarize related efforts in the following." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 268, + 553, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 268, + 553, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 268, + 553, + 663 + ], + "type": "text", + "content": "Open-Vocabulary Semantic Segmentation (OVSS). OVSS uses VLMs like CLIP for text-based segmentation. However, CLIP's global pooling layers limit its ability to generate dense pixel-level, language-aligned features [31, 79]. To address this, several training-free methods densify CLIP without altering its parameters, preserving its image-language alignment. MaskCLIP [79] was the first to replace the attention pooling layer with a convolutional layer. Later works further densify CLIP by aggregating multiple image views [31, 66], integrating priors from other vision foundation models for better pooling [35, 66], extracting different information from the self-attention layers [3, 23, 37], or removing anomaly tokens [2]. Seg-TTO [14] leverages an LLM to generate discriminative attributes for rare classes in specialized domains (e.g., medical sciences, earth monitoring) that use technical terminology (e.g., mediastinum), operating in a test-time optimization framework while maintaining zero-shot capabilities. Other approaches fine-tune CLIP for segmentation tasks using weak supervision through image-level tags, captions [21, 48], or class-agnostic object masks [21, 49]. Others employ full supervision [12, 36, 38, 74] on densely annotated datasets like COCO Stuff [5], following strategies similar to early inductive zero-shot semantic segmentation methods [4, 67]. Among them, some fine-tune CLIP's image encoder [36, 38, 74], others combine this with additional backbones and segmentation heads [38], while some learn side networks while keeping CLIP frozen [74]. Many other approaches have been proposed in this active research area [24, 68, 72, 73, 76]. While effective, these methods require labeled data and computational resources, limiting their practicality in resource-constrained scenarios." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": "Our work focuses on training-free approaches, evaluating FLOSS on three key methods: MaskCLIP [79], CLIP-DINOiser [66] and NACLIP [23], while being applicable to any text-based segmentation framework." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21472" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 323 + ], + "type": "text", + "content": "Prompting in VLMs. CLIP's image-language interface has inspired the adoption of various prompt engineering practices from LLMs to better reveal the knowledge encapsulated in the vision encoder and boost performance. The seminal CLIP work [47] introduced dataset-specific templates (e.g., \"a photo of \") to bridge the gap between training captions and test prompts for zero-shot classification. Subsequent methods have improved prompts by crowdsourcing templates [1], adjusting class names [21, 60], or modifying templates for tasks like object detection [21, 22] and occupancy prediction [60]. While effective, manual prompt tuning is labor-intensive, leading to automated strategies like adding random words or characters to the original template [50], enriching class names with WordNet concepts [20] or synonyms [39], learning more informative class names [27], and using LLMs to generate more expressive descriptions starting from simple class names or retrieve them from a dataset [51] and further use them as prompts. Some approaches also learn prompts for task-specific fitting [45] or test-time adaptation [54], and others mine them with external LLMs [16]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 327, + 296, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 399 + ], + "type": "text", + "content": "Our approach differs by selecting expert templates for each class in an unsupervised manner, starting from predefined templates. While previous methods have focused on image classification, we apply this strategy to semantic segmentation using CLIP's original ImageNet templates, which have been used by default in prior works." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 420, + 141, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 420, + 141, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 420, + 141, + 433 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 445, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 296, + 594 + ], + "type": "text", + "content": "OVSS aims to produce a soft-segmentation map " + }, + { + "bbox": [ + 55, + 445, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{S}}\\in [0,1]^{H\\times W\\times K}" + }, + { + "bbox": [ + 55, + 445, + 296, + 594 + ], + "type": "text", + "content": " given an image " + }, + { + "bbox": [ + 55, + 445, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{l}\\in \\mathbb{R}^{H\\times W\\times 3}" + }, + { + "bbox": [ + 55, + 445, + 296, + 594 + ], + "type": "text", + "content": " and a set of classes " + }, + { + "bbox": [ + 55, + 445, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{C}_k\\}_{k\\in [1,K]}" + }, + { + "bbox": [ + 55, + 445, + 296, + 594 + ], + "type": "text", + "content": " expressed in natural language. The final segmentation map " + }, + { + "bbox": [ + 55, + 445, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{P}}" + }, + { + "bbox": [ + 55, + 445, + 296, + 594 + ], + "type": "text", + "content": " is derived by applying arg max(·) on " + }, + { + "bbox": [ + 55, + 445, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{S}}" + }, + { + "bbox": [ + 55, + 445, + 296, + 594 + ], + "type": "text", + "content": " along the " + }, + { + "bbox": [ + 55, + 445, + 296, + 594 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 445, + 296, + 594 + ], + "type": "text", + "content": " -dimension. Leveraging the aligned vision encoder " + }, + { + "bbox": [ + 55, + 445, + 296, + 594 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 55, + 445, + 296, + 594 + ], + "type": "text", + "content": " and text encoder " + }, + { + "bbox": [ + 55, + 445, + 296, + 594 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 55, + 445, + 296, + 594 + ], + "type": "text", + "content": " of CLIP, the standard practice consists in classifying an image patch by comparing its vision embedding to the text embedding of class-specific prompts constructed using predefined templates " + }, + { + "bbox": [ + 55, + 445, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{T}_m\\}_{m\\in [1,M]}" + }, + { + "bbox": [ + 55, + 445, + 296, + 594 + ], + "type": "text", + "content": ". An example of prompt is \"a bright photo of a car\", denoted" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{m,k}" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " for compactness. Instead of a single template, practitioners use \"template-averaging\" to increase robustness and improve overall performance by averaging the embeddings of prompts obtained using all available templates [40, 46, 47, 50], which is a form of ensembling on the text encoder side. This writes " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\bar{c}_k = \\frac{1}{M}\\sum_{m = 1}^{M}g(\\mathcal{P}_{m,k})" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\bar{c}_k" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " is the resulting representation of class " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "C_k" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": ". Subsequently, for each patch, the predicted class is the one that maximizes the cosine similarity between the visual and the" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 72, + 386, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 386, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 386, + 84 + ], + "type": "text", + "content": "text embeddings:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 384, + 87, + 555, + 106 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 87, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 384, + 87, + 555, + 106 + ], + "type": "interline_equation", + "content": "\\hat {k} = \\arg \\max _ {k} f (\\mathcal {V}) ^ {\\mathrm {T}} \\bar {c} _ {k}, \\tag {1}", + "image_path": "3f59776a2081bacf1b7b7d487ffc5020f30ec48c60bc2d0ebecd18fcbb0301bc.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 110, + 555, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 110, + 555, + 145 + ], + "spans": [ + { + "bbox": [ + 313, + 110, + 555, + 145 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 110, + 555, + 145 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 110, + 555, + 145 + ], + "type": "text", + "content": " represents the visual patch. All embeddings are " + }, + { + "bbox": [ + 313, + 110, + 555, + 145 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 313, + 110, + 555, + 145 + ], + "type": "text", + "content": "-normalized. The text embeddings of " + }, + { + "bbox": [ + 313, + 110, + 555, + 145 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 110, + 555, + 145 + ], + "type": "text", + "content": " classes define a classifier that we denote:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 363, + 151, + 555, + 165 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 151, + 555, + 165 + ], + "spans": [ + { + "bbox": [ + 363, + 151, + 555, + 165 + ], + "type": "interline_equation", + "content": "\\boldsymbol {W} \\left(\\left\\{\\mathcal {T} _ {m} \\right\\} _ {m \\in [ 1, M ]}\\right) = \\left\\{\\bar {c} _ {k} \\right\\} _ {k \\in [ 1, K ]}. \\tag {2}", + "image_path": "48461d5678e389044dbbbbe7293da3e544392875a9a9b91aa730c04074c89cba.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 168, + 555, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 555, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 555, + 204 + ], + "type": "text", + "content": "As in the above equation, the classifier can be represented as a function of the templates used to compute the average embeddings of all classes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 208, + 555, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 208, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 208, + 555, + 338 + ], + "type": "text", + "content": "Empirical observations on single templates. OVSS models [23, 66, 68, 79] typically rely on the " + }, + { + "bbox": [ + 313, + 208, + 555, + 338 + ], + "type": "inline_equation", + "content": "M = 80" + }, + { + "bbox": [ + 313, + 208, + 555, + 338 + ], + "type": "text", + "content": " ImageNet templates of CLIP [47] to compute the average classifier of Eq. (2), which we refer to as " + }, + { + "bbox": [ + 313, + 208, + 555, + 338 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 313, + 208, + 555, + 338 + ], + "type": "text", + "content": " for brevity. While using " + }, + { + "bbox": [ + 313, + 208, + 555, + 338 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 313, + 208, + 555, + 338 + ], + "type": "text", + "content": " was shown to be robust [47, 79], the templates were originally hand-crafted for zero-shot classification on ImageNet [47], and their individual effect on class-wise performance remains unexplored for OVSS. Rather than " + }, + { + "bbox": [ + 313, + 208, + 555, + 338 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 313, + 208, + 555, + 338 + ], + "type": "text", + "content": " that uses all " + }, + { + "bbox": [ + 313, + 208, + 555, + 338 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 208, + 555, + 338 + ], + "type": "text", + "content": " templates, one can construct " + }, + { + "bbox": [ + 313, + 208, + 555, + 338 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 208, + 555, + 338 + ], + "type": "text", + "content": " distinct classifiers, each utilizing a single template " + }, + { + "bbox": [ + 313, + 208, + 555, + 338 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_m" + }, + { + "bbox": [ + 313, + 208, + 555, + 338 + ], + "type": "text", + "content": ". This writes:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 372, + 344, + 553, + 358 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 344, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 372, + 344, + 553, + 358 + ], + "type": "interline_equation", + "content": "\\boldsymbol {W} \\left(\\mathcal {T} _ {m}\\right) = \\left\\{g \\left(\\mathcal {P} _ {m, k}\\right) \\right\\} _ {k \\in [ 1, K ]}. \\tag {3}", + "image_path": "751e8fb8d5f75b142304f414c8cb52707933249895c56cb5e62d6a67f3d16e9d.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 361, + 554, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 554, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 554, + 456 + ], + "type": "text", + "content": "We design a systematic experiment to evaluate the performance of single-template classifiers " + }, + { + "bbox": [ + 313, + 361, + 554, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{W}(\\mathcal{T}_m)" + }, + { + "bbox": [ + 313, + 361, + 554, + 456 + ], + "type": "text", + "content": " from Eq. (3). Fig. 1 shows the per-class performance on Cityscapes [13] of each of the 80 single-template classifiers (plotted as colored dots, e.g., ●), where colors encode the template identifier (i.e., m), along with the average " + }, + { + "bbox": [ + 313, + 361, + 554, + 456 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 313, + 361, + 554, + 456 + ], + "type": "text", + "content": " classifier (plotted as —). This leads to two key observations:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 457, + 555, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 457, + 555, + 505 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 555, + 505 + ], + "type": "text", + "content": "a) For each class, there exist single-template classifiers outperforming " + }, + { + "bbox": [ + 313, + 457, + 555, + 505 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 313, + 457, + 555, + 505 + ], + "type": "text", + "content": " on this class; we refer to the corresponding templates as class-experts. More precisely, the set of class-experts is defined as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 321, + 510, + 553, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 510, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 321, + 510, + 553, + 533 + ], + "type": "interline_equation", + "content": "\\mathcal {E} _ {k} = \\left\\{\\mathcal {T} _ {m} \\mid \\Omega_ {k} \\left(\\boldsymbol {W} \\left(\\mathcal {T} _ {m}\\right)\\right) > \\Omega_ {k} \\left(\\boldsymbol {W} _ {\\text {C L I P}}\\right), m \\in [ 1, M ] \\right\\}, \\tag {4}", + "image_path": "4070b4aaae1b7b1863cd972df8a481ffc02bd4ebe0c5fbe29e94ed99ff8ee713.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 534, + 531, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 531, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 531, + 545 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 534, + 531, + 545 + ], + "type": "inline_equation", + "content": "\\Omega_{k}(\\cdot)" + }, + { + "bbox": [ + 313, + 534, + 531, + 545 + ], + "type": "text", + "content": " denotes the IoU performance on class " + }, + { + "bbox": [ + 313, + 534, + 531, + 545 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 534, + 531, + 545 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 546, + 554, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 554, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 554, + 582 + ], + "type": "text", + "content": "b) A single-template classifier excelling on one class may exhibit suboptimal performance on others, implying that for " + }, + { + "bbox": [ + 313, + 546, + 554, + 582 + ], + "type": "inline_equation", + "content": "k, l \\in [1, K]" + }, + { + "bbox": [ + 313, + 546, + 554, + 582 + ], + "type": "text", + "content": " the expert sets " + }, + { + "bbox": [ + 313, + 546, + 554, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_k" + }, + { + "bbox": [ + 313, + 546, + 554, + 582 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 546, + 554, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_l" + }, + { + "bbox": [ + 313, + 546, + 554, + 582 + ], + "type": "text", + "content": " may differ." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 582, + 554, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 640 + ], + "type": "text", + "content": "Following the above observations, class-experts are readily identified when ground truth labels are accessible, through direct performance comparison. We refer readers to Appendix A.3 for similar observations across datasets and models." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": "However, we are interested in the case where an unlabeled dataset is provided. Therefore, two questions stem from the previous empirical observations: (i) How to identify class-experts in an unsupervised, training-free manner? (ii) Given these identified class-experts, how to derive the final semantic predictions?" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21473" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 85, + 299, + 169 + ], + "blocks": [ + { + "bbox": [ + 107, + 72, + 249, + 83 + ], + "lines": [ + { + "bbox": [ + 107, + 72, + 249, + 83 + ], + "spans": [ + { + "bbox": [ + 107, + 72, + 249, + 83 + ], + "type": "text", + "content": "Unsupervised identification of class-experts" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 85, + 299, + 169 + ], + "lines": [ + { + "bbox": [ + 59, + 85, + 299, + 169 + ], + "spans": [ + { + "bbox": [ + 59, + 85, + 299, + 169 + ], + "type": "image", + "image_path": "8894a9fd171e353ebbc1bfe019331e61c5666bae754d5c2136c107ed1ca1a873.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 310, + 87, + 553, + 167 + ], + "blocks": [ + { + "bbox": [ + 408, + 73, + 455, + 83 + ], + "lines": [ + { + "bbox": [ + 408, + 73, + 455, + 83 + ], + "spans": [ + { + "bbox": [ + 408, + 73, + 455, + 83 + ], + "type": "text", + "content": "Expert Fusion" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 87, + 553, + 167 + ], + "lines": [ + { + "bbox": [ + 310, + 87, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 310, + 87, + 553, + 167 + ], + "type": "image", + "image_path": "8e8322105906b9164149c821e09e9c34d9f4406143d13ba5d567163264ce7c59.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 191, + 555, + 258 + ], + "lines": [ + { + "bbox": [ + 54, + 191, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 54, + 191, + 555, + 258 + ], + "type": "text", + "content": "Figure 2. Overview of FLOSS. Left: for each class, selected expert templates " + }, + { + "bbox": [ + 54, + 191, + 555, + 258 + ], + "type": "inline_equation", + "content": "(\\checkmark)" + }, + { + "bbox": [ + 54, + 191, + 555, + 258 + ], + "type": "text", + "content": " construct the estimated set " + }, + { + "bbox": [ + 54, + 191, + 555, + 258 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{E}}" + }, + { + "bbox": [ + 54, + 191, + 555, + 258 + ], + "type": "text", + "content": ", which classifies the input image and yield the soft-segmentation map " + }, + { + "bbox": [ + 54, + 191, + 555, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 54, + 191, + 555, + 258 + ], + "type": "text", + "content": ". In this example we set " + }, + { + "bbox": [ + 54, + 191, + 555, + 258 + ], + "type": "inline_equation", + "content": "N = 3" + }, + { + "bbox": [ + 54, + 191, + 555, + 258 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 54, + 191, + 555, + 258 + ], + "type": "inline_equation", + "content": "\\mathrm{Top - N}(\\dots)" + }, + { + "bbox": [ + 54, + 191, + 555, + 258 + ], + "type": "text", + "content": " template indices corresponding to lowest entropy entries to select. Right: we explain the fusion strategy for two cases. For the pixel " + }, + { + "bbox": [ + 54, + 191, + 555, + 258 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 54, + 191, + 555, + 258 + ], + "type": "text", + "content": ", the set of experts that predict their own class of expertise " + }, + { + "bbox": [ + 54, + 191, + 555, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{X}(x_1)" + }, + { + "bbox": [ + 54, + 191, + 555, + 258 + ], + "type": "text", + "content": " contains 2 experts; consequently, the decision is taken by looking at the maximum softmax scores of those two, leading to the decision of classifying " + }, + { + "bbox": [ + 54, + 191, + 555, + 258 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 54, + 191, + 555, + 258 + ], + "type": "text", + "content": " as \"person\" with the higher score. For the pixel " + }, + { + "bbox": [ + 54, + 191, + 555, + 258 + ], + "type": "inline_equation", + "content": "x_{2}" + }, + { + "bbox": [ + 54, + 191, + 555, + 258 + ], + "type": "text", + "content": ", there is no expert predicting its own class of expertise, i.e., " + }, + { + "bbox": [ + 54, + 191, + 555, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{X}(x_2) = \\emptyset" + }, + { + "bbox": [ + 54, + 191, + 555, + 258 + ], + "type": "text", + "content": ", and the decision is taken by the default classifier " + }, + { + "bbox": [ + 54, + 191, + 555, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 54, + 191, + 555, + 258 + ], + "type": "text", + "content": "; the pixel " + }, + { + "bbox": [ + 54, + 191, + 555, + 258 + ], + "type": "inline_equation", + "content": "x_{2}" + }, + { + "bbox": [ + 54, + 191, + 555, + 258 + ], + "type": "text", + "content": " thus takes the scores from " + }, + { + "bbox": [ + 54, + 191, + 555, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{\\mathrm{CLIP}}(x_2)" + }, + { + "bbox": [ + 54, + 191, + 555, + 258 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 277, + 111, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 111, + 290 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 111, + 290 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 299, + 296, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 299, + 296, + 406 + ], + "spans": [ + { + "bbox": [ + 55, + 299, + 296, + 406 + ], + "type": "text", + "content": "Given access to a dataset " + }, + { + "bbox": [ + 55, + 299, + 296, + 406 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 299, + 296, + 406 + ], + "type": "text", + "content": " of unlabeled images, our goal is to improve the performance of OVSS using the aforementioned class-experts, without training or access to any labels. To do so, for each class " + }, + { + "bbox": [ + 55, + 299, + 296, + 406 + ], + "type": "inline_equation", + "content": "k\\in [1,K]" + }, + { + "bbox": [ + 55, + 299, + 296, + 406 + ], + "type": "text", + "content": " our method first identifies a small set of class-experts among the " + }, + { + "bbox": [ + 55, + 299, + 296, + 406 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 299, + 296, + 406 + ], + "type": "text", + "content": " predefined ImageNet templates of CLIP (Sec. 4.1) by leveraging unsupervised metrics. It then builds on a simple scheme to fuse class-experts' predictions into a single OVSS semantic output (Sec. 4.2)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 415, + 280, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 280, + 428 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 280, + 428 + ], + "type": "text", + "content": "4.1. Unsupervised identification of class-experts" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 433, + 295, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 295, + 612 + ], + "type": "text", + "content": "Referring to our definition of class-experts in Sec. 3, for each of the " + }, + { + "bbox": [ + 55, + 433, + 295, + 612 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 433, + 295, + 612 + ], + "type": "text", + "content": " classes, our goal is to estimate a set of class-experts " + }, + { + "bbox": [ + 55, + 433, + 295, + 612 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{E}}_k" + }, + { + "bbox": [ + 55, + 433, + 295, + 612 + ], + "type": "text", + "content": ", among the " + }, + { + "bbox": [ + 55, + 433, + 295, + 612 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 433, + 295, + 612 + ], + "type": "text", + "content": " pre-defined CLIP templates. Being in an unsupervised setting, we have no access to the true classifier performance (i.e., " + }, + { + "bbox": [ + 55, + 433, + 295, + 612 + ], + "type": "inline_equation", + "content": "\\Omega(\\cdot)" + }, + { + "bbox": [ + 55, + 433, + 295, + 612 + ], + "type": "text", + "content": " in Eq. (4)) and therefore propose relying on unsupervised metrics which act as performance proxy [62, 70]. We leverage entropy as it is an established measure of uncertainty [18, 30], commonly adopted in deep learning literature [32, 34, 61]. The entropy value of a softmax prediction indicates a classifier's confidence, which empirically provides a reliable indicator of prediction quality. In our setting, entropy measures the confidence of single-template classifiers for each class, helping identify class-experts. We also study other unsupervised metrics in the ablations (Sec. 5.2)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 613, + 295, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 613, + 295, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 613, + 295, + 673 + ], + "type": "text", + "content": "In detail, we estimate class-experts as the set of individual templates whose associated classifiers (cf., Eq. (3)) exhibit low entropy. Specifically, for each template " + }, + { + "bbox": [ + 55, + 613, + 295, + 673 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 613, + 295, + 673 + ], + "type": "text", + "content": " and class " + }, + { + "bbox": [ + 55, + 613, + 295, + 673 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 613, + 295, + 673 + ], + "type": "text", + "content": ", we compute the average entropy " + }, + { + "bbox": [ + 55, + 613, + 295, + 673 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{m,k}" + }, + { + "bbox": [ + 55, + 613, + 295, + 673 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 55, + 613, + 295, + 673 + ], + "type": "inline_equation", + "content": "C_{m,k}" + }, + { + "bbox": [ + 55, + 613, + 295, + 673 + ], + "type": "text", + "content": " pixels predicted as class " + }, + { + "bbox": [ + 55, + 613, + 295, + 673 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 613, + 295, + 673 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 55, + 613, + 295, + 673 + ], + "type": "inline_equation", + "content": "\\pmb{W}(\\mathcal{T}_m)" + }, + { + "bbox": [ + 55, + 613, + 295, + 673 + ], + "type": "text", + "content": " in all images of " + }, + { + "bbox": [ + 55, + 613, + 295, + 673 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 613, + 295, + 673 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 683, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 683, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 104, + 683, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {H} _ {m, k} = - \\frac {1}{C _ {m , k}} \\sum_ {i = 1} ^ {C _ {m, k}} \\mathbf {q} _ {i} ^ {\\mathrm {T}} \\log (\\mathbf {q} _ {i}), \\tag {5}", + "image_path": "c393c3705def316b6131fc78b664a52a9ac8eb2df70a20c372c0e44ae1967ce2.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 278, + 555, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 278, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 555, + 340 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 278, + 555, + 340 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_i\\in \\Delta^{K - 1}" + }, + { + "bbox": [ + 313, + 278, + 555, + 340 + ], + "type": "text", + "content": " is pixel-wise probability prediction obtained by applying softmax(·) over " + }, + { + "bbox": [ + 313, + 278, + 555, + 340 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 278, + 555, + 340 + ], + "type": "text", + "content": " class-wise cosine similarities, and " + }, + { + "bbox": [ + 313, + 278, + 555, + 340 + ], + "type": "inline_equation", + "content": "\\Delta^{K - 1}" + }, + { + "bbox": [ + 313, + 278, + 555, + 340 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 313, + 278, + 555, + 340 + ], + "type": "inline_equation", + "content": "(K - 1)" + }, + { + "bbox": [ + 313, + 278, + 555, + 340 + ], + "type": "text", + "content": "-simplex in " + }, + { + "bbox": [ + 313, + 278, + 555, + 340 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^K" + }, + { + "bbox": [ + 313, + 278, + 555, + 340 + ], + "type": "text", + "content": ". Subsequently, for class " + }, + { + "bbox": [ + 313, + 278, + 555, + 340 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 278, + 555, + 340 + ], + "type": "text", + "content": ", the estimated set of experts is obtained as the " + }, + { + "bbox": [ + 313, + 278, + 555, + 340 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 278, + 555, + 340 + ], + "type": "text", + "content": " templates yielding the lowest entropy:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 345, + 348, + 553, + 367 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 348, + 553, + 367 + ], + "spans": [ + { + "bbox": [ + 345, + 348, + 553, + 367 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {E}} _ {k} = \\left\\{\\mathcal {T} _ {\\hat {m}} \\mid \\hat {m} \\in \\operatorname {T o p - N} (\\underset {m} {\\arg \\operatorname {s o r t}} \\mathcal {H} _ {m, k}) \\right\\}, \\tag {6}", + "image_path": "214c7298009845e6feaa1a33d2a5462bf0cd19c13084cb7eb81a19193efdd171.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 377, + 555, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 555, + 500 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 555, + 500 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 377, + 555, + 500 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{E}}_k" + }, + { + "bbox": [ + 313, + 377, + 555, + 500 + ], + "type": "text", + "content": " is expected to be an estimation of " + }, + { + "bbox": [ + 313, + 377, + 555, + 500 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_k" + }, + { + "bbox": [ + 313, + 377, + 555, + 500 + ], + "type": "text", + "content": " (Eq. (4)); the arg sort operator returns the sorted list of " + }, + { + "bbox": [ + 313, + 377, + 555, + 500 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 377, + 555, + 500 + ], + "type": "text", + "content": " template indices so that " + }, + { + "bbox": [ + 313, + 377, + 555, + 500 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{m,k}" + }, + { + "bbox": [ + 313, + 377, + 555, + 500 + ], + "type": "text", + "content": " are in ascending order. Finally, " + }, + { + "bbox": [ + 313, + 377, + 555, + 500 + ], + "type": "inline_equation", + "content": "\\mathrm{Top - N}(\\cdot)" + }, + { + "bbox": [ + 313, + 377, + 555, + 500 + ], + "type": "text", + "content": " thus returns " + }, + { + "bbox": [ + 313, + 377, + 555, + 500 + ], + "type": "inline_equation", + "content": "\\mathbf{N}" + }, + { + "bbox": [ + 313, + 377, + 555, + 500 + ], + "type": "text", + "content": " template indices corresponding to the lowest entropy entries. In practice, we use " + }, + { + "bbox": [ + 313, + 377, + 555, + 500 + ], + "type": "inline_equation", + "content": "N = 4" + }, + { + "bbox": [ + 313, + 377, + 555, + 500 + ], + "type": "text", + "content": " and validate this choice through an ablation study in our experimental section. Following Eq. (2), the expert classifier of class " + }, + { + "bbox": [ + 313, + 377, + 555, + 500 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 377, + 555, + 500 + ], + "type": "text", + "content": ", denoted " + }, + { + "bbox": [ + 313, + 377, + 555, + 500 + ], + "type": "inline_equation", + "content": "\\pmb {W}(\\hat{\\mathcal{E}}_k)" + }, + { + "bbox": [ + 313, + 377, + 555, + 500 + ], + "type": "text", + "content": ", is constructed by averaging the embeddings obtained from the " + }, + { + "bbox": [ + 313, + 377, + 555, + 500 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 377, + 555, + 500 + ], + "type": "text", + "content": " templates in " + }, + { + "bbox": [ + 313, + 377, + 555, + 500 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{E}}_k" + }, + { + "bbox": [ + 313, + 377, + 555, + 500 + ], + "type": "text", + "content": " for each of the " + }, + { + "bbox": [ + 313, + 377, + 555, + 500 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 377, + 555, + 500 + ], + "type": "text", + "content": " class names." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 501, + 555, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 501, + 555, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 501, + 555, + 611 + ], + "type": "text", + "content": "The above process leads to " + }, + { + "bbox": [ + 313, + 501, + 555, + 611 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 501, + 555, + 611 + ], + "type": "text", + "content": " classifiers " + }, + { + "bbox": [ + 313, + 501, + 555, + 611 + ], + "type": "inline_equation", + "content": "\\{\\pmb{W}(\\hat{\\mathcal{E}}_k)\\}_{k\\in [1,K]}" + }, + { + "bbox": [ + 313, + 501, + 555, + 611 + ], + "type": "text", + "content": ", all obtained in a fully unsupervised manner, and each is expected to excel on one specific class. However, it is important to note that although a classifier is expert of a single class, it outputs a full semantic map containing all " + }, + { + "bbox": [ + 313, + 501, + 555, + 611 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 501, + 555, + 611 + ], + "type": "text", + "content": " classes. Therefore, the above formulation produces " + }, + { + "bbox": [ + 313, + 501, + 555, + 611 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 501, + 555, + 611 + ], + "type": "text", + "content": " soft-segmentation maps, " + }, + { + "bbox": [ + 313, + 501, + 555, + 611 + ], + "type": "inline_equation", + "content": "\\{\\hat{\\mathbf{S}}_k\\}_{k\\in [1,K]}" + }, + { + "bbox": [ + 313, + 501, + 555, + 611 + ], + "type": "text", + "content": ". To output a single OVSS map, we introduce a simple scheme to fuse all expert predictions." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 619, + 468, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 619, + 468, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 619, + 468, + 632 + ], + "type": "text", + "content": "4.2. Fusion of expert predictions" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 637, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 637, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 637, + 555, + 714 + ], + "type": "text", + "content": "To fuse all predicted soft-segmentation maps " + }, + { + "bbox": [ + 313, + 637, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\{\\hat{\\mathbf{S}}_k\\}_{k\\in [1,K]}" + }, + { + "bbox": [ + 313, + 637, + 555, + 714 + ], + "type": "text", + "content": " into a single output " + }, + { + "bbox": [ + 313, + 637, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{S}}" + }, + { + "bbox": [ + 313, + 637, + 555, + 714 + ], + "type": "text", + "content": ", we rely on the fact that each map " + }, + { + "bbox": [ + 313, + 637, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{S}}_k" + }, + { + "bbox": [ + 313, + 637, + 555, + 714 + ], + "type": "text", + "content": " excels at segmenting class " + }, + { + "bbox": [ + 313, + 637, + 555, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 637, + 555, + 714 + ], + "type": "text", + "content": ". Therefore, for each pixel, the class assigned is the one having the highest probability among the expert classifiers which predicted their own class of expertise. Occasionally, when no expert predicts its own" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21474" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "content": "class of expertise for a given pixel, we simply revert to using the " + }, + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "content": " classifier. In practice, we observe that only " + }, + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "inline_equation", + "content": "\\approx 2\\%" + }, + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "content": " of pixels fall under the latter case. Therefore, the vast majority of decisions are carried out by experts." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 296, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 296, + 190 + ], + "type": "text", + "content": "Formally, for a given pixel " + }, + { + "bbox": [ + 55, + 121, + 296, + 190 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 121, + 296, + 190 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 121, + 296, + 190 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{S}}_k(x)" + }, + { + "bbox": [ + 55, + 121, + 296, + 190 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 55, + 121, + 296, + 190 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 121, + 296, + 190 + ], + "type": "text", + "content": "-dimensional softmax probability of the expert " + }, + { + "bbox": [ + 55, + 121, + 296, + 190 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 121, + 296, + 190 + ], + "type": "text", + "content": " at the position of " + }, + { + "bbox": [ + 55, + 121, + 296, + 190 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 121, + 296, + 190 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 55, + 121, + 296, + 190 + ], + "type": "inline_equation", + "content": "\\mathcal{X}(x) = \\{k \\mid \\arg \\max_{i \\in [1, K]} \\hat{\\mathbf{S}}_k(x)[i] = k\\}" + }, + { + "bbox": [ + 55, + 121, + 296, + 190 + ], + "type": "text", + "content": " be the index set of experts that predicted their own class of expertise for " + }, + { + "bbox": [ + 55, + 121, + 296, + 190 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 121, + 296, + 190 + ], + "type": "text", + "content": ". The prediction of " + }, + { + "bbox": [ + 55, + 121, + 296, + 190 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 121, + 296, + 190 + ], + "type": "text", + "content": " is determined as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 63, + 198, + 294, + 252 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 198, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 63, + 198, + 294, + 252 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {P}} (x) = \\left\\{ \\begin{array}{l l} k ^ {*} = \\underset {k \\in \\mathcal {X} (x)} {\\arg \\max } \\hat {\\mathbf {S}} _ {k} (x) [ k ], & \\text {i f} \\mathcal {X} (x) \\neq \\varnothing , \\\\ k ^ {\\dagger} = \\underset {k \\in [ 1, K ]} {\\arg \\max } \\mathbf {S} _ {\\operatorname {C L I P}} (x) [ k ], & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {7}", + "image_path": "8c3c1b1effc8f387a39418f65a4c772f74f47df28e3d291499fc5689bea07139.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 253, + 296, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 253, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 54, + 253, + 296, + 373 + ], + "type": "text", + "content": "Fig. 2 illustrates the unsupervised identification of class-experts as well as the fusion strategy. Overall, our method incurs computational overhead primarily from computing cosine similarities for the " + }, + { + "bbox": [ + 54, + 253, + 296, + 373 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 54, + 253, + 296, + 373 + ], + "type": "text", + "content": " class-experts, which scales with the number of classes and becomes more significant when " + }, + { + "bbox": [ + 54, + 253, + 296, + 373 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 54, + 253, + 296, + 373 + ], + "type": "text", + "content": " is very large (" + }, + { + "bbox": [ + 54, + 253, + 296, + 373 + ], + "type": "inline_equation", + "content": ">100" + }, + { + "bbox": [ + 54, + 253, + 296, + 373 + ], + "type": "text", + "content": "). However, this overhead is constrained to the similarity computations since the vision encoder is shared across all experts, requiring only a single forward pass regardless of the number of classes. We provide computational details in Appendix A.2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 382, + 138, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 382, + 138, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 382, + 138, + 396 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 403, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 296, + 521 + ], + "type": "text", + "content": "Experimental setup. We evaluate our method on three state-of-the-art open-vocabulary semantic segmentation models: the seminal MaskCLIP [79] and two recent models CLIP-DINOiser [66] and NACLIP [23]. CLIP-DINOiser trains a convolution layer using only 1K unlabeled images from ImageNet [15] for fast inference on pixel-level segmentation tasks, effectively imitating the DINO priors for pooling guidance and removing the need for DINO encoder at runtime. MaskCLIP and NACLIP are training-free adaptation methods of CLIP." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "text", + "content": "For a fair comparison, we report the version of NACLIP without refinement. We primarily select training-free models that keep the image encoder frozen and produce unaltered, dense image-language features. This enables us to assess the impact of the class-experts identification strategy by using the original CLIP pre-trained features and ImageNet templates. Unless otherwise stated, all models utilize CLIP ViT-B/16 backbone. CLIP-DINOiser and MaskCLIP are based on OpenCLIP [11], while NACLIP employs OpenAI's original CLIP [47]. For all models - CLIP-DINOiser, NACLIP, and MaskCLIP- we follow the authors' evaluation protocol. For CLIP-DINOiser and MaskCLIP, we resize images to 448 pixels on the shortest side with a sliding window of size 448 and stride 224. For NACLIP, images are resized to 336 pixels (except for Cityscapes which are resized to 560 pixels) with a 224 window and stride 112." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 315, + 70, + 554, + 234 + ], + "blocks": [ + { + "bbox": [ + 315, + 70, + 554, + 234 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 554, + 234 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 554, + 234 + ], + "type": "table", + "html": "
MethodCSVOC20PC59ADEStuffAvg
CLIP [47]6.749.111.23.25.715.2
GroupViT [71]11.179.723.49.215.327.7
CLIP-Surgery [37]31.4--12.921.9-
GEM [3]--32.615.7--
SCLIP [63]32.280.434.216.122.437.1
CLIP-DIY [65]11.679.719.89.913.326.9
TCL [7]23.177.530.314.919.633.1
ReCo [53]21.157.822.311.214.825.4
MaskCLIP [79]25.061.825.514.217.528.7
+ FLOSS25.861.826.214.917.829.3
NACLIP [23]35.579.735.217.423.338.2
+ FLOSS37.080.235.918.423.639.0
CLIP-DINOiser [66]31.380.935.920.024.638.5
+ FLOSS34.682.336.220.724.739.7
", + "image_path": "f051b3f4d1281ea70252ffb49bf3af721a805db56259c88c8fbe96cc2aafbfcf.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 243, + 555, + 321 + ], + "lines": [ + { + "bbox": [ + 313, + 243, + 555, + 321 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 555, + 321 + ], + "type": "text", + "content": "Table 1. OVSS across datasets and models. We report mIoU metric for eight OVSS baselines and three models either using the average templates (i.e., MaskCLIP, NACLIP, and CLIP-DINOiser) or with our method (i.e., +FLOSS). FLOSS consistently improves OVSS models across datasets of varying complexity, from urban scenes (Cityscapes, 19 classes) to general objects (COCO-Stuff, 171 classes). Bold highlights the best performance." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 340, + 555, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 555, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 555, + 507 + ], + "type": "text", + "content": "Datasets and Metric. We evaluate FLOSS across multiple semantic segmentation benchmarks: PASCAL CONTEXT 59 [42] (PC59), PASCAL VOC 20 [17] (VOC20), COCO-Stuff [5] (Stuff), ADE20K [78] (ADE), and Cityscapes [13] (CS). Note that these datasets exhibit different ontology and semantic granularity, ranging from 19 classes (CS) to 171 classes (Stuff). Additionally, we demonstrate the generalization capabilities of our class-experts identified on Cityscapes to other driving datasets such as BDD-100K [75] and MAPILLARY [44] as well as ACDC [52] under Night, Fog, Rain and Snow conditions. Importantly, our approach is training-free and uses no labels. Performance is measured using mean Intersection over Union (mIoU)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 510, + 556, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 583 + ], + "type": "text", + "content": "Implementation details. The only hyperparameter of FLOSS, being the number of class-experts to identify in Eq. (6), is set to " + }, + { + "bbox": [ + 313, + 510, + 556, + 583 + ], + "type": "inline_equation", + "content": "N = 4" + }, + { + "bbox": [ + 313, + 510, + 556, + 583 + ], + "type": "text", + "content": " in all experiments. While our method is training-free, class-experts are estimated using the training set of each dataset, while performance is always reported on the corresponding validation set." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 588, + 397, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 397, + 600 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 397, + 600 + ], + "type": "text", + "content": "5.1. Main results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Tab. 1 reports the mIoU of OVSS baselines across five datasets showing FLOSS plugged into three different models: MaskCLIP, NACLIP and CLIP-DINOiser. The performance gains vary with dataset complexity, showing smaller improvements on VOC20 " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "[+0.05, +0.5, +1.4]" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": " or PC59 " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "[+0.7, +0.7, +0.3]" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": " than on CS " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "[+0.8, +1.5, +3.3]" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": ". We attribute this to VOC20 and PC59 being closer to ImageNet which was leveraged in CLIP [47] to engineer the set of 80 templates used in all models. The benefit of our method" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21475" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 295, + 115 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 295, + 115 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 295, + 115 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 295, + 115 + ], + "type": "table", + "html": "
MethodCSVOC20PC59ADEStuffAvg
MaskCLIP + FLOSS51%54%56%57%56%55%
NACLIP + FLOSS49%20%57%57%52%47%
CLIP-DINOiser + FLOSS62%41%57%45%44%50%
", + "image_path": "f0227326c31fe3ebe4ddc064854bdd999b0c22b39b3ed846e1c041074c9c17c1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 167, + 296, + 255 + ], + "blocks": [ + { + "bbox": [ + 55, + 124, + 295, + 158 + ], + "lines": [ + { + "bbox": [ + 55, + 124, + 295, + 158 + ], + "spans": [ + { + "bbox": [ + 55, + 124, + 295, + 158 + ], + "type": "text", + "content": "Table 2. Quality of our estimated class-experts. We measure the quality of our estimated experts by evaluating their intersection with the true experts, cf. Eq. (8)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 167, + 296, + 255 + ], + "lines": [ + { + "bbox": [ + 56, + 167, + 296, + 255 + ], + "spans": [ + { + "bbox": [ + 56, + 167, + 296, + 255 + ], + "type": "table", + "html": "
Methodin domain CSout of domain
NightFogRainSnowBDDMAPAvg
MaskCLIP24.513.320.321.019.922.725.820.5
+ FLOSS25.813.921.722.321.322.926.821.5
NACLIP35.523.332.930.732.631.435.331.0
+ FLOSS37.023.632.531.033.632.336.531.6
CLIP-DINOiser31.312.724.127.927.231.335.226.4
+ FLOSS34.616.629.029.829.432.236.628.9
", + "image_path": "905a10623f4db0ba0380402e2a1773150cc53ed1975652074d978e0ac7bacc7f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 264, + 296, + 331 + ], + "lines": [ + { + "bbox": [ + 55, + 264, + 296, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 296, + 331 + ], + "type": "text", + "content": "Table 3. Generalization of experts across datasets. We evaluate the transferability of experts identified on the Cityscapes (CS) dataset to unseen datasets that share the same semantic classes and exhibit changes in lighting (Night), weather (Fog, Rain, Snow), or location and scenery (BDD, MAP). Results show a consistent improvement when using FLOSS, across almost all settings." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 350, + 295, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 295, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 295, + 387 + ], + "type": "text", + "content": "is evident on challenging benchmarks like ADE and Stuff, which have a large number of semantic categories making accurate segmentation particularly hard." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 389, + 296, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 296, + 426 + ], + "type": "text", + "content": "Quality of experts. We evaluate the quality " + }, + { + "bbox": [ + 55, + 389, + 296, + 426 + ], + "type": "inline_equation", + "content": "\\hat{\\rho}_k" + }, + { + "bbox": [ + 55, + 389, + 296, + 426 + ], + "type": "text", + "content": " of our experts by measuring, for each class " + }, + { + "bbox": [ + 55, + 389, + 296, + 426 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 389, + 296, + 426 + ], + "type": "text", + "content": ", the intersection of the estimated class-expert " + }, + { + "bbox": [ + 55, + 389, + 296, + 426 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{E}}_k" + }, + { + "bbox": [ + 55, + 389, + 296, + 426 + ], + "type": "text", + "content": " with the true set of experts " + }, + { + "bbox": [ + 55, + 389, + 296, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_k" + }, + { + "bbox": [ + 55, + 389, + 296, + 426 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 83, + 434, + 295, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 434, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 83, + 434, + 295, + 464 + ], + "type": "interline_equation", + "content": "\\hat {\\rho} _ {k} = 1 0 0 \\times \\frac {\\left| \\hat {\\mathcal {E}} _ {k} \\cap \\mathcal {E} _ {k} \\right|}{\\left| \\hat {\\mathcal {E}} _ {k} \\right|} = 1 0 0 \\times \\frac {\\left| \\hat {\\mathcal {E}} _ {k} \\cap \\mathcal {E} _ {k} \\right|}{N}. \\tag {8}", + "image_path": "dc69b4536d039507e488abfda1abd1781aad9d82a6d7405d8b12c156ace498ee.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 471, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 471, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 296, + 567 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 55, + 471, + 296, + 567 + ], + "type": "inline_equation", + "content": "|.|" + }, + { + "bbox": [ + 55, + 471, + 296, + 567 + ], + "type": "text", + "content": " denotes the cardinality of a set. Quality averaged over classes is reported in Tab. 2 and shows that regardless of the model used, we correctly estimate approximately half of the true experts. In some scenarios, such as PASCAL VOC 20 with NACLIP, even when only " + }, + { + "bbox": [ + 55, + 471, + 296, + 567 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 55, + 471, + 296, + 567 + ], + "type": "text", + "content": " of predicted experts are true experts, it is sufficient to bring some improvements (" + }, + { + "bbox": [ + 55, + 471, + 296, + 567 + ], + "type": "inline_equation", + "content": "+0.5" + }, + { + "bbox": [ + 55, + 471, + 296, + 567 + ], + "type": "text", + "content": " mIoU in Tab. 1). Per-class quality is reported in Appendix A.4, showing some variability." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "Generalization to unseen datasets. We now evaluate the generalization of experts to other unseen datasets sharing the same label set but exhibiting distribution shifts. In Tab. 3, we report performance of FLOSS when experts are identified on Cityscapes and used to evaluate performance on six unseen datasets which include four ACDC conditions (Night, Fog, Rain, Snow), BDD-100K, and MAPILLARY. Overall, we consistently improve the performance with notable gains of up to " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "+4.9" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": " mIoU for CLIP-DINOiser on ACDC Fog. This demonstrates that our class-wise experts are readily usable on other datasets sharing the same semantic classes, despite the presence of distribution shifts." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 314, + 68, + 433, + 148 + ], + "blocks": [ + { + "bbox": [ + 314, + 68, + 433, + 148 + ], + "lines": [ + { + "bbox": [ + 314, + 68, + 433, + 148 + ], + "spans": [ + { + "bbox": [ + 314, + 68, + 433, + 148 + ], + "type": "image", + "image_path": "a3551e18aa80179087cf1d94d21fdec9e5e993b0e2b3c7c1c3d0e3b297761cb0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 148, + 399, + 159 + ], + "lines": [ + { + "bbox": [ + 351, + 148, + 399, + 159 + ], + "spans": [ + { + "bbox": [ + 351, + 148, + 399, + 159 + ], + "type": "text", + "content": "(a) Cityscapes" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 168, + 555, + 236 + ], + "lines": [ + { + "bbox": [ + 313, + 168, + 555, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 555, + 236 + ], + "type": "text", + "content": "Figure 3. Effect of the number of images. We evaluate CLIP-DINOiser with FLOSS on urban scenes (a) and general object segmentation (b), when varying the number of images used to identify class-experts. Results show the average of 10 seeds, with standard deviation. Despite larger variation, we note the robustness of our method, even when having access to very few unlabeled images." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 435, + 70, + 553, + 148 + ], + "blocks": [ + { + "bbox": [ + 435, + 70, + 553, + 148 + ], + "lines": [ + { + "bbox": [ + 435, + 70, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 435, + 70, + 553, + 148 + ], + "type": "image", + "image_path": "955bcc5f50b34a848238d5a50be45dadf422d01cd017d2dcab2dfaca88143656.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 459, + 148, + 531, + 158 + ], + "lines": [ + { + "bbox": [ + 459, + 148, + 531, + 158 + ], + "spans": [ + { + "bbox": [ + 459, + 148, + 531, + 158 + ], + "type": "text", + "content": "(b) PASCAL VOC 20" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 262, + 555, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 555, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 555, + 467 + ], + "type": "text", + "content": "Low-data regimes. To investigate the data efficiency of FLOSS, we study its performance in lower data regimes. Fig. 3 reports the mIoU of CLIP-DINOiser on Cityscapes and PASCAL VOC 20, as the number of images sampled from their respective training sets (containing 2,975 and 1,464 images, respectively) varies. The figure shows the average and standard deviation over 10 seeds (i.e., randomly sampling 10 different image sets). On both datasets, the general trend shows that our method benefits from accessing more images, which stems from the better estimation of class-experts. Remarkably, our method performs well in a very low-data regime, outperforming the default CLIP-DINOiser with only 1 image on Cityscapes and around 25 images on PASCAL VOC 20. We conjecture that the latter requires more images due to its lower per-image class density compared to Cityscapes, which exhibits higher variability and density per image." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 475, + 554, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 554, + 560 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 554, + 560 + ], + "type": "text", + "content": "Low-data regimes & domain shift. To further corroborate its practicality, we evaluate FLOSS under stricter conditions combining both low-data regimes and domain shift. Specifically, we use only a single CS image to identify experts that are tested on 6 shifted domains. Tab. 4 provides evidence on the effectiveness of FLOSS in this strict setting, leading to an average improvement of " + }, + { + "bbox": [ + 313, + 475, + 554, + 560 + ], + "type": "inline_equation", + "content": "1.9\\%" + }, + { + "bbox": [ + 313, + 475, + 554, + 560 + ], + "type": "text", + "content": " across OOD datasets." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": "Expert transferability. We investigate the cross-dataset transferability of class-experts by evaluating whether experts identified on one dataset can effectively improve performance on semantically related but distinct datasets. We employ COCO-Stuff (171 classes) to identify class-experts for CLIP-DINOiser and apply them to Cityscapes, PASCAL VOC 20, and PASCAL CONTEXT 59, which share 10, 16, and 37 common classes with COCO-Stuff, respectively. For non-overlapping classes, we use the default average classifier " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": ". As reported in Tab. 5, FLOSS shows consistent improvements, confirming expert transferability across datasets." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21476" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 295, + 114 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 295, + 114 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 295, + 114 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 295, + 114 + ], + "type": "table", + "html": "
Methodin domain CSout of domain
NightFogRainSnowBDDMAPAvg
CLIP-DINOiser31.312.724.127.927.231.335.226.4
+ FLOSS33.216.228.229.328.731.735.628.3
", + "image_path": "7a39889922c6d7faa44ce922202182fb10a2113e64d03e0d70441ec39c15b22f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 191, + 296, + 244 + ], + "blocks": [ + { + "bbox": [ + 55, + 123, + 295, + 179 + ], + "lines": [ + { + "bbox": [ + 55, + 123, + 295, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 123, + 295, + 179 + ], + "type": "text", + "content": "Table 4. FLOSS with 1 image & domain shift. Experts are identified using a single Cityscapes (CS) image and evaluated on unseen datasets that share the same semantic classes and exhibit changes in lighting (Night), weather (Fog, Rain, Snow), or location and scenery (BDD, MAP). We observe a consistent boost." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 191, + 296, + 244 + ], + "lines": [ + { + "bbox": [ + 56, + 191, + 296, + 244 + ], + "spans": [ + { + "bbox": [ + 56, + 191, + 296, + 244 + ], + "type": "table", + "html": "
Methodin-domain Stuffout of domain
CS commonallVOC20 commonallPC59 commonall
CLIP-DINOiser24.634.531.180.580.839.336.0
+ FLOSS24.735.832.282.382.440.036.2
", + "image_path": "1fe61797bfb1597d562dfcc0bea8de282b6a715118bfd086f55369743a38c001.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 330, + 149, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 330, + 149, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 330, + 149, + 342 + ], + "type": "text", + "content": "5.2. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 347, + 295, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 347, + 295, + 526 + ], + "spans": [ + { + "bbox": [ + 54, + 347, + 295, + 526 + ], + "type": "text", + "content": "Fusion strategies for expert predictions. We investigate alternative strategies to fuse the expert predictions (cf., Sec. 4.2), reporting performance on Cityscapes in Tab. 6. Simply averaging all soft segmentation maps (\"Average-all\") irrespective of the templates class of expertise leads to suboptimal results across all models. When accounting for templates' expertise in fusion, conflicts arise often as several experts may predict their own class of expertise. We therefore explore three strategies to resolve conflicts: \"Default\" which falls back to " + }, + { + "bbox": [ + 54, + 347, + 295, + 526 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 54, + 347, + 295, + 526 + ], + "type": "text", + "content": " predictions, \"Average\" which averages the probability vectors of conflicting experts, and \"Highest\" which selects the most confident expert as described in Sec. 4.2. The latter consistently yields the best performance across all models, validating our fusion scheme." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 532, + 295, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 295, + 592 + ], + "type": "text", + "content": "Metrics for expert identification. We replace entropy in Eq. (6) with other unsupervised metrics to act as proxies of the template class-wise performance, reporting performance on Cityscapes using the CLIP-DINOiser model in Fig. 4a." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 592, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 55, + 592, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 592, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 295, + 616 + ], + "type": "text", + "content": "(i) Avg. Probability computes the average probability assigned to class " + }, + { + "bbox": [ + 55, + 592, + 295, + 616 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 592, + 295, + 616 + ], + "type": "text", + "content": " across all pixels predicted as class " + }, + { + "bbox": [ + 55, + 592, + 295, + 616 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 592, + 295, + 616 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 616, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 295, + 664 + ], + "type": "text", + "content": "(ii) " + }, + { + "bbox": [ + 55, + 616, + 295, + 664 + ], + "type": "inline_equation", + "content": "MaNo" + }, + { + "bbox": [ + 55, + 616, + 295, + 664 + ], + "type": "text", + "content": " [70] is adapted as it was originally designed for assessing general model performance in image classification, to evaluate class-wise performance in semantic segmentation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "(iii) \"ITI\" [25] is the Inter-class separability to Intra-class similarity ratio, and captures both the separability of the classes in the feature space as well as the ability to keep samples of the same class close together." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "type": "table", + "bbox": [ + 329, + 70, + 543, + 146 + ], + "blocks": [ + { + "bbox": [ + 55, + 252, + 296, + 308 + ], + "lines": [ + { + "bbox": [ + 55, + 252, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 296, + 308 + ], + "type": "text", + "content": "Table 5. Cross-dataset expert transferability. Class-experts identified on COCO-Stuff are applied to CS, VOC20, and PC59. Results are reported for common classes shared with COCO-Stuff and all classes. FLOSS demonstrates consistent improvements across all target datasets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 329, + 70, + 543, + 146 + ], + "lines": [ + { + "bbox": [ + 329, + 70, + 543, + 146 + ], + "spans": [ + { + "bbox": [ + 329, + 70, + 543, + 146 + ], + "type": "table", + "html": "
FusionConflictMaskCLIP +FLOSSNACLIP +FLOSSCLIP-DINOiser +FLOSS
Average-all24.736.631.2
ExpertDefault24.836.632.0
ExpertAverage25.636.734.4
ExpertHighest25.837.034.6
", + "image_path": "6fea1e91f20c515277267c1165e8571ff5789ed7d0df6cdf457c1e538003c189.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 154, + 555, + 188 + ], + "lines": [ + { + "bbox": [ + 313, + 154, + 555, + 188 + ], + "spans": [ + { + "bbox": [ + 313, + 154, + 555, + 188 + ], + "type": "text", + "content": "Table 6. Strategies for the fusion of expert predictions. Exploring various fusion strategies on Cityscapes shows the benefit of our strategy, which outperforms all others across models." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 315, + 200, + 422, + 268 + ], + "blocks": [ + { + "bbox": [ + 315, + 200, + 422, + 268 + ], + "lines": [ + { + "bbox": [ + 315, + 200, + 422, + 268 + ], + "spans": [ + { + "bbox": [ + 315, + 200, + 422, + 268 + ], + "type": "image", + "image_path": "35f018e50f043a0a4c394456be68202bbc9065a31f331cacd0ad8061f09e7407.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 270, + 399, + 280 + ], + "lines": [ + { + "bbox": [ + 339, + 270, + 399, + 280 + ], + "spans": [ + { + "bbox": [ + 339, + 270, + 399, + 280 + ], + "type": "text", + "content": "(a) Expert metrics" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 425, + 200, + 554, + 284 + ], + "blocks": [ + { + "bbox": [ + 425, + 200, + 554, + 284 + ], + "lines": [ + { + "bbox": [ + 425, + 200, + 554, + 284 + ], + "spans": [ + { + "bbox": [ + 425, + 200, + 554, + 284 + ], + "type": "image", + "image_path": "5cd0243995006990d162def40ea3ee48a4b6dee6952de3490583bee4abf28408.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 458, + 285, + 522, + 295 + ], + "lines": [ + { + "bbox": [ + 458, + 285, + 522, + 295 + ], + "spans": [ + { + "bbox": [ + 458, + 285, + 522, + 295 + ], + "type": "text", + "content": "(b) Effect of Top-N" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 366, + 395, + 502, + 491 + ], + "blocks": [ + { + "bbox": [ + 313, + 304, + 555, + 384 + ], + "lines": [ + { + "bbox": [ + 313, + 304, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 555, + 384 + ], + "type": "text", + "content": "Figure 4. Ablation of experts. We ablate our choices of metrics to identify experts as well as the number of experts per class. (a) shows unsupervised metrics to identify the experts, showing that our choice of entropy yields the best performance, although 'Avg. Probability' performs on par. (b) reports performance of FLOSS when varying the number of " + }, + { + "bbox": [ + 313, + 304, + 555, + 384 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 304, + 555, + 384 + ], + "type": "text", + "content": " experts. Results are reported on the Cityscapes dataset using the CLIP-DINOiser model." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 366, + 395, + 502, + 491 + ], + "lines": [ + { + "bbox": [ + 366, + 395, + 502, + 491 + ], + "spans": [ + { + "bbox": [ + 366, + 395, + 502, + 491 + ], + "type": "image", + "image_path": "104618af780225ba9347ec74934024773ff5ddc0e7c15be94356703eec0c6d69.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 499, + 555, + 555 + ], + "lines": [ + { + "bbox": [ + 313, + 499, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 499, + 555, + 555 + ], + "type": "text", + "content": "Figure 5. Effect of class-expert quality. We report the performance of FLOSS in the oracle-based setting, where each class-wise set of 4 experts contains a ratio " + }, + { + "bbox": [ + 313, + 499, + 555, + 555 + ], + "type": "inline_equation", + "content": "(\\rho)" + }, + { + "bbox": [ + 313, + 499, + 555, + 555 + ], + "type": "text", + "content": " of true experts. The experiment is conducted on Cityscapes using the CLIP-DINOiser model." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 577, + 494, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 577, + 494, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 577, + 494, + 589 + ], + "type": "text", + "content": "These metrics are detailed in Appendix A.5." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 590, + 553, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 553, + 613 + ], + "type": "text", + "content": "Overall, our entropy measure (cf. Eq. (5)) is the most effective, although Avg. Probability performs on par." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 617, + 555, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 677 + ], + "type": "text", + "content": "Impact of class-expert identification quality. At the core of our method is the identification of class-experts. While we demonstrated there are templates which are class-expert, for each class there exist also under-performing templates (cf. Fig. 1) which, if selected, could harm the performance." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": "To measure the impact of experts' quality on FLOSS, we devise an oracle-based experiment using " + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "inline_equation", + "content": "N = 4" + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": " templates per-class, where we have access to ground truth la" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21477" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 70, + 282, + 149 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 282, + 149 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 282, + 149 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 282, + 149 + ], + "type": "table", + "html": "
MethodViT-B/16ViT-L/14
CSPC59Avg.CSPC59Avg.
SCLIP [63]32.234.233.221.325.223.3
GEM [3]32.635.634.127.128.127.6
NACLIP [23]35.535.235.431.432.131.8
+ FLOSS37.035.936.532.432.432.4
", + "image_path": "840e7a385f70f7b5a999234090df76da968e3692fa5bf6995f02f0a260f4e4ee.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 158, + 295, + 191 + ], + "lines": [ + { + "bbox": [ + 55, + 158, + 295, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 158, + 295, + 191 + ], + "type": "text", + "content": "Table 7. Effect of backbone. Results (mIoU) show that NACLIP + FLOSS outperforms NACLIP across ViT-B/16 and ViT-L/14 CLIP backbones." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 224, + 295, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 224, + 295, + 416 + ], + "spans": [ + { + "bbox": [ + 54, + 224, + 295, + 416 + ], + "type": "text", + "content": "bels to identify the true class-experts. We randomly sample a fixed ratio " + }, + { + "bbox": [ + 54, + 224, + 295, + 416 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 54, + 224, + 295, + 416 + ], + "type": "text", + "content": " of templates among true experts (i.e., " + }, + { + "bbox": [ + 54, + 224, + 295, + 416 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_k" + }, + { + "bbox": [ + 54, + 224, + 295, + 416 + ], + "type": "text", + "content": " from Eq. (4)) and the remaining ones among non-experts (i.e., " + }, + { + "bbox": [ + 54, + 224, + 295, + 416 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{T}_m\\} \\backslash \\mathcal{E}_k" + }, + { + "bbox": [ + 54, + 224, + 295, + 416 + ], + "type": "text", + "content": "). For example, a ratio of " + }, + { + "bbox": [ + 54, + 224, + 295, + 416 + ], + "type": "inline_equation", + "content": "\\rho = 75\\%" + }, + { + "bbox": [ + 54, + 224, + 295, + 416 + ], + "type": "text", + "content": " indicates that 3 templates are experts and 1 is non-expert. The outcome of this experiment on CS using the CLIP-DINOiser model is reported in Fig. 5 and shows an expected performance boost when the ratio of true experts increases. It also highlights that " + }, + { + "bbox": [ + 54, + 224, + 295, + 416 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 54, + 224, + 295, + 416 + ], + "type": "text", + "content": " of true experts is sufficient for FLOSS to surpass the original CLIP-DINOiser. Interestingly, the performance of CLIP-DINOiser + FLOSS in Tab. 1 (34.6) slightly surpasses the performance at " + }, + { + "bbox": [ + 54, + 224, + 295, + 416 + ], + "type": "inline_equation", + "content": "\\rho = 100\\%" + }, + { + "bbox": [ + 54, + 224, + 295, + 416 + ], + "type": "text", + "content": " shown in Fig. 5 (34.1). This however does not indicate that our estimated templates in Tab. 1 are all true experts - as reflected by Tab. 2 - but rather that our correctly estimated experts are on average better than random true experts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 421, + 296, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 421, + 296, + 494 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 296, + 494 + ], + "type": "text", + "content": "To further measure the upper bound of CLIP-DINOiser + FLOSS, we implement an oracle version using only the best true class-expert (i.e., " + }, + { + "bbox": [ + 55, + 421, + 296, + 494 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{T}_{\\tilde{m}} \\mid \\tilde{m} \\in \\mathrm{Top-N}(\\Omega_k(W(\\mathcal{T}_m)))\\}" + }, + { + "bbox": [ + 55, + 421, + 296, + 494 + ], + "type": "text", + "content": "). On CS, the oracle with " + }, + { + "bbox": [ + 55, + 421, + 296, + 494 + ], + "type": "inline_equation", + "content": "N \\in \\{1, 2, 3, 4\\}" + }, + { + "bbox": [ + 55, + 421, + 296, + 494 + ], + "type": "text", + "content": " achieves respectively 37.1/37.0/37.3/37.5%, showing the potential for improvement assuming the identification of better experts." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 507, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 507, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 507, + 295, + 604 + ], + "type": "text", + "content": "FLOSS using the validation set. First, we investigate the effectiveness of FLOSS in a transductive setting. Using the validation set of CS for expert selection, CLIP-DINOiser + FLOSS achieves " + }, + { + "bbox": [ + 55, + 507, + 295, + 604 + ], + "type": "inline_equation", + "content": "34.4\\%" + }, + { + "bbox": [ + 55, + 507, + 295, + 604 + ], + "type": "text", + "content": " mIoU on the same set, compared to " + }, + { + "bbox": [ + 55, + 507, + 295, + 604 + ], + "type": "inline_equation", + "content": "31.3\\%" + }, + { + "bbox": [ + 55, + 507, + 295, + 604 + ], + "type": "text", + "content": " for CLIP-DINOiser. Second, using a portion of the validation set to select experts, i.e., 5/25/50/75 val images, we achieve 32.8/33.0/33.0/33.3% on the rest of the validation images, compared to " + }, + { + "bbox": [ + 55, + 507, + 295, + 604 + ], + "type": "inline_equation", + "content": "30.9\\%" + }, + { + "bbox": [ + 55, + 507, + 295, + 604 + ], + "type": "text", + "content": " for CLIP-DINOiser alone." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "content": "Impact of different backbones. Tab. 7 demonstrates the effectiveness of our proposed FLOSS when integrated with NACLIP using the ViT-L/14 backbone. Our approach achieves consistent improvements, reaching " + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "inline_equation", + "content": "32.4\\%" + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "content": " mIoU on both Cityscapes and PC59 datasets, leading to a 0.6 gain in average performance over NACLIP. These results validate the effectiveness of our method when using large-scale vision transformers." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 314, + 70, + 553, + 148 + ], + "blocks": [ + { + "bbox": [ + 314, + 70, + 553, + 148 + ], + "lines": [ + { + "bbox": [ + 314, + 70, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 314, + 70, + 553, + 148 + ], + "type": "image", + "image_path": "3be22c283219e3b72b7b87d86c2aa877d663ef3468c30965790df2fce340d449.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 155, + 555, + 221 + ], + "lines": [ + { + "bbox": [ + 313, + 155, + 555, + 221 + ], + "spans": [ + { + "bbox": [ + 313, + 155, + 555, + 221 + ], + "type": "text", + "content": "Figure 6. Performance of best-class experts. Performance gains (IoU) when using best class-experts versus standard template-averaging. Results are shown for CLIP-DINOiser, MaskCLIP and NACLIP. The bars represent the IoU difference between using best class-experts (selected with ground-truth labels) and the default template-averaging approach." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 243, + 485, + 256 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 485, + 256 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 485, + 256 + ], + "type": "text", + "content": "6. Perspectives and Future Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 263, + 555, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 263, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 555, + 384 + ], + "type": "text", + "content": "Our experiments show that entropy is a strong estimator of class-experts. However, future research could develop even better proxies for prediction accuracy. Template-based methods appear especially promising, with significant potential for improving OVSS performance. As illustrated in Fig. 6, there is still a large gap to close—for example, selecting the best expert for the \"sky\" class could improve IoU by over 30 percentage points. These \"upper-bounds\" could be further increased by using more or augmented templates beyond the 80 currently used." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 396, + 388, + 409 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 388, + 409 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 388, + 409 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 418, + 555, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 555, + 621 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 555, + 621 + ], + "type": "text", + "content": "We propose a new task for improving OVSS models without having access to labels and without training, from a prompt template perspective. Our work is motivated by an intriguing observation: for each class, some templates excel in segmenting it. We refer to these templates as class-experts. Given a set of unlabeled images, our FLOSS uses the entropy of predictions as an unsupervised metric to identify these class-experts, thus not requiring any training or labels. Once selected, we propose a simple scheme for fusing the predictions of the class-experts, resulting in an improved overall OVSS performance in the inductive setting, even in the presence of distribution shifts. Additionally, we corroborate the effectiveness of FLOSS, which benefits from being plug-and-play, and further show that only few unlabeled images can be sufficient for expert selection. Furthermore, the oracle performance shows the potential of the class-experts, paving the way for future research." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 625, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 625, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 625, + 556, + 715 + ], + "type": "text", + "content": "Acknowledgment. This research was partially funded by the French Agence Nationale de la Recherche (ANR) with the project SIGHT (ANR-20-CE23-0016) and supported by ELSA - European Lighthouse on Secure and Safe AI funded by the European Union under grant agreement No. 101070617. We sincerely thank Telecom Paris for providing the resources necessary to run our experiments and Nacereddine Laddaoui for his help with infrastructure. We are also grateful to Ivan Lopes for proofreading." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21478" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Stephen H Bach, Victor Sanh, Zheng-Xin Yong, Albert Webson, Colin Raffel, Nihal V Nayak, Abheesht Sharma, Taewoon Kim, M Saiful Bari, Thibault Fevry, et al. Prompt-source: An integrated development environment and repository for natural language prompts. In ACL, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 146, + 296, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 146, + 296, + 179 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 296, + 179 + ], + "type": "text", + "content": "[2] Sule Bai, Yong Liu, Yifei Han, Haoji Zhang, and Yansong Tang. Self-calibrated clip for training-free open-vocabulary segmentation. arXiv preprint arXiv:2411.15869, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 180, + 295, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 180, + 295, + 223 + ], + "spans": [ + { + "bbox": [ + 62, + 180, + 295, + 223 + ], + "type": "text", + "content": "[3] Walid Bousselham, Felix Petersen, Vittorio Ferrari, and Hilde Kuehne. Grounding everything: Emerging localization properties in vision-language transformers. In CVPR, 2024. 2, 5, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 224, + 294, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 224, + 294, + 256 + ], + "spans": [ + { + "bbox": [ + 62, + 224, + 294, + 256 + ], + "type": "text", + "content": "[4] Maxime Bucher, Tuan-Hung Vu, Matthieu Cord, and Patrick Pérez. Zero-shot semantic segmentation. In NeurlPS, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 258, + 295, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 295, + 290 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 295, + 290 + ], + "type": "text", + "content": "[5] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In CVPR, 2018. 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 291, + 294, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 291, + 294, + 334 + ], + "spans": [ + { + "bbox": [ + 62, + 291, + 294, + 334 + ], + "type": "text", + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 335, + 294, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 335, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 62, + 335, + 294, + 369 + ], + "type": "text", + "content": "[7] Junbum Cha, Jonghwan Mun, and Byungseok Roh. Learning to generate text-grounded mask for open-world semantic segmentation from only image-text pairs. In CVPR, 2023. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 370, + 294, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 370, + 294, + 412 + ], + "spans": [ + { + "bbox": [ + 62, + 370, + 294, + 412 + ], + "type": "text", + "content": "[8] Guangyi Chen, Weiran Yao, Xiangchen Song, Xinyue Li, Yongming Rao, and Kun Zhang. Plot: Prompt learning with optimal transport for vision-language models. In ICLR, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 414, + 294, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 414, + 294, + 457 + ], + "spans": [ + { + "bbox": [ + 62, + 414, + 294, + 457 + ], + "type": "text", + "content": "[9] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. TPAMI, 2017. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 459, + 294, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 459, + 294, + 501 + ], + "spans": [ + { + "bbox": [ + 57, + 459, + 294, + 501 + ], + "type": "text", + "content": "[10] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 502, + 294, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 502, + 294, + 556 + ], + "spans": [ + { + "bbox": [ + 57, + 502, + 294, + 556 + ], + "type": "text", + "content": "[11] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, 2023. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 558, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 558, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 57, + 558, + 294, + 601 + ], + "type": "text", + "content": "[12] Seokju Cho, Heeseong Shin, Sunghwan Hong, Anurag Arnab, Paul Hongsuck Seo, and Seungryong Kim. Cat-seg: Cost aggregation for open-vocabulary semantic segmentation. In CVPR, 2024. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 602, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 294, + 656 + ], + "type": "text", + "content": "[13] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 3, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 294, + 713 + ], + "type": "text", + "content": "[14] Ulindu De Silva, Didula Samaraweera, Sasini Wanigathunga, Kavindu Kariyawasam, Kanchana Ranasinghe, Muzammal Naseer, and Ranga Rodrigo. Test-time optimization for domain adaptive open vocabulary segmentation. arXiv preprint arXiv:2501.04696, 2025. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 703 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 107, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 149 + ], + "type": "text", + "content": "[16] Reza Esfandiarpoor, Cristina Menghini, and Stephen H Bach. If clip could talk: Understanding vision-language model representations through their preferred concept descriptions. In EMNLP, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 150, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 150, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 150, + 553, + 183 + ], + "type": "text", + "content": "[17] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results, 2012. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 184, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 184, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 184, + 553, + 205 + ], + "type": "text", + "content": "[18] Yarin Gal. Uncertainty in deep learning. PhD Thesis, University of Cambridge, 2016. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 206, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 206, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 317, + 206, + 553, + 249 + ], + "type": "text", + "content": "[19] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. IJCV, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 251, + 553, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 303 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 303 + ], + "type": "text", + "content": "[20] Yunhao Ge, Jie Ren, Andrew Gallagher, Yuxiao Wang, Ming-Hsuan Yang, Hartwig Adam, Laurent Itti, Balaji Lakshminarayanan, and Jiaping Zhao. Improving zero-shot generalization and robustness of multi-modal models. In CVPR, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 304, + 553, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 553, + 337 + ], + "type": "text", + "content": "[21] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In ECCV, 2022. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 338, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 370 + ], + "type": "text", + "content": "[22] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In ICLR, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 372, + 553, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 404 + ], + "type": "text", + "content": "[23] Sina Hajimiri, Ismail Ben Ayed, and Jose Dolz. Pay attention to your neighbours: Training-free open-vocabulary semantic segmentation. In WACV, 2025. 1, 2, 3, 5, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 404, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 437 + ], + "type": "text", + "content": "[24] Cong Han, Yujie Zhong, Dengjie Li, Kai Han, and Lin Ma. Open-vocabulary semantic segmentation with decoupled one-pass network. In ICCV, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 438, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 553, + 480 + ], + "type": "text", + "content": "[25] Dapeng Hu, Jian Liang, Jun Hao Liew, Chuhui Xue, Song Bai, and Xinchao Wang. Mixed samples as probes for unsupervised model selection in domain adaptation. In NeurIPS, 2023. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 482, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 553, + 525 + ], + "type": "text", + "content": "[26] Xuefeng Hu, Ke Zhang, Lu Xia, Albert Chen, Jiajia Luo, Yuyin Sun, Ken Wang, Nan Qiao, Xiao Zeng, Min Sun, et al. Reclip: Refine contrastive language image pre-training with source free domain adaptation. In WACV, 2024. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 526, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 553, + 558 + ], + "type": "text", + "content": "[27] Haiwen Huang, Songyou Peng, Dan Zhang, and Andreas Geiger. Renovating names in open-vocabulary segmentation benchmarks. In NeurIPS, 2024. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 559, + 553, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 591 + ], + "type": "text", + "content": "[28] Tony Huang, Jack Chu, and Fangyun Wei. Unsupervised prompt learning for vision-language models. arXiv preprint arXiv:2204.03649, 2022. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 592, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 592, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 592, + 553, + 624 + ], + "type": "text", + "content": "[29] Yunshi Huang, Fereshteh Shakeri, Jose Dolz, Malik Boudiaf, Houda Bahig, and Ismail Ben Ayed. Lp++: A surprisingly strong linear probe for few-shot clip. In CVPR, 2024. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 625, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 658 + ], + "type": "text", + "content": "[30] Eyke Hüllermeier and Willem Waegeman. *Aleatoric and epistemic uncertainty in machine learning: An introduction to concepts and methods.* ML, 2021. 4" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 658, + 553, + 703 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 703 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 703 + ], + "type": "text", + "content": "[31] Krishna Murthy Jatavallabhula, Alihusein Kuwajerwala, Qiao Gu, Mohd Omama, Tao Chen, Shuang Li, Ganesh Iyer, Soroush Saryazdi, Nikhil Keetha, Ayush Tewari, et al. ConceptFusion: Open-set multimodal 3d mapping. In RSS, 2023." + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21479" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "text", + "content": "[32] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? In NeurIPS, 2017. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "type": "text", + "content": "[33] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In CVPR, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "type": "text", + "content": "[34] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. In NeurIPS, 2017. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "type": "text", + "content": "[35] Mengcheng Lan, Chaofeng Chen, Yiping Ke, Xinjiang Wang, Litong Feng, and Wayne Zhang. Proxyclip: Proxy attention improves clip for open-vocabulary segmentation. In ECCV, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 251 + ], + "type": "text", + "content": "[36] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. In ICLR, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 254, + 294, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 286 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 286 + ], + "type": "text", + "content": "[37] Yi Li, Hualiang Wang, Yiqun Duan, Jiheng Zhang, and Xiaomeng Li. A closer look at the explainability of contrastive language-image pre-training. PR, 2025. 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 288, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 288, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 288, + 294, + 330 + ], + "type": "text", + "content": "[38] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In CVPR, 2023. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "type": "text", + "content": "[39] Yuqi Lin, Minghao Chen, Wenxiao Wang, Boxi Wu, Ke Li, Binbin Lin, Haifeng Liu, and Xiaofei He. Clip is also an efficient segmenter: A text-driven approach for weakly supervised semantic segmentation. In CVPR, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 377, + 294, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 294, + 399 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 294, + 399 + ], + "type": "text", + "content": "[40] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. In ICLR, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 400, + 294, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 400, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 400, + 294, + 443 + ], + "type": "text", + "content": "[41] Matthias Minderer, Josip Djolonga, Rob Romijnders, Frances Hubis, Xiaohua Zhai, Neil Houlsby, Dustin Tran, and Mario Lucic. Revisiting the calibration of modern neural networks. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 445, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 487 + ], + "type": "text", + "content": "[42] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The role of context for object detection and semantic segmentation in the wild. In CVPR, 2014. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 490, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 294, + 533 + ], + "type": "text", + "content": "[43] Jishnu Mukhoti, Tsung-Yu Lin, Omid Poursaeed, Rui Wang, Ashish Shah, Philip HS Torr, and Ser-Nam Lim. Open vocabulary semantic segmentation with patch aligned contrastive learning. In CVPR, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 535, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 294, + 567 + ], + "type": "text", + "content": "[44] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In ICCV, 2017. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 568, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 600 + ], + "type": "text", + "content": "[45] Sarah Parisot, Yongxin Yang, and Steven McDonagh. Learning to name classes for vision and language models. In CVPR, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 602, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 635 + ], + "type": "text", + "content": "[46] Sarah Pratt, Ian Covert, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. In ICCV, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 636, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 689 + ], + "type": "text", + "content": "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[48] Kanchana Ranasinghe, Brandon McKinzie, Sachin Ravi, Yinfei Yang, Alexander Toshev, and Jonathon Shlens. Per" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "text", + "content": "ceptual grouping in contrastive vision-language models. In ICCV, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[49] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In CVPR, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "text", + "content": "[50] Karsten Roth, Jae Myung Kim, A Koepke, Oriol Vinyals, Cordelia Schmid, and Zeynep Akata. Waffling around for performance: Visual classification with random words and broad concepts. In ICCV, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 186, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 217 + ], + "type": "text", + "content": "[51] Ohad Rubin, Jonathan Herzig, and Jonathan Berant. Learning to retrieve prompts for in-context learning. In *NAACL-HLT*, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 220, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 251 + ], + "type": "text", + "content": "[52] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding. In CVPR, 2021. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 253, + 553, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 284 + ], + "type": "text", + "content": "[53] Gyungin Shin, Weidi Xie, and Samuel Albanie. Reco: Retrieve and co-segment for zero-shot transfer. In NeurIPS, 2022. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 287, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 330 + ], + "type": "text", + "content": "[54] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. In NeurIPS, 2022. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 332, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 364 + ], + "type": "text", + "content": "[55] Julio Silva-Rodriguez, Sina Hajimiri, Ismail Ben Ayed, and Jose Dolz. A closer look at the few-shot adaptation of large vision-language models. In CVPR, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 365, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 397 + ], + "type": "text", + "content": "[56] Vladan Stojnic, Yannis Kalantidis, and Giorgos Tolias. Label propagation for zero-shot classification with vision-language models. In CVPR, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 399, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 430 + ], + "type": "text", + "content": "[57] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In ICCV, 2021. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 433, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 553, + 465 + ], + "type": "text", + "content": "[58] Weijie Tu, Weijian Deng, and Tom Gedeon. A closer look at the robustness of contrastive language-image pre-training (clip). In NeurIPS, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 468, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 468, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 468, + 553, + 499 + ], + "type": "text", + "content": "[59] Weijie Tu, Weijian Deng, Dylan Campbell, Stephen Gould, and Tom Gedeon. An empirical study into what matters for calibrating vision-language models. In ICML, 2024. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 501, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 553, + 544 + ], + "type": "text", + "content": "[60] Antonin Vobecky, Oriane Simeoni, David Hurych, Spyridon Gidaris, Andrei Bursuc, Patrick Pérez, and Josef Sivic. Pop-3d: Open-vocabulary 3d occupancy prediction from images. In NeurIPS, 2023. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 545, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 588 + ], + "type": "text", + "content": "[61] Tuan-Hung Vu, Himalaya Jain, Maxime Bucher, Matthieu Cord, and Patrick Pérez. Advent: Adversarial entropy minimization for domain adaptation in semantic segmentation. In CVPR, 2019. 4" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 590, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 623 + ], + "type": "text", + "content": "[62] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In ICLR, 2021. 4" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 624, + 553, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 655 + ], + "type": "text", + "content": "[63] Feng Wang, Jieru Mei, and Alan Yuille. Sclip: Rethinking self-attention for dense vision-language inference. In ECCV, 2024. 1, 5, 8" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "type": "text", + "content": "[64] Zhengbo Wang, Jian Liang, Lijun Sheng, Ran He, Zilei Wang, and Tieniu Tan. A hard-to-beat baseline for training-free CLIP-based adaptation. In ICLR, 2024. 2" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[65] Monika Wysoczańska, Michael Ramamonjisoa, Tomasz Trzeciński, and Oriane Simeoni. Clip-diy: Clip dense infer" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21480" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "ence yields open-vocabulary semantic segmentation for-free. In WACV, 2024. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 148 + ], + "type": "text", + "content": "[66] Monika Wysoczanska, Oriane Simeoni, Michael Ramamonjisoa, Andrei Bursuc, Tomasz Trzciński, and Patrick Pérez. Clip-dinoiser: Teaching clip a few dino tricks for open-vocabulary semantic segmentation. In ECCV, 2024. 1, 2, 3, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 193 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 193 + ], + "type": "text", + "content": "[67] Yongqin Xian, Subhabrata Choudhury, Yang He, Bernt Schiele, and Zeynep Akata. Semantic projection network for zero-and few-label semantic segmentation. In CVPR, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 194, + 294, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 294, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 294, + 227 + ], + "type": "text", + "content": "[68] Bin Xie, Jiale Cao, Jin Xie, Fahad Shahbaz Khan, and Yanwei Pang. Sed: A simple encoder-decoder for open-vocabulary semantic segmentation. In CVPR, 2024. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 228, + 294, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 270 + ], + "type": "text", + "content": "[69] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. In NeurIPS, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 272, + 294, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 315 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 315 + ], + "type": "text", + "content": "[70] Renchunzi Xie, Ambroise Odonnat, Vasilii Feofanov, Weijian Deng, Jianfeng Zhang, and Bo An. Mano: Exploiting matrix norm for unsupervised accuracy estimation under distribution shifts. In NeurIPS, 2024. 4, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 316, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 358 + ], + "type": "text", + "content": "[71] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. Groupvit: Semantic segmentation emerges from text supervision. In CVPR, 2022. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 360, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 294, + 403 + ], + "type": "text", + "content": "[72] Jilan Xu, Junlin Hou, Yuejie Zhang, Rui Feng, Yi Wang, Yu Qiao, and Weidi Xie. Learning open-vocabulary semantic segmentation models from natural language supervision. In CVPR, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 404, + 294, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 447 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 447 + ], + "type": "text", + "content": "[73] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In CVPR, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 449, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 294, + 480 + ], + "type": "text", + "content": "[74] Mengde Xu, Zheng Zhang, Fangyun Wei, Han Hu, and Xi-ang Bai. Side adapter network for open-vocabulary semantic segmentation. In CVPR, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 482, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 482, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 294, + 525 + ], + "type": "text", + "content": "[75] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darryll. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2020. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 526, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 294, + 568 + ], + "type": "text", + "content": "[76] Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang-Chieh Chen. Convolutions die hard: Open-vocabulary segmentation with single frozen convolutional clip. In NeurIPS, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 570, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 294, + 613 + ], + "type": "text", + "content": "[77] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. In ECCV, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 614, + 294, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 655 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 655 + ], + "type": "text", + "content": "[78] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. IJCV, 2019. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 658, + 294, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 680 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 680 + ], + "type": "text", + "content": "[79] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In ECCV, 2022. 1, 2, 3, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "text", + "content": "[80] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In CVPR, 2022. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 184 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 105 + ], + "type": "text", + "content": "[81] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 107, + 553, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 138 + ], + "type": "text", + "content": "[82] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. In ICCV, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "text", + "content": "[83] Xiangyang Zhu, Renrui Zhang, Bowei He, Aojun Zhou, Dong Wang, Bin Zhao, and Peng Gao. Not all features matter: Enhancing few-shot clip with adaptive prior refinement. In ICCV, 2023. 2" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21481" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/FLSeg_ Enhancing Privacy and Robustness in Federated Learning under Heterogeneous Data via Model Segmentation/a4f90945-afa0-4f01-8edd-1b86f5d9d098_content_list.json b/2025/FLSeg_ Enhancing Privacy and Robustness in Federated Learning under Heterogeneous Data via Model Segmentation/a4f90945-afa0-4f01-8edd-1b86f5d9d098_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7ddbf786415d207ffde0865b735f5c1052c90220 --- /dev/null +++ b/2025/FLSeg_ Enhancing Privacy and Robustness in Federated Learning under Heterogeneous Data via Model Segmentation/a4f90945-afa0-4f01-8edd-1b86f5d9d098_content_list.json @@ -0,0 +1,1748 @@ +[ + { + "type": "text", + "text": "FLSeg: Enhancing Privacy and Robustness in Federated Learning under Heterogeneous Data via Model Segmentation", + "text_level": 1, + "bbox": [ + 132, + 130, + 867, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zichun $\\mathbf{Su}^{1\\dagger}$ Zhi Lu $^{1\\dagger}$ Yutong Wu $^{1}$ Renfei Shen $^{1}$ Songfeng Lu $^{1,2*}$", + "bbox": [ + 207, + 202, + 785, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Cyber Science and Engineering, Huazhong University of Science and Technology $^{2}$ Shenzhen Huazhong University of Science and Technology Research Institute", + "bbox": [ + 133, + 220, + 864, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "suzc25@m.fudan.edu.cn, {luzhi,wyt,shenrenfei,lusongfeng}@hust.edu.cn", + "bbox": [ + 192, + 258, + 792, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 309, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Federated Learning (FL) enables collaborative training of a global model without data sharing, yet it faces critical challenges from privacy leakage and Byzantine attacks. Existing privacy-preserving robust FL frameworks suffer from three key limitations: high computational costs, restricted application of Robust Aggregation Rules (RAR), and inadequate handling of data heterogeneity. To address these limitations, we propose the FLSeg framework, which leverages Segment Exchange and Segment Aggregation to avoid excessive encryption computations while enabling unrestricted use of any RAR. Additionally, a regularization term in local training balances personalization with global model performance, effectively adapting to heterogeneous data. Our theoretical and experimental analyses demonstrate FLSeg's semi-honest security and computational efficiency: it achieves client and server time complexities of $O(\\ell)$ and $O(n\\ell)$ , with empirical results showing significantly reduced computational times. Extensive experiments further confirm FLSeg's robustness across diverse heterogeneous and adversarial scenarios.", + "bbox": [ + 88, + 339, + 485, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 671, + 220, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Federated Learning (FL) allows organizations to collaboratively train models without sharing sensitive data. By exchanging model updates instead of raw data, FL enhances privacy and efficiency, making it ideal for fields like finance [18], healthcare [16], and smart cities [31]. However, FL is currently threatened by two major issues: privacy leakage and Byzantine attacks. Privacy leakage occurs when the server can infer clients' sensitive training data by analyzing the model updates [13, 43]. Byzantine attacks, on the other hand, involve malicious clients submitting incorrect or manipulated updates, which can disrupt or compromise the global model's performance [23, 34]. To address", + "bbox": [ + 89, + 696, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "these challenges, corresponding techniques have been developed. For privacy protection, methods like Homomorphic Encryption (HE) [29], Secure Multi-Party Computation (SMPC) [5], and Differential Privacy (DP) [35, 39] aim to safeguard training data by ensuring that only encrypted or obfuscated information is shared. For robustness against Byzantine attacks, various Robust Aggregation Rule (RAR), such as Krum [4], DnC [33], and RFA [32], help identify and exclude malicious updates, thereby protecting the global model from adversarial influence.", + "bbox": [ + 511, + 310, + 905, + 460 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, achieving both robustness and privacy simultaneously presents a significant challenge, as privacy-preserving methods can unintentionally protect malicious updates from detection [27]. To address this issue, a series of Privacy-preserving and Robust FL (PRFL) frameworks have been proposed. These methods explore various architectures and protocols or select encryption schemes that are compatible with specific RAR [2, 25, 28, 29, 38].", + "bbox": [ + 511, + 460, + 905, + 582 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite these methods balancing robustness and privacy, they still have notable limitations: (1) Privacy-preserving techniques such as HE and SMPC introduce significant computational overhead [27]. (2) These methods impose restrictions on the use of RAR, supporting only simple, single RAR, which are vulnerable to adaptive attacks such as Min-Max [33] and IPM [40]. (3) Gradient heterogeneity caused by data heterogeneity significantly degrades both model performance and the accuracy of RAR [22]. However, existing PRFL methods largely overlook this issue.", + "bbox": [ + 511, + 583, + 905, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Therefore, we aim to address the following question: \"How can we efficiently preserve privacy while avoiding restrictions on RAR implementation and mitigating the impact of data heterogeneity on robustness and performance?\" To this end, we draw inspiration from existing theoretical insights: reducing data dimensionality can alleviate the curse of dimensionality and gradient heterogeneity, thereby enhancing robustness [26, 33]. Based on this principle, we propose FLSeg, a model segmentation-based framework. We design a Segment Exchange privacy protocol to achieve privacy preservation without incurring excessive computa", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author, † Equal contribution", + "bbox": [ + 114, + 886, + 346, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3916", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tional overhead. Additionally, we introduce a Segment Aggregation algorithm, which applies RAR operations at the low-dimensional model segments to mitigate gradient heterogeneity and the curse of dimensionality while supporting arbitrary RAR. Furthermore, we incorporate a regularization term into local training to facilitate the learning of personalized models, further improving performance under data heterogeneity. Our contributions are threefold:", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A Novel Privacy-Preserving Robust FL Framework. To the best of our knowledge, FLSeg is the first PRFL framework that does not restrict the choice of RAR, while also mitigating the impact of heterogeneous data.", + "- Computationally Efficient Privacy Protection. We theoretically demonstrate that our FLSeg satisfies semihonest security. We further provide a detailed analysis of FLSeg's computational time complexity, with the client-side and server-side complexities being $O(\\ell)$ and $O(n\\ell)$ , respectively. In practical tests, FLSeg's client computation time is only 233 ms, compared to 1696 ms for SOTA method ACORN [2]. The server computation time per client is only 78 ms, whereas ACORN require 181 ms per client.", + "- Extensive Experimental Analysis. Extensive experiments on three benchmark datasets show that our FLSeg can significantly improve the performance under varying degrees of data heterogeneity and in Byzantine attack, e.g., the accuracy of ACORN is $48.21\\%$ , while FLSeg achieves $64.59\\%$ under a $20\\%$ proportion of Min-Max attackers and on non-IID CIFAR-10 datasets. Further inference attack experiments also verified the privacy protection capability of FLSeg." + ], + "bbox": [ + 89, + 214, + 480, + 560 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 573, + 238, + 588 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Privacy Threats and Defences. Existing studies have demonstrated that gradient sharing in FL poses severe privacy risks. DLG attacks [42, 43] and Inverting Gradients attacks [13] can precisely reconstruct raw training data from client-uploaded gradients.", + "bbox": [ + 89, + 598, + 480, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To mitigate these risks, privacy-preserving techniques such as DP [37], HE [25], and SMPC [5] have been proposed to obfuscate or encrypt gradients. However, these approaches have notable limitations: DP inevitably reduce model utility due to noise injection, while HE and SMPC incur high computational overhead due to complex encryption structures and protocols [1].", + "bbox": [ + 89, + 674, + 480, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Byzantine Attack and Defence. Byzantine attacks are initiated by malicious clients who manipulate or send adversarial model updates to degrade model performance [23]. A common type of such attacks is data poisoning, such as LabelFlipping [10], which manipulates training data to mislead the model into learning incorrect associations. Additionally, adaptive attacks, such as Min-Max [33] and IPM [40], optimize the magnitude and direction of malicious", + "bbox": [ + 89, + 780, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "gradient perturbations to maximize the disruption of the global FL model while simultaneously evading detection.", + "bbox": [ + 511, + 90, + 903, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To mitigate these risks, RAR modify the standard mean aggregation method [30] by performing statistical or feature analysis on the collected updates to filter out malicious ones. Examples include methods based on Euclidean distance, such as Krum [4] and Bulyan [15]; those leveraging coordinate values, such as RFA [32] and Median [41]; and approaches utilizing spectral anomaly detection, such as DnC [33]. Beyond RAR, Robust FL integrate auxiliary techniques for enhanced defense. FLTrust [8] uses a root dataset to assess the quality of client updates, ClippedClustering [22] leverages historical gradient information to constrain and clip gradients before aggregation, and GAS [26] proposes a gradient splitting technique to reduce data dimensionality and enhance the robustness of RAR. However, these defenses do not secure the gradients during transmission, rendering them vulnerable to privacy threats.", + "bbox": [ + 511, + 121, + 903, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Privacy-preserving Robust FL. PRFL integrates privacy-preserving techniques with RAR through specially designed protocols. PEFL [25] and ShieldFL [29] employ HE to ensure privacy while using cosine similarity to detect malicious updates. FFL [17] proposes a parameter mixing method to protect privacy while applying cosine similarity to the mixed model for robustness. ACORN [2] and RoFL [28] utilize the SecAgg SMPC protocol to safeguard the privacy of inputs while employing zero-knowledge proofs to bound the norm of model updates, thereby mitigating the impact of malicious updates. However, the high computational cost, restricted support for RAR, and lack of consideration for data heterogeneity limit their applicability.", + "bbox": [ + 511, + 363, + 903, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 513, + 571, + 651, + 585 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Notion. For any positive integer $n \\in \\mathbb{N}^+$ , we denote the set $\\{1, \\ldots, n\\}$ by $[n]$ . The cardinality of a set $w$ is denoted by $|w|$ . We denote the $l_2$ norm of vector $w$ by $\\| w \\|$ .", + "bbox": [ + 511, + 595, + 903, + 642 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Model Segmentation. For a model $w$ , the mode of segmentation refers to a partitioning of the set of model parameters, $\\{\\gamma^{(1)},\\gamma^{(2)},\\dots,\\gamma^{(d)}\\}$ , where different segments are disjoint, and $w = \\bigcup_{k = 1}^{d}\\gamma^{(k)}$ .", + "bbox": [ + 511, + 642, + 903, + 705 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "System Setup. We consider a classic FL scenario [30], involving $n$ clients, denoted by $\\{c_i\\}_{i \\in [n]}$ , and a single server $S$ . $c_i$ has a unique local dataset $D_i$ that is non-IID. $S$ starts with initial model weights $w_0$ and distributes the model to each $c_i$ . $c_i$ train the model on their datasets and send their updates $w_i^t$ to the server $S$ , which aggregates these updates and proceeds to the next round. After $T$ rounds, this process results in a global model $w^T$ .", + "bbox": [ + 511, + 704, + 903, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Threat Model. Our system faces two primary threats. First, in line with [8, 11], there may be Byzantine clients within the system who manipulate their uploaded model to disrupt training. They can launch targeted attacks such as LabelFlipping [10], and powerful adaptive attacks like Min", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3917", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/36b96f68677cefb5702ab0523844da931eaec0dc81188ae03e627f21df1580fd.jpg", + "image_caption": [ + "Figure 1. The Overview of FLSeg." + ], + "image_footnote": [], + "bbox": [ + 94, + 87, + 903, + 338 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Max [33]. Second, the server is semi-honest, adhering to the aggregation protocol while attempting to infer private local data from the client's updates [29]. As in prior work [2, 25, 27, 29], we assume that the adversary can control a number of clients $f < (n - 1) / 2$ and semi-honest server and Byzantine clients do not collude.", + "bbox": [ + 89, + 390, + 483, + 481 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 89, + 500, + 225, + 517 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Intuition", + "text_level": 1, + "bbox": [ + 89, + 527, + 194, + 541 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing RAR methods commonly suffer from the curse of dimensionality [15], where malicious gradients can hide within high-dimensional vectors. Additionally, data heterogeneity introduces gradient heterogeneity [19], further increasing the difficulty of detection. Studies have shown that reducing input dimensionality can help alleviate this issue, for example, by processing gradients through sub-sampling [33], grouping [20], or splitting [26].", + "bbox": [ + 89, + 551, + 483, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Based on this insight, we propose an intuitive approach: by fixing the mode of segmentation, each client selects a partner to exchange model segments, while the server applies the same segmentation mode during aggregation. In this way, as long as the server is unaware of the specific exchange pairs or the exchanged segments, it cannot reconstruct the client models, thereby ensuring privacy protection. Additionally, the server apply RAR on each segment, which not only reduces the curse of dimensionality and gradient heterogeneity but also ensures that each segment to which the RAR is applied originates from different clients, thereby ensuring the correctness of the RAR implementation. Besides, although these processes mitigate gradient heterogeneity caused by Non-IID data [15, 20], they fail to address the performance degradation induced by Non-", + "bbox": [ + 89, + 674, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "IID data. To tackle this issue, we incorporate local regularization techniques from personalized FL [24, 36] to improve model performance without altering the overall process. The whole process is illustrated in Fig. 1.", + "bbox": [ + 511, + 390, + 905, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.2. Non-IID Data: Learn Personalized Model", + "text_level": 1, + "bbox": [ + 511, + 458, + 870, + 472 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Heterogeneous data can cause the global model to fail to adapt to all clients' data distributions, leading to performance degradation. Existing personalized FL methods address this issue by balancing the global and local models to integrate the generalization ability of the global model with the uniqueness of local data [9, 24, 36]. Inspired by this approach, we modify each client's local training process by allowing the client to maintain a personalized model $v_{i}$ , with its objective function defined in Eq. (1).", + "bbox": [ + 511, + 479, + 906, + 617 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {v _ {i}} h _ {i} \\left(v _ {i}, w ^ {t}\\right) = \\mathcal {L} _ {i} \\left(v _ {i}\\right) + \\frac {\\lambda}{2} \\left\\| v _ {i} - w ^ {t} \\right\\| ^ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 627, + 906, + 656 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $w^{t}$ represents the global model received in each round; $\\mathcal{L}_i(w) = \\mathbb{E}_{\\xi_i \\sim \\mathcal{D}_i}[f_i(w; \\xi_i)]$ denotes the loss function on $c_{i}$ 's local data; $\\lambda$ is a regularization parameter that balances the influence of the personalized and global models. When $\\lambda \\to +\\infty$ , the personalized model converges to the global model, whereas $\\lambda \\to 0$ leads to a more distinct personalized effect.", + "bbox": [ + 511, + 659, + 905, + 765 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The detailed process is provided in Algorithm 1. During training, the personalized model is updated synchronously with the global model. Each client retains $v_{i}^{t + 1}$ locally for local tasks and prepares the model update $w_{i}^{t}$ for the subsequent Segment Exchange.", + "bbox": [ + 511, + 766, + 905, + 842 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.3. Privacy: Segment Exchange", + "text_level": 1, + "bbox": [ + 511, + 848, + 766, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To prevent the server from obtaining each client's complete model, the client mixes its model by exchanging segments", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3918", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Algorithm 1: Local Training", + "text_level": 1, + "bbox": [ + 102, + 95, + 295, + 109 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Input : Personalized model $v_{i}^{t}$ , model received from server $w^{t}$ , regularization coefficient $\\lambda$ , learning rate $\\eta$ , local epochs $E$", + "bbox": [ + 109, + 112, + 442, + 157 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Output: Updates $w_{i}^{t + 1}$ , personalized model $v_{i}^{t + 1}$", + "bbox": [ + 111, + 157, + 434, + 174 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1 $w_{i}^{t + 1}\\gets w^{t},v_{i}^{t + 1}\\leftarrow v_{i}^{t};$", + "2 for epoch $\\leftarrow 0$ to $E - 1$ do", + "3 /*update personalized model \\*/ $v_{i}^{t + 1}\\gets v_{i}^{t + 1} - \\eta (\\nabla \\mathcal{L}_{i}(v_{i}^{t + 1}) + \\lambda (v_{i}^{t + 1} - w^{t}))$", + "4 $\\begin{array}{r}\\left\\lfloor \\begin{array}{c}/*\\mathrm{update~global~model}*/\\\\ w_i^{t + 1}\\gets w_i^{t + 1} - \\eta \\nabla \\mathcal{L}_i(w_i^{t + 1}); \\end{array} \\right. \\end{array}$", + "5 return $w_{i}^{t + 1}, v_{i}^{t + 1}$" + ], + "bbox": [ + 98, + 172, + 457, + 290 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with a partner before submitting local updates. However, to ensure that clients do not reveal their models to their exchange partners during this process, we design a secure model segment exchange protocol, detailed in Algorithm 2.", + "bbox": [ + 89, + 323, + 482, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An simple example is illustrated in Fig. 2. The two parties first use the Diffie-Hellman key exchange to agree on a mask $m$ that determines which segments are exchanged. $w_{i}^{-m}$ is encrypted using client $c_{i}$ 's key $r_{i}$ , ensuring that $c_{j}$ cannot access $c_{i}$ 's private information while still enabling segment exchange through the properties of the XOR operation. Meanwhile, the server can decrypt the client key to obtain the mixed model. However, since it does not know $m$ or the exchange partner, it cannot recover the original client model, thereby preserving privacy.", + "bbox": [ + 89, + 383, + 483, + 536 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1c85a6a0161442e27f2aed6784eaae72fed3269303ba0b7ac9d0515c4d4ce0b9.jpg", + "image_caption": [ + "Figure 2. The Workflow of Segment Exchange." + ], + "image_footnote": [], + "bbox": [ + 104, + 547, + 483, + 690 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.4. Robustness: Segment Aggregation", + "text_level": 1, + "bbox": [ + 89, + 741, + 390, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To mitigate the curse of dimensionality, FLSeg partitions the mixed model according to the mode of segmentation and then applies RAR to each segment, rather than performing RAR directly on the entire mixed model. This approach not only ensures that each segment in the RAR originates from different clients but also effectively mitigates the challenges posed by high input dimensionality. The detailed procedure is provided in Algorithm 3.", + "bbox": [ + 89, + 763, + 482, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After grouping segments based on the mode of segmen", + "bbox": [ + 109, + 885, + 480, + 901 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 2: Segment Exchange Protocal" + ], + "code_body": "Input: Client $c_{i}$ 's model update $w_{i}$ ,server's public-private key pair $(pk_S,sk_S)$ ,the exchange partner $C_j$ \n1 client $c_{i}$ and $c_{j}$ \n2 generate client's key $r_i$ . \n3 perform Diffie-Hellman key exchange with $c_{j}$ to obtain mask $m$ . \n4 $w_{i}^{m} / w_{i}^{-m}\\gets$ Set the $k$ -th segment $\\gamma^{(k)} = 0$ of $w_{i}$ if the $k$ -th bit of $m$ is $1 / 0$ . \n5 exchange $w_{i}^{-m}\\oplus r_{i}$ and $\\mathrm{Enc}(r_i,pk_S)$ with $c_{j}$ . \n6 compute $(e_i)_{\\mathrm{mix}} = w_i^m\\oplus (w_j^{-m}\\oplus r_j)$ $= (w_{i})_{\\mathrm{mix}}\\oplus r_{j}$ . \n7 send $(e_i)_{\\mathrm{mix}}$ ,Enc $(r_j,pk_S)$ to server S; \n8 server S \n9 decrypts each client $c_{i}$ 's key $r_j = \\mathrm{Dec}(\\mathrm{Enc}(r_j,pk_S),sk_S)$ . \n10 decrypts each client $c_{i}$ 's mixed model $(w_{i})_{\\mathrm{mix}}\\leftarrow (e_{i})_{\\mathrm{mix}}\\oplus r_{j};$", + "bbox": [ + 516, + 111, + 903, + 401 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 3: Segment Aggregation" + ], + "code_body": "Input: uploaded mixed models $\\{(w_{i})_{\\mathrm{mix}}\\}_{i\\in [n]}$ the mode of segmentation $\\{\\gamma^{(1)},\\dots ,\\gamma^{(d)}\\}$ Output: the aggregation result $w$ \n1 for each model $(w_{i})_{\\mathrm{mix}}$ do \n2 split model into segments $\\{\\gamma_i^{(1)},\\ldots ,\\gamma_i^{(d)}\\} \\gets (w_i)_{\\mathrm{mix}};$ \n3 for each segment $\\gamma^{(k)}\\in \\{\\gamma^{(1)},\\dots ,\\gamma^{(d)}\\}$ do \n4 apply the RAR to the segments to establish a benchmark $\\overline{\\gamma}^{(k)}\\gets \\mathrm{RAR}(\\{\\gamma_i^{(k)}\\}_{i\\in [n]})$ \n5 calculate the similarity $s_i^{(k)}$ for each segment $\\gamma_{i}^{(k)}$ with the benchmark $\\overline{\\gamma}^{(k)}$ using Eq. (2); \n6 compute each model's similarity score as $s_i = \\sum_{k = 1}^d s_i^{(k)}$ and select the top $n - 2f$ models with the highest scores to form set $I$ \n7 aggregate $w\\gets \\sum_{i\\in I}(w_i)_{\\mathrm{mix}}$ \n8 return $w$", + "bbox": [ + 516, + 441, + 903, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tation, $S$ applies RAR to each segment group to obtain the benchmark $\\overline{\\gamma}^{(k)}$ and uses it as an honest reference to compute the similarity $s_i^{(k)}$ of each segment $\\gamma_i^{(k)}$ using Eq. (2), then sums the segment similarities to obtain the model's score $s_i$ for this round.", + "bbox": [ + 511, + 748, + 906, + 825 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns _ {i} ^ {(k)} = \\frac {\\left(\\gamma_ {i} ^ {(k)} - \\mu\\right) \\cdot \\left(\\bar {\\gamma} ^ {(k)} - \\mu\\right)}{\\| \\gamma_ {i} ^ {(k)} - \\mu \\| \\| (\\bar {\\gamma} ^ {(k)} - \\mu) \\|}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 835, + 903, + 877 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $z = |\\gamma^{(k)}|$ is the number of parameters of $\\gamma_{i}^{(k)}$ and", + "bbox": [ + 511, + 882, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3919", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\overline{\\gamma}^{(k)}$ , and $\\mu = \\frac{\\sum_{j = 1}^{z}\\gamma_{i,j}^{(k)} + \\sum_{j = 1}^{z}\\overline{\\gamma}_{j}^{(k)}}{2z}$ is the average value of all parameters in $\\gamma_i^{(k)}$ and $\\overline{\\gamma}^{(k)}$", + "bbox": [ + 89, + 87, + 483, + 130 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since, in the worst case, each honest client's partner may be a malicious client, $S$ select the top $n - 2f$ clients with the highest scores for aggregation.", + "bbox": [ + 89, + 130, + 483, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.5. Summary", + "text_level": 1, + "bbox": [ + 89, + 186, + 202, + 203 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The mode of segmentation in FLSeg remains unchanged throughout the process. Before training begins, the server distributes its public key $pk_{S}$ to all clients. In each training round, clients first perform Local Training (Algorithm 1), then conduct Segment Exchange to obtain the mixed model (Algorithm 2), and finally, the server applies Segment Aggregation (Algorithm 3) to the received global models. The overall process is detailed in the algorithm provided in the Appendix A.", + "bbox": [ + 89, + 209, + 483, + 345 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "FLSeg does not restrict how clients select partners; however, to prevent honest clients from continuously collaborating with malicious clients, we design a reputation system based on computed similarity scores. Specifically, each client $c_{i}$ initializes its reputation list $R_{i} = \\{r_{1} = 0,\\dots ,r_{n} = 0\\}$ . After each round of aggregation, the server normalizes the similarity scores of each model using Eq. (3) and sends these scores to the corresponding clients. A lower score for an honest client indicates that its model may have been influenced by malicious updates.", + "bbox": [ + 89, + 345, + 483, + 498 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns _ {i} ^ {\\prime} = 2 \\left(\\frac {s _ {i} - \\operatorname* {m i n} \\left(\\left\\{s _ {i} \\right\\} _ {i \\in [ n ]}\\right)}{\\operatorname* {m a x} \\left(\\left\\{s _ {i} \\right\\} _ {i \\in [ n ]}\\right) - \\operatorname* {m i n} \\left(\\left\\{s _ {i} \\right\\} _ {i \\in [ n ]}\\right)}\\right) - 1 \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 521, + 482, + 558 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Clients use these scores to update the reputation values of their partners as $r_j = r_j + s_i'$ . In subsequent rounds, clients can select partners with higher reputation values from their reputation lists for Segment Exchange. Practical test results for the reputation system are provided in Appendix B.", + "bbox": [ + 89, + 569, + 483, + 659 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Analysis", + "text_level": 1, + "bbox": [ + 89, + 674, + 187, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Semi-Honest Security", + "text_level": 1, + "bbox": [ + 89, + 699, + 294, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use the Universally Composable Security Framework [7] to prove that FLSeg satisfies semi-honest security, meaning that FLSeg does not leak any information as long as the adversaries follow the protocol. We first present definition of semi-honest security [14]:", + "bbox": [ + 89, + 722, + 483, + 799 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Definition 1 (Real-World). For protocol $\\pi$ , there exists a semi-honest adversary $\\mathcal{A}$ that can control a subset of participants $\\mathcal{C}$ . $\\mathcal{A}$ will follow the protocol $\\pi$ but will attempt to infer the private inputs $\\{x_1, \\ldots, x_n\\}$ based on the available information. The probability distribution of the inference results is denoted as $\\text{Real}_{\\pi, \\mathcal{A}}(\\mathcal{P}, \\{x_1, \\ldots, x_n\\})$ .", + "bbox": [ + 89, + 809, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Definition 2 (Ideal-World). Assume there exists an ideal functionality $\\mathcal{F}$ that can execute any interaction in protocol $\\pi$ with absolute security. There also exists a simulator Sim that simulates interactions $\\mathcal{F}$ in the same manner as the real-world adversary $\\mathcal{A}$ . The probability distribution of the inference results is denoted as Ideal $_{\\mathcal{F},\\text{Sim}}(\\mathcal{P},\\{x_1,\\ldots,x_n\\})$ .", + "bbox": [ + 511, + 90, + 905, + 183 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Definition 3 (Semi-Honest). If $\\text{Real}_{\\pi, \\mathcal{A}}(\\mathcal{P}, \\{x_1, \\ldots, x_n\\})$ and $\\text{Ideal}_{\\mathcal{F}, \\text{Sim}}(\\mathcal{P}, \\{x_1, \\ldots, x_n\\})$ are computationally indistinguishable to an honest-but-curious adversary $\\mathcal{A}$ , then protocol $\\pi$ is considered semi-honest secure.", + "bbox": [ + 511, + 186, + 905, + 247 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Theorem 1. Assuming the $FLSeg$ protocol is denoted as $\\pi$ , the participants of the protocol $\\mathcal{P} = \\{c_1, \\ldots, c_n, S\\}$ include all clients and the server, and the private inputs are the gradients of each honest client $\\{w_1, \\ldots, w_n\\}$ . Then $FLSeg$ is semi-honest secure.", + "bbox": [ + 511, + 253, + 905, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proof. We use the Hybrid Argument [12]. By gradually replacing components of the protocol's multiple interaction stages with the ideal functionality $\\mathcal{F}$ , we demonstrate that each stage's replacement does not significantly alter the probability distribution of the adversary's inference results. Assuming $\\mathcal{T}$ represents the entire parameter space of the model, we define the following three hybrids:", + "bbox": [ + 511, + 335, + 905, + 441 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Hybrid 0: The original FLSeg Framework.", + "- Hybrid 1: Similar to Hybrid 0, but replaces $w_{i}^{-m} \\oplus r_{i}$ exchanged by clients with random values $\\phi_{i}^{(1)} \\leftarrow \\mathcal{T}$ .", + "- Hybrid 2: Similar to Hybrid 1, but replaces $(w_{i})_{\\mathrm{mix}}$ transmitted to the server with random values $\\phi_{i}^{(2)} \\stackrel{\\S}{\\leftarrow} \\mathcal{T}$ ." + ], + "bbox": [ + 513, + 441, + 903, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Lemma 1 (Hybrid 1). Hybrid 0 and Hybrid 1 are computationally indistinguishable if FLSeg is semi-honest secure.", + "bbox": [ + 511, + 534, + 905, + 564 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proof. Since $r_i$ and $\\phi_i^{(1)}$ follow the same distribution, $w_i^{(\\neg m)} \\oplus r_i$ is also distributed identically to $\\phi_i^{(1)}$ . Thus, the advantage of $\\mathcal{A}$ is given by Eq. (4).", + "bbox": [ + 511, + 573, + 905, + 625 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left| \\mathbb {P} [ \\mathcal {A} (\\phi_ {i} ^ {(1)}) = 1 ] - \\mathbb {P} [ \\mathcal {A} (w _ {i} ^ {- m} \\oplus r _ {i}) = 1 ] \\right| \\leq \\operatorname {n e g} (d) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 628, + 905, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{neg}(d)$ is a negligible function. Hence, Hybrid 0 and Hybrid 1 are computationally indistinguishable.", + "bbox": [ + 511, + 661, + 905, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Lemma 2 (Hybrid 2). Hybrid 1 and Hybrid 2 are computationally indistinguishable if FLSeg is semi-honest secure.", + "bbox": [ + 511, + 700, + 905, + 731 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proof. Since the adversary $\\mathcal{A}$ cannot control both the client and the server simultaneously, $j$ and $m$ are random to $\\mathcal{S}$ . If $\\mathcal{S}$ attempts to recover the private $w_{i}$ by random matching, the probability of success is $\\frac{1}{n} \\cdot \\frac{1}{2^d}$ . Thus, the advantage of $\\mathcal{A}$ is given by Eq. (5).", + "bbox": [ + 511, + 739, + 905, + 816 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left| \\mathbb {P} [ \\mathcal {A} ((w _ {i}) _ {\\mathrm {m i x}}) = 1 ] - \\mathbb {P} [ \\mathcal {A} (\\phi_ {i} ^ {(2)}) = 1 ] \\right| \\leq \\frac {1}{n} \\cdot \\frac {1}{2 ^ {d}} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 821, + 903, + 851 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since this advantage is negligible for a polynomial-time adversary, it can be ignored. Thus, Hybrid 1 and Hybrid 2 are computationally indistinguishable.", + "bbox": [ + 511, + 854, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3920", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/deb1ad587ccc2dc4ab6d860926534d1b4b8631e46a7549baa6ca1ca1bb4d2628.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SecAgg [5]Bell et al. [3]ACORN-detect [2]ShieldFL [29]PEFL [25]FLSeg
ClientComplexityO(n2+ ℓn)O(log2 n + ℓ log n)O(log2 n + ℓ log n)O(ℓ(TEnc + TDec))O(ℓ(TEnc + TDec))O(ℓ)
Overhead2,557 ms1,696 ms1,729 ms231,541 ms215,667 ms233 ms
ServerComplexityO(ℓn2)O(n(log2 n + ℓ log n))O(n(log2 n + ℓ log n))O(ℓnTDec)O(ℓn(TEnc + TDec))O(ℓn)
Overhead192 ms/c182 ms/c181 ms/c78,685 ms/c790,032 ms/c78 ms/c
Robustness○○◇◇◇◇
", + "bbox": [ + 135, + 88, + 864, + 191 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison of computational time complexity and time overhead. $n$ represents the number of clients, $\\ell$ denotes the size of the model parameters. In calculating computation time overhead, let $n = {10}^{3}$ and $\\ell = {10}^{6}$ . ${T}_{\\text{Enc }}$ and ${T}_{\\text{Dec }}$ represent the encryption and decryption constant overhead in HE. ”/c” denotes the average time overhead incurred by the server for each client. $\\bigcirc$ indicates the absence of robustness, $\\mathbf{D}$ denotes support for fixed RAR only,and $\\bullet$ denotes support for arbitrary RAR. We omit overheads unrelated to privacy protection,such as training and aggregation.", + "bbox": [ + 88, + 200, + 906, + 272 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Through the above hybrid experiments, we have gradually proven the information transmitted from the real-world to the ideal-world is computationally indistinguishable from random values. Thus, we have proven Theorem 1. $\\square$", + "bbox": [ + 88, + 297, + 485, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Efficiency in privacy protection", + "text_level": 1, + "bbox": [ + 89, + 371, + 369, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Assume $n$ represents the number of clients and $\\ell$ denotes the size of the model parameters. The computational overhead on the client side for FLSeg primarily consists of the following components:", + "bbox": [ + 89, + 393, + 483, + 455 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$\\mathcal{O}(\\ell^2)\\colon m\\gets$ Diffie-Hellman key exchange,", + "- $\\mathcal{O}(\\ell): \\text{Calculation of } w_i^m, w_i^{-m}, w_i^{-m} \\oplus r_i, (e_i)_{\\text{mix}}$", + "$\\mathcal{O}(\\ell):\\mathrm{Enc}(r_j,pk_{\\mathcal{S}})$" + ], + "bbox": [ + 89, + 457, + 444, + 505 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, we introduce an optimization technique: instead of directly using Diffie-Hellman to generate $m$ or encrypting $r_j$ , which incurs significant computational time costs, we generate respective random seeds $\\mathsf{seed}_{r_j}$ and $\\mathsf{seed}_m$ . Both clients then use a Pseudo-Random Number Generator (PRNG) [5] to produce $r_j$ and $m$ , and only the encrypted seed $\\mathrm{Enc}(\\mathsf{seed}_{r_j},pk_S)$ is sent. This approach reduces the overhead of Diffie-Hellman and asymmetric encryption to a constant level, depending solely on the seed length. Thus, the primary computational overhead is shifted to generating $r_j$ using PRNG $(\\mathsf{seed}_{r_j}) = r_j$ and performing XOR operations, with a complexity of $\\mathcal{O}(\\ell)$ . Similarly, the decryption overhead on the server is minimized, focusing mainly on decrypting the mixed model $(w_i)_{\\mathrm{mix}}$ using XOR operations, resulting in a complexity of $\\mathcal{O}(\\ell n)$ .", + "bbox": [ + 88, + 507, + 482, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To demonstrate the advantages of FLSeg in efficient privacy protection, we compare the computational time complexity and time overhead of FLSeg with SOTA privacy-preserving FL that also satisfy semi-honest security, including SecAgg [5], Bell et al. [3], ShieldFL [29], RoFL [28], and PEFL [25]. The comparison results are presented in Tab. 1. The time complexity of ShieldFL and PEFL is asymptotically $O(\\ell)$ ; however, the Paillier HE they employ incurs substantial constant overheads, $T_{\\mathrm{Enc}}$ and $T_{\\mathrm{Dec}}$ . SecAgg, ACORN, and Bell et al. are aggregation schemes based on SMPC techniques, requiring the compu", + "bbox": [ + 88, + 734, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tation of pairwise masking for all other clients. To more intuitively demonstrate the advantages of FLSeg, we tested each method's runtime in a practical setting with $n = 10^3$ and $\\ell = 10^6$ . The results show that FLSeg achieves an $8 \\times$ improvement in client-side and a $2 \\times$ improvement in server-side computation compared to the fastest method. See Appendix C for the experimental setup.", + "bbox": [ + 511, + 297, + 906, + 405 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 511, + 416, + 645, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Setups", + "text_level": 1, + "bbox": [ + 511, + 440, + 602, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets and Models. Our experiments are conducted on three datasets—FEMNIST [6], CIFAR-10, and CIFAR-100 [21]—using a Multilayer Perceptron (MLP), Convolutional Neural Network (CNN), and SqueezeNet, respectively. For FEMNIST, the dataset is naturally non-IID and consists of 199 unique clients. For CIFAR-10 and CIFAR-100, we set the number of clients, $n = 100$ and $n = 50$ , and generate non-IID data distributions across clients using a Dirichlet distribution by following [24, 26, 36]. The concentration parameter $\\alpha$ of the Dirichlet distribution controls the degree of data heterogeneity, with lower values indicating higher heterogeneity. For each client, $90\\%$ of their data is allocated for training, while the remaining $10\\%$ is reserved for testing.", + "bbox": [ + 511, + 462, + 906, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluated attacks. We select the classic targeted attack, LabelFlipping (LP) [10]; two powerful adaptive attacks, Min-Max [33] and IPM [40]; and two privacy inference attacks, iDLG [42] and Inverting-Gradients [13].", + "bbox": [ + 511, + 659, + 905, + 719 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared Methods. We consider the FL baseline FedAvg [30]; three prominent RARs: Krum [4], DnC [33], and RFA [32]; three robust FL methods: GAS [26], FLTrust [8], and ClippedCluster [22]; and three privacy-preserving robust FL methods: ACORN [2], ShieldFL [29], and PEFL [25].", + "bbox": [ + 511, + 719, + 905, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation. We use average top-1 accuracy of each client to evaluate the performance of global models. We run each experiment for five times and report the mean and standard deviation of the highest accuracy during the training process.", + "bbox": [ + 511, + 810, + 905, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "All hyperparameter settings for the compared methods", + "bbox": [ + 531, + 885, + 905, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3921", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/839c2736f2e641be349dda54084e3b4f964785139e0c9c7f5a6d1e3b30c1ed00.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetAttackBaselineRARRobust FLPrivacy-preserving Robust FLOur
FedAvgKrumDnCRFAGASFLTrustClipped ClusteringACORNShieldFLPEFLFLSeg (Krum)FLSeg (DnC)
CIFAR-10-52.16±0.9753.86±0.9053.32±0.1353.52±0.4851.16±0.2853.28±0.7053.44±1.5949.08±1.9752.61±0.2746.31±0.8860.45±0.8263.99±0.53
LF47.46±0.6947.55±1.9246.50±1.5555.20±1.0647.96±1.9250.61±0.6150.85±1.4347.08±0.3348.37±0.8544.50±1.2359.62±0.7663.68±0.86
IPM49.03±0.1846.44±0.3351.82±0.4535.07±1.6243.55±1.6050.52±1.9550.93±1.5545.36±2.2149.53±1.8641.07±3.0259.43±1.0661.28±0.91
Min-Max35.89±1.5745.46±1.4950.62±1.4349.61±0.6049.69±1.7148.25±0.7049.36±1.9748.21±2.0338.51±3.0912.36±0.4860.82±0.5764.59±1.02
FEMNIST-70.17±1.8169.60±0.2369.95±0.8868.76±0.9369.95±0.4737.06±1.4669.31±1.1273.14±0.8173.33±0.9059.78±1.7177.12±0.6777.88±0.31
LF68.08±1.5870.37±0.4469.27±1.5668.70±0.7668.29±1.3133.88±0.6269.90±0.5169.78±0.3770.82±0.5452.68±1.9873.44±0.7776.94±0.45
IPM66.69±1.0659.88±1.9067.99±1.8455.02±0.8555.39±0.2838.81±1.0967.07±1.0168.44±2.0770.34±1.0350.37±1.6373.94±1.1574.95±0.87
Min-Max67.46±1.5265.17±1.8568.62±1.3865.38±1.3265.85±1.1236.70±0.2768.05±1.5171.29±1.0870.19±2.1121.32±0.2276.13±1.6277.74±0.65
CIFAR-100-35.92±0.6734.80±0.5434.48±0.3834.34±0.2425.64±0.9531.53±0.4233.32±0.1335.36±0.5634.70±0.752.86±0.4754.45±0.8153.55±0.59
LF30.83±0.3334.37±0.4933.46±0.2729.28±0.5826.98±0.3731.11±0.4432.99±0.4229.85±0.9132.53±0.613.21±0.9642.32±0.6543.79±0.49
IPM25.54±1.2119.29±0.7331.38±0.5315.61±0.6425.89±0.4529.43±0.2632.44±0.5925.55±0.4823.62±0.333.24±1.2841.56±0.7545.13±1.09
Min-Max22.31±1.0216.38±0.2934.23±0.6219.46±0.9327.21±0.3830.16±1.2534.73±0.6721.04±0.4334.39±0.473.20±0.4146.05±0.3451.11±0.21
", + "bbox": [ + 94, + 88, + 903, + 297 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Top-1 Accuracy(%) under Different Byzantine Attack.", + "bbox": [ + 308, + 308, + 684, + 321 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and attacks, along with detailed experimental configurations, are provided in Appendix D.", + "bbox": [ + 89, + 349, + 482, + 380 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2. Robustness", + "text_level": 1, + "bbox": [ + 89, + 388, + 215, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2 shows the accuracy under different Byzantine attackers, with data heterogeneity set at a Dirichlet concentration parameter of $\\alpha = 0.5$ on CIFAR-10 and CIFAR-100, and attackers comprising $20\\%$ of the total number of clients. FLSeg can utilize different RAR for Segment Aggregation; we denote the RAR used by adding it in parentheses after the method name, for example, FLSeg(DnC).", + "bbox": [ + 89, + 411, + 483, + 517 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1bbd762c6f67a253ac894703353fcf710ba92cf2a0cb3c207053f8954c63f8ec.jpg", + "image_caption": [ + "Figure 3. Top-1 Accuracy(%) under Different Ratio of Attacker on CIFAR-10." + ], + "image_footnote": [], + "bbox": [ + 94, + 529, + 483, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "From the empirical results, we observe that FLSeg maintains stable performance under various Byzantine attacks. Additionally, we find that FLSeg's defensive effectiveness is significantly influenced by the choice of RAR. For instance, using DnC, a singular value decomposition (SVD)-based spectral method can achieve higher accuracy than Euclidean distance-based methods like Krum. We also found that the combination of RAR with FLSeg significantly outperforms the original RAR method, especially in complex models, indicating that FLSeg effectively mitigates the curse of dimensionality. For example, Krum nearly fails under Min-Max attacks on CIFAR-100, achieving an accuracy of only $16.38\\%$ , but when combined with FLSeg, it achieves a much improved accuracy of $46.05\\%$ .", + "bbox": [ + 89, + 689, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 3 shows the accuracy under different numbers of Min-Max attackers, with FLSeg demonstrating greater stability against varying numbers of attackers. Additionally, it is evident that other privacy-preserving robust FL methods experience a notable performance drop as the number of attackers increases. This decline occurs because the metrics these methods use to filter malicious gradients are constrained by their privacy-preserving techniques, making them vulnerable to strong adaptive attacks. In contrast, FLSeg mitigates the curse of dimensionality through Segment Aggregation and enhances resistance to malicious gradients by integrating diverse RAR.", + "bbox": [ + 511, + 349, + 906, + 531 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.3. Heterogeneous Data", + "text_level": 1, + "bbox": [ + 511, + 542, + 705, + 559 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 4 shows the accuracy under different levels of data heterogeneity on CIFAR-10. We fix the ratio of attackers at $20\\%$ using Min-Max attackers and control the level of data heterogeneity by adjusting the Dirichlet $\\alpha$ . An $\\alpha$ of 0.1 represents an extreme case where each client may only possess one or two labels out of ten possible categories, while at $\\alpha = 10$ , the data distribution is nearly IID.", + "bbox": [ + 511, + 566, + 906, + 672 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3ebce31097df6781b87d72afdeda67326be4371ae087ff471861f6ef332455fa.jpg", + "image_caption": [ + "Figure 4. Top-1 Accuracy(%) under Different Levels of Data Heterogeneity on CIFAR-10." + ], + "image_footnote": [], + "bbox": [ + 516, + 686, + 906, + 797 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We observe that FLSeg achieves significantly better performance than other methods when $\\alpha = 0.1$ . This is because other methods were not designed to account for the", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3922", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "potential impact of data heterogeneity, whereas FLSeg enables clients to better adapt to their local datasets under non-IID conditions by learning personalized models, thereby achieving improved performance. However, as the data tends to be balanced ( $\\alpha = 10$ ), the gains brought by personalization will gradually decrease, but the performance of FLSeg is still better than that of other baselines. This is because applying RAR to model segments effectively alleviates the curse of dimensionality.", + "bbox": [ + 89, + 90, + 483, + 227 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.4. Privacy", + "text_level": 1, + "bbox": [ + 89, + 239, + 186, + 253 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To verify FLSeg's privacy protection capability against a semi-honest server, we present visual results of attacks using DLG and Inverting-Gradient, with and without our defense, on FEMNIST and CIFAR-100 in Fig. 5. Without privacy protection (Origin), the server can easily reconstruct training data. In contrast, when attacking the mixed model after Segment Exchange, only meaningless information can be extracted. Furthermore, in Sec. 5.1, we prove that the probability of the server reconstructing the original client model is negligible, thereby ensuring client privacy.", + "bbox": [ + 89, + 261, + 483, + 412 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f2274569cecfff567e9a04e0cf92367ac1d434d41653225ad7799ed2ad55ebcf.jpg", + "image_caption": [ + "Figure 5. Attack results of iDLG and Inverting-Gradients on CIFAR-10 and FEMNIST, respectively." + ], + "image_footnote": [], + "bbox": [ + 96, + 428, + 478, + 568 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.5.Impact of Segment Size", + "text_level": 1, + "bbox": [ + 89, + 633, + 316, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To investigate the impact of segment size on privacy protection and model performance, we vary the segment size $|\\gamma|$ across $\\{10^2, 10^3, 10^4, 10^5, 10^6, 2904970(\\ell)\\}$ under MinMax and Inverting-Gradient attacks on the heterogeneous CIFAR-10 dataset. We use the same settings as in the robustness experiments to evaluate the model's top-1 accuracy and perform an inference attack on the model updates of a single client from one training round. The experimental results are presented in Tab. 3.", + "bbox": [ + 89, + 657, + 483, + 792 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/01cf9f1bfc4018be3cd8a51dfe1b6e210076f8d54aaaea8445c88f1a949ea1cb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
|γ|102103104105106l
Accuracy63.93%64.33%65.72%60.93%59.67%57.75%
Inference
", + "bbox": [ + 94, + 806, + 480, + 873 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Accuracy and Privacy of FLSeg(Dnc) under different $|\\gamma |$", + "bbox": [ + 91, + 882, + 478, + 897 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We observe that the accuracy of $\\mathrm{FLSeg(DnC)}$ initially increases and then slightly decreases. We hypothesize that this occurs because benchmarks computed from smaller segments lack representativeness, while larger segments fail to mitigate the curse of dimensionality. Moreover, as the segment size increases, the reconstructed attack results reveal more information about the original data, since larger segments concentrate more original gradient information in the mixed model.", + "bbox": [ + 511, + 90, + 906, + 226 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.6. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 234, + 666, + 251 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We finally study how each component contributes to the overall performance of FLSeg. Specifically, we measure the following scenarios:", + "bbox": [ + 511, + 257, + 905, + 301 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "FedAvg. The baseline method.", + "- PT-only. Adds local personalized training to FedAvg.", + "- SA-only. Adds Segment Aggregation to FedAvg.", + "- FLSeg. The complete FLSeg method." + ], + "bbox": [ + 513, + 303, + 880, + 362 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Note that we did not include Segment Exchange as its purpose is to protect privacy without compromising the correctness of subsequent aggregation.", + "bbox": [ + 511, + 363, + 905, + 409 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/10619e13be0213b5a4578037d6316a5769ad9b4d581a47aabd718a80e2b38037.jpg", + "image_caption": [ + "Figure 6. Ablation study." + ], + "image_footnote": [], + "bbox": [ + 514, + 419, + 906, + 506 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conducted ablation experiments under four scenarios combining with/without $20\\%$ Min-Max attackers and IID/non-IID $(\\alpha = 0.5)$ CIFAR-10 dataset, with the results shown in Fig. 6. We observe that, in all scenarios, Segment Aggregation effectively resists Byzantine attacks. Meanwhile, the incorporation of personalized training does not degrade SA-only performance under IID data, and improves the model performance under non-IID data.", + "bbox": [ + 511, + 545, + 906, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 679, + 633, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we identified three key challenges in privacy-preserving robust FL: high computational costs, limited flexibility in RAR, and data heterogeneity. To overcome these, we proposed FLSeg, which uses Segment Exchange and Segment Aggregation to ensure privacy without limiting aggregation choices and employs a regularization term to enhance adaptation to heterogeneous data. Theoretical and experimental results validate FLSeg's security, robustness, and resilience to data heterogeneity.", + "bbox": [ + 511, + 703, + 905, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the future, we will explore stronger threat models, such as malicious server, or extend the privacy-preserving segment exchange framework to decentralized settings.", + "bbox": [ + 511, + 840, + 905, + 886 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3923", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This research was supported in part by the Major Research Plan of Hubei Province (2023BAA027), the Changsha Key Science and Technology Project under the \"Unveiling and Commanding\" Initiative (kq2503009), the Science, Technology and Innovation Commission of Shenzhen Municipality (GJHZ20240218114659027), the Key Research and Development Plan of Hubei Province (2024BAB049), and the Fundamental Research Funds for the Central Universities (YCJJ20252336).", + "bbox": [ + 89, + 113, + 483, + 239 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 265, + 187, + 280 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Abbas Acar, Hidayet Aksu, A Selcuk Uluagac, and Mauro Conti. A survey on homomorphic encryption schemes: Theory and implementation. ACM Computing Surveys (Csur), 51(4):1-35, 2018. 2", + "[2] James Bell, Adrià Gascon, Tancrède Lepoint, Baiyu Li, Sarah Meiklejohn, Mariana Raykova, and Cathie Yun. ACORN: input validation for secure aggregation. In 32nd USENIX Security Symposium, USENIX Security 2023, Anaheim, CA, USA, August 9-11, 2023, pages 4805-4822. USENIX Association, 2023. 1, 2, 3, 6", + "[3] James Henry Bell, Kallista A. Bonawitz, Adrià Gasón, Tancrède Lepoint, and Mariana Raykova. Secure single-server aggregation with (poly)logarithmic overhead. In CCS '20: 2020 ACM SIGSAC Conference on Computer and Communications Security, Virtual Event, USA, November 9-13, 2020, pages 1253-1269. ACM, 2020. 6", + "[4] Peva Blanchard, El Mahdi El Mhamdi, Rachid Guerraoui, and Julien Stainer. Machine learning with adversaries: Byzantine tolerant gradient descent. Advances in neural information processing systems, 30, 2017. 1, 2, 6", + "[5] Keith Bonawitz, Vladimir Ivanov, Ben Kreuter, Antonio Marcedone, H Brendan McMahan, Sarvar Patel, Daniel Ramage, Aaron Segal, and Karn Seth. Practical secure aggregation for privacy-preserving machine learning. In proceedings of the 2017 ACM SIGSAC Conference on Computer and Communications Security, pages 1175-1191, 2017. 1, 2, 6", + "[6] Sebastian Caldas, Sai Meher Karthik Duddu, Peter Wu, Tian Li, Jakub Konečný, H Brendan McMahan, Virginia Smith, and Ameet Talwalkar. Leaf: A benchmark for federated settings. arXiv preprint arXiv:1812.01097, 2018. 6", + "[7] Ran Canetti. Universally composable security: A new paradigm for cryptographic protocols. In FOCS, pages 136-145. IEEE Computer Society, 2001. 5", + "[8] Xiaoyu Cao, Minghong Fang, Jia Liu, and Neil Zhenqiang Gong. Fltrust: Byzantine-robust federated learning via trust bootstrapping. In NDSS. The Internet Society, 2021. 2, 6", + "[9] Alireza Fallah, Aryan Mokhtari, and Asuman E. Ozdaglar. Personalized federated learning with theoretical guarantees: A model-agnostic meta-learning approach. In NeurIPS, 2020. 3", + "[10] Minghong Fang, Xiaoyu Cao, Jinyuan Jia, and Neil Gong. Local model poisoning attacks to {Byzantine-Robust} federated learning. In 29th USENIX security symposium (USENIX Security 20), pages 1605-1622, 2020. 2, 6" + ], + "bbox": [ + 93, + 289, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Minghong Fang, Xiaoyu Cao, Jinyuan Jia, and Neil Zhenqiang Gong. Local model poisoning attacks to byzantine-robust federated learning. In USENIX Security Symposium, pages 1605–1622. USENIX Association, 2020. 2", + "[12] Marc Fischlin and Arno Mittelbach. An overview of the hybrid argument. IACR Cryptol. ePrint Arch., page 88, 2021. 5", + "[13] Jonas Geiping, Hartmut Bauermeister, Hannah Droge, and Michael Moeller. Inverting gradients-how easy is it to break privacy in federated learning? Advances in neural information processing systems, 33:16937-16947, 2020. 1, 2, 6", + "[14] Oded Goldreich, Silvio Micali, and Avi Wigderson. How to play any mental game or A completeness theorem for protocols with honest majority. In STOC, pages 218-229. ACM, 1987. 5", + "[15] Rachid Guerraoui, Sébastien Rouault, et al. The hidden vulnerability of distributed learning in byzantium. In International Conference on Machine Learning, pages 3521-3530. PMLR, 2018. 2, 3", + "[16] Anbu Huang, Yang Liu, Tianjian Chen, Yongkai Zhou, Quan Sun, Hongfeng Chai, and Qiang Yang. Starfl: Hybrid federated learning architecture for smart urban computing. ACM Trans. Intell. Syst. Technol., 12(4):43:1-43:23, 2021. 1", + "[17] Najeeb Moharram Jebreel, Josep Domingo-Ferrer, Alberto Blanco-Justicia, and David Sánchez. Enhanced security and privacy via fragmented federated learning. IEEE Transactions on Neural Networks and Learning Systems, 2022. 2", + "[18] Peter Kairouz, H Brendan McMahan, Brendan Avent, Aurélien Bellet, Mehdi Bennis, Arjun Nitin Bhagoji, Kallista Bonawitz, Zachary Charles, Graham Cormode, Rachel Cummings, et al. Advances and open problems in federated learning. Foundations and trends® in machine learning, 14(1-2): 1-210, 2021. 1", + "[19] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank Reddi, Sebastian Stich, and Ananda Theertha Suresh. SCAFFOLD: Stochastic controlled averaging for federated learning. In Proceedings of the 37th International Conference on Machine Learning, pages 5132-5143. PMLR, 2020. 3", + "[20] Sai Praneeth Karimireddy, Lie He, and Martin Jaggi. Byzantine-robust learning on heterogeneous datasets via bucketing. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022. 3", + "[21] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6", + "[22] Shenghui Li, Edith C-H Ngai, and Thiemo Voigt. An experimental study of byzantine-robust aggregation schemes in federated learning. IEEE Transactions on Big Data, 2023. 1, 2, 6", + "[23] Shenghui Li, Edith Ngai, Fanghua Ye, Li Ju, Tianru Zhang, and Thiemo Voigt. Blades: A unified benchmark suite for byzantine attacks and defenses in federated learning. In 2024 IEEE/ACM Ninth International Conference on Internet-of-Things Design and Implementation (IoTDI), 2024. 1, 2", + "[24] Tian Li, Shengyuan Hu, Ahmad Beirami, and Virginia Smith. Ditto: Fair and robust federated learning through per" + ], + "bbox": [ + 514, + 92, + 903, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3924", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sonalization. In International conference on machine learning, pages 6357-6368. PMLR, 2021. 3, 6", + "[25] Xiaoyuan Liu, Hongwei Li, Guowen Xu, Zongqi Chen, Xiaoming Huang, and Rongxing Lu. Privacy-enhanced federated learning against poisoning adversaries. IEEE Transactions on Information Forensics and Security, 16:4574-4588, 2021. 1, 2, 3, 6", + "[26] Yuchen Liu, Chen Chen, Lingjuan Lyu, Fangzhao Wu, Sai Wu, and Gang Chen. Byzantine-robust learning on heterogeneous data via gradient splitting. In International Conference on Machine Learning, pages 21404–21425. PMLR, 2023. 1, 2, 3, 6", + "[27] Zhi Lu, SongFeng Lu, YongQuan Cui, XueMing Tang, and JunJun Wu. Split aggregation: Lightweight privacy-preserving federated learning resistant to byzantine attacks. IEEE Transactions on Information Forensics and Security, 19:5575-5590, 2024. 1, 3", + "[28] Hidde Lycklama, Lukas Burkhalter, Alexander Viand, Nicolas Kuchler, and Anwar Hithnawi. Rofl: Robustness of secure federated learning. In SP, pages 453-476. IEEE, 2023. 1, 2, 6", + "[29] Zhuoran Ma, Jianfeng Ma, Yinbin Miao, Yingjiu Li, and Robert H Deng. Shieldfl: Mitigating model poisoning attacks in privacy-preserving federated learning. IEEE Transactions on Information Forensics and Security, 17:1639-1654, 2022. 1, 2, 3, 6", + "[30] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 2, 6", + "[31] Dinh C. Nguyen, Quoc-Viet Pham, Pubudu N. Pathirana, Ming Ding, Aruna Seneviratne, Zihuai Lin, Octavia A. Dobre, and Won-Joo Hwang. Federated learning for smart healthcare: A survey. ACM Comput. Surv., 55(3):60:1-60:37, 2023. 1", + "[32] Krishna Pillutla, Sham M Kakade, and Zaid Harchaoui. Robust aggregation for federated learning. IEEE Transactions on Signal Processing, 70:1142-1154, 2022. 1, 2, 6", + "[33] Virat Shejwalkar and Amir Houmansadr. Manipulating the byzantine: Optimizing model poisoning attacks and defenses for federated learning. In NDSS, 2021. 1, 2, 3, 6", + "[34] Virat Shejwalkar, Amir Houmansadr, Peter Kairouz, and Daniel Ramage. Back to the drawing board: A critical evaluation of poisoning attacks on production federated learning. In SP, pages 1354-1371. IEEE, 2022. 1", + "[35] Lichao Sun, Jianwei Qian, and Xun Chen. LDP-FL: practical private aggregation in federated learning with local differential privacy. In IJCAI, pages 1571–1578. ijcai.org, 2021. 1", + "[36] Canh T Dinh, Nguyen Tran, and Josh Nguyen. Personalized federated learning with moreau envelopes. Advances in Neural Information Processing Systems, 33:21394-21405, 2020. 3, 6", + "[37] Kang Wei, Jun Li, Ming Ding, Chuan Ma, Hang Su, Bo Zhang, and H. Vincent Poor. User-level privacy-preserving federated learning: Analysis and performance optimization. IEEE Trans. Mob. Comput., 21(9):3388-3401, 2022. 2" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Jiahui Wu, Weizhe Zhang, and Fucai Luo. On the security of \"lsfl: A lightweight and secure federated learning scheme for edge computing\". IEEE Trans. Inf. Forensics Secur., 19: 3481-3482, 2024. 1", + "[39] Jue Xiao, Xueming Tang, and Songfeng Lu. Privacy-preserving federated class-incremental learning. IEEE Transactions on Machine Learning in Communications and Networking, 2:150-168, 2024. 1", + "[40] Cong Xie, Oluwasanmi Koyejo, and Indranil Gupta. Fall of empires: Breaking byzantine-tolerant sgd by inner product manipulation. In Uncertainty in Artificial Intelligence, pages 261-270. PMLR, 2020. 1, 2, 6", + "[41] Dong Yin, Yudong Chen, Ramchandran Kannan, and Peter Bartlett. Byzantine-robust distributed learning: Towards optimal statistical rates. In International Conference on Machine Learning, pages 5650–5659. Pmlr, 2018. 2", + "[42] Bo Zhao, Konda Reddy Mopuri, and Hakan Bilen. idlg: Improved deep leakage from gradients. CoRR, abs/2001.02610, 2020. 2, 6", + "[43] Ligeng Zhu, Zhijian Liu, and Song Han. Deep leakage from gradients. Advances in neural information processing systems, 32, 2019. 1, 2" + ], + "bbox": [ + 516, + 90, + 903, + 402 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3925", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/FLSeg_ Enhancing Privacy and Robustness in Federated Learning under Heterogeneous Data via Model Segmentation/a4f90945-afa0-4f01-8edd-1b86f5d9d098_model.json b/2025/FLSeg_ Enhancing Privacy and Robustness in Federated Learning under Heterogeneous Data via Model Segmentation/a4f90945-afa0-4f01-8edd-1b86f5d9d098_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a6acd8611ac7f0df5c163a48529b9b9e487fb893 --- /dev/null +++ b/2025/FLSeg_ Enhancing Privacy and Robustness in Federated Learning under Heterogeneous Data via Model Segmentation/a4f90945-afa0-4f01-8edd-1b86f5d9d098_model.json @@ -0,0 +1,2387 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.133, + 0.131, + 0.868, + 0.177 + ], + "angle": 0, + "content": "FLSeg: Enhancing Privacy and Robustness in Federated Learning under Heterogeneous Data via Model Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.209, + 0.203, + 0.786, + 0.221 + ], + "angle": 0, + "content": "Zichun \\(\\mathbf{Su}^{1\\dagger}\\) Zhi Lu\\(^{1\\dagger}\\) Yutong Wu\\(^{1}\\) Renfei Shen\\(^{1}\\) Songfeng Lu\\(^{1,2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.221, + 0.865, + 0.257 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Cyber Science and Engineering, Huazhong University of Science and Technology \\(^{2}\\)Shenzhen Huazhong University of Science and Technology Research Institute" + }, + { + "type": "text", + "bbox": [ + 0.194, + 0.26, + 0.794, + 0.275 + ], + "angle": 0, + "content": "suzc25@m.fudan.edu.cn, {luzhi,wyt,shenrenfei,lusongfeng}@hust.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.31, + 0.327, + 0.326 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.34, + 0.486, + 0.644 + ], + "angle": 0, + "content": "Federated Learning (FL) enables collaborative training of a global model without data sharing, yet it faces critical challenges from privacy leakage and Byzantine attacks. Existing privacy-preserving robust FL frameworks suffer from three key limitations: high computational costs, restricted application of Robust Aggregation Rules (RAR), and inadequate handling of data heterogeneity. To address these limitations, we propose the FLSeg framework, which leverages Segment Exchange and Segment Aggregation to avoid excessive encryption computations while enabling unrestricted use of any RAR. Additionally, a regularization term in local training balances personalization with global model performance, effectively adapting to heterogeneous data. Our theoretical and experimental analyses demonstrate FLSeg's semi-honest security and computational efficiency: it achieves client and server time complexities of \\(O(\\ell)\\) and \\(O(n\\ell)\\), with empirical results showing significantly reduced computational times. Extensive experiments further confirm FLSeg's robustness across diverse heterogeneous and adversarial scenarios." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.672, + 0.222, + 0.687 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.697, + 0.485, + 0.879 + ], + "angle": 0, + "content": "Federated Learning (FL) allows organizations to collaboratively train models without sharing sensitive data. By exchanging model updates instead of raw data, FL enhances privacy and efficiency, making it ideal for fields like finance [18], healthcare [16], and smart cities [31]. However, FL is currently threatened by two major issues: privacy leakage and Byzantine attacks. Privacy leakage occurs when the server can infer clients' sensitive training data by analyzing the model updates [13, 43]. Byzantine attacks, on the other hand, involve malicious clients submitting incorrect or manipulated updates, which can disrupt or compromise the global model's performance [23, 34]. To address" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.311, + 0.906, + 0.461 + ], + "angle": 0, + "content": "these challenges, corresponding techniques have been developed. For privacy protection, methods like Homomorphic Encryption (HE) [29], Secure Multi-Party Computation (SMPC) [5], and Differential Privacy (DP) [35, 39] aim to safeguard training data by ensuring that only encrypted or obfuscated information is shared. For robustness against Byzantine attacks, various Robust Aggregation Rule (RAR), such as Krum [4], DnC [33], and RFA [32], help identify and exclude malicious updates, thereby protecting the global model from adversarial influence." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.906, + 0.583 + ], + "angle": 0, + "content": "However, achieving both robustness and privacy simultaneously presents a significant challenge, as privacy-preserving methods can unintentionally protect malicious updates from detection [27]. To address this issue, a series of Privacy-preserving and Robust FL (PRFL) frameworks have been proposed. These methods explore various architectures and protocols or select encryption schemes that are compatible with specific RAR [2, 25, 28, 29, 38]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.906, + 0.734 + ], + "angle": 0, + "content": "Despite these methods balancing robustness and privacy, they still have notable limitations: (1) Privacy-preserving techniques such as HE and SMPC introduce significant computational overhead [27]. (2) These methods impose restrictions on the use of RAR, supporting only simple, single RAR, which are vulnerable to adaptive attacks such as Min-Max [33] and IPM [40]. (3) Gradient heterogeneity caused by data heterogeneity significantly degrades both model performance and the accuracy of RAR [22]. However, existing PRFL methods largely overlook this issue." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Therefore, we aim to address the following question: \"How can we efficiently preserve privacy while avoiding restrictions on RAR implementation and mitigating the impact of data heterogeneity on robustness and performance?\" To this end, we draw inspiration from existing theoretical insights: reducing data dimensionality can alleviate the curse of dimensionality and gradient heterogeneity, thereby enhancing robustness [26, 33]. Based on this principle, we propose FLSeg, a model segmentation-based framework. We design a Segment Exchange privacy protocol to achieve privacy preservation without incurring excessive computa" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.887, + 0.348, + 0.901 + ], + "angle": 0, + "content": "* Corresponding author, † Equal contribution" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "3916" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "tional overhead. Additionally, we introduce a Segment Aggregation algorithm, which applies RAR operations at the low-dimensional model segments to mitigate gradient heterogeneity and the curse of dimensionality while supporting arbitrary RAR. Furthermore, we incorporate a regularization term into local training to facilitate the learning of personalized models, further improving performance under data heterogeneity. Our contributions are threefold:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.215, + 0.482, + 0.273 + ], + "angle": 0, + "content": "- A Novel Privacy-Preserving Robust FL Framework. To the best of our knowledge, FLSeg is the first PRFL framework that does not restrict the choice of RAR, while also mitigating the impact of heterogeneous data." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.275, + 0.482, + 0.424 + ], + "angle": 0, + "content": "- Computationally Efficient Privacy Protection. We theoretically demonstrate that our FLSeg satisfies semihonest security. We further provide a detailed analysis of FLSeg's computational time complexity, with the client-side and server-side complexities being \\( O(\\ell) \\) and \\( O(n\\ell) \\), respectively. In practical tests, FLSeg's client computation time is only 233 ms, compared to 1696 ms for SOTA method ACORN [2]. The server computation time per client is only 78 ms, whereas ACORN require 181 ms per client." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.426, + 0.482, + 0.561 + ], + "angle": 0, + "content": "- Extensive Experimental Analysis. Extensive experiments on three benchmark datasets show that our FLSeg can significantly improve the performance under varying degrees of data heterogeneity and in Byzantine attack, e.g., the accuracy of ACORN is \\(48.21\\%\\), while FLSeg achieves \\(64.59\\%\\) under a \\(20\\%\\) proportion of Min-Max attackers and on non-IID CIFAR-10 datasets. Further inference attack experiments also verified the privacy protection capability of FLSeg." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.215, + 0.482, + 0.561 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.574, + 0.24, + 0.589 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.482, + 0.674 + ], + "angle": 0, + "content": "Privacy Threats and Defences. Existing studies have demonstrated that gradient sharing in FL poses severe privacy risks. DLG attacks [42, 43] and Inverting Gradients attacks [13] can precisely reconstruct raw training data from client-uploaded gradients." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.482, + 0.78 + ], + "angle": 0, + "content": "To mitigate these risks, privacy-preserving techniques such as DP [37], HE [25], and SMPC [5] have been proposed to obfuscate or encrypt gradients. However, these approaches have notable limitations: DP inevitably reduce model utility due to noise injection, while HE and SMPC incur high computational overhead due to complex encryption structures and protocols [1]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Byzantine Attack and Defence. Byzantine attacks are initiated by malicious clients who manipulate or send adversarial model updates to degrade model performance [23]. A common type of such attacks is data poisoning, such as LabelFlipping [10], which manipulates training data to mislead the model into learning incorrect associations. Additionally, adaptive attacks, such as Min-Max [33] and IPM [40], optimize the magnitude and direction of malicious" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "gradient perturbations to maximize the disruption of the global FL model while simultaneously evading detection." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.122, + 0.905, + 0.363 + ], + "angle": 0, + "content": "To mitigate these risks, RAR modify the standard mean aggregation method [30] by performing statistical or feature analysis on the collected updates to filter out malicious ones. Examples include methods based on Euclidean distance, such as Krum [4] and Bulyan [15]; those leveraging coordinate values, such as RFA [32] and Median [41]; and approaches utilizing spectral anomaly detection, such as DnC [33]. Beyond RAR, Robust FL integrate auxiliary techniques for enhanced defense. FLTrust [8] uses a root dataset to assess the quality of client updates, ClippedClustering [22] leverages historical gradient information to constrain and clip gradients before aggregation, and GAS [26] proposes a gradient splitting technique to reduce data dimensionality and enhance the robustness of RAR. However, these defenses do not secure the gradients during transmission, rendering them vulnerable to privacy threats." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.905, + 0.56 + ], + "angle": 0, + "content": "Privacy-preserving Robust FL. PRFL integrates privacy-preserving techniques with RAR through specially designed protocols. PEFL [25] and ShieldFL [29] employ HE to ensure privacy while using cosine similarity to detect malicious updates. FFL [17] proposes a parameter mixing method to protect privacy while applying cosine similarity to the mixed model for robustness. ACORN [2] and RoFL [28] utilize the SecAgg SMPC protocol to safeguard the privacy of inputs while employing zero-knowledge proofs to bound the norm of model updates, thereby mitigating the impact of malicious updates. However, the high computational cost, restricted support for RAR, and lack of consideration for data heterogeneity limit their applicability." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.572, + 0.652, + 0.587 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.597, + 0.905, + 0.643 + ], + "angle": 0, + "content": "Notion. For any positive integer \\( n \\in \\mathbb{N}^+ \\), we denote the set \\( \\{1, \\ldots, n\\} \\) by \\( [n] \\). The cardinality of a set \\( w \\) is denoted by \\( |w| \\). We denote the \\( l_2 \\) norm of vector \\( w \\) by \\( \\| w \\| \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.643, + 0.905, + 0.706 + ], + "angle": 0, + "content": "Model Segmentation. For a model \\(w\\), the mode of segmentation refers to a partitioning of the set of model parameters, \\(\\{\\gamma^{(1)},\\gamma^{(2)},\\dots,\\gamma^{(d)}\\}\\), where different segments are disjoint, and \\(w = \\bigcup_{k = 1}^{d}\\gamma^{(k)}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.905, + 0.825 + ], + "angle": 0, + "content": "System Setup. We consider a classic FL scenario [30], involving \\( n \\) clients, denoted by \\( \\{c_i\\}_{i \\in [n]} \\), and a single server \\( S \\). \\( c_i \\) has a unique local dataset \\( D_i \\) that is non-IID. \\( S \\) starts with initial model weights \\( w_0 \\) and distributes the model to each \\( c_i \\). \\( c_i \\) train the model on their datasets and send their updates \\( w_i^t \\) to the server \\( S \\), which aggregates these updates and proceeds to the next round. After \\( T \\) rounds, this process results in a global model \\( w^T \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Threat Model. Our system faces two primary threats. First, in line with [8, 11], there may be Byzantine clients within the system who manipulate their uploaded model to disrupt training. They can launch targeted attacks such as LabelFlipping [10], and powerful adaptive attacks like Min" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3917" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.088, + 0.905, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.394, + 0.35, + 0.604, + 0.365 + ], + "angle": 0, + "content": "Figure 1. The Overview of FLSeg." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.391, + 0.484, + 0.482 + ], + "angle": 0, + "content": "Max [33]. Second, the server is semi-honest, adhering to the aggregation protocol while attempting to infer private local data from the client's updates [29]. As in prior work [2, 25, 27, 29], we assume that the adversary can control a number of clients \\( f < (n - 1) / 2 \\) and semi-honest server and Byzantine clients do not collude." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.501, + 0.226, + 0.518 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.528, + 0.196, + 0.542 + ], + "angle": 0, + "content": "4.1. Intuition" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.552, + 0.484, + 0.674 + ], + "angle": 0, + "content": "Existing RAR methods commonly suffer from the curse of dimensionality [15], where malicious gradients can hide within high-dimensional vectors. Additionally, data heterogeneity introduces gradient heterogeneity [19], further increasing the difficulty of detection. Studies have shown that reducing input dimensionality can help alleviate this issue, for example, by processing gradients through sub-sampling [33], grouping [20], or splitting [26]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Based on this insight, we propose an intuitive approach: by fixing the mode of segmentation, each client selects a partner to exchange model segments, while the server applies the same segmentation mode during aggregation. In this way, as long as the server is unaware of the specific exchange pairs or the exchanged segments, it cannot reconstruct the client models, thereby ensuring privacy protection. Additionally, the server apply RAR on each segment, which not only reduces the curse of dimensionality and gradient heterogeneity but also ensures that each segment to which the RAR is applied originates from different clients, thereby ensuring the correctness of the RAR implementation. Besides, although these processes mitigate gradient heterogeneity caused by Non-IID data [15, 20], they fail to address the performance degradation induced by Non-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.391, + 0.906, + 0.453 + ], + "angle": 0, + "content": "IID data. To tackle this issue, we incorporate local regularization techniques from personalized FL [24, 36] to improve model performance without altering the overall process. The whole process is illustrated in Fig. 1." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.459, + 0.871, + 0.473 + ], + "angle": 0, + "content": "4.2. Non-IID Data: Learn Personalized Model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.481, + 0.907, + 0.618 + ], + "angle": 0, + "content": "Heterogeneous data can cause the global model to fail to adapt to all clients' data distributions, leading to performance degradation. Existing personalized FL methods address this issue by balancing the global and local models to integrate the generalization ability of the global model with the uniqueness of local data [9, 24, 36]. Inspired by this approach, we modify each client's local training process by allowing the client to maintain a personalized model \\( v_{i} \\), with its objective function defined in Eq. (1)." + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.628, + 0.907, + 0.657 + ], + "angle": 0, + "content": "\\[\n\\min _ {v _ {i}} h _ {i} \\left(v _ {i}, w ^ {t}\\right) = \\mathcal {L} _ {i} \\left(v _ {i}\\right) + \\frac {\\lambda}{2} \\left\\| v _ {i} - w ^ {t} \\right\\| ^ {2}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.906, + 0.766 + ], + "angle": 0, + "content": "where \\( w^{t} \\) represents the global model received in each round; \\( \\mathcal{L}_i(w) = \\mathbb{E}_{\\xi_i \\sim \\mathcal{D}_i}[f_i(w; \\xi_i)] \\) denotes the loss function on \\( c_{i} \\)'s local data; \\( \\lambda \\) is a regularization parameter that balances the influence of the personalized and global models. When \\( \\lambda \\to +\\infty \\), the personalized model converges to the global model, whereas \\( \\lambda \\to 0 \\) leads to a more distinct personalized effect." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.767, + 0.906, + 0.843 + ], + "angle": 0, + "content": "The detailed process is provided in Algorithm 1. During training, the personalized model is updated synchronously with the global model. Each client retains \\( v_{i}^{t + 1} \\) locally for local tasks and prepares the model update \\( w_{i}^{t} \\) for the subsequent Segment Exchange." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.767, + 0.866 + ], + "angle": 0, + "content": "4.3. Privacy: Segment Exchange" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "To prevent the server from obtaining each client's complete model, the client mixes its model by exchanging segments" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3918" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.103, + 0.096, + 0.297, + 0.11 + ], + "angle": 0, + "content": "Algorithm 1: Local Training" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.113, + 0.444, + 0.158 + ], + "angle": 0, + "content": "Input : Personalized model \\( v_{i}^{t} \\), model received from server \\( w^{t} \\), regularization coefficient \\( \\lambda \\), learning rate \\( \\eta \\), local epochs \\( E \\)" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.158, + 0.436, + 0.175 + ], + "angle": 0, + "content": "Output: Updates \\( w_{i}^{t + 1} \\), personalized model \\( v_{i}^{t + 1} \\)" + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.174, + 0.276, + 0.19 + ], + "angle": 0, + "content": "1 \\(w_{i}^{t + 1}\\gets w^{t},v_{i}^{t + 1}\\leftarrow v_{i}^{t};\\)" + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.19, + 0.298, + 0.204 + ], + "angle": 0, + "content": "2 for epoch \\(\\leftarrow 0\\) to \\(E - 1\\) do" + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.205, + 0.458, + 0.237 + ], + "angle": 0, + "content": "3 /*update personalized model \\*/ \\(v_{i}^{t + 1}\\gets v_{i}^{t + 1} - \\eta (\\nabla \\mathcal{L}_{i}(v_{i}^{t + 1}) + \\lambda (v_{i}^{t + 1} - w^{t}))\\)" + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.238, + 0.354, + 0.27 + ], + "angle": 0, + "content": "4 \\(\\begin{array}{r}\\left\\lfloor \\begin{array}{c}/*\\mathrm{update~global~model}*/\\\\ w_i^{t + 1}\\gets w_i^{t + 1} - \\eta \\nabla \\mathcal{L}_i(w_i^{t + 1}); \\end{array} \\right. \\end{array}\\)" + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.272, + 0.235, + 0.291 + ], + "angle": 0, + "content": "5 return \\( w_{i}^{t + 1}, v_{i}^{t + 1} \\)" + }, + { + "type": "list", + "bbox": [ + 0.099, + 0.174, + 0.458, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.324, + 0.483, + 0.384 + ], + "angle": 0, + "content": "with a partner before submitting local updates. However, to ensure that clients do not reveal their models to their exchange partners during this process, we design a secure model segment exchange protocol, detailed in Algorithm 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.385, + 0.484, + 0.537 + ], + "angle": 0, + "content": "An simple example is illustrated in Fig. 2. The two parties first use the Diffie-Hellman key exchange to agree on a mask \\( m \\) that determines which segments are exchanged. \\( w_{i}^{-m} \\) is encrypted using client \\( c_{i} \\)'s key \\( r_{i} \\), ensuring that \\( c_{j} \\) cannot access \\( c_{i} \\)'s private information while still enabling segment exchange through the properties of the XOR operation. Meanwhile, the server can decrypt the client key to obtain the mixed model. However, since it does not know \\( m \\) or the exchange partner, it cannot recover the original client model, thereby preserving privacy." + }, + { + "type": "image", + "bbox": [ + 0.105, + 0.549, + 0.484, + 0.691 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.7, + 0.428, + 0.715 + ], + "angle": 0, + "content": "Figure 2. The Workflow of Segment Exchange." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.742, + 0.391, + 0.759 + ], + "angle": 0, + "content": "4.4. Robustness: Segment Aggregation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.764, + 0.483, + 0.885 + ], + "angle": 0, + "content": "To mitigate the curse of dimensionality, FLSeg partitions the mixed model according to the mode of segmentation and then applies RAR to each segment, rather than performing RAR directly on the entire mixed model. This approach not only ensures that each segment in the RAR originates from different clients but also effectively mitigates the challenges posed by high input dimensionality. The detailed procedure is provided in Algorithm 3." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.482, + 0.902 + ], + "angle": 0, + "content": "After grouping segments based on the mode of segmen" + }, + { + "type": "code_caption", + "bbox": [ + 0.525, + 0.095, + 0.807, + 0.11 + ], + "angle": 0, + "content": "Algorithm 2: Segment Exchange Protocal" + }, + { + "type": "algorithm", + "bbox": [ + 0.517, + 0.112, + 0.905, + 0.402 + ], + "angle": 0, + "content": "Input: Client \\(c_{i}\\) 's model update \\(w_{i}\\) ,server's public-private key pair \\((pk_S,sk_S)\\) ,the exchange partner \\(C_j\\) \n1 client \\(c_{i}\\) and \\(c_{j}\\) \n2 generate client's key \\(r_i\\) . \n3 perform Diffie-Hellman key exchange with \\(c_{j}\\) to obtain mask \\(m\\) . \n4 \\(w_{i}^{m} / w_{i}^{-m}\\gets\\) Set the \\(k\\) -th segment \\(\\gamma^{(k)} = 0\\) of \\(w_{i}\\) if the \\(k\\) -th bit of \\(m\\) is \\(1 / 0\\) . \n5 exchange \\(w_{i}^{-m}\\oplus r_{i}\\) and \\(\\mathrm{Enc}(r_i,pk_S)\\) with \\(c_{j}\\) . \n6 compute \\((e_i)_{\\mathrm{mix}} = w_i^m\\oplus (w_j^{-m}\\oplus r_j)\\) \\(= (w_{i})_{\\mathrm{mix}}\\oplus r_{j}\\) . \n7 send \\((e_i)_{\\mathrm{mix}}\\) ,Enc \\((r_j,pk_S)\\) to server S; \n8 server S \n9 decrypts each client \\(c_{i}\\) 's key \\(r_j = \\mathrm{Dec}(\\mathrm{Enc}(r_j,pk_S),sk_S)\\) . \n10 decrypts each client \\(c_{i}\\) 's mixed model \\((w_{i})_{\\mathrm{mix}}\\leftarrow (e_{i})_{\\mathrm{mix}}\\oplus r_{j};\\)" + }, + { + "type": "code_caption", + "bbox": [ + 0.525, + 0.425, + 0.765, + 0.44 + ], + "angle": 0, + "content": "Algorithm 3: Segment Aggregation" + }, + { + "type": "algorithm", + "bbox": [ + 0.517, + 0.442, + 0.904, + 0.717 + ], + "angle": 0, + "content": "Input: uploaded mixed models \\(\\{(w_{i})_{\\mathrm{mix}}\\}_{i\\in [n]}\\) the mode of segmentation \\(\\{\\gamma^{(1)},\\dots ,\\gamma^{(d)}\\}\\) Output: the aggregation result \\(w\\) \n1 for each model \\((w_{i})_{\\mathrm{mix}}\\) do \n2 split model into segments \\(\\{\\gamma_i^{(1)},\\ldots ,\\gamma_i^{(d)}\\} \\gets (w_i)_{\\mathrm{mix}};\\) \n3 for each segment \\(\\gamma^{(k)}\\in \\{\\gamma^{(1)},\\dots ,\\gamma^{(d)}\\}\\) do \n4 apply the RAR to the segments to establish a benchmark \\(\\overline{\\gamma}^{(k)}\\gets \\mathrm{RAR}(\\{\\gamma_i^{(k)}\\}_{i\\in [n]})\\) \n5 calculate the similarity \\(s_i^{(k)}\\) for each segment \\(\\gamma_{i}^{(k)}\\) with the benchmark \\(\\overline{\\gamma}^{(k)}\\) using Eq. (2); \n6 compute each model's similarity score as \\(s_i = \\sum_{k = 1}^d s_i^{(k)}\\) and select the top \\(n - 2f\\) models with the highest scores to form set \\(I\\) \n7 aggregate \\(w\\gets \\sum_{i\\in I}(w_i)_{\\mathrm{mix}}\\) \n8 return \\(w\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.749, + 0.907, + 0.826 + ], + "angle": 0, + "content": "tation, \\(S\\) applies RAR to each segment group to obtain the benchmark \\(\\overline{\\gamma}^{(k)}\\) and uses it as an honest reference to compute the similarity \\(s_i^{(k)}\\) of each segment \\(\\gamma_i^{(k)}\\) using Eq. (2), then sums the segment similarities to obtain the model's score \\(s_i\\) for this round." + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.837, + 0.905, + 0.878 + ], + "angle": 0, + "content": "\\[\ns _ {i} ^ {(k)} = \\frac {\\left(\\gamma_ {i} ^ {(k)} - \\mu\\right) \\cdot \\left(\\bar {\\gamma} ^ {(k)} - \\mu\\right)}{\\| \\gamma_ {i} ^ {(k)} - \\mu \\| \\| (\\bar {\\gamma} ^ {(k)} - \\mu) \\|}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.883, + 0.907, + 0.904 + ], + "angle": 0, + "content": "where \\(z = |\\gamma^{(k)}|\\) is the number of parameters of \\(\\gamma_{i}^{(k)}\\) and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3919" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.088, + 0.484, + 0.131 + ], + "angle": 0, + "content": "\\(\\overline{\\gamma}^{(k)}\\) , and \\(\\mu = \\frac{\\sum_{j = 1}^{z}\\gamma_{i,j}^{(k)} + \\sum_{j = 1}^{z}\\overline{\\gamma}_{j}^{(k)}}{2z}\\) is the average value of all parameters in \\(\\gamma_i^{(k)}\\) and \\(\\overline{\\gamma}^{(k)}\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.131, + 0.484, + 0.178 + ], + "angle": 0, + "content": "Since, in the worst case, each honest client's partner may be a malicious client, \\( S \\) select the top \\( n - 2f \\) clients with the highest scores for aggregation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.187, + 0.204, + 0.204 + ], + "angle": 0, + "content": "4.5. Summary" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.21, + 0.484, + 0.346 + ], + "angle": 0, + "content": "The mode of segmentation in FLSeg remains unchanged throughout the process. Before training begins, the server distributes its public key \\(pk_{S}\\) to all clients. In each training round, clients first perform Local Training (Algorithm 1), then conduct Segment Exchange to obtain the mixed model (Algorithm 2), and finally, the server applies Segment Aggregation (Algorithm 3) to the received global models. The overall process is detailed in the algorithm provided in the Appendix A." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.347, + 0.484, + 0.499 + ], + "angle": 0, + "content": "FLSeg does not restrict how clients select partners; however, to prevent honest clients from continuously collaborating with malicious clients, we design a reputation system based on computed similarity scores. Specifically, each client \\( c_{i} \\) initializes its reputation list \\( R_{i} = \\{r_{1} = 0,\\dots ,r_{n} = 0\\} \\). After each round of aggregation, the server normalizes the similarity scores of each model using Eq. (3) and sends these scores to the corresponding clients. A lower score for an honest client indicates that its model may have been influenced by malicious updates." + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.522, + 0.483, + 0.559 + ], + "angle": 0, + "content": "\\[\ns _ {i} ^ {\\prime} = 2 \\left(\\frac {s _ {i} - \\operatorname* {m i n} \\left(\\left\\{s _ {i} \\right\\} _ {i \\in [ n ]}\\right)}{\\operatorname* {m a x} \\left(\\left\\{s _ {i} \\right\\} _ {i \\in [ n ]}\\right) - \\operatorname* {m i n} \\left(\\left\\{s _ {i} \\right\\} _ {i \\in [ n ]}\\right)}\\right) - 1 \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.57, + 0.484, + 0.66 + ], + "angle": 0, + "content": "Clients use these scores to update the reputation values of their partners as \\( r_j = r_j + s_i' \\). In subsequent rounds, clients can select partners with higher reputation values from their reputation lists for Segment Exchange. Practical test results for the reputation system are provided in Appendix B." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.675, + 0.188, + 0.692 + ], + "angle": 0, + "content": "5. Analysis" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.7, + 0.295, + 0.717 + ], + "angle": 0, + "content": "5.1. Semi-Honest Security" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.723, + 0.484, + 0.8 + ], + "angle": 0, + "content": "We use the Universally Composable Security Framework [7] to prove that FLSeg satisfies semi-honest security, meaning that FLSeg does not leak any information as long as the adversaries follow the protocol. We first present definition of semi-honest security [14]:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Definition 1 (Real-World). For protocol \\(\\pi\\), there exists a semi-honest adversary \\(\\mathcal{A}\\) that can control a subset of participants \\(\\mathcal{C}\\). \\(\\mathcal{A}\\) will follow the protocol \\(\\pi\\) but will attempt to infer the private inputs \\(\\{x_1, \\ldots, x_n\\}\\) based on the available information. The probability distribution of the inference results is denoted as \\(\\text{Real}_{\\pi, \\mathcal{A}}(\\mathcal{P}, \\{x_1, \\ldots, x_n\\})\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.184 + ], + "angle": 0, + "content": "Definition 2 (Ideal-World). Assume there exists an ideal functionality \\(\\mathcal{F}\\) that can execute any interaction in protocol \\(\\pi\\) with absolute security. There also exists a simulator Sim that simulates interactions \\(\\mathcal{F}\\) in the same manner as the real-world adversary \\(\\mathcal{A}\\). The probability distribution of the inference results is denoted as Ideal\\(_{\\mathcal{F},\\text{Sim}}(\\mathcal{P},\\{x_1,\\ldots,x_n\\})\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.188, + 0.906, + 0.248 + ], + "angle": 0, + "content": "Definition 3 (Semi-Honest). If \\( \\text{Real}_{\\pi, \\mathcal{A}}(\\mathcal{P}, \\{x_1, \\ldots, x_n\\}) \\) and \\( \\text{Ideal}_{\\mathcal{F}, \\text{Sim}}(\\mathcal{P}, \\{x_1, \\ldots, x_n\\}) \\) are computationally indistinguishable to an honest-but-curious adversary \\( \\mathcal{A} \\), then protocol \\( \\pi \\) is considered semi-honest secure." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.255, + 0.906, + 0.331 + ], + "angle": 0, + "content": "Theorem 1. Assuming the \\(FLSeg\\) protocol is denoted as \\(\\pi\\), the participants of the protocol \\(\\mathcal{P} = \\{c_1, \\ldots, c_n, S\\}\\) include all clients and the server, and the private inputs are the gradients of each honest client \\(\\{w_1, \\ldots, w_n\\}\\). Then \\(FLSeg\\) is semi-honest secure." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.337, + 0.906, + 0.443 + ], + "angle": 0, + "content": "Proof. We use the Hybrid Argument [12]. By gradually replacing components of the protocol's multiple interaction stages with the ideal functionality \\(\\mathcal{F}\\), we demonstrate that each stage's replacement does not significantly alter the probability distribution of the adversary's inference results. Assuming \\(\\mathcal{T}\\) represents the entire parameter space of the model, we define the following three hybrids:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.443, + 0.815, + 0.457 + ], + "angle": 0, + "content": "- Hybrid 0: The original FLSeg Framework." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.457, + 0.904, + 0.492 + ], + "angle": 0, + "content": "- Hybrid 1: Similar to Hybrid 0, but replaces \\(w_{i}^{-m} \\oplus r_{i}\\) exchanged by clients with random values \\(\\phi_{i}^{(1)} \\leftarrow \\mathcal{T}\\)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.492, + 0.904, + 0.528 + ], + "angle": 0, + "content": "- Hybrid 2: Similar to Hybrid 1, but replaces \\((w_{i})_{\\mathrm{mix}}\\) transmitted to the server with random values \\(\\phi_{i}^{(2)} \\stackrel{\\S}{\\leftarrow} \\mathcal{T}\\)." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.443, + 0.904, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.535, + 0.906, + 0.565 + ], + "angle": 0, + "content": "Lemma 1 (Hybrid 1). Hybrid 0 and Hybrid 1 are computationally indistinguishable if FLSeg is semi-honest secure." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.574, + 0.906, + 0.625 + ], + "angle": 0, + "content": "Proof. Since \\( r_i \\) and \\( \\phi_i^{(1)} \\) follow the same distribution, \\( w_i^{(\\neg m)} \\oplus r_i \\) is also distributed identically to \\( \\phi_i^{(1)} \\). Thus, the advantage of \\( \\mathcal{A} \\) is given by Eq. (4)." + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.63, + 0.906, + 0.656 + ], + "angle": 0, + "content": "\\[\n\\left| \\mathbb {P} [ \\mathcal {A} (\\phi_ {i} ^ {(1)}) = 1 ] - \\mathbb {P} [ \\mathcal {A} (w _ {i} ^ {- m} \\oplus r _ {i}) = 1 ] \\right| \\leq \\operatorname {n e g} (d) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.662, + 0.906, + 0.693 + ], + "angle": 0, + "content": "where \\(\\mathrm{neg}(d)\\) is a negligible function. Hence, Hybrid 0 and Hybrid 1 are computationally indistinguishable." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.701, + 0.906, + 0.732 + ], + "angle": 0, + "content": "Lemma 2 (Hybrid 2). Hybrid 1 and Hybrid 2 are computationally indistinguishable if FLSeg is semi-honest secure." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.74, + 0.906, + 0.817 + ], + "angle": 0, + "content": "Proof. Since the adversary \\(\\mathcal{A}\\) cannot control both the client and the server simultaneously, \\(j\\) and \\(m\\) are random to \\(\\mathcal{S}\\). If \\(\\mathcal{S}\\) attempts to recover the private \\(w_{i}\\) by random matching, the probability of success is \\(\\frac{1}{n} \\cdot \\frac{1}{2^d}\\). Thus, the advantage of \\(\\mathcal{A}\\) is given by Eq. (5)." + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.822, + 0.905, + 0.852 + ], + "angle": 0, + "content": "\\[\n\\left| \\mathbb {P} [ \\mathcal {A} ((w _ {i}) _ {\\mathrm {m i x}}) = 1 ] - \\mathbb {P} [ \\mathcal {A} (\\phi_ {i} ^ {(2)}) = 1 ] \\right| \\leq \\frac {1}{n} \\cdot \\frac {1}{2 ^ {d}} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Since this advantage is negligible for a polynomial-time adversary, it can be ignored. Thus, Hybrid 1 and Hybrid 2 are computationally indistinguishable." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3920" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.136, + 0.089, + 0.865, + 0.193 + ], + "angle": 0, + "content": "
SecAgg [5]Bell et al. [3]ACORN-detect [2]ShieldFL [29]PEFL [25]FLSeg
ClientComplexityO(n2+ ℓn)O(log2 n + ℓ log n)O(log2 n + ℓ log n)O(ℓ(TEnc + TDec))O(ℓ(TEnc + TDec))O(ℓ)
Overhead2,557 ms1,696 ms1,729 ms231,541 ms215,667 ms233 ms
ServerComplexityO(ℓn2)O(n(log2 n + ℓ log n))O(n(log2 n + ℓ log n))O(ℓnTDec)O(ℓn(TEnc + TDec))O(ℓn)
Overhead192 ms/c182 ms/c181 ms/c78,685 ms/c790,032 ms/c78 ms/c
Robustness○○◇◇◇◇
" + }, + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.202, + 0.908, + 0.273 + ], + "angle": 0, + "content": "Table 1. Comparison of computational time complexity and time overhead. \\( n \\) represents the number of clients, \\( \\ell \\) denotes the size of the model parameters. In calculating computation time overhead, let \\( n = {10}^{3} \\) and \\( \\ell = {10}^{6} \\) . \\( {T}_{\\text{Enc }} \\) and \\( {T}_{\\text{Dec }} \\) represent the encryption and decryption constant overhead in HE. ”/c” denotes the average time overhead incurred by the server for each client. \\( \\bigcirc \\) indicates the absence of robustness, \\( \\mathbf{D} \\) denotes support for fixed RAR only,and \\( \\bullet \\) denotes support for arbitrary RAR. We omit overheads unrelated to privacy protection,such as training and aggregation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.299, + 0.486, + 0.361 + ], + "angle": 0, + "content": "Through the above hybrid experiments, we have gradually proven the information transmitted from the real-world to the ideal-world is computationally indistinguishable from random values. Thus, we have proven Theorem 1. \\(\\square\\)" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.372, + 0.37, + 0.389 + ], + "angle": 0, + "content": "5.2. Efficiency in privacy protection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.395, + 0.484, + 0.456 + ], + "angle": 0, + "content": "Assume \\( n \\) represents the number of clients and \\( \\ell \\) denotes the size of the model parameters. The computational overhead on the client side for FLSeg primarily consists of the following components:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.458, + 0.395, + 0.474 + ], + "angle": 0, + "content": "\\(\\mathcal{O}(\\ell^2)\\colon m\\gets\\) Diffie-Hellman key exchange," + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.474, + 0.445, + 0.49 + ], + "angle": 0, + "content": "- \\(\\mathcal{O}(\\ell): \\text{Calculation of } w_i^m, w_i^{-m}, w_i^{-m} \\oplus r_i, (e_i)_{\\text{mix}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.49, + 0.245, + 0.506 + ], + "angle": 0, + "content": "\\(\\mathcal{O}(\\ell):\\mathrm{Enc}(r_j,pk_{\\mathcal{S}})\\)" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.458, + 0.445, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.508, + 0.483, + 0.734 + ], + "angle": 0, + "content": "Here, we introduce an optimization technique: instead of directly using Diffie-Hellman to generate \\(m\\) or encrypting \\(r_j\\), which incurs significant computational time costs, we generate respective random seeds \\(\\mathsf{seed}_{r_j}\\) and \\(\\mathsf{seed}_m\\). Both clients then use a Pseudo-Random Number Generator (PRNG) [5] to produce \\(r_j\\) and \\(m\\), and only the encrypted seed \\(\\mathrm{Enc}(\\mathsf{seed}_{r_j},pk_S)\\) is sent. This approach reduces the overhead of Diffie-Hellman and asymmetric encryption to a constant level, depending solely on the seed length. Thus, the primary computational overhead is shifted to generating \\(r_j\\) using PRNG \\((\\mathsf{seed}_{r_j}) = r_j\\) and performing XOR operations, with a complexity of \\(\\mathcal{O}(\\ell)\\). Similarly, the decryption overhead on the server is minimized, focusing mainly on decrypting the mixed model \\((w_i)_{\\mathrm{mix}}\\) using XOR operations, resulting in a complexity of \\(\\mathcal{O}(\\ell n)\\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.735, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To demonstrate the advantages of FLSeg in efficient privacy protection, we compare the computational time complexity and time overhead of FLSeg with SOTA privacy-preserving FL that also satisfy semi-honest security, including SecAgg [5], Bell et al. [3], ShieldFL [29], RoFL [28], and PEFL [25]. The comparison results are presented in Tab. 1. The time complexity of ShieldFL and PEFL is asymptotically \\(O(\\ell)\\); however, the Paillier HE they employ incurs substantial constant overheads, \\(T_{\\mathrm{Enc}}\\) and \\(T_{\\mathrm{Dec}}\\). SecAgg, ACORN, and Bell et al. are aggregation schemes based on SMPC techniques, requiring the compu" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.299, + 0.907, + 0.406 + ], + "angle": 0, + "content": "tation of pairwise masking for all other clients. To more intuitively demonstrate the advantages of FLSeg, we tested each method's runtime in a practical setting with \\( n = 10^3 \\) and \\( \\ell = 10^6 \\). The results show that FLSeg achieves an \\( 8 \\times \\) improvement in client-side and a \\( 2 \\times \\) improvement in server-side computation compared to the fastest method. See Appendix C for the experimental setup." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.417, + 0.646, + 0.433 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.441, + 0.603, + 0.457 + ], + "angle": 0, + "content": "6.1. Setups" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.463, + 0.907, + 0.659 + ], + "angle": 0, + "content": "Datasets and Models. Our experiments are conducted on three datasets—FEMNIST [6], CIFAR-10, and CIFAR-100 [21]—using a Multilayer Perceptron (MLP), Convolutional Neural Network (CNN), and SqueezeNet, respectively. For FEMNIST, the dataset is naturally non-IID and consists of 199 unique clients. For CIFAR-10 and CIFAR-100, we set the number of clients, \\( n = 100 \\) and \\( n = 50 \\), and generate non-IID data distributions across clients using a Dirichlet distribution by following [24, 26, 36]. The concentration parameter \\( \\alpha \\) of the Dirichlet distribution controls the degree of data heterogeneity, with lower values indicating higher heterogeneity. For each client, \\( 90\\% \\) of their data is allocated for training, while the remaining \\( 10\\% \\) is reserved for testing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.906, + 0.72 + ], + "angle": 0, + "content": "Evaluated attacks. We select the classic targeted attack, LabelFlipping (LP) [10]; two powerful adaptive attacks, Min-Max [33] and IPM [40]; and two privacy inference attacks, iDLG [42] and Inverting-Gradients [13]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.809 + ], + "angle": 0, + "content": "Compared Methods. We consider the FL baseline FedAvg [30]; three prominent RARs: Krum [4], DnC [33], and RFA [32]; three robust FL methods: GAS [26], FLTrust [8], and ClippedCluster [22]; and three privacy-preserving robust FL methods: ACORN [2], ShieldFL [29], and PEFL [25]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.885 + ], + "angle": 0, + "content": "Evaluation. We use average top-1 accuracy of each client to evaluate the performance of global models. We run each experiment for five times and report the mean and standard deviation of the highest accuracy during the training process." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.906, + 0.902 + ], + "angle": 0, + "content": "All hyperparameter settings for the compared methods" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3921" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.904, + 0.299 + ], + "angle": 0, + "content": "
DatasetAttackBaselineRARRobust FLPrivacy-preserving Robust FLOur
FedAvgKrumDnCRFAGASFLTrustClipped ClusteringACORNShieldFLPEFLFLSeg (Krum)FLSeg (DnC)
CIFAR-10-52.16±0.9753.86±0.9053.32±0.1353.52±0.4851.16±0.2853.28±0.7053.44±1.5949.08±1.9752.61±0.2746.31±0.8860.45±0.8263.99±0.53
LF47.46±0.6947.55±1.9246.50±1.5555.20±1.0647.96±1.9250.61±0.6150.85±1.4347.08±0.3348.37±0.8544.50±1.2359.62±0.7663.68±0.86
IPM49.03±0.1846.44±0.3351.82±0.4535.07±1.6243.55±1.6050.52±1.9550.93±1.5545.36±2.2149.53±1.8641.07±3.0259.43±1.0661.28±0.91
Min-Max35.89±1.5745.46±1.4950.62±1.4349.61±0.6049.69±1.7148.25±0.7049.36±1.9748.21±2.0338.51±3.0912.36±0.4860.82±0.5764.59±1.02
FEMNIST-70.17±1.8169.60±0.2369.95±0.8868.76±0.9369.95±0.4737.06±1.4669.31±1.1273.14±0.8173.33±0.9059.78±1.7177.12±0.6777.88±0.31
LF68.08±1.5870.37±0.4469.27±1.5668.70±0.7668.29±1.3133.88±0.6269.90±0.5169.78±0.3770.82±0.5452.68±1.9873.44±0.7776.94±0.45
IPM66.69±1.0659.88±1.9067.99±1.8455.02±0.8555.39±0.2838.81±1.0967.07±1.0168.44±2.0770.34±1.0350.37±1.6373.94±1.1574.95±0.87
Min-Max67.46±1.5265.17±1.8568.62±1.3865.38±1.3265.85±1.1236.70±0.2768.05±1.5171.29±1.0870.19±2.1121.32±0.2276.13±1.6277.74±0.65
CIFAR-100-35.92±0.6734.80±0.5434.48±0.3834.34±0.2425.64±0.9531.53±0.4233.32±0.1335.36±0.5634.70±0.752.86±0.4754.45±0.8153.55±0.59
LF30.83±0.3334.37±0.4933.46±0.2729.28±0.5826.98±0.3731.11±0.4432.99±0.4229.85±0.9132.53±0.613.21±0.9642.32±0.6543.79±0.49
IPM25.54±1.2119.29±0.7331.38±0.5315.61±0.6425.89±0.4529.43±0.2632.44±0.5925.55±0.4823.62±0.333.24±1.2841.56±0.7545.13±1.09
Min-Max22.31±1.0216.38±0.2934.23±0.6219.46±0.9327.21±0.3830.16±1.2534.73±0.6721.04±0.4334.39±0.473.20±0.4146.05±0.3451.11±0.21
" + }, + { + "type": "table_caption", + "bbox": [ + 0.31, + 0.309, + 0.686, + 0.323 + ], + "angle": 0, + "content": "Table 2. Top-1 Accuracy(%) under Different Byzantine Attack." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.35, + 0.483, + 0.381 + ], + "angle": 0, + "content": "and attacks, along with detailed experimental configurations, are provided in Appendix D." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.39, + 0.216, + 0.403 + ], + "angle": 0, + "content": "6.2. Robustness" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.412, + 0.484, + 0.518 + ], + "angle": 0, + "content": "Table 2 shows the accuracy under different Byzantine attackers, with data heterogeneity set at a Dirichlet concentration parameter of \\(\\alpha = 0.5\\) on CIFAR-10 and CIFAR-100, and attackers comprising \\(20\\%\\) of the total number of clients. FLSeg can utilize different RAR for Segment Aggregation; we denote the RAR used by adding it in parentheses after the method name, for example, FLSeg(DnC)." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.53, + 0.484, + 0.635 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.645, + 0.483, + 0.672 + ], + "angle": 0, + "content": "Figure 3. Top-1 Accuracy(%) under Different Ratio of Attacker on CIFAR-10." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.901 + ], + "angle": 0, + "content": "From the empirical results, we observe that FLSeg maintains stable performance under various Byzantine attacks. Additionally, we find that FLSeg's defensive effectiveness is significantly influenced by the choice of RAR. For instance, using DnC, a singular value decomposition (SVD)-based spectral method can achieve higher accuracy than Euclidean distance-based methods like Krum. We also found that the combination of RAR with FLSeg significantly outperforms the original RAR method, especially in complex models, indicating that FLSeg effectively mitigates the curse of dimensionality. For example, Krum nearly fails under Min-Max attacks on CIFAR-100, achieving an accuracy of only \\(16.38\\%\\), but when combined with FLSeg, it achieves a much improved accuracy of \\(46.05\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.35, + 0.907, + 0.532 + ], + "angle": 0, + "content": "Figure 3 shows the accuracy under different numbers of Min-Max attackers, with FLSeg demonstrating greater stability against varying numbers of attackers. Additionally, it is evident that other privacy-preserving robust FL methods experience a notable performance drop as the number of attackers increases. This decline occurs because the metrics these methods use to filter malicious gradients are constrained by their privacy-preserving techniques, making them vulnerable to strong adaptive attacks. In contrast, FLSeg mitigates the curse of dimensionality through Segment Aggregation and enhances resistance to malicious gradients by integrating diverse RAR." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.544, + 0.706, + 0.56 + ], + "angle": 0, + "content": "6.3. Heterogeneous Data" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.907, + 0.673 + ], + "angle": 0, + "content": "Figure 4 shows the accuracy under different levels of data heterogeneity on CIFAR-10. We fix the ratio of attackers at \\(20\\%\\) using Min-Max attackers and control the level of data heterogeneity by adjusting the Dirichlet \\(\\alpha\\). An \\(\\alpha\\) of 0.1 represents an extreme case where each client may only possess one or two labels out of ten possible categories, while at \\(\\alpha = 10\\), the data distribution is nearly IID." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.687, + 0.907, + 0.799 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.809, + 0.907, + 0.838 + ], + "angle": 0, + "content": "Figure 4. Top-1 Accuracy(%) under Different Levels of Data Heterogeneity on CIFAR-10." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We observe that FLSeg achieves significantly better performance than other methods when \\(\\alpha = 0.1\\). This is because other methods were not designed to account for the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3922" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.228 + ], + "angle": 0, + "content": "potential impact of data heterogeneity, whereas FLSeg enables clients to better adapt to their local datasets under non-IID conditions by learning personalized models, thereby achieving improved performance. However, as the data tends to be balanced (\\(\\alpha = 10\\)), the gains brought by personalization will gradually decrease, but the performance of FLSeg is still better than that of other baselines. This is because applying RAR to model segments effectively alleviates the curse of dimensionality." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.24, + 0.187, + 0.255 + ], + "angle": 0, + "content": "6.4. Privacy" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.262, + 0.485, + 0.414 + ], + "angle": 0, + "content": "To verify FLSeg's privacy protection capability against a semi-honest server, we present visual results of attacks using DLG and Inverting-Gradient, with and without our defense, on FEMNIST and CIFAR-100 in Fig. 5. Without privacy protection (Origin), the server can easily reconstruct training data. In contrast, when attacking the mixed model after Segment Exchange, only meaningless information can be extracted. Furthermore, in Sec. 5.1, we prove that the probability of the server reconstructing the original client model is negligible, thereby ensuring client privacy." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.429, + 0.48, + 0.569 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.578, + 0.483, + 0.607 + ], + "angle": 0, + "content": "Figure 5. Attack results of iDLG and Inverting-Gradients on CIFAR-10 and FEMNIST, respectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.635, + 0.317, + 0.65 + ], + "angle": 0, + "content": "6.5.Impact of Segment Size" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.484, + 0.794 + ], + "angle": 0, + "content": "To investigate the impact of segment size on privacy protection and model performance, we vary the segment size \\( |\\gamma| \\) across \\( \\{10^2, 10^3, 10^4, 10^5, 10^6, 2904970(\\ell)\\} \\) under MinMax and Inverting-Gradient attacks on the heterogeneous CIFAR-10 dataset. We use the same settings as in the robustness experiments to evaluate the model's top-1 accuracy and perform an inference attack on the model updates of a single client from one training round. The experimental results are presented in Tab. 3." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.807, + 0.481, + 0.874 + ], + "angle": 0, + "content": "
|γ|102103104105106l
Accuracy63.93%64.33%65.72%60.93%59.67%57.75%
Inference
" + }, + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.883, + 0.48, + 0.898 + ], + "angle": 0, + "content": "Table 3. Accuracy and Privacy of FLSeg(Dnc) under different \\( |\\gamma | \\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.227 + ], + "angle": 0, + "content": "We observe that the accuracy of \\(\\mathrm{FLSeg(DnC)}\\) initially increases and then slightly decreases. We hypothesize that this occurs because benchmarks computed from smaller segments lack representativeness, while larger segments fail to mitigate the curse of dimensionality. Moreover, as the segment size increases, the reconstructed attack results reveal more information about the original data, since larger segments concentrate more original gradient information in the mixed model." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.236, + 0.667, + 0.252 + ], + "angle": 0, + "content": "6.6. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.258, + 0.906, + 0.303 + ], + "angle": 0, + "content": "We finally study how each component contributes to the overall performance of FLSeg. Specifically, we measure the following scenarios:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.304, + 0.734, + 0.317 + ], + "angle": 0, + "content": "FedAvg. The baseline method." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.318, + 0.882, + 0.333 + ], + "angle": 0, + "content": "- PT-only. Adds local personalized training to FedAvg." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.334, + 0.854, + 0.348 + ], + "angle": 0, + "content": "- SA-only. Adds Segment Aggregation to FedAvg." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.349, + 0.78, + 0.363 + ], + "angle": 0, + "content": "- FLSeg. The complete FLSeg method." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.304, + 0.882, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.906, + 0.41 + ], + "angle": 0, + "content": "Note that we did not include Segment Exchange as its purpose is to protect privacy without compromising the correctness of subsequent aggregation." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.42, + 0.907, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.634, + 0.516, + 0.785, + 0.53 + ], + "angle": 0, + "content": "Figure 6. Ablation study." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.546, + 0.907, + 0.667 + ], + "angle": 0, + "content": "We conducted ablation experiments under four scenarios combining with/without \\(20\\%\\) Min-Max attackers and IID/non-IID \\((\\alpha = 0.5)\\) CIFAR-10 dataset, with the results shown in Fig. 6. We observe that, in all scenarios, Segment Aggregation effectively resists Byzantine attacks. Meanwhile, the incorporation of personalized training does not degrade SA-only performance under IID data, and improves the model performance under non-IID data." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.68, + 0.634, + 0.695 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.906, + 0.84 + ], + "angle": 0, + "content": "In this work, we identified three key challenges in privacy-preserving robust FL: high computational costs, limited flexibility in RAR, and data heterogeneity. To overcome these, we proposed FLSeg, which uses Segment Exchange and Segment Aggregation to ensure privacy without limiting aggregation choices and employs a regularization term to enhance adaptation to heterogeneous data. Theoretical and experimental results validate FLSeg's security, robustness, and resilience to data heterogeneity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.906, + 0.887 + ], + "angle": 0, + "content": "In the future, we will explore stronger threat models, such as malicious server, or extend the privacy-preserving segment exchange framework to decentralized settings." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3923" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.484, + 0.241 + ], + "angle": 0, + "content": "This research was supported in part by the Major Research Plan of Hubei Province (2023BAA027), the Changsha Key Science and Technology Project under the \"Unveiling and Commanding\" Initiative (kq2503009), the Science, Technology and Innovation Commission of Shenzhen Municipality (GJHZ20240218114659027), the Key Research and Development Plan of Hubei Province (2024BAB049), and the Fundamental Research Funds for the Central Universities (YCJJ20252336)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.266, + 0.188, + 0.281 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.29, + 0.484, + 0.345 + ], + "angle": 0, + "content": "[1] Abbas Acar, Hidayet Aksu, A Selcuk Uluagac, and Mauro Conti. A survey on homomorphic encryption schemes: Theory and implementation. ACM Computing Surveys (Csur), 51(4):1-35, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.347, + 0.484, + 0.429 + ], + "angle": 0, + "content": "[2] James Bell, Adrià Gascon, Tancrède Lepoint, Baiyu Li, Sarah Meiklejohn, Mariana Raykova, and Cathie Yun. ACORN: input validation for secure aggregation. In 32nd USENIX Security Symposium, USENIX Security 2023, Anaheim, CA, USA, August 9-11, 2023, pages 4805-4822. USENIX Association, 2023. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.43, + 0.484, + 0.512 + ], + "angle": 0, + "content": "[3] James Henry Bell, Kallista A. Bonawitz, Adrià Gasón, Tancrède Lepoint, and Mariana Raykova. Secure single-server aggregation with (poly)logarithmic overhead. In CCS '20: 2020 ACM SIGSAC Conference on Computer and Communications Security, Virtual Event, USA, November 9-13, 2020, pages 1253-1269. ACM, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.513, + 0.484, + 0.568 + ], + "angle": 0, + "content": "[4] Peva Blanchard, El Mahdi El Mhamdi, Rachid Guerraoui, and Julien Stainer. Machine learning with adversaries: Byzantine tolerant gradient descent. Advances in neural information processing systems, 30, 2017. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.568, + 0.484, + 0.651 + ], + "angle": 0, + "content": "[5] Keith Bonawitz, Vladimir Ivanov, Ben Kreuter, Antonio Marcedone, H Brendan McMahan, Sarvar Patel, Daniel Ramage, Aaron Segal, and Karn Seth. Practical secure aggregation for privacy-preserving machine learning. In proceedings of the 2017 ACM SIGSAC Conference on Computer and Communications Security, pages 1175-1191, 2017. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.652, + 0.484, + 0.707 + ], + "angle": 0, + "content": "[6] Sebastian Caldas, Sai Meher Karthik Duddu, Peter Wu, Tian Li, Jakub Konečný, H Brendan McMahan, Virginia Smith, and Ameet Talwalkar. Leaf: A benchmark for federated settings. arXiv preprint arXiv:1812.01097, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.708, + 0.484, + 0.748 + ], + "angle": 0, + "content": "[7] Ran Canetti. Universally composable security: A new paradigm for cryptographic protocols. In FOCS, pages 136-145. IEEE Computer Society, 2001. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.749, + 0.484, + 0.79 + ], + "angle": 0, + "content": "[8] Xiaoyu Cao, Minghong Fang, Jia Liu, and Neil Zhenqiang Gong. Fltrust: Byzantine-robust federated learning via trust bootstrapping. In NDSS. The Internet Society, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.791, + 0.484, + 0.845 + ], + "angle": 0, + "content": "[9] Alireza Fallah, Aryan Mokhtari, and Asuman E. Ozdaglar. Personalized federated learning with theoretical guarantees: A model-agnostic meta-learning approach. In NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[10] Minghong Fang, Xiaoyu Cao, Jinyuan Jia, and Neil Gong. Local model poisoning attacks to {Byzantine-Robust} federated learning. In 29th USENIX security symposium (USENIX Security 20), pages 1605-1622, 2020. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.29, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.149 + ], + "angle": 0, + "content": "[11] Minghong Fang, Xiaoyu Cao, Jinyuan Jia, and Neil Zhenqiang Gong. Local model poisoning attacks to byzantine-robust federated learning. In USENIX Security Symposium, pages 1605–1622. USENIX Association, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[12] Marc Fischlin and Arno Mittelbach. An overview of the hybrid argument. IACR Cryptol. ePrint Arch., page 88, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[13] Jonas Geiping, Hartmut Bauermeister, Hannah Droge, and Michael Moeller. Inverting gradients-how easy is it to break privacy in federated learning? Advances in neural information processing systems, 33:16937-16947, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[14] Oded Goldreich, Silvio Micali, and Avi Wigderson. How to play any mental game or A completeness theorem for protocols with honest majority. In STOC, pages 218-229. ACM, 1987. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.361 + ], + "angle": 0, + "content": "[15] Rachid Guerraoui, Sébastien Rouault, et al. The hidden vulnerability of distributed learning in byzantium. In International Conference on Machine Learning, pages 3521-3530. PMLR, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.419 + ], + "angle": 0, + "content": "[16] Anbu Huang, Yang Liu, Tianjian Chen, Yongkai Zhou, Quan Sun, Hongfeng Chai, and Qiang Yang. Starfl: Hybrid federated learning architecture for smart urban computing. ACM Trans. Intell. Syst. Technol., 12(4):43:1-43:23, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.476 + ], + "angle": 0, + "content": "[17] Najeeb Moharram Jebreel, Josep Domingo-Ferrer, Alberto Blanco-Justicia, and David Sánchez. Enhanced security and privacy via fragmented federated learning. IEEE Transactions on Neural Networks and Learning Systems, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.477, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[18] Peter Kairouz, H Brendan McMahan, Brendan Avent, Aurélien Bellet, Mehdi Bennis, Arjun Nitin Bhagoji, Kallista Bonawitz, Zachary Charles, Graham Cormode, Rachel Cummings, et al. Advances and open problems in federated learning. Foundations and trends® in machine learning, 14(1-2): 1-210, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.561, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[19] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank Reddi, Sebastian Stich, and Ananda Theertha Suresh. SCAFFOLD: Stochastic controlled averaging for federated learning. In Proceedings of the 37th International Conference on Machine Learning, pages 5132-5143. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.645, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[20] Sai Praneeth Karimireddy, Lie He, and Martin Jaggi. Byzantine-robust learning on heterogeneous datasets via bucketing. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.716, + 0.905, + 0.745 + ], + "angle": 0, + "content": "[21] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[22] Shenghui Li, Edith C-H Ngai, and Thiemo Voigt. An experimental study of byzantine-robust aggregation schemes in federated learning. IEEE Transactions on Big Data, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[23] Shenghui Li, Edith Ngai, Fanghua Ye, Li Ju, Tianru Zhang, and Thiemo Voigt. Blades: A unified benchmark suite for byzantine attacks and defenses in federated learning. In 2024 IEEE/ACM Ninth International Conference on Internet-of-Things Design and Implementation (IoTDI), 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.873, + 0.905, + 0.903 + ], + "angle": 0, + "content": "[24] Tian Li, Shengyuan Hu, Ahmad Beirami, and Virginia Smith. Ditto: Fair and robust federated learning through per" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3924" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "sonalization. In International conference on machine learning, pages 6357-6368. PMLR, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.191 + ], + "angle": 0, + "content": "[25] Xiaoyuan Liu, Hongwei Li, Guowen Xu, Zongqi Chen, Xiaoming Huang, and Rongxing Lu. Privacy-enhanced federated learning against poisoning adversaries. IEEE Transactions on Information Forensics and Security, 16:4574-4588, 2021. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[26] Yuchen Liu, Chen Chen, Lingjuan Lyu, Fangzhao Wu, Sai Wu, and Gang Chen. Byzantine-robust learning on heterogeneous data via gradient splitting. In International Conference on Machine Learning, pages 21404–21425. PMLR, 2023. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[27] Zhi Lu, SongFeng Lu, YongQuan Cui, XueMing Tang, and JunJun Wu. Split aggregation: Lightweight privacy-preserving federated learning resistant to byzantine attacks. IEEE Transactions on Information Forensics and Security, 19:5575-5590, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[28] Hidde Lycklama, Lukas Burkhalter, Alexander Viand, Nicolas Kuchler, and Anwar Hithnawi. Rofl: Robustness of secure federated learning. In SP, pages 453-476. IEEE, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[29] Zhuoran Ma, Jianfeng Ma, Yinbin Miao, Yingjiu Li, and Robert H Deng. Shieldfl: Mitigating model poisoning attacks in privacy-preserving federated learning. IEEE Transactions on Information Forensics and Security, 17:1639-1654, 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.461, + 0.482, + 0.53 + ], + "angle": 0, + "content": "[30] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.532, + 0.482, + 0.601 + ], + "angle": 0, + "content": "[31] Dinh C. Nguyen, Quoc-Viet Pham, Pubudu N. Pathirana, Ming Ding, Aruna Seneviratne, Zihuai Lin, Octavia A. Dobre, and Won-Joo Hwang. Federated learning for smart healthcare: A survey. ACM Comput. Surv., 55(3):60:1-60:37, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.603, + 0.482, + 0.645 + ], + "angle": 0, + "content": "[32] Krishna Pillutla, Sham M Kakade, and Zaid Harchaoui. Robust aggregation for federated learning. IEEE Transactions on Signal Processing, 70:1142-1154, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[33] Virat Shejwalkar and Amir Houmansadr. Manipulating the byzantine: Optimizing model poisoning attacks and defenses for federated learning. In NDSS, 2021. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[34] Virat Shejwalkar, Amir Houmansadr, Peter Kairouz, and Daniel Ramage. Back to the drawing board: A critical evaluation of poisoning attacks on production federated learning. In SP, pages 1354-1371. IEEE, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[35] Lichao Sun, Jianwei Qian, and Xun Chen. LDP-FL: practical private aggregation in federated learning with local differential privacy. In IJCAI, pages 1571–1578. ijcai.org, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[36] Canh T Dinh, Nguyen Tran, and Josh Nguyen. Personalized federated learning with moreau envelopes. Advances in Neural Information Processing Systems, 33:21394-21405, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[37] Kang Wei, Jun Li, Ming Ding, Chuan Ma, Hang Su, Bo Zhang, and H. Vincent Poor. User-level privacy-preserving federated learning: Analysis and performance optimization. IEEE Trans. Mob. Comput., 21(9):3388-3401, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[38] Jiahui Wu, Weizhe Zhang, and Fucai Luo. On the security of \"lsfl: A lightweight and secure federated learning scheme for edge computing\". IEEE Trans. Inf. Forensics Secur., 19: 3481-3482, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[39] Jue Xiao, Xueming Tang, and Songfeng Lu. Privacy-preserving federated class-incremental learning. IEEE Transactions on Machine Learning in Communications and Networking, 2:150-168, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[40] Cong Xie, Oluwasanmi Koyejo, and Indranil Gupta. Fall of empires: Breaking byzantine-tolerant sgd by inner product manipulation. In Uncertainty in Artificial Intelligence, pages 261-270. PMLR, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[41] Dong Yin, Yudong Chen, Ramchandran Kannan, and Peter Bartlett. Byzantine-robust distributed learning: Towards optimal statistical rates. In International Conference on Machine Learning, pages 5650–5659. Pmlr, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[42] Bo Zhao, Konda Reddy Mopuri, and Hakan Bilen. idlg: Improved deep leakage from gradients. CoRR, abs/2001.02610, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[43] Ligeng Zhu, Zhijian Liu, and Song Han. Deep leakage from gradients. Advances in neural information processing systems, 32, 2019. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3925" + } + ] +] \ No newline at end of file diff --git a/2025/FLSeg_ Enhancing Privacy and Robustness in Federated Learning under Heterogeneous Data via Model Segmentation/a4f90945-afa0-4f01-8edd-1b86f5d9d098_origin.pdf b/2025/FLSeg_ Enhancing Privacy and Robustness in Federated Learning under Heterogeneous Data via Model Segmentation/a4f90945-afa0-4f01-8edd-1b86f5d9d098_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1e50c492e69672025355883d53fed558743fb5b1 --- /dev/null +++ b/2025/FLSeg_ Enhancing Privacy and Robustness in Federated Learning under Heterogeneous Data via Model Segmentation/a4f90945-afa0-4f01-8edd-1b86f5d9d098_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cb2efe99071594018604331ffc975f4772c43c9fca013b1f7eefc8eb17073785 +size 1486960 diff --git a/2025/FLSeg_ Enhancing Privacy and Robustness in Federated Learning under Heterogeneous Data via Model Segmentation/full.md b/2025/FLSeg_ Enhancing Privacy and Robustness in Federated Learning under Heterogeneous Data via Model Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0539e5a2ddbb59bdb516f535fbd08212380e4206 --- /dev/null +++ b/2025/FLSeg_ Enhancing Privacy and Robustness in Federated Learning under Heterogeneous Data via Model Segmentation/full.md @@ -0,0 +1,352 @@ +# FLSeg: Enhancing Privacy and Robustness in Federated Learning under Heterogeneous Data via Model Segmentation + +Zichun $\mathbf{Su}^{1\dagger}$ Zhi Lu $^{1\dagger}$ Yutong Wu $^{1}$ Renfei Shen $^{1}$ Songfeng Lu $^{1,2*}$ + +$^{1}$ School of Cyber Science and Engineering, Huazhong University of Science and Technology $^{2}$ Shenzhen Huazhong University of Science and Technology Research Institute + +suzc25@m.fudan.edu.cn, {luzhi,wyt,shenrenfei,lusongfeng}@hust.edu.cn + +# Abstract + +Federated Learning (FL) enables collaborative training of a global model without data sharing, yet it faces critical challenges from privacy leakage and Byzantine attacks. Existing privacy-preserving robust FL frameworks suffer from three key limitations: high computational costs, restricted application of Robust Aggregation Rules (RAR), and inadequate handling of data heterogeneity. To address these limitations, we propose the FLSeg framework, which leverages Segment Exchange and Segment Aggregation to avoid excessive encryption computations while enabling unrestricted use of any RAR. Additionally, a regularization term in local training balances personalization with global model performance, effectively adapting to heterogeneous data. Our theoretical and experimental analyses demonstrate FLSeg's semi-honest security and computational efficiency: it achieves client and server time complexities of $O(\ell)$ and $O(n\ell)$ , with empirical results showing significantly reduced computational times. Extensive experiments further confirm FLSeg's robustness across diverse heterogeneous and adversarial scenarios. + +# 1. Introduction + +Federated Learning (FL) allows organizations to collaboratively train models without sharing sensitive data. By exchanging model updates instead of raw data, FL enhances privacy and efficiency, making it ideal for fields like finance [18], healthcare [16], and smart cities [31]. However, FL is currently threatened by two major issues: privacy leakage and Byzantine attacks. Privacy leakage occurs when the server can infer clients' sensitive training data by analyzing the model updates [13, 43]. Byzantine attacks, on the other hand, involve malicious clients submitting incorrect or manipulated updates, which can disrupt or compromise the global model's performance [23, 34]. To address + +these challenges, corresponding techniques have been developed. For privacy protection, methods like Homomorphic Encryption (HE) [29], Secure Multi-Party Computation (SMPC) [5], and Differential Privacy (DP) [35, 39] aim to safeguard training data by ensuring that only encrypted or obfuscated information is shared. For robustness against Byzantine attacks, various Robust Aggregation Rule (RAR), such as Krum [4], DnC [33], and RFA [32], help identify and exclude malicious updates, thereby protecting the global model from adversarial influence. + +However, achieving both robustness and privacy simultaneously presents a significant challenge, as privacy-preserving methods can unintentionally protect malicious updates from detection [27]. To address this issue, a series of Privacy-preserving and Robust FL (PRFL) frameworks have been proposed. These methods explore various architectures and protocols or select encryption schemes that are compatible with specific RAR [2, 25, 28, 29, 38]. + +Despite these methods balancing robustness and privacy, they still have notable limitations: (1) Privacy-preserving techniques such as HE and SMPC introduce significant computational overhead [27]. (2) These methods impose restrictions on the use of RAR, supporting only simple, single RAR, which are vulnerable to adaptive attacks such as Min-Max [33] and IPM [40]. (3) Gradient heterogeneity caused by data heterogeneity significantly degrades both model performance and the accuracy of RAR [22]. However, existing PRFL methods largely overlook this issue. + +Therefore, we aim to address the following question: "How can we efficiently preserve privacy while avoiding restrictions on RAR implementation and mitigating the impact of data heterogeneity on robustness and performance?" To this end, we draw inspiration from existing theoretical insights: reducing data dimensionality can alleviate the curse of dimensionality and gradient heterogeneity, thereby enhancing robustness [26, 33]. Based on this principle, we propose FLSeg, a model segmentation-based framework. We design a Segment Exchange privacy protocol to achieve privacy preservation without incurring excessive computa + +tional overhead. Additionally, we introduce a Segment Aggregation algorithm, which applies RAR operations at the low-dimensional model segments to mitigate gradient heterogeneity and the curse of dimensionality while supporting arbitrary RAR. Furthermore, we incorporate a regularization term into local training to facilitate the learning of personalized models, further improving performance under data heterogeneity. Our contributions are threefold: + +- A Novel Privacy-Preserving Robust FL Framework. To the best of our knowledge, FLSeg is the first PRFL framework that does not restrict the choice of RAR, while also mitigating the impact of heterogeneous data. +- Computationally Efficient Privacy Protection. We theoretically demonstrate that our FLSeg satisfies semihonest security. We further provide a detailed analysis of FLSeg's computational time complexity, with the client-side and server-side complexities being $O(\ell)$ and $O(n\ell)$ , respectively. In practical tests, FLSeg's client computation time is only 233 ms, compared to 1696 ms for SOTA method ACORN [2]. The server computation time per client is only 78 ms, whereas ACORN require 181 ms per client. +- Extensive Experimental Analysis. Extensive experiments on three benchmark datasets show that our FLSeg can significantly improve the performance under varying degrees of data heterogeneity and in Byzantine attack, e.g., the accuracy of ACORN is $48.21\%$ , while FLSeg achieves $64.59\%$ under a $20\%$ proportion of Min-Max attackers and on non-IID CIFAR-10 datasets. Further inference attack experiments also verified the privacy protection capability of FLSeg. + +# 2. Related Works + +Privacy Threats and Defences. Existing studies have demonstrated that gradient sharing in FL poses severe privacy risks. DLG attacks [42, 43] and Inverting Gradients attacks [13] can precisely reconstruct raw training data from client-uploaded gradients. + +To mitigate these risks, privacy-preserving techniques such as DP [37], HE [25], and SMPC [5] have been proposed to obfuscate or encrypt gradients. However, these approaches have notable limitations: DP inevitably reduce model utility due to noise injection, while HE and SMPC incur high computational overhead due to complex encryption structures and protocols [1]. + +Byzantine Attack and Defence. Byzantine attacks are initiated by malicious clients who manipulate or send adversarial model updates to degrade model performance [23]. A common type of such attacks is data poisoning, such as LabelFlipping [10], which manipulates training data to mislead the model into learning incorrect associations. Additionally, adaptive attacks, such as Min-Max [33] and IPM [40], optimize the magnitude and direction of malicious + +gradient perturbations to maximize the disruption of the global FL model while simultaneously evading detection. + +To mitigate these risks, RAR modify the standard mean aggregation method [30] by performing statistical or feature analysis on the collected updates to filter out malicious ones. Examples include methods based on Euclidean distance, such as Krum [4] and Bulyan [15]; those leveraging coordinate values, such as RFA [32] and Median [41]; and approaches utilizing spectral anomaly detection, such as DnC [33]. Beyond RAR, Robust FL integrate auxiliary techniques for enhanced defense. FLTrust [8] uses a root dataset to assess the quality of client updates, ClippedClustering [22] leverages historical gradient information to constrain and clip gradients before aggregation, and GAS [26] proposes a gradient splitting technique to reduce data dimensionality and enhance the robustness of RAR. However, these defenses do not secure the gradients during transmission, rendering them vulnerable to privacy threats. + +Privacy-preserving Robust FL. PRFL integrates privacy-preserving techniques with RAR through specially designed protocols. PEFL [25] and ShieldFL [29] employ HE to ensure privacy while using cosine similarity to detect malicious updates. FFL [17] proposes a parameter mixing method to protect privacy while applying cosine similarity to the mixed model for robustness. ACORN [2] and RoFL [28] utilize the SecAgg SMPC protocol to safeguard the privacy of inputs while employing zero-knowledge proofs to bound the norm of model updates, thereby mitigating the impact of malicious updates. However, the high computational cost, restricted support for RAR, and lack of consideration for data heterogeneity limit their applicability. + +# 3. Preliminaries + +Notion. For any positive integer $n \in \mathbb{N}^+$ , we denote the set $\{1, \ldots, n\}$ by $[n]$ . The cardinality of a set $w$ is denoted by $|w|$ . We denote the $l_2$ norm of vector $w$ by $\| w \|$ . + +Model Segmentation. For a model $w$ , the mode of segmentation refers to a partitioning of the set of model parameters, $\{\gamma^{(1)},\gamma^{(2)},\dots,\gamma^{(d)}\}$ , where different segments are disjoint, and $w = \bigcup_{k = 1}^{d}\gamma^{(k)}$ . + +System Setup. We consider a classic FL scenario [30], involving $n$ clients, denoted by $\{c_i\}_{i \in [n]}$ , and a single server $S$ . $c_i$ has a unique local dataset $D_i$ that is non-IID. $S$ starts with initial model weights $w_0$ and distributes the model to each $c_i$ . $c_i$ train the model on their datasets and send their updates $w_i^t$ to the server $S$ , which aggregates these updates and proceeds to the next round. After $T$ rounds, this process results in a global model $w^T$ . + +Threat Model. Our system faces two primary threats. First, in line with [8, 11], there may be Byzantine clients within the system who manipulate their uploaded model to disrupt training. They can launch targeted attacks such as LabelFlipping [10], and powerful adaptive attacks like Min + +![](images/36b96f68677cefb5702ab0523844da931eaec0dc81188ae03e627f21df1580fd.jpg) +Figure 1. The Overview of FLSeg. + +Max [33]. Second, the server is semi-honest, adhering to the aggregation protocol while attempting to infer private local data from the client's updates [29]. As in prior work [2, 25, 27, 29], we assume that the adversary can control a number of clients $f < (n - 1) / 2$ and semi-honest server and Byzantine clients do not collude. + +# 4. Methodology + +# 4.1. Intuition + +Existing RAR methods commonly suffer from the curse of dimensionality [15], where malicious gradients can hide within high-dimensional vectors. Additionally, data heterogeneity introduces gradient heterogeneity [19], further increasing the difficulty of detection. Studies have shown that reducing input dimensionality can help alleviate this issue, for example, by processing gradients through sub-sampling [33], grouping [20], or splitting [26]. + +Based on this insight, we propose an intuitive approach: by fixing the mode of segmentation, each client selects a partner to exchange model segments, while the server applies the same segmentation mode during aggregation. In this way, as long as the server is unaware of the specific exchange pairs or the exchanged segments, it cannot reconstruct the client models, thereby ensuring privacy protection. Additionally, the server apply RAR on each segment, which not only reduces the curse of dimensionality and gradient heterogeneity but also ensures that each segment to which the RAR is applied originates from different clients, thereby ensuring the correctness of the RAR implementation. Besides, although these processes mitigate gradient heterogeneity caused by Non-IID data [15, 20], they fail to address the performance degradation induced by Non- + +IID data. To tackle this issue, we incorporate local regularization techniques from personalized FL [24, 36] to improve model performance without altering the overall process. The whole process is illustrated in Fig. 1. + +# 4.2. Non-IID Data: Learn Personalized Model + +Heterogeneous data can cause the global model to fail to adapt to all clients' data distributions, leading to performance degradation. Existing personalized FL methods address this issue by balancing the global and local models to integrate the generalization ability of the global model with the uniqueness of local data [9, 24, 36]. Inspired by this approach, we modify each client's local training process by allowing the client to maintain a personalized model $v_{i}$ , with its objective function defined in Eq. (1). + +$$ +\min _ {v _ {i}} h _ {i} \left(v _ {i}, w ^ {t}\right) = \mathcal {L} _ {i} \left(v _ {i}\right) + \frac {\lambda}{2} \left\| v _ {i} - w ^ {t} \right\| ^ {2}, \tag {1} +$$ + +where $w^{t}$ represents the global model received in each round; $\mathcal{L}_i(w) = \mathbb{E}_{\xi_i \sim \mathcal{D}_i}[f_i(w; \xi_i)]$ denotes the loss function on $c_{i}$ 's local data; $\lambda$ is a regularization parameter that balances the influence of the personalized and global models. When $\lambda \to +\infty$ , the personalized model converges to the global model, whereas $\lambda \to 0$ leads to a more distinct personalized effect. + +The detailed process is provided in Algorithm 1. During training, the personalized model is updated synchronously with the global model. Each client retains $v_{i}^{t + 1}$ locally for local tasks and prepares the model update $w_{i}^{t}$ for the subsequent Segment Exchange. + +# 4.3. Privacy: Segment Exchange + +To prevent the server from obtaining each client's complete model, the client mixes its model by exchanging segments + +# Algorithm 1: Local Training + +Input : Personalized model $v_{i}^{t}$ , model received from server $w^{t}$ , regularization coefficient $\lambda$ , learning rate $\eta$ , local epochs $E$ + +Output: Updates $w_{i}^{t + 1}$ , personalized model $v_{i}^{t + 1}$ + +1 $w_{i}^{t + 1}\gets w^{t},v_{i}^{t + 1}\leftarrow v_{i}^{t};$ +2 for epoch $\leftarrow 0$ to $E - 1$ do +3 /*update personalized model \*/ $v_{i}^{t + 1}\gets v_{i}^{t + 1} - \eta (\nabla \mathcal{L}_{i}(v_{i}^{t + 1}) + \lambda (v_{i}^{t + 1} - w^{t}))$ +4 $\begin{array}{r}\left\lfloor \begin{array}{c}/*\mathrm{update~global~model}*/\\ w_i^{t + 1}\gets w_i^{t + 1} - \eta \nabla \mathcal{L}_i(w_i^{t + 1}); \end{array} \right. \end{array}$ +5 return $w_{i}^{t + 1}, v_{i}^{t + 1}$ + +with a partner before submitting local updates. However, to ensure that clients do not reveal their models to their exchange partners during this process, we design a secure model segment exchange protocol, detailed in Algorithm 2. + +An simple example is illustrated in Fig. 2. The two parties first use the Diffie-Hellman key exchange to agree on a mask $m$ that determines which segments are exchanged. $w_{i}^{-m}$ is encrypted using client $c_{i}$ 's key $r_{i}$ , ensuring that $c_{j}$ cannot access $c_{i}$ 's private information while still enabling segment exchange through the properties of the XOR operation. Meanwhile, the server can decrypt the client key to obtain the mixed model. However, since it does not know $m$ or the exchange partner, it cannot recover the original client model, thereby preserving privacy. + +![](images/1c85a6a0161442e27f2aed6784eaae72fed3269303ba0b7ac9d0515c4d4ce0b9.jpg) +Figure 2. The Workflow of Segment Exchange. + +# 4.4. Robustness: Segment Aggregation + +To mitigate the curse of dimensionality, FLSeg partitions the mixed model according to the mode of segmentation and then applies RAR to each segment, rather than performing RAR directly on the entire mixed model. This approach not only ensures that each segment in the RAR originates from different clients but also effectively mitigates the challenges posed by high input dimensionality. The detailed procedure is provided in Algorithm 3. + +After grouping segments based on the mode of segmen + +Algorithm 2: Segment Exchange Protocal +Input: Client $c_{i}$ 's model update $w_{i}$ ,server's public-private key pair $(pk_S,sk_S)$ ,the exchange partner $C_j$ +1 client $c_{i}$ and $c_{j}$ +2 generate client's key $r_i$ . +3 perform Diffie-Hellman key exchange with $c_{j}$ to obtain mask $m$ . +4 $w_{i}^{m} / w_{i}^{-m}\gets$ Set the $k$ -th segment $\gamma^{(k)} = 0$ of $w_{i}$ if the $k$ -th bit of $m$ is $1 / 0$ . +5 exchange $w_{i}^{-m}\oplus r_{i}$ and $\mathrm{Enc}(r_i,pk_S)$ with $c_{j}$ . +6 compute $(e_i)_{\mathrm{mix}} = w_i^m\oplus (w_j^{-m}\oplus r_j)$ $= (w_{i})_{\mathrm{mix}}\oplus r_{j}$ . +7 send $(e_i)_{\mathrm{mix}}$ ,Enc $(r_j,pk_S)$ to server S; +8 server S +9 decrypts each client $c_{i}$ 's key $r_j = \mathrm{Dec}(\mathrm{Enc}(r_j,pk_S),sk_S)$ . +10 decrypts each client $c_{i}$ 's mixed model $(w_{i})_{\mathrm{mix}}\leftarrow (e_{i})_{\mathrm{mix}}\oplus r_{j};$ + +Algorithm 3: Segment Aggregation +Input: uploaded mixed models $\{(w_{i})_{\mathrm{mix}}\}_{i\in [n]}$ the mode of segmentation $\{\gamma^{(1)},\dots ,\gamma^{(d)}\}$ Output: the aggregation result $w$ +1 for each model $(w_{i})_{\mathrm{mix}}$ do +2 split model into segments $\{\gamma_i^{(1)},\ldots ,\gamma_i^{(d)}\} \gets (w_i)_{\mathrm{mix}};$ +3 for each segment $\gamma^{(k)}\in \{\gamma^{(1)},\dots ,\gamma^{(d)}\}$ do +4 apply the RAR to the segments to establish a benchmark $\overline{\gamma}^{(k)}\gets \mathrm{RAR}(\{\gamma_i^{(k)}\}_{i\in [n]})$ +5 calculate the similarity $s_i^{(k)}$ for each segment $\gamma_{i}^{(k)}$ with the benchmark $\overline{\gamma}^{(k)}$ using Eq. (2); +6 compute each model's similarity score as $s_i = \sum_{k = 1}^d s_i^{(k)}$ and select the top $n - 2f$ models with the highest scores to form set $I$ +7 aggregate $w\gets \sum_{i\in I}(w_i)_{\mathrm{mix}}$ +8 return $w$ + +tation, $S$ applies RAR to each segment group to obtain the benchmark $\overline{\gamma}^{(k)}$ and uses it as an honest reference to compute the similarity $s_i^{(k)}$ of each segment $\gamma_i^{(k)}$ using Eq. (2), then sums the segment similarities to obtain the model's score $s_i$ for this round. + +$$ +s _ {i} ^ {(k)} = \frac {\left(\gamma_ {i} ^ {(k)} - \mu\right) \cdot \left(\bar {\gamma} ^ {(k)} - \mu\right)}{\| \gamma_ {i} ^ {(k)} - \mu \| \| (\bar {\gamma} ^ {(k)} - \mu) \|}, \tag {2} +$$ + +where $z = |\gamma^{(k)}|$ is the number of parameters of $\gamma_{i}^{(k)}$ and + +$\overline{\gamma}^{(k)}$ , and $\mu = \frac{\sum_{j = 1}^{z}\gamma_{i,j}^{(k)} + \sum_{j = 1}^{z}\overline{\gamma}_{j}^{(k)}}{2z}$ is the average value of all parameters in $\gamma_i^{(k)}$ and $\overline{\gamma}^{(k)}$ + +Since, in the worst case, each honest client's partner may be a malicious client, $S$ select the top $n - 2f$ clients with the highest scores for aggregation. + +# 4.5. Summary + +The mode of segmentation in FLSeg remains unchanged throughout the process. Before training begins, the server distributes its public key $pk_{S}$ to all clients. In each training round, clients first perform Local Training (Algorithm 1), then conduct Segment Exchange to obtain the mixed model (Algorithm 2), and finally, the server applies Segment Aggregation (Algorithm 3) to the received global models. The overall process is detailed in the algorithm provided in the Appendix A. + +FLSeg does not restrict how clients select partners; however, to prevent honest clients from continuously collaborating with malicious clients, we design a reputation system based on computed similarity scores. Specifically, each client $c_{i}$ initializes its reputation list $R_{i} = \{r_{1} = 0,\dots ,r_{n} = 0\}$ . After each round of aggregation, the server normalizes the similarity scores of each model using Eq. (3) and sends these scores to the corresponding clients. A lower score for an honest client indicates that its model may have been influenced by malicious updates. + +$$ +s _ {i} ^ {\prime} = 2 \left(\frac {s _ {i} - \operatorname* {m i n} \left(\left\{s _ {i} \right\} _ {i \in [ n ]}\right)}{\operatorname* {m a x} \left(\left\{s _ {i} \right\} _ {i \in [ n ]}\right) - \operatorname* {m i n} \left(\left\{s _ {i} \right\} _ {i \in [ n ]}\right)}\right) - 1 \tag {3} +$$ + +Clients use these scores to update the reputation values of their partners as $r_j = r_j + s_i'$ . In subsequent rounds, clients can select partners with higher reputation values from their reputation lists for Segment Exchange. Practical test results for the reputation system are provided in Appendix B. + +# 5. Analysis + +# 5.1. Semi-Honest Security + +We use the Universally Composable Security Framework [7] to prove that FLSeg satisfies semi-honest security, meaning that FLSeg does not leak any information as long as the adversaries follow the protocol. We first present definition of semi-honest security [14]: + +Definition 1 (Real-World). For protocol $\pi$ , there exists a semi-honest adversary $\mathcal{A}$ that can control a subset of participants $\mathcal{C}$ . $\mathcal{A}$ will follow the protocol $\pi$ but will attempt to infer the private inputs $\{x_1, \ldots, x_n\}$ based on the available information. The probability distribution of the inference results is denoted as $\text{Real}_{\pi, \mathcal{A}}(\mathcal{P}, \{x_1, \ldots, x_n\})$ . + +Definition 2 (Ideal-World). Assume there exists an ideal functionality $\mathcal{F}$ that can execute any interaction in protocol $\pi$ with absolute security. There also exists a simulator Sim that simulates interactions $\mathcal{F}$ in the same manner as the real-world adversary $\mathcal{A}$ . The probability distribution of the inference results is denoted as Ideal $_{\mathcal{F},\text{Sim}}(\mathcal{P},\{x_1,\ldots,x_n\})$ . + +Definition 3 (Semi-Honest). If $\text{Real}_{\pi, \mathcal{A}}(\mathcal{P}, \{x_1, \ldots, x_n\})$ and $\text{Ideal}_{\mathcal{F}, \text{Sim}}(\mathcal{P}, \{x_1, \ldots, x_n\})$ are computationally indistinguishable to an honest-but-curious adversary $\mathcal{A}$ , then protocol $\pi$ is considered semi-honest secure. + +Theorem 1. Assuming the $FLSeg$ protocol is denoted as $\pi$ , the participants of the protocol $\mathcal{P} = \{c_1, \ldots, c_n, S\}$ include all clients and the server, and the private inputs are the gradients of each honest client $\{w_1, \ldots, w_n\}$ . Then $FLSeg$ is semi-honest secure. + +Proof. We use the Hybrid Argument [12]. By gradually replacing components of the protocol's multiple interaction stages with the ideal functionality $\mathcal{F}$ , we demonstrate that each stage's replacement does not significantly alter the probability distribution of the adversary's inference results. Assuming $\mathcal{T}$ represents the entire parameter space of the model, we define the following three hybrids: + +- Hybrid 0: The original FLSeg Framework. +- Hybrid 1: Similar to Hybrid 0, but replaces $w_{i}^{-m} \oplus r_{i}$ exchanged by clients with random values $\phi_{i}^{(1)} \leftarrow \mathcal{T}$ . +- Hybrid 2: Similar to Hybrid 1, but replaces $(w_{i})_{\mathrm{mix}}$ transmitted to the server with random values $\phi_{i}^{(2)} \stackrel{\S}{\leftarrow} \mathcal{T}$ . + +Lemma 1 (Hybrid 1). Hybrid 0 and Hybrid 1 are computationally indistinguishable if FLSeg is semi-honest secure. + +Proof. Since $r_i$ and $\phi_i^{(1)}$ follow the same distribution, $w_i^{(\neg m)} \oplus r_i$ is also distributed identically to $\phi_i^{(1)}$ . Thus, the advantage of $\mathcal{A}$ is given by Eq. (4). + +$$ +\left| \mathbb {P} [ \mathcal {A} (\phi_ {i} ^ {(1)}) = 1 ] - \mathbb {P} [ \mathcal {A} (w _ {i} ^ {- m} \oplus r _ {i}) = 1 ] \right| \leq \operatorname {n e g} (d) \tag {4} +$$ + +where $\mathrm{neg}(d)$ is a negligible function. Hence, Hybrid 0 and Hybrid 1 are computationally indistinguishable. + +Lemma 2 (Hybrid 2). Hybrid 1 and Hybrid 2 are computationally indistinguishable if FLSeg is semi-honest secure. + +Proof. Since the adversary $\mathcal{A}$ cannot control both the client and the server simultaneously, $j$ and $m$ are random to $\mathcal{S}$ . If $\mathcal{S}$ attempts to recover the private $w_{i}$ by random matching, the probability of success is $\frac{1}{n} \cdot \frac{1}{2^d}$ . Thus, the advantage of $\mathcal{A}$ is given by Eq. (5). + +$$ +\left| \mathbb {P} [ \mathcal {A} ((w _ {i}) _ {\mathrm {m i x}}) = 1 ] - \mathbb {P} [ \mathcal {A} (\phi_ {i} ^ {(2)}) = 1 ] \right| \leq \frac {1}{n} \cdot \frac {1}{2 ^ {d}} \tag {5} +$$ + +Since this advantage is negligible for a polynomial-time adversary, it can be ignored. Thus, Hybrid 1 and Hybrid 2 are computationally indistinguishable. + +
SecAgg [5]Bell et al. [3]ACORN-detect [2]ShieldFL [29]PEFL [25]FLSeg
ClientComplexityO(n2+ ℓn)O(log2 n + ℓ log n)O(log2 n + ℓ log n)O(ℓ(TEnc + TDec))O(ℓ(TEnc + TDec))O(ℓ)
Overhead2,557 ms1,696 ms1,729 ms231,541 ms215,667 ms233 ms
ServerComplexityO(ℓn2)O(n(log2 n + ℓ log n))O(n(log2 n + ℓ log n))O(ℓnTDec)O(ℓn(TEnc + TDec))O(ℓn)
Overhead192 ms/c182 ms/c181 ms/c78,685 ms/c790,032 ms/c78 ms/c
Robustness○○◇◇◇◇
+ +Table 1. Comparison of computational time complexity and time overhead. $n$ represents the number of clients, $\ell$ denotes the size of the model parameters. In calculating computation time overhead, let $n = {10}^{3}$ and $\ell = {10}^{6}$ . ${T}_{\text{Enc }}$ and ${T}_{\text{Dec }}$ represent the encryption and decryption constant overhead in HE. ”/c” denotes the average time overhead incurred by the server for each client. $\bigcirc$ indicates the absence of robustness, $\mathbf{D}$ denotes support for fixed RAR only,and $\bullet$ denotes support for arbitrary RAR. We omit overheads unrelated to privacy protection,such as training and aggregation. + +Through the above hybrid experiments, we have gradually proven the information transmitted from the real-world to the ideal-world is computationally indistinguishable from random values. Thus, we have proven Theorem 1. $\square$ + +# 5.2. Efficiency in privacy protection + +Assume $n$ represents the number of clients and $\ell$ denotes the size of the model parameters. The computational overhead on the client side for FLSeg primarily consists of the following components: + +$\mathcal{O}(\ell^2)\colon m\gets$ Diffie-Hellman key exchange, +- $\mathcal{O}(\ell): \text{Calculation of } w_i^m, w_i^{-m}, w_i^{-m} \oplus r_i, (e_i)_{\text{mix}}$ +$\mathcal{O}(\ell):\mathrm{Enc}(r_j,pk_{\mathcal{S}})$ + +Here, we introduce an optimization technique: instead of directly using Diffie-Hellman to generate $m$ or encrypting $r_j$ , which incurs significant computational time costs, we generate respective random seeds $\mathsf{seed}_{r_j}$ and $\mathsf{seed}_m$ . Both clients then use a Pseudo-Random Number Generator (PRNG) [5] to produce $r_j$ and $m$ , and only the encrypted seed $\mathrm{Enc}(\mathsf{seed}_{r_j},pk_S)$ is sent. This approach reduces the overhead of Diffie-Hellman and asymmetric encryption to a constant level, depending solely on the seed length. Thus, the primary computational overhead is shifted to generating $r_j$ using PRNG $(\mathsf{seed}_{r_j}) = r_j$ and performing XOR operations, with a complexity of $\mathcal{O}(\ell)$ . Similarly, the decryption overhead on the server is minimized, focusing mainly on decrypting the mixed model $(w_i)_{\mathrm{mix}}$ using XOR operations, resulting in a complexity of $\mathcal{O}(\ell n)$ . + +To demonstrate the advantages of FLSeg in efficient privacy protection, we compare the computational time complexity and time overhead of FLSeg with SOTA privacy-preserving FL that also satisfy semi-honest security, including SecAgg [5], Bell et al. [3], ShieldFL [29], RoFL [28], and PEFL [25]. The comparison results are presented in Tab. 1. The time complexity of ShieldFL and PEFL is asymptotically $O(\ell)$ ; however, the Paillier HE they employ incurs substantial constant overheads, $T_{\mathrm{Enc}}$ and $T_{\mathrm{Dec}}$ . SecAgg, ACORN, and Bell et al. are aggregation schemes based on SMPC techniques, requiring the compu + +tation of pairwise masking for all other clients. To more intuitively demonstrate the advantages of FLSeg, we tested each method's runtime in a practical setting with $n = 10^3$ and $\ell = 10^6$ . The results show that FLSeg achieves an $8 \times$ improvement in client-side and a $2 \times$ improvement in server-side computation compared to the fastest method. See Appendix C for the experimental setup. + +# 6. Experiments + +# 6.1. Setups + +Datasets and Models. Our experiments are conducted on three datasets—FEMNIST [6], CIFAR-10, and CIFAR-100 [21]—using a Multilayer Perceptron (MLP), Convolutional Neural Network (CNN), and SqueezeNet, respectively. For FEMNIST, the dataset is naturally non-IID and consists of 199 unique clients. For CIFAR-10 and CIFAR-100, we set the number of clients, $n = 100$ and $n = 50$ , and generate non-IID data distributions across clients using a Dirichlet distribution by following [24, 26, 36]. The concentration parameter $\alpha$ of the Dirichlet distribution controls the degree of data heterogeneity, with lower values indicating higher heterogeneity. For each client, $90\%$ of their data is allocated for training, while the remaining $10\%$ is reserved for testing. + +Evaluated attacks. We select the classic targeted attack, LabelFlipping (LP) [10]; two powerful adaptive attacks, Min-Max [33] and IPM [40]; and two privacy inference attacks, iDLG [42] and Inverting-Gradients [13]. + +Compared Methods. We consider the FL baseline FedAvg [30]; three prominent RARs: Krum [4], DnC [33], and RFA [32]; three robust FL methods: GAS [26], FLTrust [8], and ClippedCluster [22]; and three privacy-preserving robust FL methods: ACORN [2], ShieldFL [29], and PEFL [25]. + +Evaluation. We use average top-1 accuracy of each client to evaluate the performance of global models. We run each experiment for five times and report the mean and standard deviation of the highest accuracy during the training process. + +All hyperparameter settings for the compared methods + +
DatasetAttackBaselineRARRobust FLPrivacy-preserving Robust FLOur
FedAvgKrumDnCRFAGASFLTrustClipped ClusteringACORNShieldFLPEFLFLSeg (Krum)FLSeg (DnC)
CIFAR-10-52.16±0.9753.86±0.9053.32±0.1353.52±0.4851.16±0.2853.28±0.7053.44±1.5949.08±1.9752.61±0.2746.31±0.8860.45±0.8263.99±0.53
LF47.46±0.6947.55±1.9246.50±1.5555.20±1.0647.96±1.9250.61±0.6150.85±1.4347.08±0.3348.37±0.8544.50±1.2359.62±0.7663.68±0.86
IPM49.03±0.1846.44±0.3351.82±0.4535.07±1.6243.55±1.6050.52±1.9550.93±1.5545.36±2.2149.53±1.8641.07±3.0259.43±1.0661.28±0.91
Min-Max35.89±1.5745.46±1.4950.62±1.4349.61±0.6049.69±1.7148.25±0.7049.36±1.9748.21±2.0338.51±3.0912.36±0.4860.82±0.5764.59±1.02
FEMNIST-70.17±1.8169.60±0.2369.95±0.8868.76±0.9369.95±0.4737.06±1.4669.31±1.1273.14±0.8173.33±0.9059.78±1.7177.12±0.6777.88±0.31
LF68.08±1.5870.37±0.4469.27±1.5668.70±0.7668.29±1.3133.88±0.6269.90±0.5169.78±0.3770.82±0.5452.68±1.9873.44±0.7776.94±0.45
IPM66.69±1.0659.88±1.9067.99±1.8455.02±0.8555.39±0.2838.81±1.0967.07±1.0168.44±2.0770.34±1.0350.37±1.6373.94±1.1574.95±0.87
Min-Max67.46±1.5265.17±1.8568.62±1.3865.38±1.3265.85±1.1236.70±0.2768.05±1.5171.29±1.0870.19±2.1121.32±0.2276.13±1.6277.74±0.65
CIFAR-100-35.92±0.6734.80±0.5434.48±0.3834.34±0.2425.64±0.9531.53±0.4233.32±0.1335.36±0.5634.70±0.752.86±0.4754.45±0.8153.55±0.59
LF30.83±0.3334.37±0.4933.46±0.2729.28±0.5826.98±0.3731.11±0.4432.99±0.4229.85±0.9132.53±0.613.21±0.9642.32±0.6543.79±0.49
IPM25.54±1.2119.29±0.7331.38±0.5315.61±0.6425.89±0.4529.43±0.2632.44±0.5925.55±0.4823.62±0.333.24±1.2841.56±0.7545.13±1.09
Min-Max22.31±1.0216.38±0.2934.23±0.6219.46±0.9327.21±0.3830.16±1.2534.73±0.6721.04±0.4334.39±0.473.20±0.4146.05±0.3451.11±0.21
+ +Table 2. Top-1 Accuracy(%) under Different Byzantine Attack. + +and attacks, along with detailed experimental configurations, are provided in Appendix D. + +# 6.2. Robustness + +Table 2 shows the accuracy under different Byzantine attackers, with data heterogeneity set at a Dirichlet concentration parameter of $\alpha = 0.5$ on CIFAR-10 and CIFAR-100, and attackers comprising $20\%$ of the total number of clients. FLSeg can utilize different RAR for Segment Aggregation; we denote the RAR used by adding it in parentheses after the method name, for example, FLSeg(DnC). + +![](images/1bbd762c6f67a253ac894703353fcf710ba92cf2a0cb3c207053f8954c63f8ec.jpg) +Figure 3. Top-1 Accuracy(%) under Different Ratio of Attacker on CIFAR-10. + +From the empirical results, we observe that FLSeg maintains stable performance under various Byzantine attacks. Additionally, we find that FLSeg's defensive effectiveness is significantly influenced by the choice of RAR. For instance, using DnC, a singular value decomposition (SVD)-based spectral method can achieve higher accuracy than Euclidean distance-based methods like Krum. We also found that the combination of RAR with FLSeg significantly outperforms the original RAR method, especially in complex models, indicating that FLSeg effectively mitigates the curse of dimensionality. For example, Krum nearly fails under Min-Max attacks on CIFAR-100, achieving an accuracy of only $16.38\%$ , but when combined with FLSeg, it achieves a much improved accuracy of $46.05\%$ . + +Figure 3 shows the accuracy under different numbers of Min-Max attackers, with FLSeg demonstrating greater stability against varying numbers of attackers. Additionally, it is evident that other privacy-preserving robust FL methods experience a notable performance drop as the number of attackers increases. This decline occurs because the metrics these methods use to filter malicious gradients are constrained by their privacy-preserving techniques, making them vulnerable to strong adaptive attacks. In contrast, FLSeg mitigates the curse of dimensionality through Segment Aggregation and enhances resistance to malicious gradients by integrating diverse RAR. + +# 6.3. Heterogeneous Data + +Figure 4 shows the accuracy under different levels of data heterogeneity on CIFAR-10. We fix the ratio of attackers at $20\%$ using Min-Max attackers and control the level of data heterogeneity by adjusting the Dirichlet $\alpha$ . An $\alpha$ of 0.1 represents an extreme case where each client may only possess one or two labels out of ten possible categories, while at $\alpha = 10$ , the data distribution is nearly IID. + +![](images/3ebce31097df6781b87d72afdeda67326be4371ae087ff471861f6ef332455fa.jpg) +Figure 4. Top-1 Accuracy(%) under Different Levels of Data Heterogeneity on CIFAR-10. + +We observe that FLSeg achieves significantly better performance than other methods when $\alpha = 0.1$ . This is because other methods were not designed to account for the + +potential impact of data heterogeneity, whereas FLSeg enables clients to better adapt to their local datasets under non-IID conditions by learning personalized models, thereby achieving improved performance. However, as the data tends to be balanced ( $\alpha = 10$ ), the gains brought by personalization will gradually decrease, but the performance of FLSeg is still better than that of other baselines. This is because applying RAR to model segments effectively alleviates the curse of dimensionality. + +# 6.4. Privacy + +To verify FLSeg's privacy protection capability against a semi-honest server, we present visual results of attacks using DLG and Inverting-Gradient, with and without our defense, on FEMNIST and CIFAR-100 in Fig. 5. Without privacy protection (Origin), the server can easily reconstruct training data. In contrast, when attacking the mixed model after Segment Exchange, only meaningless information can be extracted. Furthermore, in Sec. 5.1, we prove that the probability of the server reconstructing the original client model is negligible, thereby ensuring client privacy. + +![](images/f2274569cecfff567e9a04e0cf92367ac1d434d41653225ad7799ed2ad55ebcf.jpg) +Figure 5. Attack results of iDLG and Inverting-Gradients on CIFAR-10 and FEMNIST, respectively. + +# 6.5.Impact of Segment Size + +To investigate the impact of segment size on privacy protection and model performance, we vary the segment size $|\gamma|$ across $\{10^2, 10^3, 10^4, 10^5, 10^6, 2904970(\ell)\}$ under MinMax and Inverting-Gradient attacks on the heterogeneous CIFAR-10 dataset. We use the same settings as in the robustness experiments to evaluate the model's top-1 accuracy and perform an inference attack on the model updates of a single client from one training round. The experimental results are presented in Tab. 3. + +
|γ|102103104105106l
Accuracy63.93%64.33%65.72%60.93%59.67%57.75%
Inference
+ +Table 3. Accuracy and Privacy of FLSeg(Dnc) under different $|\gamma |$ + +We observe that the accuracy of $\mathrm{FLSeg(DnC)}$ initially increases and then slightly decreases. We hypothesize that this occurs because benchmarks computed from smaller segments lack representativeness, while larger segments fail to mitigate the curse of dimensionality. Moreover, as the segment size increases, the reconstructed attack results reveal more information about the original data, since larger segments concentrate more original gradient information in the mixed model. + +# 6.6. Ablation Study + +We finally study how each component contributes to the overall performance of FLSeg. Specifically, we measure the following scenarios: + +FedAvg. The baseline method. +- PT-only. Adds local personalized training to FedAvg. +- SA-only. Adds Segment Aggregation to FedAvg. +- FLSeg. The complete FLSeg method. + +Note that we did not include Segment Exchange as its purpose is to protect privacy without compromising the correctness of subsequent aggregation. + +![](images/10619e13be0213b5a4578037d6316a5769ad9b4d581a47aabd718a80e2b38037.jpg) +Figure 6. Ablation study. + +We conducted ablation experiments under four scenarios combining with/without $20\%$ Min-Max attackers and IID/non-IID $(\alpha = 0.5)$ CIFAR-10 dataset, with the results shown in Fig. 6. We observe that, in all scenarios, Segment Aggregation effectively resists Byzantine attacks. Meanwhile, the incorporation of personalized training does not degrade SA-only performance under IID data, and improves the model performance under non-IID data. + +# 7. Conclusion + +In this work, we identified three key challenges in privacy-preserving robust FL: high computational costs, limited flexibility in RAR, and data heterogeneity. To overcome these, we proposed FLSeg, which uses Segment Exchange and Segment Aggregation to ensure privacy without limiting aggregation choices and employs a regularization term to enhance adaptation to heterogeneous data. Theoretical and experimental results validate FLSeg's security, robustness, and resilience to data heterogeneity. + +In the future, we will explore stronger threat models, such as malicious server, or extend the privacy-preserving segment exchange framework to decentralized settings. + +# Acknowledgement + +This research was supported in part by the Major Research Plan of Hubei Province (2023BAA027), the Changsha Key Science and Technology Project under the "Unveiling and Commanding" Initiative (kq2503009), the Science, Technology and Innovation Commission of Shenzhen Municipality (GJHZ20240218114659027), the Key Research and Development Plan of Hubei Province (2024BAB049), and the Fundamental Research Funds for the Central Universities (YCJJ20252336). + +# References + +[1] Abbas Acar, Hidayet Aksu, A Selcuk Uluagac, and Mauro Conti. A survey on homomorphic encryption schemes: Theory and implementation. ACM Computing Surveys (Csur), 51(4):1-35, 2018. 2 +[2] James Bell, Adrià Gascon, Tancrède Lepoint, Baiyu Li, Sarah Meiklejohn, Mariana Raykova, and Cathie Yun. ACORN: input validation for secure aggregation. In 32nd USENIX Security Symposium, USENIX Security 2023, Anaheim, CA, USA, August 9-11, 2023, pages 4805-4822. USENIX Association, 2023. 1, 2, 3, 6 +[3] James Henry Bell, Kallista A. Bonawitz, Adrià Gasón, Tancrède Lepoint, and Mariana Raykova. Secure single-server aggregation with (poly)logarithmic overhead. In CCS '20: 2020 ACM SIGSAC Conference on Computer and Communications Security, Virtual Event, USA, November 9-13, 2020, pages 1253-1269. ACM, 2020. 6 +[4] Peva Blanchard, El Mahdi El Mhamdi, Rachid Guerraoui, and Julien Stainer. Machine learning with adversaries: Byzantine tolerant gradient descent. Advances in neural information processing systems, 30, 2017. 1, 2, 6 +[5] Keith Bonawitz, Vladimir Ivanov, Ben Kreuter, Antonio Marcedone, H Brendan McMahan, Sarvar Patel, Daniel Ramage, Aaron Segal, and Karn Seth. Practical secure aggregation for privacy-preserving machine learning. In proceedings of the 2017 ACM SIGSAC Conference on Computer and Communications Security, pages 1175-1191, 2017. 1, 2, 6 +[6] Sebastian Caldas, Sai Meher Karthik Duddu, Peter Wu, Tian Li, Jakub Konečný, H Brendan McMahan, Virginia Smith, and Ameet Talwalkar. Leaf: A benchmark for federated settings. arXiv preprint arXiv:1812.01097, 2018. 6 +[7] Ran Canetti. Universally composable security: A new paradigm for cryptographic protocols. In FOCS, pages 136-145. IEEE Computer Society, 2001. 5 +[8] Xiaoyu Cao, Minghong Fang, Jia Liu, and Neil Zhenqiang Gong. Fltrust: Byzantine-robust federated learning via trust bootstrapping. In NDSS. The Internet Society, 2021. 2, 6 +[9] Alireza Fallah, Aryan Mokhtari, and Asuman E. Ozdaglar. Personalized federated learning with theoretical guarantees: A model-agnostic meta-learning approach. In NeurIPS, 2020. 3 +[10] Minghong Fang, Xiaoyu Cao, Jinyuan Jia, and Neil Gong. Local model poisoning attacks to {Byzantine-Robust} federated learning. In 29th USENIX security symposium (USENIX Security 20), pages 1605-1622, 2020. 2, 6 + +[11] Minghong Fang, Xiaoyu Cao, Jinyuan Jia, and Neil Zhenqiang Gong. Local model poisoning attacks to byzantine-robust federated learning. In USENIX Security Symposium, pages 1605–1622. USENIX Association, 2020. 2 +[12] Marc Fischlin and Arno Mittelbach. An overview of the hybrid argument. IACR Cryptol. ePrint Arch., page 88, 2021. 5 +[13] Jonas Geiping, Hartmut Bauermeister, Hannah Droge, and Michael Moeller. Inverting gradients-how easy is it to break privacy in federated learning? Advances in neural information processing systems, 33:16937-16947, 2020. 1, 2, 6 +[14] Oded Goldreich, Silvio Micali, and Avi Wigderson. How to play any mental game or A completeness theorem for protocols with honest majority. In STOC, pages 218-229. ACM, 1987. 5 +[15] Rachid Guerraoui, Sébastien Rouault, et al. The hidden vulnerability of distributed learning in byzantium. In International Conference on Machine Learning, pages 3521-3530. PMLR, 2018. 2, 3 +[16] Anbu Huang, Yang Liu, Tianjian Chen, Yongkai Zhou, Quan Sun, Hongfeng Chai, and Qiang Yang. Starfl: Hybrid federated learning architecture for smart urban computing. ACM Trans. Intell. Syst. Technol., 12(4):43:1-43:23, 2021. 1 +[17] Najeeb Moharram Jebreel, Josep Domingo-Ferrer, Alberto Blanco-Justicia, and David Sánchez. Enhanced security and privacy via fragmented federated learning. IEEE Transactions on Neural Networks and Learning Systems, 2022. 2 +[18] Peter Kairouz, H Brendan McMahan, Brendan Avent, Aurélien Bellet, Mehdi Bennis, Arjun Nitin Bhagoji, Kallista Bonawitz, Zachary Charles, Graham Cormode, Rachel Cummings, et al. Advances and open problems in federated learning. Foundations and trends® in machine learning, 14(1-2): 1-210, 2021. 1 +[19] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank Reddi, Sebastian Stich, and Ananda Theertha Suresh. SCAFFOLD: Stochastic controlled averaging for federated learning. In Proceedings of the 37th International Conference on Machine Learning, pages 5132-5143. PMLR, 2020. 3 +[20] Sai Praneeth Karimireddy, Lie He, and Martin Jaggi. Byzantine-robust learning on heterogeneous datasets via bucketing. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022. 3 +[21] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6 +[22] Shenghui Li, Edith C-H Ngai, and Thiemo Voigt. An experimental study of byzantine-robust aggregation schemes in federated learning. IEEE Transactions on Big Data, 2023. 1, 2, 6 +[23] Shenghui Li, Edith Ngai, Fanghua Ye, Li Ju, Tianru Zhang, and Thiemo Voigt. Blades: A unified benchmark suite for byzantine attacks and defenses in federated learning. In 2024 IEEE/ACM Ninth International Conference on Internet-of-Things Design and Implementation (IoTDI), 2024. 1, 2 +[24] Tian Li, Shengyuan Hu, Ahmad Beirami, and Virginia Smith. Ditto: Fair and robust federated learning through per + +sonalization. In International conference on machine learning, pages 6357-6368. PMLR, 2021. 3, 6 +[25] Xiaoyuan Liu, Hongwei Li, Guowen Xu, Zongqi Chen, Xiaoming Huang, and Rongxing Lu. Privacy-enhanced federated learning against poisoning adversaries. IEEE Transactions on Information Forensics and Security, 16:4574-4588, 2021. 1, 2, 3, 6 +[26] Yuchen Liu, Chen Chen, Lingjuan Lyu, Fangzhao Wu, Sai Wu, and Gang Chen. Byzantine-robust learning on heterogeneous data via gradient splitting. In International Conference on Machine Learning, pages 21404–21425. PMLR, 2023. 1, 2, 3, 6 +[27] Zhi Lu, SongFeng Lu, YongQuan Cui, XueMing Tang, and JunJun Wu. Split aggregation: Lightweight privacy-preserving federated learning resistant to byzantine attacks. IEEE Transactions on Information Forensics and Security, 19:5575-5590, 2024. 1, 3 +[28] Hidde Lycklama, Lukas Burkhalter, Alexander Viand, Nicolas Kuchler, and Anwar Hithnawi. Rofl: Robustness of secure federated learning. In SP, pages 453-476. IEEE, 2023. 1, 2, 6 +[29] Zhuoran Ma, Jianfeng Ma, Yinbin Miao, Yingjiu Li, and Robert H Deng. Shieldfl: Mitigating model poisoning attacks in privacy-preserving federated learning. IEEE Transactions on Information Forensics and Security, 17:1639-1654, 2022. 1, 2, 3, 6 +[30] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 2, 6 +[31] Dinh C. Nguyen, Quoc-Viet Pham, Pubudu N. Pathirana, Ming Ding, Aruna Seneviratne, Zihuai Lin, Octavia A. Dobre, and Won-Joo Hwang. Federated learning for smart healthcare: A survey. ACM Comput. Surv., 55(3):60:1-60:37, 2023. 1 +[32] Krishna Pillutla, Sham M Kakade, and Zaid Harchaoui. Robust aggregation for federated learning. IEEE Transactions on Signal Processing, 70:1142-1154, 2022. 1, 2, 6 +[33] Virat Shejwalkar and Amir Houmansadr. Manipulating the byzantine: Optimizing model poisoning attacks and defenses for federated learning. In NDSS, 2021. 1, 2, 3, 6 +[34] Virat Shejwalkar, Amir Houmansadr, Peter Kairouz, and Daniel Ramage. Back to the drawing board: A critical evaluation of poisoning attacks on production federated learning. In SP, pages 1354-1371. IEEE, 2022. 1 +[35] Lichao Sun, Jianwei Qian, and Xun Chen. LDP-FL: practical private aggregation in federated learning with local differential privacy. In IJCAI, pages 1571–1578. ijcai.org, 2021. 1 +[36] Canh T Dinh, Nguyen Tran, and Josh Nguyen. Personalized federated learning with moreau envelopes. Advances in Neural Information Processing Systems, 33:21394-21405, 2020. 3, 6 +[37] Kang Wei, Jun Li, Ming Ding, Chuan Ma, Hang Su, Bo Zhang, and H. Vincent Poor. User-level privacy-preserving federated learning: Analysis and performance optimization. IEEE Trans. Mob. Comput., 21(9):3388-3401, 2022. 2 + +[38] Jiahui Wu, Weizhe Zhang, and Fucai Luo. On the security of "lsfl: A lightweight and secure federated learning scheme for edge computing". IEEE Trans. Inf. Forensics Secur., 19: 3481-3482, 2024. 1 +[39] Jue Xiao, Xueming Tang, and Songfeng Lu. Privacy-preserving federated class-incremental learning. IEEE Transactions on Machine Learning in Communications and Networking, 2:150-168, 2024. 1 +[40] Cong Xie, Oluwasanmi Koyejo, and Indranil Gupta. Fall of empires: Breaking byzantine-tolerant sgd by inner product manipulation. In Uncertainty in Artificial Intelligence, pages 261-270. PMLR, 2020. 1, 2, 6 +[41] Dong Yin, Yudong Chen, Ramchandran Kannan, and Peter Bartlett. Byzantine-robust distributed learning: Towards optimal statistical rates. In International Conference on Machine Learning, pages 5650–5659. Pmlr, 2018. 2 +[42] Bo Zhao, Konda Reddy Mopuri, and Hakan Bilen. idlg: Improved deep leakage from gradients. CoRR, abs/2001.02610, 2020. 2, 6 +[43] Ligeng Zhu, Zhijian Liu, and Song Han. Deep leakage from gradients. Advances in neural information processing systems, 32, 2019. 1, 2 \ No newline at end of file diff --git a/2025/FLSeg_ Enhancing Privacy and Robustness in Federated Learning under Heterogeneous Data via Model Segmentation/images.zip b/2025/FLSeg_ Enhancing Privacy and Robustness in Federated Learning under Heterogeneous Data via Model Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7841d5085f2e1ca326f7c5fccc789a718bfc4866 --- /dev/null +++ b/2025/FLSeg_ Enhancing Privacy and Robustness in Federated Learning under Heterogeneous Data via Model Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:72c55a976e0f17a6022c229064a194fd58cacb36443c152b1692a1d73d089b51 +size 551843 diff --git a/2025/FLSeg_ Enhancing Privacy and Robustness in Federated Learning under Heterogeneous Data via Model Segmentation/layout.json b/2025/FLSeg_ Enhancing Privacy and Robustness in Federated Learning under Heterogeneous Data via Model Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1935c6697eb4c02652ee47ef2e6f012c03297664 --- /dev/null +++ b/2025/FLSeg_ Enhancing Privacy and Robustness in Federated Learning under Heterogeneous Data via Model Segmentation/layout.json @@ -0,0 +1,11452 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 81, + 103, + 531, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 103, + 531, + 140 + ], + "spans": [ + { + "bbox": [ + 81, + 103, + 531, + 140 + ], + "type": "text", + "content": "FLSeg: Enhancing Privacy and Robustness in Federated Learning under Heterogeneous Data via Model Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 127, + 160, + 481, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 160, + 481, + 175 + ], + "spans": [ + { + "bbox": [ + 127, + 160, + 481, + 175 + ], + "type": "text", + "content": "Zichun " + }, + { + "bbox": [ + 127, + 160, + 481, + 175 + ], + "type": "inline_equation", + "content": "\\mathbf{Su}^{1\\dagger}" + }, + { + "bbox": [ + 127, + 160, + 481, + 175 + ], + "type": "text", + "content": " Zhi Lu" + }, + { + "bbox": [ + 127, + 160, + 481, + 175 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 127, + 160, + 481, + 175 + ], + "type": "text", + "content": " Yutong Wu" + }, + { + "bbox": [ + 127, + 160, + 481, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 127, + 160, + 481, + 175 + ], + "type": "text", + "content": " Renfei Shen" + }, + { + "bbox": [ + 127, + 160, + 481, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 127, + 160, + 481, + 175 + ], + "type": "text", + "content": " Songfeng Lu" + }, + { + "bbox": [ + 127, + 160, + 481, + 175 + ], + "type": "inline_equation", + "content": "^{1,2*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 82, + 175, + 529, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 175, + 529, + 203 + ], + "spans": [ + { + "bbox": [ + 82, + 175, + 529, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 82, + 175, + 529, + 203 + ], + "type": "text", + "content": "School of Cyber Science and Engineering, Huazhong University of Science and Technology " + }, + { + "bbox": [ + 82, + 175, + 529, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 82, + 175, + 529, + 203 + ], + "type": "text", + "content": "Shenzhen Huazhong University of Science and Technology Research Institute" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 118, + 205, + 485, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 205, + 485, + 217 + ], + "spans": [ + { + "bbox": [ + 118, + 205, + 485, + 217 + ], + "type": "text", + "content": "suzc25@m.fudan.edu.cn, {luzhi,wyt,shenrenfei,lusongfeng}@hust.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 245, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 245, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 151, + 245, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 269, + 297, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 269, + 297, + 510 + ], + "spans": [ + { + "bbox": [ + 54, + 269, + 297, + 510 + ], + "type": "text", + "content": "Federated Learning (FL) enables collaborative training of a global model without data sharing, yet it faces critical challenges from privacy leakage and Byzantine attacks. Existing privacy-preserving robust FL frameworks suffer from three key limitations: high computational costs, restricted application of Robust Aggregation Rules (RAR), and inadequate handling of data heterogeneity. To address these limitations, we propose the FLSeg framework, which leverages Segment Exchange and Segment Aggregation to avoid excessive encryption computations while enabling unrestricted use of any RAR. Additionally, a regularization term in local training balances personalization with global model performance, effectively adapting to heterogeneous data. Our theoretical and experimental analyses demonstrate FLSeg's semi-honest security and computational efficiency: it achieves client and server time complexities of " + }, + { + "bbox": [ + 54, + 269, + 297, + 510 + ], + "type": "inline_equation", + "content": "O(\\ell)" + }, + { + "bbox": [ + 54, + 269, + 297, + 510 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 269, + 297, + 510 + ], + "type": "inline_equation", + "content": "O(n\\ell)" + }, + { + "bbox": [ + 54, + 269, + 297, + 510 + ], + "type": "text", + "content": ", with empirical results showing significantly reduced computational times. Extensive experiments further confirm FLSeg's robustness across diverse heterogeneous and adversarial scenarios." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 532, + 135, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 135, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 135, + 544 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 552, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 296, + 696 + ], + "type": "text", + "content": "Federated Learning (FL) allows organizations to collaboratively train models without sharing sensitive data. By exchanging model updates instead of raw data, FL enhances privacy and efficiency, making it ideal for fields like finance [18], healthcare [16], and smart cities [31]. However, FL is currently threatened by two major issues: privacy leakage and Byzantine attacks. Privacy leakage occurs when the server can infer clients' sensitive training data by analyzing the model updates [13, 43]. Byzantine attacks, on the other hand, involve malicious clients submitting incorrect or manipulated updates, which can disrupt or compromise the global model's performance [23, 34]. To address" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 246, + 554, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 246, + 554, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 554, + 365 + ], + "type": "text", + "content": "these challenges, corresponding techniques have been developed. For privacy protection, methods like Homomorphic Encryption (HE) [29], Secure Multi-Party Computation (SMPC) [5], and Differential Privacy (DP) [35, 39] aim to safeguard training data by ensuring that only encrypted or obfuscated information is shared. For robustness against Byzantine attacks, various Robust Aggregation Rule (RAR), such as Krum [4], DnC [33], and RFA [32], help identify and exclude malicious updates, thereby protecting the global model from adversarial influence." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 365, + 554, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 554, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 554, + 461 + ], + "type": "text", + "content": "However, achieving both robustness and privacy simultaneously presents a significant challenge, as privacy-preserving methods can unintentionally protect malicious updates from detection [27]. To address this issue, a series of Privacy-preserving and Robust FL (PRFL) frameworks have been proposed. These methods explore various architectures and protocols or select encryption schemes that are compatible with specific RAR [2, 25, 28, 29, 38]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 462, + 554, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 554, + 581 + ], + "type": "text", + "content": "Despite these methods balancing robustness and privacy, they still have notable limitations: (1) Privacy-preserving techniques such as HE and SMPC introduce significant computational overhead [27]. (2) These methods impose restrictions on the use of RAR, supporting only simple, single RAR, which are vulnerable to adaptive attacks such as Min-Max [33] and IPM [40]. (3) Gradient heterogeneity caused by data heterogeneity significantly degrades both model performance and the accuracy of RAR [22]. However, existing PRFL methods largely overlook this issue." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "Therefore, we aim to address the following question: \"How can we efficiently preserve privacy while avoiding restrictions on RAR implementation and mitigating the impact of data heterogeneity on robustness and performance?\" To this end, we draw inspiration from existing theoretical insights: reducing data dimensionality can alleviate the curse of dimensionality and gradient heterogeneity, thereby enhancing robustness [26, 33]. Based on this principle, we propose FLSeg, a model segmentation-based framework. We design a Segment Exchange privacy protocol to achieve privacy preservation without incurring excessive computa" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 212, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 212, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 212, + 713 + ], + "type": "text", + "content": "* Corresponding author, † Equal contribution" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3916" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "tional overhead. Additionally, we introduce a Segment Aggregation algorithm, which applies RAR operations at the low-dimensional model segments to mitigate gradient heterogeneity and the curse of dimensionality while supporting arbitrary RAR. Furthermore, we incorporate a regularization term into local training to facilitate the learning of personalized models, further improving performance under data heterogeneity. Our contributions are threefold:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 170, + 294, + 444 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 55, + 170, + 294, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 170, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 170, + 294, + 216 + ], + "type": "text", + "content": "- A Novel Privacy-Preserving Robust FL Framework. To the best of our knowledge, FLSeg is the first PRFL framework that does not restrict the choice of RAR, while also mitigating the impact of heterogeneous data." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 217, + 294, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 294, + 335 + ], + "type": "text", + "content": "- Computationally Efficient Privacy Protection. We theoretically demonstrate that our FLSeg satisfies semihonest security. We further provide a detailed analysis of FLSeg's computational time complexity, with the client-side and server-side complexities being " + }, + { + "bbox": [ + 55, + 217, + 294, + 335 + ], + "type": "inline_equation", + "content": "O(\\ell)" + }, + { + "bbox": [ + 55, + 217, + 294, + 335 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 217, + 294, + 335 + ], + "type": "inline_equation", + "content": "O(n\\ell)" + }, + { + "bbox": [ + 55, + 217, + 294, + 335 + ], + "type": "text", + "content": ", respectively. In practical tests, FLSeg's client computation time is only 233 ms, compared to 1696 ms for SOTA method ACORN [2]. The server computation time per client is only 78 ms, whereas ACORN require 181 ms per client." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 337, + 294, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 337, + 294, + 444 + ], + "spans": [ + { + "bbox": [ + 55, + 337, + 294, + 444 + ], + "type": "text", + "content": "- Extensive Experimental Analysis. Extensive experiments on three benchmark datasets show that our FLSeg can significantly improve the performance under varying degrees of data heterogeneity and in Byzantine attack, e.g., the accuracy of ACORN is " + }, + { + "bbox": [ + 55, + 337, + 294, + 444 + ], + "type": "inline_equation", + "content": "48.21\\%" + }, + { + "bbox": [ + 55, + 337, + 294, + 444 + ], + "type": "text", + "content": ", while FLSeg achieves " + }, + { + "bbox": [ + 55, + 337, + 294, + 444 + ], + "type": "inline_equation", + "content": "64.59\\%" + }, + { + "bbox": [ + 55, + 337, + 294, + 444 + ], + "type": "text", + "content": " under a " + }, + { + "bbox": [ + 55, + 337, + 294, + 444 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 55, + 337, + 294, + 444 + ], + "type": "text", + "content": " proportion of Min-Max attackers and on non-IID CIFAR-10 datasets. Further inference attack experiments also verified the privacy protection capability of FLSeg." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 454, + 146, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 454, + 146, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 146, + 466 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 474, + 294, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 294, + 533 + ], + "type": "text", + "content": "Privacy Threats and Defences. Existing studies have demonstrated that gradient sharing in FL poses severe privacy risks. DLG attacks [42, 43] and Inverting Gradients attacks [13] can precisely reconstruct raw training data from client-uploaded gradients." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 534, + 294, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 294, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 294, + 617 + ], + "type": "text", + "content": "To mitigate these risks, privacy-preserving techniques such as DP [37], HE [25], and SMPC [5] have been proposed to obfuscate or encrypt gradients. However, these approaches have notable limitations: DP inevitably reduce model utility due to noise injection, while HE and SMPC incur high computational overhead due to complex encryption structures and protocols [1]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 618, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 294, + 713 + ], + "type": "text", + "content": "Byzantine Attack and Defence. Byzantine attacks are initiated by malicious clients who manipulate or send adversarial model updates to degrade model performance [23]. A common type of such attacks is data poisoning, such as LabelFlipping [10], which manipulates training data to mislead the model into learning incorrect associations. Additionally, adaptive attacks, such as Min-Max [33] and IPM [40], optimize the magnitude and direction of malicious" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": "gradient perturbations to maximize the disruption of the global FL model while simultaneously evading detection." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 96, + 553, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 553, + 287 + ], + "type": "text", + "content": "To mitigate these risks, RAR modify the standard mean aggregation method [30] by performing statistical or feature analysis on the collected updates to filter out malicious ones. Examples include methods based on Euclidean distance, such as Krum [4] and Bulyan [15]; those leveraging coordinate values, such as RFA [32] and Median [41]; and approaches utilizing spectral anomaly detection, such as DnC [33]. Beyond RAR, Robust FL integrate auxiliary techniques for enhanced defense. FLTrust [8] uses a root dataset to assess the quality of client updates, ClippedClustering [22] leverages historical gradient information to constrain and clip gradients before aggregation, and GAS [26] proposes a gradient splitting technique to reduce data dimensionality and enhance the robustness of RAR. However, these defenses do not secure the gradients during transmission, rendering them vulnerable to privacy threats." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 288, + 553, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 553, + 443 + ], + "type": "text", + "content": "Privacy-preserving Robust FL. PRFL integrates privacy-preserving techniques with RAR through specially designed protocols. PEFL [25] and ShieldFL [29] employ HE to ensure privacy while using cosine similarity to detect malicious updates. FFL [17] proposes a parameter mixing method to protect privacy while applying cosine similarity to the mixed model for robustness. ACORN [2] and RoFL [28] utilize the SecAgg SMPC protocol to safeguard the privacy of inputs while employing zero-knowledge proofs to bound the norm of model updates, thereby mitigating the impact of malicious updates. However, the high computational cost, restricted support for RAR, and lack of consideration for data heterogeneity limit their applicability." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 453, + 399, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 453, + 399, + 464 + ], + "spans": [ + { + "bbox": [ + 314, + 453, + 399, + 464 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 472, + 553, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 553, + 509 + ], + "type": "text", + "content": "Notion. For any positive integer " + }, + { + "bbox": [ + 313, + 472, + 553, + 509 + ], + "type": "inline_equation", + "content": "n \\in \\mathbb{N}^+" + }, + { + "bbox": [ + 313, + 472, + 553, + 509 + ], + "type": "text", + "content": ", we denote the set " + }, + { + "bbox": [ + 313, + 472, + 553, + 509 + ], + "type": "inline_equation", + "content": "\\{1, \\ldots, n\\}" + }, + { + "bbox": [ + 313, + 472, + 553, + 509 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 313, + 472, + 553, + 509 + ], + "type": "inline_equation", + "content": "[n]" + }, + { + "bbox": [ + 313, + 472, + 553, + 509 + ], + "type": "text", + "content": ". The cardinality of a set " + }, + { + "bbox": [ + 313, + 472, + 553, + 509 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 472, + 553, + 509 + ], + "type": "text", + "content": " is denoted by " + }, + { + "bbox": [ + 313, + 472, + 553, + 509 + ], + "type": "inline_equation", + "content": "|w|" + }, + { + "bbox": [ + 313, + 472, + 553, + 509 + ], + "type": "text", + "content": ". We denote the " + }, + { + "bbox": [ + 313, + 472, + 553, + 509 + ], + "type": "inline_equation", + "content": "l_2" + }, + { + "bbox": [ + 313, + 472, + 553, + 509 + ], + "type": "text", + "content": " norm of vector " + }, + { + "bbox": [ + 313, + 472, + 553, + 509 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 472, + 553, + 509 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 313, + 472, + 553, + 509 + ], + "type": "inline_equation", + "content": "\\| w \\|" + }, + { + "bbox": [ + 313, + 472, + 553, + 509 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 509, + 553, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 553, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 553, + 559 + ], + "type": "text", + "content": "Model Segmentation. For a model " + }, + { + "bbox": [ + 313, + 509, + 553, + 559 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 509, + 553, + 559 + ], + "type": "text", + "content": ", the mode of segmentation refers to a partitioning of the set of model parameters, " + }, + { + "bbox": [ + 313, + 509, + 553, + 559 + ], + "type": "inline_equation", + "content": "\\{\\gamma^{(1)},\\gamma^{(2)},\\dots,\\gamma^{(d)}\\}" + }, + { + "bbox": [ + 313, + 509, + 553, + 559 + ], + "type": "text", + "content": ", where different segments are disjoint, and " + }, + { + "bbox": [ + 313, + 509, + 553, + 559 + ], + "type": "inline_equation", + "content": "w = \\bigcup_{k = 1}^{d}\\gamma^{(k)}" + }, + { + "bbox": [ + 313, + 509, + 553, + 559 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 558, + 553, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 553, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 553, + 653 + ], + "type": "text", + "content": "System Setup. We consider a classic FL scenario [30], involving " + }, + { + "bbox": [ + 313, + 558, + 553, + 653 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 558, + 553, + 653 + ], + "type": "text", + "content": " clients, denoted by " + }, + { + "bbox": [ + 313, + 558, + 553, + 653 + ], + "type": "inline_equation", + "content": "\\{c_i\\}_{i \\in [n]}" + }, + { + "bbox": [ + 313, + 558, + 553, + 653 + ], + "type": "text", + "content": ", and a single server " + }, + { + "bbox": [ + 313, + 558, + 553, + 653 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 558, + 553, + 653 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 558, + 553, + 653 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 313, + 558, + 553, + 653 + ], + "type": "text", + "content": " has a unique local dataset " + }, + { + "bbox": [ + 313, + 558, + 553, + 653 + ], + "type": "inline_equation", + "content": "D_i" + }, + { + "bbox": [ + 313, + 558, + 553, + 653 + ], + "type": "text", + "content": " that is non-IID. " + }, + { + "bbox": [ + 313, + 558, + 553, + 653 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 558, + 553, + 653 + ], + "type": "text", + "content": " starts with initial model weights " + }, + { + "bbox": [ + 313, + 558, + 553, + 653 + ], + "type": "inline_equation", + "content": "w_0" + }, + { + "bbox": [ + 313, + 558, + 553, + 653 + ], + "type": "text", + "content": " and distributes the model to each " + }, + { + "bbox": [ + 313, + 558, + 553, + 653 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 313, + 558, + 553, + 653 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 558, + 553, + 653 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 313, + 558, + 553, + 653 + ], + "type": "text", + "content": " train the model on their datasets and send their updates " + }, + { + "bbox": [ + 313, + 558, + 553, + 653 + ], + "type": "inline_equation", + "content": "w_i^t" + }, + { + "bbox": [ + 313, + 558, + 553, + 653 + ], + "type": "text", + "content": " to the server " + }, + { + "bbox": [ + 313, + 558, + 553, + 653 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 558, + 553, + 653 + ], + "type": "text", + "content": ", which aggregates these updates and proceeds to the next round. After " + }, + { + "bbox": [ + 313, + 558, + 553, + 653 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 558, + 553, + 653 + ], + "type": "text", + "content": " rounds, this process results in a global model " + }, + { + "bbox": [ + 313, + 558, + 553, + 653 + ], + "type": "inline_equation", + "content": "w^T" + }, + { + "bbox": [ + 313, + 558, + 553, + 653 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "Threat Model. Our system faces two primary threats. First, in line with [8, 11], there may be Byzantine clients within the system who manipulate their uploaded model to disrupt training. They can launch targeted attacks such as LabelFlipping [10], and powerful adaptive attacks like Min" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3917" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 69, + 553, + 268 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 553, + 268 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 553, + 268 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 553, + 268 + ], + "type": "image", + "image_path": "36b96f68677cefb5702ab0523844da931eaec0dc81188ae03e627f21df1580fd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 241, + 277, + 369, + 289 + ], + "lines": [ + { + "bbox": [ + 241, + 277, + 369, + 289 + ], + "spans": [ + { + "bbox": [ + 241, + 277, + 369, + 289 + ], + "type": "text", + "content": "Figure 1. The Overview of FLSeg." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 309, + 296, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 309, + 296, + 381 + ], + "spans": [ + { + "bbox": [ + 55, + 309, + 296, + 381 + ], + "type": "text", + "content": "Max [33]. Second, the server is semi-honest, adhering to the aggregation protocol while attempting to infer private local data from the client's updates [29]. As in prior work [2, 25, 27, 29], we assume that the adversary can control a number of clients " + }, + { + "bbox": [ + 55, + 309, + 296, + 381 + ], + "type": "inline_equation", + "content": "f < (n - 1) / 2" + }, + { + "bbox": [ + 55, + 309, + 296, + 381 + ], + "type": "text", + "content": " and semi-honest server and Byzantine clients do not collude." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 396, + 138, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 138, + 410 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 138, + 410 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 418, + 119, + 429 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 119, + 429 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 119, + 429 + ], + "type": "text", + "content": "4.1. Intuition" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 437, + 296, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 296, + 533 + ], + "type": "text", + "content": "Existing RAR methods commonly suffer from the curse of dimensionality [15], where malicious gradients can hide within high-dimensional vectors. Additionally, data heterogeneity introduces gradient heterogeneity [19], further increasing the difficulty of detection. Studies have shown that reducing input dimensionality can help alleviate this issue, for example, by processing gradients through sub-sampling [33], grouping [20], or splitting [26]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": "Based on this insight, we propose an intuitive approach: by fixing the mode of segmentation, each client selects a partner to exchange model segments, while the server applies the same segmentation mode during aggregation. In this way, as long as the server is unaware of the specific exchange pairs or the exchanged segments, it cannot reconstruct the client models, thereby ensuring privacy protection. Additionally, the server apply RAR on each segment, which not only reduces the curse of dimensionality and gradient heterogeneity but also ensures that each segment to which the RAR is applied originates from different clients, thereby ensuring the correctness of the RAR implementation. Besides, although these processes mitigate gradient heterogeneity caused by Non-IID data [15, 20], they fail to address the performance degradation induced by Non-" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 309, + 554, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 309, + 554, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 309, + 554, + 358 + ], + "type": "text", + "content": "IID data. To tackle this issue, we incorporate local regularization techniques from personalized FL [24, 36] to improve model performance without altering the overall process. The whole process is illustrated in Fig. 1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 363, + 533, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 533, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 533, + 374 + ], + "type": "text", + "content": "4.2. Non-IID Data: Learn Personalized Model" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 380, + 555, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 555, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 555, + 489 + ], + "type": "text", + "content": "Heterogeneous data can cause the global model to fail to adapt to all clients' data distributions, leading to performance degradation. Existing personalized FL methods address this issue by balancing the global and local models to integrate the generalization ability of the global model with the uniqueness of local data [9, 24, 36]. Inspired by this approach, we modify each client's local training process by allowing the client to maintain a personalized model " + }, + { + "bbox": [ + 313, + 380, + 555, + 489 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 313, + 380, + 555, + 489 + ], + "type": "text", + "content": ", with its objective function defined in Eq. (1)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 348, + 497, + 555, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 497, + 555, + 520 + ], + "spans": [ + { + "bbox": [ + 348, + 497, + 555, + 520 + ], + "type": "interline_equation", + "content": "\\min _ {v _ {i}} h _ {i} \\left(v _ {i}, w ^ {t}\\right) = \\mathcal {L} _ {i} \\left(v _ {i}\\right) + \\frac {\\lambda}{2} \\left\\| v _ {i} - w ^ {t} \\right\\| ^ {2}, \\tag {1}", + "image_path": "ca4adf61e6877d9bed8b915e063ed81d8f232fcb1b3a685b4a450e79094e11cc.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 522, + 554, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 554, + 606 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 554, + 606 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 522, + 554, + 606 + ], + "type": "inline_equation", + "content": "w^{t}" + }, + { + "bbox": [ + 313, + 522, + 554, + 606 + ], + "type": "text", + "content": " represents the global model received in each round; " + }, + { + "bbox": [ + 313, + 522, + 554, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_i(w) = \\mathbb{E}_{\\xi_i \\sim \\mathcal{D}_i}[f_i(w; \\xi_i)]" + }, + { + "bbox": [ + 313, + 522, + 554, + 606 + ], + "type": "text", + "content": " denotes the loss function on " + }, + { + "bbox": [ + 313, + 522, + 554, + 606 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 313, + 522, + 554, + 606 + ], + "type": "text", + "content": "'s local data; " + }, + { + "bbox": [ + 313, + 522, + 554, + 606 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 522, + 554, + 606 + ], + "type": "text", + "content": " is a regularization parameter that balances the influence of the personalized and global models. When " + }, + { + "bbox": [ + 313, + 522, + 554, + 606 + ], + "type": "inline_equation", + "content": "\\lambda \\to +\\infty" + }, + { + "bbox": [ + 313, + 522, + 554, + 606 + ], + "type": "text", + "content": ", the personalized model converges to the global model, whereas " + }, + { + "bbox": [ + 313, + 522, + 554, + 606 + ], + "type": "inline_equation", + "content": "\\lambda \\to 0" + }, + { + "bbox": [ + 313, + 522, + 554, + 606 + ], + "type": "text", + "content": " leads to a more distinct personalized effect." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 607, + 554, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 607, + 554, + 667 + ], + "spans": [ + { + "bbox": [ + 313, + 607, + 554, + 667 + ], + "type": "text", + "content": "The detailed process is provided in Algorithm 1. During training, the personalized model is updated synchronously with the global model. Each client retains " + }, + { + "bbox": [ + 313, + 607, + 554, + 667 + ], + "type": "inline_equation", + "content": "v_{i}^{t + 1}" + }, + { + "bbox": [ + 313, + 607, + 554, + 667 + ], + "type": "text", + "content": " locally for local tasks and prepares the model update " + }, + { + "bbox": [ + 313, + 607, + 554, + 667 + ], + "type": "inline_equation", + "content": "w_{i}^{t}" + }, + { + "bbox": [ + 313, + 607, + 554, + 667 + ], + "type": "text", + "content": " for the subsequent Segment Exchange." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 672, + 469, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 469, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 469, + 685 + ], + "type": "text", + "content": "4.3. Privacy: Segment Exchange" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "To prevent the server from obtaining each client's complete model, the client mixes its model by exchanging segments" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3918" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 76, + 181, + 87 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 76, + 181, + 87 + ], + "spans": [ + { + "bbox": [ + 63, + 76, + 181, + 87 + ], + "type": "text", + "content": "Algorithm 1: Local Training" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 89, + 271, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 89, + 271, + 125 + ], + "spans": [ + { + "bbox": [ + 67, + 89, + 271, + 125 + ], + "type": "text", + "content": "Input : Personalized model " + }, + { + "bbox": [ + 67, + 89, + 271, + 125 + ], + "type": "inline_equation", + "content": "v_{i}^{t}" + }, + { + "bbox": [ + 67, + 89, + 271, + 125 + ], + "type": "text", + "content": ", model received from server " + }, + { + "bbox": [ + 67, + 89, + 271, + 125 + ], + "type": "inline_equation", + "content": "w^{t}" + }, + { + "bbox": [ + 67, + 89, + 271, + 125 + ], + "type": "text", + "content": ", regularization coefficient " + }, + { + "bbox": [ + 67, + 89, + 271, + 125 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 67, + 89, + 271, + 125 + ], + "type": "text", + "content": ", learning rate " + }, + { + "bbox": [ + 67, + 89, + 271, + 125 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 67, + 89, + 271, + 125 + ], + "type": "text", + "content": ", local epochs " + }, + { + "bbox": [ + 67, + 89, + 271, + 125 + ], + "type": "inline_equation", + "content": "E" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 68, + 125, + 266, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 125, + 266, + 138 + ], + "spans": [ + { + "bbox": [ + 68, + 125, + 266, + 138 + ], + "type": "text", + "content": "Output: Updates " + }, + { + "bbox": [ + 68, + 125, + 266, + 138 + ], + "type": "inline_equation", + "content": "w_{i}^{t + 1}" + }, + { + "bbox": [ + 68, + 125, + 266, + 138 + ], + "type": "text", + "content": ", personalized model " + }, + { + "bbox": [ + 68, + 125, + 266, + 138 + ], + "type": "inline_equation", + "content": "v_{i}^{t + 1}" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 60, + 137, + 280, + 230 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 60, + 137, + 168, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 137, + 168, + 150 + ], + "spans": [ + { + "bbox": [ + 60, + 137, + 168, + 150 + ], + "type": "text", + "content": "1 " + }, + { + "bbox": [ + 60, + 137, + 168, + 150 + ], + "type": "inline_equation", + "content": "w_{i}^{t + 1}\\gets w^{t},v_{i}^{t + 1}\\leftarrow v_{i}^{t};" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 60, + 150, + 182, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 150, + 182, + 161 + ], + "spans": [ + { + "bbox": [ + 60, + 150, + 182, + 161 + ], + "type": "text", + "content": "2 for epoch " + }, + { + "bbox": [ + 60, + 150, + 182, + 161 + ], + "type": "inline_equation", + "content": "\\leftarrow 0" + }, + { + "bbox": [ + 60, + 150, + 182, + 161 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 60, + 150, + 182, + 161 + ], + "type": "inline_equation", + "content": "E - 1" + }, + { + "bbox": [ + 60, + 150, + 182, + 161 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 60, + 162, + 280, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 162, + 280, + 187 + ], + "spans": [ + { + "bbox": [ + 60, + 162, + 280, + 187 + ], + "type": "text", + "content": "3 /*update personalized model \\*/ " + }, + { + "bbox": [ + 60, + 162, + 280, + 187 + ], + "type": "inline_equation", + "content": "v_{i}^{t + 1}\\gets v_{i}^{t + 1} - \\eta (\\nabla \\mathcal{L}_{i}(v_{i}^{t + 1}) + \\lambda (v_{i}^{t + 1} - w^{t}))" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 60, + 188, + 216, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 188, + 216, + 213 + ], + "spans": [ + { + "bbox": [ + 60, + 188, + 216, + 213 + ], + "type": "text", + "content": "4 " + }, + { + "bbox": [ + 60, + 188, + 216, + 213 + ], + "type": "inline_equation", + "content": "\\begin{array}{r}\\left\\lfloor \\begin{array}{c}/*\\mathrm{update~global~model}*/\\\\ w_i^{t + 1}\\gets w_i^{t + 1} - \\eta \\nabla \\mathcal{L}_i(w_i^{t + 1}); \\end{array} \\right. \\end{array}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 60, + 215, + 143, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 215, + 143, + 230 + ], + "spans": [ + { + "bbox": [ + 60, + 215, + 143, + 230 + ], + "type": "text", + "content": "5 return " + }, + { + "bbox": [ + 60, + 215, + 143, + 230 + ], + "type": "inline_equation", + "content": "w_{i}^{t + 1}, v_{i}^{t + 1}" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 256, + 295, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 295, + 304 + ], + "type": "text", + "content": "with a partner before submitting local updates. However, to ensure that clients do not reveal their models to their exchange partners during this process, we design a secure model segment exchange protocol, detailed in Algorithm 2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 304, + 296, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 296, + 425 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 296, + 425 + ], + "type": "text", + "content": "An simple example is illustrated in Fig. 2. The two parties first use the Diffie-Hellman key exchange to agree on a mask " + }, + { + "bbox": [ + 55, + 304, + 296, + 425 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 304, + 296, + 425 + ], + "type": "text", + "content": " that determines which segments are exchanged. " + }, + { + "bbox": [ + 55, + 304, + 296, + 425 + ], + "type": "inline_equation", + "content": "w_{i}^{-m}" + }, + { + "bbox": [ + 55, + 304, + 296, + 425 + ], + "type": "text", + "content": " is encrypted using client " + }, + { + "bbox": [ + 55, + 304, + 296, + 425 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 55, + 304, + 296, + 425 + ], + "type": "text", + "content": "'s key " + }, + { + "bbox": [ + 55, + 304, + 296, + 425 + ], + "type": "inline_equation", + "content": "r_{i}" + }, + { + "bbox": [ + 55, + 304, + 296, + 425 + ], + "type": "text", + "content": ", ensuring that " + }, + { + "bbox": [ + 55, + 304, + 296, + 425 + ], + "type": "inline_equation", + "content": "c_{j}" + }, + { + "bbox": [ + 55, + 304, + 296, + 425 + ], + "type": "text", + "content": " cannot access " + }, + { + "bbox": [ + 55, + 304, + 296, + 425 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 55, + 304, + 296, + 425 + ], + "type": "text", + "content": "'s private information while still enabling segment exchange through the properties of the XOR operation. Meanwhile, the server can decrypt the client key to obtain the mixed model. However, since it does not know " + }, + { + "bbox": [ + 55, + 304, + 296, + 425 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 304, + 296, + 425 + ], + "type": "text", + "content": " or the exchange partner, it cannot recover the original client model, thereby preserving privacy." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 64, + 434, + 296, + 547 + ], + "blocks": [ + { + "bbox": [ + 64, + 434, + 296, + 547 + ], + "lines": [ + { + "bbox": [ + 64, + 434, + 296, + 547 + ], + "spans": [ + { + "bbox": [ + 64, + 434, + 296, + 547 + ], + "type": "image", + "image_path": "1c85a6a0161442e27f2aed6784eaae72fed3269303ba0b7ac9d0515c4d4ce0b9.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 554, + 261, + 566 + ], + "lines": [ + { + "bbox": [ + 88, + 554, + 261, + 566 + ], + "spans": [ + { + "bbox": [ + 88, + 554, + 261, + 566 + ], + "type": "text", + "content": "Figure 2. The Workflow of Segment Exchange." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 587, + 239, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 239, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 239, + 601 + ], + "type": "text", + "content": "4.4. Robustness: Segment Aggregation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "type": "text", + "content": "To mitigate the curse of dimensionality, FLSeg partitions the mixed model according to the mode of segmentation and then applies RAR to each segment, rather than performing RAR directly on the entire mixed model. This approach not only ensures that each segment in the RAR originates from different clients but also effectively mitigates the challenges posed by high input dimensionality. The detailed procedure is provided in Algorithm 3." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 67, + 701, + 294, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 294, + 714 + ], + "type": "text", + "content": "After grouping segments based on the mode of segmen" + } + ] + } + ], + "index": 15 + }, + { + "type": "code", + "bbox": [ + 316, + 88, + 553, + 318 + ], + "blocks": [ + { + "bbox": [ + 321, + 75, + 493, + 87 + ], + "lines": [ + { + "bbox": [ + 321, + 75, + 493, + 87 + ], + "spans": [ + { + "bbox": [ + 321, + 75, + 493, + 87 + ], + "type": "text", + "content": "Algorithm 2: Segment Exchange Protocal" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "lines": [ + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "spans": [ + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "text", + "content": "Input: Client " + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "text", + "content": " 's model update " + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "text", + "content": " ,server's public-private key pair " + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "inline_equation", + "content": "(pk_S,sk_S)" + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "text", + "content": " ,the exchange partner " + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "inline_equation", + "content": "C_j" + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "text", + "content": " \n1 client " + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "inline_equation", + "content": "c_{j}" + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "text", + "content": " \n2 generate client's key " + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "text", + "content": " . \n3 perform Diffie-Hellman key exchange with " + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "inline_equation", + "content": "c_{j}" + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "text", + "content": " to obtain mask " + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "text", + "content": " . \n4 " + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "inline_equation", + "content": "w_{i}^{m} / w_{i}^{-m}\\gets" + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "text", + "content": " Set the " + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "text", + "content": " -th segment " + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "inline_equation", + "content": "\\gamma^{(k)} = 0" + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "text", + "content": " if the " + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "text", + "content": " -th bit of " + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "inline_equation", + "content": "1 / 0" + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "text", + "content": " . \n5 exchange " + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "inline_equation", + "content": "w_{i}^{-m}\\oplus r_{i}" + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "inline_equation", + "content": "\\mathrm{Enc}(r_i,pk_S)" + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "inline_equation", + "content": "c_{j}" + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "text", + "content": " . \n6 compute " + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "inline_equation", + "content": "(e_i)_{\\mathrm{mix}} = w_i^m\\oplus (w_j^{-m}\\oplus r_j)" + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "inline_equation", + "content": "= (w_{i})_{\\mathrm{mix}}\\oplus r_{j}" + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "text", + "content": " . \n7 send " + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "inline_equation", + "content": "(e_i)_{\\mathrm{mix}}" + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "text", + "content": " ,Enc " + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "inline_equation", + "content": "(r_j,pk_S)" + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "text", + "content": " to server S; \n8 server S \n9 decrypts each client " + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "text", + "content": " 's key " + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "inline_equation", + "content": "r_j = \\mathrm{Dec}(\\mathrm{Enc}(r_j,pk_S),sk_S)" + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "text", + "content": " . \n10 decrypts each client " + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "text", + "content": " 's mixed model " + }, + { + "bbox": [ + 316, + 88, + 553, + 318 + ], + "type": "inline_equation", + "content": "(w_{i})_{\\mathrm{mix}}\\leftarrow (e_{i})_{\\mathrm{mix}}\\oplus r_{j};" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "code_body" + } + ], + "index": 17, + "sub_type": "algorithm" + }, + { + "type": "code", + "bbox": [ + 316, + 350, + 553, + 567 + ], + "blocks": [ + { + "bbox": [ + 321, + 336, + 468, + 348 + ], + "lines": [ + { + "bbox": [ + 321, + 336, + 468, + 348 + ], + "spans": [ + { + "bbox": [ + 321, + 336, + 468, + 348 + ], + "type": "text", + "content": "Algorithm 3: Segment Aggregation" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 316, + 350, + 553, + 567 + ], + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 567 + ], + "type": "text", + "content": "Input: uploaded mixed models " + }, + { + "bbox": [ + 316, + 350, + 553, + 567 + ], + "type": "inline_equation", + "content": "\\{(w_{i})_{\\mathrm{mix}}\\}_{i\\in [n]}" + }, + { + "bbox": [ + 316, + 350, + 553, + 567 + ], + "type": "text", + "content": " the mode of segmentation " + }, + { + "bbox": [ + 316, + 350, + 553, + 567 + ], + "type": "inline_equation", + "content": "\\{\\gamma^{(1)},\\dots ,\\gamma^{(d)}\\}" + }, + { + "bbox": [ + 316, + 350, + 553, + 567 + ], + "type": "text", + "content": " Output: the aggregation result " + }, + { + "bbox": [ + 316, + 350, + 553, + 567 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 316, + 350, + 553, + 567 + ], + "type": "text", + "content": " \n1 for each model " + }, + { + "bbox": [ + 316, + 350, + 553, + 567 + ], + "type": "inline_equation", + "content": "(w_{i})_{\\mathrm{mix}}" + }, + { + "bbox": [ + 316, + 350, + 553, + 567 + ], + "type": "text", + "content": " do \n2 split model into segments " + }, + { + "bbox": [ + 316, + 350, + 553, + 567 + ], + "type": "inline_equation", + "content": "\\{\\gamma_i^{(1)},\\ldots ,\\gamma_i^{(d)}\\} \\gets (w_i)_{\\mathrm{mix}};" + }, + { + "bbox": [ + 316, + 350, + 553, + 567 + ], + "type": "text", + "content": " \n3 for each segment " + }, + { + "bbox": [ + 316, + 350, + 553, + 567 + ], + "type": "inline_equation", + "content": "\\gamma^{(k)}\\in \\{\\gamma^{(1)},\\dots ,\\gamma^{(d)}\\}" + }, + { + "bbox": [ + 316, + 350, + 553, + 567 + ], + "type": "text", + "content": " do \n4 apply the RAR to the segments to establish a benchmark " + }, + { + "bbox": [ + 316, + 350, + 553, + 567 + ], + "type": "inline_equation", + "content": "\\overline{\\gamma}^{(k)}\\gets \\mathrm{RAR}(\\{\\gamma_i^{(k)}\\}_{i\\in [n]})" + }, + { + "bbox": [ + 316, + 350, + 553, + 567 + ], + "type": "text", + "content": " \n5 calculate the similarity " + }, + { + "bbox": [ + 316, + 350, + 553, + 567 + ], + "type": "inline_equation", + "content": "s_i^{(k)}" + }, + { + "bbox": [ + 316, + 350, + 553, + 567 + ], + "type": "text", + "content": " for each segment " + }, + { + "bbox": [ + 316, + 350, + 553, + 567 + ], + "type": "inline_equation", + "content": "\\gamma_{i}^{(k)}" + }, + { + "bbox": [ + 316, + 350, + 553, + 567 + ], + "type": "text", + "content": " with the benchmark " + }, + { + "bbox": [ + 316, + 350, + 553, + 567 + ], + "type": "inline_equation", + "content": "\\overline{\\gamma}^{(k)}" + }, + { + "bbox": [ + 316, + 350, + 553, + 567 + ], + "type": "text", + "content": " using Eq. (2); \n6 compute each model's similarity score as " + }, + { + "bbox": [ + 316, + 350, + 553, + 567 + ], + "type": "inline_equation", + "content": "s_i = \\sum_{k = 1}^d s_i^{(k)}" + }, + { + "bbox": [ + 316, + 350, + 553, + 567 + ], + "type": "text", + "content": " and select the top " + }, + { + "bbox": [ + 316, + 350, + 553, + 567 + ], + "type": "inline_equation", + "content": "n - 2f" + }, + { + "bbox": [ + 316, + 350, + 553, + 567 + ], + "type": "text", + "content": " models with the highest scores to form set " + }, + { + "bbox": [ + 316, + 350, + 553, + 567 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 316, + 350, + 553, + 567 + ], + "type": "text", + "content": " \n7 aggregate " + }, + { + "bbox": [ + 316, + 350, + 553, + 567 + ], + "type": "inline_equation", + "content": "w\\gets \\sum_{i\\in I}(w_i)_{\\mathrm{mix}}" + }, + { + "bbox": [ + 316, + 350, + 553, + 567 + ], + "type": "text", + "content": " \n8 return " + }, + { + "bbox": [ + 316, + 350, + 553, + 567 + ], + "type": "inline_equation", + "content": "w" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "code_body" + } + ], + "index": 19, + "sub_type": "algorithm" + }, + { + "bbox": [ + 313, + 593, + 555, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 593, + 555, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 593, + 555, + 654 + ], + "type": "text", + "content": "tation, " + }, + { + "bbox": [ + 313, + 593, + 555, + 654 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 593, + 555, + 654 + ], + "type": "text", + "content": " applies RAR to each segment group to obtain the benchmark " + }, + { + "bbox": [ + 313, + 593, + 555, + 654 + ], + "type": "inline_equation", + "content": "\\overline{\\gamma}^{(k)}" + }, + { + "bbox": [ + 313, + 593, + 555, + 654 + ], + "type": "text", + "content": " and uses it as an honest reference to compute the similarity " + }, + { + "bbox": [ + 313, + 593, + 555, + 654 + ], + "type": "inline_equation", + "content": "s_i^{(k)}" + }, + { + "bbox": [ + 313, + 593, + 555, + 654 + ], + "type": "text", + "content": " of each segment " + }, + { + "bbox": [ + 313, + 593, + 555, + 654 + ], + "type": "inline_equation", + "content": "\\gamma_i^{(k)}" + }, + { + "bbox": [ + 313, + 593, + 555, + 654 + ], + "type": "text", + "content": " using Eq. (2), then sums the segment similarities to obtain the model's score " + }, + { + "bbox": [ + 313, + 593, + 555, + 654 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 313, + 593, + 555, + 654 + ], + "type": "text", + "content": " for this round." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 366, + 662, + 553, + 695 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 662, + 553, + 695 + ], + "spans": [ + { + "bbox": [ + 366, + 662, + 553, + 695 + ], + "type": "interline_equation", + "content": "s _ {i} ^ {(k)} = \\frac {\\left(\\gamma_ {i} ^ {(k)} - \\mu\\right) \\cdot \\left(\\bar {\\gamma} ^ {(k)} - \\mu\\right)}{\\| \\gamma_ {i} ^ {(k)} - \\mu \\| \\| (\\bar {\\gamma} ^ {(k)} - \\mu) \\|}, \\tag {2}", + "image_path": "609ae51017ebf9249009d33ba5c0456412b66a469a737009c80a8bd1994d82cf.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 699, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 699, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 699, + 555, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 699, + 555, + 715 + ], + "type": "inline_equation", + "content": "z = |\\gamma^{(k)}|" + }, + { + "bbox": [ + 313, + 699, + 555, + 715 + ], + "type": "text", + "content": " is the number of parameters of " + }, + { + "bbox": [ + 313, + 699, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\gamma_{i}^{(k)}" + }, + { + "bbox": [ + 313, + 699, + 555, + 715 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3919" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 69, + 296, + 103 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 69, + 296, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 69, + 296, + 103 + ], + "type": "inline_equation", + "content": "\\overline{\\gamma}^{(k)}" + }, + { + "bbox": [ + 55, + 69, + 296, + 103 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 55, + 69, + 296, + 103 + ], + "type": "inline_equation", + "content": "\\mu = \\frac{\\sum_{j = 1}^{z}\\gamma_{i,j}^{(k)} + \\sum_{j = 1}^{z}\\overline{\\gamma}_{j}^{(k)}}{2z}" + }, + { + "bbox": [ + 55, + 69, + 296, + 103 + ], + "type": "text", + "content": " is the average value of all parameters in " + }, + { + "bbox": [ + 55, + 69, + 296, + 103 + ], + "type": "inline_equation", + "content": "\\gamma_i^{(k)}" + }, + { + "bbox": [ + 55, + 69, + 296, + 103 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 69, + 296, + 103 + ], + "type": "inline_equation", + "content": "\\overline{\\gamma}^{(k)}" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 103, + 296, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 103, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 103, + 296, + 140 + ], + "type": "text", + "content": "Since, in the worst case, each honest client's partner may be a malicious client, " + }, + { + "bbox": [ + 55, + 103, + 296, + 140 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 103, + 296, + 140 + ], + "type": "text", + "content": " select the top " + }, + { + "bbox": [ + 55, + 103, + 296, + 140 + ], + "type": "inline_equation", + "content": "n - 2f" + }, + { + "bbox": [ + 55, + 103, + 296, + 140 + ], + "type": "text", + "content": " clients with the highest scores for aggregation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 148, + 124, + 161 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 148, + 124, + 161 + ], + "spans": [ + { + "bbox": [ + 55, + 148, + 124, + 161 + ], + "type": "text", + "content": "4.5. Summary" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 166, + 296, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 166, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 166, + 296, + 274 + ], + "type": "text", + "content": "The mode of segmentation in FLSeg remains unchanged throughout the process. Before training begins, the server distributes its public key " + }, + { + "bbox": [ + 55, + 166, + 296, + 274 + ], + "type": "inline_equation", + "content": "pk_{S}" + }, + { + "bbox": [ + 55, + 166, + 296, + 274 + ], + "type": "text", + "content": " to all clients. In each training round, clients first perform Local Training (Algorithm 1), then conduct Segment Exchange to obtain the mixed model (Algorithm 2), and finally, the server applies Segment Aggregation (Algorithm 3) to the received global models. The overall process is detailed in the algorithm provided in the Appendix A." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 274, + 296, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 274, + 296, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 296, + 395 + ], + "type": "text", + "content": "FLSeg does not restrict how clients select partners; however, to prevent honest clients from continuously collaborating with malicious clients, we design a reputation system based on computed similarity scores. Specifically, each client " + }, + { + "bbox": [ + 55, + 274, + 296, + 395 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 55, + 274, + 296, + 395 + ], + "type": "text", + "content": " initializes its reputation list " + }, + { + "bbox": [ + 55, + 274, + 296, + 395 + ], + "type": "inline_equation", + "content": "R_{i} = \\{r_{1} = 0,\\dots ,r_{n} = 0\\}" + }, + { + "bbox": [ + 55, + 274, + 296, + 395 + ], + "type": "text", + "content": ". After each round of aggregation, the server normalizes the similarity scores of each model using Eq. (3) and sends these scores to the corresponding clients. A lower score for an honest client indicates that its model may have been influenced by malicious updates." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 80, + 413, + 295, + 442 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 413, + 295, + 442 + ], + "spans": [ + { + "bbox": [ + 80, + 413, + 295, + 442 + ], + "type": "interline_equation", + "content": "s _ {i} ^ {\\prime} = 2 \\left(\\frac {s _ {i} - \\operatorname* {m i n} \\left(\\left\\{s _ {i} \\right\\} _ {i \\in [ n ]}\\right)}{\\operatorname* {m a x} \\left(\\left\\{s _ {i} \\right\\} _ {i \\in [ n ]}\\right) - \\operatorname* {m i n} \\left(\\left\\{s _ {i} \\right\\} _ {i \\in [ n ]}\\right)}\\right) - 1 \\tag {3}", + "image_path": "2d00c3caea9bd91e92fd206c2cb0a32476e0bea503192157bab0ae464c651e6e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 451, + 296, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 296, + 522 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 296, + 522 + ], + "type": "text", + "content": "Clients use these scores to update the reputation values of their partners as " + }, + { + "bbox": [ + 55, + 451, + 296, + 522 + ], + "type": "inline_equation", + "content": "r_j = r_j + s_i'" + }, + { + "bbox": [ + 55, + 451, + 296, + 522 + ], + "type": "text", + "content": ". In subsequent rounds, clients can select partners with higher reputation values from their reputation lists for Segment Exchange. Practical test results for the reputation system are provided in Appendix B." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 534, + 115, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 115, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 115, + 548 + ], + "type": "text", + "content": "5. Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 554, + 180, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 554, + 180, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 554, + 180, + 567 + ], + "type": "text", + "content": "5.1. Semi-Honest Security" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 572, + 296, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 572, + 296, + 633 + ], + "spans": [ + { + "bbox": [ + 55, + 572, + 296, + 633 + ], + "type": "text", + "content": "We use the Universally Composable Security Framework [7] to prove that FLSeg satisfies semi-honest security, meaning that FLSeg does not leak any information as long as the adversaries follow the protocol. We first present definition of semi-honest security [14]:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "text", + "content": "Definition 1 (Real-World). For protocol " + }, + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "text", + "content": ", there exists a semi-honest adversary " + }, + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "text", + "content": " that can control a subset of participants " + }, + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "text", + "content": " will follow the protocol " + }, + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "text", + "content": " but will attempt to infer the private inputs " + }, + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\{x_1, \\ldots, x_n\\}" + }, + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "text", + "content": " based on the available information. The probability distribution of the inference results is denoted as " + }, + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\text{Real}_{\\pi, \\mathcal{A}}(\\mathcal{P}, \\{x_1, \\ldots, x_n\\})" + }, + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "type": "text", + "content": "Definition 2 (Ideal-World). Assume there exists an ideal functionality " + }, + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "type": "text", + "content": " that can execute any interaction in protocol " + }, + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "type": "text", + "content": " with absolute security. There also exists a simulator Sim that simulates interactions " + }, + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "type": "text", + "content": " in the same manner as the real-world adversary " + }, + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "type": "text", + "content": ". The probability distribution of the inference results is denoted as Ideal" + }, + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "type": "inline_equation", + "content": "_{\\mathcal{F},\\text{Sim}}(\\mathcal{P},\\{x_1,\\ldots,x_n\\})" + }, + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 148, + 554, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 148, + 554, + 196 + ], + "spans": [ + { + "bbox": [ + 313, + 148, + 554, + 196 + ], + "type": "text", + "content": "Definition 3 (Semi-Honest). If " + }, + { + "bbox": [ + 313, + 148, + 554, + 196 + ], + "type": "inline_equation", + "content": "\\text{Real}_{\\pi, \\mathcal{A}}(\\mathcal{P}, \\{x_1, \\ldots, x_n\\})" + }, + { + "bbox": [ + 313, + 148, + 554, + 196 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 148, + 554, + 196 + ], + "type": "inline_equation", + "content": "\\text{Ideal}_{\\mathcal{F}, \\text{Sim}}(\\mathcal{P}, \\{x_1, \\ldots, x_n\\})" + }, + { + "bbox": [ + 313, + 148, + 554, + 196 + ], + "type": "text", + "content": " are computationally indistinguishable to an honest-but-curious adversary " + }, + { + "bbox": [ + 313, + 148, + 554, + 196 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 148, + 554, + 196 + ], + "type": "text", + "content": ", then protocol " + }, + { + "bbox": [ + 313, + 148, + 554, + 196 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 313, + 148, + 554, + 196 + ], + "type": "text", + "content": " is considered semi-honest secure." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 201, + 554, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 201, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 554, + 262 + ], + "type": "text", + "content": "Theorem 1. Assuming the " + }, + { + "bbox": [ + 313, + 201, + 554, + 262 + ], + "type": "inline_equation", + "content": "FLSeg" + }, + { + "bbox": [ + 313, + 201, + 554, + 262 + ], + "type": "text", + "content": " protocol is denoted as " + }, + { + "bbox": [ + 313, + 201, + 554, + 262 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 313, + 201, + 554, + 262 + ], + "type": "text", + "content": ", the participants of the protocol " + }, + { + "bbox": [ + 313, + 201, + 554, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\{c_1, \\ldots, c_n, S\\}" + }, + { + "bbox": [ + 313, + 201, + 554, + 262 + ], + "type": "text", + "content": " include all clients and the server, and the private inputs are the gradients of each honest client " + }, + { + "bbox": [ + 313, + 201, + 554, + 262 + ], + "type": "inline_equation", + "content": "\\{w_1, \\ldots, w_n\\}" + }, + { + "bbox": [ + 313, + 201, + 554, + 262 + ], + "type": "text", + "content": ". Then " + }, + { + "bbox": [ + 313, + 201, + 554, + 262 + ], + "type": "inline_equation", + "content": "FLSeg" + }, + { + "bbox": [ + 313, + 201, + 554, + 262 + ], + "type": "text", + "content": " is semi-honest secure." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 266, + 554, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 554, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 554, + 350 + ], + "type": "text", + "content": "Proof. We use the Hybrid Argument [12]. By gradually replacing components of the protocol's multiple interaction stages with the ideal functionality " + }, + { + "bbox": [ + 313, + 266, + 554, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 266, + 554, + 350 + ], + "type": "text", + "content": ", we demonstrate that each stage's replacement does not significantly alter the probability distribution of the adversary's inference results. Assuming " + }, + { + "bbox": [ + 313, + 266, + 554, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 266, + 554, + 350 + ], + "type": "text", + "content": " represents the entire parameter space of the model, we define the following three hybrids:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 350, + 553, + 418 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 314, + 350, + 498, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 350, + 498, + 361 + ], + "spans": [ + { + "bbox": [ + 314, + 350, + 498, + 361 + ], + "type": "text", + "content": "- Hybrid 0: The original FLSeg Framework." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 361, + 553, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 361, + 553, + 389 + ], + "spans": [ + { + "bbox": [ + 314, + 361, + 553, + 389 + ], + "type": "text", + "content": "- Hybrid 1: Similar to Hybrid 0, but replaces " + }, + { + "bbox": [ + 314, + 361, + 553, + 389 + ], + "type": "inline_equation", + "content": "w_{i}^{-m} \\oplus r_{i}" + }, + { + "bbox": [ + 314, + 361, + 553, + 389 + ], + "type": "text", + "content": " exchanged by clients with random values " + }, + { + "bbox": [ + 314, + 361, + 553, + 389 + ], + "type": "inline_equation", + "content": "\\phi_{i}^{(1)} \\leftarrow \\mathcal{T}" + }, + { + "bbox": [ + 314, + 361, + 553, + 389 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 389, + 553, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 389, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 314, + 389, + 553, + 418 + ], + "type": "text", + "content": "- Hybrid 2: Similar to Hybrid 1, but replaces " + }, + { + "bbox": [ + 314, + 389, + 553, + 418 + ], + "type": "inline_equation", + "content": "(w_{i})_{\\mathrm{mix}}" + }, + { + "bbox": [ + 314, + 389, + 553, + 418 + ], + "type": "text", + "content": " transmitted to the server with random values " + }, + { + "bbox": [ + 314, + 389, + 553, + 418 + ], + "type": "inline_equation", + "content": "\\phi_{i}^{(2)} \\stackrel{\\S}{\\leftarrow} \\mathcal{T}" + }, + { + "bbox": [ + 314, + 389, + 553, + 418 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 423, + 554, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 554, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 554, + 447 + ], + "type": "text", + "content": "Lemma 1 (Hybrid 1). Hybrid 0 and Hybrid 1 are computationally indistinguishable if FLSeg is semi-honest secure." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 454, + 554, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 554, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 554, + 495 + ], + "type": "text", + "content": "Proof. Since " + }, + { + "bbox": [ + 313, + 454, + 554, + 495 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 313, + 454, + 554, + 495 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 454, + 554, + 495 + ], + "type": "inline_equation", + "content": "\\phi_i^{(1)}" + }, + { + "bbox": [ + 313, + 454, + 554, + 495 + ], + "type": "text", + "content": " follow the same distribution, " + }, + { + "bbox": [ + 313, + 454, + 554, + 495 + ], + "type": "inline_equation", + "content": "w_i^{(\\neg m)} \\oplus r_i" + }, + { + "bbox": [ + 313, + 454, + 554, + 495 + ], + "type": "text", + "content": " is also distributed identically to " + }, + { + "bbox": [ + 313, + 454, + 554, + 495 + ], + "type": "inline_equation", + "content": "\\phi_i^{(1)}" + }, + { + "bbox": [ + 313, + 454, + 554, + 495 + ], + "type": "text", + "content": ". Thus, the advantage of " + }, + { + "bbox": [ + 313, + 454, + 554, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 454, + 554, + 495 + ], + "type": "text", + "content": " is given by Eq. (4)." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 323, + 498, + 554, + 519 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 498, + 554, + 519 + ], + "spans": [ + { + "bbox": [ + 323, + 498, + 554, + 519 + ], + "type": "interline_equation", + "content": "\\left| \\mathbb {P} [ \\mathcal {A} (\\phi_ {i} ^ {(1)}) = 1 ] - \\mathbb {P} [ \\mathcal {A} (w _ {i} ^ {- m} \\oplus r _ {i}) = 1 ] \\right| \\leq \\operatorname {n e g} (d) \\tag {4}", + "image_path": "47f9c0740dc91e88134c4076cf7ed34aacad11a03f37c11db1cbbef8a07d67a0.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 524, + 554, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 524, + 554, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 524, + 554, + 548 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 524, + 554, + 548 + ], + "type": "inline_equation", + "content": "\\mathrm{neg}(d)" + }, + { + "bbox": [ + 313, + 524, + 554, + 548 + ], + "type": "text", + "content": " is a negligible function. Hence, Hybrid 0 and Hybrid 1 are computationally indistinguishable." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 555, + 554, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 554, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 554, + 579 + ], + "type": "text", + "content": "Lemma 2 (Hybrid 2). Hybrid 1 and Hybrid 2 are computationally indistinguishable if FLSeg is semi-honest secure." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 586, + 554, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 554, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 554, + 647 + ], + "type": "text", + "content": "Proof. Since the adversary " + }, + { + "bbox": [ + 313, + 586, + 554, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 586, + 554, + 647 + ], + "type": "text", + "content": " cannot control both the client and the server simultaneously, " + }, + { + "bbox": [ + 313, + 586, + 554, + 647 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 586, + 554, + 647 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 586, + 554, + 647 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 586, + 554, + 647 + ], + "type": "text", + "content": " are random to " + }, + { + "bbox": [ + 313, + 586, + 554, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 313, + 586, + 554, + 647 + ], + "type": "text", + "content": ". If " + }, + { + "bbox": [ + 313, + 586, + 554, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 313, + 586, + 554, + 647 + ], + "type": "text", + "content": " attempts to recover the private " + }, + { + "bbox": [ + 313, + 586, + 554, + 647 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 313, + 586, + 554, + 647 + ], + "type": "text", + "content": " by random matching, the probability of success is " + }, + { + "bbox": [ + 313, + 586, + 554, + 647 + ], + "type": "inline_equation", + "content": "\\frac{1}{n} \\cdot \\frac{1}{2^d}" + }, + { + "bbox": [ + 313, + 586, + 554, + 647 + ], + "type": "text", + "content": ". Thus, the advantage of " + }, + { + "bbox": [ + 313, + 586, + 554, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 586, + 554, + 647 + ], + "type": "text", + "content": " is given by Eq. (5)." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 328, + 651, + 553, + 674 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 651, + 553, + 674 + ], + "spans": [ + { + "bbox": [ + 328, + 651, + 553, + 674 + ], + "type": "interline_equation", + "content": "\\left| \\mathbb {P} [ \\mathcal {A} ((w _ {i}) _ {\\mathrm {m i x}}) = 1 ] - \\mathbb {P} [ \\mathcal {A} (\\phi_ {i} ^ {(2)}) = 1 ] \\right| \\leq \\frac {1}{n} \\cdot \\frac {1}{2 ^ {d}} \\tag {5}", + "image_path": "c7ea59e03ff5c3c83696a35fb178210787ad8f36c27bce1dde1a652d81bd994e.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": "Since this advantage is negligible for a polynomial-time adversary, it can be ignored. Thus, Hybrid 1 and Hybrid 2 are computationally indistinguishable." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3920" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 83, + 70, + 529, + 152 + ], + "blocks": [ + { + "bbox": [ + 83, + 70, + 529, + 152 + ], + "lines": [ + { + "bbox": [ + 83, + 70, + 529, + 152 + ], + "spans": [ + { + "bbox": [ + 83, + 70, + 529, + 152 + ], + "type": "table", + "html": "
SecAgg [5]Bell et al. [3]ACORN-detect [2]ShieldFL [29]PEFL [25]FLSeg
ClientComplexityO(n2+ ℓn)O(log2 n + ℓ log n)O(log2 n + ℓ log n)O(ℓ(TEnc + TDec))O(ℓ(TEnc + TDec))O(ℓ)
Overhead2,557 ms1,696 ms1,729 ms231,541 ms215,667 ms233 ms
ServerComplexityO(ℓn2)O(n(log2 n + ℓ log n))O(n(log2 n + ℓ log n))O(ℓnTDec)O(ℓn(TEnc + TDec))O(ℓn)
Overhead192 ms/c182 ms/c181 ms/c78,685 ms/c790,032 ms/c78 ms/c
Robustness○○◇◇◇◇
", + "image_path": "deb1ad587ccc2dc4ab6d860926534d1b4b8631e46a7549baa6ca1ca1bb4d2628.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 159, + 555, + 216 + ], + "lines": [ + { + "bbox": [ + 54, + 159, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 54, + 159, + 555, + 216 + ], + "type": "text", + "content": "Table 1. Comparison of computational time complexity and time overhead. " + }, + { + "bbox": [ + 54, + 159, + 555, + 216 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 54, + 159, + 555, + 216 + ], + "type": "text", + "content": " represents the number of clients, " + }, + { + "bbox": [ + 54, + 159, + 555, + 216 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 54, + 159, + 555, + 216 + ], + "type": "text", + "content": " denotes the size of the model parameters. In calculating computation time overhead, let " + }, + { + "bbox": [ + 54, + 159, + 555, + 216 + ], + "type": "inline_equation", + "content": "n = {10}^{3}" + }, + { + "bbox": [ + 54, + 159, + 555, + 216 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 159, + 555, + 216 + ], + "type": "inline_equation", + "content": "\\ell = {10}^{6}" + }, + { + "bbox": [ + 54, + 159, + 555, + 216 + ], + "type": "text", + "content": " . " + }, + { + "bbox": [ + 54, + 159, + 555, + 216 + ], + "type": "inline_equation", + "content": "{T}_{\\text{Enc }}" + }, + { + "bbox": [ + 54, + 159, + 555, + 216 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 159, + 555, + 216 + ], + "type": "inline_equation", + "content": "{T}_{\\text{Dec }}" + }, + { + "bbox": [ + 54, + 159, + 555, + 216 + ], + "type": "text", + "content": " represent the encryption and decryption constant overhead in HE. ”/c” denotes the average time overhead incurred by the server for each client. " + }, + { + "bbox": [ + 54, + 159, + 555, + 216 + ], + "type": "inline_equation", + "content": "\\bigcirc" + }, + { + "bbox": [ + 54, + 159, + 555, + 216 + ], + "type": "text", + "content": " indicates the absence of robustness, " + }, + { + "bbox": [ + 54, + 159, + 555, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 54, + 159, + 555, + 216 + ], + "type": "text", + "content": " denotes support for fixed RAR only,and " + }, + { + "bbox": [ + 54, + 159, + 555, + 216 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 54, + 159, + 555, + 216 + ], + "type": "text", + "content": " denotes support for arbitrary RAR. We omit overheads unrelated to privacy protection,such as training and aggregation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 236, + 297, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 236, + 297, + 285 + ], + "spans": [ + { + "bbox": [ + 54, + 236, + 297, + 285 + ], + "type": "text", + "content": "Through the above hybrid experiments, we have gradually proven the information transmitted from the real-world to the ideal-world is computationally indistinguishable from random values. Thus, we have proven Theorem 1. " + }, + { + "bbox": [ + 54, + 236, + 297, + 285 + ], + "type": "inline_equation", + "content": "\\square" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 294, + 226, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 294, + 226, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 226, + 308 + ], + "type": "text", + "content": "5.2. Efficiency in privacy protection" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 312, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 296, + 361 + ], + "type": "text", + "content": "Assume " + }, + { + "bbox": [ + 55, + 312, + 296, + 361 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 312, + 296, + 361 + ], + "type": "text", + "content": " represents the number of clients and " + }, + { + "bbox": [ + 55, + 312, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 55, + 312, + 296, + 361 + ], + "type": "text", + "content": " denotes the size of the model parameters. The computational overhead on the client side for FLSeg primarily consists of the following components:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 362, + 272, + 400 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 55, + 362, + 241, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 362, + 241, + 375 + ], + "spans": [ + { + "bbox": [ + 55, + 362, + 241, + 375 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(\\ell^2)\\colon m\\gets" + }, + { + "bbox": [ + 55, + 362, + 241, + 375 + ], + "type": "text", + "content": " Diffie-Hellman key exchange," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 375, + 272, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 375, + 272, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 375, + 272, + 388 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 55, + 375, + 272, + 388 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(\\ell): \\text{Calculation of } w_i^m, w_i^{-m}, w_i^{-m} \\oplus r_i, (e_i)_{\\text{mix}}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 149, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 149, + 400 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 149, + 400 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(\\ell):\\mathrm{Enc}(r_j,pk_{\\mathcal{S}})" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 54, + 402, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 402, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 54, + 402, + 295, + 581 + ], + "type": "text", + "content": "Here, we introduce an optimization technique: instead of directly using Diffie-Hellman to generate " + }, + { + "bbox": [ + 54, + 402, + 295, + 581 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 54, + 402, + 295, + 581 + ], + "type": "text", + "content": " or encrypting " + }, + { + "bbox": [ + 54, + 402, + 295, + 581 + ], + "type": "inline_equation", + "content": "r_j" + }, + { + "bbox": [ + 54, + 402, + 295, + 581 + ], + "type": "text", + "content": ", which incurs significant computational time costs, we generate respective random seeds " + }, + { + "bbox": [ + 54, + 402, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\mathsf{seed}_{r_j}" + }, + { + "bbox": [ + 54, + 402, + 295, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 402, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\mathsf{seed}_m" + }, + { + "bbox": [ + 54, + 402, + 295, + 581 + ], + "type": "text", + "content": ". Both clients then use a Pseudo-Random Number Generator (PRNG) [5] to produce " + }, + { + "bbox": [ + 54, + 402, + 295, + 581 + ], + "type": "inline_equation", + "content": "r_j" + }, + { + "bbox": [ + 54, + 402, + 295, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 402, + 295, + 581 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 54, + 402, + 295, + 581 + ], + "type": "text", + "content": ", and only the encrypted seed " + }, + { + "bbox": [ + 54, + 402, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\mathrm{Enc}(\\mathsf{seed}_{r_j},pk_S)" + }, + { + "bbox": [ + 54, + 402, + 295, + 581 + ], + "type": "text", + "content": " is sent. This approach reduces the overhead of Diffie-Hellman and asymmetric encryption to a constant level, depending solely on the seed length. Thus, the primary computational overhead is shifted to generating " + }, + { + "bbox": [ + 54, + 402, + 295, + 581 + ], + "type": "inline_equation", + "content": "r_j" + }, + { + "bbox": [ + 54, + 402, + 295, + 581 + ], + "type": "text", + "content": " using PRNG " + }, + { + "bbox": [ + 54, + 402, + 295, + 581 + ], + "type": "inline_equation", + "content": "(\\mathsf{seed}_{r_j}) = r_j" + }, + { + "bbox": [ + 54, + 402, + 295, + 581 + ], + "type": "text", + "content": " and performing XOR operations, with a complexity of " + }, + { + "bbox": [ + 54, + 402, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(\\ell)" + }, + { + "bbox": [ + 54, + 402, + 295, + 581 + ], + "type": "text", + "content": ". Similarly, the decryption overhead on the server is minimized, focusing mainly on decrypting the mixed model " + }, + { + "bbox": [ + 54, + 402, + 295, + 581 + ], + "type": "inline_equation", + "content": "(w_i)_{\\mathrm{mix}}" + }, + { + "bbox": [ + 54, + 402, + 295, + 581 + ], + "type": "text", + "content": " using XOR operations, resulting in a complexity of " + }, + { + "bbox": [ + 54, + 402, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(\\ell n)" + }, + { + "bbox": [ + 54, + 402, + 295, + 581 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "text", + "content": "To demonstrate the advantages of FLSeg in efficient privacy protection, we compare the computational time complexity and time overhead of FLSeg with SOTA privacy-preserving FL that also satisfy semi-honest security, including SecAgg [5], Bell et al. [3], ShieldFL [29], RoFL [28], and PEFL [25]. The comparison results are presented in Tab. 1. The time complexity of ShieldFL and PEFL is asymptotically " + }, + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "O(\\ell)" + }, + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "text", + "content": "; however, the Paillier HE they employ incurs substantial constant overheads, " + }, + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{Enc}}" + }, + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{Dec}}" + }, + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "text", + "content": ". SecAgg, ACORN, and Bell et al. are aggregation schemes based on SMPC techniques, requiring the compu" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 236, + 555, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 555, + 321 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 555, + 321 + ], + "type": "text", + "content": "tation of pairwise masking for all other clients. To more intuitively demonstrate the advantages of FLSeg, we tested each method's runtime in a practical setting with " + }, + { + "bbox": [ + 313, + 236, + 555, + 321 + ], + "type": "inline_equation", + "content": "n = 10^3" + }, + { + "bbox": [ + 313, + 236, + 555, + 321 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 236, + 555, + 321 + ], + "type": "inline_equation", + "content": "\\ell = 10^6" + }, + { + "bbox": [ + 313, + 236, + 555, + 321 + ], + "type": "text", + "content": ". The results show that FLSeg achieves an " + }, + { + "bbox": [ + 313, + 236, + 555, + 321 + ], + "type": "inline_equation", + "content": "8 \\times" + }, + { + "bbox": [ + 313, + 236, + 555, + 321 + ], + "type": "text", + "content": " improvement in client-side and a " + }, + { + "bbox": [ + 313, + 236, + 555, + 321 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 313, + 236, + 555, + 321 + ], + "type": "text", + "content": " improvement in server-side computation compared to the fastest method. See Appendix C for the experimental setup." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 330, + 395, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 330, + 395, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 395, + 342 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 349, + 369, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 369, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 369, + 361 + ], + "type": "text", + "content": "6.1. Setups" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 366, + 555, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 555, + 521 + ], + "type": "text", + "content": "Datasets and Models. Our experiments are conducted on three datasets—FEMNIST [6], CIFAR-10, and CIFAR-100 [21]—using a Multilayer Perceptron (MLP), Convolutional Neural Network (CNN), and SqueezeNet, respectively. For FEMNIST, the dataset is naturally non-IID and consists of 199 unique clients. For CIFAR-10 and CIFAR-100, we set the number of clients, " + }, + { + "bbox": [ + 313, + 366, + 555, + 521 + ], + "type": "inline_equation", + "content": "n = 100" + }, + { + "bbox": [ + 313, + 366, + 555, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 366, + 555, + 521 + ], + "type": "inline_equation", + "content": "n = 50" + }, + { + "bbox": [ + 313, + 366, + 555, + 521 + ], + "type": "text", + "content": ", and generate non-IID data distributions across clients using a Dirichlet distribution by following [24, 26, 36]. The concentration parameter " + }, + { + "bbox": [ + 313, + 366, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 366, + 555, + 521 + ], + "type": "text", + "content": " of the Dirichlet distribution controls the degree of data heterogeneity, with lower values indicating higher heterogeneity. For each client, " + }, + { + "bbox": [ + 313, + 366, + 555, + 521 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 313, + 366, + 555, + 521 + ], + "type": "text", + "content": " of their data is allocated for training, while the remaining " + }, + { + "bbox": [ + 313, + 366, + 555, + 521 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 366, + 555, + 521 + ], + "type": "text", + "content": " is reserved for testing." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 522, + 554, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 554, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 554, + 570 + ], + "type": "text", + "content": "Evaluated attacks. We select the classic targeted attack, LabelFlipping (LP) [10]; two powerful adaptive attacks, Min-Max [33] and IPM [40]; and two privacy inference attacks, iDLG [42] and Inverting-Gradients [13]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 570, + 554, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 640 + ], + "type": "text", + "content": "Compared Methods. We consider the FL baseline FedAvg [30]; three prominent RARs: Krum [4], DnC [33], and RFA [32]; three robust FL methods: GAS [26], FLTrust [8], and ClippedCluster [22]; and three privacy-preserving robust FL methods: ACORN [2], ShieldFL [29], and PEFL [25]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 642, + 554, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 700 + ], + "type": "text", + "content": "Evaluation. We use average top-1 accuracy of each client to evaluate the performance of global models. We run each experiment for five times and report the mean and standard deviation of the highest accuracy during the training process." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 325, + 701, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 554, + 714 + ], + "type": "text", + "content": "All hyperparameter settings for the compared methods" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3921" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 553, + 236 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 236 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 236 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 236 + ], + "type": "table", + "html": "
DatasetAttackBaselineRARRobust FLPrivacy-preserving Robust FLOur
FedAvgKrumDnCRFAGASFLTrustClipped ClusteringACORNShieldFLPEFLFLSeg (Krum)FLSeg (DnC)
CIFAR-10-52.16±0.9753.86±0.9053.32±0.1353.52±0.4851.16±0.2853.28±0.7053.44±1.5949.08±1.9752.61±0.2746.31±0.8860.45±0.8263.99±0.53
LF47.46±0.6947.55±1.9246.50±1.5555.20±1.0647.96±1.9250.61±0.6150.85±1.4347.08±0.3348.37±0.8544.50±1.2359.62±0.7663.68±0.86
IPM49.03±0.1846.44±0.3351.82±0.4535.07±1.6243.55±1.6050.52±1.9550.93±1.5545.36±2.2149.53±1.8641.07±3.0259.43±1.0661.28±0.91
Min-Max35.89±1.5745.46±1.4950.62±1.4349.61±0.6049.69±1.7148.25±0.7049.36±1.9748.21±2.0338.51±3.0912.36±0.4860.82±0.5764.59±1.02
FEMNIST-70.17±1.8169.60±0.2369.95±0.8868.76±0.9369.95±0.4737.06±1.4669.31±1.1273.14±0.8173.33±0.9059.78±1.7177.12±0.6777.88±0.31
LF68.08±1.5870.37±0.4469.27±1.5668.70±0.7668.29±1.3133.88±0.6269.90±0.5169.78±0.3770.82±0.5452.68±1.9873.44±0.7776.94±0.45
IPM66.69±1.0659.88±1.9067.99±1.8455.02±0.8555.39±0.2838.81±1.0967.07±1.0168.44±2.0770.34±1.0350.37±1.6373.94±1.1574.95±0.87
Min-Max67.46±1.5265.17±1.8568.62±1.3865.38±1.3265.85±1.1236.70±0.2768.05±1.5171.29±1.0870.19±2.1121.32±0.2276.13±1.6277.74±0.65
CIFAR-100-35.92±0.6734.80±0.5434.48±0.3834.34±0.2425.64±0.9531.53±0.4233.32±0.1335.36±0.5634.70±0.752.86±0.4754.45±0.8153.55±0.59
LF30.83±0.3334.37±0.4933.46±0.2729.28±0.5826.98±0.3731.11±0.4432.99±0.4229.85±0.9132.53±0.613.21±0.9642.32±0.6543.79±0.49
IPM25.54±1.2119.29±0.7331.38±0.5315.61±0.6425.89±0.4529.43±0.2632.44±0.5925.55±0.4823.62±0.333.24±1.2841.56±0.7545.13±1.09
Min-Max22.31±1.0216.38±0.2934.23±0.6219.46±0.9327.21±0.3830.16±1.2534.73±0.6721.04±0.4334.39±0.473.20±0.4146.05±0.3451.11±0.21
", + "image_path": "839c2736f2e641be349dda54084e3b4f964785139e0c9c7f5a6d1e3b30c1ed00.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 189, + 244, + 419, + 255 + ], + "lines": [ + { + "bbox": [ + 189, + 244, + 419, + 255 + ], + "spans": [ + { + "bbox": [ + 189, + 244, + 419, + 255 + ], + "type": "text", + "content": "Table 2. Top-1 Accuracy(%) under Different Byzantine Attack." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 277, + 295, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 295, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 295, + 301 + ], + "type": "text", + "content": "and attacks, along with detailed experimental configurations, are provided in Appendix D." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 308, + 132, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 308, + 132, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 132, + 319 + ], + "type": "text", + "content": "6.2. Robustness" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 326, + 296, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 326, + 296, + 410 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 296, + 410 + ], + "type": "text", + "content": "Table 2 shows the accuracy under different Byzantine attackers, with data heterogeneity set at a Dirichlet concentration parameter of " + }, + { + "bbox": [ + 55, + 326, + 296, + 410 + ], + "type": "inline_equation", + "content": "\\alpha = 0.5" + }, + { + "bbox": [ + 55, + 326, + 296, + 410 + ], + "type": "text", + "content": " on CIFAR-10 and CIFAR-100, and attackers comprising " + }, + { + "bbox": [ + 55, + 326, + 296, + 410 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 55, + 326, + 296, + 410 + ], + "type": "text", + "content": " of the total number of clients. FLSeg can utilize different RAR for Segment Aggregation; we denote the RAR used by adding it in parentheses after the method name, for example, FLSeg(DnC)." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 58, + 419, + 296, + 502 + ], + "blocks": [ + { + "bbox": [ + 58, + 419, + 296, + 502 + ], + "lines": [ + { + "bbox": [ + 58, + 419, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 58, + 419, + 296, + 502 + ], + "type": "image", + "image_path": "1bbd762c6f67a253ac894703353fcf710ba92cf2a0cb3c207053f8954c63f8ec.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 510, + 295, + 532 + ], + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 532 + ], + "type": "text", + "content": "Figure 3. Top-1 Accuracy(%) under Different Ratio of Attacker on CIFAR-10." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "content": "From the empirical results, we observe that FLSeg maintains stable performance under various Byzantine attacks. Additionally, we find that FLSeg's defensive effectiveness is significantly influenced by the choice of RAR. For instance, using DnC, a singular value decomposition (SVD)-based spectral method can achieve higher accuracy than Euclidean distance-based methods like Krum. We also found that the combination of RAR with FLSeg significantly outperforms the original RAR method, especially in complex models, indicating that FLSeg effectively mitigates the curse of dimensionality. For example, Krum nearly fails under Min-Max attacks on CIFAR-100, achieving an accuracy of only " + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "inline_equation", + "content": "16.38\\%" + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "content": ", but when combined with FLSeg, it achieves a much improved accuracy of " + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "inline_equation", + "content": "46.05\\%" + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 277, + 555, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 555, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 555, + 421 + ], + "type": "text", + "content": "Figure 3 shows the accuracy under different numbers of Min-Max attackers, with FLSeg demonstrating greater stability against varying numbers of attackers. Additionally, it is evident that other privacy-preserving robust FL methods experience a notable performance drop as the number of attackers increases. This decline occurs because the metrics these methods use to filter malicious gradients are constrained by their privacy-preserving techniques, making them vulnerable to strong adaptive attacks. In contrast, FLSeg mitigates the curse of dimensionality through Segment Aggregation and enhances resistance to malicious gradients by integrating diverse RAR." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 430, + 432, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 432, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 432, + 443 + ], + "type": "text", + "content": "6.3. Heterogeneous Data" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 449, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 533 + ], + "type": "text", + "content": "Figure 4 shows the accuracy under different levels of data heterogeneity on CIFAR-10. We fix the ratio of attackers at " + }, + { + "bbox": [ + 313, + 449, + 555, + 533 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 449, + 555, + 533 + ], + "type": "text", + "content": " using Min-Max attackers and control the level of data heterogeneity by adjusting the Dirichlet " + }, + { + "bbox": [ + 313, + 449, + 555, + 533 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 449, + 555, + 533 + ], + "type": "text", + "content": ". An " + }, + { + "bbox": [ + 313, + 449, + 555, + 533 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 449, + 555, + 533 + ], + "type": "text", + "content": " of 0.1 represents an extreme case where each client may only possess one or two labels out of ten possible categories, while at " + }, + { + "bbox": [ + 313, + 449, + 555, + 533 + ], + "type": "inline_equation", + "content": "\\alpha = 10" + }, + { + "bbox": [ + 313, + 449, + 555, + 533 + ], + "type": "text", + "content": ", the data distribution is nearly IID." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 544, + 555, + 632 + ], + "blocks": [ + { + "bbox": [ + 316, + 544, + 555, + 632 + ], + "lines": [ + { + "bbox": [ + 316, + 544, + 555, + 632 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 555, + 632 + ], + "type": "image", + "image_path": "3ebce31097df6781b87d72afdeda67326be4371ae087ff471861f6ef332455fa.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 640, + 555, + 663 + ], + "lines": [ + { + "bbox": [ + 313, + 640, + 555, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 640, + 555, + 663 + ], + "type": "text", + "content": "Figure 4. Top-1 Accuracy(%) under Different Levels of Data Heterogeneity on CIFAR-10." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "We observe that FLSeg achieves significantly better performance than other methods when " + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\alpha = 0.1" + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": ". This is because other methods were not designed to account for the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3922" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": "potential impact of data heterogeneity, whereas FLSeg enables clients to better adapt to their local datasets under non-IID conditions by learning personalized models, thereby achieving improved performance. However, as the data tends to be balanced (" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\alpha = 10" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": "), the gains brought by personalization will gradually decrease, but the performance of FLSeg is still better than that of other baselines. This is because applying RAR to model segments effectively alleviates the curse of dimensionality." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 190, + 114, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 190, + 114, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 114, + 201 + ], + "type": "text", + "content": "6.4. Privacy" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 207, + 296, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 207, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 207, + 296, + 327 + ], + "type": "text", + "content": "To verify FLSeg's privacy protection capability against a semi-honest server, we present visual results of attacks using DLG and Inverting-Gradient, with and without our defense, on FEMNIST and CIFAR-100 in Fig. 5. Without privacy protection (Origin), the server can easily reconstruct training data. In contrast, when attacking the mixed model after Segment Exchange, only meaningless information can be extracted. Furthermore, in Sec. 5.1, we prove that the probability of the server reconstructing the original client model is negligible, thereby ensuring client privacy." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 59, + 339, + 293, + 450 + ], + "blocks": [ + { + "bbox": [ + 59, + 339, + 293, + 450 + ], + "lines": [ + { + "bbox": [ + 59, + 339, + 293, + 450 + ], + "spans": [ + { + "bbox": [ + 59, + 339, + 293, + 450 + ], + "type": "image", + "image_path": "f2274569cecfff567e9a04e0cf92367ac1d434d41653225ad7799ed2ad55ebcf.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 457, + 295, + 480 + ], + "lines": [ + { + "bbox": [ + 55, + 457, + 295, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 457, + 295, + 480 + ], + "type": "text", + "content": "Figure 5. Attack results of iDLG and Inverting-Gradients on CIFAR-10 and FEMNIST, respectively." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 502, + 194, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 194, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 194, + 514 + ], + "type": "text", + "content": "6.5.Impact of Segment Size" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": "To investigate the impact of segment size on privacy protection and model performance, we vary the segment size " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "|\\gamma|" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": " across " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\{10^2, 10^3, 10^4, 10^5, 10^6, 2904970(\\ell)\\}" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": " under MinMax and Inverting-Gradient attacks on the heterogeneous CIFAR-10 dataset. We use the same settings as in the robustness experiments to evaluate the model's top-1 accuracy and perform an inference attack on the model updates of a single client from one training round. The experimental results are presented in Tab. 3." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 58, + 639, + 294, + 692 + ], + "blocks": [ + { + "bbox": [ + 58, + 639, + 294, + 692 + ], + "lines": [ + { + "bbox": [ + 58, + 639, + 294, + 692 + ], + "spans": [ + { + "bbox": [ + 58, + 639, + 294, + 692 + ], + "type": "table", + "html": "
|γ|102103104105106l
Accuracy63.93%64.33%65.72%60.93%59.67%57.75%
Inference
", + "image_path": "01cf9f1bfc4018be3cd8a51dfe1b6e210076f8d54aaaea8445c88f1a949ea1cb.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 699, + 293, + 711 + ], + "lines": [ + { + "bbox": [ + 56, + 699, + 293, + 711 + ], + "spans": [ + { + "bbox": [ + 56, + 699, + 293, + 711 + ], + "type": "text", + "content": "Table 3. Accuracy and Privacy of FLSeg(Dnc) under different " + }, + { + "bbox": [ + 56, + 699, + 293, + 711 + ], + "type": "inline_equation", + "content": "|\\gamma |" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "text", + "content": "We observe that the accuracy of " + }, + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "inline_equation", + "content": "\\mathrm{FLSeg(DnC)}" + }, + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "text", + "content": " initially increases and then slightly decreases. We hypothesize that this occurs because benchmarks computed from smaller segments lack representativeness, while larger segments fail to mitigate the curse of dimensionality. Moreover, as the segment size increases, the reconstructed attack results reveal more information about the original data, since larger segments concentrate more original gradient information in the mixed model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 186, + 408, + 199 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 186, + 408, + 199 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 408, + 199 + ], + "type": "text", + "content": "6.6. Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 204, + 554, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 554, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 554, + 239 + ], + "type": "text", + "content": "We finally study how each component contributes to the overall performance of FLSeg. Specifically, we measure the following scenarios:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 240, + 539, + 287 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 314, + 240, + 449, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 240, + 449, + 251 + ], + "spans": [ + { + "bbox": [ + 314, + 240, + 449, + 251 + ], + "type": "text", + "content": "FedAvg. The baseline method." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 251, + 539, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 251, + 539, + 263 + ], + "spans": [ + { + "bbox": [ + 314, + 251, + 539, + 263 + ], + "type": "text", + "content": "- PT-only. Adds local personalized training to FedAvg." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 264, + 522, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 264, + 522, + 275 + ], + "spans": [ + { + "bbox": [ + 314, + 264, + 522, + 275 + ], + "type": "text", + "content": "- SA-only. Adds Segment Aggregation to FedAvg." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 276, + 477, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 276, + 477, + 287 + ], + "spans": [ + { + "bbox": [ + 314, + 276, + 477, + 287 + ], + "type": "text", + "content": "- FLSeg. The complete FLSeg method." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 288, + 554, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 554, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 554, + 324 + ], + "type": "text", + "content": "Note that we did not include Segment Exchange as its purpose is to protect privacy without compromising the correctness of subsequent aggregation." + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 315, + 332, + 555, + 401 + ], + "blocks": [ + { + "bbox": [ + 315, + 332, + 555, + 401 + ], + "lines": [ + { + "bbox": [ + 315, + 332, + 555, + 401 + ], + "spans": [ + { + "bbox": [ + 315, + 332, + 555, + 401 + ], + "type": "image", + "image_path": "10619e13be0213b5a4578037d6316a5769ad9b4d581a47aabd718a80e2b38037.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 408, + 480, + 419 + ], + "lines": [ + { + "bbox": [ + 388, + 408, + 480, + 419 + ], + "spans": [ + { + "bbox": [ + 388, + 408, + 480, + 419 + ], + "type": "text", + "content": "Figure 6. Ablation study." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 432, + 555, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 555, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 555, + 528 + ], + "type": "text", + "content": "We conducted ablation experiments under four scenarios combining with/without " + }, + { + "bbox": [ + 313, + 432, + 555, + 528 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 432, + 555, + 528 + ], + "type": "text", + "content": " Min-Max attackers and IID/non-IID " + }, + { + "bbox": [ + 313, + 432, + 555, + 528 + ], + "type": "inline_equation", + "content": "(\\alpha = 0.5)" + }, + { + "bbox": [ + 313, + 432, + 555, + 528 + ], + "type": "text", + "content": " CIFAR-10 dataset, with the results shown in Fig. 6. We observe that, in all scenarios, Segment Aggregation effectively resists Byzantine attacks. Meanwhile, the incorporation of personalized training does not degrade SA-only performance under IID data, and improves the model performance under non-IID data." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 538, + 388, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 388, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 388, + 550 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 557, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 554, + 665 + ], + "type": "text", + "content": "In this work, we identified three key challenges in privacy-preserving robust FL: high computational costs, limited flexibility in RAR, and data heterogeneity. To overcome these, we proposed FLSeg, which uses Segment Exchange and Segment Aggregation to ensure privacy without limiting aggregation choices and employs a regularization term to enhance adaptation to heterogeneous data. Theoretical and experimental results validate FLSeg's security, robustness, and resilience to data heterogeneity." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 666, + 554, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 554, + 702 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 554, + 702 + ], + "type": "text", + "content": "In the future, we will explore stronger threat models, such as malicious server, or extend the privacy-preserving segment exchange framework to decentralized settings." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3923" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 190 + ], + "type": "text", + "content": "This research was supported in part by the Major Research Plan of Hubei Province (2023BAA027), the Changsha Key Science and Technology Project under the \"Unveiling and Commanding\" Initiative (kq2503009), the Science, Technology and Innovation Commission of Shenzhen Municipality (GJHZ20240218114659027), the Key Research and Development Plan of Hubei Province (2024BAB049), and the Fundamental Research Funds for the Central Universities (YCJJ20252336)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 210, + 115, + 222 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 210, + 115, + 222 + ], + "spans": [ + { + "bbox": [ + 56, + 210, + 115, + 222 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 229, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 229, + 296, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 229, + 296, + 273 + ], + "spans": [ + { + "bbox": [ + 61, + 229, + 296, + 273 + ], + "type": "text", + "content": "[1] Abbas Acar, Hidayet Aksu, A Selcuk Uluagac, and Mauro Conti. A survey on homomorphic encryption schemes: Theory and implementation. ACM Computing Surveys (Csur), 51(4):1-35, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 274, + 296, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 274, + 296, + 339 + ], + "spans": [ + { + "bbox": [ + 61, + 274, + 296, + 339 + ], + "type": "text", + "content": "[2] James Bell, Adrià Gascon, Tancrède Lepoint, Baiyu Li, Sarah Meiklejohn, Mariana Raykova, and Cathie Yun. ACORN: input validation for secure aggregation. In 32nd USENIX Security Symposium, USENIX Security 2023, Anaheim, CA, USA, August 9-11, 2023, pages 4805-4822. USENIX Association, 2023. 1, 2, 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 340, + 296, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 340, + 296, + 405 + ], + "spans": [ + { + "bbox": [ + 62, + 340, + 296, + 405 + ], + "type": "text", + "content": "[3] James Henry Bell, Kallista A. Bonawitz, Adrià Gasón, Tancrède Lepoint, and Mariana Raykova. Secure single-server aggregation with (poly)logarithmic overhead. In CCS '20: 2020 ACM SIGSAC Conference on Computer and Communications Security, Virtual Event, USA, November 9-13, 2020, pages 1253-1269. ACM, 2020. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 406, + 296, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 406, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 62, + 406, + 296, + 449 + ], + "type": "text", + "content": "[4] Peva Blanchard, El Mahdi El Mhamdi, Rachid Guerraoui, and Julien Stainer. Machine learning with adversaries: Byzantine tolerant gradient descent. Advances in neural information processing systems, 30, 2017. 1, 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 449, + 296, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 449, + 296, + 515 + ], + "spans": [ + { + "bbox": [ + 62, + 449, + 296, + 515 + ], + "type": "text", + "content": "[5] Keith Bonawitz, Vladimir Ivanov, Ben Kreuter, Antonio Marcedone, H Brendan McMahan, Sarvar Patel, Daniel Ramage, Aaron Segal, and Karn Seth. Practical secure aggregation for privacy-preserving machine learning. In proceedings of the 2017 ACM SIGSAC Conference on Computer and Communications Security, pages 1175-1191, 2017. 1, 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 516, + 296, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 516, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 62, + 516, + 296, + 559 + ], + "type": "text", + "content": "[6] Sebastian Caldas, Sai Meher Karthik Duddu, Peter Wu, Tian Li, Jakub Konečný, H Brendan McMahan, Virginia Smith, and Ameet Talwalkar. Leaf: A benchmark for federated settings. arXiv preprint arXiv:1812.01097, 2018. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 560, + 296, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 560, + 296, + 592 + ], + "spans": [ + { + "bbox": [ + 62, + 560, + 296, + 592 + ], + "type": "text", + "content": "[7] Ran Canetti. Universally composable security: A new paradigm for cryptographic protocols. In FOCS, pages 136-145. IEEE Computer Society, 2001. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 593, + 296, + 625 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 593, + 296, + 625 + ], + "spans": [ + { + "bbox": [ + 62, + 593, + 296, + 625 + ], + "type": "text", + "content": "[8] Xiaoyu Cao, Minghong Fang, Jia Liu, and Neil Zhenqiang Gong. Fltrust: Byzantine-robust federated learning via trust bootstrapping. In NDSS. The Internet Society, 2021. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 626, + 296, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 626, + 296, + 669 + ], + "spans": [ + { + "bbox": [ + 62, + 626, + 296, + 669 + ], + "type": "text", + "content": "[9] Alireza Fallah, Aryan Mokhtari, and Asuman E. Ozdaglar. Personalized federated learning with theoretical guarantees: A model-agnostic meta-learning approach. In NeurIPS, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "text", + "content": "[10] Minghong Fang, Xiaoyu Cao, Jinyuan Jia, and Neil Gong. Local model poisoning attacks to {Byzantine-Robust} federated learning. In 29th USENIX security symposium (USENIX Security 20), pages 1605-1622, 2020. 2, 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 553, + 715 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 118 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 118 + ], + "type": "text", + "content": "[11] Minghong Fang, Xiaoyu Cao, Jinyuan Jia, and Neil Zhenqiang Gong. Local model poisoning attacks to byzantine-robust federated learning. In USENIX Security Symposium, pages 1605–1622. USENIX Association, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 118, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 150 + ], + "type": "text", + "content": "[12] Marc Fischlin and Arno Mittelbach. An overview of the hybrid argument. IACR Cryptol. ePrint Arch., page 88, 2021. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "type": "text", + "content": "[13] Jonas Geiping, Hartmut Bauermeister, Hannah Droge, and Michael Moeller. Inverting gradients-how easy is it to break privacy in federated learning? Advances in neural information processing systems, 33:16937-16947, 2020. 1, 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 198, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 240 + ], + "type": "text", + "content": "[14] Oded Goldreich, Silvio Micali, and Avi Wigderson. How to play any mental game or A completeness theorem for protocols with honest majority. In STOC, pages 218-229. ACM, 1987. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "type": "text", + "content": "[15] Rachid Guerraoui, Sébastien Rouault, et al. The hidden vulnerability of distributed learning in byzantium. In International Conference on Machine Learning, pages 3521-3530. PMLR, 2018. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 287, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 331 + ], + "type": "text", + "content": "[16] Anbu Huang, Yang Liu, Tianjian Chen, Yongkai Zhou, Quan Sun, Hongfeng Chai, and Qiang Yang. Starfl: Hybrid federated learning architecture for smart urban computing. ACM Trans. Intell. Syst. Technol., 12(4):43:1-43:23, 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 332, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 376 + ], + "type": "text", + "content": "[17] Najeeb Moharram Jebreel, Josep Domingo-Ferrer, Alberto Blanco-Justicia, and David Sánchez. Enhanced security and privacy via fragmented federated learning. IEEE Transactions on Neural Networks and Learning Systems, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 377, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 377, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 317, + 377, + 553, + 442 + ], + "type": "text", + "content": "[18] Peter Kairouz, H Brendan McMahan, Brendan Avent, Aurélien Bellet, Mehdi Bennis, Arjun Nitin Bhagoji, Kallista Bonawitz, Zachary Charles, Graham Cormode, Rachel Cummings, et al. Advances and open problems in federated learning. Foundations and trends® in machine learning, 14(1-2): 1-210, 2021. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 444, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 444, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 444, + 553, + 510 + ], + "type": "text", + "content": "[19] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank Reddi, Sebastian Stich, and Ananda Theertha Suresh. SCAFFOLD: Stochastic controlled averaging for federated learning. In Proceedings of the 37th International Conference on Machine Learning, pages 5132-5143. PMLR, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 510, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 510, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 553, + 566 + ], + "type": "text", + "content": "[20] Sai Praneeth Karimireddy, Lie He, and Martin Jaggi. Byzantine-robust learning on heterogeneous datasets via bucketing. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 567, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 567, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 315, + 567, + 553, + 590 + ], + "type": "text", + "content": "[21] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "type": "text", + "content": "[22] Shenghui Li, Edith C-H Ngai, and Thiemo Voigt. An experimental study of byzantine-robust aggregation schemes in federated learning. IEEE Transactions on Big Data, 2023. 1, 2, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 635, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 690 + ], + "type": "text", + "content": "[23] Shenghui Li, Edith Ngai, Fanghua Ye, Li Ju, Tianru Zhang, and Thiemo Voigt. Blades: A unified benchmark suite for byzantine attacks and defenses in federated learning. In 2024 IEEE/ACM Ninth International Conference on Internet-of-Things Design and Implementation (IoTDI), 2024. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 691, + 553, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 691, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 315, + 691, + 553, + 715 + ], + "type": "text", + "content": "[24] Tian Li, Shengyuan Hu, Ahmad Beirami, and Virginia Smith. Ditto: Fair and robust federated learning through per" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3924" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "sonalization. In International conference on machine learning, pages 6357-6368. PMLR, 2021. 3, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "text", + "content": "[25] Xiaoyuan Liu, Hongwei Li, Guowen Xu, Zongqi Chen, Xiaoming Huang, and Rongxing Lu. Privacy-enhanced federated learning against poisoning adversaries. IEEE Transactions on Information Forensics and Security, 16:4574-4588, 2021. 1, 2, 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 206 + ], + "type": "text", + "content": "[26] Yuchen Liu, Chen Chen, Lingjuan Lyu, Fangzhao Wu, Sai Wu, and Gang Chen. Byzantine-robust learning on heterogeneous data via gradient splitting. In International Conference on Machine Learning, pages 21404–21425. PMLR, 2023. 1, 2, 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "text", + "content": "[27] Zhi Lu, SongFeng Lu, YongQuan Cui, XueMing Tang, and JunJun Wu. Split aggregation: Lightweight privacy-preserving federated learning resistant to byzantine attacks. IEEE Transactions on Information Forensics and Security, 19:5575-5590, 2024. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 308 + ], + "type": "text", + "content": "[28] Hidde Lycklama, Lukas Burkhalter, Alexander Viand, Nicolas Kuchler, and Anwar Hithnawi. Rofl: Robustness of secure federated learning. In SP, pages 453-476. IEEE, 2023. 1, 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 309, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 363 + ], + "type": "text", + "content": "[29] Zhuoran Ma, Jianfeng Ma, Yinbin Miao, Yingjiu Li, and Robert H Deng. Shieldfl: Mitigating model poisoning attacks in privacy-preserving federated learning. IEEE Transactions on Information Forensics and Security, 17:1639-1654, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 294, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 419 + ], + "type": "text", + "content": "[30] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 421, + 294, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 421, + 294, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 421, + 294, + 475 + ], + "type": "text", + "content": "[31] Dinh C. Nguyen, Quoc-Viet Pham, Pubudu N. Pathirana, Ming Ding, Aruna Seneviratne, Zihuai Lin, Octavia A. Dobre, and Won-Joo Hwang. Federated learning for smart healthcare: A survey. ACM Comput. Surv., 55(3):60:1-60:37, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 477, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 477, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 477, + 294, + 510 + ], + "type": "text", + "content": "[32] Krishna Pillutla, Sham M Kakade, and Zaid Harchaoui. Robust aggregation for federated learning. IEEE Transactions on Signal Processing, 70:1142-1154, 2022. 1, 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 512, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 544 + ], + "type": "text", + "content": "[33] Virat Shejwalkar and Amir Houmansadr. Manipulating the byzantine: Optimizing model poisoning attacks and defenses for federated learning. In NDSS, 2021. 1, 2, 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 545, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 589 + ], + "type": "text", + "content": "[34] Virat Shejwalkar, Amir Houmansadr, Peter Kairouz, and Daniel Ramage. Back to the drawing board: A critical evaluation of poisoning attacks on production federated learning. In SP, pages 1354-1371. IEEE, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 590, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 624 + ], + "type": "text", + "content": "[35] Lichao Sun, Jianwei Qian, and Xun Chen. LDP-FL: practical private aggregation in federated learning with local differential privacy. In IJCAI, pages 1571–1578. ijcai.org, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "type": "text", + "content": "[36] Canh T Dinh, Nguyen Tran, and Josh Nguyen. Personalized federated learning with moreau envelopes. Advances in Neural Information Processing Systems, 33:21394-21405, 2020. 3, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "type": "text", + "content": "[37] Kang Wei, Jun Li, Ming Ding, Chuan Ma, Hang Su, Bo Zhang, and H. Vincent Poor. User-level privacy-preserving federated learning: Analysis and performance optimization. IEEE Trans. Mob. Comput., 21(9):3388-3401, 2022. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 319 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "text", + "content": "[38] Jiahui Wu, Weizhe Zhang, and Fucai Luo. On the security of \"lsfl: A lightweight and secure federated learning scheme for edge computing\". IEEE Trans. Inf. Forensics Secur., 19: 3481-3482, 2024. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "type": "text", + "content": "[39] Jue Xiao, Xueming Tang, and Songfeng Lu. Privacy-preserving federated class-incremental learning. IEEE Transactions on Machine Learning in Communications and Networking, 2:150-168, 2024. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "text", + "content": "[40] Cong Xie, Oluwasanmi Koyejo, and Indranil Gupta. Fall of empires: Breaking byzantine-tolerant sgd by inner product manipulation. In Uncertainty in Artificial Intelligence, pages 261-270. PMLR, 2020. 1, 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "text", + "content": "[41] Dong Yin, Yudong Chen, Ramchandran Kannan, and Peter Bartlett. Byzantine-robust distributed learning: Towards optimal statistical rates. In International Conference on Machine Learning, pages 5650–5659. Pmlr, 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 253, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 285 + ], + "type": "text", + "content": "[42] Bo Zhao, Konda Reddy Mopuri, and Hakan Bilen. idlg: Improved deep leakage from gradients. CoRR, abs/2001.02610, 2020. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 286, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 319 + ], + "type": "text", + "content": "[43] Ligeng Zhu, Zhijian Liu, and Song Han. Deep leakage from gradients. Advances in neural information processing systems, 32, 2019. 1, 2" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3925" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/FOLDER_ Accelerating Multi-Modal Large Language Models with Enhanced Performance/7a17b26d-c0b2-4285-9105-7de16ee42381_content_list.json b/2025/FOLDER_ Accelerating Multi-Modal Large Language Models with Enhanced Performance/7a17b26d-c0b2-4285-9105-7de16ee42381_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8e35b6cf169c68097ddd9feffd73fb55e93a31d0 --- /dev/null +++ b/2025/FOLDER_ Accelerating Multi-Modal Large Language Models with Enhanced Performance/7a17b26d-c0b2-4285-9105-7de16ee42381_content_list.json @@ -0,0 +1,1602 @@ +[ + { + "type": "text", + "text": "FOLDER: Accelerating Multi-modal Large Language Models with Enhanced Performance", + "text_level": 1, + "bbox": [ + 107, + 128, + 888, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haicheng Wang $^{1,2,4*}$ , Zhemeng Yu $^{1,2*}$ , Gabriele Spadaro $^{2,3}$ , Chen Ju $^{4}$ , Victor Quétu $^{2}$ , Shuai Xiao $^{4\\boxtimes}$ , Enzo Tartaglione $^{2\\boxtimes}$", + "bbox": [ + 158, + 202, + 833, + 238 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$^{1}$ Paris Elite Institute of Technology, Shanghai Jiao Tong University $^{3}$ University of Turin", + "$^{2}$ LTCI, Télecom Paris, Institut Polytechnique de Paris $^{4}$ Taobao & Tmall Group of Alibaba" + ], + "bbox": [ + 133, + 238, + 862, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "anakinSkywalker@sjtu.edu.cn, shuai.xsh@alibaba-inc.com, enzo.tartaglione@telecom-paris.fr", + "bbox": [ + 109, + 277, + 885, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 267, + 325, + 346, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, Multi-modal Large Language Models (MLLMs) have shown remarkable effectiveness for multi-modal tasks due to their ability of cross-modal understanding. However, processing long sequences of visual tokens extracted from visual backbones poses challenges for deployment in real-time applications. To address this issue, we introduce FOLDER, a simple yet effective plug-and-play module designed to reduce the length of the visual token sequence, mitigating computational and memory demands during both training and inference. Through a comprehensive analysis of the token reduction process in the vision encoder, we analyze the information loss introduced by different reduction strategies and develop FOLDER to preserve key information while removing visual redundancy. We show the effectiveness of FOLDER by integrating it into the visual backbone of various MLLMs, significantly accelerating the inference phase. Furthermore, we evaluate its utility as a training accelerator or even performance booster for MLLMs. FOLDER achieves comparable or even better performance than the original models, while dramatically reducing complexity by removing up to $70\\%$ of visual tokens. Our code is available at FOLDER Repository.", + "bbox": [ + 89, + 358, + 483, + 705 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 726, + 220, + 742 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multi-modal Large Language Models (MLLMs) excel in tasks like image captioning [12] and visual question answering [33] by learning joint visual-text representations. State-of-the-art models such as GPT-4V [1] and Claude 3 [2] show strong multimodal understanding ability. However, these models face concrete challenges in terms of computational efficiency, particularly when processing visual inputs. Modern MLLMs typically generate long sequences of im", + "bbox": [ + 89, + 752, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b8d374ca51938da28fb9a98ee201711701d9c1ee0f9c56bfa8c0278ab60e80dc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 325, + 906, + 507 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/309f1efceee5e179e46f7b793e4da6cf6bf2a14af7c63e298fa05ad387252967.jpg", + "image_caption": [ + "Figure 1. FOLDER as Accelerator & Booster. As a plug-and-play module, FOLDER can be used in both training and inference, with considerable acceleration and even a performance boost." + ], + "image_footnote": [], + "bbox": [ + 517, + 510, + 903, + 636 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "age tokens through their visual backbones. In some application scenarios, this problem becomes even more evident: recent advancements in high-resolution-supporting MLLMs [11, 71] might involve over 2000 visual tokens. Similarly, challenges also occur in multi-visual-expert architectures [32, 66, 70] that leverage multiple visual encoders. These issues are even more severe in video understanding [30, 40, 45] due to the extra temporal complexity. The resulting token explosion hinders real-time deployment because of the quadratic cost of the attention mechanism. Several studies have been conducted to address this issue. Existing approaches can be roughly divided into two", + "bbox": [ + 511, + 717, + 906, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*: Equal contribution, ${}^{20}$ : Corresponding author.", + "bbox": [ + 112, + 886, + 380, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23614", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "categories. The first category focuses on training phase solutions, such as Q-Former [39], pixel-shuffle [10] and pooling-based techniques [7]. However, these methods often suffer from performance degradation and exhibit limited scalability, due to their custom design for specific architectures. The second category aims to design plug-and-play token reduction modules for the inference phase [9, 27, 31, 65, 75, 84]. Specifically, [9, 75, 84] reduce visual tokens inside LLM. Though effective, these methods are often hard to implement, require adaptation for existing LLM acceleration modules [13], and prove unsuitable for multi-round conversations, thus greatly limiting their usage. Meanwhile, [27, 31, 65] identify the inherent redundancy in visual tokens, but fail to quantify the information loss during the token reduction process. Thus their reduction strategies are often arbitrary (uniform reduction for each layer), leading to sub-optimal outcomes.", + "bbox": [ + 89, + 90, + 480, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To find a reduction strategy that can be universally applicable and best preserve the performance, we propose a preliminary investigation into the vision backbone, to answer a fundamental question: Where does information loss originate? To model it, we identify three key factors that impact information loss during token removal: Reduction Impact, Propagation Effect, and Aggregation Method. Through a comprehensive empirical study of these factors, we devise a strategy that enhances the preservation of information while significantly reducing redundancy in the vision backbone. FOLDER is consequently designed to implement this strategy with minimal computational overhead.", + "bbox": [ + 89, + 349, + 482, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our plug-and-play solution can be easily integrated into the visual backbones of various MLLMs, effectively reducing the number of visual tokens while incurring negligible information loss (Fig. 1). We first incorporate FOLDER during MLLMs inference, reducing over $60\\%$ of visual tokens and achieving comparable, or even gains on performance across multiple tasks. Furthermore, when integrated into MLLMs pre-training, our module can be used as a training accelerator, as well as an effective regularization term, leading to remarkable performance improvements across all benchmarks, with a reduction ratio up to $70\\%$ . Our method is on-the-field proven to be a flexible, nearly lossless, and highly efficient token reduction strategy, offering a dual-purpose solution to the visual token reduction challenge. Our contributions can be summarized as follows.", + "bbox": [ + 89, + 535, + 482, + 761 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We conduct an in-depth analysis for information decay during token reduction in the vision backbone, identifying and quantifying key factors involved (Sec. 3.1, Sec. 3.2 and Sec. 3.3). These insights offer a clear understanding of how token reduction impacts the information flow.", + "- Leveraging the above analysis, we develop a simple yet effective plug-and-play visual token reduction module for MLLMs (Sec. 3.4). This novel token reduction strategy aggressively reduces the number of tokens only in the" + ], + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "last blocks of the visual encoder. On various models and benchmarks, FOLDER speeds up the inference $1.7 - 2.4 \\times$ with negligible loss or even improvement (Sec. 4.2.1).", + "bbox": [ + 526, + 90, + 903, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We show that FOLDER can also be used in MLLM training (Sec. 4.2.2), to speed-up while boosting performance, serving as a dual-solution for both training and inference.", + "bbox": [ + 513, + 136, + 903, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 195, + 663, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-modal Large Language Models. MLLMs excel at understanding multi-modal information [1, 3, 7, 39, 41, 45, 48, 67-69, 71, 83], with common architectures comprising i.) a Visual Backbone, ii.) a Vision-Language Connector, and iii.) a Pre-trained LLM. However, integrating visual data significantly increases computational cost due to the large number of visual tokens. For example, LLaVA1.5 [50] processes 576 tokens for a $336 \\times 336$ image, while VideoLLaVA [45] exceeds 2000 tokens with just 8 frames at $224 \\times 224$ . Recent efforts to enhance MLLMs with multiple vision towers [32, 66, 70] further exacerbate this issue. MLLMs Acceleration. To accelerate large models, many methods focus on system-level optimizations, such as FlashAttention [14], vLLM [13], and RingAttention [49]. Model-centric techniques like knowledge distillation [20, 72], quantization [21, 74], and pruning [60, 64, 72] also reduce size and cost. However, these approaches overlook redundant data, which unnecessarily prolongs sequences.", + "bbox": [ + 511, + 220, + 903, + 492 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Token Reduction in MLLMs. Some preliminary studies are conducted on token reduction for Vision Transformers [6, 34, 44, 85]. In the context of MLLMs, several token reduction methods, including Q-Former [39], resampler [11], pooling [7] and some other methods [29, 42, 54, 78, 79] are proposed to reduce visual tokens during the training process. However, these approaches often suffer from performance degradation and a lack of scalability.", + "bbox": [ + 511, + 493, + 903, + 613 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Meanwhile, some studies try to handle the token reduction problem during inference. In particular, ToMe [6] proposes a training-free token merging strategy for unimodal image classification. Based on that, Turbo [31] proposes an improved merging strategy considering both mutual redundancy and semantic value, which serves as a plug-in module in MLLMs. Recently, FastV [9] recognizes the visual attention sparsity in LLM, thus pruning visual tokens inside the LLM based on the task-oriented attention importance. Though effective, these methods have clear drawbacks. For ToMe/Turbo-like methods [6, 27, 31, 63, 65, 77] that reduce tokens in the vision encoder, most of them adopt uniform progressive merging across layers, which is arbitrary and inflexible. While FastV-like methods [9, 51, 75, 84] that cut tokens inside LLM layers mainly suffer from 2 problems: i.) They often require interactions with instruction tokens, making it unsuitable for multi-round/long-term visual QA. ii.) Since these approaches are inserted in LLM and often require specific hyperparameters, it is difficult to adapt", + "bbox": [ + 511, + 613, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23615", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f2312d72b13a11838068d5e60f83ba4379c9e0cae5892cb0c0c600a59cd120f8.jpg", + "image_caption": [ + "Figure 2. Minimum Number of Tokens $k$ to Preserve Energy $t_E$ Across Blocks. We evaluate on 3 different $t_E$ for every block." + ], + "image_footnote": [], + "bbox": [ + 106, + 95, + 473, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "for complex LLM modules [14] and hard to deploy. Besides, few approaches have been verified to be effective in both training and training-free inference ways. To address the above issues, we seek to find an effective, information-guided, and user-friendly token reduction method that can be easily applied in both training and inference.", + "bbox": [ + 89, + 334, + 483, + 428 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 443, + 181, + 458 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we establish FOLDER by first answering three key questions for token reduction in vision backbone: i.) how many tokens to reduce (Sec. 3.1), ii.) in which block (Sec. 3.2) and iii.) through which aggregation method (Sec. 3.3). Then, we present our method in Sec. 3.4.", + "bbox": [ + 89, + 470, + 483, + 545 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Empirical Setup. To answer these three crucial questions, we define a common empirical setup. Specifically, we use a pre-trained 12-block ViT-B [16] model and conduct the empirical experiments on ImageNet-1k [15]. Results on more diverse setups are provided in the appendix.", + "bbox": [ + 89, + 546, + 483, + 623 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Token Reduction Impact", + "text_level": 1, + "bbox": [ + 89, + 635, + 316, + 651 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To explore the link between token reduction and information loss, we use Singular Value Decomposition (SVD) to track the energy of the reduced token sequence. More precisely, given a token sequence $\\mathbf{X} \\in \\mathbb{R}^{n \\times d}$ (with $n$ being the number of tokens and $d$ their size), we can decompose it by applying SVD on $\\mathbf{X}^T$ :", + "bbox": [ + 89, + 659, + 483, + 750 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} ^ {T} = \\mathbf {U} \\boldsymbol {\\Sigma} \\mathbf {V} ^ {T}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 230, + 762, + 482, + 780 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{U} \\in \\mathbb{R}^{d \\times d}$ and $\\mathbf{V} \\in \\mathbb{R}^{n \\times n}$ are two orthogonal matrices and $\\pmb{\\Sigma} \\in \\mathbb{R}^{d \\times n}$ is a diagonal matrix, with the singular values $\\sigma_{i} = \\pmb{\\Sigma}_{ii}$ representing the variances (energy) in the new compositional directions. In this way, SVD yields the optimal energy concentration that can be ordered by their contribution to the total variance (energy) [18]. Based on this, we can estimate how much variance we preserve by", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/62e69817cfadc5278ff9eb4633366b3fa4035d059732be2bb28f9968c9028b17.jpg", + "image_caption": [ + "Figure 3. EMD Distance Between Reduced and Original Output Distributions under 3 Reduction Ratios. We compare the EMD distance by exerting token reduction on different blocks." + ], + "image_footnote": [], + "bbox": [ + 531, + 95, + 893, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "only taking the largest $k$ singular values, as:", + "bbox": [ + 511, + 345, + 805, + 362 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nE (k) = \\frac {\\sum_ {i = 1} ^ {k} \\sigma_ {i}}{\\sum_ {i = 1} ^ {n} \\sigma_ {i}}, \\quad \\sigma_ {i} \\geq \\sigma_ {i + 1} \\quad \\forall i. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 372, + 906, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "By definition, SVD provides an upper bound on preserved energy when keeping $k$ tokens. Thus, by setting a threshold $t_E$ , we can estimate the required $k$ for each block as:", + "bbox": [ + 511, + 417, + 906, + 464 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nk = \\min \\left\\{k \\mid E (k) \\geq t _ {E} \\right\\}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 474, + 903, + 492 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Observations. With a threshold $t_E$ , $k$ varies across blocks. We monitor $k$ under this threshold and find, as shown in Fig. 2, that later blocks require far fewer tokens than earlier ones to maintain the same energy. This highlights the potential for block-adaptive token reduction, suggesting the first and last blocks as key candidates.", + "bbox": [ + 511, + 503, + 905, + 594 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Token Propagation Effect", + "text_level": 1, + "bbox": [ + 511, + 604, + 746, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While token reduction impact provides an upper bound on information preservation, it overlooks inter-block dependencies and therefore may not adequately reflect the effects on subsequent blocks. In transformer architectures, errors from earlier blocks can propagate and amplify through successive layers [5], potentially affecting final performance.", + "bbox": [ + 511, + 626, + 906, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To analyze this propagation effect, we examine how token reduction at different blocks affects the final output distribution. Thus, by comparing the original output token distribution (without reduction) and the one obtained by operating token reduction only in one block $b$ , we can show the presence of such a propagation effect.", + "bbox": [ + 511, + 715, + 905, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To measure this distortion, we consider output tokens from the original model $\\mathbf{Y} \\in \\mathbb{R}^{n \\times d}$ and the ones obtained when $n - k$ tokens are reduced in the block $b$ : $\\tilde{\\mathbf{Y}}_b \\in \\mathbb{R}^{k \\times d}$ . Then, we define the empirical distributions $P_Y$ and $P_{\\tilde{\\mathbf{Y}}_b}$ supported on $\\mathbf{Y}$ and $\\tilde{\\mathbf{Y}}_b$ respectively, and we evaluate Earth Mover's Distance (EMD) [61] between them:", + "bbox": [ + 511, + 806, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23616", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f86748aa632ab4874bd8b5a9a22d644632e297e60eeec252628922aa0b48b986.jpg", + "image_caption": [ + "Figure 4. EMD & Accuracy Comparison for Aggregation Methods. We compare three aggregation methods under different reduction ratios on both EMD and Accuracy." + ], + "image_footnote": [], + "bbox": [ + 99, + 90, + 477, + 267 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {E M D} \\left(P _ {Y}, P _ {\\tilde {Y} _ {b}}\\right) = \\min _ {\\gamma} \\langle \\gamma , \\mathbf {M} \\rangle_ {F}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 353, + 482, + 377 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{M} = (d_{ij})_{n\\times k} = (\\| y_i - \\tilde{y}_{bj}\\| _2)_{n\\times k}$ denotes the metric cost matrix, with $y_{i}$ and $\\tilde{y}_{bj}$ being points in the support of distributions $P_Y$ and $P_{\\tilde{Y}_b}$ respectively. $\\gamma = (\\gamma_{ij})_{n\\times k}$ represents the optimal transport matrix, which satisfies the non-negative constraint $\\gamma_{ij}\\geq 0,\\forall i,j$ and the marginal constraints $\\sum_{j = 1}^{k}\\gamma_{ij} = P_Y(y_i)$ and $\\sum_{i = 1}^{n}\\gamma_{ij} = P_{\\tilde{Y}_b}(\\tilde{y}_{bj})$", + "bbox": [ + 89, + 386, + 483, + 481 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Observations. Due to the propagation effect, token reduction at different blocks causes varying levels of output distortion. We monitor EMD (Eq. (4)) when reducing the same number of tokens per block and find, as shown in Fig.3, that early-block reductions lead to significantly higher EMD values. This confirms that distortion amplifies across layers, making early blocks unsuitable for token reduction, contrary to Fig.2. Conversely, with a massive $75\\%$ reduction, EMD remains low in the last layers. This suggests that token reduction is most effective at the network's end, even with a high reduction rate. Results on BLIP [38] (Tab. 7) and appendix further support this conclusion.", + "bbox": [ + 89, + 481, + 483, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Token Aggregation Method", + "text_level": 1, + "bbox": [ + 89, + 676, + 336, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Sec. 3.1, we use SVD to estimate the upper bound of information preservation. However, its computational complexity prevents direct application for token reduction. Following prior works [6, 31], we decompose the process into two stages: (i) Token Matching, grouping tokens based on a matching function, and (ii) Token Aggregation, merging grouped tokens. While matching functions have been explored [6, 31], aggregation remains understudied. Formally, given a set of tokens $\\{\\mathbf{x}_1,\\mathbf{x}_2,\\dots ,\\mathbf{x}_m\\}$ to be merged into $\\mathbf{y}$ , we define the aggregation operation as follows:", + "bbox": [ + 89, + 699, + 483, + 851 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {y} = \\sum_ {i = 1} ^ {m} \\alpha_ {i} \\mathbf {x} _ {i}, \\text {s . t .} \\sum_ {i = 1} ^ {m} \\alpha_ {i} = 1. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 864, + 482, + 904 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on Eq. (5), we formulate three different token aggregation approaches: i.) Average Merging, with $\\alpha_{i} = \\frac{1}{m},\\forall i$ ii.) Weighted Merging, with $\\alpha_{i} = \\frac{\\|x_{i}\\|_{2}}{\\sum_{j = 1}^{m}\\|x_{j}\\|_{2}},\\forall i;$ iii.) Direct Dropping, with $\\alpha_{i} = 0,\\forall i\\neq i_{max},\\alpha_{i_{max}} = 1$ where $i_{max} = \\underset {i}{\\operatorname{argmax}}\\| x_i\\| _2$ . Here, the norm $\\| \\cdot \\| _2$ is used as the importance score for tokens.", + "bbox": [ + 511, + 90, + 903, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Observations. Similarly to Sec. 3.2, we compare aggregation methods using EMD. As shown in Tab. 9 and Fig. 4, both merging methods yield lower discrepancy than Direct Dropping. Given its comparable performance to Weighted Merging with greater simplicity, we select Average Merging as our preferred method. Additionally, Fig. 4 shows that a lower EMD indeed correlates with higher accuracy. Further supporting results are in Tab. 9 and appendix.", + "bbox": [ + 511, + 194, + 906, + 315 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4.FOLDER", + "text_level": 1, + "bbox": [ + 513, + 329, + 625, + 343 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motivation. Based on the above empirical analyses, we deduce that: i.) Token reduction should take place in the last blocks, where aggressive reduction is permitted. ii.) Merging is better than dropping as an aggregation method. Therefore, our goal is to design an algorithm that can reduce as many tokens as possible in one block. Previous works [63] mainly leverage clustering for token merging, which is slow and unsuitable for parallel computing. To avoid such issues, we refer to bipartite matching [6], which is however limited by a maximum reduction ratio of $1/2$ . Thus, to aggressively reduce tokens in one block, we propose FOLDER, an advanced bipartite matching strategy that can reduce an arbitrary number of tokens.", + "bbox": [ + 511, + 353, + 906, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Strategy. As illustrated in Fig. 5, our strategy focuses on the last blocks (or the last) in the vision backbone. In some cases, the target reduction $r$ exceeds half of the total tokens $\\lfloor n / 2 \\rfloor$ in a block, a scenario we call reduction overflow. To handle this, FOLDER automatically applies an iterative \"FOLD\" operation. As detailed in Alg. 1, if reduction overflow occurs in an iteration $(r_{\\text{remain}} > \\lfloor n' / 2 \\rfloor$ in line 7), we reduce tokens by half $(r_{\\text{fold}} = \\lfloor n' / 2 \\rfloor)$ . The process begins by splitting the token input sequence into equal-sized partitions: $\\mathbb{A}$ and $\\mathbb{B}$ (line 9). Then, each token in $\\mathbb{A}$ is matched to its most similar counterpart in $\\mathbb{B}$ based on a Matching Function [6, 31] $S : (\\mathbb{R}^d, \\mathbb{R}^d) \\to \\mathbb{R}$ (line 10). A visual representation is provided in Fig. 5, where the \"Partition & Match\" operation is highlighted. These matches are then sorted based on their matching scores $(\\stackrel{s_{ij}}{\\rightarrow})$ , and only the top $r_{\\text{fold}}$ matches are preserved (line 11). At this point, by merging these matches in $\\mathbf{M}_{top}$ from $\\mathbb{A}$ to $\\mathbb{B}$ , we obtain two new sets: $\\mathbb{A}'$ and $\\mathbb{B}'$ . Namely, $\\mathbb{B}'$ updates $\\mathbb{B}$ by replacing original tokens with matched tokens aggregated using a specific Aggregation Method (Sec. 3.3). $\\mathbb{A}'$ , instead, contains the tokens in $\\mathbb{A}$ whose matches were discarded in line 11. Clearly, if the reduction overflow occurs, $r_{\\text{fold}} = \\lfloor n' / 2 \\rfloor$ and thus $\\mathbb{A}' = \\emptyset$ (\"First Fold\" of Fig. 5). The concatenation between", + "bbox": [ + 511, + 551, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23617", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7e5a3493389f1845588d7d7081c9d90be78a61c80906d5e10ae784b2b260648c.jpg", + "image_caption": [ + "Figure 5. Pipeline of FOLDER. As a plug-and-play module, FOLDER is integrated into the final blocks of the vision backbone (last two here). To deal with reduction overflow, FOLDER automatically executes another FOLD operation when the expected reduction is more than half. The last FOLD, which escapes from reduction overflow, merges tokens according to the remaining reduction numbers." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 326, + 244 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8139d0587137e2031d0ddd89611f00975e58341de69122e62abe804805f2fd7d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 326, + 88, + 607, + 247 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2522f93e5cd5212032ccd98adc056bea429ed78212b25b78108d0003c3036331.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 609, + 88, + 903, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathbb{A}^{\\prime}$ and $\\mathbb{B}^{\\prime}$ represents the input token sequence for the next \"FOLD\" iteration, in which $r_{\\mathrm{remain}}$ tokens are going to be removed. This process continues through successive folds (line 3) until the overflow is no longer encountered. In this case, $r_{\\mathrm{fold}} = r_{\\mathrm{remain}}$ (line 7), and $r_{\\mathrm{remain}}$ for the next iteration will be equal to 0 (line 8). A visual representation of this case is provided in Fig. 5 as the \"Last FOLD\" iteration. Note that in the last FOLD, to fit for the positional encoding in LLM, instead of directly concatenating $\\mathbb{A}^{\\prime}$ and $\\mathbb{B}^{\\prime}$ , we rearrange the tokens into original order. More details are provided in the appendix.", + "bbox": [ + 88, + 321, + 480, + 487 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Discussion. Though simple, the insight contained in FOLDER is non-trivial. Empirically, by analysing the information flow in the vision backbone, we draw the key conclusion that token reduction should be done only in the last blocks, which is overlooked by previous works [6, 27, 31, 65]. Structurally, by generalizing bipartite matching, we specifically design a highly efficient token reduction strategy for MLLMs architecture that allows parallel computing and minimum deployment effort. Functionally, FOLDER focuses on reducing the inherent visual redundancy in the vision backbone, which is task agnostic. Thus FOLDER can superpose on task-oriented/attention-sparsity methods like FastV [9] to enjoy further acceleration (Tab. 1). Besides, FOLDER is among the very few approaches [75] that are verified to be effective in both training and training-free ways (Tab. 1 and Tab. 3), showing its superior scalability.", + "bbox": [ + 89, + 488, + 482, + 731 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 734, + 222, + 752 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we present the results obtained by our method. By using FOLDER, we reduce the number of visual tokens to accelerate both training and inference of the Multi-modal Large Language Model (MLLM).", + "bbox": [ + 89, + 760, + 482, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Evaluation & Benchmarks", + "text_level": 1, + "bbox": [ + 89, + 832, + 333, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Models. We evaluate FOLDER as training-free plug-in on various MLLMs, including image understanding models LLaVA1.5-7B/13B [50], LLaVA-OV-7B [37],", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: FOLDER applied in one block" + ], + "code_body": "Input: Token sequence $\\mathbf{X}\\in \\mathbb{R}^{n\\times d}$ reduced num $r$ Output: Reduced token sequence $\\mathbf{X}'\\in \\mathbb{R}^{(n - r)\\times d}$ \n1 $r_{\\mathrm{remain}} = r$ . \n2 $\\mathbf{X}' = \\mathbf{X}$ .. \n3 while $r_{\\mathrm{remain}} > 0$ do \n4 $\\begin{array}{rl}{\\mathbf{\\Lambda}_{\\mathbf{X}^{\\prime}},r_{\\mathrm{remain}} = \\mathrm{FOLD}(\\mathbf{X}^{\\prime},r_{\\mathrm{remain}});} & {} \\end{array}$ \n5 \n6 Function FOLD $(\\mathbf{X}'\\in \\mathbb{R}^{n'\\times d},r_{\\mathrm{remain}})$ .. \n7 $r_{\\mathrm{fold}} = \\min (\\lfloor n' / 2\\rfloor ,r_{\\mathrm{remain}});$ \n8 $r_{\\mathrm{remain}} = r_{\\mathrm{remain}} - r_{\\mathrm{fold}};$ \n9 $\\{\\mathbb{A},\\mathbb{B}\\} \\Leftarrow \\mathrm{Split~X'}$ into 2 equal-sized partitions; \n10 M $\\Leftarrow$ Compute $\\{(a_i\\xrightarrow{s_{ij}}b_j)\\mid$ $b_{j} = \\arg \\max s_{ij} = S(a_{i},b_{j}),\\forall a_{i}\\in \\mathbb{A}\\} ;$ \n11 $\\mathbf{M}_{top}\\Leftarrow$ Sort M based on $\\stackrel {s_{ij}}{\\to}$ & keep Top $r_\\mathrm{fold}$ . \n12 $\\{\\mathbb{A}',\\mathbb{B}'\\} \\Leftarrow$ Merge A to B according to Mtop; \n13 return Rearrange(A',B'), $r_{\\mathrm{remain}}$", + "bbox": [ + 516, + 343, + 903, + 621 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Minigpt4v2 [7], VILA1.5-13B [46], WePOINTS1.5-7b [53], VITA1.5-7B [24], Idefics2-8B [35], and Flamingov2-7B [4], multi-vision-tower model MMVP [70], video understanding model Video-LLaVA [45] and point cloud understanding model PointLLMv1.2-7B [76]. For merging strategy ablation, we adopt BLIP [38] for the image captioning task. For MLLM training, we use LLaVA1.5-13B [50]. Training details are provided in the appendix.", + "bbox": [ + 511, + 655, + 906, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Benchmarks. To quantify the ability of the accelerated MLLMs, we conduct thorough experiments on a wide range of tasks with VLMEvalkit [17], including $Y/N$ tasks like MME [22], HallusionBench [26], POPE [80]; MCQ tasks like MMBench-EN [81], CCBench [81], A-OKVQA [62], ScienceQAIMG [55], SEEDBenchIMG [36], MMMU [82], MMStar [8], Video-MME [23], RealWorld-QA [25], LongVideoBench [73],", + "bbox": [ + 511, + 779, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23618", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ba75cb6fd19586931c9c7ea9fa1f6fb007972c8db89fdc4652ca4f1bef42a3d8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodReduct RatioSpeed UpMMBench ENMM StarMMEA-OK VQAHallusion BenchMM MUOCR VQASEED BenchIMGScience QAPOPEAvg
Original-7B0%1×62.832.71338.978.835.632.252.460.268.179.756.9
Pooling50%1.5×59.530.31308.677.135.331.744.859.268.084.055.6
FastV [9]1.3×63.332.41345.178.636.528.753.159.467.881.056.8
Turbo [31]1.5×60.430.91311.277.235.528.035.758.167.583.054.2
Ours1.5×62.432.11338.278.338.334.048.559.568.585.457.4
Pooling66%1.7×57.629.71308.274.232.330.738.757.667.380.753.5
FastV [9]1.5×62.532.01353.277.737.430.052.658.368.279.356.5
Pyramidddrop [75]1.6×61.331.21338.776.938.629.348.656.968.381.956.0
SparseVLM [84]1.6×61.331.11333.077.133.333.138.257.667.482.354.8
Turbo [31]1.7×60.131.51301.678.034.924.734.457.867.285.053.8
Ours1.7×61.430.31350.077.939.231.346.159.768.385.456.7
FastV [9]75%1.6×61.231.01321.676.637.532.050.857.268.177.755.8
Ours+FastV1.7×61.431.21325.578.039.132.648.859.268.682.356.8
Original-13B*0%1×66.630.91371.177.036.134.054.959.468.886.458.3
Pooling50%1.5×64.131.21316.676.235.130.054.757.669.085.757.0
FastV [9]1.3×66.431.11386.375.536.232.654.958.968.485.457.9
Turbo [31]1.5×65.030.61200.278.226.932.747.958.869.786.155.6
Ours1.5×65.431.21383.778.636.434.456.259.170.586.958.8
Pooling66%1.6×62.831.01250.673.529.532.545.654.569.782.154.4
FastV [9]1.5×66.331.21352.275.035.532.354.458.068.383.557.2
Pyramidddrop [75]1.6×65.331.11313.075.835.332.052.558.069.483.556.9
SparseVLM [84]1.6×65.231.11272.677.031.132.745.858.069.583.155.7
Turbo [31]1.6×64.731.21168.376.325.933.347.958.570.585.555.3
Ours1.6×65.831.71366.977.333.835.052.658.870.786.158.1
FastV [9]75%1.6×64.231.71321.374.236.931.353.356.468.182.156.4
Ours+FastV1.8×65.132.31368.677.835.332.055.258.871.185.858.2
", + "bbox": [ + 94, + 88, + 903, + 425 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7e58a098dd1779d463baf93a9eee9a29272d9856a0421573a5821d149b967201.jpg", + "table_caption": [ + "Table 1. Results on LLaVA1.5 7B/13B. We highlight the best result for each benchmark under the same reduction ratio. All baseline evaluations are reproduced using official codes on VLMEvalkit [17]. *Retrained to provide a fair comparison in Tab. 3." + ], + "table_footnote": [], + "table_body": "
MethodReduct RatioSpeed UpMMBench ENMM StarMMEHallusion BenchOCR VQAMM MUPOPESEED BenchIMGScience QACC BenchAvg
Original0%1×64.232.11403.646.353.334.286.760.670.018.853.6
Turbo [31]50%1.5×62.832.01352.944.047.634.085.658.669.117.951.9
Ours1.5×63.032.11362.346.250.534.785.759.570.418.652.8
Turbo [31]66%1.7×62.530.71341.846.246.633.384.958.169.718.851.8
Ours1.7×63.631.91368.247.947.234.085.358.770.719.652.8
", + "bbox": [ + 93, + 469, + 901, + 563 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Results on MMVP (MLLM with Multiple Vision-towers (CLIP+DINOv2)).", + "bbox": [ + 235, + 565, + 759, + 579 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MLVU [86]; and VQA tasks like OCRVQA [57], InfoVQA [56], MMBench-Video [19], TempCompass [52]. For BLIP, we evaluate on COCO image captioning [47].", + "bbox": [ + 89, + 598, + 482, + 645 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 89, + 654, + 230, + 669 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the following experiments, we show the effectiveness of FOLDER on different MLLMs. If not specified, we only apply FOLDER in the last layer of the visual encoder.", + "bbox": [ + 89, + 676, + 483, + 722 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.1. Inference Acceleration on MLLM", + "text_level": 1, + "bbox": [ + 89, + 729, + 372, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate FOLDER's plug-and-play acceleration performance in inference phase on various image understanding MLLMs [7, 24, 46, 50, 53] and multi-vision-tower based model [70]. If not specified, we evaluate MLLM's speed-up based on the time of the first output token.", + "bbox": [ + 89, + 750, + 482, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Single-Vision-Tower MLLM. Tab. 1 compares FOLDER with previous plug-and-play SOTA methods [9, 31, 75, 84] on LLaVA1.5. FOLDER achieves the best performance-speed trade-off, even surpassing original performance with over $50\\%$ reduction on several benchmarks. Compared to", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Turbo [31], which applies uniform merging across the vision encoder, FOLDER delivers superior results under the same reduction ratio, especially on visually dense tasks like OCRVQA. This confirms our key insight: token reduction is most effective when applied only in the final blocks.", + "bbox": [ + 511, + 599, + 906, + 674 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "While FastV [9] performs similarly to FOLDER on some benchmarks, its limitations (Sec. 2) make it inherently slower and harder to implement. In contrast, FOLDER is a plug-and-play module requiring minimal adjustments to integrate with vision backbones. By reducing visual tokens before interacting with the LLM, FOLDER mitigates architectural complexities. Notably, since FOLDER and FastV tackle token reduction from different angles (visual redundancy vs. task-oriented attention), their combination (50% × 50%) enables further reduction while maintaining performance. Similarly, LLM token reduction methods [75, 84] also face deployment challenges.", + "bbox": [ + 511, + 674, + 908, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To validate our approach, we replaced the matching function [31] with naive mean pooling. The result in Tab. 1", + "bbox": [ + 511, + 856, + 906, + 886 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23619", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/82298cdc84a38cb4ced8d142acc55fb5bb746ae64f9fcdb622efa4f9448cf6a2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TrainTrain ReductInfer ReductSpeed UpMMBench ENMM StarMMEA-OK VQAInfo VQAOCR VQASEED BenchIMGScience QAPOPEAvg
Original0%0%(1,1)×66.630.91371.177.015.554.959.468.886.458.7
Pyramiddrop [75]66%0%(1.5, 1.5)×65.733.01354.279.515.357.461.270.687.959.8
Pyramiddrop [75]66%66%(1.5, 1.5)×66.732.51359.177.515.656.659.070.883.058.6
Ours0%66%(1,1.6)×65.831.71366.977.315.857.658.870.786.159.1
Ours66%0%(1.5,1)×66.632.01407.778.916.657.760.969.388.360.1
Ours66%66%(1.5,1.6)×67.832.51393.479.516.457.961.371.586.960.4
Crossget [65]FT75%(-,1.8)×58.129.11176.968.513.243.652.661.980.051.8
Prumerge [63]FT75%(-,1.7)×58.728.21135.572.413.049.855.366.382.353.6
Pixel-shuffle [10]75%75%(1.7,1.8)×63.828.81343.575.414.854.557.467.386.457.3
Ours0%75%(1,1.8)×64.531.81308.676.716.250.259.071.185.057.8
Ours75%0%(1.7,1)×63.330.71359.579.115.856.260.265.687.858.5
Ours75%75%(1.7,1.8)×65.230.71325.779.515.252.359.567.487.258.1
", + "bbox": [ + 94, + 87, + 903, + 253 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b1084fe788182f0f9e5cbd2e985a8cd7c043655e268321576a64657f27981cb2.jpg", + "table_caption": [ + "Table 3. Training with FOLDER. We plug FOLDER into LLaVA1.5-13B during training (Train Reduct>0) and inference (Infer Reduct>0). FT means finetuning on trained model (for LLaVA-Prumerge [63] and Crossget [65]). All results are reproduced/retrained using official codes/checkpoints. We highlight the best result on each benchmark." + ], + "table_footnote": [], + "table_body": "
MethodReduct RatioSpeed UpMMB-V OverallMMB-V PerceptionVideo MME
Original0%1×1.041.0431.8
Turbo65%2.5×1.000.9929.8
Ours2.4×1.041.0531.9
Turbo75%2.8×0.980.9829.0
Ours2.7×1.031.0430.7
", + "bbox": [ + 94, + 311, + 480, + 407 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c323e61bf7f66af8bdc8d7c5b866ca91981807be903ed2b68ce46445756a737e.jpg", + "table_caption": [ + "Table 4. Results on VideoLLaVA." + ], + "table_footnote": [], + "table_body": "
MethodnframeMMB-V OverallVideo MMELongVideo Bench*Temp CompassMLVU
Original641.5158.756.662.569.6
Ours-75%641.4758.855.060.668.0
Ours-75%1281.4959.557.560.970.2
Ours-75%2561.5359.458.361.370.7
", + "bbox": [ + 91, + 446, + 482, + 515 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "shows the superiority of matching over pooling. More results on MiniGPT-4v2, non-LLaVA models like Idefics2-8B and Flamingov2-7B, and point cloud understanding model PointLLMv1.2-7B are shown in the appendix.", + "bbox": [ + 89, + 563, + 482, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multi-Vision-Tower MLLM. Multi-vision-tower-based models [32, 66, 70] enhance MLLMs' visual ability but aggravate the length issue. This makes our method particularly advantageous in this context. In Tab. 2 we evaluate on MMVP [70] with CLIP [59] and DINOv2 [58] as vision towers. By applying FOLDER, we trade a significant reduction in image token length (from 1152 to 576/384) with an acceptable drop. Note that MMVP is trained by interleaving vision tower features, which may cause positional confusion. Despite that, FOLDER still achieves a satisfying acceleration-performance trade-off.", + "bbox": [ + 88, + 626, + 482, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Video Understanding MLLM. Video-based MLLMs [30, 40, 45] can generate sequences exceeding 2000 tokens for a few frames, limiting their applicability. As shown in Tab. 4, by applying FOLDER on Video-LLaVA, we can reduce visual tokens up to $65\\%$ without any performance degradation, and up to $75\\%$ with only slight & acceptable drops on some benchmarks. Notably, we achieve $2.7 \\times$ infer", + "bbox": [ + 88, + 794, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/048ec043420148ea47a8c8aea3394f39f6a9c00188d8227a1c31506fa0f99ee6.jpg", + "table_caption": [ + "Table 5. Results on LLaVA-OV-7B across more video benchmarks. *Acc of videos with duration > 180s." + ], + "table_footnote": [], + "table_body": "
MethodReduct RatioTrainInference
Time(h)Speed UpTime(s)Speed UpMem
Original0%1871×0.3651×81.4G
FastV--0.2751.3×81.4G
Turbo50%1391.34×0.2501.5×56.6G
Ours1411.33×0.2501.5×56.6G
FastV--0.2551.4×81.4G
Turbo66%1231.52×0.2301.6×48.6G
Ours1241.51×0.2301.6×48.7G
FastV--0.2251.6×81.4G
Turbo75%1121.67×0.2051.8×44.6G
Ours1131.65×0.2051.8×44.7G
", + "bbox": [ + 514, + 311, + 916, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Acceleration For Training & Inference. We evaluate training (A100-80G GPU hours) and inference acceleration (on 1st-token) on llava1.5-13B. The memory is evaluated by its peak value on H20 with batch-size 40. Note that FastV [9] can only be used during inference.", + "bbox": [ + 511, + 494, + 906, + 565 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ence speedup with $75\\%$ token reduction, still outperforming Turbo with $66\\%$ reduction ratio. Furthermore, Tab. 5 shows that on LLaVA-OV-7B, with a $75\\%$ reduction ratio within each frame, using only half of the visual tokens $(128\\times 25\\%)$ compared to the original setup, we achieve even better performance, especially on long video understanding.", + "bbox": [ + 511, + 580, + 906, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.2. Training Acceleration on MLLM", + "text_level": 1, + "bbox": [ + 511, + 681, + 790, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As a universal plug-and-play module for MLLM, FOLDER can also be integrated into MLLM pretraining and SFT. Results on LLaVA1.5-13B are in Tab. 3. Surprisingly, when incorporating FOLDER with a $66\\%$ reduction ratio, LLaVA-1.5-13B exhibits comprehensive improvements across all benchmarks, even including visually dense tasks such as OCRVQA ( $\\uparrow 3\\%$ ) and InfoVQA ( $\\uparrow 0.9\\%$ ). This suggests high-frequency noise in visual sequences can be mitigated by merging similar visual tokens. Moreover, [9] highlights attention sparsity in LLMs' visual tokens, reinforcing the need to shorten visual sequences for better focus.", + "bbox": [ + 511, + 702, + 906, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Furthermore, training on reduced sequences maintains flexibility during inference, as the original sequence can still", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23620", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4c6a30fe890b4c5968de7b3afa8171b943c23ab008f19216b4fe48678cd0de8a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodReduct RatioThroughput im/sSpeed UpCIDErB@4
Original0%22.11×133.339.7
Uniform40.21.82×128.837.9
Last-360%38.41.74×129.538.2
Last-237.91.71×129.438.2
Last-137.51.71×132.239.0
Uniform47.42.14×127.237.5
Last-370%44.22.00×128.438.2
Last-243.71.98×130.138.4
Last-143.41.96×131.038.9
", + "bbox": [ + 106, + 85, + 467, + 229 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/98ec0ee9c5b8e3175037f518bb33215acdc7e77f2edc4f4decf78e0fade7a19e.jpg", + "table_caption": [ + "Table 7. Ablation on Merging Position. We evaluate BLIP [38] on captioning task. Last- $n$ refers to uniform reduction in last $n$ blocks, uniform refers to uniform reduction in every block." + ], + "table_footnote": [], + "table_body": "
MethodMMBPOPEMMEGQASciQA
VILA1.5-13b [46]74.384.91433.764.578.9
+FOLDER-50%72.884.01445.362.078.0
VITA1.5-7b [24]80.989.11728.947.895.8
+FOLDER-50%78.188.81688.350.894.1
WePOINTS1.5-7b [53]84.086.81616.059.295.0
+FOLDER-50%83.487.01642.659.194.3
", + "bbox": [ + 93, + 294, + 496, + 386 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "be used, leading to improved results on visually demanding tasks. Similar to masked augmentation [28, 43], FOLDER acts as a regularizer but goes beyond random dropping by accelerating both training and inference.", + "bbox": [ + 89, + 421, + 482, + 482 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Compared to Pyramidddrop [75], FOLDER achieves better performance on most benchmarks. It also outperforms finetuning-based methods [63, 65] even without additional training. By reducing tokens before the LLM, FOLDER serves as a flexible alternative to training-phase modules like pixel-shuffle, enhancing both efficiency and scalability.", + "bbox": [ + 89, + 484, + 482, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Speed & Memory Evaluation. In Tab. 6, we evaluate FOLDER's acceleration performance on LLaVA1.5-13B during training and inference. By reducing $66\\%$ of tokens, FOLDER achieves $1.5 \\times$ faster training on $8 \\times \\mathrm{A}100 - 80\\mathrm{G}$ and $1.6 \\times$ speedup during inference on H20-96G. Unlike FastV, which prunes tokens within the LLM, FOLDER removes them beforehand, significantly lowering peak memory usage. With a batch size of 40 on H20-96G, it saves $40\\%$ memory while maintaining a $66\\%$ reduction ratio, enabling nearly $3 \\times$ larger batches—especially beneficial for training.", + "bbox": [ + 89, + 577, + 482, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2.3. Simplified FOLDER for Inference Acceleration", + "text_level": 1, + "bbox": [ + 89, + 743, + 465, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As illustrated in Fig. 5, FOLDER is inserted between ATTN and FFN, requiring vision encoder modifications. Since the best results are achieved by reducing tokens only in the last block (Sec. 3.4), a simpler approach is inserting FOLDER at the output. Tab. 8 shows results on 3 SOTA MLLMs (VILA1.5 [46], WePOINTS1.5 [53], VITA1.5 [24]) with $50\\%$ reduction ratio and $1.4\\times$ acceleration, while maintaining or surpassing original performance on several benchmarks. More comparisons are in the appendix.", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0c43f90833358ae7c2f4475b98265af180c8d8232dbe90eb7b01b743b223c18d.jpg", + "table_caption": [ + "Table 8. FOLDER training-free plug-in for more MLLMs." + ], + "table_footnote": [], + "table_body": "
MethodReduct RatioB@4CIDErSPICE
Original0%133.339.723.8
Avg132.039.023.6
Drop50%130.038.723.3
Weighted132.239.123.6
Avg131.038.923.5
Drop60%128.938.223.1
Weighted131.539.123.4
", + "bbox": [ + 563, + 82, + 854, + 202 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 9. Ablation Study on Aggregation Operation. On BLIP [38] captioning task, we test the performance of 3 aggregation operations discussed in Sec. 3.3. For fair comparison, we adopt the aggregation only in the last block.", + "bbox": [ + 511, + 205, + 903, + 262 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 275, + 666, + 291 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct ablation studies on BLIP [38] for image captioning to validate the merging position and aggregation methods in Sec. 3. More ablations are in the appendix.", + "bbox": [ + 511, + 297, + 903, + 343 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Merging Position. Preliminary ViT tests in (Fig. 2, Fig. 3) show that due to propagation effects (Sec. 3.2) and diminishing reduction impact (Sec. 3.1), merging more tokens in the last blocks is preferable. Based on this, we test four strategies: uniform merging across all blocks and in the last $1/2/3$ blocks. Tab. 7 presents the results which confirm that merging in the last block yields the best performance. While the vision encoder's acceleration slightly affects BLIP, in MLLMs, where LLM processing dominates $(10 - 15 \\times$ inferior), merging position has minimal impact on speed-up.", + "bbox": [ + 511, + 344, + 905, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Aggregation Method. In Sec. 3.3, we explore three distinct aggregation methods: Dropping, Average, and Weighted Average based on vector norms. As illustrated in Tab. 9, both averaging methods outperform direct dropping, as directly dropping tokens causes greater information loss. To reduce module complexity, we used averaging in all MLLM experiments, as the differences are minimal.", + "bbox": [ + 511, + 496, + 905, + 601 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 617, + 633, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce FOLDER, a plug-and-play module for efficient visual token reduction in MLLMs. Through empirical analysis of the token reduction process, we develop a strategy that performs aggressive token reduction in the vision backbone's final layers while preserving key information. Our method achieves over $70\\%$ visual token reduction with $1.8 \\times$ inference speedup and $1.65 \\times$ training acceleration. FOLDER maintains or improves performance across tasks including video understanding, demonstrating its effectiveness in making MLLMs more computationally efficient without capability loss.", + "bbox": [ + 511, + 643, + 905, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. Part of this work was funded by Hi!PARIS Center on Data Analytics and Artificial Intelligence, by the European Union's Horizon Europe Research and Innovation Programme under grant agreement No. 101120237 (ELIAS), and by the French National Research Agency (ANR) in the framework of the JCJC project \"BANERA\" under Grant ANR-24-CE23-4369.", + "bbox": [ + 511, + 810, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23621", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 2", + "[2] Anthropic. The claude 3 model family: Opus, sonnet, haiku, 2024. 1", + "[3] Kirolos Ataallah, Xiaoqian Shen, Eslam Abdelrahman, Essam Sleiman, Mingchen Zhuge, Jian Ding, Deyao Zhu, Jürgen Schmidhuber, and Mohamed Elhoseiny. Goldfish: Vision-language understanding of arbitrarily long videos, 2024. 2", + "[4] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. Openflamingo: An opensource framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. 5", + "[5] Adam Block, Dylan J Foster, Akshay Krishnamurthy, Max Simchowitz, and Cyril Zhang. Butterfly effects of sgd noise: Error amplification in behavior cloning and autoregression. arXiv preprint arXiv:2310.11428, 2023. 3", + "[6] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461, 2022. 2, 4, 5", + "[7] Jun Chen, Deyao Zhu, Xiaoqian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023.2,5,6", + "[8] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024.5", + "[9] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. arXiv preprint arXiv:2403.06764, 2024. 2, 5, 6, 7", + "[10] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhang-wei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. Science China Information Sciences, 67(12):220101, 2024. 2, 7", + "[11] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024. 1, 2", + "[12] Kanzhi Cheng, Wenpo Song, Zheng Ma, Wenhao Zhu, Zixuan Zhu, and Jianbing Zhang. Beyond generic: Enhancing" + ], + "bbox": [ + 93, + 114, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "image captioning with real-world knowledge using vision-language pre-training model. In Proceedings of the 31st ACM International Conference on Multimedia, pages 5038-5047, 2023. 1", + "[13] Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning, 2023. 2", + "[14] Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness, 2022. 2, 3", + "[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 3", + "[16] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3", + "[17] Haodong Duan, Junming Yang, Yuxuan Qiao, Xinyu Fang, Lin Chen, Yuan Liu, Xiaoyi Dong, Yuhang Zang, Pan Zhang, Jiaqi Wang, et al. Vlmevalkit: An open-source toolkit for evaluating large multi-modality models. In Proceedings of the 32nd ACM international conference on multimedia, pages 11198-11201, 2024. 5, 6", + "[18] Carl Eckart and Gale Young. The approximation of one matrix by another of lower rank. Psychometrika, 1(3):211-218, 1936. 3", + "[19] Xinyu Fang, Kangrui Mao, Haodong Duan, Xiangyu Zhao, Yining Li, Dahua Lin, and Kai Chen. Mmbench-video: A long-form multi-shot benchmark for holistic video understanding. arXiv preprint arXiv:2406.14515, 2024. 6", + "[20] Zhiyuan Fang, Jianfeng Wang, Xiaowei Hu, Lijuan Wang, Yezhou Yang, and Zicheng Liu. Compressing visual-linguistic model via knowledge distillation. In ICCV, 2021. 2", + "[21] Elias Frantar, Saleh Ashkboos, Torsten Hoefler, and Dan Alistarh. Gptq: Accurate post-training quantization for generative pre-trained transformers. arXiv preprint arXiv:2210.17323, 2022. 2", + "[22] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 5", + "[23] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 5", + "[24] Chaoyou Fu, Haojia Lin, Xiong Wang, Yi-Fan Zhang, Yunhang Shen, Xiaoyu Liu, Yangze Li, Zuwei Long, Heting Gao, Ke Li, et al. Vita-1.5: Towards gpt-4o level real-time vision and speech interaction. arXiv preprint arXiv:2501.01957, 2025. 5, 6, 8", + "[25] Grok. Grok1.5 homepage. https://x.ai/blog/grok-1.5v, 2024. Accessed: 2024-11-14. 5", + "[26] Tianrui Guan, Fuxiao Liu, Xiyang Wu, Ruiqi Xian, Zongxia Li, Xiaoyu Liu, Xijun Wang, Lichang Chen, Furong Huang," + ], + "bbox": [ + 516, + 92, + 905, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23622", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Yaser Yacoob, et al. Hallusionbench: an advanced diagnostic suite for entangled language hallucination and visual illusion in large vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14375-14385, 2024. 5", + "[27] Yuhang Han, Xuyang Liu, Pengxiang Ding, Donglin Wang, Honggang Chen, Qingsen Yan, and Siteng Huang. Rethinking token reduction in mllms: Towards a unified paradigm for training-free acceleration. arXiv preprint arXiv:2411.17686, 2024. 2, 5", + "[28] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 8", + "[29] Wenxuan Huang, Zijie Zhai, Yunhang Shen, Shaoshen Cao, Fei Zhao, Xiangfeng Xu, Zheyu Ye, and Shaohui Lin. Dynamic-llava: Efficient multimodal large language models via dynamic vision-language context sparsification. arXiv preprint arXiv:2412.00876, 2024. 2", + "[30] Peng Jin, Ryuichi Takanobu, Wancai Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13700-13710, 2024. 1, 7", + "[31] Chen Ju, Haicheng Wang, Haozhe Cheng, Xu Chen, Zhonghua Zhai, Weilin Huang, Jinsong Lan, Shuai Xiao, and Bo Zheng. Turbo: Informativity-driven acceleration plug-in for vision-language large models. In European Conference on Computer Vision, pages 436-455. Springer, 2025. 2, 4, 5, 6", + "[32] Oguzhan Fatih Kar, Alessio Tonioni, Petra Poklukar, Achin Kulshrestha, Amir Zamir, and Federico Tombari. Brave: Broadening the visual encoding of vision-language models. arXiv preprint arXiv:2404.07204, 2024. 1, 2, 7", + "[33] Zaid Khan, Vijay Kumar BG, Samuel Schulter, Xiang Yu, Yun Fu, and Manmohan Chandraker. Q: How to specialize large vision-language models to data-scarce vqa tasks? a: Self-train on unlabeled images! In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15005-15015, 2023. 1", + "[34] Zhenglun Kong, Peiyan Dong, Xiaolong Ma, Xin Meng, Mengshu Sun, Wei Niu, Xuan Shen, Geng Yuan, Bin Ren, Minghai Qin, Hao Tang, and Yanzhi Wang. Spvit: Enabling faster vision transformers via soft token pruning, 2022. 2", + "[35] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? Advances in Neural Information Processing Systems, 37:87874-87907, 2024. 5", + "[36] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 5", + "[37] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 5" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 4, 5, 8", + "[39] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 2", + "[40] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 1, 7", + "[41] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and performant baseline for vision and language, 2019. 2", + "[42] Wentong Li, Yuqian Yuan, Jian Liu, Dongqi Tang, Song Wang, Jie Qin, Jianke Zhu, and Lei Zhang. Tokenpacker: Efficient visual projector for multimodal llm. arXiv preprint arXiv:2407.02392, 2024. 2", + "[43] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pre-training via masking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 23390-23400, 2023. 8", + "[44] Youwei Liang, Chongjian Ge, Zhan Tong, Yibing Song, Jue Wang, and Pengtao Xie. Not all patches are what you need: Expediting vision transformers via token reorganizations, 2022. 2", + "[45] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 1, 2, 5, 7", + "[46] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 26689-26699, 2024. 5, 6, 8", + "[47] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6", + "[48] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2", + "[49] Hao Liu, Matei Zaharia, and Pieter Abbeel. Ring attention with blockwise transformers for near-infinite context, 2023. 2", + "[50] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 2, 5, 6", + "[51] Ting Liu, Liangtao Shi, Richang Hong, Yue Hu, Quanjun Yin, and Linfeng Zhang. Multi-stage vision token dropping: Towards efficient multimodal large language model. arXiv preprint arXiv:2411.10803, 2024. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23623", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Yuanxin Liu, Shicheng Li, Yi Liu, Yuxiang Wang, Shuhuai Ren, Lei Li, Sishuo Chen, Xu Sun, and Lu Hou. Tempcompass: Do video llms really understand videos? arXiv preprint arXiv:2403.00476, 2024. 6", + "[53] Yuan Liu, Le Tian, Xiao Zhou, Xinyu Gao, Kavio Yu, Yang Yu, and Jie Zhou. Points1. 5: Building a vision-language model towards real world applications. arXiv preprint arXiv:2412.08443, 2024. 5, 6, 8", + "[54] Zhijian Liu, Ligeng Zhu, Baifeng Shi, Zhuoyang Zhang, Yuming Lou, Shang Yang, Haocheng Xi, Shiyi Cao, Yuxian Gu, Dacheng Li, et al. Nvila: Efficient frontier visual language models. arXiv preprint arXiv:2412.04468, 2024. 2", + "[55] Pan Lu, Swaroop Mishra, Tony Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In The 36th Conference on Neural Information Processing Systems (NeurIPS), 2022. 5", + "[56] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1697-1706, 2022. 6", + "[57] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering by reading text in images. In ICDAR, 2019. 6", + "[58] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 7", + "[59] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7", + "[60] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. NeurIPS, 2021. 2", + "[61] Yossi Rubner, Carlo Tomasi, and Leonidas J Guibas. The earth mover's distance as a metric for image retrieval. International journal of computer vision, 40:99-121, 2000. 3", + "[62] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge. In European conference on computer vision, pages 146–162. Springer, 2022. 5", + "[63] Yuzhang Shang, Mu Cai, Bingxin Xu, Yong Jae Lee, and Yan Yan. Llava-prumerge: Adaptive token reduction for efficient large multimodal models. arXiv preprint arXiv:2403.15388, 2024. 2, 4, 7, 8", + "[64] Dachuan Shi, Chaofan Tao, Ying Jin, Zhendong Yang, Chun Yuan, and Jiaqi Wang. Upop: Unified and progressive pruning for compressing vision-language transformers. arXiv preprint arXiv:2301.13741, 2023. 2" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[65] Dachuan Shi, Chaofan Tao, Anyi Rao, Zhendong Yang, Chun Yuan, and Jiaqi Wang. Crossget: Cross-guided ensemble of tokens for accelerating vision-language transformers. arXiv preprint arXiv:2305.17455, 2023. 2, 5, 7, 8", + "[66] Min Shi, Fuxiao Liu, Shihao Wang, Shijia Liao, Subhashree Radhakrishnan, De-An Huang, Hongxu Yin, Karan Sapra, Yaser Yacoob, Humphrey Shi, et al. Eagle: Exploring the design space for multimodal llms with mixture of encoders. arXiv preprint arXiv:2408.15998, 2024. 1, 2, 7", + "[67] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. FLAVA: A foundational language and vision alignment model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2", + "[68] Hao Tan and Mohit Bansal. LXMERT: Learning cross-modality encoder representations from transformers. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, EMNLP-IJCNLP, November 3-7, 2019, 2019.", + "[69] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 2", + "[70] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9568-9578, 2024. 1, 2, 5, 6, 7", + "[71] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 2", + "[72] Tiannan Wang, Wangchunshu Zhou, Yan Zeng, and Xinsong Zhang. Efficientvlm: Fast and accurate vision-language models via knowledge distillation and modal-adaptive pruning. arXiv preprint arXiv:2210.07795, 2022. 2", + "[73] Haoning Wu, Dongxu Li, Bei Chen, and Junnan Li. Longvideobench: A benchmark for long-context interleaved video-language understanding. Advances in Neural Information Processing Systems, 37:28828-28857, 2024. 5", + "[74] Guangxuan Xiao, Ji Lin, Mickael Seznec, Hao Wu, Julien Demouth, and Song Han. Smoothquant: Accurate and efficient post-training quantization for large language models. In International Conference on Machine Learning, pages 38087-38099. PMLR, 2023. 2", + "[75] Long Xing, Qidong Huang, Xiaoyi Dong, Jiajie Lu, Pan Zhang, Yuhang Zang, Yuhang Cao, Conghui He, Jiaqi Wang, Feng Wu, et al. Pyramiddrop: Accelerating your large vision-language models via pyramid visual redundancy reduction. arXiv preprint arXiv:2410.17247, 2024. 2, 5, 6, 7, 8", + "[76] Runsen Xu, Xiaolong Wang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Pointllm: Empowering large language models to understand point clouds. In European" + ], + "bbox": [ + 516, + 90, + 906, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "23624", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision, pages 131-147. Springer, 2024. 5", + "[77] Senqiao Yang, Yukang Chen, Zhuotao Tian, Chengyao Wang, Jingyao Li, Bei Yu, and Jiaya Jia. Visionzip: Longer is better but not necessary in vision language models. arXiv preprint arXiv:2412.04467, 2024. 2", + "[78] Xubing Ye, Yukang Gan, Yixiao Ge, Xiao-Ping Zhang, and Yansong Tang. Atp-llava: Adaptive token pruning for large vision language models. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 24972–24982, 2025. 2", + "[79] Xubing Ye, Yukang Gan, Xiaoke Huang, Yixiao Ge, and Yansong Tang. Voco-llama: Towards vision compression with large language models. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 29836–29846, 2025. 2", + "[80] Li Yifan, Du Yifan, Zhou Kun, Wang Jinpeng, Xin Zhao Wayne, and Wen Ji-Rong. Evaluating object hallucination in large vision-language models. In The 2023 Conference on Empirical Methods in Natural Language Processing, 2023. 5", + "[81] Liu Yuan, Duan Haodong, Zhang Yuanhan, Li Bo, Zhang Songyang, Zhao Wangbo, Yuan Yike, Wang Jiaqi, He Conghui, Liu Ziwei, Chen Kai, and Lin Dahua. Mmbench: Is your multi-modal model an all-around player? arXiv:2307.06281, 2023. 5", + "[82] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Ren-liang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of CVPR, 2024. 5", + "[83] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding, 2023. 2", + "[84] Yuan Zhang, Chun-Kai Fan, Junpeng Ma, Wenzhao Zheng, Tao Huang, Kuan Cheng, Denis Gudovskiy, Tomoyuki Okuno, Yohei Nakata, Kurt Keutzer, et al. Sparsevm: Visual token sparsification for efficient vision-language model inference. arXiv preprint arXiv:2410.04417, 2024. 2, 6", + "[85] Yiwu Zhong, Zhuoming Liu, Yin Li, and Liwei Wang. Aim: Adaptive inference of multi-modal llms via token merging and pruning. arXiv preprint arXiv:2412.03248, 2024. 2", + "[86] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. MLvu: A comprehensive benchmark for multitask long video understanding. arXiv e-prints, pages arXiv-2406, 2024. 6" + ], + "bbox": [ + 91, + 90, + 482, + 796 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "23625", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/FOLDER_ Accelerating Multi-Modal Large Language Models with Enhanced Performance/7a17b26d-c0b2-4285-9105-7de16ee42381_model.json b/2025/FOLDER_ Accelerating Multi-Modal Large Language Models with Enhanced Performance/7a17b26d-c0b2-4285-9105-7de16ee42381_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ad95648fd63231db197200abe3af20c172a1e6eb --- /dev/null +++ b/2025/FOLDER_ Accelerating Multi-Modal Large Language Models with Enhanced Performance/7a17b26d-c0b2-4285-9105-7de16ee42381_model.json @@ -0,0 +1,2578 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.109, + 0.13, + 0.89, + 0.175 + ], + "angle": 0, + "content": "FOLDER: Accelerating Multi-modal Large Language Models with Enhanced Performance" + }, + { + "type": "text", + "bbox": [ + 0.16, + 0.203, + 0.834, + 0.239 + ], + "angle": 0, + "content": "Haicheng Wang\\(^{1,2,4*}\\), Zhemeng Yu\\(^{1,2*}\\), Gabriele Spadaro\\(^{2,3}\\), Chen Ju\\(^{4}\\), Victor Quétu\\(^{2}\\), Shuai Xiao\\(^{4\\boxtimes}\\), Enzo Tartaglione\\(^{2\\boxtimes}\\)" + }, + { + "type": "text", + "bbox": [ + 0.138, + 0.239, + 0.856, + 0.257 + ], + "angle": 0, + "content": "\\(^{1}\\) Paris Elite Institute of Technology, Shanghai Jiao Tong University \\(^{3}\\) University of Turin" + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.257, + 0.863, + 0.275 + ], + "angle": 0, + "content": "\\(^{2}\\) LTCI, Télecom Paris, Institut Polytechnique de Paris \\(^{4}\\) Taobao & Tmall Group of Alibaba" + }, + { + "type": "list", + "bbox": [ + 0.135, + 0.239, + 0.863, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.278, + 0.886, + 0.293 + ], + "angle": 0, + "content": "anakinSkywalker@sjtu.edu.cn, shuai.xsh@alibaba-inc.com, enzo.tartaglione@telecom-paris.fr" + }, + { + "type": "title", + "bbox": [ + 0.268, + 0.327, + 0.347, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.359, + 0.485, + 0.707 + ], + "angle": 0, + "content": "Recently, Multi-modal Large Language Models (MLLMs) have shown remarkable effectiveness for multi-modal tasks due to their ability of cross-modal understanding. However, processing long sequences of visual tokens extracted from visual backbones poses challenges for deployment in real-time applications. To address this issue, we introduce FOLDER, a simple yet effective plug-and-play module designed to reduce the length of the visual token sequence, mitigating computational and memory demands during both training and inference. Through a comprehensive analysis of the token reduction process in the vision encoder, we analyze the information loss introduced by different reduction strategies and develop FOLDER to preserve key information while removing visual redundancy. We show the effectiveness of FOLDER by integrating it into the visual backbone of various MLLMs, significantly accelerating the inference phase. Furthermore, we evaluate its utility as a training accelerator or even performance booster for MLLMs. FOLDER achieves comparable or even better performance than the original models, while dramatically reducing complexity by removing up to \\(70\\%\\) of visual tokens. Our code is available at FOLDER Repository." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.727, + 0.222, + 0.743 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.753, + 0.484, + 0.876 + ], + "angle": 0, + "content": "Multi-modal Large Language Models (MLLMs) excel in tasks like image captioning [12] and visual question answering [33] by learning joint visual-text representations. State-of-the-art models such as GPT-4V [1] and Claude 3 [2] show strong multimodal understanding ability. However, these models face concrete challenges in terms of computational efficiency, particularly when processing visual inputs. Modern MLLMs typically generate long sequences of im" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.327, + 0.907, + 0.508 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.511, + 0.904, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.646, + 0.907, + 0.69 + ], + "angle": 0, + "content": "Figure 1. FOLDER as Accelerator & Booster. As a plug-and-play module, FOLDER can be used in both training and inference, with considerable acceleration and even a performance boost." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.718, + 0.907, + 0.901 + ], + "angle": 0, + "content": "age tokens through their visual backbones. In some application scenarios, this problem becomes even more evident: recent advancements in high-resolution-supporting MLLMs [11, 71] might involve over 2000 visual tokens. Similarly, challenges also occur in multi-visual-expert architectures [32, 66, 70] that leverage multiple visual encoders. These issues are even more severe in video understanding [30, 40, 45] due to the extra temporal complexity. The resulting token explosion hinders real-time deployment because of the quadratic cost of the attention mechanism. Several studies have been conducted to address this issue. Existing approaches can be roughly divided into two" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.887, + 0.381, + 0.901 + ], + "angle": 0, + "content": "*: Equal contribution, \\( {}^{20} \\) : Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23614" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.348 + ], + "angle": 0, + "content": "categories. The first category focuses on training phase solutions, such as Q-Former [39], pixel-shuffle [10] and pooling-based techniques [7]. However, these methods often suffer from performance degradation and exhibit limited scalability, due to their custom design for specific architectures. The second category aims to design plug-and-play token reduction modules for the inference phase [9, 27, 31, 65, 75, 84]. Specifically, [9, 75, 84] reduce visual tokens inside LLM. Though effective, these methods are often hard to implement, require adaptation for existing LLM acceleration modules [13], and prove unsuitable for multi-round conversations, thus greatly limiting their usage. Meanwhile, [27, 31, 65] identify the inherent redundancy in visual tokens, but fail to quantify the information loss during the token reduction process. Thus their reduction strategies are often arbitrary (uniform reduction for each layer), leading to sub-optimal outcomes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.351, + 0.483, + 0.533 + ], + "angle": 0, + "content": "To find a reduction strategy that can be universally applicable and best preserve the performance, we propose a preliminary investigation into the vision backbone, to answer a fundamental question: Where does information loss originate? To model it, we identify three key factors that impact information loss during token removal: Reduction Impact, Propagation Effect, and Aggregation Method. Through a comprehensive empirical study of these factors, we devise a strategy that enhances the preservation of information while significantly reducing redundancy in the vision backbone. FOLDER is consequently designed to implement this strategy with minimal computational overhead." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.536, + 0.483, + 0.762 + ], + "angle": 0, + "content": "Our plug-and-play solution can be easily integrated into the visual backbones of various MLLMs, effectively reducing the number of visual tokens while incurring negligible information loss (Fig. 1). We first incorporate FOLDER during MLLMs inference, reducing over \\(60\\%\\) of visual tokens and achieving comparable, or even gains on performance across multiple tasks. Furthermore, when integrated into MLLMs pre-training, our module can be used as a training accelerator, as well as an effective regularization term, leading to remarkable performance improvements across all benchmarks, with a reduction ratio up to \\(70\\%\\). Our method is on-the-field proven to be a flexible, nearly lossless, and highly efficient token reduction strategy, offering a dual-purpose solution to the visual token reduction challenge. Our contributions can be summarized as follows." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.765, + 0.483, + 0.84 + ], + "angle": 0, + "content": "- We conduct an in-depth analysis for information decay during token reduction in the vision backbone, identifying and quantifying key factors involved (Sec. 3.1, Sec. 3.2 and Sec. 3.3). These insights offer a clear understanding of how token reduction impacts the information flow." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "- Leveraging the above analysis, we develop a simple yet effective plug-and-play visual token reduction module for MLLMs (Sec. 3.4). This novel token reduction strategy aggressively reduces the number of tokens only in the" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.765, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "last blocks of the visual encoder. On various models and benchmarks, FOLDER speeds up the inference \\(1.7 - 2.4 \\times\\) with negligible loss or even improvement (Sec. 4.2.1)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.137, + 0.905, + 0.183 + ], + "angle": 0, + "content": "- We show that FOLDER can also be used in MLLM training (Sec. 4.2.2), to speed-up while boosting performance, serving as a dual-solution for both training and inference." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.196, + 0.664, + 0.212 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.221, + 0.905, + 0.493 + ], + "angle": 0, + "content": "Multi-modal Large Language Models. MLLMs excel at understanding multi-modal information [1, 3, 7, 39, 41, 45, 48, 67-69, 71, 83], with common architectures comprising i.) a Visual Backbone, ii.) a Vision-Language Connector, and iii.) a Pre-trained LLM. However, integrating visual data significantly increases computational cost due to the large number of visual tokens. For example, LLaVA1.5 [50] processes 576 tokens for a \\(336 \\times 336\\) image, while VideoLLaVA [45] exceeds 2000 tokens with just 8 frames at \\(224 \\times 224\\). Recent efforts to enhance MLLMs with multiple vision towers [32, 66, 70] further exacerbate this issue. MLLMs Acceleration. To accelerate large models, many methods focus on system-level optimizations, such as FlashAttention [14], vLLM [13], and RingAttention [49]. Model-centric techniques like knowledge distillation [20, 72], quantization [21, 74], and pruning [60, 64, 72] also reduce size and cost. However, these approaches overlook redundant data, which unnecessarily prolongs sequences." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.494, + 0.905, + 0.614 + ], + "angle": 0, + "content": "Token Reduction in MLLMs. Some preliminary studies are conducted on token reduction for Vision Transformers [6, 34, 44, 85]. In the context of MLLMs, several token reduction methods, including Q-Former [39], resampler [11], pooling [7] and some other methods [29, 42, 54, 78, 79] are proposed to reduce visual tokens during the training process. However, these approaches often suffer from performance degradation and a lack of scalability." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.614, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Meanwhile, some studies try to handle the token reduction problem during inference. In particular, ToMe [6] proposes a training-free token merging strategy for unimodal image classification. Based on that, Turbo [31] proposes an improved merging strategy considering both mutual redundancy and semantic value, which serves as a plug-in module in MLLMs. Recently, FastV [9] recognizes the visual attention sparsity in LLM, thus pruning visual tokens inside the LLM based on the task-oriented attention importance. Though effective, these methods have clear drawbacks. For ToMe/Turbo-like methods [6, 27, 31, 63, 65, 77] that reduce tokens in the vision encoder, most of them adopt uniform progressive merging across layers, which is arbitrary and inflexible. While FastV-like methods [9, 51, 75, 84] that cut tokens inside LLM layers mainly suffer from 2 problems: i.) They often require interactions with instruction tokens, making it unsuitable for multi-round/long-term visual QA. ii.) Since these approaches are inserted in LLM and often require specific hyperparameters, it is difficult to adapt" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23615" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.096, + 0.475, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.282, + 0.483, + 0.311 + ], + "angle": 0, + "content": "Figure 2. Minimum Number of Tokens \\( k \\) to Preserve Energy \\( t_E \\) Across Blocks. We evaluate on 3 different \\( t_E \\) for every block." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.335, + 0.484, + 0.429 + ], + "angle": 0, + "content": "for complex LLM modules [14] and hard to deploy. Besides, few approaches have been verified to be effective in both training and training-free inference ways. To address the above issues, we seek to find an effective, information-guided, and user-friendly token reduction method that can be easily applied in both training and inference." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.444, + 0.183, + 0.459 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.471, + 0.484, + 0.546 + ], + "angle": 0, + "content": "In this section, we establish FOLDER by first answering three key questions for token reduction in vision backbone: i.) how many tokens to reduce (Sec. 3.1), ii.) in which block (Sec. 3.2) and iii.) through which aggregation method (Sec. 3.3). Then, we present our method in Sec. 3.4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.547, + 0.484, + 0.624 + ], + "angle": 0, + "content": "Empirical Setup. To answer these three crucial questions, we define a common empirical setup. Specifically, we use a pre-trained 12-block ViT-B [16] model and conduct the empirical experiments on ImageNet-1k [15]. Results on more diverse setups are provided in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.636, + 0.318, + 0.652 + ], + "angle": 0, + "content": "3.1. Token Reduction Impact" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.484, + 0.751 + ], + "angle": 0, + "content": "To explore the link between token reduction and information loss, we use Singular Value Decomposition (SVD) to track the energy of the reduced token sequence. More precisely, given a token sequence \\(\\mathbf{X} \\in \\mathbb{R}^{n \\times d}\\) (with \\(n\\) being the number of tokens and \\(d\\) their size), we can decompose it by applying SVD on \\(\\mathbf{X}^T\\):" + }, + { + "type": "equation", + "bbox": [ + 0.232, + 0.763, + 0.483, + 0.781 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} ^ {T} = \\mathbf {U} \\boldsymbol {\\Sigma} \\mathbf {V} ^ {T}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.903 + ], + "angle": 0, + "content": "where \\(\\mathbf{U} \\in \\mathbb{R}^{d \\times d}\\) and \\(\\mathbf{V} \\in \\mathbb{R}^{n \\times n}\\) are two orthogonal matrices and \\(\\pmb{\\Sigma} \\in \\mathbb{R}^{d \\times n}\\) is a diagonal matrix, with the singular values \\(\\sigma_{i} = \\pmb{\\Sigma}_{ii}\\) representing the variances (energy) in the new compositional directions. In this way, SVD yields the optimal energy concentration that can be ordered by their contribution to the total variance (energy) [18]. Based on this, we can estimate how much variance we preserve by" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.096, + 0.895, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.282, + 0.907, + 0.325 + ], + "angle": 0, + "content": "Figure 3. EMD Distance Between Reduced and Original Output Distributions under 3 Reduction Ratios. We compare the EMD distance by exerting token reduction on different blocks." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.347, + 0.807, + 0.363 + ], + "angle": 0, + "content": "only taking the largest \\(k\\) singular values, as:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.373, + 0.907, + 0.412 + ], + "angle": 0, + "content": "\\[\nE (k) = \\frac {\\sum_ {i = 1} ^ {k} \\sigma_ {i}}{\\sum_ {i = 1} ^ {n} \\sigma_ {i}}, \\quad \\sigma_ {i} \\geq \\sigma_ {i + 1} \\quad \\forall i. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.419, + 0.907, + 0.465 + ], + "angle": 0, + "content": "By definition, SVD provides an upper bound on preserved energy when keeping \\( k \\) tokens. Thus, by setting a threshold \\( t_E \\), we can estimate the required \\( k \\) for each block as:" + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.476, + 0.905, + 0.493 + ], + "angle": 0, + "content": "\\[\nk = \\min \\left\\{k \\mid E (k) \\geq t _ {E} \\right\\}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.504, + 0.906, + 0.595 + ], + "angle": 0, + "content": "Observations. With a threshold \\( t_E \\), \\( k \\) varies across blocks. We monitor \\( k \\) under this threshold and find, as shown in Fig. 2, that later blocks require far fewer tokens than earlier ones to maintain the same energy. This highlights the potential for block-adaptive token reduction, suggesting the first and last blocks as key candidates." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.605, + 0.747, + 0.621 + ], + "angle": 0, + "content": "3.2. Token Propagation Effect" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.627, + 0.907, + 0.717 + ], + "angle": 0, + "content": "While token reduction impact provides an upper bound on information preservation, it overlooks inter-block dependencies and therefore may not adequately reflect the effects on subsequent blocks. In transformer architectures, errors from earlier blocks can propagate and amplify through successive layers [5], potentially affecting final performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.717, + 0.906, + 0.807 + ], + "angle": 0, + "content": "To analyze this propagation effect, we examine how token reduction at different blocks affects the final output distribution. Thus, by comparing the original output token distribution (without reduction) and the one obtained by operating token reduction only in one block \\( b \\), we can show the presence of such a propagation effect." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.808, + 0.907, + 0.901 + ], + "angle": 0, + "content": "To measure this distortion, we consider output tokens from the original model \\(\\mathbf{Y} \\in \\mathbb{R}^{n \\times d}\\) and the ones obtained when \\(n - k\\) tokens are reduced in the block \\(b\\): \\(\\tilde{\\mathbf{Y}}_b \\in \\mathbb{R}^{k \\times d}\\). Then, we define the empirical distributions \\(P_Y\\) and \\(P_{\\tilde{\\mathbf{Y}}_b}\\) supported on \\(\\mathbf{Y}\\) and \\(\\tilde{\\mathbf{Y}}_b\\) respectively, and we evaluate Earth Mover's Distance (EMD) [61] between them:" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23616" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.092, + 0.478, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.28, + 0.483, + 0.325 + ], + "angle": 0, + "content": "Figure 4. EMD & Accuracy Comparison for Aggregation Methods. We compare three aggregation methods under different reduction ratios on both EMD and Accuracy." + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.354, + 0.483, + 0.378 + ], + "angle": 0, + "content": "\\[\n\\operatorname {E M D} \\left(P _ {Y}, P _ {\\tilde {Y} _ {b}}\\right) = \\min _ {\\gamma} \\langle \\gamma , \\mathbf {M} \\rangle_ {F}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.387, + 0.484, + 0.482 + ], + "angle": 0, + "content": "where \\(\\mathbf{M} = (d_{ij})_{n\\times k} = (\\| y_i - \\tilde{y}_{bj}\\| _2)_{n\\times k}\\) denotes the metric cost matrix, with \\(y_{i}\\) and \\(\\tilde{y}_{bj}\\) being points in the support of distributions \\(P_Y\\) and \\(P_{\\tilde{Y}_b}\\) respectively. \\(\\gamma = (\\gamma_{ij})_{n\\times k}\\) represents the optimal transport matrix, which satisfies the non-negative constraint \\(\\gamma_{ij}\\geq 0,\\forall i,j\\) and the marginal constraints \\(\\sum_{j = 1}^{k}\\gamma_{ij} = P_Y(y_i)\\) and \\(\\sum_{i = 1}^{n}\\gamma_{ij} = P_{\\tilde{Y}_b}(\\tilde{y}_{bj})\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.482, + 0.484, + 0.664 + ], + "angle": 0, + "content": "Observations. Due to the propagation effect, token reduction at different blocks causes varying levels of output distortion. We monitor EMD (Eq. (4)) when reducing the same number of tokens per block and find, as shown in Fig.3, that early-block reductions lead to significantly higher EMD values. This confirms that distortion amplifies across layers, making early blocks unsuitable for token reduction, contrary to Fig.2. Conversely, with a massive \\(75\\%\\) reduction, EMD remains low in the last layers. This suggests that token reduction is most effective at the network's end, even with a high reduction rate. Results on BLIP [38] (Tab. 7) and appendix further support this conclusion." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.677, + 0.338, + 0.693 + ], + "angle": 0, + "content": "3.3. Token Aggregation Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.7, + 0.484, + 0.852 + ], + "angle": 0, + "content": "In Sec. 3.1, we use SVD to estimate the upper bound of information preservation. However, its computational complexity prevents direct application for token reduction. Following prior works [6, 31], we decompose the process into two stages: (i) Token Matching, grouping tokens based on a matching function, and (ii) Token Aggregation, merging grouped tokens. While matching functions have been explored [6, 31], aggregation remains understudied. Formally, given a set of tokens \\(\\{\\mathbf{x}_1,\\mathbf{x}_2,\\dots ,\\mathbf{x}_m\\}\\) to be merged into \\(\\mathbf{y}\\), we define the aggregation operation as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.865, + 0.483, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\mathbf {y} = \\sum_ {i = 1} ^ {m} \\alpha_ {i} \\mathbf {x} _ {i}, \\text {s . t .} \\sum_ {i = 1} ^ {m} \\alpha_ {i} = 1. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.193 + ], + "angle": 0, + "content": "Based on Eq. (5), we formulate three different token aggregation approaches: i.) Average Merging, with \\(\\alpha_{i} = \\frac{1}{m},\\forall i\\) ii.) Weighted Merging, with \\(\\alpha_{i} = \\frac{\\|x_{i}\\|_{2}}{\\sum_{j = 1}^{m}\\|x_{j}\\|_{2}},\\forall i;\\) iii.) Direct Dropping, with \\(\\alpha_{i} = 0,\\forall i\\neq i_{max},\\alpha_{i_{max}} = 1\\) where \\(i_{max} = \\underset {i}{\\operatorname{argmax}}\\| x_i\\| _2\\) . Here, the norm \\(\\| \\cdot \\| _2\\) is used as the importance score for tokens." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.195, + 0.907, + 0.316 + ], + "angle": 0, + "content": "Observations. Similarly to Sec. 3.2, we compare aggregation methods using EMD. As shown in Tab. 9 and Fig. 4, both merging methods yield lower discrepancy than Direct Dropping. Given its comparable performance to Weighted Merging with greater simplicity, we select Average Merging as our preferred method. Additionally, Fig. 4 shows that a lower EMD indeed correlates with higher accuracy. Further supporting results are in Tab. 9 and appendix." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.33, + 0.627, + 0.344 + ], + "angle": 0, + "content": "3.4.FOLDER" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.354, + 0.907, + 0.549 + ], + "angle": 0, + "content": "Motivation. Based on the above empirical analyses, we deduce that: i.) Token reduction should take place in the last blocks, where aggressive reduction is permitted. ii.) Merging is better than dropping as an aggregation method. Therefore, our goal is to design an algorithm that can reduce as many tokens as possible in one block. Previous works [63] mainly leverage clustering for token merging, which is slow and unsuitable for parallel computing. To avoid such issues, we refer to bipartite matching [6], which is however limited by a maximum reduction ratio of \\(1/2\\). Thus, to aggressively reduce tokens in one block, we propose FOLDER, an advanced bipartite matching strategy that can reduce an arbitrary number of tokens." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.552, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Strategy. As illustrated in Fig. 5, our strategy focuses on the last blocks (or the last) in the vision backbone. In some cases, the target reduction \\( r \\) exceeds half of the total tokens \\( \\lfloor n / 2 \\rfloor \\) in a block, a scenario we call reduction overflow. To handle this, FOLDER automatically applies an iterative \"FOLD\" operation. As detailed in Alg. 1, if reduction overflow occurs in an iteration \\( (r_{\\text{remain}} > \\lfloor n' / 2 \\rfloor \\) in line 7), we reduce tokens by half \\( (r_{\\text{fold}} = \\lfloor n' / 2 \\rfloor) \\). The process begins by splitting the token input sequence into equal-sized partitions: \\( \\mathbb{A} \\) and \\( \\mathbb{B} \\) (line 9). Then, each token in \\( \\mathbb{A} \\) is matched to its most similar counterpart in \\( \\mathbb{B} \\) based on a Matching Function [6, 31] \\( S : (\\mathbb{R}^d, \\mathbb{R}^d) \\to \\mathbb{R} \\) (line 10). A visual representation is provided in Fig. 5, where the \"Partition & Match\" operation is highlighted. These matches are then sorted based on their matching scores \\( (\\stackrel{s_{ij}}{\\rightarrow}) \\), and only the top \\( r_{\\text{fold}} \\) matches are preserved (line 11). At this point, by merging these matches in \\( \\mathbf{M}_{top} \\) from \\( \\mathbb{A} \\) to \\( \\mathbb{B} \\), we obtain two new sets: \\( \\mathbb{A}' \\) and \\( \\mathbb{B}' \\). Namely, \\( \\mathbb{B}' \\) updates \\( \\mathbb{B} \\) by replacing original tokens with matched tokens aggregated using a specific Aggregation Method (Sec. 3.3). \\( \\mathbb{A}' \\), instead, contains the tokens in \\( \\mathbb{A} \\) whose matches were discarded in line 11. Clearly, if the reduction overflow occurs, \\( r_{\\text{fold}} = \\lfloor n' / 2 \\rfloor \\) and thus \\( \\mathbb{A}' = \\emptyset \\) (\"First Fold\" of Fig. 5). The concatenation between" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23617" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.327, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.327, + 0.089, + 0.609, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.089, + 0.904, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.261, + 0.908, + 0.304 + ], + "angle": 0, + "content": "Figure 5. Pipeline of FOLDER. As a plug-and-play module, FOLDER is integrated into the final blocks of the vision backbone (last two here). To deal with reduction overflow, FOLDER automatically executes another FOLD operation when the expected reduction is more than half. The last FOLD, which escapes from reduction overflow, merges tokens according to the remaining reduction numbers." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.322, + 0.482, + 0.488 + ], + "angle": 0, + "content": "\\(\\mathbb{A}^{\\prime}\\) and \\(\\mathbb{B}^{\\prime}\\) represents the input token sequence for the next \"FOLD\" iteration, in which \\(r_{\\mathrm{remain}}\\) tokens are going to be removed. This process continues through successive folds (line 3) until the overflow is no longer encountered. In this case, \\(r_{\\mathrm{fold}} = r_{\\mathrm{remain}}\\) (line 7), and \\(r_{\\mathrm{remain}}\\) for the next iteration will be equal to 0 (line 8). A visual representation of this case is provided in Fig. 5 as the \"Last FOLD\" iteration. Note that in the last FOLD, to fit for the positional encoding in LLM, instead of directly concatenating \\(\\mathbb{A}^{\\prime}\\) and \\(\\mathbb{B}^{\\prime}\\), we rearrange the tokens into original order. More details are provided in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.489, + 0.483, + 0.732 + ], + "angle": 0, + "content": "Discussion. Though simple, the insight contained in FOLDER is non-trivial. Empirically, by analysing the information flow in the vision backbone, we draw the key conclusion that token reduction should be done only in the last blocks, which is overlooked by previous works [6, 27, 31, 65]. Structurally, by generalizing bipartite matching, we specifically design a highly efficient token reduction strategy for MLLMs architecture that allows parallel computing and minimum deployment effort. Functionally, FOLDER focuses on reducing the inherent visual redundancy in the vision backbone, which is task agnostic. Thus FOLDER can superpose on task-oriented/attention-sparsity methods like FastV [9] to enjoy further acceleration (Tab. 1). Besides, FOLDER is among the very few approaches [75] that are verified to be effective in both training and training-free ways (Tab. 1 and Tab. 3), showing its superior scalability." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.736, + 0.223, + 0.753 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.761, + 0.483, + 0.823 + ], + "angle": 0, + "content": "In this section, we present the results obtained by our method. By using FOLDER, we reduce the number of visual tokens to accelerate both training and inference of the Multi-modal Large Language Model (MLLM)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.833, + 0.334, + 0.849 + ], + "angle": 0, + "content": "4.1. Evaluation & Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Evaluation Models. We evaluate FOLDER as training-free plug-in on various MLLMs, including image understanding models LLaVA1.5-7B/13B [50], LLaVA-OV-7B [37]," + }, + { + "type": "code_caption", + "bbox": [ + 0.523, + 0.326, + 0.823, + 0.342 + ], + "angle": 0, + "content": "Algorithm 1: FOLDER applied in one block" + }, + { + "type": "algorithm", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.622 + ], + "angle": 0, + "content": "Input: Token sequence \\(\\mathbf{X}\\in \\mathbb{R}^{n\\times d}\\) reduced num \\(r\\) Output: Reduced token sequence \\(\\mathbf{X}'\\in \\mathbb{R}^{(n - r)\\times d}\\) \n1 \\(r_{\\mathrm{remain}} = r\\) . \n2 \\(\\mathbf{X}' = \\mathbf{X}\\) .. \n3 while \\(r_{\\mathrm{remain}} > 0\\) do \n4 \\(\\begin{array}{rl}{\\mathbf{\\Lambda}_{\\mathbf{X}^{\\prime}},r_{\\mathrm{remain}} = \\mathrm{FOLD}(\\mathbf{X}^{\\prime},r_{\\mathrm{remain}});} & {} \\end{array}\\) \n5 \n6 Function FOLD \\((\\mathbf{X}'\\in \\mathbb{R}^{n'\\times d},r_{\\mathrm{remain}})\\) .. \n7 \\(r_{\\mathrm{fold}} = \\min (\\lfloor n' / 2\\rfloor ,r_{\\mathrm{remain}});\\) \n8 \\(r_{\\mathrm{remain}} = r_{\\mathrm{remain}} - r_{\\mathrm{fold}};\\) \n9 \\(\\{\\mathbb{A},\\mathbb{B}\\} \\Leftarrow \\mathrm{Split~X'}\\) into 2 equal-sized partitions; \n10 M \\(\\Leftarrow\\) Compute \\(\\{(a_i\\xrightarrow{s_{ij}}b_j)\\mid\\) \\(b_{j} = \\arg \\max s_{ij} = S(a_{i},b_{j}),\\forall a_{i}\\in \\mathbb{A}\\} ;\\) \n11 \\(\\mathbf{M}_{top}\\Leftarrow\\) Sort M based on \\(\\stackrel {s_{ij}}{\\to}\\) & keep Top \\(r_\\mathrm{fold}\\) . \n12 \\(\\{\\mathbb{A}',\\mathbb{B}'\\} \\Leftarrow\\) Merge A to B according to Mtop; \n13 return Rearrange(A',B'), \\(r_{\\mathrm{remain}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.656, + 0.907, + 0.779 + ], + "angle": 0, + "content": "Minigpt4v2 [7], VILA1.5-13B [46], WePOINTS1.5-7b [53], VITA1.5-7B [24], Idefics2-8B [35], and Flamingov2-7B [4], multi-vision-tower model MMVP [70], video understanding model Video-LLaVA [45] and point cloud understanding model PointLLMv1.2-7B [76]. For merging strategy ablation, we adopt BLIP [38] for the image captioning task. For MLLM training, we use LLaVA1.5-13B [50]. Training details are provided in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Benchmarks. To quantify the ability of the accelerated MLLMs, we conduct thorough experiments on a wide range of tasks with VLMEvalkit [17], including \\( Y/N \\) tasks like MME [22], HallusionBench [26], POPE [80]; MCQ tasks like MMBench-EN [81], CCBench [81], A-OKVQA [62], ScienceQAIMG [55], SEEDBenchIMG [36], MMMU [82], MMStar [8], Video-MME [23], RealWorld-QA [25], LongVideoBench [73]," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23618" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.904, + 0.426 + ], + "angle": 0, + "content": "
MethodReduct RatioSpeed UpMMBench ENMM StarMMEA-OK VQAHallusion BenchMM MUOCR VQASEED BenchIMGScience QAPOPEAvg
Original-7B0%1×62.832.71338.978.835.632.252.460.268.179.756.9
Pooling50%1.5×59.530.31308.677.135.331.744.859.268.084.055.6
FastV [9]1.3×63.332.41345.178.636.528.753.159.467.881.056.8
Turbo [31]1.5×60.430.91311.277.235.528.035.758.167.583.054.2
Ours1.5×62.432.11338.278.338.334.048.559.568.585.457.4
Pooling66%1.7×57.629.71308.274.232.330.738.757.667.380.753.5
FastV [9]1.5×62.532.01353.277.737.430.052.658.368.279.356.5
Pyramidddrop [75]1.6×61.331.21338.776.938.629.348.656.968.381.956.0
SparseVLM [84]1.6×61.331.11333.077.133.333.138.257.667.482.354.8
Turbo [31]1.7×60.131.51301.678.034.924.734.457.867.285.053.8
Ours1.7×61.430.31350.077.939.231.346.159.768.385.456.7
FastV [9]75%1.6×61.231.01321.676.637.532.050.857.268.177.755.8
Ours+FastV1.7×61.431.21325.578.039.132.648.859.268.682.356.8
Original-13B*0%1×66.630.91371.177.036.134.054.959.468.886.458.3
Pooling50%1.5×64.131.21316.676.235.130.054.757.669.085.757.0
FastV [9]1.3×66.431.11386.375.536.232.654.958.968.485.457.9
Turbo [31]1.5×65.030.61200.278.226.932.747.958.869.786.155.6
Ours1.5×65.431.21383.778.636.434.456.259.170.586.958.8
Pooling66%1.6×62.831.01250.673.529.532.545.654.569.782.154.4
FastV [9]1.5×66.331.21352.275.035.532.354.458.068.383.557.2
Pyramidddrop [75]1.6×65.331.11313.075.835.332.052.558.069.483.556.9
SparseVLM [84]1.6×65.231.11272.677.031.132.745.858.069.583.155.7
Turbo [31]1.6×64.731.21168.376.325.933.347.958.570.585.555.3
Ours1.6×65.831.71366.977.333.835.052.658.870.786.158.1
FastV [9]75%1.6×64.231.71321.374.236.931.353.356.468.182.156.4
Ours+FastV1.8×65.132.31368.677.835.332.055.258.871.185.858.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.429, + 0.907, + 0.459 + ], + "angle": 0, + "content": "Table 1. Results on LLaVA1.5 7B/13B. We highlight the best result for each benchmark under the same reduction ratio. All baseline evaluations are reproduced using official codes on VLMEvalkit [17]. *Retrained to provide a fair comparison in Tab. 3." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.47, + 0.902, + 0.564 + ], + "angle": 0, + "content": "
MethodReduct RatioSpeed UpMMBench ENMM StarMMEHallusion BenchOCR VQAMM MUPOPESEED BenchIMGScience QACC BenchAvg
Original0%1×64.232.11403.646.353.334.286.760.670.018.853.6
Turbo [31]50%1.5×62.832.01352.944.047.634.085.658.669.117.951.9
Ours1.5×63.032.11362.346.250.534.785.759.570.418.652.8
Turbo [31]66%1.7×62.530.71341.846.246.633.384.958.169.718.851.8
Ours1.7×63.631.91368.247.947.234.085.358.770.719.652.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.236, + 0.566, + 0.761, + 0.58 + ], + "angle": 0, + "content": "Table 2. Results on MMVP (MLLM with Multiple Vision-towers (CLIP+DINOv2))." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.646 + ], + "angle": 0, + "content": "MLVU [86]; and VQA tasks like OCRVQA [57], InfoVQA [56], MMBench-Video [19], TempCompass [52]. For BLIP, we evaluate on COCO image captioning [47]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.655, + 0.231, + 0.67 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.677, + 0.484, + 0.723 + ], + "angle": 0, + "content": "In the following experiments, we show the effectiveness of FOLDER on different MLLMs. If not specified, we only apply FOLDER in the last layer of the visual encoder." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.731, + 0.373, + 0.745 + ], + "angle": 0, + "content": "4.2.1. Inference Acceleration on MLLM" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.825 + ], + "angle": 0, + "content": "We evaluate FOLDER's plug-and-play acceleration performance in inference phase on various image understanding MLLMs [7, 24, 46, 50, 53] and multi-vision-tower based model [70]. If not specified, we evaluate MLLM's speed-up based on the time of the first output token." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Single-Vision-Tower MLLM. Tab. 1 compares FOLDER with previous plug-and-play SOTA methods [9, 31, 75, 84] on LLaVA1.5. FOLDER achieves the best performance-speed trade-off, even surpassing original performance with over \\(50\\%\\) reduction on several benchmarks. Compared to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.907, + 0.675 + ], + "angle": 0, + "content": "Turbo [31], which applies uniform merging across the vision encoder, FOLDER delivers superior results under the same reduction ratio, especially on visually dense tasks like OCRVQA. This confirms our key insight: token reduction is most effective when applied only in the final blocks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.856 + ], + "angle": 0, + "content": "While FastV [9] performs similarly to FOLDER on some benchmarks, its limitations (Sec. 2) make it inherently slower and harder to implement. In contrast, FOLDER is a plug-and-play module requiring minimal adjustments to integrate with vision backbones. By reducing visual tokens before interacting with the LLM, FOLDER mitigates architectural complexities. Notably, since FOLDER and FastV tackle token reduction from different angles (visual redundancy vs. task-oriented attention), their combination (50% × 50%) enables further reduction while maintaining performance. Similarly, LLM token reduction methods [75, 84] also face deployment challenges." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.857, + 0.908, + 0.887 + ], + "angle": 0, + "content": "To validate our approach, we replaced the matching function [31] with naive mean pooling. The result in Tab. 1" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23619" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.088, + 0.905, + 0.254 + ], + "angle": 0, + "content": "
TrainTrain ReductInfer ReductSpeed UpMMBench ENMM StarMMEA-OK VQAInfo VQAOCR VQASEED BenchIMGScience QAPOPEAvg
Original0%0%(1,1)×66.630.91371.177.015.554.959.468.886.458.7
Pyramiddrop [75]66%0%(1.5, 1.5)×65.733.01354.279.515.357.461.270.687.959.8
Pyramiddrop [75]66%66%(1.5, 1.5)×66.732.51359.177.515.656.659.070.883.058.6
Ours0%66%(1,1.6)×65.831.71366.977.315.857.658.870.786.159.1
Ours66%0%(1.5,1)×66.632.01407.778.916.657.760.969.388.360.1
Ours66%66%(1.5,1.6)×67.832.51393.479.516.457.961.371.586.960.4
Crossget [65]FT75%(-,1.8)×58.129.11176.968.513.243.652.661.980.051.8
Prumerge [63]FT75%(-,1.7)×58.728.21135.572.413.049.855.366.382.353.6
Pixel-shuffle [10]75%75%(1.7,1.8)×63.828.81343.575.414.854.557.467.386.457.3
Ours0%75%(1,1.8)×64.531.81308.676.716.250.259.071.185.057.8
Ours75%0%(1.7,1)×63.330.71359.579.115.856.260.265.687.858.5
Ours75%75%(1.7,1.8)×65.230.71325.779.515.252.359.567.487.258.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.258, + 0.908, + 0.3 + ], + "angle": 0, + "content": "Table 3. Training with FOLDER. We plug FOLDER into LLaVA1.5-13B during training (Train Reduct>0) and inference (Infer Reduct>0). FT means finetuning on trained model (for LLaVA-Prumerge [63] and Crossget [65]). All results are reproduced/retrained using official codes/checkpoints. We highlight the best result on each benchmark." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.313, + 0.482, + 0.409 + ], + "angle": 0, + "content": "
MethodReduct RatioSpeed UpMMB-V OverallMMB-V PerceptionVideo MME
Original0%1×1.041.0431.8
Turbo65%2.5×1.000.9929.8
Ours2.4×1.041.0531.9
Turbo75%2.8×0.980.9829.0
Ours2.7×1.031.0430.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.182, + 0.414, + 0.393, + 0.427 + ], + "angle": 0, + "content": "Table 4. Results on VideoLLaVA." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.447, + 0.483, + 0.516 + ], + "angle": 0, + "content": "
MethodnframeMMB-V OverallVideo MMELongVideo Bench*Temp CompassMLVU
Original641.5158.756.662.569.6
Ours-75%641.4758.855.060.668.0
Ours-75%1281.4959.557.560.970.2
Ours-75%2561.5359.458.361.370.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.519, + 0.483, + 0.546 + ], + "angle": 0, + "content": "Table 5. Results on LLaVA-OV-7B across more video benchmarks. *Acc of videos with duration > 180s." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.564, + 0.483, + 0.625 + ], + "angle": 0, + "content": "shows the superiority of matching over pooling. More results on MiniGPT-4v2, non-LLaVA models like Idefics2-8B and Flamingov2-7B, and point cloud understanding model PointLLMv1.2-7B are shown in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.627, + 0.483, + 0.793 + ], + "angle": 0, + "content": "Multi-Vision-Tower MLLM. Multi-vision-tower-based models [32, 66, 70] enhance MLLMs' visual ability but aggravate the length issue. This makes our method particularly advantageous in this context. In Tab. 2 we evaluate on MMVP [70] with CLIP [59] and DINOv2 [58] as vision towers. By applying FOLDER, we trade a significant reduction in image token length (from 1152 to 576/384) with an acceptable drop. Note that MMVP is trained by interleaving vision tower features, which may cause positional confusion. Despite that, FOLDER still achieves a satisfying acceleration-performance trade-off." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.795, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Video Understanding MLLM. Video-based MLLMs [30, 40, 45] can generate sequences exceeding 2000 tokens for a few frames, limiting their applicability. As shown in Tab. 4, by applying FOLDER on Video-LLaVA, we can reduce visual tokens up to \\(65\\%\\) without any performance degradation, and up to \\(75\\%\\) with only slight & acceptable drops on some benchmarks. Notably, we achieve \\(2.7 \\times\\) infer" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.313, + 0.918, + 0.495 + ], + "angle": 0, + "content": "
MethodReduct RatioTrainInference
Time(h)Speed UpTime(s)Speed UpMem
Original0%1871×0.3651×81.4G
FastV--0.2751.3×81.4G
Turbo50%1391.34×0.2501.5×56.6G
Ours1411.33×0.2501.5×56.6G
FastV--0.2551.4×81.4G
Turbo66%1231.52×0.2301.6×48.6G
Ours1241.51×0.2301.6×48.7G
FastV--0.2251.6×81.4G
Turbo75%1121.67×0.2051.8×44.6G
Ours1131.65×0.2051.8×44.7G
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.496, + 0.907, + 0.566 + ], + "angle": 0, + "content": "Table 6. Acceleration For Training & Inference. We evaluate training (A100-80G GPU hours) and inference acceleration (on 1st-token) on llava1.5-13B. The memory is evaluated by its peak value on H20 with batch-size 40. Note that FastV [9] can only be used during inference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.907, + 0.673 + ], + "angle": 0, + "content": "ence speedup with \\(75\\%\\) token reduction, still outperforming Turbo with \\(66\\%\\) reduction ratio. Furthermore, Tab. 5 shows that on LLaVA-OV-7B, with a \\(75\\%\\) reduction ratio within each frame, using only half of the visual tokens \\((128\\times 25\\%)\\) compared to the original setup, we achieve even better performance, especially on long video understanding." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.683, + 0.791, + 0.699 + ], + "angle": 0, + "content": "4.2.2. Training Acceleration on MLLM" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.703, + 0.907, + 0.87 + ], + "angle": 0, + "content": "As a universal plug-and-play module for MLLM, FOLDER can also be integrated into MLLM pretraining and SFT. Results on LLaVA1.5-13B are in Tab. 3. Surprisingly, when incorporating FOLDER with a \\(66\\%\\) reduction ratio, LLaVA-1.5-13B exhibits comprehensive improvements across all benchmarks, even including visually dense tasks such as OCRVQA (\\(\\uparrow 3\\%\\)) and InfoVQA (\\(\\uparrow 0.9\\%\\)). This suggests high-frequency noise in visual sequences can be mitigated by merging similar visual tokens. Moreover, [9] highlights attention sparsity in LLMs' visual tokens, reinforcing the need to shorten visual sequences for better focus." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Furthermore, training on reduced sequences maintains flexibility during inference, as the original sequence can still" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23620" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.107, + 0.086, + 0.468, + 0.23 + ], + "angle": 0, + "content": "
MethodReduct RatioThroughput im/sSpeed UpCIDErB@4
Original0%22.11×133.339.7
Uniform40.21.82×128.837.9
Last-360%38.41.74×129.538.2
Last-237.91.71×129.438.2
Last-137.51.71×132.239.0
Uniform47.42.14×127.237.5
Last-370%44.22.00×128.438.2
Last-243.71.98×130.138.4
Last-143.41.96×131.038.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.234, + 0.483, + 0.276 + ], + "angle": 0, + "content": "Table 7. Ablation on Merging Position. We evaluate BLIP [38] on captioning task. Last- \\( n \\) refers to uniform reduction in last \\( n \\) blocks, uniform refers to uniform reduction in every block." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.295, + 0.498, + 0.387 + ], + "angle": 0, + "content": "
MethodMMBPOPEMMEGQASciQA
VILA1.5-13b [46]74.384.91433.764.578.9
+FOLDER-50%72.884.01445.362.078.0
VITA1.5-7b [24]80.989.11728.947.895.8
+FOLDER-50%78.188.81688.350.894.1
WePOINTS1.5-7b [53]84.086.81616.059.295.0
+FOLDER-50%83.487.01642.659.194.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.103, + 0.391, + 0.471, + 0.405 + ], + "angle": 0, + "content": "Table 8. FOLDER training-free plug-in for more MLLMs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.422, + 0.483, + 0.483 + ], + "angle": 0, + "content": "be used, leading to improved results on visually demanding tasks. Similar to masked augmentation [28, 43], FOLDER acts as a regularizer but goes beyond random dropping by accelerating both training and inference." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.485, + 0.483, + 0.577 + ], + "angle": 0, + "content": "Compared to Pyramidddrop [75], FOLDER achieves better performance on most benchmarks. It also outperforms finetuning-based methods [63, 65] even without additional training. By reducing tokens before the LLM, FOLDER serves as a flexible alternative to training-phase modules like pixel-shuffle, enhancing both efficiency and scalability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.578, + 0.483, + 0.73 + ], + "angle": 0, + "content": "Speed & Memory Evaluation. In Tab. 6, we evaluate FOLDER's acceleration performance on LLaVA1.5-13B during training and inference. By reducing \\(66\\%\\) of tokens, FOLDER achieves \\(1.5 \\times\\) faster training on \\(8 \\times \\mathrm{A}100 - 80\\mathrm{G}\\) and \\(1.6 \\times\\) speedup during inference on H20-96G. Unlike FastV, which prunes tokens within the LLM, FOLDER removes them beforehand, significantly lowering peak memory usage. With a batch size of 40 on H20-96G, it saves \\(40\\%\\) memory while maintaining a \\(66\\%\\) reduction ratio, enabling nearly \\(3 \\times\\) larger batches—especially beneficial for training." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.744, + 0.466, + 0.759 + ], + "angle": 0, + "content": "4.2.3. Simplified FOLDER for Inference Acceleration" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "As illustrated in Fig. 5, FOLDER is inserted between ATTN and FFN, requiring vision encoder modifications. Since the best results are achieved by reducing tokens only in the last block (Sec. 3.4), a simpler approach is inserting FOLDER at the output. Tab. 8 shows results on 3 SOTA MLLMs (VILA1.5 [46], WePOINTS1.5 [53], VITA1.5 [24]) with \\(50\\%\\) reduction ratio and \\(1.4\\times\\) acceleration, while maintaining or surpassing original performance on several benchmarks. More comparisons are in the appendix." + }, + { + "type": "table", + "bbox": [ + 0.565, + 0.083, + 0.855, + 0.203 + ], + "angle": 0, + "content": "
MethodReduct RatioB@4CIDErSPICE
Original0%133.339.723.8
Avg132.039.023.6
Drop50%130.038.723.3
Weighted132.239.123.6
Avg131.038.923.5
Drop60%128.938.223.1
Weighted131.539.123.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.207, + 0.905, + 0.263 + ], + "angle": 0, + "content": "Table 9. Ablation Study on Aggregation Operation. On BLIP [38] captioning task, we test the performance of 3 aggregation operations discussed in Sec. 3.3. For fair comparison, we adopt the aggregation only in the last block." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.276, + 0.667, + 0.292 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.299, + 0.905, + 0.344 + ], + "angle": 0, + "content": "We conduct ablation studies on BLIP [38] for image captioning to validate the merging position and aggregation methods in Sec. 3. More ablations are in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.345, + 0.906, + 0.496 + ], + "angle": 0, + "content": "Merging Position. Preliminary ViT tests in (Fig. 2, Fig. 3) show that due to propagation effects (Sec. 3.2) and diminishing reduction impact (Sec. 3.1), merging more tokens in the last blocks is preferable. Based on this, we test four strategies: uniform merging across all blocks and in the last \\(1/2/3\\) blocks. Tab. 7 presents the results which confirm that merging in the last block yields the best performance. While the vision encoder's acceleration slightly affects BLIP, in MLLMs, where LLM processing dominates \\((10 - 15 \\times\\) inferior), merging position has minimal impact on speed-up." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.497, + 0.906, + 0.602 + ], + "angle": 0, + "content": "Aggregation Method. In Sec. 3.3, we explore three distinct aggregation methods: Dropping, Average, and Weighted Average based on vector norms. As illustrated in Tab. 9, both averaging methods outperform direct dropping, as directly dropping tokens causes greater information loss. To reduce module complexity, we used averaging in all MLLM experiments, as the differences are minimal." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.618, + 0.634, + 0.634 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.906, + 0.81 + ], + "angle": 0, + "content": "We introduce FOLDER, a plug-and-play module for efficient visual token reduction in MLLMs. Through empirical analysis of the token reduction process, we develop a strategy that performs aggressive token reduction in the vision backbone's final layers while preserving key information. Our method achieves over \\(70\\%\\) visual token reduction with \\(1.8 \\times\\) inference speedup and \\(1.65 \\times\\) training acceleration. FOLDER maintains or improves performance across tasks including video understanding, demonstrating its effectiveness in making MLLMs more computationally efficient without capability loss." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. Part of this work was funded by Hi!PARIS Center on Data Analytics and Artificial Intelligence, by the European Union's Horizon Europe Research and Innovation Programme under grant agreement No. 101120237 (ELIAS), and by the French National Research Agency (ANR) in the framework of the JCJC project \"BANERA\" under Grant ANR-24-CE23-4369." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23621" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.185, + 0.482, + 0.212 + ], + "angle": 0, + "content": "[2] Anthropic. The claude 3 model family: Opus, sonnet, haiku, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.215, + 0.483, + 0.282 + ], + "angle": 0, + "content": "[3] Kirolos Ataallah, Xiaoqian Shen, Eslam Abdelrahman, Essam Sleiman, Mingchen Zhuge, Jian Ding, Deyao Zhu, Jürgen Schmidhuber, and Mohamed Elhoseiny. Goldfish: Vision-language understanding of arbitrarily long videos, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.284, + 0.483, + 0.367 + ], + "angle": 0, + "content": "[4] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. Openflamingo: An opensource framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.368, + 0.482, + 0.424 + ], + "angle": 0, + "content": "[5] Adam Block, Dylan J Foster, Akshay Krishnamurthy, Max Simchowitz, and Cyril Zhang. Butterfly effects of sgd noise: Error amplification in behavior cloning and autoregression. arXiv preprint arXiv:2310.11428, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.425, + 0.483, + 0.479 + ], + "angle": 0, + "content": "[6] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461, 2022. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.483, + 0.564 + ], + "angle": 0, + "content": "[7] Jun Chen, Deyao Zhu, Xiaoqian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023.2,5,6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.483, + 0.633 + ], + "angle": 0, + "content": "[8] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.635, + 0.483, + 0.703 + ], + "angle": 0, + "content": "[9] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. arXiv preprint arXiv:2403.06764, 2024. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.705, + 0.483, + 0.787 + ], + "angle": 0, + "content": "[10] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhang-wei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. Science China Information Sciences, 67(12):220101, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[11] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[12] Kanzhi Cheng, Wenpo Song, Zheng Ma, Wenhao Zhu, Zixuan Zhu, and Jianbing Zhang. Beyond generic: Enhancing" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.906, + 0.148 + ], + "angle": 0, + "content": "image captioning with real-world knowledge using vision-language pre-training model. In Proceedings of the 31st ACM International Conference on Multimedia, pages 5038-5047, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.151, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[13] Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.181, + 0.905, + 0.222 + ], + "angle": 0, + "content": "[14] Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.224, + 0.905, + 0.28 + ], + "angle": 0, + "content": "[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.282, + 0.905, + 0.323 + ], + "angle": 0, + "content": "[16] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.326, + 0.905, + 0.409 + ], + "angle": 0, + "content": "[17] Haodong Duan, Junming Yang, Yuxuan Qiao, Xinyu Fang, Lin Chen, Yuan Liu, Xiaoyi Dong, Yuhang Zang, Pan Zhang, Jiaqi Wang, et al. Vlmevalkit: An open-source toolkit for evaluating large multi-modality models. In Proceedings of the 32nd ACM international conference on multimedia, pages 11198-11201, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.411, + 0.905, + 0.452 + ], + "angle": 0, + "content": "[18] Carl Eckart and Gale Young. The approximation of one matrix by another of lower rank. Psychometrika, 1(3):211-218, 1936. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.454, + 0.905, + 0.511 + ], + "angle": 0, + "content": "[19] Xinyu Fang, Kangrui Mao, Haodong Duan, Xiangyu Zhao, Yining Li, Dahua Lin, and Kai Chen. Mmbench-video: A long-form multi-shot benchmark for holistic video understanding. arXiv preprint arXiv:2406.14515, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.513, + 0.905, + 0.567 + ], + "angle": 0, + "content": "[20] Zhiyuan Fang, Jianfeng Wang, Xiaowei Hu, Lijuan Wang, Yezhou Yang, and Zicheng Liu. Compressing visual-linguistic model via knowledge distillation. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.571, + 0.905, + 0.625 + ], + "angle": 0, + "content": "[21] Elias Frantar, Saleh Ashkboos, Torsten Hoefler, and Dan Alistarh. Gptq: Accurate post-training quantization for generative pre-trained transformers. arXiv preprint arXiv:2210.17323, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.628, + 0.905, + 0.697 + ], + "angle": 0, + "content": "[22] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.7, + 0.905, + 0.769 + ], + "angle": 0, + "content": "[23] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.771, + 0.905, + 0.84 + ], + "angle": 0, + "content": "[24] Chaoyou Fu, Haojia Lin, Xiong Wang, Yi-Fan Zhang, Yunhang Shen, Xiaoyu Liu, Yangze Li, Zuwei Long, Heting Gao, Ke Li, et al. Vita-1.5: Towards gpt-4o level real-time vision and speech interaction. arXiv preprint arXiv:2501.01957, 2025. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.843, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[25] Grok. Grok1.5 homepage. https://x.ai/blog/grok-1.5v, 2024. Accessed: 2024-11-14. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[26] Tianrui Guan, Fuxiao Liu, Xiyang Wu, Ruiqi Xian, Zongxia Li, Xiaoyu Liu, Xijun Wang, Lichang Chen, Furong Huang," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23622" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.161 + ], + "angle": 0, + "content": "Yaser Yacoob, et al. Hallusionbench: an advanced diagnostic suite for entangled language hallucination and visual illusion in large vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14375-14385, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.483, + 0.232 + ], + "angle": 0, + "content": "[27] Yuhang Han, Xuyang Liu, Pengxiang Ding, Donglin Wang, Honggang Chen, Qingsen Yan, and Siteng Huang. Rethinking token reduction in mllms: Towards a unified paradigm for training-free acceleration. arXiv preprint arXiv:2411.17686, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.483, + 0.3 + ], + "angle": 0, + "content": "[28] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.482, + 0.37 + ], + "angle": 0, + "content": "[29] Wenxuan Huang, Zijie Zhai, Yunhang Shen, Shaoshen Cao, Fei Zhao, Xiangfeng Xu, Zheyu Ye, and Shaohui Lin. Dynamic-llava: Efficient multimodal large language models via dynamic vision-language context sparsification. arXiv preprint arXiv:2412.00876, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[30] Peng Jin, Ryuichi Takanobu, Wancai Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13700-13710, 2024. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.456, + 0.482, + 0.536 + ], + "angle": 0, + "content": "[31] Chen Ju, Haicheng Wang, Haozhe Cheng, Xu Chen, Zhonghua Zhai, Weilin Huang, Jinsong Lan, Shuai Xiao, and Bo Zheng. Turbo: Informativity-driven acceleration plug-in for vision-language large models. In European Conference on Computer Vision, pages 436-455. Springer, 2025. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.538, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[32] Oguzhan Fatih Kar, Alessio Tonioni, Petra Poklukar, Achin Kulshrestha, Amir Zamir, and Federico Tombari. Brave: Broadening the visual encoding of vision-language models. arXiv preprint arXiv:2404.07204, 2024. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.678 + ], + "angle": 0, + "content": "[33] Zaid Khan, Vijay Kumar BG, Samuel Schulter, Xiang Yu, Yun Fu, and Manmohan Chandraker. Q: How to specialize large vision-language models to data-scarce vqa tasks? a: Self-train on unlabeled images! In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15005-15015, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.734 + ], + "angle": 0, + "content": "[34] Zhenglun Kong, Peiyan Dong, Xiaolong Ma, Xin Meng, Mengshu Sun, Wei Niu, Xuan Shen, Geng Yuan, Bin Ren, Minghai Qin, Hao Tang, and Yanzhi Wang. Spvit: Enabling faster vision transformers via soft token pruning, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[35] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? Advances in Neural Information Processing Systems, 37:87874-87907, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[36] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[37] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[38] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.231 + ], + "angle": 0, + "content": "[39] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.232, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[40] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.905, + 0.33 + ], + "angle": 0, + "content": "[41] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and performant baseline for vision and language, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.331, + 0.905, + 0.384 + ], + "angle": 0, + "content": "[42] Wentong Li, Yuqian Yuan, Jian Liu, Dongqi Tang, Song Wang, Jie Qin, Jianke Zhu, and Lei Zhang. Tokenpacker: Efficient visual projector for multimodal llm. arXiv preprint arXiv:2407.02392, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.385, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[43] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pre-training via masking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 23390-23400, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.455, + 0.905, + 0.509 + ], + "angle": 0, + "content": "[44] Youwei Liang, Chongjian Ge, Zhan Tong, Yibing Song, Jue Wang, and Pengtao Xie. Not all patches are what you need: Expediting vision transformers via token reorganizations, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[45] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.905, + 0.635 + ], + "angle": 0, + "content": "[46] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 26689-26699, 2024. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.636, + 0.905, + 0.72 + ], + "angle": 0, + "content": "[47] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.721, + 0.905, + 0.748 + ], + "angle": 0, + "content": "[48] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[49] Hao Liu, Matei Zaharia, and Pieter Abbeel. Ring attention with blockwise transformers for near-infinite context, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[50] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[51] Ting Liu, Liangtao Shi, Richang Hong, Yue Hu, Quanjun Yin, and Linfeng Zhang. Multi-stage vision token dropping: Towards efficient multimodal large language model. arXiv preprint arXiv:2411.10803, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23623" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[52] Yuanxin Liu, Shicheng Li, Yi Liu, Yuxiang Wang, Shuhuai Ren, Lei Li, Sishuo Chen, Xu Sun, and Lu Hou. Tempcompass: Do video llms really understand videos? arXiv preprint arXiv:2403.00476, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.151, + 0.482, + 0.207 + ], + "angle": 0, + "content": "[53] Yuan Liu, Le Tian, Xiao Zhou, Xinyu Gao, Kavio Yu, Yang Yu, and Jie Zhou. Points1. 5: Building a vision-language model towards real world applications. arXiv preprint arXiv:2412.08443, 2024. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.209, + 0.482, + 0.264 + ], + "angle": 0, + "content": "[54] Zhijian Liu, Ligeng Zhu, Baifeng Shi, Zhuoyang Zhang, Yuming Lou, Shang Yang, Haocheng Xi, Shiyi Cao, Yuxian Gu, Dacheng Li, et al. Nvila: Efficient frontier visual language models. arXiv preprint arXiv:2412.04468, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.267, + 0.482, + 0.349 + ], + "angle": 0, + "content": "[55] Pan Lu, Swaroop Mishra, Tony Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In The 36th Conference on Neural Information Processing Systems (NeurIPS), 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.353, + 0.482, + 0.408 + ], + "angle": 0, + "content": "[56] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1697-1706, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.411, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[57] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering by reading text in images. In ICDAR, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.482, + 0.524 + ], + "angle": 0, + "content": "[58] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.527, + 0.482, + 0.609 + ], + "angle": 0, + "content": "[59] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.613, + 0.482, + 0.667 + ], + "angle": 0, + "content": "[60] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.671, + 0.482, + 0.712 + ], + "angle": 0, + "content": "[61] Yossi Rubner, Carlo Tomasi, and Leonidas J Guibas. The earth mover's distance as a metric for image retrieval. International journal of computer vision, 40:99-121, 2000. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.716, + 0.482, + 0.784 + ], + "angle": 0, + "content": "[62] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge. In European conference on computer vision, pages 146–162. Springer, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.787, + 0.482, + 0.841 + ], + "angle": 0, + "content": "[63] Yuzhang Shang, Mu Cai, Bingxin Xu, Yong Jae Lee, and Yan Yan. Llava-prumerge: Adaptive token reduction for efficient large multimodal models. arXiv preprint arXiv:2403.15388, 2024. 2, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[64] Dachuan Shi, Chaofan Tao, Ying Jin, Zhendong Yang, Chun Yuan, and Jiaqi Wang. Upop: Unified and progressive pruning for compressing vision-language transformers. arXiv preprint arXiv:2301.13741, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[65] Dachuan Shi, Chaofan Tao, Anyi Rao, Zhendong Yang, Chun Yuan, and Jiaqi Wang. Crossget: Cross-guided ensemble of tokens for accelerating vision-language transformers. arXiv preprint arXiv:2305.17455, 2023. 2, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[66] Min Shi, Fuxiao Liu, Shihao Wang, Shijia Liao, Subhashree Radhakrishnan, De-An Huang, Hongxu Yin, Karan Sapra, Yaser Yacoob, Humphrey Shi, et al. Eagle: Exploring the design space for multimodal llms with mixture of encoders. arXiv preprint arXiv:2408.15998, 2024. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[67] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. FLAVA: A foundational language and vision alignment model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[68] Hao Tan and Mohit Bansal. LXMERT: Learning cross-modality encoder representations from transformers. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, EMNLP-IJCNLP, November 3-7, 2019, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[69] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.907, + 0.516 + ], + "angle": 0, + "content": "[70] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9568-9578, 2024. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.586 + ], + "angle": 0, + "content": "[71] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.589, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[72] Tiannan Wang, Wangchunshu Zhou, Yan Zeng, and Xinsong Zhang. Efficientvlm: Fast and accurate vision-language models via knowledge distillation and modal-adaptive pruning. arXiv preprint arXiv:2210.07795, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[73] Haoning Wu, Dongxu Li, Bei Chen, and Junnan Li. Longvideobench: A benchmark for long-context interleaved video-language understanding. Advances in Neural Information Processing Systems, 37:28828-28857, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[74] Guangxuan Xiao, Ji Lin, Mickael Seznec, Hao Wu, Julien Demouth, and Song Han. Smoothquant: Accurate and efficient post-training quantization for large language models. In International Conference on Machine Learning, pages 38087-38099. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.905, + 0.855 + ], + "angle": 0, + "content": "[75] Long Xing, Qidong Huang, Xiaoyi Dong, Jiajie Lu, Pan Zhang, Yuhang Zang, Yuhang Cao, Conghui He, Jiaqi Wang, Feng Wu, et al. Pyramiddrop: Accelerating your large vision-language models via pyramid visual redundancy reduction. arXiv preprint arXiv:2410.17247, 2024. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[76] Runsen Xu, Xiaolong Wang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Pointllm: Empowering large language models to understand point clouds. In European" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23624" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "Conference on Computer Vision, pages 131-147. Springer, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[77] Senqiao Yang, Yukang Chen, Zhuotao Tian, Chengyao Wang, Jingyao Li, Bei Yu, and Jiaya Jia. Visionzip: Longer is better but not necessary in vision language models. arXiv preprint arXiv:2412.04467, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[78] Xubing Ye, Yukang Gan, Yixiao Ge, Xiao-Ping Zhang, and Yansong Tang. Atp-llava: Adaptive token pruning for large vision language models. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 24972–24982, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[79] Xubing Ye, Yukang Gan, Xiaoke Huang, Yixiao Ge, and Yansong Tang. Voco-llama: Towards vision compression with large language models. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 29836–29846, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.319, + 0.482, + 0.387 + ], + "angle": 0, + "content": "[80] Li Yifan, Du Yifan, Zhou Kun, Wang Jinpeng, Xin Zhao Wayne, and Wen Ji-Rong. Evaluating object hallucination in large vision-language models. In The 2023 Conference on Empirical Methods in Natural Language Processing, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.39, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[81] Liu Yuan, Duan Haodong, Zhang Yuanhan, Li Bo, Zhang Songyang, Zhao Wangbo, Yuan Yike, Wang Jiaqi, He Conghui, Liu Ziwei, Chen Kai, and Lin Dahua. Mmbench: Is your multi-modal model an all-around player? arXiv:2307.06281, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.461, + 0.482, + 0.57 + ], + "angle": 0, + "content": "[82] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Ren-liang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of CVPR, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.572, + 0.482, + 0.613 + ], + "angle": 0, + "content": "[83] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.615, + 0.482, + 0.684 + ], + "angle": 0, + "content": "[84] Yuan Zhang, Chun-Kai Fan, Junpeng Ma, Wenzhao Zheng, Tao Huang, Kuan Cheng, Denis Gudovskiy, Tomoyuki Okuno, Yohei Nakata, Kurt Keutzer, et al. Sparsevm: Visual token sparsification for efficient vision-language model inference. arXiv preprint arXiv:2410.04417, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.686, + 0.482, + 0.727 + ], + "angle": 0, + "content": "[85] Yiwu Zhong, Zhuoming Liu, Yin Li, and Liwei Wang. Aim: Adaptive inference of multi-modal llms via token merging and pruning. arXiv preprint arXiv:2412.03248, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.728, + 0.482, + 0.797 + ], + "angle": 0, + "content": "[86] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. MLvu: A comprehensive benchmark for multitask long video understanding. arXiv e-prints, pages arXiv-2406, 2024. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.797 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23625" + } + ] +] \ No newline at end of file diff --git a/2025/FOLDER_ Accelerating Multi-Modal Large Language Models with Enhanced Performance/7a17b26d-c0b2-4285-9105-7de16ee42381_origin.pdf b/2025/FOLDER_ Accelerating Multi-Modal Large Language Models with Enhanced Performance/7a17b26d-c0b2-4285-9105-7de16ee42381_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4540d241b365d95c151fdd6e80f7f2b8363c2d87 --- /dev/null +++ b/2025/FOLDER_ Accelerating Multi-Modal Large Language Models with Enhanced Performance/7a17b26d-c0b2-4285-9105-7de16ee42381_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8db7052a1d57799a4f38988f5ca3ca8a44aed6edd9cd6c7a3f2906b9ee7d097a +size 1124897 diff --git a/2025/FOLDER_ Accelerating Multi-Modal Large Language Models with Enhanced Performance/full.md b/2025/FOLDER_ Accelerating Multi-Modal Large Language Models with Enhanced Performance/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e4c6f65ce2e695ef4cb8b36a19ac9e78c52c80c0 --- /dev/null +++ b/2025/FOLDER_ Accelerating Multi-Modal Large Language Models with Enhanced Performance/full.md @@ -0,0 +1,348 @@ +# FOLDER: Accelerating Multi-modal Large Language Models with Enhanced Performance + +Haicheng Wang $^{1,2,4*}$ , Zhemeng Yu $^{1,2*}$ , Gabriele Spadaro $^{2,3}$ , Chen Ju $^{4}$ , Victor Quétu $^{2}$ , Shuai Xiao $^{4\boxtimes}$ , Enzo Tartaglione $^{2\boxtimes}$ + +$^{1}$ Paris Elite Institute of Technology, Shanghai Jiao Tong University $^{3}$ University of Turin +$^{2}$ LTCI, Télecom Paris, Institut Polytechnique de Paris $^{4}$ Taobao & Tmall Group of Alibaba + +anakinSkywalker@sjtu.edu.cn, shuai.xsh@alibaba-inc.com, enzo.tartaglione@telecom-paris.fr + +# Abstract + +Recently, Multi-modal Large Language Models (MLLMs) have shown remarkable effectiveness for multi-modal tasks due to their ability of cross-modal understanding. However, processing long sequences of visual tokens extracted from visual backbones poses challenges for deployment in real-time applications. To address this issue, we introduce FOLDER, a simple yet effective plug-and-play module designed to reduce the length of the visual token sequence, mitigating computational and memory demands during both training and inference. Through a comprehensive analysis of the token reduction process in the vision encoder, we analyze the information loss introduced by different reduction strategies and develop FOLDER to preserve key information while removing visual redundancy. We show the effectiveness of FOLDER by integrating it into the visual backbone of various MLLMs, significantly accelerating the inference phase. Furthermore, we evaluate its utility as a training accelerator or even performance booster for MLLMs. FOLDER achieves comparable or even better performance than the original models, while dramatically reducing complexity by removing up to $70\%$ of visual tokens. Our code is available at FOLDER Repository. + +# 1. Introduction + +Multi-modal Large Language Models (MLLMs) excel in tasks like image captioning [12] and visual question answering [33] by learning joint visual-text representations. State-of-the-art models such as GPT-4V [1] and Claude 3 [2] show strong multimodal understanding ability. However, these models face concrete challenges in terms of computational efficiency, particularly when processing visual inputs. Modern MLLMs typically generate long sequences of im + +![](images/b8d374ca51938da28fb9a98ee201711701d9c1ee0f9c56bfa8c0278ab60e80dc.jpg) + +![](images/309f1efceee5e179e46f7b793e4da6cf6bf2a14af7c63e298fa05ad387252967.jpg) +Figure 1. FOLDER as Accelerator & Booster. As a plug-and-play module, FOLDER can be used in both training and inference, with considerable acceleration and even a performance boost. + +age tokens through their visual backbones. In some application scenarios, this problem becomes even more evident: recent advancements in high-resolution-supporting MLLMs [11, 71] might involve over 2000 visual tokens. Similarly, challenges also occur in multi-visual-expert architectures [32, 66, 70] that leverage multiple visual encoders. These issues are even more severe in video understanding [30, 40, 45] due to the extra temporal complexity. The resulting token explosion hinders real-time deployment because of the quadratic cost of the attention mechanism. Several studies have been conducted to address this issue. Existing approaches can be roughly divided into two + +categories. The first category focuses on training phase solutions, such as Q-Former [39], pixel-shuffle [10] and pooling-based techniques [7]. However, these methods often suffer from performance degradation and exhibit limited scalability, due to their custom design for specific architectures. The second category aims to design plug-and-play token reduction modules for the inference phase [9, 27, 31, 65, 75, 84]. Specifically, [9, 75, 84] reduce visual tokens inside LLM. Though effective, these methods are often hard to implement, require adaptation for existing LLM acceleration modules [13], and prove unsuitable for multi-round conversations, thus greatly limiting their usage. Meanwhile, [27, 31, 65] identify the inherent redundancy in visual tokens, but fail to quantify the information loss during the token reduction process. Thus their reduction strategies are often arbitrary (uniform reduction for each layer), leading to sub-optimal outcomes. + +To find a reduction strategy that can be universally applicable and best preserve the performance, we propose a preliminary investigation into the vision backbone, to answer a fundamental question: Where does information loss originate? To model it, we identify three key factors that impact information loss during token removal: Reduction Impact, Propagation Effect, and Aggregation Method. Through a comprehensive empirical study of these factors, we devise a strategy that enhances the preservation of information while significantly reducing redundancy in the vision backbone. FOLDER is consequently designed to implement this strategy with minimal computational overhead. + +Our plug-and-play solution can be easily integrated into the visual backbones of various MLLMs, effectively reducing the number of visual tokens while incurring negligible information loss (Fig. 1). We first incorporate FOLDER during MLLMs inference, reducing over $60\%$ of visual tokens and achieving comparable, or even gains on performance across multiple tasks. Furthermore, when integrated into MLLMs pre-training, our module can be used as a training accelerator, as well as an effective regularization term, leading to remarkable performance improvements across all benchmarks, with a reduction ratio up to $70\%$ . Our method is on-the-field proven to be a flexible, nearly lossless, and highly efficient token reduction strategy, offering a dual-purpose solution to the visual token reduction challenge. Our contributions can be summarized as follows. + +- We conduct an in-depth analysis for information decay during token reduction in the vision backbone, identifying and quantifying key factors involved (Sec. 3.1, Sec. 3.2 and Sec. 3.3). These insights offer a clear understanding of how token reduction impacts the information flow. +- Leveraging the above analysis, we develop a simple yet effective plug-and-play visual token reduction module for MLLMs (Sec. 3.4). This novel token reduction strategy aggressively reduces the number of tokens only in the + +last blocks of the visual encoder. On various models and benchmarks, FOLDER speeds up the inference $1.7 - 2.4 \times$ with negligible loss or even improvement (Sec. 4.2.1). + +- We show that FOLDER can also be used in MLLM training (Sec. 4.2.2), to speed-up while boosting performance, serving as a dual-solution for both training and inference. + +# 2. Related Works + +Multi-modal Large Language Models. MLLMs excel at understanding multi-modal information [1, 3, 7, 39, 41, 45, 48, 67-69, 71, 83], with common architectures comprising i.) a Visual Backbone, ii.) a Vision-Language Connector, and iii.) a Pre-trained LLM. However, integrating visual data significantly increases computational cost due to the large number of visual tokens. For example, LLaVA1.5 [50] processes 576 tokens for a $336 \times 336$ image, while VideoLLaVA [45] exceeds 2000 tokens with just 8 frames at $224 \times 224$ . Recent efforts to enhance MLLMs with multiple vision towers [32, 66, 70] further exacerbate this issue. MLLMs Acceleration. To accelerate large models, many methods focus on system-level optimizations, such as FlashAttention [14], vLLM [13], and RingAttention [49]. Model-centric techniques like knowledge distillation [20, 72], quantization [21, 74], and pruning [60, 64, 72] also reduce size and cost. However, these approaches overlook redundant data, which unnecessarily prolongs sequences. + +Token Reduction in MLLMs. Some preliminary studies are conducted on token reduction for Vision Transformers [6, 34, 44, 85]. In the context of MLLMs, several token reduction methods, including Q-Former [39], resampler [11], pooling [7] and some other methods [29, 42, 54, 78, 79] are proposed to reduce visual tokens during the training process. However, these approaches often suffer from performance degradation and a lack of scalability. + +Meanwhile, some studies try to handle the token reduction problem during inference. In particular, ToMe [6] proposes a training-free token merging strategy for unimodal image classification. Based on that, Turbo [31] proposes an improved merging strategy considering both mutual redundancy and semantic value, which serves as a plug-in module in MLLMs. Recently, FastV [9] recognizes the visual attention sparsity in LLM, thus pruning visual tokens inside the LLM based on the task-oriented attention importance. Though effective, these methods have clear drawbacks. For ToMe/Turbo-like methods [6, 27, 31, 63, 65, 77] that reduce tokens in the vision encoder, most of them adopt uniform progressive merging across layers, which is arbitrary and inflexible. While FastV-like methods [9, 51, 75, 84] that cut tokens inside LLM layers mainly suffer from 2 problems: i.) They often require interactions with instruction tokens, making it unsuitable for multi-round/long-term visual QA. ii.) Since these approaches are inserted in LLM and often require specific hyperparameters, it is difficult to adapt + +![](images/f2312d72b13a11838068d5e60f83ba4379c9e0cae5892cb0c0c600a59cd120f8.jpg) +Figure 2. Minimum Number of Tokens $k$ to Preserve Energy $t_E$ Across Blocks. We evaluate on 3 different $t_E$ for every block. + +for complex LLM modules [14] and hard to deploy. Besides, few approaches have been verified to be effective in both training and training-free inference ways. To address the above issues, we seek to find an effective, information-guided, and user-friendly token reduction method that can be easily applied in both training and inference. + +# 3. Method + +In this section, we establish FOLDER by first answering three key questions for token reduction in vision backbone: i.) how many tokens to reduce (Sec. 3.1), ii.) in which block (Sec. 3.2) and iii.) through which aggregation method (Sec. 3.3). Then, we present our method in Sec. 3.4. + +Empirical Setup. To answer these three crucial questions, we define a common empirical setup. Specifically, we use a pre-trained 12-block ViT-B [16] model and conduct the empirical experiments on ImageNet-1k [15]. Results on more diverse setups are provided in the appendix. + +# 3.1. Token Reduction Impact + +To explore the link between token reduction and information loss, we use Singular Value Decomposition (SVD) to track the energy of the reduced token sequence. More precisely, given a token sequence $\mathbf{X} \in \mathbb{R}^{n \times d}$ (with $n$ being the number of tokens and $d$ their size), we can decompose it by applying SVD on $\mathbf{X}^T$ : + +$$ +\mathbf {X} ^ {T} = \mathbf {U} \boldsymbol {\Sigma} \mathbf {V} ^ {T}, \tag {1} +$$ + +where $\mathbf{U} \in \mathbb{R}^{d \times d}$ and $\mathbf{V} \in \mathbb{R}^{n \times n}$ are two orthogonal matrices and $\pmb{\Sigma} \in \mathbb{R}^{d \times n}$ is a diagonal matrix, with the singular values $\sigma_{i} = \pmb{\Sigma}_{ii}$ representing the variances (energy) in the new compositional directions. In this way, SVD yields the optimal energy concentration that can be ordered by their contribution to the total variance (energy) [18]. Based on this, we can estimate how much variance we preserve by + +![](images/62e69817cfadc5278ff9eb4633366b3fa4035d059732be2bb28f9968c9028b17.jpg) +Figure 3. EMD Distance Between Reduced and Original Output Distributions under 3 Reduction Ratios. We compare the EMD distance by exerting token reduction on different blocks. + +only taking the largest $k$ singular values, as: + +$$ +E (k) = \frac {\sum_ {i = 1} ^ {k} \sigma_ {i}}{\sum_ {i = 1} ^ {n} \sigma_ {i}}, \quad \sigma_ {i} \geq \sigma_ {i + 1} \quad \forall i. \tag {2} +$$ + +By definition, SVD provides an upper bound on preserved energy when keeping $k$ tokens. Thus, by setting a threshold $t_E$ , we can estimate the required $k$ for each block as: + +$$ +k = \min \left\{k \mid E (k) \geq t _ {E} \right\}. \tag {3} +$$ + +Observations. With a threshold $t_E$ , $k$ varies across blocks. We monitor $k$ under this threshold and find, as shown in Fig. 2, that later blocks require far fewer tokens than earlier ones to maintain the same energy. This highlights the potential for block-adaptive token reduction, suggesting the first and last blocks as key candidates. + +# 3.2. Token Propagation Effect + +While token reduction impact provides an upper bound on information preservation, it overlooks inter-block dependencies and therefore may not adequately reflect the effects on subsequent blocks. In transformer architectures, errors from earlier blocks can propagate and amplify through successive layers [5], potentially affecting final performance. + +To analyze this propagation effect, we examine how token reduction at different blocks affects the final output distribution. Thus, by comparing the original output token distribution (without reduction) and the one obtained by operating token reduction only in one block $b$ , we can show the presence of such a propagation effect. + +To measure this distortion, we consider output tokens from the original model $\mathbf{Y} \in \mathbb{R}^{n \times d}$ and the ones obtained when $n - k$ tokens are reduced in the block $b$ : $\tilde{\mathbf{Y}}_b \in \mathbb{R}^{k \times d}$ . Then, we define the empirical distributions $P_Y$ and $P_{\tilde{\mathbf{Y}}_b}$ supported on $\mathbf{Y}$ and $\tilde{\mathbf{Y}}_b$ respectively, and we evaluate Earth Mover's Distance (EMD) [61] between them: + +![](images/f86748aa632ab4874bd8b5a9a22d644632e297e60eeec252628922aa0b48b986.jpg) +Figure 4. EMD & Accuracy Comparison for Aggregation Methods. We compare three aggregation methods under different reduction ratios on both EMD and Accuracy. + +$$ +\operatorname {E M D} \left(P _ {Y}, P _ {\tilde {Y} _ {b}}\right) = \min _ {\gamma} \langle \gamma , \mathbf {M} \rangle_ {F}, \tag {4} +$$ + +where $\mathbf{M} = (d_{ij})_{n\times k} = (\| y_i - \tilde{y}_{bj}\| _2)_{n\times k}$ denotes the metric cost matrix, with $y_{i}$ and $\tilde{y}_{bj}$ being points in the support of distributions $P_Y$ and $P_{\tilde{Y}_b}$ respectively. $\gamma = (\gamma_{ij})_{n\times k}$ represents the optimal transport matrix, which satisfies the non-negative constraint $\gamma_{ij}\geq 0,\forall i,j$ and the marginal constraints $\sum_{j = 1}^{k}\gamma_{ij} = P_Y(y_i)$ and $\sum_{i = 1}^{n}\gamma_{ij} = P_{\tilde{Y}_b}(\tilde{y}_{bj})$ + +Observations. Due to the propagation effect, token reduction at different blocks causes varying levels of output distortion. We monitor EMD (Eq. (4)) when reducing the same number of tokens per block and find, as shown in Fig.3, that early-block reductions lead to significantly higher EMD values. This confirms that distortion amplifies across layers, making early blocks unsuitable for token reduction, contrary to Fig.2. Conversely, with a massive $75\%$ reduction, EMD remains low in the last layers. This suggests that token reduction is most effective at the network's end, even with a high reduction rate. Results on BLIP [38] (Tab. 7) and appendix further support this conclusion. + +# 3.3. Token Aggregation Method + +In Sec. 3.1, we use SVD to estimate the upper bound of information preservation. However, its computational complexity prevents direct application for token reduction. Following prior works [6, 31], we decompose the process into two stages: (i) Token Matching, grouping tokens based on a matching function, and (ii) Token Aggregation, merging grouped tokens. While matching functions have been explored [6, 31], aggregation remains understudied. Formally, given a set of tokens $\{\mathbf{x}_1,\mathbf{x}_2,\dots ,\mathbf{x}_m\}$ to be merged into $\mathbf{y}$ , we define the aggregation operation as follows: + +$$ +\mathbf {y} = \sum_ {i = 1} ^ {m} \alpha_ {i} \mathbf {x} _ {i}, \text {s . t .} \sum_ {i = 1} ^ {m} \alpha_ {i} = 1. \tag {5} +$$ + +Based on Eq. (5), we formulate three different token aggregation approaches: i.) Average Merging, with $\alpha_{i} = \frac{1}{m},\forall i$ ii.) Weighted Merging, with $\alpha_{i} = \frac{\|x_{i}\|_{2}}{\sum_{j = 1}^{m}\|x_{j}\|_{2}},\forall i;$ iii.) Direct Dropping, with $\alpha_{i} = 0,\forall i\neq i_{max},\alpha_{i_{max}} = 1$ where $i_{max} = \underset {i}{\operatorname{argmax}}\| x_i\| _2$ . Here, the norm $\| \cdot \| _2$ is used as the importance score for tokens. + +Observations. Similarly to Sec. 3.2, we compare aggregation methods using EMD. As shown in Tab. 9 and Fig. 4, both merging methods yield lower discrepancy than Direct Dropping. Given its comparable performance to Weighted Merging with greater simplicity, we select Average Merging as our preferred method. Additionally, Fig. 4 shows that a lower EMD indeed correlates with higher accuracy. Further supporting results are in Tab. 9 and appendix. + +# 3.4.FOLDER + +Motivation. Based on the above empirical analyses, we deduce that: i.) Token reduction should take place in the last blocks, where aggressive reduction is permitted. ii.) Merging is better than dropping as an aggregation method. Therefore, our goal is to design an algorithm that can reduce as many tokens as possible in one block. Previous works [63] mainly leverage clustering for token merging, which is slow and unsuitable for parallel computing. To avoid such issues, we refer to bipartite matching [6], which is however limited by a maximum reduction ratio of $1/2$ . Thus, to aggressively reduce tokens in one block, we propose FOLDER, an advanced bipartite matching strategy that can reduce an arbitrary number of tokens. + +Strategy. As illustrated in Fig. 5, our strategy focuses on the last blocks (or the last) in the vision backbone. In some cases, the target reduction $r$ exceeds half of the total tokens $\lfloor n / 2 \rfloor$ in a block, a scenario we call reduction overflow. To handle this, FOLDER automatically applies an iterative "FOLD" operation. As detailed in Alg. 1, if reduction overflow occurs in an iteration $(r_{\text{remain}} > \lfloor n' / 2 \rfloor$ in line 7), we reduce tokens by half $(r_{\text{fold}} = \lfloor n' / 2 \rfloor)$ . The process begins by splitting the token input sequence into equal-sized partitions: $\mathbb{A}$ and $\mathbb{B}$ (line 9). Then, each token in $\mathbb{A}$ is matched to its most similar counterpart in $\mathbb{B}$ based on a Matching Function [6, 31] $S : (\mathbb{R}^d, \mathbb{R}^d) \to \mathbb{R}$ (line 10). A visual representation is provided in Fig. 5, where the "Partition & Match" operation is highlighted. These matches are then sorted based on their matching scores $(\stackrel{s_{ij}}{\rightarrow})$ , and only the top $r_{\text{fold}}$ matches are preserved (line 11). At this point, by merging these matches in $\mathbf{M}_{top}$ from $\mathbb{A}$ to $\mathbb{B}$ , we obtain two new sets: $\mathbb{A}'$ and $\mathbb{B}'$ . Namely, $\mathbb{B}'$ updates $\mathbb{B}$ by replacing original tokens with matched tokens aggregated using a specific Aggregation Method (Sec. 3.3). $\mathbb{A}'$ , instead, contains the tokens in $\mathbb{A}$ whose matches were discarded in line 11. Clearly, if the reduction overflow occurs, $r_{\text{fold}} = \lfloor n' / 2 \rfloor$ and thus $\mathbb{A}' = \emptyset$ ("First Fold" of Fig. 5). The concatenation between + +![](images/7e5a3493389f1845588d7d7081c9d90be78a61c80906d5e10ae784b2b260648c.jpg) +Figure 5. Pipeline of FOLDER. As a plug-and-play module, FOLDER is integrated into the final blocks of the vision backbone (last two here). To deal with reduction overflow, FOLDER automatically executes another FOLD operation when the expected reduction is more than half. The last FOLD, which escapes from reduction overflow, merges tokens according to the remaining reduction numbers. + +![](images/8139d0587137e2031d0ddd89611f00975e58341de69122e62abe804805f2fd7d.jpg) + +![](images/2522f93e5cd5212032ccd98adc056bea429ed78212b25b78108d0003c3036331.jpg) + +$\mathbb{A}^{\prime}$ and $\mathbb{B}^{\prime}$ represents the input token sequence for the next "FOLD" iteration, in which $r_{\mathrm{remain}}$ tokens are going to be removed. This process continues through successive folds (line 3) until the overflow is no longer encountered. In this case, $r_{\mathrm{fold}} = r_{\mathrm{remain}}$ (line 7), and $r_{\mathrm{remain}}$ for the next iteration will be equal to 0 (line 8). A visual representation of this case is provided in Fig. 5 as the "Last FOLD" iteration. Note that in the last FOLD, to fit for the positional encoding in LLM, instead of directly concatenating $\mathbb{A}^{\prime}$ and $\mathbb{B}^{\prime}$ , we rearrange the tokens into original order. More details are provided in the appendix. + +Discussion. Though simple, the insight contained in FOLDER is non-trivial. Empirically, by analysing the information flow in the vision backbone, we draw the key conclusion that token reduction should be done only in the last blocks, which is overlooked by previous works [6, 27, 31, 65]. Structurally, by generalizing bipartite matching, we specifically design a highly efficient token reduction strategy for MLLMs architecture that allows parallel computing and minimum deployment effort. Functionally, FOLDER focuses on reducing the inherent visual redundancy in the vision backbone, which is task agnostic. Thus FOLDER can superpose on task-oriented/attention-sparsity methods like FastV [9] to enjoy further acceleration (Tab. 1). Besides, FOLDER is among the very few approaches [75] that are verified to be effective in both training and training-free ways (Tab. 1 and Tab. 3), showing its superior scalability. + +# 4. Experiments + +In this section, we present the results obtained by our method. By using FOLDER, we reduce the number of visual tokens to accelerate both training and inference of the Multi-modal Large Language Model (MLLM). + +# 4.1. Evaluation & Benchmarks + +Evaluation Models. We evaluate FOLDER as training-free plug-in on various MLLMs, including image understanding models LLaVA1.5-7B/13B [50], LLaVA-OV-7B [37], + +Algorithm 1: FOLDER applied in one block +Input: Token sequence $\mathbf{X}\in \mathbb{R}^{n\times d}$ reduced num $r$ Output: Reduced token sequence $\mathbf{X}'\in \mathbb{R}^{(n - r)\times d}$ +1 $r_{\mathrm{remain}} = r$ . +2 $\mathbf{X}' = \mathbf{X}$ .. +3 while $r_{\mathrm{remain}} > 0$ do +4 $\begin{array}{rl}{\mathbf{\Lambda}_{\mathbf{X}^{\prime}},r_{\mathrm{remain}} = \mathrm{FOLD}(\mathbf{X}^{\prime},r_{\mathrm{remain}});} & {} \end{array}$ +5 +6 Function FOLD $(\mathbf{X}'\in \mathbb{R}^{n'\times d},r_{\mathrm{remain}})$ .. +7 $r_{\mathrm{fold}} = \min (\lfloor n' / 2\rfloor ,r_{\mathrm{remain}});$ +8 $r_{\mathrm{remain}} = r_{\mathrm{remain}} - r_{\mathrm{fold}};$ +9 $\{\mathbb{A},\mathbb{B}\} \Leftarrow \mathrm{Split~X'}$ into 2 equal-sized partitions; +10 M $\Leftarrow$ Compute $\{(a_i\xrightarrow{s_{ij}}b_j)\mid$ $b_{j} = \arg \max s_{ij} = S(a_{i},b_{j}),\forall a_{i}\in \mathbb{A}\} ;$ +11 $\mathbf{M}_{top}\Leftarrow$ Sort M based on $\stackrel {s_{ij}}{\to}$ & keep Top $r_\mathrm{fold}$ . +12 $\{\mathbb{A}',\mathbb{B}'\} \Leftarrow$ Merge A to B according to Mtop; +13 return Rearrange(A',B'), $r_{\mathrm{remain}}$ + +Minigpt4v2 [7], VILA1.5-13B [46], WePOINTS1.5-7b [53], VITA1.5-7B [24], Idefics2-8B [35], and Flamingov2-7B [4], multi-vision-tower model MMVP [70], video understanding model Video-LLaVA [45] and point cloud understanding model PointLLMv1.2-7B [76]. For merging strategy ablation, we adopt BLIP [38] for the image captioning task. For MLLM training, we use LLaVA1.5-13B [50]. Training details are provided in the appendix. + +Benchmarks. To quantify the ability of the accelerated MLLMs, we conduct thorough experiments on a wide range of tasks with VLMEvalkit [17], including $Y/N$ tasks like MME [22], HallusionBench [26], POPE [80]; MCQ tasks like MMBench-EN [81], CCBench [81], A-OKVQA [62], ScienceQAIMG [55], SEEDBenchIMG [36], MMMU [82], MMStar [8], Video-MME [23], RealWorld-QA [25], LongVideoBench [73], + +
MethodReduct RatioSpeed UpMMBench ENMM StarMMEA-OK VQAHallusion BenchMM MUOCR VQASEED BenchIMGScience QAPOPEAvg
Original-7B0%1×62.832.71338.978.835.632.252.460.268.179.756.9
Pooling50%1.5×59.530.31308.677.135.331.744.859.268.084.055.6
FastV [9]1.3×63.332.41345.178.636.528.753.159.467.881.056.8
Turbo [31]1.5×60.430.91311.277.235.528.035.758.167.583.054.2
Ours1.5×62.432.11338.278.338.334.048.559.568.585.457.4
Pooling66%1.7×57.629.71308.274.232.330.738.757.667.380.753.5
FastV [9]1.5×62.532.01353.277.737.430.052.658.368.279.356.5
Pyramidddrop [75]1.6×61.331.21338.776.938.629.348.656.968.381.956.0
SparseVLM [84]1.6×61.331.11333.077.133.333.138.257.667.482.354.8
Turbo [31]1.7×60.131.51301.678.034.924.734.457.867.285.053.8
Ours1.7×61.430.31350.077.939.231.346.159.768.385.456.7
FastV [9]75%1.6×61.231.01321.676.637.532.050.857.268.177.755.8
Ours+FastV1.7×61.431.21325.578.039.132.648.859.268.682.356.8
Original-13B*0%1×66.630.91371.177.036.134.054.959.468.886.458.3
Pooling50%1.5×64.131.21316.676.235.130.054.757.669.085.757.0
FastV [9]1.3×66.431.11386.375.536.232.654.958.968.485.457.9
Turbo [31]1.5×65.030.61200.278.226.932.747.958.869.786.155.6
Ours1.5×65.431.21383.778.636.434.456.259.170.586.958.8
Pooling66%1.6×62.831.01250.673.529.532.545.654.569.782.154.4
FastV [9]1.5×66.331.21352.275.035.532.354.458.068.383.557.2
Pyramidddrop [75]1.6×65.331.11313.075.835.332.052.558.069.483.556.9
SparseVLM [84]1.6×65.231.11272.677.031.132.745.858.069.583.155.7
Turbo [31]1.6×64.731.21168.376.325.933.347.958.570.585.555.3
Ours1.6×65.831.71366.977.333.835.052.658.870.786.158.1
FastV [9]75%1.6×64.231.71321.374.236.931.353.356.468.182.156.4
Ours+FastV1.8×65.132.31368.677.835.332.055.258.871.185.858.2
+ +Table 1. Results on LLaVA1.5 7B/13B. We highlight the best result for each benchmark under the same reduction ratio. All baseline evaluations are reproduced using official codes on VLMEvalkit [17]. *Retrained to provide a fair comparison in Tab. 3. + +
MethodReduct RatioSpeed UpMMBench ENMM StarMMEHallusion BenchOCR VQAMM MUPOPESEED BenchIMGScience QACC BenchAvg
Original0%1×64.232.11403.646.353.334.286.760.670.018.853.6
Turbo [31]50%1.5×62.832.01352.944.047.634.085.658.669.117.951.9
Ours1.5×63.032.11362.346.250.534.785.759.570.418.652.8
Turbo [31]66%1.7×62.530.71341.846.246.633.384.958.169.718.851.8
Ours1.7×63.631.91368.247.947.234.085.358.770.719.652.8
+ +Table 2. Results on MMVP (MLLM with Multiple Vision-towers (CLIP+DINOv2)). + +MLVU [86]; and VQA tasks like OCRVQA [57], InfoVQA [56], MMBench-Video [19], TempCompass [52]. For BLIP, we evaluate on COCO image captioning [47]. + +# 4.2. Main Results + +In the following experiments, we show the effectiveness of FOLDER on different MLLMs. If not specified, we only apply FOLDER in the last layer of the visual encoder. + +# 4.2.1. Inference Acceleration on MLLM + +We evaluate FOLDER's plug-and-play acceleration performance in inference phase on various image understanding MLLMs [7, 24, 46, 50, 53] and multi-vision-tower based model [70]. If not specified, we evaluate MLLM's speed-up based on the time of the first output token. + +Single-Vision-Tower MLLM. Tab. 1 compares FOLDER with previous plug-and-play SOTA methods [9, 31, 75, 84] on LLaVA1.5. FOLDER achieves the best performance-speed trade-off, even surpassing original performance with over $50\%$ reduction on several benchmarks. Compared to + +Turbo [31], which applies uniform merging across the vision encoder, FOLDER delivers superior results under the same reduction ratio, especially on visually dense tasks like OCRVQA. This confirms our key insight: token reduction is most effective when applied only in the final blocks. + +While FastV [9] performs similarly to FOLDER on some benchmarks, its limitations (Sec. 2) make it inherently slower and harder to implement. In contrast, FOLDER is a plug-and-play module requiring minimal adjustments to integrate with vision backbones. By reducing visual tokens before interacting with the LLM, FOLDER mitigates architectural complexities. Notably, since FOLDER and FastV tackle token reduction from different angles (visual redundancy vs. task-oriented attention), their combination (50% × 50%) enables further reduction while maintaining performance. Similarly, LLM token reduction methods [75, 84] also face deployment challenges. + +To validate our approach, we replaced the matching function [31] with naive mean pooling. The result in Tab. 1 + +
TrainTrain ReductInfer ReductSpeed UpMMBench ENMM StarMMEA-OK VQAInfo VQAOCR VQASEED BenchIMGScience QAPOPEAvg
Original0%0%(1,1)×66.630.91371.177.015.554.959.468.886.458.7
Pyramiddrop [75]66%0%(1.5, 1.5)×65.733.01354.279.515.357.461.270.687.959.8
Pyramiddrop [75]66%66%(1.5, 1.5)×66.732.51359.177.515.656.659.070.883.058.6
Ours0%66%(1,1.6)×65.831.71366.977.315.857.658.870.786.159.1
Ours66%0%(1.5,1)×66.632.01407.778.916.657.760.969.388.360.1
Ours66%66%(1.5,1.6)×67.832.51393.479.516.457.961.371.586.960.4
Crossget [65]FT75%(-,1.8)×58.129.11176.968.513.243.652.661.980.051.8
Prumerge [63]FT75%(-,1.7)×58.728.21135.572.413.049.855.366.382.353.6
Pixel-shuffle [10]75%75%(1.7,1.8)×63.828.81343.575.414.854.557.467.386.457.3
Ours0%75%(1,1.8)×64.531.81308.676.716.250.259.071.185.057.8
Ours75%0%(1.7,1)×63.330.71359.579.115.856.260.265.687.858.5
Ours75%75%(1.7,1.8)×65.230.71325.779.515.252.359.567.487.258.1
+ +Table 3. Training with FOLDER. We plug FOLDER into LLaVA1.5-13B during training (Train Reduct>0) and inference (Infer Reduct>0). FT means finetuning on trained model (for LLaVA-Prumerge [63] and Crossget [65]). All results are reproduced/retrained using official codes/checkpoints. We highlight the best result on each benchmark. + +
MethodReduct RatioSpeed UpMMB-V OverallMMB-V PerceptionVideo MME
Original0%1×1.041.0431.8
Turbo65%2.5×1.000.9929.8
Ours2.4×1.041.0531.9
Turbo75%2.8×0.980.9829.0
Ours2.7×1.031.0430.7
+ +Table 4. Results on VideoLLaVA. + +
MethodnframeMMB-V OverallVideo MMELongVideo Bench*Temp CompassMLVU
Original641.5158.756.662.569.6
Ours-75%641.4758.855.060.668.0
Ours-75%1281.4959.557.560.970.2
Ours-75%2561.5359.458.361.370.7
+ +shows the superiority of matching over pooling. More results on MiniGPT-4v2, non-LLaVA models like Idefics2-8B and Flamingov2-7B, and point cloud understanding model PointLLMv1.2-7B are shown in the appendix. + +Multi-Vision-Tower MLLM. Multi-vision-tower-based models [32, 66, 70] enhance MLLMs' visual ability but aggravate the length issue. This makes our method particularly advantageous in this context. In Tab. 2 we evaluate on MMVP [70] with CLIP [59] and DINOv2 [58] as vision towers. By applying FOLDER, we trade a significant reduction in image token length (from 1152 to 576/384) with an acceptable drop. Note that MMVP is trained by interleaving vision tower features, which may cause positional confusion. Despite that, FOLDER still achieves a satisfying acceleration-performance trade-off. + +Video Understanding MLLM. Video-based MLLMs [30, 40, 45] can generate sequences exceeding 2000 tokens for a few frames, limiting their applicability. As shown in Tab. 4, by applying FOLDER on Video-LLaVA, we can reduce visual tokens up to $65\%$ without any performance degradation, and up to $75\%$ with only slight & acceptable drops on some benchmarks. Notably, we achieve $2.7 \times$ infer + +Table 5. Results on LLaVA-OV-7B across more video benchmarks. *Acc of videos with duration > 180s. + +
MethodReduct RatioTrainInference
Time(h)Speed UpTime(s)Speed UpMem
Original0%1871×0.3651×81.4G
FastV--0.2751.3×81.4G
Turbo50%1391.34×0.2501.5×56.6G
Ours1411.33×0.2501.5×56.6G
FastV--0.2551.4×81.4G
Turbo66%1231.52×0.2301.6×48.6G
Ours1241.51×0.2301.6×48.7G
FastV--0.2251.6×81.4G
Turbo75%1121.67×0.2051.8×44.6G
Ours1131.65×0.2051.8×44.7G
+ +Table 6. Acceleration For Training & Inference. We evaluate training (A100-80G GPU hours) and inference acceleration (on 1st-token) on llava1.5-13B. The memory is evaluated by its peak value on H20 with batch-size 40. Note that FastV [9] can only be used during inference. + +ence speedup with $75\%$ token reduction, still outperforming Turbo with $66\%$ reduction ratio. Furthermore, Tab. 5 shows that on LLaVA-OV-7B, with a $75\%$ reduction ratio within each frame, using only half of the visual tokens $(128\times 25\%)$ compared to the original setup, we achieve even better performance, especially on long video understanding. + +# 4.2.2. Training Acceleration on MLLM + +As a universal plug-and-play module for MLLM, FOLDER can also be integrated into MLLM pretraining and SFT. Results on LLaVA1.5-13B are in Tab. 3. Surprisingly, when incorporating FOLDER with a $66\%$ reduction ratio, LLaVA-1.5-13B exhibits comprehensive improvements across all benchmarks, even including visually dense tasks such as OCRVQA ( $\uparrow 3\%$ ) and InfoVQA ( $\uparrow 0.9\%$ ). This suggests high-frequency noise in visual sequences can be mitigated by merging similar visual tokens. Moreover, [9] highlights attention sparsity in LLMs' visual tokens, reinforcing the need to shorten visual sequences for better focus. + +Furthermore, training on reduced sequences maintains flexibility during inference, as the original sequence can still + +
MethodReduct RatioThroughput im/sSpeed UpCIDErB@4
Original0%22.11×133.339.7
Uniform40.21.82×128.837.9
Last-360%38.41.74×129.538.2
Last-237.91.71×129.438.2
Last-137.51.71×132.239.0
Uniform47.42.14×127.237.5
Last-370%44.22.00×128.438.2
Last-243.71.98×130.138.4
Last-143.41.96×131.038.9
+ +Table 7. Ablation on Merging Position. We evaluate BLIP [38] on captioning task. Last- $n$ refers to uniform reduction in last $n$ blocks, uniform refers to uniform reduction in every block. + +
MethodMMBPOPEMMEGQASciQA
VILA1.5-13b [46]74.384.91433.764.578.9
+FOLDER-50%72.884.01445.362.078.0
VITA1.5-7b [24]80.989.11728.947.895.8
+FOLDER-50%78.188.81688.350.894.1
WePOINTS1.5-7b [53]84.086.81616.059.295.0
+FOLDER-50%83.487.01642.659.194.3
+ +be used, leading to improved results on visually demanding tasks. Similar to masked augmentation [28, 43], FOLDER acts as a regularizer but goes beyond random dropping by accelerating both training and inference. + +Compared to Pyramidddrop [75], FOLDER achieves better performance on most benchmarks. It also outperforms finetuning-based methods [63, 65] even without additional training. By reducing tokens before the LLM, FOLDER serves as a flexible alternative to training-phase modules like pixel-shuffle, enhancing both efficiency and scalability. + +Speed & Memory Evaluation. In Tab. 6, we evaluate FOLDER's acceleration performance on LLaVA1.5-13B during training and inference. By reducing $66\%$ of tokens, FOLDER achieves $1.5 \times$ faster training on $8 \times \mathrm{A}100 - 80\mathrm{G}$ and $1.6 \times$ speedup during inference on H20-96G. Unlike FastV, which prunes tokens within the LLM, FOLDER removes them beforehand, significantly lowering peak memory usage. With a batch size of 40 on H20-96G, it saves $40\%$ memory while maintaining a $66\%$ reduction ratio, enabling nearly $3 \times$ larger batches—especially beneficial for training. + +# 4.2.3. Simplified FOLDER for Inference Acceleration + +As illustrated in Fig. 5, FOLDER is inserted between ATTN and FFN, requiring vision encoder modifications. Since the best results are achieved by reducing tokens only in the last block (Sec. 3.4), a simpler approach is inserting FOLDER at the output. Tab. 8 shows results on 3 SOTA MLLMs (VILA1.5 [46], WePOINTS1.5 [53], VITA1.5 [24]) with $50\%$ reduction ratio and $1.4\times$ acceleration, while maintaining or surpassing original performance on several benchmarks. More comparisons are in the appendix. + +Table 8. FOLDER training-free plug-in for more MLLMs. + +
MethodReduct RatioB@4CIDErSPICE
Original0%133.339.723.8
Avg132.039.023.6
Drop50%130.038.723.3
Weighted132.239.123.6
Avg131.038.923.5
Drop60%128.938.223.1
Weighted131.539.123.4
+ +Table 9. Ablation Study on Aggregation Operation. On BLIP [38] captioning task, we test the performance of 3 aggregation operations discussed in Sec. 3.3. For fair comparison, we adopt the aggregation only in the last block. + +# 4.3. Ablation Study + +We conduct ablation studies on BLIP [38] for image captioning to validate the merging position and aggregation methods in Sec. 3. More ablations are in the appendix. + +Merging Position. Preliminary ViT tests in (Fig. 2, Fig. 3) show that due to propagation effects (Sec. 3.2) and diminishing reduction impact (Sec. 3.1), merging more tokens in the last blocks is preferable. Based on this, we test four strategies: uniform merging across all blocks and in the last $1/2/3$ blocks. Tab. 7 presents the results which confirm that merging in the last block yields the best performance. While the vision encoder's acceleration slightly affects BLIP, in MLLMs, where LLM processing dominates $(10 - 15 \times$ inferior), merging position has minimal impact on speed-up. + +Aggregation Method. In Sec. 3.3, we explore three distinct aggregation methods: Dropping, Average, and Weighted Average based on vector norms. As illustrated in Tab. 9, both averaging methods outperform direct dropping, as directly dropping tokens causes greater information loss. To reduce module complexity, we used averaging in all MLLM experiments, as the differences are minimal. + +# 5. Conclusion + +We introduce FOLDER, a plug-and-play module for efficient visual token reduction in MLLMs. Through empirical analysis of the token reduction process, we develop a strategy that performs aggressive token reduction in the vision backbone's final layers while preserving key information. Our method achieves over $70\%$ visual token reduction with $1.8 \times$ inference speedup and $1.65 \times$ training acceleration. FOLDER maintains or improves performance across tasks including video understanding, demonstrating its effectiveness in making MLLMs more computationally efficient without capability loss. + +Acknowledgements. Part of this work was funded by Hi!PARIS Center on Data Analytics and Artificial Intelligence, by the European Union's Horizon Europe Research and Innovation Programme under grant agreement No. 101120237 (ELIAS), and by the French National Research Agency (ANR) in the framework of the JCJC project "BANERA" under Grant ANR-24-CE23-4369. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 2 +[2] Anthropic. The claude 3 model family: Opus, sonnet, haiku, 2024. 1 +[3] Kirolos Ataallah, Xiaoqian Shen, Eslam Abdelrahman, Essam Sleiman, Mingchen Zhuge, Jian Ding, Deyao Zhu, Jürgen Schmidhuber, and Mohamed Elhoseiny. Goldfish: Vision-language understanding of arbitrarily long videos, 2024. 2 +[4] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. Openflamingo: An opensource framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. 5 +[5] Adam Block, Dylan J Foster, Akshay Krishnamurthy, Max Simchowitz, and Cyril Zhang. Butterfly effects of sgd noise: Error amplification in behavior cloning and autoregression. arXiv preprint arXiv:2310.11428, 2023. 3 +[6] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461, 2022. 2, 4, 5 +[7] Jun Chen, Deyao Zhu, Xiaoqian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023.2,5,6 +[8] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024.5 +[9] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. arXiv preprint arXiv:2403.06764, 2024. 2, 5, 6, 7 +[10] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhang-wei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. Science China Information Sciences, 67(12):220101, 2024. 2, 7 +[11] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024. 1, 2 +[12] Kanzhi Cheng, Wenpo Song, Zheng Ma, Wenhao Zhu, Zixuan Zhu, and Jianbing Zhang. Beyond generic: Enhancing + +image captioning with real-world knowledge using vision-language pre-training model. In Proceedings of the 31st ACM International Conference on Multimedia, pages 5038-5047, 2023. 1 +[13] Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning, 2023. 2 +[14] Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness, 2022. 2, 3 +[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 3 +[16] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3 +[17] Haodong Duan, Junming Yang, Yuxuan Qiao, Xinyu Fang, Lin Chen, Yuan Liu, Xiaoyi Dong, Yuhang Zang, Pan Zhang, Jiaqi Wang, et al. Vlmevalkit: An open-source toolkit for evaluating large multi-modality models. In Proceedings of the 32nd ACM international conference on multimedia, pages 11198-11201, 2024. 5, 6 +[18] Carl Eckart and Gale Young. The approximation of one matrix by another of lower rank. Psychometrika, 1(3):211-218, 1936. 3 +[19] Xinyu Fang, Kangrui Mao, Haodong Duan, Xiangyu Zhao, Yining Li, Dahua Lin, and Kai Chen. Mmbench-video: A long-form multi-shot benchmark for holistic video understanding. arXiv preprint arXiv:2406.14515, 2024. 6 +[20] Zhiyuan Fang, Jianfeng Wang, Xiaowei Hu, Lijuan Wang, Yezhou Yang, and Zicheng Liu. Compressing visual-linguistic model via knowledge distillation. In ICCV, 2021. 2 +[21] Elias Frantar, Saleh Ashkboos, Torsten Hoefler, and Dan Alistarh. Gptq: Accurate post-training quantization for generative pre-trained transformers. arXiv preprint arXiv:2210.17323, 2022. 2 +[22] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 5 +[23] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 5 +[24] Chaoyou Fu, Haojia Lin, Xiong Wang, Yi-Fan Zhang, Yunhang Shen, Xiaoyu Liu, Yangze Li, Zuwei Long, Heting Gao, Ke Li, et al. Vita-1.5: Towards gpt-4o level real-time vision and speech interaction. arXiv preprint arXiv:2501.01957, 2025. 5, 6, 8 +[25] Grok. Grok1.5 homepage. https://x.ai/blog/grok-1.5v, 2024. Accessed: 2024-11-14. 5 +[26] Tianrui Guan, Fuxiao Liu, Xiyang Wu, Ruiqi Xian, Zongxia Li, Xiaoyu Liu, Xijun Wang, Lichang Chen, Furong Huang, + +Yaser Yacoob, et al. Hallusionbench: an advanced diagnostic suite for entangled language hallucination and visual illusion in large vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14375-14385, 2024. 5 +[27] Yuhang Han, Xuyang Liu, Pengxiang Ding, Donglin Wang, Honggang Chen, Qingsen Yan, and Siteng Huang. Rethinking token reduction in mllms: Towards a unified paradigm for training-free acceleration. arXiv preprint arXiv:2411.17686, 2024. 2, 5 +[28] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 8 +[29] Wenxuan Huang, Zijie Zhai, Yunhang Shen, Shaoshen Cao, Fei Zhao, Xiangfeng Xu, Zheyu Ye, and Shaohui Lin. Dynamic-llava: Efficient multimodal large language models via dynamic vision-language context sparsification. arXiv preprint arXiv:2412.00876, 2024. 2 +[30] Peng Jin, Ryuichi Takanobu, Wancai Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13700-13710, 2024. 1, 7 +[31] Chen Ju, Haicheng Wang, Haozhe Cheng, Xu Chen, Zhonghua Zhai, Weilin Huang, Jinsong Lan, Shuai Xiao, and Bo Zheng. Turbo: Informativity-driven acceleration plug-in for vision-language large models. In European Conference on Computer Vision, pages 436-455. Springer, 2025. 2, 4, 5, 6 +[32] Oguzhan Fatih Kar, Alessio Tonioni, Petra Poklukar, Achin Kulshrestha, Amir Zamir, and Federico Tombari. Brave: Broadening the visual encoding of vision-language models. arXiv preprint arXiv:2404.07204, 2024. 1, 2, 7 +[33] Zaid Khan, Vijay Kumar BG, Samuel Schulter, Xiang Yu, Yun Fu, and Manmohan Chandraker. Q: How to specialize large vision-language models to data-scarce vqa tasks? a: Self-train on unlabeled images! In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15005-15015, 2023. 1 +[34] Zhenglun Kong, Peiyan Dong, Xiaolong Ma, Xin Meng, Mengshu Sun, Wei Niu, Xuan Shen, Geng Yuan, Bin Ren, Minghai Qin, Hao Tang, and Yanzhi Wang. Spvit: Enabling faster vision transformers via soft token pruning, 2022. 2 +[35] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? Advances in Neural Information Processing Systems, 37:87874-87907, 2024. 5 +[36] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 5 +[37] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 5 + +[38] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 4, 5, 8 +[39] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 2 +[40] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 1, 7 +[41] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and performant baseline for vision and language, 2019. 2 +[42] Wentong Li, Yuqian Yuan, Jian Liu, Dongqi Tang, Song Wang, Jie Qin, Jianke Zhu, and Lei Zhang. Tokenpacker: Efficient visual projector for multimodal llm. arXiv preprint arXiv:2407.02392, 2024. 2 +[43] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pre-training via masking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 23390-23400, 2023. 8 +[44] Youwei Liang, Chongjian Ge, Zhan Tong, Yibing Song, Jue Wang, and Pengtao Xie. Not all patches are what you need: Expediting vision transformers via token reorganizations, 2022. 2 +[45] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 1, 2, 5, 7 +[46] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 26689-26699, 2024. 5, 6, 8 +[47] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6 +[48] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2 +[49] Hao Liu, Matei Zaharia, and Pieter Abbeel. Ring attention with blockwise transformers for near-infinite context, 2023. 2 +[50] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 2, 5, 6 +[51] Ting Liu, Liangtao Shi, Richang Hong, Yue Hu, Quanjun Yin, and Linfeng Zhang. Multi-stage vision token dropping: Towards efficient multimodal large language model. arXiv preprint arXiv:2411.10803, 2024. 2 + +[52] Yuanxin Liu, Shicheng Li, Yi Liu, Yuxiang Wang, Shuhuai Ren, Lei Li, Sishuo Chen, Xu Sun, and Lu Hou. Tempcompass: Do video llms really understand videos? arXiv preprint arXiv:2403.00476, 2024. 6 +[53] Yuan Liu, Le Tian, Xiao Zhou, Xinyu Gao, Kavio Yu, Yang Yu, and Jie Zhou. Points1. 5: Building a vision-language model towards real world applications. arXiv preprint arXiv:2412.08443, 2024. 5, 6, 8 +[54] Zhijian Liu, Ligeng Zhu, Baifeng Shi, Zhuoyang Zhang, Yuming Lou, Shang Yang, Haocheng Xi, Shiyi Cao, Yuxian Gu, Dacheng Li, et al. Nvila: Efficient frontier visual language models. arXiv preprint arXiv:2412.04468, 2024. 2 +[55] Pan Lu, Swaroop Mishra, Tony Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In The 36th Conference on Neural Information Processing Systems (NeurIPS), 2022. 5 +[56] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1697-1706, 2022. 6 +[57] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering by reading text in images. In ICDAR, 2019. 6 +[58] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 7 +[59] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7 +[60] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. NeurIPS, 2021. 2 +[61] Yossi Rubner, Carlo Tomasi, and Leonidas J Guibas. The earth mover's distance as a metric for image retrieval. International journal of computer vision, 40:99-121, 2000. 3 +[62] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge. In European conference on computer vision, pages 146–162. Springer, 2022. 5 +[63] Yuzhang Shang, Mu Cai, Bingxin Xu, Yong Jae Lee, and Yan Yan. Llava-prumerge: Adaptive token reduction for efficient large multimodal models. arXiv preprint arXiv:2403.15388, 2024. 2, 4, 7, 8 +[64] Dachuan Shi, Chaofan Tao, Ying Jin, Zhendong Yang, Chun Yuan, and Jiaqi Wang. Upop: Unified and progressive pruning for compressing vision-language transformers. arXiv preprint arXiv:2301.13741, 2023. 2 + +[65] Dachuan Shi, Chaofan Tao, Anyi Rao, Zhendong Yang, Chun Yuan, and Jiaqi Wang. Crossget: Cross-guided ensemble of tokens for accelerating vision-language transformers. arXiv preprint arXiv:2305.17455, 2023. 2, 5, 7, 8 +[66] Min Shi, Fuxiao Liu, Shihao Wang, Shijia Liao, Subhashree Radhakrishnan, De-An Huang, Hongxu Yin, Karan Sapra, Yaser Yacoob, Humphrey Shi, et al. Eagle: Exploring the design space for multimodal llms with mixture of encoders. arXiv preprint arXiv:2408.15998, 2024. 1, 2, 7 +[67] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. FLAVA: A foundational language and vision alignment model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2 +[68] Hao Tan and Mohit Bansal. LXMERT: Learning cross-modality encoder representations from transformers. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, EMNLP-IJCNLP, November 3-7, 2019, 2019. +[69] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 2 +[70] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9568-9578, 2024. 1, 2, 5, 6, 7 +[71] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 2 +[72] Tiannan Wang, Wangchunshu Zhou, Yan Zeng, and Xinsong Zhang. Efficientvlm: Fast and accurate vision-language models via knowledge distillation and modal-adaptive pruning. arXiv preprint arXiv:2210.07795, 2022. 2 +[73] Haoning Wu, Dongxu Li, Bei Chen, and Junnan Li. Longvideobench: A benchmark for long-context interleaved video-language understanding. Advances in Neural Information Processing Systems, 37:28828-28857, 2024. 5 +[74] Guangxuan Xiao, Ji Lin, Mickael Seznec, Hao Wu, Julien Demouth, and Song Han. Smoothquant: Accurate and efficient post-training quantization for large language models. In International Conference on Machine Learning, pages 38087-38099. PMLR, 2023. 2 +[75] Long Xing, Qidong Huang, Xiaoyi Dong, Jiajie Lu, Pan Zhang, Yuhang Zang, Yuhang Cao, Conghui He, Jiaqi Wang, Feng Wu, et al. Pyramiddrop: Accelerating your large vision-language models via pyramid visual redundancy reduction. arXiv preprint arXiv:2410.17247, 2024. 2, 5, 6, 7, 8 +[76] Runsen Xu, Xiaolong Wang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Pointllm: Empowering large language models to understand point clouds. In European + +Conference on Computer Vision, pages 131-147. Springer, 2024. 5 +[77] Senqiao Yang, Yukang Chen, Zhuotao Tian, Chengyao Wang, Jingyao Li, Bei Yu, and Jiaya Jia. Visionzip: Longer is better but not necessary in vision language models. arXiv preprint arXiv:2412.04467, 2024. 2 +[78] Xubing Ye, Yukang Gan, Yixiao Ge, Xiao-Ping Zhang, and Yansong Tang. Atp-llava: Adaptive token pruning for large vision language models. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 24972–24982, 2025. 2 +[79] Xubing Ye, Yukang Gan, Xiaoke Huang, Yixiao Ge, and Yansong Tang. Voco-llama: Towards vision compression with large language models. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 29836–29846, 2025. 2 +[80] Li Yifan, Du Yifan, Zhou Kun, Wang Jinpeng, Xin Zhao Wayne, and Wen Ji-Rong. Evaluating object hallucination in large vision-language models. In The 2023 Conference on Empirical Methods in Natural Language Processing, 2023. 5 +[81] Liu Yuan, Duan Haodong, Zhang Yuanhan, Li Bo, Zhang Songyang, Zhao Wangbo, Yuan Yike, Wang Jiaqi, He Conghui, Liu Ziwei, Chen Kai, and Lin Dahua. Mmbench: Is your multi-modal model an all-around player? arXiv:2307.06281, 2023. 5 +[82] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Ren-liang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of CVPR, 2024. 5 +[83] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding, 2023. 2 +[84] Yuan Zhang, Chun-Kai Fan, Junpeng Ma, Wenzhao Zheng, Tao Huang, Kuan Cheng, Denis Gudovskiy, Tomoyuki Okuno, Yohei Nakata, Kurt Keutzer, et al. Sparsevm: Visual token sparsification for efficient vision-language model inference. arXiv preprint arXiv:2410.04417, 2024. 2, 6 +[85] Yiwu Zhong, Zhuoming Liu, Yin Li, and Liwei Wang. Aim: Adaptive inference of multi-modal llms via token merging and pruning. arXiv preprint arXiv:2412.03248, 2024. 2 +[86] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. MLvu: A comprehensive benchmark for multitask long video understanding. arXiv e-prints, pages arXiv-2406, 2024. 6 \ No newline at end of file diff --git a/2025/FOLDER_ Accelerating Multi-Modal Large Language Models with Enhanced Performance/images.zip b/2025/FOLDER_ Accelerating Multi-Modal Large Language Models with Enhanced Performance/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..22cdff3e08381153f2b137fc134a7cf050d1e015 --- /dev/null +++ b/2025/FOLDER_ Accelerating Multi-Modal Large Language Models with Enhanced Performance/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1f7481e0f5a8fb6ce3d1dc94c28cb775b32a82392ca76a7eacf28f2fb921ed14 +size 723323 diff --git a/2025/FOLDER_ Accelerating Multi-Modal Large Language Models with Enhanced Performance/layout.json b/2025/FOLDER_ Accelerating Multi-Modal Large Language Models with Enhanced Performance/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..aa7bdc9c71d5be9dd8b4ff5faae5f2fbe2d0cb83 --- /dev/null +++ b/2025/FOLDER_ Accelerating Multi-Modal Large Language Models with Enhanced Performance/layout.json @@ -0,0 +1,10564 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 66, + 102, + 544, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 102, + 544, + 138 + ], + "spans": [ + { + "bbox": [ + 66, + 102, + 544, + 138 + ], + "type": "text", + "content": "FOLDER: Accelerating Multi-modal Large Language Models with Enhanced Performance" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 97, + 160, + 510, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 160, + 510, + 189 + ], + "spans": [ + { + "bbox": [ + 97, + 160, + 510, + 189 + ], + "type": "text", + "content": "Haicheng Wang" + }, + { + "bbox": [ + 97, + 160, + 510, + 189 + ], + "type": "inline_equation", + "content": "^{1,2,4*}" + }, + { + "bbox": [ + 97, + 160, + 510, + 189 + ], + "type": "text", + "content": ", Zhemeng Yu" + }, + { + "bbox": [ + 97, + 160, + 510, + 189 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 97, + 160, + 510, + 189 + ], + "type": "text", + "content": ", Gabriele Spadaro" + }, + { + "bbox": [ + 97, + 160, + 510, + 189 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 97, + 160, + 510, + 189 + ], + "type": "text", + "content": ", Chen Ju" + }, + { + "bbox": [ + 97, + 160, + 510, + 189 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 97, + 160, + 510, + 189 + ], + "type": "text", + "content": ", Victor Quétu" + }, + { + "bbox": [ + 97, + 160, + 510, + 189 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 97, + 160, + 510, + 189 + ], + "type": "text", + "content": ", Shuai Xiao" + }, + { + "bbox": [ + 97, + 160, + 510, + 189 + ], + "type": "inline_equation", + "content": "^{4\\boxtimes}" + }, + { + "bbox": [ + 97, + 160, + 510, + 189 + ], + "type": "text", + "content": ", Enzo Tartaglione" + }, + { + "bbox": [ + 97, + 160, + 510, + 189 + ], + "type": "inline_equation", + "content": "^{2\\boxtimes}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 82, + 189, + 528, + 217 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 84, + 189, + 523, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 189, + 523, + 203 + ], + "spans": [ + { + "bbox": [ + 84, + 189, + 523, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 84, + 189, + 523, + 203 + ], + "type": "text", + "content": " Paris Elite Institute of Technology, Shanghai Jiao Tong University " + }, + { + "bbox": [ + 84, + 189, + 523, + 203 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 84, + 189, + 523, + 203 + ], + "type": "text", + "content": " University of Turin" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 82, + 203, + 528, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 203, + 528, + 217 + ], + "spans": [ + { + "bbox": [ + 82, + 203, + 528, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 82, + 203, + 528, + 217 + ], + "type": "text", + "content": " LTCI, Télecom Paris, Institut Polytechnique de Paris " + }, + { + "bbox": [ + 82, + 203, + 528, + 217 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 82, + 203, + 528, + 217 + ], + "type": "text", + "content": " Taobao & Tmall Group of Alibaba" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 67, + 220, + 542, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 220, + 542, + 232 + ], + "spans": [ + { + "bbox": [ + 67, + 220, + 542, + 232 + ], + "type": "text", + "content": "anakinSkywalker@sjtu.edu.cn, shuai.xsh@alibaba-inc.com, enzo.tartaglione@telecom-paris.fr" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 164, + 258, + 212, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 258, + 212, + 272 + ], + "spans": [ + { + "bbox": [ + 164, + 258, + 212, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 284, + 296, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 284, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 296, + 559 + ], + "type": "text", + "content": "Recently, Multi-modal Large Language Models (MLLMs) have shown remarkable effectiveness for multi-modal tasks due to their ability of cross-modal understanding. However, processing long sequences of visual tokens extracted from visual backbones poses challenges for deployment in real-time applications. To address this issue, we introduce FOLDER, a simple yet effective plug-and-play module designed to reduce the length of the visual token sequence, mitigating computational and memory demands during both training and inference. Through a comprehensive analysis of the token reduction process in the vision encoder, we analyze the information loss introduced by different reduction strategies and develop FOLDER to preserve key information while removing visual redundancy. We show the effectiveness of FOLDER by integrating it into the visual backbone of various MLLMs, significantly accelerating the inference phase. Furthermore, we evaluate its utility as a training accelerator or even performance booster for MLLMs. FOLDER achieves comparable or even better performance than the original models, while dramatically reducing complexity by removing up to " + }, + { + "bbox": [ + 55, + 284, + 296, + 559 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 55, + 284, + 296, + 559 + ], + "type": "text", + "content": " of visual tokens. Our code is available at FOLDER Repository." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 575, + 135, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 135, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 135, + 588 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 596, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 596, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 596, + 296, + 693 + ], + "type": "text", + "content": "Multi-modal Large Language Models (MLLMs) excel in tasks like image captioning [12] and visual question answering [33] by learning joint visual-text representations. State-of-the-art models such as GPT-4V [1] and Claude 3 [2] show strong multimodal understanding ability. However, these models face concrete challenges in terms of computational efficiency, particularly when processing visual inputs. Modern MLLMs typically generate long sequences of im" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 316, + 258, + 555, + 402 + ], + "blocks": [ + { + "bbox": [ + 316, + 258, + 555, + 402 + ], + "lines": [ + { + "bbox": [ + 316, + 258, + 555, + 402 + ], + "spans": [ + { + "bbox": [ + 316, + 258, + 555, + 402 + ], + "type": "image", + "image_path": "b8d374ca51938da28fb9a98ee201711701d9c1ee0f9c56bfa8c0278ab60e80dc.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 317, + 404, + 553, + 504 + ], + "blocks": [ + { + "bbox": [ + 317, + 404, + 553, + 504 + ], + "lines": [ + { + "bbox": [ + 317, + 404, + 553, + 504 + ], + "spans": [ + { + "bbox": [ + 317, + 404, + 553, + 504 + ], + "type": "image", + "image_path": "309f1efceee5e179e46f7b793e4da6cf6bf2a14af7c63e298fa05ad387252967.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 511, + 555, + 546 + ], + "lines": [ + { + "bbox": [ + 313, + 511, + 555, + 546 + ], + "spans": [ + { + "bbox": [ + 313, + 511, + 555, + 546 + ], + "type": "text", + "content": "Figure 1. FOLDER as Accelerator & Booster. As a plug-and-play module, FOLDER can be used in both training and inference, with considerable acceleration and even a performance boost." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 568, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 555, + 713 + ], + "type": "text", + "content": "age tokens through their visual backbones. In some application scenarios, this problem becomes even more evident: recent advancements in high-resolution-supporting MLLMs [11, 71] might involve over 2000 visual tokens. Similarly, challenges also occur in multi-visual-expert architectures [32, 66, 70] that leverage multiple visual encoders. These issues are even more severe in video understanding [30, 40, 45] due to the extra temporal complexity. The resulting token explosion hinders real-time deployment because of the quadratic cost of the attention mechanism. Several studies have been conducted to address this issue. Existing approaches can be roughly divided into two" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 702, + 233, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 702, + 233, + 713 + ], + "spans": [ + { + "bbox": [ + 69, + 702, + 233, + 713 + ], + "type": "text", + "content": "*: Equal contribution, " + }, + { + "bbox": [ + 69, + 702, + 233, + 713 + ], + "type": "inline_equation", + "content": "{}^{20}" + }, + { + "bbox": [ + 69, + 702, + 233, + 713 + ], + "type": "text", + "content": " : Corresponding author." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23614" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "type": "text", + "content": "categories. The first category focuses on training phase solutions, such as Q-Former [39], pixel-shuffle [10] and pooling-based techniques [7]. However, these methods often suffer from performance degradation and exhibit limited scalability, due to their custom design for specific architectures. The second category aims to design plug-and-play token reduction modules for the inference phase [9, 27, 31, 65, 75, 84]. Specifically, [9, 75, 84] reduce visual tokens inside LLM. Though effective, these methods are often hard to implement, require adaptation for existing LLM acceleration modules [13], and prove unsuitable for multi-round conversations, thus greatly limiting their usage. Meanwhile, [27, 31, 65] identify the inherent redundancy in visual tokens, but fail to quantify the information loss during the token reduction process. Thus their reduction strategies are often arbitrary (uniform reduction for each layer), leading to sub-optimal outcomes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 277, + 295, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 295, + 422 + ], + "type": "text", + "content": "To find a reduction strategy that can be universally applicable and best preserve the performance, we propose a preliminary investigation into the vision backbone, to answer a fundamental question: Where does information loss originate? To model it, we identify three key factors that impact information loss during token removal: Reduction Impact, Propagation Effect, and Aggregation Method. Through a comprehensive empirical study of these factors, we devise a strategy that enhances the preservation of information while significantly reducing redundancy in the vision backbone. FOLDER is consequently designed to implement this strategy with minimal computational overhead." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 424, + 295, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 424, + 295, + 603 + ], + "spans": [ + { + "bbox": [ + 55, + 424, + 295, + 603 + ], + "type": "text", + "content": "Our plug-and-play solution can be easily integrated into the visual backbones of various MLLMs, effectively reducing the number of visual tokens while incurring negligible information loss (Fig. 1). We first incorporate FOLDER during MLLMs inference, reducing over " + }, + { + "bbox": [ + 55, + 424, + 295, + 603 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 55, + 424, + 295, + 603 + ], + "type": "text", + "content": " of visual tokens and achieving comparable, or even gains on performance across multiple tasks. Furthermore, when integrated into MLLMs pre-training, our module can be used as a training accelerator, as well as an effective regularization term, leading to remarkable performance improvements across all benchmarks, with a reduction ratio up to " + }, + { + "bbox": [ + 55, + 424, + 295, + 603 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 55, + 424, + 295, + 603 + ], + "type": "text", + "content": ". Our method is on-the-field proven to be a flexible, nearly lossless, and highly efficient token reduction strategy, offering a dual-purpose solution to the visual token reduction challenge. Our contributions can be summarized as follows." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 605, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 665 + ], + "type": "text", + "content": "- We conduct an in-depth analysis for information decay during token reduction in the vision backbone, identifying and quantifying key factors involved (Sec. 3.1, Sec. 3.2 and Sec. 3.3). These insights offer a clear understanding of how token reduction impacts the information flow." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "- Leveraging the above analysis, we develop a simple yet effective plug-and-play visual token reduction module for MLLMs (Sec. 3.4). This novel token reduction strategy aggressively reduces the number of tokens only in the" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 322, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 553, + 108 + ], + "type": "text", + "content": "last blocks of the visual encoder. On various models and benchmarks, FOLDER speeds up the inference " + }, + { + "bbox": [ + 322, + 72, + 553, + 108 + ], + "type": "inline_equation", + "content": "1.7 - 2.4 \\times" + }, + { + "bbox": [ + 322, + 72, + 553, + 108 + ], + "type": "text", + "content": " with negligible loss or even improvement (Sec. 4.2.1)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 108, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 108, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 314, + 108, + 553, + 144 + ], + "type": "text", + "content": "- We show that FOLDER can also be used in MLLM training (Sec. 4.2.2), to speed-up while boosting performance, serving as a dual-solution for both training and inference." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 155, + 406, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 155, + 406, + 167 + ], + "spans": [ + { + "bbox": [ + 314, + 155, + 406, + 167 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 175, + 553, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 175, + 553, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 553, + 390 + ], + "type": "text", + "content": "Multi-modal Large Language Models. MLLMs excel at understanding multi-modal information [1, 3, 7, 39, 41, 45, 48, 67-69, 71, 83], with common architectures comprising i.) a Visual Backbone, ii.) a Vision-Language Connector, and iii.) a Pre-trained LLM. However, integrating visual data significantly increases computational cost due to the large number of visual tokens. For example, LLaVA1.5 [50] processes 576 tokens for a " + }, + { + "bbox": [ + 313, + 175, + 553, + 390 + ], + "type": "inline_equation", + "content": "336 \\times 336" + }, + { + "bbox": [ + 313, + 175, + 553, + 390 + ], + "type": "text", + "content": " image, while VideoLLaVA [45] exceeds 2000 tokens with just 8 frames at " + }, + { + "bbox": [ + 313, + 175, + 553, + 390 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 313, + 175, + 553, + 390 + ], + "type": "text", + "content": ". Recent efforts to enhance MLLMs with multiple vision towers [32, 66, 70] further exacerbate this issue. MLLMs Acceleration. To accelerate large models, many methods focus on system-level optimizations, such as FlashAttention [14], vLLM [13], and RingAttention [49]. Model-centric techniques like knowledge distillation [20, 72], quantization [21, 74], and pruning [60, 64, 72] also reduce size and cost. However, these approaches overlook redundant data, which unnecessarily prolongs sequences." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 391, + 553, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 391, + 553, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 553, + 486 + ], + "type": "text", + "content": "Token Reduction in MLLMs. Some preliminary studies are conducted on token reduction for Vision Transformers [6, 34, 44, 85]. In the context of MLLMs, several token reduction methods, including Q-Former [39], resampler [11], pooling [7] and some other methods [29, 42, 54, 78, 79] are proposed to reduce visual tokens during the training process. However, these approaches often suffer from performance degradation and a lack of scalability." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 486, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 553, + 713 + ], + "type": "text", + "content": "Meanwhile, some studies try to handle the token reduction problem during inference. In particular, ToMe [6] proposes a training-free token merging strategy for unimodal image classification. Based on that, Turbo [31] proposes an improved merging strategy considering both mutual redundancy and semantic value, which serves as a plug-in module in MLLMs. Recently, FastV [9] recognizes the visual attention sparsity in LLM, thus pruning visual tokens inside the LLM based on the task-oriented attention importance. Though effective, these methods have clear drawbacks. For ToMe/Turbo-like methods [6, 27, 31, 63, 65, 77] that reduce tokens in the vision encoder, most of them adopt uniform progressive merging across layers, which is arbitrary and inflexible. While FastV-like methods [9, 51, 75, 84] that cut tokens inside LLM layers mainly suffer from 2 problems: i.) They often require interactions with instruction tokens, making it unsuitable for multi-round/long-term visual QA. ii.) Since these approaches are inserted in LLM and often require specific hyperparameters, it is difficult to adapt" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23615" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 76, + 290, + 216 + ], + "blocks": [ + { + "bbox": [ + 65, + 76, + 290, + 216 + ], + "lines": [ + { + "bbox": [ + 65, + 76, + 290, + 216 + ], + "spans": [ + { + "bbox": [ + 65, + 76, + 290, + 216 + ], + "type": "image", + "image_path": "f2312d72b13a11838068d5e60f83ba4379c9e0cae5892cb0c0c600a59cd120f8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 223, + 295, + 246 + ], + "lines": [ + { + "bbox": [ + 55, + 223, + 295, + 246 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 295, + 246 + ], + "type": "text", + "content": "Figure 2. Minimum Number of Tokens " + }, + { + "bbox": [ + 55, + 223, + 295, + 246 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 223, + 295, + 246 + ], + "type": "text", + "content": " to Preserve Energy " + }, + { + "bbox": [ + 55, + 223, + 295, + 246 + ], + "type": "inline_equation", + "content": "t_E" + }, + { + "bbox": [ + 55, + 223, + 295, + 246 + ], + "type": "text", + "content": " Across Blocks. We evaluate on 3 different " + }, + { + "bbox": [ + 55, + 223, + 295, + 246 + ], + "type": "inline_equation", + "content": "t_E" + }, + { + "bbox": [ + 55, + 223, + 295, + 246 + ], + "type": "text", + "content": " for every block." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 265, + 296, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 265, + 296, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 296, + 339 + ], + "type": "text", + "content": "for complex LLM modules [14] and hard to deploy. Besides, few approaches have been verified to be effective in both training and training-free inference ways. To address the above issues, we seek to find an effective, information-guided, and user-friendly token reduction method that can be easily applied in both training and inference." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 351, + 111, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 351, + 111, + 363 + ], + "spans": [ + { + "bbox": [ + 55, + 351, + 111, + 363 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 373, + 296, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 296, + 432 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 296, + 432 + ], + "type": "text", + "content": "In this section, we establish FOLDER by first answering three key questions for token reduction in vision backbone: i.) how many tokens to reduce (Sec. 3.1), ii.) in which block (Sec. 3.2) and iii.) through which aggregation method (Sec. 3.3). Then, we present our method in Sec. 3.4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 433, + 296, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 296, + 494 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 296, + 494 + ], + "type": "text", + "content": "Empirical Setup. To answer these three crucial questions, we define a common empirical setup. Specifically, we use a pre-trained 12-block ViT-B [16] model and conduct the empirical experiments on ImageNet-1k [15]. Results on more diverse setups are provided in the appendix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 503, + 194, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 194, + 516 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 194, + 516 + ], + "type": "text", + "content": "3.1. Token Reduction Impact" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 522, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 296, + 594 + ], + "type": "text", + "content": "To explore the link between token reduction and information loss, we use Singular Value Decomposition (SVD) to track the energy of the reduced token sequence. More precisely, given a token sequence " + }, + { + "bbox": [ + 55, + 522, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{n \\times d}" + }, + { + "bbox": [ + 55, + 522, + 296, + 594 + ], + "type": "text", + "content": " (with " + }, + { + "bbox": [ + 55, + 522, + 296, + 594 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 522, + 296, + 594 + ], + "type": "text", + "content": " being the number of tokens and " + }, + { + "bbox": [ + 55, + 522, + 296, + 594 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 522, + 296, + 594 + ], + "type": "text", + "content": " their size), we can decompose it by applying SVD on " + }, + { + "bbox": [ + 55, + 522, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^T" + }, + { + "bbox": [ + 55, + 522, + 296, + 594 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 604, + 295, + 618 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 604, + 295, + 618 + ], + "spans": [ + { + "bbox": [ + 141, + 604, + 295, + 618 + ], + "type": "interline_equation", + "content": "\\mathbf {X} ^ {T} = \\mathbf {U} \\boldsymbol {\\Sigma} \\mathbf {V} ^ {T}, \\tag {1}", + "image_path": "4a742d402600a2e831493383d8b28c96886ec9bb986648f14d06c5bc7aa1ebb5.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{U} \\in \\mathbb{R}^{d \\times d}" + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{V} \\in \\mathbb{R}^{n \\times n}" + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": " are two orthogonal matrices and " + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma} \\in \\mathbb{R}^{d \\times n}" + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": " is a diagonal matrix, with the singular values " + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\sigma_{i} = \\pmb{\\Sigma}_{ii}" + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": " representing the variances (energy) in the new compositional directions. In this way, SVD yields the optimal energy concentration that can be ordered by their contribution to the total variance (energy) [18]. Based on this, we can estimate how much variance we preserve by" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 325, + 76, + 547, + 216 + ], + "blocks": [ + { + "bbox": [ + 325, + 76, + 547, + 216 + ], + "lines": [ + { + "bbox": [ + 325, + 76, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 325, + 76, + 547, + 216 + ], + "type": "image", + "image_path": "62e69817cfadc5278ff9eb4633366b3fa4035d059732be2bb28f9968c9028b17.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 223, + 555, + 257 + ], + "lines": [ + { + "bbox": [ + 313, + 223, + 555, + 257 + ], + "spans": [ + { + "bbox": [ + 313, + 223, + 555, + 257 + ], + "type": "text", + "content": "Figure 3. EMD Distance Between Reduced and Original Output Distributions under 3 Reduction Ratios. We compare the EMD distance by exerting token reduction on different blocks." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 274, + 493, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 493, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 493, + 287 + ], + "type": "text", + "content": "only taking the largest " + }, + { + "bbox": [ + 313, + 274, + 493, + 287 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 274, + 493, + 287 + ], + "type": "text", + "content": " singular values, as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 353, + 295, + 555, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 295, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 353, + 295, + 555, + 326 + ], + "type": "interline_equation", + "content": "E (k) = \\frac {\\sum_ {i = 1} ^ {k} \\sigma_ {i}}{\\sum_ {i = 1} ^ {n} \\sigma_ {i}}, \\quad \\sigma_ {i} \\geq \\sigma_ {i + 1} \\quad \\forall i. \\tag {2}", + "image_path": "dd94da1ecc551c49e9cc47abfa0e0f042320d18a05670f74bf2e33a5b80bfa5e.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 331, + 555, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 555, + 368 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 555, + 368 + ], + "type": "text", + "content": "By definition, SVD provides an upper bound on preserved energy when keeping " + }, + { + "bbox": [ + 313, + 331, + 555, + 368 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 331, + 555, + 368 + ], + "type": "text", + "content": " tokens. Thus, by setting a threshold " + }, + { + "bbox": [ + 313, + 331, + 555, + 368 + ], + "type": "inline_equation", + "content": "t_E" + }, + { + "bbox": [ + 313, + 331, + 555, + 368 + ], + "type": "text", + "content": ", we can estimate the required " + }, + { + "bbox": [ + 313, + 331, + 555, + 368 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 331, + 555, + 368 + ], + "type": "text", + "content": " for each block as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 378, + 376, + 553, + 390 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 376, + 553, + 390 + ], + "spans": [ + { + "bbox": [ + 378, + 376, + 553, + 390 + ], + "type": "interline_equation", + "content": "k = \\min \\left\\{k \\mid E (k) \\geq t _ {E} \\right\\}. \\tag {3}", + "image_path": "831181cde982d656ef002dafdd81ed4434cf6cf61716468a148e7ecf1c066292.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 399, + 554, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 554, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 554, + 471 + ], + "type": "text", + "content": "Observations. With a threshold " + }, + { + "bbox": [ + 313, + 399, + 554, + 471 + ], + "type": "inline_equation", + "content": "t_E" + }, + { + "bbox": [ + 313, + 399, + 554, + 471 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 399, + 554, + 471 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 399, + 554, + 471 + ], + "type": "text", + "content": " varies across blocks. We monitor " + }, + { + "bbox": [ + 313, + 399, + 554, + 471 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 399, + 554, + 471 + ], + "type": "text", + "content": " under this threshold and find, as shown in Fig. 2, that later blocks require far fewer tokens than earlier ones to maintain the same energy. This highlights the potential for block-adaptive token reduction, suggesting the first and last blocks as key candidates." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 479, + 457, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 457, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 457, + 491 + ], + "type": "text", + "content": "3.2. Token Propagation Effect" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 496, + 555, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 555, + 567 + ], + "type": "text", + "content": "While token reduction impact provides an upper bound on information preservation, it overlooks inter-block dependencies and therefore may not adequately reflect the effects on subsequent blocks. In transformer architectures, errors from earlier blocks can propagate and amplify through successive layers [5], potentially affecting final performance." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 567, + 554, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 554, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 554, + 639 + ], + "type": "text", + "content": "To analyze this propagation effect, we examine how token reduction at different blocks affects the final output distribution. Thus, by comparing the original output token distribution (without reduction) and the one obtained by operating token reduction only in one block " + }, + { + "bbox": [ + 313, + 567, + 554, + 639 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 567, + 554, + 639 + ], + "type": "text", + "content": ", we can show the presence of such a propagation effect." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 639, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 555, + 713 + ], + "type": "text", + "content": "To measure this distortion, we consider output tokens from the original model " + }, + { + "bbox": [ + 313, + 639, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{Y} \\in \\mathbb{R}^{n \\times d}" + }, + { + "bbox": [ + 313, + 639, + 555, + 713 + ], + "type": "text", + "content": " and the ones obtained when " + }, + { + "bbox": [ + 313, + 639, + 555, + 713 + ], + "type": "inline_equation", + "content": "n - k" + }, + { + "bbox": [ + 313, + 639, + 555, + 713 + ], + "type": "text", + "content": " tokens are reduced in the block " + }, + { + "bbox": [ + 313, + 639, + 555, + 713 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 639, + 555, + 713 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 313, + 639, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{Y}}_b \\in \\mathbb{R}^{k \\times d}" + }, + { + "bbox": [ + 313, + 639, + 555, + 713 + ], + "type": "text", + "content": ". Then, we define the empirical distributions " + }, + { + "bbox": [ + 313, + 639, + 555, + 713 + ], + "type": "inline_equation", + "content": "P_Y" + }, + { + "bbox": [ + 313, + 639, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 639, + 555, + 713 + ], + "type": "inline_equation", + "content": "P_{\\tilde{\\mathbf{Y}}_b}" + }, + { + "bbox": [ + 313, + 639, + 555, + 713 + ], + "type": "text", + "content": " supported on " + }, + { + "bbox": [ + 313, + 639, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 313, + 639, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 639, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{Y}}_b" + }, + { + "bbox": [ + 313, + 639, + 555, + 713 + ], + "type": "text", + "content": " respectively, and we evaluate Earth Mover's Distance (EMD) [61] between them:" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23616" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 72, + 292, + 212 + ], + "blocks": [ + { + "bbox": [ + 61, + 72, + 292, + 212 + ], + "lines": [ + { + "bbox": [ + 61, + 72, + 292, + 212 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 292, + 212 + ], + "type": "image", + "image_path": "f86748aa632ab4874bd8b5a9a22d644632e297e60eeec252628922aa0b48b986.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 221, + 295, + 257 + ], + "lines": [ + { + "bbox": [ + 55, + 221, + 295, + 257 + ], + "spans": [ + { + "bbox": [ + 55, + 221, + 295, + 257 + ], + "type": "text", + "content": "Figure 4. EMD & Accuracy Comparison for Aggregation Methods. We compare three aggregation methods under different reduction ratios on both EMD and Accuracy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 107, + 280, + 295, + 299 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 280, + 295, + 299 + ], + "spans": [ + { + "bbox": [ + 107, + 280, + 295, + 299 + ], + "type": "interline_equation", + "content": "\\operatorname {E M D} \\left(P _ {Y}, P _ {\\tilde {Y} _ {b}}\\right) = \\min _ {\\gamma} \\langle \\gamma , \\mathbf {M} \\rangle_ {F}, \\tag {4}", + "image_path": "f61b4160dce8f471ea029132a70606a7f527e67b3d49163cd3ce38209fc9d706.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 306, + 296, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 306, + 296, + 381 + ], + "spans": [ + { + "bbox": [ + 55, + 306, + 296, + 381 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 306, + 296, + 381 + ], + "type": "inline_equation", + "content": "\\mathbf{M} = (d_{ij})_{n\\times k} = (\\| y_i - \\tilde{y}_{bj}\\| _2)_{n\\times k}" + }, + { + "bbox": [ + 55, + 306, + 296, + 381 + ], + "type": "text", + "content": " denotes the metric cost matrix, with " + }, + { + "bbox": [ + 55, + 306, + 296, + 381 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 55, + 306, + 296, + 381 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 306, + 296, + 381 + ], + "type": "inline_equation", + "content": "\\tilde{y}_{bj}" + }, + { + "bbox": [ + 55, + 306, + 296, + 381 + ], + "type": "text", + "content": " being points in the support of distributions " + }, + { + "bbox": [ + 55, + 306, + 296, + 381 + ], + "type": "inline_equation", + "content": "P_Y" + }, + { + "bbox": [ + 55, + 306, + 296, + 381 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 306, + 296, + 381 + ], + "type": "inline_equation", + "content": "P_{\\tilde{Y}_b}" + }, + { + "bbox": [ + 55, + 306, + 296, + 381 + ], + "type": "text", + "content": " respectively. " + }, + { + "bbox": [ + 55, + 306, + 296, + 381 + ], + "type": "inline_equation", + "content": "\\gamma = (\\gamma_{ij})_{n\\times k}" + }, + { + "bbox": [ + 55, + 306, + 296, + 381 + ], + "type": "text", + "content": " represents the optimal transport matrix, which satisfies the non-negative constraint " + }, + { + "bbox": [ + 55, + 306, + 296, + 381 + ], + "type": "inline_equation", + "content": "\\gamma_{ij}\\geq 0,\\forall i,j" + }, + { + "bbox": [ + 55, + 306, + 296, + 381 + ], + "type": "text", + "content": " and the marginal constraints " + }, + { + "bbox": [ + 55, + 306, + 296, + 381 + ], + "type": "inline_equation", + "content": "\\sum_{j = 1}^{k}\\gamma_{ij} = P_Y(y_i)" + }, + { + "bbox": [ + 55, + 306, + 296, + 381 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 306, + 296, + 381 + ], + "type": "inline_equation", + "content": "\\sum_{i = 1}^{n}\\gamma_{ij} = P_{\\tilde{Y}_b}(\\tilde{y}_{bj})" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 381, + 296, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 381, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 381, + 296, + 525 + ], + "type": "text", + "content": "Observations. Due to the propagation effect, token reduction at different blocks causes varying levels of output distortion. We monitor EMD (Eq. (4)) when reducing the same number of tokens per block and find, as shown in Fig.3, that early-block reductions lead to significantly higher EMD values. This confirms that distortion amplifies across layers, making early blocks unsuitable for token reduction, contrary to Fig.2. Conversely, with a massive " + }, + { + "bbox": [ + 55, + 381, + 296, + 525 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 55, + 381, + 296, + 525 + ], + "type": "text", + "content": " reduction, EMD remains low in the last layers. This suggests that token reduction is most effective at the network's end, even with a high reduction rate. Results on BLIP [38] (Tab. 7) and appendix further support this conclusion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 536, + 206, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 536, + 206, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 206, + 548 + ], + "type": "text", + "content": "3.3. Token Aggregation Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 554, + 296, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 554, + 296, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 554, + 296, + 674 + ], + "type": "text", + "content": "In Sec. 3.1, we use SVD to estimate the upper bound of information preservation. However, its computational complexity prevents direct application for token reduction. Following prior works [6, 31], we decompose the process into two stages: (i) Token Matching, grouping tokens based on a matching function, and (ii) Token Aggregation, merging grouped tokens. While matching functions have been explored [6, 31], aggregation remains understudied. Formally, given a set of tokens " + }, + { + "bbox": [ + 55, + 554, + 296, + 674 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_1,\\mathbf{x}_2,\\dots ,\\mathbf{x}_m\\}" + }, + { + "bbox": [ + 55, + 554, + 296, + 674 + ], + "type": "text", + "content": " to be merged into " + }, + { + "bbox": [ + 55, + 554, + 296, + 674 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 55, + 554, + 296, + 674 + ], + "type": "text", + "content": ", we define the aggregation operation as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 110, + 685, + 295, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 685, + 295, + 716 + ], + "spans": [ + { + "bbox": [ + 110, + 685, + 295, + 716 + ], + "type": "interline_equation", + "content": "\\mathbf {y} = \\sum_ {i = 1} ^ {m} \\alpha_ {i} \\mathbf {x} _ {i}, \\text {s . t .} \\sum_ {i = 1} ^ {m} \\alpha_ {i} = 1. \\tag {5}", + "image_path": "036a980e35d105f23015c8dac61f99682a48b03bd69bcc67cbf73e562ad19c90.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 152 + ], + "type": "text", + "content": "Based on Eq. (5), we formulate three different token aggregation approaches: i.) Average Merging, with " + }, + { + "bbox": [ + 313, + 72, + 553, + 152 + ], + "type": "inline_equation", + "content": "\\alpha_{i} = \\frac{1}{m},\\forall i" + }, + { + "bbox": [ + 313, + 72, + 553, + 152 + ], + "type": "text", + "content": " ii.) Weighted Merging, with " + }, + { + "bbox": [ + 313, + 72, + 553, + 152 + ], + "type": "inline_equation", + "content": "\\alpha_{i} = \\frac{\\|x_{i}\\|_{2}}{\\sum_{j = 1}^{m}\\|x_{j}\\|_{2}},\\forall i;" + }, + { + "bbox": [ + 313, + 72, + 553, + 152 + ], + "type": "text", + "content": " iii.) Direct Dropping, with " + }, + { + "bbox": [ + 313, + 72, + 553, + 152 + ], + "type": "inline_equation", + "content": "\\alpha_{i} = 0,\\forall i\\neq i_{max},\\alpha_{i_{max}} = 1" + }, + { + "bbox": [ + 313, + 72, + 553, + 152 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 72, + 553, + 152 + ], + "type": "inline_equation", + "content": "i_{max} = \\underset {i}{\\operatorname{argmax}}\\| x_i\\| _2" + }, + { + "bbox": [ + 313, + 72, + 553, + 152 + ], + "type": "text", + "content": " . Here, the norm " + }, + { + "bbox": [ + 313, + 72, + 553, + 152 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\| _2" + }, + { + "bbox": [ + 313, + 72, + 553, + 152 + ], + "type": "text", + "content": " is used as the importance score for tokens." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 154, + 555, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 154, + 555, + 250 + ], + "spans": [ + { + "bbox": [ + 313, + 154, + 555, + 250 + ], + "type": "text", + "content": "Observations. Similarly to Sec. 3.2, we compare aggregation methods using EMD. As shown in Tab. 9 and Fig. 4, both merging methods yield lower discrepancy than Direct Dropping. Given its comparable performance to Weighted Merging with greater simplicity, we select Average Merging as our preferred method. Additionally, Fig. 4 shows that a lower EMD indeed correlates with higher accuracy. Further supporting results are in Tab. 9 and appendix." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 261, + 383, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 261, + 383, + 272 + ], + "spans": [ + { + "bbox": [ + 314, + 261, + 383, + 272 + ], + "type": "text", + "content": "3.4.FOLDER" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 280, + 555, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 280, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 280, + 555, + 434 + ], + "type": "text", + "content": "Motivation. Based on the above empirical analyses, we deduce that: i.) Token reduction should take place in the last blocks, where aggressive reduction is permitted. ii.) Merging is better than dropping as an aggregation method. Therefore, our goal is to design an algorithm that can reduce as many tokens as possible in one block. Previous works [63] mainly leverage clustering for token merging, which is slow and unsuitable for parallel computing. To avoid such issues, we refer to bipartite matching [6], which is however limited by a maximum reduction ratio of " + }, + { + "bbox": [ + 313, + 280, + 555, + 434 + ], + "type": "inline_equation", + "content": "1/2" + }, + { + "bbox": [ + 313, + 280, + 555, + 434 + ], + "type": "text", + "content": ". Thus, to aggressively reduce tokens in one block, we propose FOLDER, an advanced bipartite matching strategy that can reduce an arbitrary number of tokens." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "text", + "content": "Strategy. As illustrated in Fig. 5, our strategy focuses on the last blocks (or the last) in the vision backbone. In some cases, the target reduction " + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "text", + "content": " exceeds half of the total tokens " + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\lfloor n / 2 \\rfloor" + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "text", + "content": " in a block, a scenario we call reduction overflow. To handle this, FOLDER automatically applies an iterative \"FOLD\" operation. As detailed in Alg. 1, if reduction overflow occurs in an iteration " + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "inline_equation", + "content": "(r_{\\text{remain}} > \\lfloor n' / 2 \\rfloor" + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "text", + "content": " in line 7), we reduce tokens by half " + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "inline_equation", + "content": "(r_{\\text{fold}} = \\lfloor n' / 2 \\rfloor)" + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "text", + "content": ". The process begins by splitting the token input sequence into equal-sized partitions: " + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{A}" + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{B}" + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "text", + "content": " (line 9). Then, each token in " + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{A}" + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "text", + "content": " is matched to its most similar counterpart in " + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{B}" + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "text", + "content": " based on a Matching Function [6, 31] " + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "inline_equation", + "content": "S : (\\mathbb{R}^d, \\mathbb{R}^d) \\to \\mathbb{R}" + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "text", + "content": " (line 10). A visual representation is provided in Fig. 5, where the \"Partition & Match\" operation is highlighted. These matches are then sorted based on their matching scores " + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "inline_equation", + "content": "(\\stackrel{s_{ij}}{\\rightarrow})" + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "text", + "content": ", and only the top " + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "inline_equation", + "content": "r_{\\text{fold}}" + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "text", + "content": " matches are preserved (line 11). At this point, by merging these matches in " + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{top}" + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{A}" + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{B}" + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "text", + "content": ", we obtain two new sets: " + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{A}'" + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{B}'" + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "text", + "content": ". Namely, " + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{B}'" + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "text", + "content": " updates " + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{B}" + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "text", + "content": " by replacing original tokens with matched tokens aggregated using a specific Aggregation Method (Sec. 3.3). " + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{A}'" + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "text", + "content": ", instead, contains the tokens in " + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{A}" + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "text", + "content": " whose matches were discarded in line 11. Clearly, if the reduction overflow occurs, " + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "inline_equation", + "content": "r_{\\text{fold}} = \\lfloor n' / 2 \\rfloor" + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "text", + "content": " and thus " + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{A}' = \\emptyset" + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "text", + "content": " (\"First Fold\" of Fig. 5). The concatenation between" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23617" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 200, + 194 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 200, + 194 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 200, + 194 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 200, + 194 + ], + "type": "image", + "image_path": "7e5a3493389f1845588d7d7081c9d90be78a61c80906d5e10ae784b2b260648c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 206, + 555, + 240 + ], + "lines": [ + { + "bbox": [ + 54, + 206, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 54, + 206, + 555, + 240 + ], + "type": "text", + "content": "Figure 5. Pipeline of FOLDER. As a plug-and-play module, FOLDER is integrated into the final blocks of the vision backbone (last two here). To deal with reduction overflow, FOLDER automatically executes another FOLD operation when the expected reduction is more than half. The last FOLD, which escapes from reduction overflow, merges tokens according to the remaining reduction numbers." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 200, + 70, + 372, + 196 + ], + "blocks": [ + { + "bbox": [ + 200, + 70, + 372, + 196 + ], + "lines": [ + { + "bbox": [ + 200, + 70, + 372, + 196 + ], + "spans": [ + { + "bbox": [ + 200, + 70, + 372, + 196 + ], + "type": "image", + "image_path": "8139d0587137e2031d0ddd89611f00975e58341de69122e62abe804805f2fd7d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 373, + 70, + 553, + 198 + ], + "blocks": [ + { + "bbox": [ + 373, + 70, + 553, + 198 + ], + "lines": [ + { + "bbox": [ + 373, + 70, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 373, + 70, + 553, + 198 + ], + "type": "image", + "image_path": "2522f93e5cd5212032ccd98adc056bea429ed78212b25b78108d0003c3036331.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 255, + 294, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 255, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 54, + 255, + 294, + 386 + ], + "type": "inline_equation", + "content": "\\mathbb{A}^{\\prime}" + }, + { + "bbox": [ + 54, + 255, + 294, + 386 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 255, + 294, + 386 + ], + "type": "inline_equation", + "content": "\\mathbb{B}^{\\prime}" + }, + { + "bbox": [ + 54, + 255, + 294, + 386 + ], + "type": "text", + "content": " represents the input token sequence for the next \"FOLD\" iteration, in which " + }, + { + "bbox": [ + 54, + 255, + 294, + 386 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{remain}}" + }, + { + "bbox": [ + 54, + 255, + 294, + 386 + ], + "type": "text", + "content": " tokens are going to be removed. This process continues through successive folds (line 3) until the overflow is no longer encountered. In this case, " + }, + { + "bbox": [ + 54, + 255, + 294, + 386 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{fold}} = r_{\\mathrm{remain}}" + }, + { + "bbox": [ + 54, + 255, + 294, + 386 + ], + "type": "text", + "content": " (line 7), and " + }, + { + "bbox": [ + 54, + 255, + 294, + 386 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{remain}}" + }, + { + "bbox": [ + 54, + 255, + 294, + 386 + ], + "type": "text", + "content": " for the next iteration will be equal to 0 (line 8). A visual representation of this case is provided in Fig. 5 as the \"Last FOLD\" iteration. Note that in the last FOLD, to fit for the positional encoding in LLM, instead of directly concatenating " + }, + { + "bbox": [ + 54, + 255, + 294, + 386 + ], + "type": "inline_equation", + "content": "\\mathbb{A}^{\\prime}" + }, + { + "bbox": [ + 54, + 255, + 294, + 386 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 255, + 294, + 386 + ], + "type": "inline_equation", + "content": "\\mathbb{B}^{\\prime}" + }, + { + "bbox": [ + 54, + 255, + 294, + 386 + ], + "type": "text", + "content": ", we rearrange the tokens into original order. More details are provided in the appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 387, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 387, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 295, + 579 + ], + "type": "text", + "content": "Discussion. Though simple, the insight contained in FOLDER is non-trivial. Empirically, by analysing the information flow in the vision backbone, we draw the key conclusion that token reduction should be done only in the last blocks, which is overlooked by previous works [6, 27, 31, 65]. Structurally, by generalizing bipartite matching, we specifically design a highly efficient token reduction strategy for MLLMs architecture that allows parallel computing and minimum deployment effort. Functionally, FOLDER focuses on reducing the inherent visual redundancy in the vision backbone, which is task agnostic. Thus FOLDER can superpose on task-oriented/attention-sparsity methods like FastV [9] to enjoy further acceleration (Tab. 1). Besides, FOLDER is among the very few approaches [75] that are verified to be effective in both training and training-free ways (Tab. 1 and Tab. 3), showing its superior scalability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 136, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 136, + 596 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 136, + 596 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 602, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 602, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 602, + 295, + 651 + ], + "type": "text", + "content": "In this section, we present the results obtained by our method. By using FOLDER, we reduce the number of visual tokens to accelerate both training and inference of the Multi-modal Large Language Model (MLLM)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 659, + 204, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 204, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 204, + 672 + ], + "type": "text", + "content": "4.1. Evaluation & Benchmarks" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "Evaluation Models. We evaluate FOLDER as training-free plug-in on various MLLMs, including image understanding models LLaVA1.5-7B/13B [50], LLaVA-OV-7B [37]," + } + ] + } + ], + "index": 9 + }, + { + "type": "code", + "bbox": [ + 316, + 272, + 553, + 492 + ], + "blocks": [ + { + "bbox": [ + 320, + 258, + 503, + 270 + ], + "lines": [ + { + "bbox": [ + 320, + 258, + 503, + 270 + ], + "spans": [ + { + "bbox": [ + 320, + 258, + 503, + 270 + ], + "type": "text", + "content": "Algorithm 1: FOLDER applied in one block" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 316, + 272, + 553, + 492 + ], + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 492 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 492 + ], + "type": "text", + "content": "Input: Token sequence " + }, + { + "bbox": [ + 316, + 272, + 553, + 492 + ], + "type": "inline_equation", + "content": "\\mathbf{X}\\in \\mathbb{R}^{n\\times d}" + }, + { + "bbox": [ + 316, + 272, + 553, + 492 + ], + "type": "text", + "content": " reduced num " + }, + { + "bbox": [ + 316, + 272, + 553, + 492 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 316, + 272, + 553, + 492 + ], + "type": "text", + "content": " Output: Reduced token sequence " + }, + { + "bbox": [ + 316, + 272, + 553, + 492 + ], + "type": "inline_equation", + "content": "\\mathbf{X}'\\in \\mathbb{R}^{(n - r)\\times d}" + }, + { + "bbox": [ + 316, + 272, + 553, + 492 + ], + "type": "text", + "content": " \n1 " + }, + { + "bbox": [ + 316, + 272, + 553, + 492 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{remain}} = r" + }, + { + "bbox": [ + 316, + 272, + 553, + 492 + ], + "type": "text", + "content": " . \n2 " + }, + { + "bbox": [ + 316, + 272, + 553, + 492 + ], + "type": "inline_equation", + "content": "\\mathbf{X}' = \\mathbf{X}" + }, + { + "bbox": [ + 316, + 272, + 553, + 492 + ], + "type": "text", + "content": " .. \n3 while " + }, + { + "bbox": [ + 316, + 272, + 553, + 492 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{remain}} > 0" + }, + { + "bbox": [ + 316, + 272, + 553, + 492 + ], + "type": "text", + "content": " do \n4 " + }, + { + "bbox": [ + 316, + 272, + 553, + 492 + ], + "type": "inline_equation", + "content": "\\begin{array}{rl}{\\mathbf{\\Lambda}_{\\mathbf{X}^{\\prime}},r_{\\mathrm{remain}} = \\mathrm{FOLD}(\\mathbf{X}^{\\prime},r_{\\mathrm{remain}});} & {} \\end{array}" + }, + { + "bbox": [ + 316, + 272, + 553, + 492 + ], + "type": "text", + "content": " \n5 \n6 Function FOLD " + }, + { + "bbox": [ + 316, + 272, + 553, + 492 + ], + "type": "inline_equation", + "content": "(\\mathbf{X}'\\in \\mathbb{R}^{n'\\times d},r_{\\mathrm{remain}})" + }, + { + "bbox": [ + 316, + 272, + 553, + 492 + ], + "type": "text", + "content": " .. \n7 " + }, + { + "bbox": [ + 316, + 272, + 553, + 492 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{fold}} = \\min (\\lfloor n' / 2\\rfloor ,r_{\\mathrm{remain}});" + }, + { + "bbox": [ + 316, + 272, + 553, + 492 + ], + "type": "text", + "content": " \n8 " + }, + { + "bbox": [ + 316, + 272, + 553, + 492 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{remain}} = r_{\\mathrm{remain}} - r_{\\mathrm{fold}};" + }, + { + "bbox": [ + 316, + 272, + 553, + 492 + ], + "type": "text", + "content": " \n9 " + }, + { + "bbox": [ + 316, + 272, + 553, + 492 + ], + "type": "inline_equation", + "content": "\\{\\mathbb{A},\\mathbb{B}\\} \\Leftarrow \\mathrm{Split~X'}" + }, + { + "bbox": [ + 316, + 272, + 553, + 492 + ], + "type": "text", + "content": " into 2 equal-sized partitions; \n10 M " + }, + { + "bbox": [ + 316, + 272, + 553, + 492 + ], + "type": "inline_equation", + "content": "\\Leftarrow" + }, + { + "bbox": [ + 316, + 272, + 553, + 492 + ], + "type": "text", + "content": " Compute " + }, + { + "bbox": [ + 316, + 272, + 553, + 492 + ], + "type": "inline_equation", + "content": "\\{(a_i\\xrightarrow{s_{ij}}b_j)\\mid" + }, + { + "bbox": [ + 316, + 272, + 553, + 492 + ], + "type": "inline_equation", + "content": "b_{j} = \\arg \\max s_{ij} = S(a_{i},b_{j}),\\forall a_{i}\\in \\mathbb{A}\\} ;" + }, + { + "bbox": [ + 316, + 272, + 553, + 492 + ], + "type": "text", + "content": " \n11 " + }, + { + "bbox": [ + 316, + 272, + 553, + 492 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{top}\\Leftarrow" + }, + { + "bbox": [ + 316, + 272, + 553, + 492 + ], + "type": "text", + "content": " Sort M based on " + }, + { + "bbox": [ + 316, + 272, + 553, + 492 + ], + "type": "inline_equation", + "content": "\\stackrel {s_{ij}}{\\to}" + }, + { + "bbox": [ + 316, + 272, + 553, + 492 + ], + "type": "text", + "content": " & keep Top " + }, + { + "bbox": [ + 316, + 272, + 553, + 492 + ], + "type": "inline_equation", + "content": "r_\\mathrm{fold}" + }, + { + "bbox": [ + 316, + 272, + 553, + 492 + ], + "type": "text", + "content": " . \n12 " + }, + { + "bbox": [ + 316, + 272, + 553, + 492 + ], + "type": "inline_equation", + "content": "\\{\\mathbb{A}',\\mathbb{B}'\\} \\Leftarrow" + }, + { + "bbox": [ + 316, + 272, + 553, + 492 + ], + "type": "text", + "content": " Merge A to B according to Mtop; \n13 return Rearrange(A',B'), " + }, + { + "bbox": [ + 316, + 272, + 553, + 492 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{remain}}" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "code_body" + } + ], + "index": 11, + "sub_type": "algorithm" + }, + { + "bbox": [ + 313, + 519, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 519, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 519, + 555, + 616 + ], + "type": "text", + "content": "Minigpt4v2 [7], VILA1.5-13B [46], WePOINTS1.5-7b [53], VITA1.5-7B [24], Idefics2-8B [35], and Flamingov2-7B [4], multi-vision-tower model MMVP [70], video understanding model Video-LLaVA [45] and point cloud understanding model PointLLMv1.2-7B [76]. For merging strategy ablation, we adopt BLIP [38] for the image captioning task. For MLLM training, we use LLaVA1.5-13B [50]. Training details are provided in the appendix." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": "Benchmarks. To quantify the ability of the accelerated MLLMs, we conduct thorough experiments on a wide range of tasks with VLMEvalkit [17], including " + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "inline_equation", + "content": "Y/N" + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": " tasks like MME [22], HallusionBench [26], POPE [80]; MCQ tasks like MMBench-EN [81], CCBench [81], A-OKVQA [62], ScienceQAIMG [55], SEEDBenchIMG [36], MMMU [82], MMStar [8], Video-MME [23], RealWorld-QA [25], LongVideoBench [73]," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23618" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 553, + 337 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 337 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 337 + ], + "type": "table", + "html": "
MethodReduct RatioSpeed UpMMBench ENMM StarMMEA-OK VQAHallusion BenchMM MUOCR VQASEED BenchIMGScience QAPOPEAvg
Original-7B0%1×62.832.71338.978.835.632.252.460.268.179.756.9
Pooling50%1.5×59.530.31308.677.135.331.744.859.268.084.055.6
FastV [9]1.3×63.332.41345.178.636.528.753.159.467.881.056.8
Turbo [31]1.5×60.430.91311.277.235.528.035.758.167.583.054.2
Ours1.5×62.432.11338.278.338.334.048.559.568.585.457.4
Pooling66%1.7×57.629.71308.274.232.330.738.757.667.380.753.5
FastV [9]1.5×62.532.01353.277.737.430.052.658.368.279.356.5
Pyramidddrop [75]1.6×61.331.21338.776.938.629.348.656.968.381.956.0
SparseVLM [84]1.6×61.331.11333.077.133.333.138.257.667.482.354.8
Turbo [31]1.7×60.131.51301.678.034.924.734.457.867.285.053.8
Ours1.7×61.430.31350.077.939.231.346.159.768.385.456.7
FastV [9]75%1.6×61.231.01321.676.637.532.050.857.268.177.755.8
Ours+FastV1.7×61.431.21325.578.039.132.648.859.268.682.356.8
Original-13B*0%1×66.630.91371.177.036.134.054.959.468.886.458.3
Pooling50%1.5×64.131.21316.676.235.130.054.757.669.085.757.0
FastV [9]1.3×66.431.11386.375.536.232.654.958.968.485.457.9
Turbo [31]1.5×65.030.61200.278.226.932.747.958.869.786.155.6
Ours1.5×65.431.21383.778.636.434.456.259.170.586.958.8
Pooling66%1.6×62.831.01250.673.529.532.545.654.569.782.154.4
FastV [9]1.5×66.331.21352.275.035.532.354.458.068.383.557.2
Pyramidddrop [75]1.6×65.331.11313.075.835.332.052.558.069.483.556.9
SparseVLM [84]1.6×65.231.11272.677.031.132.745.858.069.583.155.7
Turbo [31]1.6×64.731.21168.376.325.933.347.958.570.585.555.3
Ours1.6×65.831.71366.977.333.835.052.658.870.786.158.1
FastV [9]75%1.6×64.231.71321.374.236.931.353.356.468.182.156.4
Ours+FastV1.8×65.132.31368.677.835.332.055.258.871.185.858.2
", + "image_path": "ba75cb6fd19586931c9c7ea9fa1f6fb007972c8db89fdc4652ca4f1bef42a3d8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 372, + 552, + 446 + ], + "blocks": [ + { + "bbox": [ + 55, + 339, + 555, + 363 + ], + "lines": [ + { + "bbox": [ + 55, + 339, + 555, + 363 + ], + "spans": [ + { + "bbox": [ + 55, + 339, + 555, + 363 + ], + "type": "text", + "content": "Table 1. Results on LLaVA1.5 7B/13B. We highlight the best result for each benchmark under the same reduction ratio. All baseline evaluations are reproduced using official codes on VLMEvalkit [17]. *Retrained to provide a fair comparison in Tab. 3." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 372, + 552, + 446 + ], + "lines": [ + { + "bbox": [ + 57, + 372, + 552, + 446 + ], + "spans": [ + { + "bbox": [ + 57, + 372, + 552, + 446 + ], + "type": "table", + "html": "
MethodReduct RatioSpeed UpMMBench ENMM StarMMEHallusion BenchOCR VQAMM MUPOPESEED BenchIMGScience QACC BenchAvg
Original0%1×64.232.11403.646.353.334.286.760.670.018.853.6
Turbo [31]50%1.5×62.832.01352.944.047.634.085.658.669.117.951.9
Ours1.5×63.032.11362.346.250.534.785.759.570.418.652.8
Turbo [31]66%1.7×62.530.71341.846.246.633.384.958.169.718.851.8
Ours1.7×63.631.91368.247.947.234.085.358.770.719.652.8
", + "image_path": "7e58a098dd1779d463baf93a9eee9a29272d9856a0421573a5821d149b967201.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 144, + 448, + 465, + 459 + ], + "lines": [ + { + "bbox": [ + 144, + 448, + 465, + 459 + ], + "spans": [ + { + "bbox": [ + 144, + 448, + 465, + 459 + ], + "type": "text", + "content": "Table 2. Results on MMVP (MLLM with Multiple Vision-towers (CLIP+DINOv2))." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 474, + 295, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 511 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 511 + ], + "type": "text", + "content": "MLVU [86]; and VQA tasks like OCRVQA [57], InfoVQA [56], MMBench-Video [19], TempCompass [52]. For BLIP, we evaluate on COCO image captioning [47]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 518, + 141, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 141, + 530 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 141, + 530 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 536, + 296, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 536, + 296, + 572 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 296, + 572 + ], + "type": "text", + "content": "In the following experiments, we show the effectiveness of FOLDER on different MLLMs. If not specified, we only apply FOLDER in the last layer of the visual encoder." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 578, + 228, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 228, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 228, + 590 + ], + "type": "text", + "content": "4.2.1. Inference Acceleration on MLLM" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 653 + ], + "type": "text", + "content": "We evaluate FOLDER's plug-and-play acceleration performance in inference phase on various image understanding MLLMs [7, 24, 46, 50, 53] and multi-vision-tower based model [70]. If not specified, we evaluate MLLM's speed-up based on the time of the first output token." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "Single-Vision-Tower MLLM. Tab. 1 compares FOLDER with previous plug-and-play SOTA methods [9, 31, 75, 84] on LLaVA1.5. FOLDER achieves the best performance-speed trade-off, even surpassing original performance with over " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " reduction on several benchmarks. Compared to" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 475, + 555, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 555, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 555, + 534 + ], + "type": "text", + "content": "Turbo [31], which applies uniform merging across the vision encoder, FOLDER delivers superior results under the same reduction ratio, especially on visually dense tasks like OCRVQA. This confirms our key insight: token reduction is most effective when applied only in the final blocks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 534, + 556, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 677 + ], + "type": "text", + "content": "While FastV [9] performs similarly to FOLDER on some benchmarks, its limitations (Sec. 2) make it inherently slower and harder to implement. In contrast, FOLDER is a plug-and-play module requiring minimal adjustments to integrate with vision backbones. By reducing visual tokens before interacting with the LLM, FOLDER mitigates architectural complexities. Notably, since FOLDER and FastV tackle token reduction from different angles (visual redundancy vs. task-oriented attention), their combination (50% × 50%) enables further reduction while maintaining performance. Similarly, LLM token reduction methods [75, 84] also face deployment challenges." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 678, + 555, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 678, + 555, + 702 + ], + "spans": [ + { + "bbox": [ + 313, + 678, + 555, + 702 + ], + "type": "text", + "content": "To validate our approach, we replaced the matching function [31] with naive mean pooling. The result in Tab. 1" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23619" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 69, + 553, + 201 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 553, + 201 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 553, + 201 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 553, + 201 + ], + "type": "table", + "html": "
TrainTrain ReductInfer ReductSpeed UpMMBench ENMM StarMMEA-OK VQAInfo VQAOCR VQASEED BenchIMGScience QAPOPEAvg
Original0%0%(1,1)×66.630.91371.177.015.554.959.468.886.458.7
Pyramiddrop [75]66%0%(1.5, 1.5)×65.733.01354.279.515.357.461.270.687.959.8
Pyramiddrop [75]66%66%(1.5, 1.5)×66.732.51359.177.515.656.659.070.883.058.6
Ours0%66%(1,1.6)×65.831.71366.977.315.857.658.870.786.159.1
Ours66%0%(1.5,1)×66.632.01407.778.916.657.760.969.388.360.1
Ours66%66%(1.5,1.6)×67.832.51393.479.516.457.961.371.586.960.4
Crossget [65]FT75%(-,1.8)×58.129.11176.968.513.243.652.661.980.051.8
Prumerge [63]FT75%(-,1.7)×58.728.21135.572.413.049.855.366.382.353.6
Pixel-shuffle [10]75%75%(1.7,1.8)×63.828.81343.575.414.854.557.467.386.457.3
Ours0%75%(1,1.8)×64.531.81308.676.716.250.259.071.185.057.8
Ours75%0%(1.7,1)×63.330.71359.579.115.856.260.265.687.858.5
Ours75%75%(1.7,1.8)×65.230.71325.779.515.252.359.567.487.258.1
", + "image_path": "82298cdc84a38cb4ced8d142acc55fb5bb746ae64f9fcdb622efa4f9448cf6a2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 247, + 294, + 323 + ], + "blocks": [ + { + "bbox": [ + 55, + 204, + 555, + 237 + ], + "lines": [ + { + "bbox": [ + 55, + 204, + 555, + 237 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 555, + 237 + ], + "type": "text", + "content": "Table 3. Training with FOLDER. We plug FOLDER into LLaVA1.5-13B during training (Train Reduct>0) and inference (Infer Reduct>0). FT means finetuning on trained model (for LLaVA-Prumerge [63] and Crossget [65]). All results are reproduced/retrained using official codes/checkpoints. We highlight the best result on each benchmark." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 247, + 294, + 323 + ], + "lines": [ + { + "bbox": [ + 58, + 247, + 294, + 323 + ], + "spans": [ + { + "bbox": [ + 58, + 247, + 294, + 323 + ], + "type": "table", + "html": "
MethodReduct RatioSpeed UpMMB-V OverallMMB-V PerceptionVideo MME
Original0%1×1.041.0431.8
Turbo65%2.5×1.000.9929.8
Ours2.4×1.041.0531.9
Turbo75%2.8×0.980.9829.0
Ours2.7×1.031.0430.7
", + "image_path": "b1084fe788182f0f9e5cbd2e985a8cd7c043655e268321576a64657f27981cb2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 56, + 354, + 295, + 408 + ], + "blocks": [ + { + "bbox": [ + 111, + 327, + 240, + 338 + ], + "lines": [ + { + "bbox": [ + 111, + 327, + 240, + 338 + ], + "spans": [ + { + "bbox": [ + 111, + 327, + 240, + 338 + ], + "type": "text", + "content": "Table 4. Results on VideoLLaVA." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 354, + 295, + 408 + ], + "lines": [ + { + "bbox": [ + 56, + 354, + 295, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 295, + 408 + ], + "type": "table", + "html": "
MethodnframeMMB-V OverallVideo MMELongVideo Bench*Temp CompassMLVU
Original641.5158.756.662.569.6
Ours-75%641.4758.855.060.668.0
Ours-75%1281.4959.557.560.970.2
Ours-75%2561.5359.458.361.370.7
", + "image_path": "c323e61bf7f66af8bdc8d7c5b866ca91981807be903ed2b68ce46445756a737e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 446, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 446, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 446, + 295, + 495 + ], + "type": "text", + "content": "shows the superiority of matching over pooling. More results on MiniGPT-4v2, non-LLaVA models like Idefics2-8B and Flamingov2-7B, and point cloud understanding model PointLLMv1.2-7B are shown in the appendix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 496, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 496, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 54, + 496, + 295, + 628 + ], + "type": "text", + "content": "Multi-Vision-Tower MLLM. Multi-vision-tower-based models [32, 66, 70] enhance MLLMs' visual ability but aggravate the length issue. This makes our method particularly advantageous in this context. In Tab. 2 we evaluate on MMVP [70] with CLIP [59] and DINOv2 [58] as vision towers. By applying FOLDER, we trade a significant reduction in image token length (from 1152 to 576/384) with an acceptable drop. Note that MMVP is trained by interleaving vision tower features, which may cause positional confusion. Despite that, FOLDER still achieves a satisfying acceleration-performance trade-off." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 629, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 629, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 629, + 295, + 713 + ], + "type": "text", + "content": "Video Understanding MLLM. Video-based MLLMs [30, 40, 45] can generate sequences exceeding 2000 tokens for a few frames, limiting their applicability. As shown in Tab. 4, by applying FOLDER on Video-LLaVA, we can reduce visual tokens up to " + }, + { + "bbox": [ + 54, + 629, + 295, + 713 + ], + "type": "inline_equation", + "content": "65\\%" + }, + { + "bbox": [ + 54, + 629, + 295, + 713 + ], + "type": "text", + "content": " without any performance degradation, and up to " + }, + { + "bbox": [ + 54, + 629, + 295, + 713 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 54, + 629, + 295, + 713 + ], + "type": "text", + "content": " with only slight & acceptable drops on some benchmarks. Notably, we achieve " + }, + { + "bbox": [ + 54, + 629, + 295, + 713 + ], + "type": "inline_equation", + "content": "2.7 \\times" + }, + { + "bbox": [ + 54, + 629, + 295, + 713 + ], + "type": "text", + "content": " infer" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 315, + 247, + 561, + 392 + ], + "blocks": [ + { + "bbox": [ + 55, + 411, + 295, + 432 + ], + "lines": [ + { + "bbox": [ + 55, + 411, + 295, + 432 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 295, + 432 + ], + "type": "text", + "content": "Table 5. Results on LLaVA-OV-7B across more video benchmarks. *Acc of videos with duration > 180s." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 247, + 561, + 392 + ], + "lines": [ + { + "bbox": [ + 315, + 247, + 561, + 392 + ], + "spans": [ + { + "bbox": [ + 315, + 247, + 561, + 392 + ], + "type": "table", + "html": "
MethodReduct RatioTrainInference
Time(h)Speed UpTime(s)Speed UpMem
Original0%1871×0.3651×81.4G
FastV--0.2751.3×81.4G
Turbo50%1391.34×0.2501.5×56.6G
Ours1411.33×0.2501.5×56.6G
FastV--0.2551.4×81.4G
Turbo66%1231.52×0.2301.6×48.6G
Ours1241.51×0.2301.6×48.7G
FastV--0.2251.6×81.4G
Turbo75%1121.67×0.2051.8×44.6G
Ours1131.65×0.2051.8×44.7G
", + "image_path": "048ec043420148ea47a8c8aea3394f39f6a9c00188d8227a1c31506fa0f99ee6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 392, + 555, + 448 + ], + "lines": [ + { + "bbox": [ + 313, + 392, + 555, + 448 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 555, + 448 + ], + "type": "text", + "content": "Table 6. Acceleration For Training & Inference. We evaluate training (A100-80G GPU hours) and inference acceleration (on 1st-token) on llava1.5-13B. The memory is evaluated by its peak value on H20 with batch-size 40. Note that FastV [9] can only be used during inference." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 460, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 533 + ], + "type": "text", + "content": "ence speedup with " + }, + { + "bbox": [ + 313, + 460, + 555, + 533 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 313, + 460, + 555, + 533 + ], + "type": "text", + "content": " token reduction, still outperforming Turbo with " + }, + { + "bbox": [ + 313, + 460, + 555, + 533 + ], + "type": "inline_equation", + "content": "66\\%" + }, + { + "bbox": [ + 313, + 460, + 555, + 533 + ], + "type": "text", + "content": " reduction ratio. Furthermore, Tab. 5 shows that on LLaVA-OV-7B, with a " + }, + { + "bbox": [ + 313, + 460, + 555, + 533 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 313, + 460, + 555, + 533 + ], + "type": "text", + "content": " reduction ratio within each frame, using only half of the visual tokens " + }, + { + "bbox": [ + 313, + 460, + 555, + 533 + ], + "type": "inline_equation", + "content": "(128\\times 25\\%)" + }, + { + "bbox": [ + 313, + 460, + 555, + 533 + ], + "type": "text", + "content": " compared to the original setup, we achieve even better performance, especially on long video understanding." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 540, + 484, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 484, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 484, + 553 + ], + "type": "text", + "content": "4.2.2. Training Acceleration on MLLM" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 556, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 555, + 689 + ], + "type": "text", + "content": "As a universal plug-and-play module for MLLM, FOLDER can also be integrated into MLLM pretraining and SFT. Results on LLaVA1.5-13B are in Tab. 3. Surprisingly, when incorporating FOLDER with a " + }, + { + "bbox": [ + 313, + 556, + 555, + 689 + ], + "type": "inline_equation", + "content": "66\\%" + }, + { + "bbox": [ + 313, + 556, + 555, + 689 + ], + "type": "text", + "content": " reduction ratio, LLaVA-1.5-13B exhibits comprehensive improvements across all benchmarks, even including visually dense tasks such as OCRVQA (" + }, + { + "bbox": [ + 313, + 556, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\uparrow 3\\%" + }, + { + "bbox": [ + 313, + 556, + 555, + 689 + ], + "type": "text", + "content": ") and InfoVQA (" + }, + { + "bbox": [ + 313, + 556, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\uparrow 0.9\\%" + }, + { + "bbox": [ + 313, + 556, + 555, + 689 + ], + "type": "text", + "content": "). This suggests high-frequency noise in visual sequences can be mitigated by merging similar visual tokens. Moreover, [9] highlights attention sparsity in LLMs' visual tokens, reinforcing the need to shorten visual sequences for better focus." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Furthermore, training on reduced sequences maintains flexibility during inference, as the original sequence can still" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23620" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 68, + 286, + 182 + ], + "blocks": [ + { + "bbox": [ + 65, + 68, + 286, + 182 + ], + "lines": [ + { + "bbox": [ + 65, + 68, + 286, + 182 + ], + "spans": [ + { + "bbox": [ + 65, + 68, + 286, + 182 + ], + "type": "table", + "html": "
MethodReduct RatioThroughput im/sSpeed UpCIDErB@4
Original0%22.11×133.339.7
Uniform40.21.82×128.837.9
Last-360%38.41.74×129.538.2
Last-237.91.71×129.438.2
Last-137.51.71×132.239.0
Uniform47.42.14×127.237.5
Last-370%44.22.00×128.438.2
Last-243.71.98×130.138.4
Last-143.41.96×131.038.9
", + "image_path": "4c6a30fe890b4c5968de7b3afa8171b943c23ab008f19216b4fe48678cd0de8a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 233, + 304, + 306 + ], + "blocks": [ + { + "bbox": [ + 55, + 185, + 295, + 218 + ], + "lines": [ + { + "bbox": [ + 55, + 185, + 295, + 218 + ], + "spans": [ + { + "bbox": [ + 55, + 185, + 295, + 218 + ], + "type": "text", + "content": "Table 7. Ablation on Merging Position. We evaluate BLIP [38] on captioning task. Last- " + }, + { + "bbox": [ + 55, + 185, + 295, + 218 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 185, + 295, + 218 + ], + "type": "text", + "content": " refers to uniform reduction in last " + }, + { + "bbox": [ + 55, + 185, + 295, + 218 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 185, + 295, + 218 + ], + "type": "text", + "content": " blocks, uniform refers to uniform reduction in every block." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 233, + 304, + 306 + ], + "lines": [ + { + "bbox": [ + 57, + 233, + 304, + 306 + ], + "spans": [ + { + "bbox": [ + 57, + 233, + 304, + 306 + ], + "type": "table", + "html": "
MethodMMBPOPEMMEGQASciQA
VILA1.5-13b [46]74.384.91433.764.578.9
+FOLDER-50%72.884.01445.362.078.0
VITA1.5-7b [24]80.989.11728.947.895.8
+FOLDER-50%78.188.81688.350.894.1
WePOINTS1.5-7b [53]84.086.81616.059.295.0
+FOLDER-50%83.487.01642.659.194.3
", + "image_path": "98ec0ee9c5b8e3175037f518bb33215acdc7e77f2edc4f4decf78e0fade7a19e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 334, + 295, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 334, + 295, + 382 + ], + "spans": [ + { + "bbox": [ + 55, + 334, + 295, + 382 + ], + "type": "text", + "content": "be used, leading to improved results on visually demanding tasks. Similar to masked augmentation [28, 43], FOLDER acts as a regularizer but goes beyond random dropping by accelerating both training and inference." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 384, + 295, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 295, + 456 + ], + "type": "text", + "content": "Compared to Pyramidddrop [75], FOLDER achieves better performance on most benchmarks. It also outperforms finetuning-based methods [63, 65] even without additional training. By reducing tokens before the LLM, FOLDER serves as a flexible alternative to training-phase modules like pixel-shuffle, enhancing both efficiency and scalability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 457, + 295, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 457, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 457, + 295, + 578 + ], + "type": "text", + "content": "Speed & Memory Evaluation. In Tab. 6, we evaluate FOLDER's acceleration performance on LLaVA1.5-13B during training and inference. By reducing " + }, + { + "bbox": [ + 55, + 457, + 295, + 578 + ], + "type": "inline_equation", + "content": "66\\%" + }, + { + "bbox": [ + 55, + 457, + 295, + 578 + ], + "type": "text", + "content": " of tokens, FOLDER achieves " + }, + { + "bbox": [ + 55, + 457, + 295, + 578 + ], + "type": "inline_equation", + "content": "1.5 \\times" + }, + { + "bbox": [ + 55, + 457, + 295, + 578 + ], + "type": "text", + "content": " faster training on " + }, + { + "bbox": [ + 55, + 457, + 295, + 578 + ], + "type": "inline_equation", + "content": "8 \\times \\mathrm{A}100 - 80\\mathrm{G}" + }, + { + "bbox": [ + 55, + 457, + 295, + 578 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 457, + 295, + 578 + ], + "type": "inline_equation", + "content": "1.6 \\times" + }, + { + "bbox": [ + 55, + 457, + 295, + 578 + ], + "type": "text", + "content": " speedup during inference on H20-96G. Unlike FastV, which prunes tokens within the LLM, FOLDER removes them beforehand, significantly lowering peak memory usage. With a batch size of 40 on H20-96G, it saves " + }, + { + "bbox": [ + 55, + 457, + 295, + 578 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 55, + 457, + 295, + 578 + ], + "type": "text", + "content": " memory while maintaining a " + }, + { + "bbox": [ + 55, + 457, + 295, + 578 + ], + "type": "inline_equation", + "content": "66\\%" + }, + { + "bbox": [ + 55, + 457, + 295, + 578 + ], + "type": "text", + "content": " reduction ratio, enabling nearly " + }, + { + "bbox": [ + 55, + 457, + 295, + 578 + ], + "type": "inline_equation", + "content": "3 \\times" + }, + { + "bbox": [ + 55, + 457, + 295, + 578 + ], + "type": "text", + "content": " larger batches—especially beneficial for training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 589, + 285, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 589, + 285, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 589, + 285, + 601 + ], + "type": "text", + "content": "4.2.3. Simplified FOLDER for Inference Acceleration" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "As illustrated in Fig. 5, FOLDER is inserted between ATTN and FFN, requiring vision encoder modifications. Since the best results are achieved by reducing tokens only in the last block (Sec. 3.4), a simpler approach is inserting FOLDER at the output. Tab. 8 shows results on 3 SOTA MLLMs (VILA1.5 [46], WePOINTS1.5 [53], VITA1.5 [24]) with " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " reduction ratio and " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "1.4\\times" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " acceleration, while maintaining or surpassing original performance on several benchmarks. More comparisons are in the appendix." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 345, + 65, + 523, + 160 + ], + "blocks": [ + { + "bbox": [ + 63, + 309, + 288, + 320 + ], + "lines": [ + { + "bbox": [ + 63, + 309, + 288, + 320 + ], + "spans": [ + { + "bbox": [ + 63, + 309, + 288, + 320 + ], + "type": "text", + "content": "Table 8. FOLDER training-free plug-in for more MLLMs." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 345, + 65, + 523, + 160 + ], + "lines": [ + { + "bbox": [ + 345, + 65, + 523, + 160 + ], + "spans": [ + { + "bbox": [ + 345, + 65, + 523, + 160 + ], + "type": "table", + "html": "
MethodReduct RatioB@4CIDErSPICE
Original0%133.339.723.8
Avg132.039.023.6
Drop50%130.038.723.3
Weighted132.239.123.6
Avg131.038.923.5
Drop60%128.938.223.1
Weighted131.539.123.4
", + "image_path": "0c43f90833358ae7c2f4475b98265af180c8d8232dbe90eb7b01b743b223c18d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 163, + 553, + 208 + ], + "lines": [ + { + "bbox": [ + 313, + 163, + 553, + 208 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 553, + 208 + ], + "type": "text", + "content": "Table 9. Ablation Study on Aggregation Operation. On BLIP [38] captioning task, we test the performance of 3 aggregation operations discussed in Sec. 3.3. For fair comparison, we adopt the aggregation only in the last block." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 218, + 408, + 231 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 218, + 408, + 231 + ], + "spans": [ + { + "bbox": [ + 313, + 218, + 408, + 231 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 236, + 553, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 553, + 272 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 553, + 272 + ], + "type": "text", + "content": "We conduct ablation studies on BLIP [38] for image captioning to validate the merging position and aggregation methods in Sec. 3. More ablations are in the appendix." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 273, + 554, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 273, + 554, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 554, + 392 + ], + "type": "text", + "content": "Merging Position. Preliminary ViT tests in (Fig. 2, Fig. 3) show that due to propagation effects (Sec. 3.2) and diminishing reduction impact (Sec. 3.1), merging more tokens in the last blocks is preferable. Based on this, we test four strategies: uniform merging across all blocks and in the last " + }, + { + "bbox": [ + 313, + 273, + 554, + 392 + ], + "type": "inline_equation", + "content": "1/2/3" + }, + { + "bbox": [ + 313, + 273, + 554, + 392 + ], + "type": "text", + "content": " blocks. Tab. 7 presents the results which confirm that merging in the last block yields the best performance. While the vision encoder's acceleration slightly affects BLIP, in MLLMs, where LLM processing dominates " + }, + { + "bbox": [ + 313, + 273, + 554, + 392 + ], + "type": "inline_equation", + "content": "(10 - 15 \\times" + }, + { + "bbox": [ + 313, + 273, + 554, + 392 + ], + "type": "text", + "content": " inferior), merging position has minimal impact on speed-up." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 393, + 554, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 554, + 476 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 554, + 476 + ], + "type": "text", + "content": "Aggregation Method. In Sec. 3.3, we explore three distinct aggregation methods: Dropping, Average, and Weighted Average based on vector norms. As illustrated in Tab. 9, both averaging methods outperform direct dropping, as directly dropping tokens causes greater information loss. To reduce module complexity, we used averaging in all MLLM experiments, as the differences are minimal." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 489, + 388, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 489, + 388, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 489, + 388, + 502 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 510, + 554, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 554, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 554, + 641 + ], + "type": "text", + "content": "We introduce FOLDER, a plug-and-play module for efficient visual token reduction in MLLMs. Through empirical analysis of the token reduction process, we develop a strategy that performs aggressive token reduction in the vision backbone's final layers while preserving key information. Our method achieves over " + }, + { + "bbox": [ + 313, + 510, + 554, + 641 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 313, + 510, + 554, + 641 + ], + "type": "text", + "content": " visual token reduction with " + }, + { + "bbox": [ + 313, + 510, + 554, + 641 + ], + "type": "inline_equation", + "content": "1.8 \\times" + }, + { + "bbox": [ + 313, + 510, + 554, + 641 + ], + "type": "text", + "content": " inference speedup and " + }, + { + "bbox": [ + 313, + 510, + 554, + 641 + ], + "type": "inline_equation", + "content": "1.65 \\times" + }, + { + "bbox": [ + 313, + 510, + 554, + 641 + ], + "type": "text", + "content": " training acceleration. FOLDER maintains or improves performance across tasks including video understanding, demonstrating its effectiveness in making MLLMs more computationally efficient without capability loss." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": "Acknowledgements. Part of this work was funded by Hi!PARIS Center on Data Analytics and Artificial Intelligence, by the European Union's Horizon Europe Research and Innovation Programme under grant agreement No. 101120237 (ELIAS), and by the French National Research Agency (ANR) in the framework of the JCJC project \"BANERA\" under Grant ANR-24-CE23-4369." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23621" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 146, + 294, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 146, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 294, + 167 + ], + "type": "text", + "content": "[2] Anthropic. The claude 3 model family: Opus, sonnet, haiku, 2024. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 170, + 295, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 170, + 295, + 223 + ], + "spans": [ + { + "bbox": [ + 62, + 170, + 295, + 223 + ], + "type": "text", + "content": "[3] Kirolos Ataallah, Xiaoqian Shen, Eslam Abdelrahman, Essam Sleiman, Mingchen Zhuge, Jian Ding, Deyao Zhu, Jürgen Schmidhuber, and Mohamed Elhoseiny. Goldfish: Vision-language understanding of arbitrarily long videos, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 224, + 295, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 224, + 295, + 290 + ], + "spans": [ + { + "bbox": [ + 62, + 224, + 295, + 290 + ], + "type": "text", + "content": "[4] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. Openflamingo: An opensource framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 291, + 294, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 291, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 62, + 291, + 294, + 335 + ], + "type": "text", + "content": "[5] Adam Block, Dylan J Foster, Akshay Krishnamurthy, Max Simchowitz, and Cyril Zhang. Butterfly effects of sgd noise: Error amplification in behavior cloning and autoregression. arXiv preprint arXiv:2310.11428, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 336, + 295, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 336, + 295, + 379 + ], + "spans": [ + { + "bbox": [ + 62, + 336, + 295, + 379 + ], + "type": "text", + "content": "[6] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461, 2022. 2, 4, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 380, + 295, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 295, + 446 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 295, + 446 + ], + "type": "text", + "content": "[7] Jun Chen, Deyao Zhu, Xiaoqian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023.2,5,6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 447, + 295, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 295, + 501 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 295, + 501 + ], + "type": "text", + "content": "[8] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024.5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 502, + 295, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 295, + 556 + ], + "type": "text", + "content": "[9] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. arXiv preprint arXiv:2403.06764, 2024. 2, 5, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 558, + 295, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 558, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 57, + 558, + 295, + 623 + ], + "type": "text", + "content": "[10] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhang-wei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. Science China Information Sciences, 67(12):220101, 2024. 2, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 624, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 690 + ], + "type": "text", + "content": "[11] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "text", + "content": "[12] Kanzhi Cheng, Wenpo Song, Zheng Ma, Wenhao Zhu, Zixuan Zhu, and Jianbing Zhang. Beyond generic: Enhancing" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 554, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 554, + 117 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 554, + 117 + ], + "type": "text", + "content": "image captioning with real-world knowledge using vision-language pre-training model. In Proceedings of the 31st ACM International Conference on Multimedia, pages 5038-5047, 2023. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 119, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 119, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 119, + 553, + 140 + ], + "type": "text", + "content": "[13] Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 143, + 553, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 143, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 316, + 143, + 553, + 175 + ], + "type": "text", + "content": "[14] Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness, 2022. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 177, + 553, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 177, + 553, + 221 + ], + "spans": [ + { + "bbox": [ + 316, + 177, + 553, + 221 + ], + "type": "text", + "content": "[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 223, + 553, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 223, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 316, + 223, + 553, + 255 + ], + "type": "text", + "content": "[16] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 258, + 553, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 258, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 316, + 258, + 553, + 323 + ], + "type": "text", + "content": "[17] Haodong Duan, Junming Yang, Yuxuan Qiao, Xinyu Fang, Lin Chen, Yuan Liu, Xiaoyi Dong, Yuhang Zang, Pan Zhang, Jiaqi Wang, et al. Vlmevalkit: An open-source toolkit for evaluating large multi-modality models. In Proceedings of the 32nd ACM international conference on multimedia, pages 11198-11201, 2024. 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 325, + 553, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 325, + 553, + 357 + ], + "spans": [ + { + "bbox": [ + 316, + 325, + 553, + 357 + ], + "type": "text", + "content": "[18] Carl Eckart and Gale Young. The approximation of one matrix by another of lower rank. Psychometrika, 1(3):211-218, 1936. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 359, + 553, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 359, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 316, + 359, + 553, + 404 + ], + "type": "text", + "content": "[19] Xinyu Fang, Kangrui Mao, Haodong Duan, Xiangyu Zhao, Yining Li, Dahua Lin, and Kai Chen. Mmbench-video: A long-form multi-shot benchmark for holistic video understanding. arXiv preprint arXiv:2406.14515, 2024. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 406, + 553, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 406, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 316, + 406, + 553, + 449 + ], + "type": "text", + "content": "[20] Zhiyuan Fang, Jianfeng Wang, Xiaowei Hu, Lijuan Wang, Yezhou Yang, and Zicheng Liu. Compressing visual-linguistic model via knowledge distillation. In ICCV, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 452, + 553, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 452, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 316, + 452, + 553, + 495 + ], + "type": "text", + "content": "[21] Elias Frantar, Saleh Ashkboos, Torsten Hoefler, and Dan Alistarh. Gptq: Accurate post-training quantization for generative pre-trained transformers. arXiv preprint arXiv:2210.17323, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 497, + 553, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 497, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 316, + 497, + 553, + 552 + ], + "type": "text", + "content": "[22] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 554, + 553, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 554, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 316, + 554, + 553, + 609 + ], + "type": "text", + "content": "[23] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 610, + 553, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 610, + 553, + 665 + ], + "spans": [ + { + "bbox": [ + 316, + 610, + 553, + 665 + ], + "type": "text", + "content": "[24] Chaoyou Fu, Haojia Lin, Xiong Wang, Yi-Fan Zhang, Yunhang Shen, Xiaoyu Liu, Yangze Li, Zuwei Long, Heting Gao, Ke Li, et al. Vita-1.5: Towards gpt-4o level real-time vision and speech interaction. arXiv preprint arXiv:2501.01957, 2025. 5, 6, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 667, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 667, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 667, + 553, + 689 + ], + "type": "text", + "content": "[25] Grok. Grok1.5 homepage. https://x.ai/blog/grok-1.5v, 2024. Accessed: 2024-11-14. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "text", + "content": "[26] Tianrui Guan, Fuxiao Liu, Xiyang Wu, Ruiqi Xian, Zongxia Li, Xiaoyu Liu, Xijun Wang, Lichang Chen, Furong Huang," + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23622" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 127 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 127 + ], + "type": "text", + "content": "Yaser Yacoob, et al. Hallusionbench: an advanced diagnostic suite for entangled language hallucination and visual illusion in large vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14375-14385, 2024. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 295, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 295, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 295, + 183 + ], + "type": "text", + "content": "[27] Yuhang Han, Xuyang Liu, Pengxiang Ding, Donglin Wang, Honggang Chen, Qingsen Yan, and Siteng Huang. Rethinking token reduction in mllms: Towards a unified paradigm for training-free acceleration. arXiv preprint arXiv:2411.17686, 2024. 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 184, + 295, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 295, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 295, + 237 + ], + "type": "text", + "content": "[28] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 239, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 293 + ], + "type": "text", + "content": "[29] Wenxuan Huang, Zijie Zhai, Yunhang Shen, Shaoshen Cao, Fei Zhao, Xiangfeng Xu, Zheyu Ye, and Shaohui Lin. Dynamic-llava: Efficient multimodal large language models via dynamic vision-language context sparsification. arXiv preprint arXiv:2412.00876, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 294, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 294, + 358 + ], + "type": "text", + "content": "[30] Peng Jin, Ryuichi Takanobu, Wancai Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13700-13710, 2024. 1, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 361, + 294, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 294, + 424 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 294, + 424 + ], + "type": "text", + "content": "[31] Chen Ju, Haicheng Wang, Haozhe Cheng, Xu Chen, Zhonghua Zhai, Weilin Huang, Jinsong Lan, Shuai Xiao, and Bo Zheng. Turbo: Informativity-driven acceleration plug-in for vision-language large models. In European Conference on Computer Vision, pages 436-455. Springer, 2025. 2, 4, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 426, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 470 + ], + "type": "text", + "content": "[32] Oguzhan Fatih Kar, Alessio Tonioni, Petra Poklukar, Achin Kulshrestha, Amir Zamir, and Federico Tombari. Brave: Broadening the visual encoding of vision-language models. arXiv preprint arXiv:2404.07204, 2024. 1, 2, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 471, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 536 + ], + "type": "text", + "content": "[33] Zaid Khan, Vijay Kumar BG, Samuel Schulter, Xiang Yu, Yun Fu, and Manmohan Chandraker. Q: How to specialize large vision-language models to data-scarce vqa tasks? a: Self-train on unlabeled images! In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15005-15015, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 537, + 294, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 581 + ], + "type": "text", + "content": "[34] Zhenglun Kong, Peiyan Dong, Xiaolong Ma, Xin Meng, Mengshu Sun, Wei Niu, Xuan Shen, Geng Yuan, Bin Ren, Minghai Qin, Hao Tang, and Yanzhi Wang. Spvit: Enabling faster vision transformers via soft token pruning, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 582, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 294, + 624 + ], + "type": "text", + "content": "[35] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? Advances in Neural Information Processing Systems, 37:87874-87907, 2024. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 625, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 668 + ], + "type": "text", + "content": "[36] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[37] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 5" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[38] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 4, 5, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 129, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 182 + ], + "type": "text", + "content": "[39] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 183, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 183, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 316, + 183, + 553, + 227 + ], + "type": "text", + "content": "[40] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 1, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 228, + 553, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 261 + ], + "type": "text", + "content": "[41] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and performant baseline for vision and language, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 262, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 553, + 304 + ], + "type": "text", + "content": "[42] Wentong Li, Yuqian Yuan, Jian Liu, Dongqi Tang, Song Wang, Jie Qin, Jianke Zhu, and Lei Zhang. Tokenpacker: Efficient visual projector for multimodal llm. arXiv preprint arXiv:2407.02392, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 304, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 553, + 358 + ], + "type": "text", + "content": "[43] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pre-training via masking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 23390-23400, 2023. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 360, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 360, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 360, + 553, + 403 + ], + "type": "text", + "content": "[44] Youwei Liang, Chongjian Ge, Zhan Tong, Yibing Song, Jue Wang, and Pengtao Xie. Not all patches are what you need: Expediting vision transformers via token reorganizations, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 404, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 448 + ], + "type": "text", + "content": "[45] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 1, 2, 5, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 449, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 502 + ], + "type": "text", + "content": "[46] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 26689-26699, 2024. 5, 6, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 503, + 553, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 503, + 553, + 570 + ], + "spans": [ + { + "bbox": [ + 316, + 503, + 553, + 570 + ], + "type": "text", + "content": "[47] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 571, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 571, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 316, + 571, + 553, + 592 + ], + "type": "text", + "content": "[48] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 593, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 553, + 624 + ], + "type": "text", + "content": "[49] Hao Liu, Matei Zaharia, and Pieter Abbeel. Ring attention with blockwise transformers for near-infinite context, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "type": "text", + "content": "[50] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 2, 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[51] Ting Liu, Liangtao Shi, Richang Hong, Yue Hu, Quanjun Yin, and Linfeng Zhang. Multi-stage vision token dropping: Towards efficient multimodal large language model. arXiv preprint arXiv:2411.10803, 2024. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23623" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[52] Yuanxin Liu, Shicheng Li, Yi Liu, Yuxiang Wang, Shuhuai Ren, Lei Li, Sishuo Chen, Xu Sun, and Lu Hou. Tempcompass: Do video llms really understand videos? arXiv preprint arXiv:2403.00476, 2024. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 119, + 294, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 119, + 294, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 119, + 294, + 163 + ], + "type": "text", + "content": "[53] Yuan Liu, Le Tian, Xiao Zhou, Xinyu Gao, Kavio Yu, Yang Yu, and Jie Zhou. Points1. 5: Building a vision-language model towards real world applications. arXiv preprint arXiv:2412.08443, 2024. 5, 6, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 165, + 294, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 165, + 294, + 209 + ], + "spans": [ + { + "bbox": [ + 56, + 165, + 294, + 209 + ], + "type": "text", + "content": "[54] Zhijian Liu, Ligeng Zhu, Baifeng Shi, Zhuoyang Zhang, Yuming Lou, Shang Yang, Haocheng Xi, Shiyi Cao, Yuxian Gu, Dacheng Li, et al. Nvila: Efficient frontier visual language models. arXiv preprint arXiv:2412.04468, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 211, + 294, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 211, + 294, + 276 + ], + "spans": [ + { + "bbox": [ + 56, + 211, + 294, + 276 + ], + "type": "text", + "content": "[55] Pan Lu, Swaroop Mishra, Tony Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In The 36th Conference on Neural Information Processing Systems (NeurIPS), 2022. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 279, + 294, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 279, + 294, + 323 + ], + "spans": [ + { + "bbox": [ + 56, + 279, + 294, + 323 + ], + "type": "text", + "content": "[56] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1697-1706, 2022. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 325, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 325, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 325, + 294, + 358 + ], + "type": "text", + "content": "[57] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering by reading text in images. In ICDAR, 2019. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 360, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 294, + 415 + ], + "type": "text", + "content": "[58] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 417, + 294, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 417, + 294, + 482 + ], + "spans": [ + { + "bbox": [ + 56, + 417, + 294, + 482 + ], + "type": "text", + "content": "[59] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 485, + 294, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 485, + 294, + 528 + ], + "spans": [ + { + "bbox": [ + 56, + 485, + 294, + 528 + ], + "type": "text", + "content": "[60] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. NeurIPS, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 531, + 294, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 531, + 294, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 531, + 294, + 563 + ], + "type": "text", + "content": "[61] Yossi Rubner, Carlo Tomasi, and Leonidas J Guibas. The earth mover's distance as a metric for image retrieval. International journal of computer vision, 40:99-121, 2000. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 567, + 294, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 294, + 620 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 294, + 620 + ], + "type": "text", + "content": "[62] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge. In European conference on computer vision, pages 146–162. Springer, 2022. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 623, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 623, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 56, + 623, + 294, + 666 + ], + "type": "text", + "content": "[63] Yuzhang Shang, Mu Cai, Bingxin Xu, Yong Jae Lee, and Yan Yan. Llava-prumerge: Adaptive token reduction for efficient large multimodal models. arXiv preprint arXiv:2403.15388, 2024. 2, 4, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "type": "text", + "content": "[64] Dachuan Shi, Chaofan Tao, Ying Jin, Zhendong Yang, Chun Yuan, and Jiaqi Wang. Upop: Unified and progressive pruning for compressing vision-language transformers. arXiv preprint arXiv:2301.13741, 2023. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "text", + "content": "[65] Dachuan Shi, Chaofan Tao, Anyi Rao, Zhendong Yang, Chun Yuan, and Jiaqi Wang. Crossget: Cross-guided ensemble of tokens for accelerating vision-language transformers. arXiv preprint arXiv:2305.17455, 2023. 2, 5, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 173 + ], + "type": "text", + "content": "[66] Min Shi, Fuxiao Liu, Shihao Wang, Shijia Liao, Subhashree Radhakrishnan, De-An Huang, Hongxu Yin, Karan Sapra, Yaser Yacoob, Humphrey Shi, et al. Eagle: Exploring the design space for multimodal llms with mixture of encoders. arXiv preprint arXiv:2408.15998, 2024. 1, 2, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "type": "text", + "content": "[67] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. FLAVA: A foundational language and vision alignment model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 230, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 296 + ], + "type": "text", + "content": "[68] Hao Tan and Mohit Bansal. LXMERT: Learning cross-modality encoder representations from transformers. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, EMNLP-IJCNLP, November 3-7, 2019, 2019." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 298, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 553, + 352 + ], + "type": "text", + "content": "[69] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 354, + 555, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 555, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 555, + 408 + ], + "type": "text", + "content": "[70] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9568-9578, 2024. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 411, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 464 + ], + "type": "text", + "content": "[71] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 466, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 466, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 466, + 553, + 510 + ], + "type": "text", + "content": "[72] Tiannan Wang, Wangchunshu Zhou, Yan Zeng, and Xinsong Zhang. Efficientvlm: Fast and accurate vision-language models via knowledge distillation and modal-adaptive pruning. arXiv preprint arXiv:2210.07795, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 512, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 553, + 555 + ], + "type": "text", + "content": "[73] Haoning Wu, Dongxu Li, Bei Chen, and Junnan Li. Longvideobench: A benchmark for long-context interleaved video-language understanding. Advances in Neural Information Processing Systems, 37:28828-28857, 2024. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 557, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 553, + 610 + ], + "type": "text", + "content": "[74] Guangxuan Xiao, Ji Lin, Mickael Seznec, Hao Wu, Julien Demouth, and Song Han. Smoothquant: Accurate and efficient post-training quantization for large language models. In International Conference on Machine Learning, pages 38087-38099. PMLR, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 613, + 553, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 677 + ], + "type": "text", + "content": "[75] Long Xing, Qidong Huang, Xiaoyi Dong, Jiajie Lu, Pan Zhang, Yuhang Zang, Yuhang Cao, Conghui He, Jiaqi Wang, Feng Wu, et al. Pyramiddrop: Accelerating your large vision-language models via pyramid visual redundancy reduction. arXiv preprint arXiv:2410.17247, 2024. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[76] Runsen Xu, Xiaolong Wang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Pointllm: Empowering large language models to understand point clouds. In European" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23624" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 631 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "Conference on Computer Vision, pages 131-147. Springer, 2024. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[77] Senqiao Yang, Yukang Chen, Zhuotao Tian, Chengyao Wang, Jingyao Li, Bei Yu, and Jiaya Jia. Visionzip: Longer is better but not necessary in vision language models. arXiv preprint arXiv:2412.04467, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 195 + ], + "type": "text", + "content": "[78] Xubing Ye, Yukang Gan, Yixiao Ge, Xiao-Ping Zhang, and Yansong Tang. Atp-llava: Adaptive token pruning for large vision language models. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 24972–24982, 2025. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "type": "text", + "content": "[79] Xubing Ye, Yukang Gan, Xiaoke Huang, Yixiao Ge, and Yansong Tang. Voco-llama: Towards vision compression with large language models. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 29836–29846, 2025. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 252, + 294, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 252, + 294, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 252, + 294, + 306 + ], + "type": "text", + "content": "[80] Li Yifan, Du Yifan, Zhou Kun, Wang Jinpeng, Xin Zhao Wayne, and Wen Ji-Rong. Evaluating object hallucination in large vision-language models. In The 2023 Conference on Empirical Methods in Natural Language Processing, 2023. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 308, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 294, + 363 + ], + "type": "text", + "content": "[81] Liu Yuan, Duan Haodong, Zhang Yuanhan, Li Bo, Zhang Songyang, Zhao Wangbo, Yuan Yike, Wang Jiaqi, He Conghui, Liu Ziwei, Chen Kai, and Lin Dahua. Mmbench: Is your multi-modal model an all-around player? arXiv:2307.06281, 2023. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 294, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 451 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 451 + ], + "type": "text", + "content": "[82] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Ren-liang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of CVPR, 2024. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 453, + 294, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 453, + 294, + 485 + ], + "spans": [ + { + "bbox": [ + 56, + 453, + 294, + 485 + ], + "type": "text", + "content": "[83] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 487, + 294, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 487, + 294, + 541 + ], + "spans": [ + { + "bbox": [ + 56, + 487, + 294, + 541 + ], + "type": "text", + "content": "[84] Yuan Zhang, Chun-Kai Fan, Junpeng Ma, Wenzhao Zheng, Tao Huang, Kuan Cheng, Denis Gudovskiy, Tomoyuki Okuno, Yohei Nakata, Kurt Keutzer, et al. Sparsevm: Visual token sparsification for efficient vision-language model inference. arXiv preprint arXiv:2410.04417, 2024. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 543, + 294, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 543, + 294, + 575 + ], + "spans": [ + { + "bbox": [ + 56, + 543, + 294, + 575 + ], + "type": "text", + "content": "[85] Yiwu Zhong, Zhuoming Liu, Yin Li, and Liwei Wang. Aim: Adaptive inference of multi-modal llms via token merging and pruning. arXiv preprint arXiv:2412.03248, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 576, + 294, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 576, + 294, + 631 + ], + "spans": [ + { + "bbox": [ + 56, + 576, + 294, + 631 + ], + "type": "text", + "content": "[86] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. MLvu: A comprehensive benchmark for multitask long video understanding. arXiv e-prints, pages arXiv-2406, 2024. 6" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23625" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/FPEM_ Face Prior Enhanced Facial Attractiveness Prediction for Live Videos with Face Retouching/f47f8d6d-8a6c-491b-8f19-08123eb17e5c_content_list.json b/2025/FPEM_ Face Prior Enhanced Facial Attractiveness Prediction for Live Videos with Face Retouching/f47f8d6d-8a6c-491b-8f19-08123eb17e5c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d80597c9fdf9bf1bd826ec10a505fd27278268b6 --- /dev/null +++ b/2025/FPEM_ Face Prior Enhanced Facial Attractiveness Prediction for Live Videos with Face Retouching/f47f8d6d-8a6c-491b-8f19-08123eb17e5c_content_list.json @@ -0,0 +1,1783 @@ +[ + { + "type": "text", + "text": "FPEM: Face Prior Enhanced Facial Attractiveness Prediction for Live Videos with Face Retouching", + "text_level": 1, + "bbox": [ + 187, + 130, + 810, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hui Li $^{1\\dagger}$ , Xiaoyu Ren $^{1,2\\dagger}$ , Hongjiu Yu $^{1}$ , Ying Chen $^{1*}$ , Kai Li $^{1}$ , L Wang $^{1}$ , Xiongkuo Min $^{2}$ , Huiyu Duan $^{2}$ , Guangtao Zhai $^{2}$ , Xu Liu $^{1}$", + "bbox": [ + 232, + 202, + 759, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Alibaba Group 2Shanghai Jiao Tong University", + "bbox": [ + 352, + 239, + 643, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{lh401704,naomi.rxy,yuhongjiu.yhj,YingChen,kaishi.lk,zhuangshu.wlb,steven.liu}@taobao.com", + "bbox": [ + 104, + 258, + 890, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{minxiongkuo,huiyuduan,zhaiguangtao}@sjtu.edu.cn", + "bbox": [ + 284, + 277, + 707, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Facial attractiveness prediction (FAP) has long been an important computer vision task, which could be widely applied in live videos with facial retouching. However, previous FAP datasets are either small or closed-source. Moreover, the corresponding FAP models exhibit limited generalization and adaptation ability. To overcome these limitations, we introduce the first large-scale FAP dataset LiveBeauty specifically designed for live video scenarios wherein face images may be real-time processed for aesthetics purposes. 10,000 face images are collected directly from a live streaming platform, with 200,000 corresponding attractiveness annotations obtained from a well-devised subjective experiment, making LiveBeauty the largest open-access FAP dataset. Based on the built dataset, a novel FAP method named Facial Prior Enhanced Multi-modal model (FPEM) is proposed to measure the attractiveness of facial images. Extensive experiments conducted on both LiveBeauty and other open-source FAP datasets demonstrate that our proposed method achieves state-of-the-art performance. The dataset will be released at https://github.com/EstellaLH/FPEM.", + "bbox": [ + 89, + 357, + 485, + 674 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 704, + 220, + 720 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the rapid development of live streaming, short-form video, and advanced face retouching within such applications, the perceptual quality of faces becomes an important factor for the overall visual experience of users, which is particularly essential for video platforms, as it can significantly affect the average viewer watch time and even GMV (Gross Merchandise Volume). A robust metric system for evaluating facial clarity and aesthetics is essential to monitor both the overall service quality of the platform and", + "bbox": [ + 89, + 729, + 483, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c0f4409e3000fe150b3c0f06d40319855c6edc3c7b11e9f8b00c7d4d939d8af5.jpg", + "image_caption": [ + "Figure 1. SROCC vs PLCC comparison on LiveBeauty and SCUT-FBP5500 [27]. FPEM, denoted by red triangular, outperforms other methods (CNN-ER [2], 2D-FAP [32], REX-INCEP [2], Inception-V3 [49], ResNeXt-50 [56]) on both datasets." + ], + "image_footnote": [], + "bbox": [ + 516, + 333, + 705, + 464 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9ef842e72aa080ba2473e92ee4f8bded9293a0e5964df33b70d5ea313b2a5ef7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 333, + 901, + 464 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the visual appeal changes or improvements of individual streaming channels. The clarity-related assessment tasks of facial quality could be categorized into Biometric Face Image Quality Assessment (BFIQA) and Generic Face Image Quality Assessment (GFIQA). BFIQA focuses on the identity information of faces for down-stream tasks such as face recognition [1, 37, 40], while GFIQA emphasizes the visual perceptual quality of faces [6, 47, 52]. The aesthetics-related assessment of facial quality is often referred to as Facial Attractiveness Prediction (FAP) [12, 27, 28]. The goal of FAP is to quantify the attractiveness of facial images, which is vital for assessing the effectiveness of facial retouching settings and offers valuable insights for algorithm development.", + "bbox": [ + 509, + 566, + 906, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As shown in Tab. 1, numerous FAP datasets have been proposed in recent years, many of which encompass a rich diversity in age, gender, face expression and pose [9, 17, 24, 27, 53, 61, 63, 64]. However, most of these datasets are either small or unavailable due to privacy and ethical concerns. In video streaming, especially live video, it is common practice to use facial retouching algorithms to improve the facial aesthetics of content presented to audiences.", + "bbox": [ + 509, + 779, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 107, + 875, + 235, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Equal contributions.", + "bbox": [ + 109, + 887, + 223, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11458", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/ca3069fa4b8f337408f5304b04c1b63d80beffaee6d275c4d0c6e61fc47a7f17.jpg", + "table_caption": [ + "Table 1. Overview of common facial attractiveness prediction datasets." + ], + "table_footnote": [], + "table_body": "
DatasetYearRating Scale# ImageGenderEthnicity# AnnotatorPublic
Gray et al. [17]2010[-3, 3]2,056F MCaucasian30✓
MBW+FBW [61]2014[1, 5]10,400F MMulti30-
De Vries et al. [9]2015Binary9364F MAsian20-
LSFBD [63]2016[1, 5]20,000F M---
SCUT-FBP5500 [27]2018[1, 5]5,500F MMulti60✓
Tong et al. [53]2020Binary4,512F M-20-
LSAFBD [64]2020[1, 5]20,000FAsian20-
MEBeauty [24]2022[1, 10]2,550F MMulti300✓
LiveBeauty (proposed)2024[1, 5]10,000F MAsian20✓
", + "bbox": [ + 99, + 114, + 898, + 262 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The differences among these algorithms can lead to unique data distributions that vary considerably from those found in unprocessed or well-edited face images obtained from the Internet. Meanwhile, the effectiveness of face retouching is largely dependent on the computational resources available, which in turn affects these data characteristics. Since the training data impacts the generalization ability of data-driven deep learning methods, the lack of a FAP dataset specifically collected from live videos with authentic face retouching process may lead to limited performance in practical video scenarios.", + "bbox": [ + 88, + 287, + 480, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Meanwhile, there is no widely recognized FAP model specifically designed for live videos. Early methods relying on hand-crafted beauty descriptors [22, 30] tend to exhibit limited performance due to the inadequate feature representation capability. The success of the deep-learning based methods, e.g. Convolution Neural Networks (CNNs) [2, 32, 58] and Transformers [29, 39], have driven progress in the FAP task. Specifically, [39] first introduced a Vision Transformer into the FAP task and proposed a geometric prior guided network leveraging facial landmarks. Although their approach demonstrates promising performance, the straightforward feature concatenation and simplistic prior struggle to adaptively model individual facial features, thereby limiting its applicability and effectiveness in real-world applications.", + "bbox": [ + 91, + 455, + 482, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the lack of FAP datasets considering the impact of face retouching, we first construct a large-scale facial attractiveness dataset LiveBeauty, which includes 10,000 face images collected from a popular video streaming platform, with the corresponding facial attractiveness annotations obtained from 20 annotators.", + "bbox": [ + 89, + 685, + 480, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, based on the established dataset, we propose a no-reference FAP model termed Facial Prior Enhanced Multi-modal model (FPEM), to predict the facial attractiveness by extracting and integrating holistic facial prior knowledge and multi-modal aesthetic semantic features. Specifically, we first devise a Personalized Attractiveness Prior Module (PAPM) to integrate extensive facial prior knowledge, in which multi-scale visual features", + "bbox": [ + 89, + 779, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "extracted by the Swin Transformer [35] and facial prior features extracted by the pretrained FaceNet [43] are integrated via a Cross-Attention block. To obtain facial beauty attributes, we also introduce a Multi-modal Attractiveness Encoder Module (MAEM) based on CLIP [42] to extract multi-modal aesthetic semantic features. Furthermore, these extracted features are adaptively integrated through a fusion module. Extensive experimental results confirm that our proposed method achieves the best performance compared to other state-of-the-art (SOTA) methods on both our LiveBeauty dataset and other FAP datasets, as shown in Fig. 1.", + "bbox": [ + 511, + 287, + 903, + 467 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our contributions are as follow:", + "bbox": [ + 529, + 469, + 841, + 483 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We construct the LiveBeauty dataset, the first large-scale, open-source facial attractiveness prediction dataset sourced from a real-world live video platform featuring faces enhanced for aesthetics, making it suitable for applications wherein facial aesthetics are vital and therefore need to be well modeled and monitored.", + "- We propose a multi-modal method specifically designed for FAP, predicting attractiveness scores for face images. By integrating visual and textual modalities, this method can provide a more reliable prediction of facial appeal.", + "- We enhance the representation capability of our model by leveraging facial prior knowledge, and further employing selective fusion module to integrate and refine the extracted visual features." + ], + "bbox": [ + 513, + 486, + 903, + 696 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 715, + 661, + 729 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Aesthetic Assessment of Generic Images", + "text_level": 1, + "bbox": [ + 511, + 739, + 856, + 757 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image Aesthetic Assessment (IAA) task aims to assess image quality from the aesthetic perspective. The corresponding IAA methods have evolved from employing handcrafted features, the modeling of photographic rules [8], global image layout [36] or typical objects [25] in perceived images [10] to data-driven approaches based on deep learning techniques that deliver superior performance [18-21]. Notably, these IAA methods primarily concentrate on generic images rather than specifically targeting facial regions.", + "bbox": [ + 511, + 763, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11459", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Quality Assessment of Face Images", + "text_level": 1, + "bbox": [ + 89, + 90, + 395, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Facial clarity related assessment include both BFIQA and GFIQA. BFIQA was originated from biometric research focusing on evaluating the suitability of facial images for recognition, initially relying on hand-crafted features [11, 50], and recent advancing to deep learning approaches with regression and embedding-based solutions [4, 38]. Nonetheless, Chen et al. [7] noted these methods may not generalize well on other facial quality and aesthetic tasks due to the emphasis on facial biometric information. In contrast, GFIQA is defined to evaluate the conventional perceptual quality of face images. GFIQA methods have advanced through the use of transformer-based [7, 34] and multi-modal approaches [48]. Clearly, GFIQA focuses on modeling the degradation of facial regions, such as issues related to blur and noise, and is therefore orthogonal to facial attractiveness.", + "bbox": [ + 89, + 114, + 483, + 356 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Attractiveness Assessment of Face Images", + "text_level": 1, + "bbox": [ + 89, + 375, + 447, + 391 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the context of this paper, facial aesthetics primarily pertains to attractiveness. To measure the facial attractiveness of face images, early handcrafted Facial Attractiveness Prediction (FAP) methods predominantly rely on manually designed features, e.g. geometric features [12, 22, 30], traditional image descriptors [12, 22, 27], holistic descriptors [12], etc. However, these methods exhibit limited performance due to difficulties in presenting comprehensive characteristics of facial attractiveness.", + "bbox": [ + 89, + 398, + 483, + 535 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With the advancement in deep learning, FAP solutions based on Convolution Neural Networks (CNNs) show better performance. Inspired by psychological studies, specific architectures of CNNs are designed [5, 57]. The applications of Label Distribution Learning (LDL) [13, 31, 33] achieve better performance. Some methods [16, 46, 55] combine geometric prior knowledge containing landmarks, parsing maps, etc., while the others [28, 60] integrate attractiveness-related attributes such as gender, race, etc. In addition, the ensemble solutions of different loss functions like regression loss and classification loss [59] and different models including ResNeXt50 and Inception-V3 [2] are further explored. However, these CNN-based methods primarily focus on local features, limiting their capacity for long-term modeling. To address this problem, transformer-based solutions [3, 14] are proposed, while Liu et al. [29] and Gan et al. [15] combine transformers with CNNs to leverage both global and local features related to facial attractiveness. Nevertheless, these approaches have limited generalization and adaptation ability in more challenging scenarios, due to the biased predictions from training data skewed towards certain demographics, and using straightforward single-modal modeling techniques which rely on simplistic concatenation of visual features.", + "bbox": [ + 89, + 537, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5b96182c2ccf8cfdbbe76555526f15d55a289d08dc13595c1d73437b8f57134d.jpg", + "image_caption": [ + "Figure 2. Examples of the face images in our LiveBeauty dataset (from left to right are examples rated 1 to 5). The first row is male and second row is female." + ], + "image_footnote": [], + "bbox": [ + 535, + 92, + 885, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Dataset", + "text_level": 1, + "bbox": [ + 511, + 279, + 602, + 294 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Images Collection", + "text_level": 1, + "bbox": [ + 511, + 304, + 687, + 319 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To accurately reflect the distribution of facial attractiveness of in-captured videos, we first collect large-scale raw face images from one of the most popular live streaming platforms. Specifically, we retrieve replays for live channels of all streamers conducted on the platform in March 2024 with relatively high Page Views (PVs). Given that each channel lasts approximately 2 to 10 hours, we sample one image per hour for the first three hours of each channel, resulting in three images sampled from each channel. Then, a SOTA face detection model [65] is employed to generate the bounding box (bbox) of the face region.", + "bbox": [ + 511, + 325, + 906, + 491 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Data cleaning is further conducted to eliminate samples containing animated characters, virtual avatars, and samples from the same individual, to ensure the quality and diversity of the dataset. Finally, 10,000 images are collected, ensuring 10,000 different faces from over 9,500 live channels. Fig. 2 shows some examples of the face images in our LiveBeauty dataset.", + "bbox": [ + 511, + 491, + 905, + 597 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Subjective Experiments", + "text_level": 1, + "bbox": [ + 511, + 606, + 732, + 622 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We recruit 20 subjects in our experiments, including 14 females and 6 males. This ratio is determined by the actual customer demographics of the live platform. Face stimulus are displayed on an iPhone 14 Pro Max with 6.7-inch Organic Light-Emitting Diode (OLED) screen, supporting a resolution up to $2,796 \\times 1,290$ . Subjective experiments are conducted in a well-controlled lab environment with fixed settings [44], i.e., lighting, viewing distance, viewing angle, etc.", + "bbox": [ + 511, + 628, + 905, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We split the experiment into 200 sessions with 50 images per session. Each face image is displayed for 10 seconds and followed by a one-second gray screen mask. Before the experiment starting, we offer a brief explanation and 50 extra training samples for subjects to familiarize them with the task. During the experiment, subjects are asked to evaluate the facial attractiveness of the stimuli from 1 to 5, defined as: \"unappealing\", \"less favorable\", \"ordinary\", \"online influencer\" and \"celebrity\", with detailed descriptions", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11460", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/bf14c43aac92b4d22ef2b3368d64627042e8ae27de4b91c394d5a94b8aa41be3.jpg", + "image_caption": [ + "Figure 3. Illustration of the LiveBeauty MOS distribution. Skewness=-0.470 Kurtosis=1.498" + ], + "image_footnote": [], + "bbox": [ + 112, + 88, + 462, + 241 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of makeup and skin condition. After each session, subjects are asked to take at least 5 minutes of break. All subjects participate in all sessions, and each subject is allowed to participate in no more than 5 sessions each day. Ultimately, all 10,000 images are evaluated by 20 subjects, resulting in a total of 200,000 $(10,000 \\times 20)$ scores collected.", + "bbox": [ + 89, + 287, + 483, + 380 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Data Analysis", + "text_level": 1, + "bbox": [ + 89, + 388, + 235, + 404 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "According to [44], we employ subject post-screening with outlier ratio and Spearman Rank Correlation Coefficient (SROCC). Annotations that differ by more than two levels from the mode of all annotations are considered potential outliers and will be meticulously removed by the experts. A subject whose rated scores with SROCC $< 0.75$ or outlier ratio $>2\\%$ will be marked as unreliable and will be rejected. Ultimately, 20 subjects are retained. We compute the Mean Opinion Score (MOS) as the attractiveness label of each face image:", + "bbox": [ + 89, + 411, + 483, + 563 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {M O S} _ {j} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} r _ {i j}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 575, + 482, + 617 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{MOS}_j$ represents the MOS for the $j$ -th face image, $N$ is the number of the valid annotation scores rated by subjects, and $r_{ij}$ is the rated score of the $i$ -th subject on the $j$ -th face image.", + "bbox": [ + 89, + 627, + 482, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "MOS distribution of our LiveBeauty is shown in Fig. 3. It can be observed that the distributions of MOSs exhibit a Gaussian-like shape. Skewness and kurtosis values of three MOS distributions are also shown in Fig. 3, which indicate that the distributions are characterized by thin tails and the MOS is concentrated around 3. This aligns with real-world facial attractiveness distributions, i.e., most individuals possess average facial attractiveness.", + "bbox": [ + 89, + 688, + 482, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The annotation distribution of four attractive levels of each face sample is shown in Fig. 4. Although attractiveness is highly subjective and influenced by personal preference, it can be observed that the annotation scores for each face are concentrated around the three levels closest to the MOS (e.g., 3.85 comes from five 3, thirteen 4, and two 5).", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8cc5d4d81ef50b1584debd21ec78c61797f00ce540a77f0478176e5772d57e7a.jpg", + "image_caption": [ + "Figure 4. Annotation score distribution on four attractive level." + ], + "image_footnote": [], + "bbox": [ + 517, + 88, + 903, + 239 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Proposed Method", + "text_level": 1, + "bbox": [ + 511, + 273, + 687, + 290 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As illustrated in Fig. 5, our Facial Prior Enhanced Multi-modal (FPEM) method utilizes a two-stage training strategy, and consists of three primary modules.", + "bbox": [ + 511, + 299, + 903, + 345 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The first stage, namely Preliminary Training Phase, includes the Personalized Attractiveness Prior Module (PAPM in Sec. 4.1.1) and the Multi-modal Attractiveness Encoder Module (MAEM in Sec. 4.1.2). PAPM receives the image $I$ , from which a Swin Transformer $E_{MS}$ extracts multi-scale feature $f_{m}$ , and a frozen FaceNet $E_{FP}$ extracts face-aware feature $f_{s}$ . Subsequently, a Cross-Attention block incorporates $f_{m}$ and $f_{s}$ to form the personalized attractiveness feature $f_{p}$ , which are then projected to a preliminary attractiveness score $\\widehat{Q}_1$ . The MAEM contains an image encoder $E_{I}$ and a text encoder $E_{T}$ as in CLIP [42], receives the image $I$ and progressive texts, leverages visual embeddings $f_{i}$ and textual embeddings $f_{t}$ , and obtains another preliminary score $\\widehat{Q}_2$ through a Similarity Regression (SR) operation.", + "bbox": [ + 511, + 345, + 906, + 574 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the second stage, namely Selective Fusion Phase, a Selective Fusion Module (SFM in Sec. 4.2) dynamically refines the extracted visual embeddings $f_{i}$ to $f_{i}^{\\prime}$ based on the personalized attractiveness feature $f_{p}$ . Then the SR strategy is used to predict the final score $\\widehat{Q}$ .", + "bbox": [ + 511, + 575, + 905, + 654 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Preliminary Training Phase", + "text_level": 1, + "bbox": [ + 511, + 665, + 764, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this stage, PAPM and MAEM are trained independently without considering the SFM, and MAEM employs the SR strategy for regression during training.", + "bbox": [ + 511, + 688, + 905, + 736 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1.1. Personalized Attractiveness Prior Module", + "text_level": 1, + "bbox": [ + 511, + 744, + 849, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "**Stem Swin.** Considering that the visual perception progress is perceived from low level to high level, the hierarchical features from consecutive transformer stages are fully leveraged. Specifically, the features $f_{r_i} \\in \\mathbb{R}^{H_i \\times W_i \\times C_i}$ extracted from $i$ -th stage are reshaped into same size, and are then concatenated as $f_{\\Sigma} \\in \\mathbb{R}^{H_4 \\times W_4 \\times C_{\\Sigma}}$ . Meanwhile, $f_{\\Sigma}$ is refined by a Channel Attention block from channel perspective (which is shown in Fig. 6), and then flattened as $f_m \\in \\mathbb{R}^{N_4 \\times C_{\\Sigma}}$ .", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11461", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/324f40abd091c23171097a2525a108107b16200a9f71dbaf42cbb6eb88c4a023.jpg", + "image_caption": [ + "Figure 5. Architecture of the Facial Prior Enhanced Multi-modal model (FPEM), including the Personalized Attractiveness Prior Module (PAPM), the Multi-modal Attractiveness Encoder Module (MAEM) and the Selective Fusion Module (SFM)." + ], + "image_footnote": [], + "bbox": [ + 112, + 98, + 890, + 367 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Cross-Attention. Facial prior knowledge, i.e., face-aware features $f_{s}$ extracted by FaceNet, are utilized in this block to guide the learning of facial attractiveness. Specifically, the refined features $f_{m}$ extracted by the stem Swin are fused with the face-aware features $f_{s}$ through cross-attention. As shown in Fig. 6, $f_{s}$ is encoded as the query $Q$ via a Linear layer and a Layer Normalization (LN) layer, while $f_{m}$ is encoded as the key $K$ and value $V$ via a LN layer. Further, a Multi-Head Cross-Attention (MHCA) mechanism is applied, which is computed as:", + "bbox": [ + 88, + 439, + 483, + 590 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} \\operatorname {h e a d} _ {i} = \\psi \\left(\\frac {Q \\cdot K ^ {T}}{\\sqrt {d}}\\right) \\cdot V \\cdot W _ {i}, i \\in \\{1, \\dots , h \\} \\\\ \\operatorname {M H C A} (Q, K, V) = \\operatorname {C o n c a t} (\\operatorname {h e a d} _ {i}) W _ {o}, \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 598, + 482, + 641 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\psi (\\cdot)$ denotes the softmax function, $W_{o}$ denotes the projection parameter matrix and $d$ denotes the query dimension. Let FFN indicates the Feed-Forward Network, AP denotes Average Pooling, the output feature $f_{p}\\in \\mathbb{R}^{1\\times C}$ is computed as:", + "bbox": [ + 89, + 648, + 483, + 724 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} f _ {c} = f _ {m} + \\mathrm {M H C A} (Q, K, V) \\\\ f _ {p} = \\mathrm {A P} \\left(f _ {c} + \\text {F F N} \\left(\\mathrm {L N} \\left(f _ {c}\\right)\\right)\\right). \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 168, + 732, + 482, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let MLP represent the Multi Layer Perceptron, then the attractiveness score $\\widehat{Q}_1$ can be predicted by:", + "bbox": [ + 89, + 780, + 483, + 811 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {Q} _ {1} = \\operatorname {M L P} \\left(f _ {p}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 816, + 482, + 837 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.2. Multi-modal Attractiveness Encoder Module", + "text_level": 1, + "bbox": [ + 89, + 851, + 449, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inspired by [66], MAEM is devised to utilize multi-modal aesthetic semantic features. We utilize five attractiveness", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/22e36a6f14da6bdfacc30bf404eaf2ee5197c36afb4662059e959d96b30ce523.jpg", + "image_caption": [ + "Figure 6. Structure of the Channel Attention and the Cross-Attention Block." + ], + "image_footnote": [], + "bbox": [ + 532, + 431, + 705, + 616 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4efb84d53be049b81efa9c9cf066a409987082ea995876dbae46e5ee7f355b99.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 714, + 430, + 883, + 618 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "levels $a \\in \\mathcal{A} = \\{1,2,3,4,5\\} = \\{\\text{\"bad\", \"poor\", \"fair\", \"good\", \"perfect\"}\\}$ . Then the text template can be formed as \"a photo of a person with $\\{a\\}$ attractiveness\", resulting in five candidate text descriptions $\\mathcal{T}$ .", + "bbox": [ + 511, + 666, + 906, + 727 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MAEM contains a frozen text encoder $E_{T}$ (GPT-2) and a trainable image encoder $E_{I}$ (ViT-B/16), where $E_{T}$ receives five candidate text descriptions $\\mathcal{T}$ to form the textual embeddings $f_{t}^{\\prime}\\in \\mathbb{R}^{5\\times C}$ , while $E_{I}$ receives the image $I$ to form the visual embeddings $f_{i}\\in \\mathbb{R}^{1\\times C}$ .", + "bbox": [ + 511, + 727, + 905, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Subsequently, the textual embeddings $f_{t}^{\\prime}$ and visual embeddings $f_{i}$ are fed into Similar Regression (SR). In particular, the cosine similarity between $f_{t}$ and $f_{i}$ is calculated to obtain the attractiveness level probability distribution $\\hat{p}$ :", + "bbox": [ + 511, + 803, + 905, + 864 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {p} = \\psi \\left(\\frac {f _ {i} \\cdot f _ {t}}{\\| f _ {i} \\| \\times \\| f _ {t} \\|}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 871, + 903, + 905 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11462", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\psi (\\cdot)$ denotes the softmax function. Then the attractiveness scores $\\widehat{Q}_2$ can be predicted by:", + "bbox": [ + 89, + 90, + 482, + 125 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {Q} _ {2} = \\sum_ {a = 1} ^ {5} \\hat {p} (a | I) \\cdot a, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 128, + 482, + 170 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $a$ denotes the attractiveness level from 1 to 5, and $\\hat{p}(a|\\cdot)$ is the probability of the level $a$ to be estimated.", + "bbox": [ + 89, + 176, + 482, + 208 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Selective Fusion Phase", + "text_level": 1, + "bbox": [ + 89, + 215, + 302, + 229 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this stage, we freeze the parameters of MAEM and PAPM, which are trained in the previous preliminary training phase. Then the Selective Fusion Model (SFM) is trained to dynamically integrate holistic facial prior knowledge and multi-modal aesthetic semantic features. To achieve this goal, we first fuse the visual embeddings $f_{i}$ and personalized attractiveness feature $f_{p}$ via an element-wise addition. Inspired by [26], a compact feature $f_{z} \\in \\mathbb{R}^{1 \\times d}$ is further generated to facilitate adaptive selection guidance for these facial features, which is produced by a Squeeze Layer implemented utilizing a Fully Connected layer (FC) and a ReLU activation function:", + "bbox": [ + 89, + 237, + 483, + 417 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nf _ {z} = \\operatorname {R e L U} \\left(F C \\left(f _ {i} + f _ {p}\\right)\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 426, + 482, + 444 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$f_{z}$ are then projected to two adaptive weights $\\mathbf{a} \\in \\mathbb{R}^{1 \\times C}$ and $\\mathbf{b} \\in \\mathbb{R}^{1 \\times C}$ for $f_{i}$ and $f_{p}$ via two FC layers. A Softmax function is employed on $\\mathbf{a}$ and $\\mathbf{b}$ on their channel-wise digits to ensure $\\mathbf{a} + \\mathbf{b} = \\mathbf{1}$ . Afterwards, the optimized visual embeddings $f_{i}'$ can be obtained by the channel-wise weighted addition with $\\mathbf{a}$ and $\\mathbf{b}$ :", + "bbox": [ + 89, + 450, + 483, + 540 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nf _ {i} ^ {\\prime} = f _ {i} \\cdot \\mathbf {a} + f _ {p} \\cdot \\mathbf {b}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 549, + 482, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Subsequently, the refined $f_{i}^{\\prime}$ and textual embeddings $f_{t}$ are input into Similar Regression to calculate the final prediction $\\widehat{Q}$ , which is consistent with the MAEM.", + "bbox": [ + 89, + 573, + 483, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Training objectives", + "text_level": 1, + "bbox": [ + 89, + 626, + 274, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Loss for PAPM. The L1 loss between $\\widehat{Q}_1$ and the ground truth label $Q$ is calculated as the attractiveness prediction loss function $\\mathcal{L}_1$ , which can be formulated as:", + "bbox": [ + 89, + 648, + 483, + 695 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {1} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left| \\widehat {Q} _ {1 _ {i}} - Q _ {i} \\right|, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 700, + 482, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $n$ denotes number of training samples.", + "bbox": [ + 89, + 747, + 390, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Loss for MAEM. To jointly optimize ranking and regression, we incorporate a merged ranking loss $\\mathcal{L}_R$ along with the L1 scoring loss $\\mathcal{L}_S$ . Specifically, $\\mathcal{L}_R$ is composed of a fidelity loss $\\mathcal{L}_{R_1}$ and a two-direction ranking loss $\\mathcal{L}_{R_2}$ . For the fidelity loss $\\mathcal{L}_{R_1}$ , given an image pair $(I_i, I_j)$ , the binary label according to their ground truth $(Q_i, Q_j)$ is defined as:", + "bbox": [ + 89, + 762, + 483, + 854 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nR \\left(I _ {i}, I _ {j}\\right) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} Q _ {i} \\geq Q _ {j} \\\\ 0 & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 863, + 482, + 904 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7757488bf5e6d34838806709c2ac8158ef1145c54fb8814a76cea56b27eddeae.jpg", + "image_caption": [ + "Figure 7. MOS distribution of three benchmark datasets." + ], + "image_footnote": [], + "bbox": [ + 535, + 90, + 875, + 236 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "According to the Thurstone's model [51], the probability of $I_{i}$ perceiving more attractive than $I_{j}$ is estimated as:", + "bbox": [ + 511, + 272, + 906, + 304 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {R} \\left(I _ {i}, I _ {j}\\right) = \\Psi \\left(\\frac {\\widehat {Q} _ {2 i} - \\widehat {Q} _ {2 j}}{\\sqrt {2}}\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 313, + 903, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\Psi (\\cdot)$ denotes the standard normal cumulative distribution function, and its variance is set to 1. Then the fidelity loss [54] can be calculated as:", + "bbox": [ + 511, + 359, + 905, + 404 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {R _ {1}} = 1 - \\sqrt {R \\left(I _ {i} , I _ {j}\\right) \\widehat {R} \\left(I _ {i} , I _ {j}\\right)} \\tag {12} \\\\ - \\sqrt {(1 - R (I _ {i} , I _ {j})) (1 - \\widehat {R} (I _ {i} , I _ {j}))}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 414, + 903, + 468 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the two-direction ranking loss $\\mathcal{L}_{R_2}$ , given an image $I_i$ , $\\widehat{S}_{i,j}$ denotes the cosine similarity score of $j_{th}$ attractiveness level and $\\widehat{S}_{i,o} = \\max \\left\\{\\widehat{S}_{i,j} \\mid j \\in \\{1,2,3,4,5\\}\\right\\}$ . Since the human perception of facial attractiveness is gradual and rarely exhibits multiple peaks [62], we further obtain the ranking information by ensuring that the attractiveness level probability distribution only exhibits one peak and decreases in two directions (leftward and rightward). Then the two-direction pairwise ranking loss can be formulated as:", + "bbox": [ + 511, + 477, + 906, + 636 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} \\mathcal {L} _ {\\text {l e f t}} = - \\sum_ {j = 2} ^ {o} \\log \\frac {\\exp \\left(\\widehat {S} _ {i , j}\\right)}{\\exp \\left(\\widehat {S} _ {i , j}\\right) + \\exp \\left(\\widehat {S} _ {i , j - 1}\\right)} \\\\ \\mathcal {L} _ {\\text {r i g h t}} = - \\sum_ {j = o} ^ {A - 1} \\log \\frac {\\exp \\left(\\widehat {S} _ {i , j}\\right)}{\\exp \\left(\\widehat {S} _ {i , j}\\right) + \\exp \\left(\\widehat {S} _ {i , j + 1}\\right)} \\\\ \\mathcal {L} _ {R _ {2}} = - \\left(\\mathcal {L} _ {\\text {l e f t}} + \\mathcal {L} _ {\\text {r i g h t}}\\right), \\end{array} \\right. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 657, + 905, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $A$ denotes the number of attractiveness levels, and the corresponding labels of $\\widehat{S}_{i,j-1}$ , $\\widehat{S}_{i,j}$ and $\\widehat{S}_{i,j+1}$ are 0, 1, 0, respectively. In the end, the total loss of MAEM can be defined as $\\mathcal{L}_2 = \\mathcal{L}_S + \\lambda_1\\mathcal{L}_{R_1} + \\lambda_2\\mathcal{L}_{R_2}$ , where $\\lambda_1$ and $\\lambda_2$ are empirically set as 1.", + "bbox": [ + 511, + 736, + 905, + 811 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Loss for SFM. The loss used in the second stage between the ground truth label $Q$ and the final attractiveness score $\\hat{Q}$ is the L1 Loss:", + "bbox": [ + 511, + 811, + 905, + 854 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {3} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left| \\widehat {Q} _ {i} - Q _ {i} \\right|. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 864, + 903, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11463", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5eaf98d016f45e8364d6403673bd9a44b23da84fd0fe09add6599518303d8ec1.jpg", + "table_caption": [ + "Table 2. Performance comparison on LiveBeauty, MEBeauty [24] and SCUT-FBP5500 [27]. Best in red and second-best in blue." + ], + "table_footnote": [], + "table_body": "
AspectsMethodsLiveBeautyMEBeauty [24]SCUT-FBP5500 [27]
SROCC↑PLCC↑KROCC↑SROCC↑PLCC↑KROCC↑SROCC↑PLCC↑KROCC↑
BaselineViT-B [23]0.8970.8470.7330.6200.6590.4430.8690.8800.686
ResNeXt-50 [56]0.9070.8790.7460.7000.7430.5160.8990.9110.727
Inception-V3 [49]0.9060.8800.7450.6760.7270.4920.9050.9180.736
IAAAVA-MLSP [21]0.8330.8150.6500.6020.6430.4310.7640.7940.566
TANet [18]0.8910.8580.7280.6510.6900.4670.8700.8820.686
Dele-Trans [19]0.8880.8410.7210.6110.6510.4400.8700.8600.682
EAT [20]0.8930.8670.7280.6400.6970.4620.8530.8720.665
FAPComboNet [58]0.9020.8710.7410.6510.6920.4720.8970.9070.725
2D-FAP [32]0.8960.8540.7330.6900.7190.5060.9030.9150.734
REX-INCEP [2]0.9080.8840.7500.7020.7390.5140.9070.9170.739
CNN-ER [2]0.9130.8880.7570.7060.7480.5190.9100.9220.745
MEBeauty [24]0.8260.7960.6430.6490.6770.4670.7990.8020.599
Ours0.9240.8960.7730.7830.7950.5930.9300.9320.774
Improvement+1.1%+0.8%+1.6%+7.7%+4.7%+7.4%+2.0%+1.0%+2.9%
", + "bbox": [ + 99, + 114, + 898, + 338 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 362, + 222, + 380 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Benchmark Datasets", + "text_level": 1, + "bbox": [ + 89, + 388, + 285, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The proposed LiveBeauty dataset and two open-source FAP datasets including SCUT-FBP5500 [27] and MEBeauty [24] are selected as benchmark datasets. The MOS distributions of them are shown in Fig. 7. We split SCUT-FBP5500 and MEBeauty with $60\\% - 40\\%$ and $80\\% - 20\\%$ train-test ratio separately, which is kept same with their own protocol, and LiveBeauty is split using the $90\\% - 10\\%$ protocol.", + "bbox": [ + 89, + 410, + 483, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 542, + 307, + 558 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Model initialization. In MAEM, ViT-B/16 and GPT-2 are adopted as the image encoder and the text encoder separately, which are initialized by [42]. In PAPM, Swin-T is adopted as the trainable image encoder, where only four stages are reserved and initialized with parameters pretrained on multiple face-related tasks [41].", + "bbox": [ + 89, + 564, + 482, + 655 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Hyper parameters. During the whole training process, AdamW is selected as the optimizer, and the learning rate scheduler is set with linear warm-up and cosine annealing scheme. The base learning rate varies with different training phase. Each training phase lasts 50 epochs, and the batch size is set as 32. Details can be found in the supplementary.", + "bbox": [ + 89, + 656, + 482, + 746 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Competitors & Metrics", + "text_level": 1, + "bbox": [ + 89, + 756, + 305, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To confirm the effectiveness of the proposed FPEM method, we select baseline models and several SOTA FAP and IAA methods, including ComboNet [58], 2D-FAP [32], REX-INCEP [2], CNN-ER [2], MEBeauty [24], AVAMLSP [21], TANet [18], Dele-Trans [19] and EAT [20]. All the models are retrained ourselves with the default parameters defined in their papers. Three criteria are adopted to evaluate the performance of these methods, i.e., Spear", + "bbox": [ + 89, + 779, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/28d2034fb63b1395875df6d64800e3e84f7cc87ef0b6ea83451044ba25a643b3.jpg", + "image_caption": [ + "(a)LiveBeauty." + ], + "image_footnote": [], + "bbox": [ + 524, + 367, + 638, + 455 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/66cb23b930b5d467e635259eac8fc61419aac3ff5d028a63c71dc4fa16887585.jpg", + "image_caption": [ + "(b) MEBeauty." + ], + "image_footnote": [], + "bbox": [ + 651, + 367, + 767, + 454 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cd149623c33da7a322f9c693cb927d6e0111a2a3bbef640be7776cc98cd3a04c.jpg", + "image_caption": [ + "(c) SCUT-FBP5500.", + "Figure 8. Scatter plots of the FPEM predicted scores vs. MOSs on three benchmark datasets." + ], + "image_footnote": [], + "bbox": [ + 779, + 367, + 893, + 454 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "man Rank Order Correlation Coefficient (SROCC), Pearson Linear Correlation Coefficient (PLCC) and Kendall Rank Order Correlation Coefficient (KROCC).", + "bbox": [ + 511, + 521, + 905, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Performance", + "text_level": 1, + "bbox": [ + 511, + 575, + 650, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The experimental results on three FAP datasets are shown in Tab. 2, from which we could draw the following conclusions: (i) Our proposed method achieves the first place and surpasses the second place by about 0.011, 0.077, 0.020 in terms of SROCC values on LiveBeauty, MEBeauty and SCUT-FBP5500 respectively, which demonstrates its superiority. (ii) The IAA methods are inferior to the FAP methods, which manifests that the generic aesthetic assessment methods overlook the facial features involved in the subjective nature of facial attractiveness, leading to limited performance on FAP tasks. (iii) The performance of all methods drops on MEBeauty because the number of training samples is restricted and the faces are ethnically diverse in MEBeauty, indicating a large diversity in facial attractiveness. (iv) The significant improvement on MEBeauty over other datasets can be largely attributed to the enhancement brought by face priors, which is more effective in scenarios with insufficient training data.", + "bbox": [ + 511, + 598, + 906, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 8 presents scatter plots illustrating the relationship between the predicted attractiveness scores (vertical axis)", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11464", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/817f1fab70160ebe7b1e1339d43c9dc47a8bff84aa6861481461e1af5f5145fd.jpg", + "table_caption": [ + "Table 3. Ablation study on LiveBeauty test set. $E_{FP}$ denotes Cross-Attention with $f_p$ , MS denotes Multi-Scale fusion operation, SR denotes Similarity Regression strategy, $\\mathcal{L}_S$ denotes the L1 scoring loss, $\\mathcal{L}_R$ denotes the total ranking loss $\\mathcal{L}_R = \\mathcal{L}_{R_1} + \\mathcal{L}_{R_2}$ ." + ], + "table_footnote": [], + "table_body": "
ModulesMethodsMetrics
SROCC↑PLCC↑KROCC↑
PAPMonlyEMS0.90450.87100.7454
w/oEFp0.90780.87140.7475
w/o MS0.90890.88080.7487
PAPM0.90960.88100.7497
MAEMzero-shot CLIP0.38380.35820.2692
w/o SR0.88250.82260.7123
w/oLS0.91610.84330.7607
w/oLR0.91860.89110.7653
MAEM0.91980.86310.7653
ALLw/o 2-stage0.91890.88890.7652
w/o SFM0.91970.89610.7641
all combined0.92430.89620.7732
", + "bbox": [ + 99, + 143, + 477, + 339 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "obtained by FPEM and the MOSs (horizontal axis) in the test sets of three datasets. The MOSs of MEBeauty are divided by two to align the MOS range with others. The color of the scatter points from blue to red indicates the density of points from low to high. In each sub-figure, a second-order polynomial nonlinear fitting curve is superimposed. It can be observed that the fitted curves are close to the diagonal line, demonstrating the effectiveness of FPEM.", + "bbox": [ + 89, + 359, + 483, + 482 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 494, + 243, + 510 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct elaborate ablation study to investigate the contributions of different modules in our proposed method, which is shown in Tab. 3. Conclusions can be drawn that: (i) In the ablation study of PAPM, we can observe that leveraging facial prior and multi-scale features are both helpful. (ii) The result of ablation study of MAEM indicates that the zero-shot CLIP has limited generalization ability in FAP tasks, which confirms its shortcoming in aesthetic evaluation [45]. Specifically, for MAEM w/o SR, we leverage concatenation of the visual features and the textual features with highest similarity, and utilize MLP for regression. We observe a significant performance drop without SR, which manifests that specifically devised regression strategy is important for CLIP to adapt to regression tasks. Meanwhile, the joint optimization of the ranking and regression also leads to the improvement of SROCC. (iii) In the ablation study of ALL modules, we can see that employing all modules yields the best performance, which confirms the effectiveness of using SFM to adaptively integrate multi-modal aesthetic features with holistic facial prior knowledge.", + "bbox": [ + 89, + 517, + 483, + 820 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.6. Cross-Dataset Study", + "text_level": 1, + "bbox": [ + 89, + 832, + 284, + 848 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Considering the diversity of facial attractiveness perception, the cross-dataset validation is carried out to validate the generalization ability of these methods. Specifically,", + "bbox": [ + 89, + 854, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/122ed23cc017e8d75824ba76220e7065a2fe64cf599d4428c017f21929be0854.jpg", + "table_caption": [ + "Table 4. Cross-Dataset study trained on LiveBeauty and tested on the entire MEBeauty dataset and the entire SCUT-FBP5500 dataset. Best in red and second-best in blue." + ], + "table_footnote": [], + "table_body": "
MethodsMEBeauty [24]SCUT-FBP5500 [27]
SROCC↑PLCC↑KROCC↑SROCC↑PLCC↑KROCC↑
ViT-B [23]0.5730.5830.4020.5530.5480.394
ResNeXt-50 [56]0.5350.5460.3720.4780.4670.335
Inception-V3 [49]0.5750.5810.4040.4840.4800.339
AVA-MLSP [21]0.5870.5970.4140.4000.4120.275
TANet [18]0.5510.5290.3840.4010.4110.277
Dele-Trans[19]0.5760.5720.4020.4800.4760.335
EAT[20]0.6030.6240.4270.4840.4660.338
ComboNet [58]0.5400.5480.3750.3840.3750.264
2D-FAP [32]0.5560.4890.3900.4440.4450.306
REX-INCEP [2]0.5880.5920.4130.4730.4710.332
CNN-ER [2]0.5890.5960.4140.5020.4990.352
MEBeauty [24]0.5200.5160.3610.3660.3390.251
FPEM (Ours)0.6300.6290.4470.5990.5770.434
Improvement+2.7%+0.5%+2.0%+4.6%+2.9%+4.0%
", + "bbox": [ + 519, + 130, + 898, + 338 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4e071a24c33ca62e0349d2f7f47b84ec55b558130e25fffb8fd9031ffdfd5f5d.jpg", + "table_caption": [ + "Table 5. Cross-Dataset Ablation Study on each module. All models are trained on LiveBeauty and tested on the whole MEBeauty." + ], + "table_footnote": [], + "table_body": "
PAPMMAEMSFMSROCC↑PLCC↑KROCC↑
✓0.61340.61410.4349
✓0.62290.62680.4244
✓✓✓0.62990.62910.4472
", + "bbox": [ + 544, + 391, + 870, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "all the mentioned models are trained on the proposed LiveBeauty dataset and subsequently tested on the entire SCUTFBP5500 and the entire MEBeauty.", + "bbox": [ + 511, + 469, + 905, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Tab. 4, the proposed method outperforms all other methods on both two datasets, which demonstrates its strong generalization ability. Moreover, ablation results of the generalizability on MEBeauty are presented in Tab. 5. It can be observed that MAEM surpasses PAPM by leveraging multi-modal information. The generalizability is further improved by fusing their features through SFM, which indicates that integrating facial prior containing rich contextual and individual information is effective not only for improving performance but also for enhancing generalizability.", + "bbox": [ + 511, + 513, + 906, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 679, + 633, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we tackle the problem of facial attractiveness prediction in live videos that undergo face retouching. To address the existing gap in datasets, we introduce LiveBeauty, the first large-scale FAP dataset featuring beautiful face images for aesthetic purposes. This dataset is high-quality and authentic, collected from real-world environments, and is currently the largest open-access facial attractiveness dataset. We further propose FPEM, a multimodal FAP framework that selectively fuses holistic facial priors and aesthetic semantic features. Extensive experiments demonstrate the advantages of FPEM over state-of-the-art approaches. Meanwhile, LiveBeauty and FPEM can be effectively leveraged for FAP in other video applications.", + "bbox": [ + 511, + 704, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11465", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Lacey Best-Rowden and Anil K Jain. Learning face image quality from human assessments. IEEE Transactions on Information Forensics and Security, 13(12):3064-3077, 2018. 1", + "[2] Fares Bougourzi, Fadi Dornaika, and Abdelmalik Taleb-Ahmed. Deep learning based face beauty prediction via dynamic robust losses and ensemble regression. Knowledge-Based Systems, 242:108246, 2022. 1, 2, 3, 7, 8", + "[3] Djamel Eddine Boukhari, Ali Chemsa, and Riadh Ajgou. Facial beauty prediction based on vision transformer. International Journal of Electrical and Electronic Engineering & Telecommunications, pages 2319-2518, 2023. 3", + "[4] Fadi Boutros, Meiling Fang, Marcel Klemt, Biying Fu, and Naser Damer. Cr-fiaq: face image quality assessment by learning sample relative classifiability. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5836-5845, 2023. 3", + "[5] Kerang Cao, Kwang-nam Choi, Hoekyung Jung, and Lini Duan. Deep learning for facial beauty prediction. Information, 11(8):391, 2020. 3", + "[6] Nicolas Chahine, Stefania Calarasanu, Davide Garcia-Civiero, Theo Cayla, Sira Ferradans, and Jean Ponce. An image quality assessment dataset for portraits. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9968-9978, 2023. 1", + "[7] Wei-Ting Chen, Gurunandan Krishnan, Qiang Gao, SyYen Kuo, Sizhou Ma, and Jian Wang. Dsl-fqa: Assessing facial image quality via dual-set degradation learning and landmark-guided transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2931-2941, 2024. 3", + "[8] Ritendra Datta, Dhiraj Joshi, Jia Li, and James Z Wang. Studying aesthetics in photographic images using a computational approach. In Proceedings of the European Conference on Computer Vision, pages 288-301, 2006. 2", + "[9] Harm De Vries and Jason Yosinski. Can deep learning help you find the perfect match? arXiv preprint arXiv:1505.00359, 2015. 1, 2", + "[10] Yubin Deng, Chen Change Loy, and Xiaou Tang. Image aesthetic assessment: An experimental survey. IEEE Signal Processing Magazine, 34(4):80-106, 2017. 2", + "[11] Abhishek Dutta, Raymond Veldhuis, and Luuk Spreeuwers. A bayesian model for predicting face recognition performance using image quality. In Proceeding of the International Joint Conference on Biometrics, pages 1-8, 2014. 3", + "[12] Yael Eisenthal, Gideon Dror, and Eytan Ruppin. Facial attractiveness: Beauty and the machine. Neural Computation, 18(1):119-142, 2006. 1, 3", + "[13] Yang-Yu Fan, Shu Liu, Bo Li, Zhe Guo, Ashok Samal, Jun Wan, and Stan Z Li. Label distribution-based facial attractiveness computation by deep residual learning. IEEE Transactions on Multimedia, 20(8):2196-2208, 2017. 3", + "[14] Junying Gan, Xiaoshan Xie, Guohui He, and Heng Luo. Transbls: transformer combined with broad learning system" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for facial beauty prediction. Applied Intelligence, 53(21): 26110-26125, 2023. 3", + "[15] Junying Gan, Xiaoshan Xie, Yikui Zhai, Guohui He, Chaoyun Mai, and Heng Luo. Facial beauty prediction fusing transfer learning and broad learning system. Soft Computing, 27(18):13391-13404, 2023. 3", + "[16] Lian Gao, Weixin Li, Zehua Huang, Di Huang, and Yunhong Wang. Automatic facial attractiveness prediction by deep multi-task learning. In Proceedings of the IEEE International Conference on Pattern Recognition, pages 3592-3597, 2018. 3", + "[17] Douglas Gray, Kai Yu, Wei Xu, and Yihong Gong. Predicting facial beauty without landmarks. In Proceedings of the European Conference on Computer Vision, pages 434-447, 2010. 1, 2", + "[18] Shuai He, Yongchang Zhang, Rui Xie, Dongxiang Jiang, and Anlong Ming. Rethinking image aesthetics assessment: Models, datasets and benchmarks. In Proceedings of the International Joint Conference on Artificial Intelligence, pages 942–948, 2022. 2, 7, 8", + "[19] Shuai He, Anlong Ming, Yaqi Li, Jinyuan Sun, ShunTian Zheng, and Huadong Ma. Thinking image color aesthetics assessment: Models, datasets and benchmarks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21838-21847, 2023. 7, 8", + "[20] Shuai He, Anlong Ming, Shuntian Zheng, Haobin Zhong, and Huadong Ma. Eat: An enhancer for aesthetics-oriented transformers. In Proceedings of the ACM International Conference on Multimedia, pages 1023-1032, 2023. 7, 8", + "[21] Vlad Hosu, Bastian Goldlucke, and Dietmar Saupe. Effective aesthetics prediction with multi-level spatially pooled features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9375-9383, 2019. 2, 7, 8", + "[22] Tharun J. Iyer, Rahul K, Ruban Nersisson, Zhemin Zhuang, Alex Noel Joseph Raj, and Imthiaz Refayee. Machine learning-based facial beauty prediction and analysis of frontal facial images using facial landmarks and traditional image descriptors. Computational Intelligence and Neuroscience, 2021(1):4423407, 2021. 2, 3", + "[23] Alexander Kolesnikov, Alexey Dosovitskiy, Dirk Weissenborn, Georg Heigold, Jakob Uszkoreit, Lucas Beyer, Matthias Minderer, Mostafa Dehghani, Neil Houlsby, Sylvain Gelly, Thomas Unterthiner, and Xiaohua Zhai. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2021. 7, 8", + "[24] Irina Lebedeva, Yi Guo, and Fangli Ying. Mebeauty: a multi-ethnic facial beauty dataset in-the-wild. *Neural Computing and Applications*, pages 1–15, 2022. 1, 2, 7, 8", + "[25] Congcong Li, Andrew Gallagher, Alexander C. Loui, and Tsuhan Chen. Aesthetic quality assessment of consumer photos with faces. In Proceedings of the IEEE International Conference on Image Processing, pages 3221-3224, 2010. 2", + "[26] Xiang Li, Wenhai Wang, Xiaolin Hu, and Jian Yang. Selective kernel networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 510-519, 2019. 6" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11466", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Lingyu Liang, Luojun Lin, Lianwen Jin, Duorui Xie, and Mengru Li. Scut-fbp5500: A diverse benchmark dataset for multi-paradigm facial beauty prediction. In Proceedings of the IEEE International Conference on Pattern Recognition, pages 1598-1603, 2018. 1, 2, 3, 7, 8", + "[28] Luojun Lin, Lingyu Liang, Lianwen Jin, and Weijie Chen. Attribute-aware convolutional neural networks for facial beauty prediction. In Proceedings of the Information Joint Conference on Artificial Intelligence, pages 847-853, 2019. 1, 3", + "[29] Qipeng Liu, Luojun Lin, Zhifeng Shen, and Yuanlong Yu. Fbpformer: Dynamic convolutional transformer for global-local-contextual facial beauty prediction. In Proceedings of the International Conference on Artificial Neural Networks, pages 223-235, 2023. 2, 3", + "[30] Shu Liu, Yang-Yu Fan, Zhe Guo, Ashok Samal, and Afan Ali. A landmark-based data-driven approach on 2.5 d facial attractiveness computation. Neurocomputing, 238:168-178, 2017. 2, 3", + "[31] Shu Liu, Bo Li, Yang-Yu Fan, Zhe Quo, and Ashok Samal. Facial attractiveness computation by label distribution learning with deep cnn and geometric features. In Proceedings of the IEEE International Conference on Multimedia and Expo, pages 1344-1349, 2017. 3", + "[32] Shu Liu, Enquan Huang, Yan Xu, Kexuan Wang, Xiaoyan Kui, Tao Lei, and Hongying Meng. Lightweight facial attractiveness prediction using dual label distribution. arXiv preprint arXiv:2212.01742, 2022. 1, 2, 7, 8", + "[33] Shu Liu, Enquan Huang, Yan Xu, Kexuan Wang, Xiaoyan Kui, Tao Lei, and Hongying Meng. Lightweight facial attractiveness prediction using dual label distribution. arXiv preprint arXiv:2212.01742, 2022. 3", + "[34] Tie Liu, Shengxi Li, Mai Xu, Li Yang, and Xiaofei Wang. Assessing face image quality: A large-scale database and a transformer method. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3", + "[35] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2", + "[36] Naila Murray, Luca Marchesotti, and Florent Perronnin. Ava: A large-scale database for aesthetic visual analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2408-2415, 2012. 2", + "[37] Fu-Zhao Ou, Xingyu Chen, Ruixin Zhang, Yuge Huang, Shaoxin Li, Jilin Li, Yong Li, Liujuan Cao, and Yuan-Gen Wang. Sdd-fqa: Unsupervised face image quality assessment with similarity distribution distance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7670–7679, 2021. 1", + "[38] Fu-Zhao Ou, Chongyi Li, Shiqi Wang, and Sam Kwong. Clib-fiqa: Face image quality assessment with confidence calibration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1694-1704, 2024. 3", + "[39] Tianhao Peng, Mu Li, Fangmei Chen, Yong Xu, and David Zhang. Geometric prior guided hybrid deep neural network" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for facial beauty analysis. CAAI Transactions on Intelligence Technology, 9(2):467-480, 2024. 2", + "[40] P Jonathon Phillips, J Ross Beveridge, David S Bolme, Bruce A Draper, Geof H Givens, Yui Man Lui, Su Cheng, Mohammad Nayeem Teli, and Hao Zhang. On the existence of face quality measures. In Proceedings of the IEEE International Conference on Biometrics: Theory, Applications and Systems, pages 1-8, 2013. 1", + "[41] Lixiong Qin, Mei Wang, Chao Deng, Ke Wang, Xi Chen, Jiani Hu, and Weihong Deng. Swinface: a multi-task transformer for face recognition, expression recognition, age estimation and attribute estimation. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 7", + "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proceedings of the International Conference on Machine Learning, pages 8748-8763, 2021. 2, 4, 7", + "[43] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 815-823, 2015. 2", + "[44] B Series. Methodology for the subjective assessment of the quality of television pictures. Recommendation ITU-R BT, 500(13), 2012. 3, 4", + "[45] Xiangfei Sheng, Leida Li, Pengfei Chen, Jinjian Wu, Weisheng Dong, Yuzhe Yang, Liwu Xu, Yaqian Li, and Guangming Shi. Aesclip: Multi-attribute contrastive learning for image aesthetics assessment. In Proceedings of the ACM International Conference on Multimedia, page 1117–1126, 2023. 8", + "[46] Shengjie Shi, Fei Gao, Xuantong Meng, Xingxin Xu, and Jingjie Zhu. Improving facial attractiveness prediction via co-attention learning. In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing, pages 4045–4049, 2019. 3", + "[47] Shaolin Su, Hanhe Lin, Vlad Hosu, Oliver Wiedemann, Jinqiu Sun, Yu Zhu, Hantao Liu, Yanning Zhang, and Dietmar Saupe. Going the extra mile in face image quality assessment: A novel database and model. IEEE Transactions on Multimedia, 2023. 1", + "[48] Wei Sun, Weixia Zhang, Yanwei Jiang, Haoning Wu, Zicheng Zhang, Jun Jia, Yingjie Zhou, Zhongpeng Ji, Xiongkuo Min, Weisi Lin, and Zhai Guangtao. Dual-branch network for portrait image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2024. 3", + "[49] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2826, 2016. 1, 7, 8", + "[50] Philipp Terhorst, Jan Niklas Kolf, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. Ser-fiq: Unsupervised estimation of face image quality based on stochastic embedding" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11467", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "robustness. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5651-5660, 2020. 3", + "[51] Louis L Thurstone. A law of comparative judgment. In *Scaling*, pages 81–92. Routledge, 2017. 6", + "[52] Yu Tian, Zhangkai Ni, Baoliang Chen, Shiqi Wang, Hanli Wang, and Sam Kwong. Generalized visual quality assessment of gan-generated face images. arXiv preprint arXiv:2201.11975, 2022. 1", + "[53] Song Tong, Xuefeng Liang, Takatsune Kumada, and Sunao Iwaki. Putative ratios of facial attractiveness in a deep neural network. Vision Research, 178:86-99, 2021. 1, 2", + "[54] Ming-Feng Tsai, Tie-Yan Liu, Tao Qin, Hsin-Hsi Chen, and Wei-Ying Ma. Frank: a ranking method with fidelity loss. In Proceedings of the ACM SIGIR Conference on Research and Development in Information Retrieval, pages 383-390, 2007. 6", + "[55] Qinjie Xiao, You Wu, Dinghong Wang, Yong-Liang Yang, and Xiaogang Jin. Beauty3dfacenet: deep geometry and texture fusion for 3d facial attractiveness prediction. Computers & Graphics, 98:11-18, 2021. 3", + "[56] Saining Xie, Ross Girshick, Piotr Dolkar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1492-1500, 2017. 1, 7, 8", + "[57] Jie Xu, Lianwen Jin, Lingyu Liang, Ziyong Feng, Duorui Xie, and Huiyun Mao. Facial attractiveness prediction using psychologically inspired convolutional neural network (pCNN). In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing, pages 1657-1661, 2017. 3", + "[58] Lu Xu and Jinhai Xiang. Comboloss for facial attractiveness analysis with squeeze-and-excitation networks. arXiv preprint arXiv:2010.10721, 2020. 2, 7, 8", + "[59] Lu Xu and Jinhai Xiang. Comboloss for facial attractiveness analysis with squeeze-and-excitation networks. arXiv preprint arXiv:2010.10721, 2020. 3", + "[60] Lu Xu, Heng Fan, and Jinhai Xiang. Hierarchical multi-task network for race, gender and facial attractiveness recognition. In Proceedings of the IEEE International Conference on Image Processing, pages 3861-3865, 2019. 3", + "[61] Haibin Yan. Cost-sensitive ordinal regression for fully automatic facial beauty assessment. Neurocomputing, 129:334-342, 2014. 1, 2", + "[62] Qinkai Yu, Jianyang Xie, Anh Nguyen, He Zhao, Jiong Zhang, Huazhu Fu, Yitian Zhao, Yalin Zheng, and Yanda Meng. Clip-dr: Textual knowledge-guided diabetic retinopathy grading with ranking-aware prompting. arXiv preprint arXiv:2407.04068, 2024. 6", + "[63] Yikui Zhai, Yu Huang, Ying Xu, Junying Zeng, Fei Yu, and Junying Gan. Benchmark of a large scale database for facial beauty prediction. In Proceedings of the International Conference on Intelligent Information Processing, pages 1-5, 2016. 1, 2", + "[64] Yikui Zhai, Yu Huang, Ying Xu, Junying Gan, He Cao, Wenbo Deng, Ruggero Donida Labati, Vincenzo Piuri, and" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Fabio Scotti. Asian female facial beauty prediction using deep neural networks via transfer learning and multi-channel feature fusion. IEEE Access, 8:56892-56907, 2020. 1, 2", + "[65] Shifeng Zhang, Xiangyu Zhu, Zhen Lei, Hailin Shi, Xiaobo Wang, and Stan Z Li. Faceboxes: A cpu real-time face detector with high accuracy. In Proceedings of the IEEE International Joint Conference on Biometrics, pages 1-9, 2017. 3", + "[66] Weixia Zhang, Guangtao Zhai, Ying Wei, Xiaokang Yang, and Kede Ma. Blind image quality assessment via vision-language correspondence: A multitask learning perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14071-14081, 2023. 5" + ], + "bbox": [ + 516, + 92, + 903, + 287 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11468", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/FPEM_ Face Prior Enhanced Facial Attractiveness Prediction for Live Videos with Face Retouching/f47f8d6d-8a6c-491b-8f19-08123eb17e5c_model.json b/2025/FPEM_ Face Prior Enhanced Facial Attractiveness Prediction for Live Videos with Face Retouching/f47f8d6d-8a6c-491b-8f19-08123eb17e5c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f8abd427c2326da562e0347a9f210baaebd08ea2 --- /dev/null +++ b/2025/FPEM_ Face Prior Enhanced Facial Attractiveness Prediction for Live Videos with Face Retouching/f47f8d6d-8a6c-491b-8f19-08123eb17e5c_model.json @@ -0,0 +1,2598 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.189, + 0.131, + 0.812, + 0.177 + ], + "angle": 0, + "content": "FPEM: Face Prior Enhanced Facial Attractiveness Prediction for Live Videos with Face Retouching" + }, + { + "type": "text", + "bbox": [ + 0.233, + 0.203, + 0.761, + 0.24 + ], + "angle": 0, + "content": "Hui Li\\(^{1\\dagger}\\), Xiaoyu Ren\\(^{1,2\\dagger}\\), Hongjiu Yu\\(^{1}\\), Ying Chen\\(^{1*}\\), Kai Li\\(^{1}\\), L Wang\\(^{1}\\), Xiongkuo Min\\(^{2}\\), Huiyu Duan\\(^{2}\\), Guangtao Zhai\\(^{2}\\), Xu Liu\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.354, + 0.241, + 0.645, + 0.257 + ], + "angle": 0, + "content": "1 Alibaba Group 2Shanghai Jiao Tong University" + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.26, + 0.892, + 0.275 + ], + "angle": 0, + "content": "{lh401704,naomi.rxy,yuhongjiu.yhj,YingChen,kaishi.lk,zhuangshu.wlb,steven.liu}@taobao.com" + }, + { + "type": "text", + "bbox": [ + 0.285, + 0.278, + 0.708, + 0.293 + ], + "angle": 0, + "content": "{minxiongkuo,huiyuduan,zhaiguangtao}@sjtu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.358, + 0.486, + 0.675 + ], + "angle": 0, + "content": "Facial attractiveness prediction (FAP) has long been an important computer vision task, which could be widely applied in live videos with facial retouching. However, previous FAP datasets are either small or closed-source. Moreover, the corresponding FAP models exhibit limited generalization and adaptation ability. To overcome these limitations, we introduce the first large-scale FAP dataset LiveBeauty specifically designed for live video scenarios wherein face images may be real-time processed for aesthetics purposes. 10,000 face images are collected directly from a live streaming platform, with 200,000 corresponding attractiveness annotations obtained from a well-devised subjective experiment, making LiveBeauty the largest open-access FAP dataset. Based on the built dataset, a novel FAP method named Facial Prior Enhanced Multi-modal model (FPEM) is proposed to measure the attractiveness of facial images. Extensive experiments conducted on both LiveBeauty and other open-source FAP datasets demonstrate that our proposed method achieves state-of-the-art performance. The dataset will be released at https://github.com/EstellaLH/FPEM." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.705, + 0.222, + 0.721 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.73, + 0.485, + 0.866 + ], + "angle": 0, + "content": "With the rapid development of live streaming, short-form video, and advanced face retouching within such applications, the perceptual quality of faces becomes an important factor for the overall visual experience of users, which is particularly essential for video platforms, as it can significantly affect the average viewer watch time and even GMV (Gross Merchandise Volume). A robust metric system for evaluating facial clarity and aesthetics is essential to monitor both the overall service quality of the platform and" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.334, + 0.707, + 0.465 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.334, + 0.902, + 0.465 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.48, + 0.907, + 0.537 + ], + "angle": 0, + "content": "Figure 1. SROCC vs PLCC comparison on LiveBeauty and SCUT-FBP5500 [27]. FPEM, denoted by red triangular, outperforms other methods (CNN-ER [2], 2D-FAP [32], REX-INCEP [2], Inception-V3 [49], ResNeXt-50 [56]) on both datasets." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.567, + 0.907, + 0.779 + ], + "angle": 0, + "content": "the visual appeal changes or improvements of individual streaming channels. The clarity-related assessment tasks of facial quality could be categorized into Biometric Face Image Quality Assessment (BFIQA) and Generic Face Image Quality Assessment (GFIQA). BFIQA focuses on the identity information of faces for down-stream tasks such as face recognition [1, 37, 40], while GFIQA emphasizes the visual perceptual quality of faces [6, 47, 52]. The aesthetics-related assessment of facial quality is often referred to as Facial Attractiveness Prediction (FAP) [12, 27, 28]. The goal of FAP is to quantify the attractiveness of facial images, which is vital for assessing the effectiveness of facial retouching settings and offers valuable insights for algorithm development." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "As shown in Tab. 1, numerous FAP datasets have been proposed in recent years, many of which encompass a rich diversity in age, gender, face expression and pose [9, 17, 24, 27, 53, 61, 63, 64]. However, most of these datasets are either small or unavailable due to privacy and ethical concerns. In video streaming, especially live video, it is common practice to use facial retouching algorithms to improve the facial aesthetics of content presented to audiences." + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.236, + 0.888 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.225, + 0.9 + ], + "angle": 0, + "content": "Equal contributions." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.236, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11458" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.289, + 0.09, + 0.709, + 0.104 + ], + "angle": 0, + "content": "Table 1. Overview of common facial attractiveness prediction datasets." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.115, + 0.9, + 0.263 + ], + "angle": 0, + "content": "
DatasetYearRating Scale# ImageGenderEthnicity# AnnotatorPublic
Gray et al. [17]2010[-3, 3]2,056F MCaucasian30✓
MBW+FBW [61]2014[1, 5]10,400F MMulti30-
De Vries et al. [9]2015Binary9364F MAsian20-
LSFBD [63]2016[1, 5]20,000F M---
SCUT-FBP5500 [27]2018[1, 5]5,500F MMulti60✓
Tong et al. [53]2020Binary4,512F M-20-
LSAFBD [64]2020[1, 5]20,000FAsian20-
MEBeauty [24]2022[1, 10]2,550F MMulti300✓
LiveBeauty (proposed)2024[1, 5]10,000F MAsian20✓
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.288, + 0.482, + 0.452 + ], + "angle": 0, + "content": "The differences among these algorithms can lead to unique data distributions that vary considerably from those found in unprocessed or well-edited face images obtained from the Internet. Meanwhile, the effectiveness of face retouching is largely dependent on the computational resources available, which in turn affects these data characteristics. Since the training data impacts the generalization ability of data-driven deep learning methods, the lack of a FAP dataset specifically collected from live videos with authentic face retouching process may lead to limited performance in practical video scenarios." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.457, + 0.483, + 0.683 + ], + "angle": 0, + "content": "Meanwhile, there is no widely recognized FAP model specifically designed for live videos. Early methods relying on hand-crafted beauty descriptors [22, 30] tend to exhibit limited performance due to the inadequate feature representation capability. The success of the deep-learning based methods, e.g. Convolution Neural Networks (CNNs) [2, 32, 58] and Transformers [29, 39], have driven progress in the FAP task. Specifically, [39] first introduced a Vision Transformer into the FAP task and proposed a geometric prior guided network leveraging facial landmarks. Although their approach demonstrates promising performance, the straightforward feature concatenation and simplistic prior struggle to adaptively model individual facial features, thereby limiting its applicability and effectiveness in real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.686, + 0.482, + 0.777 + ], + "angle": 0, + "content": "To address the lack of FAP datasets considering the impact of face retouching, we first construct a large-scale facial attractiveness dataset LiveBeauty, which includes 10,000 face images collected from a popular video streaming platform, with the corresponding facial attractiveness annotations obtained from 20 annotators." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Furthermore, based on the established dataset, we propose a no-reference FAP model termed Facial Prior Enhanced Multi-modal model (FPEM), to predict the facial attractiveness by extracting and integrating holistic facial prior knowledge and multi-modal aesthetic semantic features. Specifically, we first devise a Personalized Attractiveness Prior Module (PAPM) to integrate extensive facial prior knowledge, in which multi-scale visual features" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.905, + 0.468 + ], + "angle": 0, + "content": "extracted by the Swin Transformer [35] and facial prior features extracted by the pretrained FaceNet [43] are integrated via a Cross-Attention block. To obtain facial beauty attributes, we also introduce a Multi-modal Attractiveness Encoder Module (MAEM) based on CLIP [42] to extract multi-modal aesthetic semantic features. Furthermore, these extracted features are adaptively integrated through a fusion module. Extensive experimental results confirm that our proposed method achieves the best performance compared to other state-of-the-art (SOTA) methods on both our LiveBeauty dataset and other FAP datasets, as shown in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.47, + 0.842, + 0.484 + ], + "angle": 0, + "content": "To summarize, our contributions are as follow:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.487, + 0.905, + 0.576 + ], + "angle": 0, + "content": "- We construct the LiveBeauty dataset, the first large-scale, open-source facial attractiveness prediction dataset sourced from a real-world live video platform featuring faces enhanced for aesthetics, making it suitable for applications wherein facial aesthetics are vital and therefore need to be well modeled and monitored." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.578, + 0.905, + 0.637 + ], + "angle": 0, + "content": "- We propose a multi-modal method specifically designed for FAP, predicting attractiveness scores for face images. By integrating visual and textual modalities, this method can provide a more reliable prediction of facial appeal." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.638, + 0.904, + 0.697 + ], + "angle": 0, + "content": "- We enhance the representation capability of our model by leveraging facial prior knowledge, and further employing selective fusion module to integrate and refine the extracted visual features." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.487, + 0.905, + 0.697 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.716, + 0.663, + 0.731 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.741, + 0.857, + 0.758 + ], + "angle": 0, + "content": "2.1. Aesthetic Assessment of Generic Images" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Image Aesthetic Assessment (IAA) task aims to assess image quality from the aesthetic perspective. The corresponding IAA methods have evolved from employing handcrafted features, the modeling of photographic rules [8], global image layout [36] or typical objects [25] in perceived images [10] to data-driven approaches based on deep learning techniques that deliver superior performance [18-21]. Notably, these IAA methods primarily concentrate on generic images rather than specifically targeting facial regions." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11459" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.396, + 0.108 + ], + "angle": 0, + "content": "2.2. Quality Assessment of Face Images" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.357 + ], + "angle": 0, + "content": "Facial clarity related assessment include both BFIQA and GFIQA. BFIQA was originated from biometric research focusing on evaluating the suitability of facial images for recognition, initially relying on hand-crafted features [11, 50], and recent advancing to deep learning approaches with regression and embedding-based solutions [4, 38]. Nonetheless, Chen et al. [7] noted these methods may not generalize well on other facial quality and aesthetic tasks due to the emphasis on facial biometric information. In contrast, GFIQA is defined to evaluate the conventional perceptual quality of face images. GFIQA methods have advanced through the use of transformer-based [7, 34] and multi-modal approaches [48]. Clearly, GFIQA focuses on modeling the degradation of facial regions, such as issues related to blur and noise, and is therefore orthogonal to facial attractiveness." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.375, + 0.449, + 0.392 + ], + "angle": 0, + "content": "2.3. Attractiveness Assessment of Face Images" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.4, + 0.484, + 0.536 + ], + "angle": 0, + "content": "In the context of this paper, facial aesthetics primarily pertains to attractiveness. To measure the facial attractiveness of face images, early handcrafted Facial Attractiveness Prediction (FAP) methods predominantly rely on manually designed features, e.g. geometric features [12, 22, 30], traditional image descriptors [12, 22, 27], holistic descriptors [12], etc. However, these methods exhibit limited performance due to difficulties in presenting comprehensive characteristics of facial attractiveness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.484, + 0.901 + ], + "angle": 0, + "content": "With the advancement in deep learning, FAP solutions based on Convolution Neural Networks (CNNs) show better performance. Inspired by psychological studies, specific architectures of CNNs are designed [5, 57]. The applications of Label Distribution Learning (LDL) [13, 31, 33] achieve better performance. Some methods [16, 46, 55] combine geometric prior knowledge containing landmarks, parsing maps, etc., while the others [28, 60] integrate attractiveness-related attributes such as gender, race, etc. In addition, the ensemble solutions of different loss functions like regression loss and classification loss [59] and different models including ResNeXt50 and Inception-V3 [2] are further explored. However, these CNN-based methods primarily focus on local features, limiting their capacity for long-term modeling. To address this problem, transformer-based solutions [3, 14] are proposed, while Liu et al. [29] and Gan et al. [15] combine transformers with CNNs to leverage both global and local features related to facial attractiveness. Nevertheless, these approaches have limited generalization and adaptation ability in more challenging scenarios, due to the biased predictions from training data skewed towards certain demographics, and using straightforward single-modal modeling techniques which rely on simplistic concatenation of visual features." + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.093, + 0.887, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.226, + 0.907, + 0.268 + ], + "angle": 0, + "content": "Figure 2. Examples of the face images in our LiveBeauty dataset (from left to right are examples rated 1 to 5). The first row is male and second row is female." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.28, + 0.603, + 0.295 + ], + "angle": 0, + "content": "3. Dataset" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.305, + 0.688, + 0.32 + ], + "angle": 0, + "content": "3.1. Images Collection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.326, + 0.907, + 0.492 + ], + "angle": 0, + "content": "To accurately reflect the distribution of facial attractiveness of in-captured videos, we first collect large-scale raw face images from one of the most popular live streaming platforms. Specifically, we retrieve replays for live channels of all streamers conducted on the platform in March 2024 with relatively high Page Views (PVs). Given that each channel lasts approximately 2 to 10 hours, we sample one image per hour for the first three hours of each channel, resulting in three images sampled from each channel. Then, a SOTA face detection model [65] is employed to generate the bounding box (bbox) of the face region." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.906, + 0.598 + ], + "angle": 0, + "content": "Data cleaning is further conducted to eliminate samples containing animated characters, virtual avatars, and samples from the same individual, to ensure the quality and diversity of the dataset. Finally, 10,000 images are collected, ensuring 10,000 different faces from over 9,500 live channels. Fig. 2 shows some examples of the face images in our LiveBeauty dataset." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.607, + 0.733, + 0.623 + ], + "angle": 0, + "content": "3.2. Subjective Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.906, + 0.763 + ], + "angle": 0, + "content": "We recruit 20 subjects in our experiments, including 14 females and 6 males. This ratio is determined by the actual customer demographics of the live platform. Face stimulus are displayed on an iPhone 14 Pro Max with 6.7-inch Organic Light-Emitting Diode (OLED) screen, supporting a resolution up to \\(2,796 \\times 1,290\\). Subjective experiments are conducted in a well-controlled lab environment with fixed settings [44], i.e., lighting, viewing distance, viewing angle, etc." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We split the experiment into 200 sessions with 50 images per session. Each face image is displayed for 10 seconds and followed by a one-second gray screen mask. Before the experiment starting, we offer a brief explanation and 50 extra training samples for subjects to familiarize them with the task. During the experiment, subjects are asked to evaluate the facial attractiveness of the stimuli from 1 to 5, defined as: \"unappealing\", \"less favorable\", \"ordinary\", \"online influencer\" and \"celebrity\", with detailed descriptions" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11460" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.089, + 0.464, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.251, + 0.462, + 0.28 + ], + "angle": 0, + "content": "Figure 3. Illustration of the LiveBeauty MOS distribution. Skewness=-0.470 Kurtosis=1.498" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.289, + 0.484, + 0.381 + ], + "angle": 0, + "content": "of makeup and skin condition. After each session, subjects are asked to take at least 5 minutes of break. All subjects participate in all sessions, and each subject is allowed to participate in no more than 5 sessions each day. Ultimately, all 10,000 images are evaluated by 20 subjects, resulting in a total of 200,000 \\((10,000 \\times 20)\\) scores collected." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.39, + 0.236, + 0.405 + ], + "angle": 0, + "content": "3.3. Data Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.412, + 0.484, + 0.564 + ], + "angle": 0, + "content": "According to [44], we employ subject post-screening with outlier ratio and Spearman Rank Correlation Coefficient (SROCC). Annotations that differ by more than two levels from the mode of all annotations are considered potential outliers and will be meticulously removed by the experts. A subject whose rated scores with SROCC \\(< 0.75\\) or outlier ratio \\(>2\\%\\) will be marked as unreliable and will be rejected. Ultimately, 20 subjects are retained. We compute the Mean Opinion Score (MOS) as the attractiveness label of each face image:" + }, + { + "type": "equation", + "bbox": [ + 0.216, + 0.576, + 0.483, + 0.618 + ], + "angle": 0, + "content": "\\[\n\\mathrm {M O S} _ {j} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} r _ {i j}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.483, + 0.688 + ], + "angle": 0, + "content": "where \\(\\mathrm{MOS}_j\\) represents the MOS for the \\(j\\)-th face image, \\(N\\) is the number of the valid annotation scores rated by subjects, and \\(r_{ij}\\) is the rated score of the \\(i\\)-th subject on the \\(j\\)-th face image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.81 + ], + "angle": 0, + "content": "MOS distribution of our LiveBeauty is shown in Fig. 3. It can be observed that the distributions of MOSs exhibit a Gaussian-like shape. Skewness and kurtosis values of three MOS distributions are also shown in Fig. 3, which indicate that the distributions are characterized by thin tails and the MOS is concentrated around 3. This aligns with real-world facial attractiveness distributions, i.e., most individuals possess average facial attractiveness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The annotation distribution of four attractive levels of each face sample is shown in Fig. 4. Although attractiveness is highly subjective and influenced by personal preference, it can be observed that the annotation scores for each face are concentrated around the three levels closest to the MOS (e.g., 3.85 comes from five 3, thirteen 4, and two 5)." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.522, + 0.251, + 0.896, + 0.266 + ], + "angle": 0, + "content": "Figure 4. Annotation score distribution on four attractive level." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.274, + 0.688, + 0.291 + ], + "angle": 0, + "content": "4. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.3, + 0.905, + 0.346 + ], + "angle": 0, + "content": "As illustrated in Fig. 5, our Facial Prior Enhanced Multi-modal (FPEM) method utilizes a two-stage training strategy, and consists of three primary modules." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.347, + 0.907, + 0.575 + ], + "angle": 0, + "content": "The first stage, namely Preliminary Training Phase, includes the Personalized Attractiveness Prior Module (PAPM in Sec. 4.1.1) and the Multi-modal Attractiveness Encoder Module (MAEM in Sec. 4.1.2). PAPM receives the image \\(I\\), from which a Swin Transformer \\(E_{MS}\\) extracts multi-scale feature \\(f_{m}\\), and a frozen FaceNet \\(E_{FP}\\) extracts face-aware feature \\(f_{s}\\). Subsequently, a Cross-Attention block incorporates \\(f_{m}\\) and \\(f_{s}\\) to form the personalized attractiveness feature \\(f_{p}\\), which are then projected to a preliminary attractiveness score \\(\\widehat{Q}_1\\). The MAEM contains an image encoder \\(E_{I}\\) and a text encoder \\(E_{T}\\) as in CLIP [42], receives the image \\(I\\) and progressive texts, leverages visual embeddings \\(f_{i}\\) and textual embeddings \\(f_{t}\\), and obtains another preliminary score \\(\\widehat{Q}_2\\) through a Similarity Regression (SR) operation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.577, + 0.906, + 0.655 + ], + "angle": 0, + "content": "In the second stage, namely Selective Fusion Phase, a Selective Fusion Module (SFM in Sec. 4.2) dynamically refines the extracted visual embeddings \\( f_{i} \\) to \\( f_{i}^{\\prime} \\) based on the personalized attractiveness feature \\( f_{p} \\). Then the SR strategy is used to predict the final score \\( \\widehat{Q} \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.666, + 0.766, + 0.683 + ], + "angle": 0, + "content": "4.1. Preliminary Training Phase" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.906, + 0.737 + ], + "angle": 0, + "content": "In this stage, PAPM and MAEM are trained independently without considering the SFM, and MAEM employs the SR strategy for regression during training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.746, + 0.851, + 0.76 + ], + "angle": 0, + "content": "4.1.1. Personalized Attractiveness Prior Module" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "**Stem Swin.** Considering that the visual perception progress is perceived from low level to high level, the hierarchical features from consecutive transformer stages are fully leveraged. Specifically, the features \\( f_{r_i} \\in \\mathbb{R}^{H_i \\times W_i \\times C_i} \\) extracted from \\( i \\)-th stage are reshaped into same size, and are then concatenated as \\( f_{\\Sigma} \\in \\mathbb{R}^{H_4 \\times W_4 \\times C_{\\Sigma}} \\). Meanwhile, \\( f_{\\Sigma} \\) is refined by a Channel Attention block from channel perspective (which is shown in Fig. 6), and then flattened as \\( f_m \\in \\mathbb{R}^{N_4 \\times C_{\\Sigma}} \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11461" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.099, + 0.891, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.385, + 0.908, + 0.416 + ], + "angle": 0, + "content": "Figure 5. Architecture of the Facial Prior Enhanced Multi-modal model (FPEM), including the Personalized Attractiveness Prior Module (PAPM), the Multi-modal Attractiveness Encoder Module (MAEM) and the Selective Fusion Module (SFM)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.44, + 0.484, + 0.592 + ], + "angle": 0, + "content": "Cross-Attention. Facial prior knowledge, i.e., face-aware features \\( f_{s} \\) extracted by FaceNet, are utilized in this block to guide the learning of facial attractiveness. Specifically, the refined features \\( f_{m} \\) extracted by the stem Swin are fused with the face-aware features \\( f_{s} \\) through cross-attention. As shown in Fig. 6, \\( f_{s} \\) is encoded as the query \\( Q \\) via a Linear layer and a Layer Normalization (LN) layer, while \\( f_{m} \\) is encoded as the key \\( K \\) and value \\( V \\) via a LN layer. Further, a Multi-Head Cross-Attention (MHCA) mechanism is applied, which is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.599, + 0.483, + 0.642 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} \\operatorname {h e a d} _ {i} = \\psi \\left(\\frac {Q \\cdot K ^ {T}}{\\sqrt {d}}\\right) \\cdot V \\cdot W _ {i}, i \\in \\{1, \\dots , h \\} \\\\ \\operatorname {M H C A} (Q, K, V) = \\operatorname {C o n c a t} (\\operatorname {h e a d} _ {i}) W _ {o}, \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.649, + 0.484, + 0.725 + ], + "angle": 0, + "content": "where \\(\\psi (\\cdot)\\) denotes the softmax function, \\(W_{o}\\) denotes the projection parameter matrix and \\(d\\) denotes the query dimension. Let FFN indicates the Feed-Forward Network, AP denotes Average Pooling, the output feature \\(f_{p}\\in \\mathbb{R}^{1\\times C}\\) is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.169, + 0.733, + 0.483, + 0.775 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} f _ {c} = f _ {m} + \\mathrm {M H C A} (Q, K, V) \\\\ f _ {p} = \\mathrm {A P} \\left(f _ {c} + \\text {F F N} \\left(\\mathrm {L N} \\left(f _ {c}\\right)\\right)\\right). \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.813 + ], + "angle": 0, + "content": "Let MLP represent the Multi Layer Perceptron, then the attractiveness score \\(\\widehat{Q}_1\\) can be predicted by:" + }, + { + "type": "equation", + "bbox": [ + 0.229, + 0.818, + 0.483, + 0.838 + ], + "angle": 0, + "content": "\\[\n\\widehat {Q} _ {1} = \\operatorname {M L P} \\left(f _ {p}\\right). \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.852, + 0.45, + 0.866 + ], + "angle": 0, + "content": "4.1.2. Multi-modal Attractiveness Encoder Module" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Inspired by [66], MAEM is devised to utilize multi-modal aesthetic semantic features. We utilize five attractiveness" + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.432, + 0.707, + 0.617 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.431, + 0.885, + 0.619 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.627, + 0.907, + 0.656 + ], + "angle": 0, + "content": "Figure 6. Structure of the Channel Attention and the Cross-Attention Block." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.667, + 0.907, + 0.728 + ], + "angle": 0, + "content": "levels \\( a \\in \\mathcal{A} = \\{1,2,3,4,5\\} = \\{\\text{\"bad\", \"poor\", \"fair\", \"good\", \"perfect\"}\\} \\). Then the text template can be formed as \"a photo of a person with \\(\\{a\\}\\) attractiveness\", resulting in five candidate text descriptions \\(\\mathcal{T}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.728, + 0.906, + 0.804 + ], + "angle": 0, + "content": "MAEM contains a frozen text encoder \\( E_{T} \\) (GPT-2) and a trainable image encoder \\( E_{I} \\) (ViT-B/16), where \\( E_{T} \\) receives five candidate text descriptions \\( \\mathcal{T} \\) to form the textual embeddings \\( f_{t}^{\\prime}\\in \\mathbb{R}^{5\\times C} \\), while \\( E_{I} \\) receives the image \\( I \\) to form the visual embeddings \\( f_{i}\\in \\mathbb{R}^{1\\times C} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.804, + 0.906, + 0.866 + ], + "angle": 0, + "content": "Subsequently, the textual embeddings \\( f_{t}^{\\prime} \\) and visual embeddings \\( f_{i} \\) are fed into Similar Regression (SR). In particular, the cosine similarity between \\( f_{t} \\) and \\( f_{i} \\) is calculated to obtain the attractiveness level probability distribution \\( \\hat{p} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.635, + 0.872, + 0.905, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\hat {p} = \\psi \\left(\\frac {f _ {i} \\cdot f _ {t}}{\\| f _ {i} \\| \\times \\| f _ {t} \\|}\\right), \\tag {5}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "11462" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.125 + ], + "angle": 0, + "content": "where \\(\\psi (\\cdot)\\) denotes the softmax function. Then the attractiveness scores \\(\\widehat{Q}_2\\) can be predicted by:" + }, + { + "type": "equation", + "bbox": [ + 0.216, + 0.13, + 0.483, + 0.171 + ], + "angle": 0, + "content": "\\[\n\\widehat {Q} _ {2} = \\sum_ {a = 1} ^ {5} \\hat {p} (a | I) \\cdot a, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.178, + 0.483, + 0.209 + ], + "angle": 0, + "content": "where \\(a\\) denotes the attractiveness level from 1 to 5, and \\(\\hat{p}(a|\\cdot)\\) is the probability of the level \\(a\\) to be estimated." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.216, + 0.303, + 0.231 + ], + "angle": 0, + "content": "4.2. Selective Fusion Phase" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.238, + 0.485, + 0.419 + ], + "angle": 0, + "content": "In this stage, we freeze the parameters of MAEM and PAPM, which are trained in the previous preliminary training phase. Then the Selective Fusion Model (SFM) is trained to dynamically integrate holistic facial prior knowledge and multi-modal aesthetic semantic features. To achieve this goal, we first fuse the visual embeddings \\( f_{i} \\) and personalized attractiveness feature \\( f_{p} \\) via an element-wise addition. Inspired by [26], a compact feature \\( f_{z} \\in \\mathbb{R}^{1 \\times d} \\) is further generated to facilitate adaptive selection guidance for these facial features, which is produced by a Squeeze Layer implemented utilizing a Fully Connected layer (FC) and a ReLU activation function:" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.427, + 0.483, + 0.445 + ], + "angle": 0, + "content": "\\[\nf _ {z} = \\operatorname {R e L U} \\left(F C \\left(f _ {i} + f _ {p}\\right)\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.451, + 0.484, + 0.541 + ], + "angle": 0, + "content": "\\(f_{z}\\) are then projected to two adaptive weights \\(\\mathbf{a} \\in \\mathbb{R}^{1 \\times C}\\) and \\(\\mathbf{b} \\in \\mathbb{R}^{1 \\times C}\\) for \\(f_{i}\\) and \\(f_{p}\\) via two FC layers. A Softmax function is employed on \\(\\mathbf{a}\\) and \\(\\mathbf{b}\\) on their channel-wise digits to ensure \\(\\mathbf{a} + \\mathbf{b} = \\mathbf{1}\\). Afterwards, the optimized visual embeddings \\(f_{i}'\\) can be obtained by the channel-wise weighted addition with \\(\\mathbf{a}\\) and \\(\\mathbf{b}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.219, + 0.55, + 0.483, + 0.567 + ], + "angle": 0, + "content": "\\[\nf _ {i} ^ {\\prime} = f _ {i} \\cdot \\mathbf {a} + f _ {p} \\cdot \\mathbf {b}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.574, + 0.484, + 0.62 + ], + "angle": 0, + "content": "Subsequently, the refined \\( f_{i}^{\\prime} \\) and textual embeddings \\( f_{t} \\) are input into Similar Regression to calculate the final prediction \\( \\widehat{Q} \\), which is consistent with the MAEM." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.627, + 0.275, + 0.643 + ], + "angle": 0, + "content": "4.3. Training objectives" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.65, + 0.484, + 0.696 + ], + "angle": 0, + "content": "Loss for PAPM. The L1 loss between \\(\\widehat{Q}_1\\) and the ground truth label \\(Q\\) is calculated as the attractiveness prediction loss function \\(\\mathcal{L}_1\\), which can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.702, + 0.483, + 0.741 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {1} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left| \\widehat {Q} _ {1 _ {i}} - Q _ {i} \\right|, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.748, + 0.391, + 0.763 + ], + "angle": 0, + "content": "where \\(n\\) denotes number of training samples." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.763, + 0.484, + 0.855 + ], + "angle": 0, + "content": "Loss for MAEM. To jointly optimize ranking and regression, we incorporate a merged ranking loss \\(\\mathcal{L}_R\\) along with the L1 scoring loss \\(\\mathcal{L}_S\\). Specifically, \\(\\mathcal{L}_R\\) is composed of a fidelity loss \\(\\mathcal{L}_{R_1}\\) and a two-direction ranking loss \\(\\mathcal{L}_{R_2}\\). For the fidelity loss \\(\\mathcal{L}_{R_1}\\), given an image pair \\((I_i, I_j)\\), the binary label according to their ground truth \\((Q_i, Q_j)\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.864, + 0.483, + 0.905 + ], + "angle": 0, + "content": "\\[\nR \\left(I _ {i}, I _ {j}\\right) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} Q _ {i} \\geq Q _ {j} \\\\ 0 & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {10}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.092, + 0.877, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.245, + 0.878, + 0.26 + ], + "angle": 0, + "content": "Figure 7. MOS distribution of three benchmark datasets." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.273, + 0.907, + 0.305 + ], + "angle": 0, + "content": "According to the Thurstone's model [51], the probability of \\( I_{i} \\) perceiving more attractive than \\( I_{j} \\) is estimated as:" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.314, + 0.905, + 0.351 + ], + "angle": 0, + "content": "\\[\n\\widehat {R} \\left(I _ {i}, I _ {j}\\right) = \\Psi \\left(\\frac {\\widehat {Q} _ {2 i} - \\widehat {Q} _ {2 j}}{\\sqrt {2}}\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.36, + 0.906, + 0.405 + ], + "angle": 0, + "content": "where \\(\\Psi (\\cdot)\\) denotes the standard normal cumulative distribution function, and its variance is set to 1. Then the fidelity loss [54] can be calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.415, + 0.905, + 0.469 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {R _ {1}} = 1 - \\sqrt {R \\left(I _ {i} , I _ {j}\\right) \\widehat {R} \\left(I _ {i} , I _ {j}\\right)} \\tag {12} \\\\ - \\sqrt {(1 - R (I _ {i} , I _ {j})) (1 - \\widehat {R} (I _ {i} , I _ {j}))}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.478, + 0.907, + 0.637 + ], + "angle": 0, + "content": "For the two-direction ranking loss \\(\\mathcal{L}_{R_2}\\), given an image \\(I_i\\), \\(\\widehat{S}_{i,j}\\) denotes the cosine similarity score of \\(j_{th}\\) attractiveness level and \\(\\widehat{S}_{i,o} = \\max \\left\\{\\widehat{S}_{i,j} \\mid j \\in \\{1,2,3,4,5\\}\\right\\}\\). Since the human perception of facial attractiveness is gradual and rarely exhibits multiple peaks [62], we further obtain the ranking information by ensuring that the attractiveness level probability distribution only exhibits one peak and decreases in two directions (leftward and rightward). Then the two-direction pairwise ranking loss can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.659, + 0.906, + 0.728 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} \\mathcal {L} _ {\\text {l e f t}} = - \\sum_ {j = 2} ^ {o} \\log \\frac {\\exp \\left(\\widehat {S} _ {i , j}\\right)}{\\exp \\left(\\widehat {S} _ {i , j}\\right) + \\exp \\left(\\widehat {S} _ {i , j - 1}\\right)} \\\\ \\mathcal {L} _ {\\text {r i g h t}} = - \\sum_ {j = o} ^ {A - 1} \\log \\frac {\\exp \\left(\\widehat {S} _ {i , j}\\right)}{\\exp \\left(\\widehat {S} _ {i , j}\\right) + \\exp \\left(\\widehat {S} _ {i , j + 1}\\right)} \\\\ \\mathcal {L} _ {R _ {2}} = - \\left(\\mathcal {L} _ {\\text {l e f t}} + \\mathcal {L} _ {\\text {r i g h t}}\\right), \\end{array} \\right. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.737, + 0.906, + 0.812 + ], + "angle": 0, + "content": "where \\(A\\) denotes the number of attractiveness levels, and the corresponding labels of \\(\\widehat{S}_{i,j-1}\\), \\(\\widehat{S}_{i,j}\\) and \\(\\widehat{S}_{i,j+1}\\) are 0, 1, 0, respectively. In the end, the total loss of MAEM can be defined as \\(\\mathcal{L}_2 = \\mathcal{L}_S + \\lambda_1\\mathcal{L}_{R_1} + \\lambda_2\\mathcal{L}_{R_2}\\), where \\(\\lambda_1\\) and \\(\\lambda_2\\) are empirically set as 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.812, + 0.906, + 0.856 + ], + "angle": 0, + "content": "Loss for SFM. The loss used in the second stage between the ground truth label \\( Q \\) and the final attractiveness score \\( \\hat{Q} \\) is the L1 Loss:" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.865, + 0.905, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {3} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left| \\widehat {Q} _ {i} - Q _ {i} \\right|. \\tag {14}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11463" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.116, + 0.089, + 0.88, + 0.104 + ], + "angle": 0, + "content": "Table 2. Performance comparison on LiveBeauty, MEBeauty [24] and SCUT-FBP5500 [27]. Best in red and second-best in blue." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.116, + 0.9, + 0.339 + ], + "angle": 0, + "content": "
AspectsMethodsLiveBeautyMEBeauty [24]SCUT-FBP5500 [27]
SROCC↑PLCC↑KROCC↑SROCC↑PLCC↑KROCC↑SROCC↑PLCC↑KROCC↑
BaselineViT-B [23]0.8970.8470.7330.6200.6590.4430.8690.8800.686
ResNeXt-50 [56]0.9070.8790.7460.7000.7430.5160.8990.9110.727
Inception-V3 [49]0.9060.8800.7450.6760.7270.4920.9050.9180.736
IAAAVA-MLSP [21]0.8330.8150.6500.6020.6430.4310.7640.7940.566
TANet [18]0.8910.8580.7280.6510.6900.4670.8700.8820.686
Dele-Trans [19]0.8880.8410.7210.6110.6510.4400.8700.8600.682
EAT [20]0.8930.8670.7280.6400.6970.4620.8530.8720.665
FAPComboNet [58]0.9020.8710.7410.6510.6920.4720.8970.9070.725
2D-FAP [32]0.8960.8540.7330.6900.7190.5060.9030.9150.734
REX-INCEP [2]0.9080.8840.7500.7020.7390.5140.9070.9170.739
CNN-ER [2]0.9130.8880.7570.7060.7480.5190.9100.9220.745
MEBeauty [24]0.8260.7960.6430.6490.6770.4670.7990.8020.599
Ours0.9240.8960.7730.7830.7950.5930.9300.9320.774
Improvement+1.1%+0.8%+1.6%+7.7%+4.7%+7.4%+2.0%+1.0%+2.9%
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.363, + 0.223, + 0.381 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.389, + 0.287, + 0.403 + ], + "angle": 0, + "content": "5.1. Benchmark Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.411, + 0.484, + 0.533 + ], + "angle": 0, + "content": "The proposed LiveBeauty dataset and two open-source FAP datasets including SCUT-FBP5500 [27] and MEBeauty [24] are selected as benchmark datasets. The MOS distributions of them are shown in Fig. 7. We split SCUT-FBP5500 and MEBeauty with \\(60\\% - 40\\%\\) and \\(80\\% - 20\\%\\) train-test ratio separately, which is kept same with their own protocol, and LiveBeauty is split using the \\(90\\% - 10\\%\\) protocol." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.543, + 0.308, + 0.559 + ], + "angle": 0, + "content": "5.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.565, + 0.483, + 0.656 + ], + "angle": 0, + "content": "Model initialization. In MAEM, ViT-B/16 and GPT-2 are adopted as the image encoder and the text encoder separately, which are initialized by [42]. In PAPM, Swin-T is adopted as the trainable image encoder, where only four stages are reserved and initialized with parameters pretrained on multiple face-related tasks [41]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.657, + 0.483, + 0.747 + ], + "angle": 0, + "content": "Hyper parameters. During the whole training process, AdamW is selected as the optimizer, and the learning rate scheduler is set with linear warm-up and cosine annealing scheme. The base learning rate varies with different training phase. Each training phase lasts 50 epochs, and the batch size is set as 32. Details can be found in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.757, + 0.307, + 0.773 + ], + "angle": 0, + "content": "5.3. Competitors & Metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To confirm the effectiveness of the proposed FPEM method, we select baseline models and several SOTA FAP and IAA methods, including ComboNet [58], 2D-FAP [32], REX-INCEP [2], CNN-ER [2], MEBeauty [24], AVAMLSP [21], TANet [18], Dele-Trans [19] and EAT [20]. All the models are retrained ourselves with the default parameters defined in their papers. Three criteria are adopted to evaluate the performance of these methods, i.e., Spear" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.368, + 0.64, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.462, + 0.625, + 0.474 + ], + "angle": 0, + "content": "(a)LiveBeauty." + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.368, + 0.768, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.462, + 0.751, + 0.475 + ], + "angle": 0, + "content": "(b) MEBeauty." + }, + { + "type": "image", + "bbox": [ + 0.781, + 0.368, + 0.895, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.784, + 0.462, + 0.892, + 0.474 + ], + "angle": 0, + "content": "(c) SCUT-FBP5500." + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.487, + 0.906, + 0.514 + ], + "angle": 0, + "content": "Figure 8. Scatter plots of the FPEM predicted scores vs. MOSs on three benchmark datasets." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.522, + 0.906, + 0.568 + ], + "angle": 0, + "content": "man Rank Order Correlation Coefficient (SROCC), Pearson Linear Correlation Coefficient (PLCC) and Kendall Rank Order Correlation Coefficient (KROCC)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.577, + 0.651, + 0.592 + ], + "angle": 0, + "content": "5.4. Performance" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.87 + ], + "angle": 0, + "content": "The experimental results on three FAP datasets are shown in Tab. 2, from which we could draw the following conclusions: (i) Our proposed method achieves the first place and surpasses the second place by about 0.011, 0.077, 0.020 in terms of SROCC values on LiveBeauty, MEBeauty and SCUT-FBP5500 respectively, which demonstrates its superiority. (ii) The IAA methods are inferior to the FAP methods, which manifests that the generic aesthetic assessment methods overlook the facial features involved in the subjective nature of facial attractiveness, leading to limited performance on FAP tasks. (iii) The performance of all methods drops on MEBeauty because the number of training samples is restricted and the faces are ethnically diverse in MEBeauty, indicating a large diversity in facial attractiveness. (iv) The significant improvement on MEBeauty over other datasets can be largely attributed to the enhancement brought by face priors, which is more effective in scenarios with insufficient training data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Fig. 8 presents scatter plots illustrating the relationship between the predicted attractiveness scores (vertical axis)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11464" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.077, + 0.483, + 0.135 + ], + "angle": 0, + "content": "Table 3. Ablation study on LiveBeauty test set. \\( E_{FP} \\) denotes Cross-Attention with \\( f_p \\), MS denotes Multi-Scale fusion operation, SR denotes Similarity Regression strategy, \\( \\mathcal{L}_S \\) denotes the L1 scoring loss, \\( \\mathcal{L}_R \\) denotes the total ranking loss \\( \\mathcal{L}_R = \\mathcal{L}_{R_1} + \\mathcal{L}_{R_2} \\)." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.145, + 0.478, + 0.34 + ], + "angle": 0, + "content": "
ModulesMethodsMetrics
SROCC↑PLCC↑KROCC↑
PAPMonlyEMS0.90450.87100.7454
w/oEFp0.90780.87140.7475
w/o MS0.90890.88080.7487
PAPM0.90960.88100.7497
MAEMzero-shot CLIP0.38380.35820.2692
w/o SR0.88250.82260.7123
w/oLS0.91610.84330.7607
w/oLR0.91860.89110.7653
MAEM0.91980.86310.7653
ALLw/o 2-stage0.91890.88890.7652
w/o SFM0.91970.89610.7641
all combined0.92430.89620.7732
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.361, + 0.484, + 0.483 + ], + "angle": 0, + "content": "obtained by FPEM and the MOSs (horizontal axis) in the test sets of three datasets. The MOSs of MEBeauty are divided by two to align the MOS range with others. The color of the scatter points from blue to red indicates the density of points from low to high. In each sub-figure, a second-order polynomial nonlinear fitting curve is superimposed. It can be observed that the fitted curves are close to the diagonal line, demonstrating the effectiveness of FPEM." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.495, + 0.245, + 0.511 + ], + "angle": 0, + "content": "5.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.518, + 0.484, + 0.821 + ], + "angle": 0, + "content": "We conduct elaborate ablation study to investigate the contributions of different modules in our proposed method, which is shown in Tab. 3. Conclusions can be drawn that: (i) In the ablation study of PAPM, we can observe that leveraging facial prior and multi-scale features are both helpful. (ii) The result of ablation study of MAEM indicates that the zero-shot CLIP has limited generalization ability in FAP tasks, which confirms its shortcoming in aesthetic evaluation [45]. Specifically, for MAEM w/o SR, we leverage concatenation of the visual features and the textual features with highest similarity, and utilize MLP for regression. We observe a significant performance drop without SR, which manifests that specifically devised regression strategy is important for CLIP to adapt to regression tasks. Meanwhile, the joint optimization of the ranking and regression also leads to the improvement of SROCC. (iii) In the ablation study of ALL modules, we can see that employing all modules yields the best performance, which confirms the effectiveness of using SFM to adaptively integrate multi-modal aesthetic features with holistic facial prior knowledge." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.285, + 0.849 + ], + "angle": 0, + "content": "5.6. Cross-Dataset Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Considering the diversity of facial attractiveness perception, the cross-dataset validation is carried out to validate the generalization ability of these methods. Specifically," + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.077, + 0.907, + 0.12 + ], + "angle": 0, + "content": "Table 4. Cross-Dataset study trained on LiveBeauty and tested on the entire MEBeauty dataset and the entire SCUT-FBP5500 dataset. Best in red and second-best in blue." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.131, + 0.9, + 0.339 + ], + "angle": 0, + "content": "
MethodsMEBeauty [24]SCUT-FBP5500 [27]
SROCC↑PLCC↑KROCC↑SROCC↑PLCC↑KROCC↑
ViT-B [23]0.5730.5830.4020.5530.5480.394
ResNeXt-50 [56]0.5350.5460.3720.4780.4670.335
Inception-V3 [49]0.5750.5810.4040.4840.4800.339
AVA-MLSP [21]0.5870.5970.4140.4000.4120.275
TANet [18]0.5510.5290.3840.4010.4110.277
Dele-Trans[19]0.5760.5720.4020.4800.4760.335
EAT[20]0.6030.6240.4270.4840.4660.338
ComboNet [58]0.5400.5480.3750.3840.3750.264
2D-FAP [32]0.5560.4890.3900.4440.4450.306
REX-INCEP [2]0.5880.5920.4130.4730.4710.332
CNN-ER [2]0.5890.5960.4140.5020.4990.352
MEBeauty [24]0.5200.5160.3610.3660.3390.251
FPEM (Ours)0.6300.6290.4470.5990.5770.434
Improvement+2.7%+0.5%+2.0%+4.6%+2.9%+4.0%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.353, + 0.907, + 0.381 + ], + "angle": 0, + "content": "Table 5. Cross-Dataset Ablation Study on each module. All models are trained on LiveBeauty and tested on the whole MEBeauty." + }, + { + "type": "table", + "bbox": [ + 0.545, + 0.392, + 0.872, + 0.452 + ], + "angle": 0, + "content": "
PAPMMAEMSFMSROCC↑PLCC↑KROCC↑
✓0.61340.61410.4349
✓0.62290.62680.4244
✓✓✓0.62990.62910.4472
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.47, + 0.906, + 0.514 + ], + "angle": 0, + "content": "all the mentioned models are trained on the proposed LiveBeauty dataset and subsequently tested on the entire SCUTFBP5500 and the entire MEBeauty." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.515, + 0.907, + 0.668 + ], + "angle": 0, + "content": "As shown in Tab. 4, the proposed method outperforms all other methods on both two datasets, which demonstrates its strong generalization ability. Moreover, ablation results of the generalizability on MEBeauty are presented in Tab. 5. It can be observed that MAEM surpasses PAPM by leveraging multi-modal information. The generalizability is further improved by fusing their features through SFM, which indicates that integrating facial prior containing rich contextual and individual information is effective not only for improving performance but also for enhancing generalizability." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.68, + 0.634, + 0.695 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In this paper, we tackle the problem of facial attractiveness prediction in live videos that undergo face retouching. To address the existing gap in datasets, we introduce LiveBeauty, the first large-scale FAP dataset featuring beautiful face images for aesthetic purposes. This dataset is high-quality and authentic, collected from real-world environments, and is currently the largest open-access facial attractiveness dataset. We further propose FPEM, a multimodal FAP framework that selectively fuses holistic facial priors and aesthetic semantic features. Extensive experiments demonstrate the advantages of FPEM over state-of-the-art approaches. Meanwhile, LiveBeauty and FPEM can be effectively leveraged for FAP in other video applications." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11465" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.169 + ], + "angle": 0, + "content": "[1] Lacey Best-Rowden and Anil K Jain. Learning face image quality from human assessments. IEEE Transactions on Information Forensics and Security, 13(12):3064-3077, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.227 + ], + "angle": 0, + "content": "[2] Fares Bougourzi, Fadi Dornaika, and Abdelmalik Taleb-Ahmed. Deep learning based face beauty prediction via dynamic robust losses and ensemble regression. Knowledge-Based Systems, 242:108246, 2022. 1, 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.23, + 0.483, + 0.285 + ], + "angle": 0, + "content": "[3] Djamel Eddine Boukhari, Ali Chemsa, and Riadh Ajgou. Facial beauty prediction based on vision transformer. International Journal of Electrical and Electronic Engineering & Telecommunications, pages 2319-2518, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.287, + 0.484, + 0.356 + ], + "angle": 0, + "content": "[4] Fadi Boutros, Meiling Fang, Marcel Klemt, Biying Fu, and Naser Damer. Cr-fiaq: face image quality assessment by learning sample relative classifiability. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5836-5845, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.359, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[5] Kerang Cao, Kwang-nam Choi, Hoekyung Jung, and Lini Duan. Deep learning for facial beauty prediction. Information, 11(8):391, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.402, + 0.483, + 0.47 + ], + "angle": 0, + "content": "[6] Nicolas Chahine, Stefania Calarasanu, Davide Garcia-Civiero, Theo Cayla, Sira Ferradans, and Jean Ponce. An image quality assessment dataset for portraits. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9968-9978, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.473, + 0.484, + 0.556 + ], + "angle": 0, + "content": "[7] Wei-Ting Chen, Gurunandan Krishnan, Qiang Gao, SyYen Kuo, Sizhou Ma, and Jian Wang. Dsl-fqa: Assessing facial image quality via dual-set degradation learning and landmark-guided transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2931-2941, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.558, + 0.482, + 0.613 + ], + "angle": 0, + "content": "[8] Ritendra Datta, Dhiraj Joshi, Jia Li, and James Z Wang. Studying aesthetics in photographic images using a computational approach. In Proceedings of the European Conference on Computer Vision, pages 288-301, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.615, + 0.482, + 0.655 + ], + "angle": 0, + "content": "[9] Harm De Vries and Jason Yosinski. Can deep learning help you find the perfect match? arXiv preprint arXiv:1505.00359, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.657, + 0.483, + 0.699 + ], + "angle": 0, + "content": "[10] Yubin Deng, Chen Change Loy, and Xiaou Tang. Image aesthetic assessment: An experimental survey. IEEE Signal Processing Magazine, 34(4):80-106, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.702, + 0.483, + 0.769 + ], + "angle": 0, + "content": "[11] Abhishek Dutta, Raymond Veldhuis, and Luuk Spreeuwers. A bayesian model for predicting face recognition performance using image quality. In Proceeding of the International Joint Conference on Biometrics, pages 1-8, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.772, + 0.483, + 0.813 + ], + "angle": 0, + "content": "[12] Yael Eisenthal, Gideon Dror, and Eytan Ruppin. Facial attractiveness: Beauty and the machine. Neural Computation, 18(1):119-142, 2006. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.816, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[13] Yang-Yu Fan, Shu Liu, Bo Li, Zhe Guo, Ashok Samal, Jun Wan, and Stan Z Li. Label distribution-based facial attractiveness computation by deep residual learning. IEEE Transactions on Multimedia, 20(8):2196-2208, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[14] Junying Gan, Xiaoshan Xie, Guohui He, and Heng Luo. Transbls: transformer combined with broad learning system" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "for facial beauty prediction. Applied Intelligence, 53(21): 26110-26125, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[15] Junying Gan, Xiaoshan Xie, Yikui Zhai, Guohui He, Chaoyun Mai, and Heng Luo. Facial beauty prediction fusing transfer learning and broad learning system. Soft Computing, 27(18):13391-13404, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.905, + 0.244 + ], + "angle": 0, + "content": "[16] Lian Gao, Weixin Li, Zehua Huang, Di Huang, and Yunhong Wang. Automatic facial attractiveness prediction by deep multi-task learning. In Proceedings of the IEEE International Conference on Pattern Recognition, pages 3592-3597, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.3 + ], + "angle": 0, + "content": "[17] Douglas Gray, Kai Yu, Wei Xu, and Yihong Gong. Predicting facial beauty without landmarks. In Proceedings of the European Conference on Computer Vision, pages 434-447, 2010. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[18] Shuai He, Yongchang Zhang, Rui Xie, Dongxiang Jiang, and Anlong Ming. Rethinking image aesthetics assessment: Models, datasets and benchmarks. In Proceedings of the International Joint Conference on Artificial Intelligence, pages 942–948, 2022. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.373, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[19] Shuai He, Anlong Ming, Yaqi Li, Jinyuan Sun, ShunTian Zheng, and Huadong Ma. Thinking image color aesthetics assessment: Models, datasets and benchmarks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21838-21847, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[20] Shuai He, Anlong Ming, Shuntian Zheng, Haobin Zhong, and Huadong Ma. Eat: An enhancer for aesthetics-oriented transformers. In Proceedings of the ACM International Conference on Multimedia, pages 1023-1032, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.905, + 0.564 + ], + "angle": 0, + "content": "[21] Vlad Hosu, Bastian Goldlucke, and Dietmar Saupe. Effective aesthetics prediction with multi-level spatially pooled features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9375-9383, 2019. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.568, + 0.905, + 0.648 + ], + "angle": 0, + "content": "[22] Tharun J. Iyer, Rahul K, Ruban Nersisson, Zhemin Zhuang, Alex Noel Joseph Raj, and Imthiaz Refayee. Machine learning-based facial beauty prediction and analysis of frontal facial images using facial landmarks and traditional image descriptors. Computational Intelligence and Neuroscience, 2021(1):4423407, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.651, + 0.905, + 0.746 + ], + "angle": 0, + "content": "[23] Alexander Kolesnikov, Alexey Dosovitskiy, Dirk Weissenborn, Georg Heigold, Jakob Uszkoreit, Lucas Beyer, Matthias Minderer, Mostafa Dehghani, Neil Houlsby, Sylvain Gelly, Thomas Unterthiner, and Xiaohua Zhai. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2021. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.748, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[24] Irina Lebedeva, Yi Guo, and Fangli Ying. Mebeauty: a multi-ethnic facial beauty dataset in-the-wild. *Neural Computing and Applications*, pages 1–15, 2022. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[25] Congcong Li, Andrew Gallagher, Alexander C. Loui, and Tsuhan Chen. Aesthetic quality assessment of consumer photos with faces. In Proceedings of the IEEE International Conference on Image Processing, pages 3221-3224, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[26] Xiang Li, Wenhai Wang, Xiaolin Hu, and Jian Yang. Selective kernel networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 510-519, 2019. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11466" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[27] Lingyu Liang, Luojun Lin, Lianwen Jin, Duorui Xie, and Mengru Li. Scut-fbp5500: A diverse benchmark dataset for multi-paradigm facial beauty prediction. In Proceedings of the IEEE International Conference on Pattern Recognition, pages 1598-1603, 2018. 1, 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.483, + 0.232 + ], + "angle": 0, + "content": "[28] Luojun Lin, Lingyu Liang, Lianwen Jin, and Weijie Chen. Attribute-aware convolutional neural networks for facial beauty prediction. In Proceedings of the Information Joint Conference on Artificial Intelligence, pages 847-853, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.484, + 0.302 + ], + "angle": 0, + "content": "[29] Qipeng Liu, Luojun Lin, Zhifeng Shen, and Yuanlong Yu. Fbpformer: Dynamic convolutional transformer for global-local-contextual facial beauty prediction. In Proceedings of the International Conference on Artificial Neural Networks, pages 223-235, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.483, + 0.357 + ], + "angle": 0, + "content": "[30] Shu Liu, Yang-Yu Fan, Zhe Guo, Ashok Samal, and Afan Ali. A landmark-based data-driven approach on 2.5 d facial attractiveness computation. Neurocomputing, 238:168-178, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.359, + 0.484, + 0.426 + ], + "angle": 0, + "content": "[31] Shu Liu, Bo Li, Yang-Yu Fan, Zhe Quo, and Ashok Samal. Facial attractiveness computation by label distribution learning with deep cnn and geometric features. In Proceedings of the IEEE International Conference on Multimedia and Expo, pages 1344-1349, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.483, + 0.482 + ], + "angle": 0, + "content": "[32] Shu Liu, Enquan Huang, Yan Xu, Kexuan Wang, Xiaoyan Kui, Tao Lei, and Hongying Meng. Lightweight facial attractiveness prediction using dual label distribution. arXiv preprint arXiv:2212.01742, 2022. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.483, + 0.538 + ], + "angle": 0, + "content": "[33] Shu Liu, Enquan Huang, Yan Xu, Kexuan Wang, Xiaoyan Kui, Tao Lei, and Hongying Meng. Lightweight facial attractiveness prediction using dual label distribution. arXiv preprint arXiv:2212.01742, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.483, + 0.594 + ], + "angle": 0, + "content": "[34] Tie Liu, Shengxi Li, Mai Xu, Li Yang, and Xiaofei Wang. Assessing face image quality: A large-scale database and a transformer method. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.483, + 0.663 + ], + "angle": 0, + "content": "[35] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.483, + 0.72 + ], + "angle": 0, + "content": "[36] Naila Murray, Luca Marchesotti, and Florent Perronnin. Ava: A large-scale database for aesthetic visual analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2408-2415, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.721, + 0.483, + 0.803 + ], + "angle": 0, + "content": "[37] Fu-Zhao Ou, Xingyu Chen, Ruixin Zhang, Yuge Huang, Shaoxin Li, Jilin Li, Yong Li, Liujuan Cao, and Yuan-Gen Wang. Sdd-fqa: Unsupervised face image quality assessment with similarity distribution distance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7670–7679, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.805, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[38] Fu-Zhao Ou, Chongyi Li, Shiqi Wang, and Sam Kwong. Clib-fiqa: Face image quality assessment with confidence calibration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1694-1704, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[39] Tianhao Peng, Mu Li, Fangmei Chen, Yong Xu, and David Zhang. Geometric prior guided hybrid deep neural network" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "for facial beauty analysis. CAAI Transactions on Intelligence Technology, 9(2):467-480, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.207 + ], + "angle": 0, + "content": "[40] P Jonathon Phillips, J Ross Beveridge, David S Bolme, Bruce A Draper, Geof H Givens, Yui Man Lui, Su Cheng, Mohammad Nayeem Teli, and Hao Zhang. On the existence of face quality measures. In Proceedings of the IEEE International Conference on Biometrics: Theory, Applications and Systems, pages 1-8, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.208, + 0.905, + 0.277 + ], + "angle": 0, + "content": "[41] Lixiong Qin, Mei Wang, Chao Deng, Ke Wang, Xi Chen, Jiani Hu, and Weihong Deng. Swinface: a multi-task transformer for face recognition, expression recognition, age estimation and attribute estimation. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.361 + ], + "angle": 0, + "content": "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proceedings of the International Conference on Machine Learning, pages 8748-8763, 2021. 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[43] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 815-823, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.905, + 0.475 + ], + "angle": 0, + "content": "[44] B Series. Methodology for the subjective assessment of the quality of television pictures. Recommendation ITU-R BT, 500(13), 2012. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.907, + 0.559 + ], + "angle": 0, + "content": "[45] Xiangfei Sheng, Leida Li, Pengfei Chen, Jinjian Wu, Weisheng Dong, Yuzhe Yang, Liwu Xu, Yaqian Li, and Guangming Shi. Aesclip: Multi-attribute contrastive learning for image aesthetics assessment. In Proceedings of the ACM International Conference on Multimedia, page 1117–1126, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[46] Shengjie Shi, Fei Gao, Xuantong Meng, Xingxin Xu, and Jingjie Zhu. Improving facial attractiveness prediction via co-attention learning. In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing, pages 4045–4049, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[47] Shaolin Su, Hanhe Lin, Vlad Hosu, Oliver Wiedemann, Jinqiu Sun, Yu Zhu, Hantao Liu, Yanning Zhang, and Dietmar Saupe. Going the extra mile in face image quality assessment: A novel database and model. IEEE Transactions on Multimedia, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.703, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[48] Wei Sun, Weixia Zhang, Yanwei Jiang, Haoning Wu, Zicheng Zhang, Jun Jia, Yingjie Zhou, Zhongpeng Ji, Xiongkuo Min, Weisi Lin, and Zhai Guangtao. Dual-branch network for portrait image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[49] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2826, 2016. 1, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[50] Philipp Terhorst, Jan Niklas Kolf, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. Ser-fiq: Unsupervised estimation of face image quality based on stochastic embedding" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "11467" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.133 + ], + "angle": 0, + "content": "robustness. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5651-5660, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.163 + ], + "angle": 0, + "content": "[51] Louis L Thurstone. A law of comparative judgment. In *Scaling*, pages 81–92. Routledge, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.484, + 0.22 + ], + "angle": 0, + "content": "[52] Yu Tian, Zhangkai Ni, Baoliang Chen, Shiqi Wang, Hanli Wang, and Sam Kwong. Generalized visual quality assessment of gan-generated face images. arXiv preprint arXiv:2201.11975, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.222, + 0.483, + 0.262 + ], + "angle": 0, + "content": "[53] Song Tong, Xuefeng Liang, Takatsune Kumada, and Sunao Iwaki. Putative ratios of facial attractiveness in a deep neural network. Vision Research, 178:86-99, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.483, + 0.332 + ], + "angle": 0, + "content": "[54] Ming-Feng Tsai, Tie-Yan Liu, Tao Qin, Hsin-Hsi Chen, and Wei-Ying Ma. Frank: a ranking method with fidelity loss. In Proceedings of the ACM SIGIR Conference on Research and Development in Information Retrieval, pages 383-390, 2007. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.483, + 0.39 + ], + "angle": 0, + "content": "[55] Qinjie Xiao, You Wu, Dinghong Wang, Yong-Liang Yang, and Xiaogang Jin. Beauty3dfacenet: deep geometry and texture fusion for 3d facial attractiveness prediction. Computers & Graphics, 98:11-18, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.483, + 0.46 + ], + "angle": 0, + "content": "[56] Saining Xie, Ross Girshick, Piotr Dolkar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1492-1500, 2017. 1, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.483, + 0.544 + ], + "angle": 0, + "content": "[57] Jie Xu, Lianwen Jin, Lingyu Liang, Ziyong Feng, Duorui Xie, and Huiyun Mao. Facial attractiveness prediction using psychologically inspired convolutional neural network (pCNN). In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing, pages 1657-1661, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.547, + 0.483, + 0.588 + ], + "angle": 0, + "content": "[58] Lu Xu and Jinhai Xiang. Comboloss for facial attractiveness analysis with squeeze-and-excitation networks. arXiv preprint arXiv:2010.10721, 2020. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.483, + 0.631 + ], + "angle": 0, + "content": "[59] Lu Xu and Jinhai Xiang. Comboloss for facial attractiveness analysis with squeeze-and-excitation networks. arXiv preprint arXiv:2010.10721, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.483, + 0.688 + ], + "angle": 0, + "content": "[60] Lu Xu, Heng Fan, and Jinhai Xiang. Hierarchical multi-task network for race, gender and facial attractiveness recognition. In Proceedings of the IEEE International Conference on Image Processing, pages 3861-3865, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.483, + 0.73 + ], + "angle": 0, + "content": "[61] Haibin Yan. Cost-sensitive ordinal regression for fully automatic facial beauty assessment. Neurocomputing, 129:334-342, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[62] Qinkai Yu, Jianyang Xie, Anh Nguyen, He Zhao, Jiong Zhang, Huazhu Fu, Yitian Zhao, Yalin Zheng, and Yanda Meng. Clip-dr: Textual knowledge-guided diabetic retinopathy grading with ranking-aware prompting. arXiv preprint arXiv:2407.04068, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[63] Yikui Zhai, Yu Huang, Ying Xu, Junying Zeng, Fei Yu, and Junying Gan. Benchmark of a large scale database for facial beauty prediction. In Proceedings of the International Conference on Intelligent Information Processing, pages 1-5, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[64] Yikui Zhai, Yu Huang, Ying Xu, Junying Gan, He Cao, Wenbo Deng, Ruggero Donida Labati, Vincenzo Piuri, and" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Fabio Scotti. Asian female facial beauty prediction using deep neural networks via transfer learning and multi-channel feature fusion. IEEE Access, 8:56892-56907, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[65] Shifeng Zhang, Xiangyu Zhu, Zhen Lei, Hailin Shi, Xiaobo Wang, and Stan Z Li. Faceboxes: A cpu real-time face detector with high accuracy. In Proceedings of the IEEE International Joint Conference on Biometrics, pages 1-9, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.206, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[66] Weixia Zhang, Guangtao Zhai, Ying Wei, Xiaokang Yang, and Kede Ma. Blind image quality assessment via vision-language correspondence: A multitask learning perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14071-14081, 2023. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "11468" + } + ] +] \ No newline at end of file diff --git a/2025/FPEM_ Face Prior Enhanced Facial Attractiveness Prediction for Live Videos with Face Retouching/f47f8d6d-8a6c-491b-8f19-08123eb17e5c_origin.pdf b/2025/FPEM_ Face Prior Enhanced Facial Attractiveness Prediction for Live Videos with Face Retouching/f47f8d6d-8a6c-491b-8f19-08123eb17e5c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7025f5afdddde33fe4a156b2cad5e9fb13729d42 --- /dev/null +++ b/2025/FPEM_ Face Prior Enhanced Facial Attractiveness Prediction for Live Videos with Face Retouching/f47f8d6d-8a6c-491b-8f19-08123eb17e5c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9cc67af925aeb79d571c52ed3eb9e41c024e50c0720657e5bb7ecc957378b383 +size 4551185 diff --git a/2025/FPEM_ Face Prior Enhanced Facial Attractiveness Prediction for Live Videos with Face Retouching/full.md b/2025/FPEM_ Face Prior Enhanced Facial Attractiveness Prediction for Live Videos with Face Retouching/full.md new file mode 100644 index 0000000000000000000000000000000000000000..89c23e11499afcdc777a2fc279877deeca8e5e6d --- /dev/null +++ b/2025/FPEM_ Face Prior Enhanced Facial Attractiveness Prediction for Live Videos with Face Retouching/full.md @@ -0,0 +1,372 @@ +# FPEM: Face Prior Enhanced Facial Attractiveness Prediction for Live Videos with Face Retouching + +Hui Li $^{1\dagger}$ , Xiaoyu Ren $^{1,2\dagger}$ , Hongjiu Yu $^{1}$ , Ying Chen $^{1*}$ , Kai Li $^{1}$ , L Wang $^{1}$ , Xiongkuo Min $^{2}$ , Huiyu Duan $^{2}$ , Guangtao Zhai $^{2}$ , Xu Liu $^{1}$ + +1 Alibaba Group 2Shanghai Jiao Tong University + +{lh401704,naomi.rxy,yuhongjiu.yhj,YingChen,kaishi.lk,zhuangshu.wlb,steven.liu}@taobao.com + +{minxiongkuo,huiyuduan,zhaiguangtao}@sjtu.edu.cn + +# Abstract + +Facial attractiveness prediction (FAP) has long been an important computer vision task, which could be widely applied in live videos with facial retouching. However, previous FAP datasets are either small or closed-source. Moreover, the corresponding FAP models exhibit limited generalization and adaptation ability. To overcome these limitations, we introduce the first large-scale FAP dataset LiveBeauty specifically designed for live video scenarios wherein face images may be real-time processed for aesthetics purposes. 10,000 face images are collected directly from a live streaming platform, with 200,000 corresponding attractiveness annotations obtained from a well-devised subjective experiment, making LiveBeauty the largest open-access FAP dataset. Based on the built dataset, a novel FAP method named Facial Prior Enhanced Multi-modal model (FPEM) is proposed to measure the attractiveness of facial images. Extensive experiments conducted on both LiveBeauty and other open-source FAP datasets demonstrate that our proposed method achieves state-of-the-art performance. The dataset will be released at https://github.com/EstellaLH/FPEM. + +# 1. Introduction + +With the rapid development of live streaming, short-form video, and advanced face retouching within such applications, the perceptual quality of faces becomes an important factor for the overall visual experience of users, which is particularly essential for video platforms, as it can significantly affect the average viewer watch time and even GMV (Gross Merchandise Volume). A robust metric system for evaluating facial clarity and aesthetics is essential to monitor both the overall service quality of the platform and + +![](images/c0f4409e3000fe150b3c0f06d40319855c6edc3c7b11e9f8b00c7d4d939d8af5.jpg) +Figure 1. SROCC vs PLCC comparison on LiveBeauty and SCUT-FBP5500 [27]. FPEM, denoted by red triangular, outperforms other methods (CNN-ER [2], 2D-FAP [32], REX-INCEP [2], Inception-V3 [49], ResNeXt-50 [56]) on both datasets. + +![](images/9ef842e72aa080ba2473e92ee4f8bded9293a0e5964df33b70d5ea313b2a5ef7.jpg) + +the visual appeal changes or improvements of individual streaming channels. The clarity-related assessment tasks of facial quality could be categorized into Biometric Face Image Quality Assessment (BFIQA) and Generic Face Image Quality Assessment (GFIQA). BFIQA focuses on the identity information of faces for down-stream tasks such as face recognition [1, 37, 40], while GFIQA emphasizes the visual perceptual quality of faces [6, 47, 52]. The aesthetics-related assessment of facial quality is often referred to as Facial Attractiveness Prediction (FAP) [12, 27, 28]. The goal of FAP is to quantify the attractiveness of facial images, which is vital for assessing the effectiveness of facial retouching settings and offers valuable insights for algorithm development. + +As shown in Tab. 1, numerous FAP datasets have been proposed in recent years, many of which encompass a rich diversity in age, gender, face expression and pose [9, 17, 24, 27, 53, 61, 63, 64]. However, most of these datasets are either small or unavailable due to privacy and ethical concerns. In video streaming, especially live video, it is common practice to use facial retouching algorithms to improve the facial aesthetics of content presented to audiences. + +Table 1. Overview of common facial attractiveness prediction datasets. + +
DatasetYearRating Scale# ImageGenderEthnicity# AnnotatorPublic
Gray et al. [17]2010[-3, 3]2,056F MCaucasian30✓
MBW+FBW [61]2014[1, 5]10,400F MMulti30-
De Vries et al. [9]2015Binary9364F MAsian20-
LSFBD [63]2016[1, 5]20,000F M---
SCUT-FBP5500 [27]2018[1, 5]5,500F MMulti60✓
Tong et al. [53]2020Binary4,512F M-20-
LSAFBD [64]2020[1, 5]20,000FAsian20-
MEBeauty [24]2022[1, 10]2,550F MMulti300✓
LiveBeauty (proposed)2024[1, 5]10,000F MAsian20✓
+ +The differences among these algorithms can lead to unique data distributions that vary considerably from those found in unprocessed or well-edited face images obtained from the Internet. Meanwhile, the effectiveness of face retouching is largely dependent on the computational resources available, which in turn affects these data characteristics. Since the training data impacts the generalization ability of data-driven deep learning methods, the lack of a FAP dataset specifically collected from live videos with authentic face retouching process may lead to limited performance in practical video scenarios. + +Meanwhile, there is no widely recognized FAP model specifically designed for live videos. Early methods relying on hand-crafted beauty descriptors [22, 30] tend to exhibit limited performance due to the inadequate feature representation capability. The success of the deep-learning based methods, e.g. Convolution Neural Networks (CNNs) [2, 32, 58] and Transformers [29, 39], have driven progress in the FAP task. Specifically, [39] first introduced a Vision Transformer into the FAP task and proposed a geometric prior guided network leveraging facial landmarks. Although their approach demonstrates promising performance, the straightforward feature concatenation and simplistic prior struggle to adaptively model individual facial features, thereby limiting its applicability and effectiveness in real-world applications. + +To address the lack of FAP datasets considering the impact of face retouching, we first construct a large-scale facial attractiveness dataset LiveBeauty, which includes 10,000 face images collected from a popular video streaming platform, with the corresponding facial attractiveness annotations obtained from 20 annotators. + +Furthermore, based on the established dataset, we propose a no-reference FAP model termed Facial Prior Enhanced Multi-modal model (FPEM), to predict the facial attractiveness by extracting and integrating holistic facial prior knowledge and multi-modal aesthetic semantic features. Specifically, we first devise a Personalized Attractiveness Prior Module (PAPM) to integrate extensive facial prior knowledge, in which multi-scale visual features + +extracted by the Swin Transformer [35] and facial prior features extracted by the pretrained FaceNet [43] are integrated via a Cross-Attention block. To obtain facial beauty attributes, we also introduce a Multi-modal Attractiveness Encoder Module (MAEM) based on CLIP [42] to extract multi-modal aesthetic semantic features. Furthermore, these extracted features are adaptively integrated through a fusion module. Extensive experimental results confirm that our proposed method achieves the best performance compared to other state-of-the-art (SOTA) methods on both our LiveBeauty dataset and other FAP datasets, as shown in Fig. 1. + +To summarize, our contributions are as follow: + +- We construct the LiveBeauty dataset, the first large-scale, open-source facial attractiveness prediction dataset sourced from a real-world live video platform featuring faces enhanced for aesthetics, making it suitable for applications wherein facial aesthetics are vital and therefore need to be well modeled and monitored. +- We propose a multi-modal method specifically designed for FAP, predicting attractiveness scores for face images. By integrating visual and textual modalities, this method can provide a more reliable prediction of facial appeal. +- We enhance the representation capability of our model by leveraging facial prior knowledge, and further employing selective fusion module to integrate and refine the extracted visual features. + +# 2. Related Works + +# 2.1. Aesthetic Assessment of Generic Images + +Image Aesthetic Assessment (IAA) task aims to assess image quality from the aesthetic perspective. The corresponding IAA methods have evolved from employing handcrafted features, the modeling of photographic rules [8], global image layout [36] or typical objects [25] in perceived images [10] to data-driven approaches based on deep learning techniques that deliver superior performance [18-21]. Notably, these IAA methods primarily concentrate on generic images rather than specifically targeting facial regions. + +# 2.2. Quality Assessment of Face Images + +Facial clarity related assessment include both BFIQA and GFIQA. BFIQA was originated from biometric research focusing on evaluating the suitability of facial images for recognition, initially relying on hand-crafted features [11, 50], and recent advancing to deep learning approaches with regression and embedding-based solutions [4, 38]. Nonetheless, Chen et al. [7] noted these methods may not generalize well on other facial quality and aesthetic tasks due to the emphasis on facial biometric information. In contrast, GFIQA is defined to evaluate the conventional perceptual quality of face images. GFIQA methods have advanced through the use of transformer-based [7, 34] and multi-modal approaches [48]. Clearly, GFIQA focuses on modeling the degradation of facial regions, such as issues related to blur and noise, and is therefore orthogonal to facial attractiveness. + +# 2.3. Attractiveness Assessment of Face Images + +In the context of this paper, facial aesthetics primarily pertains to attractiveness. To measure the facial attractiveness of face images, early handcrafted Facial Attractiveness Prediction (FAP) methods predominantly rely on manually designed features, e.g. geometric features [12, 22, 30], traditional image descriptors [12, 22, 27], holistic descriptors [12], etc. However, these methods exhibit limited performance due to difficulties in presenting comprehensive characteristics of facial attractiveness. + +With the advancement in deep learning, FAP solutions based on Convolution Neural Networks (CNNs) show better performance. Inspired by psychological studies, specific architectures of CNNs are designed [5, 57]. The applications of Label Distribution Learning (LDL) [13, 31, 33] achieve better performance. Some methods [16, 46, 55] combine geometric prior knowledge containing landmarks, parsing maps, etc., while the others [28, 60] integrate attractiveness-related attributes such as gender, race, etc. In addition, the ensemble solutions of different loss functions like regression loss and classification loss [59] and different models including ResNeXt50 and Inception-V3 [2] are further explored. However, these CNN-based methods primarily focus on local features, limiting their capacity for long-term modeling. To address this problem, transformer-based solutions [3, 14] are proposed, while Liu et al. [29] and Gan et al. [15] combine transformers with CNNs to leverage both global and local features related to facial attractiveness. Nevertheless, these approaches have limited generalization and adaptation ability in more challenging scenarios, due to the biased predictions from training data skewed towards certain demographics, and using straightforward single-modal modeling techniques which rely on simplistic concatenation of visual features. + +![](images/5b96182c2ccf8cfdbbe76555526f15d55a289d08dc13595c1d73437b8f57134d.jpg) +Figure 2. Examples of the face images in our LiveBeauty dataset (from left to right are examples rated 1 to 5). The first row is male and second row is female. + +# 3. Dataset + +# 3.1. Images Collection + +To accurately reflect the distribution of facial attractiveness of in-captured videos, we first collect large-scale raw face images from one of the most popular live streaming platforms. Specifically, we retrieve replays for live channels of all streamers conducted on the platform in March 2024 with relatively high Page Views (PVs). Given that each channel lasts approximately 2 to 10 hours, we sample one image per hour for the first three hours of each channel, resulting in three images sampled from each channel. Then, a SOTA face detection model [65] is employed to generate the bounding box (bbox) of the face region. + +Data cleaning is further conducted to eliminate samples containing animated characters, virtual avatars, and samples from the same individual, to ensure the quality and diversity of the dataset. Finally, 10,000 images are collected, ensuring 10,000 different faces from over 9,500 live channels. Fig. 2 shows some examples of the face images in our LiveBeauty dataset. + +# 3.2. Subjective Experiments + +We recruit 20 subjects in our experiments, including 14 females and 6 males. This ratio is determined by the actual customer demographics of the live platform. Face stimulus are displayed on an iPhone 14 Pro Max with 6.7-inch Organic Light-Emitting Diode (OLED) screen, supporting a resolution up to $2,796 \times 1,290$ . Subjective experiments are conducted in a well-controlled lab environment with fixed settings [44], i.e., lighting, viewing distance, viewing angle, etc. + +We split the experiment into 200 sessions with 50 images per session. Each face image is displayed for 10 seconds and followed by a one-second gray screen mask. Before the experiment starting, we offer a brief explanation and 50 extra training samples for subjects to familiarize them with the task. During the experiment, subjects are asked to evaluate the facial attractiveness of the stimuli from 1 to 5, defined as: "unappealing", "less favorable", "ordinary", "online influencer" and "celebrity", with detailed descriptions + +![](images/bf14c43aac92b4d22ef2b3368d64627042e8ae27de4b91c394d5a94b8aa41be3.jpg) +Figure 3. Illustration of the LiveBeauty MOS distribution. Skewness=-0.470 Kurtosis=1.498 + +of makeup and skin condition. After each session, subjects are asked to take at least 5 minutes of break. All subjects participate in all sessions, and each subject is allowed to participate in no more than 5 sessions each day. Ultimately, all 10,000 images are evaluated by 20 subjects, resulting in a total of 200,000 $(10,000 \times 20)$ scores collected. + +# 3.3. Data Analysis + +According to [44], we employ subject post-screening with outlier ratio and Spearman Rank Correlation Coefficient (SROCC). Annotations that differ by more than two levels from the mode of all annotations are considered potential outliers and will be meticulously removed by the experts. A subject whose rated scores with SROCC $< 0.75$ or outlier ratio $>2\%$ will be marked as unreliable and will be rejected. Ultimately, 20 subjects are retained. We compute the Mean Opinion Score (MOS) as the attractiveness label of each face image: + +$$ +\mathrm {M O S} _ {j} = \frac {1}{N} \sum_ {i = 1} ^ {N} r _ {i j}, \tag {1} +$$ + +where $\mathrm{MOS}_j$ represents the MOS for the $j$ -th face image, $N$ is the number of the valid annotation scores rated by subjects, and $r_{ij}$ is the rated score of the $i$ -th subject on the $j$ -th face image. + +MOS distribution of our LiveBeauty is shown in Fig. 3. It can be observed that the distributions of MOSs exhibit a Gaussian-like shape. Skewness and kurtosis values of three MOS distributions are also shown in Fig. 3, which indicate that the distributions are characterized by thin tails and the MOS is concentrated around 3. This aligns with real-world facial attractiveness distributions, i.e., most individuals possess average facial attractiveness. + +The annotation distribution of four attractive levels of each face sample is shown in Fig. 4. Although attractiveness is highly subjective and influenced by personal preference, it can be observed that the annotation scores for each face are concentrated around the three levels closest to the MOS (e.g., 3.85 comes from five 3, thirteen 4, and two 5). + +![](images/8cc5d4d81ef50b1584debd21ec78c61797f00ce540a77f0478176e5772d57e7a.jpg) +Figure 4. Annotation score distribution on four attractive level. + +# 4. Proposed Method + +As illustrated in Fig. 5, our Facial Prior Enhanced Multi-modal (FPEM) method utilizes a two-stage training strategy, and consists of three primary modules. + +The first stage, namely Preliminary Training Phase, includes the Personalized Attractiveness Prior Module (PAPM in Sec. 4.1.1) and the Multi-modal Attractiveness Encoder Module (MAEM in Sec. 4.1.2). PAPM receives the image $I$ , from which a Swin Transformer $E_{MS}$ extracts multi-scale feature $f_{m}$ , and a frozen FaceNet $E_{FP}$ extracts face-aware feature $f_{s}$ . Subsequently, a Cross-Attention block incorporates $f_{m}$ and $f_{s}$ to form the personalized attractiveness feature $f_{p}$ , which are then projected to a preliminary attractiveness score $\widehat{Q}_1$ . The MAEM contains an image encoder $E_{I}$ and a text encoder $E_{T}$ as in CLIP [42], receives the image $I$ and progressive texts, leverages visual embeddings $f_{i}$ and textual embeddings $f_{t}$ , and obtains another preliminary score $\widehat{Q}_2$ through a Similarity Regression (SR) operation. + +In the second stage, namely Selective Fusion Phase, a Selective Fusion Module (SFM in Sec. 4.2) dynamically refines the extracted visual embeddings $f_{i}$ to $f_{i}^{\prime}$ based on the personalized attractiveness feature $f_{p}$ . Then the SR strategy is used to predict the final score $\widehat{Q}$ . + +# 4.1. Preliminary Training Phase + +In this stage, PAPM and MAEM are trained independently without considering the SFM, and MAEM employs the SR strategy for regression during training. + +# 4.1.1. Personalized Attractiveness Prior Module + +**Stem Swin.** Considering that the visual perception progress is perceived from low level to high level, the hierarchical features from consecutive transformer stages are fully leveraged. Specifically, the features $f_{r_i} \in \mathbb{R}^{H_i \times W_i \times C_i}$ extracted from $i$ -th stage are reshaped into same size, and are then concatenated as $f_{\Sigma} \in \mathbb{R}^{H_4 \times W_4 \times C_{\Sigma}}$ . Meanwhile, $f_{\Sigma}$ is refined by a Channel Attention block from channel perspective (which is shown in Fig. 6), and then flattened as $f_m \in \mathbb{R}^{N_4 \times C_{\Sigma}}$ . + +![](images/324f40abd091c23171097a2525a108107b16200a9f71dbaf42cbb6eb88c4a023.jpg) +Figure 5. Architecture of the Facial Prior Enhanced Multi-modal model (FPEM), including the Personalized Attractiveness Prior Module (PAPM), the Multi-modal Attractiveness Encoder Module (MAEM) and the Selective Fusion Module (SFM). + +Cross-Attention. Facial prior knowledge, i.e., face-aware features $f_{s}$ extracted by FaceNet, are utilized in this block to guide the learning of facial attractiveness. Specifically, the refined features $f_{m}$ extracted by the stem Swin are fused with the face-aware features $f_{s}$ through cross-attention. As shown in Fig. 6, $f_{s}$ is encoded as the query $Q$ via a Linear layer and a Layer Normalization (LN) layer, while $f_{m}$ is encoded as the key $K$ and value $V$ via a LN layer. Further, a Multi-Head Cross-Attention (MHCA) mechanism is applied, which is computed as: + +$$ +\left\{ \begin{array}{l} \operatorname {h e a d} _ {i} = \psi \left(\frac {Q \cdot K ^ {T}}{\sqrt {d}}\right) \cdot V \cdot W _ {i}, i \in \{1, \dots , h \} \\ \operatorname {M H C A} (Q, K, V) = \operatorname {C o n c a t} (\operatorname {h e a d} _ {i}) W _ {o}, \end{array} \right. \tag {2} +$$ + +where $\psi (\cdot)$ denotes the softmax function, $W_{o}$ denotes the projection parameter matrix and $d$ denotes the query dimension. Let FFN indicates the Feed-Forward Network, AP denotes Average Pooling, the output feature $f_{p}\in \mathbb{R}^{1\times C}$ is computed as: + +$$ +\left\{ \begin{array}{l} f _ {c} = f _ {m} + \mathrm {M H C A} (Q, K, V) \\ f _ {p} = \mathrm {A P} \left(f _ {c} + \text {F F N} \left(\mathrm {L N} \left(f _ {c}\right)\right)\right). \end{array} \right. \tag {3} +$$ + +Let MLP represent the Multi Layer Perceptron, then the attractiveness score $\widehat{Q}_1$ can be predicted by: + +$$ +\widehat {Q} _ {1} = \operatorname {M L P} \left(f _ {p}\right). \tag {4} +$$ + +# 4.1.2. Multi-modal Attractiveness Encoder Module + +Inspired by [66], MAEM is devised to utilize multi-modal aesthetic semantic features. We utilize five attractiveness + +![](images/22e36a6f14da6bdfacc30bf404eaf2ee5197c36afb4662059e959d96b30ce523.jpg) +Figure 6. Structure of the Channel Attention and the Cross-Attention Block. + +![](images/4efb84d53be049b81efa9c9cf066a409987082ea995876dbae46e5ee7f355b99.jpg) + +levels $a \in \mathcal{A} = \{1,2,3,4,5\} = \{\text{"bad", "poor", "fair", "good", "perfect"}\}$ . Then the text template can be formed as "a photo of a person with $\{a\}$ attractiveness", resulting in five candidate text descriptions $\mathcal{T}$ . + +MAEM contains a frozen text encoder $E_{T}$ (GPT-2) and a trainable image encoder $E_{I}$ (ViT-B/16), where $E_{T}$ receives five candidate text descriptions $\mathcal{T}$ to form the textual embeddings $f_{t}^{\prime}\in \mathbb{R}^{5\times C}$ , while $E_{I}$ receives the image $I$ to form the visual embeddings $f_{i}\in \mathbb{R}^{1\times C}$ . + +Subsequently, the textual embeddings $f_{t}^{\prime}$ and visual embeddings $f_{i}$ are fed into Similar Regression (SR). In particular, the cosine similarity between $f_{t}$ and $f_{i}$ is calculated to obtain the attractiveness level probability distribution $\hat{p}$ : + +$$ +\hat {p} = \psi \left(\frac {f _ {i} \cdot f _ {t}}{\| f _ {i} \| \times \| f _ {t} \|}\right), \tag {5} +$$ + +where $\psi (\cdot)$ denotes the softmax function. Then the attractiveness scores $\widehat{Q}_2$ can be predicted by: + +$$ +\widehat {Q} _ {2} = \sum_ {a = 1} ^ {5} \hat {p} (a | I) \cdot a, \tag {6} +$$ + +where $a$ denotes the attractiveness level from 1 to 5, and $\hat{p}(a|\cdot)$ is the probability of the level $a$ to be estimated. + +# 4.2. Selective Fusion Phase + +In this stage, we freeze the parameters of MAEM and PAPM, which are trained in the previous preliminary training phase. Then the Selective Fusion Model (SFM) is trained to dynamically integrate holistic facial prior knowledge and multi-modal aesthetic semantic features. To achieve this goal, we first fuse the visual embeddings $f_{i}$ and personalized attractiveness feature $f_{p}$ via an element-wise addition. Inspired by [26], a compact feature $f_{z} \in \mathbb{R}^{1 \times d}$ is further generated to facilitate adaptive selection guidance for these facial features, which is produced by a Squeeze Layer implemented utilizing a Fully Connected layer (FC) and a ReLU activation function: + +$$ +f _ {z} = \operatorname {R e L U} \left(F C \left(f _ {i} + f _ {p}\right)\right), \tag {7} +$$ + +$f_{z}$ are then projected to two adaptive weights $\mathbf{a} \in \mathbb{R}^{1 \times C}$ and $\mathbf{b} \in \mathbb{R}^{1 \times C}$ for $f_{i}$ and $f_{p}$ via two FC layers. A Softmax function is employed on $\mathbf{a}$ and $\mathbf{b}$ on their channel-wise digits to ensure $\mathbf{a} + \mathbf{b} = \mathbf{1}$ . Afterwards, the optimized visual embeddings $f_{i}'$ can be obtained by the channel-wise weighted addition with $\mathbf{a}$ and $\mathbf{b}$ : + +$$ +f _ {i} ^ {\prime} = f _ {i} \cdot \mathbf {a} + f _ {p} \cdot \mathbf {b}. \tag {8} +$$ + +Subsequently, the refined $f_{i}^{\prime}$ and textual embeddings $f_{t}$ are input into Similar Regression to calculate the final prediction $\widehat{Q}$ , which is consistent with the MAEM. + +# 4.3. Training objectives + +Loss for PAPM. The L1 loss between $\widehat{Q}_1$ and the ground truth label $Q$ is calculated as the attractiveness prediction loss function $\mathcal{L}_1$ , which can be formulated as: + +$$ +\mathcal {L} _ {1} = \frac {1}{n} \sum_ {i = 1} ^ {n} \left| \widehat {Q} _ {1 _ {i}} - Q _ {i} \right|, \tag {9} +$$ + +where $n$ denotes number of training samples. + +Loss for MAEM. To jointly optimize ranking and regression, we incorporate a merged ranking loss $\mathcal{L}_R$ along with the L1 scoring loss $\mathcal{L}_S$ . Specifically, $\mathcal{L}_R$ is composed of a fidelity loss $\mathcal{L}_{R_1}$ and a two-direction ranking loss $\mathcal{L}_{R_2}$ . For the fidelity loss $\mathcal{L}_{R_1}$ , given an image pair $(I_i, I_j)$ , the binary label according to their ground truth $(Q_i, Q_j)$ is defined as: + +$$ +R \left(I _ {i}, I _ {j}\right) = \left\{ \begin{array}{l l} 1 & \text {i f} Q _ {i} \geq Q _ {j} \\ 0 & \text {o t h e r w i s e .} \end{array} \right. \tag {10} +$$ + +![](images/7757488bf5e6d34838806709c2ac8158ef1145c54fb8814a76cea56b27eddeae.jpg) +Figure 7. MOS distribution of three benchmark datasets. + +According to the Thurstone's model [51], the probability of $I_{i}$ perceiving more attractive than $I_{j}$ is estimated as: + +$$ +\widehat {R} \left(I _ {i}, I _ {j}\right) = \Psi \left(\frac {\widehat {Q} _ {2 i} - \widehat {Q} _ {2 j}}{\sqrt {2}}\right), \tag {11} +$$ + +where $\Psi (\cdot)$ denotes the standard normal cumulative distribution function, and its variance is set to 1. Then the fidelity loss [54] can be calculated as: + +$$ +\begin{array}{l} \mathcal {L} _ {R _ {1}} = 1 - \sqrt {R \left(I _ {i} , I _ {j}\right) \widehat {R} \left(I _ {i} , I _ {j}\right)} \tag {12} \\ - \sqrt {(1 - R (I _ {i} , I _ {j})) (1 - \widehat {R} (I _ {i} , I _ {j}))}. \\ \end{array} +$$ + +For the two-direction ranking loss $\mathcal{L}_{R_2}$ , given an image $I_i$ , $\widehat{S}_{i,j}$ denotes the cosine similarity score of $j_{th}$ attractiveness level and $\widehat{S}_{i,o} = \max \left\{\widehat{S}_{i,j} \mid j \in \{1,2,3,4,5\}\right\}$ . Since the human perception of facial attractiveness is gradual and rarely exhibits multiple peaks [62], we further obtain the ranking information by ensuring that the attractiveness level probability distribution only exhibits one peak and decreases in two directions (leftward and rightward). Then the two-direction pairwise ranking loss can be formulated as: + +$$ +\left\{ \begin{array}{l} \mathcal {L} _ {\text {l e f t}} = - \sum_ {j = 2} ^ {o} \log \frac {\exp \left(\widehat {S} _ {i , j}\right)}{\exp \left(\widehat {S} _ {i , j}\right) + \exp \left(\widehat {S} _ {i , j - 1}\right)} \\ \mathcal {L} _ {\text {r i g h t}} = - \sum_ {j = o} ^ {A - 1} \log \frac {\exp \left(\widehat {S} _ {i , j}\right)}{\exp \left(\widehat {S} _ {i , j}\right) + \exp \left(\widehat {S} _ {i , j + 1}\right)} \\ \mathcal {L} _ {R _ {2}} = - \left(\mathcal {L} _ {\text {l e f t}} + \mathcal {L} _ {\text {r i g h t}}\right), \end{array} \right. \tag {13} +$$ + +where $A$ denotes the number of attractiveness levels, and the corresponding labels of $\widehat{S}_{i,j-1}$ , $\widehat{S}_{i,j}$ and $\widehat{S}_{i,j+1}$ are 0, 1, 0, respectively. In the end, the total loss of MAEM can be defined as $\mathcal{L}_2 = \mathcal{L}_S + \lambda_1\mathcal{L}_{R_1} + \lambda_2\mathcal{L}_{R_2}$ , where $\lambda_1$ and $\lambda_2$ are empirically set as 1. + +Loss for SFM. The loss used in the second stage between the ground truth label $Q$ and the final attractiveness score $\hat{Q}$ is the L1 Loss: + +$$ +\mathcal {L} _ {3} = \frac {1}{n} \sum_ {i = 1} ^ {n} \left| \widehat {Q} _ {i} - Q _ {i} \right|. \tag {14} +$$ + +Table 2. Performance comparison on LiveBeauty, MEBeauty [24] and SCUT-FBP5500 [27]. Best in red and second-best in blue. + +
AspectsMethodsLiveBeautyMEBeauty [24]SCUT-FBP5500 [27]
SROCC↑PLCC↑KROCC↑SROCC↑PLCC↑KROCC↑SROCC↑PLCC↑KROCC↑
BaselineViT-B [23]0.8970.8470.7330.6200.6590.4430.8690.8800.686
ResNeXt-50 [56]0.9070.8790.7460.7000.7430.5160.8990.9110.727
Inception-V3 [49]0.9060.8800.7450.6760.7270.4920.9050.9180.736
IAAAVA-MLSP [21]0.8330.8150.6500.6020.6430.4310.7640.7940.566
TANet [18]0.8910.8580.7280.6510.6900.4670.8700.8820.686
Dele-Trans [19]0.8880.8410.7210.6110.6510.4400.8700.8600.682
EAT [20]0.8930.8670.7280.6400.6970.4620.8530.8720.665
FAPComboNet [58]0.9020.8710.7410.6510.6920.4720.8970.9070.725
2D-FAP [32]0.8960.8540.7330.6900.7190.5060.9030.9150.734
REX-INCEP [2]0.9080.8840.7500.7020.7390.5140.9070.9170.739
CNN-ER [2]0.9130.8880.7570.7060.7480.5190.9100.9220.745
MEBeauty [24]0.8260.7960.6430.6490.6770.4670.7990.8020.599
Ours0.9240.8960.7730.7830.7950.5930.9300.9320.774
Improvement+1.1%+0.8%+1.6%+7.7%+4.7%+7.4%+2.0%+1.0%+2.9%
+ +# 5. Experiments + +# 5.1. Benchmark Datasets + +The proposed LiveBeauty dataset and two open-source FAP datasets including SCUT-FBP5500 [27] and MEBeauty [24] are selected as benchmark datasets. The MOS distributions of them are shown in Fig. 7. We split SCUT-FBP5500 and MEBeauty with $60\% - 40\%$ and $80\% - 20\%$ train-test ratio separately, which is kept same with their own protocol, and LiveBeauty is split using the $90\% - 10\%$ protocol. + +# 5.2. Implementation Details + +Model initialization. In MAEM, ViT-B/16 and GPT-2 are adopted as the image encoder and the text encoder separately, which are initialized by [42]. In PAPM, Swin-T is adopted as the trainable image encoder, where only four stages are reserved and initialized with parameters pretrained on multiple face-related tasks [41]. + +Hyper parameters. During the whole training process, AdamW is selected as the optimizer, and the learning rate scheduler is set with linear warm-up and cosine annealing scheme. The base learning rate varies with different training phase. Each training phase lasts 50 epochs, and the batch size is set as 32. Details can be found in the supplementary. + +# 5.3. Competitors & Metrics + +To confirm the effectiveness of the proposed FPEM method, we select baseline models and several SOTA FAP and IAA methods, including ComboNet [58], 2D-FAP [32], REX-INCEP [2], CNN-ER [2], MEBeauty [24], AVAMLSP [21], TANet [18], Dele-Trans [19] and EAT [20]. All the models are retrained ourselves with the default parameters defined in their papers. Three criteria are adopted to evaluate the performance of these methods, i.e., Spear + +![](images/28d2034fb63b1395875df6d64800e3e84f7cc87ef0b6ea83451044ba25a643b3.jpg) +(a)LiveBeauty. + +![](images/66cb23b930b5d467e635259eac8fc61419aac3ff5d028a63c71dc4fa16887585.jpg) +(b) MEBeauty. + +![](images/cd149623c33da7a322f9c693cb927d6e0111a2a3bbef640be7776cc98cd3a04c.jpg) +(c) SCUT-FBP5500. +Figure 8. Scatter plots of the FPEM predicted scores vs. MOSs on three benchmark datasets. + +man Rank Order Correlation Coefficient (SROCC), Pearson Linear Correlation Coefficient (PLCC) and Kendall Rank Order Correlation Coefficient (KROCC). + +# 5.4. Performance + +The experimental results on three FAP datasets are shown in Tab. 2, from which we could draw the following conclusions: (i) Our proposed method achieves the first place and surpasses the second place by about 0.011, 0.077, 0.020 in terms of SROCC values on LiveBeauty, MEBeauty and SCUT-FBP5500 respectively, which demonstrates its superiority. (ii) The IAA methods are inferior to the FAP methods, which manifests that the generic aesthetic assessment methods overlook the facial features involved in the subjective nature of facial attractiveness, leading to limited performance on FAP tasks. (iii) The performance of all methods drops on MEBeauty because the number of training samples is restricted and the faces are ethnically diverse in MEBeauty, indicating a large diversity in facial attractiveness. (iv) The significant improvement on MEBeauty over other datasets can be largely attributed to the enhancement brought by face priors, which is more effective in scenarios with insufficient training data. + +Fig. 8 presents scatter plots illustrating the relationship between the predicted attractiveness scores (vertical axis) + +Table 3. Ablation study on LiveBeauty test set. $E_{FP}$ denotes Cross-Attention with $f_p$ , MS denotes Multi-Scale fusion operation, SR denotes Similarity Regression strategy, $\mathcal{L}_S$ denotes the L1 scoring loss, $\mathcal{L}_R$ denotes the total ranking loss $\mathcal{L}_R = \mathcal{L}_{R_1} + \mathcal{L}_{R_2}$ . + +
ModulesMethodsMetrics
SROCC↑PLCC↑KROCC↑
PAPMonlyEMS0.90450.87100.7454
w/oEFp0.90780.87140.7475
w/o MS0.90890.88080.7487
PAPM0.90960.88100.7497
MAEMzero-shot CLIP0.38380.35820.2692
w/o SR0.88250.82260.7123
w/oLS0.91610.84330.7607
w/oLR0.91860.89110.7653
MAEM0.91980.86310.7653
ALLw/o 2-stage0.91890.88890.7652
w/o SFM0.91970.89610.7641
all combined0.92430.89620.7732
+ +obtained by FPEM and the MOSs (horizontal axis) in the test sets of three datasets. The MOSs of MEBeauty are divided by two to align the MOS range with others. The color of the scatter points from blue to red indicates the density of points from low to high. In each sub-figure, a second-order polynomial nonlinear fitting curve is superimposed. It can be observed that the fitted curves are close to the diagonal line, demonstrating the effectiveness of FPEM. + +# 5.5. Ablation Study + +We conduct elaborate ablation study to investigate the contributions of different modules in our proposed method, which is shown in Tab. 3. Conclusions can be drawn that: (i) In the ablation study of PAPM, we can observe that leveraging facial prior and multi-scale features are both helpful. (ii) The result of ablation study of MAEM indicates that the zero-shot CLIP has limited generalization ability in FAP tasks, which confirms its shortcoming in aesthetic evaluation [45]. Specifically, for MAEM w/o SR, we leverage concatenation of the visual features and the textual features with highest similarity, and utilize MLP for regression. We observe a significant performance drop without SR, which manifests that specifically devised regression strategy is important for CLIP to adapt to regression tasks. Meanwhile, the joint optimization of the ranking and regression also leads to the improvement of SROCC. (iii) In the ablation study of ALL modules, we can see that employing all modules yields the best performance, which confirms the effectiveness of using SFM to adaptively integrate multi-modal aesthetic features with holistic facial prior knowledge. + +# 5.6. Cross-Dataset Study + +Considering the diversity of facial attractiveness perception, the cross-dataset validation is carried out to validate the generalization ability of these methods. Specifically, + +Table 4. Cross-Dataset study trained on LiveBeauty and tested on the entire MEBeauty dataset and the entire SCUT-FBP5500 dataset. Best in red and second-best in blue. + +
MethodsMEBeauty [24]SCUT-FBP5500 [27]
SROCC↑PLCC↑KROCC↑SROCC↑PLCC↑KROCC↑
ViT-B [23]0.5730.5830.4020.5530.5480.394
ResNeXt-50 [56]0.5350.5460.3720.4780.4670.335
Inception-V3 [49]0.5750.5810.4040.4840.4800.339
AVA-MLSP [21]0.5870.5970.4140.4000.4120.275
TANet [18]0.5510.5290.3840.4010.4110.277
Dele-Trans[19]0.5760.5720.4020.4800.4760.335
EAT[20]0.6030.6240.4270.4840.4660.338
ComboNet [58]0.5400.5480.3750.3840.3750.264
2D-FAP [32]0.5560.4890.3900.4440.4450.306
REX-INCEP [2]0.5880.5920.4130.4730.4710.332
CNN-ER [2]0.5890.5960.4140.5020.4990.352
MEBeauty [24]0.5200.5160.3610.3660.3390.251
FPEM (Ours)0.6300.6290.4470.5990.5770.434
Improvement+2.7%+0.5%+2.0%+4.6%+2.9%+4.0%
+ +Table 5. Cross-Dataset Ablation Study on each module. All models are trained on LiveBeauty and tested on the whole MEBeauty. + +
PAPMMAEMSFMSROCC↑PLCC↑KROCC↑
✓0.61340.61410.4349
✓0.62290.62680.4244
✓✓✓0.62990.62910.4472
+ +all the mentioned models are trained on the proposed LiveBeauty dataset and subsequently tested on the entire SCUTFBP5500 and the entire MEBeauty. + +As shown in Tab. 4, the proposed method outperforms all other methods on both two datasets, which demonstrates its strong generalization ability. Moreover, ablation results of the generalizability on MEBeauty are presented in Tab. 5. It can be observed that MAEM surpasses PAPM by leveraging multi-modal information. The generalizability is further improved by fusing their features through SFM, which indicates that integrating facial prior containing rich contextual and individual information is effective not only for improving performance but also for enhancing generalizability. + +# 6. Conclusion + +In this paper, we tackle the problem of facial attractiveness prediction in live videos that undergo face retouching. To address the existing gap in datasets, we introduce LiveBeauty, the first large-scale FAP dataset featuring beautiful face images for aesthetic purposes. This dataset is high-quality and authentic, collected from real-world environments, and is currently the largest open-access facial attractiveness dataset. We further propose FPEM, a multimodal FAP framework that selectively fuses holistic facial priors and aesthetic semantic features. Extensive experiments demonstrate the advantages of FPEM over state-of-the-art approaches. Meanwhile, LiveBeauty and FPEM can be effectively leveraged for FAP in other video applications. + +# References + +[1] Lacey Best-Rowden and Anil K Jain. Learning face image quality from human assessments. IEEE Transactions on Information Forensics and Security, 13(12):3064-3077, 2018. 1 +[2] Fares Bougourzi, Fadi Dornaika, and Abdelmalik Taleb-Ahmed. Deep learning based face beauty prediction via dynamic robust losses and ensemble regression. Knowledge-Based Systems, 242:108246, 2022. 1, 2, 3, 7, 8 +[3] Djamel Eddine Boukhari, Ali Chemsa, and Riadh Ajgou. Facial beauty prediction based on vision transformer. International Journal of Electrical and Electronic Engineering & Telecommunications, pages 2319-2518, 2023. 3 +[4] Fadi Boutros, Meiling Fang, Marcel Klemt, Biying Fu, and Naser Damer. Cr-fiaq: face image quality assessment by learning sample relative classifiability. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5836-5845, 2023. 3 +[5] Kerang Cao, Kwang-nam Choi, Hoekyung Jung, and Lini Duan. Deep learning for facial beauty prediction. Information, 11(8):391, 2020. 3 +[6] Nicolas Chahine, Stefania Calarasanu, Davide Garcia-Civiero, Theo Cayla, Sira Ferradans, and Jean Ponce. An image quality assessment dataset for portraits. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9968-9978, 2023. 1 +[7] Wei-Ting Chen, Gurunandan Krishnan, Qiang Gao, SyYen Kuo, Sizhou Ma, and Jian Wang. Dsl-fqa: Assessing facial image quality via dual-set degradation learning and landmark-guided transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2931-2941, 2024. 3 +[8] Ritendra Datta, Dhiraj Joshi, Jia Li, and James Z Wang. Studying aesthetics in photographic images using a computational approach. In Proceedings of the European Conference on Computer Vision, pages 288-301, 2006. 2 +[9] Harm De Vries and Jason Yosinski. Can deep learning help you find the perfect match? arXiv preprint arXiv:1505.00359, 2015. 1, 2 +[10] Yubin Deng, Chen Change Loy, and Xiaou Tang. Image aesthetic assessment: An experimental survey. IEEE Signal Processing Magazine, 34(4):80-106, 2017. 2 +[11] Abhishek Dutta, Raymond Veldhuis, and Luuk Spreeuwers. A bayesian model for predicting face recognition performance using image quality. In Proceeding of the International Joint Conference on Biometrics, pages 1-8, 2014. 3 +[12] Yael Eisenthal, Gideon Dror, and Eytan Ruppin. Facial attractiveness: Beauty and the machine. Neural Computation, 18(1):119-142, 2006. 1, 3 +[13] Yang-Yu Fan, Shu Liu, Bo Li, Zhe Guo, Ashok Samal, Jun Wan, and Stan Z Li. Label distribution-based facial attractiveness computation by deep residual learning. IEEE Transactions on Multimedia, 20(8):2196-2208, 2017. 3 +[14] Junying Gan, Xiaoshan Xie, Guohui He, and Heng Luo. Transbls: transformer combined with broad learning system + +for facial beauty prediction. Applied Intelligence, 53(21): 26110-26125, 2023. 3 +[15] Junying Gan, Xiaoshan Xie, Yikui Zhai, Guohui He, Chaoyun Mai, and Heng Luo. Facial beauty prediction fusing transfer learning and broad learning system. Soft Computing, 27(18):13391-13404, 2023. 3 +[16] Lian Gao, Weixin Li, Zehua Huang, Di Huang, and Yunhong Wang. Automatic facial attractiveness prediction by deep multi-task learning. In Proceedings of the IEEE International Conference on Pattern Recognition, pages 3592-3597, 2018. 3 +[17] Douglas Gray, Kai Yu, Wei Xu, and Yihong Gong. Predicting facial beauty without landmarks. In Proceedings of the European Conference on Computer Vision, pages 434-447, 2010. 1, 2 +[18] Shuai He, Yongchang Zhang, Rui Xie, Dongxiang Jiang, and Anlong Ming. Rethinking image aesthetics assessment: Models, datasets and benchmarks. In Proceedings of the International Joint Conference on Artificial Intelligence, pages 942–948, 2022. 2, 7, 8 +[19] Shuai He, Anlong Ming, Yaqi Li, Jinyuan Sun, ShunTian Zheng, and Huadong Ma. Thinking image color aesthetics assessment: Models, datasets and benchmarks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21838-21847, 2023. 7, 8 +[20] Shuai He, Anlong Ming, Shuntian Zheng, Haobin Zhong, and Huadong Ma. Eat: An enhancer for aesthetics-oriented transformers. In Proceedings of the ACM International Conference on Multimedia, pages 1023-1032, 2023. 7, 8 +[21] Vlad Hosu, Bastian Goldlucke, and Dietmar Saupe. Effective aesthetics prediction with multi-level spatially pooled features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9375-9383, 2019. 2, 7, 8 +[22] Tharun J. Iyer, Rahul K, Ruban Nersisson, Zhemin Zhuang, Alex Noel Joseph Raj, and Imthiaz Refayee. Machine learning-based facial beauty prediction and analysis of frontal facial images using facial landmarks and traditional image descriptors. Computational Intelligence and Neuroscience, 2021(1):4423407, 2021. 2, 3 +[23] Alexander Kolesnikov, Alexey Dosovitskiy, Dirk Weissenborn, Georg Heigold, Jakob Uszkoreit, Lucas Beyer, Matthias Minderer, Mostafa Dehghani, Neil Houlsby, Sylvain Gelly, Thomas Unterthiner, and Xiaohua Zhai. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2021. 7, 8 +[24] Irina Lebedeva, Yi Guo, and Fangli Ying. Mebeauty: a multi-ethnic facial beauty dataset in-the-wild. *Neural Computing and Applications*, pages 1–15, 2022. 1, 2, 7, 8 +[25] Congcong Li, Andrew Gallagher, Alexander C. Loui, and Tsuhan Chen. Aesthetic quality assessment of consumer photos with faces. In Proceedings of the IEEE International Conference on Image Processing, pages 3221-3224, 2010. 2 +[26] Xiang Li, Wenhai Wang, Xiaolin Hu, and Jian Yang. Selective kernel networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 510-519, 2019. 6 + +[27] Lingyu Liang, Luojun Lin, Lianwen Jin, Duorui Xie, and Mengru Li. Scut-fbp5500: A diverse benchmark dataset for multi-paradigm facial beauty prediction. In Proceedings of the IEEE International Conference on Pattern Recognition, pages 1598-1603, 2018. 1, 2, 3, 7, 8 +[28] Luojun Lin, Lingyu Liang, Lianwen Jin, and Weijie Chen. Attribute-aware convolutional neural networks for facial beauty prediction. In Proceedings of the Information Joint Conference on Artificial Intelligence, pages 847-853, 2019. 1, 3 +[29] Qipeng Liu, Luojun Lin, Zhifeng Shen, and Yuanlong Yu. Fbpformer: Dynamic convolutional transformer for global-local-contextual facial beauty prediction. In Proceedings of the International Conference on Artificial Neural Networks, pages 223-235, 2023. 2, 3 +[30] Shu Liu, Yang-Yu Fan, Zhe Guo, Ashok Samal, and Afan Ali. A landmark-based data-driven approach on 2.5 d facial attractiveness computation. Neurocomputing, 238:168-178, 2017. 2, 3 +[31] Shu Liu, Bo Li, Yang-Yu Fan, Zhe Quo, and Ashok Samal. Facial attractiveness computation by label distribution learning with deep cnn and geometric features. In Proceedings of the IEEE International Conference on Multimedia and Expo, pages 1344-1349, 2017. 3 +[32] Shu Liu, Enquan Huang, Yan Xu, Kexuan Wang, Xiaoyan Kui, Tao Lei, and Hongying Meng. Lightweight facial attractiveness prediction using dual label distribution. arXiv preprint arXiv:2212.01742, 2022. 1, 2, 7, 8 +[33] Shu Liu, Enquan Huang, Yan Xu, Kexuan Wang, Xiaoyan Kui, Tao Lei, and Hongying Meng. Lightweight facial attractiveness prediction using dual label distribution. arXiv preprint arXiv:2212.01742, 2022. 3 +[34] Tie Liu, Shengxi Li, Mai Xu, Li Yang, and Xiaofei Wang. Assessing face image quality: A large-scale database and a transformer method. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3 +[35] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2 +[36] Naila Murray, Luca Marchesotti, and Florent Perronnin. Ava: A large-scale database for aesthetic visual analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2408-2415, 2012. 2 +[37] Fu-Zhao Ou, Xingyu Chen, Ruixin Zhang, Yuge Huang, Shaoxin Li, Jilin Li, Yong Li, Liujuan Cao, and Yuan-Gen Wang. Sdd-fqa: Unsupervised face image quality assessment with similarity distribution distance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7670–7679, 2021. 1 +[38] Fu-Zhao Ou, Chongyi Li, Shiqi Wang, and Sam Kwong. Clib-fiqa: Face image quality assessment with confidence calibration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1694-1704, 2024. 3 +[39] Tianhao Peng, Mu Li, Fangmei Chen, Yong Xu, and David Zhang. Geometric prior guided hybrid deep neural network + +for facial beauty analysis. CAAI Transactions on Intelligence Technology, 9(2):467-480, 2024. 2 +[40] P Jonathon Phillips, J Ross Beveridge, David S Bolme, Bruce A Draper, Geof H Givens, Yui Man Lui, Su Cheng, Mohammad Nayeem Teli, and Hao Zhang. On the existence of face quality measures. In Proceedings of the IEEE International Conference on Biometrics: Theory, Applications and Systems, pages 1-8, 2013. 1 +[41] Lixiong Qin, Mei Wang, Chao Deng, Ke Wang, Xi Chen, Jiani Hu, and Weihong Deng. Swinface: a multi-task transformer for face recognition, expression recognition, age estimation and attribute estimation. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 7 +[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proceedings of the International Conference on Machine Learning, pages 8748-8763, 2021. 2, 4, 7 +[43] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 815-823, 2015. 2 +[44] B Series. Methodology for the subjective assessment of the quality of television pictures. Recommendation ITU-R BT, 500(13), 2012. 3, 4 +[45] Xiangfei Sheng, Leida Li, Pengfei Chen, Jinjian Wu, Weisheng Dong, Yuzhe Yang, Liwu Xu, Yaqian Li, and Guangming Shi. Aesclip: Multi-attribute contrastive learning for image aesthetics assessment. In Proceedings of the ACM International Conference on Multimedia, page 1117–1126, 2023. 8 +[46] Shengjie Shi, Fei Gao, Xuantong Meng, Xingxin Xu, and Jingjie Zhu. Improving facial attractiveness prediction via co-attention learning. In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing, pages 4045–4049, 2019. 3 +[47] Shaolin Su, Hanhe Lin, Vlad Hosu, Oliver Wiedemann, Jinqiu Sun, Yu Zhu, Hantao Liu, Yanning Zhang, and Dietmar Saupe. Going the extra mile in face image quality assessment: A novel database and model. IEEE Transactions on Multimedia, 2023. 1 +[48] Wei Sun, Weixia Zhang, Yanwei Jiang, Haoning Wu, Zicheng Zhang, Jun Jia, Yingjie Zhou, Zhongpeng Ji, Xiongkuo Min, Weisi Lin, and Zhai Guangtao. Dual-branch network for portrait image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2024. 3 +[49] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2826, 2016. 1, 7, 8 +[50] Philipp Terhorst, Jan Niklas Kolf, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. Ser-fiq: Unsupervised estimation of face image quality based on stochastic embedding + +robustness. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5651-5660, 2020. 3 +[51] Louis L Thurstone. A law of comparative judgment. In *Scaling*, pages 81–92. Routledge, 2017. 6 +[52] Yu Tian, Zhangkai Ni, Baoliang Chen, Shiqi Wang, Hanli Wang, and Sam Kwong. Generalized visual quality assessment of gan-generated face images. arXiv preprint arXiv:2201.11975, 2022. 1 +[53] Song Tong, Xuefeng Liang, Takatsune Kumada, and Sunao Iwaki. Putative ratios of facial attractiveness in a deep neural network. Vision Research, 178:86-99, 2021. 1, 2 +[54] Ming-Feng Tsai, Tie-Yan Liu, Tao Qin, Hsin-Hsi Chen, and Wei-Ying Ma. Frank: a ranking method with fidelity loss. In Proceedings of the ACM SIGIR Conference on Research and Development in Information Retrieval, pages 383-390, 2007. 6 +[55] Qinjie Xiao, You Wu, Dinghong Wang, Yong-Liang Yang, and Xiaogang Jin. Beauty3dfacenet: deep geometry and texture fusion for 3d facial attractiveness prediction. Computers & Graphics, 98:11-18, 2021. 3 +[56] Saining Xie, Ross Girshick, Piotr Dolkar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1492-1500, 2017. 1, 7, 8 +[57] Jie Xu, Lianwen Jin, Lingyu Liang, Ziyong Feng, Duorui Xie, and Huiyun Mao. Facial attractiveness prediction using psychologically inspired convolutional neural network (pCNN). In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing, pages 1657-1661, 2017. 3 +[58] Lu Xu and Jinhai Xiang. Comboloss for facial attractiveness analysis with squeeze-and-excitation networks. arXiv preprint arXiv:2010.10721, 2020. 2, 7, 8 +[59] Lu Xu and Jinhai Xiang. Comboloss for facial attractiveness analysis with squeeze-and-excitation networks. arXiv preprint arXiv:2010.10721, 2020. 3 +[60] Lu Xu, Heng Fan, and Jinhai Xiang. Hierarchical multi-task network for race, gender and facial attractiveness recognition. In Proceedings of the IEEE International Conference on Image Processing, pages 3861-3865, 2019. 3 +[61] Haibin Yan. Cost-sensitive ordinal regression for fully automatic facial beauty assessment. Neurocomputing, 129:334-342, 2014. 1, 2 +[62] Qinkai Yu, Jianyang Xie, Anh Nguyen, He Zhao, Jiong Zhang, Huazhu Fu, Yitian Zhao, Yalin Zheng, and Yanda Meng. Clip-dr: Textual knowledge-guided diabetic retinopathy grading with ranking-aware prompting. arXiv preprint arXiv:2407.04068, 2024. 6 +[63] Yikui Zhai, Yu Huang, Ying Xu, Junying Zeng, Fei Yu, and Junying Gan. Benchmark of a large scale database for facial beauty prediction. In Proceedings of the International Conference on Intelligent Information Processing, pages 1-5, 2016. 1, 2 +[64] Yikui Zhai, Yu Huang, Ying Xu, Junying Gan, He Cao, Wenbo Deng, Ruggero Donida Labati, Vincenzo Piuri, and + +Fabio Scotti. Asian female facial beauty prediction using deep neural networks via transfer learning and multi-channel feature fusion. IEEE Access, 8:56892-56907, 2020. 1, 2 +[65] Shifeng Zhang, Xiangyu Zhu, Zhen Lei, Hailin Shi, Xiaobo Wang, and Stan Z Li. Faceboxes: A cpu real-time face detector with high accuracy. In Proceedings of the IEEE International Joint Conference on Biometrics, pages 1-9, 2017. 3 +[66] Weixia Zhang, Guangtao Zhai, Ying Wei, Xiaokang Yang, and Kede Ma. Blind image quality assessment via vision-language correspondence: A multitask learning perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14071-14081, 2023. 5 \ No newline at end of file diff --git a/2025/FPEM_ Face Prior Enhanced Facial Attractiveness Prediction for Live Videos with Face Retouching/images.zip b/2025/FPEM_ Face Prior Enhanced Facial Attractiveness Prediction for Live Videos with Face Retouching/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..73d499f8e6f2f16bbd0380bec58050350dac79e5 --- /dev/null +++ b/2025/FPEM_ Face Prior Enhanced Facial Attractiveness Prediction for Live Videos with Face Retouching/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f8026a9f64da54ea43e99ce9c1295d807a07c27ebfeaec1d843d0d7f98ed18c4 +size 655310 diff --git a/2025/FPEM_ Face Prior Enhanced Facial Attractiveness Prediction for Live Videos with Face Retouching/layout.json b/2025/FPEM_ Face Prior Enhanced Facial Attractiveness Prediction for Live Videos with Face Retouching/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b08de3984d2db4d8b1404dc47004136905cd09b0 --- /dev/null +++ b/2025/FPEM_ Face Prior Enhanced Facial Attractiveness Prediction for Live Videos with Face Retouching/layout.json @@ -0,0 +1,10526 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 115, + 103, + 496, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 103, + 496, + 140 + ], + "spans": [ + { + "bbox": [ + 115, + 103, + 496, + 140 + ], + "type": "text", + "content": "FPEM: Face Prior Enhanced Facial Attractiveness Prediction for Live Videos with Face Retouching" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 142, + 160, + 465, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 160, + 465, + 190 + ], + "spans": [ + { + "bbox": [ + 142, + 160, + 465, + 190 + ], + "type": "text", + "content": "Hui Li" + }, + { + "bbox": [ + 142, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 142, + 160, + 465, + 190 + ], + "type": "text", + "content": ", Xiaoyu Ren" + }, + { + "bbox": [ + 142, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + }, + { + "bbox": [ + 142, + 160, + 465, + 190 + ], + "type": "text", + "content": ", Hongjiu Yu" + }, + { + "bbox": [ + 142, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 142, + 160, + 465, + 190 + ], + "type": "text", + "content": ", Ying Chen" + }, + { + "bbox": [ + 142, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 142, + 160, + 465, + 190 + ], + "type": "text", + "content": ", Kai Li" + }, + { + "bbox": [ + 142, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 142, + 160, + 465, + 190 + ], + "type": "text", + "content": ", L Wang" + }, + { + "bbox": [ + 142, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 142, + 160, + 465, + 190 + ], + "type": "text", + "content": ", Xiongkuo Min" + }, + { + "bbox": [ + 142, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 142, + 160, + 465, + 190 + ], + "type": "text", + "content": ", Huiyu Duan" + }, + { + "bbox": [ + 142, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 142, + 160, + 465, + 190 + ], + "type": "text", + "content": ", Guangtao Zhai" + }, + { + "bbox": [ + 142, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 142, + 160, + 465, + 190 + ], + "type": "text", + "content": ", Xu Liu" + }, + { + "bbox": [ + 142, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 216, + 190, + 394, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 190, + 394, + 203 + ], + "spans": [ + { + "bbox": [ + 216, + 190, + 394, + 203 + ], + "type": "text", + "content": "1 Alibaba Group 2Shanghai Jiao Tong University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 64, + 205, + 545, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 205, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 64, + 205, + 545, + 217 + ], + "type": "text", + "content": "{lh401704,naomi.rxy,yuhongjiu.yhj,YingChen,kaishi.lk,zhuangshu.wlb,steven.liu}@taobao.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 174, + 220, + 433, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 220, + 433, + 232 + ], + "spans": [ + { + "bbox": [ + 174, + 220, + 433, + 232 + ], + "type": "text", + "content": "{minxiongkuo,huiyuduan,zhaiguangtao}@sjtu.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 283, + 297, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 297, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 297, + 534 + ], + "type": "text", + "content": "Facial attractiveness prediction (FAP) has long been an important computer vision task, which could be widely applied in live videos with facial retouching. However, previous FAP datasets are either small or closed-source. Moreover, the corresponding FAP models exhibit limited generalization and adaptation ability. To overcome these limitations, we introduce the first large-scale FAP dataset LiveBeauty specifically designed for live video scenarios wherein face images may be real-time processed for aesthetics purposes. 10,000 face images are collected directly from a live streaming platform, with 200,000 corresponding attractiveness annotations obtained from a well-devised subjective experiment, making LiveBeauty the largest open-access FAP dataset. Based on the built dataset, a novel FAP method named Facial Prior Enhanced Multi-modal model (FPEM) is proposed to measure the attractiveness of facial images. Extensive experiments conducted on both LiveBeauty and other open-source FAP datasets demonstrate that our proposed method achieves state-of-the-art performance. The dataset will be released at https://github.com/EstellaLH/FPEM." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 558, + 135, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 558, + 135, + 571 + ], + "spans": [ + { + "bbox": [ + 56, + 558, + 135, + 571 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 578, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 296, + 685 + ], + "type": "text", + "content": "With the rapid development of live streaming, short-form video, and advanced face retouching within such applications, the perceptual quality of faces becomes an important factor for the overall visual experience of users, which is particularly essential for video platforms, as it can significantly affect the average viewer watch time and even GMV (Gross Merchandise Volume). A robust metric system for evaluating facial clarity and aesthetics is essential to monitor both the overall service quality of the platform and" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 264, + 432, + 368 + ], + "blocks": [ + { + "bbox": [ + 316, + 264, + 432, + 368 + ], + "lines": [ + { + "bbox": [ + 316, + 264, + 432, + 368 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 432, + 368 + ], + "type": "image", + "image_path": "c0f4409e3000fe150b3c0f06d40319855c6edc3c7b11e9f8b00c7d4d939d8af5.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 380, + 555, + 425 + ], + "lines": [ + { + "bbox": [ + 313, + 380, + 555, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 555, + 425 + ], + "type": "text", + "content": "Figure 1. SROCC vs PLCC comparison on LiveBeauty and SCUT-FBP5500 [27]. FPEM, denoted by red triangular, outperforms other methods (CNN-ER [2], 2D-FAP [32], REX-INCEP [2], Inception-V3 [49], ResNeXt-50 [56]) on both datasets." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 436, + 264, + 552, + 368 + ], + "blocks": [ + { + "bbox": [ + 436, + 264, + 552, + 368 + ], + "lines": [ + { + "bbox": [ + 436, + 264, + 552, + 368 + ], + "spans": [ + { + "bbox": [ + 436, + 264, + 552, + 368 + ], + "type": "image", + "image_path": "9ef842e72aa080ba2473e92ee4f8bded9293a0e5964df33b70d5ea313b2a5ef7.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 449, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 449, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 312, + 449, + 555, + 616 + ], + "type": "text", + "content": "the visual appeal changes or improvements of individual streaming channels. The clarity-related assessment tasks of facial quality could be categorized into Biometric Face Image Quality Assessment (BFIQA) and Generic Face Image Quality Assessment (GFIQA). BFIQA focuses on the identity information of faces for down-stream tasks such as face recognition [1, 37, 40], while GFIQA emphasizes the visual perceptual quality of faces [6, 47, 52]. The aesthetics-related assessment of facial quality is often referred to as Facial Attractiveness Prediction (FAP) [12, 27, 28]. The goal of FAP is to quantify the attractiveness of facial images, which is vital for assessing the effectiveness of facial retouching settings and offers valuable insights for algorithm development." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 617, + 556, + 715 + ], + "type": "text", + "content": "As shown in Tab. 1, numerous FAP datasets have been proposed in recent years, many of which encompass a rich diversity in age, gender, face expression and pose [9, 17, 24, 27, 53, 61, 63, 64]. However, most of these datasets are either small or unavailable due to privacy and ethical concerns. In video streaming, especially live video, it is common practice to use facial retouching algorithms to improve the facial aesthetics of content presented to audiences." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 144, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 144, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 144, + 703 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 67, + 703, + 137, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 137, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 137, + 712 + ], + "type": "text", + "content": "Equal contributions." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11458" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 91, + 550, + 208 + ], + "blocks": [ + { + "bbox": [ + 176, + 71, + 433, + 82 + ], + "lines": [ + { + "bbox": [ + 176, + 71, + 433, + 82 + ], + "spans": [ + { + "bbox": [ + 176, + 71, + 433, + 82 + ], + "type": "text", + "content": "Table 1. Overview of common facial attractiveness prediction datasets." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 91, + 550, + 208 + ], + "lines": [ + { + "bbox": [ + 61, + 91, + 550, + 208 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 550, + 208 + ], + "type": "table", + "html": "
DatasetYearRating Scale# ImageGenderEthnicity# AnnotatorPublic
Gray et al. [17]2010[-3, 3]2,056F MCaucasian30✓
MBW+FBW [61]2014[1, 5]10,400F MMulti30-
De Vries et al. [9]2015Binary9364F MAsian20-
LSFBD [63]2016[1, 5]20,000F M---
SCUT-FBP5500 [27]2018[1, 5]5,500F MMulti60✓
Tong et al. [53]2020Binary4,512F M-20-
LSAFBD [64]2020[1, 5]20,000FAsian20-
MEBeauty [24]2022[1, 10]2,550F MMulti300✓
LiveBeauty (proposed)2024[1, 5]10,000F MAsian20✓
", + "image_path": "ca3069fa4b8f337408f5304b04c1b63d80beffaee6d275c4d0c6e61fc47a7f17.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 228, + 294, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 228, + 294, + 357 + ], + "spans": [ + { + "bbox": [ + 54, + 228, + 294, + 357 + ], + "type": "text", + "content": "The differences among these algorithms can lead to unique data distributions that vary considerably from those found in unprocessed or well-edited face images obtained from the Internet. Meanwhile, the effectiveness of face retouching is largely dependent on the computational resources available, which in turn affects these data characteristics. Since the training data impacts the generalization ability of data-driven deep learning methods, the lack of a FAP dataset specifically collected from live videos with authentic face retouching process may lead to limited performance in practical video scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 361, + 295, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 295, + 540 + ], + "type": "text", + "content": "Meanwhile, there is no widely recognized FAP model specifically designed for live videos. Early methods relying on hand-crafted beauty descriptors [22, 30] tend to exhibit limited performance due to the inadequate feature representation capability. The success of the deep-learning based methods, e.g. Convolution Neural Networks (CNNs) [2, 32, 58] and Transformers [29, 39], have driven progress in the FAP task. Specifically, [39] first introduced a Vision Transformer into the FAP task and proposed a geometric prior guided network leveraging facial landmarks. Although their approach demonstrates promising performance, the straightforward feature concatenation and simplistic prior struggle to adaptively model individual facial features, thereby limiting its applicability and effectiveness in real-world applications." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 543, + 294, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 543, + 294, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 543, + 294, + 615 + ], + "type": "text", + "content": "To address the lack of FAP datasets considering the impact of face retouching, we first construct a large-scale facial attractiveness dataset LiveBeauty, which includes 10,000 face images collected from a popular video streaming platform, with the corresponding facial attractiveness annotations obtained from 20 annotators." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 617, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 294, + 713 + ], + "type": "text", + "content": "Furthermore, based on the established dataset, we propose a no-reference FAP model termed Facial Prior Enhanced Multi-modal model (FPEM), to predict the facial attractiveness by extracting and integrating holistic facial prior knowledge and multi-modal aesthetic semantic features. Specifically, we first devise a Personalized Attractiveness Prior Module (PAPM) to integrate extensive facial prior knowledge, in which multi-scale visual features" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 228, + 553, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 553, + 370 + ], + "type": "text", + "content": "extracted by the Swin Transformer [35] and facial prior features extracted by the pretrained FaceNet [43] are integrated via a Cross-Attention block. To obtain facial beauty attributes, we also introduce a Multi-modal Attractiveness Encoder Module (MAEM) based on CLIP [42] to extract multi-modal aesthetic semantic features. Furthermore, these extracted features are adaptively integrated through a fusion module. Extensive experimental results confirm that our proposed method achieves the best performance compared to other state-of-the-art (SOTA) methods on both our LiveBeauty dataset and other FAP datasets, as shown in Fig. 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 324, + 372, + 515, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 372, + 515, + 383 + ], + "spans": [ + { + "bbox": [ + 324, + 372, + 515, + 383 + ], + "type": "text", + "content": "To summarize, our contributions are as follow:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 385, + 553, + 552 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 314, + 385, + 553, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 385, + 553, + 456 + ], + "spans": [ + { + "bbox": [ + 314, + 385, + 553, + 456 + ], + "type": "text", + "content": "- We construct the LiveBeauty dataset, the first large-scale, open-source facial attractiveness prediction dataset sourced from a real-world live video platform featuring faces enhanced for aesthetics, making it suitable for applications wherein facial aesthetics are vital and therefore need to be well modeled and monitored." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 457, + 553, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 457, + 553, + 504 + ], + "spans": [ + { + "bbox": [ + 314, + 457, + 553, + 504 + ], + "type": "text", + "content": "- We propose a multi-modal method specifically designed for FAP, predicting attractiveness scores for face images. By integrating visual and textual modalities, this method can provide a more reliable prediction of facial appeal." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 505, + 553, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 505, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 314, + 505, + 553, + 552 + ], + "type": "text", + "content": "- We enhance the representation capability of our model by leveraging facial prior knowledge, and further employing selective fusion module to integrate and refine the extracted visual features." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 567, + 405, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 567, + 405, + 578 + ], + "spans": [ + { + "bbox": [ + 314, + 567, + 405, + 578 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 586, + 524, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 524, + 600 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 524, + 600 + ], + "type": "text", + "content": "2.1. Aesthetic Assessment of Generic Images" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "text", + "content": "Image Aesthetic Assessment (IAA) task aims to assess image quality from the aesthetic perspective. The corresponding IAA methods have evolved from employing handcrafted features, the modeling of photographic rules [8], global image layout [36] or typical objects [25] in perceived images [10] to data-driven approaches based on deep learning techniques that deliver superior performance [18-21]. Notably, these IAA methods primarily concentrate on generic images rather than specifically targeting facial regions." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11459" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 242, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 242, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 242, + 85 + ], + "type": "text", + "content": "2.2. Quality Assessment of Face Images" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 282 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 282 + ], + "type": "text", + "content": "Facial clarity related assessment include both BFIQA and GFIQA. BFIQA was originated from biometric research focusing on evaluating the suitability of facial images for recognition, initially relying on hand-crafted features [11, 50], and recent advancing to deep learning approaches with regression and embedding-based solutions [4, 38]. Nonetheless, Chen et al. [7] noted these methods may not generalize well on other facial quality and aesthetic tasks due to the emphasis on facial biometric information. In contrast, GFIQA is defined to evaluate the conventional perceptual quality of face images. GFIQA methods have advanced through the use of transformer-based [7, 34] and multi-modal approaches [48]. Clearly, GFIQA focuses on modeling the degradation of facial regions, such as issues related to blur and noise, and is therefore orthogonal to facial attractiveness." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 297, + 274, + 310 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 274, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 274, + 310 + ], + "type": "text", + "content": "2.3. Attractiveness Assessment of Face Images" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 316, + 296, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 296, + 424 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 296, + 424 + ], + "type": "text", + "content": "In the context of this paper, facial aesthetics primarily pertains to attractiveness. To measure the facial attractiveness of face images, early handcrafted Facial Attractiveness Prediction (FAP) methods predominantly rely on manually designed features, e.g. geometric features [12, 22, 30], traditional image descriptors [12, 22, 27], holistic descriptors [12], etc. However, these methods exhibit limited performance due to difficulties in presenting comprehensive characteristics of facial attractiveness." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 426, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 713 + ], + "type": "text", + "content": "With the advancement in deep learning, FAP solutions based on Convolution Neural Networks (CNNs) show better performance. Inspired by psychological studies, specific architectures of CNNs are designed [5, 57]. The applications of Label Distribution Learning (LDL) [13, 31, 33] achieve better performance. Some methods [16, 46, 55] combine geometric prior knowledge containing landmarks, parsing maps, etc., while the others [28, 60] integrate attractiveness-related attributes such as gender, race, etc. In addition, the ensemble solutions of different loss functions like regression loss and classification loss [59] and different models including ResNeXt50 and Inception-V3 [2] are further explored. However, these CNN-based methods primarily focus on local features, limiting their capacity for long-term modeling. To address this problem, transformer-based solutions [3, 14] are proposed, while Liu et al. [29] and Gan et al. [15] combine transformers with CNNs to leverage both global and local features related to facial attractiveness. Nevertheless, these approaches have limited generalization and adaptation ability in more challenging scenarios, due to the biased predictions from training data skewed towards certain demographics, and using straightforward single-modal modeling techniques which rely on simplistic concatenation of visual features." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 328, + 73, + 542, + 167 + ], + "blocks": [ + { + "bbox": [ + 328, + 73, + 542, + 167 + ], + "lines": [ + { + "bbox": [ + 328, + 73, + 542, + 167 + ], + "spans": [ + { + "bbox": [ + 328, + 73, + 542, + 167 + ], + "type": "image", + "image_path": "5b96182c2ccf8cfdbbe76555526f15d55a289d08dc13595c1d73437b8f57134d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 178, + 555, + 212 + ], + "lines": [ + { + "bbox": [ + 313, + 178, + 555, + 212 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 555, + 212 + ], + "type": "text", + "content": "Figure 2. Examples of the face images in our LiveBeauty dataset (from left to right are examples rated 1 to 5). The first row is male and second row is female." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 221, + 369, + 233 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 221, + 369, + 233 + ], + "spans": [ + { + "bbox": [ + 313, + 221, + 369, + 233 + ], + "type": "text", + "content": "3. Dataset" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 241, + 421, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 241, + 421, + 253 + ], + "spans": [ + { + "bbox": [ + 313, + 241, + 421, + 253 + ], + "type": "text", + "content": "3.1. Images Collection" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 258, + 555, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 555, + 389 + ], + "type": "text", + "content": "To accurately reflect the distribution of facial attractiveness of in-captured videos, we first collect large-scale raw face images from one of the most popular live streaming platforms. Specifically, we retrieve replays for live channels of all streamers conducted on the platform in March 2024 with relatively high Page Views (PVs). Given that each channel lasts approximately 2 to 10 hours, we sample one image per hour for the first three hours of each channel, resulting in three images sampled from each channel. Then, a SOTA face detection model [65] is employed to generate the bounding box (bbox) of the face region." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 389, + 554, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 554, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 554, + 473 + ], + "type": "text", + "content": "Data cleaning is further conducted to eliminate samples containing animated characters, virtual avatars, and samples from the same individual, to ensure the quality and diversity of the dataset. Finally, 10,000 images are collected, ensuring 10,000 different faces from over 9,500 live channels. Fig. 2 shows some examples of the face images in our LiveBeauty dataset." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 480, + 448, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 448, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 448, + 493 + ], + "type": "text", + "content": "3.2. Subjective Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 498, + 554, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 604 + ], + "type": "text", + "content": "We recruit 20 subjects in our experiments, including 14 females and 6 males. This ratio is determined by the actual customer demographics of the live platform. Face stimulus are displayed on an iPhone 14 Pro Max with 6.7-inch Organic Light-Emitting Diode (OLED) screen, supporting a resolution up to " + }, + { + "bbox": [ + 313, + 498, + 554, + 604 + ], + "type": "inline_equation", + "content": "2,796 \\times 1,290" + }, + { + "bbox": [ + 313, + 498, + 554, + 604 + ], + "type": "text", + "content": ". Subjective experiments are conducted in a well-controlled lab environment with fixed settings [44], i.e., lighting, viewing distance, viewing angle, etc." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "We split the experiment into 200 sessions with 50 images per session. Each face image is displayed for 10 seconds and followed by a one-second gray screen mask. Before the experiment starting, we offer a brief explanation and 50 extra training samples for subjects to familiarize them with the task. During the experiment, subjects are asked to evaluate the facial attractiveness of the stimuli from 1 to 5, defined as: \"unappealing\", \"less favorable\", \"ordinary\", \"online influencer\" and \"celebrity\", with detailed descriptions" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11460" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 70, + 283, + 191 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 283, + 191 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 283, + 191 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 283, + 191 + ], + "type": "image", + "image_path": "bf14c43aac92b4d22ef2b3368d64627042e8ae27de4b91c394d5a94b8aa41be3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 198, + 282, + 221 + ], + "lines": [ + { + "bbox": [ + 69, + 198, + 282, + 221 + ], + "spans": [ + { + "bbox": [ + 69, + 198, + 282, + 221 + ], + "type": "text", + "content": "Figure 3. Illustration of the LiveBeauty MOS distribution. Skewness=-0.470 Kurtosis=1.498" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 228, + 296, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 296, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 296, + 301 + ], + "type": "text", + "content": "of makeup and skin condition. After each session, subjects are asked to take at least 5 minutes of break. All subjects participate in all sessions, and each subject is allowed to participate in no more than 5 sessions each day. Ultimately, all 10,000 images are evaluated by 20 subjects, resulting in a total of 200,000 " + }, + { + "bbox": [ + 55, + 228, + 296, + 301 + ], + "type": "inline_equation", + "content": "(10,000 \\times 20)" + }, + { + "bbox": [ + 55, + 228, + 296, + 301 + ], + "type": "text", + "content": " scores collected." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 308, + 144, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 308, + 144, + 320 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 144, + 320 + ], + "type": "text", + "content": "3.3. Data Analysis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 326, + 296, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 326, + 296, + 446 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 296, + 446 + ], + "type": "text", + "content": "According to [44], we employ subject post-screening with outlier ratio and Spearman Rank Correlation Coefficient (SROCC). Annotations that differ by more than two levels from the mode of all annotations are considered potential outliers and will be meticulously removed by the experts. A subject whose rated scores with SROCC " + }, + { + "bbox": [ + 55, + 326, + 296, + 446 + ], + "type": "inline_equation", + "content": "< 0.75" + }, + { + "bbox": [ + 55, + 326, + 296, + 446 + ], + "type": "text", + "content": " or outlier ratio " + }, + { + "bbox": [ + 55, + 326, + 296, + 446 + ], + "type": "inline_equation", + "content": ">2\\%" + }, + { + "bbox": [ + 55, + 326, + 296, + 446 + ], + "type": "text", + "content": " will be marked as unreliable and will be rejected. Ultimately, 20 subjects are retained. We compute the Mean Opinion Score (MOS) as the attractiveness label of each face image:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 456, + 295, + 489 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 456, + 295, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 456, + 295, + 489 + ], + "type": "interline_equation", + "content": "\\mathrm {M O S} _ {j} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} r _ {i j}, \\tag {1}", + "image_path": "f613077bbe3d92234cf9210a855d781ee743128f89e093690417ff0b4a9b1be1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 497, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 295, + 544 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 497, + 295, + 544 + ], + "type": "inline_equation", + "content": "\\mathrm{MOS}_j" + }, + { + "bbox": [ + 55, + 497, + 295, + 544 + ], + "type": "text", + "content": " represents the MOS for the " + }, + { + "bbox": [ + 55, + 497, + 295, + 544 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 497, + 295, + 544 + ], + "type": "text", + "content": "-th face image, " + }, + { + "bbox": [ + 55, + 497, + 295, + 544 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 497, + 295, + 544 + ], + "type": "text", + "content": " is the number of the valid annotation scores rated by subjects, and " + }, + { + "bbox": [ + 55, + 497, + 295, + 544 + ], + "type": "inline_equation", + "content": "r_{ij}" + }, + { + "bbox": [ + 55, + 497, + 295, + 544 + ], + "type": "text", + "content": " is the rated score of the " + }, + { + "bbox": [ + 55, + 497, + 295, + 544 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 497, + 295, + 544 + ], + "type": "text", + "content": "-th subject on the " + }, + { + "bbox": [ + 55, + 497, + 295, + 544 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 497, + 295, + 544 + ], + "type": "text", + "content": "-th face image." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "text", + "content": "MOS distribution of our LiveBeauty is shown in Fig. 3. It can be observed that the distributions of MOSs exhibit a Gaussian-like shape. Skewness and kurtosis values of three MOS distributions are also shown in Fig. 3, which indicate that the distributions are characterized by thin tails and the MOS is concentrated around 3. This aligns with real-world facial attractiveness distributions, i.e., most individuals possess average facial attractiveness." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "The annotation distribution of four attractive levels of each face sample is shown in Fig. 4. Although attractiveness is highly subjective and influenced by personal preference, it can be observed that the annotation scores for each face are concentrated around the three levels closest to the MOS (e.g., 3.85 comes from five 3, thirteen 4, and two 5)." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 70, + 553, + 190 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 553, + 190 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 190 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 190 + ], + "type": "image", + "image_path": "8cc5d4d81ef50b1584debd21ec78c61797f00ce540a77f0478176e5772d57e7a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 319, + 198, + 548, + 210 + ], + "lines": [ + { + "bbox": [ + 319, + 198, + 548, + 210 + ], + "spans": [ + { + "bbox": [ + 319, + 198, + 548, + 210 + ], + "type": "text", + "content": "Figure 4. Annotation score distribution on four attractive level." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 217, + 421, + 230 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 421, + 230 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 421, + 230 + ], + "type": "text", + "content": "4. Proposed Method" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 237, + 553, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 237, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 237, + 553, + 274 + ], + "type": "text", + "content": "As illustrated in Fig. 5, our Facial Prior Enhanced Multi-modal (FPEM) method utilizes a two-stage training strategy, and consists of three primary modules." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 274, + 555, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 555, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 555, + 455 + ], + "type": "text", + "content": "The first stage, namely Preliminary Training Phase, includes the Personalized Attractiveness Prior Module (PAPM in Sec. 4.1.1) and the Multi-modal Attractiveness Encoder Module (MAEM in Sec. 4.1.2). PAPM receives the image " + }, + { + "bbox": [ + 313, + 274, + 555, + 455 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 274, + 555, + 455 + ], + "type": "text", + "content": ", from which a Swin Transformer " + }, + { + "bbox": [ + 313, + 274, + 555, + 455 + ], + "type": "inline_equation", + "content": "E_{MS}" + }, + { + "bbox": [ + 313, + 274, + 555, + 455 + ], + "type": "text", + "content": " extracts multi-scale feature " + }, + { + "bbox": [ + 313, + 274, + 555, + 455 + ], + "type": "inline_equation", + "content": "f_{m}" + }, + { + "bbox": [ + 313, + 274, + 555, + 455 + ], + "type": "text", + "content": ", and a frozen FaceNet " + }, + { + "bbox": [ + 313, + 274, + 555, + 455 + ], + "type": "inline_equation", + "content": "E_{FP}" + }, + { + "bbox": [ + 313, + 274, + 555, + 455 + ], + "type": "text", + "content": " extracts face-aware feature " + }, + { + "bbox": [ + 313, + 274, + 555, + 455 + ], + "type": "inline_equation", + "content": "f_{s}" + }, + { + "bbox": [ + 313, + 274, + 555, + 455 + ], + "type": "text", + "content": ". Subsequently, a Cross-Attention block incorporates " + }, + { + "bbox": [ + 313, + 274, + 555, + 455 + ], + "type": "inline_equation", + "content": "f_{m}" + }, + { + "bbox": [ + 313, + 274, + 555, + 455 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 274, + 555, + 455 + ], + "type": "inline_equation", + "content": "f_{s}" + }, + { + "bbox": [ + 313, + 274, + 555, + 455 + ], + "type": "text", + "content": " to form the personalized attractiveness feature " + }, + { + "bbox": [ + 313, + 274, + 555, + 455 + ], + "type": "inline_equation", + "content": "f_{p}" + }, + { + "bbox": [ + 313, + 274, + 555, + 455 + ], + "type": "text", + "content": ", which are then projected to a preliminary attractiveness score " + }, + { + "bbox": [ + 313, + 274, + 555, + 455 + ], + "type": "inline_equation", + "content": "\\widehat{Q}_1" + }, + { + "bbox": [ + 313, + 274, + 555, + 455 + ], + "type": "text", + "content": ". The MAEM contains an image encoder " + }, + { + "bbox": [ + 313, + 274, + 555, + 455 + ], + "type": "inline_equation", + "content": "E_{I}" + }, + { + "bbox": [ + 313, + 274, + 555, + 455 + ], + "type": "text", + "content": " and a text encoder " + }, + { + "bbox": [ + 313, + 274, + 555, + 455 + ], + "type": "inline_equation", + "content": "E_{T}" + }, + { + "bbox": [ + 313, + 274, + 555, + 455 + ], + "type": "text", + "content": " as in CLIP [42], receives the image " + }, + { + "bbox": [ + 313, + 274, + 555, + 455 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 274, + 555, + 455 + ], + "type": "text", + "content": " and progressive texts, leverages visual embeddings " + }, + { + "bbox": [ + 313, + 274, + 555, + 455 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 313, + 274, + 555, + 455 + ], + "type": "text", + "content": " and textual embeddings " + }, + { + "bbox": [ + 313, + 274, + 555, + 455 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 313, + 274, + 555, + 455 + ], + "type": "text", + "content": ", and obtains another preliminary score " + }, + { + "bbox": [ + 313, + 274, + 555, + 455 + ], + "type": "inline_equation", + "content": "\\widehat{Q}_2" + }, + { + "bbox": [ + 313, + 274, + 555, + 455 + ], + "type": "text", + "content": " through a Similarity Regression (SR) operation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 456, + 554, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 554, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 554, + 518 + ], + "type": "text", + "content": "In the second stage, namely Selective Fusion Phase, a Selective Fusion Module (SFM in Sec. 4.2) dynamically refines the extracted visual embeddings " + }, + { + "bbox": [ + 313, + 456, + 554, + 518 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 313, + 456, + 554, + 518 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 456, + 554, + 518 + ], + "type": "inline_equation", + "content": "f_{i}^{\\prime}" + }, + { + "bbox": [ + 313, + 456, + 554, + 518 + ], + "type": "text", + "content": " based on the personalized attractiveness feature " + }, + { + "bbox": [ + 313, + 456, + 554, + 518 + ], + "type": "inline_equation", + "content": "f_{p}" + }, + { + "bbox": [ + 313, + 456, + 554, + 518 + ], + "type": "text", + "content": ". Then the SR strategy is used to predict the final score " + }, + { + "bbox": [ + 313, + 456, + 554, + 518 + ], + "type": "inline_equation", + "content": "\\widehat{Q}" + }, + { + "bbox": [ + 313, + 456, + 554, + 518 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 527, + 468, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 527, + 468, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 527, + 468, + 540 + ], + "type": "text", + "content": "4.1. Preliminary Training Phase" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 545, + 554, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 583 + ], + "type": "text", + "content": "In this stage, PAPM and MAEM are trained independently without considering the SFM, and MAEM employs the SR strategy for regression during training." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 590, + 520, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 520, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 520, + 601 + ], + "type": "text", + "content": "4.1.1. Personalized Attractiveness Prior Module" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "**Stem Swin.** Considering that the visual perception progress is perceived from low level to high level, the hierarchical features from consecutive transformer stages are fully leveraged. Specifically, the features " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "f_{r_i} \\in \\mathbb{R}^{H_i \\times W_i \\times C_i}" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": " extracted from " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "-th stage are reshaped into same size, and are then concatenated as " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "f_{\\Sigma} \\in \\mathbb{R}^{H_4 \\times W_4 \\times C_{\\Sigma}}" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": ". Meanwhile, " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "f_{\\Sigma}" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": " is refined by a Channel Attention block from channel perspective (which is shown in Fig. 6), and then flattened as " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "f_m \\in \\mathbb{R}^{N_4 \\times C_{\\Sigma}}" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11461" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 78, + 545, + 291 + ], + "blocks": [ + { + "bbox": [ + 69, + 78, + 545, + 291 + ], + "lines": [ + { + "bbox": [ + 69, + 78, + 545, + 291 + ], + "spans": [ + { + "bbox": [ + 69, + 78, + 545, + 291 + ], + "type": "image", + "image_path": "324f40abd091c23171097a2525a108107b16200a9f71dbaf42cbb6eb88c4a023.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 304, + 555, + 329 + ], + "lines": [ + { + "bbox": [ + 54, + 304, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 54, + 304, + 555, + 329 + ], + "type": "text", + "content": "Figure 5. Architecture of the Facial Prior Enhanced Multi-modal model (FPEM), including the Personalized Attractiveness Prior Module (PAPM), the Multi-modal Attractiveness Encoder Module (MAEM) and the Selective Fusion Module (SFM)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 348, + 296, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 348, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 54, + 348, + 296, + 468 + ], + "type": "text", + "content": "Cross-Attention. Facial prior knowledge, i.e., face-aware features " + }, + { + "bbox": [ + 54, + 348, + 296, + 468 + ], + "type": "inline_equation", + "content": "f_{s}" + }, + { + "bbox": [ + 54, + 348, + 296, + 468 + ], + "type": "text", + "content": " extracted by FaceNet, are utilized in this block to guide the learning of facial attractiveness. Specifically, the refined features " + }, + { + "bbox": [ + 54, + 348, + 296, + 468 + ], + "type": "inline_equation", + "content": "f_{m}" + }, + { + "bbox": [ + 54, + 348, + 296, + 468 + ], + "type": "text", + "content": " extracted by the stem Swin are fused with the face-aware features " + }, + { + "bbox": [ + 54, + 348, + 296, + 468 + ], + "type": "inline_equation", + "content": "f_{s}" + }, + { + "bbox": [ + 54, + 348, + 296, + 468 + ], + "type": "text", + "content": " through cross-attention. As shown in Fig. 6, " + }, + { + "bbox": [ + 54, + 348, + 296, + 468 + ], + "type": "inline_equation", + "content": "f_{s}" + }, + { + "bbox": [ + 54, + 348, + 296, + 468 + ], + "type": "text", + "content": " is encoded as the query " + }, + { + "bbox": [ + 54, + 348, + 296, + 468 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 54, + 348, + 296, + 468 + ], + "type": "text", + "content": " via a Linear layer and a Layer Normalization (LN) layer, while " + }, + { + "bbox": [ + 54, + 348, + 296, + 468 + ], + "type": "inline_equation", + "content": "f_{m}" + }, + { + "bbox": [ + 54, + 348, + 296, + 468 + ], + "type": "text", + "content": " is encoded as the key " + }, + { + "bbox": [ + 54, + 348, + 296, + 468 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 54, + 348, + 296, + 468 + ], + "type": "text", + "content": " and value " + }, + { + "bbox": [ + 54, + 348, + 296, + 468 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 54, + 348, + 296, + 468 + ], + "type": "text", + "content": " via a LN layer. Further, a Multi-Head Cross-Attention (MHCA) mechanism is applied, which is computed as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 81, + 474, + 295, + 508 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 474, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 81, + 474, + 295, + 508 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} \\operatorname {h e a d} _ {i} = \\psi \\left(\\frac {Q \\cdot K ^ {T}}{\\sqrt {d}}\\right) \\cdot V \\cdot W _ {i}, i \\in \\{1, \\dots , h \\} \\\\ \\operatorname {M H C A} (Q, K, V) = \\operatorname {C o n c a t} (\\operatorname {h e a d} _ {i}) W _ {o}, \\end{array} \\right. \\tag {2}", + "image_path": "494bc945ed667a75e3a11c877ed47d8c4e858760b3538cf0c5f445fd3ce08472.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 514, + 296, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 296, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 296, + 574 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 514, + 296, + 574 + ], + "type": "inline_equation", + "content": "\\psi (\\cdot)" + }, + { + "bbox": [ + 55, + 514, + 296, + 574 + ], + "type": "text", + "content": " denotes the softmax function, " + }, + { + "bbox": [ + 55, + 514, + 296, + 574 + ], + "type": "inline_equation", + "content": "W_{o}" + }, + { + "bbox": [ + 55, + 514, + 296, + 574 + ], + "type": "text", + "content": " denotes the projection parameter matrix and " + }, + { + "bbox": [ + 55, + 514, + 296, + 574 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 514, + 296, + 574 + ], + "type": "text", + "content": " denotes the query dimension. Let FFN indicates the Feed-Forward Network, AP denotes Average Pooling, the output feature " + }, + { + "bbox": [ + 55, + 514, + 296, + 574 + ], + "type": "inline_equation", + "content": "f_{p}\\in \\mathbb{R}^{1\\times C}" + }, + { + "bbox": [ + 55, + 514, + 296, + 574 + ], + "type": "text", + "content": " is computed as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 103, + 580, + 295, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 580, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 103, + 580, + 295, + 613 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} f _ {c} = f _ {m} + \\mathrm {M H C A} (Q, K, V) \\\\ f _ {p} = \\mathrm {A P} \\left(f _ {c} + \\text {F F N} \\left(\\mathrm {L N} \\left(f _ {c}\\right)\\right)\\right). \\end{array} \\right. \\tag {3}", + "image_path": "4438d252277fa8d3061d7d6f915d9cf6666b265acb7d54312c7f49ec6f018c08.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 618, + 296, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 643 + ], + "type": "text", + "content": "Let MLP represent the Multi Layer Perceptron, then the attractiveness score " + }, + { + "bbox": [ + 55, + 618, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\widehat{Q}_1" + }, + { + "bbox": [ + 55, + 618, + 296, + 643 + ], + "type": "text", + "content": " can be predicted by:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 140, + 647, + 295, + 663 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 647, + 295, + 663 + ], + "spans": [ + { + "bbox": [ + 140, + 647, + 295, + 663 + ], + "type": "interline_equation", + "content": "\\widehat {Q} _ {1} = \\operatorname {M L P} \\left(f _ {p}\\right). \\tag {4}", + "image_path": "6160b8f35bd53bd4432adf35fdc267be3c09f1a486f4f462a4c1c06a4bfcca5f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 674, + 275, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 674, + 275, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 674, + 275, + 685 + ], + "type": "text", + "content": "4.1.2. Multi-modal Attractiveness Encoder Module" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Inspired by [66], MAEM is devised to utilize multi-modal aesthetic semantic features. We utilize five attractiveness" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 326, + 342, + 432, + 488 + ], + "blocks": [ + { + "bbox": [ + 326, + 342, + 432, + 488 + ], + "lines": [ + { + "bbox": [ + 326, + 342, + 432, + 488 + ], + "spans": [ + { + "bbox": [ + 326, + 342, + 432, + 488 + ], + "type": "image", + "image_path": "22e36a6f14da6bdfacc30bf404eaf2ee5197c36afb4662059e959d96b30ce523.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 496, + 555, + 519 + ], + "lines": [ + { + "bbox": [ + 313, + 496, + 555, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 555, + 519 + ], + "type": "text", + "content": "Figure 6. Structure of the Channel Attention and the Cross-Attention Block." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 437, + 341, + 541, + 490 + ], + "blocks": [ + { + "bbox": [ + 437, + 341, + 541, + 490 + ], + "lines": [ + { + "bbox": [ + 437, + 341, + 541, + 490 + ], + "spans": [ + { + "bbox": [ + 437, + 341, + 541, + 490 + ], + "type": "image", + "image_path": "4efb84d53be049b81efa9c9cf066a409987082ea995876dbae46e5ee7f355b99.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 528, + 555, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 555, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 555, + 576 + ], + "type": "text", + "content": "levels " + }, + { + "bbox": [ + 313, + 528, + 555, + 576 + ], + "type": "inline_equation", + "content": "a \\in \\mathcal{A} = \\{1,2,3,4,5\\} = \\{\\text{\"bad\", \"poor\", \"fair\", \"good\", \"perfect\"}\\}" + }, + { + "bbox": [ + 313, + 528, + 555, + 576 + ], + "type": "text", + "content": ". Then the text template can be formed as \"a photo of a person with " + }, + { + "bbox": [ + 313, + 528, + 555, + 576 + ], + "type": "inline_equation", + "content": "\\{a\\}" + }, + { + "bbox": [ + 313, + 528, + 555, + 576 + ], + "type": "text", + "content": " attractiveness\", resulting in five candidate text descriptions " + }, + { + "bbox": [ + 313, + 528, + 555, + 576 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 528, + 555, + 576 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 576, + 554, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 554, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 554, + 636 + ], + "type": "text", + "content": "MAEM contains a frozen text encoder " + }, + { + "bbox": [ + 313, + 576, + 554, + 636 + ], + "type": "inline_equation", + "content": "E_{T}" + }, + { + "bbox": [ + 313, + 576, + 554, + 636 + ], + "type": "text", + "content": " (GPT-2) and a trainable image encoder " + }, + { + "bbox": [ + 313, + 576, + 554, + 636 + ], + "type": "inline_equation", + "content": "E_{I}" + }, + { + "bbox": [ + 313, + 576, + 554, + 636 + ], + "type": "text", + "content": " (ViT-B/16), where " + }, + { + "bbox": [ + 313, + 576, + 554, + 636 + ], + "type": "inline_equation", + "content": "E_{T}" + }, + { + "bbox": [ + 313, + 576, + 554, + 636 + ], + "type": "text", + "content": " receives five candidate text descriptions " + }, + { + "bbox": [ + 313, + 576, + 554, + 636 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 576, + 554, + 636 + ], + "type": "text", + "content": " to form the textual embeddings " + }, + { + "bbox": [ + 313, + 576, + 554, + 636 + ], + "type": "inline_equation", + "content": "f_{t}^{\\prime}\\in \\mathbb{R}^{5\\times C}" + }, + { + "bbox": [ + 313, + 576, + 554, + 636 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 313, + 576, + 554, + 636 + ], + "type": "inline_equation", + "content": "E_{I}" + }, + { + "bbox": [ + 313, + 576, + 554, + 636 + ], + "type": "text", + "content": " receives the image " + }, + { + "bbox": [ + 313, + 576, + 554, + 636 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 576, + 554, + 636 + ], + "type": "text", + "content": " to form the visual embeddings " + }, + { + "bbox": [ + 313, + 576, + 554, + 636 + ], + "type": "inline_equation", + "content": "f_{i}\\in \\mathbb{R}^{1\\times C}" + }, + { + "bbox": [ + 313, + 576, + 554, + 636 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 636, + 554, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 636, + 554, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 636, + 554, + 685 + ], + "type": "text", + "content": "Subsequently, the textual embeddings " + }, + { + "bbox": [ + 313, + 636, + 554, + 685 + ], + "type": "inline_equation", + "content": "f_{t}^{\\prime}" + }, + { + "bbox": [ + 313, + 636, + 554, + 685 + ], + "type": "text", + "content": " and visual embeddings " + }, + { + "bbox": [ + 313, + 636, + 554, + 685 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 313, + 636, + 554, + 685 + ], + "type": "text", + "content": " are fed into Similar Regression (SR). In particular, the cosine similarity between " + }, + { + "bbox": [ + 313, + 636, + 554, + 685 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 313, + 636, + 554, + 685 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 636, + 554, + 685 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 313, + 636, + 554, + 685 + ], + "type": "text", + "content": " is calculated to obtain the attractiveness level probability distribution " + }, + { + "bbox": [ + 313, + 636, + 554, + 685 + ], + "type": "inline_equation", + "content": "\\hat{p}" + }, + { + "bbox": [ + 313, + 636, + 554, + 685 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 388, + 690, + 553, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 690, + 553, + 717 + ], + "spans": [ + { + "bbox": [ + 388, + 690, + 553, + 717 + ], + "type": "interline_equation", + "content": "\\hat {p} = \\psi \\left(\\frac {f _ {i} \\cdot f _ {t}}{\\| f _ {i} \\| \\times \\| f _ {t} \\|}\\right), \\tag {5}", + "image_path": "f925b16d242699602917135273155ea180c669f50538374c9c42e04777859d49.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "11462" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 99 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 99 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 99 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 295, + 99 + ], + "type": "inline_equation", + "content": "\\psi (\\cdot)" + }, + { + "bbox": [ + 55, + 72, + 295, + 99 + ], + "type": "text", + "content": " denotes the softmax function. Then the attractiveness scores " + }, + { + "bbox": [ + 55, + 72, + 295, + 99 + ], + "type": "inline_equation", + "content": "\\widehat{Q}_2" + }, + { + "bbox": [ + 55, + 72, + 295, + 99 + ], + "type": "text", + "content": " can be predicted by:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 102, + 295, + 135 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 102, + 295, + 135 + ], + "spans": [ + { + "bbox": [ + 132, + 102, + 295, + 135 + ], + "type": "interline_equation", + "content": "\\widehat {Q} _ {2} = \\sum_ {a = 1} ^ {5} \\hat {p} (a | I) \\cdot a, \\tag {6}", + "image_path": "a1ef14cfefdd536e3df6a610b8dbbc72b7493dfacf2a7a9f7fe8f60c8863e2b8.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 140, + 295, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 140, + 295, + 165 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 295, + 165 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 140, + 295, + 165 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 55, + 140, + 295, + 165 + ], + "type": "text", + "content": " denotes the attractiveness level from 1 to 5, and " + }, + { + "bbox": [ + 55, + 140, + 295, + 165 + ], + "type": "inline_equation", + "content": "\\hat{p}(a|\\cdot)" + }, + { + "bbox": [ + 55, + 140, + 295, + 165 + ], + "type": "text", + "content": " is the probability of the level " + }, + { + "bbox": [ + 55, + 140, + 295, + 165 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 55, + 140, + 295, + 165 + ], + "type": "text", + "content": " to be estimated." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 171, + 185, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 171, + 185, + 182 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 185, + 182 + ], + "type": "text", + "content": "4.2. Selective Fusion Phase" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 188, + 296, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 188, + 296, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 188, + 296, + 331 + ], + "type": "text", + "content": "In this stage, we freeze the parameters of MAEM and PAPM, which are trained in the previous preliminary training phase. Then the Selective Fusion Model (SFM) is trained to dynamically integrate holistic facial prior knowledge and multi-modal aesthetic semantic features. To achieve this goal, we first fuse the visual embeddings " + }, + { + "bbox": [ + 55, + 188, + 296, + 331 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 55, + 188, + 296, + 331 + ], + "type": "text", + "content": " and personalized attractiveness feature " + }, + { + "bbox": [ + 55, + 188, + 296, + 331 + ], + "type": "inline_equation", + "content": "f_{p}" + }, + { + "bbox": [ + 55, + 188, + 296, + 331 + ], + "type": "text", + "content": " via an element-wise addition. Inspired by [26], a compact feature " + }, + { + "bbox": [ + 55, + 188, + 296, + 331 + ], + "type": "inline_equation", + "content": "f_{z} \\in \\mathbb{R}^{1 \\times d}" + }, + { + "bbox": [ + 55, + 188, + 296, + 331 + ], + "type": "text", + "content": " is further generated to facilitate adaptive selection guidance for these facial features, which is produced by a Squeeze Layer implemented utilizing a Fully Connected layer (FC) and a ReLU activation function:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 338, + 295, + 352 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 338, + 295, + 352 + ], + "spans": [ + { + "bbox": [ + 116, + 338, + 295, + 352 + ], + "type": "interline_equation", + "content": "f _ {z} = \\operatorname {R e L U} \\left(F C \\left(f _ {i} + f _ {p}\\right)\\right), \\tag {7}", + "image_path": "bc0a3338292542e3450ca603c2d1777cc116000e779826aa57421f97260d16d6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 357, + 296, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 296, + 428 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 296, + 428 + ], + "type": "inline_equation", + "content": "f_{z}" + }, + { + "bbox": [ + 55, + 357, + 296, + 428 + ], + "type": "text", + "content": " are then projected to two adaptive weights " + }, + { + "bbox": [ + 55, + 357, + 296, + 428 + ], + "type": "inline_equation", + "content": "\\mathbf{a} \\in \\mathbb{R}^{1 \\times C}" + }, + { + "bbox": [ + 55, + 357, + 296, + 428 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 357, + 296, + 428 + ], + "type": "inline_equation", + "content": "\\mathbf{b} \\in \\mathbb{R}^{1 \\times C}" + }, + { + "bbox": [ + 55, + 357, + 296, + 428 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 357, + 296, + 428 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 55, + 357, + 296, + 428 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 357, + 296, + 428 + ], + "type": "inline_equation", + "content": "f_{p}" + }, + { + "bbox": [ + 55, + 357, + 296, + 428 + ], + "type": "text", + "content": " via two FC layers. A Softmax function is employed on " + }, + { + "bbox": [ + 55, + 357, + 296, + 428 + ], + "type": "inline_equation", + "content": "\\mathbf{a}" + }, + { + "bbox": [ + 55, + 357, + 296, + 428 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 357, + 296, + 428 + ], + "type": "inline_equation", + "content": "\\mathbf{b}" + }, + { + "bbox": [ + 55, + 357, + 296, + 428 + ], + "type": "text", + "content": " on their channel-wise digits to ensure " + }, + { + "bbox": [ + 55, + 357, + 296, + 428 + ], + "type": "inline_equation", + "content": "\\mathbf{a} + \\mathbf{b} = \\mathbf{1}" + }, + { + "bbox": [ + 55, + 357, + 296, + 428 + ], + "type": "text", + "content": ". Afterwards, the optimized visual embeddings " + }, + { + "bbox": [ + 55, + 357, + 296, + 428 + ], + "type": "inline_equation", + "content": "f_{i}'" + }, + { + "bbox": [ + 55, + 357, + 296, + 428 + ], + "type": "text", + "content": " can be obtained by the channel-wise weighted addition with " + }, + { + "bbox": [ + 55, + 357, + 296, + 428 + ], + "type": "inline_equation", + "content": "\\mathbf{a}" + }, + { + "bbox": [ + 55, + 357, + 296, + 428 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 357, + 296, + 428 + ], + "type": "inline_equation", + "content": "\\mathbf{b}" + }, + { + "bbox": [ + 55, + 357, + 296, + 428 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 435, + 295, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 435, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 134, + 435, + 295, + 449 + ], + "type": "interline_equation", + "content": "f _ {i} ^ {\\prime} = f _ {i} \\cdot \\mathbf {a} + f _ {p} \\cdot \\mathbf {b}. \\tag {8}", + "image_path": "55afed1821e1939bf0b73bb85e75a01cddc9e2fa6c407f03b1d6a48ac54c6286.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 454, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 454, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 296, + 491 + ], + "type": "text", + "content": "Subsequently, the refined " + }, + { + "bbox": [ + 55, + 454, + 296, + 491 + ], + "type": "inline_equation", + "content": "f_{i}^{\\prime}" + }, + { + "bbox": [ + 55, + 454, + 296, + 491 + ], + "type": "text", + "content": " and textual embeddings " + }, + { + "bbox": [ + 55, + 454, + 296, + 491 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 55, + 454, + 296, + 491 + ], + "type": "text", + "content": " are input into Similar Regression to calculate the final prediction " + }, + { + "bbox": [ + 55, + 454, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\widehat{Q}" + }, + { + "bbox": [ + 55, + 454, + 296, + 491 + ], + "type": "text", + "content": ", which is consistent with the MAEM." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 496, + 168, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 168, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 168, + 509 + ], + "type": "text", + "content": "4.3. Training objectives" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 514, + 296, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 296, + 551 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 296, + 551 + ], + "type": "text", + "content": "Loss for PAPM. The L1 loss between " + }, + { + "bbox": [ + 55, + 514, + 296, + 551 + ], + "type": "inline_equation", + "content": "\\widehat{Q}_1" + }, + { + "bbox": [ + 55, + 514, + 296, + 551 + ], + "type": "text", + "content": " and the ground truth label " + }, + { + "bbox": [ + 55, + 514, + 296, + 551 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 514, + 296, + 551 + ], + "type": "text", + "content": " is calculated as the attractiveness prediction loss function " + }, + { + "bbox": [ + 55, + 514, + 296, + 551 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 55, + 514, + 296, + 551 + ], + "type": "text", + "content": ", which can be formulated as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 123, + 555, + 295, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 555, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 123, + 555, + 295, + 586 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {1} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left| \\widehat {Q} _ {1 _ {i}} - Q _ {i} \\right|, \\tag {9}", + "image_path": "e075992e6988420b5e742731169662f1ab6453f8c2a80dd3a856d3438912cc11.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 592, + 239, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 592, + 239, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 239, + 604 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 592, + 239, + 604 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 592, + 239, + 604 + ], + "type": "text", + "content": " denotes number of training samples." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 604, + 296, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 296, + 677 + ], + "type": "text", + "content": "Loss for MAEM. To jointly optimize ranking and regression, we incorporate a merged ranking loss " + }, + { + "bbox": [ + 55, + 604, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_R" + }, + { + "bbox": [ + 55, + 604, + 296, + 677 + ], + "type": "text", + "content": " along with the L1 scoring loss " + }, + { + "bbox": [ + 55, + 604, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_S" + }, + { + "bbox": [ + 55, + 604, + 296, + 677 + ], + "type": "text", + "content": ". Specifically, " + }, + { + "bbox": [ + 55, + 604, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_R" + }, + { + "bbox": [ + 55, + 604, + 296, + 677 + ], + "type": "text", + "content": " is composed of a fidelity loss " + }, + { + "bbox": [ + 55, + 604, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{R_1}" + }, + { + "bbox": [ + 55, + 604, + 296, + 677 + ], + "type": "text", + "content": " and a two-direction ranking loss " + }, + { + "bbox": [ + 55, + 604, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{R_2}" + }, + { + "bbox": [ + 55, + 604, + 296, + 677 + ], + "type": "text", + "content": ". For the fidelity loss " + }, + { + "bbox": [ + 55, + 604, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{R_1}" + }, + { + "bbox": [ + 55, + 604, + 296, + 677 + ], + "type": "text", + "content": ", given an image pair " + }, + { + "bbox": [ + 55, + 604, + 296, + 677 + ], + "type": "inline_equation", + "content": "(I_i, I_j)" + }, + { + "bbox": [ + 55, + 604, + 296, + 677 + ], + "type": "text", + "content": ", the binary label according to their ground truth " + }, + { + "bbox": [ + 55, + 604, + 296, + 677 + ], + "type": "inline_equation", + "content": "(Q_i, Q_j)" + }, + { + "bbox": [ + 55, + 604, + 296, + 677 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 110, + 684, + 295, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 684, + 295, + 716 + ], + "spans": [ + { + "bbox": [ + 110, + 684, + 295, + 716 + ], + "type": "interline_equation", + "content": "R \\left(I _ {i}, I _ {j}\\right) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} Q _ {i} \\geq Q _ {j} \\\\ 0 & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {10}", + "image_path": "39fac5b5a3e57ed1970ee016d9d871f311adc32528910b182e166559d8270d1d.jpg" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 328, + 72, + 536, + 187 + ], + "blocks": [ + { + "bbox": [ + 328, + 72, + 536, + 187 + ], + "lines": [ + { + "bbox": [ + 328, + 72, + 536, + 187 + ], + "spans": [ + { + "bbox": [ + 328, + 72, + 536, + 187 + ], + "type": "image", + "image_path": "7757488bf5e6d34838806709c2ac8158ef1145c54fb8814a76cea56b27eddeae.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 194, + 537, + 205 + ], + "lines": [ + { + "bbox": [ + 331, + 194, + 537, + 205 + ], + "spans": [ + { + "bbox": [ + 331, + 194, + 537, + 205 + ], + "type": "text", + "content": "Figure 7. MOS distribution of three benchmark datasets." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 216, + 555, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 555, + 241 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 555, + 241 + ], + "type": "text", + "content": "According to the Thurstone's model [51], the probability of " + }, + { + "bbox": [ + 313, + 216, + 555, + 241 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 313, + 216, + 555, + 241 + ], + "type": "text", + "content": " perceiving more attractive than " + }, + { + "bbox": [ + 313, + 216, + 555, + 241 + ], + "type": "inline_equation", + "content": "I_{j}" + }, + { + "bbox": [ + 313, + 216, + 555, + 241 + ], + "type": "text", + "content": " is estimated as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 375, + 248, + 553, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 248, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 375, + 248, + 553, + 277 + ], + "type": "interline_equation", + "content": "\\widehat {R} \\left(I _ {i}, I _ {j}\\right) = \\Psi \\left(\\frac {\\widehat {Q} _ {2 i} - \\widehat {Q} _ {2 j}}{\\sqrt {2}}\\right), \\tag {11}", + "image_path": "4f212058c868543b61a6847755ce59df3e975e94100e7318180a7def718d59e4.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 285, + 554, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 554, + 320 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 554, + 320 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 285, + 554, + 320 + ], + "type": "inline_equation", + "content": "\\Psi (\\cdot)" + }, + { + "bbox": [ + 313, + 285, + 554, + 320 + ], + "type": "text", + "content": " denotes the standard normal cumulative distribution function, and its variance is set to 1. Then the fidelity loss [54] can be calculated as:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 349, + 328, + 553, + 371 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 328, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 349, + 328, + 553, + 371 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {R _ {1}} = 1 - \\sqrt {R \\left(I _ {i} , I _ {j}\\right) \\widehat {R} \\left(I _ {i} , I _ {j}\\right)} \\tag {12} \\\\ - \\sqrt {(1 - R (I _ {i} , I _ {j})) (1 - \\widehat {R} (I _ {i} , I _ {j}))}. \\\\ \\end{array}", + "image_path": "1360b5ddb0e0557e02f26c96ac00ad1ede666cba0dbdd21026160d479586ccde.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 378, + 555, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 555, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 555, + 504 + ], + "type": "text", + "content": "For the two-direction ranking loss " + }, + { + "bbox": [ + 313, + 378, + 555, + 504 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{R_2}" + }, + { + "bbox": [ + 313, + 378, + 555, + 504 + ], + "type": "text", + "content": ", given an image " + }, + { + "bbox": [ + 313, + 378, + 555, + 504 + ], + "type": "inline_equation", + "content": "I_i" + }, + { + "bbox": [ + 313, + 378, + 555, + 504 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 378, + 555, + 504 + ], + "type": "inline_equation", + "content": "\\widehat{S}_{i,j}" + }, + { + "bbox": [ + 313, + 378, + 555, + 504 + ], + "type": "text", + "content": " denotes the cosine similarity score of " + }, + { + "bbox": [ + 313, + 378, + 555, + 504 + ], + "type": "inline_equation", + "content": "j_{th}" + }, + { + "bbox": [ + 313, + 378, + 555, + 504 + ], + "type": "text", + "content": " attractiveness level and " + }, + { + "bbox": [ + 313, + 378, + 555, + 504 + ], + "type": "inline_equation", + "content": "\\widehat{S}_{i,o} = \\max \\left\\{\\widehat{S}_{i,j} \\mid j \\in \\{1,2,3,4,5\\}\\right\\}" + }, + { + "bbox": [ + 313, + 378, + 555, + 504 + ], + "type": "text", + "content": ". Since the human perception of facial attractiveness is gradual and rarely exhibits multiple peaks [62], we further obtain the ranking information by ensuring that the attractiveness level probability distribution only exhibits one peak and decreases in two directions (leftward and rightward). Then the two-direction pairwise ranking loss can be formulated as:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 328, + 521, + 554, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 521, + 554, + 576 + ], + "spans": [ + { + "bbox": [ + 328, + 521, + 554, + 576 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} \\mathcal {L} _ {\\text {l e f t}} = - \\sum_ {j = 2} ^ {o} \\log \\frac {\\exp \\left(\\widehat {S} _ {i , j}\\right)}{\\exp \\left(\\widehat {S} _ {i , j}\\right) + \\exp \\left(\\widehat {S} _ {i , j - 1}\\right)} \\\\ \\mathcal {L} _ {\\text {r i g h t}} = - \\sum_ {j = o} ^ {A - 1} \\log \\frac {\\exp \\left(\\widehat {S} _ {i , j}\\right)}{\\exp \\left(\\widehat {S} _ {i , j}\\right) + \\exp \\left(\\widehat {S} _ {i , j + 1}\\right)} \\\\ \\mathcal {L} _ {R _ {2}} = - \\left(\\mathcal {L} _ {\\text {l e f t}} + \\mathcal {L} _ {\\text {r i g h t}}\\right), \\end{array} \\right. \\tag {13}", + "image_path": "669ce97a2a83610f4308df0c2a880c4d80990f212fff856d86344b7ba899972c.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 583, + 554, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 583, + 554, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 583, + 554, + 643 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 583, + 554, + 643 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 583, + 554, + 643 + ], + "type": "text", + "content": " denotes the number of attractiveness levels, and the corresponding labels of " + }, + { + "bbox": [ + 313, + 583, + 554, + 643 + ], + "type": "inline_equation", + "content": "\\widehat{S}_{i,j-1}" + }, + { + "bbox": [ + 313, + 583, + 554, + 643 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 583, + 554, + 643 + ], + "type": "inline_equation", + "content": "\\widehat{S}_{i,j}" + }, + { + "bbox": [ + 313, + 583, + 554, + 643 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 583, + 554, + 643 + ], + "type": "inline_equation", + "content": "\\widehat{S}_{i,j+1}" + }, + { + "bbox": [ + 313, + 583, + 554, + 643 + ], + "type": "text", + "content": " are 0, 1, 0, respectively. In the end, the total loss of MAEM can be defined as " + }, + { + "bbox": [ + 313, + 583, + 554, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2 = \\mathcal{L}_S + \\lambda_1\\mathcal{L}_{R_1} + \\lambda_2\\mathcal{L}_{R_2}" + }, + { + "bbox": [ + 313, + 583, + 554, + 643 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 583, + 554, + 643 + ], + "type": "inline_equation", + "content": "\\lambda_1" + }, + { + "bbox": [ + 313, + 583, + 554, + 643 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 583, + 554, + 643 + ], + "type": "inline_equation", + "content": "\\lambda_2" + }, + { + "bbox": [ + 313, + 583, + 554, + 643 + ], + "type": "text", + "content": " are empirically set as 1." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 643, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 643, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 643, + 554, + 677 + ], + "type": "text", + "content": "Loss for SFM. The loss used in the second stage between the ground truth label " + }, + { + "bbox": [ + 313, + 643, + 554, + 677 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 643, + 554, + 677 + ], + "type": "text", + "content": " and the final attractiveness score " + }, + { + "bbox": [ + 313, + 643, + 554, + 677 + ], + "type": "inline_equation", + "content": "\\hat{Q}" + }, + { + "bbox": [ + 313, + 643, + 554, + 677 + ], + "type": "text", + "content": " is the L1 Loss:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 383, + 685, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 685, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 383, + 685, + 553, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {3} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left| \\widehat {Q} _ {i} - Q _ {i} \\right|. \\tag {14}", + "image_path": "8f2b298330304df929c3c76523772260fd155303e80df7b682bda3590fbcdd5c.jpg" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11463" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 91, + 550, + 268 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 538, + 82 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 538, + 82 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 538, + 82 + ], + "type": "text", + "content": "Table 2. Performance comparison on LiveBeauty, MEBeauty [24] and SCUT-FBP5500 [27]. Best in red and second-best in blue." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 91, + 550, + 268 + ], + "lines": [ + { + "bbox": [ + 61, + 91, + 550, + 268 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 550, + 268 + ], + "type": "table", + "html": "
AspectsMethodsLiveBeautyMEBeauty [24]SCUT-FBP5500 [27]
SROCC↑PLCC↑KROCC↑SROCC↑PLCC↑KROCC↑SROCC↑PLCC↑KROCC↑
BaselineViT-B [23]0.8970.8470.7330.6200.6590.4430.8690.8800.686
ResNeXt-50 [56]0.9070.8790.7460.7000.7430.5160.8990.9110.727
Inception-V3 [49]0.9060.8800.7450.6760.7270.4920.9050.9180.736
IAAAVA-MLSP [21]0.8330.8150.6500.6020.6430.4310.7640.7940.566
TANet [18]0.8910.8580.7280.6510.6900.4670.8700.8820.686
Dele-Trans [19]0.8880.8410.7210.6110.6510.4400.8700.8600.682
EAT [20]0.8930.8670.7280.6400.6970.4620.8530.8720.665
FAPComboNet [58]0.9020.8710.7410.6510.6920.4720.8970.9070.725
2D-FAP [32]0.8960.8540.7330.6900.7190.5060.9030.9150.734
REX-INCEP [2]0.9080.8840.7500.7020.7390.5140.9070.9170.739
CNN-ER [2]0.9130.8880.7570.7060.7480.5190.9100.9220.745
MEBeauty [24]0.8260.7960.6430.6490.6770.4670.7990.8020.599
Ours0.9240.8960.7730.7830.7950.5930.9300.9320.774
Improvement+1.1%+0.8%+1.6%+7.7%+4.7%+7.4%+2.0%+1.0%+2.9%
", + "image_path": "5eaf98d016f45e8364d6403673bd9a44b23da84fd0fe09add6599518303d8ec1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 287, + 136, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 287, + 136, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 287, + 136, + 301 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 308, + 175, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 308, + 175, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 175, + 319 + ], + "type": "text", + "content": "5.1. Benchmark Datasets" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 325, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 325, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 325, + 296, + 422 + ], + "type": "text", + "content": "The proposed LiveBeauty dataset and two open-source FAP datasets including SCUT-FBP5500 [27] and MEBeauty [24] are selected as benchmark datasets. The MOS distributions of them are shown in Fig. 7. We split SCUT-FBP5500 and MEBeauty with " + }, + { + "bbox": [ + 55, + 325, + 296, + 422 + ], + "type": "inline_equation", + "content": "60\\% - 40\\%" + }, + { + "bbox": [ + 55, + 325, + 296, + 422 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 325, + 296, + 422 + ], + "type": "inline_equation", + "content": "80\\% - 20\\%" + }, + { + "bbox": [ + 55, + 325, + 296, + 422 + ], + "type": "text", + "content": " train-test ratio separately, which is kept same with their own protocol, and LiveBeauty is split using the " + }, + { + "bbox": [ + 55, + 325, + 296, + 422 + ], + "type": "inline_equation", + "content": "90\\% - 10\\%" + }, + { + "bbox": [ + 55, + 325, + 296, + 422 + ], + "type": "text", + "content": " protocol." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 430, + 188, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 188, + 442 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 188, + 442 + ], + "type": "text", + "content": "5.2. Implementation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 447, + 295, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 295, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 295, + 519 + ], + "type": "text", + "content": "Model initialization. In MAEM, ViT-B/16 and GPT-2 are adopted as the image encoder and the text encoder separately, which are initialized by [42]. In PAPM, Swin-T is adopted as the trainable image encoder, where only four stages are reserved and initialized with parameters pretrained on multiple face-related tasks [41]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 520, + 295, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 295, + 591 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 295, + 591 + ], + "type": "text", + "content": "Hyper parameters. During the whole training process, AdamW is selected as the optimizer, and the learning rate scheduler is set with linear warm-up and cosine annealing scheme. The base learning rate varies with different training phase. Each training phase lasts 50 epochs, and the batch size is set as 32. Details can be found in the supplementary." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 599, + 187, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 187, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 187, + 612 + ], + "type": "text", + "content": "5.3. Competitors & Metrics" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": "To confirm the effectiveness of the proposed FPEM method, we select baseline models and several SOTA FAP and IAA methods, including ComboNet [58], 2D-FAP [32], REX-INCEP [2], CNN-ER [2], MEBeauty [24], AVAMLSP [21], TANet [18], Dele-Trans [19] and EAT [20]. All the models are retrained ourselves with the default parameters defined in their papers. Three criteria are adopted to evaluate the performance of these methods, i.e., Spear" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 321, + 291, + 391, + 361 + ], + "blocks": [ + { + "bbox": [ + 321, + 291, + 391, + 361 + ], + "lines": [ + { + "bbox": [ + 321, + 291, + 391, + 361 + ], + "spans": [ + { + "bbox": [ + 321, + 291, + 391, + 361 + ], + "type": "image", + "image_path": "28d2034fb63b1395875df6d64800e3e84f7cc87ef0b6ea83451044ba25a643b3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 365, + 382, + 375 + ], + "lines": [ + { + "bbox": [ + 331, + 365, + 382, + 375 + ], + "spans": [ + { + "bbox": [ + 331, + 365, + 382, + 375 + ], + "type": "text", + "content": "(a)LiveBeauty." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 399, + 291, + 470, + 360 + ], + "blocks": [ + { + "bbox": [ + 399, + 291, + 470, + 360 + ], + "lines": [ + { + "bbox": [ + 399, + 291, + 470, + 360 + ], + "spans": [ + { + "bbox": [ + 399, + 291, + 470, + 360 + ], + "type": "image", + "image_path": "66cb23b930b5d467e635259eac8fc61419aac3ff5d028a63c71dc4fa16887585.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 365, + 459, + 376 + ], + "lines": [ + { + "bbox": [ + 410, + 365, + 459, + 376 + ], + "spans": [ + { + "bbox": [ + 410, + 365, + 459, + 376 + ], + "type": "text", + "content": "(b) MEBeauty." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 477, + 291, + 547, + 360 + ], + "blocks": [ + { + "bbox": [ + 477, + 291, + 547, + 360 + ], + "lines": [ + { + "bbox": [ + 477, + 291, + 547, + 360 + ], + "spans": [ + { + "bbox": [ + 477, + 291, + 547, + 360 + ], + "type": "image", + "image_path": "cd149623c33da7a322f9c693cb927d6e0111a2a3bbef640be7776cc98cd3a04c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 479, + 365, + 545, + 375 + ], + "lines": [ + { + "bbox": [ + 479, + 365, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 479, + 365, + 545, + 375 + ], + "type": "text", + "content": "(c) SCUT-FBP5500." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 385, + 554, + 407 + ], + "lines": [ + { + "bbox": [ + 313, + 385, + 554, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 554, + 407 + ], + "type": "text", + "content": "Figure 8. Scatter plots of the FPEM predicted scores vs. MOSs on three benchmark datasets." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 413, + 554, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 554, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 554, + 449 + ], + "type": "text", + "content": "man Rank Order Correlation Coefficient (SROCC), Pearson Linear Correlation Coefficient (PLCC) and Kendall Rank Order Correlation Coefficient (KROCC)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 456, + 398, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 398, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 398, + 468 + ], + "type": "text", + "content": "5.4. Performance" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 474, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 689 + ], + "type": "text", + "content": "The experimental results on three FAP datasets are shown in Tab. 2, from which we could draw the following conclusions: (i) Our proposed method achieves the first place and surpasses the second place by about 0.011, 0.077, 0.020 in terms of SROCC values on LiveBeauty, MEBeauty and SCUT-FBP5500 respectively, which demonstrates its superiority. (ii) The IAA methods are inferior to the FAP methods, which manifests that the generic aesthetic assessment methods overlook the facial features involved in the subjective nature of facial attractiveness, leading to limited performance on FAP tasks. (iii) The performance of all methods drops on MEBeauty because the number of training samples is restricted and the faces are ethnically diverse in MEBeauty, indicating a large diversity in facial attractiveness. (iv) The significant improvement on MEBeauty over other datasets can be largely attributed to the enhancement brought by face priors, which is more effective in scenarios with insufficient training data." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "Fig. 8 presents scatter plots illustrating the relationship between the predicted attractiveness scores (vertical axis)" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11464" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 114, + 292, + 269 + ], + "blocks": [ + { + "bbox": [ + 55, + 60, + 295, + 106 + ], + "lines": [ + { + "bbox": [ + 55, + 60, + 295, + 106 + ], + "spans": [ + { + "bbox": [ + 55, + 60, + 295, + 106 + ], + "type": "text", + "content": "Table 3. Ablation study on LiveBeauty test set. " + }, + { + "bbox": [ + 55, + 60, + 295, + 106 + ], + "type": "inline_equation", + "content": "E_{FP}" + }, + { + "bbox": [ + 55, + 60, + 295, + 106 + ], + "type": "text", + "content": " denotes Cross-Attention with " + }, + { + "bbox": [ + 55, + 60, + 295, + 106 + ], + "type": "inline_equation", + "content": "f_p" + }, + { + "bbox": [ + 55, + 60, + 295, + 106 + ], + "type": "text", + "content": ", MS denotes Multi-Scale fusion operation, SR denotes Similarity Regression strategy, " + }, + { + "bbox": [ + 55, + 60, + 295, + 106 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_S" + }, + { + "bbox": [ + 55, + 60, + 295, + 106 + ], + "type": "text", + "content": " denotes the L1 scoring loss, " + }, + { + "bbox": [ + 55, + 60, + 295, + 106 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_R" + }, + { + "bbox": [ + 55, + 60, + 295, + 106 + ], + "type": "text", + "content": " denotes the total ranking loss " + }, + { + "bbox": [ + 55, + 60, + 295, + 106 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_R = \\mathcal{L}_{R_1} + \\mathcal{L}_{R_2}" + }, + { + "bbox": [ + 55, + 60, + 295, + 106 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 114, + 292, + 269 + ], + "lines": [ + { + "bbox": [ + 61, + 114, + 292, + 269 + ], + "spans": [ + { + "bbox": [ + 61, + 114, + 292, + 269 + ], + "type": "table", + "html": "
ModulesMethodsMetrics
SROCC↑PLCC↑KROCC↑
PAPMonlyEMS0.90450.87100.7454
w/oEFp0.90780.87140.7475
w/o MS0.90890.88080.7487
PAPM0.90960.88100.7497
MAEMzero-shot CLIP0.38380.35820.2692
w/o SR0.88250.82260.7123
w/oLS0.91610.84330.7607
w/oLR0.91860.89110.7653
MAEM0.91980.86310.7653
ALLw/o 2-stage0.91890.88890.7652
w/o SFM0.91970.89610.7641
all combined0.92430.89620.7732
", + "image_path": "817f1fab70160ebe7b1e1339d43c9dc47a8bff84aa6861481461e1af5f5145fd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 285, + 296, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 296, + 382 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 296, + 382 + ], + "type": "text", + "content": "obtained by FPEM and the MOSs (horizontal axis) in the test sets of three datasets. The MOSs of MEBeauty are divided by two to align the MOS range with others. The color of the scatter points from blue to red indicates the density of points from low to high. In each sub-figure, a second-order polynomial nonlinear fitting curve is superimposed. It can be observed that the fitted curves are close to the diagonal line, demonstrating the effectiveness of FPEM." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 392, + 149, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 149, + 404 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 149, + 404 + ], + "type": "text", + "content": "5.5. Ablation Study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 410, + 296, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 296, + 650 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 296, + 650 + ], + "type": "text", + "content": "We conduct elaborate ablation study to investigate the contributions of different modules in our proposed method, which is shown in Tab. 3. Conclusions can be drawn that: (i) In the ablation study of PAPM, we can observe that leveraging facial prior and multi-scale features are both helpful. (ii) The result of ablation study of MAEM indicates that the zero-shot CLIP has limited generalization ability in FAP tasks, which confirms its shortcoming in aesthetic evaluation [45]. Specifically, for MAEM w/o SR, we leverage concatenation of the visual features and the textual features with highest similarity, and utilize MLP for regression. We observe a significant performance drop without SR, which manifests that specifically devised regression strategy is important for CLIP to adapt to regression tasks. Meanwhile, the joint optimization of the ranking and regression also leads to the improvement of SROCC. (iii) In the ablation study of ALL modules, we can see that employing all modules yields the best performance, which confirms the effectiveness of using SFM to adaptively integrate multi-modal aesthetic features with holistic facial prior knowledge." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 659, + 174, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 174, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 174, + 672 + ], + "type": "text", + "content": "5.6. Cross-Dataset Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "content": "Considering the diversity of facial attractiveness perception, the cross-dataset validation is carried out to validate the generalization ability of these methods. Specifically," + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 318, + 103, + 550, + 268 + ], + "blocks": [ + { + "bbox": [ + 313, + 60, + 555, + 95 + ], + "lines": [ + { + "bbox": [ + 313, + 60, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 60, + 555, + 95 + ], + "type": "text", + "content": "Table 4. Cross-Dataset study trained on LiveBeauty and tested on the entire MEBeauty dataset and the entire SCUT-FBP5500 dataset. Best in red and second-best in blue." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 103, + 550, + 268 + ], + "lines": [ + { + "bbox": [ + 318, + 103, + 550, + 268 + ], + "spans": [ + { + "bbox": [ + 318, + 103, + 550, + 268 + ], + "type": "table", + "html": "
MethodsMEBeauty [24]SCUT-FBP5500 [27]
SROCC↑PLCC↑KROCC↑SROCC↑PLCC↑KROCC↑
ViT-B [23]0.5730.5830.4020.5530.5480.394
ResNeXt-50 [56]0.5350.5460.3720.4780.4670.335
Inception-V3 [49]0.5750.5810.4040.4840.4800.339
AVA-MLSP [21]0.5870.5970.4140.4000.4120.275
TANet [18]0.5510.5290.3840.4010.4110.277
Dele-Trans[19]0.5760.5720.4020.4800.4760.335
EAT[20]0.6030.6240.4270.4840.4660.338
ComboNet [58]0.5400.5480.3750.3840.3750.264
2D-FAP [32]0.5560.4890.3900.4440.4450.306
REX-INCEP [2]0.5880.5920.4130.4730.4710.332
CNN-ER [2]0.5890.5960.4140.5020.4990.352
MEBeauty [24]0.5200.5160.3610.3660.3390.251
FPEM (Ours)0.6300.6290.4470.5990.5770.434
Improvement+2.7%+0.5%+2.0%+4.6%+2.9%+4.0%
", + "image_path": "122ed23cc017e8d75824ba76220e7065a2fe64cf599d4428c017f21929be0854.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 333, + 310, + 533, + 357 + ], + "blocks": [ + { + "bbox": [ + 313, + 279, + 555, + 301 + ], + "lines": [ + { + "bbox": [ + 313, + 279, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 555, + 301 + ], + "type": "text", + "content": "Table 5. Cross-Dataset Ablation Study on each module. All models are trained on LiveBeauty and tested on the whole MEBeauty." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 333, + 310, + 533, + 357 + ], + "lines": [ + { + "bbox": [ + 333, + 310, + 533, + 357 + ], + "spans": [ + { + "bbox": [ + 333, + 310, + 533, + 357 + ], + "type": "table", + "html": "
PAPMMAEMSFMSROCC↑PLCC↑KROCC↑
✓0.61340.61410.4349
✓0.62290.62680.4244
✓✓✓0.62990.62910.4472
", + "image_path": "4e071a24c33ca62e0349d2f7f47b84ec55b558130e25fffb8fd9031ffdfd5f5d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 372, + 554, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 554, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 554, + 407 + ], + "type": "text", + "content": "all the mentioned models are trained on the proposed LiveBeauty dataset and subsequently tested on the entire SCUTFBP5500 and the entire MEBeauty." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 407, + 555, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 555, + 529 + ], + "type": "text", + "content": "As shown in Tab. 4, the proposed method outperforms all other methods on both two datasets, which demonstrates its strong generalization ability. Moreover, ablation results of the generalizability on MEBeauty are presented in Tab. 5. It can be observed that MAEM surpasses PAPM by leveraging multi-modal information. The generalizability is further improved by fusing their features through SFM, which indicates that integrating facial prior containing rich contextual and individual information is effective not only for improving performance but also for enhancing generalizability." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 538, + 388, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 388, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 388, + 550 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 558, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 715 + ], + "type": "text", + "content": "In this paper, we tackle the problem of facial attractiveness prediction in live videos that undergo face retouching. To address the existing gap in datasets, we introduce LiveBeauty, the first large-scale FAP dataset featuring beautiful face images for aesthetic purposes. This dataset is high-quality and authentic, collected from real-world environments, and is currently the largest open-access facial attractiveness dataset. We further propose FPEM, a multimodal FAP framework that selectively fuses holistic facial priors and aesthetic semantic features. Extensive experiments demonstrate the advantages of FPEM over state-of-the-art approaches. Meanwhile, LiveBeauty and FPEM can be effectively leveraged for FAP in other video applications." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11465" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "type": "text", + "content": "[1] Lacey Best-Rowden and Anil K Jain. Learning face image quality from human assessments. IEEE Transactions on Information Forensics and Security, 13(12):3064-3077, 2018. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 179 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 179 + ], + "type": "text", + "content": "[2] Fares Bougourzi, Fadi Dornaika, and Abdelmalik Taleb-Ahmed. Deep learning based face beauty prediction via dynamic robust losses and ensemble regression. Knowledge-Based Systems, 242:108246, 2022. 1, 2, 3, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 182, + 295, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 295, + 225 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 295, + 225 + ], + "type": "text", + "content": "[3] Djamel Eddine Boukhari, Ali Chemsa, and Riadh Ajgou. Facial beauty prediction based on vision transformer. International Journal of Electrical and Electronic Engineering & Telecommunications, pages 2319-2518, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 227, + 296, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 227, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 62, + 227, + 296, + 281 + ], + "type": "text", + "content": "[4] Fadi Boutros, Meiling Fang, Marcel Klemt, Biying Fu, and Naser Damer. Cr-fiaq: face image quality assessment by learning sample relative classifiability. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5836-5845, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 284, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 284, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 62, + 284, + 294, + 316 + ], + "type": "text", + "content": "[5] Kerang Cao, Kwang-nam Choi, Hoekyung Jung, and Lini Duan. Deep learning for facial beauty prediction. Information, 11(8):391, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 318, + 295, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 318, + 295, + 372 + ], + "spans": [ + { + "bbox": [ + 62, + 318, + 295, + 372 + ], + "type": "text", + "content": "[6] Nicolas Chahine, Stefania Calarasanu, Davide Garcia-Civiero, Theo Cayla, Sira Ferradans, and Jean Ponce. An image quality assessment dataset for portraits. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9968-9978, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 374, + 296, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 374, + 296, + 440 + ], + "spans": [ + { + "bbox": [ + 62, + 374, + 296, + 440 + ], + "type": "text", + "content": "[7] Wei-Ting Chen, Gurunandan Krishnan, Qiang Gao, SyYen Kuo, Sizhou Ma, and Jian Wang. Dsl-fqa: Assessing facial image quality via dual-set degradation learning and landmark-guided transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2931-2941, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 441, + 294, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 441, + 294, + 485 + ], + "spans": [ + { + "bbox": [ + 62, + 441, + 294, + 485 + ], + "type": "text", + "content": "[8] Ritendra Datta, Dhiraj Joshi, Jia Li, and James Z Wang. Studying aesthetics in photographic images using a computational approach. In Proceedings of the European Conference on Computer Vision, pages 288-301, 2006. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 487, + 294, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 487, + 294, + 518 + ], + "spans": [ + { + "bbox": [ + 62, + 487, + 294, + 518 + ], + "type": "text", + "content": "[9] Harm De Vries and Jason Yosinski. Can deep learning help you find the perfect match? arXiv preprint arXiv:1505.00359, 2015. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 520, + 295, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 520, + 295, + 553 + ], + "spans": [ + { + "bbox": [ + 57, + 520, + 295, + 553 + ], + "type": "text", + "content": "[10] Yubin Deng, Chen Change Loy, and Xiaou Tang. Image aesthetic assessment: An experimental survey. IEEE Signal Processing Magazine, 34(4):80-106, 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 555, + 295, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 555, + 295, + 609 + ], + "spans": [ + { + "bbox": [ + 57, + 555, + 295, + 609 + ], + "type": "text", + "content": "[11] Abhishek Dutta, Raymond Veldhuis, and Luuk Spreeuwers. A bayesian model for predicting face recognition performance using image quality. In Proceeding of the International Joint Conference on Biometrics, pages 1-8, 2014. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 611, + 295, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 611, + 295, + 643 + ], + "spans": [ + { + "bbox": [ + 57, + 611, + 295, + 643 + ], + "type": "text", + "content": "[12] Yael Eisenthal, Gideon Dror, and Eytan Ruppin. Facial attractiveness: Beauty and the machine. Neural Computation, 18(1):119-142, 2006. 1, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 646, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 646, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 646, + 295, + 689 + ], + "type": "text", + "content": "[13] Yang-Yu Fan, Shu Liu, Bo Li, Zhe Guo, Ashok Samal, Jun Wan, and Stan Z Li. Label distribution-based facial attractiveness computation by deep residual learning. IEEE Transactions on Multimedia, 20(8):2196-2208, 2017. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[14] Junying Gan, Xiaoshan Xie, Guohui He, and Heng Luo. Transbls: transformer combined with broad learning system" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "text", + "content": "for facial beauty prediction. Applied Intelligence, 53(21): 26110-26125, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "text", + "content": "[15] Junying Gan, Xiaoshan Xie, Yikui Zhai, Guohui He, Chaoyun Mai, and Heng Luo. Facial beauty prediction fusing transfer learning and broad learning system. Soft Computing, 27(18):13391-13404, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 140, + 553, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 193 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 193 + ], + "type": "text", + "content": "[16] Lian Gao, Weixin Li, Zehua Huang, Di Huang, and Yunhong Wang. Automatic facial attractiveness prediction by deep multi-task learning. In Proceedings of the IEEE International Conference on Pattern Recognition, pages 3592-3597, 2018. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 195, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 237 + ], + "type": "text", + "content": "[17] Douglas Gray, Kai Yu, Wei Xu, and Yihong Gong. Predicting facial beauty without landmarks. In Proceedings of the European Conference on Computer Vision, pages 434-447, 2010. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "text", + "content": "[18] Shuai He, Yongchang Zhang, Rui Xie, Dongxiang Jiang, and Anlong Ming. Rethinking image aesthetics assessment: Models, datasets and benchmarks. In Proceedings of the International Joint Conference on Artificial Intelligence, pages 942–948, 2022. 2, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 295, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 295, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 317, + 295, + 553, + 348 + ], + "type": "text", + "content": "[19] Shuai He, Anlong Ming, Yaqi Li, Jinyuan Sun, ShunTian Zheng, and Huadong Ma. Thinking image color aesthetics assessment: Models, datasets and benchmarks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21838-21847, 2023. 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "text", + "content": "[20] Shuai He, Anlong Ming, Shuntian Zheng, Haobin Zhong, and Huadong Ma. Eat: An enhancer for aesthetics-oriented transformers. In Proceedings of the ACM International Conference on Multimedia, pages 1023-1032, 2023. 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 394, + 553, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 553, + 446 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 553, + 446 + ], + "type": "text", + "content": "[21] Vlad Hosu, Bastian Goldlucke, and Dietmar Saupe. Effective aesthetics prediction with multi-level spatially pooled features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9375-9383, 2019. 2, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 449, + 553, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 553, + 513 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 553, + 513 + ], + "type": "text", + "content": "[22] Tharun J. Iyer, Rahul K, Ruban Nersisson, Zhemin Zhuang, Alex Noel Joseph Raj, and Imthiaz Refayee. Machine learning-based facial beauty prediction and analysis of frontal facial images using facial landmarks and traditional image descriptors. Computational Intelligence and Neuroscience, 2021(1):4423407, 2021. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 515, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 515, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 317, + 515, + 553, + 590 + ], + "type": "text", + "content": "[23] Alexander Kolesnikov, Alexey Dosovitskiy, Dirk Weissenborn, Georg Heigold, Jakob Uszkoreit, Lucas Beyer, Matthias Minderer, Mostafa Dehghani, Neil Houlsby, Sylvain Gelly, Thomas Unterthiner, and Xiaohua Zhai. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2021. 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 592, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 592, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 592, + 553, + 624 + ], + "type": "text", + "content": "[24] Irina Lebedeva, Yi Guo, and Fangli Ying. Mebeauty: a multi-ethnic facial beauty dataset in-the-wild. *Neural Computing and Applications*, pages 1–15, 2022. 1, 2, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "type": "text", + "content": "[25] Congcong Li, Andrew Gallagher, Alexander C. Loui, and Tsuhan Chen. Aesthetic quality assessment of consumer photos with faces. In Proceedings of the IEEE International Conference on Image Processing, pages 3221-3224, 2010. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[26] Xiang Li, Wenhai Wang, Xiaolin Hu, and Jian Yang. Selective kernel networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 510-519, 2019. 6" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11466" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "text", + "content": "[27] Lingyu Liang, Luojun Lin, Lianwen Jin, Duorui Xie, and Mengru Li. Scut-fbp5500: A diverse benchmark dataset for multi-paradigm facial beauty prediction. In Proceedings of the IEEE International Conference on Pattern Recognition, pages 1598-1603, 2018. 1, 2, 3, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 295, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 295, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 295, + 183 + ], + "type": "text", + "content": "[28] Luojun Lin, Lingyu Liang, Lianwen Jin, and Weijie Chen. Attribute-aware convolutional neural networks for facial beauty prediction. In Proceedings of the Information Joint Conference on Artificial Intelligence, pages 847-853, 2019. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 184, + 296, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 296, + 239 + ], + "type": "text", + "content": "[29] Qipeng Liu, Luojun Lin, Zhifeng Shen, and Yuanlong Yu. Fbpformer: Dynamic convolutional transformer for global-local-contextual facial beauty prediction. In Proceedings of the International Conference on Artificial Neural Networks, pages 223-235, 2023. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 239, + 295, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 295, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 295, + 282 + ], + "type": "text", + "content": "[30] Shu Liu, Yang-Yu Fan, Zhe Guo, Ashok Samal, and Afan Ali. A landmark-based data-driven approach on 2.5 d facial attractiveness computation. Neurocomputing, 238:168-178, 2017. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 284, + 296, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 284, + 296, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 284, + 296, + 337 + ], + "type": "text", + "content": "[31] Shu Liu, Bo Li, Yang-Yu Fan, Zhe Quo, and Ashok Samal. Facial attractiveness computation by label distribution learning with deep cnn and geometric features. In Proceedings of the IEEE International Conference on Multimedia and Expo, pages 1344-1349, 2017. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 338, + 295, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 295, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 295, + 381 + ], + "type": "text", + "content": "[32] Shu Liu, Enquan Huang, Yan Xu, Kexuan Wang, Xiaoyan Kui, Tao Lei, and Hongying Meng. Lightweight facial attractiveness prediction using dual label distribution. arXiv preprint arXiv:2212.01742, 2022. 1, 2, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 383, + 295, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 295, + 426 + ], + "type": "text", + "content": "[33] Shu Liu, Enquan Huang, Yan Xu, Kexuan Wang, Xiaoyan Kui, Tao Lei, and Hongying Meng. Lightweight facial attractiveness prediction using dual label distribution. arXiv preprint arXiv:2212.01742, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 426, + 295, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 295, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 295, + 470 + ], + "type": "text", + "content": "[34] Tie Liu, Shengxi Li, Mai Xu, Li Yang, and Xiaofei Wang. Assessing face image quality: A large-scale database and a transformer method. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 471, + 295, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 295, + 525 + ], + "type": "text", + "content": "[35] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 526, + 295, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 295, + 570 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 295, + 570 + ], + "type": "text", + "content": "[36] Naila Murray, Luca Marchesotti, and Florent Perronnin. Ava: A large-scale database for aesthetic visual analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2408-2415, 2012. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 571, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 571, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 571, + 295, + 635 + ], + "type": "text", + "content": "[37] Fu-Zhao Ou, Xingyu Chen, Ruixin Zhang, Yuge Huang, Shaoxin Li, Jilin Li, Yong Li, Liujuan Cao, and Yuan-Gen Wang. Sdd-fqa: Unsupervised face image quality assessment with similarity distribution distance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7670–7679, 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 637, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 637, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 637, + 295, + 690 + ], + "type": "text", + "content": "[38] Fu-Zhao Ou, Chongyi Li, Shiqi Wang, and Sam Kwong. Clib-fiqa: Face image quality assessment with confidence calibration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1694-1704, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "text", + "content": "[39] Tianhao Peng, Mu Li, Fangmei Chen, Yong Xu, and David Zhang. Geometric prior guided hybrid deep neural network" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "for facial beauty analysis. CAAI Transactions on Intelligence Technology, 9(2):467-480, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 163 + ], + "type": "text", + "content": "[40] P Jonathon Phillips, J Ross Beveridge, David S Bolme, Bruce A Draper, Geof H Givens, Yui Man Lui, Su Cheng, Mohammad Nayeem Teli, and Hao Zhang. On the existence of face quality measures. In Proceedings of the IEEE International Conference on Biometrics: Theory, Applications and Systems, pages 1-8, 2013. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 164, + 553, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 164, + 553, + 219 + ], + "spans": [ + { + "bbox": [ + 316, + 164, + 553, + 219 + ], + "type": "text", + "content": "[41] Lixiong Qin, Mei Wang, Chao Deng, Ke Wang, Xi Chen, Jiani Hu, and Weihong Deng. Swinface: a multi-task transformer for face recognition, expression recognition, age estimation and attribute estimation. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 220, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 285 + ], + "type": "text", + "content": "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proceedings of the International Conference on Machine Learning, pages 8748-8763, 2021. 2, 4, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 287, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 341 + ], + "type": "text", + "content": "[43] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 815-823, 2015. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 342, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 376 + ], + "type": "text", + "content": "[44] B Series. Methodology for the subjective assessment of the quality of television pictures. Recommendation ITU-R BT, 500(13), 2012. 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 377, + 555, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 555, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 555, + 442 + ], + "type": "text", + "content": "[45] Xiangfei Sheng, Leida Li, Pengfei Chen, Jinjian Wu, Weisheng Dong, Yuzhe Yang, Liwu Xu, Yaqian Li, and Guangming Shi. Aesclip: Multi-attribute contrastive learning for image aesthetics assessment. In Proceedings of the ACM International Conference on Multimedia, page 1117–1126, 2023. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 445, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 553, + 499 + ], + "type": "text", + "content": "[46] Shengjie Shi, Fei Gao, Xuantong Meng, Xingxin Xu, and Jingjie Zhu. Improving facial attractiveness prediction via co-attention learning. In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing, pages 4045–4049, 2019. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 500, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 553, + 555 + ], + "type": "text", + "content": "[47] Shaolin Su, Hanhe Lin, Vlad Hosu, Oliver Wiedemann, Jinqiu Sun, Yu Zhu, Hantao Liu, Yanning Zhang, and Dietmar Saupe. Going the extra mile in face image quality assessment: A novel database and model. IEEE Transactions on Multimedia, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 556, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 556, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 556, + 553, + 623 + ], + "type": "text", + "content": "[48] Wei Sun, Weixia Zhang, Yanwei Jiang, Haoning Wu, Zicheng Zhang, Jun Jia, Yingjie Zhou, Zhongpeng Ji, Xiongkuo Min, Weisi Lin, and Zhai Guangtao. Dual-branch network for portrait image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 624, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 679 + ], + "type": "text", + "content": "[49] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2826, 2016. 1, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[50] Philipp Terhorst, Jan Niklas Kolf, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. Ser-fiq: Unsupervised estimation of face image quality based on stochastic embedding" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11467" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 105 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 105 + ], + "type": "text", + "content": "robustness. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5651-5660, 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 129 + ], + "type": "text", + "content": "[51] Louis L Thurstone. A law of comparative judgment. In *Scaling*, pages 81–92. Routledge, 2017. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 130, + 296, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 296, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 296, + 174 + ], + "type": "text", + "content": "[52] Yu Tian, Zhangkai Ni, Baoliang Chen, Shiqi Wang, Hanli Wang, and Sam Kwong. Generalized visual quality assessment of gan-generated face images. arXiv preprint arXiv:2201.11975, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 295, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 295, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 295, + 207 + ], + "type": "text", + "content": "[53] Song Tong, Xuefeng Liang, Takatsune Kumada, and Sunao Iwaki. Putative ratios of facial attractiveness in a deep neural network. Vision Research, 178:86-99, 2021. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 209, + 295, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 295, + 262 + ], + "type": "text", + "content": "[54] Ming-Feng Tsai, Tie-Yan Liu, Tao Qin, Hsin-Hsi Chen, and Wei-Ying Ma. Frank: a ranking method with fidelity loss. In Proceedings of the ACM SIGIR Conference on Research and Development in Information Retrieval, pages 383-390, 2007. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 264, + 295, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 295, + 308 + ], + "type": "text", + "content": "[55] Qinjie Xiao, You Wu, Dinghong Wang, Yong-Liang Yang, and Xiaogang Jin. Beauty3dfacenet: deep geometry and texture fusion for 3d facial attractiveness prediction. Computers & Graphics, 98:11-18, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 310, + 295, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 295, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 295, + 364 + ], + "type": "text", + "content": "[56] Saining Xie, Ross Girshick, Piotr Dolkar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1492-1500, 2017. 1, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 365, + 295, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 295, + 430 + ], + "type": "text", + "content": "[57] Jie Xu, Lianwen Jin, Lingyu Liang, Ziyong Feng, Duorui Xie, and Huiyun Mao. Facial attractiveness prediction using psychologically inspired convolutional neural network (pCNN). In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing, pages 1657-1661, 2017. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 433, + 295, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 433, + 295, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 433, + 295, + 465 + ], + "type": "text", + "content": "[58] Lu Xu and Jinhai Xiang. Comboloss for facial attractiveness analysis with squeeze-and-excitation networks. arXiv preprint arXiv:2010.10721, 2020. 2, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 467, + 295, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 295, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 295, + 499 + ], + "type": "text", + "content": "[59] Lu Xu and Jinhai Xiang. Comboloss for facial attractiveness analysis with squeeze-and-excitation networks. arXiv preprint arXiv:2010.10721, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 501, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 295, + 544 + ], + "type": "text", + "content": "[60] Lu Xu, Heng Fan, and Jinhai Xiang. Hierarchical multi-task network for race, gender and facial attractiveness recognition. In Proceedings of the IEEE International Conference on Image Processing, pages 3861-3865, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 545, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 295, + 578 + ], + "type": "text", + "content": "[61] Haibin Yan. Cost-sensitive ordinal regression for fully automatic facial beauty assessment. Neurocomputing, 129:334-342, 2014. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 579, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 295, + 633 + ], + "type": "text", + "content": "[62] Qinkai Yu, Jianyang Xie, Anh Nguyen, He Zhao, Jiong Zhang, Huazhu Fu, Yitian Zhao, Yalin Zheng, and Yanda Meng. Clip-dr: Textual knowledge-guided diabetic retinopathy grading with ranking-aware prompting. arXiv preprint arXiv:2407.04068, 2024. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 635, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 295, + 689 + ], + "type": "text", + "content": "[63] Yikui Zhai, Yu Huang, Ying Xu, Junying Zeng, Fei Yu, and Junying Gan. Benchmark of a large scale database for facial beauty prediction. In Proceedings of the International Conference on Intelligent Information Processing, pages 1-5, 2016. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "text", + "content": "[64] Yikui Zhai, Yu Huang, Ying Xu, Junying Gan, He Cao, Wenbo Deng, Ruggero Donida Labati, Vincenzo Piuri, and" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 228 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "text", + "content": "Fabio Scotti. Asian female facial beauty prediction using deep neural networks via transfer learning and multi-channel feature fusion. IEEE Access, 8:56892-56907, 2020. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 107, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 160 + ], + "type": "text", + "content": "[65] Shifeng Zhang, Xiangyu Zhu, Zhen Lei, Hailin Shi, Xiaobo Wang, and Stan Z Li. Faceboxes: A cpu real-time face detector with high accuracy. In Proceedings of the IEEE International Joint Conference on Biometrics, pages 1-9, 2017. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 163, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 228 + ], + "type": "text", + "content": "[66] Weixia Zhang, Guangtao Zhai, Ying Wei, Xiaokang Yang, and Kede Ma. Blind image quality assessment via vision-language correspondence: A multitask learning perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14071-14081, 2023. 5" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "11468" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/FREE-Merging_ Fourier Transform for Efficient Model Merging/f3aa67c5-53d3-463b-8730-22ad4f95fe7f_content_list.json b/2025/FREE-Merging_ Fourier Transform for Efficient Model Merging/f3aa67c5-53d3-463b-8730-22ad4f95fe7f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7018c3443a5b8503d7da74f7a05a5a6e4e67cd38 --- /dev/null +++ b/2025/FREE-Merging_ Fourier Transform for Efficient Model Merging/f3aa67c5-53d3-463b-8730-22ad4f95fe7f_content_list.json @@ -0,0 +1,1759 @@ +[ + { + "type": "text", + "text": "FREE-Merging: Fourier Transform for Efficient Model Merging", + "text_level": 1, + "bbox": [ + 171, + 130, + 826, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shenghe Zheng \nHarbin Institute of Technology \nshenghez.zheng@gmail.com", + "bbox": [ + 225, + 181, + 470, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hongzhi Wang* \nHarbin Institute of Technology \nwangzh@hit.edu.cn", + "bbox": [ + 526, + 181, + 771, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the rapid growth of deep learning, there is an increasing availability of open-source models for various tasks. However, single fine-tuned models often fall short of meeting the diverse needs of users. Model merging has thus emerged as an efficient method to integrate the capabilities of existing models into a unified model. Nevertheless, existing model merging methods face challenging trade-offs between performance and deployment costs, primarily due to task interference. For the first time, we reveal that task interference is evident in the frequency domain of model parameters, yet current efforts only focus on spatial domain solutions, which are largely ineffective in addressing frequency domain interference. To mitigate the impact of frequency domain interference, we propose FR-Merging, an innovative method that effectively filters harmful frequency domain interference on the backbone with minimal computational overhead. Since performance loss is inevitable with cost-free methods, we propose a lightweight task-specific expert module that dynamically compensates for information loss during merging. This proposed framework, FREE-Merging (FR-Merging with experts), strikes a balanced trade-off between training cost, inference latency, storage requirements, and performance. We demonstrate the effectiveness of both FR-Merging and FREE-Merging on multiple tasks across CV, NLP, and Multi-Modal domains and show that they can be flexibly adapted to specific needs. Our code is available at: https://github.com/Zhengsh123/FREE-Merging.", + "bbox": [ + 91, + 301, + 483, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 756, + 220, + 772 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the current era of deep learning, the pretrain-finetune paradigm is a standard procedure [9, 37, 40]. However, edge applications often lack powerful computing resources and data, favoring ready-to-use models for specific scenarios. The recent growth of open-source platforms like Hugging-Face [43] has made this feasible. Yet, practical demands", + "bbox": [ + 89, + 782, + 482, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "are often not fully covered by a single existing model and are typically a union of several models [23, 52]. However, running separate models for each subtask increases storage and inference costs, especially with large models. Model merging attempts to combine existing models to build a model capable of addressing all target tasks [21, 33, 48]. This method reduces training costs, data privacy issues in multi-task learning (MTL) [20, 56] and deployment costs, garnering widespread attention.", + "bbox": [ + 511, + 268, + 906, + 405 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While model merging aims to reduce training costs, it often suffers from performance loss due to task interference [20]. To mitigate this issue, existing methods develop along three directions. The first direction computes merging coefficients based on weights or data [23, 33]. The second uses the parameter differences before and after finetuning (i.e. task vectors) for merging [21, 48, 52, 58]. The third method introduces independent knowledge for each task [20, 29]. While the third way balances performance and costs, it still faces two critical limitations. First, it neglects backbone optimization during merging, resulting in subpar performance. Second, it requires high computation and storage, making it impractical for edge deployment.", + "bbox": [ + 511, + 409, + 908, + 606 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We argue that an effective model merging method requires addressing both of the above issues. First, since the backbone constitutes the majority of parameters and thus determines performance, an efficient backbone merging method is essential to reduce performance loss. Then, given the no free lunch theorem, cost-free merging cannot avoid task interference. Therefore, introducing task-specific experts is necessary. However, they must remain lightweight to avoid storage and inference overhead. Thus, we propose FREE-Merging, a two-stage method that ensures an efficient backbone and lightweight experts as shown in Fig 1.", + "bbox": [ + 511, + 609, + 908, + 776 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "For backbone merging, the key lies in eliminating task interference [51]. We identify a compelling phenomenon in which task interference is evident in the frequency domain of model parameters. And we present in Sec. 4 that frequency-domain interference is strongly correlated with performance. Notably, existing methods, which focus only on spatial-domain operations, fail to address frequency-domain interference, leading to suboptimal outcomes. Ex", + "bbox": [ + 511, + 779, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author", + "bbox": [ + 109, + 886, + 238, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3863", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c1b7af58661b1fa1c1d29f024989d2bbdfd80d0cb91e6a2bd8d7578b6f72c1f7.jpg", + "image_caption": [ + "Figure 1. The workflow of FREE-Merging involves two main steps. First, FR-Merging utilizes high-pass filtering to remove harmful specialized information from each model, thereby constructing a high-quality merged backbone. Second, it employs lightweight extraction of task experts, which are dynamically added during inference to mitigate the impact of task interference." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 315, + 282 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7394ca043757c5a26b29325fd0558863920099268d48ac200488707e554e3b53.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 318, + 88, + 661, + 284 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/53aa4ab86d7788d51c11954ead466d1c0288307f670260c7616f7550dd37db05.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 89, + 905, + 284 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experimental evidence reveals that task interference is severe in the low-frequency region. This occurs because, high-frequency signals represent fine-grained variations, while low-frequency parts capture the global structure [18], which is more likely to contain task-specific information that leads to task interference. Leveraging the inherent redundancy in fine-tuned parameters [27, 47, 53], we propose to directly filter out low-frequency parts with severe task interference. As detailed in Sec. 5.2, this high-pass filtering substantially enhances generalization with minimal performance loss, preventing task interference during merging. Our lightweight yet highly effective approach, FR-Merging, is the first to apply Fourier filtering to neural network parameters, showing its potential for broader applications.", + "bbox": [ + 88, + 364, + 482, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "After obtaining a high-performance backbone through FR-Merging, we use a lightweight task-specific expert to recover the inevitable information loss with minimal storage. Unlike existing methods that store extensive knowledge, we rescale low-frequency signals, which store task-specific information, retaining task expertise with only $1\\%$ of the parameters. Inspired by MoE [39, 57], we employ a router for dynamic expert routing, enhancing model flexibility.", + "bbox": [ + 88, + 580, + 482, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, we introduce FREE-Merging, which utilizes FR-Merging to reduce task interference and incorporates lightweight experts to eliminate information loss of model merging. Our main contributions are: 1). We identify frequency-domain task interference as a critical factor that greatly impacts merging performance but remains challenging to address through spatial-domain methods. 2). To tackle frequency-domain interference, we propose FR-Merging, which reduces interference while preserving performance, constructing the merged backbone. 3). We theoretically validate the necessity of introducing new information during merging and propose an effective expert construction method. 4). Extensive experiments show that", + "bbox": [ + 89, + 704, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "FR-Merging and FREE-Merging effectively enhance performance across vision, language, and multi-modal tasks.", + "bbox": [ + 511, + 364, + 903, + 396 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 407, + 653, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Model Merging. Model merging integrates existing models to handle multiple tasks without training [21, 23, 49], but faces task interference challenges [36, 48]. Simple averaging [44] causes great performance degradation. Methods like Fisher-Merging [33] and RegMean [23] use matrices to determine merging coefficients but are computationally expensive or data-intensive, limiting edge deployment. Task Arithmetic [21] merges task vectors instead of weights, while Ties-Merging [48] resolves parameter conflicts, and AdaMerging [52] automates coefficient selection. However, these approaches have data distribution requirements and limited applicability. Techniques such as DARE [53] and PCB-Merging [12] mitigate performance drops but still face task conflicts. EMR-Merging [20] and Twin-Merging [29] store task-specific knowledge but require large storage and neglect backbone optimization. Our method uses high-pass filtering and lightweight experts to balance merging costs and effectiveness across various tasks. For further details, please refer to Appendix C.2.", + "bbox": [ + 511, + 431, + 906, + 719 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Model Ensemble. Model ensemble combines outputs from multiple models [11, 15, 22], but with large models, it faces storage and inference challenges. In contrast, Model merging enables a single model to solve multiple tasks, significantly reducing storage and inference costs.", + "bbox": [ + 511, + 719, + 903, + 794 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-task Learning. Multi-task learning (MTL) aims to solve multiple tasks using a single model [4, 28, 41], but in the era of large models, it faces challenges like high training costs, data privacy issues, and expertise requirements [52]. In contrast, model merging uses existing open-source models to create a multi-task model with little or no training, significantly reducing deployment costs.", + "bbox": [ + 511, + 795, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3864", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c735b87a90dd98bf2bf51f5bae2e5f1120d9374a56864f33f8f381651c3d0e21.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 93, + 90, + 480, + 200 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ab7a92b21a70607cbd9d59de2649a84f37f062b33d1e32d56d05aaf436854fc8.jpg", + "image_caption": [ + "(a)", + "(b)", + "Figure 2. a) The frequency domain amplitude power distribution of an attention head across different fine-tuned ViT-B-32 models. b) The correlation between the merging performance of ViT-B-32 and the variance of frequency domain amplitude power. Normalized accuracy is defined as the ratio of merged model performance to fine-tuned model performance on the same task." + ], + "image_footnote": [], + "bbox": [ + 93, + 212, + 480, + 323 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 450, + 230, + 467 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Notation. Assume that $f_{\\theta}(x) \\to y$ represents a neural network parameterized by $\\theta$ , where $x \\in R^{m}$ is the input and $y \\in R^{n}$ is the output. Recent studies have focused on model merging via task vectors. For task $k$ , the task vector is defined as $v_{k} = \\theta_{k} - \\theta_{pre}$ , where $\\theta_{k}$ is the fine-tuned model for task $k$ , and $\\theta_{pre}$ is the corresponding pre-trained model. Additionally, $e_{k}$ represents the task expert that retains the ability to solve the task $k$ with few parameters.", + "bbox": [ + 89, + 477, + 482, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem Definition. The objective of model merging is to combine models $\\{\\theta_k\\}_{k=1}^K$ for tasks $T = \\{t_i\\}_{i=1}^K$ into a single model $\\theta_m$ that can handle all tasks in $T$ . We focus on a practical scenario where a shared pre-trained model $\\theta_{pre}$ serves as the base for all fine-tuned models. It is formulated as $\\theta_m = \\lambda \\sum_{k=1}^K \\hat{G}(\\theta_k)$ , and with task vector, it becomes $\\theta_m = \\theta_{pre} + \\lambda \\Sigma_{k=1}^K G(v_k)$ , where $\\lambda$ is the coefficient, and $\\hat{G}/G$ is transformation on the parameters. Assuming a shared pre-trained model aligns with trends toward foundation models, like LLaMa [40] for language and CLIP [37] for vision, which are then fine-tuned for specific tasks.", + "bbox": [ + 89, + 599, + 482, + 767 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Task Interference in Frequency Domain", + "text_level": 1, + "bbox": [ + 89, + 782, + 449, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Task interference is a key factor that leads to the decrease in merging performance. While current methods focus on addressing interference in the spatial domain [48, 50, 53], we observe that task vectors from different fine-tuned models also exhibit great misalignment in the frequency domain.", + "bbox": [ + 89, + 809, + 482, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 2(a), we visualize the amplitude power", + "bbox": [ + 106, + 885, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/3ef97cf4dde10b5f478606464124ebaa33a7c12d746f6e3ed6b686c98c93b412.jpg", + "table_caption": [ + "Table 1. The impact of current methods on the mean variance of frequency-domain amplitude across 8 image classification tasks." + ], + "table_footnote": [], + "table_body": "
MethodViT-B/32ViT-L/14
Task Arithmetic [21]0.0590.110
DARE [53]0.057(↓3%)0.108(↓2%)
Ties-Merging [48]0.058(↓2%)0.109(↓1%)
Breadcrumbs [7]0.058(↓2%)0.107(↓3%)
PCB-Merging [12]0.056(↓5%)0.107(↓3%)
FR-Merging(ours)0.045(↓24%)0.088(↓20%)
", + "bbox": [ + 540, + 127, + 879, + 228 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "distribution in the frequency domain of task vectors from a ViT-B-32 attention head fine-tuned on various tasks [21]. Here, greater distance from the center represents a higher frequency. Since amplitude indicates signal strength, merging two signals with vastly different strengths inevitably results in the weaker signal being masked. Therefore, task interference in the low-frequency part are severe.", + "bbox": [ + 511, + 250, + 903, + 356 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Furthermore, in Fig.2(b), we train multiple groups of models on eight image tasks [21] and find a clear negative correlation between the variance of amplitude power among models and the normalized merging performance (details can be found in Appendix D.1), confirming the great impact of frequency-domain interference on merging performance.", + "bbox": [ + 511, + 356, + 905, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Additionally, as shown in Table 1, we show that existing works addressing task interference in the spatial domain fail to effectively mitigate frequency-domain task interference. Therefore, we propose our approach to tackle this issue.", + "bbox": [ + 511, + 446, + 903, + 507 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "5. Methodology", + "text_level": 1, + "bbox": [ + 513, + 518, + 648, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section presents FREE-Merging: Sec. 5.1 explains the need for a two-part design, Sec. 5.2 and Sec. 5.3 cover the FR-Merging for the backbone and the lightweight expert module, respectively, followed by the FREE-Merging in Sec. 5.4 and a complexity analysis in Sec. 5.5.", + "bbox": [ + 511, + 544, + 905, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "5.1. Motivation", + "text_level": 1, + "bbox": [ + 511, + 628, + 635, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first analyze why both merging optimization for the backbone and the introduction of experts are necessary.", + "bbox": [ + 511, + 650, + 903, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "First, backbone optimization is crucial but overlooked by current expert-based merging ways. As the backbone holds most parameters, its performance is key to the overall success. For instance, upgrading from LLaMa2 [40] to LLaMa3 [13] greatly improves results even with the identical fine-tuning. Thus, we propose a training-free merging method to reduce task interference in the backbone.", + "bbox": [ + 511, + 680, + 905, + 784 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Next, we present a theorem showing the necessity of task experts. Based on Theorem 5.1, introducing additional information is essential. For proof, see Appendix B, and the effectiveness of task experts is proved in Appendix D.4.", + "bbox": [ + 511, + 786, + 905, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Theorem 5.1 Model Merging has no free lunch. Merged model $\\theta_{m}$ cannot simultaneously retain the capabilities of $\\{\\theta_k\\}_{k=1}^K$ without introducing additional information.", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3865", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "5.2. FR-Merging", + "text_level": 1, + "bbox": [ + 89, + 90, + 225, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose an effective backbone merging method by addressing task interference in the low-frequency fine-tuned parameters as discussed in Sec. 4, which is hard to resolve in the spatial domain. Given the high redundancy of fine-tuned parameters [53], we apply high-pass filtering to directly remove the interfering regions. Let $v(x,y)$ be the task vector and $G(x,y)$ the transformed result. Then we have:", + "bbox": [ + 89, + 112, + 483, + 218 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nG (x, y) = \\mathcal {F} ^ {- 1} \\{H (\\eta , \\gamma) \\cdot \\mathcal {F} \\{v (x, y) \\} \\}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 224, + 482, + 242 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 89, + 250, + 137, + 263 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nH (\\eta , \\gamma) = \\left\\{ \\begin{array}{l l} 1, & \\sqrt {\\eta^ {2} + \\gamma^ {2}} \\geq D _ {0} \\\\ 0, & \\sqrt {\\eta^ {2} + \\gamma^ {2}} < D _ {0} \\end{array} \\right., \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 268, + 482, + 308 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{F}$ and $\\mathcal{F}^{-1}$ represent the Fourier and inverse Fourier transform, respectively. $D_0$ is the adjustable cutoff frequency. And $\\eta$ and $\\gamma$ are the distances from the frequency domain center along the $x$ - and $y$ -axes.", + "bbox": [ + 89, + 316, + 483, + 376 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next, we analyze why high-pass filtering is effective from both intuitive and experimental perspectives. Fine-tuned weights occupy distinct positions in the loss landscape, and linear interpolation often leads to high-loss regions, causing task interference [1, 50]. To address this, we minimize model differences while preserving performance, bringing them closer in the loss landscape to increase the likelihood of the merged model falling into a loss basin [45].", + "bbox": [ + 89, + 377, + 483, + 496 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Results in Sec. 4 show great differences in low-frequency regions. This is because, if we consider neural network parameters as signals carrying training information, inspired by image processing [3, 35], high-frequency signals represent sharp variations, while low-frequency parts define the overall structure [5]. Since most information resides in the structure, filtering out low-frequency signals removes key task-specific information. Then, the likelihood that the interpolation falls within the loss basin increases [1].", + "bbox": [ + 89, + 497, + 483, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Experimentally, we validate that high-pass filtering enhances model generalization with minimal performance loss as shown in Fig 3. We fine-tune ViT-B/16 [37] on 30 visual tasks following [20]. We find that removing low-frequency information results in a smaller performance drop (diagonal) compared to the improvement in generalization (off-diagonal). This indicates that low-frequency parts carry task-specific information that boosts task performance but reduces generalization. Thus, applying filtering like FR-Merging helps alleviate task interference during merging.", + "bbox": [ + 89, + 633, + 483, + 784 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To maintain consistent output after merging, we define the merging coefficient $\\lambda_{i}$ for the task vector $v_{i}$ as:", + "bbox": [ + 89, + 785, + 483, + 814 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\lambda_ {i} = \\mathbb {E} (v _ {i}) \\left(\\sum_ {j = 1} ^ {K} \\mathbb {E} (v _ {j})\\right) ^ {- 1}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 821, + 482, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbb{E}$ represents the average. Overall, we construct a merged backbone in this part that minimizes task conflicts.", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/66191ca272013351a4299d6a43952fbf48fa82abb40d63ac0e1a23a9b84c2275.jpg", + "image_caption": [ + "Figure 3. High-pass filtering improves performance over the unfiltered version. Rows represent the added task vectors, while columns are the test datasets. The results show a significant boost in generalization with only a slight decline in task performance." + ], + "image_footnote": [], + "bbox": [ + 535, + 90, + 879, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.3. Lightweight Expert Extraction", + "text_level": 1, + "bbox": [ + 511, + 395, + 785, + 411 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To compensate for information lost due to discarding low-frequency signals, we utilize lightweight experts added during inference. As analyzed in Sec. 5.2, low-frequency signals contain task-specific information, but preserving them requires an inefficient inverse Fourier transform during each inference. Thus, we propose a more efficient way.", + "bbox": [ + 511, + 417, + 905, + 507 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since fine-tuning leads to task specialization, the most changed parameters can be approximated as having the greatest impact [32]. After reasonable rescaling, these selected parameters serve as task experts, minimizing storage (typically around $1\\%$ of parameters) without extra inference computation. The validity is detailed in Appendix D.4.", + "bbox": [ + 511, + 507, + 905, + 598 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although magnitude-based extraction [48] exists, limited research on rescaling hinders compression. As rescaling is crucial to expert extraction, we propose a rescaling method. Let $M(v_{i},d)$ represent the top- $d$ ( $d$ is a percentage) parameters for task vector $v_{i}$ , $\\mu_{i}$ be the rescale factor, and $e(v_{i})$ be the expert for task $t_{i}$ . Then we have:", + "bbox": [ + 511, + 598, + 905, + 688 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ne \\left(v _ {i}\\right) = \\mu_ {i} M \\left(v _ {i}, d\\right), \\mu_ {i} = - \\frac {\\mathbb {E} \\left(M \\left(v _ {i} , d\\right)\\right) \\cdot \\log (d)}{\\lambda_ {i} \\cdot \\mathbb {E} \\left(v _ {i}\\right)}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 695, + 903, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The goal is to maintain a consistent output after extraction to preserve performance. Details are in Appendix D.5.", + "bbox": [ + 511, + 734, + 905, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In summary, we efficiently extract task-specific experts to integrate into the backbone, mitigating task conflicts without excessive storage or inference overhead.", + "bbox": [ + 511, + 765, + 905, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.4. FREE-Merging", + "text_level": 1, + "bbox": [ + 511, + 816, + 669, + 834 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This section presents FREE-Merging, combining both parts to make it suitable for multiple tasks without affecting inference speed. Inspired by MoE [39], a router dynamically assigns tasks to activated experts. The router can take", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3866", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/88035de48d64a604e7638c0e6834371651ba0277cff9739898a1e8005c2830cd.jpg", + "table_caption": [ + "Table 2. Comparison of model merging performance using ViT-B/32 (B) and ViT-L/14 (L) on eight visual tasks evaluating by accuracy." + ], + "table_footnote": [], + "table_body": "
MethodBase ModelSUN397CarsRESISC45EuroSATSVHNGTSRBMNISTDTDAvg.
BLBLBLBLBLBLBLBLBL
Individual75.382.377.792.496.197.499.710097.598.198.799.299.799.779.484.190.594.2
Traditional MTL73.980.874.490.693.996.398.296.395.897.698.999.199.599.677.984.488.993.5
Weight Averaging65.372.163.481.671.482.671.791.964.278.252.870.787.597.150.162.865.879.6
Fisher Merging [33]68.669.269.288.670.787.566.493.572.980.651.174.887.993.359.970.068.382.2
Task Arithmetic [21]63.874.162.182.172.086.777.693.874.487.965.186.894.098.952.265.670.184.5
Ties-Merging [48]64.876.562.985.074.389.378.995.783.190.371.483.397.699.056.268.873.686.0
Breadcrumbs [7]63.074.961.084.975.788.684.795.483.489.876.490.198.199.257.968.275.086.4
PCB-Merging [12]65.575.864.186.078.188.680.296.084.788.077.190.998.099.158.470.075.886.9
FR-Merging(ours)66.276.464.587.077.290.290.196.885.492.082.392.898.599.360.071.578.188.3
AdaMerging++ [52]66.679.468.390.382.291.694.297.489.693.489.097.598.399.060.679.281.191.0
Twin-Merging [29]73.682.671.790.392.194.999.399.495.396.797.298.099.199.474.080.187.892.7
EMR-Merging [20]74.182.372.790.891.995.399.499.595.896.996.998.199.199.572.180.187.792.8
FREE-Merging(ours)77.183.578.292.493.496.299.599.696.398.198.298.799.599.775.481.789.793.7
", + "bbox": [ + 94, + 109, + 906, + 310 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 Workflow of FREE-Merging", + "text_level": 1, + "bbox": [ + 93, + 334, + 372, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input: Task vectors $\\{v_{i}\\}_{i = 1}^{K}$ , pre-trained model $\\theta_{pre}$ , router $R$ , and input data $X$ .", + "bbox": [ + 93, + 354, + 480, + 385 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: FR-Merging:", + "$\\triangleright$ Only excute once.", + "2: Get backbone $\\theta_{m} = \\theta_{pre} + \\sum_{i=1}^{K} \\lambda_{i} G(v_{i})$ , where $G$ and $\\lambda$ are defined by Eq. 1 and Eq. 3, respectively.", + "3: Expert Extraction:", + "$\\triangleright$ Only execute once.", + "4: Get task experts $\\{e_i\\}_{i = 1}^K$ as Eq. 4.", + "5: Inference:", + "6: for $x \\in X$ do", + "7: $[w_{1},\\dots w_{K}]\\gets argmax(R(x))$", + "8: $\\theta_{*} = \\theta_{m} + \\sum_{i = 1}^{K}w_{i}e_{i}$", + "9: $Y\\gets Y\\cup f(x;\\theta_{*})$", + "10: end for" + ], + "bbox": [ + 101, + 387, + 480, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Output: Output $Y$ for input $X$ .", + "bbox": [ + 93, + 553, + 305, + 568 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "various forms [24], allowing for more flexibility in handling different tasks. Detailed discussions about router can be found in Appendix D.6. For detailed algorithms, see Alg 1.", + "bbox": [ + 89, + 598, + 482, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "First, we obtain the merged backbone (line 2) and the task experts (line 4). Then, during the inference, we dynamically route the experts based on inputs (lines 7-9).", + "bbox": [ + 89, + 643, + 482, + 689 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.5. Time Complexity and Storage Space Analysis", + "text_level": 1, + "bbox": [ + 89, + 696, + 475, + 714 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This section analyzes the merging time complexity and storage requirements. We aim to minimize both for practical use. Assume there are $n$ models to merge, each with $m$ parameters and requiring storage space $s$ .", + "bbox": [ + 89, + 719, + 482, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The time complexity includes the merging and inference part. In the merging part, FR-Merging has a complexity of $O(nmlogm)$ [14] while expert extraction is $O(nm)$ , making the overall merging complexity $O(nmlogm)$ , which is manageable. For instance, merging three 10B models requires only $O(10^{12})$ addition operations, a minor load for modern devices. Moreover, with only a lightweight router added, the extra inference cost is negligible.", + "bbox": [ + 89, + 780, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, we analyze storage cost. We typically save only about $1\\%$ additional parameters per model, requiring $(1 + 0.01n)s$ storage, which is far less than the requirement of $ns$ for the model ensemble. For large models, reducing storage requirements is crucial for alleviating storage pressure on edge users. For detailed analysis, please refer to Sec 6.7.", + "bbox": [ + 511, + 334, + 906, + 425 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 511, + 436, + 645, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we evaluate the proposed methods across various tasks. We compare FR-Merging with other cost-free methods (which do not require additional training or data), including Weight Averaging, Fisher Merging [33], Task Arithmetic [21], Ties-Merging [48], Bubbles [7], and PCB-Merging [12]. We also compare our FREE-Merging with popular methods, including RegMean [23], AdaMerging [52], EMR-Merging [20], and Twin-Merging [29]. The original EMR-Merging uses a perfect router; however, we use an imperfect router for realism. Some experiments include traditional MTL with data from all tasks as a baseline. The evaluation covers CV, NLP, and multi-modal tasks.", + "bbox": [ + 511, + 462, + 906, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6.1. Merging Vision Models", + "text_level": 1, + "bbox": [ + 511, + 652, + 728, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Merging 8 ViTs. We select the two visual encoders of CLIP, ViT-B/32 and ViT-L/14 [37], as the pre-trained models and choose eight image classification tasks as target tasks: SUN397, Cars, RESISC45, EuroSAT, SVHN, GT-SRB, MNIST, and DTD [52]. Additionally, an MLP is used as the lightweight router. Detailed information on the datasets and baselines can be found in the Appendix C.1.", + "bbox": [ + 511, + 672, + 906, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2 presents the comparative results. First, FR-Merging outperforms other cost-free methods, improving ViT-B/32 by $2.5\\%$ and ViT-L/14 by $1.5\\%$ over PCB-Merging [12], demonstrating its effectiveness in reducing performance loss caused by task interference. Through FR-Merging, we can construct a merged backbone that can be used independently at a low cost. Here, we focus on the basic FR-Merging. In Sec. 6.6, we discuss further im", + "bbox": [ + 511, + 780, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3867", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/26fe7aa9bb5dbfb6eeb4bc24f2ab7cc890cc3488caefa089a434a7100987540a.jpg", + "table_caption": [ + "Table 3. Comparison of model merging results using ViT-B/16 on 30 visual tasks. Additional Cost indicates whether additional training or additional data is required during model merging." + ], + "table_footnote": [], + "table_body": "
MethodAdditional CostAvg. Acc
IndividualX93.02
Weight AveragingX42.51
Task Arithmetic [21]X48.88
Ties-Merging [48]X37.53
Breadcrumbs [7]X43.57
FR-Merging(ours)X53.90
RegMean [23]✓68.13
AdaMerging [52]✓60.24
Twin-Merging [29]✓75.52
EMR-Merging [20]✓76.39
FREE-Merging(ours)✓79.67
", + "bbox": [ + 117, + 135, + 454, + 305 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a804f2ae98b06a8237374aa3f6563f9c252e7962508f7dd4a5bed7568ca42fc3.jpg", + "table_caption": [ + "Table 4. Language model merging results: RoBERTa fine-tuned on 8 discriminative tasks, T0-3B on 11 discriminative tasks with $IA^3$ , and Qwen-14B on 3 generative tasks with LoRA." + ], + "table_footnote": [], + "table_body": "
MethodBase ModelFull RoBERTaPEFTAvg.
T0-3BQwen-14B
Individual85.5571.3572.0776.32
Weight Averaging51.3458.0166.8058.71
Task Arithmetic[21]66.6563.9166.4065.65
Ties-Merging [48]64.0466.4167.4665.97
Breadcrumbs [7]68.1954.6466.7763.20
PCB-Merging [12]67.8666.1066.6966.39
FR-Merging(ours)70.0266.8868.0068.30
RegMean [23]70.0158.0167.5265.18
EMR-Merging [20]74.2067.1170.9870.76
Twin-Merging [29]78.2966.7071.6872.22
FREE-Merging(ours)80.1668.6872.7873.87
", + "bbox": [ + 94, + 368, + 488, + 563 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "provements achieved by combining it with existing methods. Then, compared to methods requiring extra information, FREE-Merging also improves performance by $2\\%$ , leveraging lightweight computation and efficient expert extraction, requiring only $1\\%$ additional storage per model, outperforming $3\\%$ of EMR-Merging [20] and $2\\%$ of Twin-Merging [29]. Thus, FREE-Merging provides an excellent balance between storage, computation, and performance.", + "bbox": [ + 88, + 588, + 482, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Merging 30 ViTs. We select ViT-B/16 as the pre-trained model and evaluate the model merging performance across 30 image classification tasks following [20].", + "bbox": [ + 89, + 713, + 483, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3 presents the results, with detailed analysis provided in Appendix E.1. It is evident that our proposed FR-Merging and FREE-Merging exhibit remarkable performance. Although merging a large number of models is uncommon in practice, this experiment underscores the effectiveness of Fourier filtering and lightweight task experts in mitigating performance loss caused by task interference. These results highlight the robustness and broad applicability of our methods across diverse scenarios.", + "bbox": [ + 88, + 763, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/758c314aec4e7191d7eb469dc59390f3a718cab1bfd8a784360791238241199c.jpg", + "table_caption": [ + "Table 5. Comparison of performance of different model merging methods after full fine-tuning LLaMa2-13B on three tasks: Instruction Following (LM), Math, and Code Generation." + ], + "table_footnote": [], + "table_body": "
MethodAlpacaEvalGSM8KMBPPAvg.
LM Fine-tuned [46]88.945.931.455.4
Math Fine-tuned [30]22.363.523.036.5
Code Fine-tuned [31]16.218.827.020.7
Task Arithmetic [21]72.148.3040.1
TIES-Merging [48]77.566.927.257.2
DARE [53]77.562.729.456.5
DELLA-Merging [8]80.461.831.457.9
PCB-Merging [12]81.260.630.757.5
FR-Merging(ours)87.564.332.861.5
Twin-Merging [29]83.561.430.458.4
EMR-Merging [20]82.662.331.258.7
FREE-Merging(ours)88.165.731.861.8
", + "bbox": [ + 521, + 141, + 895, + 311 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.2. Merging Language Models", + "text_level": 1, + "bbox": [ + 511, + 332, + 756, + 348 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Merging Medium-sized Language Models. We investigate the generalization ability of our method on medium-scale language models, using RoBERTa [59] as the pretrained model with full fine-tuning on eight discriminative tasks [48]. Details can be found in Appendix C.1.", + "bbox": [ + 511, + 354, + 906, + 430 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The average merging results of the eight tasks are presented in Table 4, with complete results in Appendix E.2. It can be observed that, FR-Merging outperforms other cost-free methods by nearly $2\\%$ . Additionally, FREE-Merging also exhibits remarkable effectiveness, achieving optimal results on almost all datasets and surpassing current methods in average performance. This highlights the great potential of our proposed methods for language models.", + "bbox": [ + 511, + 430, + 908, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Merging PEFT models. PEFT is a popular method that reduces fine-tuning costs [10], but storing PEFT parameters for multiple tasks remains space-inefficient [55]. The effectiveness of our approach on PEFT demonstrates the generalization capability and offers an efficient deployment way.", + "bbox": [ + 511, + 551, + 905, + 627 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Firstly, we use T0-3B [38] as the pre-trained model and $(IA)^{3}$ [26] as the PEFT method, fine-tuning on 11 discriminative tasks [20]. Our proposed FR-Merging and FREE-Merging improve performance by nearly $1\\%$ , as shown in Table 4. Next, we employ Qwen-14B [2] with LoRA [19] as the PEFT way, fine-tuning on three generative tasks [29]. The results show that our methods lead to a $2\\%$ advantage in both cost-free and full frameworks. This confirms the feasibility with PEFT and strong generalization of our methods. Full results are available in the Appendix E.2.", + "bbox": [ + 511, + 628, + 906, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Merging Large Language models. We explore large model merging using LLaMa2-13B [40] as the pre-trained model, with WizardLM [46] for instruction following, WizardMath [30] for math, and WizardCoder-Python [31] for coding. Experiments are conducted on AlpacaEval (instruction following), GSM8K (math), and MBPP (code) [53].", + "bbox": [ + 511, + 779, + 908, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 5 shows that our proposed FR-Merging even surpasses existing methods that rely on additional information.", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3868", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8642d43f40db61de4bea72337382dd561b78d11659556c5e06eb41fae70875a5.jpg", + "table_caption": [ + "Table 6. Results of merging multi-modal BEiT3 models on five vision-language tasks." + ], + "table_footnote": [], + "table_body": "
MethodsTask MetricCOCO-Retrieval Accuracy(↑)COCO-CaptioningImageNet-1k Classification Accuracy(↑)NLVR2 Accuracy(↑)VQAv2 Accuracy(↑)
BLEU4(↑)CIDEr(↑)METEOR(↑)ROUGE-L(↑)
Individual0.84560.3941.3370.3110.6010.85370.77650.8439
Weight Averaging0.18930.0310.0010.1150.1590.67710.28000.6285
Task Arithmetic [21]0.31770.0330.0000.1180.1760.67320.38090.6933
Ties-Merging [48]0.39290.0290.0010.1080.1670.69780.32060.6717
Breadcrumbs [7]0.18930.0110.0120.0650.1140.63230.60730.6823
FR-Merging(ours)0.70100.0160.0190.0960.1410.69920.64070.6965
EMR-Merging [20]0.79460.2891.0600.2720.5340.77420.74750.7211
Twin-Merging [29]0.79370.3251.0780.2780.5580.77300.77620.7232
FREE-Merging(ours)0.80930.3471.1950.2890.5720.78500.81680.7297
", + "bbox": [ + 96, + 114, + 903, + 258 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/50d6947d74695332a096a8c8ec33943b8c5967a9e46af2b5aa16628036d988bc.jpg", + "image_caption": [ + "Figure 4. Visualization of the loss landscape on EuroSAT before and after high-pass filtering (top vs. bottom) in the model merging of the fine-tuned ViT-B/32 on EuroSAT and DTD tasks [21]." + ], + "image_footnote": [], + "bbox": [ + 104, + 280, + 460, + 486 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This indicates that our findings in FR-Merging regarding low-frequency signals and task interference also hold when extended to large models. This is crucial for practical usage of large models. Additionally, our proposed FREE-Merging also demonstrates strong advantages, enabling efficient cross-domain deployment of large models.", + "bbox": [ + 89, + 571, + 482, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.3. Merging Multi-Modal Models", + "text_level": 1, + "bbox": [ + 89, + 676, + 356, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We merge various fine-tuned BEiT3-base [42] models across five tasks: ImageNet-1k (Classification), VQAv2 (Visual Question Answering), NLVR2 (Reasoning), COCO Captioning (Image Captioning), and COCO Retrieval (Image-Text Retrieval). COCO Captioning is evaluated with BLEU4, CIDEr, METEOR, and ROUGE-L, while the others use accuracy [20]. Details are in Appendix C.1.", + "bbox": [ + 89, + 702, + 482, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6 shows that our methods gain great improvements in multi-modal models. Unlike vision or language model merging, which often involve similar tasks, such as classification, the task similarity here is relatively low. The strong performance highlights the generalization ability of our methods, making them widely applicable.", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/818bb12cda6fcf28321d285d560a71489ef0dcdc1eb459a13105ecb45bc4d16a.jpg", + "table_caption": [ + "Table 7. Ablation of FREE-Merging components on 8 visual tasks." + ], + "table_footnote": [], + "table_body": "
FR-MergingEqua.3Top-K ExpertScalingViT-B/32ViT-B/16ViT-L/14
XXXX70.1075.2484.50
✓XXX76.9681.3488.21
X✓XX71.4377.4184.20
✓✓XX78.1082.5488.30
XX✓X65.2468.9870.13
✓✓✓X79.3884.3189.53
XX✓✓82.3286.8790.47
✓✓✓✓89.6891.0993.70
", + "bbox": [ + 517, + 297, + 901, + 393 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.4. Fourier Transform Analysis", + "text_level": 1, + "bbox": [ + 511, + 415, + 764, + 431 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we explain why Fourier high-pass filtering improves merging performance from an optimization perspective. Previous work [25] shows that landscape geometry impacts model generalization. Fig. 4 demonstrates that high-pass filtering widens the valley in the loss landscape (bottom plot). Since fine-tuned models from the same pretrained model are closer in the landscape [34], a wider valley increases the probability of merged results falling within the loss basin as shown in the bottom plot of Fig. 4, leading to better performance. Without filtering, the landscape is flat at a high level, which has little effect on model merging and only highlights the advantage of pre-training [17].", + "bbox": [ + 511, + 438, + 906, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.5. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 632, + 666, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we conduct ablation experiments on the components of our method. All experiments are performed on ViT-B/32 with eight classification tasks as target tasks.", + "bbox": [ + 511, + 656, + 906, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation on FR-Merging. We analyze the impact of FR-Merging on the backbone and overall performance. Table 7 shows that the performance of various models greatly improves after employing FR-Merging. We then compare high-pass filtering with other cost-free methods, including Ties [48], Top-K, Task Arithmetic [21], outlier dropout [7], EMR-Merging [20], and PCB-Merging [12], as well as low-pass and band-pass filtering, with results in Fig. 5a. For fairness, all experts use $1\\%$ additional parameters.", + "bbox": [ + 511, + 703, + 905, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our proposed FR-Merging outperforms other cost-free methods on both the backbone and with task experts, minimizing task interference. Examining the different filtering methods, we find that low-pass filtering leads to great per", + "bbox": [ + 511, + 840, + 905, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3869", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/59fdbfc950c358e2b25b51dacb4bb240a829ea58170cf802bcae9e2c29a103ad.jpg", + "image_caption": [ + "(a) Ablation on FR-Merging." + ], + "image_footnote": [], + "bbox": [ + 99, + 102, + 356, + 265 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cda7c8558fef8b6b097b6692b1c16000d2be81238b533d7640ab996d79654359.jpg", + "image_caption": [ + "(b) Ablation on Task Expert." + ], + "image_footnote": [], + "bbox": [ + 370, + 99, + 630, + 265 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e20f278e0c623054f3f87346772bc71f55dfae6ccb70aa07ed87a51d957bc6d0.jpg", + "image_caption": [ + "(c) Router Analysis.", + "Figure 5. a) The ablation experiments on FR-Merging show that our proposed methods achieve optimal performance. b) Performance of experts at different sparsity rates. c) In experiments with random routers, our approach demonstrates optimal performance." + ], + "image_footnote": [], + "bbox": [ + 640, + 92, + 898, + 266 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/90cb0b51041d63fb0f47fc7e3c746e12cb30dc59b676674dda7ab4e01f571a86.jpg", + "table_caption": [ + "Table 8. Transferability analysis of FR-Merging on 8 visual tasks." + ], + "table_footnote": [], + "table_body": "
MethodFR-MergingViT-B/32ViT-L/14
Ties-Merging [48]X73.686.0
Ties-Merging [48]✓78.388.4
Breadcrumbs [7]X75.086.4
Breadcrumbs [7]✓78.888.9
TSVM [16]X83.991.9
TSVM [16]✓84.892.5
", + "bbox": [ + 102, + 358, + 468, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "formance degradation, confirming that task interference in low-frequency regions is severe. Band-pass filtering also performs worse than FR-Merging, and this is because some high-frequency signals preserve abrupt changes in the signal and their removal harms the performance of the model.", + "bbox": [ + 89, + 476, + 482, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation on Merging Coefficients. We conduct an ablation study on our merging coefficients as in Eq. 3, with results in Table 7. Using them instead of simple averaging yields slight improvements by considering output variations, benefiting tasks with varying fine-tuning magnitudes.", + "bbox": [ + 89, + 551, + 482, + 627 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation on Task Expert. We conduct an ablation study on the expert extraction way. Table 7 shows that incorporating our experts and rescaling method results in significant improvements, confirming their effectiveness. To validate the superiority of our method, we evaluate expert-only performance with varying parameter proportions, comparing Top-K selection, Bernoulli selection [53], SVD extraction [29], and our Top-K selection with rescaling. Fig. 5b shows that our method excels, particularly under high sparsity.", + "bbox": [ + 89, + 627, + 482, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Router Analysis. We analyze the proposed router method, aiming for strong merging performance despite sub-optimal routing. Since router inaccuracies may arise from limited training resources or edge deployment issues, our methods maintain performance in real-world scenarios. Using a random router to simulate inaccuracy, we compare our method with existing router-based approaches. Our method outperforms Twin-Merging by $13.5\\%$ and EMR-Merging by $8.5\\%$ on average, as shown in Fig. 5c, indicating that our methods", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7245e16f0f9be52b8044c6425480daf09fab74831f625055b9ec66cfe8b9aa7c.jpg", + "table_caption": [ + "Table 9. Trade-off of performance, training, inference, and storage costs using ViT-L/14 across 8 image classification tasks." + ], + "table_footnote": [], + "table_body": "
MethodTraining TokensTraining CostStorage Param.Inference Cost (/1000 items)Acc.
Individual002.4B31.0s94.2
MTL304M12h304M31.0s93.5
FR-Merging00304M31.0s88.2
Twin-Merging0.9M147s352M32.2s92.7
FREE-Merging0.9M147s328M32.2s93.7
", + "bbox": [ + 516, + 373, + 905, + 459 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "construct a reasonable backbone and generalizable experts.", + "bbox": [ + 511, + 477, + 901, + 492 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.6. Transferability Analysis", + "text_level": 1, + "bbox": [ + 511, + 502, + 735, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our proposed FR-Merging can integrate with other methods, as it addresses frequency-domain task interference that current methods overlook. As shown in Table 8, combining various methods with FR-Merging leads to performance improvements. Therefore, FR-Merging can be integrated with more approaches to achieve greater benefits.", + "bbox": [ + 511, + 525, + 905, + 616 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.7. Speed and Storage Space Analysis", + "text_level": 1, + "bbox": [ + 511, + 626, + 812, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This part analyzes training costs, inference speed, and storage as shown in Table 9. FR-Merging achieves strong performance with significantly reducing storage costs without extra training or inference costs, while FREE-Merging slightly increases overhead for better performance. Both of these two methods enhance deep learning deployment.", + "bbox": [ + 511, + 648, + 905, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 753, + 633, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose FR-Merging and FREE-Merging to efficiently build a multi-task model by merging existing ones. FR-Merging constructs a backbone without training by high-pass filtering to remove information that dominates task interference. Meanwhile, FREE-Merging employs lightweight experts to offset information loss during merging. Our method balances training, storage, inference costs, and performance, showing high practical value.", + "bbox": [ + 511, + 779, + 905, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3870", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 188, + 187, + 204 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Samuel Ainsworth, Jonathan Hayase, and Siddhartha Srinivasa. Git re-basin: Merging models modulo permutation symmetries. In The Eleventh International Conference on Learning Representations, 2023. 4", + "[2] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 6", + "[3] Tim-Oliver Buchholz and Florian Jug. Fourier image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1846-1854, 2022. 4", + "[4] Shijie Chen, Yu Zhang, and Qiang Yang. Multi-task learning in natural language processing: An overview. ACM Computing Surveys, 56(12):1-32, 2024. 2", + "[5] Tianyi Chu, Jiafu Chen, Jiakai Sun, Shuobin Lian, Zhizhong Wang, Zhiwen Zuo, Lei Zhao, Wei Xing, and Dongming Lu. Rethinking fast fourier convolution in image inpainting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23195-23205, 2023. 4", + "[6] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instructionfinetuned language models. Journal of Machine Learning Research, 25(70):1-53, 2024. 3, 4", + "[7] MohammadReza Davari and Eugene Belilovsky. Model breadcrumbs: Scaling multi-task model merging with sparse masks. arXiv preprint arXiv:2312.06795, 2023. 3, 5, 6, 7, 8, 9, 10", + "[8] Pala Tej Deep, Rishabh Bhardwaj, and Soujanya Poria. Della-merging: Reducing interference in model merging through magnitude-based sampling. arXiv preprint arXiv:2406.11617, 2024. 6", + "[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota, 2019. Association for Computational Linguistics. 1", + "[10] Ning Ding, Yujia Qin, Guang Yang, Fuchao Wei, Zonghan Yang, Yusheng Su, Shengding Hu, Yulin Chen, Chi-Min Chan, Weize Chen, et al. Parameter-efficient fine-tuning of large-scale pre-trained language models. Nature Machine Intelligence, 5(3):220-235, 2023. 6", + "[11] Xibin Dong, Zhiwen Yu, Wenming Cao, Yifan Shi, and Qianli Ma. A survey on ensemble learning. Frontiers of Computer Science, 14:241-258, 2020. 2" + ], + "bbox": [ + 93, + 213, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "This work is supported by National Natural Science Foundation of China (NSFC) (62232005, 62202126); the National Key Research and Development Program of China (2021YFB3300502).", + "[12] Guodong Du, Junlin Lee, Jing Li, Runhua Jiang, Yifei Guo, Shuyang Yu, Hanting Liu, Sim Kuan Goh, Ho-Kin Tang, Daojing He, et al. Parameter competition balancing for model merging. arXiv preprint arXiv:2410.02396, 2024. 2, 3, 5, 6, 7, 8, 9, 10", + "[13] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 3", + "[14] Pierre Duhamel and Martin Vetterli. Fast fourier transforms: a tutorial review and a state of the art. Signal processing, 19 (4):259-299, 1990. 5", + "[15] Mudasir A Ganaie, Minghui Hu, Ashwani Kumar Malik, Muhammad Tanveer, and Ponnuthurai N Suganthan. Ensemble deep learning: A review. Engineering Applications of Artificial Intelligence, 115:105151, 2022. 2", + "[16] Antonio Andrea Gargiulo, Donato Crisostomi, Maria Sofia Bucarelli, Simone Scardapane, Fabrizio Silvestri, and Emanuele Rodola. Task singular vectors: Reducing task interference in model merging. arXiv preprint arXiv:2412.00081, 2024. 8", + "[17] Timur Garipov, Pavel Izmailov, Dmitrii Podoprikhin, Dmitry P Vetrov, and Andrew G Wilson. Loss surfaces, mode connectivity, and fast ensembling of dnns. Advances in neural information processing systems, 31, 2018. 7", + "[18] Rafael C Gonzales and Paul Wintz. Digital image processing. Addison-Wesley Longman Publishing Co., Inc., 1987. 2", + "[19] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 6", + "[20] Chenyu Huang, Peng Ye, Tao Chen, Tong He, Xi-angyu Yue, and Wanli Ouyang. Emr-merging: Tuning-free high-performance model merging. arXiv preprint arXiv:2405.17461, 2024. 1, 2, 4, 5, 6, 7, 3, 8, 9, 10", + "[21] Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Suchin Gururangan, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. Editing models with task arithmetic. arXiv preprint arXiv:2212.04089, 2022. 1, 2, 3, 5, 6, 7, 4, 8, 9, 10", + "[22] Dongfu Jiang, Xiang Ren, and Bill Yuchen Lin. Lmblender: Ensembling large language models with pairwise ranking and generative fusion. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 14165-14178, 2023. 2", + "[23] Xisen Jin, Xiang Ren, Daniel Preotiac-Pietro, and Pengxiang Cheng. Dataless knowledge fusion by merging weights of language models. arXiv preprint arXiv:2212.09849, 2022. 1, 2, 5, 6, 3, 8, 9, 10", + "[24] Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan Firat, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. Gshard: Scaling giant models with conditional computation and automatic sharding. In International Conference on Learning Representations, 2021. 5" + ], + "bbox": [ + 93, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3871", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Hao Li, Zheng Xu, Gavin Taylor, Christoph Studer, and Tom Goldstein. Visualizing the loss landscape of neural nets. Advances in neural information processing systems, 31, 2018. 7", + "[26] Haokun Liu, Derek Tam, Mohammed Muqeeth, Jay Mohta, Tenghao Huang, Mohit Bansal, and Colin A Raffel. Few-shot parameter-efficient fine-tuning is better and cheaper than in-context learning. Advances in Neural Information Processing Systems, 35:1950-1965, 2022. 6", + "[27] James Liu, Guangxuan Xiao, Kai Li, Jason D Lee, Song Han, Tri Dao, and Tianle Cai. Bitdelta: Your fine-tune may only be worth one bit. Advances in Neural Information Processing Systems, 37:13579-13600, 2024. 2", + "[28] Yajing Liu, Yuning Lu, Hao Liu, Yaozu An, Zhuoran Xu, Zhuokun Yao, Baofeng Zhang, Zhiwei Xiong, and Chenguang Gui. Hierarchical prompt learning for multi-task learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10888-10898, 2023. 2", + "[29] Zhenyi Lu, Chenghao Fan, Wei Wei, Xiaoye Qu, Dangyang Chen, and Yu Cheng. Twin-merging: Dynamic integration of modular expertise in model merging. arXiv preprint arXiv:2406.15479, 2024. 1, 2, 5, 6, 7, 8, 3, 9, 10", + "[30] Haipeng Luo, Qingfeng Sun, Can Xu, Pu Zhao, Jianguang Lou, Chongyang Tao, Xiubo Geng, Qingwei Lin, Shifeng Chen, and Dongmei Zhang. Wizardmath: Empowering mathematical reasoning for large language models via reinforced evol-instruct. arXiv preprint arXiv:2308.09583, 2023. 6", + "[31] Ziyang Luo, Can Xu, Pu Zhao, Qingfeng Sun, Xiubo Geng, Wenxiang Hu, Chongyang Tao, Jing Ma, Qingwei Lin, and Daxin Jiang. Wizardcoder: Empowering code large language models with evol-instruct. In The Twelfth International Conference on Learning Representations, 2024. 6", + "[32] Xinyin Ma, Gongfan Fang, and Xinchao Wang. Llm-pruner: On the structural pruning of large language models. Advances in neural information processing systems, 36:21702-21720, 2023. 4", + "[33] Michael S Matena and Colin A Raffel. Merging models with fisher-weighted averaging. Advances in Neural Information Processing Systems, 35:17703-17716, 2022. 1, 2, 5, 9", + "[34] Behnam Neyshabur, Hanie Sedghi, and Chiyuan Zhang. What is being transferred in transfer learning? Advances in neural information processing systems, 33:512-523, 2020. 7", + "[35] Tan Nguyen, Minh Pham, Tam Nguyen, Khai Nguyen, Stanley Osher, and Nhat Ho. Fourierformer: Transformer meets generalized fourier integral theorem. Advances in Neural Information Processing Systems, 35:29319-29335, 2022. 4", + "[36] Biqing Qi, Fangyuan Li, Zhen Wang, Junqi Gao, Dong Li, Peng Ye, and Bowen Zhou. Less is more: Efficient model merging with binary task switch. arXiv preprint arXiv:2412.00054, 2024. 2", + "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 3, 4, 5" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Victor Sanh, Albert Webson, Colin Raffel, Stephen Bach, Lintang Sutawika, Zaid Alyafeai, Antoine Chaffin, Arnaud Stiegler, Arun Raja, Manan Dey, et al. Multitask prompted training enables zero-shot task generalization. In International Conference on Learning Representations, 2022. 6", + "[39] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. In International Conference on Learning Representations, 2016. 2, 4", + "[40] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 3, 6", + "[41] Simon Vandenhende, Stamatios Georgoulis, Wouter Van Gansbeke, Marc Proesmans, Dengxin Dai, and Luc Van Gool. Multi-task learning for dense prediction tasks: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(7):3614-3633, 2021. 2", + "[42] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, et al. Image as a foreign language: Beit pretraining for vision and vision-language tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19175-19186, 2023. 7", + "[43] T Wolf. Huggingface's transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771, 2019. 1", + "[44] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In International conference on machine learning, pages 23965-23998. PMLR, 2022. 2", + "[45] Chengyue Wu, Teng Wang, Yixiao Ge, Zeyu Lu, Ruisong Zhou, Ying Shan, and Ping Luo. $\\pi$ -tuning: Transferring multimodal foundation models with optimal multi-task interpolation. In International Conference on Machine Learning, pages 37713-37727. PMLR, 2023. 4", + "[46] Can Xu, Qingfeng Sun, Kai Zheng, Xiubo Geng, Pu Zhao, Jiazhan Feng, Chongyang Tao, Qingwei Lin, and Daxin Jiang. Wizardlm: Empowering large pre-trained language models to follow complex instructions. In The Twelfth International Conference on Learning Representations, 2024. 6", + "[47] Jing Xu and Jingzhao Zhang. Random masking finds winning tickets for parameter efficient fine-tuning. In *Forty-first International Conference on Machine Learning*. 2", + "[48] Prateek Yadav, Derek Tam, Leshem Choshen, Colin A Raffel, and Mohit Bansal. Ties-merging: Resolving interference when merging models. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3, 4, 5, 6, 7, 8, 9, 10", + "[49] Enneng Yang, Li Shen, Guibing Guo, Xingwei Wang, Xiaochun Cao, Jie Zhang, and Dacheng Tao. Model merging" + ], + "bbox": [ + 516, + 90, + 903, + 902 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3872", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "in llms, mllms, and beyond: Methods, theories, applications and opportunities. arXiv preprint arXiv:2408.07666, 2024. 2", + "[50] Enneng Yang, Li Shen, Zhenyi Wang, Guibing Guo, Xiaojun Chen, Xingwei Wang, and Dacheng Tao. Representation surgery for multi-task model merging. arXiv preprint arXiv:2402.02705, 2024. 3, 4", + "[51] Enneng Yang, Li Shen, Zhenyi Wang, Guibing Guo, Xiao-jun Chen, Xingwei Wang, and Dacheng Tao. Representation surgery for multi-task model merging. In *Forty-first International Conference on Machine Learning*, 2024. 1", + "[52] Enneng Yang, Zhenyi Wang, Li Shen, Shiwei Liu, Guibing Guo, Xingwei Wang, and Dacheng Tao. Adamerging: Adaptive model merging for multi-task learning. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 5, 6, 3, 9", + "[53] Le Yu, Bowen Yu, Haiyang Yu, Fei Huang, and Yongbin Li. Language models are super mario: Absorbing abilities from homologous models as a free lunch. In *Forty-first International Conference on Machine Learning*, 2024. 2, 3, 4, 6, 8, 10", + "[54] Kerem Zaman, Leshem Choshen, and Shashank Srivastava. Fuse to forget: Bias reduction and selective memorization through model fusion. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 18763-18783, Miami, Florida, USA, 2024. Association for Computational Linguistics. 8", + "[55] Jinghan Zhang, Junteng Liu, Junxian He, et al. Composing parameter-efficient modules with arithmetic operation. Advances in Neural Information Processing Systems, 36: 12589-12610, 2023. 6", + "[56] Yu Zhang and Qiang Yang. A survey on multi-task learning. IEEE transactions on knowledge and data engineering, 34 (12):5586-5609, 2021. 1", + "[57] Yanqi Zhou, Tao Lei, Hanxiao Liu, Nan Du, Yanping Huang, Vincent Zhao, Andrew M Dai, Quoc V Le, James Laudon, et al. Mixture-of-experts with expert choice routing. Advances in Neural Information Processing Systems, 35:7103-7114, 2022. 2", + "[58] Yuyan Zhou, Liang Song, Bingning Wang, and Weipeng Chen. Metagpt: Merging large language models using model exclusive task arithmetic. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 1711-1724, 2024. 1", + "[59] Liu Zhuang, Lin Wayne, Shi Ya, and Zhao Jun. A robustly optimized BERT pre-training approach with post-training. In Proceedings of the 20th Chinese National Conference on Computational Linguistics, pages 1218-1227, Huhhot, China, 2021. Chinese Information Processing Society of China. 6" + ], + "bbox": [ + 91, + 90, + 482, + 796 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "3873", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/FREE-Merging_ Fourier Transform for Efficient Model Merging/f3aa67c5-53d3-463b-8730-22ad4f95fe7f_model.json b/2025/FREE-Merging_ Fourier Transform for Efficient Model Merging/f3aa67c5-53d3-463b-8730-22ad4f95fe7f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..deedd10c83524d6dbd36f2102852c2856f2ba2b2 --- /dev/null +++ b/2025/FREE-Merging_ Fourier Transform for Efficient Model Merging/f3aa67c5-53d3-463b-8730-22ad4f95fe7f_model.json @@ -0,0 +1,2598 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.172, + 0.131, + 0.828, + 0.154 + ], + "angle": 0, + "content": "FREE-Merging: Fourier Transform for Efficient Model Merging" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.182, + 0.471, + 0.234 + ], + "angle": 0, + "content": "Shenghe Zheng \nHarbin Institute of Technology \nshenghez.zheng@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.182, + 0.772, + 0.234 + ], + "angle": 0, + "content": "Hongzhi Wang* \nHarbin Institute of Technology \nwangzh@hit.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.303, + 0.485, + 0.726 + ], + "angle": 0, + "content": "With the rapid growth of deep learning, there is an increasing availability of open-source models for various tasks. However, single fine-tuned models often fall short of meeting the diverse needs of users. Model merging has thus emerged as an efficient method to integrate the capabilities of existing models into a unified model. Nevertheless, existing model merging methods face challenging trade-offs between performance and deployment costs, primarily due to task interference. For the first time, we reveal that task interference is evident in the frequency domain of model parameters, yet current efforts only focus on spatial domain solutions, which are largely ineffective in addressing frequency domain interference. To mitigate the impact of frequency domain interference, we propose FR-Merging, an innovative method that effectively filters harmful frequency domain interference on the backbone with minimal computational overhead. Since performance loss is inevitable with cost-free methods, we propose a lightweight task-specific expert module that dynamically compensates for information loss during merging. This proposed framework, FREE-Merging (FR-Merging with experts), strikes a balanced trade-off between training cost, inference latency, storage requirements, and performance. We demonstrate the effectiveness of both FR-Merging and FREE-Merging on multiple tasks across CV, NLP, and Multi-Modal domains and show that they can be flexibly adapted to specific needs. Our code is available at: https://github.com/Zhengsh123/FREE-Merging." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.757, + 0.222, + 0.773 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.784, + 0.483, + 0.875 + ], + "angle": 0, + "content": "In the current era of deep learning, the pretrain-finetune paradigm is a standard procedure [9, 37, 40]. However, edge applications often lack powerful computing resources and data, favoring ready-to-use models for specific scenarios. The recent growth of open-source platforms like Hugging-Face [43] has made this feasible. Yet, practical demands" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.27, + 0.907, + 0.406 + ], + "angle": 0, + "content": "are often not fully covered by a single existing model and are typically a union of several models [23, 52]. However, running separate models for each subtask increases storage and inference costs, especially with large models. Model merging attempts to combine existing models to build a model capable of addressing all target tasks [21, 33, 48]. This method reduces training costs, data privacy issues in multi-task learning (MTL) [20, 56] and deployment costs, garnering widespread attention." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.909, + 0.607 + ], + "angle": 0, + "content": "While model merging aims to reduce training costs, it often suffers from performance loss due to task interference [20]. To mitigate this issue, existing methods develop along three directions. The first direction computes merging coefficients based on weights or data [23, 33]. The second uses the parameter differences before and after finetuning (i.e. task vectors) for merging [21, 48, 52, 58]. The third method introduces independent knowledge for each task [20, 29]. While the third way balances performance and costs, it still faces two critical limitations. First, it neglects backbone optimization during merging, resulting in subpar performance. Second, it requires high computation and storage, making it impractical for edge deployment." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.611, + 0.909, + 0.777 + ], + "angle": 0, + "content": "We argue that an effective model merging method requires addressing both of the above issues. First, since the backbone constitutes the majority of parameters and thus determines performance, an efficient backbone merging method is essential to reduce performance loss. Then, given the no free lunch theorem, cost-free merging cannot avoid task interference. Therefore, introducing task-specific experts is necessary. However, they must remain lightweight to avoid storage and inference overhead. Thus, we propose FREE-Merging, a two-stage method that ensures an efficient backbone and lightweight experts as shown in Fig 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.901 + ], + "angle": 0, + "content": "For backbone merging, the key lies in eliminating task interference [51]. We identify a compelling phenomenon in which task interference is evident in the frequency domain of model parameters. And we present in Sec. 4 that frequency-domain interference is strongly correlated with performance. Notably, existing methods, which focus only on spatial-domain operations, fail to address frequency-domain interference, leading to suboptimal outcomes. Ex" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.239, + 0.9 + ], + "angle": 0, + "content": "*Corresponding Author" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3863" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.316, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.32, + 0.089, + 0.663, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.09, + 0.906, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.297, + 0.908, + 0.34 + ], + "angle": 0, + "content": "Figure 1. The workflow of FREE-Merging involves two main steps. First, FR-Merging utilizes high-pass filtering to remove harmful specialized information from each model, thereby constructing a high-quality merged backbone. Second, it employs lightweight extraction of task experts, which are dynamically added during inference to mitigate the impact of task interference." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.366, + 0.483, + 0.577 + ], + "angle": 0, + "content": "Experimental evidence reveals that task interference is severe in the low-frequency region. This occurs because, high-frequency signals represent fine-grained variations, while low-frequency parts capture the global structure [18], which is more likely to contain task-specific information that leads to task interference. Leveraging the inherent redundancy in fine-tuned parameters [27, 47, 53], we propose to directly filter out low-frequency parts with severe task interference. As detailed in Sec. 5.2, this high-pass filtering substantially enhances generalization with minimal performance loss, preventing task interference during merging. Our lightweight yet highly effective approach, FR-Merging, is the first to apply Fourier filtering to neural network parameters, showing its potential for broader applications." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.581, + 0.483, + 0.702 + ], + "angle": 0, + "content": "After obtaining a high-performance backbone through FR-Merging, we use a lightweight task-specific expert to recover the inevitable information loss with minimal storage. Unlike existing methods that store extensive knowledge, we rescale low-frequency signals, which store task-specific information, retaining task expertise with only \\(1\\%\\) of the parameters. Inspired by MoE [39, 57], we employ a router for dynamic expert routing, enhancing model flexibility." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In summary, we introduce FREE-Merging, which utilizes FR-Merging to reduce task interference and incorporates lightweight experts to eliminate information loss of model merging. Our main contributions are: 1). We identify frequency-domain task interference as a critical factor that greatly impacts merging performance but remains challenging to address through spatial-domain methods. 2). To tackle frequency-domain interference, we propose FR-Merging, which reduces interference while preserving performance, constructing the merged backbone. 3). We theoretically validate the necessity of introducing new information during merging and propose an effective expert construction method. 4). Extensive experiments show that" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.366, + 0.905, + 0.397 + ], + "angle": 0, + "content": "FR-Merging and FREE-Merging effectively enhance performance across vision, language, and multi-modal tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.408, + 0.655, + 0.423 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.433, + 0.907, + 0.72 + ], + "angle": 0, + "content": "Model Merging. Model merging integrates existing models to handle multiple tasks without training [21, 23, 49], but faces task interference challenges [36, 48]. Simple averaging [44] causes great performance degradation. Methods like Fisher-Merging [33] and RegMean [23] use matrices to determine merging coefficients but are computationally expensive or data-intensive, limiting edge deployment. Task Arithmetic [21] merges task vectors instead of weights, while Ties-Merging [48] resolves parameter conflicts, and AdaMerging [52] automates coefficient selection. However, these approaches have data distribution requirements and limited applicability. Techniques such as DARE [53] and PCB-Merging [12] mitigate performance drops but still face task conflicts. EMR-Merging [20] and Twin-Merging [29] store task-specific knowledge but require large storage and neglect backbone optimization. Our method uses high-pass filtering and lightweight experts to balance merging costs and effectiveness across various tasks. For further details, please refer to Appendix C.2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.905, + 0.795 + ], + "angle": 0, + "content": "Model Ensemble. Model ensemble combines outputs from multiple models [11, 15, 22], but with large models, it faces storage and inference challenges. In contrast, Model merging enables a single model to solve multiple tasks, significantly reducing storage and inference costs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Multi-task Learning. Multi-task learning (MTL) aims to solve multiple tasks using a single model [4, 28, 41], but in the era of large models, it faces challenges like high training costs, data privacy issues, and expertise requirements [52]. In contrast, model merging uses existing open-source models to create a multi-task model with little or no training, significantly reducing deployment costs." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3864" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.092, + 0.482, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.289, + 0.203, + 0.304, + 0.213 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.213, + 0.482, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.289, + 0.324, + 0.304, + 0.333 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.342, + 0.485, + 0.426 + ], + "angle": 0, + "content": "Figure 2. a) The frequency domain amplitude power distribution of an attention head across different fine-tuned ViT-B-32 models. b) The correlation between the merging performance of ViT-B-32 and the variance of frequency domain amplitude power. Normalized accuracy is defined as the ratio of merged model performance to fine-tuned model performance on the same task." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.452, + 0.231, + 0.468 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.478, + 0.483, + 0.599 + ], + "angle": 0, + "content": "Notation. Assume that \\( f_{\\theta}(x) \\to y \\) represents a neural network parameterized by \\( \\theta \\), where \\( x \\in R^{m} \\) is the input and \\( y \\in R^{n} \\) is the output. Recent studies have focused on model merging via task vectors. For task \\( k \\), the task vector is defined as \\( v_{k} = \\theta_{k} - \\theta_{pre} \\), where \\( \\theta_{k} \\) is the fine-tuned model for task \\( k \\), and \\( \\theta_{pre} \\) is the corresponding pre-trained model. Additionally, \\( e_{k} \\) represents the task expert that retains the ability to solve the task \\( k \\) with few parameters." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.6, + 0.483, + 0.768 + ], + "angle": 0, + "content": "Problem Definition. The objective of model merging is to combine models \\(\\{\\theta_k\\}_{k=1}^K\\) for tasks \\(T = \\{t_i\\}_{i=1}^K\\) into a single model \\(\\theta_m\\) that can handle all tasks in \\(T\\). We focus on a practical scenario where a shared pre-trained model \\(\\theta_{pre}\\) serves as the base for all fine-tuned models. It is formulated as \\(\\theta_m = \\lambda \\sum_{k=1}^K \\hat{G}(\\theta_k)\\), and with task vector, it becomes \\(\\theta_m = \\theta_{pre} + \\lambda \\Sigma_{k=1}^K G(v_k)\\), where \\(\\lambda\\) is the coefficient, and \\(\\hat{G}/G\\) is transformation on the parameters. Assuming a shared pre-trained model aligns with trends toward foundation models, like LLaMa [40] for language and CLIP [37] for vision, which are then fine-tuned for specific tasks." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.784, + 0.45, + 0.801 + ], + "angle": 0, + "content": "4. Task Interference in Frequency Domain" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.885 + ], + "angle": 0, + "content": "Task interference is a key factor that leads to the decrease in merging performance. While current methods focus on addressing interference in the spatial domain [48, 50, 53], we observe that task vectors from different fine-tuned models also exhibit great misalignment in the frequency domain." + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "As shown in Fig. 2(a), we visualize the amplitude power" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.907, + 0.118 + ], + "angle": 0, + "content": "Table 1. The impact of current methods on the mean variance of frequency-domain amplitude across 8 image classification tasks." + }, + { + "type": "table", + "bbox": [ + 0.541, + 0.128, + 0.88, + 0.229 + ], + "angle": 0, + "content": "
MethodViT-B/32ViT-L/14
Task Arithmetic [21]0.0590.110
DARE [53]0.057(↓3%)0.108(↓2%)
Ties-Merging [48]0.058(↓2%)0.109(↓1%)
Breadcrumbs [7]0.058(↓2%)0.107(↓3%)
PCB-Merging [12]0.056(↓5%)0.107(↓3%)
FR-Merging(ours)0.045(↓24%)0.088(↓20%)
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.251, + 0.905, + 0.357 + ], + "angle": 0, + "content": "distribution in the frequency domain of task vectors from a ViT-B-32 attention head fine-tuned on various tasks [21]. Here, greater distance from the center represents a higher frequency. Since amplitude indicates signal strength, merging two signals with vastly different strengths inevitably results in the weaker signal being masked. Therefore, task interference in the low-frequency part are severe." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.357, + 0.906, + 0.447 + ], + "angle": 0, + "content": "Furthermore, in Fig.2(b), we train multiple groups of models on eight image tasks [21] and find a clear negative correlation between the variance of amplitude power among models and the normalized merging performance (details can be found in Appendix D.1), confirming the great impact of frequency-domain interference on merging performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.905, + 0.508 + ], + "angle": 0, + "content": "Additionally, as shown in Table 1, we show that existing works addressing task interference in the spatial domain fail to effectively mitigate frequency-domain task interference. Therefore, we propose our approach to tackle this issue." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.52, + 0.649, + 0.538 + ], + "angle": 0, + "content": "5. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.545, + 0.906, + 0.621 + ], + "angle": 0, + "content": "This section presents FREE-Merging: Sec. 5.1 explains the need for a two-part design, Sec. 5.2 and Sec. 5.3 cover the FR-Merging for the backbone and the lightweight expert module, respectively, followed by the FREE-Merging in Sec. 5.4 and a complexity analysis in Sec. 5.5." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.629, + 0.636, + 0.643 + ], + "angle": 0, + "content": "5.1. Motivation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.651, + 0.905, + 0.681 + ], + "angle": 0, + "content": "We first analyze why both merging optimization for the backbone and the introduction of experts are necessary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.681, + 0.906, + 0.785 + ], + "angle": 0, + "content": "First, backbone optimization is crucial but overlooked by current expert-based merging ways. As the backbone holds most parameters, its performance is key to the overall success. For instance, upgrading from LLaMa2 [40] to LLaMa3 [13] greatly improves results even with the identical fine-tuning. Thus, we propose a training-free merging method to reduce task interference in the backbone." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.787, + 0.906, + 0.848 + ], + "angle": 0, + "content": "Next, we present a theorem showing the necessity of task experts. Based on Theorem 5.1, introducing additional information is essential. For proof, see Appendix B, and the effectiveness of task experts is proved in Appendix D.4." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Theorem 5.1 Model Merging has no free lunch. Merged model \\(\\theta_{m}\\) cannot simultaneously retain the capabilities of \\(\\{\\theta_k\\}_{k=1}^K\\) without introducing additional information." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3865" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.226, + 0.108 + ], + "angle": 0, + "content": "5.2. FR-Merging" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.485, + 0.219 + ], + "angle": 0, + "content": "We propose an effective backbone merging method by addressing task interference in the low-frequency fine-tuned parameters as discussed in Sec. 4, which is hard to resolve in the spatial domain. Given the high redundancy of fine-tuned parameters [53], we apply high-pass filtering to directly remove the interfering regions. Let \\( v(x,y) \\) be the task vector and \\( G(x,y) \\) the transformed result. Then we have:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.226, + 0.483, + 0.243 + ], + "angle": 0, + "content": "\\[\nG (x, y) = \\mathcal {F} ^ {- 1} \\{H (\\eta , \\gamma) \\cdot \\mathcal {F} \\{v (x, y) \\} \\}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.251, + 0.138, + 0.264 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.269, + 0.483, + 0.309 + ], + "angle": 0, + "content": "\\[\nH (\\eta , \\gamma) = \\left\\{ \\begin{array}{l l} 1, & \\sqrt {\\eta^ {2} + \\gamma^ {2}} \\geq D _ {0} \\\\ 0, & \\sqrt {\\eta^ {2} + \\gamma^ {2}} < D _ {0} \\end{array} \\right., \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.317, + 0.484, + 0.377 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}\\) and \\(\\mathcal{F}^{-1}\\) represent the Fourier and inverse Fourier transform, respectively. \\(D_0\\) is the adjustable cutoff frequency. And \\(\\eta\\) and \\(\\gamma\\) are the distances from the frequency domain center along the \\(x\\)- and \\(y\\)-axes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.378, + 0.484, + 0.497 + ], + "angle": 0, + "content": "Next, we analyze why high-pass filtering is effective from both intuitive and experimental perspectives. Fine-tuned weights occupy distinct positions in the loss landscape, and linear interpolation often leads to high-loss regions, causing task interference [1, 50]. To address this, we minimize model differences while preserving performance, bringing them closer in the loss landscape to increase the likelihood of the merged model falling into a loss basin [45]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.498, + 0.484, + 0.634 + ], + "angle": 0, + "content": "Results in Sec. 4 show great differences in low-frequency regions. This is because, if we consider neural network parameters as signals carrying training information, inspired by image processing [3, 35], high-frequency signals represent sharp variations, while low-frequency parts define the overall structure [5]. Since most information resides in the structure, filtering out low-frequency signals removes key task-specific information. Then, the likelihood that the interpolation falls within the loss basin increases [1]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.634, + 0.484, + 0.785 + ], + "angle": 0, + "content": "Experimentally, we validate that high-pass filtering enhances model generalization with minimal performance loss as shown in Fig 3. We fine-tune ViT-B/16 [37] on 30 visual tasks following [20]. We find that removing low-frequency information results in a smaller performance drop (diagonal) compared to the improvement in generalization (off-diagonal). This indicates that low-frequency parts carry task-specific information that boosts task performance but reduces generalization. Thus, applying filtering like FR-Merging helps alleviate task interference during merging." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.786, + 0.484, + 0.815 + ], + "angle": 0, + "content": "To maintain consistent output after merging, we define the merging coefficient \\(\\lambda_{i}\\) for the task vector \\(v_{i}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.822, + 0.483, + 0.865 + ], + "angle": 0, + "content": "\\[\n\\lambda_ {i} = \\mathbb {E} (v _ {i}) \\left(\\sum_ {j = 1} ^ {K} \\mathbb {E} (v _ {j})\\right) ^ {- 1}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathbb{E}\\) represents the average. Overall, we construct a merged backbone in this part that minimizes task conflicts." + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.091, + 0.88, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.317, + 0.907, + 0.374 + ], + "angle": 0, + "content": "Figure 3. High-pass filtering improves performance over the unfiltered version. Rows represent the added task vectors, while columns are the test datasets. The results show a significant boost in generalization with only a slight decline in task performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.396, + 0.787, + 0.412 + ], + "angle": 0, + "content": "5.3. Lightweight Expert Extraction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.418, + 0.906, + 0.508 + ], + "angle": 0, + "content": "To compensate for information lost due to discarding low-frequency signals, we utilize lightweight experts added during inference. As analyzed in Sec. 5.2, low-frequency signals contain task-specific information, but preserving them requires an inefficient inverse Fourier transform during each inference. Thus, we propose a more efficient way." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.906, + 0.599 + ], + "angle": 0, + "content": "Since fine-tuning leads to task specialization, the most changed parameters can be approximated as having the greatest impact [32]. After reasonable rescaling, these selected parameters serve as task experts, minimizing storage (typically around \\(1\\%\\) of parameters) without extra inference computation. The validity is detailed in Appendix D.4." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.906, + 0.689 + ], + "angle": 0, + "content": "Although magnitude-based extraction [48] exists, limited research on rescaling hinders compression. As rescaling is crucial to expert extraction, we propose a rescaling method. Let \\( M(v_{i},d) \\) represent the top- \\( d \\) (\\( d \\) is a percentage) parameters for task vector \\( v_{i} \\), \\( \\mu_{i} \\) be the rescale factor, and \\( e(v_{i}) \\) be the expert for task \\( t_{i} \\). Then we have:" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.696, + 0.905, + 0.73 + ], + "angle": 0, + "content": "\\[\ne \\left(v _ {i}\\right) = \\mu_ {i} M \\left(v _ {i}, d\\right), \\mu_ {i} = - \\frac {\\mathbb {E} \\left(M \\left(v _ {i} , d\\right)\\right) \\cdot \\log (d)}{\\lambda_ {i} \\cdot \\mathbb {E} \\left(v _ {i}\\right)}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.736, + 0.906, + 0.765 + ], + "angle": 0, + "content": "The goal is to maintain a consistent output after extraction to preserve performance. Details are in Appendix D.5." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.906, + 0.81 + ], + "angle": 0, + "content": "In summary, we efficiently extract task-specific experts to integrate into the backbone, mitigating task conflicts without excessive storage or inference overhead." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.818, + 0.671, + 0.835 + ], + "angle": 0, + "content": "5.4. FREE-Merging" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "This section presents FREE-Merging, combining both parts to make it suitable for multiple tasks without affecting inference speed. Inspired by MoE [39], a router dynamically assigns tasks to activated experts. The router can take" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3866" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.099, + 0.09, + 0.898, + 0.105 + ], + "angle": 0, + "content": "Table 2. Comparison of model merging performance using ViT-B/32 (B) and ViT-L/14 (L) on eight visual tasks evaluating by accuracy." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.11, + 0.907, + 0.311 + ], + "angle": 0, + "content": "
MethodBase ModelSUN397CarsRESISC45EuroSATSVHNGTSRBMNISTDTDAvg.
BLBLBLBLBLBLBLBLBL
Individual75.382.377.792.496.197.499.710097.598.198.799.299.799.779.484.190.594.2
Traditional MTL73.980.874.490.693.996.398.296.395.897.698.999.199.599.677.984.488.993.5
Weight Averaging65.372.163.481.671.482.671.791.964.278.252.870.787.597.150.162.865.879.6
Fisher Merging [33]68.669.269.288.670.787.566.493.572.980.651.174.887.993.359.970.068.382.2
Task Arithmetic [21]63.874.162.182.172.086.777.693.874.487.965.186.894.098.952.265.670.184.5
Ties-Merging [48]64.876.562.985.074.389.378.995.783.190.371.483.397.699.056.268.873.686.0
Breadcrumbs [7]63.074.961.084.975.788.684.795.483.489.876.490.198.199.257.968.275.086.4
PCB-Merging [12]65.575.864.186.078.188.680.296.084.788.077.190.998.099.158.470.075.886.9
FR-Merging(ours)66.276.464.587.077.290.290.196.885.492.082.392.898.599.360.071.578.188.3
AdaMerging++ [52]66.679.468.390.382.291.694.297.489.693.489.097.598.399.060.679.281.191.0
Twin-Merging [29]73.682.671.790.392.194.999.399.495.396.797.298.099.199.474.080.187.892.7
EMR-Merging [20]74.182.372.790.891.995.399.499.595.896.996.998.199.199.572.180.187.792.8
FREE-Merging(ours)77.183.578.292.493.496.299.599.696.398.198.298.799.599.775.481.789.793.7
" + }, + { + "type": "title", + "bbox": [ + 0.094, + 0.335, + 0.373, + 0.35 + ], + "angle": 0, + "content": "Algorithm 1 Workflow of FREE-Merging" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.355, + 0.482, + 0.386 + ], + "angle": 0, + "content": "Input: Task vectors \\(\\{v_{i}\\}_{i = 1}^{K}\\), pre-trained model \\(\\theta_{pre}\\), router \\(R\\), and input data \\(X\\)." + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.388, + 0.216, + 0.4 + ], + "angle": 0, + "content": "1: FR-Merging:" + }, + { + "type": "text", + "bbox": [ + 0.378, + 0.389, + 0.477, + 0.399 + ], + "angle": 0, + "content": "\\(\\triangleright\\) Only excute once." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.4, + 0.482, + 0.431 + ], + "angle": 0, + "content": "2: Get backbone \\(\\theta_{m} = \\theta_{pre} + \\sum_{i=1}^{K} \\lambda_{i} G(v_{i})\\), where \\(G\\) and \\(\\lambda\\) are defined by Eq. 1 and Eq. 3, respectively." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.433, + 0.254, + 0.446 + ], + "angle": 0, + "content": "3: Expert Extraction:" + }, + { + "type": "text", + "bbox": [ + 0.378, + 0.433, + 0.477, + 0.445 + ], + "angle": 0, + "content": "\\(\\triangleright\\) Only execute once." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.446, + 0.343, + 0.462 + ], + "angle": 0, + "content": "4: Get task experts \\(\\{e_i\\}_{i = 1}^K\\) as Eq. 4." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.463, + 0.194, + 0.475 + ], + "angle": 0, + "content": "5: Inference:" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.477, + 0.214, + 0.49 + ], + "angle": 0, + "content": "6: for \\( x \\in X \\) do" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.491, + 0.341, + 0.507 + ], + "angle": 0, + "content": "7: \\([w_{1},\\dots w_{K}]\\gets argmax(R(x))\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.507, + 0.288, + 0.524 + ], + "angle": 0, + "content": "8: \\(\\theta_{*} = \\theta_{m} + \\sum_{i = 1}^{K}w_{i}e_{i}\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.524, + 0.262, + 0.539 + ], + "angle": 0, + "content": "9: \\(Y\\gets Y\\cup f(x;\\theta_{*})\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.54, + 0.174, + 0.552 + ], + "angle": 0, + "content": "10: end for" + }, + { + "type": "list", + "bbox": [ + 0.102, + 0.388, + 0.482, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.554, + 0.306, + 0.569 + ], + "angle": 0, + "content": "Output: Output \\( Y \\) for input \\( X \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.644 + ], + "angle": 0, + "content": "various forms [24], allowing for more flexibility in handling different tasks. Detailed discussions about router can be found in Appendix D.6. For detailed algorithms, see Alg 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.69 + ], + "angle": 0, + "content": "First, we obtain the merged backbone (line 2) and the task experts (line 4). Then, during the inference, we dynamically route the experts based on inputs (lines 7-9)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.698, + 0.476, + 0.715 + ], + "angle": 0, + "content": "5.5. Time Complexity and Storage Space Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.78 + ], + "angle": 0, + "content": "This section analyzes the merging time complexity and storage requirements. We aim to minimize both for practical use. Assume there are \\( n \\) models to merge, each with \\( m \\) parameters and requiring storage space \\( s \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.902 + ], + "angle": 0, + "content": "The time complexity includes the merging and inference part. In the merging part, FR-Merging has a complexity of \\( O(nmlogm) \\) [14] while expert extraction is \\( O(nm) \\), making the overall merging complexity \\( O(nmlogm) \\), which is manageable. For instance, merging three 10B models requires only \\( O(10^{12}) \\) addition operations, a minor load for modern devices. Moreover, with only a lightweight router added, the extra inference cost is negligible." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.335, + 0.907, + 0.426 + ], + "angle": 0, + "content": "Next, we analyze storage cost. We typically save only about \\(1\\%\\) additional parameters per model, requiring \\((1 + 0.01n)s\\) storage, which is far less than the requirement of \\(ns\\) for the model ensemble. For large models, reducing storage requirements is crucial for alleviating storage pressure on edge users. For detailed analysis, please refer to Sec 6.7." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.438, + 0.646, + 0.455 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.463, + 0.907, + 0.644 + ], + "angle": 0, + "content": "In this section, we evaluate the proposed methods across various tasks. We compare FR-Merging with other cost-free methods (which do not require additional training or data), including Weight Averaging, Fisher Merging [33], Task Arithmetic [21], Ties-Merging [48], Bubbles [7], and PCB-Merging [12]. We also compare our FREE-Merging with popular methods, including RegMean [23], AdaMerging [52], EMR-Merging [20], and Twin-Merging [29]. The original EMR-Merging uses a perfect router; however, we use an imperfect router for realism. Some experiments include traditional MTL with data from all tasks as a baseline. The evaluation covers CV, NLP, and multi-modal tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.653, + 0.729, + 0.669 + ], + "angle": 0, + "content": "6.1. Merging Vision Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.907, + 0.78 + ], + "angle": 0, + "content": "Merging 8 ViTs. We select the two visual encoders of CLIP, ViT-B/32 and ViT-L/14 [37], as the pre-trained models and choose eight image classification tasks as target tasks: SUN397, Cars, RESISC45, EuroSAT, SVHN, GT-SRB, MNIST, and DTD [52]. Additionally, an MLP is used as the lightweight router. Detailed information on the datasets and baselines can be found in the Appendix C.1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Table 2 presents the comparative results. First, FR-Merging outperforms other cost-free methods, improving ViT-B/32 by \\(2.5\\%\\) and ViT-L/14 by \\(1.5\\%\\) over PCB-Merging [12], demonstrating its effectiveness in reducing performance loss caused by task interference. Through FR-Merging, we can construct a merged backbone that can be used independently at a low cost. Here, we focus on the basic FR-Merging. In Sec. 6.6, we discuss further im" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3867" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.133 + ], + "angle": 0, + "content": "Table 3. Comparison of model merging results using ViT-B/16 on 30 visual tasks. Additional Cost indicates whether additional training or additional data is required during model merging." + }, + { + "type": "table", + "bbox": [ + 0.119, + 0.136, + 0.455, + 0.306 + ], + "angle": 0, + "content": "
MethodAdditional CostAvg. Acc
IndividualX93.02
Weight AveragingX42.51
Task Arithmetic [21]X48.88
Ties-Merging [48]X37.53
Breadcrumbs [7]X43.57
FR-Merging(ours)X53.90
RegMean [23]✓68.13
AdaMerging [52]✓60.24
Twin-Merging [29]✓75.52
EMR-Merging [20]✓76.39
FREE-Merging(ours)✓79.67
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.322, + 0.485, + 0.365 + ], + "angle": 0, + "content": "Table 4. Language model merging results: RoBERTa fine-tuned on 8 discriminative tasks, T0-3B on 11 discriminative tasks with \\( IA^3 \\), and Qwen-14B on 3 generative tasks with LoRA." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.369, + 0.489, + 0.564 + ], + "angle": 0, + "content": "
MethodBase ModelFull RoBERTaPEFTAvg.
T0-3BQwen-14B
Individual85.5571.3572.0776.32
Weight Averaging51.3458.0166.8058.71
Task Arithmetic[21]66.6563.9166.4065.65
Ties-Merging [48]64.0466.4167.4665.97
Breadcrumbs [7]68.1954.6466.7763.20
PCB-Merging [12]67.8666.1066.6966.39
FR-Merging(ours)70.0266.8868.0068.30
RegMean [23]70.0158.0167.5265.18
EMR-Merging [20]74.2067.1170.9870.76
Twin-Merging [29]78.2966.7071.6872.22
FREE-Merging(ours)80.1668.6872.7873.87
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.589, + 0.483, + 0.71 + ], + "angle": 0, + "content": "provements achieved by combining it with existing methods. Then, compared to methods requiring extra information, FREE-Merging also improves performance by \\(2\\%\\), leveraging lightweight computation and efficient expert extraction, requiring only \\(1\\%\\) additional storage per model, outperforming \\(3\\%\\) of EMR-Merging [20] and \\(2\\%\\) of Twin-Merging [29]. Thus, FREE-Merging provides an excellent balance between storage, computation, and performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.714, + 0.484, + 0.76 + ], + "angle": 0, + "content": "Merging 30 ViTs. We select ViT-B/16 as the pre-trained model and evaluate the model merging performance across 30 image classification tasks following [20]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Table 3 presents the results, with detailed analysis provided in Appendix E.1. It is evident that our proposed FR-Merging and FREE-Merging exhibit remarkable performance. Although merging a large number of models is uncommon in practice, this experiment underscores the effectiveness of Fourier filtering and lightweight task experts in mitigating performance loss caused by task interference. These results highlight the robustness and broad applicability of our methods across diverse scenarios." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.907, + 0.133 + ], + "angle": 0, + "content": "Table 5. Comparison of performance of different model merging methods after full fine-tuning LLaMa2-13B on three tasks: Instruction Following (LM), Math, and Code Generation." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.142, + 0.897, + 0.312 + ], + "angle": 0, + "content": "
MethodAlpacaEvalGSM8KMBPPAvg.
LM Fine-tuned [46]88.945.931.455.4
Math Fine-tuned [30]22.363.523.036.5
Code Fine-tuned [31]16.218.827.020.7
Task Arithmetic [21]72.148.3040.1
TIES-Merging [48]77.566.927.257.2
DARE [53]77.562.729.456.5
DELLA-Merging [8]80.461.831.457.9
PCB-Merging [12]81.260.630.757.5
FR-Merging(ours)87.564.332.861.5
Twin-Merging [29]83.561.430.458.4
EMR-Merging [20]82.662.331.258.7
FREE-Merging(ours)88.165.731.861.8
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.333, + 0.757, + 0.349 + ], + "angle": 0, + "content": "6.2. Merging Language Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.355, + 0.907, + 0.431 + ], + "angle": 0, + "content": "Merging Medium-sized Language Models. We investigate the generalization ability of our method on medium-scale language models, using RoBERTa [59] as the pretrained model with full fine-tuning on eight discriminative tasks [48]. Details can be found in Appendix C.1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.909, + 0.552 + ], + "angle": 0, + "content": "The average merging results of the eight tasks are presented in Table 4, with complete results in Appendix E.2. It can be observed that, FR-Merging outperforms other cost-free methods by nearly \\(2\\%\\). Additionally, FREE-Merging also exhibits remarkable effectiveness, achieving optimal results on almost all datasets and surpassing current methods in average performance. This highlights the great potential of our proposed methods for language models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.552, + 0.906, + 0.628 + ], + "angle": 0, + "content": "Merging PEFT models. PEFT is a popular method that reduces fine-tuning costs [10], but storing PEFT parameters for multiple tasks remains space-inefficient [55]. The effectiveness of our approach on PEFT demonstrates the generalization capability and offers an efficient deployment way." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.629, + 0.907, + 0.78 + ], + "angle": 0, + "content": "Firstly, we use T0-3B [38] as the pre-trained model and \\((IA)^{3}\\) [26] as the PEFT method, fine-tuning on 11 discriminative tasks [20]. Our proposed FR-Merging and FREE-Merging improve performance by nearly \\(1\\%\\), as shown in Table 4. Next, we employ Qwen-14B [2] with LoRA [19] as the PEFT way, fine-tuning on three generative tasks [29]. The results show that our methods lead to a \\(2\\%\\) advantage in both cost-free and full frameworks. This confirms the feasibility with PEFT and strong generalization of our methods. Full results are available in the Appendix E.2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.87 + ], + "angle": 0, + "content": "Merging Large Language models. We explore large model merging using LLaMa2-13B [40] as the pre-trained model, with WizardLM [46] for instruction following, WizardMath [30] for math, and WizardCoder-Python [31] for coding. Experiments are conducted on AlpacaEval (instruction following), GSM8K (math), and MBPP (code) [53]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Table 5 shows that our proposed FR-Merging even surpasses existing methods that rely on additional information." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3868" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.243, + 0.09, + 0.754, + 0.104 + ], + "angle": 0, + "content": "Table 6. Results of merging multi-modal BEiT3 models on five vision-language tasks." + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.115, + 0.904, + 0.26 + ], + "angle": 0, + "content": "
MethodsTask MetricCOCO-Retrieval Accuracy(↑)COCO-CaptioningImageNet-1k Classification Accuracy(↑)NLVR2 Accuracy(↑)VQAv2 Accuracy(↑)
BLEU4(↑)CIDEr(↑)METEOR(↑)ROUGE-L(↑)
Individual0.84560.3941.3370.3110.6010.85370.77650.8439
Weight Averaging0.18930.0310.0010.1150.1590.67710.28000.6285
Task Arithmetic [21]0.31770.0330.0000.1180.1760.67320.38090.6933
Ties-Merging [48]0.39290.0290.0010.1080.1670.69780.32060.6717
Breadcrumbs [7]0.18930.0110.0120.0650.1140.63230.60730.6823
FR-Merging(ours)0.70100.0160.0190.0960.1410.69920.64070.6965
EMR-Merging [20]0.79460.2891.0600.2720.5340.77420.74750.7211
Twin-Merging [29]0.79370.3251.0780.2780.5580.77300.77620.7232
FREE-Merging(ours)0.80930.3471.1950.2890.5720.78500.81680.7297
" + }, + { + "type": "image", + "bbox": [ + 0.105, + 0.281, + 0.462, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.499, + 0.483, + 0.542 + ], + "angle": 0, + "content": "Figure 4. Visualization of the loss landscape on EuroSAT before and after high-pass filtering (top vs. bottom) in the model merging of the fine-tuned ViT-B/32 on EuroSAT and DTD tasks [21]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.572, + 0.483, + 0.664 + ], + "angle": 0, + "content": "This indicates that our findings in FR-Merging regarding low-frequency signals and task interference also hold when extended to large models. This is crucial for practical usage of large models. Additionally, our proposed FREE-Merging also demonstrates strong advantages, enabling efficient cross-domain deployment of large models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.678, + 0.357, + 0.695 + ], + "angle": 0, + "content": "6.3. Merging Multi-Modal Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.703, + 0.483, + 0.809 + ], + "angle": 0, + "content": "We merge various fine-tuned BEiT3-base [42] models across five tasks: ImageNet-1k (Classification), VQAv2 (Visual Question Answering), NLVR2 (Reasoning), COCO Captioning (Image Captioning), and COCO Retrieval (Image-Text Retrieval). COCO Captioning is evaluated with BLEU4, CIDEr, METEOR, and ROUGE-L, while the others use accuracy [20]. Details are in Appendix C.1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Table 6 shows that our methods gain great improvements in multi-modal models. Unlike vision or language model merging, which often involve similar tasks, such as classification, the task similarity here is relatively low. The strong performance highlights the generalization ability of our methods, making them widely applicable." + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.275, + 0.905, + 0.289 + ], + "angle": 0, + "content": "Table 7. Ablation of FREE-Merging components on 8 visual tasks." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.299, + 0.903, + 0.394 + ], + "angle": 0, + "content": "
FR-MergingEqua.3Top-K ExpertScalingViT-B/32ViT-B/16ViT-L/14
XXXX70.1075.2484.50
✓XXX76.9681.3488.21
X✓XX71.4377.4184.20
✓✓XX78.1082.5488.30
XX✓X65.2468.9870.13
✓✓✓X79.3884.3189.53
XX✓✓82.3286.8790.47
✓✓✓✓89.6891.0993.70
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.416, + 0.766, + 0.432 + ], + "angle": 0, + "content": "6.4. Fourier Transform Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.439, + 0.907, + 0.621 + ], + "angle": 0, + "content": "In this section, we explain why Fourier high-pass filtering improves merging performance from an optimization perspective. Previous work [25] shows that landscape geometry impacts model generalization. Fig. 4 demonstrates that high-pass filtering widens the valley in the loss landscape (bottom plot). Since fine-tuned models from the same pretrained model are closer in the landscape [34], a wider valley increases the probability of merged results falling within the loss basin as shown in the bottom plot of Fig. 4, leading to better performance. Without filtering, the landscape is flat at a high level, which has little effect on model merging and only highlights the advantage of pre-training [17]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.633, + 0.667, + 0.65 + ], + "angle": 0, + "content": "6.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.657, + 0.907, + 0.702 + ], + "angle": 0, + "content": "In this section, we conduct ablation experiments on the components of our method. All experiments are performed on ViT-B/32 with eight classification tasks as target tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.906, + 0.839 + ], + "angle": 0, + "content": "Ablation on FR-Merging. We analyze the impact of FR-Merging on the backbone and overall performance. Table 7 shows that the performance of various models greatly improves after employing FR-Merging. We then compare high-pass filtering with other cost-free methods, including Ties [48], Top-K, Task Arithmetic [21], outlier dropout [7], EMR-Merging [20], and PCB-Merging [12], as well as low-pass and band-pass filtering, with results in Fig. 5a. For fairness, all experts use \\(1\\%\\) additional parameters." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Our proposed FR-Merging outperforms other cost-free methods on both the backbone and with task experts, minimizing task interference. Examining the different filtering methods, we find that low-pass filtering leads to great per" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3869" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.103, + 0.357, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.151, + 0.271, + 0.305, + 0.284 + ], + "angle": 0, + "content": "(a) Ablation on FR-Merging." + }, + { + "type": "image", + "bbox": [ + 0.372, + 0.101, + 0.631, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.423, + 0.271, + 0.576, + 0.284 + ], + "angle": 0, + "content": "(b) Ablation on Task Expert." + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.093, + 0.9, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.717, + 0.271, + 0.824, + 0.284 + ], + "angle": 0, + "content": "(c) Router Analysis." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.293, + 0.908, + 0.322 + ], + "angle": 0, + "content": "Figure 5. a) The ablation experiments on FR-Merging show that our proposed methods achieve optimal performance. b) Performance of experts at different sparsity rates. c) In experiments with random routers, our approach demonstrates optimal performance." + }, + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.342, + 0.482, + 0.356 + ], + "angle": 0, + "content": "Table 8. Transferability analysis of FR-Merging on 8 visual tasks." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.359, + 0.469, + 0.461 + ], + "angle": 0, + "content": "
MethodFR-MergingViT-B/32ViT-L/14
Ties-Merging [48]X73.686.0
Ties-Merging [48]✓78.388.4
Breadcrumbs [7]X75.086.4
Breadcrumbs [7]✓78.888.9
TSVM [16]X83.991.9
TSVM [16]✓84.892.5
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.477, + 0.483, + 0.552 + ], + "angle": 0, + "content": "formance degradation, confirming that task interference in low-frequency regions is severe. Band-pass filtering also performs worse than FR-Merging, and this is because some high-frequency signals preserve abrupt changes in the signal and their removal harms the performance of the model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.552, + 0.483, + 0.628 + ], + "angle": 0, + "content": "Ablation on Merging Coefficients. We conduct an ablation study on our merging coefficients as in Eq. 3, with results in Table 7. Using them instead of simple averaging yields slight improvements by considering output variations, benefiting tasks with varying fine-tuning magnitudes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.483, + 0.765 + ], + "angle": 0, + "content": "Ablation on Task Expert. We conduct an ablation study on the expert extraction way. Table 7 shows that incorporating our experts and rescaling method results in significant improvements, confirming their effectiveness. To validate the superiority of our method, we evaluate expert-only performance with varying parameter proportions, comparing Top-K selection, Bernoulli selection [53], SVD extraction [29], and our Top-K selection with rescaling. Fig. 5b shows that our method excels, particularly under high sparsity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Router Analysis. We analyze the proposed router method, aiming for strong merging performance despite sub-optimal routing. Since router inaccuracies may arise from limited training resources or edge deployment issues, our methods maintain performance in real-world scenarios. Using a random router to simulate inaccuracy, we compare our method with existing router-based approaches. Our method outperforms Twin-Merging by \\(13.5\\%\\) and EMR-Merging by \\(8.5\\%\\) on average, as shown in Fig. 5c, indicating that our methods" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.342, + 0.906, + 0.37 + ], + "angle": 0, + "content": "Table 9. Trade-off of performance, training, inference, and storage costs using ViT-L/14 across 8 image classification tasks." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.374, + 0.906, + 0.46 + ], + "angle": 0, + "content": "
MethodTraining TokensTraining CostStorage Param.Inference Cost (/1000 items)Acc.
Individual002.4B31.0s94.2
MTL304M12h304M31.0s93.5
FR-Merging00304M31.0s88.2
Twin-Merging0.9M147s352M32.2s92.7
FREE-Merging0.9M147s328M32.2s93.7
" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.478, + 0.903, + 0.493 + ], + "angle": 0, + "content": "construct a reasonable backbone and generalizable experts." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.503, + 0.736, + 0.519 + ], + "angle": 0, + "content": "6.6. Transferability Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.526, + 0.906, + 0.617 + ], + "angle": 0, + "content": "Our proposed FR-Merging can integrate with other methods, as it addresses frequency-domain task interference that current methods overlook. As shown in Table 8, combining various methods with FR-Merging leads to performance improvements. Therefore, FR-Merging can be integrated with more approaches to achieve greater benefits." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.627, + 0.813, + 0.644 + ], + "angle": 0, + "content": "6.7. Speed and Storage Space Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.649, + 0.906, + 0.741 + ], + "angle": 0, + "content": "This part analyzes training costs, inference speed, and storage as shown in Table 9. FR-Merging achieves strong performance with significantly reducing storage costs without extra training or inference costs, while FREE-Merging slightly increases overhead for better performance. Both of these two methods enhance deep learning deployment." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.755, + 0.634, + 0.77 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.902 + ], + "angle": 0, + "content": "In this paper, we propose FR-Merging and FREE-Merging to efficiently build a multi-task model by merging existing ones. FR-Merging constructs a backbone without training by high-pass filtering to remove information that dominates task interference. Meanwhile, FREE-Merging employs lightweight experts to offset information loss during merging. Our method balances training, storage, inference costs, and performance, showing high practical value." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3870" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.484, + 0.176 + ], + "angle": 0, + "content": "This work is supported by National Natural Science Foundation of China (NSFC) (62232005, 62202126); the National Key Research and Development Program of China (2021YFB3300502)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.189, + 0.188, + 0.205 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.214, + 0.484, + 0.269 + ], + "angle": 0, + "content": "[1] Samuel Ainsworth, Jonathan Hayase, and Siddhartha Srinivasa. Git re-basin: Merging models modulo permutation symmetries. In The Eleventh International Conference on Learning Representations, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.271, + 0.484, + 0.325 + ], + "angle": 0, + "content": "[2] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.327, + 0.484, + 0.38 + ], + "angle": 0, + "content": "[3] Tim-Oliver Buchholz and Florian Jug. Fourier image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1846-1854, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.383, + 0.483, + 0.424 + ], + "angle": 0, + "content": "[4] Shijie Chen, Yu Zhang, and Qiang Yang. Multi-task learning in natural language processing: An overview. ACM Computing Surveys, 56(12):1-32, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.426, + 0.484, + 0.495 + ], + "angle": 0, + "content": "[5] Tianyi Chu, Jiafu Chen, Jiakai Sun, Shuobin Lian, Zhizhong Wang, Zhiwen Zuo, Lei Zhao, Wei Xing, and Dongming Lu. Rethinking fast fourier convolution in image inpainting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23195-23205, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.496, + 0.484, + 0.564 + ], + "angle": 0, + "content": "[6] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instructionfinetuned language models. Journal of Machine Learning Research, 25(70):1-53, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.484, + 0.62 + ], + "angle": 0, + "content": "[7] MohammadReza Davari and Eugene Belilovsky. Model breadcrumbs: Scaling multi-task model merging with sparse masks. arXiv preprint arXiv:2312.06795, 2023. 3, 5, 6, 7, 8, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.622, + 0.484, + 0.676 + ], + "angle": 0, + "content": "[8] Pala Tej Deep, Rishabh Bhardwaj, and Soujanya Poria. Della-merging: Reducing interference in model merging through magnitude-based sampling. arXiv preprint arXiv:2406.11617, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.678, + 0.484, + 0.789 + ], + "angle": 0, + "content": "[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota, 2019. Association for Computational Linguistics. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.484, + 0.858 + ], + "angle": 0, + "content": "[10] Ning Ding, Yujia Qin, Guang Yang, Fuchao Wei, Zonghan Yang, Yusheng Su, Shengding Hu, Yulin Chen, Chi-Min Chan, Weize Chen, et al. Parameter-efficient fine-tuning of large-scale pre-trained language models. Nature Machine Intelligence, 5(3):220-235, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[11] Xibin Dong, Zhiwen Yu, Wenming Cao, Yifan Shi, and Qianli Ma. A survey on ensemble learning. Frontiers of Computer Science, 14:241-258, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.214, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.162 + ], + "angle": 0, + "content": "[12] Guodong Du, Junlin Lee, Jing Li, Runhua Jiang, Yifei Guo, Shuyang Yu, Hanting Liu, Sim Kuan Goh, Ho-Kin Tang, Daojing He, et al. Parameter competition balancing for model merging. arXiv preprint arXiv:2410.02396, 2024. 2, 3, 5, 6, 7, 8, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.906, + 0.234 + ], + "angle": 0, + "content": "[13] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.236, + 0.906, + 0.278 + ], + "angle": 0, + "content": "[14] Pierre Duhamel and Martin Vetterli. Fast fourier transforms: a tutorial review and a state of the art. Signal processing, 19 (4):259-299, 1990. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.281, + 0.908, + 0.336 + ], + "angle": 0, + "content": "[15] Mudasir A Ganaie, Minghui Hu, Ashwani Kumar Malik, Muhammad Tanveer, and Ponnuthurai N Suganthan. Ensemble deep learning: A review. Engineering Applications of Artificial Intelligence, 115:105151, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.339, + 0.906, + 0.408 + ], + "angle": 0, + "content": "[16] Antonio Andrea Gargiulo, Donato Crisostomi, Maria Sofia Bucarelli, Simone Scardapane, Fabrizio Silvestri, and Emanuele Rodola. Task singular vectors: Reducing task interference in model merging. arXiv preprint arXiv:2412.00081, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.411, + 0.906, + 0.467 + ], + "angle": 0, + "content": "[17] Timur Garipov, Pavel Izmailov, Dmitrii Podoprikhin, Dmitry P Vetrov, and Andrew G Wilson. Loss surfaces, mode connectivity, and fast ensembling of dnns. Advances in neural information processing systems, 31, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.906, + 0.51 + ], + "angle": 0, + "content": "[18] Rafael C Gonzales and Paul Wintz. Digital image processing. Addison-Wesley Longman Publishing Co., Inc., 1987. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.513, + 0.906, + 0.569 + ], + "angle": 0, + "content": "[19] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.571, + 0.906, + 0.627 + ], + "angle": 0, + "content": "[20] Chenyu Huang, Peng Ye, Tao Chen, Tong He, Xi-angyu Yue, and Wanli Ouyang. Emr-merging: Tuning-free high-performance model merging. arXiv preprint arXiv:2405.17461, 2024. 1, 2, 4, 5, 6, 7, 3, 8, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.63, + 0.906, + 0.686 + ], + "angle": 0, + "content": "[21] Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Suchin Gururangan, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. Editing models with task arithmetic. arXiv preprint arXiv:2212.04089, 2022. 1, 2, 3, 5, 6, 7, 4, 8, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.688, + 0.906, + 0.757 + ], + "angle": 0, + "content": "[22] Dongfu Jiang, Xiang Ren, and Bill Yuchen Lin. Lmblender: Ensembling large language models with pairwise ranking and generative fusion. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 14165-14178, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.759, + 0.906, + 0.815 + ], + "angle": 0, + "content": "[23] Xisen Jin, Xiang Ren, Daniel Preotiac-Pietro, and Pengxiang Cheng. Dataless knowledge fusion by merging weights of language models. arXiv preprint arXiv:2212.09849, 2022. 1, 2, 5, 6, 3, 8, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.906, + 0.899 + ], + "angle": 0, + "content": "[24] Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan Firat, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. Gshard: Scaling giant models with conditional computation and automatic sharding. In International Conference on Learning Representations, 2021. 5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3871" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[25] Hao Li, Zheng Xu, Gavin Taylor, Christoph Studer, and Tom Goldstein. Visualizing the loss landscape of neural nets. Advances in neural information processing systems, 31, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.148, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[26] Haokun Liu, Derek Tam, Mohammed Muqeeth, Jay Mohta, Tenghao Huang, Mohit Bansal, and Colin A Raffel. Few-shot parameter-efficient fine-tuning is better and cheaper than in-context learning. Advances in Neural Information Processing Systems, 35:1950-1965, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.482, + 0.273 + ], + "angle": 0, + "content": "[27] James Liu, Guangxuan Xiao, Kai Li, Jason D Lee, Song Han, Tri Dao, and Tianle Cai. Bitdelta: Your fine-tune may only be worth one bit. Advances in Neural Information Processing Systems, 37:13579-13600, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.356 + ], + "angle": 0, + "content": "[28] Yajing Liu, Yuning Lu, Hao Liu, Yaozu An, Zhuoran Xu, Zhuokun Yao, Baofeng Zhang, Zhiwei Xiong, and Chenguang Gui. Hierarchical prompt learning for multi-task learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10888-10898, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.358, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[29] Zhenyi Lu, Chenghao Fan, Wei Wei, Xiaoye Qu, Dangyang Chen, and Yu Cheng. Twin-merging: Dynamic integration of modular expertise in model merging. arXiv preprint arXiv:2406.15479, 2024. 1, 2, 5, 6, 7, 8, 3, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.482, + 0.495 + ], + "angle": 0, + "content": "[30] Haipeng Luo, Qingfeng Sun, Can Xu, Pu Zhao, Jianguang Lou, Chongyang Tao, Xiubo Geng, Qingwei Lin, Shifeng Chen, and Dongmei Zhang. Wizardmath: Empowering mathematical reasoning for large language models via reinforced evol-instruct. arXiv preprint arXiv:2308.09583, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.482, + 0.566 + ], + "angle": 0, + "content": "[31] Ziyang Luo, Can Xu, Pu Zhao, Qingfeng Sun, Xiubo Geng, Wenxiang Hu, Chongyang Tao, Jing Ma, Qingwei Lin, and Daxin Jiang. Wizardcoder: Empowering code large language models with evol-instruct. In The Twelfth International Conference on Learning Representations, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.482, + 0.621 + ], + "angle": 0, + "content": "[32] Xinyin Ma, Gongfan Fang, and Xinchao Wang. Llm-pruner: On the structural pruning of large language models. Advances in neural information processing systems, 36:21702-21720, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.622, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[33] Michael S Matena and Colin A Raffel. Merging models with fisher-weighted averaging. Advances in Neural Information Processing Systems, 35:17703-17716, 2022. 1, 2, 5, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.482, + 0.705 + ], + "angle": 0, + "content": "[34] Behnam Neyshabur, Hanie Sedghi, and Chiyuan Zhang. What is being transferred in transfer learning? Advances in neural information processing systems, 33:512-523, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.482, + 0.762 + ], + "angle": 0, + "content": "[35] Tan Nguyen, Minh Pham, Tam Nguyen, Khai Nguyen, Stanley Osher, and Nhat Ho. Fourierformer: Transformer meets generalized fourier integral theorem. Advances in Neural Information Processing Systems, 35:29319-29335, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.763, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[36] Biqing Qi, Fangyuan Li, Zhen Wang, Junqi Gao, Dong Li, Peng Ye, and Bowen Zhou. Less is more: Efficient model merging with binary task switch. arXiv preprint arXiv:2412.00054, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 3, 4, 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[38] Victor Sanh, Albert Webson, Colin Raffel, Stephen Bach, Lintang Sutawika, Zaid Alyafeai, Antoine Chaffin, Arnaud Stiegler, Arun Raja, Manan Dey, et al. Multitask prompted training enables zero-shot task generalization. In International Conference on Learning Representations, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[39] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. In International Conference on Learning Representations, 2016. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[40] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[41] Simon Vandenhende, Stamatios Georgoulis, Wouter Van Gansbeke, Marc Proesmans, Dengxin Dai, and Luc Van Gool. Multi-task learning for dense prediction tasks: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(7):3614-3633, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.375, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[42] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, et al. Image as a foreign language: Beit pretraining for vision and vision-language tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19175-19186, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.475, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[43] T Wolf. Huggingface's transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.518, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[44] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In International conference on machine learning, pages 23965-23998. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.617, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[45] Chengyue Wu, Teng Wang, Yixiao Ge, Zeyu Lu, Ruisong Zhou, Ying Shan, and Ping Luo. \\(\\pi\\)-tuning: Transferring multimodal foundation models with optimal multi-task interpolation. In International Conference on Machine Learning, pages 37713-37727. PMLR, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.688, + 0.905, + 0.77 + ], + "angle": 0, + "content": "[46] Can Xu, Qingfeng Sun, Kai Zheng, Xiubo Geng, Pu Zhao, Jiazhan Feng, Chongyang Tao, Qingwei Lin, and Daxin Jiang. Wizardlm: Empowering large pre-trained language models to follow complex instructions. In The Twelfth International Conference on Learning Representations, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.772, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[47] Jing Xu and Jingzhao Zhang. Random masking finds winning tickets for parameter efficient fine-tuning. In *Forty-first International Conference on Machine Learning*. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.816, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[48] Prateek Yadav, Derek Tam, Leshem Choshen, Colin A Raffel, and Mohit Bansal. Ties-merging: Resolving interference when merging models. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3, 4, 5, 6, 7, 8, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.903 + ], + "angle": 0, + "content": "[49] Enneng Yang, Li Shen, Guibing Guo, Xingwei Wang, Xiaochun Cao, Jie Zhang, and Dacheng Tao. Model merging" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3872" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.133 + ], + "angle": 0, + "content": "in llms, mllms, and beyond: Methods, theories, applications and opportunities. arXiv preprint arXiv:2408.07666, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[50] Enneng Yang, Li Shen, Zhenyi Wang, Guibing Guo, Xiaojun Chen, Xingwei Wang, and Dacheng Tao. Representation surgery for multi-task model merging. arXiv preprint arXiv:2402.02705, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[51] Enneng Yang, Li Shen, Zhenyi Wang, Guibing Guo, Xiao-jun Chen, Xingwei Wang, and Dacheng Tao. Representation surgery for multi-task model merging. In *Forty-first International Conference on Machine Learning*, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[52] Enneng Yang, Zhenyi Wang, Li Shen, Shiwei Liu, Guibing Guo, Xingwei Wang, and Dacheng Tao. Adamerging: Adaptive model merging for multi-task learning. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 5, 6, 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.387 + ], + "angle": 0, + "content": "[53] Le Yu, Bowen Yu, Haiyang Yu, Fei Huang, and Yongbin Li. Language models are super mario: Absorbing abilities from homologous models as a free lunch. In *Forty-first International Conference on Machine Learning*, 2024. 2, 3, 4, 6, 8, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.39, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[54] Kerem Zaman, Leshem Choshen, and Shashank Srivastava. Fuse to forget: Bias reduction and selective memorization through model fusion. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 18763-18783, Miami, Florida, USA, 2024. Association for Computational Linguistics. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.475, + 0.482, + 0.529 + ], + "angle": 0, + "content": "[55] Jinghan Zhang, Junteng Liu, Junxian He, et al. Composing parameter-efficient modules with arithmetic operation. Advances in Neural Information Processing Systems, 36: 12589-12610, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.531, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[56] Yu Zhang and Qiang Yang. A survey on multi-task learning. IEEE transactions on knowledge and data engineering, 34 (12):5586-5609, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.574, + 0.482, + 0.642 + ], + "angle": 0, + "content": "[57] Yanqi Zhou, Tao Lei, Hanxiao Liu, Nan Du, Yanping Huang, Vincent Zhao, Andrew M Dai, Quoc V Le, James Laudon, et al. Mixture-of-experts with expert choice routing. Advances in Neural Information Processing Systems, 35:7103-7114, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.645, + 0.482, + 0.713 + ], + "angle": 0, + "content": "[58] Yuyan Zhou, Liang Song, Bingning Wang, and Weipeng Chen. Metagpt: Merging large language models using model exclusive task arithmetic. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 1711-1724, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.715, + 0.482, + 0.797 + ], + "angle": 0, + "content": "[59] Liu Zhuang, Lin Wayne, Shi Ya, and Zhao Jun. A robustly optimized BERT pre-training approach with post-training. In Proceedings of the 20th Chinese National Conference on Computational Linguistics, pages 1218-1227, Huhhot, China, 2021. Chinese Information Processing Society of China. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.797 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3873" + } + ] +] \ No newline at end of file diff --git a/2025/FREE-Merging_ Fourier Transform for Efficient Model Merging/f3aa67c5-53d3-463b-8730-22ad4f95fe7f_origin.pdf b/2025/FREE-Merging_ Fourier Transform for Efficient Model Merging/f3aa67c5-53d3-463b-8730-22ad4f95fe7f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..39b2493eacfd3432537e5018c891f52fa4538ba1 --- /dev/null +++ b/2025/FREE-Merging_ Fourier Transform for Efficient Model Merging/f3aa67c5-53d3-463b-8730-22ad4f95fe7f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:578b03fc661bc6eedffd877ea036fa79134484add3be03bd137f6516535d38d8 +size 1403732 diff --git a/2025/FREE-Merging_ Fourier Transform for Efficient Model Merging/full.md b/2025/FREE-Merging_ Fourier Transform for Efficient Model Merging/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5c626fcd5d561c0f0dc42da53163e932975ffefb --- /dev/null +++ b/2025/FREE-Merging_ Fourier Transform for Efficient Model Merging/full.md @@ -0,0 +1,362 @@ +# FREE-Merging: Fourier Transform for Efficient Model Merging + +Shenghe Zheng +Harbin Institute of Technology +shenghez.zheng@gmail.com + +Hongzhi Wang* +Harbin Institute of Technology +wangzh@hit.edu.cn + +# Abstract + +With the rapid growth of deep learning, there is an increasing availability of open-source models for various tasks. However, single fine-tuned models often fall short of meeting the diverse needs of users. Model merging has thus emerged as an efficient method to integrate the capabilities of existing models into a unified model. Nevertheless, existing model merging methods face challenging trade-offs between performance and deployment costs, primarily due to task interference. For the first time, we reveal that task interference is evident in the frequency domain of model parameters, yet current efforts only focus on spatial domain solutions, which are largely ineffective in addressing frequency domain interference. To mitigate the impact of frequency domain interference, we propose FR-Merging, an innovative method that effectively filters harmful frequency domain interference on the backbone with minimal computational overhead. Since performance loss is inevitable with cost-free methods, we propose a lightweight task-specific expert module that dynamically compensates for information loss during merging. This proposed framework, FREE-Merging (FR-Merging with experts), strikes a balanced trade-off between training cost, inference latency, storage requirements, and performance. We demonstrate the effectiveness of both FR-Merging and FREE-Merging on multiple tasks across CV, NLP, and Multi-Modal domains and show that they can be flexibly adapted to specific needs. Our code is available at: https://github.com/Zhengsh123/FREE-Merging. + +# 1. Introduction + +In the current era of deep learning, the pretrain-finetune paradigm is a standard procedure [9, 37, 40]. However, edge applications often lack powerful computing resources and data, favoring ready-to-use models for specific scenarios. The recent growth of open-source platforms like Hugging-Face [43] has made this feasible. Yet, practical demands + +are often not fully covered by a single existing model and are typically a union of several models [23, 52]. However, running separate models for each subtask increases storage and inference costs, especially with large models. Model merging attempts to combine existing models to build a model capable of addressing all target tasks [21, 33, 48]. This method reduces training costs, data privacy issues in multi-task learning (MTL) [20, 56] and deployment costs, garnering widespread attention. + +While model merging aims to reduce training costs, it often suffers from performance loss due to task interference [20]. To mitigate this issue, existing methods develop along three directions. The first direction computes merging coefficients based on weights or data [23, 33]. The second uses the parameter differences before and after finetuning (i.e. task vectors) for merging [21, 48, 52, 58]. The third method introduces independent knowledge for each task [20, 29]. While the third way balances performance and costs, it still faces two critical limitations. First, it neglects backbone optimization during merging, resulting in subpar performance. Second, it requires high computation and storage, making it impractical for edge deployment. + +We argue that an effective model merging method requires addressing both of the above issues. First, since the backbone constitutes the majority of parameters and thus determines performance, an efficient backbone merging method is essential to reduce performance loss. Then, given the no free lunch theorem, cost-free merging cannot avoid task interference. Therefore, introducing task-specific experts is necessary. However, they must remain lightweight to avoid storage and inference overhead. Thus, we propose FREE-Merging, a two-stage method that ensures an efficient backbone and lightweight experts as shown in Fig 1. + +For backbone merging, the key lies in eliminating task interference [51]. We identify a compelling phenomenon in which task interference is evident in the frequency domain of model parameters. And we present in Sec. 4 that frequency-domain interference is strongly correlated with performance. Notably, existing methods, which focus only on spatial-domain operations, fail to address frequency-domain interference, leading to suboptimal outcomes. Ex + +![](images/c1b7af58661b1fa1c1d29f024989d2bbdfd80d0cb91e6a2bd8d7578b6f72c1f7.jpg) +Figure 1. The workflow of FREE-Merging involves two main steps. First, FR-Merging utilizes high-pass filtering to remove harmful specialized information from each model, thereby constructing a high-quality merged backbone. Second, it employs lightweight extraction of task experts, which are dynamically added during inference to mitigate the impact of task interference. + +![](images/7394ca043757c5a26b29325fd0558863920099268d48ac200488707e554e3b53.jpg) + +![](images/53aa4ab86d7788d51c11954ead466d1c0288307f670260c7616f7550dd37db05.jpg) + +Experimental evidence reveals that task interference is severe in the low-frequency region. This occurs because, high-frequency signals represent fine-grained variations, while low-frequency parts capture the global structure [18], which is more likely to contain task-specific information that leads to task interference. Leveraging the inherent redundancy in fine-tuned parameters [27, 47, 53], we propose to directly filter out low-frequency parts with severe task interference. As detailed in Sec. 5.2, this high-pass filtering substantially enhances generalization with minimal performance loss, preventing task interference during merging. Our lightweight yet highly effective approach, FR-Merging, is the first to apply Fourier filtering to neural network parameters, showing its potential for broader applications. + +After obtaining a high-performance backbone through FR-Merging, we use a lightweight task-specific expert to recover the inevitable information loss with minimal storage. Unlike existing methods that store extensive knowledge, we rescale low-frequency signals, which store task-specific information, retaining task expertise with only $1\%$ of the parameters. Inspired by MoE [39, 57], we employ a router for dynamic expert routing, enhancing model flexibility. + +In summary, we introduce FREE-Merging, which utilizes FR-Merging to reduce task interference and incorporates lightweight experts to eliminate information loss of model merging. Our main contributions are: 1). We identify frequency-domain task interference as a critical factor that greatly impacts merging performance but remains challenging to address through spatial-domain methods. 2). To tackle frequency-domain interference, we propose FR-Merging, which reduces interference while preserving performance, constructing the merged backbone. 3). We theoretically validate the necessity of introducing new information during merging and propose an effective expert construction method. 4). Extensive experiments show that + +FR-Merging and FREE-Merging effectively enhance performance across vision, language, and multi-modal tasks. + +# 2. Related Work + +Model Merging. Model merging integrates existing models to handle multiple tasks without training [21, 23, 49], but faces task interference challenges [36, 48]. Simple averaging [44] causes great performance degradation. Methods like Fisher-Merging [33] and RegMean [23] use matrices to determine merging coefficients but are computationally expensive or data-intensive, limiting edge deployment. Task Arithmetic [21] merges task vectors instead of weights, while Ties-Merging [48] resolves parameter conflicts, and AdaMerging [52] automates coefficient selection. However, these approaches have data distribution requirements and limited applicability. Techniques such as DARE [53] and PCB-Merging [12] mitigate performance drops but still face task conflicts. EMR-Merging [20] and Twin-Merging [29] store task-specific knowledge but require large storage and neglect backbone optimization. Our method uses high-pass filtering and lightweight experts to balance merging costs and effectiveness across various tasks. For further details, please refer to Appendix C.2. + +Model Ensemble. Model ensemble combines outputs from multiple models [11, 15, 22], but with large models, it faces storage and inference challenges. In contrast, Model merging enables a single model to solve multiple tasks, significantly reducing storage and inference costs. + +Multi-task Learning. Multi-task learning (MTL) aims to solve multiple tasks using a single model [4, 28, 41], but in the era of large models, it faces challenges like high training costs, data privacy issues, and expertise requirements [52]. In contrast, model merging uses existing open-source models to create a multi-task model with little or no training, significantly reducing deployment costs. + +![](images/c735b87a90dd98bf2bf51f5bae2e5f1120d9374a56864f33f8f381651c3d0e21.jpg) + +![](images/ab7a92b21a70607cbd9d59de2649a84f37f062b33d1e32d56d05aaf436854fc8.jpg) +(a) +(b) +Figure 2. a) The frequency domain amplitude power distribution of an attention head across different fine-tuned ViT-B-32 models. b) The correlation between the merging performance of ViT-B-32 and the variance of frequency domain amplitude power. Normalized accuracy is defined as the ratio of merged model performance to fine-tuned model performance on the same task. + +# 3. Preliminaries + +Notation. Assume that $f_{\theta}(x) \to y$ represents a neural network parameterized by $\theta$ , where $x \in R^{m}$ is the input and $y \in R^{n}$ is the output. Recent studies have focused on model merging via task vectors. For task $k$ , the task vector is defined as $v_{k} = \theta_{k} - \theta_{pre}$ , where $\theta_{k}$ is the fine-tuned model for task $k$ , and $\theta_{pre}$ is the corresponding pre-trained model. Additionally, $e_{k}$ represents the task expert that retains the ability to solve the task $k$ with few parameters. + +Problem Definition. The objective of model merging is to combine models $\{\theta_k\}_{k=1}^K$ for tasks $T = \{t_i\}_{i=1}^K$ into a single model $\theta_m$ that can handle all tasks in $T$ . We focus on a practical scenario where a shared pre-trained model $\theta_{pre}$ serves as the base for all fine-tuned models. It is formulated as $\theta_m = \lambda \sum_{k=1}^K \hat{G}(\theta_k)$ , and with task vector, it becomes $\theta_m = \theta_{pre} + \lambda \Sigma_{k=1}^K G(v_k)$ , where $\lambda$ is the coefficient, and $\hat{G}/G$ is transformation on the parameters. Assuming a shared pre-trained model aligns with trends toward foundation models, like LLaMa [40] for language and CLIP [37] for vision, which are then fine-tuned for specific tasks. + +# 4. Task Interference in Frequency Domain + +Task interference is a key factor that leads to the decrease in merging performance. While current methods focus on addressing interference in the spatial domain [48, 50, 53], we observe that task vectors from different fine-tuned models also exhibit great misalignment in the frequency domain. + +As shown in Fig. 2(a), we visualize the amplitude power + +Table 1. The impact of current methods on the mean variance of frequency-domain amplitude across 8 image classification tasks. + +
MethodViT-B/32ViT-L/14
Task Arithmetic [21]0.0590.110
DARE [53]0.057(↓3%)0.108(↓2%)
Ties-Merging [48]0.058(↓2%)0.109(↓1%)
Breadcrumbs [7]0.058(↓2%)0.107(↓3%)
PCB-Merging [12]0.056(↓5%)0.107(↓3%)
FR-Merging(ours)0.045(↓24%)0.088(↓20%)
+ +distribution in the frequency domain of task vectors from a ViT-B-32 attention head fine-tuned on various tasks [21]. Here, greater distance from the center represents a higher frequency. Since amplitude indicates signal strength, merging two signals with vastly different strengths inevitably results in the weaker signal being masked. Therefore, task interference in the low-frequency part are severe. + +Furthermore, in Fig.2(b), we train multiple groups of models on eight image tasks [21] and find a clear negative correlation between the variance of amplitude power among models and the normalized merging performance (details can be found in Appendix D.1), confirming the great impact of frequency-domain interference on merging performance. + +Additionally, as shown in Table 1, we show that existing works addressing task interference in the spatial domain fail to effectively mitigate frequency-domain task interference. Therefore, we propose our approach to tackle this issue. + +# 5. Methodology + +This section presents FREE-Merging: Sec. 5.1 explains the need for a two-part design, Sec. 5.2 and Sec. 5.3 cover the FR-Merging for the backbone and the lightweight expert module, respectively, followed by the FREE-Merging in Sec. 5.4 and a complexity analysis in Sec. 5.5. + +# 5.1. Motivation + +We first analyze why both merging optimization for the backbone and the introduction of experts are necessary. + +First, backbone optimization is crucial but overlooked by current expert-based merging ways. As the backbone holds most parameters, its performance is key to the overall success. For instance, upgrading from LLaMa2 [40] to LLaMa3 [13] greatly improves results even with the identical fine-tuning. Thus, we propose a training-free merging method to reduce task interference in the backbone. + +Next, we present a theorem showing the necessity of task experts. Based on Theorem 5.1, introducing additional information is essential. For proof, see Appendix B, and the effectiveness of task experts is proved in Appendix D.4. + +Theorem 5.1 Model Merging has no free lunch. Merged model $\theta_{m}$ cannot simultaneously retain the capabilities of $\{\theta_k\}_{k=1}^K$ without introducing additional information. + +# 5.2. FR-Merging + +We propose an effective backbone merging method by addressing task interference in the low-frequency fine-tuned parameters as discussed in Sec. 4, which is hard to resolve in the spatial domain. Given the high redundancy of fine-tuned parameters [53], we apply high-pass filtering to directly remove the interfering regions. Let $v(x,y)$ be the task vector and $G(x,y)$ the transformed result. Then we have: + +$$ +G (x, y) = \mathcal {F} ^ {- 1} \{H (\eta , \gamma) \cdot \mathcal {F} \{v (x, y) \} \}, \tag {1} +$$ + +where + +$$ +H (\eta , \gamma) = \left\{ \begin{array}{l l} 1, & \sqrt {\eta^ {2} + \gamma^ {2}} \geq D _ {0} \\ 0, & \sqrt {\eta^ {2} + \gamma^ {2}} < D _ {0} \end{array} \right., \tag {2} +$$ + +where $\mathcal{F}$ and $\mathcal{F}^{-1}$ represent the Fourier and inverse Fourier transform, respectively. $D_0$ is the adjustable cutoff frequency. And $\eta$ and $\gamma$ are the distances from the frequency domain center along the $x$ - and $y$ -axes. + +Next, we analyze why high-pass filtering is effective from both intuitive and experimental perspectives. Fine-tuned weights occupy distinct positions in the loss landscape, and linear interpolation often leads to high-loss regions, causing task interference [1, 50]. To address this, we minimize model differences while preserving performance, bringing them closer in the loss landscape to increase the likelihood of the merged model falling into a loss basin [45]. + +Results in Sec. 4 show great differences in low-frequency regions. This is because, if we consider neural network parameters as signals carrying training information, inspired by image processing [3, 35], high-frequency signals represent sharp variations, while low-frequency parts define the overall structure [5]. Since most information resides in the structure, filtering out low-frequency signals removes key task-specific information. Then, the likelihood that the interpolation falls within the loss basin increases [1]. + +Experimentally, we validate that high-pass filtering enhances model generalization with minimal performance loss as shown in Fig 3. We fine-tune ViT-B/16 [37] on 30 visual tasks following [20]. We find that removing low-frequency information results in a smaller performance drop (diagonal) compared to the improvement in generalization (off-diagonal). This indicates that low-frequency parts carry task-specific information that boosts task performance but reduces generalization. Thus, applying filtering like FR-Merging helps alleviate task interference during merging. + +To maintain consistent output after merging, we define the merging coefficient $\lambda_{i}$ for the task vector $v_{i}$ as: + +$$ +\lambda_ {i} = \mathbb {E} (v _ {i}) \left(\sum_ {j = 1} ^ {K} \mathbb {E} (v _ {j})\right) ^ {- 1}, \tag {3} +$$ + +where $\mathbb{E}$ represents the average. Overall, we construct a merged backbone in this part that minimizes task conflicts. + +![](images/66191ca272013351a4299d6a43952fbf48fa82abb40d63ac0e1a23a9b84c2275.jpg) +Figure 3. High-pass filtering improves performance over the unfiltered version. Rows represent the added task vectors, while columns are the test datasets. The results show a significant boost in generalization with only a slight decline in task performance. + +# 5.3. Lightweight Expert Extraction + +To compensate for information lost due to discarding low-frequency signals, we utilize lightweight experts added during inference. As analyzed in Sec. 5.2, low-frequency signals contain task-specific information, but preserving them requires an inefficient inverse Fourier transform during each inference. Thus, we propose a more efficient way. + +Since fine-tuning leads to task specialization, the most changed parameters can be approximated as having the greatest impact [32]. After reasonable rescaling, these selected parameters serve as task experts, minimizing storage (typically around $1\%$ of parameters) without extra inference computation. The validity is detailed in Appendix D.4. + +Although magnitude-based extraction [48] exists, limited research on rescaling hinders compression. As rescaling is crucial to expert extraction, we propose a rescaling method. Let $M(v_{i},d)$ represent the top- $d$ ( $d$ is a percentage) parameters for task vector $v_{i}$ , $\mu_{i}$ be the rescale factor, and $e(v_{i})$ be the expert for task $t_{i}$ . Then we have: + +$$ +e \left(v _ {i}\right) = \mu_ {i} M \left(v _ {i}, d\right), \mu_ {i} = - \frac {\mathbb {E} \left(M \left(v _ {i} , d\right)\right) \cdot \log (d)}{\lambda_ {i} \cdot \mathbb {E} \left(v _ {i}\right)}. \tag {4} +$$ + +The goal is to maintain a consistent output after extraction to preserve performance. Details are in Appendix D.5. + +In summary, we efficiently extract task-specific experts to integrate into the backbone, mitigating task conflicts without excessive storage or inference overhead. + +# 5.4. FREE-Merging + +This section presents FREE-Merging, combining both parts to make it suitable for multiple tasks without affecting inference speed. Inspired by MoE [39], a router dynamically assigns tasks to activated experts. The router can take + +Table 2. Comparison of model merging performance using ViT-B/32 (B) and ViT-L/14 (L) on eight visual tasks evaluating by accuracy. + +
MethodBase ModelSUN397CarsRESISC45EuroSATSVHNGTSRBMNISTDTDAvg.
BLBLBLBLBLBLBLBLBL
Individual75.382.377.792.496.197.499.710097.598.198.799.299.799.779.484.190.594.2
Traditional MTL73.980.874.490.693.996.398.296.395.897.698.999.199.599.677.984.488.993.5
Weight Averaging65.372.163.481.671.482.671.791.964.278.252.870.787.597.150.162.865.879.6
Fisher Merging [33]68.669.269.288.670.787.566.493.572.980.651.174.887.993.359.970.068.382.2
Task Arithmetic [21]63.874.162.182.172.086.777.693.874.487.965.186.894.098.952.265.670.184.5
Ties-Merging [48]64.876.562.985.074.389.378.995.783.190.371.483.397.699.056.268.873.686.0
Breadcrumbs [7]63.074.961.084.975.788.684.795.483.489.876.490.198.199.257.968.275.086.4
PCB-Merging [12]65.575.864.186.078.188.680.296.084.788.077.190.998.099.158.470.075.886.9
FR-Merging(ours)66.276.464.587.077.290.290.196.885.492.082.392.898.599.360.071.578.188.3
AdaMerging++ [52]66.679.468.390.382.291.694.297.489.693.489.097.598.399.060.679.281.191.0
Twin-Merging [29]73.682.671.790.392.194.999.399.495.396.797.298.099.199.474.080.187.892.7
EMR-Merging [20]74.182.372.790.891.995.399.499.595.896.996.998.199.199.572.180.187.792.8
FREE-Merging(ours)77.183.578.292.493.496.299.599.696.398.198.298.799.599.775.481.789.793.7
+ +# Algorithm 1 Workflow of FREE-Merging + +Input: Task vectors $\{v_{i}\}_{i = 1}^{K}$ , pre-trained model $\theta_{pre}$ , router $R$ , and input data $X$ . + +1: FR-Merging: +$\triangleright$ Only excute once. +2: Get backbone $\theta_{m} = \theta_{pre} + \sum_{i=1}^{K} \lambda_{i} G(v_{i})$ , where $G$ and $\lambda$ are defined by Eq. 1 and Eq. 3, respectively. +3: Expert Extraction: +$\triangleright$ Only execute once. +4: Get task experts $\{e_i\}_{i = 1}^K$ as Eq. 4. +5: Inference: +6: for $x \in X$ do +7: $[w_{1},\dots w_{K}]\gets argmax(R(x))$ +8: $\theta_{*} = \theta_{m} + \sum_{i = 1}^{K}w_{i}e_{i}$ +9: $Y\gets Y\cup f(x;\theta_{*})$ +10: end for + +Output: Output $Y$ for input $X$ . + +various forms [24], allowing for more flexibility in handling different tasks. Detailed discussions about router can be found in Appendix D.6. For detailed algorithms, see Alg 1. + +First, we obtain the merged backbone (line 2) and the task experts (line 4). Then, during the inference, we dynamically route the experts based on inputs (lines 7-9). + +# 5.5. Time Complexity and Storage Space Analysis + +This section analyzes the merging time complexity and storage requirements. We aim to minimize both for practical use. Assume there are $n$ models to merge, each with $m$ parameters and requiring storage space $s$ . + +The time complexity includes the merging and inference part. In the merging part, FR-Merging has a complexity of $O(nmlogm)$ [14] while expert extraction is $O(nm)$ , making the overall merging complexity $O(nmlogm)$ , which is manageable. For instance, merging three 10B models requires only $O(10^{12})$ addition operations, a minor load for modern devices. Moreover, with only a lightweight router added, the extra inference cost is negligible. + +Next, we analyze storage cost. We typically save only about $1\%$ additional parameters per model, requiring $(1 + 0.01n)s$ storage, which is far less than the requirement of $ns$ for the model ensemble. For large models, reducing storage requirements is crucial for alleviating storage pressure on edge users. For detailed analysis, please refer to Sec 6.7. + +# 6. Experiments + +In this section, we evaluate the proposed methods across various tasks. We compare FR-Merging with other cost-free methods (which do not require additional training or data), including Weight Averaging, Fisher Merging [33], Task Arithmetic [21], Ties-Merging [48], Bubbles [7], and PCB-Merging [12]. We also compare our FREE-Merging with popular methods, including RegMean [23], AdaMerging [52], EMR-Merging [20], and Twin-Merging [29]. The original EMR-Merging uses a perfect router; however, we use an imperfect router for realism. Some experiments include traditional MTL with data from all tasks as a baseline. The evaluation covers CV, NLP, and multi-modal tasks. + +# 6.1. Merging Vision Models + +Merging 8 ViTs. We select the two visual encoders of CLIP, ViT-B/32 and ViT-L/14 [37], as the pre-trained models and choose eight image classification tasks as target tasks: SUN397, Cars, RESISC45, EuroSAT, SVHN, GT-SRB, MNIST, and DTD [52]. Additionally, an MLP is used as the lightweight router. Detailed information on the datasets and baselines can be found in the Appendix C.1. + +Table 2 presents the comparative results. First, FR-Merging outperforms other cost-free methods, improving ViT-B/32 by $2.5\%$ and ViT-L/14 by $1.5\%$ over PCB-Merging [12], demonstrating its effectiveness in reducing performance loss caused by task interference. Through FR-Merging, we can construct a merged backbone that can be used independently at a low cost. Here, we focus on the basic FR-Merging. In Sec. 6.6, we discuss further im + +Table 3. Comparison of model merging results using ViT-B/16 on 30 visual tasks. Additional Cost indicates whether additional training or additional data is required during model merging. + +
MethodAdditional CostAvg. Acc
IndividualX93.02
Weight AveragingX42.51
Task Arithmetic [21]X48.88
Ties-Merging [48]X37.53
Breadcrumbs [7]X43.57
FR-Merging(ours)X53.90
RegMean [23]✓68.13
AdaMerging [52]✓60.24
Twin-Merging [29]✓75.52
EMR-Merging [20]✓76.39
FREE-Merging(ours)✓79.67
+ +Table 4. Language model merging results: RoBERTa fine-tuned on 8 discriminative tasks, T0-3B on 11 discriminative tasks with $IA^3$ , and Qwen-14B on 3 generative tasks with LoRA. + +
MethodBase ModelFull RoBERTaPEFTAvg.
T0-3BQwen-14B
Individual85.5571.3572.0776.32
Weight Averaging51.3458.0166.8058.71
Task Arithmetic[21]66.6563.9166.4065.65
Ties-Merging [48]64.0466.4167.4665.97
Breadcrumbs [7]68.1954.6466.7763.20
PCB-Merging [12]67.8666.1066.6966.39
FR-Merging(ours)70.0266.8868.0068.30
RegMean [23]70.0158.0167.5265.18
EMR-Merging [20]74.2067.1170.9870.76
Twin-Merging [29]78.2966.7071.6872.22
FREE-Merging(ours)80.1668.6872.7873.87
+ +provements achieved by combining it with existing methods. Then, compared to methods requiring extra information, FREE-Merging also improves performance by $2\%$ , leveraging lightweight computation and efficient expert extraction, requiring only $1\%$ additional storage per model, outperforming $3\%$ of EMR-Merging [20] and $2\%$ of Twin-Merging [29]. Thus, FREE-Merging provides an excellent balance between storage, computation, and performance. + +Merging 30 ViTs. We select ViT-B/16 as the pre-trained model and evaluate the model merging performance across 30 image classification tasks following [20]. + +Table 3 presents the results, with detailed analysis provided in Appendix E.1. It is evident that our proposed FR-Merging and FREE-Merging exhibit remarkable performance. Although merging a large number of models is uncommon in practice, this experiment underscores the effectiveness of Fourier filtering and lightweight task experts in mitigating performance loss caused by task interference. These results highlight the robustness and broad applicability of our methods across diverse scenarios. + +Table 5. Comparison of performance of different model merging methods after full fine-tuning LLaMa2-13B on three tasks: Instruction Following (LM), Math, and Code Generation. + +
MethodAlpacaEvalGSM8KMBPPAvg.
LM Fine-tuned [46]88.945.931.455.4
Math Fine-tuned [30]22.363.523.036.5
Code Fine-tuned [31]16.218.827.020.7
Task Arithmetic [21]72.148.3040.1
TIES-Merging [48]77.566.927.257.2
DARE [53]77.562.729.456.5
DELLA-Merging [8]80.461.831.457.9
PCB-Merging [12]81.260.630.757.5
FR-Merging(ours)87.564.332.861.5
Twin-Merging [29]83.561.430.458.4
EMR-Merging [20]82.662.331.258.7
FREE-Merging(ours)88.165.731.861.8
+ +# 6.2. Merging Language Models + +Merging Medium-sized Language Models. We investigate the generalization ability of our method on medium-scale language models, using RoBERTa [59] as the pretrained model with full fine-tuning on eight discriminative tasks [48]. Details can be found in Appendix C.1. + +The average merging results of the eight tasks are presented in Table 4, with complete results in Appendix E.2. It can be observed that, FR-Merging outperforms other cost-free methods by nearly $2\%$ . Additionally, FREE-Merging also exhibits remarkable effectiveness, achieving optimal results on almost all datasets and surpassing current methods in average performance. This highlights the great potential of our proposed methods for language models. + +Merging PEFT models. PEFT is a popular method that reduces fine-tuning costs [10], but storing PEFT parameters for multiple tasks remains space-inefficient [55]. The effectiveness of our approach on PEFT demonstrates the generalization capability and offers an efficient deployment way. + +Firstly, we use T0-3B [38] as the pre-trained model and $(IA)^{3}$ [26] as the PEFT method, fine-tuning on 11 discriminative tasks [20]. Our proposed FR-Merging and FREE-Merging improve performance by nearly $1\%$ , as shown in Table 4. Next, we employ Qwen-14B [2] with LoRA [19] as the PEFT way, fine-tuning on three generative tasks [29]. The results show that our methods lead to a $2\%$ advantage in both cost-free and full frameworks. This confirms the feasibility with PEFT and strong generalization of our methods. Full results are available in the Appendix E.2. + +Merging Large Language models. We explore large model merging using LLaMa2-13B [40] as the pre-trained model, with WizardLM [46] for instruction following, WizardMath [30] for math, and WizardCoder-Python [31] for coding. Experiments are conducted on AlpacaEval (instruction following), GSM8K (math), and MBPP (code) [53]. + +Table 5 shows that our proposed FR-Merging even surpasses existing methods that rely on additional information. + +Table 6. Results of merging multi-modal BEiT3 models on five vision-language tasks. + +
MethodsTask MetricCOCO-Retrieval Accuracy(↑)COCO-CaptioningImageNet-1k Classification Accuracy(↑)NLVR2 Accuracy(↑)VQAv2 Accuracy(↑)
BLEU4(↑)CIDEr(↑)METEOR(↑)ROUGE-L(↑)
Individual0.84560.3941.3370.3110.6010.85370.77650.8439
Weight Averaging0.18930.0310.0010.1150.1590.67710.28000.6285
Task Arithmetic [21]0.31770.0330.0000.1180.1760.67320.38090.6933
Ties-Merging [48]0.39290.0290.0010.1080.1670.69780.32060.6717
Breadcrumbs [7]0.18930.0110.0120.0650.1140.63230.60730.6823
FR-Merging(ours)0.70100.0160.0190.0960.1410.69920.64070.6965
EMR-Merging [20]0.79460.2891.0600.2720.5340.77420.74750.7211
Twin-Merging [29]0.79370.3251.0780.2780.5580.77300.77620.7232
FREE-Merging(ours)0.80930.3471.1950.2890.5720.78500.81680.7297
+ +![](images/50d6947d74695332a096a8c8ec33943b8c5967a9e46af2b5aa16628036d988bc.jpg) +Figure 4. Visualization of the loss landscape on EuroSAT before and after high-pass filtering (top vs. bottom) in the model merging of the fine-tuned ViT-B/32 on EuroSAT and DTD tasks [21]. + +This indicates that our findings in FR-Merging regarding low-frequency signals and task interference also hold when extended to large models. This is crucial for practical usage of large models. Additionally, our proposed FREE-Merging also demonstrates strong advantages, enabling efficient cross-domain deployment of large models. + +# 6.3. Merging Multi-Modal Models + +We merge various fine-tuned BEiT3-base [42] models across five tasks: ImageNet-1k (Classification), VQAv2 (Visual Question Answering), NLVR2 (Reasoning), COCO Captioning (Image Captioning), and COCO Retrieval (Image-Text Retrieval). COCO Captioning is evaluated with BLEU4, CIDEr, METEOR, and ROUGE-L, while the others use accuracy [20]. Details are in Appendix C.1. + +Table 6 shows that our methods gain great improvements in multi-modal models. Unlike vision or language model merging, which often involve similar tasks, such as classification, the task similarity here is relatively low. The strong performance highlights the generalization ability of our methods, making them widely applicable. + +Table 7. Ablation of FREE-Merging components on 8 visual tasks. + +
FR-MergingEqua.3Top-K ExpertScalingViT-B/32ViT-B/16ViT-L/14
XXXX70.1075.2484.50
✓XXX76.9681.3488.21
X✓XX71.4377.4184.20
✓✓XX78.1082.5488.30
XX✓X65.2468.9870.13
✓✓✓X79.3884.3189.53
XX✓✓82.3286.8790.47
✓✓✓✓89.6891.0993.70
+ +# 6.4. Fourier Transform Analysis + +In this section, we explain why Fourier high-pass filtering improves merging performance from an optimization perspective. Previous work [25] shows that landscape geometry impacts model generalization. Fig. 4 demonstrates that high-pass filtering widens the valley in the loss landscape (bottom plot). Since fine-tuned models from the same pretrained model are closer in the landscape [34], a wider valley increases the probability of merged results falling within the loss basin as shown in the bottom plot of Fig. 4, leading to better performance. Without filtering, the landscape is flat at a high level, which has little effect on model merging and only highlights the advantage of pre-training [17]. + +# 6.5. Ablation Study + +In this section, we conduct ablation experiments on the components of our method. All experiments are performed on ViT-B/32 with eight classification tasks as target tasks. + +Ablation on FR-Merging. We analyze the impact of FR-Merging on the backbone and overall performance. Table 7 shows that the performance of various models greatly improves after employing FR-Merging. We then compare high-pass filtering with other cost-free methods, including Ties [48], Top-K, Task Arithmetic [21], outlier dropout [7], EMR-Merging [20], and PCB-Merging [12], as well as low-pass and band-pass filtering, with results in Fig. 5a. For fairness, all experts use $1\%$ additional parameters. + +Our proposed FR-Merging outperforms other cost-free methods on both the backbone and with task experts, minimizing task interference. Examining the different filtering methods, we find that low-pass filtering leads to great per + +![](images/59fdbfc950c358e2b25b51dacb4bb240a829ea58170cf802bcae9e2c29a103ad.jpg) +(a) Ablation on FR-Merging. + +![](images/cda7c8558fef8b6b097b6692b1c16000d2be81238b533d7640ab996d79654359.jpg) +(b) Ablation on Task Expert. + +![](images/e20f278e0c623054f3f87346772bc71f55dfae6ccb70aa07ed87a51d957bc6d0.jpg) +(c) Router Analysis. +Figure 5. a) The ablation experiments on FR-Merging show that our proposed methods achieve optimal performance. b) Performance of experts at different sparsity rates. c) In experiments with random routers, our approach demonstrates optimal performance. + +Table 8. Transferability analysis of FR-Merging on 8 visual tasks. + +
MethodFR-MergingViT-B/32ViT-L/14
Ties-Merging [48]X73.686.0
Ties-Merging [48]✓78.388.4
Breadcrumbs [7]X75.086.4
Breadcrumbs [7]✓78.888.9
TSVM [16]X83.991.9
TSVM [16]✓84.892.5
+ +formance degradation, confirming that task interference in low-frequency regions is severe. Band-pass filtering also performs worse than FR-Merging, and this is because some high-frequency signals preserve abrupt changes in the signal and their removal harms the performance of the model. + +Ablation on Merging Coefficients. We conduct an ablation study on our merging coefficients as in Eq. 3, with results in Table 7. Using them instead of simple averaging yields slight improvements by considering output variations, benefiting tasks with varying fine-tuning magnitudes. + +Ablation on Task Expert. We conduct an ablation study on the expert extraction way. Table 7 shows that incorporating our experts and rescaling method results in significant improvements, confirming their effectiveness. To validate the superiority of our method, we evaluate expert-only performance with varying parameter proportions, comparing Top-K selection, Bernoulli selection [53], SVD extraction [29], and our Top-K selection with rescaling. Fig. 5b shows that our method excels, particularly under high sparsity. + +Router Analysis. We analyze the proposed router method, aiming for strong merging performance despite sub-optimal routing. Since router inaccuracies may arise from limited training resources or edge deployment issues, our methods maintain performance in real-world scenarios. Using a random router to simulate inaccuracy, we compare our method with existing router-based approaches. Our method outperforms Twin-Merging by $13.5\%$ and EMR-Merging by $8.5\%$ on average, as shown in Fig. 5c, indicating that our methods + +Table 9. Trade-off of performance, training, inference, and storage costs using ViT-L/14 across 8 image classification tasks. + +
MethodTraining TokensTraining CostStorage Param.Inference Cost (/1000 items)Acc.
Individual002.4B31.0s94.2
MTL304M12h304M31.0s93.5
FR-Merging00304M31.0s88.2
Twin-Merging0.9M147s352M32.2s92.7
FREE-Merging0.9M147s328M32.2s93.7
+ +construct a reasonable backbone and generalizable experts. + +# 6.6. Transferability Analysis + +Our proposed FR-Merging can integrate with other methods, as it addresses frequency-domain task interference that current methods overlook. As shown in Table 8, combining various methods with FR-Merging leads to performance improvements. Therefore, FR-Merging can be integrated with more approaches to achieve greater benefits. + +# 6.7. Speed and Storage Space Analysis + +This part analyzes training costs, inference speed, and storage as shown in Table 9. FR-Merging achieves strong performance with significantly reducing storage costs without extra training or inference costs, while FREE-Merging slightly increases overhead for better performance. Both of these two methods enhance deep learning deployment. + +# 7. Conclusion + +In this paper, we propose FR-Merging and FREE-Merging to efficiently build a multi-task model by merging existing ones. FR-Merging constructs a backbone without training by high-pass filtering to remove information that dominates task interference. Meanwhile, FREE-Merging employs lightweight experts to offset information loss during merging. Our method balances training, storage, inference costs, and performance, showing high practical value. + +# Acknowledgements + +# References + +[1] Samuel Ainsworth, Jonathan Hayase, and Siddhartha Srinivasa. Git re-basin: Merging models modulo permutation symmetries. In The Eleventh International Conference on Learning Representations, 2023. 4 +[2] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 6 +[3] Tim-Oliver Buchholz and Florian Jug. Fourier image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1846-1854, 2022. 4 +[4] Shijie Chen, Yu Zhang, and Qiang Yang. Multi-task learning in natural language processing: An overview. ACM Computing Surveys, 56(12):1-32, 2024. 2 +[5] Tianyi Chu, Jiafu Chen, Jiakai Sun, Shuobin Lian, Zhizhong Wang, Zhiwen Zuo, Lei Zhao, Wei Xing, and Dongming Lu. Rethinking fast fourier convolution in image inpainting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23195-23205, 2023. 4 +[6] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instructionfinetuned language models. Journal of Machine Learning Research, 25(70):1-53, 2024. 3, 4 +[7] MohammadReza Davari and Eugene Belilovsky. Model breadcrumbs: Scaling multi-task model merging with sparse masks. arXiv preprint arXiv:2312.06795, 2023. 3, 5, 6, 7, 8, 9, 10 +[8] Pala Tej Deep, Rishabh Bhardwaj, and Soujanya Poria. Della-merging: Reducing interference in model merging through magnitude-based sampling. arXiv preprint arXiv:2406.11617, 2024. 6 +[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota, 2019. Association for Computational Linguistics. 1 +[10] Ning Ding, Yujia Qin, Guang Yang, Fuchao Wei, Zonghan Yang, Yusheng Su, Shengding Hu, Yulin Chen, Chi-Min Chan, Weize Chen, et al. Parameter-efficient fine-tuning of large-scale pre-trained language models. Nature Machine Intelligence, 5(3):220-235, 2023. 6 +[11] Xibin Dong, Zhiwen Yu, Wenming Cao, Yifan Shi, and Qianli Ma. A survey on ensemble learning. Frontiers of Computer Science, 14:241-258, 2020. 2 + +This work is supported by National Natural Science Foundation of China (NSFC) (62232005, 62202126); the National Key Research and Development Program of China (2021YFB3300502). +[12] Guodong Du, Junlin Lee, Jing Li, Runhua Jiang, Yifei Guo, Shuyang Yu, Hanting Liu, Sim Kuan Goh, Ho-Kin Tang, Daojing He, et al. Parameter competition balancing for model merging. arXiv preprint arXiv:2410.02396, 2024. 2, 3, 5, 6, 7, 8, 9, 10 +[13] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 3 +[14] Pierre Duhamel and Martin Vetterli. Fast fourier transforms: a tutorial review and a state of the art. Signal processing, 19 (4):259-299, 1990. 5 +[15] Mudasir A Ganaie, Minghui Hu, Ashwani Kumar Malik, Muhammad Tanveer, and Ponnuthurai N Suganthan. Ensemble deep learning: A review. Engineering Applications of Artificial Intelligence, 115:105151, 2022. 2 +[16] Antonio Andrea Gargiulo, Donato Crisostomi, Maria Sofia Bucarelli, Simone Scardapane, Fabrizio Silvestri, and Emanuele Rodola. Task singular vectors: Reducing task interference in model merging. arXiv preprint arXiv:2412.00081, 2024. 8 +[17] Timur Garipov, Pavel Izmailov, Dmitrii Podoprikhin, Dmitry P Vetrov, and Andrew G Wilson. Loss surfaces, mode connectivity, and fast ensembling of dnns. Advances in neural information processing systems, 31, 2018. 7 +[18] Rafael C Gonzales and Paul Wintz. Digital image processing. Addison-Wesley Longman Publishing Co., Inc., 1987. 2 +[19] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 6 +[20] Chenyu Huang, Peng Ye, Tao Chen, Tong He, Xi-angyu Yue, and Wanli Ouyang. Emr-merging: Tuning-free high-performance model merging. arXiv preprint arXiv:2405.17461, 2024. 1, 2, 4, 5, 6, 7, 3, 8, 9, 10 +[21] Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Suchin Gururangan, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. Editing models with task arithmetic. arXiv preprint arXiv:2212.04089, 2022. 1, 2, 3, 5, 6, 7, 4, 8, 9, 10 +[22] Dongfu Jiang, Xiang Ren, and Bill Yuchen Lin. Lmblender: Ensembling large language models with pairwise ranking and generative fusion. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 14165-14178, 2023. 2 +[23] Xisen Jin, Xiang Ren, Daniel Preotiac-Pietro, and Pengxiang Cheng. Dataless knowledge fusion by merging weights of language models. arXiv preprint arXiv:2212.09849, 2022. 1, 2, 5, 6, 3, 8, 9, 10 +[24] Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan Firat, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. Gshard: Scaling giant models with conditional computation and automatic sharding. In International Conference on Learning Representations, 2021. 5 + +[25] Hao Li, Zheng Xu, Gavin Taylor, Christoph Studer, and Tom Goldstein. Visualizing the loss landscape of neural nets. Advances in neural information processing systems, 31, 2018. 7 +[26] Haokun Liu, Derek Tam, Mohammed Muqeeth, Jay Mohta, Tenghao Huang, Mohit Bansal, and Colin A Raffel. Few-shot parameter-efficient fine-tuning is better and cheaper than in-context learning. Advances in Neural Information Processing Systems, 35:1950-1965, 2022. 6 +[27] James Liu, Guangxuan Xiao, Kai Li, Jason D Lee, Song Han, Tri Dao, and Tianle Cai. Bitdelta: Your fine-tune may only be worth one bit. Advances in Neural Information Processing Systems, 37:13579-13600, 2024. 2 +[28] Yajing Liu, Yuning Lu, Hao Liu, Yaozu An, Zhuoran Xu, Zhuokun Yao, Baofeng Zhang, Zhiwei Xiong, and Chenguang Gui. Hierarchical prompt learning for multi-task learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10888-10898, 2023. 2 +[29] Zhenyi Lu, Chenghao Fan, Wei Wei, Xiaoye Qu, Dangyang Chen, and Yu Cheng. Twin-merging: Dynamic integration of modular expertise in model merging. arXiv preprint arXiv:2406.15479, 2024. 1, 2, 5, 6, 7, 8, 3, 9, 10 +[30] Haipeng Luo, Qingfeng Sun, Can Xu, Pu Zhao, Jianguang Lou, Chongyang Tao, Xiubo Geng, Qingwei Lin, Shifeng Chen, and Dongmei Zhang. Wizardmath: Empowering mathematical reasoning for large language models via reinforced evol-instruct. arXiv preprint arXiv:2308.09583, 2023. 6 +[31] Ziyang Luo, Can Xu, Pu Zhao, Qingfeng Sun, Xiubo Geng, Wenxiang Hu, Chongyang Tao, Jing Ma, Qingwei Lin, and Daxin Jiang. Wizardcoder: Empowering code large language models with evol-instruct. In The Twelfth International Conference on Learning Representations, 2024. 6 +[32] Xinyin Ma, Gongfan Fang, and Xinchao Wang. Llm-pruner: On the structural pruning of large language models. Advances in neural information processing systems, 36:21702-21720, 2023. 4 +[33] Michael S Matena and Colin A Raffel. Merging models with fisher-weighted averaging. Advances in Neural Information Processing Systems, 35:17703-17716, 2022. 1, 2, 5, 9 +[34] Behnam Neyshabur, Hanie Sedghi, and Chiyuan Zhang. What is being transferred in transfer learning? Advances in neural information processing systems, 33:512-523, 2020. 7 +[35] Tan Nguyen, Minh Pham, Tam Nguyen, Khai Nguyen, Stanley Osher, and Nhat Ho. Fourierformer: Transformer meets generalized fourier integral theorem. Advances in Neural Information Processing Systems, 35:29319-29335, 2022. 4 +[36] Biqing Qi, Fangyuan Li, Zhen Wang, Junqi Gao, Dong Li, Peng Ye, and Bowen Zhou. Less is more: Efficient model merging with binary task switch. arXiv preprint arXiv:2412.00054, 2024. 2 +[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 3, 4, 5 + +[38] Victor Sanh, Albert Webson, Colin Raffel, Stephen Bach, Lintang Sutawika, Zaid Alyafeai, Antoine Chaffin, Arnaud Stiegler, Arun Raja, Manan Dey, et al. Multitask prompted training enables zero-shot task generalization. In International Conference on Learning Representations, 2022. 6 +[39] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. In International Conference on Learning Representations, 2016. 2, 4 +[40] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 3, 6 +[41] Simon Vandenhende, Stamatios Georgoulis, Wouter Van Gansbeke, Marc Proesmans, Dengxin Dai, and Luc Van Gool. Multi-task learning for dense prediction tasks: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(7):3614-3633, 2021. 2 +[42] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, et al. Image as a foreign language: Beit pretraining for vision and vision-language tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19175-19186, 2023. 7 +[43] T Wolf. Huggingface's transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771, 2019. 1 +[44] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In International conference on machine learning, pages 23965-23998. PMLR, 2022. 2 +[45] Chengyue Wu, Teng Wang, Yixiao Ge, Zeyu Lu, Ruisong Zhou, Ying Shan, and Ping Luo. $\pi$ -tuning: Transferring multimodal foundation models with optimal multi-task interpolation. In International Conference on Machine Learning, pages 37713-37727. PMLR, 2023. 4 +[46] Can Xu, Qingfeng Sun, Kai Zheng, Xiubo Geng, Pu Zhao, Jiazhan Feng, Chongyang Tao, Qingwei Lin, and Daxin Jiang. Wizardlm: Empowering large pre-trained language models to follow complex instructions. In The Twelfth International Conference on Learning Representations, 2024. 6 +[47] Jing Xu and Jingzhao Zhang. Random masking finds winning tickets for parameter efficient fine-tuning. In *Forty-first International Conference on Machine Learning*. 2 +[48] Prateek Yadav, Derek Tam, Leshem Choshen, Colin A Raffel, and Mohit Bansal. Ties-merging: Resolving interference when merging models. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3, 4, 5, 6, 7, 8, 9, 10 +[49] Enneng Yang, Li Shen, Guibing Guo, Xingwei Wang, Xiaochun Cao, Jie Zhang, and Dacheng Tao. Model merging + +in llms, mllms, and beyond: Methods, theories, applications and opportunities. arXiv preprint arXiv:2408.07666, 2024. 2 +[50] Enneng Yang, Li Shen, Zhenyi Wang, Guibing Guo, Xiaojun Chen, Xingwei Wang, and Dacheng Tao. Representation surgery for multi-task model merging. arXiv preprint arXiv:2402.02705, 2024. 3, 4 +[51] Enneng Yang, Li Shen, Zhenyi Wang, Guibing Guo, Xiao-jun Chen, Xingwei Wang, and Dacheng Tao. Representation surgery for multi-task model merging. In *Forty-first International Conference on Machine Learning*, 2024. 1 +[52] Enneng Yang, Zhenyi Wang, Li Shen, Shiwei Liu, Guibing Guo, Xingwei Wang, and Dacheng Tao. Adamerging: Adaptive model merging for multi-task learning. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 5, 6, 3, 9 +[53] Le Yu, Bowen Yu, Haiyang Yu, Fei Huang, and Yongbin Li. Language models are super mario: Absorbing abilities from homologous models as a free lunch. In *Forty-first International Conference on Machine Learning*, 2024. 2, 3, 4, 6, 8, 10 +[54] Kerem Zaman, Leshem Choshen, and Shashank Srivastava. Fuse to forget: Bias reduction and selective memorization through model fusion. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 18763-18783, Miami, Florida, USA, 2024. Association for Computational Linguistics. 8 +[55] Jinghan Zhang, Junteng Liu, Junxian He, et al. Composing parameter-efficient modules with arithmetic operation. Advances in Neural Information Processing Systems, 36: 12589-12610, 2023. 6 +[56] Yu Zhang and Qiang Yang. A survey on multi-task learning. IEEE transactions on knowledge and data engineering, 34 (12):5586-5609, 2021. 1 +[57] Yanqi Zhou, Tao Lei, Hanxiao Liu, Nan Du, Yanping Huang, Vincent Zhao, Andrew M Dai, Quoc V Le, James Laudon, et al. Mixture-of-experts with expert choice routing. Advances in Neural Information Processing Systems, 35:7103-7114, 2022. 2 +[58] Yuyan Zhou, Liang Song, Bingning Wang, and Weipeng Chen. Metagpt: Merging large language models using model exclusive task arithmetic. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 1711-1724, 2024. 1 +[59] Liu Zhuang, Lin Wayne, Shi Ya, and Zhao Jun. A robustly optimized BERT pre-training approach with post-training. In Proceedings of the 20th Chinese National Conference on Computational Linguistics, pages 1218-1227, Huhhot, China, 2021. Chinese Information Processing Society of China. 6 \ No newline at end of file diff --git a/2025/FREE-Merging_ Fourier Transform for Efficient Model Merging/images.zip b/2025/FREE-Merging_ Fourier Transform for Efficient Model Merging/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..71e293f228ace42226022d9d274338d71e220410 --- /dev/null +++ b/2025/FREE-Merging_ Fourier Transform for Efficient Model Merging/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5edad7fe2bb6c5064ec7d316011883f004ae278db7ad3643e04d9e40359c21d2 +size 772510 diff --git a/2025/FREE-Merging_ Fourier Transform for Efficient Model Merging/layout.json b/2025/FREE-Merging_ Fourier Transform for Efficient Model Merging/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..490a0417f59f36c51512c42fa78abe7c1965d7f1 --- /dev/null +++ b/2025/FREE-Merging_ Fourier Transform for Efficient Model Merging/layout.json @@ -0,0 +1,9606 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 105, + 103, + 506, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 103, + 506, + 121 + ], + "spans": [ + { + "bbox": [ + 105, + 103, + 506, + 121 + ], + "type": "text", + "content": "FREE-Merging: Fourier Transform for Efficient Model Merging" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 144, + 288, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 144, + 288, + 185 + ], + "spans": [ + { + "bbox": [ + 138, + 144, + 288, + 185 + ], + "type": "text", + "content": "Shenghe Zheng \nHarbin Institute of Technology \nshenghez.zheng@gmail.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 322, + 144, + 472, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 144, + 472, + 185 + ], + "spans": [ + { + "bbox": [ + 322, + 144, + 472, + 185 + ], + "type": "text", + "content": "Hongzhi Wang* \nHarbin Institute of Technology \nwangzh@hit.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 239, + 296, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 296, + 574 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 296, + 574 + ], + "type": "text", + "content": "With the rapid growth of deep learning, there is an increasing availability of open-source models for various tasks. However, single fine-tuned models often fall short of meeting the diverse needs of users. Model merging has thus emerged as an efficient method to integrate the capabilities of existing models into a unified model. Nevertheless, existing model merging methods face challenging trade-offs between performance and deployment costs, primarily due to task interference. For the first time, we reveal that task interference is evident in the frequency domain of model parameters, yet current efforts only focus on spatial domain solutions, which are largely ineffective in addressing frequency domain interference. To mitigate the impact of frequency domain interference, we propose FR-Merging, an innovative method that effectively filters harmful frequency domain interference on the backbone with minimal computational overhead. Since performance loss is inevitable with cost-free methods, we propose a lightweight task-specific expert module that dynamically compensates for information loss during merging. This proposed framework, FREE-Merging (FR-Merging with experts), strikes a balanced trade-off between training cost, inference latency, storage requirements, and performance. We demonstrate the effectiveness of both FR-Merging and FREE-Merging on multiple tasks across CV, NLP, and Multi-Modal domains and show that they can be flexibly adapted to specific needs. Our code is available at: https://github.com/Zhengsh123/FREE-Merging." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 599, + 135, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 599, + 135, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 599, + 135, + 612 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 620, + 295, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 620, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 620, + 295, + 693 + ], + "type": "text", + "content": "In the current era of deep learning, the pretrain-finetune paradigm is a standard procedure [9, 37, 40]. However, edge applications often lack powerful computing resources and data, favoring ready-to-use models for specific scenarios. The recent growth of open-source platforms like Hugging-Face [43] has made this feasible. Yet, practical demands" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 213, + 555, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 321 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 321 + ], + "type": "text", + "content": "are often not fully covered by a single existing model and are typically a union of several models [23, 52]. However, running separate models for each subtask increases storage and inference costs, especially with large models. Model merging attempts to combine existing models to build a model capable of addressing all target tasks [21, 33, 48]. This method reduces training costs, data privacy issues in multi-task learning (MTL) [20, 56] and deployment costs, garnering widespread attention." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 324, + 556, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 556, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 556, + 480 + ], + "type": "text", + "content": "While model merging aims to reduce training costs, it often suffers from performance loss due to task interference [20]. To mitigate this issue, existing methods develop along three directions. The first direction computes merging coefficients based on weights or data [23, 33]. The second uses the parameter differences before and after finetuning (i.e. task vectors) for merging [21, 48, 52, 58]. The third method introduces independent knowledge for each task [20, 29]. While the third way balances performance and costs, it still faces two critical limitations. First, it neglects backbone optimization during merging, resulting in subpar performance. Second, it requires high computation and storage, making it impractical for edge deployment." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 483, + 556, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 556, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 556, + 615 + ], + "type": "text", + "content": "We argue that an effective model merging method requires addressing both of the above issues. First, since the backbone constitutes the majority of parameters and thus determines performance, an efficient backbone merging method is essential to reduce performance loss. Then, given the no free lunch theorem, cost-free merging cannot avoid task interference. Therefore, introducing task-specific experts is necessary. However, they must remain lightweight to avoid storage and inference overhead. Thus, we propose FREE-Merging, a two-stage method that ensures an efficient backbone and lightweight experts as shown in Fig 1." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": "For backbone merging, the key lies in eliminating task interference [51]. We identify a compelling phenomenon in which task interference is evident in the frequency domain of model parameters. And we present in Sec. 4 that frequency-domain interference is strongly correlated with performance. Notably, existing methods, which focus only on spatial-domain operations, fail to address frequency-domain interference, leading to suboptimal outcomes. Ex" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 146, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 146, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 146, + 712 + ], + "type": "text", + "content": "*Corresponding Author" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3863" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 193, + 224 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 193, + 224 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 193, + 224 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 193, + 224 + ], + "type": "image", + "image_path": "c1b7af58661b1fa1c1d29f024989d2bbdfd80d0cb91e6a2bd8d7578b6f72c1f7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 235, + 555, + 269 + ], + "lines": [ + { + "bbox": [ + 54, + 235, + 555, + 269 + ], + "spans": [ + { + "bbox": [ + 54, + 235, + 555, + 269 + ], + "type": "text", + "content": "Figure 1. The workflow of FREE-Merging involves two main steps. First, FR-Merging utilizes high-pass filtering to remove harmful specialized information from each model, thereby constructing a high-quality merged backbone. Second, it employs lightweight extraction of task experts, which are dynamically added during inference to mitigate the impact of task interference." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 195, + 70, + 405, + 225 + ], + "blocks": [ + { + "bbox": [ + 195, + 70, + 405, + 225 + ], + "lines": [ + { + "bbox": [ + 195, + 70, + 405, + 225 + ], + "spans": [ + { + "bbox": [ + 195, + 70, + 405, + 225 + ], + "type": "image", + "image_path": "7394ca043757c5a26b29325fd0558863920099268d48ac200488707e554e3b53.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 408, + 71, + 554, + 225 + ], + "blocks": [ + { + "bbox": [ + 408, + 71, + 554, + 225 + ], + "lines": [ + { + "bbox": [ + 408, + 71, + 554, + 225 + ], + "spans": [ + { + "bbox": [ + 408, + 71, + 554, + 225 + ], + "type": "image", + "image_path": "53aa4ab86d7788d51c11954ead466d1c0288307f670260c7616f7550dd37db05.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 289, + 295, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 289, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 54, + 289, + 295, + 456 + ], + "type": "text", + "content": "Experimental evidence reveals that task interference is severe in the low-frequency region. This occurs because, high-frequency signals represent fine-grained variations, while low-frequency parts capture the global structure [18], which is more likely to contain task-specific information that leads to task interference. Leveraging the inherent redundancy in fine-tuned parameters [27, 47, 53], we propose to directly filter out low-frequency parts with severe task interference. As detailed in Sec. 5.2, this high-pass filtering substantially enhances generalization with minimal performance loss, preventing task interference during merging. Our lightweight yet highly effective approach, FR-Merging, is the first to apply Fourier filtering to neural network parameters, showing its potential for broader applications." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 460, + 295, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 460, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 54, + 460, + 295, + 555 + ], + "type": "text", + "content": "After obtaining a high-performance backbone through FR-Merging, we use a lightweight task-specific expert to recover the inevitable information loss with minimal storage. Unlike existing methods that store extensive knowledge, we rescale low-frequency signals, which store task-specific information, retaining task expertise with only " + }, + { + "bbox": [ + 54, + 460, + 295, + 555 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 54, + 460, + 295, + 555 + ], + "type": "text", + "content": " of the parameters. Inspired by MoE [39, 57], we employ a router for dynamic expert routing, enhancing model flexibility." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": "In summary, we introduce FREE-Merging, which utilizes FR-Merging to reduce task interference and incorporates lightweight experts to eliminate information loss of model merging. Our main contributions are: 1). We identify frequency-domain task interference as a critical factor that greatly impacts merging performance but remains challenging to address through spatial-domain methods. 2). To tackle frequency-domain interference, we propose FR-Merging, which reduces interference while preserving performance, constructing the merged backbone. 3). We theoretically validate the necessity of introducing new information during merging and propose an effective expert construction method. 4). Extensive experiments show that" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 289, + 553, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 553, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 553, + 314 + ], + "type": "text", + "content": "FR-Merging and FREE-Merging effectively enhance performance across vision, language, and multi-modal tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 323, + 400, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 400, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 400, + 335 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 342, + 555, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 570 + ], + "type": "text", + "content": "Model Merging. Model merging integrates existing models to handle multiple tasks without training [21, 23, 49], but faces task interference challenges [36, 48]. Simple averaging [44] causes great performance degradation. Methods like Fisher-Merging [33] and RegMean [23] use matrices to determine merging coefficients but are computationally expensive or data-intensive, limiting edge deployment. Task Arithmetic [21] merges task vectors instead of weights, while Ties-Merging [48] resolves parameter conflicts, and AdaMerging [52] automates coefficient selection. However, these approaches have data distribution requirements and limited applicability. Techniques such as DARE [53] and PCB-Merging [12] mitigate performance drops but still face task conflicts. EMR-Merging [20] and Twin-Merging [29] store task-specific knowledge but require large storage and neglect backbone optimization. Our method uses high-pass filtering and lightweight experts to balance merging costs and effectiveness across various tasks. For further details, please refer to Appendix C.2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 570, + 553, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 553, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 553, + 629 + ], + "type": "text", + "content": "Model Ensemble. Model ensemble combines outputs from multiple models [11, 15, 22], but with large models, it faces storage and inference challenges. In contrast, Model merging enables a single model to solve multiple tasks, significantly reducing storage and inference costs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "Multi-task Learning. Multi-task learning (MTL) aims to solve multiple tasks using a single model [4, 28, 41], but in the era of large models, it faces challenges like high training costs, data privacy issues, and expertise requirements [52]. In contrast, model merging uses existing open-source models to create a multi-task model with little or no training, significantly reducing deployment costs." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3864" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 72, + 294, + 159 + ], + "blocks": [ + { + "bbox": [ + 57, + 72, + 294, + 159 + ], + "lines": [ + { + "bbox": [ + 57, + 72, + 294, + 159 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 294, + 159 + ], + "type": "image", + "image_path": "c735b87a90dd98bf2bf51f5bae2e5f1120d9374a56864f33f8f381651c3d0e21.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 57, + 168, + 294, + 256 + ], + "blocks": [ + { + "bbox": [ + 176, + 160, + 186, + 168 + ], + "lines": [ + { + "bbox": [ + 176, + 160, + 186, + 168 + ], + "spans": [ + { + "bbox": [ + 176, + 160, + 186, + 168 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 168, + 294, + 256 + ], + "lines": [ + { + "bbox": [ + 57, + 168, + 294, + 256 + ], + "spans": [ + { + "bbox": [ + 57, + 168, + 294, + 256 + ], + "type": "image", + "image_path": "ab7a92b21a70607cbd9d59de2649a84f37f062b33d1e32d56d05aaf436854fc8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 176, + 256, + 186, + 263 + ], + "lines": [ + { + "bbox": [ + 176, + 256, + 186, + 263 + ], + "spans": [ + { + "bbox": [ + 176, + 256, + 186, + 263 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 270, + 296, + 337 + ], + "lines": [ + { + "bbox": [ + 55, + 270, + 296, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 296, + 337 + ], + "type": "text", + "content": "Figure 2. a) The frequency domain amplitude power distribution of an attention head across different fine-tuned ViT-B-32 models. b) The correlation between the merging performance of ViT-B-32 and the variance of frequency domain amplitude power. Normalized accuracy is defined as the ratio of merged model performance to fine-tuned model performance on the same task." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 357, + 141, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 141, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 141, + 370 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 378, + 295, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 295, + 474 + ], + "type": "text", + "content": "Notation. Assume that " + }, + { + "bbox": [ + 55, + 378, + 295, + 474 + ], + "type": "inline_equation", + "content": "f_{\\theta}(x) \\to y" + }, + { + "bbox": [ + 55, + 378, + 295, + 474 + ], + "type": "text", + "content": " represents a neural network parameterized by " + }, + { + "bbox": [ + 55, + 378, + 295, + 474 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 378, + 295, + 474 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 378, + 295, + 474 + ], + "type": "inline_equation", + "content": "x \\in R^{m}" + }, + { + "bbox": [ + 55, + 378, + 295, + 474 + ], + "type": "text", + "content": " is the input and " + }, + { + "bbox": [ + 55, + 378, + 295, + 474 + ], + "type": "inline_equation", + "content": "y \\in R^{n}" + }, + { + "bbox": [ + 55, + 378, + 295, + 474 + ], + "type": "text", + "content": " is the output. Recent studies have focused on model merging via task vectors. For task " + }, + { + "bbox": [ + 55, + 378, + 295, + 474 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 378, + 295, + 474 + ], + "type": "text", + "content": ", the task vector is defined as " + }, + { + "bbox": [ + 55, + 378, + 295, + 474 + ], + "type": "inline_equation", + "content": "v_{k} = \\theta_{k} - \\theta_{pre}" + }, + { + "bbox": [ + 55, + 378, + 295, + 474 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 378, + 295, + 474 + ], + "type": "inline_equation", + "content": "\\theta_{k}" + }, + { + "bbox": [ + 55, + 378, + 295, + 474 + ], + "type": "text", + "content": " is the fine-tuned model for task " + }, + { + "bbox": [ + 55, + 378, + 295, + 474 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 378, + 295, + 474 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 378, + 295, + 474 + ], + "type": "inline_equation", + "content": "\\theta_{pre}" + }, + { + "bbox": [ + 55, + 378, + 295, + 474 + ], + "type": "text", + "content": " is the corresponding pre-trained model. Additionally, " + }, + { + "bbox": [ + 55, + 378, + 295, + 474 + ], + "type": "inline_equation", + "content": "e_{k}" + }, + { + "bbox": [ + 55, + 378, + 295, + 474 + ], + "type": "text", + "content": " represents the task expert that retains the ability to solve the task " + }, + { + "bbox": [ + 55, + 378, + 295, + 474 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 378, + 295, + 474 + ], + "type": "text", + "content": " with few parameters." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 475, + 295, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 295, + 608 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 295, + 608 + ], + "type": "text", + "content": "Problem Definition. The objective of model merging is to combine models " + }, + { + "bbox": [ + 55, + 475, + 295, + 608 + ], + "type": "inline_equation", + "content": "\\{\\theta_k\\}_{k=1}^K" + }, + { + "bbox": [ + 55, + 475, + 295, + 608 + ], + "type": "text", + "content": " for tasks " + }, + { + "bbox": [ + 55, + 475, + 295, + 608 + ], + "type": "inline_equation", + "content": "T = \\{t_i\\}_{i=1}^K" + }, + { + "bbox": [ + 55, + 475, + 295, + 608 + ], + "type": "text", + "content": " into a single model " + }, + { + "bbox": [ + 55, + 475, + 295, + 608 + ], + "type": "inline_equation", + "content": "\\theta_m" + }, + { + "bbox": [ + 55, + 475, + 295, + 608 + ], + "type": "text", + "content": " that can handle all tasks in " + }, + { + "bbox": [ + 55, + 475, + 295, + 608 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 475, + 295, + 608 + ], + "type": "text", + "content": ". We focus on a practical scenario where a shared pre-trained model " + }, + { + "bbox": [ + 55, + 475, + 295, + 608 + ], + "type": "inline_equation", + "content": "\\theta_{pre}" + }, + { + "bbox": [ + 55, + 475, + 295, + 608 + ], + "type": "text", + "content": " serves as the base for all fine-tuned models. It is formulated as " + }, + { + "bbox": [ + 55, + 475, + 295, + 608 + ], + "type": "inline_equation", + "content": "\\theta_m = \\lambda \\sum_{k=1}^K \\hat{G}(\\theta_k)" + }, + { + "bbox": [ + 55, + 475, + 295, + 608 + ], + "type": "text", + "content": ", and with task vector, it becomes " + }, + { + "bbox": [ + 55, + 475, + 295, + 608 + ], + "type": "inline_equation", + "content": "\\theta_m = \\theta_{pre} + \\lambda \\Sigma_{k=1}^K G(v_k)" + }, + { + "bbox": [ + 55, + 475, + 295, + 608 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 475, + 295, + 608 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 475, + 295, + 608 + ], + "type": "text", + "content": " is the coefficient, and " + }, + { + "bbox": [ + 55, + 475, + 295, + 608 + ], + "type": "inline_equation", + "content": "\\hat{G}/G" + }, + { + "bbox": [ + 55, + 475, + 295, + 608 + ], + "type": "text", + "content": " is transformation on the parameters. Assuming a shared pre-trained model aligns with trends toward foundation models, like LLaMa [40] for language and CLIP [37] for vision, which are then fine-tuned for specific tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 620, + 275, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 620, + 275, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 620, + 275, + 634 + ], + "type": "text", + "content": "4. Task Interference in Frequency Domain" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 641, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 700 + ], + "type": "text", + "content": "Task interference is a key factor that leads to the decrease in merging performance. While current methods focus on addressing interference in the spatial domain [48, 50, 53], we observe that task vectors from different fine-tuned models also exhibit great misalignment in the frequency domain." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 65, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 65, + 701, + 295, + 714 + ], + "type": "text", + "content": "As shown in Fig. 2(a), we visualize the amplitude power" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 331, + 101, + 538, + 181 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 555, + 93 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 93 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 93 + ], + "type": "text", + "content": "Table 1. The impact of current methods on the mean variance of frequency-domain amplitude across 8 image classification tasks." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 331, + 101, + 538, + 181 + ], + "lines": [ + { + "bbox": [ + 331, + 101, + 538, + 181 + ], + "spans": [ + { + "bbox": [ + 331, + 101, + 538, + 181 + ], + "type": "table", + "html": "
MethodViT-B/32ViT-L/14
Task Arithmetic [21]0.0590.110
DARE [53]0.057(↓3%)0.108(↓2%)
Ties-Merging [48]0.058(↓2%)0.109(↓1%)
Breadcrumbs [7]0.058(↓2%)0.107(↓3%)
PCB-Merging [12]0.056(↓5%)0.107(↓3%)
FR-Merging(ours)0.045(↓24%)0.088(↓20%)
", + "image_path": "3ef97cf4dde10b5f478606464124ebaa33a7c12d746f6e3ed6b686c98c93b412.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 198, + 553, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 553, + 282 + ], + "type": "text", + "content": "distribution in the frequency domain of task vectors from a ViT-B-32 attention head fine-tuned on various tasks [21]. Here, greater distance from the center represents a higher frequency. Since amplitude indicates signal strength, merging two signals with vastly different strengths inevitably results in the weaker signal being masked. Therefore, task interference in the low-frequency part are severe." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 282, + 554, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 554, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 554, + 354 + ], + "type": "text", + "content": "Furthermore, in Fig.2(b), we train multiple groups of models on eight image tasks [21] and find a clear negative correlation between the variance of amplitude power among models and the normalized merging performance (details can be found in Appendix D.1), confirming the great impact of frequency-domain interference on merging performance." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 354, + 553, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 553, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 553, + 402 + ], + "type": "text", + "content": "Additionally, as shown in Table 1, we show that existing works addressing task interference in the spatial domain fail to effectively mitigate frequency-domain task interference. Therefore, we propose our approach to tackle this issue." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 411, + 397, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 411, + 397, + 426 + ], + "spans": [ + { + "bbox": [ + 314, + 411, + 397, + 426 + ], + "type": "text", + "content": "5. Methodology" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 431, + 554, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 554, + 491 + ], + "type": "text", + "content": "This section presents FREE-Merging: Sec. 5.1 explains the need for a two-part design, Sec. 5.2 and Sec. 5.3 cover the FR-Merging for the backbone and the lightweight expert module, respectively, followed by the FREE-Merging in Sec. 5.4 and a complexity analysis in Sec. 5.5." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 498, + 389, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 389, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 389, + 509 + ], + "type": "text", + "content": "5.1. Motivation" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 515, + 553, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 515, + 553, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 515, + 553, + 539 + ], + "type": "text", + "content": "We first analyze why both merging optimization for the backbone and the introduction of experts are necessary." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 539, + 554, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 539, + 554, + 621 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 554, + 621 + ], + "type": "text", + "content": "First, backbone optimization is crucial but overlooked by current expert-based merging ways. As the backbone holds most parameters, its performance is key to the overall success. For instance, upgrading from LLaMa2 [40] to LLaMa3 [13] greatly improves results even with the identical fine-tuning. Thus, we propose a training-free merging method to reduce task interference in the backbone." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 623, + 554, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 623, + 554, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 623, + 554, + 671 + ], + "type": "text", + "content": "Next, we present a theorem showing the necessity of task experts. Based on Theorem 5.1, introducing additional information is essential. For proof, see Appendix B, and the effectiveness of task experts is proved in Appendix D.4." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "Theorem 5.1 Model Merging has no free lunch. Merged model " + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\theta_{m}" + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": " cannot simultaneously retain the capabilities of " + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\{\\theta_k\\}_{k=1}^K" + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": " without introducing additional information." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3865" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 138, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 138, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 138, + 85 + ], + "type": "text", + "content": "5.2. FR-Merging" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 173 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 173 + ], + "type": "text", + "content": "We propose an effective backbone merging method by addressing task interference in the low-frequency fine-tuned parameters as discussed in Sec. 4, which is hard to resolve in the spatial domain. Given the high redundancy of fine-tuned parameters [53], we apply high-pass filtering to directly remove the interfering regions. Let " + }, + { + "bbox": [ + 55, + 89, + 296, + 173 + ], + "type": "inline_equation", + "content": "v(x,y)" + }, + { + "bbox": [ + 55, + 89, + 296, + 173 + ], + "type": "text", + "content": " be the task vector and " + }, + { + "bbox": [ + 55, + 89, + 296, + 173 + ], + "type": "inline_equation", + "content": "G(x,y)" + }, + { + "bbox": [ + 55, + 89, + 296, + 173 + ], + "type": "text", + "content": " the transformed result. Then we have:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 93, + 178, + 295, + 192 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 178, + 295, + 192 + ], + "spans": [ + { + "bbox": [ + 93, + 178, + 295, + 192 + ], + "type": "interline_equation", + "content": "G (x, y) = \\mathcal {F} ^ {- 1} \\{H (\\eta , \\gamma) \\cdot \\mathcal {F} \\{v (x, y) \\} \\}, \\tag {1}", + "image_path": "bd647350094cc4ad8e08f0c6ecd5dc8757b7f3d09d92d6fc442614a38e04c75b.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 198, + 84, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 198, + 84, + 209 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 84, + 209 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 95, + 213, + 295, + 244 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 213, + 295, + 244 + ], + "spans": [ + { + "bbox": [ + 95, + 213, + 295, + 244 + ], + "type": "interline_equation", + "content": "H (\\eta , \\gamma) = \\left\\{ \\begin{array}{l l} 1, & \\sqrt {\\eta^ {2} + \\gamma^ {2}} \\geq D _ {0} \\\\ 0, & \\sqrt {\\eta^ {2} + \\gamma^ {2}} < D _ {0} \\end{array} \\right., \\tag {2}", + "image_path": "2033ba92bb1e35e371841f08a5efe3bde57aa4e13a450c3c06070590bb57a0ca.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 251, + 296, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 296, + 298 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 296, + 298 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 251, + 296, + 298 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 251, + 296, + 298 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 251, + 296, + 298 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{-1}" + }, + { + "bbox": [ + 55, + 251, + 296, + 298 + ], + "type": "text", + "content": " represent the Fourier and inverse Fourier transform, respectively. " + }, + { + "bbox": [ + 55, + 251, + 296, + 298 + ], + "type": "inline_equation", + "content": "D_0" + }, + { + "bbox": [ + 55, + 251, + 296, + 298 + ], + "type": "text", + "content": " is the adjustable cutoff frequency. And " + }, + { + "bbox": [ + 55, + 251, + 296, + 298 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 251, + 296, + 298 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 251, + 296, + 298 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 251, + 296, + 298 + ], + "type": "text", + "content": " are the distances from the frequency domain center along the " + }, + { + "bbox": [ + 55, + 251, + 296, + 298 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 251, + 296, + 298 + ], + "type": "text", + "content": "- and " + }, + { + "bbox": [ + 55, + 251, + 296, + 298 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 251, + 296, + 298 + ], + "type": "text", + "content": "-axes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 299, + 296, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 299, + 296, + 393 + ], + "spans": [ + { + "bbox": [ + 55, + 299, + 296, + 393 + ], + "type": "text", + "content": "Next, we analyze why high-pass filtering is effective from both intuitive and experimental perspectives. Fine-tuned weights occupy distinct positions in the loss landscape, and linear interpolation often leads to high-loss regions, causing task interference [1, 50]. To address this, we minimize model differences while preserving performance, bringing them closer in the loss landscape to increase the likelihood of the merged model falling into a loss basin [45]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 394, + 296, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 394, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 394, + 296, + 502 + ], + "type": "text", + "content": "Results in Sec. 4 show great differences in low-frequency regions. This is because, if we consider neural network parameters as signals carrying training information, inspired by image processing [3, 35], high-frequency signals represent sharp variations, while low-frequency parts define the overall structure [5]. Since most information resides in the structure, filtering out low-frequency signals removes key task-specific information. Then, the likelihood that the interpolation falls within the loss basin increases [1]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 502, + 296, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 296, + 621 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 296, + 621 + ], + "type": "text", + "content": "Experimentally, we validate that high-pass filtering enhances model generalization with minimal performance loss as shown in Fig 3. We fine-tune ViT-B/16 [37] on 30 visual tasks following [20]. We find that removing low-frequency information results in a smaller performance drop (diagonal) compared to the improvement in generalization (off-diagonal). This indicates that low-frequency parts carry task-specific information that boosts task performance but reduces generalization. Thus, applying filtering like FR-Merging helps alleviate task interference during merging." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 622, + 296, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 622, + 296, + 645 + ], + "spans": [ + { + "bbox": [ + 55, + 622, + 296, + 645 + ], + "type": "text", + "content": "To maintain consistent output after merging, we define the merging coefficient " + }, + { + "bbox": [ + 55, + 622, + 296, + 645 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + }, + { + "bbox": [ + 55, + 622, + 296, + 645 + ], + "type": "text", + "content": " for the task vector " + }, + { + "bbox": [ + 55, + 622, + 296, + 645 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 55, + 622, + 296, + 645 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 120, + 651, + 295, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 651, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 120, + 651, + 295, + 685 + ], + "type": "interline_equation", + "content": "\\lambda_ {i} = \\mathbb {E} (v _ {i}) \\left(\\sum_ {j = 1} ^ {K} \\mathbb {E} (v _ {j})\\right) ^ {- 1}, \\tag {3}", + "image_path": "41d0b6c7eb72454c323709949ede4423b280014553482005c8e19ed36e255919.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{E}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " represents the average. Overall, we construct a merged backbone in this part that minimizes task conflicts." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 328, + 72, + 538, + 247 + ], + "blocks": [ + { + "bbox": [ + 328, + 72, + 538, + 247 + ], + "lines": [ + { + "bbox": [ + 328, + 72, + 538, + 247 + ], + "spans": [ + { + "bbox": [ + 328, + 72, + 538, + 247 + ], + "type": "image", + "image_path": "66191ca272013351a4299d6a43952fbf48fa82abb40d63ac0e1a23a9b84c2275.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 251, + 555, + 296 + ], + "lines": [ + { + "bbox": [ + 313, + 251, + 555, + 296 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 555, + 296 + ], + "type": "text", + "content": "Figure 3. High-pass filtering improves performance over the unfiltered version. Rows represent the added task vectors, while columns are the test datasets. The results show a significant boost in generalization with only a slight decline in task performance." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 313, + 481, + 326 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 313, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 313, + 481, + 326 + ], + "type": "text", + "content": "5.3. Lightweight Expert Extraction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 331, + 554, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 554, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 554, + 402 + ], + "type": "text", + "content": "To compensate for information lost due to discarding low-frequency signals, we utilize lightweight experts added during inference. As analyzed in Sec. 5.2, low-frequency signals contain task-specific information, but preserving them requires an inefficient inverse Fourier transform during each inference. Thus, we propose a more efficient way." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 402, + 554, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 554, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 554, + 474 + ], + "type": "text", + "content": "Since fine-tuning leads to task specialization, the most changed parameters can be approximated as having the greatest impact [32]. After reasonable rescaling, these selected parameters serve as task experts, minimizing storage (typically around " + }, + { + "bbox": [ + 313, + 402, + 554, + 474 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 402, + 554, + 474 + ], + "type": "text", + "content": " of parameters) without extra inference computation. The validity is detailed in Appendix D.4." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 474, + 554, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 554, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 554, + 545 + ], + "type": "text", + "content": "Although magnitude-based extraction [48] exists, limited research on rescaling hinders compression. As rescaling is crucial to expert extraction, we propose a rescaling method. Let " + }, + { + "bbox": [ + 313, + 474, + 554, + 545 + ], + "type": "inline_equation", + "content": "M(v_{i},d)" + }, + { + "bbox": [ + 313, + 474, + 554, + 545 + ], + "type": "text", + "content": " represent the top- " + }, + { + "bbox": [ + 313, + 474, + 554, + 545 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 474, + 554, + 545 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 474, + 554, + 545 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 474, + 554, + 545 + ], + "type": "text", + "content": " is a percentage) parameters for task vector " + }, + { + "bbox": [ + 313, + 474, + 554, + 545 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 313, + 474, + 554, + 545 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 474, + 554, + 545 + ], + "type": "inline_equation", + "content": "\\mu_{i}" + }, + { + "bbox": [ + 313, + 474, + 554, + 545 + ], + "type": "text", + "content": " be the rescale factor, and " + }, + { + "bbox": [ + 313, + 474, + 554, + 545 + ], + "type": "inline_equation", + "content": "e(v_{i})" + }, + { + "bbox": [ + 313, + 474, + 554, + 545 + ], + "type": "text", + "content": " be the expert for task " + }, + { + "bbox": [ + 313, + 474, + 554, + 545 + ], + "type": "inline_equation", + "content": "t_{i}" + }, + { + "bbox": [ + 313, + 474, + 554, + 545 + ], + "type": "text", + "content": ". Then we have:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 323, + 551, + 553, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 551, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 323, + 551, + 553, + 578 + ], + "type": "interline_equation", + "content": "e \\left(v _ {i}\\right) = \\mu_ {i} M \\left(v _ {i}, d\\right), \\mu_ {i} = - \\frac {\\mathbb {E} \\left(M \\left(v _ {i} , d\\right)\\right) \\cdot \\log (d)}{\\lambda_ {i} \\cdot \\mathbb {E} \\left(v _ {i}\\right)}. \\tag {4}", + "image_path": "67bcc4cc56cdfec0f5229eafa64b36c00fab728fc550807787e1bc5da6ecfb00.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 582, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 605 + ], + "type": "text", + "content": "The goal is to maintain a consistent output after extraction to preserve performance. Details are in Appendix D.5." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 606, + 554, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 554, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 554, + 641 + ], + "type": "text", + "content": "In summary, we efficiently extract task-specific experts to integrate into the backbone, mitigating task conflicts without excessive storage or inference overhead." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 647, + 410, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 410, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 410, + 661 + ], + "type": "text", + "content": "5.4. FREE-Merging" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "This section presents FREE-Merging, combining both parts to make it suitable for multiple tasks without affecting inference speed. Inspired by MoE [39], a router dynamically assigns tasks to activated experts. The router can take" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3866" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 87, + 555, + 246 + ], + "blocks": [ + { + "bbox": [ + 60, + 71, + 549, + 83 + ], + "lines": [ + { + "bbox": [ + 60, + 71, + 549, + 83 + ], + "spans": [ + { + "bbox": [ + 60, + 71, + 549, + 83 + ], + "type": "text", + "content": "Table 2. Comparison of model merging performance using ViT-B/32 (B) and ViT-L/14 (L) on eight visual tasks evaluating by accuracy." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 87, + 555, + 246 + ], + "lines": [ + { + "bbox": [ + 58, + 87, + 555, + 246 + ], + "spans": [ + { + "bbox": [ + 58, + 87, + 555, + 246 + ], + "type": "table", + "html": "
MethodBase ModelSUN397CarsRESISC45EuroSATSVHNGTSRBMNISTDTDAvg.
BLBLBLBLBLBLBLBLBL
Individual75.382.377.792.496.197.499.710097.598.198.799.299.799.779.484.190.594.2
Traditional MTL73.980.874.490.693.996.398.296.395.897.698.999.199.599.677.984.488.993.5
Weight Averaging65.372.163.481.671.482.671.791.964.278.252.870.787.597.150.162.865.879.6
Fisher Merging [33]68.669.269.288.670.787.566.493.572.980.651.174.887.993.359.970.068.382.2
Task Arithmetic [21]63.874.162.182.172.086.777.693.874.487.965.186.894.098.952.265.670.184.5
Ties-Merging [48]64.876.562.985.074.389.378.995.783.190.371.483.397.699.056.268.873.686.0
Breadcrumbs [7]63.074.961.084.975.788.684.795.483.489.876.490.198.199.257.968.275.086.4
PCB-Merging [12]65.575.864.186.078.188.680.296.084.788.077.190.998.099.158.470.075.886.9
FR-Merging(ours)66.276.464.587.077.290.290.196.885.492.082.392.898.599.360.071.578.188.3
AdaMerging++ [52]66.679.468.390.382.291.694.297.489.693.489.097.598.399.060.679.281.191.0
Twin-Merging [29]73.682.671.790.392.194.999.399.495.396.797.298.099.199.474.080.187.892.7
EMR-Merging [20]74.182.372.790.891.995.399.499.595.896.996.998.199.199.572.180.187.792.8
FREE-Merging(ours)77.183.578.292.493.496.299.599.696.398.198.298.799.599.775.481.789.793.7
", + "image_path": "88035de48d64a604e7638c0e6834371651ba0277cff9739898a1e8005c2830cd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 265, + 228, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 265, + 228, + 277 + ], + "spans": [ + { + "bbox": [ + 57, + 265, + 228, + 277 + ], + "type": "text", + "content": "Algorithm 1 Workflow of FREE-Merging" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 281, + 294, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 281, + 294, + 305 + ], + "spans": [ + { + "bbox": [ + 57, + 281, + 294, + 305 + ], + "type": "text", + "content": "Input: Task vectors " + }, + { + "bbox": [ + 57, + 281, + 294, + 305 + ], + "type": "inline_equation", + "content": "\\{v_{i}\\}_{i = 1}^{K}" + }, + { + "bbox": [ + 57, + 281, + 294, + 305 + ], + "type": "text", + "content": ", pre-trained model " + }, + { + "bbox": [ + 57, + 281, + 294, + 305 + ], + "type": "inline_equation", + "content": "\\theta_{pre}" + }, + { + "bbox": [ + 57, + 281, + 294, + 305 + ], + "type": "text", + "content": ", router " + }, + { + "bbox": [ + 57, + 281, + 294, + 305 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 57, + 281, + 294, + 305 + ], + "type": "text", + "content": ", and input data " + }, + { + "bbox": [ + 57, + 281, + 294, + 305 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 57, + 281, + 294, + 305 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 307, + 294, + 437 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 63, + 307, + 132, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 307, + 132, + 316 + ], + "spans": [ + { + "bbox": [ + 63, + 307, + 132, + 316 + ], + "type": "text", + "content": "1: FR-Merging:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 231, + 308, + 291, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 308, + 291, + 316 + ], + "spans": [ + { + "bbox": [ + 231, + 308, + 291, + 316 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 231, + 308, + 291, + 316 + ], + "type": "text", + "content": " Only excute once." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 316, + 294, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 316, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 62, + 316, + 294, + 341 + ], + "type": "text", + "content": "2: Get backbone " + }, + { + "bbox": [ + 62, + 316, + 294, + 341 + ], + "type": "inline_equation", + "content": "\\theta_{m} = \\theta_{pre} + \\sum_{i=1}^{K} \\lambda_{i} G(v_{i})" + }, + { + "bbox": [ + 62, + 316, + 294, + 341 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 62, + 316, + 294, + 341 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 62, + 316, + 294, + 341 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 62, + 316, + 294, + 341 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 62, + 316, + 294, + 341 + ], + "type": "text", + "content": " are defined by Eq. 1 and Eq. 3, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 342, + 155, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 342, + 155, + 353 + ], + "spans": [ + { + "bbox": [ + 62, + 342, + 155, + 353 + ], + "type": "text", + "content": "3: Expert Extraction:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 231, + 342, + 291, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 342, + 291, + 352 + ], + "spans": [ + { + "bbox": [ + 231, + 342, + 291, + 352 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 231, + 342, + 291, + 352 + ], + "type": "text", + "content": " Only execute once." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 353, + 209, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 353, + 209, + 365 + ], + "spans": [ + { + "bbox": [ + 62, + 353, + 209, + 365 + ], + "type": "text", + "content": "4: Get task experts " + }, + { + "bbox": [ + 62, + 353, + 209, + 365 + ], + "type": "inline_equation", + "content": "\\{e_i\\}_{i = 1}^K" + }, + { + "bbox": [ + 62, + 353, + 209, + 365 + ], + "type": "text", + "content": " as Eq. 4." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 366, + 118, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 366, + 118, + 376 + ], + "spans": [ + { + "bbox": [ + 62, + 366, + 118, + 376 + ], + "type": "text", + "content": "5: Inference:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 377, + 130, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 377, + 130, + 388 + ], + "spans": [ + { + "bbox": [ + 62, + 377, + 130, + 388 + ], + "type": "text", + "content": "6: for " + }, + { + "bbox": [ + 62, + 377, + 130, + 388 + ], + "type": "inline_equation", + "content": "x \\in X" + }, + { + "bbox": [ + 62, + 377, + 130, + 388 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 388, + 208, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 388, + 208, + 401 + ], + "spans": [ + { + "bbox": [ + 62, + 388, + 208, + 401 + ], + "type": "text", + "content": "7: " + }, + { + "bbox": [ + 62, + 388, + 208, + 401 + ], + "type": "inline_equation", + "content": "[w_{1},\\dots w_{K}]\\gets argmax(R(x))" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 401, + 176, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 401, + 176, + 415 + ], + "spans": [ + { + "bbox": [ + 62, + 401, + 176, + 415 + ], + "type": "text", + "content": "8: " + }, + { + "bbox": [ + 62, + 401, + 176, + 415 + ], + "type": "inline_equation", + "content": "\\theta_{*} = \\theta_{m} + \\sum_{i = 1}^{K}w_{i}e_{i}" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 415, + 160, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 160, + 426 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 160, + 426 + ], + "type": "text", + "content": "9: " + }, + { + "bbox": [ + 62, + 415, + 160, + 426 + ], + "type": "inline_equation", + "content": "Y\\gets Y\\cup f(x;\\theta_{*})" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 62, + 427, + 106, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 427, + 106, + 437 + ], + "spans": [ + { + "bbox": [ + 62, + 427, + 106, + 437 + ], + "type": "text", + "content": "10: end for" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 57, + 438, + 187, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 438, + 187, + 450 + ], + "spans": [ + { + "bbox": [ + 57, + 438, + 187, + 450 + ], + "type": "text", + "content": "Output: Output " + }, + { + "bbox": [ + 57, + 438, + 187, + 450 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 57, + 438, + 187, + 450 + ], + "type": "text", + "content": " for input " + }, + { + "bbox": [ + 57, + 438, + 187, + 450 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 57, + 438, + 187, + 450 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 474, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 510 + ], + "type": "text", + "content": "various forms [24], allowing for more flexibility in handling different tasks. Detailed discussions about router can be found in Appendix D.6. For detailed algorithms, see Alg 1." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 510, + 295, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 546 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 546 + ], + "type": "text", + "content": "First, we obtain the merged backbone (line 2) and the task experts (line 4). Then, during the inference, we dynamically route the experts based on inputs (lines 7-9)." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 552, + 291, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 291, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 291, + 566 + ], + "type": "text", + "content": "5.5. Time Complexity and Storage Space Analysis" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 570, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 617 + ], + "type": "text", + "content": "This section analyzes the merging time complexity and storage requirements. We aim to minimize both for practical use. Assume there are " + }, + { + "bbox": [ + 55, + 570, + 295, + 617 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 570, + 295, + 617 + ], + "type": "text", + "content": " models to merge, each with " + }, + { + "bbox": [ + 55, + 570, + 295, + 617 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 570, + 295, + 617 + ], + "type": "text", + "content": " parameters and requiring storage space " + }, + { + "bbox": [ + 55, + 570, + 295, + 617 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 570, + 295, + 617 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": "The time complexity includes the merging and inference part. In the merging part, FR-Merging has a complexity of " + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "inline_equation", + "content": "O(nmlogm)" + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": " [14] while expert extraction is " + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "inline_equation", + "content": "O(nm)" + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": ", making the overall merging complexity " + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "inline_equation", + "content": "O(nmlogm)" + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": ", which is manageable. For instance, merging three 10B models requires only " + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "inline_equation", + "content": "O(10^{12})" + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": " addition operations, a minor load for modern devices. Moreover, with only a lightweight router added, the extra inference cost is negligible." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 265, + 555, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 265, + 555, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 555, + 337 + ], + "type": "text", + "content": "Next, we analyze storage cost. We typically save only about " + }, + { + "bbox": [ + 313, + 265, + 555, + 337 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 265, + 555, + 337 + ], + "type": "text", + "content": " additional parameters per model, requiring " + }, + { + "bbox": [ + 313, + 265, + 555, + 337 + ], + "type": "inline_equation", + "content": "(1 + 0.01n)s" + }, + { + "bbox": [ + 313, + 265, + 555, + 337 + ], + "type": "text", + "content": " storage, which is far less than the requirement of " + }, + { + "bbox": [ + 313, + 265, + 555, + 337 + ], + "type": "inline_equation", + "content": "ns" + }, + { + "bbox": [ + 313, + 265, + 555, + 337 + ], + "type": "text", + "content": " for the model ensemble. For large models, reducing storage requirements is crucial for alleviating storage pressure on edge users. For detailed analysis, please refer to Sec 6.7." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 346, + 395, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 395, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 395, + 360 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 366, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 555, + 510 + ], + "type": "text", + "content": "In this section, we evaluate the proposed methods across various tasks. We compare FR-Merging with other cost-free methods (which do not require additional training or data), including Weight Averaging, Fisher Merging [33], Task Arithmetic [21], Ties-Merging [48], Bubbles [7], and PCB-Merging [12]. We also compare our FREE-Merging with popular methods, including RegMean [23], AdaMerging [52], EMR-Merging [20], and Twin-Merging [29]. The original EMR-Merging uses a perfect router; however, we use an imperfect router for realism. Some experiments include traditional MTL with data from all tasks as a baseline. The evaluation covers CV, NLP, and multi-modal tasks." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 517, + 446, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 446, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 446, + 529 + ], + "type": "text", + "content": "6.1. Merging Vision Models" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 533, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 617 + ], + "type": "text", + "content": "Merging 8 ViTs. We select the two visual encoders of CLIP, ViT-B/32 and ViT-L/14 [37], as the pre-trained models and choose eight image classification tasks as target tasks: SUN397, Cars, RESISC45, EuroSAT, SVHN, GT-SRB, MNIST, and DTD [52]. Additionally, an MLP is used as the lightweight router. Detailed information on the datasets and baselines can be found in the Appendix C.1." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "type": "text", + "content": "Table 2 presents the comparative results. First, FR-Merging outperforms other cost-free methods, improving ViT-B/32 by " + }, + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "type": "text", + "content": " and ViT-L/14 by " + }, + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "type": "text", + "content": " over PCB-Merging [12], demonstrating its effectiveness in reducing performance loss caused by task interference. Through FR-Merging, we can construct a merged backbone that can be used independently at a low cost. Here, we focus on the basic FR-Merging. In Sec. 6.6, we discuss further im" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3867" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 72, + 107, + 278, + 242 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "type": "text", + "content": "Table 3. Comparison of model merging results using ViT-B/16 on 30 visual tasks. Additional Cost indicates whether additional training or additional data is required during model merging." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 72, + 107, + 278, + 242 + ], + "lines": [ + { + "bbox": [ + 72, + 107, + 278, + 242 + ], + "spans": [ + { + "bbox": [ + 72, + 107, + 278, + 242 + ], + "type": "table", + "html": "
MethodAdditional CostAvg. Acc
IndividualX93.02
Weight AveragingX42.51
Task Arithmetic [21]X48.88
Ties-Merging [48]X37.53
Breadcrumbs [7]X43.57
FR-Merging(ours)X53.90
RegMean [23]✓68.13
AdaMerging [52]✓60.24
Twin-Merging [29]✓75.52
EMR-Merging [20]✓76.39
FREE-Merging(ours)✓79.67
", + "image_path": "26fe7aa9bb5dbfb6eeb4bc24f2ab7cc890cc3488caefa089a434a7100987540a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 58, + 292, + 299, + 446 + ], + "blocks": [ + { + "bbox": [ + 55, + 255, + 296, + 289 + ], + "lines": [ + { + "bbox": [ + 55, + 255, + 296, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 296, + 289 + ], + "type": "text", + "content": "Table 4. Language model merging results: RoBERTa fine-tuned on 8 discriminative tasks, T0-3B on 11 discriminative tasks with " + }, + { + "bbox": [ + 55, + 255, + 296, + 289 + ], + "type": "inline_equation", + "content": "IA^3" + }, + { + "bbox": [ + 55, + 255, + 296, + 289 + ], + "type": "text", + "content": ", and Qwen-14B on 3 generative tasks with LoRA." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 292, + 299, + 446 + ], + "lines": [ + { + "bbox": [ + 58, + 292, + 299, + 446 + ], + "spans": [ + { + "bbox": [ + 58, + 292, + 299, + 446 + ], + "type": "table", + "html": "
MethodBase ModelFull RoBERTaPEFTAvg.
T0-3BQwen-14B
Individual85.5571.3572.0776.32
Weight Averaging51.3458.0166.8058.71
Task Arithmetic[21]66.6563.9166.4065.65
Ties-Merging [48]64.0466.4167.4665.97
Breadcrumbs [7]68.1954.6466.7763.20
PCB-Merging [12]67.8666.1066.6966.39
FR-Merging(ours)70.0266.8868.0068.30
RegMean [23]70.0158.0167.5265.18
EMR-Merging [20]74.2067.1170.9870.76
Twin-Merging [29]78.2966.7071.6872.22
FREE-Merging(ours)80.1668.6872.7873.87
", + "image_path": "a804f2ae98b06a8237374aa3f6563f9c252e7962508f7dd4a5bed7568ca42fc3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 466, + 295, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 466, + 295, + 562 + ], + "spans": [ + { + "bbox": [ + 54, + 466, + 295, + 562 + ], + "type": "text", + "content": "provements achieved by combining it with existing methods. Then, compared to methods requiring extra information, FREE-Merging also improves performance by " + }, + { + "bbox": [ + 54, + 466, + 295, + 562 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 54, + 466, + 295, + 562 + ], + "type": "text", + "content": ", leveraging lightweight computation and efficient expert extraction, requiring only " + }, + { + "bbox": [ + 54, + 466, + 295, + 562 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 54, + 466, + 295, + 562 + ], + "type": "text", + "content": " additional storage per model, outperforming " + }, + { + "bbox": [ + 54, + 466, + 295, + 562 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 54, + 466, + 295, + 562 + ], + "type": "text", + "content": " of EMR-Merging [20] and " + }, + { + "bbox": [ + 54, + 466, + 295, + 562 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 54, + 466, + 295, + 562 + ], + "type": "text", + "content": " of Twin-Merging [29]. Thus, FREE-Merging provides an excellent balance between storage, computation, and performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 565, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 565, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 565, + 296, + 601 + ], + "type": "text", + "content": "Merging 30 ViTs. We select ViT-B/16 as the pre-trained model and evaluate the model merging performance across 30 image classification tasks following [20]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "type": "text", + "content": "Table 3 presents the results, with detailed analysis provided in Appendix E.1. It is evident that our proposed FR-Merging and FREE-Merging exhibit remarkable performance. Although merging a large number of models is uncommon in practice, this experiment underscores the effectiveness of Fourier filtering and lightweight task experts in mitigating performance loss caused by task interference. These results highlight the robustness and broad applicability of our methods across diverse scenarios." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 319, + 112, + 548, + 247 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 105 + ], + "type": "text", + "content": "Table 5. Comparison of performance of different model merging methods after full fine-tuning LLaMa2-13B on three tasks: Instruction Following (LM), Math, and Code Generation." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 112, + 548, + 247 + ], + "lines": [ + { + "bbox": [ + 319, + 112, + 548, + 247 + ], + "spans": [ + { + "bbox": [ + 319, + 112, + 548, + 247 + ], + "type": "table", + "html": "
MethodAlpacaEvalGSM8KMBPPAvg.
LM Fine-tuned [46]88.945.931.455.4
Math Fine-tuned [30]22.363.523.036.5
Code Fine-tuned [31]16.218.827.020.7
Task Arithmetic [21]72.148.3040.1
TIES-Merging [48]77.566.927.257.2
DARE [53]77.562.729.456.5
DELLA-Merging [8]80.461.831.457.9
PCB-Merging [12]81.260.630.757.5
FR-Merging(ours)87.564.332.861.5
Twin-Merging [29]83.561.430.458.4
EMR-Merging [20]82.662.331.258.7
FREE-Merging(ours)88.165.731.861.8
", + "image_path": "758c314aec4e7191d7eb469dc59390f3a718cab1bfd8a784360791238241199c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 263, + 463, + 276 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 263, + 463, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 463, + 276 + ], + "type": "text", + "content": "6.2. Merging Language Models" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 281, + 555, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 341 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 341 + ], + "type": "text", + "content": "Merging Medium-sized Language Models. We investigate the generalization ability of our method on medium-scale language models, using RoBERTa [59] as the pretrained model with full fine-tuning on eight discriminative tasks [48]. Details can be found in Appendix C.1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 341, + 556, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 556, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 556, + 437 + ], + "type": "text", + "content": "The average merging results of the eight tasks are presented in Table 4, with complete results in Appendix E.2. It can be observed that, FR-Merging outperforms other cost-free methods by nearly " + }, + { + "bbox": [ + 313, + 341, + 556, + 437 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 313, + 341, + 556, + 437 + ], + "type": "text", + "content": ". Additionally, FREE-Merging also exhibits remarkable effectiveness, achieving optimal results on almost all datasets and surpassing current methods in average performance. This highlights the great potential of our proposed methods for language models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 437, + 554, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 497 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 497 + ], + "type": "text", + "content": "Merging PEFT models. PEFT is a popular method that reduces fine-tuning costs [10], but storing PEFT parameters for multiple tasks remains space-inefficient [55]. The effectiveness of our approach on PEFT demonstrates the generalization capability and offers an efficient deployment way." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 498, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 617 + ], + "type": "text", + "content": "Firstly, we use T0-3B [38] as the pre-trained model and " + }, + { + "bbox": [ + 313, + 498, + 555, + 617 + ], + "type": "inline_equation", + "content": "(IA)^{3}" + }, + { + "bbox": [ + 313, + 498, + 555, + 617 + ], + "type": "text", + "content": " [26] as the PEFT method, fine-tuning on 11 discriminative tasks [20]. Our proposed FR-Merging and FREE-Merging improve performance by nearly " + }, + { + "bbox": [ + 313, + 498, + 555, + 617 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 498, + 555, + 617 + ], + "type": "text", + "content": ", as shown in Table 4. Next, we employ Qwen-14B [2] with LoRA [19] as the PEFT way, fine-tuning on three generative tasks [29]. The results show that our methods lead to a " + }, + { + "bbox": [ + 313, + 498, + 555, + 617 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 313, + 498, + 555, + 617 + ], + "type": "text", + "content": " advantage in both cost-free and full frameworks. This confirms the feasibility with PEFT and strong generalization of our methods. Full results are available in the Appendix E.2." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 617, + 556, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 689 + ], + "type": "text", + "content": "Merging Large Language models. We explore large model merging using LLaMa2-13B [40] as the pre-trained model, with WizardLM [46] for instruction following, WizardMath [30] for math, and WizardCoder-Python [31] for coding. Experiments are conducted on AlpacaEval (instruction following), GSM8K (math), and MBPP (code) [53]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "Table 5 shows that our proposed FR-Merging even surpasses existing methods that rely on additional information." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3868" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 91, + 553, + 205 + ], + "blocks": [ + { + "bbox": [ + 148, + 71, + 461, + 82 + ], + "lines": [ + { + "bbox": [ + 148, + 71, + 461, + 82 + ], + "spans": [ + { + "bbox": [ + 148, + 71, + 461, + 82 + ], + "type": "text", + "content": "Table 6. Results of merging multi-modal BEiT3 models on five vision-language tasks." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 91, + 553, + 205 + ], + "lines": [ + { + "bbox": [ + 59, + 91, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 59, + 91, + 553, + 205 + ], + "type": "table", + "html": "
MethodsTask MetricCOCO-Retrieval Accuracy(↑)COCO-CaptioningImageNet-1k Classification Accuracy(↑)NLVR2 Accuracy(↑)VQAv2 Accuracy(↑)
BLEU4(↑)CIDEr(↑)METEOR(↑)ROUGE-L(↑)
Individual0.84560.3941.3370.3110.6010.85370.77650.8439
Weight Averaging0.18930.0310.0010.1150.1590.67710.28000.6285
Task Arithmetic [21]0.31770.0330.0000.1180.1760.67320.38090.6933
Ties-Merging [48]0.39290.0290.0010.1080.1670.69780.32060.6717
Breadcrumbs [7]0.18930.0110.0120.0650.1140.63230.60730.6823
FR-Merging(ours)0.70100.0160.0190.0960.1410.69920.64070.6965
EMR-Merging [20]0.79460.2891.0600.2720.5340.77420.74750.7211
Twin-Merging [29]0.79370.3251.0780.2780.5580.77300.77620.7232
FREE-Merging(ours)0.80930.3471.1950.2890.5720.78500.81680.7297
", + "image_path": "8642d43f40db61de4bea72337382dd561b78d11659556c5e06eb41fae70875a5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 64, + 222, + 282, + 385 + ], + "blocks": [ + { + "bbox": [ + 64, + 222, + 282, + 385 + ], + "lines": [ + { + "bbox": [ + 64, + 222, + 282, + 385 + ], + "spans": [ + { + "bbox": [ + 64, + 222, + 282, + 385 + ], + "type": "image", + "image_path": "50d6947d74695332a096a8c8ec33943b8c5967a9e46af2b5aa16628036d988bc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 395, + 295, + 429 + ], + "lines": [ + { + "bbox": [ + 55, + 395, + 295, + 429 + ], + "spans": [ + { + "bbox": [ + 55, + 395, + 295, + 429 + ], + "type": "text", + "content": "Figure 4. Visualization of the loss landscape on EuroSAT before and after high-pass filtering (top vs. bottom) in the model merging of the fine-tuned ViT-B/32 on EuroSAT and DTD tasks [21]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 453, + 295, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 295, + 525 + ], + "type": "text", + "content": "This indicates that our findings in FR-Merging regarding low-frequency signals and task interference also hold when extended to large models. This is crucial for practical usage of large models. Additionally, our proposed FREE-Merging also demonstrates strong advantages, enabling efficient cross-domain deployment of large models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 536, + 218, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 536, + 218, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 218, + 550 + ], + "type": "text", + "content": "6.3. Merging Multi-Modal Models" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 556, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 295, + 640 + ], + "type": "text", + "content": "We merge various fine-tuned BEiT3-base [42] models across five tasks: ImageNet-1k (Classification), VQAv2 (Visual Question Answering), NLVR2 (Reasoning), COCO Captioning (Image Captioning), and COCO Retrieval (Image-Text Retrieval). COCO Captioning is evaluated with BLEU4, CIDEr, METEOR, and ROUGE-L, while the others use accuracy [20]. Details are in Appendix C.1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Table 6 shows that our methods gain great improvements in multi-modal models. Unlike vision or language model merging, which often involve similar tasks, such as classification, the task similarity here is relatively low. The strong performance highlights the generalization ability of our methods, making them widely applicable." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 236, + 552, + 312 + ], + "blocks": [ + { + "bbox": [ + 314, + 217, + 553, + 228 + ], + "lines": [ + { + "bbox": [ + 314, + 217, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 314, + 217, + 553, + 228 + ], + "type": "text", + "content": "Table 7. Ablation of FREE-Merging components on 8 visual tasks." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 236, + 552, + 312 + ], + "lines": [ + { + "bbox": [ + 317, + 236, + 552, + 312 + ], + "spans": [ + { + "bbox": [ + 317, + 236, + 552, + 312 + ], + "type": "table", + "html": "
FR-MergingEqua.3Top-K ExpertScalingViT-B/32ViT-B/16ViT-L/14
XXXX70.1075.2484.50
✓XXX76.9681.3488.21
X✓XX71.4377.4184.20
✓✓XX78.1082.5488.30
XX✓X65.2468.9870.13
✓✓✓X79.3884.3189.53
XX✓✓82.3286.8790.47
✓✓✓✓89.6891.0993.70
", + "image_path": "818bb12cda6fcf28321d285d560a71489ef0dcdc1eb459a13105ecb45bc4d16a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 329, + 468, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 468, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 468, + 342 + ], + "type": "text", + "content": "6.4. Fourier Transform Analysis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 347, + 555, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 555, + 491 + ], + "type": "text", + "content": "In this section, we explain why Fourier high-pass filtering improves merging performance from an optimization perspective. Previous work [25] shows that landscape geometry impacts model generalization. Fig. 4 demonstrates that high-pass filtering widens the valley in the loss landscape (bottom plot). Since fine-tuned models from the same pretrained model are closer in the landscape [34], a wider valley increases the probability of merged results falling within the loss basin as shown in the bottom plot of Fig. 4, leading to better performance. Without filtering, the landscape is flat at a high level, which has little effect on model merging and only highlights the advantage of pre-training [17]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 501, + 408, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 501, + 408, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 501, + 408, + 514 + ], + "type": "text", + "content": "6.5. Ablation Study" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 520, + 555, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 555, + 555 + ], + "type": "text", + "content": "In this section, we conduct ablation experiments on the components of our method. All experiments are performed on ViT-B/32 with eight classification tasks as target tasks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 557, + 554, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 554, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 554, + 664 + ], + "type": "text", + "content": "Ablation on FR-Merging. We analyze the impact of FR-Merging on the backbone and overall performance. Table 7 shows that the performance of various models greatly improves after employing FR-Merging. We then compare high-pass filtering with other cost-free methods, including Ties [48], Top-K, Task Arithmetic [21], outlier dropout [7], EMR-Merging [20], and PCB-Merging [12], as well as low-pass and band-pass filtering, with results in Fig. 5a. For fairness, all experts use " + }, + { + "bbox": [ + 313, + 557, + 554, + 664 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 557, + 554, + 664 + ], + "type": "text", + "content": " additional parameters." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "type": "text", + "content": "Our proposed FR-Merging outperforms other cost-free methods on both the backbone and with task experts, minimizing task interference. Examining the different filtering methods, we find that low-pass filtering leads to great per" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3869" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 81, + 218, + 210 + ], + "blocks": [ + { + "bbox": [ + 61, + 81, + 218, + 210 + ], + "lines": [ + { + "bbox": [ + 61, + 81, + 218, + 210 + ], + "spans": [ + { + "bbox": [ + 61, + 81, + 218, + 210 + ], + "type": "image", + "image_path": "59fdbfc950c358e2b25b51dacb4bb240a829ea58170cf802bcae9e2c29a103ad.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 92, + 214, + 186, + 224 + ], + "lines": [ + { + "bbox": [ + 92, + 214, + 186, + 224 + ], + "spans": [ + { + "bbox": [ + 92, + 214, + 186, + 224 + ], + "type": "text", + "content": "(a) Ablation on FR-Merging." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 227, + 79, + 386, + 210 + ], + "blocks": [ + { + "bbox": [ + 227, + 79, + 386, + 210 + ], + "lines": [ + { + "bbox": [ + 227, + 79, + 386, + 210 + ], + "spans": [ + { + "bbox": [ + 227, + 79, + 386, + 210 + ], + "type": "image", + "image_path": "cda7c8558fef8b6b097b6692b1c16000d2be81238b533d7640ab996d79654359.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 258, + 214, + 352, + 224 + ], + "lines": [ + { + "bbox": [ + 258, + 214, + 352, + 224 + ], + "spans": [ + { + "bbox": [ + 258, + 214, + 352, + 224 + ], + "type": "text", + "content": "(b) Ablation on Task Expert." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 392, + 73, + 550, + 211 + ], + "blocks": [ + { + "bbox": [ + 392, + 73, + 550, + 211 + ], + "lines": [ + { + "bbox": [ + 392, + 73, + 550, + 211 + ], + "spans": [ + { + "bbox": [ + 392, + 73, + 550, + 211 + ], + "type": "image", + "image_path": "e20f278e0c623054f3f87346772bc71f55dfae6ccb70aa07ed87a51d957bc6d0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 438, + 214, + 504, + 224 + ], + "lines": [ + { + "bbox": [ + 438, + 214, + 504, + 224 + ], + "spans": [ + { + "bbox": [ + 438, + 214, + 504, + 224 + ], + "type": "text", + "content": "(c) Router Analysis." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 232, + 555, + 255 + ], + "lines": [ + { + "bbox": [ + 55, + 232, + 555, + 255 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 555, + 255 + ], + "type": "text", + "content": "Figure 5. a) The ablation experiments on FR-Merging show that our proposed methods achieve optimal performance. b) Performance of experts at different sparsity rates. c) In experiments with random routers, our approach demonstrates optimal performance." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 63, + 284, + 287, + 365 + ], + "blocks": [ + { + "bbox": [ + 56, + 270, + 294, + 281 + ], + "lines": [ + { + "bbox": [ + 56, + 270, + 294, + 281 + ], + "spans": [ + { + "bbox": [ + 56, + 270, + 294, + 281 + ], + "type": "text", + "content": "Table 8. Transferability analysis of FR-Merging on 8 visual tasks." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 284, + 287, + 365 + ], + "lines": [ + { + "bbox": [ + 63, + 284, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 63, + 284, + 287, + 365 + ], + "type": "table", + "html": "
MethodFR-MergingViT-B/32ViT-L/14
Ties-Merging [48]X73.686.0
Ties-Merging [48]✓78.388.4
Breadcrumbs [7]X75.086.4
Breadcrumbs [7]✓78.888.9
TSVM [16]X83.991.9
TSVM [16]✓84.892.5
", + "image_path": "90cb0b51041d63fb0f47fc7e3c746e12cb30dc59b676674dda7ab4e01f571a86.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 377, + 295, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 377, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 295, + 437 + ], + "type": "text", + "content": "formance degradation, confirming that task interference in low-frequency regions is severe. Band-pass filtering also performs worse than FR-Merging, and this is because some high-frequency signals preserve abrupt changes in the signal and their removal harms the performance of the model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 437, + 295, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 497 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 497 + ], + "type": "text", + "content": "Ablation on Merging Coefficients. We conduct an ablation study on our merging coefficients as in Eq. 3, with results in Table 7. Using them instead of simple averaging yields slight improvements by considering output variations, benefiting tasks with varying fine-tuning magnitudes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 497, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 295, + 605 + ], + "type": "text", + "content": "Ablation on Task Expert. We conduct an ablation study on the expert extraction way. Table 7 shows that incorporating our experts and rescaling method results in significant improvements, confirming their effectiveness. To validate the superiority of our method, we evaluate expert-only performance with varying parameter proportions, comparing Top-K selection, Bernoulli selection [53], SVD extraction [29], and our Top-K selection with rescaling. Fig. 5b shows that our method excels, particularly under high sparsity." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "Router Analysis. We analyze the proposed router method, aiming for strong merging performance despite sub-optimal routing. Since router inaccuracies may arise from limited training resources or edge deployment issues, our methods maintain performance in real-world scenarios. Using a random router to simulate inaccuracy, we compare our method with existing router-based approaches. Our method outperforms Twin-Merging by " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "13.5\\%" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " and EMR-Merging by " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "8.5\\%" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " on average, as shown in Fig. 5c, indicating that our methods" + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 316, + 296, + 554, + 364 + ], + "blocks": [ + { + "bbox": [ + 313, + 270, + 554, + 293 + ], + "lines": [ + { + "bbox": [ + 313, + 270, + 554, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 554, + 293 + ], + "type": "text", + "content": "Table 9. Trade-off of performance, training, inference, and storage costs using ViT-L/14 across 8 image classification tasks." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 296, + 554, + 364 + ], + "lines": [ + { + "bbox": [ + 316, + 296, + 554, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 296, + 554, + 364 + ], + "type": "table", + "html": "
MethodTraining TokensTraining CostStorage Param.Inference Cost (/1000 items)Acc.
Individual002.4B31.0s94.2
MTL304M12h304M31.0s93.5
FR-Merging00304M31.0s88.2
Twin-Merging0.9M147s352M32.2s92.7
FREE-Merging0.9M147s328M32.2s93.7
", + "image_path": "7245e16f0f9be52b8044c6425480daf09fab74831f625055b9ec66cfe8b9aa7c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 378, + 552, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 552, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 552, + 390 + ], + "type": "text", + "content": "construct a reasonable backbone and generalizable experts." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 398, + 450, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 398, + 450, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 450, + 411 + ], + "type": "text", + "content": "6.6. Transferability Analysis" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 416, + 554, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 416, + 554, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 416, + 554, + 488 + ], + "type": "text", + "content": "Our proposed FR-Merging can integrate with other methods, as it addresses frequency-domain task interference that current methods overlook. As shown in Table 8, combining various methods with FR-Merging leads to performance improvements. Therefore, FR-Merging can be integrated with more approaches to achieve greater benefits." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 496, + 497, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 497, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 497, + 510 + ], + "type": "text", + "content": "6.7. Speed and Storage Space Analysis" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 514, + 554, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 554, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 554, + 586 + ], + "type": "text", + "content": "This part analyzes training costs, inference speed, and storage as shown in Table 9. FR-Merging achieves strong performance with significantly reducing storage costs without extra training or inference costs, while FREE-Merging slightly increases overhead for better performance. Both of these two methods enhance deep learning deployment." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 597, + 388, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 388, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 388, + 609 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 617, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 714 + ], + "type": "text", + "content": "In this paper, we propose FR-Merging and FREE-Merging to efficiently build a multi-task model by merging existing ones. FR-Merging constructs a backbone without training by high-pass filtering to remove information that dominates task interference. Meanwhile, FREE-Merging employs lightweight experts to offset information loss during merging. Our method balances training, storage, inference costs, and performance, showing high practical value." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3870" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 149, + 115, + 162 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 149, + 115, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 149, + 115, + 162 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 169, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 169, + 296, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 169, + 296, + 213 + ], + "spans": [ + { + "bbox": [ + 61, + 169, + 296, + 213 + ], + "type": "text", + "content": "[1] Samuel Ainsworth, Jonathan Hayase, and Siddhartha Srinivasa. Git re-basin: Merging models modulo permutation symmetries. In The Eleventh International Conference on Learning Representations, 2023. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 214, + 296, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 214, + 296, + 257 + ], + "spans": [ + { + "bbox": [ + 61, + 214, + 296, + 257 + ], + "type": "text", + "content": "[2] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 258, + 296, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 258, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 61, + 258, + 296, + 300 + ], + "type": "text", + "content": "[3] Tim-Oliver Buchholz and Florian Jug. Fourier image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1846-1854, 2022. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 303, + 295, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 303, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 61, + 303, + 295, + 335 + ], + "type": "text", + "content": "[4] Shijie Chen, Yu Zhang, and Qiang Yang. Multi-task learning in natural language processing: An overview. ACM Computing Surveys, 56(12):1-32, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 337, + 296, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 337, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 61, + 337, + 296, + 392 + ], + "type": "text", + "content": "[5] Tianyi Chu, Jiafu Chen, Jiakai Sun, Shuobin Lian, Zhizhong Wang, Zhiwen Zuo, Lei Zhao, Wei Xing, and Dongming Lu. Rethinking fast fourier convolution in image inpainting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23195-23205, 2023. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 392, + 296, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 296, + 446 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 296, + 446 + ], + "type": "text", + "content": "[6] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instructionfinetuned language models. Journal of Machine Learning Research, 25(70):1-53, 2024. 3, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 447, + 296, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 296, + 491 + ], + "type": "text", + "content": "[7] MohammadReza Davari and Eugene Belilovsky. Model breadcrumbs: Scaling multi-task model merging with sparse masks. arXiv preprint arXiv:2312.06795, 2023. 3, 5, 6, 7, 8, 9, 10" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 492, + 296, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 492, + 296, + 535 + ], + "spans": [ + { + "bbox": [ + 62, + 492, + 296, + 535 + ], + "type": "text", + "content": "[8] Pala Tej Deep, Rishabh Bhardwaj, and Soujanya Poria. Della-merging: Reducing interference in model merging through magnitude-based sampling. arXiv preprint arXiv:2406.11617, 2024. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 536, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 536, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 62, + 536, + 296, + 624 + ], + "type": "text", + "content": "[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota, 2019. Association for Computational Linguistics. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 625, + 296, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 296, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 296, + 679 + ], + "type": "text", + "content": "[10] Ning Ding, Yujia Qin, Guang Yang, Fuchao Wei, Zonghan Yang, Yusheng Su, Shengding Hu, Yulin Chen, Chi-Min Chan, Weize Chen, et al. Parameter-efficient fine-tuning of large-scale pre-trained language models. Nature Machine Intelligence, 5(3):220-235, 2023. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "type": "text", + "content": "[11] Xibin Dong, Zhiwen Yu, Wenming Cao, Yifan Shi, and Qianli Ma. A survey on ensemble learning. Frontiers of Computer Science, 14:241-258, 2020. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 55, + 91, + 296, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 139 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 139 + ], + "type": "text", + "content": "This work is supported by National Natural Science Foundation of China (NSFC) (62232005, 62202126); the National Key Research and Development Program of China (2021YFB3300502)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "type": "text", + "content": "[12] Guodong Du, Junlin Lee, Jing Li, Runhua Jiang, Yifei Guo, Shuyang Yu, Hanting Liu, Sim Kuan Goh, Ho-Kin Tang, Daojing He, et al. Parameter competition balancing for model merging. arXiv preprint arXiv:2410.02396, 2024. 2, 3, 5, 6, 7, 8, 9, 10" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 130, + 554, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 554, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 554, + 185 + ], + "type": "text", + "content": "[13] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 186, + 554, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 554, + 220 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 554, + 220 + ], + "type": "text", + "content": "[14] Pierre Duhamel and Martin Vetterli. Fast fourier transforms: a tutorial review and a state of the art. Signal processing, 19 (4):259-299, 1990. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 222, + 555, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 222, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 316, + 222, + 555, + 266 + ], + "type": "text", + "content": "[15] Mudasir A Ganaie, Minghui Hu, Ashwani Kumar Malik, Muhammad Tanveer, and Ponnuthurai N Suganthan. Ensemble deep learning: A review. Engineering Applications of Artificial Intelligence, 115:105151, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 268, + 554, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 268, + 554, + 323 + ], + "spans": [ + { + "bbox": [ + 316, + 268, + 554, + 323 + ], + "type": "text", + "content": "[16] Antonio Andrea Gargiulo, Donato Crisostomi, Maria Sofia Bucarelli, Simone Scardapane, Fabrizio Silvestri, and Emanuele Rodola. Task singular vectors: Reducing task interference in model merging. arXiv preprint arXiv:2412.00081, 2024. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 325, + 554, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 325, + 554, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 325, + 554, + 369 + ], + "type": "text", + "content": "[17] Timur Garipov, Pavel Izmailov, Dmitrii Podoprikhin, Dmitry P Vetrov, and Andrew G Wilson. Loss surfaces, mode connectivity, and fast ensembling of dnns. Advances in neural information processing systems, 31, 2018. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 371, + 554, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 554, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 554, + 403 + ], + "type": "text", + "content": "[18] Rafael C Gonzales and Paul Wintz. Digital image processing. Addison-Wesley Longman Publishing Co., Inc., 1987. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 406, + 554, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 406, + 554, + 450 + ], + "spans": [ + { + "bbox": [ + 316, + 406, + 554, + 450 + ], + "type": "text", + "content": "[19] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 452, + 554, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 452, + 554, + 496 + ], + "spans": [ + { + "bbox": [ + 316, + 452, + 554, + 496 + ], + "type": "text", + "content": "[20] Chenyu Huang, Peng Ye, Tao Chen, Tong He, Xi-angyu Yue, and Wanli Ouyang. Emr-merging: Tuning-free high-performance model merging. arXiv preprint arXiv:2405.17461, 2024. 1, 2, 4, 5, 6, 7, 3, 8, 9, 10" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 498, + 554, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 554, + 543 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 554, + 543 + ], + "type": "text", + "content": "[21] Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Suchin Gururangan, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. Editing models with task arithmetic. arXiv preprint arXiv:2212.04089, 2022. 1, 2, 3, 5, 6, 7, 4, 8, 9, 10" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 544, + 554, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 554, + 599 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 554, + 599 + ], + "type": "text", + "content": "[22] Dongfu Jiang, Xiang Ren, and Bill Yuchen Lin. Lmblender: Ensembling large language models with pairwise ranking and generative fusion. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 14165-14178, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 601, + 554, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 554, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 554, + 645 + ], + "type": "text", + "content": "[23] Xisen Jin, Xiang Ren, Daniel Preotiac-Pietro, and Pengxiang Cheng. Dataless knowledge fusion by merging weights of language models. arXiv preprint arXiv:2212.09849, 2022. 1, 2, 5, 6, 3, 8, 9, 10" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 647, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 554, + 712 + ], + "type": "text", + "content": "[24] Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan Firat, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. Gshard: Scaling giant models with conditional computation and automatic sharding. In International Conference on Learning Representations, 2021. 5" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3871" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "type": "text", + "content": "[25] Hao Li, Zheng Xu, Gavin Taylor, Christoph Studer, and Tom Goldstein. Visualizing the loss landscape of neural nets. Advances in neural information processing systems, 31, 2018. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 117, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 117, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 117, + 294, + 172 + ], + "type": "text", + "content": "[26] Haokun Liu, Derek Tam, Mohammed Muqeeth, Jay Mohta, Tenghao Huang, Mohit Bansal, and Colin A Raffel. Few-shot parameter-efficient fine-tuning is better and cheaper than in-context learning. Advances in Neural Information Processing Systems, 35:1950-1965, 2022. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 173, + 294, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 294, + 216 + ], + "type": "text", + "content": "[27] James Liu, Guangxuan Xiao, Kai Li, Jason D Lee, Song Han, Tri Dao, and Tianle Cai. Bitdelta: Your fine-tune may only be worth one bit. Advances in Neural Information Processing Systems, 37:13579-13600, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 281 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 281 + ], + "type": "text", + "content": "[28] Yajing Liu, Yuning Lu, Hao Liu, Yaozu An, Zhuoran Xu, Zhuokun Yao, Baofeng Zhang, Zhiwei Xiong, and Chenguang Gui. Hierarchical prompt learning for multi-task learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10888-10898, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 283, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 294, + 326 + ], + "type": "text", + "content": "[29] Zhenyi Lu, Chenghao Fan, Wei Wei, Xiaoye Qu, Dangyang Chen, and Yu Cheng. Twin-merging: Dynamic integration of modular expertise in model merging. arXiv preprint arXiv:2406.15479, 2024. 1, 2, 5, 6, 7, 8, 3, 9, 10" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 392 + ], + "type": "text", + "content": "[30] Haipeng Luo, Qingfeng Sun, Can Xu, Pu Zhao, Jianguang Lou, Chongyang Tao, Xiubo Geng, Qingwei Lin, Shifeng Chen, and Dongmei Zhang. Wizardmath: Empowering mathematical reasoning for large language models via reinforced evol-instruct. arXiv preprint arXiv:2308.09583, 2023. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 393, + 294, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 294, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 294, + 448 + ], + "type": "text", + "content": "[31] Ziyang Luo, Can Xu, Pu Zhao, Qingfeng Sun, Xiubo Geng, Wenxiang Hu, Chongyang Tao, Jing Ma, Qingwei Lin, and Daxin Jiang. Wizardcoder: Empowering code large language models with evol-instruct. In The Twelfth International Conference on Learning Representations, 2024. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 449, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 294, + 491 + ], + "type": "text", + "content": "[32] Xinyin Ma, Gongfan Fang, and Xinchao Wang. Llm-pruner: On the structural pruning of large language models. Advances in neural information processing systems, 36:21702-21720, 2023. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 492, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 294, + 525 + ], + "type": "text", + "content": "[33] Michael S Matena and Colin A Raffel. Merging models with fisher-weighted averaging. Advances in Neural Information Processing Systems, 35:17703-17716, 2022. 1, 2, 5, 9" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 526, + 294, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 294, + 558 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 294, + 558 + ], + "type": "text", + "content": "[34] Behnam Neyshabur, Hanie Sedghi, and Chiyuan Zhang. What is being transferred in transfer learning? Advances in neural information processing systems, 33:512-523, 2020. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 559, + 294, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 603 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 603 + ], + "type": "text", + "content": "[35] Tan Nguyen, Minh Pham, Tam Nguyen, Khai Nguyen, Stanley Osher, and Nhat Ho. Fourierformer: Transformer meets generalized fourier integral theorem. Advances in Neural Information Processing Systems, 35:29319-29335, 2022. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 604, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 604, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 604, + 294, + 646 + ], + "type": "text", + "content": "[36] Biqing Qi, Fangyuan Li, Zhen Wang, Junqi Gao, Dong Li, Peng Ye, and Bowen Zhou. Less is more: Efficient model merging with binary task switch. arXiv preprint arXiv:2412.00054, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "text", + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 3, 4, 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 715 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "text", + "content": "[38] Victor Sanh, Albert Webson, Colin Raffel, Stephen Bach, Lintang Sutawika, Zaid Alyafeai, Antoine Chaffin, Arnaud Stiegler, Arun Raja, Manan Dey, et al. Multitask prompted training enables zero-shot task generalization. In International Conference on Learning Representations, 2022. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 184 + ], + "type": "text", + "content": "[39] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. In International Conference on Learning Representations, 2016. 2, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 185, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 553, + 240 + ], + "type": "text", + "content": "[40] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 241, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 297 + ], + "type": "text", + "content": "[41] Simon Vandenhende, Stamatios Georgoulis, Wouter Van Gansbeke, Marc Proesmans, Dengxin Dai, and Luc Van Gool. Multi-task learning for dense prediction tasks: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(7):3614-3633, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 297, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 374 + ], + "type": "text", + "content": "[42] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, et al. Image as a foreign language: Beit pretraining for vision and vision-language tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19175-19186, 2023. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 376, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 408 + ], + "type": "text", + "content": "[43] T Wolf. Huggingface's transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771, 2019. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 410, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 410, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 410, + 553, + 487 + ], + "type": "text", + "content": "[44] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In International conference on machine learning, pages 23965-23998. PMLR, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 488, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 488, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 488, + 553, + 544 + ], + "type": "text", + "content": "[45] Chengyue Wu, Teng Wang, Yixiao Ge, Zeyu Lu, Ruisong Zhou, Ying Shan, and Ping Luo. " + }, + { + "bbox": [ + 316, + 488, + 553, + 544 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 316, + 488, + 553, + 544 + ], + "type": "text", + "content": "-tuning: Transferring multimodal foundation models with optimal multi-task interpolation. In International Conference on Machine Learning, pages 37713-37727. PMLR, 2023. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 544, + 553, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 553, + 609 + ], + "type": "text", + "content": "[46] Can Xu, Qingfeng Sun, Kai Zheng, Xiubo Geng, Pu Zhao, Jiazhan Feng, Chongyang Tao, Qingwei Lin, and Daxin Jiang. Wizardlm: Empowering large pre-trained language models to follow complex instructions. In The Twelfth International Conference on Learning Representations, 2024. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 611, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 611, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 611, + 553, + 644 + ], + "type": "text", + "content": "[47] Jing Xu and Jingzhao Zhang. Random masking finds winning tickets for parameter efficient fine-tuning. In *Forty-first International Conference on Machine Learning*. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 646, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 646, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 646, + 553, + 689 + ], + "type": "text", + "content": "[48] Prateek Yadav, Derek Tam, Leshem Choshen, Colin A Raffel, and Mohit Bansal. Ties-merging: Resolving interference when merging models. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3, 4, 5, 6, 7, 8, 9, 10" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 691, + 553, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 715 + ], + "type": "text", + "content": "[49] Enneng Yang, Li Shen, Guibing Guo, Xingwei Wang, Xiaochun Cao, Jie Zhang, and Dacheng Tao. Model merging" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3872" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 631 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "type": "text", + "content": "in llms, mllms, and beyond: Methods, theories, applications and opportunities. arXiv preprint arXiv:2408.07666, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "text", + "content": "[50] Enneng Yang, Li Shen, Zhenyi Wang, Guibing Guo, Xiaojun Chen, Xingwei Wang, and Dacheng Tao. Representation surgery for multi-task model merging. arXiv preprint arXiv:2402.02705, 2024. 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "type": "text", + "content": "[51] Enneng Yang, Li Shen, Zhenyi Wang, Guibing Guo, Xiao-jun Chen, Xingwei Wang, and Dacheng Tao. Representation surgery for multi-task model merging. In *Forty-first International Conference on Machine Learning*, 2024. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "type": "text", + "content": "[52] Enneng Yang, Zhenyi Wang, Li Shen, Shiwei Liu, Guibing Guo, Xingwei Wang, and Dacheng Tao. Adamerging: Adaptive model merging for multi-task learning. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 5, 6, 3, 9" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 294, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 306 + ], + "type": "text", + "content": "[53] Le Yu, Bowen Yu, Haiyang Yu, Fei Huang, and Yongbin Li. Language models are super mario: Absorbing abilities from homologous models as a free lunch. In *Forty-first International Conference on Machine Learning*, 2024. 2, 3, 4, 6, 8, 10" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 308, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 294, + 374 + ], + "type": "text", + "content": "[54] Kerem Zaman, Leshem Choshen, and Shashank Srivastava. Fuse to forget: Bias reduction and selective memorization through model fusion. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 18763-18783, Miami, Florida, USA, 2024. Association for Computational Linguistics. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 294, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 418 + ], + "type": "text", + "content": "[55] Jinghan Zhang, Junteng Liu, Junxian He, et al. Composing parameter-efficient modules with arithmetic operation. Advances in Neural Information Processing Systems, 36: 12589-12610, 2023. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 420, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 420, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 420, + 294, + 453 + ], + "type": "text", + "content": "[56] Yu Zhang and Qiang Yang. A survey on multi-task learning. IEEE transactions on knowledge and data engineering, 34 (12):5586-5609, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 454, + 294, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 454, + 294, + 508 + ], + "spans": [ + { + "bbox": [ + 56, + 454, + 294, + 508 + ], + "type": "text", + "content": "[57] Yanqi Zhou, Tao Lei, Hanxiao Liu, Nan Du, Yanping Huang, Vincent Zhao, Andrew M Dai, Quoc V Le, James Laudon, et al. Mixture-of-experts with expert choice routing. Advances in Neural Information Processing Systems, 35:7103-7114, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 510, + 294, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 510, + 294, + 564 + ], + "spans": [ + { + "bbox": [ + 56, + 510, + 294, + 564 + ], + "type": "text", + "content": "[58] Yuyan Zhou, Liang Song, Bingning Wang, and Weipeng Chen. Metagpt: Merging large language models using model exclusive task arithmetic. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 1711-1724, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 566, + 294, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 566, + 294, + 631 + ], + "spans": [ + { + "bbox": [ + 56, + 566, + 294, + 631 + ], + "type": "text", + "content": "[59] Liu Zhuang, Lin Wayne, Shi Ya, and Zhao Jun. A robustly optimized BERT pre-training approach with post-training. In Proceedings of the 20th Chinese National Conference on Computational Linguistics, pages 1218-1227, Huhhot, China, 2021. Chinese Information Processing Society of China. 6" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3873" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/FRET_ Feature Redundancy Elimination for Test Time Adaptation/013e8d33-0504-4dd6-a3e9-326514c12e30_content_list.json b/2025/FRET_ Feature Redundancy Elimination for Test Time Adaptation/013e8d33-0504-4dd6-a3e9-326514c12e30_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..523603bc83f18422f464b0e8d661a309bf5e9ed8 --- /dev/null +++ b/2025/FRET_ Feature Redundancy Elimination for Test Time Adaptation/013e8d33-0504-4dd6-a3e9-326514c12e30_content_list.json @@ -0,0 +1,1848 @@ +[ + { + "type": "text", + "text": "FRET: Feature Redundancy Elimination for Test Time Adaptation", + "text_level": 1, + "bbox": [ + 161, + 130, + 834, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Linjing You $^{1,*}$ Jiabao Lu $^{1,*}$ Xiayuan Huang $^{2,\\dagger}$ Xiangli Nie $^{1,\\dagger}$", + "bbox": [ + 228, + 178, + 763, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Institute of Automation, Chinese Academy of Sciences, Beijing, China", + "bbox": [ + 217, + 198, + 781, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ College of Science, Beijing Forestry University, Beijing, China", + "bbox": [ + 243, + 215, + 750, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{youlinjing2023,lujiabao2025,xiangliNie}@ia.ac.cn", + "bbox": [ + 194, + 234, + 803, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "huangxiayuan@bjfu.edu.cn", + "bbox": [ + 356, + 253, + 640, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "* Equal contribution, †Corresponding author", + "bbox": [ + 377, + 268, + 614, + 280 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 321, + 326, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Test-Time Adaptation (TTA) aims to enhance the generalization of deep learning models when faced with test data that exhibits distribution shifts from the training data. In this context, only a pre-trained model and unlabeled test data are available, making it particularly relevant for privacy-sensitive applications. In practice, we observe that feature redundancy in embeddings tends to increase as domain shifts intensify in TTA. However, existing TTA methods often overlook this redundancy, which can hinder the model's adaptability to new data. To address this issue, we introduce Feature Redundancy Elimination for Test-time Adaptation (FRET), a novel perspective for TTA. A straightforward approach (S-FRET) is to directly minimize the feature redundancy score as an optimization objective to improve adaptation. Despite its simplicity and effectiveness, S-FRET struggles with label shifts, limiting its robustness in real-world scenarios. To mitigate this limitation, we further propose Graph-based FRET (G-FRET), which integrates a Graph Convolutional Network (GCN) with contrastive learning. This design not only reduces feature redundancy but also enhances feature discriminability in both the representation and prediction layers. Extensive experiments across multiple model architectures, tasks, and datasets demonstrate the effectiveness of S-FRET and show that G-FRET achieves state-of-the-art performance. Further analysis reveals that G-FRET enables the model to extract non-redundant and highly discriminative features during inference, thereby facilitating more robust test-time adaptation. The code is available at https://github.com/youlj109/fret.", + "bbox": [ + 88, + 356, + 485, + 810 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 843, + 220, + 858 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, the emergence of deep neural networks (DNNs) [12, 18] has significantly propelled the advancement in a", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/13e6146631ac6f03fd39d1fa7363a0ba87ab939cb5b67b48fca8e42a497f032f.jpg", + "image_caption": [ + "Figure 1. An intuitive demonstration of the relationship between embedded feature redundancy and distribution migration. Top: Second-order feature relation graph of embedding features. Bottom: Plot of feature redundancy versus the level of corruption." + ], + "image_footnote": [], + "bbox": [ + 535, + 329, + 877, + 573 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vast range of human tasks [1, 15, 38]. However, the basic assumption that training and test data are drawn independently and identically (i.i.d.) may not be suitable in many real-world scenarios where the test/target distribution diverges from the training/source distribution [50]. This phenomenon, known as distribution shift, often leads to poor performance [8]. To tackle this issue, a variety of generalization or adaptation techniques have been developed to improve model robustness against distribution shifts [24, 54, 60]. Among these, Test Time Adaptation (TTA) [10, 40, 41] has been the focus of recent and active research. TTA has the advantage of only requiring access to the pre-trained model from the training/source domain and unlabeled test data. Therefore, TTA is a more secure and practical solution especially when models are publicly available.", + "bbox": [ + 509, + 672, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2120", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "available but the training data and training process remain inaccessible due to privacy and resource restrictions [24].", + "bbox": [ + 89, + 90, + 480, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing TTA methods can be broadly categorized into batch normalization calibration methods, pseudo-labeling methods, consistency training methods, and clustering-based training methods, based on different adaptation strategies [24]. Batch normalization (BN) calibration methods [53] posit that the statistics in the BN layers represent domain-specific knowledge. To bridge the domain gap, it is suggested to replace the training BN statistics with new statistics estimated over the entire target domain. Pseudolabeling methods [59] employ various filtering strategies, such as thresholding or entropy-based approaches, to obtain reliable pseudo-labels. This process helps minimize discrepancies between predicted labels and pseudo-labels. Consistency training methods [39] aim to improve the stability of network predictions or features by addressing variations in input data, such as noise or perturbations, and changes in model parameters. Clustering-based training methods [19] focus on reducing uncertainty in target network predictions by leveraging clustering techniques to organize target features into distinct groups, thereby enhancing the robustness of the model.", + "bbox": [ + 91, + 125, + 483, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite the empirical success of existing methods, they overlook feature redundancy in embeddings, which hinders generalization to new data [22]. As shown in Fig. 1, we visualize feature redundancy in ResNet-18 [12] using second-order feature relations (i.e., covariance matrix) on the CIFAR10-C dataset [13] under distribution shift [25]. Each image is embedded into a 512-dimensional vector; stacking $n$ embeddings yields an $n \\times 512$ matrix. We compute the absolute covariance matrix $(512 \\times 512)$ and visualize it as a heatmap — redder areas indicate higher feature correlation, hence greater redundancy. The top two subfigures in Fig. 1 compare the source and target domains, revealing much higher redundancy in the latter due to distribution shift. To quantify this, we remove the diagonal of the covariance matrix and sum the absolute values of the remaining elements (see Eq. (1)). As shown at the bottom of Fig. 1, redundancy increases with corruption severity, suggesting that features with lower redundancy may be better suited for adaptation to the target domain.", + "bbox": [ + 91, + 444, + 483, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building on the limitations identified in the existing literature and the observations above, we provide a novel perspective for test-time adaptation—Feature Redundancy Elimination for TTA (FRET)—which differs from existing methods. Due to the inherently differentiable nature of the redundancy metric, a straightforward and computationally efficient approach is to directly minimize the feature redundancy score. We refer to this simple approach as S-FRET. However, since S-FRET lacks the ability to perceive label distributions, it effectively addresses covariate shift but fails to handle label shift—a scenario where the tar", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "get domain exhibits a label distribution that deviates from the source domain (e.g., long-tailed data). This discrepancy can significantly degrade model performance and remains a critical challenge [3, 16, 32].", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome S-FRET's limitation in handling label shift, we further propose G-FRET, which integrates a Graph Convolutional Network (GCN). Unlike FRET, which only minimizes redundancy at the representation layer, G-FRET leverages GCN to incorporate both attention and redundant feature relationships into the representation and prediction layers. Specifically, G-FRET learns non-redundant and discriminative representation through contrastive distillation between attention and redundant representation in the representation-layer. In the prediction layer, attention predictions are optimized through entropy minimization, while ensuring that they are distinct from redundant predictions by negative learning. By maximizing the discriminability of attention relationships while eliminating redundancy in both the representation and prediction layers, G-FRET effectively reduces test-time feature redundancy and improves discrimination when handling complex distribution-shifted test data.", + "bbox": [ + 511, + 152, + 906, + 424 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We summarize the contributions of this paper as follows.", + "bbox": [ + 532, + 426, + 903, + 441 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Potential Way: provide a novel perspective for test-time adaptation — Feature Redundancy Elimination for TTA (FRET) — which differs from existing methods. It adapts the model to distribution shifts by reducing the redundancy of embedded features during testing. To the best of our knowledge, this is the first attempt to leverage feature redundancy for TTA.", + "- Innovative Methods: We propose two novel methods, S-FRET and G-FRET, to eliminate feature redundancy in test-time embeddings. S-FRET directly minimizes the feature redundancy score, offering a lightweight solution. G-FRET extends this by decomposing feature relations into attention and redundancy components, enabling joint optimization of redundancy elimination and class discriminability in real-world scenarios.", + "- Extensive Evaluation: Comprehensive experiments across multiple model architectures, tasks, and datasets demonstrate the effectiveness of S-FRET and show that G-FRET achieves state-of-the-art performance. Further in-depth analysis reveals that G-FRET facilitates the extraction of non-redundant and discriminative embedding features during testing, ultimately enhancing model adaptation to the test-time target domain." + ], + "bbox": [ + 513, + 446, + 903, + 794 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Preliminary and Notations", + "text_level": 1, + "bbox": [ + 511, + 813, + 761, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We briefly revisit Test-Time Adaptation and Feature Redundancy Elimination in this section for the convenience of introducing methods, and put detailed related work discussions into Sec. 7 due to page limits.", + "bbox": [ + 511, + 839, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2121", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b212cf9628f98fd8cde427cb2b685556eff20199b436d43bf8a4dfd8a3f9de70.jpg", + "image_caption": [ + "Figure 2. Redundancy scores of 15 corruption types in CIFAR-10/100-C and ImageNet-C across 5 corruption levels, as well as for source and target domains in PACS, OfficeHome, VLCS, and DomainNet." + ], + "image_footnote": [], + "bbox": [ + 133, + 88, + 318, + 213 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6dd2ddf82f35c52220a5a119a5822625c5201de76d2762f2cd4671979169c65b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 318, + 88, + 501, + 213 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1c64f3a491115c11a74b19ffadce1749266f4663facd651a460844a6558aff84.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 88, + 684, + 213 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5feb3184358798be6c31190903e333107c91a28c2b51d1ab20fa19683c01caf8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 683, + 88, + 864, + 213 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Test-Time Adaptation", + "text_level": 1, + "bbox": [ + 89, + 258, + 295, + 275 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the test-time adaptation (TTA) scenario, we have access only to unlabeled data from the target domain in an online manner and a pre-trained model from the source domain. Specifically, let $\\{(x_i,y_i)\\}_{i = 1}^{n_s}\\subset \\mathcal{X}_s\\times \\mathcal{Y}_s$ represent the labeled source domain dataset with labels $\\mathcal{V}_s$ and $\\{x_j\\}_{j = 1}^{n_t}\\subset \\mathcal{X}_t$ represent a batch of unlabeled target data. The model, trained on the source domain dataset and parameterized by $\\theta$ , is denoted as $g_{\\theta} = h(f(\\cdot)): \\mathcal{X}_s\\to \\mathcal{Y}_s$ , where $f$ is the backbone encoder and $h$ denotes the linear classification head. Our objective is to adapt this model to the target domain $\\mathcal{X}_t$ , which exhibits a different distribution from that of the source domain $\\mathcal{X}_s$ . During testing, for each instance $x_{j}\\in \\mathcal{X}_{t}$ , let the output of $f$ and $h$ be denoted as embedding $z_{j} = f(x_{j})\\in \\mathbb{R}^{d}$ and logs $p_j = h(z_j)\\in \\mathbb{R}^C$ , respectively, where $d$ is the dimension of the embeddings and $C$ is the number of classes.", + "bbox": [ + 89, + 282, + 483, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Feature Redundancy Elimination", + "text_level": 1, + "bbox": [ + 89, + 536, + 382, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Feature redundancy is a key concern in both feature extraction and feature selection [26, 43]. Following previous work [58], we remove the diagonal elements of the covariance matrix and calculate the sum of the absolute values of the remaining elements to derive the redundancy scores $R_{s}$ as:", + "bbox": [ + 89, + 559, + 483, + 633 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nR _ {s} = \\left\\| \\tilde {Z} ^ {T} \\tilde {Z} - I _ {d} \\right\\| _ {1} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 646, + 480, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $I_{d}$ is the identity matrix, $\\| \\cdot \\| _1$ denotes the 1-norm of a matrix and $\\tilde{Z}_{:,j} = \\frac{Z_{:,j}}{\\|Z_{:,j}\\|_2}$ denotes the matrix obtained by normalizing each column.", + "bbox": [ + 89, + 684, + 483, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Feature Redundancy Under Domain Shift", + "text_level": 1, + "bbox": [ + 89, + 750, + 465, + 766 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We conduct a systematic investigation into the behavior of feature redundancy under domain shift.", + "bbox": [ + 89, + 775, + 483, + 805 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Empirical Study.", + "text_level": 1, + "bbox": [ + 89, + 816, + 256, + 834 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 2, across approximately 50 sub-datasets, the feature redundancy score $R_{s}$ consistently increases with the severity of domain shift. (Note: $R_{s}$ is computed using models with BN parameters updated from test data", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "rather than those fixed during training [36].) This empirical trend suggests that domain shift induces stronger correlations among features. To the best of our knowledge, this phenomenon has not been systematically explored in prior work.", + "bbox": [ + 511, + 260, + 905, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Theoretical Explanation.", + "text_level": 1, + "bbox": [ + 511, + 344, + 740, + 361 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Intuitively, higher covariance elements indicate that the features fail to span the full vector space. Furthermore, we provide the following theoretical result. Missing proofs and detailed explanations are provided in Sec. 9", + "bbox": [ + 511, + 366, + 905, + 426 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Theorem 1 (Redundancy Elimination Improves Generalization). Under the empirical observation that embeddings from the source domain exhibit lower feature redundancy, i.e., $\\mathcal{R}_s(Z_t) > \\mathcal{R}_s(Z_s)$ , and by leveraging Theorem 2 from [55], we can derive the following generalization error bound: Let $\\mathcal{H}$ be a hypothesis class of VC-dimension $d_v$ . If $\\hat{h} \\in \\mathcal{H}$ minimizes the empirical error $\\hat{\\epsilon}_s(h)$ on $\\mathcal{X}_s$ and $h_t^* = \\arg \\min_{h \\in \\mathcal{H}} \\epsilon_t(h)$ is the optimal hypothesis on $\\mathcal{X}_t$ , with the assumption that all hypotheses are L-Lipschitz continuous, then $\\forall \\delta \\in (0,1)$ , with probability with at least $1 - \\delta$ the following inequality holds:", + "bbox": [ + 511, + 435, + 906, + 602 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\epsilon_ {t} (\\hat {h}) - \\epsilon_ {t} (h _ {t} ^ {*}) \\\\ \\leq \\mathcal {O} (\\sqrt {\\| \\mu_ {s} - \\mu_ {t} \\| _ {F} ^ {2} + R _ {s} (Z _ {s}) ^ {2} + R _ {s} (Z _ {t}) ^ {2}}) + C \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 609, + 877, + 656 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $C = 2\\sqrt{\\frac{d_v\\log(2n_s) - \\log(\\delta)}{2n_s}} + 2\\gamma$ and $\\gamma = \\min_{h\\in \\mathcal{H}}\\{\\epsilon_s(h(t)) + \\epsilon_t(h(t))\\}$ . $\\mu_{s}$ and $\\mu_t$ denote the means of the source and test embeddings, respectively. We use $\\mathcal{O}(\\cdot)$ to hide the constant dependence. Since $\\mathcal{R}_s(Z_t) > \\mathcal{R}_s(Z_s)$ , this implies that reducing $\\mathcal{R}_s(Z_t)$ can effectively decrease the generalization error.", + "bbox": [ + 511, + 667, + 905, + 765 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. The Proposed Approach", + "text_level": 1, + "bbox": [ + 511, + 777, + 741, + 795 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. S-FRET", + "text_level": 1, + "bbox": [ + 511, + 801, + 614, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To eliminate feature redundancy in the target domain, a simple yet effective approach is to directly minimize the feature redundancy score $R_{s}$ , as defined in Eq. (1), leading to the proposed method S-FRET (Simple and Straightforward FRET). The loss function for S-FRET can be written as:", + "bbox": [ + 511, + 825, + 905, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2122", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b7081ca10dda8bd484b3ce6ea5bcf124602d584b14194d34e556de7e34681a36.jpg", + "image_caption": [ + "Figure 3. The pipeline of our proposed Graph Convolutional Network based Feature Redundancy Elimination for Test Time Adaptation (G-FRET) method. For each test sample, G-FRET learns non-redundant and discriminative representation through contrastive distillation in the representation-layer. Additionally, in the prediction layer, attention predictions are optimized through entropy minimization, while ensuring that they are distinct from redundant predictions by using negative learning." + ], + "image_footnote": [], + "bbox": [ + 133, + 87, + 867, + 303 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S F R E T} = \\left\\| \\tilde {Z} ^ {T} \\tilde {Z} - I _ {d} \\right\\| _ {1} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 388, + 482, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This method directly minimizes the redundancy score derived from feature correlation, providing a computationally efficient solution that avoids complex architectural modifications. By focusing on the feature space of representation-layer, S-FRET enables the model to extract non-redundant representations, thereby enhancing its adaptability to the target domain (see Fig. 4).", + "bbox": [ + 89, + 422, + 483, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, since the redundancy score does not incorporate information about class distributions, S-FRET primarily addresses covariate shift but lacks the discriminative capability to handle label shift—a critical challenge where the target domain exhibits a label distribution that deviates significantly from the source domain (e.g., long-tailed data), leading to substantial performance degradation (see Tab. 4) [3, 16, 32].", + "bbox": [ + 89, + 529, + 483, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. G-FRET", + "text_level": 1, + "bbox": [ + 89, + 660, + 194, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2.1. Overview of G-FRET", + "text_level": 1, + "bbox": [ + 89, + 683, + 287, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To eliminate feature redundancy and address label shift in the target domain, we further propose a novel two-layer TTA method, namely Graph Convolutional Network based Feature Redundancy Elimination for Test Time Adaptation (G-FRET), which leverages GCN to simultaneously model feature relationships and sample-specific discriminative information in both representation and prediction layers. Fig. 3 provides an overview of our G-FRET framework.", + "bbox": [ + 89, + 703, + 482, + 823 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The key idea of G-FRET is to adapt the pre-tarined model during testing by extracting discriminative and nonredundant embedding features from both the representation and prediction layers. Firstly (Sec. 4.2.2), we apply a mask matrix to the second-order feature relation graph.", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This process yields two separate graphs: an attention graph, which highlights significant feature relationships, and a redundant graph, which identifies redundant feature relationships. Next, we derive the attention/redundancy representation and attention/redundancy prediction by utilizing a single-layer GCN to integrate the model's embeddings with the attention and redundant graphs. This process allows us to explicitly separate attention and redundancy relations in both representation and prediction layer. Then (Sec. 4.2.3), in the representation layer, contrastive learning techniques are employed to enhance the discriminability of the attention representation while ensuring it remains distant from the redundant representation. At last (Sec. 4.2.4), in the prediction layer, attention predictions are optimized through entropy minimization, while ensuring that they are distinct from redundant predictions by negative learning. Through this two-layer process, we achieve discriminative and nonredundant embedding features that effectively adapt to the target domain.", + "bbox": [ + 511, + 376, + 906, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2.2. Attention Part and Redundancy Part", + "text_level": 1, + "bbox": [ + 511, + 676, + 816, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a batch of $n_t$ data instances from the target domain, we obtain the first-order $n_t \\times d$ embedding feature matrix $Z \\in \\mathbb{R}^{n_t \\times d}$ and the second-order feature relation graph $G_F = Z^T Z \\in \\mathbb{R}^{d \\times d}$ . Following the state-of-the-art feature selection method SOFT [58] (see Appendix 7.3 for more details), we apply a symmetric mask matrix $M_M \\in \\mathbb{R}^{d \\times d}$ to $G_F$ . This decomposition of $G_F$ yields two distinct graphs: the attention graph $G_A$ and the redundancy graph $G_R$ . $G_A$ captures the attention relations (i.e., important relation) among features, while $G_R$ represents the redundant relation between them. Formally, they are calculated as follows:", + "bbox": [ + 511, + 698, + 906, + 878 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nG _ {A} = G _ {F} \\odot M _ {M}; G _ {R} = G _ {F} - G _ {A} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 885, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2123", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\odot$ is the element-wise product. However, differently from SOFT, which focuses on feature selection by deriving an optimal mask matrix $M_M$ , our objective is to extract non-redundant features $Z$ . For ideally non-redundant features, the second-order feature relation graph should be a diagonal matrix, indicating that each feature is independent of all others. Therefore, we set $M_M = I_d$ to ensure that $G_A$ captures only the essential feature relations, while $G_R$ includes all remaining relations as redundant relations. Here, $I_d \\in \\mathbb{R}^{d \\times d}$ is the identity matrix. In addition, in some scenarios with prior knowledge, such as knowing which features may be more important, or needing to maintain special structural relations between features, some adaptations can be made by simply modifying the mask matrix $M_M$ .", + "bbox": [ + 89, + 90, + 483, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, in order to obtain attention/redundancy representation and attention/redundancy prediction based on the attention and redundancy graphs, we combine the normalized graph matrix with the classifier head $h$ to form a one-layer Graph Convolutional Network (GCN [17]), which can help generate attention representations $R_{A}$ and attention predictions $P_{A}$ that contain both data information and feature relationship information:", + "bbox": [ + 89, + 303, + 483, + 424 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR _ {A} = Z D _ {A} ^ {- 1 / 2} G _ {A} D _ {A} ^ {- 1 / 2}; P _ {A} = R _ {A} \\theta^ {h} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 438, + 482, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $D_A^{-1/2}G_A D_A^{-1/2}$ is the normalized graph matrix of graph $G_A$ , $D_A$ is the degree matrix of the graph, and $\\theta^h$ are the parameters of the classifier head $h$ . Similarly, we can obtain redundancy representations $R_R$ and predictions $P_R$ from redundancy graph $G_R$ :", + "bbox": [ + 89, + 470, + 483, + 551 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR _ {R} = Z D _ {R} ^ {- 1 / 2} G _ {R} D _ {R} ^ {- 1 / 2}; P _ {R} = R _ {R} \\theta^ {h} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 563, + 482, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Therefore, we decompose the model's original presentation and prediction into two components: the attention part and the redundancy part. This decomposition enables a distinct focus on important relationships while identifying redundant ones. Specifically, $R_{A}$ and $P_{A}$ focus solely on the feature relationships within the attention part, while $R_{R}$ and $P_{R}$ are concerned exclusively with the feature relationships within the redundancy part. Based on this decomposition, our goal is to ensure that the attention part adapts effectively to the target domain. Simultaneously, we aim to separate redundancy part from the attention component in both the representation and prediction layers.", + "bbox": [ + 89, + 597, + 483, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.3. Representation-Layer Redundancy Elimination", + "text_level": 1, + "bbox": [ + 89, + 790, + 467, + 805 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the above representations $R_{A}$ and $R_{R}$ , we expect that the attention representations possess class discriminability to handle label shift, while remaining distinct from redundancy representations to eliminate feature redundancy. To achieve this, we leverage contrastive learning [23, 30] as follows:", + "bbox": [ + 89, + 809, + 483, + 898 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {R} = \\\\ - \\sum_ {i = 1} ^ {n _ {t}} \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(R _ {A _ {i}} , c _ {o}\\right)\\right)}{\\sum_ {j = 1} ^ {C} \\exp \\left(\\operatorname {s i m} \\left(R _ {A _ {i}} , c _ {j}\\right)\\right) + \\exp \\left(\\operatorname {s i m} \\left(R _ {A _ {i}} , R _ {R _ {i}}\\right)\\right)} \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 114, + 913, + 186 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{sim}(\\cdot ,\\cdot)$ denotes cosine similarity, $\\{c_j\\}_{j = 1}^C$ denote the class centers in the attention representation space, $C$ is the number of clusters, and $c_{o}$ represents the center of the cluster that contains $R_{A_i}$ . $c_{j}$ is calculated as follows:", + "bbox": [ + 511, + 196, + 906, + 258 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nc _ {j} = \\frac {\\sum_ {i} z _ {i} \\cdot \\mathbf {1} [ \\operatorname {a r g m a x} \\left(p _ {i}\\right) = j ]}{\\sum_ {i} \\mathbf {1} [ \\operatorname {a r g m a x} \\left(p _ {i}\\right) = j ]} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 263, + 906, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{1}(\\cdot)$ is an indicator function, outputting value 1 if the argument is true or 0 otherwise.", + "bbox": [ + 511, + 303, + 903, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By minimizing $\\mathcal{L}_R$ , the model simultaneously enhances the discriminability of attention representations while ensuring their distinction from redundancy representations. This process enables the model to extract robust, discriminative, and non-redundant embedding features in the representation layer, facilitating effective adaptation to the target domain.", + "bbox": [ + 511, + 333, + 905, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.4. Prediction-Layer Redundancy Elimination", + "text_level": 1, + "bbox": [ + 511, + 445, + 856, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the prediction layer, to enhance the discriminability of attention predictions while reducing the influence of redundant predictions, we combine the principles of entropy minimization and negative learning into a unified loss function:", + "bbox": [ + 511, + 463, + 906, + 523 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {P} = - \\sum_ {i = 1} ^ {N} \\sigma \\left(P _ {A _ {i}}\\right) \\log \\sigma \\left(P _ {A _ {i}}\\right) - \\sum_ {i = 1} ^ {N} \\sigma \\left(P _ {R _ {i}}\\right) \\log \\sigma \\left(1 - P _ {A _ {i}}\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 530, + 910, + 584 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\sigma$ denotes the softmax operation. Here, the first term minimizes the entropy of attention predictions $P_A$ , encouraging sharper and more confident predictions, while the second term applies negative learning to redundant predictions $P_R$ , penalizing them to ensure they remain distinct from the attention predictions.", + "bbox": [ + 511, + 584, + 905, + 675 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.5. Training Loss Function", + "text_level": 1, + "bbox": [ + 511, + 681, + 722, + 695 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By combining Eq. (6) and Eq. (8), our overall loss function for G-FRET can be written as:", + "bbox": [ + 511, + 699, + 905, + 729 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {G F R E T} = \\mathcal {L} _ {R} + \\lambda \\mathcal {L} _ {P} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 744, + 903, + 760 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ is the balancing hyperparameter that controls the trade-off between $\\mathcal{L}_R$ and $\\mathcal{L}_P$ .", + "bbox": [ + 511, + 763, + 905, + 795 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During testing, adaptation is performed online: upon receiving instances $\\{x_{i}\\}_{i = 1}^{n_{t}}$ at time $t$ , the model is updated with parameters from the previous instances. The model then generates the prediction $\\{p_i = g(x_i)\\}_{i = 1}^{n_t}$ for the new instances and updates using Eq. (9) with a single step of gradient descent. Adaptation continues as long as test data is available.", + "bbox": [ + 511, + 795, + 905, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2124", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments and Analysis", + "text_level": 1, + "bbox": [ + 89, + 89, + 334, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we conduct extensive experiments to explore the following key questions: (1) Is Test-Time Adaptation (TTA) based on feature redundancy elimination truly effective? (Sec. 5.2) (2) Do the underlying mechanisms of S-FRET and G-FRET align with our theoretical assumptions? (Sec. 5.3) (3) Additionally, we perform ablation studies and evaluate the scalability of our method in more complex scenarios. Sec. 5.4)", + "bbox": [ + 89, + 114, + 483, + 236 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 244, + 282, + 261 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate adaptation performance on two tasks: image domain adaptation and image corruption adaptation. For domain adaptation, we use the PACS [21] and OfficeHome [48] datasets; for corruption adaptation, we use CIFAR-10-C, CIFAR-100-C, and ImageNet-C [13]. To simulate label shifts, we modify CIFAR-100-C to follow a long-tail distribution, denoted as CIFAR-100-C-LT [7]. Comparison methods include BN [36], TENT [49], EATA [28], SAR [29], TSD [51], TIPI [27], and TEA [57], with ResNet-18 and ResNet-50 [12] as backbones. All results are averaged over 5 runs with different seeds, and each method undergoes independent hyperparameter tuning for fair comparison. Implementation details are provided in Sec. 10, and further results and analysis are in Sec. 11.", + "bbox": [ + 89, + 267, + 483, + 478 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Comparison with State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 89, + 486, + 454, + 502 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2.1. Domain Generalization", + "text_level": 1, + "bbox": [ + 89, + 508, + 302, + 522 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 1 presents the accuracy comparison of S-FRET and G-FRET with other state-of-the-art TTA methods on the PACS and OfficeHome datasets based on ResNet-18 and ResNet-50 backbones. The table shows that S-FRET, through its simple yet effective optimization, attains competitive performance comparable to most existing TTA methods, while G-FRET achieves superior performance across both datasets and backbone types. Specifically, on the PACS dataset, G-FRET improves upon the baseline non-adaptive source model by $6.67\\%$ and $5.45\\%$ with ResNet-18 and ResNet-50, respectively. For OfficeHome, the improvements are $1.80\\%$ and $2.34\\%$ , respectively. Compared with state-of-the-art TTA methods, G-FRET achieves the highest or second-highest accuracy overall, particularly when using the ResNet-50 as backbone, G-FRET consistently achieves the highest accuracy across all domains. In contrast, other methods fail to achieve consistently high accuracy across different domains and backbones. This is caused by G-FRET's ability to consider both feature discriminability and redundancy.", + "bbox": [ + 89, + 527, + 483, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2.2. Image Corruption", + "text_level": 1, + "bbox": [ + 89, + 835, + 264, + 851 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 2 compares the performance of S-FRET, G-FRET, and other state-of-the-art methods on CIFAR-10-C, CIFAR100-C, and ImageNet-C at corruption level 5. Two set", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a2ef422711455d64288d1a4e634a575df47c4198505be2a6655d91c251030807.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BackboneMethodPACSAvgOfficeHomeAvg
ACPPSACPRR
ResNet-18Source [12]78.37±0.0077.20±0.0095.03±0.0076.54±0.0081.84±0.0056.45±0.0048.02±0.0071.34±0.0072.23±0.0062.01±0.00
BN [36]81.01±0.0080.67±0.2795.14±0.1573.83±0.1882.66±0.0855.66±0.1749.22±0.2370.77±0.1072.45±0.1962.01±0.12
SAR [29]83.62±0.0081.47±0.0095.14±0.1580.44±0.1585.17±0.0057.30±0.1550.54±0.2570.71±0.1572.75±0.2562.82±0.12
EATA [28]82.06±0.8581.11±1.0395.16±0.2973.39±1.0082.93±0.4656.42±0.4149.75±0.4171.50±0.3572.85±0.4162.63±0.11
TENT [49]83.54±1.7882.26±1.6995.54±0.3181.05±1.7885.60±0.7557.25±0.3250.45±0.2672.18±0.3873.09±0.3463.24±0.13
TSD [51]87.05±0.4786.94±0.6295.98±0.2582.56±1.1286.13±0.4457.80±0.3450.03±0.6971.45±0.4270.92±0.4662.55±0.06
TEA [57]87.04±0.5486.54±0.8596.17±0.3682.17±1.0887.98±0.4857.07±0.8850.47±0.1871.84±0.1872.85±0.1063.06±0.27
TTP [57]84.77±0.1883.48±0.3896.17±0.1884.85±0.2584.85±0.2557.25±0.3250.45±0.2671.45±0.3572.75±0.2562.62±0.19
S-FRET86.25±0.3086.66±0.3795.96±0.2575.90±0.2486.20±0.6656.26±0.2149.86±0.3871.65±0.0372.72±0.1462.62±0.09
G-FRET87.20±0.3587.19±0.2496.28±0.3683.36±1.4388.51±0.3957.49±0.3451.28±0.4273.04±0.2973.40±0.3563.81±0.11
ResNet-50Source [12]83.89±0.0081.02±0.0096.17±0.0078.04±0.0084.78±0.0064.85±0.0052.20±0.0075.04±0.0075.88±0.0067.01±0.00
BN [36]85.19±0.3585.49±0.3096.60±0.1972.25±0.3384.88±0.0863.30±0.2053.02±0.3273.59±0.2375.10±0.2966.25±0.05
SAR [29]85.24±0.3785.49±0.3096.60±0.1976.02±0.8485.84±0.2664.78±0.1756.30±0.5774.83±0.2776.18±0.3068.02±0.19
EATA [28]85.14±0.3185.36±0.2096.54±0.1972.27±0.1184.83±0.1164.25±0.4154.09±0.5374.25±0.3675.72±0.4367.08±0.22
TTP [51]84.77±0.1884.97±0.1996.17±0.2978.53±0.3785.66±0.2563.33±0.3054.99±0.3474.16±0.2975.16±0.2968.74±0.14
TSD [51]80.93±0.8380.94±0.2797.56±0.2781.56±1.5289.97±0.1865.33±0.3056.99±0.3476.12±0.3676.52±0.2468.74±0.14
TEA [57]88.51±0.4387.75±0.5197.17±0.2481.45±1.0288.72±0.2665.62±0.4557.55±0.6475.80±0.3976.85±0.5169.25±0.25
TTP [27]87.62±0.7386.97±0.9697.01±1.1179.66±1.5087.81±0.5664.92±0.2956.47±0.3775.31±0.2776.74±0.3768.36±0.04
S-FRET89.90±0.6089.00±0.4897.46±0.3277.45±0.9188.42±0.1964.42±0.2055.00±0.3475.56±0.2476.41±0.1467.80±0.13
G-FRET90.99±0.8790.00±0.2597.59±0.2982.32±0.5990.23±0.5366.12±0.4157.63±0.3976.46±0.3477.19±0.2369.35±0.20
", + "bbox": [ + 514, + 89, + 905, + 218 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tings are used: continuous adaptation for CIFAR-10/100-C (15 corruption types applied sequentially), and independent adaptation for ImageNet-C (each corruption type evaluated separately). S-FRET significantly outperforms the non-adaptive baseline, with average gains of $22.84\\%$ on CIFAR-10-C, $17.52\\%$ on CIFAR-100-C, and $14.4\\%$ on ImageNet-C. G-FRET consistently surpasses S-FRET by leveraging both feature redundancy and label discriminability, whereas S-FRET only addresses redundancy. Under continuous adaptation, methods not explicitly reducing redundancy perform similarly to S-FRET but worse than G-FRET. This highlights the complementary roles of redundancy reduction and class discriminability in improving TTA performance—a finding further supported by our ablation studies.", + "bbox": [ + 511, + 316, + 906, + 544 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e141d0bd97b82b9850b0b83c8ace3b27ce66811c9724f037adf3281c616a95b2.jpg", + "table_caption": [ + "Table 1. Accuracy comparison of different TTA methods on PACS and OfficeHome datasets based on ResNet-18 and ResNet-50 backbones, including mean and standard deviation of 5 runs with different random seeds. The best results are highlighted in boldface, and the second ones are underlined." + ], + "table_footnote": [], + "table_body": "
MethodCIFAR-10-CCIFAR-100-CImageNet-C
1-78-151-151-78-151-151-78-151-15
Source [12]38.4661.5950.8023.5937.4931.017.7420.7914.70
BN [36]70.4776.4373.6545.7050.5748.3017.8436.1227.59
TENT [49]73.0679.2576.3647.7753.4450.7927.0542.6935.39
EATA [28]70.6076.4273.7047.6453.4050.7131.2346.1539.19
SAR [29]70.6176.5073.7546.6651.2149.0930.3745.7038.55
TIPI [27]73.7680.2177.2147.5052.0649.9327.7842.6935.73
TEA [57]71.7077.4774.7846.0750.6448.5122.8540.7232.38
TSD [51]72.5377.9675.4345.8751.0348.6220.9338.1530.11
S-FRET71.6077.2874.6346.3150.4648.5319.5937.4329.10
G-FRET73.7680.3277.2648.2353.5551.0628.6043.9636.79
", + "bbox": [ + 516, + 553, + 906, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Accuracy comparison of different TTA methods on CIFAR-10/100-C and ImageNet-C datasets at damage level of 5, with 15 types of damage applied two experimental settings: continuous adaptation and independent adaptation.", + "bbox": [ + 511, + 696, + 905, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2.3. Efficiency Comparison", + "text_level": 1, + "bbox": [ + 511, + 761, + 718, + 776 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition to accuracy, we also evaluate the runtime efficiency of different methods to assess the computational performance of S-FRET and G-FRET. As shown in Tab. 3, \"Training\" indicates whether the method involves retraining. ERM represents the source model without adaptation, while the BN method directly updates the batch normalization (BN) parameters, eliminating the need for retraining. Consequently, BN is the most computationally ef", + "bbox": [ + 511, + 780, + 905, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2125", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/843d1baee09aa7806517ecf957c99552e5386b8b52f6edd8e05a7d5d051ddeb7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Cost Time (s)
BackboneERMBNTENTEATASARTSDTIPITEAS-FRETG-FRET
Traning××✓✓✓✓✓✓✓✓
ResNet-18393648567411416265846118
ResNet-5064661131091851532641376102193
VIT-B/161411622102203452952813181185310
AVG81881241282011872361738111207
", + "bbox": [ + 89, + 88, + 485, + 174 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9f21847f3f53b1c25de002eac37afc0bbf502488b452e403243a936ad5195370.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Maximum GPU memory usage (MB)
Backbone TrainingSource ×BN ×TENT ✓EATA ✓SAR ✓TSD ✓TIPI ✓TEA ✓S-FRET ✓G-FRET ✓
ResNet-1890390427893713354831703427822127894781
ResNet-50772773541161536005635660711595154489163
VIT-B/161218-64657324662010584137651730064698045
AVG965-488857305391670377551382449027330
", + "bbox": [ + 89, + 175, + 485, + 248 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ficient method. However, as reported in Tabs. 1 and 2, BN consistently exhibits the lowest accuracy. Among the training methods, S-FRET achieves the lowest runtime and the second-lowest memory consumption across various backbones, while G-FRET exhibits average performance compared to other baselines. These results indicate that S-FRET, as a simple yet effective approach, provides a competitive accuracy-efficiency trade-off. Meanwhile, G-FRET, despite its moderate computational cost, achieves the highest accuracy, demonstrating its effectiveness in balancing efficiency and adaptation performance.", + "bbox": [ + 88, + 289, + 483, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Analysis of Underlying Mechanisms", + "text_level": 1, + "bbox": [ + 89, + 467, + 403, + 483 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3.1. Relation between Feature Redundancy Elimination and Generalizability Enhancement", + "text_level": 1, + "bbox": [ + 89, + 489, + 482, + 520 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The first mechanism to be validated is whether there exists a genuine relationship between feature redundancy and generalizability enhancement. We track the evolution of redundancy scores (RS), loss, and accuracy across increasing adaptation steps. As shown in Fig. 4, S-FRET achieves a consistent reduction in feature redundancy as adaptation progresses, accompanied by stable accuracy improvement, demonstrating the detrimental effect of redundancy on generalization and the efficacy of redundancy elimination. For G-FRET, as the number of adaptation steps increases, there is also a consistent reduction in the redundancy of the extracted embedded features, accompanied by a significant decrease in loss and a notable improvement in accuracy. In summary, both S-FRET and G-FRET has been demonstrated to be highly effective in minimizing redundancy, and redundancy elimination is strongly associated with enhancement of model generalizability.", + "bbox": [ + 89, + 523, + 483, + 781 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3.2. Label Shift", + "text_level": 1, + "bbox": [ + 89, + 790, + 215, + 804 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Is S-FRET truly affected by label shift? Tab. 4 presents the performance of S-FRET and G-FRET under test-time label shifts with long-tailed distributions. Here, the imbalance factor denotes the ratio between the most and least frequent class samples—higher values indicate more skewed class distributions. As shown in Tab. 4, S-FRET achieves", + "bbox": [ + 89, + 809, + 483, + 898 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6c6b5265d711d6457d126007c110dfeff8d1862945693eed37960cf10550ff8c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 88, + 694, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c23fa039f6d590527edb54ccc5d1d6eaf1d786be7889ed8aef9e563eea240d89.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 88, + 885, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/650d8b70e82159da38af4a0948d2ae08e21f8e9ca1d932842fdc69326e495eaf.jpg", + "image_caption": [ + "Figure 4. The relationship of the reduction of redundancy and the enhancement of generalizability attained by S-FRET and G-FRET on CIFAR-10-C and CIFAR-100-C datasets." + ], + "image_footnote": [], + "bbox": [ + 517, + 188, + 710, + 286 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/34eb8cf712c46e6b4bac357f0b5afa0656e0a287feaef868635eee2574bfc985.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 189, + 901, + 286 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2318bd43fa8472ac48e1d5a03c6e6ee98311079a777ec1952892df1a9f511466.jpg", + "table_caption": [ + "Table 3. Efficiency comparison of TTA methods on ImageNet-C (Gaussian Noise) across multiple backbone networks" + ], + "table_footnote": [], + "table_body": "
CIFAR-100-C-LT
Imbalance FactorERMBNSAREATATENTTSDTIPITEAS-FRETG-FRET
131.0148.3049.0950.7150.7948.6248.5149.9348.5351.06
1030.3947.1447.6947.8650.1247.7447.5049.5447.4750.23
10030.1346.0748.2146.0849.8046.9547.1349.5146.6250.27
AVG30.5147.1748.3348.2150.2447.7747.7149.6647.5450.52
", + "bbox": [ + 514, + 347, + 908, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Performance comparison on CIFAR-100-C-LT with different imbalance factors based on ResNet-18.", + "bbox": [ + 511, + 422, + 905, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "competitive performance on CIFAR-100-C under balanced class distributions. However, its accuracy degrades significantly as the imbalance factor increases, with $1.91\\%$ drops on CIFAR-100-C, aligning with the limitations of conventional TTA methods. In contrast, G-FRET delivers more robust performance across all imbalance factors. This highlights G-FRET's unique capability to address label shift in real-world scenarios through contrastive learning.", + "bbox": [ + 511, + 460, + 906, + 582 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3.3. Discriminability Analysis by tSNE", + "text_level": 1, + "bbox": [ + 511, + 589, + 799, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Does G-FRET truly improve discriminability compared to S-FRET? Fig. 5 shows the t-SNE [46] visualization of embedded features before and after adaptation by S-FRET and G-FRET. It is evident that before adaptation, data points from different classes are intermixed without clear separation. After adaptation by S-FRET, although class separation increases, the overlapping regions between classes still lack distinct boundaries. In contrast, G-FRET achieves more distinct cluster separation and sharper decision boundaries, demonstrating its superior capability in learning discriminative features under distribution shifts. Extended experimental results on the CIFAR-10-C are provided in Sec. 11.3.", + "bbox": [ + 511, + 608, + 906, + 790 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Ablation and Extension Study", + "text_level": 1, + "bbox": [ + 511, + 799, + 779, + 815 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4.1. Effects of Components in G-FRET", + "text_level": 1, + "bbox": [ + 511, + 821, + 800, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 5 presents the impact of the four components of G-FRET on model performance. (1) Compared to the source baseline, each individual component contributes positively to the model's performance, with respective gains of $0.30\\%$ ,", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2126", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0e5f608ff42ee0fd764a83fc8a4fc3bcd272cafcca40fa70c73cc7306cbb583f.jpg", + "image_caption": [ + "Figure 5. Discriminability visualization of embedded feature before and after adaptation by S-FRET and G-FRET on the PACS dataset." + ], + "image_footnote": [], + "bbox": [ + 112, + 88, + 460, + 191 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/105868d4fcfca7721b115c65b697fde154470f93009aa89bf97ddad3a8045421.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
OfficeHome
ComponentsACPRAVG↑
Source56.4548.0271.3472.2362.010.00
IndividualLRA55.8349.6770.8372.8962.310.30
R56.9850.4272.3372.7363.121.11
LPA56.7350.7772.1372.9963.161.15
R55.8349.3070.6272.5162.070.06
Ablationw/o LRA57.7251.1872.9973.0863.741.73
R56.7051.1672.0773.1963.281.27
w/o LPA57.1150.5272.4072.9463.241.23
R57.7751.4072.9472.9663.771.76
AttentionLR(A)+LP(A)56.7050.9772.3673.1263.291.28
RedundantLR(R)+LP(R)57.0350.4972.2772.9463.181.17
AllAll57.4951.2873.0474.4063.811.80
", + "bbox": [ + 94, + 248, + 480, + 435 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$1.11\\%$ , $1.15\\%$ , and $0.06\\%$ . Notably, the redundancy component in the representation layer, $\\mathcal{L}_R(R)$ , and the attention component in the prediction layer, $\\mathcal{L}_P(A)$ , show the most significant impact, with $\\mathcal{L}_R(R)$ achieving a $1.15\\%$ improvement. This highlights the importance of redundancy reduction for TTA. (2) The ablation study further confirms the significance of each component, indicating that their removal negatively impacts model performance. (3) Either the redundant or attention components can individually improves model performance, with respective gains of $1.28\\%$ and $1.17\\%$ . When the two components are jointly considered, they yield an even greater improvement. (4) Overall, the experimental results further validate the necessity of considering feature redundancy during the TTA process.", + "bbox": [ + 88, + 507, + 482, + 720 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4.2. Parameter Sensitivity", + "text_level": 1, + "bbox": [ + 89, + 739, + 287, + 757 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our method G-FRET involves one hyperparameter: the trade-off parameter in Eq. (9), $\\lambda$ . As shown in Fig. 6, G-FRET outperforms the source model across a wide range of values for $\\lambda$ . We also investigate the impact of the learning rate (lr) on the experimental results. The best performance is achieved with a learning rate of around $1 \\times 10^{-4}$ ; higher learning rates tend to destabilize the training process and reduce the effectiveness of model, which is also observed in other TTA methods [27-29, 49, 51, 57].", + "bbox": [ + 88, + 763, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f944b598b12597c7347c86bc326038e3545a79d13d558ff2d9debf695301f638.jpg", + "image_caption": [ + "Figure 6. Parameter sensitivity analysis G-FRET on the Office-Home dataset with ResNet-18 backbone in domain Art. Results on complex/large dataset." + ], + "image_footnote": [], + "bbox": [ + 555, + 90, + 712, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/682e904363ec7b741ee93f836a6f1258a5f181409bfd8774db041464b98c81ae.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 90, + 864, + 198 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/54db8fd2247d7710134b2c99da50ba2ff596a7a611a254f191df36a1d1ecc6d6.jpg", + "table_caption": [ + "Table 5. Effects of components in G-FRET, test on the Office-Home dataset with ResNet-18 backbone. A: Attention part, R: Redundancy Part." + ], + "table_footnote": [], + "table_body": "
DatasetsResNet-18[12]+S-FRET+G-FRETResNet-50[12]+S-FRET+G-FRETViT-B/16[6]+S-FRET+G-FRET
VLCS[45]73.6773.9075.7374.9075.4876.4276.4477.1878.56
DomainNet[34]39.1339.2640.2243.5843.6944.2249.9851.2251.50
ImageNet-C[13]14.7029.1036.7918.1534.6542.8439.8360.3662.50
Results with ViT-B/16 as backbone.
DatasetsSourceTENTPLCEATASARTIPITEATSDDeYOS-FRETG-FRET
CIFAR10-C71.4871.4871.4673.6771.4871.4871.4871.4777.8677.9573.17
CIFAR10-C-CO71.4871.4671.4472.5470.3371.4571.6371.6974.6475.4774.02
CIFAR100-C51.7152.7251.7062.4754.2952.3652.3151.7459.5763.4657.65
CIFAR100-C-CO51.7151.7751.7956.2652.5051.7951.9251.9054.0756.6952.14
PACS87.0387.7386.1787.0787.1287.8787.3783.4388.0787.1988.12
OfficeHome76.1176.8676.5076.7476.6676.9876.1575.4976.7876.5577.62
AVG68.2568.6768.1871.4668.7368.6568.4867.6271.8372.8870.45
Results on Multimodal task
DatasetsCLIPTENTTPTCLIPArTTWATT-PWATT-SS-FRET
VLCS80.9680.8572.5380.8981.0581.5981.69
PACS95.1295.2293.0895.3596.0796.0896.09
OfficeHome80.2880.4781.9180.4281.8281.7181.67
AVG85.4585.5182.5185.5586.3186.4686.48
", + "bbox": [ + 516, + 247, + 903, + 406 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Scalability on complex/large datasets, different backbones and multimoda task.", + "bbox": [ + 511, + 417, + 903, + 445 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4.3. Scalability on complex/large datasets, different backbones and multimoda task.", + "text_level": 1, + "bbox": [ + 511, + 458, + 903, + 488 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We validate our method on large-scale datasets, diverse backbones, and a multimodal task to demonstrate its robustness and architecture-independence. As shown in Tab. 6, reducing embedding redundancy leads to notably larger gains on ViTs than on CNNs, likely due to the higher redundancy in Transformer representations. In the multimodal CLIP setting [31], S-FRET continues to deliver consistent improvements by explicitly aligning embedding distributions, unlike CLIP's similarity-based matching. Notably, on CIFAR with ViT, S-FRET significantly outperforms G-FRET, suggesting that the effectiveness of redundancy reduction varies across backbone types and datasets.", + "bbox": [ + 511, + 494, + 906, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 691, + 633, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a novel approach for test-time adaptation (TTA) — Feature Redundancy Elimination for TTA (FRET) — which focuses on reducing feature redundancy during testing to address distribution shifts. This is the first attempt to leverage feature redundancy for TTA. We introduce two methods: S-FRET, which directly minimizes the feature redundancy score, and G-FRET, which optimizes both redundancy elimination and class discriminability by decomposing feature relations into attention and redundancy components. Experiments across various tasks and datasets show the effectiveness of our methods, demonstrating the value of feature redundancy elimination in TTA.", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2127", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Author Contributions and Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 89, + 480, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Linjing You initiated this work and was primarily responsible for proposing the core idea, methodology, theoretical analysis, experimental design, and discussion of key findings. Jiabao Lu conducted most of the experiments and contributed to the interpretation of experimental results. Xiayuan Huang participated in experimental discussions and provided feedback on the initial manuscript. Xiangli Nie offered valuable suggestions for the refinement and polishing of the paper.", + "bbox": [ + 89, + 114, + 483, + 251 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We would also like to sincerely thank all the anonymous reviewers and the editorial committee (ACS) whose insightful comments significantly improved the theoretical rigor, experimental depth, and overall quality of this work.", + "bbox": [ + 89, + 251, + 482, + 311 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This research was supported by Prof. Huang's project The Fundamental Research Funds for the Central Universities (Grant No. XJJSKYQD202538), Prof. Nie's project The National Key Research and Development Program of China (Grant No. 2024YFB4709100), and partially funded by the Beijing Nova Program (Grant Nos. 20220484070 and 20240484498).", + "bbox": [ + 89, + 311, + 483, + 416 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 428, + 187, + 444 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ivan Anishchenko, Samuel J Pellock, Tamuka M Chidyausiku, Theresa A Ramelot, Sergey Ovchinnikov, Jingzhou Hao, Khushboo Bafna, Christoffer Norn, Alex Kang, Asim K Bera, et al. De novo protein design by deep network hallucination. Nature, 600(7889):547-552, 2021. 1", + "[2] Gavin Brown, Adam Pocock, Ming-Jie Zhao, and Mikel Luján. Conditional likelihood maximisation: a unifying framework for information theoretic feature selection. The journal of machine learning research, 13(1):27-66, 2012. 1", + "[3] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label distribution-aware margin loss. In Proc. the Advances in Neural Information Processing Systems (NeurIPS), pages 1567-1578, 2019. 2, 4", + "[4] Xue-wen Chen and Jong Cheol Jeong. Enhanced recursive feature elimination. In Sixth international conference on machine learning and applications (ICMLA 2007), pages 429-435. IEEE, 2007. 1", + "[5] Ian Connick Covert, Wei Qiu, Mingyu Lu, Na Yoon Kim, Nathan J White, and Su-In Lee. Learning to maximize mutual information for dynamic feature selection. In International Conference on Machine Learning, pages 6424-6447. PMLR, 2023. 1", + "[6] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.8", + "[7] C. Du, Y. Wang, S. Song, and G. Huang. Probabilistic contrastive learning for long-tailed visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(9):5890-5904, 2024. 6, 3", + "[8] Tongtong Fang, Nan Lu, Gang Niu, and Masashi Sugiyama. Rethinking importance weighting for deep learning under" + ], + "bbox": [ + 99, + 453, + 482, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "distribution shift. Advances in neural information processing systems, 33:11996-12007, 2020. 1", + "[9] Ronald A Fisher. The use of multiple measurements in taxonomic problems. Annals of eugenics, 7(2):179-188, 1936. 1", + "[10] Taesik Gong, Yewon Kim, Taeckyung Lee, Sorn Chottananurak, and Sung-Ju Lee. Sotta: Robust test-time adaptation on noisy data streams. Advances in Neural Information Processing Systems, 36, 2024. 1", + "[11] Jie Gui, Zhenan Sun, Shuiwang Ji, Dacheng Tao, and Tieniu Tan. Feature selection based on structured sparsity: A comprehensive study. IEEE transactions on neural networks and learning systems, 28(7):1490-1507, 2016. 1", + "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 2, 6, 8, 3, 7, 9, 10", + "[13] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019. 2, 6, 8, 3", + "[14] Chenping Hou, Ruidong Fan, Ling-Li Zeng, and Dewen Hu. Adaptive feature selection with augmented attributes. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):9306-9324, 2023. 1", + "[15] John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Žídek, Anna Potapenko, et al. Highly accurate protein structure prediction with alphafold. nature, 596(7873):583-589, 2021. 1", + "[16] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. In Proc. the International Conference on Learning Representations (ICLR), 2019. 2, 4", + "[17] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 5", + "[18] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. nature, 521(7553):436-444, 2015. 1", + "[19] Jonghyun Lee, Dahuin Jung, Saehyung Lee, Junsung Park, Juhyeon Shin, Uiwon Hwang, and Sungroh Yoon. Entropy is not enough for test-time adaptation: From the perspective of disentangled factors. In The Twelfth International Conference on Learning Representations, 2024. 2", + "[20] Taeckyung Lee, Sorn Chottananurak, Taesik Gong, and Sung-Ju Lee. Aetta: Label-free accuracy estimation for test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28643-28652, 2024. 3", + "[21] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. In Proceedings of the IEEE international conference on computer vision, pages 5542-5550, 2017. 6, 3", + "[22] Jundong Li, Kewei Cheng, Suhang Wang, Fred Morstatter, Robert P Trevino, Jiliang Tang, and Huan Liu. Feature selection: A data perspective. ACM computing surveys (CSUR), 50(6):1-45, 2017. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2128", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Junnan Li, Pan Zhou, Caiming Xiong, and Steven C.H. Hoi. Prototypical contrastive learning of unsupervised representations. In ICLR, 2021. 5", + "[24] Jian Liang, Ran He, and Tieniu Tan. A comprehensive survey on test-time adaptation under distribution shifts. International Journal of Computer Vision, pages 1-34, 2024. 1, 2", + "[25] Duy MH Nguyen, Hoang Nguyen, Nghiem Diep, Tan Ngoc Pham, Tri Cao, Binh Nguyen, Paul Swoboda, Nhat Ho, Shadi Albarqouni, Pengtao Xie, et al. Lvm-med: Learning large-scale self-supervised vision models for medical imaging via second-order graph matching. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[26] Wamidh K Mutlag, Shaker K Ali, Zahoor M Aydam, and Bahaa H Taher. Feature extraction methods: a review. In Journal of Physics: Conference Series, page 012028. IOP Publishing, 2020. 3, 1", + "[27] A Tuan Nguyen, Thanh Nguyen-Tang, Ser-Nam Lim, and Philip HS Torr. Tipi: Test time adaptation with transformation invariance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24162-24171, 2023. 6, 8, 1, 3, 7, 9, 10", + "[28] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In International conference on machine learning, pages 16888-16905. PMLR, 2022. 6, 1, 2, 3, 7, 8, 9, 10", + "[29] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yaofo Chen, Peilin Zhao, and Mingkui Tan. Towards stable test-time adaptation in dynamic wild world. In International Conference on Learning Representations, 2023. 6, 8, 1, 3, 7, 9, 10", + "[30] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 5", + "[31] David Osowiechi, Mehrdad Noori, Gustavo Vargas Hakim, Moslem Yazdanpanah, Ali Bahri, Milad Cheraghalikhani, Sahar Dastani, Farzad Beizaee, Ismail Ayed, and Christian Desrosiers. Watt: Weight average test time adaptation of clip. Advances in neural information processing systems, 37: 48015-48044, 2024. 8", + "[32] Sung Kyu Park, Seunghan Yang, Jaegul Choo, and Sungrack Yun. Label shift adapter for test-time adaptation under covariate and label shifts. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 16375-16385, 2023. 2, 4", + "[33] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 3", + "[34] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1406-1415, 2019. 8, 3" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Sharan Ramjee and Aly El Gamal. Efficient wrapper feature selection using autoencoder and model based elimination. arXiv preprint arXiv:1905.11592, 2019. 1", + "[36] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. Advances in neural information processing systems, 33:11539-11551, 2020. 3, 6, 1, 7, 8, 9, 10", + "[37] Claude Elwood Shannon. A mathematical theory of communication. The Bell system technical journal, 27(3):379-423, 1948. 2", + "[38] David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering the game of go with deep neural networks and tree search. nature, 529(7587):484-489, 2016. 1", + "[39] Samarth Sinha, Peter Gehler, Francesco Locatello, and Bernt Schiele. Test: Test-time self-training under distribution shift. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2759-2769, 2023. 2", + "[40] Yongyi Su, Xun Xu, and Kui Jia. Towards real-world test-time adaptation: Tri-net self-training with balanced normalization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 15126-15135, 2024. 1", + "[41] Yongyi Su, Xun Xu, Tianrui Li, and Kui Jia. Revisiting realistic test-time training: Sequential inference and adaptation by anchored clustering regularized self-training. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1", + "[42] Yue Tan, Chen Chen, Weiming Zhuang, Xin Dong, Lingjuan Lyu, and Guodong Long. Is heterogeneity notorious? taming heterogeneity to handle test-time shift in federated learning. Advances in Neural Information Processing Systems, 36, 2024. 1", + "[43] Dipti Theng and Kishor K Bhoyar. Feature selection techniques for machine learning: a survey of more than two decades of research. Knowledge and Information Systems, 66(3):1575-1637, 2024. 3, 1", + "[44] Robert Tibshirani. Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society Series B: Statistical Methodology, 58(1):267-288, 1996. 1", + "[45] Antonio Torralba and Alexei A Efros. Unbiased look at dataset bias. In CVPR 2011, pages 1521-1528. IEEE, 2011. 8, 3", + "[46] Laurens van der Maaten and Geoffrey E. Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9: 2579-2605, 2008. 7", + "[47] B Venkatesh and J Anuradha. A review of feature selection and its methods. Cybernetics and information technologies, 19(1):3-26, 2019. 1", + "[48] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5018-5027, 2017. 6, 3" + ], + "bbox": [ + 516, + 90, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2129", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726, 2020. 6, 8, 1, 2, 3, 7, 9, 10", + "[50] Mei Wang and Weihong Deng. Deep visual domain adaptation: A survey. Neurocomputing, 312:135-153, 2018. 1", + "[51] Shuai Wang, Daoan Zhang, Zipei Yan, Jianguo Zhang, and Rui Li. Feature alignment and uniformity for test time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20050-20060, 2023. 6, 8, 1, 2, 3, 7, 9, 10", + "[52] Svante Wold, Kim Esbensen, and Paul Geladi. Principal component analysis. Chemometrics and intelligent laboratory systems, 2(1-3):37-52, 1987. 1", + "[53] Yanan Wu, Zhixiang Chi, Yang Wang, Konstantinos N Plataniotis, and Songhe Feng. Test-time domain adaptation by learning domain-aware batch normalization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 15961-15969, 2024. 2", + "[54] Kaichao You, Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Universal domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2720-2729, 2019. 1", + "[55] Linjing You, Jiabao Lu, and Xiayuan Huang. Test-time correlation alignment. In Forty-second International Conference on Machine Learning, 2025. 3, 2", + "[56] Longhui Yuan, Binhui Xie, and Shuang Li. Robust test-time adaptation in dynamic scenarios. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15922-15932, 2023. 1", + "[57] Yige Yuan, Bingbing Xu, Liang Hou, Fei Sun, Huawei Shen, and Xueqi Cheng. Tea: Test-time energy adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23901-23911, 2024. 6, 8, 1, 3, 7, 9, 10", + "[58] Han Yue, Jundong Li, and Hongfu Liu. Second-order unsupervised feature selection via knowledge contrastive distillation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3, 4, 1, 2", + "[59] Longbin Zeng, Jiayi Han, Liang Du, and Weiyang Ding. Rethinking precision of pseudo label: Test-time adaptation via complementary learning. Pattern Recognition Letters, 177: 96-102, 2024. 2", + "[60] Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. Domain generalization: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4396-4415, 2022. 1" + ], + "bbox": [ + 91, + 92, + 480, + 753 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2130", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/FRET_ Feature Redundancy Elimination for Test Time Adaptation/013e8d33-0504-4dd6-a3e9-326514c12e30_model.json b/2025/FRET_ Feature Redundancy Elimination for Test Time Adaptation/013e8d33-0504-4dd6-a3e9-326514c12e30_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f297048e708a3e2a9f8bbef1e3611cc381d94ba9 --- /dev/null +++ b/2025/FRET_ Feature Redundancy Elimination for Test Time Adaptation/013e8d33-0504-4dd6-a3e9-326514c12e30_model.json @@ -0,0 +1,2488 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.163, + 0.131, + 0.836, + 0.154 + ], + "angle": 0, + "content": "FRET: Feature Redundancy Elimination for Test Time Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.179, + 0.764, + 0.199 + ], + "angle": 0, + "content": "Linjing You\\(^{1,*}\\) Jiabao Lu\\(^{1,*}\\) Xiayuan Huang\\(^{2,\\dagger}\\) Xiangli Nie\\(^{1,\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.199, + 0.782, + 0.217 + ], + "angle": 0, + "content": "\\(^{1}\\)Institute of Automation, Chinese Academy of Sciences, Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.217, + 0.751, + 0.235 + ], + "angle": 0, + "content": "\\(^{2}\\)College of Science, Beijing Forestry University, Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.196, + 0.235, + 0.804, + 0.253 + ], + "angle": 0, + "content": "{youlinjing2023,lujiabao2025,xiangliNie}@ia.ac.cn" + }, + { + "type": "text", + "bbox": [ + 0.357, + 0.254, + 0.641, + 0.269 + ], + "angle": 0, + "content": "huangxiayuan@bjfu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.378, + 0.269, + 0.616, + 0.281 + ], + "angle": 0, + "content": "* Equal contribution, †Corresponding author" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.322, + 0.327, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.357, + 0.486, + 0.811 + ], + "angle": 0, + "content": "Test-Time Adaptation (TTA) aims to enhance the generalization of deep learning models when faced with test data that exhibits distribution shifts from the training data. In this context, only a pre-trained model and unlabeled test data are available, making it particularly relevant for privacy-sensitive applications. In practice, we observe that feature redundancy in embeddings tends to increase as domain shifts intensify in TTA. However, existing TTA methods often overlook this redundancy, which can hinder the model's adaptability to new data. To address this issue, we introduce Feature Redundancy Elimination for Test-time Adaptation (FRET), a novel perspective for TTA. A straightforward approach (S-FRET) is to directly minimize the feature redundancy score as an optimization objective to improve adaptation. Despite its simplicity and effectiveness, S-FRET struggles with label shifts, limiting its robustness in real-world scenarios. To mitigate this limitation, we further propose Graph-based FRET (G-FRET), which integrates a Graph Convolutional Network (GCN) with contrastive learning. This design not only reduces feature redundancy but also enhances feature discriminability in both the representation and prediction layers. Extensive experiments across multiple model architectures, tasks, and datasets demonstrate the effectiveness of S-FRET and show that G-FRET achieves state-of-the-art performance. Further analysis reveals that G-FRET enables the model to extract non-redundant and highly discriminative features during inference, thereby facilitating more robust test-time adaptation. The code is available at https://github.com/youlj109/fret." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.844, + 0.222, + 0.859 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Recently, the emergence of deep neural networks (DNNs) [12, 18] has significantly propelled the advancement in a" + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.33, + 0.878, + 0.574 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.585, + 0.908, + 0.643 + ], + "angle": 0, + "content": "Figure 1. An intuitive demonstration of the relationship between embedded feature redundancy and distribution migration. Top: Second-order feature relation graph of embedding features. Bottom: Plot of feature redundancy versus the level of corruption." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.674, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Vast range of human tasks [1, 15, 38]. However, the basic assumption that training and test data are drawn independently and identically (i.i.d.) may not be suitable in many real-world scenarios where the test/target distribution diverges from the training/source distribution [50]. This phenomenon, known as distribution shift, often leads to poor performance [8]. To tackle this issue, a variety of generalization or adaptation techniques have been developed to improve model robustness against distribution shifts [24, 54, 60]. Among these, Test Time Adaptation (TTA) [10, 40, 41] has been the focus of recent and active research. TTA has the advantage of only requiring access to the pre-trained model from the training/source domain and unlabeled test data. Therefore, TTA is a more secure and practical solution especially when models are publicly available." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2120" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.122 + ], + "angle": 0, + "content": "available but the training data and training process remain inaccessible due to privacy and resource restrictions [24]." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.125, + 0.484, + 0.442 + ], + "angle": 0, + "content": "Existing TTA methods can be broadly categorized into batch normalization calibration methods, pseudo-labeling methods, consistency training methods, and clustering-based training methods, based on different adaptation strategies [24]. Batch normalization (BN) calibration methods [53] posit that the statistics in the BN layers represent domain-specific knowledge. To bridge the domain gap, it is suggested to replace the training BN statistics with new statistics estimated over the entire target domain. Pseudolabeling methods [59] employ various filtering strategies, such as thresholding or entropy-based approaches, to obtain reliable pseudo-labels. This process helps minimize discrepancies between predicted labels and pseudo-labels. Consistency training methods [39] aim to improve the stability of network predictions or features by addressing variations in input data, such as noise or perturbations, and changes in model parameters. Clustering-based training methods [19] focus on reducing uncertainty in target network predictions by leveraging clustering techniques to organize target features into distinct groups, thereby enhancing the robustness of the model." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.445, + 0.484, + 0.732 + ], + "angle": 0, + "content": "Despite the empirical success of existing methods, they overlook feature redundancy in embeddings, which hinders generalization to new data [22]. As shown in Fig. 1, we visualize feature redundancy in ResNet-18 [12] using second-order feature relations (i.e., covariance matrix) on the CIFAR10-C dataset [13] under distribution shift [25]. Each image is embedded into a 512-dimensional vector; stacking \\( n \\) embeddings yields an \\( n \\times 512 \\) matrix. We compute the absolute covariance matrix \\( (512 \\times 512) \\) and visualize it as a heatmap — redder areas indicate higher feature correlation, hence greater redundancy. The top two subfigures in Fig. 1 compare the source and target domains, revealing much higher redundancy in the latter due to distribution shift. To quantify this, we remove the diagonal of the covariance matrix and sum the absolute values of the remaining elements (see Eq. (1)). As shown at the bottom of Fig. 1, redundancy increases with corruption severity, suggesting that features with lower redundancy may be better suited for adaptation to the target domain." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Building on the limitations identified in the existing literature and the observations above, we provide a novel perspective for test-time adaptation—Feature Redundancy Elimination for TTA (FRET)—which differs from existing methods. Due to the inherently differentiable nature of the redundancy metric, a straightforward and computationally efficient approach is to directly minimize the feature redundancy score. We refer to this simple approach as S-FRET. However, since S-FRET lacks the ability to perceive label distributions, it effectively addresses covariate shift but fails to handle label shift—a scenario where the tar" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "get domain exhibits a label distribution that deviates from the source domain (e.g., long-tailed data). This discrepancy can significantly degrade model performance and remains a critical challenge [3, 16, 32]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.153, + 0.907, + 0.425 + ], + "angle": 0, + "content": "To overcome S-FRET's limitation in handling label shift, we further propose G-FRET, which integrates a Graph Convolutional Network (GCN). Unlike FRET, which only minimizes redundancy at the representation layer, G-FRET leverages GCN to incorporate both attention and redundant feature relationships into the representation and prediction layers. Specifically, G-FRET learns non-redundant and discriminative representation through contrastive distillation between attention and redundant representation in the representation-layer. In the prediction layer, attention predictions are optimized through entropy minimization, while ensuring that they are distinct from redundant predictions by negative learning. By maximizing the discriminability of attention relationships while eliminating redundancy in both the representation and prediction layers, G-FRET effectively reduces test-time feature redundancy and improves discrimination when handling complex distribution-shifted test data." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.427, + 0.904, + 0.442 + ], + "angle": 0, + "content": "We summarize the contributions of this paper as follows." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.448, + 0.905, + 0.553 + ], + "angle": 0, + "content": "- Potential Way: provide a novel perspective for test-time adaptation — Feature Redundancy Elimination for TTA (FRET) — which differs from existing methods. It adapts the model to distribution shifts by reducing the redundancy of embedded features during testing. To the best of our knowledge, this is the first attempt to leverage feature redundancy for TTA." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.554, + 0.905, + 0.674 + ], + "angle": 0, + "content": "- Innovative Methods: We propose two novel methods, S-FRET and G-FRET, to eliminate feature redundancy in test-time embeddings. S-FRET directly minimizes the feature redundancy score, offering a lightweight solution. G-FRET extends this by decomposing feature relations into attention and redundancy components, enabling joint optimization of redundancy elimination and class discriminability in real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.675, + 0.905, + 0.795 + ], + "angle": 0, + "content": "- Extensive Evaluation: Comprehensive experiments across multiple model architectures, tasks, and datasets demonstrate the effectiveness of S-FRET and show that G-FRET achieves state-of-the-art performance. Further in-depth analysis reveals that G-FRET facilitates the extraction of non-redundant and discriminative embedding features during testing, ultimately enhancing model adaptation to the test-time target domain." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.448, + 0.905, + 0.795 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.814, + 0.762, + 0.831 + ], + "angle": 0, + "content": "2. Preliminary and Notations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.901 + ], + "angle": 0, + "content": "We briefly revisit Test-Time Adaptation and Feature Redundancy Elimination in this section for the convenience of introducing methods, and put detailed related work discussions into Sec. 7 due to page limits." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2121" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.089, + 0.32, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.32, + 0.089, + 0.502, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.089, + 0.685, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.684, + 0.089, + 0.865, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.225, + 0.908, + 0.253 + ], + "angle": 0, + "content": "Figure 2. Redundancy scores of 15 corruption types in CIFAR-10/100-C and ImageNet-C across 5 corruption levels, as well as for source and target domains in PACS, OfficeHome, VLCS, and DomainNet." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.26, + 0.296, + 0.276 + ], + "angle": 0, + "content": "2.1. Test-Time Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.283, + 0.484, + 0.525 + ], + "angle": 0, + "content": "In the test-time adaptation (TTA) scenario, we have access only to unlabeled data from the target domain in an online manner and a pre-trained model from the source domain. Specifically, let \\(\\{(x_i,y_i)\\}_{i = 1}^{n_s}\\subset \\mathcal{X}_s\\times \\mathcal{Y}_s\\) represent the labeled source domain dataset with labels \\(\\mathcal{V}_s\\) and \\(\\{x_j\\}_{j = 1}^{n_t}\\subset \\mathcal{X}_t\\) represent a batch of unlabeled target data. The model, trained on the source domain dataset and parameterized by \\(\\theta\\), is denoted as \\(g_{\\theta} = h(f(\\cdot)): \\mathcal{X}_s\\to \\mathcal{Y}_s\\), where \\(f\\) is the backbone encoder and \\(h\\) denotes the linear classification head. Our objective is to adapt this model to the target domain \\(\\mathcal{X}_t\\), which exhibits a different distribution from that of the source domain \\(\\mathcal{X}_s\\). During testing, for each instance \\(x_{j}\\in \\mathcal{X}_{t}\\), let the output of \\(f\\) and \\(h\\) be denoted as embedding \\(z_{j} = f(x_{j})\\in \\mathbb{R}^{d}\\) and logs \\(p_j = h(z_j)\\in \\mathbb{R}^C\\), respectively, where \\(d\\) is the dimension of the embeddings and \\(C\\) is the number of classes." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.537, + 0.383, + 0.552 + ], + "angle": 0, + "content": "2.2. Feature Redundancy Elimination" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.56, + 0.484, + 0.635 + ], + "angle": 0, + "content": "Feature redundancy is a key concern in both feature extraction and feature selection [26, 43]. Following previous work [58], we remove the diagonal elements of the covariance matrix and calculate the sum of the absolute values of the remaining elements to derive the redundancy scores \\( R_{s} \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.647, + 0.482, + 0.673 + ], + "angle": 0, + "content": "\\[\nR _ {s} = \\left\\| \\tilde {Z} ^ {T} \\tilde {Z} - I _ {d} \\right\\| _ {1} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.685, + 0.484, + 0.735 + ], + "angle": 0, + "content": "where \\(I_{d}\\) is the identity matrix, \\(\\| \\cdot \\| _1\\) denotes the 1-norm of a matrix and \\(\\tilde{Z}_{:,j} = \\frac{Z_{:,j}}{\\|Z_{:,j}\\|_2}\\) denotes the matrix obtained by normalizing each column." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.75, + 0.466, + 0.767 + ], + "angle": 0, + "content": "3. Feature Redundancy Under Domain Shift" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.776, + 0.484, + 0.806 + ], + "angle": 0, + "content": "We conduct a systematic investigation into the behavior of feature redundancy under domain shift." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.258, + 0.835 + ], + "angle": 0, + "content": "3.1. Empirical Study." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "As shown in Fig. 2, across approximately 50 sub-datasets, the feature redundancy score \\( R_{s} \\) consistently increases with the severity of domain shift. (Note: \\( R_{s} \\) is computed using models with BN parameters updated from test data" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.261, + 0.906, + 0.336 + ], + "angle": 0, + "content": "rather than those fixed during training [36].) This empirical trend suggests that domain shift induces stronger correlations among features. To the best of our knowledge, this phenomenon has not been systematically explored in prior work." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.345, + 0.741, + 0.362 + ], + "angle": 0, + "content": "3.2. Theoretical Explanation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.367, + 0.906, + 0.428 + ], + "angle": 0, + "content": "Intuitively, higher covariance elements indicate that the features fail to span the full vector space. Furthermore, we provide the following theoretical result. Missing proofs and detailed explanations are provided in Sec. 9" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.436, + 0.907, + 0.603 + ], + "angle": 0, + "content": "Theorem 1 (Redundancy Elimination Improves Generalization). Under the empirical observation that embeddings from the source domain exhibit lower feature redundancy, i.e., \\(\\mathcal{R}_s(Z_t) > \\mathcal{R}_s(Z_s)\\), and by leveraging Theorem 2 from [55], we can derive the following generalization error bound: Let \\(\\mathcal{H}\\) be a hypothesis class of VC-dimension \\(d_v\\). If \\(\\hat{h} \\in \\mathcal{H}\\) minimizes the empirical error \\(\\hat{\\epsilon}_s(h)\\) on \\(\\mathcal{X}_s\\) and \\(h_t^* = \\arg \\min_{h \\in \\mathcal{H}} \\epsilon_t(h)\\) is the optimal hypothesis on \\(\\mathcal{X}_t\\), with the assumption that all hypotheses are L-Lipschitz continuous, then \\(\\forall \\delta \\in (0,1)\\), with probability with at least \\(1 - \\delta\\) the following inequality holds:" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.61, + 0.878, + 0.657 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\epsilon_ {t} (\\hat {h}) - \\epsilon_ {t} (h _ {t} ^ {*}) \\\\ \\leq \\mathcal {O} (\\sqrt {\\| \\mu_ {s} - \\mu_ {t} \\| _ {F} ^ {2} + R _ {s} (Z _ {s}) ^ {2} + R _ {s} (Z _ {t}) ^ {2}}) + C \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.668, + 0.906, + 0.766 + ], + "angle": 0, + "content": "where \\( C = 2\\sqrt{\\frac{d_v\\log(2n_s) - \\log(\\delta)}{2n_s}} + 2\\gamma \\) and \\( \\gamma = \\min_{h\\in \\mathcal{H}}\\{\\epsilon_s(h(t)) + \\epsilon_t(h(t))\\} \\). \\( \\mu_{s} \\) and \\( \\mu_t \\) denote the means of the source and test embeddings, respectively. We use \\( \\mathcal{O}(\\cdot) \\) to hide the constant dependence. Since \\( \\mathcal{R}_s(Z_t) > \\mathcal{R}_s(Z_s) \\), this implies that reducing \\( \\mathcal{R}_s(Z_t) \\) can effectively decrease the generalization error." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.779, + 0.742, + 0.796 + ], + "angle": 0, + "content": "4. The Proposed Approach" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.803, + 0.615, + 0.818 + ], + "angle": 0, + "content": "4.1. S-FRET" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.9 + ], + "angle": 0, + "content": "To eliminate feature redundancy in the target domain, a simple yet effective approach is to directly minimize the feature redundancy score \\( R_{s} \\), as defined in Eq. (1), leading to the proposed method S-FRET (Simple and Straightforward FRET). The loss function for S-FRET can be written as:" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2122" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.088, + 0.868, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.313, + 0.908, + 0.371 + ], + "angle": 0, + "content": "Figure 3. The pipeline of our proposed Graph Convolutional Network based Feature Redundancy Elimination for Test Time Adaptation (G-FRET) method. For each test sample, G-FRET learns non-redundant and discriminative representation through contrastive distillation in the representation-layer. Additionally, in the prediction layer, attention predictions are optimized through entropy minimization, while ensuring that they are distinct from redundant predictions by using negative learning." + }, + { + "type": "equation", + "bbox": [ + 0.196, + 0.389, + 0.483, + 0.419 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S F R E T} = \\left\\| \\tilde {Z} ^ {T} \\tilde {Z} - I _ {d} \\right\\| _ {1} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.423, + 0.484, + 0.529 + ], + "angle": 0, + "content": "This method directly minimizes the redundancy score derived from feature correlation, providing a computationally efficient solution that avoids complex architectural modifications. By focusing on the feature space of representation-layer, S-FRET enables the model to extract non-redundant representations, thereby enhancing its adaptability to the target domain (see Fig. 4)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.53, + 0.484, + 0.651 + ], + "angle": 0, + "content": "However, since the redundancy score does not incorporate information about class distributions, S-FRET primarily addresses covariate shift but lacks the discriminative capability to handle label shift—a critical challenge where the target domain exhibits a label distribution that deviates significantly from the source domain (e.g., long-tailed data), leading to substantial performance degradation (see Tab. 4) [3, 16, 32]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.661, + 0.196, + 0.675 + ], + "angle": 0, + "content": "4.2. G-FRET" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.684, + 0.288, + 0.699 + ], + "angle": 0, + "content": "4.2.1. Overview of G-FRET" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.824 + ], + "angle": 0, + "content": "To eliminate feature redundancy and address label shift in the target domain, we further propose a novel two-layer TTA method, namely Graph Convolutional Network based Feature Redundancy Elimination for Test Time Adaptation (G-FRET), which leverages GCN to simultaneously model feature relationships and sample-specific discriminative information in both representation and prediction layers. Fig. 3 provides an overview of our G-FRET framework." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The key idea of G-FRET is to adapt the pre-tarined model during testing by extracting discriminative and nonredundant embedding features from both the representation and prediction layers. Firstly (Sec. 4.2.2), we apply a mask matrix to the second-order feature relation graph." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.377, + 0.907, + 0.664 + ], + "angle": 0, + "content": "This process yields two separate graphs: an attention graph, which highlights significant feature relationships, and a redundant graph, which identifies redundant feature relationships. Next, we derive the attention/redundancy representation and attention/redundancy prediction by utilizing a single-layer GCN to integrate the model's embeddings with the attention and redundant graphs. This process allows us to explicitly separate attention and redundancy relations in both representation and prediction layer. Then (Sec. 4.2.3), in the representation layer, contrastive learning techniques are employed to enhance the discriminability of the attention representation while ensuring it remains distant from the redundant representation. At last (Sec. 4.2.4), in the prediction layer, attention predictions are optimized through entropy minimization, while ensuring that they are distinct from redundant predictions by negative learning. Through this two-layer process, we achieve discriminative and nonredundant embedding features that effectively adapt to the target domain." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.678, + 0.818, + 0.693 + ], + "angle": 0, + "content": "4.2.2. Attention Part and Redundancy Part" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.699, + 0.907, + 0.879 + ], + "angle": 0, + "content": "Given a batch of \\( n_t \\) data instances from the target domain, we obtain the first-order \\( n_t \\times d \\) embedding feature matrix \\( Z \\in \\mathbb{R}^{n_t \\times d} \\) and the second-order feature relation graph \\( G_F = Z^T Z \\in \\mathbb{R}^{d \\times d} \\). Following the state-of-the-art feature selection method SOFT [58] (see Appendix 7.3 for more details), we apply a symmetric mask matrix \\( M_M \\in \\mathbb{R}^{d \\times d} \\) to \\( G_F \\). This decomposition of \\( G_F \\) yields two distinct graphs: the attention graph \\( G_A \\) and the redundancy graph \\( G_R \\). \\( G_A \\) captures the attention relations (i.e., important relation) among features, while \\( G_R \\) represents the redundant relation between them. Formally, they are calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "\\[\nG _ {A} = G _ {F} \\odot M _ {M}; G _ {R} = G _ {F} - G _ {A} \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2123" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.304 + ], + "angle": 0, + "content": "where \\(\\odot\\) is the element-wise product. However, differently from SOFT, which focuses on feature selection by deriving an optimal mask matrix \\(M_M\\), our objective is to extract non-redundant features \\(Z\\). For ideally non-redundant features, the second-order feature relation graph should be a diagonal matrix, indicating that each feature is independent of all others. Therefore, we set \\(M_M = I_d\\) to ensure that \\(G_A\\) captures only the essential feature relations, while \\(G_R\\) includes all remaining relations as redundant relations. Here, \\(I_d \\in \\mathbb{R}^{d \\times d}\\) is the identity matrix. In addition, in some scenarios with prior knowledge, such as knowing which features may be more important, or needing to maintain special structural relations between features, some adaptations can be made by simply modifying the mask matrix \\(M_M\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.304, + 0.484, + 0.425 + ], + "angle": 0, + "content": "Then, in order to obtain attention/redundancy representation and attention/redundancy prediction based on the attention and redundancy graphs, we combine the normalized graph matrix with the classifier head \\( h \\) to form a one-layer Graph Convolutional Network (GCN [17]), which can help generate attention representations \\( R_{A} \\) and attention predictions \\( P_{A} \\) that contain both data information and feature relationship information:" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.439, + 0.483, + 0.459 + ], + "angle": 0, + "content": "\\[\nR _ {A} = Z D _ {A} ^ {- 1 / 2} G _ {A} D _ {A} ^ {- 1 / 2}; P _ {A} = R _ {A} \\theta^ {h} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.472, + 0.484, + 0.552 + ], + "angle": 0, + "content": "where \\(D_A^{-1/2}G_A D_A^{-1/2}\\) is the normalized graph matrix of graph \\(G_A\\), \\(D_A\\) is the degree matrix of the graph, and \\(\\theta^h\\) are the parameters of the classifier head \\(h\\). Similarly, we can obtain redundancy representations \\(R_R\\) and predictions \\(P_R\\) from redundancy graph \\(G_R\\):" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.564, + 0.483, + 0.584 + ], + "angle": 0, + "content": "\\[\nR _ {R} = Z D _ {R} ^ {- 1 / 2} G _ {R} D _ {R} ^ {- 1 / 2}; P _ {R} = R _ {R} \\theta^ {h} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.484, + 0.78 + ], + "angle": 0, + "content": "Therefore, we decompose the model's original presentation and prediction into two components: the attention part and the redundancy part. This decomposition enables a distinct focus on important relationships while identifying redundant ones. Specifically, \\( R_{A} \\) and \\( P_{A} \\) focus solely on the feature relationships within the attention part, while \\( R_{R} \\) and \\( P_{R} \\) are concerned exclusively with the feature relationships within the redundancy part. Based on this decomposition, our goal is to ensure that the attention part adapts effectively to the target domain. Simultaneously, we aim to separate redundancy part from the attention component in both the representation and prediction layers." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.791, + 0.468, + 0.806 + ], + "angle": 0, + "content": "4.2.3. Representation-Layer Redundancy Elimination" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.9 + ], + "angle": 0, + "content": "With the above representations \\( R_{A} \\) and \\( R_{R} \\), we expect that the attention representations possess class discriminability to handle label shift, while remaining distinct from redundancy representations to eliminate feature redundancy. To achieve this, we leverage contrastive learning [23, 30] as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.115, + 0.915, + 0.188 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {R} = \\\\ - \\sum_ {i = 1} ^ {n _ {t}} \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(R _ {A _ {i}} , c _ {o}\\right)\\right)}{\\sum_ {j = 1} ^ {C} \\exp \\left(\\operatorname {s i m} \\left(R _ {A _ {i}} , c _ {j}\\right)\\right) + \\exp \\left(\\operatorname {s i m} \\left(R _ {A _ {i}} , R _ {R _ {i}}\\right)\\right)} \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.197, + 0.907, + 0.259 + ], + "angle": 0, + "content": "where \\(\\mathrm{sim}(\\cdot ,\\cdot)\\) denotes cosine similarity, \\(\\{c_j\\}_{j = 1}^C\\) denote the class centers in the attention representation space, \\(C\\) is the number of clusters, and \\(c_{o}\\) represents the center of the cluster that contains \\(R_{A_i}\\). \\(c_{j}\\) is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.264, + 0.907, + 0.299 + ], + "angle": 0, + "content": "\\[\nc _ {j} = \\frac {\\sum_ {i} z _ {i} \\cdot \\mathbf {1} [ \\operatorname {a r g m a x} \\left(p _ {i}\\right) = j ]}{\\sum_ {i} \\mathbf {1} [ \\operatorname {a r g m a x} \\left(p _ {i}\\right) = j ]} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.304, + 0.905, + 0.334 + ], + "angle": 0, + "content": "where \\(\\mathbf{1}(\\cdot)\\) is an indicator function, outputting value 1 if the argument is true or 0 otherwise." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.334, + 0.906, + 0.438 + ], + "angle": 0, + "content": "By minimizing \\(\\mathcal{L}_R\\), the model simultaneously enhances the discriminability of attention representations while ensuring their distinction from redundancy representations. This process enables the model to extract robust, discriminative, and non-redundant embedding features in the representation layer, facilitating effective adaptation to the target domain." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.446, + 0.857, + 0.461 + ], + "angle": 0, + "content": "4.2.4. Prediction-Layer Redundancy Elimination" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.464, + 0.907, + 0.525 + ], + "angle": 0, + "content": "In the prediction layer, to enhance the discriminability of attention predictions while reducing the influence of redundant predictions, we combine the principles of entropy minimization and negative learning into a unified loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.531, + 0.911, + 0.585 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {P} = - \\sum_ {i = 1} ^ {N} \\sigma \\left(P _ {A _ {i}}\\right) \\log \\sigma \\left(P _ {A _ {i}}\\right) - \\sum_ {i = 1} ^ {N} \\sigma \\left(P _ {R _ {i}}\\right) \\log \\sigma \\left(1 - P _ {A _ {i}}\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.585, + 0.906, + 0.676 + ], + "angle": 0, + "content": "where \\(\\sigma\\) denotes the softmax operation. Here, the first term minimizes the entropy of attention predictions \\(P_A\\), encouraging sharper and more confident predictions, while the second term applies negative learning to redundant predictions \\(P_R\\), penalizing them to ensure they remain distinct from the attention predictions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.682, + 0.723, + 0.696 + ], + "angle": 0, + "content": "4.2.5. Training Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.7, + 0.906, + 0.73 + ], + "angle": 0, + "content": "By combining Eq. (6) and Eq. (8), our overall loss function for G-FRET can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.629, + 0.746, + 0.905, + 0.761 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {G F R E T} = \\mathcal {L} _ {R} + \\lambda \\mathcal {L} _ {P} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.796 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is the balancing hyperparameter that controls the trade-off between \\(\\mathcal{L}_R\\) and \\(\\mathcal{L}_P\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.9 + ], + "angle": 0, + "content": "During testing, adaptation is performed online: upon receiving instances \\(\\{x_{i}\\}_{i = 1}^{n_{t}}\\) at time \\(t\\), the model is updated with parameters from the previous instances. The model then generates the prediction \\(\\{p_i = g(x_i)\\}_{i = 1}^{n_t}\\) for the new instances and updates using Eq. (9) with a single step of gradient descent. Adaptation continues as long as test data is available." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2124" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.09, + 0.336, + 0.108 + ], + "angle": 0, + "content": "5. Experiments and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.237 + ], + "angle": 0, + "content": "In this section, we conduct extensive experiments to explore the following key questions: (1) Is Test-Time Adaptation (TTA) based on feature redundancy elimination truly effective? (Sec. 5.2) (2) Do the underlying mechanisms of S-FRET and G-FRET align with our theoretical assumptions? (Sec. 5.3) (3) Additionally, we perform ablation studies and evaluate the scalability of our method in more complex scenarios. Sec. 5.4)" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.245, + 0.283, + 0.262 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.268, + 0.484, + 0.479 + ], + "angle": 0, + "content": "We evaluate adaptation performance on two tasks: image domain adaptation and image corruption adaptation. For domain adaptation, we use the PACS [21] and OfficeHome [48] datasets; for corruption adaptation, we use CIFAR-10-C, CIFAR-100-C, and ImageNet-C [13]. To simulate label shifts, we modify CIFAR-100-C to follow a long-tail distribution, denoted as CIFAR-100-C-LT [7]. Comparison methods include BN [36], TENT [49], EATA [28], SAR [29], TSD [51], TIPI [27], and TEA [57], with ResNet-18 and ResNet-50 [12] as backbones. All results are averaged over 5 runs with different seeds, and each method undergoes independent hyperparameter tuning for fair comparison. Implementation details are provided in Sec. 10, and further results and analysis are in Sec. 11." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.487, + 0.455, + 0.503 + ], + "angle": 0, + "content": "5.2. Comparison with State-of-the-art Methods" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.509, + 0.303, + 0.523 + ], + "angle": 0, + "content": "5.2.1. Domain Generalization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.528, + 0.484, + 0.831 + ], + "angle": 0, + "content": "Tab. 1 presents the accuracy comparison of S-FRET and G-FRET with other state-of-the-art TTA methods on the PACS and OfficeHome datasets based on ResNet-18 and ResNet-50 backbones. The table shows that S-FRET, through its simple yet effective optimization, attains competitive performance comparable to most existing TTA methods, while G-FRET achieves superior performance across both datasets and backbone types. Specifically, on the PACS dataset, G-FRET improves upon the baseline non-adaptive source model by \\(6.67\\%\\) and \\(5.45\\%\\) with ResNet-18 and ResNet-50, respectively. For OfficeHome, the improvements are \\(1.80\\%\\) and \\(2.34\\%\\), respectively. Compared with state-of-the-art TTA methods, G-FRET achieves the highest or second-highest accuracy overall, particularly when using the ResNet-50 as backbone, G-FRET consistently achieves the highest accuracy across all domains. In contrast, other methods fail to achieve consistently high accuracy across different domains and backbones. This is caused by G-FRET's ability to consider both feature discriminability and redundancy." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.837, + 0.265, + 0.852 + ], + "angle": 0, + "content": "5.2.2. Image Corruption" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Tab. 2 compares the performance of S-FRET, G-FRET, and other state-of-the-art methods on CIFAR-10-C, CIFAR100-C, and ImageNet-C at corruption level 5. Two set" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.09, + 0.906, + 0.219 + ], + "angle": 0, + "content": "
BackboneMethodPACSAvgOfficeHomeAvg
ACPPSACPRR
ResNet-18Source [12]78.37±0.0077.20±0.0095.03±0.0076.54±0.0081.84±0.0056.45±0.0048.02±0.0071.34±0.0072.23±0.0062.01±0.00
BN [36]81.01±0.0080.67±0.2795.14±0.1573.83±0.1882.66±0.0855.66±0.1749.22±0.2370.77±0.1072.45±0.1962.01±0.12
SAR [29]83.62±0.0081.47±0.0095.14±0.1580.44±0.1585.17±0.0057.30±0.1550.54±0.2570.71±0.1572.75±0.2562.82±0.12
EATA [28]82.06±0.8581.11±1.0395.16±0.2973.39±1.0082.93±0.4656.42±0.4149.75±0.4171.50±0.3572.85±0.4162.63±0.11
TENT [49]83.54±1.7882.26±1.6995.54±0.3181.05±1.7885.60±0.7557.25±0.3250.45±0.2672.18±0.3873.09±0.3463.24±0.13
TSD [51]87.05±0.4786.94±0.6295.98±0.2582.56±1.1286.13±0.4457.80±0.3450.03±0.6971.45±0.4270.92±0.4662.55±0.06
TEA [57]87.04±0.5486.54±0.8596.17±0.3682.17±1.0887.98±0.4857.07±0.8850.47±0.1871.84±0.1872.85±0.1063.06±0.27
TTP [57]84.77±0.1883.48±0.3896.17±0.1884.85±0.2584.85±0.2557.25±0.3250.45±0.2671.45±0.3572.75±0.2562.62±0.19
S-FRET86.25±0.3086.66±0.3795.96±0.2575.90±0.2486.20±0.6656.26±0.2149.86±0.3871.65±0.0372.72±0.1462.62±0.09
G-FRET87.20±0.3587.19±0.2496.28±0.3683.36±1.4388.51±0.3957.49±0.3451.28±0.4273.04±0.2973.40±0.3563.81±0.11
ResNet-50Source [12]83.89±0.0081.02±0.0096.17±0.0078.04±0.0084.78±0.0064.85±0.0052.20±0.0075.04±0.0075.88±0.0067.01±0.00
BN [36]85.19±0.3585.49±0.3096.60±0.1972.25±0.3384.88±0.0863.30±0.2053.02±0.3273.59±0.2375.10±0.2966.25±0.05
SAR [29]85.24±0.3785.49±0.3096.60±0.1976.02±0.8485.84±0.2664.78±0.1756.30±0.5774.83±0.2776.18±0.3068.02±0.19
EATA [28]85.14±0.3185.36±0.2096.54±0.1972.27±0.1184.83±0.1164.25±0.4154.09±0.5374.25±0.3675.72±0.4367.08±0.22
TTP [51]84.77±0.1884.97±0.1996.17±0.2978.53±0.3785.66±0.2563.33±0.3054.99±0.3474.16±0.2975.16±0.2968.74±0.14
TSD [51]80.93±0.8380.94±0.2797.56±0.2781.56±1.5289.97±0.1865.33±0.3056.99±0.3476.12±0.3676.52±0.2468.74±0.14
TEA [57]88.51±0.4387.75±0.5197.17±0.2481.45±1.0288.72±0.2665.62±0.4557.55±0.6475.80±0.3976.85±0.5169.25±0.25
TTP [27]87.62±0.7386.97±0.9697.01±1.1179.66±1.5087.81±0.5664.92±0.2956.47±0.3775.31±0.2776.74±0.3768.36±0.04
S-FRET89.90±0.6089.00±0.4897.46±0.3277.45±0.9188.42±0.1964.42±0.2055.00±0.3475.56±0.2476.41±0.1467.80±0.13
G-FRET90.99±0.8790.00±0.2597.59±0.2982.32±0.5990.23±0.5366.12±0.4157.63±0.3976.46±0.3477.19±0.2369.35±0.20
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.228, + 0.907, + 0.298 + ], + "angle": 0, + "content": "Table 1. Accuracy comparison of different TTA methods on PACS and OfficeHome datasets based on ResNet-18 and ResNet-50 backbones, including mean and standard deviation of 5 runs with different random seeds. The best results are highlighted in boldface, and the second ones are underlined." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.318, + 0.907, + 0.545 + ], + "angle": 0, + "content": "tings are used: continuous adaptation for CIFAR-10/100-C (15 corruption types applied sequentially), and independent adaptation for ImageNet-C (each corruption type evaluated separately). S-FRET significantly outperforms the non-adaptive baseline, with average gains of \\(22.84\\%\\) on CIFAR-10-C, \\(17.52\\%\\) on CIFAR-100-C, and \\(14.4\\%\\) on ImageNet-C. G-FRET consistently surpasses S-FRET by leveraging both feature redundancy and label discriminability, whereas S-FRET only addresses redundancy. Under continuous adaptation, methods not explicitly reducing redundancy perform similarly to S-FRET but worse than G-FRET. This highlights the complementary roles of redundancy reduction and class discriminability in improving TTA performance—a finding further supported by our ablation studies." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.554, + 0.907, + 0.687 + ], + "angle": 0, + "content": "
MethodCIFAR-10-CCIFAR-100-CImageNet-C
1-78-151-151-78-151-151-78-151-15
Source [12]38.4661.5950.8023.5937.4931.017.7420.7914.70
BN [36]70.4776.4373.6545.7050.5748.3017.8436.1227.59
TENT [49]73.0679.2576.3647.7753.4450.7927.0542.6935.39
EATA [28]70.6076.4273.7047.6453.4050.7131.2346.1539.19
SAR [29]70.6176.5073.7546.6651.2149.0930.3745.7038.55
TIPI [27]73.7680.2177.2147.5052.0649.9327.7842.6935.73
TEA [57]71.7077.4774.7846.0750.6448.5122.8540.7232.38
TSD [51]72.5377.9675.4345.8751.0348.6220.9338.1530.11
S-FRET71.6077.2874.6346.3150.4648.5319.5937.4329.10
G-FRET73.7680.3277.2648.2353.5551.0628.6043.9636.79
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.698, + 0.906, + 0.754 + ], + "angle": 0, + "content": "Table 2. Accuracy comparison of different TTA methods on CIFAR-10/100-C and ImageNet-C datasets at damage level of 5, with 15 types of damage applied two experimental settings: continuous adaptation and independent adaptation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.762, + 0.72, + 0.777 + ], + "angle": 0, + "content": "5.2.3. Efficiency Comparison" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.906, + 0.903 + ], + "angle": 0, + "content": "In addition to accuracy, we also evaluate the runtime efficiency of different methods to assess the computational performance of S-FRET and G-FRET. As shown in Tab. 3, \"Training\" indicates whether the method involves retraining. ERM represents the source model without adaptation, while the BN method directly updates the batch normalization (BN) parameters, eliminating the need for retraining. Consequently, BN is the most computationally ef" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2125" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.091, + 0.089, + 0.486, + 0.175 + ], + "angle": 0, + "content": "
Cost Time (s)
BackboneERMBNTENTEATASARTSDTIPITEAS-FRETG-FRET
Traning××✓✓✓✓✓✓✓✓
ResNet-18393648567411416265846118
ResNet-5064661131091851532641376102193
VIT-B/161411622102203452952813181185310
AVG81881241282011872361738111207
" + }, + { + "type": "table", + "bbox": [ + 0.091, + 0.176, + 0.486, + 0.249 + ], + "angle": 0, + "content": "
Maximum GPU memory usage (MB)
Backbone TrainingSource ×BN ×TENT ✓EATA ✓SAR ✓TSD ✓TIPI ✓TEA ✓S-FRET ✓G-FRET ✓
ResNet-1890390427893713354831703427822127894781
ResNet-50772773541161536005635660711595154489163
VIT-B/161218-64657324662010584137651730064698045
AVG965-488857305391670377551382449027330
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.251, + 0.483, + 0.279 + ], + "angle": 0, + "content": "Table 3. Efficiency comparison of TTA methods on ImageNet-C (Gaussian Noise) across multiple backbone networks" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.29, + 0.485, + 0.456 + ], + "angle": 0, + "content": "ficient method. However, as reported in Tabs. 1 and 2, BN consistently exhibits the lowest accuracy. Among the training methods, S-FRET achieves the lowest runtime and the second-lowest memory consumption across various backbones, while G-FRET exhibits average performance compared to other baselines. These results indicate that S-FRET, as a simple yet effective approach, provides a competitive accuracy-efficiency trade-off. Meanwhile, G-FRET, despite its moderate computational cost, achieves the highest accuracy, demonstrating its effectiveness in balancing efficiency and adaptation performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.468, + 0.405, + 0.484 + ], + "angle": 0, + "content": "5.3. Analysis of Underlying Mechanisms" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.49, + 0.483, + 0.521 + ], + "angle": 0, + "content": "5.3.1. Relation between Feature Redundancy Elimination and Generalizability Enhancement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.524, + 0.484, + 0.782 + ], + "angle": 0, + "content": "The first mechanism to be validated is whether there exists a genuine relationship between feature redundancy and generalizability enhancement. We track the evolution of redundancy scores (RS), loss, and accuracy across increasing adaptation steps. As shown in Fig. 4, S-FRET achieves a consistent reduction in feature redundancy as adaptation progresses, accompanied by stable accuracy improvement, demonstrating the detrimental effect of redundancy on generalization and the efficacy of redundancy elimination. For G-FRET, as the number of adaptation steps increases, there is also a consistent reduction in the redundancy of the extracted embedded features, accompanied by a significant decrease in loss and a notable improvement in accuracy. In summary, both S-FRET and G-FRET has been demonstrated to be highly effective in minimizing redundancy, and redundancy elimination is strongly associated with enhancement of model generalizability." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.791, + 0.216, + 0.805 + ], + "angle": 0, + "content": "5.3.2. Label Shift" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.9 + ], + "angle": 0, + "content": "Is S-FRET truly affected by label shift? Tab. 4 presents the performance of S-FRET and G-FRET under test-time label shifts with long-tailed distributions. Here, the imbalance factor denotes the ratio between the most and least frequent class samples—higher values indicate more skewed class distributions. As shown in Tab. 4, S-FRET achieves" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.089, + 0.695, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.089, + 0.887, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.189, + 0.712, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.19, + 0.903, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.298, + 0.907, + 0.339 + ], + "angle": 0, + "content": "Figure 4. The relationship of the reduction of redundancy and the enhancement of generalizability attained by S-FRET and G-FRET on CIFAR-10-C and CIFAR-100-C datasets." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.348, + 0.909, + 0.413 + ], + "angle": 0, + "content": "
CIFAR-100-C-LT
Imbalance FactorERMBNSAREATATENTTSDTIPITEAS-FRETG-FRET
131.0148.3049.0950.7150.7948.6248.5149.9348.5351.06
1030.3947.1447.6947.8650.1247.7447.5049.5447.4750.23
10030.1346.0748.2146.0849.8046.9547.1349.5146.6250.27
AVG30.5147.1748.3348.2150.2447.7747.7149.6647.5450.52
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.424, + 0.906, + 0.451 + ], + "angle": 0, + "content": "Table 4. Performance comparison on CIFAR-100-C-LT with different imbalance factors based on ResNet-18." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.461, + 0.907, + 0.583 + ], + "angle": 0, + "content": "competitive performance on CIFAR-100-C under balanced class distributions. However, its accuracy degrades significantly as the imbalance factor increases, with \\(1.91\\%\\) drops on CIFAR-100-C, aligning with the limitations of conventional TTA methods. In contrast, G-FRET delivers more robust performance across all imbalance factors. This highlights G-FRET's unique capability to address label shift in real-world scenarios through contrastive learning." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.59, + 0.8, + 0.605 + ], + "angle": 0, + "content": "5.3.3. Discriminability Analysis by tSNE" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.609, + 0.907, + 0.791 + ], + "angle": 0, + "content": "Does G-FRET truly improve discriminability compared to S-FRET? Fig. 5 shows the t-SNE [46] visualization of embedded features before and after adaptation by S-FRET and G-FRET. It is evident that before adaptation, data points from different classes are intermixed without clear separation. After adaptation by S-FRET, although class separation increases, the overlapping regions between classes still lack distinct boundaries. In contrast, G-FRET achieves more distinct cluster separation and sharper decision boundaries, demonstrating its superior capability in learning discriminative features under distribution shifts. Extended experimental results on the CIFAR-10-C are provided in Sec. 11.3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.8, + 0.78, + 0.816 + ], + "angle": 0, + "content": "5.4. Ablation and Extension Study" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.822, + 0.802, + 0.837 + ], + "angle": 0, + "content": "5.4.1. Effects of Components in G-FRET" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Tab. 5 presents the impact of the four components of G-FRET on model performance. (1) Compared to the source baseline, each individual component contributes positively to the model's performance, with respective gains of \\(0.30\\%\\)," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2126" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.089, + 0.462, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.203, + 0.483, + 0.245 + ], + "angle": 0, + "content": "Figure 5. Discriminability visualization of embedded feature before and after adaptation by S-FRET and G-FRET on the PACS dataset." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.249, + 0.481, + 0.436 + ], + "angle": 0, + "content": "
OfficeHome
ComponentsACPRAVG↑
Source56.4548.0271.3472.2362.010.00
IndividualLRA55.8349.6770.8372.8962.310.30
R56.9850.4272.3372.7363.121.11
LPA56.7350.7772.1372.9963.161.15
R55.8349.3070.6272.5162.070.06
Ablationw/o LRA57.7251.1872.9973.0863.741.73
R56.7051.1672.0773.1963.281.27
w/o LPA57.1150.5272.4072.9463.241.23
R57.7751.4072.9472.9663.771.76
AttentionLR(A)+LP(A)56.7050.9772.3673.1263.291.28
RedundantLR(R)+LP(R)57.0350.4972.2772.9463.181.17
AllAll57.4951.2873.0474.4063.811.80
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.447, + 0.483, + 0.489 + ], + "angle": 0, + "content": "Table 5. Effects of components in G-FRET, test on the Office-Home dataset with ResNet-18 backbone. A: Attention part, R: Redundancy Part." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.508, + 0.483, + 0.721 + ], + "angle": 0, + "content": "\\(1.11\\%\\), \\(1.15\\%\\), and \\(0.06\\%\\). Notably, the redundancy component in the representation layer, \\(\\mathcal{L}_R(R)\\), and the attention component in the prediction layer, \\(\\mathcal{L}_P(A)\\), show the most significant impact, with \\(\\mathcal{L}_R(R)\\) achieving a \\(1.15\\%\\) improvement. This highlights the importance of redundancy reduction for TTA. (2) The ablation study further confirms the significance of each component, indicating that their removal negatively impacts model performance. (3) Either the redundant or attention components can individually improves model performance, with respective gains of \\(1.28\\%\\) and \\(1.17\\%\\). When the two components are jointly considered, they yield an even greater improvement. (4) Overall, the experimental results further validate the necessity of considering feature redundancy during the TTA process." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.741, + 0.288, + 0.758 + ], + "angle": 0, + "content": "5.4.2. Parameter Sensitivity" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Our method G-FRET involves one hyperparameter: the trade-off parameter in Eq. (9), \\(\\lambda\\). As shown in Fig. 6, G-FRET outperforms the source model across a wide range of values for \\(\\lambda\\). We also investigate the impact of the learning rate (lr) on the experimental results. The best performance is achieved with a learning rate of around \\(1 \\times 10^{-4}\\); higher learning rates tend to destabilize the training process and reduce the effectiveness of model, which is also observed in other TTA methods [27-29, 49, 51, 57]." + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.091, + 0.713, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.091, + 0.865, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.213, + 0.905, + 0.248 + ], + "angle": 0, + "content": "Figure 6. Parameter sensitivity analysis G-FRET on the Office-Home dataset with ResNet-18 backbone in domain Art. Results on complex/large dataset." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.248, + 0.905, + 0.407 + ], + "angle": 0, + "content": "
DatasetsResNet-18[12]+S-FRET+G-FRETResNet-50[12]+S-FRET+G-FRETViT-B/16[6]+S-FRET+G-FRET
VLCS[45]73.6773.9075.7374.9075.4876.4276.4477.1878.56
DomainNet[34]39.1339.2640.2243.5843.6944.2249.9851.2251.50
ImageNet-C[13]14.7029.1036.7918.1534.6542.8439.8360.3662.50
Results with ViT-B/16 as backbone.
DatasetsSourceTENTPLCEATASARTIPITEATSDDeYOS-FRETG-FRET
CIFAR10-C71.4871.4871.4673.6771.4871.4871.4871.4777.8677.9573.17
CIFAR10-C-CO71.4871.4671.4472.5470.3371.4571.6371.6974.6475.4774.02
CIFAR100-C51.7152.7251.7062.4754.2952.3652.3151.7459.5763.4657.65
CIFAR100-C-CO51.7151.7751.7956.2652.5051.7951.9251.9054.0756.6952.14
PACS87.0387.7386.1787.0787.1287.8787.3783.4388.0787.1988.12
OfficeHome76.1176.8676.5076.7476.6676.9876.1575.4976.7876.5577.62
AVG68.2568.6768.1871.4668.7368.6568.4867.6271.8372.8870.45
Results on Multimodal task
DatasetsCLIPTENTTPTCLIPArTTWATT-PWATT-SS-FRET
VLCS80.9680.8572.5380.8981.0581.5981.69
PACS95.1295.2293.0895.3596.0796.0896.09
OfficeHome80.2880.4781.9180.4281.8281.7181.67
AVG85.4585.5182.5185.5586.3186.4686.48
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.419, + 0.905, + 0.446 + ], + "angle": 0, + "content": "Table 6. Scalability on complex/large datasets, different backbones and multimoda task." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.459, + 0.905, + 0.489 + ], + "angle": 0, + "content": "5.4.3. Scalability on complex/large datasets, different backbones and multimoda task." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.495, + 0.907, + 0.678 + ], + "angle": 0, + "content": "We validate our method on large-scale datasets, diverse backbones, and a multimodal task to demonstrate its robustness and architecture-independence. As shown in Tab. 6, reducing embedding redundancy leads to notably larger gains on ViTs than on CNNs, likely due to the higher redundancy in Transformer representations. In the multimodal CLIP setting [31], S-FRET continues to deliver consistent improvements by explicitly aligning embedding distributions, unlike CLIP's similarity-based matching. Notably, on CIFAR with ViT, S-FRET significantly outperforms G-FRET, suggesting that the effectiveness of redundancy reduction varies across backbone types and datasets." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.693, + 0.634, + 0.71 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We propose a novel approach for test-time adaptation (TTA) — Feature Redundancy Elimination for TTA (FRET) — which focuses on reducing feature redundancy during testing to address distribution shifts. This is the first attempt to leverage feature redundancy for TTA. We introduce two methods: S-FRET, which directly minimizes the feature redundancy score, and G-FRET, which optimizes both redundancy elimination and class discriminability by decomposing feature relations into attention and redundancy components. Experiments across various tasks and datasets show the effectiveness of our methods, demonstrating the value of feature redundancy elimination in TTA." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2127" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.09, + 0.482, + 0.108 + ], + "angle": 0, + "content": "Author Contributions and Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.252 + ], + "angle": 0, + "content": "Linjing You initiated this work and was primarily responsible for proposing the core idea, methodology, theoretical analysis, experimental design, and discussion of key findings. Jiabao Lu conducted most of the experiments and contributed to the interpretation of experimental results. Xiayuan Huang participated in experimental discussions and provided feedback on the initial manuscript. Xiangli Nie offered valuable suggestions for the refinement and polishing of the paper." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.252, + 0.483, + 0.312 + ], + "angle": 0, + "content": "We would also like to sincerely thank all the anonymous reviewers and the editorial committee (ACS) whose insightful comments significantly improved the theoretical rigor, experimental depth, and overall quality of this work." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.312, + 0.484, + 0.417 + ], + "angle": 0, + "content": "This research was supported by Prof. Huang's project The Fundamental Research Funds for the Central Universities (Grant No. XJJSKYQD202538), Prof. Nie's project The National Key Research and Development Program of China (Grant No. 2024YFB4709100), and partially funded by the Beijing Nova Program (Grant Nos. 20220484070 and 20240484498)." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.429, + 0.188, + 0.445 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.454, + 0.483, + 0.524 + ], + "angle": 0, + "content": "[1] Ivan Anishchenko, Samuel J Pellock, Tamuka M Chidyausiku, Theresa A Ramelot, Sergey Ovchinnikov, Jingzhou Hao, Khushboo Bafna, Christoffer Norn, Alex Kang, Asim K Bera, et al. De novo protein design by deep network hallucination. Nature, 600(7889):547-552, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.525, + 0.483, + 0.58 + ], + "angle": 0, + "content": "[2] Gavin Brown, Adam Pocock, Ming-Jie Zhao, and Mikel Luján. Conditional likelihood maximisation: a unifying framework for information theoretic feature selection. The journal of machine learning research, 13(1):27-66, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.58, + 0.482, + 0.649 + ], + "angle": 0, + "content": "[3] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label distribution-aware margin loss. In Proc. the Advances in Neural Information Processing Systems (NeurIPS), pages 1567-1578, 2019. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.65, + 0.483, + 0.704 + ], + "angle": 0, + "content": "[4] Xue-wen Chen and Jong Cheol Jeong. Enhanced recursive feature elimination. In Sixth international conference on machine learning and applications (ICMLA 2007), pages 429-435. IEEE, 2007. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.706, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[5] Ian Connick Covert, Wei Qiu, Mingyu Lu, Na Yoon Kim, Nathan J White, and Su-In Lee. Learning to maximize mutual information for dynamic feature selection. In International Conference on Machine Learning, pages 6424-6447. PMLR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.776, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[6] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.818, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[7] C. Du, Y. Wang, S. Song, and G. Huang. Probabilistic contrastive learning for long-tailed visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(9):5890-5904, 2024. 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[8] Tongtong Fang, Nan Lu, Gang Niu, and Masashi Sugiyama. Rethinking importance weighting for deep learning under" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.454, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "distribution shift. Advances in neural information processing systems, 33:11996-12007, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.123, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[9] Ronald A Fisher. The use of multiple measurements in taxonomic problems. Annals of eugenics, 7(2):179-188, 1936. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.22 + ], + "angle": 0, + "content": "[10] Taesik Gong, Yewon Kim, Taeckyung Lee, Sorn Chottananurak, and Sung-Ju Lee. Sotta: Robust test-time adaptation on noisy data streams. Advances in Neural Information Processing Systems, 36, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.222, + 0.905, + 0.277 + ], + "angle": 0, + "content": "[11] Jie Gui, Zhenan Sun, Shuiwang Ji, Dacheng Tao, and Tieniu Tan. Feature selection based on structured sparsity: A comprehensive study. IEEE transactions on neural networks and learning systems, 28(7):1490-1507, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.279, + 0.905, + 0.333 + ], + "angle": 0, + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 2, 6, 8, 3, 7, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.335, + 0.905, + 0.376 + ], + "angle": 0, + "content": "[13] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019. 2, 6, 8, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.379, + 0.905, + 0.433 + ], + "angle": 0, + "content": "[14] Chenping Hou, Ruidong Fan, Ling-Li Zeng, and Dewen Hu. Adaptive feature selection with augmented attributes. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):9306-9324, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.435, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[15] John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Žídek, Anna Potapenko, et al. Highly accurate protein structure prediction with alphafold. nature, 596(7873):583-589, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[16] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. In Proc. the International Conference on Learning Representations (ICLR), 2019. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[17] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.618, + 0.905, + 0.646 + ], + "angle": 0, + "content": "[18] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. nature, 521(7553):436-444, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.648, + 0.905, + 0.717 + ], + "angle": 0, + "content": "[19] Jonghyun Lee, Dahuin Jung, Saehyung Lee, Junsung Park, Juhyeon Shin, Uiwon Hwang, and Sungroh Yoon. Entropy is not enough for test-time adaptation: From the perspective of disentangled factors. In The Twelfth International Conference on Learning Representations, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.719, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[20] Taeckyung Lee, Sorn Chottananurak, Taesik Gong, and Sung-Ju Lee. Aetta: Label-free accuracy estimation for test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28643-28652, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[21] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. In Proceedings of the IEEE international conference on computer vision, pages 5542-5550, 2017. 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[22] Jundong Li, Kewei Cheng, Suhang Wang, Fred Morstatter, Robert P Trevino, Jiliang Tang, and Huan Liu. Feature selection: A data perspective. ACM computing surveys (CSUR), 50(6):1-45, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2128" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[23] Junnan Li, Pan Zhou, Caiming Xiong, and Steven C.H. Hoi. Prototypical contrastive learning of unsupervised representations. In ICLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[24] Jian Liang, Ran He, and Tieniu Tan. A comprehensive survey on test-time adaptation under distribution shifts. International Journal of Computer Vision, pages 1-34, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.482, + 0.276 + ], + "angle": 0, + "content": "[25] Duy MH Nguyen, Hoang Nguyen, Nghiem Diep, Tan Ngoc Pham, Tri Cao, Binh Nguyen, Paul Swoboda, Nhat Ho, Shadi Albarqouni, Pengtao Xie, et al. Lvm-med: Learning large-scale self-supervised vision models for medical imaging via second-order graph matching. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.333 + ], + "angle": 0, + "content": "[26] Wamidh K Mutlag, Shaker K Ali, Zahoor M Aydam, and Bahaa H Taher. Feature extraction methods: a review. In Journal of Physics: Conference Series, page 012028. IOP Publishing, 2020. 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.335, + 0.482, + 0.404 + ], + "angle": 0, + "content": "[27] A Tuan Nguyen, Thanh Nguyen-Tang, Ser-Nam Lim, and Philip HS Torr. Tipi: Test time adaptation with transformation invariance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24162-24171, 2023. 6, 8, 1, 3, 7, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.482, + 0.475 + ], + "angle": 0, + "content": "[28] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In International conference on machine learning, pages 16888-16905. PMLR, 2022. 6, 1, 2, 3, 7, 8, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.482, + 0.546 + ], + "angle": 0, + "content": "[29] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yaofo Chen, Peilin Zhao, and Mingkui Tan. Towards stable test-time adaptation in dynamic wild world. In International Conference on Learning Representations, 2023. 6, 8, 1, 3, 7, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.589 + ], + "angle": 0, + "content": "[30] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.591, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[31] David Osowiechi, Mehrdad Noori, Gustavo Vargas Hakim, Moslem Yazdanpanah, Ali Bahri, Milad Cheraghalikhani, Sahar Dastani, Farzad Beizaee, Ismail Ayed, and Christian Desrosiers. Watt: Weight average test time adaptation of clip. Advances in neural information processing systems, 37: 48015-48044, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[32] Sung Kyu Park, Seunghan Yang, Jaegul Choo, and Sungrack Yun. Label shift adapter for test-time adaptation under covariate and label shifts. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 16375-16385, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.747, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[33] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[34] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1406-1415, 2019. 8, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[35] Sharan Ramjee and Aly El Gamal. Efficient wrapper feature selection using autoencoder and model based elimination. arXiv preprint arXiv:1905.11592, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[36] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. Advances in neural information processing systems, 33:11539-11551, 2020. 3, 6, 1, 7, 8, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[37] Claude Elwood Shannon. A mathematical theory of communication. The Bell system technical journal, 27(3):379-423, 1948. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[38] David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering the game of go with deep neural networks and tree search. nature, 529(7587):484-489, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.333, + 0.905, + 0.402 + ], + "angle": 0, + "content": "[39] Samarth Sinha, Peter Gehler, Francesco Locatello, and Bernt Schiele. Test: Test-time self-training under distribution shift. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2759-2769, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.404, + 0.905, + 0.461 + ], + "angle": 0, + "content": "[40] Yongyi Su, Xun Xu, and Kui Jia. Towards real-world test-time adaptation: Tri-net self-training with balanced normalization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 15126-15135, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.905, + 0.53 + ], + "angle": 0, + "content": "[41] Yongyi Su, Xun Xu, Tianrui Li, and Kui Jia. Revisiting realistic test-time training: Sequential inference and adaptation by anchored clustering regularized self-training. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.532, + 0.905, + 0.6 + ], + "angle": 0, + "content": "[42] Yue Tan, Chen Chen, Weiming Zhuang, Xin Dong, Lingjuan Lyu, and Guodong Long. Is heterogeneity notorious? taming heterogeneity to handle test-time shift in federated learning. Advances in Neural Information Processing Systems, 36, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.602, + 0.905, + 0.658 + ], + "angle": 0, + "content": "[43] Dipti Theng and Kishor K Bhoyar. Feature selection techniques for machine learning: a survey of more than two decades of research. Knowledge and Information Systems, 66(3):1575-1637, 2024. 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[44] Robert Tibshirani. Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society Series B: Statistical Methodology, 58(1):267-288, 1996. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.703, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[45] Antonio Torralba and Alexei A Efros. Unbiased look at dataset bias. In CVPR 2011, pages 1521-1528. IEEE, 2011. 8, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[46] Laurens van der Maaten and Geoffrey E. Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9: 2579-2605, 2008. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[47] B Venkatesh and J Anuradha. A review of feature selection and its methods. Cybernetics and information technologies, 19(1):3-26, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[48] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5018-5027, 2017. 6, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2129" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[49] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726, 2020. 6, 8, 1, 2, 3, 7, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.175 + ], + "angle": 0, + "content": "[50] Mei Wang and Weihong Deng. Deep visual domain adaptation: A survey. Neurocomputing, 312:135-153, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.246 + ], + "angle": 0, + "content": "[51] Shuai Wang, Daoan Zhang, Zipei Yan, Jianguo Zhang, and Rui Li. Feature alignment and uniformity for test time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20050-20060, 2023. 6, 8, 1, 2, 3, 7, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[52] Svante Wold, Kim Esbensen, and Paul Geladi. Principal component analysis. Chemometrics and intelligent laboratory systems, 2(1-3):37-52, 1987. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.359 + ], + "angle": 0, + "content": "[53] Yanan Wu, Zhixiang Chi, Yang Wang, Konstantinos N Plataniotis, and Songhe Feng. Test-time domain adaptation by learning domain-aware batch normalization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 15961-15969, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[54] Kaichao You, Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Universal domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2720-2729, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.419, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[55] Linjing You, Jiabao Lu, and Xiayuan Huang. Test-time correlation alignment. In Forty-second International Conference on Machine Learning, 2025. 3, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[56] Longhui Yuan, Binhui Xie, and Shuang Li. Robust test-time adaptation in dynamic scenarios. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15922-15932, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.518, + 0.482, + 0.585 + ], + "angle": 0, + "content": "[57] Yige Yuan, Bingbing Xu, Liang Hou, Fei Sun, Huawei Shen, and Xueqi Cheng. Tea: Test-time energy adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23901-23911, 2024. 6, 8, 1, 3, 7, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.588, + 0.482, + 0.642 + ], + "angle": 0, + "content": "[58] Han Yue, Jundong Li, and Hongfu Liu. Second-order unsupervised feature selection via knowledge contrastive distillation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3, 4, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.645, + 0.482, + 0.699 + ], + "angle": 0, + "content": "[59] Longbin Zeng, Jiayi Han, Liang Du, and Weiyang Ding. Rethinking precision of pseudo label: Test-time adaptation via complementary learning. Pattern Recognition Letters, 177: 96-102, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.702, + 0.482, + 0.755 + ], + "angle": 0, + "content": "[60] Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. Domain generalization: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4396-4415, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.755 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2130" + } + ] +] \ No newline at end of file diff --git a/2025/FRET_ Feature Redundancy Elimination for Test Time Adaptation/013e8d33-0504-4dd6-a3e9-326514c12e30_origin.pdf b/2025/FRET_ Feature Redundancy Elimination for Test Time Adaptation/013e8d33-0504-4dd6-a3e9-326514c12e30_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..52a1235ac13623f07a726bc94ff315fa2e9cf4ec --- /dev/null +++ b/2025/FRET_ Feature Redundancy Elimination for Test Time Adaptation/013e8d33-0504-4dd6-a3e9-326514c12e30_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4297b1d2d55708d8af3966417d5b76b13d120fc426c34df3105b2d97853eb6af +size 16711305 diff --git a/2025/FRET_ Feature Redundancy Elimination for Test Time Adaptation/full.md b/2025/FRET_ Feature Redundancy Elimination for Test Time Adaptation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5af206e19a86b7961e591e3a5396c9bf7fcb133c --- /dev/null +++ b/2025/FRET_ Feature Redundancy Elimination for Test Time Adaptation/full.md @@ -0,0 +1,365 @@ +# FRET: Feature Redundancy Elimination for Test Time Adaptation + +Linjing You $^{1,*}$ Jiabao Lu $^{1,*}$ Xiayuan Huang $^{2,\dagger}$ Xiangli Nie $^{1,\dagger}$ + +$^{1}$ Institute of Automation, Chinese Academy of Sciences, Beijing, China + +$^{2}$ College of Science, Beijing Forestry University, Beijing, China + +{youlinjing2023,lujiabao2025,xiangliNie}@ia.ac.cn + +huangxiayuan@bjfu.edu.cn + +* Equal contribution, †Corresponding author + +# Abstract + +Test-Time Adaptation (TTA) aims to enhance the generalization of deep learning models when faced with test data that exhibits distribution shifts from the training data. In this context, only a pre-trained model and unlabeled test data are available, making it particularly relevant for privacy-sensitive applications. In practice, we observe that feature redundancy in embeddings tends to increase as domain shifts intensify in TTA. However, existing TTA methods often overlook this redundancy, which can hinder the model's adaptability to new data. To address this issue, we introduce Feature Redundancy Elimination for Test-time Adaptation (FRET), a novel perspective for TTA. A straightforward approach (S-FRET) is to directly minimize the feature redundancy score as an optimization objective to improve adaptation. Despite its simplicity and effectiveness, S-FRET struggles with label shifts, limiting its robustness in real-world scenarios. To mitigate this limitation, we further propose Graph-based FRET (G-FRET), which integrates a Graph Convolutional Network (GCN) with contrastive learning. This design not only reduces feature redundancy but also enhances feature discriminability in both the representation and prediction layers. Extensive experiments across multiple model architectures, tasks, and datasets demonstrate the effectiveness of S-FRET and show that G-FRET achieves state-of-the-art performance. Further analysis reveals that G-FRET enables the model to extract non-redundant and highly discriminative features during inference, thereby facilitating more robust test-time adaptation. The code is available at https://github.com/youlj109/fret. + +# 1. Introduction + +Recently, the emergence of deep neural networks (DNNs) [12, 18] has significantly propelled the advancement in a + +![](images/13e6146631ac6f03fd39d1fa7363a0ba87ab939cb5b67b48fca8e42a497f032f.jpg) +Figure 1. An intuitive demonstration of the relationship between embedded feature redundancy and distribution migration. Top: Second-order feature relation graph of embedding features. Bottom: Plot of feature redundancy versus the level of corruption. + +Vast range of human tasks [1, 15, 38]. However, the basic assumption that training and test data are drawn independently and identically (i.i.d.) may not be suitable in many real-world scenarios where the test/target distribution diverges from the training/source distribution [50]. This phenomenon, known as distribution shift, often leads to poor performance [8]. To tackle this issue, a variety of generalization or adaptation techniques have been developed to improve model robustness against distribution shifts [24, 54, 60]. Among these, Test Time Adaptation (TTA) [10, 40, 41] has been the focus of recent and active research. TTA has the advantage of only requiring access to the pre-trained model from the training/source domain and unlabeled test data. Therefore, TTA is a more secure and practical solution especially when models are publicly available. + +available but the training data and training process remain inaccessible due to privacy and resource restrictions [24]. + +Existing TTA methods can be broadly categorized into batch normalization calibration methods, pseudo-labeling methods, consistency training methods, and clustering-based training methods, based on different adaptation strategies [24]. Batch normalization (BN) calibration methods [53] posit that the statistics in the BN layers represent domain-specific knowledge. To bridge the domain gap, it is suggested to replace the training BN statistics with new statistics estimated over the entire target domain. Pseudolabeling methods [59] employ various filtering strategies, such as thresholding or entropy-based approaches, to obtain reliable pseudo-labels. This process helps minimize discrepancies between predicted labels and pseudo-labels. Consistency training methods [39] aim to improve the stability of network predictions or features by addressing variations in input data, such as noise or perturbations, and changes in model parameters. Clustering-based training methods [19] focus on reducing uncertainty in target network predictions by leveraging clustering techniques to organize target features into distinct groups, thereby enhancing the robustness of the model. + +Despite the empirical success of existing methods, they overlook feature redundancy in embeddings, which hinders generalization to new data [22]. As shown in Fig. 1, we visualize feature redundancy in ResNet-18 [12] using second-order feature relations (i.e., covariance matrix) on the CIFAR10-C dataset [13] under distribution shift [25]. Each image is embedded into a 512-dimensional vector; stacking $n$ embeddings yields an $n \times 512$ matrix. We compute the absolute covariance matrix $(512 \times 512)$ and visualize it as a heatmap — redder areas indicate higher feature correlation, hence greater redundancy. The top two subfigures in Fig. 1 compare the source and target domains, revealing much higher redundancy in the latter due to distribution shift. To quantify this, we remove the diagonal of the covariance matrix and sum the absolute values of the remaining elements (see Eq. (1)). As shown at the bottom of Fig. 1, redundancy increases with corruption severity, suggesting that features with lower redundancy may be better suited for adaptation to the target domain. + +Building on the limitations identified in the existing literature and the observations above, we provide a novel perspective for test-time adaptation—Feature Redundancy Elimination for TTA (FRET)—which differs from existing methods. Due to the inherently differentiable nature of the redundancy metric, a straightforward and computationally efficient approach is to directly minimize the feature redundancy score. We refer to this simple approach as S-FRET. However, since S-FRET lacks the ability to perceive label distributions, it effectively addresses covariate shift but fails to handle label shift—a scenario where the tar + +get domain exhibits a label distribution that deviates from the source domain (e.g., long-tailed data). This discrepancy can significantly degrade model performance and remains a critical challenge [3, 16, 32]. + +To overcome S-FRET's limitation in handling label shift, we further propose G-FRET, which integrates a Graph Convolutional Network (GCN). Unlike FRET, which only minimizes redundancy at the representation layer, G-FRET leverages GCN to incorporate both attention and redundant feature relationships into the representation and prediction layers. Specifically, G-FRET learns non-redundant and discriminative representation through contrastive distillation between attention and redundant representation in the representation-layer. In the prediction layer, attention predictions are optimized through entropy minimization, while ensuring that they are distinct from redundant predictions by negative learning. By maximizing the discriminability of attention relationships while eliminating redundancy in both the representation and prediction layers, G-FRET effectively reduces test-time feature redundancy and improves discrimination when handling complex distribution-shifted test data. + +We summarize the contributions of this paper as follows. + +- Potential Way: provide a novel perspective for test-time adaptation — Feature Redundancy Elimination for TTA (FRET) — which differs from existing methods. It adapts the model to distribution shifts by reducing the redundancy of embedded features during testing. To the best of our knowledge, this is the first attempt to leverage feature redundancy for TTA. +- Innovative Methods: We propose two novel methods, S-FRET and G-FRET, to eliminate feature redundancy in test-time embeddings. S-FRET directly minimizes the feature redundancy score, offering a lightweight solution. G-FRET extends this by decomposing feature relations into attention and redundancy components, enabling joint optimization of redundancy elimination and class discriminability in real-world scenarios. +- Extensive Evaluation: Comprehensive experiments across multiple model architectures, tasks, and datasets demonstrate the effectiveness of S-FRET and show that G-FRET achieves state-of-the-art performance. Further in-depth analysis reveals that G-FRET facilitates the extraction of non-redundant and discriminative embedding features during testing, ultimately enhancing model adaptation to the test-time target domain. + +# 2. Preliminary and Notations + +We briefly revisit Test-Time Adaptation and Feature Redundancy Elimination in this section for the convenience of introducing methods, and put detailed related work discussions into Sec. 7 due to page limits. + +![](images/b212cf9628f98fd8cde427cb2b685556eff20199b436d43bf8a4dfd8a3f9de70.jpg) +Figure 2. Redundancy scores of 15 corruption types in CIFAR-10/100-C and ImageNet-C across 5 corruption levels, as well as for source and target domains in PACS, OfficeHome, VLCS, and DomainNet. + +![](images/6dd2ddf82f35c52220a5a119a5822625c5201de76d2762f2cd4671979169c65b.jpg) + +![](images/1c64f3a491115c11a74b19ffadce1749266f4663facd651a460844a6558aff84.jpg) + +![](images/5feb3184358798be6c31190903e333107c91a28c2b51d1ab20fa19683c01caf8.jpg) + +# 2.1. Test-Time Adaptation + +In the test-time adaptation (TTA) scenario, we have access only to unlabeled data from the target domain in an online manner and a pre-trained model from the source domain. Specifically, let $\{(x_i,y_i)\}_{i = 1}^{n_s}\subset \mathcal{X}_s\times \mathcal{Y}_s$ represent the labeled source domain dataset with labels $\mathcal{V}_s$ and $\{x_j\}_{j = 1}^{n_t}\subset \mathcal{X}_t$ represent a batch of unlabeled target data. The model, trained on the source domain dataset and parameterized by $\theta$ , is denoted as $g_{\theta} = h(f(\cdot)): \mathcal{X}_s\to \mathcal{Y}_s$ , where $f$ is the backbone encoder and $h$ denotes the linear classification head. Our objective is to adapt this model to the target domain $\mathcal{X}_t$ , which exhibits a different distribution from that of the source domain $\mathcal{X}_s$ . During testing, for each instance $x_{j}\in \mathcal{X}_{t}$ , let the output of $f$ and $h$ be denoted as embedding $z_{j} = f(x_{j})\in \mathbb{R}^{d}$ and logs $p_j = h(z_j)\in \mathbb{R}^C$ , respectively, where $d$ is the dimension of the embeddings and $C$ is the number of classes. + +# 2.2. Feature Redundancy Elimination + +Feature redundancy is a key concern in both feature extraction and feature selection [26, 43]. Following previous work [58], we remove the diagonal elements of the covariance matrix and calculate the sum of the absolute values of the remaining elements to derive the redundancy scores $R_{s}$ as: + +$$ +R _ {s} = \left\| \tilde {Z} ^ {T} \tilde {Z} - I _ {d} \right\| _ {1} \tag {1} +$$ + +where $I_{d}$ is the identity matrix, $\| \cdot \| _1$ denotes the 1-norm of a matrix and $\tilde{Z}_{:,j} = \frac{Z_{:,j}}{\|Z_{:,j}\|_2}$ denotes the matrix obtained by normalizing each column. + +# 3. Feature Redundancy Under Domain Shift + +We conduct a systematic investigation into the behavior of feature redundancy under domain shift. + +# 3.1. Empirical Study. + +As shown in Fig. 2, across approximately 50 sub-datasets, the feature redundancy score $R_{s}$ consistently increases with the severity of domain shift. (Note: $R_{s}$ is computed using models with BN parameters updated from test data + +rather than those fixed during training [36].) This empirical trend suggests that domain shift induces stronger correlations among features. To the best of our knowledge, this phenomenon has not been systematically explored in prior work. + +# 3.2. Theoretical Explanation. + +Intuitively, higher covariance elements indicate that the features fail to span the full vector space. Furthermore, we provide the following theoretical result. Missing proofs and detailed explanations are provided in Sec. 9 + +Theorem 1 (Redundancy Elimination Improves Generalization). Under the empirical observation that embeddings from the source domain exhibit lower feature redundancy, i.e., $\mathcal{R}_s(Z_t) > \mathcal{R}_s(Z_s)$ , and by leveraging Theorem 2 from [55], we can derive the following generalization error bound: Let $\mathcal{H}$ be a hypothesis class of VC-dimension $d_v$ . If $\hat{h} \in \mathcal{H}$ minimizes the empirical error $\hat{\epsilon}_s(h)$ on $\mathcal{X}_s$ and $h_t^* = \arg \min_{h \in \mathcal{H}} \epsilon_t(h)$ is the optimal hypothesis on $\mathcal{X}_t$ , with the assumption that all hypotheses are L-Lipschitz continuous, then $\forall \delta \in (0,1)$ , with probability with at least $1 - \delta$ the following inequality holds: + +$$ +\begin{array}{l} \epsilon_ {t} (\hat {h}) - \epsilon_ {t} (h _ {t} ^ {*}) \\ \leq \mathcal {O} (\sqrt {\| \mu_ {s} - \mu_ {t} \| _ {F} ^ {2} + R _ {s} (Z _ {s}) ^ {2} + R _ {s} (Z _ {t}) ^ {2}}) + C \\ \end{array} +$$ + +where $C = 2\sqrt{\frac{d_v\log(2n_s) - \log(\delta)}{2n_s}} + 2\gamma$ and $\gamma = \min_{h\in \mathcal{H}}\{\epsilon_s(h(t)) + \epsilon_t(h(t))\}$ . $\mu_{s}$ and $\mu_t$ denote the means of the source and test embeddings, respectively. We use $\mathcal{O}(\cdot)$ to hide the constant dependence. Since $\mathcal{R}_s(Z_t) > \mathcal{R}_s(Z_s)$ , this implies that reducing $\mathcal{R}_s(Z_t)$ can effectively decrease the generalization error. + +# 4. The Proposed Approach + +# 4.1. S-FRET + +To eliminate feature redundancy in the target domain, a simple yet effective approach is to directly minimize the feature redundancy score $R_{s}$ , as defined in Eq. (1), leading to the proposed method S-FRET (Simple and Straightforward FRET). The loss function for S-FRET can be written as: + +![](images/b7081ca10dda8bd484b3ce6ea5bcf124602d584b14194d34e556de7e34681a36.jpg) +Figure 3. The pipeline of our proposed Graph Convolutional Network based Feature Redundancy Elimination for Test Time Adaptation (G-FRET) method. For each test sample, G-FRET learns non-redundant and discriminative representation through contrastive distillation in the representation-layer. Additionally, in the prediction layer, attention predictions are optimized through entropy minimization, while ensuring that they are distinct from redundant predictions by using negative learning. + +$$ +\mathcal {L} _ {S F R E T} = \left\| \tilde {Z} ^ {T} \tilde {Z} - I _ {d} \right\| _ {1} \tag {2} +$$ + +This method directly minimizes the redundancy score derived from feature correlation, providing a computationally efficient solution that avoids complex architectural modifications. By focusing on the feature space of representation-layer, S-FRET enables the model to extract non-redundant representations, thereby enhancing its adaptability to the target domain (see Fig. 4). + +However, since the redundancy score does not incorporate information about class distributions, S-FRET primarily addresses covariate shift but lacks the discriminative capability to handle label shift—a critical challenge where the target domain exhibits a label distribution that deviates significantly from the source domain (e.g., long-tailed data), leading to substantial performance degradation (see Tab. 4) [3, 16, 32]. + +# 4.2. G-FRET + +# 4.2.1. Overview of G-FRET + +To eliminate feature redundancy and address label shift in the target domain, we further propose a novel two-layer TTA method, namely Graph Convolutional Network based Feature Redundancy Elimination for Test Time Adaptation (G-FRET), which leverages GCN to simultaneously model feature relationships and sample-specific discriminative information in both representation and prediction layers. Fig. 3 provides an overview of our G-FRET framework. + +The key idea of G-FRET is to adapt the pre-tarined model during testing by extracting discriminative and nonredundant embedding features from both the representation and prediction layers. Firstly (Sec. 4.2.2), we apply a mask matrix to the second-order feature relation graph. + +This process yields two separate graphs: an attention graph, which highlights significant feature relationships, and a redundant graph, which identifies redundant feature relationships. Next, we derive the attention/redundancy representation and attention/redundancy prediction by utilizing a single-layer GCN to integrate the model's embeddings with the attention and redundant graphs. This process allows us to explicitly separate attention and redundancy relations in both representation and prediction layer. Then (Sec. 4.2.3), in the representation layer, contrastive learning techniques are employed to enhance the discriminability of the attention representation while ensuring it remains distant from the redundant representation. At last (Sec. 4.2.4), in the prediction layer, attention predictions are optimized through entropy minimization, while ensuring that they are distinct from redundant predictions by negative learning. Through this two-layer process, we achieve discriminative and nonredundant embedding features that effectively adapt to the target domain. + +# 4.2.2. Attention Part and Redundancy Part + +Given a batch of $n_t$ data instances from the target domain, we obtain the first-order $n_t \times d$ embedding feature matrix $Z \in \mathbb{R}^{n_t \times d}$ and the second-order feature relation graph $G_F = Z^T Z \in \mathbb{R}^{d \times d}$ . Following the state-of-the-art feature selection method SOFT [58] (see Appendix 7.3 for more details), we apply a symmetric mask matrix $M_M \in \mathbb{R}^{d \times d}$ to $G_F$ . This decomposition of $G_F$ yields two distinct graphs: the attention graph $G_A$ and the redundancy graph $G_R$ . $G_A$ captures the attention relations (i.e., important relation) among features, while $G_R$ represents the redundant relation between them. Formally, they are calculated as follows: + +$$ +G _ {A} = G _ {F} \odot M _ {M}; G _ {R} = G _ {F} - G _ {A} \tag {3} +$$ + +where $\odot$ is the element-wise product. However, differently from SOFT, which focuses on feature selection by deriving an optimal mask matrix $M_M$ , our objective is to extract non-redundant features $Z$ . For ideally non-redundant features, the second-order feature relation graph should be a diagonal matrix, indicating that each feature is independent of all others. Therefore, we set $M_M = I_d$ to ensure that $G_A$ captures only the essential feature relations, while $G_R$ includes all remaining relations as redundant relations. Here, $I_d \in \mathbb{R}^{d \times d}$ is the identity matrix. In addition, in some scenarios with prior knowledge, such as knowing which features may be more important, or needing to maintain special structural relations between features, some adaptations can be made by simply modifying the mask matrix $M_M$ . + +Then, in order to obtain attention/redundancy representation and attention/redundancy prediction based on the attention and redundancy graphs, we combine the normalized graph matrix with the classifier head $h$ to form a one-layer Graph Convolutional Network (GCN [17]), which can help generate attention representations $R_{A}$ and attention predictions $P_{A}$ that contain both data information and feature relationship information: + +$$ +R _ {A} = Z D _ {A} ^ {- 1 / 2} G _ {A} D _ {A} ^ {- 1 / 2}; P _ {A} = R _ {A} \theta^ {h} \tag {4} +$$ + +where $D_A^{-1/2}G_A D_A^{-1/2}$ is the normalized graph matrix of graph $G_A$ , $D_A$ is the degree matrix of the graph, and $\theta^h$ are the parameters of the classifier head $h$ . Similarly, we can obtain redundancy representations $R_R$ and predictions $P_R$ from redundancy graph $G_R$ : + +$$ +R _ {R} = Z D _ {R} ^ {- 1 / 2} G _ {R} D _ {R} ^ {- 1 / 2}; P _ {R} = R _ {R} \theta^ {h} \tag {5} +$$ + +Therefore, we decompose the model's original presentation and prediction into two components: the attention part and the redundancy part. This decomposition enables a distinct focus on important relationships while identifying redundant ones. Specifically, $R_{A}$ and $P_{A}$ focus solely on the feature relationships within the attention part, while $R_{R}$ and $P_{R}$ are concerned exclusively with the feature relationships within the redundancy part. Based on this decomposition, our goal is to ensure that the attention part adapts effectively to the target domain. Simultaneously, we aim to separate redundancy part from the attention component in both the representation and prediction layers. + +# 4.2.3. Representation-Layer Redundancy Elimination + +With the above representations $R_{A}$ and $R_{R}$ , we expect that the attention representations possess class discriminability to handle label shift, while remaining distinct from redundancy representations to eliminate feature redundancy. To achieve this, we leverage contrastive learning [23, 30] as follows: + +$$ +\begin{array}{l} \mathcal {L} _ {R} = \\ - \sum_ {i = 1} ^ {n _ {t}} \log \frac {\exp \left(\operatorname {s i m} \left(R _ {A _ {i}} , c _ {o}\right)\right)}{\sum_ {j = 1} ^ {C} \exp \left(\operatorname {s i m} \left(R _ {A _ {i}} , c _ {j}\right)\right) + \exp \left(\operatorname {s i m} \left(R _ {A _ {i}} , R _ {R _ {i}}\right)\right)} \tag {6} \\ \end{array} +$$ + +where $\mathrm{sim}(\cdot ,\cdot)$ denotes cosine similarity, $\{c_j\}_{j = 1}^C$ denote the class centers in the attention representation space, $C$ is the number of clusters, and $c_{o}$ represents the center of the cluster that contains $R_{A_i}$ . $c_{j}$ is calculated as follows: + +$$ +c _ {j} = \frac {\sum_ {i} z _ {i} \cdot \mathbf {1} [ \operatorname {a r g m a x} \left(p _ {i}\right) = j ]}{\sum_ {i} \mathbf {1} [ \operatorname {a r g m a x} \left(p _ {i}\right) = j ]} \tag {7} +$$ + +where $\mathbf{1}(\cdot)$ is an indicator function, outputting value 1 if the argument is true or 0 otherwise. + +By minimizing $\mathcal{L}_R$ , the model simultaneously enhances the discriminability of attention representations while ensuring their distinction from redundancy representations. This process enables the model to extract robust, discriminative, and non-redundant embedding features in the representation layer, facilitating effective adaptation to the target domain. + +# 4.2.4. Prediction-Layer Redundancy Elimination + +In the prediction layer, to enhance the discriminability of attention predictions while reducing the influence of redundant predictions, we combine the principles of entropy minimization and negative learning into a unified loss function: + +$$ +\mathcal {L} _ {P} = - \sum_ {i = 1} ^ {N} \sigma \left(P _ {A _ {i}}\right) \log \sigma \left(P _ {A _ {i}}\right) - \sum_ {i = 1} ^ {N} \sigma \left(P _ {R _ {i}}\right) \log \sigma \left(1 - P _ {A _ {i}}\right) \tag {8} +$$ + +where $\sigma$ denotes the softmax operation. Here, the first term minimizes the entropy of attention predictions $P_A$ , encouraging sharper and more confident predictions, while the second term applies negative learning to redundant predictions $P_R$ , penalizing them to ensure they remain distinct from the attention predictions. + +# 4.2.5. Training Loss Function + +By combining Eq. (6) and Eq. (8), our overall loss function for G-FRET can be written as: + +$$ +\mathcal {L} _ {G F R E T} = \mathcal {L} _ {R} + \lambda \mathcal {L} _ {P} \tag {9} +$$ + +where $\lambda$ is the balancing hyperparameter that controls the trade-off between $\mathcal{L}_R$ and $\mathcal{L}_P$ . + +During testing, adaptation is performed online: upon receiving instances $\{x_{i}\}_{i = 1}^{n_{t}}$ at time $t$ , the model is updated with parameters from the previous instances. The model then generates the prediction $\{p_i = g(x_i)\}_{i = 1}^{n_t}$ for the new instances and updates using Eq. (9) with a single step of gradient descent. Adaptation continues as long as test data is available. + +# 5. Experiments and Analysis + +In this section, we conduct extensive experiments to explore the following key questions: (1) Is Test-Time Adaptation (TTA) based on feature redundancy elimination truly effective? (Sec. 5.2) (2) Do the underlying mechanisms of S-FRET and G-FRET align with our theoretical assumptions? (Sec. 5.3) (3) Additionally, we perform ablation studies and evaluate the scalability of our method in more complex scenarios. Sec. 5.4) + +# 5.1. Experimental Setup + +We evaluate adaptation performance on two tasks: image domain adaptation and image corruption adaptation. For domain adaptation, we use the PACS [21] and OfficeHome [48] datasets; for corruption adaptation, we use CIFAR-10-C, CIFAR-100-C, and ImageNet-C [13]. To simulate label shifts, we modify CIFAR-100-C to follow a long-tail distribution, denoted as CIFAR-100-C-LT [7]. Comparison methods include BN [36], TENT [49], EATA [28], SAR [29], TSD [51], TIPI [27], and TEA [57], with ResNet-18 and ResNet-50 [12] as backbones. All results are averaged over 5 runs with different seeds, and each method undergoes independent hyperparameter tuning for fair comparison. Implementation details are provided in Sec. 10, and further results and analysis are in Sec. 11. + +# 5.2. Comparison with State-of-the-art Methods + +# 5.2.1. Domain Generalization + +Tab. 1 presents the accuracy comparison of S-FRET and G-FRET with other state-of-the-art TTA methods on the PACS and OfficeHome datasets based on ResNet-18 and ResNet-50 backbones. The table shows that S-FRET, through its simple yet effective optimization, attains competitive performance comparable to most existing TTA methods, while G-FRET achieves superior performance across both datasets and backbone types. Specifically, on the PACS dataset, G-FRET improves upon the baseline non-adaptive source model by $6.67\%$ and $5.45\%$ with ResNet-18 and ResNet-50, respectively. For OfficeHome, the improvements are $1.80\%$ and $2.34\%$ , respectively. Compared with state-of-the-art TTA methods, G-FRET achieves the highest or second-highest accuracy overall, particularly when using the ResNet-50 as backbone, G-FRET consistently achieves the highest accuracy across all domains. In contrast, other methods fail to achieve consistently high accuracy across different domains and backbones. This is caused by G-FRET's ability to consider both feature discriminability and redundancy. + +# 5.2.2. Image Corruption + +Tab. 2 compares the performance of S-FRET, G-FRET, and other state-of-the-art methods on CIFAR-10-C, CIFAR100-C, and ImageNet-C at corruption level 5. Two set + +
BackboneMethodPACSAvgOfficeHomeAvg
ACPPSACPRR
ResNet-18Source [12]78.37±0.0077.20±0.0095.03±0.0076.54±0.0081.84±0.0056.45±0.0048.02±0.0071.34±0.0072.23±0.0062.01±0.00
BN [36]81.01±0.0080.67±0.2795.14±0.1573.83±0.1882.66±0.0855.66±0.1749.22±0.2370.77±0.1072.45±0.1962.01±0.12
SAR [29]83.62±0.0081.47±0.0095.14±0.1580.44±0.1585.17±0.0057.30±0.1550.54±0.2570.71±0.1572.75±0.2562.82±0.12
EATA [28]82.06±0.8581.11±1.0395.16±0.2973.39±1.0082.93±0.4656.42±0.4149.75±0.4171.50±0.3572.85±0.4162.63±0.11
TENT [49]83.54±1.7882.26±1.6995.54±0.3181.05±1.7885.60±0.7557.25±0.3250.45±0.2672.18±0.3873.09±0.3463.24±0.13
TSD [51]87.05±0.4786.94±0.6295.98±0.2582.56±1.1286.13±0.4457.80±0.3450.03±0.6971.45±0.4270.92±0.4662.55±0.06
TEA [57]87.04±0.5486.54±0.8596.17±0.3682.17±1.0887.98±0.4857.07±0.8850.47±0.1871.84±0.1872.85±0.1063.06±0.27
TTP [57]84.77±0.1883.48±0.3896.17±0.1884.85±0.2584.85±0.2557.25±0.3250.45±0.2671.45±0.3572.75±0.2562.62±0.19
S-FRET86.25±0.3086.66±0.3795.96±0.2575.90±0.2486.20±0.6656.26±0.2149.86±0.3871.65±0.0372.72±0.1462.62±0.09
G-FRET87.20±0.3587.19±0.2496.28±0.3683.36±1.4388.51±0.3957.49±0.3451.28±0.4273.04±0.2973.40±0.3563.81±0.11
ResNet-50Source [12]83.89±0.0081.02±0.0096.17±0.0078.04±0.0084.78±0.0064.85±0.0052.20±0.0075.04±0.0075.88±0.0067.01±0.00
BN [36]85.19±0.3585.49±0.3096.60±0.1972.25±0.3384.88±0.0863.30±0.2053.02±0.3273.59±0.2375.10±0.2966.25±0.05
SAR [29]85.24±0.3785.49±0.3096.60±0.1976.02±0.8485.84±0.2664.78±0.1756.30±0.5774.83±0.2776.18±0.3068.02±0.19
EATA [28]85.14±0.3185.36±0.2096.54±0.1972.27±0.1184.83±0.1164.25±0.4154.09±0.5374.25±0.3675.72±0.4367.08±0.22
TTP [51]84.77±0.1884.97±0.1996.17±0.2978.53±0.3785.66±0.2563.33±0.3054.99±0.3474.16±0.2975.16±0.2968.74±0.14
TSD [51]80.93±0.8380.94±0.2797.56±0.2781.56±1.5289.97±0.1865.33±0.3056.99±0.3476.12±0.3676.52±0.2468.74±0.14
TEA [57]88.51±0.4387.75±0.5197.17±0.2481.45±1.0288.72±0.2665.62±0.4557.55±0.6475.80±0.3976.85±0.5169.25±0.25
TTP [27]87.62±0.7386.97±0.9697.01±1.1179.66±1.5087.81±0.5664.92±0.2956.47±0.3775.31±0.2776.74±0.3768.36±0.04
S-FRET89.90±0.6089.00±0.4897.46±0.3277.45±0.9188.42±0.1964.42±0.2055.00±0.3475.56±0.2476.41±0.1467.80±0.13
G-FRET90.99±0.8790.00±0.2597.59±0.2982.32±0.5990.23±0.5366.12±0.4157.63±0.3976.46±0.3477.19±0.2369.35±0.20
+ +tings are used: continuous adaptation for CIFAR-10/100-C (15 corruption types applied sequentially), and independent adaptation for ImageNet-C (each corruption type evaluated separately). S-FRET significantly outperforms the non-adaptive baseline, with average gains of $22.84\%$ on CIFAR-10-C, $17.52\%$ on CIFAR-100-C, and $14.4\%$ on ImageNet-C. G-FRET consistently surpasses S-FRET by leveraging both feature redundancy and label discriminability, whereas S-FRET only addresses redundancy. Under continuous adaptation, methods not explicitly reducing redundancy perform similarly to S-FRET but worse than G-FRET. This highlights the complementary roles of redundancy reduction and class discriminability in improving TTA performance—a finding further supported by our ablation studies. + +Table 1. Accuracy comparison of different TTA methods on PACS and OfficeHome datasets based on ResNet-18 and ResNet-50 backbones, including mean and standard deviation of 5 runs with different random seeds. The best results are highlighted in boldface, and the second ones are underlined. + +
MethodCIFAR-10-CCIFAR-100-CImageNet-C
1-78-151-151-78-151-151-78-151-15
Source [12]38.4661.5950.8023.5937.4931.017.7420.7914.70
BN [36]70.4776.4373.6545.7050.5748.3017.8436.1227.59
TENT [49]73.0679.2576.3647.7753.4450.7927.0542.6935.39
EATA [28]70.6076.4273.7047.6453.4050.7131.2346.1539.19
SAR [29]70.6176.5073.7546.6651.2149.0930.3745.7038.55
TIPI [27]73.7680.2177.2147.5052.0649.9327.7842.6935.73
TEA [57]71.7077.4774.7846.0750.6448.5122.8540.7232.38
TSD [51]72.5377.9675.4345.8751.0348.6220.9338.1530.11
S-FRET71.6077.2874.6346.3150.4648.5319.5937.4329.10
G-FRET73.7680.3277.2648.2353.5551.0628.6043.9636.79
+ +Table 2. Accuracy comparison of different TTA methods on CIFAR-10/100-C and ImageNet-C datasets at damage level of 5, with 15 types of damage applied two experimental settings: continuous adaptation and independent adaptation. + +# 5.2.3. Efficiency Comparison + +In addition to accuracy, we also evaluate the runtime efficiency of different methods to assess the computational performance of S-FRET and G-FRET. As shown in Tab. 3, "Training" indicates whether the method involves retraining. ERM represents the source model without adaptation, while the BN method directly updates the batch normalization (BN) parameters, eliminating the need for retraining. Consequently, BN is the most computationally ef + +
Cost Time (s)
BackboneERMBNTENTEATASARTSDTIPITEAS-FRETG-FRET
Traning××✓✓✓✓✓✓✓✓
ResNet-18393648567411416265846118
ResNet-5064661131091851532641376102193
VIT-B/161411622102203452952813181185310
AVG81881241282011872361738111207
+ +
Maximum GPU memory usage (MB)
Backbone TrainingSource ×BN ×TENT ✓EATA ✓SAR ✓TSD ✓TIPI ✓TEA ✓S-FRET ✓G-FRET ✓
ResNet-1890390427893713354831703427822127894781
ResNet-50772773541161536005635660711595154489163
VIT-B/161218-64657324662010584137651730064698045
AVG965-488857305391670377551382449027330
+ +ficient method. However, as reported in Tabs. 1 and 2, BN consistently exhibits the lowest accuracy. Among the training methods, S-FRET achieves the lowest runtime and the second-lowest memory consumption across various backbones, while G-FRET exhibits average performance compared to other baselines. These results indicate that S-FRET, as a simple yet effective approach, provides a competitive accuracy-efficiency trade-off. Meanwhile, G-FRET, despite its moderate computational cost, achieves the highest accuracy, demonstrating its effectiveness in balancing efficiency and adaptation performance. + +# 5.3. Analysis of Underlying Mechanisms + +# 5.3.1. Relation between Feature Redundancy Elimination and Generalizability Enhancement + +The first mechanism to be validated is whether there exists a genuine relationship between feature redundancy and generalizability enhancement. We track the evolution of redundancy scores (RS), loss, and accuracy across increasing adaptation steps. As shown in Fig. 4, S-FRET achieves a consistent reduction in feature redundancy as adaptation progresses, accompanied by stable accuracy improvement, demonstrating the detrimental effect of redundancy on generalization and the efficacy of redundancy elimination. For G-FRET, as the number of adaptation steps increases, there is also a consistent reduction in the redundancy of the extracted embedded features, accompanied by a significant decrease in loss and a notable improvement in accuracy. In summary, both S-FRET and G-FRET has been demonstrated to be highly effective in minimizing redundancy, and redundancy elimination is strongly associated with enhancement of model generalizability. + +# 5.3.2. Label Shift + +Is S-FRET truly affected by label shift? Tab. 4 presents the performance of S-FRET and G-FRET under test-time label shifts with long-tailed distributions. Here, the imbalance factor denotes the ratio between the most and least frequent class samples—higher values indicate more skewed class distributions. As shown in Tab. 4, S-FRET achieves + +![](images/6c6b5265d711d6457d126007c110dfeff8d1862945693eed37960cf10550ff8c.jpg) + +![](images/c23fa039f6d590527edb54ccc5d1d6eaf1d786be7889ed8aef9e563eea240d89.jpg) + +![](images/650d8b70e82159da38af4a0948d2ae08e21f8e9ca1d932842fdc69326e495eaf.jpg) +Figure 4. The relationship of the reduction of redundancy and the enhancement of generalizability attained by S-FRET and G-FRET on CIFAR-10-C and CIFAR-100-C datasets. + +![](images/34eb8cf712c46e6b4bac357f0b5afa0656e0a287feaef868635eee2574bfc985.jpg) + +Table 3. Efficiency comparison of TTA methods on ImageNet-C (Gaussian Noise) across multiple backbone networks + +
CIFAR-100-C-LT
Imbalance FactorERMBNSAREATATENTTSDTIPITEAS-FRETG-FRET
131.0148.3049.0950.7150.7948.6248.5149.9348.5351.06
1030.3947.1447.6947.8650.1247.7447.5049.5447.4750.23
10030.1346.0748.2146.0849.8046.9547.1349.5146.6250.27
AVG30.5147.1748.3348.2150.2447.7747.7149.6647.5450.52
+ +Table 4. Performance comparison on CIFAR-100-C-LT with different imbalance factors based on ResNet-18. + +competitive performance on CIFAR-100-C under balanced class distributions. However, its accuracy degrades significantly as the imbalance factor increases, with $1.91\%$ drops on CIFAR-100-C, aligning with the limitations of conventional TTA methods. In contrast, G-FRET delivers more robust performance across all imbalance factors. This highlights G-FRET's unique capability to address label shift in real-world scenarios through contrastive learning. + +# 5.3.3. Discriminability Analysis by tSNE + +Does G-FRET truly improve discriminability compared to S-FRET? Fig. 5 shows the t-SNE [46] visualization of embedded features before and after adaptation by S-FRET and G-FRET. It is evident that before adaptation, data points from different classes are intermixed without clear separation. After adaptation by S-FRET, although class separation increases, the overlapping regions between classes still lack distinct boundaries. In contrast, G-FRET achieves more distinct cluster separation and sharper decision boundaries, demonstrating its superior capability in learning discriminative features under distribution shifts. Extended experimental results on the CIFAR-10-C are provided in Sec. 11.3. + +# 5.4. Ablation and Extension Study + +# 5.4.1. Effects of Components in G-FRET + +Tab. 5 presents the impact of the four components of G-FRET on model performance. (1) Compared to the source baseline, each individual component contributes positively to the model's performance, with respective gains of $0.30\%$ , + +![](images/0e5f608ff42ee0fd764a83fc8a4fc3bcd272cafcca40fa70c73cc7306cbb583f.jpg) +Figure 5. Discriminability visualization of embedded feature before and after adaptation by S-FRET and G-FRET on the PACS dataset. + +
OfficeHome
ComponentsACPRAVG↑
Source56.4548.0271.3472.2362.010.00
IndividualLRA55.8349.6770.8372.8962.310.30
R56.9850.4272.3372.7363.121.11
LPA56.7350.7772.1372.9963.161.15
R55.8349.3070.6272.5162.070.06
Ablationw/o LRA57.7251.1872.9973.0863.741.73
R56.7051.1672.0773.1963.281.27
w/o LPA57.1150.5272.4072.9463.241.23
R57.7751.4072.9472.9663.771.76
AttentionLR(A)+LP(A)56.7050.9772.3673.1263.291.28
RedundantLR(R)+LP(R)57.0350.4972.2772.9463.181.17
AllAll57.4951.2873.0474.4063.811.80
+ +$1.11\%$ , $1.15\%$ , and $0.06\%$ . Notably, the redundancy component in the representation layer, $\mathcal{L}_R(R)$ , and the attention component in the prediction layer, $\mathcal{L}_P(A)$ , show the most significant impact, with $\mathcal{L}_R(R)$ achieving a $1.15\%$ improvement. This highlights the importance of redundancy reduction for TTA. (2) The ablation study further confirms the significance of each component, indicating that their removal negatively impacts model performance. (3) Either the redundant or attention components can individually improves model performance, with respective gains of $1.28\%$ and $1.17\%$ . When the two components are jointly considered, they yield an even greater improvement. (4) Overall, the experimental results further validate the necessity of considering feature redundancy during the TTA process. + +# 5.4.2. Parameter Sensitivity + +Our method G-FRET involves one hyperparameter: the trade-off parameter in Eq. (9), $\lambda$ . As shown in Fig. 6, G-FRET outperforms the source model across a wide range of values for $\lambda$ . We also investigate the impact of the learning rate (lr) on the experimental results. The best performance is achieved with a learning rate of around $1 \times 10^{-4}$ ; higher learning rates tend to destabilize the training process and reduce the effectiveness of model, which is also observed in other TTA methods [27-29, 49, 51, 57]. + +![](images/f944b598b12597c7347c86bc326038e3545a79d13d558ff2d9debf695301f638.jpg) +Figure 6. Parameter sensitivity analysis G-FRET on the Office-Home dataset with ResNet-18 backbone in domain Art. Results on complex/large dataset. + +![](images/682e904363ec7b741ee93f836a6f1258a5f181409bfd8774db041464b98c81ae.jpg) + +Table 5. Effects of components in G-FRET, test on the Office-Home dataset with ResNet-18 backbone. A: Attention part, R: Redundancy Part. + +
DatasetsResNet-18[12]+S-FRET+G-FRETResNet-50[12]+S-FRET+G-FRETViT-B/16[6]+S-FRET+G-FRET
VLCS[45]73.6773.9075.7374.9075.4876.4276.4477.1878.56
DomainNet[34]39.1339.2640.2243.5843.6944.2249.9851.2251.50
ImageNet-C[13]14.7029.1036.7918.1534.6542.8439.8360.3662.50
Results with ViT-B/16 as backbone.
DatasetsSourceTENTPLCEATASARTIPITEATSDDeYOS-FRETG-FRET
CIFAR10-C71.4871.4871.4673.6771.4871.4871.4871.4777.8677.9573.17
CIFAR10-C-CO71.4871.4671.4472.5470.3371.4571.6371.6974.6475.4774.02
CIFAR100-C51.7152.7251.7062.4754.2952.3652.3151.7459.5763.4657.65
CIFAR100-C-CO51.7151.7751.7956.2652.5051.7951.9251.9054.0756.6952.14
PACS87.0387.7386.1787.0787.1287.8787.3783.4388.0787.1988.12
OfficeHome76.1176.8676.5076.7476.6676.9876.1575.4976.7876.5577.62
AVG68.2568.6768.1871.4668.7368.6568.4867.6271.8372.8870.45
Results on Multimodal task
DatasetsCLIPTENTTPTCLIPArTTWATT-PWATT-SS-FRET
VLCS80.9680.8572.5380.8981.0581.5981.69
PACS95.1295.2293.0895.3596.0796.0896.09
OfficeHome80.2880.4781.9180.4281.8281.7181.67
AVG85.4585.5182.5185.5586.3186.4686.48
+ +Table 6. Scalability on complex/large datasets, different backbones and multimoda task. + +# 5.4.3. Scalability on complex/large datasets, different backbones and multimoda task. + +We validate our method on large-scale datasets, diverse backbones, and a multimodal task to demonstrate its robustness and architecture-independence. As shown in Tab. 6, reducing embedding redundancy leads to notably larger gains on ViTs than on CNNs, likely due to the higher redundancy in Transformer representations. In the multimodal CLIP setting [31], S-FRET continues to deliver consistent improvements by explicitly aligning embedding distributions, unlike CLIP's similarity-based matching. Notably, on CIFAR with ViT, S-FRET significantly outperforms G-FRET, suggesting that the effectiveness of redundancy reduction varies across backbone types and datasets. + +# 6. Conclusion + +We propose a novel approach for test-time adaptation (TTA) — Feature Redundancy Elimination for TTA (FRET) — which focuses on reducing feature redundancy during testing to address distribution shifts. This is the first attempt to leverage feature redundancy for TTA. We introduce two methods: S-FRET, which directly minimizes the feature redundancy score, and G-FRET, which optimizes both redundancy elimination and class discriminability by decomposing feature relations into attention and redundancy components. Experiments across various tasks and datasets show the effectiveness of our methods, demonstrating the value of feature redundancy elimination in TTA. + +# Author Contributions and Acknowledgements + +Linjing You initiated this work and was primarily responsible for proposing the core idea, methodology, theoretical analysis, experimental design, and discussion of key findings. Jiabao Lu conducted most of the experiments and contributed to the interpretation of experimental results. Xiayuan Huang participated in experimental discussions and provided feedback on the initial manuscript. Xiangli Nie offered valuable suggestions for the refinement and polishing of the paper. + +We would also like to sincerely thank all the anonymous reviewers and the editorial committee (ACS) whose insightful comments significantly improved the theoretical rigor, experimental depth, and overall quality of this work. + +This research was supported by Prof. Huang's project The Fundamental Research Funds for the Central Universities (Grant No. XJJSKYQD202538), Prof. Nie's project The National Key Research and Development Program of China (Grant No. 2024YFB4709100), and partially funded by the Beijing Nova Program (Grant Nos. 20220484070 and 20240484498). + +# References + +[1] Ivan Anishchenko, Samuel J Pellock, Tamuka M Chidyausiku, Theresa A Ramelot, Sergey Ovchinnikov, Jingzhou Hao, Khushboo Bafna, Christoffer Norn, Alex Kang, Asim K Bera, et al. De novo protein design by deep network hallucination. Nature, 600(7889):547-552, 2021. 1 +[2] Gavin Brown, Adam Pocock, Ming-Jie Zhao, and Mikel Luján. Conditional likelihood maximisation: a unifying framework for information theoretic feature selection. The journal of machine learning research, 13(1):27-66, 2012. 1 +[3] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label distribution-aware margin loss. In Proc. the Advances in Neural Information Processing Systems (NeurIPS), pages 1567-1578, 2019. 2, 4 +[4] Xue-wen Chen and Jong Cheol Jeong. Enhanced recursive feature elimination. In Sixth international conference on machine learning and applications (ICMLA 2007), pages 429-435. IEEE, 2007. 1 +[5] Ian Connick Covert, Wei Qiu, Mingyu Lu, Na Yoon Kim, Nathan J White, and Su-In Lee. Learning to maximize mutual information for dynamic feature selection. In International Conference on Machine Learning, pages 6424-6447. PMLR, 2023. 1 +[6] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.8 +[7] C. Du, Y. Wang, S. Song, and G. Huang. Probabilistic contrastive learning for long-tailed visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(9):5890-5904, 2024. 6, 3 +[8] Tongtong Fang, Nan Lu, Gang Niu, and Masashi Sugiyama. Rethinking importance weighting for deep learning under + +distribution shift. Advances in neural information processing systems, 33:11996-12007, 2020. 1 +[9] Ronald A Fisher. The use of multiple measurements in taxonomic problems. Annals of eugenics, 7(2):179-188, 1936. 1 +[10] Taesik Gong, Yewon Kim, Taeckyung Lee, Sorn Chottananurak, and Sung-Ju Lee. Sotta: Robust test-time adaptation on noisy data streams. Advances in Neural Information Processing Systems, 36, 2024. 1 +[11] Jie Gui, Zhenan Sun, Shuiwang Ji, Dacheng Tao, and Tieniu Tan. Feature selection based on structured sparsity: A comprehensive study. IEEE transactions on neural networks and learning systems, 28(7):1490-1507, 2016. 1 +[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 2, 6, 8, 3, 7, 9, 10 +[13] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019. 2, 6, 8, 3 +[14] Chenping Hou, Ruidong Fan, Ling-Li Zeng, and Dewen Hu. Adaptive feature selection with augmented attributes. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):9306-9324, 2023. 1 +[15] John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Žídek, Anna Potapenko, et al. Highly accurate protein structure prediction with alphafold. nature, 596(7873):583-589, 2021. 1 +[16] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. In Proc. the International Conference on Learning Representations (ICLR), 2019. 2, 4 +[17] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 5 +[18] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. nature, 521(7553):436-444, 2015. 1 +[19] Jonghyun Lee, Dahuin Jung, Saehyung Lee, Junsung Park, Juhyeon Shin, Uiwon Hwang, and Sungroh Yoon. Entropy is not enough for test-time adaptation: From the perspective of disentangled factors. In The Twelfth International Conference on Learning Representations, 2024. 2 +[20] Taeckyung Lee, Sorn Chottananurak, Taesik Gong, and Sung-Ju Lee. Aetta: Label-free accuracy estimation for test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28643-28652, 2024. 3 +[21] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. In Proceedings of the IEEE international conference on computer vision, pages 5542-5550, 2017. 6, 3 +[22] Jundong Li, Kewei Cheng, Suhang Wang, Fred Morstatter, Robert P Trevino, Jiliang Tang, and Huan Liu. Feature selection: A data perspective. ACM computing surveys (CSUR), 50(6):1-45, 2017. 2 + +[23] Junnan Li, Pan Zhou, Caiming Xiong, and Steven C.H. Hoi. Prototypical contrastive learning of unsupervised representations. In ICLR, 2021. 5 +[24] Jian Liang, Ran He, and Tieniu Tan. A comprehensive survey on test-time adaptation under distribution shifts. International Journal of Computer Vision, pages 1-34, 2024. 1, 2 +[25] Duy MH Nguyen, Hoang Nguyen, Nghiem Diep, Tan Ngoc Pham, Tri Cao, Binh Nguyen, Paul Swoboda, Nhat Ho, Shadi Albarqouni, Pengtao Xie, et al. Lvm-med: Learning large-scale self-supervised vision models for medical imaging via second-order graph matching. Advances in Neural Information Processing Systems, 36, 2024. 2 +[26] Wamidh K Mutlag, Shaker K Ali, Zahoor M Aydam, and Bahaa H Taher. Feature extraction methods: a review. In Journal of Physics: Conference Series, page 012028. IOP Publishing, 2020. 3, 1 +[27] A Tuan Nguyen, Thanh Nguyen-Tang, Ser-Nam Lim, and Philip HS Torr. Tipi: Test time adaptation with transformation invariance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24162-24171, 2023. 6, 8, 1, 3, 7, 9, 10 +[28] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In International conference on machine learning, pages 16888-16905. PMLR, 2022. 6, 1, 2, 3, 7, 8, 9, 10 +[29] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yaofo Chen, Peilin Zhao, and Mingkui Tan. Towards stable test-time adaptation in dynamic wild world. In International Conference on Learning Representations, 2023. 6, 8, 1, 3, 7, 9, 10 +[30] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 5 +[31] David Osowiechi, Mehrdad Noori, Gustavo Vargas Hakim, Moslem Yazdanpanah, Ali Bahri, Milad Cheraghalikhani, Sahar Dastani, Farzad Beizaee, Ismail Ayed, and Christian Desrosiers. Watt: Weight average test time adaptation of clip. Advances in neural information processing systems, 37: 48015-48044, 2024. 8 +[32] Sung Kyu Park, Seunghan Yang, Jaegul Choo, and Sungrack Yun. Label shift adapter for test-time adaptation under covariate and label shifts. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 16375-16385, 2023. 2, 4 +[33] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 3 +[34] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1406-1415, 2019. 8, 3 + +[35] Sharan Ramjee and Aly El Gamal. Efficient wrapper feature selection using autoencoder and model based elimination. arXiv preprint arXiv:1905.11592, 2019. 1 +[36] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. Advances in neural information processing systems, 33:11539-11551, 2020. 3, 6, 1, 7, 8, 9, 10 +[37] Claude Elwood Shannon. A mathematical theory of communication. The Bell system technical journal, 27(3):379-423, 1948. 2 +[38] David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering the game of go with deep neural networks and tree search. nature, 529(7587):484-489, 2016. 1 +[39] Samarth Sinha, Peter Gehler, Francesco Locatello, and Bernt Schiele. Test: Test-time self-training under distribution shift. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2759-2769, 2023. 2 +[40] Yongyi Su, Xun Xu, and Kui Jia. Towards real-world test-time adaptation: Tri-net self-training with balanced normalization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 15126-15135, 2024. 1 +[41] Yongyi Su, Xun Xu, Tianrui Li, and Kui Jia. Revisiting realistic test-time training: Sequential inference and adaptation by anchored clustering regularized self-training. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1 +[42] Yue Tan, Chen Chen, Weiming Zhuang, Xin Dong, Lingjuan Lyu, and Guodong Long. Is heterogeneity notorious? taming heterogeneity to handle test-time shift in federated learning. Advances in Neural Information Processing Systems, 36, 2024. 1 +[43] Dipti Theng and Kishor K Bhoyar. Feature selection techniques for machine learning: a survey of more than two decades of research. Knowledge and Information Systems, 66(3):1575-1637, 2024. 3, 1 +[44] Robert Tibshirani. Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society Series B: Statistical Methodology, 58(1):267-288, 1996. 1 +[45] Antonio Torralba and Alexei A Efros. Unbiased look at dataset bias. In CVPR 2011, pages 1521-1528. IEEE, 2011. 8, 3 +[46] Laurens van der Maaten and Geoffrey E. Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9: 2579-2605, 2008. 7 +[47] B Venkatesh and J Anuradha. A review of feature selection and its methods. Cybernetics and information technologies, 19(1):3-26, 2019. 1 +[48] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5018-5027, 2017. 6, 3 + +[49] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726, 2020. 6, 8, 1, 2, 3, 7, 9, 10 +[50] Mei Wang and Weihong Deng. Deep visual domain adaptation: A survey. Neurocomputing, 312:135-153, 2018. 1 +[51] Shuai Wang, Daoan Zhang, Zipei Yan, Jianguo Zhang, and Rui Li. Feature alignment and uniformity for test time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20050-20060, 2023. 6, 8, 1, 2, 3, 7, 9, 10 +[52] Svante Wold, Kim Esbensen, and Paul Geladi. Principal component analysis. Chemometrics and intelligent laboratory systems, 2(1-3):37-52, 1987. 1 +[53] Yanan Wu, Zhixiang Chi, Yang Wang, Konstantinos N Plataniotis, and Songhe Feng. Test-time domain adaptation by learning domain-aware batch normalization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 15961-15969, 2024. 2 +[54] Kaichao You, Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Universal domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2720-2729, 2019. 1 +[55] Linjing You, Jiabao Lu, and Xiayuan Huang. Test-time correlation alignment. In Forty-second International Conference on Machine Learning, 2025. 3, 2 +[56] Longhui Yuan, Binhui Xie, and Shuang Li. Robust test-time adaptation in dynamic scenarios. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15922-15932, 2023. 1 +[57] Yige Yuan, Bingbing Xu, Liang Hou, Fei Sun, Huawei Shen, and Xueqi Cheng. Tea: Test-time energy adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23901-23911, 2024. 6, 8, 1, 3, 7, 9, 10 +[58] Han Yue, Jundong Li, and Hongfu Liu. Second-order unsupervised feature selection via knowledge contrastive distillation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3, 4, 1, 2 +[59] Longbin Zeng, Jiayi Han, Liang Du, and Weiyang Ding. Rethinking precision of pseudo label: Test-time adaptation via complementary learning. Pattern Recognition Letters, 177: 96-102, 2024. 2 +[60] Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. Domain generalization: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4396-4415, 2022. 1 \ No newline at end of file diff --git a/2025/FRET_ Feature Redundancy Elimination for Test Time Adaptation/images.zip b/2025/FRET_ Feature Redundancy Elimination for Test Time Adaptation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9f97a5e710d7630e598e6586e00334716a3ada33 --- /dev/null +++ b/2025/FRET_ Feature Redundancy Elimination for Test Time Adaptation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:07f378866d78e276ba6e87957c3e50f5c5ae45c525cc95486e1e491b0ff4d8fa +size 624827 diff --git a/2025/FRET_ Feature Redundancy Elimination for Test Time Adaptation/layout.json b/2025/FRET_ Feature Redundancy Elimination for Test Time Adaptation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cb1c39465bdc2c495a856bc27e3d3bb27560922a --- /dev/null +++ b/2025/FRET_ Feature Redundancy Elimination for Test Time Adaptation/layout.json @@ -0,0 +1,10082 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 99, + 103, + 511, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 103, + 511, + 121 + ], + "spans": [ + { + "bbox": [ + 99, + 103, + 511, + 121 + ], + "type": "text", + "content": "FRET: Feature Redundancy Elimination for Test Time Adaptation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 140, + 141, + 467, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 141, + 467, + 157 + ], + "spans": [ + { + "bbox": [ + 140, + 141, + 467, + 157 + ], + "type": "text", + "content": "Linjing You" + }, + { + "bbox": [ + 140, + 141, + 467, + 157 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 140, + 141, + 467, + 157 + ], + "type": "text", + "content": " Jiabao Lu" + }, + { + "bbox": [ + 140, + 141, + 467, + 157 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 140, + 141, + 467, + 157 + ], + "type": "text", + "content": " Xiayuan Huang" + }, + { + "bbox": [ + 140, + 141, + 467, + 157 + ], + "type": "inline_equation", + "content": "^{2,\\dagger}" + }, + { + "bbox": [ + 140, + 141, + 467, + 157 + ], + "type": "text", + "content": " Xiangli Nie" + }, + { + "bbox": [ + 140, + 141, + 467, + 157 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 157, + 478, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 157, + 478, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 157, + 478, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 133, + 157, + 478, + 171 + ], + "type": "text", + "content": "Institute of Automation, Chinese Academy of Sciences, Beijing, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 149, + 171, + 459, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 171, + 459, + 186 + ], + "spans": [ + { + "bbox": [ + 149, + 171, + 459, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 149, + 171, + 459, + 186 + ], + "type": "text", + "content": "College of Science, Beijing Forestry University, Beijing, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 119, + 186, + 492, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 186, + 492, + 200 + ], + "spans": [ + { + "bbox": [ + 119, + 186, + 492, + 200 + ], + "type": "text", + "content": "{youlinjing2023,lujiabao2025,xiangliNie}@ia.ac.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 218, + 201, + 392, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 218, + 201, + 392, + 213 + ], + "spans": [ + { + "bbox": [ + 218, + 201, + 392, + 213 + ], + "type": "text", + "content": "huangxiayuan@bjfu.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 231, + 213, + 376, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 213, + 376, + 222 + ], + "spans": [ + { + "bbox": [ + 231, + 213, + 376, + 222 + ], + "type": "text", + "content": "* Equal contribution, †Corresponding author" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "spans": [ + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 282, + 297, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 282, + 297, + 642 + ], + "spans": [ + { + "bbox": [ + 54, + 282, + 297, + 642 + ], + "type": "text", + "content": "Test-Time Adaptation (TTA) aims to enhance the generalization of deep learning models when faced with test data that exhibits distribution shifts from the training data. In this context, only a pre-trained model and unlabeled test data are available, making it particularly relevant for privacy-sensitive applications. In practice, we observe that feature redundancy in embeddings tends to increase as domain shifts intensify in TTA. However, existing TTA methods often overlook this redundancy, which can hinder the model's adaptability to new data. To address this issue, we introduce Feature Redundancy Elimination for Test-time Adaptation (FRET), a novel perspective for TTA. A straightforward approach (S-FRET) is to directly minimize the feature redundancy score as an optimization objective to improve adaptation. Despite its simplicity and effectiveness, S-FRET struggles with label shifts, limiting its robustness in real-world scenarios. To mitigate this limitation, we further propose Graph-based FRET (G-FRET), which integrates a Graph Convolutional Network (GCN) with contrastive learning. This design not only reduces feature redundancy but also enhances feature discriminability in both the representation and prediction layers. Extensive experiments across multiple model architectures, tasks, and datasets demonstrate the effectiveness of S-FRET and show that G-FRET achieves state-of-the-art performance. Further analysis reveals that G-FRET enables the model to extract non-redundant and highly discriminative features during inference, thereby facilitating more robust test-time adaptation. The code is available at https://github.com/youlj109/fret." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 668, + 135, + 680 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 668, + 135, + 680 + ], + "spans": [ + { + "bbox": [ + 55, + 668, + 135, + 680 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Recently, the emergence of deep neural networks (DNNs) [12, 18] has significantly propelled the advancement in a" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 328, + 261, + 537, + 454 + ], + "blocks": [ + { + "bbox": [ + 328, + 261, + 537, + 454 + ], + "lines": [ + { + "bbox": [ + 328, + 261, + 537, + 454 + ], + "spans": [ + { + "bbox": [ + 328, + 261, + 537, + 454 + ], + "type": "image", + "image_path": "13e6146631ac6f03fd39d1fa7363a0ba87ab939cb5b67b48fca8e42a497f032f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 463, + 555, + 509 + ], + "lines": [ + { + "bbox": [ + 313, + 463, + 555, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 555, + 509 + ], + "type": "text", + "content": "Figure 1. An intuitive demonstration of the relationship between embedded feature redundancy and distribution migration. Top: Second-order feature relation graph of embedding features. Bottom: Plot of feature redundancy versus the level of corruption." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 533, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 533, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 533, + 555, + 715 + ], + "type": "text", + "content": "Vast range of human tasks [1, 15, 38]. However, the basic assumption that training and test data are drawn independently and identically (i.i.d.) may not be suitable in many real-world scenarios where the test/target distribution diverges from the training/source distribution [50]. This phenomenon, known as distribution shift, often leads to poor performance [8]. To tackle this issue, a variety of generalization or adaptation techniques have been developed to improve model robustness against distribution shifts [24, 54, 60]. Among these, Test Time Adaptation (TTA) [10, 40, 41] has been the focus of recent and active research. TTA has the advantage of only requiring access to the pre-trained model from the training/source domain and unlabeled test data. Therefore, TTA is a more secure and practical solution especially when models are publicly available." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "2120" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 96 + ], + "type": "text", + "content": "available but the training data and training process remain inaccessible due to privacy and resource restrictions [24]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 99, + 296, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 99, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 56, + 99, + 296, + 350 + ], + "type": "text", + "content": "Existing TTA methods can be broadly categorized into batch normalization calibration methods, pseudo-labeling methods, consistency training methods, and clustering-based training methods, based on different adaptation strategies [24]. Batch normalization (BN) calibration methods [53] posit that the statistics in the BN layers represent domain-specific knowledge. To bridge the domain gap, it is suggested to replace the training BN statistics with new statistics estimated over the entire target domain. Pseudolabeling methods [59] employ various filtering strategies, such as thresholding or entropy-based approaches, to obtain reliable pseudo-labels. This process helps minimize discrepancies between predicted labels and pseudo-labels. Consistency training methods [39] aim to improve the stability of network predictions or features by addressing variations in input data, such as noise or perturbations, and changes in model parameters. Clustering-based training methods [19] focus on reducing uncertainty in target network predictions by leveraging clustering techniques to organize target features into distinct groups, thereby enhancing the robustness of the model." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 352, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 352, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 56, + 352, + 296, + 579 + ], + "type": "text", + "content": "Despite the empirical success of existing methods, they overlook feature redundancy in embeddings, which hinders generalization to new data [22]. As shown in Fig. 1, we visualize feature redundancy in ResNet-18 [12] using second-order feature relations (i.e., covariance matrix) on the CIFAR10-C dataset [13] under distribution shift [25]. Each image is embedded into a 512-dimensional vector; stacking " + }, + { + "bbox": [ + 56, + 352, + 296, + 579 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 56, + 352, + 296, + 579 + ], + "type": "text", + "content": " embeddings yields an " + }, + { + "bbox": [ + 56, + 352, + 296, + 579 + ], + "type": "inline_equation", + "content": "n \\times 512" + }, + { + "bbox": [ + 56, + 352, + 296, + 579 + ], + "type": "text", + "content": " matrix. We compute the absolute covariance matrix " + }, + { + "bbox": [ + 56, + 352, + 296, + 579 + ], + "type": "inline_equation", + "content": "(512 \\times 512)" + }, + { + "bbox": [ + 56, + 352, + 296, + 579 + ], + "type": "text", + "content": " and visualize it as a heatmap — redder areas indicate higher feature correlation, hence greater redundancy. The top two subfigures in Fig. 1 compare the source and target domains, revealing much higher redundancy in the latter due to distribution shift. To quantify this, we remove the diagonal of the covariance matrix and sum the absolute values of the remaining elements (see Eq. (1)). As shown at the bottom of Fig. 1, redundancy increases with corruption severity, suggesting that features with lower redundancy may be better suited for adaptation to the target domain." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "Building on the limitations identified in the existing literature and the observations above, we provide a novel perspective for test-time adaptation—Feature Redundancy Elimination for TTA (FRET)—which differs from existing methods. Due to the inherently differentiable nature of the redundancy metric, a straightforward and computationally efficient approach is to directly minimize the feature redundancy score. We refer to this simple approach as S-FRET. However, since S-FRET lacks the ability to perceive label distributions, it effectively addresses covariate shift but fails to handle label shift—a scenario where the tar" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "get domain exhibits a label distribution that deviates from the source domain (e.g., long-tailed data). This discrepancy can significantly degrade model performance and remains a critical challenge [3, 16, 32]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 121, + 555, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 555, + 336 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 555, + 336 + ], + "type": "text", + "content": "To overcome S-FRET's limitation in handling label shift, we further propose G-FRET, which integrates a Graph Convolutional Network (GCN). Unlike FRET, which only minimizes redundancy at the representation layer, G-FRET leverages GCN to incorporate both attention and redundant feature relationships into the representation and prediction layers. Specifically, G-FRET learns non-redundant and discriminative representation through contrastive distillation between attention and redundant representation in the representation-layer. In the prediction layer, attention predictions are optimized through entropy minimization, while ensuring that they are distinct from redundant predictions by negative learning. By maximizing the discriminability of attention relationships while eliminating redundancy in both the representation and prediction layers, G-FRET effectively reduces test-time feature redundancy and improves discrimination when handling complex distribution-shifted test data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 326, + 338, + 553, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 338, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 326, + 338, + 553, + 350 + ], + "type": "text", + "content": "We summarize the contributions of this paper as follows." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 354, + 553, + 629 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 354, + 553, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 354, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 314, + 354, + 553, + 437 + ], + "type": "text", + "content": "- Potential Way: provide a novel perspective for test-time adaptation — Feature Redundancy Elimination for TTA (FRET) — which differs from existing methods. It adapts the model to distribution shifts by reducing the redundancy of embedded features during testing. To the best of our knowledge, this is the first attempt to leverage feature redundancy for TTA." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 438, + 553, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 438, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 314, + 438, + 553, + 533 + ], + "type": "text", + "content": "- Innovative Methods: We propose two novel methods, S-FRET and G-FRET, to eliminate feature redundancy in test-time embeddings. S-FRET directly minimizes the feature redundancy score, offering a lightweight solution. G-FRET extends this by decomposing feature relations into attention and redundancy components, enabling joint optimization of redundancy elimination and class discriminability in real-world scenarios." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 534, + 553, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 534, + 553, + 629 + ], + "spans": [ + { + "bbox": [ + 314, + 534, + 553, + 629 + ], + "type": "text", + "content": "- Extensive Evaluation: Comprehensive experiments across multiple model architectures, tasks, and datasets demonstrate the effectiveness of S-FRET and show that G-FRET achieves state-of-the-art performance. Further in-depth analysis reveals that G-FRET facilitates the extraction of non-redundant and discriminative embedding features during testing, ultimately enhancing model adaptation to the test-time target domain." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 644, + 466, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 644, + 466, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 644, + 466, + 658 + ], + "type": "text", + "content": "2. Preliminary and Notations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": "We briefly revisit Test-Time Adaptation and Feature Redundancy Elimination in this section for the convenience of introducing methods, and put detailed related work discussions into Sec. 7 due to page limits." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2121" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 70, + 195, + 169 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 195, + 169 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 195, + 169 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 195, + 169 + ], + "type": "image", + "image_path": "b212cf9628f98fd8cde427cb2b685556eff20199b436d43bf8a4dfd8a3f9de70.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 178, + 555, + 200 + ], + "lines": [ + { + "bbox": [ + 55, + 178, + 555, + 200 + ], + "spans": [ + { + "bbox": [ + 55, + 178, + 555, + 200 + ], + "type": "text", + "content": "Figure 2. Redundancy scores of 15 corruption types in CIFAR-10/100-C and ImageNet-C across 5 corruption levels, as well as for source and target domains in PACS, OfficeHome, VLCS, and DomainNet." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 195, + 70, + 307, + 169 + ], + "blocks": [ + { + "bbox": [ + 195, + 70, + 307, + 169 + ], + "lines": [ + { + "bbox": [ + 195, + 70, + 307, + 169 + ], + "spans": [ + { + "bbox": [ + 195, + 70, + 307, + 169 + ], + "type": "image", + "image_path": "6dd2ddf82f35c52220a5a119a5822625c5201de76d2762f2cd4671979169c65b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 306, + 70, + 419, + 169 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 419, + 169 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 419, + 169 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 419, + 169 + ], + "type": "image", + "image_path": "1c64f3a491115c11a74b19ffadce1749266f4663facd651a460844a6558aff84.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 418, + 70, + 529, + 169 + ], + "blocks": [ + { + "bbox": [ + 418, + 70, + 529, + 169 + ], + "lines": [ + { + "bbox": [ + 418, + 70, + 529, + 169 + ], + "spans": [ + { + "bbox": [ + 418, + 70, + 529, + 169 + ], + "type": "image", + "image_path": "5feb3184358798be6c31190903e333107c91a28c2b51d1ab20fa19683c01caf8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 205, + 181, + 218 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 181, + 218 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 181, + 218 + ], + "type": "text", + "content": "2.1. Test-Time Adaptation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 224, + 296, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 224, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 296, + 415 + ], + "type": "text", + "content": "In the test-time adaptation (TTA) scenario, we have access only to unlabeled data from the target domain in an online manner and a pre-trained model from the source domain. Specifically, let " + }, + { + "bbox": [ + 55, + 224, + 296, + 415 + ], + "type": "inline_equation", + "content": "\\{(x_i,y_i)\\}_{i = 1}^{n_s}\\subset \\mathcal{X}_s\\times \\mathcal{Y}_s" + }, + { + "bbox": [ + 55, + 224, + 296, + 415 + ], + "type": "text", + "content": " represent the labeled source domain dataset with labels " + }, + { + "bbox": [ + 55, + 224, + 296, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_s" + }, + { + "bbox": [ + 55, + 224, + 296, + 415 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 224, + 296, + 415 + ], + "type": "inline_equation", + "content": "\\{x_j\\}_{j = 1}^{n_t}\\subset \\mathcal{X}_t" + }, + { + "bbox": [ + 55, + 224, + 296, + 415 + ], + "type": "text", + "content": " represent a batch of unlabeled target data. The model, trained on the source domain dataset and parameterized by " + }, + { + "bbox": [ + 55, + 224, + 296, + 415 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 224, + 296, + 415 + ], + "type": "text", + "content": ", is denoted as " + }, + { + "bbox": [ + 55, + 224, + 296, + 415 + ], + "type": "inline_equation", + "content": "g_{\\theta} = h(f(\\cdot)): \\mathcal{X}_s\\to \\mathcal{Y}_s" + }, + { + "bbox": [ + 55, + 224, + 296, + 415 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 224, + 296, + 415 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 224, + 296, + 415 + ], + "type": "text", + "content": " is the backbone encoder and " + }, + { + "bbox": [ + 55, + 224, + 296, + 415 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 224, + 296, + 415 + ], + "type": "text", + "content": " denotes the linear classification head. Our objective is to adapt this model to the target domain " + }, + { + "bbox": [ + 55, + 224, + 296, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_t" + }, + { + "bbox": [ + 55, + 224, + 296, + 415 + ], + "type": "text", + "content": ", which exhibits a different distribution from that of the source domain " + }, + { + "bbox": [ + 55, + 224, + 296, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_s" + }, + { + "bbox": [ + 55, + 224, + 296, + 415 + ], + "type": "text", + "content": ". During testing, for each instance " + }, + { + "bbox": [ + 55, + 224, + 296, + 415 + ], + "type": "inline_equation", + "content": "x_{j}\\in \\mathcal{X}_{t}" + }, + { + "bbox": [ + 55, + 224, + 296, + 415 + ], + "type": "text", + "content": ", let the output of " + }, + { + "bbox": [ + 55, + 224, + 296, + 415 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 224, + 296, + 415 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 224, + 296, + 415 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 224, + 296, + 415 + ], + "type": "text", + "content": " be denoted as embedding " + }, + { + "bbox": [ + 55, + 224, + 296, + 415 + ], + "type": "inline_equation", + "content": "z_{j} = f(x_{j})\\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 55, + 224, + 296, + 415 + ], + "type": "text", + "content": " and logs " + }, + { + "bbox": [ + 55, + 224, + 296, + 415 + ], + "type": "inline_equation", + "content": "p_j = h(z_j)\\in \\mathbb{R}^C" + }, + { + "bbox": [ + 55, + 224, + 296, + 415 + ], + "type": "text", + "content": ", respectively, where " + }, + { + "bbox": [ + 55, + 224, + 296, + 415 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 224, + 296, + 415 + ], + "type": "text", + "content": " is the dimension of the embeddings and " + }, + { + "bbox": [ + 55, + 224, + 296, + 415 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 224, + 296, + 415 + ], + "type": "text", + "content": " is the number of classes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 425, + 234, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 234, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 234, + 437 + ], + "type": "text", + "content": "2.2. Feature Redundancy Elimination" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 443, + 296, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 443, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 296, + 502 + ], + "type": "text", + "content": "Feature redundancy is a key concern in both feature extraction and feature selection [26, 43]. Following previous work [58], we remove the diagonal elements of the covariance matrix and calculate the sum of the absolute values of the remaining elements to derive the redundancy scores " + }, + { + "bbox": [ + 55, + 443, + 296, + 502 + ], + "type": "inline_equation", + "content": "R_{s}" + }, + { + "bbox": [ + 55, + 443, + 296, + 502 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 512, + 294, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 294, + 533 + ], + "type": "interline_equation", + "content": "R _ {s} = \\left\\| \\tilde {Z} ^ {T} \\tilde {Z} - I _ {d} \\right\\| _ {1} \\tag {1}", + "image_path": "cfc7c2d4f7a6011ac895e4b4cc56e567eb7a8e2c9c77ce8426cd99fd94cceade.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 542, + 296, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 542, + 296, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 296, + 582 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 542, + 296, + 582 + ], + "type": "inline_equation", + "content": "I_{d}" + }, + { + "bbox": [ + 55, + 542, + 296, + 582 + ], + "type": "text", + "content": " is the identity matrix, " + }, + { + "bbox": [ + 55, + 542, + 296, + 582 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\| _1" + }, + { + "bbox": [ + 55, + 542, + 296, + 582 + ], + "type": "text", + "content": " denotes the 1-norm of a matrix and " + }, + { + "bbox": [ + 55, + 542, + 296, + 582 + ], + "type": "inline_equation", + "content": "\\tilde{Z}_{:,j} = \\frac{Z_{:,j}}{\\|Z_{:,j}\\|_2}" + }, + { + "bbox": [ + 55, + 542, + 296, + 582 + ], + "type": "text", + "content": " denotes the matrix obtained by normalizing each column." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 594, + 285, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 285, + 607 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 285, + 607 + ], + "type": "text", + "content": "3. Feature Redundancy Under Domain Shift" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 614, + 296, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 614, + 296, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 614, + 296, + 638 + ], + "type": "text", + "content": "We conduct a systematic investigation into the behavior of feature redundancy under domain shift." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 647, + 157, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 157, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 157, + 661 + ], + "type": "text", + "content": "3.1. Empirical Study." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "As shown in Fig. 2, across approximately 50 sub-datasets, the feature redundancy score " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "R_{s}" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " consistently increases with the severity of domain shift. (Note: " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "R_{s}" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " is computed using models with BN parameters updated from test data" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 206, + 554, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 206, + 554, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 206, + 554, + 266 + ], + "type": "text", + "content": "rather than those fixed during training [36].) This empirical trend suggests that domain shift induces stronger correlations among features. To the best of our knowledge, this phenomenon has not been systematically explored in prior work." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 273, + 453, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 273, + 453, + 286 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 453, + 286 + ], + "type": "text", + "content": "3.2. Theoretical Explanation." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 290, + 554, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 290, + 554, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 290, + 554, + 338 + ], + "type": "text", + "content": "Intuitively, higher covariance elements indicate that the features fail to span the full vector space. Furthermore, we provide the following theoretical result. Missing proofs and detailed explanations are provided in Sec. 9" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 345, + 555, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 345, + 555, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 555, + 477 + ], + "type": "text", + "content": "Theorem 1 (Redundancy Elimination Improves Generalization). Under the empirical observation that embeddings from the source domain exhibit lower feature redundancy, i.e., " + }, + { + "bbox": [ + 313, + 345, + 555, + 477 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_s(Z_t) > \\mathcal{R}_s(Z_s)" + }, + { + "bbox": [ + 313, + 345, + 555, + 477 + ], + "type": "text", + "content": ", and by leveraging Theorem 2 from [55], we can derive the following generalization error bound: Let " + }, + { + "bbox": [ + 313, + 345, + 555, + 477 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 345, + 555, + 477 + ], + "type": "text", + "content": " be a hypothesis class of VC-dimension " + }, + { + "bbox": [ + 313, + 345, + 555, + 477 + ], + "type": "inline_equation", + "content": "d_v" + }, + { + "bbox": [ + 313, + 345, + 555, + 477 + ], + "type": "text", + "content": ". If " + }, + { + "bbox": [ + 313, + 345, + 555, + 477 + ], + "type": "inline_equation", + "content": "\\hat{h} \\in \\mathcal{H}" + }, + { + "bbox": [ + 313, + 345, + 555, + 477 + ], + "type": "text", + "content": " minimizes the empirical error " + }, + { + "bbox": [ + 313, + 345, + 555, + 477 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_s(h)" + }, + { + "bbox": [ + 313, + 345, + 555, + 477 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 313, + 345, + 555, + 477 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_s" + }, + { + "bbox": [ + 313, + 345, + 555, + 477 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 345, + 555, + 477 + ], + "type": "inline_equation", + "content": "h_t^* = \\arg \\min_{h \\in \\mathcal{H}} \\epsilon_t(h)" + }, + { + "bbox": [ + 313, + 345, + 555, + 477 + ], + "type": "text", + "content": " is the optimal hypothesis on " + }, + { + "bbox": [ + 313, + 345, + 555, + 477 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_t" + }, + { + "bbox": [ + 313, + 345, + 555, + 477 + ], + "type": "text", + "content": ", with the assumption that all hypotheses are L-Lipschitz continuous, then " + }, + { + "bbox": [ + 313, + 345, + 555, + 477 + ], + "type": "inline_equation", + "content": "\\forall \\delta \\in (0,1)" + }, + { + "bbox": [ + 313, + 345, + 555, + 477 + ], + "type": "text", + "content": ", with probability with at least " + }, + { + "bbox": [ + 313, + 345, + 555, + 477 + ], + "type": "inline_equation", + "content": "1 - \\delta" + }, + { + "bbox": [ + 313, + 345, + 555, + 477 + ], + "type": "text", + "content": " the following inequality holds:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 331, + 483, + 537, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 483, + 537, + 520 + ], + "spans": [ + { + "bbox": [ + 331, + 483, + 537, + 520 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\epsilon_ {t} (\\hat {h}) - \\epsilon_ {t} (h _ {t} ^ {*}) \\\\ \\leq \\mathcal {O} (\\sqrt {\\| \\mu_ {s} - \\mu_ {t} \\| _ {F} ^ {2} + R _ {s} (Z _ {s}) ^ {2} + R _ {s} (Z _ {t}) ^ {2}}) + C \\\\ \\end{array}", + "image_path": "c98404b938a88c54c3ba1dfcbae5c52d0c0a692f38c9f5c71b5d323e3ed68f44.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 529, + 554, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 554, + 606 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 554, + 606 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 529, + 554, + 606 + ], + "type": "inline_equation", + "content": "C = 2\\sqrt{\\frac{d_v\\log(2n_s) - \\log(\\delta)}{2n_s}} + 2\\gamma" + }, + { + "bbox": [ + 313, + 529, + 554, + 606 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 529, + 554, + 606 + ], + "type": "inline_equation", + "content": "\\gamma = \\min_{h\\in \\mathcal{H}}\\{\\epsilon_s(h(t)) + \\epsilon_t(h(t))\\}" + }, + { + "bbox": [ + 313, + 529, + 554, + 606 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 529, + 554, + 606 + ], + "type": "inline_equation", + "content": "\\mu_{s}" + }, + { + "bbox": [ + 313, + 529, + 554, + 606 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 529, + 554, + 606 + ], + "type": "inline_equation", + "content": "\\mu_t" + }, + { + "bbox": [ + 313, + 529, + 554, + 606 + ], + "type": "text", + "content": " denote the means of the source and test embeddings, respectively. We use " + }, + { + "bbox": [ + 313, + 529, + 554, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(\\cdot)" + }, + { + "bbox": [ + 313, + 529, + 554, + 606 + ], + "type": "text", + "content": " to hide the constant dependence. Since " + }, + { + "bbox": [ + 313, + 529, + 554, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_s(Z_t) > \\mathcal{R}_s(Z_s)" + }, + { + "bbox": [ + 313, + 529, + 554, + 606 + ], + "type": "text", + "content": ", this implies that reducing " + }, + { + "bbox": [ + 313, + 529, + 554, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_s(Z_t)" + }, + { + "bbox": [ + 313, + 529, + 554, + 606 + ], + "type": "text", + "content": " can effectively decrease the generalization error." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 616, + 454, + 630 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 454, + 630 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 454, + 630 + ], + "type": "text", + "content": "4. The Proposed Approach" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 635, + 376, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 376, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 376, + 647 + ], + "type": "text", + "content": "4.1. S-FRET" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 654, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 712 + ], + "type": "text", + "content": "To eliminate feature redundancy in the target domain, a simple yet effective approach is to directly minimize the feature redundancy score " + }, + { + "bbox": [ + 313, + 654, + 554, + 712 + ], + "type": "inline_equation", + "content": "R_{s}" + }, + { + "bbox": [ + 313, + 654, + 554, + 712 + ], + "type": "text", + "content": ", as defined in Eq. (1), leading to the proposed method S-FRET (Simple and Straightforward FRET). The loss function for S-FRET can be written as:" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2122" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 69, + 531, + 240 + ], + "blocks": [ + { + "bbox": [ + 82, + 69, + 531, + 240 + ], + "lines": [ + { + "bbox": [ + 82, + 69, + 531, + 240 + ], + "spans": [ + { + "bbox": [ + 82, + 69, + 531, + 240 + ], + "type": "image", + "image_path": "b7081ca10dda8bd484b3ce6ea5bcf124602d584b14194d34e556de7e34681a36.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 247, + 555, + 293 + ], + "lines": [ + { + "bbox": [ + 54, + 247, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 54, + 247, + 555, + 293 + ], + "type": "text", + "content": "Figure 3. The pipeline of our proposed Graph Convolutional Network based Feature Redundancy Elimination for Test Time Adaptation (G-FRET) method. For each test sample, G-FRET learns non-redundant and discriminative representation through contrastive distillation in the representation-layer. Additionally, in the prediction layer, attention predictions are optimized through entropy minimization, while ensuring that they are distinct from redundant predictions by using negative learning." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 119, + 308, + 295, + 331 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 308, + 295, + 331 + ], + "spans": [ + { + "bbox": [ + 119, + 308, + 295, + 331 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S F R E T} = \\left\\| \\tilde {Z} ^ {T} \\tilde {Z} - I _ {d} \\right\\| _ {1} \\tag {2}", + "image_path": "fdd617fb3d39d561ec5c1bfbead712b0c92a1b474afb0c3cf86fe210b5a5311e.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 335, + 296, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 296, + 418 + ], + "type": "text", + "content": "This method directly minimizes the redundancy score derived from feature correlation, providing a computationally efficient solution that avoids complex architectural modifications. By focusing on the feature space of representation-layer, S-FRET enables the model to extract non-redundant representations, thereby enhancing its adaptability to the target domain (see Fig. 4)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 419, + 296, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 296, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 296, + 515 + ], + "type": "text", + "content": "However, since the redundancy score does not incorporate information about class distributions, S-FRET primarily addresses covariate shift but lacks the discriminative capability to handle label shift—a critical challenge where the target domain exhibits a label distribution that deviates significantly from the source domain (e.g., long-tailed data), leading to substantial performance degradation (see Tab. 4) [3, 16, 32]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 523, + 119, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 523, + 119, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 523, + 119, + 534 + ], + "type": "text", + "content": "4.2. G-FRET" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 541, + 176, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 541, + 176, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 541, + 176, + 553 + ], + "type": "text", + "content": "4.2.1. Overview of G-FRET" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 557, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 652 + ], + "type": "text", + "content": "To eliminate feature redundancy and address label shift in the target domain, we further propose a novel two-layer TTA method, namely Graph Convolutional Network based Feature Redundancy Elimination for Test Time Adaptation (G-FRET), which leverages GCN to simultaneously model feature relationships and sample-specific discriminative information in both representation and prediction layers. Fig. 3 provides an overview of our G-FRET framework." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "The key idea of G-FRET is to adapt the pre-tarined model during testing by extracting discriminative and nonredundant embedding features from both the representation and prediction layers. Firstly (Sec. 4.2.2), we apply a mask matrix to the second-order feature relation graph." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 298, + 555, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 298, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 298, + 555, + 525 + ], + "type": "text", + "content": "This process yields two separate graphs: an attention graph, which highlights significant feature relationships, and a redundant graph, which identifies redundant feature relationships. Next, we derive the attention/redundancy representation and attention/redundancy prediction by utilizing a single-layer GCN to integrate the model's embeddings with the attention and redundant graphs. This process allows us to explicitly separate attention and redundancy relations in both representation and prediction layer. Then (Sec. 4.2.3), in the representation layer, contrastive learning techniques are employed to enhance the discriminability of the attention representation while ensuring it remains distant from the redundant representation. At last (Sec. 4.2.4), in the prediction layer, attention predictions are optimized through entropy minimization, while ensuring that they are distinct from redundant predictions by negative learning. Through this two-layer process, we achieve discriminative and nonredundant embedding features that effectively adapt to the target domain." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 536, + 500, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 500, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 500, + 548 + ], + "type": "text", + "content": "4.2.2. Attention Part and Redundancy Part" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 553, + 555, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 553, + 555, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 553, + 555, + 696 + ], + "type": "text", + "content": "Given a batch of " + }, + { + "bbox": [ + 313, + 553, + 555, + 696 + ], + "type": "inline_equation", + "content": "n_t" + }, + { + "bbox": [ + 313, + 553, + 555, + 696 + ], + "type": "text", + "content": " data instances from the target domain, we obtain the first-order " + }, + { + "bbox": [ + 313, + 553, + 555, + 696 + ], + "type": "inline_equation", + "content": "n_t \\times d" + }, + { + "bbox": [ + 313, + 553, + 555, + 696 + ], + "type": "text", + "content": " embedding feature matrix " + }, + { + "bbox": [ + 313, + 553, + 555, + 696 + ], + "type": "inline_equation", + "content": "Z \\in \\mathbb{R}^{n_t \\times d}" + }, + { + "bbox": [ + 313, + 553, + 555, + 696 + ], + "type": "text", + "content": " and the second-order feature relation graph " + }, + { + "bbox": [ + 313, + 553, + 555, + 696 + ], + "type": "inline_equation", + "content": "G_F = Z^T Z \\in \\mathbb{R}^{d \\times d}" + }, + { + "bbox": [ + 313, + 553, + 555, + 696 + ], + "type": "text", + "content": ". Following the state-of-the-art feature selection method SOFT [58] (see Appendix 7.3 for more details), we apply a symmetric mask matrix " + }, + { + "bbox": [ + 313, + 553, + 555, + 696 + ], + "type": "inline_equation", + "content": "M_M \\in \\mathbb{R}^{d \\times d}" + }, + { + "bbox": [ + 313, + 553, + 555, + 696 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 553, + 555, + 696 + ], + "type": "inline_equation", + "content": "G_F" + }, + { + "bbox": [ + 313, + 553, + 555, + 696 + ], + "type": "text", + "content": ". This decomposition of " + }, + { + "bbox": [ + 313, + 553, + 555, + 696 + ], + "type": "inline_equation", + "content": "G_F" + }, + { + "bbox": [ + 313, + 553, + 555, + 696 + ], + "type": "text", + "content": " yields two distinct graphs: the attention graph " + }, + { + "bbox": [ + 313, + 553, + 555, + 696 + ], + "type": "inline_equation", + "content": "G_A" + }, + { + "bbox": [ + 313, + 553, + 555, + 696 + ], + "type": "text", + "content": " and the redundancy graph " + }, + { + "bbox": [ + 313, + 553, + 555, + 696 + ], + "type": "inline_equation", + "content": "G_R" + }, + { + "bbox": [ + 313, + 553, + 555, + 696 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 553, + 555, + 696 + ], + "type": "inline_equation", + "content": "G_A" + }, + { + "bbox": [ + 313, + 553, + 555, + 696 + ], + "type": "text", + "content": " captures the attention relations (i.e., important relation) among features, while " + }, + { + "bbox": [ + 313, + 553, + 555, + 696 + ], + "type": "inline_equation", + "content": "G_R" + }, + { + "bbox": [ + 313, + 553, + 555, + 696 + ], + "type": "text", + "content": " represents the redundant relation between them. Formally, they are calculated as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 359, + 701, + 555, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 359, + 701, + 555, + 714 + ], + "type": "interline_equation", + "content": "G _ {A} = G _ {F} \\odot M _ {M}; G _ {R} = G _ {F} - G _ {A} \\tag {3}", + "image_path": "79239c4679a1f9af28f782a0ca715d6acc67fa0bea3dba14ebc2b270adb12396.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2123" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": " is the element-wise product. However, differently from SOFT, which focuses on feature selection by deriving an optimal mask matrix " + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "inline_equation", + "content": "M_M" + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": ", our objective is to extract non-redundant features " + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": ". For ideally non-redundant features, the second-order feature relation graph should be a diagonal matrix, indicating that each feature is independent of all others. Therefore, we set " + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "inline_equation", + "content": "M_M = I_d" + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": " to ensure that " + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "inline_equation", + "content": "G_A" + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": " captures only the essential feature relations, while " + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "inline_equation", + "content": "G_R" + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": " includes all remaining relations as redundant relations. Here, " + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "inline_equation", + "content": "I_d \\in \\mathbb{R}^{d \\times d}" + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": " is the identity matrix. In addition, in some scenarios with prior knowledge, such as knowing which features may be more important, or needing to maintain special structural relations between features, some adaptations can be made by simply modifying the mask matrix " + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "inline_equation", + "content": "M_M" + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 240, + 296, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 240, + 296, + 336 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 296, + 336 + ], + "type": "text", + "content": "Then, in order to obtain attention/redundancy representation and attention/redundancy prediction based on the attention and redundancy graphs, we combine the normalized graph matrix with the classifier head " + }, + { + "bbox": [ + 55, + 240, + 296, + 336 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 240, + 296, + 336 + ], + "type": "text", + "content": " to form a one-layer Graph Convolutional Network (GCN [17]), which can help generate attention representations " + }, + { + "bbox": [ + 55, + 240, + 296, + 336 + ], + "type": "inline_equation", + "content": "R_{A}" + }, + { + "bbox": [ + 55, + 240, + 296, + 336 + ], + "type": "text", + "content": " and attention predictions " + }, + { + "bbox": [ + 55, + 240, + 296, + 336 + ], + "type": "inline_equation", + "content": "P_{A}" + }, + { + "bbox": [ + 55, + 240, + 296, + 336 + ], + "type": "text", + "content": " that contain both data information and feature relationship information:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 94, + 347, + 295, + 363 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 347, + 295, + 363 + ], + "spans": [ + { + "bbox": [ + 94, + 347, + 295, + 363 + ], + "type": "interline_equation", + "content": "R _ {A} = Z D _ {A} ^ {- 1 / 2} G _ {A} D _ {A} ^ {- 1 / 2}; P _ {A} = R _ {A} \\theta^ {h} \\tag {4}", + "image_path": "a7d5c2a4c76df6028b848e7554c648b0fd90f2d48021afde9d7f06207c05c0d9.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 373, + 296, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 296, + 437 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 373, + 296, + 437 + ], + "type": "inline_equation", + "content": "D_A^{-1/2}G_A D_A^{-1/2}" + }, + { + "bbox": [ + 55, + 373, + 296, + 437 + ], + "type": "text", + "content": " is the normalized graph matrix of graph " + }, + { + "bbox": [ + 55, + 373, + 296, + 437 + ], + "type": "inline_equation", + "content": "G_A" + }, + { + "bbox": [ + 55, + 373, + 296, + 437 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 373, + 296, + 437 + ], + "type": "inline_equation", + "content": "D_A" + }, + { + "bbox": [ + 55, + 373, + 296, + 437 + ], + "type": "text", + "content": " is the degree matrix of the graph, and " + }, + { + "bbox": [ + 55, + 373, + 296, + 437 + ], + "type": "inline_equation", + "content": "\\theta^h" + }, + { + "bbox": [ + 55, + 373, + 296, + 437 + ], + "type": "text", + "content": " are the parameters of the classifier head " + }, + { + "bbox": [ + 55, + 373, + 296, + 437 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 373, + 296, + 437 + ], + "type": "text", + "content": ". Similarly, we can obtain redundancy representations " + }, + { + "bbox": [ + 55, + 373, + 296, + 437 + ], + "type": "inline_equation", + "content": "R_R" + }, + { + "bbox": [ + 55, + 373, + 296, + 437 + ], + "type": "text", + "content": " and predictions " + }, + { + "bbox": [ + 55, + 373, + 296, + 437 + ], + "type": "inline_equation", + "content": "P_R" + }, + { + "bbox": [ + 55, + 373, + 296, + 437 + ], + "type": "text", + "content": " from redundancy graph " + }, + { + "bbox": [ + 55, + 373, + 296, + 437 + ], + "type": "inline_equation", + "content": "G_R" + }, + { + "bbox": [ + 55, + 373, + 296, + 437 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 94, + 446, + 295, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 446, + 295, + 462 + ], + "spans": [ + { + "bbox": [ + 94, + 446, + 295, + 462 + ], + "type": "interline_equation", + "content": "R _ {R} = Z D _ {R} ^ {- 1 / 2} G _ {R} D _ {R} ^ {- 1 / 2}; P _ {R} = R _ {R} \\theta^ {h} \\tag {5}", + "image_path": "8c53eaa90e593fa49d70503abbe3de128f9816c0b1dfb6d171146e48d31ad715.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 473, + 296, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 296, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 296, + 617 + ], + "type": "text", + "content": "Therefore, we decompose the model's original presentation and prediction into two components: the attention part and the redundancy part. This decomposition enables a distinct focus on important relationships while identifying redundant ones. Specifically, " + }, + { + "bbox": [ + 55, + 473, + 296, + 617 + ], + "type": "inline_equation", + "content": "R_{A}" + }, + { + "bbox": [ + 55, + 473, + 296, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 473, + 296, + 617 + ], + "type": "inline_equation", + "content": "P_{A}" + }, + { + "bbox": [ + 55, + 473, + 296, + 617 + ], + "type": "text", + "content": " focus solely on the feature relationships within the attention part, while " + }, + { + "bbox": [ + 55, + 473, + 296, + 617 + ], + "type": "inline_equation", + "content": "R_{R}" + }, + { + "bbox": [ + 55, + 473, + 296, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 473, + 296, + 617 + ], + "type": "inline_equation", + "content": "P_{R}" + }, + { + "bbox": [ + 55, + 473, + 296, + 617 + ], + "type": "text", + "content": " are concerned exclusively with the feature relationships within the redundancy part. Based on this decomposition, our goal is to ensure that the attention part adapts effectively to the target domain. Simultaneously, we aim to separate redundancy part from the attention component in both the representation and prediction layers." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 626, + 286, + 638 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 626, + 286, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 626, + 286, + 638 + ], + "type": "text", + "content": "4.2.3. Representation-Layer Redundancy Elimination" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 641, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 712 + ], + "type": "text", + "content": "With the above representations " + }, + { + "bbox": [ + 55, + 641, + 296, + 712 + ], + "type": "inline_equation", + "content": "R_{A}" + }, + { + "bbox": [ + 55, + 641, + 296, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 641, + 296, + 712 + ], + "type": "inline_equation", + "content": "R_{R}" + }, + { + "bbox": [ + 55, + 641, + 296, + 712 + ], + "type": "text", + "content": ", we expect that the attention representations possess class discriminability to handle label shift, while remaining distinct from redundancy representations to eliminate feature redundancy. To achieve this, we leverage contrastive learning [23, 30] as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 91, + 559, + 148 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 91, + 559, + 148 + ], + "spans": [ + { + "bbox": [ + 315, + 91, + 559, + 148 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {R} = \\\\ - \\sum_ {i = 1} ^ {n _ {t}} \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(R _ {A _ {i}} , c _ {o}\\right)\\right)}{\\sum_ {j = 1} ^ {C} \\exp \\left(\\operatorname {s i m} \\left(R _ {A _ {i}} , c _ {j}\\right)\\right) + \\exp \\left(\\operatorname {s i m} \\left(R _ {A _ {i}} , R _ {R _ {i}}\\right)\\right)} \\tag {6} \\\\ \\end{array}", + "image_path": "ebfc965bf324dcb5fdbcf2f43ced2f134b0176f46664546c66ec1362235a1d2d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 156, + 555, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 555, + 205 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 156, + 555, + 205 + ], + "type": "inline_equation", + "content": "\\mathrm{sim}(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 313, + 156, + 555, + 205 + ], + "type": "text", + "content": " denotes cosine similarity, " + }, + { + "bbox": [ + 313, + 156, + 555, + 205 + ], + "type": "inline_equation", + "content": "\\{c_j\\}_{j = 1}^C" + }, + { + "bbox": [ + 313, + 156, + 555, + 205 + ], + "type": "text", + "content": " denote the class centers in the attention representation space, " + }, + { + "bbox": [ + 313, + 156, + 555, + 205 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 156, + 555, + 205 + ], + "type": "text", + "content": " is the number of clusters, and " + }, + { + "bbox": [ + 313, + 156, + 555, + 205 + ], + "type": "inline_equation", + "content": "c_{o}" + }, + { + "bbox": [ + 313, + 156, + 555, + 205 + ], + "type": "text", + "content": " represents the center of the cluster that contains " + }, + { + "bbox": [ + 313, + 156, + 555, + 205 + ], + "type": "inline_equation", + "content": "R_{A_i}" + }, + { + "bbox": [ + 313, + 156, + 555, + 205 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 156, + 555, + 205 + ], + "type": "inline_equation", + "content": "c_{j}" + }, + { + "bbox": [ + 313, + 156, + 555, + 205 + ], + "type": "text", + "content": " is calculated as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 367, + 209, + 555, + 236 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 209, + 555, + 236 + ], + "spans": [ + { + "bbox": [ + 367, + 209, + 555, + 236 + ], + "type": "interline_equation", + "content": "c _ {j} = \\frac {\\sum_ {i} z _ {i} \\cdot \\mathbf {1} [ \\operatorname {a r g m a x} \\left(p _ {i}\\right) = j ]}{\\sum_ {i} \\mathbf {1} [ \\operatorname {a r g m a x} \\left(p _ {i}\\right) = j ]} \\tag {7}", + "image_path": "addd72451ce7ea9ba0e707767af2b404f8a9419b7fac780a27384b75cbd384a8.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 240, + 553, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 240, + 553, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 553, + 264 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 240, + 553, + 264 + ], + "type": "inline_equation", + "content": "\\mathbf{1}(\\cdot)" + }, + { + "bbox": [ + 313, + 240, + 553, + 264 + ], + "type": "text", + "content": " is an indicator function, outputting value 1 if the argument is true or 0 otherwise." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 264, + 554, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 554, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 554, + 346 + ], + "type": "text", + "content": "By minimizing " + }, + { + "bbox": [ + 313, + 264, + 554, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_R" + }, + { + "bbox": [ + 313, + 264, + 554, + 346 + ], + "type": "text", + "content": ", the model simultaneously enhances the discriminability of attention representations while ensuring their distinction from redundancy representations. This process enables the model to extract robust, discriminative, and non-redundant embedding features in the representation layer, facilitating effective adaptation to the target domain." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 353, + 524, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 353, + 524, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 353, + 524, + 365 + ], + "type": "text", + "content": "4.2.4. Prediction-Layer Redundancy Elimination" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 367, + 555, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 555, + 415 + ], + "type": "text", + "content": "In the prediction layer, to enhance the discriminability of attention predictions while reducing the influence of redundant predictions, we combine the principles of entropy minimization and negative learning into a unified loss function:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 420, + 557, + 463 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 420, + 557, + 463 + ], + "spans": [ + { + "bbox": [ + 314, + 420, + 557, + 463 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {P} = - \\sum_ {i = 1} ^ {N} \\sigma \\left(P _ {A _ {i}}\\right) \\log \\sigma \\left(P _ {A _ {i}}\\right) - \\sum_ {i = 1} ^ {N} \\sigma \\left(P _ {R _ {i}}\\right) \\log \\sigma \\left(1 - P _ {A _ {i}}\\right) \\tag {8}", + "image_path": "11d045dae8c1ae39648bc57ed8a3edf13c0455ff853a418ff6dc55e921639662.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 463, + 554, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 554, + 535 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 554, + 535 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 463, + 554, + 535 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 463, + 554, + 535 + ], + "type": "text", + "content": " denotes the softmax operation. Here, the first term minimizes the entropy of attention predictions " + }, + { + "bbox": [ + 313, + 463, + 554, + 535 + ], + "type": "inline_equation", + "content": "P_A" + }, + { + "bbox": [ + 313, + 463, + 554, + 535 + ], + "type": "text", + "content": ", encouraging sharper and more confident predictions, while the second term applies negative learning to redundant predictions " + }, + { + "bbox": [ + 313, + 463, + 554, + 535 + ], + "type": "inline_equation", + "content": "P_R" + }, + { + "bbox": [ + 313, + 463, + 554, + 535 + ], + "type": "text", + "content": ", penalizing them to ensure they remain distinct from the attention predictions." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 540, + 442, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 442, + 551 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 442, + 551 + ], + "type": "text", + "content": "4.2.5. Training Loss Function" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 554, + 554, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 554, + 554, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 554, + 554, + 578 + ], + "type": "text", + "content": "By combining Eq. (6) and Eq. (8), our overall loss function for G-FRET can be written as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 384, + 590, + 553, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 590, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 384, + 590, + 553, + 602 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {G F R E T} = \\mathcal {L} _ {R} + \\lambda \\mathcal {L} _ {P} \\tag {9}", + "image_path": "41f128a43cf712c66a0e7d6d2da753cc1666c7cf88dd9993977a9200b808a688.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 605, + 554, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 630 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 630 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 605, + 554, + 630 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 605, + 554, + 630 + ], + "type": "text", + "content": " is the balancing hyperparameter that controls the trade-off between " + }, + { + "bbox": [ + 313, + 605, + 554, + 630 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_R" + }, + { + "bbox": [ + 313, + 605, + 554, + 630 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 605, + 554, + 630 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_P" + }, + { + "bbox": [ + 313, + 605, + 554, + 630 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 630, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 712 + ], + "type": "text", + "content": "During testing, adaptation is performed online: upon receiving instances " + }, + { + "bbox": [ + 313, + 630, + 554, + 712 + ], + "type": "inline_equation", + "content": "\\{x_{i}\\}_{i = 1}^{n_{t}}" + }, + { + "bbox": [ + 313, + 630, + 554, + 712 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 313, + 630, + 554, + 712 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 630, + 554, + 712 + ], + "type": "text", + "content": ", the model is updated with parameters from the previous instances. The model then generates the prediction " + }, + { + "bbox": [ + 313, + 630, + 554, + 712 + ], + "type": "inline_equation", + "content": "\\{p_i = g(x_i)\\}_{i = 1}^{n_t}" + }, + { + "bbox": [ + 313, + 630, + 554, + 712 + ], + "type": "text", + "content": " for the new instances and updates using Eq. (9) with a single step of gradient descent. Adaptation continues as long as test data is available." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "2124" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 205, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 205, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 205, + 85 + ], + "type": "text", + "content": "5. Experiments and Analysis" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 187 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 187 + ], + "type": "text", + "content": "In this section, we conduct extensive experiments to explore the following key questions: (1) Is Test-Time Adaptation (TTA) based on feature redundancy elimination truly effective? (Sec. 5.2) (2) Do the underlying mechanisms of S-FRET and G-FRET align with our theoretical assumptions? (Sec. 5.3) (3) Additionally, we perform ablation studies and evaluate the scalability of our method in more complex scenarios. Sec. 5.4)" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 194, + 173, + 207 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 194, + 173, + 207 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 173, + 207 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 212, + 296, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 212, + 296, + 379 + ], + "spans": [ + { + "bbox": [ + 55, + 212, + 296, + 379 + ], + "type": "text", + "content": "We evaluate adaptation performance on two tasks: image domain adaptation and image corruption adaptation. For domain adaptation, we use the PACS [21] and OfficeHome [48] datasets; for corruption adaptation, we use CIFAR-10-C, CIFAR-100-C, and ImageNet-C [13]. To simulate label shifts, we modify CIFAR-100-C to follow a long-tail distribution, denoted as CIFAR-100-C-LT [7]. Comparison methods include BN [36], TENT [49], EATA [28], SAR [29], TSD [51], TIPI [27], and TEA [57], with ResNet-18 and ResNet-50 [12] as backbones. All results are averaged over 5 runs with different seeds, and each method undergoes independent hyperparameter tuning for fair comparison. Implementation details are provided in Sec. 10, and further results and analysis are in Sec. 11." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 385, + 278, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 385, + 278, + 398 + ], + "spans": [ + { + "bbox": [ + 55, + 385, + 278, + 398 + ], + "type": "text", + "content": "5.2. Comparison with State-of-the-art Methods" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 403, + 185, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 185, + 414 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 185, + 414 + ], + "type": "text", + "content": "5.2.1. Domain Generalization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 418, + 296, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 296, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 296, + 658 + ], + "type": "text", + "content": "Tab. 1 presents the accuracy comparison of S-FRET and G-FRET with other state-of-the-art TTA methods on the PACS and OfficeHome datasets based on ResNet-18 and ResNet-50 backbones. The table shows that S-FRET, through its simple yet effective optimization, attains competitive performance comparable to most existing TTA methods, while G-FRET achieves superior performance across both datasets and backbone types. Specifically, on the PACS dataset, G-FRET improves upon the baseline non-adaptive source model by " + }, + { + "bbox": [ + 55, + 418, + 296, + 658 + ], + "type": "inline_equation", + "content": "6.67\\%" + }, + { + "bbox": [ + 55, + 418, + 296, + 658 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 418, + 296, + 658 + ], + "type": "inline_equation", + "content": "5.45\\%" + }, + { + "bbox": [ + 55, + 418, + 296, + 658 + ], + "type": "text", + "content": " with ResNet-18 and ResNet-50, respectively. For OfficeHome, the improvements are " + }, + { + "bbox": [ + 55, + 418, + 296, + 658 + ], + "type": "inline_equation", + "content": "1.80\\%" + }, + { + "bbox": [ + 55, + 418, + 296, + 658 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 418, + 296, + 658 + ], + "type": "inline_equation", + "content": "2.34\\%" + }, + { + "bbox": [ + 55, + 418, + 296, + 658 + ], + "type": "text", + "content": ", respectively. Compared with state-of-the-art TTA methods, G-FRET achieves the highest or second-highest accuracy overall, particularly when using the ResNet-50 as backbone, G-FRET consistently achieves the highest accuracy across all domains. In contrast, other methods fail to achieve consistently high accuracy across different domains and backbones. This is caused by G-FRET's ability to consider both feature discriminability and redundancy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 662, + 162, + 674 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 662, + 162, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 662, + 162, + 674 + ], + "type": "text", + "content": "5.2.2. Image Corruption" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Tab. 2 compares the performance of S-FRET, G-FRET, and other state-of-the-art methods on CIFAR-10-C, CIFAR100-C, and ImageNet-C at corruption level 5. Two set" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 315, + 71, + 554, + 173 + ], + "blocks": [ + { + "bbox": [ + 315, + 71, + 554, + 173 + ], + "lines": [ + { + "bbox": [ + 315, + 71, + 554, + 173 + ], + "spans": [ + { + "bbox": [ + 315, + 71, + 554, + 173 + ], + "type": "table", + "html": "
BackboneMethodPACSAvgOfficeHomeAvg
ACPPSACPRR
ResNet-18Source [12]78.37±0.0077.20±0.0095.03±0.0076.54±0.0081.84±0.0056.45±0.0048.02±0.0071.34±0.0072.23±0.0062.01±0.00
BN [36]81.01±0.0080.67±0.2795.14±0.1573.83±0.1882.66±0.0855.66±0.1749.22±0.2370.77±0.1072.45±0.1962.01±0.12
SAR [29]83.62±0.0081.47±0.0095.14±0.1580.44±0.1585.17±0.0057.30±0.1550.54±0.2570.71±0.1572.75±0.2562.82±0.12
EATA [28]82.06±0.8581.11±1.0395.16±0.2973.39±1.0082.93±0.4656.42±0.4149.75±0.4171.50±0.3572.85±0.4162.63±0.11
TENT [49]83.54±1.7882.26±1.6995.54±0.3181.05±1.7885.60±0.7557.25±0.3250.45±0.2672.18±0.3873.09±0.3463.24±0.13
TSD [51]87.05±0.4786.94±0.6295.98±0.2582.56±1.1286.13±0.4457.80±0.3450.03±0.6971.45±0.4270.92±0.4662.55±0.06
TEA [57]87.04±0.5486.54±0.8596.17±0.3682.17±1.0887.98±0.4857.07±0.8850.47±0.1871.84±0.1872.85±0.1063.06±0.27
TTP [57]84.77±0.1883.48±0.3896.17±0.1884.85±0.2584.85±0.2557.25±0.3250.45±0.2671.45±0.3572.75±0.2562.62±0.19
S-FRET86.25±0.3086.66±0.3795.96±0.2575.90±0.2486.20±0.6656.26±0.2149.86±0.3871.65±0.0372.72±0.1462.62±0.09
G-FRET87.20±0.3587.19±0.2496.28±0.3683.36±1.4388.51±0.3957.49±0.3451.28±0.4273.04±0.2973.40±0.3563.81±0.11
ResNet-50Source [12]83.89±0.0081.02±0.0096.17±0.0078.04±0.0084.78±0.0064.85±0.0052.20±0.0075.04±0.0075.88±0.0067.01±0.00
BN [36]85.19±0.3585.49±0.3096.60±0.1972.25±0.3384.88±0.0863.30±0.2053.02±0.3273.59±0.2375.10±0.2966.25±0.05
SAR [29]85.24±0.3785.49±0.3096.60±0.1976.02±0.8485.84±0.2664.78±0.1756.30±0.5774.83±0.2776.18±0.3068.02±0.19
EATA [28]85.14±0.3185.36±0.2096.54±0.1972.27±0.1184.83±0.1164.25±0.4154.09±0.5374.25±0.3675.72±0.4367.08±0.22
TTP [51]84.77±0.1884.97±0.1996.17±0.2978.53±0.3785.66±0.2563.33±0.3054.99±0.3474.16±0.2975.16±0.2968.74±0.14
TSD [51]80.93±0.8380.94±0.2797.56±0.2781.56±1.5289.97±0.1865.33±0.3056.99±0.3476.12±0.3676.52±0.2468.74±0.14
TEA [57]88.51±0.4387.75±0.5197.17±0.2481.45±1.0288.72±0.2665.62±0.4557.55±0.6475.80±0.3976.85±0.5169.25±0.25
TTP [27]87.62±0.7386.97±0.9697.01±1.1179.66±1.5087.81±0.5664.92±0.2956.47±0.3775.31±0.2776.74±0.3768.36±0.04
S-FRET89.90±0.6089.00±0.4897.46±0.3277.45±0.9188.42±0.1964.42±0.2055.00±0.3475.56±0.2476.41±0.1467.80±0.13
G-FRET90.99±0.8790.00±0.2597.59±0.2982.32±0.5990.23±0.5366.12±0.4157.63±0.3976.46±0.3477.19±0.2369.35±0.20
", + "image_path": "a2ef422711455d64288d1a4e634a575df47c4198505be2a6655d91c251030807.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 251, + 555, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 555, + 431 + ], + "type": "text", + "content": "tings are used: continuous adaptation for CIFAR-10/100-C (15 corruption types applied sequentially), and independent adaptation for ImageNet-C (each corruption type evaluated separately). S-FRET significantly outperforms the non-adaptive baseline, with average gains of " + }, + { + "bbox": [ + 313, + 251, + 555, + 431 + ], + "type": "inline_equation", + "content": "22.84\\%" + }, + { + "bbox": [ + 313, + 251, + 555, + 431 + ], + "type": "text", + "content": " on CIFAR-10-C, " + }, + { + "bbox": [ + 313, + 251, + 555, + 431 + ], + "type": "inline_equation", + "content": "17.52\\%" + }, + { + "bbox": [ + 313, + 251, + 555, + 431 + ], + "type": "text", + "content": " on CIFAR-100-C, and " + }, + { + "bbox": [ + 313, + 251, + 555, + 431 + ], + "type": "inline_equation", + "content": "14.4\\%" + }, + { + "bbox": [ + 313, + 251, + 555, + 431 + ], + "type": "text", + "content": " on ImageNet-C. G-FRET consistently surpasses S-FRET by leveraging both feature redundancy and label discriminability, whereas S-FRET only addresses redundancy. Under continuous adaptation, methods not explicitly reducing redundancy perform similarly to S-FRET but worse than G-FRET. This highlights the complementary roles of redundancy reduction and class discriminability in improving TTA performance—a finding further supported by our ablation studies." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 316, + 438, + 555, + 544 + ], + "blocks": [ + { + "bbox": [ + 313, + 180, + 555, + 236 + ], + "lines": [ + { + "bbox": [ + 313, + 180, + 555, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 555, + 236 + ], + "type": "text", + "content": "Table 1. Accuracy comparison of different TTA methods on PACS and OfficeHome datasets based on ResNet-18 and ResNet-50 backbones, including mean and standard deviation of 5 runs with different random seeds. The best results are highlighted in boldface, and the second ones are underlined." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 438, + 555, + 544 + ], + "lines": [ + { + "bbox": [ + 316, + 438, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 555, + 544 + ], + "type": "table", + "html": "
MethodCIFAR-10-CCIFAR-100-CImageNet-C
1-78-151-151-78-151-151-78-151-15
Source [12]38.4661.5950.8023.5937.4931.017.7420.7914.70
BN [36]70.4776.4373.6545.7050.5748.3017.8436.1227.59
TENT [49]73.0679.2576.3647.7753.4450.7927.0542.6935.39
EATA [28]70.6076.4273.7047.6453.4050.7131.2346.1539.19
SAR [29]70.6176.5073.7546.6651.2149.0930.3745.7038.55
TIPI [27]73.7680.2177.2147.5052.0649.9327.7842.6935.73
TEA [57]71.7077.4774.7846.0750.6448.5122.8540.7232.38
TSD [51]72.5377.9675.4345.8751.0348.6220.9338.1530.11
S-FRET71.6077.2874.6346.3150.4648.5319.5937.4329.10
G-FRET73.7680.3277.2648.2353.5551.0628.6043.9636.79
", + "image_path": "e141d0bd97b82b9850b0b83c8ace3b27ce66811c9724f037adf3281c616a95b2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 552, + 554, + 597 + ], + "lines": [ + { + "bbox": [ + 313, + 552, + 554, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 554, + 597 + ], + "type": "text", + "content": "Table 2. Accuracy comparison of different TTA methods on CIFAR-10/100-C and ImageNet-C datasets at damage level of 5, with 15 types of damage applied two experimental settings: continuous adaptation and independent adaptation." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 603, + 440, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 603, + 440, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 603, + 440, + 615 + ], + "type": "text", + "content": "5.2.3. Efficiency Comparison" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 618, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 554, + 715 + ], + "type": "text", + "content": "In addition to accuracy, we also evaluate the runtime efficiency of different methods to assess the computational performance of S-FRET and G-FRET. As shown in Tab. 3, \"Training\" indicates whether the method involves retraining. ERM represents the source model without adaptation, while the BN method directly updates the batch normalization (BN) parameters, eliminating the need for retraining. Consequently, BN is the most computationally ef" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "2125" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 70, + 297, + 138 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 138 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 138 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 138 + ], + "type": "table", + "html": "
Cost Time (s)
BackboneERMBNTENTEATASARTSDTIPITEAS-FRETG-FRET
Traning××✓✓✓✓✓✓✓✓
ResNet-18393648567411416265846118
ResNet-5064661131091851532641376102193
VIT-B/161411622102203452952813181185310
AVG81881241282011872361738111207
", + "image_path": "843d1baee09aa7806517ecf957c99552e5386b8b52f6edd8e05a7d5d051ddeb7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 55, + 139, + 297, + 197 + ], + "blocks": [ + { + "bbox": [ + 55, + 139, + 297, + 197 + ], + "lines": [ + { + "bbox": [ + 55, + 139, + 297, + 197 + ], + "spans": [ + { + "bbox": [ + 55, + 139, + 297, + 197 + ], + "type": "table", + "html": "
Maximum GPU memory usage (MB)
Backbone TrainingSource ×BN ×TENT ✓EATA ✓SAR ✓TSD ✓TIPI ✓TEA ✓S-FRET ✓G-FRET ✓
ResNet-1890390427893713354831703427822127894781
ResNet-50772773541161536005635660711595154489163
VIT-B/161218-64657324662010584137651730064698045
AVG965-488857305391670377551382449027330
", + "image_path": "9f21847f3f53b1c25de002eac37afc0bbf502488b452e403243a936ad5195370.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 229, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 229, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 54, + 229, + 296, + 361 + ], + "type": "text", + "content": "ficient method. However, as reported in Tabs. 1 and 2, BN consistently exhibits the lowest accuracy. Among the training methods, S-FRET achieves the lowest runtime and the second-lowest memory consumption across various backbones, while G-FRET exhibits average performance compared to other baselines. These results indicate that S-FRET, as a simple yet effective approach, provides a competitive accuracy-efficiency trade-off. Meanwhile, G-FRET, despite its moderate computational cost, achieves the highest accuracy, demonstrating its effectiveness in balancing efficiency and adaptation performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 370, + 247, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 370, + 247, + 383 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 247, + 383 + ], + "type": "text", + "content": "5.3. Analysis of Underlying Mechanisms" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 388, + 295, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 295, + 412 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 295, + 412 + ], + "type": "text", + "content": "5.3.1. Relation between Feature Redundancy Elimination and Generalizability Enhancement" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 415, + 296, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 296, + 619 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 296, + 619 + ], + "type": "text", + "content": "The first mechanism to be validated is whether there exists a genuine relationship between feature redundancy and generalizability enhancement. We track the evolution of redundancy scores (RS), loss, and accuracy across increasing adaptation steps. As shown in Fig. 4, S-FRET achieves a consistent reduction in feature redundancy as adaptation progresses, accompanied by stable accuracy improvement, demonstrating the detrimental effect of redundancy on generalization and the efficacy of redundancy elimination. For G-FRET, as the number of adaptation steps increases, there is also a consistent reduction in the redundancy of the extracted embedded features, accompanied by a significant decrease in loss and a notable improvement in accuracy. In summary, both S-FRET and G-FRET has been demonstrated to be highly effective in minimizing redundancy, and redundancy elimination is strongly associated with enhancement of model generalizability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 626, + 132, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 626, + 132, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 626, + 132, + 637 + ], + "type": "text", + "content": "5.3.2. Label Shift" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 641, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 712 + ], + "type": "text", + "content": "Is S-FRET truly affected by label shift? Tab. 4 presents the performance of S-FRET and G-FRET under test-time label shifts with long-tailed distributions. Here, the imbalance factor denotes the ratio between the most and least frequent class samples—higher values indicate more skewed class distributions. As shown in Tab. 4, S-FRET achieves" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 70, + 425, + 148 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 425, + 148 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 425, + 148 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 425, + 148 + ], + "type": "image", + "image_path": "6c6b5265d711d6457d126007c110dfeff8d1862945693eed37960cf10550ff8c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 433, + 70, + 542, + 148 + ], + "blocks": [ + { + "bbox": [ + 433, + 70, + 542, + 148 + ], + "lines": [ + { + "bbox": [ + 433, + 70, + 542, + 148 + ], + "spans": [ + { + "bbox": [ + 433, + 70, + 542, + 148 + ], + "type": "image", + "image_path": "c23fa039f6d590527edb54ccc5d1d6eaf1d786be7889ed8aef9e563eea240d89.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 149, + 435, + 227 + ], + "blocks": [ + { + "bbox": [ + 317, + 149, + 435, + 227 + ], + "lines": [ + { + "bbox": [ + 317, + 149, + 435, + 227 + ], + "spans": [ + { + "bbox": [ + 317, + 149, + 435, + 227 + ], + "type": "image", + "image_path": "650d8b70e82159da38af4a0948d2ae08e21f8e9ca1d932842fdc69326e495eaf.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 236, + 555, + 268 + ], + "lines": [ + { + "bbox": [ + 313, + 236, + 555, + 268 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 555, + 268 + ], + "type": "text", + "content": "Figure 4. The relationship of the reduction of redundancy and the enhancement of generalizability attained by S-FRET and G-FRET on CIFAR-10-C and CIFAR-100-C datasets." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 436, + 150, + 552, + 227 + ], + "blocks": [ + { + "bbox": [ + 436, + 150, + 552, + 227 + ], + "lines": [ + { + "bbox": [ + 436, + 150, + 552, + 227 + ], + "spans": [ + { + "bbox": [ + 436, + 150, + 552, + 227 + ], + "type": "image", + "image_path": "34eb8cf712c46e6b4bac357f0b5afa0656e0a287feaef868635eee2574bfc985.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 315, + 275, + 556, + 327 + ], + "blocks": [ + { + "bbox": [ + 55, + 198, + 295, + 220 + ], + "lines": [ + { + "bbox": [ + 55, + 198, + 295, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 295, + 220 + ], + "type": "text", + "content": "Table 3. Efficiency comparison of TTA methods on ImageNet-C (Gaussian Noise) across multiple backbone networks" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 275, + 556, + 327 + ], + "lines": [ + { + "bbox": [ + 315, + 275, + 556, + 327 + ], + "spans": [ + { + "bbox": [ + 315, + 275, + 556, + 327 + ], + "type": "table", + "html": "
CIFAR-100-C-LT
Imbalance FactorERMBNSAREATATENTTSDTIPITEAS-FRETG-FRET
131.0148.3049.0950.7150.7948.6248.5149.9348.5351.06
1030.3947.1447.6947.8650.1247.7447.5049.5447.4750.23
10030.1346.0748.2146.0849.8046.9547.1349.5146.6250.27
AVG30.5147.1748.3348.2150.2447.7747.7149.6647.5450.52
", + "image_path": "2318bd43fa8472ac48e1d5a03c6e6ee98311079a777ec1952892df1a9f511466.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 335, + 554, + 357 + ], + "lines": [ + { + "bbox": [ + 313, + 335, + 554, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 554, + 357 + ], + "type": "text", + "content": "Table 4. Performance comparison on CIFAR-100-C-LT with different imbalance factors based on ResNet-18." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 365, + 555, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 461 + ], + "type": "text", + "content": "competitive performance on CIFAR-100-C under balanced class distributions. However, its accuracy degrades significantly as the imbalance factor increases, with " + }, + { + "bbox": [ + 313, + 365, + 555, + 461 + ], + "type": "inline_equation", + "content": "1.91\\%" + }, + { + "bbox": [ + 313, + 365, + 555, + 461 + ], + "type": "text", + "content": " drops on CIFAR-100-C, aligning with the limitations of conventional TTA methods. In contrast, G-FRET delivers more robust performance across all imbalance factors. This highlights G-FRET's unique capability to address label shift in real-world scenarios through contrastive learning." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 467, + 489, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 489, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 489, + 479 + ], + "type": "text", + "content": "5.3.3. Discriminability Analysis by tSNE" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 482, + 555, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 555, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 555, + 626 + ], + "type": "text", + "content": "Does G-FRET truly improve discriminability compared to S-FRET? Fig. 5 shows the t-SNE [46] visualization of embedded features before and after adaptation by S-FRET and G-FRET. It is evident that before adaptation, data points from different classes are intermixed without clear separation. After adaptation by S-FRET, although class separation increases, the overlapping regions between classes still lack distinct boundaries. In contrast, G-FRET achieves more distinct cluster separation and sharper decision boundaries, demonstrating its superior capability in learning discriminative features under distribution shifts. Extended experimental results on the CIFAR-10-C are provided in Sec. 11.3." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 633, + 477, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 633, + 477, + 646 + ], + "spans": [ + { + "bbox": [ + 313, + 633, + 477, + 646 + ], + "type": "text", + "content": "5.4. Ablation and Extension Study" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 651, + 490, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 651, + 490, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 651, + 490, + 662 + ], + "type": "text", + "content": "5.4.1. Effects of Components in G-FRET" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "Tab. 5 presents the impact of the four components of G-FRET on model performance. (1) Compared to the source baseline, each individual component contributes positively to the model's performance, with respective gains of " + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "inline_equation", + "content": "0.30\\%" + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2126" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 70, + 282, + 152 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 282, + 152 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 282, + 152 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 282, + 152 + ], + "type": "image", + "image_path": "0e5f608ff42ee0fd764a83fc8a4fc3bcd272cafcca40fa70c73cc7306cbb583f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 160, + 295, + 194 + ], + "lines": [ + { + "bbox": [ + 55, + 160, + 295, + 194 + ], + "spans": [ + { + "bbox": [ + 55, + 160, + 295, + 194 + ], + "type": "text", + "content": "Figure 5. Discriminability visualization of embedded feature before and after adaptation by S-FRET and G-FRET on the PACS dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 197, + 294, + 345 + ], + "blocks": [ + { + "bbox": [ + 58, + 197, + 294, + 345 + ], + "lines": [ + { + "bbox": [ + 58, + 197, + 294, + 345 + ], + "spans": [ + { + "bbox": [ + 58, + 197, + 294, + 345 + ], + "type": "table", + "html": "
OfficeHome
ComponentsACPRAVG↑
Source56.4548.0271.3472.2362.010.00
IndividualLRA55.8349.6770.8372.8962.310.30
R56.9850.4272.3372.7363.121.11
LPA56.7350.7772.1372.9963.161.15
R55.8349.3070.6272.5162.070.06
Ablationw/o LRA57.7251.1872.9973.0863.741.73
R56.7051.1672.0773.1963.281.27
w/o LPA57.1150.5272.4072.9463.241.23
R57.7751.4072.9472.9663.771.76
AttentionLR(A)+LP(A)56.7050.9772.3673.1263.291.28
RedundantLR(R)+LP(R)57.0350.4972.2772.9463.181.17
AllAll57.4951.2873.0474.4063.811.80
", + "image_path": "105868d4fcfca7721b115c65b697fde154470f93009aa89bf97ddad3a8045421.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 402, + 295, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 402, + 295, + 571 + ], + "spans": [ + { + "bbox": [ + 54, + 402, + 295, + 571 + ], + "type": "inline_equation", + "content": "1.11\\%" + }, + { + "bbox": [ + 54, + 402, + 295, + 571 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 402, + 295, + 571 + ], + "type": "inline_equation", + "content": "1.15\\%" + }, + { + "bbox": [ + 54, + 402, + 295, + 571 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 402, + 295, + 571 + ], + "type": "inline_equation", + "content": "0.06\\%" + }, + { + "bbox": [ + 54, + 402, + 295, + 571 + ], + "type": "text", + "content": ". Notably, the redundancy component in the representation layer, " + }, + { + "bbox": [ + 54, + 402, + 295, + 571 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_R(R)" + }, + { + "bbox": [ + 54, + 402, + 295, + 571 + ], + "type": "text", + "content": ", and the attention component in the prediction layer, " + }, + { + "bbox": [ + 54, + 402, + 295, + 571 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_P(A)" + }, + { + "bbox": [ + 54, + 402, + 295, + 571 + ], + "type": "text", + "content": ", show the most significant impact, with " + }, + { + "bbox": [ + 54, + 402, + 295, + 571 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_R(R)" + }, + { + "bbox": [ + 54, + 402, + 295, + 571 + ], + "type": "text", + "content": " achieving a " + }, + { + "bbox": [ + 54, + 402, + 295, + 571 + ], + "type": "inline_equation", + "content": "1.15\\%" + }, + { + "bbox": [ + 54, + 402, + 295, + 571 + ], + "type": "text", + "content": " improvement. This highlights the importance of redundancy reduction for TTA. (2) The ablation study further confirms the significance of each component, indicating that their removal negatively impacts model performance. (3) Either the redundant or attention components can individually improves model performance, with respective gains of " + }, + { + "bbox": [ + 54, + 402, + 295, + 571 + ], + "type": "inline_equation", + "content": "1.28\\%" + }, + { + "bbox": [ + 54, + 402, + 295, + 571 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 402, + 295, + 571 + ], + "type": "inline_equation", + "content": "1.17\\%" + }, + { + "bbox": [ + 54, + 402, + 295, + 571 + ], + "type": "text", + "content": ". When the two components are jointly considered, they yield an even greater improvement. (4) Overall, the experimental results further validate the necessity of considering feature redundancy during the TTA process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 586, + 176, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 176, + 600 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 176, + 600 + ], + "type": "text", + "content": "5.4.2. Parameter Sensitivity" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 295, + 713 + ], + "type": "text", + "content": "Our method G-FRET involves one hyperparameter: the trade-off parameter in Eq. (9), " + }, + { + "bbox": [ + 54, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 54, + 605, + 295, + 713 + ], + "type": "text", + "content": ". As shown in Fig. 6, G-FRET outperforms the source model across a wide range of values for " + }, + { + "bbox": [ + 54, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 54, + 605, + 295, + 713 + ], + "type": "text", + "content": ". We also investigate the impact of the learning rate (lr) on the experimental results. The best performance is achieved with a learning rate of around " + }, + { + "bbox": [ + 54, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 54, + 605, + 295, + 713 + ], + "type": "text", + "content": "; higher learning rates tend to destabilize the training process and reduce the effectiveness of model, which is also observed in other TTA methods [27-29, 49, 51, 57]." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 340, + 72, + 436, + 156 + ], + "blocks": [ + { + "bbox": [ + 340, + 72, + 436, + 156 + ], + "lines": [ + { + "bbox": [ + 340, + 72, + 436, + 156 + ], + "spans": [ + { + "bbox": [ + 340, + 72, + 436, + 156 + ], + "type": "image", + "image_path": "f944b598b12597c7347c86bc326038e3545a79d13d558ff2d9debf695301f638.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 168, + 553, + 196 + ], + "lines": [ + { + "bbox": [ + 313, + 168, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 553, + 196 + ], + "type": "text", + "content": "Figure 6. Parameter sensitivity analysis G-FRET on the Office-Home dataset with ResNet-18 backbone in domain Art. Results on complex/large dataset." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 440, + 72, + 529, + 157 + ], + "blocks": [ + { + "bbox": [ + 440, + 72, + 529, + 157 + ], + "lines": [ + { + "bbox": [ + 440, + 72, + 529, + 157 + ], + "spans": [ + { + "bbox": [ + 440, + 72, + 529, + 157 + ], + "type": "image", + "image_path": "682e904363ec7b741ee93f836a6f1258a5f181409bfd8774db041464b98c81ae.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 316, + 196, + 553, + 322 + ], + "blocks": [ + { + "bbox": [ + 55, + 354, + 295, + 387 + ], + "lines": [ + { + "bbox": [ + 55, + 354, + 295, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 295, + 387 + ], + "type": "text", + "content": "Table 5. Effects of components in G-FRET, test on the Office-Home dataset with ResNet-18 backbone. A: Attention part, R: Redundancy Part." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 196, + 553, + 322 + ], + "lines": [ + { + "bbox": [ + 316, + 196, + 553, + 322 + ], + "spans": [ + { + "bbox": [ + 316, + 196, + 553, + 322 + ], + "type": "table", + "html": "
DatasetsResNet-18[12]+S-FRET+G-FRETResNet-50[12]+S-FRET+G-FRETViT-B/16[6]+S-FRET+G-FRET
VLCS[45]73.6773.9075.7374.9075.4876.4276.4477.1878.56
DomainNet[34]39.1339.2640.2243.5843.6944.2249.9851.2251.50
ImageNet-C[13]14.7029.1036.7918.1534.6542.8439.8360.3662.50
Results with ViT-B/16 as backbone.
DatasetsSourceTENTPLCEATASARTIPITEATSDDeYOS-FRETG-FRET
CIFAR10-C71.4871.4871.4673.6771.4871.4871.4871.4777.8677.9573.17
CIFAR10-C-CO71.4871.4671.4472.5470.3371.4571.6371.6974.6475.4774.02
CIFAR100-C51.7152.7251.7062.4754.2952.3652.3151.7459.5763.4657.65
CIFAR100-C-CO51.7151.7751.7956.2652.5051.7951.9251.9054.0756.6952.14
PACS87.0387.7386.1787.0787.1287.8787.3783.4388.0787.1988.12
OfficeHome76.1176.8676.5076.7476.6676.9876.1575.4976.7876.5577.62
AVG68.2568.6768.1871.4668.7368.6568.4867.6271.8372.8870.45
Results on Multimodal task
DatasetsCLIPTENTTPTCLIPArTTWATT-PWATT-SS-FRET
VLCS80.9680.8572.5380.8981.0581.5981.69
PACS95.1295.2293.0895.3596.0796.0896.09
OfficeHome80.2880.4781.9180.4281.8281.7181.67
AVG85.4585.5182.5185.5586.3186.4686.48
", + "image_path": "54db8fd2247d7710134b2c99da50ba2ff596a7a611a254f191df36a1d1ecc6d6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 331, + 553, + 353 + ], + "lines": [ + { + "bbox": [ + 313, + 331, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 553, + 353 + ], + "type": "text", + "content": "Table 6. Scalability on complex/large datasets, different backbones and multimoda task." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 363, + 553, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 553, + 387 + ], + "type": "text", + "content": "5.4.3. Scalability on complex/large datasets, different backbones and multimoda task." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 392, + 555, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 555, + 536 + ], + "type": "text", + "content": "We validate our method on large-scale datasets, diverse backbones, and a multimodal task to demonstrate its robustness and architecture-independence. As shown in Tab. 6, reducing embedding redundancy leads to notably larger gains on ViTs than on CNNs, likely due to the higher redundancy in Transformer representations. In the multimodal CLIP setting [31], S-FRET continues to deliver consistent improvements by explicitly aligning embedding distributions, unlike CLIP's similarity-based matching. Notably, on CIFAR with ViT, S-FRET significantly outperforms G-FRET, suggesting that the effectiveness of redundancy reduction varies across backbone types and datasets." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 548, + 388, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 388, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 388, + 562 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "We propose a novel approach for test-time adaptation (TTA) — Feature Redundancy Elimination for TTA (FRET) — which focuses on reducing feature redundancy during testing to address distribution shifts. This is the first attempt to leverage feature redundancy for TTA. We introduce two methods: S-FRET, which directly minimizes the feature redundancy score, and G-FRET, which optimizes both redundancy elimination and class discriminability by decomposing feature relations into attention and redundancy components. Experiments across various tasks and datasets show the effectiveness of our methods, demonstrating the value of feature redundancy elimination in TTA." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2127" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 294, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 294, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 294, + 85 + ], + "type": "text", + "content": "Author Contributions and Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 199 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 199 + ], + "type": "text", + "content": "Linjing You initiated this work and was primarily responsible for proposing the core idea, methodology, theoretical analysis, experimental design, and discussion of key findings. Jiabao Lu conducted most of the experiments and contributed to the interpretation of experimental results. Xiayuan Huang participated in experimental discussions and provided feedback on the initial manuscript. Xiangli Nie offered valuable suggestions for the refinement and polishing of the paper." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 199, + 295, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 199, + 295, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 199, + 295, + 247 + ], + "type": "text", + "content": "We would also like to sincerely thank all the anonymous reviewers and the editorial committee (ACS) whose insightful comments significantly improved the theoretical rigor, experimental depth, and overall quality of this work." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 247, + 296, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 296, + 330 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 296, + 330 + ], + "type": "text", + "content": "This research was supported by Prof. Huang's project The Fundamental Research Funds for the Central Universities (Grant No. XJJSKYQD202538), Prof. Nie's project The National Key Research and Development Program of China (Grant No. 2024YFB4709100), and partially funded by the Beijing Nova Program (Grant Nos. 20220484070 and 20240484498)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 339, + 115, + 352 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 339, + 115, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 339, + 115, + 352 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 359, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 359, + 295, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 359, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 61, + 359, + 295, + 415 + ], + "type": "text", + "content": "[1] Ivan Anishchenko, Samuel J Pellock, Tamuka M Chidyausiku, Theresa A Ramelot, Sergey Ovchinnikov, Jingzhou Hao, Khushboo Bafna, Christoffer Norn, Alex Kang, Asim K Bera, et al. De novo protein design by deep network hallucination. Nature, 600(7889):547-552, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 415, + 295, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 415, + 295, + 459 + ], + "spans": [ + { + "bbox": [ + 61, + 415, + 295, + 459 + ], + "type": "text", + "content": "[2] Gavin Brown, Adam Pocock, Ming-Jie Zhao, and Mikel Luján. Conditional likelihood maximisation: a unifying framework for information theoretic feature selection. The journal of machine learning research, 13(1):27-66, 2012. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 459, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 459, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 62, + 459, + 294, + 514 + ], + "type": "text", + "content": "[3] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label distribution-aware margin loss. In Proc. the Advances in Neural Information Processing Systems (NeurIPS), pages 1567-1578, 2019. 2, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 514, + 295, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 295, + 557 + ], + "type": "text", + "content": "[4] Xue-wen Chen and Jong Cheol Jeong. Enhanced recursive feature elimination. In Sixth international conference on machine learning and applications (ICMLA 2007), pages 429-435. IEEE, 2007. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 559, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 559, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 559, + 295, + 612 + ], + "type": "text", + "content": "[5] Ian Connick Covert, Wei Qiu, Mingyu Lu, Na Yoon Kim, Nathan J White, and Su-In Lee. Learning to maximize mutual information for dynamic feature selection. In International Conference on Machine Learning, pages 6424-6447. PMLR, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 614, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 614, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 62, + 614, + 295, + 646 + ], + "type": "text", + "content": "[6] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 647, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 295, + 690 + ], + "type": "text", + "content": "[7] C. Du, Y. Wang, S. Song, and G. Huang. Probabilistic contrastive learning for long-tailed visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(9):5890-5904, 2024. 6, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 62, + 691, + 295, + 714 + ], + "type": "text", + "content": "[8] Tongtong Fang, Nan Lu, Gang Niu, and Masashi Sugiyama. Rethinking importance weighting for deep learning under" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "distribution shift. Advances in neural information processing systems, 33:11996-12007, 2020. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 319, + 97, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 97, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 319, + 97, + 553, + 128 + ], + "type": "text", + "content": "[9] Ronald A Fisher. The use of multiple measurements in taxonomic problems. Annals of eugenics, 7(2):179-188, 1936. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 130, + 553, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 174 + ], + "type": "text", + "content": "[10] Taesik Gong, Yewon Kim, Taeckyung Lee, Sorn Chottananurak, and Sung-Ju Lee. Sotta: Robust test-time adaptation on noisy data streams. Advances in Neural Information Processing Systems, 36, 2024. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 175, + 553, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 219 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 219 + ], + "type": "text", + "content": "[11] Jie Gui, Zhenan Sun, Shuiwang Ji, Dacheng Tao, and Tieniu Tan. Feature selection based on structured sparsity: A comprehensive study. IEEE transactions on neural networks and learning systems, 28(7):1490-1507, 2016. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 220, + 553, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 263 + ], + "type": "text", + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 2, 6, 8, 3, 7, 9, 10" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 265, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 265, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 265, + 553, + 297 + ], + "type": "text", + "content": "[13] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019. 2, 6, 8, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 300, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 300, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 300, + 553, + 342 + ], + "type": "text", + "content": "[14] Chenping Hou, Ruidong Fan, Ling-Li Zeng, and Dewen Hu. Adaptive feature selection with augmented attributes. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):9306-9324, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 344, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 344, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 344, + 553, + 398 + ], + "type": "text", + "content": "[15] John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Žídek, Anna Potapenko, et al. Highly accurate protein structure prediction with alphafold. nature, 596(7873):583-589, 2021. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 399, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 454 + ], + "type": "text", + "content": "[16] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. In Proc. the International Conference on Learning Representations (ICLR), 2019. 2, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 456, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 487 + ], + "type": "text", + "content": "[17] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 489, + 553, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 489, + 553, + 511 + ], + "spans": [ + { + "bbox": [ + 316, + 489, + 553, + 511 + ], + "type": "text", + "content": "[18] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. nature, 521(7553):436-444, 2015. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 513, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 513, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 513, + 553, + 567 + ], + "type": "text", + "content": "[19] Jonghyun Lee, Dahuin Jung, Saehyung Lee, Junsung Park, Juhyeon Shin, Uiwon Hwang, and Sungroh Yoon. Entropy is not enough for test-time adaptation: From the perspective of disentangled factors. In The Twelfth International Conference on Learning Representations, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 569, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 569, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 569, + 553, + 622 + ], + "type": "text", + "content": "[20] Taeckyung Lee, Sorn Chottananurak, Taesik Gong, and Sung-Ju Lee. Aetta: Label-free accuracy estimation for test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28643-28652, 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 624, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 668 + ], + "type": "text", + "content": "[21] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. In Proceedings of the IEEE international conference on computer vision, pages 5542-5550, 2017. 6, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[22] Jundong Li, Kewei Cheng, Suhang Wang, Fred Morstatter, Robert P Trevino, Jiliang Tang, and Huan Liu. Feature selection: A data perspective. ACM computing surveys (CSUR), 50(6):1-45, 2017. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2128" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[23] Junnan Li, Pan Zhou, Caiming Xiong, and Steven C.H. Hoi. Prototypical contrastive learning of unsupervised representations. In ICLR, 2021. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "text", + "content": "[24] Jian Liang, Ran He, and Tieniu Tan. A comprehensive survey on test-time adaptation under distribution shifts. International Journal of Computer Vision, pages 1-34, 2024. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 218 + ], + "type": "text", + "content": "[25] Duy MH Nguyen, Hoang Nguyen, Nghiem Diep, Tan Ngoc Pham, Tri Cao, Binh Nguyen, Paul Swoboda, Nhat Ho, Shadi Albarqouni, Pengtao Xie, et al. Lvm-med: Learning large-scale self-supervised vision models for medical imaging via second-order graph matching. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 263 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 263 + ], + "type": "text", + "content": "[26] Wamidh K Mutlag, Shaker K Ali, Zahoor M Aydam, and Bahaa H Taher. Feature extraction methods: a review. In Journal of Physics: Conference Series, page 012028. IOP Publishing, 2020. 3, 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 265, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 265, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 265, + 294, + 319 + ], + "type": "text", + "content": "[27] A Tuan Nguyen, Thanh Nguyen-Tang, Ser-Nam Lim, and Philip HS Torr. Tipi: Test time adaptation with transformation invariance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24162-24171, 2023. 6, 8, 1, 3, 7, 9, 10" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 321, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 294, + 376 + ], + "type": "text", + "content": "[28] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In International conference on machine learning, pages 16888-16905. PMLR, 2022. 6, 1, 2, 3, 7, 8, 9, 10" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 377, + 294, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 294, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 294, + 432 + ], + "type": "text", + "content": "[29] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yaofo Chen, Peilin Zhao, and Mingkui Tan. Towards stable test-time adaptation in dynamic wild world. In International Conference on Learning Representations, 2023. 6, 8, 1, 3, 7, 9, 10" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 434, + 294, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 466 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 466 + ], + "type": "text", + "content": "[30] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 468, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 294, + 533 + ], + "type": "text", + "content": "[31] David Osowiechi, Mehrdad Noori, Gustavo Vargas Hakim, Moslem Yazdanpanah, Ali Bahri, Milad Cheraghalikhani, Sahar Dastani, Farzad Beizaee, Ismail Ayed, and Christian Desrosiers. Watt: Weight average test time adaptation of clip. Advances in neural information processing systems, 37: 48015-48044, 2024. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 535, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 294, + 589 + ], + "type": "text", + "content": "[32] Sung Kyu Park, Seunghan Yang, Jaegul Choo, and Sungrack Yun. Label shift adapter for test-time adaptation under covariate and label shifts. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 16375-16385, 2023. 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 591, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 591, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 591, + 294, + 656 + ], + "type": "text", + "content": "[33] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[34] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1406-1415, 2019. 8, 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "text", + "content": "[35] Sharan Ramjee and Aly El Gamal. Efficient wrapper feature selection using autoencoder and model based elimination. arXiv preprint arXiv:1905.11592, 2019. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "type": "text", + "content": "[36] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. Advances in neural information processing systems, 33:11539-11551, 2020. 3, 6, 1, 7, 8, 9, 10" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 163, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 195 + ], + "type": "text", + "content": "[37] Claude Elwood Shannon. A mathematical theory of communication. The Bell system technical journal, 27(3):379-423, 1948. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 197, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 262 + ], + "type": "text", + "content": "[38] David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering the game of go with deep neural networks and tree search. nature, 529(7587):484-489, 2016. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 263, + 553, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 263, + 553, + 318 + ], + "spans": [ + { + "bbox": [ + 316, + 263, + 553, + 318 + ], + "type": "text", + "content": "[39] Samarth Sinha, Peter Gehler, Francesco Locatello, and Bernt Schiele. Test: Test-time self-training under distribution shift. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2759-2769, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 319, + 553, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 365 + ], + "type": "text", + "content": "[40] Yongyi Su, Xun Xu, and Kui Jia. Towards real-world test-time adaptation: Tri-net self-training with balanced normalization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 15126-15135, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 365, + 553, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 419 + ], + "type": "text", + "content": "[41] Yongyi Su, Xun Xu, Tianrui Li, and Kui Jia. Revisiting realistic test-time training: Sequential inference and adaptation by anchored clustering regularized self-training. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 421, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 421, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 421, + 553, + 475 + ], + "type": "text", + "content": "[42] Yue Tan, Chen Chen, Weiming Zhuang, Xin Dong, Lingjuan Lyu, and Guodong Long. Is heterogeneity notorious? taming heterogeneity to handle test-time shift in federated learning. Advances in Neural Information Processing Systems, 36, 2024. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 476, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 476, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 476, + 553, + 521 + ], + "type": "text", + "content": "[43] Dipti Theng and Kishor K Bhoyar. Feature selection techniques for machine learning: a survey of more than two decades of research. Knowledge and Information Systems, 66(3):1575-1637, 2024. 3, 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 522, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 553, + 555 + ], + "type": "text", + "content": "[44] Robert Tibshirani. Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society Series B: Statistical Methodology, 58(1):267-288, 1996. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 556, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 556, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 556, + 553, + 589 + ], + "type": "text", + "content": "[45] Antonio Torralba and Alexei A Efros. Unbiased look at dataset bias. In CVPR 2011, pages 1521-1528. IEEE, 2011. 8, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 590, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 623 + ], + "type": "text", + "content": "[46] Laurens van der Maaten and Geoffrey E. Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9: 2579-2605, 2008. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 624, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 657 + ], + "type": "text", + "content": "[47] B Venkatesh and J Anuradha. A review of feature selection and its methods. Cybernetics and information technologies, 19(1):3-26, 2019. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[48] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5018-5027, 2017. 6, 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2129" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 597 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[49] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726, 2020. 6, 8, 1, 2, 3, 7, 9, 10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 138 + ], + "type": "text", + "content": "[50] Mei Wang and Weihong Deng. Deep visual domain adaptation: A survey. Neurocomputing, 312:135-153, 2018. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 194 + ], + "type": "text", + "content": "[51] Shuai Wang, Daoan Zhang, Zipei Yan, Jianguo Zhang, and Rui Li. Feature alignment and uniformity for test time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20050-20060, 2023. 6, 8, 1, 2, 3, 7, 9, 10" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 228 + ], + "type": "text", + "content": "[52] Svante Wold, Kim Esbensen, and Paul Geladi. Principal component analysis. Chemometrics and intelligent laboratory systems, 2(1-3):37-52, 1987. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 230, + 294, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 284 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 284 + ], + "type": "text", + "content": "[53] Yanan Wu, Zhixiang Chi, Yang Wang, Konstantinos N Plataniotis, and Songhe Feng. Test-time domain adaptation by learning domain-aware batch normalization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 15961-15969, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "type": "text", + "content": "[54] Kaichao You, Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Universal domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2720-2729, 2019. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 331, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 363 + ], + "type": "text", + "content": "[55] Linjing You, Jiabao Lu, and Xiayuan Huang. Test-time correlation alignment. In Forty-second International Conference on Machine Learning, 2025. 3, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 365, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 408 + ], + "type": "text", + "content": "[56] Longhui Yuan, Binhui Xie, and Shuang Li. Robust test-time adaptation in dynamic scenarios. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15922-15932, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 410, + 294, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 410, + 294, + 463 + ], + "spans": [ + { + "bbox": [ + 56, + 410, + 294, + 463 + ], + "type": "text", + "content": "[57] Yige Yuan, Bingbing Xu, Liang Hou, Fei Sun, Huawei Shen, and Xueqi Cheng. Tea: Test-time energy adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23901-23911, 2024. 6, 8, 1, 3, 7, 9, 10" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 465, + 294, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 465, + 294, + 508 + ], + "spans": [ + { + "bbox": [ + 56, + 465, + 294, + 508 + ], + "type": "text", + "content": "[58] Han Yue, Jundong Li, and Hongfu Liu. Second-order unsupervised feature selection via knowledge contrastive distillation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3, 4, 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 510, + 294, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 510, + 294, + 553 + ], + "spans": [ + { + "bbox": [ + 56, + 510, + 294, + 553 + ], + "type": "text", + "content": "[59] Longbin Zeng, Jiayi Han, Liang Du, and Weiyang Ding. Rethinking precision of pseudo label: Test-time adaptation via complementary learning. Pattern Recognition Letters, 177: 96-102, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 555, + 294, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 294, + 597 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 294, + 597 + ], + "type": "text", + "content": "[60] Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. Domain generalization: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4396-4415, 2022. 1" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2130" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/FROSS_ Faster-Than-Real-Time Online 3D Semantic Scene Graph Generation from RGB-D Images/f4b25ed8-2175-4e29-a597-e457f4795510_content_list.json b/2025/FROSS_ Faster-Than-Real-Time Online 3D Semantic Scene Graph Generation from RGB-D Images/f4b25ed8-2175-4e29-a597-e457f4795510_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e4f71b3aa15f5700f533fe4a38212692b1cd0cf1 --- /dev/null +++ b/2025/FROSS_ Faster-Than-Real-Time Online 3D Semantic Scene Graph Generation from RGB-D Images/f4b25ed8-2175-4e29-a597-e457f4795510_content_list.json @@ -0,0 +1,1499 @@ +[ + { + "type": "text", + "text": "FROSS: Faster-than-Real-Time Online 3D Semantic Scene Graph Generation from RGB-D Images", + "text_level": 1, + "bbox": [ + 197, + 130, + 799, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hao-Yu Hou1,3", + "bbox": [ + 133, + 202, + 256, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chun-Yi Lee2", + "bbox": [ + 299, + 203, + 410, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Motoharu Sonogashira3", + "bbox": [ + 450, + 203, + 642, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yasutomo Kawanishi3", + "bbox": [ + 684, + 203, + 859, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ National Tsing Hua University", + "bbox": [ + 173, + 238, + 424, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ National Taiwan University", + "bbox": [ + 473, + 238, + 699, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ RIKEN", + "bbox": [ + 746, + 238, + 818, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The ability to abstract complex 3D environments into simplified and structured representations is crucial across various domains. 3D semantic scene graphs (SSGs) achieve this by representing objects as nodes and their interrelationships as edges, facilitating high-level scene understanding. Existing methods for 3D SSG generation, however, face significant challenges, including high computational demands and non-incremental processing that hinder their suitability for real-time open-world applications. To address this issue, we propose FROSS (Faster-than-Real-Time Online 3D Semantic Scene Graph Generation), an innovative approach for online and faster-than-realtime 3D SSG generation that leverages the direct lifting of 2D scene graphs to 3D space and represents objects as 3D Gaussian distributions. This framework eliminates the dependency on precise and computationally-intensive point cloud processing. Furthermore, we extend the Replica dataset with inter-object relationship annotations, creating the ReplicaSSG dataset for comprehensive evaluation of FROSS. The experimental results from evaluations on ReplicaSSG and 3DSSG datasets show that FROSS can achieve superior performance while operating significantly faster than prior 3D SSG generation methods. Our implementation and dataset are publicly available at https://github.com/Howardkhh/FROSS.", + "bbox": [ + 89, + 324, + 483, + 703 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 722, + 220, + 738 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The abstraction of complex 3D environments into simplified representations in real time serves as a crucial capability across multiple application domains. The incremental generation of such representations in real time holds particular significance for fields ranging from robotics and computer vision to augmented reality (AR), where systems must perceive and analyze the environment instantaneously. 3D scene graphs (SGs) [1, 9] accomplish this abstraction by representing physical objects as nodes and their interrelationships as edges within a structured framework. 3D se", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3ff1488ecdb0ccc9ab8fe0361855f9b2bdd519b3a410a1a4ce91d248d387df87.jpg", + "image_caption": [ + "Figure 1. We introduce FROSS, an online real-time 3D semantic scene graph generation method that leverages and integrates 2D scene graphs. FROSS represents objects as 3D Gaussian distributions and operates without requiring 3D reconstruction." + ], + "image_footnote": [], + "bbox": [ + 516, + 304, + 906, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mantic scene graphs (SSGs) [31] extend this representation with an emphasis on semantic relationships, such as positional configurations between objects (e.g., a cup on a table) or functional interactions (e.g., a person seated in a chair), which enables high-level scene comprehension, reasoning, and interaction. Real-world applications, however, present open-world challenges where environments often exceed known spatial boundaries and contain previously unseen spaces [27]. For instance, a room may connect to multiple adjacent spaces or corridors, which necessitates continuous", + "bbox": [ + 511, + 750, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28818", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "adaptation and incremental scene understanding. This construction process must maintain high efficiency to generate structured environmental representations without creating performance bottlenecks. Meeting the processing demands requires the generation of 3D SSG to operate at faster-than-real-time speeds, with all computations completed between successive image frames from a high-speed sensor stream.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Executing all processes at faster-than-real-time rates confers multiple advantages: (1) prevention of bottlenecks in the 3D SSG generation framework, (2) facilitation of smooth operation on resource-constrained hardware such as edge computing devices, and (3) preservation of computational resources for additional system components. Contemporary 3D SSG generation approaches predominantly focus on offline construction, which requires access to the complete scene data, either instance-segmented point clouds [31, 41], a complete image sequence [9], or both [33]. This assumption of full scene availability conflicts with incremental environmental exploration. Although several studies [34, 35] have developed methods for incremental point cloud reconstruction and segmentation to achieve online and real-time 3D SSG generation, their aggregate system latency remains substantial, even with the utilization of multiple CPU cores. This constraint stems mainly from the substantial computational demands of environmental mapping and point cloud processing, restricting their practical application in real-time systems with limited computational power or other concurrent processes. Given these limitations, the aim of this study is to develop a method for faster-than-real-time online SSG generation.", + "bbox": [ + 91, + 200, + 482, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A fundamental insight for generating 3D SSGs emerges from the observation that precise object pose and shape information is not essential. Given that 3D SSGs provide a high-level semantic understanding of the environment, their functional implications remain valid with approximate object locations. For instance, in robotic applications, high-level planning can be successfully executed using relative spatial relationships [6, 10, 29]. Similarly, in 3D scene synthesis and room layout generation [3, 5, 7, 39, 40, 45], 3D SSGs serve as foundational structures from which more detailed content can be generated. These application scenarios eliminate the need for precise environmental mapping techniques such as Simultaneous Localization and Mapping (SLAM) [2] and additional point cloud processing, both of which contribute significant computational overhead.", + "bbox": [ + 89, + 550, + 482, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This observation motivates an alternative approach for online and real-time 3D SSG generation: inferring 2D SGs from images and subsequently lifting them into 3D space, thus circumventing the computational burden of environmental mapping and point cloud processing. To the best of our knowledge, the current literature has not adequately investigated this research direction except [16]. Such unexplored territory presents significant challenges, includ", + "bbox": [ + 89, + 779, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ing the handling of incomplete 2D object observations, the achievement of real-time SG generation, and the effective integration of multi-view information into complete 3D SSGs without object redundancy. These challenges, therefore, provide promising avenues for further innovative research contributions.", + "bbox": [ + 511, + 90, + 903, + 180 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In light of these considerations, we introduce FROSS (Faster-than-Real-time Online 3D Semantic Scene Graph Generation), a methodology that constructs 3D SSGs by lifting 2D SGs to 3D space and integrating them into a comprehensive representation. To achieve real-time processing without requiring precise object localization, FROSS initiates the process by constructing 2D SGs from detected objects in images. A key innovation of FROSS lies in its utilization of Gaussian distributions to model objects in 3D space, which facilitates the integration of objects observed across multiple images into a coherent 3D SSG, as illustrated in Figure 1. Specifically, FROSS represents detected objects through Gaussian distributions, lifts these distributions from 2D SGs into 3D space, and merges them according to their predicted classes and statistical distances. We validate FROSS on the 3DSSG dataset [31] and conduct additional evaluations on ReplicaSSG, an extended version of the Replica dataset [28] that we have enhanced with object relationship annotations. ReplicaSSG incorporates object classes defined in Visual Genome [18], a widely-adopted 2D SG dataset. The main contributions of the paper can be summarized as follows:", + "bbox": [ + 511, + 181, + 906, + 512 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce FROSS, an innovative methodology for online real-time generation of 3D SSGs. FROSS demonstrates superior performance and significantly faster processing speeds compared to existing baseline methods.", + "- We propose a new merging algorithm based on Gaussian distributions for 3D object integration. This algorithm effectively prevents the proliferation of duplicate or incorrectly detected objects as the input image sequence expands, thereby maintaining computational efficiency.", + "- We develop ReplicaSSG, an extension of the Replica dataset with 3D SSG annotations. Comprehensive evaluations of FROSS on both the 3DSSG and ReplicaSSG datasets demonstrate FROSS's effectiveness and computational efficiency over diverse environmental scenarios." + ], + "bbox": [ + 513, + 513, + 903, + 724 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 739, + 653, + 753 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SGs [12, 13, 36, 38] constitute a fundamental data structure for object and relationship representation within images following their introduction by [13]. This representation has demonstrated substantial utility across various scene understanding tasks [8, 14, 15, 19, 25, 37]. The concept later evolved to encompass 3D spaces, where objects may incorporate attributes such as class labels, visual features, shape characteristics, and affordances. Research in 3D SGs developed into two branches of representation: SSGs intro", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28819", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "duced in [9, 16, 31] and hierarchical scene graphs (HSGs) proposed in [1, 11, 26]. Both frameworks represent 3D objects as nodes with pairwise relationships defined through edges. However, these approaches differ fundamentally in their relationship definitions. SSGs concentrate on semantic relationships between objects, while HSGs organize objects into hierarchical layers that include buildings, rooms, and constituent objects, where edges connect only nodes in identical or adjacent layers. Compared to HSGs, SSGs enable higher-level reasoning capabilities through rich semantic relationship representations. The extension beyond spatial relationships allows SSGs to be applied to broader domains such as robotics, 3D computer vision, and AR.", + "bbox": [ + 89, + 90, + 480, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Offline 3D Semantic Scene Graph Generation", + "text_level": 1, + "bbox": [ + 89, + 299, + 478, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The initial development of 3D SSG generation focused primarily on offline settings that required either complete instance-segmented point clouds [20, 31-33, 41, 43] or image sequences [9, 33] as inputs. Such requirements limited their applicability to scenarios with incremental data collection, as they necessitated complete data reprocessing. Despite these limitations, these studies established foundations for 3D SSG generation. The research in [9] pioneered 3D SSG generation by predicting 3D object ellipsoids from image collections and utilizing their geometry and appearance features through a recurrent neural network to extract inter-object relationships. The study in [31] advanced this field through the introduction of the 3DSSG dataset, which enhanced the 3RScan dataset [30] with SSG annotations. This dataset represents the first large-scale collection with dense annotations and has become the primary benchmark for evaluating 3D SSG generation methods. The authors of [31] further contributed a generation methodology by utilizing instance-segmented point clouds as input and leveraged PointNet [24] for feature extraction as well as a graph convolutional network [17] for label prediction.", + "bbox": [ + 91, + 321, + 483, + 640 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Online 3D Semantic Scene Graph Generation", + "text_level": 1, + "bbox": [ + 89, + 651, + 475, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Several previous studies have explored 3D SG generation in online settings, with some achieving real-time performance. The research in [16] introduced a framework that predicts local 3D SSGs from RGB-D images and integrates them incrementally into a global representation. However, their computational requirements render real-time execution infeasible. SceneGraphFusion [34] and the study in [35] developed methods for online point cloud reconstruction and segmentation, followed by feature extraction for local SSG prediction and global fusion. These frameworks can achieve real-time performance through multi-threaded execution of parallel components. In contrast to these existing methods, our proposed FROSS framework avoids precise object localization by approximating objects with Gaussian distributions. This approximation approach enables faster-than", + "bbox": [ + 89, + 674, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "real-time online 3D SSG generation, allowing the end-to-end processing latency to be below the sensor data acquisition period while maintaining single-threaded execution.", + "bbox": [ + 511, + 90, + 903, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 511, + 151, + 648, + 169 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Definition", + "text_level": 1, + "bbox": [ + 511, + 176, + 696, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The primary problem concerned in this study is the online generation of 3D SSGs for environments where the complete scene structure remains unknown a priori. Given a sequence of input images, the primary objective is to construct a 3D SSG $\\mathcal{G}$ of the target environment. The graph $\\mathcal{G}$ consists of a set of nodes $\\nu$ and their corresponding directed edges $\\mathcal{E}$ . Each node $v_{i} = (c_{i},\\mu_{i}^{3\\mathrm{D}},\\Sigma_{i}^{3\\mathrm{D}})\\in \\mathcal{V}$ is indexed by $i\\in \\{1,\\ldots ,N\\}$ , representing a unique object in the scene. Every object has a class label $c_{i}\\in \\mathcal{C}$ , where $\\mathcal{C}$ is the set of object categories. While different approaches may incorporate various node attributes, our formulation represents each object as a 3D Gaussian distribution parameterized by mean vector $\\mu_i^{3\\mathrm{D}}\\in \\mathbb{R}^3$ and covariance matrix $\\Sigma_i^{3\\mathrm{D}}\\in \\mathbb{R}^{3\\times 3}$ . The edges $e_{i\\rightarrow j}\\in \\mathcal{E}$ connect node $v_{i}$ to node $v_{j}$ with relationship $r_{i\\rightarrow j}\\in \\mathcal{R}$ , where $\\mathcal{R}$ represents the set of predefined relationship types. Furthermore, the algorithm must be able to maintain online processing capabilities to update $\\mathcal{G}$ continuously whenever new image data becomes available.", + "bbox": [ + 511, + 199, + 906, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Overview of Framework", + "text_level": 1, + "bbox": [ + 511, + 481, + 738, + 496 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 2 presents an overview of the proposed FROSS framework, which processes image sequences with corresponding depth information and camera transformations as input. The framework begins with the EGTR [12] 2D SG generation model, with the object detection backbone model replaced with RT-DETR [21, 44]. This RT-DETR object detector is a state-of-the-art real-time detection model, which preserves intermediate self-attention features for subsequent relationship extraction. FROSS then converts object bounding boxes to 2D Gaussian distributions and projects them into 3D space to generate local SSGs, as elaborated in Section 3.3. These Gaussian distributions serve as efficient approximations of both object positions and their spatial distributions for rapid processing. These local graphs are then integrated into a global SSG through the proposed merging algorithm that is described in detail in Section 3.4.", + "bbox": [ + 511, + 503, + 906, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Lifting 2D SG to 3D", + "text_level": 1, + "bbox": [ + 511, + 772, + 705, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The 2D SGs generated from RT-DETR [21, 44] and EGTR [12] enable FROSS to construct local 3D SSGs $\\mathcal{G}_{\\mathrm{local}} = (\\mathcal{V}_{\\mathrm{local}},\\mathcal{E}_{\\mathrm{local}})$ through the integration of depth information and camera transformation. This process involves two key steps: (1) the conversion of object bounding boxes to 2D Gaussians, and (2) their subsequent back-projection into 3D space while preserving inter-object relationships.", + "bbox": [ + 511, + 794, + 905, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28820", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c9c0137962ed863d7f4ef0b316022ea0a35d687b329881ea4b6ee1d68fbaf661.jpg", + "image_caption": [ + "Figure 2. An overview of the FROSS framework: (a) The process initiates with object detection via RT-DETR [44] from an RGB-D image and its associated camera pose. (b) The EGTR model [12] extracts inter-object relationships through utilization of preserved self-attention features from RT-DETR. (c) 2D Gaussian distributions derived from detected bounding boxes undergo spatial transformation into 3D space for local 3D SSG construction (Section 3.3). (d) The resultant local 3D SSG is subsequently incorporated into the global 3D SSG through our proposed merging algorithm (Section 3.4)." + ], + "image_footnote": [], + "bbox": [ + 125, + 89, + 851, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2D Gaussian Representation of Bounding Boxes. Although several methods exist for lifting 2D bounding boxes to 3D space, including point cloud back-projection and joint optimization of 3D ellipsoids [9], the Gaussian distribution approach emerges as the most computationally efficient solution among these alternatives. The representation of objects via Gaussian distributions requires the specification of mean $\\mu_i^{2\\mathrm{D}}$ and covariance matrix $\\Sigma_i^{2\\mathrm{D}}$ of the 2D object node $v_i^{2\\mathrm{D}}$ , which would correspond to a 3D node $v_i \\in \\mathcal{V}_{\\mathrm{local}}$ . Based on the formulation in [22], FROSS models bounding boxes as 2D uniform distributions. The mean values correspond to the bounding box centers, while the covariance matrices maintain consistency with the uniform distributions. The covariance matrix of $v_i^{2\\mathrm{D}}$ is formulated as:", + "bbox": [ + 89, + 364, + 483, + 578 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Sigma_ {i} ^ {\\mathrm {2 D}} = \\frac {1}{1 2} \\left[ \\begin{array}{c c} W _ {i} ^ {2} & 0 \\\\ 0 & H _ {i} ^ {2} \\end{array} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 583, + 482, + 617 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $W_{i}$ and $H_{i}$ are 2D bounding box's width and height. 3D Back-Projection of Gaussian Distributions. Given the 2D Gaussians, their corresponding depths at the Gaussian centroids, and camera transformation matrices, the mean positions $\\mu_{i}^{3\\mathrm{D}}$ and covariances $\\Sigma_{i}^{3\\mathrm{D}}$ of the Gaussians in the 3D space can be approximated. The 3D mean $\\mu_{i}^{3\\mathrm{D}}$ can be derived from the 2D mean $\\mu_{i}^{2\\mathrm{D}} = (p_{x}, p_{y})^{\\top}$ and its corresponding depth value $z$ through the following equation:", + "bbox": [ + 89, + 621, + 483, + 743 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {i} ^ {\\mathrm {3 D}} = R \\cdot K ^ {- 1} \\cdot \\left(p _ {x}, p _ {y}, 1\\right) ^ {\\top} \\cdot z + \\mathbf {t}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 160, + 747, + 482, + 766 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $K \\in \\mathbb{R}^{3 \\times 3}$ represents the camera intrinsic matrix, and $R \\in \\mathbb{R}^{3 \\times 3}$ and $\\mathbf{t} \\in \\mathbb{R}^3$ denote the rotation matrix and translation vector of the current camera, respectively. The approximation of the 3D covariance matrix utilizes the Jacobian $J$ of the local affine approximation of the projective transformation [46], which is defined as the following:", + "bbox": [ + 89, + 771, + 483, + 863 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nJ = \\left[ \\begin{array}{c c c} \\frac {f _ {x}}{z} & 0 & - \\frac {f _ {x} p _ {x}}{z ^ {2}} \\\\ 0 & \\frac {f _ {y}}{z} & - \\frac {f _ {y} p _ {y}}{z ^ {2}} \\end{array} \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 867, + 482, + 905 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{x}$ and $f_{y}$ represent the focal lengths in $x$ and $y$ dimensions. In the original literature of [46], the Jacobian $J$ is utilized to project 3D covariance matrices onto the 2D image plane. However, our approach inverts this process by back-projecting the 2D covariance matrix $\\Sigma_{i}^{2\\mathrm{D}}$ into 3D space as $\\bar{\\Sigma}_{i}^{3\\mathrm{D}}$ via the pseudo-inverse $J^{+}$ of the Jacobian:", + "bbox": [ + 511, + 364, + 906, + 457 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\Sigma} _ {i} ^ {\\mathrm {3 D}} = R J ^ {+} \\Sigma_ {i} ^ {\\mathrm {2 D}} J ^ {+ \\top} R ^ {\\top}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 464, + 906, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The naive definition alone is insufficient since the original equation is inherently designed for projection from 3D to 2D space. Simply inverting the equation results in a 3D Gaussian that lacks variance information along the depth axis, as the 2D covariance matrix does not capture depth-related variance. This limitation necessitates the specification of variance in the depth dimension. We hypothesize that this variance approximates the average variance of the other dimensions. The specification has to be applied prior to the rotation matrix $R$ , leading to the following equations:", + "bbox": [ + 511, + 492, + 906, + 645 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Sigma_ {i} ^ {\\mathrm {3 D / /}} = J ^ {+} \\Sigma_ {i} ^ {\\mathrm {2 D}} J ^ {+ \\top}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 652, + 906, + 671 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Sigma_ {i} ^ {\\mathrm {3 D} \\prime} = \\Sigma_ {i} ^ {\\mathrm {3 D} \\prime \\prime} + \\left[ \\begin{array}{l l l} 0 & 0 & 0 \\\\ 0 & 0 & 0 \\\\ 0 & 0 & \\frac {(\\Sigma_ {i} ^ {\\mathrm {3 D} \\prime \\prime}) _ {1 , 1} + (\\Sigma_ {i} ^ {\\mathrm {3 D} \\prime \\prime}) _ {2 , 2}}{2} \\end{array} \\right]. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 674, + 905, + 728 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this formulation, $(\\Sigma_{i}^{3\\mathrm{D} / \\prime})_{m,n}\\in \\mathbb{R}$ represents the element at position $(m,n)$ of matrix $\\Sigma_i^{3\\mathrm{D} / \\prime}$ . Finally, the covariance matrix $\\Sigma_{i}^{3\\mathrm{D}^{\\prime}}$ is transformed to world coordinates using the rotation matrix $R$ , resulting in the following expression:", + "bbox": [ + 511, + 736, + 906, + 797 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Sigma_ {i} ^ {\\mathrm {3 D}} = R \\Sigma_ {i} ^ {\\mathrm {3 D} \\prime} R ^ {\\top}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 645, + 805, + 906, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Merging 3D SSGs", + "text_level": 1, + "bbox": [ + 511, + 833, + 689, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With a local 3D SSG $\\mathcal{G}_{\\mathrm{local}} = (\\mathcal{V}_{\\mathrm{local}},\\mathcal{E}_{\\mathrm{local}})$ constructed, it is then integrated with the global 3D SSG $\\mathcal{G}_{\\mathrm{global}}$ through Gaussian merging and relationship accumulation. Figure 3", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28821", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "illustrates the proposed merging process. Specifically, it initializes a queue with nodes from $\\nu_{\\mathrm{local}}$ and evaluates potential merging opportunities through distance calculations. For each node $v_{i}$ in this queue, FROSS calculates the Hellinger distances [22] between the Gaussian distributions of $v_{i}$ and all other nodes $v_{j}$ of the same semantic class, in both the local node queue and $\\nu_{\\mathrm{global}}$ , to identify object correspondence. The algorithm merges nodes with Hellinger distances below the threshold $\\delta_{\\mathrm{d}}$ and adds unmerged nodes to $\\nu_{\\mathrm{global}}$ as new objects. The Hellinger distance $H_{D}(i,j)$ between Gaussian distributions $\\mathcal{N}(\\mu_i,\\Sigma_i)$ and $\\mathcal{N}(\\mu_j,\\Sigma_j)$ is quantified through the following equation, wherein $B_{D}(i,j)$ represents the Bhattacharyya distance:", + "bbox": [ + 89, + 90, + 483, + 287 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nH _ {D} (i, j) = \\sqrt {1 - \\exp (- B _ {D} (i , j))}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 295, + 482, + 315 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nB _ {D} (i, j) = \\frac {1}{8} \\Delta \\mu_ {i j} ^ {T} \\Sigma^ {- 1} \\Delta \\mu_ {i j} + \\frac {1}{2} \\ln \\left(\\frac {\\operatorname* {d e t} \\Sigma}{\\sqrt {\\operatorname* {d e t} \\Sigma_ {i} \\operatorname* {d e t} \\Sigma_ {j}}}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 316, + 480, + 366 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Delta \\mu_{ij} = \\mu_i - \\mu_j$ and $\\Sigma = \\frac{1}{2} (\\Sigma_i + \\Sigma_j)$ . The fusion of these nodes occurs through a weighted integration of their respective Gaussian distributions [4]. The weighting coefficients for this integration process reflect both spatial coverage and detection confidence across multiple viewpoints. Objects detected from diverse angular perspectives and spatial positions receive proportionally higher weights, thus mitigating viewpoint bias and ensuring robust three-dimensional representation. Given that weights $w_i$ and $w_j$ represent the detection frequencies of their respective nodes, the mean vector $\\mu_k$ and covariance matrix $\\Sigma_k$ for the resultant merged node are determined via the formulations:", + "bbox": [ + 89, + 377, + 483, + 559 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {k} = \\frac {w _ {i} \\mu_ {i} + w _ {j} \\mu_ {j}}{w _ {i} + w _ {j}}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 566, + 482, + 597 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Sigma_ {k} = \\frac {w _ {i} \\Sigma_ {i} + w _ {j} \\Sigma_ {j}}{w _ {i} + w _ {j}} + \\frac {w _ {i} w _ {j} \\left(\\mu_ {i} - \\mu_ {j}\\right) \\left(\\mu_ {i} - \\mu_ {j}\\right) ^ {\\top}}{\\left(w _ {i} + w _ {j}\\right) ^ {2}}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 599, + 482, + 635 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mu_{i},\\mu_{j}$ are the mean vectors, and $\\Sigma_{i},\\Sigma_{j}$ are the covariance matrices of the respective Gaussian distributions to be merged. All relationship edges originally connected to $v_{i}$ and $v_{j}$ are then redirected to the newly merged node $v_{k}$ . The final relationship prediction between two global object nodes is selected through majority voting, choosing the most frequent prediction across viewpoints.", + "bbox": [ + 511, + 90, + 905, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental Results", + "text_level": 1, + "bbox": [ + 511, + 210, + 718, + 228 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The evaluation of FROSS encompasses two 3D SSG datasets: 3DSSG [31] and our developed ReplicaSSG datasets. Section 4.1 introduces the datasets, baseline SSG generation methods, and evaluation metrics. Section 4.2 presents the implementation details. Section 4.3 provides comprehensive comparisons between FROSS and existing SSG generation methods on the 3DSSG dataset. Moreover, the effectiveness of our approach is further demonstrated through qualitative results in Section 4.4. We further provide runtime analyses on the ReplicaSSG dataset in Section 4.5, along with additional ablation studies in Section 4.6. More quantitative and qualitative analyses are provided in the supplementary material.", + "bbox": [ + 511, + 237, + 906, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Evaluation Setup", + "text_level": 1, + "bbox": [ + 511, + 444, + 681, + 459 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.1. Datasets", + "text_level": 1, + "bbox": [ + 511, + 465, + 617, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3DSSG. The 3DSSG dataset [31] extends 3RScan [30], which encompasses 1,482 scans of indoor environments with their corresponding RGB-D image sequences, 3D meshes, and dense instance segmentation. The dataset contains 363,555 images, with an average of 245 images per scan. 3DSSG augments the base dataset with object attributes, hierarchical category labels, and directed edges that describe inter-object semantic relationships such as 'standing on,' 'attached to,' and 'same color.' This dataset has established a primary benchmark for numerous", + "bbox": [ + 511, + 484, + 908, + 636 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/32fdd0c43df1d873c6024bd66ff45d0cd00e1cad2bbe51d739b17ac8c4e08fbb.jpg", + "image_caption": [ + "Figure 3. Illustration of the object merging process using the proposed algorithm described in Section 3.4. (a) Starting with an input image, a set of objects with predicted categories and bounding boxes is generated. (b) These objects are then lifted into 3D spaces to form the object nodes in $\\mathcal{V}_{\\mathrm{local}}$ with colors matching their bounding boxes (yellow, cyan, and blue), while pre-existing objects in $\\mathcal{V}_{\\mathrm{global}}$ are shown in green. (c) The merging algorithm combines the objects linked by the arrows, resulting in the updated global objects, represented in red." + ], + "image_footnote": [], + "bbox": [ + 173, + 660, + 823, + 829 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28822", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "studies [20, 31-35, 41-43]. Our experiments utilize the publicly available training and testing splits. To address the class imbalance in both object and relationship labels, this study adopts the commonly used category definitions in [31, 34, 35]. Specifically, the original 160 object categories map to 20 categories from NYUv2 [23], while seven of the 26 predicate categories remain. Following this categorical mapping, 3DSSG contains a total of 21,974 distinct objects and 16,324 inter-object relationships. This extensive collection of relationships offers a substantial foundation for spatial and semantic analyses. Nevertheless, the low image quality and low frame rate of 3DSSG hinders the comprehensive evaluation of FROSS at its full operational capacity. This necessitates our development of ReplicaSSG, which is an enhanced adaptation of the Replica dataset [28].", + "bbox": [ + 89, + 90, + 480, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ReplicaSSG. ReplicaSSG extends the Replica dataset [28] with newly annotated object relationships. The original Replica provides high-quality reconstructions of indoor environments with instance-segmented meshes and photorealistic renderings, and is particularly suitable for SSG evaluation. Although Replica encompasses only 18 scenes, which precludes its use for training purposes, it serves as an effective evaluation benchmark. In the evaluation of FROSS, seven scenes serve as validation environments for hyperparameter optimization, while the remaining 11 scenes function as performance assessment platforms. Rather than preserving Replica's original object category definitions, ReplicaSSG adopts the classification system from Visual Genome [18], a 2D SG dataset. This label mapping from Replica to Visual Genome facilitates zero-shot transfer from 2D SG models, as this approach eliminates the requirement for additional training on the Replica dataset. The dataset splits and label mapping are available in the released code. In addition, detailed statistics for the dataset are provided in the supplementary materials (Section 8). As a result, ReplicaSSG offers high-quality instance-segmented meshes with comprehensive relationship annotations.", + "bbox": [ + 91, + 321, + 483, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.2. Baseline Methods", + "text_level": 1, + "bbox": [ + 89, + 672, + 261, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluated FROSS on the 3DSSG dataset against several baseline methods for SSG generation. The baselines include three image-based methods, IMP [36], VGfM [9] and Kim's construction framework [16], alongside three point cloud-based methods: 3DSSG [31], SGFN [34] and the approach presented in [35]. Kim's 3D object representation and merging mechanism are integrated into FROSS for its baseline implementation, while other baseline implementations follow $[35]^1$ . In the ablation studies, we investigate the impact of using ground truth 2D SGs and camera trajectories on the ReplicaSSG dataset.", + "bbox": [ + 89, + 695, + 482, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.3. Matching Object Predictions to Ground Truth", + "text_level": 1, + "bbox": [ + 511, + 90, + 880, + 106 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The evaluation methodology follows [35] for matching predicted object instances with ground truth counterparts. As FROSS generates predictions without explicit point cloud output, we establish evaluation metrics using back-projected 3D points. These points derive from the central region $(H / 2\\times W / 2$ pixels) of each object's 2D bounding box. The matching process associates each predicted point with its nearest ground truth point within a 0.1-meter radius. A predicted object establishes correspondence with a ground truth object under two criteria, described as follows:", + "bbox": [ + 511, + 109, + 903, + 261 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. The majority (more than $50\\%$ ) of its back-projected 3D points must each have their nearest ground truth point neighbor within the corresponding ground truth object.", + "2. The ratio between the second-largest and largest ground truth object overlap counts must remain below $75\\%$ ." + ], + "bbox": [ + 511, + 262, + 903, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In contrast to the other baseline methods, which may permit multiple predictions to be matched with a single ground truth object due to possible over-segmentation in input instance-segmented point clouds, our proposed FROSS framework enforces a stricter one-to-one correspondence between predicted and ground truth objects. Since FROSS merges identical objects into a single prediction, matching multiple predictions to the same ground truth object is not required. The one-to-one correspondence enforces the integrity of the proposed merging algorithm, ensuring accurate association between predicted and ground truth objects.", + "bbox": [ + 511, + 339, + 903, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.4. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 511, + 513, + 692, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The evaluation of SSG generation employs recall metrics for relationships, objects, and predicate, a standard approach in both 2D and 3D SG evaluation [13, 31, 34, 35, 38]. Object recall quantifies the proportion of ground truth instances matched to predictions with correct category labels. Predicate recall computes the proportion of correctly classified predicates between detected objects, regardless of the object classes. Relationship recall measures the proportion of correctly identified triplets (subject, object, predicate) that meet two criteria: valid detection of both subject and object, and correct classification of their connecting predicate. To mitigate the impact of class imbalance, we report mean recall (mRecall) to provide a more equitable evaluation across all classes. Note that a more detailed explanation of the utilized recall metric is offered in the supplementary materials (Section 6).", + "bbox": [ + 511, + 534, + 906, + 776 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 787, + 730, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For our 2D SG generation model, we employ EGTR [12] with RT-DETRv2-M [21, 44] as the object detection backbone. The 2D SGs from the 3DSSG dataset [31] serve as training data specifically for 3DSSG experiments, while the Visual Genome dataset [18] functions as the training foundation for ReplicaSSG experiments. Our methodology ad", + "bbox": [ + 511, + 809, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1https://github.com/ShunChengWu/3DSSG/tree/4b783ec", + "bbox": [ + 107, + 886, + 454, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28823", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/daaa41bb09e5a7de51469946a89cfcf48375c08bc2ad387f58dcef5aa3e70345.jpg", + "table_caption": [ + "Table 1. Performance comparison of 3D SSG generation methods on the 3DSSG dataset and the end-to-end latency without environmental mapping reported in their original literature, along with their respective input modalities. The best and second-best results are highlighted in red, and blue, respectively. Latencies are reported in milliseconds." + ], + "table_footnote": [], + "table_body": "
MethodRecall (%)mRecall (%)LatencyInput Modality
Rel.Obj.Pred.Obj.Pred.
IMP [36]19.749.520.934.713.8-RGB-D
VGfM [9]19.650.020.434.811.0250RGB
3DSSG [31]12.937.422.026.214.4-Point Cloud
SGFN [34]22.051.627.537.724.0161RGB-D
Wu [35]23.353.828.443.826.6191RGB-D2
Kim [16]9.159.07.151.08.0310RGB-D
FROSS (Ours)27.962.433.063.818.07RGB-D
", + "bbox": [ + 89, + 179, + 480, + 301 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "heres to the training protocol established in EGTR, which initiates with pre-training the object detector for standard object detection tasks. All hyperparameters were determined through grid search evaluation on the validation split, with particular emphasis on relationship recall optimization. Our implementation applies a confidence threshold of 0.7 for object filtering and retains only the top ten relationships per 2D SG. For Gaussian merging operations, the Hellinger distance threshold $\\delta_{\\mathrm{d}}$ is established at 0.85. The main experimental evaluations utilize ground truth trajectories to guide the SSG generation process. The impact of estimated trajectories are further analyzed via ablation studies in Section 4.6.2.", + "bbox": [ + 88, + 328, + 482, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.Quantitative Results", + "text_level": 1, + "bbox": [ + 89, + 536, + 284, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1 presents a performance and latency comparison between FROSS and various representative baseline methodologies. The results reveal that FROSS achieves the highest performance among all baseline methods with much lower processing latency. This substantiates the advantages of lifting scene graphs from 2D images over direct point cloud reasoning [31, 34, 35], as point clouds can sometimes present challenges with occlusion, sparsity, and reconstruction quality. The superior performance compared to the other image-based methods [9, 16, 36] can be attributed to our Gaussian-based merging strategy. In contrast, IMP [36] employs a voting mechanism for object classification, while VGfM [9] uses a basic recurrent neural network, which may restrict their object and relationship classification accuracy. Kim's baseline [16] achieves similar object recall to FROSS due to the same 2D SG generation pipeline. However, its merging mechanism fails to suppress duplicate detections, which hinders relationship aggregation and leads to significantly lower relationship and predicate recall. These results collectively substantiate the effectiveness of FROSS", + "bbox": [ + 91, + 559, + 482, + 863 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2ecb1e704d98206bcb0bd47ca757d87c49a85d1e1357c7e6e6c0915b5262fb19.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison between FROSS and Wu [35] on the 3DSSG dataset. Only representative objects are visualized. Errors are marked in red, with ground truth label shown in parentheses. Node colors correspond to the respective scene graphs." + ], + "image_footnote": [], + "bbox": [ + 517, + 89, + 903, + 247 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in combining 2D information with Gaussian merging for 3D SSG generation.", + "bbox": [ + 511, + 339, + 906, + 371 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Qualitative Results", + "text_level": 1, + "bbox": [ + 511, + 378, + 696, + 395 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 4 presents the 3D SSG generation results from FROSS, demonstrating its ability to approximate object positions and shapes within a scene through Gaussian distributions, while simultaneously extracting their relationships. The unified objects highlight the effectiveness of the proposed merging algorithm, which consolidates predictions from multiple views, reducing redundant objects and processing time. This representation method offers a high-level scene understanding for downstream applications. In comparison, Wu's method [35] frequently misclassifies objects with similar geometry (e.g., \"curtain\" vs. \"door\", \"table\" vs. \"sofa\") and confuses semantically close relationships (e.g., \"standing on\" vs. \"supported by\"). This performance gap shows FROSS's advantage in utilizing richer visual information, in contrast to Wu [35], which relies solely on cropped object image for visual feature extraction. Additional qualitative results from FROSS are provided in the supplementary materials (Section 7.2).", + "bbox": [ + 511, + 401, + 908, + 674 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Runtime Analysis", + "text_level": 1, + "bbox": [ + 511, + 681, + 687, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2 examines the latency and frames-per-second (FPS) performance of FROSS, evaluated on an AMD Ryzen 9 7950X processor with an NVIDIA GeForce RTX 4090 GPU. This analysis serves to validate the computational efficiency of FROSS in accordance with its design objective of achieving faster-than-real-time processing capabilities. The evaluation restricts processing to a single CPU core and one GPU to assess system efficiency under limited computational resources. The metrics represent averages across 14,400 frames from four ReplicaSSG test scenes. The RT-DETR object detector and the EGTR relationship extraction module exhibit the highest individual latency. On the other hand, the 3D SSG merging algorithm contributes", + "bbox": [ + 511, + 704, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "2 Although Wu's method [35] supports RGB inputs, the reported results are based on predictions using dense reconstruction from RGB-D images.", + "bbox": [ + 89, + 875, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28824", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/99608f2383151275bcdeece81691b9aa90a0682dedf8c1f2f7d05d13036a0658.jpg", + "table_caption": [ + "Table 2. Runtime analysis of the key components of FROSS. Obj. Det. refers to the object detection model RT-DETR [21, 44], while Rel. Ext. corresponds to the EGTR relationship extractor [12]. Merging 3D SSGs represents the process described in Section 3.4. Only components with significant latency impact are included." + ], + "table_footnote": [], + "table_body": "
Mean Latency (in milliseconds)
Obj. Det.Rel. Ext.Merging 3D SSGsFPS
2.314.510.12144.09
", + "bbox": [ + 125, + 170, + 444, + 217 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "minimal computational overhead. Compared to the previous methods for online real-time 3D SSG generation methods [16, 34, 35], FROSS demonstrates significantly reduced end-to-end latency and increased FPS, which establishes a substantial improvement in terms of speed and efficiency.", + "bbox": [ + 89, + 236, + 483, + 313 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 321, + 243, + 338 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6.1. Using Ground Truth 2D Scene Graphs", + "text_level": 1, + "bbox": [ + 89, + 343, + 405, + 358 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the ablation analysis presented in Table 3, we investigate the influence of 2D SG quality on FROSS performance, given that FROSS relies heavily on accurate 2D SG predictions to generate 3D SSGs. The results indicate that the accuracy of the 2D SG has a significant impact on the final 3D SSG quality. In both the 3DSSG and ReplicaSSG datasets, FROSS achieves substantial gains in recall when ground truth 2D SG is used instead of predicted 2D SG. This performance difference is particularly pronounced in the ReplicaSSG dataset, where the baseline FROSS method achieves a relatively low recall score compared to the ground truth-assisted version. The discrepancy stems from differences between ReplicaSSG and Visual Genome [18]. Since the 2D SG model was originally trained on Visual Genome, the model may not generalize effectively to datasets with different characteristics. This leads to lower recall in both object and relationship detection within the ReplicaSSG dataset. This finding suggests the untapped potential for FROSS via domain-adaptive 2D SG models, with current results merely representing a lower bound on the framework's capabilities.", + "bbox": [ + 89, + 361, + 483, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6.2. Using Estimated Camera Trajectory", + "text_level": 1, + "bbox": [ + 89, + 671, + 390, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further validate FROSS's effectiveness using estimated camera trajectories instead of ground truth poses to assess real-world applicability. This evaluation is crucial as", + "bbox": [ + 89, + 689, + 482, + 736 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8eee3ccb36d9006763b37a6372c60d3c06988e7da0308080492a5bffe6b03dad.jpg", + "table_caption": [ + "Table 3. Analysis of FROSS with predicted and ground truth 2D SGs as input on both the 3DSSG and ReplicaSSG datasets. 'w/ GT' denotes the version of FROSS utilizing ground truth 2D SGs." + ], + "table_footnote": [], + "table_body": "
DatasetMethodRecall (%)mRecall (%)
Rel.Obj.Pred.Obj.Pred.
3DSSG [31]FROSS27.962.433.063.818.0
FROSS w/ GT55.888.656.093.356.8
ReplicaSSGFROSS22.326.127.828.820.4
FROSS w/ GT67.689.167.692.682.6
", + "bbox": [ + 94, + 811, + 478, + 897 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b98c3e3c8b392e5e1506893bc185755645f599e1a2a1acedc9af63b541c33174.jpg", + "table_caption": [ + "Table 4. Performance comparison of FROSS with the predicted and the ground truth camera trajectories as input on the ReplicaSSG dataset. Please note that 'w/ SLAM' denotes the version of FROSS utilizing trajectories estimated by ORB-SLAM3 [2]." + ], + "table_footnote": [], + "table_body": "
MethodRecall (%)mRecall (%)
Rel.Obj.Pred.Obj.Pred.
FROSS22.326.127.828.820.4
FROSS w/ SLAM22.725.827.227.720.1
", + "bbox": [ + 519, + 156, + 900, + 220 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/bc6c32be9d66a70866285e3bb04d5d137ead62eabccf4d873f7b68af6010e7aa.jpg", + "table_caption": [ + "Table 5. Recall on the validation split of the 3DSSG dataset with different Hellinger distance thresholds." + ], + "table_footnote": [], + "table_body": "
Threshold δd0.60.650.70.750.80.850.9
Recall (%)Rel.15.116.520.022.725.930.029.3
Obj.69.569.869.168.965.761.955.4
Pred.16.217.921.824.928.833.634.3
", + "bbox": [ + 519, + 273, + 900, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "most practical deployments must rely on estimated camera poses rather than perfect ground truth information. Table 4 presents the impact of using estimated trajectories generated by ORB-SLAM3 [2] on the ReplicaSSG dataset. The evaluation results demonstrate a slight improvement in relationship recall, while object and predicate recall exhibit minor decreases. This indicates that FROSS remains robust to inaccuracies in camera trajectory estimation. Note that a more detailed analysis of ORB-SLAM3's trajectory estimation errors on the ReplicaSSG dataset is provided for reference in the supplementary materials (Section 7.4).", + "bbox": [ + 511, + 351, + 906, + 517 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6.3. Effect of Hellinger Distance Threshold", + "text_level": 1, + "bbox": [ + 511, + 523, + 826, + 539 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5 presents the effect of varying the Hellinger distance threshold $\\delta_{\\mathrm{d}}$ . A lower threshold (e.g., 0.65) enforces stricter merging criteria, which preserves object distinctions and leads to higher object recall. In contrast, a higher threshold (e.g., 0.85) enables more extensive relationship aggregation between objects, which improves relationship performance.", + "bbox": [ + 511, + 542, + 905, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 647, + 633, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This study presented FROSS, an innovative framework for real-time generation of 3D SSG in online environments. By representing objects as 3D Gaussian distributions, FROSS achieved computational efficiency that exceeded real-time requirements. The framework integrated a real-time object detection and relationship extraction pipeline to generate 2D SGs, which underwent 3D back-projection and incremental integration into a global SSG through our proposed merging algorithm. Experimental evaluations on existing benchmark and our newly proposed dataset, ReplicaSSG, demonstrated that FROSS surpassed existing methods in both performance metrics and computational speed by significant margins. These advancements established possibilities for real-time SSG applications in open-world environments, particularly relevant to robotics and AR domains.", + "bbox": [ + 511, + 672, + 906, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28825", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was partly supported by the Japan Society for the Promotion of Science KAKENHI Grant Numbers JP24H00733. The authors gratefully acknowledge the support from the National Science and Technology Council (NSTC) in Taiwan under grant numbers NSTC 114-2221-E-002-069-MY3, NSTC 113-2221-E-002-212-MY3, NSTC 114-2218-E-A49-026, and NSTC 114-2640-E-002-006. The authors would also like to express their appreciation for the donation of the GPUs from NVIDIA Corporation and NVIDIA AI Technology Center (NVAITC) used in this work. Furthermore, the authors extend their gratitude to the National Center for High-Performance Computing for providing the necessary computational and storage resources.", + "bbox": [ + 89, + 114, + 485, + 327 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 343, + 187, + 359 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Iro Armeni, Zhi-Yang He, JunYoung Gwak, Amir R Zamir, Martin Fischer, Jitendra Malik, and Silvio Savarese. 3D scene graph: A structure for unified semantics, 3D space, and camera. In Int. Conf. Comput. Vis., pages 5664-5673, 2019. 1, 3", + "[2] Carlos Campos, Richard Elvira, Juan J Gómez Rodríguez, José MM Montiel, and Juan D Tardós. ORB-SLAM3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Trans. Robotics, 37(6):1874-1890, 2021. 2, 8", + "[3] Ata Çelen, Guo Han, Konrad Schindler, Luc Van Gool, Iro Armeni, Anton Obukhov, and Xi Wang. I-Design: Personalized llm interior designer. arXiv:2404.02838, 2024. 2", + "[4] David F Crouse, Peter Willett, Krishna Pattipati, and Lennart Svensson. A look at gaussian mixture reduction algorithms. In Int. Conf. Inform. Fusion, pages 1-8. IEEE, 2011. 5", + "[5] Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and William Yang Wang. LayoutGPT: Compositional visual planning and generation with large language models. Adv. Neural Inform. Process. Syst., 36, 2024. 2", + "[6] Maria Fox and Derek Long. PDDL2.1: An extension to PDDL for expressing temporal planning domains. J. Artificial Intell. Research, 20:61-124, 2003. 2", + "[7] Gege Gao, Weiyang Liu, Anpei Chen, Andreas Geiger, and Bernhard Schölkopf. GraphDreamer: Compositional 3D scene synthesis from scene graphs. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2", + "[8] Lizhao Gao, Bo Wang, and Wenmin Wang. Image captioning with scene-graph based semantic concepts. In Int. Conf. on Mach. Learn. Comput., pages 225-229, 2018. 2", + "[9] Paul Gay, James Stuart, and Alessio Del Bue. Visual graphs from motion (VGfM): Scene understanding with object geometry reasoning. In ACCV, pages 330-346. Springer, 2019. 1, 2, 3, 4, 6, 7", + "[10] Jörg Hoffmann. FF: The fast-forward planning system. AI magazine, 22(3):57-57, 2001. 2" + ], + "bbox": [ + 93, + 368, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Nathan Hughes, Yun Chang, and Luca Carlone. Hydra: A real-time spatial perception system for 3D scene graph construction and optimization. In Robotics: Sci. Syst., 2022. 3", + "[12] Jinbae Im, JeongYeon Nam, Nokyung Park, Hyungmin Lee, and Seunghyun Park. EGTR: Extracting graph from transformer for scene graph generation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 24229-24238, 2024. 2, 3, 4, 6, 8, 1", + "[13] Justin Johnson, Ranjay Krishna, Michael Stark, Li-Jia Li, David Shamma, Michael Bernstein, and Li Fei-Fei. Image retrieval using scene graphs. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3668-3678, 2015. 2, 6", + "[14] Justin Johnson, Agrim Gupta, and Li Fei-Fei. Image generation from scene graphs. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1219–1228, 2018. 2", + "[15] Dong-Jin Kim, Jinsoo Choi, Tae-Hyun Oh, and In So Kweon. Dense relational captioning: Triple-stream networks for relationship-based captioning. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6271-6280, 2019. 2", + "[16] Ue-Hwan Kim, Jin-Man Park, Taek-Jin Song, and Jong-Hwan Kim. 3-D scene graph: A sparse and semantic representation of physical environments for intelligent agents. IEEE Trans. Cybernetics, 50(12):4921-4933, 2019. 2, 3, 6, 7, 8", + "[17] Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In Int. Conf. Learn. Represent., 2017. 3", + "[18] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual Genome: Connecting language and vision using crowdsourced dense image annotations. Int. J. Comput. Vis., 123:32-73, 2017. 2, 6, 8, 3", + "[19] Linjie Li, Zhe Gan, Yu Cheng, and Jingjing Liu. Relation-aware graph attention network for visual question answering. In Int. Conf. Comput. Vis., pages 10313-10322, 2019. 2", + "[20] Yuanyuan Liu, Chengjiang Long, Zhaoxuan Zhang, Bokai Liu, Qiang Zhang, Baocai Yin, and Xin Yang. Explore contextual information for 3D scene graph generation. IEEE Trans. Vis. Comput. Graph., 29(12):5556-5568, 2022. 3, 6", + "[21] Wenyu Lv, Yian Zhao, Qinyao Chang, Kui Huang, Guanzhong Wang, and Yi Liu. RT-DETRv2: Improved baseline with bag-of-freebies for real-time detection transformer. arXiv:2407.17140, 2024. 3, 6, 8", + "[22] Jeffri Murragarra-Llerena, Lucas N. Kirsten, Luis Felipe Zeni, and Claudio R. Jung. Probabilistic intersection-over-union for training and evaluation of oriented object detectors. IEEE Trans. Image Process., 33:671-681, 2024. 4, 5", + "[23] Pushmeet Kohli Nathan Silberman, Derek Hoiem and Rob Fergus. Indoor segmentation and support inference from RGBD images. In Eur. Conf. Comput. Vis., 2012. 6", + "[24] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. PointNet: Deep learning on point sets for 3D classification and segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 652-660, 2017. 3", + "[25] Tianwen Qian, Jingjing Chen, Shaoxiang Chen, Bo Wu, and Yu-Gang Jiang. Scene graph refinement network for vi" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28826", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sual question answering. IEEE Trans. Multimedia, 25:3950-3961, 2022. 2", + "[26] Antoni Rosinol, Andrew Violette, Marcus Abate, Nathan Hughes, Yun Chang, Jingnan Shi, Arjun Gupta, and Luca Carlone. Kimera: From SLAM to spatial perception with 3D dynamic scene graphs. Int. J. Robotics Research, 40(12-14):1510-1546, 2021. 3", + "[27] Dhruv Shah, Benjamin Eysenbach, Gregory Kahn, Nicholas Rhinehart, and Sergey Levine. ViNG: Learning Open-World Navigation with Visual Goals. In IEEE Int. Conf. Robotics Auto., 2021. 1", + "[28] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke M. Strasdat, Renzo De Nardi, Michael Goesele, Steven Lovegrove, and Richard Newcombe. The Replica dataset: A digital replica of indoor spaces. arXiv:1906.05797, 2019. 2, 6", + "[29] Sinan Tan, Mengmeng Ge, Di Guo, Huaping Liu, and Fuchun Sun. Knowledge-based embodied question answering. IEEE Trans. Pattern Anal. Mach. Intell., 45(10):11948-11960, 2023. 2", + "[30] Johanna Wald, Armen Avetisyan, Nassir Navab, Federico Tombari, and Matthias Niessner. RIO: 3D object instance relocalization in changing indoor environments. In Int. Conf. Comput. Vis., 2019. 3, 5", + "[31] Johanna Wald, Helisa Dhamo, Nassir Navab, and Federico Tombari. Learning 3D semantic scene graphs from 3D indoor reconstructions. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3961-3970, 2020. 1, 2, 3, 5, 6, 7, 8", + "[32] Johanna Wald, Nassir Navab, and Federico Tombari. Learning 3D semantic scene graphs with instance embeddings. Int. J. Comput. Vis., 130(3):630-651, 2022.", + "[33] Ziqin Wang, Bowen Cheng, Lichen Zhao, Dong Xu, Yang Tang, and Lu Sheng. VL-SAT: Visual-linguistic semantics assisted training for 3D semantic scene graph prediction in point cloud. In IEEE Conf. Comput. Vis. Pattern Recog., pages 21560–21569, 2023. 2, 3", + "[34] Shun-Cheng Wu, Johanna Wald, Keisuke Tateno, Nassir Navab, and Federico Tombari. SceneGraphFusion: Incremental 3D scene graph prediction from rgb-d sequences. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7515-7525, 2021. 2, 3, 6, 7, 8, 1", + "[35] Shun-Cheng Wu, Keisuke Tateno, Nassir Navab, and Federico Tombari. Incremental 3D semantic scene graph prediction from RGB sequences. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5064-5074, 2023. 2, 3, 6, 7, 8, 1", + "[36] Danfei Xu, Yuke Zhu, Christopher B Choy, and Li Fei-Fei. Scene graph generation by iterative message passing. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5410-5419, 2017. 2, 6, 7, 1", + "[37] Xu Yang, Kaihua Tang, Hanwang Zhang, and Jianfei Cai. Auto-encoding scene graphs for image captioning. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10685-10694, 2019. 2" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Rowan Zellers, Mark Yatskar, Sam Thomson, and Yejin Choi. Neural Motifs: Scene graph parsing with global context. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5831-5840, 2018. 2, 6, 1, 3", + "[39] Guangyao Zhai, Evin Pinar Örnek, Shun-Cheng Wu, Yan Di, Federico Tombari, Nassir Navab, and Benjamin Busam. Commonsscenes: Generating commonsense 3d indoor scenes with scene graph diffusion. In Adv. Neural Inform. Process. Syst., 2023. 2", + "[40] Guangyao Zhai, Evin Pinar Örnek, Dave Zhenyu Chen, Ruotong Liao, Yan Di, Nassir Navab, Federico Tombari, and Benjamin Busam. Echoscene: Indoor scene generation via information echo over scene graph diffusion. In Eur. Conf. Comput. Vis., pages 167-184. Springer, 2024. 2", + "[41] Chaoyi Zhang, Jianhui Yu, Yang Song, and Weidong Cai. Exploiting edge-oriented reasoning for 3D point-based scene graph analysis. In IEEE Conf. Comput. Vis. Pattern Recog., pages 9705-9715, 2021. 2, 3, 6", + "[42] Chaoyi Zhang, Xitong Yang, Ji Hou, Kris Kitani, Weidong Cai, and Fu-Jen Chu. EgoSG: Learning 3D scene graphs from egocentric RGB-D sequences. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2535-2545, 2024.", + "[43] Shoulong Zhang, Aimin Hao, Hong Qin, et al. Knowledge-inspired 3D scene graph prediction in point cloud. Adv. Neural Inform. Process. Syst., 34:18620-18632, 2021. 3, 6", + "[44] Yian Zhao, Wenyu Lv, Shangliang Xu, Jinman Wei, Guanzhong Wang, Qingqing Dang, Yi Liu, and Jie Chen. DETRs beat YOLOs on real-time object detection. In IEEE Conf. Comput. Vis. Pattern Recog., pages 16965-16974, 2024. 3, 4, 6, 8, 2", + "[45] Kaizhi Zheng, Xiaotong Chen, Xuehai He, Jing Gu, Linjie Li, Zhengyuan Yang, Kevin Lin, Jianfeng Wang, Lijuan Wang, and Xin Eric Wang. EditRoom: LLM-parameterized graph diffusion for composable 3D room layout editing. arXiv:2410.12836, 2024. 2", + "[46] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. EWA splatting. IEEE Trans. Vis. Comput. Graph., 8(3):223-238, 2002. 4" + ], + "bbox": [ + 516, + 92, + 903, + 628 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28827", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/FROSS_ Faster-Than-Real-Time Online 3D Semantic Scene Graph Generation from RGB-D Images/f4b25ed8-2175-4e29-a597-e457f4795510_model.json b/2025/FROSS_ Faster-Than-Real-Time Online 3D Semantic Scene Graph Generation from RGB-D Images/f4b25ed8-2175-4e29-a597-e457f4795510_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d8ff151b6f517dbdb887d21066e2da92ec4eae8e --- /dev/null +++ b/2025/FROSS_ Faster-Than-Real-Time Online 3D Semantic Scene Graph Generation from RGB-D Images/f4b25ed8-2175-4e29-a597-e457f4795510_model.json @@ -0,0 +1,2046 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.199, + 0.131, + 0.8, + 0.177 + ], + "angle": 0, + "content": "FROSS: Faster-than-Real-Time Online 3D Semantic Scene Graph Generation from RGB-D Images" + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.203, + 0.258, + 0.221 + ], + "angle": 0, + "content": "Hao-Yu Hou1,3" + }, + { + "type": "text", + "bbox": [ + 0.3, + 0.204, + 0.411, + 0.221 + ], + "angle": 0, + "content": "Chun-Yi Lee2" + }, + { + "type": "text", + "bbox": [ + 0.452, + 0.204, + 0.643, + 0.222 + ], + "angle": 0, + "content": "Motoharu Sonogashira3" + }, + { + "type": "text", + "bbox": [ + 0.685, + 0.204, + 0.861, + 0.221 + ], + "angle": 0, + "content": "Yasutomo Kawanishi3" + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.239, + 0.426, + 0.257 + ], + "angle": 0, + "content": "\\(^{1}\\)National Tsing Hua University" + }, + { + "type": "text", + "bbox": [ + 0.474, + 0.239, + 0.7, + 0.257 + ], + "angle": 0, + "content": "\\(^{2}\\)National Taiwan University" + }, + { + "type": "text", + "bbox": [ + 0.748, + 0.239, + 0.819, + 0.256 + ], + "angle": 0, + "content": "\\(^{3}\\)RIKEN" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.325, + 0.485, + 0.704 + ], + "angle": 0, + "content": "The ability to abstract complex 3D environments into simplified and structured representations is crucial across various domains. 3D semantic scene graphs (SSGs) achieve this by representing objects as nodes and their interrelationships as edges, facilitating high-level scene understanding. Existing methods for 3D SSG generation, however, face significant challenges, including high computational demands and non-incremental processing that hinder their suitability for real-time open-world applications. To address this issue, we propose FROSS (Faster-than-Real-Time Online 3D Semantic Scene Graph Generation), an innovative approach for online and faster-than-realtime 3D SSG generation that leverages the direct lifting of 2D scene graphs to 3D space and represents objects as 3D Gaussian distributions. This framework eliminates the dependency on precise and computationally-intensive point cloud processing. Furthermore, we extend the Replica dataset with inter-object relationship annotations, creating the ReplicaSSG dataset for comprehensive evaluation of FROSS. The experimental results from evaluations on ReplicaSSG and 3DSSG datasets show that FROSS can achieve superior performance while operating significantly faster than prior 3D SSG generation methods. Our implementation and dataset are publicly available at https://github.com/Howardkhh/FROSS." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.723, + 0.222, + 0.739 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The abstraction of complex 3D environments into simplified representations in real time serves as a crucial capability across multiple application domains. The incremental generation of such representations in real time holds particular significance for fields ranging from robotics and computer vision to augmented reality (AR), where systems must perceive and analyze the environment instantaneously. 3D scene graphs (SGs) [1, 9] accomplish this abstraction by representing physical objects as nodes and their interrelationships as edges within a structured framework. 3D se" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.305, + 0.907, + 0.665 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.673, + 0.907, + 0.73 + ], + "angle": 0, + "content": "Figure 1. We introduce FROSS, an online real-time 3D semantic scene graph generation method that leverages and integrates 2D scene graphs. FROSS represents objects as 3D Gaussian distributions and operates without requiring 3D reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.903 + ], + "angle": 0, + "content": "mantic scene graphs (SSGs) [31] extend this representation with an emphasis on semantic relationships, such as positional configurations between objects (e.g., a cup on a table) or functional interactions (e.g., a person seated in a chair), which enables high-level scene comprehension, reasoning, and interaction. Real-world applications, however, present open-world challenges where environments often exceed known spatial boundaries and contain previously unseen spaces [27]. For instance, a room may connect to multiple adjacent spaces or corridors, which necessitates continuous" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28818" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "adaptation and incremental scene understanding. This construction process must maintain high efficiency to generate structured environmental representations without creating performance bottlenecks. Meeting the processing demands requires the generation of 3D SSG to operate at faster-than-real-time speeds, with all computations completed between successive image frames from a high-speed sensor stream." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.201, + 0.483, + 0.548 + ], + "angle": 0, + "content": "Executing all processes at faster-than-real-time rates confers multiple advantages: (1) prevention of bottlenecks in the 3D SSG generation framework, (2) facilitation of smooth operation on resource-constrained hardware such as edge computing devices, and (3) preservation of computational resources for additional system components. Contemporary 3D SSG generation approaches predominantly focus on offline construction, which requires access to the complete scene data, either instance-segmented point clouds [31, 41], a complete image sequence [9], or both [33]. This assumption of full scene availability conflicts with incremental environmental exploration. Although several studies [34, 35] have developed methods for incremental point cloud reconstruction and segmentation to achieve online and real-time 3D SSG generation, their aggregate system latency remains substantial, even with the utilization of multiple CPU cores. This constraint stems mainly from the substantial computational demands of environmental mapping and point cloud processing, restricting their practical application in real-time systems with limited computational power or other concurrent processes. Given these limitations, the aim of this study is to develop a method for faster-than-real-time online SSG generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.483, + 0.777 + ], + "angle": 0, + "content": "A fundamental insight for generating 3D SSGs emerges from the observation that precise object pose and shape information is not essential. Given that 3D SSGs provide a high-level semantic understanding of the environment, their functional implications remain valid with approximate object locations. For instance, in robotic applications, high-level planning can be successfully executed using relative spatial relationships [6, 10, 29]. Similarly, in 3D scene synthesis and room layout generation [3, 5, 7, 39, 40, 45], 3D SSGs serve as foundational structures from which more detailed content can be generated. These application scenarios eliminate the need for precise environmental mapping techniques such as Simultaneous Localization and Mapping (SLAM) [2] and additional point cloud processing, both of which contribute significant computational overhead." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.902 + ], + "angle": 0, + "content": "This observation motivates an alternative approach for online and real-time 3D SSG generation: inferring 2D SGs from images and subsequently lifting them into 3D space, thus circumventing the computational burden of environmental mapping and point cloud processing. To the best of our knowledge, the current literature has not adequately investigated this research direction except [16]. Such unexplored territory presents significant challenges, includ" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.181 + ], + "angle": 0, + "content": "ing the handling of incomplete 2D object observations, the achievement of real-time SG generation, and the effective integration of multi-view information into complete 3D SSGs without object redundancy. These challenges, therefore, provide promising avenues for further innovative research contributions." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.183, + 0.907, + 0.513 + ], + "angle": 0, + "content": "In light of these considerations, we introduce FROSS (Faster-than-Real-time Online 3D Semantic Scene Graph Generation), a methodology that constructs 3D SSGs by lifting 2D SGs to 3D space and integrating them into a comprehensive representation. To achieve real-time processing without requiring precise object localization, FROSS initiates the process by constructing 2D SGs from detected objects in images. A key innovation of FROSS lies in its utilization of Gaussian distributions to model objects in 3D space, which facilitates the integration of objects observed across multiple images into a coherent 3D SSG, as illustrated in Figure 1. Specifically, FROSS represents detected objects through Gaussian distributions, lifts these distributions from 2D SGs into 3D space, and merges them according to their predicted classes and statistical distances. We validate FROSS on the 3DSSG dataset [31] and conduct additional evaluations on ReplicaSSG, an extended version of the Replica dataset [28] that we have enhanced with object relationship annotations. ReplicaSSG incorporates object classes defined in Visual Genome [18], a widely-adopted 2D SG dataset. The main contributions of the paper can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.515, + 0.905, + 0.574 + ], + "angle": 0, + "content": "- We introduce FROSS, an innovative methodology for online real-time generation of 3D SSGs. FROSS demonstrates superior performance and significantly faster processing speeds compared to existing baseline methods." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.576, + 0.905, + 0.65 + ], + "angle": 0, + "content": "- We propose a new merging algorithm based on Gaussian distributions for 3D object integration. This algorithm effectively prevents the proliferation of duplicate or incorrectly detected objects as the input image sequence expands, thereby maintaining computational efficiency." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.651, + 0.905, + 0.726 + ], + "angle": 0, + "content": "- We develop ReplicaSSG, an extension of the Replica dataset with 3D SSG annotations. Comprehensive evaluations of FROSS on both the 3DSSG and ReplicaSSG datasets demonstrate FROSS's effectiveness and computational efficiency over diverse environmental scenarios." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.515, + 0.905, + 0.726 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.74, + 0.655, + 0.755 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "SGs [12, 13, 36, 38] constitute a fundamental data structure for object and relationship representation within images following their introduction by [13]. This representation has demonstrated substantial utility across various scene understanding tasks [8, 14, 15, 19, 25, 37]. The concept later evolved to encompass 3D spaces, where objects may incorporate attributes such as class labels, visual features, shape characteristics, and affordances. Research in 3D SGs developed into two branches of representation: SSGs intro" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28819" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.288 + ], + "angle": 0, + "content": "duced in [9, 16, 31] and hierarchical scene graphs (HSGs) proposed in [1, 11, 26]. Both frameworks represent 3D objects as nodes with pairwise relationships defined through edges. However, these approaches differ fundamentally in their relationship definitions. SSGs concentrate on semantic relationships between objects, while HSGs organize objects into hierarchical layers that include buildings, rooms, and constituent objects, where edges connect only nodes in identical or adjacent layers. Compared to HSGs, SSGs enable higher-level reasoning capabilities through rich semantic relationship representations. The extension beyond spatial relationships allows SSGs to be applied to broader domains such as robotics, 3D computer vision, and AR." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.3, + 0.479, + 0.317 + ], + "angle": 0, + "content": "2.1. Offline 3D Semantic Scene Graph Generation" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.323, + 0.485, + 0.641 + ], + "angle": 0, + "content": "The initial development of 3D SSG generation focused primarily on offline settings that required either complete instance-segmented point clouds [20, 31-33, 41, 43] or image sequences [9, 33] as inputs. Such requirements limited their applicability to scenarios with incremental data collection, as they necessitated complete data reprocessing. Despite these limitations, these studies established foundations for 3D SSG generation. The research in [9] pioneered 3D SSG generation by predicting 3D object ellipsoids from image collections and utilizing their geometry and appearance features through a recurrent neural network to extract inter-object relationships. The study in [31] advanced this field through the introduction of the 3DSSG dataset, which enhanced the 3RScan dataset [30] with SSG annotations. This dataset represents the first large-scale collection with dense annotations and has become the primary benchmark for evaluating 3D SSG generation methods. The authors of [31] further contributed a generation methodology by utilizing instance-segmented point clouds as input and leveraged PointNet [24] for feature extraction as well as a graph convolutional network [17] for label prediction." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.652, + 0.476, + 0.669 + ], + "angle": 0, + "content": "2.2. Online 3D Semantic Scene Graph Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Several previous studies have explored 3D SG generation in online settings, with some achieving real-time performance. The research in [16] introduced a framework that predicts local 3D SSGs from RGB-D images and integrates them incrementally into a global representation. However, their computational requirements render real-time execution infeasible. SceneGraphFusion [34] and the study in [35] developed methods for online point cloud reconstruction and segmentation, followed by feature extraction for local SSG prediction and global fusion. These frameworks can achieve real-time performance through multi-threaded execution of parallel components. In contrast to these existing methods, our proposed FROSS framework avoids precise object localization by approximating objects with Gaussian distributions. This approximation approach enables faster-than" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.138 + ], + "angle": 0, + "content": "real-time online 3D SSG generation, allowing the end-to-end processing latency to be below the sensor data acquisition period while maintaining single-threaded execution." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.152, + 0.65, + 0.17 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.178, + 0.697, + 0.193 + ], + "angle": 0, + "content": "3.1. Problem Definition" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.2, + 0.907, + 0.472 + ], + "angle": 0, + "content": "The primary problem concerned in this study is the online generation of 3D SSGs for environments where the complete scene structure remains unknown a priori. Given a sequence of input images, the primary objective is to construct a 3D SSG \\(\\mathcal{G}\\) of the target environment. The graph \\(\\mathcal{G}\\) consists of a set of nodes \\(\\nu\\) and their corresponding directed edges \\(\\mathcal{E}\\). Each node \\(v_{i} = (c_{i},\\mu_{i}^{3\\mathrm{D}},\\Sigma_{i}^{3\\mathrm{D}})\\in \\mathcal{V}\\) is indexed by \\(i\\in \\{1,\\ldots ,N\\}\\), representing a unique object in the scene. Every object has a class label \\(c_{i}\\in \\mathcal{C}\\), where \\(\\mathcal{C}\\) is the set of object categories. While different approaches may incorporate various node attributes, our formulation represents each object as a 3D Gaussian distribution parameterized by mean vector \\(\\mu_i^{3\\mathrm{D}}\\in \\mathbb{R}^3\\) and covariance matrix \\(\\Sigma_i^{3\\mathrm{D}}\\in \\mathbb{R}^{3\\times 3}\\). The edges \\(e_{i\\rightarrow j}\\in \\mathcal{E}\\) connect node \\(v_{i}\\) to node \\(v_{j}\\) with relationship \\(r_{i\\rightarrow j}\\in \\mathcal{R}\\), where \\(\\mathcal{R}\\) represents the set of predefined relationship types. Furthermore, the algorithm must be able to maintain online processing capabilities to update \\(\\mathcal{G}\\) continuously whenever new image data becomes available." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.482, + 0.74, + 0.497 + ], + "angle": 0, + "content": "3.2. Overview of Framework" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.505, + 0.907, + 0.763 + ], + "angle": 0, + "content": "Figure 2 presents an overview of the proposed FROSS framework, which processes image sequences with corresponding depth information and camera transformations as input. The framework begins with the EGTR [12] 2D SG generation model, with the object detection backbone model replaced with RT-DETR [21, 44]. This RT-DETR object detector is a state-of-the-art real-time detection model, which preserves intermediate self-attention features for subsequent relationship extraction. FROSS then converts object bounding boxes to 2D Gaussian distributions and projects them into 3D space to generate local SSGs, as elaborated in Section 3.3. These Gaussian distributions serve as efficient approximations of both object positions and their spatial distributions for rapid processing. These local graphs are then integrated into a global SSG through the proposed merging algorithm that is described in detail in Section 3.4." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.773, + 0.707, + 0.789 + ], + "angle": 0, + "content": "3.3. Lifting 2D SG to 3D" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.903 + ], + "angle": 0, + "content": "The 2D SGs generated from RT-DETR [21, 44] and EGTR [12] enable FROSS to construct local 3D SSGs \\(\\mathcal{G}_{\\mathrm{local}} = (\\mathcal{V}_{\\mathrm{local}},\\mathcal{E}_{\\mathrm{local}})\\) through the integration of depth information and camera transformation. This process involves two key steps: (1) the conversion of object bounding boxes to 2D Gaussians, and (2) their subsequent back-projection into 3D space while preserving inter-object relationships." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28820" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.126, + 0.09, + 0.852, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.282, + 0.908, + 0.354 + ], + "angle": 0, + "content": "Figure 2. An overview of the FROSS framework: (a) The process initiates with object detection via RT-DETR [44] from an RGB-D image and its associated camera pose. (b) The EGTR model [12] extracts inter-object relationships through utilization of preserved self-attention features from RT-DETR. (c) 2D Gaussian distributions derived from detected bounding boxes undergo spatial transformation into 3D space for local 3D SSG construction (Section 3.3). (d) The resultant local 3D SSG is subsequently incorporated into the global 3D SSG through our proposed merging algorithm (Section 3.4)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.366, + 0.484, + 0.579 + ], + "angle": 0, + "content": "2D Gaussian Representation of Bounding Boxes. Although several methods exist for lifting 2D bounding boxes to 3D space, including point cloud back-projection and joint optimization of 3D ellipsoids [9], the Gaussian distribution approach emerges as the most computationally efficient solution among these alternatives. The representation of objects via Gaussian distributions requires the specification of mean \\(\\mu_i^{2\\mathrm{D}}\\) and covariance matrix \\(\\Sigma_i^{2\\mathrm{D}}\\) of the 2D object node \\(v_i^{2\\mathrm{D}}\\), which would correspond to a 3D node \\(v_i \\in \\mathcal{V}_{\\mathrm{local}}\\). Based on the formulation in [22], FROSS models bounding boxes as 2D uniform distributions. The mean values correspond to the bounding box centers, while the covariance matrices maintain consistency with the uniform distributions. The covariance matrix of \\(v_i^{2\\mathrm{D}}\\) is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.204, + 0.584, + 0.483, + 0.618 + ], + "angle": 0, + "content": "\\[\n\\Sigma_ {i} ^ {\\mathrm {2 D}} = \\frac {1}{1 2} \\left[ \\begin{array}{c c} W _ {i} ^ {2} & 0 \\\\ 0 & H _ {i} ^ {2} \\end{array} \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.622, + 0.484, + 0.744 + ], + "angle": 0, + "content": "where \\( W_{i} \\) and \\( H_{i} \\) are 2D bounding box's width and height. 3D Back-Projection of Gaussian Distributions. Given the 2D Gaussians, their corresponding depths at the Gaussian centroids, and camera transformation matrices, the mean positions \\( \\mu_{i}^{3\\mathrm{D}} \\) and covariances \\( \\Sigma_{i}^{3\\mathrm{D}} \\) of the Gaussians in the 3D space can be approximated. The 3D mean \\( \\mu_{i}^{3\\mathrm{D}} \\) can be derived from the 2D mean \\( \\mu_{i}^{2\\mathrm{D}} = (p_{x}, p_{y})^{\\top} \\) and its corresponding depth value \\( z \\) through the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.161, + 0.748, + 0.483, + 0.767 + ], + "angle": 0, + "content": "\\[\n\\mu_ {i} ^ {\\mathrm {3 D}} = R \\cdot K ^ {- 1} \\cdot \\left(p _ {x}, p _ {y}, 1\\right) ^ {\\top} \\cdot z + \\mathbf {t}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.772, + 0.484, + 0.864 + ], + "angle": 0, + "content": "where \\(K \\in \\mathbb{R}^{3 \\times 3}\\) represents the camera intrinsic matrix, and \\(R \\in \\mathbb{R}^{3 \\times 3}\\) and \\(\\mathbf{t} \\in \\mathbb{R}^3\\) denote the rotation matrix and translation vector of the current camera, respectively. The approximation of the 3D covariance matrix utilizes the Jacobian \\(J\\) of the local affine approximation of the projective transformation [46], which is defined as the following:" + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.868, + 0.483, + 0.906 + ], + "angle": 0, + "content": "\\[\nJ = \\left[ \\begin{array}{c c c} \\frac {f _ {x}}{z} & 0 & - \\frac {f _ {x} p _ {x}}{z ^ {2}} \\\\ 0 & \\frac {f _ {y}}{z} & - \\frac {f _ {y} p _ {y}}{z ^ {2}} \\end{array} \\right], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.366, + 0.907, + 0.458 + ], + "angle": 0, + "content": "where \\( f_{x} \\) and \\( f_{y} \\) represent the focal lengths in \\( x \\) and \\( y \\) dimensions. In the original literature of [46], the Jacobian \\( J \\) is utilized to project 3D covariance matrices onto the 2D image plane. However, our approach inverts this process by back-projecting the 2D covariance matrix \\( \\Sigma_{i}^{2\\mathrm{D}} \\) into 3D space as \\( \\bar{\\Sigma}_{i}^{3\\mathrm{D}} \\) via the pseudo-inverse \\( J^{+} \\) of the Jacobian:" + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.465, + 0.907, + 0.485 + ], + "angle": 0, + "content": "\\[\n\\bar {\\Sigma} _ {i} ^ {\\mathrm {3 D}} = R J ^ {+} \\Sigma_ {i} ^ {\\mathrm {2 D}} J ^ {+ \\top} R ^ {\\top}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.907, + 0.646 + ], + "angle": 0, + "content": "The naive definition alone is insufficient since the original equation is inherently designed for projection from 3D to 2D space. Simply inverting the equation results in a 3D Gaussian that lacks variance information along the depth axis, as the 2D covariance matrix does not capture depth-related variance. This limitation necessitates the specification of variance in the depth dimension. We hypothesize that this variance approximates the average variance of the other dimensions. The specification has to be applied prior to the rotation matrix \\( R \\), leading to the following equations:" + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.653, + 0.907, + 0.672 + ], + "angle": 0, + "content": "\\[\n\\Sigma_ {i} ^ {\\mathrm {3 D / /}} = J ^ {+} \\Sigma_ {i} ^ {\\mathrm {2 D}} J ^ {+ \\top}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.675, + 0.906, + 0.729 + ], + "angle": 0, + "content": "\\[\n\\Sigma_ {i} ^ {\\mathrm {3 D} \\prime} = \\Sigma_ {i} ^ {\\mathrm {3 D} \\prime \\prime} + \\left[ \\begin{array}{l l l} 0 & 0 & 0 \\\\ 0 & 0 & 0 \\\\ 0 & 0 & \\frac {(\\Sigma_ {i} ^ {\\mathrm {3 D} \\prime \\prime}) _ {1 , 1} + (\\Sigma_ {i} ^ {\\mathrm {3 D} \\prime \\prime}) _ {2 , 2}}{2} \\end{array} \\right]. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.737, + 0.908, + 0.799 + ], + "angle": 0, + "content": "In this formulation, \\((\\Sigma_{i}^{3\\mathrm{D} / \\prime})_{m,n}\\in \\mathbb{R}\\) represents the element at position \\((m,n)\\) of matrix \\(\\Sigma_i^{3\\mathrm{D} / \\prime}\\). Finally, the covariance matrix \\(\\Sigma_{i}^{3\\mathrm{D}^{\\prime}}\\) is transformed to world coordinates using the rotation matrix \\(R\\), resulting in the following expression:" + }, + { + "type": "equation", + "bbox": [ + 0.647, + 0.806, + 0.907, + 0.825 + ], + "angle": 0, + "content": "\\[\n\\Sigma_ {i} ^ {\\mathrm {3 D}} = R \\Sigma_ {i} ^ {\\mathrm {3 D} \\prime} R ^ {\\top}. \\tag {7}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.834, + 0.691, + 0.851 + ], + "angle": 0, + "content": "3.4. Merging 3D SSGs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.855, + 0.908, + 0.903 + ], + "angle": 0, + "content": "With a local 3D SSG \\(\\mathcal{G}_{\\mathrm{local}} = (\\mathcal{V}_{\\mathrm{local}},\\mathcal{E}_{\\mathrm{local}})\\) constructed, it is then integrated with the global 3D SSG \\(\\mathcal{G}_{\\mathrm{global}}\\) through Gaussian merging and relationship accumulation. Figure 3" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28821" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.289 + ], + "angle": 0, + "content": "illustrates the proposed merging process. Specifically, it initializes a queue with nodes from \\(\\nu_{\\mathrm{local}}\\) and evaluates potential merging opportunities through distance calculations. For each node \\(v_{i}\\) in this queue, FROSS calculates the Hellinger distances [22] between the Gaussian distributions of \\(v_{i}\\) and all other nodes \\(v_{j}\\) of the same semantic class, in both the local node queue and \\(\\nu_{\\mathrm{global}}\\), to identify object correspondence. The algorithm merges nodes with Hellinger distances below the threshold \\(\\delta_{\\mathrm{d}}\\) and adds unmerged nodes to \\(\\nu_{\\mathrm{global}}\\) as new objects. The Hellinger distance \\(H_{D}(i,j)\\) between Gaussian distributions \\(\\mathcal{N}(\\mu_i,\\Sigma_i)\\) and \\(\\mathcal{N}(\\mu_j,\\Sigma_j)\\) is quantified through the following equation, wherein \\(B_{D}(i,j)\\) represents the Bhattacharyya distance:" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.296, + 0.483, + 0.316 + ], + "angle": 0, + "content": "\\[\nH _ {D} (i, j) = \\sqrt {1 - \\exp (- B _ {D} (i , j))}, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.109, + 0.318, + 0.482, + 0.367 + ], + "angle": 0, + "content": "\\[\nB _ {D} (i, j) = \\frac {1}{8} \\Delta \\mu_ {i j} ^ {T} \\Sigma^ {- 1} \\Delta \\mu_ {i j} + \\frac {1}{2} \\ln \\left(\\frac {\\operatorname* {d e t} \\Sigma}{\\sqrt {\\operatorname* {d e t} \\Sigma_ {i} \\operatorname* {d e t} \\Sigma_ {j}}}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.378, + 0.484, + 0.56 + ], + "angle": 0, + "content": "where \\(\\Delta \\mu_{ij} = \\mu_i - \\mu_j\\) and \\(\\Sigma = \\frac{1}{2} (\\Sigma_i + \\Sigma_j)\\). The fusion of these nodes occurs through a weighted integration of their respective Gaussian distributions [4]. The weighting coefficients for this integration process reflect both spatial coverage and detection confidence across multiple viewpoints. Objects detected from diverse angular perspectives and spatial positions receive proportionally higher weights, thus mitigating viewpoint bias and ensuring robust three-dimensional representation. Given that weights \\(w_i\\) and \\(w_j\\) represent the detection frequencies of their respective nodes, the mean vector \\(\\mu_k\\) and covariance matrix \\(\\Sigma_k\\) for the resultant merged node are determined via the formulations:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.567, + 0.483, + 0.598 + ], + "angle": 0, + "content": "\\[\n\\mu_ {k} = \\frac {w _ {i} \\mu_ {i} + w _ {j} \\mu_ {j}}{w _ {i} + w _ {j}}, \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.6, + 0.483, + 0.636 + ], + "angle": 0, + "content": "\\[\n\\Sigma_ {k} = \\frac {w _ {i} \\Sigma_ {i} + w _ {j} \\Sigma_ {j}}{w _ {i} + w _ {j}} + \\frac {w _ {i} w _ {j} \\left(\\mu_ {i} - \\mu_ {j}\\right) \\left(\\mu_ {i} - \\mu_ {j}\\right) ^ {\\top}}{\\left(w _ {i} + w _ {j}\\right) ^ {2}}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.198 + ], + "angle": 0, + "content": "where \\(\\mu_{i},\\mu_{j}\\) are the mean vectors, and \\(\\Sigma_{i},\\Sigma_{j}\\) are the covariance matrices of the respective Gaussian distributions to be merged. All relationship edges originally connected to \\(v_{i}\\) and \\(v_{j}\\) are then redirected to the newly merged node \\(v_{k}\\). The final relationship prediction between two global object nodes is selected through majority voting, choosing the most frequent prediction across viewpoints." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.212, + 0.72, + 0.229 + ], + "angle": 0, + "content": "4. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.238, + 0.907, + 0.435 + ], + "angle": 0, + "content": "The evaluation of FROSS encompasses two 3D SSG datasets: 3DSSG [31] and our developed ReplicaSSG datasets. Section 4.1 introduces the datasets, baseline SSG generation methods, and evaluation metrics. Section 4.2 presents the implementation details. Section 4.3 provides comprehensive comparisons between FROSS and existing SSG generation methods on the 3DSSG dataset. Moreover, the effectiveness of our approach is further demonstrated through qualitative results in Section 4.4. We further provide runtime analyses on the ReplicaSSG dataset in Section 4.5, along with additional ablation studies in Section 4.6. More quantitative and qualitative analyses are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.445, + 0.683, + 0.46 + ], + "angle": 0, + "content": "4.1. Evaluation Setup" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.467, + 0.619, + 0.48 + ], + "angle": 0, + "content": "4.1.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.486, + 0.909, + 0.637 + ], + "angle": 0, + "content": "3DSSG. The 3DSSG dataset [31] extends 3RScan [30], which encompasses 1,482 scans of indoor environments with their corresponding RGB-D image sequences, 3D meshes, and dense instance segmentation. The dataset contains 363,555 images, with an average of 245 images per scan. 3DSSG augments the base dataset with object attributes, hierarchical category labels, and directed edges that describe inter-object semantic relationships such as 'standing on,' 'attached to,' and 'same color.' This dataset has established a primary benchmark for numerous" + }, + { + "type": "image", + "bbox": [ + 0.174, + 0.661, + 0.825, + 0.83 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.842, + 0.908, + 0.898 + ], + "angle": 0, + "content": "Figure 3. Illustration of the object merging process using the proposed algorithm described in Section 3.4. (a) Starting with an input image, a set of objects with predicted categories and bounding boxes is generated. (b) These objects are then lifted into 3D spaces to form the object nodes in \\(\\mathcal{V}_{\\mathrm{local}}\\) with colors matching their bounding boxes (yellow, cyan, and blue), while pre-existing objects in \\(\\mathcal{V}_{\\mathrm{global}}\\) are shown in green. (c) The merging algorithm combines the objects linked by the arrows, resulting in the updated global objects, represented in red." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28822" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.318 + ], + "angle": 0, + "content": "studies [20, 31-35, 41-43]. Our experiments utilize the publicly available training and testing splits. To address the class imbalance in both object and relationship labels, this study adopts the commonly used category definitions in [31, 34, 35]. Specifically, the original 160 object categories map to 20 categories from NYUv2 [23], while seven of the 26 predicate categories remain. Following this categorical mapping, 3DSSG contains a total of 21,974 distinct objects and 16,324 inter-object relationships. This extensive collection of relationships offers a substantial foundation for spatial and semantic analyses. Nevertheless, the low image quality and low frame rate of 3DSSG hinders the comprehensive evaluation of FROSS at its full operational capacity. This necessitates our development of ReplicaSSG, which is an enhanced adaptation of the Replica dataset [28]." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.322, + 0.485, + 0.655 + ], + "angle": 0, + "content": "ReplicaSSG. ReplicaSSG extends the Replica dataset [28] with newly annotated object relationships. The original Replica provides high-quality reconstructions of indoor environments with instance-segmented meshes and photorealistic renderings, and is particularly suitable for SSG evaluation. Although Replica encompasses only 18 scenes, which precludes its use for training purposes, it serves as an effective evaluation benchmark. In the evaluation of FROSS, seven scenes serve as validation environments for hyperparameter optimization, while the remaining 11 scenes function as performance assessment platforms. Rather than preserving Replica's original object category definitions, ReplicaSSG adopts the classification system from Visual Genome [18], a 2D SG dataset. This label mapping from Replica to Visual Genome facilitates zero-shot transfer from 2D SG models, as this approach eliminates the requirement for additional training on the Replica dataset. The dataset splits and label mapping are available in the released code. In addition, detailed statistics for the dataset are provided in the supplementary materials (Section 8). As a result, ReplicaSSG offers high-quality instance-segmented meshes with comprehensive relationship annotations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.674, + 0.262, + 0.688 + ], + "angle": 0, + "content": "4.1.2. Baseline Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.696, + 0.483, + 0.862 + ], + "angle": 0, + "content": "We evaluated FROSS on the 3DSSG dataset against several baseline methods for SSG generation. The baselines include three image-based methods, IMP [36], VGfM [9] and Kim's construction framework [16], alongside three point cloud-based methods: 3DSSG [31], SGFN [34] and the approach presented in [35]. Kim's 3D object representation and merging mechanism are integrated into FROSS for its baseline implementation, while other baseline implementations follow \\([35]^1\\). In the ablation studies, we investigate the impact of using ground truth 2D SGs and camera trajectories on the ReplicaSSG dataset." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.092, + 0.882, + 0.107 + ], + "angle": 0, + "content": "4.1.3. Matching Object Predictions to Ground Truth" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.111, + 0.905, + 0.262 + ], + "angle": 0, + "content": "The evaluation methodology follows [35] for matching predicted object instances with ground truth counterparts. As FROSS generates predictions without explicit point cloud output, we establish evaluation metrics using back-projected 3D points. These points derive from the central region \\((H / 2\\times W / 2\\) pixels) of each object's 2D bounding box. The matching process associates each predicted point with its nearest ground truth point within a 0.1-meter radius. A predicted object establishes correspondence with a ground truth object under two criteria, described as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.263, + 0.905, + 0.308 + ], + "angle": 0, + "content": "1. The majority (more than \\(50\\%\\)) of its back-projected 3D points must each have their nearest ground truth point neighbor within the corresponding ground truth object." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.309, + 0.905, + 0.337 + ], + "angle": 0, + "content": "2. The ratio between the second-largest and largest ground truth object overlap counts must remain below \\(75\\%\\)." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.263, + 0.905, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.34, + 0.905, + 0.505 + ], + "angle": 0, + "content": "In contrast to the other baseline methods, which may permit multiple predictions to be matched with a single ground truth object due to possible over-segmentation in input instance-segmented point clouds, our proposed FROSS framework enforces a stricter one-to-one correspondence between predicted and ground truth objects. Since FROSS merges identical objects into a single prediction, matching multiple predictions to the same ground truth object is not required. The one-to-one correspondence enforces the integrity of the proposed merging algorithm, ensuring accurate association between predicted and ground truth objects." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.515, + 0.694, + 0.529 + ], + "angle": 0, + "content": "4.1.4. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.907, + 0.777 + ], + "angle": 0, + "content": "The evaluation of SSG generation employs recall metrics for relationships, objects, and predicate, a standard approach in both 2D and 3D SG evaluation [13, 31, 34, 35, 38]. Object recall quantifies the proportion of ground truth instances matched to predictions with correct category labels. Predicate recall computes the proportion of correctly classified predicates between detected objects, regardless of the object classes. Relationship recall measures the proportion of correctly identified triplets (subject, object, predicate) that meet two criteria: valid detection of both subject and object, and correct classification of their connecting predicate. To mitigate the impact of class imbalance, we report mean recall (mRecall) to provide a more equitable evaluation across all classes. Note that a more detailed explanation of the utilized recall metric is offered in the supplementary materials (Section 6)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.788, + 0.731, + 0.803 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.905, + 0.901 + ], + "angle": 0, + "content": "For our 2D SG generation model, we employ EGTR [12] with RT-DETRv2-M [21, 44] as the object detection backbone. The 2D SGs from the 3DSSG dataset [31] serve as training data specifically for 3DSSG experiments, while the Visual Genome dataset [18] functions as the training foundation for ReplicaSSG experiments. Our methodology ad" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.455, + 0.9 + ], + "angle": 0, + "content": "1https://github.com/ShunChengWu/3DSSG/tree/4b783ec" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28823" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.483, + 0.173 + ], + "angle": 0, + "content": "Table 1. Performance comparison of 3D SSG generation methods on the 3DSSG dataset and the end-to-end latency without environmental mapping reported in their original literature, along with their respective input modalities. The best and second-best results are highlighted in red, and blue, respectively. Latencies are reported in milliseconds." + }, + { + "type": "table", + "bbox": [ + 0.09, + 0.18, + 0.482, + 0.302 + ], + "angle": 0, + "content": "
MethodRecall (%)mRecall (%)LatencyInput Modality
Rel.Obj.Pred.Obj.Pred.
IMP [36]19.749.520.934.713.8-RGB-D
VGfM [9]19.650.020.434.811.0250RGB
3DSSG [31]12.937.422.026.214.4-Point Cloud
SGFN [34]22.051.627.537.724.0161RGB-D
Wu [35]23.353.828.443.826.6191RGB-D2
Kim [16]9.159.07.151.08.0310RGB-D
FROSS (Ours)27.962.433.063.818.07RGB-D
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.329, + 0.483, + 0.525 + ], + "angle": 0, + "content": "heres to the training protocol established in EGTR, which initiates with pre-training the object detector for standard object detection tasks. All hyperparameters were determined through grid search evaluation on the validation split, with particular emphasis on relationship recall optimization. Our implementation applies a confidence threshold of 0.7 for object filtering and retains only the top ten relationships per 2D SG. For Gaussian merging operations, the Hellinger distance threshold \\(\\delta_{\\mathrm{d}}\\) is established at 0.85. The main experimental evaluations utilize ground truth trajectories to guide the SSG generation process. The impact of estimated trajectories are further analyzed via ablation studies in Section 4.6.2." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.537, + 0.285, + 0.552 + ], + "angle": 0, + "content": "4.3.Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.56, + 0.483, + 0.864 + ], + "angle": 0, + "content": "Table 1 presents a performance and latency comparison between FROSS and various representative baseline methodologies. The results reveal that FROSS achieves the highest performance among all baseline methods with much lower processing latency. This substantiates the advantages of lifting scene graphs from 2D images over direct point cloud reasoning [31, 34, 35], as point clouds can sometimes present challenges with occlusion, sparsity, and reconstruction quality. The superior performance compared to the other image-based methods [9, 16, 36] can be attributed to our Gaussian-based merging strategy. In contrast, IMP [36] employs a voting mechanism for object classification, while VGfM [9] uses a basic recurrent neural network, which may restrict their object and relationship classification accuracy. Kim's baseline [16] achieves similar object recall to FROSS due to the same 2D SG generation pipeline. However, its merging mechanism fails to suppress duplicate detections, which hinders relationship aggregation and leads to significantly lower relationship and predicate recall. These results collectively substantiate the effectiveness of FROSS" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.09, + 0.905, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.259, + 0.907, + 0.316 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison between FROSS and Wu [35] on the 3DSSG dataset. Only representative objects are visualized. Errors are marked in red, with ground truth label shown in parentheses. Node colors correspond to the respective scene graphs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.34, + 0.907, + 0.372 + ], + "angle": 0, + "content": "in combining 2D information with Gaussian merging for 3D SSG generation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.38, + 0.697, + 0.396 + ], + "angle": 0, + "content": "4.4. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.402, + 0.909, + 0.675 + ], + "angle": 0, + "content": "Figure 4 presents the 3D SSG generation results from FROSS, demonstrating its ability to approximate object positions and shapes within a scene through Gaussian distributions, while simultaneously extracting their relationships. The unified objects highlight the effectiveness of the proposed merging algorithm, which consolidates predictions from multiple views, reducing redundant objects and processing time. This representation method offers a high-level scene understanding for downstream applications. In comparison, Wu's method [35] frequently misclassifies objects with similar geometry (e.g., \"curtain\" vs. \"door\", \"table\" vs. \"sofa\") and confuses semantically close relationships (e.g., \"standing on\" vs. \"supported by\"). This performance gap shows FROSS's advantage in utilizing richer visual information, in contrast to Wu [35], which relies solely on cropped object image for visual feature extraction. Additional qualitative results from FROSS are provided in the supplementary materials (Section 7.2)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.683, + 0.688, + 0.699 + ], + "angle": 0, + "content": "4.5. Runtime Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Table 2 examines the latency and frames-per-second (FPS) performance of FROSS, evaluated on an AMD Ryzen 9 7950X processor with an NVIDIA GeForce RTX 4090 GPU. This analysis serves to validate the computational efficiency of FROSS in accordance with its design objective of achieving faster-than-real-time processing capabilities. The evaluation restricts processing to a single CPU core and one GPU to assess system efficiency under limited computational resources. The metrics represent averages across 14,400 frames from four ReplicaSSG test scenes. The RT-DETR object detector and the EGTR relationship extraction module exhibit the highest individual latency. On the other hand, the 3D SSG merging algorithm contributes" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.483, + 0.903 + ], + "angle": 0, + "content": "2 Although Wu's method [35] supports RGB inputs, the reported results are based on predictions using dense reconstruction from RGB-D images." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28824" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.482, + 0.16 + ], + "angle": 0, + "content": "Table 2. Runtime analysis of the key components of FROSS. Obj. Det. refers to the object detection model RT-DETR [21, 44], while Rel. Ext. corresponds to the EGTR relationship extractor [12]. Merging 3D SSGs represents the process described in Section 3.4. Only components with significant latency impact are included." + }, + { + "type": "table", + "bbox": [ + 0.127, + 0.171, + 0.446, + 0.218 + ], + "angle": 0, + "content": "
Mean Latency (in milliseconds)
Obj. Det.Rel. Ext.Merging 3D SSGsFPS
2.314.510.12144.09
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.237, + 0.484, + 0.314 + ], + "angle": 0, + "content": "minimal computational overhead. Compared to the previous methods for online real-time 3D SSG generation methods [16, 34, 35], FROSS demonstrates significantly reduced end-to-end latency and increased FPS, which establishes a substantial improvement in terms of speed and efficiency." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.322, + 0.245, + 0.339 + ], + "angle": 0, + "content": "4.6. Ablation Study" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.344, + 0.406, + 0.359 + ], + "angle": 0, + "content": "4.6.1. Using Ground Truth 2D Scene Graphs" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.362, + 0.484, + 0.665 + ], + "angle": 0, + "content": "In the ablation analysis presented in Table 3, we investigate the influence of 2D SG quality on FROSS performance, given that FROSS relies heavily on accurate 2D SG predictions to generate 3D SSGs. The results indicate that the accuracy of the 2D SG has a significant impact on the final 3D SSG quality. In both the 3DSSG and ReplicaSSG datasets, FROSS achieves substantial gains in recall when ground truth 2D SG is used instead of predicted 2D SG. This performance difference is particularly pronounced in the ReplicaSSG dataset, where the baseline FROSS method achieves a relatively low recall score compared to the ground truth-assisted version. The discrepancy stems from differences between ReplicaSSG and Visual Genome [18]. Since the 2D SG model was originally trained on Visual Genome, the model may not generalize effectively to datasets with different characteristics. This leads to lower recall in both object and relationship detection within the ReplicaSSG dataset. This finding suggests the untapped potential for FROSS via domain-adaptive 2D SG models, with current results merely representing a lower bound on the framework's capabilities." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.672, + 0.391, + 0.687 + ], + "angle": 0, + "content": "4.6.2. Using Estimated Camera Trajectory" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.737 + ], + "angle": 0, + "content": "We further validate FROSS's effectiveness using estimated camera trajectories instead of ground truth poses to assess real-world applicability. This evaluation is crucial as" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.758, + 0.484, + 0.801 + ], + "angle": 0, + "content": "Table 3. Analysis of FROSS with predicted and ground truth 2D SGs as input on both the 3DSSG and ReplicaSSG datasets. 'w/ GT' denotes the version of FROSS utilizing ground truth 2D SGs." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.812, + 0.48, + 0.898 + ], + "angle": 0, + "content": "
DatasetMethodRecall (%)mRecall (%)
Rel.Obj.Pred.Obj.Pred.
3DSSG [31]FROSS27.962.433.063.818.0
FROSS w/ GT55.888.656.093.356.8
ReplicaSSGFROSS22.326.127.828.820.4
FROSS w/ GT67.689.167.692.682.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.905, + 0.147 + ], + "angle": 0, + "content": "Table 4. Performance comparison of FROSS with the predicted and the ground truth camera trajectories as input on the ReplicaSSG dataset. Please note that 'w/ SLAM' denotes the version of FROSS utilizing trajectories estimated by ORB-SLAM3 [2]." + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.157, + 0.901, + 0.221 + ], + "angle": 0, + "content": "
MethodRecall (%)mRecall (%)
Rel.Obj.Pred.Obj.Pred.
FROSS22.326.127.828.820.4
FROSS w/ SLAM22.725.827.227.720.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.235, + 0.905, + 0.263 + ], + "angle": 0, + "content": "Table 5. Recall on the validation split of the 3DSSG dataset with different Hellinger distance thresholds." + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.275, + 0.901, + 0.332 + ], + "angle": 0, + "content": "
Threshold δd0.60.650.70.750.80.850.9
Recall (%)Rel.15.116.520.022.725.930.029.3
Obj.69.569.869.168.965.761.955.4
Pred.16.217.921.824.928.833.634.3
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.352, + 0.907, + 0.518 + ], + "angle": 0, + "content": "most practical deployments must rely on estimated camera poses rather than perfect ground truth information. Table 4 presents the impact of using estimated trajectories generated by ORB-SLAM3 [2] on the ReplicaSSG dataset. The evaluation results demonstrate a slight improvement in relationship recall, while object and predicate recall exhibit minor decreases. This indicates that FROSS remains robust to inaccuracies in camera trajectory estimation. Note that a more detailed analysis of ORB-SLAM3's trajectory estimation errors on the ReplicaSSG dataset is provided for reference in the supplementary materials (Section 7.4)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.525, + 0.828, + 0.54 + ], + "angle": 0, + "content": "4.6.3. Effect of Hellinger Distance Threshold" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.544, + 0.906, + 0.635 + ], + "angle": 0, + "content": "Table 5 presents the effect of varying the Hellinger distance threshold \\(\\delta_{\\mathrm{d}}\\). A lower threshold (e.g., 0.65) enforces stricter merging criteria, which preserves object distinctions and leads to higher object recall. In contrast, a higher threshold (e.g., 0.85) enables more extensive relationship aggregation between objects, which improves relationship performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.648, + 0.634, + 0.664 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.907, + 0.9 + ], + "angle": 0, + "content": "This study presented FROSS, an innovative framework for real-time generation of 3D SSG in online environments. By representing objects as 3D Gaussian distributions, FROSS achieved computational efficiency that exceeded real-time requirements. The framework integrated a real-time object detection and relationship extraction pipeline to generate 2D SGs, which underwent 3D back-projection and incremental integration into a global SSG through our proposed merging algorithm. Experimental evaluations on existing benchmark and our newly proposed dataset, ReplicaSSG, demonstrated that FROSS surpassed existing methods in both performance metrics and computational speed by significant margins. These advancements established possibilities for real-time SSG applications in open-world environments, particularly relevant to robotics and AR domains." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28825" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.486, + 0.328 + ], + "angle": 0, + "content": "This work was partly supported by the Japan Society for the Promotion of Science KAKENHI Grant Numbers JP24H00733. The authors gratefully acknowledge the support from the National Science and Technology Council (NSTC) in Taiwan under grant numbers NSTC 114-2221-E-002-069-MY3, NSTC 113-2221-E-002-212-MY3, NSTC 114-2218-E-A49-026, and NSTC 114-2640-E-002-006. The authors would also like to express their appreciation for the donation of the GPUs from NVIDIA Corporation and NVIDIA AI Technology Center (NVAITC) used in this work. Furthermore, the authors extend their gratitude to the National Center for High-Performance Computing for providing the necessary computational and storage resources." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.344, + 0.188, + 0.36 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.369, + 0.484, + 0.438 + ], + "angle": 0, + "content": "[1] Iro Armeni, Zhi-Yang He, JunYoung Gwak, Amir R Zamir, Martin Fischer, Jitendra Malik, and Silvio Savarese. 3D scene graph: A structure for unified semantics, 3D space, and camera. In Int. Conf. Comput. Vis., pages 5664-5673, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.441, + 0.484, + 0.51 + ], + "angle": 0, + "content": "[2] Carlos Campos, Richard Elvira, Juan J Gómez Rodríguez, José MM Montiel, and Juan D Tardós. ORB-SLAM3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Trans. Robotics, 37(6):1874-1890, 2021. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.513, + 0.484, + 0.554 + ], + "angle": 0, + "content": "[3] Ata Çelen, Guo Han, Konrad Schindler, Luc Van Gool, Iro Armeni, Anton Obukhov, and Xi Wang. I-Design: Personalized llm interior designer. arXiv:2404.02838, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.556, + 0.482, + 0.598 + ], + "angle": 0, + "content": "[4] David F Crouse, Peter Willett, Krishna Pattipati, and Lennart Svensson. A look at gaussian mixture reduction algorithms. In Int. Conf. Inform. Fusion, pages 1-8. IEEE, 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.6, + 0.484, + 0.669 + ], + "angle": 0, + "content": "[5] Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and William Yang Wang. LayoutGPT: Compositional visual planning and generation with large language models. Adv. Neural Inform. Process. Syst., 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.671, + 0.482, + 0.712 + ], + "angle": 0, + "content": "[6] Maria Fox and Derek Long. PDDL2.1: An extension to PDDL for expressing temporal planning domains. J. Artificial Intell. Research, 20:61-124, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.715, + 0.484, + 0.77 + ], + "angle": 0, + "content": "[7] Gege Gao, Weiyang Liu, Anpei Chen, Andreas Geiger, and Bernhard Schölkopf. GraphDreamer: Compositional 3D scene synthesis from scene graphs. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.772, + 0.484, + 0.814 + ], + "angle": 0, + "content": "[8] Lizhao Gao, Bo Wang, and Wenmin Wang. Image captioning with scene-graph based semantic concepts. In Int. Conf. on Mach. Learn. Comput., pages 225-229, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.816, + 0.484, + 0.87 + ], + "angle": 0, + "content": "[9] Paul Gay, James Stuart, and Alessio Del Bue. Visual graphs from motion (VGfM): Scene understanding with object geometry reasoning. In ACCV, pages 330-346. Springer, 2019. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.485, + 0.901 + ], + "angle": 0, + "content": "[10] Jörg Hoffmann. FF: The fast-forward planning system. AI magazine, 22(3):57-57, 2001. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.369, + 0.485, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[11] Nathan Hughes, Yun Chang, and Luca Carlone. Hydra: A real-time spatial perception system for 3D scene graph construction and optimization. In Robotics: Sci. Syst., 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[12] Jinbae Im, JeongYeon Nam, Nokyung Park, Hyungmin Lee, and Seunghyun Park. EGTR: Extracting graph from transformer for scene graph generation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 24229-24238, 2024. 2, 3, 4, 6, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[13] Justin Johnson, Ranjay Krishna, Michael Stark, Li-Jia Li, David Shamma, Michael Bernstein, and Li Fei-Fei. Image retrieval using scene graphs. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3668-3678, 2015. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[14] Justin Johnson, Agrim Gupta, and Li Fei-Fei. Image generation from scene graphs. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1219–1228, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.361 + ], + "angle": 0, + "content": "[15] Dong-Jin Kim, Jinsoo Choi, Tae-Hyun Oh, and In So Kweon. Dense relational captioning: Triple-stream networks for relationship-based captioning. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6271-6280, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[16] Ue-Hwan Kim, Jin-Man Park, Taek-Jin Song, and Jong-Hwan Kim. 3-D scene graph: A sparse and semantic representation of physical environments for intelligent agents. IEEE Trans. Cybernetics, 50(12):4921-4933, 2019. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.474 + ], + "angle": 0, + "content": "[17] Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In Int. Conf. Learn. Represent., 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.477, + 0.905, + 0.558 + ], + "angle": 0, + "content": "[18] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual Genome: Connecting language and vision using crowdsourced dense image annotations. Int. J. Comput. Vis., 123:32-73, 2017. 2, 6, 8, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.904, + 0.602 + ], + "angle": 0, + "content": "[19] Linjie Li, Zhe Gan, Yu Cheng, and Jingjing Liu. Relation-aware graph attention network for visual question answering. In Int. Conf. Comput. Vis., pages 10313-10322, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.604, + 0.905, + 0.658 + ], + "angle": 0, + "content": "[20] Yuanyuan Liu, Chengjiang Long, Zhaoxuan Zhang, Bokai Liu, Qiang Zhang, Baocai Yin, and Xin Yang. Explore contextual information for 3D scene graph generation. IEEE Trans. Vis. Comput. Graph., 29(12):5556-5568, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[21] Wenyu Lv, Yian Zhao, Qinyao Chang, Kui Huang, Guanzhong Wang, and Yi Liu. RT-DETRv2: Improved baseline with bag-of-freebies for real-time detection transformer. arXiv:2407.17140, 2024. 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.717, + 0.905, + 0.772 + ], + "angle": 0, + "content": "[22] Jeffri Murragarra-Llerena, Lucas N. Kirsten, Luis Felipe Zeni, and Claudio R. Jung. Probabilistic intersection-over-union for training and evaluation of oriented object detectors. IEEE Trans. Image Process., 33:671-681, 2024. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[23] Pushmeet Kohli Nathan Silberman, Derek Hoiem and Rob Fergus. Indoor segmentation and support inference from RGBD images. In Eur. Conf. Comput. Vis., 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[24] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. PointNet: Deep learning on point sets for 3D classification and segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 652-660, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Tianwen Qian, Jingjing Chen, Shaoxiang Chen, Bo Wu, and Yu-Gang Jiang. Scene graph refinement network for vi" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28826" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "sual question answering. IEEE Trans. Multimedia, 25:3950-3961, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[26] Antoni Rosinol, Andrew Violette, Marcus Abate, Nathan Hughes, Yun Chang, Jingnan Shi, Arjun Gupta, and Luca Carlone. Kimera: From SLAM to spatial perception with 3D dynamic scene graphs. Int. J. Robotics Research, 40(12-14):1510-1546, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.191, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[27] Dhruv Shah, Benjamin Eysenbach, Gregory Kahn, Nicholas Rhinehart, and Sergey Levine. ViNG: Learning Open-World Navigation with Visual Goals. In IEEE Int. Conf. Robotics Auto., 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.482, + 0.384 + ], + "angle": 0, + "content": "[28] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke M. Strasdat, Renzo De Nardi, Michael Goesele, Steven Lovegrove, and Richard Newcombe. The Replica dataset: A digital replica of indoor spaces. arXiv:1906.05797, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.385, + 0.482, + 0.439 + ], + "angle": 0, + "content": "[29] Sinan Tan, Mengmeng Ge, Di Guo, Huaping Liu, and Fuchun Sun. Knowledge-based embodied question answering. IEEE Trans. Pattern Anal. Mach. Intell., 45(10):11948-11960, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.441, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[30] Johanna Wald, Armen Avetisyan, Nassir Navab, Federico Tombari, and Matthias Niessner. RIO: 3D object instance relocalization in changing indoor environments. In Int. Conf. Comput. Vis., 2019. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[31] Johanna Wald, Helisa Dhamo, Nassir Navab, and Federico Tombari. Learning 3D semantic scene graphs from 3D indoor reconstructions. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3961-3970, 2020. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[32] Johanna Wald, Nassir Navab, and Federico Tombari. Learning 3D semantic scene graphs with instance embeddings. Int. J. Comput. Vis., 130(3):630-651, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[33] Ziqin Wang, Bowen Cheng, Lichen Zhao, Dong Xu, Yang Tang, and Lu Sheng. VL-SAT: Visual-linguistic semantics assisted training for 3D semantic scene graph prediction in point cloud. In IEEE Conf. Comput. Vis. Pattern Recog., pages 21560–21569, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.482, + 0.733 + ], + "angle": 0, + "content": "[34] Shun-Cheng Wu, Johanna Wald, Keisuke Tateno, Nassir Navab, and Federico Tombari. SceneGraphFusion: Incremental 3D scene graph prediction from rgb-d sequences. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7515-7525, 2021. 2, 3, 6, 7, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.734, + 0.482, + 0.789 + ], + "angle": 0, + "content": "[35] Shun-Cheng Wu, Keisuke Tateno, Nassir Navab, and Federico Tombari. Incremental 3D semantic scene graph prediction from RGB sequences. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5064-5074, 2023. 2, 3, 6, 7, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[36] Danfei Xu, Yuke Zhu, Christopher B Choy, and Li Fei-Fei. Scene graph generation by iterative message passing. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5410-5419, 2017. 2, 6, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[37] Xu Yang, Kaihua Tang, Hanwang Zhang, and Jianfei Cai. Auto-encoding scene graphs for image captioning. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10685-10694, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[38] Rowan Zellers, Mark Yatskar, Sam Thomson, and Yejin Choi. Neural Motifs: Scene graph parsing with global context. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5831-5840, 2018. 2, 6, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[39] Guangyao Zhai, Evin Pinar Örnek, Shun-Cheng Wu, Yan Di, Federico Tombari, Nassir Navab, and Benjamin Busam. Commonsscenes: Generating commonsense 3d indoor scenes with scene graph diffusion. In Adv. Neural Inform. Process. Syst., 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[40] Guangyao Zhai, Evin Pinar Örnek, Dave Zhenyu Chen, Ruotong Liao, Yan Di, Nassir Navab, Federico Tombari, and Benjamin Busam. Echoscene: Indoor scene generation via information echo over scene graph diffusion. In Eur. Conf. Comput. Vis., pages 167-184. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.29, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[41] Chaoyi Zhang, Jianhui Yu, Yang Song, and Weidong Cai. Exploiting edge-oriented reasoning for 3D point-based scene graph analysis. In IEEE Conf. Comput. Vis. Pattern Recog., pages 9705-9715, 2021. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.347, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[42] Chaoyi Zhang, Xitong Yang, Ji Hou, Kris Kitani, Weidong Cai, and Fu-Jen Chu. EgoSG: Learning 3D scene graphs from egocentric RGB-D sequences. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2535-2545, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.403, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[43] Shoulong Zhang, Aimin Hao, Hong Qin, et al. Knowledge-inspired 3D scene graph prediction in point cloud. Adv. Neural Inform. Process. Syst., 34:18620-18632, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.446, + 0.905, + 0.515 + ], + "angle": 0, + "content": "[44] Yian Zhao, Wenyu Lv, Shangliang Xu, Jinman Wei, Guanzhong Wang, Qingqing Dang, Yi Liu, and Jie Chen. DETRs beat YOLOs on real-time object detection. In IEEE Conf. Comput. Vis. Pattern Recog., pages 16965-16974, 2024. 3, 4, 6, 8, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.517, + 0.905, + 0.584 + ], + "angle": 0, + "content": "[45] Kaizhi Zheng, Xiaotong Chen, Xuehai He, Jing Gu, Linjie Li, Zhengyuan Yang, Kevin Lin, Jianfeng Wang, Lijuan Wang, and Xin Eric Wang. EditRoom: LLM-parameterized graph diffusion for composable 3D room layout editing. arXiv:2410.12836, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.586, + 0.905, + 0.629 + ], + "angle": 0, + "content": "[46] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. EWA splatting. IEEE Trans. Vis. Comput. Graph., 8(3):223-238, 2002. 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.629 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28827" + } + ] +] \ No newline at end of file diff --git a/2025/FROSS_ Faster-Than-Real-Time Online 3D Semantic Scene Graph Generation from RGB-D Images/f4b25ed8-2175-4e29-a597-e457f4795510_origin.pdf b/2025/FROSS_ Faster-Than-Real-Time Online 3D Semantic Scene Graph Generation from RGB-D Images/f4b25ed8-2175-4e29-a597-e457f4795510_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a5df32a754bf9fe4d31850009553266a9a45a72a --- /dev/null +++ b/2025/FROSS_ Faster-Than-Real-Time Online 3D Semantic Scene Graph Generation from RGB-D Images/f4b25ed8-2175-4e29-a597-e457f4795510_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5a1f6e9ffbffb224c12d40a6549f6a06e14766bb75785d2b8d64a8071cf9e334 +size 14883340 diff --git a/2025/FROSS_ Faster-Than-Real-Time Online 3D Semantic Scene Graph Generation from RGB-D Images/full.md b/2025/FROSS_ Faster-Than-Real-Time Online 3D Semantic Scene Graph Generation from RGB-D Images/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c1813df92dbc38f34fd02b4bfe276c63b590dff7 --- /dev/null +++ b/2025/FROSS_ Faster-Than-Real-Time Online 3D Semantic Scene Graph Generation from RGB-D Images/full.md @@ -0,0 +1,299 @@ +# FROSS: Faster-than-Real-Time Online 3D Semantic Scene Graph Generation from RGB-D Images + +Hao-Yu Hou1,3 + +Chun-Yi Lee2 + +Motoharu Sonogashira3 + +Yasutomo Kawanishi3 + +$^{1}$ National Tsing Hua University + +$^{2}$ National Taiwan University + +$^{3}$ RIKEN + +# Abstract + +The ability to abstract complex 3D environments into simplified and structured representations is crucial across various domains. 3D semantic scene graphs (SSGs) achieve this by representing objects as nodes and their interrelationships as edges, facilitating high-level scene understanding. Existing methods for 3D SSG generation, however, face significant challenges, including high computational demands and non-incremental processing that hinder their suitability for real-time open-world applications. To address this issue, we propose FROSS (Faster-than-Real-Time Online 3D Semantic Scene Graph Generation), an innovative approach for online and faster-than-realtime 3D SSG generation that leverages the direct lifting of 2D scene graphs to 3D space and represents objects as 3D Gaussian distributions. This framework eliminates the dependency on precise and computationally-intensive point cloud processing. Furthermore, we extend the Replica dataset with inter-object relationship annotations, creating the ReplicaSSG dataset for comprehensive evaluation of FROSS. The experimental results from evaluations on ReplicaSSG and 3DSSG datasets show that FROSS can achieve superior performance while operating significantly faster than prior 3D SSG generation methods. Our implementation and dataset are publicly available at https://github.com/Howardkhh/FROSS. + +# 1. Introduction + +The abstraction of complex 3D environments into simplified representations in real time serves as a crucial capability across multiple application domains. The incremental generation of such representations in real time holds particular significance for fields ranging from robotics and computer vision to augmented reality (AR), where systems must perceive and analyze the environment instantaneously. 3D scene graphs (SGs) [1, 9] accomplish this abstraction by representing physical objects as nodes and their interrelationships as edges within a structured framework. 3D se + +![](images/3ff1488ecdb0ccc9ab8fe0361855f9b2bdd519b3a410a1a4ce91d248d387df87.jpg) +Figure 1. We introduce FROSS, an online real-time 3D semantic scene graph generation method that leverages and integrates 2D scene graphs. FROSS represents objects as 3D Gaussian distributions and operates without requiring 3D reconstruction. + +mantic scene graphs (SSGs) [31] extend this representation with an emphasis on semantic relationships, such as positional configurations between objects (e.g., a cup on a table) or functional interactions (e.g., a person seated in a chair), which enables high-level scene comprehension, reasoning, and interaction. Real-world applications, however, present open-world challenges where environments often exceed known spatial boundaries and contain previously unseen spaces [27]. For instance, a room may connect to multiple adjacent spaces or corridors, which necessitates continuous + +adaptation and incremental scene understanding. This construction process must maintain high efficiency to generate structured environmental representations without creating performance bottlenecks. Meeting the processing demands requires the generation of 3D SSG to operate at faster-than-real-time speeds, with all computations completed between successive image frames from a high-speed sensor stream. + +Executing all processes at faster-than-real-time rates confers multiple advantages: (1) prevention of bottlenecks in the 3D SSG generation framework, (2) facilitation of smooth operation on resource-constrained hardware such as edge computing devices, and (3) preservation of computational resources for additional system components. Contemporary 3D SSG generation approaches predominantly focus on offline construction, which requires access to the complete scene data, either instance-segmented point clouds [31, 41], a complete image sequence [9], or both [33]. This assumption of full scene availability conflicts with incremental environmental exploration. Although several studies [34, 35] have developed methods for incremental point cloud reconstruction and segmentation to achieve online and real-time 3D SSG generation, their aggregate system latency remains substantial, even with the utilization of multiple CPU cores. This constraint stems mainly from the substantial computational demands of environmental mapping and point cloud processing, restricting their practical application in real-time systems with limited computational power or other concurrent processes. Given these limitations, the aim of this study is to develop a method for faster-than-real-time online SSG generation. + +A fundamental insight for generating 3D SSGs emerges from the observation that precise object pose and shape information is not essential. Given that 3D SSGs provide a high-level semantic understanding of the environment, their functional implications remain valid with approximate object locations. For instance, in robotic applications, high-level planning can be successfully executed using relative spatial relationships [6, 10, 29]. Similarly, in 3D scene synthesis and room layout generation [3, 5, 7, 39, 40, 45], 3D SSGs serve as foundational structures from which more detailed content can be generated. These application scenarios eliminate the need for precise environmental mapping techniques such as Simultaneous Localization and Mapping (SLAM) [2] and additional point cloud processing, both of which contribute significant computational overhead. + +This observation motivates an alternative approach for online and real-time 3D SSG generation: inferring 2D SGs from images and subsequently lifting them into 3D space, thus circumventing the computational burden of environmental mapping and point cloud processing. To the best of our knowledge, the current literature has not adequately investigated this research direction except [16]. Such unexplored territory presents significant challenges, includ + +ing the handling of incomplete 2D object observations, the achievement of real-time SG generation, and the effective integration of multi-view information into complete 3D SSGs without object redundancy. These challenges, therefore, provide promising avenues for further innovative research contributions. + +In light of these considerations, we introduce FROSS (Faster-than-Real-time Online 3D Semantic Scene Graph Generation), a methodology that constructs 3D SSGs by lifting 2D SGs to 3D space and integrating them into a comprehensive representation. To achieve real-time processing without requiring precise object localization, FROSS initiates the process by constructing 2D SGs from detected objects in images. A key innovation of FROSS lies in its utilization of Gaussian distributions to model objects in 3D space, which facilitates the integration of objects observed across multiple images into a coherent 3D SSG, as illustrated in Figure 1. Specifically, FROSS represents detected objects through Gaussian distributions, lifts these distributions from 2D SGs into 3D space, and merges them according to their predicted classes and statistical distances. We validate FROSS on the 3DSSG dataset [31] and conduct additional evaluations on ReplicaSSG, an extended version of the Replica dataset [28] that we have enhanced with object relationship annotations. ReplicaSSG incorporates object classes defined in Visual Genome [18], a widely-adopted 2D SG dataset. The main contributions of the paper can be summarized as follows: + +- We introduce FROSS, an innovative methodology for online real-time generation of 3D SSGs. FROSS demonstrates superior performance and significantly faster processing speeds compared to existing baseline methods. +- We propose a new merging algorithm based on Gaussian distributions for 3D object integration. This algorithm effectively prevents the proliferation of duplicate or incorrectly detected objects as the input image sequence expands, thereby maintaining computational efficiency. +- We develop ReplicaSSG, an extension of the Replica dataset with 3D SSG annotations. Comprehensive evaluations of FROSS on both the 3DSSG and ReplicaSSG datasets demonstrate FROSS's effectiveness and computational efficiency over diverse environmental scenarios. + +# 2. Related Work + +SGs [12, 13, 36, 38] constitute a fundamental data structure for object and relationship representation within images following their introduction by [13]. This representation has demonstrated substantial utility across various scene understanding tasks [8, 14, 15, 19, 25, 37]. The concept later evolved to encompass 3D spaces, where objects may incorporate attributes such as class labels, visual features, shape characteristics, and affordances. Research in 3D SGs developed into two branches of representation: SSGs intro + +duced in [9, 16, 31] and hierarchical scene graphs (HSGs) proposed in [1, 11, 26]. Both frameworks represent 3D objects as nodes with pairwise relationships defined through edges. However, these approaches differ fundamentally in their relationship definitions. SSGs concentrate on semantic relationships between objects, while HSGs organize objects into hierarchical layers that include buildings, rooms, and constituent objects, where edges connect only nodes in identical or adjacent layers. Compared to HSGs, SSGs enable higher-level reasoning capabilities through rich semantic relationship representations. The extension beyond spatial relationships allows SSGs to be applied to broader domains such as robotics, 3D computer vision, and AR. + +# 2.1. Offline 3D Semantic Scene Graph Generation + +The initial development of 3D SSG generation focused primarily on offline settings that required either complete instance-segmented point clouds [20, 31-33, 41, 43] or image sequences [9, 33] as inputs. Such requirements limited their applicability to scenarios with incremental data collection, as they necessitated complete data reprocessing. Despite these limitations, these studies established foundations for 3D SSG generation. The research in [9] pioneered 3D SSG generation by predicting 3D object ellipsoids from image collections and utilizing their geometry and appearance features through a recurrent neural network to extract inter-object relationships. The study in [31] advanced this field through the introduction of the 3DSSG dataset, which enhanced the 3RScan dataset [30] with SSG annotations. This dataset represents the first large-scale collection with dense annotations and has become the primary benchmark for evaluating 3D SSG generation methods. The authors of [31] further contributed a generation methodology by utilizing instance-segmented point clouds as input and leveraged PointNet [24] for feature extraction as well as a graph convolutional network [17] for label prediction. + +# 2.2. Online 3D Semantic Scene Graph Generation + +Several previous studies have explored 3D SG generation in online settings, with some achieving real-time performance. The research in [16] introduced a framework that predicts local 3D SSGs from RGB-D images and integrates them incrementally into a global representation. However, their computational requirements render real-time execution infeasible. SceneGraphFusion [34] and the study in [35] developed methods for online point cloud reconstruction and segmentation, followed by feature extraction for local SSG prediction and global fusion. These frameworks can achieve real-time performance through multi-threaded execution of parallel components. In contrast to these existing methods, our proposed FROSS framework avoids precise object localization by approximating objects with Gaussian distributions. This approximation approach enables faster-than + +real-time online 3D SSG generation, allowing the end-to-end processing latency to be below the sensor data acquisition period while maintaining single-threaded execution. + +# 3. Methodology + +# 3.1. Problem Definition + +The primary problem concerned in this study is the online generation of 3D SSGs for environments where the complete scene structure remains unknown a priori. Given a sequence of input images, the primary objective is to construct a 3D SSG $\mathcal{G}$ of the target environment. The graph $\mathcal{G}$ consists of a set of nodes $\nu$ and their corresponding directed edges $\mathcal{E}$ . Each node $v_{i} = (c_{i},\mu_{i}^{3\mathrm{D}},\Sigma_{i}^{3\mathrm{D}})\in \mathcal{V}$ is indexed by $i\in \{1,\ldots ,N\}$ , representing a unique object in the scene. Every object has a class label $c_{i}\in \mathcal{C}$ , where $\mathcal{C}$ is the set of object categories. While different approaches may incorporate various node attributes, our formulation represents each object as a 3D Gaussian distribution parameterized by mean vector $\mu_i^{3\mathrm{D}}\in \mathbb{R}^3$ and covariance matrix $\Sigma_i^{3\mathrm{D}}\in \mathbb{R}^{3\times 3}$ . The edges $e_{i\rightarrow j}\in \mathcal{E}$ connect node $v_{i}$ to node $v_{j}$ with relationship $r_{i\rightarrow j}\in \mathcal{R}$ , where $\mathcal{R}$ represents the set of predefined relationship types. Furthermore, the algorithm must be able to maintain online processing capabilities to update $\mathcal{G}$ continuously whenever new image data becomes available. + +# 3.2. Overview of Framework + +Figure 2 presents an overview of the proposed FROSS framework, which processes image sequences with corresponding depth information and camera transformations as input. The framework begins with the EGTR [12] 2D SG generation model, with the object detection backbone model replaced with RT-DETR [21, 44]. This RT-DETR object detector is a state-of-the-art real-time detection model, which preserves intermediate self-attention features for subsequent relationship extraction. FROSS then converts object bounding boxes to 2D Gaussian distributions and projects them into 3D space to generate local SSGs, as elaborated in Section 3.3. These Gaussian distributions serve as efficient approximations of both object positions and their spatial distributions for rapid processing. These local graphs are then integrated into a global SSG through the proposed merging algorithm that is described in detail in Section 3.4. + +# 3.3. Lifting 2D SG to 3D + +The 2D SGs generated from RT-DETR [21, 44] and EGTR [12] enable FROSS to construct local 3D SSGs $\mathcal{G}_{\mathrm{local}} = (\mathcal{V}_{\mathrm{local}},\mathcal{E}_{\mathrm{local}})$ through the integration of depth information and camera transformation. This process involves two key steps: (1) the conversion of object bounding boxes to 2D Gaussians, and (2) their subsequent back-projection into 3D space while preserving inter-object relationships. + +![](images/c9c0137962ed863d7f4ef0b316022ea0a35d687b329881ea4b6ee1d68fbaf661.jpg) +Figure 2. An overview of the FROSS framework: (a) The process initiates with object detection via RT-DETR [44] from an RGB-D image and its associated camera pose. (b) The EGTR model [12] extracts inter-object relationships through utilization of preserved self-attention features from RT-DETR. (c) 2D Gaussian distributions derived from detected bounding boxes undergo spatial transformation into 3D space for local 3D SSG construction (Section 3.3). (d) The resultant local 3D SSG is subsequently incorporated into the global 3D SSG through our proposed merging algorithm (Section 3.4). + +2D Gaussian Representation of Bounding Boxes. Although several methods exist for lifting 2D bounding boxes to 3D space, including point cloud back-projection and joint optimization of 3D ellipsoids [9], the Gaussian distribution approach emerges as the most computationally efficient solution among these alternatives. The representation of objects via Gaussian distributions requires the specification of mean $\mu_i^{2\mathrm{D}}$ and covariance matrix $\Sigma_i^{2\mathrm{D}}$ of the 2D object node $v_i^{2\mathrm{D}}$ , which would correspond to a 3D node $v_i \in \mathcal{V}_{\mathrm{local}}$ . Based on the formulation in [22], FROSS models bounding boxes as 2D uniform distributions. The mean values correspond to the bounding box centers, while the covariance matrices maintain consistency with the uniform distributions. The covariance matrix of $v_i^{2\mathrm{D}}$ is formulated as: + +$$ +\Sigma_ {i} ^ {\mathrm {2 D}} = \frac {1}{1 2} \left[ \begin{array}{c c} W _ {i} ^ {2} & 0 \\ 0 & H _ {i} ^ {2} \end{array} \right], \tag {1} +$$ + +where $W_{i}$ and $H_{i}$ are 2D bounding box's width and height. 3D Back-Projection of Gaussian Distributions. Given the 2D Gaussians, their corresponding depths at the Gaussian centroids, and camera transformation matrices, the mean positions $\mu_{i}^{3\mathrm{D}}$ and covariances $\Sigma_{i}^{3\mathrm{D}}$ of the Gaussians in the 3D space can be approximated. The 3D mean $\mu_{i}^{3\mathrm{D}}$ can be derived from the 2D mean $\mu_{i}^{2\mathrm{D}} = (p_{x}, p_{y})^{\top}$ and its corresponding depth value $z$ through the following equation: + +$$ +\mu_ {i} ^ {\mathrm {3 D}} = R \cdot K ^ {- 1} \cdot \left(p _ {x}, p _ {y}, 1\right) ^ {\top} \cdot z + \mathbf {t}, \tag {2} +$$ + +where $K \in \mathbb{R}^{3 \times 3}$ represents the camera intrinsic matrix, and $R \in \mathbb{R}^{3 \times 3}$ and $\mathbf{t} \in \mathbb{R}^3$ denote the rotation matrix and translation vector of the current camera, respectively. The approximation of the 3D covariance matrix utilizes the Jacobian $J$ of the local affine approximation of the projective transformation [46], which is defined as the following: + +$$ +J = \left[ \begin{array}{c c c} \frac {f _ {x}}{z} & 0 & - \frac {f _ {x} p _ {x}}{z ^ {2}} \\ 0 & \frac {f _ {y}}{z} & - \frac {f _ {y} p _ {y}}{z ^ {2}} \end{array} \right], \tag {3} +$$ + +where $f_{x}$ and $f_{y}$ represent the focal lengths in $x$ and $y$ dimensions. In the original literature of [46], the Jacobian $J$ is utilized to project 3D covariance matrices onto the 2D image plane. However, our approach inverts this process by back-projecting the 2D covariance matrix $\Sigma_{i}^{2\mathrm{D}}$ into 3D space as $\bar{\Sigma}_{i}^{3\mathrm{D}}$ via the pseudo-inverse $J^{+}$ of the Jacobian: + +$$ +\bar {\Sigma} _ {i} ^ {\mathrm {3 D}} = R J ^ {+} \Sigma_ {i} ^ {\mathrm {2 D}} J ^ {+ \top} R ^ {\top}. \tag {4} +$$ + +The naive definition alone is insufficient since the original equation is inherently designed for projection from 3D to 2D space. Simply inverting the equation results in a 3D Gaussian that lacks variance information along the depth axis, as the 2D covariance matrix does not capture depth-related variance. This limitation necessitates the specification of variance in the depth dimension. We hypothesize that this variance approximates the average variance of the other dimensions. The specification has to be applied prior to the rotation matrix $R$ , leading to the following equations: + +$$ +\Sigma_ {i} ^ {\mathrm {3 D / /}} = J ^ {+} \Sigma_ {i} ^ {\mathrm {2 D}} J ^ {+ \top}, \tag {5} +$$ + +$$ +\Sigma_ {i} ^ {\mathrm {3 D} \prime} = \Sigma_ {i} ^ {\mathrm {3 D} \prime \prime} + \left[ \begin{array}{l l l} 0 & 0 & 0 \\ 0 & 0 & 0 \\ 0 & 0 & \frac {(\Sigma_ {i} ^ {\mathrm {3 D} \prime \prime}) _ {1 , 1} + (\Sigma_ {i} ^ {\mathrm {3 D} \prime \prime}) _ {2 , 2}}{2} \end{array} \right]. \tag {6} +$$ + +In this formulation, $(\Sigma_{i}^{3\mathrm{D} / \prime})_{m,n}\in \mathbb{R}$ represents the element at position $(m,n)$ of matrix $\Sigma_i^{3\mathrm{D} / \prime}$ . Finally, the covariance matrix $\Sigma_{i}^{3\mathrm{D}^{\prime}}$ is transformed to world coordinates using the rotation matrix $R$ , resulting in the following expression: + +$$ +\Sigma_ {i} ^ {\mathrm {3 D}} = R \Sigma_ {i} ^ {\mathrm {3 D} \prime} R ^ {\top}. \tag {7} +$$ + +# 3.4. Merging 3D SSGs + +With a local 3D SSG $\mathcal{G}_{\mathrm{local}} = (\mathcal{V}_{\mathrm{local}},\mathcal{E}_{\mathrm{local}})$ constructed, it is then integrated with the global 3D SSG $\mathcal{G}_{\mathrm{global}}$ through Gaussian merging and relationship accumulation. Figure 3 + +illustrates the proposed merging process. Specifically, it initializes a queue with nodes from $\nu_{\mathrm{local}}$ and evaluates potential merging opportunities through distance calculations. For each node $v_{i}$ in this queue, FROSS calculates the Hellinger distances [22] between the Gaussian distributions of $v_{i}$ and all other nodes $v_{j}$ of the same semantic class, in both the local node queue and $\nu_{\mathrm{global}}$ , to identify object correspondence. The algorithm merges nodes with Hellinger distances below the threshold $\delta_{\mathrm{d}}$ and adds unmerged nodes to $\nu_{\mathrm{global}}$ as new objects. The Hellinger distance $H_{D}(i,j)$ between Gaussian distributions $\mathcal{N}(\mu_i,\Sigma_i)$ and $\mathcal{N}(\mu_j,\Sigma_j)$ is quantified through the following equation, wherein $B_{D}(i,j)$ represents the Bhattacharyya distance: + +$$ +H _ {D} (i, j) = \sqrt {1 - \exp (- B _ {D} (i , j))}, \tag {8} +$$ + +$$ +B _ {D} (i, j) = \frac {1}{8} \Delta \mu_ {i j} ^ {T} \Sigma^ {- 1} \Delta \mu_ {i j} + \frac {1}{2} \ln \left(\frac {\operatorname* {d e t} \Sigma}{\sqrt {\operatorname* {d e t} \Sigma_ {i} \operatorname* {d e t} \Sigma_ {j}}}\right), \tag {9} +$$ + +where $\Delta \mu_{ij} = \mu_i - \mu_j$ and $\Sigma = \frac{1}{2} (\Sigma_i + \Sigma_j)$ . The fusion of these nodes occurs through a weighted integration of their respective Gaussian distributions [4]. The weighting coefficients for this integration process reflect both spatial coverage and detection confidence across multiple viewpoints. Objects detected from diverse angular perspectives and spatial positions receive proportionally higher weights, thus mitigating viewpoint bias and ensuring robust three-dimensional representation. Given that weights $w_i$ and $w_j$ represent the detection frequencies of their respective nodes, the mean vector $\mu_k$ and covariance matrix $\Sigma_k$ for the resultant merged node are determined via the formulations: + +$$ +\mu_ {k} = \frac {w _ {i} \mu_ {i} + w _ {j} \mu_ {j}}{w _ {i} + w _ {j}}, \tag {10} +$$ + +$$ +\Sigma_ {k} = \frac {w _ {i} \Sigma_ {i} + w _ {j} \Sigma_ {j}}{w _ {i} + w _ {j}} + \frac {w _ {i} w _ {j} \left(\mu_ {i} - \mu_ {j}\right) \left(\mu_ {i} - \mu_ {j}\right) ^ {\top}}{\left(w _ {i} + w _ {j}\right) ^ {2}}, \tag {11} +$$ + +where $\mu_{i},\mu_{j}$ are the mean vectors, and $\Sigma_{i},\Sigma_{j}$ are the covariance matrices of the respective Gaussian distributions to be merged. All relationship edges originally connected to $v_{i}$ and $v_{j}$ are then redirected to the newly merged node $v_{k}$ . The final relationship prediction between two global object nodes is selected through majority voting, choosing the most frequent prediction across viewpoints. + +# 4. Experimental Results + +The evaluation of FROSS encompasses two 3D SSG datasets: 3DSSG [31] and our developed ReplicaSSG datasets. Section 4.1 introduces the datasets, baseline SSG generation methods, and evaluation metrics. Section 4.2 presents the implementation details. Section 4.3 provides comprehensive comparisons between FROSS and existing SSG generation methods on the 3DSSG dataset. Moreover, the effectiveness of our approach is further demonstrated through qualitative results in Section 4.4. We further provide runtime analyses on the ReplicaSSG dataset in Section 4.5, along with additional ablation studies in Section 4.6. More quantitative and qualitative analyses are provided in the supplementary material. + +# 4.1. Evaluation Setup + +# 4.1.1. Datasets + +3DSSG. The 3DSSG dataset [31] extends 3RScan [30], which encompasses 1,482 scans of indoor environments with their corresponding RGB-D image sequences, 3D meshes, and dense instance segmentation. The dataset contains 363,555 images, with an average of 245 images per scan. 3DSSG augments the base dataset with object attributes, hierarchical category labels, and directed edges that describe inter-object semantic relationships such as 'standing on,' 'attached to,' and 'same color.' This dataset has established a primary benchmark for numerous + +![](images/32fdd0c43df1d873c6024bd66ff45d0cd00e1cad2bbe51d739b17ac8c4e08fbb.jpg) +Figure 3. Illustration of the object merging process using the proposed algorithm described in Section 3.4. (a) Starting with an input image, a set of objects with predicted categories and bounding boxes is generated. (b) These objects are then lifted into 3D spaces to form the object nodes in $\mathcal{V}_{\mathrm{local}}$ with colors matching their bounding boxes (yellow, cyan, and blue), while pre-existing objects in $\mathcal{V}_{\mathrm{global}}$ are shown in green. (c) The merging algorithm combines the objects linked by the arrows, resulting in the updated global objects, represented in red. + +studies [20, 31-35, 41-43]. Our experiments utilize the publicly available training and testing splits. To address the class imbalance in both object and relationship labels, this study adopts the commonly used category definitions in [31, 34, 35]. Specifically, the original 160 object categories map to 20 categories from NYUv2 [23], while seven of the 26 predicate categories remain. Following this categorical mapping, 3DSSG contains a total of 21,974 distinct objects and 16,324 inter-object relationships. This extensive collection of relationships offers a substantial foundation for spatial and semantic analyses. Nevertheless, the low image quality and low frame rate of 3DSSG hinders the comprehensive evaluation of FROSS at its full operational capacity. This necessitates our development of ReplicaSSG, which is an enhanced adaptation of the Replica dataset [28]. + +ReplicaSSG. ReplicaSSG extends the Replica dataset [28] with newly annotated object relationships. The original Replica provides high-quality reconstructions of indoor environments with instance-segmented meshes and photorealistic renderings, and is particularly suitable for SSG evaluation. Although Replica encompasses only 18 scenes, which precludes its use for training purposes, it serves as an effective evaluation benchmark. In the evaluation of FROSS, seven scenes serve as validation environments for hyperparameter optimization, while the remaining 11 scenes function as performance assessment platforms. Rather than preserving Replica's original object category definitions, ReplicaSSG adopts the classification system from Visual Genome [18], a 2D SG dataset. This label mapping from Replica to Visual Genome facilitates zero-shot transfer from 2D SG models, as this approach eliminates the requirement for additional training on the Replica dataset. The dataset splits and label mapping are available in the released code. In addition, detailed statistics for the dataset are provided in the supplementary materials (Section 8). As a result, ReplicaSSG offers high-quality instance-segmented meshes with comprehensive relationship annotations. + +# 4.1.2. Baseline Methods + +We evaluated FROSS on the 3DSSG dataset against several baseline methods for SSG generation. The baselines include three image-based methods, IMP [36], VGfM [9] and Kim's construction framework [16], alongside three point cloud-based methods: 3DSSG [31], SGFN [34] and the approach presented in [35]. Kim's 3D object representation and merging mechanism are integrated into FROSS for its baseline implementation, while other baseline implementations follow $[35]^1$ . In the ablation studies, we investigate the impact of using ground truth 2D SGs and camera trajectories on the ReplicaSSG dataset. + +# 4.1.3. Matching Object Predictions to Ground Truth + +The evaluation methodology follows [35] for matching predicted object instances with ground truth counterparts. As FROSS generates predictions without explicit point cloud output, we establish evaluation metrics using back-projected 3D points. These points derive from the central region $(H / 2\times W / 2$ pixels) of each object's 2D bounding box. The matching process associates each predicted point with its nearest ground truth point within a 0.1-meter radius. A predicted object establishes correspondence with a ground truth object under two criteria, described as follows: + +1. The majority (more than $50\%$ ) of its back-projected 3D points must each have their nearest ground truth point neighbor within the corresponding ground truth object. +2. The ratio between the second-largest and largest ground truth object overlap counts must remain below $75\%$ . + +In contrast to the other baseline methods, which may permit multiple predictions to be matched with a single ground truth object due to possible over-segmentation in input instance-segmented point clouds, our proposed FROSS framework enforces a stricter one-to-one correspondence between predicted and ground truth objects. Since FROSS merges identical objects into a single prediction, matching multiple predictions to the same ground truth object is not required. The one-to-one correspondence enforces the integrity of the proposed merging algorithm, ensuring accurate association between predicted and ground truth objects. + +# 4.1.4. Evaluation Metrics + +The evaluation of SSG generation employs recall metrics for relationships, objects, and predicate, a standard approach in both 2D and 3D SG evaluation [13, 31, 34, 35, 38]. Object recall quantifies the proportion of ground truth instances matched to predictions with correct category labels. Predicate recall computes the proportion of correctly classified predicates between detected objects, regardless of the object classes. Relationship recall measures the proportion of correctly identified triplets (subject, object, predicate) that meet two criteria: valid detection of both subject and object, and correct classification of their connecting predicate. To mitigate the impact of class imbalance, we report mean recall (mRecall) to provide a more equitable evaluation across all classes. Note that a more detailed explanation of the utilized recall metric is offered in the supplementary materials (Section 6). + +# 4.2. Implementation Details + +For our 2D SG generation model, we employ EGTR [12] with RT-DETRv2-M [21, 44] as the object detection backbone. The 2D SGs from the 3DSSG dataset [31] serve as training data specifically for 3DSSG experiments, while the Visual Genome dataset [18] functions as the training foundation for ReplicaSSG experiments. Our methodology ad + +Table 1. Performance comparison of 3D SSG generation methods on the 3DSSG dataset and the end-to-end latency without environmental mapping reported in their original literature, along with their respective input modalities. The best and second-best results are highlighted in red, and blue, respectively. Latencies are reported in milliseconds. + +
MethodRecall (%)mRecall (%)LatencyInput Modality
Rel.Obj.Pred.Obj.Pred.
IMP [36]19.749.520.934.713.8-RGB-D
VGfM [9]19.650.020.434.811.0250RGB
3DSSG [31]12.937.422.026.214.4-Point Cloud
SGFN [34]22.051.627.537.724.0161RGB-D
Wu [35]23.353.828.443.826.6191RGB-D2
Kim [16]9.159.07.151.08.0310RGB-D
FROSS (Ours)27.962.433.063.818.07RGB-D
+ +heres to the training protocol established in EGTR, which initiates with pre-training the object detector for standard object detection tasks. All hyperparameters were determined through grid search evaluation on the validation split, with particular emphasis on relationship recall optimization. Our implementation applies a confidence threshold of 0.7 for object filtering and retains only the top ten relationships per 2D SG. For Gaussian merging operations, the Hellinger distance threshold $\delta_{\mathrm{d}}$ is established at 0.85. The main experimental evaluations utilize ground truth trajectories to guide the SSG generation process. The impact of estimated trajectories are further analyzed via ablation studies in Section 4.6.2. + +# 4.3.Quantitative Results + +Table 1 presents a performance and latency comparison between FROSS and various representative baseline methodologies. The results reveal that FROSS achieves the highest performance among all baseline methods with much lower processing latency. This substantiates the advantages of lifting scene graphs from 2D images over direct point cloud reasoning [31, 34, 35], as point clouds can sometimes present challenges with occlusion, sparsity, and reconstruction quality. The superior performance compared to the other image-based methods [9, 16, 36] can be attributed to our Gaussian-based merging strategy. In contrast, IMP [36] employs a voting mechanism for object classification, while VGfM [9] uses a basic recurrent neural network, which may restrict their object and relationship classification accuracy. Kim's baseline [16] achieves similar object recall to FROSS due to the same 2D SG generation pipeline. However, its merging mechanism fails to suppress duplicate detections, which hinders relationship aggregation and leads to significantly lower relationship and predicate recall. These results collectively substantiate the effectiveness of FROSS + +![](images/2ecb1e704d98206bcb0bd47ca757d87c49a85d1e1357c7e6e6c0915b5262fb19.jpg) +Figure 4. Qualitative comparison between FROSS and Wu [35] on the 3DSSG dataset. Only representative objects are visualized. Errors are marked in red, with ground truth label shown in parentheses. Node colors correspond to the respective scene graphs. + +in combining 2D information with Gaussian merging for 3D SSG generation. + +# 4.4. Qualitative Results + +Figure 4 presents the 3D SSG generation results from FROSS, demonstrating its ability to approximate object positions and shapes within a scene through Gaussian distributions, while simultaneously extracting their relationships. The unified objects highlight the effectiveness of the proposed merging algorithm, which consolidates predictions from multiple views, reducing redundant objects and processing time. This representation method offers a high-level scene understanding for downstream applications. In comparison, Wu's method [35] frequently misclassifies objects with similar geometry (e.g., "curtain" vs. "door", "table" vs. "sofa") and confuses semantically close relationships (e.g., "standing on" vs. "supported by"). This performance gap shows FROSS's advantage in utilizing richer visual information, in contrast to Wu [35], which relies solely on cropped object image for visual feature extraction. Additional qualitative results from FROSS are provided in the supplementary materials (Section 7.2). + +# 4.5. Runtime Analysis + +Table 2 examines the latency and frames-per-second (FPS) performance of FROSS, evaluated on an AMD Ryzen 9 7950X processor with an NVIDIA GeForce RTX 4090 GPU. This analysis serves to validate the computational efficiency of FROSS in accordance with its design objective of achieving faster-than-real-time processing capabilities. The evaluation restricts processing to a single CPU core and one GPU to assess system efficiency under limited computational resources. The metrics represent averages across 14,400 frames from four ReplicaSSG test scenes. The RT-DETR object detector and the EGTR relationship extraction module exhibit the highest individual latency. On the other hand, the 3D SSG merging algorithm contributes + +Table 2. Runtime analysis of the key components of FROSS. Obj. Det. refers to the object detection model RT-DETR [21, 44], while Rel. Ext. corresponds to the EGTR relationship extractor [12]. Merging 3D SSGs represents the process described in Section 3.4. Only components with significant latency impact are included. + +
Mean Latency (in milliseconds)
Obj. Det.Rel. Ext.Merging 3D SSGsFPS
2.314.510.12144.09
+ +minimal computational overhead. Compared to the previous methods for online real-time 3D SSG generation methods [16, 34, 35], FROSS demonstrates significantly reduced end-to-end latency and increased FPS, which establishes a substantial improvement in terms of speed and efficiency. + +# 4.6. Ablation Study + +# 4.6.1. Using Ground Truth 2D Scene Graphs + +In the ablation analysis presented in Table 3, we investigate the influence of 2D SG quality on FROSS performance, given that FROSS relies heavily on accurate 2D SG predictions to generate 3D SSGs. The results indicate that the accuracy of the 2D SG has a significant impact on the final 3D SSG quality. In both the 3DSSG and ReplicaSSG datasets, FROSS achieves substantial gains in recall when ground truth 2D SG is used instead of predicted 2D SG. This performance difference is particularly pronounced in the ReplicaSSG dataset, where the baseline FROSS method achieves a relatively low recall score compared to the ground truth-assisted version. The discrepancy stems from differences between ReplicaSSG and Visual Genome [18]. Since the 2D SG model was originally trained on Visual Genome, the model may not generalize effectively to datasets with different characteristics. This leads to lower recall in both object and relationship detection within the ReplicaSSG dataset. This finding suggests the untapped potential for FROSS via domain-adaptive 2D SG models, with current results merely representing a lower bound on the framework's capabilities. + +# 4.6.2. Using Estimated Camera Trajectory + +We further validate FROSS's effectiveness using estimated camera trajectories instead of ground truth poses to assess real-world applicability. This evaluation is crucial as + +Table 3. Analysis of FROSS with predicted and ground truth 2D SGs as input on both the 3DSSG and ReplicaSSG datasets. 'w/ GT' denotes the version of FROSS utilizing ground truth 2D SGs. + +
DatasetMethodRecall (%)mRecall (%)
Rel.Obj.Pred.Obj.Pred.
3DSSG [31]FROSS27.962.433.063.818.0
FROSS w/ GT55.888.656.093.356.8
ReplicaSSGFROSS22.326.127.828.820.4
FROSS w/ GT67.689.167.692.682.6
+ +Table 4. Performance comparison of FROSS with the predicted and the ground truth camera trajectories as input on the ReplicaSSG dataset. Please note that 'w/ SLAM' denotes the version of FROSS utilizing trajectories estimated by ORB-SLAM3 [2]. + +
MethodRecall (%)mRecall (%)
Rel.Obj.Pred.Obj.Pred.
FROSS22.326.127.828.820.4
FROSS w/ SLAM22.725.827.227.720.1
+ +Table 5. Recall on the validation split of the 3DSSG dataset with different Hellinger distance thresholds. + +
Threshold δd0.60.650.70.750.80.850.9
Recall (%)Rel.15.116.520.022.725.930.029.3
Obj.69.569.869.168.965.761.955.4
Pred.16.217.921.824.928.833.634.3
+ +most practical deployments must rely on estimated camera poses rather than perfect ground truth information. Table 4 presents the impact of using estimated trajectories generated by ORB-SLAM3 [2] on the ReplicaSSG dataset. The evaluation results demonstrate a slight improvement in relationship recall, while object and predicate recall exhibit minor decreases. This indicates that FROSS remains robust to inaccuracies in camera trajectory estimation. Note that a more detailed analysis of ORB-SLAM3's trajectory estimation errors on the ReplicaSSG dataset is provided for reference in the supplementary materials (Section 7.4). + +# 4.6.3. Effect of Hellinger Distance Threshold + +Table 5 presents the effect of varying the Hellinger distance threshold $\delta_{\mathrm{d}}$ . A lower threshold (e.g., 0.65) enforces stricter merging criteria, which preserves object distinctions and leads to higher object recall. In contrast, a higher threshold (e.g., 0.85) enables more extensive relationship aggregation between objects, which improves relationship performance. + +# 5. Conclusion + +This study presented FROSS, an innovative framework for real-time generation of 3D SSG in online environments. By representing objects as 3D Gaussian distributions, FROSS achieved computational efficiency that exceeded real-time requirements. The framework integrated a real-time object detection and relationship extraction pipeline to generate 2D SGs, which underwent 3D back-projection and incremental integration into a global SSG through our proposed merging algorithm. Experimental evaluations on existing benchmark and our newly proposed dataset, ReplicaSSG, demonstrated that FROSS surpassed existing methods in both performance metrics and computational speed by significant margins. These advancements established possibilities for real-time SSG applications in open-world environments, particularly relevant to robotics and AR domains. + +# Acknowledgments + +This work was partly supported by the Japan Society for the Promotion of Science KAKENHI Grant Numbers JP24H00733. The authors gratefully acknowledge the support from the National Science and Technology Council (NSTC) in Taiwan under grant numbers NSTC 114-2221-E-002-069-MY3, NSTC 113-2221-E-002-212-MY3, NSTC 114-2218-E-A49-026, and NSTC 114-2640-E-002-006. The authors would also like to express their appreciation for the donation of the GPUs from NVIDIA Corporation and NVIDIA AI Technology Center (NVAITC) used in this work. Furthermore, the authors extend their gratitude to the National Center for High-Performance Computing for providing the necessary computational and storage resources. + +# References + +[1] Iro Armeni, Zhi-Yang He, JunYoung Gwak, Amir R Zamir, Martin Fischer, Jitendra Malik, and Silvio Savarese. 3D scene graph: A structure for unified semantics, 3D space, and camera. In Int. Conf. Comput. Vis., pages 5664-5673, 2019. 1, 3 +[2] Carlos Campos, Richard Elvira, Juan J Gómez Rodríguez, José MM Montiel, and Juan D Tardós. ORB-SLAM3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Trans. Robotics, 37(6):1874-1890, 2021. 2, 8 +[3] Ata Çelen, Guo Han, Konrad Schindler, Luc Van Gool, Iro Armeni, Anton Obukhov, and Xi Wang. I-Design: Personalized llm interior designer. arXiv:2404.02838, 2024. 2 +[4] David F Crouse, Peter Willett, Krishna Pattipati, and Lennart Svensson. A look at gaussian mixture reduction algorithms. In Int. Conf. Inform. Fusion, pages 1-8. IEEE, 2011. 5 +[5] Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and William Yang Wang. LayoutGPT: Compositional visual planning and generation with large language models. Adv. Neural Inform. Process. Syst., 36, 2024. 2 +[6] Maria Fox and Derek Long. PDDL2.1: An extension to PDDL for expressing temporal planning domains. J. Artificial Intell. Research, 20:61-124, 2003. 2 +[7] Gege Gao, Weiyang Liu, Anpei Chen, Andreas Geiger, and Bernhard Schölkopf. GraphDreamer: Compositional 3D scene synthesis from scene graphs. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2 +[8] Lizhao Gao, Bo Wang, and Wenmin Wang. Image captioning with scene-graph based semantic concepts. In Int. Conf. on Mach. Learn. Comput., pages 225-229, 2018. 2 +[9] Paul Gay, James Stuart, and Alessio Del Bue. Visual graphs from motion (VGfM): Scene understanding with object geometry reasoning. In ACCV, pages 330-346. Springer, 2019. 1, 2, 3, 4, 6, 7 +[10] Jörg Hoffmann. FF: The fast-forward planning system. AI magazine, 22(3):57-57, 2001. 2 + +[11] Nathan Hughes, Yun Chang, and Luca Carlone. Hydra: A real-time spatial perception system for 3D scene graph construction and optimization. In Robotics: Sci. Syst., 2022. 3 +[12] Jinbae Im, JeongYeon Nam, Nokyung Park, Hyungmin Lee, and Seunghyun Park. EGTR: Extracting graph from transformer for scene graph generation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 24229-24238, 2024. 2, 3, 4, 6, 8, 1 +[13] Justin Johnson, Ranjay Krishna, Michael Stark, Li-Jia Li, David Shamma, Michael Bernstein, and Li Fei-Fei. Image retrieval using scene graphs. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3668-3678, 2015. 2, 6 +[14] Justin Johnson, Agrim Gupta, and Li Fei-Fei. Image generation from scene graphs. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1219–1228, 2018. 2 +[15] Dong-Jin Kim, Jinsoo Choi, Tae-Hyun Oh, and In So Kweon. Dense relational captioning: Triple-stream networks for relationship-based captioning. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6271-6280, 2019. 2 +[16] Ue-Hwan Kim, Jin-Man Park, Taek-Jin Song, and Jong-Hwan Kim. 3-D scene graph: A sparse and semantic representation of physical environments for intelligent agents. IEEE Trans. Cybernetics, 50(12):4921-4933, 2019. 2, 3, 6, 7, 8 +[17] Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In Int. Conf. Learn. Represent., 2017. 3 +[18] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual Genome: Connecting language and vision using crowdsourced dense image annotations. Int. J. Comput. Vis., 123:32-73, 2017. 2, 6, 8, 3 +[19] Linjie Li, Zhe Gan, Yu Cheng, and Jingjing Liu. Relation-aware graph attention network for visual question answering. In Int. Conf. Comput. Vis., pages 10313-10322, 2019. 2 +[20] Yuanyuan Liu, Chengjiang Long, Zhaoxuan Zhang, Bokai Liu, Qiang Zhang, Baocai Yin, and Xin Yang. Explore contextual information for 3D scene graph generation. IEEE Trans. Vis. Comput. Graph., 29(12):5556-5568, 2022. 3, 6 +[21] Wenyu Lv, Yian Zhao, Qinyao Chang, Kui Huang, Guanzhong Wang, and Yi Liu. RT-DETRv2: Improved baseline with bag-of-freebies for real-time detection transformer. arXiv:2407.17140, 2024. 3, 6, 8 +[22] Jeffri Murragarra-Llerena, Lucas N. Kirsten, Luis Felipe Zeni, and Claudio R. Jung. Probabilistic intersection-over-union for training and evaluation of oriented object detectors. IEEE Trans. Image Process., 33:671-681, 2024. 4, 5 +[23] Pushmeet Kohli Nathan Silberman, Derek Hoiem and Rob Fergus. Indoor segmentation and support inference from RGBD images. In Eur. Conf. Comput. Vis., 2012. 6 +[24] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. PointNet: Deep learning on point sets for 3D classification and segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 652-660, 2017. 3 +[25] Tianwen Qian, Jingjing Chen, Shaoxiang Chen, Bo Wu, and Yu-Gang Jiang. Scene graph refinement network for vi + +sual question answering. IEEE Trans. Multimedia, 25:3950-3961, 2022. 2 +[26] Antoni Rosinol, Andrew Violette, Marcus Abate, Nathan Hughes, Yun Chang, Jingnan Shi, Arjun Gupta, and Luca Carlone. Kimera: From SLAM to spatial perception with 3D dynamic scene graphs. Int. J. Robotics Research, 40(12-14):1510-1546, 2021. 3 +[27] Dhruv Shah, Benjamin Eysenbach, Gregory Kahn, Nicholas Rhinehart, and Sergey Levine. ViNG: Learning Open-World Navigation with Visual Goals. In IEEE Int. Conf. Robotics Auto., 2021. 1 +[28] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke M. Strasdat, Renzo De Nardi, Michael Goesele, Steven Lovegrove, and Richard Newcombe. The Replica dataset: A digital replica of indoor spaces. arXiv:1906.05797, 2019. 2, 6 +[29] Sinan Tan, Mengmeng Ge, Di Guo, Huaping Liu, and Fuchun Sun. Knowledge-based embodied question answering. IEEE Trans. Pattern Anal. Mach. Intell., 45(10):11948-11960, 2023. 2 +[30] Johanna Wald, Armen Avetisyan, Nassir Navab, Federico Tombari, and Matthias Niessner. RIO: 3D object instance relocalization in changing indoor environments. In Int. Conf. Comput. Vis., 2019. 3, 5 +[31] Johanna Wald, Helisa Dhamo, Nassir Navab, and Federico Tombari. Learning 3D semantic scene graphs from 3D indoor reconstructions. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3961-3970, 2020. 1, 2, 3, 5, 6, 7, 8 +[32] Johanna Wald, Nassir Navab, and Federico Tombari. Learning 3D semantic scene graphs with instance embeddings. Int. J. Comput. Vis., 130(3):630-651, 2022. +[33] Ziqin Wang, Bowen Cheng, Lichen Zhao, Dong Xu, Yang Tang, and Lu Sheng. VL-SAT: Visual-linguistic semantics assisted training for 3D semantic scene graph prediction in point cloud. In IEEE Conf. Comput. Vis. Pattern Recog., pages 21560–21569, 2023. 2, 3 +[34] Shun-Cheng Wu, Johanna Wald, Keisuke Tateno, Nassir Navab, and Federico Tombari. SceneGraphFusion: Incremental 3D scene graph prediction from rgb-d sequences. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7515-7525, 2021. 2, 3, 6, 7, 8, 1 +[35] Shun-Cheng Wu, Keisuke Tateno, Nassir Navab, and Federico Tombari. Incremental 3D semantic scene graph prediction from RGB sequences. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5064-5074, 2023. 2, 3, 6, 7, 8, 1 +[36] Danfei Xu, Yuke Zhu, Christopher B Choy, and Li Fei-Fei. Scene graph generation by iterative message passing. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5410-5419, 2017. 2, 6, 7, 1 +[37] Xu Yang, Kaihua Tang, Hanwang Zhang, and Jianfei Cai. Auto-encoding scene graphs for image captioning. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10685-10694, 2019. 2 + +[38] Rowan Zellers, Mark Yatskar, Sam Thomson, and Yejin Choi. Neural Motifs: Scene graph parsing with global context. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5831-5840, 2018. 2, 6, 1, 3 +[39] Guangyao Zhai, Evin Pinar Örnek, Shun-Cheng Wu, Yan Di, Federico Tombari, Nassir Navab, and Benjamin Busam. Commonsscenes: Generating commonsense 3d indoor scenes with scene graph diffusion. In Adv. Neural Inform. Process. Syst., 2023. 2 +[40] Guangyao Zhai, Evin Pinar Örnek, Dave Zhenyu Chen, Ruotong Liao, Yan Di, Nassir Navab, Federico Tombari, and Benjamin Busam. Echoscene: Indoor scene generation via information echo over scene graph diffusion. In Eur. Conf. Comput. Vis., pages 167-184. Springer, 2024. 2 +[41] Chaoyi Zhang, Jianhui Yu, Yang Song, and Weidong Cai. Exploiting edge-oriented reasoning for 3D point-based scene graph analysis. In IEEE Conf. Comput. Vis. Pattern Recog., pages 9705-9715, 2021. 2, 3, 6 +[42] Chaoyi Zhang, Xitong Yang, Ji Hou, Kris Kitani, Weidong Cai, and Fu-Jen Chu. EgoSG: Learning 3D scene graphs from egocentric RGB-D sequences. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2535-2545, 2024. +[43] Shoulong Zhang, Aimin Hao, Hong Qin, et al. Knowledge-inspired 3D scene graph prediction in point cloud. Adv. Neural Inform. Process. Syst., 34:18620-18632, 2021. 3, 6 +[44] Yian Zhao, Wenyu Lv, Shangliang Xu, Jinman Wei, Guanzhong Wang, Qingqing Dang, Yi Liu, and Jie Chen. DETRs beat YOLOs on real-time object detection. In IEEE Conf. Comput. Vis. Pattern Recog., pages 16965-16974, 2024. 3, 4, 6, 8, 2 +[45] Kaizhi Zheng, Xiaotong Chen, Xuehai He, Jing Gu, Linjie Li, Zhengyuan Yang, Kevin Lin, Jianfeng Wang, Lijuan Wang, and Xin Eric Wang. EditRoom: LLM-parameterized graph diffusion for composable 3D room layout editing. arXiv:2410.12836, 2024. 2 +[46] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. EWA splatting. IEEE Trans. Vis. Comput. Graph., 8(3):223-238, 2002. 4 \ No newline at end of file diff --git a/2025/FROSS_ Faster-Than-Real-Time Online 3D Semantic Scene Graph Generation from RGB-D Images/images.zip b/2025/FROSS_ Faster-Than-Real-Time Online 3D Semantic Scene Graph Generation from RGB-D Images/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..649b5941db08270289c564ee40a26b2ac7200565 --- /dev/null +++ b/2025/FROSS_ Faster-Than-Real-Time Online 3D Semantic Scene Graph Generation from RGB-D Images/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2718d61051fef222b6494bc87450ed6dcf4c123d2f4efe8352da12e636e16266 +size 391910 diff --git a/2025/FROSS_ Faster-Than-Real-Time Online 3D Semantic Scene Graph Generation from RGB-D Images/layout.json b/2025/FROSS_ Faster-Than-Real-Time Online 3D Semantic Scene Graph Generation from RGB-D Images/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..75bdbafd05a754ce57b2d49a81ddd1730e2c3e5b --- /dev/null +++ b/2025/FROSS_ Faster-Than-Real-Time Online 3D Semantic Scene Graph Generation from RGB-D Images/layout.json @@ -0,0 +1,7808 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 121, + 103, + 489, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 103, + 489, + 140 + ], + "spans": [ + { + "bbox": [ + 121, + 103, + 489, + 140 + ], + "type": "text", + "content": "FROSS: Faster-than-Real-Time Online 3D Semantic Scene Graph Generation from RGB-D Images" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 82, + 160, + 157, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 160, + 157, + 175 + ], + "spans": [ + { + "bbox": [ + 82, + 160, + 157, + 175 + ], + "type": "text", + "content": "Hao-Yu Hou1,3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 183, + 161, + 251, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 161, + 251, + 175 + ], + "spans": [ + { + "bbox": [ + 183, + 161, + 251, + 175 + ], + "type": "text", + "content": "Chun-Yi Lee2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 276, + 161, + 393, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 161, + 393, + 175 + ], + "spans": [ + { + "bbox": [ + 276, + 161, + 393, + 175 + ], + "type": "text", + "content": "Motoharu Sonogashira3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 419, + 161, + 526, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 161, + 526, + 175 + ], + "spans": [ + { + "bbox": [ + 419, + 161, + 526, + 175 + ], + "type": "text", + "content": "Yasutomo Kawanishi3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 189, + 260, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 189, + 260, + 203 + ], + "spans": [ + { + "bbox": [ + 106, + 189, + 260, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 106, + 189, + 260, + 203 + ], + "type": "text", + "content": "National Tsing Hua University" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 290, + 189, + 428, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 189, + 428, + 203 + ], + "spans": [ + { + "bbox": [ + 290, + 189, + 428, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 290, + 189, + 428, + 203 + ], + "type": "text", + "content": "National Taiwan University" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 457, + 189, + 501, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 457, + 189, + 501, + 202 + ], + "spans": [ + { + "bbox": [ + 457, + 189, + 501, + 202 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 457, + 189, + 501, + 202 + ], + "type": "text", + "content": "RIKEN" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 257, + 296, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 257, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 296, + 557 + ], + "type": "text", + "content": "The ability to abstract complex 3D environments into simplified and structured representations is crucial across various domains. 3D semantic scene graphs (SSGs) achieve this by representing objects as nodes and their interrelationships as edges, facilitating high-level scene understanding. Existing methods for 3D SSG generation, however, face significant challenges, including high computational demands and non-incremental processing that hinder their suitability for real-time open-world applications. To address this issue, we propose FROSS (Faster-than-Real-Time Online 3D Semantic Scene Graph Generation), an innovative approach for online and faster-than-realtime 3D SSG generation that leverages the direct lifting of 2D scene graphs to 3D space and represents objects as 3D Gaussian distributions. This framework eliminates the dependency on precise and computationally-intensive point cloud processing. Furthermore, we extend the Replica dataset with inter-object relationship annotations, creating the ReplicaSSG dataset for comprehensive evaluation of FROSS. The experimental results from evaluations on ReplicaSSG and 3DSSG datasets show that FROSS can achieve superior performance while operating significantly faster than prior 3D SSG generation methods. Our implementation and dataset are publicly available at https://github.com/Howardkhh/FROSS." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 572, + 135, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 572, + 135, + 585 + ], + "spans": [ + { + "bbox": [ + 56, + 572, + 135, + 585 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "The abstraction of complex 3D environments into simplified representations in real time serves as a crucial capability across multiple application domains. The incremental generation of such representations in real time holds particular significance for fields ranging from robotics and computer vision to augmented reality (AR), where systems must perceive and analyze the environment instantaneously. 3D scene graphs (SGs) [1, 9] accomplish this abstraction by representing physical objects as nodes and their interrelationships as edges within a structured framework. 3D se" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 316, + 241, + 555, + 526 + ], + "blocks": [ + { + "bbox": [ + 316, + 241, + 555, + 526 + ], + "lines": [ + { + "bbox": [ + 316, + 241, + 555, + 526 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 555, + 526 + ], + "type": "image", + "image_path": "3ff1488ecdb0ccc9ab8fe0361855f9b2bdd519b3a410a1a4ce91d248d387df87.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 533, + 555, + 578 + ], + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 578 + ], + "type": "text", + "content": "Figure 1. We introduce FROSS, an online real-time 3D semantic scene graph generation method that leverages and integrates 2D scene graphs. FROSS represents objects as 3D Gaussian distributions and operates without requiring 3D reconstruction." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": "mantic scene graphs (SSGs) [31] extend this representation with an emphasis on semantic relationships, such as positional configurations between objects (e.g., a cup on a table) or functional interactions (e.g., a person seated in a chair), which enables high-level scene comprehension, reasoning, and interaction. Real-world applications, however, present open-world challenges where environments often exceed known spatial boundaries and contain previously unseen spaces [27]. For instance, a room may connect to multiple adjacent spaces or corridors, which necessitates continuous" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28818" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "adaptation and incremental scene understanding. This construction process must maintain high efficiency to generate structured environmental representations without creating performance bottlenecks. Meeting the processing demands requires the generation of 3D SSG to operate at faster-than-real-time speeds, with all computations completed between successive image frames from a high-speed sensor stream." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 159, + 295, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 159, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 56, + 159, + 295, + 434 + ], + "type": "text", + "content": "Executing all processes at faster-than-real-time rates confers multiple advantages: (1) prevention of bottlenecks in the 3D SSG generation framework, (2) facilitation of smooth operation on resource-constrained hardware such as edge computing devices, and (3) preservation of computational resources for additional system components. Contemporary 3D SSG generation approaches predominantly focus on offline construction, which requires access to the complete scene data, either instance-segmented point clouds [31, 41], a complete image sequence [9], or both [33]. This assumption of full scene availability conflicts with incremental environmental exploration. Although several studies [34, 35] have developed methods for incremental point cloud reconstruction and segmentation to achieve online and real-time 3D SSG generation, their aggregate system latency remains substantial, even with the utilization of multiple CPU cores. This constraint stems mainly from the substantial computational demands of environmental mapping and point cloud processing, restricting their practical application in real-time systems with limited computational power or other concurrent processes. Given these limitations, the aim of this study is to develop a method for faster-than-real-time online SSG generation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 436, + 295, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 295, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 295, + 615 + ], + "type": "text", + "content": "A fundamental insight for generating 3D SSGs emerges from the observation that precise object pose and shape information is not essential. Given that 3D SSGs provide a high-level semantic understanding of the environment, their functional implications remain valid with approximate object locations. For instance, in robotic applications, high-level planning can be successfully executed using relative spatial relationships [6, 10, 29]. Similarly, in 3D scene synthesis and room layout generation [3, 5, 7, 39, 40, 45], 3D SSGs serve as foundational structures from which more detailed content can be generated. These application scenarios eliminate the need for precise environmental mapping techniques such as Simultaneous Localization and Mapping (SLAM) [2] and additional point cloud processing, both of which contribute significant computational overhead." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": "This observation motivates an alternative approach for online and real-time 3D SSG generation: inferring 2D SGs from images and subsequently lifting them into 3D space, thus circumventing the computational burden of environmental mapping and point cloud processing. To the best of our knowledge, the current literature has not adequately investigated this research direction except [16]. Such unexplored territory presents significant challenges, includ" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 143 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 143 + ], + "type": "text", + "content": "ing the handling of incomplete 2D object observations, the achievement of real-time SG generation, and the effective integration of multi-view information into complete 3D SSGs without object redundancy. These challenges, therefore, provide promising avenues for further innovative research contributions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 144, + 555, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 555, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 555, + 406 + ], + "type": "text", + "content": "In light of these considerations, we introduce FROSS (Faster-than-Real-time Online 3D Semantic Scene Graph Generation), a methodology that constructs 3D SSGs by lifting 2D SGs to 3D space and integrating them into a comprehensive representation. To achieve real-time processing without requiring precise object localization, FROSS initiates the process by constructing 2D SGs from detected objects in images. A key innovation of FROSS lies in its utilization of Gaussian distributions to model objects in 3D space, which facilitates the integration of objects observed across multiple images into a coherent 3D SSG, as illustrated in Figure 1. Specifically, FROSS represents detected objects through Gaussian distributions, lifts these distributions from 2D SGs into 3D space, and merges them according to their predicted classes and statistical distances. We validate FROSS on the 3DSSG dataset [31] and conduct additional evaluations on ReplicaSSG, an extended version of the Replica dataset [28] that we have enhanced with object relationship annotations. ReplicaSSG incorporates object classes defined in Visual Genome [18], a widely-adopted 2D SG dataset. The main contributions of the paper can be summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 407, + 553, + 574 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 407, + 553, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 407, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 314, + 407, + 553, + 454 + ], + "type": "text", + "content": "- We introduce FROSS, an innovative methodology for online real-time generation of 3D SSGs. FROSS demonstrates superior performance and significantly faster processing speeds compared to existing baseline methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 456, + 553, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 456, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 315, + 456, + 553, + 514 + ], + "type": "text", + "content": "- We propose a new merging algorithm based on Gaussian distributions for 3D object integration. This algorithm effectively prevents the proliferation of duplicate or incorrectly detected objects as the input image sequence expands, thereby maintaining computational efficiency." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 515, + 553, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 515, + 553, + 574 + ], + "spans": [ + { + "bbox": [ + 314, + 515, + 553, + 574 + ], + "type": "text", + "content": "- We develop ReplicaSSG, an extension of the Replica dataset with 3D SSG annotations. Comprehensive evaluations of FROSS on both the 3DSSG and ReplicaSSG datasets demonstrate FROSS's effectiveness and computational efficiency over diverse environmental scenarios." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 586, + 400, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 586, + 400, + 597 + ], + "spans": [ + { + "bbox": [ + 314, + 586, + 400, + 597 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "SGs [12, 13, 36, 38] constitute a fundamental data structure for object and relationship representation within images following their introduction by [13]. This representation has demonstrated substantial utility across various scene understanding tasks [8, 14, 15, 19, 25, 37]. The concept later evolved to encompass 3D spaces, where objects may incorporate attributes such as class labels, visual features, shape characteristics, and affordances. Research in 3D SGs developed into two branches of representation: SSGs intro" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28819" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": "duced in [9, 16, 31] and hierarchical scene graphs (HSGs) proposed in [1, 11, 26]. Both frameworks represent 3D objects as nodes with pairwise relationships defined through edges. However, these approaches differ fundamentally in their relationship definitions. SSGs concentrate on semantic relationships between objects, while HSGs organize objects into hierarchical layers that include buildings, rooms, and constituent objects, where edges connect only nodes in identical or adjacent layers. Compared to HSGs, SSGs enable higher-level reasoning capabilities through rich semantic relationship representations. The extension beyond spatial relationships allows SSGs to be applied to broader domains such as robotics, 3D computer vision, and AR." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 237, + 293, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 237, + 293, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 237, + 293, + 251 + ], + "type": "text", + "content": "2.1. Offline 3D Semantic Scene Graph Generation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 255, + 296, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 255, + 296, + 507 + ], + "spans": [ + { + "bbox": [ + 56, + 255, + 296, + 507 + ], + "type": "text", + "content": "The initial development of 3D SSG generation focused primarily on offline settings that required either complete instance-segmented point clouds [20, 31-33, 41, 43] or image sequences [9, 33] as inputs. Such requirements limited their applicability to scenarios with incremental data collection, as they necessitated complete data reprocessing. Despite these limitations, these studies established foundations for 3D SSG generation. The research in [9] pioneered 3D SSG generation by predicting 3D object ellipsoids from image collections and utilizing their geometry and appearance features through a recurrent neural network to extract inter-object relationships. The study in [31] advanced this field through the introduction of the 3DSSG dataset, which enhanced the 3RScan dataset [30] with SSG annotations. This dataset represents the first large-scale collection with dense annotations and has become the primary benchmark for evaluating 3D SSG generation methods. The authors of [31] further contributed a generation methodology by utilizing instance-segmented point clouds as input and leveraged PointNet [24] for feature extraction as well as a graph convolutional network [17] for label prediction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 516, + 291, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 291, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 291, + 529 + ], + "type": "text", + "content": "2.2. Online 3D Semantic Scene Graph Generation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": "Several previous studies have explored 3D SG generation in online settings, with some achieving real-time performance. The research in [16] introduced a framework that predicts local 3D SSGs from RGB-D images and integrates them incrementally into a global representation. However, their computational requirements render real-time execution infeasible. SceneGraphFusion [34] and the study in [35] developed methods for online point cloud reconstruction and segmentation, followed by feature extraction for local SSG prediction and global fusion. These frameworks can achieve real-time performance through multi-threaded execution of parallel components. In contrast to these existing methods, our proposed FROSS framework avoids precise object localization by approximating objects with Gaussian distributions. This approximation approach enables faster-than" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "text", + "content": "real-time online 3D SSG generation, allowing the end-to-end processing latency to be below the sensor data acquisition period while maintaining single-threaded execution." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 120, + 397, + 134 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 120, + 397, + 134 + ], + "spans": [ + { + "bbox": [ + 313, + 120, + 397, + 134 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 140, + 426, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 140, + 426, + 152 + ], + "spans": [ + { + "bbox": [ + 313, + 140, + 426, + 152 + ], + "type": "text", + "content": "3.1. Problem Definition" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 158, + 555, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 158, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 158, + 555, + 373 + ], + "type": "text", + "content": "The primary problem concerned in this study is the online generation of 3D SSGs for environments where the complete scene structure remains unknown a priori. Given a sequence of input images, the primary objective is to construct a 3D SSG " + }, + { + "bbox": [ + 313, + 158, + 555, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 158, + 555, + 373 + ], + "type": "text", + "content": " of the target environment. The graph " + }, + { + "bbox": [ + 313, + 158, + 555, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 158, + 555, + 373 + ], + "type": "text", + "content": " consists of a set of nodes " + }, + { + "bbox": [ + 313, + 158, + 555, + 373 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 313, + 158, + 555, + 373 + ], + "type": "text", + "content": " and their corresponding directed edges " + }, + { + "bbox": [ + 313, + 158, + 555, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 158, + 555, + 373 + ], + "type": "text", + "content": ". Each node " + }, + { + "bbox": [ + 313, + 158, + 555, + 373 + ], + "type": "inline_equation", + "content": "v_{i} = (c_{i},\\mu_{i}^{3\\mathrm{D}},\\Sigma_{i}^{3\\mathrm{D}})\\in \\mathcal{V}" + }, + { + "bbox": [ + 313, + 158, + 555, + 373 + ], + "type": "text", + "content": " is indexed by " + }, + { + "bbox": [ + 313, + 158, + 555, + 373 + ], + "type": "inline_equation", + "content": "i\\in \\{1,\\ldots ,N\\}" + }, + { + "bbox": [ + 313, + 158, + 555, + 373 + ], + "type": "text", + "content": ", representing a unique object in the scene. Every object has a class label " + }, + { + "bbox": [ + 313, + 158, + 555, + 373 + ], + "type": "inline_equation", + "content": "c_{i}\\in \\mathcal{C}" + }, + { + "bbox": [ + 313, + 158, + 555, + 373 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 158, + 555, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 158, + 555, + 373 + ], + "type": "text", + "content": " is the set of object categories. While different approaches may incorporate various node attributes, our formulation represents each object as a 3D Gaussian distribution parameterized by mean vector " + }, + { + "bbox": [ + 313, + 158, + 555, + 373 + ], + "type": "inline_equation", + "content": "\\mu_i^{3\\mathrm{D}}\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 158, + 555, + 373 + ], + "type": "text", + "content": " and covariance matrix " + }, + { + "bbox": [ + 313, + 158, + 555, + 373 + ], + "type": "inline_equation", + "content": "\\Sigma_i^{3\\mathrm{D}}\\in \\mathbb{R}^{3\\times 3}" + }, + { + "bbox": [ + 313, + 158, + 555, + 373 + ], + "type": "text", + "content": ". The edges " + }, + { + "bbox": [ + 313, + 158, + 555, + 373 + ], + "type": "inline_equation", + "content": "e_{i\\rightarrow j}\\in \\mathcal{E}" + }, + { + "bbox": [ + 313, + 158, + 555, + 373 + ], + "type": "text", + "content": " connect node " + }, + { + "bbox": [ + 313, + 158, + 555, + 373 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 313, + 158, + 555, + 373 + ], + "type": "text", + "content": " to node " + }, + { + "bbox": [ + 313, + 158, + 555, + 373 + ], + "type": "inline_equation", + "content": "v_{j}" + }, + { + "bbox": [ + 313, + 158, + 555, + 373 + ], + "type": "text", + "content": " with relationship " + }, + { + "bbox": [ + 313, + 158, + 555, + 373 + ], + "type": "inline_equation", + "content": "r_{i\\rightarrow j}\\in \\mathcal{R}" + }, + { + "bbox": [ + 313, + 158, + 555, + 373 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 158, + 555, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 313, + 158, + 555, + 373 + ], + "type": "text", + "content": " represents the set of predefined relationship types. Furthermore, the algorithm must be able to maintain online processing capabilities to update " + }, + { + "bbox": [ + 313, + 158, + 555, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 158, + 555, + 373 + ], + "type": "text", + "content": " continuously whenever new image data becomes available." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 381, + 452, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 381, + 452, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 452, + 393 + ], + "type": "text", + "content": "3.2. Overview of Framework" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 399, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 555, + 604 + ], + "type": "text", + "content": "Figure 2 presents an overview of the proposed FROSS framework, which processes image sequences with corresponding depth information and camera transformations as input. The framework begins with the EGTR [12] 2D SG generation model, with the object detection backbone model replaced with RT-DETR [21, 44]. This RT-DETR object detector is a state-of-the-art real-time detection model, which preserves intermediate self-attention features for subsequent relationship extraction. FROSS then converts object bounding boxes to 2D Gaussian distributions and projects them into 3D space to generate local SSGs, as elaborated in Section 3.3. These Gaussian distributions serve as efficient approximations of both object positions and their spatial distributions for rapid processing. These local graphs are then integrated into a global SSG through the proposed merging algorithm that is described in detail in Section 3.4." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 612, + 432, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 432, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 432, + 624 + ], + "type": "text", + "content": "3.3. Lifting 2D SG to 3D" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "content": "The 2D SGs generated from RT-DETR [21, 44] and EGTR [12] enable FROSS to construct local 3D SSGs " + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\mathrm{local}} = (\\mathcal{V}_{\\mathrm{local}},\\mathcal{E}_{\\mathrm{local}})" + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "content": " through the integration of depth information and camera transformation. This process involves two key steps: (1) the conversion of object bounding boxes to 2D Gaussians, and (2) their subsequent back-projection into 3D space while preserving inter-object relationships." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28820" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 71, + 521, + 216 + ], + "blocks": [ + { + "bbox": [ + 77, + 71, + 521, + 216 + ], + "lines": [ + { + "bbox": [ + 77, + 71, + 521, + 216 + ], + "spans": [ + { + "bbox": [ + 77, + 71, + 521, + 216 + ], + "type": "image", + "image_path": "c9c0137962ed863d7f4ef0b316022ea0a35d687b329881ea4b6ee1d68fbaf661.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 223, + 555, + 280 + ], + "lines": [ + { + "bbox": [ + 54, + 223, + 555, + 280 + ], + "spans": [ + { + "bbox": [ + 54, + 223, + 555, + 280 + ], + "type": "text", + "content": "Figure 2. An overview of the FROSS framework: (a) The process initiates with object detection via RT-DETR [44] from an RGB-D image and its associated camera pose. (b) The EGTR model [12] extracts inter-object relationships through utilization of preserved self-attention features from RT-DETR. (c) 2D Gaussian distributions derived from detected bounding boxes undergo spatial transformation into 3D space for local 3D SSG construction (Section 3.3). (d) The resultant local 3D SSG is subsequently incorporated into the global 3D SSG through our proposed merging algorithm (Section 3.4)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 289, + 296, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 296, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 296, + 458 + ], + "type": "text", + "content": "2D Gaussian Representation of Bounding Boxes. Although several methods exist for lifting 2D bounding boxes to 3D space, including point cloud back-projection and joint optimization of 3D ellipsoids [9], the Gaussian distribution approach emerges as the most computationally efficient solution among these alternatives. The representation of objects via Gaussian distributions requires the specification of mean " + }, + { + "bbox": [ + 55, + 289, + 296, + 458 + ], + "type": "inline_equation", + "content": "\\mu_i^{2\\mathrm{D}}" + }, + { + "bbox": [ + 55, + 289, + 296, + 458 + ], + "type": "text", + "content": " and covariance matrix " + }, + { + "bbox": [ + 55, + 289, + 296, + 458 + ], + "type": "inline_equation", + "content": "\\Sigma_i^{2\\mathrm{D}}" + }, + { + "bbox": [ + 55, + 289, + 296, + 458 + ], + "type": "text", + "content": " of the 2D object node " + }, + { + "bbox": [ + 55, + 289, + 296, + 458 + ], + "type": "inline_equation", + "content": "v_i^{2\\mathrm{D}}" + }, + { + "bbox": [ + 55, + 289, + 296, + 458 + ], + "type": "text", + "content": ", which would correspond to a 3D node " + }, + { + "bbox": [ + 55, + 289, + 296, + 458 + ], + "type": "inline_equation", + "content": "v_i \\in \\mathcal{V}_{\\mathrm{local}}" + }, + { + "bbox": [ + 55, + 289, + 296, + 458 + ], + "type": "text", + "content": ". Based on the formulation in [22], FROSS models bounding boxes as 2D uniform distributions. The mean values correspond to the bounding box centers, while the covariance matrices maintain consistency with the uniform distributions. The covariance matrix of " + }, + { + "bbox": [ + 55, + 289, + 296, + 458 + ], + "type": "inline_equation", + "content": "v_i^{2\\mathrm{D}}" + }, + { + "bbox": [ + 55, + 289, + 296, + 458 + ], + "type": "text", + "content": " is formulated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 124, + 462, + 295, + 489 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 462, + 295, + 489 + ], + "spans": [ + { + "bbox": [ + 124, + 462, + 295, + 489 + ], + "type": "interline_equation", + "content": "\\Sigma_ {i} ^ {\\mathrm {2 D}} = \\frac {1}{1 2} \\left[ \\begin{array}{c c} W _ {i} ^ {2} & 0 \\\\ 0 & H _ {i} ^ {2} \\end{array} \\right], \\tag {1}", + "image_path": "f5724d8bfd800df9fb238eb750bc0dac9b4bc33526113f8cad0defa85895e629.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 492, + 296, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 492, + 296, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 492, + 296, + 589 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 492, + 296, + 589 + ], + "type": "inline_equation", + "content": "W_{i}" + }, + { + "bbox": [ + 55, + 492, + 296, + 589 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 492, + 296, + 589 + ], + "type": "inline_equation", + "content": "H_{i}" + }, + { + "bbox": [ + 55, + 492, + 296, + 589 + ], + "type": "text", + "content": " are 2D bounding box's width and height. 3D Back-Projection of Gaussian Distributions. Given the 2D Gaussians, their corresponding depths at the Gaussian centroids, and camera transformation matrices, the mean positions " + }, + { + "bbox": [ + 55, + 492, + 296, + 589 + ], + "type": "inline_equation", + "content": "\\mu_{i}^{3\\mathrm{D}}" + }, + { + "bbox": [ + 55, + 492, + 296, + 589 + ], + "type": "text", + "content": " and covariances " + }, + { + "bbox": [ + 55, + 492, + 296, + 589 + ], + "type": "inline_equation", + "content": "\\Sigma_{i}^{3\\mathrm{D}}" + }, + { + "bbox": [ + 55, + 492, + 296, + 589 + ], + "type": "text", + "content": " of the Gaussians in the 3D space can be approximated. The 3D mean " + }, + { + "bbox": [ + 55, + 492, + 296, + 589 + ], + "type": "inline_equation", + "content": "\\mu_{i}^{3\\mathrm{D}}" + }, + { + "bbox": [ + 55, + 492, + 296, + 589 + ], + "type": "text", + "content": " can be derived from the 2D mean " + }, + { + "bbox": [ + 55, + 492, + 296, + 589 + ], + "type": "inline_equation", + "content": "\\mu_{i}^{2\\mathrm{D}} = (p_{x}, p_{y})^{\\top}" + }, + { + "bbox": [ + 55, + 492, + 296, + 589 + ], + "type": "text", + "content": " and its corresponding depth value " + }, + { + "bbox": [ + 55, + 492, + 296, + 589 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 492, + 296, + 589 + ], + "type": "text", + "content": " through the following equation:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 98, + 592, + 295, + 607 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 592, + 295, + 607 + ], + "spans": [ + { + "bbox": [ + 98, + 592, + 295, + 607 + ], + "type": "interline_equation", + "content": "\\mu_ {i} ^ {\\mathrm {3 D}} = R \\cdot K ^ {- 1} \\cdot \\left(p _ {x}, p _ {y}, 1\\right) ^ {\\top} \\cdot z + \\mathbf {t}, \\tag {2}", + "image_path": "8ff18f5da029137134f5913f3111172204ec825864ed7a4c52373768c8910a65.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 611, + 296, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 296, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 296, + 684 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 611, + 296, + 684 + ], + "type": "inline_equation", + "content": "K \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 55, + 611, + 296, + 684 + ], + "type": "text", + "content": " represents the camera intrinsic matrix, and " + }, + { + "bbox": [ + 55, + 611, + 296, + 684 + ], + "type": "inline_equation", + "content": "R \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 55, + 611, + 296, + 684 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 611, + 296, + 684 + ], + "type": "inline_equation", + "content": "\\mathbf{t} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 611, + 296, + 684 + ], + "type": "text", + "content": " denote the rotation matrix and translation vector of the current camera, respectively. The approximation of the 3D covariance matrix utilizes the Jacobian " + }, + { + "bbox": [ + 55, + 611, + 296, + 684 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 55, + 611, + 296, + 684 + ], + "type": "text", + "content": " of the local affine approximation of the projective transformation [46], which is defined as the following:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 122, + 687, + 295, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 687, + 295, + 717 + ], + "spans": [ + { + "bbox": [ + 122, + 687, + 295, + 717 + ], + "type": "interline_equation", + "content": "J = \\left[ \\begin{array}{c c c} \\frac {f _ {x}}{z} & 0 & - \\frac {f _ {x} p _ {x}}{z ^ {2}} \\\\ 0 & \\frac {f _ {y}}{z} & - \\frac {f _ {y} p _ {y}}{z ^ {2}} \\end{array} \\right], \\tag {3}", + "image_path": "7efb96268f91c47e661dec5091a0bf9525bdb11fcdaf275b2ceba9f952f7afe0.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 289, + 555, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 555, + 362 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 555, + 362 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 289, + 555, + 362 + ], + "type": "inline_equation", + "content": "f_{x}" + }, + { + "bbox": [ + 313, + 289, + 555, + 362 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 289, + 555, + 362 + ], + "type": "inline_equation", + "content": "f_{y}" + }, + { + "bbox": [ + 313, + 289, + 555, + 362 + ], + "type": "text", + "content": " represent the focal lengths in " + }, + { + "bbox": [ + 313, + 289, + 555, + 362 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 289, + 555, + 362 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 289, + 555, + 362 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 289, + 555, + 362 + ], + "type": "text", + "content": " dimensions. In the original literature of [46], the Jacobian " + }, + { + "bbox": [ + 313, + 289, + 555, + 362 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 313, + 289, + 555, + 362 + ], + "type": "text", + "content": " is utilized to project 3D covariance matrices onto the 2D image plane. However, our approach inverts this process by back-projecting the 2D covariance matrix " + }, + { + "bbox": [ + 313, + 289, + 555, + 362 + ], + "type": "inline_equation", + "content": "\\Sigma_{i}^{2\\mathrm{D}}" + }, + { + "bbox": [ + 313, + 289, + 555, + 362 + ], + "type": "text", + "content": " into 3D space as " + }, + { + "bbox": [ + 313, + 289, + 555, + 362 + ], + "type": "inline_equation", + "content": "\\bar{\\Sigma}_{i}^{3\\mathrm{D}}" + }, + { + "bbox": [ + 313, + 289, + 555, + 362 + ], + "type": "text", + "content": " via the pseudo-inverse " + }, + { + "bbox": [ + 313, + 289, + 555, + 362 + ], + "type": "inline_equation", + "content": "J^{+}" + }, + { + "bbox": [ + 313, + 289, + 555, + 362 + ], + "type": "text", + "content": " of the Jacobian:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 381, + 368, + 555, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 368, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 381, + 368, + 555, + 384 + ], + "type": "interline_equation", + "content": "\\bar {\\Sigma} _ {i} ^ {\\mathrm {3 D}} = R J ^ {+} \\Sigma_ {i} ^ {\\mathrm {2 D}} J ^ {+ \\top} R ^ {\\top}. \\tag {4}", + "image_path": "900e302a4c4ecc2e3ae2ababf8e44029a9d2e4569e3eb8117183e41bc6d70724.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 390, + 555, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 555, + 511 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 555, + 511 + ], + "type": "text", + "content": "The naive definition alone is insufficient since the original equation is inherently designed for projection from 3D to 2D space. Simply inverting the equation results in a 3D Gaussian that lacks variance information along the depth axis, as the 2D covariance matrix does not capture depth-related variance. This limitation necessitates the specification of variance in the depth dimension. We hypothesize that this variance approximates the average variance of the other dimensions. The specification has to be applied prior to the rotation matrix " + }, + { + "bbox": [ + 313, + 390, + 555, + 511 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 390, + 555, + 511 + ], + "type": "text", + "content": ", leading to the following equations:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 338, + 517, + 555, + 532 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 517, + 555, + 532 + ], + "spans": [ + { + "bbox": [ + 338, + 517, + 555, + 532 + ], + "type": "interline_equation", + "content": "\\Sigma_ {i} ^ {\\mathrm {3 D / /}} = J ^ {+} \\Sigma_ {i} ^ {\\mathrm {2 D}} J ^ {+ \\top}, \\tag {5}", + "image_path": "c3749f63a18a7cf1be4791281f9aa4c67402257a5d4efd35bee6bc883453e70c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 340, + 534, + 554, + 577 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 534, + 554, + 577 + ], + "spans": [ + { + "bbox": [ + 340, + 534, + 554, + 577 + ], + "type": "interline_equation", + "content": "\\Sigma_ {i} ^ {\\mathrm {3 D} \\prime} = \\Sigma_ {i} ^ {\\mathrm {3 D} \\prime \\prime} + \\left[ \\begin{array}{l l l} 0 & 0 & 0 \\\\ 0 & 0 & 0 \\\\ 0 & 0 & \\frac {(\\Sigma_ {i} ^ {\\mathrm {3 D} \\prime \\prime}) _ {1 , 1} + (\\Sigma_ {i} ^ {\\mathrm {3 D} \\prime \\prime}) _ {2 , 2}}{2} \\end{array} \\right]. \\tag {6}", + "image_path": "bd7c5bcf3dd08ef38f988ff5b24078c7c8474dad2bcafbe0a947312dd9ab4bae.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 583, + 555, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 583, + 555, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 583, + 555, + 632 + ], + "type": "text", + "content": "In this formulation, " + }, + { + "bbox": [ + 313, + 583, + 555, + 632 + ], + "type": "inline_equation", + "content": "(\\Sigma_{i}^{3\\mathrm{D} / \\prime})_{m,n}\\in \\mathbb{R}" + }, + { + "bbox": [ + 313, + 583, + 555, + 632 + ], + "type": "text", + "content": " represents the element at position " + }, + { + "bbox": [ + 313, + 583, + 555, + 632 + ], + "type": "inline_equation", + "content": "(m,n)" + }, + { + "bbox": [ + 313, + 583, + 555, + 632 + ], + "type": "text", + "content": " of matrix " + }, + { + "bbox": [ + 313, + 583, + 555, + 632 + ], + "type": "inline_equation", + "content": "\\Sigma_i^{3\\mathrm{D} / \\prime}" + }, + { + "bbox": [ + 313, + 583, + 555, + 632 + ], + "type": "text", + "content": ". Finally, the covariance matrix " + }, + { + "bbox": [ + 313, + 583, + 555, + 632 + ], + "type": "inline_equation", + "content": "\\Sigma_{i}^{3\\mathrm{D}^{\\prime}}" + }, + { + "bbox": [ + 313, + 583, + 555, + 632 + ], + "type": "text", + "content": " is transformed to world coordinates using the rotation matrix " + }, + { + "bbox": [ + 313, + 583, + 555, + 632 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 583, + 555, + 632 + ], + "type": "text", + "content": ", resulting in the following expression:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 395, + 638, + 555, + 653 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 638, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 395, + 638, + 555, + 653 + ], + "type": "interline_equation", + "content": "\\Sigma_ {i} ^ {\\mathrm {3 D}} = R \\Sigma_ {i} ^ {\\mathrm {3 D} \\prime} R ^ {\\top}. \\tag {7}", + "image_path": "e11c93cc4d216da7de1c57d3a00acc93ecc3c82fd91dc11fb4f2810d845116ee.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 660, + 422, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 660, + 422, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 660, + 422, + 673 + ], + "type": "text", + "content": "3.4. Merging 3D SSGs" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "With a local 3D SSG " + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\mathrm{local}} = (\\mathcal{V}_{\\mathrm{local}},\\mathcal{E}_{\\mathrm{local}})" + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": " constructed, it is then integrated with the global 3D SSG " + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\mathrm{global}}" + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": " through Gaussian merging and relationship accumulation. Figure 3" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28821" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": "illustrates the proposed merging process. Specifically, it initializes a queue with nodes from " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "\\nu_{\\mathrm{local}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": " and evaluates potential merging opportunities through distance calculations. For each node " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": " in this queue, FROSS calculates the Hellinger distances [22] between the Gaussian distributions of " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": " and all other nodes " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "v_{j}" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": " of the same semantic class, in both the local node queue and " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "\\nu_{\\mathrm{global}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": ", to identify object correspondence. The algorithm merges nodes with Hellinger distances below the threshold " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathrm{d}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": " and adds unmerged nodes to " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "\\nu_{\\mathrm{global}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": " as new objects. The Hellinger distance " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "H_{D}(i,j)" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": " between Gaussian distributions " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mu_i,\\Sigma_i)" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mu_j,\\Sigma_j)" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": " is quantified through the following equation, wherein " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "B_{D}(i,j)" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": " represents the Bhattacharyya distance:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 234, + 295, + 250 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 234, + 295, + 250 + ], + "spans": [ + { + "bbox": [ + 65, + 234, + 295, + 250 + ], + "type": "interline_equation", + "content": "H _ {D} (i, j) = \\sqrt {1 - \\exp (- B _ {D} (i , j))}, \\tag {8}", + "image_path": "cb78b7cf5163a3b26dfdf3b6e59a3bda5a1c89b25d590050a691dbb96439359d.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 251, + 294, + 290 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 251, + 294, + 290 + ], + "spans": [ + { + "bbox": [ + 66, + 251, + 294, + 290 + ], + "type": "interline_equation", + "content": "B _ {D} (i, j) = \\frac {1}{8} \\Delta \\mu_ {i j} ^ {T} \\Sigma^ {- 1} \\Delta \\mu_ {i j} + \\frac {1}{2} \\ln \\left(\\frac {\\operatorname* {d e t} \\Sigma}{\\sqrt {\\operatorname* {d e t} \\Sigma_ {i} \\operatorname* {d e t} \\Sigma_ {j}}}\\right), \\tag {9}", + "image_path": "6433ba2d88ecc0a56f61e678201af793223d3b95566061160877527da9e5b500.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 299, + 296, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 299, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 299, + 296, + 443 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 299, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\Delta \\mu_{ij} = \\mu_i - \\mu_j" + }, + { + "bbox": [ + 55, + 299, + 296, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 299, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\Sigma = \\frac{1}{2} (\\Sigma_i + \\Sigma_j)" + }, + { + "bbox": [ + 55, + 299, + 296, + 443 + ], + "type": "text", + "content": ". The fusion of these nodes occurs through a weighted integration of their respective Gaussian distributions [4]. The weighting coefficients for this integration process reflect both spatial coverage and detection confidence across multiple viewpoints. Objects detected from diverse angular perspectives and spatial positions receive proportionally higher weights, thus mitigating viewpoint bias and ensuring robust three-dimensional representation. Given that weights " + }, + { + "bbox": [ + 55, + 299, + 296, + 443 + ], + "type": "inline_equation", + "content": "w_i" + }, + { + "bbox": [ + 55, + 299, + 296, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 299, + 296, + 443 + ], + "type": "inline_equation", + "content": "w_j" + }, + { + "bbox": [ + 55, + 299, + 296, + 443 + ], + "type": "text", + "content": " represent the detection frequencies of their respective nodes, the mean vector " + }, + { + "bbox": [ + 55, + 299, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\mu_k" + }, + { + "bbox": [ + 55, + 299, + 296, + 443 + ], + "type": "text", + "content": " and covariance matrix " + }, + { + "bbox": [ + 55, + 299, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\Sigma_k" + }, + { + "bbox": [ + 55, + 299, + 296, + 443 + ], + "type": "text", + "content": " for the resultant merged node are determined via the formulations:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 449, + 295, + 473 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 449, + 295, + 473 + ], + "spans": [ + { + "bbox": [ + 62, + 449, + 295, + 473 + ], + "type": "interline_equation", + "content": "\\mu_ {k} = \\frac {w _ {i} \\mu_ {i} + w _ {j} \\mu_ {j}}{w _ {i} + w _ {j}}, \\tag {10}", + "image_path": "238117e9036b1b535ea76518629a20c144fc94f415be5da4e935c8a64f3f0ce1.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 475, + 295, + 503 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 475, + 295, + 503 + ], + "spans": [ + { + "bbox": [ + 61, + 475, + 295, + 503 + ], + "type": "interline_equation", + "content": "\\Sigma_ {k} = \\frac {w _ {i} \\Sigma_ {i} + w _ {j} \\Sigma_ {j}}{w _ {i} + w _ {j}} + \\frac {w _ {i} w _ {j} \\left(\\mu_ {i} - \\mu_ {j}\\right) \\left(\\mu_ {i} - \\mu_ {j}\\right) ^ {\\top}}{\\left(w _ {i} + w _ {j}\\right) ^ {2}}, \\tag {11}", + "image_path": "e0a0f9399812135994062302dbb288aff34ed0790352a02332a97d670999a2fd.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "inline_equation", + "content": "\\mu_{i},\\mu_{j}" + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "content": " are the mean vectors, and " + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "inline_equation", + "content": "\\Sigma_{i},\\Sigma_{j}" + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "content": " are the covariance matrices of the respective Gaussian distributions to be merged. All relationship edges originally connected to " + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "inline_equation", + "content": "v_{j}" + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "content": " are then redirected to the newly merged node " + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "inline_equation", + "content": "v_{k}" + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "content": ". The final relationship prediction between two global object nodes is selected through majority voting, choosing the most frequent prediction across viewpoints." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 167, + 440, + 181 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 167, + 440, + 181 + ], + "spans": [ + { + "bbox": [ + 313, + 167, + 440, + 181 + ], + "type": "text", + "content": "4. Experimental Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 188, + 555, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 188, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 188, + 555, + 344 + ], + "type": "text", + "content": "The evaluation of FROSS encompasses two 3D SSG datasets: 3DSSG [31] and our developed ReplicaSSG datasets. Section 4.1 introduces the datasets, baseline SSG generation methods, and evaluation metrics. Section 4.2 presents the implementation details. Section 4.3 provides comprehensive comparisons between FROSS and existing SSG generation methods on the 3DSSG dataset. Moreover, the effectiveness of our approach is further demonstrated through qualitative results in Section 4.4. We further provide runtime analyses on the ReplicaSSG dataset in Section 4.5, along with additional ablation studies in Section 4.6. More quantitative and qualitative analyses are provided in the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 352, + 417, + 364 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 352, + 417, + 364 + ], + "spans": [ + { + "bbox": [ + 313, + 352, + 417, + 364 + ], + "type": "text", + "content": "4.1. Evaluation Setup" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 369, + 378, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 378, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 378, + 380 + ], + "type": "text", + "content": "4.1.1. Datasets" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 384, + 556, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 556, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 556, + 504 + ], + "type": "text", + "content": "3DSSG. The 3DSSG dataset [31] extends 3RScan [30], which encompasses 1,482 scans of indoor environments with their corresponding RGB-D image sequences, 3D meshes, and dense instance segmentation. The dataset contains 363,555 images, with an average of 245 images per scan. 3DSSG augments the base dataset with object attributes, hierarchical category labels, and directed edges that describe inter-object semantic relationships such as 'standing on,' 'attached to,' and 'same color.' This dataset has established a primary benchmark for numerous" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 106, + 523, + 504, + 657 + ], + "blocks": [ + { + "bbox": [ + 106, + 523, + 504, + 657 + ], + "lines": [ + { + "bbox": [ + 106, + 523, + 504, + 657 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 504, + 657 + ], + "type": "image", + "image_path": "32fdd0c43df1d873c6024bd66ff45d0cd00e1cad2bbe51d739b17ac8c4e08fbb.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 666, + 555, + 711 + ], + "lines": [ + { + "bbox": [ + 54, + 666, + 555, + 711 + ], + "spans": [ + { + "bbox": [ + 54, + 666, + 555, + 711 + ], + "type": "text", + "content": "Figure 3. Illustration of the object merging process using the proposed algorithm described in Section 3.4. (a) Starting with an input image, a set of objects with predicted categories and bounding boxes is generated. (b) These objects are then lifted into 3D spaces to form the object nodes in " + }, + { + "bbox": [ + 54, + 666, + 555, + 711 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{\\mathrm{local}}" + }, + { + "bbox": [ + 54, + 666, + 555, + 711 + ], + "type": "text", + "content": " with colors matching their bounding boxes (yellow, cyan, and blue), while pre-existing objects in " + }, + { + "bbox": [ + 54, + 666, + 555, + 711 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{\\mathrm{global}}" + }, + { + "bbox": [ + 54, + 666, + 555, + 711 + ], + "type": "text", + "content": " are shown in green. (c) The merging algorithm combines the objects linked by the arrows, resulting in the updated global objects, represented in red." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28822" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "content": "studies [20, 31-35, 41-43]. Our experiments utilize the publicly available training and testing splits. To address the class imbalance in both object and relationship labels, this study adopts the commonly used category definitions in [31, 34, 35]. Specifically, the original 160 object categories map to 20 categories from NYUv2 [23], while seven of the 26 predicate categories remain. Following this categorical mapping, 3DSSG contains a total of 21,974 distinct objects and 16,324 inter-object relationships. This extensive collection of relationships offers a substantial foundation for spatial and semantic analyses. Nevertheless, the low image quality and low frame rate of 3DSSG hinders the comprehensive evaluation of FROSS at its full operational capacity. This necessitates our development of ReplicaSSG, which is an enhanced adaptation of the Replica dataset [28]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 255, + 296, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 255, + 296, + 518 + ], + "spans": [ + { + "bbox": [ + 56, + 255, + 296, + 518 + ], + "type": "text", + "content": "ReplicaSSG. ReplicaSSG extends the Replica dataset [28] with newly annotated object relationships. The original Replica provides high-quality reconstructions of indoor environments with instance-segmented meshes and photorealistic renderings, and is particularly suitable for SSG evaluation. Although Replica encompasses only 18 scenes, which precludes its use for training purposes, it serves as an effective evaluation benchmark. In the evaluation of FROSS, seven scenes serve as validation environments for hyperparameter optimization, while the remaining 11 scenes function as performance assessment platforms. Rather than preserving Replica's original object category definitions, ReplicaSSG adopts the classification system from Visual Genome [18], a 2D SG dataset. This label mapping from Replica to Visual Genome facilitates zero-shot transfer from 2D SG models, as this approach eliminates the requirement for additional training on the Replica dataset. The dataset splits and label mapping are available in the released code. In addition, detailed statistics for the dataset are provided in the supplementary materials (Section 8). As a result, ReplicaSSG offers high-quality instance-segmented meshes with comprehensive relationship annotations." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 533, + 160, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 160, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 160, + 544 + ], + "type": "text", + "content": "4.1.2. Baseline Methods" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 551, + 295, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 295, + 682 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 295, + 682 + ], + "type": "text", + "content": "We evaluated FROSS on the 3DSSG dataset against several baseline methods for SSG generation. The baselines include three image-based methods, IMP [36], VGfM [9] and Kim's construction framework [16], alongside three point cloud-based methods: 3DSSG [31], SGFN [34] and the approach presented in [35]. Kim's 3D object representation and merging mechanism are integrated into FROSS for its baseline implementation, while other baseline implementations follow " + }, + { + "bbox": [ + 55, + 551, + 295, + 682 + ], + "type": "inline_equation", + "content": "[35]^1" + }, + { + "bbox": [ + 55, + 551, + 295, + 682 + ], + "type": "text", + "content": ". In the ablation studies, we investigate the impact of using ground truth 2D SGs and camera trajectories on the ReplicaSSG dataset." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 539, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 539, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 539, + 84 + ], + "type": "text", + "content": "4.1.3. Matching Object Predictions to Ground Truth" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 87, + 553, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 87, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 313, + 87, + 553, + 207 + ], + "type": "text", + "content": "The evaluation methodology follows [35] for matching predicted object instances with ground truth counterparts. As FROSS generates predictions without explicit point cloud output, we establish evaluation metrics using back-projected 3D points. These points derive from the central region " + }, + { + "bbox": [ + 313, + 87, + 553, + 207 + ], + "type": "inline_equation", + "content": "(H / 2\\times W / 2" + }, + { + "bbox": [ + 313, + 87, + 553, + 207 + ], + "type": "text", + "content": " pixels) of each object's 2D bounding box. The matching process associates each predicted point with its nearest ground truth point within a 0.1-meter radius. A predicted object establishes correspondence with a ground truth object under two criteria, described as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 208, + 553, + 266 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 208, + 553, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 208, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 314, + 208, + 553, + 243 + ], + "type": "text", + "content": "1. The majority (more than " + }, + { + "bbox": [ + 314, + 208, + 553, + 243 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 314, + 208, + 553, + 243 + ], + "type": "text", + "content": ") of its back-projected 3D points must each have their nearest ground truth point neighbor within the corresponding ground truth object." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 244, + 553, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 244, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 244, + 553, + 266 + ], + "type": "text", + "content": "2. The ratio between the second-largest and largest ground truth object overlap counts must remain below " + }, + { + "bbox": [ + 313, + 244, + 553, + 266 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 313, + 244, + 553, + 266 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 269, + 553, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 269, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 269, + 553, + 399 + ], + "type": "text", + "content": "In contrast to the other baseline methods, which may permit multiple predictions to be matched with a single ground truth object due to possible over-segmentation in input instance-segmented point clouds, our proposed FROSS framework enforces a stricter one-to-one correspondence between predicted and ground truth objects. Since FROSS merges identical objects into a single prediction, matching multiple predictions to the same ground truth object is not required. The one-to-one correspondence enforces the integrity of the proposed merging algorithm, ensuring accurate association between predicted and ground truth objects." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 407, + 424, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 424, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 424, + 418 + ], + "type": "text", + "content": "4.1.4. Evaluation Metrics" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 423, + 555, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 555, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 555, + 615 + ], + "type": "text", + "content": "The evaluation of SSG generation employs recall metrics for relationships, objects, and predicate, a standard approach in both 2D and 3D SG evaluation [13, 31, 34, 35, 38]. Object recall quantifies the proportion of ground truth instances matched to predictions with correct category labels. Predicate recall computes the proportion of correctly classified predicates between detected objects, regardless of the object classes. Relationship recall measures the proportion of correctly identified triplets (subject, object, predicate) that meet two criteria: valid detection of both subject and object, and correct classification of their connecting predicate. To mitigate the impact of class imbalance, we report mean recall (mRecall) to provide a more equitable evaluation across all classes. Note that a more detailed explanation of the utilized recall metric is offered in the supplementary materials (Section 6)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 624, + 447, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 447, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 447, + 635 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 641, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 553, + 713 + ], + "type": "text", + "content": "For our 2D SG generation model, we employ EGTR [12] with RT-DETRv2-M [21, 44] as the object detection backbone. The 2D SGs from the 3DSSG dataset [31] serve as training data specifically for 3DSSG experiments, while the Visual Genome dataset [18] functions as the training foundation for ReplicaSSG experiments. Our methodology ad" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 278, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 278, + 712 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 278, + 712 + ], + "type": "text", + "content": "1https://github.com/ShunChengWu/3DSSG/tree/4b783ec" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28823" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 142, + 294, + 239 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 295, + 137 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 295, + 137 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 295, + 137 + ], + "type": "text", + "content": "Table 1. Performance comparison of 3D SSG generation methods on the 3DSSG dataset and the end-to-end latency without environmental mapping reported in their original literature, along with their respective input modalities. The best and second-best results are highlighted in red, and blue, respectively. Latencies are reported in milliseconds." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 142, + 294, + 239 + ], + "lines": [ + { + "bbox": [ + 55, + 142, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 142, + 294, + 239 + ], + "type": "table", + "html": "
MethodRecall (%)mRecall (%)LatencyInput Modality
Rel.Obj.Pred.Obj.Pred.
IMP [36]19.749.520.934.713.8-RGB-D
VGfM [9]19.650.020.434.811.0250RGB
3DSSG [31]12.937.422.026.214.4-Point Cloud
SGFN [34]22.051.627.537.724.0161RGB-D
Wu [35]23.353.828.443.826.6191RGB-D2
Kim [16]9.159.07.151.08.0310RGB-D
FROSS (Ours)27.962.433.063.818.07RGB-D
", + "image_path": "daaa41bb09e5a7de51469946a89cfcf48375c08bc2ad387f58dcef5aa3e70345.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 260, + 295, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 260, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 54, + 260, + 295, + 415 + ], + "type": "text", + "content": "heres to the training protocol established in EGTR, which initiates with pre-training the object detector for standard object detection tasks. All hyperparameters were determined through grid search evaluation on the validation split, with particular emphasis on relationship recall optimization. Our implementation applies a confidence threshold of 0.7 for object filtering and retains only the top ten relationships per 2D SG. For Gaussian merging operations, the Hellinger distance threshold " + }, + { + "bbox": [ + 54, + 260, + 295, + 415 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathrm{d}}" + }, + { + "bbox": [ + 54, + 260, + 295, + 415 + ], + "type": "text", + "content": " is established at 0.85. The main experimental evaluations utilize ground truth trajectories to guide the SSG generation process. The impact of estimated trajectories are further analyzed via ablation studies in Section 4.6.2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 425, + 174, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 174, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 174, + 437 + ], + "type": "text", + "content": "4.3.Quantitative Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 443, + 295, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 443, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 56, + 443, + 295, + 684 + ], + "type": "text", + "content": "Table 1 presents a performance and latency comparison between FROSS and various representative baseline methodologies. The results reveal that FROSS achieves the highest performance among all baseline methods with much lower processing latency. This substantiates the advantages of lifting scene graphs from 2D images over direct point cloud reasoning [31, 34, 35], as point clouds can sometimes present challenges with occlusion, sparsity, and reconstruction quality. The superior performance compared to the other image-based methods [9, 16, 36] can be attributed to our Gaussian-based merging strategy. In contrast, IMP [36] employs a voting mechanism for object classification, while VGfM [9] uses a basic recurrent neural network, which may restrict their object and relationship classification accuracy. Kim's baseline [16] achieves similar object recall to FROSS due to the same 2D SG generation pipeline. However, its merging mechanism fails to suppress duplicate detections, which hinders relationship aggregation and leads to significantly lower relationship and predicate recall. These results collectively substantiate the effectiveness of FROSS" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 317, + 71, + 553, + 196 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 553, + 196 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 553, + 196 + ], + "type": "image", + "image_path": "2ecb1e704d98206bcb0bd47ca757d87c49a85d1e1357c7e6e6c0915b5262fb19.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 205, + 555, + 250 + ], + "lines": [ + { + "bbox": [ + 313, + 205, + 555, + 250 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 555, + 250 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison between FROSS and Wu [35] on the 3DSSG dataset. Only representative objects are visualized. Errors are marked in red, with ground truth label shown in parentheses. Node colors correspond to the respective scene graphs." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 269, + 555, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 269, + 555, + 294 + ], + "spans": [ + { + "bbox": [ + 313, + 269, + 555, + 294 + ], + "type": "text", + "content": "in combining 2D information with Gaussian merging for 3D SSG generation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 300, + 426, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 426, + 313 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 426, + 313 + ], + "type": "text", + "content": "4.4. Qualitative Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 318, + 556, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 556, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 556, + 534 + ], + "type": "text", + "content": "Figure 4 presents the 3D SSG generation results from FROSS, demonstrating its ability to approximate object positions and shapes within a scene through Gaussian distributions, while simultaneously extracting their relationships. The unified objects highlight the effectiveness of the proposed merging algorithm, which consolidates predictions from multiple views, reducing redundant objects and processing time. This representation method offers a high-level scene understanding for downstream applications. In comparison, Wu's method [35] frequently misclassifies objects with similar geometry (e.g., \"curtain\" vs. \"door\", \"table\" vs. \"sofa\") and confuses semantically close relationships (e.g., \"standing on\" vs. \"supported by\"). This performance gap shows FROSS's advantage in utilizing richer visual information, in contrast to Wu [35], which relies solely on cropped object image for visual feature extraction. Additional qualitative results from FROSS are provided in the supplementary materials (Section 7.2)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 540, + 421, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 421, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 421, + 553 + ], + "type": "text", + "content": "4.5. Runtime Analysis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": "Table 2 examines the latency and frames-per-second (FPS) performance of FROSS, evaluated on an AMD Ryzen 9 7950X processor with an NVIDIA GeForce RTX 4090 GPU. This analysis serves to validate the computational efficiency of FROSS in accordance with its design objective of achieving faster-than-real-time processing capabilities. The evaluation restricts processing to a single CPU core and one GPU to assess system efficiency under limited computational resources. The metrics represent averages across 14,400 frames from four ReplicaSSG test scenes. The RT-DETR object detector and the EGTR relationship extraction module exhibit the highest individual latency. On the other hand, the 3D SSG merging algorithm contributes" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 715 + ], + "type": "text", + "content": "2 Although Wu's method [35] supports RGB inputs, the reported results are based on predictions using dense reconstruction from RGB-D images." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28824" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 77, + 135, + 272, + 172 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 294, + 126 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 294, + 126 + ], + "type": "text", + "content": "Table 2. Runtime analysis of the key components of FROSS. Obj. Det. refers to the object detection model RT-DETR [21, 44], while Rel. Ext. corresponds to the EGTR relationship extractor [12]. Merging 3D SSGs represents the process described in Section 3.4. Only components with significant latency impact are included." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 77, + 135, + 272, + 172 + ], + "lines": [ + { + "bbox": [ + 77, + 135, + 272, + 172 + ], + "spans": [ + { + "bbox": [ + 77, + 135, + 272, + 172 + ], + "type": "table", + "html": "
Mean Latency (in milliseconds)
Obj. Det.Rel. Ext.Merging 3D SSGsFPS
2.314.510.12144.09
", + "image_path": "99608f2383151275bcdeece81691b9aa90a0682dedf8c1f2f7d05d13036a0658.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 187, + 296, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 187, + 296, + 248 + ], + "spans": [ + { + "bbox": [ + 55, + 187, + 296, + 248 + ], + "type": "text", + "content": "minimal computational overhead. Compared to the previous methods for online real-time 3D SSG generation methods [16, 34, 35], FROSS demonstrates significantly reduced end-to-end latency and increased FPS, which establishes a substantial improvement in terms of speed and efficiency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 255, + 149, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 149, + 268 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 149, + 268 + ], + "type": "text", + "content": "4.6. Ablation Study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 272, + 248, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 272, + 248, + 284 + ], + "spans": [ + { + "bbox": [ + 55, + 272, + 248, + 284 + ], + "type": "text", + "content": "4.6.1. Using Ground Truth 2D Scene Graphs" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 286, + 296, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 286, + 296, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 286, + 296, + 526 + ], + "type": "text", + "content": "In the ablation analysis presented in Table 3, we investigate the influence of 2D SG quality on FROSS performance, given that FROSS relies heavily on accurate 2D SG predictions to generate 3D SSGs. The results indicate that the accuracy of the 2D SG has a significant impact on the final 3D SSG quality. In both the 3DSSG and ReplicaSSG datasets, FROSS achieves substantial gains in recall when ground truth 2D SG is used instead of predicted 2D SG. This performance difference is particularly pronounced in the ReplicaSSG dataset, where the baseline FROSS method achieves a relatively low recall score compared to the ground truth-assisted version. The discrepancy stems from differences between ReplicaSSG and Visual Genome [18]. Since the 2D SG model was originally trained on Visual Genome, the model may not generalize effectively to datasets with different characteristics. This leads to lower recall in both object and relationship detection within the ReplicaSSG dataset. This finding suggests the untapped potential for FROSS via domain-adaptive 2D SG models, with current results merely representing a lower bound on the framework's capabilities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 532, + 239, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 239, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 239, + 544 + ], + "type": "text", + "content": "4.6.2. Using Estimated Camera Trajectory" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 546, + 295, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 583 + ], + "type": "text", + "content": "We further validate FROSS's effectiveness using estimated camera trajectories instead of ground truth poses to assess real-world applicability. This evaluation is crucial as" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 58, + 643, + 293, + 711 + ], + "blocks": [ + { + "bbox": [ + 55, + 600, + 296, + 634 + ], + "lines": [ + { + "bbox": [ + 55, + 600, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 296, + 634 + ], + "type": "text", + "content": "Table 3. Analysis of FROSS with predicted and ground truth 2D SGs as input on both the 3DSSG and ReplicaSSG datasets. 'w/ GT' denotes the version of FROSS utilizing ground truth 2D SGs." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 643, + 293, + 711 + ], + "lines": [ + { + "bbox": [ + 58, + 643, + 293, + 711 + ], + "spans": [ + { + "bbox": [ + 58, + 643, + 293, + 711 + ], + "type": "table", + "html": "
DatasetMethodRecall (%)mRecall (%)
Rel.Obj.Pred.Obj.Pred.
3DSSG [31]FROSS27.962.433.063.818.0
FROSS w/ GT55.888.656.093.356.8
ReplicaSSGFROSS22.326.127.828.820.4
FROSS w/ GT67.689.167.692.682.6
", + "image_path": "8eee3ccb36d9006763b37a6372c60d3c06988e7da0308080492a5bffe6b03dad.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 318, + 124, + 551, + 175 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 553, + 116 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 553, + 116 + ], + "type": "text", + "content": "Table 4. Performance comparison of FROSS with the predicted and the ground truth camera trajectories as input on the ReplicaSSG dataset. Please note that 'w/ SLAM' denotes the version of FROSS utilizing trajectories estimated by ORB-SLAM3 [2]." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 124, + 551, + 175 + ], + "lines": [ + { + "bbox": [ + 318, + 124, + 551, + 175 + ], + "spans": [ + { + "bbox": [ + 318, + 124, + 551, + 175 + ], + "type": "table", + "html": "
MethodRecall (%)mRecall (%)
Rel.Obj.Pred.Obj.Pred.
FROSS22.326.127.828.820.4
FROSS w/ SLAM22.725.827.227.720.1
", + "image_path": "b98c3e3c8b392e5e1506893bc185755645f599e1a2a1acedc9af63b541c33174.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 318, + 217, + 551, + 262 + ], + "blocks": [ + { + "bbox": [ + 313, + 186, + 553, + 208 + ], + "lines": [ + { + "bbox": [ + 313, + 186, + 553, + 208 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 553, + 208 + ], + "type": "text", + "content": "Table 5. Recall on the validation split of the 3DSSG dataset with different Hellinger distance thresholds." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 217, + 551, + 262 + ], + "lines": [ + { + "bbox": [ + 318, + 217, + 551, + 262 + ], + "spans": [ + { + "bbox": [ + 318, + 217, + 551, + 262 + ], + "type": "table", + "html": "
Threshold δd0.60.650.70.750.80.850.9
Recall (%)Rel.15.116.520.022.725.930.029.3
Obj.69.569.869.168.965.761.955.4
Pred.16.217.921.824.928.833.634.3
", + "image_path": "bc6c32be9d66a70866285e3bb04d5d137ead62eabccf4d873f7b68af6010e7aa.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 278, + 555, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 278, + 555, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 555, + 410 + ], + "type": "text", + "content": "most practical deployments must rely on estimated camera poses rather than perfect ground truth information. Table 4 presents the impact of using estimated trajectories generated by ORB-SLAM3 [2] on the ReplicaSSG dataset. The evaluation results demonstrate a slight improvement in relationship recall, while object and predicate recall exhibit minor decreases. This indicates that FROSS remains robust to inaccuracies in camera trajectory estimation. Note that a more detailed analysis of ORB-SLAM3's trajectory estimation errors on the ReplicaSSG dataset is provided for reference in the supplementary materials (Section 7.4)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 415, + 506, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 506, + 427 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 506, + 427 + ], + "type": "text", + "content": "4.6.3. Effect of Hellinger Distance Threshold" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 430, + 554, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 554, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 554, + 502 + ], + "type": "text", + "content": "Table 5 presents the effect of varying the Hellinger distance threshold " + }, + { + "bbox": [ + 313, + 430, + 554, + 502 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathrm{d}}" + }, + { + "bbox": [ + 313, + 430, + 554, + 502 + ], + "type": "text", + "content": ". A lower threshold (e.g., 0.65) enforces stricter merging criteria, which preserves object distinctions and leads to higher object recall. In contrast, a higher threshold (e.g., 0.85) enables more extensive relationship aggregation between objects, which improves relationship performance." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 513, + 388, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 513, + 388, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 513, + 388, + 525 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 533, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 712 + ], + "type": "text", + "content": "This study presented FROSS, an innovative framework for real-time generation of 3D SSG in online environments. By representing objects as 3D Gaussian distributions, FROSS achieved computational efficiency that exceeded real-time requirements. The framework integrated a real-time object detection and relationship extraction pipeline to generate 2D SGs, which underwent 3D back-projection and incremental integration into a global SSG through our proposed merging algorithm. Experimental evaluations on existing benchmark and our newly proposed dataset, ReplicaSSG, demonstrated that FROSS surpassed existing methods in both performance metrics and computational speed by significant margins. These advancements established possibilities for real-time SSG applications in open-world environments, particularly relevant to robotics and AR domains." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28825" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 259 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 259 + ], + "type": "text", + "content": "This work was partly supported by the Japan Society for the Promotion of Science KAKENHI Grant Numbers JP24H00733. The authors gratefully acknowledge the support from the National Science and Technology Council (NSTC) in Taiwan under grant numbers NSTC 114-2221-E-002-069-MY3, NSTC 113-2221-E-002-212-MY3, NSTC 114-2218-E-A49-026, and NSTC 114-2640-E-002-006. The authors would also like to express their appreciation for the donation of the GPUs from NVIDIA Corporation and NVIDIA AI Technology Center (NVAITC) used in this work. Furthermore, the authors extend their gratitude to the National Center for High-Performance Computing for providing the necessary computational and storage resources." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 272, + 115, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 115, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 115, + 285 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 292, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 292, + 296, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 292, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 61, + 292, + 296, + 346 + ], + "type": "text", + "content": "[1] Iro Armeni, Zhi-Yang He, JunYoung Gwak, Amir R Zamir, Martin Fischer, Jitendra Malik, and Silvio Savarese. 3D scene graph: A structure for unified semantics, 3D space, and camera. In Int. Conf. Comput. Vis., pages 5664-5673, 2019. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 349, + 296, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 349, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 61, + 349, + 296, + 403 + ], + "type": "text", + "content": "[2] Carlos Campos, Richard Elvira, Juan J Gómez Rodríguez, José MM Montiel, and Juan D Tardós. ORB-SLAM3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Trans. Robotics, 37(6):1874-1890, 2021. 2, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 406, + 296, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 406, + 296, + 438 + ], + "spans": [ + { + "bbox": [ + 62, + 406, + 296, + 438 + ], + "type": "text", + "content": "[3] Ata Çelen, Guo Han, Konrad Schindler, Luc Van Gool, Iro Armeni, Anton Obukhov, and Xi Wang. I-Design: Personalized llm interior designer. arXiv:2404.02838, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 440, + 294, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 440, + 294, + 473 + ], + "spans": [ + { + "bbox": [ + 62, + 440, + 294, + 473 + ], + "type": "text", + "content": "[4] David F Crouse, Peter Willett, Krishna Pattipati, and Lennart Svensson. A look at gaussian mixture reduction algorithms. In Int. Conf. Inform. Fusion, pages 1-8. IEEE, 2011. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 475, + 296, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 475, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 62, + 475, + 296, + 529 + ], + "type": "text", + "content": "[5] Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and William Yang Wang. LayoutGPT: Compositional visual planning and generation with large language models. Adv. Neural Inform. Process. Syst., 36, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 531, + 294, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 531, + 294, + 563 + ], + "spans": [ + { + "bbox": [ + 62, + 531, + 294, + 563 + ], + "type": "text", + "content": "[6] Maria Fox and Derek Long. PDDL2.1: An extension to PDDL for expressing temporal planning domains. J. Artificial Intell. Research, 20:61-124, 2003. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 566, + 296, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 566, + 296, + 609 + ], + "spans": [ + { + "bbox": [ + 62, + 566, + 296, + 609 + ], + "type": "text", + "content": "[7] Gege Gao, Weiyang Liu, Anpei Chen, Andreas Geiger, and Bernhard Schölkopf. GraphDreamer: Compositional 3D scene synthesis from scene graphs. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 611, + 296, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 611, + 296, + 644 + ], + "spans": [ + { + "bbox": [ + 62, + 611, + 296, + 644 + ], + "type": "text", + "content": "[8] Lizhao Gao, Bo Wang, and Wenmin Wang. Image captioning with scene-graph based semantic concepts. In Int. Conf. on Mach. Learn. Comput., pages 225-229, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 646, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 646, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 646, + 296, + 689 + ], + "type": "text", + "content": "[9] Paul Gay, James Stuart, and Alessio Del Bue. Visual graphs from motion (VGfM): Scene understanding with object geometry reasoning. In ACCV, pages 330-346. Springer, 2019. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "type": "text", + "content": "[10] Jörg Hoffmann. FF: The fast-forward planning system. AI magazine, 22(3):57-57, 2001. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[11] Nathan Hughes, Yun Chang, and Luca Carlone. Hydra: A real-time spatial perception system for 3D scene graph construction and optimization. In Robotics: Sci. Syst., 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 108, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 162 + ], + "type": "text", + "content": "[12] Jinbae Im, JeongYeon Nam, Nokyung Park, Hyungmin Lee, and Seunghyun Park. EGTR: Extracting graph from transformer for scene graph generation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 24229-24238, 2024. 2, 3, 4, 6, 8, 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 207 + ], + "type": "text", + "content": "[13] Justin Johnson, Ranjay Krishna, Michael Stark, Li-Jia Li, David Shamma, Michael Bernstein, and Li Fei-Fei. Image retrieval using scene graphs. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3668-3678, 2015. 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 209, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 241 + ], + "type": "text", + "content": "[14] Justin Johnson, Agrim Gupta, and Li Fei-Fei. Image generation from scene graphs. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1219–1228, 2018. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "type": "text", + "content": "[15] Dong-Jin Kim, Jinsoo Choi, Tae-Hyun Oh, and In So Kweon. Dense relational captioning: Triple-stream networks for relationship-based captioning. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6271-6280, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 287, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 341 + ], + "type": "text", + "content": "[16] Ue-Hwan Kim, Jin-Man Park, Taek-Jin Song, and Jong-Hwan Kim. 3-D scene graph: A sparse and semantic representation of physical environments for intelligent agents. IEEE Trans. Cybernetics, 50(12):4921-4933, 2019. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 343, + 553, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 375 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 375 + ], + "type": "text", + "content": "[17] Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In Int. Conf. Learn. Represent., 2017. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 377, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 377, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 317, + 377, + 553, + 441 + ], + "type": "text", + "content": "[18] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual Genome: Connecting language and vision using crowdsourced dense image annotations. Int. J. Comput. Vis., 123:32-73, 2017. 2, 6, 8, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 444, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 553, + 476 + ], + "type": "text", + "content": "[19] Linjie Li, Zhe Gan, Yu Cheng, and Jingjing Liu. Relation-aware graph attention network for visual question answering. In Int. Conf. Comput. Vis., pages 10313-10322, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 478, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 478, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 317, + 478, + 553, + 521 + ], + "type": "text", + "content": "[20] Yuanyuan Liu, Chengjiang Long, Zhaoxuan Zhang, Bokai Liu, Qiang Zhang, Baocai Yin, and Xin Yang. Explore contextual information for 3D scene graph generation. IEEE Trans. Vis. Comput. Graph., 29(12):5556-5568, 2022. 3, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 523, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 566 + ], + "type": "text", + "content": "[21] Wenyu Lv, Yian Zhao, Qinyao Chang, Kui Huang, Guanzhong Wang, and Yi Liu. RT-DETRv2: Improved baseline with bag-of-freebies for real-time detection transformer. arXiv:2407.17140, 2024. 3, 6, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 567, + 553, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 553, + 611 + ], + "type": "text", + "content": "[22] Jeffri Murragarra-Llerena, Lucas N. Kirsten, Luis Felipe Zeni, and Claudio R. Jung. Probabilistic intersection-over-union for training and evaluation of oriented object detectors. IEEE Trans. Image Process., 33:671-681, 2024. 4, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 613, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 645 + ], + "type": "text", + "content": "[23] Pushmeet Kohli Nathan Silberman, Derek Hoiem and Rob Fergus. Indoor segmentation and support inference from RGBD images. In Eur. Conf. Comput. Vis., 2012. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[24] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. PointNet: Deep learning on point sets for 3D classification and segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 652-660, 2017. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[25] Tianwen Qian, Jingjing Chen, Shaoxiang Chen, Bo Wu, and Yu-Gang Jiang. Scene graph refinement network for vi" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28826" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "sual question answering. IEEE Trans. Multimedia, 25:3950-3961, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "type": "text", + "content": "[26] Antoni Rosinol, Andrew Violette, Marcus Abate, Nathan Hughes, Yun Chang, Jingnan Shi, Arjun Gupta, and Luca Carlone. Kimera: From SLAM to spatial perception with 3D dynamic scene graphs. Int. J. Robotics Research, 40(12-14):1510-1546, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 151, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 151, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 151, + 294, + 194 + ], + "type": "text", + "content": "[27] Dhruv Shah, Benjamin Eysenbach, Gregory Kahn, Nicholas Rhinehart, and Sergey Levine. ViNG: Learning Open-World Navigation with Visual Goals. In IEEE Int. Conf. Robotics Auto., 2021. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 195, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 294, + 304 + ], + "type": "text", + "content": "[28] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke M. Strasdat, Renzo De Nardi, Michael Goesele, Steven Lovegrove, and Richard Newcombe. The Replica dataset: A digital replica of indoor spaces. arXiv:1906.05797, 2019. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 304, + 294, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 294, + 347 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 294, + 347 + ], + "type": "text", + "content": "[29] Sinan Tan, Mengmeng Ge, Di Guo, Huaping Liu, and Fuchun Sun. Knowledge-based embodied question answering. IEEE Trans. Pattern Anal. Mach. Intell., 45(10):11948-11960, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 349, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 349, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 349, + 294, + 392 + ], + "type": "text", + "content": "[30] Johanna Wald, Armen Avetisyan, Nassir Navab, Federico Tombari, and Matthias Niessner. RIO: 3D object instance relocalization in changing indoor environments. In Int. Conf. Comput. Vis., 2019. 3, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 393, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 294, + 437 + ], + "type": "text", + "content": "[31] Johanna Wald, Helisa Dhamo, Nassir Navab, and Federico Tombari. Learning 3D semantic scene graphs from 3D indoor reconstructions. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3961-3970, 2020. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 437, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 470 + ], + "type": "text", + "content": "[32] Johanna Wald, Nassir Navab, and Federico Tombari. Learning 3D semantic scene graphs with instance embeddings. Int. J. Comput. Vis., 130(3):630-651, 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "type": "text", + "content": "[33] Ziqin Wang, Bowen Cheng, Lichen Zhao, Dong Xu, Yang Tang, and Lu Sheng. VL-SAT: Visual-linguistic semantics assisted training for 3D semantic scene graph prediction in point cloud. In IEEE Conf. Comput. Vis. Pattern Recog., pages 21560–21569, 2023. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 526, + 294, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 294, + 580 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 294, + 580 + ], + "type": "text", + "content": "[34] Shun-Cheng Wu, Johanna Wald, Keisuke Tateno, Nassir Navab, and Federico Tombari. SceneGraphFusion: Incremental 3D scene graph prediction from rgb-d sequences. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7515-7525, 2021. 2, 3, 6, 7, 8, 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 581, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 581, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 581, + 294, + 624 + ], + "type": "text", + "content": "[35] Shun-Cheng Wu, Keisuke Tateno, Nassir Navab, and Federico Tombari. Incremental 3D semantic scene graph prediction from RGB sequences. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5064-5074, 2023. 2, 3, 6, 7, 8, 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 625, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 668 + ], + "type": "text", + "content": "[36] Danfei Xu, Yuke Zhu, Christopher B Choy, and Li Fei-Fei. Scene graph generation by iterative message passing. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5410-5419, 2017. 2, 6, 7, 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "type": "text", + "content": "[37] Xu Yang, Kaihua Tang, Hanwang Zhang, and Jianfei Cai. Auto-encoding scene graphs for image captioning. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10685-10694, 2019. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 498 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[38] Rowan Zellers, Mark Yatskar, Sam Thomson, and Yejin Choi. Neural Motifs: Scene graph parsing with global context. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5831-5840, 2018. 2, 6, 1, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 172 + ], + "type": "text", + "content": "[39] Guangyao Zhai, Evin Pinar Örnek, Shun-Cheng Wu, Yan Di, Federico Tombari, Nassir Navab, and Benjamin Busam. Commonsscenes: Generating commonsense 3d indoor scenes with scene graph diffusion. In Adv. Neural Inform. Process. Syst., 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 174, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 553, + 228 + ], + "type": "text", + "content": "[40] Guangyao Zhai, Evin Pinar Örnek, Dave Zhenyu Chen, Ruotong Liao, Yan Di, Nassir Navab, Federico Tombari, and Benjamin Busam. Echoscene: Indoor scene generation via information echo over scene graph diffusion. In Eur. Conf. Comput. Vis., pages 167-184. Springer, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 229, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 229, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 229, + 553, + 274 + ], + "type": "text", + "content": "[41] Chaoyi Zhang, Jianhui Yu, Yang Song, and Weidong Cai. Exploiting edge-oriented reasoning for 3D point-based scene graph analysis. In IEEE Conf. Comput. Vis. Pattern Recog., pages 9705-9715, 2021. 2, 3, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 274, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 553, + 319 + ], + "type": "text", + "content": "[42] Chaoyi Zhang, Xitong Yang, Ji Hou, Kris Kitani, Weidong Cai, and Fu-Jen Chu. EgoSG: Learning 3D scene graphs from egocentric RGB-D sequences. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2535-2545, 2024." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 319, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 352 + ], + "type": "text", + "content": "[43] Shoulong Zhang, Aimin Hao, Hong Qin, et al. Knowledge-inspired 3D scene graph prediction in point cloud. Adv. Neural Inform. Process. Syst., 34:18620-18632, 2021. 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 353, + 553, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 353, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 316, + 353, + 553, + 407 + ], + "type": "text", + "content": "[44] Yian Zhao, Wenyu Lv, Shangliang Xu, Jinman Wei, Guanzhong Wang, Qingqing Dang, Yi Liu, and Jie Chen. DETRs beat YOLOs on real-time object detection. In IEEE Conf. Comput. Vis. Pattern Recog., pages 16965-16974, 2024. 3, 4, 6, 8, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 409, + 553, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 409, + 553, + 462 + ], + "spans": [ + { + "bbox": [ + 316, + 409, + 553, + 462 + ], + "type": "text", + "content": "[45] Kaizhi Zheng, Xiaotong Chen, Xuehai He, Jing Gu, Linjie Li, Zhengyuan Yang, Kevin Lin, Jianfeng Wang, Lijuan Wang, and Xin Eric Wang. EditRoom: LLM-parameterized graph diffusion for composable 3D room layout editing. arXiv:2410.12836, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 464, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 464, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 464, + 553, + 498 + ], + "type": "text", + "content": "[46] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. EWA splatting. IEEE Trans. Vis. Comput. Graph., 8(3):223-238, 2002. 4" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28827" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/FVGen_ Accelerating Novel-View Synthesis with Adversarial Video Diffusion Distillation/381cfeb4-88c2-44b1-a694-860cc7f57161_content_list.json b/2025/FVGen_ Accelerating Novel-View Synthesis with Adversarial Video Diffusion Distillation/381cfeb4-88c2-44b1-a694-860cc7f57161_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8136275c1574f54f360e2d49a629ad90aa26caca --- /dev/null +++ b/2025/FVGen_ Accelerating Novel-View Synthesis with Adversarial Video Diffusion Distillation/381cfeb4-88c2-44b1-a694-860cc7f57161_content_list.json @@ -0,0 +1,1658 @@ +[ + { + "type": "text", + "text": "FVGen: Accelerating Novel-View Synthesis with Adversarial Video Diffusion Distillation", + "text_level": 1, + "bbox": [ + 111, + 128, + 885, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wenbin Teng $^{1,2}$ , Gonglin Chen $^{1,2}$ , Haiwei Chen $^{1,2}$ , Yajie Zhao $^{1,2*}$ $^{1}$ Institute for Creative Technologies \n $^{2}$ University of Southern California \n{wenbinte, gonglinc}@usc.edu, {chenh, zhao}@ict.usc.edu,", + "bbox": [ + 192, + 202, + 803, + 257 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f58be604aa60f583d0eb5e36ffcade1f90c56d6b51ae048204c4c6422c873ff7.jpg", + "image_caption": [ + "Figure 1. We present FVGen, a method of fast novel-view synthesis for 3D scene reconstruction with sparse inputs. Recent method [52] unleash the generation capabilities of video diffusion models conditioned on a prior 3D cue for dense view creation, but always suffers from a long sampling time. We propose a novel framework that can significantly reduce the sampling time and maintain the generation quality." + ], + "image_footnote": [], + "bbox": [ + 133, + 295, + 864, + 540 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 614, + 326, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent progress in 3D reconstruction has enabled realistic 3D models from dense image captures, yet challenges persist with sparse views, often leading to artifacts in unseen areas. Recent works leverage Video Diffusion Models (VDMs) to generate dense observations, filling the gaps when only sparse views are available for 3D reconstruction tasks. A significant limitation of these methods is their slow sampling speed when using VDMs. In this paper, we present FVGen, a novel framework that addresses this challenge by enabling fast novel view synthesis using VDMs in as few as four sampling steps. We propose a novel video diffusion model distillation method that distills a multi-step denoising teacher model into a few-step denoising student model using Generative Adversarial Networks (GANs) and softened reverse KL-divergence minimization. Extensive ex", + "bbox": [ + 88, + 647, + 485, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "periments on real-world datasets show that, compared to previous works, our framework generates the same number of novel views with similar (or even better) visual quality while reducing sampling time by more than $90\\%$ . FVGen significantly improves time efficiency for downstream reconstruction tasks, particularly when working with sparse input views (more than 2) where pre-trained VDMs need to be run multiple times to achieve better spatial coverage. Project Page: https://wbteng9526.github.io/fvgen/", + "bbox": [ + 511, + 614, + 908, + 752 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 768, + 645, + 784 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reconstructing 3D scenes from 2D images remains a pivotal research area at the intersection of computer vision and graphics. This field has extensive social impact, with broad applications in autonomous navigation, augmented reality, and more. With the recent development of neural rendering techniques such as NeRF [29] and 3DGS [16], creating high-quality 3D assets from 2D images has become more", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 236, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26095", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "accessible. However, the success of 3D reconstruction depends on dense observations, which are often difficult to obtain in real-world scenarios. This limitation makes it challenging to apply these methods in many practical applications.", + "bbox": [ + 89, + 90, + 480, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Introducing generative models into 3D reconstruction can fill gaps in missing views when only sparse input views are available. However, novel view synthesis from sparse views, especially in the settings with less than 2 views, is inherently an ill-posed problem. A bulk of previous works [7, 8, 30, 43, 52] have found that priors learned from large image and video diffusion models [10, 11] are rich enough to inpaint detailed and reasonable information in the many under-observed areas in the reconstructed 3D representations. We are primarily interested in the use of video diffusion models (VDM) to perform novel view synthesis. What bridges VDM to novel view synthesis is its ability to synthesize continuous views following a camera trajectories between the observed views. The spatio-temporal consistency in these smooth video trajectories, as demonstrated in several works [6, 7, 25, 44, 52], effectively addresses the challenging \"Janus Problem\" that typically limits generation quality in other multi-view settings. However, VDM is a computationally expensive method because of the iterative sampling process that characterized all denoising diffusion methods. The long generation time makes VDM-based generation methods unsuitable for certain real-world downstream tasks, such as dynamic 3D reconstruction, where new consistent views must be generated continuously. This limitation also affects large-scale 3D reconstruction that requires synthesizing a significant number of novel views.", + "bbox": [ + 91, + 172, + 482, + 563 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we show that VDM-based novel view synthesis can be significantly accelerated for free. Without any compromise on the generation quality, we speed up diffusion-based synthesis of novel view by a factor of $90\\%$ with a 4-step student model trained under a GAN objective [23]. Acceleration of diffusion models has been actively studied in the past, but previous techniques may not be well suited to our task due to two reasons: First, we need a method specifically for accelerating video diffusion models, rather than image diffusion models. Second, the method must generalize effectively when trained on multi-view video datasets, which contain significantly less data than general video datasets. To accelerate a diffusion model, recent methods [22, 28, 34, 46, 49] propose minimizing the distributional difference between a few-step student generation model and the original multi-step teacher model. However, these methods are restricted to images with specific domains. The most related work to us is [50], which generates more video frames with faster speed by optimizing a distribution matching distillation (DMD) loss. However, this method requires an effective initialization of the student model. This is completed by generating a noise-", + "bbox": [ + 91, + 569, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "latent pair through an ODE solver [36] and training the student model by optimizing a regression loss [49, 50]. The generation process usually takes a very long time and the student's ability is limited by the teacher's generation quality. More importantly, we have observed that DMD optimizations tend to be unstable and have the tendency towards mode collapse, when trained under multi-view video data. This may be due to the inherent concept of DMD loss optimization being the reverse KL-divergence minimization, which tends to be mode-seeking and zeroing out the modes of teacher's that are not relatively dominant.", + "bbox": [ + 511, + 90, + 903, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To solve the previous problems with VDM acceleration, we propose a novel framework, named FVGen, that performs novel views generation with a fast video diffusion generation with as few as 4 steps. First of all, different from previous GAN-based distillation methods [34, 51], we find it beneficial to initial the few-step student model by training a GAN objective, where the student model is regarded as a generator of fake samples, and, instead of training a separate discriminator as in previous works, we leverage a pre-trained teacher model as the discriminator. Second, we propose the soften reverse KL-divergence to solve the unstable optimization of DMD loss. The soften reverse KL-divergence maintains similar mode-seeking behavior as reverse KL-divergence but with more robust distillation: it prevents the student from ignoring entire space of teacher's distribution, thereby preserving more of teacher's knowledge, which we find particularly beneficial in the context of limited data.", + "bbox": [ + 511, + 258, + 903, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Through effective and robust distillation, FVGen is capable of generating video sequences with comparable or better visual quality as ViewCrafter [52] but with significantly less time. FVGen is trained with Dl3DV-10K [24] dataset and extensive experiments on real world dataset including MipNeRF360 [1] and Tanks-and-Temple [18] demonstrate that our method achieves overall better performance than other state-of-the-art (SOTA) novel view synthesis methods and diffusion distillation methods.", + "bbox": [ + 511, + 531, + 903, + 666 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 511, + 684, + 663, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Diffusion-based Novel View Synthesis", + "text_level": 1, + "bbox": [ + 511, + 710, + 836, + 726 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "NeRF [29] and 3DGS [16] excel at novel view synthesis with dense input views. However, reconstructing scenes from sparse views requires additional priors. Diffusion models [10, 32] can generate realistic pseudo input views from sparse original inputs. Several works use Score-Distillation Sampling [21, 30, 39, 42] to create 3D objects from text prompts or single images by distilling latent diffusion model [32] knowledge into 3D representations. While NeRF, 3DGS, and SDS typically require time-consuming per-scene optimization, view-dependent diffusion models [8, 19, 20, 35] offer an alternative. These mod", + "bbox": [ + 511, + 734, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26096", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1c45046d4adef763a2176185c0c5c2871cb8b26708cda36532c9666f1a6d6155.jpg", + "image_caption": [ + "Figure 2. Overview of student initialization. We initialize our student model by training a GAN objective. The student model $G_{\\theta}$ , initialized with the weights of teacher model, uses few-step denoising to generate fake videos to fool the discriminator $D$ . The fixed teacher model, together with a trainable discriminator, are optimized to differentiate between fake and real videos." + ], + "image_footnote": [], + "bbox": [ + 112, + 90, + 509, + 220 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/38940fdad6faed9e8b2fbb5423e64d5561e0ad715ae0fbf445b5a143f3873c3a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 93, + 883, + 220 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "els directly generate 3D-consistent novel views conditioned on input images but directly lifting 2D images to 3D input therefore face challenges when scaling to larger scenes.", + "bbox": [ + 88, + 287, + 480, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An alternative approach leverages video diffusion models [6, 7, 25, 38, 52] for multi-view image generation. These methods fine-tune pre-trained latent video diffusion models (LVDM) [2, 4, 12, 44, 48] using multi-view images with both 2D and 3D guidance. The 2D guidance typically involves semantic features like CLIP [14] embeddings, while 3D guidance comes from sparse input cues. For instance, [52] uses DUSt3R [40] to build an initial point cloud and guides generation by combining point cloud renders with camera poses. Similarly, [6] conditions generation on feedforward 3DGS renders, enabling 360-degree scene generation with longer frame sequences through multi-step refinement. However, these methods suffer from speed issues caused by video diffusion sampling when generating novel views.", + "bbox": [ + 91, + 335, + 483, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Diffusion Distillation", + "text_level": 1, + "bbox": [ + 89, + 574, + 287, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models typically require numerous denoising steps to generate high-quality samples, which makes them computationally intensive. To accelerate generation, distillation methods train a student model that mimics the teacher diffusion model's behavior while using fewer sampling steps. Using fewer sampling steps reduces quality, so existing works incorporate adversarial training [3, 15, 22, 27, 33, 34, 46] to enhance the student model's output. However, simple GAN training cannot fully bridge the gap between student and teacher distributions, causing the student's generated output to remain noticeably different from the original diffusion model. To address this problem, Distribution Matching Distillation (DMD) [49, 51] optimizes the reverse KL-divergence. For tractable optimization, [49, 51] leverages the gradient of DMD loss, which represents the difference between student and teacher score functions. For fast video generation, [50] extends DMD to video diffusion models like CogVideoX [48]. While these methods work well for general image and video generation tasks, they suffer from mode collapse and training insta", + "bbox": [ + 91, + 598, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "bility when applied to view synthesis on relatively small datasets like multi-view video datasets.", + "bbox": [ + 511, + 287, + 906, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 335, + 651, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Video Diffusion Model", + "text_level": 1, + "bbox": [ + 511, + 359, + 723, + 375 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A video diffusion model is a generative model designed to synthesize realistic video sequences by learning the complex temporal and spatial patterns within video data. Based on the diffusion process, these models iteratively apply noise to video frames and then learn to reverse this process to generate new sequences. The forward diffusion process progressively adds Gaussian noise to the input clean sample $\\pmb{x}_0 \\sim p(\\pmb{x})$ over $T$ timesteps, leading to highly noisy data $\\pmb{x}_T \\sim \\mathcal{N}(0, I)$ . This process is formulated as:", + "bbox": [ + 511, + 383, + 906, + 521 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {t - 1}\\right) = \\mathcal {N} \\left(\\boldsymbol {x} _ {t}; \\sqrt {\\alpha_ {t}} \\boldsymbol {x} _ {t - 1}, (1 - \\alpha_ {t}) \\boldsymbol {I}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 537, + 906, + 555 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha_{t}$ and $\\pmb{x}_t$ are the noise strength and noisy data at timestep $t$ . The model then learns a denoising function, often parameterized by a neural network $\\epsilon_{\\theta}(\\pmb{x}_t,t)$ , to reverse this process and generate realistic video frames, where the optimization function is defined by the MSE loss:", + "bbox": [ + 511, + 568, + 906, + 643 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d i f f}} (\\theta) = \\mathbb {E} _ {t \\sim \\mathcal {U} (0, 1), \\epsilon \\sim \\mathcal {N} (0, I), \\boldsymbol {x} \\sim p, c} \\| \\epsilon - \\epsilon_ {\\theta} (\\boldsymbol {x} _ {t}, t, c) \\| ^ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 657, + 903, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $c$ is the condition embedding of diffusion model which is usually represented as a text or image prompt. Instead of generating sequences at full resolution, Latent Video Diffusion Model (LVDM) first encodes video data $\\pmb{x} \\in \\mathbb{R}^{F \\times C \\times H \\times W}$ into latent space using a pre-trained VAE [17] encoder: $\\pmb{z} = \\mathcal{E}(\\pmb{x})$ , where $\\pmb{z} \\in \\mathbb{R}^{F \\times C \\times h \\times w}$ . The compression will mitigate computational complexity while maintaining video generation quality.", + "bbox": [ + 511, + 699, + 906, + 820 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Distribution Matching Distillation", + "text_level": 1, + "bbox": [ + 511, + 830, + 810, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Distribution matching distillation adopts the idea of variational score distillation [42] that is optimized to match the distribution of student and teacher model by minimizing the", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26097", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/bb23e5dca17e672a494befbd10bd2c91d30b95edd42880458fc1671f4374fa0a.jpg", + "image_caption": [ + "Figure 3. Overview of Distribution Matching Distillation (DMD). The gradient is the difference between the teacher score and fake score w.r.t. the student model output. We apply soften reverse KL-divergence for stable training and avoiding mode-collapse." + ], + "image_footnote": [], + "bbox": [ + 112, + 85, + 890, + 265 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "reverse KL-divergence. Given a student model $G$ parameterized by $\\theta$ , the gradient of reverse KL-divergence is", + "bbox": [ + 89, + 316, + 483, + 349 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\nabla \\mathcal {L} _ {\\mathrm {D M D}} = \\mathbb {E} _ {t} \\left(\\nabla_ {\\theta} \\mathbf {D} _ {\\text {R e v e r s e - K L}} \\left(p _ {\\text {f a k e}, t} \\| p _ {\\text {r e a l}, t}\\right)\\right) \\\\ = - \\mathbb {E} _ {t} \\left(\\left[ s _ {\\text {r e a l}} \\left(F \\left(G _ {\\theta} (\\boldsymbol {z}), t\\right), t\\right) \\right. \\right. \\\\ \\left. - s _ {\\text {f a k e}} \\left(F \\left(G _ {\\theta} (\\boldsymbol {z}), t\\right), t\\right) \\frac {d G _ {\\theta} (\\boldsymbol {z})}{d \\theta} \\right]\\left. \\right), \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 357, + 482, + 462 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $p_{\\mathrm{real}}$ and $p_{\\mathrm{fake}}$ are the real and fake distribution. $z \\sim \\mathcal{N}(0, I)$ is the initial random Gaussian noise. $F$ is the diffusion forward function determined by Eq. (1). $s_{\\mathrm{real}}$ and $s_{\\mathrm{fake}}$ are the real and fake score function and defined as follows according to [37]:", + "bbox": [ + 89, + 470, + 483, + 546 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns \\left(\\boldsymbol {x} _ {t}, t\\right) = \\nabla_ {\\boldsymbol {x} _ {t}} \\log p _ {t} \\left(\\boldsymbol {x} _ {t}\\right) = - \\frac {\\boldsymbol {x} _ {t} - \\alpha_ {t} \\mu \\left(\\boldsymbol {x} _ {t} , t\\right)}{\\sigma_ {t} ^ {2}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 549, + 482, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $p_t$ is the distribution of the noisy sample $x$ . $\\mu$ is the clean sample prediction of the diffusion model. While optimizing student model $G_{\\theta}$ with gradient descent formulated in Eq. (3), the gradients of $\\mu_{\\mathrm{real}}$ are stopped and $\\mu_{\\mathrm{fake}}$ is dynamically optimized with the output of $G_{\\theta}$ with regular diffusion loss similar to Eq. (2)", + "bbox": [ + 89, + 590, + 483, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 89, + 693, + 181, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose Fast Video Generation (FVGen) that distills the generation capability of a multi-step teacher 3D-VDM into a few-step student 3D-VDM. The teacher 3D-VDM used in our experiments is the novel-view synthesis model ViewCrafter [52]. From a set of observed images $I^{\\mathrm{obs}}$ and observed cameras $\\pi^{\\mathrm{obs}}$ , ViewCrafter learns a conditional distribution $I^{\\mathrm{tgt}} \\sim p(I^{\\mathrm{tgt}}|I^{\\mathrm{obs}},\\pi^{\\mathrm{obs}},\\pi^{\\mathrm{tgt}})$ , where $I^{\\mathrm{tgt}} = \\{I_i'\\}_{i=1}^L$ is a sequence of $L$ images corresponding to the target views $\\pi^{\\mathrm{tgt}}$ . The student model FVGen is trained to sample from the same conditional distribution with a much faster inference speed. The architectures of FVGen are shown in Figure 2 and 3. For notation simplicity purpose,", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we ignore including $I^{\\mathrm{obs}}$ , $\\pi^{\\mathrm{obs}}$ , $\\pi^{\\mathrm{tgt}}$ in the future formula as they are all the conditional inputs for both teacher and student diffusion model.", + "bbox": [ + 511, + 316, + 903, + 362 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first initialize the parameters of student model by training with a generative adversarial network (GAN) objective, leveraging the teacher model as discriminator (Section 4.1). After that, we continue to train the student model by optimizing DMD loss and diffusion loss to minimize the soften reverse KL-divergence between the distribution of student and teacher model (Section 4.2).", + "bbox": [ + 511, + 363, + 905, + 468 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Student Initialization with GAN Training", + "text_level": 1, + "bbox": [ + 511, + 478, + 867, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Directly training student model with DMD (Eq. 3) will cause training collapse as it is difficult to naively distill multi-step denoising into few-step. [49, 50] proposed to create noise-latent pairs by an ODE solver [36] and minimize a regression loss with student model for initialization. However, we found that it is very time consuming to generate large amount of video data pairs and the student model will also be restricted by the teacher's limitations. Inspired by [23, 51], we propose to initialize the student model by training a GAN objective with the real images. Specifically, the student model $(G_{\\theta})$ generates fake video samples that fool the teacher model who serves as a discriminator (denoted as $D$ ) by minimizing the generator loss $\\mathcal{L}_G$ . The discriminator $D$ classifies real samples from generated fake samples by maximizing $-\\mathcal{L}_D$ . The min-max game adversarial optimization is formulated as:", + "bbox": [ + 511, + 500, + 906, + 741 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal{L}_{D} = \\mathbb{E}_{\\substack{\\boldsymbol {x}\\sim p_{\\text{real}}\\\\ t\\sim \\mathcal{U}[0,T]}}[\\log f\\left(D(F(\\boldsymbol {x},t))\\right)] \\\\ - \\mathbb {E} _ {\\substack {\\boldsymbol {z} \\sim \\mathcal {N} (0, I) \\\\ t \\sim \\mathcal {U} [ 0, T ]}} \\left[ \\log f \\left(D \\left(F \\left(G _ {\\theta} (\\boldsymbol {z}), t\\right)\\right)\\right) \\right], \\tag{5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 744, + 903, + 800 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal{L}_{G} = \\operatorname{\\mathbb{E}}_{\\substack{\\boldsymbol {z}\\sim \\mathcal{N}(0,\\boldsymbol {I})\\\\ t\\sim \\mathcal{U}[0,T]}}[\\log f\\left(D(F(G_{\\theta}(\\boldsymbol {z}),t))\\right)],\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 803, + 857, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $t$ is the time-step sampling and $F$ is the forward noise-injection model similar to Eq. (3). Discriminator $D$ retrieves the output feature map of the middle block of diffusion model UNet and passes it into the classifier $f$ . To", + "bbox": [ + 511, + 839, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26098", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "fit with video data, $f$ is composed of a 3D Convolutional Neural Network that compresses the feature map into classification logits. Here, we use the original sample $\\pmb{x}$ as the input to the discriminator instead of the generated sample $\\pmb{x}$ from teacher model such that the student model would not be restricted by the generation capability of teacher model. For the purpose of training stability, recent methods [6, 52] predict the velocity field $\\pmb{v}_{\\theta}$ , and we convert it into the sample prediction with $\\pmb{x}_0 = \\sqrt{\\alpha_t\\pmb{x}_t} -\\sqrt{1 - \\alpha_t}\\pmb{v}_{\\theta}$ . Although the generated video samples are still very blurry, the GAN training scheme provides an effective initialization of the student model, which paves the way for more effective training with distribution matching loss.", + "bbox": [ + 89, + 90, + 483, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Distribution Matching with Soften Reverse KL-Divergence Minimization", + "text_level": 1, + "bbox": [ + 89, + 294, + 483, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As discussed in Section 3.2, the core concept of distribution matching distillation (DMD) [49, 51] is to minimize the reverse KL-divergence between student and teacher distributions. While reverse KL is popular for its mode-seeking focus, it has notable limitations in the context of generative modeling, its tendency toward mode collapse. Since reverse KL rewards the student for zeroing out probability in any region where the teacher's density is low, the student can effectively ignore parts where the teacher's distribution are not dominant. If the teacher's distribution $p_{\\mathrm{real}}(\\pmb{x})$ has multiple modes of varying height, minimizing ReverseKL $(p_{\\mathrm{fake}} \\| p_{\\mathrm{real}})$ may lead $p_{\\mathrm{fake}}$ to concentrate on a subset of those modes and assign negligible mass to others. Therefore, the exclusive nature of reverse KL-divergence makes it prone to mode dropping and a lack of coverage of the full distribution.", + "bbox": [ + 89, + 330, + 483, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address the shortcomings of pure reverse KL, we propose to leverage a Soften reverse KL-divergence (Soften RKL) [9, 47]. Soften RKL refers to a modified divergence that retains the mode-seeking bias of reverse KL but softens its exclusion other relatively lower-probability regions, thereby mitigating mode collapse. Compared to reverse KL formulated in Eq. (3), the Soften RKL can be formulated as:", + "bbox": [ + 89, + 573, + 483, + 679 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {D} _ {\\text {S o f t e n - R K L}} \\left(p _ {\\text {f a k e}, t} \\| p _ {\\text {r e a l}, t}\\right) = \\mathbf {D} _ {K L} \\left(\\frac {1}{2} p _ {\\text {r e a l}} + \\frac {1}{2} p _ {\\text {f a k e}} \\| p _ {\\text {r e a l}}\\right), \\\\ = p _ {\\text {f a k e}} (r + 1) \\log \\left(\\frac {1}{2} + \\frac {1}{2 r}\\right), \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 685, + 480, + 767 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $r = r(\\pmb{x},t) \\coloneqq p_{\\mathrm{real},t}(\\pmb{x}) / p_{\\mathrm{fake},t}(\\pmb{x})$ is the density ratio. Here $\\pmb{x}$ denotes the noise-injection version of the output of student model formulated as $F(G_{\\theta}(\\pmb{z}),t)$ with $t \\sim \\mathcal{U}(0,T)$ . Thanks to our optimized GAN objective in Section 4.1, the density ratio could be approximated by the following:", + "bbox": [ + 89, + 773, + 483, + 864 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr (\\boldsymbol {x}, t) = \\frac {p _ {\\text {r e a l} , t} (\\boldsymbol {x})}{p _ {\\text {f a k e} , t} (\\boldsymbol {x})} \\approx \\frac {f (D (\\boldsymbol {x} , t))}{1 - f (D (\\boldsymbol {x} , t))}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 869, + 483, + 905 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In summary, the well-optimized GAN objective not only provides an efficient weight initialization of the student model, but also provides a direct estimate of the density ratio. Here, instead of directly comparing $p_{\\mathrm{fake}}$ and $p_{\\mathrm{real}}$ as in $\\mathbf{D}_{KL}(p_{\\mathrm{fake}}||p_{\\mathrm{real}})$ , we compare an even mixture of $p_{\\mathrm{real}}$ and $p_{\\mathrm{fake}}$ to $p_{\\mathrm{real}}$ . Soften RKL maintains the mode-seeking property of reverse KL, while penalizes more on mode dropping. This criterion yields a more robust distillation since it prevents the student from ignoring the entire teacher distribution, thereby preserving more of teacher's knowledge. Through experiments, we find that the soften RKL yields more stable training and more realistic generations. Computing the divergence directly is generally intractable, the gradient with respect to $\\theta$ is formulated as follows:", + "bbox": [ + 511, + 90, + 906, + 303 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\nabla \\mathcal {L} _ {\\mathrm {D M D}} ^ {\\text {S o f t t e n - R K L}} = \\tag {8} \\\\ - \\mathbb {E} _ {t} \\left(\\frac {1}{r (\\boldsymbol {x} , t)} \\left[ s _ {\\text {r e a l}} (\\boldsymbol {x}, t) - s _ {\\text {f a k e}} (\\boldsymbol {x}, t) \\right] \\frac {d G _ {\\theta} (\\boldsymbol {z})}{d \\theta}\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 314, + 903, + 376 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{x} = F(G_{\\theta}(z),t)$ as previously defined. Apart from the soften RKL divergence optimization, we dynamically update the fake diffusion model $\\mu_{\\mathrm{fake}}$ to adjust to change of student distribution in accordance with the original setting of DMD [49]. The loss term is identical to Eq. (2). We summarize the full training procedure in Alg. 1", + "bbox": [ + 511, + 388, + 906, + 481 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [ + "Algorithm 1 FVGen Training Procedure" + ], + "code_body": "Require: Few-step timesteps $\\mathcal{T} = \\{0, t_1, t_2, \\dots, t_Q\\}$ , pretrained teacher model $\\mu_{\\mathrm{real}}$ , discriminator classifier $D$ dataset $\\mathcal{D}$ .", + "guess_lang": "txt", + "bbox": [ + 514, + 515, + 905, + 559 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Initialize student model $G_{\\theta}$ with $\\mu_{\\mathrm{real}}$ .", + "2: Initialize fake score function with $\\mu_{\\mathrm{fake}}$", + "3: while training do", + "4: Sample a video from dataset $\\pmb{x}_0 \\sim \\mathcal{D}$", + "5: Add noise with timestep $t \\sim \\mathcal{T}$ : $\\pmb{x}_t = \\sqrt{\\alpha_t} \\pmb{x}_0 + \\sqrt{1 - \\alpha_t} \\pmb{\\epsilon}, \\quad \\pmb{\\epsilon} \\sim \\mathcal{N}(0, \\pmb{I})$", + "6: Predict with student: $\\hat{\\pmb{x}}_0 = G_\\theta (\\pmb {x}_t,t)$", + "7: Add noise $\\epsilon' \\sim \\mathcal{N}(0, I)$ with timestep $\\tau \\sim \\mathcal{U}(0, T)$ to real and fake sample:", + "8: $\\hat{\\pmb{x}}_{\\tau} = \\sqrt{\\alpha_{\\tau}}\\hat{\\pmb{x}}_0 + \\sqrt{1 - \\alpha_{\\tau}}\\pmb{\\epsilon}'$", + "9: $\\pmb{x}_{\\tau} = \\sqrt{\\alpha_{\\tau}}\\pmb{x}_0 + \\sqrt{1 - \\alpha_{\\tau}}\\pmb{\\epsilon}'$", + "10: Update $G_{\\theta}$ and $D$ with Eq. (5).", + "11: end while", + "12: Output trained student $G_{\\theta}$", + "13: while training do", + "14: Repeat step 4 - 8", + "15: Update $G_{\\theta}$ with DMD Loss: Eq. (8)", + "16: Add noise $\\pmb{\\epsilon}'' \\sim \\mathcal{N}(0, \\pmb{I})$ with $t_1 \\sim \\mathcal{U}(0, T)$ to $\\hat{\\pmb{x}}_0$ :", + "17: $\\hat{\\pmb{x}}_{t_1} = \\sqrt{\\alpha_{t_1}}\\hat{\\pmb{x}}_0 + \\sqrt{1 - \\alpha_{t_1}}\\pmb{\\epsilon}^{\\prime \\prime}$", + "18: Update $\\mu_{\\mathrm{fake}}$ with denoising loss: Eq. (2)", + "19: end while" + ], + "bbox": [ + 521, + 561, + 903, + 876 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26099", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a5173ad0018a8a98b79691071d9dc58bb5ed4d7e6d02a33f35ce1446b8ba633e.jpg", + "image_caption": [ + "Figure 4. Qualitative Results. We compare FVGen with SOTA novel view synthesis method on MipNeRF360 [1] and Tanks-and-Temples [18] datasets." + ], + "image_footnote": [], + "bbox": [ + 93, + 92, + 903, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 454, + 223, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 89, + 479, + 266, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. All of our student model, teacher model and fake score function is initialized with the parameters of the sparse model of ViewCrafter [52]. We first initialize the student model by training GAN objective. The parameters of teacher model is fixed, and the student model and GAN classifier are trained for 4000 iterations with a two-scale update rule inspired by [51]. Next, we optimize DMD loss by continuously training student model and the fake score function for another 5000 iterations with the same two-scale update rule. The whole pipeline is trained on 8 NVIDIA H100 with batch size 4. The training last about 1 day. We utilize AdamW optimizer [26] and set learning rates of all optimizers to be $5 \\times 10^{-5}$ . The model is trained on images with a resolution of $512 \\times 320$ .", + "bbox": [ + 88, + 501, + 483, + 713 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. FVGen is trained with DL3DV-10K [24], a real-world, scene-level video dataset that includes more than 10K long video clips. Similar to [52], we construct point cloud and ground truth video pairs using DL3DV dataset. We randomly sample short video clips with a random frame stride less than 5. The first frame and last frame are used to construct the prior point cloud with DUSt3R [40]. We use PyTorch3d [31], an efficient renderer to construct the point cloud renders with the camera poses of intermediate frames. Through this process, we create 20,000 training data pairs. We validate our method", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "on two public datasets Tanks-and-Temple [18] and MipN-eRF360 [1]", + "bbox": [ + 511, + 455, + 903, + 486 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline Methods and Evaluation Metrics. To evaluate our proposed method, we compare our method with several state-of-the-art methods: ViewCrafter [52], MVsplat360 [6]. We also compare our video diffusion distillation method with two state-of-the-art methods: CausVid [50] and DMD2 [51]. We use PSNR [13], SSIM [41], LPIPS [53] to measure reconstruction quality. In addition, we also reported the distribution metric, i.e. Frechet Inception Distance (FID), which compares the distribution of generated views and ground truth views. Apart from the visual quality comparison,", + "bbox": [ + 511, + 503, + 906, + 670 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Novel View Synthesis Comparison", + "text_level": 1, + "bbox": [ + 511, + 679, + 812, + 695 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Similar to the evaluation setting of MVSplat360 [6], for each scene of both MipNeRF360 [1] and Tanks-and-Temples [18], we select 5 views that are far from each other but could cover the whole scene. We use our method and baseline method to generate 56 views sampled from the natural camera trajectory (14 views between each 2 input views).", + "bbox": [ + 511, + 700, + 906, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Results The qualitative comparisons with SOTA novel view synthesis methods are shown in Figure 4. DepthSplat is a feed-forward generalizable Gaussian splatting method. When there are relatively low overlap between input frames, DepthSplat would exhibit ob", + "bbox": [ + 511, + 825, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26100", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3593d096fbd141b0bd7d7e363319734fea7531952e193b1721eaca22c323bdfe.jpg", + "image_caption": [ + "Input Views" + ], + "image_footnote": [], + "bbox": [ + 98, + 112, + 194, + 196 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3bc694cfe6a24c0a192576afb84f426f4a59e100418d05a94f94b7daa655dbfa.jpg", + "image_caption": [ + "ViewCrafter + DMD2" + ], + "image_footnote": [], + "bbox": [ + 194, + 112, + 370, + 196 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7ed713d6573a8d58858221afbe310e8d34b9d6b4d10f7d397f68892f5ba57fa9.jpg", + "image_caption": [ + "ViewCrafter + CausVid" + ], + "image_footnote": [], + "bbox": [ + 370, + 112, + 547, + 196 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ea21bd79ce63a0e33b213f2987fa8414ff3a7f703e5d8b418a01ca7ef1b79771.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 547, + 112, + 722, + 196 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dc3664f6f3a5da2c21d8c2d8d3e2c57a9bb8fef29635d735152febda79fe914f.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 722, + 112, + 903, + 196 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b50b0fda2a94f4464748d47e6eed07c821be76a45a5a4b5f6e8d8b858b0a5dc6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 198, + 194, + 295 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/89e6b038db7333ff3d1125e105fcfaa2fa9b59f152c215a08a67e5062c40c368.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 194, + 198, + 370, + 295 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9e23cd413eec3e8b8d44edfb1080969d2684934b0b1b23e24c0e6dba0ff6359e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 370, + 198, + 547, + 295 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ae3ed8211ff5172647b4ff3d95890ab90dde83b558d0761947afd5ea651e1b62.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 547, + 198, + 722, + 295 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/caec3d14189f46947e91b50e8e2342b49914703da329894418d13f8b70ca80b0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 722, + 198, + 903, + 295 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/83ca9a1348f07e041d5f0c085bfc8385339e29fc49b1fe86e58016e2517d53f8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 296, + 194, + 393 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6ef51fe1dbb98a0cbda6fc693ef3f0d17637e45f12f14cc00b990a96981e1582.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 194, + 296, + 370, + 393 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/40295a4d29f721bdcd815a5a3f759482433c5c75152c23602ccaf20578c22839.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 370, + 296, + 547, + 393 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/864656901dbeaefc5fed444ec8ff99d0930a3513424cc0da1f392923dc1d814c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 547, + 296, + 722, + 393 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c4fc1b7603d36534b0d1efd782c4e45483ba707c8ef5a511a93995deaac73bb6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 722, + 296, + 903, + 393 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6dcf859b41a54aa21d081b9996a0ec71f4c82ad19bd8825bd88c175249ab7949.jpg", + "image_caption": [ + "Figure 5. Qualitative Results. We compare FVGen with baseline diffusion distillation method on MipNeRF360 [1] and Tanks-and-Temples [18] datasets." + ], + "image_footnote": [], + "bbox": [ + 96, + 393, + 194, + 479 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9c98280ab97cbbb32f8aac3ed708707be30cbc0c540a411c6f2c711836b5a934.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 194, + 393, + 370, + 479 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/42119404f3c8ef2c82ceec63ea954ee11e42d5951782be40dd53f8b712c7fcb5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 370, + 393, + 547, + 479 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0c83a9787bb16cf32e23c4e8e5f7555014098b34faa93f9e28dddb519611a8aa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 547, + 393, + 722, + 479 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/547326389256b57ba8ee5d23b49790ab3e0f4111f22c1a1b29634e05d7b31219.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 722, + 393, + 903, + 479 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "vious artifacts due to inconsistent depth scale and the issue of floating Gaussians. In comparison, MVSplat360 [6] uses VDM to refine the Gaussian splattering renders and remove the floaters. However, the current MVSplat360 only supports training with low resolution inputs and the current results present significant blurriness. Compared to DepthSplat and MVSplat360, ViewCrafter [52] generates photo-realistic novel views with higher resolution. However, ViewCrafter is limited by the generation speed. Given generating 16 total frames, the average inference time is 13.2 seconds. In comparison, our method achieves a 10x speedup, generating the same number of frames in approximately 1 second and achieves similar visual results. The qualitative results prove that our method is able to distill the multi-step sampling into a few-step sampling process and achieve results that are similar to or even better than teacher ViewCrafter model with the student model.", + "bbox": [ + 88, + 532, + 485, + 790 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Analysis Table 1 presents quantitative comparisons on both datasets of FVGen and other novel view synthesis methods. Except for ViewCrafter [52], our method significantly surpasses other SOTA methods in terms of both perceptual and distribution metrics. Compared with ViewCrafter, our method yields similar quality", + "bbox": [ + 89, + 809, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/aeccd7fc704cf3c75f0643159c883bc9adf4435aeb31914530bf7335c5ee5b2c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PSNR ↑SSIM ↑LPIPS ↓FID ↓Time ↓
mip-360DepthSplat11.230.2130.71532.454.2
MVSplat36012.280.2850.68225.6987.2
ViewCrafter16.350.3460.43316.2866.3
Ours16.280.3520.42917.445.1
TNTDepthSplat12.430.2630.67735.884.3
MVSplat36014.180.3010.53225.2387.3
ViewCrafter18.690.4020.20823.9465.9
Ours18.720.4110.21023.645.0
", + "bbox": [ + 531, + 531, + 890, + 679 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Quantitative Results. We report four quantitative metrics for image quality comparison on 2 separate datasets: Mip-NeRF360 [1] (denoted as mip-360 in the table) and Tanks-and-Temple [18] (denoted as TNT in the table). We highlight the best results in red and second-best in yellow.", + "bbox": [ + 511, + 689, + 906, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "but with more than $90\\%$ faster in generation speed. The quantitative results also highlights the effectiveness of our proposed method.", + "bbox": [ + 511, + 763, + 906, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Diffusion Distillation Comparison", + "text_level": 1, + "bbox": [ + 511, + 816, + 808, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also compare FVGen with SOTA diffusion distillation methods. Similar to the 2-view evaluation setting of DepthSplat [45] and MVSplat [5], we select 2 views that are far from each other in both MipNeRF360 [1] and Tanks-and-", + "bbox": [ + 511, + 839, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26101", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Temples [18] dataset. We use our method and baseline method to generate 16 views samples from the natural camera trajectory (14 intermediate views and 2 input views).", + "bbox": [ + 89, + 90, + 482, + 137 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Results First we perform DMD2 [51] on ViewCrafter for few-step distillation. DMD2 applies end-to-end training of student model, GAN discriminator and fake score function. We find that this training scheme is not stable and present significant variance so that it is hard to converge, leading to blurry results as visually depicted in Figure 5. In addition, we compare our method with a recent work, Caus-Vid [50], on ViewCrafter distillation. CausVid [50] creates a small dataset with ODE solver and initialize the student model by optimizing a regression loss. Caus-Vid also leverages DMD [49] for 4-step video generation. However, the qualitative results indicate that our GAN object training scheme provides better initialization compared to the regression training scheme of ODE pairs.", + "bbox": [ + 89, + 165, + 483, + 378 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3202923bbcee39e56cd53b81833c3c74fee0dedfed137afd09a8516eb4c8a25d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PSNR ↑SSIM ↑LPIPS ↓FID ↓
mip-360VC+DMD29.290.1840.83635.59
VC+CausVid15.770.3360.44119.27
Ours16.280.3520.42917.44
TNTVC+DMD210.270.2090.73137.67
VC+CausVid17.330.4050.23224.92
Ours18.720.4110.21023.64
", + "bbox": [ + 109, + 393, + 464, + 515 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Quantitative Analysis Table 2 presents quantitative comparisons with other diffusion distillation methods, and further proves the superiority of our method. Our method differs with DMD2 [51] in several ways: 1) we apply a 3D discriminator on video data, 2) we perform soften reverse-KL divergence for DMD optimization, 3) our GAN training and DMD training are kept separate for more stable training and more accurate density ratio calculation. Through training DMD2 [51], we found a unstable training process and we recorded the last model before the training collapsed. In comparison, CausVid [50] presents more stable training than DMD2 [51] but still suffer from mode collapse. Our method surpasses the other two baseline methods in both perceptual and distribution metrics.", + "bbox": [ + 89, + 628, + 483, + 840 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 856, + 254, + 871 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we analyze the contributions of each module of our proposed method in Table 3 and Figure 6.", + "bbox": [ + 89, + 878, + 482, + 910 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/355022c16312b43a447ca8b9c11b1f4b65e1bb370356153c1d1e7fe0e691767c.jpg", + "table_caption": [ + "Table 2. Quantitative Results. We report four quantitative metrics for image quality comparison on 2 separate datasets: Mip-NeRF360 [1] (denoted as mip-360 in the table) and Tanks-and-Temple [18] (denoted as TNT in the table). VC stands for ViewCrafter [52] for space saving purpose. $\\mathbf{VC} + \\mathbf{X}$ means applying diffusion acceleration method X on ViewCrafter." + ], + "table_footnote": [], + "table_body": "
GANDMDSoften RKLPSNR ↑SSIM ↑LPIPS ↓FID ↓
✗✓✓8.620.1540.88040.17
✓✗✗16.230.3690.37521.48
✓✓✗16.850.3850.33721.05
✓✓✓17.500.3820.32020.54
", + "bbox": [ + 527, + 88, + 893, + 174 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation studies. Quantitative analysis of different components of FVGen.", + "bbox": [ + 511, + 184, + 903, + 213 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3c8c9a8a076a58ae51952b2685b688dc1f2745054087fc585e5bc404e5073279.jpg", + "image_caption": [ + "Figure 6. Ablation Studies. Visualization of the contribution of different components of FVGen. Our full model presents the most visual similarity between the ground truth." + ], + "image_footnote": [], + "bbox": [ + 514, + 232, + 903, + 318 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Assessing student initialization. As it can be seen from the first column of Figure 6 and Table 3, student initialization is a pivotal module of our architecture. Without student initialization with GAN optimization, the student model would fail to generate samples close to the teacher distribution, thus the teacher score function is not reliable for distillation.", + "bbox": [ + 511, + 386, + 906, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Assessing distribution matching. Optimization of DMD continues to align the student model distribution with teacher model. Comparing with full model, although the perceptual metrics are not significantly degraded, qualitative results show that without distribution matching, the generated results appear significant blurriness and artifacts.", + "bbox": [ + 511, + 506, + 905, + 598 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Assessing soften reverse KL-divergence. As discussed in Section 4.2, soften reverse KL-divergence enables more stable and robust training. Therefore, the DMD optimization achieves better convergence and generated results are more distilled into teacher distribution.", + "bbox": [ + 511, + 614, + 905, + 688 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Limitations", + "text_level": 1, + "bbox": [ + 511, + 707, + 635, + 722 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We identify several limitations in our current work. First, although FVGen achieves similar visual quality compared to ViewCrafter, it is limited by the drawbacks of ViewCrafter: e.g., degradation in structural integrity and consistency when experiencing extremely sparse inputs. Second, FVGen benefits from the synergy of three video diffusion models; therefore, due to computational constraints, we were only able to train the model with videos up to only 16 frames, which may not provide enough coverage for extremely large scenes. Therefore, our future work would address on the limitations mentioned above.", + "bbox": [ + 511, + 734, + 906, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26102", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 281, + 187, + 297 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-NeRF: A Multiscale Representation for Anti-Aliasing Neural Radiance Fields. Technical report, 2021. 2, 6, 7, 8", + "[2] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 3", + "[3] Dar-Yen Chen, Hmrishav Bandyopadhyay, Kai Zou, and Yi-Zhe Song. Nitrofusion: High-fidelity single-step diffusion through dynamic adversarial training. arXiv preprint arXiv:2412.02030, 2024. 3", + "[4] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videographer2: Overcoming data limitations for high-quality video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7310-7320, 2024. 3", + "[5] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splattering from sparse multi-view images. arXiv preprint arXiv:2403.14627, 2024. 7", + "[6] Yuedong Chen, Chuanxia Zheng, Haofei Xu, Bohan Zhuang, Andrea Vedaldi, Tat-Jen Cham, and Jianfei Cai. Mvsplat360: Feed-forward 360 scene synthesis from sparse views. 2024. 2, 3, 5, 6, 7", + "[7] Zilong Chen, Yikai Wang, Feng Wang, Zhengyi Wang, and Huaping Liu. V3d: Video diffusion models are effective 3d generators. arXiv preprint arXiv:2403.06738, 2024. 2, 3", + "[8] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T Barron, and Ben Poole. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv preprint arXiv:2405.10314, 2024. 2", + "[9] Dibya Ghosh. KI divergence - intuition and examples, n.d. Accessed: 2025-03-07. 5", + "[10] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2" + ], + "bbox": [ + 93, + 306, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Supported by the Intelligence Advanced Research Projects Activity (IARPA) via Department of Interior/ Interior Business Center (DOI/IBC) contract number 140D0423C0075. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon. Disclaimer: The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of IARPA, DOI/IBC, or the U.S. Government.", + "[11] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv:2204.03458, 2022. 2", + "[12] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022.3", + "[13] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssm. In 2010 20th international conference on pattern recognition, pages 2366-2369. IEEE, 2010. 6", + "[14] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 3", + "[15] Minguk Kang, Richard Zhang, Connelly Barnes, Sylvain Paris, Suha Kwak, Jaesik Park, Eli Shechtman, Jun-Yan Zhu, and Taesung Park. Distilling diffusion models into conditional gans. In European Conference on Computer Vision, pages 428-447. Springer, 2024. 3", + "[16] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 2", + "[17] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3", + "[18] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36 (4):1-13, 2017. 2, 6, 7, 8", + "[19] Xin Kong, Shikun Liu, Xiaoyang Lyu, Marwan Taher, Xiaojuan Qi, and Andrew J Davison. Eschernet: A generative model for scalable view synthesis. arXiv preprint arXiv:2402.03908, 2024. 2", + "[20] Xinyang Li, Zhangyu Lai, Linning Xu, Yansong Qu, Liujuan Cao, Shengchuan Zhang, Bo Dai, and Rongrong Ji. Director3d: Real-world camera trajectory and 3d scene generation from text. arXiv preprint arXiv:2406.17601, 2024. 2", + "[21] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 2", + "[22] Shanchuan Lin, Anran Wang, and Xiao Yang. Sdxlightning: Progressive adversarial diffusion distillation. arXiv preprint arXiv:2402.13929, 2024. 2, 3", + "[23] Shanchuan Lin, Xin Xia, Yuxi Ren, Ceyuan Yang, Xuefeng Xiao, and Lu Jiang. Diffusion adversarial post-training for one-step video generation. arXiv preprint arXiv:2501.08316, 2025. 2, 4", + "[24] Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22160–22169, 2024. 2, 6" + ], + "bbox": [ + 93, + 92, + 903, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26103", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Fangfu Liu, Wenqiang Sun, Hanyang Wang, Yikai Wang, Haowen Sun, Junliang Ye, Jun Zhang, and Yueqi Duan. Reconx: Reconstruct any scene from sparse views with video diffusion model. arXiv preprint arXiv:2408.16767, 2024. 2, 3", + "[26] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6", + "[27] Yihong Luo, Xiaolong Chen, Xinghua Qu, Tianyang Hu, and Jing Tang. You only sample once: Taming one-step text-to-image synthesis by self-cooperative diffusion gans. arXiv preprint arXiv:2403.12931, 2024. 3", + "[28] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14297-14306, 2023. 2", + "[29] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2", + "[30] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2", + "[31] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020. 6", + "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2", + "[33] Axel Sauer, Frederic Boesel, Tim Dockhorn, Andreas Blattmann, Patrick Esser, and Robin Rombach. Fast high-resolution image synthesis with latent adversarial diffusion distillation. In SIGGRAPH Asia 2024 Conference Papers, pages 1-11, 2024. 3", + "[34] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In European Conference on Computer Vision, pages 87-103. Springer, 2024. 2, 3", + "[35] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 2", + "[36] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2, 4", + "[37] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 4", + "[38] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. In European Conference on Computer Vision, pages 439-457. Springer, 2025. 3" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12619-12629, 2023. 2", + "[40] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 3, 6", + "[41] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6", + "[42] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 2, 3", + "[43] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21551-21561, 2024. 2", + "[44] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Gongye Liu, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Dynamiccafter: Animating open-domain images with video diffusion priors. In European Conference on Computer Vision, pages 399-417. Springer, 2025. 2, 3", + "[45] Haofei Xu, Songyou Peng, Fangjinhua Wang, Hermann Blum, Daniel Barath, Andreas Geiger, and Marc Pollefeys. Depthsplat: Connecting gaussian splatting and depth. arXiv preprint arXiv:2410.13862, 2024. 7", + "[46] Yanwu Xu, Yang Zhao, Zhisheng Xiao, and Tingbo Hou. Ufogen: You forward once large scale text-to-image generation via diffusion gans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8196-8206, 2024. 2, 3", + "[47] Yilun Xu, Weili Nie, and Arash Vahdat. One-step diffusion models with $f$ -divergence distribution matching. arXiv preprint arXiv:2502.15681, 2025. 5", + "[48] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024.3", + "[49] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6613-6623, 2024. 2, 3, 4, 5, 8", + "[50] Tianwei Yin, Qiang Zhang, Richard Zhang, William T Freeman, Fredo Durand, Eli Shechtman, and Xun Huang. From slow bidirectional to fast causal video generators. arXiv preprint arXiv:2412.07772, 2024. 2, 3, 4, 6, 8" + ], + "bbox": [ + 516, + 90, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26104", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and Bill Freeman. Improved distribution matching distillation for fast image synthesis. Advances in Neural Information Processing Systems, 37: 47455-47487, 2025. 2, 3, 4, 5, 6, 8", + "[52] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024. 1, 2, 3, 4, 5, 6, 7, 8", + "[53] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + ], + "bbox": [ + 91, + 90, + 480, + 301 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "26105", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/FVGen_ Accelerating Novel-View Synthesis with Adversarial Video Diffusion Distillation/381cfeb4-88c2-44b1-a694-860cc7f57161_model.json b/2025/FVGen_ Accelerating Novel-View Synthesis with Adversarial Video Diffusion Distillation/381cfeb4-88c2-44b1-a694-860cc7f57161_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a45bc9899d75da4cab1e52ecea2a6ecb5e66534d --- /dev/null +++ b/2025/FVGen_ Accelerating Novel-View Synthesis with Adversarial Video Diffusion Distillation/381cfeb4-88c2-44b1-a694-860cc7f57161_model.json @@ -0,0 +1,2422 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.112, + 0.13, + 0.887, + 0.175 + ], + "angle": 0, + "content": "FVGen: Accelerating Novel-View Synthesis with Adversarial Video Diffusion Distillation" + }, + { + "type": "text", + "bbox": [ + 0.194, + 0.203, + 0.805, + 0.258 + ], + "angle": 0, + "content": "Wenbin Teng\\(^{1,2}\\), Gonglin Chen\\(^{1,2}\\), Haiwei Chen\\(^{1,2}\\), Yajie Zhao\\(^{1,2*}\\) \n\\(^{1}\\)Institute for Creative Technologies \n\\(^{2}\\)University of Southern California \n{wenbinte, gonglinc}@usc.edu, {chenh, zhao}@ict.usc.edu," + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.296, + 0.865, + 0.541 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.558, + 0.908, + 0.613 + ], + "angle": 0, + "content": "Figure 1. We present FVGen, a method of fast novel-view synthesis for 3D scene reconstruction with sparse inputs. Recent method [52] unleash the generation capabilities of video diffusion models conditioned on a prior 3D cue for dense view creation, but always suffers from a long sampling time. We propose a novel framework that can significantly reduce the sampling time and maintain the generation quality." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.615, + 0.327, + 0.63 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.648, + 0.486, + 0.875 + ], + "angle": 0, + "content": "Recent progress in 3D reconstruction has enabled realistic 3D models from dense image captures, yet challenges persist with sparse views, often leading to artifacts in unseen areas. Recent works leverage Video Diffusion Models (VDMs) to generate dense observations, filling the gaps when only sparse views are available for 3D reconstruction tasks. A significant limitation of these methods is their slow sampling speed when using VDMs. In this paper, we present FVGen, a novel framework that addresses this challenge by enabling fast novel view synthesis using VDMs in as few as four sampling steps. We propose a novel video diffusion model distillation method that distills a multi-step denoising teacher model into a few-step denoising student model using Generative Adversarial Networks (GANs) and softened reverse KL-divergence minimization. Extensive ex" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.616, + 0.909, + 0.753 + ], + "angle": 0, + "content": "periments on real-world datasets show that, compared to previous works, our framework generates the same number of novel views with similar (or even better) visual quality while reducing sampling time by more than \\(90\\%\\). FVGen significantly improves time efficiency for downstream reconstruction tasks, particularly when working with sparse input views (more than 2) where pre-trained VDMs need to be run multiple times to achieve better spatial coverage. Project Page: https://wbteng9526.github.io/fvgen/" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.769, + 0.646, + 0.785 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Reconstructing 3D scenes from 2D images remains a pivotal research area at the intersection of computer vision and graphics. This field has extensive social impact, with broad applications in autonomous navigation, augmented reality, and more. With the recent development of neural rendering techniques such as NeRF [29] and 3DGS [16], creating high-quality 3D assets from 2D images has become more" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.237, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26095" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.166 + ], + "angle": 0, + "content": "accessible. However, the success of 3D reconstruction depends on dense observations, which are often difficult to obtain in real-world scenarios. This limitation makes it challenging to apply these methods in many practical applications." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.173, + 0.483, + 0.564 + ], + "angle": 0, + "content": "Introducing generative models into 3D reconstruction can fill gaps in missing views when only sparse input views are available. However, novel view synthesis from sparse views, especially in the settings with less than 2 views, is inherently an ill-posed problem. A bulk of previous works [7, 8, 30, 43, 52] have found that priors learned from large image and video diffusion models [10, 11] are rich enough to inpaint detailed and reasonable information in the many under-observed areas in the reconstructed 3D representations. We are primarily interested in the use of video diffusion models (VDM) to perform novel view synthesis. What bridges VDM to novel view synthesis is its ability to synthesize continuous views following a camera trajectories between the observed views. The spatio-temporal consistency in these smooth video trajectories, as demonstrated in several works [6, 7, 25, 44, 52], effectively addresses the challenging \"Janus Problem\" that typically limits generation quality in other multi-view settings. However, VDM is a computationally expensive method because of the iterative sampling process that characterized all denoising diffusion methods. The long generation time makes VDM-based generation methods unsuitable for certain real-world downstream tasks, such as dynamic 3D reconstruction, where new consistent views must be generated continuously. This limitation also affects large-scale 3D reconstruction that requires synthesizing a significant number of novel views." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.57, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this paper, we show that VDM-based novel view synthesis can be significantly accelerated for free. Without any compromise on the generation quality, we speed up diffusion-based synthesis of novel view by a factor of \\(90\\%\\) with a 4-step student model trained under a GAN objective [23]. Acceleration of diffusion models has been actively studied in the past, but previous techniques may not be well suited to our task due to two reasons: First, we need a method specifically for accelerating video diffusion models, rather than image diffusion models. Second, the method must generalize effectively when trained on multi-view video datasets, which contain significantly less data than general video datasets. To accelerate a diffusion model, recent methods [22, 28, 34, 46, 49] propose minimizing the distributional difference between a few-step student generation model and the original multi-step teacher model. However, these methods are restricted to images with specific domains. The most related work to us is [50], which generates more video frames with faster speed by optimizing a distribution matching distillation (DMD) loss. However, this method requires an effective initialization of the student model. This is completed by generating a noise-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.257 + ], + "angle": 0, + "content": "latent pair through an ODE solver [36] and training the student model by optimizing a regression loss [49, 50]. The generation process usually takes a very long time and the student's ability is limited by the teacher's generation quality. More importantly, we have observed that DMD optimizations tend to be unstable and have the tendency towards mode collapse, when trained under multi-view video data. This may be due to the inherent concept of DMD loss optimization being the reverse KL-divergence minimization, which tends to be mode-seeking and zeroing out the modes of teacher's that are not relatively dominant." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.259, + 0.905, + 0.529 + ], + "angle": 0, + "content": "To solve the previous problems with VDM acceleration, we propose a novel framework, named FVGen, that performs novel views generation with a fast video diffusion generation with as few as 4 steps. First of all, different from previous GAN-based distillation methods [34, 51], we find it beneficial to initial the few-step student model by training a GAN objective, where the student model is regarded as a generator of fake samples, and, instead of training a separate discriminator as in previous works, we leverage a pre-trained teacher model as the discriminator. Second, we propose the soften reverse KL-divergence to solve the unstable optimization of DMD loss. The soften reverse KL-divergence maintains similar mode-seeking behavior as reverse KL-divergence but with more robust distillation: it prevents the student from ignoring entire space of teacher's distribution, thereby preserving more of teacher's knowledge, which we find particularly beneficial in the context of limited data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.532, + 0.905, + 0.667 + ], + "angle": 0, + "content": "Through effective and robust distillation, FVGen is capable of generating video sequences with comparable or better visual quality as ViewCrafter [52] but with significantly less time. FVGen is trained with Dl3DV-10K [24] dataset and extensive experiments on real world dataset including MipNeRF360 [1] and Tanks-and-Temple [18] demonstrate that our method achieves overall better performance than other state-of-the-art (SOTA) novel view synthesis methods and diffusion distillation methods." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.685, + 0.665, + 0.701 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.712, + 0.838, + 0.727 + ], + "angle": 0, + "content": "2.1. Diffusion-based Novel View Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.905, + 0.901 + ], + "angle": 0, + "content": "NeRF [29] and 3DGS [16] excel at novel view synthesis with dense input views. However, reconstructing scenes from sparse views requires additional priors. Diffusion models [10, 32] can generate realistic pseudo input views from sparse original inputs. Several works use Score-Distillation Sampling [21, 30, 39, 42] to create 3D objects from text prompts or single images by distilling latent diffusion model [32] knowledge into 3D representations. While NeRF, 3DGS, and SDS typically require time-consuming per-scene optimization, view-dependent diffusion models [8, 19, 20, 35] offer an alternative. These mod" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26096" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.092, + 0.511, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.094, + 0.885, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.232, + 0.908, + 0.276 + ], + "angle": 0, + "content": "Figure 2. Overview of student initialization. We initialize our student model by training a GAN objective. The student model \\(G_{\\theta}\\), initialized with the weights of teacher model, uses few-step denoising to generate fake videos to fool the discriminator \\(D\\). The fixed teacher model, together with a trainable discriminator, are optimized to differentiate between fake and real videos." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.289, + 0.482, + 0.334 + ], + "angle": 0, + "content": "els directly generate 3D-consistent novel views conditioned on input images but directly lifting 2D images to 3D input therefore face challenges when scaling to larger scenes." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.336, + 0.485, + 0.56 + ], + "angle": 0, + "content": "An alternative approach leverages video diffusion models [6, 7, 25, 38, 52] for multi-view image generation. These methods fine-tune pre-trained latent video diffusion models (LVDM) [2, 4, 12, 44, 48] using multi-view images with both 2D and 3D guidance. The 2D guidance typically involves semantic features like CLIP [14] embeddings, while 3D guidance comes from sparse input cues. For instance, [52] uses DUSt3R [40] to build an initial point cloud and guides generation by combining point cloud renders with camera poses. Similarly, [6] conditions generation on feedforward 3DGS renders, enabling 360-degree scene generation with longer frame sequences through multi-step refinement. However, these methods suffer from speed issues caused by video diffusion sampling when generating novel views." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.575, + 0.288, + 0.59 + ], + "angle": 0, + "content": "2.2. Diffusion Distillation" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.599, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Diffusion models typically require numerous denoising steps to generate high-quality samples, which makes them computationally intensive. To accelerate generation, distillation methods train a student model that mimics the teacher diffusion model's behavior while using fewer sampling steps. Using fewer sampling steps reduces quality, so existing works incorporate adversarial training [3, 15, 22, 27, 33, 34, 46] to enhance the student model's output. However, simple GAN training cannot fully bridge the gap between student and teacher distributions, causing the student's generated output to remain noticeably different from the original diffusion model. To address this problem, Distribution Matching Distillation (DMD) [49, 51] optimizes the reverse KL-divergence. For tractable optimization, [49, 51] leverages the gradient of DMD loss, which represents the difference between student and teacher score functions. For fast video generation, [50] extends DMD to video diffusion models like CogVideoX [48]. While these methods work well for general image and video generation tasks, they suffer from mode collapse and training insta" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.289, + 0.907, + 0.318 + ], + "angle": 0, + "content": "bility when applied to view synthesis on relatively small datasets like multi-view video datasets." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.336, + 0.652, + 0.351 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.361, + 0.724, + 0.376 + ], + "angle": 0, + "content": "3.1. Video Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.384, + 0.907, + 0.522 + ], + "angle": 0, + "content": "A video diffusion model is a generative model designed to synthesize realistic video sequences by learning the complex temporal and spatial patterns within video data. Based on the diffusion process, these models iteratively apply noise to video frames and then learn to reverse this process to generate new sequences. The forward diffusion process progressively adds Gaussian noise to the input clean sample \\( \\pmb{x}_0 \\sim p(\\pmb{x}) \\) over \\( T \\) timesteps, leading to highly noisy data \\( \\pmb{x}_T \\sim \\mathcal{N}(0, I) \\). This process is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.538, + 0.907, + 0.556 + ], + "angle": 0, + "content": "\\[\nq \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {t - 1}\\right) = \\mathcal {N} \\left(\\boldsymbol {x} _ {t}; \\sqrt {\\alpha_ {t}} \\boldsymbol {x} _ {t - 1}, (1 - \\alpha_ {t}) \\boldsymbol {I}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.907, + 0.645 + ], + "angle": 0, + "content": "where \\(\\alpha_{t}\\) and \\(\\pmb{x}_t\\) are the noise strength and noisy data at timestep \\(t\\). The model then learns a denoising function, often parameterized by a neural network \\(\\epsilon_{\\theta}(\\pmb{x}_t,t)\\), to reverse this process and generate realistic video frames, where the optimization function is defined by the MSE loss:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.658, + 0.905, + 0.698 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d i f f}} (\\theta) = \\mathbb {E} _ {t \\sim \\mathcal {U} (0, 1), \\epsilon \\sim \\mathcal {N} (0, I), \\boldsymbol {x} \\sim p, c} \\| \\epsilon - \\epsilon_ {\\theta} (\\boldsymbol {x} _ {t}, t, c) \\| ^ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.7, + 0.907, + 0.821 + ], + "angle": 0, + "content": "where \\(c\\) is the condition embedding of diffusion model which is usually represented as a text or image prompt. Instead of generating sequences at full resolution, Latent Video Diffusion Model (LVDM) first encodes video data \\(\\pmb{x} \\in \\mathbb{R}^{F \\times C \\times H \\times W}\\) into latent space using a pre-trained VAE [17] encoder: \\(\\pmb{z} = \\mathcal{E}(\\pmb{x})\\), where \\(\\pmb{z} \\in \\mathbb{R}^{F \\times C \\times h \\times w}\\). The compression will mitigate computational complexity while maintaining video generation quality." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.832, + 0.812, + 0.849 + ], + "angle": 0, + "content": "3.2. Distribution Matching Distillation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Distribution matching distillation adopts the idea of variational score distillation [42] that is optimized to match the distribution of student and teacher model by minimizing the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26097" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.087, + 0.891, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.276, + 0.908, + 0.307 + ], + "angle": 0, + "content": "Figure 3. Overview of Distribution Matching Distillation (DMD). The gradient is the difference between the teacher score and fake score w.r.t. the student model output. We apply soften reverse KL-divergence for stable training and avoiding mode-collapse." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.484, + 0.35 + ], + "angle": 0, + "content": "reverse KL-divergence. Given a student model \\(G\\) parameterized by \\(\\theta\\), the gradient of reverse KL-divergence is" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.358, + 0.483, + 0.463 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\nabla \\mathcal {L} _ {\\mathrm {D M D}} = \\mathbb {E} _ {t} \\left(\\nabla_ {\\theta} \\mathbf {D} _ {\\text {R e v e r s e - K L}} \\left(p _ {\\text {f a k e}, t} \\| p _ {\\text {r e a l}, t}\\right)\\right) \\\\ = - \\mathbb {E} _ {t} \\left(\\left[ s _ {\\text {r e a l}} \\left(F \\left(G _ {\\theta} (\\boldsymbol {z}), t\\right), t\\right) \\right. \\right. \\\\ \\left. - s _ {\\text {f a k e}} \\left(F \\left(G _ {\\theta} (\\boldsymbol {z}), t\\right), t\\right) \\frac {d G _ {\\theta} (\\boldsymbol {z})}{d \\theta} \\right]\\left. \\right), \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.471, + 0.484, + 0.547 + ], + "angle": 0, + "content": "where \\( p_{\\mathrm{real}} \\) and \\( p_{\\mathrm{fake}} \\) are the real and fake distribution. \\( z \\sim \\mathcal{N}(0, I) \\) is the initial random Gaussian noise. \\( F \\) is the diffusion forward function determined by Eq. (1). \\( s_{\\mathrm{real}} \\) and \\( s_{\\mathrm{fake}} \\) are the real and fake score function and defined as follows according to [37]:" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.55, + 0.483, + 0.582 + ], + "angle": 0, + "content": "\\[\ns \\left(\\boldsymbol {x} _ {t}, t\\right) = \\nabla_ {\\boldsymbol {x} _ {t}} \\log p _ {t} \\left(\\boldsymbol {x} _ {t}\\right) = - \\frac {\\boldsymbol {x} _ {t} - \\alpha_ {t} \\mu \\left(\\boldsymbol {x} _ {t} , t\\right)}{\\sigma_ {t} ^ {2}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.591, + 0.484, + 0.682 + ], + "angle": 0, + "content": "where \\( p_t \\) is the distribution of the noisy sample \\( x \\). \\( \\mu \\) is the clean sample prediction of the diffusion model. While optimizing student model \\( G_{\\theta} \\) with gradient descent formulated in Eq. (3), the gradients of \\( \\mu_{\\mathrm{real}} \\) are stopped and \\( \\mu_{\\mathrm{fake}} \\) is dynamically optimized with the output of \\( G_{\\theta} \\) with regular diffusion loss similar to Eq. (2)" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.694, + 0.182, + 0.71 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We propose Fast Video Generation (FVGen) that distills the generation capability of a multi-step teacher 3D-VDM into a few-step student 3D-VDM. The teacher 3D-VDM used in our experiments is the novel-view synthesis model ViewCrafter [52]. From a set of observed images \\( I^{\\mathrm{obs}} \\) and observed cameras \\( \\pi^{\\mathrm{obs}} \\), ViewCrafter learns a conditional distribution \\( I^{\\mathrm{tgt}} \\sim p(I^{\\mathrm{tgt}}|I^{\\mathrm{obs}},\\pi^{\\mathrm{obs}},\\pi^{\\mathrm{tgt}}) \\), where \\( I^{\\mathrm{tgt}} = \\{I_i'\\}_{i=1}^L \\) is a sequence of \\( L \\) images corresponding to the target views \\( \\pi^{\\mathrm{tgt}} \\). The student model FVGen is trained to sample from the same conditional distribution with a much faster inference speed. The architectures of FVGen are shown in Figure 2 and 3. For notation simplicity purpose," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.318, + 0.905, + 0.363 + ], + "angle": 0, + "content": "we ignore including \\( I^{\\mathrm{obs}} \\), \\( \\pi^{\\mathrm{obs}} \\), \\( \\pi^{\\mathrm{tgt}} \\) in the future formula as they are all the conditional inputs for both teacher and student diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.906, + 0.469 + ], + "angle": 0, + "content": "We first initialize the parameters of student model by training with a generative adversarial network (GAN) objective, leveraging the teacher model as discriminator (Section 4.1). After that, we continue to train the student model by optimizing DMD loss and diffusion loss to minimize the soften reverse KL-divergence between the distribution of student and teacher model (Section 4.2)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.479, + 0.868, + 0.495 + ], + "angle": 0, + "content": "4.1. Student Initialization with GAN Training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.5, + 0.907, + 0.742 + ], + "angle": 0, + "content": "Directly training student model with DMD (Eq. 3) will cause training collapse as it is difficult to naively distill multi-step denoising into few-step. [49, 50] proposed to create noise-latent pairs by an ODE solver [36] and minimize a regression loss with student model for initialization. However, we found that it is very time consuming to generate large amount of video data pairs and the student model will also be restricted by the teacher's limitations. Inspired by [23, 51], we propose to initialize the student model by training a GAN objective with the real images. Specifically, the student model \\((G_{\\theta})\\) generates fake video samples that fool the teacher model who serves as a discriminator (denoted as \\(D\\)) by minimizing the generator loss \\(\\mathcal{L}_G\\). The discriminator \\(D\\) classifies real samples from generated fake samples by maximizing \\(-\\mathcal{L}_D\\). The min-max game adversarial optimization is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.746, + 0.905, + 0.801 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal{L}_{D} = \\mathbb{E}_{\\substack{\\boldsymbol {x}\\sim p_{\\text{real}}\\\\ t\\sim \\mathcal{U}[0,T]}}[\\log f\\left(D(F(\\boldsymbol {x},t))\\right)] \\\\ - \\mathbb {E} _ {\\substack {\\boldsymbol {z} \\sim \\mathcal {N} (0, I) \\\\ t \\sim \\mathcal {U} [ 0, T ]}} \\left[ \\log f \\left(D \\left(F \\left(G _ {\\theta} (\\boldsymbol {z}), t\\right)\\right)\\right) \\right], \\tag{5} \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.804, + 0.858, + 0.833 + ], + "angle": 0, + "content": "\\[\n\\mathcal{L}_{G} = \\operatorname{\\mathbb{E}}_{\\substack{\\boldsymbol {z}\\sim \\mathcal{N}(0,\\boldsymbol {I})\\\\ t\\sim \\mathcal{U}[0,T]}}[\\log f\\left(D(F(G_{\\theta}(\\boldsymbol {z}),t))\\right)],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\( t \\) is the time-step sampling and \\( F \\) is the forward noise-injection model similar to Eq. (3). Discriminator \\( D \\) retrieves the output feature map of the middle block of diffusion model UNet and passes it into the classifier \\( f \\). To" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26098" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.289 + ], + "angle": 0, + "content": "fit with video data, \\( f \\) is composed of a 3D Convolutional Neural Network that compresses the feature map into classification logits. Here, we use the original sample \\( \\pmb{x} \\) as the input to the discriminator instead of the generated sample \\( \\pmb{x} \\) from teacher model such that the student model would not be restricted by the generation capability of teacher model. For the purpose of training stability, recent methods [6, 52] predict the velocity field \\( \\pmb{v}_{\\theta} \\), and we convert it into the sample prediction with \\( \\pmb{x}_0 = \\sqrt{\\alpha_t\\pmb{x}_t} -\\sqrt{1 - \\alpha_t}\\pmb{v}_{\\theta} \\). Although the generated video samples are still very blurry, the GAN training scheme provides an effective initialization of the student model, which paves the way for more effective training with distribution matching loss." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.295, + 0.484, + 0.327 + ], + "angle": 0, + "content": "4.2. Distribution Matching with Soften Reverse KL-Divergence Minimization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.332, + 0.484, + 0.572 + ], + "angle": 0, + "content": "As discussed in Section 3.2, the core concept of distribution matching distillation (DMD) [49, 51] is to minimize the reverse KL-divergence between student and teacher distributions. While reverse KL is popular for its mode-seeking focus, it has notable limitations in the context of generative modeling, its tendency toward mode collapse. Since reverse KL rewards the student for zeroing out probability in any region where the teacher's density is low, the student can effectively ignore parts where the teacher's distribution are not dominant. If the teacher's distribution \\( p_{\\mathrm{real}}(\\pmb{x}) \\) has multiple modes of varying height, minimizing ReverseKL \\( (p_{\\mathrm{fake}} \\| p_{\\mathrm{real}}) \\) may lead \\( p_{\\mathrm{fake}} \\) to concentrate on a subset of those modes and assign negligible mass to others. Therefore, the exclusive nature of reverse KL-divergence makes it prone to mode dropping and a lack of coverage of the full distribution." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.574, + 0.484, + 0.68 + ], + "angle": 0, + "content": "To address the shortcomings of pure reverse KL, we propose to leverage a Soften reverse KL-divergence (Soften RKL) [9, 47]. Soften RKL refers to a modified divergence that retains the mode-seeking bias of reverse KL but softens its exclusion other relatively lower-probability regions, thereby mitigating mode collapse. Compared to reverse KL formulated in Eq. (3), the Soften RKL can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.686, + 0.482, + 0.768 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {D} _ {\\text {S o f t e n - R K L}} \\left(p _ {\\text {f a k e}, t} \\| p _ {\\text {r e a l}, t}\\right) = \\mathbf {D} _ {K L} \\left(\\frac {1}{2} p _ {\\text {r e a l}} + \\frac {1}{2} p _ {\\text {f a k e}} \\| p _ {\\text {r e a l}}\\right), \\\\ = p _ {\\text {f a k e}} (r + 1) \\log \\left(\\frac {1}{2} + \\frac {1}{2 r}\\right), \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.775, + 0.484, + 0.866 + ], + "angle": 0, + "content": "where \\(r = r(\\pmb{x},t) \\coloneqq p_{\\mathrm{real},t}(\\pmb{x}) / p_{\\mathrm{fake},t}(\\pmb{x})\\) is the density ratio. Here \\(\\pmb{x}\\) denotes the noise-injection version of the output of student model formulated as \\(F(G_{\\theta}(\\pmb{z}),t)\\) with \\(t \\sim \\mathcal{U}(0,T)\\). Thanks to our optimized GAN objective in Section 4.1, the density ratio could be approximated by the following:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.871, + 0.484, + 0.906 + ], + "angle": 0, + "content": "\\[\nr (\\boldsymbol {x}, t) = \\frac {p _ {\\text {r e a l} , t} (\\boldsymbol {x})}{p _ {\\text {f a k e} , t} (\\boldsymbol {x})} \\approx \\frac {f (D (\\boldsymbol {x} , t))}{1 - f (D (\\boldsymbol {x} , t))}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.304 + ], + "angle": 0, + "content": "In summary, the well-optimized GAN objective not only provides an efficient weight initialization of the student model, but also provides a direct estimate of the density ratio. Here, instead of directly comparing \\( p_{\\mathrm{fake}} \\) and \\( p_{\\mathrm{real}} \\) as in \\( \\mathbf{D}_{KL}(p_{\\mathrm{fake}}||p_{\\mathrm{real}}) \\), we compare an even mixture of \\( p_{\\mathrm{real}} \\) and \\( p_{\\mathrm{fake}} \\) to \\( p_{\\mathrm{real}} \\). Soften RKL maintains the mode-seeking property of reverse KL, while penalizes more on mode dropping. This criterion yields a more robust distillation since it prevents the student from ignoring the entire teacher distribution, thereby preserving more of teacher's knowledge. Through experiments, we find that the soften RKL yields more stable training and more realistic generations. Computing the divergence directly is generally intractable, the gradient with respect to \\( \\theta \\) is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.315, + 0.905, + 0.377 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\nabla \\mathcal {L} _ {\\mathrm {D M D}} ^ {\\text {S o f t t e n - R K L}} = \\tag {8} \\\\ - \\mathbb {E} _ {t} \\left(\\frac {1}{r (\\boldsymbol {x} , t)} \\left[ s _ {\\text {r e a l}} (\\boldsymbol {x}, t) - s _ {\\text {f a k e}} (\\boldsymbol {x}, t) \\right] \\frac {d G _ {\\theta} (\\boldsymbol {z})}{d \\theta}\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.389, + 0.907, + 0.482 + ], + "angle": 0, + "content": "where \\(\\pmb{x} = F(G_{\\theta}(z),t)\\) as previously defined. Apart from the soften RKL divergence optimization, we dynamically update the fake diffusion model \\(\\mu_{\\mathrm{fake}}\\) to adjust to change of student distribution in accordance with the original setting of DMD [49]. The loss term is identical to Eq. (2). We summarize the full training procedure in Alg. 1" + }, + { + "type": "code_caption", + "bbox": [ + 0.515, + 0.496, + 0.785, + 0.512 + ], + "angle": 0, + "content": "Algorithm 1 FVGen Training Procedure" + }, + { + "type": "code", + "bbox": [ + 0.516, + 0.516, + 0.906, + 0.56 + ], + "angle": 0, + "content": "Require: Few-step timesteps \\(\\mathcal{T} = \\{0, t_1, t_2, \\dots, t_Q\\}\\), pretrained teacher model \\(\\mu_{\\mathrm{real}}\\), discriminator classifier \\(D\\) dataset \\(\\mathcal{D}\\)." + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.563, + 0.797, + 0.577 + ], + "angle": 0, + "content": "1: Initialize student model \\( G_{\\theta} \\) with \\( \\mu_{\\mathrm{real}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.578, + 0.807, + 0.592 + ], + "angle": 0, + "content": "2: Initialize fake score function with \\(\\mu_{\\mathrm{fake}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.594, + 0.66, + 0.607 + ], + "angle": 0, + "content": "3: while training do" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.608, + 0.814, + 0.623 + ], + "angle": 0, + "content": "4: Sample a video from dataset \\( \\pmb{x}_0 \\sim \\mathcal{D} \\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.623, + 0.905, + 0.652 + ], + "angle": 0, + "content": "5: Add noise with timestep \\(t \\sim \\mathcal{T}\\): \\(\\pmb{x}_t = \\sqrt{\\alpha_t} \\pmb{x}_0 + \\sqrt{1 - \\alpha_t} \\pmb{\\epsilon}, \\quad \\pmb{\\epsilon} \\sim \\mathcal{N}(0, \\pmb{I})\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.653, + 0.812, + 0.667 + ], + "angle": 0, + "content": "6: Predict with student: \\(\\hat{\\pmb{x}}_0 = G_\\theta (\\pmb {x}_t,t)\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.668, + 0.904, + 0.697 + ], + "angle": 0, + "content": "7: Add noise \\(\\epsilon' \\sim \\mathcal{N}(0, I)\\) with timestep \\(\\tau \\sim \\mathcal{U}(0, T)\\) to real and fake sample:" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.698, + 0.754, + 0.714 + ], + "angle": 0, + "content": "8: \\(\\hat{\\pmb{x}}_{\\tau} = \\sqrt{\\alpha_{\\tau}}\\hat{\\pmb{x}}_0 + \\sqrt{1 - \\alpha_{\\tau}}\\pmb{\\epsilon}'\\)" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.715, + 0.753, + 0.729 + ], + "angle": 0, + "content": "9: \\(\\pmb{x}_{\\tau} = \\sqrt{\\alpha_{\\tau}}\\pmb{x}_0 + \\sqrt{1 - \\alpha_{\\tau}}\\pmb{\\epsilon}'\\)" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.729, + 0.77, + 0.743 + ], + "angle": 0, + "content": "10: Update \\(G_{\\theta}\\) and \\(D\\) with Eq. (5)." + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.745, + 0.613, + 0.757 + ], + "angle": 0, + "content": "11: end while" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.759, + 0.721, + 0.773 + ], + "angle": 0, + "content": "12: Output trained student \\( G_{\\theta} \\)" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.775, + 0.66, + 0.788 + ], + "angle": 0, + "content": "13: while training do" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.789, + 0.68, + 0.803 + ], + "angle": 0, + "content": "14: Repeat step 4 - 8" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.804, + 0.807, + 0.818 + ], + "angle": 0, + "content": "15: Update \\(G_{\\theta}\\) with DMD Loss: Eq. (8)" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.819, + 0.899, + 0.834 + ], + "angle": 0, + "content": "16: Add noise \\(\\pmb{\\epsilon}'' \\sim \\mathcal{N}(0, \\pmb{I})\\) with \\(t_1 \\sim \\mathcal{U}(0, T)\\) to \\(\\hat{\\pmb{x}}_0\\):" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.835, + 0.77, + 0.85 + ], + "angle": 0, + "content": "17: \\(\\hat{\\pmb{x}}_{t_1} = \\sqrt{\\alpha_{t_1}}\\hat{\\pmb{x}}_0 + \\sqrt{1 - \\alpha_{t_1}}\\pmb{\\epsilon}^{\\prime \\prime}\\)" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.85, + 0.836, + 0.865 + ], + "angle": 0, + "content": "18: Update \\(\\mu_{\\mathrm{fake}}\\) with denoising loss: Eq. (2)" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.865, + 0.613, + 0.877 + ], + "angle": 0, + "content": "19: end while" + }, + { + "type": "list", + "bbox": [ + 0.522, + 0.563, + 0.905, + 0.877 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26099" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.093, + 0.905, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.413, + 0.908, + 0.444 + ], + "angle": 0, + "content": "Figure 4. Qualitative Results. We compare FVGen with SOTA novel view synthesis method on MipNeRF360 [1] and Tanks-and-Temples [18] datasets." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.455, + 0.224, + 0.472 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.48, + 0.267, + 0.496 + ], + "angle": 0, + "content": "5.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.502, + 0.484, + 0.714 + ], + "angle": 0, + "content": "Implementation Details. All of our student model, teacher model and fake score function is initialized with the parameters of the sparse model of ViewCrafter [52]. We first initialize the student model by training GAN objective. The parameters of teacher model is fixed, and the student model and GAN classifier are trained for 4000 iterations with a two-scale update rule inspired by [51]. Next, we optimize DMD loss by continuously training student model and the fake score function for another 5000 iterations with the same two-scale update rule. The whole pipeline is trained on 8 NVIDIA H100 with batch size 4. The training last about 1 day. We utilize AdamW optimizer [26] and set learning rates of all optimizers to be \\(5 \\times 10^{-5}\\). The model is trained on images with a resolution of \\(512 \\times 320\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Datasets. FVGen is trained with DL3DV-10K [24], a real-world, scene-level video dataset that includes more than 10K long video clips. Similar to [52], we construct point cloud and ground truth video pairs using DL3DV dataset. We randomly sample short video clips with a random frame stride less than 5. The first frame and last frame are used to construct the prior point cloud with DUSt3R [40]. We use PyTorch3d [31], an efficient renderer to construct the point cloud renders with the camera poses of intermediate frames. Through this process, we create 20,000 training data pairs. We validate our method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.456, + 0.905, + 0.487 + ], + "angle": 0, + "content": "on two public datasets Tanks-and-Temple [18] and MipN-eRF360 [1]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.505, + 0.907, + 0.671 + ], + "angle": 0, + "content": "Baseline Methods and Evaluation Metrics. To evaluate our proposed method, we compare our method with several state-of-the-art methods: ViewCrafter [52], MVsplat360 [6]. We also compare our video diffusion distillation method with two state-of-the-art methods: CausVid [50] and DMD2 [51]. We use PSNR [13], SSIM [41], LPIPS [53] to measure reconstruction quality. In addition, we also reported the distribution metric, i.e. Frechet Inception Distance (FID), which compares the distribution of generated views and ground truth views. Apart from the visual quality comparison," + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.68, + 0.813, + 0.696 + ], + "angle": 0, + "content": "5.2. Novel View Synthesis Comparison" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.701, + 0.907, + 0.807 + ], + "angle": 0, + "content": "Similar to the evaluation setting of MVSplat360 [6], for each scene of both MipNeRF360 [1] and Tanks-and-Temples [18], we select 5 views that are far from each other but could cover the whole scene. We use our method and baseline method to generate 56 views sampled from the natural camera trajectory (14 views between each 2 input views)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Qualitative Results The qualitative comparisons with SOTA novel view synthesis methods are shown in Figure 4. DepthSplat is a feed-forward generalizable Gaussian splatting method. When there are relatively low overlap between input frames, DepthSplat would exhibit ob" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26100" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.097, + 0.19, + 0.111 + ], + "angle": 0, + "content": "Input Views" + }, + { + "type": "image_caption", + "bbox": [ + 0.208, + 0.097, + 0.361, + 0.111 + ], + "angle": 0, + "content": "ViewCrafter + DMD2" + }, + { + "type": "image_caption", + "bbox": [ + 0.375, + 0.097, + 0.54, + 0.111 + ], + "angle": 0, + "content": "ViewCrafter + CausVid" + }, + { + "type": "image_caption", + "bbox": [ + 0.617, + 0.097, + 0.652, + 0.11 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.765, + 0.097, + 0.86, + 0.11 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.113, + 0.196, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.196, + 0.113, + 0.371, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.113, + 0.548, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.548, + 0.113, + 0.723, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.723, + 0.113, + 0.905, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.199, + 0.196, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.196, + 0.199, + 0.372, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.372, + 0.199, + 0.548, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.548, + 0.199, + 0.723, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.723, + 0.199, + 0.905, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.297, + 0.196, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.196, + 0.297, + 0.372, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.372, + 0.297, + 0.548, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.548, + 0.297, + 0.723, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.723, + 0.297, + 0.905, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.395, + 0.196, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.196, + 0.395, + 0.372, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.372, + 0.395, + 0.548, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.548, + 0.395, + 0.723, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.723, + 0.395, + 0.905, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.492, + 0.907, + 0.521 + ], + "angle": 0, + "content": "Figure 5. Qualitative Results. We compare FVGen with baseline diffusion distillation method on MipNeRF360 [1] and Tanks-and-Temples [18] datasets." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.534, + 0.486, + 0.791 + ], + "angle": 0, + "content": "vious artifacts due to inconsistent depth scale and the issue of floating Gaussians. In comparison, MVSplat360 [6] uses VDM to refine the Gaussian splattering renders and remove the floaters. However, the current MVSplat360 only supports training with low resolution inputs and the current results present significant blurriness. Compared to DepthSplat and MVSplat360, ViewCrafter [52] generates photo-realistic novel views with higher resolution. However, ViewCrafter is limited by the generation speed. Given generating 16 total frames, the average inference time is 13.2 seconds. In comparison, our method achieves a 10x speedup, generating the same number of frames in approximately 1 second and achieves similar visual results. The qualitative results prove that our method is able to distill the multi-step sampling into a few-step sampling process and achieve results that are similar to or even better than teacher ViewCrafter model with the student model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Quantitative Analysis Table 1 presents quantitative comparisons on both datasets of FVGen and other novel view synthesis methods. Except for ViewCrafter [52], our method significantly surpasses other SOTA methods in terms of both perceptual and distribution metrics. Compared with ViewCrafter, our method yields similar quality" + }, + { + "type": "table", + "bbox": [ + 0.532, + 0.532, + 0.892, + 0.68 + ], + "angle": 0, + "content": "
PSNR ↑SSIM ↑LPIPS ↓FID ↓Time ↓
mip-360DepthSplat11.230.2130.71532.454.2
MVSplat36012.280.2850.68225.6987.2
ViewCrafter16.350.3460.43316.2866.3
Ours16.280.3520.42917.445.1
TNTDepthSplat12.430.2630.67735.884.3
MVSplat36014.180.3010.53225.2387.3
ViewCrafter18.690.4020.20823.9465.9
Ours18.720.4110.21023.645.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.69, + 0.907, + 0.759 + ], + "angle": 0, + "content": "Table 1. Quantitative Results. We report four quantitative metrics for image quality comparison on 2 separate datasets: Mip-NeRF360 [1] (denoted as mip-360 in the table) and Tanks-and-Temple [18] (denoted as TNT in the table). We highlight the best results in red and second-best in yellow." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.764, + 0.907, + 0.809 + ], + "angle": 0, + "content": "but with more than \\(90\\%\\) faster in generation speed. The quantitative results also highlights the effectiveness of our proposed method." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.818, + 0.81, + 0.835 + ], + "angle": 0, + "content": "5.3. Diffusion Distillation Comparison" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We also compare FVGen with SOTA diffusion distillation methods. Similar to the 2-view evaluation setting of DepthSplat [45] and MVSplat [5], we select 2 views that are far from each other in both MipNeRF360 [1] and Tanks-and-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26101" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.138 + ], + "angle": 0, + "content": "Temples [18] dataset. We use our method and baseline method to generate 16 views samples from the natural camera trajectory (14 intermediate views and 2 input views)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.166, + 0.484, + 0.379 + ], + "angle": 0, + "content": "Qualitative Results First we perform DMD2 [51] on ViewCrafter for few-step distillation. DMD2 applies end-to-end training of student model, GAN discriminator and fake score function. We find that this training scheme is not stable and present significant variance so that it is hard to converge, leading to blurry results as visually depicted in Figure 5. In addition, we compare our method with a recent work, Caus-Vid [50], on ViewCrafter distillation. CausVid [50] creates a small dataset with ODE solver and initialize the student model by optimizing a regression loss. Caus-Vid also leverages DMD [49] for 4-step video generation. However, the qualitative results indicate that our GAN object training scheme provides better initialization compared to the regression training scheme of ODE pairs." + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.395, + 0.465, + 0.516 + ], + "angle": 0, + "content": "
PSNR ↑SSIM ↑LPIPS ↓FID ↓
mip-360VC+DMD29.290.1840.83635.59
VC+CausVid15.770.3360.44119.27
Ours16.280.3520.42917.44
TNTVC+DMD210.270.2090.73137.67
VC+CausVid17.330.4050.23224.92
Ours18.720.4110.21023.64
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.526, + 0.483, + 0.61 + ], + "angle": 0, + "content": "Table 2. Quantitative Results. We report four quantitative metrics for image quality comparison on 2 separate datasets: Mip-NeRF360 [1] (denoted as mip-360 in the table) and Tanks-and-Temple [18] (denoted as TNT in the table). VC stands for ViewCrafter [52] for space saving purpose. \\(\\mathbf{VC} + \\mathbf{X}\\) means applying diffusion acceleration method X on ViewCrafter." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.484, + 0.842 + ], + "angle": 0, + "content": "Quantitative Analysis Table 2 presents quantitative comparisons with other diffusion distillation methods, and further proves the superiority of our method. Our method differs with DMD2 [51] in several ways: 1) we apply a 3D discriminator on video data, 2) we perform soften reverse-KL divergence for DMD optimization, 3) our GAN training and DMD training are kept separate for more stable training and more accurate density ratio calculation. Through training DMD2 [51], we found a unstable training process and we recorded the last model before the training collapsed. In comparison, CausVid [50] presents more stable training than DMD2 [51] but still suffer from mode collapse. Our method surpasses the other two baseline methods in both perceptual and distribution metrics." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.857, + 0.255, + 0.872 + ], + "angle": 0, + "content": "5.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.88, + 0.483, + 0.911 + ], + "angle": 0, + "content": "In this section, we analyze the contributions of each module of our proposed method in Table 3 and Figure 6." + }, + { + "type": "table", + "bbox": [ + 0.529, + 0.089, + 0.895, + 0.175 + ], + "angle": 0, + "content": "
GANDMDSoften RKLPSNR ↑SSIM ↑LPIPS ↓FID ↓
✗✓✓8.620.1540.88040.17
✓✗✗16.230.3690.37521.48
✓✓✗16.850.3850.33721.05
✓✓✓17.500.3820.32020.54
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.185, + 0.905, + 0.214 + ], + "angle": 0, + "content": "Table 3. Ablation studies. Quantitative analysis of different components of FVGen." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.233, + 0.905, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.328, + 0.908, + 0.371 + ], + "angle": 0, + "content": "Figure 6. Ablation Studies. Visualization of the contribution of different components of FVGen. Our full model presents the most visual similarity between the ground truth." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.387, + 0.907, + 0.492 + ], + "angle": 0, + "content": "Assessing student initialization. As it can be seen from the first column of Figure 6 and Table 3, student initialization is a pivotal module of our architecture. Without student initialization with GAN optimization, the student model would fail to generate samples close to the teacher distribution, thus the teacher score function is not reliable for distillation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.906, + 0.599 + ], + "angle": 0, + "content": "Assessing distribution matching. Optimization of DMD continues to align the student model distribution with teacher model. Comparing with full model, although the perceptual metrics are not significantly degraded, qualitative results show that without distribution matching, the generated results appear significant blurriness and artifacts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.615, + 0.906, + 0.689 + ], + "angle": 0, + "content": "Assessing soften reverse KL-divergence. As discussed in Section 4.2, soften reverse KL-divergence enables more stable and robust training. Therefore, the DMD optimization achieves better convergence and generated results are more distilled into teacher distribution." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.708, + 0.637, + 0.723 + ], + "angle": 0, + "content": "6. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.9 + ], + "angle": 0, + "content": "We identify several limitations in our current work. First, although FVGen achieves similar visual quality compared to ViewCrafter, it is limited by the drawbacks of ViewCrafter: e.g., degradation in structural integrity and consistency when experiencing extremely sparse inputs. Second, FVGen benefits from the synergy of three video diffusion models; therefore, due to computational constraints, we were only able to train the model with videos up to only 16 frames, which may not provide enough coverage for extremely large scenes. Therefore, our future work would address on the limitations mentioned above." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26102" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.271, + 0.108 + ], + "angle": 0, + "content": "7. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.484, + 0.254 + ], + "angle": 0, + "content": "Supported by the Intelligence Advanced Research Projects Activity (IARPA) via Department of Interior/ Interior Business Center (DOI/IBC) contract number 140D0423C0075. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon. Disclaimer: The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of IARPA, DOI/IBC, or the U.S. Government." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.282, + 0.188, + 0.298 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.307, + 0.484, + 0.363 + ], + "angle": 0, + "content": "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-NeRF: A Multiscale Representation for Anti-Aliasing Neural Radiance Fields. Technical report, 2021. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.364, + 0.484, + 0.446 + ], + "angle": 0, + "content": "[2] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.448, + 0.484, + 0.504 + ], + "angle": 0, + "content": "[3] Dar-Yen Chen, Hmrishav Bandyopadhyay, Kai Zou, and Yi-Zhe Song. Nitrofusion: High-fidelity single-step diffusion through dynamic adversarial training. arXiv preprint arXiv:2412.02030, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.505, + 0.484, + 0.588 + ], + "angle": 0, + "content": "[4] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videographer2: Overcoming data limitations for high-quality video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7310-7320, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.59, + 0.484, + 0.657 + ], + "angle": 0, + "content": "[5] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splattering from sparse multi-view images. arXiv preprint arXiv:2403.14627, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.66, + 0.484, + 0.715 + ], + "angle": 0, + "content": "[6] Yuedong Chen, Chuanxia Zheng, Haofei Xu, Bohan Zhuang, Andrea Vedaldi, Tat-Jen Cham, and Jianfei Cai. Mvsplat360: Feed-forward 360 scene synthesis from sparse views. 2024. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.717, + 0.484, + 0.759 + ], + "angle": 0, + "content": "[7] Zilong Chen, Yikai Wang, Feng Wang, Zhengyi Wang, and Huaping Liu. V3d: Video diffusion models are effective 3d generators. arXiv preprint arXiv:2403.06738, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.76, + 0.484, + 0.829 + ], + "angle": 0, + "content": "[8] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T Barron, and Ben Poole. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv preprint arXiv:2405.10314, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.83, + 0.484, + 0.857 + ], + "angle": 0, + "content": "[9] Dibya Ghosh. KI divergence - intuition and examples, n.d. Accessed: 2025-03-07. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[10] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.307, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[11] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv:2204.03458, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[12] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[13] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssm. In 2010 20th international conference on pattern recognition, pages 2366-2369. IEEE, 2010. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.234, + 0.905, + 0.313 + ], + "angle": 0, + "content": "[14] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[15] Minguk Kang, Richard Zhang, Connelly Barnes, Sylvain Paris, Suha Kwak, Jaesik Park, Eli Shechtman, Jun-Yan Zhu, and Taesung Park. Distilling diffusion models into conditional gans. In European Conference on Computer Vision, pages 428-447. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.905, + 0.441 + ], + "angle": 0, + "content": "[16] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.469 + ], + "angle": 0, + "content": "[17] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.525 + ], + "angle": 0, + "content": "[18] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36 (4):1-13, 2017. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[19] Xin Kong, Shikun Liu, Xiaoyang Lyu, Marwan Taher, Xiaojuan Qi, and Andrew J Davison. Eschernet: A generative model for scalable view synthesis. arXiv preprint arXiv:2402.03908, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.581, + 0.905, + 0.637 + ], + "angle": 0, + "content": "[20] Xinyang Li, Zhangyu Lai, Linning Xu, Yansong Qu, Liujuan Cao, Shengchuan Zhang, Bo Dai, and Rongrong Ji. Director3d: Real-world camera trajectory and 3d scene generation from text. arXiv preprint arXiv:2406.17601, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.638, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[21] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.721, + 0.905, + 0.762 + ], + "angle": 0, + "content": "[22] Shanchuan Lin, Anran Wang, and Xiao Yang. Sdxlightning: Progressive adversarial diffusion distillation. arXiv preprint arXiv:2402.13929, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.763, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[23] Shanchuan Lin, Xin Xia, Yuxi Ren, Ceyuan Yang, Xuefeng Xiao, and Lu Jiang. Diffusion adversarial post-training for one-step video generation. arXiv preprint arXiv:2501.08316, 2025. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[24] Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22160–22169, 2024. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26103" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.16 + ], + "angle": 0, + "content": "[25] Fangfu Liu, Wenqiang Sun, Hanyang Wang, Yikai Wang, Haowen Sun, Junliang Ye, Jun Zhang, and Yueqi Duan. Reconx: Reconstruct any scene from sparse views with video diffusion model. arXiv preprint arXiv:2408.16767, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.192 + ], + "angle": 0, + "content": "[26] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[27] Yihong Luo, Xiaolong Chen, Xinghua Qu, Tianyang Hu, and Jing Tang. You only sample once: Taming one-step text-to-image synthesis by self-cooperative diffusion gans. arXiv preprint arXiv:2403.12931, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[28] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14297-14306, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[29] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.482, + 0.418 + ], + "angle": 0, + "content": "[30] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.475 + ], + "angle": 0, + "content": "[31] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.478, + 0.482, + 0.546 + ], + "angle": 0, + "content": "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.617 + ], + "angle": 0, + "content": "[33] Axel Sauer, Frederic Boesel, Tim Dockhorn, Andreas Blattmann, Patrick Esser, and Robin Rombach. Fast high-resolution image synthesis with latent adversarial diffusion distillation. In SIGGRAPH Asia 2024 Conference Papers, pages 1-11, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[34] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In European Conference on Computer Vision, pages 87-103. Springer, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.482, + 0.716 + ], + "angle": 0, + "content": "[35] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[36] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[37] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[38] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. In European Conference on Computer Vision, pages 439-457. Springer, 2025. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[39] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12619-12629, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.231 + ], + "angle": 0, + "content": "[40] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[41] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.905, + 0.356 + ], + "angle": 0, + "content": "[42] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[43] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21551-21561, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[44] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Gongye Liu, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Dynamiccafter: Animating open-domain images with video diffusion priors. In European Conference on Computer Vision, pages 399-417. Springer, 2025. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[45] Haofei Xu, Songyou Peng, Fangjinhua Wang, Hermann Blum, Daniel Barath, Andreas Geiger, and Marc Pollefeys. Depthsplat: Connecting gaussian splatting and depth. arXiv preprint arXiv:2410.13862, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[46] Yanwu Xu, Yang Zhao, Zhisheng Xiao, and Tingbo Hou. Ufogen: You forward once large scale text-to-image generation via diffusion gans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8196-8206, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.905, + 0.692 + ], + "angle": 0, + "content": "[47] Yilun Xu, Weili Nie, and Arash Vahdat. One-step diffusion models with \\(f\\)-divergence distribution matching. arXiv preprint arXiv:2502.15681, 2025. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.693, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[48] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[49] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6613-6623, 2024. 2, 3, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[50] Tianwei Yin, Qiang Zhang, Richard Zhang, William T Freeman, Fredo Durand, Eli Shechtman, and Xun Huang. From slow bidirectional to fast causal video generators. arXiv preprint arXiv:2412.07772, 2024. 2, 3, 4, 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26104" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[51] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and Bill Freeman. Improved distribution matching distillation for fast image synthesis. Advances in Neural Information Processing Systems, 37: 47455-47487, 2025. 2, 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[52] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.303 + ], + "angle": 0, + "content": "[53] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "26105" + } + ] +] \ No newline at end of file diff --git a/2025/FVGen_ Accelerating Novel-View Synthesis with Adversarial Video Diffusion Distillation/381cfeb4-88c2-44b1-a694-860cc7f57161_origin.pdf b/2025/FVGen_ Accelerating Novel-View Synthesis with Adversarial Video Diffusion Distillation/381cfeb4-88c2-44b1-a694-860cc7f57161_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c2feeac81edeb7db94bb006601730c7b56c297bc --- /dev/null +++ b/2025/FVGen_ Accelerating Novel-View Synthesis with Adversarial Video Diffusion Distillation/381cfeb4-88c2-44b1-a694-860cc7f57161_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:30df57995b6d9079e932aae8459a7fd18f60198d41ea494c5b3857af3d6b56a7 +size 49814622 diff --git a/2025/FVGen_ Accelerating Novel-View Synthesis with Adversarial Video Diffusion Distillation/full.md b/2025/FVGen_ Accelerating Novel-View Synthesis with Adversarial Video Diffusion Distillation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..70e2316ceaa649c68492e1ea3b1d58c5c4223c8a --- /dev/null +++ b/2025/FVGen_ Accelerating Novel-View Synthesis with Adversarial Video Diffusion Distillation/full.md @@ -0,0 +1,338 @@ +# FVGen: Accelerating Novel-View Synthesis with Adversarial Video Diffusion Distillation + +Wenbin Teng $^{1,2}$ , Gonglin Chen $^{1,2}$ , Haiwei Chen $^{1,2}$ , Yajie Zhao $^{1,2*}$ $^{1}$ Institute for Creative Technologies + $^{2}$ University of Southern California +{wenbinte, gonglinc}@usc.edu, {chenh, zhao}@ict.usc.edu, + +![](images/f58be604aa60f583d0eb5e36ffcade1f90c56d6b51ae048204c4c6422c873ff7.jpg) +Figure 1. We present FVGen, a method of fast novel-view synthesis for 3D scene reconstruction with sparse inputs. Recent method [52] unleash the generation capabilities of video diffusion models conditioned on a prior 3D cue for dense view creation, but always suffers from a long sampling time. We propose a novel framework that can significantly reduce the sampling time and maintain the generation quality. + +# Abstract + +Recent progress in 3D reconstruction has enabled realistic 3D models from dense image captures, yet challenges persist with sparse views, often leading to artifacts in unseen areas. Recent works leverage Video Diffusion Models (VDMs) to generate dense observations, filling the gaps when only sparse views are available for 3D reconstruction tasks. A significant limitation of these methods is their slow sampling speed when using VDMs. In this paper, we present FVGen, a novel framework that addresses this challenge by enabling fast novel view synthesis using VDMs in as few as four sampling steps. We propose a novel video diffusion model distillation method that distills a multi-step denoising teacher model into a few-step denoising student model using Generative Adversarial Networks (GANs) and softened reverse KL-divergence minimization. Extensive ex + +periments on real-world datasets show that, compared to previous works, our framework generates the same number of novel views with similar (or even better) visual quality while reducing sampling time by more than $90\%$ . FVGen significantly improves time efficiency for downstream reconstruction tasks, particularly when working with sparse input views (more than 2) where pre-trained VDMs need to be run multiple times to achieve better spatial coverage. Project Page: https://wbteng9526.github.io/fvgen/ + +# 1. Introduction + +Reconstructing 3D scenes from 2D images remains a pivotal research area at the intersection of computer vision and graphics. This field has extensive social impact, with broad applications in autonomous navigation, augmented reality, and more. With the recent development of neural rendering techniques such as NeRF [29] and 3DGS [16], creating high-quality 3D assets from 2D images has become more + +accessible. However, the success of 3D reconstruction depends on dense observations, which are often difficult to obtain in real-world scenarios. This limitation makes it challenging to apply these methods in many practical applications. + +Introducing generative models into 3D reconstruction can fill gaps in missing views when only sparse input views are available. However, novel view synthesis from sparse views, especially in the settings with less than 2 views, is inherently an ill-posed problem. A bulk of previous works [7, 8, 30, 43, 52] have found that priors learned from large image and video diffusion models [10, 11] are rich enough to inpaint detailed and reasonable information in the many under-observed areas in the reconstructed 3D representations. We are primarily interested in the use of video diffusion models (VDM) to perform novel view synthesis. What bridges VDM to novel view synthesis is its ability to synthesize continuous views following a camera trajectories between the observed views. The spatio-temporal consistency in these smooth video trajectories, as demonstrated in several works [6, 7, 25, 44, 52], effectively addresses the challenging "Janus Problem" that typically limits generation quality in other multi-view settings. However, VDM is a computationally expensive method because of the iterative sampling process that characterized all denoising diffusion methods. The long generation time makes VDM-based generation methods unsuitable for certain real-world downstream tasks, such as dynamic 3D reconstruction, where new consistent views must be generated continuously. This limitation also affects large-scale 3D reconstruction that requires synthesizing a significant number of novel views. + +In this paper, we show that VDM-based novel view synthesis can be significantly accelerated for free. Without any compromise on the generation quality, we speed up diffusion-based synthesis of novel view by a factor of $90\%$ with a 4-step student model trained under a GAN objective [23]. Acceleration of diffusion models has been actively studied in the past, but previous techniques may not be well suited to our task due to two reasons: First, we need a method specifically for accelerating video diffusion models, rather than image diffusion models. Second, the method must generalize effectively when trained on multi-view video datasets, which contain significantly less data than general video datasets. To accelerate a diffusion model, recent methods [22, 28, 34, 46, 49] propose minimizing the distributional difference between a few-step student generation model and the original multi-step teacher model. However, these methods are restricted to images with specific domains. The most related work to us is [50], which generates more video frames with faster speed by optimizing a distribution matching distillation (DMD) loss. However, this method requires an effective initialization of the student model. This is completed by generating a noise- + +latent pair through an ODE solver [36] and training the student model by optimizing a regression loss [49, 50]. The generation process usually takes a very long time and the student's ability is limited by the teacher's generation quality. More importantly, we have observed that DMD optimizations tend to be unstable and have the tendency towards mode collapse, when trained under multi-view video data. This may be due to the inherent concept of DMD loss optimization being the reverse KL-divergence minimization, which tends to be mode-seeking and zeroing out the modes of teacher's that are not relatively dominant. + +To solve the previous problems with VDM acceleration, we propose a novel framework, named FVGen, that performs novel views generation with a fast video diffusion generation with as few as 4 steps. First of all, different from previous GAN-based distillation methods [34, 51], we find it beneficial to initial the few-step student model by training a GAN objective, where the student model is regarded as a generator of fake samples, and, instead of training a separate discriminator as in previous works, we leverage a pre-trained teacher model as the discriminator. Second, we propose the soften reverse KL-divergence to solve the unstable optimization of DMD loss. The soften reverse KL-divergence maintains similar mode-seeking behavior as reverse KL-divergence but with more robust distillation: it prevents the student from ignoring entire space of teacher's distribution, thereby preserving more of teacher's knowledge, which we find particularly beneficial in the context of limited data. + +Through effective and robust distillation, FVGen is capable of generating video sequences with comparable or better visual quality as ViewCrafter [52] but with significantly less time. FVGen is trained with Dl3DV-10K [24] dataset and extensive experiments on real world dataset including MipNeRF360 [1] and Tanks-and-Temple [18] demonstrate that our method achieves overall better performance than other state-of-the-art (SOTA) novel view synthesis methods and diffusion distillation methods. + +# 2. Related Works + +# 2.1. Diffusion-based Novel View Synthesis + +NeRF [29] and 3DGS [16] excel at novel view synthesis with dense input views. However, reconstructing scenes from sparse views requires additional priors. Diffusion models [10, 32] can generate realistic pseudo input views from sparse original inputs. Several works use Score-Distillation Sampling [21, 30, 39, 42] to create 3D objects from text prompts or single images by distilling latent diffusion model [32] knowledge into 3D representations. While NeRF, 3DGS, and SDS typically require time-consuming per-scene optimization, view-dependent diffusion models [8, 19, 20, 35] offer an alternative. These mod + +![](images/1c45046d4adef763a2176185c0c5c2871cb8b26708cda36532c9666f1a6d6155.jpg) +Figure 2. Overview of student initialization. We initialize our student model by training a GAN objective. The student model $G_{\theta}$ , initialized with the weights of teacher model, uses few-step denoising to generate fake videos to fool the discriminator $D$ . The fixed teacher model, together with a trainable discriminator, are optimized to differentiate between fake and real videos. + +![](images/38940fdad6faed9e8b2fbb5423e64d5561e0ad715ae0fbf445b5a143f3873c3a.jpg) + +els directly generate 3D-consistent novel views conditioned on input images but directly lifting 2D images to 3D input therefore face challenges when scaling to larger scenes. + +An alternative approach leverages video diffusion models [6, 7, 25, 38, 52] for multi-view image generation. These methods fine-tune pre-trained latent video diffusion models (LVDM) [2, 4, 12, 44, 48] using multi-view images with both 2D and 3D guidance. The 2D guidance typically involves semantic features like CLIP [14] embeddings, while 3D guidance comes from sparse input cues. For instance, [52] uses DUSt3R [40] to build an initial point cloud and guides generation by combining point cloud renders with camera poses. Similarly, [6] conditions generation on feedforward 3DGS renders, enabling 360-degree scene generation with longer frame sequences through multi-step refinement. However, these methods suffer from speed issues caused by video diffusion sampling when generating novel views. + +# 2.2. Diffusion Distillation + +Diffusion models typically require numerous denoising steps to generate high-quality samples, which makes them computationally intensive. To accelerate generation, distillation methods train a student model that mimics the teacher diffusion model's behavior while using fewer sampling steps. Using fewer sampling steps reduces quality, so existing works incorporate adversarial training [3, 15, 22, 27, 33, 34, 46] to enhance the student model's output. However, simple GAN training cannot fully bridge the gap between student and teacher distributions, causing the student's generated output to remain noticeably different from the original diffusion model. To address this problem, Distribution Matching Distillation (DMD) [49, 51] optimizes the reverse KL-divergence. For tractable optimization, [49, 51] leverages the gradient of DMD loss, which represents the difference between student and teacher score functions. For fast video generation, [50] extends DMD to video diffusion models like CogVideoX [48]. While these methods work well for general image and video generation tasks, they suffer from mode collapse and training insta + +bility when applied to view synthesis on relatively small datasets like multi-view video datasets. + +# 3. Preliminaries + +# 3.1. Video Diffusion Model + +A video diffusion model is a generative model designed to synthesize realistic video sequences by learning the complex temporal and spatial patterns within video data. Based on the diffusion process, these models iteratively apply noise to video frames and then learn to reverse this process to generate new sequences. The forward diffusion process progressively adds Gaussian noise to the input clean sample $\pmb{x}_0 \sim p(\pmb{x})$ over $T$ timesteps, leading to highly noisy data $\pmb{x}_T \sim \mathcal{N}(0, I)$ . This process is formulated as: + +$$ +q \left(\boldsymbol {x} _ {t} \mid \boldsymbol {x} _ {t - 1}\right) = \mathcal {N} \left(\boldsymbol {x} _ {t}; \sqrt {\alpha_ {t}} \boldsymbol {x} _ {t - 1}, (1 - \alpha_ {t}) \boldsymbol {I}\right), \tag {1} +$$ + +where $\alpha_{t}$ and $\pmb{x}_t$ are the noise strength and noisy data at timestep $t$ . The model then learns a denoising function, often parameterized by a neural network $\epsilon_{\theta}(\pmb{x}_t,t)$ , to reverse this process and generate realistic video frames, where the optimization function is defined by the MSE loss: + +$$ +\mathcal {L} _ {\text {d i f f}} (\theta) = \mathbb {E} _ {t \sim \mathcal {U} (0, 1), \epsilon \sim \mathcal {N} (0, I), \boldsymbol {x} \sim p, c} \| \epsilon - \epsilon_ {\theta} (\boldsymbol {x} _ {t}, t, c) \| ^ {2}, \tag {2} +$$ + +where $c$ is the condition embedding of diffusion model which is usually represented as a text or image prompt. Instead of generating sequences at full resolution, Latent Video Diffusion Model (LVDM) first encodes video data $\pmb{x} \in \mathbb{R}^{F \times C \times H \times W}$ into latent space using a pre-trained VAE [17] encoder: $\pmb{z} = \mathcal{E}(\pmb{x})$ , where $\pmb{z} \in \mathbb{R}^{F \times C \times h \times w}$ . The compression will mitigate computational complexity while maintaining video generation quality. + +# 3.2. Distribution Matching Distillation + +Distribution matching distillation adopts the idea of variational score distillation [42] that is optimized to match the distribution of student and teacher model by minimizing the + +![](images/bb23e5dca17e672a494befbd10bd2c91d30b95edd42880458fc1671f4374fa0a.jpg) +Figure 3. Overview of Distribution Matching Distillation (DMD). The gradient is the difference between the teacher score and fake score w.r.t. the student model output. We apply soften reverse KL-divergence for stable training and avoiding mode-collapse. + +reverse KL-divergence. Given a student model $G$ parameterized by $\theta$ , the gradient of reverse KL-divergence is + +$$ +\begin{array}{l} \nabla \mathcal {L} _ {\mathrm {D M D}} = \mathbb {E} _ {t} \left(\nabla_ {\theta} \mathbf {D} _ {\text {R e v e r s e - K L}} \left(p _ {\text {f a k e}, t} \| p _ {\text {r e a l}, t}\right)\right) \\ = - \mathbb {E} _ {t} \left(\left[ s _ {\text {r e a l}} \left(F \left(G _ {\theta} (\boldsymbol {z}), t\right), t\right) \right. \right. \\ \left. - s _ {\text {f a k e}} \left(F \left(G _ {\theta} (\boldsymbol {z}), t\right), t\right) \frac {d G _ {\theta} (\boldsymbol {z})}{d \theta} \right]\left. \right), \tag {3} \\ \end{array} +$$ + +where $p_{\mathrm{real}}$ and $p_{\mathrm{fake}}$ are the real and fake distribution. $z \sim \mathcal{N}(0, I)$ is the initial random Gaussian noise. $F$ is the diffusion forward function determined by Eq. (1). $s_{\mathrm{real}}$ and $s_{\mathrm{fake}}$ are the real and fake score function and defined as follows according to [37]: + +$$ +s \left(\boldsymbol {x} _ {t}, t\right) = \nabla_ {\boldsymbol {x} _ {t}} \log p _ {t} \left(\boldsymbol {x} _ {t}\right) = - \frac {\boldsymbol {x} _ {t} - \alpha_ {t} \mu \left(\boldsymbol {x} _ {t} , t\right)}{\sigma_ {t} ^ {2}}, \tag {4} +$$ + +where $p_t$ is the distribution of the noisy sample $x$ . $\mu$ is the clean sample prediction of the diffusion model. While optimizing student model $G_{\theta}$ with gradient descent formulated in Eq. (3), the gradients of $\mu_{\mathrm{real}}$ are stopped and $\mu_{\mathrm{fake}}$ is dynamically optimized with the output of $G_{\theta}$ with regular diffusion loss similar to Eq. (2) + +# 4. Method + +We propose Fast Video Generation (FVGen) that distills the generation capability of a multi-step teacher 3D-VDM into a few-step student 3D-VDM. The teacher 3D-VDM used in our experiments is the novel-view synthesis model ViewCrafter [52]. From a set of observed images $I^{\mathrm{obs}}$ and observed cameras $\pi^{\mathrm{obs}}$ , ViewCrafter learns a conditional distribution $I^{\mathrm{tgt}} \sim p(I^{\mathrm{tgt}}|I^{\mathrm{obs}},\pi^{\mathrm{obs}},\pi^{\mathrm{tgt}})$ , where $I^{\mathrm{tgt}} = \{I_i'\}_{i=1}^L$ is a sequence of $L$ images corresponding to the target views $\pi^{\mathrm{tgt}}$ . The student model FVGen is trained to sample from the same conditional distribution with a much faster inference speed. The architectures of FVGen are shown in Figure 2 and 3. For notation simplicity purpose, + +we ignore including $I^{\mathrm{obs}}$ , $\pi^{\mathrm{obs}}$ , $\pi^{\mathrm{tgt}}$ in the future formula as they are all the conditional inputs for both teacher and student diffusion model. + +We first initialize the parameters of student model by training with a generative adversarial network (GAN) objective, leveraging the teacher model as discriminator (Section 4.1). After that, we continue to train the student model by optimizing DMD loss and diffusion loss to minimize the soften reverse KL-divergence between the distribution of student and teacher model (Section 4.2). + +# 4.1. Student Initialization with GAN Training + +Directly training student model with DMD (Eq. 3) will cause training collapse as it is difficult to naively distill multi-step denoising into few-step. [49, 50] proposed to create noise-latent pairs by an ODE solver [36] and minimize a regression loss with student model for initialization. However, we found that it is very time consuming to generate large amount of video data pairs and the student model will also be restricted by the teacher's limitations. Inspired by [23, 51], we propose to initialize the student model by training a GAN objective with the real images. Specifically, the student model $(G_{\theta})$ generates fake video samples that fool the teacher model who serves as a discriminator (denoted as $D$ ) by minimizing the generator loss $\mathcal{L}_G$ . The discriminator $D$ classifies real samples from generated fake samples by maximizing $-\mathcal{L}_D$ . The min-max game adversarial optimization is formulated as: + +$$ +\begin{array}{l} \mathcal{L}_{D} = \mathbb{E}_{\substack{\boldsymbol {x}\sim p_{\text{real}}\\ t\sim \mathcal{U}[0,T]}}[\log f\left(D(F(\boldsymbol {x},t))\right)] \\ - \mathbb {E} _ {\substack {\boldsymbol {z} \sim \mathcal {N} (0, I) \\ t \sim \mathcal {U} [ 0, T ]}} \left[ \log f \left(D \left(F \left(G _ {\theta} (\boldsymbol {z}), t\right)\right)\right) \right], \tag{5} \\ \end{array} +$$ + +$$ +\mathcal{L}_{G} = \operatorname{\mathbb{E}}_{\substack{\boldsymbol {z}\sim \mathcal{N}(0,\boldsymbol {I})\\ t\sim \mathcal{U}[0,T]}}[\log f\left(D(F(G_{\theta}(\boldsymbol {z}),t))\right)], +$$ + +where $t$ is the time-step sampling and $F$ is the forward noise-injection model similar to Eq. (3). Discriminator $D$ retrieves the output feature map of the middle block of diffusion model UNet and passes it into the classifier $f$ . To + +fit with video data, $f$ is composed of a 3D Convolutional Neural Network that compresses the feature map into classification logits. Here, we use the original sample $\pmb{x}$ as the input to the discriminator instead of the generated sample $\pmb{x}$ from teacher model such that the student model would not be restricted by the generation capability of teacher model. For the purpose of training stability, recent methods [6, 52] predict the velocity field $\pmb{v}_{\theta}$ , and we convert it into the sample prediction with $\pmb{x}_0 = \sqrt{\alpha_t\pmb{x}_t} -\sqrt{1 - \alpha_t}\pmb{v}_{\theta}$ . Although the generated video samples are still very blurry, the GAN training scheme provides an effective initialization of the student model, which paves the way for more effective training with distribution matching loss. + +# 4.2. Distribution Matching with Soften Reverse KL-Divergence Minimization + +As discussed in Section 3.2, the core concept of distribution matching distillation (DMD) [49, 51] is to minimize the reverse KL-divergence between student and teacher distributions. While reverse KL is popular for its mode-seeking focus, it has notable limitations in the context of generative modeling, its tendency toward mode collapse. Since reverse KL rewards the student for zeroing out probability in any region where the teacher's density is low, the student can effectively ignore parts where the teacher's distribution are not dominant. If the teacher's distribution $p_{\mathrm{real}}(\pmb{x})$ has multiple modes of varying height, minimizing ReverseKL $(p_{\mathrm{fake}} \| p_{\mathrm{real}})$ may lead $p_{\mathrm{fake}}$ to concentrate on a subset of those modes and assign negligible mass to others. Therefore, the exclusive nature of reverse KL-divergence makes it prone to mode dropping and a lack of coverage of the full distribution. + +To address the shortcomings of pure reverse KL, we propose to leverage a Soften reverse KL-divergence (Soften RKL) [9, 47]. Soften RKL refers to a modified divergence that retains the mode-seeking bias of reverse KL but softens its exclusion other relatively lower-probability regions, thereby mitigating mode collapse. Compared to reverse KL formulated in Eq. (3), the Soften RKL can be formulated as: + +$$ +\begin{array}{l} \mathbf {D} _ {\text {S o f t e n - R K L}} \left(p _ {\text {f a k e}, t} \| p _ {\text {r e a l}, t}\right) = \mathbf {D} _ {K L} \left(\frac {1}{2} p _ {\text {r e a l}} + \frac {1}{2} p _ {\text {f a k e}} \| p _ {\text {r e a l}}\right), \\ = p _ {\text {f a k e}} (r + 1) \log \left(\frac {1}{2} + \frac {1}{2 r}\right), \tag {6} \\ \end{array} +$$ + +where $r = r(\pmb{x},t) \coloneqq p_{\mathrm{real},t}(\pmb{x}) / p_{\mathrm{fake},t}(\pmb{x})$ is the density ratio. Here $\pmb{x}$ denotes the noise-injection version of the output of student model formulated as $F(G_{\theta}(\pmb{z}),t)$ with $t \sim \mathcal{U}(0,T)$ . Thanks to our optimized GAN objective in Section 4.1, the density ratio could be approximated by the following: + +$$ +r (\boldsymbol {x}, t) = \frac {p _ {\text {r e a l} , t} (\boldsymbol {x})}{p _ {\text {f a k e} , t} (\boldsymbol {x})} \approx \frac {f (D (\boldsymbol {x} , t))}{1 - f (D (\boldsymbol {x} , t))}. \tag {7} +$$ + +In summary, the well-optimized GAN objective not only provides an efficient weight initialization of the student model, but also provides a direct estimate of the density ratio. Here, instead of directly comparing $p_{\mathrm{fake}}$ and $p_{\mathrm{real}}$ as in $\mathbf{D}_{KL}(p_{\mathrm{fake}}||p_{\mathrm{real}})$ , we compare an even mixture of $p_{\mathrm{real}}$ and $p_{\mathrm{fake}}$ to $p_{\mathrm{real}}$ . Soften RKL maintains the mode-seeking property of reverse KL, while penalizes more on mode dropping. This criterion yields a more robust distillation since it prevents the student from ignoring the entire teacher distribution, thereby preserving more of teacher's knowledge. Through experiments, we find that the soften RKL yields more stable training and more realistic generations. Computing the divergence directly is generally intractable, the gradient with respect to $\theta$ is formulated as follows: + +$$ +\begin{array}{l} \nabla \mathcal {L} _ {\mathrm {D M D}} ^ {\text {S o f t t e n - R K L}} = \tag {8} \\ - \mathbb {E} _ {t} \left(\frac {1}{r (\boldsymbol {x} , t)} \left[ s _ {\text {r e a l}} (\boldsymbol {x}, t) - s _ {\text {f a k e}} (\boldsymbol {x}, t) \right] \frac {d G _ {\theta} (\boldsymbol {z})}{d \theta}\right), \\ \end{array} +$$ + +where $\pmb{x} = F(G_{\theta}(z),t)$ as previously defined. Apart from the soften RKL divergence optimization, we dynamically update the fake diffusion model $\mu_{\mathrm{fake}}$ to adjust to change of student distribution in accordance with the original setting of DMD [49]. The loss term is identical to Eq. (2). We summarize the full training procedure in Alg. 1 + +Algorithm 1 FVGen Training Procedure +```txt +Require: Few-step timesteps $\mathcal{T} = \{0, t_1, t_2, \dots, t_Q\}$ , pretrained teacher model $\mu_{\mathrm{real}}$ , discriminator classifier $D$ dataset $\mathcal{D}$ . +``` + +1: Initialize student model $G_{\theta}$ with $\mu_{\mathrm{real}}$ . +2: Initialize fake score function with $\mu_{\mathrm{fake}}$ +3: while training do +4: Sample a video from dataset $\pmb{x}_0 \sim \mathcal{D}$ +5: Add noise with timestep $t \sim \mathcal{T}$ : $\pmb{x}_t = \sqrt{\alpha_t} \pmb{x}_0 + \sqrt{1 - \alpha_t} \pmb{\epsilon}, \quad \pmb{\epsilon} \sim \mathcal{N}(0, \pmb{I})$ +6: Predict with student: $\hat{\pmb{x}}_0 = G_\theta (\pmb {x}_t,t)$ +7: Add noise $\epsilon' \sim \mathcal{N}(0, I)$ with timestep $\tau \sim \mathcal{U}(0, T)$ to real and fake sample: +8: $\hat{\pmb{x}}_{\tau} = \sqrt{\alpha_{\tau}}\hat{\pmb{x}}_0 + \sqrt{1 - \alpha_{\tau}}\pmb{\epsilon}'$ +9: $\pmb{x}_{\tau} = \sqrt{\alpha_{\tau}}\pmb{x}_0 + \sqrt{1 - \alpha_{\tau}}\pmb{\epsilon}'$ +10: Update $G_{\theta}$ and $D$ with Eq. (5). +11: end while +12: Output trained student $G_{\theta}$ +13: while training do +14: Repeat step 4 - 8 +15: Update $G_{\theta}$ with DMD Loss: Eq. (8) +16: Add noise $\pmb{\epsilon}'' \sim \mathcal{N}(0, \pmb{I})$ with $t_1 \sim \mathcal{U}(0, T)$ to $\hat{\pmb{x}}_0$ : +17: $\hat{\pmb{x}}_{t_1} = \sqrt{\alpha_{t_1}}\hat{\pmb{x}}_0 + \sqrt{1 - \alpha_{t_1}}\pmb{\epsilon}^{\prime \prime}$ +18: Update $\mu_{\mathrm{fake}}$ with denoising loss: Eq. (2) +19: end while + +![](images/a5173ad0018a8a98b79691071d9dc58bb5ed4d7e6d02a33f35ce1446b8ba633e.jpg) +Figure 4. Qualitative Results. We compare FVGen with SOTA novel view synthesis method on MipNeRF360 [1] and Tanks-and-Temples [18] datasets. + +# 5. Experiments + +# 5.1. Experiment Setup + +Implementation Details. All of our student model, teacher model and fake score function is initialized with the parameters of the sparse model of ViewCrafter [52]. We first initialize the student model by training GAN objective. The parameters of teacher model is fixed, and the student model and GAN classifier are trained for 4000 iterations with a two-scale update rule inspired by [51]. Next, we optimize DMD loss by continuously training student model and the fake score function for another 5000 iterations with the same two-scale update rule. The whole pipeline is trained on 8 NVIDIA H100 with batch size 4. The training last about 1 day. We utilize AdamW optimizer [26] and set learning rates of all optimizers to be $5 \times 10^{-5}$ . The model is trained on images with a resolution of $512 \times 320$ . + +Datasets. FVGen is trained with DL3DV-10K [24], a real-world, scene-level video dataset that includes more than 10K long video clips. Similar to [52], we construct point cloud and ground truth video pairs using DL3DV dataset. We randomly sample short video clips with a random frame stride less than 5. The first frame and last frame are used to construct the prior point cloud with DUSt3R [40]. We use PyTorch3d [31], an efficient renderer to construct the point cloud renders with the camera poses of intermediate frames. Through this process, we create 20,000 training data pairs. We validate our method + +on two public datasets Tanks-and-Temple [18] and MipN-eRF360 [1] + +Baseline Methods and Evaluation Metrics. To evaluate our proposed method, we compare our method with several state-of-the-art methods: ViewCrafter [52], MVsplat360 [6]. We also compare our video diffusion distillation method with two state-of-the-art methods: CausVid [50] and DMD2 [51]. We use PSNR [13], SSIM [41], LPIPS [53] to measure reconstruction quality. In addition, we also reported the distribution metric, i.e. Frechet Inception Distance (FID), which compares the distribution of generated views and ground truth views. Apart from the visual quality comparison, + +# 5.2. Novel View Synthesis Comparison + +Similar to the evaluation setting of MVSplat360 [6], for each scene of both MipNeRF360 [1] and Tanks-and-Temples [18], we select 5 views that are far from each other but could cover the whole scene. We use our method and baseline method to generate 56 views sampled from the natural camera trajectory (14 views between each 2 input views). + +Qualitative Results The qualitative comparisons with SOTA novel view synthesis methods are shown in Figure 4. DepthSplat is a feed-forward generalizable Gaussian splatting method. When there are relatively low overlap between input frames, DepthSplat would exhibit ob + +![](images/3593d096fbd141b0bd7d7e363319734fea7531952e193b1721eaca22c323bdfe.jpg) +Input Views + +![](images/3bc694cfe6a24c0a192576afb84f426f4a59e100418d05a94f94b7daa655dbfa.jpg) +ViewCrafter + DMD2 + +![](images/7ed713d6573a8d58858221afbe310e8d34b9d6b4d10f7d397f68892f5ba57fa9.jpg) +ViewCrafter + CausVid + +![](images/ea21bd79ce63a0e33b213f2987fa8414ff3a7f703e5d8b418a01ca7ef1b79771.jpg) +Ours + +![](images/dc3664f6f3a5da2c21d8c2d8d3e2c57a9bb8fef29635d735152febda79fe914f.jpg) +Ground Truth + +![](images/b50b0fda2a94f4464748d47e6eed07c821be76a45a5a4b5f6e8d8b858b0a5dc6.jpg) + +![](images/89e6b038db7333ff3d1125e105fcfaa2fa9b59f152c215a08a67e5062c40c368.jpg) + +![](images/9e23cd413eec3e8b8d44edfb1080969d2684934b0b1b23e24c0e6dba0ff6359e.jpg) + +![](images/ae3ed8211ff5172647b4ff3d95890ab90dde83b558d0761947afd5ea651e1b62.jpg) + +![](images/caec3d14189f46947e91b50e8e2342b49914703da329894418d13f8b70ca80b0.jpg) + +![](images/83ca9a1348f07e041d5f0c085bfc8385339e29fc49b1fe86e58016e2517d53f8.jpg) + +![](images/6ef51fe1dbb98a0cbda6fc693ef3f0d17637e45f12f14cc00b990a96981e1582.jpg) + +![](images/40295a4d29f721bdcd815a5a3f759482433c5c75152c23602ccaf20578c22839.jpg) + +![](images/864656901dbeaefc5fed444ec8ff99d0930a3513424cc0da1f392923dc1d814c.jpg) + +![](images/c4fc1b7603d36534b0d1efd782c4e45483ba707c8ef5a511a93995deaac73bb6.jpg) + +![](images/6dcf859b41a54aa21d081b9996a0ec71f4c82ad19bd8825bd88c175249ab7949.jpg) +Figure 5. Qualitative Results. We compare FVGen with baseline diffusion distillation method on MipNeRF360 [1] and Tanks-and-Temples [18] datasets. + +![](images/9c98280ab97cbbb32f8aac3ed708707be30cbc0c540a411c6f2c711836b5a934.jpg) + +![](images/42119404f3c8ef2c82ceec63ea954ee11e42d5951782be40dd53f8b712c7fcb5.jpg) + +![](images/0c83a9787bb16cf32e23c4e8e5f7555014098b34faa93f9e28dddb519611a8aa.jpg) + +![](images/547326389256b57ba8ee5d23b49790ab3e0f4111f22c1a1b29634e05d7b31219.jpg) + +vious artifacts due to inconsistent depth scale and the issue of floating Gaussians. In comparison, MVSplat360 [6] uses VDM to refine the Gaussian splattering renders and remove the floaters. However, the current MVSplat360 only supports training with low resolution inputs and the current results present significant blurriness. Compared to DepthSplat and MVSplat360, ViewCrafter [52] generates photo-realistic novel views with higher resolution. However, ViewCrafter is limited by the generation speed. Given generating 16 total frames, the average inference time is 13.2 seconds. In comparison, our method achieves a 10x speedup, generating the same number of frames in approximately 1 second and achieves similar visual results. The qualitative results prove that our method is able to distill the multi-step sampling into a few-step sampling process and achieve results that are similar to or even better than teacher ViewCrafter model with the student model. + +Quantitative Analysis Table 1 presents quantitative comparisons on both datasets of FVGen and other novel view synthesis methods. Except for ViewCrafter [52], our method significantly surpasses other SOTA methods in terms of both perceptual and distribution metrics. Compared with ViewCrafter, our method yields similar quality + +
PSNR ↑SSIM ↑LPIPS ↓FID ↓Time ↓
mip-360DepthSplat11.230.2130.71532.454.2
MVSplat36012.280.2850.68225.6987.2
ViewCrafter16.350.3460.43316.2866.3
Ours16.280.3520.42917.445.1
TNTDepthSplat12.430.2630.67735.884.3
MVSplat36014.180.3010.53225.2387.3
ViewCrafter18.690.4020.20823.9465.9
Ours18.720.4110.21023.645.0
+ +Table 1. Quantitative Results. We report four quantitative metrics for image quality comparison on 2 separate datasets: Mip-NeRF360 [1] (denoted as mip-360 in the table) and Tanks-and-Temple [18] (denoted as TNT in the table). We highlight the best results in red and second-best in yellow. + +but with more than $90\%$ faster in generation speed. The quantitative results also highlights the effectiveness of our proposed method. + +# 5.3. Diffusion Distillation Comparison + +We also compare FVGen with SOTA diffusion distillation methods. Similar to the 2-view evaluation setting of DepthSplat [45] and MVSplat [5], we select 2 views that are far from each other in both MipNeRF360 [1] and Tanks-and- + +Temples [18] dataset. We use our method and baseline method to generate 16 views samples from the natural camera trajectory (14 intermediate views and 2 input views). + +Qualitative Results First we perform DMD2 [51] on ViewCrafter for few-step distillation. DMD2 applies end-to-end training of student model, GAN discriminator and fake score function. We find that this training scheme is not stable and present significant variance so that it is hard to converge, leading to blurry results as visually depicted in Figure 5. In addition, we compare our method with a recent work, Caus-Vid [50], on ViewCrafter distillation. CausVid [50] creates a small dataset with ODE solver and initialize the student model by optimizing a regression loss. Caus-Vid also leverages DMD [49] for 4-step video generation. However, the qualitative results indicate that our GAN object training scheme provides better initialization compared to the regression training scheme of ODE pairs. + +
PSNR ↑SSIM ↑LPIPS ↓FID ↓
mip-360VC+DMD29.290.1840.83635.59
VC+CausVid15.770.3360.44119.27
Ours16.280.3520.42917.44
TNTVC+DMD210.270.2090.73137.67
VC+CausVid17.330.4050.23224.92
Ours18.720.4110.21023.64
+ +Quantitative Analysis Table 2 presents quantitative comparisons with other diffusion distillation methods, and further proves the superiority of our method. Our method differs with DMD2 [51] in several ways: 1) we apply a 3D discriminator on video data, 2) we perform soften reverse-KL divergence for DMD optimization, 3) our GAN training and DMD training are kept separate for more stable training and more accurate density ratio calculation. Through training DMD2 [51], we found a unstable training process and we recorded the last model before the training collapsed. In comparison, CausVid [50] presents more stable training than DMD2 [51] but still suffer from mode collapse. Our method surpasses the other two baseline methods in both perceptual and distribution metrics. + +# 5.4. Ablation Studies + +In this section, we analyze the contributions of each module of our proposed method in Table 3 and Figure 6. + +Table 2. Quantitative Results. We report four quantitative metrics for image quality comparison on 2 separate datasets: Mip-NeRF360 [1] (denoted as mip-360 in the table) and Tanks-and-Temple [18] (denoted as TNT in the table). VC stands for ViewCrafter [52] for space saving purpose. $\mathbf{VC} + \mathbf{X}$ means applying diffusion acceleration method X on ViewCrafter. + +
GANDMDSoften RKLPSNR ↑SSIM ↑LPIPS ↓FID ↓
✗✓✓8.620.1540.88040.17
✓✗✗16.230.3690.37521.48
✓✓✗16.850.3850.33721.05
✓✓✓17.500.3820.32020.54
+ +Table 3. Ablation studies. Quantitative analysis of different components of FVGen. + +![](images/3c8c9a8a076a58ae51952b2685b688dc1f2745054087fc585e5bc404e5073279.jpg) +Figure 6. Ablation Studies. Visualization of the contribution of different components of FVGen. Our full model presents the most visual similarity between the ground truth. + +Assessing student initialization. As it can be seen from the first column of Figure 6 and Table 3, student initialization is a pivotal module of our architecture. Without student initialization with GAN optimization, the student model would fail to generate samples close to the teacher distribution, thus the teacher score function is not reliable for distillation. + +Assessing distribution matching. Optimization of DMD continues to align the student model distribution with teacher model. Comparing with full model, although the perceptual metrics are not significantly degraded, qualitative results show that without distribution matching, the generated results appear significant blurriness and artifacts. + +Assessing soften reverse KL-divergence. As discussed in Section 4.2, soften reverse KL-divergence enables more stable and robust training. Therefore, the DMD optimization achieves better convergence and generated results are more distilled into teacher distribution. + +# 6. Limitations + +We identify several limitations in our current work. First, although FVGen achieves similar visual quality compared to ViewCrafter, it is limited by the drawbacks of ViewCrafter: e.g., degradation in structural integrity and consistency when experiencing extremely sparse inputs. Second, FVGen benefits from the synergy of three video diffusion models; therefore, due to computational constraints, we were only able to train the model with videos up to only 16 frames, which may not provide enough coverage for extremely large scenes. Therefore, our future work would address on the limitations mentioned above. + +# 7. Acknowledgement + +# References + +[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-NeRF: A Multiscale Representation for Anti-Aliasing Neural Radiance Fields. Technical report, 2021. 2, 6, 7, 8 +[2] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 3 +[3] Dar-Yen Chen, Hmrishav Bandyopadhyay, Kai Zou, and Yi-Zhe Song. Nitrofusion: High-fidelity single-step diffusion through dynamic adversarial training. arXiv preprint arXiv:2412.02030, 2024. 3 +[4] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videographer2: Overcoming data limitations for high-quality video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7310-7320, 2024. 3 +[5] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splattering from sparse multi-view images. arXiv preprint arXiv:2403.14627, 2024. 7 +[6] Yuedong Chen, Chuanxia Zheng, Haofei Xu, Bohan Zhuang, Andrea Vedaldi, Tat-Jen Cham, and Jianfei Cai. Mvsplat360: Feed-forward 360 scene synthesis from sparse views. 2024. 2, 3, 5, 6, 7 +[7] Zilong Chen, Yikai Wang, Feng Wang, Zhengyi Wang, and Huaping Liu. V3d: Video diffusion models are effective 3d generators. arXiv preprint arXiv:2403.06738, 2024. 2, 3 +[8] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T Barron, and Ben Poole. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv preprint arXiv:2405.10314, 2024. 2 +[9] Dibya Ghosh. KI divergence - intuition and examples, n.d. Accessed: 2025-03-07. 5 +[10] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2 + +Supported by the Intelligence Advanced Research Projects Activity (IARPA) via Department of Interior/ Interior Business Center (DOI/IBC) contract number 140D0423C0075. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon. Disclaimer: The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of IARPA, DOI/IBC, or the U.S. Government. +[11] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv:2204.03458, 2022. 2 +[12] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022.3 +[13] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssm. In 2010 20th international conference on pattern recognition, pages 2366-2369. IEEE, 2010. 6 +[14] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 3 +[15] Minguk Kang, Richard Zhang, Connelly Barnes, Sylvain Paris, Suha Kwak, Jaesik Park, Eli Shechtman, Jun-Yan Zhu, and Taesung Park. Distilling diffusion models into conditional gans. In European Conference on Computer Vision, pages 428-447. Springer, 2024. 3 +[16] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 2 +[17] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3 +[18] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36 (4):1-13, 2017. 2, 6, 7, 8 +[19] Xin Kong, Shikun Liu, Xiaoyang Lyu, Marwan Taher, Xiaojuan Qi, and Andrew J Davison. Eschernet: A generative model for scalable view synthesis. arXiv preprint arXiv:2402.03908, 2024. 2 +[20] Xinyang Li, Zhangyu Lai, Linning Xu, Yansong Qu, Liujuan Cao, Shengchuan Zhang, Bo Dai, and Rongrong Ji. Director3d: Real-world camera trajectory and 3d scene generation from text. arXiv preprint arXiv:2406.17601, 2024. 2 +[21] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 2 +[22] Shanchuan Lin, Anran Wang, and Xiao Yang. Sdxlightning: Progressive adversarial diffusion distillation. arXiv preprint arXiv:2402.13929, 2024. 2, 3 +[23] Shanchuan Lin, Xin Xia, Yuxi Ren, Ceyuan Yang, Xuefeng Xiao, and Lu Jiang. Diffusion adversarial post-training for one-step video generation. arXiv preprint arXiv:2501.08316, 2025. 2, 4 +[24] Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22160–22169, 2024. 2, 6 + +[25] Fangfu Liu, Wenqiang Sun, Hanyang Wang, Yikai Wang, Haowen Sun, Junliang Ye, Jun Zhang, and Yueqi Duan. Reconx: Reconstruct any scene from sparse views with video diffusion model. arXiv preprint arXiv:2408.16767, 2024. 2, 3 +[26] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[27] Yihong Luo, Xiaolong Chen, Xinghua Qu, Tianyang Hu, and Jing Tang. You only sample once: Taming one-step text-to-image synthesis by self-cooperative diffusion gans. arXiv preprint arXiv:2403.12931, 2024. 3 +[28] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14297-14306, 2023. 2 +[29] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2 +[30] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2 +[31] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020. 6 +[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2 +[33] Axel Sauer, Frederic Boesel, Tim Dockhorn, Andreas Blattmann, Patrick Esser, and Robin Rombach. Fast high-resolution image synthesis with latent adversarial diffusion distillation. In SIGGRAPH Asia 2024 Conference Papers, pages 1-11, 2024. 3 +[34] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In European Conference on Computer Vision, pages 87-103. Springer, 2024. 2, 3 +[35] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 2 +[36] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2, 4 +[37] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 4 +[38] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. In European Conference on Computer Vision, pages 439-457. Springer, 2025. 3 + +[39] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12619-12629, 2023. 2 +[40] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 3, 6 +[41] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6 +[42] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 2, 3 +[43] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21551-21561, 2024. 2 +[44] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Gongye Liu, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Dynamiccafter: Animating open-domain images with video diffusion priors. In European Conference on Computer Vision, pages 399-417. Springer, 2025. 2, 3 +[45] Haofei Xu, Songyou Peng, Fangjinhua Wang, Hermann Blum, Daniel Barath, Andreas Geiger, and Marc Pollefeys. Depthsplat: Connecting gaussian splatting and depth. arXiv preprint arXiv:2410.13862, 2024. 7 +[46] Yanwu Xu, Yang Zhao, Zhisheng Xiao, and Tingbo Hou. Ufogen: You forward once large scale text-to-image generation via diffusion gans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8196-8206, 2024. 2, 3 +[47] Yilun Xu, Weili Nie, and Arash Vahdat. One-step diffusion models with $f$ -divergence distribution matching. arXiv preprint arXiv:2502.15681, 2025. 5 +[48] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024.3 +[49] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6613-6623, 2024. 2, 3, 4, 5, 8 +[50] Tianwei Yin, Qiang Zhang, Richard Zhang, William T Freeman, Fredo Durand, Eli Shechtman, and Xun Huang. From slow bidirectional to fast causal video generators. arXiv preprint arXiv:2412.07772, 2024. 2, 3, 4, 6, 8 + +[51] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and Bill Freeman. Improved distribution matching distillation for fast image synthesis. Advances in Neural Information Processing Systems, 37: 47455-47487, 2025. 2, 3, 4, 5, 6, 8 +[52] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024. 1, 2, 3, 4, 5, 6, 7, 8 +[53] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6 \ No newline at end of file diff --git a/2025/FVGen_ Accelerating Novel-View Synthesis with Adversarial Video Diffusion Distillation/images.zip b/2025/FVGen_ Accelerating Novel-View Synthesis with Adversarial Video Diffusion Distillation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..378a0148e91d3a3d2b54eb1aace61fa29ff4bf4b --- /dev/null +++ b/2025/FVGen_ Accelerating Novel-View Synthesis with Adversarial Video Diffusion Distillation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3777a749cb376748adb459b684a6b1be0d406be4395291311145a29d47fa03b3 +size 859515 diff --git a/2025/FVGen_ Accelerating Novel-View Synthesis with Adversarial Video Diffusion Distillation/layout.json b/2025/FVGen_ Accelerating Novel-View Synthesis with Adversarial Video Diffusion Distillation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ea9f896dc3241ecbb1bae836d38f18ef5ec47d83 --- /dev/null +++ b/2025/FVGen_ Accelerating Novel-View Synthesis with Adversarial Video Diffusion Distillation/layout.json @@ -0,0 +1,9571 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 68, + 102, + 542, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 102, + 542, + 138 + ], + "spans": [ + { + "bbox": [ + 68, + 102, + 542, + 138 + ], + "type": "text", + "content": "FVGen: Accelerating Novel-View Synthesis with Adversarial Video Diffusion Distillation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 118, + 160, + 492, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 160, + 492, + 204 + ], + "spans": [ + { + "bbox": [ + 118, + 160, + 492, + 204 + ], + "type": "text", + "content": "Wenbin Teng" + }, + { + "bbox": [ + 118, + 160, + 492, + 204 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 118, + 160, + 492, + 204 + ], + "type": "text", + "content": ", Gonglin Chen" + }, + { + "bbox": [ + 118, + 160, + 492, + 204 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 118, + 160, + 492, + 204 + ], + "type": "text", + "content": ", Haiwei Chen" + }, + { + "bbox": [ + 118, + 160, + 492, + 204 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 118, + 160, + 492, + 204 + ], + "type": "text", + "content": ", Yajie Zhao" + }, + { + "bbox": [ + 118, + 160, + 492, + 204 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 118, + 160, + 492, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 118, + 160, + 492, + 204 + ], + "type": "text", + "content": "Institute for Creative Technologies \n" + }, + { + "bbox": [ + 118, + 160, + 492, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 118, + 160, + 492, + 204 + ], + "type": "text", + "content": "University of Southern California \n{wenbinte, gonglinc}@usc.edu, {chenh, zhao}@ict.usc.edu," + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 82, + 234, + 529, + 428 + ], + "blocks": [ + { + "bbox": [ + 82, + 234, + 529, + 428 + ], + "lines": [ + { + "bbox": [ + 82, + 234, + 529, + 428 + ], + "spans": [ + { + "bbox": [ + 82, + 234, + 529, + 428 + ], + "type": "image", + "image_path": "f58be604aa60f583d0eb5e36ffcade1f90c56d6b51ae048204c4c6422c873ff7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 441, + 555, + 485 + ], + "lines": [ + { + "bbox": [ + 54, + 441, + 555, + 485 + ], + "spans": [ + { + "bbox": [ + 54, + 441, + 555, + 485 + ], + "type": "text", + "content": "Figure 1. We present FVGen, a method of fast novel-view synthesis for 3D scene reconstruction with sparse inputs. Recent method [52] unleash the generation capabilities of video diffusion models conditioned on a prior 3D cue for dense view creation, but always suffers from a long sampling time. We propose a novel framework that can significantly reduce the sampling time and maintain the generation quality." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 487, + 200, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 487, + 200, + 498 + ], + "spans": [ + { + "bbox": [ + 151, + 487, + 200, + 498 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 513, + 297, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 513, + 297, + 693 + ], + "spans": [ + { + "bbox": [ + 54, + 513, + 297, + 693 + ], + "type": "text", + "content": "Recent progress in 3D reconstruction has enabled realistic 3D models from dense image captures, yet challenges persist with sparse views, often leading to artifacts in unseen areas. Recent works leverage Video Diffusion Models (VDMs) to generate dense observations, filling the gaps when only sparse views are available for 3D reconstruction tasks. A significant limitation of these methods is their slow sampling speed when using VDMs. In this paper, we present FVGen, a novel framework that addresses this challenge by enabling fast novel view synthesis using VDMs in as few as four sampling steps. We propose a novel video diffusion model distillation method that distills a multi-step denoising teacher model into a few-step denoising student model using Generative Adversarial Networks (GANs) and softened reverse KL-divergence minimization. Extensive ex" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 487, + 556, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 556, + 596 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 556, + 596 + ], + "type": "text", + "content": "periments on real-world datasets show that, compared to previous works, our framework generates the same number of novel views with similar (or even better) visual quality while reducing sampling time by more than " + }, + { + "bbox": [ + 313, + 487, + 556, + 596 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 313, + 487, + 556, + 596 + ], + "type": "text", + "content": ". FVGen significantly improves time efficiency for downstream reconstruction tasks, particularly when working with sparse input views (more than 2) where pre-trained VDMs need to be run multiple times to achieve better spatial coverage. Project Page: https://wbteng9526.github.io/fvgen/" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 609, + 395, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 609, + 395, + 621 + ], + "spans": [ + { + "bbox": [ + 314, + 609, + 395, + 621 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "Reconstructing 3D scenes from 2D images remains a pivotal research area at the intersection of computer vision and graphics. This field has extensive social impact, with broad applications in autonomous navigation, augmented reality, and more. With the recent development of neural rendering techniques such as NeRF [29] and 3DGS [16], creating high-quality 3D assets from 2D images has become more" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 145, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 145, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 145, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26095" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "content": "accessible. However, the success of 3D reconstruction depends on dense observations, which are often difficult to obtain in real-world scenarios. This limitation makes it challenging to apply these methods in many practical applications." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 137, + 295, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 137, + 295, + 446 + ], + "spans": [ + { + "bbox": [ + 56, + 137, + 295, + 446 + ], + "type": "text", + "content": "Introducing generative models into 3D reconstruction can fill gaps in missing views when only sparse input views are available. However, novel view synthesis from sparse views, especially in the settings with less than 2 views, is inherently an ill-posed problem. A bulk of previous works [7, 8, 30, 43, 52] have found that priors learned from large image and video diffusion models [10, 11] are rich enough to inpaint detailed and reasonable information in the many under-observed areas in the reconstructed 3D representations. We are primarily interested in the use of video diffusion models (VDM) to perform novel view synthesis. What bridges VDM to novel view synthesis is its ability to synthesize continuous views following a camera trajectories between the observed views. The spatio-temporal consistency in these smooth video trajectories, as demonstrated in several works [6, 7, 25, 44, 52], effectively addresses the challenging \"Janus Problem\" that typically limits generation quality in other multi-view settings. However, VDM is a computationally expensive method because of the iterative sampling process that characterized all denoising diffusion methods. The long generation time makes VDM-based generation methods unsuitable for certain real-world downstream tasks, such as dynamic 3D reconstruction, where new consistent views must be generated continuously. This limitation also affects large-scale 3D reconstruction that requires synthesizing a significant number of novel views." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 451, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 451, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 451, + 295, + 713 + ], + "type": "text", + "content": "In this paper, we show that VDM-based novel view synthesis can be significantly accelerated for free. Without any compromise on the generation quality, we speed up diffusion-based synthesis of novel view by a factor of " + }, + { + "bbox": [ + 56, + 451, + 295, + 713 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 56, + 451, + 295, + 713 + ], + "type": "text", + "content": " with a 4-step student model trained under a GAN objective [23]. Acceleration of diffusion models has been actively studied in the past, but previous techniques may not be well suited to our task due to two reasons: First, we need a method specifically for accelerating video diffusion models, rather than image diffusion models. Second, the method must generalize effectively when trained on multi-view video datasets, which contain significantly less data than general video datasets. To accelerate a diffusion model, recent methods [22, 28, 34, 46, 49] propose minimizing the distributional difference between a few-step student generation model and the original multi-step teacher model. However, these methods are restricted to images with specific domains. The most related work to us is [50], which generates more video frames with faster speed by optimizing a distribution matching distillation (DMD) loss. However, this method requires an effective initialization of the student model. This is completed by generating a noise-" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "text", + "content": "latent pair through an ODE solver [36] and training the student model by optimizing a regression loss [49, 50]. The generation process usually takes a very long time and the student's ability is limited by the teacher's generation quality. More importantly, we have observed that DMD optimizations tend to be unstable and have the tendency towards mode collapse, when trained under multi-view video data. This may be due to the inherent concept of DMD loss optimization being the reverse KL-divergence minimization, which tends to be mode-seeking and zeroing out the modes of teacher's that are not relatively dominant." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 205, + 553, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 553, + 418 + ], + "type": "text", + "content": "To solve the previous problems with VDM acceleration, we propose a novel framework, named FVGen, that performs novel views generation with a fast video diffusion generation with as few as 4 steps. First of all, different from previous GAN-based distillation methods [34, 51], we find it beneficial to initial the few-step student model by training a GAN objective, where the student model is regarded as a generator of fake samples, and, instead of training a separate discriminator as in previous works, we leverage a pre-trained teacher model as the discriminator. Second, we propose the soften reverse KL-divergence to solve the unstable optimization of DMD loss. The soften reverse KL-divergence maintains similar mode-seeking behavior as reverse KL-divergence but with more robust distillation: it prevents the student from ignoring entire space of teacher's distribution, thereby preserving more of teacher's knowledge, which we find particularly beneficial in the context of limited data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 421, + 553, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 553, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 553, + 528 + ], + "type": "text", + "content": "Through effective and robust distillation, FVGen is capable of generating video sequences with comparable or better visual quality as ViewCrafter [52] but with significantly less time. FVGen is trained with Dl3DV-10K [24] dataset and extensive experiments on real world dataset including MipNeRF360 [1] and Tanks-and-Temple [18] demonstrate that our method achieves overall better performance than other state-of-the-art (SOTA) novel view synthesis methods and diffusion distillation methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 542, + 406, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 542, + 406, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 406, + 555 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 563, + 512, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 512, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 512, + 575 + ], + "type": "text", + "content": "2.1. Diffusion-based Novel View Synthesis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "text", + "content": "NeRF [29] and 3DGS [16] excel at novel view synthesis with dense input views. However, reconstructing scenes from sparse views requires additional priors. Diffusion models [10, 32] can generate realistic pseudo input views from sparse original inputs. Several works use Score-Distillation Sampling [21, 30, 39, 42] to create 3D objects from text prompts or single images by distilling latent diffusion model [32] knowledge into 3D representations. While NeRF, 3DGS, and SDS typically require time-consuming per-scene optimization, view-dependent diffusion models [8, 19, 20, 35] offer an alternative. These mod" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "26096" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 72, + 312, + 175 + ], + "blocks": [ + { + "bbox": [ + 69, + 72, + 312, + 175 + ], + "lines": [ + { + "bbox": [ + 69, + 72, + 312, + 175 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 312, + 175 + ], + "type": "image", + "image_path": "1c45046d4adef763a2176185c0c5c2871cb8b26708cda36532c9666f1a6d6155.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 183, + 555, + 218 + ], + "lines": [ + { + "bbox": [ + 54, + 183, + 555, + 218 + ], + "spans": [ + { + "bbox": [ + 54, + 183, + 555, + 218 + ], + "type": "text", + "content": "Figure 2. Overview of student initialization. We initialize our student model by training a GAN objective. The student model " + }, + { + "bbox": [ + 54, + 183, + 555, + 218 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + }, + { + "bbox": [ + 54, + 183, + 555, + 218 + ], + "type": "text", + "content": ", initialized with the weights of teacher model, uses few-step denoising to generate fake videos to fool the discriminator " + }, + { + "bbox": [ + 54, + 183, + 555, + 218 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 54, + 183, + 555, + 218 + ], + "type": "text", + "content": ". The fixed teacher model, together with a trainable discriminator, are optimized to differentiate between fake and real videos." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 315, + 74, + 541, + 175 + ], + "blocks": [ + { + "bbox": [ + 315, + 74, + 541, + 175 + ], + "lines": [ + { + "bbox": [ + 315, + 74, + 541, + 175 + ], + "spans": [ + { + "bbox": [ + 315, + 74, + 541, + 175 + ], + "type": "image", + "image_path": "38940fdad6faed9e8b2fbb5423e64d5561e0ad715ae0fbf445b5a143f3873c3a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 228, + 294, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 228, + 294, + 264 + ], + "spans": [ + { + "bbox": [ + 54, + 228, + 294, + 264 + ], + "type": "text", + "content": "els directly generate 3D-consistent novel views conditioned on input images but directly lifting 2D images to 3D input therefore face challenges when scaling to larger scenes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 266, + 296, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 266, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 266, + 296, + 443 + ], + "type": "text", + "content": "An alternative approach leverages video diffusion models [6, 7, 25, 38, 52] for multi-view image generation. These methods fine-tune pre-trained latent video diffusion models (LVDM) [2, 4, 12, 44, 48] using multi-view images with both 2D and 3D guidance. The 2D guidance typically involves semantic features like CLIP [14] embeddings, while 3D guidance comes from sparse input cues. For instance, [52] uses DUSt3R [40] to build an initial point cloud and guides generation by combining point cloud renders with camera poses. Similarly, [6] conditions generation on feedforward 3DGS renders, enabling 360-degree scene generation with longer frame sequences through multi-step refinement. However, these methods suffer from speed issues caused by video diffusion sampling when generating novel views." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 455, + 176, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 176, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 176, + 467 + ], + "type": "text", + "content": "2.2. Diffusion Distillation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 474, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 474, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 474, + 296, + 713 + ], + "type": "text", + "content": "Diffusion models typically require numerous denoising steps to generate high-quality samples, which makes them computationally intensive. To accelerate generation, distillation methods train a student model that mimics the teacher diffusion model's behavior while using fewer sampling steps. Using fewer sampling steps reduces quality, so existing works incorporate adversarial training [3, 15, 22, 27, 33, 34, 46] to enhance the student model's output. However, simple GAN training cannot fully bridge the gap between student and teacher distributions, causing the student's generated output to remain noticeably different from the original diffusion model. To address this problem, Distribution Matching Distillation (DMD) [49, 51] optimizes the reverse KL-divergence. For tractable optimization, [49, 51] leverages the gradient of DMD loss, which represents the difference between student and teacher score functions. For fast video generation, [50] extends DMD to video diffusion models like CogVideoX [48]. While these methods work well for general image and video generation tasks, they suffer from mode collapse and training insta" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 228, + 555, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 251 + ], + "type": "text", + "content": "bility when applied to view synthesis on relatively small datasets like multi-view video datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 266, + 399, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 399, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 399, + 277 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 285, + 443, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 443, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 443, + 297 + ], + "type": "text", + "content": "3.1. Video Diffusion Model" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 304, + 555, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 555, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 555, + 413 + ], + "type": "text", + "content": "A video diffusion model is a generative model designed to synthesize realistic video sequences by learning the complex temporal and spatial patterns within video data. Based on the diffusion process, these models iteratively apply noise to video frames and then learn to reverse this process to generate new sequences. The forward diffusion process progressively adds Gaussian noise to the input clean sample " + }, + { + "bbox": [ + 313, + 304, + 555, + 413 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0 \\sim p(\\pmb{x})" + }, + { + "bbox": [ + 313, + 304, + 555, + 413 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 313, + 304, + 555, + 413 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 304, + 555, + 413 + ], + "type": "text", + "content": " timesteps, leading to highly noisy data " + }, + { + "bbox": [ + 313, + 304, + 555, + 413 + ], + "type": "inline_equation", + "content": "\\pmb{x}_T \\sim \\mathcal{N}(0, I)" + }, + { + "bbox": [ + 313, + 304, + 555, + 413 + ], + "type": "text", + "content": ". This process is formulated as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 343, + 426, + 555, + 440 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 426, + 555, + 440 + ], + "spans": [ + { + "bbox": [ + 343, + 426, + 555, + 440 + ], + "type": "interline_equation", + "content": "q \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {t - 1}\\right) = \\mathcal {N} \\left(\\boldsymbol {x} _ {t}; \\sqrt {\\alpha_ {t}} \\boldsymbol {x} _ {t - 1}, (1 - \\alpha_ {t}) \\boldsymbol {I}\\right), \\tag {1}", + "image_path": "30ab431faec0899218b4cba2177e1a78ff316be80a7e9e42a179a7f8f697f98a.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 450, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 450, + 555, + 510 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 313, + 450, + 555, + 510 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 450, + 555, + 510 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t" + }, + { + "bbox": [ + 313, + 450, + 555, + 510 + ], + "type": "text", + "content": " are the noise strength and noisy data at timestep " + }, + { + "bbox": [ + 313, + 450, + 555, + 510 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 450, + 555, + 510 + ], + "type": "text", + "content": ". The model then learns a denoising function, often parameterized by a neural network " + }, + { + "bbox": [ + 313, + 450, + 555, + 510 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\pmb{x}_t,t)" + }, + { + "bbox": [ + 313, + 450, + 555, + 510 + ], + "type": "text", + "content": ", to reverse this process and generate realistic video frames, where the optimization function is defined by the MSE loss:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 322, + 521, + 553, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 521, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 322, + 521, + 553, + 552 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d i f f}} (\\theta) = \\mathbb {E} _ {t \\sim \\mathcal {U} (0, 1), \\epsilon \\sim \\mathcal {N} (0, I), \\boldsymbol {x} \\sim p, c} \\| \\epsilon - \\epsilon_ {\\theta} (\\boldsymbol {x} _ {t}, t, c) \\| ^ {2}, \\tag {2}", + "image_path": "2901d96472d51b25668093af8f0be2158953252d50eaf8b03cd027afa6928ddf.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 554, + 555, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 554, + 555, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 554, + 555, + 650 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 554, + 555, + 650 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 554, + 555, + 650 + ], + "type": "text", + "content": " is the condition embedding of diffusion model which is usually represented as a text or image prompt. Instead of generating sequences at full resolution, Latent Video Diffusion Model (LVDM) first encodes video data " + }, + { + "bbox": [ + 313, + 554, + 555, + 650 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\in \\mathbb{R}^{F \\times C \\times H \\times W}" + }, + { + "bbox": [ + 313, + 554, + 555, + 650 + ], + "type": "text", + "content": " into latent space using a pre-trained VAE [17] encoder: " + }, + { + "bbox": [ + 313, + 554, + 555, + 650 + ], + "type": "inline_equation", + "content": "\\pmb{z} = \\mathcal{E}(\\pmb{x})" + }, + { + "bbox": [ + 313, + 554, + 555, + 650 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 554, + 555, + 650 + ], + "type": "inline_equation", + "content": "\\pmb{z} \\in \\mathbb{R}^{F \\times C \\times h \\times w}" + }, + { + "bbox": [ + 313, + 554, + 555, + 650 + ], + "type": "text", + "content": ". The compression will mitigate computational complexity while maintaining video generation quality." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 658, + 496, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 496, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 496, + 672 + ], + "type": "text", + "content": "3.2. Distribution Matching Distillation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "Distribution matching distillation adopts the idea of variational score distillation [42] that is optimized to match the distribution of student and teacher model by minimizing the" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26097" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 68, + 545, + 210 + ], + "blocks": [ + { + "bbox": [ + 69, + 68, + 545, + 210 + ], + "lines": [ + { + "bbox": [ + 69, + 68, + 545, + 210 + ], + "spans": [ + { + "bbox": [ + 69, + 68, + 545, + 210 + ], + "type": "image", + "image_path": "bb23e5dca17e672a494befbd10bd2c91d30b95edd42880458fc1671f4374fa0a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 218, + 555, + 243 + ], + "lines": [ + { + "bbox": [ + 54, + 218, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 54, + 218, + 555, + 243 + ], + "type": "text", + "content": "Figure 3. Overview of Distribution Matching Distillation (DMD). The gradient is the difference between the teacher score and fake score w.r.t. the student model output. We apply soften reverse KL-divergence for stable training and avoiding mode-collapse." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 251, + 296, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 296, + 277 + ], + "type": "text", + "content": "reverse KL-divergence. Given a student model " + }, + { + "bbox": [ + 55, + 251, + 296, + 277 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 251, + 296, + 277 + ], + "type": "text", + "content": " parameterized by " + }, + { + "bbox": [ + 55, + 251, + 296, + 277 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 251, + 296, + 277 + ], + "type": "text", + "content": ", the gradient of reverse KL-divergence is" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 83, + 283, + 295, + 366 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 283, + 295, + 366 + ], + "spans": [ + { + "bbox": [ + 83, + 283, + 295, + 366 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\nabla \\mathcal {L} _ {\\mathrm {D M D}} = \\mathbb {E} _ {t} \\left(\\nabla_ {\\theta} \\mathbf {D} _ {\\text {R e v e r s e - K L}} \\left(p _ {\\text {f a k e}, t} \\| p _ {\\text {r e a l}, t}\\right)\\right) \\\\ = - \\mathbb {E} _ {t} \\left(\\left[ s _ {\\text {r e a l}} \\left(F \\left(G _ {\\theta} (\\boldsymbol {z}), t\\right), t\\right) \\right. \\right. \\\\ \\left. - s _ {\\text {f a k e}} \\left(F \\left(G _ {\\theta} (\\boldsymbol {z}), t\\right), t\\right) \\frac {d G _ {\\theta} (\\boldsymbol {z})}{d \\theta} \\right]\\left. \\right), \\tag {3} \\\\ \\end{array}", + "image_path": "f5551245e2e84aeb84df867906e44ec4d8966cb6a5694ec9c3deb29cb6b5bd6f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 373, + 296, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 296, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 296, + 433 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 373, + 296, + 433 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{real}}" + }, + { + "bbox": [ + 55, + 373, + 296, + 433 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 373, + 296, + 433 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{fake}}" + }, + { + "bbox": [ + 55, + 373, + 296, + 433 + ], + "type": "text", + "content": " are the real and fake distribution. " + }, + { + "bbox": [ + 55, + 373, + 296, + 433 + ], + "type": "inline_equation", + "content": "z \\sim \\mathcal{N}(0, I)" + }, + { + "bbox": [ + 55, + 373, + 296, + 433 + ], + "type": "text", + "content": " is the initial random Gaussian noise. " + }, + { + "bbox": [ + 55, + 373, + 296, + 433 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 373, + 296, + 433 + ], + "type": "text", + "content": " is the diffusion forward function determined by Eq. (1). " + }, + { + "bbox": [ + 55, + 373, + 296, + 433 + ], + "type": "inline_equation", + "content": "s_{\\mathrm{real}}" + }, + { + "bbox": [ + 55, + 373, + 296, + 433 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 373, + 296, + 433 + ], + "type": "inline_equation", + "content": "s_{\\mathrm{fake}}" + }, + { + "bbox": [ + 55, + 373, + 296, + 433 + ], + "type": "text", + "content": " are the real and fake score function and defined as follows according to [37]:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 70, + 435, + 295, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 435, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 70, + 435, + 295, + 460 + ], + "type": "interline_equation", + "content": "s \\left(\\boldsymbol {x} _ {t}, t\\right) = \\nabla_ {\\boldsymbol {x} _ {t}} \\log p _ {t} \\left(\\boldsymbol {x} _ {t}\\right) = - \\frac {\\boldsymbol {x} _ {t} - \\alpha_ {t} \\mu \\left(\\boldsymbol {x} _ {t} , t\\right)}{\\sigma_ {t} ^ {2}}, \\tag {4}", + "image_path": "0c4e03da9d987dfb670ab4d59b7f6bd65dafc89bd0df518c0343480945bb6f8b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 468, + 296, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 296, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 296, + 540 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 468, + 296, + 540 + ], + "type": "inline_equation", + "content": "p_t" + }, + { + "bbox": [ + 55, + 468, + 296, + 540 + ], + "type": "text", + "content": " is the distribution of the noisy sample " + }, + { + "bbox": [ + 55, + 468, + 296, + 540 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 468, + 296, + 540 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 468, + 296, + 540 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 468, + 296, + 540 + ], + "type": "text", + "content": " is the clean sample prediction of the diffusion model. While optimizing student model " + }, + { + "bbox": [ + 55, + 468, + 296, + 540 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + }, + { + "bbox": [ + 55, + 468, + 296, + 540 + ], + "type": "text", + "content": " with gradient descent formulated in Eq. (3), the gradients of " + }, + { + "bbox": [ + 55, + 468, + 296, + 540 + ], + "type": "inline_equation", + "content": "\\mu_{\\mathrm{real}}" + }, + { + "bbox": [ + 55, + 468, + 296, + 540 + ], + "type": "text", + "content": " are stopped and " + }, + { + "bbox": [ + 55, + 468, + 296, + 540 + ], + "type": "inline_equation", + "content": "\\mu_{\\mathrm{fake}}" + }, + { + "bbox": [ + 55, + 468, + 296, + 540 + ], + "type": "text", + "content": " is dynamically optimized with the output of " + }, + { + "bbox": [ + 55, + 468, + 296, + 540 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + }, + { + "bbox": [ + 55, + 468, + 296, + 540 + ], + "type": "text", + "content": " with regular diffusion loss similar to Eq. (2)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 549, + 111, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 549, + 111, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 549, + 111, + 562 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "We propose Fast Video Generation (FVGen) that distills the generation capability of a multi-step teacher 3D-VDM into a few-step student 3D-VDM. The teacher 3D-VDM used in our experiments is the novel-view synthesis model ViewCrafter [52]. From a set of observed images " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "I^{\\mathrm{obs}}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " and observed cameras " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\pi^{\\mathrm{obs}}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": ", ViewCrafter learns a conditional distribution " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "I^{\\mathrm{tgt}} \\sim p(I^{\\mathrm{tgt}}|I^{\\mathrm{obs}},\\pi^{\\mathrm{obs}},\\pi^{\\mathrm{tgt}})" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "I^{\\mathrm{tgt}} = \\{I_i'\\}_{i=1}^L" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " is a sequence of " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " images corresponding to the target views " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\pi^{\\mathrm{tgt}}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": ". The student model FVGen is trained to sample from the same conditional distribution with a much faster inference speed. The architectures of FVGen are shown in Figure 2 and 3. For notation simplicity purpose," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 251, + 553, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 553, + 287 + ], + "type": "text", + "content": "we ignore including " + }, + { + "bbox": [ + 313, + 251, + 553, + 287 + ], + "type": "inline_equation", + "content": "I^{\\mathrm{obs}}" + }, + { + "bbox": [ + 313, + 251, + 553, + 287 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 251, + 553, + 287 + ], + "type": "inline_equation", + "content": "\\pi^{\\mathrm{obs}}" + }, + { + "bbox": [ + 313, + 251, + 553, + 287 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 251, + 553, + 287 + ], + "type": "inline_equation", + "content": "\\pi^{\\mathrm{tgt}}" + }, + { + "bbox": [ + 313, + 251, + 553, + 287 + ], + "type": "text", + "content": " in the future formula as they are all the conditional inputs for both teacher and student diffusion model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 288, + 554, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 554, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 554, + 371 + ], + "type": "text", + "content": "We first initialize the parameters of student model by training with a generative adversarial network (GAN) objective, leveraging the teacher model as discriminator (Section 4.1). After that, we continue to train the student model by optimizing DMD loss and diffusion loss to minimize the soften reverse KL-divergence between the distribution of student and teacher model (Section 4.2)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 379, + 531, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 379, + 531, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 531, + 392 + ], + "type": "text", + "content": "4.1. Student Initialization with GAN Training" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 396, + 555, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 555, + 587 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 555, + 587 + ], + "type": "text", + "content": "Directly training student model with DMD (Eq. 3) will cause training collapse as it is difficult to naively distill multi-step denoising into few-step. [49, 50] proposed to create noise-latent pairs by an ODE solver [36] and minimize a regression loss with student model for initialization. However, we found that it is very time consuming to generate large amount of video data pairs and the student model will also be restricted by the teacher's limitations. Inspired by [23, 51], we propose to initialize the student model by training a GAN objective with the real images. Specifically, the student model " + }, + { + "bbox": [ + 313, + 396, + 555, + 587 + ], + "type": "inline_equation", + "content": "(G_{\\theta})" + }, + { + "bbox": [ + 313, + 396, + 555, + 587 + ], + "type": "text", + "content": " generates fake video samples that fool the teacher model who serves as a discriminator (denoted as " + }, + { + "bbox": [ + 313, + 396, + 555, + 587 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 396, + 555, + 587 + ], + "type": "text", + "content": ") by minimizing the generator loss " + }, + { + "bbox": [ + 313, + 396, + 555, + 587 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_G" + }, + { + "bbox": [ + 313, + 396, + 555, + 587 + ], + "type": "text", + "content": ". The discriminator " + }, + { + "bbox": [ + 313, + 396, + 555, + 587 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 396, + 555, + 587 + ], + "type": "text", + "content": " classifies real samples from generated fake samples by maximizing " + }, + { + "bbox": [ + 313, + 396, + 555, + 587 + ], + "type": "inline_equation", + "content": "-\\mathcal{L}_D" + }, + { + "bbox": [ + 313, + 396, + 555, + 587 + ], + "type": "text", + "content": ". The min-max game adversarial optimization is formulated as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 343, + 590, + 553, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 590, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 343, + 590, + 553, + 634 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal{L}_{D} = \\mathbb{E}_{\\substack{\\boldsymbol {x}\\sim p_{\\text{real}}\\\\ t\\sim \\mathcal{U}[0,T]}}[\\log f\\left(D(F(\\boldsymbol {x},t))\\right)] \\\\ - \\mathbb {E} _ {\\substack {\\boldsymbol {z} \\sim \\mathcal {N} (0, I) \\\\ t \\sim \\mathcal {U} [ 0, T ]}} \\left[ \\log f \\left(D \\left(F \\left(G _ {\\theta} (\\boldsymbol {z}), t\\right)\\right)\\right) \\right], \\tag{5} \\\\ \\end{array}", + "image_path": "f9982087c8b33c6dbf29ab78483487f142dc612c71cd537399a3838fb0f442bd.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 344, + 636, + 525, + 659 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 636, + 525, + 659 + ], + "spans": [ + { + "bbox": [ + 344, + 636, + 525, + 659 + ], + "type": "interline_equation", + "content": "\\mathcal{L}_{G} = \\operatorname{\\mathbb{E}}_{\\substack{\\boldsymbol {z}\\sim \\mathcal{N}(0,\\boldsymbol {I})\\\\ t\\sim \\mathcal{U}[0,T]}}[\\log f\\left(D(F(G_{\\theta}(\\boldsymbol {z}),t))\\right)],", + "image_path": "8686b8e352b9d212758eabd45e367aeb88495cebd1f6b176b2b67a6110e85318.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": " is the time-step sampling and " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": " is the forward noise-injection model similar to Eq. (3). Discriminator " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": " retrieves the output feature map of the middle block of diffusion model UNet and passes it into the classifier " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": ". To" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26098" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": "fit with video data, " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": " is composed of a 3D Convolutional Neural Network that compresses the feature map into classification logits. Here, we use the original sample " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": " as the input to the discriminator instead of the generated sample " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": " from teacher model such that the student model would not be restricted by the generation capability of teacher model. For the purpose of training stability, recent methods [6, 52] predict the velocity field " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "\\pmb{v}_{\\theta}" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": ", and we convert it into the sample prediction with " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0 = \\sqrt{\\alpha_t\\pmb{x}_t} -\\sqrt{1 - \\alpha_t}\\pmb{v}_{\\theta}" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": ". Although the generated video samples are still very blurry, the GAN training scheme provides an effective initialization of the student model, which paves the way for more effective training with distribution matching loss." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 233, + 296, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 233, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 233, + 296, + 258 + ], + "type": "text", + "content": "4.2. Distribution Matching with Soften Reverse KL-Divergence Minimization" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 262, + 296, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 296, + 453 + ], + "type": "text", + "content": "As discussed in Section 3.2, the core concept of distribution matching distillation (DMD) [49, 51] is to minimize the reverse KL-divergence between student and teacher distributions. While reverse KL is popular for its mode-seeking focus, it has notable limitations in the context of generative modeling, its tendency toward mode collapse. Since reverse KL rewards the student for zeroing out probability in any region where the teacher's density is low, the student can effectively ignore parts where the teacher's distribution are not dominant. If the teacher's distribution " + }, + { + "bbox": [ + 55, + 262, + 296, + 453 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{real}}(\\pmb{x})" + }, + { + "bbox": [ + 55, + 262, + 296, + 453 + ], + "type": "text", + "content": " has multiple modes of varying height, minimizing ReverseKL " + }, + { + "bbox": [ + 55, + 262, + 296, + 453 + ], + "type": "inline_equation", + "content": "(p_{\\mathrm{fake}} \\| p_{\\mathrm{real}})" + }, + { + "bbox": [ + 55, + 262, + 296, + 453 + ], + "type": "text", + "content": " may lead " + }, + { + "bbox": [ + 55, + 262, + 296, + 453 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{fake}}" + }, + { + "bbox": [ + 55, + 262, + 296, + 453 + ], + "type": "text", + "content": " to concentrate on a subset of those modes and assign negligible mass to others. Therefore, the exclusive nature of reverse KL-divergence makes it prone to mode dropping and a lack of coverage of the full distribution." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 454, + 296, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 454, + 296, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 296, + 538 + ], + "type": "text", + "content": "To address the shortcomings of pure reverse KL, we propose to leverage a Soften reverse KL-divergence (Soften RKL) [9, 47]. Soften RKL refers to a modified divergence that retains the mode-seeking bias of reverse KL but softens its exclusion other relatively lower-probability regions, thereby mitigating mode collapse. Compared to reverse KL formulated in Eq. (3), the Soften RKL can be formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 543, + 294, + 608 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 543, + 294, + 608 + ], + "spans": [ + { + "bbox": [ + 58, + 543, + 294, + 608 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {D} _ {\\text {S o f t e n - R K L}} \\left(p _ {\\text {f a k e}, t} \\| p _ {\\text {r e a l}, t}\\right) = \\mathbf {D} _ {K L} \\left(\\frac {1}{2} p _ {\\text {r e a l}} + \\frac {1}{2} p _ {\\text {f a k e}} \\| p _ {\\text {r e a l}}\\right), \\\\ = p _ {\\text {f a k e}} (r + 1) \\log \\left(\\frac {1}{2} + \\frac {1}{2 r}\\right), \\tag {6} \\\\ \\end{array}", + "image_path": "ae5309cca43101c723523603790b75819af3d0a9eac7e39546910e0b4876147a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 613, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 613, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 613, + 296, + 685 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 613, + 296, + 685 + ], + "type": "inline_equation", + "content": "r = r(\\pmb{x},t) \\coloneqq p_{\\mathrm{real},t}(\\pmb{x}) / p_{\\mathrm{fake},t}(\\pmb{x})" + }, + { + "bbox": [ + 55, + 613, + 296, + 685 + ], + "type": "text", + "content": " is the density ratio. Here " + }, + { + "bbox": [ + 55, + 613, + 296, + 685 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 55, + 613, + 296, + 685 + ], + "type": "text", + "content": " denotes the noise-injection version of the output of student model formulated as " + }, + { + "bbox": [ + 55, + 613, + 296, + 685 + ], + "type": "inline_equation", + "content": "F(G_{\\theta}(\\pmb{z}),t)" + }, + { + "bbox": [ + 55, + 613, + 296, + 685 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 613, + 296, + 685 + ], + "type": "inline_equation", + "content": "t \\sim \\mathcal{U}(0,T)" + }, + { + "bbox": [ + 55, + 613, + 296, + 685 + ], + "type": "text", + "content": ". Thanks to our optimized GAN objective in Section 4.1, the density ratio could be approximated by the following:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 93, + 689, + 296, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 689, + 296, + 717 + ], + "spans": [ + { + "bbox": [ + 93, + 689, + 296, + 717 + ], + "type": "interline_equation", + "content": "r (\\boldsymbol {x}, t) = \\frac {p _ {\\text {r e a l} , t} (\\boldsymbol {x})}{p _ {\\text {f a k e} , t} (\\boldsymbol {x})} \\approx \\frac {f (D (\\boldsymbol {x} , t))}{1 - f (D (\\boldsymbol {x} , t))}. \\tag {7}", + "image_path": "ae0aed14687d28257c07592ee6e16af147a8a078fff12b10ee5fba3e0b59ce5c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": "In summary, the well-optimized GAN objective not only provides an efficient weight initialization of the student model, but also provides a direct estimate of the density ratio. Here, instead of directly comparing " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{fake}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{real}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": " as in " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{KL}(p_{\\mathrm{fake}}||p_{\\mathrm{real}})" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": ", we compare an even mixture of " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{real}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{fake}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{real}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": ". Soften RKL maintains the mode-seeking property of reverse KL, while penalizes more on mode dropping. This criterion yields a more robust distillation since it prevents the student from ignoring the entire teacher distribution, thereby preserving more of teacher's knowledge. Through experiments, we find that the soften RKL yields more stable training and more realistic generations. Computing the divergence directly is generally intractable, the gradient with respect to " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": " is formulated as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 330, + 249, + 553, + 298 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 249, + 553, + 298 + ], + "spans": [ + { + "bbox": [ + 330, + 249, + 553, + 298 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\nabla \\mathcal {L} _ {\\mathrm {D M D}} ^ {\\text {S o f t t e n - R K L}} = \\tag {8} \\\\ - \\mathbb {E} _ {t} \\left(\\frac {1}{r (\\boldsymbol {x} , t)} \\left[ s _ {\\text {r e a l}} (\\boldsymbol {x}, t) - s _ {\\text {f a k e}} (\\boldsymbol {x}, t) \\right] \\frac {d G _ {\\theta} (\\boldsymbol {z})}{d \\theta}\\right), \\\\ \\end{array}", + "image_path": "71f612cbdb1b7ac19e536359afbeb107b19286b023cfaaebd0e4e99ccf7bccb6.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 308, + 555, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 555, + 381 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 308, + 555, + 381 + ], + "type": "inline_equation", + "content": "\\pmb{x} = F(G_{\\theta}(z),t)" + }, + { + "bbox": [ + 313, + 308, + 555, + 381 + ], + "type": "text", + "content": " as previously defined. Apart from the soften RKL divergence optimization, we dynamically update the fake diffusion model " + }, + { + "bbox": [ + 313, + 308, + 555, + 381 + ], + "type": "inline_equation", + "content": "\\mu_{\\mathrm{fake}}" + }, + { + "bbox": [ + 313, + 308, + 555, + 381 + ], + "type": "text", + "content": " to adjust to change of student distribution in accordance with the original setting of DMD [49]. The loss term is identical to Eq. (2). We summarize the full training procedure in Alg. 1" + } + ] + } + ], + "index": 9 + }, + { + "type": "code", + "bbox": [ + 315, + 408, + 554, + 443 + ], + "blocks": [ + { + "bbox": [ + 315, + 392, + 480, + 405 + ], + "lines": [ + { + "bbox": [ + 315, + 392, + 480, + 405 + ], + "spans": [ + { + "bbox": [ + 315, + 392, + 480, + 405 + ], + "type": "text", + "content": "Algorithm 1 FVGen Training Procedure" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 315, + 408, + 554, + 443 + ], + "lines": [ + { + "bbox": [ + 315, + 408, + 554, + 443 + ], + "spans": [ + { + "bbox": [ + 315, + 408, + 554, + 443 + ], + "type": "text", + "content": "Require: Few-step timesteps " + }, + { + "bbox": [ + 315, + 408, + 554, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{0, t_1, t_2, \\dots, t_Q\\}" + }, + { + "bbox": [ + 315, + 408, + 554, + 443 + ], + "type": "text", + "content": ", pretrained teacher model " + }, + { + "bbox": [ + 315, + 408, + 554, + 443 + ], + "type": "inline_equation", + "content": "\\mu_{\\mathrm{real}}" + }, + { + "bbox": [ + 315, + 408, + 554, + 443 + ], + "type": "text", + "content": ", discriminator classifier " + }, + { + "bbox": [ + 315, + 408, + 554, + 443 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 315, + 408, + 554, + 443 + ], + "type": "text", + "content": " dataset " + }, + { + "bbox": [ + 315, + 408, + 554, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 315, + 408, + 554, + 443 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "code_body" + } + ], + "index": 11, + "sub_type": "code", + "guess_lang": "txt" + }, + { + "bbox": [ + 319, + 445, + 553, + 694 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 321, + 445, + 487, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 445, + 487, + 456 + ], + "spans": [ + { + "bbox": [ + 321, + 445, + 487, + 456 + ], + "type": "text", + "content": "1: Initialize student model " + }, + { + "bbox": [ + 321, + 445, + 487, + 456 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + }, + { + "bbox": [ + 321, + 445, + 487, + 456 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 321, + 445, + 487, + 456 + ], + "type": "inline_equation", + "content": "\\mu_{\\mathrm{real}}" + }, + { + "bbox": [ + 321, + 445, + 487, + 456 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 321, + 457, + 493, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 457, + 493, + 468 + ], + "spans": [ + { + "bbox": [ + 321, + 457, + 493, + 468 + ], + "type": "text", + "content": "2: Initialize fake score function with " + }, + { + "bbox": [ + 321, + 457, + 493, + 468 + ], + "type": "inline_equation", + "content": "\\mu_{\\mathrm{fake}}" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 321, + 470, + 403, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 470, + 403, + 480 + ], + "spans": [ + { + "bbox": [ + 321, + 470, + 403, + 480 + ], + "type": "text", + "content": "3: while training do" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 481, + 498, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 481, + 498, + 493 + ], + "spans": [ + { + "bbox": [ + 321, + 481, + 498, + 493 + ], + "type": "text", + "content": "4: Sample a video from dataset " + }, + { + "bbox": [ + 321, + 481, + 498, + 493 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0 \\sim \\mathcal{D}" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 493, + 553, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 493, + 553, + 516 + ], + "spans": [ + { + "bbox": [ + 321, + 493, + 553, + 516 + ], + "type": "text", + "content": "5: Add noise with timestep " + }, + { + "bbox": [ + 321, + 493, + 553, + 516 + ], + "type": "inline_equation", + "content": "t \\sim \\mathcal{T}" + }, + { + "bbox": [ + 321, + 493, + 553, + 516 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 321, + 493, + 553, + 516 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t = \\sqrt{\\alpha_t} \\pmb{x}_0 + \\sqrt{1 - \\alpha_t} \\pmb{\\epsilon}, \\quad \\pmb{\\epsilon} \\sim \\mathcal{N}(0, \\pmb{I})" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 517, + 496, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 517, + 496, + 528 + ], + "spans": [ + { + "bbox": [ + 321, + 517, + 496, + 528 + ], + "type": "text", + "content": "6: Predict with student: " + }, + { + "bbox": [ + 321, + 517, + 496, + 528 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}_0 = G_\\theta (\\pmb {x}_t,t)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 529, + 553, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 529, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 321, + 529, + 553, + 552 + ], + "type": "text", + "content": "7: Add noise " + }, + { + "bbox": [ + 321, + 529, + 553, + 552 + ], + "type": "inline_equation", + "content": "\\epsilon' \\sim \\mathcal{N}(0, I)" + }, + { + "bbox": [ + 321, + 529, + 553, + 552 + ], + "type": "text", + "content": " with timestep " + }, + { + "bbox": [ + 321, + 529, + 553, + 552 + ], + "type": "inline_equation", + "content": "\\tau \\sim \\mathcal{U}(0, T)" + }, + { + "bbox": [ + 321, + 529, + 553, + 552 + ], + "type": "text", + "content": " to real and fake sample:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 552, + 461, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 552, + 461, + 565 + ], + "spans": [ + { + "bbox": [ + 321, + 552, + 461, + 565 + ], + "type": "text", + "content": "8: " + }, + { + "bbox": [ + 321, + 552, + 461, + 565 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}_{\\tau} = \\sqrt{\\alpha_{\\tau}}\\hat{\\pmb{x}}_0 + \\sqrt{1 - \\alpha_{\\tau}}\\pmb{\\epsilon}'" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 566, + 460, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 566, + 460, + 577 + ], + "spans": [ + { + "bbox": [ + 321, + 566, + 460, + 577 + ], + "type": "text", + "content": "9: " + }, + { + "bbox": [ + 321, + 566, + 460, + 577 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{\\tau} = \\sqrt{\\alpha_{\\tau}}\\pmb{x}_0 + \\sqrt{1 - \\alpha_{\\tau}}\\pmb{\\epsilon}'" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 577, + 471, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 577, + 471, + 588 + ], + "spans": [ + { + "bbox": [ + 321, + 577, + 471, + 588 + ], + "type": "text", + "content": "10: Update " + }, + { + "bbox": [ + 321, + 577, + 471, + 588 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + }, + { + "bbox": [ + 321, + 577, + 471, + 588 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 321, + 577, + 471, + 588 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 321, + 577, + 471, + 588 + ], + "type": "text", + "content": " with Eq. (5)." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 590, + 375, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 590, + 375, + 599 + ], + "spans": [ + { + "bbox": [ + 321, + 590, + 375, + 599 + ], + "type": "text", + "content": "11: end while" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 319, + 601, + 441, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 601, + 441, + 612 + ], + "spans": [ + { + "bbox": [ + 319, + 601, + 441, + 612 + ], + "type": "text", + "content": "12: Output trained student " + }, + { + "bbox": [ + 319, + 601, + 441, + 612 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 319, + 613, + 403, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 613, + 403, + 624 + ], + "spans": [ + { + "bbox": [ + 319, + 613, + 403, + 624 + ], + "type": "text", + "content": "13: while training do" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 319, + 624, + 416, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 624, + 416, + 635 + ], + "spans": [ + { + "bbox": [ + 319, + 624, + 416, + 635 + ], + "type": "text", + "content": "14: Repeat step 4 - 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 319, + 636, + 493, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 636, + 493, + 647 + ], + "spans": [ + { + "bbox": [ + 319, + 636, + 493, + 647 + ], + "type": "text", + "content": "15: Update " + }, + { + "bbox": [ + 319, + 636, + 493, + 647 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + }, + { + "bbox": [ + 319, + 636, + 493, + 647 + ], + "type": "text", + "content": " with DMD Loss: Eq. (8)" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 319, + 648, + 550, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 648, + 550, + 660 + ], + "spans": [ + { + "bbox": [ + 319, + 648, + 550, + 660 + ], + "type": "text", + "content": "16: Add noise " + }, + { + "bbox": [ + 319, + 648, + 550, + 660 + ], + "type": "inline_equation", + "content": "\\pmb{\\epsilon}'' \\sim \\mathcal{N}(0, \\pmb{I})" + }, + { + "bbox": [ + 319, + 648, + 550, + 660 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 319, + 648, + 550, + 660 + ], + "type": "inline_equation", + "content": "t_1 \\sim \\mathcal{U}(0, T)" + }, + { + "bbox": [ + 319, + 648, + 550, + 660 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 319, + 648, + 550, + 660 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}_0" + }, + { + "bbox": [ + 319, + 648, + 550, + 660 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 319, + 661, + 471, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 661, + 471, + 673 + ], + "spans": [ + { + "bbox": [ + 319, + 661, + 471, + 673 + ], + "type": "text", + "content": "17: " + }, + { + "bbox": [ + 319, + 661, + 471, + 673 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}_{t_1} = \\sqrt{\\alpha_{t_1}}\\hat{\\pmb{x}}_0 + \\sqrt{1 - \\alpha_{t_1}}\\pmb{\\epsilon}^{\\prime \\prime}" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 319, + 673, + 511, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 673, + 511, + 685 + ], + "spans": [ + { + "bbox": [ + 319, + 673, + 511, + 685 + ], + "type": "text", + "content": "18: Update " + }, + { + "bbox": [ + 319, + 673, + 511, + 685 + ], + "type": "inline_equation", + "content": "\\mu_{\\mathrm{fake}}" + }, + { + "bbox": [ + 319, + 673, + 511, + 685 + ], + "type": "text", + "content": " with denoising loss: Eq. (2)" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 319, + 685, + 375, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 685, + 375, + 694 + ], + "spans": [ + { + "bbox": [ + 319, + 685, + 375, + 694 + ], + "type": "text", + "content": "19: end while" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26099" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 73, + 553, + 319 + ], + "blocks": [ + { + "bbox": [ + 57, + 73, + 553, + 319 + ], + "lines": [ + { + "bbox": [ + 57, + 73, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 57, + 73, + 553, + 319 + ], + "type": "image", + "image_path": "a5173ad0018a8a98b79691071d9dc58bb5ed4d7e6d02a33f35ce1446b8ba633e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 327, + 555, + 351 + ], + "lines": [ + { + "bbox": [ + 55, + 327, + 555, + 351 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 555, + 351 + ], + "type": "text", + "content": "Figure 4. Qualitative Results. We compare FVGen with SOTA novel view synthesis method on MipNeRF360 [1] and Tanks-and-Temples [18] datasets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 360, + 137, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 137, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 137, + 373 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 380, + 163, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 163, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 163, + 392 + ], + "type": "text", + "content": "5.1. Experiment Setup" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 397, + 296, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 397, + 296, + 565 + ], + "spans": [ + { + "bbox": [ + 54, + 397, + 296, + 565 + ], + "type": "text", + "content": "Implementation Details. All of our student model, teacher model and fake score function is initialized with the parameters of the sparse model of ViewCrafter [52]. We first initialize the student model by training GAN objective. The parameters of teacher model is fixed, and the student model and GAN classifier are trained for 4000 iterations with a two-scale update rule inspired by [51]. Next, we optimize DMD loss by continuously training student model and the fake score function for another 5000 iterations with the same two-scale update rule. The whole pipeline is trained on 8 NVIDIA H100 with batch size 4. The training last about 1 day. We utilize AdamW optimizer [26] and set learning rates of all optimizers to be " + }, + { + "bbox": [ + 54, + 397, + 296, + 565 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-5}" + }, + { + "bbox": [ + 54, + 397, + 296, + 565 + ], + "type": "text", + "content": ". The model is trained on images with a resolution of " + }, + { + "bbox": [ + 54, + 397, + 296, + 565 + ], + "type": "inline_equation", + "content": "512 \\times 320" + }, + { + "bbox": [ + 54, + 397, + 296, + 565 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "Datasets. FVGen is trained with DL3DV-10K [24], a real-world, scene-level video dataset that includes more than 10K long video clips. Similar to [52], we construct point cloud and ground truth video pairs using DL3DV dataset. We randomly sample short video clips with a random frame stride less than 5. The first frame and last frame are used to construct the prior point cloud with DUSt3R [40]. We use PyTorch3d [31], an efficient renderer to construct the point cloud renders with the camera poses of intermediate frames. Through this process, we create 20,000 training data pairs. We validate our method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 361, + 553, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 553, + 385 + ], + "type": "text", + "content": "on two public datasets Tanks-and-Temple [18] and MipN-eRF360 [1]" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 399, + 555, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 555, + 531 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 555, + 531 + ], + "type": "text", + "content": "Baseline Methods and Evaluation Metrics. To evaluate our proposed method, we compare our method with several state-of-the-art methods: ViewCrafter [52], MVsplat360 [6]. We also compare our video diffusion distillation method with two state-of-the-art methods: CausVid [50] and DMD2 [51]. We use PSNR [13], SSIM [41], LPIPS [53] to measure reconstruction quality. In addition, we also reported the distribution metric, i.e. Frechet Inception Distance (FID), which compares the distribution of generated views and ground truth views. Apart from the visual quality comparison," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 538, + 497, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 497, + 551 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 497, + 551 + ], + "type": "text", + "content": "5.2. Novel View Synthesis Comparison" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 555, + 555, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 555, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 555, + 639 + ], + "type": "text", + "content": "Similar to the evaluation setting of MVSplat360 [6], for each scene of both MipNeRF360 [1] and Tanks-and-Temples [18], we select 5 views that are far from each other but could cover the whole scene. We use our method and baseline method to generate 56 views sampled from the natural camera trajectory (14 views between each 2 input views)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "content": "Qualitative Results The qualitative comparisons with SOTA novel view synthesis methods are shown in Figure 4. DepthSplat is a feed-forward generalizable Gaussian splatting method. When there are relatively low overlap between input frames, DepthSplat would exhibit ob" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26100" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 89, + 119, + 156 + ], + "blocks": [ + { + "bbox": [ + 63, + 76, + 116, + 87 + ], + "lines": [ + { + "bbox": [ + 63, + 76, + 116, + 87 + ], + "spans": [ + { + "bbox": [ + 63, + 76, + 116, + 87 + ], + "type": "text", + "content": "Input Views" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 89, + 119, + 156 + ], + "lines": [ + { + "bbox": [ + 60, + 89, + 119, + 156 + ], + "spans": [ + { + "bbox": [ + 60, + 89, + 119, + 156 + ], + "type": "image", + "image_path": "3593d096fbd141b0bd7d7e363319734fea7531952e193b1721eaca22c323bdfe.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 119, + 89, + 227, + 156 + ], + "blocks": [ + { + "bbox": [ + 127, + 76, + 220, + 87 + ], + "lines": [ + { + "bbox": [ + 127, + 76, + 220, + 87 + ], + "spans": [ + { + "bbox": [ + 127, + 76, + 220, + 87 + ], + "type": "text", + "content": "ViewCrafter + DMD2" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 119, + 89, + 227, + 156 + ], + "lines": [ + { + "bbox": [ + 119, + 89, + 227, + 156 + ], + "spans": [ + { + "bbox": [ + 119, + 89, + 227, + 156 + ], + "type": "image", + "image_path": "3bc694cfe6a24c0a192576afb84f426f4a59e100418d05a94f94b7daa655dbfa.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 227, + 89, + 335, + 156 + ], + "blocks": [ + { + "bbox": [ + 229, + 76, + 330, + 87 + ], + "lines": [ + { + "bbox": [ + 229, + 76, + 330, + 87 + ], + "spans": [ + { + "bbox": [ + 229, + 76, + 330, + 87 + ], + "type": "text", + "content": "ViewCrafter + CausVid" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 227, + 89, + 335, + 156 + ], + "lines": [ + { + "bbox": [ + 227, + 89, + 335, + 156 + ], + "spans": [ + { + "bbox": [ + 227, + 89, + 335, + 156 + ], + "type": "image", + "image_path": "7ed713d6573a8d58858221afbe310e8d34b9d6b4d10f7d397f68892f5ba57fa9.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 335, + 89, + 442, + 156 + ], + "blocks": [ + { + "bbox": [ + 377, + 76, + 399, + 87 + ], + "lines": [ + { + "bbox": [ + 377, + 76, + 399, + 87 + ], + "spans": [ + { + "bbox": [ + 377, + 76, + 399, + 87 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 335, + 89, + 442, + 156 + ], + "lines": [ + { + "bbox": [ + 335, + 89, + 442, + 156 + ], + "spans": [ + { + "bbox": [ + 335, + 89, + 442, + 156 + ], + "type": "image", + "image_path": "ea21bd79ce63a0e33b213f2987fa8414ff3a7f703e5d8b418a01ca7ef1b79771.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 442, + 89, + 553, + 156 + ], + "blocks": [ + { + "bbox": [ + 468, + 76, + 526, + 87 + ], + "lines": [ + { + "bbox": [ + 468, + 76, + 526, + 87 + ], + "spans": [ + { + "bbox": [ + 468, + 76, + 526, + 87 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 442, + 89, + 553, + 156 + ], + "lines": [ + { + "bbox": [ + 442, + 89, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 442, + 89, + 553, + 156 + ], + "type": "image", + "image_path": "dc3664f6f3a5da2c21d8c2d8d3e2c57a9bb8fef29635d735152febda79fe914f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 59, + 157, + 119, + 234 + ], + "blocks": [ + { + "bbox": [ + 59, + 157, + 119, + 234 + ], + "lines": [ + { + "bbox": [ + 59, + 157, + 119, + 234 + ], + "spans": [ + { + "bbox": [ + 59, + 157, + 119, + 234 + ], + "type": "image", + "image_path": "b50b0fda2a94f4464748d47e6eed07c821be76a45a5a4b5f6e8d8b858b0a5dc6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 119, + 157, + 227, + 234 + ], + "blocks": [ + { + "bbox": [ + 119, + 157, + 227, + 234 + ], + "lines": [ + { + "bbox": [ + 119, + 157, + 227, + 234 + ], + "spans": [ + { + "bbox": [ + 119, + 157, + 227, + 234 + ], + "type": "image", + "image_path": "89e6b038db7333ff3d1125e105fcfaa2fa9b59f152c215a08a67e5062c40c368.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 227, + 157, + 335, + 234 + ], + "blocks": [ + { + "bbox": [ + 227, + 157, + 335, + 234 + ], + "lines": [ + { + "bbox": [ + 227, + 157, + 335, + 234 + ], + "spans": [ + { + "bbox": [ + 227, + 157, + 335, + 234 + ], + "type": "image", + "image_path": "9e23cd413eec3e8b8d44edfb1080969d2684934b0b1b23e24c0e6dba0ff6359e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 335, + 157, + 442, + 234 + ], + "blocks": [ + { + "bbox": [ + 335, + 157, + 442, + 234 + ], + "lines": [ + { + "bbox": [ + 335, + 157, + 442, + 234 + ], + "spans": [ + { + "bbox": [ + 335, + 157, + 442, + 234 + ], + "type": "image", + "image_path": "ae3ed8211ff5172647b4ff3d95890ab90dde83b558d0761947afd5ea651e1b62.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 442, + 157, + 553, + 234 + ], + "blocks": [ + { + "bbox": [ + 442, + 157, + 553, + 234 + ], + "lines": [ + { + "bbox": [ + 442, + 157, + 553, + 234 + ], + "spans": [ + { + "bbox": [ + 442, + 157, + 553, + 234 + ], + "type": "image", + "image_path": "caec3d14189f46947e91b50e8e2342b49914703da329894418d13f8b70ca80b0.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 59, + 235, + 119, + 312 + ], + "blocks": [ + { + "bbox": [ + 59, + 235, + 119, + 312 + ], + "lines": [ + { + "bbox": [ + 59, + 235, + 119, + 312 + ], + "spans": [ + { + "bbox": [ + 59, + 235, + 119, + 312 + ], + "type": "image", + "image_path": "83ca9a1348f07e041d5f0c085bfc8385339e29fc49b1fe86e58016e2517d53f8.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 119, + 235, + 227, + 312 + ], + "blocks": [ + { + "bbox": [ + 119, + 235, + 227, + 312 + ], + "lines": [ + { + "bbox": [ + 119, + 235, + 227, + 312 + ], + "spans": [ + { + "bbox": [ + 119, + 235, + 227, + 312 + ], + "type": "image", + "image_path": "6ef51fe1dbb98a0cbda6fc693ef3f0d17637e45f12f14cc00b990a96981e1582.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 227, + 235, + 335, + 312 + ], + "blocks": [ + { + "bbox": [ + 227, + 235, + 335, + 312 + ], + "lines": [ + { + "bbox": [ + 227, + 235, + 335, + 312 + ], + "spans": [ + { + "bbox": [ + 227, + 235, + 335, + 312 + ], + "type": "image", + "image_path": "40295a4d29f721bdcd815a5a3f759482433c5c75152c23602ccaf20578c22839.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 335, + 235, + 442, + 312 + ], + "blocks": [ + { + "bbox": [ + 335, + 235, + 442, + 312 + ], + "lines": [ + { + "bbox": [ + 335, + 235, + 442, + 312 + ], + "spans": [ + { + "bbox": [ + 335, + 235, + 442, + 312 + ], + "type": "image", + "image_path": "864656901dbeaefc5fed444ec8ff99d0930a3513424cc0da1f392923dc1d814c.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 442, + 235, + 553, + 312 + ], + "blocks": [ + { + "bbox": [ + 442, + 235, + 553, + 312 + ], + "lines": [ + { + "bbox": [ + 442, + 235, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 442, + 235, + 553, + 312 + ], + "type": "image", + "image_path": "c4fc1b7603d36534b0d1efd782c4e45483ba707c8ef5a511a93995deaac73bb6.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 59, + 312, + 119, + 380 + ], + "blocks": [ + { + "bbox": [ + 59, + 312, + 119, + 380 + ], + "lines": [ + { + "bbox": [ + 59, + 312, + 119, + 380 + ], + "spans": [ + { + "bbox": [ + 59, + 312, + 119, + 380 + ], + "type": "image", + "image_path": "6dcf859b41a54aa21d081b9996a0ec71f4c82ad19bd8825bd88c175249ab7949.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 389, + 555, + 412 + ], + "lines": [ + { + "bbox": [ + 55, + 389, + 555, + 412 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 555, + 412 + ], + "type": "text", + "content": "Figure 5. Qualitative Results. We compare FVGen with baseline diffusion distillation method on MipNeRF360 [1] and Tanks-and-Temples [18] datasets." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 119, + 312, + 227, + 380 + ], + "blocks": [ + { + "bbox": [ + 119, + 312, + 227, + 380 + ], + "lines": [ + { + "bbox": [ + 119, + 312, + 227, + 380 + ], + "spans": [ + { + "bbox": [ + 119, + 312, + 227, + 380 + ], + "type": "image", + "image_path": "9c98280ab97cbbb32f8aac3ed708707be30cbc0c540a411c6f2c711836b5a934.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 227, + 312, + 335, + 380 + ], + "blocks": [ + { + "bbox": [ + 227, + 312, + 335, + 380 + ], + "lines": [ + { + "bbox": [ + 227, + 312, + 335, + 380 + ], + "spans": [ + { + "bbox": [ + 227, + 312, + 335, + 380 + ], + "type": "image", + "image_path": "42119404f3c8ef2c82ceec63ea954ee11e42d5951782be40dd53f8b712c7fcb5.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 335, + 312, + 442, + 380 + ], + "blocks": [ + { + "bbox": [ + 335, + 312, + 442, + 380 + ], + "lines": [ + { + "bbox": [ + 335, + 312, + 442, + 380 + ], + "spans": [ + { + "bbox": [ + 335, + 312, + 442, + 380 + ], + "type": "image", + "image_path": "0c83a9787bb16cf32e23c4e8e5f7555014098b34faa93f9e28dddb519611a8aa.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 442, + 312, + 553, + 380 + ], + "blocks": [ + { + "bbox": [ + 442, + 312, + 553, + 380 + ], + "lines": [ + { + "bbox": [ + 442, + 312, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 442, + 312, + 553, + 380 + ], + "type": "image", + "image_path": "547326389256b57ba8ee5d23b49790ab3e0f4111f22c1a1b29634e05d7b31219.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 54, + 422, + 297, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 422, + 297, + 626 + ], + "spans": [ + { + "bbox": [ + 54, + 422, + 297, + 626 + ], + "type": "text", + "content": "vious artifacts due to inconsistent depth scale and the issue of floating Gaussians. In comparison, MVSplat360 [6] uses VDM to refine the Gaussian splattering renders and remove the floaters. However, the current MVSplat360 only supports training with low resolution inputs and the current results present significant blurriness. Compared to DepthSplat and MVSplat360, ViewCrafter [52] generates photo-realistic novel views with higher resolution. However, ViewCrafter is limited by the generation speed. Given generating 16 total frames, the average inference time is 13.2 seconds. In comparison, our method achieves a 10x speedup, generating the same number of frames in approximately 1 second and achieves similar visual results. The qualitative results prove that our method is able to distill the multi-step sampling into a few-step sampling process and achieve results that are similar to or even better than teacher ViewCrafter model with the student model." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "text", + "content": "Quantitative Analysis Table 1 presents quantitative comparisons on both datasets of FVGen and other novel view synthesis methods. Except for ViewCrafter [52], our method significantly surpasses other SOTA methods in terms of both perceptual and distribution metrics. Compared with ViewCrafter, our method yields similar quality" + } + ] + } + ], + "index": 27 + }, + { + "type": "table", + "bbox": [ + 325, + 421, + 545, + 538 + ], + "blocks": [ + { + "bbox": [ + 325, + 421, + 545, + 538 + ], + "lines": [ + { + "bbox": [ + 325, + 421, + 545, + 538 + ], + "spans": [ + { + "bbox": [ + 325, + 421, + 545, + 538 + ], + "type": "table", + "html": "
PSNR ↑SSIM ↑LPIPS ↓FID ↓Time ↓
mip-360DepthSplat11.230.2130.71532.454.2
MVSplat36012.280.2850.68225.6987.2
ViewCrafter16.350.3460.43316.2866.3
Ours16.280.3520.42917.445.1
TNTDepthSplat12.430.2630.67735.884.3
MVSplat36014.180.3010.53225.2387.3
ViewCrafter18.690.4020.20823.9465.9
Ours18.720.4110.21023.645.0
", + "image_path": "aeccd7fc704cf3c75f0643159c883bc9adf4435aeb31914530bf7335c5ee5b2c.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "table_body" + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 546, + 555, + 601 + ], + "lines": [ + { + "bbox": [ + 313, + 546, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 555, + 601 + ], + "type": "text", + "content": "Table 1. Quantitative Results. We report four quantitative metrics for image quality comparison on 2 separate datasets: Mip-NeRF360 [1] (denoted as mip-360 in the table) and Tanks-and-Temple [18] (denoted as TNT in the table). We highlight the best results in red and second-best in yellow." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 605, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 640 + ], + "type": "text", + "content": "but with more than " + }, + { + "bbox": [ + 313, + 605, + 555, + 640 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 313, + 605, + 555, + 640 + ], + "type": "text", + "content": " faster in generation speed. The quantitative results also highlights the effectiveness of our proposed method." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 647, + 495, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 495, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 495, + 661 + ], + "type": "text", + "content": "5.3. Diffusion Distillation Comparison" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "We also compare FVGen with SOTA diffusion distillation methods. Similar to the 2-view evaluation setting of DepthSplat [45] and MVSplat [5], we select 2 views that are far from each other in both MipNeRF360 [1] and Tanks-and-" + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26101" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": "Temples [18] dataset. We use our method and baseline method to generate 16 views samples from the natural camera trajectory (14 intermediate views and 2 input views)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 131, + 296, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 131, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 131, + 296, + 300 + ], + "type": "text", + "content": "Qualitative Results First we perform DMD2 [51] on ViewCrafter for few-step distillation. DMD2 applies end-to-end training of student model, GAN discriminator and fake score function. We find that this training scheme is not stable and present significant variance so that it is hard to converge, leading to blurry results as visually depicted in Figure 5. In addition, we compare our method with a recent work, Caus-Vid [50], on ViewCrafter distillation. CausVid [50] creates a small dataset with ODE solver and initialize the student model by optimizing a regression loss. Caus-Vid also leverages DMD [49] for 4-step video generation. However, the qualitative results indicate that our GAN object training scheme provides better initialization compared to the regression training scheme of ODE pairs." + } + ] + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 67, + 312, + 284, + 408 + ], + "blocks": [ + { + "bbox": [ + 67, + 312, + 284, + 408 + ], + "lines": [ + { + "bbox": [ + 67, + 312, + 284, + 408 + ], + "spans": [ + { + "bbox": [ + 67, + 312, + 284, + 408 + ], + "type": "table", + "html": "
PSNR ↑SSIM ↑LPIPS ↓FID ↓
mip-360VC+DMD29.290.1840.83635.59
VC+CausVid15.770.3360.44119.27
Ours16.280.3520.42917.44
TNTVC+DMD210.270.2090.73137.67
VC+CausVid17.330.4050.23224.92
Ours18.720.4110.21023.64
", + "image_path": "3202923bbcee39e56cd53b81833c3c74fee0dedfed137afd09a8516eb4c8a25d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 498, + 296, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 666 + ], + "type": "text", + "content": "Quantitative Analysis Table 2 presents quantitative comparisons with other diffusion distillation methods, and further proves the superiority of our method. Our method differs with DMD2 [51] in several ways: 1) we apply a 3D discriminator on video data, 2) we perform soften reverse-KL divergence for DMD optimization, 3) our GAN training and DMD training are kept separate for more stable training and more accurate density ratio calculation. Through training DMD2 [51], we found a unstable training process and we recorded the last model before the training collapsed. In comparison, CausVid [50] presents more stable training than DMD2 [51] but still suffer from mode collapse. Our method surpasses the other two baseline methods in both perceptual and distribution metrics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 678, + 156, + 690 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 678, + 156, + 690 + ], + "spans": [ + { + "bbox": [ + 55, + 678, + 156, + 690 + ], + "type": "text", + "content": "5.4. Ablation Studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 696, + 295, + 721 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 696, + 295, + 721 + ], + "spans": [ + { + "bbox": [ + 55, + 696, + 295, + 721 + ], + "type": "text", + "content": "In this section, we analyze the contributions of each module of our proposed method in Table 3 and Figure 6." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 323, + 70, + 547, + 138 + ], + "blocks": [ + { + "bbox": [ + 55, + 416, + 295, + 483 + ], + "lines": [ + { + "bbox": [ + 55, + 416, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 416, + 295, + 483 + ], + "type": "text", + "content": "Table 2. Quantitative Results. We report four quantitative metrics for image quality comparison on 2 separate datasets: Mip-NeRF360 [1] (denoted as mip-360 in the table) and Tanks-and-Temple [18] (denoted as TNT in the table). VC stands for ViewCrafter [52] for space saving purpose. " + }, + { + "bbox": [ + 55, + 416, + 295, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{VC} + \\mathbf{X}" + }, + { + "bbox": [ + 55, + 416, + 295, + 483 + ], + "type": "text", + "content": " means applying diffusion acceleration method X on ViewCrafter." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 70, + 547, + 138 + ], + "lines": [ + { + "bbox": [ + 323, + 70, + 547, + 138 + ], + "spans": [ + { + "bbox": [ + 323, + 70, + 547, + 138 + ], + "type": "table", + "html": "
GANDMDSoften RKLPSNR ↑SSIM ↑LPIPS ↓FID ↓
✗✓✓8.620.1540.88040.17
✓✗✗16.230.3690.37521.48
✓✓✗16.850.3850.33721.05
✓✓✓17.500.3820.32020.54
", + "image_path": "355022c16312b43a447ca8b9c11b1f4b65e1bb370356153c1d1e7fe0e691767c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 146, + 553, + 169 + ], + "lines": [ + { + "bbox": [ + 313, + 146, + 553, + 169 + ], + "spans": [ + { + "bbox": [ + 313, + 146, + 553, + 169 + ], + "type": "text", + "content": "Table 3. Ablation studies. Quantitative analysis of different components of FVGen." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 315, + 184, + 553, + 252 + ], + "blocks": [ + { + "bbox": [ + 315, + 184, + 553, + 252 + ], + "lines": [ + { + "bbox": [ + 315, + 184, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 315, + 184, + 553, + 252 + ], + "type": "image", + "image_path": "3c8c9a8a076a58ae51952b2685b688dc1f2745054087fc585e5bc404e5073279.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 259, + 555, + 293 + ], + "lines": [ + { + "bbox": [ + 313, + 259, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 259, + 555, + 293 + ], + "type": "text", + "content": "Figure 6. Ablation Studies. Visualization of the contribution of different components of FVGen. Our full model presents the most visual similarity between the ground truth." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 306, + 555, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 555, + 389 + ], + "type": "text", + "content": "Assessing student initialization. As it can be seen from the first column of Figure 6 and Table 3, student initialization is a pivotal module of our architecture. Without student initialization with GAN optimization, the student model would fail to generate samples close to the teacher distribution, thus the teacher score function is not reliable for distillation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 401, + 554, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 554, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 554, + 474 + ], + "type": "text", + "content": "Assessing distribution matching. Optimization of DMD continues to align the student model distribution with teacher model. Comparing with full model, although the perceptual metrics are not significantly degraded, qualitative results show that without distribution matching, the generated results appear significant blurriness and artifacts." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 487, + 554, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 554, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 554, + 545 + ], + "type": "text", + "content": "Assessing soften reverse KL-divergence. As discussed in Section 4.2, soften reverse KL-divergence enables more stable and robust training. Therefore, the DMD optimization achieves better convergence and generated results are more distilled into teacher distribution." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 560, + 389, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 560, + 389, + 572 + ], + "spans": [ + { + "bbox": [ + 313, + 560, + 389, + 572 + ], + "type": "text", + "content": "6. Limitations" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 582, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 712 + ], + "type": "text", + "content": "We identify several limitations in our current work. First, although FVGen achieves similar visual quality compared to ViewCrafter, it is limited by the drawbacks of ViewCrafter: e.g., degradation in structural integrity and consistency when experiencing extremely sparse inputs. Second, FVGen benefits from the synergy of three video diffusion models; therefore, due to computational constraints, we were only able to train the model with videos up to only 16 frames, which may not provide enough coverage for extremely large scenes. Therefore, our future work would address on the limitations mentioned above." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26102" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "text", + "content": "7. Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 223, + 115, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 223, + 115, + 236 + ], + "spans": [ + { + "bbox": [ + 56, + 223, + 115, + 236 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 243, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 243, + 296, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 243, + 296, + 287 + ], + "spans": [ + { + "bbox": [ + 61, + 243, + 296, + 287 + ], + "type": "text", + "content": "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-NeRF: A Multiscale Representation for Anti-Aliasing Neural Radiance Fields. Technical report, 2021. 2, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 288, + 296, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 288, + 296, + 353 + ], + "spans": [ + { + "bbox": [ + 61, + 288, + 296, + 353 + ], + "type": "text", + "content": "[2] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 354, + 296, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 354, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 62, + 354, + 296, + 399 + ], + "type": "text", + "content": "[3] Dar-Yen Chen, Hmrishav Bandyopadhyay, Kai Zou, and Yi-Zhe Song. Nitrofusion: High-fidelity single-step diffusion through dynamic adversarial training. arXiv preprint arXiv:2412.02030, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 399, + 296, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 399, + 296, + 465 + ], + "spans": [ + { + "bbox": [ + 62, + 399, + 296, + 465 + ], + "type": "text", + "content": "[4] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videographer2: Overcoming data limitations for high-quality video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7310-7320, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 467, + 296, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 467, + 296, + 520 + ], + "spans": [ + { + "bbox": [ + 62, + 467, + 296, + 520 + ], + "type": "text", + "content": "[5] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splattering from sparse multi-view images. arXiv preprint arXiv:2403.14627, 2024. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 522, + 296, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 522, + 296, + 566 + ], + "spans": [ + { + "bbox": [ + 62, + 522, + 296, + 566 + ], + "type": "text", + "content": "[6] Yuedong Chen, Chuanxia Zheng, Haofei Xu, Bohan Zhuang, Andrea Vedaldi, Tat-Jen Cham, and Jianfei Cai. Mvsplat360: Feed-forward 360 scene synthesis from sparse views. 2024. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 567, + 296, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 567, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 62, + 567, + 296, + 601 + ], + "type": "text", + "content": "[7] Zilong Chen, Yikai Wang, Feng Wang, Zhengyi Wang, and Huaping Liu. V3d: Video diffusion models are effective 3d generators. arXiv preprint arXiv:2403.06738, 2024. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 601, + 296, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 601, + 296, + 656 + ], + "spans": [ + { + "bbox": [ + 62, + 601, + 296, + 656 + ], + "type": "text", + "content": "[8] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T Barron, and Ben Poole. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv preprint arXiv:2405.10314, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 657, + 296, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 657, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 62, + 657, + 296, + 678 + ], + "type": "text", + "content": "[9] Dibya Ghosh. KI divergence - intuition and examples, n.d. Accessed: 2025-03-07. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 680, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 296, + 714 + ], + "type": "text", + "content": "[10] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 55, + 90, + 296, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 201 + ], + "type": "text", + "content": "Supported by the Intelligence Advanced Research Projects Activity (IARPA) via Department of Interior/ Interior Business Center (DOI/IBC) contract number 140D0423C0075. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon. Disclaimer: The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of IARPA, DOI/IBC, or the U.S. Government." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[11] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv:2204.03458, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 107, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 149 + ], + "type": "text", + "content": "[12] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022.3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 151, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 553, + 184 + ], + "type": "text", + "content": "[13] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssm. In 2010 20th international conference on pattern recognition, pages 2366-2369. IEEE, 2010. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 185, + 553, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 185, + 553, + 247 + ], + "spans": [ + { + "bbox": [ + 317, + 185, + 553, + 247 + ], + "type": "text", + "content": "[14] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 250, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 553, + 304 + ], + "type": "text", + "content": "[15] Minguk Kang, Richard Zhang, Connelly Barnes, Sylvain Paris, Suha Kwak, Jaesik Park, Eli Shechtman, Jun-Yan Zhu, and Taesung Park. Distilling diffusion models into conditional gans. In European Conference on Computer Vision, pages 428-447. Springer, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 305, + 553, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 553, + 349 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 553, + 349 + ], + "type": "text", + "content": "[16] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 350, + 553, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 371 + ], + "type": "text", + "content": "[17] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "type": "text", + "content": "[18] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36 (4):1-13, 2017. 2, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 416, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 553, + 459 + ], + "type": "text", + "content": "[19] Xin Kong, Shikun Liu, Xiaoyang Lyu, Marwan Taher, Xiaojuan Qi, and Andrew J Davison. Eschernet: A generative model for scalable view synthesis. arXiv preprint arXiv:2402.03908, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 460, + 553, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 553, + 504 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 553, + 504 + ], + "type": "text", + "content": "[20] Xinyang Li, Zhangyu Lai, Linning Xu, Yansong Qu, Liujuan Cao, Shengchuan Zhang, Bo Dai, and Rongrong Ji. Director3d: Real-world camera trajectory and 3d scene generation from text. arXiv preprint arXiv:2406.17601, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 505, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 505, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 317, + 505, + 553, + 569 + ], + "type": "text", + "content": "[21] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 571, + 553, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 571, + 553, + 603 + ], + "spans": [ + { + "bbox": [ + 317, + 571, + 553, + 603 + ], + "type": "text", + "content": "[22] Shanchuan Lin, Anran Wang, and Xiao Yang. Sdxlightning: Progressive adversarial diffusion distillation. arXiv preprint arXiv:2402.13929, 2024. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 604, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 604, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 317, + 604, + 553, + 646 + ], + "type": "text", + "content": "[23] Shanchuan Lin, Xin Xia, Yuxi Ren, Ceyuan Yang, Xuefeng Xiao, and Lu Jiang. Diffusion adversarial post-training for one-step video generation. arXiv preprint arXiv:2501.08316, 2025. 2, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 647, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 714 + ], + "type": "text", + "content": "[24] Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22160–22169, 2024. 2, 6" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26103" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "type": "text", + "content": "[25] Fangfu Liu, Wenqiang Sun, Hanyang Wang, Yikai Wang, Haowen Sun, Junliang Ye, Jun Zhang, and Yueqi Duan. Reconx: Reconstruct any scene from sparse views with video diffusion model. arXiv preprint arXiv:2408.16767, 2024. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 152 + ], + "type": "text", + "content": "[26] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "type": "text", + "content": "[27] Yihong Luo, Xiaolong Chen, Xinghua Qu, Tianyang Hu, and Jing Tang. You only sample once: Taming one-step text-to-image synthesis by self-cooperative diffusion gans. arXiv preprint arXiv:2403.12931, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 252 + ], + "type": "text", + "content": "[28] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14297-14306, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "text", + "content": "[29] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 299, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 294, + 331 + ], + "type": "text", + "content": "[30] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "type": "text", + "content": "[31] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 378, + 294, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 378, + 294, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 378, + 294, + 432 + ], + "type": "text", + "content": "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 434, + 294, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 488 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 488 + ], + "type": "text", + "content": "[33] Axel Sauer, Frederic Boesel, Tim Dockhorn, Andreas Blattmann, Patrick Esser, and Robin Rombach. Fast high-resolution image synthesis with latent adversarial diffusion distillation. In SIGGRAPH Asia 2024 Conference Papers, pages 1-11, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 490, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 294, + 533 + ], + "type": "text", + "content": "[34] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In European Conference on Computer Vision, pages 87-103. Springer, 2024. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 535, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 294, + 567 + ], + "type": "text", + "content": "[35] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 568, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 601 + ], + "type": "text", + "content": "[36] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 603, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 646 + ], + "type": "text", + "content": "[37] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 647, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 713 + ], + "type": "text", + "content": "[38] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. In European Conference on Computer Vision, pages 439-457. Springer, 2025. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "text", + "content": "[39] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12619-12629, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 182 + ], + "type": "text", + "content": "[40] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 184, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 227 + ], + "type": "text", + "content": "[41] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 228, + 553, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 281 + ], + "type": "text", + "content": "[42] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 283, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 553, + 348 + ], + "type": "text", + "content": "[43] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21551-21561, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 350, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 415 + ], + "type": "text", + "content": "[44] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Gongye Liu, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Dynamiccafter: Animating open-domain images with video diffusion priors. In European Conference on Computer Vision, pages 399-417. Springer, 2025. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 416, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 553, + 459 + ], + "type": "text", + "content": "[45] Haofei Xu, Songyou Peng, Fangjinhua Wang, Hermann Blum, Daniel Barath, Andreas Geiger, and Marc Pollefeys. Depthsplat: Connecting gaussian splatting and depth. arXiv preprint arXiv:2410.13862, 2024. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 460, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 514 + ], + "type": "text", + "content": "[46] Yanwu Xu, Yang Zhao, Zhisheng Xiao, and Tingbo Hou. Ufogen: You forward once large scale text-to-image generation via diffusion gans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8196-8206, 2024. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 515, + 553, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 553, + 548 + ], + "type": "text", + "content": "[47] Yilun Xu, Weili Nie, and Arash Vahdat. One-step diffusion models with " + }, + { + "bbox": [ + 316, + 515, + 553, + 548 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 316, + 515, + 553, + 548 + ], + "type": "text", + "content": "-divergence distribution matching. arXiv preprint arXiv:2502.15681, 2025. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 548, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 602 + ], + "type": "text", + "content": "[48] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024.3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 603, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 553, + 668 + ], + "type": "text", + "content": "[49] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6613-6623, 2024. 2, 3, 4, 5, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[50] Tianwei Yin, Qiang Zhang, Richard Zhang, William T Freeman, Fredo Durand, Eli Shechtman, and Xun Huang. From slow bidirectional to fast causal video generators. arXiv preprint arXiv:2412.07772, 2024. 2, 3, 4, 6, 8" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26104" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 239 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[51] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and Bill Freeman. Improved distribution matching distillation for fast image synthesis. Advances in Neural Information Processing Systems, 37: 47455-47487, 2025. 2, 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "type": "text", + "content": "[52] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 239 + ], + "type": "text", + "content": "[53] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26105" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/FW-Merging_ Scaling Model Merging with Frank-Wolfe Optimization/56e48ade-9cb7-4688-a0be-edb71b4d459b_content_list.json b/2025/FW-Merging_ Scaling Model Merging with Frank-Wolfe Optimization/56e48ade-9cb7-4688-a0be-edb71b4d459b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..13cc9aa73cc235cbdd7dc3ae93f8b9ecb6559fd4 --- /dev/null +++ b/2025/FW-Merging_ Scaling Model Merging with Frank-Wolfe Optimization/56e48ade-9cb7-4688-a0be-edb71b4d459b_content_list.json @@ -0,0 +1,1694 @@ +[ + { + "type": "text", + "text": "FW-Merging: Scaling Model Merging with Frank-Wolfe Optimization", + "text_level": 1, + "bbox": [ + 143, + 130, + 852, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hao (Mark) Chen1 Shell Xu Hu2 Wayne Luk1 Timothy Hospedales2 Hongxiang Fan1 \n1Imperial College London, UK 2Samsung AI Center, Cambridge, UK", + "bbox": [ + 120, + 179, + 870, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "hcl620@ic.ac.uk shell.hu@samsung.com w.luk@imperial.ac.uk t.hospedales@samsung.com hongxiang.fan@imperial.ac.uk", + "bbox": [ + 117, + 219, + 874, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 263, + 326, + 280 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Model merging has emerged as a promising approach for multi-task learning (MTL), offering a data-efficient alternative to conventional fine-tuning. However, with the rapid development of the open-source AI ecosystem and the increasing availability of fine-tuned foundation models, existing model merging methods face two key limitations: (i) They are primarily designed for in-house fine-tuned models, making them less adaptable to diverse model sources with partially unknown model and task information, (ii) They struggle to scale effectively when merging numerous model checkpoints. To address these challenges, we formulate model merging as a constrained optimization problem and introduce a novel approach: Frank-Wolfe Merging (FW-Merging). Inspired by Frank-Wolfe optimization, our approach iteratively selects the most relevant model in the pool to minimize a linear approximation of the objective function and then executes a local merging similar to the Frank-Wolfe update. The objective function is designed to capture the desired behavior of the target-merged model, while the fine-tuned candidate models define the constraint set. More importantly, FW-Merging serves as an orthogonal technique for existing merging methods, seamlessly integrating with them to further enhance accuracy performance. Our experiments show that FW-Merging scales across diverse model sources, remaining stable with 16 irrelevant models and improving by $15.3\\%$ with 16 relevant models on 20 CV tasks, while maintaining constant memory overhead—unlike the linear overhead of data-informed merging methods. Compared with the state-of-the-art approaches, FW-Merging surpasses the data-free merging method by $32.8\\%$ and outperforms the data-informed Adamerging by $8.39\\%$ when merging 20 ViT models. Our code is open-sourced at here.", + "bbox": [ + 89, + 295, + 485, + 794 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 821, + 220, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multi-task learning (MTL)-based fine-tuning adapts a single pre-trained Large Language Model (LLM) for multiple", + "bbox": [ + 89, + 847, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "downstream applications, reducing the deployment overhead of separately fine-tuning multiple models [53]. However, it still demands a large amount of high-quality data, which might only exist in the private domain [32], and significant compute resources [36]. To mitigate these issues, model merging has emerged as a promising technique for fusing fine-tuned models within the parameter space [14, 48, 50]. Existing model merging methods can be broadly classified into two categories: data-free methods [14, 15, 48], and data-informed methods [50, 51], which optimize merge coefficients based on additional data.", + "bbox": [ + 511, + 266, + 903, + 431 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While these approaches have proven effective, several key limitations hinder their scalability and broader adoption. First, these methods adjust merging coefficients based on the known capabilities of the models on specific tasks to optimize performance, which is less robust when dealing with diverse model sources with unknown information1. This is primarily caused by the inability to distinguish high-quality models from poorly fine-tuned ones in an unknown model setting. Second, when scaling up these approaches to a large number of unknown models, these methods struggle and can result in significant performance degradation. As shown in Figure 1a, our profiling study demonstrates a performance reduction ranging from $18.9\\%$ to $64.4\\%$ . These limitations are further amplified by the fast-growing open-source AI ecosystem, where platforms such as Hugging Face have driven a surge in the release of powerful LLMs with many lacking complete information. Given that merging open-source models has repeatedly shown the potential to produce top-ranking LLMs on major model leaderboards [13], developing scalable and robust merging techniques is essential to harness the growing number of open-source models, further enhancing performance and widening the adoption of model merging.", + "bbox": [ + 511, + 433, + 903, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To effectively scale model merging and leverage the vast collection of open-sourced models with unknown capabilities, the new model merging method must exhibit two fundamental scaling properties: 1) as more irrelevant models", + "bbox": [ + 511, + 780, + 903, + 840 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 805, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1This paper refers unknown information to: 1) when open-source models are partially assessed on limited benchmarks, leaving their performance on other tasks unknown and costly to evaluate, and 2) when models contain misleading information, polluting the merging process.", + "bbox": [ + 511, + 851, + 903, + 900 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "*Corresponding Authors: Shell Xu Hu and Hongxiang Fan", + "bbox": [ + 107, + 886, + 419, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3390", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3da232e398f85080f32a85987872d1e92462d5cc7f196443b585abe00fb321bb.jpg", + "image_caption": [ + "(a) 4 CV tasks with irrelevant model addition." + ], + "image_footnote": [], + "bbox": [ + 98, + 90, + 357, + 239 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e75d2e7c3bb354a233d93ddc1b22cb9fdf5316c7c4616b34d65c4b71078a47d7.jpg", + "image_caption": [ + "(b) 20 CV tasks with relevant model addition." + ], + "image_footnote": [], + "bbox": [ + 367, + 90, + 630, + 239 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/953a1664a8cf8cb344efc734e23da2d593146a26f1633a292b439cf709e91a12.jpg", + "image_caption": [ + "(c) NLP tasks with noisy model addition.", + "Figure 1. Performance scaling of FW-Merging across CV tasks. (a) demonstrate robustness to irrelevant models, while (b) show improved performance with relevant models. (c) analyzes performance degradation when incorporating a noisy model initialized from a different pre-trained checkpoint. Detailed results and experimental setup are discussed in Section 4.4." + ], + "image_footnote": [], + "bbox": [ + 638, + 92, + 898, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "are added to the merging pool, the performance should remain unaffected, and 2) as more relevant models are added to the merging pool, the performance should steadily increase, converging towards the optimal performance. To this end, we revisit model merging and formulate it as a constrained optimization problem, where the objective function dictates the desirable behavior of the final merged model, and fine-tuned checkpoints form the constraint set. Inspired by Frank-Wolfe optimization, we introduce Frank-Wolfe Merging (FW-Merging), an iterative algorithm designed to enhance merging efficiency while maintaining robustness at scale. FW-Merging comprises three principal stages in each iteration: (1) Relevance Evaluation: Instead of merging models arbitrarily, we obtain the linear approximation of the objective function using gradients of the current model, revealing the most beneficial direction for improvement. (2) Model Selection: The most relevant checkpoints are selected from the constraint set by minimizing the linear approximation, ensuring that each step incorporates task-specific knowledge with minimal interference. (3) Knowledge Integration: The selected checkpoint is integrated using an orthogonal merging method, striking a balance between adaptation and stability in the merged model.", + "bbox": [ + 91, + 323, + 483, + 670 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We demonstrate the effectiveness of FW-Merging with a diverse pool of fine-tuned checkpoints across various language and vision tasks, compared to both data-free and data-informed model merging methods as well as traditional MTL-based fine-tuning. As shown in Figure 1, FW-Merging satisfy our two fundamental scaling properties: accuracy performance does not drop when 16 irrelevant models are added (compared to a $49\\%$ drop in task-arithmetic) and steadily improves by $15.3\\%$ when 16 relevant models are included. Additionally, FW-Merging requires only constant memory overhead, as it selects and merges a fixed number of models at a time. In contrast, methods that optimize merging coefficients [50] or resolve parameter interference [48] must store all models in memory, leading to linear overhead. Moreover, FW-Merging", + "bbox": [ + 89, + 674, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "exhibits greater robustness to noisy models lacking critical information, such as their initialization point. As shown in Figure 1c, FW-Merging experiences minimal performance degradation when a misinitialized model is introduced, whereas Ties suffers a performance drop of up to $3.2\\%$ . FW-Merging outperforms state-of-the-art data-free merging method by $32.8\\%$ and the data-informed method Adamerging by $8.39\\%$ when merging 20 ViT models. On the language benchmarks, FW-Merging achieves $6.3\\%$ improvement over the best model merging method across discriminative and generative tasks, while even surpassing the performance of traditional MTL using only $3.4\\%$ of its required data. Our results position FW-Merging as an effective solution to scale model merging to the next level.", + "bbox": [ + 511, + 323, + 906, + 534 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 532, + 536, + 857, + 550 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Identify scalability and robustness issues in existing model merging techniques through experiments, highlighting the urgent need for large-scale model merging.", + "- Formulate model merging as a constrained optimization problem with an objective function that explicitly captures the desired behavior of the final merged model.", + "- Introduce Frank-Wolfe Merging, a novel iterative method that autonomously guides the merged model toward an optimized direction, even with large sets of black-box open-source checkpoints.", + "- Evaluate our proposed approach on extensive benchmarks, demonstrating its effectiveness and scalability." + ], + "bbox": [ + 513, + 551, + 903, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 752, + 653, + 768 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Efficient Multi-Task Learning. In traditional Multi-Task Learning (MTL), a single model is trained on a dataset containing multiple tasks to enable the model to acquire diverse capabilities [2]. However, a significant challenge in traditional MTL is the issue of negative transfer [16]. To mitigate this, architecture-based approaches have been developed, such as parameter sparsification [26, 39] and shared structure modularization [29, 30]. On the optimization side,", + "bbox": [ + 511, + 779, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3391", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "methods to resolve gradient conflicts [6, 54] and domination of gradient or learning rate [5, 24] have been proposed. With the rise of Large Language Models (LLMs), MTL faces additional challenges, particularly the high computational costs. To address these challenges, strategies like parameter-efficient fine-tuning [11, 20, 21] and memory-efficient fine-tuning [8, 22, 31] have been introduced to minimize both memory and computational resource usage. More recently, model merging has emerged as a promising approach to make MTL more compute- and data-efficient.", + "bbox": [ + 89, + 90, + 483, + 243 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Model Merging. While pre-merging methods prepare favorable conditions for merging, during-merging techniques combine multiple neural networks into a single model while retaining or enhancing their capabilities [49]. In this work, we focus on during-merging methods. Early insights into neural network landscapes [10] revealed that linear interpolation between models exposes useful loss surface properties, laying the foundation for weight averaging—a core merging technique. Simple averaging widens optima and improves generalization [15], evolving into advanced methods like model soups [46] and heterogeneous model merging. Recent advances introduce more structured approaches, such as Fisher-Weighted Averaging [37], which incorporates Fisher information to weight parameters more effectively, and Permutation Alignment methods like Git Re-Basin [1], which address weight permutation symmetries. Interference Resolution techniques, including TIES [25] and DOGE [45], mitigate parameter conflicts either through explicit alignment or projective gradient descent. Task Arithmetic [33] enables weight-space operations to combine task-specific behaviors in language models, while Diversity-Aware Merging, such as DARE [23], leverages model diversity to improve sparse-to-dense integration. In contrast to the data-free methods mentioned above, data-informed methods [40, 50, 51] optimize merging coefficients using additional data. Model merging is impactful for LLMs, enabling efficient knowledge integration without full retraining, facilitating distributed finetuning [44], multi-task learning [35], and cost-effective model adaptation.", + "bbox": [ + 91, + 261, + 483, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 729, + 181, + 746 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary: Frank-Wolfe algorithm", + "text_level": 1, + "bbox": [ + 89, + 755, + 410, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The Frank-Wolfe (FW) algorithm [9], also known as the conditional gradient method, is an iterative optimization algorithm for constrained optimization problems of the form:", + "bbox": [ + 89, + 777, + 482, + 824 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {x \\in \\mathcal {C}} f (x) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 251, + 835, + 482, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $f$ is a continuously differentiable function, and $\\mathcal{C}$ is a compact convex set. The algorithm follows an elegant", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "geometric intuition: at each iteration $t$ , FW first identifies which vertex of $\\mathcal{C}$ yields the steepest descent direction and then moves towards this vertex to decrease the value of the objective function. More specifically, FW algorithm:", + "bbox": [ + 511, + 90, + 905, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "1. Constructs a linear subproblem of the original optimization (a.k.a. linear minimization oracle) using first-order Taylor expansion at the point $x_{t}$ :", + "bbox": [ + 514, + 152, + 906, + 196 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {L M O} \\left(\\mathcal {C}, x _ {t}\\right) := \\underset {s \\in \\mathcal {C}} {\\arg \\min } \\langle s, \\nabla f \\left(x _ {t}\\right) \\rangle \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 205, + 906, + 231 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "2. Finds the vertex $s_t$ of the feasible set $\\mathcal{C}$ by picking $s_t \\in \\mathrm{LMO}(\\mathcal{C}, x_t)$ .", + "3. Takes a careful step from the current point $x_{t}$ towards this direction $s_t - x_t$ , maintaining feasibility through the convex combination: $x_{t + 1} = (1 - \\gamma_t)x_t + \\gamma_t s_t$ ." + ], + "bbox": [ + 511, + 237, + 903, + 313 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The step size $\\gamma_t \\in [0,1]$ can be chosen by line search", + "bbox": [ + 531, + 313, + 883, + 329 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\gamma_ {t} = \\underset {\\gamma \\in [ 0, 1 ]} {\\arg \\min } f \\left((1 - \\gamma) x _ {t} + \\gamma s _ {t}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 335, + 903, + 367 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "which ensures a sufficient decrease in $f(\\cdot)$ at each FW step.", + "bbox": [ + 511, + 376, + 903, + 391 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To determine when to stop, the $FW$ gap is used to measure the suboptimality in terms of the proximity to the best solution of LMO:", + "bbox": [ + 511, + 391, + 903, + 434 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ng _ {t} := \\max _ {s \\in \\mathcal {C}} \\langle - \\nabla f (x _ {t}), s _ {t} - x _ {t} \\rangle , \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 444, + 903, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "which is non-negative by definition.", + "bbox": [ + 511, + 474, + 753, + 489 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Frank-Wolfe Model Merging", + "text_level": 1, + "bbox": [ + 511, + 497, + 772, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We consider the problem of fine-tuning a pre-trained foundation model on new tasks. Given a pre-trained model $\\theta_0$ and previously fine-tuned models $\\{\\theta_1^*,\\theta_2^*,\\dots ,\\theta_n^*\\}$ , we aim to fine-tune the $(n + 1)$ -th model $\\theta_{n + 1}$ on new tasks as a convex combination of the previous models with optimal merging coefficients.", + "bbox": [ + 511, + 518, + 905, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To this end, we propose a Frank-Wolfe based model merging framework, which is described as follows.", + "bbox": [ + 511, + 609, + 905, + 640 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\lambda} \\ell \\left(\\sum_ {i = 1} ^ {n} \\lambda_ {i} \\theta_ {i} ^ {*}\\right) \\quad \\text {s . t .} \\quad \\sum_ {i = 1} ^ {n} \\lambda_ {i} = 1, \\quad \\lambda_ {i} \\geq 0, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 648, + 903, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\lambda = \\{\\lambda_1, \\dots, \\lambda_n\\}$ are the merging coefficients, and $\\ell$ is a loss function formulated for a specific goal, such as aligning the merged and individual models [45] or satisfying a task objective [50].", + "bbox": [ + 511, + 695, + 905, + 755 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Potentially, a scaling issue of this formulation appears when the number of fine-tuned models $n$ is large, since we need to keep all the fine-tuned models in memory. To address this, we propose a reformulation of the problem eq. (5) as follows:", + "bbox": [ + 511, + 755, + 905, + 829 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta \\in \\mathcal {M}} \\ell (\\theta), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 673, + 840, + 903, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{M} \\coloneqq \\mathrm{conv}(\\{\\theta_i^*\\}_{i=1}^n)$ is the convex hull of the set of previously fine-tuned models.", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3392", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/18f1e2940049aa7754c9381767ffdb58bccee810266c06bb4df81659d2a94216.jpg", + "image_caption": [ + "Figure 2. Illustration of model merging methods. $\\Theta_{A}$ is an irrelevant model, while $\\Theta_{B}$ and $\\Theta_{C}$ are relevant models. Darker regions indicate higher objective function loss. Task Arithmetic treats all task vectors equally, failing to move optimally. Adamerging assigns different coefficients, moving towards more desirable direction but suffer from slow convergence due to interference from $\\Theta_{a}A$ . FW-Merging iteratively selects the most relevant model to merge and adapts step sizes, efficiently reaching the optimum after $T$ iterations." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 277, + 257 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/78d1279bd2ea75d76c428caa5197a219cd2617832cc9b90ba2ce66874201f725.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 300, + 88, + 486, + 256 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2c48118c9254e59a646f332c633a407a89b5952a889e64411c5ac0f0044437cf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 508, + 88, + 694, + 257 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5fa98ac24c3c56b40d8c7c6d370663f6d3fcc490c0150901bb21ab21df69c251.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 88, + 905, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proposition 1. The optimization problems in equations (5) and (6) are equivalent.", + "text_level": 1, + "bbox": [ + 89, + 334, + 480, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proof. By definition of convex hull, any point $\\theta \\in \\mathcal{M}$ can be written as a convex combination of the vertices $\\{\\theta_i^*\\}_{i=1}^n$ , i.e., $\\theta = \\sum_{i=1}^{n} \\lambda_i \\theta_i^*$ where $\\lambda \\in \\Delta^n := \\{\\lambda \\in \\mathbb{R}^n \\mid \\sum_{i=1}^{n} \\lambda_i = 1, \\lambda_i \\geq 0\\}$ . Therefore:", + "bbox": [ + 89, + 373, + 482, + 435 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta \\in \\mathcal {M}} \\ell (\\theta) = \\min _ {\\lambda \\in \\Delta^ {n}} \\ell \\Bigl (\\sum_ {i = 1} ^ {n} \\lambda_ {i} \\theta_ {i} ^ {*} \\Bigr).\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 444, + 393, + 483 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This shows that any solution of one problem can be mapped to a solution of the other problem with the same objective value.", + "bbox": [ + 89, + 492, + 483, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the FW algorithm requires the initial solution to be an interior point of the constraint set, we add the initial solution $\\theta_0$ to form a new constraint set $\\mathcal{M} \\coloneqq \\mathrm{conv}(\\mathcal{M} \\cup \\{\\theta_0\\})$ , which we still denote as $\\mathcal{M}$ for simple notations. A nice property of this reformulation is that the LMO can be simplified to", + "bbox": [ + 89, + 547, + 483, + 638 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {L M O} \\left(\\left\\{\\theta_ {i} ^ {*} \\right\\} _ {i = 1} ^ {n}, \\theta_ {t}\\right) = \\underset {s \\in \\left\\{\\theta_ {1} ^ {*}, \\dots , \\theta_ {n} ^ {*} \\right\\}} {\\arg \\min } \\langle \\nabla \\ell \\left(\\theta_ {t}\\right), s \\rangle \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 648, + 482, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This is because for linear programming problems over convex sets, the optimal solution is always attained at the vertices of the constraint set. Algorithm 1 details the steps, and Figure 2 provides an overview.", + "bbox": [ + 89, + 685, + 483, + 746 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Design choices of the algorithm", + "text_level": 1, + "bbox": [ + 89, + 753, + 366, + 771 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The above algorithm illustrates the key ingredients of Frank-Wolfe merging: LMO, stopping criterion $g_{t}$ , line search routine, and the merging function. We discuss in this section the design choices of these components.", + "bbox": [ + 89, + 776, + 483, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Merging function The main deviation from the classical FW algorithm is the reinterpretation of the FW update $x_{t + 1} = (1 - \\gamma_t)x_t + \\gamma_t s_t$ as a local merging between $\\theta_t$", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: Frank-Wolfe Merging:" + ], + "code_body": "Input: Initial solution $\\theta_0$ ; Fine-tuned checkpoints $\\{\\theta_i^*\\}_{i=1}^n$ ; FW budget $T$ . \nOutput: Merged model $\\theta_{merged}^*$ . \n1: if $\\theta_0 \\notin \\mathcal{M}$ then $\\mathcal{M} := \\mathcal{M} \\cup \\{\\theta_0\\}$ \n2: for $t = 0 \\dots T$ do \n3: Let $s_t := \\mathrm{LMO}(\\theta_t)$ and $d_t := s_t - \\theta_t$ \n4: if $g_t := \\langle -\\nabla \\ell(\\theta_t), d_t \\rangle \\leq \\epsilon$ then \n5: return $\\theta_{merged}^* \\gets \\theta_t$ \n6: end if \n7: Line-search: $\\gamma_t \\in \\arg \\min_{\\gamma \\in [0,1]} \\ell(\\theta_t + \\gamma d_t)$ \n8: Update: $\\theta_{t+1} := \\mathrm{MergeFn}(\\theta_t, s_t, \\gamma_t)$ \n9: end for \n10: return $\\theta_{merged}^* \\gets \\theta_T$", + "bbox": [ + 526, + 354, + 874, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and $s_t$ . We denote by MergeFn the customizable merging function as long as the merged model stays in the convex hull $\\mathcal{M}$ . The most straightforward merging function is the convex combination:", + "bbox": [ + 511, + 599, + 906, + 659 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {M e r g e F n} \\left(\\theta_ {t}, s _ {t}, \\gamma_ {t}\\right) := \\left(1 - \\gamma_ {t}\\right) \\theta_ {t} + \\gamma_ {t} s _ {t}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 674, + 906, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which corresponds to the Task-Arithmetic [33] method. The step size $\\gamma_{t}$ makes sure the merged model stays in $\\mathcal{M}$ .", + "bbox": [ + 511, + 703, + 905, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It is natural to ask whether other existing model merging methods, such as TIES-Merging [25] and DARE [23] could also be used as MergeFn. The problem with these sophisticated merging methods is that the merged model might leave the constraint set, and thus violate the assumption of maintaining feasibility required by the classical FW theory. We verified in practice that these less rigorous merging functions might achieve better performance in certain cases but they generally cause more stability issues. Therefore, we do not consider these merging functions from the current comparison.", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3393", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hard FW v.s. Soft FW In the case of deep learning, the optimization problem is non-convex, additional efforts are needed to better characterize the loss landscape and prevent the LMO from being dominated by one or a few fine-tuned models, which occurs because the linear approximation of $\\ell(\\theta)$ is an inaccurate sketch of the original objective function. Instead of relying on the argmin of linear subproblems, we fetch the top- $k$ vertices of LMO, $\\{\\tilde{s}_j\\}_{j=1}^k$ . A more subtle top- $k$ operation can be performed in a task-wise fashion if the original objective function involves multi-tasks.", + "bbox": [ + 89, + 90, + 480, + 241 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the top- $k$ vertices, we now go back to eq. (5) to obtain the optimal merging coefficients $\\{\\lambda_j^*\\}_{j=1}^k$ . Note that this inner optimization is a reduced version of original eq. (5) because hosting $k$ models in memory would not be a problem. We also remove the line search step as this gives a new merging function of the form", + "bbox": [ + 89, + 242, + 480, + 335 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {M e r g e F n} \\left(\\theta_ {t}, \\left\\{\\tilde {s} _ {j} \\right\\} _ {j = 1} ^ {k}, \\left\\{\\lambda_ {j} ^ {*} \\right\\} _ {j = 1} ^ {k}\\right) := \\theta_ {t} + \\sum_ {j = 1} ^ {k} \\lambda_ {j} ^ {*} \\left(\\tilde {s} _ {j} - \\theta_ {t}\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 347, + 480, + 405 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We call this oracle soft LMO in comparison to the argmin version which we call hard LMO.", + "bbox": [ + 89, + 417, + 480, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Task-wise LMO v.s. layer-wise LMO The naive implementation of FW-Merging would vectorize the whole model weights $\\theta$ and then solve LMO. We call this taskwise LMO. Since different layers contribute differently to model performance [52], a layer-wise LMO may yield better model merging. To incorporate this, the constraint set is redefined as a Cartesian product of convex hulls for each layer: $\\mathcal{M} := \\mathcal{M}_1 \\times \\dots \\times \\mathcal{M}_L$ , where $L$ is the number of layers and $\\mathcal{M}_l := \\mathrm{conv}\\left(\\left\\{\\theta_i^{*,l}\\right\\}_{i=1}^n\\right)$ . The LMO is then conducted layer-wise:", + "bbox": [ + 89, + 468, + 483, + 626 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {L M O} \\left(\\left\\{\\theta_ {i} ^ {* , l} \\right\\} _ {i = 1} ^ {n}, D, \\theta_ {t} ^ {l}\\right) = \\underset {s ^ {l} \\in \\left\\{\\theta_ {1} ^ {* , l}, \\dots , \\theta_ {n} ^ {* , l} \\right\\}} {\\arg \\min } \\left\\langle \\nabla \\ell \\left(\\theta^ {t}\\right) ^ {l}, s ^ {l} \\right\\rangle . \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 635, + 480, + 683 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This version can be viewed as a block-coordinate Frank-Wolfe algorithm [19], which is applied when the problem has a natural decomposition into blocks.", + "bbox": [ + 89, + 695, + 480, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 756, + 223, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 89, + 781, + 266, + 797 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "**Benchmarks.** Our primary objective is to evaluate the effectiveness of our method in scenarios where the number of models greatly exceeds the number of evaluation tasks, and each model's capabilities are unknown in advance.", + "bbox": [ + 89, + 803, + 480, + 864 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Vision Tasks: Following the setting of TALL [43], we use 20 ViT-B/32 models, each fine-tuned on a different vision task. The number of models to be merged is intentionally set to be significantly larger than the number of evaluation tasks, allowing us to assess the scalability of model merging methods. The evaluation benchmarks consist of four tasks: SUN397 [47], Stanford Cars [18], GTSRB [38], and DTD [7].", + "- Language Discriminative Tasks: We prepare 8 RoBERTa checkpoints [27] fine-tuned on eight tasks from the GLUE benchmark, following the practice in [28]. The merged model is then evaluated on four tasks from the GLUE benchmark [42]: MNLI, QNLI, QQP, and RTE.", + "- Language Generative Tasks: We collect 16 LLaMA2-7B models [41] fine-tuned with LoRA [12] on various tasks from Hugging Face. These models have unknown and uncontrolled capabilities, making them equivalently black-box models. Our goal is to evaluate the robustness of model merging methods in this challenging setting. The evaluation benchmarks include CNN/DM summarization [34], PubMedQA [17], and HumanEval [4]." + ], + "bbox": [ + 513, + 90, + 903, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. For vision tasks, we report the classification accuracy. Following [28], we report the average normalized score for language tasks to account for differences in task-specific score ranges to account for variations in task-specific score ranges. The normalized score is computed as $\\mathrm{Score}_{\\mathrm{normalized}} = \\frac{1}{T}\\sum_{t=1}^{T}\\frac{\\mathrm{Score}(f(\\pmb{\\theta}^*))}{\\mathrm{Score}(f_t(\\pmb{\\theta}_t))}$ .", + "bbox": [ + 511, + 430, + 903, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We compare FW-Merging with both data-informed and data-free model merging methods. For data-informed model merging, we compare FW-Merging against Adamerging [50], Surgery [51], and Concrete Merging [40]. To ensure a fair comparison, these methods are trained only on the same tasks as FW-Merging. For data-free model merging, we compare FW-Merging with Fisher Merging, Weight Averaging, RegMean Merging, Task Arithmetic [33], Ties-Merging [48], and DARE Merging [23] across both language and vision tasks. Additionally, we fine-tune one model on the discriminative language benchmark and another on the generative language benchmark to serve as additional baselines.", + "bbox": [ + 511, + 546, + 906, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementations. We implement two FW-Merging variants: $\\mathrm{FW}_{\\text {hard }}$ , which uses hard FW with layer-wise LMO, and $\\mathrm{FW}_{\\text {soft }}$ , which employs soft FW with task-wise LMO (Section 3.3). For $\\mathrm{FW}_{\\text {soft }}$ , layer-wise coefficients are optimized via gradient descent on the training dataset to solve eq. 5, differing from Adamerging [50] by minimizing cross-entropy loss on training data rather than entropy on test samples. On language benchmarks, the training dataset consists of 100 samples per task, and $\\mathrm{FW}_{\\text {hard }}$ runs", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "2For inner optimization, we use projected gradient descent with a projection of $\\{\\lambda_j\\}_{j = 1}^k$ onto the simplex after each gradient update.", + "bbox": [ + 89, + 875, + 480, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3394", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/6e9a39bcb8e1ae9ca1b914d1d63ad6b890def48ed03940924eb68b72e57070fb.jpg", + "table_caption": [ + "Table 1. Performance on 4 Discriminative Tasks when merging 8 RoBERTa and 3 Generative Tasks when merging 16 LLaMA2-7B." + ], + "table_footnote": [], + "table_body": "
Method4 Disc. Tasks (8 Models)3 Gen. Tasks (16 Models)Avg. Normalized Score
Pretrained49.677.163.4
Traditional MTL73.181.277.2
Task Arithmetic (w/ DARE)77.316.847.1
Ties-Merging (w/ DARE)75.646.661.1
Task Arithmetic80.875.978.4
Ties-Merging64.378.571.4
FWhard (Ours)85.481.183.1
", + "bbox": [ + 127, + 104, + 859, + 246 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/30b5700ec75b6f45d6eca874b5717dda76abb713225c2e2e2241cdaad0701793.jpg", + "table_caption": [ + "Table 2. Performance on 4 CV Tasks when merging 20 ViT-B/32." + ], + "table_footnote": [], + "table_body": "
MethodSUN397CarsGTSRBDTDAvg.
Pretrained62.359.732.643.849.6
DARE (TIES)5.92.316.711.89.2
Task Arithmetic20.412.229.822.321.2
Ties-Merging51.036.257.740.646.4
Weight Averaging64.259.643.146.553.4
Fisher Merging64.663.843.046.954.6
RegMean65.562.259.753.960.3
LW Concrete AM62.560.388.054.766.3
Adamerging66.470.195.164.073.9
Surgery69.771.896.673.477.9
FWhard (Ours)66.569.995.164.574.0
FWsoft (Ours)72.974.896.876.080.1
", + "bbox": [ + 94, + 266, + 478, + 459 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "for 10 iterations, initialized with Task Arithmetic's merged model. For vision tasks, it runs for 3 iterations. $FW_{\\text{soft}}$ runs for 15 iterations on vision benchmarks, initialized with the pre-trained model. Training datasets consist of 100 samples from MNLI, QNLI, QQP, and RTE [42] for discriminative tasks; CNN/DM [34], CodeAlpaca-20k [3], and PubMedQA [17] for generative tasks; and SUN397 [47], Stanford Cars [18], GTSRB [38], and DTD [7] for vision tasks.", + "bbox": [ + 89, + 468, + 482, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with Model Merging Methods", + "text_level": 1, + "bbox": [ + 89, + 599, + 455, + 617 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate FW-Merging against both data-informed and data-free model merging approaches across language and vision benchmarks. Table 1 reports the results for language tasks, including both discriminative and generative settings, while Table 2 presents results on vision benchmarks.", + "bbox": [ + 89, + 622, + 483, + 699 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Language Tasks. $\\mathsf{FW}_{\\text{hard}}$ achieves the highest average normalized score across language benchmarks, consistently surpassing prior model merging baselines. Specifically, $\\mathsf{FW}_{\\text{hard}}$ improves upon Task Arithmetic by 4.6 points, Ties-Merging by 11.7, and DARE (Ties) by 9.8 on discriminative tasks. Table 4 shows that $\\mathsf{FW}_{\\text{hard}}$ also outperforms data-informed Adamerging by 5.9 points. For generative tasks, $\\mathsf{FW}_{\\text{hard}}$ outperforms Task Arithmetic by 5.2 points, Ties-Merging by 2.6, and DARE (Ties) by 34.5. Interestingly, while Task Arithmetic outperforms Ties-Merging on discriminative tasks by a margin of 16.5 points, it lags behind by 2.6 points on the more challenging generative tasks.", + "bbox": [ + 89, + 719, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This discrepancy likely arises from increased interference among task vectors as more checkpoints are merged. Unlike Ties-Merging, which explicitly resolves merging conflicts, Task Arithmetic lacks a reconciliation mechanism, making it more susceptible to such interference. In contrast, $\\mathsf{FW}_{\\text {hard }}$ consistently outperforms both Ties-Merging and Task Arithmetic by selectively merging only the most relevant model parameters in each iteration. This targeted approach effectively mitigates interference, leading to more stable and robust performance across both discriminative and generative tasks.", + "bbox": [ + 511, + 253, + 906, + 420 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Vision Tasks. $\\mathsf{FW}_{\\mathrm{soft}}$ achieves state-of-the-art performance across multiple vision benchmarks, surpassing data-informed methods like Adamerging and Surgery. As shown in Table 2, $\\mathsf{FW}_{\\mathrm{hard}}$ surpasses Adamerging, Concrete Merging, and all data-free merging methods in overall performance. Additionally, $\\mathsf{FW}_{\\mathrm{soft}}$ attains the highest accuracy (80.1%), outperforming Adamerging by 6.2% and Surgery by 2.2%. Unlike Surgery, which requires additional task-specific parameters and multiple forward passes per inference, our approach efficiently adapts to diverse visual tasks without increasing storage or inference complexity.", + "bbox": [ + 511, + 441, + 903, + 608 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In general, data-free merging methods show significantly lower performance compared to data-informed approaches while merging a large number of models, when the models' capabilities do not precisely align with the evaluation tasks. This limitation arises because data-free methods treat all models equally, merging them without considering their unique capabilities, which amplifies interference between models. In contrast, data-informed merging methods achieve superior performance by optimizing merging coefficients on datasets as they allow for explicit control over desirable capabilities. FW-Merging, in particular, enhances scalability via hard model selection based on the linear approximation minimization.", + "bbox": [ + 511, + 609, + 906, + 805 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparison with Traditional MTL", + "text_level": 1, + "bbox": [ + 511, + 816, + 818, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare FW-Merging with models fine-tuned using traditional MTL on discriminative and generative tasks. In each case, one single model is fine-tuned across all tasks, with performance and computational cost reported", + "bbox": [ + 511, + 839, + 905, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3395", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a9ea11b279109757e9c4edc4c98bd47662828c950b6ce506ea6ec9d449e036c0.jpg", + "table_caption": [ + "Table 3. Merging Methods' Performance vs. Number of Models when Adding Relevant vs. Irrelevant Models." + ], + "table_footnote": [], + "table_body": "
#Models4 CV Tasks\nWhen "Irrelevant" Models Added20 CV Tasks\nWhen "Relevant" Models Added
DARETaskTiesAMFWsoftDARETaskTiesAMFWsoft
473.670.365.775.274.157.359.260.259.659.2
664.164.164.474.673.954.058.461.664.063.2
848.557.163.673.474.048.257.263.964.764.5
1040.652.962.972.873.943.555.063.965.466.2
1231.747.961.971.474.139.152.363.765.167.5
1423.640.157.769.974.036.148.962.064.168.0
1617.933.352.568.674.132.345.359.762.768.3
1812.225.949.264.674.027.840.357.360.768.4
209.221.246.461.074.224.236.355.658.968.3
", + "bbox": [ + 187, + 104, + 803, + 294 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/df2de3ebb3c8975898230d9a993b3f46bc15f236e32ff6dcedd62fd22858c3b4.jpg", + "image_caption": [ + "Figure 3. Linear Approximation of the Objective Function of Model Checkpoints Across Different Tasks in a Frank-Wolfe Iteration. The x-axis represents the checkpoints, and each graph shows the linear approximation result for each task." + ], + "image_footnote": [], + "bbox": [ + 96, + 296, + 903, + 450 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/46371d62766429ea62f50e0ac32b2d296a83b675910a089dd4cdf913e08aabbd.jpg", + "table_caption": [ + "Table 4. Costs and Perf. of methods on NLP discriminative tasks." + ], + "table_footnote": [], + "table_body": "
MethodData Samples/TaskTime CostPerf.
Traditional MTL2.9K4.2h73.1
Data-free Merging0080.8
Data-informed Merging1.6K2min79.5
FW Merging1002min85.4
", + "bbox": [ + 94, + 510, + 478, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in Table 1 and Table 4. Traditional MTL achieves an average score of 77.2, lower than that of FW-Merging (83.1). On discriminative tasks, MTL scores 73.1, trailing FW-Merging (85.4). For generative tasks, MTL scores 81.2, while FW-Merging closely matches it at 80.8, suggesting that FW-Merging's performance matches that of traditional MTL.", + "bbox": [ + 89, + 595, + 483, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Table 4, FW-Merging demonstrates a substantial advantage in efficiency. Traditional MTL requires fine-tuning on 2.9K samples per task and takes 4.2 hours of training time, which is computationally intensive. In contrast, FW-Merging only requires 100 training samples per task and completes the merging process in just 2 minutes. This huge reduction in computational cost underscores the effectiveness of FW-Merging compared to traditional MTL. Moreover, FW-Merging has a key advantage over traditional MTL: while MTL requires a large volume of high-quality data for optimal performance, FW-Merging needs only a small set of data because: $I$ ) it optimizes merging coefficients based on models' character", + "bbox": [ + 88, + 704, + 485, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "istics, which simplifies the optimization space, and 2) it uses model weights as inputs, which are much more information-dense representations than data, enabling more efficient objective learning.", + "bbox": [ + 511, + 500, + 906, + 559 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "FW-Merging is a post-training technique that does not require access to original training data, making it ideal for privacy-sensitive or data-scarce scenarios. Overall, the results suggest that FW-Merging is a scalable, efficient alternative to traditional MTL, providing comparable performance at a reduced computational cost.", + "bbox": [ + 511, + 560, + 908, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Scaling to More Models and Tasks", + "text_level": 1, + "bbox": [ + 511, + 662, + 815, + 680 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We investigate the performance scaling of different merging methods with the number of models, as shown in Figure 1a, Figure 1b, and Table 3. In large-scale model merging, models from open-source platforms vary in quality. To simulate this, we use 20 ViT-B/32 models fine-tuned on tasks that are either included in the evaluation benchmark or not. A model is irrelevant if its fine-tuning dataset does not match the training split of the evaluation task, and relevant if it matches. To ensure fair comparison, the total number of training iterations run by $\\mathrm{FW}_{\\mathrm{soft}}$ is the same as that of Adamreging.", + "bbox": [ + 509, + 702, + 906, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Table 3, adding irrelevant models sharply reduces the performance of data-free methods: DARE by", + "bbox": [ + 511, + 869, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3396", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5fb679ee7d74e01196f6a15c7c4ded41938208856404d448ed7b56a00187e0be.jpg", + "table_caption": [ + "Table 5. Ablation on design variants of FW-Merging." + ], + "table_footnote": [], + "table_body": "
Coefficient λMethodLMOScore
Vision Tasks
OptimizedFWsoftLayer-wise79.7
OptimizedFWsoftTask-wise80.1
UnoptimizedFWsoftLayer-wise69.8
UnoptimizedFWsoftTask-wise70.3
-FWhardLayer-wise74.0
-FWhardTask-wise73.7
NLP Discriminative Tasks
-FWhardLayer-wise85.4
-FWhardTask-wise78.2
", + "bbox": [ + 112, + 103, + 450, + 292 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$64.4\\%$ , Task Arithmetic by $49.1\\%$ , and Ties by $19.1\\%$ , likely due to task interference and equal treatment of all models. Data-informed methods degrade less, with Adamerging dropping by $14.2\\%$ . In contrast, $\\mathsf{FW}_{\\mathrm{soft}}$ remains highly stable, fluctuating only from $73.9\\%$ to $74.1\\%$ as more models are added. In Figure 3, we examine the linear approximation of different checkpoints for a specific task and find that the model fine-tuned on the task consistently yields the most negative linear approximation. This indicates that in the Frank-Wolfe update, the most relevant checkpoint is chosen as the direction for merging, allowing FW-Merging to iteratively improve the merged model in the optimized direction within the constraint set. The inner product between gradients and model parameters serves as a reliable indicator of model relevance, with minimized computational cost, further demonstrating FW-Merging's scalability even in the presence of irrelevant models.", + "bbox": [ + 89, + 301, + 483, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Adding relevant models should ideally improve performance, but data-free methods still degrade as shown in Table 3: DARE by $33.1\\%$ , Task Arithmetic by $22.9\\%$ , and Ties by $4.6\\%$ , with Ties performing best by mitigating parameter conflicts. Data-informed methods like Adamerging fluctuate between $58.9\\%$ and $64.7\\%$ as merging complexity increases, whereas $\\mathsf{FW}_{\\mathrm{soft}}$ steadily improves from $59.2\\%$ to $68.3\\%$ by iteratively selecting the most relevant models, facilitating smoother convergence. These results underscore FW-Merging's effectiveness as a scalable solution for large-scale model merging.", + "bbox": [ + 89, + 561, + 483, + 728 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 739, + 254, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Design variants. Table 5 compares the design variants of FW-Merging (Section 3.3). Task-wise LMO aligns better with $\\mathrm{FW}_{\\mathrm{soft}}$ , improving performance slightly by 0.5 points over layer-wise LMO, while layer-wise LMO is more effective for $\\mathrm{FW}_{\\mathrm{hard}}$ , especially on language tasks, yielding a 7.2-point gain. This is likely because $\\mathrm{FW}_{\\mathrm{soft}}$ optimizes layer-wise coefficients during merging, reducing the impact of layer-wise selection.", + "bbox": [ + 89, + 763, + 482, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\mathrm{FW}_{\\mathrm{soft}}$ excels when merging a large number of models,", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ae19002c886a8aea4bfcd1afe05fbcf7f6e685988213e5ef1f9be3da3dff54f2.jpg", + "image_caption": [ + "(a) Initial soluiton" + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 700, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/35343851102dddc913b651adead6e9a37c74a06a25753f212018c0ed25712f9b.jpg", + "image_caption": [ + "(b) Merging functions" + ], + "image_footnote": [], + "bbox": [ + 718, + 90, + 900, + 183 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 4. Ablation on FW-Merging. (a) reports accuracies on the vision benchmark, while (b) on vision and language benchmarks.", + "bbox": [ + 511, + 202, + 903, + 229 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "outperforming $\\mathrm{FW}_{\\mathrm{hard}}$ by up to 6.7 points. Its ability to select multiple optimal directions per iteration allows it to navigate the parameter space efficiently.", + "bbox": [ + 511, + 234, + 903, + 280 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Optimizing merging coefficients $\\lambda$ further improves performance by up to 9.9 points, underscoring the importance of weighting model parameters based on their relevance.", + "bbox": [ + 511, + 281, + 903, + 325 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Initial solution. We examine the effect of initialization on FW-Merging. An ideal initial solution should either (1) be closer to the global optimum or (2) expand the constraint set with a more meaningful search space. As shown in Figure 4a, initializing FW-Merging with the Adamerging result improves performance compared to starting from the pre-trained model, likely because Adamerging is closer to the optimal point. In contrast, task arithmetic leads to worse performance than the pre-trained model, potentially due to its poor performance on vision tasks $(21.2\\%)$ , suggesting it starts further from the optimum. Consequently, more FW iterations are required to achieve convergence.", + "bbox": [ + 511, + 344, + 906, + 526 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Flexibility of merging functions. Although only a restricted set of merging functions ensure that FW-Merging remains within the convex hull, we demonstrate the flexibility of FW-Merging by showing its ability to enhance alternative merging functions. As shown in Figure 4b, applying FW-Merging with both Task Arithmetic and Ties-Merging improves performance on NLP and vision tasks, even though Ties-Merging does not necessarily stay within the convex hull. This suggests that FW-Merging remains effective across different merging functions.", + "bbox": [ + 511, + 544, + 906, + 695 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 708, + 633, + 723 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we extend model merging to a more challenging setting where the merging pool consists of a large number of black-box fine-tuned checkpoints. While existing methods require prior knowledge of model details to achieve optimized performance, our proposed Frank-Wolfe Merging (FW-Merging) scales effectively with a large number of black-box models, iteratively refining the merged model towards the optimal point defined by an objective function. Experiments demonstrate that FW-Merging achieves superior performance and scalability, paving the way for next-generation model merging.", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3397", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The support of the United Kingdom EPSRC (grant number UKRI256, EP/V028251/1, EP/N031768/1, EP/S030069/1, and EP/X036006/1), Intel, and AMD is gratefully acknowledged.", + "bbox": [ + 89, + 114, + 483, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 193, + 187, + 209 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Samuel K Ainsworth, Jonathan Hayase, Siddhartha Srinivasan, Khaled Saab, and Stefano Fusi. Git re-basin: Merging models modulo permutation symmetries. arXiv preprint arXiv:2209.04836, 2022. 3", + "[2] Rich Caruana. Multitask learning. Machine learning, 28: 41-75, 1997. 2", + "[3] Sahil Chaudhary. Code alpaca: An instruction-following llama model for code generation. https://github.com/sahil280114/codealpaca, 2023.6", + "[4] Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott Gray, Nick Ryder, Mikhail Pavlov, Alethea Power, Lukasz Kaiser, Mohammad Bavarian, Clemens Winter, Philippe Tillet, Felipe Petroski Such, Dave Cummings, Matthias Plappert, Fotios Chantzis, Elizabeth Barnes, Ariel Herbert-Voss, William Hebgen Guss, Alex Nichol, Alex Paino, Nikolas Tezak, Jie Tang, Igor Babuschkin, Suchir Balaji, Shantanu Jain, William Saunders, Christopher Hesse, Andrew N. Carr, Jan Leike, Josh Achiam, Vedant Misra, Evan Morikawa, Alec Radford, Matthew Knight, Miles Brundage, Mira Murati, Katie Mayer, Peter Welinder, Bob McGrew, Dario Amodei, Sam McCandlish, Ilya Sutskever, and Wojciech Zaremba. Evaluating large language models trained on code. 2021. 5", + "[5] Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks. In International conference on machine learning, pages 794-803. PMLR, 2018. 3", + "[6] Zhao Chen, Jiquan Ngiam, Yanping Huang, Thang Luong, Henrik Kretzschmar, Yuning Chai, and Dragomir Anguelov. Just pick a sign: Optimizing deep multitask models with gradient sign dropout. Advances in Neural Information Processing Systems, 33:2039-2050, 2020. 3", + "[7] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3606-3613, 2014. 5, 6", + "[8] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: Efficient finetuning of quantized llms. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[9] Marguerite Frank, Philip Wolfe, et al. An algorithm for quadratic programming. Naval research logistics quarterly, 3(1-2):95-110, 1956. 3" + ], + "bbox": [ + 99, + 217, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Ian J Goodfellow, Oriol Vinyals, and Andrew M Saxe. Qualitatively characterizing neural network optimization problems. arXiv preprint arXiv:1412.6544, 2014. 3", + "[11] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3", + "[12] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 5", + "[13] Hugging Face. Open LLM Leaderboard, 2024. Accessed: March 4, 2025. 1", + "[14] Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Suchin Gururangan, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. Editing models with task arithmetic. arXiv preprint arXiv:2212.04089, 2022. 1", + "[15] Pavel Izmailov, Dmitrii Podoprikhin, Timur Garipov, Dmitry Vetrov, and Andrew Gordon Wilson. Averaging weights leads to wider optima and better generalization. arXiv preprint arXiv:1803.05407, 2018. 1, 3", + "[16] Junguang Jiang, Baixu Chen, Junwei Pan, Ximei Wang, Dapeng Liu, Jie Jiang, and Mingsheng Long. Forkmerge: Mitigating negative transfer in auxiliary-task learning. Advances in neural information processing systems, 36:30367-30389, 2023. 2", + "[17] Qiao Jin, Bhuwan Dhingra, Zhengping Liu, William Cohen, and Xinghua Lu. Pubmedqa: A dataset for biomedical research question answering. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2567-2577, 2019. 5, 6", + "[18] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013. 5, 6", + "[19] Simon Lacoste-Julien, Martin Jaggi, Mark Schmidt, and Patrick Pletscher. Block-coordinate frank-wolfe optimization for structural svms. In International Conference on Machine Learning, pages 53-61. PMLR, 2013. 5", + "[20] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 3", + "[21] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 3", + "[22] Yixiao Li, Yifan Yu, Chen Liang, Pengcheng He, Nikos Karampatziakis, Weizhu Chen, and Tuo Zhao. Loftq: Lorafine-tuning-aware quantization for large language models. arXiv preprint arXiv:2310.08659, 2023. 3", + "[23] Fuzhao Liu, Xiaotian Hu, Chengyu Chen, Zhijian Wang, Shuxin Xiao, Zhen Wang, Zhongxiang Wang, Chuanxin Xie, Eric Xing, and Song Han. Dare: Diversity-aware model merging for sparse-to-dense mixture-of-experts. arXiv preprint arXiv:2402.10887, 2024. 3, 4, 5" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3398", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Liyang Liu, Yi Li, Zhanghui Kuang, J Xue, Yimin Chen, Wenming Yang, Qingmin Liao, and Wayne Zhang. Towards impartial multi-task learning. iclr, 2021. 3", + "[25] Prateek Yadav Liu et al. Resolving interference when merging models. arXiv preprint arXiv:2306.01708, 2023. 3, 4", + "[26] Shikun Liu, Edward Johns, and Andrew J Davison. End-to-end multi-task learning with attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1871-1880, 2019. 2", + "[27] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettle-moyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 5", + "[28] Zhenyi Lu, Chenghao Fan, Wei Wei, Xiaoye Qu, Dangyang Chen, and Yu Cheng. Twin-merging: Dynamic integration of modular expertise in model merging. Advances in Neural Information Processing Systems, 37:78905–78935, 2025. 5", + "[29] Jiaqi Ma, Zhe Zhao, Xinyang Yi, Jilin Chen, Lichan Hong, and Ed H Chi. Modeling task relationships in multi-task learning with multi-gate mixture-of-experts. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1930-1939, 2018. 2", + "[30] Jiaqi Ma, Zhe Zhao, Jilin Chen, Ang Li, Lichan Hong, and Ed H Chi. Snr: Sub-network routing for flexible parameter sharing in multi-task learning. In Proceedings of the AAAI conference on artificial intelligence, pages 216-223, 2019. 2", + "[31] Sadhika Malladi, Tianyu Gao, Eshaan Nichani, Alex Damian, Jason D Lee, Danqi Chen, and Sanjeev Arora. Finetuning language models with just forward passes. Advances in Neural Information Processing Systems, 36:53038-53075, 2023. 3", + "[32] Priyanka Mary Mammen. Federated learning: Opportunities and challenges. arXiv preprint arXiv:2101.05428, 2021. 1", + "[33] Eric Mitchell, Kenton Lee, Alexander Khazatsky, Christopher D Manning, and Chelsea Finn. Editing models with task arithmetic. arXiv preprint arXiv:2212.04089, 2022. 3, 4, 5", + "[34] Ramesh Nallapati, Bowen Zhou, Caglar Gulcehre, Bing Xiang, et al. Abstractive text summarization using sequence-to-sequence rnns and beyond. arXiv preprint arXiv:1602.06023, 2016.5, 6", + "[35] Shraman Prakash, Rohit Girdhar, Devamanyu Hazarika, Joao Carreira, and Andrew Zisserman. Unified model for image, video, audio and language tasks. arXiv preprint arXiv:2307.16184, 2023. 3", + "[36] Samyam Rajbhandari, Jeff Rasley, Olatunjri Ruwase, and Yuxiong He. Zero: Memory optimizations toward training trillion parameter models. In SC20: International Conference for High Performance Computing, Networking, Storage and Analysis, pages 1-16. IEEE, 2020. 1", + "[37] Sidak Pal Singh and Martin Jaggi. Merging models with fisher-weighted averaging. arXiv preprint arXiv:2111.09832, 2021. 3", + "[38] Johannes Stallkamp, Marc Schlipsing, Jan Salmen, and Christian Igel. The german traffic sign recognition benchmark: a multi-class classification competition. In *The 2011" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "international joint conference on neural networks, pages 1453-1460. IEEE, 2011. 5, 6", + "[39] Ximeng Sun, Rameswar Panda, Rogerio Feris, and Kate Saenko. Adashare: Learning what to share for efficient deep multi-task learning. Advances in Neural Information Processing Systems, 33:8728-8740, 2020. 2", + "[40] Anke Tang, Li Shen, Yong Luo, Liang Ding, Han Hu, Bo Du, and Dacheng Tao. Concrete subspace learning based interference elimination for multi-task model fusion. arXiv preprint arXiv:2312.06173, 2023. 3, 5", + "[41] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 5", + "[42] Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R Bowman. Glue: A multi-task benchmark and analysis platform for natural language understanding. arXiv preprint arXiv:1804.07461, 2018. 5, 6", + "[43] Ke Wang, Nikolaos Dimitriadis, Guillermo Ortiz-Jimenez, François Fleuret, and Pascal Frossard. Localizing task information for improved model merging and compression. arXiv preprint arXiv:2405.07813, 2024. 5", + "[44] Zirui Wang, Daniel Duckworth, Shashank Nag, Kshitij Patil, Jonathon Shlens, Ekin D Cubuk, Barret Zoph, Joseph Campbell, Mitchell Wortsman, and Google Brain. lo-fi: distributed fine-tuning without communication. arXiv preprint arXiv:2210.11948, 2022. 3", + "[45] Yongxian Wei, Anke Tang, Li Shen, Feng Xiong, Chun Yuan, and Xiaochun Cao. Modeling multi-task model merging as adaptive projective gradient descent. arXiv preprint arXiv:2501.01230, 2025. 3", + "[46] Mitchell Wortsman, Gabriel Ilharco, Samir Yitzhak Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. arXiv preprint arXiv:2203.05482, 2022. 3", + "[47] Jianxiong Xiao, Krista A Ehinger, James Hays, Antonio Torralba, and Aude Oliva. Sun database: Exploring a large collection of scene categories. International Journal of Computer Vision, 119:3-22, 2016. 5, 6", + "[48] Prateek Yadav, Derek Tam, Leshem Choshen, Colin A Raffel, and Mohit Bansal. Ties-merging: Resolving interference when merging models. Advances in Neural Information Processing Systems, 36:7093-7115, 2023. 1, 2, 5", + "[49] Enneng Yang, Li Shen, Guibing Guo, Xingwei Wang, Xiaochun Cao, Jie Zhang, and Dacheng Tao. Model merging in llms, mllms, and beyond: Methods, theories, applications and opportunities, 2024. URL https://arxiv.org/abs/2408.07666, 2408.3", + "[50] Enneng Yang, Zhenyi Wang, Li Shen, Shiwei Liu, Guibing Guo, Xingwei Wang, and Dacheng Tao. Adamerging: Adaptive model merging for multi-task learning. arXiv preprint arXiv:2310.02575, 2023. 1, 2, 3, 5", + "[51] Enneng Yang, Li Shen, Zhenyi Wang, Guibing Guo, Xiaojun Chen, Xingwei Wang, and Dacheng Tao. Representa" + ], + "bbox": [ + 516, + 92, + 903, + 902 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3399", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tion surgery for multi-task model merging. arXiv preprint arXiv:2402.02705, 2024. 1, 3, 5", + "[52] Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How transferable are features in deep neural networks? Advances in neural information processing systems, 27, 2014. 5", + "[53] Jun Yu, Yutong Dai, Xiaokang Liu, Jin Huang, Yishan Shen, Ke Zhang, Rong Zhou, Eashan Adhikarla, Wenxuan Ye, Yixin Liu, et al. Unleashing the power of multitask learning: A comprehensive survey spanning traditional, deep, and pretrained foundation model eras. arXiv preprint arXiv:2404.18961, 2024. 1", + "[54] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. Advances in neural information processing systems, 33:5824-5836, 2020. 3" + ], + "bbox": [ + 91, + 90, + 483, + 316 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "3400", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/FW-Merging_ Scaling Model Merging with Frank-Wolfe Optimization/56e48ade-9cb7-4688-a0be-edb71b4d459b_model.json b/2025/FW-Merging_ Scaling Model Merging with Frank-Wolfe Optimization/56e48ade-9cb7-4688-a0be-edb71b4d459b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9bbf8cfc957db74f631f25c7795c7372cec807c4 --- /dev/null +++ b/2025/FW-Merging_ Scaling Model Merging with Frank-Wolfe Optimization/56e48ade-9cb7-4688-a0be-edb71b4d459b_model.json @@ -0,0 +1,2411 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.807, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.145, + 0.131, + 0.853, + 0.153 + ], + "angle": 0, + "content": "FW-Merging: Scaling Model Merging with Frank-Wolfe Optimization" + }, + { + "type": "text", + "bbox": [ + 0.121, + 0.18, + 0.871, + 0.218 + ], + "angle": 0, + "content": "Hao (Mark) Chen1 Shell Xu Hu2 Wayne Luk1 Timothy Hospedales2 Hongxiang Fan1 \n1Imperial College London, UK 2Samsung AI Center, Cambridge, UK" + }, + { + "type": "text", + "bbox": [ + 0.119, + 0.22, + 0.875, + 0.252 + ], + "angle": 0, + "content": "hcl620@ic.ac.uk shell.hu@samsung.com w.luk@imperial.ac.uk t.hospedales@samsung.com hongxiang.fan@imperial.ac.uk" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.265, + 0.327, + 0.281 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.296, + 0.486, + 0.795 + ], + "angle": 0, + "content": "Model merging has emerged as a promising approach for multi-task learning (MTL), offering a data-efficient alternative to conventional fine-tuning. However, with the rapid development of the open-source AI ecosystem and the increasing availability of fine-tuned foundation models, existing model merging methods face two key limitations: (i) They are primarily designed for in-house fine-tuned models, making them less adaptable to diverse model sources with partially unknown model and task information, (ii) They struggle to scale effectively when merging numerous model checkpoints. To address these challenges, we formulate model merging as a constrained optimization problem and introduce a novel approach: Frank-Wolfe Merging (FW-Merging). Inspired by Frank-Wolfe optimization, our approach iteratively selects the most relevant model in the pool to minimize a linear approximation of the objective function and then executes a local merging similar to the Frank-Wolfe update. The objective function is designed to capture the desired behavior of the target-merged model, while the fine-tuned candidate models define the constraint set. More importantly, FW-Merging serves as an orthogonal technique for existing merging methods, seamlessly integrating with them to further enhance accuracy performance. Our experiments show that FW-Merging scales across diverse model sources, remaining stable with 16 irrelevant models and improving by \\(15.3\\%\\) with 16 relevant models on 20 CV tasks, while maintaining constant memory overhead—unlike the linear overhead of data-informed merging methods. Compared with the state-of-the-art approaches, FW-Merging surpasses the data-free merging method by \\(32.8\\%\\) and outperforms the data-informed Adamerging by \\(8.39\\%\\) when merging 20 ViT models. Our code is open-sourced at here." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.823, + 0.222, + 0.839 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.848, + 0.483, + 0.88 + ], + "angle": 0, + "content": "Multi-task learning (MTL)-based fine-tuning adapts a single pre-trained Large Language Model (LLM) for multiple" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.267, + 0.905, + 0.433 + ], + "angle": 0, + "content": "downstream applications, reducing the deployment overhead of separately fine-tuning multiple models [53]. However, it still demands a large amount of high-quality data, which might only exist in the private domain [32], and significant compute resources [36]. To mitigate these issues, model merging has emerged as a promising technique for fusing fine-tuned models within the parameter space [14, 48, 50]. Existing model merging methods can be broadly classified into two categories: data-free methods [14, 15, 48], and data-informed methods [50, 51], which optimize merge coefficients based on additional data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.434, + 0.905, + 0.78 + ], + "angle": 0, + "content": "While these approaches have proven effective, several key limitations hinder their scalability and broader adoption. First, these methods adjust merging coefficients based on the known capabilities of the models on specific tasks to optimize performance, which is less robust when dealing with diverse model sources with unknown information1. This is primarily caused by the inability to distinguish high-quality models from poorly fine-tuned ones in an unknown model setting. Second, when scaling up these approaches to a large number of unknown models, these methods struggle and can result in significant performance degradation. As shown in Figure 1a, our profiling study demonstrates a performance reduction ranging from \\(18.9\\%\\) to \\(64.4\\%\\). These limitations are further amplified by the fast-growing open-source AI ecosystem, where platforms such as Hugging Face have driven a surge in the release of powerful LLMs with many lacking complete information. Given that merging open-source models has repeatedly shown the potential to produce top-ranking LLMs on major model leaderboards [13], developing scalable and robust merging techniques is essential to harness the growing number of open-source models, further enhancing performance and widening the adoption of model merging." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.905, + 0.841 + ], + "angle": 0, + "content": "To effectively scale model merging and leverage the vast collection of open-sourced models with unknown capabilities, the new model merging method must exhibit two fundamental scaling properties: 1) as more irrelevant models" + }, + { + "type": "page_footnote", + "bbox": [ + 0.512, + 0.852, + 0.905, + 0.901 + ], + "angle": 0, + "content": "1This paper refers unknown information to: 1) when open-source models are partially assessed on limited benchmarks, leaving their performance on other tasks unknown and costly to evaluate, and 2) when models contain misleading information, polluting the merging process." + }, + { + "type": "footer", + "bbox": [ + 0.109, + 0.887, + 0.421, + 0.901 + ], + "angle": 0, + "content": "*Corresponding Authors: Shell Xu Hu and Hongxiang Fan" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3390" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.092, + 0.359, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.248, + 0.348, + 0.26 + ], + "angle": 0, + "content": "(a) 4 CV tasks with irrelevant model addition." + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.092, + 0.631, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.378, + 0.248, + 0.62, + 0.26 + ], + "angle": 0, + "content": "(b) 20 CV tasks with relevant model addition." + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.093, + 0.9, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.661, + 0.248, + 0.878, + 0.26 + ], + "angle": 0, + "content": "(c) NLP tasks with noisy model addition." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.273, + 0.908, + 0.316 + ], + "angle": 0, + "content": "Figure 1. Performance scaling of FW-Merging across CV tasks. (a) demonstrate robustness to irrelevant models, while (b) show improved performance with relevant models. (c) analyzes performance degradation when incorporating a noisy model initialized from a different pre-trained checkpoint. Detailed results and experimental setup are discussed in Section 4.4." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.324, + 0.485, + 0.671 + ], + "angle": 0, + "content": "are added to the merging pool, the performance should remain unaffected, and 2) as more relevant models are added to the merging pool, the performance should steadily increase, converging towards the optimal performance. To this end, we revisit model merging and formulate it as a constrained optimization problem, where the objective function dictates the desirable behavior of the final merged model, and fine-tuned checkpoints form the constraint set. Inspired by Frank-Wolfe optimization, we introduce Frank-Wolfe Merging (FW-Merging), an iterative algorithm designed to enhance merging efficiency while maintaining robustness at scale. FW-Merging comprises three principal stages in each iteration: (1) Relevance Evaluation: Instead of merging models arbitrarily, we obtain the linear approximation of the objective function using gradients of the current model, revealing the most beneficial direction for improvement. (2) Model Selection: The most relevant checkpoints are selected from the constraint set by minimizing the linear approximation, ensuring that each step incorporates task-specific knowledge with minimal interference. (3) Knowledge Integration: The selected checkpoint is integrated using an orthogonal merging method, striking a balance between adaptation and stability in the merged model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We demonstrate the effectiveness of FW-Merging with a diverse pool of fine-tuned checkpoints across various language and vision tasks, compared to both data-free and data-informed model merging methods as well as traditional MTL-based fine-tuning. As shown in Figure 1, FW-Merging satisfy our two fundamental scaling properties: accuracy performance does not drop when 16 irrelevant models are added (compared to a \\(49\\%\\) drop in task-arithmetic) and steadily improves by \\(15.3\\%\\) when 16 relevant models are included. Additionally, FW-Merging requires only constant memory overhead, as it selects and merges a fixed number of models at a time. In contrast, methods that optimize merging coefficients [50] or resolve parameter interference [48] must store all models in memory, leading to linear overhead. Moreover, FW-Merging" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.324, + 0.907, + 0.535 + ], + "angle": 0, + "content": "exhibits greater robustness to noisy models lacking critical information, such as their initialization point. As shown in Figure 1c, FW-Merging experiences minimal performance degradation when a misinitialized model is introduced, whereas Ties suffers a performance drop of up to \\(3.2\\%\\). FW-Merging outperforms state-of-the-art data-free merging method by \\(32.8\\%\\) and the data-informed method Adamerging by \\(8.39\\%\\) when merging 20 ViT models. On the language benchmarks, FW-Merging achieves \\(6.3\\%\\) improvement over the best model merging method across discriminative and generative tasks, while even surpassing the performance of traditional MTL using only \\(3.4\\%\\) of its required data. Our results position FW-Merging as an effective solution to scale model merging to the next level." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.537, + 0.859, + 0.551 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.553, + 0.905, + 0.599 + ], + "angle": 0, + "content": "- Identify scalability and robustness issues in existing model merging techniques through experiments, highlighting the urgent need for large-scale model merging." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.599, + 0.905, + 0.643 + ], + "angle": 0, + "content": "- Formulate model merging as a constrained optimization problem with an objective function that explicitly captures the desired behavior of the final merged model." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.643, + 0.905, + 0.704 + ], + "angle": 0, + "content": "- Introduce Frank-Wolfe Merging, a novel iterative method that autonomously guides the merged model toward an optimized direction, even with large sets of black-box open-source checkpoints." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.704, + 0.905, + 0.734 + ], + "angle": 0, + "content": "- Evaluate our proposed approach on extensive benchmarks, demonstrating its effectiveness and scalability." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.553, + 0.905, + 0.734 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.753, + 0.655, + 0.769 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Efficient Multi-Task Learning. In traditional Multi-Task Learning (MTL), a single model is trained on a dataset containing multiple tasks to enable the model to acquire diverse capabilities [2]. However, a significant challenge in traditional MTL is the issue of negative transfer [16]. To mitigate this, architecture-based approaches have been developed, such as parameter sparsification [26, 39] and shared structure modularization [29, 30]. On the optimization side," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3391" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.244 + ], + "angle": 0, + "content": "methods to resolve gradient conflicts [6, 54] and domination of gradient or learning rate [5, 24] have been proposed. With the rise of Large Language Models (LLMs), MTL faces additional challenges, particularly the high computational costs. To address these challenges, strategies like parameter-efficient fine-tuning [11, 20, 21] and memory-efficient fine-tuning [8, 22, 31] have been introduced to minimize both memory and computational resource usage. More recently, model merging has emerged as a promising approach to make MTL more compute- and data-efficient." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.262, + 0.485, + 0.718 + ], + "angle": 0, + "content": "Model Merging. While pre-merging methods prepare favorable conditions for merging, during-merging techniques combine multiple neural networks into a single model while retaining or enhancing their capabilities [49]. In this work, we focus on during-merging methods. Early insights into neural network landscapes [10] revealed that linear interpolation between models exposes useful loss surface properties, laying the foundation for weight averaging—a core merging technique. Simple averaging widens optima and improves generalization [15], evolving into advanced methods like model soups [46] and heterogeneous model merging. Recent advances introduce more structured approaches, such as Fisher-Weighted Averaging [37], which incorporates Fisher information to weight parameters more effectively, and Permutation Alignment methods like Git Re-Basin [1], which address weight permutation symmetries. Interference Resolution techniques, including TIES [25] and DOGE [45], mitigate parameter conflicts either through explicit alignment or projective gradient descent. Task Arithmetic [33] enables weight-space operations to combine task-specific behaviors in language models, while Diversity-Aware Merging, such as DARE [23], leverages model diversity to improve sparse-to-dense integration. In contrast to the data-free methods mentioned above, data-informed methods [40, 50, 51] optimize merging coefficients using additional data. Model merging is impactful for LLMs, enabling efficient knowledge integration without full retraining, facilitating distributed finetuning [44], multi-task learning [35], and cost-effective model adaptation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.731, + 0.183, + 0.747 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.756, + 0.411, + 0.773 + ], + "angle": 0, + "content": "3.1. Preliminary: Frank-Wolfe algorithm" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.779, + 0.483, + 0.825 + ], + "angle": 0, + "content": "The Frank-Wolfe (FW) algorithm [9], also known as the conditional gradient method, is an iterative optimization algorithm for constrained optimization problems of the form:" + }, + { + "type": "equation", + "bbox": [ + 0.253, + 0.837, + 0.483, + 0.86 + ], + "angle": 0, + "content": "\\[\n\\min _ {x \\in \\mathcal {C}} f (x) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\( f \\) is a continuously differentiable function, and \\( \\mathcal{C} \\) is a compact convex set. The algorithm follows an elegant" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.152 + ], + "angle": 0, + "content": "geometric intuition: at each iteration \\( t \\), FW first identifies which vertex of \\( \\mathcal{C} \\) yields the steepest descent direction and then moves towards this vertex to decrease the value of the objective function. More specifically, FW algorithm:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.153, + 0.907, + 0.198 + ], + "angle": 0, + "content": "1. Constructs a linear subproblem of the original optimization (a.k.a. linear minimization oracle) using first-order Taylor expansion at the point \\( x_{t} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.206, + 0.907, + 0.232 + ], + "angle": 0, + "content": "\\[\n\\operatorname {L M O} \\left(\\mathcal {C}, x _ {t}\\right) := \\underset {s \\in \\mathcal {C}} {\\arg \\min } \\langle s, \\nabla f \\left(x _ {t}\\right) \\rangle \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.238, + 0.905, + 0.268 + ], + "angle": 0, + "content": "2. Finds the vertex \\( s_t \\) of the feasible set \\( \\mathcal{C} \\) by picking \\( s_t \\in \\mathrm{LMO}(\\mathcal{C}, x_t) \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.269, + 0.905, + 0.314 + ], + "angle": 0, + "content": "3. Takes a careful step from the current point \\( x_{t} \\) towards this direction \\( s_t - x_t \\), maintaining feasibility through the convex combination: \\( x_{t + 1} = (1 - \\gamma_t)x_t + \\gamma_t s_t \\)." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.238, + 0.905, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.314, + 0.885, + 0.33 + ], + "angle": 0, + "content": "The step size \\(\\gamma_t \\in [0,1]\\) can be chosen by line search" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.337, + 0.905, + 0.368 + ], + "angle": 0, + "content": "\\[\n\\gamma_ {t} = \\underset {\\gamma \\in [ 0, 1 ]} {\\arg \\min } f \\left((1 - \\gamma) x _ {t} + \\gamma s _ {t}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.377, + 0.905, + 0.392 + ], + "angle": 0, + "content": "which ensures a sufficient decrease in \\( f(\\cdot) \\) at each FW step." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.392, + 0.905, + 0.435 + ], + "angle": 0, + "content": "To determine when to stop, the \\( FW \\) gap is used to measure the suboptimality in terms of the proximity to the best solution of LMO:" + }, + { + "type": "equation", + "bbox": [ + 0.603, + 0.445, + 0.905, + 0.467 + ], + "angle": 0, + "content": "\\[\ng _ {t} := \\max _ {s \\in \\mathcal {C}} \\langle - \\nabla f (x _ {t}), s _ {t} - x _ {t} \\rangle , \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.475, + 0.754, + 0.491 + ], + "angle": 0, + "content": "which is non-negative by definition." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.498, + 0.774, + 0.515 + ], + "angle": 0, + "content": "3.2. Frank-Wolfe Model Merging" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.906, + 0.61 + ], + "angle": 0, + "content": "We consider the problem of fine-tuning a pre-trained foundation model on new tasks. Given a pre-trained model \\(\\theta_0\\) and previously fine-tuned models \\(\\{\\theta_1^*,\\theta_2^*,\\dots ,\\theta_n^*\\}\\), we aim to fine-tune the \\((n + 1)\\)-th model \\(\\theta_{n + 1}\\) on new tasks as a convex combination of the previous models with optimal merging coefficients." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.611, + 0.906, + 0.641 + ], + "angle": 0, + "content": "To this end, we propose a Frank-Wolfe based model merging framework, which is described as follows." + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.649, + 0.905, + 0.687 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\lambda} \\ell \\left(\\sum_ {i = 1} ^ {n} \\lambda_ {i} \\theta_ {i} ^ {*}\\right) \\quad \\text {s . t .} \\quad \\sum_ {i = 1} ^ {n} \\lambda_ {i} = 1, \\quad \\lambda_ {i} \\geq 0, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.696, + 0.906, + 0.756 + ], + "angle": 0, + "content": "where \\(\\lambda = \\{\\lambda_1, \\dots, \\lambda_n\\}\\) are the merging coefficients, and \\(\\ell\\) is a loss function formulated for a specific goal, such as aligning the merged and individual models [45] or satisfying a task objective [50]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.756, + 0.906, + 0.83 + ], + "angle": 0, + "content": "Potentially, a scaling issue of this formulation appears when the number of fine-tuned models \\( n \\) is large, since we need to keep all the fine-tuned models in memory. To address this, we propose a reformulation of the problem eq. (5) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.674, + 0.841, + 0.905, + 0.863 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta \\in \\mathcal {M}} \\ell (\\theta), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathcal{M} \\coloneqq \\mathrm{conv}(\\{\\theta_i^*\\}_{i=1}^n)\\) is the convex hull of the set of previously fine-tuned models." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3392" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.278, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.089, + 0.487, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.089, + 0.696, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.089, + 0.906, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.27, + 0.907, + 0.327 + ], + "angle": 0, + "content": "Figure 2. Illustration of model merging methods. \\(\\Theta_{A}\\) is an irrelevant model, while \\(\\Theta_{B}\\) and \\(\\Theta_{C}\\) are relevant models. Darker regions indicate higher objective function loss. Task Arithmetic treats all task vectors equally, failing to move optimally. Adamerging assigns different coefficients, moving towards more desirable direction but suffer from slow convergence due to interference from \\(\\Theta_{a}A\\). FW-Merging iteratively selects the most relevant model to merge and adapts step sizes, efficiently reaching the optimum after \\(T\\) iterations." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.335, + 0.482, + 0.366 + ], + "angle": 0, + "content": "Proposition 1. The optimization problems in equations (5) and (6) are equivalent." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.374, + 0.483, + 0.436 + ], + "angle": 0, + "content": "Proof. By definition of convex hull, any point \\(\\theta \\in \\mathcal{M}\\) can be written as a convex combination of the vertices \\(\\{\\theta_i^*\\}_{i=1}^n\\), i.e., \\(\\theta = \\sum_{i=1}^{n} \\lambda_i \\theta_i^*\\) where \\(\\lambda \\in \\Delta^n := \\{\\lambda \\in \\mathbb{R}^n \\mid \\sum_{i=1}^{n} \\lambda_i = 1, \\lambda_i \\geq 0\\}\\). Therefore:" + }, + { + "type": "equation", + "bbox": [ + 0.178, + 0.445, + 0.395, + 0.484 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta \\in \\mathcal {M}} \\ell (\\theta) = \\min _ {\\lambda \\in \\Delta^ {n}} \\ell \\Bigl (\\sum_ {i = 1} ^ {n} \\lambda_ {i} \\theta_ {i} ^ {*} \\Bigr).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.493, + 0.484, + 0.538 + ], + "angle": 0, + "content": "This shows that any solution of one problem can be mapped to a solution of the other problem with the same objective value." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.548, + 0.484, + 0.64 + ], + "angle": 0, + "content": "Since the FW algorithm requires the initial solution to be an interior point of the constraint set, we add the initial solution \\(\\theta_0\\) to form a new constraint set \\(\\mathcal{M} \\coloneqq \\mathrm{conv}(\\mathcal{M} \\cup \\{\\theta_0\\})\\), which we still denote as \\(\\mathcal{M}\\) for simple notations. A nice property of this reformulation is that the LMO can be simplified to" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.649, + 0.483, + 0.677 + ], + "angle": 0, + "content": "\\[\n\\operatorname {L M O} \\left(\\left\\{\\theta_ {i} ^ {*} \\right\\} _ {i = 1} ^ {n}, \\theta_ {t}\\right) = \\underset {s \\in \\left\\{\\theta_ {1} ^ {*}, \\dots , \\theta_ {n} ^ {*} \\right\\}} {\\arg \\min } \\langle \\nabla \\ell \\left(\\theta_ {t}\\right), s \\rangle \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.686, + 0.484, + 0.747 + ], + "angle": 0, + "content": "This is because for linear programming problems over convex sets, the optimal solution is always attained at the vertices of the constraint set. Algorithm 1 details the steps, and Figure 2 provides an overview." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.755, + 0.367, + 0.772 + ], + "angle": 0, + "content": "3.3. Design choices of the algorithm" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.777, + 0.484, + 0.839 + ], + "angle": 0, + "content": "The above algorithm illustrates the key ingredients of Frank-Wolfe merging: LMO, stopping criterion \\( g_{t} \\), line search routine, and the merging function. We discuss in this section the design choices of these components." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Merging function The main deviation from the classical FW algorithm is the reinterpretation of the FW update \\( x_{t + 1} = (1 - \\gamma_t)x_t + \\gamma_t s_t \\) as a local merging between \\( \\theta_t \\)" + }, + { + "type": "code_caption", + "bbox": [ + 0.523, + 0.338, + 0.771, + 0.354 + ], + "angle": 0, + "content": "Algorithm 1: Frank-Wolfe Merging:" + }, + { + "type": "algorithm", + "bbox": [ + 0.527, + 0.355, + 0.875, + 0.567 + ], + "angle": 0, + "content": "Input: Initial solution \\(\\theta_0\\); Fine-tuned checkpoints \\(\\{\\theta_i^*\\}_{i=1}^n\\); FW budget \\(T\\). \nOutput: Merged model \\(\\theta_{merged}^*\\). \n1: if \\(\\theta_0 \\notin \\mathcal{M}\\) then \\(\\mathcal{M} := \\mathcal{M} \\cup \\{\\theta_0\\}\\) \n2: for \\(t = 0 \\dots T\\) do \n3: Let \\(s_t := \\mathrm{LMO}(\\theta_t)\\) and \\(d_t := s_t - \\theta_t\\) \n4: if \\(g_t := \\langle -\\nabla \\ell(\\theta_t), d_t \\rangle \\leq \\epsilon\\) then \n5: return \\(\\theta_{merged}^* \\gets \\theta_t\\) \n6: end if \n7: Line-search: \\(\\gamma_t \\in \\arg \\min_{\\gamma \\in [0,1]} \\ell(\\theta_t + \\gamma d_t)\\) \n8: Update: \\(\\theta_{t+1} := \\mathrm{MergeFn}(\\theta_t, s_t, \\gamma_t)\\) \n9: end for \n10: return \\(\\theta_{merged}^* \\gets \\theta_T\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.601, + 0.907, + 0.66 + ], + "angle": 0, + "content": "and \\( s_t \\). We denote by MergeFn the customizable merging function as long as the merged model stays in the convex hull \\( \\mathcal{M} \\). The most straightforward merging function is the convex combination:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.675, + 0.907, + 0.693 + ], + "angle": 0, + "content": "\\[\n\\operatorname {M e r g e F n} \\left(\\theta_ {t}, s _ {t}, \\gamma_ {t}\\right) := \\left(1 - \\gamma_ {t}\\right) \\theta_ {t} + \\gamma_ {t} s _ {t}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.906, + 0.734 + ], + "angle": 0, + "content": "which corresponds to the Task-Arithmetic [33] method. The step size \\(\\gamma_{t}\\) makes sure the merged model stays in \\(\\mathcal{M}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "It is natural to ask whether other existing model merging methods, such as TIES-Merging [25] and DARE [23] could also be used as MergeFn. The problem with these sophisticated merging methods is that the merged model might leave the constraint set, and thus violate the assumption of maintaining feasibility required by the classical FW theory. We verified in practice that these less rigorous merging functions might achieve better performance in certain cases but they generally cause more stability issues. Therefore, we do not consider these merging functions from the current comparison." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3393" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.482, + 0.242 + ], + "angle": 0, + "content": "Hard FW v.s. Soft FW In the case of deep learning, the optimization problem is non-convex, additional efforts are needed to better characterize the loss landscape and prevent the LMO from being dominated by one or a few fine-tuned models, which occurs because the linear approximation of \\(\\ell(\\theta)\\) is an inaccurate sketch of the original objective function. Instead of relying on the argmin of linear subproblems, we fetch the top-\\(k\\) vertices of LMO, \\(\\{\\tilde{s}_j\\}_{j=1}^k\\). A more subtle top-\\(k\\) operation can be performed in a task-wise fashion if the original objective function involves multi-tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.243, + 0.482, + 0.336 + ], + "angle": 0, + "content": "Given the top-\\(k\\) vertices, we now go back to eq. (5) to obtain the optimal merging coefficients \\(\\{\\lambda_j^*\\}_{j=1}^k\\). Note that this inner optimization is a reduced version of original eq. (5) because hosting \\(k\\) models in memory would not be a problem. We also remove the line search step as this gives a new merging function of the form" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.348, + 0.482, + 0.406 + ], + "angle": 0, + "content": "\\[\n\\operatorname {M e r g e F n} \\left(\\theta_ {t}, \\left\\{\\tilde {s} _ {j} \\right\\} _ {j = 1} ^ {k}, \\left\\{\\lambda_ {j} ^ {*} \\right\\} _ {j = 1} ^ {k}\\right) := \\theta_ {t} + \\sum_ {j = 1} ^ {k} \\lambda_ {j} ^ {*} \\left(\\tilde {s} _ {j} - \\theta_ {t}\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.418, + 0.482, + 0.448 + ], + "angle": 0, + "content": "We call this oracle soft LMO in comparison to the argmin version which we call hard LMO." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.469, + 0.484, + 0.627 + ], + "angle": 0, + "content": "Task-wise LMO v.s. layer-wise LMO The naive implementation of FW-Merging would vectorize the whole model weights \\(\\theta\\) and then solve LMO. We call this taskwise LMO. Since different layers contribute differently to model performance [52], a layer-wise LMO may yield better model merging. To incorporate this, the constraint set is redefined as a Cartesian product of convex hulls for each layer: \\(\\mathcal{M} := \\mathcal{M}_1 \\times \\dots \\times \\mathcal{M}_L\\), where \\(L\\) is the number of layers and \\(\\mathcal{M}_l := \\mathrm{conv}\\left(\\left\\{\\theta_i^{*,l}\\right\\}_{i=1}^n\\right)\\). The LMO is then conducted layer-wise:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.636, + 0.482, + 0.684 + ], + "angle": 0, + "content": "\\[\n\\operatorname {L M O} \\left(\\left\\{\\theta_ {i} ^ {* , l} \\right\\} _ {i = 1} ^ {n}, D, \\theta_ {t} ^ {l}\\right) = \\underset {s ^ {l} \\in \\left\\{\\theta_ {1} ^ {* , l}, \\dots , \\theta_ {n} ^ {* , l} \\right\\}} {\\arg \\min } \\left\\langle \\nabla \\ell \\left(\\theta^ {t}\\right) ^ {l}, s ^ {l} \\right\\rangle . \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.696, + 0.482, + 0.743 + ], + "angle": 0, + "content": "This version can be viewed as a block-coordinate Frank-Wolfe algorithm [19], which is applied when the problem has a natural decomposition into blocks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.757, + 0.224, + 0.774 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.782, + 0.267, + 0.798 + ], + "angle": 0, + "content": "4.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.804, + 0.482, + 0.866 + ], + "angle": 0, + "content": "**Benchmarks.** Our primary objective is to evaluate the effectiveness of our method in scenarios where the number of models greatly exceeds the number of evaluation tasks, and each model's capabilities are unknown in advance." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.091, + 0.905, + 0.212 + ], + "angle": 0, + "content": "- Vision Tasks: Following the setting of TALL [43], we use 20 ViT-B/32 models, each fine-tuned on a different vision task. The number of models to be merged is intentionally set to be significantly larger than the number of evaluation tasks, allowing us to assess the scalability of model merging methods. The evaluation benchmarks consist of four tasks: SUN397 [47], Stanford Cars [18], GTSRB [38], and DTD [7]." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.213, + 0.905, + 0.287 + ], + "angle": 0, + "content": "- Language Discriminative Tasks: We prepare 8 RoBERTa checkpoints [27] fine-tuned on eight tasks from the GLUE benchmark, following the practice in [28]. The merged model is then evaluated on four tasks from the GLUE benchmark [42]: MNLI, QNLI, QQP, and RTE." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.288, + 0.904, + 0.409 + ], + "angle": 0, + "content": "- Language Generative Tasks: We collect 16 LLaMA2-7B models [41] fine-tuned with LoRA [12] on various tasks from Hugging Face. These models have unknown and uncontrolled capabilities, making them equivalently black-box models. Our goal is to evaluate the robustness of model merging methods in this challenging setting. The evaluation benchmarks include CNN/DM summarization [34], PubMedQA [17], and HumanEval [4]." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.091, + 0.905, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.431, + 0.905, + 0.528 + ], + "angle": 0, + "content": "Metrics. For vision tasks, we report the classification accuracy. Following [28], we report the average normalized score for language tasks to account for differences in task-specific score ranges to account for variations in task-specific score ranges. The normalized score is computed as \\(\\mathrm{Score}_{\\mathrm{normalized}} = \\frac{1}{T}\\sum_{t=1}^{T}\\frac{\\mathrm{Score}(f(\\pmb{\\theta}^*))}{\\mathrm{Score}(f_t(\\pmb{\\theta}_t))}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.547, + 0.907, + 0.743 + ], + "angle": 0, + "content": "Baselines. We compare FW-Merging with both data-informed and data-free model merging methods. For data-informed model merging, we compare FW-Merging against Adamerging [50], Surgery [51], and Concrete Merging [40]. To ensure a fair comparison, these methods are trained only on the same tasks as FW-Merging. For data-free model merging, we compare FW-Merging with Fisher Merging, Weight Averaging, RegMean Merging, Task Arithmetic [33], Ties-Merging [48], and DARE Merging [23] across both language and vision tasks. Additionally, we fine-tune one model on the discriminative language benchmark and another on the generative language benchmark to serve as additional baselines." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Implementations. We implement two FW-Merging variants: \\(\\mathrm{FW}_{\\text {hard }}\\), which uses hard FW with layer-wise LMO, and \\(\\mathrm{FW}_{\\text {soft }}\\), which employs soft FW with task-wise LMO (Section 3.3). For \\(\\mathrm{FW}_{\\text {soft }}\\), layer-wise coefficients are optimized via gradient descent on the training dataset to solve eq. 5, differing from Adamerging [50] by minimizing cross-entropy loss on training data rather than entropy on test samples. On language benchmarks, the training dataset consists of 100 samples per task, and \\(\\mathrm{FW}_{\\text {hard }}\\) runs" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.875, + 0.482, + 0.903 + ], + "angle": 0, + "content": "2For inner optimization, we use projected gradient descent with a projection of \\(\\{\\lambda_j\\}_{j = 1}^k\\) onto the simplex after each gradient update." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3394" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.11, + 0.089, + 0.887, + 0.103 + ], + "angle": 0, + "content": "Table 1. Performance on 4 Discriminative Tasks when merging 8 RoBERTa and 3 Generative Tasks when merging 16 LLaMA2-7B." + }, + { + "type": "table", + "bbox": [ + 0.129, + 0.106, + 0.86, + 0.247 + ], + "angle": 0, + "content": "
Method4 Disc. Tasks (8 Models)3 Gen. Tasks (16 Models)Avg. Normalized Score
Pretrained49.677.163.4
Traditional MTL73.181.277.2
Task Arithmetic (w/ DARE)77.316.847.1
Ties-Merging (w/ DARE)75.646.661.1
Task Arithmetic80.875.978.4
Ties-Merging64.378.571.4
FWhard (Ours)85.481.183.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.094, + 0.253, + 0.48, + 0.266 + ], + "angle": 0, + "content": "Table 2. Performance on 4 CV Tasks when merging 20 ViT-B/32." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.267, + 0.48, + 0.46 + ], + "angle": 0, + "content": "
MethodSUN397CarsGTSRBDTDAvg.
Pretrained62.359.732.643.849.6
DARE (TIES)5.92.316.711.89.2
Task Arithmetic20.412.229.822.321.2
Ties-Merging51.036.257.740.646.4
Weight Averaging64.259.643.146.553.4
Fisher Merging64.663.843.046.954.6
RegMean65.562.259.753.960.3
LW Concrete AM62.560.388.054.766.3
Adamerging66.470.195.164.073.9
Surgery69.771.896.673.477.9
FWhard (Ours)66.569.995.164.574.0
FWsoft (Ours)72.974.896.876.080.1
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.469, + 0.483, + 0.592 + ], + "angle": 0, + "content": "for 10 iterations, initialized with Task Arithmetic's merged model. For vision tasks, it runs for 3 iterations. \\(FW_{\\text{soft}}\\) runs for 15 iterations on vision benchmarks, initialized with the pre-trained model. Training datasets consist of 100 samples from MNLI, QNLI, QQP, and RTE [42] for discriminative tasks; CNN/DM [34], CodeAlpaca-20k [3], and PubMedQA [17] for generative tasks; and SUN397 [47], Stanford Cars [18], GTSRB [38], and DTD [7] for vision tasks." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.601, + 0.456, + 0.618 + ], + "angle": 0, + "content": "4.2. Comparison with Model Merging Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.623, + 0.484, + 0.7 + ], + "angle": 0, + "content": "We evaluate FW-Merging against both data-informed and data-free model merging approaches across language and vision benchmarks. Table 1 reports the results for language tasks, including both discriminative and generative settings, while Table 2 presents results on vision benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Language Tasks. \\( \\mathsf{FW}_{\\text{hard}} \\) achieves the highest average normalized score across language benchmarks, consistently surpassing prior model merging baselines. Specifically, \\( \\mathsf{FW}_{\\text{hard}} \\) improves upon Task Arithmetic by 4.6 points, Ties-Merging by 11.7, and DARE (Ties) by 9.8 on discriminative tasks. Table 4 shows that \\( \\mathsf{FW}_{\\text{hard}} \\) also outperforms data-informed Adamerging by 5.9 points. For generative tasks, \\( \\mathsf{FW}_{\\text{hard}} \\) outperforms Task Arithmetic by 5.2 points, Ties-Merging by 2.6, and DARE (Ties) by 34.5. Interestingly, while Task Arithmetic outperforms Ties-Merging on discriminative tasks by a margin of 16.5 points, it lags behind by 2.6 points on the more challenging generative tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.254, + 0.907, + 0.421 + ], + "angle": 0, + "content": "This discrepancy likely arises from increased interference among task vectors as more checkpoints are merged. Unlike Ties-Merging, which explicitly resolves merging conflicts, Task Arithmetic lacks a reconciliation mechanism, making it more susceptible to such interference. In contrast, \\(\\mathsf{FW}_{\\text {hard }}\\) consistently outperforms both Ties-Merging and Task Arithmetic by selectively merging only the most relevant model parameters in each iteration. This targeted approach effectively mitigates interference, leading to more stable and robust performance across both discriminative and generative tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.443, + 0.905, + 0.609 + ], + "angle": 0, + "content": "Vision Tasks. \\(\\mathsf{FW}_{\\mathrm{soft}}\\) achieves state-of-the-art performance across multiple vision benchmarks, surpassing data-informed methods like Adamerging and Surgery. As shown in Table 2, \\(\\mathsf{FW}_{\\mathrm{hard}}\\) surpasses Adamerging, Concrete Merging, and all data-free merging methods in overall performance. Additionally, \\(\\mathsf{FW}_{\\mathrm{soft}}\\) attains the highest accuracy (80.1%), outperforming Adamerging by 6.2% and Surgery by 2.2%. Unlike Surgery, which requires additional task-specific parameters and multiple forward passes per inference, our approach efficiently adapts to diverse visual tasks without increasing storage or inference complexity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.907, + 0.806 + ], + "angle": 0, + "content": "In general, data-free merging methods show significantly lower performance compared to data-informed approaches while merging a large number of models, when the models' capabilities do not precisely align with the evaluation tasks. This limitation arises because data-free methods treat all models equally, merging them without considering their unique capabilities, which amplifies interference between models. In contrast, data-informed merging methods achieve superior performance by optimizing merging coefficients on datasets as they allow for explicit control over desirable capabilities. FW-Merging, in particular, enhances scalability via hard model selection based on the linear approximation minimization." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.818, + 0.82, + 0.835 + ], + "angle": 0, + "content": "4.3. Comparison with Traditional MTL" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.903 + ], + "angle": 0, + "content": "We compare FW-Merging with models fine-tuned using traditional MTL on discriminative and generative tasks. In each case, one single model is fine-tuned across all tasks, with performance and computational cost reported" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "3395" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.173, + 0.089, + 0.825, + 0.103 + ], + "angle": 0, + "content": "Table 3. Merging Methods' Performance vs. Number of Models when Adding Relevant vs. Irrelevant Models." + }, + { + "type": "table", + "bbox": [ + 0.188, + 0.106, + 0.805, + 0.295 + ], + "angle": 0, + "content": "
#Models4 CV Tasks\nWhen "Irrelevant" Models Added20 CV Tasks\nWhen "Relevant" Models Added
DARETaskTiesAMFWsoftDARETaskTiesAMFWsoft
473.670.365.775.274.157.359.260.259.659.2
664.164.164.474.673.954.058.461.664.063.2
848.557.163.673.474.048.257.263.964.764.5
1040.652.962.972.873.943.555.063.965.466.2
1231.747.961.971.474.139.152.363.765.167.5
1423.640.157.769.974.036.148.962.064.168.0
1617.933.352.568.674.132.345.359.762.768.3
1812.225.949.264.674.027.840.357.360.768.4
209.221.246.461.074.224.236.355.658.968.3
" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.297, + 0.904, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.463, + 0.907, + 0.491 + ], + "angle": 0, + "content": "Figure 3. Linear Approximation of the Objective Function of Model Checkpoints Across Different Tasks in a Frank-Wolfe Iteration. The x-axis represents the checkpoints, and each graph shows the linear approximation result for each task." + }, + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.497, + 0.482, + 0.51 + ], + "angle": 0, + "content": "Table 4. Costs and Perf. of methods on NLP discriminative tasks." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.511, + 0.48, + 0.587 + ], + "angle": 0, + "content": "
MethodData Samples/TaskTime CostPerf.
Traditional MTL2.9K4.2h73.1
Data-free Merging0080.8
Data-informed Merging1.6K2min79.5
FW Merging1002min85.4
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.484, + 0.701 + ], + "angle": 0, + "content": "in Table 1 and Table 4. Traditional MTL achieves an average score of 77.2, lower than that of FW-Merging (83.1). On discriminative tasks, MTL scores 73.1, trailing FW-Merging (85.4). For generative tasks, MTL scores 81.2, while FW-Merging closely matches it at 80.8, suggesting that FW-Merging's performance matches that of traditional MTL." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.705, + 0.486, + 0.902 + ], + "angle": 0, + "content": "As shown in Table 4, FW-Merging demonstrates a substantial advantage in efficiency. Traditional MTL requires fine-tuning on 2.9K samples per task and takes 4.2 hours of training time, which is computationally intensive. In contrast, FW-Merging only requires 100 training samples per task and completes the merging process in just 2 minutes. This huge reduction in computational cost underscores the effectiveness of FW-Merging compared to traditional MTL. Moreover, FW-Merging has a key advantage over traditional MTL: while MTL requires a large volume of high-quality data for optimal performance, FW-Merging needs only a small set of data because: \\(I\\)) it optimizes merging coefficients based on models' character" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.5, + 0.907, + 0.56 + ], + "angle": 0, + "content": "istics, which simplifies the optimization space, and 2) it uses model weights as inputs, which are much more information-dense representations than data, enabling more efficient objective learning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.561, + 0.909, + 0.652 + ], + "angle": 0, + "content": "FW-Merging is a post-training technique that does not require access to original training data, making it ideal for privacy-sensitive or data-scarce scenarios. Overall, the results suggest that FW-Merging is a scalable, efficient alternative to traditional MTL, providing comparable performance at a reduced computational cost." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.664, + 0.816, + 0.681 + ], + "angle": 0, + "content": "4.4. Scaling to More Models and Tasks" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.703, + 0.907, + 0.87 + ], + "angle": 0, + "content": "We investigate the performance scaling of different merging methods with the number of models, as shown in Figure 1a, Figure 1b, and Table 3. In large-scale model merging, models from open-source platforms vary in quality. To simulate this, we use 20 ViT-B/32 models fine-tuned on tasks that are either included in the evaluation benchmark or not. A model is irrelevant if its fine-tuning dataset does not match the training split of the evaluation task, and relevant if it matches. To ensure fair comparison, the total number of training iterations run by \\( \\mathrm{FW}_{\\mathrm{soft}} \\) is the same as that of Adamreging." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.909, + 0.903 + ], + "angle": 0, + "content": "As shown in Table 3, adding irrelevant models sharply reduces the performance of data-free methods: DARE by" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3396" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.123, + 0.09, + 0.449, + 0.103 + ], + "angle": 0, + "content": "Table 5. Ablation on design variants of FW-Merging." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.104, + 0.451, + 0.294 + ], + "angle": 0, + "content": "
Coefficient λMethodLMOScore
Vision Tasks
OptimizedFWsoftLayer-wise79.7
OptimizedFWsoftTask-wise80.1
UnoptimizedFWsoftLayer-wise69.8
UnoptimizedFWsoftTask-wise70.3
-FWhardLayer-wise74.0
-FWhardTask-wise73.7
NLP Discriminative Tasks
-FWhardLayer-wise85.4
-FWhardTask-wise78.2
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.303, + 0.484, + 0.56 + ], + "angle": 0, + "content": "\\(64.4\\%\\), Task Arithmetic by \\(49.1\\%\\), and Ties by \\(19.1\\%\\), likely due to task interference and equal treatment of all models. Data-informed methods degrade less, with Adamerging dropping by \\(14.2\\%\\). In contrast, \\(\\mathsf{FW}_{\\mathrm{soft}}\\) remains highly stable, fluctuating only from \\(73.9\\%\\) to \\(74.1\\%\\) as more models are added. In Figure 3, we examine the linear approximation of different checkpoints for a specific task and find that the model fine-tuned on the task consistently yields the most negative linear approximation. This indicates that in the Frank-Wolfe update, the most relevant checkpoint is chosen as the direction for merging, allowing FW-Merging to iteratively improve the merged model in the optimized direction within the constraint set. The inner product between gradients and model parameters serves as a reliable indicator of model relevance, with minimized computational cost, further demonstrating FW-Merging's scalability even in the presence of irrelevant models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.562, + 0.484, + 0.729 + ], + "angle": 0, + "content": "Adding relevant models should ideally improve performance, but data-free methods still degrade as shown in Table 3: DARE by \\(33.1\\%\\), Task Arithmetic by \\(22.9\\%\\), and Ties by \\(4.6\\%\\), with Ties performing best by mitigating parameter conflicts. Data-informed methods like Adamerging fluctuate between \\(58.9\\%\\) and \\(64.7\\%\\) as merging complexity increases, whereas \\(\\mathsf{FW}_{\\mathrm{soft}}\\) steadily improves from \\(59.2\\%\\) to \\(68.3\\%\\) by iteratively selecting the most relevant models, facilitating smoother convergence. These results underscore FW-Merging's effectiveness as a scalable solution for large-scale model merging." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.741, + 0.255, + 0.755 + ], + "angle": 0, + "content": "4.5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.764, + 0.483, + 0.884 + ], + "angle": 0, + "content": "Design variants. Table 5 compares the design variants of FW-Merging (Section 3.3). Task-wise LMO aligns better with \\(\\mathrm{FW}_{\\mathrm{soft}}\\), improving performance slightly by 0.5 points over layer-wise LMO, while layer-wise LMO is more effective for \\(\\mathrm{FW}_{\\mathrm{hard}}\\), especially on language tasks, yielding a 7.2-point gain. This is likely because \\(\\mathrm{FW}_{\\mathrm{soft}}\\) optimizes layer-wise coefficients during merging, reducing the impact of layer-wise selection." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "\\(\\mathrm{FW}_{\\mathrm{soft}}\\) excels when merging a large number of models," + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.092, + 0.701, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.189, + 0.659, + 0.201 + ], + "angle": 0, + "content": "(a) Initial soluiton" + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.092, + 0.901, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.752, + 0.189, + 0.869, + 0.201 + ], + "angle": 0, + "content": "(b) Merging functions" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.203, + 0.905, + 0.231 + ], + "angle": 0, + "content": "Figure 4. Ablation on FW-Merging. (a) reports accuracies on the vision benchmark, while (b) on vision and language benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.236, + 0.905, + 0.281 + ], + "angle": 0, + "content": "outperforming \\(\\mathrm{FW}_{\\mathrm{hard}}\\) by up to 6.7 points. Its ability to select multiple optimal directions per iteration allows it to navigate the parameter space efficiently." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.282, + 0.905, + 0.327 + ], + "angle": 0, + "content": "Optimizing merging coefficients \\(\\lambda\\) further improves performance by up to 9.9 points, underscoring the importance of weighting model parameters based on their relevance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.345, + 0.907, + 0.527 + ], + "angle": 0, + "content": "Initial solution. We examine the effect of initialization on FW-Merging. An ideal initial solution should either (1) be closer to the global optimum or (2) expand the constraint set with a more meaningful search space. As shown in Figure 4a, initializing FW-Merging with the Adamerging result improves performance compared to starting from the pre-trained model, likely because Adamerging is closer to the optimal point. In contrast, task arithmetic leads to worse performance than the pre-trained model, potentially due to its poor performance on vision tasks \\((21.2\\%)\\), suggesting it starts further from the optimum. Consequently, more FW iterations are required to achieve convergence." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.545, + 0.907, + 0.696 + ], + "angle": 0, + "content": "Flexibility of merging functions. Although only a restricted set of merging functions ensure that FW-Merging remains within the convex hull, we demonstrate the flexibility of FW-Merging by showing its ability to enhance alternative merging functions. As shown in Figure 4b, applying FW-Merging with both Task Arithmetic and Ties-Merging improves performance on NLP and vision tasks, even though Ties-Merging does not necessarily stay within the convex hull. This suggests that FW-Merging remains effective across different merging functions." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.709, + 0.634, + 0.724 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this work, we extend model merging to a more challenging setting where the merging pool consists of a large number of black-box fine-tuned checkpoints. While existing methods require prior knowledge of model details to achieve optimized performance, our proposed Frank-Wolfe Merging (FW-Merging) scales effectively with a large number of black-box models, iteratively refining the merged model towards the optimal point defined by an objective function. Experiments demonstrate that FW-Merging achieves superior performance and scalability, paving the way for next-generation model merging." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3397" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.271, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.484, + 0.178 + ], + "angle": 0, + "content": "The support of the United Kingdom EPSRC (grant number UKRI256, EP/V028251/1, EP/N031768/1, EP/S030069/1, and EP/X036006/1), Intel, and AMD is gratefully acknowledged." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.194, + 0.188, + 0.21 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.218, + 0.484, + 0.274 + ], + "angle": 0, + "content": "[1] Samuel K Ainsworth, Jonathan Hayase, Siddhartha Srinivasan, Khaled Saab, and Stefano Fusi. Git re-basin: Merging models modulo permutation symmetries. arXiv preprint arXiv:2209.04836, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.276, + 0.484, + 0.303 + ], + "angle": 0, + "content": "[2] Rich Caruana. Multitask learning. Machine learning, 28: 41-75, 1997. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.306, + 0.484, + 0.348 + ], + "angle": 0, + "content": "[3] Sahil Chaudhary. Code alpaca: An instruction-following llama model for code generation. https://github.com/sahil280114/codealpaca, 2023.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.35, + 0.485, + 0.599 + ], + "angle": 0, + "content": "[4] Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott Gray, Nick Ryder, Mikhail Pavlov, Alethea Power, Lukasz Kaiser, Mohammad Bavarian, Clemens Winter, Philippe Tillet, Felipe Petroski Such, Dave Cummings, Matthias Plappert, Fotios Chantzis, Elizabeth Barnes, Ariel Herbert-Voss, William Hebgen Guss, Alex Nichol, Alex Paino, Nikolas Tezak, Jie Tang, Igor Babuschkin, Suchir Balaji, Shantanu Jain, William Saunders, Christopher Hesse, Andrew N. Carr, Jan Leike, Josh Achiam, Vedant Misra, Evan Morikawa, Alec Radford, Matthew Knight, Miles Brundage, Mira Murati, Katie Mayer, Peter Welinder, Bob McGrew, Dario Amodei, Sam McCandlish, Ilya Sutskever, and Wojciech Zaremba. Evaluating large language models trained on code. 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.601, + 0.484, + 0.671 + ], + "angle": 0, + "content": "[5] Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks. In International conference on machine learning, pages 794-803. PMLR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.673, + 0.484, + 0.741 + ], + "angle": 0, + "content": "[6] Zhao Chen, Jiquan Ngiam, Yanping Huang, Thang Luong, Henrik Kretzschmar, Yuning Chai, and Dragomir Anguelov. Just pick a sign: Optimizing deep multitask models with gradient sign dropout. Advances in Neural Information Processing Systems, 33:2039-2050, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.744, + 0.484, + 0.8 + ], + "angle": 0, + "content": "[7] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3606-3613, 2014. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.802, + 0.484, + 0.856 + ], + "angle": 0, + "content": "[8] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: Efficient finetuning of quantized llms. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.859, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[9] Marguerite Frank, Philip Wolfe, et al. An algorithm for quadratic programming. Naval research logistics quarterly, 3(1-2):95-110, 1956. 3" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.218, + 0.485, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[10] Ian J Goodfellow, Oriol Vinyals, and Andrew M Saxe. Qualitatively characterizing neural network optimization problems. arXiv preprint arXiv:1412.6544, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.193 + ], + "angle": 0, + "content": "[11] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.195, + 0.905, + 0.25 + ], + "angle": 0, + "content": "[12] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.253, + 0.904, + 0.28 + ], + "angle": 0, + "content": "[13] Hugging Face. Open LLM Leaderboard, 2024. Accessed: March 4, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.283, + 0.905, + 0.338 + ], + "angle": 0, + "content": "[14] Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Suchin Gururangan, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. Editing models with task arithmetic. arXiv preprint arXiv:2212.04089, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.341, + 0.905, + 0.397 + ], + "angle": 0, + "content": "[15] Pavel Izmailov, Dmitrii Podoprikhin, Timur Garipov, Dmitry Vetrov, and Andrew Gordon Wilson. Averaging weights leads to wider optima and better generalization. arXiv preprint arXiv:1803.05407, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.399, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[16] Junguang Jiang, Baixu Chen, Junwei Pan, Ximei Wang, Dapeng Liu, Jie Jiang, and Mingsheng Long. Forkmerge: Mitigating negative transfer in auxiliary-task learning. Advances in neural information processing systems, 36:30367-30389, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.47, + 0.905, + 0.567 + ], + "angle": 0, + "content": "[17] Qiao Jin, Bhuwan Dhingra, Zhengping Liu, William Cohen, and Xinghua Lu. Pubmedqa: A dataset for biomedical research question answering. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2567-2577, 2019. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.569, + 0.905, + 0.626 + ], + "angle": 0, + "content": "[18] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.628, + 0.905, + 0.683 + ], + "angle": 0, + "content": "[19] Simon Lacoste-Julien, Martin Jaggi, Mark Schmidt, and Patrick Pletscher. Block-coordinate frank-wolfe optimization for structural svms. In International Conference on Machine Learning, pages 53-61. PMLR, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.686, + 0.905, + 0.727 + ], + "angle": 0, + "content": "[20] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.73, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[21] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[22] Yixiao Li, Yifan Yu, Chen Liang, Pengcheng He, Nikos Karampatziakis, Weizhu Chen, and Tuo Zhao. Loftq: Lorafine-tuning-aware quantization for large language models. arXiv preprint arXiv:2310.08659, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[23] Fuzhao Liu, Xiaotian Hu, Chengyu Chen, Zhijian Wang, Shuxin Xiao, Zhen Wang, Zhongxiang Wang, Chuanxin Xie, Eric Xing, and Song Han. Dare: Diversity-aware model merging for sparse-to-dense mixture-of-experts. arXiv preprint arXiv:2402.10887, 2024. 3, 4, 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3398" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[24] Liyang Liu, Yi Li, Zhanghui Kuang, J Xue, Yimin Chen, Wenming Yang, Qingmin Liao, and Wayne Zhang. Towards impartial multi-task learning. iclr, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.483, + 0.163 + ], + "angle": 0, + "content": "[25] Prateek Yadav Liu et al. Resolving interference when merging models. arXiv preprint arXiv:2306.01708, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.484, + 0.218 + ], + "angle": 0, + "content": "[26] Shikun Liu, Edward Johns, and Andrew J Davison. End-to-end multi-task learning with attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1871-1880, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.483, + 0.287 + ], + "angle": 0, + "content": "[27] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettle-moyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.287, + 0.483, + 0.344 + ], + "angle": 0, + "content": "[28] Zhenyi Lu, Chenghao Fan, Wei Wei, Xiaoye Qu, Dangyang Chen, and Yu Cheng. Twin-merging: Dynamic integration of modular expertise in model merging. Advances in Neural Information Processing Systems, 37:78905–78935, 2025. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.483, + 0.426 + ], + "angle": 0, + "content": "[29] Jiaqi Ma, Zhe Zhao, Xinyang Yi, Jilin Chen, Lichan Hong, and Ed H Chi. Modeling task relationships in multi-task learning with multi-gate mixture-of-experts. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1930-1939, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.426, + 0.483, + 0.483 + ], + "angle": 0, + "content": "[30] Jiaqi Ma, Zhe Zhao, Jilin Chen, Ang Li, Lichan Hong, and Ed H Chi. Snr: Sub-network routing for flexible parameter sharing in multi-task learning. In Proceedings of the AAAI conference on artificial intelligence, pages 216-223, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.483, + 0.483, + 0.551 + ], + "angle": 0, + "content": "[31] Sadhika Malladi, Tianyu Gao, Eshaan Nichani, Alex Damian, Jason D Lee, Danqi Chen, and Sanjeev Arora. Finetuning language models with just forward passes. Advances in Neural Information Processing Systems, 36:53038-53075, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.551, + 0.483, + 0.581 + ], + "angle": 0, + "content": "[32] Priyanka Mary Mammen. Federated learning: Opportunities and challenges. arXiv preprint arXiv:2101.05428, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.483, + 0.635 + ], + "angle": 0, + "content": "[33] Eric Mitchell, Kenton Lee, Alexander Khazatsky, Christopher D Manning, and Chelsea Finn. Editing models with task arithmetic. arXiv preprint arXiv:2212.04089, 2022. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.635, + 0.483, + 0.692 + ], + "angle": 0, + "content": "[34] Ramesh Nallapati, Bowen Zhou, Caglar Gulcehre, Bing Xiang, et al. Abstractive text summarization using sequence-to-sequence rnns and beyond. arXiv preprint arXiv:1602.06023, 2016.5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.483, + 0.748 + ], + "angle": 0, + "content": "[35] Shraman Prakash, Rohit Girdhar, Devamanyu Hazarika, Joao Carreira, and Andrew Zisserman. Unified model for image, video, audio and language tasks. arXiv preprint arXiv:2307.16184, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.483, + 0.817 + ], + "angle": 0, + "content": "[36] Samyam Rajbhandari, Jeff Rasley, Olatunjri Ruwase, and Yuxiong He. Zero: Memory optimizations toward training trillion parameter models. In SC20: International Conference for High Performance Computing, Networking, Storage and Analysis, pages 1-16. IEEE, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.483, + 0.859 + ], + "angle": 0, + "content": "[37] Sidak Pal Singh and Martin Jaggi. Merging models with fisher-weighted averaging. arXiv preprint arXiv:2111.09832, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[38] Johannes Stallkamp, Marc Schlipsing, Jan Salmen, and Christian Igel. The german traffic sign recognition benchmark: a multi-class classification competition. In *The 2011" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "international joint conference on neural networks, pages 1453-1460. IEEE, 2011. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[39] Ximeng Sun, Rameswar Panda, Rogerio Feris, and Kate Saenko. Adashare: Learning what to share for efficient deep multi-task learning. Advances in Neural Information Processing Systems, 33:8728-8740, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.235 + ], + "angle": 0, + "content": "[40] Anke Tang, Li Shen, Yong Luo, Liang Ding, Han Hu, Bo Du, and Dacheng Tao. Concrete subspace learning based interference elimination for multi-task model fusion. arXiv preprint arXiv:2312.06173, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[41] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.362 + ], + "angle": 0, + "content": "[42] Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R Bowman. Glue: A multi-task benchmark and analysis platform for natural language understanding. arXiv preprint arXiv:1804.07461, 2018. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.419 + ], + "angle": 0, + "content": "[43] Ke Wang, Nikolaos Dimitriadis, Guillermo Ortiz-Jimenez, François Fleuret, and Pascal Frossard. Localizing task information for improved model merging and compression. arXiv preprint arXiv:2405.07813, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.419, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[44] Zirui Wang, Daniel Duckworth, Shashank Nag, Kshitij Patil, Jonathon Shlens, Ekin D Cubuk, Barret Zoph, Joseph Campbell, Mitchell Wortsman, and Google Brain. lo-fi: distributed fine-tuning without communication. arXiv preprint arXiv:2210.11948, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.546 + ], + "angle": 0, + "content": "[45] Yongxian Wei, Anke Tang, Li Shen, Feng Xiong, Chun Yuan, and Xiaochun Cao. Modeling multi-task model merging as adaptive projective gradient descent. arXiv preprint arXiv:2501.01230, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.546, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[46] Mitchell Wortsman, Gabriel Ilharco, Samir Yitzhak Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. arXiv preprint arXiv:2203.05482, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.631, + 0.905, + 0.688 + ], + "angle": 0, + "content": "[47] Jianxiong Xiao, Krista A Ehinger, James Hays, Antonio Torralba, and Aude Oliva. Sun database: Exploring a large collection of scene categories. International Journal of Computer Vision, 119:3-22, 2016. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.688, + 0.905, + 0.745 + ], + "angle": 0, + "content": "[48] Prateek Yadav, Derek Tam, Leshem Choshen, Colin A Raffel, and Mohit Bansal. Ties-merging: Resolving interference when merging models. Advances in Neural Information Processing Systems, 36:7093-7115, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.745, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[49] Enneng Yang, Li Shen, Guibing Guo, Xingwei Wang, Xiaochun Cao, Jie Zhang, and Dacheng Tao. Model merging in llms, mllms, and beyond: Methods, theories, applications and opportunities, 2024. URL https://arxiv.org/abs/2408.07666, 2408.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.816, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[50] Enneng Yang, Zhenyi Wang, Li Shen, Shiwei Liu, Guibing Guo, Xingwei Wang, and Dacheng Tao. Adamerging: Adaptive model merging for multi-task learning. arXiv preprint arXiv:2310.02575, 2023. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.872, + 0.905, + 0.903 + ], + "angle": 0, + "content": "[51] Enneng Yang, Li Shen, Zhenyi Wang, Guibing Guo, Xiaojun Chen, Xingwei Wang, and Dacheng Tao. Representa" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3399" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "tion surgery for multi-task model merging. arXiv preprint arXiv:2402.02705, 2024. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.176 + ], + "angle": 0, + "content": "[52] Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How transferable are features in deep neural networks? Advances in neural information processing systems, 27, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.484, + 0.261 + ], + "angle": 0, + "content": "[53] Jun Yu, Yutong Dai, Xiaokang Liu, Jin Huang, Yishan Shen, Ke Zhang, Rong Zhou, Eashan Adhikarla, Wenxuan Ye, Yixin Liu, et al. Unleashing the power of multitask learning: A comprehensive survey spanning traditional, deep, and pretrained foundation model eras. arXiv preprint arXiv:2404.18961, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.484, + 0.318 + ], + "angle": 0, + "content": "[54] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. Advances in neural information processing systems, 33:5824-5836, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "3400" + } + ] +] \ No newline at end of file diff --git a/2025/FW-Merging_ Scaling Model Merging with Frank-Wolfe Optimization/56e48ade-9cb7-4688-a0be-edb71b4d459b_origin.pdf b/2025/FW-Merging_ Scaling Model Merging with Frank-Wolfe Optimization/56e48ade-9cb7-4688-a0be-edb71b4d459b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8b705d88577c976e3dbf4c8363479555bad2ec98 --- /dev/null +++ b/2025/FW-Merging_ Scaling Model Merging with Frank-Wolfe Optimization/56e48ade-9cb7-4688-a0be-edb71b4d459b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9be24e84c0106ece887189c0e9bab1a5f4d167979e89cd6f85d172e486ced96d +size 1296807 diff --git a/2025/FW-Merging_ Scaling Model Merging with Frank-Wolfe Optimization/full.md b/2025/FW-Merging_ Scaling Model Merging with Frank-Wolfe Optimization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..278cc7d7cf758c105a5ccf9a4873a94fd7ca7a39 --- /dev/null +++ b/2025/FW-Merging_ Scaling Model Merging with Frank-Wolfe Optimization/full.md @@ -0,0 +1,355 @@ +# FW-Merging: Scaling Model Merging with Frank-Wolfe Optimization + +Hao (Mark) Chen1 Shell Xu Hu2 Wayne Luk1 Timothy Hospedales2 Hongxiang Fan1 +1Imperial College London, UK 2Samsung AI Center, Cambridge, UK + +hcl620@ic.ac.uk shell.hu@samsung.com w.luk@imperial.ac.uk t.hospedales@samsung.com hongxiang.fan@imperial.ac.uk + +# Abstract + +Model merging has emerged as a promising approach for multi-task learning (MTL), offering a data-efficient alternative to conventional fine-tuning. However, with the rapid development of the open-source AI ecosystem and the increasing availability of fine-tuned foundation models, existing model merging methods face two key limitations: (i) They are primarily designed for in-house fine-tuned models, making them less adaptable to diverse model sources with partially unknown model and task information, (ii) They struggle to scale effectively when merging numerous model checkpoints. To address these challenges, we formulate model merging as a constrained optimization problem and introduce a novel approach: Frank-Wolfe Merging (FW-Merging). Inspired by Frank-Wolfe optimization, our approach iteratively selects the most relevant model in the pool to minimize a linear approximation of the objective function and then executes a local merging similar to the Frank-Wolfe update. The objective function is designed to capture the desired behavior of the target-merged model, while the fine-tuned candidate models define the constraint set. More importantly, FW-Merging serves as an orthogonal technique for existing merging methods, seamlessly integrating with them to further enhance accuracy performance. Our experiments show that FW-Merging scales across diverse model sources, remaining stable with 16 irrelevant models and improving by $15.3\%$ with 16 relevant models on 20 CV tasks, while maintaining constant memory overhead—unlike the linear overhead of data-informed merging methods. Compared with the state-of-the-art approaches, FW-Merging surpasses the data-free merging method by $32.8\%$ and outperforms the data-informed Adamerging by $8.39\%$ when merging 20 ViT models. Our code is open-sourced at here. + +# 1. Introduction + +Multi-task learning (MTL)-based fine-tuning adapts a single pre-trained Large Language Model (LLM) for multiple + +downstream applications, reducing the deployment overhead of separately fine-tuning multiple models [53]. However, it still demands a large amount of high-quality data, which might only exist in the private domain [32], and significant compute resources [36]. To mitigate these issues, model merging has emerged as a promising technique for fusing fine-tuned models within the parameter space [14, 48, 50]. Existing model merging methods can be broadly classified into two categories: data-free methods [14, 15, 48], and data-informed methods [50, 51], which optimize merge coefficients based on additional data. + +While these approaches have proven effective, several key limitations hinder their scalability and broader adoption. First, these methods adjust merging coefficients based on the known capabilities of the models on specific tasks to optimize performance, which is less robust when dealing with diverse model sources with unknown information1. This is primarily caused by the inability to distinguish high-quality models from poorly fine-tuned ones in an unknown model setting. Second, when scaling up these approaches to a large number of unknown models, these methods struggle and can result in significant performance degradation. As shown in Figure 1a, our profiling study demonstrates a performance reduction ranging from $18.9\%$ to $64.4\%$ . These limitations are further amplified by the fast-growing open-source AI ecosystem, where platforms such as Hugging Face have driven a surge in the release of powerful LLMs with many lacking complete information. Given that merging open-source models has repeatedly shown the potential to produce top-ranking LLMs on major model leaderboards [13], developing scalable and robust merging techniques is essential to harness the growing number of open-source models, further enhancing performance and widening the adoption of model merging. + +To effectively scale model merging and leverage the vast collection of open-sourced models with unknown capabilities, the new model merging method must exhibit two fundamental scaling properties: 1) as more irrelevant models + +![](images/3da232e398f85080f32a85987872d1e92462d5cc7f196443b585abe00fb321bb.jpg) +(a) 4 CV tasks with irrelevant model addition. + +![](images/e75d2e7c3bb354a233d93ddc1b22cb9fdf5316c7c4616b34d65c4b71078a47d7.jpg) +(b) 20 CV tasks with relevant model addition. + +![](images/953a1664a8cf8cb344efc734e23da2d593146a26f1633a292b439cf709e91a12.jpg) +(c) NLP tasks with noisy model addition. +Figure 1. Performance scaling of FW-Merging across CV tasks. (a) demonstrate robustness to irrelevant models, while (b) show improved performance with relevant models. (c) analyzes performance degradation when incorporating a noisy model initialized from a different pre-trained checkpoint. Detailed results and experimental setup are discussed in Section 4.4. + +are added to the merging pool, the performance should remain unaffected, and 2) as more relevant models are added to the merging pool, the performance should steadily increase, converging towards the optimal performance. To this end, we revisit model merging and formulate it as a constrained optimization problem, where the objective function dictates the desirable behavior of the final merged model, and fine-tuned checkpoints form the constraint set. Inspired by Frank-Wolfe optimization, we introduce Frank-Wolfe Merging (FW-Merging), an iterative algorithm designed to enhance merging efficiency while maintaining robustness at scale. FW-Merging comprises three principal stages in each iteration: (1) Relevance Evaluation: Instead of merging models arbitrarily, we obtain the linear approximation of the objective function using gradients of the current model, revealing the most beneficial direction for improvement. (2) Model Selection: The most relevant checkpoints are selected from the constraint set by minimizing the linear approximation, ensuring that each step incorporates task-specific knowledge with minimal interference. (3) Knowledge Integration: The selected checkpoint is integrated using an orthogonal merging method, striking a balance between adaptation and stability in the merged model. + +We demonstrate the effectiveness of FW-Merging with a diverse pool of fine-tuned checkpoints across various language and vision tasks, compared to both data-free and data-informed model merging methods as well as traditional MTL-based fine-tuning. As shown in Figure 1, FW-Merging satisfy our two fundamental scaling properties: accuracy performance does not drop when 16 irrelevant models are added (compared to a $49\%$ drop in task-arithmetic) and steadily improves by $15.3\%$ when 16 relevant models are included. Additionally, FW-Merging requires only constant memory overhead, as it selects and merges a fixed number of models at a time. In contrast, methods that optimize merging coefficients [50] or resolve parameter interference [48] must store all models in memory, leading to linear overhead. Moreover, FW-Merging + +exhibits greater robustness to noisy models lacking critical information, such as their initialization point. As shown in Figure 1c, FW-Merging experiences minimal performance degradation when a misinitialized model is introduced, whereas Ties suffers a performance drop of up to $3.2\%$ . FW-Merging outperforms state-of-the-art data-free merging method by $32.8\%$ and the data-informed method Adamerging by $8.39\%$ when merging 20 ViT models. On the language benchmarks, FW-Merging achieves $6.3\%$ improvement over the best model merging method across discriminative and generative tasks, while even surpassing the performance of traditional MTL using only $3.4\%$ of its required data. Our results position FW-Merging as an effective solution to scale model merging to the next level. + +Our contributions can be summarized as follows: + +- Identify scalability and robustness issues in existing model merging techniques through experiments, highlighting the urgent need for large-scale model merging. +- Formulate model merging as a constrained optimization problem with an objective function that explicitly captures the desired behavior of the final merged model. +- Introduce Frank-Wolfe Merging, a novel iterative method that autonomously guides the merged model toward an optimized direction, even with large sets of black-box open-source checkpoints. +- Evaluate our proposed approach on extensive benchmarks, demonstrating its effectiveness and scalability. + +# 2. Related Work + +Efficient Multi-Task Learning. In traditional Multi-Task Learning (MTL), a single model is trained on a dataset containing multiple tasks to enable the model to acquire diverse capabilities [2]. However, a significant challenge in traditional MTL is the issue of negative transfer [16]. To mitigate this, architecture-based approaches have been developed, such as parameter sparsification [26, 39] and shared structure modularization [29, 30]. On the optimization side, + +methods to resolve gradient conflicts [6, 54] and domination of gradient or learning rate [5, 24] have been proposed. With the rise of Large Language Models (LLMs), MTL faces additional challenges, particularly the high computational costs. To address these challenges, strategies like parameter-efficient fine-tuning [11, 20, 21] and memory-efficient fine-tuning [8, 22, 31] have been introduced to minimize both memory and computational resource usage. More recently, model merging has emerged as a promising approach to make MTL more compute- and data-efficient. + +Model Merging. While pre-merging methods prepare favorable conditions for merging, during-merging techniques combine multiple neural networks into a single model while retaining or enhancing their capabilities [49]. In this work, we focus on during-merging methods. Early insights into neural network landscapes [10] revealed that linear interpolation between models exposes useful loss surface properties, laying the foundation for weight averaging—a core merging technique. Simple averaging widens optima and improves generalization [15], evolving into advanced methods like model soups [46] and heterogeneous model merging. Recent advances introduce more structured approaches, such as Fisher-Weighted Averaging [37], which incorporates Fisher information to weight parameters more effectively, and Permutation Alignment methods like Git Re-Basin [1], which address weight permutation symmetries. Interference Resolution techniques, including TIES [25] and DOGE [45], mitigate parameter conflicts either through explicit alignment or projective gradient descent. Task Arithmetic [33] enables weight-space operations to combine task-specific behaviors in language models, while Diversity-Aware Merging, such as DARE [23], leverages model diversity to improve sparse-to-dense integration. In contrast to the data-free methods mentioned above, data-informed methods [40, 50, 51] optimize merging coefficients using additional data. Model merging is impactful for LLMs, enabling efficient knowledge integration without full retraining, facilitating distributed finetuning [44], multi-task learning [35], and cost-effective model adaptation. + +# 3. Method + +# 3.1. Preliminary: Frank-Wolfe algorithm + +The Frank-Wolfe (FW) algorithm [9], also known as the conditional gradient method, is an iterative optimization algorithm for constrained optimization problems of the form: + +$$ +\min _ {x \in \mathcal {C}} f (x) \tag {1} +$$ + +where $f$ is a continuously differentiable function, and $\mathcal{C}$ is a compact convex set. The algorithm follows an elegant + +geometric intuition: at each iteration $t$ , FW first identifies which vertex of $\mathcal{C}$ yields the steepest descent direction and then moves towards this vertex to decrease the value of the objective function. More specifically, FW algorithm: + +1. Constructs a linear subproblem of the original optimization (a.k.a. linear minimization oracle) using first-order Taylor expansion at the point $x_{t}$ : + +$$ +\operatorname {L M O} \left(\mathcal {C}, x _ {t}\right) := \underset {s \in \mathcal {C}} {\arg \min } \langle s, \nabla f \left(x _ {t}\right) \rangle \tag {2} +$$ + +2. Finds the vertex $s_t$ of the feasible set $\mathcal{C}$ by picking $s_t \in \mathrm{LMO}(\mathcal{C}, x_t)$ . +3. Takes a careful step from the current point $x_{t}$ towards this direction $s_t - x_t$ , maintaining feasibility through the convex combination: $x_{t + 1} = (1 - \gamma_t)x_t + \gamma_t s_t$ . + +The step size $\gamma_t \in [0,1]$ can be chosen by line search + +$$ +\gamma_ {t} = \underset {\gamma \in [ 0, 1 ]} {\arg \min } f \left((1 - \gamma) x _ {t} + \gamma s _ {t}\right), \tag {3} +$$ + +which ensures a sufficient decrease in $f(\cdot)$ at each FW step. + +To determine when to stop, the $FW$ gap is used to measure the suboptimality in terms of the proximity to the best solution of LMO: + +$$ +g _ {t} := \max _ {s \in \mathcal {C}} \langle - \nabla f (x _ {t}), s _ {t} - x _ {t} \rangle , \tag {4} +$$ + +which is non-negative by definition. + +# 3.2. Frank-Wolfe Model Merging + +We consider the problem of fine-tuning a pre-trained foundation model on new tasks. Given a pre-trained model $\theta_0$ and previously fine-tuned models $\{\theta_1^*,\theta_2^*,\dots ,\theta_n^*\}$ , we aim to fine-tune the $(n + 1)$ -th model $\theta_{n + 1}$ on new tasks as a convex combination of the previous models with optimal merging coefficients. + +To this end, we propose a Frank-Wolfe based model merging framework, which is described as follows. + +$$ +\min _ {\lambda} \ell \left(\sum_ {i = 1} ^ {n} \lambda_ {i} \theta_ {i} ^ {*}\right) \quad \text {s . t .} \quad \sum_ {i = 1} ^ {n} \lambda_ {i} = 1, \quad \lambda_ {i} \geq 0, \tag {5} +$$ + +where $\lambda = \{\lambda_1, \dots, \lambda_n\}$ are the merging coefficients, and $\ell$ is a loss function formulated for a specific goal, such as aligning the merged and individual models [45] or satisfying a task objective [50]. + +Potentially, a scaling issue of this formulation appears when the number of fine-tuned models $n$ is large, since we need to keep all the fine-tuned models in memory. To address this, we propose a reformulation of the problem eq. (5) as follows: + +$$ +\min _ {\theta \in \mathcal {M}} \ell (\theta), \tag {6} +$$ + +where $\mathcal{M} \coloneqq \mathrm{conv}(\{\theta_i^*\}_{i=1}^n)$ is the convex hull of the set of previously fine-tuned models. + +![](images/18f1e2940049aa7754c9381767ffdb58bccee810266c06bb4df81659d2a94216.jpg) +Figure 2. Illustration of model merging methods. $\Theta_{A}$ is an irrelevant model, while $\Theta_{B}$ and $\Theta_{C}$ are relevant models. Darker regions indicate higher objective function loss. Task Arithmetic treats all task vectors equally, failing to move optimally. Adamerging assigns different coefficients, moving towards more desirable direction but suffer from slow convergence due to interference from $\Theta_{a}A$ . FW-Merging iteratively selects the most relevant model to merge and adapts step sizes, efficiently reaching the optimum after $T$ iterations. + +![](images/78d1279bd2ea75d76c428caa5197a219cd2617832cc9b90ba2ce66874201f725.jpg) + +![](images/2c48118c9254e59a646f332c633a407a89b5952a889e64411c5ac0f0044437cf.jpg) + +![](images/5fa98ac24c3c56b40d8c7c6d370663f6d3fcc490c0150901bb21ab21df69c251.jpg) + +# Proposition 1. The optimization problems in equations (5) and (6) are equivalent. + +Proof. By definition of convex hull, any point $\theta \in \mathcal{M}$ can be written as a convex combination of the vertices $\{\theta_i^*\}_{i=1}^n$ , i.e., $\theta = \sum_{i=1}^{n} \lambda_i \theta_i^*$ where $\lambda \in \Delta^n := \{\lambda \in \mathbb{R}^n \mid \sum_{i=1}^{n} \lambda_i = 1, \lambda_i \geq 0\}$ . Therefore: + +$$ +\min _ {\theta \in \mathcal {M}} \ell (\theta) = \min _ {\lambda \in \Delta^ {n}} \ell \Bigl (\sum_ {i = 1} ^ {n} \lambda_ {i} \theta_ {i} ^ {*} \Bigr). +$$ + +This shows that any solution of one problem can be mapped to a solution of the other problem with the same objective value. + +Since the FW algorithm requires the initial solution to be an interior point of the constraint set, we add the initial solution $\theta_0$ to form a new constraint set $\mathcal{M} \coloneqq \mathrm{conv}(\mathcal{M} \cup \{\theta_0\})$ , which we still denote as $\mathcal{M}$ for simple notations. A nice property of this reformulation is that the LMO can be simplified to + +$$ +\operatorname {L M O} \left(\left\{\theta_ {i} ^ {*} \right\} _ {i = 1} ^ {n}, \theta_ {t}\right) = \underset {s \in \left\{\theta_ {1} ^ {*}, \dots , \theta_ {n} ^ {*} \right\}} {\arg \min } \langle \nabla \ell \left(\theta_ {t}\right), s \rangle \tag {7} +$$ + +This is because for linear programming problems over convex sets, the optimal solution is always attained at the vertices of the constraint set. Algorithm 1 details the steps, and Figure 2 provides an overview. + +# 3.3. Design choices of the algorithm + +The above algorithm illustrates the key ingredients of Frank-Wolfe merging: LMO, stopping criterion $g_{t}$ , line search routine, and the merging function. We discuss in this section the design choices of these components. + +Merging function The main deviation from the classical FW algorithm is the reinterpretation of the FW update $x_{t + 1} = (1 - \gamma_t)x_t + \gamma_t s_t$ as a local merging between $\theta_t$ + +Algorithm 1: Frank-Wolfe Merging: +Input: Initial solution $\theta_0$ ; Fine-tuned checkpoints $\{\theta_i^*\}_{i=1}^n$ ; FW budget $T$ . +Output: Merged model $\theta_{merged}^*$ . +1: if $\theta_0 \notin \mathcal{M}$ then $\mathcal{M} := \mathcal{M} \cup \{\theta_0\}$ +2: for $t = 0 \dots T$ do +3: Let $s_t := \mathrm{LMO}(\theta_t)$ and $d_t := s_t - \theta_t$ +4: if $g_t := \langle -\nabla \ell(\theta_t), d_t \rangle \leq \epsilon$ then +5: return $\theta_{merged}^* \gets \theta_t$ +6: end if +7: Line-search: $\gamma_t \in \arg \min_{\gamma \in [0,1]} \ell(\theta_t + \gamma d_t)$ +8: Update: $\theta_{t+1} := \mathrm{MergeFn}(\theta_t, s_t, \gamma_t)$ +9: end for +10: return $\theta_{merged}^* \gets \theta_T$ + +and $s_t$ . We denote by MergeFn the customizable merging function as long as the merged model stays in the convex hull $\mathcal{M}$ . The most straightforward merging function is the convex combination: + +$$ +\operatorname {M e r g e F n} \left(\theta_ {t}, s _ {t}, \gamma_ {t}\right) := \left(1 - \gamma_ {t}\right) \theta_ {t} + \gamma_ {t} s _ {t}, \tag {8} +$$ + +which corresponds to the Task-Arithmetic [33] method. The step size $\gamma_{t}$ makes sure the merged model stays in $\mathcal{M}$ . + +It is natural to ask whether other existing model merging methods, such as TIES-Merging [25] and DARE [23] could also be used as MergeFn. The problem with these sophisticated merging methods is that the merged model might leave the constraint set, and thus violate the assumption of maintaining feasibility required by the classical FW theory. We verified in practice that these less rigorous merging functions might achieve better performance in certain cases but they generally cause more stability issues. Therefore, we do not consider these merging functions from the current comparison. + +Hard FW v.s. Soft FW In the case of deep learning, the optimization problem is non-convex, additional efforts are needed to better characterize the loss landscape and prevent the LMO from being dominated by one or a few fine-tuned models, which occurs because the linear approximation of $\ell(\theta)$ is an inaccurate sketch of the original objective function. Instead of relying on the argmin of linear subproblems, we fetch the top- $k$ vertices of LMO, $\{\tilde{s}_j\}_{j=1}^k$ . A more subtle top- $k$ operation can be performed in a task-wise fashion if the original objective function involves multi-tasks. + +Given the top- $k$ vertices, we now go back to eq. (5) to obtain the optimal merging coefficients $\{\lambda_j^*\}_{j=1}^k$ . Note that this inner optimization is a reduced version of original eq. (5) because hosting $k$ models in memory would not be a problem. We also remove the line search step as this gives a new merging function of the form + +$$ +\operatorname {M e r g e F n} \left(\theta_ {t}, \left\{\tilde {s} _ {j} \right\} _ {j = 1} ^ {k}, \left\{\lambda_ {j} ^ {*} \right\} _ {j = 1} ^ {k}\right) := \theta_ {t} + \sum_ {j = 1} ^ {k} \lambda_ {j} ^ {*} \left(\tilde {s} _ {j} - \theta_ {t}\right). \tag {9} +$$ + +We call this oracle soft LMO in comparison to the argmin version which we call hard LMO. + +Task-wise LMO v.s. layer-wise LMO The naive implementation of FW-Merging would vectorize the whole model weights $\theta$ and then solve LMO. We call this taskwise LMO. Since different layers contribute differently to model performance [52], a layer-wise LMO may yield better model merging. To incorporate this, the constraint set is redefined as a Cartesian product of convex hulls for each layer: $\mathcal{M} := \mathcal{M}_1 \times \dots \times \mathcal{M}_L$ , where $L$ is the number of layers and $\mathcal{M}_l := \mathrm{conv}\left(\left\{\theta_i^{*,l}\right\}_{i=1}^n\right)$ . The LMO is then conducted layer-wise: + +$$ +\operatorname {L M O} \left(\left\{\theta_ {i} ^ {* , l} \right\} _ {i = 1} ^ {n}, D, \theta_ {t} ^ {l}\right) = \underset {s ^ {l} \in \left\{\theta_ {1} ^ {* , l}, \dots , \theta_ {n} ^ {* , l} \right\}} {\arg \min } \left\langle \nabla \ell \left(\theta^ {t}\right) ^ {l}, s ^ {l} \right\rangle . \tag {10} +$$ + +This version can be viewed as a block-coordinate Frank-Wolfe algorithm [19], which is applied when the problem has a natural decomposition into blocks. + +# 4. Experiments + +# 4.1. Experiment Setup + +**Benchmarks.** Our primary objective is to evaluate the effectiveness of our method in scenarios where the number of models greatly exceeds the number of evaluation tasks, and each model's capabilities are unknown in advance. + +- Vision Tasks: Following the setting of TALL [43], we use 20 ViT-B/32 models, each fine-tuned on a different vision task. The number of models to be merged is intentionally set to be significantly larger than the number of evaluation tasks, allowing us to assess the scalability of model merging methods. The evaluation benchmarks consist of four tasks: SUN397 [47], Stanford Cars [18], GTSRB [38], and DTD [7]. +- Language Discriminative Tasks: We prepare 8 RoBERTa checkpoints [27] fine-tuned on eight tasks from the GLUE benchmark, following the practice in [28]. The merged model is then evaluated on four tasks from the GLUE benchmark [42]: MNLI, QNLI, QQP, and RTE. +- Language Generative Tasks: We collect 16 LLaMA2-7B models [41] fine-tuned with LoRA [12] on various tasks from Hugging Face. These models have unknown and uncontrolled capabilities, making them equivalently black-box models. Our goal is to evaluate the robustness of model merging methods in this challenging setting. The evaluation benchmarks include CNN/DM summarization [34], PubMedQA [17], and HumanEval [4]. + +Metrics. For vision tasks, we report the classification accuracy. Following [28], we report the average normalized score for language tasks to account for differences in task-specific score ranges to account for variations in task-specific score ranges. The normalized score is computed as $\mathrm{Score}_{\mathrm{normalized}} = \frac{1}{T}\sum_{t=1}^{T}\frac{\mathrm{Score}(f(\pmb{\theta}^*))}{\mathrm{Score}(f_t(\pmb{\theta}_t))}$ . + +Baselines. We compare FW-Merging with both data-informed and data-free model merging methods. For data-informed model merging, we compare FW-Merging against Adamerging [50], Surgery [51], and Concrete Merging [40]. To ensure a fair comparison, these methods are trained only on the same tasks as FW-Merging. For data-free model merging, we compare FW-Merging with Fisher Merging, Weight Averaging, RegMean Merging, Task Arithmetic [33], Ties-Merging [48], and DARE Merging [23] across both language and vision tasks. Additionally, we fine-tune one model on the discriminative language benchmark and another on the generative language benchmark to serve as additional baselines. + +Implementations. We implement two FW-Merging variants: $\mathrm{FW}_{\text {hard }}$ , which uses hard FW with layer-wise LMO, and $\mathrm{FW}_{\text {soft }}$ , which employs soft FW with task-wise LMO (Section 3.3). For $\mathrm{FW}_{\text {soft }}$ , layer-wise coefficients are optimized via gradient descent on the training dataset to solve eq. 5, differing from Adamerging [50] by minimizing cross-entropy loss on training data rather than entropy on test samples. On language benchmarks, the training dataset consists of 100 samples per task, and $\mathrm{FW}_{\text {hard }}$ runs + +Table 1. Performance on 4 Discriminative Tasks when merging 8 RoBERTa and 3 Generative Tasks when merging 16 LLaMA2-7B. + +
Method4 Disc. Tasks (8 Models)3 Gen. Tasks (16 Models)Avg. Normalized Score
Pretrained49.677.163.4
Traditional MTL73.181.277.2
Task Arithmetic (w/ DARE)77.316.847.1
Ties-Merging (w/ DARE)75.646.661.1
Task Arithmetic80.875.978.4
Ties-Merging64.378.571.4
FWhard (Ours)85.481.183.1
+ +Table 2. Performance on 4 CV Tasks when merging 20 ViT-B/32. + +
MethodSUN397CarsGTSRBDTDAvg.
Pretrained62.359.732.643.849.6
DARE (TIES)5.92.316.711.89.2
Task Arithmetic20.412.229.822.321.2
Ties-Merging51.036.257.740.646.4
Weight Averaging64.259.643.146.553.4
Fisher Merging64.663.843.046.954.6
RegMean65.562.259.753.960.3
LW Concrete AM62.560.388.054.766.3
Adamerging66.470.195.164.073.9
Surgery69.771.896.673.477.9
FWhard (Ours)66.569.995.164.574.0
FWsoft (Ours)72.974.896.876.080.1
+ +for 10 iterations, initialized with Task Arithmetic's merged model. For vision tasks, it runs for 3 iterations. $FW_{\text{soft}}$ runs for 15 iterations on vision benchmarks, initialized with the pre-trained model. Training datasets consist of 100 samples from MNLI, QNLI, QQP, and RTE [42] for discriminative tasks; CNN/DM [34], CodeAlpaca-20k [3], and PubMedQA [17] for generative tasks; and SUN397 [47], Stanford Cars [18], GTSRB [38], and DTD [7] for vision tasks. + +# 4.2. Comparison with Model Merging Methods + +We evaluate FW-Merging against both data-informed and data-free model merging approaches across language and vision benchmarks. Table 1 reports the results for language tasks, including both discriminative and generative settings, while Table 2 presents results on vision benchmarks. + +Language Tasks. $\mathsf{FW}_{\text{hard}}$ achieves the highest average normalized score across language benchmarks, consistently surpassing prior model merging baselines. Specifically, $\mathsf{FW}_{\text{hard}}$ improves upon Task Arithmetic by 4.6 points, Ties-Merging by 11.7, and DARE (Ties) by 9.8 on discriminative tasks. Table 4 shows that $\mathsf{FW}_{\text{hard}}$ also outperforms data-informed Adamerging by 5.9 points. For generative tasks, $\mathsf{FW}_{\text{hard}}$ outperforms Task Arithmetic by 5.2 points, Ties-Merging by 2.6, and DARE (Ties) by 34.5. Interestingly, while Task Arithmetic outperforms Ties-Merging on discriminative tasks by a margin of 16.5 points, it lags behind by 2.6 points on the more challenging generative tasks. + +This discrepancy likely arises from increased interference among task vectors as more checkpoints are merged. Unlike Ties-Merging, which explicitly resolves merging conflicts, Task Arithmetic lacks a reconciliation mechanism, making it more susceptible to such interference. In contrast, $\mathsf{FW}_{\text {hard }}$ consistently outperforms both Ties-Merging and Task Arithmetic by selectively merging only the most relevant model parameters in each iteration. This targeted approach effectively mitigates interference, leading to more stable and robust performance across both discriminative and generative tasks. + +Vision Tasks. $\mathsf{FW}_{\mathrm{soft}}$ achieves state-of-the-art performance across multiple vision benchmarks, surpassing data-informed methods like Adamerging and Surgery. As shown in Table 2, $\mathsf{FW}_{\mathrm{hard}}$ surpasses Adamerging, Concrete Merging, and all data-free merging methods in overall performance. Additionally, $\mathsf{FW}_{\mathrm{soft}}$ attains the highest accuracy (80.1%), outperforming Adamerging by 6.2% and Surgery by 2.2%. Unlike Surgery, which requires additional task-specific parameters and multiple forward passes per inference, our approach efficiently adapts to diverse visual tasks without increasing storage or inference complexity. + +In general, data-free merging methods show significantly lower performance compared to data-informed approaches while merging a large number of models, when the models' capabilities do not precisely align with the evaluation tasks. This limitation arises because data-free methods treat all models equally, merging them without considering their unique capabilities, which amplifies interference between models. In contrast, data-informed merging methods achieve superior performance by optimizing merging coefficients on datasets as they allow for explicit control over desirable capabilities. FW-Merging, in particular, enhances scalability via hard model selection based on the linear approximation minimization. + +# 4.3. Comparison with Traditional MTL + +We compare FW-Merging with models fine-tuned using traditional MTL on discriminative and generative tasks. In each case, one single model is fine-tuned across all tasks, with performance and computational cost reported + +Table 3. Merging Methods' Performance vs. Number of Models when Adding Relevant vs. Irrelevant Models. + +
#Models4 CV Tasks +When "Irrelevant" Models Added20 CV Tasks +When "Relevant" Models Added
DARETaskTiesAMFWsoftDARETaskTiesAMFWsoft
473.670.365.775.274.157.359.260.259.659.2
664.164.164.474.673.954.058.461.664.063.2
848.557.163.673.474.048.257.263.964.764.5
1040.652.962.972.873.943.555.063.965.466.2
1231.747.961.971.474.139.152.363.765.167.5
1423.640.157.769.974.036.148.962.064.168.0
1617.933.352.568.674.132.345.359.762.768.3
1812.225.949.264.674.027.840.357.360.768.4
209.221.246.461.074.224.236.355.658.968.3
+ +![](images/df2de3ebb3c8975898230d9a993b3f46bc15f236e32ff6dcedd62fd22858c3b4.jpg) +Figure 3. Linear Approximation of the Objective Function of Model Checkpoints Across Different Tasks in a Frank-Wolfe Iteration. The x-axis represents the checkpoints, and each graph shows the linear approximation result for each task. + +Table 4. Costs and Perf. of methods on NLP discriminative tasks. + +
MethodData Samples/TaskTime CostPerf.
Traditional MTL2.9K4.2h73.1
Data-free Merging0080.8
Data-informed Merging1.6K2min79.5
FW Merging1002min85.4
+ +in Table 1 and Table 4. Traditional MTL achieves an average score of 77.2, lower than that of FW-Merging (83.1). On discriminative tasks, MTL scores 73.1, trailing FW-Merging (85.4). For generative tasks, MTL scores 81.2, while FW-Merging closely matches it at 80.8, suggesting that FW-Merging's performance matches that of traditional MTL. + +As shown in Table 4, FW-Merging demonstrates a substantial advantage in efficiency. Traditional MTL requires fine-tuning on 2.9K samples per task and takes 4.2 hours of training time, which is computationally intensive. In contrast, FW-Merging only requires 100 training samples per task and completes the merging process in just 2 minutes. This huge reduction in computational cost underscores the effectiveness of FW-Merging compared to traditional MTL. Moreover, FW-Merging has a key advantage over traditional MTL: while MTL requires a large volume of high-quality data for optimal performance, FW-Merging needs only a small set of data because: $I$ ) it optimizes merging coefficients based on models' character + +istics, which simplifies the optimization space, and 2) it uses model weights as inputs, which are much more information-dense representations than data, enabling more efficient objective learning. + +FW-Merging is a post-training technique that does not require access to original training data, making it ideal for privacy-sensitive or data-scarce scenarios. Overall, the results suggest that FW-Merging is a scalable, efficient alternative to traditional MTL, providing comparable performance at a reduced computational cost. + +# 4.4. Scaling to More Models and Tasks + +We investigate the performance scaling of different merging methods with the number of models, as shown in Figure 1a, Figure 1b, and Table 3. In large-scale model merging, models from open-source platforms vary in quality. To simulate this, we use 20 ViT-B/32 models fine-tuned on tasks that are either included in the evaluation benchmark or not. A model is irrelevant if its fine-tuning dataset does not match the training split of the evaluation task, and relevant if it matches. To ensure fair comparison, the total number of training iterations run by $\mathrm{FW}_{\mathrm{soft}}$ is the same as that of Adamreging. + +As shown in Table 3, adding irrelevant models sharply reduces the performance of data-free methods: DARE by + +Table 5. Ablation on design variants of FW-Merging. + +
Coefficient λMethodLMOScore
Vision Tasks
OptimizedFWsoftLayer-wise79.7
OptimizedFWsoftTask-wise80.1
UnoptimizedFWsoftLayer-wise69.8
UnoptimizedFWsoftTask-wise70.3
-FWhardLayer-wise74.0
-FWhardTask-wise73.7
NLP Discriminative Tasks
-FWhardLayer-wise85.4
-FWhardTask-wise78.2
+ +$64.4\%$ , Task Arithmetic by $49.1\%$ , and Ties by $19.1\%$ , likely due to task interference and equal treatment of all models. Data-informed methods degrade less, with Adamerging dropping by $14.2\%$ . In contrast, $\mathsf{FW}_{\mathrm{soft}}$ remains highly stable, fluctuating only from $73.9\%$ to $74.1\%$ as more models are added. In Figure 3, we examine the linear approximation of different checkpoints for a specific task and find that the model fine-tuned on the task consistently yields the most negative linear approximation. This indicates that in the Frank-Wolfe update, the most relevant checkpoint is chosen as the direction for merging, allowing FW-Merging to iteratively improve the merged model in the optimized direction within the constraint set. The inner product between gradients and model parameters serves as a reliable indicator of model relevance, with minimized computational cost, further demonstrating FW-Merging's scalability even in the presence of irrelevant models. + +Adding relevant models should ideally improve performance, but data-free methods still degrade as shown in Table 3: DARE by $33.1\%$ , Task Arithmetic by $22.9\%$ , and Ties by $4.6\%$ , with Ties performing best by mitigating parameter conflicts. Data-informed methods like Adamerging fluctuate between $58.9\%$ and $64.7\%$ as merging complexity increases, whereas $\mathsf{FW}_{\mathrm{soft}}$ steadily improves from $59.2\%$ to $68.3\%$ by iteratively selecting the most relevant models, facilitating smoother convergence. These results underscore FW-Merging's effectiveness as a scalable solution for large-scale model merging. + +# 4.5. Ablation Studies + +Design variants. Table 5 compares the design variants of FW-Merging (Section 3.3). Task-wise LMO aligns better with $\mathrm{FW}_{\mathrm{soft}}$ , improving performance slightly by 0.5 points over layer-wise LMO, while layer-wise LMO is more effective for $\mathrm{FW}_{\mathrm{hard}}$ , especially on language tasks, yielding a 7.2-point gain. This is likely because $\mathrm{FW}_{\mathrm{soft}}$ optimizes layer-wise coefficients during merging, reducing the impact of layer-wise selection. + +$\mathrm{FW}_{\mathrm{soft}}$ excels when merging a large number of models, + +![](images/ae19002c886a8aea4bfcd1afe05fbcf7f6e685988213e5ef1f9be3da3dff54f2.jpg) +(a) Initial soluiton + +![](images/35343851102dddc913b651adead6e9a37c74a06a25753f212018c0ed25712f9b.jpg) +(b) Merging functions + +Figure 4. Ablation on FW-Merging. (a) reports accuracies on the vision benchmark, while (b) on vision and language benchmarks. + +outperforming $\mathrm{FW}_{\mathrm{hard}}$ by up to 6.7 points. Its ability to select multiple optimal directions per iteration allows it to navigate the parameter space efficiently. + +Optimizing merging coefficients $\lambda$ further improves performance by up to 9.9 points, underscoring the importance of weighting model parameters based on their relevance. + +Initial solution. We examine the effect of initialization on FW-Merging. An ideal initial solution should either (1) be closer to the global optimum or (2) expand the constraint set with a more meaningful search space. As shown in Figure 4a, initializing FW-Merging with the Adamerging result improves performance compared to starting from the pre-trained model, likely because Adamerging is closer to the optimal point. In contrast, task arithmetic leads to worse performance than the pre-trained model, potentially due to its poor performance on vision tasks $(21.2\%)$ , suggesting it starts further from the optimum. Consequently, more FW iterations are required to achieve convergence. + +Flexibility of merging functions. Although only a restricted set of merging functions ensure that FW-Merging remains within the convex hull, we demonstrate the flexibility of FW-Merging by showing its ability to enhance alternative merging functions. As shown in Figure 4b, applying FW-Merging with both Task Arithmetic and Ties-Merging improves performance on NLP and vision tasks, even though Ties-Merging does not necessarily stay within the convex hull. This suggests that FW-Merging remains effective across different merging functions. + +# 5. Conclusion + +In this work, we extend model merging to a more challenging setting where the merging pool consists of a large number of black-box fine-tuned checkpoints. While existing methods require prior knowledge of model details to achieve optimized performance, our proposed Frank-Wolfe Merging (FW-Merging) scales effectively with a large number of black-box models, iteratively refining the merged model towards the optimal point defined by an objective function. Experiments demonstrate that FW-Merging achieves superior performance and scalability, paving the way for next-generation model merging. + +# 6. Acknowledgement + +The support of the United Kingdom EPSRC (grant number UKRI256, EP/V028251/1, EP/N031768/1, EP/S030069/1, and EP/X036006/1), Intel, and AMD is gratefully acknowledged. + +# References + +[1] Samuel K Ainsworth, Jonathan Hayase, Siddhartha Srinivasan, Khaled Saab, and Stefano Fusi. Git re-basin: Merging models modulo permutation symmetries. arXiv preprint arXiv:2209.04836, 2022. 3 +[2] Rich Caruana. Multitask learning. Machine learning, 28: 41-75, 1997. 2 +[3] Sahil Chaudhary. Code alpaca: An instruction-following llama model for code generation. https://github.com/sahil280114/codealpaca, 2023.6 +[4] Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott Gray, Nick Ryder, Mikhail Pavlov, Alethea Power, Lukasz Kaiser, Mohammad Bavarian, Clemens Winter, Philippe Tillet, Felipe Petroski Such, Dave Cummings, Matthias Plappert, Fotios Chantzis, Elizabeth Barnes, Ariel Herbert-Voss, William Hebgen Guss, Alex Nichol, Alex Paino, Nikolas Tezak, Jie Tang, Igor Babuschkin, Suchir Balaji, Shantanu Jain, William Saunders, Christopher Hesse, Andrew N. Carr, Jan Leike, Josh Achiam, Vedant Misra, Evan Morikawa, Alec Radford, Matthew Knight, Miles Brundage, Mira Murati, Katie Mayer, Peter Welinder, Bob McGrew, Dario Amodei, Sam McCandlish, Ilya Sutskever, and Wojciech Zaremba. Evaluating large language models trained on code. 2021. 5 +[5] Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks. In International conference on machine learning, pages 794-803. PMLR, 2018. 3 +[6] Zhao Chen, Jiquan Ngiam, Yanping Huang, Thang Luong, Henrik Kretzschmar, Yuning Chai, and Dragomir Anguelov. Just pick a sign: Optimizing deep multitask models with gradient sign dropout. Advances in Neural Information Processing Systems, 33:2039-2050, 2020. 3 +[7] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3606-3613, 2014. 5, 6 +[8] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: Efficient finetuning of quantized llms. Advances in Neural Information Processing Systems, 36, 2024. 3 +[9] Marguerite Frank, Philip Wolfe, et al. An algorithm for quadratic programming. Naval research logistics quarterly, 3(1-2):95-110, 1956. 3 + +[10] Ian J Goodfellow, Oriol Vinyals, and Andrew M Saxe. Qualitatively characterizing neural network optimization problems. arXiv preprint arXiv:1412.6544, 2014. 3 +[11] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3 +[12] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 5 +[13] Hugging Face. Open LLM Leaderboard, 2024. Accessed: March 4, 2025. 1 +[14] Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Suchin Gururangan, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. Editing models with task arithmetic. arXiv preprint arXiv:2212.04089, 2022. 1 +[15] Pavel Izmailov, Dmitrii Podoprikhin, Timur Garipov, Dmitry Vetrov, and Andrew Gordon Wilson. Averaging weights leads to wider optima and better generalization. arXiv preprint arXiv:1803.05407, 2018. 1, 3 +[16] Junguang Jiang, Baixu Chen, Junwei Pan, Ximei Wang, Dapeng Liu, Jie Jiang, and Mingsheng Long. Forkmerge: Mitigating negative transfer in auxiliary-task learning. Advances in neural information processing systems, 36:30367-30389, 2023. 2 +[17] Qiao Jin, Bhuwan Dhingra, Zhengping Liu, William Cohen, and Xinghua Lu. Pubmedqa: A dataset for biomedical research question answering. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2567-2577, 2019. 5, 6 +[18] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013. 5, 6 +[19] Simon Lacoste-Julien, Martin Jaggi, Mark Schmidt, and Patrick Pletscher. Block-coordinate frank-wolfe optimization for structural svms. In International Conference on Machine Learning, pages 53-61. PMLR, 2013. 5 +[20] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 3 +[21] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 3 +[22] Yixiao Li, Yifan Yu, Chen Liang, Pengcheng He, Nikos Karampatziakis, Weizhu Chen, and Tuo Zhao. Loftq: Lorafine-tuning-aware quantization for large language models. arXiv preprint arXiv:2310.08659, 2023. 3 +[23] Fuzhao Liu, Xiaotian Hu, Chengyu Chen, Zhijian Wang, Shuxin Xiao, Zhen Wang, Zhongxiang Wang, Chuanxin Xie, Eric Xing, and Song Han. Dare: Diversity-aware model merging for sparse-to-dense mixture-of-experts. arXiv preprint arXiv:2402.10887, 2024. 3, 4, 5 + +[24] Liyang Liu, Yi Li, Zhanghui Kuang, J Xue, Yimin Chen, Wenming Yang, Qingmin Liao, and Wayne Zhang. Towards impartial multi-task learning. iclr, 2021. 3 +[25] Prateek Yadav Liu et al. Resolving interference when merging models. arXiv preprint arXiv:2306.01708, 2023. 3, 4 +[26] Shikun Liu, Edward Johns, and Andrew J Davison. End-to-end multi-task learning with attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1871-1880, 2019. 2 +[27] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettle-moyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 5 +[28] Zhenyi Lu, Chenghao Fan, Wei Wei, Xiaoye Qu, Dangyang Chen, and Yu Cheng. Twin-merging: Dynamic integration of modular expertise in model merging. Advances in Neural Information Processing Systems, 37:78905–78935, 2025. 5 +[29] Jiaqi Ma, Zhe Zhao, Xinyang Yi, Jilin Chen, Lichan Hong, and Ed H Chi. Modeling task relationships in multi-task learning with multi-gate mixture-of-experts. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1930-1939, 2018. 2 +[30] Jiaqi Ma, Zhe Zhao, Jilin Chen, Ang Li, Lichan Hong, and Ed H Chi. Snr: Sub-network routing for flexible parameter sharing in multi-task learning. In Proceedings of the AAAI conference on artificial intelligence, pages 216-223, 2019. 2 +[31] Sadhika Malladi, Tianyu Gao, Eshaan Nichani, Alex Damian, Jason D Lee, Danqi Chen, and Sanjeev Arora. Finetuning language models with just forward passes. Advances in Neural Information Processing Systems, 36:53038-53075, 2023. 3 +[32] Priyanka Mary Mammen. Federated learning: Opportunities and challenges. arXiv preprint arXiv:2101.05428, 2021. 1 +[33] Eric Mitchell, Kenton Lee, Alexander Khazatsky, Christopher D Manning, and Chelsea Finn. Editing models with task arithmetic. arXiv preprint arXiv:2212.04089, 2022. 3, 4, 5 +[34] Ramesh Nallapati, Bowen Zhou, Caglar Gulcehre, Bing Xiang, et al. Abstractive text summarization using sequence-to-sequence rnns and beyond. arXiv preprint arXiv:1602.06023, 2016.5, 6 +[35] Shraman Prakash, Rohit Girdhar, Devamanyu Hazarika, Joao Carreira, and Andrew Zisserman. Unified model for image, video, audio and language tasks. arXiv preprint arXiv:2307.16184, 2023. 3 +[36] Samyam Rajbhandari, Jeff Rasley, Olatunjri Ruwase, and Yuxiong He. Zero: Memory optimizations toward training trillion parameter models. In SC20: International Conference for High Performance Computing, Networking, Storage and Analysis, pages 1-16. IEEE, 2020. 1 +[37] Sidak Pal Singh and Martin Jaggi. Merging models with fisher-weighted averaging. arXiv preprint arXiv:2111.09832, 2021. 3 +[38] Johannes Stallkamp, Marc Schlipsing, Jan Salmen, and Christian Igel. The german traffic sign recognition benchmark: a multi-class classification competition. In *The 2011 + +international joint conference on neural networks, pages 1453-1460. IEEE, 2011. 5, 6 +[39] Ximeng Sun, Rameswar Panda, Rogerio Feris, and Kate Saenko. Adashare: Learning what to share for efficient deep multi-task learning. Advances in Neural Information Processing Systems, 33:8728-8740, 2020. 2 +[40] Anke Tang, Li Shen, Yong Luo, Liang Ding, Han Hu, Bo Du, and Dacheng Tao. Concrete subspace learning based interference elimination for multi-task model fusion. arXiv preprint arXiv:2312.06173, 2023. 3, 5 +[41] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 5 +[42] Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R Bowman. Glue: A multi-task benchmark and analysis platform for natural language understanding. arXiv preprint arXiv:1804.07461, 2018. 5, 6 +[43] Ke Wang, Nikolaos Dimitriadis, Guillermo Ortiz-Jimenez, François Fleuret, and Pascal Frossard. Localizing task information for improved model merging and compression. arXiv preprint arXiv:2405.07813, 2024. 5 +[44] Zirui Wang, Daniel Duckworth, Shashank Nag, Kshitij Patil, Jonathon Shlens, Ekin D Cubuk, Barret Zoph, Joseph Campbell, Mitchell Wortsman, and Google Brain. lo-fi: distributed fine-tuning without communication. arXiv preprint arXiv:2210.11948, 2022. 3 +[45] Yongxian Wei, Anke Tang, Li Shen, Feng Xiong, Chun Yuan, and Xiaochun Cao. Modeling multi-task model merging as adaptive projective gradient descent. arXiv preprint arXiv:2501.01230, 2025. 3 +[46] Mitchell Wortsman, Gabriel Ilharco, Samir Yitzhak Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. arXiv preprint arXiv:2203.05482, 2022. 3 +[47] Jianxiong Xiao, Krista A Ehinger, James Hays, Antonio Torralba, and Aude Oliva. Sun database: Exploring a large collection of scene categories. International Journal of Computer Vision, 119:3-22, 2016. 5, 6 +[48] Prateek Yadav, Derek Tam, Leshem Choshen, Colin A Raffel, and Mohit Bansal. Ties-merging: Resolving interference when merging models. Advances in Neural Information Processing Systems, 36:7093-7115, 2023. 1, 2, 5 +[49] Enneng Yang, Li Shen, Guibing Guo, Xingwei Wang, Xiaochun Cao, Jie Zhang, and Dacheng Tao. Model merging in llms, mllms, and beyond: Methods, theories, applications and opportunities, 2024. URL https://arxiv.org/abs/2408.07666, 2408.3 +[50] Enneng Yang, Zhenyi Wang, Li Shen, Shiwei Liu, Guibing Guo, Xingwei Wang, and Dacheng Tao. Adamerging: Adaptive model merging for multi-task learning. arXiv preprint arXiv:2310.02575, 2023. 1, 2, 3, 5 +[51] Enneng Yang, Li Shen, Zhenyi Wang, Guibing Guo, Xiaojun Chen, Xingwei Wang, and Dacheng Tao. Representa + +tion surgery for multi-task model merging. arXiv preprint arXiv:2402.02705, 2024. 1, 3, 5 +[52] Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How transferable are features in deep neural networks? Advances in neural information processing systems, 27, 2014. 5 +[53] Jun Yu, Yutong Dai, Xiaokang Liu, Jin Huang, Yishan Shen, Ke Zhang, Rong Zhou, Eashan Adhikarla, Wenxuan Ye, Yixin Liu, et al. Unleashing the power of multitask learning: A comprehensive survey spanning traditional, deep, and pretrained foundation model eras. arXiv preprint arXiv:2404.18961, 2024. 1 +[54] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. Advances in neural information processing systems, 33:5824-5836, 2020. 3 \ No newline at end of file diff --git a/2025/FW-Merging_ Scaling Model Merging with Frank-Wolfe Optimization/images.zip b/2025/FW-Merging_ Scaling Model Merging with Frank-Wolfe Optimization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..21bea3fa4935ac95a53aa6770ce1a00fc7c03224 --- /dev/null +++ b/2025/FW-Merging_ Scaling Model Merging with Frank-Wolfe Optimization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c2cfd7e56f028c32ec6c794496c1cc2c0aa3339bb4fc8a0d622d9c62d34627fe +size 528444 diff --git a/2025/FW-Merging_ Scaling Model Merging with Frank-Wolfe Optimization/layout.json b/2025/FW-Merging_ Scaling Model Merging with Frank-Wolfe Optimization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a02f0b6365be76d002e440549f05c2134cd9a8cb --- /dev/null +++ b/2025/FW-Merging_ Scaling Model Merging with Frank-Wolfe Optimization/layout.json @@ -0,0 +1,9645 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 88, + 103, + 522, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 103, + 522, + 121 + ], + "spans": [ + { + "bbox": [ + 88, + 103, + 522, + 121 + ], + "type": "text", + "content": "FW-Merging: Scaling Model Merging with Frank-Wolfe Optimization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 74, + 142, + 533, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 142, + 533, + 172 + ], + "spans": [ + { + "bbox": [ + 74, + 142, + 533, + 172 + ], + "type": "text", + "content": "Hao (Mark) Chen1 Shell Xu Hu2 Wayne Luk1 Timothy Hospedales2 Hongxiang Fan1 \n1Imperial College London, UK 2Samsung AI Center, Cambridge, UK" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 72, + 174, + 535, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 174, + 535, + 199 + ], + "spans": [ + { + "bbox": [ + 72, + 174, + 535, + 199 + ], + "type": "text", + "content": "hcl620@ic.ac.uk shell.hu@samsung.com w.luk@imperial.ac.uk t.hospedales@samsung.com hongxiang.fan@imperial.ac.uk" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 209, + 200, + 222 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 209, + 200, + 222 + ], + "spans": [ + { + "bbox": [ + 151, + 209, + 200, + 222 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 234, + 297, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 234, + 297, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 234, + 297, + 629 + ], + "type": "text", + "content": "Model merging has emerged as a promising approach for multi-task learning (MTL), offering a data-efficient alternative to conventional fine-tuning. However, with the rapid development of the open-source AI ecosystem and the increasing availability of fine-tuned foundation models, existing model merging methods face two key limitations: (i) They are primarily designed for in-house fine-tuned models, making them less adaptable to diverse model sources with partially unknown model and task information, (ii) They struggle to scale effectively when merging numerous model checkpoints. To address these challenges, we formulate model merging as a constrained optimization problem and introduce a novel approach: Frank-Wolfe Merging (FW-Merging). Inspired by Frank-Wolfe optimization, our approach iteratively selects the most relevant model in the pool to minimize a linear approximation of the objective function and then executes a local merging similar to the Frank-Wolfe update. The objective function is designed to capture the desired behavior of the target-merged model, while the fine-tuned candidate models define the constraint set. More importantly, FW-Merging serves as an orthogonal technique for existing merging methods, seamlessly integrating with them to further enhance accuracy performance. Our experiments show that FW-Merging scales across diverse model sources, remaining stable with 16 irrelevant models and improving by " + }, + { + "bbox": [ + 55, + 234, + 297, + 629 + ], + "type": "inline_equation", + "content": "15.3\\%" + }, + { + "bbox": [ + 55, + 234, + 297, + 629 + ], + "type": "text", + "content": " with 16 relevant models on 20 CV tasks, while maintaining constant memory overhead—unlike the linear overhead of data-informed merging methods. Compared with the state-of-the-art approaches, FW-Merging surpasses the data-free merging method by " + }, + { + "bbox": [ + 55, + 234, + 297, + 629 + ], + "type": "inline_equation", + "content": "32.8\\%" + }, + { + "bbox": [ + 55, + 234, + 297, + 629 + ], + "type": "text", + "content": " and outperforms the data-informed Adamerging by " + }, + { + "bbox": [ + 55, + 234, + 297, + 629 + ], + "type": "inline_equation", + "content": "8.39\\%" + }, + { + "bbox": [ + 55, + 234, + 297, + 629 + ], + "type": "text", + "content": " when merging 20 ViT models. Our code is open-sourced at here." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 651, + 135, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 651, + 135, + 664 + ], + "spans": [ + { + "bbox": [ + 56, + 651, + 135, + 664 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 671, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 295, + 696 + ], + "type": "text", + "content": "Multi-task learning (MTL)-based fine-tuning adapts a single pre-trained Large Language Model (LLM) for multiple" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 211, + 553, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 211, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 553, + 342 + ], + "type": "text", + "content": "downstream applications, reducing the deployment overhead of separately fine-tuning multiple models [53]. However, it still demands a large amount of high-quality data, which might only exist in the private domain [32], and significant compute resources [36]. To mitigate these issues, model merging has emerged as a promising technique for fusing fine-tuned models within the parameter space [14, 48, 50]. Existing model merging methods can be broadly classified into two categories: data-free methods [14, 15, 48], and data-informed methods [50, 51], which optimize merge coefficients based on additional data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 343, + 553, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 343, + 553, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 343, + 553, + 617 + ], + "type": "text", + "content": "While these approaches have proven effective, several key limitations hinder their scalability and broader adoption. First, these methods adjust merging coefficients based on the known capabilities of the models on specific tasks to optimize performance, which is less robust when dealing with diverse model sources with unknown information1. This is primarily caused by the inability to distinguish high-quality models from poorly fine-tuned ones in an unknown model setting. Second, when scaling up these approaches to a large number of unknown models, these methods struggle and can result in significant performance degradation. As shown in Figure 1a, our profiling study demonstrates a performance reduction ranging from " + }, + { + "bbox": [ + 313, + 343, + 553, + 617 + ], + "type": "inline_equation", + "content": "18.9\\%" + }, + { + "bbox": [ + 313, + 343, + 553, + 617 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 343, + 553, + 617 + ], + "type": "inline_equation", + "content": "64.4\\%" + }, + { + "bbox": [ + 313, + 343, + 553, + 617 + ], + "type": "text", + "content": ". These limitations are further amplified by the fast-growing open-source AI ecosystem, where platforms such as Hugging Face have driven a surge in the release of powerful LLMs with many lacking complete information. Given that merging open-source models has repeatedly shown the potential to produce top-ranking LLMs on major model leaderboards [13], developing scalable and robust merging techniques is essential to harness the growing number of open-source models, further enhancing performance and widening the adoption of model merging." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 618, + 553, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 553, + 666 + ], + "type": "text", + "content": "To effectively scale model merging and leverage the vast collection of open-sourced models with unknown capabilities, the new model merging method must exhibit two fundamental scaling properties: 1) as more irrelevant models" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 313, + 674, + 553, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 674, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 674, + 553, + 713 + ], + "type": "text", + "content": "1This paper refers unknown information to: 1) when open-source models are partially assessed on limited benchmarks, leaving their performance on other tasks unknown and costly to evaluate, and 2) when models contain misleading information, polluting the merging process." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 66, + 702, + 257, + 713 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 257, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 257, + 713 + ], + "type": "text", + "content": "*Corresponding Authors: Shell Xu Hu and Hongxiang Fan" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3390" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 72, + 219, + 190 + ], + "blocks": [ + { + "bbox": [ + 60, + 72, + 219, + 190 + ], + "lines": [ + { + "bbox": [ + 60, + 72, + 219, + 190 + ], + "spans": [ + { + "bbox": [ + 60, + 72, + 219, + 190 + ], + "type": "image", + "image_path": "3da232e398f85080f32a85987872d1e92462d5cc7f196443b585abe00fb321bb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 65, + 196, + 212, + 205 + ], + "lines": [ + { + "bbox": [ + 65, + 196, + 212, + 205 + ], + "spans": [ + { + "bbox": [ + 65, + 196, + 212, + 205 + ], + "type": "text", + "content": "(a) 4 CV tasks with irrelevant model addition." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 225, + 72, + 386, + 190 + ], + "blocks": [ + { + "bbox": [ + 225, + 72, + 386, + 190 + ], + "lines": [ + { + "bbox": [ + 225, + 72, + 386, + 190 + ], + "spans": [ + { + "bbox": [ + 225, + 72, + 386, + 190 + ], + "type": "image", + "image_path": "e75d2e7c3bb354a233d93ddc1b22cb9fdf5316c7c4616b34d65c4b71078a47d7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 231, + 196, + 379, + 205 + ], + "lines": [ + { + "bbox": [ + 231, + 196, + 379, + 205 + ], + "spans": [ + { + "bbox": [ + 231, + 196, + 379, + 205 + ], + "type": "text", + "content": "(b) 20 CV tasks with relevant model addition." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 391, + 73, + 550, + 191 + ], + "blocks": [ + { + "bbox": [ + 391, + 73, + 550, + 191 + ], + "lines": [ + { + "bbox": [ + 391, + 73, + 550, + 191 + ], + "spans": [ + { + "bbox": [ + 391, + 73, + 550, + 191 + ], + "type": "image", + "image_path": "953a1664a8cf8cb344efc734e23da2d593146a26f1633a292b439cf709e91a12.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 404, + 196, + 537, + 205 + ], + "lines": [ + { + "bbox": [ + 404, + 196, + 537, + 205 + ], + "spans": [ + { + "bbox": [ + 404, + 196, + 537, + 205 + ], + "type": "text", + "content": "(c) NLP tasks with noisy model addition." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 216, + 555, + 250 + ], + "lines": [ + { + "bbox": [ + 55, + 216, + 555, + 250 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 555, + 250 + ], + "type": "text", + "content": "Figure 1. Performance scaling of FW-Merging across CV tasks. (a) demonstrate robustness to irrelevant models, while (b) show improved performance with relevant models. (c) analyzes performance degradation when incorporating a noisy model initialized from a different pre-trained checkpoint. Detailed results and experimental setup are discussed in Section 4.4." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 256, + 296, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 256, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 56, + 256, + 296, + 531 + ], + "type": "text", + "content": "are added to the merging pool, the performance should remain unaffected, and 2) as more relevant models are added to the merging pool, the performance should steadily increase, converging towards the optimal performance. To this end, we revisit model merging and formulate it as a constrained optimization problem, where the objective function dictates the desirable behavior of the final merged model, and fine-tuned checkpoints form the constraint set. Inspired by Frank-Wolfe optimization, we introduce Frank-Wolfe Merging (FW-Merging), an iterative algorithm designed to enhance merging efficiency while maintaining robustness at scale. FW-Merging comprises three principal stages in each iteration: (1) Relevance Evaluation: Instead of merging models arbitrarily, we obtain the linear approximation of the objective function using gradients of the current model, revealing the most beneficial direction for improvement. (2) Model Selection: The most relevant checkpoints are selected from the constraint set by minimizing the linear approximation, ensuring that each step incorporates task-specific knowledge with minimal interference. (3) Knowledge Integration: The selected checkpoint is integrated using an orthogonal merging method, striking a balance between adaptation and stability in the merged model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "text", + "content": "We demonstrate the effectiveness of FW-Merging with a diverse pool of fine-tuned checkpoints across various language and vision tasks, compared to both data-free and data-informed model merging methods as well as traditional MTL-based fine-tuning. As shown in Figure 1, FW-Merging satisfy our two fundamental scaling properties: accuracy performance does not drop when 16 irrelevant models are added (compared to a " + }, + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "inline_equation", + "content": "49\\%" + }, + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "text", + "content": " drop in task-arithmetic) and steadily improves by " + }, + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "inline_equation", + "content": "15.3\\%" + }, + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "text", + "content": " when 16 relevant models are included. Additionally, FW-Merging requires only constant memory overhead, as it selects and merges a fixed number of models at a time. In contrast, methods that optimize merging coefficients [50] or resolve parameter interference [48] must store all models in memory, leading to linear overhead. Moreover, FW-Merging" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 256, + 555, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 256, + 555, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 256, + 555, + 423 + ], + "type": "text", + "content": "exhibits greater robustness to noisy models lacking critical information, such as their initialization point. As shown in Figure 1c, FW-Merging experiences minimal performance degradation when a misinitialized model is introduced, whereas Ties suffers a performance drop of up to " + }, + { + "bbox": [ + 313, + 256, + 555, + 423 + ], + "type": "inline_equation", + "content": "3.2\\%" + }, + { + "bbox": [ + 313, + 256, + 555, + 423 + ], + "type": "text", + "content": ". FW-Merging outperforms state-of-the-art data-free merging method by " + }, + { + "bbox": [ + 313, + 256, + 555, + 423 + ], + "type": "inline_equation", + "content": "32.8\\%" + }, + { + "bbox": [ + 313, + 256, + 555, + 423 + ], + "type": "text", + "content": " and the data-informed method Adamerging by " + }, + { + "bbox": [ + 313, + 256, + 555, + 423 + ], + "type": "inline_equation", + "content": "8.39\\%" + }, + { + "bbox": [ + 313, + 256, + 555, + 423 + ], + "type": "text", + "content": " when merging 20 ViT models. On the language benchmarks, FW-Merging achieves " + }, + { + "bbox": [ + 313, + 256, + 555, + 423 + ], + "type": "inline_equation", + "content": "6.3\\%" + }, + { + "bbox": [ + 313, + 256, + 555, + 423 + ], + "type": "text", + "content": " improvement over the best model merging method across discriminative and generative tasks, while even surpassing the performance of traditional MTL using only " + }, + { + "bbox": [ + 313, + 256, + 555, + 423 + ], + "type": "inline_equation", + "content": "3.4\\%" + }, + { + "bbox": [ + 313, + 256, + 555, + 423 + ], + "type": "text", + "content": " of its required data. Our results position FW-Merging as an effective solution to scale model merging to the next level." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 326, + 425, + 525, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 425, + 525, + 436 + ], + "spans": [ + { + "bbox": [ + 326, + 425, + 525, + 436 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 437, + 553, + 581 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 314, + 437, + 553, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 437, + 553, + 474 + ], + "spans": [ + { + "bbox": [ + 314, + 437, + 553, + 474 + ], + "type": "text", + "content": "- Identify scalability and robustness issues in existing model merging techniques through experiments, highlighting the urgent need for large-scale model merging." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 474, + 553, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 474, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 314, + 474, + 553, + 509 + ], + "type": "text", + "content": "- Formulate model merging as a constrained optimization problem with an objective function that explicitly captures the desired behavior of the final merged model." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 509, + 553, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 509, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 314, + 509, + 553, + 557 + ], + "type": "text", + "content": "- Introduce Frank-Wolfe Merging, a novel iterative method that autonomously guides the merged model toward an optimized direction, even with large sets of black-box open-source checkpoints." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 557, + 553, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 557, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 314, + 557, + 553, + 581 + ], + "type": "text", + "content": "- Evaluate our proposed approach on extensive benchmarks, demonstrating its effectiveness and scalability." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 596, + 400, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 596, + 400, + 609 + ], + "spans": [ + { + "bbox": [ + 314, + 596, + 400, + 609 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "content": "Efficient Multi-Task Learning. In traditional Multi-Task Learning (MTL), a single model is trained on a dataset containing multiple tasks to enable the model to acquire diverse capabilities [2]. However, a significant challenge in traditional MTL is the issue of negative transfer [16]. To mitigate this, architecture-based approaches have been developed, such as parameter sparsification [26, 39] and shared structure modularization [29, 30]. On the optimization side," + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3391" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "content": "methods to resolve gradient conflicts [6, 54] and domination of gradient or learning rate [5, 24] have been proposed. With the rise of Large Language Models (LLMs), MTL faces additional challenges, particularly the high computational costs. To address these challenges, strategies like parameter-efficient fine-tuning [11, 20, 21] and memory-efficient fine-tuning [8, 22, 31] have been introduced to minimize both memory and computational resource usage. More recently, model merging has emerged as a promising approach to make MTL more compute- and data-efficient." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 207, + 296, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 207, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 207, + 296, + 568 + ], + "type": "text", + "content": "Model Merging. While pre-merging methods prepare favorable conditions for merging, during-merging techniques combine multiple neural networks into a single model while retaining or enhancing their capabilities [49]. In this work, we focus on during-merging methods. Early insights into neural network landscapes [10] revealed that linear interpolation between models exposes useful loss surface properties, laying the foundation for weight averaging—a core merging technique. Simple averaging widens optima and improves generalization [15], evolving into advanced methods like model soups [46] and heterogeneous model merging. Recent advances introduce more structured approaches, such as Fisher-Weighted Averaging [37], which incorporates Fisher information to weight parameters more effectively, and Permutation Alignment methods like Git Re-Basin [1], which address weight permutation symmetries. Interference Resolution techniques, including TIES [25] and DOGE [45], mitigate parameter conflicts either through explicit alignment or projective gradient descent. Task Arithmetic [33] enables weight-space operations to combine task-specific behaviors in language models, while Diversity-Aware Merging, such as DARE [23], leverages model diversity to improve sparse-to-dense integration. In contrast to the data-free methods mentioned above, data-informed methods [40, 50, 51] optimize merging coefficients using additional data. Model merging is impactful for LLMs, enabling efficient knowledge integration without full retraining, facilitating distributed finetuning [44], multi-task learning [35], and cost-effective model adaptation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 578, + 111, + 591 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 111, + 591 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 111, + 591 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 598, + 251, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 598, + 251, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 598, + 251, + 612 + ], + "type": "text", + "content": "3.1. Preliminary: Frank-Wolfe algorithm" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 616, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 295, + 653 + ], + "type": "text", + "content": "The Frank-Wolfe (FW) algorithm [9], also known as the conditional gradient method, is an iterative optimization algorithm for constrained optimization problems of the form:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 154, + 662, + 295, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 662, + 295, + 681 + ], + "spans": [ + { + "bbox": [ + 154, + 662, + 295, + 681 + ], + "type": "interline_equation", + "content": "\\min _ {x \\in \\mathcal {C}} f (x) \\tag {1}", + "image_path": "aae980370ce2264dc3854e6ed301b1ee210d5c0a153d3d22fc2289cd9c1f1243.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " is a continuously differentiable function, and " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " is a compact convex set. The algorithm follows an elegant" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": "geometric intuition: at each iteration " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": ", FW first identifies which vertex of " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": " yields the steepest descent direction and then moves towards this vertex to decrease the value of the objective function. More specifically, FW algorithm:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 121, + 555, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 121, + 555, + 156 + ], + "spans": [ + { + "bbox": [ + 315, + 121, + 555, + 156 + ], + "type": "text", + "content": "1. Constructs a linear subproblem of the original optimization (a.k.a. linear minimization oracle) using first-order Taylor expansion at the point " + }, + { + "bbox": [ + 315, + 121, + 555, + 156 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 315, + 121, + 555, + 156 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 365, + 163, + 555, + 183 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 163, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 365, + 163, + 555, + 183 + ], + "type": "interline_equation", + "content": "\\operatorname {L M O} \\left(\\mathcal {C}, x _ {t}\\right) := \\underset {s \\in \\mathcal {C}} {\\arg \\min } \\langle s, \\nabla f \\left(x _ {t}\\right) \\rangle \\tag {2}", + "image_path": "10f27106d405334bcf3510380199bf957b64bc9e5fc3d2011e3eeb2c9cb44c34.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 188, + 553, + 248 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 313, + 188, + 553, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 188, + 553, + 212 + ], + "spans": [ + { + "bbox": [ + 313, + 188, + 553, + 212 + ], + "type": "text", + "content": "2. Finds the vertex " + }, + { + "bbox": [ + 313, + 188, + 553, + 212 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 313, + 188, + 553, + 212 + ], + "type": "text", + "content": " of the feasible set " + }, + { + "bbox": [ + 313, + 188, + 553, + 212 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 188, + 553, + 212 + ], + "type": "text", + "content": " by picking " + }, + { + "bbox": [ + 313, + 188, + 553, + 212 + ], + "type": "inline_equation", + "content": "s_t \\in \\mathrm{LMO}(\\mathcal{C}, x_t)" + }, + { + "bbox": [ + 313, + 188, + 553, + 212 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 213, + 553, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 553, + 248 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 553, + 248 + ], + "type": "text", + "content": "3. Takes a careful step from the current point " + }, + { + "bbox": [ + 313, + 213, + 553, + 248 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 313, + 213, + 553, + 248 + ], + "type": "text", + "content": " towards this direction " + }, + { + "bbox": [ + 313, + 213, + 553, + 248 + ], + "type": "inline_equation", + "content": "s_t - x_t" + }, + { + "bbox": [ + 313, + 213, + 553, + 248 + ], + "type": "text", + "content": ", maintaining feasibility through the convex combination: " + }, + { + "bbox": [ + 313, + 213, + 553, + 248 + ], + "type": "inline_equation", + "content": "x_{t + 1} = (1 - \\gamma_t)x_t + \\gamma_t s_t" + }, + { + "bbox": [ + 313, + 213, + 553, + 248 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 325, + 248, + 541, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 248, + 541, + 261 + ], + "spans": [ + { + "bbox": [ + 325, + 248, + 541, + 261 + ], + "type": "text", + "content": "The step size " + }, + { + "bbox": [ + 325, + 248, + 541, + 261 + ], + "type": "inline_equation", + "content": "\\gamma_t \\in [0,1]" + }, + { + "bbox": [ + 325, + 248, + 541, + 261 + ], + "type": "text", + "content": " can be chosen by line search" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 361, + 266, + 553, + 291 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 266, + 553, + 291 + ], + "spans": [ + { + "bbox": [ + 361, + 266, + 553, + 291 + ], + "type": "interline_equation", + "content": "\\gamma_ {t} = \\underset {\\gamma \\in [ 0, 1 ]} {\\arg \\min } f \\left((1 - \\gamma) x _ {t} + \\gamma s _ {t}\\right), \\tag {3}", + "image_path": "67167ad39100adb7811f4e8bab5888954819b33a1ff0ea9d3b0ccd52a0b992cf.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 298, + 553, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 298, + 553, + 310 + ], + "spans": [ + { + "bbox": [ + 313, + 298, + 553, + 310 + ], + "type": "text", + "content": "which ensures a sufficient decrease in " + }, + { + "bbox": [ + 313, + 298, + 553, + 310 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 313, + 298, + 553, + 310 + ], + "type": "text", + "content": " at each FW step." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 310, + 553, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 310, + 553, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 310, + 553, + 344 + ], + "type": "text", + "content": "To determine when to stop, the " + }, + { + "bbox": [ + 313, + 310, + 553, + 344 + ], + "type": "inline_equation", + "content": "FW" + }, + { + "bbox": [ + 313, + 310, + 553, + 344 + ], + "type": "text", + "content": " gap is used to measure the suboptimality in terms of the proximity to the best solution of LMO:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 369, + 352, + 553, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 352, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 369, + 352, + 553, + 369 + ], + "type": "interline_equation", + "content": "g _ {t} := \\max _ {s \\in \\mathcal {C}} \\langle - \\nabla f (x _ {t}), s _ {t} - x _ {t} \\rangle , \\tag {4}", + "image_path": "f3206fe9d2992695a85208308d116259de4a65891b73cc3cecd8dc8f67273ead.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 376, + 461, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 461, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 461, + 388 + ], + "type": "text", + "content": "which is non-negative by definition." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 394, + 473, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 394, + 473, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 473, + 407 + ], + "type": "text", + "content": "3.2. Frank-Wolfe Model Merging" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 411, + 554, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 554, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 554, + 483 + ], + "type": "text", + "content": "We consider the problem of fine-tuning a pre-trained foundation model on new tasks. Given a pre-trained model " + }, + { + "bbox": [ + 313, + 411, + 554, + 483 + ], + "type": "inline_equation", + "content": "\\theta_0" + }, + { + "bbox": [ + 313, + 411, + 554, + 483 + ], + "type": "text", + "content": " and previously fine-tuned models " + }, + { + "bbox": [ + 313, + 411, + 554, + 483 + ], + "type": "inline_equation", + "content": "\\{\\theta_1^*,\\theta_2^*,\\dots ,\\theta_n^*\\}" + }, + { + "bbox": [ + 313, + 411, + 554, + 483 + ], + "type": "text", + "content": ", we aim to fine-tune the " + }, + { + "bbox": [ + 313, + 411, + 554, + 483 + ], + "type": "inline_equation", + "content": "(n + 1)" + }, + { + "bbox": [ + 313, + 411, + 554, + 483 + ], + "type": "text", + "content": "-th model " + }, + { + "bbox": [ + 313, + 411, + 554, + 483 + ], + "type": "inline_equation", + "content": "\\theta_{n + 1}" + }, + { + "bbox": [ + 313, + 411, + 554, + 483 + ], + "type": "text", + "content": " on new tasks as a convex combination of the previous models with optimal merging coefficients." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 483, + 554, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 554, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 554, + 507 + ], + "type": "text", + "content": "To this end, we propose a Frank-Wolfe based model merging framework, which is described as follows." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 330, + 514, + 553, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 514, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 330, + 514, + 553, + 544 + ], + "type": "interline_equation", + "content": "\\min _ {\\lambda} \\ell \\left(\\sum_ {i = 1} ^ {n} \\lambda_ {i} \\theta_ {i} ^ {*}\\right) \\quad \\text {s . t .} \\quad \\sum_ {i = 1} ^ {n} \\lambda_ {i} = 1, \\quad \\lambda_ {i} \\geq 0, \\tag {5}", + "image_path": "b37aa1ae65e03a4b9fac080053c23068d93c26db80d05357c0f8441634de4f2b.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 551, + 554, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 551, + 554, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 551, + 554, + 598 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 551, + 554, + 598 + ], + "type": "inline_equation", + "content": "\\lambda = \\{\\lambda_1, \\dots, \\lambda_n\\}" + }, + { + "bbox": [ + 313, + 551, + 554, + 598 + ], + "type": "text", + "content": " are the merging coefficients, and " + }, + { + "bbox": [ + 313, + 551, + 554, + 598 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 313, + 551, + 554, + 598 + ], + "type": "text", + "content": " is a loss function formulated for a specific goal, such as aligning the merged and individual models [45] or satisfying a task objective [50]." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 598, + 554, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 598, + 554, + 657 + ], + "spans": [ + { + "bbox": [ + 313, + 598, + 554, + 657 + ], + "type": "text", + "content": "Potentially, a scaling issue of this formulation appears when the number of fine-tuned models " + }, + { + "bbox": [ + 313, + 598, + 554, + 657 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 598, + 554, + 657 + ], + "type": "text", + "content": " is large, since we need to keep all the fine-tuned models in memory. To address this, we propose a reformulation of the problem eq. (5) as follows:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 412, + 666, + 553, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 412, + 666, + 553, + 683 + ], + "spans": [ + { + "bbox": [ + 412, + 666, + 553, + 683 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta \\in \\mathcal {M}} \\ell (\\theta), \\tag {6}", + "image_path": "7ac8f5568f82dbc1c8fbb050f5518e755eed17e665745f31920be8326dacd0de.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{M} \\coloneqq \\mathrm{conv}(\\{\\theta_i^*\\}_{i=1}^n)" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": " is the convex hull of the set of previously fine-tuned models." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3392" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 170, + 204 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 170, + 204 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 170, + 204 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 170, + 204 + ], + "type": "image", + "image_path": "18f1e2940049aa7754c9381767ffdb58bccee810266c06bb4df81659d2a94216.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 213, + 555, + 258 + ], + "lines": [ + { + "bbox": [ + 54, + 213, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 54, + 213, + 555, + 258 + ], + "type": "text", + "content": "Figure 2. Illustration of model merging methods. " + }, + { + "bbox": [ + 54, + 213, + 555, + 258 + ], + "type": "inline_equation", + "content": "\\Theta_{A}" + }, + { + "bbox": [ + 54, + 213, + 555, + 258 + ], + "type": "text", + "content": " is an irrelevant model, while " + }, + { + "bbox": [ + 54, + 213, + 555, + 258 + ], + "type": "inline_equation", + "content": "\\Theta_{B}" + }, + { + "bbox": [ + 54, + 213, + 555, + 258 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 213, + 555, + 258 + ], + "type": "inline_equation", + "content": "\\Theta_{C}" + }, + { + "bbox": [ + 54, + 213, + 555, + 258 + ], + "type": "text", + "content": " are relevant models. Darker regions indicate higher objective function loss. Task Arithmetic treats all task vectors equally, failing to move optimally. Adamerging assigns different coefficients, moving towards more desirable direction but suffer from slow convergence due to interference from " + }, + { + "bbox": [ + 54, + 213, + 555, + 258 + ], + "type": "inline_equation", + "content": "\\Theta_{a}A" + }, + { + "bbox": [ + 54, + 213, + 555, + 258 + ], + "type": "text", + "content": ". FW-Merging iteratively selects the most relevant model to merge and adapts step sizes, efficiently reaching the optimum after " + }, + { + "bbox": [ + 54, + 213, + 555, + 258 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 54, + 213, + 555, + 258 + ], + "type": "text", + "content": " iterations." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 184, + 70, + 298, + 203 + ], + "blocks": [ + { + "bbox": [ + 184, + 70, + 298, + 203 + ], + "lines": [ + { + "bbox": [ + 184, + 70, + 298, + 203 + ], + "spans": [ + { + "bbox": [ + 184, + 70, + 298, + 203 + ], + "type": "image", + "image_path": "78d1279bd2ea75d76c428caa5197a219cd2617832cc9b90ba2ce66874201f725.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 311, + 70, + 425, + 204 + ], + "blocks": [ + { + "bbox": [ + 311, + 70, + 425, + 204 + ], + "lines": [ + { + "bbox": [ + 311, + 70, + 425, + 204 + ], + "spans": [ + { + "bbox": [ + 311, + 70, + 425, + 204 + ], + "type": "image", + "image_path": "2c48118c9254e59a646f332c633a407a89b5952a889e64411c5ac0f0044437cf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 434, + 70, + 554, + 204 + ], + "blocks": [ + { + "bbox": [ + 434, + 70, + 554, + 204 + ], + "lines": [ + { + "bbox": [ + 434, + 70, + 554, + 204 + ], + "spans": [ + { + "bbox": [ + 434, + 70, + 554, + 204 + ], + "type": "image", + "image_path": "5fa98ac24c3c56b40d8c7c6d370663f6d3fcc490c0150901bb21ab21df69c251.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 265, + 294, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 265, + 294, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 294, + 289 + ], + "type": "text", + "content": "Proposition 1. The optimization problems in equations (5) and (6) are equivalent." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 296, + 295, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 296, + 295, + 345 + ], + "spans": [ + { + "bbox": [ + 55, + 296, + 295, + 345 + ], + "type": "text", + "content": "Proof. By definition of convex hull, any point " + }, + { + "bbox": [ + 55, + 296, + 295, + 345 + ], + "type": "inline_equation", + "content": "\\theta \\in \\mathcal{M}" + }, + { + "bbox": [ + 55, + 296, + 295, + 345 + ], + "type": "text", + "content": " can be written as a convex combination of the vertices " + }, + { + "bbox": [ + 55, + 296, + 295, + 345 + ], + "type": "inline_equation", + "content": "\\{\\theta_i^*\\}_{i=1}^n" + }, + { + "bbox": [ + 55, + 296, + 295, + 345 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 55, + 296, + 295, + 345 + ], + "type": "inline_equation", + "content": "\\theta = \\sum_{i=1}^{n} \\lambda_i \\theta_i^*" + }, + { + "bbox": [ + 55, + 296, + 295, + 345 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 296, + 295, + 345 + ], + "type": "inline_equation", + "content": "\\lambda \\in \\Delta^n := \\{\\lambda \\in \\mathbb{R}^n \\mid \\sum_{i=1}^{n} \\lambda_i = 1, \\lambda_i \\geq 0\\}" + }, + { + "bbox": [ + 55, + 296, + 295, + 345 + ], + "type": "text", + "content": ". Therefore:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 108, + 352, + 241, + 383 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 352, + 241, + 383 + ], + "spans": [ + { + "bbox": [ + 108, + 352, + 241, + 383 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta \\in \\mathcal {M}} \\ell (\\theta) = \\min _ {\\lambda \\in \\Delta^ {n}} \\ell \\Bigl (\\sum_ {i = 1} ^ {n} \\lambda_ {i} \\theta_ {i} ^ {*} \\Bigr).", + "image_path": "8540ee3b40c737741a6ca0186e7e0a394209ad704a3dca6f6b597953795ea484.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 390, + 296, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 296, + 426 + ], + "type": "text", + "content": "This shows that any solution of one problem can be mapped to a solution of the other problem with the same objective value." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 434, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 296, + 506 + ], + "type": "text", + "content": "Since the FW algorithm requires the initial solution to be an interior point of the constraint set, we add the initial solution " + }, + { + "bbox": [ + 55, + 434, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\theta_0" + }, + { + "bbox": [ + 55, + 434, + 296, + 506 + ], + "type": "text", + "content": " to form a new constraint set " + }, + { + "bbox": [ + 55, + 434, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{M} \\coloneqq \\mathrm{conv}(\\mathcal{M} \\cup \\{\\theta_0\\})" + }, + { + "bbox": [ + 55, + 434, + 296, + 506 + ], + "type": "text", + "content": ", which we still denote as " + }, + { + "bbox": [ + 55, + 434, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 434, + 296, + 506 + ], + "type": "text", + "content": " for simple notations. A nice property of this reformulation is that the LMO can be simplified to" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 83, + 514, + 295, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 514, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 83, + 514, + 295, + 536 + ], + "type": "interline_equation", + "content": "\\operatorname {L M O} \\left(\\left\\{\\theta_ {i} ^ {*} \\right\\} _ {i = 1} ^ {n}, \\theta_ {t}\\right) = \\underset {s \\in \\left\\{\\theta_ {1} ^ {*}, \\dots , \\theta_ {n} ^ {*} \\right\\}} {\\arg \\min } \\langle \\nabla \\ell \\left(\\theta_ {t}\\right), s \\rangle \\tag {7}", + "image_path": "abc3d9d31802ba5e6d2b66c733145f51ba8e4aa94deddc3924ae588c2973fc77.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 543, + 296, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 543, + 296, + 591 + ], + "spans": [ + { + "bbox": [ + 55, + 543, + 296, + 591 + ], + "type": "text", + "content": "This is because for linear programming problems over convex sets, the optimal solution is always attained at the vertices of the constraint set. Algorithm 1 details the steps, and Figure 2 provides an overview." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 597, + 224, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 224, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 224, + 611 + ], + "type": "text", + "content": "3.3. Design choices of the algorithm" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 615, + 296, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 615, + 296, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 615, + 296, + 664 + ], + "type": "text", + "content": "The above algorithm illustrates the key ingredients of Frank-Wolfe merging: LMO, stopping criterion " + }, + { + "bbox": [ + 55, + 615, + 296, + 664 + ], + "type": "inline_equation", + "content": "g_{t}" + }, + { + "bbox": [ + 55, + 615, + 296, + 664 + ], + "type": "text", + "content": ", line search routine, and the merging function. We discuss in this section the design choices of these components." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Merging function The main deviation from the classical FW algorithm is the reinterpretation of the FW update " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "x_{t + 1} = (1 - \\gamma_t)x_t + \\gamma_t s_t" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " as a local merging between " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\theta_t" + } + ] + } + ], + "index": 14 + }, + { + "type": "code", + "bbox": [ + 322, + 281, + 535, + 449 + ], + "blocks": [ + { + "bbox": [ + 320, + 267, + 471, + 280 + ], + "lines": [ + { + "bbox": [ + 320, + 267, + 471, + 280 + ], + "spans": [ + { + "bbox": [ + 320, + 267, + 471, + 280 + ], + "type": "text", + "content": "Algorithm 1: Frank-Wolfe Merging:" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 322, + 281, + 535, + 449 + ], + "lines": [ + { + "bbox": [ + 322, + 281, + 535, + 449 + ], + "spans": [ + { + "bbox": [ + 322, + 281, + 535, + 449 + ], + "type": "text", + "content": "Input: Initial solution " + }, + { + "bbox": [ + 322, + 281, + 535, + 449 + ], + "type": "inline_equation", + "content": "\\theta_0" + }, + { + "bbox": [ + 322, + 281, + 535, + 449 + ], + "type": "text", + "content": "; Fine-tuned checkpoints " + }, + { + "bbox": [ + 322, + 281, + 535, + 449 + ], + "type": "inline_equation", + "content": "\\{\\theta_i^*\\}_{i=1}^n" + }, + { + "bbox": [ + 322, + 281, + 535, + 449 + ], + "type": "text", + "content": "; FW budget " + }, + { + "bbox": [ + 322, + 281, + 535, + 449 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 322, + 281, + 535, + 449 + ], + "type": "text", + "content": ". \nOutput: Merged model " + }, + { + "bbox": [ + 322, + 281, + 535, + 449 + ], + "type": "inline_equation", + "content": "\\theta_{merged}^*" + }, + { + "bbox": [ + 322, + 281, + 535, + 449 + ], + "type": "text", + "content": ". \n1: if " + }, + { + "bbox": [ + 322, + 281, + 535, + 449 + ], + "type": "inline_equation", + "content": "\\theta_0 \\notin \\mathcal{M}" + }, + { + "bbox": [ + 322, + 281, + 535, + 449 + ], + "type": "text", + "content": " then " + }, + { + "bbox": [ + 322, + 281, + 535, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{M} := \\mathcal{M} \\cup \\{\\theta_0\\}" + }, + { + "bbox": [ + 322, + 281, + 535, + 449 + ], + "type": "text", + "content": " \n2: for " + }, + { + "bbox": [ + 322, + 281, + 535, + 449 + ], + "type": "inline_equation", + "content": "t = 0 \\dots T" + }, + { + "bbox": [ + 322, + 281, + 535, + 449 + ], + "type": "text", + "content": " do \n3: Let " + }, + { + "bbox": [ + 322, + 281, + 535, + 449 + ], + "type": "inline_equation", + "content": "s_t := \\mathrm{LMO}(\\theta_t)" + }, + { + "bbox": [ + 322, + 281, + 535, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 322, + 281, + 535, + 449 + ], + "type": "inline_equation", + "content": "d_t := s_t - \\theta_t" + }, + { + "bbox": [ + 322, + 281, + 535, + 449 + ], + "type": "text", + "content": " \n4: if " + }, + { + "bbox": [ + 322, + 281, + 535, + 449 + ], + "type": "inline_equation", + "content": "g_t := \\langle -\\nabla \\ell(\\theta_t), d_t \\rangle \\leq \\epsilon" + }, + { + "bbox": [ + 322, + 281, + 535, + 449 + ], + "type": "text", + "content": " then \n5: return " + }, + { + "bbox": [ + 322, + 281, + 535, + 449 + ], + "type": "inline_equation", + "content": "\\theta_{merged}^* \\gets \\theta_t" + }, + { + "bbox": [ + 322, + 281, + 535, + 449 + ], + "type": "text", + "content": " \n6: end if \n7: Line-search: " + }, + { + "bbox": [ + 322, + 281, + 535, + 449 + ], + "type": "inline_equation", + "content": "\\gamma_t \\in \\arg \\min_{\\gamma \\in [0,1]} \\ell(\\theta_t + \\gamma d_t)" + }, + { + "bbox": [ + 322, + 281, + 535, + 449 + ], + "type": "text", + "content": " \n8: Update: " + }, + { + "bbox": [ + 322, + 281, + 535, + 449 + ], + "type": "inline_equation", + "content": "\\theta_{t+1} := \\mathrm{MergeFn}(\\theta_t, s_t, \\gamma_t)" + }, + { + "bbox": [ + 322, + 281, + 535, + 449 + ], + "type": "text", + "content": " \n9: end for \n10: return " + }, + { + "bbox": [ + 322, + 281, + 535, + 449 + ], + "type": "inline_equation", + "content": "\\theta_{merged}^* \\gets \\theta_T" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "code_body" + } + ], + "index": 16, + "sub_type": "algorithm" + }, + { + "bbox": [ + 313, + 475, + 555, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 555, + 522 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 555, + 522 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 313, + 475, + 555, + 522 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 313, + 475, + 555, + 522 + ], + "type": "text", + "content": ". We denote by MergeFn the customizable merging function as long as the merged model stays in the convex hull " + }, + { + "bbox": [ + 313, + 475, + 555, + 522 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 475, + 555, + 522 + ], + "type": "text", + "content": ". The most straightforward merging function is the convex combination:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 347, + 534, + 555, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 534, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 347, + 534, + 555, + 548 + ], + "type": "interline_equation", + "content": "\\operatorname {M e r g e F n} \\left(\\theta_ {t}, s _ {t}, \\gamma_ {t}\\right) := \\left(1 - \\gamma_ {t}\\right) \\theta_ {t} + \\gamma_ {t} s _ {t}, \\tag {8}", + "image_path": "df9dbba0edc8336971a20f5ab6dfd87c5512f1790519c6680941e0de26790407.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 557, + 554, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 554, + 581 + ], + "type": "text", + "content": "which corresponds to the Task-Arithmetic [33] method. The step size " + }, + { + "bbox": [ + 313, + 557, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\gamma_{t}" + }, + { + "bbox": [ + 313, + 557, + 554, + 581 + ], + "type": "text", + "content": " makes sure the merged model stays in " + }, + { + "bbox": [ + 313, + 557, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 557, + 554, + 581 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "It is natural to ask whether other existing model merging methods, such as TIES-Merging [25] and DARE [23] could also be used as MergeFn. The problem with these sophisticated merging methods is that the merged model might leave the constraint set, and thus violate the assumption of maintaining feasibility required by the classical FW theory. We verified in practice that these less rigorous merging functions might achieve better performance in certain cases but they generally cause more stability issues. Therefore, we do not consider these merging functions from the current comparison." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3393" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": "Hard FW v.s. Soft FW In the case of deep learning, the optimization problem is non-convex, additional efforts are needed to better characterize the loss landscape and prevent the LMO from being dominated by one or a few fine-tuned models, which occurs because the linear approximation of " + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "inline_equation", + "content": "\\ell(\\theta)" + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": " is an inaccurate sketch of the original objective function. Instead of relying on the argmin of linear subproblems, we fetch the top-" + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": " vertices of LMO, " + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "inline_equation", + "content": "\\{\\tilde{s}_j\\}_{j=1}^k" + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": ". A more subtle top-" + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": " operation can be performed in a task-wise fashion if the original objective function involves multi-tasks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 192, + 294, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 192, + 294, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 294, + 266 + ], + "type": "text", + "content": "Given the top-" + }, + { + "bbox": [ + 55, + 192, + 294, + 266 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 192, + 294, + 266 + ], + "type": "text", + "content": " vertices, we now go back to eq. (5) to obtain the optimal merging coefficients " + }, + { + "bbox": [ + 55, + 192, + 294, + 266 + ], + "type": "inline_equation", + "content": "\\{\\lambda_j^*\\}_{j=1}^k" + }, + { + "bbox": [ + 55, + 192, + 294, + 266 + ], + "type": "text", + "content": ". Note that this inner optimization is a reduced version of original eq. (5) because hosting " + }, + { + "bbox": [ + 55, + 192, + 294, + 266 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 192, + 294, + 266 + ], + "type": "text", + "content": " models in memory would not be a problem. We also remove the line search step as this gives a new merging function of the form" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 275, + 294, + 321 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 294, + 321 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 294, + 321 + ], + "type": "interline_equation", + "content": "\\operatorname {M e r g e F n} \\left(\\theta_ {t}, \\left\\{\\tilde {s} _ {j} \\right\\} _ {j = 1} ^ {k}, \\left\\{\\lambda_ {j} ^ {*} \\right\\} _ {j = 1} ^ {k}\\right) := \\theta_ {t} + \\sum_ {j = 1} ^ {k} \\lambda_ {j} ^ {*} \\left(\\tilde {s} _ {j} - \\theta_ {t}\\right). \\tag {9}", + "image_path": "9d1b7accd3ffc37f756117bc4859d607edefb53c63216295b86571e612c93f0a.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 331, + 294, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 294, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 294, + 354 + ], + "type": "text", + "content": "We call this oracle soft LMO in comparison to the argmin version which we call hard LMO." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 371, + 296, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 296, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 296, + 496 + ], + "type": "text", + "content": "Task-wise LMO v.s. layer-wise LMO The naive implementation of FW-Merging would vectorize the whole model weights " + }, + { + "bbox": [ + 55, + 371, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 371, + 296, + 496 + ], + "type": "text", + "content": " and then solve LMO. We call this taskwise LMO. Since different layers contribute differently to model performance [52], a layer-wise LMO may yield better model merging. To incorporate this, the constraint set is redefined as a Cartesian product of convex hulls for each layer: " + }, + { + "bbox": [ + 55, + 371, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\mathcal{M} := \\mathcal{M}_1 \\times \\dots \\times \\mathcal{M}_L" + }, + { + "bbox": [ + 55, + 371, + 296, + 496 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 371, + 296, + 496 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 371, + 296, + 496 + ], + "type": "text", + "content": " is the number of layers and " + }, + { + "bbox": [ + 55, + 371, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_l := \\mathrm{conv}\\left(\\left\\{\\theta_i^{*,l}\\right\\}_{i=1}^n\\right)" + }, + { + "bbox": [ + 55, + 371, + 296, + 496 + ], + "type": "text", + "content": ". The LMO is then conducted layer-wise:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 63, + 503, + 294, + 541 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 503, + 294, + 541 + ], + "spans": [ + { + "bbox": [ + 63, + 503, + 294, + 541 + ], + "type": "interline_equation", + "content": "\\operatorname {L M O} \\left(\\left\\{\\theta_ {i} ^ {* , l} \\right\\} _ {i = 1} ^ {n}, D, \\theta_ {t} ^ {l}\\right) = \\underset {s ^ {l} \\in \\left\\{\\theta_ {1} ^ {* , l}, \\dots , \\theta_ {n} ^ {* , l} \\right\\}} {\\arg \\min } \\left\\langle \\nabla \\ell \\left(\\theta^ {t}\\right) ^ {l}, s ^ {l} \\right\\rangle . \\tag {10}", + "image_path": "de29b75be5ca4b9c469c467fc190e2e71ec00f88ac098e5ba70ce93c1e6aa4df.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 551, + 294, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 294, + 588 + ], + "type": "text", + "content": "This version can be viewed as a block-coordinate Frank-Wolfe algorithm [19], which is applied when the problem has a natural decomposition into blocks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 599, + 137, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 137, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 137, + 613 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 619, + 163, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 619, + 163, + 632 + ], + "spans": [ + { + "bbox": [ + 55, + 619, + 163, + 632 + ], + "type": "text", + "content": "4.1. Experiment Setup" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 636, + 294, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 636, + 294, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 636, + 294, + 685 + ], + "type": "text", + "content": "**Benchmarks.** Our primary objective is to evaluate the effectiveness of our method in scenarios where the number of models greatly exceeds the number of evaluation tasks, and each model's capabilities are unknown in advance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 72, + 553, + 323 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 314, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 553, + 167 + ], + "type": "text", + "content": "- Vision Tasks: Following the setting of TALL [43], we use 20 ViT-B/32 models, each fine-tuned on a different vision task. The number of models to be merged is intentionally set to be significantly larger than the number of evaluation tasks, allowing us to assess the scalability of model merging methods. The evaluation benchmarks consist of four tasks: SUN397 [47], Stanford Cars [18], GTSRB [38], and DTD [7]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 168, + 553, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 168, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 314, + 168, + 553, + 227 + ], + "type": "text", + "content": "- Language Discriminative Tasks: We prepare 8 RoBERTa checkpoints [27] fine-tuned on eight tasks from the GLUE benchmark, following the practice in [28]. The merged model is then evaluated on four tasks from the GLUE benchmark [42]: MNLI, QNLI, QQP, and RTE." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 228, + 553, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 228, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 314, + 228, + 553, + 323 + ], + "type": "text", + "content": "- Language Generative Tasks: We collect 16 LLaMA2-7B models [41] fine-tuned with LoRA [12] on various tasks from Hugging Face. These models have unknown and uncontrolled capabilities, making them equivalently black-box models. Our goal is to evaluate the robustness of model merging methods in this challenging setting. The evaluation benchmarks include CNN/DM summarization [34], PubMedQA [17], and HumanEval [4]." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 341, + 553, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 553, + 418 + ], + "type": "text", + "content": "Metrics. For vision tasks, we report the classification accuracy. Following [28], we report the average normalized score for language tasks to account for differences in task-specific score ranges to account for variations in task-specific score ranges. The normalized score is computed as " + }, + { + "bbox": [ + 313, + 341, + 553, + 418 + ], + "type": "inline_equation", + "content": "\\mathrm{Score}_{\\mathrm{normalized}} = \\frac{1}{T}\\sum_{t=1}^{T}\\frac{\\mathrm{Score}(f(\\pmb{\\theta}^*))}{\\mathrm{Score}(f_t(\\pmb{\\theta}_t))}" + }, + { + "bbox": [ + 313, + 341, + 553, + 418 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 433, + 555, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 555, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 555, + 588 + ], + "type": "text", + "content": "Baselines. We compare FW-Merging with both data-informed and data-free model merging methods. For data-informed model merging, we compare FW-Merging against Adamerging [50], Surgery [51], and Concrete Merging [40]. To ensure a fair comparison, these methods are trained only on the same tasks as FW-Merging. For data-free model merging, we compare FW-Merging with Fisher Merging, Weight Averaging, RegMean Merging, Task Arithmetic [33], Ties-Merging [48], and DARE Merging [23] across both language and vision tasks. Additionally, we fine-tune one model on the discriminative language benchmark and another on the generative language benchmark to serve as additional baselines." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "Implementations. We implement two FW-Merging variants: " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{FW}_{\\text {hard }}" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": ", which uses hard FW with layer-wise LMO, and " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{FW}_{\\text {soft }}" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": ", which employs soft FW with task-wise LMO (Section 3.3). For " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{FW}_{\\text {soft }}" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": ", layer-wise coefficients are optimized via gradient descent on the training dataset to solve eq. 5, differing from Adamerging [50] by minimizing cross-entropy loss on training data rather than entropy on test samples. On language benchmarks, the training dataset consists of 100 samples per task, and " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{FW}_{\\text {hard }}" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": " runs" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 294, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 294, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 294, + 715 + ], + "type": "text", + "content": "2For inner optimization, we use projected gradient descent with a projection of " + }, + { + "bbox": [ + 55, + 693, + 294, + 715 + ], + "type": "inline_equation", + "content": "\\{\\lambda_j\\}_{j = 1}^k" + }, + { + "bbox": [ + 55, + 693, + 294, + 715 + ], + "type": "text", + "content": " onto the simplex after each gradient update." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3394" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 78, + 83, + 526, + 195 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 542, + 81 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 542, + 81 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 542, + 81 + ], + "type": "text", + "content": "Table 1. Performance on 4 Discriminative Tasks when merging 8 RoBERTa and 3 Generative Tasks when merging 16 LLaMA2-7B." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 78, + 83, + 526, + 195 + ], + "lines": [ + { + "bbox": [ + 78, + 83, + 526, + 195 + ], + "spans": [ + { + "bbox": [ + 78, + 83, + 526, + 195 + ], + "type": "table", + "html": "
Method4 Disc. Tasks (8 Models)3 Gen. Tasks (16 Models)Avg. Normalized Score
Pretrained49.677.163.4
Traditional MTL73.181.277.2
Task Arithmetic (w/ DARE)77.316.847.1
Ties-Merging (w/ DARE)75.646.661.1
Task Arithmetic80.875.978.4
Ties-Merging64.378.571.4
FWhard (Ours)85.481.183.1
", + "image_path": "6e9a39bcb8e1ae9ca1b914d1d63ad6b890def48ed03940924eb68b72e57070fb.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 58, + 211, + 293, + 364 + ], + "blocks": [ + { + "bbox": [ + 57, + 200, + 293, + 210 + ], + "lines": [ + { + "bbox": [ + 57, + 200, + 293, + 210 + ], + "spans": [ + { + "bbox": [ + 57, + 200, + 293, + 210 + ], + "type": "text", + "content": "Table 2. Performance on 4 CV Tasks when merging 20 ViT-B/32." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 211, + 293, + 364 + ], + "lines": [ + { + "bbox": [ + 58, + 211, + 293, + 364 + ], + "spans": [ + { + "bbox": [ + 58, + 211, + 293, + 364 + ], + "type": "table", + "html": "
MethodSUN397CarsGTSRBDTDAvg.
Pretrained62.359.732.643.849.6
DARE (TIES)5.92.316.711.89.2
Task Arithmetic20.412.229.822.321.2
Ties-Merging51.036.257.740.646.4
Weight Averaging64.259.643.146.553.4
Fisher Merging64.663.843.046.954.6
RegMean65.562.259.753.960.3
LW Concrete AM62.560.388.054.766.3
Adamerging66.470.195.164.073.9
Surgery69.771.896.673.477.9
FWhard (Ours)66.569.995.164.574.0
FWsoft (Ours)72.974.896.876.080.1
", + "image_path": "30b5700ec75b6f45d6eca874b5717dda76abb713225c2e2e2241cdaad0701793.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 371, + 295, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 295, + 468 + ], + "type": "text", + "content": "for 10 iterations, initialized with Task Arithmetic's merged model. For vision tasks, it runs for 3 iterations. " + }, + { + "bbox": [ + 55, + 371, + 295, + 468 + ], + "type": "inline_equation", + "content": "FW_{\\text{soft}}" + }, + { + "bbox": [ + 55, + 371, + 295, + 468 + ], + "type": "text", + "content": " runs for 15 iterations on vision benchmarks, initialized with the pre-trained model. Training datasets consist of 100 samples from MNLI, QNLI, QQP, and RTE [42] for discriminative tasks; CNN/DM [34], CodeAlpaca-20k [3], and PubMedQA [17] for generative tasks; and SUN397 [47], Stanford Cars [18], GTSRB [38], and DTD [7] for vision tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 475, + 279, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 279, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 279, + 489 + ], + "type": "text", + "content": "4.2. Comparison with Model Merging Methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 493, + 296, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 493, + 296, + 554 + ], + "spans": [ + { + "bbox": [ + 55, + 493, + 296, + 554 + ], + "type": "text", + "content": "We evaluate FW-Merging against both data-informed and data-free model merging approaches across language and vision benchmarks. Table 1 reports the results for language tasks, including both discriminative and generative settings, while Table 2 presents results on vision benchmarks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": "Language Tasks. " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathsf{FW}_{\\text{hard}}" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": " achieves the highest average normalized score across language benchmarks, consistently surpassing prior model merging baselines. Specifically, " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathsf{FW}_{\\text{hard}}" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": " improves upon Task Arithmetic by 4.6 points, Ties-Merging by 11.7, and DARE (Ties) by 9.8 on discriminative tasks. Table 4 shows that " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathsf{FW}_{\\text{hard}}" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": " also outperforms data-informed Adamerging by 5.9 points. For generative tasks, " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathsf{FW}_{\\text{hard}}" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": " outperforms Task Arithmetic by 5.2 points, Ties-Merging by 2.6, and DARE (Ties) by 34.5. Interestingly, while Task Arithmetic outperforms Ties-Merging on discriminative tasks by a margin of 16.5 points, it lags behind by 2.6 points on the more challenging generative tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 201, + 555, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 201, + 555, + 333 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 555, + 333 + ], + "type": "text", + "content": "This discrepancy likely arises from increased interference among task vectors as more checkpoints are merged. Unlike Ties-Merging, which explicitly resolves merging conflicts, Task Arithmetic lacks a reconciliation mechanism, making it more susceptible to such interference. In contrast, " + }, + { + "bbox": [ + 313, + 201, + 555, + 333 + ], + "type": "inline_equation", + "content": "\\mathsf{FW}_{\\text {hard }}" + }, + { + "bbox": [ + 313, + 201, + 555, + 333 + ], + "type": "text", + "content": " consistently outperforms both Ties-Merging and Task Arithmetic by selectively merging only the most relevant model parameters in each iteration. This targeted approach effectively mitigates interference, leading to more stable and robust performance across both discriminative and generative tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 350, + 553, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 553, + 482 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 553, + 482 + ], + "type": "text", + "content": "Vision Tasks. " + }, + { + "bbox": [ + 313, + 350, + 553, + 482 + ], + "type": "inline_equation", + "content": "\\mathsf{FW}_{\\mathrm{soft}}" + }, + { + "bbox": [ + 313, + 350, + 553, + 482 + ], + "type": "text", + "content": " achieves state-of-the-art performance across multiple vision benchmarks, surpassing data-informed methods like Adamerging and Surgery. As shown in Table 2, " + }, + { + "bbox": [ + 313, + 350, + 553, + 482 + ], + "type": "inline_equation", + "content": "\\mathsf{FW}_{\\mathrm{hard}}" + }, + { + "bbox": [ + 313, + 350, + 553, + 482 + ], + "type": "text", + "content": " surpasses Adamerging, Concrete Merging, and all data-free merging methods in overall performance. Additionally, " + }, + { + "bbox": [ + 313, + 350, + 553, + 482 + ], + "type": "inline_equation", + "content": "\\mathsf{FW}_{\\mathrm{soft}}" + }, + { + "bbox": [ + 313, + 350, + 553, + 482 + ], + "type": "text", + "content": " attains the highest accuracy (80.1%), outperforming Adamerging by 6.2% and Surgery by 2.2%. Unlike Surgery, which requires additional task-specific parameters and multiple forward passes per inference, our approach efficiently adapts to diverse visual tasks without increasing storage or inference complexity." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 483, + 555, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 555, + 638 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 555, + 638 + ], + "type": "text", + "content": "In general, data-free merging methods show significantly lower performance compared to data-informed approaches while merging a large number of models, when the models' capabilities do not precisely align with the evaluation tasks. This limitation arises because data-free methods treat all models equally, merging them without considering their unique capabilities, which amplifies interference between models. In contrast, data-informed merging methods achieve superior performance by optimizing merging coefficients on datasets as they allow for explicit control over desirable capabilities. FW-Merging, in particular, enhances scalability via hard model selection based on the linear approximation minimization." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 647, + 501, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 501, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 501, + 661 + ], + "type": "text", + "content": "4.3. Comparison with Traditional MTL" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "text", + "content": "We compare FW-Merging with models fine-tuned using traditional MTL on discriminative and generative tasks. In each case, one single model is fine-tuned across all tasks, with performance and computational cost reported" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "3395" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 115, + 83, + 492, + 233 + ], + "blocks": [ + { + "bbox": [ + 105, + 70, + 504, + 81 + ], + "lines": [ + { + "bbox": [ + 105, + 70, + 504, + 81 + ], + "spans": [ + { + "bbox": [ + 105, + 70, + 504, + 81 + ], + "type": "text", + "content": "Table 3. Merging Methods' Performance vs. Number of Models when Adding Relevant vs. Irrelevant Models." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 115, + 83, + 492, + 233 + ], + "lines": [ + { + "bbox": [ + 115, + 83, + 492, + 233 + ], + "spans": [ + { + "bbox": [ + 115, + 83, + 492, + 233 + ], + "type": "table", + "html": "
#Models4 CV Tasks\nWhen "Irrelevant" Models Added20 CV Tasks\nWhen "Relevant" Models Added
DARETaskTiesAMFWsoftDARETaskTiesAMFWsoft
473.670.365.775.274.157.359.260.259.659.2
664.164.164.474.673.954.058.461.664.063.2
848.557.163.673.474.048.257.263.964.764.5
1040.652.962.972.873.943.555.063.965.466.2
1231.747.961.971.474.139.152.363.765.167.5
1423.640.157.769.974.036.148.962.064.168.0
1617.933.352.568.674.132.345.359.762.768.3
1812.225.949.264.674.027.840.357.360.768.4
209.221.246.461.074.224.236.355.658.968.3
", + "image_path": "a9ea11b279109757e9c4edc4c98bd47662828c950b6ce506ea6ec9d449e036c0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 59, + 235, + 553, + 357 + ], + "blocks": [ + { + "bbox": [ + 59, + 235, + 553, + 357 + ], + "lines": [ + { + "bbox": [ + 59, + 235, + 553, + 357 + ], + "spans": [ + { + "bbox": [ + 59, + 235, + 553, + 357 + ], + "type": "image", + "image_path": "df2de3ebb3c8975898230d9a993b3f46bc15f236e32ff6dcedd62fd22858c3b4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 366, + 555, + 388 + ], + "lines": [ + { + "bbox": [ + 55, + 366, + 555, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 555, + 388 + ], + "type": "text", + "content": "Figure 3. Linear Approximation of the Objective Function of Model Checkpoints Across Different Tasks in a Frank-Wolfe Iteration. The x-axis represents the checkpoints, and each graph shows the linear approximation result for each task." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 404, + 293, + 464 + ], + "blocks": [ + { + "bbox": [ + 56, + 393, + 294, + 403 + ], + "lines": [ + { + "bbox": [ + 56, + 393, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 294, + 403 + ], + "type": "text", + "content": "Table 4. Costs and Perf. of methods on NLP discriminative tasks." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 404, + 293, + 464 + ], + "lines": [ + { + "bbox": [ + 58, + 404, + 293, + 464 + ], + "spans": [ + { + "bbox": [ + 58, + 404, + 293, + 464 + ], + "type": "table", + "html": "
MethodData Samples/TaskTime CostPerf.
Traditional MTL2.9K4.2h73.1
Data-free Merging0080.8
Data-informed Merging1.6K2min79.5
FW Merging1002min85.4
", + "image_path": "46371d62766429ea62f50e0ac32b2d296a83b675910a089dd4cdf913e08aabbd.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 472, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 296, + 555 + ], + "type": "text", + "content": "in Table 1 and Table 4. Traditional MTL achieves an average score of 77.2, lower than that of FW-Merging (83.1). On discriminative tasks, MTL scores 73.1, trailing FW-Merging (85.4). For generative tasks, MTL scores 81.2, while FW-Merging closely matches it at 80.8, suggesting that FW-Merging's performance matches that of traditional MTL." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 558, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 558, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 558, + 297, + 714 + ], + "type": "text", + "content": "As shown in Table 4, FW-Merging demonstrates a substantial advantage in efficiency. Traditional MTL requires fine-tuning on 2.9K samples per task and takes 4.2 hours of training time, which is computationally intensive. In contrast, FW-Merging only requires 100 training samples per task and completes the merging process in just 2 minutes. This huge reduction in computational cost underscores the effectiveness of FW-Merging compared to traditional MTL. Moreover, FW-Merging has a key advantage over traditional MTL: while MTL requires a large volume of high-quality data for optimal performance, FW-Merging needs only a small set of data because: " + }, + { + "bbox": [ + 54, + 558, + 297, + 714 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 54, + 558, + 297, + 714 + ], + "type": "text", + "content": ") it optimizes merging coefficients based on models' character" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 396, + 555, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 555, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 555, + 443 + ], + "type": "text", + "content": "istics, which simplifies the optimization space, and 2) it uses model weights as inputs, which are much more information-dense representations than data, enabling more efficient objective learning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 444, + 556, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 556, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 556, + 516 + ], + "type": "text", + "content": "FW-Merging is a post-training technique that does not require access to original training data, making it ideal for privacy-sensitive or data-scarce scenarios. Overall, the results suggest that FW-Merging is a scalable, efficient alternative to traditional MTL, providing comparable performance at a reduced computational cost." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 525, + 499, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 499, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 499, + 539 + ], + "type": "text", + "content": "4.4. Scaling to More Models and Tasks" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 556, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 556, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 312, + 556, + 555, + 689 + ], + "type": "text", + "content": "We investigate the performance scaling of different merging methods with the number of models, as shown in Figure 1a, Figure 1b, and Table 3. In large-scale model merging, models from open-source platforms vary in quality. To simulate this, we use 20 ViT-B/32 models fine-tuned on tasks that are either included in the evaluation benchmark or not. A model is irrelevant if its fine-tuning dataset does not match the training split of the evaluation task, and relevant if it matches. To ensure fair comparison, the total number of training iterations run by " + }, + { + "bbox": [ + 312, + 556, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\mathrm{FW}_{\\mathrm{soft}}" + }, + { + "bbox": [ + 312, + 556, + 555, + 689 + ], + "type": "text", + "content": " is the same as that of Adamreging." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "content": "As shown in Table 3, adding irrelevant models sharply reduces the performance of data-free methods: DARE by" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3396" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 82, + 276, + 232 + ], + "blocks": [ + { + "bbox": [ + 75, + 71, + 274, + 81 + ], + "lines": [ + { + "bbox": [ + 75, + 71, + 274, + 81 + ], + "spans": [ + { + "bbox": [ + 75, + 71, + 274, + 81 + ], + "type": "text", + "content": "Table 5. Ablation on design variants of FW-Merging." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 82, + 276, + 232 + ], + "lines": [ + { + "bbox": [ + 69, + 82, + 276, + 232 + ], + "spans": [ + { + "bbox": [ + 69, + 82, + 276, + 232 + ], + "type": "table", + "html": "
Coefficient λMethodLMOScore
Vision Tasks
OptimizedFWsoftLayer-wise79.7
OptimizedFWsoftTask-wise80.1
UnoptimizedFWsoftLayer-wise69.8
UnoptimizedFWsoftTask-wise70.3
-FWhardLayer-wise74.0
-FWhardTask-wise73.7
NLP Discriminative Tasks
-FWhardLayer-wise85.4
-FWhardTask-wise78.2
", + "image_path": "5fb679ee7d74e01196f6a15c7c4ded41938208856404d448ed7b56a00187e0be.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 239, + 296, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 296, + 443 + ], + "type": "inline_equation", + "content": "64.4\\%" + }, + { + "bbox": [ + 55, + 239, + 296, + 443 + ], + "type": "text", + "content": ", Task Arithmetic by " + }, + { + "bbox": [ + 55, + 239, + 296, + 443 + ], + "type": "inline_equation", + "content": "49.1\\%" + }, + { + "bbox": [ + 55, + 239, + 296, + 443 + ], + "type": "text", + "content": ", and Ties by " + }, + { + "bbox": [ + 55, + 239, + 296, + 443 + ], + "type": "inline_equation", + "content": "19.1\\%" + }, + { + "bbox": [ + 55, + 239, + 296, + 443 + ], + "type": "text", + "content": ", likely due to task interference and equal treatment of all models. Data-informed methods degrade less, with Adamerging dropping by " + }, + { + "bbox": [ + 55, + 239, + 296, + 443 + ], + "type": "inline_equation", + "content": "14.2\\%" + }, + { + "bbox": [ + 55, + 239, + 296, + 443 + ], + "type": "text", + "content": ". In contrast, " + }, + { + "bbox": [ + 55, + 239, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\mathsf{FW}_{\\mathrm{soft}}" + }, + { + "bbox": [ + 55, + 239, + 296, + 443 + ], + "type": "text", + "content": " remains highly stable, fluctuating only from " + }, + { + "bbox": [ + 55, + 239, + 296, + 443 + ], + "type": "inline_equation", + "content": "73.9\\%" + }, + { + "bbox": [ + 55, + 239, + 296, + 443 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 239, + 296, + 443 + ], + "type": "inline_equation", + "content": "74.1\\%" + }, + { + "bbox": [ + 55, + 239, + 296, + 443 + ], + "type": "text", + "content": " as more models are added. In Figure 3, we examine the linear approximation of different checkpoints for a specific task and find that the model fine-tuned on the task consistently yields the most negative linear approximation. This indicates that in the Frank-Wolfe update, the most relevant checkpoint is chosen as the direction for merging, allowing FW-Merging to iteratively improve the merged model in the optimized direction within the constraint set. The inner product between gradients and model parameters serves as a reliable indicator of model relevance, with minimized computational cost, further demonstrating FW-Merging's scalability even in the presence of irrelevant models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 445, + 296, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 296, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 296, + 577 + ], + "type": "text", + "content": "Adding relevant models should ideally improve performance, but data-free methods still degrade as shown in Table 3: DARE by " + }, + { + "bbox": [ + 55, + 445, + 296, + 577 + ], + "type": "inline_equation", + "content": "33.1\\%" + }, + { + "bbox": [ + 55, + 445, + 296, + 577 + ], + "type": "text", + "content": ", Task Arithmetic by " + }, + { + "bbox": [ + 55, + 445, + 296, + 577 + ], + "type": "inline_equation", + "content": "22.9\\%" + }, + { + "bbox": [ + 55, + 445, + 296, + 577 + ], + "type": "text", + "content": ", and Ties by " + }, + { + "bbox": [ + 55, + 445, + 296, + 577 + ], + "type": "inline_equation", + "content": "4.6\\%" + }, + { + "bbox": [ + 55, + 445, + 296, + 577 + ], + "type": "text", + "content": ", with Ties performing best by mitigating parameter conflicts. Data-informed methods like Adamerging fluctuate between " + }, + { + "bbox": [ + 55, + 445, + 296, + 577 + ], + "type": "inline_equation", + "content": "58.9\\%" + }, + { + "bbox": [ + 55, + 445, + 296, + 577 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 445, + 296, + 577 + ], + "type": "inline_equation", + "content": "64.7\\%" + }, + { + "bbox": [ + 55, + 445, + 296, + 577 + ], + "type": "text", + "content": " as merging complexity increases, whereas " + }, + { + "bbox": [ + 55, + 445, + 296, + 577 + ], + "type": "inline_equation", + "content": "\\mathsf{FW}_{\\mathrm{soft}}" + }, + { + "bbox": [ + 55, + 445, + 296, + 577 + ], + "type": "text", + "content": " steadily improves from " + }, + { + "bbox": [ + 55, + 445, + 296, + 577 + ], + "type": "inline_equation", + "content": "59.2\\%" + }, + { + "bbox": [ + 55, + 445, + 296, + 577 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 445, + 296, + 577 + ], + "type": "inline_equation", + "content": "68.3\\%" + }, + { + "bbox": [ + 55, + 445, + 296, + 577 + ], + "type": "text", + "content": " by iteratively selecting the most relevant models, facilitating smoother convergence. These results underscore FW-Merging's effectiveness as a scalable solution for large-scale model merging." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 586, + 156, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 156, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 156, + 597 + ], + "type": "text", + "content": "4.5. Ablation Studies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "type": "text", + "content": "Design variants. Table 5 compares the design variants of FW-Merging (Section 3.3). Task-wise LMO aligns better with " + }, + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\mathrm{FW}_{\\mathrm{soft}}" + }, + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "type": "text", + "content": ", improving performance slightly by 0.5 points over layer-wise LMO, while layer-wise LMO is more effective for " + }, + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\mathrm{FW}_{\\mathrm{hard}}" + }, + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "type": "text", + "content": ", especially on language tasks, yielding a 7.2-point gain. This is likely because " + }, + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\mathrm{FW}_{\\mathrm{soft}}" + }, + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "type": "text", + "content": " optimizes layer-wise coefficients during merging, reducing the impact of layer-wise selection." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{FW}_{\\mathrm{soft}}" + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": " excels when merging a large number of models," + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 429, + 144 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 429, + 144 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 429, + 144 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 429, + 144 + ], + "type": "image", + "image_path": "ae19002c886a8aea4bfcd1afe05fbcf7f6e685988213e5ef1f9be3da3dff54f2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 149, + 403, + 159 + ], + "lines": [ + { + "bbox": [ + 343, + 149, + 403, + 159 + ], + "spans": [ + { + "bbox": [ + 343, + 149, + 403, + 159 + ], + "type": "text", + "content": "(a) Initial soluiton" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 440, + 72, + 551, + 145 + ], + "blocks": [ + { + "bbox": [ + 440, + 72, + 551, + 145 + ], + "lines": [ + { + "bbox": [ + 440, + 72, + 551, + 145 + ], + "spans": [ + { + "bbox": [ + 440, + 72, + 551, + 145 + ], + "type": "image", + "image_path": "35343851102dddc913b651adead6e9a37c74a06a25753f212018c0ed25712f9b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 460, + 149, + 531, + 159 + ], + "lines": [ + { + "bbox": [ + 460, + 149, + 531, + 159 + ], + "spans": [ + { + "bbox": [ + 460, + 149, + 531, + 159 + ], + "type": "text", + "content": "(b) Merging functions" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 160, + 553, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 160, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 313, + 160, + 553, + 182 + ], + "type": "text", + "content": "Figure 4. Ablation on FW-Merging. (a) reports accuracies on the vision benchmark, while (b) on vision and language benchmarks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 186, + 553, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 186, + 553, + 222 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 553, + 222 + ], + "type": "text", + "content": "outperforming " + }, + { + "bbox": [ + 313, + 186, + 553, + 222 + ], + "type": "inline_equation", + "content": "\\mathrm{FW}_{\\mathrm{hard}}" + }, + { + "bbox": [ + 313, + 186, + 553, + 222 + ], + "type": "text", + "content": " by up to 6.7 points. Its ability to select multiple optimal directions per iteration allows it to navigate the parameter space efficiently." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 223, + 553, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 223, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 223, + 553, + 258 + ], + "type": "text", + "content": "Optimizing merging coefficients " + }, + { + "bbox": [ + 313, + 223, + 553, + 258 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 223, + 553, + 258 + ], + "type": "text", + "content": " further improves performance by up to 9.9 points, underscoring the importance of weighting model parameters based on their relevance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 273, + 555, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 273, + 555, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 555, + 417 + ], + "type": "text", + "content": "Initial solution. We examine the effect of initialization on FW-Merging. An ideal initial solution should either (1) be closer to the global optimum or (2) expand the constraint set with a more meaningful search space. As shown in Figure 4a, initializing FW-Merging with the Adamerging result improves performance compared to starting from the pre-trained model, likely because Adamerging is closer to the optimal point. In contrast, task arithmetic leads to worse performance than the pre-trained model, potentially due to its poor performance on vision tasks " + }, + { + "bbox": [ + 313, + 273, + 555, + 417 + ], + "type": "inline_equation", + "content": "(21.2\\%)" + }, + { + "bbox": [ + 313, + 273, + 555, + 417 + ], + "type": "text", + "content": ", suggesting it starts further from the optimum. Consequently, more FW iterations are required to achieve convergence." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 431, + 555, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 555, + 551 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 555, + 551 + ], + "type": "text", + "content": "Flexibility of merging functions. Although only a restricted set of merging functions ensure that FW-Merging remains within the convex hull, we demonstrate the flexibility of FW-Merging by showing its ability to enhance alternative merging functions. As shown in Figure 4b, applying FW-Merging with both Task Arithmetic and Ties-Merging improves performance on NLP and vision tasks, even though Ties-Merging does not necessarily stay within the convex hull. This suggests that FW-Merging remains effective across different merging functions." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 561, + 388, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 561, + 388, + 573 + ], + "spans": [ + { + "bbox": [ + 314, + 561, + 388, + 573 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "In this work, we extend model merging to a more challenging setting where the merging pool consists of a large number of black-box fine-tuned checkpoints. While existing methods require prior knowledge of model details to achieve optimized performance, our proposed Frank-Wolfe Merging (FW-Merging) scales effectively with a large number of black-box models, iteratively refining the merged model towards the optimal point defined by an objective function. Experiments demonstrate that FW-Merging achieves superior performance and scalability, paving the way for next-generation model merging." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3397" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "text", + "content": "6. Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "type": "text", + "content": "The support of the United Kingdom EPSRC (grant number UKRI256, EP/V028251/1, EP/N031768/1, EP/S030069/1, and EP/X036006/1), Intel, and AMD is gratefully acknowledged." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 153, + 115, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 153, + 115, + 166 + ], + "spans": [ + { + "bbox": [ + 56, + 153, + 115, + 166 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 172, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 172, + 296, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 172, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 61, + 172, + 296, + 217 + ], + "type": "text", + "content": "[1] Samuel K Ainsworth, Jonathan Hayase, Siddhartha Srinivasan, Khaled Saab, and Stefano Fusi. Git re-basin: Merging models modulo permutation symmetries. arXiv preprint arXiv:2209.04836, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 218, + 296, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 218, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 62, + 218, + 296, + 239 + ], + "type": "text", + "content": "[2] Rich Caruana. Multitask learning. Machine learning, 28: 41-75, 1997. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 242, + 296, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 242, + 296, + 275 + ], + "spans": [ + { + "bbox": [ + 62, + 242, + 296, + 275 + ], + "type": "text", + "content": "[3] Sahil Chaudhary. Code alpaca: An instruction-following llama model for code generation. https://github.com/sahil280114/codealpaca, 2023.6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 277, + 296, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 277, + 296, + 474 + ], + "spans": [ + { + "bbox": [ + 62, + 277, + 296, + 474 + ], + "type": "text", + "content": "[4] Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott Gray, Nick Ryder, Mikhail Pavlov, Alethea Power, Lukasz Kaiser, Mohammad Bavarian, Clemens Winter, Philippe Tillet, Felipe Petroski Such, Dave Cummings, Matthias Plappert, Fotios Chantzis, Elizabeth Barnes, Ariel Herbert-Voss, William Hebgen Guss, Alex Nichol, Alex Paino, Nikolas Tezak, Jie Tang, Igor Babuschkin, Suchir Balaji, Shantanu Jain, William Saunders, Christopher Hesse, Andrew N. Carr, Jan Leike, Josh Achiam, Vedant Misra, Evan Morikawa, Alec Radford, Matthew Knight, Miles Brundage, Mira Murati, Katie Mayer, Peter Welinder, Bob McGrew, Dario Amodei, Sam McCandlish, Ilya Sutskever, and Wojciech Zaremba. Evaluating large language models trained on code. 2021. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 475, + 296, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 475, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 62, + 475, + 296, + 531 + ], + "type": "text", + "content": "[5] Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks. In International conference on machine learning, pages 794-803. PMLR, 2018. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 533, + 296, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 533, + 296, + 586 + ], + "spans": [ + { + "bbox": [ + 62, + 533, + 296, + 586 + ], + "type": "text", + "content": "[6] Zhao Chen, Jiquan Ngiam, Yanping Huang, Thang Luong, Henrik Kretzschmar, Yuning Chai, and Dragomir Anguelov. Just pick a sign: Optimizing deep multitask models with gradient sign dropout. Advances in Neural Information Processing Systems, 33:2039-2050, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 589, + 296, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 589, + 296, + 633 + ], + "spans": [ + { + "bbox": [ + 62, + 589, + 296, + 633 + ], + "type": "text", + "content": "[7] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3606-3613, 2014. 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 635, + 296, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 635, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 62, + 635, + 296, + 677 + ], + "type": "text", + "content": "[8] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: Efficient finetuning of quantized llms. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 680, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 680, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 680, + 296, + 713 + ], + "type": "text", + "content": "[9] Marguerite Frank, Philip Wolfe, et al. An algorithm for quadratic programming. Naval research logistics quarterly, 3(1-2):95-110, 1956. 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[10] Ian J Goodfellow, Oriol Vinyals, and Andrew M Saxe. Qualitatively characterizing neural network optimization problems. arXiv preprint arXiv:1412.6544, 2014. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 108, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 152 + ], + "type": "text", + "content": "[11] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 154, + 553, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 154, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 316, + 154, + 553, + 198 + ], + "type": "text", + "content": "[12] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 200, + 553, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 200, + 553, + 221 + ], + "spans": [ + { + "bbox": [ + 316, + 200, + 553, + 221 + ], + "type": "text", + "content": "[13] Hugging Face. Open LLM Leaderboard, 2024. Accessed: March 4, 2025. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 224, + 553, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 224, + 553, + 267 + ], + "spans": [ + { + "bbox": [ + 316, + 224, + 553, + 267 + ], + "type": "text", + "content": "[14] Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Suchin Gururangan, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. Editing models with task arithmetic. arXiv preprint arXiv:2212.04089, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 270, + 553, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 270, + 553, + 314 + ], + "spans": [ + { + "bbox": [ + 316, + 270, + 553, + 314 + ], + "type": "text", + "content": "[15] Pavel Izmailov, Dmitrii Podoprikhin, Timur Garipov, Dmitry Vetrov, and Andrew Gordon Wilson. Averaging weights leads to wider optima and better generalization. arXiv preprint arXiv:1803.05407, 2018. 1, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 316, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 370 + ], + "type": "text", + "content": "[16] Junguang Jiang, Baixu Chen, Junwei Pan, Ximei Wang, Dapeng Liu, Jie Jiang, and Mingsheng Long. Forkmerge: Mitigating negative transfer in auxiliary-task learning. Advances in neural information processing systems, 36:30367-30389, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 372, + 553, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 372, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 317, + 372, + 553, + 449 + ], + "type": "text", + "content": "[17] Qiao Jin, Bhuwan Dhingra, Zhengping Liu, William Cohen, and Xinghua Lu. Pubmedqa: A dataset for biomedical research question answering. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2567-2577, 2019. 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 450, + 553, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 450, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 316, + 450, + 553, + 495 + ], + "type": "text", + "content": "[18] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013. 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 497, + 553, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 497, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 316, + 497, + 553, + 540 + ], + "type": "text", + "content": "[19] Simon Lacoste-Julien, Martin Jaggi, Mark Schmidt, and Patrick Pletscher. Block-coordinate frank-wolfe optimization for structural svms. In International Conference on Machine Learning, pages 53-61. PMLR, 2013. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 543, + 553, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 543, + 553, + 575 + ], + "spans": [ + { + "bbox": [ + 316, + 543, + 553, + 575 + ], + "type": "text", + "content": "[20] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 578, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 553, + 610 + ], + "type": "text", + "content": "[21] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 613, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 656 + ], + "type": "text", + "content": "[22] Yixiao Li, Yifan Yu, Chen Liang, Pengcheng He, Nikos Karampatziakis, Weizhu Chen, and Tuo Zhao. Loftq: Lorafine-tuning-aware quantization for large language models. arXiv preprint arXiv:2310.08659, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "text", + "content": "[23] Fuzhao Liu, Xiaotian Hu, Chengyu Chen, Zhijian Wang, Shuxin Xiao, Zhen Wang, Zhongxiang Wang, Chuanxin Xie, Eric Xing, and Song Han. Dare: Diversity-aware model merging for sparse-to-dense mixture-of-experts. arXiv preprint arXiv:2402.10887, 2024. 3, 4, 5" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3398" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[24] Liyang Liu, Yi Li, Zhanghui Kuang, J Xue, Yimin Chen, Wenming Yang, Qingmin Liao, and Wayne Zhang. Towards impartial multi-task learning. iclr, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 129 + ], + "type": "text", + "content": "[25] Prateek Yadav Liu et al. Resolving interference when merging models. arXiv preprint arXiv:2306.01708, 2023. 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 129, + 296, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 296, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 296, + 172 + ], + "type": "text", + "content": "[26] Shikun Liu, Edward Johns, and Andrew J Davison. End-to-end multi-task learning with attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1871-1880, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 172, + 295, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 295, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 295, + 227 + ], + "type": "text", + "content": "[27] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettle-moyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 227, + 295, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 227, + 295, + 272 + ], + "spans": [ + { + "bbox": [ + 56, + 227, + 295, + 272 + ], + "type": "text", + "content": "[28] Zhenyi Lu, Chenghao Fan, Wei Wei, Xiaoye Qu, Dangyang Chen, and Yu Cheng. Twin-merging: Dynamic integration of modular expertise in model merging. Advances in Neural Information Processing Systems, 37:78905–78935, 2025. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 272, + 295, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 295, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 295, + 337 + ], + "type": "text", + "content": "[29] Jiaqi Ma, Zhe Zhao, Xinyang Yi, Jilin Chen, Lichan Hong, and Ed H Chi. Modeling task relationships in multi-task learning with multi-gate mixture-of-experts. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1930-1939, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 337, + 295, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 337, + 295, + 382 + ], + "spans": [ + { + "bbox": [ + 56, + 337, + 295, + 382 + ], + "type": "text", + "content": "[30] Jiaqi Ma, Zhe Zhao, Jilin Chen, Ang Li, Lichan Hong, and Ed H Chi. Snr: Sub-network routing for flexible parameter sharing in multi-task learning. In Proceedings of the AAAI conference on artificial intelligence, pages 216-223, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 382, + 295, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 382, + 295, + 436 + ], + "spans": [ + { + "bbox": [ + 56, + 382, + 295, + 436 + ], + "type": "text", + "content": "[31] Sadhika Malladi, Tianyu Gao, Eshaan Nichani, Alex Damian, Jason D Lee, Danqi Chen, and Sanjeev Arora. Finetuning language models with just forward passes. Advances in Neural Information Processing Systems, 36:53038-53075, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 436, + 295, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 436, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 56, + 436, + 295, + 460 + ], + "type": "text", + "content": "[32] Priyanka Mary Mammen. Federated learning: Opportunities and challenges. arXiv preprint arXiv:2101.05428, 2021. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 460, + 295, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 295, + 502 + ], + "type": "text", + "content": "[33] Eric Mitchell, Kenton Lee, Alexander Khazatsky, Christopher D Manning, and Chelsea Finn. Editing models with task arithmetic. arXiv preprint arXiv:2212.04089, 2022. 3, 4, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 502, + 295, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 502, + 295, + 548 + ], + "spans": [ + { + "bbox": [ + 56, + 502, + 295, + 548 + ], + "type": "text", + "content": "[34] Ramesh Nallapati, Bowen Zhou, Caglar Gulcehre, Bing Xiang, et al. Abstractive text summarization using sequence-to-sequence rnns and beyond. arXiv preprint arXiv:1602.06023, 2016.5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 548, + 295, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 295, + 592 + ], + "type": "text", + "content": "[35] Shraman Prakash, Rohit Girdhar, Devamanyu Hazarika, Joao Carreira, and Andrew Zisserman. Unified model for image, video, audio and language tasks. arXiv preprint arXiv:2307.16184, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 592, + 295, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 295, + 647 + ], + "type": "text", + "content": "[36] Samyam Rajbhandari, Jeff Rasley, Olatunjri Ruwase, and Yuxiong He. Zero: Memory optimizations toward training trillion parameter models. In SC20: International Conference for High Performance Computing, Networking, Storage and Analysis, pages 1-16. IEEE, 2020. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 647, + 295, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 295, + 680 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 295, + 680 + ], + "type": "text", + "content": "[37] Sidak Pal Singh and Martin Jaggi. Merging models with fisher-weighted averaging. arXiv preprint arXiv:2111.09832, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 680, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 295, + 713 + ], + "type": "text", + "content": "[38] Johannes Stallkamp, Marc Schlipsing, Jan Salmen, and Christian Igel. The german traffic sign recognition benchmark: a multi-class classification competition. In *The 2011" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 715 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "international joint conference on neural networks, pages 1453-1460. IEEE, 2011. 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 95, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 553, + 140 + ], + "type": "text", + "content": "[39] Ximeng Sun, Rameswar Panda, Rogerio Feris, and Kate Saenko. Adashare: Learning what to share for efficient deep multi-task learning. Advances in Neural Information Processing Systems, 33:8728-8740, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 141, + 553, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 186 + ], + "type": "text", + "content": "[40] Anke Tang, Li Shen, Yong Luo, Liang Ding, Han Hu, Bo Du, and Dacheng Tao. Concrete subspace learning based interference elimination for multi-task model fusion. arXiv preprint arXiv:2312.06173, 2023. 3, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 186, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 241 + ], + "type": "text", + "content": "[41] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 241, + 553, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 286 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 286 + ], + "type": "text", + "content": "[42] Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R Bowman. Glue: A multi-task benchmark and analysis platform for natural language understanding. arXiv preprint arXiv:1804.07461, 2018. 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 286, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 331 + ], + "type": "text", + "content": "[43] Ke Wang, Nikolaos Dimitriadis, Guillermo Ortiz-Jimenez, François Fleuret, and Pascal Frossard. Localizing task information for improved model merging and compression. arXiv preprint arXiv:2405.07813, 2024. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 331, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 387 + ], + "type": "text", + "content": "[44] Zirui Wang, Daniel Duckworth, Shashank Nag, Kshitij Patil, Jonathon Shlens, Ekin D Cubuk, Barret Zoph, Joseph Campbell, Mitchell Wortsman, and Google Brain. lo-fi: distributed fine-tuning without communication. arXiv preprint arXiv:2210.11948, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 388, + 553, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 432 + ], + "type": "text", + "content": "[45] Yongxian Wei, Anke Tang, Li Shen, Feng Xiong, Chun Yuan, and Xiaochun Cao. Modeling multi-task model merging as adaptive projective gradient descent. arXiv preprint arXiv:2501.01230, 2025. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 432, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 432, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 432, + 553, + 499 + ], + "type": "text", + "content": "[46] Mitchell Wortsman, Gabriel Ilharco, Samir Yitzhak Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. arXiv preprint arXiv:2203.05482, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 499, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 499, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 499, + 553, + 544 + ], + "type": "text", + "content": "[47] Jianxiong Xiao, Krista A Ehinger, James Hays, Antonio Torralba, and Aude Oliva. Sun database: Exploring a large collection of scene categories. International Journal of Computer Vision, 119:3-22, 2016. 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 544, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 553, + 590 + ], + "type": "text", + "content": "[48] Prateek Yadav, Derek Tam, Leshem Choshen, Colin A Raffel, and Mohit Bansal. Ties-merging: Resolving interference when merging models. Advances in Neural Information Processing Systems, 36:7093-7115, 2023. 1, 2, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 590, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 646 + ], + "type": "text", + "content": "[49] Enneng Yang, Li Shen, Guibing Guo, Xingwei Wang, Xiaochun Cao, Jie Zhang, and Dacheng Tao. Model merging in llms, mllms, and beyond: Methods, theories, applications and opportunities, 2024. URL https://arxiv.org/abs/2408.07666, 2408.3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 646, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 646, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 646, + 553, + 690 + ], + "type": "text", + "content": "[50] Enneng Yang, Zhenyi Wang, Li Shen, Shiwei Liu, Guibing Guo, Xingwei Wang, and Dacheng Tao. Adamerging: Adaptive model merging for multi-task learning. arXiv preprint arXiv:2310.02575, 2023. 1, 2, 3, 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 690, + 553, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 690, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 316, + 690, + 553, + 715 + ], + "type": "text", + "content": "[51] Enneng Yang, Li Shen, Zhenyi Wang, Guibing Guo, Xiaojun Chen, Xingwei Wang, and Dacheng Tao. Representa" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3399" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 251 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "text", + "content": "tion surgery for multi-task model merging. arXiv preprint arXiv:2402.02705, 2024. 1, 3, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 139 + ], + "type": "text", + "content": "[52] Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How transferable are features in deep neural networks? Advances in neural information processing systems, 27, 2014. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 296, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 296, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 296, + 206 + ], + "type": "text", + "content": "[53] Jun Yu, Yutong Dai, Xiaokang Liu, Jin Huang, Yishan Shen, Ke Zhang, Rong Zhou, Eashan Adhikarla, Wenxuan Ye, Yixin Liu, et al. Unleashing the power of multitask learning: A comprehensive survey spanning traditional, deep, and pretrained foundation model eras. arXiv preprint arXiv:2404.18961, 2024. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 296, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 296, + 251 + ], + "type": "text", + "content": "[54] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. Advances in neural information processing systems, 33:5824-5836, 2020. 3" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3400" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Face Retouching with Diffusion Data Generation and Spectral Restorement/2a12f446-c130-4e36-bc58-dba590a7c6d7_content_list.json b/2025/Face Retouching with Diffusion Data Generation and Spectral Restorement/2a12f446-c130-4e36-bc58-dba590a7c6d7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4d75a12f2a6f08fa13e1934e0cb9a48726dd8041 --- /dev/null +++ b/2025/Face Retouching with Diffusion Data Generation and Spectral Restorement/2a12f446-c130-4e36-bc58-dba590a7c6d7_content_list.json @@ -0,0 +1,1775 @@ +[ + { + "type": "text", + "text": "Face Retouching with Diffusion Data Generation and Spectral Restored", + "text_level": 1, + "bbox": [ + 117, + 130, + 879, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhidan Xu $^{1}$ Xiaoqin Zhang $^{1*}$ Shijian Lu $^{2*}$ $^{1}$ Wenzhou University \n $^{2}$ Nanyang Technological University", + "bbox": [ + 300, + 179, + 699, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zhidan.xu@stu.wzu.edu.cn, xqzhang@wzu.edu.cn, Shijian.Lu@ntu.edu.sg", + "bbox": [ + 197, + 236, + 792, + 252 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/666c64ac7910fd7a5e8132c36811d290b9d13d0d4b2d8cc27a5aa6b4dbb3610f.jpg", + "image_caption": [ + "Figure 1. The proposed face retouching technique can repair local facial regions and remove blemishes with superb detail preservation and global consistency in facial colour, facial texture, etc. Zoom-in for more details." + ], + "image_footnote": [], + "bbox": [ + 94, + 287, + 225, + 526 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/79ee83db86b7150308208498944780079d30097a43ecfaf3419a8f6d610fd802.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 228, + 287, + 361, + 527 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ccb564ef2ab8c699c606b82659cc4e26a704ffff62b988db2d143d4d94e5645c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 287, + 496, + 527 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9fd608ab178b3d17144a6816ca858ed8697be679c0c75cad40bd7ea7521e20ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 287, + 632, + 527 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c53ec5dd3cc3090d92c1638ee577bc11af9012b23f6033112dc6c9d6ca16212f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 289, + 767, + 527 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ffc2a90a66bad763ec078a3a22d0a5adc06b8cfd5316b106503d4b9459079f4c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 769, + 289, + 903, + 527 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 585, + 326, + 602 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Face retouching has achieved impressive performance largely driven by its wide range of applications in various real-world tasks. However, most existing works often encounters a dilemma between global consistency and local detail preservation, partially due to the lack of large-scale and high-quality training data. We address the face retouching challenge from two perspectives. First, we create a large-scale face retouching benchmark to mitigate the data scarcity issue. The benchmark comprises 25,000 pairs of high-quality facial images (before and after face retouching) that contain a variety of facial attributes and blemish types such as acne and moles. Second, we design a novel framework that introduces frequency selection and restoration (FSR) and multi-resolution fusion (MRF) that leverages frequency-aware dynamic aggregation and spatial-frequency filtering to achieve global consistency and local detail preservation concurrently. Inspired by the principle", + "bbox": [ + 86, + 619, + 485, + 876 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of JPEG compression, FSR introduces frequency-domain quantization with spatial projections to learn enhanced feature representations. MRF fuses multi-resolution features via laplacian pyramid fusion, removing large-area blemishes and preserving local fine details effectively. Extensive experiments over multiple benchmarks show that the proposed framework outperforms the state-of-the-art quantitatively and qualitatively. The created benchmark also provides valuable data for training and evaluating both existing and future face retouching networks.", + "bbox": [ + 511, + 573, + 906, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 739, + 645, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Face retouching, which aims to beautify face images by removing various blemishes and smoothing facial skin, has attracted increasing interest due to its applications in various real-world tasks in social media, virtual avatar creation, etc. Manual retouching has been widely practiced in the literature, relying on expert retouchers who balance blemish removal and the global uniformity of skin tones meticulously. However, the process of manual retouching is usually time-consuming, highly skill-dependent, and short of", + "bbox": [ + 509, + 763, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 114, + 887, + 243, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14722", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "scalability.", + "bbox": [ + 89, + 90, + 165, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Various image processing techniques have been explored for automating the face retouching process without sacrificing retouching quality. Specifically, face retouching has been widely studied in traditional image processing and computer vision research, where different filtering and rule-based methods [14, 24, 31, 32] have been adopted to tackle the face retouching challenge. In addition, dedicated techniques have also been designed for various mobile applications [12, 13, 20] which rely on image filtering as well. Despite being lightweight, these traditional techniques share two constraints: 1) they rely heavily on fine-tuning of handcrafted parameters and cannot generalize well; 2) they tend to over/under-smooth skin textures, eliminate critical details like pores, and introduce unnatural whitening effects.", + "bbox": [ + 89, + 108, + 480, + 319 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Leveraging the great success of deep learning in image enhancement, several deep neural networks have been designed and achieved promising retouching performance [7, 9, 21, 23, 25, 29, 30]. However, these methods often struggle to balance local detail repairs and global consistency in skin colour, skin texture, etc., partially due to the lack of large-scale paired training images. For example, ABPN [9] improves retouching with pyramid structures and generative models but falls short in managing fine blemishes and retaining details. StyleRetouch [23] achieves impressive retouching but often encounters scalability and efficiency issues while handling high-resolution images.", + "bbox": [ + 89, + 321, + 482, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We tackle the face retouching challenges from two perspectives. The first is to create a benchmark that can effectively train and evaluate face retouching networks. Several benchmarks such as FFHQ [8] have been adopted for retouching studies but they are not created for the task of face retouching and therefore lack comprehensive representations of real-world blemish types and facial attributes. In addition, these benchmarks comprise real face images with excessive identity features which often cause privacy concerns. We create HGFR, a high-definition face retouching benchmark that has three desirable features. First, HGFR is large-scale containing up to 25,000 pairs of face images with and without blemishes. Second, HGFR is featured by superb image quality with an image resolution of 1024 by 1024. Third, HGFR has superior diversity including a large portion of synthetic images that are generated by three finetuned diffusion models [17]. The adoption of synthetic images mitigate the privacy concerns, obviates the costly and laborious image collection processes, and allows generating large-scale blemished facial images that cover a wide spectrum of imperfections and facial attributes.", + "bbox": [ + 89, + 506, + 482, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition, we design a face retouching network that can balance the local repairs and global consistency effectively. The network achieves face retouching with two designs. The first is frequency selection and restoration (FSR) that performs multi-resolution retouching by har", + "bbox": [ + 89, + 824, + 482, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "nessing both frequency and spatial signals for addressing localized blemishes without affecting the overall image structure. The second is multi-resolution fusion (MRF) that fuses high-resolution details with low-resolution context, ensuring that large-area blemishes can be repaired while maintaining global coherence. These two designs achieve fine-grained local corrections with broad structural consistency concurrently, delivering superior face retouching performance across multiple benchmarks.", + "bbox": [ + 511, + 90, + 903, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions of this work can be summarized in three major aspects. First, we create the first large-scale, high-quality and diverse face retouching benchmark that comprise 25,000 pairs of high-definition face images that are generated by diffusion models. The benchmark captures a variety of blemishes and facial attributes which can be employed for both training and evaluating face retouching networks. Second, we design a novel face retouching network that introduces frequency selection and restoration and multi-resolution fusion for effective face retouching. Third, extensive experiments over multiple benchmarks show that our face retouching network achieves superior performance as compared with the state-of-the-art. The proposed benchmark can be effectively adopted for training and evaluating face retouching networks as well.", + "bbox": [ + 511, + 227, + 903, + 454 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 470, + 653, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Face Retouching Studies", + "text_level": 1, + "bbox": [ + 511, + 494, + 736, + 511 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Face retouching has been widely explored in the literature in two representative approaches. The first leverages traditional image filtering techniques for boosting visual quality while preserving skin textures and details, largely driven by various commercial applications such as Pixlr [20], MeituPic [13], and BeautyCam [12]. However, this approach relies heavily on manual adjustments with poor scalability. The second explores deep learning for face retouching. For example, AutoRetouch [21] adopts residual learning over a large-scale dataset (FFHQR) created from FFHQ [8]. ABPN [9] introduces a pyramid architecture to handle blemishes in high-resolution images. BPFRe [30] addresses limited training data via blemish-aware processing and progressive restoration. RetouchFormer [29] synthesizes high-quality facial images to reduce dependency on paired training data. Despite their impressive performance, these deep learning methods often struggle in maintaining global consistency and local detail preservation.", + "bbox": [ + 511, + 518, + 903, + 790 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Face Retouching Data", + "text_level": 1, + "bbox": [ + 511, + 801, + 718, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Though face retouching has been studied for years, the research community still falls short in large-scale face retouching benchmarks. The lack of such benchmarks is largely attributed to the challenges in creating paired face images with and without blemishes. Several prior studies", + "bbox": [ + 511, + 825, + 903, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14723", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ac58da3de16893fa921c48da0d49c721ff65bd7f751b7dc963032b959c89f154.jpg", + "image_caption": [ + "Figure 2. Qualitative illustration of the created large-scale, high-quality, and diverse face retouching benchmark. Zoom-in for more details." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 906, + 305 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "attempted to tackle the challenges in different ways. For example, [10] presents an exemplar-based method that collects retouching data with laser-treated facial images. However, its applicability is limited to laser-induced effects only. AutoRetouch [21] introduces FFHQR, a large-scale retouching benchmark comprising 70,000 image pairs. However, many images in FFHQR do not have blemishes since the underlying FFHQ dataset was originally collected for face generation. We utilize diffusion models to synthesize paired face images with and without blemishes, leading to a comprehensive paired dataset that can be adopted for training and evaluating face retouching networks.", + "bbox": [ + 88, + 356, + 485, + 539 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. High-Definition Retouching Benchmark", + "text_level": 1, + "bbox": [ + 89, + 553, + 449, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Manual collection of paired retouching images is a challenging task that requires sophisticated control to ensure just one variable for each pair of face images, i.e., having or not having blemishes. In addition, collecting such large-scale and diverse paired images is laborious and time-consuming. We address this challenge by synthesizing paired face images with stable diffusion models, leading to the first large-scale, high-quality, and diverse benchmark that comprises 25,000 paired face images with diverse blemish types and facial attributes. Fig. 2 shows samples of synthesized image pairs with and without blemishes.", + "bbox": [ + 88, + 579, + 485, + 747 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Data Generation Process", + "text_level": 1, + "bbox": [ + 89, + 757, + 316, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We create the proposed face retouching benchmark in two stages, including the first stage that generates face images without blemishes and the second that generates the corresponding paired image with blemishes. We generate the face images without blemishes by leveraging three fine-tuned stable diffusion models. Specifically, we design 7,920 image-generation prompts that systematically permute combinations of facial attributes—including age,", + "bbox": [ + 88, + 780, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/bfde7dc39e556beb3481f9369fb084519fb229d1fbc6039d6ac663bdd998508b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Blemish TypesNumber of Each BlemishNumber of Synthe-sized FaceNumber of Natural Face
HGFRMoles, Freckles, Acne7,92023,7601,240
", + "bbox": [ + 514, + 354, + 908, + 448 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Details of the proposed benchmark HGFR.", + "bbox": [ + 553, + 457, + 862, + 472 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "gender, aesthetics, hairstyle, skin tone, and detailed features (e.g., lips, nose, eyes, face shape, and eyebrows).", + "bbox": [ + 511, + 501, + 906, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The generation with such combinatorial facial attributes ensures that the generated face images exhibit a broad spectrum of diversity and rich facial representations while preserving a natural and unblemished appearance.", + "bbox": [ + 511, + 532, + 905, + 593 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For each clean face image generated in the first stage, we generate a corresponding paired image with blemishes via a controlled modification process. Specifically, we leverage a state-of-the-art facial attribute segmentation model [19] to first partition the clean face image into semantically meaningful regions (e.g., cheeks, nose, and neck) while deliberately excluding key areas such as hair, eyes, eyebrows, and mouth to maintain critical identity cues. We then apply the pre-trained LoRA [6] to the isolated facial regions to synthesize blemishes such as moles, freckles, and acne. Such targeted blemish synthesis ensures that the paired images accurately reflect common facial imperfections without compromising the overall facial structure and identity, thereby balancing the local repairs and global consistency. More details on the image synthesis are provided in the supplementary material.", + "bbox": [ + 511, + 594, + 908, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Data Statistics", + "text_level": 1, + "bbox": [ + 511, + 847, + 661, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The proposed benchmark HGFR consists of 25,000 image pairs, including 23,760 synthesized image pairs and 1,240", + "bbox": [ + 511, + 869, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14724", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/7b122b6906e6dcc078542654ee18f433b4f4898fa66f6da93e20ee93cf5527f8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Dataset Name# Images/PairsResolutionFace SourceTypeFace with Blemishes
PPR10K11,161 images4K-8KNaturalPortrait Retouching5%
RetouchingFFHQ>500K images1024×1024NaturalFace Recognition16%
EFrR--NaturalFace Retouching-
DigiFace-1M>1M images-SyntheticVarious poses, lighting, and backgrounds-
FRTDD--Restoration-basedTampering Detection-
FFHQ70,000 pairs1024×1024NaturalFace Generation16%
FFHQR70,000 pairs1024×1024NaturalFace Retouching16%
HGFR (Ours)25,000 pairs1024×1024Synthetic & NaturalFace Retouching83%
", + "bbox": [ + 91, + 75, + 906, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 2. Summary of existing face retouching benchmarks. ${}^{\\prime } -$ ’ Indicates that the information is unavailable. The percentages of face images with Blemishes are determined by the zero-shot CLIP classifier. More details are provided in the supplementary materials.", + "bbox": [ + 89, + 332, + 906, + 362 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "natural ones. As detailed in Tab. 1, HGFR consists of three primary types of blemishes including moles, freckles, and acne—with 7,920 image pairs in each of the three categories. Such systematic design yields a comprehensive representation of facial blemishes across 23,760 synthetic images, augmented by a small set of natural faces.", + "bbox": [ + 88, + 387, + 480, + 478 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Tab. 2 compares HGFR with several existing retouching benchmark. HGFR has two unique features. First, it has superb diversity and includes both synthetic and natural face retouching images, enabling more robust and comprehensive face retouching evaluations. Second, HGFR contains a greater variety of blemish types. Based on the classification by zero-shot CLIP [16], up to $83\\%$ images in HGFR contain blemishes while most existing benchmarks have less than $20\\%$ image containing blemishes. The much larger number of blemish instances enables more robust training and evaluation of existing and future face retouching networks.", + "bbox": [ + 88, + 479, + 482, + 645 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 89, + 661, + 181, + 678 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 3 illustrates our overall network architecture, which is composed of several key components. First, a Soft Mask Generation Branch (SMGB) automatically predicts masks to guide localized blemish repair. Next, the frequency selection and restoration (FSR) refines skin regions by leveraging both frequency and spatial domain information. Finally, the multi-resolution fusion (MRF) integrates multi-resolution outputs to reconstruct a high-quality, blemish-free portrait.", + "bbox": [ + 89, + 686, + 482, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given an input image $I_0 \\in \\mathbb{R}^{h \\times w \\times 3}$ , we first generate a Laplacian planes, $L_I = [I_0, I_1, \\dots, I_l]$ , and the corresponding high-frequency group $L_H = [H_0, H_1, \\dots, H_l]$ derived from Laplacian pyramid decomposition [1] (with $l$ , the number of downsampling operations, set to 2 by default). The FSR is then applied to each $I_i \\in L_I$ to produce", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "retouched outputs $R_{i} \\in \\mathbb{R}\\frac{h}{2^{i}}\\times \\frac{w}{2^{i}}\\times 3$ . These multi-scale results are subsequently fused via MRF to reconstruct the final high-resolution portrait. Inspired by the Laplacian pyramid [1], MRF uses progressive upsampling combined with a Conv+LeakyReLU+Conv(CLC) module to refine details and ensure seamless integration of features across scales.", + "bbox": [ + 511, + 386, + 906, + 478 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Soft Mask Generation Branch (SMGB) is designed to automatically predict a mask $M \\in \\mathbb{R}^{\\frac{h}{2^l} \\times \\frac{w}{2^l} \\times 3}$ that highlights target regions for localized repair. We adopt a U-Net architecture with skip connections [18] to fuse low-level and high-level features, thereby enhancing segmentation accuracy. Although most datasets lack explicit ground truth masks $M_{gt}$ , these can be derived by computing the difference between the input image $I_l$ and its corresponding ground truth, followed by thresholding. The SMGB contributes in two ways: first, the predicted mask $M_{pred}$ enables the FSR to concentrate on defect regions; and second, joint training with SMGB enhances the network's overall contextual and semantic awareness, leading to more consistent retouching results.", + "bbox": [ + 511, + 479, + 908, + 693 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Frequency Selection and Restoration", + "text_level": 1, + "bbox": [ + 511, + 700, + 831, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address Challenge (1) outlined in Section 1—achieving local repairs with global consistency and detail fidelity—we propose a context-aware frequency selection and restoration (FSR) framework. As shown in Fig. 3, FSR adopts a multi-task design that includes Frequency-Aware Dynamic Aggregation, Space Domain Projection, and Selective Frequency Domain-based FFN.", + "bbox": [ + 511, + 723, + 906, + 829 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1.1. Preliminaries.", + "text_level": 1, + "bbox": [ + 511, + 835, + 656, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a feature map $z \\in \\mathbb{R}^{H \\times W \\times C}$ , we apply the Discrete Fourier Transform (DFT) channel-wise to map spatial-domain features into the frequency domain. For a single", + "bbox": [ + 511, + 853, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14725", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d426f5a4a2d62c2fb338f988adcf5b91ca3071eadb76c27cd8442afa12e01900.jpg", + "image_caption": [ + "Figure 3. Overview of our proposed blemish repair framework. The input image is first decomposed into a Laplacian pyramid to obtain multi-scale representations and high-frequency components. A soft mask generation branch (SMGB) then automatically predicts masks to highlight defect regions for targeted repair. Subsequently, the frequency selection and restoration (FSR) module leverages both dynamic spectral filtering and spatial-domain projection—via learnable quantization matrices and self-attention mechanisms—to adaptively restore fine details and maintain global consistency. Finally, the multi-resolution fusion (MRF) progressively fuses multi-scale outputs through upsampling and convolutional refinement, yielding a high-quality, blemish-free portrait." + ], + "image_footnote": [], + "bbox": [ + 89, + 61, + 911, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "channel, the Fourier transform is defined as:", + "bbox": [ + 89, + 529, + 382, + 542 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF \\left(z _ {C}\\right) (u, v) = \\sum_ {h = 0} ^ {H - 1} \\sum_ {w = 0} ^ {W - 1} z _ {C} (h, w) e ^ {- j 2 \\pi \\left(\\frac {h u}{H} + \\frac {w v}{W}\\right)} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 111, + 554, + 482, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The real and imaginary parts are then extracted, allowing us to compute the magnitude spectrum $A(z_{C})(u,v)$ and phase spectrum $\\Phi (z_C)(u,v)$ via:", + "bbox": [ + 89, + 608, + 483, + 654 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA \\left(z _ {C}\\right) (u, v) = \\sqrt {\\Re \\left(F \\left(z _ {C}\\right) (u , v)\\right) ^ {2} + \\Im \\left(F \\left(z _ {C}\\right) (u , v)\\right) ^ {2}} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 666, + 482, + 699 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Phi (z _ {C}) (u, v) = \\arctan \\left[ \\frac {\\Im (F (z _ {C}) (u , v))}{\\Re (F (z _ {C}) (u , v))} \\right] \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 710, + 482, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By sequentially applying Eqs. (1) to (3) to all channels, we obtain the full-channel magnitude and phase spectra.", + "bbox": [ + 89, + 751, + 483, + 781 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.2. Frequency-Aware Dynamic Aggregation.", + "text_level": 1, + "bbox": [ + 89, + 790, + 419, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In many transformer-based methods [2, 5, 27], global selfattention can overlook localized defect characteristics. To address this, our dynamic spectral filter operates in the frequency domain to target specific imperfections (e.g., moles, freckles, acne), as shown in Fig. 4 left. We aggregate the magnitude and phase information using pointwise convolu", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tion:", + "bbox": [ + 513, + 530, + 547, + 541 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF ^ {*} (z) (u, v) = \\mathcal {G} \\left(\\operatorname {C o n v} _ {1} ^ {1} \\left(F \\left(z _ {C} \\cdot M _ {\\text {p r e d}}\\right) (u, v)\\right)\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 547, + 903, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{F}(z_C\\cdot M_{pred})(u,v)\\in$ $\\{A(z_C\\cdot M_{pred})(u,v),P(z_C\\cdot M_{pred})(u,v)\\}$ and $\\mathcal{G}(\\cdot)$ is the GEGLU activation function [22]. To adaptively weight the frequency components, we introduce learnable quantization matrices $W_{A}$ and $W_{\\Phi}$ :", + "bbox": [ + 514, + 578, + 903, + 654 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} A ^ {\\prime} (z) (u, v) = W _ {A} \\odot A (z) (u, v), \\\\ I ^ {\\prime} (z) (u, v) = W _ {I} \\odot I (z) (u, v). \\end{array} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 660, + 903, + 685 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Phi^ {\\prime} (z) (u, v) = W _ {\\Phi} \\odot \\Phi (z) (u, v).\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 680, + 818, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The updated frequency feature is then recomposed as:", + "bbox": [ + 513, + 704, + 870, + 719 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF ^ {\\prime} (z) (u, v) = A ^ {\\prime} (z) (u, v) e ^ {j \\Phi^ {\\prime} (z) (u, v)}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 726, + 903, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This filtered component is combined with the original frequency features via a residual connection:", + "bbox": [ + 511, + 752, + 903, + 781 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {F} (z) (u, v) = F ^ {\\prime} (z) (u, v) + F (z) (u, v). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 789, + 903, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Reconstruction into the spatial domain is achieved by extracting the real and imaginary parts and applying the inverse Fourier Transform:", + "bbox": [ + 511, + 814, + 903, + 858 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\Re} (z) (u, v) = \\tilde {A} (z) (u, v) \\cdot \\cos \\tilde {\\Phi} (z) (u, v), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 864, + 903, + 888 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathfrak {S}} (z) (u, v) = \\tilde {A} (z) (u, v) \\cdot \\sin \\tilde {\\Phi} (z) (u, v),\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 886, + 846, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14726", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/819a22ee551f1e3fe871f58e5db2d21248e23b830432e0f51b7168c370b6dc6e.jpg", + "image_caption": [ + "Figure 4. Details of frequency selection and restoration (FSR)" + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 483, + 175 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {z} = \\mathcal {F} ^ {- 1} \\left(\\tilde {\\Re} (z) (u, v), \\tilde {\\Im} (z) (u, v)\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 237, + 482, + 263 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The Fourier transform and its inverse are implemented via DFT and IDFT algorithms [3, 4, 34]. We denote the entire FADA process as $\\mathcal{F}\\mathcal{A}(\\cdot)$ .", + "bbox": [ + 89, + 268, + 483, + 315 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.3. Space Domain Projection.", + "text_level": 1, + "bbox": [ + 89, + 323, + 316, + 338 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After frequency-domain processing, we further refine features in the spatial domain, as shown in Fig. 4 right. For an input feature $z \\in \\mathbb{R}^{H \\times W \\times C}$ , layer normalization $\\mathcal{L}(\\cdot)$ [11] is applied to yield $\\mathcal{L}(z)$ . This normalized feature is projected into query $(Q)$ , key $(K)$ , and value $(V)$ matrices via independent frequency-domain projections:", + "bbox": [ + 89, + 342, + 483, + 434 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nQ = \\mathcal {F} \\mathcal {A} _ {Q} (\\mathcal {L} (\\tilde {z})), K = \\mathcal {F} \\mathcal {A} _ {K} (\\mathcal {L} (\\tilde {z})), V = \\mathcal {F} \\mathcal {A} _ {V} (\\mathcal {L} (\\tilde {z})), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 444, + 486, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{F}\\mathcal{A}_Q$ , $\\mathcal{F}\\mathcal{A}_K$ , and $\\mathcal{F}\\mathcal{A}_V$ denote learnable projection functions. The output of FSR is then given by:", + "bbox": [ + 89, + 476, + 483, + 506 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nz ^ {\\prime} = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {\\Lambda}} + B\\right) V, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 517, + 482, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "with $\\Lambda$ representing the channel dimension and $B$ the positional bias. Finally, a residual connection is used:", + "bbox": [ + 89, + 561, + 483, + 593 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nz ^ {*} = z ^ {\\prime} + z. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 241, + 604, + 480, + 621 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.4. Selective Frequency Domain-based FFN", + "text_level": 1, + "bbox": [ + 89, + 642, + 416, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To further enhance feature representation, as shown in Fig. 5 we design a selective frequency feed-forward network (SFFN) that adaptively retains critical frequency components for latent blemish removal. Inspired by the JPEG compression algorithm, SFFN employs a learnable quantization matrix:", + "bbox": [ + 89, + 662, + 483, + 751 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} z _ {1} = F \\left(P \\left(C o n v _ {1} ^ {1} (\\mathcal {L} (z)))\\right)\\right), \\\\ z _ {2} = F ^ {- 1} \\left(W \\odot z _ {1}\\right), \\tag {13} \\\\ z _ {\\text {o u t}} = \\mathcal {G} \\left(P ^ {- 1} \\left(z _ {2}\\right)\\right) + z ^ {*}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 761, + 482, + 843 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $P(\\cdot)$ and $P^{-1}(\\cdot)$ represent patch unfolding and folding operations. The output $z_{\\mathrm{out}}$ is then fed into subsequent layers.", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/71de102d6168d7cf02648eba3f2329e64cd15e0e610fbff7b7e635cf53361b5b.jpg", + "image_caption": [ + "Figure 5. Details of selective frequency domain-based FFN" + ], + "image_footnote": [], + "bbox": [ + 517, + 63, + 903, + 190 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6bd23f8512ed513e7538d18de33844475b45e6309dd000089f923edefecc18f8.jpg", + "image_caption": [ + "(a) Input", + "Figure 6. Visual Comparison of MRF at Different Resolutions: (a) Input image; (b) Low Resolution image, with large blemishes removed; (c) Original Resolution image, with fine blemishes removed; (d) Our method, achieving the best retouching." + ], + "image_footnote": [], + "bbox": [ + 516, + 228, + 612, + 303 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d632f7a4455f60dffa4c0a3fa3785ff5aa1d969160602204548f117ba703cd2d.jpg", + "image_caption": [ + "(b) Result on", + "Low Resolution" + ], + "image_footnote": [], + "bbox": [ + 614, + 228, + 710, + 303 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5c9a2ab99ac84dee4030e670208a86ab85603fc33ff67e5ffba81620f9e0dae7.jpg", + "image_caption": [ + "(c) Result on", + "Original Resolution" + ], + "image_footnote": [], + "bbox": [ + 709, + 228, + 805, + 303 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6b09e06d585728865577e5449d313237669699faf4798e91de5303c9e044466a.jpg", + "image_caption": [ + "(d) Our" + ], + "image_footnote": [], + "bbox": [ + 807, + 228, + 903, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Multi-Resolution Fusion", + "text_level": 1, + "bbox": [ + 511, + 412, + 736, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We observe that while high-resolution features are essential for restoring fine details, low-resolution features are more effective for repairing large-area blemishes. The differences can be observed in Fig. 6. To combine such complementary advantages, we propose MRF to fuse multi-resolution features to produce a detail-rich, blemish-free portrait. Specifically, MRF incorporates high-frequency components from the original image to compensate for the information loss during downsampling. The fusion process is defined as:", + "bbox": [ + 511, + 434, + 906, + 571 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} O _ {l - 1} = \\operatorname {C o n v} _ {3} ^ {3} \\left(\\sigma \\left(\\operatorname {C o n v} _ {3} ^ {3} \\left(\\operatorname {C a t} \\left(u p \\left(O _ {l}\\right), H _ {l - 1}\\right)\\right)\\right)\\right) \\tag {14} \\\\ + u p (R _ {l - 1}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 577, + 903, + 623 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": ", where $O_{l}$ , $O_{l-1}$ denotes the fusion result, and $O_{l} = R_{l}$ . $R_{l}$ , $R_{l-1}$ corresponds to the different resolutions of the blemish removal face images generated by Laplace image $I_{l}$ , $I_{l-1}$ through the face blemish removal network, respectively. $H_{l-1}$ is the high-frequency component of Laplace image $I_{l-1}$ , $up(\\cdot)$ denotes bilinear interpolation, $\\mathrm{Cat}(\\cdot)$ channel-wise concatenation. The CLC [15] is implemented through sequential $Conv_{3}^{3}$ layers with leaky ReLU activation, which is a convolution with a $3 \\times 3$ kernel and stride 3, and $\\sigma$ is a leaky ReLU activation (negative slope 0.2); when $l = 1$ , $O_{l-1} = O_{0}$ i.e. the final blemish-removed face image.", + "bbox": [ + 511, + 630, + 906, + 811 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 823, + 645, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 511, + 848, + 720, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We conduct extensive experiments over two benchmarks to validate our face retouching network. The", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14727", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2d54259064f5256db0da009fa71a2fef575f3a73abc799ae42cffef8e3db3caf.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetsFFHQRHGFRFFHQR + HGFR
MetricsPSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
AutoRetouch29.380.9180.076630.680.8920.082331.060.9240.0744
pix2pixHD41.180.9800.013342.370.9650.016742.680.9820.0132
ABPN44.260.9920.017647.020.9860.018947.650.9920.0178
BPFRe45.130.9930.009247.280.9890.013448.520.9930.0128
RetouchFormer45.240.9930.009047.610.9900.012848.780.9930.0105
Ours45.480.9930.008248.020.9910.012549.150.9940.0094
", + "bbox": [ + 91, + 88, + 906, + 223 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Quantitative comparison of our method with existing methods over three training setups ( $\\downarrow$ Lower is better; $\\uparrow$ Higher is better)", + "bbox": [ + 101, + 232, + 893, + 247 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ebb5f8f7675142b521bc957f45c4af8619ffeb23a0f734a09dfc8cbbc7de89d8.jpg", + "image_caption": [ + "Figure 7. Qualitative comparison of our method with several state-of-the-art methods on Internet images." + ], + "image_footnote": [], + "bbox": [ + 93, + 261, + 906, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "first is our proposed synthetic benchmark HGFR as detailed in Section 3. The second is the benchmark FFHQR [8], a large-scale face retouching benchmark consisting of 70,000 high-definition facial images paired with their retouched versions. We randomly split HGFR into 20,000 images for training, 2,500 for validation, and 2,500 for testing. For FFHQR, we follow the data splitting in [8].", + "bbox": [ + 88, + 573, + 482, + 680 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation Details. Our model and baselines are implemented in Python 3.8 using PyTorch 2.3.1 and trained on 4 NVIDIA RTX 4090 GPUs. We adopt the Adam optimizer with $\\beta_{1} = 0.9$ , $\\beta_{2} = 0.999$ , and set the initial learning rate to 0.0005 for a batch size of 8. After 100 epochs, the learning rate is reduced by a factor of 10. The default number of downsampling steps ( $l$ ) is set to 2. Training takes approximately 15 hours on HGFR, 35 hours on FFHQR, and 55 hours on FFHQR+HGFR.", + "bbox": [ + 88, + 684, + 482, + 819 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation Metrics. We evaluate the retouching with three evaluation metrics including Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM)[28], and Learned Perceptual Image Patch Similarity (LPIPS)[33]. The three metrics have been widely", + "bbox": [ + 88, + 824, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "adopted for image generation evaluations.", + "bbox": [ + 511, + 574, + 790, + 589 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Quantitative and Qualitative Experiments", + "text_level": 1, + "bbox": [ + 511, + 603, + 870, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our face retouching method with several existing work including Pix2PixHD [26], AutoRetouch [21], ABPN [9], BPFRe [30], and Retouch-Former [29]. Pix2PixHD is an image-to-image translation network and the rest are specifically designed for facial retouching. We train three retouching models by using the HGFR training images, the FFHQR training images, and a combination of the training images of the two datasets (i.e., FFHQR+HGFR). Tab. 3 shows experimental results. It can be observed that the proposed method performs clearly the best consistently across all datasets and evaluation metrics. The superior face retouching performance is largely attributed to our designed FSR and MRF, more details to be discussed in the ensuing ablation experiments.", + "bbox": [ + 509, + 627, + 906, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also conduct qualitative comparisons to further demonstrate the superiority of our designed face retouching network as illustrated in Fig. 7. It can be observed that AutoRetouch tends to introduce global noises and fails to", + "bbox": [ + 511, + 839, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14728", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f5ecec7d3205b1cd010a4b6f00b0f50cfbfa28e60bd0522fbe6498137ed08438.jpg", + "image_caption": [ + "(a) Input" + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 285, + 189 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8b7d8eb23a34d43cfcd2d655d4de4c5c9e87bdc1bea243b0e2dd2dd2daef9440.jpg", + "image_caption": [ + "(b) Retouching with FFHQR" + ], + "image_footnote": [], + "bbox": [ + 285, + 88, + 480, + 189 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/faf6a1349bbd690c121c45918d335791461ade172e753edb5a688df734a23da6.jpg", + "image_caption": [ + "(c) Retouching with HGFR", + "Figure 8. Face retouching with models trained by using FFHQR training data alone in (b), HGFR training data alone in (c) and the combination of HGFR and FFHQR training data in (d)." + ], + "image_footnote": [], + "bbox": [ + 91, + 200, + 287, + 300 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fd957c41b234c399878da7a8d426f8814b599a69ae3f41615c30b24114930f6d.jpg", + "image_caption": [ + "(d) Retouching with FFHQR+HGFR" + ], + "image_footnote": [], + "bbox": [ + 285, + 200, + 480, + 300 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "clearly delineate retouched regions while Pix2PixHD suffers from noticeable color shifts. ABPN, RetouchFormer, and BPFRe can deliver competitive retouching effects, but their outputs fall short in detail preservation and refinement. As a comparison, our method balances the local blemish repair and global facial consistency and demonstrates superior retouching experience.", + "bbox": [ + 89, + 393, + 483, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The experiments also validate the effectiveness of our benchmark from two perspectives: 1) training effective retouching network; and 2) augmenting existing benchmark for better retouching performance. Fig. 8 shows qualitative experiments. We can see that the model trained using HGFR alone performs clearly better than the one trained using FFHQR alone. The clear better retouching effect is largely because our benchmark is specifically synthesized with large-scale, high-quality, and diverse paired images. In addition, we can observe that the model using HGRR+FFHQR performs the best, demonstrating that our created benchmark can augment and complement existing benchmarks effectively.", + "bbox": [ + 89, + 501, + 483, + 698 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 710, + 243, + 726 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct ablation experiments to examine how the proposed FSR and MRF contribute to the face retouching as shown in Tab. 4. The baseline FFN only achieves 42.93 dB in PSNR and 0.967 in SSIM. Including FSR improves the two metrics to 45.79 dB and 0.980, indicating that our frequency-aware design substantially improves detail preservation. Further including MRF on top of FSR yields the best performance (47.81 dB/0.990), indicating the complementary role of MRF via multi-scale feature refinement.", + "bbox": [ + 89, + 733, + 482, + 868 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also conduct qualitative ablation experiments. As Fig. 9 shows, FSR outperforms the baseline clearly in bal", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/76c3bee8d22bad114f848b418e81923d90e6fe1a9e1d2abc1889706abe0e2fad.jpg", + "image_caption": [ + "(a) Input" + ], + "image_footnote": [], + "bbox": [ + 516, + 88, + 710, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a1d9074a402a4d60267611d041b062c39ec64afb105b2f4cf857871045139d1d.jpg", + "image_caption": [ + "(b) Baseline" + ], + "image_footnote": [], + "bbox": [ + 710, + 88, + 903, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/34c967b81367a3820f2916c8c7ff2b30cf7b45b08243c53d7838a0962daf4143.jpg", + "image_caption": [ + "(c) Baseline+FSR", + "Figure 9. Qualitative ablation studies on the proposed FSR and MRF. FSR substantially improves the detail preservation and global consistency as compared with the Baseline, and MRF further repairs residual blemishes with enhanced retouching effects." + ], + "image_footnote": [], + "bbox": [ + 516, + 204, + 710, + 306 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4e68a1c32716355ecbe6c26c7dab80d83799d4fe4a12254b25c1a6386220173b.jpg", + "image_caption": [ + "(d) Baseline+FSR+MRF" + ], + "image_footnote": [], + "bbox": [ + 710, + 204, + 903, + 306 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ancing detail preservation and global consistency, and MRF further enhances the face retouching by removing the residual blemishes. Please refer to the appendix for more detailed ablation experiments.", + "bbox": [ + 511, + 417, + 905, + 479 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3129ff7e05eb284def3a419016945df1b886554aacc996fc3c98078d26eb970b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MRFFSRFFNPSNRs/SSIMs
FFN(baseline)✓42.93/0.967
FSR✓45.79/0.980
MRF+FSR✓✓47.81/0.990
", + "bbox": [ + 514, + 494, + 905, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation Studies on FSR and MRF over HGFR.", + "bbox": [ + 539, + 571, + 877, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 599, + 633, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presents a high-definition face retouching benchmark and a novel face retouching framework that can effectively balance local detail repair with global consistency. The benchmark comprises 25,000 high-quality paired face retouching images that are generated by diffusion models, possessing superb diversity for training and evaluation face retouching networks. The framework integrates frequency selection and sestoration with multi-resolution fusion, demonstrating superior retouching performance. Extensive experiments confirm the effectiveness of our created benchmark and designed retouching framework.", + "bbox": [ + 511, + 626, + 906, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ACKNOWLEDGEMENT This work was supported in part by the National Key Research and Development Program Project of China [grant no. 2024YFC3306902], in part by the National Natural Science Foundation of China [grant no. U24A20242], and in part by the Zhejiang Provincial Natural Science Foundation [grant no. LDT23F02024F02].", + "bbox": [ + 511, + 794, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14729", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Peter J Burt and Edward H Adelson. The laplacian pyramid as a compact image code. In Readings in computer vision, pages 671-679. Elsevier, 1987. 4", + "[2] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ArXiv, abs/2010.11929, 2020. 5", + "[3] Matteo Frigo and Steven G Johnson. Fftw: An adaptive software architecture for the fft. In ICASSP, pages 1381-1384. IEEE, 1998. 6", + "[4] Junming Hou, Qi Cao, Ran Ran, Che Liu, Junling Li, and Liang-jian Deng. Bidomain modeling paradigm for pan-sharpening. In ACM MM, pages 347-357, 2023. 6", + "[5] Jie Hu, Li Shen, Samuel Albanie, Gang Sun, and Enhua Wu. Squeeze-and-excitation networks. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7132-7141, 2017. 5", + "[6] J. Edward Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. ArXiv, abs/2106.09685, 2021. 3", + "[7] Le Jiang, Yan Huang, Lianxin Xie, Wen Xue, Cheng Liu, Si Wu, and Hau-San Wong. Hunting blemishes: Language-guided high-fidelity face retouching transformer with limited paired data. In ACM MM, pages 5102-5111, 2024. 2", + "[8] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, pages 4401-4410, 2019. 2, 7", + "[9] Biwen Lei, Xiefan Guo, Hongyu Yang, Miaomiao Cui, Xuansong Xie, and Di Huang. Abpn: adaptive blend pyramid network for real-time local retouching of ultra high-resolution photo. In CVPR, pages 2108-2117, 2022. 1, 2, 7", + "[10] Tsung-Ying Lin, Yu-Ting Tsai, Tsung-Shian Huang, Wen-Chieh Lin, and Jung-Hong Chuang. Exemplar-based freckle retouching and skin tone adjustment. Computers & Graphics, 78:54-63, 2019. 3", + "[11] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 6", + "[12] Meitu Inc. Beautycam. https://meiyan.meipai.com/, 2013.2", + "[13] Meitu Inc. Meitupic. http://xiuxiu.meitu.com/, 2013.2", + "[14] Dongbo Min, Sunghwan Choi, Jiangbo Lu, Bumsub Ham, Kwanghoon Sohn, and Minh N Do. Fast global image smoothing based on weighted least squares. IEEE TIP, 23 (12):5638-5653, 2014. 2", + "[15] Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. CoRR, abs/1511.06434, 2015. 6" + ], + "bbox": [ + 93, + 114, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 4", + "[17] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2", + "[18] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 4", + "[19] Ioannis Sarridis, Christos Koutlis, Symeon Papadopoulos, and Christos Diou. Facex: Understanding face attribute classifiers through summary model explanations. In Proceedings of the 2024 International Conference on Multimedia Retrieval, pages 758-766, 2024. 3", + "[20] Ola Sevandersson. Pixlr. https://www pixlr.com/, 2008.2", + "[21] Alireza Shafaei, James J Little, and Mark Schmidt. Autoretouch: Automatic professional face retouching. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 990-998, 2021. 1, 2, 3, 7", + "[22] Noam Shazeer. Glu variants improve transformer. arXiv preprint arXiv:2002.05202, 2020. 5", + "[23] Wanchao Su, Can Wang, Chen Liu, Fangzhou Han, Hongbo Fu, and Jing Liao. Styleretoucher: Generalized portrait image retouching with gan priors. IEEE TVCG, 2024. 2", + "[24] Carlo Tomasi and Roberto Manduchi. Bilateral filtering for gray and color images. In ICCV, pages 839-846. IEEE, 1998. 2", + "[25] Sudha Velusamy, Rishubh Parihar, Raviprasad Kini, and Aniket Rege. Fabsoften: Face beautification via dynamic skin smoothing, guided feathering, and texture restoration. In CVPRW, pages 530-531, 2020. 2", + "[26] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In CVPR, pages 8798-8807, 2018. 1, 7", + "[27] X. Wang, Ross B. Girshick, Abhinav Kumar Gupta, and Kaiming He. Non-local neural networks. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7794-7803, 2017. 5", + "[28] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 13(4):600-612, 2004. 7", + "[29] Xue Wen, Lianxin Xie, Le Jiang, Tianyi Chen, Si Wu, Cheng Liu, and Hau-San Wong. Retouchformer: semi-supervised high-quality face retouching transformer with prior-based selective self-attention. In AAAI, pages 5903–5911, 2024. 1, 2, 7", + "[30] Lianxin Xie, Wen Xue, Zhen Xu, Si Wu, Zhiwen Yu, and Hau San Wong. Blemish-aware and progressive face re" + ], + "bbox": [ + 516, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14730", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "touching with limited paired data. In CVPR, pages 5599-5608, 2023. 2, 7", + "[31] Li Xu, Cewu Lu, Yi Xu, and Jiaya Jia. Image smoothing via 10 gradient minimization. In Proceedings of the 2011 SIGGRAPH Asia conference, pages 1-12, 2011. 2", + "[32] Qi Zhang, Li Xu, and Jiaya Jia. 100+ times faster weighted median filter (wmf). In CVPR, pages 2830-2837, 2014. 2", + "[33] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 7", + "[34] Man Zhou, Keyu Yan, Xueyang Fu, Aiping Liu, and Chengjun Xie. Pan-guided band-aware multi-spectral feature enhancement for pan-sharpening. IEEE Transactions on Computational Imaging, 9:238-249, 2023. 6" + ], + "bbox": [ + 91, + 90, + 482, + 305 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14731", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Face Retouching with Diffusion Data Generation and Spectral Restorement/2a12f446-c130-4e36-bc58-dba590a7c6d7_model.json b/2025/Face Retouching with Diffusion Data Generation and Spectral Restorement/2a12f446-c130-4e36-bc58-dba590a7c6d7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fc87a6dcd2c40a1a9c353f3465abe54325bf8769 --- /dev/null +++ b/2025/Face Retouching with Diffusion Data Generation and Spectral Restorement/2a12f446-c130-4e36-bc58-dba590a7c6d7_model.json @@ -0,0 +1,2255 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.119, + 0.131, + 0.88, + 0.154 + ], + "angle": 0, + "content": "Face Retouching with Diffusion Data Generation and Spectral Restored" + }, + { + "type": "text", + "bbox": [ + 0.302, + 0.18, + 0.7, + 0.235 + ], + "angle": 0, + "content": "Zhidan Xu\\(^{1}\\) Xiaoqin Zhang\\(^{1*}\\) Shijian Lu\\(^{2*}\\) \n\\(^{1}\\)Wenzhou University \n\\(^{2}\\)Nanyang Technological University" + }, + { + "type": "text", + "bbox": [ + 0.199, + 0.237, + 0.794, + 0.253 + ], + "angle": 0, + "content": "zhidan.xu@stu.wzu.edu.cn, xqzhang@wzu.edu.cn, Shijian.Lu@ntu.edu.sg" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.289, + 0.227, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.289, + 0.362, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.289, + 0.498, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.289, + 0.633, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.29, + 0.768, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.77, + 0.29, + 0.904, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.531, + 0.908, + 0.561 + ], + "angle": 0, + "content": "Figure 1. The proposed face retouching technique can repair local facial regions and remove blemishes with superb detail preservation and global consistency in facial colour, facial texture, etc. Zoom-in for more details." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.587, + 0.327, + 0.603 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.62, + 0.486, + 0.877 + ], + "angle": 0, + "content": "Face retouching has achieved impressive performance largely driven by its wide range of applications in various real-world tasks. However, most existing works often encounters a dilemma between global consistency and local detail preservation, partially due to the lack of large-scale and high-quality training data. We address the face retouching challenge from two perspectives. First, we create a large-scale face retouching benchmark to mitigate the data scarcity issue. The benchmark comprises 25,000 pairs of high-quality facial images (before and after face retouching) that contain a variety of facial attributes and blemish types such as acne and moles. Second, we design a novel framework that introduces frequency selection and restoration (FSR) and multi-resolution fusion (MRF) that leverages frequency-aware dynamic aggregation and spatial-frequency filtering to achieve global consistency and local detail preservation concurrently. Inspired by the principle" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.574, + 0.907, + 0.726 + ], + "angle": 0, + "content": "of JPEG compression, FSR introduces frequency-domain quantization with spatial projections to learn enhanced feature representations. MRF fuses multi-resolution features via laplacian pyramid fusion, removing large-area blemishes and preserving local fine details effectively. Extensive experiments over multiple benchmarks show that the proposed framework outperforms the state-of-the-art quantitatively and qualitatively. The created benchmark also provides valuable data for training and evaluating both existing and future face retouching networks." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.74, + 0.646, + 0.755 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Face retouching, which aims to beautify face images by removing various blemishes and smoothing facial skin, has attracted increasing interest due to its applications in various real-world tasks in social media, virtual avatar creation, etc. Manual retouching has been widely practiced in the literature, relying on expert retouchers who balance blemish removal and the global uniformity of skin tones meticulously. However, the process of manual retouching is usually time-consuming, highly skill-dependent, and short of" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.888, + 0.245, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "14722" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.166, + 0.106 + ], + "angle": 0, + "content": "scalability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.109, + 0.482, + 0.32 + ], + "angle": 0, + "content": "Various image processing techniques have been explored for automating the face retouching process without sacrificing retouching quality. Specifically, face retouching has been widely studied in traditional image processing and computer vision research, where different filtering and rule-based methods [14, 24, 31, 32] have been adopted to tackle the face retouching challenge. In addition, dedicated techniques have also been designed for various mobile applications [12, 13, 20] which rely on image filtering as well. Despite being lightweight, these traditional techniques share two constraints: 1) they rely heavily on fine-tuning of handcrafted parameters and cannot generalize well; 2) they tend to over/under-smooth skin textures, eliminate critical details like pores, and introduce unnatural whitening effects." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.323, + 0.483, + 0.504 + ], + "angle": 0, + "content": "Leveraging the great success of deep learning in image enhancement, several deep neural networks have been designed and achieved promising retouching performance [7, 9, 21, 23, 25, 29, 30]. However, these methods often struggle to balance local detail repairs and global consistency in skin colour, skin texture, etc., partially due to the lack of large-scale paired training images. For example, ABPN [9] improves retouching with pyramid structures and generative models but falls short in managing fine blemishes and retaining details. StyleRetouch [23] achieves impressive retouching but often encounters scalability and efficiency issues while handling high-resolution images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.507, + 0.483, + 0.824 + ], + "angle": 0, + "content": "We tackle the face retouching challenges from two perspectives. The first is to create a benchmark that can effectively train and evaluate face retouching networks. Several benchmarks such as FFHQ [8] have been adopted for retouching studies but they are not created for the task of face retouching and therefore lack comprehensive representations of real-world blemish types and facial attributes. In addition, these benchmarks comprise real face images with excessive identity features which often cause privacy concerns. We create HGFR, a high-definition face retouching benchmark that has three desirable features. First, HGFR is large-scale containing up to 25,000 pairs of face images with and without blemishes. Second, HGFR is featured by superb image quality with an image resolution of 1024 by 1024. Third, HGFR has superior diversity including a large portion of synthetic images that are generated by three finetuned diffusion models [17]. The adoption of synthetic images mitigate the privacy concerns, obviates the costly and laborious image collection processes, and allows generating large-scale blemished facial images that cover a wide spectrum of imperfections and facial attributes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.903 + ], + "angle": 0, + "content": "In addition, we design a face retouching network that can balance the local repairs and global consistency effectively. The network achieves face retouching with two designs. The first is frequency selection and restoration (FSR) that performs multi-resolution retouching by har" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.227 + ], + "angle": 0, + "content": "nessing both frequency and spatial signals for addressing localized blemishes without affecting the overall image structure. The second is multi-resolution fusion (MRF) that fuses high-resolution details with low-resolution context, ensuring that large-area blemishes can be repaired while maintaining global coherence. These two designs achieve fine-grained local corrections with broad structural consistency concurrently, delivering superior face retouching performance across multiple benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.228, + 0.905, + 0.455 + ], + "angle": 0, + "content": "The contributions of this work can be summarized in three major aspects. First, we create the first large-scale, high-quality and diverse face retouching benchmark that comprise 25,000 pairs of high-definition face images that are generated by diffusion models. The benchmark captures a variety of blemishes and facial attributes which can be employed for both training and evaluating face retouching networks. Second, we design a novel face retouching network that introduces frequency selection and restoration and multi-resolution fusion for effective face retouching. Third, extensive experiments over multiple benchmarks show that our face retouching network achieves superior performance as compared with the state-of-the-art. The proposed benchmark can be effectively adopted for training and evaluating face retouching networks as well." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.471, + 0.655, + 0.485 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.496, + 0.738, + 0.512 + ], + "angle": 0, + "content": "2.1. Face Retouching Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.519, + 0.905, + 0.791 + ], + "angle": 0, + "content": "Face retouching has been widely explored in the literature in two representative approaches. The first leverages traditional image filtering techniques for boosting visual quality while preserving skin textures and details, largely driven by various commercial applications such as Pixlr [20], MeituPic [13], and BeautyCam [12]. However, this approach relies heavily on manual adjustments with poor scalability. The second explores deep learning for face retouching. For example, AutoRetouch [21] adopts residual learning over a large-scale dataset (FFHQR) created from FFHQ [8]. ABPN [9] introduces a pyramid architecture to handle blemishes in high-resolution images. BPFRe [30] addresses limited training data via blemish-aware processing and progressive restoration. RetouchFormer [29] synthesizes high-quality facial images to reduce dependency on paired training data. Despite their impressive performance, these deep learning methods often struggle in maintaining global consistency and local detail preservation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.803, + 0.72, + 0.818 + ], + "angle": 0, + "content": "2.2. Face Retouching Data" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.903 + ], + "angle": 0, + "content": "Though face retouching has been studied for years, the research community still falls short in large-scale face retouching benchmarks. The lack of such benchmarks is largely attributed to the challenges in creating paired face images with and without blemishes. Several prior studies" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14723" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.087, + 0.907, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.316, + 0.907, + 0.331 + ], + "angle": 0, + "content": "Figure 2. Qualitative illustration of the created large-scale, high-quality, and diverse face retouching benchmark. Zoom-in for more details." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.357, + 0.486, + 0.54 + ], + "angle": 0, + "content": "attempted to tackle the challenges in different ways. For example, [10] presents an exemplar-based method that collects retouching data with laser-treated facial images. However, its applicability is limited to laser-induced effects only. AutoRetouch [21] introduces FFHQR, a large-scale retouching benchmark comprising 70,000 image pairs. However, many images in FFHQR do not have blemishes since the underlying FFHQ dataset was originally collected for face generation. We utilize diffusion models to synthesize paired face images with and without blemishes, leading to a comprehensive paired dataset that can be adopted for training and evaluating face retouching networks." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.554, + 0.45, + 0.572 + ], + "angle": 0, + "content": "3. High-Definition Retouching Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.58, + 0.486, + 0.748 + ], + "angle": 0, + "content": "Manual collection of paired retouching images is a challenging task that requires sophisticated control to ensure just one variable for each pair of face images, i.e., having or not having blemishes. In addition, collecting such large-scale and diverse paired images is laborious and time-consuming. We address this challenge by synthesizing paired face images with stable diffusion models, leading to the first large-scale, high-quality, and diverse benchmark that comprises 25,000 paired face images with diverse blemish types and facial attributes. Fig. 2 shows samples of synthesized image pairs with and without blemishes." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.758, + 0.318, + 0.773 + ], + "angle": 0, + "content": "3.1. Data Generation Process" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.781, + 0.485, + 0.903 + ], + "angle": 0, + "content": "We create the proposed face retouching benchmark in two stages, including the first stage that generates face images without blemishes and the second that generates the corresponding paired image with blemishes. We generate the face images without blemishes by leveraging three fine-tuned stable diffusion models. Specifically, we design 7,920 image-generation prompts that systematically permute combinations of facial attributes—including age," + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.355, + 0.909, + 0.449 + ], + "angle": 0, + "content": "
Blemish TypesNumber of Each BlemishNumber of Synthe-sized FaceNumber of Natural Face
HGFRMoles, Freckles, Acne7,92023,7601,240
" + }, + { + "type": "table_caption", + "bbox": [ + 0.555, + 0.458, + 0.864, + 0.473 + ], + "angle": 0, + "content": "Table 1. Details of the proposed benchmark HGFR." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.502, + 0.907, + 0.533 + ], + "angle": 0, + "content": "gender, aesthetics, hairstyle, skin tone, and detailed features (e.g., lips, nose, eyes, face shape, and eyebrows)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.533, + 0.906, + 0.594 + ], + "angle": 0, + "content": "The generation with such combinatorial facial attributes ensures that the generated face images exhibit a broad spectrum of diversity and rich facial representations while preserving a natural and unblemished appearance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.595, + 0.909, + 0.837 + ], + "angle": 0, + "content": "For each clean face image generated in the first stage, we generate a corresponding paired image with blemishes via a controlled modification process. Specifically, we leverage a state-of-the-art facial attribute segmentation model [19] to first partition the clean face image into semantically meaningful regions (e.g., cheeks, nose, and neck) while deliberately excluding key areas such as hair, eyes, eyebrows, and mouth to maintain critical identity cues. We then apply the pre-trained LoRA [6] to the isolated facial regions to synthesize blemishes such as moles, freckles, and acne. Such targeted blemish synthesis ensures that the paired images accurately reflect common facial imperfections without compromising the overall facial structure and identity, thereby balancing the local repairs and global consistency. More details on the image synthesis are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.663, + 0.863 + ], + "angle": 0, + "content": "3.2. Data Statistics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The proposed benchmark HGFR consists of 25,000 image pairs, including 23,760 synthesized image pairs and 1,240" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "14724" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.077, + 0.907, + 0.323 + ], + "angle": 0, + "content": "
Dataset Name# Images/PairsResolutionFace SourceTypeFace with Blemishes
PPR10K11,161 images4K-8KNaturalPortrait Retouching5%
RetouchingFFHQ>500K images1024×1024NaturalFace Recognition16%
EFrR--NaturalFace Retouching-
DigiFace-1M>1M images-SyntheticVarious poses, lighting, and backgrounds-
FRTDD--Restoration-basedTampering Detection-
FFHQ70,000 pairs1024×1024NaturalFace Generation16%
FFHQR70,000 pairs1024×1024NaturalFace Retouching16%
HGFR (Ours)25,000 pairs1024×1024Synthetic & NaturalFace Retouching83%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.333, + 0.907, + 0.363 + ], + "angle": 0, + "content": "Table 2. Summary of existing face retouching benchmarks. \\( {}^{\\prime } - \\) ’ Indicates that the information is unavailable. The percentages of face images with Blemishes are determined by the zero-shot CLIP classifier. More details are provided in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.388, + 0.482, + 0.479 + ], + "angle": 0, + "content": "natural ones. As detailed in Tab. 1, HGFR consists of three primary types of blemishes including moles, freckles, and acne—with 7,920 image pairs in each of the three categories. Such systematic design yields a comprehensive representation of facial blemishes across 23,760 synthetic images, augmented by a small set of natural faces." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.48, + 0.483, + 0.646 + ], + "angle": 0, + "content": "Tab. 2 compares HGFR with several existing retouching benchmark. HGFR has two unique features. First, it has superb diversity and includes both synthetic and natural face retouching images, enabling more robust and comprehensive face retouching evaluations. Second, HGFR contains a greater variety of blemish types. Based on the classification by zero-shot CLIP [16], up to \\(83\\%\\) images in HGFR contain blemishes while most existing benchmarks have less than \\(20\\%\\) image containing blemishes. The much larger number of blemish instances enables more robust training and evaluation of existing and future face retouching networks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.662, + 0.182, + 0.679 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.809 + ], + "angle": 0, + "content": "Fig. 3 illustrates our overall network architecture, which is composed of several key components. First, a Soft Mask Generation Branch (SMGB) automatically predicts masks to guide localized blemish repair. Next, the frequency selection and restoration (FSR) refines skin regions by leveraging both frequency and spatial domain information. Finally, the multi-resolution fusion (MRF) integrates multi-resolution outputs to reconstruct a high-quality, blemish-free portrait." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Given an input image \\( I_0 \\in \\mathbb{R}^{h \\times w \\times 3} \\), we first generate a Laplacian planes, \\( L_I = [I_0, I_1, \\dots, I_l] \\), and the corresponding high-frequency group \\( L_H = [H_0, H_1, \\dots, H_l] \\) derived from Laplacian pyramid decomposition [1] (with \\( l \\), the number of downsampling operations, set to 2 by default). The FSR is then applied to each \\( I_i \\in L_I \\) to produce" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.387, + 0.907, + 0.479 + ], + "angle": 0, + "content": "retouched outputs \\( R_{i} \\in \\mathbb{R}\\frac{h}{2^{i}}\\times \\frac{w}{2^{i}}\\times 3 \\). These multi-scale results are subsequently fused via MRF to reconstruct the final high-resolution portrait. Inspired by the Laplacian pyramid [1], MRF uses progressive upsampling combined with a Conv+LeakyReLU+Conv(CLC) module to refine details and ensure seamless integration of features across scales." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.48, + 0.909, + 0.694 + ], + "angle": 0, + "content": "The Soft Mask Generation Branch (SMGB) is designed to automatically predict a mask \\( M \\in \\mathbb{R}^{\\frac{h}{2^l} \\times \\frac{w}{2^l} \\times 3} \\) that highlights target regions for localized repair. We adopt a U-Net architecture with skip connections [18] to fuse low-level and high-level features, thereby enhancing segmentation accuracy. Although most datasets lack explicit ground truth masks \\( M_{gt} \\), these can be derived by computing the difference between the input image \\( I_l \\) and its corresponding ground truth, followed by thresholding. The SMGB contributes in two ways: first, the predicted mask \\( M_{pred} \\) enables the FSR to concentrate on defect regions; and second, joint training with SMGB enhances the network's overall contextual and semantic awareness, leading to more consistent retouching results." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.702, + 0.833, + 0.719 + ], + "angle": 0, + "content": "4.1. Frequency Selection and Restoration" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.724, + 0.907, + 0.83 + ], + "angle": 0, + "content": "To address Challenge (1) outlined in Section 1—achieving local repairs with global consistency and detail fidelity—we propose a context-aware frequency selection and restoration (FSR) framework. As shown in Fig. 3, FSR adopts a multi-task design that includes Frequency-Aware Dynamic Aggregation, Space Domain Projection, and Selective Frequency Domain-based FFN." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.837, + 0.658, + 0.851 + ], + "angle": 0, + "content": "4.1.1. Preliminaries." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.854, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Given a feature map \\(z \\in \\mathbb{R}^{H \\times W \\times C}\\), we apply the Discrete Fourier Transform (DFT) channel-wise to map spatial-domain features into the frequency domain. For a single" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14725" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.063, + 0.912, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.419, + 0.907, + 0.504 + ], + "angle": 0, + "content": "Figure 3. Overview of our proposed blemish repair framework. The input image is first decomposed into a Laplacian pyramid to obtain multi-scale representations and high-frequency components. A soft mask generation branch (SMGB) then automatically predicts masks to highlight defect regions for targeted repair. Subsequently, the frequency selection and restoration (FSR) module leverages both dynamic spectral filtering and spatial-domain projection—via learnable quantization matrices and self-attention mechanisms—to adaptively restore fine details and maintain global consistency. Finally, the multi-resolution fusion (MRF) progressively fuses multi-scale outputs through upsampling and convolutional refinement, yielding a high-quality, blemish-free portrait." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.53, + 0.383, + 0.543 + ], + "angle": 0, + "content": "channel, the Fourier transform is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.112, + 0.555, + 0.483, + 0.597 + ], + "angle": 0, + "content": "\\[\nF \\left(z _ {C}\\right) (u, v) = \\sum_ {h = 0} ^ {H - 1} \\sum_ {w = 0} ^ {W - 1} z _ {C} (h, w) e ^ {- j 2 \\pi \\left(\\frac {h u}{H} + \\frac {w v}{W}\\right)} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.609, + 0.484, + 0.655 + ], + "angle": 0, + "content": "The real and imaginary parts are then extracted, allowing us to compute the magnitude spectrum \\( A(z_{C})(u,v) \\) and phase spectrum \\( \\Phi (z_C)(u,v) \\) via:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.667, + 0.483, + 0.7 + ], + "angle": 0, + "content": "\\[\nA \\left(z _ {C}\\right) (u, v) = \\sqrt {\\Re \\left(F \\left(z _ {C}\\right) (u , v)\\right) ^ {2} + \\Im \\left(F \\left(z _ {C}\\right) (u , v)\\right) ^ {2}} \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.711, + 0.483, + 0.746 + ], + "angle": 0, + "content": "\\[\n\\Phi (z _ {C}) (u, v) = \\arctan \\left[ \\frac {\\Im (F (z _ {C}) (u , v))}{\\Re (F (z _ {C}) (u , v))} \\right] \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.752, + 0.484, + 0.782 + ], + "angle": 0, + "content": "By sequentially applying Eqs. (1) to (3) to all channels, we obtain the full-channel magnitude and phase spectra." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.791, + 0.421, + 0.807 + ], + "angle": 0, + "content": "4.1.2. Frequency-Aware Dynamic Aggregation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In many transformer-based methods [2, 5, 27], global selfattention can overlook localized defect characteristics. To address this, our dynamic spectral filter operates in the frequency domain to target specific imperfections (e.g., moles, freckles, acne), as shown in Fig. 4 left. We aggregate the magnitude and phase information using pointwise convolu" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.531, + 0.549, + 0.542 + ], + "angle": 0, + "content": "tion:" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.548, + 0.905, + 0.573 + ], + "angle": 0, + "content": "\\[\nF ^ {*} (z) (u, v) = \\mathcal {G} \\left(\\operatorname {C o n v} _ {1} ^ {1} \\left(F \\left(z _ {C} \\cdot M _ {\\text {p r e d}}\\right) (u, v)\\right)\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.579, + 0.905, + 0.655 + ], + "angle": 0, + "content": "where \\(\\mathrm{F}(z_C\\cdot M_{pred})(u,v)\\in\\) \\(\\{A(z_C\\cdot M_{pred})(u,v),P(z_C\\cdot M_{pred})(u,v)\\}\\) and \\(\\mathcal{G}(\\cdot)\\) is the GEGLU activation function [22]. To adaptively weight the frequency components, we introduce learnable quantization matrices \\(W_{A}\\) and \\(W_{\\Phi}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.661, + 0.905, + 0.686 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} A ^ {\\prime} (z) (u, v) = W _ {A} \\odot A (z) (u, v), \\\\ I ^ {\\prime} (z) (u, v) = W _ {I} \\odot I (z) (u, v). \\end{array} \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.681, + 0.82, + 0.698 + ], + "angle": 0, + "content": "\\[\n\\Phi^ {\\prime} (z) (u, v) = W _ {\\Phi} \\odot \\Phi (z) (u, v).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.705, + 0.871, + 0.72 + ], + "angle": 0, + "content": "The updated frequency feature is then recomposed as:" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.727, + 0.905, + 0.746 + ], + "angle": 0, + "content": "\\[\nF ^ {\\prime} (z) (u, v) = A ^ {\\prime} (z) (u, v) e ^ {j \\Phi^ {\\prime} (z) (u, v)}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.753, + 0.905, + 0.782 + ], + "angle": 0, + "content": "This filtered component is combined with the original frequency features via a residual connection:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.79, + 0.905, + 0.808 + ], + "angle": 0, + "content": "\\[\n\\tilde {F} (z) (u, v) = F ^ {\\prime} (z) (u, v) + F (z) (u, v). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.815, + 0.905, + 0.859 + ], + "angle": 0, + "content": "Reconstruction into the spatial domain is achieved by extracting the real and imaginary parts and applying the inverse Fourier Transform:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.865, + 0.905, + 0.89 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\Re} (z) (u, v) = \\tilde {A} (z) (u, v) \\cdot \\cos \\tilde {\\Phi} (z) (u, v), \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.887, + 0.847, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathfrak {S}} (z) (u, v) = \\tilde {A} (z) (u, v) \\cdot \\sin \\tilde {\\Phi} (z) (u, v),\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14726" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.485, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.185, + 0.473, + 0.2 + ], + "angle": 0, + "content": "Figure 4. Details of frequency selection and restoration (FSR)" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.238, + 0.483, + 0.264 + ], + "angle": 0, + "content": "\\[\n\\tilde {z} = \\mathcal {F} ^ {- 1} \\left(\\tilde {\\Re} (z) (u, v), \\tilde {\\Im} (z) (u, v)\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.27, + 0.484, + 0.316 + ], + "angle": 0, + "content": "The Fourier transform and its inverse are implemented via DFT and IDFT algorithms [3, 4, 34]. We denote the entire FADA process as \\(\\mathcal{F}\\mathcal{A}(\\cdot)\\)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.324, + 0.318, + 0.339 + ], + "angle": 0, + "content": "4.1.3. Space Domain Projection." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.343, + 0.484, + 0.435 + ], + "angle": 0, + "content": "After frequency-domain processing, we further refine features in the spatial domain, as shown in Fig. 4 right. For an input feature \\( z \\in \\mathbb{R}^{H \\times W \\times C} \\), layer normalization \\( \\mathcal{L}(\\cdot) \\) [11] is applied to yield \\( \\mathcal{L}(z) \\). This normalized feature is projected into query \\( (Q) \\), key \\( (K) \\), and value \\( (V) \\) matrices via independent frequency-domain projections:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.445, + 0.487, + 0.476 + ], + "angle": 0, + "content": "\\[\nQ = \\mathcal {F} \\mathcal {A} _ {Q} (\\mathcal {L} (\\tilde {z})), K = \\mathcal {F} \\mathcal {A} _ {K} (\\mathcal {L} (\\tilde {z})), V = \\mathcal {F} \\mathcal {A} _ {V} (\\mathcal {L} (\\tilde {z})), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.477, + 0.484, + 0.507 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}\\mathcal{A}_Q\\), \\(\\mathcal{F}\\mathcal{A}_K\\), and \\(\\mathcal{F}\\mathcal{A}_V\\) denote learnable projection functions. The output of FSR is then given by:" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.518, + 0.483, + 0.552 + ], + "angle": 0, + "content": "\\[\nz ^ {\\prime} = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {\\Lambda}} + B\\right) V, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.563, + 0.484, + 0.594 + ], + "angle": 0, + "content": "with \\(\\Lambda\\) representing the channel dimension and \\(B\\) the positional bias. Finally, a residual connection is used:" + }, + { + "type": "equation", + "bbox": [ + 0.243, + 0.606, + 0.482, + 0.622 + ], + "angle": 0, + "content": "\\[\nz ^ {*} = z ^ {\\prime} + z. \\tag {12}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.643, + 0.418, + 0.659 + ], + "angle": 0, + "content": "4.1.4. Selective Frequency Domain-based FFN" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.663, + 0.484, + 0.752 + ], + "angle": 0, + "content": "To further enhance feature representation, as shown in Fig. 5 we design a selective frequency feed-forward network (SFFN) that adaptively retains critical frequency components for latent blemish removal. Inspired by the JPEG compression algorithm, SFFN employs a learnable quantization matrix:" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.762, + 0.483, + 0.844 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} z _ {1} = F \\left(P \\left(C o n v _ {1} ^ {1} (\\mathcal {L} (z)))\\right)\\right), \\\\ z _ {2} = F ^ {- 1} \\left(W \\odot z _ {1}\\right), \\tag {13} \\\\ z _ {\\text {o u t}} = \\mathcal {G} \\left(P ^ {- 1} \\left(z _ {2}\\right)\\right) + z ^ {*}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.855, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\( P(\\cdot) \\) and \\( P^{-1}(\\cdot) \\) represent patch unfolding and folding operations. The output \\( z_{\\mathrm{out}} \\) is then fed into subsequent layers." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.064, + 0.905, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.202, + 0.887, + 0.217 + ], + "angle": 0, + "content": "Figure 5. Details of selective frequency domain-based FFN" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.229, + 0.614, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.304, + 0.585, + 0.314 + ], + "angle": 0, + "content": "(a) Input" + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.229, + 0.711, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.633, + 0.304, + 0.693, + 0.313 + ], + "angle": 0, + "content": "(b) Result on" + }, + { + "type": "image_caption", + "bbox": [ + 0.627, + 0.313, + 0.696, + 0.323 + ], + "angle": 0, + "content": "Low Resolution" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.229, + 0.807, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.73, + 0.304, + 0.788, + 0.313 + ], + "angle": 0, + "content": "(c) Result on" + }, + { + "type": "image_caption", + "bbox": [ + 0.717, + 0.313, + 0.8, + 0.323 + ], + "angle": 0, + "content": "Original Resolution" + }, + { + "type": "image", + "bbox": [ + 0.808, + 0.229, + 0.905, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.839, + 0.304, + 0.873, + 0.313 + ], + "angle": 0, + "content": "(d) Our" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.334, + 0.906, + 0.391 + ], + "angle": 0, + "content": "Figure 6. Visual Comparison of MRF at Different Resolutions: (a) Input image; (b) Low Resolution image, with large blemishes removed; (c) Original Resolution image, with fine blemishes removed; (d) Our method, achieving the best retouching." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.413, + 0.738, + 0.428 + ], + "angle": 0, + "content": "4.2. Multi-Resolution Fusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.435, + 0.907, + 0.572 + ], + "angle": 0, + "content": "We observe that while high-resolution features are essential for restoring fine details, low-resolution features are more effective for repairing large-area blemishes. The differences can be observed in Fig. 6. To combine such complementary advantages, we propose MRF to fuse multi-resolution features to produce a detail-rich, blemish-free portrait. Specifically, MRF incorporates high-frequency components from the original image to compensate for the information loss during downsampling. The fusion process is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.578, + 0.905, + 0.624 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} O _ {l - 1} = \\operatorname {C o n v} _ {3} ^ {3} \\left(\\sigma \\left(\\operatorname {C o n v} _ {3} ^ {3} \\left(\\operatorname {C a t} \\left(u p \\left(O _ {l}\\right), H _ {l - 1}\\right)\\right)\\right)\\right) \\tag {14} \\\\ + u p (R _ {l - 1}), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.631, + 0.907, + 0.812 + ], + "angle": 0, + "content": ", where \\( O_{l} \\), \\( O_{l-1} \\) denotes the fusion result, and \\( O_{l} = R_{l} \\). \\( R_{l} \\), \\( R_{l-1} \\) corresponds to the different resolutions of the blemish removal face images generated by Laplace image \\( I_{l} \\), \\( I_{l-1} \\) through the face blemish removal network, respectively. \\( H_{l-1} \\) is the high-frequency component of Laplace image \\( I_{l-1} \\), \\( up(\\cdot) \\) denotes bilinear interpolation, \\( \\mathrm{Cat}(\\cdot) \\) channel-wise concatenation. The CLC [15] is implemented through sequential \\( Conv_{3}^{3} \\) layers with leaky ReLU activation, which is a convolution with a \\( 3 \\times 3 \\) kernel and stride 3, and \\( \\sigma \\) is a leaky ReLU activation (negative slope 0.2); when \\( l = 1 \\), \\( O_{l-1} = O_{0} \\) i.e. the final blemish-removed face image." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.824, + 0.646, + 0.841 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.722, + 0.866 + ], + "angle": 0, + "content": "5.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Datasets. We conduct extensive experiments over two benchmarks to validate our face retouching network. The" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14727" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.908, + 0.224 + ], + "angle": 0, + "content": "
DatasetsFFHQRHGFRFFHQR + HGFR
MetricsPSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
AutoRetouch29.380.9180.076630.680.8920.082331.060.9240.0744
pix2pixHD41.180.9800.013342.370.9650.016742.680.9820.0132
ABPN44.260.9920.017647.020.9860.018947.650.9920.0178
BPFRe45.130.9930.009247.280.9890.013448.520.9930.0128
RetouchFormer45.240.9930.009047.610.9900.012848.780.9930.0105
Ours45.480.9930.008248.020.9910.012549.150.9940.0094
" + }, + { + "type": "table_caption", + "bbox": [ + 0.102, + 0.233, + 0.895, + 0.248 + ], + "angle": 0, + "content": "Table 3. Quantitative comparison of our method with existing methods over three training setups ( \\( \\downarrow \\) Lower is better; \\( \\uparrow \\) Higher is better)" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.262, + 0.907, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.186, + 0.534, + 0.809, + 0.549 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison of our method with several state-of-the-art methods on Internet images." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.574, + 0.483, + 0.681 + ], + "angle": 0, + "content": "first is our proposed synthetic benchmark HGFR as detailed in Section 3. The second is the benchmark FFHQR [8], a large-scale face retouching benchmark consisting of 70,000 high-definition facial images paired with their retouched versions. We randomly split HGFR into 20,000 images for training, 2,500 for validation, and 2,500 for testing. For FFHQR, we follow the data splitting in [8]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.685, + 0.483, + 0.82 + ], + "angle": 0, + "content": "Implementation Details. Our model and baselines are implemented in Python 3.8 using PyTorch 2.3.1 and trained on 4 NVIDIA RTX 4090 GPUs. We adopt the Adam optimizer with \\(\\beta_{1} = 0.9\\), \\(\\beta_{2} = 0.999\\), and set the initial learning rate to 0.0005 for a batch size of 8. After 100 epochs, the learning rate is reduced by a factor of 10. The default number of downsampling steps (\\(l\\)) is set to 2. Training takes approximately 15 hours on HGFR, 35 hours on FFHQR, and 55 hours on FFHQR+HGFR." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.825, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Evaluation Metrics. We evaluate the retouching with three evaluation metrics including Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM)[28], and Learned Perceptual Image Patch Similarity (LPIPS)[33]. The three metrics have been widely" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.575, + 0.791, + 0.59 + ], + "angle": 0, + "content": "adopted for image generation evaluations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.604, + 0.871, + 0.62 + ], + "angle": 0, + "content": "5.2. Quantitative and Qualitative Experiments" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.628, + 0.907, + 0.839 + ], + "angle": 0, + "content": "We compare our face retouching method with several existing work including Pix2PixHD [26], AutoRetouch [21], ABPN [9], BPFRe [30], and Retouch-Former [29]. Pix2PixHD is an image-to-image translation network and the rest are specifically designed for facial retouching. We train three retouching models by using the HGFR training images, the FFHQR training images, and a combination of the training images of the two datasets (i.e., FFHQR+HGFR). Tab. 3 shows experimental results. It can be observed that the proposed method performs clearly the best consistently across all datasets and evaluation metrics. The superior face retouching performance is largely attributed to our designed FSR and MRF, more details to be discussed in the ensuing ablation experiments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We also conduct qualitative comparisons to further demonstrate the superiority of our designed face retouching network as illustrated in Fig. 7. It can be observed that AutoRetouch tends to introduce global noises and fails to" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14728" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.287, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.17, + 0.19, + 0.211, + 0.2 + ], + "angle": 0, + "content": "(a) Input" + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.089, + 0.482, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.319, + 0.19, + 0.449, + 0.2 + ], + "angle": 0, + "content": "(b) Retouching with FFHQR" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.201, + 0.288, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.13, + 0.302, + 0.252, + 0.312 + ], + "angle": 0, + "content": "(c) Retouching with HGFR" + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.201, + 0.482, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.302, + 0.467, + 0.312 + ], + "angle": 0, + "content": "(d) Retouching with FFHQR+HGFR" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.325, + 0.483, + 0.368 + ], + "angle": 0, + "content": "Figure 8. Face retouching with models trained by using FFHQR training data alone in (b), HGFR training data alone in (c) and the combination of HGFR and FFHQR training data in (d)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.395, + 0.484, + 0.501 + ], + "angle": 0, + "content": "clearly delineate retouched regions while Pix2PixHD suffers from noticeable color shifts. ABPN, RetouchFormer, and BPFRe can deliver competitive retouching effects, but their outputs fall short in detail preservation and refinement. As a comparison, our method balances the local blemish repair and global facial consistency and demonstrates superior retouching experience." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.502, + 0.484, + 0.699 + ], + "angle": 0, + "content": "The experiments also validate the effectiveness of our benchmark from two perspectives: 1) training effective retouching network; and 2) augmenting existing benchmark for better retouching performance. Fig. 8 shows qualitative experiments. We can see that the model trained using HGFR alone performs clearly better than the one trained using FFHQR alone. The clear better retouching effect is largely because our benchmark is specifically synthesized with large-scale, high-quality, and diverse paired images. In addition, we can observe that the model using HGRR+FFHQR performs the best, demonstrating that our created benchmark can augment and complement existing benchmarks effectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.711, + 0.245, + 0.727 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.734, + 0.483, + 0.869 + ], + "angle": 0, + "content": "We conduct ablation experiments to examine how the proposed FSR and MRF contribute to the face retouching as shown in Tab. 4. The baseline FFN only achieves 42.93 dB in PSNR and 0.967 in SSIM. Including FSR improves the two metrics to 45.79 dB and 0.980, indicating that our frequency-aware design substantially improves detail preservation. Further including MRF on top of FSR yields the best performance (47.81 dB/0.990), indicating the complementary role of MRF via multi-scale feature refinement." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We also conduct qualitative ablation experiments. As Fig. 9 shows, FSR outperforms the baseline clearly in bal" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.712, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.59, + 0.193, + 0.639, + 0.205 + ], + "angle": 0, + "content": "(a) Input" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.089, + 0.905, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.774, + 0.193, + 0.841, + 0.205 + ], + "angle": 0, + "content": "(b) Baseline" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.205, + 0.712, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.566, + 0.309, + 0.663, + 0.32 + ], + "angle": 0, + "content": "(c) Baseline+FSR" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.205, + 0.905, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.741, + 0.309, + 0.874, + 0.32 + ], + "angle": 0, + "content": "(d) Baseline+FSR+MRF" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.332, + 0.906, + 0.388 + ], + "angle": 0, + "content": "Figure 9. Qualitative ablation studies on the proposed FSR and MRF. FSR substantially improves the detail preservation and global consistency as compared with the Baseline, and MRF further repairs residual blemishes with enhanced retouching effects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.418, + 0.906, + 0.48 + ], + "angle": 0, + "content": "ancing detail preservation and global consistency, and MRF further enhances the face retouching by removing the residual blemishes. Please refer to the appendix for more detailed ablation experiments." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.495, + 0.906, + 0.563 + ], + "angle": 0, + "content": "
MRFFSRFFNPSNRs/SSIMs
FFN(baseline)✓42.93/0.967
FSR✓45.79/0.980
MRF+FSR✓✓47.81/0.990
" + }, + { + "type": "table_caption", + "bbox": [ + 0.54, + 0.573, + 0.878, + 0.586 + ], + "angle": 0, + "content": "Table 4. Ablation Studies on FSR and MRF over HGFR." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.6, + 0.634, + 0.616 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.627, + 0.907, + 0.793 + ], + "angle": 0, + "content": "This paper presents a high-definition face retouching benchmark and a novel face retouching framework that can effectively balance local detail repair with global consistency. The benchmark comprises 25,000 high-quality paired face retouching images that are generated by diffusion models, possessing superb diversity for training and evaluation face retouching networks. The framework integrates frequency selection and sestoration with multi-resolution fusion, demonstrating superior retouching performance. Extensive experiments confirm the effectiveness of our created benchmark and designed retouching framework." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.901 + ], + "angle": 0, + "content": "ACKNOWLEDGEMENT This work was supported in part by the National Key Research and Development Program Project of China [grant no. 2024YFC3306902], in part by the National Natural Science Foundation of China [grant no. U24A20242], and in part by the Zhejiang Provincial Natural Science Foundation [grant no. LDT23F02024F02]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14729" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.482, + 0.157 + ], + "angle": 0, + "content": "[1] Peter J Burt and Edward H Adelson. The laplacian pyramid as a compact image code. In Readings in computer vision, pages 671-679. Elsevier, 1987. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.483, + 0.241 + ], + "angle": 0, + "content": "[2] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ArXiv, abs/2010.11929, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.243, + 0.482, + 0.284 + ], + "angle": 0, + "content": "[3] Matteo Frigo and Steven G Johnson. Fftw: An adaptive software architecture for the fft. In ICASSP, pages 1381-1384. IEEE, 1998. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.287, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[4] Junming Hou, Qi Cao, Ran Ran, Che Liu, Junling Li, and Liang-jian Deng. Bidomain modeling paradigm for pan-sharpening. In ACM MM, pages 347-357, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.33, + 0.482, + 0.384 + ], + "angle": 0, + "content": "[5] Jie Hu, Li Shen, Samuel Albanie, Gang Sun, and Enhua Wu. Squeeze-and-excitation networks. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7132-7141, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.387, + 0.482, + 0.442 + ], + "angle": 0, + "content": "[6] J. Edward Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. ArXiv, abs/2106.09685, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.444, + 0.482, + 0.5 + ], + "angle": 0, + "content": "[7] Le Jiang, Yan Huang, Lianxin Xie, Wen Xue, Cheng Liu, Si Wu, and Hau-San Wong. Hunting blemishes: Language-guided high-fidelity face retouching transformer with limited paired data. In ACM MM, pages 5102-5111, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.501, + 0.482, + 0.543 + ], + "angle": 0, + "content": "[8] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, pages 4401-4410, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.545, + 0.482, + 0.613 + ], + "angle": 0, + "content": "[9] Biwen Lei, Xiefan Guo, Hongyu Yang, Miaomiao Cui, Xuansong Xie, and Di Huang. Abpn: adaptive blend pyramid network for real-time local retouching of ultra high-resolution photo. In CVPR, pages 2108-2117, 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.615, + 0.482, + 0.67 + ], + "angle": 0, + "content": "[10] Tsung-Ying Lin, Yu-Ting Tsai, Tsung-Shian Huang, Wen-Chieh Lin, and Jung-Hong Chuang. Exemplar-based freckle retouching and skin tone adjustment. Computers & Graphics, 78:54-63, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.672, + 0.482, + 0.728 + ], + "angle": 0, + "content": "[11] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.73, + 0.481, + 0.757 + ], + "angle": 0, + "content": "[12] Meitu Inc. Beautycam. https://meiyan.meipai.com/, 2013.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.758, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[13] Meitu Inc. Meitupic. http://xiuxiu.meitu.com/, 2013.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.788, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[14] Dongbo Min, Sunghwan Choi, Jiangbo Lu, Bumsub Ham, Kwanghoon Sohn, and Minh N Do. Fast global image smoothing based on weighted least squares. IEEE TIP, 23 (12):5638-5653, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.845, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[15] Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. CoRR, abs/1511.06434, 2015. 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.176 + ], + "angle": 0, + "content": "[16] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.906, + 0.232 + ], + "angle": 0, + "content": "[17] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.234, + 0.906, + 0.317 + ], + "angle": 0, + "content": "[18] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.319, + 0.906, + 0.387 + ], + "angle": 0, + "content": "[19] Ioannis Sarridis, Christos Koutlis, Symeon Papadopoulos, and Christos Diou. Facex: Understanding face attribute classifiers through summary model explanations. In Proceedings of the 2024 International Conference on Multimedia Retrieval, pages 758-766, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.389, + 0.905, + 0.416 + ], + "angle": 0, + "content": "[20] Ola Sevandersson. Pixlr. https://www pixlr.com/, 2008.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.908, + 0.473 + ], + "angle": 0, + "content": "[21] Alireza Shafaei, James J Little, and Mark Schmidt. Autoretouch: Automatic professional face retouching. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 990-998, 2021. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.475, + 0.906, + 0.503 + ], + "angle": 0, + "content": "[22] Noam Shazeer. Glu variants improve transformer. arXiv preprint arXiv:2002.05202, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.504, + 0.906, + 0.545 + ], + "angle": 0, + "content": "[23] Wanchao Su, Can Wang, Chen Liu, Fangzhou Han, Hongbo Fu, and Jing Liao. Styleretoucher: Generalized portrait image retouching with gan priors. IEEE TVCG, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.906, + 0.586 + ], + "angle": 0, + "content": "[24] Carlo Tomasi and Roberto Manduchi. Bilateral filtering for gray and color images. In ICCV, pages 839-846. IEEE, 1998. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.589, + 0.906, + 0.645 + ], + "angle": 0, + "content": "[25] Sudha Velusamy, Rishubh Parihar, Raviprasad Kini, and Aniket Rege. Fabsoften: Face beautification via dynamic skin smoothing, guided feathering, and texture restoration. In CVPRW, pages 530-531, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.647, + 0.906, + 0.702 + ], + "angle": 0, + "content": "[26] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In CVPR, pages 8798-8807, 2018. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.704, + 0.906, + 0.758 + ], + "angle": 0, + "content": "[27] X. Wang, Ross B. Girshick, Abhinav Kumar Gupta, and Kaiming He. Non-local neural networks. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7794-7803, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.906, + 0.801 + ], + "angle": 0, + "content": "[28] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 13(4):600-612, 2004. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.906, + 0.87 + ], + "angle": 0, + "content": "[29] Xue Wen, Lianxin Xie, Le Jiang, Tianyi Chen, Si Wu, Cheng Liu, and Hau-San Wong. Retouchformer: semi-supervised high-quality face retouching transformer with prior-based selective self-attention. In AAAI, pages 5903–5911, 2024. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.906, + 0.902 + ], + "angle": 0, + "content": "[30] Lianxin Xie, Wen Xue, Zhen Xu, Si Wu, Zhiwen Yu, and Hau San Wong. Blemish-aware and progressive face re" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14730" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "touching with limited paired data. In CVPR, pages 5599-5608, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.163 + ], + "angle": 0, + "content": "[31] Li Xu, Cewu Lu, Yi Xu, and Jiaya Jia. Image smoothing via 10 gradient minimization. In Proceedings of the 2011 SIGGRAPH Asia conference, pages 1-12, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.193 + ], + "angle": 0, + "content": "[32] Qi Zhang, Li Xu, and Jiaya Jia. 100+ times faster weighted median filter (wmf). In CVPR, pages 2830-2837, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.194, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[33] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.306 + ], + "angle": 0, + "content": "[34] Man Zhou, Keyu Yan, Xueyang Fu, Aiping Liu, and Chengjun Xie. Pan-guided band-aware multi-spectral feature enhancement for pan-sharpening. IEEE Transactions on Computational Imaging, 9:238-249, 2023. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "14731" + } + ] +] \ No newline at end of file diff --git a/2025/Face Retouching with Diffusion Data Generation and Spectral Restorement/2a12f446-c130-4e36-bc58-dba590a7c6d7_origin.pdf b/2025/Face Retouching with Diffusion Data Generation and Spectral Restorement/2a12f446-c130-4e36-bc58-dba590a7c6d7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..10a0f67f6427d911960905b49e884fa232a09386 --- /dev/null +++ b/2025/Face Retouching with Diffusion Data Generation and Spectral Restorement/2a12f446-c130-4e36-bc58-dba590a7c6d7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:965738d07adf169bd508d669eee5c9dbde032972dd65e37383aa1dc0fb4076f9 +size 4037253 diff --git a/2025/Face Retouching with Diffusion Data Generation and Spectral Restorement/full.md b/2025/Face Retouching with Diffusion Data Generation and Spectral Restorement/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d8d8d4cd05a58a652289e15c6aea5dc114946792 --- /dev/null +++ b/2025/Face Retouching with Diffusion Data Generation and Spectral Restorement/full.md @@ -0,0 +1,351 @@ +# Face Retouching with Diffusion Data Generation and Spectral Restored + +Zhidan Xu $^{1}$ Xiaoqin Zhang $^{1*}$ Shijian Lu $^{2*}$ $^{1}$ Wenzhou University + $^{2}$ Nanyang Technological University + +zhidan.xu@stu.wzu.edu.cn, xqzhang@wzu.edu.cn, Shijian.Lu@ntu.edu.sg + +![](images/666c64ac7910fd7a5e8132c36811d290b9d13d0d4b2d8cc27a5aa6b4dbb3610f.jpg) +Figure 1. The proposed face retouching technique can repair local facial regions and remove blemishes with superb detail preservation and global consistency in facial colour, facial texture, etc. Zoom-in for more details. + +![](images/79ee83db86b7150308208498944780079d30097a43ecfaf3419a8f6d610fd802.jpg) + +![](images/ccb564ef2ab8c699c606b82659cc4e26a704ffff62b988db2d143d4d94e5645c.jpg) + +![](images/9fd608ab178b3d17144a6816ca858ed8697be679c0c75cad40bd7ea7521e20ed.jpg) + +![](images/c53ec5dd3cc3090d92c1638ee577bc11af9012b23f6033112dc6c9d6ca16212f.jpg) + +![](images/ffc2a90a66bad763ec078a3a22d0a5adc06b8cfd5316b106503d4b9459079f4c.jpg) + +# Abstract + +Face retouching has achieved impressive performance largely driven by its wide range of applications in various real-world tasks. However, most existing works often encounters a dilemma between global consistency and local detail preservation, partially due to the lack of large-scale and high-quality training data. We address the face retouching challenge from two perspectives. First, we create a large-scale face retouching benchmark to mitigate the data scarcity issue. The benchmark comprises 25,000 pairs of high-quality facial images (before and after face retouching) that contain a variety of facial attributes and blemish types such as acne and moles. Second, we design a novel framework that introduces frequency selection and restoration (FSR) and multi-resolution fusion (MRF) that leverages frequency-aware dynamic aggregation and spatial-frequency filtering to achieve global consistency and local detail preservation concurrently. Inspired by the principle + +of JPEG compression, FSR introduces frequency-domain quantization with spatial projections to learn enhanced feature representations. MRF fuses multi-resolution features via laplacian pyramid fusion, removing large-area blemishes and preserving local fine details effectively. Extensive experiments over multiple benchmarks show that the proposed framework outperforms the state-of-the-art quantitatively and qualitatively. The created benchmark also provides valuable data for training and evaluating both existing and future face retouching networks. + +# 1. Introduction + +Face retouching, which aims to beautify face images by removing various blemishes and smoothing facial skin, has attracted increasing interest due to its applications in various real-world tasks in social media, virtual avatar creation, etc. Manual retouching has been widely practiced in the literature, relying on expert retouchers who balance blemish removal and the global uniformity of skin tones meticulously. However, the process of manual retouching is usually time-consuming, highly skill-dependent, and short of + +scalability. + +Various image processing techniques have been explored for automating the face retouching process without sacrificing retouching quality. Specifically, face retouching has been widely studied in traditional image processing and computer vision research, where different filtering and rule-based methods [14, 24, 31, 32] have been adopted to tackle the face retouching challenge. In addition, dedicated techniques have also been designed for various mobile applications [12, 13, 20] which rely on image filtering as well. Despite being lightweight, these traditional techniques share two constraints: 1) they rely heavily on fine-tuning of handcrafted parameters and cannot generalize well; 2) they tend to over/under-smooth skin textures, eliminate critical details like pores, and introduce unnatural whitening effects. + +Leveraging the great success of deep learning in image enhancement, several deep neural networks have been designed and achieved promising retouching performance [7, 9, 21, 23, 25, 29, 30]. However, these methods often struggle to balance local detail repairs and global consistency in skin colour, skin texture, etc., partially due to the lack of large-scale paired training images. For example, ABPN [9] improves retouching with pyramid structures and generative models but falls short in managing fine blemishes and retaining details. StyleRetouch [23] achieves impressive retouching but often encounters scalability and efficiency issues while handling high-resolution images. + +We tackle the face retouching challenges from two perspectives. The first is to create a benchmark that can effectively train and evaluate face retouching networks. Several benchmarks such as FFHQ [8] have been adopted for retouching studies but they are not created for the task of face retouching and therefore lack comprehensive representations of real-world blemish types and facial attributes. In addition, these benchmarks comprise real face images with excessive identity features which often cause privacy concerns. We create HGFR, a high-definition face retouching benchmark that has three desirable features. First, HGFR is large-scale containing up to 25,000 pairs of face images with and without blemishes. Second, HGFR is featured by superb image quality with an image resolution of 1024 by 1024. Third, HGFR has superior diversity including a large portion of synthetic images that are generated by three finetuned diffusion models [17]. The adoption of synthetic images mitigate the privacy concerns, obviates the costly and laborious image collection processes, and allows generating large-scale blemished facial images that cover a wide spectrum of imperfections and facial attributes. + +In addition, we design a face retouching network that can balance the local repairs and global consistency effectively. The network achieves face retouching with two designs. The first is frequency selection and restoration (FSR) that performs multi-resolution retouching by har + +nessing both frequency and spatial signals for addressing localized blemishes without affecting the overall image structure. The second is multi-resolution fusion (MRF) that fuses high-resolution details with low-resolution context, ensuring that large-area blemishes can be repaired while maintaining global coherence. These two designs achieve fine-grained local corrections with broad structural consistency concurrently, delivering superior face retouching performance across multiple benchmarks. + +The contributions of this work can be summarized in three major aspects. First, we create the first large-scale, high-quality and diverse face retouching benchmark that comprise 25,000 pairs of high-definition face images that are generated by diffusion models. The benchmark captures a variety of blemishes and facial attributes which can be employed for both training and evaluating face retouching networks. Second, we design a novel face retouching network that introduces frequency selection and restoration and multi-resolution fusion for effective face retouching. Third, extensive experiments over multiple benchmarks show that our face retouching network achieves superior performance as compared with the state-of-the-art. The proposed benchmark can be effectively adopted for training and evaluating face retouching networks as well. + +# 2. Related Work + +# 2.1. Face Retouching Studies + +Face retouching has been widely explored in the literature in two representative approaches. The first leverages traditional image filtering techniques for boosting visual quality while preserving skin textures and details, largely driven by various commercial applications such as Pixlr [20], MeituPic [13], and BeautyCam [12]. However, this approach relies heavily on manual adjustments with poor scalability. The second explores deep learning for face retouching. For example, AutoRetouch [21] adopts residual learning over a large-scale dataset (FFHQR) created from FFHQ [8]. ABPN [9] introduces a pyramid architecture to handle blemishes in high-resolution images. BPFRe [30] addresses limited training data via blemish-aware processing and progressive restoration. RetouchFormer [29] synthesizes high-quality facial images to reduce dependency on paired training data. Despite their impressive performance, these deep learning methods often struggle in maintaining global consistency and local detail preservation. + +# 2.2. Face Retouching Data + +Though face retouching has been studied for years, the research community still falls short in large-scale face retouching benchmarks. The lack of such benchmarks is largely attributed to the challenges in creating paired face images with and without blemishes. Several prior studies + +![](images/ac58da3de16893fa921c48da0d49c721ff65bd7f751b7dc963032b959c89f154.jpg) +Figure 2. Qualitative illustration of the created large-scale, high-quality, and diverse face retouching benchmark. Zoom-in for more details. + +attempted to tackle the challenges in different ways. For example, [10] presents an exemplar-based method that collects retouching data with laser-treated facial images. However, its applicability is limited to laser-induced effects only. AutoRetouch [21] introduces FFHQR, a large-scale retouching benchmark comprising 70,000 image pairs. However, many images in FFHQR do not have blemishes since the underlying FFHQ dataset was originally collected for face generation. We utilize diffusion models to synthesize paired face images with and without blemishes, leading to a comprehensive paired dataset that can be adopted for training and evaluating face retouching networks. + +# 3. High-Definition Retouching Benchmark + +Manual collection of paired retouching images is a challenging task that requires sophisticated control to ensure just one variable for each pair of face images, i.e., having or not having blemishes. In addition, collecting such large-scale and diverse paired images is laborious and time-consuming. We address this challenge by synthesizing paired face images with stable diffusion models, leading to the first large-scale, high-quality, and diverse benchmark that comprises 25,000 paired face images with diverse blemish types and facial attributes. Fig. 2 shows samples of synthesized image pairs with and without blemishes. + +# 3.1. Data Generation Process + +We create the proposed face retouching benchmark in two stages, including the first stage that generates face images without blemishes and the second that generates the corresponding paired image with blemishes. We generate the face images without blemishes by leveraging three fine-tuned stable diffusion models. Specifically, we design 7,920 image-generation prompts that systematically permute combinations of facial attributes—including age, + +
Blemish TypesNumber of Each BlemishNumber of Synthe-sized FaceNumber of Natural Face
HGFRMoles, Freckles, Acne7,92023,7601,240
+ +Table 1. Details of the proposed benchmark HGFR. + +gender, aesthetics, hairstyle, skin tone, and detailed features (e.g., lips, nose, eyes, face shape, and eyebrows). + +The generation with such combinatorial facial attributes ensures that the generated face images exhibit a broad spectrum of diversity and rich facial representations while preserving a natural and unblemished appearance. + +For each clean face image generated in the first stage, we generate a corresponding paired image with blemishes via a controlled modification process. Specifically, we leverage a state-of-the-art facial attribute segmentation model [19] to first partition the clean face image into semantically meaningful regions (e.g., cheeks, nose, and neck) while deliberately excluding key areas such as hair, eyes, eyebrows, and mouth to maintain critical identity cues. We then apply the pre-trained LoRA [6] to the isolated facial regions to synthesize blemishes such as moles, freckles, and acne. Such targeted blemish synthesis ensures that the paired images accurately reflect common facial imperfections without compromising the overall facial structure and identity, thereby balancing the local repairs and global consistency. More details on the image synthesis are provided in the supplementary material. + +# 3.2. Data Statistics + +The proposed benchmark HGFR consists of 25,000 image pairs, including 23,760 synthesized image pairs and 1,240 + +
Dataset Name# Images/PairsResolutionFace SourceTypeFace with Blemishes
PPR10K11,161 images4K-8KNaturalPortrait Retouching5%
RetouchingFFHQ>500K images1024×1024NaturalFace Recognition16%
EFrR--NaturalFace Retouching-
DigiFace-1M>1M images-SyntheticVarious poses, lighting, and backgrounds-
FRTDD--Restoration-basedTampering Detection-
FFHQ70,000 pairs1024×1024NaturalFace Generation16%
FFHQR70,000 pairs1024×1024NaturalFace Retouching16%
HGFR (Ours)25,000 pairs1024×1024Synthetic & NaturalFace Retouching83%
+ +Table 2. Summary of existing face retouching benchmarks. ${}^{\prime } -$ ’ Indicates that the information is unavailable. The percentages of face images with Blemishes are determined by the zero-shot CLIP classifier. More details are provided in the supplementary materials. + +natural ones. As detailed in Tab. 1, HGFR consists of three primary types of blemishes including moles, freckles, and acne—with 7,920 image pairs in each of the three categories. Such systematic design yields a comprehensive representation of facial blemishes across 23,760 synthetic images, augmented by a small set of natural faces. + +Tab. 2 compares HGFR with several existing retouching benchmark. HGFR has two unique features. First, it has superb diversity and includes both synthetic and natural face retouching images, enabling more robust and comprehensive face retouching evaluations. Second, HGFR contains a greater variety of blemish types. Based on the classification by zero-shot CLIP [16], up to $83\%$ images in HGFR contain blemishes while most existing benchmarks have less than $20\%$ image containing blemishes. The much larger number of blemish instances enables more robust training and evaluation of existing and future face retouching networks. + +# 4. Method + +Fig. 3 illustrates our overall network architecture, which is composed of several key components. First, a Soft Mask Generation Branch (SMGB) automatically predicts masks to guide localized blemish repair. Next, the frequency selection and restoration (FSR) refines skin regions by leveraging both frequency and spatial domain information. Finally, the multi-resolution fusion (MRF) integrates multi-resolution outputs to reconstruct a high-quality, blemish-free portrait. + +Given an input image $I_0 \in \mathbb{R}^{h \times w \times 3}$ , we first generate a Laplacian planes, $L_I = [I_0, I_1, \dots, I_l]$ , and the corresponding high-frequency group $L_H = [H_0, H_1, \dots, H_l]$ derived from Laplacian pyramid decomposition [1] (with $l$ , the number of downsampling operations, set to 2 by default). The FSR is then applied to each $I_i \in L_I$ to produce + +retouched outputs $R_{i} \in \mathbb{R}\frac{h}{2^{i}}\times \frac{w}{2^{i}}\times 3$ . These multi-scale results are subsequently fused via MRF to reconstruct the final high-resolution portrait. Inspired by the Laplacian pyramid [1], MRF uses progressive upsampling combined with a Conv+LeakyReLU+Conv(CLC) module to refine details and ensure seamless integration of features across scales. + +The Soft Mask Generation Branch (SMGB) is designed to automatically predict a mask $M \in \mathbb{R}^{\frac{h}{2^l} \times \frac{w}{2^l} \times 3}$ that highlights target regions for localized repair. We adopt a U-Net architecture with skip connections [18] to fuse low-level and high-level features, thereby enhancing segmentation accuracy. Although most datasets lack explicit ground truth masks $M_{gt}$ , these can be derived by computing the difference between the input image $I_l$ and its corresponding ground truth, followed by thresholding. The SMGB contributes in two ways: first, the predicted mask $M_{pred}$ enables the FSR to concentrate on defect regions; and second, joint training with SMGB enhances the network's overall contextual and semantic awareness, leading to more consistent retouching results. + +# 4.1. Frequency Selection and Restoration + +To address Challenge (1) outlined in Section 1—achieving local repairs with global consistency and detail fidelity—we propose a context-aware frequency selection and restoration (FSR) framework. As shown in Fig. 3, FSR adopts a multi-task design that includes Frequency-Aware Dynamic Aggregation, Space Domain Projection, and Selective Frequency Domain-based FFN. + +# 4.1.1. Preliminaries. + +Given a feature map $z \in \mathbb{R}^{H \times W \times C}$ , we apply the Discrete Fourier Transform (DFT) channel-wise to map spatial-domain features into the frequency domain. For a single + +![](images/d426f5a4a2d62c2fb338f988adcf5b91ca3071eadb76c27cd8442afa12e01900.jpg) +Figure 3. Overview of our proposed blemish repair framework. The input image is first decomposed into a Laplacian pyramid to obtain multi-scale representations and high-frequency components. A soft mask generation branch (SMGB) then automatically predicts masks to highlight defect regions for targeted repair. Subsequently, the frequency selection and restoration (FSR) module leverages both dynamic spectral filtering and spatial-domain projection—via learnable quantization matrices and self-attention mechanisms—to adaptively restore fine details and maintain global consistency. Finally, the multi-resolution fusion (MRF) progressively fuses multi-scale outputs through upsampling and convolutional refinement, yielding a high-quality, blemish-free portrait. + +channel, the Fourier transform is defined as: + +$$ +F \left(z _ {C}\right) (u, v) = \sum_ {h = 0} ^ {H - 1} \sum_ {w = 0} ^ {W - 1} z _ {C} (h, w) e ^ {- j 2 \pi \left(\frac {h u}{H} + \frac {w v}{W}\right)} \tag {1} +$$ + +The real and imaginary parts are then extracted, allowing us to compute the magnitude spectrum $A(z_{C})(u,v)$ and phase spectrum $\Phi (z_C)(u,v)$ via: + +$$ +A \left(z _ {C}\right) (u, v) = \sqrt {\Re \left(F \left(z _ {C}\right) (u , v)\right) ^ {2} + \Im \left(F \left(z _ {C}\right) (u , v)\right) ^ {2}} \tag {2} +$$ + +$$ +\Phi (z _ {C}) (u, v) = \arctan \left[ \frac {\Im (F (z _ {C}) (u , v))}{\Re (F (z _ {C}) (u , v))} \right] \tag {3} +$$ + +By sequentially applying Eqs. (1) to (3) to all channels, we obtain the full-channel magnitude and phase spectra. + +# 4.1.2. Frequency-Aware Dynamic Aggregation. + +In many transformer-based methods [2, 5, 27], global selfattention can overlook localized defect characteristics. To address this, our dynamic spectral filter operates in the frequency domain to target specific imperfections (e.g., moles, freckles, acne), as shown in Fig. 4 left. We aggregate the magnitude and phase information using pointwise convolu + +tion: + +$$ +F ^ {*} (z) (u, v) = \mathcal {G} \left(\operatorname {C o n v} _ {1} ^ {1} \left(F \left(z _ {C} \cdot M _ {\text {p r e d}}\right) (u, v)\right)\right) \tag {4} +$$ + +where $\mathrm{F}(z_C\cdot M_{pred})(u,v)\in$ $\{A(z_C\cdot M_{pred})(u,v),P(z_C\cdot M_{pred})(u,v)\}$ and $\mathcal{G}(\cdot)$ is the GEGLU activation function [22]. To adaptively weight the frequency components, we introduce learnable quantization matrices $W_{A}$ and $W_{\Phi}$ : + +$$ +\begin{array}{l} A ^ {\prime} (z) (u, v) = W _ {A} \odot A (z) (u, v), \\ I ^ {\prime} (z) (u, v) = W _ {I} \odot I (z) (u, v). \end{array} \tag {5} +$$ + +$$ +\Phi^ {\prime} (z) (u, v) = W _ {\Phi} \odot \Phi (z) (u, v). +$$ + +The updated frequency feature is then recomposed as: + +$$ +F ^ {\prime} (z) (u, v) = A ^ {\prime} (z) (u, v) e ^ {j \Phi^ {\prime} (z) (u, v)}. \tag {6} +$$ + +This filtered component is combined with the original frequency features via a residual connection: + +$$ +\tilde {F} (z) (u, v) = F ^ {\prime} (z) (u, v) + F (z) (u, v). \tag {7} +$$ + +Reconstruction into the spatial domain is achieved by extracting the real and imaginary parts and applying the inverse Fourier Transform: + +$$ +\tilde {\Re} (z) (u, v) = \tilde {A} (z) (u, v) \cdot \cos \tilde {\Phi} (z) (u, v), \tag {8} +$$ + +$$ +\tilde {\mathfrak {S}} (z) (u, v) = \tilde {A} (z) (u, v) \cdot \sin \tilde {\Phi} (z) (u, v), +$$ + +![](images/819a22ee551f1e3fe871f58e5db2d21248e23b830432e0f51b7168c370b6dc6e.jpg) +Figure 4. Details of frequency selection and restoration (FSR) + +$$ +\tilde {z} = \mathcal {F} ^ {- 1} \left(\tilde {\Re} (z) (u, v), \tilde {\Im} (z) (u, v)\right). \tag {9} +$$ + +The Fourier transform and its inverse are implemented via DFT and IDFT algorithms [3, 4, 34]. We denote the entire FADA process as $\mathcal{F}\mathcal{A}(\cdot)$ . + +# 4.1.3. Space Domain Projection. + +After frequency-domain processing, we further refine features in the spatial domain, as shown in Fig. 4 right. For an input feature $z \in \mathbb{R}^{H \times W \times C}$ , layer normalization $\mathcal{L}(\cdot)$ [11] is applied to yield $\mathcal{L}(z)$ . This normalized feature is projected into query $(Q)$ , key $(K)$ , and value $(V)$ matrices via independent frequency-domain projections: + +$$ +Q = \mathcal {F} \mathcal {A} _ {Q} (\mathcal {L} (\tilde {z})), K = \mathcal {F} \mathcal {A} _ {K} (\mathcal {L} (\tilde {z})), V = \mathcal {F} \mathcal {A} _ {V} (\mathcal {L} (\tilde {z})), \tag {10} +$$ + +where $\mathcal{F}\mathcal{A}_Q$ , $\mathcal{F}\mathcal{A}_K$ , and $\mathcal{F}\mathcal{A}_V$ denote learnable projection functions. The output of FSR is then given by: + +$$ +z ^ {\prime} = \operatorname {s o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {\Lambda}} + B\right) V, \tag {11} +$$ + +with $\Lambda$ representing the channel dimension and $B$ the positional bias. Finally, a residual connection is used: + +$$ +z ^ {*} = z ^ {\prime} + z. \tag {12} +$$ + +# 4.1.4. Selective Frequency Domain-based FFN + +To further enhance feature representation, as shown in Fig. 5 we design a selective frequency feed-forward network (SFFN) that adaptively retains critical frequency components for latent blemish removal. Inspired by the JPEG compression algorithm, SFFN employs a learnable quantization matrix: + +$$ +\begin{array}{l} z _ {1} = F \left(P \left(C o n v _ {1} ^ {1} (\mathcal {L} (z)))\right)\right), \\ z _ {2} = F ^ {- 1} \left(W \odot z _ {1}\right), \tag {13} \\ z _ {\text {o u t}} = \mathcal {G} \left(P ^ {- 1} \left(z _ {2}\right)\right) + z ^ {*}, \\ \end{array} +$$ + +where $P(\cdot)$ and $P^{-1}(\cdot)$ represent patch unfolding and folding operations. The output $z_{\mathrm{out}}$ is then fed into subsequent layers. + +![](images/71de102d6168d7cf02648eba3f2329e64cd15e0e610fbff7b7e635cf53361b5b.jpg) +Figure 5. Details of selective frequency domain-based FFN + +![](images/6bd23f8512ed513e7538d18de33844475b45e6309dd000089f923edefecc18f8.jpg) +(a) Input +Figure 6. Visual Comparison of MRF at Different Resolutions: (a) Input image; (b) Low Resolution image, with large blemishes removed; (c) Original Resolution image, with fine blemishes removed; (d) Our method, achieving the best retouching. + +![](images/d632f7a4455f60dffa4c0a3fa3785ff5aa1d969160602204548f117ba703cd2d.jpg) +(b) Result on +Low Resolution + +![](images/5c9a2ab99ac84dee4030e670208a86ab85603fc33ff67e5ffba81620f9e0dae7.jpg) +(c) Result on +Original Resolution + +![](images/6b09e06d585728865577e5449d313237669699faf4798e91de5303c9e044466a.jpg) +(d) Our + +# 4.2. Multi-Resolution Fusion + +We observe that while high-resolution features are essential for restoring fine details, low-resolution features are more effective for repairing large-area blemishes. The differences can be observed in Fig. 6. To combine such complementary advantages, we propose MRF to fuse multi-resolution features to produce a detail-rich, blemish-free portrait. Specifically, MRF incorporates high-frequency components from the original image to compensate for the information loss during downsampling. The fusion process is defined as: + +$$ +\begin{array}{l} O _ {l - 1} = \operatorname {C o n v} _ {3} ^ {3} \left(\sigma \left(\operatorname {C o n v} _ {3} ^ {3} \left(\operatorname {C a t} \left(u p \left(O _ {l}\right), H _ {l - 1}\right)\right)\right)\right) \tag {14} \\ + u p (R _ {l - 1}), \\ \end{array} +$$ + +, where $O_{l}$ , $O_{l-1}$ denotes the fusion result, and $O_{l} = R_{l}$ . $R_{l}$ , $R_{l-1}$ corresponds to the different resolutions of the blemish removal face images generated by Laplace image $I_{l}$ , $I_{l-1}$ through the face blemish removal network, respectively. $H_{l-1}$ is the high-frequency component of Laplace image $I_{l-1}$ , $up(\cdot)$ denotes bilinear interpolation, $\mathrm{Cat}(\cdot)$ channel-wise concatenation. The CLC [15] is implemented through sequential $Conv_{3}^{3}$ layers with leaky ReLU activation, which is a convolution with a $3 \times 3$ kernel and stride 3, and $\sigma$ is a leaky ReLU activation (negative slope 0.2); when $l = 1$ , $O_{l-1} = O_{0}$ i.e. the final blemish-removed face image. + +# 5. Experiments + +# 5.1. Experimental Settings + +Datasets. We conduct extensive experiments over two benchmarks to validate our face retouching network. The + +
DatasetsFFHQRHGFRFFHQR + HGFR
MetricsPSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
AutoRetouch29.380.9180.076630.680.8920.082331.060.9240.0744
pix2pixHD41.180.9800.013342.370.9650.016742.680.9820.0132
ABPN44.260.9920.017647.020.9860.018947.650.9920.0178
BPFRe45.130.9930.009247.280.9890.013448.520.9930.0128
RetouchFormer45.240.9930.009047.610.9900.012848.780.9930.0105
Ours45.480.9930.008248.020.9910.012549.150.9940.0094
+ +Table 3. Quantitative comparison of our method with existing methods over three training setups ( $\downarrow$ Lower is better; $\uparrow$ Higher is better) + +![](images/ebb5f8f7675142b521bc957f45c4af8619ffeb23a0f734a09dfc8cbbc7de89d8.jpg) +Figure 7. Qualitative comparison of our method with several state-of-the-art methods on Internet images. + +first is our proposed synthetic benchmark HGFR as detailed in Section 3. The second is the benchmark FFHQR [8], a large-scale face retouching benchmark consisting of 70,000 high-definition facial images paired with their retouched versions. We randomly split HGFR into 20,000 images for training, 2,500 for validation, and 2,500 for testing. For FFHQR, we follow the data splitting in [8]. + +Implementation Details. Our model and baselines are implemented in Python 3.8 using PyTorch 2.3.1 and trained on 4 NVIDIA RTX 4090 GPUs. We adopt the Adam optimizer with $\beta_{1} = 0.9$ , $\beta_{2} = 0.999$ , and set the initial learning rate to 0.0005 for a batch size of 8. After 100 epochs, the learning rate is reduced by a factor of 10. The default number of downsampling steps ( $l$ ) is set to 2. Training takes approximately 15 hours on HGFR, 35 hours on FFHQR, and 55 hours on FFHQR+HGFR. + +Evaluation Metrics. We evaluate the retouching with three evaluation metrics including Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM)[28], and Learned Perceptual Image Patch Similarity (LPIPS)[33]. The three metrics have been widely + +adopted for image generation evaluations. + +# 5.2. Quantitative and Qualitative Experiments + +We compare our face retouching method with several existing work including Pix2PixHD [26], AutoRetouch [21], ABPN [9], BPFRe [30], and Retouch-Former [29]. Pix2PixHD is an image-to-image translation network and the rest are specifically designed for facial retouching. We train three retouching models by using the HGFR training images, the FFHQR training images, and a combination of the training images of the two datasets (i.e., FFHQR+HGFR). Tab. 3 shows experimental results. It can be observed that the proposed method performs clearly the best consistently across all datasets and evaluation metrics. The superior face retouching performance is largely attributed to our designed FSR and MRF, more details to be discussed in the ensuing ablation experiments. + +We also conduct qualitative comparisons to further demonstrate the superiority of our designed face retouching network as illustrated in Fig. 7. It can be observed that AutoRetouch tends to introduce global noises and fails to + +![](images/f5ecec7d3205b1cd010a4b6f00b0f50cfbfa28e60bd0522fbe6498137ed08438.jpg) +(a) Input + +![](images/8b7d8eb23a34d43cfcd2d655d4de4c5c9e87bdc1bea243b0e2dd2dd2daef9440.jpg) +(b) Retouching with FFHQR + +![](images/faf6a1349bbd690c121c45918d335791461ade172e753edb5a688df734a23da6.jpg) +(c) Retouching with HGFR +Figure 8. Face retouching with models trained by using FFHQR training data alone in (b), HGFR training data alone in (c) and the combination of HGFR and FFHQR training data in (d). + +![](images/fd957c41b234c399878da7a8d426f8814b599a69ae3f41615c30b24114930f6d.jpg) +(d) Retouching with FFHQR+HGFR + +clearly delineate retouched regions while Pix2PixHD suffers from noticeable color shifts. ABPN, RetouchFormer, and BPFRe can deliver competitive retouching effects, but their outputs fall short in detail preservation and refinement. As a comparison, our method balances the local blemish repair and global facial consistency and demonstrates superior retouching experience. + +The experiments also validate the effectiveness of our benchmark from two perspectives: 1) training effective retouching network; and 2) augmenting existing benchmark for better retouching performance. Fig. 8 shows qualitative experiments. We can see that the model trained using HGFR alone performs clearly better than the one trained using FFHQR alone. The clear better retouching effect is largely because our benchmark is specifically synthesized with large-scale, high-quality, and diverse paired images. In addition, we can observe that the model using HGRR+FFHQR performs the best, demonstrating that our created benchmark can augment and complement existing benchmarks effectively. + +# 5.3. Ablation Study + +We conduct ablation experiments to examine how the proposed FSR and MRF contribute to the face retouching as shown in Tab. 4. The baseline FFN only achieves 42.93 dB in PSNR and 0.967 in SSIM. Including FSR improves the two metrics to 45.79 dB and 0.980, indicating that our frequency-aware design substantially improves detail preservation. Further including MRF on top of FSR yields the best performance (47.81 dB/0.990), indicating the complementary role of MRF via multi-scale feature refinement. + +We also conduct qualitative ablation experiments. As Fig. 9 shows, FSR outperforms the baseline clearly in bal + +![](images/76c3bee8d22bad114f848b418e81923d90e6fe1a9e1d2abc1889706abe0e2fad.jpg) +(a) Input + +![](images/a1d9074a402a4d60267611d041b062c39ec64afb105b2f4cf857871045139d1d.jpg) +(b) Baseline + +![](images/34c967b81367a3820f2916c8c7ff2b30cf7b45b08243c53d7838a0962daf4143.jpg) +(c) Baseline+FSR +Figure 9. Qualitative ablation studies on the proposed FSR and MRF. FSR substantially improves the detail preservation and global consistency as compared with the Baseline, and MRF further repairs residual blemishes with enhanced retouching effects. + +![](images/4e68a1c32716355ecbe6c26c7dab80d83799d4fe4a12254b25c1a6386220173b.jpg) +(d) Baseline+FSR+MRF + +ancing detail preservation and global consistency, and MRF further enhances the face retouching by removing the residual blemishes. Please refer to the appendix for more detailed ablation experiments. + +
MRFFSRFFNPSNRs/SSIMs
FFN(baseline)✓42.93/0.967
FSR✓45.79/0.980
MRF+FSR✓✓47.81/0.990
+ +Table 4. Ablation Studies on FSR and MRF over HGFR. + +# 6. Conclusion + +This paper presents a high-definition face retouching benchmark and a novel face retouching framework that can effectively balance local detail repair with global consistency. The benchmark comprises 25,000 high-quality paired face retouching images that are generated by diffusion models, possessing superb diversity for training and evaluation face retouching networks. The framework integrates frequency selection and sestoration with multi-resolution fusion, demonstrating superior retouching performance. Extensive experiments confirm the effectiveness of our created benchmark and designed retouching framework. + +ACKNOWLEDGEMENT This work was supported in part by the National Key Research and Development Program Project of China [grant no. 2024YFC3306902], in part by the National Natural Science Foundation of China [grant no. U24A20242], and in part by the Zhejiang Provincial Natural Science Foundation [grant no. LDT23F02024F02]. + +# References + +[1] Peter J Burt and Edward H Adelson. The laplacian pyramid as a compact image code. In Readings in computer vision, pages 671-679. Elsevier, 1987. 4 +[2] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ArXiv, abs/2010.11929, 2020. 5 +[3] Matteo Frigo and Steven G Johnson. Fftw: An adaptive software architecture for the fft. In ICASSP, pages 1381-1384. IEEE, 1998. 6 +[4] Junming Hou, Qi Cao, Ran Ran, Che Liu, Junling Li, and Liang-jian Deng. Bidomain modeling paradigm for pan-sharpening. In ACM MM, pages 347-357, 2023. 6 +[5] Jie Hu, Li Shen, Samuel Albanie, Gang Sun, and Enhua Wu. Squeeze-and-excitation networks. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7132-7141, 2017. 5 +[6] J. Edward Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. ArXiv, abs/2106.09685, 2021. 3 +[7] Le Jiang, Yan Huang, Lianxin Xie, Wen Xue, Cheng Liu, Si Wu, and Hau-San Wong. Hunting blemishes: Language-guided high-fidelity face retouching transformer with limited paired data. In ACM MM, pages 5102-5111, 2024. 2 +[8] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, pages 4401-4410, 2019. 2, 7 +[9] Biwen Lei, Xiefan Guo, Hongyu Yang, Miaomiao Cui, Xuansong Xie, and Di Huang. Abpn: adaptive blend pyramid network for real-time local retouching of ultra high-resolution photo. In CVPR, pages 2108-2117, 2022. 1, 2, 7 +[10] Tsung-Ying Lin, Yu-Ting Tsai, Tsung-Shian Huang, Wen-Chieh Lin, and Jung-Hong Chuang. Exemplar-based freckle retouching and skin tone adjustment. Computers & Graphics, 78:54-63, 2019. 3 +[11] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 6 +[12] Meitu Inc. Beautycam. https://meiyan.meipai.com/, 2013.2 +[13] Meitu Inc. Meitupic. http://xiuxiu.meitu.com/, 2013.2 +[14] Dongbo Min, Sunghwan Choi, Jiangbo Lu, Bumsub Ham, Kwanghoon Sohn, and Minh N Do. Fast global image smoothing based on weighted least squares. IEEE TIP, 23 (12):5638-5653, 2014. 2 +[15] Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. CoRR, abs/1511.06434, 2015. 6 + +[16] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 4 +[17] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2 +[18] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 4 +[19] Ioannis Sarridis, Christos Koutlis, Symeon Papadopoulos, and Christos Diou. Facex: Understanding face attribute classifiers through summary model explanations. In Proceedings of the 2024 International Conference on Multimedia Retrieval, pages 758-766, 2024. 3 +[20] Ola Sevandersson. Pixlr. https://www pixlr.com/, 2008.2 +[21] Alireza Shafaei, James J Little, and Mark Schmidt. Autoretouch: Automatic professional face retouching. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 990-998, 2021. 1, 2, 3, 7 +[22] Noam Shazeer. Glu variants improve transformer. arXiv preprint arXiv:2002.05202, 2020. 5 +[23] Wanchao Su, Can Wang, Chen Liu, Fangzhou Han, Hongbo Fu, and Jing Liao. Styleretoucher: Generalized portrait image retouching with gan priors. IEEE TVCG, 2024. 2 +[24] Carlo Tomasi and Roberto Manduchi. Bilateral filtering for gray and color images. In ICCV, pages 839-846. IEEE, 1998. 2 +[25] Sudha Velusamy, Rishubh Parihar, Raviprasad Kini, and Aniket Rege. Fabsoften: Face beautification via dynamic skin smoothing, guided feathering, and texture restoration. In CVPRW, pages 530-531, 2020. 2 +[26] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In CVPR, pages 8798-8807, 2018. 1, 7 +[27] X. Wang, Ross B. Girshick, Abhinav Kumar Gupta, and Kaiming He. Non-local neural networks. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7794-7803, 2017. 5 +[28] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 13(4):600-612, 2004. 7 +[29] Xue Wen, Lianxin Xie, Le Jiang, Tianyi Chen, Si Wu, Cheng Liu, and Hau-San Wong. Retouchformer: semi-supervised high-quality face retouching transformer with prior-based selective self-attention. In AAAI, pages 5903–5911, 2024. 1, 2, 7 +[30] Lianxin Xie, Wen Xue, Zhen Xu, Si Wu, Zhiwen Yu, and Hau San Wong. Blemish-aware and progressive face re + +touching with limited paired data. In CVPR, pages 5599-5608, 2023. 2, 7 +[31] Li Xu, Cewu Lu, Yi Xu, and Jiaya Jia. Image smoothing via 10 gradient minimization. In Proceedings of the 2011 SIGGRAPH Asia conference, pages 1-12, 2011. 2 +[32] Qi Zhang, Li Xu, and Jiaya Jia. 100+ times faster weighted median filter (wmf). In CVPR, pages 2830-2837, 2014. 2 +[33] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 7 +[34] Man Zhou, Keyu Yan, Xueyang Fu, Aiping Liu, and Chengjun Xie. Pan-guided band-aware multi-spectral feature enhancement for pan-sharpening. IEEE Transactions on Computational Imaging, 9:238-249, 2023. 6 \ No newline at end of file diff --git a/2025/Face Retouching with Diffusion Data Generation and Spectral Restorement/images.zip b/2025/Face Retouching with Diffusion Data Generation and Spectral Restorement/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c13bfeda22ebc841f981a781d0220e4bddc27992 --- /dev/null +++ b/2025/Face Retouching with Diffusion Data Generation and Spectral Restorement/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:522245d0f935060ce05a8fa4e6c1e51ebb3e9dd67bd0619781bc24bf32d79388 +size 810929 diff --git a/2025/Face Retouching with Diffusion Data Generation and Spectral Restorement/layout.json b/2025/Face Retouching with Diffusion Data Generation and Spectral Restorement/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..14f64ad5a265546011f99962e418d39587539624 --- /dev/null +++ b/2025/Face Retouching with Diffusion Data Generation and Spectral Restorement/layout.json @@ -0,0 +1,8322 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 72, + 103, + 538, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 103, + 538, + 121 + ], + "spans": [ + { + "bbox": [ + 72, + 103, + 538, + 121 + ], + "type": "text", + "content": "Face Retouching with Diffusion Data Generation and Spectral Restored" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 184, + 142, + 428, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 142, + 428, + 186 + ], + "spans": [ + { + "bbox": [ + 184, + 142, + 428, + 186 + ], + "type": "text", + "content": "Zhidan Xu" + }, + { + "bbox": [ + 184, + 142, + 428, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 184, + 142, + 428, + 186 + ], + "type": "text", + "content": " Xiaoqin Zhang" + }, + { + "bbox": [ + 184, + 142, + 428, + 186 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 184, + 142, + 428, + 186 + ], + "type": "text", + "content": " Shijian Lu" + }, + { + "bbox": [ + 184, + 142, + 428, + 186 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 184, + 142, + 428, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 184, + 142, + 428, + 186 + ], + "type": "text", + "content": "Wenzhou University \n" + }, + { + "bbox": [ + 184, + 142, + 428, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 184, + 142, + 428, + 186 + ], + "type": "text", + "content": "Nanyang Technological University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 121, + 187, + 485, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 187, + 485, + 200 + ], + "spans": [ + { + "bbox": [ + 121, + 187, + 485, + 200 + ], + "type": "text", + "content": "zhidan.xu@stu.wzu.edu.cn, xqzhang@wzu.edu.cn, Shijian.Lu@ntu.edu.sg" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 58, + 228, + 138, + 417 + ], + "blocks": [ + { + "bbox": [ + 58, + 228, + 138, + 417 + ], + "lines": [ + { + "bbox": [ + 58, + 228, + 138, + 417 + ], + "spans": [ + { + "bbox": [ + 58, + 228, + 138, + 417 + ], + "type": "image", + "image_path": "666c64ac7910fd7a5e8132c36811d290b9d13d0d4b2d8cc27a5aa6b4dbb3610f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 420, + 555, + 444 + ], + "lines": [ + { + "bbox": [ + 55, + 420, + 555, + 444 + ], + "spans": [ + { + "bbox": [ + 55, + 420, + 555, + 444 + ], + "type": "text", + "content": "Figure 1. The proposed face retouching technique can repair local facial regions and remove blemishes with superb detail preservation and global consistency in facial colour, facial texture, etc. Zoom-in for more details." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 140, + 228, + 221, + 418 + ], + "blocks": [ + { + "bbox": [ + 140, + 228, + 221, + 418 + ], + "lines": [ + { + "bbox": [ + 140, + 228, + 221, + 418 + ], + "spans": [ + { + "bbox": [ + 140, + 228, + 221, + 418 + ], + "type": "image", + "image_path": "79ee83db86b7150308208498944780079d30097a43ecfaf3419a8f6d610fd802.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 223, + 228, + 304, + 418 + ], + "blocks": [ + { + "bbox": [ + 223, + 228, + 304, + 418 + ], + "lines": [ + { + "bbox": [ + 223, + 228, + 304, + 418 + ], + "spans": [ + { + "bbox": [ + 223, + 228, + 304, + 418 + ], + "type": "image", + "image_path": "ccb564ef2ab8c699c606b82659cc4e26a704ffff62b988db2d143d4d94e5645c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 306, + 228, + 387, + 418 + ], + "blocks": [ + { + "bbox": [ + 306, + 228, + 387, + 418 + ], + "lines": [ + { + "bbox": [ + 306, + 228, + 387, + 418 + ], + "spans": [ + { + "bbox": [ + 306, + 228, + 387, + 418 + ], + "type": "image", + "image_path": "9fd608ab178b3d17144a6816ca858ed8697be679c0c75cad40bd7ea7521e20ed.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 388, + 229, + 470, + 418 + ], + "blocks": [ + { + "bbox": [ + 388, + 229, + 470, + 418 + ], + "lines": [ + { + "bbox": [ + 388, + 229, + 470, + 418 + ], + "spans": [ + { + "bbox": [ + 388, + 229, + 470, + 418 + ], + "type": "image", + "image_path": "c53ec5dd3cc3090d92c1638ee577bc11af9012b23f6033112dc6c9d6ca16212f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 471, + 229, + 553, + 418 + ], + "blocks": [ + { + "bbox": [ + 471, + 229, + 553, + 418 + ], + "lines": [ + { + "bbox": [ + 471, + 229, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 471, + 229, + 553, + 418 + ], + "type": "image", + "image_path": "ffc2a90a66bad763ec078a3a22d0a5adc06b8cfd5316b106503d4b9459079f4c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 151, + 464, + 200, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 464, + 200, + 477 + ], + "spans": [ + { + "bbox": [ + 151, + 464, + 200, + 477 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 491, + 297, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 491, + 297, + 694 + ], + "spans": [ + { + "bbox": [ + 53, + 491, + 297, + 694 + ], + "type": "text", + "content": "Face retouching has achieved impressive performance largely driven by its wide range of applications in various real-world tasks. However, most existing works often encounters a dilemma between global consistency and local detail preservation, partially due to the lack of large-scale and high-quality training data. We address the face retouching challenge from two perspectives. First, we create a large-scale face retouching benchmark to mitigate the data scarcity issue. The benchmark comprises 25,000 pairs of high-quality facial images (before and after face retouching) that contain a variety of facial attributes and blemish types such as acne and moles. Second, we design a novel framework that introduces frequency selection and restoration (FSR) and multi-resolution fusion (MRF) that leverages frequency-aware dynamic aggregation and spatial-frequency filtering to achieve global consistency and local detail preservation concurrently. Inspired by the principle" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 454, + 555, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 555, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 555, + 574 + ], + "type": "text", + "content": "of JPEG compression, FSR introduces frequency-domain quantization with spatial projections to learn enhanced feature representations. MRF fuses multi-resolution features via laplacian pyramid fusion, removing large-area blemishes and preserving local fine details effectively. Extensive experiments over multiple benchmarks show that the proposed framework outperforms the state-of-the-art quantitatively and qualitatively. The created benchmark also provides valuable data for training and evaluating both existing and future face retouching networks." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 586, + 395, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 586, + 395, + 597 + ], + "spans": [ + { + "bbox": [ + 314, + 586, + 395, + 597 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 605, + 556, + 715 + ], + "type": "text", + "content": "Face retouching, which aims to beautify face images by removing various blemishes and smoothing facial skin, has attracted increasing interest due to its applications in various real-world tasks in social media, virtual avatar creation, etc. Manual retouching has been widely practiced in the literature, relying on expert retouchers who balance blemish removal and the global uniformity of skin tones meticulously. However, the process of manual retouching is usually time-consuming, highly skill-dependent, and short of" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 703, + 149, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 703, + 149, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 703, + 149, + 713 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "14722" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 101, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 101, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 101, + 83 + ], + "type": "text", + "content": "scalability." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 86, + 294, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 86, + 294, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 86, + 294, + 253 + ], + "type": "text", + "content": "Various image processing techniques have been explored for automating the face retouching process without sacrificing retouching quality. Specifically, face retouching has been widely studied in traditional image processing and computer vision research, where different filtering and rule-based methods [14, 24, 31, 32] have been adopted to tackle the face retouching challenge. In addition, dedicated techniques have also been designed for various mobile applications [12, 13, 20] which rely on image filtering as well. Despite being lightweight, these traditional techniques share two constraints: 1) they rely heavily on fine-tuning of handcrafted parameters and cannot generalize well; 2) they tend to over/under-smooth skin textures, eliminate critical details like pores, and introduce unnatural whitening effects." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 255, + 295, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 295, + 399 + ], + "type": "text", + "content": "Leveraging the great success of deep learning in image enhancement, several deep neural networks have been designed and achieved promising retouching performance [7, 9, 21, 23, 25, 29, 30]. However, these methods often struggle to balance local detail repairs and global consistency in skin colour, skin texture, etc., partially due to the lack of large-scale paired training images. For example, ABPN [9] improves retouching with pyramid structures and generative models but falls short in managing fine blemishes and retaining details. StyleRetouch [23] achieves impressive retouching but often encounters scalability and efficiency issues while handling high-resolution images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 401, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 295, + 652 + ], + "type": "text", + "content": "We tackle the face retouching challenges from two perspectives. The first is to create a benchmark that can effectively train and evaluate face retouching networks. Several benchmarks such as FFHQ [8] have been adopted for retouching studies but they are not created for the task of face retouching and therefore lack comprehensive representations of real-world blemish types and facial attributes. In addition, these benchmarks comprise real face images with excessive identity features which often cause privacy concerns. We create HGFR, a high-definition face retouching benchmark that has three desirable features. First, HGFR is large-scale containing up to 25,000 pairs of face images with and without blemishes. Second, HGFR is featured by superb image quality with an image resolution of 1024 by 1024. Third, HGFR has superior diversity including a large portion of synthetic images that are generated by three finetuned diffusion models [17]. The adoption of synthetic images mitigate the privacy concerns, obviates the costly and laborious image collection processes, and allows generating large-scale blemished facial images that cover a wide spectrum of imperfections and facial attributes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 653, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 715 + ], + "type": "text", + "content": "In addition, we design a face retouching network that can balance the local repairs and global consistency effectively. The network achieves face retouching with two designs. The first is frequency selection and restoration (FSR) that performs multi-resolution retouching by har" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": "nessing both frequency and spatial signals for addressing localized blemishes without affecting the overall image structure. The second is multi-resolution fusion (MRF) that fuses high-resolution details with low-resolution context, ensuring that large-area blemishes can be repaired while maintaining global coherence. These two designs achieve fine-grained local corrections with broad structural consistency concurrently, delivering superior face retouching performance across multiple benchmarks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 180, + 553, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 553, + 360 + ], + "type": "text", + "content": "The contributions of this work can be summarized in three major aspects. First, we create the first large-scale, high-quality and diverse face retouching benchmark that comprise 25,000 pairs of high-definition face images that are generated by diffusion models. The benchmark captures a variety of blemishes and facial attributes which can be employed for both training and evaluating face retouching networks. Second, we design a novel face retouching network that introduces frequency selection and restoration and multi-resolution fusion for effective face retouching. Third, extensive experiments over multiple benchmarks show that our face retouching network achieves superior performance as compared with the state-of-the-art. The proposed benchmark can be effectively adopted for training and evaluating face retouching networks as well." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 373, + 400, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 400, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 400, + 384 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 392, + 451, + 405 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 451, + 405 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 451, + 405 + ], + "type": "text", + "content": "2.1. Face Retouching Studies" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 411, + 553, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 553, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 553, + 626 + ], + "type": "text", + "content": "Face retouching has been widely explored in the literature in two representative approaches. The first leverages traditional image filtering techniques for boosting visual quality while preserving skin textures and details, largely driven by various commercial applications such as Pixlr [20], MeituPic [13], and BeautyCam [12]. However, this approach relies heavily on manual adjustments with poor scalability. The second explores deep learning for face retouching. For example, AutoRetouch [21] adopts residual learning over a large-scale dataset (FFHQR) created from FFHQ [8]. ABPN [9] introduces a pyramid architecture to handle blemishes in high-resolution images. BPFRe [30] addresses limited training data via blemish-aware processing and progressive restoration. RetouchFormer [29] synthesizes high-quality facial images to reduce dependency on paired training data. Despite their impressive performance, these deep learning methods often struggle in maintaining global consistency and local detail preservation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 635, + 440, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 440, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 440, + 647 + ], + "type": "text", + "content": "2.2. Face Retouching Data" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 654, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 715 + ], + "type": "text", + "content": "Though face retouching has been studied for years, the research community still falls short in large-scale face retouching benchmarks. The lack of such benchmarks is largely attributed to the challenges in creating paired face images with and without blemishes. Several prior studies" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14723" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 555, + 242 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 555, + 242 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 555, + 242 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 555, + 242 + ], + "type": "image", + "image_path": "ac58da3de16893fa921c48da0d49c721ff65bd7f751b7dc963032b959c89f154.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 250, + 555, + 262 + ], + "lines": [ + { + "bbox": [ + 55, + 250, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 250, + 555, + 262 + ], + "type": "text", + "content": "Figure 2. Qualitative illustration of the created large-scale, high-quality, and diverse face retouching benchmark. Zoom-in for more details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 282, + 297, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 282, + 297, + 427 + ], + "spans": [ + { + "bbox": [ + 54, + 282, + 297, + 427 + ], + "type": "text", + "content": "attempted to tackle the challenges in different ways. For example, [10] presents an exemplar-based method that collects retouching data with laser-treated facial images. However, its applicability is limited to laser-induced effects only. AutoRetouch [21] introduces FFHQR, a large-scale retouching benchmark comprising 70,000 image pairs. However, many images in FFHQR do not have blemishes since the underlying FFHQ dataset was originally collected for face generation. We utilize diffusion models to synthesize paired face images with and without blemishes, leading to a comprehensive paired dataset that can be adopted for training and evaluating face retouching networks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 438, + 275, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 275, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 275, + 453 + ], + "type": "text", + "content": "3. High-Definition Retouching Benchmark" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 459, + 297, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 459, + 297, + 592 + ], + "spans": [ + { + "bbox": [ + 54, + 459, + 297, + 592 + ], + "type": "text", + "content": "Manual collection of paired retouching images is a challenging task that requires sophisticated control to ensure just one variable for each pair of face images, i.e., having or not having blemishes. In addition, collecting such large-scale and diverse paired images is laborious and time-consuming. We address this challenge by synthesizing paired face images with stable diffusion models, leading to the first large-scale, high-quality, and diverse benchmark that comprises 25,000 paired face images with diverse blemish types and facial attributes. Fig. 2 shows samples of synthesized image pairs with and without blemishes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 600, + 194, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 194, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 194, + 612 + ], + "type": "text", + "content": "3.1. Data Generation Process" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 618, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 618, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 618, + 296, + 715 + ], + "type": "text", + "content": "We create the proposed face retouching benchmark in two stages, including the first stage that generates face images without blemishes and the second that generates the corresponding paired image with blemishes. We generate the face images without blemishes by leveraging three fine-tuned stable diffusion models. Specifically, we design 7,920 image-generation prompts that systematically permute combinations of facial attributes—including age," + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 315, + 281, + 556, + 355 + ], + "blocks": [ + { + "bbox": [ + 315, + 281, + 556, + 355 + ], + "lines": [ + { + "bbox": [ + 315, + 281, + 556, + 355 + ], + "spans": [ + { + "bbox": [ + 315, + 281, + 556, + 355 + ], + "type": "table", + "html": "
Blemish TypesNumber of Each BlemishNumber of Synthe-sized FaceNumber of Natural Face
HGFRMoles, Freckles, Acne7,92023,7601,240
", + "image_path": "bfde7dc39e556beb3481f9369fb084519fb229d1fbc6039d6ac663bdd998508b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 339, + 362, + 528, + 374 + ], + "lines": [ + { + "bbox": [ + 339, + 362, + 528, + 374 + ], + "spans": [ + { + "bbox": [ + 339, + 362, + 528, + 374 + ], + "type": "text", + "content": "Table 1. Details of the proposed benchmark HGFR." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 397, + 555, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 555, + 422 + ], + "type": "text", + "content": "gender, aesthetics, hairstyle, skin tone, and detailed features (e.g., lips, nose, eyes, face shape, and eyebrows)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 422, + 554, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 554, + 470 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 554, + 470 + ], + "type": "text", + "content": "The generation with such combinatorial facial attributes ensures that the generated face images exhibit a broad spectrum of diversity and rich facial representations while preserving a natural and unblemished appearance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 471, + 556, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 556, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 556, + 662 + ], + "type": "text", + "content": "For each clean face image generated in the first stage, we generate a corresponding paired image with blemishes via a controlled modification process. Specifically, we leverage a state-of-the-art facial attribute segmentation model [19] to first partition the clean face image into semantically meaningful regions (e.g., cheeks, nose, and neck) while deliberately excluding key areas such as hair, eyes, eyebrows, and mouth to maintain critical identity cues. We then apply the pre-trained LoRA [6] to the isolated facial regions to synthesize blemishes such as moles, freckles, and acne. Such targeted blemish synthesis ensures that the paired images accurately reflect common facial imperfections without compromising the overall facial structure and identity, thereby balancing the local repairs and global consistency. More details on the image synthesis are provided in the supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 671, + 405, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 405, + 683 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 405, + 683 + ], + "type": "text", + "content": "3.2. Data Statistics" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "content": "The proposed benchmark HGFR consists of 25,000 image pairs, including 23,760 synthesized image pairs and 1,240" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "14724" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 60, + 555, + 255 + ], + "blocks": [ + { + "bbox": [ + 56, + 60, + 555, + 255 + ], + "lines": [ + { + "bbox": [ + 56, + 60, + 555, + 255 + ], + "spans": [ + { + "bbox": [ + 56, + 60, + 555, + 255 + ], + "type": "table", + "html": "
Dataset Name# Images/PairsResolutionFace SourceTypeFace with Blemishes
PPR10K11,161 images4K-8KNaturalPortrait Retouching5%
RetouchingFFHQ>500K images1024×1024NaturalFace Recognition16%
EFrR--NaturalFace Retouching-
DigiFace-1M>1M images-SyntheticVarious poses, lighting, and backgrounds-
FRTDD--Restoration-basedTampering Detection-
FFHQ70,000 pairs1024×1024NaturalFace Generation16%
FFHQR70,000 pairs1024×1024NaturalFace Retouching16%
HGFR (Ours)25,000 pairs1024×1024Synthetic & NaturalFace Retouching83%
", + "image_path": "7b122b6906e6dcc078542654ee18f433b4f4898fa66f6da93e20ee93cf5527f8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 263, + 555, + 287 + ], + "lines": [ + { + "bbox": [ + 55, + 263, + 555, + 287 + ], + "spans": [ + { + "bbox": [ + 55, + 263, + 555, + 287 + ], + "type": "text", + "content": "Table 2. Summary of existing face retouching benchmarks. " + }, + { + "bbox": [ + 55, + 263, + 555, + 287 + ], + "type": "inline_equation", + "content": "{}^{\\prime } -" + }, + { + "bbox": [ + 55, + 263, + 555, + 287 + ], + "type": "text", + "content": " ’ Indicates that the information is unavailable. The percentages of face images with Blemishes are determined by the zero-shot CLIP classifier. More details are provided in the supplementary materials." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 307, + 294, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 307, + 294, + 379 + ], + "spans": [ + { + "bbox": [ + 54, + 307, + 294, + 379 + ], + "type": "text", + "content": "natural ones. As detailed in Tab. 1, HGFR consists of three primary types of blemishes including moles, freckles, and acne—with 7,920 image pairs in each of the three categories. Such systematic design yields a comprehensive representation of facial blemishes across 23,760 synthetic images, augmented by a small set of natural faces." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 380, + 295, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 380, + 295, + 511 + ], + "spans": [ + { + "bbox": [ + 54, + 380, + 295, + 511 + ], + "type": "text", + "content": "Tab. 2 compares HGFR with several existing retouching benchmark. HGFR has two unique features. First, it has superb diversity and includes both synthetic and natural face retouching images, enabling more robust and comprehensive face retouching evaluations. Second, HGFR contains a greater variety of blemish types. Based on the classification by zero-shot CLIP [16], up to " + }, + { + "bbox": [ + 54, + 380, + 295, + 511 + ], + "type": "inline_equation", + "content": "83\\%" + }, + { + "bbox": [ + 54, + 380, + 295, + 511 + ], + "type": "text", + "content": " images in HGFR contain blemishes while most existing benchmarks have less than " + }, + { + "bbox": [ + 54, + 380, + 295, + 511 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 54, + 380, + 295, + 511 + ], + "type": "text", + "content": " image containing blemishes. The much larger number of blemish instances enables more robust training and evaluation of existing and future face retouching networks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 524, + 111, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 524, + 111, + 537 + ], + "spans": [ + { + "bbox": [ + 55, + 524, + 111, + 537 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 544, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 640 + ], + "type": "text", + "content": "Fig. 3 illustrates our overall network architecture, which is composed of several key components. First, a Soft Mask Generation Branch (SMGB) automatically predicts masks to guide localized blemish repair. Next, the frequency selection and restoration (FSR) refines skin regions by leveraging both frequency and spatial domain information. Finally, the multi-resolution fusion (MRF) integrates multi-resolution outputs to reconstruct a high-quality, blemish-free portrait." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "Given an input image " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "I_0 \\in \\mathbb{R}^{h \\times w \\times 3}" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": ", we first generate a Laplacian planes, " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "L_I = [I_0, I_1, \\dots, I_l]" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": ", and the corresponding high-frequency group " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "L_H = [H_0, H_1, \\dots, H_l]" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " derived from Laplacian pyramid decomposition [1] (with " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": ", the number of downsampling operations, set to 2 by default). The FSR is then applied to each " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "I_i \\in L_I" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " to produce" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 306, + 555, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 555, + 379 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 555, + 379 + ], + "type": "text", + "content": "retouched outputs " + }, + { + "bbox": [ + 313, + 306, + 555, + 379 + ], + "type": "inline_equation", + "content": "R_{i} \\in \\mathbb{R}\\frac{h}{2^{i}}\\times \\frac{w}{2^{i}}\\times 3" + }, + { + "bbox": [ + 313, + 306, + 555, + 379 + ], + "type": "text", + "content": ". These multi-scale results are subsequently fused via MRF to reconstruct the final high-resolution portrait. Inspired by the Laplacian pyramid [1], MRF uses progressive upsampling combined with a Conv+LeakyReLU+Conv(CLC) module to refine details and ensure seamless integration of features across scales." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 380, + 556, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 556, + 549 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 556, + 549 + ], + "type": "text", + "content": "The Soft Mask Generation Branch (SMGB) is designed to automatically predict a mask " + }, + { + "bbox": [ + 313, + 380, + 556, + 549 + ], + "type": "inline_equation", + "content": "M \\in \\mathbb{R}^{\\frac{h}{2^l} \\times \\frac{w}{2^l} \\times 3}" + }, + { + "bbox": [ + 313, + 380, + 556, + 549 + ], + "type": "text", + "content": " that highlights target regions for localized repair. We adopt a U-Net architecture with skip connections [18] to fuse low-level and high-level features, thereby enhancing segmentation accuracy. Although most datasets lack explicit ground truth masks " + }, + { + "bbox": [ + 313, + 380, + 556, + 549 + ], + "type": "inline_equation", + "content": "M_{gt}" + }, + { + "bbox": [ + 313, + 380, + 556, + 549 + ], + "type": "text", + "content": ", these can be derived by computing the difference between the input image " + }, + { + "bbox": [ + 313, + 380, + 556, + 549 + ], + "type": "inline_equation", + "content": "I_l" + }, + { + "bbox": [ + 313, + 380, + 556, + 549 + ], + "type": "text", + "content": " and its corresponding ground truth, followed by thresholding. The SMGB contributes in two ways: first, the predicted mask " + }, + { + "bbox": [ + 313, + 380, + 556, + 549 + ], + "type": "inline_equation", + "content": "M_{pred}" + }, + { + "bbox": [ + 313, + 380, + 556, + 549 + ], + "type": "text", + "content": " enables the FSR to concentrate on defect regions; and second, joint training with SMGB enhances the network's overall contextual and semantic awareness, leading to more consistent retouching results." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 555, + 509, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 509, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 509, + 569 + ], + "type": "text", + "content": "4.1. Frequency Selection and Restoration" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 573, + 555, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 573, + 555, + 657 + ], + "spans": [ + { + "bbox": [ + 313, + 573, + 555, + 657 + ], + "type": "text", + "content": "To address Challenge (1) outlined in Section 1—achieving local repairs with global consistency and detail fidelity—we propose a context-aware frequency selection and restoration (FSR) framework. As shown in Fig. 3, FSR adopts a multi-task design that includes Frequency-Aware Dynamic Aggregation, Space Domain Projection, and Selective Frequency Domain-based FFN." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 662, + 402, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 662, + 402, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 662, + 402, + 673 + ], + "type": "text", + "content": "4.1.1. Preliminaries." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 676, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 676, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 676, + 554, + 713 + ], + "type": "text", + "content": "Given a feature map " + }, + { + "bbox": [ + 313, + 676, + 554, + 713 + ], + "type": "inline_equation", + "content": "z \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 313, + 676, + 554, + 713 + ], + "type": "text", + "content": ", we apply the Discrete Fourier Transform (DFT) channel-wise to map spatial-domain features into the frequency domain. For a single" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14725" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 49, + 558, + 323 + ], + "blocks": [ + { + "bbox": [ + 55, + 49, + 558, + 323 + ], + "lines": [ + { + "bbox": [ + 55, + 49, + 558, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 49, + 558, + 323 + ], + "type": "image", + "image_path": "d426f5a4a2d62c2fb338f988adcf5b91ca3071eadb76c27cd8442afa12e01900.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 331, + 555, + 399 + ], + "lines": [ + { + "bbox": [ + 54, + 331, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 54, + 331, + 555, + 399 + ], + "type": "text", + "content": "Figure 3. Overview of our proposed blemish repair framework. The input image is first decomposed into a Laplacian pyramid to obtain multi-scale representations and high-frequency components. A soft mask generation branch (SMGB) then automatically predicts masks to highlight defect regions for targeted repair. Subsequently, the frequency selection and restoration (FSR) module leverages both dynamic spectral filtering and spatial-domain projection—via learnable quantization matrices and self-attention mechanisms—to adaptively restore fine details and maintain global consistency. Finally, the multi-resolution fusion (MRF) progressively fuses multi-scale outputs through upsampling and convolutional refinement, yielding a high-quality, blemish-free portrait." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 419, + 234, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 234, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 234, + 430 + ], + "type": "text", + "content": "channel, the Fourier transform is defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 439, + 295, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 439, + 295, + 472 + ], + "spans": [ + { + "bbox": [ + 68, + 439, + 295, + 472 + ], + "type": "interline_equation", + "content": "F \\left(z _ {C}\\right) (u, v) = \\sum_ {h = 0} ^ {H - 1} \\sum_ {w = 0} ^ {W - 1} z _ {C} (h, w) e ^ {- j 2 \\pi \\left(\\frac {h u}{H} + \\frac {w v}{W}\\right)} \\tag {1}", + "image_path": "d1729d690d860f2aaa3ca99eacf357b4e5fc55fba4c78e3fd9d7ae87c17b49df.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 482, + 296, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 482, + 296, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 482, + 296, + 518 + ], + "type": "text", + "content": "The real and imaginary parts are then extracted, allowing us to compute the magnitude spectrum " + }, + { + "bbox": [ + 55, + 482, + 296, + 518 + ], + "type": "inline_equation", + "content": "A(z_{C})(u,v)" + }, + { + "bbox": [ + 55, + 482, + 296, + 518 + ], + "type": "text", + "content": " and phase spectrum " + }, + { + "bbox": [ + 55, + 482, + 296, + 518 + ], + "type": "inline_equation", + "content": "\\Phi (z_C)(u,v)" + }, + { + "bbox": [ + 55, + 482, + 296, + 518 + ], + "type": "text", + "content": " via:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 528, + 295, + 554 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 528, + 295, + 554 + ], + "spans": [ + { + "bbox": [ + 62, + 528, + 295, + 554 + ], + "type": "interline_equation", + "content": "A \\left(z _ {C}\\right) (u, v) = \\sqrt {\\Re \\left(F \\left(z _ {C}\\right) (u , v)\\right) ^ {2} + \\Im \\left(F \\left(z _ {C}\\right) (u , v)\\right) ^ {2}} \\tag {2}", + "image_path": "9e57ccca38d3f2b4ee2285518005d95788cad64fac5f902ad766c7e4749b0843.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 88, + 563, + 295, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 563, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 88, + 563, + 295, + 590 + ], + "type": "interline_equation", + "content": "\\Phi (z _ {C}) (u, v) = \\arctan \\left[ \\frac {\\Im (F (z _ {C}) (u , v))}{\\Re (F (z _ {C}) (u , v))} \\right] \\tag {3}", + "image_path": "0f25880bd48959ec0fe96b90b7b80ed04cc2942b6f8baa52562aa083870e272f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 595, + 296, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 595, + 296, + 619 + ], + "spans": [ + { + "bbox": [ + 55, + 595, + 296, + 619 + ], + "type": "text", + "content": "By sequentially applying Eqs. (1) to (3) to all channels, we obtain the full-channel magnitude and phase spectra." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 626, + 257, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 626, + 257, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 626, + 257, + 639 + ], + "type": "text", + "content": "4.1.2. Frequency-Aware Dynamic Aggregation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "In many transformer-based methods [2, 5, 27], global selfattention can overlook localized defect characteristics. To address this, our dynamic spectral filter operates in the frequency domain to target specific imperfections (e.g., moles, freckles, acne), as shown in Fig. 4 left. We aggregate the magnitude and phase information using pointwise convolu" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 420, + 335, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 420, + 335, + 429 + ], + "spans": [ + { + "bbox": [ + 314, + 420, + 335, + 429 + ], + "type": "text", + "content": "tion:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 326, + 434, + 553, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 434, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 326, + 434, + 553, + 453 + ], + "type": "interline_equation", + "content": "F ^ {*} (z) (u, v) = \\mathcal {G} \\left(\\operatorname {C o n v} _ {1} ^ {1} \\left(F \\left(z _ {C} \\cdot M _ {\\text {p r e d}}\\right) (u, v)\\right)\\right) \\tag {4}", + "image_path": "91addcddfb90e6354e2e09be1e55c22b19872bd3f1b0a5d35b3ebeff9912bcfd.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 458, + 553, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 458, + 553, + 518 + ], + "spans": [ + { + "bbox": [ + 315, + 458, + 553, + 518 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 315, + 458, + 553, + 518 + ], + "type": "inline_equation", + "content": "\\mathrm{F}(z_C\\cdot M_{pred})(u,v)\\in" + }, + { + "bbox": [ + 315, + 458, + 553, + 518 + ], + "type": "inline_equation", + "content": "\\{A(z_C\\cdot M_{pred})(u,v),P(z_C\\cdot M_{pred})(u,v)\\}" + }, + { + "bbox": [ + 315, + 458, + 553, + 518 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 315, + 458, + 553, + 518 + ], + "type": "inline_equation", + "content": "\\mathcal{G}(\\cdot)" + }, + { + "bbox": [ + 315, + 458, + 553, + 518 + ], + "type": "text", + "content": " is the GEGLU activation function [22]. To adaptively weight the frequency components, we introduce learnable quantization matrices " + }, + { + "bbox": [ + 315, + 458, + 553, + 518 + ], + "type": "inline_equation", + "content": "W_{A}" + }, + { + "bbox": [ + 315, + 458, + 553, + 518 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 315, + 458, + 553, + 518 + ], + "type": "inline_equation", + "content": "W_{\\Phi}" + }, + { + "bbox": [ + 315, + 458, + 553, + 518 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 365, + 523, + 553, + 543 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 523, + 553, + 543 + ], + "spans": [ + { + "bbox": [ + 365, + 523, + 553, + 543 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} A ^ {\\prime} (z) (u, v) = W _ {A} \\odot A (z) (u, v), \\\\ I ^ {\\prime} (z) (u, v) = W _ {I} \\odot I (z) (u, v). \\end{array} \\tag {5}", + "image_path": "59efef9b45cc40251d7997758965a7ab19d9741809128ac3fc1535ec23ccd497.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 366, + 539, + 501, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 539, + 501, + 552 + ], + "spans": [ + { + "bbox": [ + 366, + 539, + 501, + 552 + ], + "type": "interline_equation", + "content": "\\Phi^ {\\prime} (z) (u, v) = W _ {\\Phi} \\odot \\Phi (z) (u, v).", + "image_path": "1ed5cede7ea07074d973546c4c7c4040e07d50c3b5391373d285cd80d1733f80.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 558, + 533, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 558, + 533, + 570 + ], + "spans": [ + { + "bbox": [ + 314, + 558, + 533, + 570 + ], + "type": "text", + "content": "The updated frequency feature is then recomposed as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 355, + 575, + 553, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 575, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 355, + 575, + 553, + 590 + ], + "type": "interline_equation", + "content": "F ^ {\\prime} (z) (u, v) = A ^ {\\prime} (z) (u, v) e ^ {j \\Phi^ {\\prime} (z) (u, v)}. \\tag {6}", + "image_path": "8abf89e2d75997611f2b62d6f0cc1a47ee7b3a7fbd236d7de2b1d9d823d5f636.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 596, + 553, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 596, + 553, + 619 + ], + "spans": [ + { + "bbox": [ + 313, + 596, + 553, + 619 + ], + "type": "text", + "content": "This filtered component is combined with the original frequency features via a residual connection:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 351, + 625, + 553, + 639 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 625, + 553, + 639 + ], + "spans": [ + { + "bbox": [ + 351, + 625, + 553, + 639 + ], + "type": "interline_equation", + "content": "\\tilde {F} (z) (u, v) = F ^ {\\prime} (z) (u, v) + F (z) (u, v). \\tag {7}", + "image_path": "215cca8b50777470c18e23b264a7575c2d300eee8fb3088bd16f009612d2f18d.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 645, + 553, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 645, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 313, + 645, + 553, + 680 + ], + "type": "text", + "content": "Reconstruction into the spatial domain is achieved by extracting the real and imaginary parts and applying the inverse Fourier Transform:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 348, + 685, + 553, + 704 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 685, + 553, + 704 + ], + "spans": [ + { + "bbox": [ + 348, + 685, + 553, + 704 + ], + "type": "interline_equation", + "content": "\\tilde {\\Re} (z) (u, v) = \\tilde {A} (z) (u, v) \\cdot \\cos \\tilde {\\Phi} (z) (u, v), \\tag {8}", + "image_path": "deeab9843a0e3fa323fefcea8d05e20c80da1c783a3e08016ea40bf3a30f3de5.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 350, + 702, + 518, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 702, + 518, + 715 + ], + "spans": [ + { + "bbox": [ + 350, + 702, + 518, + 715 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathfrak {S}} (z) (u, v) = \\tilde {A} (z) (u, v) \\cdot \\sin \\tilde {\\Phi} (z) (u, v),", + "image_path": "75c3423640680edc5c1f62ba2ae3ffd4ef60bac401ef738b75c72048492f0ea0.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14726" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 296, + 139 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 296, + 139 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 296, + 139 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 296, + 139 + ], + "type": "image", + "image_path": "819a22ee551f1e3fe871f58e5db2d21248e23b830432e0f51b7168c370b6dc6e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 146, + 289, + 158 + ], + "lines": [ + { + "bbox": [ + 62, + 146, + 289, + 158 + ], + "spans": [ + { + "bbox": [ + 62, + 146, + 289, + 158 + ], + "type": "text", + "content": "Figure 4. Details of frequency selection and restoration (FSR)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 102, + 188, + 295, + 209 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 188, + 295, + 209 + ], + "spans": [ + { + "bbox": [ + 102, + 188, + 295, + 209 + ], + "type": "interline_equation", + "content": "\\tilde {z} = \\mathcal {F} ^ {- 1} \\left(\\tilde {\\Re} (z) (u, v), \\tilde {\\Im} (z) (u, v)\\right). \\tag {9}", + "image_path": "bbf8d8ba52a8146772cbfbe2cc71c38863c14bd24079f67395ca59d72c32d595.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 213, + 296, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 213, + 296, + 250 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 296, + 250 + ], + "type": "text", + "content": "The Fourier transform and its inverse are implemented via DFT and IDFT algorithms [3, 4, 34]. We denote the entire FADA process as " + }, + { + "bbox": [ + 55, + 213, + 296, + 250 + ], + "type": "inline_equation", + "content": "\\mathcal{F}\\mathcal{A}(\\cdot)" + }, + { + "bbox": [ + 55, + 213, + 296, + 250 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 256, + 194, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 194, + 268 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 194, + 268 + ], + "type": "text", + "content": "4.1.3. Space Domain Projection." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 271, + 296, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 271, + 296, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 296, + 344 + ], + "type": "text", + "content": "After frequency-domain processing, we further refine features in the spatial domain, as shown in Fig. 4 right. For an input feature " + }, + { + "bbox": [ + 55, + 271, + 296, + 344 + ], + "type": "inline_equation", + "content": "z \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 55, + 271, + 296, + 344 + ], + "type": "text", + "content": ", layer normalization " + }, + { + "bbox": [ + 55, + 271, + 296, + 344 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\cdot)" + }, + { + "bbox": [ + 55, + 271, + 296, + 344 + ], + "type": "text", + "content": " [11] is applied to yield " + }, + { + "bbox": [ + 55, + 271, + 296, + 344 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(z)" + }, + { + "bbox": [ + 55, + 271, + 296, + 344 + ], + "type": "text", + "content": ". This normalized feature is projected into query " + }, + { + "bbox": [ + 55, + 271, + 296, + 344 + ], + "type": "inline_equation", + "content": "(Q)" + }, + { + "bbox": [ + 55, + 271, + 296, + 344 + ], + "type": "text", + "content": ", key " + }, + { + "bbox": [ + 55, + 271, + 296, + 344 + ], + "type": "inline_equation", + "content": "(K)" + }, + { + "bbox": [ + 55, + 271, + 296, + 344 + ], + "type": "text", + "content": ", and value " + }, + { + "bbox": [ + 55, + 271, + 296, + 344 + ], + "type": "inline_equation", + "content": "(V)" + }, + { + "bbox": [ + 55, + 271, + 296, + 344 + ], + "type": "text", + "content": " matrices via independent frequency-domain projections:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 352, + 298, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 352, + 298, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 298, + 376 + ], + "type": "interline_equation", + "content": "Q = \\mathcal {F} \\mathcal {A} _ {Q} (\\mathcal {L} (\\tilde {z})), K = \\mathcal {F} \\mathcal {A} _ {K} (\\mathcal {L} (\\tilde {z})), V = \\mathcal {F} \\mathcal {A} _ {V} (\\mathcal {L} (\\tilde {z})), \\tag {10}", + "image_path": "a4dbac8b6501ec70cf8da92afa53a33b078d734330f0bc057036bd845a64afb5.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 377, + 296, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 377, + 296, + 401 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 296, + 401 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 377, + 296, + 401 + ], + "type": "inline_equation", + "content": "\\mathcal{F}\\mathcal{A}_Q" + }, + { + "bbox": [ + 55, + 377, + 296, + 401 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 377, + 296, + 401 + ], + "type": "inline_equation", + "content": "\\mathcal{F}\\mathcal{A}_K" + }, + { + "bbox": [ + 55, + 377, + 296, + 401 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 377, + 296, + 401 + ], + "type": "inline_equation", + "content": "\\mathcal{F}\\mathcal{A}_V" + }, + { + "bbox": [ + 55, + 377, + 296, + 401 + ], + "type": "text", + "content": " denote learnable projection functions. The output of FSR is then given by:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 113, + 410, + 295, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 410, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 113, + 410, + 295, + 437 + ], + "type": "interline_equation", + "content": "z ^ {\\prime} = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {\\Lambda}} + B\\right) V, \\tag {11}", + "image_path": "cc00271c5b495f911800db0ffb97426136ecd0159b95b380a8b4a905057e73a4.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 445, + 296, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 296, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 296, + 470 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 55, + 445, + 296, + 470 + ], + "type": "inline_equation", + "content": "\\Lambda" + }, + { + "bbox": [ + 55, + 445, + 296, + 470 + ], + "type": "text", + "content": " representing the channel dimension and " + }, + { + "bbox": [ + 55, + 445, + 296, + 470 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 445, + 296, + 470 + ], + "type": "text", + "content": " the positional bias. Finally, a residual connection is used:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 148, + 479, + 294, + 492 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 479, + 294, + 492 + ], + "spans": [ + { + "bbox": [ + 148, + 479, + 294, + 492 + ], + "type": "interline_equation", + "content": "z ^ {*} = z ^ {\\prime} + z. \\tag {12}", + "image_path": "81d0a94810a720bbb4737b6c1327208626fd6313e8159f62af5d9dfa2fc9ae35.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 509, + 255, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 255, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 255, + 521 + ], + "type": "text", + "content": "4.1.4. Selective Frequency Domain-based FFN" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 525, + 296, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 296, + 595 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 296, + 595 + ], + "type": "text", + "content": "To further enhance feature representation, as shown in Fig. 5 we design a selective frequency feed-forward network (SFFN) that adaptively retains critical frequency components for latent blemish removal. Inspired by the JPEG compression algorithm, SFFN employs a learnable quantization matrix:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 113, + 603, + 295, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 603, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 113, + 603, + 295, + 668 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} z _ {1} = F \\left(P \\left(C o n v _ {1} ^ {1} (\\mathcal {L} (z)))\\right)\\right), \\\\ z _ {2} = F ^ {- 1} \\left(W \\odot z _ {1}\\right), \\tag {13} \\\\ z _ {\\text {o u t}} = \\mathcal {G} \\left(P ^ {- 1} \\left(z _ {2}\\right)\\right) + z ^ {*}, \\\\ \\end{array}", + "image_path": "04dc0ef2f555548481139802ec13600666f6043bcf4e809e8956109f95d29da9.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "P(\\cdot)" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "P^{-1}(\\cdot)" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " represent patch unfolding and folding operations. The output " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{out}}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " is then fed into subsequent layers." + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 317, + 50, + 553, + 151 + ], + "blocks": [ + { + "bbox": [ + 317, + 50, + 553, + 151 + ], + "lines": [ + { + "bbox": [ + 317, + 50, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 317, + 50, + 553, + 151 + ], + "type": "image", + "image_path": "71de102d6168d7cf02648eba3f2329e64cd15e0e610fbff7b7e635cf53361b5b.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 325, + 159, + 542, + 171 + ], + "lines": [ + { + "bbox": [ + 325, + 159, + 542, + 171 + ], + "spans": [ + { + "bbox": [ + 325, + 159, + 542, + 171 + ], + "type": "text", + "content": "Figure 5. Details of selective frequency domain-based FFN" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 316, + 181, + 375, + 240 + ], + "blocks": [ + { + "bbox": [ + 316, + 181, + 375, + 240 + ], + "lines": [ + { + "bbox": [ + 316, + 181, + 375, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 181, + 375, + 240 + ], + "type": "image", + "image_path": "6bd23f8512ed513e7538d18de33844475b45e6309dd000089f923edefecc18f8.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 240, + 358, + 248 + ], + "lines": [ + { + "bbox": [ + 333, + 240, + 358, + 248 + ], + "spans": [ + { + "bbox": [ + 333, + 240, + 358, + 248 + ], + "type": "text", + "content": "(a) Input" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 264, + 554, + 309 + ], + "lines": [ + { + "bbox": [ + 313, + 264, + 554, + 309 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 554, + 309 + ], + "type": "text", + "content": "Figure 6. Visual Comparison of MRF at Different Resolutions: (a) Input image; (b) Low Resolution image, with large blemishes removed; (c) Original Resolution image, with fine blemishes removed; (d) Our method, achieving the best retouching." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 376, + 181, + 435, + 240 + ], + "blocks": [ + { + "bbox": [ + 376, + 181, + 435, + 240 + ], + "lines": [ + { + "bbox": [ + 376, + 181, + 435, + 240 + ], + "spans": [ + { + "bbox": [ + 376, + 181, + 435, + 240 + ], + "type": "image", + "image_path": "d632f7a4455f60dffa4c0a3fa3785ff5aa1d969160602204548f117ba703cd2d.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 387, + 240, + 424, + 247 + ], + "lines": [ + { + "bbox": [ + 387, + 240, + 424, + 247 + ], + "spans": [ + { + "bbox": [ + 387, + 240, + 424, + 247 + ], + "type": "text", + "content": "(b) Result on" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 383, + 247, + 425, + 255 + ], + "lines": [ + { + "bbox": [ + 383, + 247, + 425, + 255 + ], + "spans": [ + { + "bbox": [ + 383, + 247, + 425, + 255 + ], + "type": "text", + "content": "Low Resolution" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 434, + 181, + 493, + 240 + ], + "blocks": [ + { + "bbox": [ + 434, + 181, + 493, + 240 + ], + "lines": [ + { + "bbox": [ + 434, + 181, + 493, + 240 + ], + "spans": [ + { + "bbox": [ + 434, + 181, + 493, + 240 + ], + "type": "image", + "image_path": "5c9a2ab99ac84dee4030e670208a86ab85603fc33ff67e5ffba81620f9e0dae7.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 446, + 240, + 482, + 247 + ], + "lines": [ + { + "bbox": [ + 446, + 240, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 446, + 240, + 482, + 247 + ], + "type": "text", + "content": "(c) Result on" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 438, + 247, + 489, + 255 + ], + "lines": [ + { + "bbox": [ + 438, + 247, + 489, + 255 + ], + "spans": [ + { + "bbox": [ + 438, + 247, + 489, + 255 + ], + "type": "text", + "content": "Original Resolution" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 494, + 181, + 553, + 240 + ], + "blocks": [ + { + "bbox": [ + 494, + 181, + 553, + 240 + ], + "lines": [ + { + "bbox": [ + 494, + 181, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 494, + 181, + 553, + 240 + ], + "type": "image", + "image_path": "6b09e06d585728865577e5449d313237669699faf4798e91de5303c9e044466a.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 513, + 240, + 534, + 247 + ], + "lines": [ + { + "bbox": [ + 513, + 240, + 534, + 247 + ], + "spans": [ + { + "bbox": [ + 513, + 240, + 534, + 247 + ], + "type": "text", + "content": "(d) Our" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 327, + 451, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 451, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 451, + 338 + ], + "type": "text", + "content": "4.2. Multi-Resolution Fusion" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 344, + 555, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 344, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 344, + 555, + 453 + ], + "type": "text", + "content": "We observe that while high-resolution features are essential for restoring fine details, low-resolution features are more effective for repairing large-area blemishes. The differences can be observed in Fig. 6. To combine such complementary advantages, we propose MRF to fuse multi-resolution features to produce a detail-rich, blemish-free portrait. Specifically, MRF incorporates high-frequency components from the original image to compensate for the information loss during downsampling. The fusion process is defined as:" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 321, + 457, + 553, + 494 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 457, + 553, + 494 + ], + "spans": [ + { + "bbox": [ + 321, + 457, + 553, + 494 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} O _ {l - 1} = \\operatorname {C o n v} _ {3} ^ {3} \\left(\\sigma \\left(\\operatorname {C o n v} _ {3} ^ {3} \\left(\\operatorname {C a t} \\left(u p \\left(O _ {l}\\right), H _ {l - 1}\\right)\\right)\\right)\\right) \\tag {14} \\\\ + u p (R _ {l - 1}), \\\\ \\end{array}", + "image_path": "7c9918e6e32f6f65a0deeb0d7f7988b6719f56f6bfc5c9efbd7aa7da34c558fe.jpg" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 499, + 555, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 499, + 555, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 499, + 555, + 643 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 499, + 555, + 643 + ], + "type": "inline_equation", + "content": "O_{l}" + }, + { + "bbox": [ + 313, + 499, + 555, + 643 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 499, + 555, + 643 + ], + "type": "inline_equation", + "content": "O_{l-1}" + }, + { + "bbox": [ + 313, + 499, + 555, + 643 + ], + "type": "text", + "content": " denotes the fusion result, and " + }, + { + "bbox": [ + 313, + 499, + 555, + 643 + ], + "type": "inline_equation", + "content": "O_{l} = R_{l}" + }, + { + "bbox": [ + 313, + 499, + 555, + 643 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 499, + 555, + 643 + ], + "type": "inline_equation", + "content": "R_{l}" + }, + { + "bbox": [ + 313, + 499, + 555, + 643 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 499, + 555, + 643 + ], + "type": "inline_equation", + "content": "R_{l-1}" + }, + { + "bbox": [ + 313, + 499, + 555, + 643 + ], + "type": "text", + "content": " corresponds to the different resolutions of the blemish removal face images generated by Laplace image " + }, + { + "bbox": [ + 313, + 499, + 555, + 643 + ], + "type": "inline_equation", + "content": "I_{l}" + }, + { + "bbox": [ + 313, + 499, + 555, + 643 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 499, + 555, + 643 + ], + "type": "inline_equation", + "content": "I_{l-1}" + }, + { + "bbox": [ + 313, + 499, + 555, + 643 + ], + "type": "text", + "content": " through the face blemish removal network, respectively. " + }, + { + "bbox": [ + 313, + 499, + 555, + 643 + ], + "type": "inline_equation", + "content": "H_{l-1}" + }, + { + "bbox": [ + 313, + 499, + 555, + 643 + ], + "type": "text", + "content": " is the high-frequency component of Laplace image " + }, + { + "bbox": [ + 313, + 499, + 555, + 643 + ], + "type": "inline_equation", + "content": "I_{l-1}" + }, + { + "bbox": [ + 313, + 499, + 555, + 643 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 499, + 555, + 643 + ], + "type": "inline_equation", + "content": "up(\\cdot)" + }, + { + "bbox": [ + 313, + 499, + 555, + 643 + ], + "type": "text", + "content": " denotes bilinear interpolation, " + }, + { + "bbox": [ + 313, + 499, + 555, + 643 + ], + "type": "inline_equation", + "content": "\\mathrm{Cat}(\\cdot)" + }, + { + "bbox": [ + 313, + 499, + 555, + 643 + ], + "type": "text", + "content": " channel-wise concatenation. The CLC [15] is implemented through sequential " + }, + { + "bbox": [ + 313, + 499, + 555, + 643 + ], + "type": "inline_equation", + "content": "Conv_{3}^{3}" + }, + { + "bbox": [ + 313, + 499, + 555, + 643 + ], + "type": "text", + "content": " layers with leaky ReLU activation, which is a convolution with a " + }, + { + "bbox": [ + 313, + 499, + 555, + 643 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 313, + 499, + 555, + 643 + ], + "type": "text", + "content": " kernel and stride 3, and " + }, + { + "bbox": [ + 313, + 499, + 555, + 643 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 499, + 555, + 643 + ], + "type": "text", + "content": " is a leaky ReLU activation (negative slope 0.2); when " + }, + { + "bbox": [ + 313, + 499, + 555, + 643 + ], + "type": "inline_equation", + "content": "l = 1" + }, + { + "bbox": [ + 313, + 499, + 555, + 643 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 499, + 555, + 643 + ], + "type": "inline_equation", + "content": "O_{l-1} = O_{0}" + }, + { + "bbox": [ + 313, + 499, + 555, + 643 + ], + "type": "text", + "content": " i.e. the final blemish-removed face image." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 652, + 395, + 666 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 652, + 395, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 652, + 395, + 666 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 672, + 441, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 441, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 441, + 685 + ], + "type": "text", + "content": "5.1. Experimental Settings" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "Datasets. We conduct extensive experiments over two benchmarks to validate our face retouching network. The" + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14727" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 555, + 177 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 555, + 177 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 555, + 177 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 555, + 177 + ], + "type": "table", + "html": "
DatasetsFFHQRHGFRFFHQR + HGFR
MetricsPSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
AutoRetouch29.380.9180.076630.680.8920.082331.060.9240.0744
pix2pixHD41.180.9800.013342.370.9650.016742.680.9820.0132
ABPN44.260.9920.017647.020.9860.018947.650.9920.0178
BPFRe45.130.9930.009247.280.9890.013448.520.9930.0128
RetouchFormer45.240.9930.009047.610.9900.012848.780.9930.0105
Ours45.480.9930.008248.020.9910.012549.150.9940.0094
", + "image_path": "2d54259064f5256db0da009fa71a2fef575f3a73abc799ae42cffef8e3db3caf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 62, + 184, + 547, + 196 + ], + "lines": [ + { + "bbox": [ + 62, + 184, + 547, + 196 + ], + "spans": [ + { + "bbox": [ + 62, + 184, + 547, + 196 + ], + "type": "text", + "content": "Table 3. Quantitative comparison of our method with existing methods over three training setups ( " + }, + { + "bbox": [ + 62, + 184, + 547, + 196 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 62, + 184, + 547, + 196 + ], + "type": "text", + "content": " Lower is better; " + }, + { + "bbox": [ + 62, + 184, + 547, + 196 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 62, + 184, + 547, + 196 + ], + "type": "text", + "content": " Higher is better)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 57, + 207, + 555, + 415 + ], + "blocks": [ + { + "bbox": [ + 57, + 207, + 555, + 415 + ], + "lines": [ + { + "bbox": [ + 57, + 207, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 57, + 207, + 555, + 415 + ], + "type": "image", + "image_path": "ebb5f8f7675142b521bc957f45c4af8619ffeb23a0f734a09dfc8cbbc7de89d8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 113, + 422, + 495, + 434 + ], + "lines": [ + { + "bbox": [ + 113, + 422, + 495, + 434 + ], + "spans": [ + { + "bbox": [ + 113, + 422, + 495, + 434 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison of our method with several state-of-the-art methods on Internet images." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 454, + 295, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 454, + 295, + 539 + ], + "spans": [ + { + "bbox": [ + 54, + 454, + 295, + 539 + ], + "type": "text", + "content": "first is our proposed synthetic benchmark HGFR as detailed in Section 3. The second is the benchmark FFHQR [8], a large-scale face retouching benchmark consisting of 70,000 high-definition facial images paired with their retouched versions. We randomly split HGFR into 20,000 images for training, 2,500 for validation, and 2,500 for testing. For FFHQR, we follow the data splitting in [8]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 542, + 295, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 542, + 295, + 649 + ], + "spans": [ + { + "bbox": [ + 54, + 542, + 295, + 649 + ], + "type": "text", + "content": "Implementation Details. Our model and baselines are implemented in Python 3.8 using PyTorch 2.3.1 and trained on 4 NVIDIA RTX 4090 GPUs. We adopt the Adam optimizer with " + }, + { + "bbox": [ + 54, + 542, + 295, + 649 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 54, + 542, + 295, + 649 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 542, + 295, + 649 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.999" + }, + { + "bbox": [ + 54, + 542, + 295, + 649 + ], + "type": "text", + "content": ", and set the initial learning rate to 0.0005 for a batch size of 8. After 100 epochs, the learning rate is reduced by a factor of 10. The default number of downsampling steps (" + }, + { + "bbox": [ + 54, + 542, + 295, + 649 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 54, + 542, + 295, + 649 + ], + "type": "text", + "content": ") is set to 2. Training takes approximately 15 hours on HGFR, 35 hours on FFHQR, and 55 hours on FFHQR+HGFR." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 653, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 653, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 653, + 295, + 715 + ], + "type": "text", + "content": "Evaluation Metrics. We evaluate the retouching with three evaluation metrics including Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM)[28], and Learned Perceptual Image Patch Similarity (LPIPS)[33]. The three metrics have been widely" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 455, + 484, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 484, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 484, + 467 + ], + "type": "text", + "content": "adopted for image generation evaluations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 478, + 533, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 478, + 533, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 533, + 491 + ], + "type": "text", + "content": "5.2. Quantitative and Qualitative Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 497, + 555, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 497, + 555, + 664 + ], + "spans": [ + { + "bbox": [ + 312, + 497, + 555, + 664 + ], + "type": "text", + "content": "We compare our face retouching method with several existing work including Pix2PixHD [26], AutoRetouch [21], ABPN [9], BPFRe [30], and Retouch-Former [29]. Pix2PixHD is an image-to-image translation network and the rest are specifically designed for facial retouching. We train three retouching models by using the HGFR training images, the FFHQR training images, and a combination of the training images of the two datasets (i.e., FFHQR+HGFR). Tab. 3 shows experimental results. It can be observed that the proposed method performs clearly the best consistently across all datasets and evaluation metrics. The superior face retouching performance is largely attributed to our designed FSR and MRF, more details to be discussed in the ensuing ablation experiments." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "We also conduct qualitative comparisons to further demonstrate the superiority of our designed face retouching network as illustrated in Fig. 7. It can be observed that AutoRetouch tends to introduce global noises and fails to" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14728" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 175, + 150 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 175, + 150 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 175, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 175, + 150 + ], + "type": "image", + "image_path": "f5ecec7d3205b1cd010a4b6f00b0f50cfbfa28e60bd0522fbe6498137ed08438.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 104, + 150, + 129, + 158 + ], + "lines": [ + { + "bbox": [ + 104, + 150, + 129, + 158 + ], + "spans": [ + { + "bbox": [ + 104, + 150, + 129, + 158 + ], + "type": "text", + "content": "(a) Input" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 175, + 70, + 294, + 150 + ], + "blocks": [ + { + "bbox": [ + 175, + 70, + 294, + 150 + ], + "lines": [ + { + "bbox": [ + 175, + 70, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 175, + 70, + 294, + 150 + ], + "type": "image", + "image_path": "8b7d8eb23a34d43cfcd2d655d4de4c5c9e87bdc1bea243b0e2dd2dd2daef9440.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 195, + 150, + 274, + 158 + ], + "lines": [ + { + "bbox": [ + 195, + 150, + 274, + 158 + ], + "spans": [ + { + "bbox": [ + 195, + 150, + 274, + 158 + ], + "type": "text", + "content": "(b) Retouching with FFHQR" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 56, + 159, + 176, + 238 + ], + "blocks": [ + { + "bbox": [ + 56, + 159, + 176, + 238 + ], + "lines": [ + { + "bbox": [ + 56, + 159, + 176, + 238 + ], + "spans": [ + { + "bbox": [ + 56, + 159, + 176, + 238 + ], + "type": "image", + "image_path": "faf6a1349bbd690c121c45918d335791461ade172e753edb5a688df734a23da6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 79, + 239, + 154, + 247 + ], + "lines": [ + { + "bbox": [ + 79, + 239, + 154, + 247 + ], + "spans": [ + { + "bbox": [ + 79, + 239, + 154, + 247 + ], + "type": "text", + "content": "(c) Retouching with HGFR" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 257, + 295, + 291 + ], + "lines": [ + { + "bbox": [ + 55, + 257, + 295, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 295, + 291 + ], + "type": "text", + "content": "Figure 8. Face retouching with models trained by using FFHQR training data alone in (b), HGFR training data alone in (c) and the combination of HGFR and FFHQR training data in (d)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 175, + 159, + 294, + 238 + ], + "blocks": [ + { + "bbox": [ + 175, + 159, + 294, + 238 + ], + "lines": [ + { + "bbox": [ + 175, + 159, + 294, + 238 + ], + "spans": [ + { + "bbox": [ + 175, + 159, + 294, + 238 + ], + "type": "image", + "image_path": "fd957c41b234c399878da7a8d426f8814b599a69ae3f41615c30b24114930f6d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 239, + 285, + 247 + ], + "lines": [ + { + "bbox": [ + 184, + 239, + 285, + 247 + ], + "spans": [ + { + "bbox": [ + 184, + 239, + 285, + 247 + ], + "type": "text", + "content": "(d) Retouching with FFHQR+HGFR" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 312, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 296, + 396 + ], + "type": "text", + "content": "clearly delineate retouched regions while Pix2PixHD suffers from noticeable color shifts. ABPN, RetouchFormer, and BPFRe can deliver competitive retouching effects, but their outputs fall short in detail preservation and refinement. As a comparison, our method balances the local blemish repair and global facial consistency and demonstrates superior retouching experience." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 397, + 296, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 397, + 296, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 397, + 296, + 553 + ], + "type": "text", + "content": "The experiments also validate the effectiveness of our benchmark from two perspectives: 1) training effective retouching network; and 2) augmenting existing benchmark for better retouching performance. Fig. 8 shows qualitative experiments. We can see that the model trained using HGFR alone performs clearly better than the one trained using FFHQR alone. The clear better retouching effect is largely because our benchmark is specifically synthesized with large-scale, high-quality, and diverse paired images. In addition, we can observe that the model using HGRR+FFHQR performs the best, demonstrating that our created benchmark can augment and complement existing benchmarks effectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 563, + 149, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 149, + 575 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 149, + 575 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 581, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 295, + 688 + ], + "type": "text", + "content": "We conduct ablation experiments to examine how the proposed FSR and MRF contribute to the face retouching as shown in Tab. 4. The baseline FFN only achieves 42.93 dB in PSNR and 0.967 in SSIM. Including FSR improves the two metrics to 45.79 dB and 0.980, indicating that our frequency-aware design substantially improves detail preservation. Further including MRF on top of FSR yields the best performance (47.81 dB/0.990), indicating the complementary role of MRF via multi-scale feature refinement." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "We also conduct qualitative ablation experiments. As Fig. 9 shows, FSR outperforms the baseline clearly in bal" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 435, + 152 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 435, + 152 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 435, + 152 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 435, + 152 + ], + "type": "image", + "image_path": "76c3bee8d22bad114f848b418e81923d90e6fe1a9e1d2abc1889706abe0e2fad.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 152, + 391, + 162 + ], + "lines": [ + { + "bbox": [ + 361, + 152, + 391, + 162 + ], + "spans": [ + { + "bbox": [ + 361, + 152, + 391, + 162 + ], + "type": "text", + "content": "(a) Input" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 435, + 70, + 553, + 152 + ], + "blocks": [ + { + "bbox": [ + 435, + 70, + 553, + 152 + ], + "lines": [ + { + "bbox": [ + 435, + 70, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 435, + 70, + 553, + 152 + ], + "type": "image", + "image_path": "a1d9074a402a4d60267611d041b062c39ec64afb105b2f4cf857871045139d1d.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 152, + 514, + 162 + ], + "lines": [ + { + "bbox": [ + 473, + 152, + 514, + 162 + ], + "spans": [ + { + "bbox": [ + 473, + 152, + 514, + 162 + ], + "type": "text", + "content": "(b) Baseline" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 316, + 162, + 435, + 243 + ], + "blocks": [ + { + "bbox": [ + 316, + 162, + 435, + 243 + ], + "lines": [ + { + "bbox": [ + 316, + 162, + 435, + 243 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 435, + 243 + ], + "type": "image", + "image_path": "34c967b81367a3820f2916c8c7ff2b30cf7b45b08243c53d7838a0962daf4143.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 346, + 244, + 405, + 253 + ], + "lines": [ + { + "bbox": [ + 346, + 244, + 405, + 253 + ], + "spans": [ + { + "bbox": [ + 346, + 244, + 405, + 253 + ], + "type": "text", + "content": "(c) Baseline+FSR" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 262, + 554, + 307 + ], + "lines": [ + { + "bbox": [ + 313, + 262, + 554, + 307 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 554, + 307 + ], + "type": "text", + "content": "Figure 9. Qualitative ablation studies on the proposed FSR and MRF. FSR substantially improves the detail preservation and global consistency as compared with the Baseline, and MRF further repairs residual blemishes with enhanced retouching effects." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 435, + 162, + 553, + 243 + ], + "blocks": [ + { + "bbox": [ + 435, + 162, + 553, + 243 + ], + "lines": [ + { + "bbox": [ + 435, + 162, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 435, + 162, + 553, + 243 + ], + "type": "image", + "image_path": "4e68a1c32716355ecbe6c26c7dab80d83799d4fe4a12254b25c1a6386220173b.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 453, + 244, + 534, + 253 + ], + "lines": [ + { + "bbox": [ + 453, + 244, + 534, + 253 + ], + "spans": [ + { + "bbox": [ + 453, + 244, + 534, + 253 + ], + "type": "text", + "content": "(d) Baseline+FSR+MRF" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 331, + 554, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 554, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 554, + 380 + ], + "type": "text", + "content": "ancing detail preservation and global consistency, and MRF further enhances the face retouching by removing the residual blemishes. Please refer to the appendix for more detailed ablation experiments." + } + ] + } + ], + "index": 23 + }, + { + "type": "table", + "bbox": [ + 315, + 392, + 554, + 445 + ], + "blocks": [ + { + "bbox": [ + 315, + 392, + 554, + 445 + ], + "lines": [ + { + "bbox": [ + 315, + 392, + 554, + 445 + ], + "spans": [ + { + "bbox": [ + 315, + 392, + 554, + 445 + ], + "type": "table", + "html": "
MRFFSRFFNPSNRs/SSIMs
FFN(baseline)✓42.93/0.967
FSR✓45.79/0.980
MRF+FSR✓✓47.81/0.990
", + "image_path": "3129ff7e05eb284def3a419016945df1b886554aacc996fc3c98078d26eb970b.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 330, + 453, + 537, + 464 + ], + "lines": [ + { + "bbox": [ + 330, + 453, + 537, + 464 + ], + "spans": [ + { + "bbox": [ + 330, + 453, + 537, + 464 + ], + "type": "text", + "content": "Table 4. Ablation Studies on FSR and MRF over HGFR." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 475, + 388, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 388, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 388, + 487 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 496, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 555, + 628 + ], + "type": "text", + "content": "This paper presents a high-definition face retouching benchmark and a novel face retouching framework that can effectively balance local detail repair with global consistency. The benchmark comprises 25,000 high-quality paired face retouching images that are generated by diffusion models, possessing superb diversity for training and evaluation face retouching networks. The framework integrates frequency selection and sestoration with multi-resolution fusion, demonstrating superior retouching performance. Extensive experiments confirm the effectiveness of our created benchmark and designed retouching framework." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "type": "text", + "content": "ACKNOWLEDGEMENT This work was supported in part by the National Key Research and Development Program Project of China [grant no. 2024YFC3306902], in part by the National Natural Science Foundation of China [grant no. U24A20242], and in part by the Zhejiang Provincial Natural Science Foundation [grant no. LDT23F02024F02]." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14729" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 124 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 124 + ], + "type": "text", + "content": "[1] Peter J Burt and Edward H Adelson. The laplacian pyramid as a compact image code. In Readings in computer vision, pages 671-679. Elsevier, 1987. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 295, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 295, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 295, + 190 + ], + "type": "text", + "content": "[2] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ArXiv, abs/2010.11929, 2020. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 192, + 294, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 192, + 294, + 224 + ], + "spans": [ + { + "bbox": [ + 62, + 192, + 294, + 224 + ], + "type": "text", + "content": "[3] Matteo Frigo and Steven G Johnson. Fftw: An adaptive software architecture for the fft. In ICASSP, pages 1381-1384. IEEE, 1998. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 227, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 227, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 62, + 227, + 294, + 259 + ], + "type": "text", + "content": "[4] Junming Hou, Qi Cao, Ran Ran, Che Liu, Junling Li, and Liang-jian Deng. Bidomain modeling paradigm for pan-sharpening. In ACM MM, pages 347-357, 2023. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 261, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 261, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 261, + 294, + 304 + ], + "type": "text", + "content": "[5] Jie Hu, Li Shen, Samuel Albanie, Gang Sun, and Enhua Wu. Squeeze-and-excitation networks. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7132-7141, 2017. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 306, + 294, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 306, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 306, + 294, + 350 + ], + "type": "text", + "content": "[6] J. Edward Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. ArXiv, abs/2106.09685, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 351, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 351, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 351, + 294, + 396 + ], + "type": "text", + "content": "[7] Le Jiang, Yan Huang, Lianxin Xie, Wen Xue, Cheng Liu, Si Wu, and Hau-San Wong. Hunting blemishes: Language-guided high-fidelity face retouching transformer with limited paired data. In ACM MM, pages 5102-5111, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 396, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 396, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 396, + 294, + 430 + ], + "type": "text", + "content": "[8] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, pages 4401-4410, 2019. 2, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 431, + 294, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 431, + 294, + 485 + ], + "spans": [ + { + "bbox": [ + 62, + 431, + 294, + 485 + ], + "type": "text", + "content": "[9] Biwen Lei, Xiefan Guo, Hongyu Yang, Miaomiao Cui, Xuansong Xie, and Di Huang. Abpn: adaptive blend pyramid network for real-time local retouching of ultra high-resolution photo. In CVPR, pages 2108-2117, 2022. 1, 2, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 487, + 294, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 487, + 294, + 530 + ], + "spans": [ + { + "bbox": [ + 57, + 487, + 294, + 530 + ], + "type": "text", + "content": "[10] Tsung-Ying Lin, Yu-Ting Tsai, Tsung-Shian Huang, Wen-Chieh Lin, and Jung-Hong Chuang. Exemplar-based freckle retouching and skin tone adjustment. Computers & Graphics, 78:54-63, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 532, + 294, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 532, + 294, + 576 + ], + "spans": [ + { + "bbox": [ + 57, + 532, + 294, + 576 + ], + "type": "text", + "content": "[11] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 578, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 578, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 57, + 578, + 294, + 599 + ], + "type": "text", + "content": "[12] Meitu Inc. Beautycam. https://meiyan.meipai.com/, 2013.2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 600, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 600, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 57, + 600, + 294, + 622 + ], + "type": "text", + "content": "[13] Meitu Inc. Meitupic. http://xiuxiu.meitu.com/, 2013.2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 624, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 294, + 667 + ], + "type": "text", + "content": "[14] Dongbo Min, Sunghwan Choi, Jiangbo Lu, Bumsub Ham, Kwanghoon Sohn, and Minh N Do. Fast global image smoothing based on weighted least squares. IEEE TIP, 23 (12):5638-5653, 2014. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 669, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 669, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 669, + 294, + 712 + ], + "type": "text", + "content": "[15] Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. CoRR, abs/1511.06434, 2015. 6" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 139 + ], + "type": "text", + "content": "[16] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 140, + 554, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 554, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 554, + 183 + ], + "type": "text", + "content": "[17] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 185, + 554, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 185, + 554, + 251 + ], + "spans": [ + { + "bbox": [ + 317, + 185, + 554, + 251 + ], + "type": "text", + "content": "[18] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 252, + 554, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 252, + 554, + 306 + ], + "spans": [ + { + "bbox": [ + 317, + 252, + 554, + 306 + ], + "type": "text", + "content": "[19] Ioannis Sarridis, Christos Koutlis, Symeon Papadopoulos, and Christos Diou. Facex: Understanding face attribute classifiers through summary model explanations. In Proceedings of the 2024 International Conference on Multimedia Retrieval, pages 758-766, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 308, + 553, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 329 + ], + "type": "text", + "content": "[20] Ola Sevandersson. Pixlr. https://www pixlr.com/, 2008.2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 331, + 555, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 555, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 555, + 374 + ], + "type": "text", + "content": "[21] Alireza Shafaei, James J Little, and Mark Schmidt. Autoretouch: Automatic professional face retouching. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 990-998, 2021. 1, 2, 3, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 376, + 554, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 554, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 554, + 398 + ], + "type": "text", + "content": "[22] Noam Shazeer. Glu variants improve transformer. arXiv preprint arXiv:2002.05202, 2020. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 399, + 554, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 554, + 431 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 554, + 431 + ], + "type": "text", + "content": "[23] Wanchao Su, Can Wang, Chen Liu, Fangzhou Han, Hongbo Fu, and Jing Liao. Styleretoucher: Generalized portrait image retouching with gan priors. IEEE TVCG, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 433, + 554, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 554, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 554, + 464 + ], + "type": "text", + "content": "[24] Carlo Tomasi and Roberto Manduchi. Bilateral filtering for gray and color images. In ICCV, pages 839-846. IEEE, 1998. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 466, + 554, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 466, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 466, + 554, + 510 + ], + "type": "text", + "content": "[25] Sudha Velusamy, Rishubh Parihar, Raviprasad Kini, and Aniket Rege. Fabsoften: Face beautification via dynamic skin smoothing, guided feathering, and texture restoration. In CVPRW, pages 530-531, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 512, + 554, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 512, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 512, + 554, + 555 + ], + "type": "text", + "content": "[26] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In CVPR, pages 8798-8807, 2018. 1, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 557, + 554, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 557, + 554, + 600 + ], + "spans": [ + { + "bbox": [ + 317, + 557, + 554, + 600 + ], + "type": "text", + "content": "[27] X. Wang, Ross B. Girshick, Abhinav Kumar Gupta, and Kaiming He. Non-local neural networks. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7794-7803, 2017. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 601, + 554, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 554, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 554, + 634 + ], + "type": "text", + "content": "[28] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 13(4):600-612, 2004. 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 635, + 554, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 554, + 689 + ], + "type": "text", + "content": "[29] Xue Wen, Lianxin Xie, Le Jiang, Tianyi Chen, Si Wu, Cheng Liu, and Hau-San Wong. Retouchformer: semi-supervised high-quality face retouching transformer with prior-based selective self-attention. In AAAI, pages 5903–5911, 2024. 1, 2, 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 691, + 554, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 554, + 714 + ], + "type": "text", + "content": "[30] Lianxin Xie, Wen Xue, Zhen Xu, Si Wu, Zhiwen Yu, and Hau San Wong. Blemish-aware and progressive face re" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14730" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 242 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "touching with limited paired data. In CVPR, pages 5599-5608, 2023. 2, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "type": "text", + "content": "[31] Li Xu, Cewu Lu, Yi Xu, and Jiaya Jia. Image smoothing via 10 gradient minimization. In Proceedings of the 2011 SIGGRAPH Asia conference, pages 1-12, 2011. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 129, + 294, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 152 + ], + "type": "text", + "content": "[32] Qi Zhang, Li Xu, and Jiaya Jia. 100+ times faster weighted median filter (wmf). In CVPR, pages 2830-2837, 2014. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 153, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 153, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 153, + 294, + 195 + ], + "type": "text", + "content": "[33] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 198, + 294, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 242 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 242 + ], + "type": "text", + "content": "[34] Man Zhou, Keyu Yan, Xueyang Fu, Aiping Liu, and Chengjun Xie. Pan-guided band-aware multi-spectral feature enhancement for pan-sharpening. IEEE Transactions on Computational Imaging, 9:238-249, 2023. 6" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14731" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/FaceCraft4D_ Animated 3D Facial Avatar Generation from a Single Image/c512f2aa-6fbd-4940-89be-2245ba51ffe2_content_list.json b/2025/FaceCraft4D_ Animated 3D Facial Avatar Generation from a Single Image/c512f2aa-6fbd-4940-89be-2245ba51ffe2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b1ecd8b3496977312cbfdd9e339ca2c0d2d48420 --- /dev/null +++ b/2025/FaceCraft4D_ Animated 3D Facial Avatar Generation from a Single Image/c512f2aa-6fbd-4940-89be-2245ba51ffe2_content_list.json @@ -0,0 +1,1677 @@ +[ + { + "type": "text", + "text": "FaceCraft4D: Animated 3D Facial Avatar Generation from a Single Image", + "text_level": 1, + "bbox": [ + 124, + 130, + 875, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fei Yin $^{1,2}$ , Mallikarjun B R $^{2}$ , Chun-Han Yao $^{2}$ , Rafal K. Mantiuk $^{1\\dagger}$ , Varun Jampani $^{2\\dagger}$ , $^{1}$ University of Cambridge, $^{2}$ Stability AI", + "bbox": [ + 169, + 179, + 826, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 251, + 328, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present a novel framework for generating high-quality, animatable 4D avatar from a single image. While recent advances have shown promising results in 4D avatar creation, existing methods either require extensive multiview data or struggle with shape accuracy and identity consistency. To address these limitations, we propose a comprehensive system that leverages shape, image, and video priors to create full-view, animatable avatars. Our approach first obtains initial coarse shape through 3D-GAN inversion. Then, it enhances multiview textures using depth-guided warping signals for cross-view consistency with the help of the image diffusion model. To handle expression animation, we incorporate a video prior with synchronized driving signals across viewpoints. We further introduce a Consistent-Inconsistent training to effectively handle data inconsistencies during 4D reconstruction. Experimental results demonstrate that our method achieves superior quality compared to the prior art, while maintaining consistency across different viewpoints and expressions.", + "bbox": [ + 89, + 282, + 483, + 570 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 599, + 220, + 616 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4D Avatar generation aims to create animatable 3D head avatars, enabling the generation of consistent identities with custom viewpoints and expressions. High-quality 4D heads have significant applications in gaming, education, and film industries. However, existing 4D avatar building methods [18, 23, 27, 36] often require extensive inputs, such as multiview video with camera pose estimation, which is impractical for widespread use. Furthermore, these methods struggle to produce realistic results for extreme viewpoints and expression changes due to data scarcity, for instance the back view of the head.", + "bbox": [ + 89, + 625, + 483, + 790 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The goal of this work is to create a 4D facial avatar from just a single image. This task is challenging because a single image provides limited information, missing essential details like depth and multiple viewpoints that are crucial", + "bbox": [ + 89, + 792, + 483, + 853 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e2954b1f9e34b45a8d107e831d9d4f06ee11db2319c7dce3e2ac4d0cdd74f3b3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 250, + 924, + 362 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7d796904c96fbfd3ab6d848a903ba91d7327258b56d8dc689b592794e70319ae.jpg", + "image_caption": [ + "Figure 1. Given a single image input, our method is capable of generating 4D avatars, producing photorealistic textures and consistent expressions across multiple views. Our approach also demonstrates robust performance on challenging inputs, including cartoon characters or old photographs." + ], + "image_footnote": [], + "bbox": [ + 517, + 364, + 921, + 474 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "for constructing a complete 3D model. A single image cannot provide information about the unseen parts of a person's head or capture dynamic aspects like facial expressions. It is also impractical to train an end-to-end model due to the lack of multiview full-head animation datasets. To address these challenges, the system needs to rely on prior knowledge, like shape and expression.", + "bbox": [ + 511, + 566, + 908, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing single-image avatar generation methods come with many shortcomings, which we summarize in Tab. 1. The works that rely on a single image of novel identity at test time, typically utilize large-scale, unstructured 2D images [1] and videos [7, 39] during training to account for multiview and temporal aspects. Due to the practical challenges in collecting multiview full-head animation datasets, most single-image methods, including recent work [35], cannot reconstruct a complete 360-degree view of the head. While some approaches can generate full views [1, 40], they typically lack the ability to model facial expressions. Among methods that can animate, some methods rely solely on 2D representation [39], which makes these methods multi-view inconsistent by design. To overcome the above limitation, some methods rely on a hybrid representation", + "bbox": [ + 511, + 674, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Equal advising.", + "bbox": [ + 107, + 862, + 202, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*No biometric data was used to train, validate, or evaluate the model described in this work.", + "bbox": [ + 89, + 875, + 482, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11612", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/26e68ba3bfc2cda36fc6f0e62c01bd72b55253b0f14cc77f6b97a2c5604769ea.jpg", + "table_caption": [ + "Table 1. Comparison of methods for avatar generation and animation. The first two methods focus on general object generation, while the remaining methods are specifically designed for avatar applications." + ], + "table_footnote": [], + "table_body": "
MethodsGoalDimensionInput360° viewAnimation3D ConsistencyQuality
LDM [32]Image Generation2DSingle Image-High
DreamFusion [26], SV3D [37]Image-to-3D3DSingle Image✓LowLow
LivePortrait [10], AniPortrait [39]Video Generation2DSingle Image✓LowHigh
PanoHead [1], SPI [44]3D GAN Inversion3DSingle Image✓MediumMedium
Portrait4D [7], CAP4D [35]3D Talking Head3DSingle Image✓MediumHigh
FaceCraft4D4D AvATAR Generation3DSingle Image✓✓HighHigh
HQ3D [36], GA [27]4D AvATAR Generation3DMulti-view Videos✓✓HighHigh
", + "bbox": [ + 117, + 127, + 883, + 243 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "that combines 2D and 3D [7, 34]. However, such methods cannot handle extreme camera angles, as they rely on 2D components for high-resolution synthesis. In contrast, we aim to build our model that has pure 3D representation, and, therefore, is multiview-consistent by design.", + "bbox": [ + 88, + 261, + 480, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose FaceCraft4D, which takes a single image as input and creates a 3D Facial Avatar that can be animated and rendered in 360-degree view, as shown in Fig. 1. To tackle such an under-constrained problem, we make use of 3 different types of priors: a geometry prior, an image prior, and a video prior. We use these priors to synthesize personalized multiview images of the given identity, which are then used to train an explicit 3D model of an avatar that can be animated (see Fig. 2). We will refer to such a model as a 4D avatar.", + "bbox": [ + 88, + 337, + 482, + 487 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We start with a geometry prior from pretrained 3D-GAN [1]. Since, 3D-GAN is trained with large-scale images that comprises of all the views of the head, it has information about the complete head shape. Through 3D-GAN inversion [31], we obtain personalized shape along with an approximate texture. The resulting 3D representation lacks personalized high-quality multiview consistent texture and can not be used for creating novel expressions or animations (see Tab. 1). To improve the quality of the textures, we use of a 2D image prior. To ensure multi-view consistency and preservation of identity across the views, we employ epipolar constraints and cross-view mutual attention. The resulting multi-view images can be used to build a high-quality static 3D head model, which, however, lacks control over expressions and articulations. To animate the avatar, we utilize a 2D video prior [10], which generates training data for facial animation across different viewpoints.", + "bbox": [ + 91, + 489, + 482, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, for real-time rendering performance, we train a 3D Gaussian representation [16] that is rigged to a FLAME parametric face model [20, 27] using the multi-view videos obtained in the previous steps. To further improve the training consistency across the views and to avoid blur, we propose a COnsistent-INconsistent (COIN) training, which can capture cross-view inconsistencies in an MLP.", + "bbox": [ + 89, + 747, + 482, + 851 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, the contributions of our work are as follows:", + "bbox": [ + 109, + 854, + 480, + 868 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We present a novel framework for generating high-quality 4D head models from a single image of a face. We com", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "pare our framework with previous methods and summarize the differences in Tab. 1.", + "bbox": [ + 524, + 261, + 903, + 289 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce novel warping-based control generation signals to improve quality and consistency across different views and expressions.", + "- We introduce COIN training, which enables a 3D Gaussian model to learn from inconsistent data while preserving high-quality, consistent features." + ], + "bbox": [ + 511, + 291, + 903, + 381 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 398, + 653, + 414 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Head generation. Numerous studies have explored the reconstruction of 3D head models. To obtain sufficient information, most approaches [9, 27, 48, 48] utilize monocular video sequences, which provide multiple viewpoints of the subject's head. However, a significant limitation of these methods is their reliance on multiple viewpoints and lengthy video inputs. They also struggle to accurately reproduce less prominent features in videos, such as the back of the head.", + "bbox": [ + 511, + 425, + 906, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another approach is to train a 3D generator on a large head image dataset. EG3D [3] proposed a hybrid tri-plane representation that can quickly and efficiently render high-quality facial images. Panohead [1] and Portrait3D [40] expanded on EG3D by increasing the number of planes, enabling the generation of 3D heads from limited angles to full 360-degree views. However, most methods in this category rely on hybrid representation having a 2D super-resolution network. This impacts their generalization to extreme camera poses and suffer identity inconsistency across views.", + "bbox": [ + 511, + 563, + 905, + 712 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, the rise of text-to-image models has spurred the development of novel approaches for text-to-3D generation tasks. One such approach is Score Distillation Sampling (SDS) [26, 29], which leverages the score function to distill diffusion priors into 3D representations, enabling the generation of 3D content from text or image inputs. Headsculpt [11, 49] incorporates facial keypoint information into the text-to-3D generation method, effectively constraining the facial layout. However, the texture generated by these methods can be oversaturated, making it challenging to achieve photo-realistic head models.", + "bbox": [ + 511, + 714, + 905, + 880 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Head animation. Head animation techniques can be", + "bbox": [ + 511, + 885, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11613", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fcc560db6db8a53d342c1d8369e5acf4d14a52206f7b43ce07b7c1995d2cb790.jpg", + "image_caption": [ + "Figure 2. Overview of FaceCraft4D: Our approach begins with estimating the geometry of a single input image using a Geometry Prior. This geometry guides the synthesis of personalized multiview images, providing $360^{\\circ}$ views and varied expressions, with support from both 2D image and video priors. Since the synthesized data often exhibit inconsistency across views, we propose COIN optimization for robust 4D optimization. By fitting the model to the multiview data, we achieve our final 4D avatar. All faces in this manuscript come from the public FFHQ dataset [15]." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 897, + 270 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "broadly categorized into 2D-based and 3D-aware approaches, distinguished by their utilization of explicit camera models for rendering. Recent methods [8, 13, 30, 43] have predominantly adopted warping-based strategies, applying deformation fields to intermediate appearance features to capture facial motion characteristics. While effective for frontal and moderate poses, these approaches struggle to maintain geometric consistency during large pose variations due to their inherent lack of 3D modeling.", + "bbox": [ + 88, + 363, + 480, + 498 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To enable free-viewpoint rendering with stronger 3D consistency guarantees, researchers integrated explicit 3D representations into the head synthesis pipeline. Early approaches [17, 41] relied on 3D morphable models for representing head geometry and texture. Subsequent works [5, 14, 21, 42, 45] leveraged more sophisticated representations such as Neural Radiance Fields (NeRFs) [25] to better capture complex structures like hair and accessories. Recent advances like GAGAvatar [4, 22], which employs 3D Gaussian splatting [16], have demonstrated impressive performance in both rendering quality and speed. However, these methods are typically limited in their ability to generate full 360-degree views. Our approach enables true 360-degree view synthesis while maintaining real-time rendering capabilities and achieving superior animation quality.", + "bbox": [ + 89, + 500, + 482, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 739, + 181, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we propose a novel framework for generating high-fidelity 4D avatars from a single portrait image $I_{R}$ . Reconstructing a 3D dynamic model from a single 2D image is an inherently ill-posed problem, as critical 3D and temporal information is missing. To address this challenge, we leverage several 2D and 3D priors to compensate for the missing data.", + "bbox": [ + 89, + 763, + 482, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall pipeline, shown in Fig. 2, consists of two stages: personalized multiview generation and 4D repre", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sensation optimization. In the Personalized Multiview Generation stage, we utilize geometry priors to estimate the 3D geometry (Sec. 3.1), use image priors to enhance the texture (Sec. 3.2), and video priors to synthesize high-quality, consistent multi-view videos with dynamic expressions (Sec. 3.3). We then leverage this synthetically-generated personalized image set to optimize a robust 4D representation model (Sec. 3.4). The trained 4D model can be animated by the control parameters of a morphable model, enabling the generation of high-fidelity 4D avatars from a single portrait image.", + "bbox": [ + 511, + 363, + 906, + 530 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Geometry Prior", + "text_level": 1, + "bbox": [ + 511, + 578, + 671, + 594 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The initialization is crucial for 3D generation. A good initialization should follow the approximate geometry of the head shown in the input image, $I_{R}$ . For that, we use the prior provided by a 3D-aware generator, PanoHead [1, 40], which was trained on a large $360^{\\circ}$ head dataset. We obtain a coarse 3D reconstruction for a given input image $I_{R}$ through GAN inversion. Specifically, we first optimize the latent code $z$ within the generator's domain by minimizing a combination of pixel-wise $\\mathcal{L}_2$ loss and image-level LPIPS loss [47]. Subsequently, following [31], we keep the optimized latent code $z$ fixed while fine-tuning the 3D generator's parameters. During this fine-tuning stage, we maintain the same loss functions used in the latent optimization phase to ensure the rendered views closely match the input image. This process gives us an initial 3D head representation, though one that cannot be animated. The optimized generator is then used to synthesize a set of approximate multiview images $\\{I_i\\}$ and their corresponding depth maps $\\{D_i\\}$ , where $i$ is the view index.", + "bbox": [ + 511, + 613, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11614", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2bf14056c35e2f198145d4363f0ed09d58925f35166af031f7950102ec378c51.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 93, + 89, + 495, + 361 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0c3aebe297ab434d9984668cebe207fbe7b271a8818fb8f306b617d5cce678d0.jpg", + "image_caption": [ + "Figure 3. Image Prior: We introduce a cross-view mutual attention mechanism and epipolar constraints to enhance consistency in generated novel views. Our approach aligns reference and target images, maintaining visual coherence across viewpoints.", + "Figure 4. Triplane-based methods (e.g. PanoHead) are sensitive to focal length (scale) and exhibit significant degradation of quality as focal length decreases. In contrast, our Gaussian-based methods maintain robust performance across varying focal lengths." + ], + "image_footnote": [], + "bbox": [ + 94, + 431, + 496, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Image Prior", + "text_level": 1, + "bbox": [ + 89, + 709, + 220, + 726 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The textures synthesized from the geometry prior may lack visual quality, and the identity of the individual may not be well preserved across the views. In particular, when rendering with different focal lengths, triplane-based methods [1] exhibit substantial degradation in image quality, as illustrated in Fig. 4. We enhance the generated multiview images using 2D diffusion models [32], which have demonstrated powerful generative capabilities. Although direct image-to-image translation [24] appears to be a straightforward solution, this approach poses significant risks: the de-noising process may introduce cross-view inconsistencies", + "bbox": [ + 89, + 734, + 485, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and also modify the original semantic content. To address these risks, we propose two strategies that constrain the generation process, which we describe next and illustrate in Fig. 3.", + "bbox": [ + 511, + 90, + 906, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Cross-view mutual attention. Inspired by MasaCtrl [2], we introduce a cross-view attention mechanism to replace the standard self-attention during the generation process. This modification allows us to inject information from the reference image into the novel view. Specifically, we first perform DDIM inversion [33] to introduce noise to both the reference frame $I_{R}$ and the coarse image $I_{i}$ of the target novel view. This step preserves the overall face structure and contour. Then, the diffusion model is used to gradually denoise two images to enhance the texture. In this process, we replace the key $K$ and value $V$ matrices (calculated in self-attention) of the novel view with those derived from the reference image. Different from [2], we adopt a batch-processing strategy in which multiple images $I_{i}$ representing different viewpoints are treated as a single batch, with the reference image $I_{R}$ serving as the consistent source for all subsequent calculations. This not only streamlines the generation process but also ensures uniformity in feature injection across novel viewpoints.", + "bbox": [ + 511, + 155, + 908, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The rationale behind this approach lies in the observation that, despite variations in viewpoint, the underlying textures (represented by the attention values) should maintain consistency across the views. The cross-view attention mechanism enables the transfer of information from the reference to the novel views, enhancing both the fidelity and coherence of the generated novel views.", + "bbox": [ + 511, + 443, + 908, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Warping-based control generation. The consistency across the views is further enforced by epipolar constraints, which enforce geometric relationships between novel views and the input view. We first use cross-view mutual attention to generate the view of the back of the head. The generated back view along with the reference image are set as anchors. Then, we propagate the texture information to novel views through depth-based warping. Specifically, we lift the anchor image into 3D space and then project it to the adjacent view with depth guidance $\\{D_i\\}$ . During the projection, we form a mask to indicate the visibility of pixels to filter invisible areas using the rendered depth values. The projected image is then used to guide the texture enhancement process. The intermediate latent of the enhanced image in the diffusion model is blended with the latent of the projected image via a simple mask copy-pasting operation. By combining the latent, the resulting images benefit from both the improved consistency provided by the epipolar constraints and the enhanced quality achieved through the denoising process. We repeat this process for various azimuth angles, creating a comprehensive set of viewpoints.", + "bbox": [ + 511, + 551, + 908, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the aforementioned guidance and epipolar constraints, we can ensure coherence across the entire view", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11615", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/377bb3b1d7a5824f47a2fdc08872b959557f55bd95fa37e26c74b0c9c011193d.jpg", + "image_caption": [ + "Figure 5. COIN-Training: To address geometric and color inconsistencies between the views, we train two 3D representations: a view-consistent base model (GaussianAvatar) and an MLP, encoding inconsistencies between the views. The MLP with inconsistencies lets us robustly reconstruct the high-quality 3D base model." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 495, + 253 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "space and generate high-fidelity view-consistent images $\\{I_i^*\\}$ for further reconstruction.", + "bbox": [ + 89, + 363, + 482, + 395 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Video Prior", + "text_level": 1, + "bbox": [ + 89, + 404, + 217, + 420 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The view-consistent data created in the previous step has a static expression similar to that of reference image $I_R$ . This may not be enough to create diverse expressions as the images do not contain some of the expression-dependent information. For example, if the reference image has the mouth closed, the synthesized dataset would not have details inside the mouth region. To incorporate expression-dependent textures and geometry that may not be available in the multiview image generation stage (Sec. 3.2), we need to create multi-view facial expression data. We use LivePortrait [10] as a video prior, which can help generate multiview expressive videos. LivePortrait uses a single source image and a driving video to generate new videos where the source identity mimics the expressions and poses from the target video. Instead of directly copying the driving video's motion, one has the option of capturing the changes in relative expression but maintain the pose of the input image. Given that we have the reference and multiview-consistent enhanced images $\\{I_i^*\\}$ from the previous step, we provide these to LivePortrait as the source images together with driving videos with diverse poses and expressions (details in Sec. 4), and obtain target videos. Because the identity is consistent across the input views, the resulting videos also maintain the identity. Using identical driving videos ensures synchronized expressions.", + "bbox": [ + 89, + 428, + 483, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Robust 4D optimization with COIN training", + "text_level": 1, + "bbox": [ + 89, + 816, + 467, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The synthesized multi-view videos have high-quality textures and good cross-view consistency, but small misalignment of features and colors is unavoidable. Learning 4D representation directly on such data leads to blurry textures", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for the rendered objects. Here, we address this problem by proposing a robust training paradigm. At a high level, we learn a base representation that is common to all views, and a delta representation that is specific to each view during training. This helps in cornering inconsistencies concerning each view to separate representation. The base model will have a coarse but consistent representation learned, while the high-frequency details end up in the delta-space representation. At test time, we choose one of the views as our reference view and use its delta representation as the common high-frequency details for all the views.", + "bbox": [ + 511, + 90, + 903, + 256 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In specific, we achieve this by jointly training two 4D representations: a consistent GaussianAvatar [27], which represents the 3D structure using Gaussians bound to a FLAME-tracked [20] mesh, and a NeRF-style MLP, tasked with encoding inconsistencies between the views. As shown in Fig. 5, we design the losses that prioritize structure reconstruction in the consistent GaussianAvatar (LPIPS), and high-frequency detail reconstruction in the MLP with the view-dependent inconsistencies (L1 and SSIM). By isolating view-dependent variations in a separate model, this strategy prevents these inconsistencies from corrupting the base representation. The approach can be considered a form of robust regression, with the key difference that the view-dependent MLP ensures that inconsistencies (outliers) are spatially localized. This cannot be achieved with robust norms (e.g., L1), which do not maintain the information about the spatial position of outliers.", + "bbox": [ + 511, + 257, + 906, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We choose GaussianAvatar as our consistent representation because it lets us animate the avatar with controllable FLAME expression and pose parameters, providing a solid foundation for modeling consistent geometric and appearance features. The details of FLAME tracking and controlling are explained in Sec. 1.3 of the supplementary. To model the view-dependent variations, we design a lightweight MLP-based architecture that introduces minimal computational overhead while effectively capturing view-specific details. The network encodes the offset in the color of each Gaussian (see Fig. 5), which can be formulated as:", + "bbox": [ + 511, + 513, + 905, + 691 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nc _ {\\text {o f f s e t}} = \\operatorname {M L P} \\left(e _ {\\text {v i e w}}, c, e _ {g}; \\theta\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 696, + 903, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $e_{\\mathrm{view}} \\in \\mathbb{R}^d$ is a learnable view-specific embedding that encodes view-representation, $c$ represents the initial Gaussian color attributes from the GaussianAvatar, $e_g$ is a position embedding for each Gaussian, accounting for location-specific color variation ranges (e.g., eyes vs. mouth regions), and $\\theta$ is a vector of the trainable parameters. The final view-dependent color is computed as the sum of the base GaussianAvatar color and the offset: $c^* = c + c_{\\mathrm{offset}}$ . The MLP consists of only two layers to maintain computational efficiency and preserve the real-time rendering capabilities of the Gaussian-based representation.", + "bbox": [ + 511, + 718, + 905, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During training, we render two types of images: $I_{i}^{\\mathrm{C}}$ from", + "bbox": [ + 531, + 883, + 903, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11616", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7cce182327f9a8e2e3eb4fa9477028bf878112b3a7cb62d60a165e40a2ded385.jpg", + "image_caption": [ + "Figure 6. Qualitative comparison on static 3D head generation from a single image." + ], + "image_footnote": [], + "bbox": [ + 99, + 85, + 890, + 448 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the consistent component alone, and $I_{i}^{\\mathrm{IC}}$ from the combination of both components. To ensure accurate reconstruction of high-frequency details, we supervise $I_{i}^{\\mathrm{IC}}$ using pixel-wise losses:", + "bbox": [ + 89, + 491, + 483, + 551 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p i x e l}} = \\lambda_ {1} \\mathcal {L} 1 \\left(I _ {i} ^ {\\mathrm {I C}}, I _ {i} ^ {*}\\right) + \\lambda_ {\\text {S S I M}} \\text {S S I M} \\left(I _ {i} ^ {\\mathrm {I C}}, I _ {i} ^ {*}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 563, + 482, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To prevent the inconsistent component from dominating the representation and maintain the effectiveness of the consistent component, we introduce a structural supervision loss and a regularization term:", + "bbox": [ + 89, + 593, + 483, + 654 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s t r u c}} = \\lambda_ {\\text {L P I P S}} \\operatorname {L P I P S} \\left(I _ {i} ^ {\\mathrm {C}}, I _ {i} ^ {*}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 669, + 482, + 688 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e g}} = \\lambda_ {\\text {o f f s e t}} \\mathcal {L} 1 \\left(c _ {\\text {o f f s e t}}, 0\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 710, + 482, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where LPIPS loss ensures image-level supervision for the consistent component, and the offset regularization encourages minimal view-dependent modifications. The hyperparameters are set to $\\lambda_{1} = 0.8$ , $\\lambda_{\\mathrm{SSIM}} = 0.2$ , $\\lambda_{\\mathrm{LPIPS}} = 0.05$ and $\\lambda_{\\mathrm{offset}} = 1$ .", + "bbox": [ + 89, + 733, + 482, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During inference, we input a single fixed view-embedding for the inconsistent component. Given that our reference image $I_R$ contains superior quality and identity information compared to the generated multiview images $\\{I_i^*\\}$ , we select its corresponding view embedding for inference. Novel expressions and poses can be easily gen", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "erated by adjusting the FLAME parameters, leveraging the animation capabilities of our backbone model.", + "bbox": [ + 511, + 491, + 906, + 522 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 537, + 645, + 555 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. During the multiview generation stage, we use PanoHead [1] as our geometry prior, Cosmicman [19] as the image prior, and LivePortrait [10] as the video prior model. We render 24 views along the horizon orbit for static reconstruction. For animation, we select $I_R$ and 11 frontal views from these generated images as source images and randomly sample 8 video clips from the NerSemble [18] dataset as driving videos. The source images and driving videos are then input to the video prior model to produce multiview video data. In the 4D optimization stage, we first optimize the GaussianAvatar [27] model with the static data over 30K iterations. Next, we apply COIN-optimization techniques, finetuning the static Gaussian model over 90K iterations. The complete process takes approximately two hours to generate a single 4D avatar.", + "bbox": [ + 511, + 564, + 906, + 790 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. In the absence of ground truth data, we use commonly adopted non-reference metrics to evaluate the rendered images in terms of image quality and identity similarity. For overall image quality, we use the FID [12] score. To assess multi-view consistency, we employ the CLIP-I [28] score. For identity preservation, we calculate ID scores by measuring the cosine distance of face", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11617", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5e1438b7557703f1f8df2d2be930093745f1411125ca5f79cae0b730defa889f.jpg", + "image_caption": [ + "Figure 7. Qualitative comparison on animation of different views." + ], + "image_footnote": [], + "bbox": [ + 117, + 85, + 305, + 277 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6d13b9e1d13f030a9266c8ef6892f310f5e90c5cc05abeafb50e4cf4161d9c99.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 308, + 85, + 401, + 277 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b109ea62d5b353691c4d537a77651f346c66ee6e069629cfb0a78faf14676417.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 85, + 591, + 277 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7a7b66431ef2e601978661e80ed7d1132b11c58ba801eff2abdb3ae6e7a9238e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 594, + 85, + 802, + 277 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4853a7dc0f8bc84c4ab8f30cdd89f286eada173252fea01c25f6ca351732ca09.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 803, + 85, + 901, + 277 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8363dcdef69382e3710e1387c21569021842e0eb4fe0994d09be6e2d62d2d11a.jpg", + "table_caption": [ + "Table 2. Quantitative comparison on static 3D head generation." + ], + "table_footnote": [], + "table_body": "
MethodCLIP-I ↑ID ↑FID ↓
GaussianCube [46]0.68300.4300258.81
PanoHead [1]0.82330.4246195.28
SV3D [37]0.76560.4331234.86
Portrait3D [40]0.70660.3719302.74
Ours0.80530.5082174.36
Table 3. Quantitative comparison on 3D head animation.
MethodCLIP-I ↑ID ↑FID ↓
AntiPortrait [39]0.46530.4171364.99
Portrait-4D-V2 [7]0.52360.4592248.36
Ours0.57370.4602201.76
", + "bbox": [ + 96, + 325, + 477, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "recognition features [6] between the novel views and the reference images.", + "bbox": [ + 89, + 520, + 482, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Datasets. We conducted our analysis on a subset of facial images from FFHQ [15] dataset, a high-quality, in-the-wild facial dataset. For evaluation, we select 100 images with unoccluded faces to evaluate our model and baselines.", + "bbox": [ + 89, + 554, + 483, + 614 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Static 3D Generation Comparisons", + "text_level": 1, + "bbox": [ + 89, + 627, + 395, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. We compare our method with existing image-to-3D approaches, including a general object-focused model, SV3D [37], and three head-specific methods: the inference-based GaussianCube [46], the 3D-GAN-based PanoHead [1], and the optimization-based Portrait3D [40].", + "bbox": [ + 89, + 654, + 482, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative results. Fig. 6 shows visual comparisons where our method achieves the highest visual fidelity and identity coherence compared to other approaches. Notably, it preserves details such as teeth and earrings in side views. In contrast, SV3D, which focuses on general objects, struggles to capture accurate head geometry, often degenerating into a flat plane. GaussianCube tends to \"toonify\" the input head due to the nature of its training data. While PanoHead achieves good visual quality, it suffers from detail loss and floating artifacts during pose rotation. Portrait3D also falls short in quality, as the SDS [26] optimization introduces", + "bbox": [ + 89, + 734, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "variations that lead to blurring artifacts.", + "bbox": [ + 511, + 314, + 774, + 330 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative results. For each generated head model, we render images from five viewpoints: left-frontal, left, right-frontal, right, and frontal. All five rendered images, along with the reference image, are used to compute the CLIP-I, ID, and FID metrics. As shown in Tab. 2, our method outperforms others in both ID and FID scores. This aligns with the observed visual fidelity, as our results appear more realistic and coherent. Our method achieves higher identity similarity and better detail preservation, even in side views. The results for the MEAD multiview dataset [38] can be found in the supplementary material.", + "bbox": [ + 511, + 334, + 906, + 501 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Animation Comparisons", + "text_level": 1, + "bbox": [ + 511, + 515, + 738, + 531 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. We compare our method with one-shot video-based head reenactment approaches, including the 2D-based AniPortrait [39] and the 3D-based Portrait-4D-V2 [7]. For AniPortrait, which uses keypoints as intermediate driving signals, we modify the Euler angles of the expression image to rotate the keypoints and generate novel view videos. In the case of Portrait-4D-V2, novel views can be generated directly by editing the camera parameters.", + "bbox": [ + 511, + 542, + 906, + 664 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative results. As shown in Fig. 7, our method preserves expressions with greater accuracy. Portrait-4D-V2 lacks a back-of-head model, so when rotated to side views, the back appears empty. AniPortrait, as a 2D-based method, can achieve high texture quality. However, with large-angle rotations, AniPortrait struggles to preserve identity details and introduces noticeable artifacts in the hair and background. In contrast, our method maintains accurate geometry and expression, even with camera movement.", + "bbox": [ + 511, + 667, + 908, + 805 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative results. We test our data on unseen views (especially $60^{\\circ} - 180^{\\circ}$ ), to highlight each method's performance under varying viewing angles, as handling extreme rotations is often challenging. The results are presented in Tab. 3. Our method outperforms all baselines, demonstrating robustness across varied views. The performance", + "bbox": [ + 511, + 809, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11618", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/313be90bd12dd852f03f64c64227a372a430ba2ab777a5d8e601c53b8ba97d51.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 171, + 165 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c122fd351ad377d65f478f329105424842d4f988e2d4e2d97c02254a613d3541.jpg", + "image_caption": [ + "Warped Image" + ], + "image_footnote": [], + "bbox": [ + 187, + 90, + 256, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bcd261144454207b004104adb87ef4c5910f455d7c766f99d3fb9c32de343a50.jpg", + "image_caption": [ + "w/o Warping" + ], + "image_footnote": [], + "bbox": [ + 264, + 90, + 334, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/17e8e167a194f0873557dd97b390d9bf23e21edce6eb5566aaa09439d85bdcc0.jpg", + "image_caption": [ + "w/o MA" + ], + "image_footnote": [], + "bbox": [ + 346, + 90, + 416, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/88c00eaea3dac1381f6c9e0670bcac73027c4fa382bf7416aeb64afbb342edaf.jpg", + "image_caption": [ + "Full" + ], + "image_footnote": [], + "bbox": [ + 426, + 90, + 496, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ed7841e80558856f152deb92b73f1cbaf90b5d5513955e9f8a345126ad1d2b4a.jpg", + "image_caption": [ + "Figure 8. Ablation of multiview image generation modules." + ], + "image_footnote": [], + "bbox": [ + 102, + 194, + 184, + 277 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e50304eb204c57a4f80602f2869503c39f1193f82eb80be5a54a72704d394212.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 205, + 194, + 297, + 272 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/98c1170d45a8510af70f4e188cbff88bcb97da8c554d8f0d8a23f4babf0f98a4.jpg", + "image_caption": [ + "w/o Video Prior" + ], + "image_footnote": [], + "bbox": [ + 196, + 272, + 295, + 306 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7aabae2050151d565e62a780db6775c03367b5eec66715a7e74586ade5be315d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 307, + 194, + 390, + 273 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b3b2915c9d660ab5e193b688260a8e4ca015760adf79932434949553c1592ef9.jpg", + "image_caption": [ + "w/o COIN" + ], + "image_footnote": [], + "bbox": [ + 297, + 272, + 397, + 306 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dc449449143426ae99c281625dd1028f06f3ce8b31f53ea2c69d790baec6d52f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 194, + 496, + 273 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/08c1a764200f0d51043005c6386236ed3e06c6ec8600f4d80f6427aefe37419a.jpg", + "image_caption": [ + "Full" + ], + "image_footnote": [], + "bbox": [ + 400, + 272, + 496, + 306 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5c196a350c7472c310176fc8ce70aeb206faae6cbd826a59d637ca825c5e440c.jpg", + "image_caption": [ + "Input", + "Figure 9. Ablation of animation modules.", + "Input", + "Novel View" + ], + "image_footnote": [], + "bbox": [ + 96, + 343, + 163, + 402 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4127dca5d723113e4c933bb28e37d4e7c6c17bcc2a951a08f3485c52f1eac1f9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 165, + 344, + 230, + 402 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a107644927fae515ea994a11560366e629251892ccfb7ad688dcc80cc905d59b.jpg", + "image_caption": [ + "Input", + "Novel View", + "Figure 10. More results with diverse inputs." + ], + "image_footnote": [], + "bbox": [ + 232, + 345, + 310, + 405 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3dbd67799bd7684a42045d51ab257f56ca9a16e74dd2b53a19969db5d7baccfd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 310, + 345, + 375, + 405 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7868cc0faa3507170f6b59803cd8692bb4cefbaf9270033afc15e7cdbd91e0e5.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 377, + 345, + 434, + 402 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6392f69b9f4e1c2c849e3cfdc472430ea8278a00964b2454426cbea5b161daea.jpg", + "image_caption": [ + "Novel View" + ], + "image_footnote": [], + "bbox": [ + 439, + 345, + 493, + 401 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "of baselines degrades significantly as the viewing angle increases. Additional results and baseline comparisons are provided in the supplementary material.", + "bbox": [ + 89, + 446, + 482, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 503, + 243, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multiview image generation. We conduct an ablation study on two multi-view generation modules, as shown in Tab. 4 and Fig. 8. Incorporating the warping module allows us to preserve finer details, like tattoos, in novel views. While removing MA may lead to semantic inconsistencies, such as changes in gender. Adding mutual attention (MA) further improves fidelity and enhances identity similarity by injecting features from the reference images.", + "bbox": [ + 89, + 529, + 483, + 651 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Animation. Our 3D representation is based on GaussianAvatar [27], which is FLAME-driven, enabling direct animation through FLAME parameters. To evaluate the base performance, we remove our video prior module. As shown in Fig. 9, without the video prior, the final avatar fails to synthesize accurate expressions. For consistency modeling, we compare our model with and without the COIN training. Without COIN, training directly on inconsistent generated images results in blurred textures. With COIN training, regions such as the teeth remain sharp, and finer details, like individual hair strands, are preserved. The quantitative results in Tab. 5 further validate the effectiveness of the COIN optimization across novel views.", + "bbox": [ + 89, + 655, + 483, + 851 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Processing times. We analyzed the processing times of each stage. Our method consists of three stages: a) 3D GAN inversion (2 mins); b) Multiview image & video generation", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/79beeb0797be7a29a7fc8baa5f0a6766e80cdc56c1e27e8bcab154cf0bb90f8e.jpg", + "table_caption": [ + "Table 4. Ablation of multiview image generation modules. \"Warp\" refers to the warping-based control generation (Sec. 3.2) and MA to cross-view mutual attention (Sec. 3.2)." + ], + "table_footnote": [], + "table_body": "
MethodCLIP-I ↑ID ↑FID
w/o Warp w/o MA0.73280.4787182.27
w/o Warp0.78860.4915171.08
w/o MA0.81510.4951172.43
Full0.81620.4984166.96
", + "bbox": [ + 516, + 130, + 903, + 215 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3324c71000ce7abaf0fc0db71d278e8dc4d2e2decf6d7c5ce85ddcf379db5d27.jpg", + "table_caption": [ + "Table 5. Ablation of animation modules." + ], + "table_footnote": [], + "table_body": "
MethodCLIP-I ↑ID ↑FID ↓
w/o COIN0.76880.4952144.95
Full0.77290.5010142.80
", + "bbox": [ + 516, + 231, + 903, + 287 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e75873be5fa38fa9abc591503832299e89d4e50a7fd360efb479dd618ace1a96.jpg", + "table_caption": [ + "Table 6. Time comparison with optimization-based avatar reconstruction methods." + ], + "table_footnote": [], + "table_body": "
MethodNerFaceIMAvatarPointAvatarHeadStudio
Time (hour)54h48h6h2h
MethodMonoGAGaussianAvatarHq3davatarOurs
Time (hour)7h2h12h2.5h
", + "bbox": [ + 519, + 319, + 900, + 376 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(10 mins), FLAME parameters extraction (10 mins); and c) COIN-training to overfit multiview videos (2 hours). The generation of a single avatar $\\sim 2.5$ hours, which is comparable to or faster than other optimization-based avatar reconstruction methods, as shown in Tab. 6. At inference, we achieve real time rendering at 156 FPS with a resolution of $512 \\times 512$ .", + "bbox": [ + 511, + 388, + 905, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Diverse inputs. To validate robustness, we conduct experiments on diverse inputs, including cartoon, line drawing, and portraits with extreme camera poses. The results in Fig. 10 demonstrate consistent performance across various domains. Additional video results are available on the webpage in the supplementary materials.", + "bbox": [ + 511, + 500, + 905, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 603, + 638, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present a novel method for generating high-quality 4D portraits from a single image. By leveraging multiple priors, including 3D-GAN for geometry initialization, image prior for texture enhancement, and video prior for animation, our method can generate animated avatars with detailed textures, excellent consistency, and plausible rendition of views missing in the input. The proposed depth-guided warping ensures cross-view consistency during the generation process, while our COIN-training strategy enables high-quality 4D reconstruction from potentially inconsistent multi-view data. Extensive experiments demonstrate that our method outperforms existing approaches in terms of geometry accuracy, texture quality, and identity preservation across different viewpoints and expressions. Our framework makes high-quality 4D portrait generation more accessible by removing the requirement for multi-view data capture, paving the way for broader applications.", + "bbox": [ + 511, + 628, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11619", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sizhe An, Hongyi Xu, Yichun Shi, Guoxian Song, Umit Y Ogras, and Linjie Luo. Panohead: Geometry-aware 3d full-head synthesis in 360deg. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20950-20959, 2023. 1, 2, 3, 4, 6, 7", + "[2] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22560–22570, 2023. 4", + "[3] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16123-16133, 2022. 2", + "[4] Xuangeng Chu and Tatsuya Harada. Generalizable and animatable gaussian head avatar. arXiv preprint arXiv:2410.07971, 2024.3", + "[5] Xuangeng Chu, Yu Li, Ailing Zeng, Tianyu Yang, Lijian Lin, Yunfei Liu, and Tatsuya Harada. Gpavatar: Generalizable and precise head avatar from image (s). arXiv preprint arXiv:2401.10215, 2024.3", + "[6] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 7", + "[7] Yu Deng, Duomin Wang, Xiaohang Ren, Xingyu Chen, and Baoyuan Wang. Portrait4d: Learning one-shot 4d head avatar synthesis using synthetic data. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1, 2, 7", + "[8] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2663-2671, 2022. 3", + "[9] Philip-William Grassal, Malte Prinzler, Titus Leistner, Carsten Rother, Matthias Nießner, and Justus Thies. Neural head avatars from monocular rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18653–18664, 2022. 2", + "[10] Jianzhu Guo, Dingyun Zhang, Xiaogiang Liu, Zhizhou Zhong, Yuan Zhang, Pengfei Wan, and Di Zhang. Liveportrait: Efficient portrait animation with stitching and retargeting control. arXiv preprint arXiv:2407.03168, 2024. 2, 5, 6", + "[11] Xiao Han, Yukang Cao, Kai Han, Xiatian Zhu, Jiankang Deng, Yi-Zhe Song, Tao Xiang, and Kwan-Yee K Wong. Headsculpt: Crafting 3d head avatars with text. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[12] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilib-" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "rium. Advances in neural information processing systems, 30, 2017. 6", + "[13] Fa-Ting Hong, Longhao Zhang, Li Shen, and Dan Xu. Depth-aware generative adversarial network for talking head video generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3397-3406, 2022. 3", + "[14] Yang Hong, Bo Peng, Haiyao Xiao, Ligang Liu, and Juyong Zhang. Headnerf: A real-time nerf-based parametric head model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20374-20384, 2022. 3", + "[15] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 3, 7", + "[16] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 3", + "[17] Taras Khakhulin, Vanessa Sklyarova, Victor Lempitsky, and Egor Zakharov. Realistic one-shot mesh-based head avatars. In European Conference on Computer Vision, pages 345-362. Springer, 2022. 3", + "[18] Tobias Kirschstein, Shenhan Qian, Simon Giebenhain, Tim Walter, and Matthias Nießner. Nersemble: Multi-view radiance field reconstruction of human heads. ACM Transactions on Graphics (TOG), 42(4):1-14, 2023. 1, 6", + "[19] Shikai Li, Jianglin Fu, Kaiyuan Liu, Wentao Wang, KwanYee Lin, and Wayne Wu. Cosmicman: A text-to-image foundation model for humans. In Computer Vision and Pattern Recognition (CVPR), 2024. 6", + "[20] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6), 2017. 2, 5", + "[21] Weichuang Li, Longhao Zhang, Dong Wang, Bin Zhao, Zhi-gang Wang, Mulin Chen, Bang Zhang, Zhongjian Wang, Liefeng Bo, and Xuelong Li. One-shot high-fidelity talking-head synthesis with deformable neural radiance field. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17969-17978, 2023. 3", + "[22] Xian Liu, Xiaohang Zhan, Jiaxiang Tang, Ying Shan, Gang Zeng, Dahua Lin, Xihui Liu, and Ziwei Liu. Humangau-sian: Text-driven 3d human generation with gaussian splatt-ing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6646-6657, 2024. 3", + "[23] Shengjie Ma, Yanlin Weng, Tianjia Shao, and Kun Zhou. 3d gaussian blendshapes for head avatar animation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-10, 2024. 1", + "[24] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 4", + "[25] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf:" + ], + "bbox": [ + 516, + 93, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11620", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3", + "[26] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2, 7", + "[27] Shenhan Qian, Tobias Kirschstein, Liam Schoneveld, Davide Davoli, Simon Giebenhain, and Matthias Nießner. Gaussianavatars: Photorealistic head avatars with rigged 3d gaus-sians. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20299-20309, 2024. 1, 2, 5, 6, 8", + "[28] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6", + "[29] Amit Raj, Srinivas Kaza, Ben Poole, Michael Niemeyer, Nataniel Ruiz, Ben Mildenhall, Shiran Zada, Kfir Aberman, Michael Rubinstein, Jonathan Barron, et al. Dreambooth3d: Subject-driven text-to-3d generation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2349–2359, 2023. 2", + "[30] Yurui Ren, Ge Li, Yuanqi Chen, Thomas H Li, and Shan Liu. Pirenderer: Controllable portrait image generation via semantic neural rendering. In Proceedings of the IEEE/CVF international conference on computer vision, pages 13759-13768, 2021. 3", + "[31] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on graphics (TOG), 42(1):1-13, 2022. 2, 3", + "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 4", + "[33] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.4", + "[34] Jingxiang Sun, Xuan Wang, Lizhen Wang, Xiaoyu Li, Yong Zhang, Hongwen Zhang, and Yebin Liu. Next3d: Generative neural texture rasterization for 3d-aware head avatars. In CVPR, 2023. 2", + "[35] Felix Taubner, Ruihang Zhang, Mathieu Tuli, and David B Lindell. Cap4d: Creating animatable 4d portrait avatars with morphable multi-view diffusion models. arXiv preprint arXiv:2412.12093, 2024. 1, 2", + "[36] Kartik Teotia, Mallikarjun B R, Xingang Pan, Hyeongwoo Kim, Pablo Garrido, Mohamed Elgharib, and Christian Theobalt. HQ3davatar: High-quality implicit 3d head avatar. ACM Transactions on Graphics, 43(3):1-24, 2024. 1, 2", + "[37] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "video diffusion. arXiv preprint arXiv:2403.12008, 2024. 2, 7", + "[38] Kaisiyuan Wang, Qianyi Wu, Linsen Song, Zhuoqian Yang, Wayne Wu, Chen Qian, Ran He, Yu Qiao, and Chen Change Loy. Mead: A large-scale audio-visual dataset for emotional talking-face generation. In European Conference on Computer Vision, pages 700–717. Springer, 2020. 7", + "[39] Huawei Wei, Zejun Yang, and Zhisheng Wang. Aniportrait: Audio-driven synthesis of photorealistic portrait animation. arXiv preprint arXiv:2403.17694, 2024. 1, 2, 7", + "[40] Yiqian Wu, Hao Xu, Xiangjun Tang, Xien Chen, Siyu Tang, Zhebin Zhang, Chen Li, and Xiaogang Jin. Portrait3d: Text-guided high-quality 3d portrait generation using pyramid representation and gans prior. ACM Transactions on Graphics (TOG), 43(4):1-12, 2024. 1, 2, 3, 7", + "[41] Sicheng Xu, Jiaolong Yang, Dong Chen, Fang Wen, Yu Deng, Yunde Jia, and Xin Tong. Deep 3d portrait from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7710-7720, 2020. 3", + "[42] Zhenhui Ye, Tianyun Zhong, Yi Ren, Jiaqi Yang, Weichuang Li, Jiawei Huang, Ziyue Jiang, Jinzheng He, Rongjie Huang, Jinglin Liu, et al. Real3d-portrait: One-shot realistic 3d talking portrait synthesis. arXiv preprint arXiv:2401.08503, 2024. 3", + "[43] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. In European conference on computer vision, pages 85-101. Springer, 2022. 3", + "[44] Fei Yin, Yong Zhang, Xuan Wang, Tengfei Wang, Xiaoyu Li, Yuan Gong, Yanbo Fan, Xiaodong Cun, Ying Shan, Cengiz Oztireli, et al. 3d gan inversion with facial symmetry prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 342-351, 2023. 2", + "[45] Wangbo Yu, Yanbo Fan, Yong Zhang, Xuan Wang, Fei Yin, Yunpeng Bai, Yan-Pei Cao, Ying Shan, Yang Wu, Zhongqian Sun, et al. Nofa: Nerf-based one-shot facial avatar reconstruction. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-12, 2023. 3", + "[46] Bowen Zhang, Yiji Cheng, Jiaolong Yang, Chunyu Wang, Feng Zhao, Yansong Tang, Dong Chen, and Baining Guo. Gaussiancube: Structuring gaussian splatting using optimal transport for 3d generative modeling. arXiv preprint arXiv:2403.19655, 2024. 7", + "[47] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 3", + "[48] Yufeng Zheng, Wang Yifan, Gordon Wetzstein, Michael J Black, and Otmar Hilliges. Pointavatar: Deformable point-based head avatars from videos. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21057-21067, 2023. 2", + "[49] Zhenglin Zhou, Fan Ma, Hehe Fan, Zongxin Yang, and Yi Yang. Headstudio: Text to animatable head avatars with 3d gaussian splatting. In European Conference on Computer Vision, pages 145-163. Springer, 2024. 2" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11621", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/FaceCraft4D_ Animated 3D Facial Avatar Generation from a Single Image/c512f2aa-6fbd-4940-89be-2245ba51ffe2_model.json b/2025/FaceCraft4D_ Animated 3D Facial Avatar Generation from a Single Image/c512f2aa-6fbd-4940-89be-2245ba51ffe2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e6928b58f6df8f9ffc3e68fed09008c5dc32522c --- /dev/null +++ b/2025/FaceCraft4D_ Animated 3D Facial Avatar Generation from a Single Image/c512f2aa-6fbd-4940-89be-2245ba51ffe2_model.json @@ -0,0 +1,2420 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.125, + 0.131, + 0.877, + 0.155 + ], + "angle": 0, + "content": "FaceCraft4D: Animated 3D Facial Avatar Generation from a Single Image" + }, + { + "type": "text", + "bbox": [ + 0.171, + 0.18, + 0.828, + 0.218 + ], + "angle": 0, + "content": "Fei Yin\\(^{1,2}\\), Mallikarjun B R\\(^{2}\\), Chun-Han Yao\\(^{2}\\), Rafal K. Mantiuk\\(^{1\\dagger}\\), Varun Jampani\\(^{2\\dagger}\\), \\(^{1}\\)University of Cambridge, \\(^{2}\\)Stability AI" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.252, + 0.329, + 0.269 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.284, + 0.485, + 0.571 + ], + "angle": 0, + "content": "We present a novel framework for generating high-quality, animatable 4D avatar from a single image. While recent advances have shown promising results in 4D avatar creation, existing methods either require extensive multiview data or struggle with shape accuracy and identity consistency. To address these limitations, we propose a comprehensive system that leverages shape, image, and video priors to create full-view, animatable avatars. Our approach first obtains initial coarse shape through 3D-GAN inversion. Then, it enhances multiview textures using depth-guided warping signals for cross-view consistency with the help of the image diffusion model. To handle expression animation, we incorporate a video prior with synchronized driving signals across viewpoints. We further introduce a Consistent-Inconsistent training to effectively handle data inconsistencies during 4D reconstruction. Experimental results demonstrate that our method achieves superior quality compared to the prior art, while maintaining consistency across different viewpoints and expressions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.601, + 0.222, + 0.617 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.626, + 0.484, + 0.791 + ], + "angle": 0, + "content": "4D Avatar generation aims to create animatable 3D head avatars, enabling the generation of consistent identities with custom viewpoints and expressions. High-quality 4D heads have significant applications in gaming, education, and film industries. However, existing 4D avatar building methods [18, 23, 27, 36] often require extensive inputs, such as multiview video with camera pose estimation, which is impractical for widespread use. Furthermore, these methods struggle to produce realistic results for extreme viewpoints and expression changes due to data scarcity, for instance the back view of the head." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.793, + 0.484, + 0.854 + ], + "angle": 0, + "content": "The goal of this work is to create a 4D facial avatar from just a single image. This task is challenging because a single image provides limited information, missing essential details like depth and multiple viewpoints that are crucial" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.251, + 0.925, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.365, + 0.922, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.486, + 0.908, + 0.556 + ], + "angle": 0, + "content": "Figure 1. Given a single image input, our method is capable of generating 4D avatars, producing photorealistic textures and consistent expressions across multiple views. Our approach also demonstrates robust performance on challenging inputs, including cartoon characters or old photographs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.909, + 0.674 + ], + "angle": 0, + "content": "for constructing a complete 3D model. A single image cannot provide information about the unseen parts of a person's head or capture dynamic aspects like facial expressions. It is also impractical to train an end-to-end model due to the lack of multiview full-head animation datasets. To address these challenges, the system needs to rely on prior knowledge, like shape and expression." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Existing single-image avatar generation methods come with many shortcomings, which we summarize in Tab. 1. The works that rely on a single image of novel identity at test time, typically utilize large-scale, unstructured 2D images [1] and videos [7, 39] during training to account for multiview and temporal aspects. Due to the practical challenges in collecting multiview full-head animation datasets, most single-image methods, including recent work [35], cannot reconstruct a complete 360-degree view of the head. While some approaches can generate full views [1, 40], they typically lack the ability to model facial expressions. Among methods that can animate, some methods rely solely on 2D representation [39], which makes these methods multi-view inconsistent by design. To overcome the above limitation, some methods rely on a hybrid representation" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.863, + 0.203, + 0.876 + ], + "angle": 0, + "content": "\\(\\dagger\\) Equal advising." + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.876, + 0.483, + 0.9 + ], + "angle": 0, + "content": "*No biometric data was used to train, validate, or evaluate the model described in this work." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.863, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "11612" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 1. Comparison of methods for avatar generation and animation. The first two methods focus on general object generation, while the remaining methods are specifically designed for avatar applications." + }, + { + "type": "table", + "bbox": [ + 0.118, + 0.128, + 0.885, + 0.244 + ], + "angle": 0, + "content": "
MethodsGoalDimensionInput360° viewAnimation3D ConsistencyQuality
LDM [32]Image Generation2DSingle Image-High
DreamFusion [26], SV3D [37]Image-to-3D3DSingle Image✓LowLow
LivePortrait [10], AniPortrait [39]Video Generation2DSingle Image✓LowHigh
PanoHead [1], SPI [44]3D GAN Inversion3DSingle Image✓MediumMedium
Portrait4D [7], CAP4D [35]3D Talking Head3DSingle Image✓MediumHigh
FaceCraft4D4D AvATAR Generation3DSingle Image✓✓HighHigh
HQ3D [36], GA [27]4D AvATAR Generation3DMulti-view Videos✓✓HighHigh
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.262, + 0.482, + 0.337 + ], + "angle": 0, + "content": "that combines 2D and 3D [7, 34]. However, such methods cannot handle extreme camera angles, as they rely on 2D components for high-resolution synthesis. In contrast, we aim to build our model that has pure 3D representation, and, therefore, is multiview-consistent by design." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.338, + 0.483, + 0.488 + ], + "angle": 0, + "content": "We propose FaceCraft4D, which takes a single image as input and creates a 3D Facial Avatar that can be animated and rendered in 360-degree view, as shown in Fig. 1. To tackle such an under-constrained problem, we make use of 3 different types of priors: a geometry prior, an image prior, and a video prior. We use these priors to synthesize personalized multiview images of the given identity, which are then used to train an explicit 3D model of an avatar that can be animated (see Fig. 2). We will refer to such a model as a 4D avatar." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.49, + 0.483, + 0.746 + ], + "angle": 0, + "content": "We start with a geometry prior from pretrained 3D-GAN [1]. Since, 3D-GAN is trained with large-scale images that comprises of all the views of the head, it has information about the complete head shape. Through 3D-GAN inversion [31], we obtain personalized shape along with an approximate texture. The resulting 3D representation lacks personalized high-quality multiview consistent texture and can not be used for creating novel expressions or animations (see Tab. 1). To improve the quality of the textures, we use of a 2D image prior. To ensure multi-view consistency and preservation of identity across the views, we employ epipolar constraints and cross-view mutual attention. The resulting multi-view images can be used to build a high-quality static 3D head model, which, however, lacks control over expressions and articulations. To animate the avatar, we utilize a 2D video prior [10], which generates training data for facial animation across different viewpoints." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.748, + 0.483, + 0.852 + ], + "angle": 0, + "content": "Finally, for real-time rendering performance, we train a 3D Gaussian representation [16] that is rigged to a FLAME parametric face model [20, 27] using the multi-view videos obtained in the previous steps. To further improve the training consistency across the views and to avoid blur, we propose a COnsistent-INconsistent (COIN) training, which can capture cross-view inconsistencies in an MLP." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.855, + 0.482, + 0.869 + ], + "angle": 0, + "content": "In summary, the contributions of our work are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- We present a novel framework for generating high-quality 4D head models from a single image of a face. We com" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.262, + 0.905, + 0.29 + ], + "angle": 0, + "content": "pare our framework with previous methods and summarize the differences in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.292, + 0.905, + 0.337 + ], + "angle": 0, + "content": "- We introduce novel warping-based control generation signals to improve quality and consistency across different views and expressions." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.338, + 0.905, + 0.382 + ], + "angle": 0, + "content": "- We introduce COIN training, which enables a 3D Gaussian model to learn from inconsistent data while preserving high-quality, consistent features." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.292, + 0.905, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.399, + 0.655, + 0.415 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.426, + 0.907, + 0.56 + ], + "angle": 0, + "content": "3D Head generation. Numerous studies have explored the reconstruction of 3D head models. To obtain sufficient information, most approaches [9, 27, 48, 48] utilize monocular video sequences, which provide multiple viewpoints of the subject's head. However, a significant limitation of these methods is their reliance on multiple viewpoints and lengthy video inputs. They also struggle to accurately reproduce less prominent features in videos, such as the back of the head." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.564, + 0.906, + 0.713 + ], + "angle": 0, + "content": "Another approach is to train a 3D generator on a large head image dataset. EG3D [3] proposed a hybrid tri-plane representation that can quickly and efficiently render high-quality facial images. Panohead [1] and Portrait3D [40] expanded on EG3D by increasing the number of planes, enabling the generation of 3D heads from limited angles to full 360-degree views. However, most methods in this category rely on hybrid representation having a 2D super-resolution network. This impacts their generalization to extreme camera poses and suffer identity inconsistency across views." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.715, + 0.906, + 0.881 + ], + "angle": 0, + "content": "Recently, the rise of text-to-image models has spurred the development of novel approaches for text-to-3D generation tasks. One such approach is Score Distillation Sampling (SDS) [26, 29], which leverages the score function to distill diffusion priors into 3D representations, enabling the generation of 3D content from text or image inputs. Headsculpt [11, 49] incorporates facial keypoint information into the text-to-3D generation method, effectively constraining the facial layout. However, the texture generated by these methods can be oversaturated, making it challenging to achieve photo-realistic head models." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Head animation. Head animation techniques can be" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11613" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.898, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.28, + 0.908, + 0.351 + ], + "angle": 0, + "content": "Figure 2. Overview of FaceCraft4D: Our approach begins with estimating the geometry of a single input image using a Geometry Prior. This geometry guides the synthesis of personalized multiview images, providing \\(360^{\\circ}\\) views and varied expressions, with support from both 2D image and video priors. Since the synthesized data often exhibit inconsistency across views, we propose COIN optimization for robust 4D optimization. By fitting the model to the multiview data, we achieve our final 4D avatar. All faces in this manuscript come from the public FFHQ dataset [15]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.364, + 0.482, + 0.499 + ], + "angle": 0, + "content": "broadly categorized into 2D-based and 3D-aware approaches, distinguished by their utilization of explicit camera models for rendering. Recent methods [8, 13, 30, 43] have predominantly adopted warping-based strategies, applying deformation fields to intermediate appearance features to capture facial motion characteristics. While effective for frontal and moderate poses, these approaches struggle to maintain geometric consistency during large pose variations due to their inherent lack of 3D modeling." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.5, + 0.483, + 0.727 + ], + "angle": 0, + "content": "To enable free-viewpoint rendering with stronger 3D consistency guarantees, researchers integrated explicit 3D representations into the head synthesis pipeline. Early approaches [17, 41] relied on 3D morphable models for representing head geometry and texture. Subsequent works [5, 14, 21, 42, 45] leveraged more sophisticated representations such as Neural Radiance Fields (NeRFs) [25] to better capture complex structures like hair and accessories. Recent advances like GAGAvatar [4, 22], which employs 3D Gaussian splatting [16], have demonstrated impressive performance in both rendering quality and speed. However, these methods are typically limited in their ability to generate full 360-degree views. Our approach enables true 360-degree view synthesis while maintaining real-time rendering capabilities and achieving superior animation quality." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.74, + 0.182, + 0.755 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.87 + ], + "angle": 0, + "content": "In this work, we propose a novel framework for generating high-fidelity 4D avatars from a single portrait image \\( I_{R} \\). Reconstructing a 3D dynamic model from a single 2D image is an inherently ill-posed problem, as critical 3D and temporal information is missing. To address this challenge, we leverage several 2D and 3D priors to compensate for the missing data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "The overall pipeline, shown in Fig. 2, consists of two stages: personalized multiview generation and 4D repre" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.907, + 0.531 + ], + "angle": 0, + "content": "sensation optimization. In the Personalized Multiview Generation stage, we utilize geometry priors to estimate the 3D geometry (Sec. 3.1), use image priors to enhance the texture (Sec. 3.2), and video priors to synthesize high-quality, consistent multi-view videos with dynamic expressions (Sec. 3.3). We then leverage this synthetically-generated personalized image set to optimize a robust 4D representation model (Sec. 3.4). The trained 4D model can be animated by the control parameters of a morphable model, enabling the generation of high-fidelity 4D avatars from a single portrait image." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.579, + 0.673, + 0.595 + ], + "angle": 0, + "content": "3.1. Geometry Prior" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.901 + ], + "angle": 0, + "content": "The initialization is crucial for 3D generation. A good initialization should follow the approximate geometry of the head shown in the input image, \\( I_{R} \\). For that, we use the prior provided by a 3D-aware generator, PanoHead [1, 40], which was trained on a large \\( 360^{\\circ} \\) head dataset. We obtain a coarse 3D reconstruction for a given input image \\( I_{R} \\) through GAN inversion. Specifically, we first optimize the latent code \\( z \\) within the generator's domain by minimizing a combination of pixel-wise \\( \\mathcal{L}_2 \\) loss and image-level LPIPS loss [47]. Subsequently, following [31], we keep the optimized latent code \\( z \\) fixed while fine-tuning the 3D generator's parameters. During this fine-tuning stage, we maintain the same loss functions used in the latent optimization phase to ensure the rendered views closely match the input image. This process gives us an initial 3D head representation, though one that cannot be animated. The optimized generator is then used to synthesize a set of approximate multiview images \\( \\{I_i\\} \\) and their corresponding depth maps \\( \\{D_i\\} \\), where \\( i \\) is the view index." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11614" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.09, + 0.496, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.372, + 0.485, + 0.43 + ], + "angle": 0, + "content": "Figure 3. Image Prior: We introduce a cross-view mutual attention mechanism and epipolar constraints to enhance consistency in generated novel views. Our approach aligns reference and target images, maintaining visual coherence across viewpoints." + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.433, + 0.498, + 0.621 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.638, + 0.486, + 0.695 + ], + "angle": 0, + "content": "Figure 4. Triplane-based methods (e.g. PanoHead) are sensitive to focal length (scale) and exhibit significant degradation of quality as focal length decreases. In contrast, our Gaussian-based methods maintain robust performance across varying focal lengths." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.71, + 0.221, + 0.727 + ], + "angle": 0, + "content": "3.2. Image Prior" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.486, + 0.903 + ], + "angle": 0, + "content": "The textures synthesized from the geometry prior may lack visual quality, and the identity of the individual may not be well preserved across the views. In particular, when rendering with different focal lengths, triplane-based methods [1] exhibit substantial degradation in image quality, as illustrated in Fig. 4. We enhance the generated multiview images using 2D diffusion models [32], which have demonstrated powerful generative capabilities. Although direct image-to-image translation [24] appears to be a straightforward solution, this approach poses significant risks: the de-noising process may introduce cross-view inconsistencies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.152 + ], + "angle": 0, + "content": "and also modify the original semantic content. To address these risks, we propose two strategies that constrain the generation process, which we describe next and illustrate in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.156, + 0.909, + 0.443 + ], + "angle": 0, + "content": "Cross-view mutual attention. Inspired by MasaCtrl [2], we introduce a cross-view attention mechanism to replace the standard self-attention during the generation process. This modification allows us to inject information from the reference image into the novel view. Specifically, we first perform DDIM inversion [33] to introduce noise to both the reference frame \\( I_{R} \\) and the coarse image \\( I_{i} \\) of the target novel view. This step preserves the overall face structure and contour. Then, the diffusion model is used to gradually denoise two images to enhance the texture. In this process, we replace the key \\( K \\) and value \\( V \\) matrices (calculated in self-attention) of the novel view with those derived from the reference image. Different from [2], we adopt a batch-processing strategy in which multiple images \\( I_{i} \\) representing different viewpoints are treated as a single batch, with the reference image \\( I_{R} \\) serving as the consistent source for all subsequent calculations. This not only streamlines the generation process but also ensures uniformity in feature injection across novel viewpoints." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.444, + 0.909, + 0.549 + ], + "angle": 0, + "content": "The rationale behind this approach lies in the observation that, despite variations in viewpoint, the underlying textures (represented by the attention values) should maintain consistency across the views. The cross-view attention mechanism enables the transfer of information from the reference to the novel views, enhancing both the fidelity and coherence of the generated novel views." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.909, + 0.87 + ], + "angle": 0, + "content": "Warping-based control generation. The consistency across the views is further enforced by epipolar constraints, which enforce geometric relationships between novel views and the input view. We first use cross-view mutual attention to generate the view of the back of the head. The generated back view along with the reference image are set as anchors. Then, we propagate the texture information to novel views through depth-based warping. Specifically, we lift the anchor image into 3D space and then project it to the adjacent view with depth guidance \\(\\{D_i\\}\\). During the projection, we form a mask to indicate the visibility of pixels to filter invisible areas using the rendered depth values. The projected image is then used to guide the texture enhancement process. The intermediate latent of the enhanced image in the diffusion model is blended with the latent of the projected image via a simple mask copy-pasting operation. By combining the latent, the resulting images benefit from both the improved consistency provided by the epipolar constraints and the enhanced quality achieved through the denoising process. We repeat this process for various azimuth angles, creating a comprehensive set of viewpoints." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.908, + 0.901 + ], + "angle": 0, + "content": "With the aforementioned guidance and epipolar constraints, we can ensure coherence across the entire view" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11615" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.089, + 0.496, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.265, + 0.485, + 0.337 + ], + "angle": 0, + "content": "Figure 5. COIN-Training: To address geometric and color inconsistencies between the views, we train two 3D representations: a view-consistent base model (GaussianAvatar) and an MLP, encoding inconsistencies between the views. The MLP with inconsistencies lets us robustly reconstruct the high-quality 3D base model." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.364, + 0.483, + 0.396 + ], + "angle": 0, + "content": "space and generate high-fidelity view-consistent images \\(\\{I_i^*\\}\\) for further reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.405, + 0.218, + 0.421 + ], + "angle": 0, + "content": "3.3. Video Prior" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.429, + 0.484, + 0.807 + ], + "angle": 0, + "content": "The view-consistent data created in the previous step has a static expression similar to that of reference image \\( I_R \\). This may not be enough to create diverse expressions as the images do not contain some of the expression-dependent information. For example, if the reference image has the mouth closed, the synthesized dataset would not have details inside the mouth region. To incorporate expression-dependent textures and geometry that may not be available in the multiview image generation stage (Sec. 3.2), we need to create multi-view facial expression data. We use LivePortrait [10] as a video prior, which can help generate multiview expressive videos. LivePortrait uses a single source image and a driving video to generate new videos where the source identity mimics the expressions and poses from the target video. Instead of directly copying the driving video's motion, one has the option of capturing the changes in relative expression but maintain the pose of the input image. Given that we have the reference and multiview-consistent enhanced images \\( \\{I_i^*\\} \\) from the previous step, we provide these to LivePortrait as the source images together with driving videos with diverse poses and expressions (details in Sec. 4), and obtain target videos. Because the identity is consistent across the input views, the resulting videos also maintain the identity. Using identical driving videos ensures synchronized expressions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.468, + 0.835 + ], + "angle": 0, + "content": "3.4. Robust 4D optimization with COIN training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The synthesized multi-view videos have high-quality textures and good cross-view consistency, but small misalignment of features and colors is unavoidable. Learning 4D representation directly on such data leads to blurry textures" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.257 + ], + "angle": 0, + "content": "for the rendered objects. Here, we address this problem by proposing a robust training paradigm. At a high level, we learn a base representation that is common to all views, and a delta representation that is specific to each view during training. This helps in cornering inconsistencies concerning each view to separate representation. The base model will have a coarse but consistent representation learned, while the high-frequency details end up in the delta-space representation. At test time, we choose one of the views as our reference view and use its delta representation as the common high-frequency details for all the views." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.258, + 0.907, + 0.514 + ], + "angle": 0, + "content": "In specific, we achieve this by jointly training two 4D representations: a consistent GaussianAvatar [27], which represents the 3D structure using Gaussians bound to a FLAME-tracked [20] mesh, and a NeRF-style MLP, tasked with encoding inconsistencies between the views. As shown in Fig. 5, we design the losses that prioritize structure reconstruction in the consistent GaussianAvatar (LPIPS), and high-frequency detail reconstruction in the MLP with the view-dependent inconsistencies (L1 and SSIM). By isolating view-dependent variations in a separate model, this strategy prevents these inconsistencies from corrupting the base representation. The approach can be considered a form of robust regression, with the key difference that the view-dependent MLP ensures that inconsistencies (outliers) are spatially localized. This cannot be achieved with robust norms (e.g., L1), which do not maintain the information about the spatial position of outliers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.515, + 0.906, + 0.693 + ], + "angle": 0, + "content": "We choose GaussianAvatar as our consistent representation because it lets us animate the avatar with controllable FLAME expression and pose parameters, providing a solid foundation for modeling consistent geometric and appearance features. The details of FLAME tracking and controlling are explained in Sec. 1.3 of the supplementary. To model the view-dependent variations, we design a lightweight MLP-based architecture that introduces minimal computational overhead while effectively capturing view-specific details. The network encodes the offset in the color of each Gaussian (see Fig. 5), which can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.697, + 0.905, + 0.713 + ], + "angle": 0, + "content": "\\[\nc _ {\\text {o f f s e t}} = \\operatorname {M L P} \\left(e _ {\\text {v i e w}}, c, e _ {g}; \\theta\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.719, + 0.906, + 0.884 + ], + "angle": 0, + "content": "where \\( e_{\\mathrm{view}} \\in \\mathbb{R}^d \\) is a learnable view-specific embedding that encodes view-representation, \\( c \\) represents the initial Gaussian color attributes from the GaussianAvatar, \\( e_g \\) is a position embedding for each Gaussian, accounting for location-specific color variation ranges (e.g., eyes vs. mouth regions), and \\( \\theta \\) is a vector of the trainable parameters. The final view-dependent color is computed as the sum of the base GaussianAvatar color and the offset: \\( c^* = c + c_{\\mathrm{offset}} \\). The MLP consists of only two layers to maintain computational efficiency and preserve the real-time rendering capabilities of the Gaussian-based representation." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.885, + 0.905, + 0.902 + ], + "angle": 0, + "content": "During training, we render two types of images: \\( I_{i}^{\\mathrm{C}} \\) from" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11616" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.086, + 0.891, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.25, + 0.457, + 0.746, + 0.473 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison on static 3D head generation from a single image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.484, + 0.553 + ], + "angle": 0, + "content": "the consistent component alone, and \\( I_{i}^{\\mathrm{IC}} \\) from the combination of both components. To ensure accurate reconstruction of high-frequency details, we supervise \\( I_{i}^{\\mathrm{IC}} \\) using pixel-wise losses:" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.564, + 0.483, + 0.584 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p i x e l}} = \\lambda_ {1} \\mathcal {L} 1 \\left(I _ {i} ^ {\\mathrm {I C}}, I _ {i} ^ {*}\\right) + \\lambda_ {\\text {S S I M}} \\text {S S I M} \\left(I _ {i} ^ {\\mathrm {I C}}, I _ {i} ^ {*}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.594, + 0.484, + 0.655 + ], + "angle": 0, + "content": "To prevent the inconsistent component from dominating the representation and maintain the effectiveness of the consistent component, we introduce a structural supervision loss and a regularization term:" + }, + { + "type": "equation", + "bbox": [ + 0.189, + 0.67, + 0.483, + 0.689 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s t r u c}} = \\lambda_ {\\text {L P I P S}} \\operatorname {L P I P S} \\left(I _ {i} ^ {\\mathrm {C}}, I _ {i} ^ {*}\\right) \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.711, + 0.483, + 0.728 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e g}} = \\lambda_ {\\text {o f f s e t}} \\mathcal {L} 1 \\left(c _ {\\text {o f f s e t}}, 0\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.734, + 0.483, + 0.81 + ], + "angle": 0, + "content": "where LPIPS loss ensures image-level supervision for the consistent component, and the offset regularization encourages minimal view-dependent modifications. The hyperparameters are set to \\(\\lambda_{1} = 0.8\\), \\(\\lambda_{\\mathrm{SSIM}} = 0.2\\), \\(\\lambda_{\\mathrm{LPIPS}} = 0.05\\) and \\(\\lambda_{\\mathrm{offset}} = 1\\)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "During inference, we input a single fixed view-embedding for the inconsistent component. Given that our reference image \\( I_R \\) contains superior quality and identity information compared to the generated multiview images \\( \\{I_i^*\\} \\), we select its corresponding view embedding for inference. Novel expressions and poses can be easily gen" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.492, + 0.907, + 0.523 + ], + "angle": 0, + "content": "erated by adjusting the FLAME parameters, leveraging the animation capabilities of our backbone model." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.539, + 0.646, + 0.556 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.907, + 0.791 + ], + "angle": 0, + "content": "Implementation details. During the multiview generation stage, we use PanoHead [1] as our geometry prior, Cosmicman [19] as the image prior, and LivePortrait [10] as the video prior model. We render 24 views along the horizon orbit for static reconstruction. For animation, we select \\( I_R \\) and 11 frontal views from these generated images as source images and randomly sample 8 video clips from the NerSemble [18] dataset as driving videos. The source images and driving videos are then input to the video prior model to produce multiview video data. In the 4D optimization stage, we first optimize the GaussianAvatar [27] model with the static data over 30K iterations. Next, we apply COIN-optimization techniques, finetuning the static Gaussian model over 90K iterations. The complete process takes approximately two hours to generate a single 4D avatar." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Evaluation metrics. In the absence of ground truth data, we use commonly adopted non-reference metrics to evaluate the rendered images in terms of image quality and identity similarity. For overall image quality, we use the FID [12] score. To assess multi-view consistency, we employ the CLIP-I [28] score. For identity preservation, we calculate ID scores by measuring the cosine distance of face" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11617" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.086, + 0.306, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.309, + 0.087, + 0.402, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.087, + 0.593, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.596, + 0.087, + 0.803, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.805, + 0.087, + 0.903, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.303, + 0.28, + 0.694, + 0.295 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison on animation of different views." + }, + { + "type": "table_caption", + "bbox": [ + 0.098, + 0.313, + 0.475, + 0.327 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison on static 3D head generation." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.327, + 0.478, + 0.51 + ], + "angle": 0, + "content": "
MethodCLIP-I ↑ID ↑FID ↓
GaussianCube [46]0.68300.4300258.81
PanoHead [1]0.82330.4246195.28
SV3D [37]0.76560.4331234.86
Portrait3D [40]0.70660.3719302.74
Ours0.80530.5082174.36
Table 3. Quantitative comparison on 3D head animation.
MethodCLIP-I ↑ID ↑FID ↓
AntiPortrait [39]0.46530.4171364.99
Portrait-4D-V2 [7]0.52360.4592248.36
Ours0.57370.4602201.76
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.521, + 0.483, + 0.552 + ], + "angle": 0, + "content": "recognition features [6] between the novel views and the reference images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.555, + 0.484, + 0.616 + ], + "angle": 0, + "content": "Datasets. We conducted our analysis on a subset of facial images from FFHQ [15] dataset, a high-quality, in-the-wild facial dataset. For evaluation, we select 100 images with unoccluded faces to evaluate our model and baselines." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.628, + 0.396, + 0.645 + ], + "angle": 0, + "content": "4.1. Static 3D Generation Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.655, + 0.483, + 0.732 + ], + "angle": 0, + "content": "Baselines. We compare our method with existing image-to-3D approaches, including a general object-focused model, SV3D [37], and three head-specific methods: the inference-based GaussianCube [46], the 3D-GAN-based PanoHead [1], and the optimization-based Portrait3D [40]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Qualitative results. Fig. 6 shows visual comparisons where our method achieves the highest visual fidelity and identity coherence compared to other approaches. Notably, it preserves details such as teeth and earrings in side views. In contrast, SV3D, which focuses on general objects, struggles to capture accurate head geometry, often degenerating into a flat plane. GaussianCube tends to \"toonify\" the input head due to the nature of its training data. While PanoHead achieves good visual quality, it suffers from detail loss and floating artifacts during pose rotation. Portrait3D also falls short in quality, as the SDS [26] optimization introduces" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.315, + 0.776, + 0.331 + ], + "angle": 0, + "content": "variations that lead to blurring artifacts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.335, + 0.907, + 0.502 + ], + "angle": 0, + "content": "Quantitative results. For each generated head model, we render images from five viewpoints: left-frontal, left, right-frontal, right, and frontal. All five rendered images, along with the reference image, are used to compute the CLIP-I, ID, and FID metrics. As shown in Tab. 2, our method outperforms others in both ID and FID scores. This aligns with the observed visual fidelity, as our results appear more realistic and coherent. Our method achieves higher identity similarity and better detail preservation, even in side views. The results for the MEAD multiview dataset [38] can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.516, + 0.74, + 0.532 + ], + "angle": 0, + "content": "4.2. Animation Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.543, + 0.907, + 0.665 + ], + "angle": 0, + "content": "Baselines. We compare our method with one-shot video-based head reenactment approaches, including the 2D-based AniPortrait [39] and the 3D-based Portrait-4D-V2 [7]. For AniPortrait, which uses keypoints as intermediate driving signals, we modify the Euler angles of the expression image to rotate the keypoints and generate novel view videos. In the case of Portrait-4D-V2, novel views can be generated directly by editing the camera parameters." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.669, + 0.909, + 0.806 + ], + "angle": 0, + "content": "Qualitative results. As shown in Fig. 7, our method preserves expressions with greater accuracy. Portrait-4D-V2 lacks a back-of-head model, so when rotated to side views, the back appears empty. AniPortrait, as a 2D-based method, can achieve high texture quality. However, with large-angle rotations, AniPortrait struggles to preserve identity details and introduces noticeable artifacts in the hair and background. In contrast, our method maintains accurate geometry and expression, even with camera movement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Quantitative results. We test our data on unseen views (especially \\(60^{\\circ} - 180^{\\circ}\\)), to highlight each method's performance under varying viewing angles, as handling extreme rotations is often challenging. The results are presented in Tab. 3. Our method outperforms all baselines, demonstrating robustness across varied views. The performance" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11618" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.09, + 0.172, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.167, + 0.15, + 0.176 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.189, + 0.091, + 0.258, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.187, + 0.166, + 0.263, + 0.177 + ], + "angle": 0, + "content": "Warped Image" + }, + { + "type": "image", + "bbox": [ + 0.265, + 0.091, + 0.335, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.267, + 0.166, + 0.336, + 0.177 + ], + "angle": 0, + "content": "w/o Warping" + }, + { + "type": "image", + "bbox": [ + 0.347, + 0.092, + 0.417, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.355, + 0.166, + 0.4, + 0.176 + ], + "angle": 0, + "content": "w/o MA" + }, + { + "type": "image", + "bbox": [ + 0.427, + 0.091, + 0.498, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.444, + 0.166, + 0.466, + 0.176 + ], + "angle": 0, + "content": "Full" + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.18, + 0.463, + 0.193 + ], + "angle": 0, + "content": "Figure 8. Ablation of multiview image generation modules." + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.195, + 0.185, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.307, + 0.157, + 0.318 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.206, + 0.195, + 0.298, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.197, + 0.273, + 0.296, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.207, + 0.308, + 0.288, + 0.317 + ], + "angle": 0, + "content": "w/o Video Prior" + }, + { + "type": "image", + "bbox": [ + 0.308, + 0.195, + 0.391, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.273, + 0.398, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.322, + 0.308, + 0.376, + 0.317 + ], + "angle": 0, + "content": "w/o COIN" + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.195, + 0.498, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.401, + 0.273, + 0.498, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.435, + 0.308, + 0.456, + 0.317 + ], + "angle": 0, + "content": "Full" + }, + { + "type": "image_caption", + "bbox": [ + 0.163, + 0.326, + 0.411, + 0.34 + ], + "angle": 0, + "content": "Figure 9. Ablation of animation modules." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.344, + 0.164, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.407, + 0.146, + 0.417 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.166, + 0.345, + 0.231, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.163, + 0.406, + 0.221, + 0.417 + ], + "angle": 0, + "content": "Novel View" + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.346, + 0.311, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.26, + 0.407, + 0.287, + 0.417 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.311, + 0.346, + 0.376, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.311, + 0.407, + 0.368, + 0.417 + ], + "angle": 0, + "content": "Novel View" + }, + { + "type": "image", + "bbox": [ + 0.378, + 0.346, + 0.436, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.391, + 0.407, + 0.418, + 0.417 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.441, + 0.347, + 0.494, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.442, + 0.406, + 0.499, + 0.416 + ], + "angle": 0, + "content": "Novel View" + }, + { + "type": "image_caption", + "bbox": [ + 0.156, + 0.419, + 0.417, + 0.433 + ], + "angle": 0, + "content": "Figure 10. More results with diverse inputs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.448, + 0.483, + 0.495 + ], + "angle": 0, + "content": "of baselines degrades significantly as the viewing angle increases. Additional results and baseline comparisons are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.505, + 0.244, + 0.52 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.53, + 0.484, + 0.652 + ], + "angle": 0, + "content": "Multiview image generation. We conduct an ablation study on two multi-view generation modules, as shown in Tab. 4 and Fig. 8. Incorporating the warping module allows us to preserve finer details, like tattoos, in novel views. While removing MA may lead to semantic inconsistencies, such as changes in gender. Adding mutual attention (MA) further improves fidelity and enhances identity similarity by injecting features from the reference images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.656, + 0.484, + 0.852 + ], + "angle": 0, + "content": "Animation. Our 3D representation is based on GaussianAvatar [27], which is FLAME-driven, enabling direct animation through FLAME parameters. To evaluate the base performance, we remove our video prior module. As shown in Fig. 9, without the video prior, the final avatar fails to synthesize accurate expressions. For consistency modeling, we compare our model with and without the COIN training. Without COIN, training directly on inconsistent generated images results in blurred textures. With COIN training, regions such as the teeth remain sharp, and finer details, like individual hair strands, are preserved. The quantitative results in Tab. 5 further validate the effectiveness of the COIN optimization across novel views." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Processing times. We analyzed the processing times of each stage. Our method consists of three stages: a) 3D GAN inversion (2 mins); b) Multiview image & video generation" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.905, + 0.13 + ], + "angle": 0, + "content": "Table 4. Ablation of multiview image generation modules. \"Warp\" refers to the warping-based control generation (Sec. 3.2) and MA to cross-view mutual attention (Sec. 3.2)." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.131, + 0.905, + 0.216 + ], + "angle": 0, + "content": "
MethodCLIP-I ↑ID ↑FID
w/o Warp w/o MA0.73280.4787182.27
w/o Warp0.78860.4915171.08
w/o MA0.81510.4951172.43
Full0.81620.4984166.96
" + }, + { + "type": "table_caption", + "bbox": [ + 0.589, + 0.219, + 0.83, + 0.23 + ], + "angle": 0, + "content": "Table 5. Ablation of animation modules." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.232, + 0.905, + 0.289 + ], + "angle": 0, + "content": "
MethodCLIP-I ↑ID ↑FID ↓
w/o COIN0.76880.4952144.95
Full0.77290.5010142.80
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.292, + 0.905, + 0.318 + ], + "angle": 0, + "content": "Table 6. Time comparison with optimization-based avatar reconstruction methods." + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.32, + 0.901, + 0.377 + ], + "angle": 0, + "content": "
MethodNerFaceIMAvatarPointAvatarHeadStudio
Time (hour)54h48h6h2h
MethodMonoGAGaussianAvatarHq3davatarOurs
Time (hour)7h2h12h2.5h
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.39, + 0.906, + 0.495 + ], + "angle": 0, + "content": "(10 mins), FLAME parameters extraction (10 mins); and c) COIN-training to overfit multiview videos (2 hours). The generation of a single avatar \\(\\sim 2.5\\) hours, which is comparable to or faster than other optimization-based avatar reconstruction methods, as shown in Tab. 6. At inference, we achieve real time rendering at 156 FPS with a resolution of \\(512 \\times 512\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.5, + 0.906, + 0.591 + ], + "angle": 0, + "content": "Diverse inputs. To validate robustness, we conduct experiments on diverse inputs, including cartoon, line drawing, and portraits with extreme camera poses. The results in Fig. 10 demonstrate consistent performance across various domains. Additional video results are available on the webpage in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.604, + 0.64, + 0.62 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we present a novel method for generating high-quality 4D portraits from a single image. By leveraging multiple priors, including 3D-GAN for geometry initialization, image prior for texture enhancement, and video prior for animation, our method can generate animated avatars with detailed textures, excellent consistency, and plausible rendition of views missing in the input. The proposed depth-guided warping ensures cross-view consistency during the generation process, while our COIN-training strategy enables high-quality 4D reconstruction from potentially inconsistent multi-view data. Extensive experiments demonstrate that our method outperforms existing approaches in terms of geometry accuracy, texture quality, and identity preservation across different viewpoints and expressions. Our framework makes high-quality 4D portrait generation more accessible by removing the requirement for multi-view data capture, paving the way for broader applications." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11619" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Sizhe An, Hongyi Xu, Yichun Shi, Guoxian Song, Umit Y Ogras, and Linjie Luo. Panohead: Geometry-aware 3d full-head synthesis in 360deg. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20950-20959, 2023. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.484, + 0.257 + ], + "angle": 0, + "content": "[2] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22560–22570, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.259, + 0.483, + 0.342 + ], + "angle": 0, + "content": "[3] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16123-16133, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.344, + 0.483, + 0.384 + ], + "angle": 0, + "content": "[4] Xuangeng Chu and Tatsuya Harada. Generalizable and animatable gaussian head avatar. arXiv preprint arXiv:2410.07971, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.387, + 0.483, + 0.442 + ], + "angle": 0, + "content": "[5] Xuangeng Chu, Yu Li, Ailing Zeng, Tianyu Yang, Lijian Lin, Yunfei Liu, and Tatsuya Harada. Gpavatar: Generalizable and precise head avatar from image (s). arXiv preprint arXiv:2401.10215, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.445, + 0.483, + 0.514 + ], + "angle": 0, + "content": "[6] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.517, + 0.483, + 0.584 + ], + "angle": 0, + "content": "[7] Yu Deng, Duomin Wang, Xiaohang Ren, Xingyu Chen, and Baoyuan Wang. Portrait4d: Learning one-shot 4d head avatar synthesis using synthetic data. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.588, + 0.483, + 0.657 + ], + "angle": 0, + "content": "[8] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2663-2671, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.66, + 0.483, + 0.728 + ], + "angle": 0, + "content": "[9] Philip-William Grassal, Malte Prinzler, Titus Leistner, Carsten Rother, Matthias Nießner, and Justus Thies. Neural head avatars from monocular rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18653–18664, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.731, + 0.483, + 0.799 + ], + "angle": 0, + "content": "[10] Jianzhu Guo, Dingyun Zhang, Xiaogiang Liu, Zhizhou Zhong, Yuan Zhang, Pengfei Wan, and Di Zhang. Liveportrait: Efficient portrait animation with stitching and retargeting control. arXiv preprint arXiv:2407.03168, 2024. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[11] Xiao Han, Yukang Cao, Kai Han, Xiatian Zhu, Jiankang Deng, Yi-Zhe Song, Tao Xiang, and Kwan-Yee K Wong. Headsculpt: Crafting 3d head avatars with text. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilib-" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.094, + 0.905, + 0.119 + ], + "angle": 0, + "content": "rium. Advances in neural information processing systems, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[13] Fa-Ting Hong, Longhao Zhang, Li Shen, and Dan Xu. Depth-aware generative adversarial network for talking head video generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3397-3406, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.193, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[14] Yang Hong, Bo Peng, Haiyao Xiao, Ligang Liu, and Juyong Zhang. Headnerf: A real-time nerf-based parametric head model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20374-20384, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[15] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.321, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[16] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.378, + 0.905, + 0.433 + ], + "angle": 0, + "content": "[17] Taras Khakhulin, Vanessa Sklyarova, Victor Lempitsky, and Egor Zakharov. Realistic one-shot mesh-based head avatars. In European Conference on Computer Vision, pages 345-362. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.435, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[18] Tobias Kirschstein, Shenhan Qian, Simon Giebenhain, Tim Walter, and Matthias Nießner. Nersemble: Multi-view radiance field reconstruction of human heads. ACM Transactions on Graphics (TOG), 42(4):1-14, 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.491, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[19] Shikai Li, Jianglin Fu, Kaiyuan Liu, Wentao Wang, KwanYee Lin, and Wayne Wu. Cosmicman: A text-to-image foundation model for humans. In Computer Vision and Pattern Recognition (CVPR), 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.548, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[20] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6), 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.605, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[21] Weichuang Li, Longhao Zhang, Dong Wang, Bin Zhao, Zhi-gang Wang, Mulin Chen, Bang Zhang, Zhongjian Wang, Liefeng Bo, and Xuelong Li. One-shot high-fidelity talking-head synthesis with deformable neural radiance field. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17969-17978, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.689, + 0.905, + 0.77 + ], + "angle": 0, + "content": "[22] Xian Liu, Xiaohang Zhan, Jiaxiang Tang, Ying Shan, Gang Zeng, Dahua Lin, Xihui Liu, and Ziwei Liu. Humangau-sian: Text-driven 3d human generation with gaussian splatt-ing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6646-6657, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.774, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[23] Shengjie Ma, Yanlin Weng, Tianjia Shao, and Kun Zhou. 3d gaussian blendshapes for head avatar animation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-10, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.817, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[24] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf:" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.094, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11620" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.133 + ], + "angle": 0, + "content": "Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[26] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[27] Shenhan Qian, Tobias Kirschstein, Liam Schoneveld, Davide Davoli, Simon Giebenhain, and Matthias Nießner. Gaussianavatars: Photorealistic head avatars with rigged 3d gaus-sians. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20299-20309, 2024. 1, 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[28] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.482, + 0.431 + ], + "angle": 0, + "content": "[29] Amit Raj, Srinivas Kaza, Ben Poole, Michael Niemeyer, Nataniel Ruiz, Ben Mildenhall, Shiran Zada, Kfir Aberman, Michael Rubinstein, Jonathan Barron, et al. Dreambooth3d: Subject-driven text-to-3d generation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2349–2359, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.433, + 0.482, + 0.501 + ], + "angle": 0, + "content": "[30] Yurui Ren, Ge Li, Yuanqi Chen, Thomas H Li, and Shan Liu. Pirenderer: Controllable portrait image generation via semantic neural rendering. In Proceedings of the IEEE/CVF international conference on computer vision, pages 13759-13768, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.482, + 0.558 + ], + "angle": 0, + "content": "[31] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on graphics (TOG), 42(1):1-13, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.561, + 0.482, + 0.63 + ], + "angle": 0, + "content": "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[33] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.482, + 0.729 + ], + "angle": 0, + "content": "[34] Jingxiang Sun, Xuan Wang, Lizhen Wang, Xiaoyu Li, Yong Zhang, Hongwen Zhang, and Yebin Liu. Next3d: Generative neural texture rasterization for 3d-aware head avatars. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[35] Felix Taubner, Ruihang Zhang, Mathieu Tuli, and David B Lindell. Cap4d: Creating animatable 4d portrait avatars with morphable multi-view diffusion models. arXiv preprint arXiv:2412.12093, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[36] Kartik Teotia, Mallikarjun B R, Xingang Pan, Hyeongwoo Kim, Pablo Garrido, Mohamed Elgharib, and Christian Theobalt. HQ3davatar: High-quality implicit 3d head avatar. ACM Transactions on Graphics, 43(3):1-24, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[37] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "video diffusion. arXiv preprint arXiv:2403.12008, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[38] Kaisiyuan Wang, Qianyi Wu, Linsen Song, Zhuoqian Yang, Wayne Wu, Chen Qian, Ran He, Yu Qiao, and Chen Change Loy. Mead: A large-scale audio-visual dataset for emotional talking-face generation. In European Conference on Computer Vision, pages 700–717. Springer, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[39] Huawei Wei, Zejun Yang, and Zhisheng Wang. Aniportrait: Audio-driven synthesis of photorealistic portrait animation. arXiv preprint arXiv:2403.17694, 2024. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.301 + ], + "angle": 0, + "content": "[40] Yiqian Wu, Hao Xu, Xiangjun Tang, Xien Chen, Siyu Tang, Zhebin Zhang, Chen Li, and Xiaogang Jin. Portrait3d: Text-guided high-quality 3d portrait generation using pyramid representation and gans prior. ACM Transactions on Graphics (TOG), 43(4):1-12, 2024. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[41] Sicheng Xu, Jiaolong Yang, Dong Chen, Fang Wen, Yu Deng, Yunde Jia, and Xin Tong. Deep 3d portrait from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7710-7720, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[42] Zhenhui Ye, Tianyun Zhong, Yi Ren, Jiaqi Yang, Weichuang Li, Jiawei Huang, Ziyue Jiang, Jinzheng He, Rongjie Huang, Jinglin Liu, et al. Real3d-portrait: One-shot realistic 3d talking portrait synthesis. arXiv preprint arXiv:2401.08503, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[43] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. In European conference on computer vision, pages 85-101. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[44] Fei Yin, Yong Zhang, Xuan Wang, Tengfei Wang, Xiaoyu Li, Yuan Gong, Yanbo Fan, Xiaodong Cun, Ying Shan, Cengiz Oztireli, et al. 3d gan inversion with facial symmetry prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 342-351, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[45] Wangbo Yu, Yanbo Fan, Yong Zhang, Xuan Wang, Fei Yin, Yunpeng Bai, Yan-Pei Cao, Ying Shan, Yang Wu, Zhongqian Sun, et al. Nofa: Nerf-based one-shot facial avatar reconstruction. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-12, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.905, + 0.733 + ], + "angle": 0, + "content": "[46] Bowen Zhang, Yiji Cheng, Jiaolong Yang, Chunyu Wang, Feng Zhao, Yansong Tang, Dong Chen, and Baining Guo. Gaussiancube: Structuring gaussian splatting using optimal transport for 3d generative modeling. arXiv preprint arXiv:2403.19655, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.734, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[47] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.907, + 0.845 + ], + "angle": 0, + "content": "[48] Yufeng Zheng, Wang Yifan, Gordon Wetzstein, Michael J Black, and Otmar Hilliges. Pointavatar: Deformable point-based head avatars from videos. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21057-21067, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[49] Zhenglin Zhou, Fan Ma, Hehe Fan, Zongxin Yang, and Yi Yang. Headstudio: Text to animatable head avatars with 3d gaussian splatting. In European Conference on Computer Vision, pages 145-163. Springer, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11621" + } + ] +] \ No newline at end of file diff --git a/2025/FaceCraft4D_ Animated 3D Facial Avatar Generation from a Single Image/c512f2aa-6fbd-4940-89be-2245ba51ffe2_origin.pdf b/2025/FaceCraft4D_ Animated 3D Facial Avatar Generation from a Single Image/c512f2aa-6fbd-4940-89be-2245ba51ffe2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..eadfb660ec11f3c56cef6396463fbe4cdb3e3763 --- /dev/null +++ b/2025/FaceCraft4D_ Animated 3D Facial Avatar Generation from a Single Image/c512f2aa-6fbd-4940-89be-2245ba51ffe2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:97d1985f65936c83b93e12018a390e7c9c55edbe9a0ab3da1d16dc5959036710 +size 5233353 diff --git a/2025/FaceCraft4D_ Animated 3D Facial Avatar Generation from a Single Image/full.md b/2025/FaceCraft4D_ Animated 3D Facial Avatar Generation from a Single Image/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3b51e3f528a846e8757f0ebd2525190707eec08c --- /dev/null +++ b/2025/FaceCraft4D_ Animated 3D Facial Avatar Generation from a Single Image/full.md @@ -0,0 +1,329 @@ +# FaceCraft4D: Animated 3D Facial Avatar Generation from a Single Image + +Fei Yin $^{1,2}$ , Mallikarjun B R $^{2}$ , Chun-Han Yao $^{2}$ , Rafal K. Mantiuk $^{1\dagger}$ , Varun Jampani $^{2\dagger}$ , $^{1}$ University of Cambridge, $^{2}$ Stability AI + +# Abstract + +We present a novel framework for generating high-quality, animatable 4D avatar from a single image. While recent advances have shown promising results in 4D avatar creation, existing methods either require extensive multiview data or struggle with shape accuracy and identity consistency. To address these limitations, we propose a comprehensive system that leverages shape, image, and video priors to create full-view, animatable avatars. Our approach first obtains initial coarse shape through 3D-GAN inversion. Then, it enhances multiview textures using depth-guided warping signals for cross-view consistency with the help of the image diffusion model. To handle expression animation, we incorporate a video prior with synchronized driving signals across viewpoints. We further introduce a Consistent-Inconsistent training to effectively handle data inconsistencies during 4D reconstruction. Experimental results demonstrate that our method achieves superior quality compared to the prior art, while maintaining consistency across different viewpoints and expressions. + +# 1. Introduction + +4D Avatar generation aims to create animatable 3D head avatars, enabling the generation of consistent identities with custom viewpoints and expressions. High-quality 4D heads have significant applications in gaming, education, and film industries. However, existing 4D avatar building methods [18, 23, 27, 36] often require extensive inputs, such as multiview video with camera pose estimation, which is impractical for widespread use. Furthermore, these methods struggle to produce realistic results for extreme viewpoints and expression changes due to data scarcity, for instance the back view of the head. + +The goal of this work is to create a 4D facial avatar from just a single image. This task is challenging because a single image provides limited information, missing essential details like depth and multiple viewpoints that are crucial + +![](images/e2954b1f9e34b45a8d107e831d9d4f06ee11db2319c7dce3e2ac4d0cdd74f3b3.jpg) + +![](images/7d796904c96fbfd3ab6d848a903ba91d7327258b56d8dc689b592794e70319ae.jpg) +Figure 1. Given a single image input, our method is capable of generating 4D avatars, producing photorealistic textures and consistent expressions across multiple views. Our approach also demonstrates robust performance on challenging inputs, including cartoon characters or old photographs. + +for constructing a complete 3D model. A single image cannot provide information about the unseen parts of a person's head or capture dynamic aspects like facial expressions. It is also impractical to train an end-to-end model due to the lack of multiview full-head animation datasets. To address these challenges, the system needs to rely on prior knowledge, like shape and expression. + +Existing single-image avatar generation methods come with many shortcomings, which we summarize in Tab. 1. The works that rely on a single image of novel identity at test time, typically utilize large-scale, unstructured 2D images [1] and videos [7, 39] during training to account for multiview and temporal aspects. Due to the practical challenges in collecting multiview full-head animation datasets, most single-image methods, including recent work [35], cannot reconstruct a complete 360-degree view of the head. While some approaches can generate full views [1, 40], they typically lack the ability to model facial expressions. Among methods that can animate, some methods rely solely on 2D representation [39], which makes these methods multi-view inconsistent by design. To overcome the above limitation, some methods rely on a hybrid representation + +Table 1. Comparison of methods for avatar generation and animation. The first two methods focus on general object generation, while the remaining methods are specifically designed for avatar applications. + +
MethodsGoalDimensionInput360° viewAnimation3D ConsistencyQuality
LDM [32]Image Generation2DSingle Image-High
DreamFusion [26], SV3D [37]Image-to-3D3DSingle Image✓LowLow
LivePortrait [10], AniPortrait [39]Video Generation2DSingle Image✓LowHigh
PanoHead [1], SPI [44]3D GAN Inversion3DSingle Image✓MediumMedium
Portrait4D [7], CAP4D [35]3D Talking Head3DSingle Image✓MediumHigh
FaceCraft4D4D AvATAR Generation3DSingle Image✓✓HighHigh
HQ3D [36], GA [27]4D AvATAR Generation3DMulti-view Videos✓✓HighHigh
+ +that combines 2D and 3D [7, 34]. However, such methods cannot handle extreme camera angles, as they rely on 2D components for high-resolution synthesis. In contrast, we aim to build our model that has pure 3D representation, and, therefore, is multiview-consistent by design. + +We propose FaceCraft4D, which takes a single image as input and creates a 3D Facial Avatar that can be animated and rendered in 360-degree view, as shown in Fig. 1. To tackle such an under-constrained problem, we make use of 3 different types of priors: a geometry prior, an image prior, and a video prior. We use these priors to synthesize personalized multiview images of the given identity, which are then used to train an explicit 3D model of an avatar that can be animated (see Fig. 2). We will refer to such a model as a 4D avatar. + +We start with a geometry prior from pretrained 3D-GAN [1]. Since, 3D-GAN is trained with large-scale images that comprises of all the views of the head, it has information about the complete head shape. Through 3D-GAN inversion [31], we obtain personalized shape along with an approximate texture. The resulting 3D representation lacks personalized high-quality multiview consistent texture and can not be used for creating novel expressions or animations (see Tab. 1). To improve the quality of the textures, we use of a 2D image prior. To ensure multi-view consistency and preservation of identity across the views, we employ epipolar constraints and cross-view mutual attention. The resulting multi-view images can be used to build a high-quality static 3D head model, which, however, lacks control over expressions and articulations. To animate the avatar, we utilize a 2D video prior [10], which generates training data for facial animation across different viewpoints. + +Finally, for real-time rendering performance, we train a 3D Gaussian representation [16] that is rigged to a FLAME parametric face model [20, 27] using the multi-view videos obtained in the previous steps. To further improve the training consistency across the views and to avoid blur, we propose a COnsistent-INconsistent (COIN) training, which can capture cross-view inconsistencies in an MLP. + +In summary, the contributions of our work are as follows: + +- We present a novel framework for generating high-quality 4D head models from a single image of a face. We com + +pare our framework with previous methods and summarize the differences in Tab. 1. + +- We introduce novel warping-based control generation signals to improve quality and consistency across different views and expressions. +- We introduce COIN training, which enables a 3D Gaussian model to learn from inconsistent data while preserving high-quality, consistent features. + +# 2. Related Work + +3D Head generation. Numerous studies have explored the reconstruction of 3D head models. To obtain sufficient information, most approaches [9, 27, 48, 48] utilize monocular video sequences, which provide multiple viewpoints of the subject's head. However, a significant limitation of these methods is their reliance on multiple viewpoints and lengthy video inputs. They also struggle to accurately reproduce less prominent features in videos, such as the back of the head. + +Another approach is to train a 3D generator on a large head image dataset. EG3D [3] proposed a hybrid tri-plane representation that can quickly and efficiently render high-quality facial images. Panohead [1] and Portrait3D [40] expanded on EG3D by increasing the number of planes, enabling the generation of 3D heads from limited angles to full 360-degree views. However, most methods in this category rely on hybrid representation having a 2D super-resolution network. This impacts their generalization to extreme camera poses and suffer identity inconsistency across views. + +Recently, the rise of text-to-image models has spurred the development of novel approaches for text-to-3D generation tasks. One such approach is Score Distillation Sampling (SDS) [26, 29], which leverages the score function to distill diffusion priors into 3D representations, enabling the generation of 3D content from text or image inputs. Headsculpt [11, 49] incorporates facial keypoint information into the text-to-3D generation method, effectively constraining the facial layout. However, the texture generated by these methods can be oversaturated, making it challenging to achieve photo-realistic head models. + +Head animation. Head animation techniques can be + +![](images/fcc560db6db8a53d342c1d8369e5acf4d14a52206f7b43ce07b7c1995d2cb790.jpg) +Figure 2. Overview of FaceCraft4D: Our approach begins with estimating the geometry of a single input image using a Geometry Prior. This geometry guides the synthesis of personalized multiview images, providing $360^{\circ}$ views and varied expressions, with support from both 2D image and video priors. Since the synthesized data often exhibit inconsistency across views, we propose COIN optimization for robust 4D optimization. By fitting the model to the multiview data, we achieve our final 4D avatar. All faces in this manuscript come from the public FFHQ dataset [15]. + +broadly categorized into 2D-based and 3D-aware approaches, distinguished by their utilization of explicit camera models for rendering. Recent methods [8, 13, 30, 43] have predominantly adopted warping-based strategies, applying deformation fields to intermediate appearance features to capture facial motion characteristics. While effective for frontal and moderate poses, these approaches struggle to maintain geometric consistency during large pose variations due to their inherent lack of 3D modeling. + +To enable free-viewpoint rendering with stronger 3D consistency guarantees, researchers integrated explicit 3D representations into the head synthesis pipeline. Early approaches [17, 41] relied on 3D morphable models for representing head geometry and texture. Subsequent works [5, 14, 21, 42, 45] leveraged more sophisticated representations such as Neural Radiance Fields (NeRFs) [25] to better capture complex structures like hair and accessories. Recent advances like GAGAvatar [4, 22], which employs 3D Gaussian splatting [16], have demonstrated impressive performance in both rendering quality and speed. However, these methods are typically limited in their ability to generate full 360-degree views. Our approach enables true 360-degree view synthesis while maintaining real-time rendering capabilities and achieving superior animation quality. + +# 3. Method + +In this work, we propose a novel framework for generating high-fidelity 4D avatars from a single portrait image $I_{R}$ . Reconstructing a 3D dynamic model from a single 2D image is an inherently ill-posed problem, as critical 3D and temporal information is missing. To address this challenge, we leverage several 2D and 3D priors to compensate for the missing data. + +The overall pipeline, shown in Fig. 2, consists of two stages: personalized multiview generation and 4D repre + +sensation optimization. In the Personalized Multiview Generation stage, we utilize geometry priors to estimate the 3D geometry (Sec. 3.1), use image priors to enhance the texture (Sec. 3.2), and video priors to synthesize high-quality, consistent multi-view videos with dynamic expressions (Sec. 3.3). We then leverage this synthetically-generated personalized image set to optimize a robust 4D representation model (Sec. 3.4). The trained 4D model can be animated by the control parameters of a morphable model, enabling the generation of high-fidelity 4D avatars from a single portrait image. + +# 3.1. Geometry Prior + +The initialization is crucial for 3D generation. A good initialization should follow the approximate geometry of the head shown in the input image, $I_{R}$ . For that, we use the prior provided by a 3D-aware generator, PanoHead [1, 40], which was trained on a large $360^{\circ}$ head dataset. We obtain a coarse 3D reconstruction for a given input image $I_{R}$ through GAN inversion. Specifically, we first optimize the latent code $z$ within the generator's domain by minimizing a combination of pixel-wise $\mathcal{L}_2$ loss and image-level LPIPS loss [47]. Subsequently, following [31], we keep the optimized latent code $z$ fixed while fine-tuning the 3D generator's parameters. During this fine-tuning stage, we maintain the same loss functions used in the latent optimization phase to ensure the rendered views closely match the input image. This process gives us an initial 3D head representation, though one that cannot be animated. The optimized generator is then used to synthesize a set of approximate multiview images $\{I_i\}$ and their corresponding depth maps $\{D_i\}$ , where $i$ is the view index. + +![](images/2bf14056c35e2f198145d4363f0ed09d58925f35166af031f7950102ec378c51.jpg) + +![](images/0c3aebe297ab434d9984668cebe207fbe7b271a8818fb8f306b617d5cce678d0.jpg) +Figure 3. Image Prior: We introduce a cross-view mutual attention mechanism and epipolar constraints to enhance consistency in generated novel views. Our approach aligns reference and target images, maintaining visual coherence across viewpoints. +Figure 4. Triplane-based methods (e.g. PanoHead) are sensitive to focal length (scale) and exhibit significant degradation of quality as focal length decreases. In contrast, our Gaussian-based methods maintain robust performance across varying focal lengths. + +# 3.2. Image Prior + +The textures synthesized from the geometry prior may lack visual quality, and the identity of the individual may not be well preserved across the views. In particular, when rendering with different focal lengths, triplane-based methods [1] exhibit substantial degradation in image quality, as illustrated in Fig. 4. We enhance the generated multiview images using 2D diffusion models [32], which have demonstrated powerful generative capabilities. Although direct image-to-image translation [24] appears to be a straightforward solution, this approach poses significant risks: the de-noising process may introduce cross-view inconsistencies + +and also modify the original semantic content. To address these risks, we propose two strategies that constrain the generation process, which we describe next and illustrate in Fig. 3. + +Cross-view mutual attention. Inspired by MasaCtrl [2], we introduce a cross-view attention mechanism to replace the standard self-attention during the generation process. This modification allows us to inject information from the reference image into the novel view. Specifically, we first perform DDIM inversion [33] to introduce noise to both the reference frame $I_{R}$ and the coarse image $I_{i}$ of the target novel view. This step preserves the overall face structure and contour. Then, the diffusion model is used to gradually denoise two images to enhance the texture. In this process, we replace the key $K$ and value $V$ matrices (calculated in self-attention) of the novel view with those derived from the reference image. Different from [2], we adopt a batch-processing strategy in which multiple images $I_{i}$ representing different viewpoints are treated as a single batch, with the reference image $I_{R}$ serving as the consistent source for all subsequent calculations. This not only streamlines the generation process but also ensures uniformity in feature injection across novel viewpoints. + +The rationale behind this approach lies in the observation that, despite variations in viewpoint, the underlying textures (represented by the attention values) should maintain consistency across the views. The cross-view attention mechanism enables the transfer of information from the reference to the novel views, enhancing both the fidelity and coherence of the generated novel views. + +Warping-based control generation. The consistency across the views is further enforced by epipolar constraints, which enforce geometric relationships between novel views and the input view. We first use cross-view mutual attention to generate the view of the back of the head. The generated back view along with the reference image are set as anchors. Then, we propagate the texture information to novel views through depth-based warping. Specifically, we lift the anchor image into 3D space and then project it to the adjacent view with depth guidance $\{D_i\}$ . During the projection, we form a mask to indicate the visibility of pixels to filter invisible areas using the rendered depth values. The projected image is then used to guide the texture enhancement process. The intermediate latent of the enhanced image in the diffusion model is blended with the latent of the projected image via a simple mask copy-pasting operation. By combining the latent, the resulting images benefit from both the improved consistency provided by the epipolar constraints and the enhanced quality achieved through the denoising process. We repeat this process for various azimuth angles, creating a comprehensive set of viewpoints. + +With the aforementioned guidance and epipolar constraints, we can ensure coherence across the entire view + +![](images/377bb3b1d7a5824f47a2fdc08872b959557f55bd95fa37e26c74b0c9c011193d.jpg) +Figure 5. COIN-Training: To address geometric and color inconsistencies between the views, we train two 3D representations: a view-consistent base model (GaussianAvatar) and an MLP, encoding inconsistencies between the views. The MLP with inconsistencies lets us robustly reconstruct the high-quality 3D base model. + +space and generate high-fidelity view-consistent images $\{I_i^*\}$ for further reconstruction. + +# 3.3. Video Prior + +The view-consistent data created in the previous step has a static expression similar to that of reference image $I_R$ . This may not be enough to create diverse expressions as the images do not contain some of the expression-dependent information. For example, if the reference image has the mouth closed, the synthesized dataset would not have details inside the mouth region. To incorporate expression-dependent textures and geometry that may not be available in the multiview image generation stage (Sec. 3.2), we need to create multi-view facial expression data. We use LivePortrait [10] as a video prior, which can help generate multiview expressive videos. LivePortrait uses a single source image and a driving video to generate new videos where the source identity mimics the expressions and poses from the target video. Instead of directly copying the driving video's motion, one has the option of capturing the changes in relative expression but maintain the pose of the input image. Given that we have the reference and multiview-consistent enhanced images $\{I_i^*\}$ from the previous step, we provide these to LivePortrait as the source images together with driving videos with diverse poses and expressions (details in Sec. 4), and obtain target videos. Because the identity is consistent across the input views, the resulting videos also maintain the identity. Using identical driving videos ensures synchronized expressions. + +# 3.4. Robust 4D optimization with COIN training + +The synthesized multi-view videos have high-quality textures and good cross-view consistency, but small misalignment of features and colors is unavoidable. Learning 4D representation directly on such data leads to blurry textures + +for the rendered objects. Here, we address this problem by proposing a robust training paradigm. At a high level, we learn a base representation that is common to all views, and a delta representation that is specific to each view during training. This helps in cornering inconsistencies concerning each view to separate representation. The base model will have a coarse but consistent representation learned, while the high-frequency details end up in the delta-space representation. At test time, we choose one of the views as our reference view and use its delta representation as the common high-frequency details for all the views. + +In specific, we achieve this by jointly training two 4D representations: a consistent GaussianAvatar [27], which represents the 3D structure using Gaussians bound to a FLAME-tracked [20] mesh, and a NeRF-style MLP, tasked with encoding inconsistencies between the views. As shown in Fig. 5, we design the losses that prioritize structure reconstruction in the consistent GaussianAvatar (LPIPS), and high-frequency detail reconstruction in the MLP with the view-dependent inconsistencies (L1 and SSIM). By isolating view-dependent variations in a separate model, this strategy prevents these inconsistencies from corrupting the base representation. The approach can be considered a form of robust regression, with the key difference that the view-dependent MLP ensures that inconsistencies (outliers) are spatially localized. This cannot be achieved with robust norms (e.g., L1), which do not maintain the information about the spatial position of outliers. + +We choose GaussianAvatar as our consistent representation because it lets us animate the avatar with controllable FLAME expression and pose parameters, providing a solid foundation for modeling consistent geometric and appearance features. The details of FLAME tracking and controlling are explained in Sec. 1.3 of the supplementary. To model the view-dependent variations, we design a lightweight MLP-based architecture that introduces minimal computational overhead while effectively capturing view-specific details. The network encodes the offset in the color of each Gaussian (see Fig. 5), which can be formulated as: + +$$ +c _ {\text {o f f s e t}} = \operatorname {M L P} \left(e _ {\text {v i e w}}, c, e _ {g}; \theta\right), \tag {1} +$$ + +where $e_{\mathrm{view}} \in \mathbb{R}^d$ is a learnable view-specific embedding that encodes view-representation, $c$ represents the initial Gaussian color attributes from the GaussianAvatar, $e_g$ is a position embedding for each Gaussian, accounting for location-specific color variation ranges (e.g., eyes vs. mouth regions), and $\theta$ is a vector of the trainable parameters. The final view-dependent color is computed as the sum of the base GaussianAvatar color and the offset: $c^* = c + c_{\mathrm{offset}}$ . The MLP consists of only two layers to maintain computational efficiency and preserve the real-time rendering capabilities of the Gaussian-based representation. + +During training, we render two types of images: $I_{i}^{\mathrm{C}}$ from + +![](images/7cce182327f9a8e2e3eb4fa9477028bf878112b3a7cb62d60a165e40a2ded385.jpg) +Figure 6. Qualitative comparison on static 3D head generation from a single image. + +the consistent component alone, and $I_{i}^{\mathrm{IC}}$ from the combination of both components. To ensure accurate reconstruction of high-frequency details, we supervise $I_{i}^{\mathrm{IC}}$ using pixel-wise losses: + +$$ +\mathcal {L} _ {\text {p i x e l}} = \lambda_ {1} \mathcal {L} 1 \left(I _ {i} ^ {\mathrm {I C}}, I _ {i} ^ {*}\right) + \lambda_ {\text {S S I M}} \text {S S I M} \left(I _ {i} ^ {\mathrm {I C}}, I _ {i} ^ {*}\right) \tag {2} +$$ + +To prevent the inconsistent component from dominating the representation and maintain the effectiveness of the consistent component, we introduce a structural supervision loss and a regularization term: + +$$ +\mathcal {L} _ {\text {s t r u c}} = \lambda_ {\text {L P I P S}} \operatorname {L P I P S} \left(I _ {i} ^ {\mathrm {C}}, I _ {i} ^ {*}\right) \tag {3} +$$ + +$$ +\mathcal {L} _ {\text {r e g}} = \lambda_ {\text {o f f s e t}} \mathcal {L} 1 \left(c _ {\text {o f f s e t}}, 0\right), \tag {4} +$$ + +where LPIPS loss ensures image-level supervision for the consistent component, and the offset regularization encourages minimal view-dependent modifications. The hyperparameters are set to $\lambda_{1} = 0.8$ , $\lambda_{\mathrm{SSIM}} = 0.2$ , $\lambda_{\mathrm{LPIPS}} = 0.05$ and $\lambda_{\mathrm{offset}} = 1$ . + +During inference, we input a single fixed view-embedding for the inconsistent component. Given that our reference image $I_R$ contains superior quality and identity information compared to the generated multiview images $\{I_i^*\}$ , we select its corresponding view embedding for inference. Novel expressions and poses can be easily gen + +erated by adjusting the FLAME parameters, leveraging the animation capabilities of our backbone model. + +# 4. Experiments + +Implementation details. During the multiview generation stage, we use PanoHead [1] as our geometry prior, Cosmicman [19] as the image prior, and LivePortrait [10] as the video prior model. We render 24 views along the horizon orbit for static reconstruction. For animation, we select $I_R$ and 11 frontal views from these generated images as source images and randomly sample 8 video clips from the NerSemble [18] dataset as driving videos. The source images and driving videos are then input to the video prior model to produce multiview video data. In the 4D optimization stage, we first optimize the GaussianAvatar [27] model with the static data over 30K iterations. Next, we apply COIN-optimization techniques, finetuning the static Gaussian model over 90K iterations. The complete process takes approximately two hours to generate a single 4D avatar. + +Evaluation metrics. In the absence of ground truth data, we use commonly adopted non-reference metrics to evaluate the rendered images in terms of image quality and identity similarity. For overall image quality, we use the FID [12] score. To assess multi-view consistency, we employ the CLIP-I [28] score. For identity preservation, we calculate ID scores by measuring the cosine distance of face + +![](images/5e1438b7557703f1f8df2d2be930093745f1411125ca5f79cae0b730defa889f.jpg) +Figure 7. Qualitative comparison on animation of different views. + +![](images/6d13b9e1d13f030a9266c8ef6892f310f5e90c5cc05abeafb50e4cf4161d9c99.jpg) + +![](images/b109ea62d5b353691c4d537a77651f346c66ee6e069629cfb0a78faf14676417.jpg) + +![](images/7a7b66431ef2e601978661e80ed7d1132b11c58ba801eff2abdb3ae6e7a9238e.jpg) + +![](images/4853a7dc0f8bc84c4ab8f30cdd89f286eada173252fea01c25f6ca351732ca09.jpg) + +Table 2. Quantitative comparison on static 3D head generation. + +
MethodCLIP-I ↑ID ↑FID ↓
GaussianCube [46]0.68300.4300258.81
PanoHead [1]0.82330.4246195.28
SV3D [37]0.76560.4331234.86
Portrait3D [40]0.70660.3719302.74
Ours0.80530.5082174.36
Table 3. Quantitative comparison on 3D head animation.
MethodCLIP-I ↑ID ↑FID ↓
AntiPortrait [39]0.46530.4171364.99
Portrait-4D-V2 [7]0.52360.4592248.36
Ours0.57370.4602201.76
+ +recognition features [6] between the novel views and the reference images. + +Datasets. We conducted our analysis on a subset of facial images from FFHQ [15] dataset, a high-quality, in-the-wild facial dataset. For evaluation, we select 100 images with unoccluded faces to evaluate our model and baselines. + +# 4.1. Static 3D Generation Comparisons + +Baselines. We compare our method with existing image-to-3D approaches, including a general object-focused model, SV3D [37], and three head-specific methods: the inference-based GaussianCube [46], the 3D-GAN-based PanoHead [1], and the optimization-based Portrait3D [40]. + +Qualitative results. Fig. 6 shows visual comparisons where our method achieves the highest visual fidelity and identity coherence compared to other approaches. Notably, it preserves details such as teeth and earrings in side views. In contrast, SV3D, which focuses on general objects, struggles to capture accurate head geometry, often degenerating into a flat plane. GaussianCube tends to "toonify" the input head due to the nature of its training data. While PanoHead achieves good visual quality, it suffers from detail loss and floating artifacts during pose rotation. Portrait3D also falls short in quality, as the SDS [26] optimization introduces + +variations that lead to blurring artifacts. + +Quantitative results. For each generated head model, we render images from five viewpoints: left-frontal, left, right-frontal, right, and frontal. All five rendered images, along with the reference image, are used to compute the CLIP-I, ID, and FID metrics. As shown in Tab. 2, our method outperforms others in both ID and FID scores. This aligns with the observed visual fidelity, as our results appear more realistic and coherent. Our method achieves higher identity similarity and better detail preservation, even in side views. The results for the MEAD multiview dataset [38] can be found in the supplementary material. + +# 4.2. Animation Comparisons + +Baselines. We compare our method with one-shot video-based head reenactment approaches, including the 2D-based AniPortrait [39] and the 3D-based Portrait-4D-V2 [7]. For AniPortrait, which uses keypoints as intermediate driving signals, we modify the Euler angles of the expression image to rotate the keypoints and generate novel view videos. In the case of Portrait-4D-V2, novel views can be generated directly by editing the camera parameters. + +Qualitative results. As shown in Fig. 7, our method preserves expressions with greater accuracy. Portrait-4D-V2 lacks a back-of-head model, so when rotated to side views, the back appears empty. AniPortrait, as a 2D-based method, can achieve high texture quality. However, with large-angle rotations, AniPortrait struggles to preserve identity details and introduces noticeable artifacts in the hair and background. In contrast, our method maintains accurate geometry and expression, even with camera movement. + +Quantitative results. We test our data on unseen views (especially $60^{\circ} - 180^{\circ}$ ), to highlight each method's performance under varying viewing angles, as handling extreme rotations is often challenging. The results are presented in Tab. 3. Our method outperforms all baselines, demonstrating robustness across varied views. The performance + +![](images/313be90bd12dd852f03f64c64227a372a430ba2ab777a5d8e601c53b8ba97d51.jpg) +Input + +![](images/c122fd351ad377d65f478f329105424842d4f988e2d4e2d97c02254a613d3541.jpg) +Warped Image + +![](images/bcd261144454207b004104adb87ef4c5910f455d7c766f99d3fb9c32de343a50.jpg) +w/o Warping + +![](images/17e8e167a194f0873557dd97b390d9bf23e21edce6eb5566aaa09439d85bdcc0.jpg) +w/o MA + +![](images/88c00eaea3dac1381f6c9e0670bcac73027c4fa382bf7416aeb64afbb342edaf.jpg) +Full + +![](images/ed7841e80558856f152deb92b73f1cbaf90b5d5513955e9f8a345126ad1d2b4a.jpg) +Figure 8. Ablation of multiview image generation modules. + +![](images/e50304eb204c57a4f80602f2869503c39f1193f82eb80be5a54a72704d394212.jpg) + +![](images/98c1170d45a8510af70f4e188cbff88bcb97da8c554d8f0d8a23f4babf0f98a4.jpg) +w/o Video Prior + +![](images/7aabae2050151d565e62a780db6775c03367b5eec66715a7e74586ade5be315d.jpg) + +![](images/b3b2915c9d660ab5e193b688260a8e4ca015760adf79932434949553c1592ef9.jpg) +w/o COIN + +![](images/dc449449143426ae99c281625dd1028f06f3ce8b31f53ea2c69d790baec6d52f.jpg) + +![](images/08c1a764200f0d51043005c6386236ed3e06c6ec8600f4d80f6427aefe37419a.jpg) +Full + +![](images/5c196a350c7472c310176fc8ce70aeb206faae6cbd826a59d637ca825c5e440c.jpg) +Input +Figure 9. Ablation of animation modules. +Input +Novel View + +![](images/4127dca5d723113e4c933bb28e37d4e7c6c17bcc2a951a08f3485c52f1eac1f9.jpg) + +![](images/a107644927fae515ea994a11560366e629251892ccfb7ad688dcc80cc905d59b.jpg) +Input +Novel View +Figure 10. More results with diverse inputs. + +![](images/3dbd67799bd7684a42045d51ab257f56ca9a16e74dd2b53a19969db5d7baccfd.jpg) + +![](images/7868cc0faa3507170f6b59803cd8692bb4cefbaf9270033afc15e7cdbd91e0e5.jpg) +Input + +![](images/6392f69b9f4e1c2c849e3cfdc472430ea8278a00964b2454426cbea5b161daea.jpg) +Novel View + +of baselines degrades significantly as the viewing angle increases. Additional results and baseline comparisons are provided in the supplementary material. + +# 4.3. Ablation Study + +Multiview image generation. We conduct an ablation study on two multi-view generation modules, as shown in Tab. 4 and Fig. 8. Incorporating the warping module allows us to preserve finer details, like tattoos, in novel views. While removing MA may lead to semantic inconsistencies, such as changes in gender. Adding mutual attention (MA) further improves fidelity and enhances identity similarity by injecting features from the reference images. + +Animation. Our 3D representation is based on GaussianAvatar [27], which is FLAME-driven, enabling direct animation through FLAME parameters. To evaluate the base performance, we remove our video prior module. As shown in Fig. 9, without the video prior, the final avatar fails to synthesize accurate expressions. For consistency modeling, we compare our model with and without the COIN training. Without COIN, training directly on inconsistent generated images results in blurred textures. With COIN training, regions such as the teeth remain sharp, and finer details, like individual hair strands, are preserved. The quantitative results in Tab. 5 further validate the effectiveness of the COIN optimization across novel views. + +Processing times. We analyzed the processing times of each stage. Our method consists of three stages: a) 3D GAN inversion (2 mins); b) Multiview image & video generation + +Table 4. Ablation of multiview image generation modules. "Warp" refers to the warping-based control generation (Sec. 3.2) and MA to cross-view mutual attention (Sec. 3.2). + +
MethodCLIP-I ↑ID ↑FID
w/o Warp w/o MA0.73280.4787182.27
w/o Warp0.78860.4915171.08
w/o MA0.81510.4951172.43
Full0.81620.4984166.96
+ +Table 5. Ablation of animation modules. + +
MethodCLIP-I ↑ID ↑FID ↓
w/o COIN0.76880.4952144.95
Full0.77290.5010142.80
+ +Table 6. Time comparison with optimization-based avatar reconstruction methods. + +
MethodNerFaceIMAvatarPointAvatarHeadStudio
Time (hour)54h48h6h2h
MethodMonoGAGaussianAvatarHq3davatarOurs
Time (hour)7h2h12h2.5h
+ +(10 mins), FLAME parameters extraction (10 mins); and c) COIN-training to overfit multiview videos (2 hours). The generation of a single avatar $\sim 2.5$ hours, which is comparable to or faster than other optimization-based avatar reconstruction methods, as shown in Tab. 6. At inference, we achieve real time rendering at 156 FPS with a resolution of $512 \times 512$ . + +Diverse inputs. To validate robustness, we conduct experiments on diverse inputs, including cartoon, line drawing, and portraits with extreme camera poses. The results in Fig. 10 demonstrate consistent performance across various domains. Additional video results are available on the webpage in the supplementary materials. + +# 5. Conclusions + +In this paper, we present a novel method for generating high-quality 4D portraits from a single image. By leveraging multiple priors, including 3D-GAN for geometry initialization, image prior for texture enhancement, and video prior for animation, our method can generate animated avatars with detailed textures, excellent consistency, and plausible rendition of views missing in the input. The proposed depth-guided warping ensures cross-view consistency during the generation process, while our COIN-training strategy enables high-quality 4D reconstruction from potentially inconsistent multi-view data. Extensive experiments demonstrate that our method outperforms existing approaches in terms of geometry accuracy, texture quality, and identity preservation across different viewpoints and expressions. Our framework makes high-quality 4D portrait generation more accessible by removing the requirement for multi-view data capture, paving the way for broader applications. + +# References + +[1] Sizhe An, Hongyi Xu, Yichun Shi, Guoxian Song, Umit Y Ogras, and Linjie Luo. Panohead: Geometry-aware 3d full-head synthesis in 360deg. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20950-20959, 2023. 1, 2, 3, 4, 6, 7 +[2] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22560–22570, 2023. 4 +[3] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16123-16133, 2022. 2 +[4] Xuangeng Chu and Tatsuya Harada. Generalizable and animatable gaussian head avatar. arXiv preprint arXiv:2410.07971, 2024.3 +[5] Xuangeng Chu, Yu Li, Ailing Zeng, Tianyu Yang, Lijian Lin, Yunfei Liu, and Tatsuya Harada. Gpavatar: Generalizable and precise head avatar from image (s). arXiv preprint arXiv:2401.10215, 2024.3 +[6] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 7 +[7] Yu Deng, Duomin Wang, Xiaohang Ren, Xingyu Chen, and Baoyuan Wang. Portrait4d: Learning one-shot 4d head avatar synthesis using synthetic data. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1, 2, 7 +[8] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2663-2671, 2022. 3 +[9] Philip-William Grassal, Malte Prinzler, Titus Leistner, Carsten Rother, Matthias Nießner, and Justus Thies. Neural head avatars from monocular rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18653–18664, 2022. 2 +[10] Jianzhu Guo, Dingyun Zhang, Xiaogiang Liu, Zhizhou Zhong, Yuan Zhang, Pengfei Wan, and Di Zhang. Liveportrait: Efficient portrait animation with stitching and retargeting control. arXiv preprint arXiv:2407.03168, 2024. 2, 5, 6 +[11] Xiao Han, Yukang Cao, Kai Han, Xiatian Zhu, Jiankang Deng, Yi-Zhe Song, Tao Xiang, and Kwan-Yee K Wong. Headsculpt: Crafting 3d head avatars with text. Advances in Neural Information Processing Systems, 36, 2024. 2 +[12] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilib- + +rium. Advances in neural information processing systems, 30, 2017. 6 +[13] Fa-Ting Hong, Longhao Zhang, Li Shen, and Dan Xu. Depth-aware generative adversarial network for talking head video generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3397-3406, 2022. 3 +[14] Yang Hong, Bo Peng, Haiyao Xiao, Ligang Liu, and Juyong Zhang. Headnerf: A real-time nerf-based parametric head model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20374-20384, 2022. 3 +[15] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 3, 7 +[16] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 3 +[17] Taras Khakhulin, Vanessa Sklyarova, Victor Lempitsky, and Egor Zakharov. Realistic one-shot mesh-based head avatars. In European Conference on Computer Vision, pages 345-362. Springer, 2022. 3 +[18] Tobias Kirschstein, Shenhan Qian, Simon Giebenhain, Tim Walter, and Matthias Nießner. Nersemble: Multi-view radiance field reconstruction of human heads. ACM Transactions on Graphics (TOG), 42(4):1-14, 2023. 1, 6 +[19] Shikai Li, Jianglin Fu, Kaiyuan Liu, Wentao Wang, KwanYee Lin, and Wayne Wu. Cosmicman: A text-to-image foundation model for humans. In Computer Vision and Pattern Recognition (CVPR), 2024. 6 +[20] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6), 2017. 2, 5 +[21] Weichuang Li, Longhao Zhang, Dong Wang, Bin Zhao, Zhi-gang Wang, Mulin Chen, Bang Zhang, Zhongjian Wang, Liefeng Bo, and Xuelong Li. One-shot high-fidelity talking-head synthesis with deformable neural radiance field. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17969-17978, 2023. 3 +[22] Xian Liu, Xiaohang Zhan, Jiaxiang Tang, Ying Shan, Gang Zeng, Dahua Lin, Xihui Liu, and Ziwei Liu. Humangau-sian: Text-driven 3d human generation with gaussian splatt-ing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6646-6657, 2024. 3 +[23] Shengjie Ma, Yanlin Weng, Tianjia Shao, and Kun Zhou. 3d gaussian blendshapes for head avatar animation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-10, 2024. 1 +[24] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 4 +[25] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: + +Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3 +[26] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2, 7 +[27] Shenhan Qian, Tobias Kirschstein, Liam Schoneveld, Davide Davoli, Simon Giebenhain, and Matthias Nießner. Gaussianavatars: Photorealistic head avatars with rigged 3d gaus-sians. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20299-20309, 2024. 1, 2, 5, 6, 8 +[28] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6 +[29] Amit Raj, Srinivas Kaza, Ben Poole, Michael Niemeyer, Nataniel Ruiz, Ben Mildenhall, Shiran Zada, Kfir Aberman, Michael Rubinstein, Jonathan Barron, et al. Dreambooth3d: Subject-driven text-to-3d generation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2349–2359, 2023. 2 +[30] Yurui Ren, Ge Li, Yuanqi Chen, Thomas H Li, and Shan Liu. Pirenderer: Controllable portrait image generation via semantic neural rendering. In Proceedings of the IEEE/CVF international conference on computer vision, pages 13759-13768, 2021. 3 +[31] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on graphics (TOG), 42(1):1-13, 2022. 2, 3 +[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 4 +[33] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.4 +[34] Jingxiang Sun, Xuan Wang, Lizhen Wang, Xiaoyu Li, Yong Zhang, Hongwen Zhang, and Yebin Liu. Next3d: Generative neural texture rasterization for 3d-aware head avatars. In CVPR, 2023. 2 +[35] Felix Taubner, Ruihang Zhang, Mathieu Tuli, and David B Lindell. Cap4d: Creating animatable 4d portrait avatars with morphable multi-view diffusion models. arXiv preprint arXiv:2412.12093, 2024. 1, 2 +[36] Kartik Teotia, Mallikarjun B R, Xingang Pan, Hyeongwoo Kim, Pablo Garrido, Mohamed Elgharib, and Christian Theobalt. HQ3davatar: High-quality implicit 3d head avatar. ACM Transactions on Graphics, 43(3):1-24, 2024. 1, 2 +[37] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent + +video diffusion. arXiv preprint arXiv:2403.12008, 2024. 2, 7 +[38] Kaisiyuan Wang, Qianyi Wu, Linsen Song, Zhuoqian Yang, Wayne Wu, Chen Qian, Ran He, Yu Qiao, and Chen Change Loy. Mead: A large-scale audio-visual dataset for emotional talking-face generation. In European Conference on Computer Vision, pages 700–717. Springer, 2020. 7 +[39] Huawei Wei, Zejun Yang, and Zhisheng Wang. Aniportrait: Audio-driven synthesis of photorealistic portrait animation. arXiv preprint arXiv:2403.17694, 2024. 1, 2, 7 +[40] Yiqian Wu, Hao Xu, Xiangjun Tang, Xien Chen, Siyu Tang, Zhebin Zhang, Chen Li, and Xiaogang Jin. Portrait3d: Text-guided high-quality 3d portrait generation using pyramid representation and gans prior. ACM Transactions on Graphics (TOG), 43(4):1-12, 2024. 1, 2, 3, 7 +[41] Sicheng Xu, Jiaolong Yang, Dong Chen, Fang Wen, Yu Deng, Yunde Jia, and Xin Tong. Deep 3d portrait from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7710-7720, 2020. 3 +[42] Zhenhui Ye, Tianyun Zhong, Yi Ren, Jiaqi Yang, Weichuang Li, Jiawei Huang, Ziyue Jiang, Jinzheng He, Rongjie Huang, Jinglin Liu, et al. Real3d-portrait: One-shot realistic 3d talking portrait synthesis. arXiv preprint arXiv:2401.08503, 2024. 3 +[43] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. In European conference on computer vision, pages 85-101. Springer, 2022. 3 +[44] Fei Yin, Yong Zhang, Xuan Wang, Tengfei Wang, Xiaoyu Li, Yuan Gong, Yanbo Fan, Xiaodong Cun, Ying Shan, Cengiz Oztireli, et al. 3d gan inversion with facial symmetry prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 342-351, 2023. 2 +[45] Wangbo Yu, Yanbo Fan, Yong Zhang, Xuan Wang, Fei Yin, Yunpeng Bai, Yan-Pei Cao, Ying Shan, Yang Wu, Zhongqian Sun, et al. Nofa: Nerf-based one-shot facial avatar reconstruction. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-12, 2023. 3 +[46] Bowen Zhang, Yiji Cheng, Jiaolong Yang, Chunyu Wang, Feng Zhao, Yansong Tang, Dong Chen, and Baining Guo. Gaussiancube: Structuring gaussian splatting using optimal transport for 3d generative modeling. arXiv preprint arXiv:2403.19655, 2024. 7 +[47] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 3 +[48] Yufeng Zheng, Wang Yifan, Gordon Wetzstein, Michael J Black, and Otmar Hilliges. Pointavatar: Deformable point-based head avatars from videos. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21057-21067, 2023. 2 +[49] Zhenglin Zhou, Fan Ma, Hehe Fan, Zongxin Yang, and Yi Yang. Headstudio: Text to animatable head avatars with 3d gaussian splatting. In European Conference on Computer Vision, pages 145-163. Springer, 2024. 2 \ No newline at end of file diff --git a/2025/FaceCraft4D_ Animated 3D Facial Avatar Generation from a Single Image/images.zip b/2025/FaceCraft4D_ Animated 3D Facial Avatar Generation from a Single Image/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1b2379a92757e2f5c19fd5bf144ede8bcae39351 --- /dev/null +++ b/2025/FaceCraft4D_ Animated 3D Facial Avatar Generation from a Single Image/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ebabc8b3cc0368dcf0b27c07cc07440d322a8d101ea54d1e27b3124fcb75cbb4 +size 605301 diff --git a/2025/FaceCraft4D_ Animated 3D Facial Avatar Generation from a Single Image/layout.json b/2025/FaceCraft4D_ Animated 3D Facial Avatar Generation from a Single Image/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..63f6ca7a3fb5fb29cd38e51d85d73e34f1bff9a1 --- /dev/null +++ b/2025/FaceCraft4D_ Animated 3D Facial Avatar Generation from a Single Image/layout.json @@ -0,0 +1,8541 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 76, + 103, + 536, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 103, + 536, + 122 + ], + "spans": [ + { + "bbox": [ + 76, + 103, + 536, + 122 + ], + "type": "text", + "content": "FaceCraft4D: Animated 3D Facial Avatar Generation from a Single Image" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 142, + 506, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 142, + 506, + 172 + ], + "spans": [ + { + "bbox": [ + 104, + 142, + 506, + 172 + ], + "type": "text", + "content": "Fei Yin" + }, + { + "bbox": [ + 104, + 142, + 506, + 172 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 104, + 142, + 506, + 172 + ], + "type": "text", + "content": ", Mallikarjun B R" + }, + { + "bbox": [ + 104, + 142, + 506, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 104, + 142, + 506, + 172 + ], + "type": "text", + "content": ", Chun-Han Yao" + }, + { + "bbox": [ + 104, + 142, + 506, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 104, + 142, + 506, + 172 + ], + "type": "text", + "content": ", Rafal K. Mantiuk" + }, + { + "bbox": [ + 104, + 142, + 506, + 172 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 104, + 142, + 506, + 172 + ], + "type": "text", + "content": ", Varun Jampani" + }, + { + "bbox": [ + 104, + 142, + 506, + 172 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 104, + 142, + 506, + 172 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 104, + 142, + 506, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 104, + 142, + 506, + 172 + ], + "type": "text", + "content": "University of Cambridge, " + }, + { + "bbox": [ + 104, + 142, + 506, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 104, + 142, + 506, + 172 + ], + "type": "text", + "content": "Stability AI" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 199, + 201, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 199, + 201, + 213 + ], + "spans": [ + { + "bbox": [ + 151, + 199, + 201, + 213 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 224, + 296, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 224, + 296, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 296, + 452 + ], + "type": "text", + "content": "We present a novel framework for generating high-quality, animatable 4D avatar from a single image. While recent advances have shown promising results in 4D avatar creation, existing methods either require extensive multiview data or struggle with shape accuracy and identity consistency. To address these limitations, we propose a comprehensive system that leverages shape, image, and video priors to create full-view, animatable avatars. Our approach first obtains initial coarse shape through 3D-GAN inversion. Then, it enhances multiview textures using depth-guided warping signals for cross-view consistency with the help of the image diffusion model. To handle expression animation, we incorporate a video prior with synchronized driving signals across viewpoints. We further introduce a Consistent-Inconsistent training to effectively handle data inconsistencies during 4D reconstruction. Experimental results demonstrate that our method achieves superior quality compared to the prior art, while maintaining consistency across different viewpoints and expressions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 475, + 135, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 135, + 488 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 135, + 488 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 495, + 296, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 296, + 626 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 296, + 626 + ], + "type": "text", + "content": "4D Avatar generation aims to create animatable 3D head avatars, enabling the generation of consistent identities with custom viewpoints and expressions. High-quality 4D heads have significant applications in gaming, education, and film industries. However, existing 4D avatar building methods [18, 23, 27, 36] often require extensive inputs, such as multiview video with camera pose estimation, which is impractical for widespread use. Furthermore, these methods struggle to produce realistic results for extreme viewpoints and expression changes due to data scarcity, for instance the back view of the head." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 628, + 296, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 296, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 296, + 676 + ], + "type": "text", + "content": "The goal of this work is to create a 4D facial avatar from just a single image. This task is challenging because a single image provides limited information, missing essential details like depth and multiple viewpoints that are crucial" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 198, + 566, + 287 + ], + "blocks": [ + { + "bbox": [ + 317, + 198, + 566, + 287 + ], + "lines": [ + { + "bbox": [ + 317, + 198, + 566, + 287 + ], + "spans": [ + { + "bbox": [ + 317, + 198, + 566, + 287 + ], + "type": "image", + "image_path": "e2954b1f9e34b45a8d107e831d9d4f06ee11db2319c7dce3e2ac4d0cdd74f3b3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 289, + 564, + 376 + ], + "blocks": [ + { + "bbox": [ + 317, + 289, + 564, + 376 + ], + "lines": [ + { + "bbox": [ + 317, + 289, + 564, + 376 + ], + "spans": [ + { + "bbox": [ + 317, + 289, + 564, + 376 + ], + "type": "image", + "image_path": "7d796904c96fbfd3ab6d848a903ba91d7327258b56d8dc689b592794e70319ae.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 384, + 555, + 440 + ], + "lines": [ + { + "bbox": [ + 313, + 384, + 555, + 440 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 555, + 440 + ], + "type": "text", + "content": "Figure 1. Given a single image input, our method is capable of generating 4D avatars, producing photorealistic textures and consistent expressions across multiple views. Our approach also demonstrates robust performance on challenging inputs, including cartoon characters or old photographs." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 449, + 556, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 556, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 556, + 533 + ], + "type": "text", + "content": "for constructing a complete 3D model. A single image cannot provide information about the unseen parts of a person's head or capture dynamic aspects like facial expressions. It is also impractical to train an end-to-end model due to the lack of multiview full-head animation datasets. To address these challenges, the system needs to rely on prior knowledge, like shape and expression." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "type": "text", + "content": "Existing single-image avatar generation methods come with many shortcomings, which we summarize in Tab. 1. The works that rely on a single image of novel identity at test time, typically utilize large-scale, unstructured 2D images [1] and videos [7, 39] during training to account for multiview and temporal aspects. Due to the practical challenges in collecting multiview full-head animation datasets, most single-image methods, including recent work [35], cannot reconstruct a complete 360-degree view of the head. While some approaches can generate full views [1, 40], they typically lack the ability to model facial expressions. Among methods that can animate, some methods rely solely on 2D representation [39], which makes these methods multi-view inconsistent by design. To overcome the above limitation, some methods rely on a hybrid representation" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 683, + 124, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 683, + 124, + 693 + ], + "spans": [ + { + "bbox": [ + 66, + 683, + 124, + 693 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 66, + 683, + 124, + 693 + ], + "type": "text", + "content": " Equal advising." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 693, + 295, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 712 + ], + "type": "text", + "content": "*No biometric data was used to train, validate, or evaluate the model described in this work." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "11612" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 72, + 101, + 541, + 193 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Comparison of methods for avatar generation and animation. The first two methods focus on general object generation, while the remaining methods are specifically designed for avatar applications." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 72, + 101, + 541, + 193 + ], + "lines": [ + { + "bbox": [ + 72, + 101, + 541, + 193 + ], + "spans": [ + { + "bbox": [ + 72, + 101, + 541, + 193 + ], + "type": "table", + "html": "
MethodsGoalDimensionInput360° viewAnimation3D ConsistencyQuality
LDM [32]Image Generation2DSingle Image-High
DreamFusion [26], SV3D [37]Image-to-3D3DSingle Image✓LowLow
LivePortrait [10], AniPortrait [39]Video Generation2DSingle Image✓LowHigh
PanoHead [1], SPI [44]3D GAN Inversion3DSingle Image✓MediumMedium
Portrait4D [7], CAP4D [35]3D Talking Head3DSingle Image✓MediumHigh
FaceCraft4D4D AvATAR Generation3DSingle Image✓✓HighHigh
HQ3D [36], GA [27]4D AvATAR Generation3DMulti-view Videos✓✓HighHigh
", + "image_path": "26e68ba3bfc2cda36fc6f0e62c01bd72b55253b0f14cc77f6b97a2c5604769ea.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 207, + 294, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 207, + 294, + 266 + ], + "spans": [ + { + "bbox": [ + 54, + 207, + 294, + 266 + ], + "type": "text", + "content": "that combines 2D and 3D [7, 34]. However, such methods cannot handle extreme camera angles, as they rely on 2D components for high-resolution synthesis. In contrast, we aim to build our model that has pure 3D representation, and, therefore, is multiview-consistent by design." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 267, + 295, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 267, + 295, + 386 + ], + "spans": [ + { + "bbox": [ + 54, + 267, + 295, + 386 + ], + "type": "text", + "content": "We propose FaceCraft4D, which takes a single image as input and creates a 3D Facial Avatar that can be animated and rendered in 360-degree view, as shown in Fig. 1. To tackle such an under-constrained problem, we make use of 3 different types of priors: a geometry prior, an image prior, and a video prior. We use these priors to synthesize personalized multiview images of the given identity, which are then used to train an explicit 3D model of an avatar that can be animated (see Fig. 2). We will refer to such a model as a 4D avatar." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 388, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 295, + 590 + ], + "type": "text", + "content": "We start with a geometry prior from pretrained 3D-GAN [1]. Since, 3D-GAN is trained with large-scale images that comprises of all the views of the head, it has information about the complete head shape. Through 3D-GAN inversion [31], we obtain personalized shape along with an approximate texture. The resulting 3D representation lacks personalized high-quality multiview consistent texture and can not be used for creating novel expressions or animations (see Tab. 1). To improve the quality of the textures, we use of a 2D image prior. To ensure multi-view consistency and preservation of identity across the views, we employ epipolar constraints and cross-view mutual attention. The resulting multi-view images can be used to build a high-quality static 3D head model, which, however, lacks control over expressions and articulations. To animate the avatar, we utilize a 2D video prior [10], which generates training data for facial animation across different viewpoints." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 592, + 295, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 592, + 295, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 295, + 674 + ], + "type": "text", + "content": "Finally, for real-time rendering performance, we train a 3D Gaussian representation [16] that is rigged to a FLAME parametric face model [20, 27] using the multi-view videos obtained in the previous steps. To further improve the training consistency across the views and to avoid blur, we propose a COnsistent-INconsistent (COIN) training, which can capture cross-view inconsistencies in an MLP." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 677, + 294, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 677, + 294, + 688 + ], + "spans": [ + { + "bbox": [ + 67, + 677, + 294, + 688 + ], + "type": "text", + "content": "In summary, the contributions of our work are as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "- We present a novel framework for generating high-quality 4D head models from a single image of a face. We com" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 321, + 207, + 553, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 207, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 321, + 207, + 553, + 229 + ], + "type": "text", + "content": "pare our framework with previous methods and summarize the differences in Tab. 1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 231, + 553, + 302 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 314, + 231, + 553, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 231, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 314, + 231, + 553, + 266 + ], + "type": "text", + "content": "- We introduce novel warping-based control generation signals to improve quality and consistency across different views and expressions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 267, + 553, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 267, + 553, + 302 + ], + "spans": [ + { + "bbox": [ + 313, + 267, + 553, + 302 + ], + "type": "text", + "content": "- We introduce COIN training, which enables a 3D Gaussian model to learn from inconsistent data while preserving high-quality, consistent features." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 316, + 400, + 328 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 316, + 400, + 328 + ], + "spans": [ + { + "bbox": [ + 314, + 316, + 400, + 328 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 337, + 555, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 337, + 555, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 555, + 443 + ], + "type": "text", + "content": "3D Head generation. Numerous studies have explored the reconstruction of 3D head models. To obtain sufficient information, most approaches [9, 27, 48, 48] utilize monocular video sequences, which provide multiple viewpoints of the subject's head. However, a significant limitation of these methods is their reliance on multiple viewpoints and lengthy video inputs. They also struggle to accurately reproduce less prominent features in videos, such as the back of the head." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 446, + 554, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 446, + 554, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 446, + 554, + 564 + ], + "type": "text", + "content": "Another approach is to train a 3D generator on a large head image dataset. EG3D [3] proposed a hybrid tri-plane representation that can quickly and efficiently render high-quality facial images. Panohead [1] and Portrait3D [40] expanded on EG3D by increasing the number of planes, enabling the generation of 3D heads from limited angles to full 360-degree views. However, most methods in this category rely on hybrid representation having a 2D super-resolution network. This impacts their generalization to extreme camera poses and suffer identity inconsistency across views." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 566, + 554, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 566, + 554, + 697 + ], + "spans": [ + { + "bbox": [ + 313, + 566, + 554, + 697 + ], + "type": "text", + "content": "Recently, the rise of text-to-image models has spurred the development of novel approaches for text-to-3D generation tasks. One such approach is Score Distillation Sampling (SDS) [26, 29], which leverages the score function to distill diffusion priors into 3D representations, enabling the generation of 3D content from text or image inputs. Headsculpt [11, 49] incorporates facial keypoint information into the text-to-3D generation method, effectively constraining the facial layout. However, the texture generated by these methods can be oversaturated, making it challenging to achieve photo-realistic head models." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "type": "text", + "content": "Head animation. Head animation techniques can be" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11613" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 549, + 214 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 549, + 214 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 549, + 214 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 549, + 214 + ], + "type": "image", + "image_path": "fcc560db6db8a53d342c1d8369e5acf4d14a52206f7b43ce07b7c1995d2cb790.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 221, + 555, + 277 + ], + "lines": [ + { + "bbox": [ + 54, + 221, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 54, + 221, + 555, + 277 + ], + "type": "text", + "content": "Figure 2. Overview of FaceCraft4D: Our approach begins with estimating the geometry of a single input image using a Geometry Prior. This geometry guides the synthesis of personalized multiview images, providing " + }, + { + "bbox": [ + 54, + 221, + 555, + 277 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 54, + 221, + 555, + 277 + ], + "type": "text", + "content": " views and varied expressions, with support from both 2D image and video priors. Since the synthesized data often exhibit inconsistency across views, we propose COIN optimization for robust 4D optimization. By fitting the model to the multiview data, we achieve our final 4D avatar. All faces in this manuscript come from the public FFHQ dataset [15]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 288, + 294, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 288, + 294, + 395 + ], + "spans": [ + { + "bbox": [ + 54, + 288, + 294, + 395 + ], + "type": "text", + "content": "broadly categorized into 2D-based and 3D-aware approaches, distinguished by their utilization of explicit camera models for rendering. Recent methods [8, 13, 30, 43] have predominantly adopted warping-based strategies, applying deformation fields to intermediate appearance features to capture facial motion characteristics. While effective for frontal and moderate poses, these approaches struggle to maintain geometric consistency during large pose variations due to their inherent lack of 3D modeling." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 396, + 295, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 295, + 575 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 295, + 575 + ], + "type": "text", + "content": "To enable free-viewpoint rendering with stronger 3D consistency guarantees, researchers integrated explicit 3D representations into the head synthesis pipeline. Early approaches [17, 41] relied on 3D morphable models for representing head geometry and texture. Subsequent works [5, 14, 21, 42, 45] leveraged more sophisticated representations such as Neural Radiance Fields (NeRFs) [25] to better capture complex structures like hair and accessories. Recent advances like GAGAvatar [4, 22], which employs 3D Gaussian splatting [16], have demonstrated impressive performance in both rendering quality and speed. However, these methods are typically limited in their ability to generate full 360-degree views. Our approach enables true 360-degree view synthesis while maintaining real-time rendering capabilities and achieving superior animation quality." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 586, + 111, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 111, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 111, + 597 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "text", + "content": "In this work, we propose a novel framework for generating high-fidelity 4D avatars from a single portrait image " + }, + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "inline_equation", + "content": "I_{R}" + }, + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "text", + "content": ". Reconstructing a 3D dynamic model from a single 2D image is an inherently ill-posed problem, as critical 3D and temporal information is missing. To address this challenge, we leverage several 2D and 3D priors to compensate for the missing data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "The overall pipeline, shown in Fig. 2, consists of two stages: personalized multiview generation and 4D repre" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 288, + 555, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 555, + 420 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 555, + 420 + ], + "type": "text", + "content": "sensation optimization. In the Personalized Multiview Generation stage, we utilize geometry priors to estimate the 3D geometry (Sec. 3.1), use image priors to enhance the texture (Sec. 3.2), and video priors to synthesize high-quality, consistent multi-view videos with dynamic expressions (Sec. 3.3). We then leverage this synthetically-generated personalized image set to optimize a robust 4D representation model (Sec. 3.4). The trained 4D model can be animated by the control parameters of a morphable model, enabling the generation of high-fidelity 4D avatars from a single portrait image." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 458, + 411, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 458, + 411, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 458, + 411, + 471 + ], + "type": "text", + "content": "3.1. Geometry Prior" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": "The initialization is crucial for 3D generation. A good initialization should follow the approximate geometry of the head shown in the input image, " + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "inline_equation", + "content": "I_{R}" + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": ". For that, we use the prior provided by a 3D-aware generator, PanoHead [1, 40], which was trained on a large " + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": " head dataset. We obtain a coarse 3D reconstruction for a given input image " + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "inline_equation", + "content": "I_{R}" + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": " through GAN inversion. Specifically, we first optimize the latent code " + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": " within the generator's domain by minimizing a combination of pixel-wise " + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2" + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": " loss and image-level LPIPS loss [47]. Subsequently, following [31], we keep the optimized latent code " + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": " fixed while fine-tuning the 3D generator's parameters. During this fine-tuning stage, we maintain the same loss functions used in the latent optimization phase to ensure the rendered views closely match the input image. This process gives us an initial 3D head representation, though one that cannot be animated. The optimized generator is then used to synthesize a set of approximate multiview images " + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\{I_i\\}" + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": " and their corresponding depth maps " + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\{D_i\\}" + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": " is the view index." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11614" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 71, + 303, + 286 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 303, + 286 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 303, + 286 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 303, + 286 + ], + "type": "image", + "image_path": "2bf14056c35e2f198145d4363f0ed09d58925f35166af031f7950102ec378c51.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 342, + 304, + 491 + ], + "blocks": [ + { + "bbox": [ + 55, + 294, + 296, + 340 + ], + "lines": [ + { + "bbox": [ + 55, + 294, + 296, + 340 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 296, + 340 + ], + "type": "text", + "content": "Figure 3. Image Prior: We introduce a cross-view mutual attention mechanism and epipolar constraints to enhance consistency in generated novel views. Our approach aligns reference and target images, maintaining visual coherence across viewpoints." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 342, + 304, + 491 + ], + "lines": [ + { + "bbox": [ + 58, + 342, + 304, + 491 + ], + "spans": [ + { + "bbox": [ + 58, + 342, + 304, + 491 + ], + "type": "image", + "image_path": "0c3aebe297ab434d9984668cebe207fbe7b271a8818fb8f306b617d5cce678d0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 505, + 297, + 550 + ], + "lines": [ + { + "bbox": [ + 55, + 505, + 297, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 505, + 297, + 550 + ], + "type": "text", + "content": "Figure 4. Triplane-based methods (e.g. PanoHead) are sensitive to focal length (scale) and exhibit significant degradation of quality as focal length decreases. In contrast, our Gaussian-based methods maintain robust performance across varying focal lengths." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 562, + 135, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 562, + 135, + 575 + ], + "spans": [ + { + "bbox": [ + 55, + 562, + 135, + 575 + ], + "type": "text", + "content": "3.2. Image Prior" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "content": "The textures synthesized from the geometry prior may lack visual quality, and the identity of the individual may not be well preserved across the views. In particular, when rendering with different focal lengths, triplane-based methods [1] exhibit substantial degradation in image quality, as illustrated in Fig. 4. We enhance the generated multiview images using 2D diffusion models [32], which have demonstrated powerful generative capabilities. Although direct image-to-image translation [24] appears to be a straightforward solution, this approach poses significant risks: the de-noising process may introduce cross-view inconsistencies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 120 + ], + "type": "text", + "content": "and also modify the original semantic content. To address these risks, we propose two strategies that constrain the generation process, which we describe next and illustrate in Fig. 3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 123, + 556, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 123, + 556, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 123, + 556, + 350 + ], + "type": "text", + "content": "Cross-view mutual attention. Inspired by MasaCtrl [2], we introduce a cross-view attention mechanism to replace the standard self-attention during the generation process. This modification allows us to inject information from the reference image into the novel view. Specifically, we first perform DDIM inversion [33] to introduce noise to both the reference frame " + }, + { + "bbox": [ + 313, + 123, + 556, + 350 + ], + "type": "inline_equation", + "content": "I_{R}" + }, + { + "bbox": [ + 313, + 123, + 556, + 350 + ], + "type": "text", + "content": " and the coarse image " + }, + { + "bbox": [ + 313, + 123, + 556, + 350 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 313, + 123, + 556, + 350 + ], + "type": "text", + "content": " of the target novel view. This step preserves the overall face structure and contour. Then, the diffusion model is used to gradually denoise two images to enhance the texture. In this process, we replace the key " + }, + { + "bbox": [ + 313, + 123, + 556, + 350 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 123, + 556, + 350 + ], + "type": "text", + "content": " and value " + }, + { + "bbox": [ + 313, + 123, + 556, + 350 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 123, + 556, + 350 + ], + "type": "text", + "content": " matrices (calculated in self-attention) of the novel view with those derived from the reference image. Different from [2], we adopt a batch-processing strategy in which multiple images " + }, + { + "bbox": [ + 313, + 123, + 556, + 350 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 313, + 123, + 556, + 350 + ], + "type": "text", + "content": " representing different viewpoints are treated as a single batch, with the reference image " + }, + { + "bbox": [ + 313, + 123, + 556, + 350 + ], + "type": "inline_equation", + "content": "I_{R}" + }, + { + "bbox": [ + 313, + 123, + 556, + 350 + ], + "type": "text", + "content": " serving as the consistent source for all subsequent calculations. This not only streamlines the generation process but also ensures uniformity in feature injection across novel viewpoints." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 351, + 556, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 351, + 556, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 351, + 556, + 434 + ], + "type": "text", + "content": "The rationale behind this approach lies in the observation that, despite variations in viewpoint, the underlying textures (represented by the attention values) should maintain consistency across the views. The cross-view attention mechanism enables the transfer of information from the reference to the novel views, enhancing both the fidelity and coherence of the generated novel views." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 437, + 556, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 556, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 556, + 689 + ], + "type": "text", + "content": "Warping-based control generation. The consistency across the views is further enforced by epipolar constraints, which enforce geometric relationships between novel views and the input view. We first use cross-view mutual attention to generate the view of the back of the head. The generated back view along with the reference image are set as anchors. Then, we propagate the texture information to novel views through depth-based warping. Specifically, we lift the anchor image into 3D space and then project it to the adjacent view with depth guidance " + }, + { + "bbox": [ + 313, + 437, + 556, + 689 + ], + "type": "inline_equation", + "content": "\\{D_i\\}" + }, + { + "bbox": [ + 313, + 437, + 556, + 689 + ], + "type": "text", + "content": ". During the projection, we form a mask to indicate the visibility of pixels to filter invisible areas using the rendered depth values. The projected image is then used to guide the texture enhancement process. The intermediate latent of the enhanced image in the diffusion model is blended with the latent of the projected image via a simple mask copy-pasting operation. By combining the latent, the resulting images benefit from both the improved consistency provided by the epipolar constraints and the enhanced quality achieved through the denoising process. We repeat this process for various azimuth angles, creating a comprehensive set of viewpoints." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "With the aforementioned guidance and epipolar constraints, we can ensure coherence across the entire view" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11615" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 303, + 201 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 303, + 201 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 303, + 201 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 303, + 201 + ], + "type": "image", + "image_path": "377bb3b1d7a5824f47a2fdc08872b959557f55bd95fa37e26c74b0c9c011193d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 209, + 296, + 266 + ], + "lines": [ + { + "bbox": [ + 55, + 209, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 296, + 266 + ], + "type": "text", + "content": "Figure 5. COIN-Training: To address geometric and color inconsistencies between the views, we train two 3D representations: a view-consistent base model (GaussianAvatar) and an MLP, encoding inconsistencies between the views. The MLP with inconsistencies lets us robustly reconstruct the high-quality 3D base model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 288, + 295, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 295, + 313 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 295, + 313 + ], + "type": "text", + "content": "space and generate high-fidelity view-consistent images " + }, + { + "bbox": [ + 55, + 288, + 295, + 313 + ], + "type": "inline_equation", + "content": "\\{I_i^*\\}" + }, + { + "bbox": [ + 55, + 288, + 295, + 313 + ], + "type": "text", + "content": " for further reconstruction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 320, + 133, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 320, + 133, + 333 + ], + "spans": [ + { + "bbox": [ + 55, + 320, + 133, + 333 + ], + "type": "text", + "content": "3.3. Video Prior" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 339, + 296, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 339, + 296, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 339, + 296, + 639 + ], + "type": "text", + "content": "The view-consistent data created in the previous step has a static expression similar to that of reference image " + }, + { + "bbox": [ + 55, + 339, + 296, + 639 + ], + "type": "inline_equation", + "content": "I_R" + }, + { + "bbox": [ + 55, + 339, + 296, + 639 + ], + "type": "text", + "content": ". This may not be enough to create diverse expressions as the images do not contain some of the expression-dependent information. For example, if the reference image has the mouth closed, the synthesized dataset would not have details inside the mouth region. To incorporate expression-dependent textures and geometry that may not be available in the multiview image generation stage (Sec. 3.2), we need to create multi-view facial expression data. We use LivePortrait [10] as a video prior, which can help generate multiview expressive videos. LivePortrait uses a single source image and a driving video to generate new videos where the source identity mimics the expressions and poses from the target video. Instead of directly copying the driving video's motion, one has the option of capturing the changes in relative expression but maintain the pose of the input image. Given that we have the reference and multiview-consistent enhanced images " + }, + { + "bbox": [ + 55, + 339, + 296, + 639 + ], + "type": "inline_equation", + "content": "\\{I_i^*\\}" + }, + { + "bbox": [ + 55, + 339, + 296, + 639 + ], + "type": "text", + "content": " from the previous step, we provide these to LivePortrait as the source images together with driving videos with diverse poses and expressions (details in Sec. 4), and obtain target videos. Because the identity is consistent across the input views, the resulting videos also maintain the identity. Using identical driving videos ensures synchronized expressions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 647, + 286, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 286, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 286, + 661 + ], + "type": "text", + "content": "3.4. Robust 4D optimization with COIN training" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "The synthesized multi-view videos have high-quality textures and good cross-view consistency, but small misalignment of features and colors is unavoidable. Learning 4D representation directly on such data leads to blurry textures" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "text", + "content": "for the rendered objects. Here, we address this problem by proposing a robust training paradigm. At a high level, we learn a base representation that is common to all views, and a delta representation that is specific to each view during training. This helps in cornering inconsistencies concerning each view to separate representation. The base model will have a coarse but consistent representation learned, while the high-frequency details end up in the delta-space representation. At test time, we choose one of the views as our reference view and use its delta representation as the common high-frequency details for all the views." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 204, + 555, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 555, + 407 + ], + "type": "text", + "content": "In specific, we achieve this by jointly training two 4D representations: a consistent GaussianAvatar [27], which represents the 3D structure using Gaussians bound to a FLAME-tracked [20] mesh, and a NeRF-style MLP, tasked with encoding inconsistencies between the views. As shown in Fig. 5, we design the losses that prioritize structure reconstruction in the consistent GaussianAvatar (LPIPS), and high-frequency detail reconstruction in the MLP with the view-dependent inconsistencies (L1 and SSIM). By isolating view-dependent variations in a separate model, this strategy prevents these inconsistencies from corrupting the base representation. The approach can be considered a form of robust regression, with the key difference that the view-dependent MLP ensures that inconsistencies (outliers) are spatially localized. This cannot be achieved with robust norms (e.g., L1), which do not maintain the information about the spatial position of outliers." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 407, + 554, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 554, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 554, + 548 + ], + "type": "text", + "content": "We choose GaussianAvatar as our consistent representation because it lets us animate the avatar with controllable FLAME expression and pose parameters, providing a solid foundation for modeling consistent geometric and appearance features. The details of FLAME tracking and controlling are explained in Sec. 1.3 of the supplementary. To model the view-dependent variations, we design a lightweight MLP-based architecture that introduces minimal computational overhead while effectively capturing view-specific details. The network encodes the offset in the color of each Gaussian (see Fig. 5), which can be formulated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 375, + 552, + 553, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 552, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 375, + 552, + 553, + 564 + ], + "type": "interline_equation", + "content": "c _ {\\text {o f f s e t}} = \\operatorname {M L P} \\left(e _ {\\text {v i e w}}, c, e _ {g}; \\theta\\right), \\tag {1}", + "image_path": "22b893639cec259329f202de4f8544f7147fd4eee7df85e4fcaaefb6bb7e354d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 569, + 554, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 554, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 554, + 700 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 569, + 554, + 700 + ], + "type": "inline_equation", + "content": "e_{\\mathrm{view}} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 569, + 554, + 700 + ], + "type": "text", + "content": " is a learnable view-specific embedding that encodes view-representation, " + }, + { + "bbox": [ + 313, + 569, + 554, + 700 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 569, + 554, + 700 + ], + "type": "text", + "content": " represents the initial Gaussian color attributes from the GaussianAvatar, " + }, + { + "bbox": [ + 313, + 569, + 554, + 700 + ], + "type": "inline_equation", + "content": "e_g" + }, + { + "bbox": [ + 313, + 569, + 554, + 700 + ], + "type": "text", + "content": " is a position embedding for each Gaussian, accounting for location-specific color variation ranges (e.g., eyes vs. mouth regions), and " + }, + { + "bbox": [ + 313, + 569, + 554, + 700 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 569, + 554, + 700 + ], + "type": "text", + "content": " is a vector of the trainable parameters. The final view-dependent color is computed as the sum of the base GaussianAvatar color and the offset: " + }, + { + "bbox": [ + 313, + 569, + 554, + 700 + ], + "type": "inline_equation", + "content": "c^* = c + c_{\\mathrm{offset}}" + }, + { + "bbox": [ + 313, + 569, + 554, + 700 + ], + "type": "text", + "content": ". The MLP consists of only two layers to maintain computational efficiency and preserve the real-time rendering capabilities of the Gaussian-based representation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 325, + 700, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 700, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 700, + 553, + 714 + ], + "type": "text", + "content": "During training, we render two types of images: " + }, + { + "bbox": [ + 325, + 700, + 553, + 714 + ], + "type": "inline_equation", + "content": "I_{i}^{\\mathrm{C}}" + }, + { + "bbox": [ + 325, + 700, + 553, + 714 + ], + "type": "text", + "content": " from" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11616" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 68, + 545, + 355 + ], + "blocks": [ + { + "bbox": [ + 61, + 68, + 545, + 355 + ], + "lines": [ + { + "bbox": [ + 61, + 68, + 545, + 355 + ], + "spans": [ + { + "bbox": [ + 61, + 68, + 545, + 355 + ], + "type": "image", + "image_path": "7cce182327f9a8e2e3eb4fa9477028bf878112b3a7cb62d60a165e40a2ded385.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 361, + 456, + 374 + ], + "lines": [ + { + "bbox": [ + 153, + 361, + 456, + 374 + ], + "spans": [ + { + "bbox": [ + 153, + 361, + 456, + 374 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison on static 3D head generation from a single image." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 389, + 296, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 296, + 437 + ], + "type": "text", + "content": "the consistent component alone, and " + }, + { + "bbox": [ + 55, + 389, + 296, + 437 + ], + "type": "inline_equation", + "content": "I_{i}^{\\mathrm{IC}}" + }, + { + "bbox": [ + 55, + 389, + 296, + 437 + ], + "type": "text", + "content": " from the combination of both components. To ensure accurate reconstruction of high-frequency details, we supervise " + }, + { + "bbox": [ + 55, + 389, + 296, + 437 + ], + "type": "inline_equation", + "content": "I_{i}^{\\mathrm{IC}}" + }, + { + "bbox": [ + 55, + 389, + 296, + 437 + ], + "type": "text", + "content": " using pixel-wise losses:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 82, + 446, + 295, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 446, + 295, + 462 + ], + "spans": [ + { + "bbox": [ + 82, + 446, + 295, + 462 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p i x e l}} = \\lambda_ {1} \\mathcal {L} 1 \\left(I _ {i} ^ {\\mathrm {I C}}, I _ {i} ^ {*}\\right) + \\lambda_ {\\text {S S I M}} \\text {S S I M} \\left(I _ {i} ^ {\\mathrm {I C}}, I _ {i} ^ {*}\\right) \\tag {2}", + "image_path": "d3bc377b7a2c4d4a03c8223cbafb0ffee91ca135a3c265f35a3a5bb1d4307eb3.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 470, + 296, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 470, + 296, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 470, + 296, + 518 + ], + "type": "text", + "content": "To prevent the inconsistent component from dominating the representation and maintain the effectiveness of the consistent component, we introduce a structural supervision loss and a regularization term:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 530, + 295, + 545 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 530, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 115, + 530, + 295, + 545 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s t r u c}} = \\lambda_ {\\text {L P I P S}} \\operatorname {L P I P S} \\left(I _ {i} ^ {\\mathrm {C}}, I _ {i} ^ {*}\\right) \\tag {3}", + "image_path": "e93e62ac2f6a598fc0ad1c38683f91e3f716d8d24a5d8710432ea55b668def30.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 122, + 563, + 295, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 563, + 295, + 576 + ], + "spans": [ + { + "bbox": [ + 122, + 563, + 295, + 576 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e g}} = \\lambda_ {\\text {o f f s e t}} \\mathcal {L} 1 \\left(c _ {\\text {o f f s e t}}, 0\\right), \\tag {4}", + "image_path": "5cc7ecadf679aa0af1b86642352785856d9e1422d96c9edb349df5a61d7ee13d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 581, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 295, + 641 + ], + "type": "text", + "content": "where LPIPS loss ensures image-level supervision for the consistent component, and the offset regularization encourages minimal view-dependent modifications. The hyperparameters are set to " + }, + { + "bbox": [ + 55, + 581, + 295, + 641 + ], + "type": "inline_equation", + "content": "\\lambda_{1} = 0.8" + }, + { + "bbox": [ + 55, + 581, + 295, + 641 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 581, + 295, + 641 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{SSIM}} = 0.2" + }, + { + "bbox": [ + 55, + 581, + 295, + 641 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 581, + 295, + 641 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{LPIPS}} = 0.05" + }, + { + "bbox": [ + 55, + 581, + 295, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 581, + 295, + 641 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{offset}} = 1" + }, + { + "bbox": [ + 55, + 581, + 295, + 641 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "During inference, we input a single fixed view-embedding for the inconsistent component. Given that our reference image " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "I_R" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " contains superior quality and identity information compared to the generated multiview images " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{I_i^*\\}" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": ", we select its corresponding view embedding for inference. Novel expressions and poses can be easily gen" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 389, + 555, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 555, + 414 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 555, + 414 + ], + "type": "text", + "content": "erated by adjusting the FLAME parameters, leveraging the animation capabilities of our backbone model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 426, + 395, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 426, + 395, + 440 + ], + "spans": [ + { + "bbox": [ + 314, + 426, + 395, + 440 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 447, + 555, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 555, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 555, + 626 + ], + "type": "text", + "content": "Implementation details. During the multiview generation stage, we use PanoHead [1] as our geometry prior, Cosmicman [19] as the image prior, and LivePortrait [10] as the video prior model. We render 24 views along the horizon orbit for static reconstruction. For animation, we select " + }, + { + "bbox": [ + 313, + 447, + 555, + 626 + ], + "type": "inline_equation", + "content": "I_R" + }, + { + "bbox": [ + 313, + 447, + 555, + 626 + ], + "type": "text", + "content": " and 11 frontal views from these generated images as source images and randomly sample 8 video clips from the NerSemble [18] dataset as driving videos. The source images and driving videos are then input to the video prior model to produce multiview video data. In the 4D optimization stage, we first optimize the GaussianAvatar [27] model with the static data over 30K iterations. Next, we apply COIN-optimization techniques, finetuning the static Gaussian model over 90K iterations. The complete process takes approximately two hours to generate a single 4D avatar." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "Evaluation metrics. In the absence of ground truth data, we use commonly adopted non-reference metrics to evaluate the rendered images in terms of image quality and identity similarity. For overall image quality, we use the FID [12] score. To assess multi-view consistency, we employ the CLIP-I [28] score. For identity preservation, we calculate ID scores by measuring the cosine distance of face" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11617" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 68, + 187, + 220 + ], + "blocks": [ + { + "bbox": [ + 72, + 68, + 187, + 220 + ], + "lines": [ + { + "bbox": [ + 72, + 68, + 187, + 220 + ], + "spans": [ + { + "bbox": [ + 72, + 68, + 187, + 220 + ], + "type": "image", + "image_path": "5e1438b7557703f1f8df2d2be930093745f1411125ca5f79cae0b730defa889f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 185, + 221, + 424, + 233 + ], + "lines": [ + { + "bbox": [ + 185, + 221, + 424, + 233 + ], + "spans": [ + { + "bbox": [ + 185, + 221, + 424, + 233 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison on animation of different views." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 189, + 68, + 246, + 220 + ], + "blocks": [ + { + "bbox": [ + 189, + 68, + 246, + 220 + ], + "lines": [ + { + "bbox": [ + 189, + 68, + 246, + 220 + ], + "spans": [ + { + "bbox": [ + 189, + 68, + 246, + 220 + ], + "type": "image", + "image_path": "6d13b9e1d13f030a9266c8ef6892f310f5e90c5cc05abeafb50e4cf4161d9c99.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 253, + 68, + 362, + 220 + ], + "blocks": [ + { + "bbox": [ + 253, + 68, + 362, + 220 + ], + "lines": [ + { + "bbox": [ + 253, + 68, + 362, + 220 + ], + "spans": [ + { + "bbox": [ + 253, + 68, + 362, + 220 + ], + "type": "image", + "image_path": "b109ea62d5b353691c4d537a77651f346c66ee6e069629cfb0a78faf14676417.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 364, + 68, + 491, + 220 + ], + "blocks": [ + { + "bbox": [ + 364, + 68, + 491, + 220 + ], + "lines": [ + { + "bbox": [ + 364, + 68, + 491, + 220 + ], + "spans": [ + { + "bbox": [ + 364, + 68, + 491, + 220 + ], + "type": "image", + "image_path": "7a7b66431ef2e601978661e80ed7d1132b11c58ba801eff2abdb3ae6e7a9238e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 492, + 68, + 552, + 220 + ], + "blocks": [ + { + "bbox": [ + 492, + 68, + 552, + 220 + ], + "lines": [ + { + "bbox": [ + 492, + 68, + 552, + 220 + ], + "spans": [ + { + "bbox": [ + 492, + 68, + 552, + 220 + ], + "type": "image", + "image_path": "4853a7dc0f8bc84c4ab8f30cdd89f286eada173252fea01c25f6ca351732ca09.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 59, + 258, + 292, + 403 + ], + "blocks": [ + { + "bbox": [ + 59, + 247, + 290, + 258 + ], + "lines": [ + { + "bbox": [ + 59, + 247, + 290, + 258 + ], + "spans": [ + { + "bbox": [ + 59, + 247, + 290, + 258 + ], + "type": "text", + "content": "Table 2. Quantitative comparison on static 3D head generation." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 258, + 292, + 403 + ], + "lines": [ + { + "bbox": [ + 59, + 258, + 292, + 403 + ], + "spans": [ + { + "bbox": [ + 59, + 258, + 292, + 403 + ], + "type": "table", + "html": "
MethodCLIP-I ↑ID ↑FID ↓
GaussianCube [46]0.68300.4300258.81
PanoHead [1]0.82330.4246195.28
SV3D [37]0.76560.4331234.86
Portrait3D [40]0.70660.3719302.74
Ours0.80530.5082174.36
Table 3. Quantitative comparison on 3D head animation.
MethodCLIP-I ↑ID ↑FID ↓
AntiPortrait [39]0.46530.4171364.99
Portrait-4D-V2 [7]0.52360.4592248.36
Ours0.57370.4602201.76
", + "image_path": "8363dcdef69382e3710e1387c21569021842e0eb4fe0994d09be6e2d62d2d11a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 412, + 295, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 412, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 295, + 437 + ], + "type": "text", + "content": "recognition features [6] between the novel views and the reference images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 439, + 296, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 439, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 55, + 439, + 296, + 487 + ], + "type": "text", + "content": "Datasets. We conducted our analysis on a subset of facial images from FFHQ [15] dataset, a high-quality, in-the-wild facial dataset. For evaluation, we select 100 images with unoccluded faces to evaluate our model and baselines." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 497, + 242, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 242, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 242, + 510 + ], + "type": "text", + "content": "4.1. Static 3D Generation Comparisons" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 518, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 295, + 579 + ], + "type": "text", + "content": "Baselines. We compare our method with existing image-to-3D approaches, including a general object-focused model, SV3D [37], and three head-specific methods: the inference-based GaussianCube [46], the 3D-GAN-based PanoHead [1], and the optimization-based Portrait3D [40]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": "Qualitative results. Fig. 6 shows visual comparisons where our method achieves the highest visual fidelity and identity coherence compared to other approaches. Notably, it preserves details such as teeth and earrings in side views. In contrast, SV3D, which focuses on general objects, struggles to capture accurate head geometry, often degenerating into a flat plane. GaussianCube tends to \"toonify\" the input head due to the nature of its training data. While PanoHead achieves good visual quality, it suffers from detail loss and floating artifacts during pose rotation. Portrait3D also falls short in quality, as the SDS [26] optimization introduces" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 249, + 474, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 249, + 474, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 249, + 474, + 262 + ], + "type": "text", + "content": "variations that lead to blurring artifacts." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 265, + 555, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 265, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 555, + 397 + ], + "type": "text", + "content": "Quantitative results. For each generated head model, we render images from five viewpoints: left-frontal, left, right-frontal, right, and frontal. All five rendered images, along with the reference image, are used to compute the CLIP-I, ID, and FID metrics. As shown in Tab. 2, our method outperforms others in both ID and FID scores. This aligns with the observed visual fidelity, as our results appear more realistic and coherent. Our method achieves higher identity similarity and better detail preservation, even in side views. The results for the MEAD multiview dataset [38] can be found in the supplementary material." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 408, + 452, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 452, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 452, + 421 + ], + "type": "text", + "content": "4.2. Animation Comparisons" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 430, + 555, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 555, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 555, + 526 + ], + "type": "text", + "content": "Baselines. We compare our method with one-shot video-based head reenactment approaches, including the 2D-based AniPortrait [39] and the 3D-based Portrait-4D-V2 [7]. For AniPortrait, which uses keypoints as intermediate driving signals, we modify the Euler angles of the expression image to rotate the keypoints and generate novel view videos. In the case of Portrait-4D-V2, novel views can be generated directly by editing the camera parameters." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 529, + 556, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 556, + 638 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 556, + 638 + ], + "type": "text", + "content": "Qualitative results. As shown in Fig. 7, our method preserves expressions with greater accuracy. Portrait-4D-V2 lacks a back-of-head model, so when rotated to side views, the back appears empty. AniPortrait, as a 2D-based method, can achieve high texture quality. However, with large-angle rotations, AniPortrait struggles to preserve identity details and introduces noticeable artifacts in the hair and background. In contrast, our method maintains accurate geometry and expression, even with camera movement." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": "Quantitative results. We test our data on unseen views (especially " + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "inline_equation", + "content": "60^{\\circ} - 180^{\\circ}" + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": "), to highlight each method's performance under varying viewing angles, as handling extreme rotations is often challenging. The results are presented in Tab. 3. Our method outperforms all baselines, demonstrating robustness across varied views. The performance" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11618" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 105, + 131 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 105, + 131 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 105, + 131 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 105, + 131 + ], + "type": "image", + "image_path": "313be90bd12dd852f03f64c64227a372a430ba2ab777a5d8e601c53b8ba97d51.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 132, + 91, + 139 + ], + "lines": [ + { + "bbox": [ + 74, + 132, + 91, + 139 + ], + "spans": [ + { + "bbox": [ + 74, + 132, + 91, + 139 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 115, + 72, + 157, + 129 + ], + "blocks": [ + { + "bbox": [ + 115, + 72, + 157, + 129 + ], + "lines": [ + { + "bbox": [ + 115, + 72, + 157, + 129 + ], + "spans": [ + { + "bbox": [ + 115, + 72, + 157, + 129 + ], + "type": "image", + "image_path": "c122fd351ad377d65f478f329105424842d4f988e2d4e2d97c02254a613d3541.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 114, + 131, + 160, + 140 + ], + "lines": [ + { + "bbox": [ + 114, + 131, + 160, + 140 + ], + "spans": [ + { + "bbox": [ + 114, + 131, + 160, + 140 + ], + "type": "text", + "content": "Warped Image" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 162, + 72, + 205, + 129 + ], + "blocks": [ + { + "bbox": [ + 162, + 72, + 205, + 129 + ], + "lines": [ + { + "bbox": [ + 162, + 72, + 205, + 129 + ], + "spans": [ + { + "bbox": [ + 162, + 72, + 205, + 129 + ], + "type": "image", + "image_path": "bcd261144454207b004104adb87ef4c5910f455d7c766f99d3fb9c32de343a50.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 163, + 131, + 205, + 140 + ], + "lines": [ + { + "bbox": [ + 163, + 131, + 205, + 140 + ], + "spans": [ + { + "bbox": [ + 163, + 131, + 205, + 140 + ], + "type": "text", + "content": "w/o Warping" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 212, + 72, + 255, + 129 + ], + "blocks": [ + { + "bbox": [ + 212, + 72, + 255, + 129 + ], + "lines": [ + { + "bbox": [ + 212, + 72, + 255, + 129 + ], + "spans": [ + { + "bbox": [ + 212, + 72, + 255, + 129 + ], + "type": "image", + "image_path": "17e8e167a194f0873557dd97b390d9bf23e21edce6eb5566aaa09439d85bdcc0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 217, + 131, + 244, + 139 + ], + "lines": [ + { + "bbox": [ + 217, + 131, + 244, + 139 + ], + "spans": [ + { + "bbox": [ + 217, + 131, + 244, + 139 + ], + "type": "text", + "content": "w/o MA" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 261, + 72, + 304, + 129 + ], + "blocks": [ + { + "bbox": [ + 261, + 72, + 304, + 129 + ], + "lines": [ + { + "bbox": [ + 261, + 72, + 304, + 129 + ], + "spans": [ + { + "bbox": [ + 261, + 72, + 304, + 129 + ], + "type": "image", + "image_path": "88c00eaea3dac1381f6c9e0670bcac73027c4fa382bf7416aeb64afbb342edaf.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 271, + 131, + 285, + 139 + ], + "lines": [ + { + "bbox": [ + 271, + 131, + 285, + 139 + ], + "spans": [ + { + "bbox": [ + 271, + 131, + 285, + 139 + ], + "type": "text", + "content": "Full" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 63, + 154, + 113, + 220 + ], + "blocks": [ + { + "bbox": [ + 67, + 142, + 283, + 152 + ], + "lines": [ + { + "bbox": [ + 67, + 142, + 283, + 152 + ], + "spans": [ + { + "bbox": [ + 67, + 142, + 283, + 152 + ], + "type": "text", + "content": "Figure 8. Ablation of multiview image generation modules." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 63, + 154, + 113, + 220 + ], + "lines": [ + { + "bbox": [ + 63, + 154, + 113, + 220 + ], + "spans": [ + { + "bbox": [ + 63, + 154, + 113, + 220 + ], + "type": "image", + "image_path": "ed7841e80558856f152deb92b73f1cbaf90b5d5513955e9f8a345126ad1d2b4a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 126, + 154, + 182, + 216 + ], + "blocks": [ + { + "bbox": [ + 126, + 154, + 182, + 216 + ], + "lines": [ + { + "bbox": [ + 126, + 154, + 182, + 216 + ], + "spans": [ + { + "bbox": [ + 126, + 154, + 182, + 216 + ], + "type": "image", + "image_path": "e50304eb204c57a4f80602f2869503c39f1193f82eb80be5a54a72704d394212.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 120, + 216, + 181, + 243 + ], + "blocks": [ + { + "bbox": [ + 120, + 216, + 181, + 243 + ], + "lines": [ + { + "bbox": [ + 120, + 216, + 181, + 243 + ], + "spans": [ + { + "bbox": [ + 120, + 216, + 181, + 243 + ], + "type": "image", + "image_path": "98c1170d45a8510af70f4e188cbff88bcb97da8c554d8f0d8a23f4babf0f98a4.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 126, + 243, + 176, + 251 + ], + "lines": [ + { + "bbox": [ + 126, + 243, + 176, + 251 + ], + "spans": [ + { + "bbox": [ + 126, + 243, + 176, + 251 + ], + "type": "text", + "content": "w/o Video Prior" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 188, + 154, + 239, + 217 + ], + "blocks": [ + { + "bbox": [ + 188, + 154, + 239, + 217 + ], + "lines": [ + { + "bbox": [ + 188, + 154, + 239, + 217 + ], + "spans": [ + { + "bbox": [ + 188, + 154, + 239, + 217 + ], + "type": "image", + "image_path": "7aabae2050151d565e62a780db6775c03367b5eec66715a7e74586ade5be315d.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 182, + 216, + 243, + 243 + ], + "blocks": [ + { + "bbox": [ + 182, + 216, + 243, + 243 + ], + "lines": [ + { + "bbox": [ + 182, + 216, + 243, + 243 + ], + "spans": [ + { + "bbox": [ + 182, + 216, + 243, + 243 + ], + "type": "image", + "image_path": "b3b2915c9d660ab5e193b688260a8e4ca015760adf79932434949553c1592ef9.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 197, + 243, + 230, + 251 + ], + "lines": [ + { + "bbox": [ + 197, + 243, + 230, + 251 + ], + "spans": [ + { + "bbox": [ + 197, + 243, + 230, + 251 + ], + "type": "text", + "content": "w/o COIN" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 249, + 154, + 304, + 217 + ], + "blocks": [ + { + "bbox": [ + 249, + 154, + 304, + 217 + ], + "lines": [ + { + "bbox": [ + 249, + 154, + 304, + 217 + ], + "spans": [ + { + "bbox": [ + 249, + 154, + 304, + 217 + ], + "type": "image", + "image_path": "dc449449143426ae99c281625dd1028f06f3ce8b31f53ea2c69d790baec6d52f.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 245, + 216, + 304, + 243 + ], + "blocks": [ + { + "bbox": [ + 245, + 216, + 304, + 243 + ], + "lines": [ + { + "bbox": [ + 245, + 216, + 304, + 243 + ], + "spans": [ + { + "bbox": [ + 245, + 216, + 304, + 243 + ], + "type": "image", + "image_path": "08c1a764200f0d51043005c6386236ed3e06c6ec8600f4d80f6427aefe37419a.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 266, + 243, + 279, + 251 + ], + "lines": [ + { + "bbox": [ + 266, + 243, + 279, + 251 + ], + "spans": [ + { + "bbox": [ + 266, + 243, + 279, + 251 + ], + "type": "text", + "content": "Full" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 59, + 272, + 100, + 319 + ], + "blocks": [ + { + "bbox": [ + 78, + 243, + 96, + 251 + ], + "lines": [ + { + "bbox": [ + 78, + 243, + 96, + 251 + ], + "spans": [ + { + "bbox": [ + 78, + 243, + 96, + 251 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 99, + 258, + 251, + 269 + ], + "lines": [ + { + "bbox": [ + 99, + 258, + 251, + 269 + ], + "spans": [ + { + "bbox": [ + 99, + 258, + 251, + 269 + ], + "type": "text", + "content": "Figure 9. Ablation of animation modules." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 272, + 100, + 319 + ], + "lines": [ + { + "bbox": [ + 59, + 272, + 100, + 319 + ], + "spans": [ + { + "bbox": [ + 59, + 272, + 100, + 319 + ], + "type": "image", + "image_path": "5c196a350c7472c310176fc8ce70aeb206faae6cbd826a59d637ca825c5e440c.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 322, + 89, + 330 + ], + "lines": [ + { + "bbox": [ + 72, + 322, + 89, + 330 + ], + "spans": [ + { + "bbox": [ + 72, + 322, + 89, + 330 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 99, + 321, + 135, + 330 + ], + "lines": [ + { + "bbox": [ + 99, + 321, + 135, + 330 + ], + "spans": [ + { + "bbox": [ + 99, + 321, + 135, + 330 + ], + "type": "text", + "content": "Novel View" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 101, + 273, + 141, + 319 + ], + "blocks": [ + { + "bbox": [ + 101, + 273, + 141, + 319 + ], + "lines": [ + { + "bbox": [ + 101, + 273, + 141, + 319 + ], + "spans": [ + { + "bbox": [ + 101, + 273, + 141, + 319 + ], + "type": "image", + "image_path": "4127dca5d723113e4c933bb28e37d4e7c6c17bcc2a951a08f3485c52f1eac1f9.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 142, + 274, + 190, + 321 + ], + "blocks": [ + { + "bbox": [ + 142, + 274, + 190, + 321 + ], + "lines": [ + { + "bbox": [ + 142, + 274, + 190, + 321 + ], + "spans": [ + { + "bbox": [ + 142, + 274, + 190, + 321 + ], + "type": "image", + "image_path": "a107644927fae515ea994a11560366e629251892ccfb7ad688dcc80cc905d59b.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 322, + 175, + 330 + ], + "lines": [ + { + "bbox": [ + 159, + 322, + 175, + 330 + ], + "spans": [ + { + "bbox": [ + 159, + 322, + 175, + 330 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 190, + 322, + 225, + 330 + ], + "lines": [ + { + "bbox": [ + 190, + 322, + 225, + 330 + ], + "spans": [ + { + "bbox": [ + 190, + 322, + 225, + 330 + ], + "type": "text", + "content": "Novel View" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 95, + 331, + 255, + 342 + ], + "lines": [ + { + "bbox": [ + 95, + 331, + 255, + 342 + ], + "spans": [ + { + "bbox": [ + 95, + 331, + 255, + 342 + ], + "type": "text", + "content": "Figure 10. More results with diverse inputs." + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 190, + 274, + 230, + 321 + ], + "blocks": [ + { + "bbox": [ + 190, + 274, + 230, + 321 + ], + "lines": [ + { + "bbox": [ + 190, + 274, + 230, + 321 + ], + "spans": [ + { + "bbox": [ + 190, + 274, + 230, + 321 + ], + "type": "image", + "image_path": "3dbd67799bd7684a42045d51ab257f56ca9a16e74dd2b53a19969db5d7baccfd.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 231, + 274, + 266, + 319 + ], + "blocks": [ + { + "bbox": [ + 231, + 274, + 266, + 319 + ], + "lines": [ + { + "bbox": [ + 231, + 274, + 266, + 319 + ], + "spans": [ + { + "bbox": [ + 231, + 274, + 266, + 319 + ], + "type": "image", + "image_path": "7868cc0faa3507170f6b59803cd8692bb4cefbaf9270033afc15e7cdbd91e0e5.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 239, + 322, + 255, + 330 + ], + "lines": [ + { + "bbox": [ + 239, + 322, + 255, + 330 + ], + "spans": [ + { + "bbox": [ + 239, + 322, + 255, + 330 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 269, + 274, + 302, + 318 + ], + "blocks": [ + { + "bbox": [ + 269, + 274, + 302, + 318 + ], + "lines": [ + { + "bbox": [ + 269, + 274, + 302, + 318 + ], + "spans": [ + { + "bbox": [ + 269, + 274, + 302, + 318 + ], + "type": "image", + "image_path": "6392f69b9f4e1c2c849e3cfdc472430ea8278a00964b2454426cbea5b161daea.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 270, + 321, + 305, + 329 + ], + "lines": [ + { + "bbox": [ + 270, + 321, + 305, + 329 + ], + "spans": [ + { + "bbox": [ + 270, + 321, + 305, + 329 + ], + "type": "text", + "content": "Novel View" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "bbox": [ + 55, + 354, + 295, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 295, + 392 + ], + "type": "text", + "content": "of baselines degrades significantly as the viewing angle increases. Additional results and baseline comparisons are provided in the supplementary material." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 55, + 399, + 149, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 149, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 149, + 411 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 55, + 419, + 296, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 296, + 516 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 296, + 516 + ], + "type": "text", + "content": "Multiview image generation. We conduct an ablation study on two multi-view generation modules, as shown in Tab. 4 and Fig. 8. Incorporating the warping module allows us to preserve finer details, like tattoos, in novel views. While removing MA may lead to semantic inconsistencies, such as changes in gender. Adding mutual attention (MA) further improves fidelity and enhances identity similarity by injecting features from the reference images." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 55, + 519, + 296, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 296, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 296, + 674 + ], + "type": "text", + "content": "Animation. Our 3D representation is based on GaussianAvatar [27], which is FLAME-driven, enabling direct animation through FLAME parameters. To evaluate the base performance, we remove our video prior module. As shown in Fig. 9, without the video prior, the final avatar fails to synthesize accurate expressions. For consistency modeling, we compare our model with and without the COIN training. Without COIN, training directly on inconsistent generated images results in blurred textures. With COIN training, regions such as the teeth remain sharp, and finer details, like individual hair strands, are preserved. The quantitative results in Tab. 5 further validate the effectiveness of the COIN optimization across novel views." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Processing times. We analyzed the processing times of each stage. Our method consists of three stages: a) 3D GAN inversion (2 mins); b) Multiview image & video generation" + } + ] + } + ], + "index": 40 + }, + { + "type": "table", + "bbox": [ + 316, + 103, + 553, + 171 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 553, + 102 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 553, + 102 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 553, + 102 + ], + "type": "text", + "content": "Table 4. Ablation of multiview image generation modules. \"Warp\" refers to the warping-based control generation (Sec. 3.2) and MA to cross-view mutual attention (Sec. 3.2)." + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 103, + 553, + 171 + ], + "lines": [ + { + "bbox": [ + 316, + 103, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 103, + 553, + 171 + ], + "type": "table", + "html": "
MethodCLIP-I ↑ID ↑FID
w/o Warp w/o MA0.73280.4787182.27
w/o Warp0.78860.4915171.08
w/o MA0.81510.4951172.43
Full0.81620.4984166.96
", + "image_path": "79beeb0797be7a29a7fc8baa5f0a6766e80cdc56c1e27e8bcab154cf0bb90f8e.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "table_body" + } + ], + "index": 42 + }, + { + "type": "table", + "bbox": [ + 316, + 183, + 553, + 228 + ], + "blocks": [ + { + "bbox": [ + 360, + 173, + 507, + 182 + ], + "lines": [ + { + "bbox": [ + 360, + 173, + 507, + 182 + ], + "spans": [ + { + "bbox": [ + 360, + 173, + 507, + 182 + ], + "type": "text", + "content": "Table 5. Ablation of animation modules." + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 183, + 553, + 228 + ], + "lines": [ + { + "bbox": [ + 316, + 183, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 183, + 553, + 228 + ], + "type": "table", + "html": "
MethodCLIP-I ↑ID ↑FID ↓
w/o COIN0.76880.4952144.95
Full0.77290.5010142.80
", + "image_path": "3324c71000ce7abaf0fc0db71d278e8dc4d2e2decf6d7c5ce85ddcf379db5d27.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "table_body" + } + ], + "index": 44 + }, + { + "type": "table", + "bbox": [ + 318, + 253, + 551, + 298 + ], + "blocks": [ + { + "bbox": [ + 314, + 231, + 553, + 251 + ], + "lines": [ + { + "bbox": [ + 314, + 231, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 314, + 231, + 553, + 251 + ], + "type": "text", + "content": "Table 6. Time comparison with optimization-based avatar reconstruction methods." + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 253, + 551, + 298 + ], + "lines": [ + { + "bbox": [ + 318, + 253, + 551, + 298 + ], + "spans": [ + { + "bbox": [ + 318, + 253, + 551, + 298 + ], + "type": "table", + "html": "
MethodNerFaceIMAvatarPointAvatarHeadStudio
Time (hour)54h48h6h2h
MethodMonoGAGaussianAvatarHq3davatarOurs
Time (hour)7h2h12h2.5h
", + "image_path": "e75873be5fa38fa9abc591503832299e89d4e50a7fd360efb479dd618ace1a96.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "table_body" + } + ], + "index": 46 + }, + { + "bbox": [ + 313, + 308, + 554, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 554, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 554, + 392 + ], + "type": "text", + "content": "(10 mins), FLAME parameters extraction (10 mins); and c) COIN-training to overfit multiview videos (2 hours). The generation of a single avatar " + }, + { + "bbox": [ + 313, + 308, + 554, + 392 + ], + "type": "inline_equation", + "content": "\\sim 2.5" + }, + { + "bbox": [ + 313, + 308, + 554, + 392 + ], + "type": "text", + "content": " hours, which is comparable to or faster than other optimization-based avatar reconstruction methods, as shown in Tab. 6. At inference, we achieve real time rendering at 156 FPS with a resolution of " + }, + { + "bbox": [ + 313, + 308, + 554, + 392 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 308, + 554, + 392 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 313, + 396, + 554, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 554, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 554, + 468 + ], + "type": "text", + "content": "Diverse inputs. To validate robustness, we conduct experiments on diverse inputs, including cartoon, line drawing, and portraits with extreme camera poses. The results in Fig. 10 demonstrate consistent performance across various domains. Additional video results are available on the webpage in the supplementary materials." + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 313, + 478, + 391, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 478, + 391, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 391, + 491 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we present a novel method for generating high-quality 4D portraits from a single image. By leveraging multiple priors, including 3D-GAN for geometry initialization, image prior for texture enhancement, and video prior for animation, our method can generate animated avatars with detailed textures, excellent consistency, and plausible rendition of views missing in the input. The proposed depth-guided warping ensures cross-view consistency during the generation process, while our COIN-training strategy enables high-quality 4D reconstruction from potentially inconsistent multi-view data. Extensive experiments demonstrate that our method outperforms existing approaches in terms of geometry accuracy, texture quality, and identity preservation across different viewpoints and expressions. Our framework makes high-quality 4D portrait generation more accessible by removing the requirement for multi-view data capture, paving the way for broader applications." + } + ] + } + ], + "index": 50 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11619" + } + ] + } + ], + "index": 51 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Sizhe An, Hongyi Xu, Yichun Shi, Guoxian Song, Umit Y Ogras, and Linjie Luo. Panohead: Geometry-aware 3d full-head synthesis in 360deg. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20950-20959, 2023. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "type": "text", + "content": "[2] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22560–22570, 2023. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 205, + 295, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 295, + 270 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 295, + 270 + ], + "type": "text", + "content": "[3] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16123-16133, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 272, + 295, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 272, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 272, + 295, + 304 + ], + "type": "text", + "content": "[4] Xuangeng Chu and Tatsuya Harada. Generalizable and animatable gaussian head avatar. arXiv preprint arXiv:2410.07971, 2024.3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 306, + 295, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 306, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 306, + 295, + 350 + ], + "type": "text", + "content": "[5] Xuangeng Chu, Yu Li, Ailing Zeng, Tianyu Yang, Lijian Lin, Yunfei Liu, and Tatsuya Harada. Gpavatar: Generalizable and precise head avatar from image (s). arXiv preprint arXiv:2401.10215, 2024.3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 352, + 295, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 352, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 352, + 295, + 407 + ], + "type": "text", + "content": "[6] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 409, + 295, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 409, + 295, + 462 + ], + "spans": [ + { + "bbox": [ + 62, + 409, + 295, + 462 + ], + "type": "text", + "content": "[7] Yu Deng, Duomin Wang, Xiaohang Ren, Xingyu Chen, and Baoyuan Wang. Portrait4d: Learning one-shot 4d head avatar synthesis using synthetic data. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1, 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 465, + 295, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 465, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 62, + 465, + 295, + 520 + ], + "type": "text", + "content": "[8] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2663-2671, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 522, + 295, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 522, + 295, + 576 + ], + "spans": [ + { + "bbox": [ + 62, + 522, + 295, + 576 + ], + "type": "text", + "content": "[9] Philip-William Grassal, Malte Prinzler, Titus Leistner, Carsten Rother, Matthias Nießner, and Justus Thies. Neural head avatars from monocular rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18653–18664, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 578, + 295, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 578, + 295, + 632 + ], + "spans": [ + { + "bbox": [ + 57, + 578, + 295, + 632 + ], + "type": "text", + "content": "[10] Jianzhu Guo, Dingyun Zhang, Xiaogiang Liu, Zhizhou Zhong, Yuan Zhang, Pengfei Wan, and Di Zhang. Liveportrait: Efficient portrait animation with stitching and retargeting control. arXiv preprint arXiv:2407.03168, 2024. 2, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 635, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 678 + ], + "type": "text", + "content": "[11] Xiao Han, Yukang Cao, Kai Han, Xiatian Zhu, Jiankang Deng, Yi-Zhe Song, Tao Xiang, and Kwan-Yee K Wong. Headsculpt: Crafting 3d head avatars with text. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[12] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilib-" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 74, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 74, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 74, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 74, + 553, + 94 + ], + "type": "text", + "content": "rium. Advances in neural information processing systems, 30, 2017. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 96, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 150 + ], + "type": "text", + "content": "[13] Fa-Ting Hong, Longhao Zhang, Li Shen, and Dan Xu. Depth-aware generative adversarial network for talking head video generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3397-3406, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 152, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 553, + 206 + ], + "type": "text", + "content": "[14] Yang Hong, Bo Peng, Haiyao Xiao, Ligang Liu, and Juyong Zhang. Headnerf: A real-time nerf-based parametric head model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20374-20384, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "type": "text", + "content": "[15] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 3, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 254, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 254, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 317, + 254, + 553, + 296 + ], + "type": "text", + "content": "[16] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 299, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 299, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 317, + 299, + 553, + 342 + ], + "type": "text", + "content": "[17] Taras Khakhulin, Vanessa Sklyarova, Victor Lempitsky, and Egor Zakharov. Realistic one-shot mesh-based head avatars. In European Conference on Computer Vision, pages 345-362. Springer, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 344, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 344, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 317, + 344, + 553, + 387 + ], + "type": "text", + "content": "[18] Tobias Kirschstein, Shenhan Qian, Simon Giebenhain, Tim Walter, and Matthias Nießner. Nersemble: Multi-view radiance field reconstruction of human heads. ACM Transactions on Graphics (TOG), 42(4):1-14, 2023. 1, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 388, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 388, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 317, + 388, + 553, + 431 + ], + "type": "text", + "content": "[19] Shikai Li, Jianglin Fu, Kaiyuan Liu, Wentao Wang, KwanYee Lin, and Wayne Wu. Cosmicman: A text-to-image foundation model for humans. In Computer Vision and Pattern Recognition (CVPR), 2024. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 434, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 434, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 317, + 434, + 553, + 477 + ], + "type": "text", + "content": "[20] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6), 2017. 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 479, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 479, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 479, + 553, + 544 + ], + "type": "text", + "content": "[21] Weichuang Li, Longhao Zhang, Dong Wang, Bin Zhao, Zhi-gang Wang, Mulin Chen, Bang Zhang, Zhongjian Wang, Liefeng Bo, and Xuelong Li. One-shot high-fidelity talking-head synthesis with deformable neural radiance field. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17969-17978, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 545, + 553, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 545, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 317, + 545, + 553, + 609 + ], + "type": "text", + "content": "[22] Xian Liu, Xiaohang Zhan, Jiaxiang Tang, Ying Shan, Gang Zeng, Dahua Lin, Xihui Liu, and Ziwei Liu. Humangau-sian: Text-driven 3d human generation with gaussian splatt-ing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6646-6657, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 613, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 553, + 645 + ], + "type": "text", + "content": "[23] Shengjie Ma, Yanlin Weng, Tianjia Shao, and Kun Zhou. 3d gaussian blendshapes for head avatar animation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-10, 2024. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "type": "text", + "content": "[24] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[25] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf:" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "11620" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 105 + ], + "type": "text", + "content": "Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "type": "text", + "content": "[26] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 207 + ], + "type": "text", + "content": "[27] Shenhan Qian, Tobias Kirschstein, Liam Schoneveld, Davide Davoli, Simon Giebenhain, and Matthias Nießner. Gaussianavatars: Photorealistic head avatars with rigged 3d gaus-sians. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20299-20309, 2024. 1, 2, 5, 6, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 274 + ], + "type": "text", + "content": "[28] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 275, + 294, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 294, + 341 + ], + "type": "text", + "content": "[29] Amit Raj, Srinivas Kaza, Ben Poole, Michael Niemeyer, Nataniel Ruiz, Ben Mildenhall, Shiran Zada, Kfir Aberman, Michael Rubinstein, Jonathan Barron, et al. Dreambooth3d: Subject-driven text-to-3d generation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2349–2359, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 342, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 294, + 396 + ], + "type": "text", + "content": "[30] Yurui Ren, Ge Li, Yuanqi Chen, Thomas H Li, and Shan Liu. Pirenderer: Controllable portrait image generation via semantic neural rendering. In Proceedings of the IEEE/CVF international conference on computer vision, pages 13759-13768, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 399, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 441 + ], + "type": "text", + "content": "[31] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on graphics (TOG), 42(1):1-13, 2022. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 444, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 444, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 444, + 294, + 498 + ], + "type": "text", + "content": "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 500, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 294, + 533 + ], + "type": "text", + "content": "[33] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 534, + 294, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 294, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 294, + 577 + ], + "type": "text", + "content": "[34] Jingxiang Sun, Xuan Wang, Lizhen Wang, Xiaoyu Li, Yong Zhang, Hongwen Zhang, and Yebin Liu. Next3d: Generative neural texture rasterization for 3d-aware head avatars. In CVPR, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 622 + ], + "type": "text", + "content": "[35] Felix Taubner, Ruihang Zhang, Mathieu Tuli, and David B Lindell. Cap4d: Creating animatable 4d portrait avatars with morphable multi-view diffusion models. arXiv preprint arXiv:2412.12093, 2024. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 668 + ], + "type": "text", + "content": "[36] Kartik Teotia, Mallikarjun B R, Xingang Pan, Hyeongwoo Kim, Pablo Garrido, Mohamed Elgharib, and Christian Theobalt. HQ3davatar: High-quality implicit 3d head avatar. ACM Transactions on Graphics, 43(3):1-24, 2024. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[37] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "video diffusion. arXiv preprint arXiv:2403.12008, 2024. 2, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 95, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 553, + 150 + ], + "type": "text", + "content": "[38] Kaisiyuan Wang, Qianyi Wu, Linsen Song, Zhuoqian Yang, Wayne Wu, Chen Qian, Ran He, Yu Qiao, and Chen Change Loy. Mead: A large-scale audio-visual dataset for emotional talking-face generation. In European Conference on Computer Vision, pages 700–717. Springer, 2020. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 151, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 553, + 183 + ], + "type": "text", + "content": "[39] Huawei Wei, Zejun Yang, and Zhisheng Wang. Aniportrait: Audio-driven synthesis of photorealistic portrait animation. arXiv preprint arXiv:2403.17694, 2024. 1, 2, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 184, + 553, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 238 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 238 + ], + "type": "text", + "content": "[40] Yiqian Wu, Hao Xu, Xiangjun Tang, Xien Chen, Siyu Tang, Zhebin Zhang, Chen Li, and Xiaogang Jin. Portrait3d: Text-guided high-quality 3d portrait generation using pyramid representation and gans prior. ACM Transactions on Graphics (TOG), 43(4):1-12, 2024. 1, 2, 3, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "text", + "content": "[41] Sicheng Xu, Jiaolong Yang, Dong Chen, Fang Wen, Yu Deng, Yunde Jia, and Xin Tong. Deep 3d portrait from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7710-7720, 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 294, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 348 + ], + "type": "text", + "content": "[42] Zhenhui Ye, Tianyun Zhong, Yi Ren, Jiaqi Yang, Weichuang Li, Jiawei Huang, Ziyue Jiang, Jinzheng He, Rongjie Huang, Jinglin Liu, et al. Real3d-portrait: One-shot realistic 3d talking portrait synthesis. arXiv preprint arXiv:2401.08503, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 350, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 415 + ], + "type": "text", + "content": "[43] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. In European conference on computer vision, pages 85-101. Springer, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 415, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 470 + ], + "type": "text", + "content": "[44] Fei Yin, Yong Zhang, Xuan Wang, Tengfei Wang, Xiaoyu Li, Yuan Gong, Yanbo Fan, Xiaodong Cun, Ying Shan, Cengiz Oztireli, et al. 3d gan inversion with facial symmetry prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 342-351, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 471, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 553, + 525 + ], + "type": "text", + "content": "[45] Wangbo Yu, Yanbo Fan, Yong Zhang, Xuan Wang, Fei Yin, Yunpeng Bai, Yan-Pei Cao, Ying Shan, Yang Wu, Zhongqian Sun, et al. Nofa: Nerf-based one-shot facial avatar reconstruction. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-12, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 526, + 553, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 553, + 580 + ], + "type": "text", + "content": "[46] Bowen Zhang, Yiji Cheng, Jiaolong Yang, Chunyu Wang, Feng Zhao, Yansong Tang, Dong Chen, and Baining Guo. Gaussiancube: Structuring gaussian splatting using optimal transport for 3d generative modeling. arXiv preprint arXiv:2403.19655, 2024. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 581, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 581, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 581, + 553, + 613 + ], + "type": "text", + "content": "[47] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 614, + 555, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 555, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 555, + 669 + ], + "type": "text", + "content": "[48] Yufeng Zheng, Wang Yifan, Gordon Wetzstein, Michael J Black, and Otmar Hilliges. Pointavatar: Deformable point-based head avatars from videos. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21057-21067, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[49] Zhenglin Zhou, Fan Ma, Hehe Fan, Zongxin Yang, and Yi Yang. Headstudio: Text to animatable head avatars with 3d gaussian splatting. In European Conference on Computer Vision, pages 145-163. Springer, 2024. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "11621" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/FaceLift_ Learning Generalizable Single Image 3D Face Reconstruction from Synthetic Heads/1763108f-f851-4f26-ba5a-510a2c051a36_content_list.json b/2025/FaceLift_ Learning Generalizable Single Image 3D Face Reconstruction from Synthetic Heads/1763108f-f851-4f26-ba5a-510a2c051a36_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..852665a4dc44c3dc971e693351292ab3c7dd7fc5 --- /dev/null +++ b/2025/FaceLift_ Learning Generalizable Single Image 3D Face Reconstruction from Synthetic Heads/1763108f-f851-4f26-ba5a-510a2c051a36_content_list.json @@ -0,0 +1,2720 @@ +[ + { + "type": "text", + "text": "FaceLift: Learning Generalizable Single Image 3D Face Reconstruction from Synthetic Heads", + "text_level": 1, + "bbox": [ + 109, + 128, + 883, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Weijie Lyu $^{1*}$ Yi Zhou $^{2}$ Ming-Hsuan Yang $^{1}$ Zhixin Shu $^{2}$", + "bbox": [ + 248, + 202, + 746, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of California, Merced 2Adobe Research", + "bbox": [ + 284, + 227, + 712, + 244 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cfe088eee6b0315c59c76e4e0696b845e6a2ca48fbeffb44833afc0d3bfb61c1.jpg", + "image_caption": [ + "Single Image Input" + ], + "image_footnote": [], + "bbox": [ + 102, + 258, + 251, + 375 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4c793c5ef31a912d348c6c0cc95f77ae46f7d17e7dc33b912fa9241149b36de4.jpg", + "image_caption": [ + "High-fidelity 3D Gaussian Representation" + ], + "image_footnote": [], + "bbox": [ + 240, + 258, + 418, + 375 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dabe944618ae82f7db9e5729ef8c6fdb9b8b48395e9768665d7e5ccdb4c32e6d.jpg", + "image_caption": [ + "Full Head NVS", + "Figure 1. FaceLift transforms a single facial image into a high-fidelity 3D Gaussian head representation. Trained exclusively on synthetic 3D data, our pipeline first generates sparse, identity-preserving multiview images of the input head using a diffusion model. These sparse generated views are then fed into a transformer-based 3D Gaussian splats reconstructor, producing complete and detailed 3D head representation that generalize remarkably well to real-world human images. Project page: https://weijielyu.github.io/FaceLift." + ], + "image_footnote": [], + "bbox": [ + 424, + 258, + 566, + 375 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cab329f58b2a5b3e56294b1affd051cc3875f0fdf537c45b1b82e1be3e3c6cb4.jpg", + "image_caption": [ + "High-quality Details" + ], + "image_footnote": [], + "bbox": [ + 571, + 258, + 732, + 375 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 479, + 326, + 496 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present FaceLift, a novel feed-forward approach for generalizable high-quality 360-degree 3D head reconstruction from a single image. Our pipeline first employs a multiview latent diffusion model to generate consistent side and back views from a single facial input, which then feed into a transformer-based reconstructor that produces a comprehensive 3D Gaussian splats representation. Previous methods for monocular 3D face reconstruction often lack full view coverage or view consistency due to insufficient multi-view supervision. We address this by creating a high-quality synthetic head dataset that enables consistent supervision across viewpoints. To bridge the domain gap between synthetic training data and real-world images, we propose a simple yet effective technique that ensures the view generation process maintains fidelity to the input by learning to reconstruct the input image alongside the view generation. Despite being trained exclusively on synthetic data, our method demonstrates remarkable generalization to real-world images. Through extensive qualitative and quantitative evaluations, we show that FaceLift outperforms state-of-the-art 3D face reconstruction methods on identity preservation, detail recovery and rendering quality.", + "bbox": [ + 88, + 516, + 485, + 849 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 479, + 643, + 494 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D face reconstruction has been a central focus in computer vision and graphics research for decades, driven by its crucial applications in immersive virtual and augmented realities, VFX and gaming, digital entertainment, and next-generation telepresence systems. However, achieving high quality reconstruction from a single image remains very challenging. On one hand, the monocular face reconstruction problem is highly ill-posed – a single 2D image can be produced by countless different 3D face shapes, creating fundamental ambiguity. On the other hand, the human visual system is highly attuned to facial details, making even subtle artifacts and imperfections noticeable to the eye.", + "bbox": [ + 511, + 505, + 906, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Traditional 3D head synthesis approaches typically use parametric textured mesh models [32, 60] trained on 3D scan datasets. While these models enable basic head generation, the rendered images frequently lack fine-scale geometric detail, realistic textures, and convincing hair, limiting their perceptual realism and expressiveness. Recent breakthroughs in image generative models [19, 23] and novel view synthesis techniques [27, 40] have opened new possibilities for this research area. Leveraging these developments, recent works [1, 72] use neural 3D representations to learn effective 3D head representation from unstructured real face image datasets [25, 68]. While these datasets improve the realism and diversity of rendering results, they fail to provide multi-view supervision for modeling 3D consis", + "bbox": [ + 511, + 688, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work was done when Weijie Lyu was an intern at Adobe Research.", + "bbox": [ + 114, + 886, + 477, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12691", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9714794b1877a26ae2d9a09376437b1f1598dca3e1e561b53c0e802b8ef606e4.jpg", + "image_caption": [ + "Figure 2. Comparison with RodinHD. RodinHD [73] trains triplane diffusion with synthetic data, resulting in apparent identity loss. In contrast, FaceLift achieves better identity preservation and generalizes effectively to real human portraits." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 478, + 338 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tency causing view inconsistency and artifacts on the back of the head. Since diverse multi-view real images are hard to acquire, RodinHD [73] leverages synthetic multi-view images to train generative models that directly output 3D neural representations of the head. However, training solely on synthetic data often results in significant perceptual identity loss in the generated outputs, as demonstrated in Fig. 2.", + "bbox": [ + 89, + 401, + 483, + 507 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we present FaceLift, a pipeline for learning generalizable and high-fidelity single image to 3D face representation from synthetic head data. To achieve high quality reconstruction that preserves the input facial identity, we adopt a two-stage pipeline to first generate identity preserving multi-view images using a diffusion model [48], followed by a transformer-based feed-forward reconstructor to fuse the generated sparse views into a comprehensive 3D Gaussian representation. We train the model with synthetic images - multi-view renderings of 3D synthetic human portraits using Blender. We highlight two key techniques for generalizing to real-world images and preserving input facial identity. First, we emphasize the importance of reconstructing the input image alongside the view synthesis task in the conditional diffusion model training, which significantly improved generalization capability in testing. Second, we demonstrate that training the feed-forward reconstructor benefits from a two-stage training process: pretraining on general objects [12] to acquire a rich geometry and texture prior, followed by fine-tuning on synthetic human head data to capture head-specific geometry. With our two-stage approach, we focus on learning identity preservation in the image space during the first stage, achieving higher input fidelity compared to existing methods.", + "bbox": [ + 89, + 507, + 483, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Comparing with prior art, we achieve three key advancements: (1) robust view consistency through multi-view at", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tention and supervision, (2) improved generalization from training techniques and foundational model, ensuring accurate identity preservation, (3) high-quality facial texture and hair details via pixel-aligned Gaussian representation.", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We extensively evaluate FaceLift quantitatively and qualitatively across diverse datasets. Using real multiview studio captures [39] and an independent synthetic human dataset [8], our approach consistently surpasses previous state-of-the-art methods across all evaluation metrics. Through extensive testing on in-the-wild portrait images, we demonstrate that our method reconstructs complete 3D heads with fine-grained details, accurate identity preservation, and high visual fidelity. Comparisons and ablation studies confirm that multi-view consistent training data is crucial for high-fidelity face reconstruction. Our contributions are summarized as follows:", + "bbox": [ + 511, + 152, + 906, + 332 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose FaceLift, a framework that reconstructs a high-fidelity 3D head from a single image using view generation and feed-forward reconstructor.", + "- Despite being trained solely on synthetic human head data, our approach shows no domain gap on real-world images, highlighting both the effectiveness of synthetic data and our model's robust generalization capabilities.", + "- We construct two benchmarks on single-image to 3D full head reconstruction tasks using the publicly available datasets Cafca [8] and Ava-256 [39] to quantitatively evaluate models' performance on both reconstruction accuracy and identity preservation ability.", + "- Our comprehensive evaluation confirms that our approach achieves state-of-the-art performance in reconstruction accuracy and identity preservation." + ], + "bbox": [ + 513, + 335, + 903, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 575, + 653, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Face Reconstruction. 3D face reconstruction has been a long-standing challenge in computer vision, with substantial progress driven by diverse approaches. Vetter and Blanz [60] pioneer a method for synthesizing 3D faces by linearly blending multiple 3D templates, now widely known as blendshapes. This work establishes the foundation for 3D Morphable Models (3DMMs), which represent 3D face shapes and textures as principal components derived from scanned data. Subsequent research [5, 6, 32, 34, 47] extend this framework, enabling the generation of new 3D faces by manipulating blending coefficients. However, these methods produce mesh-based representations that lack fine details and are limited to modeling the front of the face, excluding hair and 360-degree synthesis. While 3DMM-based methods have been foundational, recent advances in deep learning, especially Generative Adversarial Networks (GANs) [19, 25, 26], have greatly improved 3D face synthesis quality. EG3D [72] uses a tri-plane NeRF representation with a pose-conditioned StyleGAN2 [26] framework. Follow-up works [3, 33] achieve single-image-to-3D gen", + "bbox": [ + 511, + 598, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12692", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "eration through GAN inversion [11]. Despite their success, these methods can only synthesize near-frontal views. To overcome this, PanoHead [1] introduces a tri-grid neural volume representation, enabling full 360-degree head synthesis. Unfortunately, it does not provide a 3D head representation for consistent multi-view rendering. Recent efforts explore alternative representations for 3D face reconstruction from sparse input, such as a single image [17, 42, 61] or few-shot images [7]. However, these methods still require pre-instance optimization. Rodin [63] and its extension RodinHD [73] employ an image-conditioned diffusion model to generate a triplane representation of a human head for full-head novel view synthesis. Nevertheless, their triplane diffusion model is limited to synthetic data and struggles to achieve high-fidelity reconstructions from real-world images. For animatable 3D head avatars generations, Morphable Diffusion [10] generates multi-view consistent images from a single image using a morphable mesh, while HeadGAP [76] generates 3D animatable head avatars using few-shot input, leveraging 3D head priors learned from large-scale data. In contrast, our work focuses on leveraging synthetic training data to produce high-fidelity, detailed 3D Gaussian head models.", + "bbox": [ + 89, + 90, + 480, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Synthetic Human Data. Capturing high-quality 3D data of real humans requires a controlled studio environment and costly photography equipment [39]. As an alternative, large-scale synthetic 3D head datasets have emerged as an effective and resource-efficient solution for tasks like human head reconstruction [8, 63, 65, 73] and photorealistic relighting [9, 71], offering a scalable way to train models without the restrictions of real-world data acquisition. Inspired by these previous works, we aim to use synthetic data to improve the model's understanding of the human head and minimize the generalization gap between synthetic data training and real-world inference.", + "bbox": [ + 89, + 443, + 480, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image or Text to 3D. Generative models have achieved remarkable success in 2D image generation with VAEs [28, 58], GANs [19, 25, 26], and diffusion models [23, 48, 54]. Building on this success, extensive research has extended these models to 3D content generation [18, 41, 43, 66]. Starting with DreamFusion [45], numerous works [36, 46, 51, 57, 64] try to distill NeRF [40, 62, 72] or 3D Gaussians [27] representation from 2D image diffusion using a Score Distillation Sampling (SDS) loss. These methods can produce high-quality results but often encounter challenges such as slow optimization, over-saturated colors, and the Janus problem. To overcome these challenges, recent works [30, 31, 35, 37, 53] generate multi-view images with high consistency, which can be directly used for 3D reconstruction with neural reconstruction methods [27, 40, 62]. However, optimizing NeRF or NeuS remains far from real-time performance. Recent advancements in large 3D reconstruction models (LRMs) [24, 30, 56, 74] offer a path", + "bbox": [ + 89, + 628, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "way to faster 3D reconstruction. Leveraging scalable transformer architectures [15, 59] and large datasets like Objverse [12, 13], these models effectively capture generalizable 3D priors. Unlike traditional pre-scene optimization methods [27, 40, 62], LRMs employ a feed-forward approach, enabling the prediction of high-quality NeRF, mesh, or 3D Gaussian representations from sparse images in under a second. However, most of these research efforts are applied to general objects, with limited or suboptimal results for 3D head reconstruction.", + "bbox": [ + 511, + 90, + 903, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 511, + 251, + 686, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 3, given a single frontal image of a human face $y$ , our goal is to reconstruct a complete 3D head $G$ , represented as 3D Gaussian splats, with detailed facial texture and preserved identity. This requires our system to have prior knowledge on the geometry structure of a human face and the ability to synthesis plausible details which are not visible in the input view. Hence, we train a multiview diffusion model $f_{D}$ on synthetic human head data to generate $N$ views $x_0^1, x_0^2, \\ldots, x_0^N$ covering $360^\\circ$ of the human head while achieving multi-view consistency and preserving identity. We choose pixel-aligned 3D Gaussians to obtain the final 3D representation. Compared to NeRFs and meshes, 3D Gaussians offer explicit volumetric primitives that better capture subtle facial geometry and fine details. Their semi-transparent kernels naturally model effects like wispy hair and translucency, which are challenging for discrete surfaces or density fields. These generated views $x_0^{1:N}$ from the diffusion model, along with their corresponding Plücker ray coordinates $\\mathbf{P}^{(1:N)}$ , are input into a transformer-based Gaussian reconstructor $f_R$ to predict a set of 3D Gaussians $G$ . Training of the Gaussian reconstructor follows a pre-training process on general objects [12] and a fine-tuning process on synthetic human head data.", + "bbox": [ + 511, + 273, + 903, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Synthetic Human Head Dataset", + "text_level": 1, + "bbox": [ + 511, + 625, + 790, + 640 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We implement a 3D head asset generation pipeline inspired by [65]. Our process begins with a collection of high-quality, artist-created 3D head meshes, which we enhance by incorporating detailed facial components, including eyes, teeth, gums, and both facial and scalp hair. We then augment these base models through rigging for pose variation and blendshape deformation for diverse facial expressions. The final head models are enriched with a set of PBR texture maps, including albedo, normal, roughness, specular, and subsurface scattering maps. At last, we dress the head model with a collection of clothing assets. The entire pipeline is implemented in Blender and the images are rendered with Cycles renderer.", + "bbox": [ + 511, + 643, + 903, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To train our networks, we render images (samples shown in Fig. 4) at $512 \\times 512$ resolution from 200 unique identities, each with 50 varied appearances, including different hairstyles, skin tones, expressions, clothes, poses, etc.", + "bbox": [ + 511, + 840, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12693", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0b6dcfa3e9d95382ee8a66e0d65d3b40dd971f7d166af00dc6902fe0699a9ddf.jpg", + "image_caption": [ + "Figure 3. Overview of FaceLift. Given a single image of a human face $y$ as input, we train an image-conditioned, multi-view diffusion model to generate novel views $x_0^1, \\ldots, x_0^N$ covering the entire head. By generating $x_0^1$ the same as $y$ and leveraging the high-quality synthetic data, our multi-view latent diffusion model can hallucinate unseen views of the human head with high-fidelity and multi-view consistency. We then train a transformer-based reconstructor $f_R$ , which takes multi-view images $x_0^{1:N}$ and their camera poses $\\mathbf{P}^{1:N}$ as input and generates 3D Gaussian splats $G$ to represent the human head." + ], + "image_footnote": [], + "bbox": [ + 112, + 88, + 883, + 219 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ce4410f083f829d2240ca808516f1e389ee172b9e65296f4a48fba896189a94e.jpg", + "image_caption": [ + "Figure 4. Synthetic data examples. Top row: six views for diffusion training. Bottom row: samples of random views for reconstructor training." + ], + "image_footnote": [], + "bbox": [ + 99, + 305, + 478, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We render our training dataset under two types of lighting conditions: (1) ambient light and (2) random HDR environment light. We render six views for each subject to train the multi-view diffusion model. For fine-tuning the transformer-based reconstructor, we render 32 views with random camera poses.", + "bbox": [ + 89, + 457, + 483, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. View Generation", + "text_level": 1, + "bbox": [ + 89, + 551, + 256, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We model the sparse view generation from a single image input as a conditional diffusion process. We use a multiview diffusion model $f_{D}$ to generate $N$ views, denoted as $X_0^1, X_0^2, \\ldots, X_0^N$ , given a single front-facing image $y$ and CLIP text embeddings $e^1, e^2, \\ldots, e^N$ corresponding to each generated view. This process is expressed as:", + "bbox": [ + 89, + 570, + 483, + 660 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{X _ {0} ^ {1}, X _ {0} ^ {2}, \\dots , X _ {0} ^ {N} \\right\\} = f _ {D} \\left(y, \\left\\{e ^ {1}, e ^ {2}, \\dots , e ^ {N} \\right\\}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 661, + 482, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We aim to learn the joint distribution of all these views, conditioning on the input image $y$ and text embedding $e^1, e^2, \\ldots, e^N$ . We denote the joint distribution as:", + "bbox": [ + 89, + 686, + 482, + 731 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {f _ {D}} \\left(x _ {0} ^ {1: N} \\mid y, e ^ {1: N}\\right) := p _ {f _ {D}} \\left(\\left\\{x _ {0} ^ {1}, \\dots x _ {0} ^ {N} \\right\\} \\mid y, \\left\\{e ^ {1}, \\dots e ^ {N} \\right\\}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 732, + 485, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "View Selection. Given a single near frontal view face image with azimuth $\\alpha$ , our multi-view diffusion model generates six views with azimuths equal to $\\{\\alpha, \\alpha \\pm 45^\\circ, \\alpha \\pm 90^\\circ, \\alpha + 180^\\circ\\}$ , covering 360 degrees of the human head. Elevation is 0 for all images. We opt for six views as the optimal balance - fewer views compromise detail quality while more views become computationally prohibitive for full head reconstruction. An ablation study comparing four, six, and eight views is presented in Sec. 5.2.", + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-view Attention. To ensure the consistency of the generated novel views, we use a multi-view attention mechanism to facilitate information propagation and implicitly encode multi-view dependencies. Our attention module encourages multi-view consistency by extending the 2D self-attention mechanism to 3D and enabling interactions across views. Instead of treating each view independently, we apply self-attention across all views simultaneously, allowing information to be shared between them. Specifically, we start with an input tensor of shape $B \\times V \\times H \\times W \\times C$ , where $B$ is the batch size, $V$ is the number of views, $H$ and $W$ denote the spatial resolution of the intermediate feature maps, and $C$ is the number of feature channels. We reshape this tensor to $B \\times VHW \\times C$ , treating all spatial locations across all views as a unified sequence of tokens for self-attention. This design allows the model to learn multi-view correlations by sharing information across views within the attention layers, enabling it to generate consistent RGB images. We provide an ablation study on the multi-view attention mechanism in the supplementary material.", + "bbox": [ + 511, + 304, + 906, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Input View Reconstruction. During Training, we enforce the first generated view to share the same camera with the input image. In other words, we reconstruct the input view in the view generation process. We find this approach, combined with the multi-view attention mechanism, significantly outperforms the alternative strategy of generating only novel views, which tends to overfit to synthetic training identities and compromises generalization capability as we will show in Sec. 5.1 and the supplementary material.", + "bbox": [ + 511, + 606, + 905, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Multi-view to 3D Gaussians Reconstruction", + "text_level": 1, + "bbox": [ + 511, + 746, + 882, + 760 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Transformer-based Reconstructor. We choose pixel-aligned 3D Gaussians as the final 3D representation. Each Gaussian $G_{i}$ is defined by position $p_i$ , scale $s_i$ , orientation $q_{i}$ , opacity $\\alpha_{i}$ , and color features $c_{i}$ . We use a transformer-based reconstructor $f_{G}$ to obtain 3D Gaussians from generated multi-view images $x_0^{1:N}$ and their corresponding Plücker ray coordinates [44] $\\mathbf{P}^{1:N}$ :", + "bbox": [ + 511, + 763, + 905, + 868 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{G _ {i} \\right\\} _ {i = 1} ^ {N H W} = \\left\\{p _ {i}, s _ {i}, q _ {i}, \\alpha_ {i}, c _ {i} \\right\\} _ {i = 1} ^ {N H W} = f _ {G} \\left(x _ {0} ^ {1: N}, \\mathrm {P} ^ {1: N}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 867, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12694", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our $f_{G}$ is a large reconstruction model [24, 74] which follows the implementation of GS-LRM [74]: the $N$ multiview images are concatenated with their Plücker ray coordinates computed from the camera intrinsic and extrinsic parameters for pose conditioning. Then, the inputs are patched by dividing the per-view feature map into non-overlapping patches with a patch size of $p$ . Each 2D patch is then flattened into a 1D vector. Finally, a linear layer $L$ is utilized to map the 1D vectors to image patch tokens:", + "bbox": [ + 89, + 90, + 480, + 226 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\mathrm {T} _ {j} ^ {n} \\right\\} _ {j = 1, 2, \\dots , \\frac {H W}{p ^ {2}}} = L \\left(\\text {P a t c h i f y} _ {p} \\left(\\text {C o n c a t} \\left(\\mathrm {I} ^ {n}, \\mathrm {P} ^ {n}\\right)\\right)\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 229, + 480, + 265 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Where $\\{\\mathrm{T}_j^n\\}$ denotes the set of patch tokens for image $n$ , totaling $\\frac{HW}{p^2}$ tokens per image. The set of multi-view image tokens $\\{\\mathrm{T}_j^n\\}$ are concatenated and processed through a chain of transformer blocks. Each transformer block is equipped with residual connections [20] and consists of PreLayerNorm [2], multi-head Self-Attention [59] and MLP. Later, the output tokens from the transformer are decoded into Gaussian parameters using a single linear layer. Then, the Gaussian parameters are unpatchified into $p^2$ Gaussians. Finally, we end up with $HW$ Gaussians for each view, where pixel encodes one 3D Gaussian.", + "bbox": [ + 89, + 265, + 482, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Two-stage Training. We find that training the transformer-based reconstructor solely on synthetic human head data leads to inferior texture details when applied to real-world images (see ablation study in Fig. 12). We suspect this limitation arises because the synthetic datasets lack geometric diversity. To address this, we propose a two-stage training approach in which the reconstructor is pre-trained on diverse object data [12] and subsequently fine-tuned using synthetic head data. The pre-training stage enables the reconstructor to learn a broad prior of diverse geometric structures, yielding more detailed and clearer textures in delicate facial regions such as the eyes, nose, and ears. The fine-tuning process then imparts specific knowledge of head structure, producing smoother and more realistic results. During training, we randomly select four input views to reconstruct a total of eight views, four input and four novel views. Following [74], we optimize the model using a combination of MSE and perceptual losses. During inference, the reconstructor processes the six-view outputs from multi-view diffusion model to reconstruct the head.", + "bbox": [ + 89, + 438, + 482, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Real-world Image Inference", + "text_level": 1, + "bbox": [ + 89, + 746, + 341, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For inference on real-world images, since their intrinsic parameters are unknown, we adopt a camera fov of $50^{\\circ}$ , same as during training. To ensure plausible outputs, we first apply an MTCNN face detector to estimate the face's size and center. The image is then resized and cropped/extended to match the average face size and center computed from the training data. We find this alignment compensates for the unknown intrinsic parameters well, ensuring plausible reconstruction results.", + "bbox": [ + 89, + 763, + 482, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 90, + 645, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 513, + 111, + 702, + 128 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Datasets. To quantitatively evaluate FaceLift, we establish two benchmarks for single-image to 3D full head reconstruction tasks using publicly available datasets: (1) Cafca dataset [8]: We select 40 subjects with 7 to 19 test camera poses each. Since the camera positions are randomly distributed, we manually select the most frontal view as input. Note that this synthetic dataset was independently developed and differs significantly from our training dataset. (2) Ava-256 dataset [39]: This studio-captured dataset contains real human subjects. We sample 10 subjects and 10 test camera poses for our evaluation. More details in supplemental. To demonstrate our system's generalization capabilities, we also evaluate on a set of in-the-wild face images for qualitative assessment.", + "bbox": [ + 511, + 130, + 903, + 342 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We compare our results against three state-of-the-art methods for single-face 3D reconstruction: GGHead [29], PanoHead [1], and Dual Encoder [4]. We perform GAN-inversion to reconstruct 3D head from a given input image using these models. To emphasize the importance of utilizing our synthetic human head data for training, we also compare our method with two methods that focus on general object reconstruction: Era3D [31] and LGM [56]. More comparison results with mesh-based methods are provided in the supplementary material.", + "bbox": [ + 511, + 345, + 903, + 496 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We further developed a baseline, Our $MV + LGM$ , which leverages the multi-view outputs generated by our diffusion model and employs LGM for reconstruction. This demonstrates that our method can be seamlessly integrated with other reconstruction frameworks to enhance performance on face reconstruction tasks. We tried to fine-tune the LGM reconstructor with our synthetic data, but it provides inferior results with incorrect geometry and artifacts compared with the original weights, which we suspect is due to training data mismatch (see details in the supplementary material).", + "bbox": [ + 511, + 500, + 903, + 650 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. We evaluate reconstruction quality using four standard metrics: PSNR, SSIM, LPIPS [75], and DreamSim [16]. To evaluate identity preservation, we perform face verification using ArcFace [14] through the DeepFace [52] implementation.", + "bbox": [ + 511, + 652, + 903, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. Both Cafca [8] and Ava-256 [39] datasets provide multi-view RGB images and corresponding camera poses. However, their camera system differs from the ones utilized in FaceLift and baselines. We recalculate the test camera extrinsic in each method's camera system. For a more accurate comparison, we use the Mediapipe facial landmark detection algorithm [38] to identify facial landmarks in both the target images and the rendered outputs, aligning them based on landmark distributions. Details are provided in the supplementary material.", + "bbox": [ + 511, + 733, + 903, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our system takes approximately 8 seconds to infer a", + "bbox": [ + 531, + 885, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12695", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c141da7ab0cf605f11c946a900ff61197c18b0d3d60bf0d0b00c5f332e395b2d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPSNR ↑SSIM ↑LPIPS ↓DreamSim ↓ArcFace ↓
GGHead [29]10.350.74060.36360.32520.2681
PanoHead [1]10.720.75940.33510.20480.2183
Dual Encoder [4]10.780.73850.39220.27850.2421
Era3D [31]13.690.72300.36620.28920.2978
LGM [56]16.520.79330.30600.15520.2557
Our MV + LGM [56]14.130.78120.29560.12820.1767
FaceLift16.610.79680.26940.10960.1573
", + "bbox": [ + 91, + 88, + 480, + 186 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2a40e3bf004896985025f7ab3ebdd754a8e5799ec01d59de5f2e98d3d4a12003.jpg", + "image_caption": [ + "Figure 5. Visual results on Cafca compared with face reconstruction methods. FaceLift renders novel views that closely match the ground truth, while other baselines often fail to reconstruct the 3D head in correct colors or geometry structures." + ], + "image_footnote": [], + "bbox": [ + 93, + 224, + 478, + 335 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/980876d00a5dbee340c3773abeaaa984769433a2e4fede1b5d8489c12e3c6b93.jpg", + "image_caption": [ + "Figure 6. Visual results on Cafca compared with general objects reconstruction methods. Comparison with general object reconstruction methods shows the importance of specialized data." + ], + "image_footnote": [], + "bbox": [ + 94, + 405, + 478, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3D Gaussian head from a single image: about 1.5 seconds for preprocessing (background removal, rescaling, etc.), 5.5 seconds for multi-view image generation, and under 1 second for 3D Gaussians reconstruction.", + "bbox": [ + 89, + 525, + 483, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Experiments on the Cafca Dataset", + "text_level": 1, + "bbox": [ + 89, + 595, + 390, + 611 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We report numerical comparison results on Cafca in Tab. 1. FaceLift performs favorably against baselines, especially on DreamSim [16] metric, indicating high-quality perceptual similarity. It also achieves better identity preservation performance, as demonstrated by a lower ArcFace [14] distance. We show visual results in Fig. 5 and Fig. 6. FaceLift yields rendering results that closely match the ground truth. Compared with other baselines, GGHead [29] does not support full-head rendering, resulting in unrealistic outputs when the view angle significantly deviates from the input. PanoHead [1] struggles with challenging hairstyles, while Dual Encoder [4] produces blurred facial textures. Additionally, Era3D [31] introduces artifacts on the back of the head, and LGM [56] yields inaccurate nose and jaw shapes, underscoring the importance of our synthetic human head data. When integrated with our multi-view diffusion approach, LGM achieves enhanced performance, demonstrating that our method can be seamlessly combined with existing baselines to boost their results.", + "bbox": [ + 89, + 614, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b01fc94fc3d1e3dcf390fa1da1529764e5941c6be2f5975402b944121cd547d0.jpg", + "table_caption": [ + "Table 1. Quantitative results on Cafca. FaceLift achieves favorable performance on all evaluation metrics." + ], + "table_footnote": [], + "table_body": "
MethodPSNR ↑SSIM ↑LPIPS ↓DreamSim ↓ArcFace ↓
Era3D [31]14.770.79630.25380.25150.3721
LGM [56]14.050.81360.24760.14960.3142
Our MV+LGM [56]15.240.82130.22920.10930.2264
FaceLift16.520.82710.22770.10650.1871
", + "bbox": [ + 517, + 88, + 898, + 155 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Quantitative results on Ava-256. FaceLift performs favorably than baseline methods in both reconstruction metrics and identity facial identity metric, showing a better generalization ability towards real-captured human images.", + "bbox": [ + 511, + 157, + 906, + 215 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f3b50a414451536fe0eff292687a3a9b3c3c283a3df72120ffbc5e162c16dddc.jpg", + "image_caption": [ + "Figure 7. Visual results on Ava-256. Compared with baselines, FaceLift provides multi-view renderings that are more realistic and similar to ground truth. Era3D fails to deliver delicate facial structures, while LGM generates inaccurate head shapes and colors." + ], + "image_footnote": [], + "bbox": [ + 516, + 223, + 898, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Experiments on the Ava-256 Dataset", + "text_level": 1, + "bbox": [ + 511, + 493, + 828, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We further evaluate FaceLift against other baselines on a studio-captured real human dataset, Ava-256 [50]. GAN-inversion based methods [1, 4, 29] fail to produce reasonable results with the test camera poses in this dataset, so we exclude these baselines. Tab. 2 shows that FaceLift outperforms all other baselines across all evaluation metrics, demonstrating superior reconstruction quality and identity preservation. It also highlights FaceLift's strong ability to generalize to real human faces. As shown in Fig. 7, FaceLift achieves more realistic head synthesis, while Era3D [31] struggles with accurate skin and hair textures, as well as facial details. LGM [56] produces inaccuracies in the nose shape. When combined with our multi-view diffusion model, LGM yields more accurate geometric structures, yet its texture quality remains inferior to that of FaceLift.", + "bbox": [ + 511, + 512, + 906, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Experiments with In-the-wild Images", + "text_level": 1, + "bbox": [ + 511, + 744, + 836, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We collect in-the-wild human face images and present qualitative results in comparison with other baselines in Fig. 8. Baseline methods often produce undesirable artifacts. For instance, PanoHead [1] frequently fails to render the back of the head and sometimes generates extra eyes at the rear. It also struggles to synthesize hair, shadows, wrinkles, and facial paint accurately, and its outputs lack multi-view consistency (e.g., the girl continues to face the camera in novel view 1 despite a changed pose). Dual Encoder [4] improves", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12696", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7fed4dcd366bfe9351ce23d69b27cd076b931dc2a56517fef93c57a8d9c4f698.jpg", + "image_caption": [ + "Input View" + ], + "image_footnote": [], + "bbox": [ + 109, + 112, + 171, + 157 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cc3371677dd01ccd6c488aaf6314b097a778b1fb3907678e08bd6f858ef84de6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 176, + 113, + 289, + 157 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2ce1d5f4766c43d796e58cfb5c4f54e5a98545564062e4aa29c84ecedf7e89da.jpg", + "image_caption": [ + "Rendered View 1", + "PanoHead [1] Dual Encoder [4] Era3D [31] LGM [56] Our MV+LGM [56] FaceLift" + ], + "image_footnote": [], + "bbox": [ + 292, + 112, + 316, + 157 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a0e7f84978759d7e7dda93d8bd10e6cb8e55621aaa464b9f31eed2265b5d2032.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 328, + 112, + 343, + 157 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a9891b6749d1924b63e624b85486e7626742cf5d96ff787910ef8bef17784071.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 356, + 112, + 397, + 157 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e0651557001a66bd835743f9bbd4ddfb8f2afd68bad881e500f925d7b48f052f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 112, + 473, + 157 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cb894c9741d5af94be99357a9541b1f968da9a5aab03fab507b18dfd33258d1a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 473, + 112, + 506, + 157 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/004aad661a7572a1ba60a0599505327955565695f369f588cb38de89fbf46cc8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 109, + 157, + 171, + 203 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1538e0c98c0a05b0bdfaf295aeb40c2136b90bba5c176bf32b5b44dd3ee2862a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 176, + 157, + 289, + 203 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7b75797307eb674cffb0dba25868dcec53b25dba705a523ecb371df2bda04e41.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 157, + 325, + 203 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c9cae66a05bfc0862caae8d76f273b4741c1571ddcec578f8b420e253aa3c336.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 330, + 157, + 403, + 203 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/faa508d719a709b37fb65f42a38a5291e89eccdd079921a80e09cdc43c983d1b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 411, + 157, + 444, + 203 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/08e558f12b6b11539b893d7aca98695a6cea72839651422915b8b7dda5b69330.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 473, + 157, + 504, + 203 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f2b978650ee2eb1ebc6285f69ce3f187ff16870a835fab61c172595f5cf2b234.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 109, + 203, + 171, + 247 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b071742ae771d3f28ea0fdddbf10492ff18d603aeed5115c423771bd930bba9e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 176, + 203, + 295, + 247 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f627882566c8698574d62225529f51c69520146e7f051b141ed8a86c64b0c295.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 203, + 321, + 247 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4d58b91842b19a59c910fdf5073fc6fb5516712408d9ded18859d6d00d3338fe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 330, + 203, + 403, + 247 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2e1d0837f706a9b08c4ab0f87b4538fa164fda9549434708f24f5b19da9cb0b9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 411, + 203, + 444, + 247 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3e39e3a2689d98d1769913a61af779a597f756aa387e342f041b090764edca3b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 473, + 203, + 524, + 247 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/696f4d052ce3c5f19bf95c64f001978babcabf997d82908af7b9d839f0bcd72c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 109, + 247, + 171, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d58917e4655c08349c2094b04122ac403cc69c0bbf4b8f18e63fa3d2e85ef385.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 176, + 247, + 294, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/72da2da62ae289ed3e43f91e6c5a56bcb74fc726d7d7908b51d1c70e6955595e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 247, + 321, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/05814bacd1f35b02ddfce6af1a0ba537fcf64dd5d453eedb3e16e1e85e51b66f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 330, + 251, + 401, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0212b1361b22cb910d7634c492aaae91cacf06ce4b1d613c990890104a6a9564.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 411, + 251, + 454, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/46e4388a529e6d40706c096f8d2775bd357b7a64aa279500e69af3824b9e0460.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 467, + 251, + 517, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d3cf9e3ba4093efd3b33ec928ce15c2f6468ba2b19b78fc7674286b0d64a06e7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 109, + 292, + 171, + 337 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2aa4ce74db8d2d98f6910bf0327bb6bdadcf34e8325cf758f06036be5a3b81b1.jpg", + "image_caption": [ + "Figure 8. Visual comparison on in-the-wild data. FaceLift demonstrates great generalization ability and robustness towards in-the-wild images, provides realistic unseen view rendering results. Era3D [31] and LGM [56] generate 3D head representation in inaccurate shape. PanoHead [1] often creates severe artifacts on the back of the head and can not handle challenging hairstyles well. Dual Encoder [4] shows improved performance on reconstructing the back of the head but exhibits more pronounced identity loss." + ], + "image_footnote": [], + "bbox": [ + 176, + 292, + 285, + 339 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f37bf9c3434ce7c15b1274dc1172f907759074931b2f3c942c728774f0b777e8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 292, + 321, + 337 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/172da7f7e278e703a269c702bd783b56d9abbce69276359ac717bc5daefd8ff2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 328, + 292, + 398, + 337 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d5ec3caa8da854c4b92fa715f1c32e590f1b172cd4dac059991d87e461296289.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 411, + 292, + 454, + 337 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/710e6f5cb150b5fc31ed817dc4f49d5752165dc91e4f9ee46ab0c84824167814.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 467, + 292, + 517, + 337 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/47c2863441c19e80b5aedf749d4dbabf62f9b6618d13d6cd235e39025d501fc1.jpg", + "image_caption": [ + "PanoHead [1] Dual Encoder [4] Era3D [31] LGM [56] Our MV+LGM [56] FaceLift" + ], + "image_footnote": [], + "bbox": [ + 532, + 112, + 589, + 203 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ace96edd69222586fb87f8347f0863719cb572d97b979a9f4d92b387d391f481.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 112, + 647, + 203 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b26439310673bc3889fa5791cd9d216ba2073ffd5d7890bb3200dd36a2795567.jpg", + "image_caption": [ + "Rendered View 2" + ], + "image_footnote": [], + "bbox": [ + 651, + 112, + 704, + 203 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/56e54984847429d6c95620228b5c37ce816762af2e9f9fb2b9d7c4e459b4255a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 112, + 756, + 203 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c26be89bef57287162a4867cf1d90aabc0bf520cbe59d16d5522094c264b15a3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 761, + 112, + 880, + 203 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7066f7b2c616b077c658c3772b0cc4429b459becb137ba38340d398269f1cba8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 537, + 203, + 589, + 247 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ddf09e78a09547973e498c9f6610325bd8a0f3463556bc30b65c666bf1b9e8d9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 203, + 647, + 247 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/da4ab45a31af054d60263a797f55f1fdea9cc121962a01793c2101d4bc962ddf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 203, + 694, + 247 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/30f20024d208f23e6a6810bb257058d1fe955ccdcb7682109ecc8a2e266c9613.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 203, + 754, + 247 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1737dc64f07ef3ed5c598bdab4e2485fbf36f624e45976982b9848a2b220ae4c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 203, + 815, + 247 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2a4dc6f3aed0654eb50d32d2a63dead013b0829af13487a75bc164186eeb4c62.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 833, + 203, + 872, + 247 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/de8e7144457bcc8acf398e4febcd56dd4d53b5fb644011ed9a52767fe2bf7468.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 537, + 247, + 589, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e4fb5e0209f1dfd285657ba099204ed1fa59ed6dbe12ea97d7b0454e6c0d201d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 247, + 647, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/438c22936c88de444262a16e1b303cdb8fa4e1be998c6b3046c83afa1c51252b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 247, + 694, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4a03516c85a51a9b9af704f43d7796eec20b9f79650a2d39321377486522779e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 247, + 754, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ccd0b9c4810f97b5138b72d27af938aabbf1d6f92a69aa52b3d18a674d579a1c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 247, + 815, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7c4715b6f10acfb1b97a2ed03f2b6778f70052155dab2044d364025dea44717d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 833, + 247, + 872, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/09ee325345a2110047a36baad93a88a19ec3f70b1c948050a89e9b86940e22a5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 537, + 292, + 589, + 337 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cb2cf711bd9d8917bc7a8245c3393acf14bbeb0dea8d1415ce422cceb5904123.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 292, + 647, + 337 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/561bd2eb28cc3aef32e81ff8a88b25dda3c6478c6d52168a6b247cff83374acc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 292, + 694, + 337 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9d71026ac245a36a82b48ad3d4ed357ba863c22dbcf5b8f61a8891df941b285b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 292, + 754, + 337 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/099fdb90dea4a100c4550f464f03d459e772735877fe98ca812b5033d7bda469.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 292, + 815, + 337 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3d4d5023c1837e395efcf3a5c7ac29203d5221a45df07dd4437f49bf435dcce9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 833, + 292, + 872, + 337 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e683f50c292b647db5c575035bab36fb43412c8b3f05847fbe142c4508d581b4.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 109, + 417, + 209, + 492 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/42170feec4828d1edb96b6c450a01e42b79e40ca51819a9474c34962adad3f6d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 109, + 492, + 209, + 568 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8b33b5ae6bdf1db4ce89133b415ee9846e52bdb3115a327c54d8f23012c5c99f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 109, + 568, + 209, + 641 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/819e68b1235ba896362723419c8cc731982f6d4871b4840efe150517ecd6f469.jpg", + "image_caption": [ + "Figure 9. Results of FaceLift on in-the-wild images. FaceLift accurately reconstructs 3D head models under challenging lighting conditions, achieving high fidelity (row 1). It captures fine facial details such as wrinkles (row 2), mustaches (row 3), and individual hairs (row 2 and row 4). Additionally, it remains robust to complex facial expressions (row 3) and various skin tones (row 4). Furthermore, it can realistically reconstruct facial paint (row 4). More results are provided in the supplementary materials." + ], + "image_footnote": [], + "bbox": [ + 109, + 642, + 209, + 715 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/31b9b6ac088ee01197536f8c4cf062a39dce10db1351de3dcfe95bffa6cf855c.jpg", + "image_caption": [ + "Novel View Renderings" + ], + "image_footnote": [], + "bbox": [ + 212, + 417, + 300, + 493 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7430f188b9bec33669050f87230094e1a2524890bd260facc91ac1a8fe599001.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 212, + 493, + 300, + 568 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7ff410bb307ecbddb194e78b5765e7674840f056624b9e72656264fdfc1c0126.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 212, + 569, + 300, + 641 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d39c3f9f266a4a8a7093718cb754410db28d686221570ccbfdbe815246d2024b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 212, + 642, + 300, + 715 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/34ae4dcb80d9089ae45d3bf989b1c6f0eb257eaffc1a0f56aa98aec975e8e7fb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 307, + 417, + 393, + 493 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/683198bc99409c4fcc6863731d3f51f4666df54bc70be6032f19bced133ff30d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 308, + 493, + 395, + 568 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e92968777d293309531cc31cf7abe476fe5428c6a0a55f910450a09bc4b3fc91.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 308, + 569, + 385, + 641 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f9ca2e1dcb33c7f2df878ac41e053cb67481f0cb8c7ca68938e5867573574167.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 308, + 642, + 403, + 715 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c1648b172f1ab59b1b24a7e8bd1309e1bad1de862f0ff6cd765edce6db96c411.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 417, + 496, + 492 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/43aa0531600db240ce378e530d06040e6fe0d2337b660ffd362a46f2f9199ab5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 492, + 496, + 568 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3f1f0c8c68e741e2c63a4825dad23ca25c7b49785138ffbcb1b40e4af4961102.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 568, + 496, + 641 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8941c9ea4b80dddf5f0ef6bb2ac47ae90a1f90ef6631159075c286a445eaa09d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 642, + 496, + 715 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d72b93463faa8e38d78ec8ecb6768e971317bcdd882f1650c99ce80ad3a31bb1.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 496, + 417, + 593, + 492 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9efa900195c53c3b5d44bdeb74d2d59a2d3c717b2fdb144950a482519b60d3d0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 492, + 593, + 568 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f7dbb95ae514a93278cc78a94fc10693c895095d85fe4cf11de5d27f27cd8985.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 568, + 593, + 641 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/87cde82c9a07dac1f957010a830505824604793aab71805061bbd7fb385c454d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 642, + 593, + 715 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a2b6ffcdbe0dd09b893bcacbf1d277be6ead5309d2e6bdb449e7abc7828eca15.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 604, + 426, + 678, + 493 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/717aa3c10193708c1b500a117a3960f9b232e022038a0f92b248ecff098bce3a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 601, + 493, + 686, + 568 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5af16ca856d791c86904c9e7c58edb17bf5a21a5fad59e4135d05f832ff28613.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 601, + 569, + 687, + 641 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/aea72c41915ba4a76ede0be99b6097c3d31b134ed8f9a5b9a5510324e84691f3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 601, + 642, + 687, + 715 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a9fa10bb12b4e86b4e2a5b9554652d13d4b5be78baa8a5d642861e887ded452b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 426, + 784, + 492 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c2635d201159ca66dfd55b6dca23cab5c5fd13d180414a0405816a4146bec799.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 492, + 784, + 568 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b6aa15dfd416726563ae38f384d19e01503a354be8033458091da1279eed2119.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 569, + 784, + 641 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b7c5b1a7728821d32edc5ea79be51b769a259ad9c1df9d44feed7e060c0f8498.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 642, + 782, + 715 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c84e8a014be493022565309f86b61032cc929027898b7e4ef17e288082d0c5ce.jpg", + "image_caption": [ + "Novel View Renderings" + ], + "image_footnote": [], + "bbox": [ + 784, + 417, + 882, + 492 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/85268e301c7331bb8d3cec1496bab75a3fd89334a98b307c18c8376e7f107e41.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 784, + 492, + 882, + 568 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d79dbaaa377eb9fd4682ec3754d698fc68c94dd35cec04bba5889d84a7d1e356.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 784, + 568, + 882, + 641 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/779e4cd33e63249246a59662c223cee50077538ef23cba0b3eb77f8ce929f627.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 784, + 642, + 882, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "back-of-head rendering but suffers from severe identity loss (row 2) and fails to accurately reconstruct face paint (row 4). Era3D [31] often produces an inaccurate head shape, particularly from side views, and offers fewer geometric details compared to FaceLift. LGM [56] generates Gaussians with inaccurate color and opacity and lacks proper facial geometry, resulting in distorted features. Baseline Our $MV + LGM$", + "bbox": [ + 89, + 787, + 485, + 893 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "shows that our multi-view diffusion model enhances LGM by providing improved facial geometry and texture details. However, the LGM reconstructor still produces Gaussians with inaccurate shapes and opacities, further underscoring the advantages of our transformer-based reconstructor.", + "bbox": [ + 511, + 787, + 906, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We present more FaceLift's novel view rendering results in Fig. 9 to demonstrate FaceLift's ability to produce high", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12697", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/845b35b8900bed5b992a65264b8854aebec4973ae4d4a4e5b4220c0fff50502d.jpg", + "image_caption": [ + "Input", + "w/o. Input View Reconstruction" + ], + "image_footnote": [], + "bbox": [ + 94, + 106, + 169, + 165 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/731b4add7905f0ca8004777d6d4fbb9e6dfd22b57b398b42dfd8cabbacbe1541.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 176, + 117, + 307, + 165 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6242894fee37ee8eaf33967c03081fb726dbc2b955e49d5adb5f583366a28a8e.jpg", + "image_caption": [ + "w. Input View Reconstruction", + "Figure 10. Importance of input view reconstruction. The diffusion model that is not trained to perform the input view reconstruction, i.e., w/o. Input View Reconstruction, overfits to synthetic training distribution, suffers from severe identity loss during inference. Trained with input view reconstruction, our method preserves the input identity and expression faithfully." + ], + "image_footnote": [], + "bbox": [ + 326, + 117, + 459, + 166 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "fidelity, realistic 3D head reconstructions with intricate details across a variety of challenging scenarios. FaceLift effectively handles faces under various lighting conditions. It can especially render realistic novel view images given a photo captured with an iPhone under dark lighting conditions (row 1 column 1), emphasizing its robustness and potential for real-world application. It reconstructs facial details with high fidelity, especially the wrinkles and folds on the face caused by extreme expression. FaceLift also excels at reconstructing challenging textures, such as mustaches and hair. Furthermore, it faithfully reconstructs facial paint, despite such data not being included in our synthetic face dataset, showcasing its strong generalization ability.", + "bbox": [ + 89, + 266, + 483, + 464 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 473, + 241, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1. Input View Reconstruction", + "text_level": 1, + "bbox": [ + 89, + 494, + 333, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct an ablation study to demonstrate the importance of reconstructing the input view during training. For comparison, we train a multi-view diffusion model that generates six novel views. In this baseline, the first generated view's elevation is adjusted from $0^{\\circ}$ to $20^{\\circ}$ , while the remaining views adopt the same camera poses as in our default setting. We refer to this variant as w/o. Input View Reconstruction. Fig. 10 presents the view generation results of the two diffusion models when applied to real-world images. Without the input view reconstruction task, the model trained on the synthetic dataset generates views within a limited distribution, leading to noticeable identity loss. Moreover, it loses its ability to preserve facial expressions and face paint. In contrast, incorporating the input view reconstruction task during training enables our diffusion model to faithfully regenerate the input view, significantly improving its generalization ability. Quantitative comparison is provided in the supplementary material.", + "bbox": [ + 89, + 513, + 483, + 785 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Number of Views", + "text_level": 1, + "bbox": [ + 89, + 791, + 261, + 805 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluate three configurations: four views (front, left, right, and back), six views (adding front-left and front-right), and eight views (further including front-top and front-bottom). Fig. 11 compares the baselines using different numbers of input views. With only four views, the reconstructor fails to capture a complete forehead; however,", + "bbox": [ + 89, + 810, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1451d24a8979813180aaf4d9c677cf29adfd574559f477841755d545a198f77e.jpg", + "image_caption": [ + "Input", + "Figure 11. Number of input views of Gaussian reconstructor. Using six views strike a good balance between reconstruction quality and computational efficiency." + ], + "image_footnote": [], + "bbox": [ + 517, + 102, + 573, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ebea9c0089956936d4ff1c3860e7348c481971082a3e3a3f3c2cf5ed2e74c72c.jpg", + "image_caption": [ + "4 Views" + ], + "image_footnote": [], + "bbox": [ + 576, + 102, + 627, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/086d8465aef1302223669b97f016f39874388dd9f05c1b3686b8f8464221a73d.jpg", + "image_caption": [ + "6 Views" + ], + "image_footnote": [], + "bbox": [ + 633, + 102, + 681, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2bcdc2093d6fdacb84b6595a329e3f1d4622a4ef56cbff5e94682bb68a0929b2.jpg", + "image_caption": [ + "8 Views" + ], + "image_footnote": [], + "bbox": [ + 686, + 102, + 735, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/56c6e3265cb6279cc3cbc601089dfc57cab62c215e8b192ed1061af25cd27787.jpg", + "image_caption": [ + "4 Views", + "6 Views" + ], + "image_footnote": [], + "bbox": [ + 738, + 102, + 816, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/da9d4abf0b20d79e2738c4cf89d56fb493257d806db2d8e9874b911e4fc72377.jpg", + "image_caption": [ + "8 Views" + ], + "image_footnote": [], + "bbox": [ + 820, + 102, + 870, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/be77612c1ac35f9ebb8413fe72ce03b65ab200b4772b39f31d749c25f2d3ecc8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 874, + 102, + 898, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/565043e59b9872106b383935b4a352ac74896a625eea24ce37e046517f146b6d.jpg", + "image_caption": [ + "Input", + "Figure 12. Two-stage training of reconstructor. Without pretraining on general objects, the reconstructor fails to produce clear textures in the reconstruction results. Meanwhile, without finetuning on synthetic human head data, the model lacks a refined understanding of facial structures, including the eyes and nose." + ], + "image_footnote": [], + "bbox": [ + 517, + 248, + 607, + 321 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b64979970ac3b38502e399df428f969f3c505664be9d449950e26c9af8a79618.jpg", + "image_caption": [ + "w/o. Pre-training" + ], + "image_footnote": [], + "bbox": [ + 614, + 248, + 707, + 321 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8b5535165c67f0ec8061d93797ac56f96a5d564889c666f7ad2a4bab4343df01.jpg", + "image_caption": [ + "w/o. Fine-tuning" + ], + "image_footnote": [], + "bbox": [ + 710, + 248, + 803, + 321 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d1ae0f9ad341904144bd05b4957a41a44ff91b0a7ad3e71407835d7af36011e7.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 805, + 248, + 898, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "with six views, it reconstructs the eyes and eyebrows more smoothly and renders challenging textures—such as facial wrinkles and ear folds—more realistically. Eight views do not offer significant visual improvements, and incur a higher computational cost in both stages. Thus, we find that six views achieve a good balance between reconstruction quality and computational efficiency.", + "bbox": [ + 511, + 407, + 905, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Two-stage Reconstructor Training", + "text_level": 1, + "bbox": [ + 511, + 518, + 812, + 534 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As illustrated in Sec. 3.3, our Gaussian reconstructor follows a two-stage training pipeline. Fig. 12 shows that pretraining on general objects helps the model learn a diverse prior of geometric structures, resulting in clearer textures on delicate facial regions. Meanwhile, fine-tuning on synthetic human head data enables the reconstructor to gain a more refined understanding of facial structure, thereby enhancing the accuracy of features such as the eyes, nose, and hair.", + "bbox": [ + 511, + 537, + 905, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 667, + 638, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose FaceLift, a feed-forward approach that lifts a single facial image to a detailed 3D reconstruction with preserved identity features. Our method uses multi-view diffusion to generate unobservable views and employs a transformer-based reconstructor to reconstruct 3D Gaussian splats, enabling high-quality novel view synthesis. To overcome the difficulty of capturing real-world multi-view human head images, we render high-quality synthetic data for training and show that, despite being trained solely on synthetic data, FaceLift can reconstruct 3D heads from real-world captured images with high fidelity. Compared with baselines [1, 4, 29, 31, 56], FaceLift generates 3D head representation with finer geometry and texture details and exhibits better identity preservation ability.", + "bbox": [ + 511, + 689, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12698", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This was supported in part by the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korean Government (MSIT) (No. RS-2024-00457882, National AI Research Lab Project).", + "bbox": [ + 89, + 114, + 483, + 175 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We appreciate the insightful discussions with Kai Zhang, Hao Tan, Zexiang Xu, Sai Bi, Sumit Chaturvedi, Hanwen Jiang, Yu-Ju Tsai, Kuan-Chih Huang, Chengxu Liu, and Dingyi Dai. We thank Nathan Carr and Kalyan Sunkavalli for their support.", + "bbox": [ + 89, + 176, + 483, + 252 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 265, + 187, + 281 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sizhe An, Hongyi Xu, Yichun Shi, Guoxian Song, Umit Y Ogras, and Linjie Luo. PanoHead: Geometry-aware 3D fullhead synthesis in $360^{\\circ}$ . In CVPR, 2023. 1, 3, 5, 6, 7, 8", + "[2] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 5", + "[3] Ananta R. Bhattarai, Matthias Nießner, and Artem Sevastopolsky. TriPlaneNet: An encoder for EG3D inversion. In WACV, 2024. 2", + "[4] Bahri Batuhan Bilecen, Ahmet Berke Gokmen, and Aysegul Dundar. Dual encoder GAN inversion for high-fidelity 3D head reconstruction from single images. In NeurIPS, 2024. 5, 6, 7, 8", + "[5] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3D faces. In SIGGRAPH, 1999. 2", + "[6] James Booth, Anastasios Roussos, Stefanos Zafeiriou, Allan Ponniah, and David Dunaway. A 3D morphable model learnt from 10,000 faces. In CVPR, 2016. 2", + "[7] Marcel C Bühler, Kripasindhu Sarkar, Tanmay Shah, Gengyan Li, Daoye Wang, Leonhard Helminger, Sergio Orts-Escalano, Dmitry Lagun, Otmar Hilliges, Thabo Beeler, et al. Preface: A data-driven volumetric prior for few-shot ultra high-resolution face synthesis. In ICCV, 2023. 3", + "[8] Marcel C Bühler, Gengyan Li, Erroll Wood, Leonhard Helminger, Xu Chen, Tanmay Shah, Daoye Wang, Stephan Garbin, Sergio Orts-Escolano, Otmar Hilliges, et al. Cafca: High-quality novel view synthesis of expressive faces from casual few-shot captures. In SIGGRAPH Asia, 2024. 2, 3, 5, 12, 15, 17", + "[9] Sumit Chaturvedi, Mengwei Ren, Yannick Hold-Geoffroy, Jingyuan Liu, Julie Dorsey, and Zhixin Shu. SynthLight: Portrait relighting with diffusion model by learning to re-render synthetic faces. In CVPR, 2025. 3", + "[10] Xiyi Chen, Marko Mihajlovic, Shaofei Wang, Sergey Prokudin, and Siyu Tang. Morphable diffusion: 3d-consistent diffusion for single-image avatar creation. In CVPR, 2024. 3", + "[11] Antonia Creswell and Anil Anthony Bharath. Inverting the generator of a generative adversarial network. In TNNLS, 2018. 3", + "[12] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana" + ], + "bbox": [ + 93, + 289, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3D objects. In CVPR, 2023. 2, 3, 5", + "[13] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-XL: A universe of $10\\mathrm{M} + 3\\mathrm{D}$ objects. In NeurIPS, 2024. 3", + "[14] Jiankang Deng, Jia Guo, Jing Yang, Niannan Xue, Irene Kotsia, and Stefanos Zafeiriou. ArcFace: Additive angular margin loss for deep face recognition. In IEEE TPAMI, 2022. 5, 6", + "[15] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 3", + "[16] Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. DreamSim: Learning new dimensions of human visual similarity using synthetic data. In NeurIPS, 2023. 5, 6", + "[17] Chen Gao, Yichang Shih, Wei-Sheng Lai, Chia-Kai Liang, and Jia-Bin Huang. Portrait neural radiance fields from a single image. arXiv preprint arXiv:2012.05903, 2020. 3", + "[18] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. GET3D: A generative model of high quality 3D textured shapes learned from images. In NeurIPS, 2022. 3", + "[19] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 1, 2, 3", + "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5", + "[21] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 16", + "[22] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 15", + "[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1, 3", + "[24] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: Large reconstruction model for single image to 3D. In ICLR, 2024. 3, 5", + "[25] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 1, 2, 3", + "[26] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In CVPR, 2020. 2, 3", + "[27] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D gaussian splatting for real-time radiance field rendering. In ACM TOG, 2023. 1, 3", + "[28] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In ICLR, 2014. 3", + "[29] Tobias Kirschstein, Simon Giebenhain, Jiapeng Tang, Markos Georgopoulos, and Matthias Nießner. GGHead: Fast and generalizable 3D gaussian heads. In SIGGRAPH Asia, 2024. 5, 6, 8" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12699", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3D: Fast text-to-3D with sparse-view generation and large reconstruction model. In ICLR, 2024. 3", + "[31] Peng Li, Yuan Liu, Xiaoxiao Long, Feihu Zhang, Cheng Lin, Mengfei Li, Xingqun Qi, Shanghang Zhang, Wenhan Luo, Ping Tan, et al. Era3D: High-resolution multiview diffusion using efficient row-wise attention. In NeurIPS, 2024. 3, 5, 6, 7, 8", + "[32] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. In ACM TOG, 2017. 1, 2", + "[33] Xueting Li, Shalini De Mello, Sifei Liu, Koki Nagano, Umar Iqbal, and Jan Kautz. Generalizable one-shot 3D neural head avatar. In NeurIPS, 2024. 2", + "[34] Jiangke Lin, Yi Yuan, Tianjia Shao, and Kun Zhou. Towards high-fidelity 3D face reconstruction from in-the-wild images using graph convolutional networks. In CVPR, 2020. 2", + "[35] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T., Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3D mesh in 45 seconds without pershape optimization. In NeurIPS, 2023. 3", + "[36] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In ICCV, 2023. 3", + "[37] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3D: Single image to 3D using cross-domain diffusion. In CVPR, 2024. 3", + "[38] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Guang Yong, Juhyun Lee, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, 2019. 5", + "[39] Julieta Martinez, Emily Kim, Javier Romero, Timur Bagautdinov, Shunsuke Saito, Shouou-I Yu, Stuart Anderson, Michael Zollhöfer, Te-Li Wang, Shaojie Bai, Chenghui Li, Shih-En Wei, Rohan Joshi, Wyatt Borsos, Tomas Simon, Jason Saragih, Paul Theodosis, Alexander Greene, Anjani Josyula, Silvio Mano Maeta, Andrew I. Jewett, Simon Venshtain, Christopher Heilman, Yueh-Tung Chen, Sidi Fu, Mohamed Ezzeldin A. Elshaer, Tingfang Du, Longhua Wu, Shen-Chi Chen, Kai Kang, Michael Wu, Youssef Emad, Steven Longay, Ashley Brewer, Hitesh Shah, James Booth, Taylor Koska, Kayla Hidle, Matt Andromalos, Joanna Hsu, Thomas Dauer, Peter Selednik, Tim Godisart, Scott Ardisson, Matthew Cipperly, Ben Humberston, Lon Farr, Bob Hansen, Peihong Guo, Dave Braun, Steven Krenn, He Wen, Lucas Evans, Natalia Fadeeva, Matthew Stewart, Gabriel Schwartz, Divam Gupta, Gyeongsik Moon, Kaiwen Guo, Yuan Dong, Yichen Xu, Takaaki Shiratori, Fabian Prada, Bernardo R. Pires, Bo Peng, Julia Buffalini, Autumn Trimble, Kevyn McPhail, Melissa Schoeller, and Yaser Sheikh. Codec Avatar Studio: Paired human captures for complete, driveable, and generalizable avatars. In NeurlPS, 2024. 2, 3, 5, 15, 17" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 3", + "[41] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. StyleSDF: High-resolution 3D-consistent image and geometry generation. In CVPR, 2022. 3", + "[42] Foivos Paraperas Papantoniou, Alexandros Lattas, Stylianos Moschoglou, and Stefanos Zafeiriou. Relightify: Relightable 3D faces from a single image via diffusion models. In ICCV, 2023. 3", + "[43] Dario Pavllo, Graham Spinks, Thomas Hofmann, Marie-Francine Moens, and Aurélien Lucchi. Convolutional generation of textured 3D meshes. In NeurIPS, 2020. 3", + "[44] Julius Plucker. XVii. on a new geometry of space. Philosophical Transactions of the Royal Society of London, 1865. 4", + "[45] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D diffusion. In ICLR, 2023. 3", + "[46] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3D object generation using both 2D and 3D diffusion priors. In ICLR, 2024. 3", + "[47] Elad Richardson, Matan Sela, and Ron Kimmel. 3D face reconstruction by learning from synthetic data. In 3DV, 2016. 2", + "[48] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3", + "[49] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 16", + "[50] Shunsuke Saito, Gabriel Schwartz, Tomas Simon, Junxuan Li, and Giljoo Nam. Relightable gaussian codec avatars. In CVPR, 2024. 6", + "[51] Kyle Sargent, Zizhang Li, Tanmay Shah, Charles Herrmann, Hong-Xing Yu, Yunzhi Zhang, Eric Ryan Chan, Dmitry Lagun, Li Fei-Fei, Deqing Sun, et al. ZeroNVS: Zero-shot 360-degree view synthesis from a single real image. In CVPR, 2024. 3", + "[52] Sefik Ilkin Serengil and Alper Ozpinar. HyperExtended LightFace: A facial attribute analysis framework. In ICEET, 2021. 5", + "[53] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: A single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 3", + "[54] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 3, 16", + "[55] Wenqiang Sun, Shuo Chen, Fangfu Liu, Zilong Chen, Yueqi Duan, Jun Zhang, and Yikai Wang. DimensionX: Create any 3D and 4D scenes from a single image with controllable video diffusion. arXiv preprint arXiv:2411.04928, 2024. 12" + ], + "bbox": [ + 514, + 90, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12700", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[56] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. LGM: Large multi-view gaussian model for high-resolution 3D content creation. In ECCV, 2024. 3, 5, 6, 7, 8, 17, 18", + "[57] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. DreamGaussian: Generative gaussian splatting for efficient 3D content creation. In ICLR, 2024. 3", + "[58] Aäron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning. In NeurIPS, 2017. 3", + "[59] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 3, 5", + "[60] Thomas Vetter and Volker Blanz. Estimating coloured 3D face models from single images: An example based approach. In ECCV, 1998. 1, 2", + "[61] Vishal Vinod, Tanmay Shah, and Dmitry Lagun. TEGLO: High fidelity canonical texture mapping from single-view images. In WACV, 2024. 3", + "[62] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In NeurIPS, 2021. 3", + "[63] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3D digital avatars using diffusion. In CVPR, 2023. 3", + "[64] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. ProlificDreamer: High-fidelity and diverse text-to-3D generation with variational score distillation. In NeurIPS, 2024. 3", + "[65] Erroll Wood, Tadas Baltrusaitis, Charlie Hewitt, Sebastian Dziadzio, Thomas J Cashman, and Jamie Shotton. Fake it till you make it: Face analysis in the wild using synthetic data alone. In ICCV, 2021. 3", + "[66] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3D generative-adversarial modeling. In NeurIPS, 2016. 3", + "[67] Kailu Wu, Fangfu Liu, Zhihan Cai, Runjie Yan, Hanyang Wang, Yating Hu, Yueqi Duan, and Kaiheng Ma. Unique3d: High-quality and efficient 3d mesh generation from a single image. arXiv preprint arXiv:2405.20343, 2024. 12, 13", + "[68] Yiqian Wu, Jing Zhang, Hongbo Fu, and Xiaogang Jin. LPFF: A portrait dataset for face generators across large poses. In ICCV, 2023. 1", + "[69] Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin Tong, and Jiaolong Yang. Structured 3d latents for scalable and versatile 3d generation. In CVPR, 2024. 12, 13", + "[70] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024. 12, 13" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[71] Yu-Ying Yeh, Koki Nagano, Sameh Khamis, Jan Kautz, Ming-Yu Liu, and Ting-Chun Wang. Learning to relight portrait images via a virtual light stage and synthetic-to-real adaptation. In ACM TOG, 2022. 3", + "[72] Ziyang Yuan, Yiming Zhu, Yu Li, Hongyu Liu, and Chun Yuan. Make encoder great again in 3D GAN inversion through geometry and occlusion-aware encoding. In ICCV, 2023. 1, 2, 3", + "[73] Bowen Zhang, Yiji Cheng, Chunyu Wang, Ting Zhang, Jiaolong Yang, Yansong Tang, Feng Zhao, Dong Chen, and Baining Guo. RodinHD: High-fidelity 3D avatar generation with diffusion models. In ECCV, 2024. 2, 3", + "[74] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. GS-LRM: Large reconstruction model for 3D gaussian splatting. In ECCV, 2024. 3, 5, 17", + "[75] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5", + "[76] Xiaozheng Zheng, Chao Wen, Zhaohu Li, Weiyi Zhang, Zhuo Su, Xu Chang, Yang Zhao, Zheng Lv, Xiaoyuan Zhang, Yongjie Zhang, Guidong Wang, and Xu Lan. Headgap: Few-shot 3d head avatar via generalizable gaussian priors. In 3DV, 2025. 3" + ], + "bbox": [ + 514, + 92, + 903, + 430 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "12701", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/FaceLift_ Learning Generalizable Single Image 3D Face Reconstruction from Synthetic Heads/1763108f-f851-4f26-ba5a-510a2c051a36_model.json b/2025/FaceLift_ Learning Generalizable Single Image 3D Face Reconstruction from Synthetic Heads/1763108f-f851-4f26-ba5a-510a2c051a36_model.json new file mode 100644 index 0000000000000000000000000000000000000000..52ef6b8566d9a2b2bfb8ba0bd944180b29964b5c --- /dev/null +++ b/2025/FaceLift_ Learning Generalizable Single Image 3D Face Reconstruction from Synthetic Heads/1763108f-f851-4f26-ba5a-510a2c051a36_model.json @@ -0,0 +1,3654 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.111, + 0.13, + 0.885, + 0.175 + ], + "angle": 0, + "content": "FaceLift: Learning Generalizable Single Image 3D Face Reconstruction from Synthetic Heads" + }, + { + "type": "text", + "bbox": [ + 0.249, + 0.203, + 0.748, + 0.223 + ], + "angle": 0, + "content": "Weijie Lyu\\(^{1*}\\) Yi Zhou\\(^{2}\\) Ming-Hsuan Yang\\(^{1}\\) Zhixin Shu\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.285, + 0.228, + 0.714, + 0.246 + ], + "angle": 0, + "content": "1University of California, Merced 2Adobe Research" + }, + { + "type": "image", + "bbox": [ + 0.103, + 0.259, + 0.252, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.379, + 0.234, + 0.392 + ], + "angle": 0, + "content": "Single Image Input" + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.259, + 0.419, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.269, + 0.379, + 0.42, + 0.403 + ], + "angle": 0, + "content": "High-fidelity 3D Gaussian Representation" + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.259, + 0.568, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.454, + 0.379, + 0.541, + 0.391 + ], + "angle": 0, + "content": "Full Head NVS" + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.259, + 0.733, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.659, + 0.379, + 0.778, + 0.392 + ], + "angle": 0, + "content": "High-quality Details" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.412, + 0.907, + 0.469 + ], + "angle": 0, + "content": "Figure 1. FaceLift transforms a single facial image into a high-fidelity 3D Gaussian head representation. Trained exclusively on synthetic 3D data, our pipeline first generates sparse, identity-preserving multiview images of the input head using a diffusion model. These sparse generated views are then fed into a transformer-based 3D Gaussian splats reconstructor, producing complete and detailed 3D head representation that generalize remarkably well to real-world human images. Project page: https://weijielyu.github.io/FaceLift." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.481, + 0.327, + 0.497 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.517, + 0.486, + 0.851 + ], + "angle": 0, + "content": "We present FaceLift, a novel feed-forward approach for generalizable high-quality 360-degree 3D head reconstruction from a single image. Our pipeline first employs a multiview latent diffusion model to generate consistent side and back views from a single facial input, which then feed into a transformer-based reconstructor that produces a comprehensive 3D Gaussian splats representation. Previous methods for monocular 3D face reconstruction often lack full view coverage or view consistency due to insufficient multi-view supervision. We address this by creating a high-quality synthetic head dataset that enables consistent supervision across viewpoints. To bridge the domain gap between synthetic training data and real-world images, we propose a simple yet effective technique that ensures the view generation process maintains fidelity to the input by learning to reconstruct the input image alongside the view generation. Despite being trained exclusively on synthetic data, our method demonstrates remarkable generalization to real-world images. Through extensive qualitative and quantitative evaluations, we show that FaceLift outperforms state-of-the-art 3D face reconstruction methods on identity preservation, detail recovery and rendering quality." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.48, + 0.645, + 0.496 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.506, + 0.907, + 0.687 + ], + "angle": 0, + "content": "3D face reconstruction has been a central focus in computer vision and graphics research for decades, driven by its crucial applications in immersive virtual and augmented realities, VFX and gaming, digital entertainment, and next-generation telepresence systems. However, achieving high quality reconstruction from a single image remains very challenging. On one hand, the monocular face reconstruction problem is highly ill-posed – a single 2D image can be produced by countless different 3D face shapes, creating fundamental ambiguity. On the other hand, the human visual system is highly attuned to facial details, making even subtle artifacts and imperfections noticeable to the eye." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Traditional 3D head synthesis approaches typically use parametric textured mesh models [32, 60] trained on 3D scan datasets. While these models enable basic head generation, the rendered images frequently lack fine-scale geometric detail, realistic textures, and convincing hair, limiting their perceptual realism and expressiveness. Recent breakthroughs in image generative models [19, 23] and novel view synthesis techniques [27, 40] have opened new possibilities for this research area. Leveraging these developments, recent works [1, 72] use neural 3D representations to learn effective 3D head representation from unstructured real face image datasets [25, 68]. While these datasets improve the realism and diversity of rendering results, they fail to provide multi-view supervision for modeling 3D consis" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.887, + 0.478, + 0.9 + ], + "angle": 0, + "content": "*Work was done when Weijie Lyu was an intern at Adobe Research." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12691" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.48, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.341, + 0.483, + 0.397 + ], + "angle": 0, + "content": "Figure 2. Comparison with RodinHD. RodinHD [73] trains triplane diffusion with synthetic data, resulting in apparent identity loss. In contrast, FaceLift achieves better identity preservation and generalizes effectively to real human portraits." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.402, + 0.484, + 0.508 + ], + "angle": 0, + "content": "tency causing view inconsistency and artifacts on the back of the head. Since diverse multi-view real images are hard to acquire, RodinHD [73] leverages synthetic multi-view images to train generative models that directly output 3D neural representations of the head. However, training solely on synthetic data often results in significant perceptual identity loss in the generated outputs, as demonstrated in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.508, + 0.484, + 0.87 + ], + "angle": 0, + "content": "In this work, we present FaceLift, a pipeline for learning generalizable and high-fidelity single image to 3D face representation from synthetic head data. To achieve high quality reconstruction that preserves the input facial identity, we adopt a two-stage pipeline to first generate identity preserving multi-view images using a diffusion model [48], followed by a transformer-based feed-forward reconstructor to fuse the generated sparse views into a comprehensive 3D Gaussian representation. We train the model with synthetic images - multi-view renderings of 3D synthetic human portraits using Blender. We highlight two key techniques for generalizing to real-world images and preserving input facial identity. First, we emphasize the importance of reconstructing the input image alongside the view synthesis task in the conditional diffusion model training, which significantly improved generalization capability in testing. Second, we demonstrate that training the feed-forward reconstructor benefits from a two-stage training process: pretraining on general objects [12] to acquire a rich geometry and texture prior, followed by fine-tuning on synthetic human head data to capture head-specific geometry. With our two-stage approach, we focus on learning identity preservation in the image space during the first stage, achieving higher input fidelity compared to existing methods." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Comparing with prior art, we achieve three key advancements: (1) robust view consistency through multi-view at" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "tention and supervision, (2) improved generalization from training techniques and foundational model, ensuring accurate identity preservation, (3) high-quality facial texture and hair details via pixel-aligned Gaussian representation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.153, + 0.907, + 0.333 + ], + "angle": 0, + "content": "We extensively evaluate FaceLift quantitatively and qualitatively across diverse datasets. Using real multiview studio captures [39] and an independent synthetic human dataset [8], our approach consistently surpasses previous state-of-the-art methods across all evaluation metrics. Through extensive testing on in-the-wild portrait images, we demonstrate that our method reconstructs complete 3D heads with fine-grained details, accurate identity preservation, and high visual fidelity. Comparisons and ablation studies confirm that multi-view consistent training data is crucial for high-fidelity face reconstruction. Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.336, + 0.905, + 0.379 + ], + "angle": 0, + "content": "- We propose FaceLift, a framework that reconstructs a high-fidelity 3D head from a single image using view generation and feed-forward reconstructor." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.381, + 0.905, + 0.44 + ], + "angle": 0, + "content": "- Despite being trained solely on synthetic human head data, our approach shows no domain gap on real-world images, highlighting both the effectiveness of synthetic data and our model's robust generalization capabilities." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.442, + 0.905, + 0.516 + ], + "angle": 0, + "content": "- We construct two benchmarks on single-image to 3D full head reconstruction tasks using the publicly available datasets Cafca [8] and Ava-256 [39] to quantitatively evaluate models' performance on both reconstruction accuracy and identity preservation ability." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.517, + 0.905, + 0.562 + ], + "angle": 0, + "content": "- Our comprehensive evaluation confirms that our approach achieves state-of-the-art performance in reconstruction accuracy and identity preservation." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.336, + 0.905, + 0.562 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.576, + 0.655, + 0.591 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Face Reconstruction. 3D face reconstruction has been a long-standing challenge in computer vision, with substantial progress driven by diverse approaches. Vetter and Blanz [60] pioneer a method for synthesizing 3D faces by linearly blending multiple 3D templates, now widely known as blendshapes. This work establishes the foundation for 3D Morphable Models (3DMMs), which represent 3D face shapes and textures as principal components derived from scanned data. Subsequent research [5, 6, 32, 34, 47] extend this framework, enabling the generation of new 3D faces by manipulating blending coefficients. However, these methods produce mesh-based representations that lack fine details and are limited to modeling the front of the face, excluding hair and 360-degree synthesis. While 3DMM-based methods have been foundational, recent advances in deep learning, especially Generative Adversarial Networks (GANs) [19, 25, 26], have greatly improved 3D face synthesis quality. EG3D [72] uses a tri-plane NeRF representation with a pose-conditioned StyleGAN2 [26] framework. Follow-up works [3, 33] achieve single-image-to-3D gen" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12692" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.438 + ], + "angle": 0, + "content": "eration through GAN inversion [11]. Despite their success, these methods can only synthesize near-frontal views. To overcome this, PanoHead [1] introduces a tri-grid neural volume representation, enabling full 360-degree head synthesis. Unfortunately, it does not provide a 3D head representation for consistent multi-view rendering. Recent efforts explore alternative representations for 3D face reconstruction from sparse input, such as a single image [17, 42, 61] or few-shot images [7]. However, these methods still require pre-instance optimization. Rodin [63] and its extension RodinHD [73] employ an image-conditioned diffusion model to generate a triplane representation of a human head for full-head novel view synthesis. Nevertheless, their triplane diffusion model is limited to synthetic data and struggles to achieve high-fidelity reconstructions from real-world images. For animatable 3D head avatars generations, Morphable Diffusion [10] generates multi-view consistent images from a single image using a morphable mesh, while HeadGAP [76] generates 3D animatable head avatars using few-shot input, leveraging 3D head priors learned from large-scale data. In contrast, our work focuses on leveraging synthetic training data to produce high-fidelity, detailed 3D Gaussian head models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.444, + 0.482, + 0.625 + ], + "angle": 0, + "content": "Synthetic Human Data. Capturing high-quality 3D data of real humans requires a controlled studio environment and costly photography equipment [39]. As an alternative, large-scale synthetic 3D head datasets have emerged as an effective and resource-efficient solution for tasks like human head reconstruction [8, 63, 65, 73] and photorealistic relighting [9, 71], offering a scalable way to train models without the restrictions of real-world data acquisition. Inspired by these previous works, we aim to use synthetic data to improve the model's understanding of the human head and minimize the generalization gap between synthetic data training and real-world inference." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Image or Text to 3D. Generative models have achieved remarkable success in 2D image generation with VAEs [28, 58], GANs [19, 25, 26], and diffusion models [23, 48, 54]. Building on this success, extensive research has extended these models to 3D content generation [18, 41, 43, 66]. Starting with DreamFusion [45], numerous works [36, 46, 51, 57, 64] try to distill NeRF [40, 62, 72] or 3D Gaussians [27] representation from 2D image diffusion using a Score Distillation Sampling (SDS) loss. These methods can produce high-quality results but often encounter challenges such as slow optimization, over-saturated colors, and the Janus problem. To overcome these challenges, recent works [30, 31, 35, 37, 53] generate multi-view images with high consistency, which can be directly used for 3D reconstruction with neural reconstruction methods [27, 40, 62]. However, optimizing NeRF or NeuS remains far from real-time performance. Recent advancements in large 3D reconstruction models (LRMs) [24, 30, 56, 74] offer a path" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.243 + ], + "angle": 0, + "content": "way to faster 3D reconstruction. Leveraging scalable transformer architectures [15, 59] and large datasets like Objverse [12, 13], these models effectively capture generalizable 3D priors. Unlike traditional pre-scene optimization methods [27, 40, 62], LRMs employ a feed-forward approach, enabling the prediction of high-quality NeRF, mesh, or 3D Gaussian representations from sparse images in under a second. However, most of these research efforts are applied to general objects, with limited or suboptimal results for 3D head reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.252, + 0.687, + 0.269 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.274, + 0.905, + 0.621 + ], + "angle": 0, + "content": "As shown in Fig. 3, given a single frontal image of a human face \\( y \\), our goal is to reconstruct a complete 3D head \\( G \\), represented as 3D Gaussian splats, with detailed facial texture and preserved identity. This requires our system to have prior knowledge on the geometry structure of a human face and the ability to synthesis plausible details which are not visible in the input view. Hence, we train a multiview diffusion model \\( f_{D} \\) on synthetic human head data to generate \\( N \\) views \\( x_0^1, x_0^2, \\ldots, x_0^N \\) covering \\( 360^\\circ \\) of the human head while achieving multi-view consistency and preserving identity. We choose pixel-aligned 3D Gaussians to obtain the final 3D representation. Compared to NeRFs and meshes, 3D Gaussians offer explicit volumetric primitives that better capture subtle facial geometry and fine details. Their semi-transparent kernels naturally model effects like wispy hair and translucency, which are challenging for discrete surfaces or density fields. These generated views \\( x_0^{1:N} \\) from the diffusion model, along with their corresponding Plücker ray coordinates \\( \\mathbf{P}^{(1:N)} \\), are input into a transformer-based Gaussian reconstructor \\( f_R \\) to predict a set of 3D Gaussians \\( G \\). Training of the Gaussian reconstructor follows a pre-training process on general objects [12] and a fine-tuning process on synthetic human head data." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.626, + 0.792, + 0.641 + ], + "angle": 0, + "content": "3.1. Synthetic Human Head Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.905, + 0.84 + ], + "angle": 0, + "content": "We implement a 3D head asset generation pipeline inspired by [65]. Our process begins with a collection of high-quality, artist-created 3D head meshes, which we enhance by incorporating detailed facial components, including eyes, teeth, gums, and both facial and scalp hair. We then augment these base models through rigging for pose variation and blendshape deformation for diverse facial expressions. The final head models are enriched with a set of PBR texture maps, including albedo, normal, roughness, specular, and subsurface scattering maps. At last, we dress the head model with a collection of clothing assets. The entire pipeline is implemented in Blender and the images are rendered with Cycles renderer." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.905, + 0.901 + ], + "angle": 0, + "content": "To train our networks, we render images (samples shown in Fig. 4) at \\(512 \\times 512\\) resolution from 200 unique identities, each with 50 varied appearances, including different hairstyles, skin tones, expressions, clothes, poses, etc." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12693" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.089, + 0.885, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.223, + 0.907, + 0.294 + ], + "angle": 0, + "content": "Figure 3. Overview of FaceLift. Given a single image of a human face \\( y \\) as input, we train an image-conditioned, multi-view diffusion model to generate novel views \\( x_0^1, \\ldots, x_0^N \\) covering the entire head. By generating \\( x_0^1 \\) the same as \\( y \\) and leveraging the high-quality synthetic data, our multi-view latent diffusion model can hallucinate unseen views of the human head with high-fidelity and multi-view consistency. We then train a transformer-based reconstructor \\( f_R \\), which takes multi-view images \\( x_0^{1:N} \\) and their camera poses \\( \\mathbf{P}^{1:N} \\) as input and generates 3D Gaussian splats \\( G \\) to represent the human head." + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.306, + 0.48, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.406, + 0.483, + 0.448 + ], + "angle": 0, + "content": "Figure 4. Synthetic data examples. Top row: six views for diffusion training. Bottom row: samples of random views for reconstructor training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.458, + 0.484, + 0.549 + ], + "angle": 0, + "content": "We render our training dataset under two types of lighting conditions: (1) ambient light and (2) random HDR environment light. We render six views for each subject to train the multi-view diffusion model. For fine-tuning the transformer-based reconstructor, we render 32 views with random camera poses." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.553, + 0.258, + 0.567 + ], + "angle": 0, + "content": "3.2. View Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.571, + 0.484, + 0.661 + ], + "angle": 0, + "content": "We model the sparse view generation from a single image input as a conditional diffusion process. We use a multiview diffusion model \\( f_{D} \\) to generate \\( N \\) views, denoted as \\( X_0^1, X_0^2, \\ldots, X_0^N \\), given a single front-facing image \\( y \\) and CLIP text embeddings \\( e^1, e^2, \\ldots, e^N \\) corresponding to each generated view. This process is expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.118, + 0.662, + 0.483, + 0.681 + ], + "angle": 0, + "content": "\\[\n\\left\\{X _ {0} ^ {1}, X _ {0} ^ {2}, \\dots , X _ {0} ^ {N} \\right\\} = f _ {D} \\left(y, \\left\\{e ^ {1}, e ^ {2}, \\dots , e ^ {N} \\right\\}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.732 + ], + "angle": 0, + "content": "We aim to learn the joint distribution of all these views, conditioning on the input image \\(y\\) and text embedding \\(e^1, e^2, \\ldots, e^N\\). We denote the joint distribution as:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.733, + 0.486, + 0.765 + ], + "angle": 0, + "content": "\\[\np _ {f _ {D}} \\left(x _ {0} ^ {1: N} \\mid y, e ^ {1: N}\\right) := p _ {f _ {D}} \\left(\\left\\{x _ {0} ^ {1}, \\dots x _ {0} ^ {N} \\right\\} \\mid y, \\left\\{e ^ {1}, \\dots e ^ {N} \\right\\}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.902 + ], + "angle": 0, + "content": "View Selection. Given a single near frontal view face image with azimuth \\(\\alpha\\), our multi-view diffusion model generates six views with azimuths equal to \\(\\{\\alpha, \\alpha \\pm 45^\\circ, \\alpha \\pm 90^\\circ, \\alpha + 180^\\circ\\}\\), covering 360 degrees of the human head. Elevation is 0 for all images. We opt for six views as the optimal balance - fewer views compromise detail quality while more views become computationally prohibitive for full head reconstruction. An ablation study comparing four, six, and eight views is presented in Sec. 5.2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.305, + 0.907, + 0.606 + ], + "angle": 0, + "content": "Multi-view Attention. To ensure the consistency of the generated novel views, we use a multi-view attention mechanism to facilitate information propagation and implicitly encode multi-view dependencies. Our attention module encourages multi-view consistency by extending the 2D self-attention mechanism to 3D and enabling interactions across views. Instead of treating each view independently, we apply self-attention across all views simultaneously, allowing information to be shared between them. Specifically, we start with an input tensor of shape \\( B \\times V \\times H \\times W \\times C \\), where \\( B \\) is the batch size, \\( V \\) is the number of views, \\( H \\) and \\( W \\) denote the spatial resolution of the intermediate feature maps, and \\( C \\) is the number of feature channels. We reshape this tensor to \\( B \\times VHW \\times C \\), treating all spatial locations across all views as a unified sequence of tokens for self-attention. This design allows the model to learn multi-view correlations by sharing information across views within the attention layers, enabling it to generate consistent RGB images. We provide an ablation study on the multi-view attention mechanism in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.607, + 0.906, + 0.743 + ], + "angle": 0, + "content": "Input View Reconstruction. During Training, we enforce the first generated view to share the same camera with the input image. In other words, we reconstruct the input view in the view generation process. We find this approach, combined with the multi-view attention mechanism, significantly outperforms the alternative strategy of generating only novel views, which tends to overfit to synthetic training identities and compromises generalization capability as we will show in Sec. 5.1 and the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.747, + 0.883, + 0.761 + ], + "angle": 0, + "content": "3.3. Multi-view to 3D Gaussians Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.869 + ], + "angle": 0, + "content": "Transformer-based Reconstructor. We choose pixel-aligned 3D Gaussians as the final 3D representation. Each Gaussian \\(G_{i}\\) is defined by position \\(p_i\\), scale \\(s_i\\), orientation \\(q_{i}\\), opacity \\(\\alpha_{i}\\), and color features \\(c_{i}\\). We use a transformer-based reconstructor \\(f_{G}\\) to obtain 3D Gaussians from generated multi-view images \\(x_0^{1:N}\\) and their corresponding Plücker ray coordinates [44] \\(\\mathbf{P}^{1:N}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.868, + 0.905, + 0.901 + ], + "angle": 0, + "content": "\\[\n\\left\\{G _ {i} \\right\\} _ {i = 1} ^ {N H W} = \\left\\{p _ {i}, s _ {i}, q _ {i}, \\alpha_ {i}, c _ {i} \\right\\} _ {i = 1} ^ {N H W} = f _ {G} \\left(x _ {0} ^ {1: N}, \\mathrm {P} ^ {1: N}\\right), \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12694" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.227 + ], + "angle": 0, + "content": "Our \\( f_{G} \\) is a large reconstruction model [24, 74] which follows the implementation of GS-LRM [74]: the \\( N \\) multiview images are concatenated with their Plücker ray coordinates computed from the camera intrinsic and extrinsic parameters for pose conditioning. Then, the inputs are patched by dividing the per-view feature map into non-overlapping patches with a patch size of \\( p \\). Each 2D patch is then flattened into a 1D vector. Finally, a linear layer \\( L \\) is utilized to map the 1D vectors to image patch tokens:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.231, + 0.482, + 0.266 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\mathrm {T} _ {j} ^ {n} \\right\\} _ {j = 1, 2, \\dots , \\frac {H W}{p ^ {2}}} = L \\left(\\text {P a t c h i f y} _ {p} \\left(\\text {C o n c a t} \\left(\\mathrm {I} ^ {n}, \\mathrm {P} ^ {n}\\right)\\right)\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.266, + 0.483, + 0.435 + ], + "angle": 0, + "content": "Where \\(\\{\\mathrm{T}_j^n\\}\\) denotes the set of patch tokens for image \\(n\\), totaling \\(\\frac{HW}{p^2}\\) tokens per image. The set of multi-view image tokens \\(\\{\\mathrm{T}_j^n\\}\\) are concatenated and processed through a chain of transformer blocks. Each transformer block is equipped with residual connections [20] and consists of PreLayerNorm [2], multi-head Self-Attention [59] and MLP. Later, the output tokens from the transformer are decoded into Gaussian parameters using a single linear layer. Then, the Gaussian parameters are unpatchified into \\(p^2\\) Gaussians. Finally, we end up with \\(HW\\) Gaussians for each view, where pixel encodes one 3D Gaussian." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.439, + 0.483, + 0.741 + ], + "angle": 0, + "content": "Two-stage Training. We find that training the transformer-based reconstructor solely on synthetic human head data leads to inferior texture details when applied to real-world images (see ablation study in Fig. 12). We suspect this limitation arises because the synthetic datasets lack geometric diversity. To address this, we propose a two-stage training approach in which the reconstructor is pre-trained on diverse object data [12] and subsequently fine-tuned using synthetic head data. The pre-training stage enables the reconstructor to learn a broad prior of diverse geometric structures, yielding more detailed and clearer textures in delicate facial regions such as the eyes, nose, and ears. The fine-tuning process then imparts specific knowledge of head structure, producing smoother and more realistic results. During training, we randomly select four input views to reconstruct a total of eight views, four input and four novel views. Following [74], we optimize the model using a combination of MSE and perceptual losses. During inference, the reconstructor processes the six-view outputs from multi-view diffusion model to reconstruct the head." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.747, + 0.342, + 0.763 + ], + "angle": 0, + "content": "3.4. Real-world Image Inference" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.9 + ], + "angle": 0, + "content": "For inference on real-world images, since their intrinsic parameters are unknown, we adopt a camera fov of \\(50^{\\circ}\\), same as during training. To ensure plausible outputs, we first apply an MTCNN face detector to estimate the face's size and center. The image is then resized and cropped/extended to match the average face size and center computed from the training data. We find this alignment compensates for the unknown intrinsic parameters well, ensuring plausible reconstruction results." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.646, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.112, + 0.703, + 0.129 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.131, + 0.905, + 0.343 + ], + "angle": 0, + "content": "Evaluation Datasets. To quantitatively evaluate FaceLift, we establish two benchmarks for single-image to 3D full head reconstruction tasks using publicly available datasets: (1) Cafca dataset [8]: We select 40 subjects with 7 to 19 test camera poses each. Since the camera positions are randomly distributed, we manually select the most frontal view as input. Note that this synthetic dataset was independently developed and differs significantly from our training dataset. (2) Ava-256 dataset [39]: This studio-captured dataset contains real human subjects. We sample 10 subjects and 10 test camera poses for our evaluation. More details in supplemental. To demonstrate our system's generalization capabilities, we also evaluate on a set of in-the-wild face images for qualitative assessment." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.347, + 0.905, + 0.497 + ], + "angle": 0, + "content": "Baselines. We compare our results against three state-of-the-art methods for single-face 3D reconstruction: GGHead [29], PanoHead [1], and Dual Encoder [4]. We perform GAN-inversion to reconstruct 3D head from a given input image using these models. To emphasize the importance of utilizing our synthetic human head data for training, we also compare our method with two methods that focus on general object reconstruction: Era3D [31] and LGM [56]. More comparison results with mesh-based methods are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.5, + 0.905, + 0.651 + ], + "angle": 0, + "content": "We further developed a baseline, Our \\(MV + LGM\\), which leverages the multi-view outputs generated by our diffusion model and employs LGM for reconstruction. This demonstrates that our method can be seamlessly integrated with other reconstruction frameworks to enhance performance on face reconstruction tasks. We tried to fine-tune the LGM reconstructor with our synthetic data, but it provides inferior results with incorrect geometry and artifacts compared with the original weights, which we suspect is due to training data mismatch (see details in the supplementary material)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.654, + 0.905, + 0.73 + ], + "angle": 0, + "content": "Evaluation Metrics. We evaluate reconstruction quality using four standard metrics: PSNR, SSIM, LPIPS [75], and DreamSim [16]. To evaluate identity preservation, we perform face verification using ArcFace [14] through the DeepFace [52] implementation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.734, + 0.905, + 0.884 + ], + "angle": 0, + "content": "Implementation Details. Both Cafca [8] and Ava-256 [39] datasets provide multi-view RGB images and corresponding camera poses. However, their camera system differs from the ones utilized in FaceLift and baselines. We recalculate the test camera extrinsic in each method's camera system. For a more accurate comparison, we use the Mediapipe facial landmark detection algorithm [38] to identify facial landmarks in both the target images and the rendered outputs, aligning them based on landmark distributions. Details are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Our system takes approximately 8 seconds to infer a" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12695" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.482, + 0.187 + ], + "angle": 0, + "content": "
MethodPSNR ↑SSIM ↑LPIPS ↓DreamSim ↓ArcFace ↓
GGHead [29]10.350.74060.36360.32520.2681
PanoHead [1]10.720.75940.33510.20480.2183
Dual Encoder [4]10.780.73850.39220.27850.2421
Era3D [31]13.690.72300.36620.28920.2978
LGM [56]16.520.79330.30600.15520.2557
Our MV + LGM [56]14.130.78120.29560.12820.1767
FaceLift16.610.79680.26940.10960.1573
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.19, + 0.483, + 0.218 + ], + "angle": 0, + "content": "Table 1. Quantitative results on Cafca. FaceLift achieves favorable performance on all evaluation metrics." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.226, + 0.479, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.341, + 0.483, + 0.397 + ], + "angle": 0, + "content": "Figure 5. Visual results on Cafca compared with face reconstruction methods. FaceLift renders novel views that closely match the ground truth, while other baselines often fail to reconstruct the 3D head in correct colors or geometry structures." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.406, + 0.479, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.47, + 0.483, + 0.513 + ], + "angle": 0, + "content": "Figure 6. Visual results on Cafca compared with general objects reconstruction methods. Comparison with general object reconstruction methods shows the importance of specialized data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.526, + 0.484, + 0.587 + ], + "angle": 0, + "content": "3D Gaussian head from a single image: about 1.5 seconds for preprocessing (background removal, rescaling, etc.), 5.5 seconds for multi-view image generation, and under 1 second for 3D Gaussians reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.596, + 0.391, + 0.612 + ], + "angle": 0, + "content": "4.2. Experiments on the Cafca Dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.615, + 0.484, + 0.903 + ], + "angle": 0, + "content": "We report numerical comparison results on Cafca in Tab. 1. FaceLift performs favorably against baselines, especially on DreamSim [16] metric, indicating high-quality perceptual similarity. It also achieves better identity preservation performance, as demonstrated by a lower ArcFace [14] distance. We show visual results in Fig. 5 and Fig. 6. FaceLift yields rendering results that closely match the ground truth. Compared with other baselines, GGHead [29] does not support full-head rendering, resulting in unrealistic outputs when the view angle significantly deviates from the input. PanoHead [1] struggles with challenging hairstyles, while Dual Encoder [4] produces blurred facial textures. Additionally, Era3D [31] introduces artifacts on the back of the head, and LGM [56] yields inaccurate nose and jaw shapes, underscoring the importance of our synthetic human head data. When integrated with our multi-view diffusion approach, LGM achieves enhanced performance, demonstrating that our method can be seamlessly combined with existing baselines to boost their results." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.9, + 0.156 + ], + "angle": 0, + "content": "
MethodPSNR ↑SSIM ↑LPIPS ↓DreamSim ↓ArcFace ↓
Era3D [31]14.770.79630.25380.25150.3721
LGM [56]14.050.81360.24760.14960.3142
Our MV+LGM [56]15.240.82130.22920.10930.2264
FaceLift16.520.82710.22770.10650.1871
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.159, + 0.907, + 0.216 + ], + "angle": 0, + "content": "Table 2. Quantitative results on Ava-256. FaceLift performs favorably than baseline methods in both reconstruction metrics and identity facial identity metric, showing a better generalization ability towards real-captured human images." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.224, + 0.899, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.425, + 0.907, + 0.482 + ], + "angle": 0, + "content": "Figure 7. Visual results on Ava-256. Compared with baselines, FaceLift provides multi-view renderings that are more realistic and similar to ground truth. Era3D fails to deliver delicate facial structures, while LGM generates inaccurate head shapes and colors." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.494, + 0.829, + 0.51 + ], + "angle": 0, + "content": "4.3. Experiments on the Ava-256 Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.513, + 0.907, + 0.74 + ], + "angle": 0, + "content": "We further evaluate FaceLift against other baselines on a studio-captured real human dataset, Ava-256 [50]. GAN-inversion based methods [1, 4, 29] fail to produce reasonable results with the test camera poses in this dataset, so we exclude these baselines. Tab. 2 shows that FaceLift outperforms all other baselines across all evaluation metrics, demonstrating superior reconstruction quality and identity preservation. It also highlights FaceLift's strong ability to generalize to real human faces. As shown in Fig. 7, FaceLift achieves more realistic head synthesis, while Era3D [31] struggles with accurate skin and hair textures, as well as facial details. LGM [56] produces inaccuracies in the nose shape. When combined with our multi-view diffusion model, LGM yields more accurate geometric structures, yet its texture quality remains inferior to that of FaceLift." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.746, + 0.837, + 0.763 + ], + "angle": 0, + "content": "4.4. Experiments with In-the-wild Images" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We collect in-the-wild human face images and present qualitative results in comparison with other baselines in Fig. 8. Baseline methods often produce undesirable artifacts. For instance, PanoHead [1] frequently fails to render the back of the head and sometimes generates extra eyes at the rear. It also struggles to synthesize hair, shadows, wrinkles, and facial paint accurately, and its outputs lack multi-view consistency (e.g., the girl continues to face the camera in novel view 1 despite a changed pose). Dual Encoder [4] improves" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12696" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.09, + 0.173, + 0.102 + ], + "angle": 0, + "content": "Input View" + }, + { + "type": "image_caption", + "bbox": [ + 0.313, + 0.092, + 0.393, + 0.102 + ], + "angle": 0, + "content": "Rendered View 1" + }, + { + "type": "image_caption", + "bbox": [ + 0.179, + 0.104, + 0.516, + 0.113 + ], + "angle": 0, + "content": "PanoHead [1] Dual Encoder [4] Era3D [31] LGM [56] Our MV+LGM [56] FaceLift" + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.113, + 0.172, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.178, + 0.114, + 0.29, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.113, + 0.317, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.329, + 0.113, + 0.344, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.357, + 0.113, + 0.398, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.113, + 0.474, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.475, + 0.113, + 0.507, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.158, + 0.172, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.178, + 0.158, + 0.29, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.158, + 0.326, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.159, + 0.404, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.413, + 0.159, + 0.445, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.474, + 0.159, + 0.506, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.204, + 0.172, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.178, + 0.204, + 0.296, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.204, + 0.322, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.204, + 0.404, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.413, + 0.204, + 0.445, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.474, + 0.204, + 0.526, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.248, + 0.172, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.178, + 0.248, + 0.295, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.248, + 0.322, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.252, + 0.403, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.412, + 0.252, + 0.455, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.252, + 0.519, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.293, + 0.172, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.178, + 0.293, + 0.287, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.293, + 0.322, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.33, + 0.293, + 0.4, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.412, + 0.293, + 0.455, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.293, + 0.519, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.667, + 0.092, + 0.747, + 0.102 + ], + "angle": 0, + "content": "Rendered View 2" + }, + { + "type": "image_caption", + "bbox": [ + 0.535, + 0.104, + 0.869, + 0.113 + ], + "angle": 0, + "content": "PanoHead [1] Dual Encoder [4] Era3D [31] LGM [56] Our MV+LGM [56] FaceLift" + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.113, + 0.59, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.592, + 0.113, + 0.648, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.113, + 0.705, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.113, + 0.758, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.762, + 0.113, + 0.882, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.204, + 0.591, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.204, + 0.648, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.204, + 0.696, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.204, + 0.756, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.204, + 0.816, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.834, + 0.204, + 0.874, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.248, + 0.591, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.248, + 0.648, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.248, + 0.696, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.248, + 0.756, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.248, + 0.816, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.834, + 0.248, + 0.874, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.293, + 0.591, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.293, + 0.648, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.293, + 0.696, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.293, + 0.756, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.293, + 0.816, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.834, + 0.293, + 0.874, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.343, + 0.907, + 0.399 + ], + "angle": 0, + "content": "Figure 8. Visual comparison on in-the-wild data. FaceLift demonstrates great generalization ability and robustness towards in-the-wild images, provides realistic unseen view rendering results. Era3D [31] and LGM [56] generate 3D head representation in inaccurate shape. PanoHead [1] often creates severe artifacts on the back of the head and can not handle challenging hairstyles well. Dual Encoder [4] shows improved performance on reconstructing the back of the head but exhibits more pronounced identity loss." + }, + { + "type": "image_caption", + "bbox": [ + 0.143, + 0.403, + 0.18, + 0.417 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.418, + 0.21, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.493, + 0.21, + 0.569 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.569, + 0.21, + 0.642 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.643, + 0.21, + 0.716 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.281, + 0.403, + 0.426, + 0.417 + ], + "angle": 0, + "content": "Novel View Renderings" + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.418, + 0.302, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.494, + 0.301, + 0.569 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.57, + 0.301, + 0.642 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.643, + 0.301, + 0.717 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.308, + 0.418, + 0.395, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.309, + 0.494, + 0.397, + 0.569 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.31, + 0.57, + 0.387, + 0.642 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.309, + 0.643, + 0.404, + 0.717 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.4, + 0.418, + 0.498, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.4, + 0.493, + 0.498, + 0.569 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.4, + 0.569, + 0.498, + 0.642 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.4, + 0.643, + 0.498, + 0.717 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.403, + 0.564, + 0.416 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.418, + 0.594, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.493, + 0.594, + 0.569 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.569, + 0.594, + 0.642 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.643, + 0.594, + 0.717 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.403, + 0.81, + 0.416 + ], + "angle": 0, + "content": "Novel View Renderings" + }, + { + "type": "image", + "bbox": [ + 0.605, + 0.428, + 0.679, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.494, + 0.687, + 0.569 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.57, + 0.688, + 0.642 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.643, + 0.688, + 0.717 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.69, + 0.428, + 0.785, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.69, + 0.493, + 0.785, + 0.569 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.69, + 0.57, + 0.785, + 0.642 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.69, + 0.643, + 0.784, + 0.717 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.785, + 0.418, + 0.883, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.785, + 0.493, + 0.883, + 0.569 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.785, + 0.569, + 0.883, + 0.642 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.785, + 0.643, + 0.883, + 0.717 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.72, + 0.907, + 0.778 + ], + "angle": 0, + "content": "Figure 9. Results of FaceLift on in-the-wild images. FaceLift accurately reconstructs 3D head models under challenging lighting conditions, achieving high fidelity (row 1). It captures fine facial details such as wrinkles (row 2), mustaches (row 3), and individual hairs (row 2 and row 4). Additionally, it remains robust to complex facial expressions (row 3) and various skin tones (row 4). Furthermore, it can realistically reconstruct facial paint (row 4). More results are provided in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.789, + 0.486, + 0.895 + ], + "angle": 0, + "content": "back-of-head rendering but suffers from severe identity loss (row 2) and fails to accurately reconstruct face paint (row 4). Era3D [31] often produces an inaccurate head shape, particularly from side views, and offers fewer geometric details compared to FaceLift. LGM [56] generates Gaussians with inaccurate color and opacity and lacks proper facial geometry, resulting in distorted features. Baseline Our \\(MV + LGM\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.789, + 0.907, + 0.865 + ], + "angle": 0, + "content": "shows that our multi-view diffusion model enhances LGM by providing improved facial geometry and texture details. However, the LGM reconstructor still produces Gaussians with inaccurate shapes and opacities, further underscoring the advantages of our transformer-based reconstructor." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We present more FaceLift's novel view rendering results in Fig. 9 to demonstrate FaceLift's ability to produce high" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12697" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.094, + 0.148, + 0.103 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image_caption", + "bbox": [ + 0.171, + 0.094, + 0.325, + 0.104 + ], + "angle": 0, + "content": "w/o. Input View Reconstruction" + }, + { + "type": "image_caption", + "bbox": [ + 0.33, + 0.094, + 0.472, + 0.104 + ], + "angle": 0, + "content": "w. Input View Reconstruction" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.107, + 0.171, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.178, + 0.118, + 0.308, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.327, + 0.118, + 0.46, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.171, + 0.483, + 0.254 + ], + "angle": 0, + "content": "Figure 10. Importance of input view reconstruction. The diffusion model that is not trained to perform the input view reconstruction, i.e., w/o. Input View Reconstruction, overfits to synthetic training distribution, suffers from severe identity loss during inference. Trained with input view reconstruction, our method preserves the input identity and expression faithfully." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.267, + 0.484, + 0.465 + ], + "angle": 0, + "content": "fidelity, realistic 3D head reconstructions with intricate details across a variety of challenging scenarios. FaceLift effectively handles faces under various lighting conditions. It can especially render realistic novel view images given a photo captured with an iPhone under dark lighting conditions (row 1 column 1), emphasizing its robustness and potential for real-world application. It reconstructs facial details with high fidelity, especially the wrinkles and folds on the face caused by extreme expression. FaceLift also excels at reconstructing challenging textures, such as mustaches and hair. Furthermore, it faithfully reconstructs facial paint, despite such data not being included in our synthetic face dataset, showcasing its strong generalization ability." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.474, + 0.243, + 0.49 + ], + "angle": 0, + "content": "5. Ablation Study" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.495, + 0.334, + 0.51 + ], + "angle": 0, + "content": "5.1. Input View Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.514, + 0.484, + 0.786 + ], + "angle": 0, + "content": "We conduct an ablation study to demonstrate the importance of reconstructing the input view during training. For comparison, we train a multi-view diffusion model that generates six novel views. In this baseline, the first generated view's elevation is adjusted from \\(0^{\\circ}\\) to \\(20^{\\circ}\\), while the remaining views adopt the same camera poses as in our default setting. We refer to this variant as w/o. Input View Reconstruction. Fig. 10 presents the view generation results of the two diffusion models when applied to real-world images. Without the input view reconstruction task, the model trained on the synthetic dataset generates views within a limited distribution, leading to noticeable identity loss. Moreover, it loses its ability to preserve facial expressions and face paint. In contrast, incorporating the input view reconstruction task during training enables our diffusion model to faithfully regenerate the input view, significantly improving its generalization ability. Quantitative comparison is provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.792, + 0.262, + 0.806 + ], + "angle": 0, + "content": "5.2. Number of Views" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We evaluate three configurations: four views (front, left, right, and back), six views (adding front-left and front-right), and eight views (further including front-top and front-bottom). Fig. 11 compares the baselines using different numbers of input views. With only four views, the reconstructor fails to capture a complete forehead; however," + }, + { + "type": "image_caption", + "bbox": [ + 0.533, + 0.094, + 0.558, + 0.102 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image_caption", + "bbox": [ + 0.582, + 0.094, + 0.617, + 0.102 + ], + "angle": 0, + "content": "4 Views" + }, + { + "type": "image_caption", + "bbox": [ + 0.636, + 0.094, + 0.672, + 0.102 + ], + "angle": 0, + "content": "6 Views" + }, + { + "type": "image_caption", + "bbox": [ + 0.691, + 0.094, + 0.726, + 0.102 + ], + "angle": 0, + "content": "8 Views" + }, + { + "type": "image_caption", + "bbox": [ + 0.746, + 0.094, + 0.78, + 0.102 + ], + "angle": 0, + "content": "4 Views" + }, + { + "type": "image_caption", + "bbox": [ + 0.8, + 0.094, + 0.835, + 0.102 + ], + "angle": 0, + "content": "6 Views" + }, + { + "type": "image_caption", + "bbox": [ + 0.855, + 0.094, + 0.889, + 0.102 + ], + "angle": 0, + "content": "8 Views" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.103, + 0.574, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.578, + 0.103, + 0.629, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.103, + 0.683, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.103, + 0.736, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.739, + 0.103, + 0.818, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.821, + 0.103, + 0.872, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.875, + 0.103, + 0.899, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.192, + 0.906, + 0.233 + ], + "angle": 0, + "content": "Figure 11. Number of input views of Gaussian reconstructor. Using six views strike a good balance between reconstruction quality and computational efficiency." + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.237, + 0.582, + 0.247 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image_caption", + "bbox": [ + 0.617, + 0.237, + 0.703, + 0.247 + ], + "angle": 0, + "content": "w/o. Pre-training" + }, + { + "type": "image_caption", + "bbox": [ + 0.716, + 0.237, + 0.799, + 0.247 + ], + "angle": 0, + "content": "w/o. Fine-tuning" + }, + { + "type": "image_caption", + "bbox": [ + 0.838, + 0.237, + 0.864, + 0.246 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.249, + 0.609, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.249, + 0.708, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.711, + 0.249, + 0.804, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.806, + 0.249, + 0.899, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.327, + 0.906, + 0.397 + ], + "angle": 0, + "content": "Figure 12. Two-stage training of reconstructor. Without pretraining on general objects, the reconstructor fails to produce clear textures in the reconstruction results. Meanwhile, without finetuning on synthetic human head data, the model lacks a refined understanding of facial structures, including the eyes and nose." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.408, + 0.906, + 0.514 + ], + "angle": 0, + "content": "with six views, it reconstructs the eyes and eyebrows more smoothly and renders challenging textures—such as facial wrinkles and ear folds—more realistically. Eight views do not offer significant visual improvements, and incur a higher computational cost in both stages. Thus, we find that six views achieve a good balance between reconstruction quality and computational efficiency." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.52, + 0.813, + 0.535 + ], + "angle": 0, + "content": "5.3. Two-stage Reconstructor Training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.906, + 0.658 + ], + "angle": 0, + "content": "As illustrated in Sec. 3.3, our Gaussian reconstructor follows a two-stage training pipeline. Fig. 12 shows that pretraining on general objects helps the model learn a diverse prior of geometric structures, resulting in clearer textures on delicate facial regions. Meanwhile, fine-tuning on synthetic human head data enables the reconstructor to gain a more refined understanding of facial structure, thereby enhancing the accuracy of features such as the eyes, nose, and hair." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.669, + 0.64, + 0.683 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We propose FaceLift, a feed-forward approach that lifts a single facial image to a detailed 3D reconstruction with preserved identity features. Our method uses multi-view diffusion to generate unobservable views and employs a transformer-based reconstructor to reconstruct 3D Gaussian splats, enabling high-quality novel view synthesis. To overcome the difficulty of capturing real-world multi-view human head images, we render high-quality synthetic data for training and show that, despite being trained solely on synthetic data, FaceLift can reconstruct 3D heads from real-world captured images with high fidelity. Compared with baselines [1, 4, 29, 31, 56], FaceLift generates 3D head representation with finer geometry and texture details and exhibits better identity preservation ability." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12698" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.271, + 0.108 + ], + "angle": 0, + "content": "7. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.484, + 0.176 + ], + "angle": 0, + "content": "This was supported in part by the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korean Government (MSIT) (No. RS-2024-00457882, National AI Research Lab Project)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.177, + 0.484, + 0.253 + ], + "angle": 0, + "content": "We appreciate the insightful discussions with Kai Zhang, Hao Tan, Zexiang Xu, Sai Bi, Sumit Chaturvedi, Hanwen Jiang, Yu-Ju Tsai, Kuan-Chih Huang, Chengxu Liu, and Dingyi Dai. We thank Nathan Carr and Kalyan Sunkavalli for their support." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.266, + 0.188, + 0.282 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.29, + 0.484, + 0.332 + ], + "angle": 0, + "content": "[1] Sizhe An, Hongyi Xu, Yichun Shi, Guoxian Song, Umit Y Ogras, and Linjie Luo. PanoHead: Geometry-aware 3D fullhead synthesis in \\(360^{\\circ}\\). In CVPR, 2023. 1, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.334, + 0.484, + 0.375 + ], + "angle": 0, + "content": "[2] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.377, + 0.483, + 0.417 + ], + "angle": 0, + "content": "[3] Ananta R. Bhattarai, Matthias Nießner, and Artem Sevastopolsky. TriPlaneNet: An encoder for EG3D inversion. In WACV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.42, + 0.483, + 0.475 + ], + "angle": 0, + "content": "[4] Bahri Batuhan Bilecen, Ahmet Berke Gokmen, and Aysegul Dundar. Dual encoder GAN inversion for high-fidelity 3D head reconstruction from single images. In NeurIPS, 2024. 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.477, + 0.483, + 0.504 + ], + "angle": 0, + "content": "[5] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3D faces. In SIGGRAPH, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.506, + 0.483, + 0.546 + ], + "angle": 0, + "content": "[6] James Booth, Anastasios Roussos, Stefanos Zafeiriou, Allan Ponniah, and David Dunaway. A 3D morphable model learnt from 10,000 faces. In CVPR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.548, + 0.484, + 0.63 + ], + "angle": 0, + "content": "[7] Marcel C Bühler, Kripasindhu Sarkar, Tanmay Shah, Gengyan Li, Daoye Wang, Leonhard Helminger, Sergio Orts-Escalano, Dmitry Lagun, Otmar Hilliges, Thabo Beeler, et al. Preface: A data-driven volumetric prior for few-shot ultra high-resolution face synthesis. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.633, + 0.484, + 0.715 + ], + "angle": 0, + "content": "[8] Marcel C Bühler, Gengyan Li, Erroll Wood, Leonhard Helminger, Xu Chen, Tanmay Shah, Daoye Wang, Stephan Garbin, Sergio Orts-Escolano, Otmar Hilliges, et al. Cafca: High-quality novel view synthesis of expressive faces from casual few-shot captures. In SIGGRAPH Asia, 2024. 2, 3, 5, 12, 15, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.717, + 0.483, + 0.772 + ], + "angle": 0, + "content": "[9] Sumit Chaturvedi, Mengwei Ren, Yannick Hold-Geoffroy, Jingyuan Liu, Julie Dorsey, and Zhixin Shu. SynthLight: Portrait relighting with diffusion model by learning to re-render synthetic faces. In CVPR, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.774, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[10] Xiyi Chen, Marko Mihajlovic, Shaofei Wang, Sergey Prokudin, and Siyu Tang. Morphable diffusion: 3d-consistent diffusion for single-image avatar creation. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.831, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[11] Antonia Creswell and Anil Anthony Bharath. Inverting the generator of a generative adversarial network. In TNNLS, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[12] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.29, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3D objects. In CVPR, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[13] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-XL: A universe of \\(10\\mathrm{M} + 3\\mathrm{D}\\) objects. In NeurIPS, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.18, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[14] Jiankang Deng, Jia Guo, Jing Yang, Niannan Xue, Irene Kotsia, and Stefanos Zafeiriou. ArcFace: Additive angular margin loss for deep face recognition. In IEEE TPAMI, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.237, + 0.905, + 0.265 + ], + "angle": 0, + "content": "[15] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.267, + 0.905, + 0.322 + ], + "angle": 0, + "content": "[16] Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. DreamSim: Learning new dimensions of human visual similarity using synthetic data. In NeurIPS, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.324, + 0.905, + 0.366 + ], + "angle": 0, + "content": "[17] Chen Gao, Yichang Shih, Wei-Sheng Lai, Chia-Kai Liang, and Jia-Bin Huang. Portrait neural radiance fields from a single image. arXiv preprint arXiv:2012.05903, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.368, + 0.905, + 0.423 + ], + "angle": 0, + "content": "[18] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. GET3D: A generative model of high quality 3D textured shapes learned from images. In NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.425, + 0.905, + 0.479 + ], + "angle": 0, + "content": "[19] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.482, + 0.905, + 0.522 + ], + "angle": 0, + "content": "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.905, + 0.567 + ], + "angle": 0, + "content": "[21] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.569, + 0.905, + 0.597 + ], + "angle": 0, + "content": "[22] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.599, + 0.905, + 0.627 + ], + "angle": 0, + "content": "[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.629, + 0.905, + 0.683 + ], + "angle": 0, + "content": "[24] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: Large reconstruction model for single image to 3D. In ICLR, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.686, + 0.905, + 0.727 + ], + "angle": 0, + "content": "[25] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.729, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[26] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In CVPR, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.773, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[27] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D gaussian splatting for real-time radiance field rendering. In ACM TOG, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.816, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[28] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In ICLR, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[29] Tobias Kirschstein, Simon Giebenhain, Jiapeng Tang, Markos Georgopoulos, and Matthias Nießner. GGHead: Fast and generalizable 3D gaussian heads. In SIGGRAPH Asia, 2024. 5, 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12699" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[30] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3D: Fast text-to-3D with sparse-view generation and large reconstruction model. In ICLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[31] Peng Li, Yuan Liu, Xiaoxiao Long, Feihu Zhang, Cheng Lin, Mengfei Li, Xingqun Qi, Shanghang Zhang, Wenhan Luo, Ping Tan, et al. Era3D: High-resolution multiview diffusion using efficient row-wise attention. In NeurIPS, 2024. 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.482, + 0.273 + ], + "angle": 0, + "content": "[32] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. In ACM TOG, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[33] Xueting Li, Shalini De Mello, Sifei Liu, Koki Nagano, Umar Iqbal, and Jan Kautz. Generalizable one-shot 3D neural head avatar. In NeurIPS, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.358 + ], + "angle": 0, + "content": "[34] Jiangke Lin, Yi Yuan, Tianjia Shao, and Kun Zhou. Towards high-fidelity 3D face reconstruction from in-the-wild images using graph convolutional networks. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.359, + 0.482, + 0.413 + ], + "angle": 0, + "content": "[35] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T., Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3D mesh in 45 seconds without pershape optimization. In NeurIPS, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.415, + 0.482, + 0.456 + ], + "angle": 0, + "content": "[36] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.457, + 0.482, + 0.525 + ], + "angle": 0, + "content": "[37] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3D: Single image to 3D using cross-domain diffusion. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.527, + 0.482, + 0.595 + ], + "angle": 0, + "content": "[38] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Guang Yong, Juhyun Lee, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.597, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[39] Julieta Martinez, Emily Kim, Javier Romero, Timur Bagautdinov, Shunsuke Saito, Shouou-I Yu, Stuart Anderson, Michael Zollhöfer, Te-Li Wang, Shaojie Bai, Chenghui Li, Shih-En Wei, Rohan Joshi, Wyatt Borsos, Tomas Simon, Jason Saragih, Paul Theodosis, Alexander Greene, Anjani Josyula, Silvio Mano Maeta, Andrew I. Jewett, Simon Venshtain, Christopher Heilman, Yueh-Tung Chen, Sidi Fu, Mohamed Ezzeldin A. Elshaer, Tingfang Du, Longhua Wu, Shen-Chi Chen, Kai Kang, Michael Wu, Youssef Emad, Steven Longay, Ashley Brewer, Hitesh Shah, James Booth, Taylor Koska, Kayla Hidle, Matt Andromalos, Joanna Hsu, Thomas Dauer, Peter Selednik, Tim Godisart, Scott Ardisson, Matthew Cipperly, Ben Humberston, Lon Farr, Bob Hansen, Peihong Guo, Dave Braun, Steven Krenn, He Wen, Lucas Evans, Natalia Fadeeva, Matthew Stewart, Gabriel Schwartz, Divam Gupta, Gyeongsik Moon, Kaiwen Guo, Yuan Dong, Yichen Xu, Takaaki Shiratori, Fabian Prada, Bernardo R. Pires, Bo Peng, Julia Buffalini, Autumn Trimble, Kevyn McPhail, Melissa Schoeller, and Yaser Sheikh. Codec Avatar Studio: Paired human captures for complete, driveable, and generalizable avatars. In NeurlPS, 2024. 2, 3, 5, 15, 17" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[40] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.149, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[41] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. StyleSDF: High-resolution 3D-consistent image and geometry generation. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.207, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[42] Foivos Paraperas Papantoniou, Alexandros Lattas, Stylianos Moschoglou, and Stefanos Zafeiriou. Relightify: Relightable 3D faces from a single image via diffusion models. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.262, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[43] Dario Pavllo, Graham Spinks, Thomas Hofmann, Marie-Francine Moens, and Aurélien Lucchi. Convolutional generation of textured 3D meshes. In NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.305, + 0.905, + 0.345 + ], + "angle": 0, + "content": "[44] Julius Plucker. XVii. on a new geometry of space. Philosophical Transactions of the Royal Society of London, 1865. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.348, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[45] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D diffusion. In ICLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.391, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[46] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3D object generation using both 2D and 3D diffusion priors. In ICLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.461, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[47] Elad Richardson, Matan Sela, and Ron Kimmel. 3D face reconstruction by learning from synthetic data. In 3DV, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.504, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[48] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.547, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[49] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.59, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[50] Shunsuke Saito, Gabriel Schwartz, Tomas Simon, Junxuan Li, and Giljoo Nam. Relightable gaussian codec avatars. In CVPR, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.632, + 0.905, + 0.7 + ], + "angle": 0, + "content": "[51] Kyle Sargent, Zizhang Li, Tanmay Shah, Charles Herrmann, Hong-Xing Yu, Yunzhi Zhang, Eric Ryan Chan, Dmitry Lagun, Li Fei-Fei, Deqing Sun, et al. ZeroNVS: Zero-shot 360-degree view synthesis from a single real image. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.703, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[52] Sefik Ilkin Serengil and Alper Ozpinar. HyperExtended LightFace: A facial attribute analysis framework. In ICEET, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.746, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[53] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: A single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.817, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[54] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 3, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.845, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[55] Wenqiang Sun, Shuo Chen, Fangfu Liu, Zilong Chen, Yueqi Duan, Jun Zhang, and Yikai Wang. DimensionX: Create any 3D and 4D scenes from a single image with controllable video diffusion. arXiv preprint arXiv:2411.04928, 2024. 12" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12700" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[56] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. LGM: Large multi-view gaussian model for high-resolution 3D content creation. In ECCV, 2024. 3, 5, 6, 7, 8, 17, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[57] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. DreamGaussian: Generative gaussian splatting for efficient 3D content creation. In ICLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[58] Aäron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning. In NeurIPS, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.288 + ], + "angle": 0, + "content": "[59] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.333 + ], + "angle": 0, + "content": "[60] Thomas Vetter and Volker Blanz. Estimating coloured 3D face models from single images: An example based approach. In ECCV, 1998. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.335, + 0.482, + 0.376 + ], + "angle": 0, + "content": "[61] Vishal Vinod, Tanmay Shah, and Dmitry Lagun. TEGLO: High fidelity canonical texture mapping from single-view images. In WACV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[62] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In NeurIPS, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.482, + 0.502 + ], + "angle": 0, + "content": "[63] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3D digital avatars using diffusion. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.482, + 0.56 + ], + "angle": 0, + "content": "[64] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. ProlificDreamer: High-fidelity and diverse text-to-3D generation with variational score distillation. In NeurIPS, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.616 + ], + "angle": 0, + "content": "[65] Erroll Wood, Tadas Baltrusaitis, Charlie Hewitt, Sebastian Dziadzio, Thomas J Cashman, and Jamie Shotton. Fake it till you make it: Face analysis in the wild using synthetic data alone. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[66] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3D generative-adversarial modeling. In NeurIPS, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[67] Kailu Wu, Fangfu Liu, Zhihan Cai, Runjie Yan, Hanyang Wang, Yating Hu, Yueqi Duan, and Kaiheng Ma. Unique3d: High-quality and efficient 3d mesh generation from a single image. arXiv preprint arXiv:2405.20343, 2024. 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[68] Yiqian Wu, Jing Zhang, Hongbo Fu, and Xiaogang Jin. LPFF: A portrait dataset for face generators across large poses. In ICCV, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[69] Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin Tong, and Jiaolong Yang. Structured 3d latents for scalable and versatile 3d generation. In CVPR, 2024. 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[70] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024. 12, 13" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[71] Yu-Ying Yeh, Koki Nagano, Sameh Khamis, Jan Kautz, Ming-Yu Liu, and Ting-Chun Wang. Learning to relight portrait images via a virtual light stage and synthetic-to-real adaptation. In ACM TOG, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.149, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[72] Ziyang Yuan, Yiming Zhu, Yu Li, Hongyu Liu, and Chun Yuan. Make encoder great again in 3D GAN inversion through geometry and occlusion-aware encoding. In ICCV, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.206, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[73] Bowen Zhang, Yiji Cheng, Chunyu Wang, Ting Zhang, Jiaolong Yang, Yansong Tang, Feng Zhao, Dong Chen, and Baining Guo. RodinHD: High-fidelity 3D avatar generation with diffusion models. In ECCV, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.263, + 0.905, + 0.316 + ], + "angle": 0, + "content": "[74] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. GS-LRM: Large reconstruction model for 3D gaussian splatting. In ECCV, 2024. 3, 5, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.319, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[75] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.362, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[76] Xiaozheng Zheng, Chao Wen, Zhaohu Li, Weiyi Zhang, Zhuo Su, Xu Chang, Yang Zhao, Zheng Lv, Xiaoyuan Zhang, Yongjie Zhang, Guidong Wang, and Xu Lan. Headgap: Few-shot 3d head avatar via generalizable gaussian priors. In 3DV, 2025. 3" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12701" + } + ] +] \ No newline at end of file diff --git a/2025/FaceLift_ Learning Generalizable Single Image 3D Face Reconstruction from Synthetic Heads/1763108f-f851-4f26-ba5a-510a2c051a36_origin.pdf b/2025/FaceLift_ Learning Generalizable Single Image 3D Face Reconstruction from Synthetic Heads/1763108f-f851-4f26-ba5a-510a2c051a36_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5703e8cd76ecad9c8cf45e2dde0be39f768d8f46 --- /dev/null +++ b/2025/FaceLift_ Learning Generalizable Single Image 3D Face Reconstruction from Synthetic Heads/1763108f-f851-4f26-ba5a-510a2c051a36_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9ccf58a63c332b2b78dd9b678ad28d1bf520625e1219bda08aae4485fd3bbffc +size 4577752 diff --git a/2025/FaceLift_ Learning Generalizable Single Image 3D Face Reconstruction from Synthetic Heads/full.md b/2025/FaceLift_ Learning Generalizable Single Image 3D Face Reconstruction from Synthetic Heads/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d3be489045cfe518f35d06e928144c46529136ce --- /dev/null +++ b/2025/FaceLift_ Learning Generalizable Single Image 3D Face Reconstruction from Synthetic Heads/full.md @@ -0,0 +1,512 @@ +# FaceLift: Learning Generalizable Single Image 3D Face Reconstruction from Synthetic Heads + +Weijie Lyu $^{1*}$ Yi Zhou $^{2}$ Ming-Hsuan Yang $^{1}$ Zhixin Shu $^{2}$ + +1University of California, Merced 2Adobe Research + +![](images/cfe088eee6b0315c59c76e4e0696b845e6a2ca48fbeffb44833afc0d3bfb61c1.jpg) +Single Image Input + +![](images/4c793c5ef31a912d348c6c0cc95f77ae46f7d17e7dc33b912fa9241149b36de4.jpg) +High-fidelity 3D Gaussian Representation + +![](images/dabe944618ae82f7db9e5729ef8c6fdb9b8b48395e9768665d7e5ccdb4c32e6d.jpg) +Full Head NVS +Figure 1. FaceLift transforms a single facial image into a high-fidelity 3D Gaussian head representation. Trained exclusively on synthetic 3D data, our pipeline first generates sparse, identity-preserving multiview images of the input head using a diffusion model. These sparse generated views are then fed into a transformer-based 3D Gaussian splats reconstructor, producing complete and detailed 3D head representation that generalize remarkably well to real-world human images. Project page: https://weijielyu.github.io/FaceLift. + +![](images/cab329f58b2a5b3e56294b1affd051cc3875f0fdf537c45b1b82e1be3e3c6cb4.jpg) +High-quality Details + +# Abstract + +We present FaceLift, a novel feed-forward approach for generalizable high-quality 360-degree 3D head reconstruction from a single image. Our pipeline first employs a multiview latent diffusion model to generate consistent side and back views from a single facial input, which then feed into a transformer-based reconstructor that produces a comprehensive 3D Gaussian splats representation. Previous methods for monocular 3D face reconstruction often lack full view coverage or view consistency due to insufficient multi-view supervision. We address this by creating a high-quality synthetic head dataset that enables consistent supervision across viewpoints. To bridge the domain gap between synthetic training data and real-world images, we propose a simple yet effective technique that ensures the view generation process maintains fidelity to the input by learning to reconstruct the input image alongside the view generation. Despite being trained exclusively on synthetic data, our method demonstrates remarkable generalization to real-world images. Through extensive qualitative and quantitative evaluations, we show that FaceLift outperforms state-of-the-art 3D face reconstruction methods on identity preservation, detail recovery and rendering quality. + +# 1. Introduction + +3D face reconstruction has been a central focus in computer vision and graphics research for decades, driven by its crucial applications in immersive virtual and augmented realities, VFX and gaming, digital entertainment, and next-generation telepresence systems. However, achieving high quality reconstruction from a single image remains very challenging. On one hand, the monocular face reconstruction problem is highly ill-posed – a single 2D image can be produced by countless different 3D face shapes, creating fundamental ambiguity. On the other hand, the human visual system is highly attuned to facial details, making even subtle artifacts and imperfections noticeable to the eye. + +Traditional 3D head synthesis approaches typically use parametric textured mesh models [32, 60] trained on 3D scan datasets. While these models enable basic head generation, the rendered images frequently lack fine-scale geometric detail, realistic textures, and convincing hair, limiting their perceptual realism and expressiveness. Recent breakthroughs in image generative models [19, 23] and novel view synthesis techniques [27, 40] have opened new possibilities for this research area. Leveraging these developments, recent works [1, 72] use neural 3D representations to learn effective 3D head representation from unstructured real face image datasets [25, 68]. While these datasets improve the realism and diversity of rendering results, they fail to provide multi-view supervision for modeling 3D consis + +![](images/9714794b1877a26ae2d9a09376437b1f1598dca3e1e561b53c0e802b8ef606e4.jpg) +Figure 2. Comparison with RodinHD. RodinHD [73] trains triplane diffusion with synthetic data, resulting in apparent identity loss. In contrast, FaceLift achieves better identity preservation and generalizes effectively to real human portraits. + +tency causing view inconsistency and artifacts on the back of the head. Since diverse multi-view real images are hard to acquire, RodinHD [73] leverages synthetic multi-view images to train generative models that directly output 3D neural representations of the head. However, training solely on synthetic data often results in significant perceptual identity loss in the generated outputs, as demonstrated in Fig. 2. + +In this work, we present FaceLift, a pipeline for learning generalizable and high-fidelity single image to 3D face representation from synthetic head data. To achieve high quality reconstruction that preserves the input facial identity, we adopt a two-stage pipeline to first generate identity preserving multi-view images using a diffusion model [48], followed by a transformer-based feed-forward reconstructor to fuse the generated sparse views into a comprehensive 3D Gaussian representation. We train the model with synthetic images - multi-view renderings of 3D synthetic human portraits using Blender. We highlight two key techniques for generalizing to real-world images and preserving input facial identity. First, we emphasize the importance of reconstructing the input image alongside the view synthesis task in the conditional diffusion model training, which significantly improved generalization capability in testing. Second, we demonstrate that training the feed-forward reconstructor benefits from a two-stage training process: pretraining on general objects [12] to acquire a rich geometry and texture prior, followed by fine-tuning on synthetic human head data to capture head-specific geometry. With our two-stage approach, we focus on learning identity preservation in the image space during the first stage, achieving higher input fidelity compared to existing methods. + +Comparing with prior art, we achieve three key advancements: (1) robust view consistency through multi-view at + +tention and supervision, (2) improved generalization from training techniques and foundational model, ensuring accurate identity preservation, (3) high-quality facial texture and hair details via pixel-aligned Gaussian representation. + +We extensively evaluate FaceLift quantitatively and qualitatively across diverse datasets. Using real multiview studio captures [39] and an independent synthetic human dataset [8], our approach consistently surpasses previous state-of-the-art methods across all evaluation metrics. Through extensive testing on in-the-wild portrait images, we demonstrate that our method reconstructs complete 3D heads with fine-grained details, accurate identity preservation, and high visual fidelity. Comparisons and ablation studies confirm that multi-view consistent training data is crucial for high-fidelity face reconstruction. Our contributions are summarized as follows: + +- We propose FaceLift, a framework that reconstructs a high-fidelity 3D head from a single image using view generation and feed-forward reconstructor. +- Despite being trained solely on synthetic human head data, our approach shows no domain gap on real-world images, highlighting both the effectiveness of synthetic data and our model's robust generalization capabilities. +- We construct two benchmarks on single-image to 3D full head reconstruction tasks using the publicly available datasets Cafca [8] and Ava-256 [39] to quantitatively evaluate models' performance on both reconstruction accuracy and identity preservation ability. +- Our comprehensive evaluation confirms that our approach achieves state-of-the-art performance in reconstruction accuracy and identity preservation. + +# 2. Related Work + +Face Reconstruction. 3D face reconstruction has been a long-standing challenge in computer vision, with substantial progress driven by diverse approaches. Vetter and Blanz [60] pioneer a method for synthesizing 3D faces by linearly blending multiple 3D templates, now widely known as blendshapes. This work establishes the foundation for 3D Morphable Models (3DMMs), which represent 3D face shapes and textures as principal components derived from scanned data. Subsequent research [5, 6, 32, 34, 47] extend this framework, enabling the generation of new 3D faces by manipulating blending coefficients. However, these methods produce mesh-based representations that lack fine details and are limited to modeling the front of the face, excluding hair and 360-degree synthesis. While 3DMM-based methods have been foundational, recent advances in deep learning, especially Generative Adversarial Networks (GANs) [19, 25, 26], have greatly improved 3D face synthesis quality. EG3D [72] uses a tri-plane NeRF representation with a pose-conditioned StyleGAN2 [26] framework. Follow-up works [3, 33] achieve single-image-to-3D gen + +eration through GAN inversion [11]. Despite their success, these methods can only synthesize near-frontal views. To overcome this, PanoHead [1] introduces a tri-grid neural volume representation, enabling full 360-degree head synthesis. Unfortunately, it does not provide a 3D head representation for consistent multi-view rendering. Recent efforts explore alternative representations for 3D face reconstruction from sparse input, such as a single image [17, 42, 61] or few-shot images [7]. However, these methods still require pre-instance optimization. Rodin [63] and its extension RodinHD [73] employ an image-conditioned diffusion model to generate a triplane representation of a human head for full-head novel view synthesis. Nevertheless, their triplane diffusion model is limited to synthetic data and struggles to achieve high-fidelity reconstructions from real-world images. For animatable 3D head avatars generations, Morphable Diffusion [10] generates multi-view consistent images from a single image using a morphable mesh, while HeadGAP [76] generates 3D animatable head avatars using few-shot input, leveraging 3D head priors learned from large-scale data. In contrast, our work focuses on leveraging synthetic training data to produce high-fidelity, detailed 3D Gaussian head models. + +Synthetic Human Data. Capturing high-quality 3D data of real humans requires a controlled studio environment and costly photography equipment [39]. As an alternative, large-scale synthetic 3D head datasets have emerged as an effective and resource-efficient solution for tasks like human head reconstruction [8, 63, 65, 73] and photorealistic relighting [9, 71], offering a scalable way to train models without the restrictions of real-world data acquisition. Inspired by these previous works, we aim to use synthetic data to improve the model's understanding of the human head and minimize the generalization gap between synthetic data training and real-world inference. + +Image or Text to 3D. Generative models have achieved remarkable success in 2D image generation with VAEs [28, 58], GANs [19, 25, 26], and diffusion models [23, 48, 54]. Building on this success, extensive research has extended these models to 3D content generation [18, 41, 43, 66]. Starting with DreamFusion [45], numerous works [36, 46, 51, 57, 64] try to distill NeRF [40, 62, 72] or 3D Gaussians [27] representation from 2D image diffusion using a Score Distillation Sampling (SDS) loss. These methods can produce high-quality results but often encounter challenges such as slow optimization, over-saturated colors, and the Janus problem. To overcome these challenges, recent works [30, 31, 35, 37, 53] generate multi-view images with high consistency, which can be directly used for 3D reconstruction with neural reconstruction methods [27, 40, 62]. However, optimizing NeRF or NeuS remains far from real-time performance. Recent advancements in large 3D reconstruction models (LRMs) [24, 30, 56, 74] offer a path + +way to faster 3D reconstruction. Leveraging scalable transformer architectures [15, 59] and large datasets like Objverse [12, 13], these models effectively capture generalizable 3D priors. Unlike traditional pre-scene optimization methods [27, 40, 62], LRMs employ a feed-forward approach, enabling the prediction of high-quality NeRF, mesh, or 3D Gaussian representations from sparse images in under a second. However, most of these research efforts are applied to general objects, with limited or suboptimal results for 3D head reconstruction. + +# 3. Proposed Method + +As shown in Fig. 3, given a single frontal image of a human face $y$ , our goal is to reconstruct a complete 3D head $G$ , represented as 3D Gaussian splats, with detailed facial texture and preserved identity. This requires our system to have prior knowledge on the geometry structure of a human face and the ability to synthesis plausible details which are not visible in the input view. Hence, we train a multiview diffusion model $f_{D}$ on synthetic human head data to generate $N$ views $x_0^1, x_0^2, \ldots, x_0^N$ covering $360^\circ$ of the human head while achieving multi-view consistency and preserving identity. We choose pixel-aligned 3D Gaussians to obtain the final 3D representation. Compared to NeRFs and meshes, 3D Gaussians offer explicit volumetric primitives that better capture subtle facial geometry and fine details. Their semi-transparent kernels naturally model effects like wispy hair and translucency, which are challenging for discrete surfaces or density fields. These generated views $x_0^{1:N}$ from the diffusion model, along with their corresponding Plücker ray coordinates $\mathbf{P}^{(1:N)}$ , are input into a transformer-based Gaussian reconstructor $f_R$ to predict a set of 3D Gaussians $G$ . Training of the Gaussian reconstructor follows a pre-training process on general objects [12] and a fine-tuning process on synthetic human head data. + +# 3.1. Synthetic Human Head Dataset + +We implement a 3D head asset generation pipeline inspired by [65]. Our process begins with a collection of high-quality, artist-created 3D head meshes, which we enhance by incorporating detailed facial components, including eyes, teeth, gums, and both facial and scalp hair. We then augment these base models through rigging for pose variation and blendshape deformation for diverse facial expressions. The final head models are enriched with a set of PBR texture maps, including albedo, normal, roughness, specular, and subsurface scattering maps. At last, we dress the head model with a collection of clothing assets. The entire pipeline is implemented in Blender and the images are rendered with Cycles renderer. + +To train our networks, we render images (samples shown in Fig. 4) at $512 \times 512$ resolution from 200 unique identities, each with 50 varied appearances, including different hairstyles, skin tones, expressions, clothes, poses, etc. + +![](images/0b6dcfa3e9d95382ee8a66e0d65d3b40dd971f7d166af00dc6902fe0699a9ddf.jpg) +Figure 3. Overview of FaceLift. Given a single image of a human face $y$ as input, we train an image-conditioned, multi-view diffusion model to generate novel views $x_0^1, \ldots, x_0^N$ covering the entire head. By generating $x_0^1$ the same as $y$ and leveraging the high-quality synthetic data, our multi-view latent diffusion model can hallucinate unseen views of the human head with high-fidelity and multi-view consistency. We then train a transformer-based reconstructor $f_R$ , which takes multi-view images $x_0^{1:N}$ and their camera poses $\mathbf{P}^{1:N}$ as input and generates 3D Gaussian splats $G$ to represent the human head. + +![](images/ce4410f083f829d2240ca808516f1e389ee172b9e65296f4a48fba896189a94e.jpg) +Figure 4. Synthetic data examples. Top row: six views for diffusion training. Bottom row: samples of random views for reconstructor training. + +We render our training dataset under two types of lighting conditions: (1) ambient light and (2) random HDR environment light. We render six views for each subject to train the multi-view diffusion model. For fine-tuning the transformer-based reconstructor, we render 32 views with random camera poses. + +# 3.2. View Generation + +We model the sparse view generation from a single image input as a conditional diffusion process. We use a multiview diffusion model $f_{D}$ to generate $N$ views, denoted as $X_0^1, X_0^2, \ldots, X_0^N$ , given a single front-facing image $y$ and CLIP text embeddings $e^1, e^2, \ldots, e^N$ corresponding to each generated view. This process is expressed as: + +$$ +\left\{X _ {0} ^ {1}, X _ {0} ^ {2}, \dots , X _ {0} ^ {N} \right\} = f _ {D} \left(y, \left\{e ^ {1}, e ^ {2}, \dots , e ^ {N} \right\}\right). \tag {1} +$$ + +We aim to learn the joint distribution of all these views, conditioning on the input image $y$ and text embedding $e^1, e^2, \ldots, e^N$ . We denote the joint distribution as: + +$$ +p _ {f _ {D}} \left(x _ {0} ^ {1: N} \mid y, e ^ {1: N}\right) := p _ {f _ {D}} \left(\left\{x _ {0} ^ {1}, \dots x _ {0} ^ {N} \right\} \mid y, \left\{e ^ {1}, \dots e ^ {N} \right\}\right). \tag {2} +$$ + +View Selection. Given a single near frontal view face image with azimuth $\alpha$ , our multi-view diffusion model generates six views with azimuths equal to $\{\alpha, \alpha \pm 45^\circ, \alpha \pm 90^\circ, \alpha + 180^\circ\}$ , covering 360 degrees of the human head. Elevation is 0 for all images. We opt for six views as the optimal balance - fewer views compromise detail quality while more views become computationally prohibitive for full head reconstruction. An ablation study comparing four, six, and eight views is presented in Sec. 5.2. + +Multi-view Attention. To ensure the consistency of the generated novel views, we use a multi-view attention mechanism to facilitate information propagation and implicitly encode multi-view dependencies. Our attention module encourages multi-view consistency by extending the 2D self-attention mechanism to 3D and enabling interactions across views. Instead of treating each view independently, we apply self-attention across all views simultaneously, allowing information to be shared between them. Specifically, we start with an input tensor of shape $B \times V \times H \times W \times C$ , where $B$ is the batch size, $V$ is the number of views, $H$ and $W$ denote the spatial resolution of the intermediate feature maps, and $C$ is the number of feature channels. We reshape this tensor to $B \times VHW \times C$ , treating all spatial locations across all views as a unified sequence of tokens for self-attention. This design allows the model to learn multi-view correlations by sharing information across views within the attention layers, enabling it to generate consistent RGB images. We provide an ablation study on the multi-view attention mechanism in the supplementary material. + +Input View Reconstruction. During Training, we enforce the first generated view to share the same camera with the input image. In other words, we reconstruct the input view in the view generation process. We find this approach, combined with the multi-view attention mechanism, significantly outperforms the alternative strategy of generating only novel views, which tends to overfit to synthetic training identities and compromises generalization capability as we will show in Sec. 5.1 and the supplementary material. + +# 3.3. Multi-view to 3D Gaussians Reconstruction + +Transformer-based Reconstructor. We choose pixel-aligned 3D Gaussians as the final 3D representation. Each Gaussian $G_{i}$ is defined by position $p_i$ , scale $s_i$ , orientation $q_{i}$ , opacity $\alpha_{i}$ , and color features $c_{i}$ . We use a transformer-based reconstructor $f_{G}$ to obtain 3D Gaussians from generated multi-view images $x_0^{1:N}$ and their corresponding Plücker ray coordinates [44] $\mathbf{P}^{1:N}$ : + +$$ +\left\{G _ {i} \right\} _ {i = 1} ^ {N H W} = \left\{p _ {i}, s _ {i}, q _ {i}, \alpha_ {i}, c _ {i} \right\} _ {i = 1} ^ {N H W} = f _ {G} \left(x _ {0} ^ {1: N}, \mathrm {P} ^ {1: N}\right), \tag {3} +$$ + +Our $f_{G}$ is a large reconstruction model [24, 74] which follows the implementation of GS-LRM [74]: the $N$ multiview images are concatenated with their Plücker ray coordinates computed from the camera intrinsic and extrinsic parameters for pose conditioning. Then, the inputs are patched by dividing the per-view feature map into non-overlapping patches with a patch size of $p$ . Each 2D patch is then flattened into a 1D vector. Finally, a linear layer $L$ is utilized to map the 1D vectors to image patch tokens: + +$$ +\left\{\mathrm {T} _ {j} ^ {n} \right\} _ {j = 1, 2, \dots , \frac {H W}{p ^ {2}}} = L \left(\text {P a t c h i f y} _ {p} \left(\text {C o n c a t} \left(\mathrm {I} ^ {n}, \mathrm {P} ^ {n}\right)\right)\right). \tag {4} +$$ + +Where $\{\mathrm{T}_j^n\}$ denotes the set of patch tokens for image $n$ , totaling $\frac{HW}{p^2}$ tokens per image. The set of multi-view image tokens $\{\mathrm{T}_j^n\}$ are concatenated and processed through a chain of transformer blocks. Each transformer block is equipped with residual connections [20] and consists of PreLayerNorm [2], multi-head Self-Attention [59] and MLP. Later, the output tokens from the transformer are decoded into Gaussian parameters using a single linear layer. Then, the Gaussian parameters are unpatchified into $p^2$ Gaussians. Finally, we end up with $HW$ Gaussians for each view, where pixel encodes one 3D Gaussian. + +Two-stage Training. We find that training the transformer-based reconstructor solely on synthetic human head data leads to inferior texture details when applied to real-world images (see ablation study in Fig. 12). We suspect this limitation arises because the synthetic datasets lack geometric diversity. To address this, we propose a two-stage training approach in which the reconstructor is pre-trained on diverse object data [12] and subsequently fine-tuned using synthetic head data. The pre-training stage enables the reconstructor to learn a broad prior of diverse geometric structures, yielding more detailed and clearer textures in delicate facial regions such as the eyes, nose, and ears. The fine-tuning process then imparts specific knowledge of head structure, producing smoother and more realistic results. During training, we randomly select four input views to reconstruct a total of eight views, four input and four novel views. Following [74], we optimize the model using a combination of MSE and perceptual losses. During inference, the reconstructor processes the six-view outputs from multi-view diffusion model to reconstruct the head. + +# 3.4. Real-world Image Inference + +For inference on real-world images, since their intrinsic parameters are unknown, we adopt a camera fov of $50^{\circ}$ , same as during training. To ensure plausible outputs, we first apply an MTCNN face detector to estimate the face's size and center. The image is then resized and cropped/extended to match the average face size and center computed from the training data. We find this alignment compensates for the unknown intrinsic parameters well, ensuring plausible reconstruction results. + +# 4. Experiments + +# 4.1. Experimental Setup + +Evaluation Datasets. To quantitatively evaluate FaceLift, we establish two benchmarks for single-image to 3D full head reconstruction tasks using publicly available datasets: (1) Cafca dataset [8]: We select 40 subjects with 7 to 19 test camera poses each. Since the camera positions are randomly distributed, we manually select the most frontal view as input. Note that this synthetic dataset was independently developed and differs significantly from our training dataset. (2) Ava-256 dataset [39]: This studio-captured dataset contains real human subjects. We sample 10 subjects and 10 test camera poses for our evaluation. More details in supplemental. To demonstrate our system's generalization capabilities, we also evaluate on a set of in-the-wild face images for qualitative assessment. + +Baselines. We compare our results against three state-of-the-art methods for single-face 3D reconstruction: GGHead [29], PanoHead [1], and Dual Encoder [4]. We perform GAN-inversion to reconstruct 3D head from a given input image using these models. To emphasize the importance of utilizing our synthetic human head data for training, we also compare our method with two methods that focus on general object reconstruction: Era3D [31] and LGM [56]. More comparison results with mesh-based methods are provided in the supplementary material. + +We further developed a baseline, Our $MV + LGM$ , which leverages the multi-view outputs generated by our diffusion model and employs LGM for reconstruction. This demonstrates that our method can be seamlessly integrated with other reconstruction frameworks to enhance performance on face reconstruction tasks. We tried to fine-tune the LGM reconstructor with our synthetic data, but it provides inferior results with incorrect geometry and artifacts compared with the original weights, which we suspect is due to training data mismatch (see details in the supplementary material). + +Evaluation Metrics. We evaluate reconstruction quality using four standard metrics: PSNR, SSIM, LPIPS [75], and DreamSim [16]. To evaluate identity preservation, we perform face verification using ArcFace [14] through the DeepFace [52] implementation. + +Implementation Details. Both Cafca [8] and Ava-256 [39] datasets provide multi-view RGB images and corresponding camera poses. However, their camera system differs from the ones utilized in FaceLift and baselines. We recalculate the test camera extrinsic in each method's camera system. For a more accurate comparison, we use the Mediapipe facial landmark detection algorithm [38] to identify facial landmarks in both the target images and the rendered outputs, aligning them based on landmark distributions. Details are provided in the supplementary material. + +Our system takes approximately 8 seconds to infer a + +
MethodPSNR ↑SSIM ↑LPIPS ↓DreamSim ↓ArcFace ↓
GGHead [29]10.350.74060.36360.32520.2681
PanoHead [1]10.720.75940.33510.20480.2183
Dual Encoder [4]10.780.73850.39220.27850.2421
Era3D [31]13.690.72300.36620.28920.2978
LGM [56]16.520.79330.30600.15520.2557
Our MV + LGM [56]14.130.78120.29560.12820.1767
FaceLift16.610.79680.26940.10960.1573
+ +![](images/2a40e3bf004896985025f7ab3ebdd754a8e5799ec01d59de5f2e98d3d4a12003.jpg) +Figure 5. Visual results on Cafca compared with face reconstruction methods. FaceLift renders novel views that closely match the ground truth, while other baselines often fail to reconstruct the 3D head in correct colors or geometry structures. + +![](images/980876d00a5dbee340c3773abeaaa984769433a2e4fede1b5d8489c12e3c6b93.jpg) +Figure 6. Visual results on Cafca compared with general objects reconstruction methods. Comparison with general object reconstruction methods shows the importance of specialized data. + +3D Gaussian head from a single image: about 1.5 seconds for preprocessing (background removal, rescaling, etc.), 5.5 seconds for multi-view image generation, and under 1 second for 3D Gaussians reconstruction. + +# 4.2. Experiments on the Cafca Dataset + +We report numerical comparison results on Cafca in Tab. 1. FaceLift performs favorably against baselines, especially on DreamSim [16] metric, indicating high-quality perceptual similarity. It also achieves better identity preservation performance, as demonstrated by a lower ArcFace [14] distance. We show visual results in Fig. 5 and Fig. 6. FaceLift yields rendering results that closely match the ground truth. Compared with other baselines, GGHead [29] does not support full-head rendering, resulting in unrealistic outputs when the view angle significantly deviates from the input. PanoHead [1] struggles with challenging hairstyles, while Dual Encoder [4] produces blurred facial textures. Additionally, Era3D [31] introduces artifacts on the back of the head, and LGM [56] yields inaccurate nose and jaw shapes, underscoring the importance of our synthetic human head data. When integrated with our multi-view diffusion approach, LGM achieves enhanced performance, demonstrating that our method can be seamlessly combined with existing baselines to boost their results. + +Table 1. Quantitative results on Cafca. FaceLift achieves favorable performance on all evaluation metrics. + +
MethodPSNR ↑SSIM ↑LPIPS ↓DreamSim ↓ArcFace ↓
Era3D [31]14.770.79630.25380.25150.3721
LGM [56]14.050.81360.24760.14960.3142
Our MV+LGM [56]15.240.82130.22920.10930.2264
FaceLift16.520.82710.22770.10650.1871
+ +Table 2. Quantitative results on Ava-256. FaceLift performs favorably than baseline methods in both reconstruction metrics and identity facial identity metric, showing a better generalization ability towards real-captured human images. + +![](images/f3b50a414451536fe0eff292687a3a9b3c3c283a3df72120ffbc5e162c16dddc.jpg) +Figure 7. Visual results on Ava-256. Compared with baselines, FaceLift provides multi-view renderings that are more realistic and similar to ground truth. Era3D fails to deliver delicate facial structures, while LGM generates inaccurate head shapes and colors. + +# 4.3. Experiments on the Ava-256 Dataset + +We further evaluate FaceLift against other baselines on a studio-captured real human dataset, Ava-256 [50]. GAN-inversion based methods [1, 4, 29] fail to produce reasonable results with the test camera poses in this dataset, so we exclude these baselines. Tab. 2 shows that FaceLift outperforms all other baselines across all evaluation metrics, demonstrating superior reconstruction quality and identity preservation. It also highlights FaceLift's strong ability to generalize to real human faces. As shown in Fig. 7, FaceLift achieves more realistic head synthesis, while Era3D [31] struggles with accurate skin and hair textures, as well as facial details. LGM [56] produces inaccuracies in the nose shape. When combined with our multi-view diffusion model, LGM yields more accurate geometric structures, yet its texture quality remains inferior to that of FaceLift. + +# 4.4. Experiments with In-the-wild Images + +We collect in-the-wild human face images and present qualitative results in comparison with other baselines in Fig. 8. Baseline methods often produce undesirable artifacts. For instance, PanoHead [1] frequently fails to render the back of the head and sometimes generates extra eyes at the rear. It also struggles to synthesize hair, shadows, wrinkles, and facial paint accurately, and its outputs lack multi-view consistency (e.g., the girl continues to face the camera in novel view 1 despite a changed pose). Dual Encoder [4] improves + +![](images/7fed4dcd366bfe9351ce23d69b27cd076b931dc2a56517fef93c57a8d9c4f698.jpg) +Input View + +![](images/cc3371677dd01ccd6c488aaf6314b097a778b1fb3907678e08bd6f858ef84de6.jpg) + +![](images/2ce1d5f4766c43d796e58cfb5c4f54e5a98545564062e4aa29c84ecedf7e89da.jpg) +Rendered View 1 +PanoHead [1] Dual Encoder [4] Era3D [31] LGM [56] Our MV+LGM [56] FaceLift + +![](images/a0e7f84978759d7e7dda93d8bd10e6cb8e55621aaa464b9f31eed2265b5d2032.jpg) + +![](images/a9891b6749d1924b63e624b85486e7626742cf5d96ff787910ef8bef17784071.jpg) + +![](images/e0651557001a66bd835743f9bbd4ddfb8f2afd68bad881e500f925d7b48f052f.jpg) + +![](images/cb894c9741d5af94be99357a9541b1f968da9a5aab03fab507b18dfd33258d1a.jpg) + +![](images/004aad661a7572a1ba60a0599505327955565695f369f588cb38de89fbf46cc8.jpg) + +![](images/1538e0c98c0a05b0bdfaf295aeb40c2136b90bba5c176bf32b5b44dd3ee2862a.jpg) + +![](images/7b75797307eb674cffb0dba25868dcec53b25dba705a523ecb371df2bda04e41.jpg) + +![](images/c9cae66a05bfc0862caae8d76f273b4741c1571ddcec578f8b420e253aa3c336.jpg) + +![](images/faa508d719a709b37fb65f42a38a5291e89eccdd079921a80e09cdc43c983d1b.jpg) + +![](images/08e558f12b6b11539b893d7aca98695a6cea72839651422915b8b7dda5b69330.jpg) + +![](images/f2b978650ee2eb1ebc6285f69ce3f187ff16870a835fab61c172595f5cf2b234.jpg) + +![](images/b071742ae771d3f28ea0fdddbf10492ff18d603aeed5115c423771bd930bba9e.jpg) + +![](images/f627882566c8698574d62225529f51c69520146e7f051b141ed8a86c64b0c295.jpg) + +![](images/4d58b91842b19a59c910fdf5073fc6fb5516712408d9ded18859d6d00d3338fe.jpg) + +![](images/2e1d0837f706a9b08c4ab0f87b4538fa164fda9549434708f24f5b19da9cb0b9.jpg) + +![](images/3e39e3a2689d98d1769913a61af779a597f756aa387e342f041b090764edca3b.jpg) + +![](images/696f4d052ce3c5f19bf95c64f001978babcabf997d82908af7b9d839f0bcd72c.jpg) + +![](images/d58917e4655c08349c2094b04122ac403cc69c0bbf4b8f18e63fa3d2e85ef385.jpg) + +![](images/72da2da62ae289ed3e43f91e6c5a56bcb74fc726d7d7908b51d1c70e6955595e.jpg) + +![](images/05814bacd1f35b02ddfce6af1a0ba537fcf64dd5d453eedb3e16e1e85e51b66f.jpg) + +![](images/0212b1361b22cb910d7634c492aaae91cacf06ce4b1d613c990890104a6a9564.jpg) + +![](images/46e4388a529e6d40706c096f8d2775bd357b7a64aa279500e69af3824b9e0460.jpg) + +![](images/d3cf9e3ba4093efd3b33ec928ce15c2f6468ba2b19b78fc7674286b0d64a06e7.jpg) + +![](images/2aa4ce74db8d2d98f6910bf0327bb6bdadcf34e8325cf758f06036be5a3b81b1.jpg) +Figure 8. Visual comparison on in-the-wild data. FaceLift demonstrates great generalization ability and robustness towards in-the-wild images, provides realistic unseen view rendering results. Era3D [31] and LGM [56] generate 3D head representation in inaccurate shape. PanoHead [1] often creates severe artifacts on the back of the head and can not handle challenging hairstyles well. Dual Encoder [4] shows improved performance on reconstructing the back of the head but exhibits more pronounced identity loss. + +![](images/f37bf9c3434ce7c15b1274dc1172f907759074931b2f3c942c728774f0b777e8.jpg) + +![](images/172da7f7e278e703a269c702bd783b56d9abbce69276359ac717bc5daefd8ff2.jpg) + +![](images/d5ec3caa8da854c4b92fa715f1c32e590f1b172cd4dac059991d87e461296289.jpg) + +![](images/710e6f5cb150b5fc31ed817dc4f49d5752165dc91e4f9ee46ab0c84824167814.jpg) + +![](images/47c2863441c19e80b5aedf749d4dbabf62f9b6618d13d6cd235e39025d501fc1.jpg) +PanoHead [1] Dual Encoder [4] Era3D [31] LGM [56] Our MV+LGM [56] FaceLift + +![](images/ace96edd69222586fb87f8347f0863719cb572d97b979a9f4d92b387d391f481.jpg) + +![](images/b26439310673bc3889fa5791cd9d216ba2073ffd5d7890bb3200dd36a2795567.jpg) +Rendered View 2 + +![](images/56e54984847429d6c95620228b5c37ce816762af2e9f9fb2b9d7c4e459b4255a.jpg) + +![](images/c26be89bef57287162a4867cf1d90aabc0bf520cbe59d16d5522094c264b15a3.jpg) + +![](images/7066f7b2c616b077c658c3772b0cc4429b459becb137ba38340d398269f1cba8.jpg) + +![](images/ddf09e78a09547973e498c9f6610325bd8a0f3463556bc30b65c666bf1b9e8d9.jpg) + +![](images/da4ab45a31af054d60263a797f55f1fdea9cc121962a01793c2101d4bc962ddf.jpg) + +![](images/30f20024d208f23e6a6810bb257058d1fe955ccdcb7682109ecc8a2e266c9613.jpg) + +![](images/1737dc64f07ef3ed5c598bdab4e2485fbf36f624e45976982b9848a2b220ae4c.jpg) + +![](images/2a4dc6f3aed0654eb50d32d2a63dead013b0829af13487a75bc164186eeb4c62.jpg) + +![](images/de8e7144457bcc8acf398e4febcd56dd4d53b5fb644011ed9a52767fe2bf7468.jpg) + +![](images/e4fb5e0209f1dfd285657ba099204ed1fa59ed6dbe12ea97d7b0454e6c0d201d.jpg) + +![](images/438c22936c88de444262a16e1b303cdb8fa4e1be998c6b3046c83afa1c51252b.jpg) + +![](images/4a03516c85a51a9b9af704f43d7796eec20b9f79650a2d39321377486522779e.jpg) + +![](images/ccd0b9c4810f97b5138b72d27af938aabbf1d6f92a69aa52b3d18a674d579a1c.jpg) + +![](images/7c4715b6f10acfb1b97a2ed03f2b6778f70052155dab2044d364025dea44717d.jpg) + +![](images/09ee325345a2110047a36baad93a88a19ec3f70b1c948050a89e9b86940e22a5.jpg) + +![](images/cb2cf711bd9d8917bc7a8245c3393acf14bbeb0dea8d1415ce422cceb5904123.jpg) + +![](images/561bd2eb28cc3aef32e81ff8a88b25dda3c6478c6d52168a6b247cff83374acc.jpg) + +![](images/9d71026ac245a36a82b48ad3d4ed357ba863c22dbcf5b8f61a8891df941b285b.jpg) + +![](images/099fdb90dea4a100c4550f464f03d459e772735877fe98ca812b5033d7bda469.jpg) + +![](images/3d4d5023c1837e395efcf3a5c7ac29203d5221a45df07dd4437f49bf435dcce9.jpg) + +![](images/e683f50c292b647db5c575035bab36fb43412c8b3f05847fbe142c4508d581b4.jpg) +Input + +![](images/42170feec4828d1edb96b6c450a01e42b79e40ca51819a9474c34962adad3f6d.jpg) + +![](images/8b33b5ae6bdf1db4ce89133b415ee9846e52bdb3115a327c54d8f23012c5c99f.jpg) + +![](images/819e68b1235ba896362723419c8cc731982f6d4871b4840efe150517ecd6f469.jpg) +Figure 9. Results of FaceLift on in-the-wild images. FaceLift accurately reconstructs 3D head models under challenging lighting conditions, achieving high fidelity (row 1). It captures fine facial details such as wrinkles (row 2), mustaches (row 3), and individual hairs (row 2 and row 4). Additionally, it remains robust to complex facial expressions (row 3) and various skin tones (row 4). Furthermore, it can realistically reconstruct facial paint (row 4). More results are provided in the supplementary materials. + +![](images/31b9b6ac088ee01197536f8c4cf062a39dce10db1351de3dcfe95bffa6cf855c.jpg) +Novel View Renderings + +![](images/7430f188b9bec33669050f87230094e1a2524890bd260facc91ac1a8fe599001.jpg) + +![](images/7ff410bb307ecbddb194e78b5765e7674840f056624b9e72656264fdfc1c0126.jpg) + +![](images/d39c3f9f266a4a8a7093718cb754410db28d686221570ccbfdbe815246d2024b.jpg) + +![](images/34ae4dcb80d9089ae45d3bf989b1c6f0eb257eaffc1a0f56aa98aec975e8e7fb.jpg) + +![](images/683198bc99409c4fcc6863731d3f51f4666df54bc70be6032f19bced133ff30d.jpg) + +![](images/e92968777d293309531cc31cf7abe476fe5428c6a0a55f910450a09bc4b3fc91.jpg) + +![](images/f9ca2e1dcb33c7f2df878ac41e053cb67481f0cb8c7ca68938e5867573574167.jpg) + +![](images/c1648b172f1ab59b1b24a7e8bd1309e1bad1de862f0ff6cd765edce6db96c411.jpg) + +![](images/43aa0531600db240ce378e530d06040e6fe0d2337b660ffd362a46f2f9199ab5.jpg) + +![](images/3f1f0c8c68e741e2c63a4825dad23ca25c7b49785138ffbcb1b40e4af4961102.jpg) + +![](images/8941c9ea4b80dddf5f0ef6bb2ac47ae90a1f90ef6631159075c286a445eaa09d.jpg) + +![](images/d72b93463faa8e38d78ec8ecb6768e971317bcdd882f1650c99ce80ad3a31bb1.jpg) +Input + +![](images/9efa900195c53c3b5d44bdeb74d2d59a2d3c717b2fdb144950a482519b60d3d0.jpg) + +![](images/f7dbb95ae514a93278cc78a94fc10693c895095d85fe4cf11de5d27f27cd8985.jpg) + +![](images/87cde82c9a07dac1f957010a830505824604793aab71805061bbd7fb385c454d.jpg) + +![](images/a2b6ffcdbe0dd09b893bcacbf1d277be6ead5309d2e6bdb449e7abc7828eca15.jpg) + +![](images/717aa3c10193708c1b500a117a3960f9b232e022038a0f92b248ecff098bce3a.jpg) + +![](images/5af16ca856d791c86904c9e7c58edb17bf5a21a5fad59e4135d05f832ff28613.jpg) + +![](images/aea72c41915ba4a76ede0be99b6097c3d31b134ed8f9a5b9a5510324e84691f3.jpg) + +![](images/a9fa10bb12b4e86b4e2a5b9554652d13d4b5be78baa8a5d642861e887ded452b.jpg) + +![](images/c2635d201159ca66dfd55b6dca23cab5c5fd13d180414a0405816a4146bec799.jpg) + +![](images/b6aa15dfd416726563ae38f384d19e01503a354be8033458091da1279eed2119.jpg) + +![](images/b7c5b1a7728821d32edc5ea79be51b769a259ad9c1df9d44feed7e060c0f8498.jpg) + +![](images/c84e8a014be493022565309f86b61032cc929027898b7e4ef17e288082d0c5ce.jpg) +Novel View Renderings + +![](images/85268e301c7331bb8d3cec1496bab75a3fd89334a98b307c18c8376e7f107e41.jpg) + +![](images/d79dbaaa377eb9fd4682ec3754d698fc68c94dd35cec04bba5889d84a7d1e356.jpg) + +![](images/779e4cd33e63249246a59662c223cee50077538ef23cba0b3eb77f8ce929f627.jpg) + +back-of-head rendering but suffers from severe identity loss (row 2) and fails to accurately reconstruct face paint (row 4). Era3D [31] often produces an inaccurate head shape, particularly from side views, and offers fewer geometric details compared to FaceLift. LGM [56] generates Gaussians with inaccurate color and opacity and lacks proper facial geometry, resulting in distorted features. Baseline Our $MV + LGM$ + +shows that our multi-view diffusion model enhances LGM by providing improved facial geometry and texture details. However, the LGM reconstructor still produces Gaussians with inaccurate shapes and opacities, further underscoring the advantages of our transformer-based reconstructor. + +We present more FaceLift's novel view rendering results in Fig. 9 to demonstrate FaceLift's ability to produce high + +![](images/845b35b8900bed5b992a65264b8854aebec4973ae4d4a4e5b4220c0fff50502d.jpg) +Input +w/o. Input View Reconstruction + +![](images/731b4add7905f0ca8004777d6d4fbb9e6dfd22b57b398b42dfd8cabbacbe1541.jpg) + +![](images/6242894fee37ee8eaf33967c03081fb726dbc2b955e49d5adb5f583366a28a8e.jpg) +w. Input View Reconstruction +Figure 10. Importance of input view reconstruction. The diffusion model that is not trained to perform the input view reconstruction, i.e., w/o. Input View Reconstruction, overfits to synthetic training distribution, suffers from severe identity loss during inference. Trained with input view reconstruction, our method preserves the input identity and expression faithfully. + +fidelity, realistic 3D head reconstructions with intricate details across a variety of challenging scenarios. FaceLift effectively handles faces under various lighting conditions. It can especially render realistic novel view images given a photo captured with an iPhone under dark lighting conditions (row 1 column 1), emphasizing its robustness and potential for real-world application. It reconstructs facial details with high fidelity, especially the wrinkles and folds on the face caused by extreme expression. FaceLift also excels at reconstructing challenging textures, such as mustaches and hair. Furthermore, it faithfully reconstructs facial paint, despite such data not being included in our synthetic face dataset, showcasing its strong generalization ability. + +# 5. Ablation Study + +# 5.1. Input View Reconstruction + +We conduct an ablation study to demonstrate the importance of reconstructing the input view during training. For comparison, we train a multi-view diffusion model that generates six novel views. In this baseline, the first generated view's elevation is adjusted from $0^{\circ}$ to $20^{\circ}$ , while the remaining views adopt the same camera poses as in our default setting. We refer to this variant as w/o. Input View Reconstruction. Fig. 10 presents the view generation results of the two diffusion models when applied to real-world images. Without the input view reconstruction task, the model trained on the synthetic dataset generates views within a limited distribution, leading to noticeable identity loss. Moreover, it loses its ability to preserve facial expressions and face paint. In contrast, incorporating the input view reconstruction task during training enables our diffusion model to faithfully regenerate the input view, significantly improving its generalization ability. Quantitative comparison is provided in the supplementary material. + +# 5.2. Number of Views + +We evaluate three configurations: four views (front, left, right, and back), six views (adding front-left and front-right), and eight views (further including front-top and front-bottom). Fig. 11 compares the baselines using different numbers of input views. With only four views, the reconstructor fails to capture a complete forehead; however, + +![](images/1451d24a8979813180aaf4d9c677cf29adfd574559f477841755d545a198f77e.jpg) +Input +Figure 11. Number of input views of Gaussian reconstructor. Using six views strike a good balance between reconstruction quality and computational efficiency. + +![](images/ebea9c0089956936d4ff1c3860e7348c481971082a3e3a3f3c2cf5ed2e74c72c.jpg) +4 Views + +![](images/086d8465aef1302223669b97f016f39874388dd9f05c1b3686b8f8464221a73d.jpg) +6 Views + +![](images/2bcdc2093d6fdacb84b6595a329e3f1d4622a4ef56cbff5e94682bb68a0929b2.jpg) +8 Views + +![](images/56c6e3265cb6279cc3cbc601089dfc57cab62c215e8b192ed1061af25cd27787.jpg) +4 Views +6 Views + +![](images/da9d4abf0b20d79e2738c4cf89d56fb493257d806db2d8e9874b911e4fc72377.jpg) +8 Views + +![](images/be77612c1ac35f9ebb8413fe72ce03b65ab200b4772b39f31d749c25f2d3ecc8.jpg) + +![](images/565043e59b9872106b383935b4a352ac74896a625eea24ce37e046517f146b6d.jpg) +Input +Figure 12. Two-stage training of reconstructor. Without pretraining on general objects, the reconstructor fails to produce clear textures in the reconstruction results. Meanwhile, without finetuning on synthetic human head data, the model lacks a refined understanding of facial structures, including the eyes and nose. + +![](images/b64979970ac3b38502e399df428f969f3c505664be9d449950e26c9af8a79618.jpg) +w/o. Pre-training + +![](images/8b5535165c67f0ec8061d93797ac56f96a5d564889c666f7ad2a4bab4343df01.jpg) +w/o. Fine-tuning + +![](images/d1ae0f9ad341904144bd05b4957a41a44ff91b0a7ad3e71407835d7af36011e7.jpg) +Ours + +with six views, it reconstructs the eyes and eyebrows more smoothly and renders challenging textures—such as facial wrinkles and ear folds—more realistically. Eight views do not offer significant visual improvements, and incur a higher computational cost in both stages. Thus, we find that six views achieve a good balance between reconstruction quality and computational efficiency. + +# 5.3. Two-stage Reconstructor Training + +As illustrated in Sec. 3.3, our Gaussian reconstructor follows a two-stage training pipeline. Fig. 12 shows that pretraining on general objects helps the model learn a diverse prior of geometric structures, resulting in clearer textures on delicate facial regions. Meanwhile, fine-tuning on synthetic human head data enables the reconstructor to gain a more refined understanding of facial structure, thereby enhancing the accuracy of features such as the eyes, nose, and hair. + +# 6. Conclusions + +We propose FaceLift, a feed-forward approach that lifts a single facial image to a detailed 3D reconstruction with preserved identity features. Our method uses multi-view diffusion to generate unobservable views and employs a transformer-based reconstructor to reconstruct 3D Gaussian splats, enabling high-quality novel view synthesis. To overcome the difficulty of capturing real-world multi-view human head images, we render high-quality synthetic data for training and show that, despite being trained solely on synthetic data, FaceLift can reconstruct 3D heads from real-world captured images with high fidelity. Compared with baselines [1, 4, 29, 31, 56], FaceLift generates 3D head representation with finer geometry and texture details and exhibits better identity preservation ability. + +# 7. Acknowledgement + +This was supported in part by the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korean Government (MSIT) (No. RS-2024-00457882, National AI Research Lab Project). + +We appreciate the insightful discussions with Kai Zhang, Hao Tan, Zexiang Xu, Sai Bi, Sumit Chaturvedi, Hanwen Jiang, Yu-Ju Tsai, Kuan-Chih Huang, Chengxu Liu, and Dingyi Dai. We thank Nathan Carr and Kalyan Sunkavalli for their support. + +# References + +[1] Sizhe An, Hongyi Xu, Yichun Shi, Guoxian Song, Umit Y Ogras, and Linjie Luo. PanoHead: Geometry-aware 3D fullhead synthesis in $360^{\circ}$ . In CVPR, 2023. 1, 3, 5, 6, 7, 8 +[2] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 5 +[3] Ananta R. Bhattarai, Matthias Nießner, and Artem Sevastopolsky. TriPlaneNet: An encoder for EG3D inversion. In WACV, 2024. 2 +[4] Bahri Batuhan Bilecen, Ahmet Berke Gokmen, and Aysegul Dundar. Dual encoder GAN inversion for high-fidelity 3D head reconstruction from single images. In NeurIPS, 2024. 5, 6, 7, 8 +[5] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3D faces. In SIGGRAPH, 1999. 2 +[6] James Booth, Anastasios Roussos, Stefanos Zafeiriou, Allan Ponniah, and David Dunaway. A 3D morphable model learnt from 10,000 faces. In CVPR, 2016. 2 +[7] Marcel C Bühler, Kripasindhu Sarkar, Tanmay Shah, Gengyan Li, Daoye Wang, Leonhard Helminger, Sergio Orts-Escalano, Dmitry Lagun, Otmar Hilliges, Thabo Beeler, et al. Preface: A data-driven volumetric prior for few-shot ultra high-resolution face synthesis. In ICCV, 2023. 3 +[8] Marcel C Bühler, Gengyan Li, Erroll Wood, Leonhard Helminger, Xu Chen, Tanmay Shah, Daoye Wang, Stephan Garbin, Sergio Orts-Escolano, Otmar Hilliges, et al. Cafca: High-quality novel view synthesis of expressive faces from casual few-shot captures. In SIGGRAPH Asia, 2024. 2, 3, 5, 12, 15, 17 +[9] Sumit Chaturvedi, Mengwei Ren, Yannick Hold-Geoffroy, Jingyuan Liu, Julie Dorsey, and Zhixin Shu. SynthLight: Portrait relighting with diffusion model by learning to re-render synthetic faces. In CVPR, 2025. 3 +[10] Xiyi Chen, Marko Mihajlovic, Shaofei Wang, Sergey Prokudin, and Siyu Tang. Morphable diffusion: 3d-consistent diffusion for single-image avatar creation. In CVPR, 2024. 3 +[11] Antonia Creswell and Anil Anthony Bharath. Inverting the generator of a generative adversarial network. In TNNLS, 2018. 3 +[12] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana + +Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3D objects. In CVPR, 2023. 2, 3, 5 +[13] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-XL: A universe of $10\mathrm{M} + 3\mathrm{D}$ objects. In NeurIPS, 2024. 3 +[14] Jiankang Deng, Jia Guo, Jing Yang, Niannan Xue, Irene Kotsia, and Stefanos Zafeiriou. ArcFace: Additive angular margin loss for deep face recognition. In IEEE TPAMI, 2022. 5, 6 +[15] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 3 +[16] Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. DreamSim: Learning new dimensions of human visual similarity using synthetic data. In NeurIPS, 2023. 5, 6 +[17] Chen Gao, Yichang Shih, Wei-Sheng Lai, Chia-Kai Liang, and Jia-Bin Huang. Portrait neural radiance fields from a single image. arXiv preprint arXiv:2012.05903, 2020. 3 +[18] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. GET3D: A generative model of high quality 3D textured shapes learned from images. In NeurIPS, 2022. 3 +[19] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 1, 2, 3 +[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5 +[21] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 16 +[22] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 15 +[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1, 3 +[24] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: Large reconstruction model for single image to 3D. In ICLR, 2024. 3, 5 +[25] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 1, 2, 3 +[26] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In CVPR, 2020. 2, 3 +[27] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D gaussian splatting for real-time radiance field rendering. In ACM TOG, 2023. 1, 3 +[28] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In ICLR, 2014. 3 +[29] Tobias Kirschstein, Simon Giebenhain, Jiapeng Tang, Markos Georgopoulos, and Matthias Nießner. GGHead: Fast and generalizable 3D gaussian heads. In SIGGRAPH Asia, 2024. 5, 6, 8 + +[30] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3D: Fast text-to-3D with sparse-view generation and large reconstruction model. In ICLR, 2024. 3 +[31] Peng Li, Yuan Liu, Xiaoxiao Long, Feihu Zhang, Cheng Lin, Mengfei Li, Xingqun Qi, Shanghang Zhang, Wenhan Luo, Ping Tan, et al. Era3D: High-resolution multiview diffusion using efficient row-wise attention. In NeurIPS, 2024. 3, 5, 6, 7, 8 +[32] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. In ACM TOG, 2017. 1, 2 +[33] Xueting Li, Shalini De Mello, Sifei Liu, Koki Nagano, Umar Iqbal, and Jan Kautz. Generalizable one-shot 3D neural head avatar. In NeurIPS, 2024. 2 +[34] Jiangke Lin, Yi Yuan, Tianjia Shao, and Kun Zhou. Towards high-fidelity 3D face reconstruction from in-the-wild images using graph convolutional networks. In CVPR, 2020. 2 +[35] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T., Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3D mesh in 45 seconds without pershape optimization. In NeurIPS, 2023. 3 +[36] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In ICCV, 2023. 3 +[37] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3D: Single image to 3D using cross-domain diffusion. In CVPR, 2024. 3 +[38] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Guang Yong, Juhyun Lee, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, 2019. 5 +[39] Julieta Martinez, Emily Kim, Javier Romero, Timur Bagautdinov, Shunsuke Saito, Shouou-I Yu, Stuart Anderson, Michael Zollhöfer, Te-Li Wang, Shaojie Bai, Chenghui Li, Shih-En Wei, Rohan Joshi, Wyatt Borsos, Tomas Simon, Jason Saragih, Paul Theodosis, Alexander Greene, Anjani Josyula, Silvio Mano Maeta, Andrew I. Jewett, Simon Venshtain, Christopher Heilman, Yueh-Tung Chen, Sidi Fu, Mohamed Ezzeldin A. Elshaer, Tingfang Du, Longhua Wu, Shen-Chi Chen, Kai Kang, Michael Wu, Youssef Emad, Steven Longay, Ashley Brewer, Hitesh Shah, James Booth, Taylor Koska, Kayla Hidle, Matt Andromalos, Joanna Hsu, Thomas Dauer, Peter Selednik, Tim Godisart, Scott Ardisson, Matthew Cipperly, Ben Humberston, Lon Farr, Bob Hansen, Peihong Guo, Dave Braun, Steven Krenn, He Wen, Lucas Evans, Natalia Fadeeva, Matthew Stewart, Gabriel Schwartz, Divam Gupta, Gyeongsik Moon, Kaiwen Guo, Yuan Dong, Yichen Xu, Takaaki Shiratori, Fabian Prada, Bernardo R. Pires, Bo Peng, Julia Buffalini, Autumn Trimble, Kevyn McPhail, Melissa Schoeller, and Yaser Sheikh. Codec Avatar Studio: Paired human captures for complete, driveable, and generalizable avatars. In NeurlPS, 2024. 2, 3, 5, 15, 17 + +[40] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 3 +[41] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. StyleSDF: High-resolution 3D-consistent image and geometry generation. In CVPR, 2022. 3 +[42] Foivos Paraperas Papantoniou, Alexandros Lattas, Stylianos Moschoglou, and Stefanos Zafeiriou. Relightify: Relightable 3D faces from a single image via diffusion models. In ICCV, 2023. 3 +[43] Dario Pavllo, Graham Spinks, Thomas Hofmann, Marie-Francine Moens, and Aurélien Lucchi. Convolutional generation of textured 3D meshes. In NeurIPS, 2020. 3 +[44] Julius Plucker. XVii. on a new geometry of space. Philosophical Transactions of the Royal Society of London, 1865. 4 +[45] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D diffusion. In ICLR, 2023. 3 +[46] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3D object generation using both 2D and 3D diffusion priors. In ICLR, 2024. 3 +[47] Elad Richardson, Matan Sela, and Ron Kimmel. 3D face reconstruction by learning from synthetic data. In 3DV, 2016. 2 +[48] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3 +[49] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 16 +[50] Shunsuke Saito, Gabriel Schwartz, Tomas Simon, Junxuan Li, and Giljoo Nam. Relightable gaussian codec avatars. In CVPR, 2024. 6 +[51] Kyle Sargent, Zizhang Li, Tanmay Shah, Charles Herrmann, Hong-Xing Yu, Yunzhi Zhang, Eric Ryan Chan, Dmitry Lagun, Li Fei-Fei, Deqing Sun, et al. ZeroNVS: Zero-shot 360-degree view synthesis from a single real image. In CVPR, 2024. 3 +[52] Sefik Ilkin Serengil and Alper Ozpinar. HyperExtended LightFace: A facial attribute analysis framework. In ICEET, 2021. 5 +[53] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: A single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 3 +[54] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 3, 16 +[55] Wenqiang Sun, Shuo Chen, Fangfu Liu, Zilong Chen, Yueqi Duan, Jun Zhang, and Yikai Wang. DimensionX: Create any 3D and 4D scenes from a single image with controllable video diffusion. arXiv preprint arXiv:2411.04928, 2024. 12 + +[56] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. LGM: Large multi-view gaussian model for high-resolution 3D content creation. In ECCV, 2024. 3, 5, 6, 7, 8, 17, 18 +[57] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. DreamGaussian: Generative gaussian splatting for efficient 3D content creation. In ICLR, 2024. 3 +[58] Aäron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning. In NeurIPS, 2017. 3 +[59] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 3, 5 +[60] Thomas Vetter and Volker Blanz. Estimating coloured 3D face models from single images: An example based approach. In ECCV, 1998. 1, 2 +[61] Vishal Vinod, Tanmay Shah, and Dmitry Lagun. TEGLO: High fidelity canonical texture mapping from single-view images. In WACV, 2024. 3 +[62] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In NeurIPS, 2021. 3 +[63] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3D digital avatars using diffusion. In CVPR, 2023. 3 +[64] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. ProlificDreamer: High-fidelity and diverse text-to-3D generation with variational score distillation. In NeurIPS, 2024. 3 +[65] Erroll Wood, Tadas Baltrusaitis, Charlie Hewitt, Sebastian Dziadzio, Thomas J Cashman, and Jamie Shotton. Fake it till you make it: Face analysis in the wild using synthetic data alone. In ICCV, 2021. 3 +[66] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3D generative-adversarial modeling. In NeurIPS, 2016. 3 +[67] Kailu Wu, Fangfu Liu, Zhihan Cai, Runjie Yan, Hanyang Wang, Yating Hu, Yueqi Duan, and Kaiheng Ma. Unique3d: High-quality and efficient 3d mesh generation from a single image. arXiv preprint arXiv:2405.20343, 2024. 12, 13 +[68] Yiqian Wu, Jing Zhang, Hongbo Fu, and Xiaogang Jin. LPFF: A portrait dataset for face generators across large poses. In ICCV, 2023. 1 +[69] Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin Tong, and Jiaolong Yang. Structured 3d latents for scalable and versatile 3d generation. In CVPR, 2024. 12, 13 +[70] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024. 12, 13 + +[71] Yu-Ying Yeh, Koki Nagano, Sameh Khamis, Jan Kautz, Ming-Yu Liu, and Ting-Chun Wang. Learning to relight portrait images via a virtual light stage and synthetic-to-real adaptation. In ACM TOG, 2022. 3 +[72] Ziyang Yuan, Yiming Zhu, Yu Li, Hongyu Liu, and Chun Yuan. Make encoder great again in 3D GAN inversion through geometry and occlusion-aware encoding. In ICCV, 2023. 1, 2, 3 +[73] Bowen Zhang, Yiji Cheng, Chunyu Wang, Ting Zhang, Jiaolong Yang, Yansong Tang, Feng Zhao, Dong Chen, and Baining Guo. RodinHD: High-fidelity 3D avatar generation with diffusion models. In ECCV, 2024. 2, 3 +[74] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. GS-LRM: Large reconstruction model for 3D gaussian splatting. In ECCV, 2024. 3, 5, 17 +[75] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5 +[76] Xiaozheng Zheng, Chao Wen, Zhaohu Li, Weiyi Zhang, Zhuo Su, Xu Chang, Yang Zhao, Zheng Lv, Xiaoyuan Zhang, Yongjie Zhang, Guidong Wang, and Xu Lan. Headgap: Few-shot 3d head avatar via generalizable gaussian priors. In 3DV, 2025. 3 \ No newline at end of file diff --git a/2025/FaceLift_ Learning Generalizable Single Image 3D Face Reconstruction from Synthetic Heads/images.zip b/2025/FaceLift_ Learning Generalizable Single Image 3D Face Reconstruction from Synthetic Heads/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4d18205098f16950d86f29c17743dde3ab494d2d --- /dev/null +++ b/2025/FaceLift_ Learning Generalizable Single Image 3D Face Reconstruction from Synthetic Heads/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:47c21cfd606c2a0df2473489d161ec5882e67755c3702736e5c806b450a83fae +size 657142 diff --git a/2025/FaceLift_ Learning Generalizable Single Image 3D Face Reconstruction from Synthetic Heads/layout.json b/2025/FaceLift_ Learning Generalizable Single Image 3D Face Reconstruction from Synthetic Heads/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..64f2874d265217d1b47b652d19ecf409df5194e1 --- /dev/null +++ b/2025/FaceLift_ Learning Generalizable Single Image 3D Face Reconstruction from Synthetic Heads/layout.json @@ -0,0 +1,13514 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 102, + 541, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 102, + 541, + 138 + ], + "spans": [ + { + "bbox": [ + 67, + 102, + 541, + 138 + ], + "type": "text", + "content": "FaceLift: Learning Generalizable Single Image 3D Face Reconstruction from Synthetic Heads" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 152, + 160, + 457, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 160, + 457, + 176 + ], + "spans": [ + { + "bbox": [ + 152, + 160, + 457, + 176 + ], + "type": "text", + "content": "Weijie Lyu" + }, + { + "bbox": [ + 152, + 160, + 457, + 176 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 152, + 160, + 457, + 176 + ], + "type": "text", + "content": " Yi Zhou" + }, + { + "bbox": [ + 152, + 160, + 457, + 176 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 152, + 160, + 457, + 176 + ], + "type": "text", + "content": " Ming-Hsuan Yang" + }, + { + "bbox": [ + 152, + 160, + 457, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 152, + 160, + 457, + 176 + ], + "type": "text", + "content": " Zhixin Shu" + }, + { + "bbox": [ + 152, + 160, + 457, + 176 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 174, + 180, + 436, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 180, + 436, + 194 + ], + "spans": [ + { + "bbox": [ + 174, + 180, + 436, + 194 + ], + "type": "text", + "content": "1University of California, Merced 2Adobe Research" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 63, + 205, + 154, + 297 + ], + "blocks": [ + { + "bbox": [ + 63, + 205, + 154, + 297 + ], + "lines": [ + { + "bbox": [ + 63, + 205, + 154, + 297 + ], + "spans": [ + { + "bbox": [ + 63, + 205, + 154, + 297 + ], + "type": "image", + "image_path": "cfe088eee6b0315c59c76e4e0696b845e6a2ca48fbeffb44833afc0d3bfb61c1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 300, + 143, + 310 + ], + "lines": [ + { + "bbox": [ + 74, + 300, + 143, + 310 + ], + "spans": [ + { + "bbox": [ + 74, + 300, + 143, + 310 + ], + "type": "text", + "content": "Single Image Input" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 147, + 205, + 256, + 297 + ], + "blocks": [ + { + "bbox": [ + 147, + 205, + 256, + 297 + ], + "lines": [ + { + "bbox": [ + 147, + 205, + 256, + 297 + ], + "spans": [ + { + "bbox": [ + 147, + 205, + 256, + 297 + ], + "type": "image", + "image_path": "4c793c5ef31a912d348c6c0cc95f77ae46f7d17e7dc33b912fa9241149b36de4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 164, + 300, + 257, + 319 + ], + "lines": [ + { + "bbox": [ + 164, + 300, + 257, + 319 + ], + "spans": [ + { + "bbox": [ + 164, + 300, + 257, + 319 + ], + "type": "text", + "content": "High-fidelity 3D Gaussian Representation" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 260, + 205, + 347, + 297 + ], + "blocks": [ + { + "bbox": [ + 260, + 205, + 347, + 297 + ], + "lines": [ + { + "bbox": [ + 260, + 205, + 347, + 297 + ], + "spans": [ + { + "bbox": [ + 260, + 205, + 347, + 297 + ], + "type": "image", + "image_path": "dabe944618ae82f7db9e5729ef8c6fdb9b8b48395e9768665d7e5ccdb4c32e6d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 277, + 300, + 331, + 309 + ], + "lines": [ + { + "bbox": [ + 277, + 300, + 331, + 309 + ], + "spans": [ + { + "bbox": [ + 277, + 300, + 331, + 309 + ], + "type": "text", + "content": "Full Head NVS" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 326, + 555, + 371 + ], + "lines": [ + { + "bbox": [ + 55, + 326, + 555, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 555, + 371 + ], + "type": "text", + "content": "Figure 1. FaceLift transforms a single facial image into a high-fidelity 3D Gaussian head representation. Trained exclusively on synthetic 3D data, our pipeline first generates sparse, identity-preserving multiview images of the input head using a diffusion model. These sparse generated views are then fed into a transformer-based 3D Gaussian splats reconstructor, producing complete and detailed 3D head representation that generalize remarkably well to real-world human images. Project page: https://weijielyu.github.io/FaceLift." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 350, + 205, + 448, + 297 + ], + "blocks": [ + { + "bbox": [ + 350, + 205, + 448, + 297 + ], + "lines": [ + { + "bbox": [ + 350, + 205, + 448, + 297 + ], + "spans": [ + { + "bbox": [ + 350, + 205, + 448, + 297 + ], + "type": "image", + "image_path": "cab329f58b2a5b3e56294b1affd051cc3875f0fdf537c45b1b82e1be3e3c6cb4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 300, + 476, + 310 + ], + "lines": [ + { + "bbox": [ + 403, + 300, + 476, + 310 + ], + "spans": [ + { + "bbox": [ + 403, + 300, + 476, + 310 + ], + "type": "text", + "content": "High-quality Details" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 152, + 380, + 200, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 380, + 200, + 393 + ], + "spans": [ + { + "bbox": [ + 152, + 380, + 200, + 393 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 54, + 409, + 297, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 409, + 297, + 673 + ], + "spans": [ + { + "bbox": [ + 54, + 409, + 297, + 673 + ], + "type": "text", + "content": "We present FaceLift, a novel feed-forward approach for generalizable high-quality 360-degree 3D head reconstruction from a single image. Our pipeline first employs a multiview latent diffusion model to generate consistent side and back views from a single facial input, which then feed into a transformer-based reconstructor that produces a comprehensive 3D Gaussian splats representation. Previous methods for monocular 3D face reconstruction often lack full view coverage or view consistency due to insufficient multi-view supervision. We address this by creating a high-quality synthetic head dataset that enables consistent supervision across viewpoints. To bridge the domain gap between synthetic training data and real-world images, we propose a simple yet effective technique that ensures the view generation process maintains fidelity to the input by learning to reconstruct the input image alongside the view generation. Despite being trained exclusively on synthetic data, our method demonstrates remarkable generalization to real-world images. Through extensive qualitative and quantitative evaluations, we show that FaceLift outperforms state-of-the-art 3D face reconstruction methods on identity preservation, detail recovery and rendering quality." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 380, + 394, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 380, + 394, + 392 + ], + "spans": [ + { + "bbox": [ + 315, + 380, + 394, + 392 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 400, + 555, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 555, + 544 + ], + "type": "text", + "content": "3D face reconstruction has been a central focus in computer vision and graphics research for decades, driven by its crucial applications in immersive virtual and augmented realities, VFX and gaming, digital entertainment, and next-generation telepresence systems. However, achieving high quality reconstruction from a single image remains very challenging. On one hand, the monocular face reconstruction problem is highly ill-posed – a single 2D image can be produced by countless different 3D face shapes, creating fundamental ambiguity. On the other hand, the human visual system is highly attuned to facial details, making even subtle artifacts and imperfections noticeable to the eye." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": "Traditional 3D head synthesis approaches typically use parametric textured mesh models [32, 60] trained on 3D scan datasets. While these models enable basic head generation, the rendered images frequently lack fine-scale geometric detail, realistic textures, and convincing hair, limiting their perceptual realism and expressiveness. Recent breakthroughs in image generative models [19, 23] and novel view synthesis techniques [27, 40] have opened new possibilities for this research area. Leveraging these developments, recent works [1, 72] use neural 3D representations to learn effective 3D head representation from unstructured real face image datasets [25, 68]. While these datasets improve the realism and diversity of rendering results, they fail to provide multi-view supervision for modeling 3D consis" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 292, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 292, + 712 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 292, + 712 + ], + "type": "text", + "content": "*Work was done when Weijie Lyu was an intern at Adobe Research." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12691" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 293, + 268 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 293, + 268 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 293, + 268 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 293, + 268 + ], + "type": "image", + "image_path": "9714794b1877a26ae2d9a09376437b1f1598dca3e1e561b53c0e802b8ef606e4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 270, + 295, + 314 + ], + "lines": [ + { + "bbox": [ + 55, + 270, + 295, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 295, + 314 + ], + "type": "text", + "content": "Figure 2. Comparison with RodinHD. RodinHD [73] trains triplane diffusion with synthetic data, resulting in apparent identity loss. In contrast, FaceLift achieves better identity preservation and generalizes effectively to real human portraits." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 318, + 296, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 318, + 296, + 402 + ], + "spans": [ + { + "bbox": [ + 55, + 318, + 296, + 402 + ], + "type": "text", + "content": "tency causing view inconsistency and artifacts on the back of the head. Since diverse multi-view real images are hard to acquire, RodinHD [73] leverages synthetic multi-view images to train generative models that directly output 3D neural representations of the head. However, training solely on synthetic data often results in significant perceptual identity loss in the generated outputs, as demonstrated in Fig. 2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 402, + 296, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 402, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 402, + 296, + 689 + ], + "type": "text", + "content": "In this work, we present FaceLift, a pipeline for learning generalizable and high-fidelity single image to 3D face representation from synthetic head data. To achieve high quality reconstruction that preserves the input facial identity, we adopt a two-stage pipeline to first generate identity preserving multi-view images using a diffusion model [48], followed by a transformer-based feed-forward reconstructor to fuse the generated sparse views into a comprehensive 3D Gaussian representation. We train the model with synthetic images - multi-view renderings of 3D synthetic human portraits using Blender. We highlight two key techniques for generalizing to real-world images and preserving input facial identity. First, we emphasize the importance of reconstructing the input image alongside the view synthesis task in the conditional diffusion model training, which significantly improved generalization capability in testing. Second, we demonstrate that training the feed-forward reconstructor benefits from a two-stage training process: pretraining on general objects [12] to acquire a rich geometry and texture prior, followed by fine-tuning on synthetic human head data to capture head-specific geometry. With our two-stage approach, we focus on learning identity preservation in the image space during the first stage, achieving higher input fidelity compared to existing methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Comparing with prior art, we achieve three key advancements: (1) robust view consistency through multi-view at" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "tention and supervision, (2) improved generalization from training techniques and foundational model, ensuring accurate identity preservation, (3) high-quality facial texture and hair details via pixel-aligned Gaussian representation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 121, + 555, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 555, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 555, + 263 + ], + "type": "text", + "content": "We extensively evaluate FaceLift quantitatively and qualitatively across diverse datasets. Using real multiview studio captures [39] and an independent synthetic human dataset [8], our approach consistently surpasses previous state-of-the-art methods across all evaluation metrics. Through extensive testing on in-the-wild portrait images, we demonstrate that our method reconstructs complete 3D heads with fine-grained details, accurate identity preservation, and high visual fidelity. Comparisons and ablation studies confirm that multi-view consistent training data is crucial for high-fidelity face reconstruction. Our contributions are summarized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 266, + 553, + 445 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 314, + 266, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 266, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 314, + 266, + 553, + 300 + ], + "type": "text", + "content": "- We propose FaceLift, a framework that reconstructs a high-fidelity 3D head from a single image using view generation and feed-forward reconstructor." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 301, + 553, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 301, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 314, + 301, + 553, + 348 + ], + "type": "text", + "content": "- Despite being trained solely on synthetic human head data, our approach shows no domain gap on real-world images, highlighting both the effectiveness of synthetic data and our model's robust generalization capabilities." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 350, + 553, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 350, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 314, + 350, + 553, + 408 + ], + "type": "text", + "content": "- We construct two benchmarks on single-image to 3D full head reconstruction tasks using the publicly available datasets Cafca [8] and Ava-256 [39] to quantitatively evaluate models' performance on both reconstruction accuracy and identity preservation ability." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 409, + 553, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 409, + 553, + 445 + ], + "spans": [ + { + "bbox": [ + 314, + 409, + 553, + 445 + ], + "type": "text", + "content": "- Our comprehensive evaluation confirms that our approach achieves state-of-the-art performance in reconstruction accuracy and identity preservation." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 456, + 400, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 456, + 400, + 468 + ], + "spans": [ + { + "bbox": [ + 314, + 456, + 400, + 468 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "text", + "content": "Face Reconstruction. 3D face reconstruction has been a long-standing challenge in computer vision, with substantial progress driven by diverse approaches. Vetter and Blanz [60] pioneer a method for synthesizing 3D faces by linearly blending multiple 3D templates, now widely known as blendshapes. This work establishes the foundation for 3D Morphable Models (3DMMs), which represent 3D face shapes and textures as principal components derived from scanned data. Subsequent research [5, 6, 32, 34, 47] extend this framework, enabling the generation of new 3D faces by manipulating blending coefficients. However, these methods produce mesh-based representations that lack fine details and are limited to modeling the front of the face, excluding hair and 360-degree synthesis. While 3DMM-based methods have been foundational, recent advances in deep learning, especially Generative Adversarial Networks (GANs) [19, 25, 26], have greatly improved 3D face synthesis quality. EG3D [72] uses a tri-plane NeRF representation with a pose-conditioned StyleGAN2 [26] framework. Follow-up works [3, 33] achieve single-image-to-3D gen" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12692" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 346 + ], + "type": "text", + "content": "eration through GAN inversion [11]. Despite their success, these methods can only synthesize near-frontal views. To overcome this, PanoHead [1] introduces a tri-grid neural volume representation, enabling full 360-degree head synthesis. Unfortunately, it does not provide a 3D head representation for consistent multi-view rendering. Recent efforts explore alternative representations for 3D face reconstruction from sparse input, such as a single image [17, 42, 61] or few-shot images [7]. However, these methods still require pre-instance optimization. Rodin [63] and its extension RodinHD [73] employ an image-conditioned diffusion model to generate a triplane representation of a human head for full-head novel view synthesis. Nevertheless, their triplane diffusion model is limited to synthetic data and struggles to achieve high-fidelity reconstructions from real-world images. For animatable 3D head avatars generations, Morphable Diffusion [10] generates multi-view consistent images from a single image using a morphable mesh, while HeadGAP [76] generates 3D animatable head avatars using few-shot input, leveraging 3D head priors learned from large-scale data. In contrast, our work focuses on leveraging synthetic training data to produce high-fidelity, detailed 3D Gaussian head models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 351, + 294, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 351, + 294, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 351, + 294, + 495 + ], + "type": "text", + "content": "Synthetic Human Data. Capturing high-quality 3D data of real humans requires a controlled studio environment and costly photography equipment [39]. As an alternative, large-scale synthetic 3D head datasets have emerged as an effective and resource-efficient solution for tasks like human head reconstruction [8, 63, 65, 73] and photorealistic relighting [9, 71], offering a scalable way to train models without the restrictions of real-world data acquisition. Inspired by these previous works, we aim to use synthetic data to improve the model's understanding of the human head and minimize the generalization gap between synthetic data training and real-world inference." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "text", + "content": "Image or Text to 3D. Generative models have achieved remarkable success in 2D image generation with VAEs [28, 58], GANs [19, 25, 26], and diffusion models [23, 48, 54]. Building on this success, extensive research has extended these models to 3D content generation [18, 41, 43, 66]. Starting with DreamFusion [45], numerous works [36, 46, 51, 57, 64] try to distill NeRF [40, 62, 72] or 3D Gaussians [27] representation from 2D image diffusion using a Score Distillation Sampling (SDS) loss. These methods can produce high-quality results but often encounter challenges such as slow optimization, over-saturated colors, and the Janus problem. To overcome these challenges, recent works [30, 31, 35, 37, 53] generate multi-view images with high consistency, which can be directly used for 3D reconstruction with neural reconstruction methods [27, 40, 62]. However, optimizing NeRF or NeuS remains far from real-time performance. Recent advancements in large 3D reconstruction models (LRMs) [24, 30, 56, 74] offer a path" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": "way to faster 3D reconstruction. Leveraging scalable transformer architectures [15, 59] and large datasets like Objverse [12, 13], these models effectively capture generalizable 3D priors. Unlike traditional pre-scene optimization methods [27, 40, 62], LRMs employ a feed-forward approach, enabling the prediction of high-quality NeRF, mesh, or 3D Gaussian representations from sparse images in under a second. However, most of these research efforts are applied to general objects, with limited or suboptimal results for 3D head reconstruction." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 199, + 420, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 199, + 420, + 213 + ], + "spans": [ + { + "bbox": [ + 313, + 199, + 420, + 213 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 217, + 553, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 553, + 491 + ], + "type": "text", + "content": "As shown in Fig. 3, given a single frontal image of a human face " + }, + { + "bbox": [ + 313, + 217, + 553, + 491 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 217, + 553, + 491 + ], + "type": "text", + "content": ", our goal is to reconstruct a complete 3D head " + }, + { + "bbox": [ + 313, + 217, + 553, + 491 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 217, + 553, + 491 + ], + "type": "text", + "content": ", represented as 3D Gaussian splats, with detailed facial texture and preserved identity. This requires our system to have prior knowledge on the geometry structure of a human face and the ability to synthesis plausible details which are not visible in the input view. Hence, we train a multiview diffusion model " + }, + { + "bbox": [ + 313, + 217, + 553, + 491 + ], + "type": "inline_equation", + "content": "f_{D}" + }, + { + "bbox": [ + 313, + 217, + 553, + 491 + ], + "type": "text", + "content": " on synthetic human head data to generate " + }, + { + "bbox": [ + 313, + 217, + 553, + 491 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 217, + 553, + 491 + ], + "type": "text", + "content": " views " + }, + { + "bbox": [ + 313, + 217, + 553, + 491 + ], + "type": "inline_equation", + "content": "x_0^1, x_0^2, \\ldots, x_0^N" + }, + { + "bbox": [ + 313, + 217, + 553, + 491 + ], + "type": "text", + "content": " covering " + }, + { + "bbox": [ + 313, + 217, + 553, + 491 + ], + "type": "inline_equation", + "content": "360^\\circ" + }, + { + "bbox": [ + 313, + 217, + 553, + 491 + ], + "type": "text", + "content": " of the human head while achieving multi-view consistency and preserving identity. We choose pixel-aligned 3D Gaussians to obtain the final 3D representation. Compared to NeRFs and meshes, 3D Gaussians offer explicit volumetric primitives that better capture subtle facial geometry and fine details. Their semi-transparent kernels naturally model effects like wispy hair and translucency, which are challenging for discrete surfaces or density fields. These generated views " + }, + { + "bbox": [ + 313, + 217, + 553, + 491 + ], + "type": "inline_equation", + "content": "x_0^{1:N}" + }, + { + "bbox": [ + 313, + 217, + 553, + 491 + ], + "type": "text", + "content": " from the diffusion model, along with their corresponding Plücker ray coordinates " + }, + { + "bbox": [ + 313, + 217, + 553, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{(1:N)}" + }, + { + "bbox": [ + 313, + 217, + 553, + 491 + ], + "type": "text", + "content": ", are input into a transformer-based Gaussian reconstructor " + }, + { + "bbox": [ + 313, + 217, + 553, + 491 + ], + "type": "inline_equation", + "content": "f_R" + }, + { + "bbox": [ + 313, + 217, + 553, + 491 + ], + "type": "text", + "content": " to predict a set of 3D Gaussians " + }, + { + "bbox": [ + 313, + 217, + 553, + 491 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 217, + 553, + 491 + ], + "type": "text", + "content": ". Training of the Gaussian reconstructor follows a pre-training process on general objects [12] and a fine-tuning process on synthetic human head data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 495, + 484, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 484, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 484, + 507 + ], + "type": "text", + "content": "3.1. Synthetic Human Head Dataset" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 510, + 553, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 553, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 553, + 665 + ], + "type": "text", + "content": "We implement a 3D head asset generation pipeline inspired by [65]. Our process begins with a collection of high-quality, artist-created 3D head meshes, which we enhance by incorporating detailed facial components, including eyes, teeth, gums, and both facial and scalp hair. We then augment these base models through rigging for pose variation and blendshape deformation for diverse facial expressions. The final head models are enriched with a set of PBR texture maps, including albedo, normal, roughness, specular, and subsurface scattering maps. At last, we dress the head model with a collection of clothing assets. The entire pipeline is implemented in Blender and the images are rendered with Cycles renderer." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "content": "To train our networks, we render images (samples shown in Fig. 4) at " + }, + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "content": " resolution from 200 unique identities, each with 50 varied appearances, including different hairstyles, skin tones, expressions, clothes, poses, etc." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12693" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 70, + 541, + 174 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 541, + 174 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 541, + 174 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 541, + 174 + ], + "type": "image", + "image_path": "0b6dcfa3e9d95382ee8a66e0d65d3b40dd971f7d166af00dc6902fe0699a9ddf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 176, + 555, + 232 + ], + "lines": [ + { + "bbox": [ + 55, + 176, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 176, + 555, + 232 + ], + "type": "text", + "content": "Figure 3. Overview of FaceLift. Given a single image of a human face " + }, + { + "bbox": [ + 55, + 176, + 555, + 232 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 176, + 555, + 232 + ], + "type": "text", + "content": " as input, we train an image-conditioned, multi-view diffusion model to generate novel views " + }, + { + "bbox": [ + 55, + 176, + 555, + 232 + ], + "type": "inline_equation", + "content": "x_0^1, \\ldots, x_0^N" + }, + { + "bbox": [ + 55, + 176, + 555, + 232 + ], + "type": "text", + "content": " covering the entire head. By generating " + }, + { + "bbox": [ + 55, + 176, + 555, + 232 + ], + "type": "inline_equation", + "content": "x_0^1" + }, + { + "bbox": [ + 55, + 176, + 555, + 232 + ], + "type": "text", + "content": " the same as " + }, + { + "bbox": [ + 55, + 176, + 555, + 232 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 176, + 555, + 232 + ], + "type": "text", + "content": " and leveraging the high-quality synthetic data, our multi-view latent diffusion model can hallucinate unseen views of the human head with high-fidelity and multi-view consistency. We then train a transformer-based reconstructor " + }, + { + "bbox": [ + 55, + 176, + 555, + 232 + ], + "type": "inline_equation", + "content": "f_R" + }, + { + "bbox": [ + 55, + 176, + 555, + 232 + ], + "type": "text", + "content": ", which takes multi-view images " + }, + { + "bbox": [ + 55, + 176, + 555, + 232 + ], + "type": "inline_equation", + "content": "x_0^{1:N}" + }, + { + "bbox": [ + 55, + 176, + 555, + 232 + ], + "type": "text", + "content": " and their camera poses " + }, + { + "bbox": [ + 55, + 176, + 555, + 232 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{1:N}" + }, + { + "bbox": [ + 55, + 176, + 555, + 232 + ], + "type": "text", + "content": " as input and generates 3D Gaussian splats " + }, + { + "bbox": [ + 55, + 176, + 555, + 232 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 176, + 555, + 232 + ], + "type": "text", + "content": " to represent the human head." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 61, + 242, + 293, + 319 + ], + "blocks": [ + { + "bbox": [ + 61, + 242, + 293, + 319 + ], + "lines": [ + { + "bbox": [ + 61, + 242, + 293, + 319 + ], + "spans": [ + { + "bbox": [ + 61, + 242, + 293, + 319 + ], + "type": "image", + "image_path": "ce4410f083f829d2240ca808516f1e389ee172b9e65296f4a48fba896189a94e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 321, + 295, + 354 + ], + "lines": [ + { + "bbox": [ + 55, + 321, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 321, + 295, + 354 + ], + "type": "text", + "content": "Figure 4. Synthetic data examples. Top row: six views for diffusion training. Bottom row: samples of random views for reconstructor training." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 362, + 296, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 362, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 362, + 296, + 434 + ], + "type": "text", + "content": "We render our training dataset under two types of lighting conditions: (1) ambient light and (2) random HDR environment light. We render six views for each subject to train the multi-view diffusion model. For fine-tuning the transformer-based reconstructor, we render 32 views with random camera poses." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 437, + 157, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 157, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 157, + 449 + ], + "type": "text", + "content": "3.2. View Generation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 452, + 296, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 452, + 296, + 523 + ], + "spans": [ + { + "bbox": [ + 55, + 452, + 296, + 523 + ], + "type": "text", + "content": "We model the sparse view generation from a single image input as a conditional diffusion process. We use a multiview diffusion model " + }, + { + "bbox": [ + 55, + 452, + 296, + 523 + ], + "type": "inline_equation", + "content": "f_{D}" + }, + { + "bbox": [ + 55, + 452, + 296, + 523 + ], + "type": "text", + "content": " to generate " + }, + { + "bbox": [ + 55, + 452, + 296, + 523 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 452, + 296, + 523 + ], + "type": "text", + "content": " views, denoted as " + }, + { + "bbox": [ + 55, + 452, + 296, + 523 + ], + "type": "inline_equation", + "content": "X_0^1, X_0^2, \\ldots, X_0^N" + }, + { + "bbox": [ + 55, + 452, + 296, + 523 + ], + "type": "text", + "content": ", given a single front-facing image " + }, + { + "bbox": [ + 55, + 452, + 296, + 523 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 452, + 296, + 523 + ], + "type": "text", + "content": " and CLIP text embeddings " + }, + { + "bbox": [ + 55, + 452, + 296, + 523 + ], + "type": "inline_equation", + "content": "e^1, e^2, \\ldots, e^N" + }, + { + "bbox": [ + 55, + 452, + 296, + 523 + ], + "type": "text", + "content": " corresponding to each generated view. This process is expressed as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 72, + 524, + 295, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 524, + 295, + 539 + ], + "spans": [ + { + "bbox": [ + 72, + 524, + 295, + 539 + ], + "type": "interline_equation", + "content": "\\left\\{X _ {0} ^ {1}, X _ {0} ^ {2}, \\dots , X _ {0} ^ {N} \\right\\} = f _ {D} \\left(y, \\left\\{e ^ {1}, e ^ {2}, \\dots , e ^ {N} \\right\\}\\right). \\tag {1}", + "image_path": "259aff94bfdaeee9e36c1118d11f8b5bb93d8e02824b2990527fb96b53d9f71d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 544, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 579 + ], + "type": "text", + "content": "We aim to learn the joint distribution of all these views, conditioning on the input image " + }, + { + "bbox": [ + 55, + 544, + 295, + 579 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 544, + 295, + 579 + ], + "type": "text", + "content": " and text embedding " + }, + { + "bbox": [ + 55, + 544, + 295, + 579 + ], + "type": "inline_equation", + "content": "e^1, e^2, \\ldots, e^N" + }, + { + "bbox": [ + 55, + 544, + 295, + 579 + ], + "type": "text", + "content": ". We denote the joint distribution as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 580, + 297, + 605 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 297, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 297, + 605 + ], + "type": "interline_equation", + "content": "p _ {f _ {D}} \\left(x _ {0} ^ {1: N} \\mid y, e ^ {1: N}\\right) := p _ {f _ {D}} \\left(\\left\\{x _ {0} ^ {1}, \\dots x _ {0} ^ {N} \\right\\} \\mid y, \\left\\{e ^ {1}, \\dots e ^ {N} \\right\\}\\right). \\tag {2}", + "image_path": "8ce75a2c83ab7ff84df2fe7333c140c194ae124aa34be24d75182d7600d17dd3.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": "View Selection. Given a single near frontal view face image with azimuth " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": ", our multi-view diffusion model generates six views with azimuths equal to " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{\\alpha, \\alpha \\pm 45^\\circ, \\alpha \\pm 90^\\circ, \\alpha + 180^\\circ\\}" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": ", covering 360 degrees of the human head. Elevation is 0 for all images. We opt for six views as the optimal balance - fewer views compromise detail quality while more views become computationally prohibitive for full head reconstruction. An ablation study comparing four, six, and eight views is presented in Sec. 5.2." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 241, + 555, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 241, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 241, + 555, + 479 + ], + "type": "text", + "content": "Multi-view Attention. To ensure the consistency of the generated novel views, we use a multi-view attention mechanism to facilitate information propagation and implicitly encode multi-view dependencies. Our attention module encourages multi-view consistency by extending the 2D self-attention mechanism to 3D and enabling interactions across views. Instead of treating each view independently, we apply self-attention across all views simultaneously, allowing information to be shared between them. Specifically, we start with an input tensor of shape " + }, + { + "bbox": [ + 313, + 241, + 555, + 479 + ], + "type": "inline_equation", + "content": "B \\times V \\times H \\times W \\times C" + }, + { + "bbox": [ + 313, + 241, + 555, + 479 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 241, + 555, + 479 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 241, + 555, + 479 + ], + "type": "text", + "content": " is the batch size, " + }, + { + "bbox": [ + 313, + 241, + 555, + 479 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 241, + 555, + 479 + ], + "type": "text", + "content": " is the number of views, " + }, + { + "bbox": [ + 313, + 241, + 555, + 479 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 241, + 555, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 241, + 555, + 479 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 241, + 555, + 479 + ], + "type": "text", + "content": " denote the spatial resolution of the intermediate feature maps, and " + }, + { + "bbox": [ + 313, + 241, + 555, + 479 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 241, + 555, + 479 + ], + "type": "text", + "content": " is the number of feature channels. We reshape this tensor to " + }, + { + "bbox": [ + 313, + 241, + 555, + 479 + ], + "type": "inline_equation", + "content": "B \\times VHW \\times C" + }, + { + "bbox": [ + 313, + 241, + 555, + 479 + ], + "type": "text", + "content": ", treating all spatial locations across all views as a unified sequence of tokens for self-attention. This design allows the model to learn multi-view correlations by sharing information across views within the attention layers, enabling it to generate consistent RGB images. We provide an ablation study on the multi-view attention mechanism in the supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 480, + 554, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 554, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 554, + 588 + ], + "type": "text", + "content": "Input View Reconstruction. During Training, we enforce the first generated view to share the same camera with the input image. In other words, we reconstruct the input view in the view generation process. We find this approach, combined with the multi-view attention mechanism, significantly outperforms the alternative strategy of generating only novel views, which tends to overfit to synthetic training identities and compromises generalization capability as we will show in Sec. 5.1 and the supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 591, + 540, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 591, + 540, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 591, + 540, + 602 + ], + "type": "text", + "content": "3.3. Multi-view to 3D Gaussians Reconstruction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 605, + 554, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 688 + ], + "type": "text", + "content": "Transformer-based Reconstructor. We choose pixel-aligned 3D Gaussians as the final 3D representation. Each Gaussian " + }, + { + "bbox": [ + 313, + 605, + 554, + 688 + ], + "type": "inline_equation", + "content": "G_{i}" + }, + { + "bbox": [ + 313, + 605, + 554, + 688 + ], + "type": "text", + "content": " is defined by position " + }, + { + "bbox": [ + 313, + 605, + 554, + 688 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 313, + 605, + 554, + 688 + ], + "type": "text", + "content": ", scale " + }, + { + "bbox": [ + 313, + 605, + 554, + 688 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 313, + 605, + 554, + 688 + ], + "type": "text", + "content": ", orientation " + }, + { + "bbox": [ + 313, + 605, + 554, + 688 + ], + "type": "inline_equation", + "content": "q_{i}" + }, + { + "bbox": [ + 313, + 605, + 554, + 688 + ], + "type": "text", + "content": ", opacity " + }, + { + "bbox": [ + 313, + 605, + 554, + 688 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 313, + 605, + 554, + 688 + ], + "type": "text", + "content": ", and color features " + }, + { + "bbox": [ + 313, + 605, + 554, + 688 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 313, + 605, + 554, + 688 + ], + "type": "text", + "content": ". We use a transformer-based reconstructor " + }, + { + "bbox": [ + 313, + 605, + 554, + 688 + ], + "type": "inline_equation", + "content": "f_{G}" + }, + { + "bbox": [ + 313, + 605, + 554, + 688 + ], + "type": "text", + "content": " to obtain 3D Gaussians from generated multi-view images " + }, + { + "bbox": [ + 313, + 605, + 554, + 688 + ], + "type": "inline_equation", + "content": "x_0^{1:N}" + }, + { + "bbox": [ + 313, + 605, + 554, + 688 + ], + "type": "text", + "content": " and their corresponding Plücker ray coordinates [44] " + }, + { + "bbox": [ + 313, + 605, + 554, + 688 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{1:N}" + }, + { + "bbox": [ + 313, + 605, + 554, + 688 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 322, + 687, + 553, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 687, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 322, + 687, + 553, + 713 + ], + "type": "interline_equation", + "content": "\\left\\{G _ {i} \\right\\} _ {i = 1} ^ {N H W} = \\left\\{p _ {i}, s _ {i}, q _ {i}, \\alpha_ {i}, c _ {i} \\right\\} _ {i = 1} ^ {N H W} = f _ {G} \\left(x _ {0} ^ {1: N}, \\mathrm {P} ^ {1: N}\\right), \\tag {3}", + "image_path": "9d834fe98d96ac99647d141f7762d7fc35921b2b6d01f3675a450b34145ac0bc.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12694" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": "Our " + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "inline_equation", + "content": "f_{G}" + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": " is a large reconstruction model [24, 74] which follows the implementation of GS-LRM [74]: the " + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": " multiview images are concatenated with their Plücker ray coordinates computed from the camera intrinsic and extrinsic parameters for pose conditioning. Then, the inputs are patched by dividing the per-view feature map into non-overlapping patches with a patch size of " + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": ". Each 2D patch is then flattened into a 1D vector. Finally, a linear layer " + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": " is utilized to map the 1D vectors to image patch tokens:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 182, + 294, + 210 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 182, + 294, + 210 + ], + "spans": [ + { + "bbox": [ + 61, + 182, + 294, + 210 + ], + "type": "interline_equation", + "content": "\\left\\{\\mathrm {T} _ {j} ^ {n} \\right\\} _ {j = 1, 2, \\dots , \\frac {H W}{p ^ {2}}} = L \\left(\\text {P a t c h i f y} _ {p} \\left(\\text {C o n c a t} \\left(\\mathrm {I} ^ {n}, \\mathrm {P} ^ {n}\\right)\\right)\\right). \\tag {4}", + "image_path": "36f8d9c1cbd45d84bb92e07c1d29bcc10a39a4ecc861cf0767bb13fb522251a0.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 210, + 295, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 210, + 295, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 210, + 295, + 344 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 55, + 210, + 295, + 344 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{T}_j^n\\}" + }, + { + "bbox": [ + 55, + 210, + 295, + 344 + ], + "type": "text", + "content": " denotes the set of patch tokens for image " + }, + { + "bbox": [ + 55, + 210, + 295, + 344 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 210, + 295, + 344 + ], + "type": "text", + "content": ", totaling " + }, + { + "bbox": [ + 55, + 210, + 295, + 344 + ], + "type": "inline_equation", + "content": "\\frac{HW}{p^2}" + }, + { + "bbox": [ + 55, + 210, + 295, + 344 + ], + "type": "text", + "content": " tokens per image. The set of multi-view image tokens " + }, + { + "bbox": [ + 55, + 210, + 295, + 344 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{T}_j^n\\}" + }, + { + "bbox": [ + 55, + 210, + 295, + 344 + ], + "type": "text", + "content": " are concatenated and processed through a chain of transformer blocks. Each transformer block is equipped with residual connections [20] and consists of PreLayerNorm [2], multi-head Self-Attention [59] and MLP. Later, the output tokens from the transformer are decoded into Gaussian parameters using a single linear layer. Then, the Gaussian parameters are unpatchified into " + }, + { + "bbox": [ + 55, + 210, + 295, + 344 + ], + "type": "inline_equation", + "content": "p^2" + }, + { + "bbox": [ + 55, + 210, + 295, + 344 + ], + "type": "text", + "content": " Gaussians. Finally, we end up with " + }, + { + "bbox": [ + 55, + 210, + 295, + 344 + ], + "type": "inline_equation", + "content": "HW" + }, + { + "bbox": [ + 55, + 210, + 295, + 344 + ], + "type": "text", + "content": " Gaussians for each view, where pixel encodes one 3D Gaussian." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 347, + 295, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 347, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 295, + 586 + ], + "type": "text", + "content": "Two-stage Training. We find that training the transformer-based reconstructor solely on synthetic human head data leads to inferior texture details when applied to real-world images (see ablation study in Fig. 12). We suspect this limitation arises because the synthetic datasets lack geometric diversity. To address this, we propose a two-stage training approach in which the reconstructor is pre-trained on diverse object data [12] and subsequently fine-tuned using synthetic head data. The pre-training stage enables the reconstructor to learn a broad prior of diverse geometric structures, yielding more detailed and clearer textures in delicate facial regions such as the eyes, nose, and ears. The fine-tuning process then imparts specific knowledge of head structure, producing smoother and more realistic results. During training, we randomly select four input views to reconstruct a total of eight views, four input and four novel views. Following [74], we optimize the model using a combination of MSE and perceptual losses. During inference, the reconstructor processes the six-view outputs from multi-view diffusion model to reconstruct the head." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 591, + 209, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 591, + 209, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 591, + 209, + 604 + ], + "type": "text", + "content": "3.4. Real-world Image Inference" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 712 + ], + "type": "text", + "content": "For inference on real-world images, since their intrinsic parameters are unknown, we adopt a camera fov of " + }, + { + "bbox": [ + 55, + 605, + 295, + 712 + ], + "type": "inline_equation", + "content": "50^{\\circ}" + }, + { + "bbox": [ + 55, + 605, + 295, + 712 + ], + "type": "text", + "content": ", same as during training. To ensure plausible outputs, we first apply an MTCNN face detector to estimate the face's size and center. The image is then resized and cropped/extended to match the average face size and center computed from the training data. We find this alignment compensates for the unknown intrinsic parameters well, ensuring plausible reconstruction results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 72, + 395, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 395, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 395, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 88, + 430, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 88, + 430, + 102 + ], + "spans": [ + { + "bbox": [ + 314, + 88, + 430, + 102 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 103, + 553, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 103, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 313, + 103, + 553, + 271 + ], + "type": "text", + "content": "Evaluation Datasets. To quantitatively evaluate FaceLift, we establish two benchmarks for single-image to 3D full head reconstruction tasks using publicly available datasets: (1) Cafca dataset [8]: We select 40 subjects with 7 to 19 test camera poses each. Since the camera positions are randomly distributed, we manually select the most frontal view as input. Note that this synthetic dataset was independently developed and differs significantly from our training dataset. (2) Ava-256 dataset [39]: This studio-captured dataset contains real human subjects. We sample 10 subjects and 10 test camera poses for our evaluation. More details in supplemental. To demonstrate our system's generalization capabilities, we also evaluate on a set of in-the-wild face images for qualitative assessment." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 274, + 553, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 553, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 553, + 393 + ], + "type": "text", + "content": "Baselines. We compare our results against three state-of-the-art methods for single-face 3D reconstruction: GGHead [29], PanoHead [1], and Dual Encoder [4]. We perform GAN-inversion to reconstruct 3D head from a given input image using these models. To emphasize the importance of utilizing our synthetic human head data for training, we also compare our method with two methods that focus on general object reconstruction: Era3D [31] and LGM [56]. More comparison results with mesh-based methods are provided in the supplementary material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 396, + 553, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 553, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 553, + 515 + ], + "type": "text", + "content": "We further developed a baseline, Our " + }, + { + "bbox": [ + 313, + 396, + 553, + 515 + ], + "type": "inline_equation", + "content": "MV + LGM" + }, + { + "bbox": [ + 313, + 396, + 553, + 515 + ], + "type": "text", + "content": ", which leverages the multi-view outputs generated by our diffusion model and employs LGM for reconstruction. This demonstrates that our method can be seamlessly integrated with other reconstruction frameworks to enhance performance on face reconstruction tasks. We tried to fine-tune the LGM reconstructor with our synthetic data, but it provides inferior results with incorrect geometry and artifacts compared with the original weights, which we suspect is due to training data mismatch (see details in the supplementary material)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 517, + 553, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 553, + 578 + ], + "type": "text", + "content": "Evaluation Metrics. We evaluate reconstruction quality using four standard metrics: PSNR, SSIM, LPIPS [75], and DreamSim [16]. To evaluate identity preservation, we perform face verification using ArcFace [14] through the DeepFace [52] implementation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 581, + 553, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 553, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 553, + 700 + ], + "type": "text", + "content": "Implementation Details. Both Cafca [8] and Ava-256 [39] datasets provide multi-view RGB images and corresponding camera poses. However, their camera system differs from the ones utilized in FaceLift and baselines. We recalculate the test camera extrinsic in each method's camera system. For a more accurate comparison, we use the Mediapipe facial landmark detection algorithm [38] to identify facial landmarks in both the target images and the rendered outputs, aligning them based on landmark distributions. Details are provided in the supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 325, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 553, + 713 + ], + "type": "text", + "content": "Our system takes approximately 8 seconds to infer a" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12695" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 294, + 148 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 294, + 148 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 294, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 294, + 148 + ], + "type": "table", + "html": "
MethodPSNR ↑SSIM ↑LPIPS ↓DreamSim ↓ArcFace ↓
GGHead [29]10.350.74060.36360.32520.2681
PanoHead [1]10.720.75940.33510.20480.2183
Dual Encoder [4]10.780.73850.39220.27850.2421
Era3D [31]13.690.72300.36620.28920.2978
LGM [56]16.520.79330.30600.15520.2557
Our MV + LGM [56]14.130.78120.29560.12820.1767
FaceLift16.610.79680.26940.10960.1573
", + "image_path": "c141da7ab0cf605f11c946a900ff61197c18b0d3d60bf0d0b00c5f332e395b2d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 57, + 178, + 293, + 266 + ], + "blocks": [ + { + "bbox": [ + 57, + 178, + 293, + 266 + ], + "lines": [ + { + "bbox": [ + 57, + 178, + 293, + 266 + ], + "spans": [ + { + "bbox": [ + 57, + 178, + 293, + 266 + ], + "type": "image", + "image_path": "2a40e3bf004896985025f7ab3ebdd754a8e5799ec01d59de5f2e98d3d4a12003.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 270, + 295, + 314 + ], + "lines": [ + { + "bbox": [ + 55, + 270, + 295, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 295, + 314 + ], + "type": "text", + "content": "Figure 5. Visual results on Cafca compared with face reconstruction methods. FaceLift renders novel views that closely match the ground truth, while other baselines often fail to reconstruct the 3D head in correct colors or geometry structures." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 58, + 321, + 293, + 369 + ], + "blocks": [ + { + "bbox": [ + 58, + 321, + 293, + 369 + ], + "lines": [ + { + "bbox": [ + 58, + 321, + 293, + 369 + ], + "spans": [ + { + "bbox": [ + 58, + 321, + 293, + 369 + ], + "type": "image", + "image_path": "980876d00a5dbee340c3773abeaaa984769433a2e4fede1b5d8489c12e3c6b93.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 372, + 295, + 406 + ], + "lines": [ + { + "bbox": [ + 55, + 372, + 295, + 406 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 295, + 406 + ], + "type": "text", + "content": "Figure 6. Visual results on Cafca compared with general objects reconstruction methods. Comparison with general object reconstruction methods shows the importance of specialized data." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 416, + 296, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 416, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 416, + 296, + 464 + ], + "type": "text", + "content": "3D Gaussian head from a single image: about 1.5 seconds for preprocessing (background removal, rescaling, etc.), 5.5 seconds for multi-view image generation, and under 1 second for 3D Gaussians reconstruction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 472, + 239, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 239, + 484 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 239, + 484 + ], + "type": "text", + "content": "4.2. Experiments on the Cafca Dataset" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 487, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 296, + 715 + ], + "type": "text", + "content": "We report numerical comparison results on Cafca in Tab. 1. FaceLift performs favorably against baselines, especially on DreamSim [16] metric, indicating high-quality perceptual similarity. It also achieves better identity preservation performance, as demonstrated by a lower ArcFace [14] distance. We show visual results in Fig. 5 and Fig. 6. FaceLift yields rendering results that closely match the ground truth. Compared with other baselines, GGHead [29] does not support full-head rendering, resulting in unrealistic outputs when the view angle significantly deviates from the input. PanoHead [1] struggles with challenging hairstyles, while Dual Encoder [4] produces blurred facial textures. Additionally, Era3D [31] introduces artifacts on the back of the head, and LGM [56] yields inaccurate nose and jaw shapes, underscoring the importance of our synthetic human head data. When integrated with our multi-view diffusion approach, LGM achieves enhanced performance, demonstrating that our method can be seamlessly combined with existing baselines to boost their results." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 550, + 123 + ], + "blocks": [ + { + "bbox": [ + 55, + 150, + 295, + 172 + ], + "lines": [ + { + "bbox": [ + 55, + 150, + 295, + 172 + ], + "spans": [ + { + "bbox": [ + 55, + 150, + 295, + 172 + ], + "type": "text", + "content": "Table 1. Quantitative results on Cafca. FaceLift achieves favorable performance on all evaluation metrics." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 70, + 550, + 123 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 550, + 123 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 550, + 123 + ], + "type": "table", + "html": "
MethodPSNR ↑SSIM ↑LPIPS ↓DreamSim ↓ArcFace ↓
Era3D [31]14.770.79630.25380.25150.3721
LGM [56]14.050.81360.24760.14960.3142
Our MV+LGM [56]15.240.82130.22920.10930.2264
FaceLift16.520.82710.22770.10650.1871
", + "image_path": "b01fc94fc3d1e3dcf390fa1da1529764e5941c6be2f5975402b944121cd547d0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 125, + 555, + 171 + ], + "lines": [ + { + "bbox": [ + 313, + 125, + 555, + 171 + ], + "spans": [ + { + "bbox": [ + 313, + 125, + 555, + 171 + ], + "type": "text", + "content": "Table 2. Quantitative results on Ava-256. FaceLift performs favorably than baseline methods in both reconstruction metrics and identity facial identity metric, showing a better generalization ability towards real-captured human images." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 316, + 177, + 550, + 335 + ], + "blocks": [ + { + "bbox": [ + 316, + 177, + 550, + 335 + ], + "lines": [ + { + "bbox": [ + 316, + 177, + 550, + 335 + ], + "spans": [ + { + "bbox": [ + 316, + 177, + 550, + 335 + ], + "type": "image", + "image_path": "f3b50a414451536fe0eff292687a3a9b3c3c283a3df72120ffbc5e162c16dddc.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 336, + 555, + 381 + ], + "lines": [ + { + "bbox": [ + 313, + 336, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 555, + 381 + ], + "type": "text", + "content": "Figure 7. Visual results on Ava-256. Compared with baselines, FaceLift provides multi-view renderings that are more realistic and similar to ground truth. Era3D fails to deliver delicate facial structures, while LGM generates inaccurate head shapes and colors." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 391, + 507, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 391, + 507, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 507, + 403 + ], + "type": "text", + "content": "4.3. Experiments on the Ava-256 Dataset" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 406, + 555, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 555, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 555, + 586 + ], + "type": "text", + "content": "We further evaluate FaceLift against other baselines on a studio-captured real human dataset, Ava-256 [50]. GAN-inversion based methods [1, 4, 29] fail to produce reasonable results with the test camera poses in this dataset, so we exclude these baselines. Tab. 2 shows that FaceLift outperforms all other baselines across all evaluation metrics, demonstrating superior reconstruction quality and identity preservation. It also highlights FaceLift's strong ability to generalize to real human faces. As shown in Fig. 7, FaceLift achieves more realistic head synthesis, while Era3D [31] struggles with accurate skin and hair textures, as well as facial details. LGM [56] produces inaccuracies in the nose shape. When combined with our multi-view diffusion model, LGM yields more accurate geometric structures, yet its texture quality remains inferior to that of FaceLift." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 590, + 512, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 512, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 512, + 604 + ], + "type": "text", + "content": "4.4. Experiments with In-the-wild Images" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "We collect in-the-wild human face images and present qualitative results in comparison with other baselines in Fig. 8. Baseline methods often produce undesirable artifacts. For instance, PanoHead [1] frequently fails to render the back of the head and sometimes generates extra eyes at the rear. It also struggles to synthesize hair, shadows, wrinkles, and facial paint accurately, and its outputs lack multi-view consistency (e.g., the girl continues to face the camera in novel view 1 despite a changed pose). Dual Encoder [4] improves" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12696" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 89, + 105, + 125 + ], + "blocks": [ + { + "bbox": [ + 69, + 71, + 105, + 80 + ], + "lines": [ + { + "bbox": [ + 69, + 71, + 105, + 80 + ], + "spans": [ + { + "bbox": [ + 69, + 71, + 105, + 80 + ], + "type": "text", + "content": "Input View" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 89, + 105, + 125 + ], + "lines": [ + { + "bbox": [ + 67, + 89, + 105, + 125 + ], + "spans": [ + { + "bbox": [ + 67, + 89, + 105, + 125 + ], + "type": "image", + "image_path": "7fed4dcd366bfe9351ce23d69b27cd076b931dc2a56517fef93c57a8d9c4f698.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 108, + 90, + 177, + 125 + ], + "blocks": [ + { + "bbox": [ + 108, + 90, + 177, + 125 + ], + "lines": [ + { + "bbox": [ + 108, + 90, + 177, + 125 + ], + "spans": [ + { + "bbox": [ + 108, + 90, + 177, + 125 + ], + "type": "image", + "image_path": "cc3371677dd01ccd6c488aaf6314b097a778b1fb3907678e08bd6f858ef84de6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 179, + 89, + 194, + 125 + ], + "blocks": [ + { + "bbox": [ + 191, + 72, + 240, + 80 + ], + "lines": [ + { + "bbox": [ + 191, + 72, + 240, + 80 + ], + "spans": [ + { + "bbox": [ + 191, + 72, + 240, + 80 + ], + "type": "text", + "content": "Rendered View 1" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 109, + 82, + 315, + 89 + ], + "lines": [ + { + "bbox": [ + 109, + 82, + 315, + 89 + ], + "spans": [ + { + "bbox": [ + 109, + 82, + 315, + 89 + ], + "type": "text", + "content": "PanoHead [1] Dual Encoder [4] Era3D [31] LGM [56] Our MV+LGM [56] FaceLift" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 179, + 89, + 194, + 125 + ], + "lines": [ + { + "bbox": [ + 179, + 89, + 194, + 125 + ], + "spans": [ + { + "bbox": [ + 179, + 89, + 194, + 125 + ], + "type": "image", + "image_path": "2ce1d5f4766c43d796e58cfb5c4f54e5a98545564062e4aa29c84ecedf7e89da.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 201, + 89, + 210, + 125 + ], + "blocks": [ + { + "bbox": [ + 201, + 89, + 210, + 125 + ], + "lines": [ + { + "bbox": [ + 201, + 89, + 210, + 125 + ], + "spans": [ + { + "bbox": [ + 201, + 89, + 210, + 125 + ], + "type": "image", + "image_path": "a0e7f84978759d7e7dda93d8bd10e6cb8e55621aaa464b9f31eed2265b5d2032.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 218, + 89, + 243, + 125 + ], + "blocks": [ + { + "bbox": [ + 218, + 89, + 243, + 125 + ], + "lines": [ + { + "bbox": [ + 218, + 89, + 243, + 125 + ], + "spans": [ + { + "bbox": [ + 218, + 89, + 243, + 125 + ], + "type": "image", + "image_path": "a9891b6749d1924b63e624b85486e7626742cf5d96ff787910ef8bef17784071.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 250, + 89, + 290, + 125 + ], + "blocks": [ + { + "bbox": [ + 250, + 89, + 290, + 125 + ], + "lines": [ + { + "bbox": [ + 250, + 89, + 290, + 125 + ], + "spans": [ + { + "bbox": [ + 250, + 89, + 290, + 125 + ], + "type": "image", + "image_path": "e0651557001a66bd835743f9bbd4ddfb8f2afd68bad881e500f925d7b48f052f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 290, + 89, + 310, + 125 + ], + "blocks": [ + { + "bbox": [ + 290, + 89, + 310, + 125 + ], + "lines": [ + { + "bbox": [ + 290, + 89, + 310, + 125 + ], + "spans": [ + { + "bbox": [ + 290, + 89, + 310, + 125 + ], + "type": "image", + "image_path": "cb894c9741d5af94be99357a9541b1f968da9a5aab03fab507b18dfd33258d1a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 67, + 125, + 105, + 161 + ], + "blocks": [ + { + "bbox": [ + 67, + 125, + 105, + 161 + ], + "lines": [ + { + "bbox": [ + 67, + 125, + 105, + 161 + ], + "spans": [ + { + "bbox": [ + 67, + 125, + 105, + 161 + ], + "type": "image", + "image_path": "004aad661a7572a1ba60a0599505327955565695f369f588cb38de89fbf46cc8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 108, + 125, + 177, + 161 + ], + "blocks": [ + { + "bbox": [ + 108, + 125, + 177, + 161 + ], + "lines": [ + { + "bbox": [ + 108, + 125, + 177, + 161 + ], + "spans": [ + { + "bbox": [ + 108, + 125, + 177, + 161 + ], + "type": "image", + "image_path": "1538e0c98c0a05b0bdfaf295aeb40c2136b90bba5c176bf32b5b44dd3ee2862a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 179, + 125, + 199, + 161 + ], + "blocks": [ + { + "bbox": [ + 179, + 125, + 199, + 161 + ], + "lines": [ + { + "bbox": [ + 179, + 125, + 199, + 161 + ], + "spans": [ + { + "bbox": [ + 179, + 125, + 199, + 161 + ], + "type": "image", + "image_path": "7b75797307eb674cffb0dba25868dcec53b25dba705a523ecb371df2bda04e41.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 202, + 125, + 247, + 161 + ], + "blocks": [ + { + "bbox": [ + 202, + 125, + 247, + 161 + ], + "lines": [ + { + "bbox": [ + 202, + 125, + 247, + 161 + ], + "spans": [ + { + "bbox": [ + 202, + 125, + 247, + 161 + ], + "type": "image", + "image_path": "c9cae66a05bfc0862caae8d76f273b4741c1571ddcec578f8b420e253aa3c336.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 252, + 125, + 272, + 161 + ], + "blocks": [ + { + "bbox": [ + 252, + 125, + 272, + 161 + ], + "lines": [ + { + "bbox": [ + 252, + 125, + 272, + 161 + ], + "spans": [ + { + "bbox": [ + 252, + 125, + 272, + 161 + ], + "type": "image", + "image_path": "faa508d719a709b37fb65f42a38a5291e89eccdd079921a80e09cdc43c983d1b.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 290, + 125, + 309, + 161 + ], + "blocks": [ + { + "bbox": [ + 290, + 125, + 309, + 161 + ], + "lines": [ + { + "bbox": [ + 290, + 125, + 309, + 161 + ], + "spans": [ + { + "bbox": [ + 290, + 125, + 309, + 161 + ], + "type": "image", + "image_path": "08e558f12b6b11539b893d7aca98695a6cea72839651422915b8b7dda5b69330.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 67, + 161, + 105, + 196 + ], + "blocks": [ + { + "bbox": [ + 67, + 161, + 105, + 196 + ], + "lines": [ + { + "bbox": [ + 67, + 161, + 105, + 196 + ], + "spans": [ + { + "bbox": [ + 67, + 161, + 105, + 196 + ], + "type": "image", + "image_path": "f2b978650ee2eb1ebc6285f69ce3f187ff16870a835fab61c172595f5cf2b234.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 108, + 161, + 181, + 196 + ], + "blocks": [ + { + "bbox": [ + 108, + 161, + 181, + 196 + ], + "lines": [ + { + "bbox": [ + 108, + 161, + 181, + 196 + ], + "spans": [ + { + "bbox": [ + 108, + 161, + 181, + 196 + ], + "type": "image", + "image_path": "b071742ae771d3f28ea0fdddbf10492ff18d603aeed5115c423771bd930bba9e.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 181, + 161, + 197, + 196 + ], + "blocks": [ + { + "bbox": [ + 181, + 161, + 197, + 196 + ], + "lines": [ + { + "bbox": [ + 181, + 161, + 197, + 196 + ], + "spans": [ + { + "bbox": [ + 181, + 161, + 197, + 196 + ], + "type": "image", + "image_path": "f627882566c8698574d62225529f51c69520146e7f051b141ed8a86c64b0c295.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 202, + 161, + 247, + 196 + ], + "blocks": [ + { + "bbox": [ + 202, + 161, + 247, + 196 + ], + "lines": [ + { + "bbox": [ + 202, + 161, + 247, + 196 + ], + "spans": [ + { + "bbox": [ + 202, + 161, + 247, + 196 + ], + "type": "image", + "image_path": "4d58b91842b19a59c910fdf5073fc6fb5516712408d9ded18859d6d00d3338fe.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 252, + 161, + 272, + 196 + ], + "blocks": [ + { + "bbox": [ + 252, + 161, + 272, + 196 + ], + "lines": [ + { + "bbox": [ + 252, + 161, + 272, + 196 + ], + "spans": [ + { + "bbox": [ + 252, + 161, + 272, + 196 + ], + "type": "image", + "image_path": "2e1d0837f706a9b08c4ab0f87b4538fa164fda9549434708f24f5b19da9cb0b9.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 290, + 161, + 321, + 196 + ], + "blocks": [ + { + "bbox": [ + 290, + 161, + 321, + 196 + ], + "lines": [ + { + "bbox": [ + 290, + 161, + 321, + 196 + ], + "spans": [ + { + "bbox": [ + 290, + 161, + 321, + 196 + ], + "type": "image", + "image_path": "3e39e3a2689d98d1769913a61af779a597f756aa387e342f041b090764edca3b.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 67, + 196, + 105, + 232 + ], + "blocks": [ + { + "bbox": [ + 67, + 196, + 105, + 232 + ], + "lines": [ + { + "bbox": [ + 67, + 196, + 105, + 232 + ], + "spans": [ + { + "bbox": [ + 67, + 196, + 105, + 232 + ], + "type": "image", + "image_path": "696f4d052ce3c5f19bf95c64f001978babcabf997d82908af7b9d839f0bcd72c.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 108, + 196, + 180, + 232 + ], + "blocks": [ + { + "bbox": [ + 108, + 196, + 180, + 232 + ], + "lines": [ + { + "bbox": [ + 108, + 196, + 180, + 232 + ], + "spans": [ + { + "bbox": [ + 108, + 196, + 180, + 232 + ], + "type": "image", + "image_path": "d58917e4655c08349c2094b04122ac403cc69c0bbf4b8f18e63fa3d2e85ef385.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 181, + 196, + 197, + 232 + ], + "blocks": [ + { + "bbox": [ + 181, + 196, + 197, + 232 + ], + "lines": [ + { + "bbox": [ + 181, + 196, + 197, + 232 + ], + "spans": [ + { + "bbox": [ + 181, + 196, + 197, + 232 + ], + "type": "image", + "image_path": "72da2da62ae289ed3e43f91e6c5a56bcb74fc726d7d7908b51d1c70e6955595e.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 202, + 199, + 246, + 232 + ], + "blocks": [ + { + "bbox": [ + 202, + 199, + 246, + 232 + ], + "lines": [ + { + "bbox": [ + 202, + 199, + 246, + 232 + ], + "spans": [ + { + "bbox": [ + 202, + 199, + 246, + 232 + ], + "type": "image", + "image_path": "05814bacd1f35b02ddfce6af1a0ba537fcf64dd5d453eedb3e16e1e85e51b66f.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 252, + 199, + 278, + 232 + ], + "blocks": [ + { + "bbox": [ + 252, + 199, + 278, + 232 + ], + "lines": [ + { + "bbox": [ + 252, + 199, + 278, + 232 + ], + "spans": [ + { + "bbox": [ + 252, + 199, + 278, + 232 + ], + "type": "image", + "image_path": "0212b1361b22cb910d7634c492aaae91cacf06ce4b1d613c990890104a6a9564.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 286, + 199, + 317, + 232 + ], + "blocks": [ + { + "bbox": [ + 286, + 199, + 317, + 232 + ], + "lines": [ + { + "bbox": [ + 286, + 199, + 317, + 232 + ], + "spans": [ + { + "bbox": [ + 286, + 199, + 317, + 232 + ], + "type": "image", + "image_path": "46e4388a529e6d40706c096f8d2775bd357b7a64aa279500e69af3824b9e0460.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 67, + 232, + 105, + 267 + ], + "blocks": [ + { + "bbox": [ + 67, + 232, + 105, + 267 + ], + "lines": [ + { + "bbox": [ + 67, + 232, + 105, + 267 + ], + "spans": [ + { + "bbox": [ + 67, + 232, + 105, + 267 + ], + "type": "image", + "image_path": "d3cf9e3ba4093efd3b33ec928ce15c2f6468ba2b19b78fc7674286b0d64a06e7.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 108, + 232, + 175, + 269 + ], + "blocks": [ + { + "bbox": [ + 108, + 232, + 175, + 269 + ], + "lines": [ + { + "bbox": [ + 108, + 232, + 175, + 269 + ], + "spans": [ + { + "bbox": [ + 108, + 232, + 175, + 269 + ], + "type": "image", + "image_path": "2aa4ce74db8d2d98f6910bf0327bb6bdadcf34e8325cf758f06036be5a3b81b1.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 271, + 555, + 316 + ], + "lines": [ + { + "bbox": [ + 55, + 271, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 555, + 316 + ], + "type": "text", + "content": "Figure 8. Visual comparison on in-the-wild data. FaceLift demonstrates great generalization ability and robustness towards in-the-wild images, provides realistic unseen view rendering results. Era3D [31] and LGM [56] generate 3D head representation in inaccurate shape. PanoHead [1] often creates severe artifacts on the back of the head and can not handle challenging hairstyles well. Dual Encoder [4] shows improved performance on reconstructing the back of the head but exhibits more pronounced identity loss." + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 181, + 232, + 197, + 267 + ], + "blocks": [ + { + "bbox": [ + 181, + 232, + 197, + 267 + ], + "lines": [ + { + "bbox": [ + 181, + 232, + 197, + 267 + ], + "spans": [ + { + "bbox": [ + 181, + 232, + 197, + 267 + ], + "type": "image", + "image_path": "f37bf9c3434ce7c15b1274dc1172f907759074931b2f3c942c728774f0b777e8.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 201, + 232, + 244, + 267 + ], + "blocks": [ + { + "bbox": [ + 201, + 232, + 244, + 267 + ], + "lines": [ + { + "bbox": [ + 201, + 232, + 244, + 267 + ], + "spans": [ + { + "bbox": [ + 201, + 232, + 244, + 267 + ], + "type": "image", + "image_path": "172da7f7e278e703a269c702bd783b56d9abbce69276359ac717bc5daefd8ff2.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 252, + 232, + 278, + 267 + ], + "blocks": [ + { + "bbox": [ + 252, + 232, + 278, + 267 + ], + "lines": [ + { + "bbox": [ + 252, + 232, + 278, + 267 + ], + "spans": [ + { + "bbox": [ + 252, + 232, + 278, + 267 + ], + "type": "image", + "image_path": "d5ec3caa8da854c4b92fa715f1c32e590f1b172cd4dac059991d87e461296289.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 286, + 232, + 317, + 267 + ], + "blocks": [ + { + "bbox": [ + 286, + 232, + 317, + 267 + ], + "lines": [ + { + "bbox": [ + 286, + 232, + 317, + 267 + ], + "spans": [ + { + "bbox": [ + 286, + 232, + 317, + 267 + ], + "type": "image", + "image_path": "710e6f5cb150b5fc31ed817dc4f49d5752165dc91e4f9ee46ab0c84824167814.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 326, + 89, + 361, + 161 + ], + "blocks": [ + { + "bbox": [ + 327, + 82, + 531, + 89 + ], + "lines": [ + { + "bbox": [ + 327, + 82, + 531, + 89 + ], + "spans": [ + { + "bbox": [ + 327, + 82, + 531, + 89 + ], + "type": "text", + "content": "PanoHead [1] Dual Encoder [4] Era3D [31] LGM [56] Our MV+LGM [56] FaceLift" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 326, + 89, + 361, + 161 + ], + "lines": [ + { + "bbox": [ + 326, + 89, + 361, + 161 + ], + "spans": [ + { + "bbox": [ + 326, + 89, + 361, + 161 + ], + "type": "image", + "image_path": "47c2863441c19e80b5aedf749d4dbabf62f9b6618d13d6cd235e39025d501fc1.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 362, + 89, + 396, + 161 + ], + "blocks": [ + { + "bbox": [ + 362, + 89, + 396, + 161 + ], + "lines": [ + { + "bbox": [ + 362, + 89, + 396, + 161 + ], + "spans": [ + { + "bbox": [ + 362, + 89, + 396, + 161 + ], + "type": "image", + "image_path": "ace96edd69222586fb87f8347f0863719cb572d97b979a9f4d92b387d391f481.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 399, + 89, + 431, + 161 + ], + "blocks": [ + { + "bbox": [ + 408, + 72, + 457, + 80 + ], + "lines": [ + { + "bbox": [ + 408, + 72, + 457, + 80 + ], + "spans": [ + { + "bbox": [ + 408, + 72, + 457, + 80 + ], + "type": "text", + "content": "Rendered View 2" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 399, + 89, + 431, + 161 + ], + "lines": [ + { + "bbox": [ + 399, + 89, + 431, + 161 + ], + "spans": [ + { + "bbox": [ + 399, + 89, + 431, + 161 + ], + "type": "image", + "image_path": "b26439310673bc3889fa5791cd9d216ba2073ffd5d7890bb3200dd36a2795567.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 434, + 89, + 463, + 161 + ], + "blocks": [ + { + "bbox": [ + 434, + 89, + 463, + 161 + ], + "lines": [ + { + "bbox": [ + 434, + 89, + 463, + 161 + ], + "spans": [ + { + "bbox": [ + 434, + 89, + 463, + 161 + ], + "type": "image", + "image_path": "56e54984847429d6c95620228b5c37ce816762af2e9f9fb2b9d7c4e459b4255a.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 466, + 89, + 539, + 161 + ], + "blocks": [ + { + "bbox": [ + 466, + 89, + 539, + 161 + ], + "lines": [ + { + "bbox": [ + 466, + 89, + 539, + 161 + ], + "spans": [ + { + "bbox": [ + 466, + 89, + 539, + 161 + ], + "type": "image", + "image_path": "c26be89bef57287162a4867cf1d90aabc0bf520cbe59d16d5522094c264b15a3.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 329, + 161, + 361, + 196 + ], + "blocks": [ + { + "bbox": [ + 329, + 161, + 361, + 196 + ], + "lines": [ + { + "bbox": [ + 329, + 161, + 361, + 196 + ], + "spans": [ + { + "bbox": [ + 329, + 161, + 361, + 196 + ], + "type": "image", + "image_path": "7066f7b2c616b077c658c3772b0cc4429b459becb137ba38340d398269f1cba8.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 362, + 161, + 396, + 196 + ], + "blocks": [ + { + "bbox": [ + 362, + 161, + 396, + 196 + ], + "lines": [ + { + "bbox": [ + 362, + 161, + 396, + 196 + ], + "spans": [ + { + "bbox": [ + 362, + 161, + 396, + 196 + ], + "type": "image", + "image_path": "ddf09e78a09547973e498c9f6610325bd8a0f3463556bc30b65c666bf1b9e8d9.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 399, + 161, + 425, + 196 + ], + "blocks": [ + { + "bbox": [ + 399, + 161, + 425, + 196 + ], + "lines": [ + { + "bbox": [ + 399, + 161, + 425, + 196 + ], + "spans": [ + { + "bbox": [ + 399, + 161, + 425, + 196 + ], + "type": "image", + "image_path": "da4ab45a31af054d60263a797f55f1fdea9cc121962a01793c2101d4bc962ddf.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 434, + 161, + 462, + 196 + ], + "blocks": [ + { + "bbox": [ + 434, + 161, + 462, + 196 + ], + "lines": [ + { + "bbox": [ + 434, + 161, + 462, + 196 + ], + "spans": [ + { + "bbox": [ + 434, + 161, + 462, + 196 + ], + "type": "image", + "image_path": "30f20024d208f23e6a6810bb257058d1fe955ccdcb7682109ecc8a2e266c9613.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 465, + 161, + 499, + 196 + ], + "blocks": [ + { + "bbox": [ + 465, + 161, + 499, + 196 + ], + "lines": [ + { + "bbox": [ + 465, + 161, + 499, + 196 + ], + "spans": [ + { + "bbox": [ + 465, + 161, + 499, + 196 + ], + "type": "image", + "image_path": "1737dc64f07ef3ed5c598bdab4e2485fbf36f624e45976982b9848a2b220ae4c.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 510, + 161, + 534, + 196 + ], + "blocks": [ + { + "bbox": [ + 510, + 161, + 534, + 196 + ], + "lines": [ + { + "bbox": [ + 510, + 161, + 534, + 196 + ], + "spans": [ + { + "bbox": [ + 510, + 161, + 534, + 196 + ], + "type": "image", + "image_path": "2a4dc6f3aed0654eb50d32d2a63dead013b0829af13487a75bc164186eeb4c62.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 329, + 196, + 361, + 232 + ], + "blocks": [ + { + "bbox": [ + 329, + 196, + 361, + 232 + ], + "lines": [ + { + "bbox": [ + 329, + 196, + 361, + 232 + ], + "spans": [ + { + "bbox": [ + 329, + 196, + 361, + 232 + ], + "type": "image", + "image_path": "de8e7144457bcc8acf398e4febcd56dd4d53b5fb644011ed9a52767fe2bf7468.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 362, + 196, + 396, + 232 + ], + "blocks": [ + { + "bbox": [ + 362, + 196, + 396, + 232 + ], + "lines": [ + { + "bbox": [ + 362, + 196, + 396, + 232 + ], + "spans": [ + { + "bbox": [ + 362, + 196, + 396, + 232 + ], + "type": "image", + "image_path": "e4fb5e0209f1dfd285657ba099204ed1fa59ed6dbe12ea97d7b0454e6c0d201d.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 399, + 196, + 425, + 232 + ], + "blocks": [ + { + "bbox": [ + 399, + 196, + 425, + 232 + ], + "lines": [ + { + "bbox": [ + 399, + 196, + 425, + 232 + ], + "spans": [ + { + "bbox": [ + 399, + 196, + 425, + 232 + ], + "type": "image", + "image_path": "438c22936c88de444262a16e1b303cdb8fa4e1be998c6b3046c83afa1c51252b.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_body" + } + ], + "index": 49 + }, + { + "type": "image", + "bbox": [ + 434, + 196, + 462, + 232 + ], + "blocks": [ + { + "bbox": [ + 434, + 196, + 462, + 232 + ], + "lines": [ + { + "bbox": [ + 434, + 196, + 462, + 232 + ], + "spans": [ + { + "bbox": [ + 434, + 196, + 462, + 232 + ], + "type": "image", + "image_path": "4a03516c85a51a9b9af704f43d7796eec20b9f79650a2d39321377486522779e.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 465, + 196, + 499, + 232 + ], + "blocks": [ + { + "bbox": [ + 465, + 196, + 499, + 232 + ], + "lines": [ + { + "bbox": [ + 465, + 196, + 499, + 232 + ], + "spans": [ + { + "bbox": [ + 465, + 196, + 499, + 232 + ], + "type": "image", + "image_path": "ccd0b9c4810f97b5138b72d27af938aabbf1d6f92a69aa52b3d18a674d579a1c.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_body" + } + ], + "index": 51 + }, + { + "type": "image", + "bbox": [ + 510, + 196, + 534, + 232 + ], + "blocks": [ + { + "bbox": [ + 510, + 196, + 534, + 232 + ], + "lines": [ + { + "bbox": [ + 510, + 196, + 534, + 232 + ], + "spans": [ + { + "bbox": [ + 510, + 196, + 534, + 232 + ], + "type": "image", + "image_path": "7c4715b6f10acfb1b97a2ed03f2b6778f70052155dab2044d364025dea44717d.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 329, + 232, + 361, + 267 + ], + "blocks": [ + { + "bbox": [ + 329, + 232, + 361, + 267 + ], + "lines": [ + { + "bbox": [ + 329, + 232, + 361, + 267 + ], + "spans": [ + { + "bbox": [ + 329, + 232, + 361, + 267 + ], + "type": "image", + "image_path": "09ee325345a2110047a36baad93a88a19ec3f70b1c948050a89e9b86940e22a5.jpg" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_body" + } + ], + "index": 53 + }, + { + "type": "image", + "bbox": [ + 362, + 232, + 396, + 267 + ], + "blocks": [ + { + "bbox": [ + 362, + 232, + 396, + 267 + ], + "lines": [ + { + "bbox": [ + 362, + 232, + 396, + 267 + ], + "spans": [ + { + "bbox": [ + 362, + 232, + 396, + 267 + ], + "type": "image", + "image_path": "cb2cf711bd9d8917bc7a8245c3393acf14bbeb0dea8d1415ce422cceb5904123.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_body" + } + ], + "index": 54 + }, + { + "type": "image", + "bbox": [ + 399, + 232, + 425, + 267 + ], + "blocks": [ + { + "bbox": [ + 399, + 232, + 425, + 267 + ], + "lines": [ + { + "bbox": [ + 399, + 232, + 425, + 267 + ], + "spans": [ + { + "bbox": [ + 399, + 232, + 425, + 267 + ], + "type": "image", + "image_path": "561bd2eb28cc3aef32e81ff8a88b25dda3c6478c6d52168a6b247cff83374acc.jpg" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_body" + } + ], + "index": 55 + }, + { + "type": "image", + "bbox": [ + 434, + 232, + 462, + 267 + ], + "blocks": [ + { + "bbox": [ + 434, + 232, + 462, + 267 + ], + "lines": [ + { + "bbox": [ + 434, + 232, + 462, + 267 + ], + "spans": [ + { + "bbox": [ + 434, + 232, + 462, + 267 + ], + "type": "image", + "image_path": "9d71026ac245a36a82b48ad3d4ed357ba863c22dbcf5b8f61a8891df941b285b.jpg" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_body" + } + ], + "index": 56 + }, + { + "type": "image", + "bbox": [ + 465, + 232, + 499, + 267 + ], + "blocks": [ + { + "bbox": [ + 465, + 232, + 499, + 267 + ], + "lines": [ + { + "bbox": [ + 465, + 232, + 499, + 267 + ], + "spans": [ + { + "bbox": [ + 465, + 232, + 499, + 267 + ], + "type": "image", + "image_path": "099fdb90dea4a100c4550f464f03d459e772735877fe98ca812b5033d7bda469.jpg" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_body" + } + ], + "index": 57 + }, + { + "type": "image", + "bbox": [ + 510, + 232, + 534, + 267 + ], + "blocks": [ + { + "bbox": [ + 510, + 232, + 534, + 267 + ], + "lines": [ + { + "bbox": [ + 510, + 232, + 534, + 267 + ], + "spans": [ + { + "bbox": [ + 510, + 232, + 534, + 267 + ], + "type": "image", + "image_path": "3d4d5023c1837e395efcf3a5c7ac29203d5221a45df07dd4437f49bf435dcce9.jpg" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_body" + } + ], + "index": 58 + }, + { + "type": "image", + "bbox": [ + 67, + 331, + 128, + 390 + ], + "blocks": [ + { + "bbox": [ + 87, + 319, + 110, + 330 + ], + "lines": [ + { + "bbox": [ + 87, + 319, + 110, + 330 + ], + "spans": [ + { + "bbox": [ + 87, + 319, + 110, + 330 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 331, + 128, + 390 + ], + "lines": [ + { + "bbox": [ + 67, + 331, + 128, + 390 + ], + "spans": [ + { + "bbox": [ + 67, + 331, + 128, + 390 + ], + "type": "image", + "image_path": "e683f50c292b647db5c575035bab36fb43412c8b3f05847fbe142c4508d581b4.jpg" + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_body" + } + ], + "index": 61 + }, + { + "type": "image", + "bbox": [ + 67, + 390, + 128, + 450 + ], + "blocks": [ + { + "bbox": [ + 67, + 390, + 128, + 450 + ], + "lines": [ + { + "bbox": [ + 67, + 390, + 128, + 450 + ], + "spans": [ + { + "bbox": [ + 67, + 390, + 128, + 450 + ], + "type": "image", + "image_path": "42170feec4828d1edb96b6c450a01e42b79e40ca51819a9474c34962adad3f6d.jpg" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "image_body" + } + ], + "index": 62 + }, + { + "type": "image", + "bbox": [ + 67, + 450, + 128, + 508 + ], + "blocks": [ + { + "bbox": [ + 67, + 450, + 128, + 508 + ], + "lines": [ + { + "bbox": [ + 67, + 450, + 128, + 508 + ], + "spans": [ + { + "bbox": [ + 67, + 450, + 128, + 508 + ], + "type": "image", + "image_path": "8b33b5ae6bdf1db4ce89133b415ee9846e52bdb3115a327c54d8f23012c5c99f.jpg" + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "image_body" + } + ], + "index": 63 + }, + { + "type": "image", + "bbox": [ + 67, + 509, + 128, + 567 + ], + "blocks": [ + { + "bbox": [ + 67, + 509, + 128, + 567 + ], + "lines": [ + { + "bbox": [ + 67, + 509, + 128, + 567 + ], + "spans": [ + { + "bbox": [ + 67, + 509, + 128, + 567 + ], + "type": "image", + "image_path": "819e68b1235ba896362723419c8cc731982f6d4871b4840efe150517ecd6f469.jpg" + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 570, + 555, + 616 + ], + "lines": [ + { + "bbox": [ + 55, + 570, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 555, + 616 + ], + "type": "text", + "content": "Figure 9. Results of FaceLift on in-the-wild images. FaceLift accurately reconstructs 3D head models under challenging lighting conditions, achieving high fidelity (row 1). It captures fine facial details such as wrinkles (row 2), mustaches (row 3), and individual hairs (row 2 and row 4). Additionally, it remains robust to complex facial expressions (row 3) and various skin tones (row 4). Furthermore, it can realistically reconstruct facial paint (row 4). More results are provided in the supplementary materials." + } + ] + } + ], + "index": 96, + "angle": 0, + "type": "image_caption" + } + ], + "index": 64 + }, + { + "type": "image", + "bbox": [ + 130, + 331, + 184, + 391 + ], + "blocks": [ + { + "bbox": [ + 171, + 319, + 260, + 330 + ], + "lines": [ + { + "bbox": [ + 171, + 319, + 260, + 330 + ], + "spans": [ + { + "bbox": [ + 171, + 319, + 260, + 330 + ], + "type": "text", + "content": "Novel View Renderings" + } + ] + } + ], + "index": 65, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 331, + 184, + 391 + ], + "lines": [ + { + "bbox": [ + 130, + 331, + 184, + 391 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 184, + 391 + ], + "type": "image", + "image_path": "31b9b6ac088ee01197536f8c4cf062a39dce10db1351de3dcfe95bffa6cf855c.jpg" + } + ] + } + ], + "index": 66, + "angle": 0, + "type": "image_body" + } + ], + "index": 66 + }, + { + "type": "image", + "bbox": [ + 130, + 391, + 184, + 450 + ], + "blocks": [ + { + "bbox": [ + 130, + 391, + 184, + 450 + ], + "lines": [ + { + "bbox": [ + 130, + 391, + 184, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 184, + 450 + ], + "type": "image", + "image_path": "7430f188b9bec33669050f87230094e1a2524890bd260facc91ac1a8fe599001.jpg" + } + ] + } + ], + "index": 67, + "angle": 0, + "type": "image_body" + } + ], + "index": 67 + }, + { + "type": "image", + "bbox": [ + 130, + 451, + 184, + 508 + ], + "blocks": [ + { + "bbox": [ + 130, + 451, + 184, + 508 + ], + "lines": [ + { + "bbox": [ + 130, + 451, + 184, + 508 + ], + "spans": [ + { + "bbox": [ + 130, + 451, + 184, + 508 + ], + "type": "image", + "image_path": "7ff410bb307ecbddb194e78b5765e7674840f056624b9e72656264fdfc1c0126.jpg" + } + ] + } + ], + "index": 68, + "angle": 0, + "type": "image_body" + } + ], + "index": 68 + }, + { + "type": "image", + "bbox": [ + 130, + 509, + 184, + 567 + ], + "blocks": [ + { + "bbox": [ + 130, + 509, + 184, + 567 + ], + "lines": [ + { + "bbox": [ + 130, + 509, + 184, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 184, + 567 + ], + "type": "image", + "image_path": "d39c3f9f266a4a8a7093718cb754410db28d686221570ccbfdbe815246d2024b.jpg" + } + ] + } + ], + "index": 69, + "angle": 0, + "type": "image_body" + } + ], + "index": 69 + }, + { + "type": "image", + "bbox": [ + 188, + 331, + 241, + 391 + ], + "blocks": [ + { + "bbox": [ + 188, + 331, + 241, + 391 + ], + "lines": [ + { + "bbox": [ + 188, + 331, + 241, + 391 + ], + "spans": [ + { + "bbox": [ + 188, + 331, + 241, + 391 + ], + "type": "image", + "image_path": "34ae4dcb80d9089ae45d3bf989b1c6f0eb257eaffc1a0f56aa98aec975e8e7fb.jpg" + } + ] + } + ], + "index": 70, + "angle": 0, + "type": "image_body" + } + ], + "index": 70 + }, + { + "type": "image", + "bbox": [ + 189, + 391, + 242, + 450 + ], + "blocks": [ + { + "bbox": [ + 189, + 391, + 242, + 450 + ], + "lines": [ + { + "bbox": [ + 189, + 391, + 242, + 450 + ], + "spans": [ + { + "bbox": [ + 189, + 391, + 242, + 450 + ], + "type": "image", + "image_path": "683198bc99409c4fcc6863731d3f51f4666df54bc70be6032f19bced133ff30d.jpg" + } + ] + } + ], + "index": 71, + "angle": 0, + "type": "image_body" + } + ], + "index": 71 + }, + { + "type": "image", + "bbox": [ + 189, + 451, + 236, + 508 + ], + "blocks": [ + { + "bbox": [ + 189, + 451, + 236, + 508 + ], + "lines": [ + { + "bbox": [ + 189, + 451, + 236, + 508 + ], + "spans": [ + { + "bbox": [ + 189, + 451, + 236, + 508 + ], + "type": "image", + "image_path": "e92968777d293309531cc31cf7abe476fe5428c6a0a55f910450a09bc4b3fc91.jpg" + } + ] + } + ], + "index": 72, + "angle": 0, + "type": "image_body" + } + ], + "index": 72 + }, + { + "type": "image", + "bbox": [ + 189, + 509, + 247, + 567 + ], + "blocks": [ + { + "bbox": [ + 189, + 509, + 247, + 567 + ], + "lines": [ + { + "bbox": [ + 189, + 509, + 247, + 567 + ], + "spans": [ + { + "bbox": [ + 189, + 509, + 247, + 567 + ], + "type": "image", + "image_path": "f9ca2e1dcb33c7f2df878ac41e053cb67481f0cb8c7ca68938e5867573574167.jpg" + } + ] + } + ], + "index": 73, + "angle": 0, + "type": "image_body" + } + ], + "index": 73 + }, + { + "type": "image", + "bbox": [ + 244, + 331, + 304, + 390 + ], + "blocks": [ + { + "bbox": [ + 244, + 331, + 304, + 390 + ], + "lines": [ + { + "bbox": [ + 244, + 331, + 304, + 390 + ], + "spans": [ + { + "bbox": [ + 244, + 331, + 304, + 390 + ], + "type": "image", + "image_path": "c1648b172f1ab59b1b24a7e8bd1309e1bad1de862f0ff6cd765edce6db96c411.jpg" + } + ] + } + ], + "index": 74, + "angle": 0, + "type": "image_body" + } + ], + "index": 74 + }, + { + "type": "image", + "bbox": [ + 244, + 390, + 304, + 450 + ], + "blocks": [ + { + "bbox": [ + 244, + 390, + 304, + 450 + ], + "lines": [ + { + "bbox": [ + 244, + 390, + 304, + 450 + ], + "spans": [ + { + "bbox": [ + 244, + 390, + 304, + 450 + ], + "type": "image", + "image_path": "43aa0531600db240ce378e530d06040e6fe0d2337b660ffd362a46f2f9199ab5.jpg" + } + ] + } + ], + "index": 75, + "angle": 0, + "type": "image_body" + } + ], + "index": 75 + }, + { + "type": "image", + "bbox": [ + 244, + 450, + 304, + 508 + ], + "blocks": [ + { + "bbox": [ + 244, + 450, + 304, + 508 + ], + "lines": [ + { + "bbox": [ + 244, + 450, + 304, + 508 + ], + "spans": [ + { + "bbox": [ + 244, + 450, + 304, + 508 + ], + "type": "image", + "image_path": "3f1f0c8c68e741e2c63a4825dad23ca25c7b49785138ffbcb1b40e4af4961102.jpg" + } + ] + } + ], + "index": 76, + "angle": 0, + "type": "image_body" + } + ], + "index": 76 + }, + { + "type": "image", + "bbox": [ + 244, + 509, + 304, + 567 + ], + "blocks": [ + { + "bbox": [ + 244, + 509, + 304, + 567 + ], + "lines": [ + { + "bbox": [ + 244, + 509, + 304, + 567 + ], + "spans": [ + { + "bbox": [ + 244, + 509, + 304, + 567 + ], + "type": "image", + "image_path": "8941c9ea4b80dddf5f0ef6bb2ac47ae90a1f90ef6631159075c286a445eaa09d.jpg" + } + ] + } + ], + "index": 77, + "angle": 0, + "type": "image_body" + } + ], + "index": 77 + }, + { + "type": "image", + "bbox": [ + 304, + 331, + 363, + 390 + ], + "blocks": [ + { + "bbox": [ + 322, + 319, + 345, + 329 + ], + "lines": [ + { + "bbox": [ + 322, + 319, + 345, + 329 + ], + "spans": [ + { + "bbox": [ + 322, + 319, + 345, + 329 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 78, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 331, + 363, + 390 + ], + "lines": [ + { + "bbox": [ + 304, + 331, + 363, + 390 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 363, + 390 + ], + "type": "image", + "image_path": "d72b93463faa8e38d78ec8ecb6768e971317bcdd882f1650c99ce80ad3a31bb1.jpg" + } + ] + } + ], + "index": 79, + "angle": 0, + "type": "image_body" + } + ], + "index": 79 + }, + { + "type": "image", + "bbox": [ + 304, + 390, + 363, + 450 + ], + "blocks": [ + { + "bbox": [ + 304, + 390, + 363, + 450 + ], + "lines": [ + { + "bbox": [ + 304, + 390, + 363, + 450 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 363, + 450 + ], + "type": "image", + "image_path": "9efa900195c53c3b5d44bdeb74d2d59a2d3c717b2fdb144950a482519b60d3d0.jpg" + } + ] + } + ], + "index": 80, + "angle": 0, + "type": "image_body" + } + ], + "index": 80 + }, + { + "type": "image", + "bbox": [ + 304, + 450, + 363, + 508 + ], + "blocks": [ + { + "bbox": [ + 304, + 450, + 363, + 508 + ], + "lines": [ + { + "bbox": [ + 304, + 450, + 363, + 508 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 363, + 508 + ], + "type": "image", + "image_path": "f7dbb95ae514a93278cc78a94fc10693c895095d85fe4cf11de5d27f27cd8985.jpg" + } + ] + } + ], + "index": 81, + "angle": 0, + "type": "image_body" + } + ], + "index": 81 + }, + { + "type": "image", + "bbox": [ + 304, + 509, + 363, + 567 + ], + "blocks": [ + { + "bbox": [ + 304, + 509, + 363, + 567 + ], + "lines": [ + { + "bbox": [ + 304, + 509, + 363, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 363, + 567 + ], + "type": "image", + "image_path": "87cde82c9a07dac1f957010a830505824604793aab71805061bbd7fb385c454d.jpg" + } + ] + } + ], + "index": 82, + "angle": 0, + "type": "image_body" + } + ], + "index": 82 + }, + { + "type": "image", + "bbox": [ + 370, + 338, + 415, + 391 + ], + "blocks": [ + { + "bbox": [ + 370, + 338, + 415, + 391 + ], + "lines": [ + { + "bbox": [ + 370, + 338, + 415, + 391 + ], + "spans": [ + { + "bbox": [ + 370, + 338, + 415, + 391 + ], + "type": "image", + "image_path": "a2b6ffcdbe0dd09b893bcacbf1d277be6ead5309d2e6bdb449e7abc7828eca15.jpg" + } + ] + } + ], + "index": 84, + "angle": 0, + "type": "image_body" + } + ], + "index": 84 + }, + { + "type": "image", + "bbox": [ + 368, + 391, + 420, + 450 + ], + "blocks": [ + { + "bbox": [ + 368, + 391, + 420, + 450 + ], + "lines": [ + { + "bbox": [ + 368, + 391, + 420, + 450 + ], + "spans": [ + { + "bbox": [ + 368, + 391, + 420, + 450 + ], + "type": "image", + "image_path": "717aa3c10193708c1b500a117a3960f9b232e022038a0f92b248ecff098bce3a.jpg" + } + ] + } + ], + "index": 85, + "angle": 0, + "type": "image_body" + } + ], + "index": 85 + }, + { + "type": "image", + "bbox": [ + 368, + 451, + 421, + 508 + ], + "blocks": [ + { + "bbox": [ + 368, + 451, + 421, + 508 + ], + "lines": [ + { + "bbox": [ + 368, + 451, + 421, + 508 + ], + "spans": [ + { + "bbox": [ + 368, + 451, + 421, + 508 + ], + "type": "image", + "image_path": "5af16ca856d791c86904c9e7c58edb17bf5a21a5fad59e4135d05f832ff28613.jpg" + } + ] + } + ], + "index": 86, + "angle": 0, + "type": "image_body" + } + ], + "index": 86 + }, + { + "type": "image", + "bbox": [ + 368, + 509, + 421, + 567 + ], + "blocks": [ + { + "bbox": [ + 368, + 509, + 421, + 567 + ], + "lines": [ + { + "bbox": [ + 368, + 509, + 421, + 567 + ], + "spans": [ + { + "bbox": [ + 368, + 509, + 421, + 567 + ], + "type": "image", + "image_path": "aea72c41915ba4a76ede0be99b6097c3d31b134ed8f9a5b9a5510324e84691f3.jpg" + } + ] + } + ], + "index": 87, + "angle": 0, + "type": "image_body" + } + ], + "index": 87 + }, + { + "type": "image", + "bbox": [ + 422, + 338, + 480, + 390 + ], + "blocks": [ + { + "bbox": [ + 422, + 338, + 480, + 390 + ], + "lines": [ + { + "bbox": [ + 422, + 338, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 422, + 338, + 480, + 390 + ], + "type": "image", + "image_path": "a9fa10bb12b4e86b4e2a5b9554652d13d4b5be78baa8a5d642861e887ded452b.jpg" + } + ] + } + ], + "index": 88, + "angle": 0, + "type": "image_body" + } + ], + "index": 88 + }, + { + "type": "image", + "bbox": [ + 422, + 390, + 480, + 450 + ], + "blocks": [ + { + "bbox": [ + 422, + 390, + 480, + 450 + ], + "lines": [ + { + "bbox": [ + 422, + 390, + 480, + 450 + ], + "spans": [ + { + "bbox": [ + 422, + 390, + 480, + 450 + ], + "type": "image", + "image_path": "c2635d201159ca66dfd55b6dca23cab5c5fd13d180414a0405816a4146bec799.jpg" + } + ] + } + ], + "index": 89, + "angle": 0, + "type": "image_body" + } + ], + "index": 89 + }, + { + "type": "image", + "bbox": [ + 422, + 451, + 480, + 508 + ], + "blocks": [ + { + "bbox": [ + 422, + 451, + 480, + 508 + ], + "lines": [ + { + "bbox": [ + 422, + 451, + 480, + 508 + ], + "spans": [ + { + "bbox": [ + 422, + 451, + 480, + 508 + ], + "type": "image", + "image_path": "b6aa15dfd416726563ae38f384d19e01503a354be8033458091da1279eed2119.jpg" + } + ] + } + ], + "index": 90, + "angle": 0, + "type": "image_body" + } + ], + "index": 90 + }, + { + "type": "image", + "bbox": [ + 422, + 509, + 479, + 567 + ], + "blocks": [ + { + "bbox": [ + 422, + 509, + 479, + 567 + ], + "lines": [ + { + "bbox": [ + 422, + 509, + 479, + 567 + ], + "spans": [ + { + "bbox": [ + 422, + 509, + 479, + 567 + ], + "type": "image", + "image_path": "b7c5b1a7728821d32edc5ea79be51b769a259ad9c1df9d44feed7e060c0f8498.jpg" + } + ] + } + ], + "index": 91, + "angle": 0, + "type": "image_body" + } + ], + "index": 91 + }, + { + "type": "image", + "bbox": [ + 480, + 331, + 540, + 390 + ], + "blocks": [ + { + "bbox": [ + 406, + 319, + 495, + 329 + ], + "lines": [ + { + "bbox": [ + 406, + 319, + 495, + 329 + ], + "spans": [ + { + "bbox": [ + 406, + 319, + 495, + 329 + ], + "type": "text", + "content": "Novel View Renderings" + } + ] + } + ], + "index": 83, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 480, + 331, + 540, + 390 + ], + "lines": [ + { + "bbox": [ + 480, + 331, + 540, + 390 + ], + "spans": [ + { + "bbox": [ + 480, + 331, + 540, + 390 + ], + "type": "image", + "image_path": "c84e8a014be493022565309f86b61032cc929027898b7e4ef17e288082d0c5ce.jpg" + } + ] + } + ], + "index": 92, + "angle": 0, + "type": "image_body" + } + ], + "index": 92 + }, + { + "type": "image", + "bbox": [ + 480, + 390, + 540, + 450 + ], + "blocks": [ + { + "bbox": [ + 480, + 390, + 540, + 450 + ], + "lines": [ + { + "bbox": [ + 480, + 390, + 540, + 450 + ], + "spans": [ + { + "bbox": [ + 480, + 390, + 540, + 450 + ], + "type": "image", + "image_path": "85268e301c7331bb8d3cec1496bab75a3fd89334a98b307c18c8376e7f107e41.jpg" + } + ] + } + ], + "index": 93, + "angle": 0, + "type": "image_body" + } + ], + "index": 93 + }, + { + "type": "image", + "bbox": [ + 480, + 450, + 540, + 508 + ], + "blocks": [ + { + "bbox": [ + 480, + 450, + 540, + 508 + ], + "lines": [ + { + "bbox": [ + 480, + 450, + 540, + 508 + ], + "spans": [ + { + "bbox": [ + 480, + 450, + 540, + 508 + ], + "type": "image", + "image_path": "d79dbaaa377eb9fd4682ec3754d698fc68c94dd35cec04bba5889d84a7d1e356.jpg" + } + ] + } + ], + "index": 94, + "angle": 0, + "type": "image_body" + } + ], + "index": 94 + }, + { + "type": "image", + "bbox": [ + 480, + 509, + 540, + 567 + ], + "blocks": [ + { + "bbox": [ + 480, + 509, + 540, + 567 + ], + "lines": [ + { + "bbox": [ + 480, + 509, + 540, + 567 + ], + "spans": [ + { + "bbox": [ + 480, + 509, + 540, + 567 + ], + "type": "image", + "image_path": "779e4cd33e63249246a59662c223cee50077538ef23cba0b3eb77f8ce929f627.jpg" + } + ] + } + ], + "index": 95, + "angle": 0, + "type": "image_body" + } + ], + "index": 95 + }, + { + "bbox": [ + 55, + 624, + 297, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 297, + 708 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 297, + 708 + ], + "type": "text", + "content": "back-of-head rendering but suffers from severe identity loss (row 2) and fails to accurately reconstruct face paint (row 4). Era3D [31] often produces an inaccurate head shape, particularly from side views, and offers fewer geometric details compared to FaceLift. LGM [56] generates Gaussians with inaccurate color and opacity and lacks proper facial geometry, resulting in distorted features. Baseline Our " + }, + { + "bbox": [ + 55, + 624, + 297, + 708 + ], + "type": "inline_equation", + "content": "MV + LGM" + } + ] + } + ], + "index": 97 + }, + { + "bbox": [ + 313, + 624, + 555, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 555, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 555, + 685 + ], + "type": "text", + "content": "shows that our multi-view diffusion model enhances LGM by providing improved facial geometry and texture details. However, the LGM reconstructor still produces Gaussians with inaccurate shapes and opacities, further underscoring the advantages of our transformer-based reconstructor." + } + ] + } + ], + "index": 98 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "We present more FaceLift's novel view rendering results in Fig. 9 to demonstrate FaceLift's ability to produce high" + } + ] + } + ], + "index": 99 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12697" + } + ] + } + ], + "index": 100 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 84, + 104, + 131 + ], + "blocks": [ + { + "bbox": [ + 71, + 74, + 90, + 81 + ], + "lines": [ + { + "bbox": [ + 71, + 74, + 90, + 81 + ], + "spans": [ + { + "bbox": [ + 71, + 74, + 90, + 81 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 104, + 74, + 198, + 82 + ], + "lines": [ + { + "bbox": [ + 104, + 74, + 198, + 82 + ], + "spans": [ + { + "bbox": [ + 104, + 74, + 198, + 82 + ], + "type": "text", + "content": "w/o. Input View Reconstruction" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 84, + 104, + 131 + ], + "lines": [ + { + "bbox": [ + 58, + 84, + 104, + 131 + ], + "spans": [ + { + "bbox": [ + 58, + 84, + 104, + 131 + ], + "type": "image", + "image_path": "845b35b8900bed5b992a65264b8854aebec4973ae4d4a4e5b4220c0fff50502d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 108, + 93, + 188, + 131 + ], + "blocks": [ + { + "bbox": [ + 108, + 93, + 188, + 131 + ], + "lines": [ + { + "bbox": [ + 108, + 93, + 188, + 131 + ], + "spans": [ + { + "bbox": [ + 108, + 93, + 188, + 131 + ], + "type": "image", + "image_path": "731b4add7905f0ca8004777d6d4fbb9e6dfd22b57b398b42dfd8cabbacbe1541.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 200, + 93, + 281, + 132 + ], + "blocks": [ + { + "bbox": [ + 201, + 74, + 288, + 82 + ], + "lines": [ + { + "bbox": [ + 201, + 74, + 288, + 82 + ], + "spans": [ + { + "bbox": [ + 201, + 74, + 288, + 82 + ], + "type": "text", + "content": "w. Input View Reconstruction" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 200, + 93, + 281, + 132 + ], + "lines": [ + { + "bbox": [ + 200, + 93, + 281, + 132 + ], + "spans": [ + { + "bbox": [ + 200, + 93, + 281, + 132 + ], + "type": "image", + "image_path": "6242894fee37ee8eaf33967c03081fb726dbc2b955e49d5adb5f583366a28a8e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 135, + 295, + 201 + ], + "lines": [ + { + "bbox": [ + 55, + 135, + 295, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 135, + 295, + 201 + ], + "type": "text", + "content": "Figure 10. Importance of input view reconstruction. The diffusion model that is not trained to perform the input view reconstruction, i.e., w/o. Input View Reconstruction, overfits to synthetic training distribution, suffers from severe identity loss during inference. Trained with input view reconstruction, our method preserves the input identity and expression faithfully." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 211, + 296, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 211, + 296, + 368 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 296, + 368 + ], + "type": "text", + "content": "fidelity, realistic 3D head reconstructions with intricate details across a variety of challenging scenarios. FaceLift effectively handles faces under various lighting conditions. It can especially render realistic novel view images given a photo captured with an iPhone under dark lighting conditions (row 1 column 1), emphasizing its robustness and potential for real-world application. It reconstructs facial details with high fidelity, especially the wrinkles and folds on the face caused by extreme expression. FaceLift also excels at reconstructing challenging textures, such as mustaches and hair. Furthermore, it faithfully reconstructs facial paint, despite such data not being included in our synthetic face dataset, showcasing its strong generalization ability." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 375, + 148, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 375, + 148, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 375, + 148, + 388 + ], + "type": "text", + "content": "5. Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 392, + 204, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 204, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 204, + 403 + ], + "type": "text", + "content": "5.1. Input View Reconstruction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 407, + 296, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 296, + 622 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 296, + 622 + ], + "type": "text", + "content": "We conduct an ablation study to demonstrate the importance of reconstructing the input view during training. For comparison, we train a multi-view diffusion model that generates six novel views. In this baseline, the first generated view's elevation is adjusted from " + }, + { + "bbox": [ + 55, + 407, + 296, + 622 + ], + "type": "inline_equation", + "content": "0^{\\circ}" + }, + { + "bbox": [ + 55, + 407, + 296, + 622 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 407, + 296, + 622 + ], + "type": "inline_equation", + "content": "20^{\\circ}" + }, + { + "bbox": [ + 55, + 407, + 296, + 622 + ], + "type": "text", + "content": ", while the remaining views adopt the same camera poses as in our default setting. We refer to this variant as w/o. Input View Reconstruction. Fig. 10 presents the view generation results of the two diffusion models when applied to real-world images. Without the input view reconstruction task, the model trained on the synthetic dataset generates views within a limited distribution, leading to noticeable identity loss. Moreover, it loses its ability to preserve facial expressions and face paint. In contrast, incorporating the input view reconstruction task during training enables our diffusion model to faithfully regenerate the input view, significantly improving its generalization ability. Quantitative comparison is provided in the supplementary material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 627, + 160, + 638 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 627, + 160, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 627, + 160, + 638 + ], + "type": "text", + "content": "5.2. Number of Views" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": "We evaluate three configurations: four views (front, left, right, and back), six views (adding front-left and front-right), and eight views (further including front-top and front-bottom). Fig. 11 compares the baselines using different numbers of input views. With only four views, the reconstructor fails to capture a complete forehead; however," + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 317, + 81, + 351, + 148 + ], + "blocks": [ + { + "bbox": [ + 326, + 74, + 341, + 80 + ], + "lines": [ + { + "bbox": [ + 326, + 74, + 341, + 80 + ], + "spans": [ + { + "bbox": [ + 326, + 74, + 341, + 80 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 81, + 351, + 148 + ], + "lines": [ + { + "bbox": [ + 317, + 81, + 351, + 148 + ], + "spans": [ + { + "bbox": [ + 317, + 81, + 351, + 148 + ], + "type": "image", + "image_path": "1451d24a8979813180aaf4d9c677cf29adfd574559f477841755d545a198f77e.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 152, + 554, + 184 + ], + "lines": [ + { + "bbox": [ + 313, + 152, + 554, + 184 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 554, + 184 + ], + "type": "text", + "content": "Figure 11. Number of input views of Gaussian reconstructor. Using six views strike a good balance between reconstruction quality and computational efficiency." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 353, + 81, + 384, + 148 + ], + "blocks": [ + { + "bbox": [ + 356, + 74, + 377, + 80 + ], + "lines": [ + { + "bbox": [ + 356, + 74, + 377, + 80 + ], + "spans": [ + { + "bbox": [ + 356, + 74, + 377, + 80 + ], + "type": "text", + "content": "4 Views" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 353, + 81, + 384, + 148 + ], + "lines": [ + { + "bbox": [ + 353, + 81, + 384, + 148 + ], + "spans": [ + { + "bbox": [ + 353, + 81, + 384, + 148 + ], + "type": "image", + "image_path": "ebea9c0089956936d4ff1c3860e7348c481971082a3e3a3f3c2cf5ed2e74c72c.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 388, + 81, + 417, + 148 + ], + "blocks": [ + { + "bbox": [ + 389, + 74, + 411, + 80 + ], + "lines": [ + { + "bbox": [ + 389, + 74, + 411, + 80 + ], + "spans": [ + { + "bbox": [ + 389, + 74, + 411, + 80 + ], + "type": "text", + "content": "6 Views" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 388, + 81, + 417, + 148 + ], + "lines": [ + { + "bbox": [ + 388, + 81, + 417, + 148 + ], + "spans": [ + { + "bbox": [ + 388, + 81, + 417, + 148 + ], + "type": "image", + "image_path": "086d8465aef1302223669b97f016f39874388dd9f05c1b3686b8f8464221a73d.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 420, + 81, + 450, + 148 + ], + "blocks": [ + { + "bbox": [ + 422, + 74, + 444, + 80 + ], + "lines": [ + { + "bbox": [ + 422, + 74, + 444, + 80 + ], + "spans": [ + { + "bbox": [ + 422, + 74, + 444, + 80 + ], + "type": "text", + "content": "8 Views" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 420, + 81, + 450, + 148 + ], + "lines": [ + { + "bbox": [ + 420, + 81, + 450, + 148 + ], + "spans": [ + { + "bbox": [ + 420, + 81, + 450, + 148 + ], + "type": "image", + "image_path": "2bcdc2093d6fdacb84b6595a329e3f1d4622a4ef56cbff5e94682bb68a0929b2.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 452, + 81, + 500, + 148 + ], + "blocks": [ + { + "bbox": [ + 456, + 74, + 477, + 80 + ], + "lines": [ + { + "bbox": [ + 456, + 74, + 477, + 80 + ], + "spans": [ + { + "bbox": [ + 456, + 74, + 477, + 80 + ], + "type": "text", + "content": "4 Views" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 489, + 74, + 511, + 80 + ], + "lines": [ + { + "bbox": [ + 489, + 74, + 511, + 80 + ], + "spans": [ + { + "bbox": [ + 489, + 74, + 511, + 80 + ], + "type": "text", + "content": "6 Views" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 452, + 81, + 500, + 148 + ], + "lines": [ + { + "bbox": [ + 452, + 81, + 500, + 148 + ], + "spans": [ + { + "bbox": [ + 452, + 81, + 500, + 148 + ], + "type": "image", + "image_path": "56c6e3265cb6279cc3cbc601089dfc57cab62c215e8b192ed1061af25cd27787.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 502, + 81, + 533, + 148 + ], + "blocks": [ + { + "bbox": [ + 523, + 74, + 544, + 80 + ], + "lines": [ + { + "bbox": [ + 523, + 74, + 544, + 80 + ], + "spans": [ + { + "bbox": [ + 523, + 74, + 544, + 80 + ], + "type": "text", + "content": "8 Views" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 502, + 81, + 533, + 148 + ], + "lines": [ + { + "bbox": [ + 502, + 81, + 533, + 148 + ], + "spans": [ + { + "bbox": [ + 502, + 81, + 533, + 148 + ], + "type": "image", + "image_path": "da9d4abf0b20d79e2738c4cf89d56fb493257d806db2d8e9874b911e4fc72377.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 535, + 81, + 550, + 148 + ], + "blocks": [ + { + "bbox": [ + 535, + 81, + 550, + 148 + ], + "lines": [ + { + "bbox": [ + 535, + 81, + 550, + 148 + ], + "spans": [ + { + "bbox": [ + 535, + 81, + 550, + 148 + ], + "type": "image", + "image_path": "be77612c1ac35f9ebb8413fe72ce03b65ab200b4772b39f31d749c25f2d3ecc8.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 317, + 197, + 372, + 255 + ], + "blocks": [ + { + "bbox": [ + 336, + 187, + 356, + 195 + ], + "lines": [ + { + "bbox": [ + 336, + 187, + 356, + 195 + ], + "spans": [ + { + "bbox": [ + 336, + 187, + 356, + 195 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 197, + 372, + 255 + ], + "lines": [ + { + "bbox": [ + 317, + 197, + 372, + 255 + ], + "spans": [ + { + "bbox": [ + 317, + 197, + 372, + 255 + ], + "type": "image", + "image_path": "565043e59b9872106b383935b4a352ac74896a625eea24ce37e046517f146b6d.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 258, + 554, + 314 + ], + "lines": [ + { + "bbox": [ + 313, + 258, + 554, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 554, + 314 + ], + "type": "text", + "content": "Figure 12. Two-stage training of reconstructor. Without pretraining on general objects, the reconstructor fails to produce clear textures in the reconstruction results. Meanwhile, without finetuning on synthetic human head data, the model lacks a refined understanding of facial structures, including the eyes and nose." + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 376, + 197, + 433, + 255 + ], + "blocks": [ + { + "bbox": [ + 377, + 187, + 430, + 195 + ], + "lines": [ + { + "bbox": [ + 377, + 187, + 430, + 195 + ], + "spans": [ + { + "bbox": [ + 377, + 187, + 430, + 195 + ], + "type": "text", + "content": "w/o. Pre-training" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 376, + 197, + 433, + 255 + ], + "lines": [ + { + "bbox": [ + 376, + 197, + 433, + 255 + ], + "spans": [ + { + "bbox": [ + 376, + 197, + 433, + 255 + ], + "type": "image", + "image_path": "b64979970ac3b38502e399df428f969f3c505664be9d449950e26c9af8a79618.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 435, + 197, + 492, + 255 + ], + "blocks": [ + { + "bbox": [ + 438, + 187, + 488, + 195 + ], + "lines": [ + { + "bbox": [ + 438, + 187, + 488, + 195 + ], + "spans": [ + { + "bbox": [ + 438, + 187, + 488, + 195 + ], + "type": "text", + "content": "w/o. Fine-tuning" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 435, + 197, + 492, + 255 + ], + "lines": [ + { + "bbox": [ + 435, + 197, + 492, + 255 + ], + "spans": [ + { + "bbox": [ + 435, + 197, + 492, + 255 + ], + "type": "image", + "image_path": "8b5535165c67f0ec8061d93797ac56f96a5d564889c666f7ad2a4bab4343df01.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 493, + 197, + 550, + 255 + ], + "blocks": [ + { + "bbox": [ + 512, + 187, + 528, + 194 + ], + "lines": [ + { + "bbox": [ + 512, + 187, + 528, + 194 + ], + "spans": [ + { + "bbox": [ + 512, + 187, + 528, + 194 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 493, + 197, + 550, + 255 + ], + "lines": [ + { + "bbox": [ + 493, + 197, + 550, + 255 + ], + "spans": [ + { + "bbox": [ + 493, + 197, + 550, + 255 + ], + "type": "image", + "image_path": "d1ae0f9ad341904144bd05b4957a41a44ff91b0a7ad3e71407835d7af36011e7.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 323, + 554, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 554, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 554, + 407 + ], + "type": "text", + "content": "with six views, it reconstructs the eyes and eyebrows more smoothly and renders challenging textures—such as facial wrinkles and ear folds—more realistically. Eight views do not offer significant visual improvements, and incur a higher computational cost in both stages. Thus, we find that six views achieve a good balance between reconstruction quality and computational efficiency." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 313, + 411, + 497, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 497, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 497, + 423 + ], + "type": "text", + "content": "5.3. Two-stage Reconstructor Training" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 313, + 426, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 554, + 521 + ], + "type": "text", + "content": "As illustrated in Sec. 3.3, our Gaussian reconstructor follows a two-stage training pipeline. Fig. 12 shows that pretraining on general objects helps the model learn a diverse prior of geometric structures, resulting in clearer textures on delicate facial regions. Meanwhile, fine-tuning on synthetic human head data enables the reconstructor to gain a more refined understanding of facial structure, thereby enhancing the accuracy of features such as the eyes, nose, and hair." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 313, + 529, + 391, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 391, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 391, + 540 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": "We propose FaceLift, a feed-forward approach that lifts a single facial image to a detailed 3D reconstruction with preserved identity features. Our method uses multi-view diffusion to generate unobservable views and employs a transformer-based reconstructor to reconstruct 3D Gaussian splats, enabling high-quality novel view synthesis. To overcome the difficulty of capturing real-world multi-view human head images, we render high-quality synthetic data for training and show that, despite being trained solely on synthetic data, FaceLift can reconstruct 3D heads from real-world captured images with high fidelity. Compared with baselines [1, 4, 29, 31, 56], FaceLift generates 3D head representation with finer geometry and texture details and exhibits better identity preservation ability." + } + ] + } + ], + "index": 41 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12698" + } + ] + } + ], + "index": 42 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "text", + "content": "7. Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 139 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 139 + ], + "type": "text", + "content": "This was supported in part by the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korean Government (MSIT) (No. RS-2024-00457882, National AI Research Lab Project)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 140, + 296, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 140, + 296, + 200 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 296, + 200 + ], + "type": "text", + "content": "We appreciate the insightful discussions with Kai Zhang, Hao Tan, Zexiang Xu, Sai Bi, Sumit Chaturvedi, Hanwen Jiang, Yu-Ju Tsai, Kuan-Chih Huang, Chengxu Liu, and Dingyi Dai. We thank Nathan Carr and Kalyan Sunkavalli for their support." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 210, + 115, + 223 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 210, + 115, + 223 + ], + "spans": [ + { + "bbox": [ + 56, + 210, + 115, + 223 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 229, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 229, + 296, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 229, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 61, + 229, + 296, + 262 + ], + "type": "text", + "content": "[1] Sizhe An, Hongyi Xu, Yichun Shi, Guoxian Song, Umit Y Ogras, and Linjie Luo. PanoHead: Geometry-aware 3D fullhead synthesis in " + }, + { + "bbox": [ + 61, + 229, + 296, + 262 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 61, + 229, + 296, + 262 + ], + "type": "text", + "content": ". In CVPR, 2023. 1, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 264, + 296, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 264, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 62, + 264, + 296, + 297 + ], + "type": "text", + "content": "[2] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 298, + 295, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 298, + 295, + 330 + ], + "spans": [ + { + "bbox": [ + 62, + 298, + 295, + 330 + ], + "type": "text", + "content": "[3] Ananta R. Bhattarai, Matthias Nießner, and Artem Sevastopolsky. TriPlaneNet: An encoder for EG3D inversion. In WACV, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 332, + 295, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 332, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 62, + 332, + 295, + 376 + ], + "type": "text", + "content": "[4] Bahri Batuhan Bilecen, Ahmet Berke Gokmen, and Aysegul Dundar. Dual encoder GAN inversion for high-fidelity 3D head reconstruction from single images. In NeurIPS, 2024. 5, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 377, + 295, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 377, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 62, + 377, + 295, + 399 + ], + "type": "text", + "content": "[5] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3D faces. In SIGGRAPH, 1999. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 400, + 295, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 400, + 295, + 432 + ], + "spans": [ + { + "bbox": [ + 62, + 400, + 295, + 432 + ], + "type": "text", + "content": "[6] James Booth, Anastasios Roussos, Stefanos Zafeiriou, Allan Ponniah, and David Dunaway. A 3D morphable model learnt from 10,000 faces. In CVPR, 2016. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 434, + 296, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 434, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 62, + 434, + 296, + 498 + ], + "type": "text", + "content": "[7] Marcel C Bühler, Kripasindhu Sarkar, Tanmay Shah, Gengyan Li, Daoye Wang, Leonhard Helminger, Sergio Orts-Escalano, Dmitry Lagun, Otmar Hilliges, Thabo Beeler, et al. Preface: A data-driven volumetric prior for few-shot ultra high-resolution face synthesis. In ICCV, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 501, + 296, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 501, + 296, + 566 + ], + "spans": [ + { + "bbox": [ + 62, + 501, + 296, + 566 + ], + "type": "text", + "content": "[8] Marcel C Bühler, Gengyan Li, Erroll Wood, Leonhard Helminger, Xu Chen, Tanmay Shah, Daoye Wang, Stephan Garbin, Sergio Orts-Escolano, Otmar Hilliges, et al. Cafca: High-quality novel view synthesis of expressive faces from casual few-shot captures. In SIGGRAPH Asia, 2024. 2, 3, 5, 12, 15, 17" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 567, + 295, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 567, + 295, + 611 + ], + "spans": [ + { + "bbox": [ + 62, + 567, + 295, + 611 + ], + "type": "text", + "content": "[9] Sumit Chaturvedi, Mengwei Ren, Yannick Hold-Geoffroy, Jingyuan Liu, Julie Dorsey, and Zhixin Shu. SynthLight: Portrait relighting with diffusion model by learning to re-render synthetic faces. In CVPR, 2025. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 613, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 295, + 656 + ], + "type": "text", + "content": "[10] Xiyi Chen, Marko Mihajlovic, Shaofei Wang, Sergey Prokudin, and Siyu Tang. Morphable diffusion: 3d-consistent diffusion for single-image avatar creation. In CVPR, 2024. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 658, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 689 + ], + "type": "text", + "content": "[11] Antonia Creswell and Anil Anthony Bharath. Inverting the generator of a generative adversarial network. In TNNLS, 2018. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "text", + "content": "[12] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3D objects. In CVPR, 2023. 2, 3, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 97, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 140 + ], + "type": "text", + "content": "[13] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-XL: A universe of " + }, + { + "bbox": [ + 316, + 97, + 553, + 140 + ], + "type": "inline_equation", + "content": "10\\mathrm{M} + 3\\mathrm{D}" + }, + { + "bbox": [ + 316, + 97, + 553, + 140 + ], + "type": "text", + "content": " objects. In NeurIPS, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 142, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 142, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 142, + 553, + 185 + ], + "type": "text", + "content": "[14] Jiankang Deng, Jia Guo, Jing Yang, Niannan Xue, Irene Kotsia, and Stefanos Zafeiriou. ArcFace: Additive angular margin loss for deep face recognition. In IEEE TPAMI, 2022. 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 187, + 553, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 187, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 316, + 187, + 553, + 209 + ], + "type": "text", + "content": "[15] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 211, + 553, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 211, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 317, + 211, + 553, + 255 + ], + "type": "text", + "content": "[16] Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. DreamSim: Learning new dimensions of human visual similarity using synthetic data. In NeurIPS, 2023. 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 256, + 553, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 256, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 316, + 256, + 553, + 289 + ], + "type": "text", + "content": "[17] Chen Gao, Yichang Shih, Wei-Sheng Lai, Chia-Kai Liang, and Jia-Bin Huang. Portrait neural radiance fields from a single image. arXiv preprint arXiv:2012.05903, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 291, + 553, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 291, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 316, + 291, + 553, + 335 + ], + "type": "text", + "content": "[18] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. GET3D: A generative model of high quality 3D textured shapes learned from images. In NeurIPS, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 336, + 553, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 336, + 553, + 379 + ], + "spans": [ + { + "bbox": [ + 316, + 336, + 553, + 379 + ], + "type": "text", + "content": "[19] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 1, 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 381, + 553, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 381, + 553, + 413 + ], + "spans": [ + { + "bbox": [ + 316, + 381, + 553, + 413 + ], + "type": "text", + "content": "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 415, + 553, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 449 + ], + "type": "text", + "content": "[21] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 16" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 450, + 553, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 450, + 553, + 472 + ], + "spans": [ + { + "bbox": [ + 316, + 450, + 553, + 472 + ], + "type": "text", + "content": "[22] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 15" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 474, + 553, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 474, + 553, + 496 + ], + "spans": [ + { + "bbox": [ + 316, + 474, + 553, + 496 + ], + "type": "text", + "content": "[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 498, + 553, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 553, + 540 + ], + "type": "text", + "content": "[24] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: Large reconstruction model for single image to 3D. In ICLR, 2024. 3, 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 543, + 553, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 543, + 553, + 575 + ], + "spans": [ + { + "bbox": [ + 316, + 543, + 553, + 575 + ], + "type": "text", + "content": "[25] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 1, 2, 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 577, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 577, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 316, + 577, + 553, + 610 + ], + "type": "text", + "content": "[26] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In CVPR, 2020. 2, 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 612, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 612, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 612, + 553, + 644 + ], + "type": "text", + "content": "[27] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D gaussian splatting for real-time radiance field rendering. In ACM TOG, 2023. 1, 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 646, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 646, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 646, + 553, + 668 + ], + "type": "text", + "content": "[28] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In ICLR, 2014. 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[29] Tobias Kirschstein, Simon Giebenhain, Jiapeng Tang, Markos Georgopoulos, and Matthias Nießner. GGHead: Fast and generalizable 3D gaussian heads. In SIGGRAPH Asia, 2024. 5, 6, 8" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12699" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[30] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3D: Fast text-to-3D with sparse-view generation and large reconstruction model. In ICLR, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 294, + 183 + ], + "type": "text", + "content": "[31] Peng Li, Yuan Liu, Xiaoxiao Long, Feihu Zhang, Cheng Lin, Mengfei Li, Xingqun Qi, Shanghang Zhang, Wenhan Luo, Ping Tan, et al. Era3D: High-resolution multiview diffusion using efficient row-wise attention. In NeurIPS, 2024. 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 184, + 294, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 294, + 216 + ], + "type": "text", + "content": "[32] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. In ACM TOG, 2017. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 249 + ], + "type": "text", + "content": "[33] Xueting Li, Shalini De Mello, Sifei Liu, Koki Nagano, Umar Iqbal, and Jan Kautz. Generalizable one-shot 3D neural head avatar. In NeurIPS, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 250, + 294, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 283 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 283 + ], + "type": "text", + "content": "[34] Jiangke Lin, Yi Yuan, Tianjia Shao, and Kun Zhou. Towards high-fidelity 3D face reconstruction from in-the-wild images using graph convolutional networks. In CVPR, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 284, + 294, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 284, + 294, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 284, + 294, + 327 + ], + "type": "text", + "content": "[35] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T., Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3D mesh in 45 seconds without pershape optimization. In NeurIPS, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 328, + 294, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 328, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 56, + 328, + 294, + 361 + ], + "type": "text", + "content": "[36] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In ICCV, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 361, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 294, + 415 + ], + "type": "text", + "content": "[37] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3D: Single image to 3D using cross-domain diffusion. In CVPR, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 417, + 294, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 417, + 294, + 471 + ], + "spans": [ + { + "bbox": [ + 56, + 417, + 294, + 471 + ], + "type": "text", + "content": "[38] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Guang Yong, Juhyun Lee, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, 2019. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 472, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 472, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 472, + 295, + 712 + ], + "type": "text", + "content": "[39] Julieta Martinez, Emily Kim, Javier Romero, Timur Bagautdinov, Shunsuke Saito, Shouou-I Yu, Stuart Anderson, Michael Zollhöfer, Te-Li Wang, Shaojie Bai, Chenghui Li, Shih-En Wei, Rohan Joshi, Wyatt Borsos, Tomas Simon, Jason Saragih, Paul Theodosis, Alexander Greene, Anjani Josyula, Silvio Mano Maeta, Andrew I. Jewett, Simon Venshtain, Christopher Heilman, Yueh-Tung Chen, Sidi Fu, Mohamed Ezzeldin A. Elshaer, Tingfang Du, Longhua Wu, Shen-Chi Chen, Kai Kang, Michael Wu, Youssef Emad, Steven Longay, Ashley Brewer, Hitesh Shah, James Booth, Taylor Koska, Kayla Hidle, Matt Andromalos, Joanna Hsu, Thomas Dauer, Peter Selednik, Tim Godisart, Scott Ardisson, Matthew Cipperly, Ben Humberston, Lon Farr, Bob Hansen, Peihong Guo, Dave Braun, Steven Krenn, He Wen, Lucas Evans, Natalia Fadeeva, Matthew Stewart, Gabriel Schwartz, Divam Gupta, Gyeongsik Moon, Kaiwen Guo, Yuan Dong, Yichen Xu, Takaaki Shiratori, Fabian Prada, Bernardo R. Pires, Bo Peng, Julia Buffalini, Autumn Trimble, Kevyn McPhail, Melissa Schoeller, and Yaser Sheikh. Codec Avatar Studio: Paired human captures for complete, driveable, and generalizable avatars. In NeurlPS, 2024. 2, 3, 5, 15, 17" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 116 + ], + "type": "text", + "content": "[40] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 118, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 118, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 315, + 118, + 553, + 162 + ], + "type": "text", + "content": "[41] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. StyleSDF: High-resolution 3D-consistent image and geometry generation. In CVPR, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 163, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 163, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 315, + 163, + 553, + 205 + ], + "type": "text", + "content": "[42] Foivos Paraperas Papantoniou, Alexandros Lattas, Stylianos Moschoglou, and Stefanos Zafeiriou. Relightify: Relightable 3D faces from a single image via diffusion models. In ICCV, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 207, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 207, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 315, + 207, + 553, + 239 + ], + "type": "text", + "content": "[43] Dario Pavllo, Graham Spinks, Thomas Hofmann, Marie-Francine Moens, and Aurélien Lucchi. Convolutional generation of textured 3D meshes. In NeurIPS, 2020. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 241, + 553, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 241, + 553, + 273 + ], + "spans": [ + { + "bbox": [ + 315, + 241, + 553, + 273 + ], + "type": "text", + "content": "[44] Julius Plucker. XVii. on a new geometry of space. Philosophical Transactions of the Royal Society of London, 1865. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 275, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 275, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 315, + 275, + 553, + 307 + ], + "type": "text", + "content": "[45] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D diffusion. In ICLR, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 309, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 309, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 315, + 309, + 553, + 364 + ], + "type": "text", + "content": "[46] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3D object generation using both 2D and 3D diffusion priors. In ICLR, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 365, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 365, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 315, + 365, + 553, + 397 + ], + "type": "text", + "content": "[47] Elad Richardson, Matan Sela, and Ron Kimmel. 3D face reconstruction by learning from synthetic data. In 3DV, 2016. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 399, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 399, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 315, + 399, + 553, + 431 + ], + "type": "text", + "content": "[48] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 433, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 433, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 315, + 433, + 553, + 465 + ], + "type": "text", + "content": "[49] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 16" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 467, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 467, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 315, + 467, + 553, + 498 + ], + "type": "text", + "content": "[50] Shunsuke Saito, Gabriel Schwartz, Tomas Simon, Junxuan Li, and Giljoo Nam. Relightable gaussian codec avatars. In CVPR, 2024. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 500, + 553, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 500, + 553, + 554 + ], + "spans": [ + { + "bbox": [ + 315, + 500, + 553, + 554 + ], + "type": "text", + "content": "[51] Kyle Sargent, Zizhang Li, Tanmay Shah, Charles Herrmann, Hong-Xing Yu, Yunzhi Zhang, Eric Ryan Chan, Dmitry Lagun, Li Fei-Fei, Deqing Sun, et al. ZeroNVS: Zero-shot 360-degree view synthesis from a single real image. In CVPR, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 556, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 556, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 315, + 556, + 553, + 589 + ], + "type": "text", + "content": "[52] Sefik Ilkin Serengil and Alper Ozpinar. HyperExtended LightFace: A facial attribute analysis framework. In ICEET, 2021. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 590, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 590, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 315, + 590, + 553, + 644 + ], + "type": "text", + "content": "[53] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: A single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 647, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 647, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 315, + 647, + 553, + 667 + ], + "type": "text", + "content": "[54] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 3, 16" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 669, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 669, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 315, + 669, + 553, + 712 + ], + "type": "text", + "content": "[55] Wenqiang Sun, Shuo Chen, Fangfu Liu, Zilong Chen, Yueqi Duan, Jun Zhang, and Yikai Wang. DimensionX: Create any 3D and 4D scenes from a single image with controllable video diffusion. arXiv preprint arXiv:2411.04928, 2024. 12" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12700" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[56] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. LGM: Large multi-view gaussian model for high-resolution 3D content creation. In ECCV, 2024. 3, 5, 6, 7, 8, 17, 18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "text", + "content": "[57] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. DreamGaussian: Generative gaussian splatting for efficient 3D content creation. In ICLR, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "type": "text", + "content": "[58] Aäron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning. In NeurIPS, 2017. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 228 + ], + "type": "text", + "content": "[59] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 3, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 230, + 294, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 263 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 263 + ], + "type": "text", + "content": "[60] Thomas Vetter and Volker Blanz. Estimating coloured 3D face models from single images: An example based approach. In ECCV, 1998. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 265, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 265, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 265, + 294, + 297 + ], + "type": "text", + "content": "[61] Vishal Vinod, Tanmay Shah, and Dmitry Lagun. TEGLO: High fidelity canonical texture mapping from single-view images. In WACV, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 299, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 294, + 342 + ], + "type": "text", + "content": "[62] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In NeurIPS, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 343, + 294, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 294, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 294, + 397 + ], + "type": "text", + "content": "[63] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3D digital avatars using diffusion. In CVPR, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 399, + 294, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 443 + ], + "type": "text", + "content": "[64] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. ProlificDreamer: High-fidelity and diverse text-to-3D generation with variational score distillation. In NeurIPS, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 445, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 487 + ], + "type": "text", + "content": "[65] Erroll Wood, Tadas Baltrusaitis, Charlie Hewitt, Sebastian Dziadzio, Thomas J Cashman, and Jamie Shotton. Fake it till you make it: Face analysis in the wild using synthetic data alone. In ICCV, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 489, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 294, + 533 + ], + "type": "text", + "content": "[66] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3D generative-adversarial modeling. In NeurIPS, 2016. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 535, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 294, + 578 + ], + "type": "text", + "content": "[67] Kailu Wu, Fangfu Liu, Zhihan Cai, Runjie Yan, Hanyang Wang, Yating Hu, Yueqi Duan, and Kaiheng Ma. Unique3d: High-quality and efficient 3d mesh generation from a single image. arXiv preprint arXiv:2405.20343, 2024. 12, 13" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 580, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 294, + 612 + ], + "type": "text", + "content": "[68] Yiqian Wu, Jing Zhang, Hongbo Fu, and Xiaogang Jin. LPFF: A portrait dataset for face generators across large poses. In ICCV, 2023. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 614, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 656 + ], + "type": "text", + "content": "[69] Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin Tong, and Jiaolong Yang. Structured 3d latents for scalable and versatile 3d generation. In CVPR, 2024. 12, 13" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[70] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024. 12, 13" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 553, + 341 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 315, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 553, + 116 + ], + "type": "text", + "content": "[71] Yu-Ying Yeh, Koki Nagano, Sameh Khamis, Jan Kautz, Ming-Yu Liu, and Ting-Chun Wang. Learning to relight portrait images via a virtual light stage and synthetic-to-real adaptation. In ACM TOG, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 315, + 118, + 553, + 161 + ], + "type": "text", + "content": "[72] Ziyang Yuan, Yiming Zhu, Yu Li, Hongyu Liu, and Chun Yuan. Make encoder great again in 3D GAN inversion through geometry and occlusion-aware encoding. In ICCV, 2023. 1, 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 163, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 163, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 315, + 163, + 553, + 206 + ], + "type": "text", + "content": "[73] Bowen Zhang, Yiji Cheng, Chunyu Wang, Ting Zhang, Jiaolong Yang, Yansong Tang, Feng Zhao, Dong Chen, and Baining Guo. RodinHD: High-fidelity 3D avatar generation with diffusion models. In ECCV, 2024. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 208, + 553, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 208, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 315, + 208, + 553, + 250 + ], + "type": "text", + "content": "[74] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. GS-LRM: Large reconstruction model for 3D gaussian splatting. In ECCV, 2024. 3, 5, 17" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 252, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 252, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 315, + 252, + 553, + 285 + ], + "type": "text", + "content": "[75] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 286, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 286, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 315, + 286, + 553, + 341 + ], + "type": "text", + "content": "[76] Xiaozheng Zheng, Chao Wen, Zhaohu Li, Weiyi Zhang, Zhuo Su, Xu Chang, Yang Zhao, Zheng Lv, Xiaoyuan Zhang, Yongjie Zhang, Guidong Wang, and Xu Lan. Headgap: Few-shot 3d head avatar via generalizable gaussian priors. In 3DV, 2025. 3" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12701" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/FaceShield_ Defending Facial Image against Deepfake Threats/4436b2e6-eb11-4750-8df1-5fa6df4f36fe_content_list.json b/2025/FaceShield_ Defending Facial Image against Deepfake Threats/4436b2e6-eb11-4750-8df1-5fa6df4f36fe_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..eb17e557eb3f2e81a30a8df7942370b0aa27bb44 --- /dev/null +++ b/2025/FaceShield_ Defending Facial Image against Deepfake Threats/4436b2e6-eb11-4750-8df1-5fa6df4f36fe_content_list.json @@ -0,0 +1,1722 @@ +[ + { + "type": "text", + "text": "FaceShield: Defending Facial Image against Deepfake Threats", + "text_level": 1, + "bbox": [ + 124, + 107, + 875, + 133 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jaehwan Jeong $^{1}$ Sumin In $^{1}$ Sieun Kim $^{1}$ Hannie Shin $^{1}$ Jongheon Jeong $^{1}$", + "bbox": [ + 178, + 154, + 816, + 172 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sang Ho Yoon² Jaewook Chung³ Sangpil Kim¹*", + "bbox": [ + 290, + 172, + 712, + 191 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Korea University $^{2}$ KAIST $^{3}$ Samsung Research", + "bbox": [ + 290, + 207, + 707, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 272, + 326, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The rising use of deepfakes in criminal activities presents a significant issue, inciting widespread controversy. While numerous studies have tackled this problem, most primarily focus on deepfake detection. These reactive solutions are insufficient as a fundamental approach for crimes where authenticity is disregarded. Existing proactive defenses also have limitations, as they are effective only for deepfake models based on specific Generative Adversarial Networks (GANs), making them less applicable in light of recent advancements in diffusion-based models. In this paper, we propose a proactive defense method named FaceShield, which introduces novel defense strategies targeting deepfakes generated by Diffusion Models (DMs) and facilitates defenses on various existing GAN-based deepfake models through facial feature extractor manipulations. Our approach consists of three main components: (i) manipulating the attention mechanism of DMs to exclude protected facial features during the denoising process, (ii) targeting prominent facial feature extraction models to enhance the robustness of our adversarial perturbation, and (iii) employing Gaussian blur and low-pass filtering techniques to improve imperceptibility while enhancing robustness against JPEG compression. Experimental results on the CelebAHQ and VGGFace2-HQ datasets demonstrate that our method achieves state-of-the-art performance against the latest deepfake models based on DMs, while also exhibiting transferability to GANs and showcasing greater imperceptibility of noise along with enhanced robustness.", + "bbox": [ + 88, + 304, + 485, + 728 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 744, + 220, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The advancement of deepfake technology and improved accessibility [3, 22, 25, 37, 51] has led to significant transformations in modern society. Due to the ease of face swapping, it has been applied across various fields, providing both entertainment and convenience. However, its powerful capability to generate realistic content has also enabled malicious users to exploit it for criminal purposes, leading", + "bbox": [ + 89, + 770, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b3f7f53c35add7f73176b70fd21a5606436e5af331ccf4421a328952c3d51214.jpg", + "image_caption": [ + "Figure 1. Protecting Face during Deepfake using FaceShield. Pure images are vulnerable to face swapping, allowing the target image's face to be easily reflected. In contrast, images protected by FaceShield conceal facial feature from deepfake." + ], + "image_footnote": [], + "bbox": [ + 535, + 282, + 882, + 431 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to the creation of fake news and various societal problems.", + "bbox": [ + 511, + 507, + 901, + 522 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address the growing concerns surrounding deepfake technology, various countermeasures have been explored, which can be broadly divided into two categories. The first is deepfake detection techniques [5, 12, 13, 31, 36], which act as passive defenses by classifying whether content is synthetic or authentic. While effective for authenticity verification, these offer only binary results and fail to address advanced threats, such as crimes using realistic fakes. In contrast, proactive defense strategies offer a more comprehensive solution. These approaches involve embedding imperceptible adversarial perturbation into face images to prevent the protected face from being effectively processed by deepfakes. However, most previous research [8, 14, 28, 35, 44] has concentrated on GAN-based models, often targeting individual models, which limits effectiveness against emerging DM-based deepfakes [20, 34, 43, 49]. Although significant research exists on image protection within DMs [6, 23, 24, 29, 30, 42] for image editing, the focus has primarily been on attacking the noising and denoising processes when an image is used as a query (Fig.2a). This leads to targeting the encoder or predicted noise post-UNet processing. However, we observe that such strategies are ineffective for DM-based deepfake models, where the source image influences the output in the form of key-value pairs through attention mechanisms (Fig.2b).", + "bbox": [ + 509, + 523, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 114, + 887, + 240, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10364", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9517ecb6a28033c29d026d26da5158d182f422be79ae3af76ac8a2a60250e3ae.jpg", + "image_caption": [ + "Figure 2. Image editing and Deepfake processes in DMs. (a) In DM-based image editing, a single image is input as a query $Q$ and edited based on a prompt condition. (b) In DM-based deepfake, two images are used, with the target image serving as the query $Q$ while the source image acts as the condition for swapping. This condition operates as key $K$ and value $V$ in the cross-attention layer." + ], + "image_footnote": [], + "bbox": [ + 96, + 90, + 465, + 205 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/708d60f488b3aaa8cced6e3abc8bf0b62b3b3a4250c2b2057081f7412204dccf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 89, + 903, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we focus on attacking state-of-the-art DM-based deepfakes while ensuring applicability to GAN-based models as well (Fig.1). Given the uncertainty surrounding deepfakes that malicious users might employ, we explore approaches to improve the extensibility across different architectures (e.g., GANs, DMs) and model transferability across different pre-trained backbones. Simultaneously, we propose a novel noise update method that enhances imperceptibility while being robust to JPEG compression.", + "bbox": [ + 88, + 276, + 480, + 412 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For DM-based deepfake attacks, we leverage the structural properties in which the conditioning image is embedded and integrated into the denoising UNet through attention mechanisms. By utilizing the IP-Adapter [43], commonly employed for inpainting, we extract effective adversarial noise from the embedding of the conditioning image. This perturbation effectively disrupts the propagation of the conditioning process, ensuring that the final output does not replicate the features of the protected image.", + "bbox": [ + 88, + 414, + 482, + 549 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To enhance the generalizability of our approach, we target two commonly used facial feature extractors. First, we attack the MTCNN model [48], which uses a cascade pyramid architecture to achieve superior performance and robust detection capabilities. Due to this, it is widely adopted not only in deepfake generation but also across various applications. We leverage the fact that the model scales images to different sizes during face detection. Our perturbation is designed to ensure robustness across various scaling factors and interpolation modes (e.g., BILINEAR, AREA), leading to superior performance compared to existing methods [19, 46]. Additionally, we target ArcFace [7], a widely adopted pre-trained model for facial feature extraction in deepfake applications. By incorporating both methods into our work, we ensure that our approach disrupts a range of deepfakes commonly used for facial landmark detection and feature extraction, thereby improving the overall robustness of our method against various deepfake systems.", + "bbox": [ + 89, + 551, + 482, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the noise updating process, we refine the perturbation using two techniques: Noise Blur, which measures differences between adjacent pixels for imperceptible refinement, and Low-pass filtering, retaining low-frequency components, enhancing robustness against JPEG compression.", + "bbox": [ + 89, + 825, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our main contributions are as follows:", + "bbox": [ + 529, + 277, + 883, + 291 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a novel attack on deepfakes based on diffusion models. To the best of our knowledge, our proposed method is the first attempt to protect images used as conditions while demonstrating robust performance across various deepfake models by targeting common facial feature extractors.", + "- We propose a novel noise update mechanism that integrates Gaussian blur technique with the projected gradient descent method, significantly enhancing imperceptibility. Additionally, we implement low-pass filtering to reduce perturbation loss rates during JPEG compression compared to existing methods.", + "- We demonstrate that our deepfake attack method is robust across various deepfake models, outperforming previous diffusion attacks by achieving higher distortion with significantly less noise." + ], + "bbox": [ + 511, + 292, + 903, + 534 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 550, + 653, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deepfake techniques. With advancements in generative models, deepfake technology has evolved into a specialized field focused on facial synthesis. Previous deepfake models, primarily based on GANs, generally follow a three-stage process: face detection and localization, feature extraction, and face swapping. Among these, studies such as [11, 40, 41, 45, 51] employ MTCNN [48] for face detection and landmark extraction, while the majority of deepfake models, including [3, 21, 22, 40, 51, 52], leverage ArcFace [7] for identity feature extraction. These steps are similarly employed in DM-based deepfakes that have emerged with the progress of diffusion models. Notable examples, including [20, 34, 49], integrate [7] to maintain identity consistency. However, recent work has focused on leveraging the capabilities of diffusion models to develop face-swapping methods [43] that achieve high performance without explicitly following previous approaches.", + "bbox": [ + 509, + 579, + 906, + 837 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Enhancing model transferability. In the research on adversarial attacks, various attempts have been made to improve transferability. [9, 38] proposed the model ensemble technique, generating adversarial examples using multiple", + "bbox": [ + 511, + 839, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10365", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "models to enhance their effectiveness on unseen models. [39] introduced a method that selectively utilizes specific layers within a model to improve transferability. Similarly, [15, 16, 50] investigated techniques that manipulate intermediate layer feature distributions or amplify activation values to prevent adversarial noise from overfitting to a particular model. Furthermore, [2, 47] explored the use of multiple pre-trained backbones within similar model architectures to enhance transferability across different backbone networks.", + "bbox": [ + 89, + 90, + 480, + 226 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 239, + 181, + 255 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose a novel pipeline, FaceShield, to safeguard facial images from being exploited by diverse Deepfake methods through conditional attacks on DMs and facial feature extractor attacks. In this section, we first introduce the foundational adversarial attack framework utilized across our approach (Sec.3.1). We then detail our method for disrupting information flow when a facial image is employed as a conditioning input in DMs (Sec.3.2). Subsequently, we present our approach for preventing accurate facial feature extraction (Sec.3.3). Finally, we introduce our adversarial noise update mechanism, designed to enhance imperceptibility and mitigate degradation from JPEG compression (Sec.3.4).", + "bbox": [ + 89, + 263, + 483, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 453, + 230, + 468 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Cross-attention mechanism. To condition generative DMs, the cross-attention mechanism is used, as shown in Fig.2. Similar to self-attention, it involves computations using the query $Q$ , key $K$ , and value $V$ . However, unlike self-attention, where $Q$ , $K$ and $V$ are derived from the same source, cross-attention conditions the process by obtaining $Q$ from the noised image $z_{t}$ through a learned linear projection $\\ell_{q}$ , while $K$ and $V$ are derived from the textual or image embedding $C_{\\mathrm{emb}}$ using learned linear projection $\\ell_{k}$ and $\\ell_{v}$ , respectively:", + "bbox": [ + 89, + 479, + 483, + 630 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nQ = \\ell_ {q} (z _ {t}), K = \\ell_ {k} \\left(C _ {\\text {e m b}}\\right), V = \\ell_ {v} \\left(C _ {\\text {e m b}}\\right), \\text {a n d} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 638, + 482, + 655 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}}\\right) V, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 657, + 482, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $d_{k}$ are the dimensions of the key vectors.", + "bbox": [ + 89, + 699, + 406, + 713 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Projected gradient descent (PGD). PGD is a widely used method for crafting adversarial examples when the user has access to the model parameters. This technique iteratively updates an adversarial perturbation by computing the gradient of a certain loss $\\mathcal{L}_{\\mathrm{adv}}$ with respect to the input. At each step, noise is added in the gradient direction while keeping the perturbation within a predefined bound, ensuring the noise is small but effective:", + "bbox": [ + 89, + 715, + 483, + 835 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\delta \\leftarrow \\operatorname {P r o j} _ {\\| \\delta \\| _ {\\infty} \\leq \\eta} \\left(\\delta - \\alpha \\cdot \\operatorname {s i g n} \\left(\\nabla_ {\\delta} \\mathcal {L} _ {\\mathrm {a d v}}\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 845, + 482, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha$ is the step size and $\\mathrm{Proj}_{\\| \\delta \\|_{\\infty} \\leq \\eta}(\\cdot)$ projects $\\delta$ onto the $\\ell_{\\infty}$ ball of radius $\\eta$ . By projecting the adversarial exam", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ple back onto the valid perturbation space, PGD maintains imperceptibility while disrupting the model's predictions.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Conditioned Face Attack", + "text_level": 1, + "bbox": [ + 511, + 132, + 741, + 148 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We now describe our approach to protecting images, specifically by disrupting the effective transfer of information when they are used as conditioning inputs in DMs. The core of our approach is to effectively interfere with key information using minimal noise, while also ensuring that the model does not overfit by accessing only a minimal number of layers to obtain gradients. To achieve this, we propose two methods that target both the initial projection phase and the final attention mechanism during the image conditioning process within latent diffusion models [27].", + "bbox": [ + 511, + 155, + 906, + 306 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Face projector attack. When images are used as conditioning inputs, they are firstly transformed into an embedding vector through a pre-trained model [26]. In this method, we access only the topmost layer $\\mathcal{P}$ of the model to disrupt the projection process, causing the image to be projected with incorrect information at the initial stage. For the attack loss function, we consider that converging to a single target value might not ensure consistent convergence speeds or balanced performance. Given that one of our main goals is to design noise applicable to various images, we design our approach to induce random divergence based on the input image, using the $\\mathcal{L}_1$ loss function in this process:", + "bbox": [ + 511, + 309, + 908, + 492 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {p r o j}} (\\delta ; x) = \\| \\mathcal {P} (x + \\delta) - \\mathcal {P} (x) \\| _ {1}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 503, + 906, + 522 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\delta$ is the adversarial noise.", + "bbox": [ + 511, + 534, + 725, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Attention disruption attack. We also focus on identifying the core vectors within the denoising UNet that are most sensitive to conditional inputs. Initially, we analyze the influence of cross-attention across each UNet layer. Based on prior research [33], which shows that different cross-attention layers respond variably to conditioning information, we investigate the impact on perturbation performance for each region. Our findings lead to the conclusion that targeting attacks near mid-layers produces more significant disruption in qualitative metrics compared to using only the up-down layers or the entire layers, as supported by our experimental results in Fig.6. Based on these insights, we propose a novel approach that specifically targets mid-layers during the attack on the diffusion process.", + "bbox": [ + 511, + 551, + 906, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To induce a mismatch in conditioning, we use the mid-layer cross-attention mechanism, as described in Eq. (1). Based on the idea that the condition is conveyed to the query through attention, we calculate the attention score to obtain the strength of attention. This is done by performing operations on the query $Q \\in \\mathbb{R}^{h \\times \\mathrm{res} \\times d}$ and key $K \\in \\mathbb{R}^{h \\times \\mathrm{seq} \\times d}$ , where $h$ (number of heads), res (resolution), seq (sequence length), and $d$ (head dimension). This is followed by a Softmax operation along the seq dimension to derive the", + "bbox": [ + 511, + 765, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10366", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3af093bf8674b87529653d3c016f5a47231ed3cfc38cb50e5b220b904a743cdd.jpg", + "image_caption": [ + "Figure 3. Overview. Our method has three main parts: (i) Conditioned face attack, which disrupts feature transfer by targeting the embedding process and the attention map variance in the cross-attention layer; (ii) Facial feature extractor attack, which decreases the probability value of face detection and causes extraction disruptions, and (iii) Enhanced noise update, which improves imperceptibility by applying Gaussian blur to regions with significant intensity changes between adjacent pixels, and increases robustness against JPEG compression distortion by encoding the noise in the low-frequency domain." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 447, + 375 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b93492d6df6d9a48bea2beacb5a6051490eeb4910fba6ddf91d9df5710568b85.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 450, + 90, + 898, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "attention map $A_{\\mathrm{map}} \\in \\mathbb{R}^{h \\times \\mathrm{res} \\times \\mathrm{seq}}$ . Exploiting this mechanism, we obtain the variance $A_{\\mathrm{var}}$ , allowing us to evaluate attention strength by the following equation:", + "bbox": [ + 89, + 469, + 482, + 516 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nA _ {\\text {v a r}} = \\frac {1}{\\text {s e q}} \\sum_ {i = 1} ^ {\\text {s e q}} \\left(A _ {\\text {m a p}} [:, i,: ] - \\bar {A} _ {\\text {m a p}}\\right) ^ {2} \\in \\mathbb {R} ^ {h \\times \\text {r e s}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 526, + 482, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\bar{A}_{\\mathrm{map}} = \\frac{1}{\\mathrm{seq}}\\sum_{i' = 1}^{\\mathrm{seq}}A_{\\mathrm{map}}[:,i',:]$ is the mean of attention map across the seq dimension.", + "bbox": [ + 89, + 579, + 482, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on them, we propose an adversarial attack strategy that maximizes $A_{\\mathrm{var}}$ , thereby preventing the proper reflection of conditional information $K$ on $Q$ . In this process, we encode the original image $x$ to use as the query $Q$ and project the same $x$ to obtain the key $K$ , which is then used to calculate $A_{\\mathrm{var}}$ . Thereafter, we find a quantile $P_{t_{\\mathrm{var}}}$ corresponding to a predefined threshold $t_{\\mathrm{var}}$ between 0 and 1 to identify the regions exhibiting weak attention. Using this $P_{t_{\\mathrm{var}}}$ , we create a mask $M_{\\mathrm{var}}$ such that values less than or equal to $P_{t_{\\mathrm{var}}}$ are set to 1, and values greater than $P_{t_{\\mathrm{var}}}$ are set to 0. This can be mathematically expressed as follows:", + "bbox": [ + 89, + 612, + 482, + 777 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM _ {\\text {v a r}} = \\mathbb {1} \\left[ A _ {\\text {v a r}} \\leq P _ {t _ {\\text {v a r}}} \\right], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 779, + 482, + 797 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbb{1}$ is the indicator random variable. Subsequently, we derive $A_{\\mathrm{var}}^{\\prime}$ from the same process, using the perturbed image $x + \\delta$ , and perform attention unequalization on the regions defined by $M_{\\mathrm{var}}$ . This method generates missing values by assigning random attention to previously unattended regions between the original images, with the loss function", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathcal{L}_{\\mathrm{attn}}$ defined as follows:", + "bbox": [ + 514, + 470, + 681, + 484 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a t t n}} (\\delta ; x, \\sigma_ {\\max }) = \\| \\left(\\sigma_ {\\max } - A _ {\\operatorname {v a r}} ^ {\\prime}\\right) \\odot M _ {\\operatorname {v a r}} \\| _ {2}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 496, + 906, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\odot$ is the Hadamard product, and $\\sigma_{\\mathrm{max}}$ denotes the maximum variance that can be obtained from the Softmax output based on the seq, following the equation:", + "bbox": [ + 511, + 525, + 905, + 571 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {\\max } = \\frac {1}{\\operatorname {s e q}} \\left(\\left(1 - \\frac {1}{\\operatorname {s e q}}\\right) ^ {2} + (\\operatorname {s e q} - 1) \\cdot \\frac {1}{\\operatorname {s e q} ^ {2}}\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 575, + 905, + 616 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the supplementary material, we provide the algorithm that outlines the method for calculating $\\mathcal{L}_{\\mathrm{attn}}$", + "bbox": [ + 511, + 625, + 905, + 656 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Facial Feature Extractor Attack", + "text_level": 1, + "bbox": [ + 511, + 664, + 795, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We design additional perturbation targeting two types of facial extraction, enhancing the applicability of our method not only to DM-based models but also to various other deep-fake architecture-based models.", + "bbox": [ + 511, + 686, + 905, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "MTCNN attack. We break down the MTCNN attack we propose into three principal stages: (i) selecting the resizing scale, (ii) enhancing robustness against interpolation, and (iii) formulating a loss function to expedite convergence. Through this process, we achieve not only various resize modes but also model transferability.", + "bbox": [ + 511, + 750, + 905, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Firstly, we select a set of appropriate resizing scales $s_i \\in S$ . This is to ensure that our attack technique effectively targets only the bounding boxes reaching the final layers of MTCNN. The suitable scale values are selected", + "bbox": [ + 511, + 840, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10367", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "among the multi-scale factors that the MTCNN model internally uses, and the detailed workings are provided in the supplementary material.", + "bbox": [ + 89, + 90, + 480, + 135 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Using the scale factor $s_i$ selected in the previous step, we next scale the input image size $D_{\\mathrm{adv}} = (h,w)$ , where $h$ and $w$ are the image's height and width, to yield $D_{\\mathrm{sc1}} = (s_i \\cdot h, s_i \\cdot w)$ . This results in an intermediate size $D_{\\mathrm{int}} = D_{\\mathrm{adv}} \\odot D_{\\mathrm{sc1}}$ obtained through element-wise multiplication. The input image $x \\in \\mathbb{R}^{c \\times h \\times w}$ is then upscaled to $D_{\\mathrm{int}}$ using NEAREST interpolation. Afterward, we downsample the image to $D_{\\mathrm{sc1}}$ via average pooling, assigning equal weights to each region referenced during interpolation. This approach runs parallel with a direct BILINEAR scaling of $D_{\\mathrm{adv}}$ to $D_{\\mathrm{sc1}}$ , thereby ensuring robust noise generation that functions effectively across various interpolation modes.", + "bbox": [ + 89, + 136, + 482, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the final stage, we perform a targeted attack on the initial P-Net $\\mathcal{T}$ to effectively disrupt the cascade pyramid structure. We pass the downsampled adversarial noise-added image $\\tilde{x}_{\\mathrm{adv}} \\in \\mathbb{R}^{c \\times s_i \\cdot h \\times s_i \\cdot w}$ through $\\mathcal{T}$ , which outputs probabilities $P_{\\mathrm{T,F}}$ for bounding boxes. To expedite the convergence of the MTCNN loss function $\\mathcal{L}_{\\mathrm{mtcnn}}$ , we propose a masking technique that leverages both the existence probabilities $P_{\\mathrm{T}}$ and the non-existence probabilities $P_{\\mathrm{F}}$ . The mask $M_{\\mathrm{prob}}$ is constructed to retain indices in $P_{\\mathrm{T}}$ that exceed the detection threshold $t_{\\mathrm{prob}}$ :", + "bbox": [ + 89, + 316, + 483, + 470 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nM _ {\\text {p r o b}} = \\mathbb {1} \\left[ P _ {\\mathrm {T}} (i, j) > t _ {\\text {p r o b}} \\right], \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 476, + 482, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbb{1}$ is the indicator random variable. Then, the $\\mathcal{L}_{\\mathrm{mtcnn}}$ converges with the mean squared error loss using $M_{\\mathrm{prob}}$ :", + "bbox": [ + 89, + 500, + 482, + 531 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {m t c n n}} (\\delta ; x, p _ {\\mathrm {g t}}) = \\left\\| \\left(\\mathcal {T} (x + \\delta) - p _ {\\mathrm {g t}}\\right) \\odot M _ {\\text {p r o b}} \\right\\| _ {2}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 537, + 482, + 555 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{T}(\\cdot) = [P_{\\mathrm{F}}, P_{\\mathrm{T}}]^T, p_{\\mathrm{gt}} = [t_{\\mathrm{prob}} + \\beta t_{\\mathrm{prob}} - \\beta]^T$ , and $\\beta$ is a value between 0 and 1. Additional details are provided along with the algorithm in the supplementary material.", + "bbox": [ + 89, + 561, + 482, + 608 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Identity attack. To effectively disrupt the accurate reflection of source face information, we target the ArcFace $\\mathcal{A}$ models [7], which are face identity embedding models widely used in deepfake systems. To improve transferability, we ensemble the most commonly used pre-trained backbones within these models. Since $\\mathcal{A}$ represents feature vectors extracted from the same person's face as vectors pointing in similar directions, we designed our approach to induce divergence from the original image $x$ by employing cosine similarity loss, thereby effectively obscuring the relevant identity information:", + "bbox": [ + 89, + 609, + 482, + 776 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {i d}} (\\delta ; x) = \\frac {\\mathcal {A} (x + \\delta) \\cdot \\mathcal {A} (x)}{\\| \\mathcal {A} (x + \\delta) \\| _ {2} \\| \\mathcal {A} (x) \\| _ {2}} - 1. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 782, + 482, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overall loss operation. Accordingly, the total loss function $\\mathcal{L}_{\\mathrm{total}}$ is defined and used as follows:", + "bbox": [ + 89, + 823, + 482, + 854 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {\\text {p r o j}} \\mathcal {L} _ {\\text {p r o j}} + \\lambda_ {\\text {a t t n}} \\mathcal {L} _ {\\text {a t t n}} \\\\ + \\lambda_ {\\mathrm {m t c n n}} \\mathcal {L} _ {\\mathrm {m t c n n}} + \\lambda_ {\\mathrm {i d}} \\mathcal {L} _ {\\mathrm {i d}}, \\tag {12} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 861, + 482, + 896 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: FaceShield" + ], + "code_body": "Input: image $x$ , steps $N$ , noise clamp $\\epsilon$ , step size $\\alpha$ , MTCNN detection threshold $t_{\\mathrm{prob}}$ , threshold weight $\\beta$ , CLIP Image Projector $\\mathcal{P}$ , Mid-layer cross-attention variance in Stable Diffusion $A_{\\mathrm{var}}'$ , MTCNN P-Network $\\mathcal{T}$ , ArcFace $\\mathcal{A}$", + "bbox": [ + 509, + 111, + 866, + 186 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Result: protected image $x_{\\mathrm{adv}}$", + "bbox": [ + 509, + 188, + 704, + 202 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [], + "code_body": "1 Initialize adversarial perturbation $\\delta \\gets 0$ , and protected image $x_{\\mathrm{adv}} \\gets x$", + "bbox": [ + 519, + 203, + 831, + 233 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2 for $n = 1,\\dots,N$ do", + "bbox": [ + 519, + 234, + 661, + 246 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [], + "code_body": "4 $\\mathcal{L}_{\\mathrm{attn}}\\gets ||(\\sigma_{\\mathrm{max}} - A_{\\mathrm{var}}^{\\prime})\\odot M_{\\mathrm{var}}||_{2},$ where $\\sigma_{\\mathrm{max}}$ derived from Eq. (8), and $M_{\\mathrm{var}}$ from Eq. (6)", + "bbox": [ + 519, + 263, + 849, + 306 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [], + "code_body": "5 $\\mathcal{L}_{\\mathrm{mtcnn}}\\gets \\left\\| (\\mathcal{T}(x + \\delta) - p_{\\mathrm{gt}})\\odot M_{\\mathrm{prob}}\\right\\| _2,$ where $p_{\\mathrm{gt}} = [t_{\\mathrm{prob}} + \\beta t_{\\mathrm{prob}} - \\beta ]^T$ , and $M_{\\mathrm{prob}}$ from Eq. (9)", + "bbox": [ + 519, + 309, + 874, + 354 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "3 $\\mathcal{L}_{\\mathrm{proj}}\\gets \\| \\mathcal{P}(x + \\delta) - \\mathcal{P}(x)\\| _1$", + "6 $\\mathcal{L}_{\\mathrm{id}}\\gets \\frac{\\mathcal{A}(x + \\delta)\\cdot\\mathcal{A}(x)}{\\|\\mathcal{A}(x + \\delta)\\|_2\\|\\mathcal{A}(x)\\|_2} -1$", + "7 Compute the total attack loss:" + ], + "bbox": [ + 519, + 247, + 874, + 388 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {\\text {p r o j}} \\mathcal {L} _ {\\text {p r o j}} + \\lambda_ {\\text {a t t n}} \\mathcal {L} _ {\\text {a t t n}} + \\\\ \\lambda_ {\\text {m t c n n}} \\mathcal {L} _ {\\text {m t c n n}} + \\lambda_ {\\text {i d}} \\mathcal {L} _ {\\text {i d}} \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 390, + 812, + 420 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "8 Update adversarial perturbation: $\\delta \\gets \\alpha \\cdot \\mathrm{sign}(\\nabla_{x_{\\mathrm{adv}}}\\mathcal{L}_{\\mathrm{total}})$", + "9 $\\delta_{\\mathrm{blur}}\\gets \\mathbf{G}$ aussianBlur(δ)", + "10 $\\delta_{\\mathrm{rgb}}^{\\prime}\\gets \\mathrm{LowPassFilter}(\\delta_{\\mathrm{blur}})$", + "11 $x_{\\mathrm{adv}}\\gets x_{\\mathrm{adv}} - \\delta_{\\mathrm{rgb}}^{\\prime}$", + "12 $x_{\\mathrm{adv}} \\gets x + \\operatorname{clip}(x_{\\mathrm{adv}} - x, -\\epsilon, \\epsilon)$" + ], + "bbox": [ + 517, + 421, + 787, + 515 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "13 end", + "bbox": [ + 514, + 517, + 563, + 529 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "14 Clip the image range: $x_{\\mathrm{adv}} \\gets \\mathrm{clip}(x_{\\mathrm{adv}}, 0, 255)$", + "bbox": [ + 514, + 530, + 849, + 546 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where each $\\lambda$ is a hyperparameter derived from grid searches to control the strength of the respective loss term. Additionally, the sign of $\\lambda$ determines the convergence or divergence of the loss function (i.e., $\\lambda_{\\mathrm{proj}}$ and $\\lambda_{\\mathrm{id}}$ are negative, while $\\lambda_{\\mathrm{attn}}$ and $\\lambda_{\\mathrm{mtcnn}}$ are positive).", + "bbox": [ + 511, + 560, + 906, + 637 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Enhanced Noise Update", + "text_level": 1, + "bbox": [ + 513, + 656, + 733, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We integrate two additional techniques into the standard PGD to enhance robustness by enabling more imperceptible noise updates and preventing the loss of information due to purification techniques.", + "bbox": [ + 511, + 681, + 906, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Gaussian blur. To enhance noise imperceptibility, we introduce a technique that constrains variations between adjacent regions, addressing the limitations of PGD methods (see Eq. (3)) that only regulate overall noise magnitude. This stems from the observation that differences between neighboring pixels can be as perceptible as the total noise itself. To achieve this, we utilize the Sobel operator [17] to emphasize areas of rapid intensity change, generating a mask $M_{\\mathrm{sob}}$ that highlights image boundaries. Gaussian blur $\\mathcal{G}(\\cdot)$ is then applied selectively to these regions during noise up-", + "bbox": [ + 511, + 750, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10368", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/68bbecd2ee370670e2083198557ae5165196c95048ac83f52e6d9cd858ddd4b3.jpg", + "image_caption": [ + "Figure 4. Qualitative Results. Protection performance across various deepfake models when our adversarial noise is applied. Models [34, 43] highlighted in the orange box typically exhibit facial distortions due to the influence described in Sec. 3.2, while those [3, 11, 20, 49] in the blue box display newly generated faces that diverge from the source image, attributed to the impact detailed in Sec. 3.3." + ], + "image_footnote": [], + "bbox": [ + 93, + 89, + 903, + 281 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "dates, ensuring smoother transitions between adjacent pixels and maintaining a consistent visual appearance:", + "bbox": [ + 89, + 340, + 482, + 372 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\delta_ {\\text {b l u r}} = \\mathcal {G} (\\delta) \\odot M _ {\\text {s o b}} + \\delta \\odot (1 - M _ {\\text {s o b}}). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 383, + 482, + 400 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Low pass filtering. To minimize information loss when saving images in JPEG format and ensure robustness to bit reduction during the compression process [10], low-frequency components are utilized. At each iteration, the newly updated adversarial noise $\\delta_{\\mathrm{rgb}} \\in \\mathbb{R}^{c \\times h \\times w}$ , where $c$ (channel), $h$ (height), and $w$ (width) undergoes a padding operation and patchification according to a predefined patch size $p$ . Then a DCT transform [1] is applied to each patch and channel, resulting in $\\delta_{\\mathrm{dct}} \\in \\mathbb{R}^{c \\times h' \\times w' \\times p \\times p}$ , where $h' = h/p$ and $w' = w/p$ , in the frequency domain. Using a low-pass filtering mask $M_{\\mathrm{lp}}$ , only the low-frequency components of the noise are retained. The noise $\\delta_{\\mathrm{rgb}}' \\in \\mathbb{R}^{c \\times h \\times w}$ is then reconstructed back into the RGB domain through an inverse transformation. The effectiveness of this approach is demonstrated through the experimental results presented in the supplementary material, and the overall operation of FaceShield is described in Algorithm 1.", + "bbox": [ + 88, + 414, + 483, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 684, + 223, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Setups", + "text_level": 1, + "bbox": [ + 89, + 709, + 179, + 726 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation details. For a fair performance comparison, we use open-source baseline [23, 24, 29, 42] and apply noise to the same dataset under identical hyperparameter settings. The corresponding results are presented in Table 2, while Table 1 provides a performance comparison on diffusion-based deepfakes [20, 34, 43, 49]. The extensibility experiments on GAN-based models [3, 11] are shown in Table 4, where, in the absence of existing attack methods for these models, we validate FaceShield's effectiveness through comparisons with the original images. In cases where the feature extractor fails to detect a face, we adjust", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the generation process to exclude facial features during reconstruction. Detailed descriptions and an analysis of the resources are provided in the supplementary material.", + "bbox": [ + 511, + 340, + 903, + 386 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We evaluate our method using two datasets: CelebA-HQ [18] and VGGFace2-HQ [4], both of which have been used in previous studies [3, 11, 34]. The former is the high-resolution version of CelebA, containing 30,000 celebrity face images, while the latter is the high-resolution version of VGGFace2, consisting of 3.3 million face images from 9,131 unique identities. For our experiments, we randomly select 200 identities from each dataset, using 100 images for the source and 100 images for the target.", + "bbox": [ + 511, + 388, + 906, + 526 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Qualitative Results", + "text_level": 1, + "bbox": [ + 511, + 535, + 696, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Performance results across deepfakes. As shown in Fig.4, FaceShield demonstrates robustness across various deepfake models. The perturbations result in either (i) pronounced artifacts reflecting non-relevant facial information instead of key features [34, 43], or (ii) a complete misinterpretation of the source face, generating a new, unrelated identity [3, 11, 20, 49].", + "bbox": [ + 511, + 560, + 905, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with state-of-the-art methods. We compare our method with baselines on DM-based deepfake model [43]. Although the methods [23, 24, 29, 42] that achieved high performance in diffusion adversarial attacks fail to induce visible changes on the deepfake model, ours demonstrates strong protective performance (Fig.5).", + "bbox": [ + 511, + 669, + 905, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.Quantitative Results", + "text_level": 1, + "bbox": [ + 511, + 768, + 707, + 786 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Automatic metrics. As shown in Table 1, we compare FaceShield to baseline methods across deepfake models [3, 11, 20, 34, 43, 49] using $L_{2}$ , Identity Score Matching [32] (ISM), and PSNR. The $L_{2}$ and PSNR metrics evaluate image quality by comparing deepfake results from clean and protected images, with higher $L_{2}$ and lower PSNR indicating more distortion. ISM measures the similarity between", + "bbox": [ + 511, + 794, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10369", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9dd095337cb007d42da9b8d5b8dcb335459245e557fc55f72cf9980b3b1b40c4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelDiffFace [20]DiffSwap [49]FaceSwap [34]IP-Adapter [43]
DatasetCelebA-HQ [18]
MethodL2↑ISM↓PSNR↓HE↑L2↑ISM↓PSNR↓HE↑L2↑ISM↓PSNR↓HE↑L2↑ISM↓PSNR↓HE↑
AdvDM [24]0.0210.47139.3684.220.0680.19928.3624.680.3030.24521.6154.520.2070.23525.3322.76
Mist [23]0.0210.46839.4433.940.0670.20128.3844.180.2870.23022.2634.780.1520.26528.2134.26
PhotoGuard [29]0.0220.46939.1943.820.0680.20128.2924.580.2820.23822.3164.440.1530.26828.1014.44
SDST [42]0.0210.47039.5124.080.0670.20728.3835.040.2740.26122.5824.680.1470.27328.4404.32
Ours0.0440.24332.0525.760.0720.16327.8336.200.3360.19420.7596.160.3500.07220.2666.60
Ours (Q=75)0.0430.25932.259-0.0700.16928.034-0.3170.20921.286-0.3260.11220.867-
DatasetVGGFace2-HQ [4]
MethodL2↑ISM↓PSNR↓HE↑L2↑ISM↓PSNR↓HE↑L2↑ISM↓PSNR↓HE↑L2↑ISM↓PSNR↓HE↑
AdvDM [24]0.0420.47933.0643.680.1050.21524.7694.780.4190.36118.5964.380.2510.27123.2502.36
Mist [23]0.0410.47833.2154.260.1020.22724.9643.940.3790.25919.6264.500.1810.29126.0704.10
PhotoGuard [29]0.0430.47932.9383.960.1100.21524.2724.180.3730.26619.6554.140.1800.29426.1573.82
SDST [42]0.0410.48333.2425.300.1070.22524.5064.580.3590.25819.9964.140.1660.29226.7844.06
Ours0.0620.27829.2046.100.1130.17724.0546.120.4530.23717.9196.160.3820.11219.4786.42
Ours (Q=75)0.0600.30829.435-0.1120.18524.201-0.4210.23718.573-0.3770.16719.618-
", + "bbox": [ + 101, + 88, + 903, + 347 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/45f719ce1a6e7f608513429d4db0b62680f6fd28e5810c1319a3b22d35cd8e30.jpg", + "image_caption": [ + "Figure 5. We generate deepfake [43] results from protected images of methods [23, 24, 29, 42]. While these fail to disrupt deepfake generation, our method causes deepfakes to malfunction." + ], + "image_footnote": [], + "bbox": [ + 93, + 412, + 480, + 526 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the source face and the deepfake output, with lower values indicating less similarity. We conduct experiments on 100 source-target pairs from CelebA-HQ [18] and VGGFace2-HQ [4], showing that FaceShield outperforms baselines across all metrics. We also analyze the noise levels in protected images using LPIPS, PSNR, and SSIM, as shown in Table 2. These image quality metrics, compared between protected and original images, show that our method consistently produces less noise than baseline methods. Additionally, we measure the Frequency Rate (FR), which indicates that most of FaceShield's noise is concentrated in low frequencies. This property helps maintain its effectiveness under JPEG compression. To verify, we compressed the protected images to JPEG Quality 75 and tested across four deepfake models. The results show that while performance slightly decreases, FaceShield still outperforms baseline methods, as shown in Table 1, Ours $(\\mathrm{Q} = 75)$ .", + "bbox": [ + 88, + 589, + 482, + 847 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Human evaluation. We conduct a human evaluation (HE) on the same methods and models, using 20 source images and 100 participants recruited via Amazon Mechanical", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/233f4091ceed9970ee450aa6b627e1db6155d0c205fedab74b42836e09372ba2.jpg", + "table_caption": [ + "Table 1. Comparison of perturbation effectiveness among baseline methods on four deepfake models using the CelebA-HQ [18] and VGGFace2-HQ [4] datasets. Our method exhibits the largest distortion in image quality (L2, PSNR) and source similarity (ISM), as well as in human evaluation (HE). Results on JPEG-compressed images (Quality factor 75) further confirm robust protection under compression." + ], + "table_footnote": [], + "table_body": "
DatasetCelebA-HQ [18]
MethodLPIPS ↓PSNR ↑SSIM ↑FR ↑HE ↑
AdvDM [24]0.421430.44760.84382.10773.86
Mist [23]0.549229.99350.86841.65834.70
PhotoGuard [29]0.551529.91270.86691.65384.82
SDST [42]0.540931.47620.90331.67675.12
Ours0.201732.62890.939418.46515.64
DatasetVGGFace2-HQ [4]
MethodLPIPS ↓PSNR ↑SSIM ↑FR ↑HE ↑
AdvDM [24]0.410830.25230.84362.06673.66
Mist [23]0.520829.90680.87211.68724.34
PhotoGuard [29]0.522129.82040.87121.68244.62
SDST [42]0.506031.35450.90921.68924.48
Ours0.194131.57990.934118.04005.28
", + "bbox": [ + 517, + 410, + 905, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. With the same step size, iterations, and noise clamping values applied, our method shows the least distortion across three image quality metrics (LPIPS, SSIM, PSNR, HE) and demonstrates a higher low-frequency content (FR).", + "bbox": [ + 511, + 601, + 906, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Turk. Participants assess two factors: protection noise visibility (Table 2) and the similarity between the source image and its deepfake output (Table 1). We employ a Likert scale from 1 to 7. For noise visibility, a score of 7 indicates the least visible noise, while for deepfake similarity, a score of 7 reflects a significant deviation from the source identity.", + "bbox": [ + 511, + 665, + 906, + 756 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 768, + 666, + 786 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of gaussian blur on noise. To evaluate the effect of Gaussian blur, one of the enhanced noise update methods, we present qualitative results in the Supplementary Material comparing the blur effect's presence and absence. From this comparison, it is clear that the noise update becomes significantly more invisible at the boundaries of abrupt changes in the noise, as detected through Sobel filtering.", + "bbox": [ + 511, + 794, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10370", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of each loss function. To demonstrate the generalizability of each loss function, we conducted an ablation study across six models using the ISM metric, as shown in Table 3. The results illustrate how FaceShield protects faces, as seen in Table 1 and Table 4, with red shading indicating performance degradation when a loss function is removed. This experiment shows that each loss function successfully impacts multiple models, and by combining them into $\\mathcal{L}_{\\mathrm{total}}$ , we cover a broader range of deepfakes.", + "bbox": [ + 89, + 90, + 483, + 227 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a3997732db5b3407d6da389ef3a9416b951b64c881bcb27b53c0800736d71fd0.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3. Each model's performance is measured using the ISM score, confirming that each loss function ensures transferability across various deepfakes. As a result, the integrated $\\mathcal{L}_{\\mathrm{total}}$ is capable of covering a wider range." + ], + "table_body": "
ISM↓DiffFaceDiffSwapFaceSwapIP-AdapterSimSwapInfoSwap
w/o Lproj0.2410.1670.2700.1350.5440.256
w/o Lattn0.2540.1700.2230.0760.1680.252
w/o Lmtconn0.2310.1740.1660.0470.1830.354
w/o Lid0.4460.1750.2170.0400.5120.430
Ltotal0.2430.1630.1940.0720.1840.237
", + "bbox": [ + 99, + 229, + 478, + 306 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Attack effectiveness of mid-layers. We qualitatively show that focusing the diffusion attack on the mid-layers of the denoising UNet [27] is more effective than applying it to the entire layers or the up/down layers, as shown in Fig.6. The experiment is conducted by applying noise $\\delta = 4 / 255$ to create protected images, which are then passed through the deepfake model [43] to compare the resulting outputs.", + "bbox": [ + 89, + 376, + 483, + 483 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6cfc6f3cb61eebe9e730a7e21f05e79c5ac08de5137498d0fbc3dc69f041465c.jpg", + "image_caption": [ + "Figure 6. A comparison of conditioned face attack results across UNet [27] layers shows the best protection when targeting mid-layer cross-attention." + ], + "image_footnote": [], + "bbox": [ + 93, + 488, + 483, + 665 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "MTCNN resize evaluation. To demonstrate the superiority of our method across various resizing modes and model transferability, we conduct an ablation study comparing it to the conventional BILINEAR scaling method [46]. We evaluate performance using different scaling methods from the OpenCV and PIL libraries, with 3,000 images from both the CelebA-HQ and VGGFace2-HQ datasets. Transferability is assessed through experiments conducted on both PyTorch and TensorFlow versions. The evaluation is based on the final bounding box detection probabilities from MTCNN [48], and the results in the supplementary material confirm that our method outperforms existing approaches.", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Applications", + "text_level": 1, + "bbox": [ + 513, + 90, + 648, + 107 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Extensibility on GAN-based deepfake models. We also conduct additional experiments on the GAN-based diffusion model [3, 11]. The experimental conditions are the same as those in Table 2, and the results, as shown in Table 4, indicate a degradation in model performance. Qualitative assessments are provided in Fig. 4.", + "bbox": [ + 511, + 114, + 906, + 205 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c7ceb026bad85236b5b3e4120d725b983e527f8b4e23cb5685999d0d22ea59f7.jpg", + "table_caption": [], + "table_footnote": [ + "Table 4. Applicability of FaceShield to other deepfake frameworks. Our method, when applied to GAN-based models, not only reduces image quality $(L_{2}$ , PSNR) but also significantly lowers source similarity (ISM)." + ], + "table_body": "
ModelSimSwap [3]InfoSwap [11]
DatasetCelebA-HQ [18]
MethodL2↑ISM↓PSNR↓L2↑ISM↓PSNR↓
Original0.0000.54480.0000.0000.43180.000
Ours0.0700.18426.9210.1290.23730.220
DatasetVGGFace2-HQ [4]
MethodL2↑ISM↓PSNR↓L2↑ISM↓PSNR↓
Original0.0000.68180.0000.0000.56580.000
Ours0.0670.31427.3050.1420.35629.044
", + "bbox": [ + 517, + 212, + 903, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Transferability with different weights. To demonstrate that FaceShield ensures robust transferability not only across structurally different models but also to models with similar architectures but different pre-trained weights, we evaluated its performance on various versions of IP-Adapter [43]. The results, which can be found in the supplementary material, confirm the superior transferability performance of our method.", + "bbox": [ + 511, + 406, + 908, + 526 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 540, + 633, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this study, we propose FaceShield, an invisible facial protection technique designed to attack various deepfakes. Through comparisons with multiple baseline methods, we demonstrate that FaceShield offers superior protection with significantly lower resource costs, particularly for deepfake models utilizing the latest diffusion techniques. Furthermore, its design integrates diverse transferability enhancement strategies, ensuring consistent performance not only across various pretrained versions but also across diffusion-based models with different architectures. This robustness extends to entirely different architectures, including GAN-based models. Additionally, by incorporating an improved noise update mechanism that ensures invisibility while minimizing information loss, FaceShield proves to be a practical and effective solution for preventing the misuse of facial images across a wide range of deepfake systems.", + "bbox": [ + 511, + 566, + 908, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and Future Work. Although we introduce a method to enhance robustness against JPEG compression and resizing, other purification techniques still exist, which may lead to the potential loss of our protective noise information. Therefore, we plan to further strengthen the noise to effectively counter a broader range of purification methods.", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10371", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgment", + "text_level": 1, + "bbox": [ + 89, + 90, + 263, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by Culture, Sports and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture, Sports and Tourism (International Collaborative Research and Global Talent Development for the Development of Copyright Management and Protection Technologies for Generative AI, RS-2024-00345025, $47\\%$ ; Research on neural watermark technology for copyright protection of generative AI 3D content, RS-2024-00348469, $25\\%$ ), the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT)(RS-2025-00521602, $25\\%$ ), Institute of Information & communications Technology Planning & Evaluation (IITP) & ITRC(Information Technology Research Center) grant funded by the Korea government(MSIT) (No.RS-2019-II190079, Artificial Intelligence Graduate School Program(Korea University), $1\\%$ ; IITP-2025-RS-2024-00436857, $1\\%$ ; IITP-2025-RS-2025-02304828, Artificial Intelligence Star Fellowship Support Program to Nurture the Best Talents, $1\\%$ ), and Artificial intelligence industrial convergence cluster development project funded by the Ministry of Science and ICT(MSIT, Korea)&Gwangju Metropolitan City.", + "bbox": [ + 89, + 114, + 485, + 449 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 458, + 187, + 474 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] N. Ahmed, T. Natarajan, and K.R. Rao. Discrete cosine transform. IEEE Transactions on Computers, C-23(1):90–93, 1974. 6", + "[2] Huanran Chen, Yichi Zhang, Yinpeng Dong, Xiao Yang, Hang Su, and Jun Zhu. Rethinking model ensemble in transfer-based adversarial attacks. In The Twelfth International Conference on Learning Representations, 2024. 3", + "[3] Renwang Chen, Xuanhong Chen, Bingbing Ni, and Yanhao Ge. Simswap: An efficient framework for high fidelity face swapping. In Proceedings of the 28th ACM international conference on multimedia, pages 2003-2011, 2020. 1, 2, 6, 8", + "[4] Xuanhong Chen, Bingbing Ni, Yutian Liu, Naiyuan Liu, Zhilin Zeng, and Hang Wang. Simswap++: Towards faster and high-quality identity swapping. IEEE Trans. Pattern Anal. Mach. Intell., 46(1):576-592, 2024. 6, 7, 8", + "[5] Jongwook Choi, Taehoon Kim, Yonghyun Jeong, Seungryul Baek, and Jongwon Choi. Exploiting style latent flows for generalizing deepfake video detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1133-1143, 2024. 1", + "[6] June Suk Choi, Kyungmin Lee, Jongheon Jeong, Saining Xie, Jinwoo Shin, and Kimin Lee. Diffusionguard: A robust defense against malicious diffusion-based image editing. arXiv preprint arXiv:2410.05694, 2024. 1", + "[7] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 2, 5" + ], + "bbox": [ + 99, + 483, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[8] Junhao Dong and Xiaohua Xie. Visually maintained image disturbance against deepfake face swapping. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 1", + "[9] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9185-9193, 2018. 2", + "[10] Gintare Karolina Dziugaite, Zoubin Ghahramani, and Daniel M Roy. A study of the effect of jpg compression on adversarial images. arXiv preprint arXiv:1608.00853, 2016. 6", + "[11] Gege Gao, Huaibo Huang, Chaoyou Fu, Zhaoyang Li, and Ran He. Information bottleneck disentanglement for identity swapping. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3404-3413, 2021. 2, 6, 8", + "[12] Xiao Guo, Xiaohong Liu, Zhiyuan Ren, Steven Grosz, Iacopo Masi, and Xiaoming Liu. Hierarchical fine-grained image forgery detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3155-3165, 2023. 1", + "[13] Baojin Huang, Zhongyuan Wang, Jifan Yang, Jiaxin Ai, Qin Zou, Qian Wang, and Dengpan Ye. Implicit identity driven deepfake face swapping detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4490-4499, 2023. 1", + "[14] Hao Huang, Yongtao Wang, Zhaoyu Chen, Yuze Zhang, Yuheng Li, Zhi Tang, Wei Chu, Jingdong Chen, Weisi Lin, and Kai-Kuang Ma. Cmua-watermark: A cross-model universal adversarial watermark for combating deepfakes. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 989-997, 2022. 1", + "[15] Qian Huang, Isay Katsman, Horace He, Zeqi Gu, Serge Belongie, and Ser-Nam Lim. Enhancing adversarial example transferability with an intermediate level attack. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4733-4742, 2019. 3", + "[16] Nathan Inkawich, Kevin Liang, Lawrence Carin, and Yiran Chen. Transferable perturbations of deep feature distributions. In International Conference on Learning Representations, 2020. 3", + "[17] Nick Kanopoulos, Nagesh Vasanthavada, and Robert L Baker. Design of an image edge detection filter using the sobel operator. IEEE Journal of solid-state circuits, 23(2): 358-367, 1988. 5", + "[18] Tero Karras. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 6, 7, 8", + "[19] Edgar Kaziakhmedov, Klim Kireev, Grigorii Melnikov, Mikhail Pautov, and Aleksandr Petiushko. Real-world attack on mtcnn face detection system. In 2019 International Multi-Conference on Engineering, Computer and Information Sciences (SIBIRCON), pages 0422-0427. IEEE, 2019. 2", + "[20] Kihong Kim, Yunho Kim, Seokju Cho, Junyoung Seo, Jisu Nam, Kychul Lee, Seung Wook Kim, and Kwanghee Lee." + ], + "bbox": [ + 514, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10372", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Difface: Diffusion-based face swapping with facial guidance. Pattern Recognit., 163:111451, 2022. 1, 2, 6, 7", + "[21] Jia Li, Zhaoyang Li, Jie Cao, Xingguang Song, and Ran He. Faceinpainter: High fidelity face adaptation to heterogeneous domains. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5089-5098, 2021. 2", + "[22] Lingzhi Li, Jianmin Bao, Hao Yang, Dong Chen, and Fang Wen. Faceshifter: Towards high fidelity and occlusion aware face swapping. arXiv preprint arXiv:1912.13457, 2019. 1, 2", + "[23] Chumeng Liang and Xiaoyu Wu. Mist: Towards improved adversarial examples for diffusion models. arXiv preprint arXiv:2305.12683, 2023. 1, 6, 7", + "[24] Chumeng Liang, Xiaoyu Wu, Yang Hua, Jiaru Zhang, Yiming Xue, Tao Song, Zhengui Xue, Ruhui Ma, and Haibing Guan. Adversarial example does good: Preventing painting imitation from diffusion models via adversarial examples. arXiv preprint arXiv:2302.04578, 2023. 1, 6, 7", + "[25] Kunlin Liu, Ivan Perov, Daiheng Gao, Nikolay Chervoniy, Wenbo Zhou, and Weiming Zhang. Deepfacelab: Integrated, flexible and extensible face-swapping framework. Pattern Recognit., 141:109628, 2020. 1", + "[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3", + "[27] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 3, 8", + "[28] Nataniel Ruiz, Sarah Adel Bargal, and Stan Sclaroff. Disrupting deepfakes: Adversarial attacks against conditional image translation networks and facial manipulation systems. In Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 236-251. Springer, 2020. 1", + "[29] Hadi Salman, Alaa Khaddaj, Guillaume Leclerc, Andrew Ilyas, and Aleksander Madry. Raising the cost of malicious ai-powered image editing. In International Conference on Machine Learning, 2023. 1, 6, 7", + "[30] Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y Zhao. Glaze: Protecting artists from style mimicry by {Text-to-Image} models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 2187–2204, 2023. 1", + "[31] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, and Yunchao Wei. Learning on gradients: Generalized artifacts representation for gan-generated images detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12105-12114, 2023. 1", + "[32] Thanh Van Le, Hao Phung, Thuan Hoang Nguyen, Quan Dao, Ngoc N Tran, and Anh Tran. Anti-dreambooth: Protecting users from personalized text-to-image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2116-2127, 2023. 6" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Andrey Voynov, Qinghao Chu, Daniel Cohen-Or, and Kfir Aberman. $\\mathfrak{p}+$ : Extended textual conditioning in text-to-image generation. arXiv preprint arXiv:2303.09522, 2023. 3", + "[34] Feifei Wang. Face swap via diffusion model. arXiv preprint arXiv:2403.01108, 2024. 1, 2, 6, 7", + "[35] Run Wang, Ziheng Huang, Zhikai Chen, Li Liu, Jing Chen, and Lina Wang. Anti-forgery: Towards a stealthy and robust deepfake disruption attack via adversarial perceptual-aware perturbations. arXiv preprint arXiv:2206.00477, 2022. 1", + "[36] Tianyi Wang and Kam Pui Chow. Noise based deepfake detection via multi-head relative-interaction. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 14548-14556, 2023. 1", + "[37] Yuhan Wang, Xu Chen, Junwei Zhu, Wenqing Chu, Ying Tai, Chengjie Wang, Jilin Li, Yongjian Wu, Feiyue Huang, and Rongrong Ji. Hififace: 3d shape and semantic prior guided high fidelity face swapping. arXiv preprint arXiv:2106.09965, 2021. 1", + "[38] Zhibo Wang, Hengchang Guo, Zhifei Zhang, Wenxin Liu, Zhan Qin, and Kui Ren. Feature importance-aware transferable adversarial attacks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7639-7648, 2021. 2", + "[39] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L Yuille. Improving transferability of adversarial examples with input diversity. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2730-2739, 2019. 3", + "[40] Chao Xu, Jiangning Zhang, Miao Hua, Qian He, Zili Yi, and Yong Liu. Region-aware face swapping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7632-7641, 2022. 2", + "[41] Yangyang Xu, Bailin Deng, Junle Wang, Yanqing Jing, Jia Pan, and Shengfeng He. High-resolution face swapping via latent semantics disentanglement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7642-7651, 2022. 2", + "[42] Haotian Xue, Chumeng Liang, Xiaoyu Wu, and Yongxin Chen. Toward effective protection against diffusion-based mimicry through score distillation. In The Twelfth International Conference on Learning Representations, 2023. 1, 6, 7", + "[43] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 1, 2, 6, 7, 8", + "[44] Chin-Yuan Yeh, Hsi-Wen Chen, Shang-Lun Tsai, and Sheng-De Wang. Disrupting image-translation-based deepfake algorithms with adversarial attacks. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision Workshops, pages 53–62, 2020. 1", + "[45] Ge Yuan, Maomao Li, Yong Zhang, and Huicheng Zheng. Reliableswap: Boosting general face swapping via reliable supervision. arXiv preprint arXiv:2306.05356, 2023. 2", + "[46] Chongyang Zhang, Yu Qi, and Hiroyuki Kameda. Multiscale perturbation fusion adversarial attack on mtconn face" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10373", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "detection system. In 2022 4th International Conference on Communications, Information System and Computer Engineering (CISCE), pages 142-146. IEEE, 2022. 2, 8", + "[47] Jianping Zhang, Zhuoer Xu, shiwen cui, Changhua Meng, Weibin Wu, and Michael Lyu. On the robustness of latent diffusion models, 2024. 3", + "[48] Kaipeng Zhang, Zhanpeng Zhang, Zhifeng Li, and Yu Qiao. Joint face detection and alignment using multitask cascaded convolutional networks. IEEE signal processing letters, 23 (10):1499-1503, 2016. 2, 8", + "[49] Wenliang Zhao, Yongming Rao, Weikang Shi, Zuyan Liu, Jie Zhou, and Jiwen Lu. Diffswap: High-fidelity and controllable face swapping via 3d-aware masked diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8568-8577, 2023. 1, 2, 6, 7", + "[50] Wen Zhou, Xin Hou, Yongjun Chen, Mengyun Tang, Xiangqi Huang, Xiang Gan, and Yong Yang. Transferable adversarial perturbations. In Proceedings of the European Conference on Computer Vision (ECCV), pages 452-467, 2018. 3", + "[51] Yuhao Zhu, Qi Li, Jian Wang, Cheng-Zhong Xu, and Zhenan Sun. One shot face swapping on megapixels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4834-4844, 2021. 1, 2", + "[52] Yixuan Zhu, Wenliang Zhao, Yansong Tang, Yongming Rao, Jie Zhou, and Jiwen Lu. Stableswap: Stable face swapping in a shared and controllable latent space. IEEE Transactions on Multimedia, 2024. 2" + ], + "bbox": [ + 91, + 90, + 482, + 500 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "10374", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/FaceShield_ Defending Facial Image against Deepfake Threats/4436b2e6-eb11-4750-8df1-5fa6df4f36fe_model.json b/2025/FaceShield_ Defending Facial Image against Deepfake Threats/4436b2e6-eb11-4750-8df1-5fa6df4f36fe_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a53b6616ef9867559baf3d7269e90d6743eacd53 --- /dev/null +++ b/2025/FaceShield_ Defending Facial Image against Deepfake Threats/4436b2e6-eb11-4750-8df1-5fa6df4f36fe_model.json @@ -0,0 +1,2400 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.125, + 0.108, + 0.877, + 0.134 + ], + "angle": 0, + "content": "FaceShield: Defending Facial Image against Deepfake Threats" + }, + { + "type": "text", + "bbox": [ + 0.179, + 0.155, + 0.817, + 0.174 + ], + "angle": 0, + "content": "Jaehwan Jeong\\(^{1}\\) Sumin In\\(^{1}\\) Sieun Kim\\(^{1}\\) Hannie Shin\\(^{1}\\) Jongheon Jeong\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.292, + 0.174, + 0.714, + 0.193 + ], + "angle": 0, + "content": "Sang Ho Yoon² Jaewook Chung³ Sangpil Kim¹*" + }, + { + "type": "text", + "bbox": [ + 0.291, + 0.208, + 0.708, + 0.228 + ], + "angle": 0, + "content": "\\(^{1}\\)Korea University \\(^{2}\\)KAIST \\(^{3}\\)Samsung Research" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.273, + 0.327, + 0.289 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.305, + 0.486, + 0.729 + ], + "angle": 0, + "content": "The rising use of deepfakes in criminal activities presents a significant issue, inciting widespread controversy. While numerous studies have tackled this problem, most primarily focus on deepfake detection. These reactive solutions are insufficient as a fundamental approach for crimes where authenticity is disregarded. Existing proactive defenses also have limitations, as they are effective only for deepfake models based on specific Generative Adversarial Networks (GANs), making them less applicable in light of recent advancements in diffusion-based models. In this paper, we propose a proactive defense method named FaceShield, which introduces novel defense strategies targeting deepfakes generated by Diffusion Models (DMs) and facilitates defenses on various existing GAN-based deepfake models through facial feature extractor manipulations. Our approach consists of three main components: (i) manipulating the attention mechanism of DMs to exclude protected facial features during the denoising process, (ii) targeting prominent facial feature extraction models to enhance the robustness of our adversarial perturbation, and (iii) employing Gaussian blur and low-pass filtering techniques to improve imperceptibility while enhancing robustness against JPEG compression. Experimental results on the CelebAHQ and VGGFace2-HQ datasets demonstrate that our method achieves state-of-the-art performance against the latest deepfake models based on DMs, while also exhibiting transferability to GANs and showcasing greater imperceptibility of noise along with enhanced robustness." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.745, + 0.222, + 0.761 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.771, + 0.485, + 0.879 + ], + "angle": 0, + "content": "The advancement of deepfake technology and improved accessibility [3, 22, 25, 37, 51] has led to significant transformations in modern society. Due to the ease of face swapping, it has been applied across various fields, providing both entertainment and convenience. However, its powerful capability to generate realistic content has also enabled malicious users to exploit it for criminal purposes, leading" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.284, + 0.883, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.441, + 0.907, + 0.497 + ], + "angle": 0, + "content": "Figure 1. Protecting Face during Deepfake using FaceShield. Pure images are vulnerable to face swapping, allowing the target image's face to be easily reflected. In contrast, images protected by FaceShield conceal facial feature from deepfake." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.902, + 0.523 + ], + "angle": 0, + "content": "to the creation of fake news and various societal problems." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.524, + 0.907, + 0.903 + ], + "angle": 0, + "content": "To address the growing concerns surrounding deepfake technology, various countermeasures have been explored, which can be broadly divided into two categories. The first is deepfake detection techniques [5, 12, 13, 31, 36], which act as passive defenses by classifying whether content is synthetic or authentic. While effective for authenticity verification, these offer only binary results and fail to address advanced threats, such as crimes using realistic fakes. In contrast, proactive defense strategies offer a more comprehensive solution. These approaches involve embedding imperceptible adversarial perturbation into face images to prevent the protected face from being effectively processed by deepfakes. However, most previous research [8, 14, 28, 35, 44] has concentrated on GAN-based models, often targeting individual models, which limits effectiveness against emerging DM-based deepfakes [20, 34, 43, 49]. Although significant research exists on image protection within DMs [6, 23, 24, 29, 30, 42] for image editing, the focus has primarily been on attacking the noising and denoising processes when an image is used as a query (Fig.2a). This leads to targeting the encoder or predicted noise post-UNet processing. However, we observe that such strategies are ineffective for DM-based deepfake models, where the source image influences the output in the form of key-value pairs through attention mechanisms (Fig.2b)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.888, + 0.241, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10364" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.091, + 0.466, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.09, + 0.904, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.221, + 0.908, + 0.266 + ], + "angle": 0, + "content": "Figure 2. Image editing and Deepfake processes in DMs. (a) In DM-based image editing, a single image is input as a query \\( Q \\) and edited based on a prompt condition. (b) In DM-based deepfake, two images are used, with the target image serving as the query \\( Q \\) while the source image acts as the condition for swapping. This condition operates as key \\( K \\) and value \\( V \\) in the cross-attention layer." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.277, + 0.482, + 0.413 + ], + "angle": 0, + "content": "In this paper, we focus on attacking state-of-the-art DM-based deepfakes while ensuring applicability to GAN-based models as well (Fig.1). Given the uncertainty surrounding deepfakes that malicious users might employ, we explore approaches to improve the extensibility across different architectures (e.g., GANs, DMs) and model transferability across different pre-trained backbones. Simultaneously, we propose a novel noise update method that enhances imperceptibility while being robust to JPEG compression." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.415, + 0.483, + 0.55 + ], + "angle": 0, + "content": "For DM-based deepfake attacks, we leverage the structural properties in which the conditioning image is embedded and integrated into the denoising UNet through attention mechanisms. By utilizing the IP-Adapter [43], commonly employed for inpainting, we extract effective adversarial noise from the embedding of the conditioning image. This perturbation effectively disrupts the propagation of the conditioning process, ensuring that the final output does not replicate the features of the protected image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.552, + 0.483, + 0.825 + ], + "angle": 0, + "content": "To enhance the generalizability of our approach, we target two commonly used facial feature extractors. First, we attack the MTCNN model [48], which uses a cascade pyramid architecture to achieve superior performance and robust detection capabilities. Due to this, it is widely adopted not only in deepfake generation but also across various applications. We leverage the fact that the model scales images to different sizes during face detection. Our perturbation is designed to ensure robustness across various scaling factors and interpolation modes (e.g., BILINEAR, AREA), leading to superior performance compared to existing methods [19, 46]. Additionally, we target ArcFace [7], a widely adopted pre-trained model for facial feature extraction in deepfake applications. By incorporating both methods into our work, we ensure that our approach disrupts a range of deepfakes commonly used for facial landmark detection and feature extraction, thereby improving the overall robustness of our method against various deepfake systems." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In the noise updating process, we refine the perturbation using two techniques: Noise Blur, which measures differences between adjacent pixels for imperceptible refinement, and Low-pass filtering, retaining low-frequency components, enhancing robustness against JPEG compression." + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.278, + 0.885, + 0.292 + ], + "angle": 0, + "content": "To summarize, our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.294, + 0.905, + 0.383 + ], + "angle": 0, + "content": "- We introduce a novel attack on deepfakes based on diffusion models. To the best of our knowledge, our proposed method is the first attempt to protect images used as conditions while demonstrating robust performance across various deepfake models by targeting common facial feature extractors." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.385, + 0.905, + 0.474 + ], + "angle": 0, + "content": "- We propose a novel noise update mechanism that integrates Gaussian blur technique with the projected gradient descent method, significantly enhancing imperceptibility. Additionally, we implement low-pass filtering to reduce perturbation loss rates during JPEG compression compared to existing methods." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.475, + 0.905, + 0.535 + ], + "angle": 0, + "content": "- We demonstrate that our deepfake attack method is robust across various deepfake models, outperforming previous diffusion attacks by achieving higher distortion with significantly less noise." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.294, + 0.905, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.551, + 0.655, + 0.567 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.58, + 0.907, + 0.838 + ], + "angle": 0, + "content": "Deepfake techniques. With advancements in generative models, deepfake technology has evolved into a specialized field focused on facial synthesis. Previous deepfake models, primarily based on GANs, generally follow a three-stage process: face detection and localization, feature extraction, and face swapping. Among these, studies such as [11, 40, 41, 45, 51] employ MTCNN [48] for face detection and landmark extraction, while the majority of deepfake models, including [3, 21, 22, 40, 51, 52], leverage ArcFace [7] for identity feature extraction. These steps are similarly employed in DM-based deepfakes that have emerged with the progress of diffusion models. Notable examples, including [20, 34, 49], integrate [7] to maintain identity consistency. However, recent work has focused on leveraging the capabilities of diffusion models to develop face-swapping methods [43] that achieve high performance without explicitly following previous approaches." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Enhancing model transferability. In the research on adversarial attacks, various attempts have been made to improve transferability. [9, 38] proposed the model ensemble technique, generating adversarial examples using multiple" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10365" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.227 + ], + "angle": 0, + "content": "models to enhance their effectiveness on unseen models. [39] introduced a method that selectively utilizes specific layers within a model to improve transferability. Similarly, [15, 16, 50] investigated techniques that manipulate intermediate layer feature distributions or amplify activation values to prevent adversarial noise from overfitting to a particular model. Furthermore, [2, 47] explored the use of multiple pre-trained backbones within similar model architectures to enhance transferability across different backbone networks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.24, + 0.182, + 0.256 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.265, + 0.484, + 0.448 + ], + "angle": 0, + "content": "We propose a novel pipeline, FaceShield, to safeguard facial images from being exploited by diverse Deepfake methods through conditional attacks on DMs and facial feature extractor attacks. In this section, we first introduce the foundational adversarial attack framework utilized across our approach (Sec.3.1). We then detail our method for disrupting information flow when a facial image is employed as a conditioning input in DMs (Sec.3.2). Subsequently, we present our approach for preventing accurate facial feature extraction (Sec.3.3). Finally, we introduce our adversarial noise update mechanism, designed to enhance imperceptibility and mitigate degradation from JPEG compression (Sec.3.4)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.454, + 0.232, + 0.469 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.48, + 0.484, + 0.631 + ], + "angle": 0, + "content": "Cross-attention mechanism. To condition generative DMs, the cross-attention mechanism is used, as shown in Fig.2. Similar to self-attention, it involves computations using the query \\( Q \\), key \\( K \\), and value \\( V \\). However, unlike self-attention, where \\( Q \\), \\( K \\) and \\( V \\) are derived from the same source, cross-attention conditions the process by obtaining \\( Q \\) from the noised image \\( z_{t} \\) through a learned linear projection \\( \\ell_{q} \\), while \\( K \\) and \\( V \\) are derived from the textual or image embedding \\( C_{\\mathrm{emb}} \\) using learned linear projection \\( \\ell_{k} \\) and \\( \\ell_{v} \\), respectively:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.639, + 0.483, + 0.656 + ], + "angle": 0, + "content": "\\[\nQ = \\ell_ {q} (z _ {t}), K = \\ell_ {k} \\left(C _ {\\text {e m b}}\\right), V = \\ell_ {v} \\left(C _ {\\text {e m b}}\\right), \\text {a n d} \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.129, + 0.658, + 0.483, + 0.693 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}}\\right) V, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.7, + 0.408, + 0.714 + ], + "angle": 0, + "content": "where \\(d_{k}\\) are the dimensions of the key vectors." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.716, + 0.484, + 0.837 + ], + "angle": 0, + "content": "Projected gradient descent (PGD). PGD is a widely used method for crafting adversarial examples when the user has access to the model parameters. This technique iteratively updates an adversarial perturbation by computing the gradient of a certain loss \\(\\mathcal{L}_{\\mathrm{adv}}\\) with respect to the input. At each step, noise is added in the gradient direction while keeping the perturbation within a predefined bound, ensuring the noise is small but effective:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.846, + 0.483, + 0.865 + ], + "angle": 0, + "content": "\\[\n\\delta \\leftarrow \\operatorname {P r o j} _ {\\| \\delta \\| _ {\\infty} \\leq \\eta} \\left(\\delta - \\alpha \\cdot \\operatorname {s i g n} \\left(\\nabla_ {\\delta} \\mathcal {L} _ {\\mathrm {a d v}}\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is the step size and \\(\\mathrm{Proj}_{\\| \\delta \\|_{\\infty} \\leq \\eta}(\\cdot)\\) projects \\(\\delta\\) onto the \\(\\ell_{\\infty}\\) ball of radius \\(\\eta\\). By projecting the adversarial exam" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "ple back onto the valid perturbation space, PGD maintains imperceptibility while disrupting the model's predictions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.133, + 0.742, + 0.149 + ], + "angle": 0, + "content": "3.2. Conditioned Face Attack" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.156, + 0.907, + 0.308 + ], + "angle": 0, + "content": "We now describe our approach to protecting images, specifically by disrupting the effective transfer of information when they are used as conditioning inputs in DMs. The core of our approach is to effectively interfere with key information using minimal noise, while also ensuring that the model does not overfit by accessing only a minimal number of layers to obtain gradients. To achieve this, we propose two methods that target both the initial projection phase and the final attention mechanism during the image conditioning process within latent diffusion models [27]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.31, + 0.909, + 0.493 + ], + "angle": 0, + "content": "Face projector attack. When images are used as conditioning inputs, they are firstly transformed into an embedding vector through a pre-trained model [26]. In this method, we access only the topmost layer \\(\\mathcal{P}\\) of the model to disrupt the projection process, causing the image to be projected with incorrect information at the initial stage. For the attack loss function, we consider that converging to a single target value might not ensure consistent convergence speeds or balanced performance. Given that one of our main goals is to design noise applicable to various images, we design our approach to induce random divergence based on the input image, using the \\(\\mathcal{L}_1\\) loss function in this process:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.505, + 0.907, + 0.523 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {p r o j}} (\\delta ; x) = \\| \\mathcal {P} (x + \\delta) - \\mathcal {P} (x) \\| _ {1}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.535, + 0.727, + 0.548 + ], + "angle": 0, + "content": "where \\(\\delta\\) is the adversarial noise." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.907, + 0.765 + ], + "angle": 0, + "content": "Attention disruption attack. We also focus on identifying the core vectors within the denoising UNet that are most sensitive to conditional inputs. Initially, we analyze the influence of cross-attention across each UNet layer. Based on prior research [33], which shows that different cross-attention layers respond variably to conditioning information, we investigate the impact on perturbation performance for each region. Our findings lead to the conclusion that targeting attacks near mid-layers produces more significant disruption in qualitative metrics compared to using only the up-down layers or the entire layers, as supported by our experimental results in Fig.6. Based on these insights, we propose a novel approach that specifically targets mid-layers during the attack on the diffusion process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.909, + 0.902 + ], + "angle": 0, + "content": "To induce a mismatch in conditioning, we use the mid-layer cross-attention mechanism, as described in Eq. (1). Based on the idea that the condition is conveyed to the query through attention, we calculate the attention score to obtain the strength of attention. This is done by performing operations on the query \\( Q \\in \\mathbb{R}^{h \\times \\mathrm{res} \\times d} \\) and key \\( K \\in \\mathbb{R}^{h \\times \\mathrm{seq} \\times d} \\), where \\( h \\) (number of heads), res (resolution), seq (sequence length), and \\( d \\) (head dimension). This is followed by a Softmax operation along the seq dimension to derive the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10366" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.09, + 0.448, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.451, + 0.091, + 0.899, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.387, + 0.907, + 0.457 + ], + "angle": 0, + "content": "Figure 3. Overview. Our method has three main parts: (i) Conditioned face attack, which disrupts feature transfer by targeting the embedding process and the attention map variance in the cross-attention layer; (ii) Facial feature extractor attack, which decreases the probability value of face detection and causes extraction disruptions, and (iii) Enhanced noise update, which improves imperceptibility by applying Gaussian blur to regions with significant intensity changes between adjacent pixels, and increases robustness against JPEG compression distortion by encoding the noise in the low-frequency domain." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.47, + 0.483, + 0.517 + ], + "angle": 0, + "content": "attention map \\(A_{\\mathrm{map}} \\in \\mathbb{R}^{h \\times \\mathrm{res} \\times \\mathrm{seq}}\\). Exploiting this mechanism, we obtain the variance \\(A_{\\mathrm{var}}\\), allowing us to evaluate attention strength by the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.111, + 0.527, + 0.483, + 0.567 + ], + "angle": 0, + "content": "\\[\nA _ {\\text {v a r}} = \\frac {1}{\\text {s e q}} \\sum_ {i = 1} ^ {\\text {s e q}} \\left(A _ {\\text {m a p}} [:, i,: ] - \\bar {A} _ {\\text {m a p}}\\right) ^ {2} \\in \\mathbb {R} ^ {h \\times \\text {r e s}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.58, + 0.483, + 0.612 + ], + "angle": 0, + "content": "where \\(\\bar{A}_{\\mathrm{map}} = \\frac{1}{\\mathrm{seq}}\\sum_{i' = 1}^{\\mathrm{seq}}A_{\\mathrm{map}}[:,i',:]\\) is the mean of attention map across the seq dimension." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.483, + 0.779 + ], + "angle": 0, + "content": "Based on them, we propose an adversarial attack strategy that maximizes \\( A_{\\mathrm{var}} \\), thereby preventing the proper reflection of conditional information \\( K \\) on \\( Q \\). In this process, we encode the original image \\( x \\) to use as the query \\( Q \\) and project the same \\( x \\) to obtain the key \\( K \\), which is then used to calculate \\( A_{\\mathrm{var}} \\). Thereafter, we find a quantile \\( P_{t_{\\mathrm{var}}} \\) corresponding to a predefined threshold \\( t_{\\mathrm{var}} \\) between 0 and 1 to identify the regions exhibiting weak attention. Using this \\( P_{t_{\\mathrm{var}}} \\), we create a mask \\( M_{\\mathrm{var}} \\) such that values less than or equal to \\( P_{t_{\\mathrm{var}}} \\) are set to 1, and values greater than \\( P_{t_{\\mathrm{var}}} \\) are set to 0. This can be mathematically expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.203, + 0.78, + 0.483, + 0.798 + ], + "angle": 0, + "content": "\\[\nM _ {\\text {v a r}} = \\mathbb {1} \\left[ A _ {\\text {v a r}} \\leq P _ {t _ {\\text {v a r}}} \\right], \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathbb{1}\\) is the indicator random variable. Subsequently, we derive \\(A_{\\mathrm{var}}^{\\prime}\\) from the same process, using the perturbed image \\(x + \\delta\\) , and perform attention unequalization on the regions defined by \\(M_{\\mathrm{var}}\\) . This method generates missing values by assigning random attention to previously unattended regions between the original images, with the loss function" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.471, + 0.682, + 0.485 + ], + "angle": 0, + "content": "\\(\\mathcal{L}_{\\mathrm{attn}}\\) defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.497, + 0.907, + 0.515 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a t t n}} (\\delta ; x, \\sigma_ {\\max }) = \\| \\left(\\sigma_ {\\max } - A _ {\\operatorname {v a r}} ^ {\\prime}\\right) \\odot M _ {\\operatorname {v a r}} \\| _ {2}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.526, + 0.906, + 0.572 + ], + "angle": 0, + "content": "where \\(\\odot\\) is the Hadamard product, and \\(\\sigma_{\\mathrm{max}}\\) denotes the maximum variance that can be obtained from the Softmax output based on the seq, following the equation:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.576, + 0.906, + 0.617 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {\\max } = \\frac {1}{\\operatorname {s e q}} \\left(\\left(1 - \\frac {1}{\\operatorname {s e q}}\\right) ^ {2} + (\\operatorname {s e q} - 1) \\cdot \\frac {1}{\\operatorname {s e q} ^ {2}}\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.626, + 0.906, + 0.657 + ], + "angle": 0, + "content": "In the supplementary material, we provide the algorithm that outlines the method for calculating \\(\\mathcal{L}_{\\mathrm{attn}}\\)" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.665, + 0.796, + 0.68 + ], + "angle": 0, + "content": "3.3. Facial Feature Extractor Attack" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.687, + 0.906, + 0.746 + ], + "angle": 0, + "content": "We design additional perturbation targeting two types of facial extraction, enhancing the applicability of our method not only to DM-based models but also to various other deep-fake architecture-based models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.906, + 0.84 + ], + "angle": 0, + "content": "MTCNN attack. We break down the MTCNN attack we propose into three principal stages: (i) selecting the resizing scale, (ii) enhancing robustness against interpolation, and (iii) formulating a loss function to expedite convergence. Through this process, we achieve not only various resize modes but also model transferability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Firstly, we select a set of appropriate resizing scales \\( s_i \\in S \\). This is to ensure that our attack technique effectively targets only the bounding boxes reaching the final layers of MTCNN. The suitable scale values are selected" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10367" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.136 + ], + "angle": 0, + "content": "among the multi-scale factors that the MTCNN model internally uses, and the detailed workings are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.137, + 0.483, + 0.318 + ], + "angle": 0, + "content": "Using the scale factor \\( s_i \\) selected in the previous step, we next scale the input image size \\( D_{\\mathrm{adv}} = (h,w) \\), where \\( h \\) and \\( w \\) are the image's height and width, to yield \\( D_{\\mathrm{sc1}} = (s_i \\cdot h, s_i \\cdot w) \\). This results in an intermediate size \\( D_{\\mathrm{int}} = D_{\\mathrm{adv}} \\odot D_{\\mathrm{sc1}} \\) obtained through element-wise multiplication. The input image \\( x \\in \\mathbb{R}^{c \\times h \\times w} \\) is then upscaled to \\( D_{\\mathrm{int}} \\) using NEAREST interpolation. Afterward, we downsample the image to \\( D_{\\mathrm{sc1}} \\) via average pooling, assigning equal weights to each region referenced during interpolation. This approach runs parallel with a direct BILINEAR scaling of \\( D_{\\mathrm{adv}} \\) to \\( D_{\\mathrm{sc1}} \\), thereby ensuring robust noise generation that functions effectively across various interpolation modes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.484, + 0.471 + ], + "angle": 0, + "content": "In the final stage, we perform a targeted attack on the initial P-Net \\(\\mathcal{T}\\) to effectively disrupt the cascade pyramid structure. We pass the downsampled adversarial noise-added image \\(\\tilde{x}_{\\mathrm{adv}} \\in \\mathbb{R}^{c \\times s_i \\cdot h \\times s_i \\cdot w}\\) through \\(\\mathcal{T}\\), which outputs probabilities \\(P_{\\mathrm{T,F}}\\) for bounding boxes. To expedite the convergence of the MTCNN loss function \\(\\mathcal{L}_{\\mathrm{mtcnn}}\\), we propose a masking technique that leverages both the existence probabilities \\(P_{\\mathrm{T}}\\) and the non-existence probabilities \\(P_{\\mathrm{F}}\\). The mask \\(M_{\\mathrm{prob}}\\) is constructed to retain indices in \\(P_{\\mathrm{T}}\\) that exceed the detection threshold \\(t_{\\mathrm{prob}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.477, + 0.483, + 0.495 + ], + "angle": 0, + "content": "\\[\nM _ {\\text {p r o b}} = \\mathbb {1} \\left[ P _ {\\mathrm {T}} (i, j) > t _ {\\text {p r o b}} \\right], \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.501, + 0.483, + 0.532 + ], + "angle": 0, + "content": "where \\(\\mathbb{1}\\) is the indicator random variable. Then, the \\(\\mathcal{L}_{\\mathrm{mtcnn}}\\) converges with the mean squared error loss using \\(M_{\\mathrm{prob}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.538, + 0.483, + 0.556 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {m t c n n}} (\\delta ; x, p _ {\\mathrm {g t}}) = \\left\\| \\left(\\mathcal {T} (x + \\delta) - p _ {\\mathrm {g t}}\\right) \\odot M _ {\\text {p r o b}} \\right\\| _ {2}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.562, + 0.483, + 0.609 + ], + "angle": 0, + "content": "where \\(\\mathcal{T}(\\cdot) = [P_{\\mathrm{F}}, P_{\\mathrm{T}}]^T, p_{\\mathrm{gt}} = [t_{\\mathrm{prob}} + \\beta t_{\\mathrm{prob}} - \\beta]^T\\), and \\(\\beta\\) is a value between 0 and 1. Additional details are provided along with the algorithm in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.611, + 0.483, + 0.777 + ], + "angle": 0, + "content": "Identity attack. To effectively disrupt the accurate reflection of source face information, we target the ArcFace \\(\\mathcal{A}\\) models [7], which are face identity embedding models widely used in deepfake systems. To improve transferability, we ensemble the most commonly used pre-trained backbones within these models. Since \\(\\mathcal{A}\\) represents feature vectors extracted from the same person's face as vectors pointing in similar directions, we designed our approach to induce divergence from the original image \\(x\\) by employing cosine similarity loss, thereby effectively obscuring the relevant identity information:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.783, + 0.483, + 0.817 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {i d}} (\\delta ; x) = \\frac {\\mathcal {A} (x + \\delta) \\cdot \\mathcal {A} (x)}{\\| \\mathcal {A} (x + \\delta) \\| _ {2} \\| \\mathcal {A} (x) \\| _ {2}} - 1. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.824, + 0.483, + 0.855 + ], + "angle": 0, + "content": "Overall loss operation. Accordingly, the total loss function \\(\\mathcal{L}_{\\mathrm{total}}\\) is defined and used as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.862, + 0.483, + 0.897 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {\\text {p r o j}} \\mathcal {L} _ {\\text {p r o j}} + \\lambda_ {\\text {a t t n}} \\mathcal {L} _ {\\text {a t t n}} \\\\ + \\lambda_ {\\mathrm {m t c n n}} \\mathcal {L} _ {\\mathrm {m t c n n}} + \\lambda_ {\\mathrm {i d}} \\mathcal {L} _ {\\mathrm {i d}}, \\tag {12} \\\\ \\end{array}\n\\]" + }, + { + "type": "code_caption", + "bbox": [ + 0.524, + 0.095, + 0.697, + 0.11 + ], + "angle": 0, + "content": "Algorithm 1: FaceShield" + }, + { + "type": "algorithm", + "bbox": [ + 0.51, + 0.112, + 0.867, + 0.188 + ], + "angle": 0, + "content": "Input: image \\(x\\), steps \\(N\\), noise clamp \\(\\epsilon\\), step size \\(\\alpha\\), MTCNN detection threshold \\(t_{\\mathrm{prob}}\\), threshold weight \\(\\beta\\), CLIP Image Projector \\(\\mathcal{P}\\), Mid-layer cross-attention variance in Stable Diffusion \\(A_{\\mathrm{var}}'\\), MTCNN P-Network \\(\\mathcal{T}\\), ArcFace \\(\\mathcal{A}\\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.189, + 0.705, + 0.203 + ], + "angle": 0, + "content": "Result: protected image \\(x_{\\mathrm{adv}}\\)" + }, + { + "type": "algorithm", + "bbox": [ + 0.52, + 0.204, + 0.833, + 0.234 + ], + "angle": 0, + "content": "1 Initialize adversarial perturbation \\(\\delta \\gets 0\\), and protected image \\(x_{\\mathrm{adv}} \\gets x\\)" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.235, + 0.662, + 0.247 + ], + "angle": 0, + "content": "2 for \\(n = 1,\\dots,N\\) do" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.248, + 0.764, + 0.264 + ], + "angle": 0, + "content": "3 \\(\\mathcal{L}_{\\mathrm{proj}}\\gets \\| \\mathcal{P}(x + \\delta) - \\mathcal{P}(x)\\| _1\\)" + }, + { + "type": "algorithm", + "bbox": [ + 0.521, + 0.264, + 0.851, + 0.308 + ], + "angle": 0, + "content": "4 \\(\\mathcal{L}_{\\mathrm{attn}}\\gets ||(\\sigma_{\\mathrm{max}} - A_{\\mathrm{var}}^{\\prime})\\odot M_{\\mathrm{var}}||_{2},\\) where \\(\\sigma_{\\mathrm{max}}\\) derived from Eq. (8), and \\(M_{\\mathrm{var}}\\) from Eq. (6)" + }, + { + "type": "algorithm", + "bbox": [ + 0.521, + 0.31, + 0.875, + 0.355 + ], + "angle": 0, + "content": "5 \\(\\mathcal{L}_{\\mathrm{mtcnn}}\\gets \\left\\| (\\mathcal{T}(x + \\delta) - p_{\\mathrm{gt}})\\odot M_{\\mathrm{prob}}\\right\\| _2,\\) where \\(p_{\\mathrm{gt}} = [t_{\\mathrm{prob}} + \\beta t_{\\mathrm{prob}} - \\beta ]^T\\) , and \\(M_{\\mathrm{prob}}\\) from Eq. (9)" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.355, + 0.756, + 0.376 + ], + "angle": 0, + "content": "6 \\(\\mathcal{L}_{\\mathrm{id}}\\gets \\frac{\\mathcal{A}(x + \\delta)\\cdot\\mathcal{A}(x)}{\\|\\mathcal{A}(x + \\delta)\\|_2\\|\\mathcal{A}(x)\\|_2} -1\\)" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.377, + 0.758, + 0.39 + ], + "angle": 0, + "content": "7 Compute the total attack loss:" + }, + { + "type": "list", + "bbox": [ + 0.521, + 0.248, + 0.875, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "equation", + "bbox": [ + 0.569, + 0.391, + 0.813, + 0.421 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {\\text {p r o j}} \\mathcal {L} _ {\\text {p r o j}} + \\lambda_ {\\text {a t t n}} \\mathcal {L} _ {\\text {a t t n}} + \\\\ \\lambda_ {\\text {m t c n n}} \\mathcal {L} _ {\\text {m t c n n}} + \\lambda_ {\\text {i d}} \\mathcal {L} _ {\\text {i d}} \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.422, + 0.774, + 0.452 + ], + "angle": 0, + "content": "8 Update adversarial perturbation: \\(\\delta \\gets \\alpha \\cdot \\mathrm{sign}(\\nabla_{x_{\\mathrm{adv}}}\\mathcal{L}_{\\mathrm{total}})\\)" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.452, + 0.745, + 0.466 + ], + "angle": 0, + "content": "9 \\(\\delta_{\\mathrm{blur}}\\gets \\mathbf{G}\\) aussianBlur(δ)" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.466, + 0.767, + 0.483 + ], + "angle": 0, + "content": "10 \\(\\delta_{\\mathrm{rgb}}^{\\prime}\\gets \\mathrm{LowPassFilter}(\\delta_{\\mathrm{blur}})\\)" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.483, + 0.699, + 0.5 + ], + "angle": 0, + "content": "11 \\(x_{\\mathrm{adv}}\\gets x_{\\mathrm{adv}} - \\delta_{\\mathrm{rgb}}^{\\prime}\\)" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.501, + 0.789, + 0.516 + ], + "angle": 0, + "content": "12 \\(x_{\\mathrm{adv}} \\gets x + \\operatorname{clip}(x_{\\mathrm{adv}} - x, -\\epsilon, \\epsilon)\\)" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.422, + 0.789, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.518, + 0.564, + 0.53 + ], + "angle": 0, + "content": "13 end" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.531, + 0.85, + 0.547 + ], + "angle": 0, + "content": "14 Clip the image range: \\( x_{\\mathrm{adv}} \\gets \\mathrm{clip}(x_{\\mathrm{adv}}, 0, 255) \\)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.561, + 0.907, + 0.638 + ], + "angle": 0, + "content": "where each \\(\\lambda\\) is a hyperparameter derived from grid searches to control the strength of the respective loss term. Additionally, the sign of \\(\\lambda\\) determines the convergence or divergence of the loss function (i.e., \\(\\lambda_{\\mathrm{proj}}\\) and \\(\\lambda_{\\mathrm{id}}\\) are negative, while \\(\\lambda_{\\mathrm{attn}}\\) and \\(\\lambda_{\\mathrm{mtcnn}}\\) are positive)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.657, + 0.734, + 0.674 + ], + "angle": 0, + "content": "3.4. Enhanced Noise Update" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.683, + 0.907, + 0.745 + ], + "angle": 0, + "content": "We integrate two additional techniques into the standard PGD to enhance robustness by enabling more imperceptible noise updates and preventing the loss of information due to purification techniques." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Gaussian blur. To enhance noise imperceptibility, we introduce a technique that constrains variations between adjacent regions, addressing the limitations of PGD methods (see Eq. (3)) that only regulate overall noise magnitude. This stems from the observation that differences between neighboring pixels can be as perceptible as the total noise itself. To achieve this, we utilize the Sobel operator [17] to emphasize areas of rapid intensity change, generating a mask \\( M_{\\mathrm{sob}} \\) that highlights image boundaries. Gaussian blur \\( \\mathcal{G}(\\cdot) \\) is then applied selectively to these regions during noise up-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10368" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.09, + 0.905, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.285, + 0.908, + 0.328 + ], + "angle": 0, + "content": "Figure 4. Qualitative Results. Protection performance across various deepfake models when our adversarial noise is applied. Models [34, 43] highlighted in the orange box typically exhibit facial distortions due to the influence described in Sec. 3.2, while those [3, 11, 20, 49] in the blue box display newly generated faces that diverge from the source image, attributed to the impact detailed in Sec. 3.3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.342, + 0.483, + 0.373 + ], + "angle": 0, + "content": "dates, ensuring smoother transitions between adjacent pixels and maintaining a consistent visual appearance:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.384, + 0.483, + 0.401 + ], + "angle": 0, + "content": "\\[\n\\delta_ {\\text {b l u r}} = \\mathcal {G} (\\delta) \\odot M _ {\\text {s o b}} + \\delta \\odot (1 - M _ {\\text {s o b}}). \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.415, + 0.485, + 0.673 + ], + "angle": 0, + "content": "Low pass filtering. To minimize information loss when saving images in JPEG format and ensure robustness to bit reduction during the compression process [10], low-frequency components are utilized. At each iteration, the newly updated adversarial noise \\(\\delta_{\\mathrm{rgb}} \\in \\mathbb{R}^{c \\times h \\times w}\\), where \\(c\\) (channel), \\(h\\) (height), and \\(w\\) (width) undergoes a padding operation and patchification according to a predefined patch size \\(p\\). Then a DCT transform [1] is applied to each patch and channel, resulting in \\(\\delta_{\\mathrm{dct}} \\in \\mathbb{R}^{c \\times h' \\times w' \\times p \\times p}\\), where \\(h' = h/p\\) and \\(w' = w/p\\), in the frequency domain. Using a low-pass filtering mask \\(M_{\\mathrm{lp}}\\), only the low-frequency components of the noise are retained. The noise \\(\\delta_{\\mathrm{rgb}}' \\in \\mathbb{R}^{c \\times h \\times w}\\) is then reconstructed back into the RGB domain through an inverse transformation. The effectiveness of this approach is demonstrated through the experimental results presented in the supplementary material, and the overall operation of FaceShield is described in Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.685, + 0.224, + 0.703 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.71, + 0.181, + 0.727 + ], + "angle": 0, + "content": "4.1. Setups" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Evaluation details. For a fair performance comparison, we use open-source baseline [23, 24, 29, 42] and apply noise to the same dataset under identical hyperparameter settings. The corresponding results are presented in Table 2, while Table 1 provides a performance comparison on diffusion-based deepfakes [20, 34, 43, 49]. The extensibility experiments on GAN-based models [3, 11] are shown in Table 4, where, in the absence of existing attack methods for these models, we validate FaceShield's effectiveness through comparisons with the original images. In cases where the feature extractor fails to detect a face, we adjust" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.342, + 0.905, + 0.387 + ], + "angle": 0, + "content": "the generation process to exclude facial features during reconstruction. Detailed descriptions and an analysis of the resources are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.39, + 0.907, + 0.527 + ], + "angle": 0, + "content": "Datasets. We evaluate our method using two datasets: CelebA-HQ [18] and VGGFace2-HQ [4], both of which have been used in previous studies [3, 11, 34]. The former is the high-resolution version of CelebA, containing 30,000 celebrity face images, while the latter is the high-resolution version of VGGFace2, consisting of 3.3 million face images from 9,131 unique identities. For our experiments, we randomly select 200 identities from each dataset, using 100 images for the source and 100 images for the target." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.536, + 0.697, + 0.552 + ], + "angle": 0, + "content": "4.2. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.561, + 0.906, + 0.667 + ], + "angle": 0, + "content": "Performance results across deepfakes. As shown in Fig.4, FaceShield demonstrates robustness across various deepfake models. The perturbations result in either (i) pronounced artifacts reflecting non-relevant facial information instead of key features [34, 43], or (ii) a complete misinterpretation of the source face, generating a new, unrelated identity [3, 11, 20, 49]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.67, + 0.906, + 0.76 + ], + "angle": 0, + "content": "Comparison with state-of-the-art methods. We compare our method with baselines on DM-based deepfake model [43]. Although the methods [23, 24, 29, 42] that achieved high performance in diffusion adversarial attacks fail to induce visible changes on the deepfake model, ours demonstrates strong protective performance (Fig.5)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.77, + 0.708, + 0.787 + ], + "angle": 0, + "content": "4.3.Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Automatic metrics. As shown in Table 1, we compare FaceShield to baseline methods across deepfake models [3, 11, 20, 34, 43, 49] using \\( L_{2} \\), Identity Score Matching [32] (ISM), and PSNR. The \\( L_{2} \\) and PSNR metrics evaluate image quality by comparing deepfake results from clean and protected images, with higher \\( L_{2} \\) and lower PSNR indicating more distortion. ISM measures the similarity between" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10369" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.102, + 0.089, + 0.905, + 0.348 + ], + "angle": 0, + "content": "
ModelDiffFace [20]DiffSwap [49]FaceSwap [34]IP-Adapter [43]
DatasetCelebA-HQ [18]
MethodL2↑ISM↓PSNR↓HE↑L2↑ISM↓PSNR↓HE↑L2↑ISM↓PSNR↓HE↑L2↑ISM↓PSNR↓HE↑
AdvDM [24]0.0210.47139.3684.220.0680.19928.3624.680.3030.24521.6154.520.2070.23525.3322.76
Mist [23]0.0210.46839.4433.940.0670.20128.3844.180.2870.23022.2634.780.1520.26528.2134.26
PhotoGuard [29]0.0220.46939.1943.820.0680.20128.2924.580.2820.23822.3164.440.1530.26828.1014.44
SDST [42]0.0210.47039.5124.080.0670.20728.3835.040.2740.26122.5824.680.1470.27328.4404.32
Ours0.0440.24332.0525.760.0720.16327.8336.200.3360.19420.7596.160.3500.07220.2666.60
Ours (Q=75)0.0430.25932.259-0.0700.16928.034-0.3170.20921.286-0.3260.11220.867-
DatasetVGGFace2-HQ [4]
MethodL2↑ISM↓PSNR↓HE↑L2↑ISM↓PSNR↓HE↑L2↑ISM↓PSNR↓HE↑L2↑ISM↓PSNR↓HE↑
AdvDM [24]0.0420.47933.0643.680.1050.21524.7694.780.4190.36118.5964.380.2510.27123.2502.36
Mist [23]0.0410.47833.2154.260.1020.22724.9643.940.3790.25919.6264.500.1810.29126.0704.10
PhotoGuard [29]0.0430.47932.9383.960.1100.21524.2724.180.3730.26619.6554.140.1800.29426.1573.82
SDST [42]0.0410.48333.2425.300.1070.22524.5064.580.3590.25819.9964.140.1660.29226.7844.06
Ours0.0620.27829.2046.100.1130.17724.0546.120.4530.23717.9196.160.3820.11219.4786.42
Ours (Q=75)0.0600.30829.435-0.1120.18524.201-0.4210.23718.573-0.3770.16719.618-
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.358, + 0.907, + 0.401 + ], + "angle": 0, + "content": "Table 1. Comparison of perturbation effectiveness among baseline methods on four deepfake models using the CelebA-HQ [18] and VGGFace2-HQ [4] datasets. Our method exhibits the largest distortion in image quality (L2, PSNR) and source similarity (ISM), as well as in human evaluation (HE). Results on JPEG-compressed images (Quality factor 75) further confirm robust protection under compression." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.413, + 0.482, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.541, + 0.483, + 0.584 + ], + "angle": 0, + "content": "Figure 5. We generate deepfake [43] results from protected images of methods [23, 24, 29, 42]. While these fail to disrupt deepfake generation, our method causes deepfakes to malfunction." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.59, + 0.483, + 0.848 + ], + "angle": 0, + "content": "the source face and the deepfake output, with lower values indicating less similarity. We conduct experiments on 100 source-target pairs from CelebA-HQ [18] and VGGFace2-HQ [4], showing that FaceShield outperforms baselines across all metrics. We also analyze the noise levels in protected images using LPIPS, PSNR, and SSIM, as shown in Table 2. These image quality metrics, compared between protected and original images, show that our method consistently produces less noise than baseline methods. Additionally, we measure the Frequency Rate (FR), which indicates that most of FaceShield's noise is concentrated in low frequencies. This property helps maintain its effectiveness under JPEG compression. To verify, we compressed the protected images to JPEG Quality 75 and tested across four deepfake models. The results show that while performance slightly decreases, FaceShield still outperforms baseline methods, as shown in Table 1, Ours \\((\\mathrm{Q} = 75)\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Human evaluation. We conduct a human evaluation (HE) on the same methods and models, using 20 source images and 100 participants recruited via Amazon Mechanical" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.411, + 0.906, + 0.591 + ], + "angle": 0, + "content": "
DatasetCelebA-HQ [18]
MethodLPIPS ↓PSNR ↑SSIM ↑FR ↑HE ↑
AdvDM [24]0.421430.44760.84382.10773.86
Mist [23]0.549229.99350.86841.65834.70
PhotoGuard [29]0.551529.91270.86691.65384.82
SDST [42]0.540931.47620.90331.67675.12
Ours0.201732.62890.939418.46515.64
DatasetVGGFace2-HQ [4]
MethodLPIPS ↓PSNR ↑SSIM ↑FR ↑HE ↑
AdvDM [24]0.410830.25230.84362.06673.66
Mist [23]0.520829.90680.87211.68724.34
PhotoGuard [29]0.522129.82040.87121.68244.62
SDST [42]0.506031.35450.90921.68924.48
Ours0.194131.57990.934118.04005.28
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.602, + 0.907, + 0.659 + ], + "angle": 0, + "content": "Table 2. With the same step size, iterations, and noise clamping values applied, our method shows the least distortion across three image quality metrics (LPIPS, SSIM, PSNR, HE) and demonstrates a higher low-frequency content (FR)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.666, + 0.907, + 0.757 + ], + "angle": 0, + "content": "Turk. Participants assess two factors: protection noise visibility (Table 2) and the similarity between the source image and its deepfake output (Table 1). We employ a Likert scale from 1 to 7. For noise visibility, a score of 7 indicates the least visible noise, while for deepfake similarity, a score of 7 reflects a significant deviation from the source identity." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.77, + 0.667, + 0.787 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Effect of gaussian blur on noise. To evaluate the effect of Gaussian blur, one of the enhanced noise update methods, we present qualitative results in the Supplementary Material comparing the blur effect's presence and absence. From this comparison, it is clear that the noise update becomes significantly more invisible at the boundaries of abrupt changes in the noise, as detected through Sobel filtering." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10370" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.228 + ], + "angle": 0, + "content": "Impact of each loss function. To demonstrate the generalizability of each loss function, we conducted an ablation study across six models using the ISM metric, as shown in Table 3. The results illustrate how FaceShield protects faces, as seen in Table 1 and Table 4, with red shading indicating performance degradation when a loss function is removed. This experiment shows that each loss function successfully impacts multiple models, and by combining them into \\(\\mathcal{L}_{\\mathrm{total}}\\), we cover a broader range of deepfakes." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.231, + 0.48, + 0.307 + ], + "angle": 0, + "content": "
ISM↓DiffFaceDiffSwapFaceSwapIP-AdapterSimSwapInfoSwap
w/o Lproj0.2410.1670.2700.1350.5440.256
w/o Lattn0.2540.1700.2230.0760.1680.252
w/o Lmtconn0.2310.1740.1660.0470.1830.354
w/o Lid0.4460.1750.2170.0400.5120.430
Ltotal0.2430.1630.1940.0720.1840.237
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.314, + 0.484, + 0.37 + ], + "angle": 0, + "content": "Table 3. Each model's performance is measured using the ISM score, confirming that each loss function ensures transferability across various deepfakes. As a result, the integrated \\(\\mathcal{L}_{\\mathrm{total}}\\) is capable of covering a wider range." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.377, + 0.484, + 0.484 + ], + "angle": 0, + "content": "Attack effectiveness of mid-layers. We qualitatively show that focusing the diffusion attack on the mid-layers of the denoising UNet [27] is more effective than applying it to the entire layers or the up/down layers, as shown in Fig.6. The experiment is conducted by applying noise \\(\\delta = 4 / 255\\) to create protected images, which are then passed through the deepfake model [43] to compare the resulting outputs." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.489, + 0.484, + 0.666 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.669, + 0.483, + 0.711 + ], + "angle": 0, + "content": "Figure 6. A comparison of conditioned face attack results across UNet [27] layers shows the best protection when targeting mid-layer cross-attention." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.902 + ], + "angle": 0, + "content": "MTCNN resize evaluation. To demonstrate the superiority of our method across various resizing modes and model transferability, we conduct an ablation study comparing it to the conventional BILINEAR scaling method [46]. We evaluate performance using different scaling methods from the OpenCV and PIL libraries, with 3,000 images from both the CelebA-HQ and VGGFace2-HQ datasets. Transferability is assessed through experiments conducted on both PyTorch and TensorFlow versions. The evaluation is based on the final bounding box detection probabilities from MTCNN [48], and the results in the supplementary material confirm that our method outperforms existing approaches." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.649, + 0.108 + ], + "angle": 0, + "content": "4.5. Applications" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.115, + 0.907, + 0.207 + ], + "angle": 0, + "content": "Extensibility on GAN-based deepfake models. We also conduct additional experiments on the GAN-based diffusion model [3, 11]. The experimental conditions are the same as those in Table 2, and the results, as shown in Table 4, indicate a degradation in model performance. Qualitative assessments are provided in Fig. 4." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.213, + 0.905, + 0.336 + ], + "angle": 0, + "content": "
ModelSimSwap [3]InfoSwap [11]
DatasetCelebA-HQ [18]
MethodL2↑ISM↓PSNR↓L2↑ISM↓PSNR↓
Original0.0000.54480.0000.0000.43180.000
Ours0.0700.18426.9210.1290.23730.220
DatasetVGGFace2-HQ [4]
MethodL2↑ISM↓PSNR↓L2↑ISM↓PSNR↓
Original0.0000.68180.0000.0000.56580.000
Ours0.0670.31427.3050.1420.35629.044
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.513, + 0.342, + 0.907, + 0.398 + ], + "angle": 0, + "content": "Table 4. Applicability of FaceShield to other deepfake frameworks. Our method, when applied to GAN-based models, not only reduces image quality \\((L_{2}\\), PSNR) but also significantly lowers source similarity (ISM)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.407, + 0.909, + 0.527 + ], + "angle": 0, + "content": "Transferability with different weights. To demonstrate that FaceShield ensures robust transferability not only across structurally different models but also to models with similar architectures but different pre-trained weights, we evaluated its performance on various versions of IP-Adapter [43]. The results, which can be found in the supplementary material, confirm the superior transferability performance of our method." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.541, + 0.634, + 0.557 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.909, + 0.809 + ], + "angle": 0, + "content": "In this study, we propose FaceShield, an invisible facial protection technique designed to attack various deepfakes. Through comparisons with multiple baseline methods, we demonstrate that FaceShield offers superior protection with significantly lower resource costs, particularly for deepfake models utilizing the latest diffusion techniques. Furthermore, its design integrates diverse transferability enhancement strategies, ensuring consistent performance not only across various pretrained versions but also across diffusion-based models with different architectures. This robustness extends to entirely different architectures, including GAN-based models. Additionally, by incorporating an improved noise update mechanism that ensures invisibility while minimizing information loss, FaceShield proves to be a practical and effective solution for preventing the misuse of facial images across a wide range of deepfake systems." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Limitations and Future Work. Although we introduce a method to enhance robustness against JPEG compression and resizing, other purification techniques still exist, which may lead to the potential loss of our protective noise information. Therefore, we plan to further strengthen the noise to effectively counter a broader range of purification methods." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10371" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.264, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.45 + ], + "angle": 0, + "content": "This work was supported by Culture, Sports and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture, Sports and Tourism (International Collaborative Research and Global Talent Development for the Development of Copyright Management and Protection Technologies for Generative AI, RS-2024-00345025, \\(47\\%\\); Research on neural watermark technology for copyright protection of generative AI 3D content, RS-2024-00348469, \\(25\\%\\)), the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT)(RS-2025-00521602, \\(25\\%\\)), Institute of Information & communications Technology Planning & Evaluation (IITP) & ITRC(Information Technology Research Center) grant funded by the Korea government(MSIT) (No.RS-2019-II190079, Artificial Intelligence Graduate School Program(Korea University), \\(1\\%\\); IITP-2025-RS-2024-00436857, \\(1\\%\\); IITP-2025-RS-2025-02304828, Artificial Intelligence Star Fellowship Support Program to Nurture the Best Talents, \\(1\\%\\)), and Artificial intelligence industrial convergence cluster development project funded by the Ministry of Science and ICT(MSIT, Korea)&Gwangju Metropolitan City." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.459, + 0.188, + 0.475 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.484, + 0.483, + 0.525 + ], + "angle": 0, + "content": "[1] N. Ahmed, T. Natarajan, and K.R. Rao. Discrete cosine transform. IEEE Transactions on Computers, C-23(1):90–93, 1974. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.526, + 0.483, + 0.58 + ], + "angle": 0, + "content": "[2] Huanran Chen, Yichi Zhang, Yinpeng Dong, Xiao Yang, Hang Su, and Jun Zhu. Rethinking model ensemble in transfer-based adversarial attacks. In The Twelfth International Conference on Learning Representations, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.582, + 0.484, + 0.65 + ], + "angle": 0, + "content": "[3] Renwang Chen, Xuanhong Chen, Bingbing Ni, and Yanhao Ge. Simswap: An efficient framework for high fidelity face swapping. In Proceedings of the 28th ACM international conference on multimedia, pages 2003-2011, 2020. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.652, + 0.483, + 0.706 + ], + "angle": 0, + "content": "[4] Xuanhong Chen, Bingbing Ni, Yutian Liu, Naiyuan Liu, Zhilin Zeng, and Hang Wang. Simswap++: Towards faster and high-quality identity swapping. IEEE Trans. Pattern Anal. Mach. Intell., 46(1):576-592, 2024. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.707, + 0.484, + 0.775 + ], + "angle": 0, + "content": "[5] Jongwook Choi, Taehoon Kim, Yonghyun Jeong, Seungryul Baek, and Jongwon Choi. Exploiting style latent flows for generalizing deepfake video detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1133-1143, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.777, + 0.483, + 0.831 + ], + "angle": 0, + "content": "[6] June Suk Choi, Kyungmin Lee, Jongheon Jeong, Saining Xie, Jinwoo Shin, and Kimin Lee. Diffusionguard: A robust defense against malicious diffusion-based image editing. arXiv preprint arXiv:2410.05694, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.832, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[7] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.484, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.907, + 0.148 + ], + "angle": 0, + "content": "[8] Junhao Dong and Xiaohua Xie. Visually maintained image disturbance against deepfake face swapping. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.149, + 0.906, + 0.217 + ], + "angle": 0, + "content": "[9] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9185-9193, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.218, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[10] Gintare Karolina Dziugaite, Zoubin Ghahramani, and Daniel M Roy. A study of the effect of jpg compression on adversarial images. arXiv preprint arXiv:1608.00853, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.273, + 0.905, + 0.342 + ], + "angle": 0, + "content": "[11] Gege Gao, Huaibo Huang, Chaoyou Fu, Zhaoyang Li, and Ran He. Information bottleneck disentanglement for identity swapping. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3404-3413, 2021. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.343, + 0.907, + 0.413 + ], + "angle": 0, + "content": "[12] Xiao Guo, Xiaohong Liu, Zhiyuan Ren, Steven Grosz, Iacopo Masi, and Xiaoming Liu. Hierarchical fine-grained image forgery detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3155-3165, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.907, + 0.483 + ], + "angle": 0, + "content": "[13] Baojin Huang, Zhongyuan Wang, Jifan Yang, Jiaxin Ai, Qin Zou, Qian Wang, and Dengpan Ye. Implicit identity driven deepfake face swapping detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4490-4499, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.484, + 0.906, + 0.566 + ], + "angle": 0, + "content": "[14] Hao Huang, Yongtao Wang, Zhaoyu Chen, Yuze Zhang, Yuheng Li, Zhi Tang, Wei Chu, Jingdong Chen, Weisi Lin, and Kai-Kuang Ma. Cmua-watermark: A cross-model universal adversarial watermark for combating deepfakes. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 989-997, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.567, + 0.906, + 0.635 + ], + "angle": 0, + "content": "[15] Qian Huang, Isay Katsman, Horace He, Zeqi Gu, Serge Belongie, and Ser-Nam Lim. Enhancing adversarial example transferability with an intermediate level attack. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4733-4742, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.636, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[16] Nathan Inkawich, Kevin Liang, Lawrence Carin, and Yiran Chen. Transferable perturbations of deep feature distributions. In International Conference on Learning Representations, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.692, + 0.905, + 0.747 + ], + "angle": 0, + "content": "[17] Nick Kanopoulos, Nagesh Vasanthavada, and Robert L Baker. Design of an image edge detection filter using the sobel operator. IEEE Journal of solid-state circuits, 23(2): 358-367, 1988. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.748, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[18] Tero Karras. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.79, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[19] Edgar Kaziakhmedov, Klim Kireev, Grigorii Melnikov, Mikhail Pautov, and Aleksandr Petiushko. Real-world attack on mtcnn face detection system. In 2019 International Multi-Conference on Engineering, Computer and Information Sciences (SIBIRCON), pages 0422-0427. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.872, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[20] Kihong Kim, Yunho Kim, Seokju Cho, Junyoung Seo, Jisu Nam, Kychul Lee, Seung Wook Kim, and Kwanghee Lee." + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10372" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Difface: Diffusion-based face swapping with facial guidance. Pattern Recognit., 163:111451, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.188 + ], + "angle": 0, + "content": "[21] Jia Li, Zhaoyang Li, Jie Cao, Xingguang Song, and Ran He. Faceinpainter: High fidelity face adaptation to heterogeneous domains. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5089-5098, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[22] Lingzhi Li, Jianmin Bao, Hao Yang, Dong Chen, and Fang Wen. Faceshifter: Towards high fidelity and occlusion aware face swapping. arXiv preprint arXiv:1912.13457, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.482, + 0.273 + ], + "angle": 0, + "content": "[23] Chumeng Liang and Xiaoyu Wu. Mist: Towards improved adversarial examples for diffusion models. arXiv preprint arXiv:2305.12683, 2023. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.343 + ], + "angle": 0, + "content": "[24] Chumeng Liang, Xiaoyu Wu, Yang Hua, Jiaru Zhang, Yiming Xue, Tao Song, Zhengui Xue, Ruhui Ma, and Haibing Guan. Adversarial example does good: Preventing painting imitation from diffusion models via adversarial examples. arXiv preprint arXiv:2302.04578, 2023. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[25] Kunlin Liu, Ivan Perov, Daiheng Gao, Nikolay Chervoniy, Wenbo Zhou, and Weiming Zhang. Deepfacelab: Integrated, flexible and extensible face-swapping framework. Pattern Recognit., 141:109628, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[27] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.636 + ], + "angle": 0, + "content": "[28] Nataniel Ruiz, Sarah Adel Bargal, and Stan Sclaroff. Disrupting deepfakes: Adversarial attacks against conditional image translation networks and facial manipulation systems. In Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 236-251. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.482, + 0.691 + ], + "angle": 0, + "content": "[29] Hadi Salman, Alaa Khaddaj, Guillaume Leclerc, Andrew Ilyas, and Aleksander Madry. Raising the cost of malicious ai-powered image editing. In International Conference on Machine Learning, 2023. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[30] Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y Zhao. Glaze: Protecting artists from style mimicry by {Text-to-Image} models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 2187–2204, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[31] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, and Yunchao Wei. Learning on gradients: Generalized artifacts representation for gan-generated images detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12105-12114, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[32] Thanh Van Le, Hao Phung, Thuan Hoang Nguyen, Quan Dao, Ngoc N Tran, and Anh Tran. Anti-dreambooth: Protecting users from personalized text-to-image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2116-2127, 2023. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.146 + ], + "angle": 0, + "content": "[33] Andrey Voynov, Qinghao Chu, Daniel Cohen-Or, and Kfir Aberman. \\(\\mathfrak{p}+\\): Extended textual conditioning in text-to-image generation. arXiv preprint arXiv:2303.09522, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[34] Feifei Wang. Face swap via diffusion model. arXiv preprint arXiv:2403.01108, 2024. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[35] Run Wang, Ziheng Huang, Zhikai Chen, Li Liu, Jing Chen, and Lina Wang. Anti-forgery: Towards a stealthy and robust deepfake disruption attack via adversarial perceptual-aware perturbations. arXiv preprint arXiv:2206.00477, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.907, + 0.289 + ], + "angle": 0, + "content": "[36] Tianyi Wang and Kam Pui Chow. Noise based deepfake detection via multi-head relative-interaction. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 14548-14556, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.293, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[37] Yuhan Wang, Xu Chen, Junwei Zhu, Wenqing Chu, Ying Tai, Chengjie Wang, Jilin Li, Yongjian Wu, Feiyue Huang, and Rongrong Ji. Hififace: 3d shape and semantic prior guided high fidelity face swapping. arXiv preprint arXiv:2106.09965, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[38] Zhibo Wang, Hengchang Guo, Zhifei Zhang, Wenxin Liu, Zhan Qin, and Kui Ren. Feature importance-aware transferable adversarial attacks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7639-7648, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[39] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L Yuille. Improving transferability of adversarial examples with input diversity. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2730-2739, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.907, + 0.56 + ], + "angle": 0, + "content": "[40] Chao Xu, Jiangning Zhang, Miao Hua, Qian He, Zili Yi, and Yong Liu. Region-aware face swapping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7632-7641, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.907, + 0.63 + ], + "angle": 0, + "content": "[41] Yangyang Xu, Bailin Deng, Junle Wang, Yanqing Jing, Jia Pan, and Shengfeng He. High-resolution face swapping via latent semantics disentanglement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7642-7651, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.905, + 0.699 + ], + "angle": 0, + "content": "[42] Haotian Xue, Chumeng Liang, Xiaoyu Wu, and Yongxin Chen. Toward effective protection against diffusion-based mimicry through score distillation. In The Twelfth International Conference on Learning Representations, 2023. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.702, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[43] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.759, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[44] Chin-Yuan Yeh, Hsi-Wen Chen, Shang-Lun Tsai, and Sheng-De Wang. Disrupting image-translation-based deepfake algorithms with adversarial attacks. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision Workshops, pages 53–62, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[45] Ge Yuan, Maomao Li, Yong Zhang, and Huicheng Zheng. Reliableswap: Boosting general face swapping via reliable supervision. arXiv preprint arXiv:2306.05356, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[46] Chongyang Zhang, Yu Qi, and Hiroyuki Kameda. Multiscale perturbation fusion adversarial attack on mtconn face" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "10373" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.124, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "detection system. In 2022 4th International Conference on Communications, Information System and Computer Engineering (CISCE), pages 142-146. IEEE, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.175 + ], + "angle": 0, + "content": "[47] Jianping Zhang, Zhuoer Xu, shiwen cui, Changhua Meng, Weibin Wu, and Michael Lyu. On the robustness of latent diffusion models, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[48] Kaipeng Zhang, Zhanpeng Zhang, Zhifeng Li, and Yu Qiao. Joint face detection and alignment using multitask cascaded convolutional networks. IEEE signal processing letters, 23 (10):1499-1503, 2016. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[49] Wenliang Zhao, Yongming Rao, Weikang Shi, Zuyan Liu, Jie Zhou, and Jiwen Lu. Diffswap: High-fidelity and controllable face swapping via 3d-aware masked diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8568-8577, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.386 + ], + "angle": 0, + "content": "[50] Wen Zhou, Xin Hou, Yongjun Chen, Mengyun Tang, Xiangqi Huang, Xiang Gan, and Yong Yang. Transferable adversarial perturbations. In Proceedings of the European Conference on Computer Vision (ECCV), pages 452-467, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.39, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[51] Yuhao Zhu, Qi Li, Jian Wang, Cheng-Zhong Xu, and Zhenan Sun. One shot face swapping on megapixels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4834-4844, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.447, + 0.482, + 0.501 + ], + "angle": 0, + "content": "[52] Yixuan Zhu, Wenliang Zhao, Yansong Tang, Yongming Rao, Jie Zhou, and Jiwen Lu. Stableswap: Stable face swapping in a shared and controllable latent space. IEEE Transactions on Multimedia, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.501 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "10374" + } + ] +] \ No newline at end of file diff --git a/2025/FaceShield_ Defending Facial Image against Deepfake Threats/4436b2e6-eb11-4750-8df1-5fa6df4f36fe_origin.pdf b/2025/FaceShield_ Defending Facial Image against Deepfake Threats/4436b2e6-eb11-4750-8df1-5fa6df4f36fe_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7453dc9ebe9a18103fc8fa49a9239b187890724f --- /dev/null +++ b/2025/FaceShield_ Defending Facial Image against Deepfake Threats/4436b2e6-eb11-4750-8df1-5fa6df4f36fe_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e003218e6cd7fec63875d74a6a72e02960068608c5d503a2a7e346888ae1f955 +size 1444358 diff --git a/2025/FaceShield_ Defending Facial Image against Deepfake Threats/full.md b/2025/FaceShield_ Defending Facial Image against Deepfake Threats/full.md new file mode 100644 index 0000000000000000000000000000000000000000..bbcfcb9c45c530ca7233a6df1b98238860960a1a --- /dev/null +++ b/2025/FaceShield_ Defending Facial Image against Deepfake Threats/full.md @@ -0,0 +1,351 @@ +# FaceShield: Defending Facial Image against Deepfake Threats + +Jaehwan Jeong $^{1}$ Sumin In $^{1}$ Sieun Kim $^{1}$ Hannie Shin $^{1}$ Jongheon Jeong $^{1}$ + +Sang Ho Yoon² Jaewook Chung³ Sangpil Kim¹* + +$^{1}$ Korea University $^{2}$ KAIST $^{3}$ Samsung Research + +# Abstract + +The rising use of deepfakes in criminal activities presents a significant issue, inciting widespread controversy. While numerous studies have tackled this problem, most primarily focus on deepfake detection. These reactive solutions are insufficient as a fundamental approach for crimes where authenticity is disregarded. Existing proactive defenses also have limitations, as they are effective only for deepfake models based on specific Generative Adversarial Networks (GANs), making them less applicable in light of recent advancements in diffusion-based models. In this paper, we propose a proactive defense method named FaceShield, which introduces novel defense strategies targeting deepfakes generated by Diffusion Models (DMs) and facilitates defenses on various existing GAN-based deepfake models through facial feature extractor manipulations. Our approach consists of three main components: (i) manipulating the attention mechanism of DMs to exclude protected facial features during the denoising process, (ii) targeting prominent facial feature extraction models to enhance the robustness of our adversarial perturbation, and (iii) employing Gaussian blur and low-pass filtering techniques to improve imperceptibility while enhancing robustness against JPEG compression. Experimental results on the CelebAHQ and VGGFace2-HQ datasets demonstrate that our method achieves state-of-the-art performance against the latest deepfake models based on DMs, while also exhibiting transferability to GANs and showcasing greater imperceptibility of noise along with enhanced robustness. + +# 1. Introduction + +The advancement of deepfake technology and improved accessibility [3, 22, 25, 37, 51] has led to significant transformations in modern society. Due to the ease of face swapping, it has been applied across various fields, providing both entertainment and convenience. However, its powerful capability to generate realistic content has also enabled malicious users to exploit it for criminal purposes, leading + +![](images/b3f7f53c35add7f73176b70fd21a5606436e5af331ccf4421a328952c3d51214.jpg) +Figure 1. Protecting Face during Deepfake using FaceShield. Pure images are vulnerable to face swapping, allowing the target image's face to be easily reflected. In contrast, images protected by FaceShield conceal facial feature from deepfake. + +to the creation of fake news and various societal problems. + +To address the growing concerns surrounding deepfake technology, various countermeasures have been explored, which can be broadly divided into two categories. The first is deepfake detection techniques [5, 12, 13, 31, 36], which act as passive defenses by classifying whether content is synthetic or authentic. While effective for authenticity verification, these offer only binary results and fail to address advanced threats, such as crimes using realistic fakes. In contrast, proactive defense strategies offer a more comprehensive solution. These approaches involve embedding imperceptible adversarial perturbation into face images to prevent the protected face from being effectively processed by deepfakes. However, most previous research [8, 14, 28, 35, 44] has concentrated on GAN-based models, often targeting individual models, which limits effectiveness against emerging DM-based deepfakes [20, 34, 43, 49]. Although significant research exists on image protection within DMs [6, 23, 24, 29, 30, 42] for image editing, the focus has primarily been on attacking the noising and denoising processes when an image is used as a query (Fig.2a). This leads to targeting the encoder or predicted noise post-UNet processing. However, we observe that such strategies are ineffective for DM-based deepfake models, where the source image influences the output in the form of key-value pairs through attention mechanisms (Fig.2b). + +![](images/9517ecb6a28033c29d026d26da5158d182f422be79ae3af76ac8a2a60250e3ae.jpg) +Figure 2. Image editing and Deepfake processes in DMs. (a) In DM-based image editing, a single image is input as a query $Q$ and edited based on a prompt condition. (b) In DM-based deepfake, two images are used, with the target image serving as the query $Q$ while the source image acts as the condition for swapping. This condition operates as key $K$ and value $V$ in the cross-attention layer. + +![](images/708d60f488b3aaa8cced6e3abc8bf0b62b3b3a4250c2b2057081f7412204dccf.jpg) + +In this paper, we focus on attacking state-of-the-art DM-based deepfakes while ensuring applicability to GAN-based models as well (Fig.1). Given the uncertainty surrounding deepfakes that malicious users might employ, we explore approaches to improve the extensibility across different architectures (e.g., GANs, DMs) and model transferability across different pre-trained backbones. Simultaneously, we propose a novel noise update method that enhances imperceptibility while being robust to JPEG compression. + +For DM-based deepfake attacks, we leverage the structural properties in which the conditioning image is embedded and integrated into the denoising UNet through attention mechanisms. By utilizing the IP-Adapter [43], commonly employed for inpainting, we extract effective adversarial noise from the embedding of the conditioning image. This perturbation effectively disrupts the propagation of the conditioning process, ensuring that the final output does not replicate the features of the protected image. + +To enhance the generalizability of our approach, we target two commonly used facial feature extractors. First, we attack the MTCNN model [48], which uses a cascade pyramid architecture to achieve superior performance and robust detection capabilities. Due to this, it is widely adopted not only in deepfake generation but also across various applications. We leverage the fact that the model scales images to different sizes during face detection. Our perturbation is designed to ensure robustness across various scaling factors and interpolation modes (e.g., BILINEAR, AREA), leading to superior performance compared to existing methods [19, 46]. Additionally, we target ArcFace [7], a widely adopted pre-trained model for facial feature extraction in deepfake applications. By incorporating both methods into our work, we ensure that our approach disrupts a range of deepfakes commonly used for facial landmark detection and feature extraction, thereby improving the overall robustness of our method against various deepfake systems. + +In the noise updating process, we refine the perturbation using two techniques: Noise Blur, which measures differences between adjacent pixels for imperceptible refinement, and Low-pass filtering, retaining low-frequency components, enhancing robustness against JPEG compression. + +To summarize, our main contributions are as follows: + +- We introduce a novel attack on deepfakes based on diffusion models. To the best of our knowledge, our proposed method is the first attempt to protect images used as conditions while demonstrating robust performance across various deepfake models by targeting common facial feature extractors. +- We propose a novel noise update mechanism that integrates Gaussian blur technique with the projected gradient descent method, significantly enhancing imperceptibility. Additionally, we implement low-pass filtering to reduce perturbation loss rates during JPEG compression compared to existing methods. +- We demonstrate that our deepfake attack method is robust across various deepfake models, outperforming previous diffusion attacks by achieving higher distortion with significantly less noise. + +# 2. Related Work + +Deepfake techniques. With advancements in generative models, deepfake technology has evolved into a specialized field focused on facial synthesis. Previous deepfake models, primarily based on GANs, generally follow a three-stage process: face detection and localization, feature extraction, and face swapping. Among these, studies such as [11, 40, 41, 45, 51] employ MTCNN [48] for face detection and landmark extraction, while the majority of deepfake models, including [3, 21, 22, 40, 51, 52], leverage ArcFace [7] for identity feature extraction. These steps are similarly employed in DM-based deepfakes that have emerged with the progress of diffusion models. Notable examples, including [20, 34, 49], integrate [7] to maintain identity consistency. However, recent work has focused on leveraging the capabilities of diffusion models to develop face-swapping methods [43] that achieve high performance without explicitly following previous approaches. + +Enhancing model transferability. In the research on adversarial attacks, various attempts have been made to improve transferability. [9, 38] proposed the model ensemble technique, generating adversarial examples using multiple + +models to enhance their effectiveness on unseen models. [39] introduced a method that selectively utilizes specific layers within a model to improve transferability. Similarly, [15, 16, 50] investigated techniques that manipulate intermediate layer feature distributions or amplify activation values to prevent adversarial noise from overfitting to a particular model. Furthermore, [2, 47] explored the use of multiple pre-trained backbones within similar model architectures to enhance transferability across different backbone networks. + +# 3. Method + +We propose a novel pipeline, FaceShield, to safeguard facial images from being exploited by diverse Deepfake methods through conditional attacks on DMs and facial feature extractor attacks. In this section, we first introduce the foundational adversarial attack framework utilized across our approach (Sec.3.1). We then detail our method for disrupting information flow when a facial image is employed as a conditioning input in DMs (Sec.3.2). Subsequently, we present our approach for preventing accurate facial feature extraction (Sec.3.3). Finally, we introduce our adversarial noise update mechanism, designed to enhance imperceptibility and mitigate degradation from JPEG compression (Sec.3.4). + +# 3.1. Preliminaries + +Cross-attention mechanism. To condition generative DMs, the cross-attention mechanism is used, as shown in Fig.2. Similar to self-attention, it involves computations using the query $Q$ , key $K$ , and value $V$ . However, unlike self-attention, where $Q$ , $K$ and $V$ are derived from the same source, cross-attention conditions the process by obtaining $Q$ from the noised image $z_{t}$ through a learned linear projection $\ell_{q}$ , while $K$ and $V$ are derived from the textual or image embedding $C_{\mathrm{emb}}$ using learned linear projection $\ell_{k}$ and $\ell_{v}$ , respectively: + +$$ +Q = \ell_ {q} (z _ {t}), K = \ell_ {k} \left(C _ {\text {e m b}}\right), V = \ell_ {v} \left(C _ {\text {e m b}}\right), \text {a n d} \tag {1} +$$ + +$$ +\operatorname {A t t e n t i o n} (Q, K, V) = \operatorname {S o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {d _ {k}}}\right) V, \tag {2} +$$ + +where $d_{k}$ are the dimensions of the key vectors. + +Projected gradient descent (PGD). PGD is a widely used method for crafting adversarial examples when the user has access to the model parameters. This technique iteratively updates an adversarial perturbation by computing the gradient of a certain loss $\mathcal{L}_{\mathrm{adv}}$ with respect to the input. At each step, noise is added in the gradient direction while keeping the perturbation within a predefined bound, ensuring the noise is small but effective: + +$$ +\delta \leftarrow \operatorname {P r o j} _ {\| \delta \| _ {\infty} \leq \eta} \left(\delta - \alpha \cdot \operatorname {s i g n} \left(\nabla_ {\delta} \mathcal {L} _ {\mathrm {a d v}}\right)\right), \tag {3} +$$ + +where $\alpha$ is the step size and $\mathrm{Proj}_{\| \delta \|_{\infty} \leq \eta}(\cdot)$ projects $\delta$ onto the $\ell_{\infty}$ ball of radius $\eta$ . By projecting the adversarial exam + +ple back onto the valid perturbation space, PGD maintains imperceptibility while disrupting the model's predictions. + +# 3.2. Conditioned Face Attack + +We now describe our approach to protecting images, specifically by disrupting the effective transfer of information when they are used as conditioning inputs in DMs. The core of our approach is to effectively interfere with key information using minimal noise, while also ensuring that the model does not overfit by accessing only a minimal number of layers to obtain gradients. To achieve this, we propose two methods that target both the initial projection phase and the final attention mechanism during the image conditioning process within latent diffusion models [27]. + +Face projector attack. When images are used as conditioning inputs, they are firstly transformed into an embedding vector through a pre-trained model [26]. In this method, we access only the topmost layer $\mathcal{P}$ of the model to disrupt the projection process, causing the image to be projected with incorrect information at the initial stage. For the attack loss function, we consider that converging to a single target value might not ensure consistent convergence speeds or balanced performance. Given that one of our main goals is to design noise applicable to various images, we design our approach to induce random divergence based on the input image, using the $\mathcal{L}_1$ loss function in this process: + +$$ +\mathcal {L} _ {\mathrm {p r o j}} (\delta ; x) = \| \mathcal {P} (x + \delta) - \mathcal {P} (x) \| _ {1}, \tag {4} +$$ + +where $\delta$ is the adversarial noise. + +Attention disruption attack. We also focus on identifying the core vectors within the denoising UNet that are most sensitive to conditional inputs. Initially, we analyze the influence of cross-attention across each UNet layer. Based on prior research [33], which shows that different cross-attention layers respond variably to conditioning information, we investigate the impact on perturbation performance for each region. Our findings lead to the conclusion that targeting attacks near mid-layers produces more significant disruption in qualitative metrics compared to using only the up-down layers or the entire layers, as supported by our experimental results in Fig.6. Based on these insights, we propose a novel approach that specifically targets mid-layers during the attack on the diffusion process. + +To induce a mismatch in conditioning, we use the mid-layer cross-attention mechanism, as described in Eq. (1). Based on the idea that the condition is conveyed to the query through attention, we calculate the attention score to obtain the strength of attention. This is done by performing operations on the query $Q \in \mathbb{R}^{h \times \mathrm{res} \times d}$ and key $K \in \mathbb{R}^{h \times \mathrm{seq} \times d}$ , where $h$ (number of heads), res (resolution), seq (sequence length), and $d$ (head dimension). This is followed by a Softmax operation along the seq dimension to derive the + +![](images/3af093bf8674b87529653d3c016f5a47231ed3cfc38cb50e5b220b904a743cdd.jpg) +Figure 3. Overview. Our method has three main parts: (i) Conditioned face attack, which disrupts feature transfer by targeting the embedding process and the attention map variance in the cross-attention layer; (ii) Facial feature extractor attack, which decreases the probability value of face detection and causes extraction disruptions, and (iii) Enhanced noise update, which improves imperceptibility by applying Gaussian blur to regions with significant intensity changes between adjacent pixels, and increases robustness against JPEG compression distortion by encoding the noise in the low-frequency domain. + +![](images/b93492d6df6d9a48bea2beacb5a6051490eeb4910fba6ddf91d9df5710568b85.jpg) + +attention map $A_{\mathrm{map}} \in \mathbb{R}^{h \times \mathrm{res} \times \mathrm{seq}}$ . Exploiting this mechanism, we obtain the variance $A_{\mathrm{var}}$ , allowing us to evaluate attention strength by the following equation: + +$$ +A _ {\text {v a r}} = \frac {1}{\text {s e q}} \sum_ {i = 1} ^ {\text {s e q}} \left(A _ {\text {m a p}} [:, i,: ] - \bar {A} _ {\text {m a p}}\right) ^ {2} \in \mathbb {R} ^ {h \times \text {r e s}}, \tag {5} +$$ + +where $\bar{A}_{\mathrm{map}} = \frac{1}{\mathrm{seq}}\sum_{i' = 1}^{\mathrm{seq}}A_{\mathrm{map}}[:,i',:]$ is the mean of attention map across the seq dimension. + +Based on them, we propose an adversarial attack strategy that maximizes $A_{\mathrm{var}}$ , thereby preventing the proper reflection of conditional information $K$ on $Q$ . In this process, we encode the original image $x$ to use as the query $Q$ and project the same $x$ to obtain the key $K$ , which is then used to calculate $A_{\mathrm{var}}$ . Thereafter, we find a quantile $P_{t_{\mathrm{var}}}$ corresponding to a predefined threshold $t_{\mathrm{var}}$ between 0 and 1 to identify the regions exhibiting weak attention. Using this $P_{t_{\mathrm{var}}}$ , we create a mask $M_{\mathrm{var}}$ such that values less than or equal to $P_{t_{\mathrm{var}}}$ are set to 1, and values greater than $P_{t_{\mathrm{var}}}$ are set to 0. This can be mathematically expressed as follows: + +$$ +M _ {\text {v a r}} = \mathbb {1} \left[ A _ {\text {v a r}} \leq P _ {t _ {\text {v a r}}} \right], \tag {6} +$$ + +where $\mathbb{1}$ is the indicator random variable. Subsequently, we derive $A_{\mathrm{var}}^{\prime}$ from the same process, using the perturbed image $x + \delta$ , and perform attention unequalization on the regions defined by $M_{\mathrm{var}}$ . This method generates missing values by assigning random attention to previously unattended regions between the original images, with the loss function + +$\mathcal{L}_{\mathrm{attn}}$ defined as follows: + +$$ +\mathcal {L} _ {\text {a t t n}} (\delta ; x, \sigma_ {\max }) = \| \left(\sigma_ {\max } - A _ {\operatorname {v a r}} ^ {\prime}\right) \odot M _ {\operatorname {v a r}} \| _ {2}, \tag {7} +$$ + +where $\odot$ is the Hadamard product, and $\sigma_{\mathrm{max}}$ denotes the maximum variance that can be obtained from the Softmax output based on the seq, following the equation: + +$$ +\sigma_ {\max } = \frac {1}{\operatorname {s e q}} \left(\left(1 - \frac {1}{\operatorname {s e q}}\right) ^ {2} + (\operatorname {s e q} - 1) \cdot \frac {1}{\operatorname {s e q} ^ {2}}\right). \tag {8} +$$ + +In the supplementary material, we provide the algorithm that outlines the method for calculating $\mathcal{L}_{\mathrm{attn}}$ + +# 3.3. Facial Feature Extractor Attack + +We design additional perturbation targeting two types of facial extraction, enhancing the applicability of our method not only to DM-based models but also to various other deep-fake architecture-based models. + +MTCNN attack. We break down the MTCNN attack we propose into three principal stages: (i) selecting the resizing scale, (ii) enhancing robustness against interpolation, and (iii) formulating a loss function to expedite convergence. Through this process, we achieve not only various resize modes but also model transferability. + +Firstly, we select a set of appropriate resizing scales $s_i \in S$ . This is to ensure that our attack technique effectively targets only the bounding boxes reaching the final layers of MTCNN. The suitable scale values are selected + +among the multi-scale factors that the MTCNN model internally uses, and the detailed workings are provided in the supplementary material. + +Using the scale factor $s_i$ selected in the previous step, we next scale the input image size $D_{\mathrm{adv}} = (h,w)$ , where $h$ and $w$ are the image's height and width, to yield $D_{\mathrm{sc1}} = (s_i \cdot h, s_i \cdot w)$ . This results in an intermediate size $D_{\mathrm{int}} = D_{\mathrm{adv}} \odot D_{\mathrm{sc1}}$ obtained through element-wise multiplication. The input image $x \in \mathbb{R}^{c \times h \times w}$ is then upscaled to $D_{\mathrm{int}}$ using NEAREST interpolation. Afterward, we downsample the image to $D_{\mathrm{sc1}}$ via average pooling, assigning equal weights to each region referenced during interpolation. This approach runs parallel with a direct BILINEAR scaling of $D_{\mathrm{adv}}$ to $D_{\mathrm{sc1}}$ , thereby ensuring robust noise generation that functions effectively across various interpolation modes. + +In the final stage, we perform a targeted attack on the initial P-Net $\mathcal{T}$ to effectively disrupt the cascade pyramid structure. We pass the downsampled adversarial noise-added image $\tilde{x}_{\mathrm{adv}} \in \mathbb{R}^{c \times s_i \cdot h \times s_i \cdot w}$ through $\mathcal{T}$ , which outputs probabilities $P_{\mathrm{T,F}}$ for bounding boxes. To expedite the convergence of the MTCNN loss function $\mathcal{L}_{\mathrm{mtcnn}}$ , we propose a masking technique that leverages both the existence probabilities $P_{\mathrm{T}}$ and the non-existence probabilities $P_{\mathrm{F}}$ . The mask $M_{\mathrm{prob}}$ is constructed to retain indices in $P_{\mathrm{T}}$ that exceed the detection threshold $t_{\mathrm{prob}}$ : + +$$ +M _ {\text {p r o b}} = \mathbb {1} \left[ P _ {\mathrm {T}} (i, j) > t _ {\text {p r o b}} \right], \tag {9} +$$ + +where $\mathbb{1}$ is the indicator random variable. Then, the $\mathcal{L}_{\mathrm{mtcnn}}$ converges with the mean squared error loss using $M_{\mathrm{prob}}$ : + +$$ +\mathcal {L} _ {\mathrm {m t c n n}} (\delta ; x, p _ {\mathrm {g t}}) = \left\| \left(\mathcal {T} (x + \delta) - p _ {\mathrm {g t}}\right) \odot M _ {\text {p r o b}} \right\| _ {2}, \tag {10} +$$ + +where $\mathcal{T}(\cdot) = [P_{\mathrm{F}}, P_{\mathrm{T}}]^T, p_{\mathrm{gt}} = [t_{\mathrm{prob}} + \beta t_{\mathrm{prob}} - \beta]^T$ , and $\beta$ is a value between 0 and 1. Additional details are provided along with the algorithm in the supplementary material. + +Identity attack. To effectively disrupt the accurate reflection of source face information, we target the ArcFace $\mathcal{A}$ models [7], which are face identity embedding models widely used in deepfake systems. To improve transferability, we ensemble the most commonly used pre-trained backbones within these models. Since $\mathcal{A}$ represents feature vectors extracted from the same person's face as vectors pointing in similar directions, we designed our approach to induce divergence from the original image $x$ by employing cosine similarity loss, thereby effectively obscuring the relevant identity information: + +$$ +\mathcal {L} _ {\mathrm {i d}} (\delta ; x) = \frac {\mathcal {A} (x + \delta) \cdot \mathcal {A} (x)}{\| \mathcal {A} (x + \delta) \| _ {2} \| \mathcal {A} (x) \| _ {2}} - 1. \tag {11} +$$ + +Overall loss operation. Accordingly, the total loss function $\mathcal{L}_{\mathrm{total}}$ is defined and used as follows: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {t o t a l}} = \lambda_ {\text {p r o j}} \mathcal {L} _ {\text {p r o j}} + \lambda_ {\text {a t t n}} \mathcal {L} _ {\text {a t t n}} \\ + \lambda_ {\mathrm {m t c n n}} \mathcal {L} _ {\mathrm {m t c n n}} + \lambda_ {\mathrm {i d}} \mathcal {L} _ {\mathrm {i d}}, \tag {12} \\ \end{array} +$$ + +Algorithm 1: FaceShield +Input: image $x$ , steps $N$ , noise clamp $\epsilon$ , step size $\alpha$ , MTCNN detection threshold $t_{\mathrm{prob}}$ , threshold weight $\beta$ , CLIP Image Projector $\mathcal{P}$ , Mid-layer cross-attention variance in Stable Diffusion $A_{\mathrm{var}}'$ , MTCNN P-Network $\mathcal{T}$ , ArcFace $\mathcal{A}$ + +Result: protected image $x_{\mathrm{adv}}$ + +1 Initialize adversarial perturbation $\delta \gets 0$ , and protected image $x_{\mathrm{adv}} \gets x$ + +2 for $n = 1,\dots,N$ do + +4 $\mathcal{L}_{\mathrm{attn}}\gets ||(\sigma_{\mathrm{max}} - A_{\mathrm{var}}^{\prime})\odot M_{\mathrm{var}}||_{2},$ where $\sigma_{\mathrm{max}}$ derived from Eq. (8), and $M_{\mathrm{var}}$ from Eq. (6) + +5 $\mathcal{L}_{\mathrm{mtcnn}}\gets \left\| (\mathcal{T}(x + \delta) - p_{\mathrm{gt}})\odot M_{\mathrm{prob}}\right\| _2,$ where $p_{\mathrm{gt}} = [t_{\mathrm{prob}} + \beta t_{\mathrm{prob}} - \beta ]^T$ , and $M_{\mathrm{prob}}$ from Eq. (9) + +3 $\mathcal{L}_{\mathrm{proj}}\gets \| \mathcal{P}(x + \delta) - \mathcal{P}(x)\| _1$ +6 $\mathcal{L}_{\mathrm{id}}\gets \frac{\mathcal{A}(x + \delta)\cdot\mathcal{A}(x)}{\|\mathcal{A}(x + \delta)\|_2\|\mathcal{A}(x)\|_2} -1$ +7 Compute the total attack loss: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {t o t a l}} = \lambda_ {\text {p r o j}} \mathcal {L} _ {\text {p r o j}} + \lambda_ {\text {a t t n}} \mathcal {L} _ {\text {a t t n}} + \\ \lambda_ {\text {m t c n n}} \mathcal {L} _ {\text {m t c n n}} + \lambda_ {\text {i d}} \mathcal {L} _ {\text {i d}} \end{array} +$$ + +8 Update adversarial perturbation: $\delta \gets \alpha \cdot \mathrm{sign}(\nabla_{x_{\mathrm{adv}}}\mathcal{L}_{\mathrm{total}})$ +9 $\delta_{\mathrm{blur}}\gets \mathbf{G}$ aussianBlur(δ) +10 $\delta_{\mathrm{rgb}}^{\prime}\gets \mathrm{LowPassFilter}(\delta_{\mathrm{blur}})$ +11 $x_{\mathrm{adv}}\gets x_{\mathrm{adv}} - \delta_{\mathrm{rgb}}^{\prime}$ +12 $x_{\mathrm{adv}} \gets x + \operatorname{clip}(x_{\mathrm{adv}} - x, -\epsilon, \epsilon)$ + +13 end + +14 Clip the image range: $x_{\mathrm{adv}} \gets \mathrm{clip}(x_{\mathrm{adv}}, 0, 255)$ + +where each $\lambda$ is a hyperparameter derived from grid searches to control the strength of the respective loss term. Additionally, the sign of $\lambda$ determines the convergence or divergence of the loss function (i.e., $\lambda_{\mathrm{proj}}$ and $\lambda_{\mathrm{id}}$ are negative, while $\lambda_{\mathrm{attn}}$ and $\lambda_{\mathrm{mtcnn}}$ are positive). + +# 3.4. Enhanced Noise Update + +We integrate two additional techniques into the standard PGD to enhance robustness by enabling more imperceptible noise updates and preventing the loss of information due to purification techniques. + +Gaussian blur. To enhance noise imperceptibility, we introduce a technique that constrains variations between adjacent regions, addressing the limitations of PGD methods (see Eq. (3)) that only regulate overall noise magnitude. This stems from the observation that differences between neighboring pixels can be as perceptible as the total noise itself. To achieve this, we utilize the Sobel operator [17] to emphasize areas of rapid intensity change, generating a mask $M_{\mathrm{sob}}$ that highlights image boundaries. Gaussian blur $\mathcal{G}(\cdot)$ is then applied selectively to these regions during noise up- + +![](images/68bbecd2ee370670e2083198557ae5165196c95048ac83f52e6d9cd858ddd4b3.jpg) +Figure 4. Qualitative Results. Protection performance across various deepfake models when our adversarial noise is applied. Models [34, 43] highlighted in the orange box typically exhibit facial distortions due to the influence described in Sec. 3.2, while those [3, 11, 20, 49] in the blue box display newly generated faces that diverge from the source image, attributed to the impact detailed in Sec. 3.3. + +dates, ensuring smoother transitions between adjacent pixels and maintaining a consistent visual appearance: + +$$ +\delta_ {\text {b l u r}} = \mathcal {G} (\delta) \odot M _ {\text {s o b}} + \delta \odot (1 - M _ {\text {s o b}}). \tag {13} +$$ + +Low pass filtering. To minimize information loss when saving images in JPEG format and ensure robustness to bit reduction during the compression process [10], low-frequency components are utilized. At each iteration, the newly updated adversarial noise $\delta_{\mathrm{rgb}} \in \mathbb{R}^{c \times h \times w}$ , where $c$ (channel), $h$ (height), and $w$ (width) undergoes a padding operation and patchification according to a predefined patch size $p$ . Then a DCT transform [1] is applied to each patch and channel, resulting in $\delta_{\mathrm{dct}} \in \mathbb{R}^{c \times h' \times w' \times p \times p}$ , where $h' = h/p$ and $w' = w/p$ , in the frequency domain. Using a low-pass filtering mask $M_{\mathrm{lp}}$ , only the low-frequency components of the noise are retained. The noise $\delta_{\mathrm{rgb}}' \in \mathbb{R}^{c \times h \times w}$ is then reconstructed back into the RGB domain through an inverse transformation. The effectiveness of this approach is demonstrated through the experimental results presented in the supplementary material, and the overall operation of FaceShield is described in Algorithm 1. + +# 4. Experiments + +# 4.1. Setups + +Evaluation details. For a fair performance comparison, we use open-source baseline [23, 24, 29, 42] and apply noise to the same dataset under identical hyperparameter settings. The corresponding results are presented in Table 2, while Table 1 provides a performance comparison on diffusion-based deepfakes [20, 34, 43, 49]. The extensibility experiments on GAN-based models [3, 11] are shown in Table 4, where, in the absence of existing attack methods for these models, we validate FaceShield's effectiveness through comparisons with the original images. In cases where the feature extractor fails to detect a face, we adjust + +the generation process to exclude facial features during reconstruction. Detailed descriptions and an analysis of the resources are provided in the supplementary material. + +Datasets. We evaluate our method using two datasets: CelebA-HQ [18] and VGGFace2-HQ [4], both of which have been used in previous studies [3, 11, 34]. The former is the high-resolution version of CelebA, containing 30,000 celebrity face images, while the latter is the high-resolution version of VGGFace2, consisting of 3.3 million face images from 9,131 unique identities. For our experiments, we randomly select 200 identities from each dataset, using 100 images for the source and 100 images for the target. + +# 4.2. Qualitative Results + +Performance results across deepfakes. As shown in Fig.4, FaceShield demonstrates robustness across various deepfake models. The perturbations result in either (i) pronounced artifacts reflecting non-relevant facial information instead of key features [34, 43], or (ii) a complete misinterpretation of the source face, generating a new, unrelated identity [3, 11, 20, 49]. + +Comparison with state-of-the-art methods. We compare our method with baselines on DM-based deepfake model [43]. Although the methods [23, 24, 29, 42] that achieved high performance in diffusion adversarial attacks fail to induce visible changes on the deepfake model, ours demonstrates strong protective performance (Fig.5). + +# 4.3.Quantitative Results + +Automatic metrics. As shown in Table 1, we compare FaceShield to baseline methods across deepfake models [3, 11, 20, 34, 43, 49] using $L_{2}$ , Identity Score Matching [32] (ISM), and PSNR. The $L_{2}$ and PSNR metrics evaluate image quality by comparing deepfake results from clean and protected images, with higher $L_{2}$ and lower PSNR indicating more distortion. ISM measures the similarity between + +
ModelDiffFace [20]DiffSwap [49]FaceSwap [34]IP-Adapter [43]
DatasetCelebA-HQ [18]
MethodL2↑ISM↓PSNR↓HE↑L2↑ISM↓PSNR↓HE↑L2↑ISM↓PSNR↓HE↑L2↑ISM↓PSNR↓HE↑
AdvDM [24]0.0210.47139.3684.220.0680.19928.3624.680.3030.24521.6154.520.2070.23525.3322.76
Mist [23]0.0210.46839.4433.940.0670.20128.3844.180.2870.23022.2634.780.1520.26528.2134.26
PhotoGuard [29]0.0220.46939.1943.820.0680.20128.2924.580.2820.23822.3164.440.1530.26828.1014.44
SDST [42]0.0210.47039.5124.080.0670.20728.3835.040.2740.26122.5824.680.1470.27328.4404.32
Ours0.0440.24332.0525.760.0720.16327.8336.200.3360.19420.7596.160.3500.07220.2666.60
Ours (Q=75)0.0430.25932.259-0.0700.16928.034-0.3170.20921.286-0.3260.11220.867-
DatasetVGGFace2-HQ [4]
MethodL2↑ISM↓PSNR↓HE↑L2↑ISM↓PSNR↓HE↑L2↑ISM↓PSNR↓HE↑L2↑ISM↓PSNR↓HE↑
AdvDM [24]0.0420.47933.0643.680.1050.21524.7694.780.4190.36118.5964.380.2510.27123.2502.36
Mist [23]0.0410.47833.2154.260.1020.22724.9643.940.3790.25919.6264.500.1810.29126.0704.10
PhotoGuard [29]0.0430.47932.9383.960.1100.21524.2724.180.3730.26619.6554.140.1800.29426.1573.82
SDST [42]0.0410.48333.2425.300.1070.22524.5064.580.3590.25819.9964.140.1660.29226.7844.06
Ours0.0620.27829.2046.100.1130.17724.0546.120.4530.23717.9196.160.3820.11219.4786.42
Ours (Q=75)0.0600.30829.435-0.1120.18524.201-0.4210.23718.573-0.3770.16719.618-
+ +![](images/45f719ce1a6e7f608513429d4db0b62680f6fd28e5810c1319a3b22d35cd8e30.jpg) +Figure 5. We generate deepfake [43] results from protected images of methods [23, 24, 29, 42]. While these fail to disrupt deepfake generation, our method causes deepfakes to malfunction. + +the source face and the deepfake output, with lower values indicating less similarity. We conduct experiments on 100 source-target pairs from CelebA-HQ [18] and VGGFace2-HQ [4], showing that FaceShield outperforms baselines across all metrics. We also analyze the noise levels in protected images using LPIPS, PSNR, and SSIM, as shown in Table 2. These image quality metrics, compared between protected and original images, show that our method consistently produces less noise than baseline methods. Additionally, we measure the Frequency Rate (FR), which indicates that most of FaceShield's noise is concentrated in low frequencies. This property helps maintain its effectiveness under JPEG compression. To verify, we compressed the protected images to JPEG Quality 75 and tested across four deepfake models. The results show that while performance slightly decreases, FaceShield still outperforms baseline methods, as shown in Table 1, Ours $(\mathrm{Q} = 75)$ . + +Human evaluation. We conduct a human evaluation (HE) on the same methods and models, using 20 source images and 100 participants recruited via Amazon Mechanical + +Table 1. Comparison of perturbation effectiveness among baseline methods on four deepfake models using the CelebA-HQ [18] and VGGFace2-HQ [4] datasets. Our method exhibits the largest distortion in image quality (L2, PSNR) and source similarity (ISM), as well as in human evaluation (HE). Results on JPEG-compressed images (Quality factor 75) further confirm robust protection under compression. + +
DatasetCelebA-HQ [18]
MethodLPIPS ↓PSNR ↑SSIM ↑FR ↑HE ↑
AdvDM [24]0.421430.44760.84382.10773.86
Mist [23]0.549229.99350.86841.65834.70
PhotoGuard [29]0.551529.91270.86691.65384.82
SDST [42]0.540931.47620.90331.67675.12
Ours0.201732.62890.939418.46515.64
DatasetVGGFace2-HQ [4]
MethodLPIPS ↓PSNR ↑SSIM ↑FR ↑HE ↑
AdvDM [24]0.410830.25230.84362.06673.66
Mist [23]0.520829.90680.87211.68724.34
PhotoGuard [29]0.522129.82040.87121.68244.62
SDST [42]0.506031.35450.90921.68924.48
Ours0.194131.57990.934118.04005.28
+ +Table 2. With the same step size, iterations, and noise clamping values applied, our method shows the least distortion across three image quality metrics (LPIPS, SSIM, PSNR, HE) and demonstrates a higher low-frequency content (FR). + +Turk. Participants assess two factors: protection noise visibility (Table 2) and the similarity between the source image and its deepfake output (Table 1). We employ a Likert scale from 1 to 7. For noise visibility, a score of 7 indicates the least visible noise, while for deepfake similarity, a score of 7 reflects a significant deviation from the source identity. + +# 4.4. Ablation Study + +Effect of gaussian blur on noise. To evaluate the effect of Gaussian blur, one of the enhanced noise update methods, we present qualitative results in the Supplementary Material comparing the blur effect's presence and absence. From this comparison, it is clear that the noise update becomes significantly more invisible at the boundaries of abrupt changes in the noise, as detected through Sobel filtering. + +Impact of each loss function. To demonstrate the generalizability of each loss function, we conducted an ablation study across six models using the ISM metric, as shown in Table 3. The results illustrate how FaceShield protects faces, as seen in Table 1 and Table 4, with red shading indicating performance degradation when a loss function is removed. This experiment shows that each loss function successfully impacts multiple models, and by combining them into $\mathcal{L}_{\mathrm{total}}$ , we cover a broader range of deepfakes. + +
ISM↓DiffFaceDiffSwapFaceSwapIP-AdapterSimSwapInfoSwap
w/o Lproj0.2410.1670.2700.1350.5440.256
w/o Lattn0.2540.1700.2230.0760.1680.252
w/o Lmtconn0.2310.1740.1660.0470.1830.354
w/o Lid0.4460.1750.2170.0400.5120.430
Ltotal0.2430.1630.1940.0720.1840.237
+ +Table 3. Each model's performance is measured using the ISM score, confirming that each loss function ensures transferability across various deepfakes. As a result, the integrated $\mathcal{L}_{\mathrm{total}}$ is capable of covering a wider range. + +Attack effectiveness of mid-layers. We qualitatively show that focusing the diffusion attack on the mid-layers of the denoising UNet [27] is more effective than applying it to the entire layers or the up/down layers, as shown in Fig.6. The experiment is conducted by applying noise $\delta = 4 / 255$ to create protected images, which are then passed through the deepfake model [43] to compare the resulting outputs. + +![](images/6cfc6f3cb61eebe9e730a7e21f05e79c5ac08de5137498d0fbc3dc69f041465c.jpg) +Figure 6. A comparison of conditioned face attack results across UNet [27] layers shows the best protection when targeting mid-layer cross-attention. + +MTCNN resize evaluation. To demonstrate the superiority of our method across various resizing modes and model transferability, we conduct an ablation study comparing it to the conventional BILINEAR scaling method [46]. We evaluate performance using different scaling methods from the OpenCV and PIL libraries, with 3,000 images from both the CelebA-HQ and VGGFace2-HQ datasets. Transferability is assessed through experiments conducted on both PyTorch and TensorFlow versions. The evaluation is based on the final bounding box detection probabilities from MTCNN [48], and the results in the supplementary material confirm that our method outperforms existing approaches. + +# 4.5. Applications + +Extensibility on GAN-based deepfake models. We also conduct additional experiments on the GAN-based diffusion model [3, 11]. The experimental conditions are the same as those in Table 2, and the results, as shown in Table 4, indicate a degradation in model performance. Qualitative assessments are provided in Fig. 4. + +
ModelSimSwap [3]InfoSwap [11]
DatasetCelebA-HQ [18]
MethodL2↑ISM↓PSNR↓L2↑ISM↓PSNR↓
Original0.0000.54480.0000.0000.43180.000
Ours0.0700.18426.9210.1290.23730.220
DatasetVGGFace2-HQ [4]
MethodL2↑ISM↓PSNR↓L2↑ISM↓PSNR↓
Original0.0000.68180.0000.0000.56580.000
Ours0.0670.31427.3050.1420.35629.044
+ +Table 4. Applicability of FaceShield to other deepfake frameworks. Our method, when applied to GAN-based models, not only reduces image quality $(L_{2}$ , PSNR) but also significantly lowers source similarity (ISM). + +Transferability with different weights. To demonstrate that FaceShield ensures robust transferability not only across structurally different models but also to models with similar architectures but different pre-trained weights, we evaluated its performance on various versions of IP-Adapter [43]. The results, which can be found in the supplementary material, confirm the superior transferability performance of our method. + +# 5. Conclusion + +In this study, we propose FaceShield, an invisible facial protection technique designed to attack various deepfakes. Through comparisons with multiple baseline methods, we demonstrate that FaceShield offers superior protection with significantly lower resource costs, particularly for deepfake models utilizing the latest diffusion techniques. Furthermore, its design integrates diverse transferability enhancement strategies, ensuring consistent performance not only across various pretrained versions but also across diffusion-based models with different architectures. This robustness extends to entirely different architectures, including GAN-based models. Additionally, by incorporating an improved noise update mechanism that ensures invisibility while minimizing information loss, FaceShield proves to be a practical and effective solution for preventing the misuse of facial images across a wide range of deepfake systems. + +Limitations and Future Work. Although we introduce a method to enhance robustness against JPEG compression and resizing, other purification techniques still exist, which may lead to the potential loss of our protective noise information. Therefore, we plan to further strengthen the noise to effectively counter a broader range of purification methods. + +# 6. Acknowledgment + +This work was supported by Culture, Sports and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture, Sports and Tourism (International Collaborative Research and Global Talent Development for the Development of Copyright Management and Protection Technologies for Generative AI, RS-2024-00345025, $47\%$ ; Research on neural watermark technology for copyright protection of generative AI 3D content, RS-2024-00348469, $25\%$ ), the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT)(RS-2025-00521602, $25\%$ ), Institute of Information & communications Technology Planning & Evaluation (IITP) & ITRC(Information Technology Research Center) grant funded by the Korea government(MSIT) (No.RS-2019-II190079, Artificial Intelligence Graduate School Program(Korea University), $1\%$ ; IITP-2025-RS-2024-00436857, $1\%$ ; IITP-2025-RS-2025-02304828, Artificial Intelligence Star Fellowship Support Program to Nurture the Best Talents, $1\%$ ), and Artificial intelligence industrial convergence cluster development project funded by the Ministry of Science and ICT(MSIT, Korea)&Gwangju Metropolitan City. + +# References + +[1] N. Ahmed, T. Natarajan, and K.R. Rao. Discrete cosine transform. IEEE Transactions on Computers, C-23(1):90–93, 1974. 6 +[2] Huanran Chen, Yichi Zhang, Yinpeng Dong, Xiao Yang, Hang Su, and Jun Zhu. Rethinking model ensemble in transfer-based adversarial attacks. In The Twelfth International Conference on Learning Representations, 2024. 3 +[3] Renwang Chen, Xuanhong Chen, Bingbing Ni, and Yanhao Ge. Simswap: An efficient framework for high fidelity face swapping. In Proceedings of the 28th ACM international conference on multimedia, pages 2003-2011, 2020. 1, 2, 6, 8 +[4] Xuanhong Chen, Bingbing Ni, Yutian Liu, Naiyuan Liu, Zhilin Zeng, and Hang Wang. Simswap++: Towards faster and high-quality identity swapping. IEEE Trans. Pattern Anal. Mach. Intell., 46(1):576-592, 2024. 6, 7, 8 +[5] Jongwook Choi, Taehoon Kim, Yonghyun Jeong, Seungryul Baek, and Jongwon Choi. Exploiting style latent flows for generalizing deepfake video detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1133-1143, 2024. 1 +[6] June Suk Choi, Kyungmin Lee, Jongheon Jeong, Saining Xie, Jinwoo Shin, and Kimin Lee. Diffusionguard: A robust defense against malicious diffusion-based image editing. arXiv preprint arXiv:2410.05694, 2024. 1 +[7] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 2, 5 + +[8] Junhao Dong and Xiaohua Xie. Visually maintained image disturbance against deepfake face swapping. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 1 +[9] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9185-9193, 2018. 2 +[10] Gintare Karolina Dziugaite, Zoubin Ghahramani, and Daniel M Roy. A study of the effect of jpg compression on adversarial images. arXiv preprint arXiv:1608.00853, 2016. 6 +[11] Gege Gao, Huaibo Huang, Chaoyou Fu, Zhaoyang Li, and Ran He. Information bottleneck disentanglement for identity swapping. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3404-3413, 2021. 2, 6, 8 +[12] Xiao Guo, Xiaohong Liu, Zhiyuan Ren, Steven Grosz, Iacopo Masi, and Xiaoming Liu. Hierarchical fine-grained image forgery detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3155-3165, 2023. 1 +[13] Baojin Huang, Zhongyuan Wang, Jifan Yang, Jiaxin Ai, Qin Zou, Qian Wang, and Dengpan Ye. Implicit identity driven deepfake face swapping detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4490-4499, 2023. 1 +[14] Hao Huang, Yongtao Wang, Zhaoyu Chen, Yuze Zhang, Yuheng Li, Zhi Tang, Wei Chu, Jingdong Chen, Weisi Lin, and Kai-Kuang Ma. Cmua-watermark: A cross-model universal adversarial watermark for combating deepfakes. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 989-997, 2022. 1 +[15] Qian Huang, Isay Katsman, Horace He, Zeqi Gu, Serge Belongie, and Ser-Nam Lim. Enhancing adversarial example transferability with an intermediate level attack. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4733-4742, 2019. 3 +[16] Nathan Inkawich, Kevin Liang, Lawrence Carin, and Yiran Chen. Transferable perturbations of deep feature distributions. In International Conference on Learning Representations, 2020. 3 +[17] Nick Kanopoulos, Nagesh Vasanthavada, and Robert L Baker. Design of an image edge detection filter using the sobel operator. IEEE Journal of solid-state circuits, 23(2): 358-367, 1988. 5 +[18] Tero Karras. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 6, 7, 8 +[19] Edgar Kaziakhmedov, Klim Kireev, Grigorii Melnikov, Mikhail Pautov, and Aleksandr Petiushko. Real-world attack on mtcnn face detection system. In 2019 International Multi-Conference on Engineering, Computer and Information Sciences (SIBIRCON), pages 0422-0427. IEEE, 2019. 2 +[20] Kihong Kim, Yunho Kim, Seokju Cho, Junyoung Seo, Jisu Nam, Kychul Lee, Seung Wook Kim, and Kwanghee Lee. + +Difface: Diffusion-based face swapping with facial guidance. Pattern Recognit., 163:111451, 2022. 1, 2, 6, 7 +[21] Jia Li, Zhaoyang Li, Jie Cao, Xingguang Song, and Ran He. Faceinpainter: High fidelity face adaptation to heterogeneous domains. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5089-5098, 2021. 2 +[22] Lingzhi Li, Jianmin Bao, Hao Yang, Dong Chen, and Fang Wen. Faceshifter: Towards high fidelity and occlusion aware face swapping. arXiv preprint arXiv:1912.13457, 2019. 1, 2 +[23] Chumeng Liang and Xiaoyu Wu. Mist: Towards improved adversarial examples for diffusion models. arXiv preprint arXiv:2305.12683, 2023. 1, 6, 7 +[24] Chumeng Liang, Xiaoyu Wu, Yang Hua, Jiaru Zhang, Yiming Xue, Tao Song, Zhengui Xue, Ruhui Ma, and Haibing Guan. Adversarial example does good: Preventing painting imitation from diffusion models via adversarial examples. arXiv preprint arXiv:2302.04578, 2023. 1, 6, 7 +[25] Kunlin Liu, Ivan Perov, Daiheng Gao, Nikolay Chervoniy, Wenbo Zhou, and Weiming Zhang. Deepfacelab: Integrated, flexible and extensible face-swapping framework. Pattern Recognit., 141:109628, 2020. 1 +[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3 +[27] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 3, 8 +[28] Nataniel Ruiz, Sarah Adel Bargal, and Stan Sclaroff. Disrupting deepfakes: Adversarial attacks against conditional image translation networks and facial manipulation systems. In Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 236-251. Springer, 2020. 1 +[29] Hadi Salman, Alaa Khaddaj, Guillaume Leclerc, Andrew Ilyas, and Aleksander Madry. Raising the cost of malicious ai-powered image editing. In International Conference on Machine Learning, 2023. 1, 6, 7 +[30] Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y Zhao. Glaze: Protecting artists from style mimicry by {Text-to-Image} models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 2187–2204, 2023. 1 +[31] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, and Yunchao Wei. Learning on gradients: Generalized artifacts representation for gan-generated images detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12105-12114, 2023. 1 +[32] Thanh Van Le, Hao Phung, Thuan Hoang Nguyen, Quan Dao, Ngoc N Tran, and Anh Tran. Anti-dreambooth: Protecting users from personalized text-to-image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2116-2127, 2023. 6 + +[33] Andrey Voynov, Qinghao Chu, Daniel Cohen-Or, and Kfir Aberman. $\mathfrak{p}+$ : Extended textual conditioning in text-to-image generation. arXiv preprint arXiv:2303.09522, 2023. 3 +[34] Feifei Wang. Face swap via diffusion model. arXiv preprint arXiv:2403.01108, 2024. 1, 2, 6, 7 +[35] Run Wang, Ziheng Huang, Zhikai Chen, Li Liu, Jing Chen, and Lina Wang. Anti-forgery: Towards a stealthy and robust deepfake disruption attack via adversarial perceptual-aware perturbations. arXiv preprint arXiv:2206.00477, 2022. 1 +[36] Tianyi Wang and Kam Pui Chow. Noise based deepfake detection via multi-head relative-interaction. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 14548-14556, 2023. 1 +[37] Yuhan Wang, Xu Chen, Junwei Zhu, Wenqing Chu, Ying Tai, Chengjie Wang, Jilin Li, Yongjian Wu, Feiyue Huang, and Rongrong Ji. Hififace: 3d shape and semantic prior guided high fidelity face swapping. arXiv preprint arXiv:2106.09965, 2021. 1 +[38] Zhibo Wang, Hengchang Guo, Zhifei Zhang, Wenxin Liu, Zhan Qin, and Kui Ren. Feature importance-aware transferable adversarial attacks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7639-7648, 2021. 2 +[39] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L Yuille. Improving transferability of adversarial examples with input diversity. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2730-2739, 2019. 3 +[40] Chao Xu, Jiangning Zhang, Miao Hua, Qian He, Zili Yi, and Yong Liu. Region-aware face swapping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7632-7641, 2022. 2 +[41] Yangyang Xu, Bailin Deng, Junle Wang, Yanqing Jing, Jia Pan, and Shengfeng He. High-resolution face swapping via latent semantics disentanglement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7642-7651, 2022. 2 +[42] Haotian Xue, Chumeng Liang, Xiaoyu Wu, and Yongxin Chen. Toward effective protection against diffusion-based mimicry through score distillation. In The Twelfth International Conference on Learning Representations, 2023. 1, 6, 7 +[43] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 1, 2, 6, 7, 8 +[44] Chin-Yuan Yeh, Hsi-Wen Chen, Shang-Lun Tsai, and Sheng-De Wang. Disrupting image-translation-based deepfake algorithms with adversarial attacks. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision Workshops, pages 53–62, 2020. 1 +[45] Ge Yuan, Maomao Li, Yong Zhang, and Huicheng Zheng. Reliableswap: Boosting general face swapping via reliable supervision. arXiv preprint arXiv:2306.05356, 2023. 2 +[46] Chongyang Zhang, Yu Qi, and Hiroyuki Kameda. Multiscale perturbation fusion adversarial attack on mtconn face + +detection system. In 2022 4th International Conference on Communications, Information System and Computer Engineering (CISCE), pages 142-146. IEEE, 2022. 2, 8 +[47] Jianping Zhang, Zhuoer Xu, shiwen cui, Changhua Meng, Weibin Wu, and Michael Lyu. On the robustness of latent diffusion models, 2024. 3 +[48] Kaipeng Zhang, Zhanpeng Zhang, Zhifeng Li, and Yu Qiao. Joint face detection and alignment using multitask cascaded convolutional networks. IEEE signal processing letters, 23 (10):1499-1503, 2016. 2, 8 +[49] Wenliang Zhao, Yongming Rao, Weikang Shi, Zuyan Liu, Jie Zhou, and Jiwen Lu. Diffswap: High-fidelity and controllable face swapping via 3d-aware masked diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8568-8577, 2023. 1, 2, 6, 7 +[50] Wen Zhou, Xin Hou, Yongjun Chen, Mengyun Tang, Xiangqi Huang, Xiang Gan, and Yong Yang. Transferable adversarial perturbations. In Proceedings of the European Conference on Computer Vision (ECCV), pages 452-467, 2018. 3 +[51] Yuhao Zhu, Qi Li, Jian Wang, Cheng-Zhong Xu, and Zhenan Sun. One shot face swapping on megapixels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4834-4844, 2021. 1, 2 +[52] Yixuan Zhu, Wenliang Zhao, Yansong Tang, Yongming Rao, Jie Zhou, and Jiwen Lu. Stableswap: Stable face swapping in a shared and controllable latent space. IEEE Transactions on Multimedia, 2024. 2 \ No newline at end of file diff --git a/2025/FaceShield_ Defending Facial Image against Deepfake Threats/images.zip b/2025/FaceShield_ Defending Facial Image against Deepfake Threats/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e14db61d01dd5ca209a1d82b85120cc11ebb1964 --- /dev/null +++ b/2025/FaceShield_ Defending Facial Image against Deepfake Threats/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3b121a8d59b0a116ee3ca3e9299b15b523867735a0434bded49c4cbfffeaaf12 +size 752371 diff --git a/2025/FaceShield_ Defending Facial Image against Deepfake Threats/layout.json b/2025/FaceShield_ Defending Facial Image against Deepfake Threats/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2c6e29805ef2204b0b8ae9bc5ef239837fc1ec08 --- /dev/null +++ b/2025/FaceShield_ Defending Facial Image against Deepfake Threats/layout.json @@ -0,0 +1,10317 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 76, + 85, + 536, + 106 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 85, + 536, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 85, + 536, + 106 + ], + "type": "text", + "content": "FaceShield: Defending Facial Image against Deepfake Threats" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 109, + 122, + 500, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 122, + 500, + 137 + ], + "spans": [ + { + "bbox": [ + 109, + 122, + 500, + 137 + ], + "type": "text", + "content": "Jaehwan Jeong" + }, + { + "bbox": [ + 109, + 122, + 500, + 137 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 109, + 122, + 500, + 137 + ], + "type": "text", + "content": " Sumin In" + }, + { + "bbox": [ + 109, + 122, + 500, + 137 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 109, + 122, + 500, + 137 + ], + "type": "text", + "content": " Sieun Kim" + }, + { + "bbox": [ + 109, + 122, + 500, + 137 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 109, + 122, + 500, + 137 + ], + "type": "text", + "content": " Hannie Shin" + }, + { + "bbox": [ + 109, + 122, + 500, + 137 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 109, + 122, + 500, + 137 + ], + "type": "text", + "content": " Jongheon Jeong" + }, + { + "bbox": [ + 109, + 122, + 500, + 137 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 178, + 137, + 436, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 137, + 436, + 152 + ], + "spans": [ + { + "bbox": [ + 178, + 137, + 436, + 152 + ], + "type": "text", + "content": "Sang Ho Yoon² Jaewook Chung³ Sangpil Kim¹*" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 178, + 164, + 433, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 164, + 433, + 180 + ], + "spans": [ + { + "bbox": [ + 178, + 164, + 433, + 180 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 178, + 164, + 433, + 180 + ], + "type": "text", + "content": "Korea University " + }, + { + "bbox": [ + 178, + 164, + 433, + 180 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 178, + 164, + 433, + 180 + ], + "type": "text", + "content": "KAIST " + }, + { + "bbox": [ + 178, + 164, + 433, + 180 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 178, + 164, + 433, + 180 + ], + "type": "text", + "content": "Samsung Research" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 216, + 200, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 216, + 200, + 228 + ], + "spans": [ + { + "bbox": [ + 151, + 216, + 200, + 228 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 241, + 297, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 241, + 297, + 577 + ], + "spans": [ + { + "bbox": [ + 54, + 241, + 297, + 577 + ], + "type": "text", + "content": "The rising use of deepfakes in criminal activities presents a significant issue, inciting widespread controversy. While numerous studies have tackled this problem, most primarily focus on deepfake detection. These reactive solutions are insufficient as a fundamental approach for crimes where authenticity is disregarded. Existing proactive defenses also have limitations, as they are effective only for deepfake models based on specific Generative Adversarial Networks (GANs), making them less applicable in light of recent advancements in diffusion-based models. In this paper, we propose a proactive defense method named FaceShield, which introduces novel defense strategies targeting deepfakes generated by Diffusion Models (DMs) and facilitates defenses on various existing GAN-based deepfake models through facial feature extractor manipulations. Our approach consists of three main components: (i) manipulating the attention mechanism of DMs to exclude protected facial features during the denoising process, (ii) targeting prominent facial feature extraction models to enhance the robustness of our adversarial perturbation, and (iii) employing Gaussian blur and low-pass filtering techniques to improve imperceptibility while enhancing robustness against JPEG compression. Experimental results on the CelebAHQ and VGGFace2-HQ datasets demonstrate that our method achieves state-of-the-art performance against the latest deepfake models based on DMs, while also exhibiting transferability to GANs and showcasing greater imperceptibility of noise along with enhanced robustness." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 590, + 135, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 135, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 135, + 602 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 610, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 610, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 610, + 296, + 696 + ], + "type": "text", + "content": "The advancement of deepfake technology and improved accessibility [3, 22, 25, 37, 51] has led to significant transformations in modern society. Due to the ease of face swapping, it has been applied across various fields, providing both entertainment and convenience. However, its powerful capability to generate realistic content has also enabled malicious users to exploit it for criminal purposes, leading" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 328, + 224, + 540, + 342 + ], + "blocks": [ + { + "bbox": [ + 328, + 224, + 540, + 342 + ], + "lines": [ + { + "bbox": [ + 328, + 224, + 540, + 342 + ], + "spans": [ + { + "bbox": [ + 328, + 224, + 540, + 342 + ], + "type": "image", + "image_path": "b3f7f53c35add7f73176b70fd21a5606436e5af331ccf4421a328952c3d51214.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 349, + 555, + 393 + ], + "lines": [ + { + "bbox": [ + 313, + 349, + 555, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 555, + 393 + ], + "type": "text", + "content": "Figure 1. Protecting Face during Deepfake using FaceShield. Pure images are vulnerable to face swapping, allowing the target image's face to be easily reflected. In contrast, images protected by FaceShield conceal facial feature from deepfake." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 402, + 552, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 552, + 414 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 552, + 414 + ], + "type": "text", + "content": "to the creation of fake news and various societal problems." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 415, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 415, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 415, + 555, + 715 + ], + "type": "text", + "content": "To address the growing concerns surrounding deepfake technology, various countermeasures have been explored, which can be broadly divided into two categories. The first is deepfake detection techniques [5, 12, 13, 31, 36], which act as passive defenses by classifying whether content is synthetic or authentic. While effective for authenticity verification, these offer only binary results and fail to address advanced threats, such as crimes using realistic fakes. In contrast, proactive defense strategies offer a more comprehensive solution. These approaches involve embedding imperceptible adversarial perturbation into face images to prevent the protected face from being effectively processed by deepfakes. However, most previous research [8, 14, 28, 35, 44] has concentrated on GAN-based models, often targeting individual models, which limits effectiveness against emerging DM-based deepfakes [20, 34, 43, 49]. Although significant research exists on image protection within DMs [6, 23, 24, 29, 30, 42] for image editing, the focus has primarily been on attacking the noising and denoising processes when an image is used as a query (Fig.2a). This leads to targeting the encoder or predicted noise post-UNet processing. However, we observe that such strategies are ineffective for DM-based deepfake models, where the source image influences the output in the form of key-value pairs through attention mechanisms (Fig.2b)." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 703, + 147, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 703, + 147, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 703, + 147, + 713 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10364" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 285, + 163 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 285, + 163 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 285, + 163 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 285, + 163 + ], + "type": "image", + "image_path": "9517ecb6a28033c29d026d26da5158d182f422be79ae3af76ac8a2a60250e3ae.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 175, + 555, + 210 + ], + "lines": [ + { + "bbox": [ + 54, + 175, + 555, + 210 + ], + "spans": [ + { + "bbox": [ + 54, + 175, + 555, + 210 + ], + "type": "text", + "content": "Figure 2. Image editing and Deepfake processes in DMs. (a) In DM-based image editing, a single image is input as a query " + }, + { + "bbox": [ + 54, + 175, + 555, + 210 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 54, + 175, + 555, + 210 + ], + "type": "text", + "content": " and edited based on a prompt condition. (b) In DM-based deepfake, two images are used, with the target image serving as the query " + }, + { + "bbox": [ + 54, + 175, + 555, + 210 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 54, + 175, + 555, + 210 + ], + "type": "text", + "content": " while the source image acts as the condition for swapping. This condition operates as key " + }, + { + "bbox": [ + 54, + 175, + 555, + 210 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 54, + 175, + 555, + 210 + ], + "type": "text", + "content": " and value " + }, + { + "bbox": [ + 54, + 175, + 555, + 210 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 54, + 175, + 555, + 210 + ], + "type": "text", + "content": " in the cross-attention layer." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 301, + 71, + 553, + 167 + ], + "blocks": [ + { + "bbox": [ + 301, + 71, + 553, + 167 + ], + "lines": [ + { + "bbox": [ + 301, + 71, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 301, + 71, + 553, + 167 + ], + "type": "image", + "image_path": "708d60f488b3aaa8cced6e3abc8bf0b62b3b3a4250c2b2057081f7412204dccf.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 219, + 294, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 219, + 294, + 327 + ], + "spans": [ + { + "bbox": [ + 54, + 219, + 294, + 327 + ], + "type": "text", + "content": "In this paper, we focus on attacking state-of-the-art DM-based deepfakes while ensuring applicability to GAN-based models as well (Fig.1). Given the uncertainty surrounding deepfakes that malicious users might employ, we explore approaches to improve the extensibility across different architectures (e.g., GANs, DMs) and model transferability across different pre-trained backbones. Simultaneously, we propose a novel noise update method that enhances imperceptibility while being robust to JPEG compression." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 328, + 295, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 328, + 295, + 435 + ], + "spans": [ + { + "bbox": [ + 54, + 328, + 295, + 435 + ], + "type": "text", + "content": "For DM-based deepfake attacks, we leverage the structural properties in which the conditioning image is embedded and integrated into the denoising UNet through attention mechanisms. By utilizing the IP-Adapter [43], commonly employed for inpainting, we extract effective adversarial noise from the embedding of the conditioning image. This perturbation effectively disrupts the propagation of the conditioning process, ensuring that the final output does not replicate the features of the protected image." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 437, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 653 + ], + "type": "text", + "content": "To enhance the generalizability of our approach, we target two commonly used facial feature extractors. First, we attack the MTCNN model [48], which uses a cascade pyramid architecture to achieve superior performance and robust detection capabilities. Due to this, it is widely adopted not only in deepfake generation but also across various applications. We leverage the fact that the model scales images to different sizes during face detection. Our perturbation is designed to ensure robustness across various scaling factors and interpolation modes (e.g., BILINEAR, AREA), leading to superior performance compared to existing methods [19, 46]. Additionally, we target ArcFace [7], a widely adopted pre-trained model for facial feature extraction in deepfake applications. By incorporating both methods into our work, we ensure that our approach disrupts a range of deepfakes commonly used for facial landmark detection and feature extraction, thereby improving the overall robustness of our method against various deepfake systems." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "text", + "content": "In the noise updating process, we refine the perturbation using two techniques: Noise Blur, which measures differences between adjacent pixels for imperceptible refinement, and Low-pass filtering, retaining low-frequency components, enhancing robustness against JPEG compression." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 324, + 220, + 541, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 220, + 541, + 231 + ], + "spans": [ + { + "bbox": [ + 324, + 220, + 541, + 231 + ], + "type": "text", + "content": "To summarize, our main contributions are as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 232, + 553, + 423 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 314, + 232, + 553, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 232, + 553, + 303 + ], + "spans": [ + { + "bbox": [ + 314, + 232, + 553, + 303 + ], + "type": "text", + "content": "- We introduce a novel attack on deepfakes based on diffusion models. To the best of our knowledge, our proposed method is the first attempt to protect images used as conditions while demonstrating robust performance across various deepfake models by targeting common facial feature extractors." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 304, + 553, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 553, + 375 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 553, + 375 + ], + "type": "text", + "content": "- We propose a novel noise update mechanism that integrates Gaussian blur technique with the projected gradient descent method, significantly enhancing imperceptibility. Additionally, we implement low-pass filtering to reduce perturbation loss rates during JPEG compression compared to existing methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 376, + 553, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 553, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 553, + 423 + ], + "type": "text", + "content": "- We demonstrate that our deepfake attack method is robust across various deepfake models, outperforming previous diffusion attacks by achieving higher distortion with significantly less noise." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 436, + 400, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 436, + 400, + 449 + ], + "spans": [ + { + "bbox": [ + 314, + 436, + 400, + 449 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 459, + 555, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 459, + 555, + 663 + ], + "spans": [ + { + "bbox": [ + 312, + 459, + 555, + 663 + ], + "type": "text", + "content": "Deepfake techniques. With advancements in generative models, deepfake technology has evolved into a specialized field focused on facial synthesis. Previous deepfake models, primarily based on GANs, generally follow a three-stage process: face detection and localization, feature extraction, and face swapping. Among these, studies such as [11, 40, 41, 45, 51] employ MTCNN [48] for face detection and landmark extraction, while the majority of deepfake models, including [3, 21, 22, 40, 51, 52], leverage ArcFace [7] for identity feature extraction. These steps are similarly employed in DM-based deepfakes that have emerged with the progress of diffusion models. Notable examples, including [20, 34, 49], integrate [7] to maintain identity consistency. However, recent work has focused on leveraging the capabilities of diffusion models to develop face-swapping methods [43] that achieve high performance without explicitly following previous approaches." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": "Enhancing model transferability. In the research on adversarial attacks, various attempts have been made to improve transferability. [9, 38] proposed the model ensemble technique, generating adversarial examples using multiple" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10365" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": "models to enhance their effectiveness on unseen models. [39] introduced a method that selectively utilizes specific layers within a model to improve transferability. Similarly, [15, 16, 50] investigated techniques that manipulate intermediate layer feature distributions or amplify activation values to prevent adversarial noise from overfitting to a particular model. Furthermore, [2, 47] explored the use of multiple pre-trained backbones within similar model architectures to enhance transferability across different backbone networks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 190, + 111, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 190, + 111, + 202 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 111, + 202 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 209, + 296, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 209, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 296, + 354 + ], + "type": "text", + "content": "We propose a novel pipeline, FaceShield, to safeguard facial images from being exploited by diverse Deepfake methods through conditional attacks on DMs and facial feature extractor attacks. In this section, we first introduce the foundational adversarial attack framework utilized across our approach (Sec.3.1). We then detail our method for disrupting information flow when a facial image is employed as a conditioning input in DMs (Sec.3.2). Subsequently, we present our approach for preventing accurate facial feature extraction (Sec.3.3). Finally, we introduce our adversarial noise update mechanism, designed to enhance imperceptibility and mitigate degradation from JPEG compression (Sec.3.4)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 359, + 141, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 359, + 141, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 359, + 141, + 371 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 380, + 296, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 296, + 499 + ], + "type": "text", + "content": "Cross-attention mechanism. To condition generative DMs, the cross-attention mechanism is used, as shown in Fig.2. Similar to self-attention, it involves computations using the query " + }, + { + "bbox": [ + 55, + 380, + 296, + 499 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 380, + 296, + 499 + ], + "type": "text", + "content": ", key " + }, + { + "bbox": [ + 55, + 380, + 296, + 499 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 380, + 296, + 499 + ], + "type": "text", + "content": ", and value " + }, + { + "bbox": [ + 55, + 380, + 296, + 499 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 380, + 296, + 499 + ], + "type": "text", + "content": ". However, unlike self-attention, where " + }, + { + "bbox": [ + 55, + 380, + 296, + 499 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 380, + 296, + 499 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 380, + 296, + 499 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 380, + 296, + 499 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 380, + 296, + 499 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 380, + 296, + 499 + ], + "type": "text", + "content": " are derived from the same source, cross-attention conditions the process by obtaining " + }, + { + "bbox": [ + 55, + 380, + 296, + 499 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 380, + 296, + 499 + ], + "type": "text", + "content": " from the noised image " + }, + { + "bbox": [ + 55, + 380, + 296, + 499 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 55, + 380, + 296, + 499 + ], + "type": "text", + "content": " through a learned linear projection " + }, + { + "bbox": [ + 55, + 380, + 296, + 499 + ], + "type": "inline_equation", + "content": "\\ell_{q}" + }, + { + "bbox": [ + 55, + 380, + 296, + 499 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 55, + 380, + 296, + 499 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 380, + 296, + 499 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 380, + 296, + 499 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 380, + 296, + 499 + ], + "type": "text", + "content": " are derived from the textual or image embedding " + }, + { + "bbox": [ + 55, + 380, + 296, + 499 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{emb}}" + }, + { + "bbox": [ + 55, + 380, + 296, + 499 + ], + "type": "text", + "content": " using learned linear projection " + }, + { + "bbox": [ + 55, + 380, + 296, + 499 + ], + "type": "inline_equation", + "content": "\\ell_{k}" + }, + { + "bbox": [ + 55, + 380, + 296, + 499 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 380, + 296, + 499 + ], + "type": "inline_equation", + "content": "\\ell_{v}" + }, + { + "bbox": [ + 55, + 380, + 296, + 499 + ], + "type": "text", + "content": ", respectively:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 506, + 295, + 519 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 506, + 295, + 519 + ], + "spans": [ + { + "bbox": [ + 67, + 506, + 295, + 519 + ], + "type": "interline_equation", + "content": "Q = \\ell_ {q} (z _ {t}), K = \\ell_ {k} \\left(C _ {\\text {e m b}}\\right), V = \\ell_ {v} \\left(C _ {\\text {e m b}}\\right), \\text {a n d} \\tag {1}", + "image_path": "d120563d571c35fcf99900e27ca87600f13f5ae4b48d2d34acb9caf9739b776b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 78, + 521, + 295, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 521, + 295, + 548 + ], + "spans": [ + { + "bbox": [ + 78, + 521, + 295, + 548 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}}\\right) V, \\tag {2}", + "image_path": "7b10548d69264580f7dc3f10a6c9f372ef0dc50f2b4d68e6c2836843c45d927a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 554, + 249, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 554, + 249, + 565 + ], + "spans": [ + { + "bbox": [ + 55, + 554, + 249, + 565 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 554, + 249, + 565 + ], + "type": "inline_equation", + "content": "d_{k}" + }, + { + "bbox": [ + 55, + 554, + 249, + 565 + ], + "type": "text", + "content": " are the dimensions of the key vectors." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 567, + 296, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 296, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 296, + 662 + ], + "type": "text", + "content": "Projected gradient descent (PGD). PGD is a widely used method for crafting adversarial examples when the user has access to the model parameters. This technique iteratively updates an adversarial perturbation by computing the gradient of a certain loss " + }, + { + "bbox": [ + 55, + 567, + 296, + 662 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{adv}}" + }, + { + "bbox": [ + 55, + 567, + 296, + 662 + ], + "type": "text", + "content": " with respect to the input. At each step, noise is added in the gradient direction while keeping the perturbation within a predefined bound, ensuring the noise is small but effective:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 85, + 670, + 295, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 670, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 85, + 670, + 295, + 685 + ], + "type": "interline_equation", + "content": "\\delta \\leftarrow \\operatorname {P r o j} _ {\\| \\delta \\| _ {\\infty} \\leq \\eta} \\left(\\delta - \\alpha \\cdot \\operatorname {s i g n} \\left(\\nabla_ {\\delta} \\mathcal {L} _ {\\mathrm {a d v}}\\right)\\right), \\tag {3}", + "image_path": "110a503849a14283a2d3cfe2e64a76aa3381b47abb04da83f3ac783730c84937.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " is the step size and " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{Proj}_{\\| \\delta \\|_{\\infty} \\leq \\eta}(\\cdot)" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " projects " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " onto the " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " ball of radius " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": ". By projecting the adversarial exam" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "ple back onto the valid perturbation space, PGD maintains imperceptibility while disrupting the model's predictions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 105, + 454, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 105, + 454, + 118 + ], + "spans": [ + { + "bbox": [ + 313, + 105, + 454, + 118 + ], + "type": "text", + "content": "3.2. Conditioned Face Attack" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 123, + 555, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 123, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 123, + 555, + 243 + ], + "type": "text", + "content": "We now describe our approach to protecting images, specifically by disrupting the effective transfer of information when they are used as conditioning inputs in DMs. The core of our approach is to effectively interfere with key information using minimal noise, while also ensuring that the model does not overfit by accessing only a minimal number of layers to obtain gradients. To achieve this, we propose two methods that target both the initial projection phase and the final attention mechanism during the image conditioning process within latent diffusion models [27]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 245, + 556, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 245, + 556, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 556, + 390 + ], + "type": "text", + "content": "Face projector attack. When images are used as conditioning inputs, they are firstly transformed into an embedding vector through a pre-trained model [26]. In this method, we access only the topmost layer " + }, + { + "bbox": [ + 313, + 245, + 556, + 390 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 245, + 556, + 390 + ], + "type": "text", + "content": " of the model to disrupt the projection process, causing the image to be projected with incorrect information at the initial stage. For the attack loss function, we consider that converging to a single target value might not ensure consistent convergence speeds or balanced performance. Given that one of our main goals is to design noise applicable to various images, we design our approach to induce random divergence based on the input image, using the " + }, + { + "bbox": [ + 313, + 245, + 556, + 390 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 313, + 245, + 556, + 390 + ], + "type": "text", + "content": " loss function in this process:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 359, + 399, + 555, + 414 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 399, + 555, + 414 + ], + "spans": [ + { + "bbox": [ + 359, + 399, + 555, + 414 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {p r o j}} (\\delta ; x) = \\| \\mathcal {P} (x + \\delta) - \\mathcal {P} (x) \\| _ {1}, \\tag {4}", + "image_path": "474e80e5ee89b2e0c67db7ae0cc5c3ce38df37be0b0ca98e7f33cae469059a50.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 423, + 444, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 444, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 444, + 434 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 423, + 444, + 434 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 423, + 444, + 434 + ], + "type": "text", + "content": " is the adversarial noise." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 437, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 605 + ], + "type": "text", + "content": "Attention disruption attack. We also focus on identifying the core vectors within the denoising UNet that are most sensitive to conditional inputs. Initially, we analyze the influence of cross-attention across each UNet layer. Based on prior research [33], which shows that different cross-attention layers respond variably to conditioning information, we investigate the impact on perturbation performance for each region. Our findings lead to the conclusion that targeting attacks near mid-layers produces more significant disruption in qualitative metrics compared to using only the up-down layers or the entire layers, as supported by our experimental results in Fig.6. Based on these insights, we propose a novel approach that specifically targets mid-layers during the attack on the diffusion process." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "type": "text", + "content": "To induce a mismatch in conditioning, we use the mid-layer cross-attention mechanism, as described in Eq. (1). Based on the idea that the condition is conveyed to the query through attention, we calculate the attention score to obtain the strength of attention. This is done by performing operations on the query " + }, + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "type": "inline_equation", + "content": "Q \\in \\mathbb{R}^{h \\times \\mathrm{res} \\times d}" + }, + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "type": "text", + "content": " and key " + }, + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "type": "inline_equation", + "content": "K \\in \\mathbb{R}^{h \\times \\mathrm{seq} \\times d}" + }, + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "type": "text", + "content": " (number of heads), res (resolution), seq (sequence length), and " + }, + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "type": "text", + "content": " (head dimension). This is followed by a Softmax operation along the seq dimension to derive the" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10366" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 274, + 297 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 274, + 297 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 274, + 297 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 274, + 297 + ], + "type": "image", + "image_path": "3af093bf8674b87529653d3c016f5a47231ed3cfc38cb50e5b220b904a743cdd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 306, + 555, + 361 + ], + "lines": [ + { + "bbox": [ + 55, + 306, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 306, + 555, + 361 + ], + "type": "text", + "content": "Figure 3. Overview. Our method has three main parts: (i) Conditioned face attack, which disrupts feature transfer by targeting the embedding process and the attention map variance in the cross-attention layer; (ii) Facial feature extractor attack, which decreases the probability value of face detection and causes extraction disruptions, and (iii) Enhanced noise update, which improves imperceptibility by applying Gaussian blur to regions with significant intensity changes between adjacent pixels, and increases robustness against JPEG compression distortion by encoding the noise in the low-frequency domain." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 276, + 72, + 550, + 297 + ], + "blocks": [ + { + "bbox": [ + 276, + 72, + 550, + 297 + ], + "lines": [ + { + "bbox": [ + 276, + 72, + 550, + 297 + ], + "spans": [ + { + "bbox": [ + 276, + 72, + 550, + 297 + ], + "type": "image", + "image_path": "b93492d6df6d9a48bea2beacb5a6051490eeb4910fba6ddf91d9df5710568b85.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 372, + 295, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 372, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 295, + 409 + ], + "type": "text", + "content": "attention map " + }, + { + "bbox": [ + 55, + 372, + 295, + 409 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{map}} \\in \\mathbb{R}^{h \\times \\mathrm{res} \\times \\mathrm{seq}}" + }, + { + "bbox": [ + 55, + 372, + 295, + 409 + ], + "type": "text", + "content": ". Exploiting this mechanism, we obtain the variance " + }, + { + "bbox": [ + 55, + 372, + 295, + 409 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{var}}" + }, + { + "bbox": [ + 55, + 372, + 295, + 409 + ], + "type": "text", + "content": ", allowing us to evaluate attention strength by the following equation:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 417, + 295, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 417, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 67, + 417, + 295, + 449 + ], + "type": "interline_equation", + "content": "A _ {\\text {v a r}} = \\frac {1}{\\text {s e q}} \\sum_ {i = 1} ^ {\\text {s e q}} \\left(A _ {\\text {m a p}} [:, i,: ] - \\bar {A} _ {\\text {m a p}}\\right) ^ {2} \\in \\mathbb {R} ^ {h \\times \\text {r e s}}, \\tag {5}", + "image_path": "31ae7bdc825da7bc4e21f709ab0867aa7b28a4fe898ca1676c5c5c44833854ef.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 459, + 295, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 459, + 295, + 484 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 295, + 484 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 459, + 295, + 484 + ], + "type": "inline_equation", + "content": "\\bar{A}_{\\mathrm{map}} = \\frac{1}{\\mathrm{seq}}\\sum_{i' = 1}^{\\mathrm{seq}}A_{\\mathrm{map}}[:,i',:]" + }, + { + "bbox": [ + 55, + 459, + 295, + 484 + ], + "type": "text", + "content": " is the mean of attention map across the seq dimension." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 485, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 295, + 616 + ], + "type": "text", + "content": "Based on them, we propose an adversarial attack strategy that maximizes " + }, + { + "bbox": [ + 55, + 485, + 295, + 616 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{var}}" + }, + { + "bbox": [ + 55, + 485, + 295, + 616 + ], + "type": "text", + "content": ", thereby preventing the proper reflection of conditional information " + }, + { + "bbox": [ + 55, + 485, + 295, + 616 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 485, + 295, + 616 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 55, + 485, + 295, + 616 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 485, + 295, + 616 + ], + "type": "text", + "content": ". In this process, we encode the original image " + }, + { + "bbox": [ + 55, + 485, + 295, + 616 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 485, + 295, + 616 + ], + "type": "text", + "content": " to use as the query " + }, + { + "bbox": [ + 55, + 485, + 295, + 616 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 485, + 295, + 616 + ], + "type": "text", + "content": " and project the same " + }, + { + "bbox": [ + 55, + 485, + 295, + 616 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 485, + 295, + 616 + ], + "type": "text", + "content": " to obtain the key " + }, + { + "bbox": [ + 55, + 485, + 295, + 616 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 485, + 295, + 616 + ], + "type": "text", + "content": ", which is then used to calculate " + }, + { + "bbox": [ + 55, + 485, + 295, + 616 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{var}}" + }, + { + "bbox": [ + 55, + 485, + 295, + 616 + ], + "type": "text", + "content": ". Thereafter, we find a quantile " + }, + { + "bbox": [ + 55, + 485, + 295, + 616 + ], + "type": "inline_equation", + "content": "P_{t_{\\mathrm{var}}}" + }, + { + "bbox": [ + 55, + 485, + 295, + 616 + ], + "type": "text", + "content": " corresponding to a predefined threshold " + }, + { + "bbox": [ + 55, + 485, + 295, + 616 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{var}}" + }, + { + "bbox": [ + 55, + 485, + 295, + 616 + ], + "type": "text", + "content": " between 0 and 1 to identify the regions exhibiting weak attention. Using this " + }, + { + "bbox": [ + 55, + 485, + 295, + 616 + ], + "type": "inline_equation", + "content": "P_{t_{\\mathrm{var}}}" + }, + { + "bbox": [ + 55, + 485, + 295, + 616 + ], + "type": "text", + "content": ", we create a mask " + }, + { + "bbox": [ + 55, + 485, + 295, + 616 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{var}}" + }, + { + "bbox": [ + 55, + 485, + 295, + 616 + ], + "type": "text", + "content": " such that values less than or equal to " + }, + { + "bbox": [ + 55, + 485, + 295, + 616 + ], + "type": "inline_equation", + "content": "P_{t_{\\mathrm{var}}}" + }, + { + "bbox": [ + 55, + 485, + 295, + 616 + ], + "type": "text", + "content": " are set to 1, and values greater than " + }, + { + "bbox": [ + 55, + 485, + 295, + 616 + ], + "type": "inline_equation", + "content": "P_{t_{\\mathrm{var}}}" + }, + { + "bbox": [ + 55, + 485, + 295, + 616 + ], + "type": "text", + "content": " are set to 0. This can be mathematically expressed as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 124, + 617, + 295, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 617, + 295, + 632 + ], + "spans": [ + { + "bbox": [ + 124, + 617, + 295, + 632 + ], + "type": "interline_equation", + "content": "M _ {\\text {v a r}} = \\mathbb {1} \\left[ A _ {\\text {v a r}} \\leq P _ {t _ {\\text {v a r}}} \\right], \\tag {6}", + "image_path": "eb64cce14ba376b1e5cd4a7c2a446820738e3e57cad8900dfee1baf4dfef3fa4.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{1}" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " is the indicator random variable. Subsequently, we derive " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{var}}^{\\prime}" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " from the same process, using the perturbed image " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "x + \\delta" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " , and perform attention unequalization on the regions defined by " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{var}}" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " . This method generates missing values by assigning random attention to previously unattended regions between the original images, with the loss function" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 373, + 417, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 373, + 417, + 384 + ], + "spans": [ + { + "bbox": [ + 315, + 373, + 417, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{attn}}" + }, + { + "bbox": [ + 315, + 373, + 417, + 384 + ], + "type": "text", + "content": " defined as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 331, + 393, + 555, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 393, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 331, + 393, + 555, + 407 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a t t n}} (\\delta ; x, \\sigma_ {\\max }) = \\| \\left(\\sigma_ {\\max } - A _ {\\operatorname {v a r}} ^ {\\prime}\\right) \\odot M _ {\\operatorname {v a r}} \\| _ {2}, \\tag {7}", + "image_path": "b8da20ed3392dc4fb2d50b5ebe4c8ad9615d22bde5d71a50ef760706ef4717e3.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 416, + 554, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 416, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 416, + 554, + 453 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 416, + 554, + 453 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 313, + 416, + 554, + 453 + ], + "type": "text", + "content": " is the Hadamard product, and " + }, + { + "bbox": [ + 313, + 416, + 554, + 453 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{max}}" + }, + { + "bbox": [ + 313, + 416, + 554, + 453 + ], + "type": "text", + "content": " denotes the maximum variance that can be obtained from the Softmax output based on the seq, following the equation:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 322, + 456, + 554, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 456, + 554, + 488 + ], + "spans": [ + { + "bbox": [ + 322, + 456, + 554, + 488 + ], + "type": "interline_equation", + "content": "\\sigma_ {\\max } = \\frac {1}{\\operatorname {s e q}} \\left(\\left(1 - \\frac {1}{\\operatorname {s e q}}\\right) ^ {2} + (\\operatorname {s e q} - 1) \\cdot \\frac {1}{\\operatorname {s e q} ^ {2}}\\right). \\tag {8}", + "image_path": "234f017c3818480e341997eeb5dc29d37f80310cda1ad1094680ae48a888a695.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 495, + 554, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 554, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 554, + 520 + ], + "type": "text", + "content": "In the supplementary material, we provide the algorithm that outlines the method for calculating " + }, + { + "bbox": [ + 313, + 495, + 554, + 520 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{attn}}" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 526, + 487, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 487, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 487, + 538 + ], + "type": "text", + "content": "3.3. Facial Feature Extractor Attack" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 544, + 554, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 554, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 554, + 590 + ], + "type": "text", + "content": "We design additional perturbation targeting two types of facial extraction, enhancing the applicability of our method not only to DM-based models but also to various other deep-fake architecture-based models." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 594, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 554, + 665 + ], + "type": "text", + "content": "MTCNN attack. We break down the MTCNN attack we propose into three principal stages: (i) selecting the resizing scale, (ii) enhancing robustness against interpolation, and (iii) formulating a loss function to expedite convergence. Through this process, we achieve not only various resize modes but also model transferability." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "content": "Firstly, we select a set of appropriate resizing scales " + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "inline_equation", + "content": "s_i \\in S" + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "content": ". This is to ensure that our attack technique effectively targets only the bounding boxes reaching the final layers of MTCNN. The suitable scale values are selected" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10367" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 107 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 107 + ], + "type": "text", + "content": "among the multi-scale factors that the MTCNN model internally uses, and the detailed workings are provided in the supplementary material." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 108, + 295, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 108, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 108, + 295, + 251 + ], + "type": "text", + "content": "Using the scale factor " + }, + { + "bbox": [ + 55, + 108, + 295, + 251 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 55, + 108, + 295, + 251 + ], + "type": "text", + "content": " selected in the previous step, we next scale the input image size " + }, + { + "bbox": [ + 55, + 108, + 295, + 251 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{adv}} = (h,w)" + }, + { + "bbox": [ + 55, + 108, + 295, + 251 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 108, + 295, + 251 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 108, + 295, + 251 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 108, + 295, + 251 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 108, + 295, + 251 + ], + "type": "text", + "content": " are the image's height and width, to yield " + }, + { + "bbox": [ + 55, + 108, + 295, + 251 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{sc1}} = (s_i \\cdot h, s_i \\cdot w)" + }, + { + "bbox": [ + 55, + 108, + 295, + 251 + ], + "type": "text", + "content": ". This results in an intermediate size " + }, + { + "bbox": [ + 55, + 108, + 295, + 251 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{int}} = D_{\\mathrm{adv}} \\odot D_{\\mathrm{sc1}}" + }, + { + "bbox": [ + 55, + 108, + 295, + 251 + ], + "type": "text", + "content": " obtained through element-wise multiplication. The input image " + }, + { + "bbox": [ + 55, + 108, + 295, + 251 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^{c \\times h \\times w}" + }, + { + "bbox": [ + 55, + 108, + 295, + 251 + ], + "type": "text", + "content": " is then upscaled to " + }, + { + "bbox": [ + 55, + 108, + 295, + 251 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{int}}" + }, + { + "bbox": [ + 55, + 108, + 295, + 251 + ], + "type": "text", + "content": " using NEAREST interpolation. Afterward, we downsample the image to " + }, + { + "bbox": [ + 55, + 108, + 295, + 251 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{sc1}}" + }, + { + "bbox": [ + 55, + 108, + 295, + 251 + ], + "type": "text", + "content": " via average pooling, assigning equal weights to each region referenced during interpolation. This approach runs parallel with a direct BILINEAR scaling of " + }, + { + "bbox": [ + 55, + 108, + 295, + 251 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{adv}}" + }, + { + "bbox": [ + 55, + 108, + 295, + 251 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 108, + 295, + 251 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{sc1}}" + }, + { + "bbox": [ + 55, + 108, + 295, + 251 + ], + "type": "text", + "content": ", thereby ensuring robust noise generation that functions effectively across various interpolation modes." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 251, + 296, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 296, + 373 + ], + "type": "text", + "content": "In the final stage, we perform a targeted attack on the initial P-Net " + }, + { + "bbox": [ + 55, + 251, + 296, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 251, + 296, + 373 + ], + "type": "text", + "content": " to effectively disrupt the cascade pyramid structure. We pass the downsampled adversarial noise-added image " + }, + { + "bbox": [ + 55, + 251, + 296, + 373 + ], + "type": "inline_equation", + "content": "\\tilde{x}_{\\mathrm{adv}} \\in \\mathbb{R}^{c \\times s_i \\cdot h \\times s_i \\cdot w}" + }, + { + "bbox": [ + 55, + 251, + 296, + 373 + ], + "type": "text", + "content": " through " + }, + { + "bbox": [ + 55, + 251, + 296, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 251, + 296, + 373 + ], + "type": "text", + "content": ", which outputs probabilities " + }, + { + "bbox": [ + 55, + 251, + 296, + 373 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{T,F}}" + }, + { + "bbox": [ + 55, + 251, + 296, + 373 + ], + "type": "text", + "content": " for bounding boxes. To expedite the convergence of the MTCNN loss function " + }, + { + "bbox": [ + 55, + 251, + 296, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{mtcnn}}" + }, + { + "bbox": [ + 55, + 251, + 296, + 373 + ], + "type": "text", + "content": ", we propose a masking technique that leverages both the existence probabilities " + }, + { + "bbox": [ + 55, + 251, + 296, + 373 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{T}}" + }, + { + "bbox": [ + 55, + 251, + 296, + 373 + ], + "type": "text", + "content": " and the non-existence probabilities " + }, + { + "bbox": [ + 55, + 251, + 296, + 373 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{F}}" + }, + { + "bbox": [ + 55, + 251, + 296, + 373 + ], + "type": "text", + "content": ". The mask " + }, + { + "bbox": [ + 55, + 251, + 296, + 373 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{prob}}" + }, + { + "bbox": [ + 55, + 251, + 296, + 373 + ], + "type": "text", + "content": " is constructed to retain indices in " + }, + { + "bbox": [ + 55, + 251, + 296, + 373 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{T}}" + }, + { + "bbox": [ + 55, + 251, + 296, + 373 + ], + "type": "text", + "content": " that exceed the detection threshold " + }, + { + "bbox": [ + 55, + 251, + 296, + 373 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{prob}}" + }, + { + "bbox": [ + 55, + 251, + 296, + 373 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 377, + 295, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 377, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 116, + 377, + 295, + 392 + ], + "type": "interline_equation", + "content": "M _ {\\text {p r o b}} = \\mathbb {1} \\left[ P _ {\\mathrm {T}} (i, j) > t _ {\\text {p r o b}} \\right], \\tag {9}", + "image_path": "4cba7fa1d868f5016dea8ae1bd15b010981996f3d34e3235632dbea057d3c336.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 396, + 295, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 295, + 421 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 295, + 421 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 396, + 295, + 421 + ], + "type": "inline_equation", + "content": "\\mathbb{1}" + }, + { + "bbox": [ + 55, + 396, + 295, + 421 + ], + "type": "text", + "content": " is the indicator random variable. Then, the " + }, + { + "bbox": [ + 55, + 396, + 295, + 421 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{mtcnn}}" + }, + { + "bbox": [ + 55, + 396, + 295, + 421 + ], + "type": "text", + "content": " converges with the mean squared error loss using " + }, + { + "bbox": [ + 55, + 396, + 295, + 421 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{prob}}" + }, + { + "bbox": [ + 55, + 396, + 295, + 421 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 63, + 426, + 295, + 440 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 426, + 295, + 440 + ], + "spans": [ + { + "bbox": [ + 63, + 426, + 295, + 440 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {m t c n n}} (\\delta ; x, p _ {\\mathrm {g t}}) = \\left\\| \\left(\\mathcal {T} (x + \\delta) - p _ {\\mathrm {g t}}\\right) \\odot M _ {\\text {p r o b}} \\right\\| _ {2}, \\tag {10}", + "image_path": "c12ed4b90f12a29a84f4a080b9825727c0ea2ecbb1abad76ff6f50fca63eb553.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 445, + 295, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 295, + 482 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 295, + 482 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 445, + 295, + 482 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(\\cdot) = [P_{\\mathrm{F}}, P_{\\mathrm{T}}]^T, p_{\\mathrm{gt}} = [t_{\\mathrm{prob}} + \\beta t_{\\mathrm{prob}} - \\beta]^T" + }, + { + "bbox": [ + 55, + 445, + 295, + 482 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 445, + 295, + 482 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 445, + 295, + 482 + ], + "type": "text", + "content": " is a value between 0 and 1. Additional details are provided along with the algorithm in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 483, + 295, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 295, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 295, + 615 + ], + "type": "text", + "content": "Identity attack. To effectively disrupt the accurate reflection of source face information, we target the ArcFace " + }, + { + "bbox": [ + 55, + 483, + 295, + 615 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 55, + 483, + 295, + 615 + ], + "type": "text", + "content": " models [7], which are face identity embedding models widely used in deepfake systems. To improve transferability, we ensemble the most commonly used pre-trained backbones within these models. Since " + }, + { + "bbox": [ + 55, + 483, + 295, + 615 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 55, + 483, + 295, + 615 + ], + "type": "text", + "content": " represents feature vectors extracted from the same person's face as vectors pointing in similar directions, we designed our approach to induce divergence from the original image " + }, + { + "bbox": [ + 55, + 483, + 295, + 615 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 483, + 295, + 615 + ], + "type": "text", + "content": " by employing cosine similarity loss, thereby effectively obscuring the relevant identity information:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 93, + 620, + 295, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 620, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 93, + 620, + 295, + 647 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {i d}} (\\delta ; x) = \\frac {\\mathcal {A} (x + \\delta) \\cdot \\mathcal {A} (x)}{\\| \\mathcal {A} (x + \\delta) \\| _ {2} \\| \\mathcal {A} (x) \\| _ {2}} - 1. \\tag {11}", + "image_path": "2a8f6e5b6e7f2eb060ad0486b20ae52dac1e977b63407494a65bb74639558f89.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 652, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 652, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 652, + 295, + 677 + ], + "type": "text", + "content": "Overall loss operation. Accordingly, the total loss function " + }, + { + "bbox": [ + 55, + 652, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{total}}" + }, + { + "bbox": [ + 55, + 652, + 295, + 677 + ], + "type": "text", + "content": " is defined and used as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 100, + 682, + 295, + 710 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 682, + 295, + 710 + ], + "spans": [ + { + "bbox": [ + 100, + 682, + 295, + 710 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {\\text {p r o j}} \\mathcal {L} _ {\\text {p r o j}} + \\lambda_ {\\text {a t t n}} \\mathcal {L} _ {\\text {a t t n}} \\\\ + \\lambda_ {\\mathrm {m t c n n}} \\mathcal {L} _ {\\mathrm {m t c n n}} + \\lambda_ {\\mathrm {i d}} \\mathcal {L} _ {\\mathrm {i d}}, \\tag {12} \\\\ \\end{array}", + "image_path": "7a2a7c4db6f9a30eac0db528612bb3f4bcc36ffa7b584c9aa973dbcbad97bd44.jpg" + } + ] + } + ], + "index": 10 + }, + { + "type": "code", + "bbox": [ + 312, + 88, + 530, + 148 + ], + "blocks": [ + { + "bbox": [ + 320, + 75, + 426, + 87 + ], + "lines": [ + { + "bbox": [ + 320, + 75, + 426, + 87 + ], + "spans": [ + { + "bbox": [ + 320, + 75, + 426, + 87 + ], + "type": "text", + "content": "Algorithm 1: FaceShield" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 312, + 88, + 530, + 148 + ], + "lines": [ + { + "bbox": [ + 312, + 88, + 530, + 148 + ], + "spans": [ + { + "bbox": [ + 312, + 88, + 530, + 148 + ], + "type": "text", + "content": "Input: image " + }, + { + "bbox": [ + 312, + 88, + 530, + 148 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 312, + 88, + 530, + 148 + ], + "type": "text", + "content": ", steps " + }, + { + "bbox": [ + 312, + 88, + 530, + 148 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 312, + 88, + 530, + 148 + ], + "type": "text", + "content": ", noise clamp " + }, + { + "bbox": [ + 312, + 88, + 530, + 148 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 312, + 88, + 530, + 148 + ], + "type": "text", + "content": ", step size " + }, + { + "bbox": [ + 312, + 88, + 530, + 148 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 312, + 88, + 530, + 148 + ], + "type": "text", + "content": ", MTCNN detection threshold " + }, + { + "bbox": [ + 312, + 88, + 530, + 148 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{prob}}" + }, + { + "bbox": [ + 312, + 88, + 530, + 148 + ], + "type": "text", + "content": ", threshold weight " + }, + { + "bbox": [ + 312, + 88, + 530, + 148 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 312, + 88, + 530, + 148 + ], + "type": "text", + "content": ", CLIP Image Projector " + }, + { + "bbox": [ + 312, + 88, + 530, + 148 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 312, + 88, + 530, + 148 + ], + "type": "text", + "content": ", Mid-layer cross-attention variance in Stable Diffusion " + }, + { + "bbox": [ + 312, + 88, + 530, + 148 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{var}}'" + }, + { + "bbox": [ + 312, + 88, + 530, + 148 + ], + "type": "text", + "content": ", MTCNN P-Network " + }, + { + "bbox": [ + 312, + 88, + 530, + 148 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 312, + 88, + 530, + 148 + ], + "type": "text", + "content": ", ArcFace " + }, + { + "bbox": [ + 312, + 88, + 530, + 148 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "code_body" + } + ], + "index": 12, + "sub_type": "algorithm" + }, + { + "bbox": [ + 312, + 149, + 431, + 160 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 149, + 431, + 160 + ], + "spans": [ + { + "bbox": [ + 312, + 149, + 431, + 160 + ], + "type": "text", + "content": "Result: protected image " + }, + { + "bbox": [ + 312, + 149, + 431, + 160 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{adv}}" + } + ] + } + ], + "index": 13 + }, + { + "type": "code", + "bbox": [ + 318, + 161, + 509, + 185 + ], + "blocks": [ + { + "bbox": [ + 318, + 161, + 509, + 185 + ], + "lines": [ + { + "bbox": [ + 318, + 161, + 509, + 185 + ], + "spans": [ + { + "bbox": [ + 318, + 161, + 509, + 185 + ], + "type": "text", + "content": "1 Initialize adversarial perturbation " + }, + { + "bbox": [ + 318, + 161, + 509, + 185 + ], + "type": "inline_equation", + "content": "\\delta \\gets 0" + }, + { + "bbox": [ + 318, + 161, + 509, + 185 + ], + "type": "text", + "content": ", and protected image " + }, + { + "bbox": [ + 318, + 161, + 509, + 185 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{adv}} \\gets x" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "code_body" + } + ], + "index": 14, + "sub_type": "algorithm" + }, + { + "bbox": [ + 318, + 186, + 405, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 186, + 405, + 195 + ], + "spans": [ + { + "bbox": [ + 318, + 186, + 405, + 195 + ], + "type": "text", + "content": "2 for " + }, + { + "bbox": [ + 318, + 186, + 405, + 195 + ], + "type": "inline_equation", + "content": "n = 1,\\dots,N" + }, + { + "bbox": [ + 318, + 186, + 405, + 195 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 15 + }, + { + "type": "code", + "bbox": [ + 318, + 209, + 520, + 243 + ], + "blocks": [ + { + "bbox": [ + 318, + 209, + 520, + 243 + ], + "lines": [ + { + "bbox": [ + 318, + 209, + 520, + 243 + ], + "spans": [ + { + "bbox": [ + 318, + 209, + 520, + 243 + ], + "type": "text", + "content": "4 " + }, + { + "bbox": [ + 318, + 209, + 520, + 243 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{attn}}\\gets ||(\\sigma_{\\mathrm{max}} - A_{\\mathrm{var}}^{\\prime})\\odot M_{\\mathrm{var}}||_{2}," + }, + { + "bbox": [ + 318, + 209, + 520, + 243 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 318, + 209, + 520, + 243 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{max}}" + }, + { + "bbox": [ + 318, + 209, + 520, + 243 + ], + "type": "text", + "content": " derived from Eq. (8), and " + }, + { + "bbox": [ + 318, + 209, + 520, + 243 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{var}}" + }, + { + "bbox": [ + 318, + 209, + 520, + 243 + ], + "type": "text", + "content": " from Eq. (6)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "code_body" + } + ], + "index": 17, + "sub_type": "algorithm" + }, + { + "type": "code", + "bbox": [ + 318, + 245, + 535, + 281 + ], + "blocks": [ + { + "bbox": [ + 318, + 245, + 535, + 281 + ], + "lines": [ + { + "bbox": [ + 318, + 245, + 535, + 281 + ], + "spans": [ + { + "bbox": [ + 318, + 245, + 535, + 281 + ], + "type": "text", + "content": "5 " + }, + { + "bbox": [ + 318, + 245, + 535, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{mtcnn}}\\gets \\left\\| (\\mathcal{T}(x + \\delta) - p_{\\mathrm{gt}})\\odot M_{\\mathrm{prob}}\\right\\| _2," + }, + { + "bbox": [ + 318, + 245, + 535, + 281 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 318, + 245, + 535, + 281 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{gt}} = [t_{\\mathrm{prob}} + \\beta t_{\\mathrm{prob}} - \\beta ]^T" + }, + { + "bbox": [ + 318, + 245, + 535, + 281 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 318, + 245, + 535, + 281 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{prob}}" + }, + { + "bbox": [ + 318, + 245, + 535, + 281 + ], + "type": "text", + "content": " from Eq. (9)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "code_body" + } + ], + "index": 18, + "sub_type": "algorithm" + }, + { + "bbox": [ + 318, + 196, + 535, + 308 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 318, + 196, + 467, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 196, + 467, + 209 + ], + "spans": [ + { + "bbox": [ + 318, + 196, + 467, + 209 + ], + "type": "text", + "content": "3 " + }, + { + "bbox": [ + 318, + 196, + 467, + 209 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{proj}}\\gets \\| \\mathcal{P}(x + \\delta) - \\mathcal{P}(x)\\| _1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 318, + 281, + 462, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 281, + 462, + 297 + ], + "spans": [ + { + "bbox": [ + 318, + 281, + 462, + 297 + ], + "type": "text", + "content": "6 " + }, + { + "bbox": [ + 318, + 281, + 462, + 297 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{id}}\\gets \\frac{\\mathcal{A}(x + \\delta)\\cdot\\mathcal{A}(x)}{\\|\\mathcal{A}(x + \\delta)\\|_2\\|\\mathcal{A}(x)\\|_2} -1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 318, + 298, + 463, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 298, + 463, + 308 + ], + "spans": [ + { + "bbox": [ + 318, + 298, + 463, + 308 + ], + "type": "text", + "content": "7 Compute the total attack loss:" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 348, + 309, + 497, + 333 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 309, + 497, + 333 + ], + "spans": [ + { + "bbox": [ + 348, + 309, + 497, + 333 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {\\text {p r o j}} \\mathcal {L} _ {\\text {p r o j}} + \\lambda_ {\\text {a t t n}} \\mathcal {L} _ {\\text {a t t n}} + \\\\ \\lambda_ {\\text {m t c n n}} \\mathcal {L} _ {\\text {m t c n n}} + \\lambda_ {\\text {i d}} \\mathcal {L} _ {\\text {i d}} \\end{array}", + "image_path": "bd7f0810716a6abdf8d831deaf5e8515e66341deadb6bfab68887d403c4e0e9c.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 334, + 482, + 408 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 318, + 334, + 473, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 334, + 473, + 357 + ], + "spans": [ + { + "bbox": [ + 318, + 334, + 473, + 357 + ], + "type": "text", + "content": "8 Update adversarial perturbation: " + }, + { + "bbox": [ + 318, + 334, + 473, + 357 + ], + "type": "inline_equation", + "content": "\\delta \\gets \\alpha \\cdot \\mathrm{sign}(\\nabla_{x_{\\mathrm{adv}}}\\mathcal{L}_{\\mathrm{total}})" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 318, + 357, + 455, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 357, + 455, + 369 + ], + "spans": [ + { + "bbox": [ + 318, + 357, + 455, + 369 + ], + "type": "text", + "content": "9 " + }, + { + "bbox": [ + 318, + 357, + 455, + 369 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathrm{blur}}\\gets \\mathbf{G}" + }, + { + "bbox": [ + 318, + 357, + 455, + 369 + ], + "type": "text", + "content": " aussianBlur(δ)" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 369, + 469, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 369, + 469, + 382 + ], + "spans": [ + { + "bbox": [ + 317, + 369, + 469, + 382 + ], + "type": "text", + "content": "10 " + }, + { + "bbox": [ + 317, + 369, + 469, + 382 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathrm{rgb}}^{\\prime}\\gets \\mathrm{LowPassFilter}(\\delta_{\\mathrm{blur}})" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 382, + 427, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 382, + 427, + 396 + ], + "spans": [ + { + "bbox": [ + 317, + 382, + 427, + 396 + ], + "type": "text", + "content": "11 " + }, + { + "bbox": [ + 317, + 382, + 427, + 396 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{adv}}\\gets x_{\\mathrm{adv}} - \\delta_{\\mathrm{rgb}}^{\\prime}" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 396, + 482, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 396, + 482, + 408 + ], + "spans": [ + { + "bbox": [ + 317, + 396, + 482, + 408 + ], + "type": "text", + "content": "12 " + }, + { + "bbox": [ + 317, + 396, + 482, + 408 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{adv}} \\gets x + \\operatorname{clip}(x_{\\mathrm{adv}} - x, -\\epsilon, \\epsilon)" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 315, + 410, + 345, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 410, + 345, + 419 + ], + "spans": [ + { + "bbox": [ + 315, + 410, + 345, + 419 + ], + "type": "text", + "content": "13 end" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 315, + 420, + 520, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 420, + 520, + 433 + ], + "spans": [ + { + "bbox": [ + 315, + 420, + 520, + 433 + ], + "type": "text", + "content": "14 Clip the image range: " + }, + { + "bbox": [ + 315, + 420, + 520, + 433 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{adv}} \\gets \\mathrm{clip}(x_{\\mathrm{adv}}, 0, 255)" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 444, + 555, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 555, + 505 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 555, + 505 + ], + "type": "text", + "content": "where each " + }, + { + "bbox": [ + 313, + 444, + 555, + 505 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 444, + 555, + 505 + ], + "type": "text", + "content": " is a hyperparameter derived from grid searches to control the strength of the respective loss term. Additionally, the sign of " + }, + { + "bbox": [ + 313, + 444, + 555, + 505 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 444, + 555, + 505 + ], + "type": "text", + "content": " determines the convergence or divergence of the loss function (i.e., " + }, + { + "bbox": [ + 313, + 444, + 555, + 505 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{proj}}" + }, + { + "bbox": [ + 313, + 444, + 555, + 505 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 444, + 555, + 505 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{id}}" + }, + { + "bbox": [ + 313, + 444, + 555, + 505 + ], + "type": "text", + "content": " are negative, while " + }, + { + "bbox": [ + 313, + 444, + 555, + 505 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{attn}}" + }, + { + "bbox": [ + 313, + 444, + 555, + 505 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 444, + 555, + 505 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{mtcnn}}" + }, + { + "bbox": [ + 313, + 444, + 555, + 505 + ], + "type": "text", + "content": " are positive)." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 314, + 520, + 449, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 520, + 449, + 533 + ], + "spans": [ + { + "bbox": [ + 314, + 520, + 449, + 533 + ], + "type": "text", + "content": "3.4. Enhanced Noise Update" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 540, + 555, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 555, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 555, + 590 + ], + "type": "text", + "content": "We integrate two additional techniques into the standard PGD to enhance robustness by enabling more imperceptible noise updates and preventing the loss of information due to purification techniques." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": "Gaussian blur. To enhance noise imperceptibility, we introduce a technique that constrains variations between adjacent regions, addressing the limitations of PGD methods (see Eq. (3)) that only regulate overall noise magnitude. This stems from the observation that differences between neighboring pixels can be as perceptible as the total noise itself. To achieve this, we utilize the Sobel operator [17] to emphasize areas of rapid intensity change, generating a mask " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{sob}}" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": " that highlights image boundaries. Gaussian blur " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{G}(\\cdot)" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": " is then applied selectively to these regions during noise up-" + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10368" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 71, + 553, + 223 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 553, + 223 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 553, + 223 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 553, + 223 + ], + "type": "image", + "image_path": "68bbecd2ee370670e2083198557ae5165196c95048ac83f52e6d9cd858ddd4b3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 225, + 555, + 259 + ], + "lines": [ + { + "bbox": [ + 55, + 225, + 555, + 259 + ], + "spans": [ + { + "bbox": [ + 55, + 225, + 555, + 259 + ], + "type": "text", + "content": "Figure 4. Qualitative Results. Protection performance across various deepfake models when our adversarial noise is applied. Models [34, 43] highlighted in the orange box typically exhibit facial distortions due to the influence described in Sec. 3.2, while those [3, 11, 20, 49] in the blue box display newly generated faces that diverge from the source image, attributed to the impact detailed in Sec. 3.3." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 270, + 295, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 295, + 295 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 295, + 295 + ], + "type": "text", + "content": "dates, ensuring smoother transitions between adjacent pixels and maintaining a consistent visual appearance:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 304, + 295, + 317 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 304, + 295, + 317 + ], + "spans": [ + { + "bbox": [ + 91, + 304, + 295, + 317 + ], + "type": "interline_equation", + "content": "\\delta_ {\\text {b l u r}} = \\mathcal {G} (\\delta) \\odot M _ {\\text {s o b}} + \\delta \\odot (1 - M _ {\\text {s o b}}). \\tag {13}", + "image_path": "7e024d45b87c666e97a2d62acd7fbc9cd613018b855962cca7c6a1867f5f955f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 328, + 296, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 328, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 54, + 328, + 296, + 533 + ], + "type": "text", + "content": "Low pass filtering. To minimize information loss when saving images in JPEG format and ensure robustness to bit reduction during the compression process [10], low-frequency components are utilized. At each iteration, the newly updated adversarial noise " + }, + { + "bbox": [ + 54, + 328, + 296, + 533 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathrm{rgb}} \\in \\mathbb{R}^{c \\times h \\times w}" + }, + { + "bbox": [ + 54, + 328, + 296, + 533 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 54, + 328, + 296, + 533 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 54, + 328, + 296, + 533 + ], + "type": "text", + "content": " (channel), " + }, + { + "bbox": [ + 54, + 328, + 296, + 533 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 54, + 328, + 296, + 533 + ], + "type": "text", + "content": " (height), and " + }, + { + "bbox": [ + 54, + 328, + 296, + 533 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 54, + 328, + 296, + 533 + ], + "type": "text", + "content": " (width) undergoes a padding operation and patchification according to a predefined patch size " + }, + { + "bbox": [ + 54, + 328, + 296, + 533 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 54, + 328, + 296, + 533 + ], + "type": "text", + "content": ". Then a DCT transform [1] is applied to each patch and channel, resulting in " + }, + { + "bbox": [ + 54, + 328, + 296, + 533 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathrm{dct}} \\in \\mathbb{R}^{c \\times h' \\times w' \\times p \\times p}" + }, + { + "bbox": [ + 54, + 328, + 296, + 533 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 54, + 328, + 296, + 533 + ], + "type": "inline_equation", + "content": "h' = h/p" + }, + { + "bbox": [ + 54, + 328, + 296, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 328, + 296, + 533 + ], + "type": "inline_equation", + "content": "w' = w/p" + }, + { + "bbox": [ + 54, + 328, + 296, + 533 + ], + "type": "text", + "content": ", in the frequency domain. Using a low-pass filtering mask " + }, + { + "bbox": [ + 54, + 328, + 296, + 533 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{lp}}" + }, + { + "bbox": [ + 54, + 328, + 296, + 533 + ], + "type": "text", + "content": ", only the low-frequency components of the noise are retained. The noise " + }, + { + "bbox": [ + 54, + 328, + 296, + 533 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathrm{rgb}}' \\in \\mathbb{R}^{c \\times h \\times w}" + }, + { + "bbox": [ + 54, + 328, + 296, + 533 + ], + "type": "text", + "content": " is then reconstructed back into the RGB domain through an inverse transformation. The effectiveness of this approach is demonstrated through the experimental results presented in the supplementary material, and the overall operation of FaceShield is described in Algorithm 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 542, + 137, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 542, + 137, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 137, + 556 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 562, + 110, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 562, + 110, + 575 + ], + "spans": [ + { + "bbox": [ + 55, + 562, + 110, + 575 + ], + "type": "text", + "content": "4.1. Setups" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "Evaluation details. For a fair performance comparison, we use open-source baseline [23, 24, 29, 42] and apply noise to the same dataset under identical hyperparameter settings. The corresponding results are presented in Table 2, while Table 1 provides a performance comparison on diffusion-based deepfakes [20, 34, 43, 49]. The extensibility experiments on GAN-based models [3, 11] are shown in Table 4, where, in the absence of existing attack methods for these models, we validate FaceShield's effectiveness through comparisons with the original images. In cases where the feature extractor fails to detect a face, we adjust" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 270, + 553, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 553, + 306 + ], + "type": "text", + "content": "the generation process to exclude facial features during reconstruction. Detailed descriptions and an analysis of the resources are provided in the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 308, + 555, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 555, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 555, + 417 + ], + "type": "text", + "content": "Datasets. We evaluate our method using two datasets: CelebA-HQ [18] and VGGFace2-HQ [4], both of which have been used in previous studies [3, 11, 34]. The former is the high-resolution version of CelebA, containing 30,000 celebrity face images, while the latter is the high-resolution version of VGGFace2, consisting of 3.3 million face images from 9,131 unique identities. For our experiments, we randomly select 200 identities from each dataset, using 100 images for the source and 100 images for the target." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 424, + 426, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 424, + 426, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 424, + 426, + 437 + ], + "type": "text", + "content": "4.2. Qualitative Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 444, + 554, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 554, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 554, + 528 + ], + "type": "text", + "content": "Performance results across deepfakes. As shown in Fig.4, FaceShield demonstrates robustness across various deepfake models. The perturbations result in either (i) pronounced artifacts reflecting non-relevant facial information instead of key features [34, 43], or (ii) a complete misinterpretation of the source face, generating a new, unrelated identity [3, 11, 20, 49]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 530, + 554, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 530, + 554, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 530, + 554, + 601 + ], + "type": "text", + "content": "Comparison with state-of-the-art methods. We compare our method with baselines on DM-based deepfake model [43]. Although the methods [23, 24, 29, 42] that achieved high performance in diffusion adversarial attacks fail to induce visible changes on the deepfake model, ours demonstrates strong protective performance (Fig.5)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 609, + 433, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 433, + 623 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 433, + 623 + ], + "type": "text", + "content": "4.3.Quantitative Results" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": "Automatic metrics. As shown in Table 1, we compare FaceShield to baseline methods across deepfake models [3, 11, 20, 34, 43, 49] using " + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": ", Identity Score Matching [32] (ISM), and PSNR. The " + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": " and PSNR metrics evaluate image quality by comparing deepfake results from clean and protected images, with higher " + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": " and lower PSNR indicating more distortion. ISM measures the similarity between" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10369" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 70, + 553, + 275 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 553, + 275 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 553, + 275 + ], + "type": "table", + "html": "
ModelDiffFace [20]DiffSwap [49]FaceSwap [34]IP-Adapter [43]
DatasetCelebA-HQ [18]
MethodL2↑ISM↓PSNR↓HE↑L2↑ISM↓PSNR↓HE↑L2↑ISM↓PSNR↓HE↑L2↑ISM↓PSNR↓HE↑
AdvDM [24]0.0210.47139.3684.220.0680.19928.3624.680.3030.24521.6154.520.2070.23525.3322.76
Mist [23]0.0210.46839.4433.940.0670.20128.3844.180.2870.23022.2634.780.1520.26528.2134.26
PhotoGuard [29]0.0220.46939.1943.820.0680.20128.2924.580.2820.23822.3164.440.1530.26828.1014.44
SDST [42]0.0210.47039.5124.080.0670.20728.3835.040.2740.26122.5824.680.1470.27328.4404.32
Ours0.0440.24332.0525.760.0720.16327.8336.200.3360.19420.7596.160.3500.07220.2666.60
Ours (Q=75)0.0430.25932.259-0.0700.16928.034-0.3170.20921.286-0.3260.11220.867-
DatasetVGGFace2-HQ [4]
MethodL2↑ISM↓PSNR↓HE↑L2↑ISM↓PSNR↓HE↑L2↑ISM↓PSNR↓HE↑L2↑ISM↓PSNR↓HE↑
AdvDM [24]0.0420.47933.0643.680.1050.21524.7694.780.4190.36118.5964.380.2510.27123.2502.36
Mist [23]0.0410.47833.2154.260.1020.22724.9643.940.3790.25919.6264.500.1810.29126.0704.10
PhotoGuard [29]0.0430.47932.9383.960.1100.21524.2724.180.3730.26619.6554.140.1800.29426.1573.82
SDST [42]0.0410.48333.2425.300.1070.22524.5064.580.3590.25819.9964.140.1660.29226.7844.06
Ours0.0620.27829.2046.100.1130.17724.0546.120.4530.23717.9196.160.3820.11219.4786.42
Ours (Q=75)0.0600.30829.435-0.1120.18524.201-0.4210.23718.573-0.3770.16719.618-
", + "image_path": "9dd095337cb007d42da9b8d5b8dcb335459245e557fc55f72cf9980b3b1b40c4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 57, + 327, + 294, + 417 + ], + "blocks": [ + { + "bbox": [ + 57, + 327, + 294, + 417 + ], + "lines": [ + { + "bbox": [ + 57, + 327, + 294, + 417 + ], + "spans": [ + { + "bbox": [ + 57, + 327, + 294, + 417 + ], + "type": "image", + "image_path": "45f719ce1a6e7f608513429d4db0b62680f6fd28e5810c1319a3b22d35cd8e30.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 428, + 295, + 462 + ], + "lines": [ + { + "bbox": [ + 55, + 428, + 295, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 428, + 295, + 462 + ], + "type": "text", + "content": "Figure 5. We generate deepfake [43] results from protected images of methods [23, 24, 29, 42]. While these fail to disrupt deepfake generation, our method causes deepfakes to malfunction." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 467, + 295, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 467, + 295, + 671 + ], + "spans": [ + { + "bbox": [ + 54, + 467, + 295, + 671 + ], + "type": "text", + "content": "the source face and the deepfake output, with lower values indicating less similarity. We conduct experiments on 100 source-target pairs from CelebA-HQ [18] and VGGFace2-HQ [4], showing that FaceShield outperforms baselines across all metrics. We also analyze the noise levels in protected images using LPIPS, PSNR, and SSIM, as shown in Table 2. These image quality metrics, compared between protected and original images, show that our method consistently produces less noise than baseline methods. Additionally, we measure the Frequency Rate (FR), which indicates that most of FaceShield's noise is concentrated in low frequencies. This property helps maintain its effectiveness under JPEG compression. To verify, we compressed the protected images to JPEG Quality 75 and tested across four deepfake models. The results show that while performance slightly decreases, FaceShield still outperforms baseline methods, as shown in Table 1, Ours " + }, + { + "bbox": [ + 54, + 467, + 295, + 671 + ], + "type": "inline_equation", + "content": "(\\mathrm{Q} = 75)" + }, + { + "bbox": [ + 54, + 467, + 295, + 671 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Human evaluation. We conduct a human evaluation (HE) on the same methods and models, using 20 source images and 100 participants recruited via Amazon Mechanical" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 317, + 325, + 554, + 468 + ], + "blocks": [ + { + "bbox": [ + 55, + 283, + 555, + 317 + ], + "lines": [ + { + "bbox": [ + 55, + 283, + 555, + 317 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 555, + 317 + ], + "type": "text", + "content": "Table 1. Comparison of perturbation effectiveness among baseline methods on four deepfake models using the CelebA-HQ [18] and VGGFace2-HQ [4] datasets. Our method exhibits the largest distortion in image quality (L2, PSNR) and source similarity (ISM), as well as in human evaluation (HE). Results on JPEG-compressed images (Quality factor 75) further confirm robust protection under compression." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 325, + 554, + 468 + ], + "lines": [ + { + "bbox": [ + 317, + 325, + 554, + 468 + ], + "spans": [ + { + "bbox": [ + 317, + 325, + 554, + 468 + ], + "type": "table", + "html": "
DatasetCelebA-HQ [18]
MethodLPIPS ↓PSNR ↑SSIM ↑FR ↑HE ↑
AdvDM [24]0.421430.44760.84382.10773.86
Mist [23]0.549229.99350.86841.65834.70
PhotoGuard [29]0.551529.91270.86691.65384.82
SDST [42]0.540931.47620.90331.67675.12
Ours0.201732.62890.939418.46515.64
DatasetVGGFace2-HQ [4]
MethodLPIPS ↓PSNR ↑SSIM ↑FR ↑HE ↑
AdvDM [24]0.410830.25230.84362.06673.66
Mist [23]0.520829.90680.87211.68724.34
PhotoGuard [29]0.522129.82040.87121.68244.62
SDST [42]0.506031.35450.90921.68924.48
Ours0.194131.57990.934118.04005.28
", + "image_path": "233f4091ceed9970ee450aa6b627e1db6155d0c205fedab74b42836e09372ba2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 476, + 555, + 521 + ], + "lines": [ + { + "bbox": [ + 313, + 476, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 476, + 555, + 521 + ], + "type": "text", + "content": "Table 2. With the same step size, iterations, and noise clamping values applied, our method shows the least distortion across three image quality metrics (LPIPS, SSIM, PSNR, HE) and demonstrates a higher low-frequency content (FR)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 527, + 555, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 527, + 555, + 599 + ], + "spans": [ + { + "bbox": [ + 313, + 527, + 555, + 599 + ], + "type": "text", + "content": "Turk. Participants assess two factors: protection noise visibility (Table 2) and the similarity between the source image and its deepfake output (Table 1). We employ a Likert scale from 1 to 7. For noise visibility, a score of 7 indicates the least visible noise, while for deepfake similarity, a score of 7 reflects a significant deviation from the source identity." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 609, + 408, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 408, + 623 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 408, + 623 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": "Effect of gaussian blur on noise. To evaluate the effect of Gaussian blur, one of the enhanced noise update methods, we present qualitative results in the Supplementary Material comparing the blur effect's presence and absence. From this comparison, it is clear that the noise update becomes significantly more invisible at the boundaries of abrupt changes in the noise, as detected through Sobel filtering." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10370" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": "Impact of each loss function. To demonstrate the generalizability of each loss function, we conducted an ablation study across six models using the ISM metric, as shown in Table 3. The results illustrate how FaceShield protects faces, as seen in Table 1 and Table 4, with red shading indicating performance degradation when a loss function is removed. This experiment shows that each loss function successfully impacts multiple models, and by combining them into " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{total}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": ", we cover a broader range of deepfakes." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 61, + 182, + 293, + 243 + ], + "blocks": [ + { + "bbox": [ + 61, + 182, + 293, + 243 + ], + "lines": [ + { + "bbox": [ + 61, + 182, + 293, + 243 + ], + "spans": [ + { + "bbox": [ + 61, + 182, + 293, + 243 + ], + "type": "table", + "html": "
ISM↓DiffFaceDiffSwapFaceSwapIP-AdapterSimSwapInfoSwap
w/o Lproj0.2410.1670.2700.1350.5440.256
w/o Lattn0.2540.1700.2230.0760.1680.252
w/o Lmtconn0.2310.1740.1660.0470.1830.354
w/o Lid0.4460.1750.2170.0400.5120.430
Ltotal0.2430.1630.1940.0720.1840.237
", + "image_path": "a3997732db5b3407d6da389ef3a9416b951b64c881bcb27b53c0800736d71fd0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 248, + 296, + 293 + ], + "lines": [ + { + "bbox": [ + 55, + 248, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 248, + 296, + 293 + ], + "type": "text", + "content": "Table 3. Each model's performance is measured using the ISM score, confirming that each loss function ensures transferability across various deepfakes. As a result, the integrated " + }, + { + "bbox": [ + 55, + 248, + 296, + 293 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{total}}" + }, + { + "bbox": [ + 55, + 248, + 296, + 293 + ], + "type": "text", + "content": " is capable of covering a wider range." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 298, + 296, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 298, + 296, + 383 + ], + "spans": [ + { + "bbox": [ + 55, + 298, + 296, + 383 + ], + "type": "text", + "content": "Attack effectiveness of mid-layers. We qualitatively show that focusing the diffusion attack on the mid-layers of the denoising UNet [27] is more effective than applying it to the entire layers or the up/down layers, as shown in Fig.6. The experiment is conducted by applying noise " + }, + { + "bbox": [ + 55, + 298, + 296, + 383 + ], + "type": "inline_equation", + "content": "\\delta = 4 / 255" + }, + { + "bbox": [ + 55, + 298, + 296, + 383 + ], + "type": "text", + "content": " to create protected images, which are then passed through the deepfake model [43] to compare the resulting outputs." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 57, + 387, + 296, + 527 + ], + "blocks": [ + { + "bbox": [ + 57, + 387, + 296, + 527 + ], + "lines": [ + { + "bbox": [ + 57, + 387, + 296, + 527 + ], + "spans": [ + { + "bbox": [ + 57, + 387, + 296, + 527 + ], + "type": "image", + "image_path": "6cfc6f3cb61eebe9e730a7e21f05e79c5ac08de5137498d0fbc3dc69f041465c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 529, + 295, + 563 + ], + "lines": [ + { + "bbox": [ + 55, + 529, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 295, + 563 + ], + "type": "text", + "content": "Figure 6. A comparison of conditioned face attack results across UNet [27] layers shows the best protection when targeting mid-layer cross-attention." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "MTCNN resize evaluation. To demonstrate the superiority of our method across various resizing modes and model transferability, we conduct an ablation study comparing it to the conventional BILINEAR scaling method [46]. We evaluate performance using different scaling methods from the OpenCV and PIL libraries, with 3,000 images from both the CelebA-HQ and VGGFace2-HQ datasets. Transferability is assessed through experiments conducted on both PyTorch and TensorFlow versions. The evaluation is based on the final bounding box detection probabilities from MTCNN [48], and the results in the supplementary material confirm that our method outperforms existing approaches." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 72, + 397, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 397, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 397, + 85 + ], + "type": "text", + "content": "4.5. Applications" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 91, + 555, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 163 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 163 + ], + "type": "text", + "content": "Extensibility on GAN-based deepfake models. We also conduct additional experiments on the GAN-based diffusion model [3, 11]. The experimental conditions are the same as those in Table 2, and the results, as shown in Table 4, indicate a degradation in model performance. Qualitative assessments are provided in Fig. 4." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 317, + 168, + 553, + 266 + ], + "blocks": [ + { + "bbox": [ + 317, + 168, + 553, + 266 + ], + "lines": [ + { + "bbox": [ + 317, + 168, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 317, + 168, + 553, + 266 + ], + "type": "table", + "html": "
ModelSimSwap [3]InfoSwap [11]
DatasetCelebA-HQ [18]
MethodL2↑ISM↓PSNR↓L2↑ISM↓PSNR↓
Original0.0000.54480.0000.0000.43180.000
Ours0.0700.18426.9210.1290.23730.220
DatasetVGGFace2-HQ [4]
MethodL2↑ISM↓PSNR↓L2↑ISM↓PSNR↓
Original0.0000.68180.0000.0000.56580.000
Ours0.0670.31427.3050.1420.35629.044
", + "image_path": "c7ceb026bad85236b5b3e4120d725b983e527f8b4e23cb5685999d0d22ea59f7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 313, + 270, + 555, + 315 + ], + "lines": [ + { + "bbox": [ + 313, + 270, + 555, + 315 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 555, + 315 + ], + "type": "text", + "content": "Table 4. Applicability of FaceShield to other deepfake frameworks. Our method, when applied to GAN-based models, not only reduces image quality " + }, + { + "bbox": [ + 313, + 270, + 555, + 315 + ], + "type": "inline_equation", + "content": "(L_{2}" + }, + { + "bbox": [ + 313, + 270, + 555, + 315 + ], + "type": "text", + "content": ", PSNR) but also significantly lowers source similarity (ISM)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 322, + 556, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 322, + 556, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 322, + 556, + 417 + ], + "type": "text", + "content": "Transferability with different weights. To demonstrate that FaceShield ensures robust transferability not only across structurally different models but also to models with similar architectures but different pre-trained weights, we evaluated its performance on various versions of IP-Adapter [43]. The results, which can be found in the supplementary material, confirm the superior transferability performance of our method." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 428, + 388, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 428, + 388, + 441 + ], + "spans": [ + { + "bbox": [ + 314, + 428, + 388, + 441 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 449, + 556, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 556, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 556, + 640 + ], + "type": "text", + "content": "In this study, we propose FaceShield, an invisible facial protection technique designed to attack various deepfakes. Through comparisons with multiple baseline methods, we demonstrate that FaceShield offers superior protection with significantly lower resource costs, particularly for deepfake models utilizing the latest diffusion techniques. Furthermore, its design integrates diverse transferability enhancement strategies, ensuring consistent performance not only across various pretrained versions but also across diffusion-based models with different architectures. This robustness extends to entirely different architectures, including GAN-based models. Additionally, by incorporating an improved noise update mechanism that ensures invisibility while minimizing information loss, FaceShield proves to be a practical and effective solution for preventing the misuse of facial images across a wide range of deepfake systems." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Limitations and Future Work. Although we introduce a method to enhance robustness against JPEG compression and resizing, other purification techniques still exist, which may lead to the potential loss of our protective noise information. Therefore, we plan to further strengthen the noise to effectively counter a broader range of purification methods." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10371" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 161, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 161, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 161, + 85 + ], + "type": "text", + "content": "6. Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 356 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 356 + ], + "type": "text", + "content": "This work was supported by Culture, Sports and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture, Sports and Tourism (International Collaborative Research and Global Talent Development for the Development of Copyright Management and Protection Technologies for Generative AI, RS-2024-00345025, " + }, + { + "bbox": [ + 55, + 91, + 297, + 356 + ], + "type": "inline_equation", + "content": "47\\%" + }, + { + "bbox": [ + 55, + 91, + 297, + 356 + ], + "type": "text", + "content": "; Research on neural watermark technology for copyright protection of generative AI 3D content, RS-2024-00348469, " + }, + { + "bbox": [ + 55, + 91, + 297, + 356 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 55, + 91, + 297, + 356 + ], + "type": "text", + "content": "), the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT)(RS-2025-00521602, " + }, + { + "bbox": [ + 55, + 91, + 297, + 356 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 55, + 91, + 297, + 356 + ], + "type": "text", + "content": "), Institute of Information & communications Technology Planning & Evaluation (IITP) & ITRC(Information Technology Research Center) grant funded by the Korea government(MSIT) (No.RS-2019-II190079, Artificial Intelligence Graduate School Program(Korea University), " + }, + { + "bbox": [ + 55, + 91, + 297, + 356 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 55, + 91, + 297, + 356 + ], + "type": "text", + "content": "; IITP-2025-RS-2024-00436857, " + }, + { + "bbox": [ + 55, + 91, + 297, + 356 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 55, + 91, + 297, + 356 + ], + "type": "text", + "content": "; IITP-2025-RS-2025-02304828, Artificial Intelligence Star Fellowship Support Program to Nurture the Best Talents, " + }, + { + "bbox": [ + 55, + 91, + 297, + 356 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 55, + 91, + 297, + 356 + ], + "type": "text", + "content": "), and Artificial intelligence industrial convergence cluster development project funded by the Ministry of Science and ICT(MSIT, Korea)&Gwangju Metropolitan City." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 363, + 115, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 363, + 115, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 363, + 115, + 376 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 383, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 61, + 383, + 295, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 383, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 61, + 383, + 295, + 415 + ], + "type": "text", + "content": "[1] N. Ahmed, T. Natarajan, and K.R. Rao. Discrete cosine transform. IEEE Transactions on Computers, C-23(1):90–93, 1974. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 416, + 295, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 416, + 295, + 459 + ], + "spans": [ + { + "bbox": [ + 61, + 416, + 295, + 459 + ], + "type": "text", + "content": "[2] Huanran Chen, Yichi Zhang, Yinpeng Dong, Xiao Yang, Hang Su, and Jun Zhu. Rethinking model ensemble in transfer-based adversarial attacks. In The Twelfth International Conference on Learning Representations, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 460, + 296, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 460, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 62, + 460, + 296, + 514 + ], + "type": "text", + "content": "[3] Renwang Chen, Xuanhong Chen, Bingbing Ni, and Yanhao Ge. Simswap: An efficient framework for high fidelity face swapping. In Proceedings of the 28th ACM international conference on multimedia, pages 2003-2011, 2020. 1, 2, 6, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 516, + 295, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 516, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 62, + 516, + 295, + 559 + ], + "type": "text", + "content": "[4] Xuanhong Chen, Bingbing Ni, Yutian Liu, Naiyuan Liu, Zhilin Zeng, and Hang Wang. Simswap++: Towards faster and high-quality identity swapping. IEEE Trans. Pattern Anal. Mach. Intell., 46(1):576-592, 2024. 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 559, + 296, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 559, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 62, + 559, + 296, + 613 + ], + "type": "text", + "content": "[5] Jongwook Choi, Taehoon Kim, Yonghyun Jeong, Seungryul Baek, and Jongwon Choi. Exploiting style latent flows for generalizing deepfake video detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1133-1143, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 615, + 295, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 615, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 62, + 615, + 295, + 658 + ], + "type": "text", + "content": "[6] June Suk Choi, Kyungmin Lee, Jongheon Jeong, Saining Xie, Jinwoo Shin, and Kimin Lee. Diffusionguard: A robust defense against malicious diffusion-based image editing. arXiv preprint arXiv:2410.05694, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 658, + 295, + 712 + ], + "type": "text", + "content": "[7] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 2, 5" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 320, + 73, + 555, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 555, + 117 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 555, + 117 + ], + "type": "text", + "content": "[8] Junhao Dong and Xiaohua Xie. Visually maintained image disturbance against deepfake face swapping. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 320, + 118, + 554, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 118, + 554, + 171 + ], + "spans": [ + { + "bbox": [ + 320, + 118, + 554, + 171 + ], + "type": "text", + "content": "[9] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9185-9193, 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 172, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 172, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 316, + 172, + 553, + 215 + ], + "type": "text", + "content": "[10] Gintare Karolina Dziugaite, Zoubin Ghahramani, and Daniel M Roy. A study of the effect of jpg compression on adversarial images. arXiv preprint arXiv:1608.00853, 2016. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 216, + 553, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 216, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 315, + 216, + 553, + 270 + ], + "type": "text", + "content": "[11] Gege Gao, Huaibo Huang, Chaoyou Fu, Zhaoyang Li, and Ran He. Information bottleneck disentanglement for identity swapping. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3404-3413, 2021. 2, 6, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 271, + 555, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 271, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 271, + 555, + 327 + ], + "type": "text", + "content": "[12] Xiao Guo, Xiaohong Liu, Zhiyuan Ren, Steven Grosz, Iacopo Masi, and Xiaoming Liu. Hierarchical fine-grained image forgery detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3155-3165, 2023. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 327, + 555, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 555, + 382 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 555, + 382 + ], + "type": "text", + "content": "[13] Baojin Huang, Zhongyuan Wang, Jifan Yang, Jiaxin Ai, Qin Zou, Qian Wang, and Dengpan Ye. Implicit identity driven deepfake face swapping detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4490-4499, 2023. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 383, + 554, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 383, + 554, + 448 + ], + "spans": [ + { + "bbox": [ + 317, + 383, + 554, + 448 + ], + "type": "text", + "content": "[14] Hao Huang, Yongtao Wang, Zhaoyu Chen, Yuze Zhang, Yuheng Li, Zhi Tang, Wei Chu, Jingdong Chen, Weisi Lin, and Kai-Kuang Ma. Cmua-watermark: A cross-model universal adversarial watermark for combating deepfakes. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 989-997, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 449, + 554, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 554, + 502 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 554, + 502 + ], + "type": "text", + "content": "[15] Qian Huang, Isay Katsman, Horace He, Zeqi Gu, Serge Belongie, and Ser-Nam Lim. Enhancing adversarial example transferability with an intermediate level attack. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4733-4742, 2019. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 503, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 503, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 317, + 503, + 553, + 547 + ], + "type": "text", + "content": "[16] Nathan Inkawich, Kevin Liang, Lawrence Carin, and Yiran Chen. Transferable perturbations of deep feature distributions. In International Conference on Learning Representations, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 548, + 553, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 553, + 591 + ], + "type": "text", + "content": "[17] Nick Kanopoulos, Nagesh Vasanthavada, and Robert L Baker. Design of an image edge detection filter using the sobel operator. IEEE Journal of solid-state circuits, 23(2): 358-367, 1988. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 592, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 592, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 317, + 592, + 553, + 624 + ], + "type": "text", + "content": "[18] Tero Karras. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 625, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 625, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 625, + 553, + 689 + ], + "type": "text", + "content": "[19] Edgar Kaziakhmedov, Klim Kireev, Grigorii Melnikov, Mikhail Pautov, and Aleksandr Petiushko. Real-world attack on mtcnn face detection system. In 2019 International Multi-Conference on Engineering, Computer and Information Sciences (SIBIRCON), pages 0422-0427. IEEE, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 690, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 690, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 690, + 553, + 714 + ], + "type": "text", + "content": "[20] Kihong Kim, Yunho Kim, Seokju Cho, Junyoung Seo, Jisu Nam, Kychul Lee, Seung Wook Kim, and Kwanghee Lee." + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10372" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "text", + "content": "Difface: Diffusion-based face swapping with facial guidance. Pattern Recognit., 163:111451, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 148 + ], + "type": "text", + "content": "[21] Jia Li, Zhaoyang Li, Jie Cao, Xingguang Song, and Ran He. Faceinpainter: High fidelity face adaptation to heterogeneous domains. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5089-5098, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 183 + ], + "type": "text", + "content": "[22] Lingzhi Li, Jianmin Bao, Hao Yang, Dong Chen, and Fang Wen. Faceshifter: Towards high fidelity and occlusion aware face swapping. arXiv preprint arXiv:1912.13457, 2019. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 184, + 294, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 294, + 216 + ], + "type": "text", + "content": "[23] Chumeng Liang and Xiaoyu Wu. Mist: Towards improved adversarial examples for diffusion models. arXiv preprint arXiv:2305.12683, 2023. 1, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 217, + 294, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 271 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 271 + ], + "type": "text", + "content": "[24] Chumeng Liang, Xiaoyu Wu, Yang Hua, Jiaru Zhang, Yiming Xue, Tao Song, Zhengui Xue, Ruhui Ma, and Haibing Guan. Adversarial example does good: Preventing painting imitation from diffusion models via adversarial examples. arXiv preprint arXiv:2302.04578, 2023. 1, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 272, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 316 + ], + "type": "text", + "content": "[25] Kunlin Liu, Ivan Perov, Daiheng Gao, Nikolay Chervoniy, Wenbo Zhou, and Weiming Zhang. Deepfacelab: Integrated, flexible and extensible face-swapping framework. Pattern Recognit., 141:109628, 2020. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 316, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 381 + ], + "type": "text", + "content": "[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 383, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 437 + ], + "type": "text", + "content": "[27] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 3, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 437, + 294, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 503 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 503 + ], + "type": "text", + "content": "[28] Nataniel Ruiz, Sarah Adel Bargal, and Stan Sclaroff. Disrupting deepfakes: Adversarial attacks against conditional image translation networks and facial manipulation systems. In Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 236-251. Springer, 2020. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 504, + 294, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 294, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 294, + 547 + ], + "type": "text", + "content": "[29] Hadi Salman, Alaa Khaddaj, Guillaume Leclerc, Andrew Ilyas, and Aleksander Madry. Raising the cost of malicious ai-powered image editing. In International Conference on Machine Learning, 2023. 1, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "type": "text", + "content": "[30] Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y Zhao. Glaze: Protecting artists from style mimicry by {Text-to-Image} models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 2187–2204, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 603, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 658 + ], + "type": "text", + "content": "[31] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, and Yunchao Wei. Learning on gradients: Generalized artifacts representation for gan-generated images detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12105-12114, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "text", + "content": "[32] Thanh Van Le, Hao Phung, Thuan Hoang Nguyen, Quan Dao, Ngoc N Tran, and Anh Tran. Anti-dreambooth: Protecting users from personalized text-to-image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2116-2127, 2023. 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "type": "text", + "content": "[33] Andrey Voynov, Qinghao Chu, Daniel Cohen-Or, and Kfir Aberman. " + }, + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "type": "inline_equation", + "content": "\\mathfrak{p}+" + }, + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "type": "text", + "content": ": Extended textual conditioning in text-to-image generation. arXiv preprint arXiv:2303.09522, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 139 + ], + "type": "text", + "content": "[34] Feifei Wang. Face swap via diffusion model. arXiv preprint arXiv:2403.01108, 2024. 1, 2, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "type": "text", + "content": "[35] Run Wang, Ziheng Huang, Zhikai Chen, Li Liu, Jing Chen, and Lina Wang. Anti-forgery: Towards a stealthy and robust deepfake disruption attack via adversarial perceptual-aware perturbations. arXiv preprint arXiv:2206.00477, 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 186, + 555, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 555, + 228 + ], + "type": "text", + "content": "[36] Tianyi Wang and Kam Pui Chow. Noise based deepfake detection via multi-head relative-interaction. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 14548-14556, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 232, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 232, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 553, + 285 + ], + "type": "text", + "content": "[37] Yuhan Wang, Xu Chen, Junwei Zhu, Wenqing Chu, Ying Tai, Chengjie Wang, Jilin Li, Yongjian Wu, Feiyue Huang, and Rongrong Ji. Hififace: 3d shape and semantic prior guided high fidelity face swapping. arXiv preprint arXiv:2106.09965, 2021. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 287, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 341 + ], + "type": "text", + "content": "[38] Zhibo Wang, Hengchang Guo, Zhifei Zhang, Wenxin Liu, Zhan Qin, and Kui Ren. Feature importance-aware transferable adversarial attacks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7639-7648, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 343, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 398 + ], + "type": "text", + "content": "[39] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L Yuille. Improving transferability of adversarial examples with input diversity. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2730-2739, 2019. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 399, + 555, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 555, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 555, + 443 + ], + "type": "text", + "content": "[40] Chao Xu, Jiangning Zhang, Miao Hua, Qian He, Zili Yi, and Yong Liu. Region-aware face swapping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7632-7641, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 444, + 555, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 555, + 498 + ], + "type": "text", + "content": "[41] Yangyang Xu, Bailin Deng, Junle Wang, Yanqing Jing, Jia Pan, and Shengfeng He. High-resolution face swapping via latent semantics disentanglement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7642-7651, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 500, + 553, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 553, + 553 + ], + "type": "text", + "content": "[42] Haotian Xue, Chumeng Liang, Xiaoyu Wu, and Yongxin Chen. Toward effective protection against diffusion-based mimicry through score distillation. In The Twelfth International Conference on Learning Representations, 2023. 1, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 555, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 553, + 599 + ], + "type": "text", + "content": "[43] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 601, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 656 + ], + "type": "text", + "content": "[44] Chin-Yuan Yeh, Hsi-Wen Chen, Shang-Lun Tsai, and Sheng-De Wang. Disrupting image-translation-based deepfake algorithms with adversarial attacks. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision Workshops, pages 53–62, 2020. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "type": "text", + "content": "[45] Ge Yuan, Maomao Li, Yong Zhang, and Huicheng Zheng. Reliableswap: Boosting general face swapping via reliable supervision. arXiv preprint arXiv:2306.05356, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[46] Chongyang Zhang, Yu Qi, and Hiroyuki Kameda. Multiscale perturbation fusion adversarial attack on mtconn face" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10373" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 396 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "type": "text", + "content": "detection system. In 2022 4th International Conference on Communications, Information System and Computer Engineering (CISCE), pages 142-146. IEEE, 2022. 2, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 138 + ], + "type": "text", + "content": "[47] Jianping Zhang, Zhuoer Xu, shiwen cui, Changhua Meng, Weibin Wu, and Michael Lyu. On the robustness of latent diffusion models, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "type": "text", + "content": "[48] Kaipeng Zhang, Zhanpeng Zhang, Zhifeng Li, and Yu Qiao. Joint face detection and alignment using multitask cascaded convolutional networks. IEEE signal processing letters, 23 (10):1499-1503, 2016. 2, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 251 + ], + "type": "text", + "content": "[49] Wenliang Zhao, Yongming Rao, Weikang Shi, Zuyan Liu, Jie Zhou, and Jiwen Lu. Diffswap: High-fidelity and controllable face swapping via 3d-aware masked diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8568-8577, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 294, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 305 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 305 + ], + "type": "text", + "content": "[50] Wen Zhou, Xin Hou, Yongjun Chen, Mengyun Tang, Xiangqi Huang, Xiang Gan, and Yong Yang. Transferable adversarial perturbations. In Proceedings of the European Conference on Computer Vision (ECCV), pages 452-467, 2018. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 308, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 294, + 352 + ], + "type": "text", + "content": "[51] Yuhao Zhu, Qi Li, Jian Wang, Cheng-Zhong Xu, and Zhenan Sun. One shot face swapping on megapixels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4834-4844, 2021. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 396 + ], + "type": "text", + "content": "[52] Yixuan Zhu, Wenliang Zhao, Yansong Tang, Yongming Rao, Jie Zhou, and Jiwen Lu. Stableswap: Stable face swapping in a shared and controllable latent space. IEEE Transactions on Multimedia, 2024. 2" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10374" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/FaceXFormer_ A Unified Transformer for Facial Analysis/43f275cd-265b-414a-a63d-4d3a6e26c5b2_content_list.json b/2025/FaceXFormer_ A Unified Transformer for Facial Analysis/43f275cd-265b-414a-a63d-4d3a6e26c5b2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2723f28250c674136f1088d390f11ab8dd3c1b31 --- /dev/null +++ b/2025/FaceXFormer_ A Unified Transformer for Facial Analysis/43f275cd-265b-414a-a63d-4d3a6e26c5b2_content_list.json @@ -0,0 +1,1763 @@ +[ + { + "type": "text", + "text": "FaceXFormer: A Unified Transformer for Facial Analysis", + "text_level": 1, + "bbox": [ + 209, + 130, + 787, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kartik Narayan* Vibashan VS* Rama Chellappa Vishal M. Patel {knaraya4, vvishnu2, rchella4, vpatel36}@jhu.edu https://kartik-3004.github.io/facexformer/", + "bbox": [ + 222, + 180, + 776, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we introduce FaceXFormer, an end-to-end unified transformer model capable of performing ten facial analysis tasks within a single framework. These tasks include face parsing, landmark detection, head pose estimation, attribute prediction, age, gender, and race estimation, facial expression recognition, face recognition, and face visibility. Traditional face analysis approaches rely on task-specific architectures and pre-processing techniques, limiting scalability and integration. In contrast, FaceXFormer employs a transformer-based encoder-decoder architecture, where each task is represented as a learnable token, enabling seamless multi-task processing within a unified model. To enhance efficiency, we introduce FaceX, a lightweight decoder with a novel bi-directional cross-attention mechanism, which jointly processes face and task tokens to learn robust and generalized facial representations. We train FaceXFormer on ten diverse face perception datasets and evaluate it against both specialized and multi-task models across multiple benchmarks, demonstrating state-of-the-art or competitive performance. Additionally, we analyze the impact of various components of FaceXFormer on performance, assess real-world robustness in \"in-the-wild\" settings, and conduct a performance comparison under different backbones. To the best of our knowledge, FaceXFormer is the first model capable of handling ten facial analysis tasks while maintaining real-time performance at 33.21 FPS.", + "bbox": [ + 88, + 301, + 485, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 741, + 220, + 757 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Face analysis is a crucial problem as it has broad range of application such as face verification and identification [89, 90], surveillance [24], face swapping [13], face editing [134], de-occlusion [117], 3D face reconstruction [108], retail [1], image generation [115] and face retrieval [119]. Facial analysis tasks (Figure 1 include face parsing [32, 104], landmarks detection [49, 132], head pose", + "bbox": [ + 89, + 767, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/011f675c36c3728dc56d0534340176552e1e5573283a668947f968a7b6f83ca8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 268, + 640, + 324 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/34e47ae5cec6fe0b54160e6ba718692cfc847898aad1c890a23470e0fd9bc08e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 268, + 767, + 324 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4fd5f53af71db7a0b2293e5ffa098e8b233bf57ffb12f4790325dc9aa8bffc32.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 769, + 268, + 890, + 324 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/320fad68ac5a570511cd16ca92b3c93e3cedbdb5934c042777b6e7f800992b61.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 325, + 640, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/824afc3c35698bb46270a631fe825c7523021bb2a20a7c3c214a39436efafa9e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 325, + 767, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d2b05e3d9db0c9a991e726c6d6b4c1116b70fbd1c454fdd19e4d16b908cbe07c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 769, + 325, + 890, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4c673f1707ab6d6e1b6ca15e0511c3e24f28bd831f58c6720571f2780c667b15.jpg", + "image_caption": [ + "Figure 1. FaceXFormer an end-to-end unified transformer model for 10 different facial analysis tasks such as face parsing, landmark detection, head pose estimation, attributes recognition, age, gender, and race estimation, facial expression recognition, face recognition, and face visibility prediction." + ], + "image_footnote": [], + "bbox": [ + 519, + 383, + 640, + 443 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/aaa3d90d9b69c1c1ef23d3370edf0460ce75e110a59205d69ab14310d6704b6d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 383, + 767, + 443 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ce8fd9dc97c7eb3f66bc9b90069f95ad19af7b4cffe63d4d648a9741d2585acf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 769, + 383, + 890, + 443 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "estimation [12, 131], facial attributes recognition [61, 69], age/gender/race estimation [8, 46], facial expression recognition [82], face recognition [37], and face visibility prediction [39, 58]. Therefore, developing a generalized and robust face model for all tasks is a crucial and longstanding problem in the face community.", + "bbox": [ + 511, + 551, + 906, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Why Unified Model? In recent years, significant advancements have been made in facial analysis, developing state-of-the-art methods and face libraries for various tasks [12, 13, 46, 117, 131, 132]. Despite these methods achieving promising performance, they cannot be integrated into a single pipeline due to their specialized model designs and task-specific pre-processing techniques. Furthermore, deploying multiple specialized models simultaneously is computationally intensive and impractical for real-time applications, leading to increased system complexity and resource consumption. These challenges emphasize the need for a unified model that can concurrently handle multiple facial analysis tasks efficiently (see Table 1). A single model capable of addressing multiple facial tasks is desirable because it: (1) learns a robust and generalized face representation capable of handling \"in-the-wild\" images; (2) intra-task modeling helps the models to learn task-invariant representation;", + "bbox": [ + 511, + 643, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution", + "bbox": [ + 109, + 887, + 217, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11369", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/1c6479d6fd4d0219ddc7c277b09279e94a29005c48a5bc91180cc35763279bd8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsFPLDHPEAttrAgeGenRaceVisExpFR
Single-Task Models
DML-CSR [127]✓
FP-LIIF [80]✓
SegFace [67]✓
Wing [22]✓
HRNet [98]✓
WHENet [131]✓
TriNet [9]✓
img2pose [3]✓
TokenHPE [121]✓
SSPL [85]✓
VOLO-D1 [40]✓
DLDL-v2 [23]✓
3DDE [94]✓
MNN [95]✓
KTN [44]✓
DMUE [82]✓
CosFace [97]✓
ArcFace [15]✓
AdaFace [37]✓
Multi-Task Models
SSP+SSG [34]✓✓
Hetero-FAE [27]✓✓✓✓✓
FairFace [35]✓✓✓
MiVOLO [40]✓✓
MTL-CNN [137]✓✓✓
ProS [17]✓✓✓
FaRL [130]✓✓✓✓✓
HyperFace [75]✓✓✓✓✓
AllinOne [76]✓✓✓✓✓✓
Swinface [71]✓✓✓✓
QFace [88]✓✓✓✓
Facetor [72]✓✓✓✓✓✓✓
FaceXFormer✓✓✓✓✓✓✓✓✓✓
", + "bbox": [ + 94, + 89, + 483, + 489 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Comparison with representative methods under different task settings. Our proposed FaceXFormer can perform various facial analysis tasks in a single model. FP - Face Parsing, LD - Landmarks Detection, HPE - Head Pose Estimation, Attr - Attributes Recognition, Age - Age, Gen - Gender, Race - Race Estimation, Exp - Facial Expression Recognition, FR - Face Recognition, and Vis - Face Visibility Prediction", + "bbox": [ + 89, + 502, + 483, + 599 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and (3) simplifies deployment pipelines by reducing computational overhead and enabling faster inference.", + "bbox": [ + 89, + 628, + 482, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Proposed FaceXFormer Architecture: To this end, we introduce FaceXFormer, an end-to-end unified model designed for ten different facial analysis tasks, as depicted in Figure 1. These tasks include face parsing, landmark detection, head pose estimation, attributes recognition, age/gender/race estimation, facial expression recognition, face recognition and face visibility prediction. FaceXFormer enables task unification by leveraging transformers and learnable tokens as its core components. Specifically, we employ a transformer-based encoder-decoder structure, where the encoder extracts hierarchical face representations and fuses them using a MLP fusion module. The fused features are then processed in the decoder, where each facial analysis task is represented by a unique learnable token, allowing for the simultaneous and effective processing of multiple tasks. In particular, we propose a lightweight de", + "bbox": [ + 88, + 659, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "coder, FaceX, which processes both face and task tokens together using bi-directional cross-attention mechanism (Section 3.2), enabling the model to learn robust face representations that generalize across various tasks. The bi-directional cross-attention mechanism enables a 2-layer lightweight decoder, allowing the model to operate in real time. After modeling the intra-task and face-token relationships in the FaceX decoder, the task tokens are fed into a unified head, which converts these task tokens into corresponding task predictions.", + "bbox": [ + 511, + 90, + 903, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our extensive experiments demonstrate that FaceX-Former achieves state-of-the-art or competitive performance compared to specialized models and existing multitask models across multiple benchmarks, while supporting more tasks than any previous multi-task model. Moreover, we show that our model effectively handles images \"in the wild\", demonstrating its robustness and generalizability across ten different tasks. This robustness is critical for real-world applications where uncontrolled conditions and diverse inputs are common. FaceXFormer achieves state-of-the-art performance at 33.21 FPS, representing a significant $69.44\\%$ speed boost over prior multi-task models, making it highly suitable for real-world applications.", + "bbox": [ + 511, + 242, + 906, + 438 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our paper's contributions are as follows:", + "bbox": [ + 532, + 439, + 883, + 453 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We introduce FaceXFormer, a unified transformer-based framework capable of simultaneously processing ten different facial analysis tasks, achieving real-time performance of 33.21 FPS.", + "2. We propose FaceX, a lightweight decoder that employs the proposed bi-directional cross-attention mechanism, enabling joint processing of face and task tokens.", + "3. We conduct extensive experiments and analyses to demonstrate that our approach achieves state-of-the-art performance with reduced inference time compared to specialized and multi-task models across multiple tasks." + ], + "bbox": [ + 513, + 454, + 903, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 633, + 653, + 648 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Facial analysis tasks: Facial analysis tasks involve face parsing [11, 32, 67, 127], landmarks detection [49, 59, 132], head pose estimation [12, 95, 121, 131], facial attributes recognition [61, 69, 85, 130], age/gender/race estimation [8, 40, 43, 46], facial expression recognition [44, 45], face recognition [15, 97] and face visibility prediction [39, 58]. These tasks hold significance in various applications such as face swapping [13, 65], face editing [134], de-occlusion [117], 3D face reconstruction [108], driver assistance [64], human-robot interaction [86], retail [1], face verification and identification [89, 90], image generation [115], image retrieval [119] and surveillance [24, 66]. Specialized models excel in their respective tasks but cannot be easily integrated with other tasks due to the need for extensive task-specific pre-processing [50, 132]. Generally, these models under-perform when applied to tasks beyond", + "bbox": [ + 511, + 657, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11370", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/7ff01b30d6774eba61d478272a32effcb7d0568fe9bcb496f4e4646796813771.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
QFace [88]Facetor [72]FaceXFormer
Tasks4710
Model Size-178.9 M109.29 M
Face PretrainedYesYesNo
FPS (fp32)-14.3033.21
DecoderQuery2Label [52]Pixel Decoder & Transformer DecoderFaceX
Decoder Layers992
", + "bbox": [ + 96, + 88, + 478, + 186 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 2. Comparison of multi-task face analysis methods.", + "bbox": [ + 114, + 196, + 455, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "their specialization as their design is specific to their designated tasks. Some works [29, 60, 124, 126] perform multiple tasks simultaneously but utilize the additional tasks for guidance or auxiliary loss calculation to enhance the performance of the primary task.", + "bbox": [ + 89, + 231, + 482, + 306 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-task learning for face analysis: HyperFace [75] and AllinOne [76] are early convolution-based models that aim to perform multiple tasks. Recent multi-task frameworks, such as QFace [88] and Facetor [72], are also inspired from DETR [10] and propose a unified model structure consisting of learnable tokens. However, these previous works differ from the proposed method in several key aspects, as summarized in Table 2. Specifically, QFace [88] employs a feature fusion module that uses stage embeddings to aggregate features from the encoder. Facetor [72] introduces a layer-attention mechanism to fuse features from different encoder layers and incorporates two separate decoders. Both methods, employ a 9-layer transformer decoder, also Facetor additionally includes a Pixel Decoder. These architectural components increase computational overhead, resulting in slower inference times. In contrast, FaceXFormer proposes a bi-directional cross-attention mechanism, which enables efficient task-specific feature extraction from face tokens resulting in a 2-layer lightweight decoder. This design choice is the primary reason for FaceXFormer's superior speed and performance. Notably, unlike previous methods, FaceXFormer does not rely on face-specific pertaining backbone.", + "bbox": [ + 91, + 309, + 483, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unified transformer models: In recent years, the rise of transformers [19, 96] have paved the way for the unification of multiple tasks within a single architecture. Unified transformer architectures are being explored across various computer vision problems, including segmentation [47, 138], visual question answering (VQA) [99, 118], tracking [102, 133], detection [103]. While these models may not achieve state-of-the-art (SOTA) performance and may under-perform compared to specialized models on some tasks, they demonstrate competitive performance across a variety of tasks. Such unification efforts have led to the development of foundational models like SAM [38], CLIP [73], LLaMA [93], GPT-3 [6], DALL-E [74], etc. However, these models are computationally intensive and not suitable for facial analysis applications that require real-time performance. Motivated by this challenge, we propose FaceXFormer: the first lightweight, transformer-based", + "bbox": [ + 89, + 643, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "model capable of performing multiple facial analysis tasks. It delivers real-time performance at 33.21 FPS and can be seamlessly integrated into existing systems providing additional annotations for the person of interest.", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. FaceXFormer", + "text_level": 1, + "bbox": [ + 511, + 165, + 655, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our framework, we follow a standard encoder-decoder structure as illustrated in Fig. 2. For an input face image $\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}$ , we extract coarse to fine-grained multiscale features $\\mathbf{S}_i$ , where $i$ belongs to the $i$ -th encoder output. To learn a unified face representation $\\mathbf{F}$ , these multiscale features are then fused using a MLP-Fusion $\\mathbf{M}$ module. Following fusion, we initialize a series of task-specific tokens $\\mathbf{T} = \\langle T_1, \\dots, T_n \\rangle$ , with each $t_i$ representing a face task. Face tokens $\\mathbf{F}$ and task tokens $\\mathbf{T}$ are then processed by a lightweight Decoder FXDec where task tokens are attended with face tokens to learn relevant task representation.", + "bbox": [ + 511, + 191, + 906, + 369 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\langle \\hat {\\mathbf {T}} \\rangle = \\mathbf {F} \\mathbf {X D e c} (\\langle \\mathbf {F}, \\mathbf {T} \\rangle ; \\mathbf {S} _ {i})\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 369, + 805, + 388 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $\\hat{\\mathbf{T}}$ represents the output task tokens. These tokens are then fed into unified heads, where each task token is refined and passed to its respective task head for prediction.", + "bbox": [ + 511, + 395, + 906, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Multi-scale Encoder", + "text_level": 1, + "bbox": [ + 511, + 450, + 705, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the encoder, we employ a multi-scale encoding strategy to address the varying feature requirements intrinsic to each face analysis task. For instance, age estimation requires a global representation, while face parsing necessitates a fine-grained representation. Given an input image $\\mathbf{I}$ , it is processed through a set of encoder layers. For each encoder layer, the output captures information at varying levels of abstraction and detail, generating multi-scale features $\\{\\mathbf{S}_i\\}_{i=1}^n$ , where $i$ ranges from 1 to 4. This results in a hierarchical structure of features, wherein each feature map $\\mathbf{S}_i$ transitions from a coarse to a fine-grained representation suitable for diverse facial analysis tasks.", + "bbox": [ + 511, + 473, + 906, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "MLP-Fusion: Assigning each feature-map $\\mathbf{S}_i$ to each face task is sub-optimal; rather, learning a unified face representation is more optimal and parameter-efficient. Following [113], we utilize a MLP-Fusion module $\\mathbf{M}$ to generate a fused face representation from the multi-scale features $\\{\\mathbf{S}_i\\}_{i=1}^n$ . In this framework, each feature map $\\mathbf{S}_i$ is initially passed through a separate MLP layer, standardizing the channel dimensions across scales to facilitate fusion. The transformed features are then concatenated and passed through a fusion MLP layer to aggregate a fused representation $\\mathbf{F}$ as follows:", + "bbox": [ + 511, + 669, + 908, + 834 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {S}} _ {i} = \\operatorname {M L P} _ {\\operatorname {p r o j}} \\left(D _ {i}, D _ {t}\\right) \\left(\\mathbf {S} _ {i}\\right), \\forall i \\in \\{1, \\dots , n \\},\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 844, + 857, + 864 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {\\text {c a t}} = \\operatorname {C o n c a t} \\left(\\hat {\\mathbf {S}} _ {1}, \\hat {\\mathbf {S}} _ {2}, \\dots , \\hat {\\mathbf {S}} _ {n}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 866, + 772, + 883 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} = \\mathbf {M L P} _ {\\text {f u s i o n}} (n D _ {t}, D _ {t}) (\\mathbf {F} _ {\\text {c a t}}),\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 886, + 772, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11371", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ed200455b8c4d7533c92c3aa23b01acb1e1a467f965197a79eb665254c14938f.jpg", + "image_caption": [ + "Figure 2. Overview of our proposed framework. The FaceXFormer employs an encoder-decoder architecture, extracting multi-scale features from the input face image $\\mathbf{I}$ , and fusing them into a unified representation $\\mathbf{F}$ via MLP-Fusion. Task tokens $\\mathbf{T}$ are processed alongside face representation $\\mathbf{F}$ in the FaceX Decoder FXDec, resulting in refined task-specific tokens $\\hat{\\mathbf{T}}$ . These refined tokens are then used for task-specific predictions by passing through the unified head. FaceXFormer performs ten tasks, including face parsing, landmark detection, head pose estimation, attribute prediction, age, gender, and race estimation, facial expression recognition, face recognition, and face visibility prediction, achieving state-of-the-art performance at a real-time FPS of 33.21." + ], + "image_footnote": [], + "bbox": [ + 94, + 94, + 888, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $D_{i}$ and $D_{t}$ are the multi-scale feature channel dimensions of $\\mathbf{S}_i$ and the target channel dimension, respectively. The MLP-fusion design ensures minimal computational overhead (983k parameters) while maintaining the ability to perform efficient feature fusion, which is crucial for real-time application based face analysis tasks.", + "bbox": [ + 88, + 474, + 483, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. FaceX Decoder", + "text_level": 1, + "bbox": [ + 89, + 575, + 243, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Detection transformer (DETR) [10] employs object tokens to learn bounding box predictions for each object. Inspired by this approach, we introduce Task Tokens, whereby each task token is designed to learn specific facial tasks leveraging the fused face representation. However, existing decoders such as DETR [10] and Deformable-DETR [136] are computationally intensive, impacting runtime significantly. To address this, we propose FaceX (FXDec) a lightweight decoder designed to efficiently model the task tokens with face tokens. Specifically, each task token learns a task-related representation by interacting with other task tokens $\\mathbf{T}$ and face tokens $\\mathbf{F}$ , enhancing the overall representation. The Lightweight Decoder comprises of three main components: 1) Task Self-Attention, 2) Task-to-Face Cross-Attention, and 3) Face-to-Task Cross-Attention as illustrated in Figure 2.", + "bbox": [ + 89, + 598, + 483, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Task Self-Attention (TSA): The Task Self-Attention module is designed to refine the task-specific representations within the set of task tokens $\\mathbf{T} = \\langle T_1, \\ldots, T_n \\rangle$ . Each task", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "token $T_{i}$ is an embedded representation that corresponds to a specific facial task. In TSA, each $T_{i}$ is updated by attending to all other task tokens to capture task-specific interactions. Formally, the updated task token $T_{i}^{\\prime}$ is computed as:", + "bbox": [ + 511, + 474, + 906, + 536 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T} _ {i} ^ {\\prime} = \\operatorname {S e l f A t t n} (\\mathbf {Q} = T _ {i}, \\mathbf {K} = \\mathbf {T}, \\mathbf {V} = \\mathbf {T}),\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 542, + 849, + 560 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where Attention denotes the multi-headed self-attention mechanism, and $\\mathbf{Q}$ , $\\mathbf{K}$ , and $\\mathbf{V}$ represent the queries, keys, and values, respectively. Therefore, TSA essentially helps the model to learn task-invariant representation.", + "bbox": [ + 511, + 566, + 906, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Task-to-Face Cross-Attention (TFCA): The Task-to-Face Cross-Attention module allows each task token to interact with the fused face representation $\\mathbf{F}$ . This enables each task token to gather information relevant to its specific facial task from the fused face features. In this module, the fused face representation $\\mathbf{F}$ acts as both key and value, while the task tokens serve as queries. The updated task token $\\hat{T}_i$ is then computed as follows:", + "bbox": [ + 511, + 642, + 908, + 763 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {T} _ {i} = \\operatorname {C r o s s A t t n} (\\mathbf {Q} = T _ {i} ^ {\\prime}, \\mathbf {K} = \\mathbf {F}, \\mathbf {V} = \\mathbf {F}),\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 768, + 852, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{T}} = \\langle \\hat{T}_1,\\dots ,\\hat{T}_n\\rangle$ is the output task token. Thus, TFCA enables direct interaction between the task-specific tokens and the compact facial features, facilitating task-focused feature extraction.", + "bbox": [ + 511, + 792, + 905, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Face-to-Task Cross-Attention (FTCA): Conversely, the Face-to-Task Cross-Attention module is designed to refine", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11372", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the fused face representation $\\mathbf{F}$ based on the information from the updated task tokens. This process aids in enhancing the face representation with task-specific details, thereby improving the extraction of overall fused representation. In FTCA, the set of updated task tokens $\\mathbf{T}' = \\{\\mathbf{T}_1'',\\mathbf{T}_2'',\\dots,\\mathbf{T}_m''\\}$ acts as both keys and values, while the fused face features $\\mathbf{F}$ serve as queries. The refined face representation $\\hat{\\mathbf{F}}$ is computed as:", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {F}} = \\operatorname {C r o s s A t t n} (\\mathbf {Q} = \\mathbf {F}, \\mathbf {K} = \\mathbf {T} ^ {\\prime}, \\mathbf {V} = \\mathbf {T} ^ {\\prime}).\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 222, + 431, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Through this inverse attention mechanism, the face representation is augmented with critical task-specific details, enabling a robust approach towards facial task unification.", + "bbox": [ + 89, + 251, + 482, + 296 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Unified-Head", + "text_level": 1, + "bbox": [ + 89, + 305, + 230, + 320 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Unified-Head, the task tokens are processed to obtain corresponding task predictions. As shown in Figure 2, the output face tokens $\\hat{\\mathbf{F}}$ and task tokens $\\hat{\\mathbf{T}}$ are processed through a Task-to-Face Cross-Attention mechanism to obtain final refined features. Then, the output tokens are fed into their corresponding task heads. The task head for landmark detection is a hourglass network, for head pose estimation is a regression MLP, and for face recognition is PartialFC [4], while the tasks of age, gender and race estimation, facial expression recognition, face visibility prediction, and attributes prediction utilize classification MLPs. For face parsing, we leverage the output $\\hat{\\mathbf{F}}$ and process it through an upsampling layer, then perform a cross-product with the face parsing token to obtain a segmentation map. The number of tokens for segmentation corresponds to the total number of classes. For landmark prediction, it corresponds to the number of landmarks (i.e., 68). For head pose estimation, the number of tokens is 9, representing the $3\\times 3$ rotation matrix. For other tasks, one token is used for each.", + "bbox": [ + 89, + 327, + 483, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Multi-Task Training", + "text_level": 1, + "bbox": [ + 89, + 625, + 284, + 640 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We aim to train FaceXFormer for multiple facial analysis tasks simultaneously, however each task requires distinct and sometimes conflicting pre-processing steps. For instance, landmark detection typically requires keypoint alignment of faces, which contradicts the needs for head pose estimation, as it may eliminate the natural variability of headposes. Due to these reasons, integrating all tasks into a single model poses significant challenges. To address this, FaceXFormer incorporates task-specific tokens designed to extract task-specific features from the fused representation. These task tokens compel the backbone to learn a unified representation capable of supporting a broad spectrum of facial analysis tasks. We employ different loss functions for each task and combine them in a joint objective for training. The final loss function is given as:", + "bbox": [ + 89, + 646, + 482, + 871 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L = \\lambda_ {s e g} L _ {s e g} + \\lambda_ {l n d} L _ {l n d} + \\lambda_ {h p e} L _ {h p e} + \\lambda_ {a t t r} L _ {a t t r} + \\lambda_ {a} L _ {a} \\\\ + \\lambda_ {g / r} L _ {g / r} + \\lambda_ {e x p} L _ {e x p} + \\lambda_ {f r} L _ {f r} + \\lambda_ {v i s} L _ {v i s} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 872, + 478, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $L_{seg}$ is the mean of dice loss [87] and Cross-Entropy (CE) loss for face parsing, $L_{lnd}$ is STAR loss [132] for landmarks prediction, $L_{hpe}$ is geodesic loss [121] for head pose estimation, $L_{g/r}$ is CE loss for gender/race estimation, $L_a$ is mean of L1 loss and CE loss for age estimation, $L_{exp}$ is CE loss for facial expression recognition, $L_{fr}$ is Arc-Face [15] loss for face recognition, and $L_{attr}$ and $L_{vis}$ are Binary Cross-Entropy with logits loss for attributes prediction and face visibility prediction respectively.", + "bbox": [ + 511, + 90, + 903, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments and Results", + "text_level": 1, + "bbox": [ + 511, + 239, + 746, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To validate the effectiveness of FaceXFormer, we conduct extensive experiments across ten diverse facial analysis tasks using a range of datasets, as described in Section 4.1. Implementation details are provided in Section 4.2 to ensure reproducibility and transparency. Finally, Sections 4.3 and 4.4 present the quantitative and qualitative results, respectively.", + "bbox": [ + 511, + 265, + 906, + 371 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Metrics", + "text_level": 1, + "bbox": [ + 511, + 380, + 712, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We perform co-training, where the model is simultaneously trained for multiple tasks using a total of 10 datasets with task-specific annotations. We conduct a comprehensive evaluation, comparing our approach with both task-specific and multi-task models. We present our results on the test sets according to the standard protocol for each task using the following datasets:", + "bbox": [ + 511, + 401, + 903, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Train: Face Parsing: CelebAMaskHQ [42]; Landmarks Detection: 300W [79]; Head Pose Estimation: 300W-LP [135]; Attributes Prediction: CelebA [54]; Facial Expression Recognition: RAF-DB [45], AffectNet [62]; Age/Gender/Race estimation: UTKFace [125], FairFace [36]; Face Recognition: MS1MV3 [25]; Visibility Prediction: COFW [7].", + "bbox": [ + 511, + 507, + 903, + 612 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Test: Face Parsing: CelebAMaskHQ [42]; Landmarks Detection: 300W [135], 300VW [83]; Head Pose Estimation: BIWI [20]; Attributes Prediction: CelebA [54], LFWA [107]; Facial Expression Recognition: RAFDB [45]; Age/Gender/Race Estimation: UTKFace [125], FairFace [36]; Face Recognition: LFW [31], CFP-FP [81], AgeDB [63], CALFW [129], CPLFW [128]; Visibility Prediction: COFW [7].", + "bbox": [ + 511, + 613, + 903, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The evaluation metrics used are the F1-score for face parsing, Normalized Mean Error (NME) for landmark prediction, Mean Absolute Error (MAE) for head pose estimation and age estimation, accuracy for facial expression recognition, attributes prediction, gender estimation, race estimation, 1:1 verification accuracy for face recognition, and recall at $80\\%$ precision for face visibility prediction.", + "bbox": [ + 511, + 734, + 903, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 848, + 730, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We train our models using a distributed PyTorch setup on eight A6000 GPUs, each equipped with 48GB of memory.", + "bbox": [ + 511, + 869, + 903, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11373", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ccbc3c101323901de6489867423bb2d6175b2de6659246449d8a1ae12bd3d8a5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodInput Res.SkinHairNoseL-EyeR-EyeL-BrowR-BrowL-LipI-MouthU-LipMean F1 ↑
Wei et al. [105]51296.491.191.987.185.080.882.591.090.687.988.43
EHANet [55]51296.093.993.786.286.583.283.190.393.888.689.53
EAGRNet [91]47396.294.994.088.689.085.785.291.295.088.990.87
AGRNet [92]47396.587.693.988.789.185.585.691.192.089.189.91
FaRL_scratch [130]51296.294.993.889.089.085.385.490.091.788.190.34
DML-CSR [127]47395.794.593.989.489.685.585.791.091.889.190.62
FP-LIIF [80]25696.495.193.788.588.584.584.390.392.187.590.09
SwinFace [71]××××××××××××
QFace [88]××××××××××××
Facetor [72]51296.696.293.989.489.186.286.390.691.689.090.89
FaceXFormer22496.495.793.890.190.386.085.990.692.189.291.01
", + "bbox": [ + 91, + 88, + 906, + 266 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/445a9223ee193bc31351a5d5e93076b8f4075243bbe7971c6457e573a7aed881.jpg", + "table_caption": [ + "Table 3. Performance comparison for face parsing on the CelebAMask-HQ dataset [42]. The symbol $\\times$ indicates that the model does not perform the corresponding task. Red = First Best, Blue = Second Best. $\\times$ indicates a model that doesn't perform the task." + ], + "table_footnote": [], + "table_body": "
MethodsExpression (RAF-DB) ↑MethodsVisibility (COFW) ↑MethodsAge (MAE) UTKFace ↓
DLP-CNN [45]80.89RCPR [7]40OR-CNN [68]5.74
gACNN [48]85.07Wu et al. [111]44.43Axel Berg et al. [5]4.55
IPA2LT [120]86.77Wu et al. [110]49.11CORAL [8]5.47
RAN [101]86.90ECT [122]63.4Gustafsson et al. [26]4.65
CovPool [2]87.003DDE [94]63.89R50-SORD [70]4.36
SCN [100]87.03MNN [95]72.12VOLO-D1 [40]4.23
DACL [21]87.78DLDL-v2 [23]4.42
KTN [44]88.07MWR [84]4.37
DMUE [82]88.76
SwinFace [71]86.54SwinFace [71]×SwinFace [71]×
QFace [88]92.86QFace [88]×QFace [88]×
Facetor [72]87.58Facetor [72]×Facetor [72]4.10
FaceXFormer88.24FaceXFormer72.56FaceXFormer4.17
", + "bbox": [ + 93, + 323, + 478, + 463 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Performance comparison on facial expression recognition, face visibility prediction and age estimation.", + "bbox": [ + 89, + 474, + 482, + 502 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The models' backbones are initialized with ImageNet pretrained weights and processes input images at a resolution of $224 \\times 224$ . We employ the AdamW optimizer with a weight decay of $1e^{-5}$ . All models are trained for 12 epochs with a batch size of 48 on each GPU, and an initial learning rate of $1e^{-4}$ , which decays by a factor of 10 at the $6^{th}$ and $10^{th}$ epochs. For some tasks, we train the model for additional epochs to reach the upper ceiling of performance. For data augmentation, we randomly apply Gaussian blur, grayscale conversion, gamma correction, occlusion, horizontal flipping, and affine transformations, such as rotation, translation and scaling. The number of FaceX decoder $N$ is set to two. To ensure stable training across tasks when using multiple datasets of varying sample sizes, we equalize the representation of each task's samples in every batch through upsampling.", + "bbox": [ + 88, + 521, + 482, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Main results", + "text_level": 1, + "bbox": [ + 89, + 772, + 223, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Table 3, Table 4, Table 6, Table 5, we present a comparative analysis of FaceXFormer against recent methods across a variety of tasks. A key highlight of our work is its unique capability to deliver promising results across multiple tasks at real-time inference speed using a single unified model. Specifically, FaceXFormer achieves state-of-the-art performance in face parsing, with a mean F1 score of 91.01 on", + "bbox": [ + 89, + 794, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CelebAMaskHQ at a resolution of $224 \\times 224$ , which is half the input size required by other SOTA methods. Furthermore, it demonstrates superior performance in head pose estimation and landmark detection, achieving a mean MAE of 3.52 and a mean NME of 4.67, respectively. Additionally, FaceXFormer provides a significant performance boost in attributes prediction and visibility prediction, achieving an accuracy of $91.83\\%$ on the CelebA dataset and $72.56\\%$ on COFW. It also performs competitively in age estimation, achieving the second-best score of 4.17, and achieves an accuracy of $88.24\\%$ in facial expression recognition. In face recognition, FaceXFormer outperforms Faceptor, achieving a mean accuracy of $95.94\\%$ compared to $95.28\\%$ . However, we observe that multi-task models generally underperform compared to specialized ones in this task. This can be attributed to conflicting training objectives, which force the model to learn identity-invariant features rather than identity-specific representations crucial for accurate recognition. The results on gender estimation across different race categories is shown in Table 8.", + "bbox": [ + 511, + 325, + 906, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Recent models such as SwinFace [71], QFace [88], and Facetor [72] also aim to unify multiple tasks but only address a subset of them. These approaches often exclude complex tasks such as segmentation, head pose estimation, and landmark prediction. Moreover, they rely on multiple decoders and computationally expensive attention mechanisms, adding to the overall computational overhead. In contrast, FaceXFormer seamlessly unifies these complex tasks using a lightweight decoder and achieves state-of-the-art performance across them at a real-time FPS of 33.21. It outperforms previous multi-task models in segmentation, head pose estimation, landmark prediction, attribute prediction, and face visibility prediction, while achieving the second-best performance in age estimation. In this work, we simultaneously train for ten heterogeneous tasks, posing a more formidable challenge than previous approaches. Despite this, FaceXFormer effectively handles multiple tasks, achieving SOTA or competitive performance in real time.", + "bbox": [ + 511, + 628, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11374", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/28239e7818336f231c7111c68a2618ff4d7112993283064b55c377bb2fc7d4c8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsHeadpose (BIWI)MethodsLandmarks (300W)MethodsCelebA Acc. ↑
YawPitchRollMAE ↓Full ↓Com ↓Chal ↓
HopeNet [78]4.816.613.274.89LAB [109]3.492.985.19PANDA-1 [123]85.43
QuatNet [30]5.494.012.944.15Wing [22]4.043.277.18LNets+ANet [53]87.33
FSA-Net [116]4.275.492.934.14DeCaFa [14]3.392.935.26SSP+SSG [34]88.24
EVA-GCN [114]6.014.782.983.98HRNet [98]3.322.875.15MOON [77]90.94
TriNet [9]4.114.753.043.97PicassoNet [106]3.583.035.81NSA [56]90.61
img2pose [3]4.563.543.243.78AVS+SAN [18]3.863.216.46MCNN-AUX [28]91.29
MNN [95]3.984.612.393.66LUVLi [39]3.232.765.16MCFA [137]91.23
MFDNet [51]3.404.682.773.62HIH [41]3.092.654.89DMM-CNN [57]91.70
TokenHPE [121]3.954.512.713.72PIPNet [33]3.192.784.89SSPL [85]91.77
WHENet [131]3.994.393.063.81SLPT [112]3.172.754.90FaRL [130]91.39
SwinFace [71]××××SwinFace [71]×××SwinFace [71]91.38
QFace [88]----QFace [88]×××QFace [88]91.56
Facector [72]××××Facector [72]3.162.754.84Facector [72]91.39
FaceXFormer3.913.972.673.52FaceXFormer3.052.664.67FaceXFormer91.83
", + "bbox": [ + 102, + 88, + 893, + 325 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cd77275a25ef12d44f7b79a7e69ab57a087cc6190b571b273e63f4adcf785f88.jpg", + "table_caption": [ + "Table 5. Performance comparison on headpose, landmark detection, and attribute recognition. The symbol $\\times$ indicates that the model does not perform the corresponding task, while - denotes that results for this dataset are not provided. Red = First Best, Blue = Second Best." + ], + "table_footnote": [], + "table_body": "
MethodLFWCFP-FPAgeDBCALFWCPLFWMean ↑
CosFace [97]99.8198.1298.1195.7692.2896.81
ArcFace [15]99.8398.2798.2895.4592.0896.78
VPL-ArcFace [16]99.8399.1198.6096.1293.4597.42
AdaFace [37]99.8399.1198.1796.0293.9397.41
SwinFace [71]99.8798.6098.1596.1093.4297.22
QFace [88]××××××
Facetor [72]99.4096.3493.6594.7592.2795.28
FaceXFormer99.6896.7596.3595.5091.4195.94
", + "bbox": [ + 93, + 382, + 480, + 488 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This success can be attributed to the efficiency of the proposed lightweight decoder, which employs a novel bidirectional cross-attention mechanism.", + "bbox": [ + 89, + 527, + 482, + 574 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Qualitative \"in-the-wild\" results", + "text_level": 1, + "bbox": [ + 89, + 582, + 375, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we present the qualitative results of FaceXFormer on randomly selected \"in-the-wild\" images. We select five random images and showcase the results for face parsing, head pose estimation, landmarks prediction, age estimation, gender and race classification, and attributes prediction in Figure 3. Notably, the model successfully performs complex tasks such as face segmentation, head pose estimation, and landmark prediction, even when input samples exhibit extreme poses, occlusions, or blurring. Furthermore, FaceXFormer can be effectively used as a tool to generate multiple annotations for each image, making it valuable for various downstream tasks. These results highlight FaceXFormer's robust performance in challenging, real-world scenarios.", + "bbox": [ + 88, + 604, + 482, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 829, + 243, + 847 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we explore the impact of different components of FaceXFormer on performance. Additionally, we demonstrate that the proposed model exhibits minimal bias", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "compared to other models by evaluating age and gender prediction across various demographics. Furthermore, we analyze the impact of using different backbones of varying sizes on the FPS and performance.", + "bbox": [ + 511, + 386, + 906, + 446 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/36f4356896acb2034ecfc224541eadea0eedecb6f3f92554eb1705ef73b9bee1.jpg", + "table_caption": [ + "Table 6. Performance comparison for face recognition." + ], + "table_footnote": [], + "table_body": "
MLP FusionSelf Attn.Cross Attn.Bi-dir. CAHPE (MAE)Lnd (NME)Attr. (Acc.)Age (MAE)
✓✓3.704.7091.214.21
✓✓17.1631.4979.9016.15
✓✓✓4.635.3089.984.84
✓✓✓3.524.6791.834.17
", + "bbox": [ + 514, + 457, + 903, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 7. Impact of various components on performance.", + "bbox": [ + 540, + 549, + 877, + 564 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Impact of various components in FaceXFormer", + "text_level": 1, + "bbox": [ + 511, + 575, + 906, + 593 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate the contribution of each component in FaceXFormer, we conduct an ablation study focusing on the importance of specific design choices and their impact on performance across various tasks. The results of these experiments are summarized in Table 7. We observe that without MLP fusion (row 1), there is a drop in performance, highlighting the importance of integrating multi-scale features to capture both global and local information essential for accurate predictions. The model performs extremely poorly (row 2) without cross-attention in the decoder, which is expected, as there is no interaction between face tokens and task tokens in this case. Introducing the proposed bidirectional cross-attention (row 4), which corresponds to FaceXFormer, in the decoder provides a significant boost compared to using standard cross-attention (row 3), yielding improvements of 1.11 MAE in head pose estimation, 0.63 NME in landmark detection, 1.85 accuracy points in attribute prediction, and 0.67 MAE in age estimation. These results demonstrate the importance of MLP fusion and bidirectional cross-attention in the FaceXFormer architecture.", + "bbox": [ + 511, + 598, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11375", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b851fa0fb1c9084a6105526b5b8f8927c62ace62808298c4e17514eb5a478a32.jpg", + "image_caption": [ + "Figure 3. FaceXFormer predictions on \"in-the-wild\" images" + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 480, + 422 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/aaccb12e6da01b590f355c10cad07152429759d6ef2327edb8cce80f3e1cc50d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelData PointsWhiteNon-whiteAverageDiscrepancy
AgeFairFace100K60.0560.6360.520.58
CLIP400M62.2561.9562.00-0.30
FaRL20M61.4961.8461.780.35
FaceXFormer5.9M58.9459.4459.340.50
GenderFairFace100K94.1594.4194.360.26
CLIP400M94.8795.7895.610.91
FaRL20M95.1695.7795.650.61
FaceXFormer5.9M95.3495.1995.22-0.09
", + "bbox": [ + 94, + 460, + 482, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Bias Analysis and Ethical Considerations", + "text_level": 1, + "bbox": [ + 89, + 617, + 442, + 632 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In our work, we utilize 17 unique datasets for training and evaluation. We obtained these datasets following the procedures stated on their respective pages and signed the license agreements if and when necessary. As we train our models on multiple datasets designed for different tasks, the subjects across different age groups, genders, and races is not equal. This imbalance may introduce bias in the model. Therefore, we provide an analysis using the FairFace [35] dataset, which is balanced in terms of age, gender and race. We follow [73] and define the \"Non-white\" group to include multiple racial categories: \"Black\", \"Indian\", \"East Asian\", \"Southeast Asian\", \"Middle Eastern\" and \"Latino\". As can be seen from Table 8, FaceXFormer shows the smallest performance discrepancy across different racial groups and exhibits minimal bias compared to other models despite being trained on fewer data points. This can be attributed to race estimation being the task in co-training.", + "bbox": [ + 88, + 643, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Performance under different backbones", + "text_level": 1, + "bbox": [ + 511, + 90, + 856, + 104 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To evaluate the impact of different backbones on performance and FPS, we conduct an ablation study comparing various backbone architectures in FaceXFormer. We categorize head pose estimation, landmark prediction, and age estimation as regression (Reg) tasks, while attribute prediction and facial expression recognition fall under classification (Cls). Additionally, face parsing is denoted as segmentation (Seg). The results are summarized in Table 9.", + "bbox": [ + 511, + 114, + 903, + 234 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3d5cbb5be63d082597a5d40787c839c76ff5066d40cd22ad4beea733084b03a0.jpg", + "table_caption": [ + "Table 8. Age and gender accuracy w.r.t race groups on FairFace" + ], + "table_footnote": [], + "table_body": "
BackboneSegRegClsFPSParams
MobileNet90.214.6488.2239.7625.32
ResNet10190.494.3788.9134.9865.54
ConvNext-B91.084.3589.0936.61110.19
Swin-B91.014.1290.0333.21109.29
", + "bbox": [ + 532, + 251, + 887, + 337 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 9. Effect of different backbones on performance and FPS.", + "bbox": [ + 519, + 347, + 898, + 361 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "From the results, we observe that ConvNeXt achieves the best performance in segmentation with an F1 score of $91.08\\%$ . The Swin Transformer backbone excels in both regression and classification tasks, with a mean error of 4.12 and a mean accuracy of $90.03\\%$ , respectively. In contrast, MobileNet demonstrates the lowest performance metrics, including an F1 score of $90.21\\%$ and a mean error of 4.64, highlighting its limitations in handling larger, more complex datasets due to its smaller receptive field compared to the Swin Transformer. The selection of the Swin Transformer as the backbone for FaceXFormer is driven by its superior scalability and global contextual understanding, both of which are essential for facial analysis tasks.", + "bbox": [ + 511, + 383, + 906, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 601, + 633, + 616 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "FaceXFormer introduces a novel end-to-end unified model that efficiently handles a wide range of facial analysis tasks in real-time. By adopting a transformer-based encoder-decoder architecture and representing each task as a learnable token, our approach seamlessly integrates multiple tasks within a single framework while maintaining minimal computational cost and fast inference times. The proposed lightweight decoder, FaceX, incorporates a novel bi-directional cross-attention mechanism, enhancing the model's ability to learn robust and generalized face representations across diverse tasks. Comprehensive experiments demonstrate that FaceXFormer achieves state-of-the-art performance across multiple facial analysis tasks, achieving a real-time FPS of 33.21. In broader applications, FaceXFormer can serve as an annotator for large-scale face datasets and can be integrated into existing facial analysis systems to provide extra information, making it a valuable tool for surveillance, subject analysis, and image retrieval.", + "bbox": [ + 511, + 628, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11376", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This research is based upon work supported in part by the Office of the Director of National Intelligence (ODNI), Intelligence Advanced Research Projects Activity (IARPA), via [2022-21102100005]. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies, either expressed or implied, of ODNI, IARPA, or the U.S. Government. The US Government is authorized to reproduce and distribute reprints for governmental purposes notwithstanding any copyright annotation therein.", + "bbox": [ + 89, + 113, + 483, + 255 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 281, + 187, + 297 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] B Abirami, TS Subashini, and V Mahavaishnavi. Gender and age prediction from real time facial images using cnn. Materials Today: Proceedings, 33:4708-4712, 2020. 1, 2", + "[2] Dinesh Acharya, Zhiwu Huang, Danda Pani Paudel, and Luc Van Gool. Covariance pooling for facial expression recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 367-374, 2018. 6", + "[3] Vitor Albiero, Xingyu Chen, Xi Yin, Guan Pang, and Tal Hassner. img2pose: Face alignment and detection via 6dof, face pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7617-7627, 2021. 2, 7", + "[4] Xiang An, Xuhan Zhu, Yuan Gao, Yang Xiao, Yongle Zhao, Ziyong Feng, Lan Wu, Bin Qin, Ming Zhang, Debing Zhang, et al. Partial fc: Training 10 million identities on a single machine. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1445-1449, 2021. 5", + "[5] Axel Berg, Magnus Oskarsson, and Mark O'Connor. Deep ordinal regression with label diversity. In 2020 25th international conference on pattern recognition (ICPR), pages 2740-2747. IEEE, 2021. 6", + "[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3", + "[7] Xavier P Burgos-Artizzu, Pietro Perona, and Piotr Dólar. Robust face landmark estimation under occlusion. In Proceedings of the IEEE international conference on computer vision, pages 1513-1520, 2013. 5, 6", + "[8] Wenzhi Cao, Vahid Mirjalili, and Sebastian Raschka. Rank consistent ordinal regression for neural networks with application to age estimation. Pattern Recognition Letters, 140:325-331, 2020. 1, 2, 6", + "[9] Zhiwen Cao, Zongcheng Chu, Dongfang Liu, and Yingjie Chen. A vector-based representation to enhance head pose estimation. In Proceedings of the IEEE/CVF Winter Conference on applications of computer vision, pages 1188-1197, 2021. 2, 7" + ], + "bbox": [ + 107, + 306, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3, 4", + "[11] Liang-Chieh Chen, Yi Yang, Jiang Wang, Wei Xu, and Alan L Yuille. Attention to scale: Scale-aware semantic image segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3640-3649, 2016. 2", + "[12] Alejandro Cobo, Roberto Valle, José M Buenapasada, and Luis Baumela. On the representation and methodology for wide and short range head pose estimation. Pattern Recognition, 149:110263, 2024. 1, 2", + "[13] Kaiwen Cui, Rongliang Wu, Fangneng Zhan, and Shijian Lu. Face transformer: Towards high fidelity and accurate face swapping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 668-677, 2023. 1, 2", + "[14] Arnaud Dapogny, Kevin Bailly, and Matthieu Cord. Decafa: Deep convolutional cascade for face alignment in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6893-6901, 2019. 7", + "[15] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 2, 5, 7", + "[16] Jiankang Deng, Jia Guo, Jing Yang, Alexandros Lattas, and Stefanos Zafeiriou. Variational prototype learning for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11906-11915, 2021. 7", + "[17] Xing Di, Yiyu Zheng, Xiaoming Liu, and Yu Cheng. Pros: Facial omni-representation learning via prototype-based self-distillation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6087-6098, 2024. 2", + "[18] Xuanyi Dong, Yan Yan, Wanli Ouyang, and Yi Yang. Style aggregated network for facial landmark detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 379-388, 2018. 7", + "[19] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.3", + "[20] Gabriele Fanelli, Matthias Dantone, Juergen Gall, Andrea Fossati, and Luc Van Gool. Random forests for real time 3d face analysis. International journal of computer vision, 101:437-458, 2013. 5", + "[21] Amir Hossein Farzaneh and Xiaojun Qi. Facial expression recognition in the wild via deep attentive center loss. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2402-2411, 2021. 6", + "[22] Zhen-Hua Feng, Josef Kittler, Muhammad Awais, Patrik Huber, and Xiao-Jun Wu. Wing loss for robust facial land" + ], + "bbox": [ + 522, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11377", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "mark localisation with convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2, 7", + "[23] Bin-Bin Gao, Xin-Xin Liu, Hong-Yu Zhou, Jianxin Wu, and Xin Geng. Learning expectation of label distribution for facial age and attractiveness estimation. arXiv preprint arXiv:2007.01771, 2020. 2, 6", + "[24] Asma El Kissi Ghalleb, Safa Boumaiza, and Najoua Essoukri Ben Amara. Demographic face profiling based on age, gender and race. In 2020 5th International Conference on Advanced Technologies for Signal and Image Processing (ATSIP), pages 1-6. IEEE, 2020. 1, 2", + "[25] Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. Ms-celeb-1m: A dataset and benchmark for large-scale face recognition. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 87-102. Springer, 2016. 5", + "[26] Fredrik K Gustafsson, Martin Danelljan, Goutam Bhat, and Thomas B Schon. Energy-based models for deep probabilistic regression. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16, pages 325-343. Springer, 2020. 6", + "[27] Hu Han, Anil K Jain, Fang Wang, Shiguang Shan, and Xilin Chen. Heterogeneous face attribute estimation: A deep multi-task learning approach. IEEE transactions on pattern analysis and machine intelligence, 40(11):2597-2609, 2017. 2", + "[28] Emily Hand and Rama Chellappa. Attributes for improved attributes: A multi-task network utilizing implicit and explicit relationships for facial attribute classification. In Proceedings of the AAAI conference on artificial intelligence, 2017. 7", + "[29] Hui-Lan Hsieh, Winston Hsu, and Yan-Ying Chen. Multi-task learning for face identification and attribute estimation. In 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2981-2985, 2017. 3", + "[30] Heng-Wei Hsu, Tung-Yu Wu, Sheng Wan, Wing Hung Wong, and Chen-Yi Lee. Quatnet: Quaternion-based head pose estimation with multiregression loss. IEEE Transactions on Multimedia, 21(4):1035-1046, 2018. 7", + "[31] Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. In Workshop on faces in 'Real-Life' Images: detection, alignment, and recognition, 2008. 5", + "[32] Aaron S Jackson, Michel Valstar, and Georgios Tzimiropoulos. A cnn cascade for landmark guided semantic part segmentation. In Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part III 14, pages 143-155. Springer, 2016. 1, 2", + "[33] Haibo Jin, Shengcai Liao, and Ling Shao. Pixel-in-pixel net: Towards efficient facial landmark detection in the wild. International Journal of Computer Vision, 129(12):3174-3194, 2021. 7" + ], + "bbox": [ + 99, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Mahdi M Kalayeh, Boqing Gong, and Mubarak Shah. Improving facial attribute prediction using semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6942-6950, 2017. 2, 7", + "[35] Kimmo Karkkainen and Jungseock Joo. Fairface: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1548-1558, 2021. 2, 8", + "[36] Kimmo Karkkainen and Jungseock Joo. Fairface: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1548-1558, 2021. 5", + "[37] Minchul Kim, Anil K Jain, and Xiaoming Liu. Adaface: Quality adaptive margin for face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18750-18759, 2022. 1, 2, 7", + "[38] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 3", + "[39] Abhinav Kumar, Tim K Marks, Wenxuan Mou, Ye Wang, Michael Jones, Anoop Cherian, Toshiaki Koike-Akino, Xiaoming Liu, and Chen Feng. Luvli face alignment: Estimating landmarks' location, uncertainty, and visibility likelihood. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8236-8246, 2020. 1, 2, 7", + "[40] Maksim Kuprashevich and Irina Tolstykh. Mivolo: Multi-input transformer for age and gender estimation. arXiv preprint arXiv:2307.04616, 2023. 2, 6", + "[41] Xing Lan, Qinghao Hu, Qiang Chen, Jian Xue, and Jian Cheng. Hih: Towards more accurate face alignment via heatmap in heatmap. arXiv preprint arXiv:2104.03100, 2021.7", + "[42] Cheng-Han Lee, Ziwei Liu, Lingyun Wu, and Ping Luo. Maskgan: Towards diverse and interactive facial image manipulation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 5, 6", + "[43] Gil Levi and Tal Hassner. Age and gender classification using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 34-42, 2015. 2", + "[44] Hangyu Li, Nannan Wang, Xinpeng Ding, Xi Yang, and Xinbo Gao. Adaptively learning facial expression representation via cf labels and distillation. IEEE Transactions on Image Processing, 30:2016-2028, 2021. 2, 6", + "[45] Shan Li, Weihong Deng, and JunPing Du. Reliable crowdsourcing and deep locality-preserving learning for expression recognition in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2852-2861, 2017. 2, 5, 6", + "[46] Wanhua Li, Xiaoke Huang, Jiwen Lu, Jianjiang Feng, and Jie Zhou. Learning probabilistic ordinal embeddings for uncertainty-aware regression. In Proceedings of the" + ], + "bbox": [ + 522, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11378", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF conference on computer vision and pattern recognition, pages 13896-13905, 2021. 1, 2", + "[47] Xiangtai Li, Haobo Yuan, Wei Li, Henghui Ding, Size Wu, Wenwei Zhang, Yining Li, Kai Chen, and Chen Change Loy. Omg-seg: Is one model good enough for all segmentation? arXiv preprint arXiv:2401.10229, 2024. 3", + "[48] Yong Li, Jiabei Zeng, Shiguang Shan, and Xilin Chen. Occlusion aware facial expression recognition using cnn with attention mechanism. IEEE Transactions on Image Processing, 28(5):2439-2450, 2018. 6", + "[49] Chunze Lin, Beier Zhu, Quan Wang, Renjie Liao, Chen Qian, Jiwen Lu, and Jie Zhou. Structure-coherent deep feature learning for robust face alignment. IEEE Transactions on Image Processing, 30:5313-5326, 2021. 1, 2", + "[50] Jinpeng Lin, Hao Yang, Dong Chen, Ming Zeng, Fang Wen, and Lu Yuan. Face parsing with roi tanh-warping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5654-5663, 2019. 2", + "[51] Hai Liu, Shuai Fang, Zhaoli Zhang, Duantengchuan Li, Ke Lin, and Jiazhang Wang. Mfdnet: Collaborative poses perception and matrix fisher distribution for head pose estimation. IEEE Transactions on Multimedia, 24:2449-2460, 2021. 7", + "[52] S Liu, L Zhang, X Yang, H Su, and J Zhu. Query2label: A simple transformer way to multi-label classification. arxiv 2021. arXiv preprint arXiv:2107.10834, 2021. 3", + "[53] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3730-3738, 2015. 7", + "[54] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), 2015. 5", + "[55] Ling Luo, Dingyu Xue, and Xinglong Feng. Ehanet: An effective hierarchical aggregation network for face parsing. Applied Sciences, 10(9):3135, 2020. 6", + "[56] Upal Mahbub, Sayantan Sarkar, and Rama Chellappa. Segment-based methods for facial attribute detection from partial faces. IEEE Transactions on Affective Computing, 11(4):601-613, 2018. 7", + "[57] Longbiao Mao, Yan Yan, Jing-Hao Xue, and Hanzi Wang. Deep multi-task multi-label cnn for effective facial attribute classification. IEEE Transactions on Affective Computing, 13(2):818-828, 2020. 7", + "[58] Chen Mi, Baoxi Yuan, Peng Ma, Yingxia Guo, Le Qi, Feng Wang, Wenbo Wu, and Lingling Wang. Visibility prediction based on landmark detection in foggy weather. In 2020 International Conference on Robots & Intelligent System (ICRIS), pages 134-137, 2020. 1, 2", + "[59] Paul Micaelli, Arash Vahdat, Hongxu Yin, Jan Kautz, and Pavlo Molchanov. Recurrence without recurrence: Stable video landmark detection with deep equilibrium models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22814-22825, 2023. 2" + ], + "bbox": [ + 99, + 92, + 482, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[60] Zuheng Ming, Junshi Xia, Muhammad Muzzamil Luqman, Jean-Christophe Burie, and Kaixing Zhao. Dynamic multitask learning for face recognition with facial expression. arXiv preprint arXiv:1911.03281, 2019. 3", + "[61] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE transactions on pattern analysis and machine intelligence, 41(8):1979–1993, 2018. 1, 2", + "[62] Ali Mollahosseini, Behzad Hasani, and Mohammad H Ma-hoor. Affectnet: A database for facial expression, valence, and arousal computing in the wild. IEEE Transactions on Affective Computing, 10(1):18-31, 2017. 5", + "[63] Stylianos Moschoglou, Athanasios Papaioannou, Christos Sagonas, Jiankang Deng, Irene Kotsia, and Stefanos Zafeiriou. Agedb: the first manually collected, in-the-wild age database. In proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 51-59, 2017. 5", + "[64] Erik Murphy-Chutorian, Anup Doshi, and Mohan Manubhai Trivedi. Head pose estimation for driver assistance systems: A robust algorithm and experimental evaluation. In 2007 IEEE intelligent transportation systems conference, pages 709-714. IEEE, 2007. 2", + "[65] Kartik Narayan, Harsh Agarwal, Kartik Thakral, Surbhi Mittal, Mayank Vatsa, and Richa Singh. Df-platter: Multiface heterogeneous deepfake dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9739-9748, 2023. 2", + "[66] Kartik Narayan, Nithin Gopalakrishnan Nair, Jennifer Xu, Rama Chellappa, and Vishal M Patel. Petalface: Parameter efficient transfer learning for low-resolution face recognition. arXiv preprint arXiv:2412.07771, 2024. 2", + "[67] Kartik Narayan, Vibashan VS, and Vishal M Patel. Segface: Face segmentation of long-tail classes. arXiv preprint arXiv:2412.08647, 2024. 2", + "[68] Zhenxing Niu, Mo Zhou, Le Wang, Xinbo Gao, and Gang Hua. Ordinal regression with multiple output cnn for age estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4920-4928, 2016. 6", + "[69] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European conference on computer vision, pages 69-84. Springer, 2016. 1, 2", + "[70] Jakub Paplham, Vojt Franc, et al. A call to reflect on evaluation practices for age estimation: Comparative analysis of the state-of-the-art and a unified benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1196-1205, 2024. 6", + "[71] Lixiong Qin, Mei Wang, Chao Deng, Ke Wang, Xi Chen, Jiani Hu, and Weihong Deng. Swinface: a multi-task transformer for face recognition, expression recognition, age estimation and attribute estimation. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 2, 6, 7", + "[72] Lixiong Qin, Mei Wang, Xuannan Liu, Yuhang Zhang, Wei Deng, Xiaoshuai Song, Weiran Xu, and Weihong Deng." + ], + "bbox": [ + 522, + 92, + 906, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11379", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Facetor: A generalist model for face perception. In European Conference on Computer Vision, pages 240-260. Springer, 2025. 2, 3, 6, 7", + "[73] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 8", + "[74] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 3", + "[75] Rajeev Ranjan, Vishal M Patel, and Rama Chellappa. Hyperface: A deep multi-task learning framework for face detection, landmark localization, pose estimation, and gender recognition. IEEE transactions on pattern analysis and machine intelligence, 41(1):121-135, 2017. 2, 3", + "[76] Rajeev Ranjan, Swami Sankaranarayanan, Carlos D Castillo, and Rama Chellappa. An all-in-one convolutional neural network for face analysis. In 2017 12th IEEE international conference on automatic face & gesture recognition (FG 2017), pages 17-24. IEEE, 2017. 2, 3", + "[77] Ethan M Rudd, Manuel Günther, and Terrance E Boult. Moon: A mixed objective optimization network for the recognition of facial attributes. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14, pages 19-35. Springer, 2016. 7", + "[78] Nataniel Ruiz, Eunjji Chong, and James M Rehg. Fine-grained head pose estimation without keypoints. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 2074-2083, 2018. 7", + "[79] Christos Sagonas, Georgios Tzimiropoulos, Stefanos Zafeiriou, and Maja Pantic. 300 faces in-the-wild challenge: The first facial landmark localization challenge. In Proceedings of the IEEE international conference on computer vision workshops, pages 397-403, 2013. 5", + "[80] Mausoom Sarkar, Mayur Hemani, Rishabh Jain, Balaji Krishnamurthy, et al. Parameter efficient local implicit image function network for face segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20970-20980, 2023. 2, 6", + "[81] S. Sengupta, J.C. Cheng, C.D. Castillo, V.M. Patel, R. Chellappa, and D.W. Jacobs. Frontal to profile face verification in the wild. In IEEE Conference on Applications of Computer Vision, 2016. 5", + "[82] Jiahui She, Yibo Hu, Hailin Shi, Jun Wang, Qiu Shen, and Tao Mei. Dive into ambiguity: Latent distribution mining and pairwise uncertainty estimation for facial expression recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6248-6257, 2021. 1, 2, 6", + "[83] Jie Shen, Stefanos Zafeiriou, Grigoris G Chrysos, Jean Kossaifi, Georgios Tzimiropoulos, and Maja Pantic. The first facial landmark tracking in-the-wild challenge: Benchmark" + ], + "bbox": [ + 99, + 92, + 482, + 898 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and results. In Proceedings of the IEEE international conference on computer vision workshops, pages 50-58, 2015. 5", + "[84] Nyeong-Ho Shin, Seon-Ho Lee, and Chang-Su Kim. Moving window regression: A novel approach to ordinal regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18760-18769, 2022. 6", + "[85] Ying Shu, Yan Yan, Si Chen, Jing-Hao Xue, Chunhua Shen, and Hanzi Wang. Learning spatial-semantic relationship for facial attribute recognition with limited labeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11916–11925, 2021. 2, 7", + "[86] Dominykas Strazdas, Jan Hintz, and Ayoub Al-Hamadi. Robo-hud: Interaction concept for contactless operation of industrial cobotic systems. Applied Sciences, 11(12):5366, 2021. 2", + "[87] Carole H Sudre, Wenqi Li, Tom Vercauteren, Sebastian Ourselin, and M Jorge Cardoso. Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: Third International Workshop, DLMIA 2017, and 7th International Workshop, ML-CDS 2017, Held in conjunction with MICCAI 2017, Quebec City, QC, Canada, September 14, Proceedings 3, pages 240-248. Springer, 2017. 5", + "[88] Haomiao Sun, Mingjie He, Shiguang Shan, Hu Han, and Xilin Chen. Task-adaptive q-face. arXiv preprint arXiv:2405.09059, 2024. 2, 3, 6, 7", + "[89] Yi Sun, Yuheng Chen, Xiaogang Wang, and Xiaou Tang. Deep learning face representation by joint identification-verification. Advances in neural information processing systems, 27, 2014. 1, 2", + "[90] Yaniv Taigman, Ming Yang, Marc'Aurelio Ranzato, and Lior Wolf. Deepface: Closing the gap to human-level performance in face verification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1701-1708, 2014. 1, 2", + "[91] Gusi Te, Yinglu Liu, Wei Hu, Hailin Shi, and Tao Mei. Edge-aware graph representation learning and reasoning for face parsing. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16, pages 258-274. Springer, 2020. 6", + "[92] Gusi Te, Wei Hu, Yinglu Liu, Hailin Shi, and Tao Mei. Agrnet: Adaptive graph representation learning and reasoning for face parsing. IEEE Transactions on Image Processing, 30:8236-8250, 2021. 6", + "[93] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3", + "[94] Roberto Valle, José M Buenapasada, Antonio Valdés, and Luis Baumela. Face alignment using a 3d deeply-initialized ensemble of regression trees. Computer Vision and Image Understanding, 189:102846, 2019. 2, 6" + ], + "bbox": [ + 524, + 92, + 903, + 898 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "11380", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[95] Roberto Valle, José M Buenapasada, and Luis Baumela. Multi-task head pose estimation in-the-wild. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43 (8):2874-2881, 2020. 2, 6, 7", + "[96] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3", + "[97] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5265-5274, 2018. 2, 7", + "[98] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 43(10):3349-3364, 2020. 2, 7", + "[99] Junke Wang, Dongdong Chen, Zuxuan Wu, Chong Luo, Luowei Zhou, Yucheng Zhao, Yujia Xie, Ce Liu, Yu-Gang Jiang, and Lu Yuan. Omnivl: One foundation model for image-language and video-language tasks. Advances in neural information processing systems, 35:5696-5710, 2022. 3", + "[100] Kai Wang, Xiaojiang Peng, Jianfei Yang, Shijian Lu, and Yu Qiao. Suppressing uncertainties for large-scale facial expression recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6897-6906, 2020. 6", + "[101] Kai Wang, Xiaojiang Peng, Jianfei Yang, Debin Meng, and Yu Qiao. Region attention networks for pose and occlusion robust facial expression recognition. IEEE Transactions on Image Processing, 29:4057-4069, 2020. 6", + "[102] Qianqian Wang, Yen-Yu Chang, Ruojin Cai, Zhengqi Li, Bharath Hariharan, Aleksander Holynski, and Noah Snavely. Tracking everything everywhere all at once. arXiv preprint arXiv:2306.05422, 2023. 3", + "[103] Zhenyu Wang, Yali Li, Xi Chen, Ser-Nam Lim, Antonio Torralba, Hengshuang Zhao, and Shengjin Wang. Detecting everything in the open world: Towards universal object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11433-11443, 2023. 3", + "[104] Zhen Wei, Yao Sun, Jinqiao Wang, Hanjiang Lai, and Si Liu. Learning adaptive receptive fields for deep image parsing network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2434-2442, 2017. 1", + "[105] Zhen Wei, Si Liu, Yao Sun, and Hefei Ling. Accurate facial image parsing at real-time speed. IEEE Transactions on Image Processing, 28(9):4659-4670, 2019. 6", + "[106] Tiancheng Wen, Zhonggan Ding, Yongqiang Yao, Yaxiong Wang, and Xueming Qian. Picassonet: Searching adaptive architecture for efficient facial landmark localization. IEEE Transactions on Neural Networks and Learning Systems, 34(12):10516-10527, 2023. 7" + ], + "bbox": [ + 93, + 92, + 482, + 898 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[107] Lior Wolf, Tal Hassner, and Yaniv Taigman. Effective unconstrained face recognition by combining multiple descriptors and learned background statistics. IEEE transactions on pattern analysis and machine intelligence, 33(10): 1978-1990, 2010. 5", + "[108] Erroll Wood, Tadas Baltrusaitis, Charlie Hewitt, Matthew Johnson, Jingjing Shen, Nikola Milosavljevic, Daniel Wilde, Stephan Garbin, Toby Sharp, Ivan Stojiljković, et al. 3d face reconstruction with dense landmarks. In European Conference on Computer Vision, pages 160-177. Springer, 2022. 1, 2", + "[109] Wayne Wu, Chen Qian, Shuo Yang, Quan Wang, Yici Cai, and Qiang Zhou. Look at boundary: A boundary-aware face alignment algorithm. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2129-2138, 2018. 7", + "[110] Yue Wu and Qiang Ji. Robust facial landmark detection under significant head poses and occlusion. In Proceedings of the IEEE International Conference on Computer Vision, pages 3658-3666, 2015. 6", + "[111] Yue Wu, Chao Gou, and Qiang Ji. Simultaneous facial landmark detection, pose and deformation estimation under facial occlusion. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3471-3480, 2017. 6", + "[112] Jiahao Xia, Weiwei Qu, Wenjian Huang, Jianguo Zhang, Xi Wang, and Min Xu. Sparse local patch transformer for robust face alignment and landmarks inherent relation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4052-4061, 2022. 7", + "[113] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. 3", + "[114] Miao Xin, Shentong Mo, and Yuanze Lin. Eva-gcn: Head pose estimation based on graph convolutional networks. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 1462-1471, 2021. 7", + "[115] Xinchen Yan, Jimei Yang, Kihyuk Sohn, and Honglak Lee. Attribute2image: Conditional image generation from visual attributes. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 776-791. Springer, 2016. 1, 2", + "[116] Tsun-Yi Yang, Yi-Ting Chen, Yen-Yu Lin, and Yung-Yu Chuang. Fsa-net: Learning fine-grained structure aggregation for head pose estimation from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1087-1096, 2019. 7", + "[117] Xiangnan Yin, Di Huang, Zehua Fu, Yunhong Wang, and Liming Chen. Segmentation-reconstruction-guided facial image de-occlusion. In 2023 IEEE 17th International Conference on Automatic Face and Gesture Recognition (FG), pages 1–8. IEEE, 2023. 1, 2", + "[118] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang," + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "11381", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 3", + "[119] Alireza Zaeemzadeh, Shabnam Ghadar, Baldo Faieta, Zhe Lin, Nazanin Rahnavard, Mubarak Shah, and Ratheesh Kalarot. Face image retrieval with attribute manipulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12116-12125, 2021. 1, 2", + "[120] Jiabei Zeng, Shiguang Shan, and Xilin Chen. Facial expression recognition with inconsistently annotated datasets. In Proceedings of the European conference on computer vision (ECCV), pages 222-237, 2018. 6", + "[121] Cheng Zhang, Hai Liu, Yongjian Deng, Bochen Xie, and Youfu Li. Tokenhpe: Learning orientation tokens for efficient head pose estimation via transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8897-8906, 2023. 2, 5, 7", + "[122] Hongwen Zhang, Qi Li, Zhenan Sun, and Yunfan Liu. Combining data-driven and model-driven methods for robust facial landmark detection. IEEE Transactions on Information Forensics and Security, 13(10):2409-2422, 2018. 6", + "[123] Ning Zhang, Manohar Paluri, Marc'Aurelio Ranzato, Trevor Darrell, and Lubomir Bourdev. Panda: Pose aligned networks for deep attribute modeling. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1637-1644, 2014. 7", + "[124] Zhanpeng Zhang, Ping Luo, Chen Change Loy, and Xiaou Tang. Facial landmark detection by deep multi-task learning. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VI 13, pages 94-108. Springer, 2014. 3", + "[125] Zhifei Zhang, Yang Song, and Hairong Qi. Age progression/regression by conditional adversarial autoencoder. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5810-5818, 2017. 5", + "[126] Rui Zhao, Tianshan Liu, Jun Xiao, Daniel PK Lun, and Kin-Man Lam. Deep multi-task learning for facial expression recognition and synthesis based on selective feature sharing. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 4412-4419. IEEE, 2021. 3", + "[127] Qingping Zheng, Jiankang Deng, Zheng Zhu, Ying Li, and Stefanos Zafeiriou. Decoupled multi-task learning with cyclical self-regulation for face parsing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4156-4165, 2022. 2, 6", + "[128] Tianyue Zheng and Weihong Deng. Cross-pose lfw: A database for studying cross-pose face recognition in unconstrained environments. Beijing University of Posts and Telecommunications, Tech. Rep, 5(7):5, 2018. 5", + "[129] Tianyue Zheng, Weihong Deng, and Jiani Hu. Crossage lfw: A database for studying cross-age face recognition in unconstrained environments. arXiv preprint arXiv:1708.08197, 2017. 5", + "[130] Yinglin Zheng, Hao Yang, Ting Zhang, Jianmin Bao, Dongdong Chen, Yangyu Huang, Lu Yuan, Dong Chen, Ming Zeng, and Fang Wen. General facial representation learning in a visual-linguistic manner. In Proceedings of the" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 13 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18697-18709, 2022. 2, 6, 7", + "[131] Yijun Zhou and James Gregson. Whenet: Real-time fine-grained estimation for wide range head pose. arXiv preprint arXiv:2005.10353, 2020. 1, 2, 7", + "[132] Zhenglin Zhou, Huaxia Li, Hong Liu, Nanyang Wang, Gang Yu, and Rongrong Ji. Star loss: Reducing semantic ambiguity in facial landmark detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15475-15484, 2023. 1, 2, 5", + "[133] Jiawen Zhu, Simiao Lai, Xin Chen, Dong Wang, and Huchuan Lu. Visual prompt multi-modal tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9516-9526, 2023. 3", + "[134] Peihao Zhu, Rameen Abdal, Yipeng Qin, and Peter Wonka. Sean: Image synthesis with semantic region-adaptive normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5104-5113, 2020. 1, 2", + "[135] Xiangyu Zhu, Zhen Lei, Xiaoming Liu, Hailin Shi, and Stan Z Li. Face alignment across large poses: A 3d solution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 146-155, 2016. 5", + "[136] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 4", + "[137] Ni Zhuang, Yan Yan, Si Chen, and Hanzi Wang. Multi-task learning of cascaded cnn for facial attribute classification. In 2018 24th International Conference on Pattern Recognition (ICPR), pages 2069-2074. IEEE, 2018. 2, 7", + "[138] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. Advances in Neural Information Processing Systems, 36, 2024. 3" + ], + "bbox": [ + 516, + 92, + 906, + 587 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "11382", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 13 + } +] \ No newline at end of file diff --git a/2025/FaceXFormer_ A Unified Transformer for Facial Analysis/43f275cd-265b-414a-a63d-4d3a6e26c5b2_model.json b/2025/FaceXFormer_ A Unified Transformer for Facial Analysis/43f275cd-265b-414a-a63d-4d3a6e26c5b2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2708df48b086e02515398b51177bd22e909ef0bf --- /dev/null +++ b/2025/FaceXFormer_ A Unified Transformer for Facial Analysis/43f275cd-265b-414a-a63d-4d3a6e26c5b2_model.json @@ -0,0 +1,3220 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.21, + 0.131, + 0.788, + 0.154 + ], + "angle": 0, + "content": "FaceXFormer: A Unified Transformer for Facial Analysis" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.181, + 0.777, + 0.234 + ], + "angle": 0, + "content": "Kartik Narayan* Vibashan VS* Rama Chellappa Vishal M. Patel {knaraya4, vvishnu2, rchella4, vpatel36}@jhu.edu https://kartik-3004.github.io/facexformer/" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.303, + 0.486, + 0.711 + ], + "angle": 0, + "content": "In this work, we introduce FaceXFormer, an end-to-end unified transformer model capable of performing ten facial analysis tasks within a single framework. These tasks include face parsing, landmark detection, head pose estimation, attribute prediction, age, gender, and race estimation, facial expression recognition, face recognition, and face visibility. Traditional face analysis approaches rely on task-specific architectures and pre-processing techniques, limiting scalability and integration. In contrast, FaceXFormer employs a transformer-based encoder-decoder architecture, where each task is represented as a learnable token, enabling seamless multi-task processing within a unified model. To enhance efficiency, we introduce FaceX, a lightweight decoder with a novel bi-directional cross-attention mechanism, which jointly processes face and task tokens to learn robust and generalized facial representations. We train FaceXFormer on ten diverse face perception datasets and evaluate it against both specialized and multi-task models across multiple benchmarks, demonstrating state-of-the-art or competitive performance. Additionally, we analyze the impact of various components of FaceXFormer on performance, assess real-world robustness in \"in-the-wild\" settings, and conduct a performance comparison under different backbones. To the best of our knowledge, FaceXFormer is the first model capable of handling ten facial analysis tasks while maintaining real-time performance at 33.21 FPS." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.742, + 0.222, + 0.758 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.768, + 0.484, + 0.875 + ], + "angle": 0, + "content": "Face analysis is a crucial problem as it has broad range of application such as face verification and identification [89, 90], surveillance [24], face swapping [13], face editing [134], de-occlusion [117], 3D face reconstruction [108], retail [1], image generation [115] and face retrieval [119]. Facial analysis tasks (Figure 1 include face parsing [32, 104], landmarks detection [49, 132], head pose" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.269, + 0.642, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.269, + 0.768, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.771, + 0.269, + 0.892, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.327, + 0.642, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.327, + 0.768, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.771, + 0.327, + 0.892, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.385, + 0.641, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.385, + 0.768, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.771, + 0.385, + 0.892, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.455, + 0.907, + 0.526 + ], + "angle": 0, + "content": "Figure 1. FaceXFormer an end-to-end unified transformer model for 10 different facial analysis tasks such as face parsing, landmark detection, head pose estimation, attributes recognition, age, gender, and race estimation, facial expression recognition, face recognition, and face visibility prediction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.907, + 0.644 + ], + "angle": 0, + "content": "estimation [12, 131], facial attributes recognition [61, 69], age/gender/race estimation [8, 46], facial expression recognition [82], face recognition [37], and face visibility prediction [39, 58]. Therefore, developing a generalized and robust face model for all tasks is a crucial and longstanding problem in the face community." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Why Unified Model? In recent years, significant advancements have been made in facial analysis, developing state-of-the-art methods and face libraries for various tasks [12, 13, 46, 117, 131, 132]. Despite these methods achieving promising performance, they cannot be integrated into a single pipeline due to their specialized model designs and task-specific pre-processing techniques. Furthermore, deploying multiple specialized models simultaneously is computationally intensive and impractical for real-time applications, leading to increased system complexity and resource consumption. These challenges emphasize the need for a unified model that can concurrently handle multiple facial analysis tasks efficiently (see Table 1). A single model capable of addressing multiple facial tasks is desirable because it: (1) learns a robust and generalized face representation capable of handling \"in-the-wild\" images; (2) intra-task modeling helps the models to learn task-invariant representation;" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.218, + 0.9 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11369" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.09, + 0.484, + 0.491 + ], + "angle": 0, + "content": "
MethodsFPLDHPEAttrAgeGenRaceVisExpFR
Single-Task Models
DML-CSR [127]✓
FP-LIIF [80]✓
SegFace [67]✓
Wing [22]✓
HRNet [98]✓
WHENet [131]✓
TriNet [9]✓
img2pose [3]✓
TokenHPE [121]✓
SSPL [85]✓
VOLO-D1 [40]✓
DLDL-v2 [23]✓
3DDE [94]✓
MNN [95]✓
KTN [44]✓
DMUE [82]✓
CosFace [97]✓
ArcFace [15]✓
AdaFace [37]✓
Multi-Task Models
SSP+SSG [34]✓✓
Hetero-FAE [27]✓✓✓✓✓
FairFace [35]✓✓✓
MiVOLO [40]✓✓
MTL-CNN [137]✓✓✓
ProS [17]✓✓✓
FaRL [130]✓✓✓✓✓
HyperFace [75]✓✓✓✓✓
AllinOne [76]✓✓✓✓✓✓
Swinface [71]✓✓✓✓
QFace [88]✓✓✓✓
Facetor [72]✓✓✓✓✓✓✓
FaceXFormer✓✓✓✓✓✓✓✓✓✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.503, + 0.484, + 0.601 + ], + "angle": 0, + "content": "Table 1. Comparison with representative methods under different task settings. Our proposed FaceXFormer can perform various facial analysis tasks in a single model. FP - Face Parsing, LD - Landmarks Detection, HPE - Head Pose Estimation, Attr - Attributes Recognition, Age - Age, Gen - Gender, Race - Race Estimation, Exp - Facial Expression Recognition, FR - Face Recognition, and Vis - Face Visibility Prediction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.658 + ], + "angle": 0, + "content": "and (3) simplifies deployment pipelines by reducing computational overhead and enabling faster inference." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.66, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Proposed FaceXFormer Architecture: To this end, we introduce FaceXFormer, an end-to-end unified model designed for ten different facial analysis tasks, as depicted in Figure 1. These tasks include face parsing, landmark detection, head pose estimation, attributes recognition, age/gender/race estimation, facial expression recognition, face recognition and face visibility prediction. FaceXFormer enables task unification by leveraging transformers and learnable tokens as its core components. Specifically, we employ a transformer-based encoder-decoder structure, where the encoder extracts hierarchical face representations and fuses them using a MLP fusion module. The fused features are then processed in the decoder, where each facial analysis task is represented by a unique learnable token, allowing for the simultaneous and effective processing of multiple tasks. In particular, we propose a lightweight de" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.242 + ], + "angle": 0, + "content": "coder, FaceX, which processes both face and task tokens together using bi-directional cross-attention mechanism (Section 3.2), enabling the model to learn robust face representations that generalize across various tasks. The bi-directional cross-attention mechanism enables a 2-layer lightweight decoder, allowing the model to operate in real time. After modeling the intra-task and face-token relationships in the FaceX decoder, the task tokens are fed into a unified head, which converts these task tokens into corresponding task predictions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.243, + 0.907, + 0.439 + ], + "angle": 0, + "content": "Our extensive experiments demonstrate that FaceX-Former achieves state-of-the-art or competitive performance compared to specialized models and existing multitask models across multiple benchmarks, while supporting more tasks than any previous multi-task model. Moreover, we show that our model effectively handles images \"in the wild\", demonstrating its robustness and generalizability across ten different tasks. This robustness is critical for real-world applications where uncontrolled conditions and diverse inputs are common. FaceXFormer achieves state-of-the-art performance at 33.21 FPS, representing a significant \\(69.44\\%\\) speed boost over prior multi-task models, making it highly suitable for real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.44, + 0.885, + 0.454 + ], + "angle": 0, + "content": "In summary, our paper's contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.455, + 0.905, + 0.514 + ], + "angle": 0, + "content": "1. We introduce FaceXFormer, a unified transformer-based framework capable of simultaneously processing ten different facial analysis tasks, achieving real-time performance of 33.21 FPS." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.515, + 0.905, + 0.56 + ], + "angle": 0, + "content": "2. We propose FaceX, a lightweight decoder that employs the proposed bi-directional cross-attention mechanism, enabling joint processing of face and task tokens." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.56, + 0.905, + 0.62 + ], + "angle": 0, + "content": "3. We conduct extensive experiments and analyses to demonstrate that our approach achieves state-of-the-art performance with reduced inference time compared to specialized and multi-task models across multiple tasks." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.455, + 0.905, + 0.62 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.634, + 0.655, + 0.649 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Facial analysis tasks: Facial analysis tasks involve face parsing [11, 32, 67, 127], landmarks detection [49, 59, 132], head pose estimation [12, 95, 121, 131], facial attributes recognition [61, 69, 85, 130], age/gender/race estimation [8, 40, 43, 46], facial expression recognition [44, 45], face recognition [15, 97] and face visibility prediction [39, 58]. These tasks hold significance in various applications such as face swapping [13, 65], face editing [134], de-occlusion [117], 3D face reconstruction [108], driver assistance [64], human-robot interaction [86], retail [1], face verification and identification [89, 90], image generation [115], image retrieval [119] and surveillance [24, 66]. Specialized models excel in their respective tasks but cannot be easily integrated with other tasks due to the need for extensive task-specific pre-processing [50, 132]. Generally, these models under-perform when applied to tasks beyond" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11370" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.089, + 0.48, + 0.188 + ], + "angle": 0, + "content": "
QFace [88]Facetor [72]FaceXFormer
Tasks4710
Model Size-178.9 M109.29 M
Face PretrainedYesYesNo
FPS (fp32)-14.3033.21
DecoderQuery2Label [52]Pixel Decoder & Transformer DecoderFaceX
Decoder Layers992
" + }, + { + "type": "table_caption", + "bbox": [ + 0.115, + 0.198, + 0.457, + 0.212 + ], + "angle": 0, + "content": "Table 2. Comparison of multi-task face analysis methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.232, + 0.483, + 0.307 + ], + "angle": 0, + "content": "their specialization as their design is specific to their designated tasks. Some works [29, 60, 124, 126] perform multiple tasks simultaneously but utilize the additional tasks for guidance or auxiliary loss calculation to enhance the performance of the primary task." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.31, + 0.484, + 0.642 + ], + "angle": 0, + "content": "Multi-task learning for face analysis: HyperFace [75] and AllinOne [76] are early convolution-based models that aim to perform multiple tasks. Recent multi-task frameworks, such as QFace [88] and Facetor [72], are also inspired from DETR [10] and propose a unified model structure consisting of learnable tokens. However, these previous works differ from the proposed method in several key aspects, as summarized in Table 2. Specifically, QFace [88] employs a feature fusion module that uses stage embeddings to aggregate features from the encoder. Facetor [72] introduces a layer-attention mechanism to fuse features from different encoder layers and incorporates two separate decoders. Both methods, employ a 9-layer transformer decoder, also Facetor additionally includes a Pixel Decoder. These architectural components increase computational overhead, resulting in slower inference times. In contrast, FaceXFormer proposes a bi-directional cross-attention mechanism, which enables efficient task-specific feature extraction from face tokens resulting in a 2-layer lightweight decoder. This design choice is the primary reason for FaceXFormer's superior speed and performance. Notably, unlike previous methods, FaceXFormer does not rely on face-specific pertaining backbone." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Unified transformer models: In recent years, the rise of transformers [19, 96] have paved the way for the unification of multiple tasks within a single architecture. Unified transformer architectures are being explored across various computer vision problems, including segmentation [47, 138], visual question answering (VQA) [99, 118], tracking [102, 133], detection [103]. While these models may not achieve state-of-the-art (SOTA) performance and may under-perform compared to specialized models on some tasks, they demonstrate competitive performance across a variety of tasks. Such unification efforts have led to the development of foundational models like SAM [38], CLIP [73], LLaMA [93], GPT-3 [6], DALL-E [74], etc. However, these models are computationally intensive and not suitable for facial analysis applications that require real-time performance. Motivated by this challenge, we propose FaceXFormer: the first lightweight, transformer-based" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "model capable of performing multiple facial analysis tasks. It delivers real-time performance at 33.21 FPS and can be seamlessly integrated into existing systems providing additional annotations for the person of interest." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.166, + 0.656, + 0.182 + ], + "angle": 0, + "content": "3. FaceXFormer" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.192, + 0.907, + 0.37 + ], + "angle": 0, + "content": "In our framework, we follow a standard encoder-decoder structure as illustrated in Fig. 2. For an input face image \\(\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}\\), we extract coarse to fine-grained multiscale features \\(\\mathbf{S}_i\\), where \\(i\\) belongs to the \\(i\\)-th encoder output. To learn a unified face representation \\(\\mathbf{F}\\), these multiscale features are then fused using a MLP-Fusion \\(\\mathbf{M}\\) module. Following fusion, we initialize a series of task-specific tokens \\(\\mathbf{T} = \\langle T_1, \\dots, T_n \\rangle\\), with each \\(t_i\\) representing a face task. Face tokens \\(\\mathbf{F}\\) and task tokens \\(\\mathbf{T}\\) are then processed by a lightweight Decoder FXDec where task tokens are attended with face tokens to learn relevant task representation." + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.371, + 0.807, + 0.389 + ], + "angle": 0, + "content": "\\[\n\\langle \\hat {\\mathbf {T}} \\rangle = \\mathbf {F} \\mathbf {X D e c} (\\langle \\mathbf {F}, \\mathbf {T} \\rangle ; \\mathbf {S} _ {i})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.396, + 0.907, + 0.442 + ], + "angle": 0, + "content": "Here, \\(\\hat{\\mathbf{T}}\\) represents the output task tokens. These tokens are then fed into unified heads, where each task token is refined and passed to its respective task head for prediction." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.452, + 0.706, + 0.466 + ], + "angle": 0, + "content": "3.1. Multi-scale Encoder" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.474, + 0.907, + 0.655 + ], + "angle": 0, + "content": "In the encoder, we employ a multi-scale encoding strategy to address the varying feature requirements intrinsic to each face analysis task. For instance, age estimation requires a global representation, while face parsing necessitates a fine-grained representation. Given an input image \\(\\mathbf{I}\\), it is processed through a set of encoder layers. For each encoder layer, the output captures information at varying levels of abstraction and detail, generating multi-scale features \\(\\{\\mathbf{S}_i\\}_{i=1}^n\\), where \\(i\\) ranges from 1 to 4. This results in a hierarchical structure of features, wherein each feature map \\(\\mathbf{S}_i\\) transitions from a coarse to a fine-grained representation suitable for diverse facial analysis tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.67, + 0.909, + 0.835 + ], + "angle": 0, + "content": "MLP-Fusion: Assigning each feature-map \\(\\mathbf{S}_i\\) to each face task is sub-optimal; rather, learning a unified face representation is more optimal and parameter-efficient. Following [113], we utilize a MLP-Fusion module \\(\\mathbf{M}\\) to generate a fused face representation from the multi-scale features \\(\\{\\mathbf{S}_i\\}_{i=1}^n\\). In this framework, each feature map \\(\\mathbf{S}_i\\) is initially passed through a separate MLP layer, standardizing the channel dimensions across scales to facilitate fusion. The transformed features are then concatenated and passed through a fusion MLP layer to aggregate a fused representation \\(\\mathbf{F}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.845, + 0.858, + 0.865 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {S}} _ {i} = \\operatorname {M L P} _ {\\operatorname {p r o j}} \\left(D _ {i}, D _ {t}\\right) \\left(\\mathbf {S} _ {i}\\right), \\forall i \\in \\{1, \\dots , n \\},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.867, + 0.774, + 0.885 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {\\text {c a t}} = \\operatorname {C o n c a t} \\left(\\hat {\\mathbf {S}} _ {1}, \\hat {\\mathbf {S}} _ {2}, \\dots , \\hat {\\mathbf {S}} _ {n}\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.562, + 0.887, + 0.774, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} = \\mathbf {M L P} _ {\\text {f u s i o n}} (n D _ {t}, D _ {t}) (\\mathbf {F} _ {\\text {c a t}}),\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11371" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.095, + 0.889, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.365, + 0.908, + 0.451 + ], + "angle": 0, + "content": "Figure 2. Overview of our proposed framework. The FaceXFormer employs an encoder-decoder architecture, extracting multi-scale features from the input face image \\(\\mathbf{I}\\), and fusing them into a unified representation \\(\\mathbf{F}\\) via MLP-Fusion. Task tokens \\(\\mathbf{T}\\) are processed alongside face representation \\(\\mathbf{F}\\) in the FaceX Decoder FXDec, resulting in refined task-specific tokens \\(\\hat{\\mathbf{T}}\\). These refined tokens are then used for task-specific predictions by passing through the unified head. FaceXFormer performs ten tasks, including face parsing, landmark detection, head pose estimation, attribute prediction, age, gender, and race estimation, facial expression recognition, face recognition, and face visibility prediction, achieving state-of-the-art performance at a real-time FPS of 33.21." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.475, + 0.484, + 0.567 + ], + "angle": 0, + "content": "where \\(D_{i}\\) and \\(D_{t}\\) are the multi-scale feature channel dimensions of \\(\\mathbf{S}_i\\) and the target channel dimension, respectively. The MLP-fusion design ensures minimal computational overhead (983k parameters) while maintaining the ability to perform efficient feature fusion, which is crucial for real-time application based face analysis tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.577, + 0.245, + 0.591 + ], + "angle": 0, + "content": "3.2. FaceX Decoder" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.484, + 0.841 + ], + "angle": 0, + "content": "Detection transformer (DETR) [10] employs object tokens to learn bounding box predictions for each object. Inspired by this approach, we introduce Task Tokens, whereby each task token is designed to learn specific facial tasks leveraging the fused face representation. However, existing decoders such as DETR [10] and Deformable-DETR [136] are computationally intensive, impacting runtime significantly. To address this, we propose FaceX (FXDec) a lightweight decoder designed to efficiently model the task tokens with face tokens. Specifically, each task token learns a task-related representation by interacting with other task tokens \\(\\mathbf{T}\\) and face tokens \\(\\mathbf{F}\\), enhancing the overall representation. The Lightweight Decoder comprises of three main components: 1) Task Self-Attention, 2) Task-to-Face Cross-Attention, and 3) Face-to-Task Cross-Attention as illustrated in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Task Self-Attention (TSA): The Task Self-Attention module is designed to refine the task-specific representations within the set of task tokens \\(\\mathbf{T} = \\langle T_1, \\ldots, T_n \\rangle\\). Each task" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.475, + 0.907, + 0.537 + ], + "angle": 0, + "content": "token \\( T_{i} \\) is an embedded representation that corresponds to a specific facial task. In TSA, each \\( T_{i} \\) is updated by attending to all other task tokens to capture task-specific interactions. Formally, the updated task token \\( T_{i}^{\\prime} \\) is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.544, + 0.851, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T} _ {i} ^ {\\prime} = \\operatorname {S e l f A t t n} (\\mathbf {Q} = T _ {i}, \\mathbf {K} = \\mathbf {T}, \\mathbf {V} = \\mathbf {T}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.907, + 0.629 + ], + "angle": 0, + "content": "where Attention denotes the multi-headed self-attention mechanism, and \\(\\mathbf{Q}\\), \\(\\mathbf{K}\\), and \\(\\mathbf{V}\\) represent the queries, keys, and values, respectively. Therefore, TSA essentially helps the model to learn task-invariant representation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.909, + 0.764 + ], + "angle": 0, + "content": "Task-to-Face Cross-Attention (TFCA): The Task-to-Face Cross-Attention module allows each task token to interact with the fused face representation \\(\\mathbf{F}\\). This enables each task token to gather information relevant to its specific facial task from the fused face features. In this module, the fused face representation \\(\\mathbf{F}\\) acts as both key and value, while the task tokens serve as queries. The updated task token \\(\\hat{T}_i\\) is then computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.77, + 0.854, + 0.789 + ], + "angle": 0, + "content": "\\[\n\\hat {T} _ {i} = \\operatorname {C r o s s A t t n} (\\mathbf {Q} = T _ {i} ^ {\\prime}, \\mathbf {K} = \\mathbf {F}, \\mathbf {V} = \\mathbf {F}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.794, + 0.906, + 0.855 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{T}} = \\langle \\hat{T}_1,\\dots ,\\hat{T}_n\\rangle\\) is the output task token. Thus, TFCA enables direct interaction between the task-specific tokens and the compact facial features, facilitating task-focused feature extraction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "Face-to-Task Cross-Attention (FTCA): Conversely, the Face-to-Task Cross-Attention module is designed to refine" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11372" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "the fused face representation \\(\\mathbf{F}\\) based on the information from the updated task tokens. This process aids in enhancing the face representation with task-specific details, thereby improving the extraction of overall fused representation. In FTCA, the set of updated task tokens \\(\\mathbf{T}' = \\{\\mathbf{T}_1'',\\mathbf{T}_2'',\\dots,\\mathbf{T}_m''\\}\\) acts as both keys and values, while the fused face features \\(\\mathbf{F}\\) serve as queries. The refined face representation \\(\\hat{\\mathbf{F}}\\) is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.223, + 0.432, + 0.24 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {F}} = \\operatorname {C r o s s A t t n} (\\mathbf {Q} = \\mathbf {F}, \\mathbf {K} = \\mathbf {T} ^ {\\prime}, \\mathbf {V} = \\mathbf {T} ^ {\\prime}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.252, + 0.483, + 0.297 + ], + "angle": 0, + "content": "Through this inverse attention mechanism, the face representation is augmented with critical task-specific details, enabling a robust approach towards facial task unification." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.306, + 0.231, + 0.321 + ], + "angle": 0, + "content": "3.3. Unified-Head" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.328, + 0.484, + 0.614 + ], + "angle": 0, + "content": "In Unified-Head, the task tokens are processed to obtain corresponding task predictions. As shown in Figure 2, the output face tokens \\(\\hat{\\mathbf{F}}\\) and task tokens \\(\\hat{\\mathbf{T}}\\) are processed through a Task-to-Face Cross-Attention mechanism to obtain final refined features. Then, the output tokens are fed into their corresponding task heads. The task head for landmark detection is a hourglass network, for head pose estimation is a regression MLP, and for face recognition is PartialFC [4], while the tasks of age, gender and race estimation, facial expression recognition, face visibility prediction, and attributes prediction utilize classification MLPs. For face parsing, we leverage the output \\(\\hat{\\mathbf{F}}\\) and process it through an upsampling layer, then perform a cross-product with the face parsing token to obtain a segmentation map. The number of tokens for segmentation corresponds to the total number of classes. For landmark prediction, it corresponds to the number of landmarks (i.e., 68). For head pose estimation, the number of tokens is 9, representing the \\(3\\times 3\\) rotation matrix. For other tasks, one token is used for each." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.625, + 0.285, + 0.641 + ], + "angle": 0, + "content": "3.4. Multi-Task Training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.647, + 0.483, + 0.872 + ], + "angle": 0, + "content": "We aim to train FaceXFormer for multiple facial analysis tasks simultaneously, however each task requires distinct and sometimes conflicting pre-processing steps. For instance, landmark detection typically requires keypoint alignment of faces, which contradicts the needs for head pose estimation, as it may eliminate the natural variability of headposes. Due to these reasons, integrating all tasks into a single model poses significant challenges. To address this, FaceXFormer incorporates task-specific tokens designed to extract task-specific features from the fused representation. These task tokens compel the backbone to learn a unified representation capable of supporting a broad spectrum of facial analysis tasks. We employ different loss functions for each task and combine them in a joint objective for training. The final loss function is given as:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.873, + 0.48, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L = \\lambda_ {s e g} L _ {s e g} + \\lambda_ {l n d} L _ {l n d} + \\lambda_ {h p e} L _ {h p e} + \\lambda_ {a t t r} L _ {a t t r} + \\lambda_ {a} L _ {a} \\\\ + \\lambda_ {g / r} L _ {g / r} + \\lambda_ {e x p} L _ {e x p} + \\lambda_ {f r} L _ {f r} + \\lambda_ {v i s} L _ {v i s} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.228 + ], + "angle": 0, + "content": "where \\( L_{seg} \\) is the mean of dice loss [87] and Cross-Entropy (CE) loss for face parsing, \\( L_{lnd} \\) is STAR loss [132] for landmarks prediction, \\( L_{hpe} \\) is geodesic loss [121] for head pose estimation, \\( L_{g/r} \\) is CE loss for gender/race estimation, \\( L_a \\) is mean of L1 loss and CE loss for age estimation, \\( L_{exp} \\) is CE loss for facial expression recognition, \\( L_{fr} \\) is Arc-Face [15] loss for face recognition, and \\( L_{attr} \\) and \\( L_{vis} \\) are Binary Cross-Entropy with logits loss for attributes prediction and face visibility prediction respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.24, + 0.748, + 0.258 + ], + "angle": 0, + "content": "4. Experiments and Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.266, + 0.907, + 0.372 + ], + "angle": 0, + "content": "To validate the effectiveness of FaceXFormer, we conduct extensive experiments across ten diverse facial analysis tasks using a range of datasets, as described in Section 4.1. Implementation details are provided in Section 4.2 to ensure reproducibility and transparency. Finally, Sections 4.3 and 4.4 present the quantitative and qualitative results, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.381, + 0.713, + 0.395 + ], + "angle": 0, + "content": "4.1. Datasets and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.402, + 0.905, + 0.508 + ], + "angle": 0, + "content": "We perform co-training, where the model is simultaneously trained for multiple tasks using a total of 10 datasets with task-specific annotations. We conduct a comprehensive evaluation, comparing our approach with both task-specific and multi-task models. We present our results on the test sets according to the standard protocol for each task using the following datasets:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.508, + 0.905, + 0.613 + ], + "angle": 0, + "content": "Train: Face Parsing: CelebAMaskHQ [42]; Landmarks Detection: 300W [79]; Head Pose Estimation: 300W-LP [135]; Attributes Prediction: CelebA [54]; Facial Expression Recognition: RAF-DB [45], AffectNet [62]; Age/Gender/Race estimation: UTKFace [125], FairFace [36]; Face Recognition: MS1MV3 [25]; Visibility Prediction: COFW [7]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.614, + 0.905, + 0.734 + ], + "angle": 0, + "content": "Test: Face Parsing: CelebAMaskHQ [42]; Landmarks Detection: 300W [135], 300VW [83]; Head Pose Estimation: BIWI [20]; Attributes Prediction: CelebA [54], LFWA [107]; Facial Expression Recognition: RAFDB [45]; Age/Gender/Race Estimation: UTKFace [125], FairFace [36]; Face Recognition: LFW [31], CFP-FP [81], AgeDB [63], CALFW [129], CPLFW [128]; Visibility Prediction: COFW [7]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.905, + 0.84 + ], + "angle": 0, + "content": "The evaluation metrics used are the F1-score for face parsing, Normalized Mean Error (NME) for landmark prediction, Mean Absolute Error (MAE) for head pose estimation and age estimation, accuracy for facial expression recognition, attributes prediction, gender estimation, race estimation, 1:1 verification accuracy for face recognition, and recall at \\(80\\%\\) precision for face visibility prediction." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.731, + 0.865 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.902 + ], + "angle": 0, + "content": "We train our models using a distributed PyTorch setup on eight A6000 GPUs, each equipped with 48GB of memory." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11373" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.908, + 0.267 + ], + "angle": 0, + "content": "
MethodInput Res.SkinHairNoseL-EyeR-EyeL-BrowR-BrowL-LipI-MouthU-LipMean F1 ↑
Wei et al. [105]51296.491.191.987.185.080.882.591.090.687.988.43
EHANet [55]51296.093.993.786.286.583.283.190.393.888.689.53
EAGRNet [91]47396.294.994.088.689.085.785.291.295.088.990.87
AGRNet [92]47396.587.693.988.789.185.585.691.192.089.189.91
FaRL_scratch [130]51296.294.993.889.089.085.385.490.091.788.190.34
DML-CSR [127]47395.794.593.989.489.685.585.791.091.889.190.62
FP-LIIF [80]25696.495.193.788.588.584.584.390.392.187.590.09
SwinFace [71]××××××××××××
QFace [88]××××××××××××
Facetor [72]51296.696.293.989.489.186.286.390.691.689.090.89
FaceXFormer22496.495.793.890.190.386.085.990.692.189.291.01
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.278, + 0.908, + 0.308 + ], + "angle": 0, + "content": "Table 3. Performance comparison for face parsing on the CelebAMask-HQ dataset [42]. The symbol \\(\\times\\) indicates that the model does not perform the corresponding task. Red = First Best, Blue = Second Best. \\(\\times\\) indicates a model that doesn't perform the task." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.324, + 0.48, + 0.464 + ], + "angle": 0, + "content": "
MethodsExpression (RAF-DB) ↑MethodsVisibility (COFW) ↑MethodsAge (MAE) UTKFace ↓
DLP-CNN [45]80.89RCPR [7]40OR-CNN [68]5.74
gACNN [48]85.07Wu et al. [111]44.43Axel Berg et al. [5]4.55
IPA2LT [120]86.77Wu et al. [110]49.11CORAL [8]5.47
RAN [101]86.90ECT [122]63.4Gustafsson et al. [26]4.65
CovPool [2]87.003DDE [94]63.89R50-SORD [70]4.36
SCN [100]87.03MNN [95]72.12VOLO-D1 [40]4.23
DACL [21]87.78DLDL-v2 [23]4.42
KTN [44]88.07MWR [84]4.37
DMUE [82]88.76
SwinFace [71]86.54SwinFace [71]×SwinFace [71]×
QFace [88]92.86QFace [88]×QFace [88]×
Facetor [72]87.58Facetor [72]×Facetor [72]4.10
FaceXFormer88.24FaceXFormer72.56FaceXFormer4.17
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.475, + 0.483, + 0.503 + ], + "angle": 0, + "content": "Table 4. Performance comparison on facial expression recognition, face visibility prediction and age estimation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.522, + 0.483, + 0.765 + ], + "angle": 0, + "content": "The models' backbones are initialized with ImageNet pretrained weights and processes input images at a resolution of \\(224 \\times 224\\). We employ the AdamW optimizer with a weight decay of \\(1e^{-5}\\). All models are trained for 12 epochs with a batch size of 48 on each GPU, and an initial learning rate of \\(1e^{-4}\\), which decays by a factor of 10 at the \\(6^{th}\\) and \\(10^{th}\\) epochs. For some tasks, we train the model for additional epochs to reach the upper ceiling of performance. For data augmentation, we randomly apply Gaussian blur, grayscale conversion, gamma correction, occlusion, horizontal flipping, and affine transformations, such as rotation, translation and scaling. The number of FaceX decoder \\(N\\) is set to two. To ensure stable training across tasks when using multiple datasets of varying sample sizes, we equalize the representation of each task's samples in every batch through upsampling." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.773, + 0.225, + 0.789 + ], + "angle": 0, + "content": "4.3. Main results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In Table 3, Table 4, Table 6, Table 5, we present a comparative analysis of FaceXFormer against recent methods across a variety of tasks. A key highlight of our work is its unique capability to deliver promising results across multiple tasks at real-time inference speed using a single unified model. Specifically, FaceXFormer achieves state-of-the-art performance in face parsing, with a mean F1 score of 91.01 on" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.326, + 0.907, + 0.629 + ], + "angle": 0, + "content": "CelebAMaskHQ at a resolution of \\(224 \\times 224\\), which is half the input size required by other SOTA methods. Furthermore, it demonstrates superior performance in head pose estimation and landmark detection, achieving a mean MAE of 3.52 and a mean NME of 4.67, respectively. Additionally, FaceXFormer provides a significant performance boost in attributes prediction and visibility prediction, achieving an accuracy of \\(91.83\\%\\) on the CelebA dataset and \\(72.56\\%\\) on COFW. It also performs competitively in age estimation, achieving the second-best score of 4.17, and achieves an accuracy of \\(88.24\\%\\) in facial expression recognition. In face recognition, FaceXFormer outperforms Faceptor, achieving a mean accuracy of \\(95.94\\%\\) compared to \\(95.28\\%\\). However, we observe that multi-task models generally underperform compared to specialized ones in this task. This can be attributed to conflicting training objectives, which force the model to learn identity-invariant features rather than identity-specific representations crucial for accurate recognition. The results on gender estimation across different race categories is shown in Table 8." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Recent models such as SwinFace [71], QFace [88], and Facetor [72] also aim to unify multiple tasks but only address a subset of them. These approaches often exclude complex tasks such as segmentation, head pose estimation, and landmark prediction. Moreover, they rely on multiple decoders and computationally expensive attention mechanisms, adding to the overall computational overhead. In contrast, FaceXFormer seamlessly unifies these complex tasks using a lightweight decoder and achieves state-of-the-art performance across them at a real-time FPS of 33.21. It outperforms previous multi-task models in segmentation, head pose estimation, landmark prediction, attribute prediction, and face visibility prediction, while achieving the second-best performance in age estimation. In this work, we simultaneously train for ten heterogeneous tasks, posing a more formidable challenge than previous approaches. Despite this, FaceXFormer effectively handles multiple tasks, achieving SOTA or competitive performance in real time." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11374" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.103, + 0.089, + 0.895, + 0.326 + ], + "angle": 0, + "content": "
MethodsHeadpose (BIWI)MethodsLandmarks (300W)MethodsCelebA Acc. ↑
YawPitchRollMAE ↓Full ↓Com ↓Chal ↓
HopeNet [78]4.816.613.274.89LAB [109]3.492.985.19PANDA-1 [123]85.43
QuatNet [30]5.494.012.944.15Wing [22]4.043.277.18LNets+ANet [53]87.33
FSA-Net [116]4.275.492.934.14DeCaFa [14]3.392.935.26SSP+SSG [34]88.24
EVA-GCN [114]6.014.782.983.98HRNet [98]3.322.875.15MOON [77]90.94
TriNet [9]4.114.753.043.97PicassoNet [106]3.583.035.81NSA [56]90.61
img2pose [3]4.563.543.243.78AVS+SAN [18]3.863.216.46MCNN-AUX [28]91.29
MNN [95]3.984.612.393.66LUVLi [39]3.232.765.16MCFA [137]91.23
MFDNet [51]3.404.682.773.62HIH [41]3.092.654.89DMM-CNN [57]91.70
TokenHPE [121]3.954.512.713.72PIPNet [33]3.192.784.89SSPL [85]91.77
WHENet [131]3.994.393.063.81SLPT [112]3.172.754.90FaRL [130]91.39
SwinFace [71]××××SwinFace [71]×××SwinFace [71]91.38
QFace [88]----QFace [88]×××QFace [88]91.56
Facector [72]××××Facector [72]3.162.754.84Facector [72]91.39
FaceXFormer3.913.972.673.52FaceXFormer3.052.664.67FaceXFormer91.83
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.336, + 0.905, + 0.365 + ], + "angle": 0, + "content": "Table 5. Performance comparison on headpose, landmark detection, and attribute recognition. The symbol \\(\\times\\) indicates that the model does not perform the corresponding task, while - denotes that results for this dataset are not provided. Red = First Best, Blue = Second Best." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.383, + 0.481, + 0.489 + ], + "angle": 0, + "content": "
MethodLFWCFP-FPAgeDBCALFWCPLFWMean ↑
CosFace [97]99.8198.1298.1195.7692.2896.81
ArcFace [15]99.8398.2798.2895.4592.0896.78
VPL-ArcFace [16]99.8399.1198.6096.1293.4597.42
AdaFace [37]99.8399.1198.1796.0293.9397.41
SwinFace [71]99.8798.6098.1596.1093.4297.22
QFace [88]××××××
Facetor [72]99.4096.3493.6594.7592.2795.28
FaceXFormer99.6896.7596.3595.5091.4195.94
" + }, + { + "type": "table_caption", + "bbox": [ + 0.124, + 0.499, + 0.45, + 0.513 + ], + "angle": 0, + "content": "Table 6. Performance comparison for face recognition." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.529, + 0.483, + 0.575 + ], + "angle": 0, + "content": "This success can be attributed to the efficiency of the proposed lightweight decoder, which employs a novel bidirectional cross-attention mechanism." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.583, + 0.377, + 0.6 + ], + "angle": 0, + "content": "4.4. Qualitative \"in-the-wild\" results" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.605, + 0.483, + 0.817 + ], + "angle": 0, + "content": "In this section, we present the qualitative results of FaceXFormer on randomly selected \"in-the-wild\" images. We select five random images and showcase the results for face parsing, head pose estimation, landmarks prediction, age estimation, gender and race classification, and attributes prediction in Figure 3. Notably, the model successfully performs complex tasks such as face segmentation, head pose estimation, and landmark prediction, even when input samples exhibit extreme poses, occlusions, or blurring. Furthermore, FaceXFormer can be effectively used as a tool to generate multiple annotations for each image, making it valuable for various downstream tasks. These results highlight FaceXFormer's robust performance in challenging, real-world scenarios." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.83, + 0.245, + 0.848 + ], + "angle": 0, + "content": "5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In this section, we explore the impact of different components of FaceXFormer on performance. Additionally, we demonstrate that the proposed model exhibits minimal bias" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.387, + 0.907, + 0.448 + ], + "angle": 0, + "content": "compared to other models by evaluating age and gender prediction across various demographics. Furthermore, we analyze the impact of using different backbones of varying sizes on the FPS and performance." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.458, + 0.905, + 0.543 + ], + "angle": 0, + "content": "
MLP FusionSelf Attn.Cross Attn.Bi-dir. CAHPE (MAE)Lnd (NME)Attr. (Acc.)Age (MAE)
✓✓3.704.7091.214.21
✓✓17.1631.4979.9016.15
✓✓✓4.635.3089.984.84
✓✓✓3.524.6791.834.17
" + }, + { + "type": "table_caption", + "bbox": [ + 0.542, + 0.55, + 0.878, + 0.565 + ], + "angle": 0, + "content": "Table 7. Impact of various components on performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.577, + 0.907, + 0.594 + ], + "angle": 0, + "content": "5.1. Impact of various components in FaceXFormer" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.901 + ], + "angle": 0, + "content": "To evaluate the contribution of each component in FaceXFormer, we conduct an ablation study focusing on the importance of specific design choices and their impact on performance across various tasks. The results of these experiments are summarized in Table 7. We observe that without MLP fusion (row 1), there is a drop in performance, highlighting the importance of integrating multi-scale features to capture both global and local information essential for accurate predictions. The model performs extremely poorly (row 2) without cross-attention in the decoder, which is expected, as there is no interaction between face tokens and task tokens in this case. Introducing the proposed bidirectional cross-attention (row 4), which corresponds to FaceXFormer, in the decoder provides a significant boost compared to using standard cross-attention (row 3), yielding improvements of 1.11 MAE in head pose estimation, 0.63 NME in landmark detection, 1.85 accuracy points in attribute prediction, and 0.67 MAE in age estimation. These results demonstrate the importance of MLP fusion and bidirectional cross-attention in the FaceXFormer architecture." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11375" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.482, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.108, + 0.433, + 0.466, + 0.447 + ], + "angle": 0, + "content": "Figure 3. FaceXFormer predictions on \"in-the-wild\" images" + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.462, + 0.483, + 0.572 + ], + "angle": 0, + "content": "
ModelData PointsWhiteNon-whiteAverageDiscrepancy
AgeFairFace100K60.0560.6360.520.58
CLIP400M62.2561.9562.00-0.30
FaRL20M61.4961.8461.780.35
FaceXFormer5.9M58.9459.4459.340.50
GenderFairFace100K94.1594.4194.360.26
CLIP400M94.8795.7895.610.91
FaRL20M95.1695.7795.650.61
FaceXFormer5.9M95.3495.1995.22-0.09
" + }, + { + "type": "table_caption", + "bbox": [ + 0.098, + 0.579, + 0.477, + 0.593 + ], + "angle": 0, + "content": "Table 8. Age and gender accuracy w.r.t race groups on FairFace" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.618, + 0.443, + 0.633 + ], + "angle": 0, + "content": "5.2. Bias Analysis and Ethical Considerations" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.644, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In our work, we utilize 17 unique datasets for training and evaluation. We obtained these datasets following the procedures stated on their respective pages and signed the license agreements if and when necessary. As we train our models on multiple datasets designed for different tasks, the subjects across different age groups, genders, and races is not equal. This imbalance may introduce bias in the model. Therefore, we provide an analysis using the FairFace [35] dataset, which is balanced in terms of age, gender and race. We follow [73] and define the \"Non-white\" group to include multiple racial categories: \"Black\", \"Indian\", \"East Asian\", \"Southeast Asian\", \"Middle Eastern\" and \"Latino\". As can be seen from Table 8, FaceXFormer shows the smallest performance discrepancy across different racial groups and exhibits minimal bias compared to other models despite being trained on fewer data points. This can be attributed to race estimation being the task in co-training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.091, + 0.857, + 0.106 + ], + "angle": 0, + "content": "5.3. Performance under different backbones" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.115, + 0.905, + 0.236 + ], + "angle": 0, + "content": "To evaluate the impact of different backbones on performance and FPS, we conduct an ablation study comparing various backbone architectures in FaceXFormer. We categorize head pose estimation, landmark prediction, and age estimation as regression (Reg) tasks, while attribute prediction and facial expression recognition fall under classification (Cls). Additionally, face parsing is denoted as segmentation (Seg). The results are summarized in Table 9." + }, + { + "type": "table", + "bbox": [ + 0.534, + 0.252, + 0.888, + 0.338 + ], + "angle": 0, + "content": "
BackboneSegRegClsFPSParams
MobileNet90.214.6488.2239.7625.32
ResNet10190.494.3788.9134.9865.54
ConvNext-B91.084.3589.0936.61110.19
Swin-B91.014.1290.0333.21109.29
" + }, + { + "type": "table_caption", + "bbox": [ + 0.52, + 0.348, + 0.899, + 0.362 + ], + "angle": 0, + "content": "Table 9. Effect of different backbones on performance and FPS." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.385, + 0.907, + 0.582 + ], + "angle": 0, + "content": "From the results, we observe that ConvNeXt achieves the best performance in segmentation with an F1 score of \\(91.08\\%\\). The Swin Transformer backbone excels in both regression and classification tasks, with a mean error of 4.12 and a mean accuracy of \\(90.03\\%\\), respectively. In contrast, MobileNet demonstrates the lowest performance metrics, including an F1 score of \\(90.21\\%\\) and a mean error of 4.64, highlighting its limitations in handling larger, more complex datasets due to its smaller receptive field compared to the Swin Transformer. The selection of the Swin Transformer as the backbone for FaceXFormer is driven by its superior scalability and global contextual understanding, both of which are essential for facial analysis tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.602, + 0.634, + 0.617 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.901 + ], + "angle": 0, + "content": "FaceXFormer introduces a novel end-to-end unified model that efficiently handles a wide range of facial analysis tasks in real-time. By adopting a transformer-based encoder-decoder architecture and representing each task as a learnable token, our approach seamlessly integrates multiple tasks within a single framework while maintaining minimal computational cost and fast inference times. The proposed lightweight decoder, FaceX, incorporates a novel bi-directional cross-attention mechanism, enhancing the model's ability to learn robust and generalized face representations across diverse tasks. Comprehensive experiments demonstrate that FaceXFormer achieves state-of-the-art performance across multiple facial analysis tasks, achieving a real-time FPS of 33.21. In broader applications, FaceXFormer can serve as an annotator for large-scale face datasets and can be integrated into existing facial analysis systems to provide extra information, making it a valuable tool for surveillance, subject analysis, and image retrieval." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11376" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.484, + 0.256 + ], + "angle": 0, + "content": "This research is based upon work supported in part by the Office of the Director of National Intelligence (ODNI), Intelligence Advanced Research Projects Activity (IARPA), via [2022-21102100005]. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies, either expressed or implied, of ODNI, IARPA, or the U.S. Government. The US Government is authorized to reproduce and distribute reprints for governmental purposes notwithstanding any copyright annotation therein." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.282, + 0.188, + 0.298 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.307, + 0.483, + 0.348 + ], + "angle": 0, + "content": "[1] B Abirami, TS Subashini, and V Mahavaishnavi. Gender and age prediction from real time facial images using cnn. Materials Today: Proceedings, 33:4708-4712, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.351, + 0.483, + 0.419 + ], + "angle": 0, + "content": "[2] Dinesh Acharya, Zhiwu Huang, Danda Pani Paudel, and Luc Van Gool. Covariance pooling for facial expression recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 367-374, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.422, + 0.483, + 0.49 + ], + "angle": 0, + "content": "[3] Vitor Albiero, Xingyu Chen, Xi Yin, Guan Pang, and Tal Hassner. img2pose: Face alignment and detection via 6dof, face pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7617-7627, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.492, + 0.483, + 0.573 + ], + "angle": 0, + "content": "[4] Xiang An, Xuhan Zhu, Yuan Gao, Yang Xiao, Yongle Zhao, Ziyong Feng, Lan Wu, Bin Qin, Ming Zhang, Debing Zhang, et al. Partial fc: Training 10 million identities on a single machine. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1445-1449, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.577, + 0.482, + 0.631 + ], + "angle": 0, + "content": "[5] Axel Berg, Magnus Oskarsson, and Mark O'Connor. Deep ordinal regression with label diversity. In 2020 25th international conference on pattern recognition (ICPR), pages 2740-2747. IEEE, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.634, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.718, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[7] Xavier P Burgos-Artizzu, Pietro Perona, and Piotr Dólar. Robust face landmark estimation under occlusion. In Proceedings of the IEEE international conference on computer vision, pages 1513-1520, 2013. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.776, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[8] Wenzhi Cao, Vahid Mirjalili, and Sebastian Raschka. Rank consistent ordinal regression for neural networks with application to age estimation. Pattern Recognition Letters, 140:325-331, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.832, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[9] Zhiwen Cao, Zongcheng Chu, Dongfang Liu, and Yingjie Chen. A vector-based representation to enhance head pose estimation. In Proceedings of the IEEE/CVF Winter Conference on applications of computer vision, pages 1188-1197, 2021. 2, 7" + }, + { + "type": "list", + "bbox": [ + 0.108, + 0.307, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[10] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.163, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[11] Liang-Chieh Chen, Yi Yang, Jiang Wang, Wei Xu, and Alan L Yuille. Attention to scale: Scale-aware semantic image segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3640-3649, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.233, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[12] Alejandro Cobo, Roberto Valle, José M Buenapasada, and Luis Baumela. On the representation and methodology for wide and short range head pose estimation. Pattern Recognition, 149:110263, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.289, + 0.905, + 0.356 + ], + "angle": 0, + "content": "[13] Kaiwen Cui, Rongliang Wu, Fangneng Zhan, and Shijian Lu. Face transformer: Towards high fidelity and accurate face swapping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 668-677, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.358, + 0.904, + 0.413 + ], + "angle": 0, + "content": "[14] Arnaud Dapogny, Kevin Bailly, and Matthieu Cord. Decafa: Deep convolutional cascade for face alignment in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6893-6901, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.415, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[15] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.484, + 0.905, + 0.551 + ], + "angle": 0, + "content": "[16] Jiankang Deng, Jia Guo, Jing Yang, Alexandros Lattas, and Stefanos Zafeiriou. Variational prototype learning for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11906-11915, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.553, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[17] Xing Di, Yiyu Zheng, Xiaoming Liu, and Yu Cheng. Pros: Facial omni-representation learning via prototype-based self-distillation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6087-6098, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.623, + 0.905, + 0.678 + ], + "angle": 0, + "content": "[18] Xuanyi Dong, Yan Yan, Wanli Ouyang, and Yi Yang. Style aggregated network for facial landmark detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 379-388, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.679, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[19] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.763, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[20] Gabriele Fanelli, Matthias Dantone, Juergen Gall, Andrea Fossati, and Luc Van Gool. Random forests for real time 3d face analysis. International journal of computer vision, 101:437-458, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.818, + 0.905, + 0.873 + ], + "angle": 0, + "content": "[21] Amir Hossein Farzaneh and Xiaojun Qi. Facial expression recognition in the wild via deep attentive center loss. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2402-2411, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[22] Zhen-Hua Feng, Josef Kittler, Muhammad Awais, Patrik Huber, and Xiao-Jun Wu. Wing loss for robust facial land" + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11377" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "mark localisation with convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.135, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[23] Bin-Bin Gao, Xin-Xin Liu, Hong-Yu Zhou, Jianxin Wu, and Xin Geng. Learning expectation of label distribution for facial age and attractiveness estimation. arXiv preprint arXiv:2007.01771, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.19, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[24] Asma El Kissi Ghalleb, Safa Boumaiza, and Najoua Essoukri Ben Amara. Demographic face profiling based on age, gender and race. In 2020 5th International Conference on Advanced Technologies for Signal and Image Processing (ATSIP), pages 1-6. IEEE, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.261, + 0.483, + 0.342 + ], + "angle": 0, + "content": "[25] Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. Ms-celeb-1m: A dataset and benchmark for large-scale face recognition. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 87-102. Springer, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.344, + 0.483, + 0.424 + ], + "angle": 0, + "content": "[26] Fredrik K Gustafsson, Martin Danelljan, Goutam Bhat, and Thomas B Schon. Energy-based models for deep probabilistic regression. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16, pages 325-343. Springer, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.428, + 0.482, + 0.495 + ], + "angle": 0, + "content": "[27] Hu Han, Anil K Jain, Fang Wang, Shiguang Shan, and Xilin Chen. Heterogeneous face attribute estimation: A deep multi-task learning approach. IEEE transactions on pattern analysis and machine intelligence, 40(11):2597-2609, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.497, + 0.482, + 0.564 + ], + "angle": 0, + "content": "[28] Emily Hand and Rama Chellappa. Attributes for improved attributes: A multi-task network utilizing implicit and explicit relationships for facial attribute classification. In Proceedings of the AAAI conference on artificial intelligence, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.567, + 0.482, + 0.634 + ], + "angle": 0, + "content": "[29] Hui-Lan Hsieh, Winston Hsu, and Yan-Ying Chen. Multi-task learning for face identification and attribute estimation. In 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2981-2985, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.637, + 0.482, + 0.691 + ], + "angle": 0, + "content": "[30] Heng-Wei Hsu, Tung-Yu Wu, Sheng Wan, Wing Hung Wong, and Chen-Yi Lee. Quatnet: Quaternion-based head pose estimation with multiregression loss. IEEE Transactions on Multimedia, 21(4):1035-1046, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.693, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[31] Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. In Workshop on faces in 'Real-Life' Images: detection, alignment, and recognition, 2008. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.763, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[32] Aaron S Jackson, Michel Valstar, and Georgios Tzimiropoulos. A cnn cascade for landmark guided semantic part segmentation. In Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part III 14, pages 143-155. Springer, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.846, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[33] Haibo Jin, Shengcai Liao, and Ling Shao. Pixel-in-pixel net: Towards efficient facial landmark detection in the wild. International Journal of Computer Vision, 129(12):3174-3194, 2021. 7" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.16 + ], + "angle": 0, + "content": "[34] Mahdi M Kalayeh, Boqing Gong, and Mubarak Shah. Improving facial attribute prediction using semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6942-6950, 2017. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.164, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[35] Kimmo Karkkainen and Jungseock Joo. Fairface: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1548-1558, 2021. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.235, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[36] Kimmo Karkkainen and Jungseock Joo. Fairface: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1548-1558, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.305, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[37] Minchul Kim, Anil K Jain, and Xiaoming Liu. Adaface: Quality adaptive margin for face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18750-18759, 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.362, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[38] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.42, + 0.905, + 0.515 + ], + "angle": 0, + "content": "[39] Abhinav Kumar, Tim K Marks, Wenxuan Mou, Ye Wang, Michael Jones, Anoop Cherian, Toshiaki Koike-Akino, Xiaoming Liu, and Chen Feng. Luvli face alignment: Estimating landmarks' location, uncertainty, and visibility likelihood. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8236-8246, 2020. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.518, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[40] Maksim Kuprashevich and Irina Tolstykh. Mivolo: Multi-input transformer for age and gender estimation. arXiv preprint arXiv:2307.04616, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.561, + 0.905, + 0.614 + ], + "angle": 0, + "content": "[41] Xing Lan, Qinghao Hu, Qiang Chen, Jian Xue, and Jian Cheng. Hih: Towards more accurate face alignment via heatmap in heatmap. arXiv preprint arXiv:2104.03100, 2021.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.617, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[42] Cheng-Han Lee, Ziwei Liu, Lingyun Wu, and Ping Luo. Maskgan: Towards diverse and interactive facial image manipulation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.675, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[43] Gil Levi and Tal Hassner. Age and gender classification using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 34-42, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.732, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[44] Hangyu Li, Nannan Wang, Xinpeng Ding, Xi Yang, and Xinbo Gao. Adaptively learning facial expression representation via cf labels and distillation. IEEE Transactions on Image Processing, 30:2016-2028, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.789, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[45] Shan Li, Weihong Deng, and JunPing Du. Reliable crowdsourcing and deep locality-preserving learning for expression recognition in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2852-2861, 2017. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[46] Wanhua Li, Xiaoke Huang, Jiwen Lu, Jianjiang Feng, and Jie Zhou. Learning probabilistic ordinal embeddings for uncertainty-aware regression. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "11378" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "IEEE/CVF conference on computer vision and pattern recognition, pages 13896-13905, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.123, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[47] Xiangtai Li, Haobo Yuan, Wei Li, Henghui Ding, Size Wu, Wenwei Zhang, Yining Li, Kai Chen, and Chen Change Loy. Omg-seg: Is one model good enough for all segmentation? arXiv preprint arXiv:2401.10229, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.181, + 0.482, + 0.236 + ], + "angle": 0, + "content": "[48] Yong Li, Jiabei Zeng, Shiguang Shan, and Xilin Chen. Occlusion aware facial expression recognition using cnn with attention mechanism. IEEE Transactions on Image Processing, 28(5):2439-2450, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.239, + 0.483, + 0.294 + ], + "angle": 0, + "content": "[49] Chunze Lin, Beier Zhu, Quan Wang, Renjie Liao, Chen Qian, Jiwen Lu, and Jie Zhou. Structure-coherent deep feature learning for robust face alignment. IEEE Transactions on Image Processing, 30:5313-5326, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.297, + 0.482, + 0.352 + ], + "angle": 0, + "content": "[50] Jinpeng Lin, Hao Yang, Dong Chen, Ming Zeng, Fang Wen, and Lu Yuan. Face parsing with roi tanh-warping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5654-5663, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.355, + 0.482, + 0.422 + ], + "angle": 0, + "content": "[51] Hai Liu, Shuai Fang, Zhaoli Zhang, Duantengchuan Li, Ke Lin, and Jiazhang Wang. Mfdnet: Collaborative poses perception and matrix fisher distribution for head pose estimation. IEEE Transactions on Multimedia, 24:2449-2460, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.426, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[52] S Liu, L Zhang, X Yang, H Su, and J Zhu. Query2label: A simple transformer way to multi-label classification. arxiv 2021. arXiv preprint arXiv:2107.10834, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.47, + 0.482, + 0.526 + ], + "angle": 0, + "content": "[53] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3730-3738, 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.528, + 0.482, + 0.582 + ], + "angle": 0, + "content": "[54] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.586, + 0.482, + 0.628 + ], + "angle": 0, + "content": "[55] Ling Luo, Dingyu Xue, and Xinglong Feng. Ehanet: An effective hierarchical aggregation network for face parsing. Applied Sciences, 10(9):3135, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.63, + 0.482, + 0.685 + ], + "angle": 0, + "content": "[56] Upal Mahbub, Sayantan Sarkar, and Rama Chellappa. Segment-based methods for facial attribute detection from partial faces. IEEE Transactions on Affective Computing, 11(4):601-613, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.688, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[57] Longbiao Mao, Yan Yan, Jing-Hao Xue, and Hanzi Wang. Deep multi-task multi-label cnn for effective facial attribute classification. IEEE Transactions on Affective Computing, 13(2):818-828, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.746, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[58] Chen Mi, Baoxi Yuan, Peng Ma, Yingxia Guo, Le Qi, Feng Wang, Wenbo Wu, and Lingling Wang. Visibility prediction based on landmark detection in foggy weather. In 2020 International Conference on Robots & Intelligent System (ICRIS), pages 134-137, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.818, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[59] Paul Micaelli, Arash Vahdat, Hongxu Yin, Jan Kautz, and Pavlo Molchanov. Recurrence without recurrence: Stable video landmark detection with deep equilibrium models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22814-22825, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[60] Zuheng Ming, Junshi Xia, Muhammad Muzzamil Luqman, Jean-Christophe Burie, and Kaixing Zhao. Dynamic multitask learning for face recognition with facial expression. arXiv preprint arXiv:1911.03281, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.15, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[61] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE transactions on pattern analysis and machine intelligence, 41(8):1979–1993, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.221, + 0.905, + 0.276 + ], + "angle": 0, + "content": "[62] Ali Mollahosseini, Behzad Hasani, and Mohammad H Ma-hoor. Affectnet: A database for facial expression, valence, and arousal computing in the wild. IEEE Transactions on Affective Computing, 10(1):18-31, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.278, + 0.905, + 0.359 + ], + "angle": 0, + "content": "[63] Stylianos Moschoglou, Athanasios Papaioannou, Christos Sagonas, Jiankang Deng, Irene Kotsia, and Stefanos Zafeiriou. Agedb: the first manually collected, in-the-wild age database. In proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 51-59, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.362, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[64] Erik Murphy-Chutorian, Anup Doshi, and Mohan Manubhai Trivedi. Head pose estimation for driver assistance systems: A robust algorithm and experimental evaluation. In 2007 IEEE intelligent transportation systems conference, pages 709-714. IEEE, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.433, + 0.907, + 0.502 + ], + "angle": 0, + "content": "[65] Kartik Narayan, Harsh Agarwal, Kartik Thakral, Surbhi Mittal, Mayank Vatsa, and Richa Singh. Df-platter: Multiface heterogeneous deepfake dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9739-9748, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.504, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[66] Kartik Narayan, Nithin Gopalakrishnan Nair, Jennifer Xu, Rama Chellappa, and Vishal M Patel. Petalface: Parameter efficient transfer learning for low-resolution face recognition. arXiv preprint arXiv:2412.07771, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.561, + 0.905, + 0.601 + ], + "angle": 0, + "content": "[67] Kartik Narayan, Vibashan VS, and Vishal M Patel. Segface: Face segmentation of long-tail classes. arXiv preprint arXiv:2412.08647, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.604, + 0.905, + 0.671 + ], + "angle": 0, + "content": "[68] Zhenxing Niu, Mo Zhou, Le Wang, Xinbo Gao, and Gang Hua. Ordinal regression with multiple output cnn for age estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4920-4928, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.675, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[69] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European conference on computer vision, pages 69-84. Springer, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.732, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[70] Jakub Paplham, Vojt Franc, et al. A call to reflect on evaluation practices for age estimation: Comparative analysis of the state-of-the-art and a unified benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1196-1205, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.803, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[71] Lixiong Qin, Mei Wang, Chao Deng, Ke Wang, Xi Chen, Jiani Hu, and Weihong Deng. Swinface: a multi-task transformer for face recognition, expression recognition, age estimation and attribute estimation. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[72] Lixiong Qin, Mei Wang, Xuannan Liu, Yuhang Zhang, Wei Deng, Xiaoshuai Song, Weiran Xu, and Weihong Deng." + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11379" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "Facetor: A generalist model for face perception. In European Conference on Computer Vision, pages 240-260. Springer, 2025. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.136, + 0.483, + 0.219 + ], + "angle": 0, + "content": "[73] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.221, + 0.483, + 0.288 + ], + "angle": 0, + "content": "[74] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.292, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[75] Rajeev Ranjan, Vishal M Patel, and Rama Chellappa. Hyperface: A deep multi-task learning framework for face detection, landmark localization, pose estimation, and gender recognition. IEEE transactions on pattern analysis and machine intelligence, 41(1):121-135, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.363, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[76] Rajeev Ranjan, Swami Sankaranarayanan, Carlos D Castillo, and Rama Chellappa. An all-in-one convolutional neural network for face analysis. In 2017 12th IEEE international conference on automatic face & gesture recognition (FG 2017), pages 17-24. IEEE, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.434, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[77] Ethan M Rudd, Manuel Günther, and Terrance E Boult. Moon: A mixed objective optimization network for the recognition of facial attributes. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14, pages 19-35. Springer, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.519, + 0.482, + 0.573 + ], + "angle": 0, + "content": "[78] Nataniel Ruiz, Eunjji Chong, and James M Rehg. Fine-grained head pose estimation without keypoints. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 2074-2083, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.576, + 0.482, + 0.645 + ], + "angle": 0, + "content": "[79] Christos Sagonas, Georgios Tzimiropoulos, Stefanos Zafeiriou, and Maja Pantic. 300 faces in-the-wild challenge: The first facial landmark localization challenge. In Proceedings of the IEEE international conference on computer vision workshops, pages 397-403, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.647, + 0.483, + 0.715 + ], + "angle": 0, + "content": "[80] Mausoom Sarkar, Mayur Hemani, Rishabh Jain, Balaji Krishnamurthy, et al. Parameter efficient local implicit image function network for face segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20970-20980, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.718, + 0.482, + 0.772 + ], + "angle": 0, + "content": "[81] S. Sengupta, J.C. Cheng, C.D. Castillo, V.M. Patel, R. Chellappa, and D.W. Jacobs. Frontal to profile face verification in the wild. In IEEE Conference on Applications of Computer Vision, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.775, + 0.482, + 0.856 + ], + "angle": 0, + "content": "[82] Jiahui She, Yibo Hu, Hailin Shi, Jun Wang, Qiu Shen, and Tao Mei. Dive into ambiguity: Latent distribution mining and pairwise uncertainty estimation for facial expression recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6248-6257, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.859, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[83] Jie Shen, Stefanos Zafeiriou, Grigoris G Chrysos, Jean Kossaifi, Georgios Tzimiropoulos, and Maja Pantic. The first facial landmark tracking in-the-wild challenge: Benchmark" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.553, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "and results. In Proceedings of the IEEE international conference on computer vision workshops, pages 50-58, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.136, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[84] Nyeong-Ho Shin, Seon-Ho Lee, and Chang-Su Kim. Moving window regression: A novel approach to ordinal regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18760-18769, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.207, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[85] Ying Shu, Yan Yan, Si Chen, Jing-Hao Xue, Chunhua Shen, and Hanzi Wang. Learning spatial-semantic relationship for facial attribute recognition with limited labeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11916–11925, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.291, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[86] Dominykas Strazdas, Jan Hintz, and Ayoub Al-Hamadi. Robo-hud: Interaction concept for contactless operation of industrial cobotic systems. Applied Sciences, 11(12):5366, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.349, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[87] Carole H Sudre, Wenqi Li, Tom Vercauteren, Sebastian Ourselin, and M Jorge Cardoso. Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: Third International Workshop, DLMIA 2017, and 7th International Workshop, ML-CDS 2017, Held in conjunction with MICCAI 2017, Quebec City, QC, Canada, September 14, Proceedings 3, pages 240-248. Springer, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.476, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[88] Haomiao Sun, Mingjie He, Shiguang Shan, Hu Han, and Xilin Chen. Task-adaptive q-face. arXiv preprint arXiv:2405.09059, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.519, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[89] Yi Sun, Yuheng Chen, Xiaogang Wang, and Xiaou Tang. Deep learning face representation by joint identification-verification. Advances in neural information processing systems, 27, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.576, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[90] Yaniv Taigman, Ming Yang, Marc'Aurelio Ranzato, and Lior Wolf. Deepface: Closing the gap to human-level performance in face verification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1701-1708, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.647, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[91] Gusi Te, Yinglu Liu, Wei Hu, Hailin Shi, and Tao Mei. Edge-aware graph representation learning and reasoning for face parsing. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16, pages 258-274. Springer, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.718, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[92] Gusi Te, Wei Hu, Yinglu Liu, Hailin Shi, and Tao Mei. Agrnet: Adaptive graph representation learning and reasoning for face parsing. IEEE Transactions on Image Processing, 30:8236-8250, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.775, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[93] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[94] Roberto Valle, José M Buenapasada, Antonio Valdés, and Luis Baumela. Face alignment using a 3d deeply-initialized ensemble of regression trees. Computer Vision and Image Understanding, 189:102846, 2019. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11380" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[95] Roberto Valle, José M Buenapasada, and Luis Baumela. Multi-task head pose estimation in-the-wild. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43 (8):2874-2881, 2020. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.15, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[96] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.208, + 0.483, + 0.276 + ], + "angle": 0, + "content": "[97] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5265-5274, 2018. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.279, + 0.482, + 0.359 + ], + "angle": 0, + "content": "[98] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 43(10):3349-3364, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.363, + 0.482, + 0.444 + ], + "angle": 0, + "content": "[99] Junke Wang, Dongdong Chen, Zuxuan Wu, Chong Luo, Luowei Zhou, Yucheng Zhao, Yujia Xie, Ce Liu, Yu-Gang Jiang, and Lu Yuan. Omnivl: One foundation model for image-language and video-language tasks. Advances in neural information processing systems, 35:5696-5710, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.447, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[100] Kai Wang, Xiaojiang Peng, Jianfei Yang, Shijian Lu, and Yu Qiao. Suppressing uncertainties for large-scale facial expression recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6897-6906, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.519, + 0.482, + 0.573 + ], + "angle": 0, + "content": "[101] Kai Wang, Xiaojiang Peng, Jianfei Yang, Debin Meng, and Yu Qiao. Region attention networks for pose and occlusion robust facial expression recognition. IEEE Transactions on Image Processing, 29:4057-4069, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.576, + 0.482, + 0.631 + ], + "angle": 0, + "content": "[102] Qianqian Wang, Yen-Yu Chang, Ruojin Cai, Zhengqi Li, Bharath Hariharan, Aleksander Holynski, and Noah Snavely. Tracking everything everywhere all at once. arXiv preprint arXiv:2306.05422, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.633, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[103] Zhenyu Wang, Yali Li, Xi Chen, Ser-Nam Lim, Antonio Torralba, Hengshuang Zhao, and Shengjin Wang. Detecting everything in the open world: Towards universal object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11433-11443, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.718, + 0.482, + 0.785 + ], + "angle": 0, + "content": "[104] Zhen Wei, Yao Sun, Jinqiao Wang, Hanjiang Lai, and Si Liu. Learning adaptive receptive fields for deep image parsing network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2434-2442, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[105] Zhen Wei, Si Liu, Yao Sun, and Hefei Ling. Accurate facial image parsing at real-time speed. IEEE Transactions on Image Processing, 28(9):4659-4670, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[106] Tiancheng Wen, Zhonggan Ding, Yongqiang Yao, Yaxiong Wang, and Xueming Qian. Picassonet: Searching adaptive architecture for efficient facial landmark localization. IEEE Transactions on Neural Networks and Learning Systems, 34(12):10516-10527, 2023. 7" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.16 + ], + "angle": 0, + "content": "[107] Lior Wolf, Tal Hassner, and Yaniv Taigman. Effective unconstrained face recognition by combining multiple descriptors and learned background statistics. IEEE transactions on pattern analysis and machine intelligence, 33(10): 1978-1990, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.163, + 0.905, + 0.244 + ], + "angle": 0, + "content": "[108] Erroll Wood, Tadas Baltrusaitis, Charlie Hewitt, Matthew Johnson, Jingjing Shen, Nikola Milosavljevic, Daniel Wilde, Stephan Garbin, Toby Sharp, Ivan Stojiljković, et al. 3d face reconstruction with dense landmarks. In European Conference on Computer Vision, pages 160-177. Springer, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.247, + 0.905, + 0.314 + ], + "angle": 0, + "content": "[109] Wayne Wu, Chen Qian, Shuo Yang, Quan Wang, Yici Cai, and Qiang Zhou. Look at boundary: A boundary-aware face alignment algorithm. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2129-2138, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[110] Yue Wu and Qiang Ji. Robust facial landmark detection under significant head poses and occlusion. In Proceedings of the IEEE International Conference on Computer Vision, pages 3658-3666, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.372, + 0.905, + 0.439 + ], + "angle": 0, + "content": "[111] Yue Wu, Chao Gou, and Qiang Ji. Simultaneous facial landmark detection, pose and deformation estimation under facial occlusion. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3471-3480, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.442, + 0.905, + 0.522 + ], + "angle": 0, + "content": "[112] Jiahao Xia, Weiwei Qu, Wenjian Huang, Jianguo Zhang, Xi Wang, and Min Xu. Sparse local patch transformer for robust face alignment and landmarks inherent relation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4052-4061, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[113] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.595, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[114] Miao Xin, Shentong Mo, and Yuanze Lin. Eva-gcn: Head pose estimation based on graph convolutional networks. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 1462-1471, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.651, + 0.905, + 0.732 + ], + "angle": 0, + "content": "[115] Xinchen Yan, Jimei Yang, Kihyuk Sohn, and Honglak Lee. Attribute2image: Conditional image generation from visual attributes. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 776-791. Springer, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.734, + 0.905, + 0.803 + ], + "angle": 0, + "content": "[116] Tsun-Yi Yang, Yi-Ting Chen, Yen-Yu Lin, and Yung-Yu Chuang. Fsa-net: Learning fine-grained structure aggregation for head pose estimation from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1087-1096, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[117] Xiangnan Yin, Di Huang, Zehua Fu, Yunhong Wang, and Liming Chen. Segmentation-reconstruction-guided facial image de-occlusion. In 2023 IEEE 17th International Conference on Automatic Face and Gesture Recognition (FG), pages 1–8. IEEE, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[118] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "11381" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[119] Alireza Zaeemzadeh, Shabnam Ghadar, Baldo Faieta, Zhe Lin, Nazanin Rahnavard, Mubarak Shah, and Ratheesh Kalarot. Face image retrieval with attribute manipulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12116-12125, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.483, + 0.262 + ], + "angle": 0, + "content": "[120] Jiabei Zeng, Shiguang Shan, and Xilin Chen. Facial expression recognition with inconsistently annotated datasets. In Proceedings of the European conference on computer vision (ECCV), pages 222-237, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.483, + 0.332 + ], + "angle": 0, + "content": "[121] Cheng Zhang, Hai Liu, Yongjian Deng, Bochen Xie, and Youfu Li. Tokenhpe: Learning orientation tokens for efficient head pose estimation via transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8897-8906, 2023. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.483, + 0.39 + ], + "angle": 0, + "content": "[122] Hongwen Zhang, Qi Li, Zhenan Sun, and Yunfan Liu. Combining data-driven and model-driven methods for robust facial landmark detection. IEEE Transactions on Information Forensics and Security, 13(10):2409-2422, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.483, + 0.461 + ], + "angle": 0, + "content": "[123] Ning Zhang, Manohar Paluri, Marc'Aurelio Ranzato, Trevor Darrell, and Lubomir Bourdev. Panda: Pose aligned networks for deep attribute modeling. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1637-1644, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.483, + 0.532 + ], + "angle": 0, + "content": "[124] Zhanpeng Zhang, Ping Luo, Chen Change Loy, and Xiaou Tang. Facial landmark detection by deep multi-task learning. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VI 13, pages 94-108. Springer, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.483, + 0.589 + ], + "angle": 0, + "content": "[125] Zhifei Zhang, Yang Song, and Hairong Qi. Age progression/regression by conditional adversarial autoencoder. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5810-5818, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.483, + 0.659 + ], + "angle": 0, + "content": "[126] Rui Zhao, Tianshan Liu, Jun Xiao, Daniel PK Lun, and Kin-Man Lam. Deep multi-task learning for facial expression recognition and synthesis based on selective feature sharing. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 4412-4419. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.66, + 0.483, + 0.731 + ], + "angle": 0, + "content": "[127] Qingping Zheng, Jiankang Deng, Zheng Zhu, Ying Li, and Stefanos Zafeiriou. Decoupled multi-task learning with cyclical self-regulation for face parsing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4156-4165, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.483, + 0.788 + ], + "angle": 0, + "content": "[128] Tianyue Zheng and Weihong Deng. Cross-pose lfw: A database for studying cross-pose face recognition in unconstrained environments. Beijing University of Posts and Telecommunications, Tech. Rep, 5(7):5, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[129] Tianyue Zheng, Weihong Deng, and Jiani Hu. Crossage lfw: A database for studying cross-age face recognition in unconstrained environments. arXiv preprint arXiv:1708.08197, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[130] Yinglin Zheng, Hao Yang, Ting Zhang, Jianmin Bao, Dongdong Chen, Yangyu Huang, Lu Yuan, Dong Chen, Ming Zeng, and Fang Wen. General facial representation learning in a visual-linguistic manner. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18697-18709, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.163 + ], + "angle": 0, + "content": "[131] Yijun Zhou and James Gregson. Whenet: Real-time fine-grained estimation for wide range head pose. arXiv preprint arXiv:2005.10353, 2020. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.907, + 0.234 + ], + "angle": 0, + "content": "[132] Zhenglin Zhou, Huaxia Li, Hong Liu, Nanyang Wang, Gang Yu, and Rongrong Ji. Star loss: Reducing semantic ambiguity in facial landmark detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15475-15484, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[133] Jiawen Zhu, Simiao Lai, Xin Chen, Dong Wang, and Huchuan Lu. Visual prompt multi-modal tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9516-9526, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[134] Peihao Zhu, Rameen Abdal, Yipeng Qin, and Peter Wonka. Sean: Image synthesis with semantic region-adaptive normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5104-5113, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[135] Xiangyu Zhu, Zhen Lei, Xiaoming Liu, Hailin Shi, and Stan Z Li. Face alignment across large poses: A 3d solution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 146-155, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[136] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.475, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[137] Ni Zhuang, Yan Yan, Si Chen, and Hanzi Wang. Multi-task learning of cascaded cnn for facial attribute classification. In 2018 24th International Conference on Pattern Recognition (ICPR), pages 2069-2074. IEEE, 2018. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.532, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[138] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.588 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11382" + } + ] +] \ No newline at end of file diff --git a/2025/FaceXFormer_ A Unified Transformer for Facial Analysis/43f275cd-265b-414a-a63d-4d3a6e26c5b2_origin.pdf b/2025/FaceXFormer_ A Unified Transformer for Facial Analysis/43f275cd-265b-414a-a63d-4d3a6e26c5b2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ed1d58ac5f934ae0a41fea98440dc3651a033afb --- /dev/null +++ b/2025/FaceXFormer_ A Unified Transformer for Facial Analysis/43f275cd-265b-414a-a63d-4d3a6e26c5b2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:031a8d0672b92237d8a55179e1df0f7f6b892b679507bc1232a6fbf950716118 +size 1127431 diff --git a/2025/FaceXFormer_ A Unified Transformer for Facial Analysis/full.md b/2025/FaceXFormer_ A Unified Transformer for Facial Analysis/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b2b9c20f042f72aaf94621ad9713642cf649cbed --- /dev/null +++ b/2025/FaceXFormer_ A Unified Transformer for Facial Analysis/full.md @@ -0,0 +1,398 @@ +# FaceXFormer: A Unified Transformer for Facial Analysis + +Kartik Narayan* Vibashan VS* Rama Chellappa Vishal M. Patel {knaraya4, vvishnu2, rchella4, vpatel36}@jhu.edu https://kartik-3004.github.io/facexformer/ + +# Abstract + +In this work, we introduce FaceXFormer, an end-to-end unified transformer model capable of performing ten facial analysis tasks within a single framework. These tasks include face parsing, landmark detection, head pose estimation, attribute prediction, age, gender, and race estimation, facial expression recognition, face recognition, and face visibility. Traditional face analysis approaches rely on task-specific architectures and pre-processing techniques, limiting scalability and integration. In contrast, FaceXFormer employs a transformer-based encoder-decoder architecture, where each task is represented as a learnable token, enabling seamless multi-task processing within a unified model. To enhance efficiency, we introduce FaceX, a lightweight decoder with a novel bi-directional cross-attention mechanism, which jointly processes face and task tokens to learn robust and generalized facial representations. We train FaceXFormer on ten diverse face perception datasets and evaluate it against both specialized and multi-task models across multiple benchmarks, demonstrating state-of-the-art or competitive performance. Additionally, we analyze the impact of various components of FaceXFormer on performance, assess real-world robustness in "in-the-wild" settings, and conduct a performance comparison under different backbones. To the best of our knowledge, FaceXFormer is the first model capable of handling ten facial analysis tasks while maintaining real-time performance at 33.21 FPS. + +# 1. Introduction + +Face analysis is a crucial problem as it has broad range of application such as face verification and identification [89, 90], surveillance [24], face swapping [13], face editing [134], de-occlusion [117], 3D face reconstruction [108], retail [1], image generation [115] and face retrieval [119]. Facial analysis tasks (Figure 1 include face parsing [32, 104], landmarks detection [49, 132], head pose + +![](images/011f675c36c3728dc56d0534340176552e1e5573283a668947f968a7b6f83ca8.jpg) + +![](images/34e47ae5cec6fe0b54160e6ba718692cfc847898aad1c890a23470e0fd9bc08e.jpg) + +![](images/4fd5f53af71db7a0b2293e5ffa098e8b233bf57ffb12f4790325dc9aa8bffc32.jpg) + +![](images/320fad68ac5a570511cd16ca92b3c93e3cedbdb5934c042777b6e7f800992b61.jpg) + +![](images/824afc3c35698bb46270a631fe825c7523021bb2a20a7c3c214a39436efafa9e.jpg) + +![](images/d2b05e3d9db0c9a991e726c6d6b4c1116b70fbd1c454fdd19e4d16b908cbe07c.jpg) + +![](images/4c673f1707ab6d6e1b6ca15e0511c3e24f28bd831f58c6720571f2780c667b15.jpg) +Figure 1. FaceXFormer an end-to-end unified transformer model for 10 different facial analysis tasks such as face parsing, landmark detection, head pose estimation, attributes recognition, age, gender, and race estimation, facial expression recognition, face recognition, and face visibility prediction. + +![](images/aaa3d90d9b69c1c1ef23d3370edf0460ce75e110a59205d69ab14310d6704b6d.jpg) + +![](images/ce8fd9dc97c7eb3f66bc9b90069f95ad19af7b4cffe63d4d648a9741d2585acf.jpg) + +estimation [12, 131], facial attributes recognition [61, 69], age/gender/race estimation [8, 46], facial expression recognition [82], face recognition [37], and face visibility prediction [39, 58]. Therefore, developing a generalized and robust face model for all tasks is a crucial and longstanding problem in the face community. + +Why Unified Model? In recent years, significant advancements have been made in facial analysis, developing state-of-the-art methods and face libraries for various tasks [12, 13, 46, 117, 131, 132]. Despite these methods achieving promising performance, they cannot be integrated into a single pipeline due to their specialized model designs and task-specific pre-processing techniques. Furthermore, deploying multiple specialized models simultaneously is computationally intensive and impractical for real-time applications, leading to increased system complexity and resource consumption. These challenges emphasize the need for a unified model that can concurrently handle multiple facial analysis tasks efficiently (see Table 1). A single model capable of addressing multiple facial tasks is desirable because it: (1) learns a robust and generalized face representation capable of handling "in-the-wild" images; (2) intra-task modeling helps the models to learn task-invariant representation; + +
MethodsFPLDHPEAttrAgeGenRaceVisExpFR
Single-Task Models
DML-CSR [127]✓
FP-LIIF [80]✓
SegFace [67]✓
Wing [22]✓
HRNet [98]✓
WHENet [131]✓
TriNet [9]✓
img2pose [3]✓
TokenHPE [121]✓
SSPL [85]✓
VOLO-D1 [40]✓
DLDL-v2 [23]✓
3DDE [94]✓
MNN [95]✓
KTN [44]✓
DMUE [82]✓
CosFace [97]✓
ArcFace [15]✓
AdaFace [37]✓
Multi-Task Models
SSP+SSG [34]✓✓
Hetero-FAE [27]✓✓✓✓✓
FairFace [35]✓✓✓
MiVOLO [40]✓✓
MTL-CNN [137]✓✓✓
ProS [17]✓✓✓
FaRL [130]✓✓✓✓✓
HyperFace [75]✓✓✓✓✓
AllinOne [76]✓✓✓✓✓✓
Swinface [71]✓✓✓✓
QFace [88]✓✓✓✓
Facetor [72]✓✓✓✓✓✓✓
FaceXFormer✓✓✓✓✓✓✓✓✓✓
+ +Table 1. Comparison with representative methods under different task settings. Our proposed FaceXFormer can perform various facial analysis tasks in a single model. FP - Face Parsing, LD - Landmarks Detection, HPE - Head Pose Estimation, Attr - Attributes Recognition, Age - Age, Gen - Gender, Race - Race Estimation, Exp - Facial Expression Recognition, FR - Face Recognition, and Vis - Face Visibility Prediction + +and (3) simplifies deployment pipelines by reducing computational overhead and enabling faster inference. + +Proposed FaceXFormer Architecture: To this end, we introduce FaceXFormer, an end-to-end unified model designed for ten different facial analysis tasks, as depicted in Figure 1. These tasks include face parsing, landmark detection, head pose estimation, attributes recognition, age/gender/race estimation, facial expression recognition, face recognition and face visibility prediction. FaceXFormer enables task unification by leveraging transformers and learnable tokens as its core components. Specifically, we employ a transformer-based encoder-decoder structure, where the encoder extracts hierarchical face representations and fuses them using a MLP fusion module. The fused features are then processed in the decoder, where each facial analysis task is represented by a unique learnable token, allowing for the simultaneous and effective processing of multiple tasks. In particular, we propose a lightweight de + +coder, FaceX, which processes both face and task tokens together using bi-directional cross-attention mechanism (Section 3.2), enabling the model to learn robust face representations that generalize across various tasks. The bi-directional cross-attention mechanism enables a 2-layer lightweight decoder, allowing the model to operate in real time. After modeling the intra-task and face-token relationships in the FaceX decoder, the task tokens are fed into a unified head, which converts these task tokens into corresponding task predictions. + +Our extensive experiments demonstrate that FaceX-Former achieves state-of-the-art or competitive performance compared to specialized models and existing multitask models across multiple benchmarks, while supporting more tasks than any previous multi-task model. Moreover, we show that our model effectively handles images "in the wild", demonstrating its robustness and generalizability across ten different tasks. This robustness is critical for real-world applications where uncontrolled conditions and diverse inputs are common. FaceXFormer achieves state-of-the-art performance at 33.21 FPS, representing a significant $69.44\%$ speed boost over prior multi-task models, making it highly suitable for real-world applications. + +In summary, our paper's contributions are as follows: + +1. We introduce FaceXFormer, a unified transformer-based framework capable of simultaneously processing ten different facial analysis tasks, achieving real-time performance of 33.21 FPS. +2. We propose FaceX, a lightweight decoder that employs the proposed bi-directional cross-attention mechanism, enabling joint processing of face and task tokens. +3. We conduct extensive experiments and analyses to demonstrate that our approach achieves state-of-the-art performance with reduced inference time compared to specialized and multi-task models across multiple tasks. + +# 2. Related Work + +Facial analysis tasks: Facial analysis tasks involve face parsing [11, 32, 67, 127], landmarks detection [49, 59, 132], head pose estimation [12, 95, 121, 131], facial attributes recognition [61, 69, 85, 130], age/gender/race estimation [8, 40, 43, 46], facial expression recognition [44, 45], face recognition [15, 97] and face visibility prediction [39, 58]. These tasks hold significance in various applications such as face swapping [13, 65], face editing [134], de-occlusion [117], 3D face reconstruction [108], driver assistance [64], human-robot interaction [86], retail [1], face verification and identification [89, 90], image generation [115], image retrieval [119] and surveillance [24, 66]. Specialized models excel in their respective tasks but cannot be easily integrated with other tasks due to the need for extensive task-specific pre-processing [50, 132]. Generally, these models under-perform when applied to tasks beyond + +
QFace [88]Facetor [72]FaceXFormer
Tasks4710
Model Size-178.9 M109.29 M
Face PretrainedYesYesNo
FPS (fp32)-14.3033.21
DecoderQuery2Label [52]Pixel Decoder & Transformer DecoderFaceX
Decoder Layers992
+ +Table 2. Comparison of multi-task face analysis methods. + +their specialization as their design is specific to their designated tasks. Some works [29, 60, 124, 126] perform multiple tasks simultaneously but utilize the additional tasks for guidance or auxiliary loss calculation to enhance the performance of the primary task. + +Multi-task learning for face analysis: HyperFace [75] and AllinOne [76] are early convolution-based models that aim to perform multiple tasks. Recent multi-task frameworks, such as QFace [88] and Facetor [72], are also inspired from DETR [10] and propose a unified model structure consisting of learnable tokens. However, these previous works differ from the proposed method in several key aspects, as summarized in Table 2. Specifically, QFace [88] employs a feature fusion module that uses stage embeddings to aggregate features from the encoder. Facetor [72] introduces a layer-attention mechanism to fuse features from different encoder layers and incorporates two separate decoders. Both methods, employ a 9-layer transformer decoder, also Facetor additionally includes a Pixel Decoder. These architectural components increase computational overhead, resulting in slower inference times. In contrast, FaceXFormer proposes a bi-directional cross-attention mechanism, which enables efficient task-specific feature extraction from face tokens resulting in a 2-layer lightweight decoder. This design choice is the primary reason for FaceXFormer's superior speed and performance. Notably, unlike previous methods, FaceXFormer does not rely on face-specific pertaining backbone. + +Unified transformer models: In recent years, the rise of transformers [19, 96] have paved the way for the unification of multiple tasks within a single architecture. Unified transformer architectures are being explored across various computer vision problems, including segmentation [47, 138], visual question answering (VQA) [99, 118], tracking [102, 133], detection [103]. While these models may not achieve state-of-the-art (SOTA) performance and may under-perform compared to specialized models on some tasks, they demonstrate competitive performance across a variety of tasks. Such unification efforts have led to the development of foundational models like SAM [38], CLIP [73], LLaMA [93], GPT-3 [6], DALL-E [74], etc. However, these models are computationally intensive and not suitable for facial analysis applications that require real-time performance. Motivated by this challenge, we propose FaceXFormer: the first lightweight, transformer-based + +model capable of performing multiple facial analysis tasks. It delivers real-time performance at 33.21 FPS and can be seamlessly integrated into existing systems providing additional annotations for the person of interest. + +# 3. FaceXFormer + +In our framework, we follow a standard encoder-decoder structure as illustrated in Fig. 2. For an input face image $\mathbf{I} \in \mathbb{R}^{H \times W \times 3}$ , we extract coarse to fine-grained multiscale features $\mathbf{S}_i$ , where $i$ belongs to the $i$ -th encoder output. To learn a unified face representation $\mathbf{F}$ , these multiscale features are then fused using a MLP-Fusion $\mathbf{M}$ module. Following fusion, we initialize a series of task-specific tokens $\mathbf{T} = \langle T_1, \dots, T_n \rangle$ , with each $t_i$ representing a face task. Face tokens $\mathbf{F}$ and task tokens $\mathbf{T}$ are then processed by a lightweight Decoder FXDec where task tokens are attended with face tokens to learn relevant task representation. + +$$ +\langle \hat {\mathbf {T}} \rangle = \mathbf {F} \mathbf {X D e c} (\langle \mathbf {F}, \mathbf {T} \rangle ; \mathbf {S} _ {i}) +$$ + +Here, $\hat{\mathbf{T}}$ represents the output task tokens. These tokens are then fed into unified heads, where each task token is refined and passed to its respective task head for prediction. + +# 3.1. Multi-scale Encoder + +In the encoder, we employ a multi-scale encoding strategy to address the varying feature requirements intrinsic to each face analysis task. For instance, age estimation requires a global representation, while face parsing necessitates a fine-grained representation. Given an input image $\mathbf{I}$ , it is processed through a set of encoder layers. For each encoder layer, the output captures information at varying levels of abstraction and detail, generating multi-scale features $\{\mathbf{S}_i\}_{i=1}^n$ , where $i$ ranges from 1 to 4. This results in a hierarchical structure of features, wherein each feature map $\mathbf{S}_i$ transitions from a coarse to a fine-grained representation suitable for diverse facial analysis tasks. + +MLP-Fusion: Assigning each feature-map $\mathbf{S}_i$ to each face task is sub-optimal; rather, learning a unified face representation is more optimal and parameter-efficient. Following [113], we utilize a MLP-Fusion module $\mathbf{M}$ to generate a fused face representation from the multi-scale features $\{\mathbf{S}_i\}_{i=1}^n$ . In this framework, each feature map $\mathbf{S}_i$ is initially passed through a separate MLP layer, standardizing the channel dimensions across scales to facilitate fusion. The transformed features are then concatenated and passed through a fusion MLP layer to aggregate a fused representation $\mathbf{F}$ as follows: + +$$ +\hat {\mathbf {S}} _ {i} = \operatorname {M L P} _ {\operatorname {p r o j}} \left(D _ {i}, D _ {t}\right) \left(\mathbf {S} _ {i}\right), \forall i \in \{1, \dots , n \}, +$$ + +$$ +\mathbf {F} _ {\text {c a t}} = \operatorname {C o n c a t} \left(\hat {\mathbf {S}} _ {1}, \hat {\mathbf {S}} _ {2}, \dots , \hat {\mathbf {S}} _ {n}\right), +$$ + +$$ +\mathbf {F} = \mathbf {M L P} _ {\text {f u s i o n}} (n D _ {t}, D _ {t}) (\mathbf {F} _ {\text {c a t}}), +$$ + +![](images/ed200455b8c4d7533c92c3aa23b01acb1e1a467f965197a79eb665254c14938f.jpg) +Figure 2. Overview of our proposed framework. The FaceXFormer employs an encoder-decoder architecture, extracting multi-scale features from the input face image $\mathbf{I}$ , and fusing them into a unified representation $\mathbf{F}$ via MLP-Fusion. Task tokens $\mathbf{T}$ are processed alongside face representation $\mathbf{F}$ in the FaceX Decoder FXDec, resulting in refined task-specific tokens $\hat{\mathbf{T}}$ . These refined tokens are then used for task-specific predictions by passing through the unified head. FaceXFormer performs ten tasks, including face parsing, landmark detection, head pose estimation, attribute prediction, age, gender, and race estimation, facial expression recognition, face recognition, and face visibility prediction, achieving state-of-the-art performance at a real-time FPS of 33.21. + +where $D_{i}$ and $D_{t}$ are the multi-scale feature channel dimensions of $\mathbf{S}_i$ and the target channel dimension, respectively. The MLP-fusion design ensures minimal computational overhead (983k parameters) while maintaining the ability to perform efficient feature fusion, which is crucial for real-time application based face analysis tasks. + +# 3.2. FaceX Decoder + +Detection transformer (DETR) [10] employs object tokens to learn bounding box predictions for each object. Inspired by this approach, we introduce Task Tokens, whereby each task token is designed to learn specific facial tasks leveraging the fused face representation. However, existing decoders such as DETR [10] and Deformable-DETR [136] are computationally intensive, impacting runtime significantly. To address this, we propose FaceX (FXDec) a lightweight decoder designed to efficiently model the task tokens with face tokens. Specifically, each task token learns a task-related representation by interacting with other task tokens $\mathbf{T}$ and face tokens $\mathbf{F}$ , enhancing the overall representation. The Lightweight Decoder comprises of three main components: 1) Task Self-Attention, 2) Task-to-Face Cross-Attention, and 3) Face-to-Task Cross-Attention as illustrated in Figure 2. + +Task Self-Attention (TSA): The Task Self-Attention module is designed to refine the task-specific representations within the set of task tokens $\mathbf{T} = \langle T_1, \ldots, T_n \rangle$ . Each task + +token $T_{i}$ is an embedded representation that corresponds to a specific facial task. In TSA, each $T_{i}$ is updated by attending to all other task tokens to capture task-specific interactions. Formally, the updated task token $T_{i}^{\prime}$ is computed as: + +$$ +\mathbf {T} _ {i} ^ {\prime} = \operatorname {S e l f A t t n} (\mathbf {Q} = T _ {i}, \mathbf {K} = \mathbf {T}, \mathbf {V} = \mathbf {T}), +$$ + +where Attention denotes the multi-headed self-attention mechanism, and $\mathbf{Q}$ , $\mathbf{K}$ , and $\mathbf{V}$ represent the queries, keys, and values, respectively. Therefore, TSA essentially helps the model to learn task-invariant representation. + +Task-to-Face Cross-Attention (TFCA): The Task-to-Face Cross-Attention module allows each task token to interact with the fused face representation $\mathbf{F}$ . This enables each task token to gather information relevant to its specific facial task from the fused face features. In this module, the fused face representation $\mathbf{F}$ acts as both key and value, while the task tokens serve as queries. The updated task token $\hat{T}_i$ is then computed as follows: + +$$ +\hat {T} _ {i} = \operatorname {C r o s s A t t n} (\mathbf {Q} = T _ {i} ^ {\prime}, \mathbf {K} = \mathbf {F}, \mathbf {V} = \mathbf {F}), +$$ + +where $\hat{\mathbf{T}} = \langle \hat{T}_1,\dots ,\hat{T}_n\rangle$ is the output task token. Thus, TFCA enables direct interaction between the task-specific tokens and the compact facial features, facilitating task-focused feature extraction. + +Face-to-Task Cross-Attention (FTCA): Conversely, the Face-to-Task Cross-Attention module is designed to refine + +the fused face representation $\mathbf{F}$ based on the information from the updated task tokens. This process aids in enhancing the face representation with task-specific details, thereby improving the extraction of overall fused representation. In FTCA, the set of updated task tokens $\mathbf{T}' = \{\mathbf{T}_1'',\mathbf{T}_2'',\dots,\mathbf{T}_m''\}$ acts as both keys and values, while the fused face features $\mathbf{F}$ serve as queries. The refined face representation $\hat{\mathbf{F}}$ is computed as: + +$$ +\hat {\mathbf {F}} = \operatorname {C r o s s A t t n} (\mathbf {Q} = \mathbf {F}, \mathbf {K} = \mathbf {T} ^ {\prime}, \mathbf {V} = \mathbf {T} ^ {\prime}). +$$ + +Through this inverse attention mechanism, the face representation is augmented with critical task-specific details, enabling a robust approach towards facial task unification. + +# 3.3. Unified-Head + +In Unified-Head, the task tokens are processed to obtain corresponding task predictions. As shown in Figure 2, the output face tokens $\hat{\mathbf{F}}$ and task tokens $\hat{\mathbf{T}}$ are processed through a Task-to-Face Cross-Attention mechanism to obtain final refined features. Then, the output tokens are fed into their corresponding task heads. The task head for landmark detection is a hourglass network, for head pose estimation is a regression MLP, and for face recognition is PartialFC [4], while the tasks of age, gender and race estimation, facial expression recognition, face visibility prediction, and attributes prediction utilize classification MLPs. For face parsing, we leverage the output $\hat{\mathbf{F}}$ and process it through an upsampling layer, then perform a cross-product with the face parsing token to obtain a segmentation map. The number of tokens for segmentation corresponds to the total number of classes. For landmark prediction, it corresponds to the number of landmarks (i.e., 68). For head pose estimation, the number of tokens is 9, representing the $3\times 3$ rotation matrix. For other tasks, one token is used for each. + +# 3.4. Multi-Task Training + +We aim to train FaceXFormer for multiple facial analysis tasks simultaneously, however each task requires distinct and sometimes conflicting pre-processing steps. For instance, landmark detection typically requires keypoint alignment of faces, which contradicts the needs for head pose estimation, as it may eliminate the natural variability of headposes. Due to these reasons, integrating all tasks into a single model poses significant challenges. To address this, FaceXFormer incorporates task-specific tokens designed to extract task-specific features from the fused representation. These task tokens compel the backbone to learn a unified representation capable of supporting a broad spectrum of facial analysis tasks. We employ different loss functions for each task and combine them in a joint objective for training. The final loss function is given as: + +$$ +\begin{array}{l} L = \lambda_ {s e g} L _ {s e g} + \lambda_ {l n d} L _ {l n d} + \lambda_ {h p e} L _ {h p e} + \lambda_ {a t t r} L _ {a t t r} + \lambda_ {a} L _ {a} \\ + \lambda_ {g / r} L _ {g / r} + \lambda_ {e x p} L _ {e x p} + \lambda_ {f r} L _ {f r} + \lambda_ {v i s} L _ {v i s} \\ \end{array} +$$ + +where $L_{seg}$ is the mean of dice loss [87] and Cross-Entropy (CE) loss for face parsing, $L_{lnd}$ is STAR loss [132] for landmarks prediction, $L_{hpe}$ is geodesic loss [121] for head pose estimation, $L_{g/r}$ is CE loss for gender/race estimation, $L_a$ is mean of L1 loss and CE loss for age estimation, $L_{exp}$ is CE loss for facial expression recognition, $L_{fr}$ is Arc-Face [15] loss for face recognition, and $L_{attr}$ and $L_{vis}$ are Binary Cross-Entropy with logits loss for attributes prediction and face visibility prediction respectively. + +# 4. Experiments and Results + +To validate the effectiveness of FaceXFormer, we conduct extensive experiments across ten diverse facial analysis tasks using a range of datasets, as described in Section 4.1. Implementation details are provided in Section 4.2 to ensure reproducibility and transparency. Finally, Sections 4.3 and 4.4 present the quantitative and qualitative results, respectively. + +# 4.1. Datasets and Metrics + +We perform co-training, where the model is simultaneously trained for multiple tasks using a total of 10 datasets with task-specific annotations. We conduct a comprehensive evaluation, comparing our approach with both task-specific and multi-task models. We present our results on the test sets according to the standard protocol for each task using the following datasets: + +Train: Face Parsing: CelebAMaskHQ [42]; Landmarks Detection: 300W [79]; Head Pose Estimation: 300W-LP [135]; Attributes Prediction: CelebA [54]; Facial Expression Recognition: RAF-DB [45], AffectNet [62]; Age/Gender/Race estimation: UTKFace [125], FairFace [36]; Face Recognition: MS1MV3 [25]; Visibility Prediction: COFW [7]. + +Test: Face Parsing: CelebAMaskHQ [42]; Landmarks Detection: 300W [135], 300VW [83]; Head Pose Estimation: BIWI [20]; Attributes Prediction: CelebA [54], LFWA [107]; Facial Expression Recognition: RAFDB [45]; Age/Gender/Race Estimation: UTKFace [125], FairFace [36]; Face Recognition: LFW [31], CFP-FP [81], AgeDB [63], CALFW [129], CPLFW [128]; Visibility Prediction: COFW [7]. + +The evaluation metrics used are the F1-score for face parsing, Normalized Mean Error (NME) for landmark prediction, Mean Absolute Error (MAE) for head pose estimation and age estimation, accuracy for facial expression recognition, attributes prediction, gender estimation, race estimation, 1:1 verification accuracy for face recognition, and recall at $80\%$ precision for face visibility prediction. + +# 4.2. Implementation Details + +We train our models using a distributed PyTorch setup on eight A6000 GPUs, each equipped with 48GB of memory. + +
MethodInput Res.SkinHairNoseL-EyeR-EyeL-BrowR-BrowL-LipI-MouthU-LipMean F1 ↑
Wei et al. [105]51296.491.191.987.185.080.882.591.090.687.988.43
EHANet [55]51296.093.993.786.286.583.283.190.393.888.689.53
EAGRNet [91]47396.294.994.088.689.085.785.291.295.088.990.87
AGRNet [92]47396.587.693.988.789.185.585.691.192.089.189.91
FaRL_scratch [130]51296.294.993.889.089.085.385.490.091.788.190.34
DML-CSR [127]47395.794.593.989.489.685.585.791.091.889.190.62
FP-LIIF [80]25696.495.193.788.588.584.584.390.392.187.590.09
SwinFace [71]××××××××××××
QFace [88]××××××××××××
Facetor [72]51296.696.293.989.489.186.286.390.691.689.090.89
FaceXFormer22496.495.793.890.190.386.085.990.692.189.291.01
+ +Table 3. Performance comparison for face parsing on the CelebAMask-HQ dataset [42]. The symbol $\times$ indicates that the model does not perform the corresponding task. Red = First Best, Blue = Second Best. $\times$ indicates a model that doesn't perform the task. + +
MethodsExpression (RAF-DB) ↑MethodsVisibility (COFW) ↑MethodsAge (MAE) UTKFace ↓
DLP-CNN [45]80.89RCPR [7]40OR-CNN [68]5.74
gACNN [48]85.07Wu et al. [111]44.43Axel Berg et al. [5]4.55
IPA2LT [120]86.77Wu et al. [110]49.11CORAL [8]5.47
RAN [101]86.90ECT [122]63.4Gustafsson et al. [26]4.65
CovPool [2]87.003DDE [94]63.89R50-SORD [70]4.36
SCN [100]87.03MNN [95]72.12VOLO-D1 [40]4.23
DACL [21]87.78DLDL-v2 [23]4.42
KTN [44]88.07MWR [84]4.37
DMUE [82]88.76
SwinFace [71]86.54SwinFace [71]×SwinFace [71]×
QFace [88]92.86QFace [88]×QFace [88]×
Facetor [72]87.58Facetor [72]×Facetor [72]4.10
FaceXFormer88.24FaceXFormer72.56FaceXFormer4.17
+ +Table 4. Performance comparison on facial expression recognition, face visibility prediction and age estimation. + +The models' backbones are initialized with ImageNet pretrained weights and processes input images at a resolution of $224 \times 224$ . We employ the AdamW optimizer with a weight decay of $1e^{-5}$ . All models are trained for 12 epochs with a batch size of 48 on each GPU, and an initial learning rate of $1e^{-4}$ , which decays by a factor of 10 at the $6^{th}$ and $10^{th}$ epochs. For some tasks, we train the model for additional epochs to reach the upper ceiling of performance. For data augmentation, we randomly apply Gaussian blur, grayscale conversion, gamma correction, occlusion, horizontal flipping, and affine transformations, such as rotation, translation and scaling. The number of FaceX decoder $N$ is set to two. To ensure stable training across tasks when using multiple datasets of varying sample sizes, we equalize the representation of each task's samples in every batch through upsampling. + +# 4.3. Main results + +In Table 3, Table 4, Table 6, Table 5, we present a comparative analysis of FaceXFormer against recent methods across a variety of tasks. A key highlight of our work is its unique capability to deliver promising results across multiple tasks at real-time inference speed using a single unified model. Specifically, FaceXFormer achieves state-of-the-art performance in face parsing, with a mean F1 score of 91.01 on + +CelebAMaskHQ at a resolution of $224 \times 224$ , which is half the input size required by other SOTA methods. Furthermore, it demonstrates superior performance in head pose estimation and landmark detection, achieving a mean MAE of 3.52 and a mean NME of 4.67, respectively. Additionally, FaceXFormer provides a significant performance boost in attributes prediction and visibility prediction, achieving an accuracy of $91.83\%$ on the CelebA dataset and $72.56\%$ on COFW. It also performs competitively in age estimation, achieving the second-best score of 4.17, and achieves an accuracy of $88.24\%$ in facial expression recognition. In face recognition, FaceXFormer outperforms Faceptor, achieving a mean accuracy of $95.94\%$ compared to $95.28\%$ . However, we observe that multi-task models generally underperform compared to specialized ones in this task. This can be attributed to conflicting training objectives, which force the model to learn identity-invariant features rather than identity-specific representations crucial for accurate recognition. The results on gender estimation across different race categories is shown in Table 8. + +Recent models such as SwinFace [71], QFace [88], and Facetor [72] also aim to unify multiple tasks but only address a subset of them. These approaches often exclude complex tasks such as segmentation, head pose estimation, and landmark prediction. Moreover, they rely on multiple decoders and computationally expensive attention mechanisms, adding to the overall computational overhead. In contrast, FaceXFormer seamlessly unifies these complex tasks using a lightweight decoder and achieves state-of-the-art performance across them at a real-time FPS of 33.21. It outperforms previous multi-task models in segmentation, head pose estimation, landmark prediction, attribute prediction, and face visibility prediction, while achieving the second-best performance in age estimation. In this work, we simultaneously train for ten heterogeneous tasks, posing a more formidable challenge than previous approaches. Despite this, FaceXFormer effectively handles multiple tasks, achieving SOTA or competitive performance in real time. + +
MethodsHeadpose (BIWI)MethodsLandmarks (300W)MethodsCelebA Acc. ↑
YawPitchRollMAE ↓Full ↓Com ↓Chal ↓
HopeNet [78]4.816.613.274.89LAB [109]3.492.985.19PANDA-1 [123]85.43
QuatNet [30]5.494.012.944.15Wing [22]4.043.277.18LNets+ANet [53]87.33
FSA-Net [116]4.275.492.934.14DeCaFa [14]3.392.935.26SSP+SSG [34]88.24
EVA-GCN [114]6.014.782.983.98HRNet [98]3.322.875.15MOON [77]90.94
TriNet [9]4.114.753.043.97PicassoNet [106]3.583.035.81NSA [56]90.61
img2pose [3]4.563.543.243.78AVS+SAN [18]3.863.216.46MCNN-AUX [28]91.29
MNN [95]3.984.612.393.66LUVLi [39]3.232.765.16MCFA [137]91.23
MFDNet [51]3.404.682.773.62HIH [41]3.092.654.89DMM-CNN [57]91.70
TokenHPE [121]3.954.512.713.72PIPNet [33]3.192.784.89SSPL [85]91.77
WHENet [131]3.994.393.063.81SLPT [112]3.172.754.90FaRL [130]91.39
SwinFace [71]××××SwinFace [71]×××SwinFace [71]91.38
QFace [88]----QFace [88]×××QFace [88]91.56
Facector [72]××××Facector [72]3.162.754.84Facector [72]91.39
FaceXFormer3.913.972.673.52FaceXFormer3.052.664.67FaceXFormer91.83
+ +Table 5. Performance comparison on headpose, landmark detection, and attribute recognition. The symbol $\times$ indicates that the model does not perform the corresponding task, while - denotes that results for this dataset are not provided. Red = First Best, Blue = Second Best. + +
MethodLFWCFP-FPAgeDBCALFWCPLFWMean ↑
CosFace [97]99.8198.1298.1195.7692.2896.81
ArcFace [15]99.8398.2798.2895.4592.0896.78
VPL-ArcFace [16]99.8399.1198.6096.1293.4597.42
AdaFace [37]99.8399.1198.1796.0293.9397.41
SwinFace [71]99.8798.6098.1596.1093.4297.22
QFace [88]××××××
Facetor [72]99.4096.3493.6594.7592.2795.28
FaceXFormer99.6896.7596.3595.5091.4195.94
+ +This success can be attributed to the efficiency of the proposed lightweight decoder, which employs a novel bidirectional cross-attention mechanism. + +# 4.4. Qualitative "in-the-wild" results + +In this section, we present the qualitative results of FaceXFormer on randomly selected "in-the-wild" images. We select five random images and showcase the results for face parsing, head pose estimation, landmarks prediction, age estimation, gender and race classification, and attributes prediction in Figure 3. Notably, the model successfully performs complex tasks such as face segmentation, head pose estimation, and landmark prediction, even when input samples exhibit extreme poses, occlusions, or blurring. Furthermore, FaceXFormer can be effectively used as a tool to generate multiple annotations for each image, making it valuable for various downstream tasks. These results highlight FaceXFormer's robust performance in challenging, real-world scenarios. + +# 5. Ablation Study + +In this section, we explore the impact of different components of FaceXFormer on performance. Additionally, we demonstrate that the proposed model exhibits minimal bias + +compared to other models by evaluating age and gender prediction across various demographics. Furthermore, we analyze the impact of using different backbones of varying sizes on the FPS and performance. + +Table 6. Performance comparison for face recognition. + +
MLP FusionSelf Attn.Cross Attn.Bi-dir. CAHPE (MAE)Lnd (NME)Attr. (Acc.)Age (MAE)
✓✓3.704.7091.214.21
✓✓17.1631.4979.9016.15
✓✓✓4.635.3089.984.84
✓✓✓3.524.6791.834.17
+ +Table 7. Impact of various components on performance. + +# 5.1. Impact of various components in FaceXFormer + +To evaluate the contribution of each component in FaceXFormer, we conduct an ablation study focusing on the importance of specific design choices and their impact on performance across various tasks. The results of these experiments are summarized in Table 7. We observe that without MLP fusion (row 1), there is a drop in performance, highlighting the importance of integrating multi-scale features to capture both global and local information essential for accurate predictions. The model performs extremely poorly (row 2) without cross-attention in the decoder, which is expected, as there is no interaction between face tokens and task tokens in this case. Introducing the proposed bidirectional cross-attention (row 4), which corresponds to FaceXFormer, in the decoder provides a significant boost compared to using standard cross-attention (row 3), yielding improvements of 1.11 MAE in head pose estimation, 0.63 NME in landmark detection, 1.85 accuracy points in attribute prediction, and 0.67 MAE in age estimation. These results demonstrate the importance of MLP fusion and bidirectional cross-attention in the FaceXFormer architecture. + +![](images/b851fa0fb1c9084a6105526b5b8f8927c62ace62808298c4e17514eb5a478a32.jpg) +Figure 3. FaceXFormer predictions on "in-the-wild" images + +
ModelData PointsWhiteNon-whiteAverageDiscrepancy
AgeFairFace100K60.0560.6360.520.58
CLIP400M62.2561.9562.00-0.30
FaRL20M61.4961.8461.780.35
FaceXFormer5.9M58.9459.4459.340.50
GenderFairFace100K94.1594.4194.360.26
CLIP400M94.8795.7895.610.91
FaRL20M95.1695.7795.650.61
FaceXFormer5.9M95.3495.1995.22-0.09
+ +# 5.2. Bias Analysis and Ethical Considerations + +In our work, we utilize 17 unique datasets for training and evaluation. We obtained these datasets following the procedures stated on their respective pages and signed the license agreements if and when necessary. As we train our models on multiple datasets designed for different tasks, the subjects across different age groups, genders, and races is not equal. This imbalance may introduce bias in the model. Therefore, we provide an analysis using the FairFace [35] dataset, which is balanced in terms of age, gender and race. We follow [73] and define the "Non-white" group to include multiple racial categories: "Black", "Indian", "East Asian", "Southeast Asian", "Middle Eastern" and "Latino". As can be seen from Table 8, FaceXFormer shows the smallest performance discrepancy across different racial groups and exhibits minimal bias compared to other models despite being trained on fewer data points. This can be attributed to race estimation being the task in co-training. + +# 5.3. Performance under different backbones + +To evaluate the impact of different backbones on performance and FPS, we conduct an ablation study comparing various backbone architectures in FaceXFormer. We categorize head pose estimation, landmark prediction, and age estimation as regression (Reg) tasks, while attribute prediction and facial expression recognition fall under classification (Cls). Additionally, face parsing is denoted as segmentation (Seg). The results are summarized in Table 9. + +Table 8. Age and gender accuracy w.r.t race groups on FairFace + +
BackboneSegRegClsFPSParams
MobileNet90.214.6488.2239.7625.32
ResNet10190.494.3788.9134.9865.54
ConvNext-B91.084.3589.0936.61110.19
Swin-B91.014.1290.0333.21109.29
+ +Table 9. Effect of different backbones on performance and FPS. + +From the results, we observe that ConvNeXt achieves the best performance in segmentation with an F1 score of $91.08\%$ . The Swin Transformer backbone excels in both regression and classification tasks, with a mean error of 4.12 and a mean accuracy of $90.03\%$ , respectively. In contrast, MobileNet demonstrates the lowest performance metrics, including an F1 score of $90.21\%$ and a mean error of 4.64, highlighting its limitations in handling larger, more complex datasets due to its smaller receptive field compared to the Swin Transformer. The selection of the Swin Transformer as the backbone for FaceXFormer is driven by its superior scalability and global contextual understanding, both of which are essential for facial analysis tasks. + +# 6. Conclusion + +FaceXFormer introduces a novel end-to-end unified model that efficiently handles a wide range of facial analysis tasks in real-time. By adopting a transformer-based encoder-decoder architecture and representing each task as a learnable token, our approach seamlessly integrates multiple tasks within a single framework while maintaining minimal computational cost and fast inference times. The proposed lightweight decoder, FaceX, incorporates a novel bi-directional cross-attention mechanism, enhancing the model's ability to learn robust and generalized face representations across diverse tasks. Comprehensive experiments demonstrate that FaceXFormer achieves state-of-the-art performance across multiple facial analysis tasks, achieving a real-time FPS of 33.21. In broader applications, FaceXFormer can serve as an annotator for large-scale face datasets and can be integrated into existing facial analysis systems to provide extra information, making it a valuable tool for surveillance, subject analysis, and image retrieval. + +# Acknowledgment + +This research is based upon work supported in part by the Office of the Director of National Intelligence (ODNI), Intelligence Advanced Research Projects Activity (IARPA), via [2022-21102100005]. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies, either expressed or implied, of ODNI, IARPA, or the U.S. Government. The US Government is authorized to reproduce and distribute reprints for governmental purposes notwithstanding any copyright annotation therein. + +# References + +[1] B Abirami, TS Subashini, and V Mahavaishnavi. Gender and age prediction from real time facial images using cnn. Materials Today: Proceedings, 33:4708-4712, 2020. 1, 2 +[2] Dinesh Acharya, Zhiwu Huang, Danda Pani Paudel, and Luc Van Gool. Covariance pooling for facial expression recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 367-374, 2018. 6 +[3] Vitor Albiero, Xingyu Chen, Xi Yin, Guan Pang, and Tal Hassner. img2pose: Face alignment and detection via 6dof, face pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7617-7627, 2021. 2, 7 +[4] Xiang An, Xuhan Zhu, Yuan Gao, Yang Xiao, Yongle Zhao, Ziyong Feng, Lan Wu, Bin Qin, Ming Zhang, Debing Zhang, et al. Partial fc: Training 10 million identities on a single machine. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1445-1449, 2021. 5 +[5] Axel Berg, Magnus Oskarsson, and Mark O'Connor. Deep ordinal regression with label diversity. In 2020 25th international conference on pattern recognition (ICPR), pages 2740-2747. IEEE, 2021. 6 +[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3 +[7] Xavier P Burgos-Artizzu, Pietro Perona, and Piotr Dólar. Robust face landmark estimation under occlusion. In Proceedings of the IEEE international conference on computer vision, pages 1513-1520, 2013. 5, 6 +[8] Wenzhi Cao, Vahid Mirjalili, and Sebastian Raschka. Rank consistent ordinal regression for neural networks with application to age estimation. Pattern Recognition Letters, 140:325-331, 2020. 1, 2, 6 +[9] Zhiwen Cao, Zongcheng Chu, Dongfang Liu, and Yingjie Chen. A vector-based representation to enhance head pose estimation. In Proceedings of the IEEE/CVF Winter Conference on applications of computer vision, pages 1188-1197, 2021. 2, 7 + +[10] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3, 4 +[11] Liang-Chieh Chen, Yi Yang, Jiang Wang, Wei Xu, and Alan L Yuille. Attention to scale: Scale-aware semantic image segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3640-3649, 2016. 2 +[12] Alejandro Cobo, Roberto Valle, José M Buenapasada, and Luis Baumela. On the representation and methodology for wide and short range head pose estimation. Pattern Recognition, 149:110263, 2024. 1, 2 +[13] Kaiwen Cui, Rongliang Wu, Fangneng Zhan, and Shijian Lu. Face transformer: Towards high fidelity and accurate face swapping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 668-677, 2023. 1, 2 +[14] Arnaud Dapogny, Kevin Bailly, and Matthieu Cord. Decafa: Deep convolutional cascade for face alignment in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6893-6901, 2019. 7 +[15] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 2, 5, 7 +[16] Jiankang Deng, Jia Guo, Jing Yang, Alexandros Lattas, and Stefanos Zafeiriou. Variational prototype learning for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11906-11915, 2021. 7 +[17] Xing Di, Yiyu Zheng, Xiaoming Liu, and Yu Cheng. Pros: Facial omni-representation learning via prototype-based self-distillation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6087-6098, 2024. 2 +[18] Xuanyi Dong, Yan Yan, Wanli Ouyang, and Yi Yang. Style aggregated network for facial landmark detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 379-388, 2018. 7 +[19] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.3 +[20] Gabriele Fanelli, Matthias Dantone, Juergen Gall, Andrea Fossati, and Luc Van Gool. Random forests for real time 3d face analysis. International journal of computer vision, 101:437-458, 2013. 5 +[21] Amir Hossein Farzaneh and Xiaojun Qi. Facial expression recognition in the wild via deep attentive center loss. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2402-2411, 2021. 6 +[22] Zhen-Hua Feng, Josef Kittler, Muhammad Awais, Patrik Huber, and Xiao-Jun Wu. Wing loss for robust facial land + +mark localisation with convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2, 7 +[23] Bin-Bin Gao, Xin-Xin Liu, Hong-Yu Zhou, Jianxin Wu, and Xin Geng. Learning expectation of label distribution for facial age and attractiveness estimation. arXiv preprint arXiv:2007.01771, 2020. 2, 6 +[24] Asma El Kissi Ghalleb, Safa Boumaiza, and Najoua Essoukri Ben Amara. Demographic face profiling based on age, gender and race. In 2020 5th International Conference on Advanced Technologies for Signal and Image Processing (ATSIP), pages 1-6. IEEE, 2020. 1, 2 +[25] Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. Ms-celeb-1m: A dataset and benchmark for large-scale face recognition. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 87-102. Springer, 2016. 5 +[26] Fredrik K Gustafsson, Martin Danelljan, Goutam Bhat, and Thomas B Schon. Energy-based models for deep probabilistic regression. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16, pages 325-343. Springer, 2020. 6 +[27] Hu Han, Anil K Jain, Fang Wang, Shiguang Shan, and Xilin Chen. Heterogeneous face attribute estimation: A deep multi-task learning approach. IEEE transactions on pattern analysis and machine intelligence, 40(11):2597-2609, 2017. 2 +[28] Emily Hand and Rama Chellappa. Attributes for improved attributes: A multi-task network utilizing implicit and explicit relationships for facial attribute classification. In Proceedings of the AAAI conference on artificial intelligence, 2017. 7 +[29] Hui-Lan Hsieh, Winston Hsu, and Yan-Ying Chen. Multi-task learning for face identification and attribute estimation. In 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2981-2985, 2017. 3 +[30] Heng-Wei Hsu, Tung-Yu Wu, Sheng Wan, Wing Hung Wong, and Chen-Yi Lee. Quatnet: Quaternion-based head pose estimation with multiregression loss. IEEE Transactions on Multimedia, 21(4):1035-1046, 2018. 7 +[31] Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. In Workshop on faces in 'Real-Life' Images: detection, alignment, and recognition, 2008. 5 +[32] Aaron S Jackson, Michel Valstar, and Georgios Tzimiropoulos. A cnn cascade for landmark guided semantic part segmentation. In Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part III 14, pages 143-155. Springer, 2016. 1, 2 +[33] Haibo Jin, Shengcai Liao, and Ling Shao. Pixel-in-pixel net: Towards efficient facial landmark detection in the wild. International Journal of Computer Vision, 129(12):3174-3194, 2021. 7 + +[34] Mahdi M Kalayeh, Boqing Gong, and Mubarak Shah. Improving facial attribute prediction using semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6942-6950, 2017. 2, 7 +[35] Kimmo Karkkainen and Jungseock Joo. Fairface: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1548-1558, 2021. 2, 8 +[36] Kimmo Karkkainen and Jungseock Joo. Fairface: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1548-1558, 2021. 5 +[37] Minchul Kim, Anil K Jain, and Xiaoming Liu. Adaface: Quality adaptive margin for face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18750-18759, 2022. 1, 2, 7 +[38] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 3 +[39] Abhinav Kumar, Tim K Marks, Wenxuan Mou, Ye Wang, Michael Jones, Anoop Cherian, Toshiaki Koike-Akino, Xiaoming Liu, and Chen Feng. Luvli face alignment: Estimating landmarks' location, uncertainty, and visibility likelihood. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8236-8246, 2020. 1, 2, 7 +[40] Maksim Kuprashevich and Irina Tolstykh. Mivolo: Multi-input transformer for age and gender estimation. arXiv preprint arXiv:2307.04616, 2023. 2, 6 +[41] Xing Lan, Qinghao Hu, Qiang Chen, Jian Xue, and Jian Cheng. Hih: Towards more accurate face alignment via heatmap in heatmap. arXiv preprint arXiv:2104.03100, 2021.7 +[42] Cheng-Han Lee, Ziwei Liu, Lingyun Wu, and Ping Luo. Maskgan: Towards diverse and interactive facial image manipulation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 5, 6 +[43] Gil Levi and Tal Hassner. Age and gender classification using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 34-42, 2015. 2 +[44] Hangyu Li, Nannan Wang, Xinpeng Ding, Xi Yang, and Xinbo Gao. Adaptively learning facial expression representation via cf labels and distillation. IEEE Transactions on Image Processing, 30:2016-2028, 2021. 2, 6 +[45] Shan Li, Weihong Deng, and JunPing Du. Reliable crowdsourcing and deep locality-preserving learning for expression recognition in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2852-2861, 2017. 2, 5, 6 +[46] Wanhua Li, Xiaoke Huang, Jiwen Lu, Jianjiang Feng, and Jie Zhou. Learning probabilistic ordinal embeddings for uncertainty-aware regression. In Proceedings of the + +IEEE/CVF conference on computer vision and pattern recognition, pages 13896-13905, 2021. 1, 2 +[47] Xiangtai Li, Haobo Yuan, Wei Li, Henghui Ding, Size Wu, Wenwei Zhang, Yining Li, Kai Chen, and Chen Change Loy. Omg-seg: Is one model good enough for all segmentation? arXiv preprint arXiv:2401.10229, 2024. 3 +[48] Yong Li, Jiabei Zeng, Shiguang Shan, and Xilin Chen. Occlusion aware facial expression recognition using cnn with attention mechanism. IEEE Transactions on Image Processing, 28(5):2439-2450, 2018. 6 +[49] Chunze Lin, Beier Zhu, Quan Wang, Renjie Liao, Chen Qian, Jiwen Lu, and Jie Zhou. Structure-coherent deep feature learning for robust face alignment. IEEE Transactions on Image Processing, 30:5313-5326, 2021. 1, 2 +[50] Jinpeng Lin, Hao Yang, Dong Chen, Ming Zeng, Fang Wen, and Lu Yuan. Face parsing with roi tanh-warping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5654-5663, 2019. 2 +[51] Hai Liu, Shuai Fang, Zhaoli Zhang, Duantengchuan Li, Ke Lin, and Jiazhang Wang. Mfdnet: Collaborative poses perception and matrix fisher distribution for head pose estimation. IEEE Transactions on Multimedia, 24:2449-2460, 2021. 7 +[52] S Liu, L Zhang, X Yang, H Su, and J Zhu. Query2label: A simple transformer way to multi-label classification. arxiv 2021. arXiv preprint arXiv:2107.10834, 2021. 3 +[53] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3730-3738, 2015. 7 +[54] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), 2015. 5 +[55] Ling Luo, Dingyu Xue, and Xinglong Feng. Ehanet: An effective hierarchical aggregation network for face parsing. Applied Sciences, 10(9):3135, 2020. 6 +[56] Upal Mahbub, Sayantan Sarkar, and Rama Chellappa. Segment-based methods for facial attribute detection from partial faces. IEEE Transactions on Affective Computing, 11(4):601-613, 2018. 7 +[57] Longbiao Mao, Yan Yan, Jing-Hao Xue, and Hanzi Wang. Deep multi-task multi-label cnn for effective facial attribute classification. IEEE Transactions on Affective Computing, 13(2):818-828, 2020. 7 +[58] Chen Mi, Baoxi Yuan, Peng Ma, Yingxia Guo, Le Qi, Feng Wang, Wenbo Wu, and Lingling Wang. Visibility prediction based on landmark detection in foggy weather. In 2020 International Conference on Robots & Intelligent System (ICRIS), pages 134-137, 2020. 1, 2 +[59] Paul Micaelli, Arash Vahdat, Hongxu Yin, Jan Kautz, and Pavlo Molchanov. Recurrence without recurrence: Stable video landmark detection with deep equilibrium models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22814-22825, 2023. 2 + +[60] Zuheng Ming, Junshi Xia, Muhammad Muzzamil Luqman, Jean-Christophe Burie, and Kaixing Zhao. Dynamic multitask learning for face recognition with facial expression. arXiv preprint arXiv:1911.03281, 2019. 3 +[61] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE transactions on pattern analysis and machine intelligence, 41(8):1979–1993, 2018. 1, 2 +[62] Ali Mollahosseini, Behzad Hasani, and Mohammad H Ma-hoor. Affectnet: A database for facial expression, valence, and arousal computing in the wild. IEEE Transactions on Affective Computing, 10(1):18-31, 2017. 5 +[63] Stylianos Moschoglou, Athanasios Papaioannou, Christos Sagonas, Jiankang Deng, Irene Kotsia, and Stefanos Zafeiriou. Agedb: the first manually collected, in-the-wild age database. In proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 51-59, 2017. 5 +[64] Erik Murphy-Chutorian, Anup Doshi, and Mohan Manubhai Trivedi. Head pose estimation for driver assistance systems: A robust algorithm and experimental evaluation. In 2007 IEEE intelligent transportation systems conference, pages 709-714. IEEE, 2007. 2 +[65] Kartik Narayan, Harsh Agarwal, Kartik Thakral, Surbhi Mittal, Mayank Vatsa, and Richa Singh. Df-platter: Multiface heterogeneous deepfake dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9739-9748, 2023. 2 +[66] Kartik Narayan, Nithin Gopalakrishnan Nair, Jennifer Xu, Rama Chellappa, and Vishal M Patel. Petalface: Parameter efficient transfer learning for low-resolution face recognition. arXiv preprint arXiv:2412.07771, 2024. 2 +[67] Kartik Narayan, Vibashan VS, and Vishal M Patel. Segface: Face segmentation of long-tail classes. arXiv preprint arXiv:2412.08647, 2024. 2 +[68] Zhenxing Niu, Mo Zhou, Le Wang, Xinbo Gao, and Gang Hua. Ordinal regression with multiple output cnn for age estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4920-4928, 2016. 6 +[69] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European conference on computer vision, pages 69-84. Springer, 2016. 1, 2 +[70] Jakub Paplham, Vojt Franc, et al. A call to reflect on evaluation practices for age estimation: Comparative analysis of the state-of-the-art and a unified benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1196-1205, 2024. 6 +[71] Lixiong Qin, Mei Wang, Chao Deng, Ke Wang, Xi Chen, Jiani Hu, and Weihong Deng. Swinface: a multi-task transformer for face recognition, expression recognition, age estimation and attribute estimation. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 2, 6, 7 +[72] Lixiong Qin, Mei Wang, Xuannan Liu, Yuhang Zhang, Wei Deng, Xiaoshuai Song, Weiran Xu, and Weihong Deng. + +Facetor: A generalist model for face perception. In European Conference on Computer Vision, pages 240-260. Springer, 2025. 2, 3, 6, 7 +[73] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 8 +[74] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 3 +[75] Rajeev Ranjan, Vishal M Patel, and Rama Chellappa. Hyperface: A deep multi-task learning framework for face detection, landmark localization, pose estimation, and gender recognition. IEEE transactions on pattern analysis and machine intelligence, 41(1):121-135, 2017. 2, 3 +[76] Rajeev Ranjan, Swami Sankaranarayanan, Carlos D Castillo, and Rama Chellappa. An all-in-one convolutional neural network for face analysis. In 2017 12th IEEE international conference on automatic face & gesture recognition (FG 2017), pages 17-24. IEEE, 2017. 2, 3 +[77] Ethan M Rudd, Manuel Günther, and Terrance E Boult. Moon: A mixed objective optimization network for the recognition of facial attributes. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14, pages 19-35. Springer, 2016. 7 +[78] Nataniel Ruiz, Eunjji Chong, and James M Rehg. Fine-grained head pose estimation without keypoints. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 2074-2083, 2018. 7 +[79] Christos Sagonas, Georgios Tzimiropoulos, Stefanos Zafeiriou, and Maja Pantic. 300 faces in-the-wild challenge: The first facial landmark localization challenge. In Proceedings of the IEEE international conference on computer vision workshops, pages 397-403, 2013. 5 +[80] Mausoom Sarkar, Mayur Hemani, Rishabh Jain, Balaji Krishnamurthy, et al. Parameter efficient local implicit image function network for face segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20970-20980, 2023. 2, 6 +[81] S. Sengupta, J.C. Cheng, C.D. Castillo, V.M. Patel, R. Chellappa, and D.W. Jacobs. Frontal to profile face verification in the wild. In IEEE Conference on Applications of Computer Vision, 2016. 5 +[82] Jiahui She, Yibo Hu, Hailin Shi, Jun Wang, Qiu Shen, and Tao Mei. Dive into ambiguity: Latent distribution mining and pairwise uncertainty estimation for facial expression recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6248-6257, 2021. 1, 2, 6 +[83] Jie Shen, Stefanos Zafeiriou, Grigoris G Chrysos, Jean Kossaifi, Georgios Tzimiropoulos, and Maja Pantic. The first facial landmark tracking in-the-wild challenge: Benchmark + +and results. In Proceedings of the IEEE international conference on computer vision workshops, pages 50-58, 2015. 5 +[84] Nyeong-Ho Shin, Seon-Ho Lee, and Chang-Su Kim. Moving window regression: A novel approach to ordinal regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18760-18769, 2022. 6 +[85] Ying Shu, Yan Yan, Si Chen, Jing-Hao Xue, Chunhua Shen, and Hanzi Wang. Learning spatial-semantic relationship for facial attribute recognition with limited labeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11916–11925, 2021. 2, 7 +[86] Dominykas Strazdas, Jan Hintz, and Ayoub Al-Hamadi. Robo-hud: Interaction concept for contactless operation of industrial cobotic systems. Applied Sciences, 11(12):5366, 2021. 2 +[87] Carole H Sudre, Wenqi Li, Tom Vercauteren, Sebastian Ourselin, and M Jorge Cardoso. Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: Third International Workshop, DLMIA 2017, and 7th International Workshop, ML-CDS 2017, Held in conjunction with MICCAI 2017, Quebec City, QC, Canada, September 14, Proceedings 3, pages 240-248. Springer, 2017. 5 +[88] Haomiao Sun, Mingjie He, Shiguang Shan, Hu Han, and Xilin Chen. Task-adaptive q-face. arXiv preprint arXiv:2405.09059, 2024. 2, 3, 6, 7 +[89] Yi Sun, Yuheng Chen, Xiaogang Wang, and Xiaou Tang. Deep learning face representation by joint identification-verification. Advances in neural information processing systems, 27, 2014. 1, 2 +[90] Yaniv Taigman, Ming Yang, Marc'Aurelio Ranzato, and Lior Wolf. Deepface: Closing the gap to human-level performance in face verification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1701-1708, 2014. 1, 2 +[91] Gusi Te, Yinglu Liu, Wei Hu, Hailin Shi, and Tao Mei. Edge-aware graph representation learning and reasoning for face parsing. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16, pages 258-274. Springer, 2020. 6 +[92] Gusi Te, Wei Hu, Yinglu Liu, Hailin Shi, and Tao Mei. Agrnet: Adaptive graph representation learning and reasoning for face parsing. IEEE Transactions on Image Processing, 30:8236-8250, 2021. 6 +[93] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3 +[94] Roberto Valle, José M Buenapasada, Antonio Valdés, and Luis Baumela. Face alignment using a 3d deeply-initialized ensemble of regression trees. Computer Vision and Image Understanding, 189:102846, 2019. 2, 6 + +[95] Roberto Valle, José M Buenapasada, and Luis Baumela. Multi-task head pose estimation in-the-wild. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43 (8):2874-2881, 2020. 2, 6, 7 +[96] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3 +[97] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5265-5274, 2018. 2, 7 +[98] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 43(10):3349-3364, 2020. 2, 7 +[99] Junke Wang, Dongdong Chen, Zuxuan Wu, Chong Luo, Luowei Zhou, Yucheng Zhao, Yujia Xie, Ce Liu, Yu-Gang Jiang, and Lu Yuan. Omnivl: One foundation model for image-language and video-language tasks. Advances in neural information processing systems, 35:5696-5710, 2022. 3 +[100] Kai Wang, Xiaojiang Peng, Jianfei Yang, Shijian Lu, and Yu Qiao. Suppressing uncertainties for large-scale facial expression recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6897-6906, 2020. 6 +[101] Kai Wang, Xiaojiang Peng, Jianfei Yang, Debin Meng, and Yu Qiao. Region attention networks for pose and occlusion robust facial expression recognition. IEEE Transactions on Image Processing, 29:4057-4069, 2020. 6 +[102] Qianqian Wang, Yen-Yu Chang, Ruojin Cai, Zhengqi Li, Bharath Hariharan, Aleksander Holynski, and Noah Snavely. Tracking everything everywhere all at once. arXiv preprint arXiv:2306.05422, 2023. 3 +[103] Zhenyu Wang, Yali Li, Xi Chen, Ser-Nam Lim, Antonio Torralba, Hengshuang Zhao, and Shengjin Wang. Detecting everything in the open world: Towards universal object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11433-11443, 2023. 3 +[104] Zhen Wei, Yao Sun, Jinqiao Wang, Hanjiang Lai, and Si Liu. Learning adaptive receptive fields for deep image parsing network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2434-2442, 2017. 1 +[105] Zhen Wei, Si Liu, Yao Sun, and Hefei Ling. Accurate facial image parsing at real-time speed. IEEE Transactions on Image Processing, 28(9):4659-4670, 2019. 6 +[106] Tiancheng Wen, Zhonggan Ding, Yongqiang Yao, Yaxiong Wang, and Xueming Qian. Picassonet: Searching adaptive architecture for efficient facial landmark localization. IEEE Transactions on Neural Networks and Learning Systems, 34(12):10516-10527, 2023. 7 + +[107] Lior Wolf, Tal Hassner, and Yaniv Taigman. Effective unconstrained face recognition by combining multiple descriptors and learned background statistics. IEEE transactions on pattern analysis and machine intelligence, 33(10): 1978-1990, 2010. 5 +[108] Erroll Wood, Tadas Baltrusaitis, Charlie Hewitt, Matthew Johnson, Jingjing Shen, Nikola Milosavljevic, Daniel Wilde, Stephan Garbin, Toby Sharp, Ivan Stojiljković, et al. 3d face reconstruction with dense landmarks. In European Conference on Computer Vision, pages 160-177. Springer, 2022. 1, 2 +[109] Wayne Wu, Chen Qian, Shuo Yang, Quan Wang, Yici Cai, and Qiang Zhou. Look at boundary: A boundary-aware face alignment algorithm. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2129-2138, 2018. 7 +[110] Yue Wu and Qiang Ji. Robust facial landmark detection under significant head poses and occlusion. In Proceedings of the IEEE International Conference on Computer Vision, pages 3658-3666, 2015. 6 +[111] Yue Wu, Chao Gou, and Qiang Ji. Simultaneous facial landmark detection, pose and deformation estimation under facial occlusion. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3471-3480, 2017. 6 +[112] Jiahao Xia, Weiwei Qu, Wenjian Huang, Jianguo Zhang, Xi Wang, and Min Xu. Sparse local patch transformer for robust face alignment and landmarks inherent relation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4052-4061, 2022. 7 +[113] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. 3 +[114] Miao Xin, Shentong Mo, and Yuanze Lin. Eva-gcn: Head pose estimation based on graph convolutional networks. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 1462-1471, 2021. 7 +[115] Xinchen Yan, Jimei Yang, Kihyuk Sohn, and Honglak Lee. Attribute2image: Conditional image generation from visual attributes. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 776-791. Springer, 2016. 1, 2 +[116] Tsun-Yi Yang, Yi-Ting Chen, Yen-Yu Lin, and Yung-Yu Chuang. Fsa-net: Learning fine-grained structure aggregation for head pose estimation from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1087-1096, 2019. 7 +[117] Xiangnan Yin, Di Huang, Zehua Fu, Yunhong Wang, and Liming Chen. Segmentation-reconstruction-guided facial image de-occlusion. In 2023 IEEE 17th International Conference on Automatic Face and Gesture Recognition (FG), pages 1–8. IEEE, 2023. 1, 2 +[118] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, + +Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 3 +[119] Alireza Zaeemzadeh, Shabnam Ghadar, Baldo Faieta, Zhe Lin, Nazanin Rahnavard, Mubarak Shah, and Ratheesh Kalarot. Face image retrieval with attribute manipulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12116-12125, 2021. 1, 2 +[120] Jiabei Zeng, Shiguang Shan, and Xilin Chen. Facial expression recognition with inconsistently annotated datasets. In Proceedings of the European conference on computer vision (ECCV), pages 222-237, 2018. 6 +[121] Cheng Zhang, Hai Liu, Yongjian Deng, Bochen Xie, and Youfu Li. Tokenhpe: Learning orientation tokens for efficient head pose estimation via transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8897-8906, 2023. 2, 5, 7 +[122] Hongwen Zhang, Qi Li, Zhenan Sun, and Yunfan Liu. Combining data-driven and model-driven methods for robust facial landmark detection. IEEE Transactions on Information Forensics and Security, 13(10):2409-2422, 2018. 6 +[123] Ning Zhang, Manohar Paluri, Marc'Aurelio Ranzato, Trevor Darrell, and Lubomir Bourdev. Panda: Pose aligned networks for deep attribute modeling. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1637-1644, 2014. 7 +[124] Zhanpeng Zhang, Ping Luo, Chen Change Loy, and Xiaou Tang. Facial landmark detection by deep multi-task learning. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VI 13, pages 94-108. Springer, 2014. 3 +[125] Zhifei Zhang, Yang Song, and Hairong Qi. Age progression/regression by conditional adversarial autoencoder. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5810-5818, 2017. 5 +[126] Rui Zhao, Tianshan Liu, Jun Xiao, Daniel PK Lun, and Kin-Man Lam. Deep multi-task learning for facial expression recognition and synthesis based on selective feature sharing. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 4412-4419. IEEE, 2021. 3 +[127] Qingping Zheng, Jiankang Deng, Zheng Zhu, Ying Li, and Stefanos Zafeiriou. Decoupled multi-task learning with cyclical self-regulation for face parsing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4156-4165, 2022. 2, 6 +[128] Tianyue Zheng and Weihong Deng. Cross-pose lfw: A database for studying cross-pose face recognition in unconstrained environments. Beijing University of Posts and Telecommunications, Tech. Rep, 5(7):5, 2018. 5 +[129] Tianyue Zheng, Weihong Deng, and Jiani Hu. Crossage lfw: A database for studying cross-age face recognition in unconstrained environments. arXiv preprint arXiv:1708.08197, 2017. 5 +[130] Yinglin Zheng, Hao Yang, Ting Zhang, Jianmin Bao, Dongdong Chen, Yangyu Huang, Lu Yuan, Dong Chen, Ming Zeng, and Fang Wen. General facial representation learning in a visual-linguistic manner. In Proceedings of the + +IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18697-18709, 2022. 2, 6, 7 +[131] Yijun Zhou and James Gregson. Whenet: Real-time fine-grained estimation for wide range head pose. arXiv preprint arXiv:2005.10353, 2020. 1, 2, 7 +[132] Zhenglin Zhou, Huaxia Li, Hong Liu, Nanyang Wang, Gang Yu, and Rongrong Ji. Star loss: Reducing semantic ambiguity in facial landmark detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15475-15484, 2023. 1, 2, 5 +[133] Jiawen Zhu, Simiao Lai, Xin Chen, Dong Wang, and Huchuan Lu. Visual prompt multi-modal tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9516-9526, 2023. 3 +[134] Peihao Zhu, Rameen Abdal, Yipeng Qin, and Peter Wonka. Sean: Image synthesis with semantic region-adaptive normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5104-5113, 2020. 1, 2 +[135] Xiangyu Zhu, Zhen Lei, Xiaoming Liu, Hailin Shi, and Stan Z Li. Face alignment across large poses: A 3d solution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 146-155, 2016. 5 +[136] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 4 +[137] Ni Zhuang, Yan Yan, Si Chen, and Hanzi Wang. Multi-task learning of cascaded cnn for facial attribute classification. In 2018 24th International Conference on Pattern Recognition (ICPR), pages 2069-2074. IEEE, 2018. 2, 7 +[138] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. Advances in Neural Information Processing Systems, 36, 2024. 3 \ No newline at end of file diff --git a/2025/FaceXFormer_ A Unified Transformer for Facial Analysis/images.zip b/2025/FaceXFormer_ A Unified Transformer for Facial Analysis/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6414c7fc9f54e427b682ba0f130a3ff41fc04cab --- /dev/null +++ b/2025/FaceXFormer_ A Unified Transformer for Facial Analysis/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d8d0f34764918f32f7c4347e50d2ad3cc6d929ef16cbaa64713b97a5c4fd9a43 +size 727880 diff --git a/2025/FaceXFormer_ A Unified Transformer for Facial Analysis/layout.json b/2025/FaceXFormer_ A Unified Transformer for Facial Analysis/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f46754549348e2393cb2d789ed93e50a87e95980 --- /dev/null +++ b/2025/FaceXFormer_ A Unified Transformer for Facial Analysis/layout.json @@ -0,0 +1,11300 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 128, + 103, + 482, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 103, + 482, + 121 + ], + "spans": [ + { + "bbox": [ + 128, + 103, + 482, + 121 + ], + "type": "text", + "content": "FaceXFormer: A Unified Transformer for Facial Analysis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 136, + 143, + 475, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 143, + 475, + 185 + ], + "spans": [ + { + "bbox": [ + 136, + 143, + 475, + 185 + ], + "type": "text", + "content": "Kartik Narayan* Vibashan VS* Rama Chellappa Vishal M. Patel {knaraya4, vvishnu2, rchella4, vpatel36}@jhu.edu https://kartik-3004.github.io/facexformer/" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 239, + 297, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 239, + 297, + 563 + ], + "spans": [ + { + "bbox": [ + 54, + 239, + 297, + 563 + ], + "type": "text", + "content": "In this work, we introduce FaceXFormer, an end-to-end unified transformer model capable of performing ten facial analysis tasks within a single framework. These tasks include face parsing, landmark detection, head pose estimation, attribute prediction, age, gender, and race estimation, facial expression recognition, face recognition, and face visibility. Traditional face analysis approaches rely on task-specific architectures and pre-processing techniques, limiting scalability and integration. In contrast, FaceXFormer employs a transformer-based encoder-decoder architecture, where each task is represented as a learnable token, enabling seamless multi-task processing within a unified model. To enhance efficiency, we introduce FaceX, a lightweight decoder with a novel bi-directional cross-attention mechanism, which jointly processes face and task tokens to learn robust and generalized facial representations. We train FaceXFormer on ten diverse face perception datasets and evaluate it against both specialized and multi-task models across multiple benchmarks, demonstrating state-of-the-art or competitive performance. Additionally, we analyze the impact of various components of FaceXFormer on performance, assess real-world robustness in \"in-the-wild\" settings, and conduct a performance comparison under different backbones. To the best of our knowledge, FaceXFormer is the first model capable of handling ten facial analysis tasks while maintaining real-time performance at 33.21 FPS." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 587, + 135, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 587, + 135, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 587, + 135, + 600 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 608, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 608, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 608, + 296, + 693 + ], + "type": "text", + "content": "Face analysis is a crucial problem as it has broad range of application such as face verification and identification [89, 90], surveillance [24], face swapping [13], face editing [134], de-occlusion [117], 3D face reconstruction [108], retail [1], image generation [115] and face retrieval [119]. Facial analysis tasks (Figure 1 include face parsing [32, 104], landmarks detection [49, 132], head pose" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 318, + 213, + 392, + 257 + ], + "blocks": [ + { + "bbox": [ + 318, + 213, + 392, + 257 + ], + "lines": [ + { + "bbox": [ + 318, + 213, + 392, + 257 + ], + "spans": [ + { + "bbox": [ + 318, + 213, + 392, + 257 + ], + "type": "image", + "image_path": "011f675c36c3728dc56d0534340176552e1e5573283a668947f968a7b6f83ca8.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 395, + 213, + 470, + 257 + ], + "blocks": [ + { + "bbox": [ + 395, + 213, + 470, + 257 + ], + "lines": [ + { + "bbox": [ + 395, + 213, + 470, + 257 + ], + "spans": [ + { + "bbox": [ + 395, + 213, + 470, + 257 + ], + "type": "image", + "image_path": "34e47ae5cec6fe0b54160e6ba718692cfc847898aad1c890a23470e0fd9bc08e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 471, + 213, + 545, + 257 + ], + "blocks": [ + { + "bbox": [ + 471, + 213, + 545, + 257 + ], + "lines": [ + { + "bbox": [ + 471, + 213, + 545, + 257 + ], + "spans": [ + { + "bbox": [ + 471, + 213, + 545, + 257 + ], + "type": "image", + "image_path": "4fd5f53af71db7a0b2293e5ffa098e8b233bf57ffb12f4790325dc9aa8bffc32.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 318, + 258, + 392, + 304 + ], + "blocks": [ + { + "bbox": [ + 318, + 258, + 392, + 304 + ], + "lines": [ + { + "bbox": [ + 318, + 258, + 392, + 304 + ], + "spans": [ + { + "bbox": [ + 318, + 258, + 392, + 304 + ], + "type": "image", + "image_path": "320fad68ac5a570511cd16ca92b3c93e3cedbdb5934c042777b6e7f800992b61.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 395, + 258, + 470, + 304 + ], + "blocks": [ + { + "bbox": [ + 395, + 258, + 470, + 304 + ], + "lines": [ + { + "bbox": [ + 395, + 258, + 470, + 304 + ], + "spans": [ + { + "bbox": [ + 395, + 258, + 470, + 304 + ], + "type": "image", + "image_path": "824afc3c35698bb46270a631fe825c7523021bb2a20a7c3c214a39436efafa9e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 471, + 258, + 545, + 304 + ], + "blocks": [ + { + "bbox": [ + 471, + 258, + 545, + 304 + ], + "lines": [ + { + "bbox": [ + 471, + 258, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 471, + 258, + 545, + 304 + ], + "type": "image", + "image_path": "d2b05e3d9db0c9a991e726c6d6b4c1116b70fbd1c454fdd19e4d16b908cbe07c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 318, + 304, + 392, + 351 + ], + "blocks": [ + { + "bbox": [ + 318, + 304, + 392, + 351 + ], + "lines": [ + { + "bbox": [ + 318, + 304, + 392, + 351 + ], + "spans": [ + { + "bbox": [ + 318, + 304, + 392, + 351 + ], + "type": "image", + "image_path": "4c673f1707ab6d6e1b6ca15e0511c3e24f28bd831f58c6720571f2780c667b15.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 360, + 555, + 416 + ], + "lines": [ + { + "bbox": [ + 313, + 360, + 555, + 416 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 555, + 416 + ], + "type": "text", + "content": "Figure 1. FaceXFormer an end-to-end unified transformer model for 10 different facial analysis tasks such as face parsing, landmark detection, head pose estimation, attributes recognition, age, gender, and race estimation, facial expression recognition, face recognition, and face visibility prediction." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 395, + 304, + 470, + 351 + ], + "blocks": [ + { + "bbox": [ + 395, + 304, + 470, + 351 + ], + "lines": [ + { + "bbox": [ + 395, + 304, + 470, + 351 + ], + "spans": [ + { + "bbox": [ + 395, + 304, + 470, + 351 + ], + "type": "image", + "image_path": "aaa3d90d9b69c1c1ef23d3370edf0460ce75e110a59205d69ab14310d6704b6d.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 471, + 304, + 545, + 351 + ], + "blocks": [ + { + "bbox": [ + 471, + 304, + 545, + 351 + ], + "lines": [ + { + "bbox": [ + 471, + 304, + 545, + 351 + ], + "spans": [ + { + "bbox": [ + 471, + 304, + 545, + 351 + ], + "type": "image", + "image_path": "ce8fd9dc97c7eb3f66bc9b90069f95ad19af7b4cffe63d4d648a9741d2585acf.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 437, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 510 + ], + "type": "text", + "content": "estimation [12, 131], facial attributes recognition [61, 69], age/gender/race estimation [8, 46], facial expression recognition [82], face recognition [37], and face visibility prediction [39, 58]. Therefore, developing a generalized and robust face model for all tasks is a crucial and longstanding problem in the face community." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": "Why Unified Model? In recent years, significant advancements have been made in facial analysis, developing state-of-the-art methods and face libraries for various tasks [12, 13, 46, 117, 131, 132]. Despite these methods achieving promising performance, they cannot be integrated into a single pipeline due to their specialized model designs and task-specific pre-processing techniques. Furthermore, deploying multiple specialized models simultaneously is computationally intensive and impractical for real-time applications, leading to increased system complexity and resource consumption. These challenges emphasize the need for a unified model that can concurrently handle multiple facial analysis tasks efficiently (see Table 1). A single model capable of addressing multiple facial tasks is desirable because it: (1) learns a robust and generalized face representation capable of handling \"in-the-wild\" images; (2) intra-task modeling helps the models to learn task-invariant representation;" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 133, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 133, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 133, + 712 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11369" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 71, + 296, + 388 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 296, + 388 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 296, + 388 + ], + "type": "table", + "html": "
MethodsFPLDHPEAttrAgeGenRaceVisExpFR
Single-Task Models
DML-CSR [127]✓
FP-LIIF [80]✓
SegFace [67]✓
Wing [22]✓
HRNet [98]✓
WHENet [131]✓
TriNet [9]✓
img2pose [3]✓
TokenHPE [121]✓
SSPL [85]✓
VOLO-D1 [40]✓
DLDL-v2 [23]✓
3DDE [94]✓
MNN [95]✓
KTN [44]✓
DMUE [82]✓
CosFace [97]✓
ArcFace [15]✓
AdaFace [37]✓
Multi-Task Models
SSP+SSG [34]✓✓
Hetero-FAE [27]✓✓✓✓✓
FairFace [35]✓✓✓
MiVOLO [40]✓✓
MTL-CNN [137]✓✓✓
ProS [17]✓✓✓
FaRL [130]✓✓✓✓✓
HyperFace [75]✓✓✓✓✓
AllinOne [76]✓✓✓✓✓✓
Swinface [71]✓✓✓✓
QFace [88]✓✓✓✓
Facetor [72]✓✓✓✓✓✓✓
FaceXFormer✓✓✓✓✓✓✓✓✓✓
", + "image_path": "1c6479d6fd4d0219ddc7c277b09279e94a29005c48a5bc91180cc35763279bd8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 398, + 296, + 475 + ], + "lines": [ + { + "bbox": [ + 55, + 398, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 398, + 296, + 475 + ], + "type": "text", + "content": "Table 1. Comparison with representative methods under different task settings. Our proposed FaceXFormer can perform various facial analysis tasks in a single model. FP - Face Parsing, LD - Landmarks Detection, HPE - Head Pose Estimation, Attr - Attributes Recognition, Age - Age, Gen - Gender, Race - Race Estimation, Exp - Facial Expression Recognition, FR - Face Recognition, and Vis - Face Visibility Prediction" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 498, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 521 + ], + "type": "text", + "content": "and (3) simplifies deployment pipelines by reducing computational overhead and enabling faster inference." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 522, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 522, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 522, + 295, + 713 + ], + "type": "text", + "content": "Proposed FaceXFormer Architecture: To this end, we introduce FaceXFormer, an end-to-end unified model designed for ten different facial analysis tasks, as depicted in Figure 1. These tasks include face parsing, landmark detection, head pose estimation, attributes recognition, age/gender/race estimation, facial expression recognition, face recognition and face visibility prediction. FaceXFormer enables task unification by leveraging transformers and learnable tokens as its core components. Specifically, we employ a transformer-based encoder-decoder structure, where the encoder extracts hierarchical face representations and fuses them using a MLP fusion module. The fused features are then processed in the decoder, where each facial analysis task is represented by a unique learnable token, allowing for the simultaneous and effective processing of multiple tasks. In particular, we propose a lightweight de" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "content": "coder, FaceX, which processes both face and task tokens together using bi-directional cross-attention mechanism (Section 3.2), enabling the model to learn robust face representations that generalize across various tasks. The bi-directional cross-attention mechanism enables a 2-layer lightweight decoder, allowing the model to operate in real time. After modeling the intra-task and face-token relationships in the FaceX decoder, the task tokens are fed into a unified head, which converts these task tokens into corresponding task predictions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 192, + 555, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 555, + 347 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 555, + 347 + ], + "type": "text", + "content": "Our extensive experiments demonstrate that FaceX-Former achieves state-of-the-art or competitive performance compared to specialized models and existing multitask models across multiple benchmarks, while supporting more tasks than any previous multi-task model. Moreover, we show that our model effectively handles images \"in the wild\", demonstrating its robustness and generalizability across ten different tasks. This robustness is critical for real-world applications where uncontrolled conditions and diverse inputs are common. FaceXFormer achieves state-of-the-art performance at 33.21 FPS, representing a significant " + }, + { + "bbox": [ + 313, + 192, + 555, + 347 + ], + "type": "inline_equation", + "content": "69.44\\%" + }, + { + "bbox": [ + 313, + 192, + 555, + 347 + ], + "type": "text", + "content": " speed boost over prior multi-task models, making it highly suitable for real-world applications." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 326, + 348, + 541, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 348, + 541, + 359 + ], + "spans": [ + { + "bbox": [ + 326, + 348, + 541, + 359 + ], + "type": "text", + "content": "In summary, our paper's contributions are as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 360, + 553, + 491 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 315, + 360, + 553, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 360, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 315, + 360, + 553, + 407 + ], + "type": "text", + "content": "1. We introduce FaceXFormer, a unified transformer-based framework capable of simultaneously processing ten different facial analysis tasks, achieving real-time performance of 33.21 FPS." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 407, + 553, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 407, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 314, + 407, + 553, + 443 + ], + "type": "text", + "content": "2. We propose FaceX, a lightweight decoder that employs the proposed bi-directional cross-attention mechanism, enabling joint processing of face and task tokens." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 443, + 553, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 443, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 314, + 443, + 553, + 491 + ], + "type": "text", + "content": "3. We conduct extensive experiments and analyses to demonstrate that our approach achieves state-of-the-art performance with reduced inference time compared to specialized and multi-task models across multiple tasks." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 502, + 400, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 502, + 400, + 514 + ], + "spans": [ + { + "bbox": [ + 314, + 502, + 400, + 514 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": "Facial analysis tasks: Facial analysis tasks involve face parsing [11, 32, 67, 127], landmarks detection [49, 59, 132], head pose estimation [12, 95, 121, 131], facial attributes recognition [61, 69, 85, 130], age/gender/race estimation [8, 40, 43, 46], facial expression recognition [44, 45], face recognition [15, 97] and face visibility prediction [39, 58]. These tasks hold significance in various applications such as face swapping [13, 65], face editing [134], de-occlusion [117], 3D face reconstruction [108], driver assistance [64], human-robot interaction [86], retail [1], face verification and identification [89, 90], image generation [115], image retrieval [119] and surveillance [24, 66]. Specialized models excel in their respective tasks but cannot be easily integrated with other tasks due to the need for extensive task-specific pre-processing [50, 132]. Generally, these models under-perform when applied to tasks beyond" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "11370" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 293, + 148 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 293, + 148 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 293, + 148 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 293, + 148 + ], + "type": "table", + "html": "
QFace [88]Facetor [72]FaceXFormer
Tasks4710
Model Size-178.9 M109.29 M
Face PretrainedYesYesNo
FPS (fp32)-14.3033.21
DecoderQuery2Label [52]Pixel Decoder & Transformer DecoderFaceX
Decoder Layers992
", + "image_path": "7ff01b30d6774eba61d478272a32effcb7d0568fe9bcb496f4e4646796813771.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 70, + 156, + 279, + 167 + ], + "lines": [ + { + "bbox": [ + 70, + 156, + 279, + 167 + ], + "spans": [ + { + "bbox": [ + 70, + 156, + 279, + 167 + ], + "type": "text", + "content": "Table 2. Comparison of multi-task face analysis methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 183, + 295, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 183, + 295, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 183, + 295, + 243 + ], + "type": "text", + "content": "their specialization as their design is specific to their designated tasks. Some works [29, 60, 124, 126] perform multiple tasks simultaneously but utilize the additional tasks for guidance or auxiliary loss calculation to enhance the performance of the primary task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 245, + 296, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 245, + 296, + 508 + ], + "spans": [ + { + "bbox": [ + 56, + 245, + 296, + 508 + ], + "type": "text", + "content": "Multi-task learning for face analysis: HyperFace [75] and AllinOne [76] are early convolution-based models that aim to perform multiple tasks. Recent multi-task frameworks, such as QFace [88] and Facetor [72], are also inspired from DETR [10] and propose a unified model structure consisting of learnable tokens. However, these previous works differ from the proposed method in several key aspects, as summarized in Table 2. Specifically, QFace [88] employs a feature fusion module that uses stage embeddings to aggregate features from the encoder. Facetor [72] introduces a layer-attention mechanism to fuse features from different encoder layers and incorporates two separate decoders. Both methods, employ a 9-layer transformer decoder, also Facetor additionally includes a Pixel Decoder. These architectural components increase computational overhead, resulting in slower inference times. In contrast, FaceXFormer proposes a bi-directional cross-attention mechanism, which enables efficient task-specific feature extraction from face tokens resulting in a 2-layer lightweight decoder. This design choice is the primary reason for FaceXFormer's superior speed and performance. Notably, unlike previous methods, FaceXFormer does not rely on face-specific pertaining backbone." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": "Unified transformer models: In recent years, the rise of transformers [19, 96] have paved the way for the unification of multiple tasks within a single architecture. Unified transformer architectures are being explored across various computer vision problems, including segmentation [47, 138], visual question answering (VQA) [99, 118], tracking [102, 133], detection [103]. While these models may not achieve state-of-the-art (SOTA) performance and may under-perform compared to specialized models on some tasks, they demonstrate competitive performance across a variety of tasks. Such unification efforts have led to the development of foundational models like SAM [38], CLIP [73], LLaMA [93], GPT-3 [6], DALL-E [74], etc. However, these models are computationally intensive and not suitable for facial analysis applications that require real-time performance. Motivated by this challenge, we propose FaceXFormer: the first lightweight, transformer-based" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "model capable of performing multiple facial analysis tasks. It delivers real-time performance at 33.21 FPS and can be seamlessly integrated into existing systems providing additional annotations for the person of interest." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 131, + 401, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 131, + 401, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 131, + 401, + 144 + ], + "type": "text", + "content": "3. FaceXFormer" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 152, + 555, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 152, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 555, + 293 + ], + "type": "text", + "content": "In our framework, we follow a standard encoder-decoder structure as illustrated in Fig. 2. For an input face image " + }, + { + "bbox": [ + 313, + 152, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 313, + 152, + 555, + 293 + ], + "type": "text", + "content": ", we extract coarse to fine-grained multiscale features " + }, + { + "bbox": [ + 313, + 152, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_i" + }, + { + "bbox": [ + 313, + 152, + 555, + 293 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 152, + 555, + 293 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 152, + 555, + 293 + ], + "type": "text", + "content": " belongs to the " + }, + { + "bbox": [ + 313, + 152, + 555, + 293 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 152, + 555, + 293 + ], + "type": "text", + "content": "-th encoder output. To learn a unified face representation " + }, + { + "bbox": [ + 313, + 152, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 313, + 152, + 555, + 293 + ], + "type": "text", + "content": ", these multiscale features are then fused using a MLP-Fusion " + }, + { + "bbox": [ + 313, + 152, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 313, + 152, + 555, + 293 + ], + "type": "text", + "content": " module. Following fusion, we initialize a series of task-specific tokens " + }, + { + "bbox": [ + 313, + 152, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{T} = \\langle T_1, \\dots, T_n \\rangle" + }, + { + "bbox": [ + 313, + 152, + 555, + 293 + ], + "type": "text", + "content": ", with each " + }, + { + "bbox": [ + 313, + 152, + 555, + 293 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 313, + 152, + 555, + 293 + ], + "type": "text", + "content": " representing a face task. Face tokens " + }, + { + "bbox": [ + 313, + 152, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 313, + 152, + 555, + 293 + ], + "type": "text", + "content": " and task tokens " + }, + { + "bbox": [ + 313, + 152, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 313, + 152, + 555, + 293 + ], + "type": "text", + "content": " are then processed by a lightweight Decoder FXDec where task tokens are attended with face tokens to learn relevant task representation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 375, + 293, + 493, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 293, + 493, + 308 + ], + "spans": [ + { + "bbox": [ + 375, + 293, + 493, + 308 + ], + "type": "interline_equation", + "content": "\\langle \\hat {\\mathbf {T}} \\rangle = \\mathbf {F} \\mathbf {X D e c} (\\langle \\mathbf {F}, \\mathbf {T} \\rangle ; \\mathbf {S} _ {i})", + "image_path": "c3160f287bbbebccb8a9f81c4d3f002c77adfab52a0cee874d865658bf22fa9b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 313, + 555, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 313, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 313, + 555, + 350 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 313, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{T}}" + }, + { + "bbox": [ + 313, + 313, + 555, + 350 + ], + "type": "text", + "content": " represents the output task tokens. These tokens are then fed into unified heads, where each task token is refined and passed to its respective task head for prediction." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 357, + 432, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 432, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 432, + 369 + ], + "type": "text", + "content": "3.1. Multi-scale Encoder" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 375, + 555, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 375, + 555, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 555, + 518 + ], + "type": "text", + "content": "In the encoder, we employ a multi-scale encoding strategy to address the varying feature requirements intrinsic to each face analysis task. For instance, age estimation requires a global representation, while face parsing necessitates a fine-grained representation. Given an input image " + }, + { + "bbox": [ + 313, + 375, + 555, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 313, + 375, + 555, + 518 + ], + "type": "text", + "content": ", it is processed through a set of encoder layers. For each encoder layer, the output captures information at varying levels of abstraction and detail, generating multi-scale features " + }, + { + "bbox": [ + 313, + 375, + 555, + 518 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{S}_i\\}_{i=1}^n" + }, + { + "bbox": [ + 313, + 375, + 555, + 518 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 375, + 555, + 518 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 375, + 555, + 518 + ], + "type": "text", + "content": " ranges from 1 to 4. This results in a hierarchical structure of features, wherein each feature map " + }, + { + "bbox": [ + 313, + 375, + 555, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_i" + }, + { + "bbox": [ + 313, + 375, + 555, + 518 + ], + "type": "text", + "content": " transitions from a coarse to a fine-grained representation suitable for diverse facial analysis tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 530, + 556, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 530, + 556, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 530, + 556, + 661 + ], + "type": "text", + "content": "MLP-Fusion: Assigning each feature-map " + }, + { + "bbox": [ + 313, + 530, + 556, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_i" + }, + { + "bbox": [ + 313, + 530, + 556, + 661 + ], + "type": "text", + "content": " to each face task is sub-optimal; rather, learning a unified face representation is more optimal and parameter-efficient. Following [113], we utilize a MLP-Fusion module " + }, + { + "bbox": [ + 313, + 530, + 556, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 313, + 530, + 556, + 661 + ], + "type": "text", + "content": " to generate a fused face representation from the multi-scale features " + }, + { + "bbox": [ + 313, + 530, + 556, + 661 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{S}_i\\}_{i=1}^n" + }, + { + "bbox": [ + 313, + 530, + 556, + 661 + ], + "type": "text", + "content": ". In this framework, each feature map " + }, + { + "bbox": [ + 313, + 530, + 556, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_i" + }, + { + "bbox": [ + 313, + 530, + 556, + 661 + ], + "type": "text", + "content": " is initially passed through a separate MLP layer, standardizing the channel dimensions across scales to facilitate fusion. The transformed features are then concatenated and passed through a fusion MLP layer to aggregate a fused representation " + }, + { + "bbox": [ + 313, + 530, + 556, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 313, + 530, + 556, + 661 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 341, + 669, + 525, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 669, + 525, + 685 + ], + "spans": [ + { + "bbox": [ + 341, + 669, + 525, + 685 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {S}} _ {i} = \\operatorname {M L P} _ {\\operatorname {p r o j}} \\left(D _ {i}, D _ {t}\\right) \\left(\\mathbf {S} _ {i}\\right), \\forall i \\in \\{1, \\dots , n \\},", + "image_path": "9fcb04e83da54866de7b562d1319321c20ce04a4b3ac6099e1b031a3f097ff99.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 343, + 686, + 473, + 700 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 686, + 473, + 700 + ], + "spans": [ + { + "bbox": [ + 343, + 686, + 473, + 700 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {\\text {c a t}} = \\operatorname {C o n c a t} \\left(\\hat {\\mathbf {S}} _ {1}, \\hat {\\mathbf {S}} _ {2}, \\dots , \\hat {\\mathbf {S}} _ {n}\\right),", + "image_path": "c2134bdc4ff34ebd829df70d4d12dc07a474d1a624c9623ebe94444714d9c18f.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 343, + 702, + 473, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 702, + 473, + 715 + ], + "spans": [ + { + "bbox": [ + 343, + 702, + 473, + 715 + ], + "type": "interline_equation", + "content": "\\mathbf {F} = \\mathbf {M L P} _ {\\text {f u s i o n}} (n D _ {t}, D _ {t}) (\\mathbf {F} _ {\\text {c a t}}),", + "image_path": "baf13878e988c0566490cc91256fab22ada88ff7bba277637fbb37a54e43e70f.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11371" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 75, + 544, + 277 + ], + "blocks": [ + { + "bbox": [ + 58, + 75, + 544, + 277 + ], + "lines": [ + { + "bbox": [ + 58, + 75, + 544, + 277 + ], + "spans": [ + { + "bbox": [ + 58, + 75, + 544, + 277 + ], + "type": "image", + "image_path": "ed200455b8c4d7533c92c3aa23b01acb1e1a467f965197a79eb665254c14938f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 289, + 555, + 357 + ], + "lines": [ + { + "bbox": [ + 54, + 289, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 54, + 289, + 555, + 357 + ], + "type": "text", + "content": "Figure 2. Overview of our proposed framework. The FaceXFormer employs an encoder-decoder architecture, extracting multi-scale features from the input face image " + }, + { + "bbox": [ + 54, + 289, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 54, + 289, + 555, + 357 + ], + "type": "text", + "content": ", and fusing them into a unified representation " + }, + { + "bbox": [ + 54, + 289, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 54, + 289, + 555, + 357 + ], + "type": "text", + "content": " via MLP-Fusion. Task tokens " + }, + { + "bbox": [ + 54, + 289, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 54, + 289, + 555, + 357 + ], + "type": "text", + "content": " are processed alongside face representation " + }, + { + "bbox": [ + 54, + 289, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 54, + 289, + 555, + 357 + ], + "type": "text", + "content": " in the FaceX Decoder FXDec, resulting in refined task-specific tokens " + }, + { + "bbox": [ + 54, + 289, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{T}}" + }, + { + "bbox": [ + 54, + 289, + 555, + 357 + ], + "type": "text", + "content": ". These refined tokens are then used for task-specific predictions by passing through the unified head. FaceXFormer performs ten tasks, including face parsing, landmark detection, head pose estimation, attribute prediction, age, gender, and race estimation, facial expression recognition, face recognition, and face visibility prediction, achieving state-of-the-art performance at a real-time FPS of 33.21." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 376, + 296, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 376, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 54, + 376, + 296, + 449 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 54, + 376, + 296, + 449 + ], + "type": "inline_equation", + "content": "D_{i}" + }, + { + "bbox": [ + 54, + 376, + 296, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 376, + 296, + 449 + ], + "type": "inline_equation", + "content": "D_{t}" + }, + { + "bbox": [ + 54, + 376, + 296, + 449 + ], + "type": "text", + "content": " are the multi-scale feature channel dimensions of " + }, + { + "bbox": [ + 54, + 376, + 296, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_i" + }, + { + "bbox": [ + 54, + 376, + 296, + 449 + ], + "type": "text", + "content": " and the target channel dimension, respectively. The MLP-fusion design ensures minimal computational overhead (983k parameters) while maintaining the ability to perform efficient feature fusion, which is crucial for real-time application based face analysis tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 456, + 149, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 149, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 149, + 468 + ], + "type": "text", + "content": "3.2. FaceX Decoder" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 474, + 296, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 296, + 666 + ], + "type": "text", + "content": "Detection transformer (DETR) [10] employs object tokens to learn bounding box predictions for each object. Inspired by this approach, we introduce Task Tokens, whereby each task token is designed to learn specific facial tasks leveraging the fused face representation. However, existing decoders such as DETR [10] and Deformable-DETR [136] are computationally intensive, impacting runtime significantly. To address this, we propose FaceX (FXDec) a lightweight decoder designed to efficiently model the task tokens with face tokens. Specifically, each task token learns a task-related representation by interacting with other task tokens " + }, + { + "bbox": [ + 55, + 474, + 296, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 55, + 474, + 296, + 666 + ], + "type": "text", + "content": " and face tokens " + }, + { + "bbox": [ + 55, + 474, + 296, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 55, + 474, + 296, + 666 + ], + "type": "text", + "content": ", enhancing the overall representation. The Lightweight Decoder comprises of three main components: 1) Task Self-Attention, 2) Task-to-Face Cross-Attention, and 3) Face-to-Task Cross-Attention as illustrated in Figure 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Task Self-Attention (TSA): The Task Self-Attention module is designed to refine the task-specific representations within the set of task tokens " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{T} = \\langle T_1, \\ldots, T_n \\rangle" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": ". Each task" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 376, + 555, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 555, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 555, + 425 + ], + "type": "text", + "content": "token " + }, + { + "bbox": [ + 313, + 376, + 555, + 425 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 313, + 376, + 555, + 425 + ], + "type": "text", + "content": " is an embedded representation that corresponds to a specific facial task. In TSA, each " + }, + { + "bbox": [ + 313, + 376, + 555, + 425 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 313, + 376, + 555, + 425 + ], + "type": "text", + "content": " is updated by attending to all other task tokens to capture task-specific interactions. Formally, the updated task token " + }, + { + "bbox": [ + 313, + 376, + 555, + 425 + ], + "type": "inline_equation", + "content": "T_{i}^{\\prime}" + }, + { + "bbox": [ + 313, + 376, + 555, + 425 + ], + "type": "text", + "content": " is computed as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 347, + 430, + 520, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 430, + 520, + 444 + ], + "spans": [ + { + "bbox": [ + 347, + 430, + 520, + 444 + ], + "type": "interline_equation", + "content": "\\mathbf {T} _ {i} ^ {\\prime} = \\operatorname {S e l f A t t n} (\\mathbf {Q} = T _ {i}, \\mathbf {K} = \\mathbf {T}, \\mathbf {V} = \\mathbf {T}),", + "image_path": "d0826a378e61ba77cf20ce9bc58604615236495f2fe622c5f2113bf06459f201.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 449, + 555, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 498 + ], + "type": "text", + "content": "where Attention denotes the multi-headed self-attention mechanism, and " + }, + { + "bbox": [ + 313, + 449, + 555, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 313, + 449, + 555, + 498 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 449, + 555, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 313, + 449, + 555, + 498 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 449, + 555, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 313, + 449, + 555, + 498 + ], + "type": "text", + "content": " represent the queries, keys, and values, respectively. Therefore, TSA essentially helps the model to learn task-invariant representation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 509, + 556, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 556, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 556, + 605 + ], + "type": "text", + "content": "Task-to-Face Cross-Attention (TFCA): The Task-to-Face Cross-Attention module allows each task token to interact with the fused face representation " + }, + { + "bbox": [ + 313, + 509, + 556, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 313, + 509, + 556, + 605 + ], + "type": "text", + "content": ". This enables each task token to gather information relevant to its specific facial task from the fused face features. In this module, the fused face representation " + }, + { + "bbox": [ + 313, + 509, + 556, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 313, + 509, + 556, + 605 + ], + "type": "text", + "content": " acts as both key and value, while the task tokens serve as queries. The updated task token " + }, + { + "bbox": [ + 313, + 509, + 556, + 605 + ], + "type": "inline_equation", + "content": "\\hat{T}_i" + }, + { + "bbox": [ + 313, + 509, + 556, + 605 + ], + "type": "text", + "content": " is then computed as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 345, + 609, + 522, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 609, + 522, + 624 + ], + "spans": [ + { + "bbox": [ + 345, + 609, + 522, + 624 + ], + "type": "interline_equation", + "content": "\\hat {T} _ {i} = \\operatorname {C r o s s A t t n} (\\mathbf {Q} = T _ {i} ^ {\\prime}, \\mathbf {K} = \\mathbf {F}, \\mathbf {V} = \\mathbf {F}),", + "image_path": "8db24d21f58d5b7b2e26ec0a3c1c7dadcd47d0e02bb11deac51b7261cfd6e333.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 628, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 554, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 628, + 554, + 677 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{T}} = \\langle \\hat{T}_1,\\dots ,\\hat{T}_n\\rangle" + }, + { + "bbox": [ + 313, + 628, + 554, + 677 + ], + "type": "text", + "content": " is the output task token. Thus, TFCA enables direct interaction between the task-specific tokens and the compact facial features, facilitating task-focused feature extraction." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Face-to-Task Cross-Attention (FTCA): Conversely, the Face-to-Task Cross-Attention module is designed to refine" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "11372" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "the fused face representation " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": " based on the information from the updated task tokens. This process aids in enhancing the face representation with task-specific details, thereby improving the extraction of overall fused representation. In FTCA, the set of updated task tokens " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "\\mathbf{T}' = \\{\\mathbf{T}_1'',\\mathbf{T}_2'',\\dots,\\mathbf{T}_m''\\}" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": " acts as both keys and values, while the fused face features " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": " serve as queries. The refined face representation " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": " is computed as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 85, + 176, + 264, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 176, + 264, + 190 + ], + "spans": [ + { + "bbox": [ + 85, + 176, + 264, + 190 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {F}} = \\operatorname {C r o s s A t t n} (\\mathbf {Q} = \\mathbf {F}, \\mathbf {K} = \\mathbf {T} ^ {\\prime}, \\mathbf {V} = \\mathbf {T} ^ {\\prime}).", + "image_path": "3a3c84e78b8ecc44266ebf31d9165ff7d28e1f8b36765e5471a6f0de92763021.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 199, + 295, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 199, + 295, + 235 + ], + "spans": [ + { + "bbox": [ + 55, + 199, + 295, + 235 + ], + "type": "text", + "content": "Through this inverse attention mechanism, the face representation is augmented with critical task-specific details, enabling a robust approach towards facial task unification." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 242, + 141, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 242, + 141, + 254 + ], + "spans": [ + { + "bbox": [ + 55, + 242, + 141, + 254 + ], + "type": "text", + "content": "3.3. Unified-Head" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 259, + 296, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 259, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 259, + 296, + 486 + ], + "type": "text", + "content": "In Unified-Head, the task tokens are processed to obtain corresponding task predictions. As shown in Figure 2, the output face tokens " + }, + { + "bbox": [ + 55, + 259, + 296, + 486 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}" + }, + { + "bbox": [ + 55, + 259, + 296, + 486 + ], + "type": "text", + "content": " and task tokens " + }, + { + "bbox": [ + 55, + 259, + 296, + 486 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{T}}" + }, + { + "bbox": [ + 55, + 259, + 296, + 486 + ], + "type": "text", + "content": " are processed through a Task-to-Face Cross-Attention mechanism to obtain final refined features. Then, the output tokens are fed into their corresponding task heads. The task head for landmark detection is a hourglass network, for head pose estimation is a regression MLP, and for face recognition is PartialFC [4], while the tasks of age, gender and race estimation, facial expression recognition, face visibility prediction, and attributes prediction utilize classification MLPs. For face parsing, we leverage the output " + }, + { + "bbox": [ + 55, + 259, + 296, + 486 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}" + }, + { + "bbox": [ + 55, + 259, + 296, + 486 + ], + "type": "text", + "content": " and process it through an upsampling layer, then perform a cross-product with the face parsing token to obtain a segmentation map. The number of tokens for segmentation corresponds to the total number of classes. For landmark prediction, it corresponds to the number of landmarks (i.e., 68). For head pose estimation, the number of tokens is 9, representing the " + }, + { + "bbox": [ + 55, + 259, + 296, + 486 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 55, + 259, + 296, + 486 + ], + "type": "text", + "content": " rotation matrix. For other tasks, one token is used for each." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 495, + 174, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 174, + 507 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 174, + 507 + ], + "type": "text", + "content": "3.4. Multi-Task Training" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 512, + 295, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 512, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 55, + 512, + 295, + 690 + ], + "type": "text", + "content": "We aim to train FaceXFormer for multiple facial analysis tasks simultaneously, however each task requires distinct and sometimes conflicting pre-processing steps. For instance, landmark detection typically requires keypoint alignment of faces, which contradicts the needs for head pose estimation, as it may eliminate the natural variability of headposes. Due to these reasons, integrating all tasks into a single model poses significant challenges. To address this, FaceXFormer incorporates task-specific tokens designed to extract task-specific features from the fused representation. These task tokens compel the backbone to learn a unified representation capable of supporting a broad spectrum of facial analysis tasks. We employ different loss functions for each task and combine them in a joint objective for training. The final loss function is given as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 691, + 293, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 691, + 293, + 715 + ], + "spans": [ + { + "bbox": [ + 67, + 691, + 293, + 715 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L = \\lambda_ {s e g} L _ {s e g} + \\lambda_ {l n d} L _ {l n d} + \\lambda_ {h p e} L _ {h p e} + \\lambda_ {a t t r} L _ {a t t r} + \\lambda_ {a} L _ {a} \\\\ + \\lambda_ {g / r} L _ {g / r} + \\lambda_ {e x p} L _ {e x p} + \\lambda_ {f r} L _ {f r} + \\lambda_ {v i s} L _ {v i s} \\\\ \\end{array}", + "image_path": "3ab07bab3e4425c529bf5a375689cc964de14e02fedf0bdce68ebd647cc7f971.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "inline_equation", + "content": "L_{seg}" + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": " is the mean of dice loss [87] and Cross-Entropy (CE) loss for face parsing, " + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "inline_equation", + "content": "L_{lnd}" + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": " is STAR loss [132] for landmarks prediction, " + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "inline_equation", + "content": "L_{hpe}" + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": " is geodesic loss [121] for head pose estimation, " + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "inline_equation", + "content": "L_{g/r}" + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": " is CE loss for gender/race estimation, " + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "inline_equation", + "content": "L_a" + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": " is mean of L1 loss and CE loss for age estimation, " + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "inline_equation", + "content": "L_{exp}" + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": " is CE loss for facial expression recognition, " + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "inline_equation", + "content": "L_{fr}" + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": " is Arc-Face [15] loss for face recognition, and " + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "inline_equation", + "content": "L_{attr}" + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "inline_equation", + "content": "L_{vis}" + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": " are Binary Cross-Entropy with logits loss for attributes prediction and face visibility prediction respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 190, + 457, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 190, + 457, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 190, + 457, + 204 + ], + "type": "text", + "content": "4. Experiments and Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 210, + 555, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 210, + 555, + 294 + ], + "spans": [ + { + "bbox": [ + 313, + 210, + 555, + 294 + ], + "type": "text", + "content": "To validate the effectiveness of FaceXFormer, we conduct extensive experiments across ten diverse facial analysis tasks using a range of datasets, as described in Section 4.1. Implementation details are provided in Section 4.2 to ensure reproducibility and transparency. Finally, Sections 4.3 and 4.4 present the quantitative and qualitative results, respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 301, + 436, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 436, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 436, + 312 + ], + "type": "text", + "content": "4.1. Datasets and Metrics" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 318, + 553, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 553, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 553, + 402 + ], + "type": "text", + "content": "We perform co-training, where the model is simultaneously trained for multiple tasks using a total of 10 datasets with task-specific annotations. We conduct a comprehensive evaluation, comparing our approach with both task-specific and multi-task models. We present our results on the test sets according to the standard protocol for each task using the following datasets:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 402, + 553, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 553, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 553, + 485 + ], + "type": "text", + "content": "Train: Face Parsing: CelebAMaskHQ [42]; Landmarks Detection: 300W [79]; Head Pose Estimation: 300W-LP [135]; Attributes Prediction: CelebA [54]; Facial Expression Recognition: RAF-DB [45], AffectNet [62]; Age/Gender/Race estimation: UTKFace [125], FairFace [36]; Face Recognition: MS1MV3 [25]; Visibility Prediction: COFW [7]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 486, + 553, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 553, + 581 + ], + "type": "text", + "content": "Test: Face Parsing: CelebAMaskHQ [42]; Landmarks Detection: 300W [135], 300VW [83]; Head Pose Estimation: BIWI [20]; Attributes Prediction: CelebA [54], LFWA [107]; Facial Expression Recognition: RAFDB [45]; Age/Gender/Race Estimation: UTKFace [125], FairFace [36]; Face Recognition: LFW [31], CFP-FP [81], AgeDB [63], CALFW [129], CPLFW [128]; Visibility Prediction: COFW [7]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 582, + 553, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 553, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 553, + 665 + ], + "type": "text", + "content": "The evaluation metrics used are the F1-score for face parsing, Normalized Mean Error (NME) for landmark prediction, Mean Absolute Error (MAE) for head pose estimation and age estimation, accuracy for facial expression recognition, attributes prediction, gender estimation, race estimation, 1:1 verification accuracy for face recognition, and recall at " + }, + { + "bbox": [ + 313, + 582, + 553, + 665 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 313, + 582, + 553, + 665 + ], + "type": "text", + "content": " precision for face visibility prediction." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 672, + 447, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 447, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 447, + 685 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "We train our models using a distributed PyTorch setup on eight A6000 GPUs, each equipped with 48GB of memory." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11373" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 555, + 211 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 555, + 211 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 555, + 211 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 555, + 211 + ], + "type": "table", + "html": "
MethodInput Res.SkinHairNoseL-EyeR-EyeL-BrowR-BrowL-LipI-MouthU-LipMean F1 ↑
Wei et al. [105]51296.491.191.987.185.080.882.591.090.687.988.43
EHANet [55]51296.093.993.786.286.583.283.190.393.888.689.53
EAGRNet [91]47396.294.994.088.689.085.785.291.295.088.990.87
AGRNet [92]47396.587.693.988.789.185.585.691.192.089.189.91
FaRL_scratch [130]51296.294.993.889.089.085.385.490.091.788.190.34
DML-CSR [127]47395.794.593.989.489.685.585.791.091.889.190.62
FP-LIIF [80]25696.495.193.788.588.584.584.390.392.187.590.09
SwinFace [71]××××××××××××
QFace [88]××××××××××××
Facetor [72]51296.696.293.989.489.186.286.390.691.689.090.89
FaceXFormer22496.495.793.890.190.386.085.990.692.189.291.01
", + "image_path": "ccbc3c101323901de6489867423bb2d6175b2de6659246449d8a1ae12bd3d8a5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 256, + 293, + 367 + ], + "blocks": [ + { + "bbox": [ + 55, + 220, + 555, + 243 + ], + "lines": [ + { + "bbox": [ + 55, + 220, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 555, + 243 + ], + "type": "text", + "content": "Table 3. Performance comparison for face parsing on the CelebAMask-HQ dataset [42]. The symbol " + }, + { + "bbox": [ + 55, + 220, + 555, + 243 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 55, + 220, + 555, + 243 + ], + "type": "text", + "content": " indicates that the model does not perform the corresponding task. Red = First Best, Blue = Second Best. " + }, + { + "bbox": [ + 55, + 220, + 555, + 243 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 55, + 220, + 555, + 243 + ], + "type": "text", + "content": " indicates a model that doesn't perform the task." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 256, + 293, + 367 + ], + "lines": [ + { + "bbox": [ + 57, + 256, + 293, + 367 + ], + "spans": [ + { + "bbox": [ + 57, + 256, + 293, + 367 + ], + "type": "table", + "html": "
MethodsExpression (RAF-DB) ↑MethodsVisibility (COFW) ↑MethodsAge (MAE) UTKFace ↓
DLP-CNN [45]80.89RCPR [7]40OR-CNN [68]5.74
gACNN [48]85.07Wu et al. [111]44.43Axel Berg et al. [5]4.55
IPA2LT [120]86.77Wu et al. [110]49.11CORAL [8]5.47
RAN [101]86.90ECT [122]63.4Gustafsson et al. [26]4.65
CovPool [2]87.003DDE [94]63.89R50-SORD [70]4.36
SCN [100]87.03MNN [95]72.12VOLO-D1 [40]4.23
DACL [21]87.78DLDL-v2 [23]4.42
KTN [44]88.07MWR [84]4.37
DMUE [82]88.76
SwinFace [71]86.54SwinFace [71]×SwinFace [71]×
QFace [88]92.86QFace [88]×QFace [88]×
Facetor [72]87.58Facetor [72]×Facetor [72]4.10
FaceXFormer88.24FaceXFormer72.56FaceXFormer4.17
", + "image_path": "445a9223ee193bc31351a5d5e93076b8f4075243bbe7971c6457e573a7aed881.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 376, + 295, + 398 + ], + "lines": [ + { + "bbox": [ + 55, + 376, + 295, + 398 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 295, + 398 + ], + "type": "text", + "content": "Table 4. Performance comparison on facial expression recognition, face visibility prediction and age estimation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 413, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 413, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 54, + 413, + 295, + 605 + ], + "type": "text", + "content": "The models' backbones are initialized with ImageNet pretrained weights and processes input images at a resolution of " + }, + { + "bbox": [ + 54, + 413, + 295, + 605 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 54, + 413, + 295, + 605 + ], + "type": "text", + "content": ". We employ the AdamW optimizer with a weight decay of " + }, + { + "bbox": [ + 54, + 413, + 295, + 605 + ], + "type": "inline_equation", + "content": "1e^{-5}" + }, + { + "bbox": [ + 54, + 413, + 295, + 605 + ], + "type": "text", + "content": ". All models are trained for 12 epochs with a batch size of 48 on each GPU, and an initial learning rate of " + }, + { + "bbox": [ + 54, + 413, + 295, + 605 + ], + "type": "inline_equation", + "content": "1e^{-4}" + }, + { + "bbox": [ + 54, + 413, + 295, + 605 + ], + "type": "text", + "content": ", which decays by a factor of 10 at the " + }, + { + "bbox": [ + 54, + 413, + 295, + 605 + ], + "type": "inline_equation", + "content": "6^{th}" + }, + { + "bbox": [ + 54, + 413, + 295, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 413, + 295, + 605 + ], + "type": "inline_equation", + "content": "10^{th}" + }, + { + "bbox": [ + 54, + 413, + 295, + 605 + ], + "type": "text", + "content": " epochs. For some tasks, we train the model for additional epochs to reach the upper ceiling of performance. For data augmentation, we randomly apply Gaussian blur, grayscale conversion, gamma correction, occlusion, horizontal flipping, and affine transformations, such as rotation, translation and scaling. The number of FaceX decoder " + }, + { + "bbox": [ + 54, + 413, + 295, + 605 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 413, + 295, + 605 + ], + "type": "text", + "content": " is set to two. To ensure stable training across tasks when using multiple datasets of varying sample sizes, we equalize the representation of each task's samples in every batch through upsampling." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 612, + 137, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 137, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 137, + 624 + ], + "type": "text", + "content": "4.3. Main results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": "In Table 3, Table 4, Table 6, Table 5, we present a comparative analysis of FaceXFormer against recent methods across a variety of tasks. A key highlight of our work is its unique capability to deliver promising results across multiple tasks at real-time inference speed using a single unified model. Specifically, FaceXFormer achieves state-of-the-art performance in face parsing, with a mean F1 score of 91.01 on" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 258, + 555, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 555, + 498 + ], + "type": "text", + "content": "CelebAMaskHQ at a resolution of " + }, + { + "bbox": [ + 313, + 258, + 555, + 498 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 313, + 258, + 555, + 498 + ], + "type": "text", + "content": ", which is half the input size required by other SOTA methods. Furthermore, it demonstrates superior performance in head pose estimation and landmark detection, achieving a mean MAE of 3.52 and a mean NME of 4.67, respectively. Additionally, FaceXFormer provides a significant performance boost in attributes prediction and visibility prediction, achieving an accuracy of " + }, + { + "bbox": [ + 313, + 258, + 555, + 498 + ], + "type": "inline_equation", + "content": "91.83\\%" + }, + { + "bbox": [ + 313, + 258, + 555, + 498 + ], + "type": "text", + "content": " on the CelebA dataset and " + }, + { + "bbox": [ + 313, + 258, + 555, + 498 + ], + "type": "inline_equation", + "content": "72.56\\%" + }, + { + "bbox": [ + 313, + 258, + 555, + 498 + ], + "type": "text", + "content": " on COFW. It also performs competitively in age estimation, achieving the second-best score of 4.17, and achieves an accuracy of " + }, + { + "bbox": [ + 313, + 258, + 555, + 498 + ], + "type": "inline_equation", + "content": "88.24\\%" + }, + { + "bbox": [ + 313, + 258, + 555, + 498 + ], + "type": "text", + "content": " in facial expression recognition. In face recognition, FaceXFormer outperforms Faceptor, achieving a mean accuracy of " + }, + { + "bbox": [ + 313, + 258, + 555, + 498 + ], + "type": "inline_equation", + "content": "95.94\\%" + }, + { + "bbox": [ + 313, + 258, + 555, + 498 + ], + "type": "text", + "content": " compared to " + }, + { + "bbox": [ + 313, + 258, + 555, + 498 + ], + "type": "inline_equation", + "content": "95.28\\%" + }, + { + "bbox": [ + 313, + 258, + 555, + 498 + ], + "type": "text", + "content": ". However, we observe that multi-task models generally underperform compared to specialized ones in this task. This can be attributed to conflicting training objectives, which force the model to learn identity-invariant features rather than identity-specific representations crucial for accurate recognition. The results on gender estimation across different race categories is shown in Table 8." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "type": "text", + "content": "Recent models such as SwinFace [71], QFace [88], and Facetor [72] also aim to unify multiple tasks but only address a subset of them. These approaches often exclude complex tasks such as segmentation, head pose estimation, and landmark prediction. Moreover, they rely on multiple decoders and computationally expensive attention mechanisms, adding to the overall computational overhead. In contrast, FaceXFormer seamlessly unifies these complex tasks using a lightweight decoder and achieves state-of-the-art performance across them at a real-time FPS of 33.21. It outperforms previous multi-task models in segmentation, head pose estimation, landmark prediction, attribute prediction, and face visibility prediction, while achieving the second-best performance in age estimation. In this work, we simultaneously train for ten heterogeneous tasks, posing a more formidable challenge than previous approaches. Despite this, FaceXFormer effectively handles multiple tasks, achieving SOTA or competitive performance in real time." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11374" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 70, + 547, + 258 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 547, + 258 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 547, + 258 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 547, + 258 + ], + "type": "table", + "html": "
MethodsHeadpose (BIWI)MethodsLandmarks (300W)MethodsCelebA Acc. ↑
YawPitchRollMAE ↓Full ↓Com ↓Chal ↓
HopeNet [78]4.816.613.274.89LAB [109]3.492.985.19PANDA-1 [123]85.43
QuatNet [30]5.494.012.944.15Wing [22]4.043.277.18LNets+ANet [53]87.33
FSA-Net [116]4.275.492.934.14DeCaFa [14]3.392.935.26SSP+SSG [34]88.24
EVA-GCN [114]6.014.782.983.98HRNet [98]3.322.875.15MOON [77]90.94
TriNet [9]4.114.753.043.97PicassoNet [106]3.583.035.81NSA [56]90.61
img2pose [3]4.563.543.243.78AVS+SAN [18]3.863.216.46MCNN-AUX [28]91.29
MNN [95]3.984.612.393.66LUVLi [39]3.232.765.16MCFA [137]91.23
MFDNet [51]3.404.682.773.62HIH [41]3.092.654.89DMM-CNN [57]91.70
TokenHPE [121]3.954.512.713.72PIPNet [33]3.192.784.89SSPL [85]91.77
WHENet [131]3.994.393.063.81SLPT [112]3.172.754.90FaRL [130]91.39
SwinFace [71]××××SwinFace [71]×××SwinFace [71]91.38
QFace [88]----QFace [88]×××QFace [88]91.56
Facector [72]××××Facector [72]3.162.754.84Facector [72]91.39
FaceXFormer3.913.972.673.52FaceXFormer3.052.664.67FaceXFormer91.83
", + "image_path": "28239e7818336f231c7111c68a2618ff4d7112993283064b55c377bb2fc7d4c8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 303, + 294, + 387 + ], + "blocks": [ + { + "bbox": [ + 55, + 266, + 553, + 289 + ], + "lines": [ + { + "bbox": [ + 55, + 266, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 553, + 289 + ], + "type": "text", + "content": "Table 5. Performance comparison on headpose, landmark detection, and attribute recognition. The symbol " + }, + { + "bbox": [ + 55, + 266, + 553, + 289 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 55, + 266, + 553, + 289 + ], + "type": "text", + "content": " indicates that the model does not perform the corresponding task, while - denotes that results for this dataset are not provided. Red = First Best, Blue = Second Best." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 303, + 294, + 387 + ], + "lines": [ + { + "bbox": [ + 57, + 303, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 57, + 303, + 294, + 387 + ], + "type": "table", + "html": "
MethodLFWCFP-FPAgeDBCALFWCPLFWMean ↑
CosFace [97]99.8198.1298.1195.7692.2896.81
ArcFace [15]99.8398.2798.2895.4592.0896.78
VPL-ArcFace [16]99.8399.1198.6096.1293.4597.42
AdaFace [37]99.8399.1198.1796.0293.9397.41
SwinFace [71]99.8798.6098.1596.1093.4297.22
QFace [88]××××××
Facetor [72]99.4096.3493.6594.7592.2795.28
FaceXFormer99.6896.7596.3595.5091.4195.94
", + "image_path": "cd77275a25ef12d44f7b79a7e69ab57a087cc6190b571b273e63f4adcf785f88.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 418, + 295, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 295, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 295, + 455 + ], + "type": "text", + "content": "This success can be attributed to the efficiency of the proposed lightweight decoder, which employs a novel bidirectional cross-attention mechanism." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 461, + 230, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 230, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 230, + 475 + ], + "type": "text", + "content": "4.4. Qualitative \"in-the-wild\" results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 479, + 295, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 479, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 54, + 479, + 295, + 647 + ], + "type": "text", + "content": "In this section, we present the qualitative results of FaceXFormer on randomly selected \"in-the-wild\" images. We select five random images and showcase the results for face parsing, head pose estimation, landmarks prediction, age estimation, gender and race classification, and attributes prediction in Figure 3. Notably, the model successfully performs complex tasks such as face segmentation, head pose estimation, and landmark prediction, even when input samples exhibit extreme poses, occlusions, or blurring. Furthermore, FaceXFormer can be effectively used as a tool to generate multiple annotations for each image, making it valuable for various downstream tasks. These results highlight FaceXFormer's robust performance in challenging, real-world scenarios." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 657, + 149, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 149, + 671 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 149, + 671 + ], + "type": "text", + "content": "5. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "In this section, we explore the impact of different components of FaceXFormer on performance. Additionally, we demonstrate that the proposed model exhibits minimal bias" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 306, + 555, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 555, + 354 + ], + "type": "text", + "content": "compared to other models by evaluating age and gender prediction across various demographics. Furthermore, we analyze the impact of using different backbones of varying sizes on the FPS and performance." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 315, + 362, + 553, + 430 + ], + "blocks": [ + { + "bbox": [ + 75, + 395, + 275, + 406 + ], + "lines": [ + { + "bbox": [ + 75, + 395, + 275, + 406 + ], + "spans": [ + { + "bbox": [ + 75, + 395, + 275, + 406 + ], + "type": "text", + "content": "Table 6. Performance comparison for face recognition." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 362, + 553, + 430 + ], + "lines": [ + { + "bbox": [ + 315, + 362, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 315, + 362, + 553, + 430 + ], + "type": "table", + "html": "
MLP FusionSelf Attn.Cross Attn.Bi-dir. CAHPE (MAE)Lnd (NME)Attr. (Acc.)Age (MAE)
✓✓3.704.7091.214.21
✓✓17.1631.4979.9016.15
✓✓✓4.635.3089.984.84
✓✓✓3.524.6791.834.17
", + "image_path": "36f4356896acb2034ecfc224541eadea0eedecb6f3f92554eb1705ef73b9bee1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 331, + 435, + 537, + 447 + ], + "lines": [ + { + "bbox": [ + 331, + 435, + 537, + 447 + ], + "spans": [ + { + "bbox": [ + 331, + 435, + 537, + 447 + ], + "type": "text", + "content": "Table 7. Impact of various components on performance." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 456, + 555, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 555, + 470 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 555, + 470 + ], + "type": "text", + "content": "5.1. Impact of various components in FaceXFormer" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "type": "text", + "content": "To evaluate the contribution of each component in FaceXFormer, we conduct an ablation study focusing on the importance of specific design choices and their impact on performance across various tasks. The results of these experiments are summarized in Table 7. We observe that without MLP fusion (row 1), there is a drop in performance, highlighting the importance of integrating multi-scale features to capture both global and local information essential for accurate predictions. The model performs extremely poorly (row 2) without cross-attention in the decoder, which is expected, as there is no interaction between face tokens and task tokens in this case. Introducing the proposed bidirectional cross-attention (row 4), which corresponds to FaceXFormer, in the decoder provides a significant boost compared to using standard cross-attention (row 3), yielding improvements of 1.11 MAE in head pose estimation, 0.63 NME in landmark detection, 1.85 accuracy points in attribute prediction, and 0.67 MAE in age estimation. These results demonstrate the importance of MLP fusion and bidirectional cross-attention in the FaceXFormer architecture." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11375" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 294, + 335 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 294, + 335 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 294, + 335 + ], + "type": "image", + "image_path": "b851fa0fb1c9084a6105526b5b8f8927c62ace62808298c4e17514eb5a478a32.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 342, + 285, + 354 + ], + "lines": [ + { + "bbox": [ + 66, + 342, + 285, + 354 + ], + "spans": [ + { + "bbox": [ + 66, + 342, + 285, + 354 + ], + "type": "text", + "content": "Figure 3. FaceXFormer predictions on \"in-the-wild\" images" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 365, + 295, + 453 + ], + "blocks": [ + { + "bbox": [ + 58, + 365, + 295, + 453 + ], + "lines": [ + { + "bbox": [ + 58, + 365, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 58, + 365, + 295, + 453 + ], + "type": "table", + "html": "
ModelData PointsWhiteNon-whiteAverageDiscrepancy
AgeFairFace100K60.0560.6360.520.58
CLIP400M62.2561.9562.00-0.30
FaRL20M61.4961.8461.780.35
FaceXFormer5.9M58.9459.4459.340.50
GenderFairFace100K94.1594.4194.360.26
CLIP400M94.8795.7895.610.91
FaRL20M95.1695.7795.650.61
FaceXFormer5.9M95.3495.1995.22-0.09
", + "image_path": "aaccb12e6da01b590f355c10cad07152429759d6ef2327edb8cce80f3e1cc50d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 489, + 271, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 489, + 271, + 501 + ], + "spans": [ + { + "bbox": [ + 55, + 489, + 271, + 501 + ], + "type": "text", + "content": "5.2. Bias Analysis and Ethical Considerations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 510, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 510, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 510, + 295, + 713 + ], + "type": "text", + "content": "In our work, we utilize 17 unique datasets for training and evaluation. We obtained these datasets following the procedures stated on their respective pages and signed the license agreements if and when necessary. As we train our models on multiple datasets designed for different tasks, the subjects across different age groups, genders, and races is not equal. This imbalance may introduce bias in the model. Therefore, we provide an analysis using the FairFace [35] dataset, which is balanced in terms of age, gender and race. We follow [73] and define the \"Non-white\" group to include multiple racial categories: \"Black\", \"Indian\", \"East Asian\", \"Southeast Asian\", \"Middle Eastern\" and \"Latino\". As can be seen from Table 8, FaceXFormer shows the smallest performance discrepancy across different racial groups and exhibits minimal bias compared to other models despite being trained on fewer data points. This can be attributed to race estimation being the task in co-training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 524, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 524, + 83 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 524, + 83 + ], + "type": "text", + "content": "5.3. Performance under different backbones" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 91, + 553, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 553, + 186 + ], + "type": "text", + "content": "To evaluate the impact of different backbones on performance and FPS, we conduct an ablation study comparing various backbone architectures in FaceXFormer. We categorize head pose estimation, landmark prediction, and age estimation as regression (Reg) tasks, while attribute prediction and facial expression recognition fall under classification (Cls). Additionally, face parsing is denoted as segmentation (Seg). The results are summarized in Table 9." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 326, + 199, + 543, + 267 + ], + "blocks": [ + { + "bbox": [ + 59, + 458, + 291, + 469 + ], + "lines": [ + { + "bbox": [ + 59, + 458, + 291, + 469 + ], + "spans": [ + { + "bbox": [ + 59, + 458, + 291, + 469 + ], + "type": "text", + "content": "Table 8. Age and gender accuracy w.r.t race groups on FairFace" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 199, + 543, + 267 + ], + "lines": [ + { + "bbox": [ + 326, + 199, + 543, + 267 + ], + "spans": [ + { + "bbox": [ + 326, + 199, + 543, + 267 + ], + "type": "table", + "html": "
BackboneSegRegClsFPSParams
MobileNet90.214.6488.2239.7625.32
ResNet10190.494.3788.9134.9865.54
ConvNext-B91.084.3589.0936.61110.19
Swin-B91.014.1290.0333.21109.29
", + "image_path": "3d5cbb5be63d082597a5d40787c839c76ff5066d40cd22ad4beea733084b03a0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 318, + 275, + 550, + 286 + ], + "lines": [ + { + "bbox": [ + 318, + 275, + 550, + 286 + ], + "spans": [ + { + "bbox": [ + 318, + 275, + 550, + 286 + ], + "type": "text", + "content": "Table 9. Effect of different backbones on performance and FPS." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 304, + 555, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 555, + 460 + ], + "type": "text", + "content": "From the results, we observe that ConvNeXt achieves the best performance in segmentation with an F1 score of " + }, + { + "bbox": [ + 313, + 304, + 555, + 460 + ], + "type": "inline_equation", + "content": "91.08\\%" + }, + { + "bbox": [ + 313, + 304, + 555, + 460 + ], + "type": "text", + "content": ". The Swin Transformer backbone excels in both regression and classification tasks, with a mean error of 4.12 and a mean accuracy of " + }, + { + "bbox": [ + 313, + 304, + 555, + 460 + ], + "type": "inline_equation", + "content": "90.03\\%" + }, + { + "bbox": [ + 313, + 304, + 555, + 460 + ], + "type": "text", + "content": ", respectively. In contrast, MobileNet demonstrates the lowest performance metrics, including an F1 score of " + }, + { + "bbox": [ + 313, + 304, + 555, + 460 + ], + "type": "inline_equation", + "content": "90.21\\%" + }, + { + "bbox": [ + 313, + 304, + 555, + 460 + ], + "type": "text", + "content": " and a mean error of 4.64, highlighting its limitations in handling larger, more complex datasets due to its smaller receptive field compared to the Swin Transformer. The selection of the Swin Transformer as the backbone for FaceXFormer is driven by its superior scalability and global contextual understanding, both of which are essential for facial analysis tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 476, + 388, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 476, + 388, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 476, + 388, + 488 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": "FaceXFormer introduces a novel end-to-end unified model that efficiently handles a wide range of facial analysis tasks in real-time. By adopting a transformer-based encoder-decoder architecture and representing each task as a learnable token, our approach seamlessly integrates multiple tasks within a single framework while maintaining minimal computational cost and fast inference times. The proposed lightweight decoder, FaceX, incorporates a novel bi-directional cross-attention mechanism, enhancing the model's ability to learn robust and generalized face representations across diverse tasks. Comprehensive experiments demonstrate that FaceXFormer achieves state-of-the-art performance across multiple facial analysis tasks, achieving a real-time FPS of 33.21. In broader applications, FaceXFormer can serve as an annotator for large-scale face datasets and can be integrated into existing facial analysis systems to provide extra information, making it a valuable tool for surveillance, subject analysis, and image retrieval." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11376" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 202 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 202 + ], + "type": "text", + "content": "This research is based upon work supported in part by the Office of the Director of National Intelligence (ODNI), Intelligence Advanced Research Projects Activity (IARPA), via [2022-21102100005]. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies, either expressed or implied, of ODNI, IARPA, or the U.S. Government. The US Government is authorized to reproduce and distribute reprints for governmental purposes notwithstanding any copyright annotation therein." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 223, + 115, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 223, + 115, + 236 + ], + "spans": [ + { + "bbox": [ + 56, + 223, + 115, + 236 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 243, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 66, + 243, + 295, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 243, + 295, + 275 + ], + "spans": [ + { + "bbox": [ + 66, + 243, + 295, + 275 + ], + "type": "text", + "content": "[1] B Abirami, TS Subashini, and V Mahavaishnavi. Gender and age prediction from real time facial images using cnn. Materials Today: Proceedings, 33:4708-4712, 2020. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 277, + 295, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 277, + 295, + 331 + ], + "spans": [ + { + "bbox": [ + 66, + 277, + 295, + 331 + ], + "type": "text", + "content": "[2] Dinesh Acharya, Zhiwu Huang, Danda Pani Paudel, and Luc Van Gool. Covariance pooling for facial expression recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 367-374, 2018. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 334, + 295, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 334, + 295, + 388 + ], + "spans": [ + { + "bbox": [ + 66, + 334, + 295, + 388 + ], + "type": "text", + "content": "[3] Vitor Albiero, Xingyu Chen, Xi Yin, Guan Pang, and Tal Hassner. img2pose: Face alignment and detection via 6dof, face pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7617-7627, 2021. 2, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 389, + 295, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 389, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 66, + 389, + 295, + 453 + ], + "type": "text", + "content": "[4] Xiang An, Xuhan Zhu, Yuan Gao, Yang Xiao, Yongle Zhao, Ziyong Feng, Lan Wu, Bin Qin, Ming Zhang, Debing Zhang, et al. Partial fc: Training 10 million identities on a single machine. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1445-1449, 2021. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 456, + 294, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 456, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 66, + 456, + 294, + 499 + ], + "type": "text", + "content": "[5] Axel Berg, Magnus Oskarsson, and Mark O'Connor. Deep ordinal regression with label diversity. In 2020 25th international conference on pattern recognition (ICPR), pages 2740-2747. IEEE, 2021. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 502, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 502, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 66, + 502, + 294, + 566 + ], + "type": "text", + "content": "[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 568, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 568, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 66, + 568, + 295, + 612 + ], + "type": "text", + "content": "[7] Xavier P Burgos-Artizzu, Pietro Perona, and Piotr Dólar. Robust face landmark estimation under occlusion. In Proceedings of the IEEE international conference on computer vision, pages 1513-1520, 2013. 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 66, + 614, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 614, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 66, + 614, + 295, + 656 + ], + "type": "text", + "content": "[8] Wenzhi Cao, Vahid Mirjalili, and Sebastian Raschka. Rank consistent ordinal regression for neural networks with application to age estimation. Pattern Recognition Letters, 140:325-331, 2020. 1, 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 66, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 66, + 658, + 295, + 712 + ], + "type": "text", + "content": "[9] Zhiwen Cao, Zongcheng Chu, Dongfang Liu, and Yingjie Chen. A vector-based representation to enhance head pose estimation. In Proceedings of the IEEE/CVF Winter Conference on applications of computer vision, pages 1188-1197, 2021. 2, 7" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 320, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 553, + 127 + ], + "type": "text", + "content": "[10] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 320, + 129, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 129, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 320, + 129, + 553, + 183 + ], + "type": "text", + "content": "[11] Liang-Chieh Chen, Yi Yang, Jiang Wang, Wei Xu, and Alan L Yuille. Attention to scale: Scale-aware semantic image segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3640-3649, 2016. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 184, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 184, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 320, + 184, + 553, + 227 + ], + "type": "text", + "content": "[12] Alejandro Cobo, Roberto Valle, José M Buenapasada, and Luis Baumela. On the representation and methodology for wide and short range head pose estimation. Pattern Recognition, 149:110263, 2024. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 228, + 553, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 228, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 320, + 228, + 553, + 281 + ], + "type": "text", + "content": "[13] Kaiwen Cui, Rongliang Wu, Fangneng Zhan, and Shijian Lu. Face transformer: Towards high fidelity and accurate face swapping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 668-677, 2023. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 283, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 283, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 320, + 283, + 553, + 327 + ], + "type": "text", + "content": "[14] Arnaud Dapogny, Kevin Bailly, and Matthieu Cord. Decafa: Deep convolutional cascade for face alignment in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6893-6901, 2019. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 328, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 328, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 320, + 328, + 553, + 381 + ], + "type": "text", + "content": "[15] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 2, 5, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 383, + 553, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 383, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 320, + 383, + 553, + 436 + ], + "type": "text", + "content": "[16] Jiankang Deng, Jia Guo, Jing Yang, Alexandros Lattas, and Stefanos Zafeiriou. Variational prototype learning for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11906-11915, 2021. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 437, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 437, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 320, + 437, + 553, + 491 + ], + "type": "text", + "content": "[17] Xing Di, Yiyu Zheng, Xiaoming Liu, and Yu Cheng. Pros: Facial omni-representation learning via prototype-based self-distillation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6087-6098, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 493, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 493, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 320, + 493, + 553, + 536 + ], + "type": "text", + "content": "[18] Xuanyi Dong, Yan Yan, Wanli Ouyang, and Yi Yang. Style aggregated network for facial landmark detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 379-388, 2018. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 537, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 537, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 320, + 537, + 553, + 602 + ], + "type": "text", + "content": "[19] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 604, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 604, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 320, + 604, + 553, + 645 + ], + "type": "text", + "content": "[20] Gabriele Fanelli, Matthias Dantone, Juergen Gall, Andrea Fossati, and Luc Van Gool. Random forests for real time 3d face analysis. International journal of computer vision, 101:437-458, 2013. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 647, + 553, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 647, + 553, + 691 + ], + "spans": [ + { + "bbox": [ + 320, + 647, + 553, + 691 + ], + "type": "text", + "content": "[21] Amir Hossein Farzaneh and Xiaojun Qi. Facial expression recognition in the wild via deep attentive center loss. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2402-2411, 2021. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 692, + 553, + 713 + ], + "type": "text", + "content": "[22] Zhen-Hua Feng, Josef Kittler, Muhammad Awais, Patrik Huber, and Xiao-Jun Wu. Wing loss for robust facial land" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11377" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 81, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 81, + 73, + 294, + 106 + ], + "type": "text", + "content": "mark localisation with convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 106, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 106, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 61, + 106, + 295, + 149 + ], + "type": "text", + "content": "[23] Bin-Bin Gao, Xin-Xin Liu, Hong-Yu Zhou, Jianxin Wu, and Xin Geng. Learning expectation of label distribution for facial age and attractiveness estimation. arXiv preprint arXiv:2007.01771, 2020. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 150, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 150, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 61, + 150, + 294, + 205 + ], + "type": "text", + "content": "[24] Asma El Kissi Ghalleb, Safa Boumaiza, and Najoua Essoukri Ben Amara. Demographic face profiling based on age, gender and race. In 2020 5th International Conference on Advanced Technologies for Signal and Image Processing (ATSIP), pages 1-6. IEEE, 2020. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 206, + 295, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 206, + 295, + 270 + ], + "spans": [ + { + "bbox": [ + 62, + 206, + 295, + 270 + ], + "type": "text", + "content": "[25] Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. Ms-celeb-1m: A dataset and benchmark for large-scale face recognition. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 87-102. Springer, 2016. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 272, + 295, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 272, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 62, + 272, + 295, + 335 + ], + "type": "text", + "content": "[26] Fredrik K Gustafsson, Martin Danelljan, Goutam Bhat, and Thomas B Schon. Energy-based models for deep probabilistic regression. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16, pages 325-343. Springer, 2020. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 338, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 338, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 62, + 338, + 294, + 392 + ], + "type": "text", + "content": "[27] Hu Han, Anil K Jain, Fang Wang, Shiguang Shan, and Xilin Chen. Heterogeneous face attribute estimation: A deep multi-task learning approach. IEEE transactions on pattern analysis and machine intelligence, 40(11):2597-2609, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 393, + 294, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 393, + 294, + 446 + ], + "spans": [ + { + "bbox": [ + 62, + 393, + 294, + 446 + ], + "type": "text", + "content": "[28] Emily Hand and Rama Chellappa. Attributes for improved attributes: A multi-task network utilizing implicit and explicit relationships for facial attribute classification. In Proceedings of the AAAI conference on artificial intelligence, 2017. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 449, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 449, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 62, + 449, + 294, + 502 + ], + "type": "text", + "content": "[29] Hui-Lan Hsieh, Winston Hsu, and Yan-Ying Chen. Multi-task learning for face identification and attribute estimation. In 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2981-2985, 2017. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 504, + 294, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 504, + 294, + 547 + ], + "spans": [ + { + "bbox": [ + 62, + 504, + 294, + 547 + ], + "type": "text", + "content": "[30] Heng-Wei Hsu, Tung-Yu Wu, Sheng Wan, Wing Hung Wong, and Chen-Yi Lee. Quatnet: Quaternion-based head pose estimation with multiregression loss. IEEE Transactions on Multimedia, 21(4):1035-1046, 2018. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 548, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 548, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 62, + 548, + 294, + 602 + ], + "type": "text", + "content": "[31] Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. In Workshop on faces in 'Real-Life' Images: detection, alignment, and recognition, 2008. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 604, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 604, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 62, + 604, + 294, + 668 + ], + "type": "text", + "content": "[32] Aaron S Jackson, Michel Valstar, and Georgios Tzimiropoulos. A cnn cascade for landmark guided semantic part segmentation. In Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part III 14, pages 143-155. Springer, 2016. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "type": "text", + "content": "[33] Haibo Jin, Shengcai Liao, and Ling Shao. Pixel-in-pixel net: Towards efficient facial landmark detection in the wild. International Journal of Computer Vision, 129(12):3174-3194, 2021. 7" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 321, + 73, + 553, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 73, + 553, + 126 + ], + "spans": [ + { + "bbox": [ + 321, + 73, + 553, + 126 + ], + "type": "text", + "content": "[34] Mahdi M Kalayeh, Boqing Gong, and Mubarak Shah. Improving facial attribute prediction using semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6942-6950, 2017. 2, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 320, + 129, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 129, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 320, + 129, + 553, + 184 + ], + "type": "text", + "content": "[35] Kimmo Karkkainen and Jungseock Joo. Fairface: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1548-1558, 2021. 2, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 186, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 186, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 321, + 186, + 553, + 239 + ], + "type": "text", + "content": "[36] Kimmo Karkkainen and Jungseock Joo. Fairface: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1548-1558, 2021. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 241, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 241, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 321, + 241, + 553, + 285 + ], + "type": "text", + "content": "[37] Minchul Kim, Anil K Jain, and Xiaoming Liu. Adaface: Quality adaptive margin for face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18750-18759, 2022. 1, 2, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 286, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 286, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 321, + 286, + 553, + 330 + ], + "type": "text", + "content": "[38] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 332, + 553, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 332, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 321, + 332, + 553, + 407 + ], + "type": "text", + "content": "[39] Abhinav Kumar, Tim K Marks, Wenxuan Mou, Ye Wang, Michael Jones, Anoop Cherian, Toshiaki Koike-Akino, Xiaoming Liu, and Chen Feng. Luvli face alignment: Estimating landmarks' location, uncertainty, and visibility likelihood. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8236-8246, 2020. 1, 2, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 410, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 410, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 321, + 410, + 553, + 442 + ], + "type": "text", + "content": "[40] Maksim Kuprashevich and Irina Tolstykh. Mivolo: Multi-input transformer for age and gender estimation. arXiv preprint arXiv:2307.04616, 2023. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 444, + 553, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 444, + 553, + 486 + ], + "spans": [ + { + "bbox": [ + 321, + 444, + 553, + 486 + ], + "type": "text", + "content": "[41] Xing Lan, Qinghao Hu, Qiang Chen, Jian Xue, and Jian Cheng. Hih: Towards more accurate face alignment via heatmap in heatmap. arXiv preprint arXiv:2104.03100, 2021.7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 488, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 488, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 321, + 488, + 553, + 533 + ], + "type": "text", + "content": "[42] Cheng-Han Lee, Ziwei Liu, Lingyun Wu, and Ping Luo. Maskgan: Towards diverse and interactive facial image manipulation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 534, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 534, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 321, + 534, + 553, + 578 + ], + "type": "text", + "content": "[43] Gil Levi and Tal Hassner. Age and gender classification using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 34-42, 2015. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 579, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 579, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 321, + 579, + 553, + 623 + ], + "type": "text", + "content": "[44] Hangyu Li, Nannan Wang, Xinpeng Ding, Xi Yang, and Xinbo Gao. Adaptively learning facial expression representation via cf labels and distillation. IEEE Transactions on Image Processing, 30:2016-2028, 2021. 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 624, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 624, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 321, + 624, + 553, + 678 + ], + "type": "text", + "content": "[45] Shan Li, Weihong Deng, and JunPing Du. Reliable crowdsourcing and deep locality-preserving learning for expression recognition in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2852-2861, 2017. 2, 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 321, + 681, + 553, + 713 + ], + "type": "text", + "content": "[46] Wanhua Li, Xiaoke Huang, Jiwen Lu, Jianjiang Feng, and Jie Zhou. Learning probabilistic ordinal embeddings for uncertainty-aware regression. In Proceedings of the" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11378" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 81, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 81, + 73, + 294, + 95 + ], + "type": "text", + "content": "IEEE/CVF conference on computer vision and pattern recognition, pages 13896-13905, 2021. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 97, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 97, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 61, + 97, + 295, + 140 + ], + "type": "text", + "content": "[47] Xiangtai Li, Haobo Yuan, Wei Li, Henghui Ding, Size Wu, Wenwei Zhang, Yining Li, Kai Chen, and Chen Change Loy. Omg-seg: Is one model good enough for all segmentation? arXiv preprint arXiv:2401.10229, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 143, + 294, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 143, + 294, + 186 + ], + "spans": [ + { + "bbox": [ + 61, + 143, + 294, + 186 + ], + "type": "text", + "content": "[48] Yong Li, Jiabei Zeng, Shiguang Shan, and Xilin Chen. Occlusion aware facial expression recognition using cnn with attention mechanism. IEEE Transactions on Image Processing, 28(5):2439-2450, 2018. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 189, + 295, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 189, + 295, + 232 + ], + "spans": [ + { + "bbox": [ + 62, + 189, + 295, + 232 + ], + "type": "text", + "content": "[49] Chunze Lin, Beier Zhu, Quan Wang, Renjie Liao, Chen Qian, Jiwen Lu, and Jie Zhou. Structure-coherent deep feature learning for robust face alignment. IEEE Transactions on Image Processing, 30:5313-5326, 2021. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 235, + 294, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 235, + 294, + 278 + ], + "spans": [ + { + "bbox": [ + 62, + 235, + 294, + 278 + ], + "type": "text", + "content": "[50] Jinpeng Lin, Hao Yang, Dong Chen, Ming Zeng, Fang Wen, and Lu Yuan. Face parsing with roi tanh-warping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5654-5663, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 281, + 294, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 281, + 294, + 334 + ], + "spans": [ + { + "bbox": [ + 62, + 281, + 294, + 334 + ], + "type": "text", + "content": "[51] Hai Liu, Shuai Fang, Zhaoli Zhang, Duantengchuan Li, Ke Lin, and Jiazhang Wang. Mfdnet: Collaborative poses perception and matrix fisher distribution for head pose estimation. IEEE Transactions on Multimedia, 24:2449-2460, 2021. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 337, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 337, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 62, + 337, + 294, + 370 + ], + "type": "text", + "content": "[52] S Liu, L Zhang, X Yang, H Su, and J Zhu. Query2label: A simple transformer way to multi-label classification. arxiv 2021. arXiv preprint arXiv:2107.10834, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 372, + 294, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 372, + 294, + 416 + ], + "spans": [ + { + "bbox": [ + 62, + 372, + 294, + 416 + ], + "type": "text", + "content": "[53] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3730-3738, 2015. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 418, + 294, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 418, + 294, + 460 + ], + "spans": [ + { + "bbox": [ + 62, + 418, + 294, + 460 + ], + "type": "text", + "content": "[54] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), 2015. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 464, + 294, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 464, + 294, + 497 + ], + "spans": [ + { + "bbox": [ + 62, + 464, + 294, + 497 + ], + "type": "text", + "content": "[55] Ling Luo, Dingyu Xue, and Xinglong Feng. Ehanet: An effective hierarchical aggregation network for face parsing. Applied Sciences, 10(9):3135, 2020. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 498, + 294, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 498, + 294, + 542 + ], + "spans": [ + { + "bbox": [ + 62, + 498, + 294, + 542 + ], + "type": "text", + "content": "[56] Upal Mahbub, Sayantan Sarkar, and Rama Chellappa. Segment-based methods for facial attribute detection from partial faces. IEEE Transactions on Affective Computing, 11(4):601-613, 2018. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 544, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 544, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 62, + 544, + 294, + 588 + ], + "type": "text", + "content": "[57] Longbiao Mao, Yan Yan, Jing-Hao Xue, and Hanzi Wang. Deep multi-task multi-label cnn for effective facial attribute classification. IEEE Transactions on Affective Computing, 13(2):818-828, 2020. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 590, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 590, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 62, + 590, + 294, + 645 + ], + "type": "text", + "content": "[58] Chen Mi, Baoxi Yuan, Peng Ma, Yingxia Guo, Le Qi, Feng Wang, Wenbo Wu, and Lingling Wang. Visibility prediction based on landmark detection in foggy weather. In 2020 International Conference on Robots & Intelligent System (ICRIS), pages 134-137, 2020. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 647, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 294, + 712 + ], + "type": "text", + "content": "[59] Paul Micaelli, Arash Vahdat, Hongxu Yin, Jan Kautz, and Pavlo Molchanov. Recurrence without recurrence: Stable video landmark detection with deep equilibrium models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22814-22825, 2023. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 321, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 321, + 73, + 553, + 116 + ], + "type": "text", + "content": "[60] Zuheng Ming, Junshi Xia, Muhammad Muzzamil Luqman, Jean-Christophe Burie, and Kaixing Zhao. Dynamic multitask learning for face recognition with facial expression. arXiv preprint arXiv:1911.03281, 2019. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 118, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 118, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 320, + 118, + 553, + 172 + ], + "type": "text", + "content": "[61] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE transactions on pattern analysis and machine intelligence, 41(8):1979–1993, 2018. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 175, + 553, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 175, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 320, + 175, + 553, + 218 + ], + "type": "text", + "content": "[62] Ali Mollahosseini, Behzad Hasani, and Mohammad H Ma-hoor. Affectnet: A database for facial expression, valence, and arousal computing in the wild. IEEE Transactions on Affective Computing, 10(1):18-31, 2017. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 220, + 553, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 220, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 320, + 220, + 553, + 284 + ], + "type": "text", + "content": "[63] Stylianos Moschoglou, Athanasios Papaioannou, Christos Sagonas, Jiankang Deng, Irene Kotsia, and Stefanos Zafeiriou. Agedb: the first manually collected, in-the-wild age database. In proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 51-59, 2017. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 286, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 286, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 320, + 286, + 553, + 341 + ], + "type": "text", + "content": "[64] Erik Murphy-Chutorian, Anup Doshi, and Mohan Manubhai Trivedi. Head pose estimation for driver assistance systems: A robust algorithm and experimental evaluation. In 2007 IEEE intelligent transportation systems conference, pages 709-714. IEEE, 2007. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 342, + 555, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 342, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 320, + 342, + 555, + 397 + ], + "type": "text", + "content": "[65] Kartik Narayan, Harsh Agarwal, Kartik Thakral, Surbhi Mittal, Mayank Vatsa, and Richa Singh. Df-platter: Multiface heterogeneous deepfake dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9739-9748, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 399, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 399, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 320, + 399, + 553, + 442 + ], + "type": "text", + "content": "[66] Kartik Narayan, Nithin Gopalakrishnan Nair, Jennifer Xu, Rama Chellappa, and Vishal M Patel. Petalface: Parameter efficient transfer learning for low-resolution face recognition. arXiv preprint arXiv:2412.07771, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 444, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 444, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 320, + 444, + 553, + 475 + ], + "type": "text", + "content": "[67] Kartik Narayan, Vibashan VS, and Vishal M Patel. Segface: Face segmentation of long-tail classes. arXiv preprint arXiv:2412.08647, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 478, + 553, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 478, + 553, + 531 + ], + "spans": [ + { + "bbox": [ + 320, + 478, + 553, + 531 + ], + "type": "text", + "content": "[68] Zhenxing Niu, Mo Zhou, Le Wang, Xinbo Gao, and Gang Hua. Ordinal regression with multiple output cnn for age estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4920-4928, 2016. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 534, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 534, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 320, + 534, + 553, + 578 + ], + "type": "text", + "content": "[69] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European conference on computer vision, pages 69-84. Springer, 2016. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 579, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 579, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 320, + 579, + 553, + 634 + ], + "type": "text", + "content": "[70] Jakub Paplham, Vojt Franc, et al. A call to reflect on evaluation practices for age estimation: Comparative analysis of the state-of-the-art and a unified benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1196-1205, 2024. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 320, + 635, + 553, + 689 + ], + "type": "text", + "content": "[71] Lixiong Qin, Mei Wang, Chao Deng, Ke Wang, Xi Chen, Jiani Hu, and Weihong Deng. Swinface: a multi-task transformer for face recognition, expression recognition, age estimation and attribute estimation. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 2, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 691, + 553, + 713 + ], + "type": "text", + "content": "[72] Lixiong Qin, Mei Wang, Xuannan Liu, Yuhang Zhang, Wei Deng, Xiaoshuai Song, Weiran Xu, and Weihong Deng." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11379" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 81, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 81, + 73, + 294, + 106 + ], + "type": "text", + "content": "Facetor: A generalist model for face perception. In European Conference on Computer Vision, pages 240-260. Springer, 2025. 2, 3, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 107, + 295, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 107, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 61, + 107, + 295, + 173 + ], + "type": "text", + "content": "[73] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 175, + 295, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 175, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 61, + 175, + 295, + 228 + ], + "type": "text", + "content": "[74] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 231, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 231, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 62, + 231, + 294, + 285 + ], + "type": "text", + "content": "[75] Rajeev Ranjan, Vishal M Patel, and Rama Chellappa. Hyperface: A deep multi-task learning framework for face detection, landmark localization, pose estimation, and gender recognition. IEEE transactions on pattern analysis and machine intelligence, 41(1):121-135, 2017. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 287, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 287, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 62, + 287, + 294, + 342 + ], + "type": "text", + "content": "[76] Rajeev Ranjan, Swami Sankaranarayanan, Carlos D Castillo, and Rama Chellappa. An all-in-one convolutional neural network for face analysis. In 2017 12th IEEE international conference on automatic face & gesture recognition (FG 2017), pages 17-24. IEEE, 2017. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 343, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 343, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 62, + 343, + 294, + 408 + ], + "type": "text", + "content": "[77] Ethan M Rudd, Manuel Günther, and Terrance E Boult. Moon: A mixed objective optimization network for the recognition of facial attributes. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14, pages 19-35. Springer, 2016. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 411, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 411, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 62, + 411, + 294, + 453 + ], + "type": "text", + "content": "[78] Nataniel Ruiz, Eunjji Chong, and James M Rehg. Fine-grained head pose estimation without keypoints. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 2074-2083, 2018. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 456, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 456, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 62, + 456, + 294, + 510 + ], + "type": "text", + "content": "[79] Christos Sagonas, Georgios Tzimiropoulos, Stefanos Zafeiriou, and Maja Pantic. 300 faces in-the-wild challenge: The first facial landmark localization challenge. In Proceedings of the IEEE international conference on computer vision workshops, pages 397-403, 2013. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 512, + 295, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 512, + 295, + 566 + ], + "spans": [ + { + "bbox": [ + 62, + 512, + 295, + 566 + ], + "type": "text", + "content": "[80] Mausoom Sarkar, Mayur Hemani, Rishabh Jain, Balaji Krishnamurthy, et al. Parameter efficient local implicit image function network for face segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20970-20980, 2023. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 568, + 294, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 568, + 294, + 611 + ], + "spans": [ + { + "bbox": [ + 62, + 568, + 294, + 611 + ], + "type": "text", + "content": "[81] S. Sengupta, J.C. Cheng, C.D. Castillo, V.M. Patel, R. Chellappa, and D.W. Jacobs. Frontal to profile face verification in the wild. In IEEE Conference on Applications of Computer Vision, 2016. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 613, + 294, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 613, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 62, + 613, + 294, + 677 + ], + "type": "text", + "content": "[82] Jiahui She, Yibo Hu, Hailin Shi, Jun Wang, Qiu Shen, and Tao Mei. Dive into ambiguity: Latent distribution mining and pairwise uncertainty estimation for facial expression recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6248-6257, 2021. 1, 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 680, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 680, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 680, + 294, + 712 + ], + "type": "text", + "content": "[83] Jie Shen, Stefanos Zafeiriou, Grigoris G Chrysos, Jean Kossaifi, Georgios Tzimiropoulos, and Maja Pantic. The first facial landmark tracking in-the-wild challenge: Benchmark" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 321, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 338, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 338, + 73, + 553, + 105 + ], + "type": "text", + "content": "and results. In Proceedings of the IEEE international conference on computer vision workshops, pages 50-58, 2015. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 321, + 107, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 107, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 321, + 107, + 553, + 161 + ], + "type": "text", + "content": "[84] Nyeong-Ho Shin, Seon-Ho Lee, and Chang-Su Kim. Moving window regression: A novel approach to ordinal regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18760-18769, 2022. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 163, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 163, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 321, + 163, + 553, + 228 + ], + "type": "text", + "content": "[85] Ying Shu, Yan Yan, Si Chen, Jing-Hao Xue, Chunhua Shen, and Hanzi Wang. Learning spatial-semantic relationship for facial attribute recognition with limited labeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11916–11925, 2021. 2, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 230, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 230, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 321, + 230, + 553, + 274 + ], + "type": "text", + "content": "[86] Dominykas Strazdas, Jan Hintz, and Ayoub Al-Hamadi. Robo-hud: Interaction concept for contactless operation of industrial cobotic systems. Applied Sciences, 11(12):5366, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 276, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 276, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 321, + 276, + 553, + 374 + ], + "type": "text", + "content": "[87] Carole H Sudre, Wenqi Li, Tom Vercauteren, Sebastian Ourselin, and M Jorge Cardoso. Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: Third International Workshop, DLMIA 2017, and 7th International Workshop, ML-CDS 2017, Held in conjunction with MICCAI 2017, Quebec City, QC, Canada, September 14, Proceedings 3, pages 240-248. Springer, 2017. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 376, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 376, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 321, + 376, + 553, + 408 + ], + "type": "text", + "content": "[88] Haomiao Sun, Mingjie He, Shiguang Shan, Hu Han, and Xilin Chen. Task-adaptive q-face. arXiv preprint arXiv:2405.09059, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 411, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 411, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 321, + 411, + 553, + 453 + ], + "type": "text", + "content": "[89] Yi Sun, Yuheng Chen, Xiaogang Wang, and Xiaou Tang. Deep learning face representation by joint identification-verification. Advances in neural information processing systems, 27, 2014. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 456, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 456, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 321, + 456, + 553, + 510 + ], + "type": "text", + "content": "[90] Yaniv Taigman, Ming Yang, Marc'Aurelio Ranzato, and Lior Wolf. Deepface: Closing the gap to human-level performance in face verification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1701-1708, 2014. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 512, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 512, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 321, + 512, + 553, + 566 + ], + "type": "text", + "content": "[91] Gusi Te, Yinglu Liu, Wei Hu, Hailin Shi, and Tao Mei. Edge-aware graph representation learning and reasoning for face parsing. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16, pages 258-274. Springer, 2020. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 568, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 568, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 321, + 568, + 553, + 610 + ], + "type": "text", + "content": "[92] Gusi Te, Wei Hu, Yinglu Liu, Hailin Shi, and Tao Mei. Agrnet: Adaptive graph representation learning and reasoning for face parsing. IEEE Transactions on Image Processing, 30:8236-8250, 2021. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 613, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 613, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 321, + 613, + 553, + 667 + ], + "type": "text", + "content": "[93] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 321, + 670, + 553, + 712 + ], + "type": "text", + "content": "[94] Roberto Valle, José M Buenapasada, Antonio Valdés, and Luis Baumela. Face alignment using a 3d deeply-initialized ensemble of regression trees. Computer Vision and Image Understanding, 189:102846, 2019. 2, 6" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "11380" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "type": "text", + "content": "[95] Roberto Valle, José M Buenapasada, and Luis Baumela. Multi-task head pose estimation in-the-wild. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43 (8):2874-2881, 2020. 2, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 118, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 118, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 61, + 118, + 294, + 162 + ], + "type": "text", + "content": "[96] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 164, + 295, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 164, + 295, + 218 + ], + "spans": [ + { + "bbox": [ + 62, + 164, + 295, + 218 + ], + "type": "text", + "content": "[97] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5265-5274, 2018. 2, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 220, + 294, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 220, + 294, + 284 + ], + "spans": [ + { + "bbox": [ + 62, + 220, + 294, + 284 + ], + "type": "text", + "content": "[98] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 43(10):3349-3364, 2020. 2, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 287, + 294, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 287, + 294, + 351 + ], + "spans": [ + { + "bbox": [ + 62, + 287, + 294, + 351 + ], + "type": "text", + "content": "[99] Junke Wang, Dongdong Chen, Zuxuan Wu, Chong Luo, Luowei Zhou, Yucheng Zhao, Yujia Xie, Ce Liu, Yu-Gang Jiang, and Lu Yuan. Omnivl: One foundation model for image-language and video-language tasks. Advances in neural information processing systems, 35:5696-5710, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 354, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 354, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 57, + 354, + 294, + 408 + ], + "type": "text", + "content": "[100] Kai Wang, Xiaojiang Peng, Jianfei Yang, Shijian Lu, and Yu Qiao. Suppressing uncertainties for large-scale facial expression recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6897-6906, 2020. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 411, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 411, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 57, + 411, + 294, + 453 + ], + "type": "text", + "content": "[101] Kai Wang, Xiaojiang Peng, Jianfei Yang, Debin Meng, and Yu Qiao. Region attention networks for pose and occlusion robust facial expression recognition. IEEE Transactions on Image Processing, 29:4057-4069, 2020. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 456, + 294, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 456, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 57, + 456, + 294, + 499 + ], + "type": "text", + "content": "[102] Qianqian Wang, Yen-Yu Chang, Ruojin Cai, Zhengqi Li, Bharath Hariharan, Aleksander Holynski, and Noah Snavely. Tracking everything everywhere all at once. arXiv preprint arXiv:2306.05422, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 501, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 501, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 57, + 501, + 294, + 566 + ], + "type": "text", + "content": "[103] Zhenyu Wang, Yali Li, Xi Chen, Ser-Nam Lim, Antonio Torralba, Hengshuang Zhao, and Shengjin Wang. Detecting everything in the open world: Towards universal object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11433-11443, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 568, + 294, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 568, + 294, + 621 + ], + "spans": [ + { + "bbox": [ + 57, + 568, + 294, + 621 + ], + "type": "text", + "content": "[104] Zhen Wei, Yao Sun, Jinqiao Wang, Hanjiang Lai, and Si Liu. Learning adaptive receptive fields for deep image parsing network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2434-2442, 2017. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 624, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 294, + 657 + ], + "type": "text", + "content": "[105] Zhen Wei, Si Liu, Yao Sun, and Hefei Ling. Accurate facial image parsing at real-time speed. IEEE Transactions on Image Processing, 28(9):4659-4670, 2019. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 294, + 712 + ], + "type": "text", + "content": "[106] Tiancheng Wen, Zhonggan Ding, Yongqiang Yao, Yaxiong Wang, and Xueming Qian. Picassonet: Searching adaptive architecture for efficient facial landmark localization. IEEE Transactions on Neural Networks and Learning Systems, 34(12):10516-10527, 2023. 7" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 126 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 126 + ], + "type": "text", + "content": "[107] Lior Wolf, Tal Hassner, and Yaniv Taigman. Effective unconstrained face recognition by combining multiple descriptors and learned background statistics. IEEE transactions on pattern analysis and machine intelligence, 33(10): 1978-1990, 2010. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 129, + 553, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 193 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 193 + ], + "type": "text", + "content": "[108] Erroll Wood, Tadas Baltrusaitis, Charlie Hewitt, Matthew Johnson, Jingjing Shen, Nikola Milosavljevic, Daniel Wilde, Stephan Garbin, Toby Sharp, Ivan Stojiljković, et al. 3d face reconstruction with dense landmarks. In European Conference on Computer Vision, pages 160-177. Springer, 2022. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 195, + 553, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 195, + 553, + 248 + ], + "spans": [ + { + "bbox": [ + 317, + 195, + 553, + 248 + ], + "type": "text", + "content": "[109] Wayne Wu, Chen Qian, Shuo Yang, Quan Wang, Yici Cai, and Qiang Zhou. Look at boundary: A boundary-aware face alignment algorithm. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2129-2138, 2018. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 251, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 293 + ], + "type": "text", + "content": "[110] Yue Wu and Qiang Ji. Robust facial landmark detection under significant head poses and occlusion. In Proceedings of the IEEE International Conference on Computer Vision, pages 3658-3666, 2015. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 294, + 553, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 294, + 553, + 347 + ], + "spans": [ + { + "bbox": [ + 317, + 294, + 553, + 347 + ], + "type": "text", + "content": "[111] Yue Wu, Chao Gou, and Qiang Ji. Simultaneous facial landmark detection, pose and deformation estimation under facial occlusion. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3471-3480, 2017. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 350, + 553, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 350, + 553, + 413 + ], + "spans": [ + { + "bbox": [ + 317, + 350, + 553, + 413 + ], + "type": "text", + "content": "[112] Jiahao Xia, Weiwei Qu, Wenjian Huang, Jianguo Zhang, Xi Wang, and Min Xu. Sparse local patch transformer for robust face alignment and landmarks inherent relation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4052-4061, 2022. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 415, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 469 + ], + "type": "text", + "content": "[113] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 471, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 471, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 471, + 553, + 514 + ], + "type": "text", + "content": "[114] Miao Xin, Shentong Mo, and Yuanze Lin. Eva-gcn: Head pose estimation based on graph convolutional networks. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 1462-1471, 2021. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 515, + 553, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 515, + 553, + 579 + ], + "spans": [ + { + "bbox": [ + 317, + 515, + 553, + 579 + ], + "type": "text", + "content": "[115] Xinchen Yan, Jimei Yang, Kihyuk Sohn, and Honglak Lee. Attribute2image: Conditional image generation from visual attributes. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 776-791. Springer, 2016. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 581, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 581, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 581, + 553, + 635 + ], + "type": "text", + "content": "[116] Tsun-Yi Yang, Yi-Ting Chen, Yen-Yu Lin, and Yung-Yu Chuang. Fsa-net: Learning fine-grained structure aggregation for head pose estimation from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1087-1096, 2019. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 636, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 553, + 690 + ], + "type": "text", + "content": "[117] Xiangnan Yin, Di Huang, Zehua Fu, Yunhong Wang, and Liming Chen. Segmentation-reconstruction-guided facial image de-occlusion. In 2023 IEEE 17th International Conference on Automatic Face and Gesture Recognition (FG), pages 1–8. IEEE, 2023. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[118] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang," + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "11381" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 81, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 81, + 72, + 294, + 106 + ], + "type": "text", + "content": "Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "type": "text", + "content": "[119] Alireza Zaeemzadeh, Shabnam Ghadar, Baldo Faieta, Zhe Lin, Nazanin Rahnavard, Mubarak Shah, and Ratheesh Kalarot. Face image retrieval with attribute manipulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12116-12125, 2021. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 295, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 295, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 295, + 207 + ], + "type": "text", + "content": "[120] Jiabei Zeng, Shiguang Shan, and Xilin Chen. Facial expression recognition with inconsistently annotated datasets. In Proceedings of the European conference on computer vision (ECCV), pages 222-237, 2018. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 295, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 295, + 262 + ], + "type": "text", + "content": "[121] Cheng Zhang, Hai Liu, Yongjian Deng, Bochen Xie, and Youfu Li. Tokenhpe: Learning orientation tokens for efficient head pose estimation via transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8897-8906, 2023. 2, 5, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 295, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 295, + 308 + ], + "type": "text", + "content": "[122] Hongwen Zhang, Qi Li, Zhenan Sun, and Yunfan Liu. Combining data-driven and model-driven methods for robust facial landmark detection. IEEE Transactions on Information Forensics and Security, 13(10):2409-2422, 2018. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 310, + 295, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 295, + 365 + ], + "type": "text", + "content": "[123] Ning Zhang, Manohar Paluri, Marc'Aurelio Ranzato, Trevor Darrell, and Lubomir Bourdev. Panda: Pose aligned networks for deep attribute modeling. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1637-1644, 2014. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 295, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 295, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 295, + 421 + ], + "type": "text", + "content": "[124] Zhanpeng Zhang, Ping Luo, Chen Change Loy, and Xiaou Tang. Facial landmark detection by deep multi-task learning. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VI 13, pages 94-108. Springer, 2014. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 295, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 295, + 466 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 295, + 466 + ], + "type": "text", + "content": "[125] Zhifei Zhang, Yang Song, and Hairong Qi. Age progression/regression by conditional adversarial autoencoder. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5810-5818, 2017. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 467, + 295, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 295, + 521 + ], + "type": "text", + "content": "[126] Rui Zhao, Tianshan Liu, Jun Xiao, Daniel PK Lun, and Kin-Man Lam. Deep multi-task learning for facial expression recognition and synthesis based on selective feature sharing. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 4412-4419. IEEE, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 522, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 522, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 522, + 295, + 578 + ], + "type": "text", + "content": "[127] Qingping Zheng, Jiankang Deng, Zheng Zhu, Ying Li, and Stefanos Zafeiriou. Decoupled multi-task learning with cyclical self-regulation for face parsing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4156-4165, 2022. 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 295, + 624 + ], + "type": "text", + "content": "[128] Tianyue Zheng and Weihong Deng. Cross-pose lfw: A database for studying cross-pose face recognition in unconstrained environments. Beijing University of Posts and Telecommunications, Tech. Rep, 5(7):5, 2018. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 667 + ], + "type": "text", + "content": "[129] Tianyue Zheng, Weihong Deng, and Jiani Hu. Crossage lfw: A database for studying cross-age face recognition in unconstrained environments. arXiv preprint arXiv:1708.08197, 2017. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 669, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 295, + 713 + ], + "type": "text", + "content": "[130] Yinglin Zheng, Hao Yang, Ting Zhang, Jianmin Bao, Dongdong Chen, Yangyu Huang, Lu Yuan, Dong Chen, Ming Zeng, and Fang Wen. General facial representation learning in a visual-linguistic manner. In Proceedings of the" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 465 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "type": "text", + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18697-18709, 2022. 2, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "text", + "content": "[131] Yijun Zhou and James Gregson. Whenet: Real-time fine-grained estimation for wide range head pose. arXiv preprint arXiv:2005.10353, 2020. 1, 2, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 130, + 555, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 555, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 555, + 185 + ], + "type": "text", + "content": "[132] Zhenglin Zhou, Huaxia Li, Hong Liu, Nanyang Wang, Gang Yu, and Rongrong Ji. Star loss: Reducing semantic ambiguity in facial landmark detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15475-15484, 2023. 1, 2, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 186, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 230 + ], + "type": "text", + "content": "[133] Jiawen Zhu, Simiao Lai, Xin Chen, Dong Wang, and Huchuan Lu. Visual prompt multi-modal tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9516-9526, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 231, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 553, + 285 + ], + "type": "text", + "content": "[134] Peihao Zhu, Rameen Abdal, Yipeng Qin, and Peter Wonka. Sean: Image synthesis with semantic region-adaptive normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5104-5113, 2020. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "type": "text", + "content": "[135] Xiangyu Zhu, Zhen Lei, Xiaoming Liu, Hailin Shi, and Stan Z Li. Face alignment across large poses: A 3d solution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 146-155, 2016. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "type": "text", + "content": "[136] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 376, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 420 + ], + "type": "text", + "content": "[137] Ni Zhuang, Yan Yan, Si Chen, and Hanzi Wang. Multi-task learning of cascaded cnn for facial attribute classification. In 2018 24th International Conference on Pattern Recognition (ICPR), pages 2069-2074. IEEE, 2018. 2, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 421, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 421, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 316, + 421, + 553, + 465 + ], + "type": "text", + "content": "[138] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11382" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Factorized Learning for Temporally Grounded Video-Language Models/0391813b-b98c-4912-884e-a194864058af_content_list.json b/2025/Factorized Learning for Temporally Grounded Video-Language Models/0391813b-b98c-4912-884e-a194864058af_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f1afbbc07cd423c832ff7d7730e8a768b459e839 --- /dev/null +++ b/2025/Factorized Learning for Temporally Grounded Video-Language Models/0391813b-b98c-4912-884e-a194864058af_content_list.json @@ -0,0 +1,1478 @@ +[ + { + "type": "text", + "text": "Factorized Learning for Temporally Grounded Video-Language Models", + "text_level": 1, + "bbox": [ + 137, + 130, + 859, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wenzheng Zeng, Difei Gao, Mike Zheng Shou†, Hwee Tou Ng† National University of Singapore", + "bbox": [ + 246, + 179, + 750, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "wenzhengzeng@u.nus.edu, {daniel.difei.gao, mike.zheng.shou}@gmail.com, nght@nus.edu.sg", + "bbox": [ + 112, + 219, + 874, + 234 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/734db46a50c501461c381372d16c4057a9a6ff5a1a51c7bb6694bb24042975d2.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 94, + 279, + 416, + 477 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2ee64aa051f07996dc9f92f80ba9a130050ff53409093b1ceb61ef052113fb3d.jpg", + "image_caption": [ + "Figure 1. (a) Performance: Our method consistently outperforms SOTA methods across various tasks (here we draw the maximum performance across methods, detailed in Sec. 6). (b) Model: We propose a new framework $\\mathrm{D}^2\\mathrm{VLM}$ , where we decompose the generation objective into a \"grounding then answering with evidence referencing\" paradigm and introduce evidence tokens to emphasize explicit event-level visual semantic capture. (c) Training Algorithm: We introduce Factorized Preference Optimization (FPO) that explicitly addresses both temporal grounding and textual response. A factorized data synthesis approach is also designed to support FPO." + ], + "image_footnote": [], + "bbox": [ + 423, + 268, + 898, + 479 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 583, + 326, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent video-language models have shown great potential for video understanding, but still struggle with accurate temporal grounding for event-level perception. We observe that two main factors in video understanding (i.e., temporal grounding and textual response) form a logical hierarchy: accurate temporal evidence grounding lays the foundation for reliable textual response. However, existing works typically handle these two tasks in a coupled manner without a clear logical structure, leading to sub-optimal objectives. We address this from a factorized learning perspective. We first propose $D^{2}$ VLM, a framework that decouples the learning of these two tasks while also emphasizing their inherent dependency. We adopt a \"grounding then answering with evidence referencing\" paradigm and introduce evidence tokens for evidence grounding, which emphasize event-level visual semantic capture beyond the focus on timestamp representation in existing works. To further facilitate the learning of these two tasks, we introduce a novel factorized", + "bbox": [ + 86, + 614, + 485, + 888 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "preference optimization (FPO) algorithm. Unlike standard preference optimization, FPO explicitly incorporates probabilistic temporal grounding modeling into the optimization objective, enabling preference learning for both temporal grounding and textual response. We also construct a synthetic dataset to address the lack of suitable datasets for factorized preference learning with explicit temporal grounding. Experiments on various tasks demonstrate the clear advantage of our approach. Our source code is available at https://github.com/nusnlp/d2v1m.", + "bbox": [ + 511, + 569, + 908, + 720 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 753, + 645, + 768 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in video-language models, especially those built upon large language models (video LLMs), have enabled remarkable progress in video understanding [10, 21, 23, 25, 31, 44]. Through their flexible video&text-int-text-out nature, video LLMs demonstrate great potential as general-purpose solvers, unifying various tasks (e.g., temporal grounding [9], dense captioning [47], and question answering [7]) into one generic \"video question answer-", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding authors", + "bbox": [ + 114, + 898, + 245, + 911 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20683", + "bbox": [ + 478, + 949, + 517, + 960 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing\" framework. Despite their effectiveness, existing video LLMs still struggle with accurate temporal grounding for event perception and localization [10, 11, 23, 31, 31], a crucial capability not only for grounding tasks themselves but also for relevant tasks that need textual answering.", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We notice that for video understanding, temporal event-level grounding and textual response are two primary tasks that exhibit distinct characteristics yet maintain strong logical dependencies. Specifically, temporal grounding focuses on precisely locating temporal events (evidence) that support answering, while textual response emphasizes accurate interpretation from the grounded evidence and generating coherent textual answers. However, existing methods [4, 13, 14, 23, 31] typically handle these two tasks in a coupled manner, with two major limitations: (1) Various special tokens are designed for temporal grounding [4, 14, 23], but their generation is mixed with text token generation without a clear logical structure, leading to coupled learning objectives. (2) More importantly, these special tokens mainly focus on timestamp representation for precise timestamp output, lacking the explicit capture of visual semantics from grounded events. In contrast, we argue that such event-level visual semantics should not be overlooked, as it could inherently serve as crucial contexts for subsequent textual answer generation, especially under the next-token prediction paradigm.", + "bbox": [ + 91, + 170, + 482, + 487 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on the aforementioned observations, we propose to address this from a factorized learning perspective. We first introduce a new framework $\\mathrm{D}^2\\mathrm{VLM}$ that decouples the learning of temporal evidence grounding and textual answering, while preserving and even strengthening their inherent dependency. Specifically, as shown in Fig. 1 (b), we decompose the model response into two sequential stages: (1) pure temporal grounding that aims to localize and capture essential visual evidence for response, followed by (2) interleaved text-evidence answer generation, where both the textual answer and temporal information are produced in an evidence-referencing manner to establish consistency with the previously grounded evidence.", + "bbox": [ + 89, + 489, + 482, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Technically, we introduce evidence tokens, a special token type dedicated to temporal evidence grounding. Different from existing designs of grounding tokens that focus on its special category and timestamp representation, our evidence tokens not only aim to determine the temporal location of the grounded event, but also emphasize the capture of event-level visual semantics, which serves as crucial context for subsequent answer generation. During the subsequent interleaved text-evidence generation, the evidence-referencing process is achieved by generating evidence tokens that align with those from the previous grounding stage. This ensures that the output information in the final response remains consistent with the initially grounded evidence while reinforcing logical coherence across stages.", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Besides providing decoupled and clearer task objectives, our design naturally fits well with the teacher-forcing autoregressive training paradigm, as subsequent textual response generation is conditioned on the correctly grounded evidence, enabling a learning shortcut for more stable training. Our experiments demonstrate that both the designed sequence generation objective and event-level visual semantic capture are essential for performance improvement, offering valuable insights for future model design.", + "bbox": [ + 511, + 90, + 903, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To further facilitate the learning of these two tasks, we propose a novel Factorized Preference Optimization (FPO) algorithm. Unlike standard preference optimization, FPO incorporates probabilistic temporal grounding modeling into the optimization objective, enabling explicit preference learning for temporal grounding in addition to standard textual preference. Meanwhile, another obstacle is that existing video preference datasets do not account for such temporal grounding aspects, making it infeasible to directly apply our proposed FPO algorithm. To overcome this limitation, we construct a synthetic dataset by introducing factorized perturbations into the original preferred response sequence. These perturbations are applied at the sub-video event level, considering two main factors: temporal grounding and textual response, along with multiple possible sub-factors. This approach ensures a structured and controllable noise generation process, where the cause and type of noise are precisely known without any manual annotation. As a result, our data synthesis process is both reliable and scalable, making it possible to effectively conduct the proposed factorized preference optimization. To the best of our knowledge, we are the first to use preference learning involving explicit temporal grounding, and moreover, in a factorized manner, as illustrated in Fig. 1 (c).", + "bbox": [ + 511, + 228, + 903, + 589 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct extensive experiments to demonstrate the superiority of our proposed framework. An intuitive comparison is shown in Fig. 1 (a), where our 3.8B model consistently outperforms existing SOTA methods (typically ranging from 3.8B to 13B in model size) across various tasks.", + "bbox": [ + 511, + 590, + 903, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 684, + 653, + 699 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Video LLMs", + "text_level": 1, + "bbox": [ + 513, + 709, + 648, + 724 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The recent development of video LLMs has shown great potential as generic video understanding assistants [21, 23, 25, 44], primarily due to their flexible video&text-in-text-out capability. Despite the effectiveness, existing video LLMs still struggle with accurate temporal event-level perception and localization (i.e., temporal grounding) [10, 13, 23, 31, 39], a crucial capability not only for grounding tasks but also for generating correct textual answers that are heavily based on grounding ability.", + "bbox": [ + 511, + 732, + 903, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Various efforts have been made to enhance the temporal grounding ability of video LLMs [4, 10, 11, 14]. Be", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20684", + "bbox": [ + 478, + 949, + 517, + 960 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sides adding targeted data for temporally sensitive fine-tuning [13, 20, 31, 37, 43], recent work usually designs different special tokens for precise temporal representation [10, 14, 23, 29, 36], where some work further combines such special tokens with extra grounding decoder [4, 11]. Despite their effectiveness, we argue that there still exist two potential limitations. First, in the generation sequence, special time tokens are interleaved with text tokens, leading to a highly coupled learning objective between temporal grounding and textual answering. Second, these special tokens are primarily designed for timestamp representations to output timestamp, lacking the ability to capture the visual semantics of grounded events.", + "bbox": [ + 89, + 90, + 480, + 286 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we aim to enhance temporally grounded video LLMs from a factorized learning perspective. We first decompose the task into learning of temporal evidence grounding and textual response while also strengthening their inherent dependency. We also propose a factorized preference learning paradigm to further facilitate the two-task learning.", + "bbox": [ + 89, + 287, + 480, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Preference Optimization", + "text_level": 1, + "bbox": [ + 89, + 400, + 316, + 416 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reinforcement learning from human feedback (RLHF) [28, 33] or from AI feedback (RLAIF) [3, 15] has been widely used to align LLMs' behavior with human preference, with PPO [27, 32, 33, 48] and DPO [30] as common implementations. While this spirit has extended to the multimodal domain [17, 35, 45], it remains underexplored for video LLMs [18, 46], especially regarding explicit temporal grounding. In this work, we introduce a factorized preference optimization (FPO) algorithm that explicitly incorporates probabilistic temporal grounding modeling into the optimization objective, enabling preference learning for both temporal grounding and textual responses. We also construct a synthetic dataset to address the lack of suitable datasets for factorized preference learning for explicit temporal grounding.", + "bbox": [ + 89, + 422, + 482, + 650 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 664, + 181, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here we will present the proposed $\\mathrm{D}^2\\mathrm{VLM}$ framework.", + "bbox": [ + 89, + 689, + 455, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 89, + 712, + 218, + 728 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We begin by first illustrating the general formulation of temporally grounded video-language models, particularly in the context of video LLMs. Given a video $V$ and a textual question (or prompt) $Q$ , the goal is to answer the question based on the given video, where the answer is typically composed of textual response together with temporal grounding information (i.e., localized temporal interval).", + "bbox": [ + 89, + 734, + 480, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For video LLMs, separate visual and text encoders will be adopted first to encode the raw inputs into a hidden space, in the format of tokens. Specifically, input video will be encoded into video tokens $F_V \\in \\mathbb{R}^{(T \\times N_v) \\times C}$ , where $T$ is the", + "bbox": [ + 89, + 839, + 480, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0d9d250db552b66f1160a86c8e187ab1a78bef520a04afaa2aa64af1439a6337.jpg", + "image_caption": [ + "Figure 2. Conceptual demonstration of the $\\mathrm{D}^2\\mathrm{VLM}$ framework." + ], + "image_footnote": [], + "bbox": [ + 524, + 87, + 893, + 189 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/dd96c16028af816f3df21c01c25c7c45fc72e42131c65b35c51d2c5e7652841c.jpg", + "image_caption": [ + "Figure 3. The visual semantic capture process of token." + ], + "image_footnote": [], + "bbox": [ + 553, + 223, + 864, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "number of frames, $N_{v}$ is the number of frame-level visual tokens, and $C$ is the number of channels. Without loss of generality, here we set $N_{v} = 1$ (i.e., $F_{V}\\in \\mathbb{R}^{T\\times C}$ , one token per frame) for illustration purposes. Similarly, input question will be encoded into text tokens $F_{Q}\\in \\mathbb{R}^{N_{t}\\times C}$ , where $N_{t}$ is the number of tokens. Then, both $F_{V}$ and $F_{Q}$ will feed into the video-aware LLM decoder, where the response will be generated in an autoregressive manner.", + "bbox": [ + 511, + 364, + 905, + 486 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. $\\mathbf{D}^2\\mathbf{VLM}$", + "text_level": 1, + "bbox": [ + 511, + 498, + 612, + 512 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We notice that temporal event-level grounding and textual response are two primary tasks in temporally grounded video understanding tasks, where they exhibit distinct characteristics yet maintain strong logical dependencies: Temporal grounding focuses on precisely locating temporal events (evidence) that support answering, while textual response emphasizes accurate evidence interpretation and generating coherent textual answers. However, existing methods typically handle these two tasks in a coupled manner, with two major limitations: (1) Various special tokens are designed for temporal grounding, but are mixed with text token generation without a clear logical structure, leading to coupled learning objectives. (2) More importantly, these special tokens mainly focus on timestamp representation to output timestamp, lacking the explicit capture of visual semantics from grounded events. In contrast, we argue that such visual semantics can actually serve as crucial contexts for subsequent textual answer generation, especially under the next-token prediction paradigm.", + "bbox": [ + 511, + 521, + 906, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Based on this insight, we propose $\\mathrm{D}^2\\mathrm{VLM}$ , where our core motivation is to decouple the tied learning objective while also preserving and even strengthening the inherent dependency between these two tasks. An overall conceptual demonstration is shown in Fig. 2. Specifically, encoded video tokens $F_{V}$ and question textual tokens $F_{Q}$ are fed", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20685", + "bbox": [ + 478, + 949, + 517, + 960 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "into the LLM decoder, which will generate response tokens autoregressively. Here, we enforce the video LLM to first conduct a pure grounding task before providing the actual answer, where the grounded events are expected to serve as essential evidence for the subsequent response. Then, the model will give the actual answer in an interleaved text-evidence manner which will be explained later.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Evidence token for grounding and more. Technically, we introduce evidence token $<\\mathrm{evi}>$ , a special token type dedicated to temporal grounding. Unlike existing designs that mainly focus on category and timestamp representation, the proposed $<\\mathrm{evi}>$ token not only identifies the temporal location of the grounded event but also emphasizes the explicit capture of event-level visual semantics. Specifically, as shown in Fig. 3, once the video LLM generates a token classified as an $<\\mathrm{evi}>$ token, we first calculate the similarity between the $<\\mathrm{evi}>$ token and each frame-level LLM-processed video token $\\tilde{F}_V \\in \\mathbb{R}^{T \\times C}$ (shares the same dimension as the input $F_V$ ). Video tokens with high similarity to the $<\\mathrm{evi}>$ token are then regarded as salient tokens. We aggregate the visual semantics from these salient tokens into the $<\\mathrm{evi}>$ token with simple implementations (i.e., average pool the salient tokens and then add it to the $<\\mathrm{evi}>$ token). Such an operation explicitly endows the $<\\mathrm{evi}>$ token with rich visual semantics of the corresponding grounded event, enabling it to truly serve as solid evidence for the subsequent response based on it. Experiments in Sec. 6 demonstrate that both the design of event-level modeling and visual semantic capture are crucial for performance improvement.", + "bbox": [ + 91, + 200, + 483, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Interleaved text-evi token response. After the pure grounding stage, the generation objective switches to answer the input question based on the grounded evidence. The answer is typically composed of textual response together with temporal grounding information (e.g., \"You capture the dog at 10s-15s and 50s-60s.\"). Here, we formulate the response objective as an interleaved text-evidence token generation process, where tokens are generated again to represent the required temporal output information in a manner akin to evidence referencing, rather than using standard textual tokens to denote timestamps only. The exact timestamps for answering can be directly obtained from the similarity calculation between the token and the frame-level video tokens $\\tilde{F}_V$ , using the frame indices of the salient video tokens (Fig. 3). To make the model aware of such transition between tasks, we introduce another special token, , to indicate the end of the evidence grounding stage and encourage the model to begin generating the interleaved text-evidence answer.", + "bbox": [ + 89, + 550, + 482, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Ideally, the generated token during this stage should also match the one in the evidence grounding phase (i.e., the final response should be consistent with the previous grounded one). We introduce an explicit constraint to", + "bbox": [ + 89, + 839, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "enhance this consistency:", + "bbox": [ + 513, + 92, + 684, + 106 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {c o n s} = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\left| F _ {< e v i > _ {k}} ^ {S _ {1}} - F _ {< e v i > _ {k}} ^ {S _ {2}} \\right|, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 108, + 903, + 148 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $S_{1}, S_{2}$ denote the pure evidence grounding stage and interleaved text-evil response stage, respectively. $k$ is the $k^{th} < \\mathrm{evi} >$ token in each stage. The $K < \\mathrm{evi} >$ tokens are generated in temporal order, each aligned with a distinct interval corresponding to a ground-truth event. Such a constraint enforces strong consistency across stages, preserving logical coherence and strengthening the alignment between evidence grounding and answer generation.", + "bbox": [ + 511, + 152, + 903, + 273 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Loss function. We adopt the following components to supervise network learning:", + "bbox": [ + 511, + 273, + 903, + 303 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {s f t} + L _ {g n d} + L _ {c o n s}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 309, + 903, + 327 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $L_{sft}$ is the standard token classification loss for supervised fine-tuning [22, 24, 38], $L_{gnd}$ is the average grounding loss for generated tokens across both stages, with each formulated as:", + "bbox": [ + 511, + 334, + 903, + 393 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {g n d} ^ {< e v i >} = \\frac {1}{T} \\sum_ {t = 1} ^ {T} B C E \\left(y ^ {t}, s i m ^ {t}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 396, + 903, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where BCE is the binary cross-entropy loss, $y^{t}$ is the frame-level ground-truth (i.e., 1/0 for foreground/background), and $sim^{t}$ is the normalized dot-product similarity between token and frame-level video token feature $\\tilde{F}_{V}^{t}$ .", + "bbox": [ + 511, + 441, + 903, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Factorized Preference Optimization", + "text_level": 1, + "bbox": [ + 511, + 513, + 836, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To further enhance the model learning on both temporal event grounding and textual response, we introduce a novel factorized preference optimization (FPO) algorithm. Unlike standard preference optimization, FPO explicitly incorporates probabilistic temporal grounding modeling into the optimization objective, enabling explicit preference learning on both temporal grounding and textual response. Here, we start with a macroscopic formulation that captures the essence of preference optimization and gradually elaborate on our new design.", + "bbox": [ + 511, + 540, + 905, + 690 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by the existing efficient preference learning algorithm DPO [30], we adopt a similar principle: encouraging the model to assign a higher probability of generating a good response than generating a bad response, which can be defined as:", + "bbox": [ + 511, + 691, + 905, + 766 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL = - \\left[ \\log \\sigma (\\log \\pi (R) - \\log \\pi (\\bar {R})) \\right], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 773, + 903, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pi (R)$ and $\\pi (\\bar{R})$ are the joint generation probabilities of preferred and dis-preferred response, respectively. Under the autoregressive paradigm, such joint generation log-probability can be further detailed as:", + "bbox": [ + 511, + 801, + 903, + 862 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\log \\pi (R) = \\sum_ {R _ {i} \\in R} \\log p \\left(R _ {i} \\mid V, Q, R _ {1: i - 1}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 869, + 903, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20686", + "bbox": [ + 478, + 949, + 517, + 960 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/26907527ed2703e7b13e7315c1451f2ce4b629306d7d475f84710ade8bdebc56.jpg", + "image_caption": [ + "Figure 4. The proposed data synthesis pipeline, where factorized perturbation is imposed to form the dispreferred data." + ], + "image_footnote": [], + "bbox": [ + 133, + 88, + 468, + 271 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/873acf6b89d7d689cad4d6c1ffb571055ddd3c59b1029f10bf5a1069777d865b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 475, + 88, + 864, + 271 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $R_{i}$ is the $i^{th}$ token of the whole model response $R$ . $V$ and $Q$ are the input video and question, respectively.", + "bbox": [ + 89, + 318, + 480, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "One of our goals is to perform preference optimization not only for textual responses but also with an explicit emphasis on temporal grounding. However, achieving such multi-factor preference learning presents two key challenges: (1) The lack of suitable preference data that explicitly considers temporal grounding while maintaining a factorized structure. We will elaborate more on this issue and introduce our newly proposed synthetic dataset as a solution in Sec. 5. (2) The challenge on building suitable objectives for explicit preference optimization on temporal grounding, particularly in quantifying the model's grounding probability (i.e., measuring the likelihood that the model will ground a specific temporal event). Many existing works introduce special tokens and regression heads to enhance the grounding ability of video LLMs. Despite their effectiveness, this deviates from the standard discrete text token prediction paradigm, making it non-trivial to model grounding likelihood for preference optimization.", + "bbox": [ + 88, + 349, + 482, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this work, we notice that our designed $\\mathrm{D}^2\\mathrm{VLM}$ framework can naturally model such event grounding probability, in a different and explicit way compared with the standard token prediction probability. Specifically, as grounding is achieved by frame-level feature similarity calculation between the $<\\mathrm{evi}>$ token and each video frame, we directly use such similarity as the frame-level grounding probability and multiply these probabilities across frames to obtain the interval-level grounding probability. Formally, for each generated $<\\mathrm{evi}>_k$ token during model response, its corresponding grounding probability $p_g([s_k,e_k])$ for a given time interval $[s_k,e_k]$ ( $s_k$ , $e_k$ denote the starting and ending time index respectively) can be formulated as:", + "bbox": [ + 89, + 619, + 482, + 816 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {g} ([ s _ {k}, e _ {k} ]) = \\prod_ {t = 1} ^ {T} \\left\\{ \\begin{array}{l l} s i m _ {k} ^ {t} & \\text {i f} s _ {k} \\leq t \\leq e _ {k} \\\\ 1 - s i m _ {k} ^ {t} & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 823, + 482, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $T$ is the total number of input video frames, $sim_k^t$ is the normalized dot-product similarity between $\\langle \\mathrm{evi} \\rangle_k$ and", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "frame-level video token feature $\\tilde{F}_V^t$ . Based on such event grounding modeling, the log-probability formulation within our proposed factorized preference optimization is given as:", + "bbox": [ + 511, + 318, + 906, + 364 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\log \\pi (R) = \\sum_ {R _ {i} \\in R} \\log p \\left(R _ {i} \\mid V, Q, R _ {1: i - 1}\\right) \\\\ + \\sum_ {R _ {k} = < e v i >} \\log p _ {g} \\left(\\left[ s _ {k}, e _ {k} \\right] \\mid V, Q, R _ {1: k - 1}\\right), \\end{array} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 382, + 903, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the first term corresponds to the standard textual token prediction modeling, while the second term represents the proposed explicit modeling on temporal event grounding. By substituting Equation (7) into Equation (4), we obtain the final formulation of our Factorized Preference Optimization (FPO). Next, we explain the process of constructing multi-factor preferred $(R)$ and dis-preferred $(\\bar{R})$ data pairs, which serve as the training data for FPO.", + "bbox": [ + 511, + 455, + 906, + 577 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Factorized Preference Data Synthesis", + "text_level": 1, + "bbox": [ + 511, + 588, + 849, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As previously mentioned, one key challenge is the lack of suitable preference data that captures multi-factor preferences—not only for textual responses but also for explicit temporal grounding. Particularly, in the video domain, preference data is typically constructed by providing a video LLM with either complete or degraded video inputs (e.g., blurred or cropped images/videos) and treating its responses as preferred or dispreferred, respectively [18]. The underlying assumption is that the model will produce inferior responses when given degraded or incomplete visual inputs. However, existing works using this approach often face some major issues: (1) The reliability of the generated preference data is relatively less controlled. Specifically, the generated preferred sample may not actually be the preferred one, and the generated dispreferred data can sometimes be as good as or even better than the preferred data [18]. As a result, post-filtering is usually required, which involves additional models and ad-hoc human interventions, increasing costs but still cannot fully ensure", + "bbox": [ + 511, + 613, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20687", + "bbox": [ + 478, + 949, + 517, + 960 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "quality. (2) This lack of control also makes it difficult to properly factorize the generated dispreferred data, including identifying the source or the reason why it is considered dispreferred. (3) Last but not least, existing video preference data mainly focuses on generating responses in pure textual format (e.g., video caption) [18, 46], while generating preference response that explicitly contains temporal grounding information (e.g., temporal event position) remains under-explored. Overall, these issues make factorized preference optimization difficult to achieve.", + "bbox": [ + 89, + 90, + 480, + 241 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To address the aforementioned limitations, we propose an automatic factorized preference data synthesis framework, which introduces factorized perturbations on the original response at the sub-video event level. As shown in Fig. 4 (a), the overall process consists of the following steps:", + "bbox": [ + 89, + 244, + 480, + 320 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Random selection of factors and sub-video events. For each given sample (video-question-answer pair), we first randomly select a factor that determines the exact type of the perturbation. As shown in Fig. 4 (b), the factors are broadly categorized into two types and each can be further divided into multiple sub-types: (1) Temporal grounding: This type of perturbation aims to reveal/simulate model response inaccuracies related to temporal event grounding. It can be further divided into sub-types, including temporal localization shift, randomly adding/deleting grounded events (corresponding to false positives and missed detections), and merging multiple events into one (corresponding to a lack of fine-grained distinction in event boundaries). (2) Textual response: This type of perturbation modifies the semantic correctness of the textual response. It includes sub-types such as distorting key information, which disrupts critical content, and repeating responses, a common failure mode observed in video LLMs.", + "bbox": [ + 89, + 321, + 482, + 593 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Next, we randomly select some sub-video events to apply factorized perturbation. A sub-video event refers to a semantically meaningful segment of the video that can be determined based on the given sample (e.g., its question or ground-truth answer). For instance, in a dense video captioning task, an event naturally corresponds to a specific captioned interval, while in an action localization task, an event can align with a single action instance. Without loss of generality, we take the dense video caption task as an illustration example (as shown in Fig. 4 (c)), as this task emphasizes both temporal grounding and textual description. Overall, such an event-level design aligns naturally with the inherent video structure, ensuring that perturbations are introduced in a way that respects the video's temporal and semantic coherence. Moreover, it enables fine-grained perturbation, allowing controlled modifications at the event level rather than applying coarse, less precisely controlled noise to the entire response sequence.", + "bbox": [ + 89, + 595, + 482, + 867 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Injecting factorized perturbations. Once the factors and sub-video events are determined, localized perturbation is", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "introduced based on the selected factors and sub-video events. If a temporal grounding factor is selected, we modify the timestamps of the selected events based on the determined sub-factor (e.g., shift the interval, delete the interval, etc.). For the factor of textual response, for each selected event, we prompt off-the-shelf LLM (e.g., Qwen [41, 42]) to generate a distorted response based on the original correct event-level response. If the \"repeat\" sub-factor is selected, we can easily achieve this by setting the response of the current event to its previous one to simulate such a response repeating phenomenon. A synthesized example is shown in Fig. 4 (c). After the perturbation synthesis, we combine the newly obtained noisy response with the original one to form the preference sample pair for factorized preference optimization. In practice, our preference data is synthesized based on the E.T. Instruct 164K dataset [23].", + "bbox": [ + 511, + 90, + 903, + 332 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Discussion. Compared to existing preference data synthesis approaches, our framework ensures that the generated dispreferred response contains specific, well-factorized noise rather than arbitrary, unpredictable degradation, and thus we can better identify the noise source without manual costs. To our knowledge, we are also the first to emphasize explicit temporal grounding in preference data, and moreover, in a multi-factorized manner. The overall design makes our factorized preference optimization in Sec. 4 feasible.", + "bbox": [ + 511, + 340, + 903, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 513, + 513, + 643, + 530 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 513, + 545, + 702, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. Following the practice of [23], we adopt the pre-trained ViT-G/14 from EVA-CLIP [8, 34], followed by a Q-Former-like feature compressor [6, 16, 23], and use Phi-3-Mini-3.8B [1] as the base LLM. Also following [23], our model is initialized from the stage-2 model of [23] and [19], and we also only fine-tune our model on the E.T. Instruct 164K dataset [23] using LoRA [12]. We conduct training on $4 \\times$ NVIDIA H100 GPUs within 1 day. More details are provided in the supplementary material.", + "bbox": [ + 511, + 575, + 903, + 712 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation tasks and datasets. We evaluate our method across various datasets and tasks. (1) E.T. Bench Grounding [23]. It consists of 5 different grounding-related tasks (e.g., moment retrieval, action localization, summarization/highlight detection, etc.) for comprehensive evaluation. (2) E.T. Bench Dense Captioning [23]. It is composed of tasks like dense video caption and step localization, where both key event grounding and textual description are required. Besides, we also conduct experiments on two widely used individual datasets, (3) Charades-STA [9] and (4) YouCook2 [47], for independent evaluation on moment retrieval and dense video captioning, respectively.", + "bbox": [ + 511, + 719, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20688", + "bbox": [ + 478, + 949, + 517, + 960 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/006892a719e3c860c8f997e69a7d55e25ffff3d3762246f740e71314cce4a1a8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodYearE.T. Bench GroundingE.T. Bench Dense Captioning
TVG\\( _{F1} \\)EPM\\( _{F1} \\)TAL\\( _{F1} \\)EVS\\( _{F1} \\)VHD\\( _{F1} \\)Avg\\( _{F1} \\)DVC\\( _{F1} \\)DVC\\( _{Sim} \\)SLC\\( _{F1} \\)SLC\\( _{Sim} \\)Avg\\( _{F1} \\)Avg\\( _{Sim} \\)
Video-ChatGPT-7B [26]ACL'247.01.315.18.428.812.18.811.35.710.27.310.8
Video-LLaVA-7B [21]EMNLP'247.01.915.00.328.910.628.015.00.98.314.511.7
LLaMA-VID-7B [19]ECCV'245.51.28.01.430.09.227.112.65.211.116.211.9
Video-LLaMA-2-7B [5]arXiv'240.10.00.00.01.50.30.614.50.015.20.314.9
PLLAVA [40]arXiv'246.91.15.70.328.98.613.310.69.711.811.511.2
VTimeLLM-7B [13]CVPR'247.61.918.215.928.914.512.413.18.76.410.69.8
TimeChat-7B [31]CVPR'2426.23.910.129.140.522.016.612.55.69.211.110.9
LITA-13B [14]ECCV'2422.24.618.029.723.919.739.717.221.012.230.414.7
VTG-LLM-7B [10]AAAI'2515.93.714.426.848.221.840.218.620.814.430.516.5
Qwen2.5-VL-7B* [2]arXiv'2546.69.332.219.968.635.339.021.425.515.132.218.2
E.T.Chat-3.8B [23]NeurIPS'2438.610.230.825.462.533.538.419.724.414.631.417.2
D²VLM-3.8B (Ours)ICCV'2560.214.433.435.268.242.348.325.526.718.137.521.8
", + "bbox": [ + 94, + 88, + 903, + 281 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ca89a56b151c617137533935c0cadb863b3eb967270ec20c0b1652098d5a6089.jpg", + "table_caption": [ + "Table 1. Performance comparison on E.T. Bench-Grounding dataset (5 sub tasks: TVG: Temporal Video Grounding; EPM: Episodic Memory; TAL: Temporal Action Localization; EVS: Extractive Video Summarization; VHD: Video Highlight Detection) and E.T. Bench-Dense Captioning dataset (2 sub tasks: DVC: Dense Video Captioning; SLC: Step Localization and Captioning). \\*: Implemented by us, with more details in the supplementary material." + ], + "table_footnote": [], + "table_body": "
MethodYearR@1(IoU=0.5)R@1(IoU=0.7)
TimeChat-7B [31]CVPR’2432.213.4
VTimeLLM-7B [13]CVPR’2427.511.4
VTimeLLM-13B [13]CVPR’2434.314.7
Momentor-7B [29]ICML’2426.611.6
HawkEye-7B [37]arXiv’2431.414.5
VTG-LLM-7B [10]AAAI’2533.815.7
TRACE-7B [11]ICLR’2540.319.4
VideoChat-T-7B [43]ICLR’2548.724.0
E.T.Chat-3.8B [23]NeurIPS’2445.920.0
D²VLM-3.8B (Ours)ICCV’2550.326.0
", + "bbox": [ + 94, + 364, + 480, + 502 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c06aab560e4321fc74ac473c8957c8f989838a8f0f3284849beb37c29c414850.jpg", + "table_caption": [ + "Table 2. Performance comparison on Charades-STA." + ], + "table_footnote": [], + "table_body": "
MethodYearF1CIDErSODA_c
TimeChat-7B [31]CVPR’2412.63.41.2
VTG-LLM-7B [10]AAAI’2517.55.01.5
TRACE-7B [11]ICLR’2522.48.12.2
D²VLM-3.8B (Ours)ICCV’2526.410.63.2
", + "bbox": [ + 94, + 536, + 480, + 618 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Performance comparison on YouCook2.", + "bbox": [ + 140, + 622, + 433, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2. Main Benchmarking Results", + "text_level": 1, + "bbox": [ + 89, + 664, + 346, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "E.T. Bench grounding. The results are shown in the left part of Tab. 1, where we compare our method with the recently published SOTA methods. It can be observed that the proposed $\\mathrm{D}^2\\mathrm{VLM}$ significantly outperforms recently published SOTA methods across various tasks, achieving an average F1 score improvement of at least $7.0\\%$ . Meanwhile, our model is relatively lightweight (3.8B parameters) compared to most state-of-the-art methods (ranging from 3.8B to 13B, with the majority being 7B parameters). This further demonstrates the superiority of our proposed method.", + "bbox": [ + 89, + 686, + 482, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "E.T. Bench dense captioning. The performance comparison is shown in the right part of Tab. 1. It can be seen that our method also shows noticeable superiority in dense captioning-related tasks, where it outperforms recent SOTA", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "methods in both temporal event grounding (at least $5.3\\%$ on average F1 score) and textual description (at least $3.6\\%$ on average Sim score). This essentially demonstrates the generalizability of our method across various scenarios.", + "bbox": [ + 511, + 367, + 903, + 428 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Charades-STA. From Tab. 2, it can be seen that our method still outperforms recent SOTA methods by large margins. Specifically, our 3.8B model outperforms counterpart [23] with the same base model by $4.4\\%$ on $\\mathrm{R}@\\mathrm{1(IoU = 0.5)}$ , and outperforms methods published in 2025 with 7B model size [10, 11, 31] by at least $1.6\\%$ on $\\mathrm{R}@\\mathrm{1(IoU = 0.5)}$ . This further demonstrates the superiority of our method in both effectiveness and efficiency.", + "bbox": [ + 511, + 429, + 906, + 549 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "YouCook2. As shown in Tab. 3, our method achieves state-of-the-art performance compared to recent SOTA approaches. Specifically, it outperforms them by at least $4\\%$ in temporal event grounding (F1) while also demonstrating superior textual description capacity, with gains of at least $2.5\\%$ on CIDEr and $1.0\\%$ on SODA.c. This further highlights the superiority of our approach in both temporal grounding and text generation.", + "bbox": [ + 511, + 550, + 905, + 671 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 681, + 679, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here we conduct ablation studies on E.T. Bench grounding and E.T. Bench Dense Captioning datasets to analyze the effect of the proposed components, in both temporal event-level grounding and textual description. For each dataset (i.e., grounding, dense captioning), we report the average performance across multiple sub-tasks. In the following, we first analyze the component effect of the newly designed generation objective and evidence token, where we do not take the proposed FPO into account. Finally, we analyze the effect of the proposed FPO.", + "bbox": [ + 511, + 703, + 906, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generation objective of $\\mathbf{D}^2\\mathbf{VLM}$ . Here, we analyze the effectiveness of our newly designed generation objective within $\\mathbf{D}^2\\mathbf{VLM}$ . From Tab. 4, it can be seen that: (1) Row-1", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20689", + "bbox": [ + 478, + 949, + 519, + 960 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/24e0b92329d9d6b81d2631416d6bc21edc0174ba6231270e0027c4508b68a2de.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Decomposed ObjectiveInterleaved Text-avi GenerationConsistency ConstraintGrounding AvgF1Dense Captioning AvgF1AvgSim
21.214.311.3
✓28.923.116.0
✓✓35.634.319.8
✓✓✓39.535.021.2
", + "bbox": [ + 94, + 88, + 594, + 176 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a758ac3a911bf9ed71da82babdc7a18bbf1007558bcb1ac06bfd344804a245af.jpg", + "table_caption": [ + "Table 4. Component effect of the designed generation objective." + ], + "table_footnote": [], + "table_body": "
FPOGrounding AvgF1Dense Captioning AvgF1AvgSim
×39.535.021.2
✓42.337.521.8
", + "bbox": [ + 617, + 88, + 861, + 176 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/65bb30731d8ae3e9dda1f838f9709e1c7ae9cdf7748fc95beb06b38e245e5848.jpg", + "table_caption": [ + "Table 5. Effect of the proposed FPO." + ], + "table_footnote": [], + "table_body": "
MethodGrounding \nAvgF1Dense \nAvgF1Captioning \nAvgSim
w/o event-level modeling26.133.416.2
w/o visual semantic capture37.127.517.7
Full design39.535.021.2
", + "bbox": [ + 112, + 219, + 459, + 303 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Component effect within evidence token design.", + "bbox": [ + 114, + 305, + 455, + 320 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "v.s. Row-2: Decompose the coupled grounding-textual response objective into grounding (by our designed token) then answer from grounded event can substantially improve the performance in both grounding (i.e., $7.7\\%$ averaged F1 on grounding tasks and $8.8\\%$ averaged F1 on dense captioning tasks) and textual description (i.e., $4.7\\%$ averaged sentence similarity on dense captioning tasks). (2) Row-2 v.s. Row-3: When we shift the answering stage from pure textual token generation into interleaved text-evi token generation, the performance can be further enhanced by large margins (at least $6.7\\%$ on grounding and $3.8\\%$ on textual description). The reason is that under such cases, the overall generation objective can be interpreted as grounding then answer with evidence referencing, where such referencing strengthens the relationship between event grounding and textual answering. (3) Row-3 v.s. Row-4: The introduced consistency constraint further boosts the performance by a noticeable margin, as it explicitly enhances the consistency between the two stages. Overall, the results validate the effectiveness of our design.", + "bbox": [ + 89, + 353, + 482, + 655 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Design of evidence token. Here we analyze the impact of the core components within the token design (i.e., the event-level modeling and the explicit visual semantic capture). The results are shown in Tab. 6, where w/o event-level modeling means each token only captures the information of individual frame (a common design in existing works [14, 23, 29]), rather than an interval-level event (e.g., 10s-15s). w/o visual semantic capture indicates visual semantics from the video tokens is not explicitly integrated into the token through the operation illustrated in Sec. 3.2 (also a common design in most of the existing works [4, 14, 29]). From Tab. 6, it can be summarized that: (1) The performance drops significantly without event-level modeling. We argue that this is because compared to individual timestamp-level modeling, event-level modeling more naturally aligns with the characteristics of", + "bbox": [ + 89, + 657, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "event-level evidence grounding and can capture more robust event-level semantics in a global perspective. (2) Explicit visual semantic capture is also crucial for enhancing model effectiveness, as it enables tokens to truly serve as solid context for subsequent answer generation, which also fits well with the autoregressive generation process. Notably, compared with grounding-only tasks, such an effect is more pronounced on dense captioning tasks, as textual description generation can truly benefit from the captured event-level visual semantics.", + "bbox": [ + 511, + 223, + 903, + 372 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of factorized preference optimization. Tab. 5 demonstrates that the proposed FPO algorithm consistently enhances performance on both temporal grounding and captioning-related tasks, with a more significant improvement in temporal grounding. This indeed validates the effectiveness of our factorized optimization objective design.", + "bbox": [ + 511, + 375, + 906, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion and Limitations", + "text_level": 1, + "bbox": [ + 511, + 481, + 769, + 496 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we uncover some key limitations of existing temporal grounded video-language models: (1) the use of coupled learning objective for temporal grounding and textual response generation, and (2) the overlooking of inherent dependency between these two tasks, especially the event-level visual semantics from the intermediate grounding result. We tackle this from a factorized learning perspective, where we propose $\\mathrm{D}^2\\mathrm{VLM}$ that decomposes the generation objective into a \"grounding then answering with evidence referencing\" paradigm, and introduce evidence tokens to emphasize explicit visual semantic capture beyond timestamp representation. Furthermore, we design a factorized preference learning algorithm, coupled with a factorized synthetic dataset, to enhance the learning of both tasks. Extensive experiments demonstrate the clear superiority of our approach across various scenarios.", + "bbox": [ + 511, + 507, + 906, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Despite achieving new state-of-the-art results, our approach still has room for improvement (e.g., F1 scores of only $14.4\\%$ on the episodic memory task and $26.4\\%$ on the YouCook2 dense video captioning task). Besides, while our factorized data generation offers good factorized controllability and directly contributes to factorized optimization, it focuses only on generating negative (dis-preferred) samples. However, generating more diverse positive samples to further enrich the preference data is also meaningful, and we leave this for future work.", + "bbox": [ + 511, + 750, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20690", + "bbox": [ + 478, + 949, + 519, + 960 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 236, + 187, + 253 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Marah Abdin, Sam Ade Jacobs, Ammar Ahmad Awan, Jyoti Aneja, Ahmed Awadallah, Hany Awadalla, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Harkirat Behl, et al. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219, 2024. 6", + "[2] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. 7", + "[3] Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, et al. Constitutional ai: Harmlessness from ai feedback. arXiv preprint arXiv:2212.08073, 2022. 3", + "[4] Qirui Chen, Shangzhe Di, and Weidi Xie. Grounded multi-hop videoqa in long-form egocentric videos. arXiv preprint arXiv:2408.14469, 2024. 2, 3, 8", + "[5] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 7", + "[6] Wenliang Dai, Junnan Li, DONGXU LI, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale N Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. In Advances in Neural Information Processing Systems, pages 49250-49267. Curran Associates, Inc., 2023. 6", + "[7] Shangzhe Di and Weidi Xie. Grounded question-answering in long egocentric videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12934–12943, 2024. 1", + "[8] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19358-19369, 2023. 6", + "[9] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In Proceedings of the IEEE international conference on computer vision, pages 5267-5275, 2017. 1, 6", + "[10] Yongxin Guo, Jingyu Liu, Mingda Li, Dingxin Cheng, Xiaoying Tang, Dianbo Sui, Qingbin Liu, Xi Chen, and Kevin Zhao. Vtg-llm: Integrating timestamp knowledge into video" + ], + "bbox": [ + 93, + 261, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "This research is supported by the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No: AISG3-RP-2022-030). We would like to acknowledge that computational work involved in this research work is partially supported by NUS IT's Research Computing group under grant number NUSREC-HPC-00001.", + "llms for enhanced video temporal grounding. arXiv preprint arXiv:2405.13382, 2024. 1, 2, 3, 7", + "[11] Yongxin Guo, Jingyu Liu, Mingda Li, Xiaoying Tang, Qingbin Liu, and Xi Chen. Trace: Temporal grounding video llm via causal event modeling. In International Conference on Learning Representations, 2025. 2, 3, 7", + "[12] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6", + "[13] Bin Huang, Xin Wang, Hong Chen, Zihan Song, and Wenwu Zhu. Vtimellm: Empower llm to grasp video moments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14271-14280, 2024. 2, 3, 7", + "[14] De-An Huang, Shijia Liao, Subhashree Radhakrishnan, Hongxu Yin, Pavlo Molchanov, Zhiding Yu, and Jan Kautz. Lita: Language instructed temporal-localization assistant. In European Conference on Computer Vision, pages 202-218. Springer, 2024. 2, 3, 7, 8", + "[15] Harrison Lee, Samrat Phatale, Hassan Mansoor, Thomas Mesnard, Johan Ferret, Kellie Lu, Colton Bishop, Ethan Hall, Victor Carbune, Abhinav Rastogi, et al. Rlaif vs. rlhf: Scaling reinforcement learning from human feedback with ai feedback. arXiv preprint arXiv:2309.00267, 2023. 3", + "[16] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 6", + "[17] Lei Li, Zhihui Xie, Mukai Li, Shunian Chen, Peiyi Wang, Liang Chen, Yazheng Yang, Benyou Wang, and Lingpeng Kong. Silkie: Preference distillation for large visual language models. arXiv preprint arXiv:2312.10665, 2023. 3", + "[18] Rui Li, Xiaohan Wang, Yuhui Zhang, Zeyu Wang, and Serena Yeung-Levy. Temporal preference optimization for long-form video understanding. arXiv preprint arXiv:2501.13919, 2025. 3, 5, 6", + "[19] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In European Conference on Computer Vision, pages 323–340. Springer, 2024. 6, 7", + "[20] Zhaowei Li, Qi Xu, Dong Zhang, Hang Song, YiQing Cai, Qi Qi, Ran Zhou, Junting Pan, Zefeng Li, Vu Tu, et al. Groundinggpt: Language enhanced multi-modal grounding model. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6657-6678, 2024. 3", + "[21] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 5971–5984, 2024. 1, 2, 7", + "[22] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2023. 4", + "[23] Ye Liu, Zongyang Ma, Zhongang Qi, Yang Wu, Ying Shan, and Chang Wen Chen. E.T. Bench: Towards open-ended" + ], + "bbox": [ + 93, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20691", + "bbox": [ + 478, + 949, + 517, + 960 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "event-level video-language understanding. In The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2024. 1, 2, 3, 6, 7, 8", + "[24] Shayne Longpre, Le Hou, Tu Vu, Albert Webson, Hyung Won Chung, Yi Tay, Denny Zhou, Quoc V Le, Barret Zoph, Jason Wei, et al. The flan collection: Designing data and methods for effective instruction tuning. In International Conference on Machine Learning, pages 22631-22648. PMLR, 2023. 4", + "[25] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12585–12602, 2024. 1, 2", + "[26] Munan Ning, Bin Zhu, Yujia Xie, Bin Lin, Jiaxi Cui, Lu Yuan, Dongdong Chen, and Li Yuan. Video-bench: A comprehensive benchmark and toolkit for evaluating video-based large language models. arXiv preprint arXiv:2311.16103, 2023. 7", + "[27] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 3", + "[28] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 3", + "[29] Long Qian, Juncheng Li, Yu Wu, Yaobo Ye, Hao Fei, TatSeng Chua, Yueting Zhuang, and Siliang Tang. Momentor: Advancing video large language model with fine-grained temporal reasoning. arXiv preprint arXiv:2402.11435, 2024. 3, 7, 8", + "[30] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36:53728-53741, 2023. 3, 4", + "[31] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14313-14323, 2024. 1, 2, 3, 7", + "[32] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 3", + "[33] Nisan Stiannon, Long Ouyang, Jeffrey Wu, Daniel Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford, Dario Amodei, and Paul F Christiano. Learning to summarize with human feedback. Advances in neural information processing systems, 33:3008-3021, 2020. 3", + "[34] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 6" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liang-Yan Gui, Yu-Xiong Wang, Yiming Yang, et al. Aligning large multimodal models with factually augmented rlhf. In Annual Meeting of the Association for Computational Linguistics, 2024. 3", + "[36] Haibo Wang, Zhiyang Xu, Yu Cheng, Shizhe Diao, Yufan Zhou, Yixin Cao, Qifan Wang, Weifeng Ge, and Lifu Huang. Grounded-videollm: Sharpening fine-grained temporal grounding in video large language models. arXiv preprint arXiv:2410.03290, 2024. 3", + "[37] Yueqian Wang, Xiaojun Meng, Jianxin Liang, Yuxuan Wang, Qun Liu, and Dongyan Zhao. Hawkeye: Training video-text llms for grounding text in videos. arXiv preprint arXiv:2403.10228, 2024. 3, 7", + "[38] Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652, 2021. 4", + "[39] Junbin Xiao, Angela Yao, Yicong Li, and Tat-Seng Chua. Can i trust your answer? visually grounded video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13204-13214, 2024. 2", + "[40] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 7", + "[41] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng He, Junyang Lin, Kai Dang, Keming Lu, Keqin Chen, Kexin Yang, Mei Li, Mingfeng Xue, Na Ni, Pei Zhang, Peng Wang, Ru Peng, Rui Men, Ruize Gao, Runji Lin, Shijie Wang, Shuai Bai, Sinan Tan, Tianhang Zhu, Tianhao Li, Tianyu Liu, Wenbin Ge, Xiaodong Deng, Xiaohuan Zhou, Xingzhang Ren, Xinyu Zhang, Xipin Wei, Xuancheng Ren, Yang Fan, Yang Yao, Yichang Zhang, Yu Wan, Yunfei Chu, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, and Zhihao Fan. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 6", + "[42] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, Le Yu, Mei Li, Mingfeng Xue, Pei Zhang, Qin Zhu, Rui Men, Runji Lin, Tianhao Li, Tingyu Xia, Xingzhang Ren, Xuancheng Ren, Yang Fan, Yang Su, Yichang Zhang, Yu Wan, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, and Zihan Qiu. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115, 2024. 6", + "[43] Xiangyu Zeng, Kunchang Li, Chenting Wang, Xinhao Li, Tianxiang Jiang, Ziang Yan, Songze Li, Yansong Shi, Zhengrong Yue, Yi Wang, et al. Timesuite: Improving mllms for long video understanding via grounded tuning. In International Conference on Learning Representations, 2025. 3, 7" + ], + "bbox": [ + 516, + 90, + 905, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20692", + "bbox": [ + 478, + 949, + 517, + 960 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[44] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 1, 2", + "[45] Mengxi Zhang, Wenhao Wu, Yu Lu, Yuxin Song, Kang Rong, Huanjin Yao, Jianbo Zhao, Fanglong Liu, Haocheng Feng, Jingdong Wang, et al. Automated multi-level preference for mllms. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 3", + "[46] Ruohong Zhang, Liangke Gui, Zhiqing Sun, Yihao Feng, Keyang Xu, Yuanhan Zhang, Di Fu, Chunyuan Li, Alexander Hauptmann, Yonatan Bisk, et al. Direct preference optimization of video large multimodal models from language model reward. arXiv preprint arXiv:2404.01258, 2024. 3, 6", + "[47] Luowei Zhou, Chenliang Xu, and Jason Corso. Towards automatic learning of procedures from web instructional videos. In Proceedings of the AAAI Conference on Artificial Intelligence, 2018. 1, 6", + "[48] Daniel M Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. Fine-tuning language models from human preferences, 2020. URL https://arxiv.org/abs, page 14, 1909. 3" + ], + "bbox": [ + 89, + 92, + 482, + 398 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20693", + "bbox": [ + 478, + 950, + 517, + 960 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Factorized Learning for Temporally Grounded Video-Language Models/0391813b-b98c-4912-884e-a194864058af_model.json b/2025/Factorized Learning for Temporally Grounded Video-Language Models/0391813b-b98c-4912-884e-a194864058af_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ce7b1930fbfd401d316bb703c2e5ac60753e0d4f --- /dev/null +++ b/2025/Factorized Learning for Temporally Grounded Video-Language Models/0391813b-b98c-4912-884e-a194864058af_model.json @@ -0,0 +1,2026 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.138, + 0.131, + 0.861, + 0.154 + ], + "angle": 0, + "content": "Factorized Learning for Temporally Grounded Video-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.247, + 0.18, + 0.751, + 0.218 + ], + "angle": 0, + "content": "Wenzheng Zeng, Difei Gao, Mike Zheng Shou†, Hwee Tou Ng† National University of Singapore" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.22, + 0.875, + 0.235 + ], + "angle": 0, + "content": "wenzhengzeng@u.nus.edu, {daniel.difei.gao, mike.zheng.shou}@gmail.com, nght@nus.edu.sg" + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.271, + 0.116, + 0.284 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.28, + 0.418, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.424, + 0.269, + 0.9, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.486, + 0.907, + 0.557 + ], + "angle": 0, + "content": "Figure 1. (a) Performance: Our method consistently outperforms SOTA methods across various tasks (here we draw the maximum performance across methods, detailed in Sec. 6). (b) Model: We propose a new framework \\(\\mathrm{D}^2\\mathrm{VLM}\\), where we decompose the generation objective into a \"grounding then answering with evidence referencing\" paradigm and introduce evidence tokens to emphasize explicit event-level visual semantic capture. (c) Training Algorithm: We introduce Factorized Preference Optimization (FPO) that explicitly addresses both temporal grounding and textual response. A factorized data synthesis approach is also designed to support FPO." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.584, + 0.327, + 0.6 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.616, + 0.486, + 0.889 + ], + "angle": 0, + "content": "Recent video-language models have shown great potential for video understanding, but still struggle with accurate temporal grounding for event-level perception. We observe that two main factors in video understanding (i.e., temporal grounding and textual response) form a logical hierarchy: accurate temporal evidence grounding lays the foundation for reliable textual response. However, existing works typically handle these two tasks in a coupled manner without a clear logical structure, leading to sub-optimal objectives. We address this from a factorized learning perspective. We first propose \\( D^{2} \\) VLM, a framework that decouples the learning of these two tasks while also emphasizing their inherent dependency. We adopt a \"grounding then answering with evidence referencing\" paradigm and introduce evidence tokens for evidence grounding, which emphasize event-level visual semantic capture beyond the focus on timestamp representation in existing works. To further facilitate the learning of these two tasks, we introduce a novel factorized" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.57, + 0.909, + 0.721 + ], + "angle": 0, + "content": "preference optimization (FPO) algorithm. Unlike standard preference optimization, FPO explicitly incorporates probabilistic temporal grounding modeling into the optimization objective, enabling preference learning for both temporal grounding and textual response. We also construct a synthetic dataset to address the lack of suitable datasets for factorized preference learning with explicit temporal grounding. Experiments on various tasks demonstrate the clear advantage of our approach. Our source code is available at https://github.com/nusnlp/d2v1m." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.754, + 0.646, + 0.769 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Recent advances in video-language models, especially those built upon large language models (video LLMs), have enabled remarkable progress in video understanding [10, 21, 23, 25, 31, 44]. Through their flexible video&text-int-text-out nature, video LLMs demonstrate great potential as general-purpose solvers, unifying various tasks (e.g., temporal grounding [9], dense captioning [47], and question answering [7]) into one generic \"video question answer-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.116, + 0.899, + 0.246, + 0.912 + ], + "angle": 0, + "content": "†Corresponding authors" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.95, + 0.519, + 0.962 + ], + "angle": 0, + "content": "20683" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "ing\" framework. Despite their effectiveness, existing video LLMs still struggle with accurate temporal grounding for event perception and localization [10, 11, 23, 31, 31], a crucial capability not only for grounding tasks themselves but also for relevant tasks that need textual answering." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.171, + 0.483, + 0.488 + ], + "angle": 0, + "content": "We notice that for video understanding, temporal event-level grounding and textual response are two primary tasks that exhibit distinct characteristics yet maintain strong logical dependencies. Specifically, temporal grounding focuses on precisely locating temporal events (evidence) that support answering, while textual response emphasizes accurate interpretation from the grounded evidence and generating coherent textual answers. However, existing methods [4, 13, 14, 23, 31] typically handle these two tasks in a coupled manner, with two major limitations: (1) Various special tokens are designed for temporal grounding [4, 14, 23], but their generation is mixed with text token generation without a clear logical structure, leading to coupled learning objectives. (2) More importantly, these special tokens mainly focus on timestamp representation for precise timestamp output, lacking the explicit capture of visual semantics from grounded events. In contrast, we argue that such event-level visual semantics should not be overlooked, as it could inherently serve as crucial contexts for subsequent textual answer generation, especially under the next-token prediction paradigm." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.491, + 0.483, + 0.687 + ], + "angle": 0, + "content": "Based on the aforementioned observations, we propose to address this from a factorized learning perspective. We first introduce a new framework \\( \\mathrm{D}^2\\mathrm{VLM} \\) that decouples the learning of temporal evidence grounding and textual answering, while preserving and even strengthening their inherent dependency. Specifically, as shown in Fig. 1 (b), we decompose the model response into two sequential stages: (1) pure temporal grounding that aims to localize and capture essential visual evidence for response, followed by (2) interleaved text-evidence answer generation, where both the textual answer and temporal information are produced in an evidence-referencing manner to establish consistency with the previously grounded evidence." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Technically, we introduce evidence tokens, a special token type dedicated to temporal evidence grounding. Different from existing designs of grounding tokens that focus on its special category and timestamp representation, our evidence tokens not only aim to determine the temporal location of the grounded event, but also emphasize the capture of event-level visual semantics, which serves as crucial context for subsequent answer generation. During the subsequent interleaved text-evidence generation, the evidence-referencing process is achieved by generating evidence tokens that align with those from the previous grounding stage. This ensures that the output information in the final response remains consistent with the initially grounded evidence while reinforcing logical coherence across stages." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.228 + ], + "angle": 0, + "content": "Besides providing decoupled and clearer task objectives, our design naturally fits well with the teacher-forcing autoregressive training paradigm, as subsequent textual response generation is conditioned on the correctly grounded evidence, enabling a learning shortcut for more stable training. Our experiments demonstrate that both the designed sequence generation objective and event-level visual semantic capture are essential for performance improvement, offering valuable insights for future model design." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.229, + 0.905, + 0.59 + ], + "angle": 0, + "content": "To further facilitate the learning of these two tasks, we propose a novel Factorized Preference Optimization (FPO) algorithm. Unlike standard preference optimization, FPO incorporates probabilistic temporal grounding modeling into the optimization objective, enabling explicit preference learning for temporal grounding in addition to standard textual preference. Meanwhile, another obstacle is that existing video preference datasets do not account for such temporal grounding aspects, making it infeasible to directly apply our proposed FPO algorithm. To overcome this limitation, we construct a synthetic dataset by introducing factorized perturbations into the original preferred response sequence. These perturbations are applied at the sub-video event level, considering two main factors: temporal grounding and textual response, along with multiple possible sub-factors. This approach ensures a structured and controllable noise generation process, where the cause and type of noise are precisely known without any manual annotation. As a result, our data synthesis process is both reliable and scalable, making it possible to effectively conduct the proposed factorized preference optimization. To the best of our knowledge, we are the first to use preference learning involving explicit temporal grounding, and moreover, in a factorized manner, as illustrated in Fig. 1 (c)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.592, + 0.905, + 0.669 + ], + "angle": 0, + "content": "We conduct extensive experiments to demonstrate the superiority of our proposed framework. An intuitive comparison is shown in Fig. 1 (a), where our 3.8B model consistently outperforms existing SOTA methods (typically ranging from 3.8B to 13B in model size) across various tasks." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.685, + 0.655, + 0.7 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.71, + 0.649, + 0.725 + ], + "angle": 0, + "content": "2.1. Video LLMs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.733, + 0.905, + 0.87 + ], + "angle": 0, + "content": "The recent development of video LLMs has shown great potential as generic video understanding assistants [21, 23, 25, 44], primarily due to their flexible video&text-in-text-out capability. Despite the effectiveness, existing video LLMs still struggle with accurate temporal event-level perception and localization (i.e., temporal grounding) [10, 13, 23, 31, 39], a crucial capability not only for grounding tasks but also for generating correct textual answers that are heavily based on grounding ability." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Various efforts have been made to enhance the temporal grounding ability of video LLMs [4, 10, 11, 14]. Be" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.95, + 0.518, + 0.962 + ], + "angle": 0, + "content": "20684" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.287 + ], + "angle": 0, + "content": "sides adding targeted data for temporally sensitive fine-tuning [13, 20, 31, 37, 43], recent work usually designs different special tokens for precise temporal representation [10, 14, 23, 29, 36], where some work further combines such special tokens with extra grounding decoder [4, 11]. Despite their effectiveness, we argue that there still exist two potential limitations. First, in the generation sequence, special time tokens are interleaved with text tokens, leading to a highly coupled learning objective between temporal grounding and textual answering. Second, these special tokens are primarily designed for timestamp representations to output timestamp, lacking the ability to capture the visual semantics of grounded events." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.288, + 0.482, + 0.395 + ], + "angle": 0, + "content": "In this work, we aim to enhance temporally grounded video LLMs from a factorized learning perspective. We first decompose the task into learning of temporal evidence grounding and textual response while also strengthening their inherent dependency. We also propose a factorized preference learning paradigm to further facilitate the two-task learning." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.401, + 0.317, + 0.417 + ], + "angle": 0, + "content": "2.2. Preference Optimization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.423, + 0.483, + 0.651 + ], + "angle": 0, + "content": "Reinforcement learning from human feedback (RLHF) [28, 33] or from AI feedback (RLAIF) [3, 15] has been widely used to align LLMs' behavior with human preference, with PPO [27, 32, 33, 48] and DPO [30] as common implementations. While this spirit has extended to the multimodal domain [17, 35, 45], it remains underexplored for video LLMs [18, 46], especially regarding explicit temporal grounding. In this work, we introduce a factorized preference optimization (FPO) algorithm that explicitly incorporates probabilistic temporal grounding modeling into the optimization objective, enabling preference learning for both temporal grounding and textual responses. We also construct a synthetic dataset to address the lack of suitable datasets for factorized preference learning for explicit temporal grounding." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.665, + 0.182, + 0.68 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.69, + 0.456, + 0.707 + ], + "angle": 0, + "content": "Here we will present the proposed \\(\\mathrm{D}^2\\mathrm{VLM}\\) framework." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.22, + 0.729 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.482, + 0.84 + ], + "angle": 0, + "content": "We begin by first illustrating the general formulation of temporally grounded video-language models, particularly in the context of video LLMs. Given a video \\( V \\) and a textual question (or prompt) \\( Q \\), the goal is to answer the question based on the given video, where the answer is typically composed of textual response together with temporal grounding information (i.e., localized temporal interval)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.482, + 0.901 + ], + "angle": 0, + "content": "For video LLMs, separate visual and text encoders will be adopted first to encode the raw inputs into a hidden space, in the format of tokens. Specifically, input video will be encoded into video tokens \\( F_V \\in \\mathbb{R}^{(T \\times N_v) \\times C} \\), where \\( T \\) is the" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.088, + 0.895, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.52, + 0.195, + 0.898, + 0.21 + ], + "angle": 0, + "content": "Figure 2. Conceptual demonstration of the \\(\\mathrm{D}^2\\mathrm{VLM}\\) framework." + }, + { + "type": "image", + "bbox": [ + 0.554, + 0.224, + 0.866, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.52, + 0.324, + 0.898, + 0.338 + ], + "angle": 0, + "content": "Figure 3. The visual semantic capture process of token." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.366, + 0.906, + 0.487 + ], + "angle": 0, + "content": "number of frames, \\(N_{v}\\) is the number of frame-level visual tokens, and \\(C\\) is the number of channels. Without loss of generality, here we set \\(N_{v} = 1\\) (i.e., \\(F_{V}\\in \\mathbb{R}^{T\\times C}\\), one token per frame) for illustration purposes. Similarly, input question will be encoded into text tokens \\(F_{Q}\\in \\mathbb{R}^{N_{t}\\times C}\\), where \\(N_{t}\\) is the number of tokens. Then, both \\(F_{V}\\) and \\(F_{Q}\\) will feed into the video-aware LLM decoder, where the response will be generated in an autoregressive manner." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.499, + 0.614, + 0.513 + ], + "angle": 0, + "content": "3.2. \\(\\mathbf{D}^2\\mathbf{VLM}\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.907, + 0.809 + ], + "angle": 0, + "content": "We notice that temporal event-level grounding and textual response are two primary tasks in temporally grounded video understanding tasks, where they exhibit distinct characteristics yet maintain strong logical dependencies: Temporal grounding focuses on precisely locating temporal events (evidence) that support answering, while textual response emphasizes accurate evidence interpretation and generating coherent textual answers. However, existing methods typically handle these two tasks in a coupled manner, with two major limitations: (1) Various special tokens are designed for temporal grounding, but are mixed with text token generation without a clear logical structure, leading to coupled learning objectives. (2) More importantly, these special tokens mainly focus on timestamp representation to output timestamp, lacking the explicit capture of visual semantics from grounded events. In contrast, we argue that such visual semantics can actually serve as crucial contexts for subsequent textual answer generation, especially under the next-token prediction paradigm." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Based on this insight, we propose \\(\\mathrm{D}^2\\mathrm{VLM}\\), where our core motivation is to decouple the tied learning objective while also preserving and even strengthening the inherent dependency between these two tasks. An overall conceptual demonstration is shown in Fig. 2. Specifically, encoded video tokens \\(F_{V}\\) and question textual tokens \\(F_{Q}\\) are fed" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.95, + 0.518, + 0.962 + ], + "angle": 0, + "content": "20685" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "into the LLM decoder, which will generate response tokens autoregressively. Here, we enforce the video LLM to first conduct a pure grounding task before providing the actual answer, where the grounded events are expected to serve as essential evidence for the subsequent response. Then, the model will give the actual answer in an interleaved text-evidence manner which will be explained later." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.201, + 0.485, + 0.548 + ], + "angle": 0, + "content": "Evidence token for grounding and more. Technically, we introduce evidence token \\(<\\mathrm{evi}>\\), a special token type dedicated to temporal grounding. Unlike existing designs that mainly focus on category and timestamp representation, the proposed \\(<\\mathrm{evi}>\\) token not only identifies the temporal location of the grounded event but also emphasizes the explicit capture of event-level visual semantics. Specifically, as shown in Fig. 3, once the video LLM generates a token classified as an \\(<\\mathrm{evi}>\\) token, we first calculate the similarity between the \\(<\\mathrm{evi}>\\) token and each frame-level LLM-processed video token \\(\\tilde{F}_V \\in \\mathbb{R}^{T \\times C}\\) (shares the same dimension as the input \\(F_V\\)). Video tokens with high similarity to the \\(<\\mathrm{evi}>\\) token are then regarded as salient tokens. We aggregate the visual semantics from these salient tokens into the \\(<\\mathrm{evi}>\\) token with simple implementations (i.e., average pool the salient tokens and then add it to the \\(<\\mathrm{evi}>\\) token). Such an operation explicitly endows the \\(<\\mathrm{evi}>\\) token with rich visual semantics of the corresponding grounded event, enabling it to truly serve as solid evidence for the subsequent response based on it. Experiments in Sec. 6 demonstrate that both the design of event-level modeling and visual semantic capture are crucial for performance improvement." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.483, + 0.837 + ], + "angle": 0, + "content": "Interleaved text-evi token response. After the pure grounding stage, the generation objective switches to answer the input question based on the grounded evidence. The answer is typically composed of textual response together with temporal grounding information (e.g., \"You capture the dog at 10s-15s and 50s-60s.\"). Here, we formulate the response objective as an interleaved text-evidence token generation process, where tokens are generated again to represent the required temporal output information in a manner akin to evidence referencing, rather than using standard textual tokens to denote timestamps only. The exact timestamps for answering can be directly obtained from the similarity calculation between the token and the frame-level video tokens \\(\\tilde{F}_V\\), using the frame indices of the salient video tokens (Fig. 3). To make the model aware of such transition between tasks, we introduce another special token, , to indicate the end of the evidence grounding stage and encourage the model to begin generating the interleaved text-evidence answer." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Ideally, the generated token during this stage should also match the one in the evidence grounding phase (i.e., the final response should be consistent with the previous grounded one). We introduce an explicit constraint to" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.093, + 0.685, + 0.107 + ], + "angle": 0, + "content": "enhance this consistency:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.109, + 0.905, + 0.149 + ], + "angle": 0, + "content": "\\[\nL _ {c o n s} = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\left| F _ {< e v i > _ {k}} ^ {S _ {1}} - F _ {< e v i > _ {k}} ^ {S _ {2}} \\right|, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.153, + 0.905, + 0.274 + ], + "angle": 0, + "content": "where \\( S_{1}, S_{2} \\) denote the pure evidence grounding stage and interleaved text-evil response stage, respectively. \\( k \\) is the \\( k^{th} < \\mathrm{evi} > \\) token in each stage. The \\( K < \\mathrm{evi} > \\) tokens are generated in temporal order, each aligned with a distinct interval corresponding to a ground-truth event. Such a constraint enforces strong consistency across stages, preserving logical coherence and strengthening the alignment between evidence grounding and answer generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.275, + 0.905, + 0.304 + ], + "angle": 0, + "content": "Loss function. We adopt the following components to supervise network learning:" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.31, + 0.905, + 0.328 + ], + "angle": 0, + "content": "\\[\nL = L _ {s f t} + L _ {g n d} + L _ {c o n s}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.335, + 0.905, + 0.395 + ], + "angle": 0, + "content": "where \\( L_{sft} \\) is the standard token classification loss for supervised fine-tuning [22, 24, 38], \\( L_{gnd} \\) is the average grounding loss for generated tokens across both stages, with each formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.397, + 0.905, + 0.437 + ], + "angle": 0, + "content": "\\[\nL _ {g n d} ^ {< e v i >} = \\frac {1}{T} \\sum_ {t = 1} ^ {T} B C E \\left(y ^ {t}, s i m ^ {t}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.443, + 0.905, + 0.505 + ], + "angle": 0, + "content": "where BCE is the binary cross-entropy loss, \\( y^{t} \\) is the frame-level ground-truth (i.e., 1/0 for foreground/background), and \\( sim^{t} \\) is the normalized dot-product similarity between token and frame-level video token feature \\( \\tilde{F}_{V}^{t} \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.515, + 0.838, + 0.533 + ], + "angle": 0, + "content": "4. Factorized Preference Optimization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.541, + 0.906, + 0.691 + ], + "angle": 0, + "content": "To further enhance the model learning on both temporal event grounding and textual response, we introduce a novel factorized preference optimization (FPO) algorithm. Unlike standard preference optimization, FPO explicitly incorporates probabilistic temporal grounding modeling into the optimization objective, enabling explicit preference learning on both temporal grounding and textual response. Here, we start with a macroscopic formulation that captures the essence of preference optimization and gradually elaborate on our new design." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.692, + 0.906, + 0.767 + ], + "angle": 0, + "content": "Inspired by the existing efficient preference learning algorithm DPO [30], we adopt a similar principle: encouraging the model to assign a higher probability of generating a good response than generating a bad response, which can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.775, + 0.905, + 0.794 + ], + "angle": 0, + "content": "\\[\nL = - \\left[ \\log \\sigma (\\log \\pi (R) - \\log \\pi (\\bar {R})) \\right], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.802, + 0.905, + 0.863 + ], + "angle": 0, + "content": "where \\(\\pi (R)\\) and \\(\\pi (\\bar{R})\\) are the joint generation probabilities of preferred and dis-preferred response, respectively. Under the autoregressive paradigm, such joint generation log-probability can be further detailed as:" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.871, + 0.905, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\log \\pi (R) = \\sum_ {R _ {i} \\in R} \\log p \\left(R _ {i} \\mid V, Q, R _ {1: i - 1}\\right), \\tag {5}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.95, + 0.519, + 0.962 + ], + "angle": 0, + "content": "20686" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.089, + 0.47, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.476, + 0.089, + 0.865, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.279, + 0.848, + 0.294 + ], + "angle": 0, + "content": "Figure 4. The proposed data synthesis pipeline, where factorized perturbation is imposed to form the dispreferred data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.319, + 0.482, + 0.35 + ], + "angle": 0, + "content": "where \\(R_{i}\\) is the \\(i^{th}\\) token of the whole model response \\(R\\). \\(V\\) and \\(Q\\) are the input video and question, respectively." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.35, + 0.483, + 0.621 + ], + "angle": 0, + "content": "One of our goals is to perform preference optimization not only for textual responses but also with an explicit emphasis on temporal grounding. However, achieving such multi-factor preference learning presents two key challenges: (1) The lack of suitable preference data that explicitly considers temporal grounding while maintaining a factorized structure. We will elaborate more on this issue and introduce our newly proposed synthetic dataset as a solution in Sec. 5. (2) The challenge on building suitable objectives for explicit preference optimization on temporal grounding, particularly in quantifying the model's grounding probability (i.e., measuring the likelihood that the model will ground a specific temporal event). Many existing works introduce special tokens and regression heads to enhance the grounding ability of video LLMs. Despite their effectiveness, this deviates from the standard discrete text token prediction paradigm, making it non-trivial to model grounding likelihood for preference optimization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.621, + 0.483, + 0.817 + ], + "angle": 0, + "content": "In this work, we notice that our designed \\(\\mathrm{D}^2\\mathrm{VLM}\\) framework can naturally model such event grounding probability, in a different and explicit way compared with the standard token prediction probability. Specifically, as grounding is achieved by frame-level feature similarity calculation between the \\(<\\mathrm{evi}>\\) token and each video frame, we directly use such similarity as the frame-level grounding probability and multiply these probabilities across frames to obtain the interval-level grounding probability. Formally, for each generated \\(<\\mathrm{evi}>_k\\) token during model response, its corresponding grounding probability \\(p_g([s_k,e_k])\\) for a given time interval \\([s_k,e_k]\\) (\\(s_k\\), \\(e_k\\) denote the starting and ending time index respectively) can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.129, + 0.824, + 0.483, + 0.865 + ], + "angle": 0, + "content": "\\[\np _ {g} ([ s _ {k}, e _ {k} ]) = \\prod_ {t = 1} ^ {T} \\left\\{ \\begin{array}{l l} s i m _ {k} ^ {t} & \\text {i f} s _ {k} \\leq t \\leq e _ {k} \\\\ 1 - s i m _ {k} ^ {t} & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Here, \\(T\\) is the total number of input video frames, \\(sim_k^t\\) is the normalized dot-product similarity between \\(\\langle \\mathrm{evi} \\rangle_k\\) and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.319, + 0.907, + 0.366 + ], + "angle": 0, + "content": "frame-level video token feature \\(\\tilde{F}_V^t\\). Based on such event grounding modeling, the log-probability formulation within our proposed factorized preference optimization is given as:" + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.383, + 0.905, + 0.452 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\log \\pi (R) = \\sum_ {R _ {i} \\in R} \\log p \\left(R _ {i} \\mid V, Q, R _ {1: i - 1}\\right) \\\\ + \\sum_ {R _ {k} = < e v i >} \\log p _ {g} \\left(\\left[ s _ {k}, e _ {k} \\right] \\mid V, Q, R _ {1: k - 1}\\right), \\end{array} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.457, + 0.907, + 0.578 + ], + "angle": 0, + "content": "where the first term corresponds to the standard textual token prediction modeling, while the second term represents the proposed explicit modeling on temporal event grounding. By substituting Equation (7) into Equation (4), we obtain the final formulation of our Factorized Preference Optimization (FPO). Next, we explain the process of constructing multi-factor preferred \\((R)\\) and dis-preferred \\((\\bar{R})\\) data pairs, which serve as the training data for FPO." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.589, + 0.851, + 0.606 + ], + "angle": 0, + "content": "5. Factorized Preference Data Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.901 + ], + "angle": 0, + "content": "As previously mentioned, one key challenge is the lack of suitable preference data that captures multi-factor preferences—not only for textual responses but also for explicit temporal grounding. Particularly, in the video domain, preference data is typically constructed by providing a video LLM with either complete or degraded video inputs (e.g., blurred or cropped images/videos) and treating its responses as preferred or dispreferred, respectively [18]. The underlying assumption is that the model will produce inferior responses when given degraded or incomplete visual inputs. However, existing works using this approach often face some major issues: (1) The reliability of the generated preference data is relatively less controlled. Specifically, the generated preferred sample may not actually be the preferred one, and the generated dispreferred data can sometimes be as good as or even better than the preferred data [18]. As a result, post-filtering is usually required, which involves additional models and ad-hoc human interventions, increasing costs but still cannot fully ensure" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.95, + 0.518, + 0.962 + ], + "angle": 0, + "content": "20687" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.242 + ], + "angle": 0, + "content": "quality. (2) This lack of control also makes it difficult to properly factorize the generated dispreferred data, including identifying the source or the reason why it is considered dispreferred. (3) Last but not least, existing video preference data mainly focuses on generating responses in pure textual format (e.g., video caption) [18, 46], while generating preference response that explicitly contains temporal grounding information (e.g., temporal event position) remains under-explored. Overall, these issues make factorized preference optimization difficult to achieve." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.245, + 0.482, + 0.321 + ], + "angle": 0, + "content": "To address the aforementioned limitations, we propose an automatic factorized preference data synthesis framework, which introduces factorized perturbations on the original response at the sub-video event level. As shown in Fig. 4 (a), the overall process consists of the following steps:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.322, + 0.483, + 0.594 + ], + "angle": 0, + "content": "Random selection of factors and sub-video events. For each given sample (video-question-answer pair), we first randomly select a factor that determines the exact type of the perturbation. As shown in Fig. 4 (b), the factors are broadly categorized into two types and each can be further divided into multiple sub-types: (1) Temporal grounding: This type of perturbation aims to reveal/simulate model response inaccuracies related to temporal event grounding. It can be further divided into sub-types, including temporal localization shift, randomly adding/deleting grounded events (corresponding to false positives and missed detections), and merging multiple events into one (corresponding to a lack of fine-grained distinction in event boundaries). (2) Textual response: This type of perturbation modifies the semantic correctness of the textual response. It includes sub-types such as distorting key information, which disrupts critical content, and repeating responses, a common failure mode observed in video LLMs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.483, + 0.868 + ], + "angle": 0, + "content": "Next, we randomly select some sub-video events to apply factorized perturbation. A sub-video event refers to a semantically meaningful segment of the video that can be determined based on the given sample (e.g., its question or ground-truth answer). For instance, in a dense video captioning task, an event naturally corresponds to a specific captioned interval, while in an action localization task, an event can align with a single action instance. Without loss of generality, we take the dense video caption task as an illustration example (as shown in Fig. 4 (c)), as this task emphasizes both temporal grounding and textual description. Overall, such an event-level design aligns naturally with the inherent video structure, ensuring that perturbations are introduced in a way that respects the video's temporal and semantic coherence. Moreover, it enables fine-grained perturbation, allowing controlled modifications at the event level rather than applying coarse, less precisely controlled noise to the entire response sequence." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Injecting factorized perturbations. Once the factors and sub-video events are determined, localized perturbation is" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.333 + ], + "angle": 0, + "content": "introduced based on the selected factors and sub-video events. If a temporal grounding factor is selected, we modify the timestamps of the selected events based on the determined sub-factor (e.g., shift the interval, delete the interval, etc.). For the factor of textual response, for each selected event, we prompt off-the-shelf LLM (e.g., Qwen [41, 42]) to generate a distorted response based on the original correct event-level response. If the \"repeat\" sub-factor is selected, we can easily achieve this by setting the response of the current event to its previous one to simulate such a response repeating phenomenon. A synthesized example is shown in Fig. 4 (c). After the perturbation synthesis, we combine the newly obtained noisy response with the original one to form the preference sample pair for factorized preference optimization. In practice, our preference data is synthesized based on the E.T. Instruct 164K dataset [23]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.341, + 0.905, + 0.476 + ], + "angle": 0, + "content": "Discussion. Compared to existing preference data synthesis approaches, our framework ensures that the generated dispreferred response contains specific, well-factorized noise rather than arbitrary, unpredictable degradation, and thus we can better identify the noise source without manual costs. To our knowledge, we are also the first to emphasize explicit temporal grounding in preference data, and moreover, in a multi-factorized manner. The overall design makes our factorized preference optimization in Sec. 4 feasible." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.514, + 0.645, + 0.531 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.546, + 0.703, + 0.563 + ], + "angle": 0, + "content": "6.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.576, + 0.905, + 0.713 + ], + "angle": 0, + "content": "Implementation details. Following the practice of [23], we adopt the pre-trained ViT-G/14 from EVA-CLIP [8, 34], followed by a Q-Former-like feature compressor [6, 16, 23], and use Phi-3-Mini-3.8B [1] as the base LLM. Also following [23], our model is initialized from the stage-2 model of [23] and [19], and we also only fine-tune our model on the E.T. Instruct 164K dataset [23] using LoRA [12]. We conduct training on \\(4 \\times\\) NVIDIA H100 GPUs within 1 day. More details are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Evaluation tasks and datasets. We evaluate our method across various datasets and tasks. (1) E.T. Bench Grounding [23]. It consists of 5 different grounding-related tasks (e.g., moment retrieval, action localization, summarization/highlight detection, etc.) for comprehensive evaluation. (2) E.T. Bench Dense Captioning [23]. It is composed of tasks like dense video caption and step localization, where both key event grounding and textual description are required. Besides, we also conduct experiments on two widely used individual datasets, (3) Charades-STA [9] and (4) YouCook2 [47], for independent evaluation on moment retrieval and dense video captioning, respectively." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.95, + 0.518, + 0.962 + ], + "angle": 0, + "content": "20688" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.904, + 0.282 + ], + "angle": 0, + "content": "
MethodYearE.T. Bench GroundingE.T. Bench Dense Captioning
TVG\\( _{F1} \\)EPM\\( _{F1} \\)TAL\\( _{F1} \\)EVS\\( _{F1} \\)VHD\\( _{F1} \\)Avg\\( _{F1} \\)DVC\\( _{F1} \\)DVC\\( _{Sim} \\)SLC\\( _{F1} \\)SLC\\( _{Sim} \\)Avg\\( _{F1} \\)Avg\\( _{Sim} \\)
Video-ChatGPT-7B [26]ACL'247.01.315.18.428.812.18.811.35.710.27.310.8
Video-LLaVA-7B [21]EMNLP'247.01.915.00.328.910.628.015.00.98.314.511.7
LLaMA-VID-7B [19]ECCV'245.51.28.01.430.09.227.112.65.211.116.211.9
Video-LLaMA-2-7B [5]arXiv'240.10.00.00.01.50.30.614.50.015.20.314.9
PLLAVA [40]arXiv'246.91.15.70.328.98.613.310.69.711.811.511.2
VTimeLLM-7B [13]CVPR'247.61.918.215.928.914.512.413.18.76.410.69.8
TimeChat-7B [31]CVPR'2426.23.910.129.140.522.016.612.55.69.211.110.9
LITA-13B [14]ECCV'2422.24.618.029.723.919.739.717.221.012.230.414.7
VTG-LLM-7B [10]AAAI'2515.93.714.426.848.221.840.218.620.814.430.516.5
Qwen2.5-VL-7B* [2]arXiv'2546.69.332.219.968.635.339.021.425.515.132.218.2
E.T.Chat-3.8B [23]NeurIPS'2438.610.230.825.462.533.538.419.724.414.631.417.2
D²VLM-3.8B (Ours)ICCV'2560.214.433.435.268.242.348.325.526.718.137.521.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.286, + 0.907, + 0.343 + ], + "angle": 0, + "content": "Table 1. Performance comparison on E.T. Bench-Grounding dataset (5 sub tasks: TVG: Temporal Video Grounding; EPM: Episodic Memory; TAL: Temporal Action Localization; EVS: Extractive Video Summarization; VHD: Video Highlight Detection) and E.T. Bench-Dense Captioning dataset (2 sub tasks: DVC: Dense Video Captioning; SLC: Step Localization and Captioning). \\*: Implemented by us, with more details in the supplementary material." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.366, + 0.482, + 0.503 + ], + "angle": 0, + "content": "
MethodYearR@1(IoU=0.5)R@1(IoU=0.7)
TimeChat-7B [31]CVPR’2432.213.4
VTimeLLM-7B [13]CVPR’2427.511.4
VTimeLLM-13B [13]CVPR’2434.314.7
Momentor-7B [29]ICML’2426.611.6
HawkEye-7B [37]arXiv’2431.414.5
VTG-LLM-7B [10]AAAI’2533.815.7
TRACE-7B [11]ICLR’2540.319.4
VideoChat-T-7B [43]ICLR’2548.724.0
E.T.Chat-3.8B [23]NeurIPS’2445.920.0
D²VLM-3.8B (Ours)ICCV’2550.326.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.13, + 0.508, + 0.444, + 0.521 + ], + "angle": 0, + "content": "Table 2. Performance comparison on Charades-STA." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.537, + 0.482, + 0.619 + ], + "angle": 0, + "content": "
MethodYearF1CIDErSODA_c
TimeChat-7B [31]CVPR’2412.63.41.2
VTG-LLM-7B [10]AAAI’2517.55.01.5
TRACE-7B [11]ICLR’2522.48.12.2
D²VLM-3.8B (Ours)ICCV’2526.410.63.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.141, + 0.623, + 0.434, + 0.637 + ], + "angle": 0, + "content": "Table 3. Performance comparison on YouCook2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.665, + 0.348, + 0.682 + ], + "angle": 0, + "content": "6.2. Main Benchmarking Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.84 + ], + "angle": 0, + "content": "E.T. Bench grounding. The results are shown in the left part of Tab. 1, where we compare our method with the recently published SOTA methods. It can be observed that the proposed \\(\\mathrm{D}^2\\mathrm{VLM}\\) significantly outperforms recently published SOTA methods across various tasks, achieving an average F1 score improvement of at least \\(7.0\\%\\). Meanwhile, our model is relatively lightweight (3.8B parameters) compared to most state-of-the-art methods (ranging from 3.8B to 13B, with the majority being 7B parameters). This further demonstrates the superiority of our proposed method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "E.T. Bench dense captioning. The performance comparison is shown in the right part of Tab. 1. It can be seen that our method also shows noticeable superiority in dense captioning-related tasks, where it outperforms recent SOTA" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.368, + 0.905, + 0.429 + ], + "angle": 0, + "content": "methods in both temporal event grounding (at least \\(5.3\\%\\) on average F1 score) and textual description (at least \\(3.6\\%\\) on average Sim score). This essentially demonstrates the generalizability of our method across various scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.43, + 0.907, + 0.55 + ], + "angle": 0, + "content": "Charades-STA. From Tab. 2, it can be seen that our method still outperforms recent SOTA methods by large margins. Specifically, our 3.8B model outperforms counterpart [23] with the same base model by \\(4.4\\%\\) on \\(\\mathrm{R}@\\mathrm{1(IoU = 0.5)}\\), and outperforms methods published in 2025 with 7B model size [10, 11, 31] by at least \\(1.6\\%\\) on \\(\\mathrm{R}@\\mathrm{1(IoU = 0.5)}\\). This further demonstrates the superiority of our method in both effectiveness and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.551, + 0.906, + 0.672 + ], + "angle": 0, + "content": "YouCook2. As shown in Tab. 3, our method achieves state-of-the-art performance compared to recent SOTA approaches. Specifically, it outperforms them by at least \\(4\\%\\) in temporal event grounding (F1) while also demonstrating superior textual description capacity, with gains of at least \\(2.5\\%\\) on CIDEr and \\(1.0\\%\\) on SODA.c. This further highlights the superiority of our approach in both temporal grounding and text generation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.682, + 0.68, + 0.697 + ], + "angle": 0, + "content": "6.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.856 + ], + "angle": 0, + "content": "Here we conduct ablation studies on E.T. Bench grounding and E.T. Bench Dense Captioning datasets to analyze the effect of the proposed components, in both temporal event-level grounding and textual description. For each dataset (i.e., grounding, dense captioning), we report the average performance across multiple sub-tasks. In the following, we first analyze the component effect of the newly designed generation objective and evidence token, where we do not take the proposed FPO into account. Finally, we analyze the effect of the proposed FPO." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Generation objective of \\(\\mathbf{D}^2\\mathbf{VLM}\\). Here, we analyze the effectiveness of our newly designed generation objective within \\(\\mathbf{D}^2\\mathbf{VLM}\\). From Tab. 4, it can be seen that: (1) Row-1" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.95, + 0.52, + 0.962 + ], + "angle": 0, + "content": "20689" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.596, + 0.178 + ], + "angle": 0, + "content": "
Decomposed ObjectiveInterleaved Text-avi GenerationConsistency ConstraintGrounding AvgF1Dense Captioning AvgF1AvgSim
21.214.311.3
✓28.923.116.0
✓✓35.634.319.8
✓✓✓39.535.021.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.155, + 0.182, + 0.536, + 0.198 + ], + "angle": 0, + "content": "Table 4. Component effect of the designed generation objective." + }, + { + "type": "table", + "bbox": [ + 0.619, + 0.089, + 0.862, + 0.178 + ], + "angle": 0, + "content": "
FPOGrounding AvgF1Dense Captioning AvgF1AvgSim
×39.535.021.2
✓42.337.521.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.63, + 0.182, + 0.851, + 0.197 + ], + "angle": 0, + "content": "Table 5. Effect of the proposed FPO." + }, + { + "type": "table", + "bbox": [ + 0.114, + 0.22, + 0.46, + 0.304 + ], + "angle": 0, + "content": "
MethodGrounding \nAvgF1Dense \nAvgF1Captioning \nAvgSim
w/o event-level modeling26.133.416.2
w/o visual semantic capture37.127.517.7
Full design39.535.021.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.116, + 0.306, + 0.456, + 0.321 + ], + "angle": 0, + "content": "Table 6. Component effect within evidence token design." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.354, + 0.483, + 0.656 + ], + "angle": 0, + "content": "v.s. Row-2: Decompose the coupled grounding-textual response objective into grounding (by our designed token) then answer from grounded event can substantially improve the performance in both grounding (i.e., \\(7.7\\%\\) averaged F1 on grounding tasks and \\(8.8\\%\\) averaged F1 on dense captioning tasks) and textual description (i.e., \\(4.7\\%\\) averaged sentence similarity on dense captioning tasks). (2) Row-2 v.s. Row-3: When we shift the answering stage from pure textual token generation into interleaved text-evi token generation, the performance can be further enhanced by large margins (at least \\(6.7\\%\\) on grounding and \\(3.8\\%\\) on textual description). The reason is that under such cases, the overall generation objective can be interpreted as grounding then answer with evidence referencing, where such referencing strengthens the relationship between event grounding and textual answering. (3) Row-3 v.s. Row-4: The introduced consistency constraint further boosts the performance by a noticeable margin, as it explicitly enhances the consistency between the two stages. Overall, the results validate the effectiveness of our design." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Design of evidence token. Here we analyze the impact of the core components within the token design (i.e., the event-level modeling and the explicit visual semantic capture). The results are shown in Tab. 6, where w/o event-level modeling means each token only captures the information of individual frame (a common design in existing works [14, 23, 29]), rather than an interval-level event (e.g., 10s-15s). w/o visual semantic capture indicates visual semantics from the video tokens is not explicitly integrated into the token through the operation illustrated in Sec. 3.2 (also a common design in most of the existing works [4, 14, 29]). From Tab. 6, it can be summarized that: (1) The performance drops significantly without event-level modeling. We argue that this is because compared to individual timestamp-level modeling, event-level modeling more naturally aligns with the characteristics of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.224, + 0.905, + 0.373 + ], + "angle": 0, + "content": "event-level evidence grounding and can capture more robust event-level semantics in a global perspective. (2) Explicit visual semantic capture is also crucial for enhancing model effectiveness, as it enables tokens to truly serve as solid context for subsequent answer generation, which also fits well with the autoregressive generation process. Notably, compared with grounding-only tasks, such an effect is more pronounced on dense captioning tasks, as textual description generation can truly benefit from the captured event-level visual semantics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.375, + 0.907, + 0.466 + ], + "angle": 0, + "content": "Effect of factorized preference optimization. Tab. 5 demonstrates that the proposed FPO algorithm consistently enhances performance on both temporal grounding and captioning-related tasks, with a more significant improvement in temporal grounding. This indeed validates the effectiveness of our factorized optimization objective design." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.482, + 0.771, + 0.497 + ], + "angle": 0, + "content": "7. Conclusion and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.907, + 0.749 + ], + "angle": 0, + "content": "In this work, we uncover some key limitations of existing temporal grounded video-language models: (1) the use of coupled learning objective for temporal grounding and textual response generation, and (2) the overlooking of inherent dependency between these two tasks, especially the event-level visual semantics from the intermediate grounding result. We tackle this from a factorized learning perspective, where we propose \\(\\mathrm{D}^2\\mathrm{VLM}\\) that decomposes the generation objective into a \"grounding then answering with evidence referencing\" paradigm, and introduce evidence tokens to emphasize explicit visual semantic capture beyond timestamp representation. Furthermore, we design a factorized preference learning algorithm, coupled with a factorized synthetic dataset, to enhance the learning of both tasks. Extensive experiments demonstrate the clear superiority of our approach across various scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Despite achieving new state-of-the-art results, our approach still has room for improvement (e.g., F1 scores of only \\(14.4\\%\\) on the episodic memory task and \\(26.4\\%\\) on the YouCook2 dense video captioning task). Besides, while our factorized data generation offers good factorized controllability and directly contributes to factorized optimization, it focuses only on generating negative (dis-preferred) samples. However, generating more diverse positive samples to further enrich the preference data is also meaningful, and we leave this for future work." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.95, + 0.52, + 0.962 + ], + "angle": 0, + "content": "20690" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.222 + ], + "angle": 0, + "content": "This research is supported by the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No: AISG3-RP-2022-030). We would like to acknowledge that computational work involved in this research work is partially supported by NUS IT's Research Computing group under grant number NUSREC-HPC-00001." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.237, + 0.188, + 0.254 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.262, + 0.484, + 0.331 + ], + "angle": 0, + "content": "[1] Marah Abdin, Sam Ade Jacobs, Ammar Ahmad Awan, Jyoti Aneja, Ahmed Awadallah, Hany Awadalla, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Harkirat Behl, et al. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.333, + 0.484, + 0.387 + ], + "angle": 0, + "content": "[2] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.39, + 0.484, + 0.458 + ], + "angle": 0, + "content": "[3] Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, et al. Constitutional ai: Harmlessness from ai feedback. arXiv preprint arXiv:2212.08073, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.461, + 0.484, + 0.502 + ], + "angle": 0, + "content": "[4] Qirui Chen, Shangzhe Di, and Weidi Xie. Grounded multi-hop videoqa in long-form egocentric videos. arXiv preprint arXiv:2408.14469, 2024. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.505, + 0.484, + 0.573 + ], + "angle": 0, + "content": "[5] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.576, + 0.484, + 0.658 + ], + "angle": 0, + "content": "[6] Wenliang Dai, Junnan Li, DONGXU LI, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale N Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. In Advances in Neural Information Processing Systems, pages 49250-49267. Curran Associates, Inc., 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.661, + 0.484, + 0.716 + ], + "angle": 0, + "content": "[7] Shangzhe Di and Weidi Xie. Grounded question-answering in long egocentric videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12934–12943, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.718, + 0.484, + 0.799 + ], + "angle": 0, + "content": "[8] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19358-19369, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.803, + 0.484, + 0.857 + ], + "angle": 0, + "content": "[9] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In Proceedings of the IEEE international conference on computer vision, pages 5267-5275, 2017. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[10] Yongxin Guo, Jingyu Liu, Mingda Li, Dingxin Cheng, Xiaoying Tang, Dianbo Sui, Qingbin Liu, Xi Chen, and Kevin Zhao. Vtg-llm: Integrating timestamp knowledge into video" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.262, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.12 + ], + "angle": 0, + "content": "llms for enhanced video temporal grounding. arXiv preprint arXiv:2405.13382, 2024. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.176 + ], + "angle": 0, + "content": "[11] Yongxin Guo, Jingyu Liu, Mingda Li, Xiaoying Tang, Qingbin Liu, and Xi Chen. Trace: Temporal grounding video llm via causal event modeling. In International Conference on Learning Representations, 2025. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.906, + 0.232 + ], + "angle": 0, + "content": "[12] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.234, + 0.906, + 0.3 + ], + "angle": 0, + "content": "[13] Bin Huang, Xin Wang, Hong Chen, Zihan Song, and Wenwu Zhu. Vtimellm: Empower llm to grasp video moments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14271-14280, 2024. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.906, + 0.37 + ], + "angle": 0, + "content": "[14] De-An Huang, Shijia Liao, Subhashree Radhakrishnan, Hongxu Yin, Pavlo Molchanov, Zhiding Yu, and Jan Kautz. Lita: Language instructed temporal-localization assistant. In European Conference on Computer Vision, pages 202-218. Springer, 2024. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.372, + 0.906, + 0.44 + ], + "angle": 0, + "content": "[15] Harrison Lee, Samrat Phatale, Hassan Mansoor, Thomas Mesnard, Johan Ferret, Kellie Lu, Colton Bishop, Ethan Hall, Victor Carbune, Abhinav Rastogi, et al. Rlaif vs. rlhf: Scaling reinforcement learning from human feedback with ai feedback. arXiv preprint arXiv:2309.00267, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.442, + 0.906, + 0.509 + ], + "angle": 0, + "content": "[16] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.511, + 0.906, + 0.567 + ], + "angle": 0, + "content": "[17] Lei Li, Zhihui Xie, Mukai Li, Shunian Chen, Peiyi Wang, Liang Chen, Yazheng Yang, Benyou Wang, and Lingpeng Kong. Silkie: Preference distillation for large visual language models. arXiv preprint arXiv:2312.10665, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.568, + 0.906, + 0.621 + ], + "angle": 0, + "content": "[18] Rui Li, Xiaohan Wang, Yuhui Zhang, Zeyu Wang, and Serena Yeung-Levy. Temporal preference optimization for long-form video understanding. arXiv preprint arXiv:2501.13919, 2025. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.623, + 0.906, + 0.678 + ], + "angle": 0, + "content": "[19] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In European Conference on Computer Vision, pages 323–340. Springer, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.679, + 0.906, + 0.761 + ], + "angle": 0, + "content": "[20] Zhaowei Li, Qi Xu, Dong Zhang, Hang Song, YiQing Cai, Qi Qi, Ran Zhou, Junting Pan, Zefeng Li, Vu Tu, et al. Groundinggpt: Language enhanced multi-modal grounding model. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6657-6678, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.763, + 0.907, + 0.831 + ], + "angle": 0, + "content": "[21] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 5971–5984, 2024. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.906, + 0.873 + ], + "angle": 0, + "content": "[22] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[23] Ye Liu, Zongyang Ma, Zhongang Qi, Yang Wu, Ying Shan, and Chang Wen Chen. E.T. Bench: Towards open-ended" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.95, + 0.518, + 0.962 + ], + "angle": 0, + "content": "20691" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "event-level video-language understanding. In The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2024. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.135, + 0.483, + 0.217 + ], + "angle": 0, + "content": "[24] Shayne Longpre, Le Hou, Tu Vu, Albert Webson, Hyung Won Chung, Yi Tay, Denny Zhou, Quoc V Le, Barret Zoph, Jason Wei, et al. The flan collection: Designing data and methods for effective instruction tuning. In International Conference on Machine Learning, pages 22631-22648. PMLR, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.483, + 0.3 + ], + "angle": 0, + "content": "[25] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12585–12602, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.301, + 0.482, + 0.37 + ], + "angle": 0, + "content": "[26] Munan Ning, Bin Zhu, Yujia Xie, Bin Lin, Jiaxi Cui, Lu Yuan, Dongdong Chen, and Li Yuan. Video-bench: A comprehensive benchmark and toolkit for evaluating video-based large language models. arXiv preprint arXiv:2311.16103, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.371, + 0.483, + 0.453 + ], + "angle": 0, + "content": "[27] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.483, + 0.537 + ], + "angle": 0, + "content": "[28] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.538, + 0.483, + 0.607 + ], + "angle": 0, + "content": "[29] Long Qian, Juncheng Li, Yu Wu, Yaobo Ye, Hao Fei, TatSeng Chua, Yueting Zhuang, and Siliang Tang. Momentor: Advancing video large language model with fine-grained temporal reasoning. arXiv preprint arXiv:2402.11435, 2024. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.608, + 0.483, + 0.677 + ], + "angle": 0, + "content": "[30] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36:53728-53741, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.678, + 0.483, + 0.747 + ], + "angle": 0, + "content": "[31] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14313-14323, 2024. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.483, + 0.789 + ], + "angle": 0, + "content": "[32] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[33] Nisan Stiannon, Long Ouyang, Jeffrey Wu, Daniel Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford, Dario Amodei, and Paul F Christiano. Learning to summarize with human feedback. Advances in neural information processing systems, 33:3008-3021, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[34] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.174 + ], + "angle": 0, + "content": "[35] Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liang-Yan Gui, Yu-Xiong Wang, Yiming Yang, et al. Aligning large multimodal models with factually augmented rlhf. In Annual Meeting of the Association for Computational Linguistics, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.176, + 0.906, + 0.245 + ], + "angle": 0, + "content": "[36] Haibo Wang, Zhiyang Xu, Yu Cheng, Shizhe Diao, Yufan Zhou, Yixin Cao, Qifan Wang, Weifeng Ge, and Lifu Huang. Grounded-videollm: Sharpening fine-grained temporal grounding in video large language models. arXiv preprint arXiv:2410.03290, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.906, + 0.3 + ], + "angle": 0, + "content": "[37] Yueqian Wang, Xiaojun Meng, Jianxin Liang, Yuxuan Wang, Qun Liu, and Dongyan Zhao. Hawkeye: Training video-text llms for grounding text in videos. arXiv preprint arXiv:2403.10228, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.906, + 0.357 + ], + "angle": 0, + "content": "[38] Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.906, + 0.426 + ], + "angle": 0, + "content": "[39] Junbin Xiao, Angela Yao, Yicong Li, and Tat-Seng Chua. Can i trust your answer? visually grounded video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13204-13214, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.906, + 0.482 + ], + "angle": 0, + "content": "[40] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.906, + 0.691 + ], + "angle": 0, + "content": "[41] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng He, Junyang Lin, Kai Dang, Keming Lu, Keqin Chen, Kexin Yang, Mei Li, Mingfeng Xue, Na Ni, Pei Zhang, Peng Wang, Ru Peng, Rui Men, Ruize Gao, Runji Lin, Shijie Wang, Shuai Bai, Sinan Tan, Tianhang Zhu, Tianhao Li, Tianyu Liu, Wenbin Ge, Xiaodong Deng, Xiaohuan Zhou, Xingzhang Ren, Xinyu Zhang, Xipin Wei, Xuancheng Ren, Yang Fan, Yang Yao, Yichang Zhang, Yu Wan, Yunfei Chu, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, and Zhihao Fan. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.906, + 0.831 + ], + "angle": 0, + "content": "[42] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, Le Yu, Mei Li, Mingfeng Xue, Pei Zhang, Qin Zhu, Rui Men, Runji Lin, Tianhao Li, Tingyu Xia, Xingzhang Ren, Xuancheng Ren, Yang Fan, Yang Su, Yichang Zhang, Yu Wan, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, and Zihan Qiu. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[43] Xiangyu Zeng, Kunchang Li, Chenting Wang, Xinhao Li, Tianxiang Jiang, Ziang Yan, Songze Li, Yansong Shi, Zhengrong Yue, Yi Wang, et al. Timesuite: Improving mllms for long video understanding via grounded tuning. In International Conference on Learning Representations, 2025. 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.95, + 0.519, + 0.962 + ], + "angle": 0, + "content": "20692" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.482, + 0.135 + ], + "angle": 0, + "content": "[44] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.136, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[45] Mengxi Zhang, Wenhao Wu, Yu Lu, Yuxin Song, Kang Rong, Huanjin Yao, Jianbo Zhao, Fanglong Liu, Haocheng Feng, Jingdong Wang, et al. Automated multi-level preference for mllms. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.207, + 0.483, + 0.275 + ], + "angle": 0, + "content": "[46] Ruohong Zhang, Liangke Gui, Zhiqing Sun, Yihao Feng, Keyang Xu, Yuanhan Zhang, Di Fu, Chunyuan Li, Alexander Hauptmann, Yonatan Bisk, et al. Direct preference optimization of video large multimodal models from language model reward. arXiv preprint arXiv:2404.01258, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.277, + 0.482, + 0.331 + ], + "angle": 0, + "content": "[47] Luowei Zhou, Chenliang Xu, and Jason Corso. Towards automatic learning of procedures from web instructional videos. In Proceedings of the AAAI Conference on Artificial Intelligence, 2018. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.333, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[48] Daniel M Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. Fine-tuning language models from human preferences, 2020. URL https://arxiv.org/abs, page 14, 1909. 3" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.483, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.951, + 0.518, + 0.962 + ], + "angle": 0, + "content": "20693" + } + ] +] \ No newline at end of file diff --git a/2025/Factorized Learning for Temporally Grounded Video-Language Models/0391813b-b98c-4912-884e-a194864058af_origin.pdf b/2025/Factorized Learning for Temporally Grounded Video-Language Models/0391813b-b98c-4912-884e-a194864058af_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bbca3936b6efeb695d5094279bca1ae726762379 --- /dev/null +++ b/2025/Factorized Learning for Temporally Grounded Video-Language Models/0391813b-b98c-4912-884e-a194864058af_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b8c6f32c5a4e9c2f8e6662e9254a9b84619d2f9ff92cebc75ebbfebf492830d6 +size 1117067 diff --git a/2025/Factorized Learning for Temporally Grounded Video-Language Models/full.md b/2025/Factorized Learning for Temporally Grounded Video-Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2717bf58ed3eeb4974c815c064f1432c56962dbe --- /dev/null +++ b/2025/Factorized Learning for Temporally Grounded Video-Language Models/full.md @@ -0,0 +1,291 @@ +# Factorized Learning for Temporally Grounded Video-Language Models + +Wenzheng Zeng, Difei Gao, Mike Zheng Shou†, Hwee Tou Ng† National University of Singapore + +wenzhengzeng@u.nus.edu, {daniel.difei.gao, mike.zheng.shou}@gmail.com, nght@nus.edu.sg + +![](images/734db46a50c501461c381372d16c4057a9a6ff5a1a51c7bb6694bb24042975d2.jpg) +(a) + +![](images/2ee64aa051f07996dc9f92f80ba9a130050ff53409093b1ceb61ef052113fb3d.jpg) +Figure 1. (a) Performance: Our method consistently outperforms SOTA methods across various tasks (here we draw the maximum performance across methods, detailed in Sec. 6). (b) Model: We propose a new framework $\mathrm{D}^2\mathrm{VLM}$ , where we decompose the generation objective into a "grounding then answering with evidence referencing" paradigm and introduce evidence tokens to emphasize explicit event-level visual semantic capture. (c) Training Algorithm: We introduce Factorized Preference Optimization (FPO) that explicitly addresses both temporal grounding and textual response. A factorized data synthesis approach is also designed to support FPO. + +# Abstract + +Recent video-language models have shown great potential for video understanding, but still struggle with accurate temporal grounding for event-level perception. We observe that two main factors in video understanding (i.e., temporal grounding and textual response) form a logical hierarchy: accurate temporal evidence grounding lays the foundation for reliable textual response. However, existing works typically handle these two tasks in a coupled manner without a clear logical structure, leading to sub-optimal objectives. We address this from a factorized learning perspective. We first propose $D^{2}$ VLM, a framework that decouples the learning of these two tasks while also emphasizing their inherent dependency. We adopt a "grounding then answering with evidence referencing" paradigm and introduce evidence tokens for evidence grounding, which emphasize event-level visual semantic capture beyond the focus on timestamp representation in existing works. To further facilitate the learning of these two tasks, we introduce a novel factorized + +preference optimization (FPO) algorithm. Unlike standard preference optimization, FPO explicitly incorporates probabilistic temporal grounding modeling into the optimization objective, enabling preference learning for both temporal grounding and textual response. We also construct a synthetic dataset to address the lack of suitable datasets for factorized preference learning with explicit temporal grounding. Experiments on various tasks demonstrate the clear advantage of our approach. Our source code is available at https://github.com/nusnlp/d2v1m. + +# 1. Introduction + +Recent advances in video-language models, especially those built upon large language models (video LLMs), have enabled remarkable progress in video understanding [10, 21, 23, 25, 31, 44]. Through their flexible video&text-int-text-out nature, video LLMs demonstrate great potential as general-purpose solvers, unifying various tasks (e.g., temporal grounding [9], dense captioning [47], and question answering [7]) into one generic "video question answer- + +ing" framework. Despite their effectiveness, existing video LLMs still struggle with accurate temporal grounding for event perception and localization [10, 11, 23, 31, 31], a crucial capability not only for grounding tasks themselves but also for relevant tasks that need textual answering. + +We notice that for video understanding, temporal event-level grounding and textual response are two primary tasks that exhibit distinct characteristics yet maintain strong logical dependencies. Specifically, temporal grounding focuses on precisely locating temporal events (evidence) that support answering, while textual response emphasizes accurate interpretation from the grounded evidence and generating coherent textual answers. However, existing methods [4, 13, 14, 23, 31] typically handle these two tasks in a coupled manner, with two major limitations: (1) Various special tokens are designed for temporal grounding [4, 14, 23], but their generation is mixed with text token generation without a clear logical structure, leading to coupled learning objectives. (2) More importantly, these special tokens mainly focus on timestamp representation for precise timestamp output, lacking the explicit capture of visual semantics from grounded events. In contrast, we argue that such event-level visual semantics should not be overlooked, as it could inherently serve as crucial contexts for subsequent textual answer generation, especially under the next-token prediction paradigm. + +Based on the aforementioned observations, we propose to address this from a factorized learning perspective. We first introduce a new framework $\mathrm{D}^2\mathrm{VLM}$ that decouples the learning of temporal evidence grounding and textual answering, while preserving and even strengthening their inherent dependency. Specifically, as shown in Fig. 1 (b), we decompose the model response into two sequential stages: (1) pure temporal grounding that aims to localize and capture essential visual evidence for response, followed by (2) interleaved text-evidence answer generation, where both the textual answer and temporal information are produced in an evidence-referencing manner to establish consistency with the previously grounded evidence. + +Technically, we introduce evidence tokens, a special token type dedicated to temporal evidence grounding. Different from existing designs of grounding tokens that focus on its special category and timestamp representation, our evidence tokens not only aim to determine the temporal location of the grounded event, but also emphasize the capture of event-level visual semantics, which serves as crucial context for subsequent answer generation. During the subsequent interleaved text-evidence generation, the evidence-referencing process is achieved by generating evidence tokens that align with those from the previous grounding stage. This ensures that the output information in the final response remains consistent with the initially grounded evidence while reinforcing logical coherence across stages. + +Besides providing decoupled and clearer task objectives, our design naturally fits well with the teacher-forcing autoregressive training paradigm, as subsequent textual response generation is conditioned on the correctly grounded evidence, enabling a learning shortcut for more stable training. Our experiments demonstrate that both the designed sequence generation objective and event-level visual semantic capture are essential for performance improvement, offering valuable insights for future model design. + +To further facilitate the learning of these two tasks, we propose a novel Factorized Preference Optimization (FPO) algorithm. Unlike standard preference optimization, FPO incorporates probabilistic temporal grounding modeling into the optimization objective, enabling explicit preference learning for temporal grounding in addition to standard textual preference. Meanwhile, another obstacle is that existing video preference datasets do not account for such temporal grounding aspects, making it infeasible to directly apply our proposed FPO algorithm. To overcome this limitation, we construct a synthetic dataset by introducing factorized perturbations into the original preferred response sequence. These perturbations are applied at the sub-video event level, considering two main factors: temporal grounding and textual response, along with multiple possible sub-factors. This approach ensures a structured and controllable noise generation process, where the cause and type of noise are precisely known without any manual annotation. As a result, our data synthesis process is both reliable and scalable, making it possible to effectively conduct the proposed factorized preference optimization. To the best of our knowledge, we are the first to use preference learning involving explicit temporal grounding, and moreover, in a factorized manner, as illustrated in Fig. 1 (c). + +We conduct extensive experiments to demonstrate the superiority of our proposed framework. An intuitive comparison is shown in Fig. 1 (a), where our 3.8B model consistently outperforms existing SOTA methods (typically ranging from 3.8B to 13B in model size) across various tasks. + +# 2. Related Work + +# 2.1. Video LLMs + +The recent development of video LLMs has shown great potential as generic video understanding assistants [21, 23, 25, 44], primarily due to their flexible video&text-in-text-out capability. Despite the effectiveness, existing video LLMs still struggle with accurate temporal event-level perception and localization (i.e., temporal grounding) [10, 13, 23, 31, 39], a crucial capability not only for grounding tasks but also for generating correct textual answers that are heavily based on grounding ability. + +Various efforts have been made to enhance the temporal grounding ability of video LLMs [4, 10, 11, 14]. Be + +sides adding targeted data for temporally sensitive fine-tuning [13, 20, 31, 37, 43], recent work usually designs different special tokens for precise temporal representation [10, 14, 23, 29, 36], where some work further combines such special tokens with extra grounding decoder [4, 11]. Despite their effectiveness, we argue that there still exist two potential limitations. First, in the generation sequence, special time tokens are interleaved with text tokens, leading to a highly coupled learning objective between temporal grounding and textual answering. Second, these special tokens are primarily designed for timestamp representations to output timestamp, lacking the ability to capture the visual semantics of grounded events. + +In this work, we aim to enhance temporally grounded video LLMs from a factorized learning perspective. We first decompose the task into learning of temporal evidence grounding and textual response while also strengthening their inherent dependency. We also propose a factorized preference learning paradigm to further facilitate the two-task learning. + +# 2.2. Preference Optimization + +Reinforcement learning from human feedback (RLHF) [28, 33] or from AI feedback (RLAIF) [3, 15] has been widely used to align LLMs' behavior with human preference, with PPO [27, 32, 33, 48] and DPO [30] as common implementations. While this spirit has extended to the multimodal domain [17, 35, 45], it remains underexplored for video LLMs [18, 46], especially regarding explicit temporal grounding. In this work, we introduce a factorized preference optimization (FPO) algorithm that explicitly incorporates probabilistic temporal grounding modeling into the optimization objective, enabling preference learning for both temporal grounding and textual responses. We also construct a synthetic dataset to address the lack of suitable datasets for factorized preference learning for explicit temporal grounding. + +# 3. Method + +Here we will present the proposed $\mathrm{D}^2\mathrm{VLM}$ framework. + +# 3.1. Preliminary + +We begin by first illustrating the general formulation of temporally grounded video-language models, particularly in the context of video LLMs. Given a video $V$ and a textual question (or prompt) $Q$ , the goal is to answer the question based on the given video, where the answer is typically composed of textual response together with temporal grounding information (i.e., localized temporal interval). + +For video LLMs, separate visual and text encoders will be adopted first to encode the raw inputs into a hidden space, in the format of tokens. Specifically, input video will be encoded into video tokens $F_V \in \mathbb{R}^{(T \times N_v) \times C}$ , where $T$ is the + +![](images/0d9d250db552b66f1160a86c8e187ab1a78bef520a04afaa2aa64af1439a6337.jpg) +Figure 2. Conceptual demonstration of the $\mathrm{D}^2\mathrm{VLM}$ framework. + +![](images/dd96c16028af816f3df21c01c25c7c45fc72e42131c65b35c51d2c5e7652841c.jpg) +Figure 3. The visual semantic capture process of token. + +number of frames, $N_{v}$ is the number of frame-level visual tokens, and $C$ is the number of channels. Without loss of generality, here we set $N_{v} = 1$ (i.e., $F_{V}\in \mathbb{R}^{T\times C}$ , one token per frame) for illustration purposes. Similarly, input question will be encoded into text tokens $F_{Q}\in \mathbb{R}^{N_{t}\times C}$ , where $N_{t}$ is the number of tokens. Then, both $F_{V}$ and $F_{Q}$ will feed into the video-aware LLM decoder, where the response will be generated in an autoregressive manner. + +# 3.2. $\mathbf{D}^2\mathbf{VLM}$ + +We notice that temporal event-level grounding and textual response are two primary tasks in temporally grounded video understanding tasks, where they exhibit distinct characteristics yet maintain strong logical dependencies: Temporal grounding focuses on precisely locating temporal events (evidence) that support answering, while textual response emphasizes accurate evidence interpretation and generating coherent textual answers. However, existing methods typically handle these two tasks in a coupled manner, with two major limitations: (1) Various special tokens are designed for temporal grounding, but are mixed with text token generation without a clear logical structure, leading to coupled learning objectives. (2) More importantly, these special tokens mainly focus on timestamp representation to output timestamp, lacking the explicit capture of visual semantics from grounded events. In contrast, we argue that such visual semantics can actually serve as crucial contexts for subsequent textual answer generation, especially under the next-token prediction paradigm. + +Based on this insight, we propose $\mathrm{D}^2\mathrm{VLM}$ , where our core motivation is to decouple the tied learning objective while also preserving and even strengthening the inherent dependency between these two tasks. An overall conceptual demonstration is shown in Fig. 2. Specifically, encoded video tokens $F_{V}$ and question textual tokens $F_{Q}$ are fed + +into the LLM decoder, which will generate response tokens autoregressively. Here, we enforce the video LLM to first conduct a pure grounding task before providing the actual answer, where the grounded events are expected to serve as essential evidence for the subsequent response. Then, the model will give the actual answer in an interleaved text-evidence manner which will be explained later. + +Evidence token for grounding and more. Technically, we introduce evidence token $<\mathrm{evi}>$ , a special token type dedicated to temporal grounding. Unlike existing designs that mainly focus on category and timestamp representation, the proposed $<\mathrm{evi}>$ token not only identifies the temporal location of the grounded event but also emphasizes the explicit capture of event-level visual semantics. Specifically, as shown in Fig. 3, once the video LLM generates a token classified as an $<\mathrm{evi}>$ token, we first calculate the similarity between the $<\mathrm{evi}>$ token and each frame-level LLM-processed video token $\tilde{F}_V \in \mathbb{R}^{T \times C}$ (shares the same dimension as the input $F_V$ ). Video tokens with high similarity to the $<\mathrm{evi}>$ token are then regarded as salient tokens. We aggregate the visual semantics from these salient tokens into the $<\mathrm{evi}>$ token with simple implementations (i.e., average pool the salient tokens and then add it to the $<\mathrm{evi}>$ token). Such an operation explicitly endows the $<\mathrm{evi}>$ token with rich visual semantics of the corresponding grounded event, enabling it to truly serve as solid evidence for the subsequent response based on it. Experiments in Sec. 6 demonstrate that both the design of event-level modeling and visual semantic capture are crucial for performance improvement. + +Interleaved text-evi token response. After the pure grounding stage, the generation objective switches to answer the input question based on the grounded evidence. The answer is typically composed of textual response together with temporal grounding information (e.g., "You capture the dog at 10s-15s and 50s-60s."). Here, we formulate the response objective as an interleaved text-evidence token generation process, where tokens are generated again to represent the required temporal output information in a manner akin to evidence referencing, rather than using standard textual tokens to denote timestamps only. The exact timestamps for answering can be directly obtained from the similarity calculation between the token and the frame-level video tokens $\tilde{F}_V$ , using the frame indices of the salient video tokens (Fig. 3). To make the model aware of such transition between tasks, we introduce another special token, , to indicate the end of the evidence grounding stage and encourage the model to begin generating the interleaved text-evidence answer. + +Ideally, the generated token during this stage should also match the one in the evidence grounding phase (i.e., the final response should be consistent with the previous grounded one). We introduce an explicit constraint to + +enhance this consistency: + +$$ +L _ {c o n s} = \frac {1}{K} \sum_ {k = 1} ^ {K} \left| F _ {< e v i > _ {k}} ^ {S _ {1}} - F _ {< e v i > _ {k}} ^ {S _ {2}} \right|, \tag {1} +$$ + +where $S_{1}, S_{2}$ denote the pure evidence grounding stage and interleaved text-evil response stage, respectively. $k$ is the $k^{th} < \mathrm{evi} >$ token in each stage. The $K < \mathrm{evi} >$ tokens are generated in temporal order, each aligned with a distinct interval corresponding to a ground-truth event. Such a constraint enforces strong consistency across stages, preserving logical coherence and strengthening the alignment between evidence grounding and answer generation. + +Loss function. We adopt the following components to supervise network learning: + +$$ +L = L _ {s f t} + L _ {g n d} + L _ {c o n s}, \tag {2} +$$ + +where $L_{sft}$ is the standard token classification loss for supervised fine-tuning [22, 24, 38], $L_{gnd}$ is the average grounding loss for generated tokens across both stages, with each formulated as: + +$$ +L _ {g n d} ^ {< e v i >} = \frac {1}{T} \sum_ {t = 1} ^ {T} B C E \left(y ^ {t}, s i m ^ {t}\right), \tag {3} +$$ + +where BCE is the binary cross-entropy loss, $y^{t}$ is the frame-level ground-truth (i.e., 1/0 for foreground/background), and $sim^{t}$ is the normalized dot-product similarity between token and frame-level video token feature $\tilde{F}_{V}^{t}$ . + +# 4. Factorized Preference Optimization + +To further enhance the model learning on both temporal event grounding and textual response, we introduce a novel factorized preference optimization (FPO) algorithm. Unlike standard preference optimization, FPO explicitly incorporates probabilistic temporal grounding modeling into the optimization objective, enabling explicit preference learning on both temporal grounding and textual response. Here, we start with a macroscopic formulation that captures the essence of preference optimization and gradually elaborate on our new design. + +Inspired by the existing efficient preference learning algorithm DPO [30], we adopt a similar principle: encouraging the model to assign a higher probability of generating a good response than generating a bad response, which can be defined as: + +$$ +L = - \left[ \log \sigma (\log \pi (R) - \log \pi (\bar {R})) \right], \tag {4} +$$ + +where $\pi (R)$ and $\pi (\bar{R})$ are the joint generation probabilities of preferred and dis-preferred response, respectively. Under the autoregressive paradigm, such joint generation log-probability can be further detailed as: + +$$ +\log \pi (R) = \sum_ {R _ {i} \in R} \log p \left(R _ {i} \mid V, Q, R _ {1: i - 1}\right), \tag {5} +$$ + +![](images/26907527ed2703e7b13e7315c1451f2ce4b629306d7d475f84710ade8bdebc56.jpg) +Figure 4. The proposed data synthesis pipeline, where factorized perturbation is imposed to form the dispreferred data. + +![](images/873acf6b89d7d689cad4d6c1ffb571055ddd3c59b1029f10bf5a1069777d865b.jpg) + +where $R_{i}$ is the $i^{th}$ token of the whole model response $R$ . $V$ and $Q$ are the input video and question, respectively. + +One of our goals is to perform preference optimization not only for textual responses but also with an explicit emphasis on temporal grounding. However, achieving such multi-factor preference learning presents two key challenges: (1) The lack of suitable preference data that explicitly considers temporal grounding while maintaining a factorized structure. We will elaborate more on this issue and introduce our newly proposed synthetic dataset as a solution in Sec. 5. (2) The challenge on building suitable objectives for explicit preference optimization on temporal grounding, particularly in quantifying the model's grounding probability (i.e., measuring the likelihood that the model will ground a specific temporal event). Many existing works introduce special tokens and regression heads to enhance the grounding ability of video LLMs. Despite their effectiveness, this deviates from the standard discrete text token prediction paradigm, making it non-trivial to model grounding likelihood for preference optimization. + +In this work, we notice that our designed $\mathrm{D}^2\mathrm{VLM}$ framework can naturally model such event grounding probability, in a different and explicit way compared with the standard token prediction probability. Specifically, as grounding is achieved by frame-level feature similarity calculation between the $<\mathrm{evi}>$ token and each video frame, we directly use such similarity as the frame-level grounding probability and multiply these probabilities across frames to obtain the interval-level grounding probability. Formally, for each generated $<\mathrm{evi}>_k$ token during model response, its corresponding grounding probability $p_g([s_k,e_k])$ for a given time interval $[s_k,e_k]$ ( $s_k$ , $e_k$ denote the starting and ending time index respectively) can be formulated as: + +$$ +p _ {g} ([ s _ {k}, e _ {k} ]) = \prod_ {t = 1} ^ {T} \left\{ \begin{array}{l l} s i m _ {k} ^ {t} & \text {i f} s _ {k} \leq t \leq e _ {k} \\ 1 - s i m _ {k} ^ {t} & \text {o t h e r w i s e} \end{array} \right. \tag {6} +$$ + +Here, $T$ is the total number of input video frames, $sim_k^t$ is the normalized dot-product similarity between $\langle \mathrm{evi} \rangle_k$ and + +frame-level video token feature $\tilde{F}_V^t$ . Based on such event grounding modeling, the log-probability formulation within our proposed factorized preference optimization is given as: + +$$ +\begin{array}{l} \log \pi (R) = \sum_ {R _ {i} \in R} \log p \left(R _ {i} \mid V, Q, R _ {1: i - 1}\right) \\ + \sum_ {R _ {k} = < e v i >} \log p _ {g} \left(\left[ s _ {k}, e _ {k} \right] \mid V, Q, R _ {1: k - 1}\right), \end{array} \tag {7} +$$ + +where the first term corresponds to the standard textual token prediction modeling, while the second term represents the proposed explicit modeling on temporal event grounding. By substituting Equation (7) into Equation (4), we obtain the final formulation of our Factorized Preference Optimization (FPO). Next, we explain the process of constructing multi-factor preferred $(R)$ and dis-preferred $(\bar{R})$ data pairs, which serve as the training data for FPO. + +# 5. Factorized Preference Data Synthesis + +As previously mentioned, one key challenge is the lack of suitable preference data that captures multi-factor preferences—not only for textual responses but also for explicit temporal grounding. Particularly, in the video domain, preference data is typically constructed by providing a video LLM with either complete or degraded video inputs (e.g., blurred or cropped images/videos) and treating its responses as preferred or dispreferred, respectively [18]. The underlying assumption is that the model will produce inferior responses when given degraded or incomplete visual inputs. However, existing works using this approach often face some major issues: (1) The reliability of the generated preference data is relatively less controlled. Specifically, the generated preferred sample may not actually be the preferred one, and the generated dispreferred data can sometimes be as good as or even better than the preferred data [18]. As a result, post-filtering is usually required, which involves additional models and ad-hoc human interventions, increasing costs but still cannot fully ensure + +quality. (2) This lack of control also makes it difficult to properly factorize the generated dispreferred data, including identifying the source or the reason why it is considered dispreferred. (3) Last but not least, existing video preference data mainly focuses on generating responses in pure textual format (e.g., video caption) [18, 46], while generating preference response that explicitly contains temporal grounding information (e.g., temporal event position) remains under-explored. Overall, these issues make factorized preference optimization difficult to achieve. + +To address the aforementioned limitations, we propose an automatic factorized preference data synthesis framework, which introduces factorized perturbations on the original response at the sub-video event level. As shown in Fig. 4 (a), the overall process consists of the following steps: + +Random selection of factors and sub-video events. For each given sample (video-question-answer pair), we first randomly select a factor that determines the exact type of the perturbation. As shown in Fig. 4 (b), the factors are broadly categorized into two types and each can be further divided into multiple sub-types: (1) Temporal grounding: This type of perturbation aims to reveal/simulate model response inaccuracies related to temporal event grounding. It can be further divided into sub-types, including temporal localization shift, randomly adding/deleting grounded events (corresponding to false positives and missed detections), and merging multiple events into one (corresponding to a lack of fine-grained distinction in event boundaries). (2) Textual response: This type of perturbation modifies the semantic correctness of the textual response. It includes sub-types such as distorting key information, which disrupts critical content, and repeating responses, a common failure mode observed in video LLMs. + +Next, we randomly select some sub-video events to apply factorized perturbation. A sub-video event refers to a semantically meaningful segment of the video that can be determined based on the given sample (e.g., its question or ground-truth answer). For instance, in a dense video captioning task, an event naturally corresponds to a specific captioned interval, while in an action localization task, an event can align with a single action instance. Without loss of generality, we take the dense video caption task as an illustration example (as shown in Fig. 4 (c)), as this task emphasizes both temporal grounding and textual description. Overall, such an event-level design aligns naturally with the inherent video structure, ensuring that perturbations are introduced in a way that respects the video's temporal and semantic coherence. Moreover, it enables fine-grained perturbation, allowing controlled modifications at the event level rather than applying coarse, less precisely controlled noise to the entire response sequence. + +Injecting factorized perturbations. Once the factors and sub-video events are determined, localized perturbation is + +introduced based on the selected factors and sub-video events. If a temporal grounding factor is selected, we modify the timestamps of the selected events based on the determined sub-factor (e.g., shift the interval, delete the interval, etc.). For the factor of textual response, for each selected event, we prompt off-the-shelf LLM (e.g., Qwen [41, 42]) to generate a distorted response based on the original correct event-level response. If the "repeat" sub-factor is selected, we can easily achieve this by setting the response of the current event to its previous one to simulate such a response repeating phenomenon. A synthesized example is shown in Fig. 4 (c). After the perturbation synthesis, we combine the newly obtained noisy response with the original one to form the preference sample pair for factorized preference optimization. In practice, our preference data is synthesized based on the E.T. Instruct 164K dataset [23]. + +Discussion. Compared to existing preference data synthesis approaches, our framework ensures that the generated dispreferred response contains specific, well-factorized noise rather than arbitrary, unpredictable degradation, and thus we can better identify the noise source without manual costs. To our knowledge, we are also the first to emphasize explicit temporal grounding in preference data, and moreover, in a multi-factorized manner. The overall design makes our factorized preference optimization in Sec. 4 feasible. + +# 6. Experiments + +# 6.1. Experimental Setup + +Implementation details. Following the practice of [23], we adopt the pre-trained ViT-G/14 from EVA-CLIP [8, 34], followed by a Q-Former-like feature compressor [6, 16, 23], and use Phi-3-Mini-3.8B [1] as the base LLM. Also following [23], our model is initialized from the stage-2 model of [23] and [19], and we also only fine-tune our model on the E.T. Instruct 164K dataset [23] using LoRA [12]. We conduct training on $4 \times$ NVIDIA H100 GPUs within 1 day. More details are provided in the supplementary material. + +Evaluation tasks and datasets. We evaluate our method across various datasets and tasks. (1) E.T. Bench Grounding [23]. It consists of 5 different grounding-related tasks (e.g., moment retrieval, action localization, summarization/highlight detection, etc.) for comprehensive evaluation. (2) E.T. Bench Dense Captioning [23]. It is composed of tasks like dense video caption and step localization, where both key event grounding and textual description are required. Besides, we also conduct experiments on two widely used individual datasets, (3) Charades-STA [9] and (4) YouCook2 [47], for independent evaluation on moment retrieval and dense video captioning, respectively. + +
MethodYearE.T. Bench GroundingE.T. Bench Dense Captioning
TVG\( _{F1} \)EPM\( _{F1} \)TAL\( _{F1} \)EVS\( _{F1} \)VHD\( _{F1} \)Avg\( _{F1} \)DVC\( _{F1} \)DVC\( _{Sim} \)SLC\( _{F1} \)SLC\( _{Sim} \)Avg\( _{F1} \)Avg\( _{Sim} \)
Video-ChatGPT-7B [26]ACL'247.01.315.18.428.812.18.811.35.710.27.310.8
Video-LLaVA-7B [21]EMNLP'247.01.915.00.328.910.628.015.00.98.314.511.7
LLaMA-VID-7B [19]ECCV'245.51.28.01.430.09.227.112.65.211.116.211.9
Video-LLaMA-2-7B [5]arXiv'240.10.00.00.01.50.30.614.50.015.20.314.9
PLLAVA [40]arXiv'246.91.15.70.328.98.613.310.69.711.811.511.2
VTimeLLM-7B [13]CVPR'247.61.918.215.928.914.512.413.18.76.410.69.8
TimeChat-7B [31]CVPR'2426.23.910.129.140.522.016.612.55.69.211.110.9
LITA-13B [14]ECCV'2422.24.618.029.723.919.739.717.221.012.230.414.7
VTG-LLM-7B [10]AAAI'2515.93.714.426.848.221.840.218.620.814.430.516.5
Qwen2.5-VL-7B* [2]arXiv'2546.69.332.219.968.635.339.021.425.515.132.218.2
E.T.Chat-3.8B [23]NeurIPS'2438.610.230.825.462.533.538.419.724.414.631.417.2
D²VLM-3.8B (Ours)ICCV'2560.214.433.435.268.242.348.325.526.718.137.521.8
+ +Table 1. Performance comparison on E.T. Bench-Grounding dataset (5 sub tasks: TVG: Temporal Video Grounding; EPM: Episodic Memory; TAL: Temporal Action Localization; EVS: Extractive Video Summarization; VHD: Video Highlight Detection) and E.T. Bench-Dense Captioning dataset (2 sub tasks: DVC: Dense Video Captioning; SLC: Step Localization and Captioning). \*: Implemented by us, with more details in the supplementary material. + +
MethodYearR@1(IoU=0.5)R@1(IoU=0.7)
TimeChat-7B [31]CVPR’2432.213.4
VTimeLLM-7B [13]CVPR’2427.511.4
VTimeLLM-13B [13]CVPR’2434.314.7
Momentor-7B [29]ICML’2426.611.6
HawkEye-7B [37]arXiv’2431.414.5
VTG-LLM-7B [10]AAAI’2533.815.7
TRACE-7B [11]ICLR’2540.319.4
VideoChat-T-7B [43]ICLR’2548.724.0
E.T.Chat-3.8B [23]NeurIPS’2445.920.0
D²VLM-3.8B (Ours)ICCV’2550.326.0
+ +Table 2. Performance comparison on Charades-STA. + +
MethodYearF1CIDErSODA_c
TimeChat-7B [31]CVPR’2412.63.41.2
VTG-LLM-7B [10]AAAI’2517.55.01.5
TRACE-7B [11]ICLR’2522.48.12.2
D²VLM-3.8B (Ours)ICCV’2526.410.63.2
+ +Table 3. Performance comparison on YouCook2. + +# 6.2. Main Benchmarking Results + +E.T. Bench grounding. The results are shown in the left part of Tab. 1, where we compare our method with the recently published SOTA methods. It can be observed that the proposed $\mathrm{D}^2\mathrm{VLM}$ significantly outperforms recently published SOTA methods across various tasks, achieving an average F1 score improvement of at least $7.0\%$ . Meanwhile, our model is relatively lightweight (3.8B parameters) compared to most state-of-the-art methods (ranging from 3.8B to 13B, with the majority being 7B parameters). This further demonstrates the superiority of our proposed method. + +E.T. Bench dense captioning. The performance comparison is shown in the right part of Tab. 1. It can be seen that our method also shows noticeable superiority in dense captioning-related tasks, where it outperforms recent SOTA + +methods in both temporal event grounding (at least $5.3\%$ on average F1 score) and textual description (at least $3.6\%$ on average Sim score). This essentially demonstrates the generalizability of our method across various scenarios. + +Charades-STA. From Tab. 2, it can be seen that our method still outperforms recent SOTA methods by large margins. Specifically, our 3.8B model outperforms counterpart [23] with the same base model by $4.4\%$ on $\mathrm{R}@\mathrm{1(IoU = 0.5)}$ , and outperforms methods published in 2025 with 7B model size [10, 11, 31] by at least $1.6\%$ on $\mathrm{R}@\mathrm{1(IoU = 0.5)}$ . This further demonstrates the superiority of our method in both effectiveness and efficiency. + +YouCook2. As shown in Tab. 3, our method achieves state-of-the-art performance compared to recent SOTA approaches. Specifically, it outperforms them by at least $4\%$ in temporal event grounding (F1) while also demonstrating superior textual description capacity, with gains of at least $2.5\%$ on CIDEr and $1.0\%$ on SODA.c. This further highlights the superiority of our approach in both temporal grounding and text generation. + +# 6.3. Ablation Studies + +Here we conduct ablation studies on E.T. Bench grounding and E.T. Bench Dense Captioning datasets to analyze the effect of the proposed components, in both temporal event-level grounding and textual description. For each dataset (i.e., grounding, dense captioning), we report the average performance across multiple sub-tasks. In the following, we first analyze the component effect of the newly designed generation objective and evidence token, where we do not take the proposed FPO into account. Finally, we analyze the effect of the proposed FPO. + +Generation objective of $\mathbf{D}^2\mathbf{VLM}$ . Here, we analyze the effectiveness of our newly designed generation objective within $\mathbf{D}^2\mathbf{VLM}$ . From Tab. 4, it can be seen that: (1) Row-1 + +
Decomposed ObjectiveInterleaved Text-avi GenerationConsistency ConstraintGrounding AvgF1Dense Captioning AvgF1AvgSim
21.214.311.3
✓28.923.116.0
✓✓35.634.319.8
✓✓✓39.535.021.2
+ +Table 4. Component effect of the designed generation objective. + +
FPOGrounding AvgF1Dense Captioning AvgF1AvgSim
×39.535.021.2
✓42.337.521.8
+ +Table 5. Effect of the proposed FPO. + +
MethodGrounding +AvgF1Dense +AvgF1Captioning +AvgSim
w/o event-level modeling26.133.416.2
w/o visual semantic capture37.127.517.7
Full design39.535.021.2
+ +Table 6. Component effect within evidence token design. + +v.s. Row-2: Decompose the coupled grounding-textual response objective into grounding (by our designed token) then answer from grounded event can substantially improve the performance in both grounding (i.e., $7.7\%$ averaged F1 on grounding tasks and $8.8\%$ averaged F1 on dense captioning tasks) and textual description (i.e., $4.7\%$ averaged sentence similarity on dense captioning tasks). (2) Row-2 v.s. Row-3: When we shift the answering stage from pure textual token generation into interleaved text-evi token generation, the performance can be further enhanced by large margins (at least $6.7\%$ on grounding and $3.8\%$ on textual description). The reason is that under such cases, the overall generation objective can be interpreted as grounding then answer with evidence referencing, where such referencing strengthens the relationship between event grounding and textual answering. (3) Row-3 v.s. Row-4: The introduced consistency constraint further boosts the performance by a noticeable margin, as it explicitly enhances the consistency between the two stages. Overall, the results validate the effectiveness of our design. + +Design of evidence token. Here we analyze the impact of the core components within the token design (i.e., the event-level modeling and the explicit visual semantic capture). The results are shown in Tab. 6, where w/o event-level modeling means each token only captures the information of individual frame (a common design in existing works [14, 23, 29]), rather than an interval-level event (e.g., 10s-15s). w/o visual semantic capture indicates visual semantics from the video tokens is not explicitly integrated into the token through the operation illustrated in Sec. 3.2 (also a common design in most of the existing works [4, 14, 29]). From Tab. 6, it can be summarized that: (1) The performance drops significantly without event-level modeling. We argue that this is because compared to individual timestamp-level modeling, event-level modeling more naturally aligns with the characteristics of + +event-level evidence grounding and can capture more robust event-level semantics in a global perspective. (2) Explicit visual semantic capture is also crucial for enhancing model effectiveness, as it enables tokens to truly serve as solid context for subsequent answer generation, which also fits well with the autoregressive generation process. Notably, compared with grounding-only tasks, such an effect is more pronounced on dense captioning tasks, as textual description generation can truly benefit from the captured event-level visual semantics. + +Effect of factorized preference optimization. Tab. 5 demonstrates that the proposed FPO algorithm consistently enhances performance on both temporal grounding and captioning-related tasks, with a more significant improvement in temporal grounding. This indeed validates the effectiveness of our factorized optimization objective design. + +# 7. Conclusion and Limitations + +In this work, we uncover some key limitations of existing temporal grounded video-language models: (1) the use of coupled learning objective for temporal grounding and textual response generation, and (2) the overlooking of inherent dependency between these two tasks, especially the event-level visual semantics from the intermediate grounding result. We tackle this from a factorized learning perspective, where we propose $\mathrm{D}^2\mathrm{VLM}$ that decomposes the generation objective into a "grounding then answering with evidence referencing" paradigm, and introduce evidence tokens to emphasize explicit visual semantic capture beyond timestamp representation. Furthermore, we design a factorized preference learning algorithm, coupled with a factorized synthetic dataset, to enhance the learning of both tasks. Extensive experiments demonstrate the clear superiority of our approach across various scenarios. + +Despite achieving new state-of-the-art results, our approach still has room for improvement (e.g., F1 scores of only $14.4\%$ on the episodic memory task and $26.4\%$ on the YouCook2 dense video captioning task). Besides, while our factorized data generation offers good factorized controllability and directly contributes to factorized optimization, it focuses only on generating negative (dis-preferred) samples. However, generating more diverse positive samples to further enrich the preference data is also meaningful, and we leave this for future work. + +# Acknowledgment + +# References + +[1] Marah Abdin, Sam Ade Jacobs, Ammar Ahmad Awan, Jyoti Aneja, Ahmed Awadallah, Hany Awadalla, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Harkirat Behl, et al. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219, 2024. 6 +[2] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. 7 +[3] Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, et al. Constitutional ai: Harmlessness from ai feedback. arXiv preprint arXiv:2212.08073, 2022. 3 +[4] Qirui Chen, Shangzhe Di, and Weidi Xie. Grounded multi-hop videoqa in long-form egocentric videos. arXiv preprint arXiv:2408.14469, 2024. 2, 3, 8 +[5] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 7 +[6] Wenliang Dai, Junnan Li, DONGXU LI, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale N Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. In Advances in Neural Information Processing Systems, pages 49250-49267. Curran Associates, Inc., 2023. 6 +[7] Shangzhe Di and Weidi Xie. Grounded question-answering in long egocentric videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12934–12943, 2024. 1 +[8] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19358-19369, 2023. 6 +[9] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In Proceedings of the IEEE international conference on computer vision, pages 5267-5275, 2017. 1, 6 +[10] Yongxin Guo, Jingyu Liu, Mingda Li, Dingxin Cheng, Xiaoying Tang, Dianbo Sui, Qingbin Liu, Xi Chen, and Kevin Zhao. Vtg-llm: Integrating timestamp knowledge into video + +This research is supported by the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No: AISG3-RP-2022-030). We would like to acknowledge that computational work involved in this research work is partially supported by NUS IT's Research Computing group under grant number NUSREC-HPC-00001. +llms for enhanced video temporal grounding. arXiv preprint arXiv:2405.13382, 2024. 1, 2, 3, 7 +[11] Yongxin Guo, Jingyu Liu, Mingda Li, Xiaoying Tang, Qingbin Liu, and Xi Chen. Trace: Temporal grounding video llm via causal event modeling. In International Conference on Learning Representations, 2025. 2, 3, 7 +[12] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6 +[13] Bin Huang, Xin Wang, Hong Chen, Zihan Song, and Wenwu Zhu. Vtimellm: Empower llm to grasp video moments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14271-14280, 2024. 2, 3, 7 +[14] De-An Huang, Shijia Liao, Subhashree Radhakrishnan, Hongxu Yin, Pavlo Molchanov, Zhiding Yu, and Jan Kautz. Lita: Language instructed temporal-localization assistant. In European Conference on Computer Vision, pages 202-218. Springer, 2024. 2, 3, 7, 8 +[15] Harrison Lee, Samrat Phatale, Hassan Mansoor, Thomas Mesnard, Johan Ferret, Kellie Lu, Colton Bishop, Ethan Hall, Victor Carbune, Abhinav Rastogi, et al. Rlaif vs. rlhf: Scaling reinforcement learning from human feedback with ai feedback. arXiv preprint arXiv:2309.00267, 2023. 3 +[16] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 6 +[17] Lei Li, Zhihui Xie, Mukai Li, Shunian Chen, Peiyi Wang, Liang Chen, Yazheng Yang, Benyou Wang, and Lingpeng Kong. Silkie: Preference distillation for large visual language models. arXiv preprint arXiv:2312.10665, 2023. 3 +[18] Rui Li, Xiaohan Wang, Yuhui Zhang, Zeyu Wang, and Serena Yeung-Levy. Temporal preference optimization for long-form video understanding. arXiv preprint arXiv:2501.13919, 2025. 3, 5, 6 +[19] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In European Conference on Computer Vision, pages 323–340. Springer, 2024. 6, 7 +[20] Zhaowei Li, Qi Xu, Dong Zhang, Hang Song, YiQing Cai, Qi Qi, Ran Zhou, Junting Pan, Zefeng Li, Vu Tu, et al. Groundinggpt: Language enhanced multi-modal grounding model. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6657-6678, 2024. 3 +[21] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 5971–5984, 2024. 1, 2, 7 +[22] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2023. 4 +[23] Ye Liu, Zongyang Ma, Zhongang Qi, Yang Wu, Ying Shan, and Chang Wen Chen. E.T. Bench: Towards open-ended + +event-level video-language understanding. In The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2024. 1, 2, 3, 6, 7, 8 +[24] Shayne Longpre, Le Hou, Tu Vu, Albert Webson, Hyung Won Chung, Yi Tay, Denny Zhou, Quoc V Le, Barret Zoph, Jason Wei, et al. The flan collection: Designing data and methods for effective instruction tuning. In International Conference on Machine Learning, pages 22631-22648. PMLR, 2023. 4 +[25] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12585–12602, 2024. 1, 2 +[26] Munan Ning, Bin Zhu, Yujia Xie, Bin Lin, Jiaxi Cui, Lu Yuan, Dongdong Chen, and Li Yuan. Video-bench: A comprehensive benchmark and toolkit for evaluating video-based large language models. arXiv preprint arXiv:2311.16103, 2023. 7 +[27] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 3 +[28] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 3 +[29] Long Qian, Juncheng Li, Yu Wu, Yaobo Ye, Hao Fei, TatSeng Chua, Yueting Zhuang, and Siliang Tang. Momentor: Advancing video large language model with fine-grained temporal reasoning. arXiv preprint arXiv:2402.11435, 2024. 3, 7, 8 +[30] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36:53728-53741, 2023. 3, 4 +[31] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14313-14323, 2024. 1, 2, 3, 7 +[32] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 3 +[33] Nisan Stiannon, Long Ouyang, Jeffrey Wu, Daniel Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford, Dario Amodei, and Paul F Christiano. Learning to summarize with human feedback. Advances in neural information processing systems, 33:3008-3021, 2020. 3 +[34] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 6 + +[35] Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liang-Yan Gui, Yu-Xiong Wang, Yiming Yang, et al. Aligning large multimodal models with factually augmented rlhf. In Annual Meeting of the Association for Computational Linguistics, 2024. 3 +[36] Haibo Wang, Zhiyang Xu, Yu Cheng, Shizhe Diao, Yufan Zhou, Yixin Cao, Qifan Wang, Weifeng Ge, and Lifu Huang. Grounded-videollm: Sharpening fine-grained temporal grounding in video large language models. arXiv preprint arXiv:2410.03290, 2024. 3 +[37] Yueqian Wang, Xiaojun Meng, Jianxin Liang, Yuxuan Wang, Qun Liu, and Dongyan Zhao. Hawkeye: Training video-text llms for grounding text in videos. arXiv preprint arXiv:2403.10228, 2024. 3, 7 +[38] Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652, 2021. 4 +[39] Junbin Xiao, Angela Yao, Yicong Li, and Tat-Seng Chua. Can i trust your answer? visually grounded video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13204-13214, 2024. 2 +[40] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 7 +[41] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng He, Junyang Lin, Kai Dang, Keming Lu, Keqin Chen, Kexin Yang, Mei Li, Mingfeng Xue, Na Ni, Pei Zhang, Peng Wang, Ru Peng, Rui Men, Ruize Gao, Runji Lin, Shijie Wang, Shuai Bai, Sinan Tan, Tianhang Zhu, Tianhao Li, Tianyu Liu, Wenbin Ge, Xiaodong Deng, Xiaohuan Zhou, Xingzhang Ren, Xinyu Zhang, Xipin Wei, Xuancheng Ren, Yang Fan, Yang Yao, Yichang Zhang, Yu Wan, Yunfei Chu, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, and Zhihao Fan. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 6 +[42] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, Le Yu, Mei Li, Mingfeng Xue, Pei Zhang, Qin Zhu, Rui Men, Runji Lin, Tianhao Li, Tingyu Xia, Xingzhang Ren, Xuancheng Ren, Yang Fan, Yang Su, Yichang Zhang, Yu Wan, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, and Zihan Qiu. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115, 2024. 6 +[43] Xiangyu Zeng, Kunchang Li, Chenting Wang, Xinhao Li, Tianxiang Jiang, Ziang Yan, Songze Li, Yansong Shi, Zhengrong Yue, Yi Wang, et al. Timesuite: Improving mllms for long video understanding via grounded tuning. In International Conference on Learning Representations, 2025. 3, 7 + +[44] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 1, 2 +[45] Mengxi Zhang, Wenhao Wu, Yu Lu, Yuxin Song, Kang Rong, Huanjin Yao, Jianbo Zhao, Fanglong Liu, Haocheng Feng, Jingdong Wang, et al. Automated multi-level preference for mllms. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 3 +[46] Ruohong Zhang, Liangke Gui, Zhiqing Sun, Yihao Feng, Keyang Xu, Yuanhan Zhang, Di Fu, Chunyuan Li, Alexander Hauptmann, Yonatan Bisk, et al. Direct preference optimization of video large multimodal models from language model reward. arXiv preprint arXiv:2404.01258, 2024. 3, 6 +[47] Luowei Zhou, Chenliang Xu, and Jason Corso. Towards automatic learning of procedures from web instructional videos. In Proceedings of the AAAI Conference on Artificial Intelligence, 2018. 1, 6 +[48] Daniel M Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. Fine-tuning language models from human preferences, 2020. URL https://arxiv.org/abs, page 14, 1909. 3 \ No newline at end of file diff --git a/2025/Factorized Learning for Temporally Grounded Video-Language Models/images.zip b/2025/Factorized Learning for Temporally Grounded Video-Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f18843871c89c98abf49d976e879e4d4122ed9ff --- /dev/null +++ b/2025/Factorized Learning for Temporally Grounded Video-Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:88f19f96c71888c0cbcfb45ddd9fc8ae9b1fb904de3f2fc397bb8b09b497b858 +size 425008 diff --git a/2025/Factorized Learning for Temporally Grounded Video-Language Models/layout.json b/2025/Factorized Learning for Temporally Grounded Video-Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1a96008b20f372e04e824b4c21b6020258112769 --- /dev/null +++ b/2025/Factorized Learning for Temporally Grounded Video-Language Models/layout.json @@ -0,0 +1,7845 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 84, + 103, + 526, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 103, + 526, + 121 + ], + "spans": [ + { + "bbox": [ + 84, + 103, + 526, + 121 + ], + "type": "text", + "content": "Factorized Learning for Temporally Grounded Video-Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 151, + 142, + 459, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 142, + 459, + 172 + ], + "spans": [ + { + "bbox": [ + 151, + 142, + 459, + 172 + ], + "type": "text", + "content": "Wenzheng Zeng, Difei Gao, Mike Zheng Shou†, Hwee Tou Ng† National University of Singapore" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 174, + 535, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 174, + 535, + 186 + ], + "spans": [ + { + "bbox": [ + 69, + 174, + 535, + 186 + ], + "type": "text", + "content": "wenzhengzeng@u.nus.edu, {daniel.difei.gao, mike.zheng.shou}@gmail.com, nght@nus.edu.sg" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 58, + 221, + 255, + 378 + ], + "blocks": [ + { + "bbox": [ + 59, + 214, + 70, + 224 + ], + "lines": [ + { + "bbox": [ + 59, + 214, + 70, + 224 + ], + "spans": [ + { + "bbox": [ + 59, + 214, + 70, + 224 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 221, + 255, + 378 + ], + "lines": [ + { + "bbox": [ + 58, + 221, + 255, + 378 + ], + "spans": [ + { + "bbox": [ + 58, + 221, + 255, + 378 + ], + "type": "image", + "image_path": "734db46a50c501461c381372d16c4057a9a6ff5a1a51c7bb6694bb24042975d2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 259, + 213, + 550, + 380 + ], + "blocks": [ + { + "bbox": [ + 259, + 213, + 550, + 380 + ], + "lines": [ + { + "bbox": [ + 259, + 213, + 550, + 380 + ], + "spans": [ + { + "bbox": [ + 259, + 213, + 550, + 380 + ], + "type": "image", + "image_path": "2ee64aa051f07996dc9f92f80ba9a130050ff53409093b1ceb61ef052113fb3d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 384, + 555, + 441 + ], + "lines": [ + { + "bbox": [ + 54, + 384, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 54, + 384, + 555, + 441 + ], + "type": "text", + "content": "Figure 1. (a) Performance: Our method consistently outperforms SOTA methods across various tasks (here we draw the maximum performance across methods, detailed in Sec. 6). (b) Model: We propose a new framework " + }, + { + "bbox": [ + 54, + 384, + 555, + 441 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^2\\mathrm{VLM}" + }, + { + "bbox": [ + 54, + 384, + 555, + 441 + ], + "type": "text", + "content": ", where we decompose the generation objective into a \"grounding then answering with evidence referencing\" paradigm and introduce evidence tokens to emphasize explicit event-level visual semantic capture. (c) Training Algorithm: We introduce Factorized Preference Optimization (FPO) that explicitly addresses both temporal grounding and textual response. A factorized data synthesis approach is also designed to support FPO." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 462, + 200, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 462, + 200, + 475 + ], + "spans": [ + { + "bbox": [ + 151, + 462, + 200, + 475 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 487, + 297, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 487, + 297, + 704 + ], + "spans": [ + { + "bbox": [ + 53, + 487, + 297, + 704 + ], + "type": "text", + "content": "Recent video-language models have shown great potential for video understanding, but still struggle with accurate temporal grounding for event-level perception. We observe that two main factors in video understanding (i.e., temporal grounding and textual response) form a logical hierarchy: accurate temporal evidence grounding lays the foundation for reliable textual response. However, existing works typically handle these two tasks in a coupled manner without a clear logical structure, leading to sub-optimal objectives. We address this from a factorized learning perspective. We first propose " + }, + { + "bbox": [ + 53, + 487, + 297, + 704 + ], + "type": "inline_equation", + "content": "D^{2}" + }, + { + "bbox": [ + 53, + 487, + 297, + 704 + ], + "type": "text", + "content": " VLM, a framework that decouples the learning of these two tasks while also emphasizing their inherent dependency. We adopt a \"grounding then answering with evidence referencing\" paradigm and introduce evidence tokens for evidence grounding, which emphasize event-level visual semantic capture beyond the focus on timestamp representation in existing works. To further facilitate the learning of these two tasks, we introduce a novel factorized" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 451, + 556, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 556, + 571 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 556, + 571 + ], + "type": "text", + "content": "preference optimization (FPO) algorithm. Unlike standard preference optimization, FPO explicitly incorporates probabilistic temporal grounding modeling into the optimization objective, enabling preference learning for both temporal grounding and textual response. We also construct a synthetic dataset to address the lack of suitable datasets for factorized preference learning with explicit temporal grounding. Experiments on various tasks demonstrate the clear advantage of our approach. Our source code is available at https://github.com/nusnlp/d2v1m." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 597, + 395, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 597, + 395, + 609 + ], + "spans": [ + { + "bbox": [ + 314, + 597, + 395, + 609 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "Recent advances in video-language models, especially those built upon large language models (video LLMs), have enabled remarkable progress in video understanding [10, 21, 23, 25, 31, 44]. Through their flexible video&text-int-text-out nature, video LLMs demonstrate great potential as general-purpose solvers, unifying various tasks (e.g., temporal grounding [9], dense captioning [47], and question answering [7]) into one generic \"video question answer-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 712, + 150, + 722 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 712, + 150, + 722 + ], + "spans": [ + { + "bbox": [ + 70, + 712, + 150, + 722 + ], + "type": "text", + "content": "†Corresponding authors" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "spans": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "type": "text", + "content": "20683" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "ing\" framework. Despite their effectiveness, existing video LLMs still struggle with accurate temporal grounding for event perception and localization [10, 11, 23, 31, 31], a crucial capability not only for grounding tasks themselves but also for relevant tasks that need textual answering." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 135, + 295, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 135, + 295, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 135, + 295, + 386 + ], + "type": "text", + "content": "We notice that for video understanding, temporal event-level grounding and textual response are two primary tasks that exhibit distinct characteristics yet maintain strong logical dependencies. Specifically, temporal grounding focuses on precisely locating temporal events (evidence) that support answering, while textual response emphasizes accurate interpretation from the grounded evidence and generating coherent textual answers. However, existing methods [4, 13, 14, 23, 31] typically handle these two tasks in a coupled manner, with two major limitations: (1) Various special tokens are designed for temporal grounding [4, 14, 23], but their generation is mixed with text token generation without a clear logical structure, leading to coupled learning objectives. (2) More importantly, these special tokens mainly focus on timestamp representation for precise timestamp output, lacking the explicit capture of visual semantics from grounded events. In contrast, we argue that such event-level visual semantics should not be overlooked, as it could inherently serve as crucial contexts for subsequent textual answer generation, especially under the next-token prediction paradigm." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 388, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 295, + 544 + ], + "type": "text", + "content": "Based on the aforementioned observations, we propose to address this from a factorized learning perspective. We first introduce a new framework " + }, + { + "bbox": [ + 55, + 388, + 295, + 544 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^2\\mathrm{VLM}" + }, + { + "bbox": [ + 55, + 388, + 295, + 544 + ], + "type": "text", + "content": " that decouples the learning of temporal evidence grounding and textual answering, while preserving and even strengthening their inherent dependency. Specifically, as shown in Fig. 1 (b), we decompose the model response into two sequential stages: (1) pure temporal grounding that aims to localize and capture essential visual evidence for response, followed by (2) interleaved text-evidence answer generation, where both the textual answer and temporal information are produced in an evidence-referencing manner to establish consistency with the previously grounded evidence." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "Technically, we introduce evidence tokens, a special token type dedicated to temporal evidence grounding. Different from existing designs of grounding tokens that focus on its special category and timestamp representation, our evidence tokens not only aim to determine the temporal location of the grounded event, but also emphasize the capture of event-level visual semantics, which serves as crucial context for subsequent answer generation. During the subsequent interleaved text-evidence generation, the evidence-referencing process is achieved by generating evidence tokens that align with those from the previous grounding stage. This ensures that the output information in the final response remains consistent with the initially grounded evidence while reinforcing logical coherence across stages." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": "Besides providing decoupled and clearer task objectives, our design naturally fits well with the teacher-forcing autoregressive training paradigm, as subsequent textual response generation is conditioned on the correctly grounded evidence, enabling a learning shortcut for more stable training. Our experiments demonstrate that both the designed sequence generation objective and event-level visual semantic capture are essential for performance improvement, offering valuable insights for future model design." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 181, + 553, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 181, + 553, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 181, + 553, + 467 + ], + "type": "text", + "content": "To further facilitate the learning of these two tasks, we propose a novel Factorized Preference Optimization (FPO) algorithm. Unlike standard preference optimization, FPO incorporates probabilistic temporal grounding modeling into the optimization objective, enabling explicit preference learning for temporal grounding in addition to standard textual preference. Meanwhile, another obstacle is that existing video preference datasets do not account for such temporal grounding aspects, making it infeasible to directly apply our proposed FPO algorithm. To overcome this limitation, we construct a synthetic dataset by introducing factorized perturbations into the original preferred response sequence. These perturbations are applied at the sub-video event level, considering two main factors: temporal grounding and textual response, along with multiple possible sub-factors. This approach ensures a structured and controllable noise generation process, where the cause and type of noise are precisely known without any manual annotation. As a result, our data synthesis process is both reliable and scalable, making it possible to effectively conduct the proposed factorized preference optimization. To the best of our knowledge, we are the first to use preference learning involving explicit temporal grounding, and moreover, in a factorized manner, as illustrated in Fig. 1 (c)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 468, + 553, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 553, + 529 + ], + "type": "text", + "content": "We conduct extensive experiments to demonstrate the superiority of our proposed framework. An intuitive comparison is shown in Fig. 1 (a), where our 3.8B model consistently outperforms existing SOTA methods (typically ranging from 3.8B to 13B in model size) across various tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 542, + 400, + 554 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 542, + 400, + 554 + ], + "spans": [ + { + "bbox": [ + 314, + 542, + 400, + 554 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 562, + 397, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 562, + 397, + 574 + ], + "spans": [ + { + "bbox": [ + 314, + 562, + 397, + 574 + ], + "type": "text", + "content": "2.1. Video LLMs" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 580, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 553, + 689 + ], + "type": "text", + "content": "The recent development of video LLMs has shown great potential as generic video understanding assistants [21, 23, 25, 44], primarily due to their flexible video&text-in-text-out capability. Despite the effectiveness, existing video LLMs still struggle with accurate temporal event-level perception and localization (i.e., temporal grounding) [10, 13, 23, 31, 39], a crucial capability not only for grounding tasks but also for generating correct textual answers that are heavily based on grounding ability." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Various efforts have been made to enhance the temporal grounding ability of video LLMs [4, 10, 11, 14]. Be" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "spans": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "type": "text", + "content": "20684" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 227 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 227 + ], + "type": "text", + "content": "sides adding targeted data for temporally sensitive fine-tuning [13, 20, 31, 37, 43], recent work usually designs different special tokens for precise temporal representation [10, 14, 23, 29, 36], where some work further combines such special tokens with extra grounding decoder [4, 11]. Despite their effectiveness, we argue that there still exist two potential limitations. First, in the generation sequence, special time tokens are interleaved with text tokens, leading to a highly coupled learning objective between temporal grounding and textual answering. Second, these special tokens are primarily designed for timestamp representations to output timestamp, lacking the ability to capture the visual semantics of grounded events." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 228, + 294, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 294, + 312 + ], + "type": "text", + "content": "In this work, we aim to enhance temporally grounded video LLMs from a factorized learning perspective. We first decompose the task into learning of temporal evidence grounding and textual response while also strengthening their inherent dependency. We also propose a factorized preference learning paradigm to further facilitate the two-task learning." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 317, + 194, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 317, + 194, + 330 + ], + "spans": [ + { + "bbox": [ + 55, + 317, + 194, + 330 + ], + "type": "text", + "content": "2.2. Preference Optimization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 335, + 295, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 295, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 295, + 515 + ], + "type": "text", + "content": "Reinforcement learning from human feedback (RLHF) [28, 33] or from AI feedback (RLAIF) [3, 15] has been widely used to align LLMs' behavior with human preference, with PPO [27, 32, 33, 48] and DPO [30] as common implementations. While this spirit has extended to the multimodal domain [17, 35, 45], it remains underexplored for video LLMs [18, 46], especially regarding explicit temporal grounding. In this work, we introduce a factorized preference optimization (FPO) algorithm that explicitly incorporates probabilistic temporal grounding modeling into the optimization objective, enabling preference learning for both temporal grounding and textual responses. We also construct a synthetic dataset to address the lack of suitable datasets for factorized preference learning for explicit temporal grounding." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 526, + 111, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 526, + 111, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 111, + 538 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 546, + 279, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 279, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 279, + 559 + ], + "type": "text", + "content": "Here we will present the proposed " + }, + { + "bbox": [ + 55, + 546, + 279, + 559 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^2\\mathrm{VLM}" + }, + { + "bbox": [ + 55, + 546, + 279, + 559 + ], + "type": "text", + "content": " framework." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 564, + 134, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 134, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 134, + 577 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 294, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 294, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 294, + 665 + ], + "type": "text", + "content": "We begin by first illustrating the general formulation of temporally grounded video-language models, particularly in the context of video LLMs. Given a video " + }, + { + "bbox": [ + 55, + 582, + 294, + 665 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 582, + 294, + 665 + ], + "type": "text", + "content": " and a textual question (or prompt) " + }, + { + "bbox": [ + 55, + 582, + 294, + 665 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 582, + 294, + 665 + ], + "type": "text", + "content": ", the goal is to answer the question based on the given video, where the answer is typically composed of textual response together with temporal grounding information (i.e., localized temporal interval)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 294, + 713 + ], + "type": "text", + "content": "For video LLMs, separate visual and text encoders will be adopted first to encode the raw inputs into a hidden space, in the format of tokens. Specifically, input video will be encoded into video tokens " + }, + { + "bbox": [ + 55, + 665, + 294, + 713 + ], + "type": "inline_equation", + "content": "F_V \\in \\mathbb{R}^{(T \\times N_v) \\times C}" + }, + { + "bbox": [ + 55, + 665, + 294, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 665, + 294, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 665, + 294, + 713 + ], + "type": "text", + "content": " is the" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 321, + 69, + 547, + 150 + ], + "blocks": [ + { + "bbox": [ + 321, + 69, + 547, + 150 + ], + "lines": [ + { + "bbox": [ + 321, + 69, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 321, + 69, + 547, + 150 + ], + "type": "image", + "image_path": "0d9d250db552b66f1160a86c8e187ab1a78bef520a04afaa2aa64af1439a6337.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 154, + 549, + 166 + ], + "lines": [ + { + "bbox": [ + 318, + 154, + 549, + 166 + ], + "spans": [ + { + "bbox": [ + 318, + 154, + 549, + 166 + ], + "type": "text", + "content": "Figure 2. Conceptual demonstration of the " + }, + { + "bbox": [ + 318, + 154, + 549, + 166 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^2\\mathrm{VLM}" + }, + { + "bbox": [ + 318, + 154, + 549, + 166 + ], + "type": "text", + "content": " framework." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 339, + 177, + 529, + 251 + ], + "blocks": [ + { + "bbox": [ + 339, + 177, + 529, + 251 + ], + "lines": [ + { + "bbox": [ + 339, + 177, + 529, + 251 + ], + "spans": [ + { + "bbox": [ + 339, + 177, + 529, + 251 + ], + "type": "image", + "image_path": "dd96c16028af816f3df21c01c25c7c45fc72e42131c65b35c51d2c5e7652841c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 256, + 549, + 267 + ], + "lines": [ + { + "bbox": [ + 318, + 256, + 549, + 267 + ], + "spans": [ + { + "bbox": [ + 318, + 256, + 549, + 267 + ], + "type": "text", + "content": "Figure 3. The visual semantic capture process of token." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 289, + 554, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 554, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 554, + 385 + ], + "type": "text", + "content": "number of frames, " + }, + { + "bbox": [ + 313, + 289, + 554, + 385 + ], + "type": "inline_equation", + "content": "N_{v}" + }, + { + "bbox": [ + 313, + 289, + 554, + 385 + ], + "type": "text", + "content": " is the number of frame-level visual tokens, and " + }, + { + "bbox": [ + 313, + 289, + 554, + 385 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 289, + 554, + 385 + ], + "type": "text", + "content": " is the number of channels. Without loss of generality, here we set " + }, + { + "bbox": [ + 313, + 289, + 554, + 385 + ], + "type": "inline_equation", + "content": "N_{v} = 1" + }, + { + "bbox": [ + 313, + 289, + 554, + 385 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 313, + 289, + 554, + 385 + ], + "type": "inline_equation", + "content": "F_{V}\\in \\mathbb{R}^{T\\times C}" + }, + { + "bbox": [ + 313, + 289, + 554, + 385 + ], + "type": "text", + "content": ", one token per frame) for illustration purposes. Similarly, input question will be encoded into text tokens " + }, + { + "bbox": [ + 313, + 289, + 554, + 385 + ], + "type": "inline_equation", + "content": "F_{Q}\\in \\mathbb{R}^{N_{t}\\times C}" + }, + { + "bbox": [ + 313, + 289, + 554, + 385 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 289, + 554, + 385 + ], + "type": "inline_equation", + "content": "N_{t}" + }, + { + "bbox": [ + 313, + 289, + 554, + 385 + ], + "type": "text", + "content": " is the number of tokens. Then, both " + }, + { + "bbox": [ + 313, + 289, + 554, + 385 + ], + "type": "inline_equation", + "content": "F_{V}" + }, + { + "bbox": [ + 313, + 289, + 554, + 385 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 289, + 554, + 385 + ], + "type": "inline_equation", + "content": "F_{Q}" + }, + { + "bbox": [ + 313, + 289, + 554, + 385 + ], + "type": "text", + "content": " will feed into the video-aware LLM decoder, where the response will be generated in an autoregressive manner." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 395, + 375, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 395, + 375, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 395, + 375, + 406 + ], + "type": "text", + "content": "3.2. " + }, + { + "bbox": [ + 313, + 395, + 375, + 406 + ], + "type": "inline_equation", + "content": "\\mathbf{D}^2\\mathbf{VLM}" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 413, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 555, + 640 + ], + "type": "text", + "content": "We notice that temporal event-level grounding and textual response are two primary tasks in temporally grounded video understanding tasks, where they exhibit distinct characteristics yet maintain strong logical dependencies: Temporal grounding focuses on precisely locating temporal events (evidence) that support answering, while textual response emphasizes accurate evidence interpretation and generating coherent textual answers. However, existing methods typically handle these two tasks in a coupled manner, with two major limitations: (1) Various special tokens are designed for temporal grounding, but are mixed with text token generation without a clear logical structure, leading to coupled learning objectives. (2) More importantly, these special tokens mainly focus on timestamp representation to output timestamp, lacking the explicit capture of visual semantics from grounded events. In contrast, we argue that such visual semantics can actually serve as crucial contexts for subsequent textual answer generation, especially under the next-token prediction paradigm." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "Based on this insight, we propose " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^2\\mathrm{VLM}" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": ", where our core motivation is to decouple the tied learning objective while also preserving and even strengthening the inherent dependency between these two tasks. An overall conceptual demonstration is shown in Fig. 2. Specifically, encoded video tokens " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "F_{V}" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": " and question textual tokens " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "F_{Q}" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": " are fed" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "spans": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "type": "text", + "content": "20685" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "into the LLM decoder, which will generate response tokens autoregressively. Here, we enforce the video LLM to first conduct a pure grounding task before providing the actual answer, where the grounded events are expected to serve as essential evidence for the subsequent response. Then, the model will give the actual answer in an interleaved text-evidence manner which will be explained later." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 159, + 296, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 159, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 56, + 159, + 296, + 434 + ], + "type": "text", + "content": "Evidence token for grounding and more. Technically, we introduce evidence token " + }, + { + "bbox": [ + 56, + 159, + 296, + 434 + ], + "type": "inline_equation", + "content": "<\\mathrm{evi}>" + }, + { + "bbox": [ + 56, + 159, + 296, + 434 + ], + "type": "text", + "content": ", a special token type dedicated to temporal grounding. Unlike existing designs that mainly focus on category and timestamp representation, the proposed " + }, + { + "bbox": [ + 56, + 159, + 296, + 434 + ], + "type": "inline_equation", + "content": "<\\mathrm{evi}>" + }, + { + "bbox": [ + 56, + 159, + 296, + 434 + ], + "type": "text", + "content": " token not only identifies the temporal location of the grounded event but also emphasizes the explicit capture of event-level visual semantics. Specifically, as shown in Fig. 3, once the video LLM generates a token classified as an " + }, + { + "bbox": [ + 56, + 159, + 296, + 434 + ], + "type": "inline_equation", + "content": "<\\mathrm{evi}>" + }, + { + "bbox": [ + 56, + 159, + 296, + 434 + ], + "type": "text", + "content": " token, we first calculate the similarity between the " + }, + { + "bbox": [ + 56, + 159, + 296, + 434 + ], + "type": "inline_equation", + "content": "<\\mathrm{evi}>" + }, + { + "bbox": [ + 56, + 159, + 296, + 434 + ], + "type": "text", + "content": " token and each frame-level LLM-processed video token " + }, + { + "bbox": [ + 56, + 159, + 296, + 434 + ], + "type": "inline_equation", + "content": "\\tilde{F}_V \\in \\mathbb{R}^{T \\times C}" + }, + { + "bbox": [ + 56, + 159, + 296, + 434 + ], + "type": "text", + "content": " (shares the same dimension as the input " + }, + { + "bbox": [ + 56, + 159, + 296, + 434 + ], + "type": "inline_equation", + "content": "F_V" + }, + { + "bbox": [ + 56, + 159, + 296, + 434 + ], + "type": "text", + "content": "). Video tokens with high similarity to the " + }, + { + "bbox": [ + 56, + 159, + 296, + 434 + ], + "type": "inline_equation", + "content": "<\\mathrm{evi}>" + }, + { + "bbox": [ + 56, + 159, + 296, + 434 + ], + "type": "text", + "content": " token are then regarded as salient tokens. We aggregate the visual semantics from these salient tokens into the " + }, + { + "bbox": [ + 56, + 159, + 296, + 434 + ], + "type": "inline_equation", + "content": "<\\mathrm{evi}>" + }, + { + "bbox": [ + 56, + 159, + 296, + 434 + ], + "type": "text", + "content": " token with simple implementations (i.e., average pool the salient tokens and then add it to the " + }, + { + "bbox": [ + 56, + 159, + 296, + 434 + ], + "type": "inline_equation", + "content": "<\\mathrm{evi}>" + }, + { + "bbox": [ + 56, + 159, + 296, + 434 + ], + "type": "text", + "content": " token). Such an operation explicitly endows the " + }, + { + "bbox": [ + 56, + 159, + 296, + 434 + ], + "type": "inline_equation", + "content": "<\\mathrm{evi}>" + }, + { + "bbox": [ + 56, + 159, + 296, + 434 + ], + "type": "text", + "content": " token with rich visual semantics of the corresponding grounded event, enabling it to truly serve as solid evidence for the subsequent response based on it. Experiments in Sec. 6 demonstrate that both the design of event-level modeling and visual semantic capture are crucial for performance improvement." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 436, + 295, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 295, + 662 + ], + "type": "text", + "content": "Interleaved text-evi token response. After the pure grounding stage, the generation objective switches to answer the input question based on the grounded evidence. The answer is typically composed of textual response together with temporal grounding information (e.g., \"You capture the dog at 10s-15s and 50s-60s.\"). Here, we formulate the response objective as an interleaved text-evidence token generation process, where tokens are generated again to represent the required temporal output information in a manner akin to evidence referencing, rather than using standard textual tokens to denote timestamps only. The exact timestamps for answering can be directly obtained from the similarity calculation between the token and the frame-level video tokens " + }, + { + "bbox": [ + 55, + 436, + 295, + 662 + ], + "type": "inline_equation", + "content": "\\tilde{F}_V" + }, + { + "bbox": [ + 55, + 436, + 295, + 662 + ], + "type": "text", + "content": ", using the frame indices of the salient video tokens (Fig. 3). To make the model aware of such transition between tasks, we introduce another special token, , to indicate the end of the evidence grounding stage and encourage the model to begin generating the interleaved text-evidence answer." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": "Ideally, the generated token during this stage should also match the one in the evidence grounding phase (i.e., the final response should be consistent with the previous grounded one). We introduce an explicit constraint to" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 73, + 419, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 73, + 419, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 73, + 419, + 84 + ], + "type": "text", + "content": "enhance this consistency:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 353, + 86, + 553, + 118 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 86, + 553, + 118 + ], + "spans": [ + { + "bbox": [ + 353, + 86, + 553, + 118 + ], + "type": "interline_equation", + "content": "L _ {c o n s} = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\left| F _ {< e v i > _ {k}} ^ {S _ {1}} - F _ {< e v i > _ {k}} ^ {S _ {2}} \\right|, \\tag {1}", + "image_path": "2f7d437ce206fff24c3038dacfd188cd0d3ae4b9af8e800aded732f726003e75.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 121, + 553, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 553, + 217 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 121, + 553, + 217 + ], + "type": "inline_equation", + "content": "S_{1}, S_{2}" + }, + { + "bbox": [ + 313, + 121, + 553, + 217 + ], + "type": "text", + "content": " denote the pure evidence grounding stage and interleaved text-evil response stage, respectively. " + }, + { + "bbox": [ + 313, + 121, + 553, + 217 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 121, + 553, + 217 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 313, + 121, + 553, + 217 + ], + "type": "inline_equation", + "content": "k^{th} < \\mathrm{evi} >" + }, + { + "bbox": [ + 313, + 121, + 553, + 217 + ], + "type": "text", + "content": " token in each stage. The " + }, + { + "bbox": [ + 313, + 121, + 553, + 217 + ], + "type": "inline_equation", + "content": "K < \\mathrm{evi} >" + }, + { + "bbox": [ + 313, + 121, + 553, + 217 + ], + "type": "text", + "content": " tokens are generated in temporal order, each aligned with a distinct interval corresponding to a ground-truth event. Such a constraint enforces strong consistency across stages, preserving logical coherence and strengthening the alignment between evidence grounding and answer generation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 217, + 553, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 553, + 240 + ], + "type": "text", + "content": "Loss function. We adopt the following components to supervise network learning:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 377, + 245, + 553, + 259 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 245, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 377, + 245, + 553, + 259 + ], + "type": "interline_equation", + "content": "L = L _ {s f t} + L _ {g n d} + L _ {c o n s}, \\tag {2}", + "image_path": "571cd4e30b0e91417ea2f8072ff77716acbcc2545b510428425ef563d2464911.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 265, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 265, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 553, + 312 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 265, + 553, + 312 + ], + "type": "inline_equation", + "content": "L_{sft}" + }, + { + "bbox": [ + 313, + 265, + 553, + 312 + ], + "type": "text", + "content": " is the standard token classification loss for supervised fine-tuning [22, 24, 38], " + }, + { + "bbox": [ + 313, + 265, + 553, + 312 + ], + "type": "inline_equation", + "content": "L_{gnd}" + }, + { + "bbox": [ + 313, + 265, + 553, + 312 + ], + "type": "text", + "content": " is the average grounding loss for generated tokens across both stages, with each formulated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 361, + 314, + 553, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 314, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 361, + 314, + 553, + 346 + ], + "type": "interline_equation", + "content": "L _ {g n d} ^ {< e v i >} = \\frac {1}{T} \\sum_ {t = 1} ^ {T} B C E \\left(y ^ {t}, s i m ^ {t}\\right), \\tag {3}", + "image_path": "4024abc6ce991458d96d118841797335a6d76010d945809055dcb95a43fdf1e2.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 350, + 553, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 553, + 399 + ], + "type": "text", + "content": "where BCE is the binary cross-entropy loss, " + }, + { + "bbox": [ + 313, + 350, + 553, + 399 + ], + "type": "inline_equation", + "content": "y^{t}" + }, + { + "bbox": [ + 313, + 350, + 553, + 399 + ], + "type": "text", + "content": " is the frame-level ground-truth (i.e., 1/0 for foreground/background), and " + }, + { + "bbox": [ + 313, + 350, + 553, + 399 + ], + "type": "inline_equation", + "content": "sim^{t}" + }, + { + "bbox": [ + 313, + 350, + 553, + 399 + ], + "type": "text", + "content": " is the normalized dot-product similarity between token and frame-level video token feature " + }, + { + "bbox": [ + 313, + 350, + 553, + 399 + ], + "type": "inline_equation", + "content": "\\tilde{F}_{V}^{t}" + }, + { + "bbox": [ + 313, + 350, + 553, + 399 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 407, + 512, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 512, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 512, + 422 + ], + "type": "text", + "content": "4. Factorized Preference Optimization" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 428, + 554, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 428, + 554, + 547 + ], + "spans": [ + { + "bbox": [ + 313, + 428, + 554, + 547 + ], + "type": "text", + "content": "To further enhance the model learning on both temporal event grounding and textual response, we introduce a novel factorized preference optimization (FPO) algorithm. Unlike standard preference optimization, FPO explicitly incorporates probabilistic temporal grounding modeling into the optimization objective, enabling explicit preference learning on both temporal grounding and textual response. Here, we start with a macroscopic formulation that captures the essence of preference optimization and gradually elaborate on our new design." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 548, + 554, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 554, + 607 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 554, + 607 + ], + "type": "text", + "content": "Inspired by the existing efficient preference learning algorithm DPO [30], we adopt a similar principle: encouraging the model to assign a higher probability of generating a good response than generating a bad response, which can be defined as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 351, + 613, + 553, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 613, + 553, + 628 + ], + "spans": [ + { + "bbox": [ + 351, + 613, + 553, + 628 + ], + "type": "interline_equation", + "content": "L = - \\left[ \\log \\sigma (\\log \\pi (R) - \\log \\pi (\\bar {R})) \\right], \\tag {4}", + "image_path": "0d9e4eea7eaaf38036f1b22ebe792f16ecf99e594713f34187c32baf86368d70.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 635, + 553, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 553, + 683 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 553, + 683 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 635, + 553, + 683 + ], + "type": "inline_equation", + "content": "\\pi (R)" + }, + { + "bbox": [ + 313, + 635, + 553, + 683 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 635, + 553, + 683 + ], + "type": "inline_equation", + "content": "\\pi (\\bar{R})" + }, + { + "bbox": [ + 313, + 635, + 553, + 683 + ], + "type": "text", + "content": " are the joint generation probabilities of preferred and dis-preferred response, respectively. Under the autoregressive paradigm, such joint generation log-probability can be further detailed as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 345, + 689, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 689, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 345, + 689, + 553, + 715 + ], + "type": "interline_equation", + "content": "\\log \\pi (R) = \\sum_ {R _ {i} \\in R} \\log p \\left(R _ {i} \\mid V, Q, R _ {1: i - 1}\\right), \\tag {5}", + "image_path": "7d07e0786aa25ba59c0a215c3149e5568ee52a70d061ad256fadf99b6b6cecb9.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "spans": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "type": "text", + "content": "20686" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 70, + 287, + 215 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 287, + 215 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 287, + 215 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 287, + 215 + ], + "type": "image", + "image_path": "26907527ed2703e7b13e7315c1451f2ce4b629306d7d475f84710ade8bdebc56.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 220, + 518, + 232 + ], + "lines": [ + { + "bbox": [ + 88, + 220, + 518, + 232 + ], + "spans": [ + { + "bbox": [ + 88, + 220, + 518, + 232 + ], + "type": "text", + "content": "Figure 4. The proposed data synthesis pipeline, where factorized perturbation is imposed to form the dispreferred data." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 291, + 70, + 529, + 215 + ], + "blocks": [ + { + "bbox": [ + 291, + 70, + 529, + 215 + ], + "lines": [ + { + "bbox": [ + 291, + 70, + 529, + 215 + ], + "spans": [ + { + "bbox": [ + 291, + 70, + 529, + 215 + ], + "type": "image", + "image_path": "873acf6b89d7d689cad4d6c1ffb571055ddd3c59b1029f10bf5a1069777d865b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 252, + 294, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 252, + 294, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 294, + 277 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 252, + 294, + 277 + ], + "type": "inline_equation", + "content": "R_{i}" + }, + { + "bbox": [ + 55, + 252, + 294, + 277 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 55, + 252, + 294, + 277 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 55, + 252, + 294, + 277 + ], + "type": "text", + "content": " token of the whole model response " + }, + { + "bbox": [ + 55, + 252, + 294, + 277 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 252, + 294, + 277 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 252, + 294, + 277 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 252, + 294, + 277 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 252, + 294, + 277 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 252, + 294, + 277 + ], + "type": "text", + "content": " are the input video and question, respectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 277, + 295, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 277, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 54, + 277, + 295, + 491 + ], + "type": "text", + "content": "One of our goals is to perform preference optimization not only for textual responses but also with an explicit emphasis on temporal grounding. However, achieving such multi-factor preference learning presents two key challenges: (1) The lack of suitable preference data that explicitly considers temporal grounding while maintaining a factorized structure. We will elaborate more on this issue and introduce our newly proposed synthetic dataset as a solution in Sec. 5. (2) The challenge on building suitable objectives for explicit preference optimization on temporal grounding, particularly in quantifying the model's grounding probability (i.e., measuring the likelihood that the model will ground a specific temporal event). Many existing works introduce special tokens and regression heads to enhance the grounding ability of video LLMs. Despite their effectiveness, this deviates from the standard discrete text token prediction paradigm, making it non-trivial to model grounding likelihood for preference optimization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 491, + 295, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 295, + 647 + ], + "type": "text", + "content": "In this work, we notice that our designed " + }, + { + "bbox": [ + 55, + 491, + 295, + 647 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^2\\mathrm{VLM}" + }, + { + "bbox": [ + 55, + 491, + 295, + 647 + ], + "type": "text", + "content": " framework can naturally model such event grounding probability, in a different and explicit way compared with the standard token prediction probability. Specifically, as grounding is achieved by frame-level feature similarity calculation between the " + }, + { + "bbox": [ + 55, + 491, + 295, + 647 + ], + "type": "inline_equation", + "content": "<\\mathrm{evi}>" + }, + { + "bbox": [ + 55, + 491, + 295, + 647 + ], + "type": "text", + "content": " token and each video frame, we directly use such similarity as the frame-level grounding probability and multiply these probabilities across frames to obtain the interval-level grounding probability. Formally, for each generated " + }, + { + "bbox": [ + 55, + 491, + 295, + 647 + ], + "type": "inline_equation", + "content": "<\\mathrm{evi}>_k" + }, + { + "bbox": [ + 55, + 491, + 295, + 647 + ], + "type": "text", + "content": " token during model response, its corresponding grounding probability " + }, + { + "bbox": [ + 55, + 491, + 295, + 647 + ], + "type": "inline_equation", + "content": "p_g([s_k,e_k])" + }, + { + "bbox": [ + 55, + 491, + 295, + 647 + ], + "type": "text", + "content": " for a given time interval " + }, + { + "bbox": [ + 55, + 491, + 295, + 647 + ], + "type": "inline_equation", + "content": "[s_k,e_k]" + }, + { + "bbox": [ + 55, + 491, + 295, + 647 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 491, + 295, + 647 + ], + "type": "inline_equation", + "content": "s_k" + }, + { + "bbox": [ + 55, + 491, + 295, + 647 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 491, + 295, + 647 + ], + "type": "inline_equation", + "content": "e_k" + }, + { + "bbox": [ + 55, + 491, + 295, + 647 + ], + "type": "text", + "content": " denote the starting and ending time index respectively) can be formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 78, + 652, + 295, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 652, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 78, + 652, + 295, + 685 + ], + "type": "interline_equation", + "content": "p _ {g} ([ s _ {k}, e _ {k} ]) = \\prod_ {t = 1} ^ {T} \\left\\{ \\begin{array}{l l} s i m _ {k} ^ {t} & \\text {i f} s _ {k} \\leq t \\leq e _ {k} \\\\ 1 - s i m _ {k} ^ {t} & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {6}", + "image_path": "3dfe9c77df8837a274e38d9417ff22bca13c7c6709c5b0015c3e688e6b3fe367.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " is the total number of input video frames, " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "sim_k^t" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " is the normalized dot-product similarity between " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{evi} \\rangle_k" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 252, + 555, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 252, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 252, + 555, + 289 + ], + "type": "text", + "content": "frame-level video token feature " + }, + { + "bbox": [ + 313, + 252, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\tilde{F}_V^t" + }, + { + "bbox": [ + 313, + 252, + 555, + 289 + ], + "type": "text", + "content": ". Based on such event grounding modeling, the log-probability formulation within our proposed factorized preference optimization is given as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 331, + 303, + 553, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 303, + 553, + 357 + ], + "spans": [ + { + "bbox": [ + 331, + 303, + 553, + 357 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\log \\pi (R) = \\sum_ {R _ {i} \\in R} \\log p \\left(R _ {i} \\mid V, Q, R _ {1: i - 1}\\right) \\\\ + \\sum_ {R _ {k} = < e v i >} \\log p _ {g} \\left(\\left[ s _ {k}, e _ {k} \\right] \\mid V, Q, R _ {1: k - 1}\\right), \\end{array} \\tag {7}", + "image_path": "bfc73ecde94241341fbfba085aa2ffb9bc15d938ae051575f9f19891fa9c6f93.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 361, + 555, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 555, + 457 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 555, + 457 + ], + "type": "text", + "content": "where the first term corresponds to the standard textual token prediction modeling, while the second term represents the proposed explicit modeling on temporal event grounding. By substituting Equation (7) into Equation (4), we obtain the final formulation of our Factorized Preference Optimization (FPO). Next, we explain the process of constructing multi-factor preferred " + }, + { + "bbox": [ + 313, + 361, + 555, + 457 + ], + "type": "inline_equation", + "content": "(R)" + }, + { + "bbox": [ + 313, + 361, + 555, + 457 + ], + "type": "text", + "content": " and dis-preferred " + }, + { + "bbox": [ + 313, + 361, + 555, + 457 + ], + "type": "inline_equation", + "content": "(\\bar{R})" + }, + { + "bbox": [ + 313, + 361, + 555, + 457 + ], + "type": "text", + "content": " data pairs, which serve as the training data for FPO." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 466, + 520, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 466, + 520, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 466, + 520, + 479 + ], + "type": "text", + "content": "5. Factorized Preference Data Synthesis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": "As previously mentioned, one key challenge is the lack of suitable preference data that captures multi-factor preferences—not only for textual responses but also for explicit temporal grounding. Particularly, in the video domain, preference data is typically constructed by providing a video LLM with either complete or degraded video inputs (e.g., blurred or cropped images/videos) and treating its responses as preferred or dispreferred, respectively [18]. The underlying assumption is that the model will produce inferior responses when given degraded or incomplete visual inputs. However, existing works using this approach often face some major issues: (1) The reliability of the generated preference data is relatively less controlled. Specifically, the generated preferred sample may not actually be the preferred one, and the generated dispreferred data can sometimes be as good as or even better than the preferred data [18]. As a result, post-filtering is usually required, which involves additional models and ad-hoc human interventions, increasing costs but still cannot fully ensure" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "spans": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "type": "text", + "content": "20687" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": "quality. (2) This lack of control also makes it difficult to properly factorize the generated dispreferred data, including identifying the source or the reason why it is considered dispreferred. (3) Last but not least, existing video preference data mainly focuses on generating responses in pure textual format (e.g., video caption) [18, 46], while generating preference response that explicitly contains temporal grounding information (e.g., temporal event position) remains under-explored. Overall, these issues make factorized preference optimization difficult to achieve." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 194, + 294, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 194, + 294, + 254 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 294, + 254 + ], + "type": "text", + "content": "To address the aforementioned limitations, we propose an automatic factorized preference data synthesis framework, which introduces factorized perturbations on the original response at the sub-video event level. As shown in Fig. 4 (a), the overall process consists of the following steps:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 255, + 295, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 295, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 295, + 470 + ], + "type": "text", + "content": "Random selection of factors and sub-video events. For each given sample (video-question-answer pair), we first randomly select a factor that determines the exact type of the perturbation. As shown in Fig. 4 (b), the factors are broadly categorized into two types and each can be further divided into multiple sub-types: (1) Temporal grounding: This type of perturbation aims to reveal/simulate model response inaccuracies related to temporal event grounding. It can be further divided into sub-types, including temporal localization shift, randomly adding/deleting grounded events (corresponding to false positives and missed detections), and merging multiple events into one (corresponding to a lack of fine-grained distinction in event boundaries). (2) Textual response: This type of perturbation modifies the semantic correctness of the textual response. It includes sub-types such as distorting key information, which disrupts critical content, and repeating responses, a common failure mode observed in video LLMs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 472, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 687 + ], + "type": "text", + "content": "Next, we randomly select some sub-video events to apply factorized perturbation. A sub-video event refers to a semantically meaningful segment of the video that can be determined based on the given sample (e.g., its question or ground-truth answer). For instance, in a dense video captioning task, an event naturally corresponds to a specific captioned interval, while in an action localization task, an event can align with a single action instance. Without loss of generality, we take the dense video caption task as an illustration example (as shown in Fig. 4 (c)), as this task emphasizes both temporal grounding and textual description. Overall, such an event-level design aligns naturally with the inherent video structure, ensuring that perturbations are introduced in a way that respects the video's temporal and semantic coherence. Moreover, it enables fine-grained perturbation, allowing controlled modifications at the event level rather than applying coarse, less precisely controlled noise to the entire response sequence." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Injecting factorized perturbations. Once the factors and sub-video events are determined, localized perturbation is" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "type": "text", + "content": "introduced based on the selected factors and sub-video events. If a temporal grounding factor is selected, we modify the timestamps of the selected events based on the determined sub-factor (e.g., shift the interval, delete the interval, etc.). For the factor of textual response, for each selected event, we prompt off-the-shelf LLM (e.g., Qwen [41, 42]) to generate a distorted response based on the original correct event-level response. If the \"repeat\" sub-factor is selected, we can easily achieve this by setting the response of the current event to its previous one to simulate such a response repeating phenomenon. A synthesized example is shown in Fig. 4 (c). After the perturbation synthesis, we combine the newly obtained noisy response with the original one to form the preference sample pair for factorized preference optimization. In practice, our preference data is synthesized based on the E.T. Instruct 164K dataset [23]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 270, + 553, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 553, + 376 + ], + "type": "text", + "content": "Discussion. Compared to existing preference data synthesis approaches, our framework ensures that the generated dispreferred response contains specific, well-factorized noise rather than arbitrary, unpredictable degradation, and thus we can better identify the noise source without manual costs. To our knowledge, we are also the first to emphasize explicit temporal grounding in preference data, and moreover, in a multi-factorized manner. The overall design makes our factorized preference optimization in Sec. 4 feasible." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 407, + 394, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 407, + 394, + 420 + ], + "spans": [ + { + "bbox": [ + 314, + 407, + 394, + 420 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 432, + 430, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 432, + 430, + 445 + ], + "spans": [ + { + "bbox": [ + 314, + 432, + 430, + 445 + ], + "type": "text", + "content": "6.1. Experimental Setup" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 456, + 553, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 553, + 564 + ], + "type": "text", + "content": "Implementation details. Following the practice of [23], we adopt the pre-trained ViT-G/14 from EVA-CLIP [8, 34], followed by a Q-Former-like feature compressor [6, 16, 23], and use Phi-3-Mini-3.8B [1] as the base LLM. Also following [23], our model is initialized from the stage-2 model of [23] and [19], and we also only fine-tune our model on the E.T. Instruct 164K dataset [23] using LoRA [12]. We conduct training on " + }, + { + "bbox": [ + 313, + 456, + 553, + 564 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 313, + 456, + 553, + 564 + ], + "type": "text", + "content": " NVIDIA H100 GPUs within 1 day. More details are provided in the supplementary material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "type": "text", + "content": "Evaluation tasks and datasets. We evaluate our method across various datasets and tasks. (1) E.T. Bench Grounding [23]. It consists of 5 different grounding-related tasks (e.g., moment retrieval, action localization, summarization/highlight detection, etc.) for comprehensive evaluation. (2) E.T. Bench Dense Captioning [23]. It is composed of tasks like dense video caption and step localization, where both key event grounding and textual description are required. Besides, we also conduct experiments on two widely used individual datasets, (3) Charades-STA [9] and (4) YouCook2 [47], for independent evaluation on moment retrieval and dense video captioning, respectively." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "spans": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "type": "text", + "content": "20688" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 553, + 223 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 223 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 223 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 223 + ], + "type": "table", + "html": "
MethodYearE.T. Bench GroundingE.T. Bench Dense Captioning
TVG\\( _{F1} \\)EPM\\( _{F1} \\)TAL\\( _{F1} \\)EVS\\( _{F1} \\)VHD\\( _{F1} \\)Avg\\( _{F1} \\)DVC\\( _{F1} \\)DVC\\( _{Sim} \\)SLC\\( _{F1} \\)SLC\\( _{Sim} \\)Avg\\( _{F1} \\)Avg\\( _{Sim} \\)
Video-ChatGPT-7B [26]ACL'247.01.315.18.428.812.18.811.35.710.27.310.8
Video-LLaVA-7B [21]EMNLP'247.01.915.00.328.910.628.015.00.98.314.511.7
LLaMA-VID-7B [19]ECCV'245.51.28.01.430.09.227.112.65.211.116.211.9
Video-LLaMA-2-7B [5]arXiv'240.10.00.00.01.50.30.614.50.015.20.314.9
PLLAVA [40]arXiv'246.91.15.70.328.98.613.310.69.711.811.511.2
VTimeLLM-7B [13]CVPR'247.61.918.215.928.914.512.413.18.76.410.69.8
TimeChat-7B [31]CVPR'2426.23.910.129.140.522.016.612.55.69.211.110.9
LITA-13B [14]ECCV'2422.24.618.029.723.919.739.717.221.012.230.414.7
VTG-LLM-7B [10]AAAI'2515.93.714.426.848.221.840.218.620.814.430.516.5
Qwen2.5-VL-7B* [2]arXiv'2546.69.332.219.968.635.339.021.425.515.132.218.2
E.T.Chat-3.8B [23]NeurIPS'2438.610.230.825.462.533.538.419.724.414.631.417.2
D²VLM-3.8B (Ours)ICCV'2560.214.433.435.268.242.348.325.526.718.137.521.8
", + "image_path": "006892a719e3c860c8f997e69a7d55e25ffff3d3762246f740e71314cce4a1a8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 289, + 294, + 398 + ], + "blocks": [ + { + "bbox": [ + 55, + 226, + 555, + 271 + ], + "lines": [ + { + "bbox": [ + 55, + 226, + 555, + 271 + ], + "spans": [ + { + "bbox": [ + 55, + 226, + 555, + 271 + ], + "type": "text", + "content": "Table 1. Performance comparison on E.T. Bench-Grounding dataset (5 sub tasks: TVG: Temporal Video Grounding; EPM: Episodic Memory; TAL: Temporal Action Localization; EVS: Extractive Video Summarization; VHD: Video Highlight Detection) and E.T. Bench-Dense Captioning dataset (2 sub tasks: DVC: Dense Video Captioning; SLC: Step Localization and Captioning). \\*: Implemented by us, with more details in the supplementary material." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 289, + 294, + 398 + ], + "lines": [ + { + "bbox": [ + 58, + 289, + 294, + 398 + ], + "spans": [ + { + "bbox": [ + 58, + 289, + 294, + 398 + ], + "type": "table", + "html": "
MethodYearR@1(IoU=0.5)R@1(IoU=0.7)
TimeChat-7B [31]CVPR’2432.213.4
VTimeLLM-7B [13]CVPR’2427.511.4
VTimeLLM-13B [13]CVPR’2434.314.7
Momentor-7B [29]ICML’2426.611.6
HawkEye-7B [37]arXiv’2431.414.5
VTG-LLM-7B [10]AAAI’2533.815.7
TRACE-7B [11]ICLR’2540.319.4
VideoChat-T-7B [43]ICLR’2548.724.0
E.T.Chat-3.8B [23]NeurIPS’2445.920.0
D²VLM-3.8B (Ours)ICCV’2550.326.0
", + "image_path": "ca89a56b151c617137533935c0cadb863b3eb967270ec20c0b1652098d5a6089.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 425, + 294, + 490 + ], + "blocks": [ + { + "bbox": [ + 79, + 402, + 271, + 412 + ], + "lines": [ + { + "bbox": [ + 79, + 402, + 271, + 412 + ], + "spans": [ + { + "bbox": [ + 79, + 402, + 271, + 412 + ], + "type": "text", + "content": "Table 2. Performance comparison on Charades-STA." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 425, + 294, + 490 + ], + "lines": [ + { + "bbox": [ + 58, + 425, + 294, + 490 + ], + "spans": [ + { + "bbox": [ + 58, + 425, + 294, + 490 + ], + "type": "table", + "html": "
MethodYearF1CIDErSODA_c
TimeChat-7B [31]CVPR’2412.63.41.2
VTG-LLM-7B [10]AAAI’2517.55.01.5
TRACE-7B [11]ICLR’2522.48.12.2
D²VLM-3.8B (Ours)ICCV’2526.410.63.2
", + "image_path": "c06aab560e4321fc74ac473c8957c8f989838a8f0f3284849beb37c29c414850.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 86, + 493, + 265, + 504 + ], + "lines": [ + { + "bbox": [ + 86, + 493, + 265, + 504 + ], + "spans": [ + { + "bbox": [ + 86, + 493, + 265, + 504 + ], + "type": "text", + "content": "Table 3. Performance comparison on YouCook2." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 526, + 212, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 526, + 212, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 212, + 540 + ], + "type": "text", + "content": "6.2. Main Benchmarking Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 544, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 665 + ], + "type": "text", + "content": "E.T. Bench grounding. The results are shown in the left part of Tab. 1, where we compare our method with the recently published SOTA methods. It can be observed that the proposed " + }, + { + "bbox": [ + 55, + 544, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^2\\mathrm{VLM}" + }, + { + "bbox": [ + 55, + 544, + 295, + 665 + ], + "type": "text", + "content": " significantly outperforms recently published SOTA methods across various tasks, achieving an average F1 score improvement of at least " + }, + { + "bbox": [ + 55, + 544, + 295, + 665 + ], + "type": "inline_equation", + "content": "7.0\\%" + }, + { + "bbox": [ + 55, + 544, + 295, + 665 + ], + "type": "text", + "content": ". Meanwhile, our model is relatively lightweight (3.8B parameters) compared to most state-of-the-art methods (ranging from 3.8B to 13B, with the majority being 7B parameters). This further demonstrates the superiority of our proposed method." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "E.T. Bench dense captioning. The performance comparison is shown in the right part of Tab. 1. It can be seen that our method also shows noticeable superiority in dense captioning-related tasks, where it outperforms recent SOTA" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 291, + 553, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 291, + 553, + 339 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 553, + 339 + ], + "type": "text", + "content": "methods in both temporal event grounding (at least " + }, + { + "bbox": [ + 313, + 291, + 553, + 339 + ], + "type": "inline_equation", + "content": "5.3\\%" + }, + { + "bbox": [ + 313, + 291, + 553, + 339 + ], + "type": "text", + "content": " on average F1 score) and textual description (at least " + }, + { + "bbox": [ + 313, + 291, + 553, + 339 + ], + "type": "inline_equation", + "content": "3.6\\%" + }, + { + "bbox": [ + 313, + 291, + 553, + 339 + ], + "type": "text", + "content": " on average Sim score). This essentially demonstrates the generalizability of our method across various scenarios." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 340, + 555, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 555, + 435 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 555, + 435 + ], + "type": "text", + "content": "Charades-STA. From Tab. 2, it can be seen that our method still outperforms recent SOTA methods by large margins. Specifically, our 3.8B model outperforms counterpart [23] with the same base model by " + }, + { + "bbox": [ + 313, + 340, + 555, + 435 + ], + "type": "inline_equation", + "content": "4.4\\%" + }, + { + "bbox": [ + 313, + 340, + 555, + 435 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 313, + 340, + 555, + 435 + ], + "type": "inline_equation", + "content": "\\mathrm{R}@\\mathrm{1(IoU = 0.5)}" + }, + { + "bbox": [ + 313, + 340, + 555, + 435 + ], + "type": "text", + "content": ", and outperforms methods published in 2025 with 7B model size [10, 11, 31] by at least " + }, + { + "bbox": [ + 313, + 340, + 555, + 435 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 313, + 340, + 555, + 435 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 313, + 340, + 555, + 435 + ], + "type": "inline_equation", + "content": "\\mathrm{R}@\\mathrm{1(IoU = 0.5)}" + }, + { + "bbox": [ + 313, + 340, + 555, + 435 + ], + "type": "text", + "content": ". This further demonstrates the superiority of our method in both effectiveness and efficiency." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 436, + 554, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 554, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 554, + 532 + ], + "type": "text", + "content": "YouCook2. As shown in Tab. 3, our method achieves state-of-the-art performance compared to recent SOTA approaches. Specifically, it outperforms them by at least " + }, + { + "bbox": [ + 313, + 436, + 554, + 532 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 313, + 436, + 554, + 532 + ], + "type": "text", + "content": " in temporal event grounding (F1) while also demonstrating superior textual description capacity, with gains of at least " + }, + { + "bbox": [ + 313, + 436, + 554, + 532 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 313, + 436, + 554, + 532 + ], + "type": "text", + "content": " on CIDEr and " + }, + { + "bbox": [ + 313, + 436, + 554, + 532 + ], + "type": "inline_equation", + "content": "1.0\\%" + }, + { + "bbox": [ + 313, + 436, + 554, + 532 + ], + "type": "text", + "content": " on SODA.c. This further highlights the superiority of our approach in both temporal grounding and text generation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 540, + 416, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 416, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 416, + 552 + ], + "type": "text", + "content": "6.3. Ablation Studies" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 557, + 555, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 677 + ], + "type": "text", + "content": "Here we conduct ablation studies on E.T. Bench grounding and E.T. Bench Dense Captioning datasets to analyze the effect of the proposed components, in both temporal event-level grounding and textual description. For each dataset (i.e., grounding, dense captioning), we report the average performance across multiple sub-tasks. In the following, we first analyze the component effect of the newly designed generation objective and evidence token, where we do not take the proposed FPO into account. Finally, we analyze the effect of the proposed FPO." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "Generation objective of " + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{D}^2\\mathbf{VLM}" + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": ". Here, we analyze the effectiveness of our newly designed generation objective within " + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{D}^2\\mathbf{VLM}" + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": ". From Tab. 4, it can be seen that: (1) Row-1" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 752, + 318, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 752, + 318, + 761 + ], + "spans": [ + { + "bbox": [ + 293, + 752, + 318, + 761 + ], + "type": "text", + "content": "20689" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 364, + 140 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 364, + 140 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 364, + 140 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 364, + 140 + ], + "type": "table", + "html": "
Decomposed ObjectiveInterleaved Text-avi GenerationConsistency ConstraintGrounding AvgF1Dense Captioning AvgF1AvgSim
21.214.311.3
✓28.923.116.0
✓✓35.634.319.8
✓✓✓39.535.021.2
", + "image_path": "24e0b92329d9d6b81d2631416d6bc21edc0174ba6231270e0027c4508b68a2de.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 378, + 70, + 527, + 140 + ], + "blocks": [ + { + "bbox": [ + 94, + 144, + 328, + 156 + ], + "lines": [ + { + "bbox": [ + 94, + 144, + 328, + 156 + ], + "spans": [ + { + "bbox": [ + 94, + 144, + 328, + 156 + ], + "type": "text", + "content": "Table 4. Component effect of the designed generation objective." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 378, + 70, + 527, + 140 + ], + "lines": [ + { + "bbox": [ + 378, + 70, + 527, + 140 + ], + "spans": [ + { + "bbox": [ + 378, + 70, + 527, + 140 + ], + "type": "table", + "html": "
FPOGrounding AvgF1Dense Captioning AvgF1AvgSim
×39.535.021.2
✓42.337.521.8
", + "image_path": "a758ac3a911bf9ed71da82babdc7a18bbf1007558bcb1ac06bfd344804a245af.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 69, + 174, + 281, + 240 + ], + "blocks": [ + { + "bbox": [ + 385, + 144, + 520, + 156 + ], + "lines": [ + { + "bbox": [ + 385, + 144, + 520, + 156 + ], + "spans": [ + { + "bbox": [ + 385, + 144, + 520, + 156 + ], + "type": "text", + "content": "Table 5. Effect of the proposed FPO." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 174, + 281, + 240 + ], + "lines": [ + { + "bbox": [ + 69, + 174, + 281, + 240 + ], + "spans": [ + { + "bbox": [ + 69, + 174, + 281, + 240 + ], + "type": "table", + "html": "
MethodGrounding \nAvgF1Dense \nAvgF1Captioning \nAvgSim
w/o event-level modeling26.133.416.2
w/o visual semantic capture37.127.517.7
Full design39.535.021.2
", + "image_path": "65bb30731d8ae3e9dda1f838f9709e1c7ae9cdf7748fc95beb06b38e245e5848.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 70, + 242, + 279, + 254 + ], + "lines": [ + { + "bbox": [ + 70, + 242, + 279, + 254 + ], + "spans": [ + { + "bbox": [ + 70, + 242, + 279, + 254 + ], + "type": "text", + "content": "Table 6. Component effect within evidence token design." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 280, + 295, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 280, + 295, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 280, + 295, + 519 + ], + "type": "text", + "content": "v.s. Row-2: Decompose the coupled grounding-textual response objective into grounding (by our designed token) then answer from grounded event can substantially improve the performance in both grounding (i.e., " + }, + { + "bbox": [ + 55, + 280, + 295, + 519 + ], + "type": "inline_equation", + "content": "7.7\\%" + }, + { + "bbox": [ + 55, + 280, + 295, + 519 + ], + "type": "text", + "content": " averaged F1 on grounding tasks and " + }, + { + "bbox": [ + 55, + 280, + 295, + 519 + ], + "type": "inline_equation", + "content": "8.8\\%" + }, + { + "bbox": [ + 55, + 280, + 295, + 519 + ], + "type": "text", + "content": " averaged F1 on dense captioning tasks) and textual description (i.e., " + }, + { + "bbox": [ + 55, + 280, + 295, + 519 + ], + "type": "inline_equation", + "content": "4.7\\%" + }, + { + "bbox": [ + 55, + 280, + 295, + 519 + ], + "type": "text", + "content": " averaged sentence similarity on dense captioning tasks). (2) Row-2 v.s. Row-3: When we shift the answering stage from pure textual token generation into interleaved text-evi token generation, the performance can be further enhanced by large margins (at least " + }, + { + "bbox": [ + 55, + 280, + 295, + 519 + ], + "type": "inline_equation", + "content": "6.7\\%" + }, + { + "bbox": [ + 55, + 280, + 295, + 519 + ], + "type": "text", + "content": " on grounding and " + }, + { + "bbox": [ + 55, + 280, + 295, + 519 + ], + "type": "inline_equation", + "content": "3.8\\%" + }, + { + "bbox": [ + 55, + 280, + 295, + 519 + ], + "type": "text", + "content": " on textual description). The reason is that under such cases, the overall generation objective can be interpreted as grounding then answer with evidence referencing, where such referencing strengthens the relationship between event grounding and textual answering. (3) Row-3 v.s. Row-4: The introduced consistency constraint further boosts the performance by a noticeable margin, as it explicitly enhances the consistency between the two stages. Overall, the results validate the effectiveness of our design." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "content": "Design of evidence token. Here we analyze the impact of the core components within the token design (i.e., the event-level modeling and the explicit visual semantic capture). The results are shown in Tab. 6, where w/o event-level modeling means each token only captures the information of individual frame (a common design in existing works [14, 23, 29]), rather than an interval-level event (e.g., 10s-15s). w/o visual semantic capture indicates visual semantics from the video tokens is not explicitly integrated into the token through the operation illustrated in Sec. 3.2 (also a common design in most of the existing works [4, 14, 29]). From Tab. 6, it can be summarized that: (1) The performance drops significantly without event-level modeling. We argue that this is because compared to individual timestamp-level modeling, event-level modeling more naturally aligns with the characteristics of" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 177, + 553, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 177, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 313, + 177, + 553, + 295 + ], + "type": "text", + "content": "event-level evidence grounding and can capture more robust event-level semantics in a global perspective. (2) Explicit visual semantic capture is also crucial for enhancing model effectiveness, as it enables tokens to truly serve as solid context for subsequent answer generation, which also fits well with the autoregressive generation process. Notably, compared with grounding-only tasks, such an effect is more pronounced on dense captioning tasks, as textual description generation can truly benefit from the captured event-level visual semantics." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 297, + 555, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 555, + 369 + ], + "type": "text", + "content": "Effect of factorized preference optimization. Tab. 5 demonstrates that the proposed FPO algorithm consistently enhances performance on both temporal grounding and captioning-related tasks, with a more significant improvement in temporal grounding. This indeed validates the effectiveness of our factorized optimization objective design." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 381, + 471, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 381, + 471, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 471, + 393 + ], + "type": "text", + "content": "7. Conclusion and Limitations" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 402, + 555, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 555, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 555, + 593 + ], + "type": "text", + "content": "In this work, we uncover some key limitations of existing temporal grounded video-language models: (1) the use of coupled learning objective for temporal grounding and textual response generation, and (2) the overlooking of inherent dependency between these two tasks, especially the event-level visual semantics from the intermediate grounding result. We tackle this from a factorized learning perspective, where we propose " + }, + { + "bbox": [ + 313, + 402, + 555, + 593 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^2\\mathrm{VLM}" + }, + { + "bbox": [ + 313, + 402, + 555, + 593 + ], + "type": "text", + "content": " that decomposes the generation objective into a \"grounding then answering with evidence referencing\" paradigm, and introduce evidence tokens to emphasize explicit visual semantic capture beyond timestamp representation. Furthermore, we design a factorized preference learning algorithm, coupled with a factorized synthetic dataset, to enhance the learning of both tasks. Extensive experiments demonstrate the clear superiority of our approach across various scenarios." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "content": "Despite achieving new state-of-the-art results, our approach still has room for improvement (e.g., F1 scores of only " + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "inline_equation", + "content": "14.4\\%" + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "content": " on the episodic memory task and " + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "inline_equation", + "content": "26.4\\%" + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "content": " on the YouCook2 dense video captioning task). Besides, while our factorized data generation offers good factorized controllability and directly contributes to factorized optimization, it focuses only on generating negative (dis-preferred) samples. However, generating more diverse positive samples to further enrich the preference data is also meaningful, and we leave this for future work." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 752, + 318, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 752, + 318, + 761 + ], + "spans": [ + { + "bbox": [ + 293, + 752, + 318, + 761 + ], + "type": "text", + "content": "20690" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 187, + 115, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 187, + 115, + 201 + ], + "spans": [ + { + "bbox": [ + 56, + 187, + 115, + 201 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 207, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 207, + 296, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 207, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 61, + 207, + 296, + 262 + ], + "type": "text", + "content": "[1] Marah Abdin, Sam Ade Jacobs, Ammar Ahmad Awan, Jyoti Aneja, Ahmed Awadallah, Hany Awadalla, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Harkirat Behl, et al. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219, 2024. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 263, + 296, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 263, + 296, + 306 + ], + "spans": [ + { + "bbox": [ + 62, + 263, + 296, + 306 + ], + "type": "text", + "content": "[2] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 308, + 296, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 308, + 296, + 362 + ], + "spans": [ + { + "bbox": [ + 62, + 308, + 296, + 362 + ], + "type": "text", + "content": "[3] Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, et al. Constitutional ai: Harmlessness from ai feedback. arXiv preprint arXiv:2212.08073, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 365, + 296, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 365, + 296, + 397 + ], + "spans": [ + { + "bbox": [ + 62, + 365, + 296, + 397 + ], + "type": "text", + "content": "[4] Qirui Chen, Shangzhe Di, and Weidi Xie. Grounded multi-hop videoqa in long-form egocentric videos. arXiv preprint arXiv:2408.14469, 2024. 2, 3, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 399, + 296, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 399, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 62, + 399, + 296, + 453 + ], + "type": "text", + "content": "[5] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 456, + 296, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 456, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 62, + 456, + 296, + 521 + ], + "type": "text", + "content": "[6] Wenliang Dai, Junnan Li, DONGXU LI, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale N Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. In Advances in Neural Information Processing Systems, pages 49250-49267. Curran Associates, Inc., 2023. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 523, + 296, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 523, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 62, + 523, + 296, + 567 + ], + "type": "text", + "content": "[7] Shangzhe Di and Weidi Xie. Grounded question-answering in long egocentric videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12934–12943, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 568, + 296, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 568, + 296, + 632 + ], + "spans": [ + { + "bbox": [ + 62, + 568, + 296, + 632 + ], + "type": "text", + "content": "[8] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19358-19369, 2023. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 635, + 296, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 635, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 62, + 635, + 296, + 678 + ], + "type": "text", + "content": "[9] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In Proceedings of the IEEE international conference on computer vision, pages 5267-5275, 2017. 1, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "type": "text", + "content": "[10] Yongxin Guo, Jingyu Liu, Mingda Li, Dingxin Cheng, Xiaoying Tang, Dianbo Sui, Qingbin Liu, Xi Chen, and Kevin Zhao. Vtg-llm: Integrating timestamp knowledge into video" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "type": "text", + "content": "This research is supported by the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No: AISG3-RP-2022-030). We would like to acknowledge that computational work involved in this research work is partially supported by NUS IT's Research Computing group under grant number NUSREC-HPC-00001." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "text", + "content": "llms for enhanced video temporal grounding. arXiv preprint arXiv:2405.13382, 2024. 1, 2, 3, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 139 + ], + "type": "text", + "content": "[11] Yongxin Guo, Jingyu Liu, Mingda Li, Xiaoying Tang, Qingbin Liu, and Xi Chen. Trace: Temporal grounding video llm via causal event modeling. In International Conference on Learning Representations, 2025. 2, 3, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 140, + 554, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 554, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 554, + 183 + ], + "type": "text", + "content": "[12] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 185, + 554, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 185, + 554, + 237 + ], + "spans": [ + { + "bbox": [ + 317, + 185, + 554, + 237 + ], + "type": "text", + "content": "[13] Bin Huang, Xin Wang, Hong Chen, Zihan Song, and Wenwu Zhu. Vtimellm: Empower llm to grasp video moments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14271-14280, 2024. 2, 3, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 239, + 554, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 554, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 554, + 293 + ], + "type": "text", + "content": "[14] De-An Huang, Shijia Liao, Subhashree Radhakrishnan, Hongxu Yin, Pavlo Molchanov, Zhiding Yu, and Jan Kautz. Lita: Language instructed temporal-localization assistant. In European Conference on Computer Vision, pages 202-218. Springer, 2024. 2, 3, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 294, + 554, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 294, + 554, + 348 + ], + "spans": [ + { + "bbox": [ + 317, + 294, + 554, + 348 + ], + "type": "text", + "content": "[15] Harrison Lee, Samrat Phatale, Hassan Mansoor, Thomas Mesnard, Johan Ferret, Kellie Lu, Colton Bishop, Ethan Hall, Victor Carbune, Abhinav Rastogi, et al. Rlaif vs. rlhf: Scaling reinforcement learning from human feedback with ai feedback. arXiv preprint arXiv:2309.00267, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 350, + 554, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 350, + 554, + 403 + ], + "spans": [ + { + "bbox": [ + 317, + 350, + 554, + 403 + ], + "type": "text", + "content": "[16] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 404, + 554, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 404, + 554, + 449 + ], + "spans": [ + { + "bbox": [ + 317, + 404, + 554, + 449 + ], + "type": "text", + "content": "[17] Lei Li, Zhihui Xie, Mukai Li, Shunian Chen, Peiyi Wang, Liang Chen, Yazheng Yang, Benyou Wang, and Lingpeng Kong. Silkie: Preference distillation for large visual language models. arXiv preprint arXiv:2312.10665, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 449, + 554, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 554, + 491 + ], + "type": "text", + "content": "[18] Rui Li, Xiaohan Wang, Yuhui Zhang, Zeyu Wang, and Serena Yeung-Levy. Temporal preference optimization for long-form video understanding. arXiv preprint arXiv:2501.13919, 2025. 3, 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 493, + 554, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 493, + 554, + 536 + ], + "spans": [ + { + "bbox": [ + 317, + 493, + 554, + 536 + ], + "type": "text", + "content": "[19] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In European Conference on Computer Vision, pages 323–340. Springer, 2024. 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 537, + 554, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 537, + 554, + 602 + ], + "spans": [ + { + "bbox": [ + 317, + 537, + 554, + 602 + ], + "type": "text", + "content": "[20] Zhaowei Li, Qi Xu, Dong Zhang, Hang Song, YiQing Cai, Qi Qi, Ran Zhou, Junting Pan, Zefeng Li, Vu Tu, et al. Groundinggpt: Language enhanced multi-modal grounding model. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6657-6678, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 604, + 555, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 604, + 555, + 658 + ], + "spans": [ + { + "bbox": [ + 317, + 604, + 555, + 658 + ], + "type": "text", + "content": "[21] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 5971–5984, 2024. 1, 2, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 658, + 554, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 554, + 691 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 554, + 691 + ], + "type": "text", + "content": "[22] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2023. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "type": "text", + "content": "[23] Ye Liu, Zongyang Ma, Zhongang Qi, Yang Wu, Ying Shan, and Chang Wen Chen. E.T. Bench: Towards open-ended" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "spans": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "type": "text", + "content": "20691" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "text", + "content": "event-level video-language understanding. In The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2024. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 171 + ], + "type": "text", + "content": "[24] Shayne Longpre, Le Hou, Tu Vu, Albert Webson, Hyung Won Chung, Yi Tay, Denny Zhou, Quoc V Le, Barret Zoph, Jason Wei, et al. The flan collection: Designing data and methods for effective instruction tuning. In International Conference on Machine Learning, pages 22631-22648. PMLR, 2023. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 295, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 295, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 295, + 237 + ], + "type": "text", + "content": "[25] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12585–12602, 2024. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 238, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 238, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 238, + 294, + 293 + ], + "type": "text", + "content": "[26] Munan Ning, Bin Zhu, Yujia Xie, Bin Lin, Jiaxi Cui, Lu Yuan, Dongdong Chen, and Li Yuan. Video-bench: A comprehensive benchmark and toolkit for evaluating video-based large language models. arXiv preprint arXiv:2311.16103, 2023. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 293, + 295, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 295, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 295, + 358 + ], + "type": "text", + "content": "[27] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 360, + 295, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 295, + 425 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 295, + 425 + ], + "type": "text", + "content": "[28] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 426, + 295, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 295, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 295, + 480 + ], + "type": "text", + "content": "[29] Long Qian, Juncheng Li, Yu Wu, Yaobo Ye, Hao Fei, TatSeng Chua, Yueting Zhuang, and Siliang Tang. Momentor: Advancing video large language model with fine-grained temporal reasoning. arXiv preprint arXiv:2402.11435, 2024. 3, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 481, + 295, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 481, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 481, + 295, + 536 + ], + "type": "text", + "content": "[30] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36:53728-53741, 2023. 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 536, + 295, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 536, + 295, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 536, + 295, + 591 + ], + "type": "text", + "content": "[31] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14313-14323, 2024. 1, 2, 3, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 592, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 295, + 624 + ], + "type": "text", + "content": "[32] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 625, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 295, + 679 + ], + "type": "text", + "content": "[33] Nisan Stiannon, Long Ouyang, Jeffrey Wu, Daniel Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford, Dario Amodei, and Paul F Christiano. Learning to summarize with human feedback. Advances in neural information processing systems, 33:3008-3021, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 680, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 295, + 713 + ], + "type": "text", + "content": "[34] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 6" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 137 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 137 + ], + "type": "text", + "content": "[35] Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liang-Yan Gui, Yu-Xiong Wang, Yiming Yang, et al. Aligning large multimodal models with factually augmented rlhf. In Annual Meeting of the Association for Computational Linguistics, 2024. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 139, + 554, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 139, + 554, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 139, + 554, + 194 + ], + "type": "text", + "content": "[36] Haibo Wang, Zhiyang Xu, Yu Cheng, Shizhe Diao, Yufan Zhou, Yixin Cao, Qifan Wang, Weifeng Ge, and Lifu Huang. Grounded-videollm: Sharpening fine-grained temporal grounding in video large language models. arXiv preprint arXiv:2410.03290, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 195, + 554, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 554, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 554, + 237 + ], + "type": "text", + "content": "[37] Yueqian Wang, Xiaojun Meng, Jianxin Liang, Yuxuan Wang, Qun Liu, and Dongyan Zhao. Hawkeye: Training video-text llms for grounding text in videos. arXiv preprint arXiv:2403.10228, 2024. 3, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 239, + 554, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 554, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 554, + 282 + ], + "type": "text", + "content": "[38] Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652, 2021. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 284, + 554, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 554, + 337 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 554, + 337 + ], + "type": "text", + "content": "[39] Junbin Xiao, Angela Yao, Yicong Li, and Tat-Seng Chua. Can i trust your answer? visually grounded video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13204-13214, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 338, + 554, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 554, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 554, + 381 + ], + "type": "text", + "content": "[40] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 383, + 554, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 554, + 547 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 554, + 547 + ], + "type": "text", + "content": "[41] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng He, Junyang Lin, Kai Dang, Keming Lu, Keqin Chen, Kexin Yang, Mei Li, Mingfeng Xue, Na Ni, Pei Zhang, Peng Wang, Ru Peng, Rui Men, Ruize Gao, Runji Lin, Shijie Wang, Shuai Bai, Sinan Tan, Tianhang Zhu, Tianhao Li, Tianyu Liu, Wenbin Ge, Xiaodong Deng, Xiaohuan Zhou, Xingzhang Ren, Xinyu Zhang, Xipin Wei, Xuancheng Ren, Yang Fan, Yang Yao, Yichang Zhang, Yu Wan, Yunfei Chu, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, and Zhihao Fan. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 548, + 554, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 554, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 554, + 658 + ], + "type": "text", + "content": "[42] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, Le Yu, Mei Li, Mingfeng Xue, Pei Zhang, Qin Zhu, Rui Men, Runji Lin, Tianhao Li, Tingyu Xia, Xingzhang Ren, Xuancheng Ren, Yang Fan, Yang Su, Yichang Zhang, Yu Wan, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, and Zihan Qiu. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115, 2024. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 658, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 554, + 713 + ], + "type": "text", + "content": "[43] Xiangyu Zeng, Kunchang Li, Chenting Wang, Xinhao Li, Tianxiang Jiang, Ziang Yan, Songze Li, Yansong Shi, Zhengrong Yue, Yi Wang, et al. Timesuite: Improving mllms for long video understanding via grounded tuning. In International Conference on Learning Representations, 2025. 3, 7" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "spans": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "type": "text", + "content": "20692" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 295, + 316 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[44] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 107, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 107, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 55, + 107, + 295, + 162 + ], + "type": "text", + "content": "[45] Mengxi Zhang, Wenhao Wu, Yu Lu, Yuxin Song, Kang Rong, Huanjin Yao, Jianbo Zhao, Fanglong Liu, Haocheng Feng, Jingdong Wang, et al. Automated multi-level preference for mllms. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 163, + 295, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 163, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 57, + 163, + 295, + 217 + ], + "type": "text", + "content": "[46] Ruohong Zhang, Liangke Gui, Zhiqing Sun, Yihao Feng, Keyang Xu, Yuanhan Zhang, Di Fu, Chunyuan Li, Alexander Hauptmann, Yonatan Bisk, et al. Direct preference optimization of video large multimodal models from language model reward. arXiv preprint arXiv:2404.01258, 2024. 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 219, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 219, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 57, + 219, + 294, + 262 + ], + "type": "text", + "content": "[47] Luowei Zhou, Chenliang Xu, and Jason Corso. Towards automatic learning of procedures from web instructional videos. In Proceedings of the AAAI Conference on Artificial Intelligence, 2018. 1, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 263, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 263, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 57, + 263, + 294, + 316 + ], + "type": "text", + "content": "[48] Daniel M Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. Fine-tuning language models from human preferences, 2020. URL https://arxiv.org/abs, page 14, 1909. 3" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 753, + 317, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 317, + 761 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 317, + 761 + ], + "type": "text", + "content": "20693" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Failure Cases Are Better Learned But Boundary Says Sorry_ Facilitating Smooth Perception Change for Accuracy-Robustness Trade-Off in Adversarial Training/9886c220-8af0-438e-8e12-4de3189d4e86_content_list.json b/2025/Failure Cases Are Better Learned But Boundary Says Sorry_ Facilitating Smooth Perception Change for Accuracy-Robustness Trade-Off in Adversarial Training/9886c220-8af0-438e-8e12-4de3189d4e86_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..dd4bec1e7d2a8a16d555675f8370a60721b53b7b --- /dev/null +++ b/2025/Failure Cases Are Better Learned But Boundary Says Sorry_ Facilitating Smooth Perception Change for Accuracy-Robustness Trade-Off in Adversarial Training/9886c220-8af0-438e-8e12-4de3189d4e86_content_list.json @@ -0,0 +1,1379 @@ +[ + { + "type": "text", + "text": "Failure Cases Are Better Learned But Boundary Says Sorry: Facilitating Smooth Perception Change for Accuracy-Robustness Trade-Off in Adversarial Training", + "text_level": 1, + "bbox": [ + 91, + 130, + 906, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yanyun Wang Li Liu*", + "bbox": [ + 398, + 203, + 599, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The Hong Kong University of Science and Technology (Guangzhou)", + "bbox": [ + 223, + 226, + 767, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 277, + 326, + 295 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Adversarial Training (AT) is one of the most effective methods to train robust Deep Neural Networks (DNNs). However, AT creates an inherent trade-off between clean accuracy and adversarial robustness, which is commonly attributed to the more complicated decision boundary caused by the insufficient learning of hard adversarial samples. In this work, we reveal a counterintuitive fact for the first time: From the perspective of perception consistency, hard adversarial samples that can still attack the robust model after AT are already learned better than those successfully defended. Thus, different from previous views, we argue that it is rather the over-sufficient learning of hard adversarial samples that degrades the decision boundary and contributes to the trade-off problem. Specifically, the excessive pursuit of perception consistency would force the model to view the perturbations as noise and ignore the information within them, which should have been utilized to induce a smoother perception transition towards the decision boundary to support its establishment to an appropriate location. In response, we define a new AT objective named Robust Perception, encouraging the model perception to change smoothly with input perturbations, based on which we propose a novel Robust Perception Adversarial Training (RPAT) method, effectively mitigating the current accuracy-robustness trade-off. Experiments on CIFAR-10, CIFAR-100, and Tiny-ImageNet with ResNet-18, PreActResNet-18, and WideResNet-34-10 demonstrate the effectiveness of our method beyond four common baselines and 12 state-of-the-art (SOTA) works. The code is available at https://github.com/FlaAI/RPAT.", + "bbox": [ + 88, + 310, + 485, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 791, + 220, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep neural networks (DNNs) have demonstrated impressive performance across various real-world applications [24, 26, 60]. However, they are inherently vulnerable to adversarial attacks [4, 39]. Specifically, even impercep", + "bbox": [ + 89, + 816, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3af3a29bf8b909c59ece5b66e79e02c7ee88bde2635c70c29200b013605ee149.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 276, + 709, + 398 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/be6a139caa4fa6f60e810e855d5df80c8bbfbb95dd96fc195ab83bf5118e09d4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 723, + 276, + 903, + 398 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5a8f87284bdeaa37efaf2340baceebfc4f83998107bbda98f0f3e654e80ec287.jpg", + "image_caption": [ + "Figure 1. Success and failure cases of AT defense from the perspectives of prediction and perception, respectively. (a) and (b): AT works through building a robust ball with radius $d$ such that the model prediction remains unchanged within it. Then if the robust $d$ -ball fails to contain the $\\epsilon$ -ball, there would still be failure cases beating the robust model under perturbation $\\delta$ within the $\\epsilon$ -ball. (c) and (d): In this work, we reveal that the smaller $d$ -ball in the failure cases is not a result of more vulnerable model perception as previous intuition. In fact, as demonstrated by the contour line, which shows the perturbation ball to the same degree of perception change, the failure cases already achieve less perception change under similar $\\delta$ within the $\\epsilon$ -ball. This implies that the issue of these failure cases is not failing to learn perception consistency following the current AT objective, but failing to support the establishment of decision boundary to an appropriate location." + ], + "image_footnote": [], + "bbox": [ + 519, + 400, + 709, + 515 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9c677d01bc3c8323c7274ecfa24a41d02b945ed5d31859ccbbddafa4d943183e.jpg", + "image_caption": [], + "image_footnote": [ + "(c) (d) Contour line of model perception" + ], + "bbox": [ + 712, + 401, + 903, + 526 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tible perturbations to inputs can cause substantial changes in the outputs [12, 17]. As this problem raises significant security concerns [45], previously, several defense strategies, such as Defense Distillation [32], Feature Squeezing [52], Randomization [51], and Input Denoising [14, 25], have been proposed to improve the adversarial robustness of DNNs. However, most of these approaches have proven", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author: avrillliu@hkust-gz.edu.cn", + "bbox": [ + 107, + 886, + 465, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4691", + "bbox": [ + 482, + 944, + 513, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ineffective against more advanced adaptive attacks [2, 42], such as Auto-Attack [7]. Currently, Adversarial Training (AT) [12, 29] is regarded as one of the most effective methods for training robust DNNs [2, 9].", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, an accuracy-robustness trade-off problem have been widely identified for AT in recent studies [3, 28, 34-36, 43, 44, 48, 54, 58]. Compared with standard training, the improvement in adversarial robustness acquired by AT is usually at the expense of the reduction in clean accuracy, which is afraid to significantly degrade the user experience in clean context and hinder the application of AT in real-world practices. The current consensus is that AT aggravates the accuracy-robustness trade-off by creating overly complex decision boundaries, primarily because models struggle to sufficiently learn from hard adversarial samples during training [6, 8, 10, 19, 27, 28, 34, 48, 53, 54]. Specifically, as the conventional AT objective encourages the model to perceive any adversarial inputs within the $\\epsilon$ -ball similar to the corresponding benign one to produce consistent predictions for them, previous works empirically assume that the wrong predictions of the hard adversarial samples are attributed to the insufficiently optimized perception of them under the current objective. Therefore, although defining \"hard sample\" from different perspectives (e.g., classification result and distance to decision boundary), previous works agree to further optimize the perception consistency on them, ending up with various learning strategies specific for them, such as data reweighting [27, 59], adaptive perturbation radius [8, 34], and non-one-hot supervised signal [10, 19].", + "bbox": [ + 91, + 154, + 483, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, in this work, we reveal that for the robust model trained by convention AT, the statistical difference in model perception from benign to hard adversarial samples causing failure defense is already smaller than to others (i.e., those successfully defended). It implies that the model has even more sufficiently learned them under the current AT objective, and should have successfully defended them if the above assumption holds. Thus, as demonstrated in Figure 1, we suggest that the key factor here is no longer the optimization of model perception but the location of the decision boundary. In that case, keeping on pursuing over-sufficient small perception change would not only help little in further reducing the prediction error, but also force the model to view the perturbations as noise and ignore the information within them, which impairs their ability in claiming robust balls with reasonable gradients to push the decision boundary to an appropriate location. As a consequence, such sample-level imbalance in boundary establishment would naturally increase its complexity, harming generalization ability and degrading the trade-off between accuracy and robustness.", + "bbox": [ + 89, + 550, + 482, + 866 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To mitigate the over-sufficient learning of model perception on the failure cases and retain effective gradient sig", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "nals w.r.t. them to facilitate the establishment of the decision boundary to an appropriate location, firstly, we define a new AT objective named Robust Perception in addition to the conventional one, encouraging smoother perception changes towards the decision boundary along with the input perturbations. Rigorous theoretical derivations also demonstrate its effectiveness in smoothening the decision boundary and releasing the current trade-off problem from the perspectives of local linearity [12] and Lipschitz regularization [1, 11, 21, 33]. Secondly, based on our new objective, we propose a novel Robust Perception Adversarial Training (RPAT) method, concurrently improving clean accuracy and adversarial robustness in AT. Extensive experiments on CIFAR-10, CIFAR-100, and Tiny-ImageNet with ResNet-18, PreActResNet-18, and WideResNet-34-10 demonstrate the advancement of RPAT beyond four common benchmarks and 12 state-of-the-art (SOTA) works.", + "bbox": [ + 511, + 90, + 903, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Background and Related Works", + "text_level": 1, + "bbox": [ + 511, + 359, + 805, + 376 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Robustness and Adversarial Training", + "text_level": 1, + "bbox": [ + 511, + 383, + 834, + 400 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Adversarial robustness reflects model vulnerability to malicious perturbations, which is commonly quantified by test accuracy under adversarial attacks [3, 39, 46]. Representatively, the Fast Gradient Sign Method (FGSM) [12] perturbs input samples along the gradient of the loss function, whereas the Projected Gradient Descent (PGD) attack [29] applies FGSM iteratively and projects the perturbation onto the $\\epsilon$ -ball. Auto-Attack [7] introduces and combines several parameter-free attacks to mitigate practical issues such as hyper-parameter tuning and gradient masking. Due to their effectiveness, PGD and Auto-Attack are widely used to assess adversarial robustness, and so does this work.", + "bbox": [ + 511, + 406, + 903, + 587 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previously, although many defense strategies including Defense Distillation [32], Feature Squeezing [52], Input Denoising [14, 25] and Randomization [51] have been developed, most rely on obfuscated gradients [2] and can be evaded by advanced adaptive attacks [42]. Nowadays, Adversarial Training [12, 29] is recognized as the most effective method to enhance adversarial robustness [2, 9, 41]. Unlike clean training on benign samples, AT directly learns on adversarily augmented data. Its objective can be formulated as the following min-max problem [29]:", + "bbox": [ + 511, + 588, + 903, + 739 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\boldsymbol {\\theta}} \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\max _ {\\| \\mathbf {x} _ {i} ^ {\\prime} - \\mathbf {x} _ {i} \\| _ {P} \\leq \\epsilon} \\mathcal {L} \\left(f _ {\\boldsymbol {\\theta}} \\left(\\mathbf {x} _ {i} ^ {\\prime}\\right), y _ {i}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 747, + 903, + 785 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $f_{\\pmb{\\theta}}(\\mathbf{x}_i) = \\arg \\max_{c=1,\\dots,C} \\mathbf{p}_c(\\mathbf{x}_i, \\pmb{\\theta})$ with $\\mathbf{p}_c$ denoting the softmax probability of class $c$ within the prediction of $\\mathbf{x}_i$ , $n$ denotes the number of training samples, and $\\mathbf{x}_i'$ is generated by adding the strongest $L_p$ -norm perturbation within the $\\epsilon$ -ball to the original sample $\\mathbf{x}_i$ . This objective aims at prediction consistency within the $\\epsilon$ -ball, encouraging the model to also produce similar perception there.", + "bbox": [ + 511, + 794, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4692", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "PGD-AT [29] is the first attempt at solving this min-max optimization iteratively by approximating the inner maximization via PGD to craft adversarial samples and then performing stochastic gradient descent (SGD) on these samples for the outer minimization. TRADES [58] first considers balancing clean accuracy and adversarial robustness in AT, which decomposes robust error into natural error and boundary error, with the latter occurring only if specific data points are sufficiently close to the decision boundary. MART [45] further refines AT by distinguishing misclassified from correctly classified samples, which suggests appending boundary error only for misclassified samples to the basic AT objective in PGD-AT. Consistency-AT [40] encourages the prediction between adversarial samples derived from different augmentations of the same benign sample to be similar, which explicitly enhances the prediction consistency further.", + "bbox": [ + 89, + 90, + 480, + 347 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Accuracy-Robustness Trade-Off in AT", + "text_level": 1, + "bbox": [ + 89, + 362, + 421, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An important problem identified for AT is the trade-off between clean accuracy and adversarial robustness [3, 28, 34-36, 43, 44, 54, 58], which means the robustness gained by AT often comes at the cost of reduced model accuracy compared with clean training. Consequently, AT can degrade user experience under benign conditions, which may seriously delay its real-world deployment. A widely acknowledged cause of this problem is that AT tends to learn more complicated decision boundaries than clean training [6, 10, 28, 34, 53, 54], thereby harming the ability of the robust model to generalize to unseen data. A number of previous works attribute this to the insufficient learning of hard adversarial samples [8, 10, 19, 27, 34, 59]. Though defining \"hard adversarial samples\" from different perspectives, such as the classification result and the distance to decision boundary, they basically reached a consensus that the model perception w.r.t. hard adversarial samples is not well-optimized, and as a consequence, fails to produce the consistent final prediction. In response, these works propose various solutions to help the model learn better prediction consistency for those hard adversarial samples.", + "bbox": [ + 89, + 385, + 482, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For instance, GAIRAT [59] proposes that hard samples nearer to the decision boundary are more critical but less robust and should be assigned a larger weight. MAIL [27] further proposes three types of probabilistic margins to quantify the closeness for better reweighting adversarial data accordingly. In contrast, MMA [8] suggests learning hard adversarial samples with reduced perturbations, which employs an adaptive $\\epsilon$ for perturbations to maximize the margin between data and the decision boundary. EWAT [20] weighs the loss for each adversarial sample proportionally to the entropy of its prediction distribution during AT to focus on those with more uncertain labels. TE [10] reveals that the model may attempt to memorize certain hard ad", + "bbox": [ + 89, + 704, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "versarial samples during AT, as it is difficult to assign high-confidence one-hot labels for them [6, 38]. SOVR [19] suggests increasing the margins of logits for certain hard samples by switching to a one-vs-the-rest loss.", + "bbox": [ + 511, + 90, + 903, + 150 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Additionally, several works also study the trade-off problem from other perspectives. AWP [49] proposes a double perturbation mechanism that flattens the loss landscape by weight perturbation to improve robust generalization. KD+SWA [5] injects more smoothening during AT by leveraging knowledge distillation to smooth the logits and performing stochastic weight averaging [18] to smooth the weights. ADR [50] employs soft labels as a guidance mechanism, accurately reflecting the distribution shift under attack during AT. CURE [13] selectively updates specific layers while preserving others to enhance the learning capacity of the model. ReBAT [47] views AT as a dynamic min-max game between the model trainer and the attacker, rebalancing the two players by either constraining the trainer's capacity or intensifying the attack.", + "bbox": [ + 511, + 152, + 903, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Motivation: AT Over-Learns Failure Cases", + "text_level": 1, + "bbox": [ + 511, + 391, + 897, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "It is basically common sense for DNN-based classification that the model learns to perceive input better and is thus able to predict it more precisely (especially when the model perception is indicated by its logits). Therefore, it seems reasonable at first glance that the previous works assume that the wrong prediction of the hard adversarial samples is due to the insufficient model perception of them, and propose various solutions upon this assumption in response.", + "bbox": [ + 511, + 417, + 903, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, to the best of our knowledge, there is no previous discussion on a simple but crucial question: What if the assumption does not hold? If the model perception on the hard adversarial samples is already relatively close to the corresponding benign ones (e.g., at a similar or even better level than non-hard ones), then it would be no longer effective enough to further optimize the model on such samples through an objective simply pursuing prediction consistency as the conventional AT. Also, in turn, this can no longer effectively benefit the further reduction of the prediction error. Surprisingly, our following proof-of-concept experiments reveal that this is exactly the case.", + "bbox": [ + 511, + 537, + 903, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The proof-of-concept experiments are conducted on CIFAR-10 dataset with ResNet-18 architecture, with the results illustrated in Figure 2. As suggested by the x-axis, we involve three cases with different pairs of training and evaluation methods, which are respectively 1) trained on clean data and evaluated with random perturbations, 2) trained and evaluated with randomly perturbed data, and 3) trained with PGD-AT and evaluated with PGD-20 adversary. The budget of the random perturbations $\\epsilon = 8 / 255$ is assigned the same as PGD, following our experimental setup detailed in Appendix A, so do other training and evaluation settings. Given that the conventional AT objective in Equation (1)", + "bbox": [ + 511, + 719, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4693", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "encourages the model to produce a similar perception for benign and adversarial samples, we utilize the similarity in model perception between the corresponding benign and perturbed samples to estimate the learning sufficiency of the perturbed samples in AT. Specifically, the model perception is represented by the logits, and the similarity is measured through the mean squared error (MSE) score. We calculate the score respectively for success and failure cases of defense, which are divided according to the correctness of the final classification results of the perturbed samples. In addition, the model accuracy on clean and perturbed data is also recorded and illustrated together.", + "bbox": [ + 89, + 90, + 480, + 272 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4c76dec9f1f0d9dc7d8a43f98f770561211e18dd756e5efa172bbaae57571c0b.jpg", + "image_caption": [ + "Figure 2. Proof-of-concept experiments for the different learning sufficiency in model perception for success and failure cases on CIFAR-10 with ResNet-18. In the x-axis, the \"Random\" for \"Training Algorithm\" means the training data is randomly perturbed within the budget $\\epsilon$ , while that for \"Perturbation Method\" indicates the attack used for evaluation on test data is random perturbation. The bars demonstrate the MSE score on logits between each pair of benign and perturbed samples, with the lines additionally showing the model accuracy on clean and perturbed data. Our results reveal a surprising phenomenon that, different from the clean training (with or without random noise), where failure cases are worse learned as we intuitively expected, AT already learns a better model perception for failure cases." + ], + "image_footnote": [], + "bbox": [ + 109, + 284, + 460, + 481 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Surprisingly, we experimentally found that the failure cases already learn better in model perception than the success cases. First, for the clean trained model, the perception on failure cases is significantly worse than on the success ones, just as we expected. Although the difference between success and failure cases has decreased for the randomly trained model, the failure cases remain worse, which suggests that non-adversarial perturbations are influential but not enough to change the trend. However, unlike the previous intuition, the model trained by PGD-AT can already perceive failure cases better than success ones. Our finding updates the current knowledge that, unlike those in clean training, the failure cases in AT do not fail simply because they are not sufficiently learned. On the contrary,", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the model even learns them harder to try to achieve the prediction and perception consistency. The smaller perception change under similar perturbations is direct evidence for this point of view.", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, as a straightforward deduction, now that we have success cases with worse perception but failure cases with better perception, there should be a statistical difference between them in the maximum perception change they can tolerate while still maintaining the consistent prediction, just as demonstrated in Figure 1. In other words, the decision boundary is established relatively closer to the failure cases, which is instead the issue we should focus on, because such sample-level imbalance in boundary establishment can naturally complicate it and thus degrade the accuracy-robustness trade-off as mentioned in Section 2.2. In that case, the over-consistent model perception between benign and adversarial samples in the failure cases rather harms the final robustness, because this is expected to force the model to view the perturbations as noise and ignore the information within them, which should have been utilized to induce a smoother perception transition towards the decision boundary to support its establishment, thus impairing the ability of such cases to claim a reasonable robust distance during AT compared with success cases. Therefore, in this work, we argue that it is not insufficient but oversufficient learning of failure cases that contributes to the more complicated decision boundary and finally results in the current trade-off problem.", + "bbox": [ + 511, + 152, + 906, + 515 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Proposed Method", + "text_level": 1, + "bbox": [ + 511, + 532, + 687, + 550 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Robust Perception: A New AT Objective", + "text_level": 1, + "bbox": [ + 511, + 559, + 861, + 574 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Section 3, the too-small perception change under perturbations caused by the over-sufficient learning of failure cases has been identified as an important factor contributing to the current trade-off problem, which in turn questions the effectiveness of the conventional AT objective on such failure cases.", + "bbox": [ + 511, + 582, + 903, + 671 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Formally, under Equation (1) which implicitly optimize the perception distance $\\| h_{\\theta}(\\mathbf{x}^{\\prime}) - h_{\\theta}(\\mathbf{x})\\| \\to 0$ between corresponding benign sample $\\mathbf{x}$ and adversarial sample $\\mathbf{x}^{\\prime}$ , the gradient of the conventional AT loss w.r.t. the perturbation, $\\nabla_{\\| \\mathbf{x}^{\\prime} - \\mathbf{x}\\|}\\mathcal{L}$ , would also approach 0. In that case, the model cannot further optimize the decision boundary using the perturbation information. Thus, a straightforward idea to mitigate this is to encourage smoother perception change towards the decision boundary along with the input perturbation, such that $\\| h_{\\theta}(\\mathbf{x}^{\\prime}) - h_{\\theta}(\\mathbf{x})\\| \\propto \\| \\mathbf{x}^{\\prime} - \\mathbf{x}\\|$ , as long as the final prediction remains unchanged. This is expected to retain an effective gradient signal in the direction of the perturbation, which enables the model to utilize this information to guide the decision boundary to a more reasonable location. Accordingly, in this section, we propose Robust", + "bbox": [ + 511, + 674, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4694", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e7a9880cec0e57341631f7022e1a3db81a4ef2c5b4bd8502106ff36fd843a4cf.jpg", + "image_caption": [ + "Figure 3. Illustration of the proposed RPAT method. In addition to the conventional AT objective, which learns adversarial samples with the original supervised labels, RPAT introduces a novel regularization term to achieve the newly defined Robust Perception objective. Given benign sample $\\mathbf{x}$ and adversarial samples $\\mathbf{x}' = \\mathbf{x} + \\Delta$ , for any interpolation samples $\\mathbf{x} + \\lambda \\cdot \\Delta$ with $\\lambda \\in [0,1]$ , it encourages the ratio of the perception changes, $z_{\\theta}(\\mathbf{x} + \\lambda \\cdot \\Delta) - z_{\\theta}(\\mathbf{x})$ and $z_{\\theta}(\\mathbf{x}') - z_{\\theta}(\\mathbf{x} + \\lambda \\cdot \\Delta)$ , respectively from $\\mathbf{x}$ to $\\mathbf{x} + \\lambda \\cdot \\Delta$ and from $\\mathbf{x} + \\lambda \\cdot \\Delta$ to $\\mathbf{x}'$ , to approach the ratio indicated by $\\lambda$ , such that the model perception would change smoothly along with the input perturbations from $\\mathbf{x}$ to $\\mathbf{x}'$ ." + ], + "image_footnote": [], + "bbox": [ + 106, + 88, + 893, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Perception as a new AT objective in addition to the conventional one to realize this idea.", + "bbox": [ + 89, + 436, + 482, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Definition 1 (Robust Perception). Provided an AT task with target model $\\theta$ , let $\\mathbf{x}$ and $\\mathbf{x}'$ be any pairs of corresponding benign and adversarial samples with $\\Delta = \\mathbf{x}' - \\mathbf{x}$ , for any hidden representation $h_{\\theta}(\\mathbf{x})$ denoting the model perception, our additional AT objective can be formulated as:", + "bbox": [ + 89, + 472, + 483, + 550 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\forall \\alpha \\in [ 0, 1 ], \\| h _ {\\pmb {\\theta}} (\\mathbf {x} + \\alpha \\cdot \\Delta) - h _ {\\pmb {\\theta}} (\\mathbf {x}) \\| = \\alpha \\cdot \\| h _ {\\pmb {\\theta}} (\\mathbf {x} ^ {\\prime}) - h _ {\\pmb {\\theta}} (\\mathbf {x}) \\|.\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 560, + 483, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To facilitate the understanding of the proposed Robust Perception, in addition to the most straightforward perspective that it encourages the change in model perception along with the adversarial perturbations, the effectiveness of this new objective can also be demonstrated from other two perspectives, local linearity [12] and Lipschitz regularization [1, 11, 21, 33], both of which also help smooth the decision boundary and mitigate the current trade-off problem. Below, we further detail them theoretically.", + "bbox": [ + 89, + 590, + 483, + 727 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Theorem 1. Let $H$ be the Hessian Matrix such that $H_{h_\\theta}(\\mathbf{x}) = \\nabla_{\\mathbf{x}}^2 h_\\theta(\\mathbf{x})$ , then with the new optimization objective of Robust Perception, we have:", + "bbox": [ + 89, + 732, + 485, + 777 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\forall \\Delta , \\Delta^ {\\top} \\cdot H _ {h _ {\\theta}} (\\mathbf {x}) \\cdot \\Delta \\rightarrow 0.\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 787, + 382, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proof. We defer the proof to Appendix B.1.", + "bbox": [ + 89, + 818, + 380, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This means Robust Perception limits the second-order (and higher-order as well) nonlinear effects within the adversarial perturbation to the model perception, such that the perception would change mainly along with the linear", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "term of the perturbation, and thus be smoother. This also aligns with the local linearity hypothesis of adversarial samples [12], which assumes that linear approximations dominate the model behavior under adversarial perturbations.", + "bbox": [ + 511, + 436, + 906, + 497 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Theorem 2. Let $J$ be the Jacobian Matrix such that $J_{h_{\\theta}}(\\mathbf{x}) = \\nabla_{\\mathbf{x}} h_{\\theta}(\\mathbf{x})$ , then with the new optimization objective of Robust Perception, we have:", + "bbox": [ + 511, + 502, + 908, + 547 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\forall \\alpha \\in [ 0, 1 ], J _ {h _ {\\boldsymbol {\\theta}}} (\\mathbf {x} + \\boldsymbol {\\alpha} \\cdot \\Delta) \\rightarrow J _ {h _ {\\boldsymbol {\\theta}}} (\\mathbf {x}),\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 560, + 844, + 578 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "then with $\\| \\cdot \\|_{\\mathrm{spec}}$ denoting the Spectral Norm and $\\gamma$ being any micro value, given $\\| J_{h_\\theta}(\\mathbf{x} + \\alpha \\cdot \\Delta) - J_{h_\\theta}(\\mathbf{x})\\|_{\\mathrm{spec}} \\leq \\gamma$ , the function $h_\\theta(\\mathbf{x})$ can be referred to as $K$ -Lipschitz with the Lipschitz constant $K$ upper-bounded by:", + "bbox": [ + 511, + 589, + 906, + 650 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nK \\leq \\sup _ {\\mathbf {x}} \\| J _ {h _ {\\theta}} (\\mathbf {x}) \\| _ {\\operatorname {s p e c}} + \\gamma .\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 661, + 807, + 685 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proof. We defer the proof to Appendix B.2.", + "bbox": [ + 511, + 696, + 803, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This theorem tells us that Robust Perception also regularizes the change $\\gamma$ in Jacobian along with the adversarial perturbation, which can be interpreted as limiting the increase of the global Lipschitz constant by $\\gamma$ under perturbation. Previous works have demonstrated that this is expected to help achieve smoother decision boundaries and improve the accuracy-robustness trade-off [1, 11, 21, 33].", + "bbox": [ + 511, + 717, + 905, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Robust Perception Adversarial Training", + "text_level": 1, + "bbox": [ + 511, + 832, + 857, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on the above considerations, in this section, we propose a novel AT method named Robust Perception Adversarial Training (RPAT). It utilizes a new additional", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4695", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "regularization term appended to the conventional AT objective to achieve the Robust Perception. Formally, given a $C$ -class classification task $(C \\geq 2)$ and a dataset $\\mathcal{D} = \\{(\\mathbf{x}_i, y_i)\\}_{i=1,\\dots,n}$ with benign sample $\\mathbf{x}_i \\in \\mathbb{R}^d$ and supervised label $y_i \\in \\{1,\\dots,C\\}$ , let $\\hat{\\mathbf{x}}_i'$ denotes the adversarial sample corresponding to $\\mathbf{x}_i$ , which is generated as $\\hat{\\mathbf{x}}_i' = \\arg \\max_{\\|\\mathbf{x}_i' - \\mathbf{x}_i\\|_p \\leq \\epsilon} \\mathbb{1}(\\arg \\max_{k=1,\\dots,C} h_\\theta(\\mathbf{x}_i'))^{(k)} \\neq y_i)$ , then based on the Definition 1 in Section 4.1, the adversarial risk to be minimized for the proposed RPAT w.r.t. $\\mathbf{x}_i$ can be formulated with the 0-1 loss [58] as:", + "bbox": [ + 89, + 90, + 483, + 244 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {R} (\\boldsymbol {\\theta}, \\mathbf {x} _ {i}) := \\mathbb {1} \\left(f _ {\\boldsymbol {\\theta}} \\left(\\hat {\\mathbf {x}} _ {i} ^ {\\prime}\\right) \\neq y _ {i}\\right) \\\\ + \\mathbb {1} \\left(\\frac {h _ {\\boldsymbol {\\theta}} \\left(\\tilde {\\mathbf {x}} _ {i}\\right) - h _ {\\boldsymbol {\\theta}} \\left(\\mathbf {x} _ {i}\\right)}{\\alpha} \\neq \\frac {h _ {\\boldsymbol {\\theta}} \\left(\\hat {\\mathbf {x}} _ {i} ^ {\\prime}\\right) - h _ {\\boldsymbol {\\theta}} \\left(\\tilde {\\mathbf {x}} _ {i}\\right)}{1 - \\alpha}\\right), \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 253, + 480, + 306 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\tilde{\\mathbf{x}}_i = \\mathbf{x}_i + \\alpha \\cdot (\\hat{\\mathbf{x}}_i' - \\mathbf{x}_i)$", + "bbox": [ + 89, + 316, + 302, + 333 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "While the proposed Robust Perception aims to improve learning efficacy for the failure cases, the above risk does not explicitly distinguish success and failure cases. This is because Robust Perception is designed to mitigate the gap of the failure cases in the effectiveness of learning compared with the success ones, rather than require any new characteristic beyond the success cases. In other words, the success cases are expected to also satisfy Robust Perception at the beginning. Therefore, considering the computational efficiency, it is not necessary to specifically exclude the success cases for the Robust Perception term.", + "bbox": [ + 88, + 333, + 482, + 498 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, by introducing practical surrogate losses to optimize the adversarial risk in Equation (2), we end up with our new RPAT method. To be specific, for the supervised learning term for adversarial samples, we adopt cross-entropy (CE) just as the conventional AT by default. Then, for the Robust Perception term, as it is formulated in a format of similarity regularization, various criterion that measures the similarity distance between the two input items are feasible options, such as mean squared error (MSE), Kullback-Leibler divergence (KL), Jensen-Shannon divergence (JS), and Cosine similarity. We simply adopt MSE loss for the basic version of RPAT here, and would also demonstrate the effectiveness of KL loss later in Section 5.3. With the logits $\\mathbf{z}(\\mathbf{x}_i,\\boldsymbol {\\theta})$ utilized as the hidden representation $h_\\theta (\\cdot)$ to denote the model perception, and $\\mathbf{p}(\\mathbf{x}_i,\\boldsymbol {\\theta})$ being the softmax output as given in Equation (1), the overall objective of RPAT can be formulated as:", + "bbox": [ + 89, + 500, + 483, + 755 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} ^ {R P A T} (\\boldsymbol {\\theta}, \\mathcal {D}, \\lambda , \\alpha) := \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left(\\mathcal {L} ^ {C E} \\left(\\mathbf {p} \\left(\\hat {\\mathbf {x}} _ {i} ^ {\\prime}, \\boldsymbol {\\theta}\\right), y _ {i}\\right) \\right. \\tag {3} \\\\ + \\lambda \\cdot \\mathcal {L} ^ {M S E} \\left( \\right.\\frac {\\mathbf {z} (\\tilde {\\mathbf {x}} _ {i} , \\boldsymbol {\\theta}) - \\mathbf {z} (\\mathbf {x} _ {i} , \\boldsymbol {\\theta})}{\\alpha} \\left| \\right.\\left| \\right. \\frac {\\mathbf {z} (\\hat {\\mathbf {x}} _ {i} ^ {\\prime} , \\boldsymbol {\\theta}) - \\mathbf {z} (\\tilde {\\mathbf {x}} _ {i} , \\boldsymbol {\\theta})}{1 - \\alpha}\\left. \\right)\\left. \\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 766, + 482, + 829 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda$ is a weight hyper-parameter to balance the conventional AT objective and our new Robust Perception objective. An intuitive illustration of the proposed RPAT method is provided in Figure 3.", + "bbox": [ + 89, + 840, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 513, + 90, + 645, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following previous AT works, the effectiveness of the proposed RPAT is evaluated on CIFAR-10, CIFAR-100 [22] and Tiny-ImageNet [23] datasets. The main experiments can be divided into two parts. First, based on Tack et al. [40], we implement RPAT upon three common AT baselines, namely PGD-AT, TRADES, and MART, as well an advanced method, Consistency-AT, with ResNet-18 [15] architecture. This is to show the general effectiveness of the proposed Robust Perception in improving the current trade-off. Second, by integrating RPAT into one of the current SOTA methods, ReBAT [47], with PreActResNet-18 [16] and WideResNet-34-10 [56] architectures, we achieve new SOTA performance beyond 12 SOTA methods on the accuracy-robustness trade-off problem. This is to demonstrate the advancement achieved by the proposed RPAT in the AT area. The detailed experimental setup is provided in Appendix A.", + "bbox": [ + 511, + 114, + 906, + 372 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Evaluation Measures", + "text_level": 1, + "bbox": [ + 511, + 383, + 712, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "There are five measures in total adopted to evaluate the experimental methods from three different perspectives. Firstly, we report clean accuracy on benign test data, which reflects the ability of the robust models to maintain performance in benign cases. Secondly, we evaluate adversarial robustness under two adversarial attack methods, namely PGD-20 [29] and Auto-Attack (AA) [7], with default settings aligning with Appendix A and random starts. Finally, to test the performance w.r.t. the current accuracy-robustness trade-off problem in AT, on the one hand, we report the mean of the clean accuracy and the robust score under Auto-Attack acquired above, and on the other hand, we introduce the Natural-Robustness Ratio (NRR) [13]:", + "bbox": [ + 511, + 406, + 906, + 602 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nN R R = \\frac {2 \\times C l e a n A c c u r a c y \\times A d v e r s a r i a l R o b u s t n e s s}{C l e a n A c c u r a c y + A d v e r s a r i a l R o b u s t n e s s},\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 612, + 895, + 645 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "which emphasizes how well the robust model strikes a balance between clean accuracy and robustness. A lower NRR value suggests that the model might prioritize one side over the other. All the results reported are averages of three runs, with the \"best\" performance of each run recorded on the best checkpoint achieving the highest PGD-20 accuracy, which ensures that the training process is isolated from AA, the test-time adversary to evaluate the robustness.", + "bbox": [ + 511, + 655, + 906, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Effectiveness in Improving Baselines", + "text_level": 1, + "bbox": [ + 511, + 787, + 830, + 803 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The proposed RPAT can be easily implemented with existing AT methods by simply appending the Robust Perception term. In this section, we illustrate the improvements that RPAT can bring to our four experimental benchmarks, namely PGD-AT, TRADES, MART, and Consistency-AT, on CIFAR-10, CIFAR-100, and Tiny-ImageNet with", + "bbox": [ + 511, + 809, + 905, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4696", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c3e6018b054b8a83a43decac3987908935d105e7e267d5588c66ea58ba3925d4.jpg", + "table_caption": [ + "Table 1. Improvements that the proposed RPAT achieved upon four benchmarks. The results are acquired with ResNet-18 and ${\\ell }_{\\infty }$ norm." + ], + "table_footnote": [], + "table_body": "
DatasetMethodBenchmarks+RPAT (Ours)
CleanPGD-20AAMeanNRRCleanPGD-20AAMeanNRR
bestbestbestbestfinalbestfinalbestfinal
CIFAR-10PGD-AT82.9250.6146.7464.83059.78283.2082.9951.2950.9748.0047.6065.60060.878
TRADES79.6751.9147.6263.64559.61080.0280.0251.2151.2147.4747.4563.74559.590
MART77.9353.0046.7062.31558.40275.4475.9552.8852.8747.0347.3061.23557.940
Consistency-AT83.4251.9647.7265.57060.71184.1284.0452.3352.1448.9849.0666.55061.911
CIFAR-100PGD-AT56.5628.8025.0240.79034.69358.2258.5829.1628.5724.8824.2341.55034.862
TRADES55.3929.3624.5139.95033.98357.5057.2529.4229.0025.0524.4541.27534.897
MART49.8330.3825.0037.41533.29550.7250.7230.3330.3325.3425.0638.03033.796
Consistency-AT58.5329.2825.3941.96035.41760.3360.5129.9729.6426.3125.9743.32036.641
Tiny-ImageNetPGD-AT46.3221.5817.0731.69524.94747.6848.3222.1322.0017.7717.9132.72525.891
TRADES46.7521.5216.6031.67524.50048.7748.8521.9821.9316.9216.7132.84525.124
MART39.7022.9017.1828.44023.98241.7641.6623.4023.2917.7917.7829.77524.951
Consistency-AT46.5421.8917.6032.07025.54149.7449.8523.1622.8418.8418.7434.29027.329
", + "bbox": [ + 94, + 114, + 903, + 316 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/40a2fbae0b216eace616d02a012d3508daf7058c29acafee77646090d21a6e77.jpg", + "table_caption": [ + "Table 2. Improvements that the proposed RPAT achieved upon four benchmarks. The results are acquired with ResNet-18 and ${\\ell }_{2}$ norm." + ], + "table_footnote": [], + "table_body": "
DatasetMethodBenchmarks+RPAT (Ours)
CleanPGD-20AAMeanNRRCleanPGD-20AAMeanNRR
bestbestbestbestfinalbestfinalbestfinal
CIFAR-10PGD-AT87.7667.9266.3677.06075.57488.2088.4568.5668.5267.6367.6277.91576.557
TRADES83.9968.6065.9374.96073.87285.1784.8068.9568.9067.6767.3776.42075.418
MART84.0968.3266.2875.18574.13084.6584.6568.6068.6066.6966.7075.67074.604
Consistency-AT88.7669.3567.4678.11076.65889.3889.5270.4170.3669.4469.5579.41078.158
CIFAR-100PGD-AT65.0041.5439.2752.13548.96065.1465.4441.6141.2839.2339.2052.18548.969
TRADES61.2543.0540.1550.7048.50562.4162.6443.5643.1040.3640.1251.38549.020
MART60.0843.4239.8549.96547.9260.5760.5743.9243.9040.3640.3650.46548.442
Consistency-AT65.1442.2839.9252.53049.50365.5466.8142.8142.5140.2640.6152.90049.880
Tiny-ImageNetPGD-AT59.3643.1240.9050.13048.43160.2860.2844.0644.0641.8941.8751.08549.430
TRADES57.6144.7042.3349.97048.80258.4658.5044.8744.6542.4242.4450.44049.165
MART56.4744.8442.2249.34548.31657.1756.9545.0244.6742.5142.1649.84048.762
Consistency-AT61.4944.9242.5952.04050.32462.1962.6445.2844.8743.1742.8552.68050.963
", + "bbox": [ + 94, + 354, + 901, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ResNet-18, demonstrating the wide effectiveness of our method. The results under $\\ell_{\\infty}$ and $\\ell_{2}$ threat models are shown in Table 1 and Table 2, respectively, where the best clean accuracy, AA robustness, as well mean and NRR for their trade-off on each of the datasets are marked in bold, while individual scores not improved are marked with underline. Note that PGD-20 is the indicator for selecting the \"best\" checkpoint as explained in Section 5.1, and the \"final\" scores are additionally provided to show that RPAT does not suffer from serious robust-overfitting, which is also evidence for appropriate algorithm design. Thus, these fields are not involved in the comparison. It can be found that RPAT achieves significant improvement for most scores among the tables, where Consistency-AT+RPAT is basically the most competitive for all the experimental groups, which demonstrates the general effectiveness of our method in mitigating the current trade-off problem.", + "bbox": [ + 89, + 580, + 480, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Advancement beyond 12 SOTAs", + "text_level": 1, + "bbox": [ + 89, + 848, + 372, + 863 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To demonstrate the advancement that the proposed RPAT contributes, we further integrate RPAT into ReBAT [47],", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "one of the current SOTA AT methods on the accuracy-robustness trade-off problem. It implements WA [18], averaging multiple points along the SGD trajectory for better generalization, through the exponential moving average (EMA) [37] strategy with a decay rate of 0.999. Thus, we also involve WA itself (i.e., with PGD-AT), as well as $\\mathrm{KD + SWA}$ [5] which shares a similar idea, for comparison. Other differences in ReBAT include $KL$ loss and a reduced learning rate decay factor from 0.1 to 0.5. We refer to our method upon ReBAT as $\\mathrm{RPAT}^{++}$ , and compare it with 12 previous SOTAs introduced in Section 2.2, respectively with PreActResNet-18 and WideResNet-34-10 as shown in Table 3 and Table 4. The clean and AA scores of the SOTAs are from the best checkpoint among their original papers, Kanai et al. [19], Wang et al. [47], and our reproduction if any (so due to the lack of previous record, ADR and CURE are excluded for PreActResNet-18). Beyond the basic CIFAR-10 results under $\\ell_{\\infty}$ norm, Table 3 involves CIFAR-100 and $\\ell_2$ norm, and Table 4 further consider the Cutmix [55] data augmentation strategy as marked by “*”, which ensures the comprehensiveness of our comparison.", + "bbox": [ + 511, + 580, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4697", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8e6de9a6b1bcf620105e396f5a7a89d55e46a412280ed19cbd3c9a8ee1a8182c.jpg", + "table_caption": [ + "Table 3. Comparison of the proposed $\\mathrm{RPAT^{+ + }}$ with current SOTAs on accuracy-robustness trade-off with PreActResNet-18. The best results are marked in bold." + ], + "table_footnote": [], + "table_body": "
NormMethodCIFAR-10CIFAR-100
CleanAAMeanNRRCleanAAMeanNRR
\\( \\ell_{\\infty} \\)WAUAI'1883.5049.8966.69562.46157.2625.8341.54535.601
MMAICLR'2085.5037.2061.35051.84460.6018.4039.50028.229
AWPNeurIPS'2081.1150.0965.60061.93354.1025.1639.63034.347
GAIRATICLR'2178.7037.7058.20050.97952.0019.8035.90028.680
KD+SWAICLR'2184.0649.8266.94062.56257.1725.6641.41535.422
EWATICML'2182.8048.2065.50060.93154.2023.5238.86032.805
MAILNeurIPS'2179.5039.6059.55052.86746.5016.7031.60024.574
TEICLR'2282.0450.1266.08062.22556.4125.8441.12535.444
SOVRICML'2381.9049.4065.65061.62852.1024.3038.20033.142
ReBATNeurIPS'2382.0950.7266.40562.70056.1327.6041.86537.004
RPAT++Ours82.6351.0066.81563.07256.8427.6842.26037.230
\\( \\ell_{2} \\)ReBATNeurIPS'2388.7971.0079.89578.90565.5842.6754.12551.701
RPAT++Ours89.0671.2680.16079.17265.6342.8554.24051.848
", + "bbox": [ + 94, + 127, + 594, + 330 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1ca8e9d1d96a6a2979fa47d40176a820aea878f852909e21e84c5a72b163e855.jpg", + "table_caption": [ + "Table 4. Comparison of ${\\mathrm{{RPAT}}}^{+ + }$ with SOTAs on CIFAR-10 with WideResNet-34-10 and ${\\ell }_{\\infty }$ norm." + ], + "table_footnote": [], + "table_body": "
MethodCleanAAMeanNRR
WAUAI'1887.6652.6570.15565.787
MMAICLR'2087.8043.1065.45057.818
AWPNeurIPS'2085.6353.3269.47565.718
GAIRATICLR'2183.0041.8062.40055.599
KD+SWAICLR'2187.4553.5970.52066.456
EWATICML'2186.0051.6068.80064.500
MAILNeurIPS'2182.2043.3062.75056.721
TEICLR'2285.9752.8869.42565.482
SOVRICML'2385.0053.1069.05065.366
ReBATNeurIPS'2385.2554.7870.01566.700
ADRICLR'2484.6753.2568.96065.381
CUREICLR'2487.0552.1069.57565.186
\\(RPAT^{++}\\)Ours86.7654.9770.86567.300
ReBAT*NeurIPS'2386.6655.6471.15067.769
\\(RPAT^{++*}\\)Ours87.5755.7971.68068.158
", + "bbox": [ + 607, + 128, + 898, + 330 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/feec7bf49567c597de2bfb94966fcd39f9b246c998a5ab54408b2f62b07b554c.jpg", + "image_caption": [ + "(a) Impact of loss function." + ], + "image_footnote": [], + "bbox": [ + 96, + 353, + 276, + 462 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a1d5beb1720d50f0b83416088093087387fc3e74d0ea117a517ad85dff5164c0.jpg", + "image_caption": [ + "(b) Impact of hyper-parameter $\\alpha$", + "Figure 4. Two ablation studies of the proposed RPAT on CIFAR-10 with ResNet-18 under $\\ell_{\\infty}$ norm. The MSE loss in (a) and the points with $\\alpha = 0.5$ in (b) correspond to the record of \"PGD-AT+RPAT\" on CIFAR-10 in Table 1." + ], + "image_footnote": [], + "bbox": [ + 276, + 353, + 478, + 463 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The results show that the proposed $\\mathrm{RPAT^{+ + }}$ achieves new SOTA performance for most of the measures beyond previous SOTAs on the accuracy-robustness trade-off problem, which strongly supports the contribution of this work.", + "bbox": [ + 89, + 575, + 483, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 648, + 243, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As the proposed RPAT can be implemented upon existing AT methods by simply appending the Robust Perception term into their original objectives, the comparison in performance between and after adding RPAT in Section 5.2 and Section 5.3 have achieved a comprehensive ablation study on the effectiveness of RPAT. Therefore, in this section, we mainly investigate the impact of two setting details, respectively the loss function adopted for the Robust Perception term (Figure 4 (a)) and the specific value for the hyperparameter $\\alpha$ (Figure 4 (b)). Besides, we further discuss an optional trick suggested by ReBAT but excluded in the above experiments in Appendix 5.3.", + "bbox": [ + 89, + 672, + 482, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Specifically, just as we expected, different loss functions to calculate the similarity between model perceptions, including MSE, KL, JS and Cosine, would not significantly", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "impact the final performance of RPAT, which suggests that the effectiveness of our method is from the successful conceptual design of Robust Perception, instead of relying on specific implementation of the regularization term. Then, for $\\alpha$ , we try 0.2 and 0.8 besides the default value 0.5 on the one hand, and also introduce beta distribution $\\beta \\sim Beta(0.75, 0.75)$ as suggested by Zhang et al. [57] on the other hand. Specifically, we try $\\alpha = \\min(\\beta, (1 - \\beta))$ and $\\alpha = \\max(\\beta, (1 - \\beta))$ , namely indicated as \"Beta-\" and \"Beta+\" in (Figure 4 (b)), as they respectively approach adversarial and benign samples more. The results show that the default $\\alpha = 0.5$ is an appropriate choice, which is reasonable because the midpoint is expected to be more representative for the ideal target that Robust Perception holds for all $\\alpha \\in [0, 1]$ .", + "bbox": [ + 511, + 354, + 906, + 582 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 601, + 633, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work studies the current accuracy-robustness trade-off problem in AT. To the best of our knowledge, this is the first work that explicitly attributes the trade-off problem to the over-sufficient (rather than insufficient) learning of hard adversarial samples. We surprisingly reveal that such hard samples are already learned better than those successfully defended in model perception, which means the conventional AT objective is no longer effective to further optimize the model w.r.t. these hard samples. Instead, it forces the model to ignore the perturbations, in turn degrading their effectiveness in supporting the establishment of decision boundary to an appropriate location. To mitigate this, our new AT objective, Robust Perception, encourages smoother perception transition towards the decision boundary along with perturbations, based on which the proposed RPAT method achieves new SOTA performance. Still, exploring better methods to approximate this ideal objective would be a promising future direction.", + "bbox": [ + 511, + 628, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4698", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the National Natural Science Foundation of China (No. 62471420), GuangDong Basic and Applied Basic Research Foundation (2025A15150122 96), and CCF-Tencent Rhino-Bird Open Research Fund.", + "bbox": [ + 89, + 113, + 483, + 172 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 195, + 187, + 210 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Abulikemu Abuduweili and Changliu Liu. Estimating neural network robustness via lipschitz constant and architecture sensitivity. In Safe and Robust Robot Learning for Operation in the Real World Workshop @ CoRL, 2024. 2, 5, 11", + "[2] Anish Athalye, Nicholas Carlini, and David Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In ICML, 2018. 2", + "[3] Tao Bai, Jinqi Luo, Jun Zhao, Bihan Wen, and Qian Wang. Recent advances in adversarial training for adversarial robustness. In IJCAI, 2021. 2, 3", + "[4] Battista Biggio, Igino Corona, Davide Maiorca, Blaine Nelson, Nedim Šrndić, Pavel Laskov, Giorgio Giacinto, and Fabio Roli. Evasion attacks against machine learning at test time. In ECML-PKDD, 2013. 1", + "[5] Tianlong Chen, Zhenyu Zhang, Sijia Liu, Shiyu Chang, and Zhangyang Wang. Robust overfitting may be mitigated by properly learned smoothening. In ICLR, 2021. 3, 7", + "[6] Minhao Cheng, Qi Lei, Pin-Yu Chen, Inderjit Dhillon, and Cho-Jui Hsieh. Cat: Customized adversarial training for improved robustness. In *IJCAI*, 2022. 2, 3", + "[7] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In ICML, 2020. 2, 6", + "[8] Gavin Weiguang Ding, Yash Sharma, Kry Yik Chau Lui, and Ruitong Huang. Mma training: Direct input space margin maximization through adversarial training. In ICLR, 2020. 2, 3", + "[9] Yinpeng Dong, Qi-An Fu, Xiao Yang, Tianyu Pang, Hang Su, Zihao Xiao, and Jun Zhu. Benchmarking adversarial robustness on image classification. In CVPR, 2020. 2", + "[10] Yinpeng Dong, Ke Xu, Xiao Yang, Tianyu Pang, Zhijie Deng, Hang Su, and Jun Zhu. Exploring memorization in adversarial training. In ICLR, 2022. 2, 3", + "[11] Mahyar Fazlyab, Taha Entesari, Aniket Roy, and Rama Chellappa. Certified robustness via dynamic margin maximization and improved lipschitz regularization. In NeurIPS, 2023. 2, 5, 11", + "[12] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In ICLR, 2015. 1, 2, 5, 11", + "[13] Shruthi Gowda, Bahram Zonooz, and Elahe Arani. Conserve-update-revise to cure generalization and robustness trade-off in adversarial training. In ICLR, 2024. 3, 6", + "[14] Chuan Guo, Mayank Rana, Moustapha Cisse, and Laurens van der Maaten. Countering adversarial images using input transformations. In ICLR, 2018. 1, 2", + "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6" + ], + "bbox": [ + 93, + 220, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In ECCV, 2016. 6", + "[17] Xiaowei Huang, Daniel Kroening, Wenjie Ruan, James Sharp, Youcheng Sun, Emese Thamo, Min Wu, and Xinping Yi. A survey of safety and trustworthiness of deep neural networks: Verification, testing, adversarial attack and defence, and interpretability. Computer Science Review, 2020. 1", + "[18] Pavel Izmailov, Dmitrii Podoprikhin, Timur Garipov, Dmitry Vetrov, and Andrew Gordon Wilson. Averaging weights leads to wider optima and better generalization. In UAI, 2018. 3, 7", + "[19] Sekitoshi Kanai, Shin'ya Yamaguchi, Masanori Yamada, Hiroshi Takahashi, Kentaro Ohno, and Yasutoshi Ida. One-vs-the-rest loss to focus on important samples in adversarial training. In ICML, 2023. 2, 3, 7", + "[20] Minseon Kim, Jihoon Tack, Jinwoo Shin, and Sung Ju Hwang. Entropy weighted adversarial training. In Adversarial Machine Learning Workshop @ ICML, 2021. 3", + "[21] Vishaal Krishnan, Al Makedah, Abed AlRahman, and Fabio Pasqualetti. Lipschitz bounds and provably robust training by laplacian smoothing. In NeurIPS, 2020. 2, 5, 11", + "[22] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. 2009. 6", + "[23] Fei-Fei Li, Andrej Karpathy, and Justin Johnson. Tiny image dataset. 2015. 6", + "[24] Zi Liang, Pinghui Wang, Ruofei Zhang, Nuo Xu, Shuo Zhang, Lifeng Xing, Haitao Bai, and Ziyang Zhou. Merge: Fast private text generation. In AAAI, 2024. 1", + "[25] Fangzhou Liao, Ming Liang, Yinpeng Dong, Tianyu Pang, Xiaolin Hu, and Jun Zhu. Defense against adversarial attacks using high-level representation guided denoiser. In CVPR, 2018. 1, 2", + "[26] Weilin Lin, Nanjun Zhou, Yanyun Wang, Jianze Li, Hui Xiong, and Li Liu. Backdoor: A comprehensive benchmark for backdoor learning in diffusion model. arXiv preprint arXiv:2502.11798, 2025. 1", + "[27] Feng Liu, Bo Han, Tongliang Liu, Chen Gong, Gang Niu, Mingyuan Zhou, Masashi Sugiyama, et al. Probabilistic margins for instance reweighting in adversarial training. In NeurIPS, 2021. 2, 3", + "[28] Xin Liu, Yichen Yang, Kun He, and John E Hopcroft. Parameter interpolation adversarial training for robust image classification. IEEE Transactions on Information Forensics and Security (TIFS), 2025. 2, 3", + "[29] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In ICLR, 2018. 2, 3, 6", + "[30] Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. In ICLR, 2018. 12", + "[31] Tianyu Pang, Xiao Yang, Yinpeng Dong, Hang Su, and Jun Zhu. Bag of tricks for adversarial training. In ICLR, 2021. 11", + "[32] Nicolas Papernot, Patrick McDaniel, Ian Goodfellow, Somesh Jha, Z Berkay Celik, and Ananthram Swami. Practi" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4699", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "cal black-box attacks against machine learning. In AsiaCCS, 2017. 1, 2", + "[33] Patricia Pauli, Anne Koch, Julian Berberich, Paul Kohler, and Frank Allgower. Training robust neural networks using lipschitz bounds. IEEE Control Systems Letters, 2021. 2, 5, 11", + "[34] Rahul Rade and Seyed-Mohsen Moosavi-Dezfooli. Reducing excessive margin to achieve a better accuracy vs. robustness trade-off. In ICLR, 2022. 2, 3", + "[35] Aditi Raghunathan, Sang Michael Xie, Fanny Yang, John Duchi, and Percy Liang. Adversarial training can hurt generalization. In Identifying and Understanding Deep Learning Phenomena Workshop @ ICML, 2019.", + "[36] Aditi Raghunathan, Sang Michael Xie, Fanny Yang, John Duchi, and Percy Liang. Understanding and mitigating the tradeoff between robustness and accuracy. In ICML, 2020. 2, 3", + "[37] Sylvestre-Alvise Rebuffi, Sven Gowal, Dan Andrei Calian, Florian Stimberg, Olivia Wiles, and Timothy A Mann. Data augmentation can improve robustness. In NeurIPS, 2021. 7", + "[38] David Stutz, Matthias Hein, and Bernt Schiele. Confidence-calibrated adversarial training: Generalizing to unseen attacks. In ICML, 2020. 3", + "[39] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In ICLR, 2014. 1, 2", + "[40] Jihoon Tack, Sihyun Yu, Jongheon Jeong, Minseon Kim, Sung Ju Hwang, and Jinwoo Shin. Consistency regularization for adversarial robustness. In AAAI, 2022. 3, 6", + "[41] Lue Tao, Lei Feng, Jinfeng Yi, Sheng-Jun Huang, and Songcan Chen. Better safe than sorry: Preventing delusive adversaries with adversarial training. In NeurIPS, 2021. 2", + "[42] Florian Tramer, Nicholas Carlini, Wieland Brendel, and Aleksander Madry. On adaptive attacks to adversarial example defenses. In NeurIPS, 2020. 2", + "[43] Dimitris Tsipras, Shibani Santurkar, Logan Engstrom, Alexander Turner, and Aleksander Madry. Robustness may be at odds with accuracy. In ICLR, 2019. 2, 3", + "[44] Haotao Wang, Tianlong Chen, Shupeng Gui, TingKuei Hu, Ji Liu, and Zhangyang Wang. Once-for-all adversarial training: In-situ tradeoff between robustness and accuracy for free. In NeurIPS, 2020. 2, 3", + "[45] Yisen Wang, Difan Zou, Jinfeng Yi, James Bailey, Xingjun Ma, and Quanquan Gu. Improving adversarial robustness requires revisiting misclassified examples. In ICLR, 2020. 1, 3", + "[46] Yanyun Wang, Dehui Du, Haibo Hu, Zi Liang, and Yuan-hao Liu. Tsfool: Crafting highly-imperceptible adversarial time series through multi-objective attack. arXiv preprint arXiv:2209.06388, 2022. 2", + "[47] Yifei Wang, Liangchen Li, Jiansheng Yang, Zhouchen Lin, and Yisen Wang. Balance, imbalance, and rebalance: Understanding robust overfitting from a minimax game perspective. In NeurIPS, 2023. 3, 6, 7", + "[48] Yanyun Wang, Li Liu, Zi Liang, Qingqing Ye, and Haibo Hu. New paradigm of adversarial training: Releasing accuracy" + ], + "bbox": [ + 91, + 90, + 482, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "robustness trade-off via dummy class. arXiv preprint arXiv:2410.12671, 2024. 2", + "[49] Dongxian Wu, Shu-Tao Xia, and Yisen Wang. Adversarial weight perturbation helps robust generalization. In NeurIPS, 2020. 3", + "[50] Yu-Yu Wu, Hung-Jui Wang, and Shang-Tse Chen. Annealing self-distillation rectification improves adversarial training. In ICLR, 2024. 3", + "[51] Cihang Xie, Jianyu Wang, Zhishuai Zhang, Zhou Ren, and Alan Yuille. Mitigating adversarial effects through randomization. In ICLR, 2018. 1, 2", + "[52] Weilin Xu, David Evans, and Yanjun Qi. Feature squeezing: Detecting adversarial examples in deep neural networks. In NDSS, 2018. 1, 2", + "[53] Shuo Yang and Chang Xu. One size does not fit all: Data-adaptive adversarial training. In ECCV, 2022. 2, 3", + "[54] Jia-Li Yin, Bin Chen, Wanqing Zhu, Bo-Hao Chen, and Ximeng Liu. Push stricter to decide better: A class-conditional feature adaptive framework for improving adversarial robustness. IEEE Transactions on Information Forensics and Security (TIFS), 2023. 2, 3", + "[55] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In ICCV, 2019. 7", + "[56] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In BMVC, 2016. 6", + "[57] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. Mixup: Beyond empirical risk minimization. In ICLR, 2019. 8", + "[58] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In ICML, 2019, 2, 3, 6, 11", + "[59] Jingfeng Zhang, Jianing Zhu, Gang Niu, Bo Han, Masashi Sugiyama, and Mohan Kankanhalli. Geometry-aware instance-reweighted adversarial training. In ICLR, 2021. 2, 3", + "[60] Ziyu Zhou, Gengyu Lyu, Yiming Huang, Zihao Wang, Ziyu Jia, and Zhen Yang. Sdformer: Transformer with spectral filter and dynamic attention for multivariate time series long-term forecasting. In IJCAI, 2024. 1" + ], + "bbox": [ + 516, + 92, + 903, + 688 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4700", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Failure Cases Are Better Learned But Boundary Says Sorry_ Facilitating Smooth Perception Change for Accuracy-Robustness Trade-Off in Adversarial Training/9886c220-8af0-438e-8e12-4de3189d4e86_model.json b/2025/Failure Cases Are Better Learned But Boundary Says Sorry_ Facilitating Smooth Perception Change for Accuracy-Robustness Trade-Off in Adversarial Training/9886c220-8af0-438e-8e12-4de3189d4e86_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e4910997462386e2025154679f6c762b8bb8a774 --- /dev/null +++ b/2025/Failure Cases Are Better Learned But Boundary Says Sorry_ Facilitating Smooth Perception Change for Accuracy-Robustness Trade-Off in Adversarial Training/9886c220-8af0-438e-8e12-4de3189d4e86_model.json @@ -0,0 +1,2057 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.131, + 0.908, + 0.178 + ], + "angle": 0, + "content": "Failure Cases Are Better Learned But Boundary Says Sorry: Facilitating Smooth Perception Change for Accuracy-Robustness Trade-Off in Adversarial Training" + }, + { + "type": "text", + "bbox": [ + 0.399, + 0.204, + 0.6, + 0.222 + ], + "angle": 0, + "content": "Yanyun Wang Li Liu*" + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.227, + 0.768, + 0.246 + ], + "angle": 0, + "content": "The Hong Kong University of Science and Technology (Guangzhou)" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.279, + 0.327, + 0.296 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.311, + 0.486, + 0.765 + ], + "angle": 0, + "content": "Adversarial Training (AT) is one of the most effective methods to train robust Deep Neural Networks (DNNs). However, AT creates an inherent trade-off between clean accuracy and adversarial robustness, which is commonly attributed to the more complicated decision boundary caused by the insufficient learning of hard adversarial samples. In this work, we reveal a counterintuitive fact for the first time: From the perspective of perception consistency, hard adversarial samples that can still attack the robust model after AT are already learned better than those successfully defended. Thus, different from previous views, we argue that it is rather the over-sufficient learning of hard adversarial samples that degrades the decision boundary and contributes to the trade-off problem. Specifically, the excessive pursuit of perception consistency would force the model to view the perturbations as noise and ignore the information within them, which should have been utilized to induce a smoother perception transition towards the decision boundary to support its establishment to an appropriate location. In response, we define a new AT objective named Robust Perception, encouraging the model perception to change smoothly with input perturbations, based on which we propose a novel Robust Perception Adversarial Training (RPAT) method, effectively mitigating the current accuracy-robustness trade-off. Experiments on CIFAR-10, CIFAR-100, and Tiny-ImageNet with ResNet-18, PreActResNet-18, and WideResNet-34-10 demonstrate the effectiveness of our method beyond four common baselines and 12 state-of-the-art (SOTA) works. The code is available at https://github.com/FlaAI/RPAT." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.792, + 0.222, + 0.808 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.818, + 0.484, + 0.88 + ], + "angle": 0, + "content": "Deep neural networks (DNNs) have demonstrated impressive performance across various real-world applications [24, 26, 60]. However, they are inherently vulnerable to adversarial attacks [4, 39]. Specifically, even impercep" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.277, + 0.71, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.725, + 0.277, + 0.905, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.401, + 0.71, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.402, + 0.905, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.522, + 0.52, + 0.9, + 0.548 + ], + "angle": 0, + "content": "(c) (d) Contour line of model perception" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.56, + 0.909, + 0.77 + ], + "angle": 0, + "content": "Figure 1. Success and failure cases of AT defense from the perspectives of prediction and perception, respectively. (a) and (b): AT works through building a robust ball with radius \\( d \\) such that the model prediction remains unchanged within it. Then if the robust \\( d \\)-ball fails to contain the \\( \\epsilon \\)-ball, there would still be failure cases beating the robust model under perturbation \\( \\delta \\) within the \\( \\epsilon \\)-ball. (c) and (d): In this work, we reveal that the smaller \\( d \\)-ball in the failure cases is not a result of more vulnerable model perception as previous intuition. In fact, as demonstrated by the contour line, which shows the perturbation ball to the same degree of perception change, the failure cases already achieve less perception change under similar \\( \\delta \\) within the \\( \\epsilon \\)-ball. This implies that the issue of these failure cases is not failing to learn perception consistency following the current AT objective, but failing to support the establishment of decision boundary to an appropriate location." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "tible perturbations to inputs can cause substantial changes in the outputs [12, 17]. As this problem raises significant security concerns [45], previously, several defense strategies, such as Defense Distillation [32], Feature Squeezing [52], Randomization [51], and Input Denoising [14, 25], have been proposed to improve the adversarial robustness of DNNs. However, most of these approaches have proven" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.466, + 0.901 + ], + "angle": 0, + "content": "*Corresponding Author: avrillliu@hkust-gz.edu.cn" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.958 + ], + "angle": 0, + "content": "4691" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "ineffective against more advanced adaptive attacks [2, 42], such as Auto-Attack [7]. Currently, Adversarial Training (AT) [12, 29] is regarded as one of the most effective methods for training robust DNNs [2, 9]." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.155, + 0.485, + 0.548 + ], + "angle": 0, + "content": "However, an accuracy-robustness trade-off problem have been widely identified for AT in recent studies [3, 28, 34-36, 43, 44, 48, 54, 58]. Compared with standard training, the improvement in adversarial robustness acquired by AT is usually at the expense of the reduction in clean accuracy, which is afraid to significantly degrade the user experience in clean context and hinder the application of AT in real-world practices. The current consensus is that AT aggravates the accuracy-robustness trade-off by creating overly complex decision boundaries, primarily because models struggle to sufficiently learn from hard adversarial samples during training [6, 8, 10, 19, 27, 28, 34, 48, 53, 54]. Specifically, as the conventional AT objective encourages the model to perceive any adversarial inputs within the \\(\\epsilon\\)-ball similar to the corresponding benign one to produce consistent predictions for them, previous works empirically assume that the wrong predictions of the hard adversarial samples are attributed to the insufficiently optimized perception of them under the current objective. Therefore, although defining \"hard sample\" from different perspectives (e.g., classification result and distance to decision boundary), previous works agree to further optimize the perception consistency on them, ending up with various learning strategies specific for them, such as data reweighting [27, 59], adaptive perturbation radius [8, 34], and non-one-hot supervised signal [10, 19]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.483, + 0.867 + ], + "angle": 0, + "content": "However, in this work, we reveal that for the robust model trained by convention AT, the statistical difference in model perception from benign to hard adversarial samples causing failure defense is already smaller than to others (i.e., those successfully defended). It implies that the model has even more sufficiently learned them under the current AT objective, and should have successfully defended them if the above assumption holds. Thus, as demonstrated in Figure 1, we suggest that the key factor here is no longer the optimization of model perception but the location of the decision boundary. In that case, keeping on pursuing over-sufficient small perception change would not only help little in further reducing the prediction error, but also force the model to view the perturbations as noise and ignore the information within them, which impairs their ability in claiming robust balls with reasonable gradients to push the decision boundary to an appropriate location. As a consequence, such sample-level imbalance in boundary establishment would naturally increase its complexity, harming generalization ability and degrading the trade-off between accuracy and robustness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To mitigate the over-sufficient learning of model perception on the failure cases and retain effective gradient sig" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.349 + ], + "angle": 0, + "content": "nals w.r.t. them to facilitate the establishment of the decision boundary to an appropriate location, firstly, we define a new AT objective named Robust Perception in addition to the conventional one, encouraging smoother perception changes towards the decision boundary along with the input perturbations. Rigorous theoretical derivations also demonstrate its effectiveness in smoothening the decision boundary and releasing the current trade-off problem from the perspectives of local linearity [12] and Lipschitz regularization [1, 11, 21, 33]. Secondly, based on our new objective, we propose a novel Robust Perception Adversarial Training (RPAT) method, concurrently improving clean accuracy and adversarial robustness in AT. Extensive experiments on CIFAR-10, CIFAR-100, and Tiny-ImageNet with ResNet-18, PreActResNet-18, and WideResNet-34-10 demonstrate the advancement of RPAT beyond four common benchmarks and 12 state-of-the-art (SOTA) works." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.361, + 0.807, + 0.377 + ], + "angle": 0, + "content": "2. Background and Related Works" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.385, + 0.836, + 0.401 + ], + "angle": 0, + "content": "2.1. Robustness and Adversarial Training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.407, + 0.905, + 0.588 + ], + "angle": 0, + "content": "Adversarial robustness reflects model vulnerability to malicious perturbations, which is commonly quantified by test accuracy under adversarial attacks [3, 39, 46]. Representatively, the Fast Gradient Sign Method (FGSM) [12] perturbs input samples along the gradient of the loss function, whereas the Projected Gradient Descent (PGD) attack [29] applies FGSM iteratively and projects the perturbation onto the \\(\\epsilon\\)-ball. Auto-Attack [7] introduces and combines several parameter-free attacks to mitigate practical issues such as hyper-parameter tuning and gradient masking. Due to their effectiveness, PGD and Auto-Attack are widely used to assess adversarial robustness, and so does this work." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.589, + 0.905, + 0.74 + ], + "angle": 0, + "content": "Previously, although many defense strategies including Defense Distillation [32], Feature Squeezing [52], Input Denoising [14, 25] and Randomization [51] have been developed, most rely on obfuscated gradients [2] and can be evaded by advanced adaptive attacks [42]. Nowadays, Adversarial Training [12, 29] is recognized as the most effective method to enhance adversarial robustness [2, 9, 41]. Unlike clean training on benign samples, AT directly learns on adversarily augmented data. Its objective can be formulated as the following min-max problem [29]:" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.748, + 0.905, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\boldsymbol {\\theta}} \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\max _ {\\| \\mathbf {x} _ {i} ^ {\\prime} - \\mathbf {x} _ {i} \\| _ {P} \\leq \\epsilon} \\mathcal {L} \\left(f _ {\\boldsymbol {\\theta}} \\left(\\mathbf {x} _ {i} ^ {\\prime}\\right), y _ {i}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.905, + 0.901 + ], + "angle": 0, + "content": "where \\( f_{\\pmb{\\theta}}(\\mathbf{x}_i) = \\arg \\max_{c=1,\\dots,C} \\mathbf{p}_c(\\mathbf{x}_i, \\pmb{\\theta}) \\) with \\( \\mathbf{p}_c \\) denoting the softmax probability of class \\( c \\) within the prediction of \\( \\mathbf{x}_i \\), \\( n \\) denotes the number of training samples, and \\( \\mathbf{x}_i' \\) is generated by adding the strongest \\( L_p \\)-norm perturbation within the \\( \\epsilon \\)-ball to the original sample \\( \\mathbf{x}_i \\). This objective aims at prediction consistency within the \\( \\epsilon \\)-ball, encouraging the model to also produce similar perception there." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4692" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.348 + ], + "angle": 0, + "content": "PGD-AT [29] is the first attempt at solving this min-max optimization iteratively by approximating the inner maximization via PGD to craft adversarial samples and then performing stochastic gradient descent (SGD) on these samples for the outer minimization. TRADES [58] first considers balancing clean accuracy and adversarial robustness in AT, which decomposes robust error into natural error and boundary error, with the latter occurring only if specific data points are sufficiently close to the decision boundary. MART [45] further refines AT by distinguishing misclassified from correctly classified samples, which suggests appending boundary error only for misclassified samples to the basic AT objective in PGD-AT. Consistency-AT [40] encourages the prediction between adversarial samples derived from different augmentations of the same benign sample to be similar, which explicitly enhances the prediction consistency further." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.363, + 0.422, + 0.379 + ], + "angle": 0, + "content": "2.2. Accuracy-Robustness Trade-Off in AT" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.386, + 0.483, + 0.703 + ], + "angle": 0, + "content": "An important problem identified for AT is the trade-off between clean accuracy and adversarial robustness [3, 28, 34-36, 43, 44, 54, 58], which means the robustness gained by AT often comes at the cost of reduced model accuracy compared with clean training. Consequently, AT can degrade user experience under benign conditions, which may seriously delay its real-world deployment. A widely acknowledged cause of this problem is that AT tends to learn more complicated decision boundaries than clean training [6, 10, 28, 34, 53, 54], thereby harming the ability of the robust model to generalize to unseen data. A number of previous works attribute this to the insufficient learning of hard adversarial samples [8, 10, 19, 27, 34, 59]. Though defining \"hard adversarial samples\" from different perspectives, such as the classification result and the distance to decision boundary, they basically reached a consensus that the model perception w.r.t. hard adversarial samples is not well-optimized, and as a consequence, fails to produce the consistent final prediction. In response, these works propose various solutions to help the model learn better prediction consistency for those hard adversarial samples." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.901 + ], + "angle": 0, + "content": "For instance, GAIRAT [59] proposes that hard samples nearer to the decision boundary are more critical but less robust and should be assigned a larger weight. MAIL [27] further proposes three types of probabilistic margins to quantify the closeness for better reweighting adversarial data accordingly. In contrast, MMA [8] suggests learning hard adversarial samples with reduced perturbations, which employs an adaptive \\(\\epsilon\\) for perturbations to maximize the margin between data and the decision boundary. EWAT [20] weighs the loss for each adversarial sample proportionally to the entropy of its prediction distribution during AT to focus on those with more uncertain labels. TE [10] reveals that the model may attempt to memorize certain hard ad" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.151 + ], + "angle": 0, + "content": "versarial samples during AT, as it is difficult to assign high-confidence one-hot labels for them [6, 38]. SOVR [19] suggests increasing the margins of logits for certain hard samples by switching to a one-vs-the-rest loss." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.153, + 0.905, + 0.379 + ], + "angle": 0, + "content": "Additionally, several works also study the trade-off problem from other perspectives. AWP [49] proposes a double perturbation mechanism that flattens the loss landscape by weight perturbation to improve robust generalization. KD+SWA [5] injects more smoothening during AT by leveraging knowledge distillation to smooth the logits and performing stochastic weight averaging [18] to smooth the weights. ADR [50] employs soft labels as a guidance mechanism, accurately reflecting the distribution shift under attack during AT. CURE [13] selectively updates specific layers while preserving others to enhance the learning capacity of the model. ReBAT [47] views AT as a dynamic min-max game between the model trainer and the attacker, rebalancing the two players by either constraining the trainer's capacity or intensifying the attack." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.392, + 0.898, + 0.409 + ], + "angle": 0, + "content": "3. Motivation: AT Over-Learns Failure Cases" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.418, + 0.905, + 0.538 + ], + "angle": 0, + "content": "It is basically common sense for DNN-based classification that the model learns to perceive input better and is thus able to predict it more precisely (especially when the model perception is indicated by its logits). Therefore, it seems reasonable at first glance that the previous works assume that the wrong prediction of the hard adversarial samples is due to the insufficient model perception of them, and propose various solutions upon this assumption in response." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.539, + 0.905, + 0.719 + ], + "angle": 0, + "content": "However, to the best of our knowledge, there is no previous discussion on a simple but crucial question: What if the assumption does not hold? If the model perception on the hard adversarial samples is already relatively close to the corresponding benign ones (e.g., at a similar or even better level than non-hard ones), then it would be no longer effective enough to further optimize the model on such samples through an objective simply pursuing prediction consistency as the conventional AT. Also, in turn, this can no longer effectively benefit the further reduction of the prediction error. Surprisingly, our following proof-of-concept experiments reveal that this is exactly the case." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.72, + 0.905, + 0.901 + ], + "angle": 0, + "content": "The proof-of-concept experiments are conducted on CIFAR-10 dataset with ResNet-18 architecture, with the results illustrated in Figure 2. As suggested by the x-axis, we involve three cases with different pairs of training and evaluation methods, which are respectively 1) trained on clean data and evaluated with random perturbations, 2) trained and evaluated with randomly perturbed data, and 3) trained with PGD-AT and evaluated with PGD-20 adversary. The budget of the random perturbations \\(\\epsilon = 8 / 255\\) is assigned the same as PGD, following our experimental setup detailed in Appendix A, so do other training and evaluation settings. Given that the conventional AT objective in Equation (1)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4693" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.273 + ], + "angle": 0, + "content": "encourages the model to produce a similar perception for benign and adversarial samples, we utilize the similarity in model perception between the corresponding benign and perturbed samples to estimate the learning sufficiency of the perturbed samples in AT. Specifically, the model perception is represented by the logits, and the similarity is measured through the mean squared error (MSE) score. We calculate the score respectively for success and failure cases of defense, which are divided according to the correctness of the final classification results of the perturbed samples. In addition, the model accuracy on clean and perturbed data is also recorded and illustrated together." + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.285, + 0.462, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.495, + 0.483, + 0.676 + ], + "angle": 0, + "content": "Figure 2. Proof-of-concept experiments for the different learning sufficiency in model perception for success and failure cases on CIFAR-10 with ResNet-18. In the x-axis, the \"Random\" for \"Training Algorithm\" means the training data is randomly perturbed within the budget \\(\\epsilon\\), while that for \"Perturbation Method\" indicates the attack used for evaluation on test data is random perturbation. The bars demonstrate the MSE score on logits between each pair of benign and perturbed samples, with the lines additionally showing the model accuracy on clean and perturbed data. Our results reveal a surprising phenomenon that, different from the clean training (with or without random noise), where failure cases are worse learned as we intuitively expected, AT already learns a better model perception for failure cases." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Surprisingly, we experimentally found that the failure cases already learn better in model perception than the success cases. First, for the clean trained model, the perception on failure cases is significantly worse than on the success ones, just as we expected. Although the difference between success and failure cases has decreased for the randomly trained model, the failure cases remain worse, which suggests that non-adversarial perturbations are influential but not enough to change the trend. However, unlike the previous intuition, the model trained by PGD-AT can already perceive failure cases better than success ones. Our finding updates the current knowledge that, unlike those in clean training, the failure cases in AT do not fail simply because they are not sufficiently learned. On the contrary," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "the model even learns them harder to try to achieve the prediction and perception consistency. The smaller perception change under similar perturbations is direct evidence for this point of view." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.153, + 0.907, + 0.516 + ], + "angle": 0, + "content": "Then, as a straightforward deduction, now that we have success cases with worse perception but failure cases with better perception, there should be a statistical difference between them in the maximum perception change they can tolerate while still maintaining the consistent prediction, just as demonstrated in Figure 1. In other words, the decision boundary is established relatively closer to the failure cases, which is instead the issue we should focus on, because such sample-level imbalance in boundary establishment can naturally complicate it and thus degrade the accuracy-robustness trade-off as mentioned in Section 2.2. In that case, the over-consistent model perception between benign and adversarial samples in the failure cases rather harms the final robustness, because this is expected to force the model to view the perturbations as noise and ignore the information within them, which should have been utilized to induce a smoother perception transition towards the decision boundary to support its establishment, thus impairing the ability of such cases to claim a reasonable robust distance during AT compared with success cases. Therefore, in this work, we argue that it is not insufficient but oversufficient learning of failure cases that contributes to the more complicated decision boundary and finally results in the current trade-off problem." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.533, + 0.688, + 0.551 + ], + "angle": 0, + "content": "4. Proposed Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.56, + 0.862, + 0.575 + ], + "angle": 0, + "content": "4.1. Robust Perception: A New AT Objective" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.905, + 0.672 + ], + "angle": 0, + "content": "In Section 3, the too-small perception change under perturbations caused by the over-sufficient learning of failure cases has been identified as an important factor contributing to the current trade-off problem, which in turn questions the effectiveness of the conventional AT objective on such failure cases." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Formally, under Equation (1) which implicitly optimize the perception distance \\(\\| h_{\\theta}(\\mathbf{x}^{\\prime}) - h_{\\theta}(\\mathbf{x})\\| \\to 0\\) between corresponding benign sample \\(\\mathbf{x}\\) and adversarial sample \\(\\mathbf{x}^{\\prime}\\), the gradient of the conventional AT loss w.r.t. the perturbation, \\(\\nabla_{\\| \\mathbf{x}^{\\prime} - \\mathbf{x}\\|}\\mathcal{L}\\), would also approach 0. In that case, the model cannot further optimize the decision boundary using the perturbation information. Thus, a straightforward idea to mitigate this is to encourage smoother perception change towards the decision boundary along with the input perturbation, such that \\(\\| h_{\\theta}(\\mathbf{x}^{\\prime}) - h_{\\theta}(\\mathbf{x})\\| \\propto \\| \\mathbf{x}^{\\prime} - \\mathbf{x}\\|\\), as long as the final prediction remains unchanged. This is expected to retain an effective gradient signal in the direction of the perturbation, which enables the model to utilize this information to guide the decision boundary to a more reasonable location. Accordingly, in this section, we propose Robust" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4694" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.089, + 0.895, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.34, + 0.908, + 0.412 + ], + "angle": 0, + "content": "Figure 3. Illustration of the proposed RPAT method. In addition to the conventional AT objective, which learns adversarial samples with the original supervised labels, RPAT introduces a novel regularization term to achieve the newly defined Robust Perception objective. Given benign sample \\(\\mathbf{x}\\) and adversarial samples \\(\\mathbf{x}' = \\mathbf{x} + \\Delta\\), for any interpolation samples \\(\\mathbf{x} + \\lambda \\cdot \\Delta\\) with \\(\\lambda \\in [0,1]\\), it encourages the ratio of the perception changes, \\(z_{\\theta}(\\mathbf{x} + \\lambda \\cdot \\Delta) - z_{\\theta}(\\mathbf{x})\\) and \\(z_{\\theta}(\\mathbf{x}') - z_{\\theta}(\\mathbf{x} + \\lambda \\cdot \\Delta)\\), respectively from \\(\\mathbf{x}\\) to \\(\\mathbf{x} + \\lambda \\cdot \\Delta\\) and from \\(\\mathbf{x} + \\lambda \\cdot \\Delta\\) to \\(\\mathbf{x}'\\), to approach the ratio indicated by \\(\\lambda\\), such that the model perception would change smoothly along with the input perturbations from \\(\\mathbf{x}\\) to \\(\\mathbf{x}'\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.437, + 0.483, + 0.467 + ], + "angle": 0, + "content": "Perception as a new AT objective in addition to the conventional one to realize this idea." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.473, + 0.484, + 0.551 + ], + "angle": 0, + "content": "Definition 1 (Robust Perception). Provided an AT task with target model \\(\\theta\\), let \\(\\mathbf{x}\\) and \\(\\mathbf{x}'\\) be any pairs of corresponding benign and adversarial samples with \\(\\Delta = \\mathbf{x}' - \\mathbf{x}\\), for any hidden representation \\(h_{\\theta}(\\mathbf{x})\\) denoting the model perception, our additional AT objective can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.561, + 0.485, + 0.58 + ], + "angle": 0, + "content": "\\[\n\\forall \\alpha \\in [ 0, 1 ], \\| h _ {\\pmb {\\theta}} (\\mathbf {x} + \\alpha \\cdot \\Delta) - h _ {\\pmb {\\theta}} (\\mathbf {x}) \\| = \\alpha \\cdot \\| h _ {\\pmb {\\theta}} (\\mathbf {x} ^ {\\prime}) - h _ {\\pmb {\\theta}} (\\mathbf {x}) \\|.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.591, + 0.484, + 0.728 + ], + "angle": 0, + "content": "To facilitate the understanding of the proposed Robust Perception, in addition to the most straightforward perspective that it encourages the change in model perception along with the adversarial perturbations, the effectiveness of this new objective can also be demonstrated from other two perspectives, local linearity [12] and Lipschitz regularization [1, 11, 21, 33], both of which also help smooth the decision boundary and mitigate the current trade-off problem. Below, we further detail them theoretically." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.733, + 0.486, + 0.779 + ], + "angle": 0, + "content": "Theorem 1. Let \\( H \\) be the Hessian Matrix such that \\( H_{h_\\theta}(\\mathbf{x}) = \\nabla_{\\mathbf{x}}^2 h_\\theta(\\mathbf{x}) \\), then with the new optimization objective of Robust Perception, we have:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.789, + 0.383, + 0.808 + ], + "angle": 0, + "content": "\\[\n\\forall \\Delta , \\Delta^ {\\top} \\cdot H _ {h _ {\\theta}} (\\mathbf {x}) \\cdot \\Delta \\rightarrow 0.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.819, + 0.381, + 0.835 + ], + "angle": 0, + "content": "Proof. We defer the proof to Appendix B.1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "This means Robust Perception limits the second-order (and higher-order as well) nonlinear effects within the adversarial perturbation to the model perception, such that the perception would change mainly along with the linear" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.437, + 0.907, + 0.498 + ], + "angle": 0, + "content": "term of the perturbation, and thus be smoother. This also aligns with the local linearity hypothesis of adversarial samples [12], which assumes that linear approximations dominate the model behavior under adversarial perturbations." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.503, + 0.909, + 0.549 + ], + "angle": 0, + "content": "Theorem 2. Let \\(J\\) be the Jacobian Matrix such that \\(J_{h_{\\theta}}(\\mathbf{x}) = \\nabla_{\\mathbf{x}} h_{\\theta}(\\mathbf{x})\\), then with the new optimization objective of Robust Perception, we have:" + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.561, + 0.846, + 0.579 + ], + "angle": 0, + "content": "\\[\n\\forall \\alpha \\in [ 0, 1 ], J _ {h _ {\\boldsymbol {\\theta}}} (\\mathbf {x} + \\boldsymbol {\\alpha} \\cdot \\Delta) \\rightarrow J _ {h _ {\\boldsymbol {\\theta}}} (\\mathbf {x}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.59, + 0.907, + 0.651 + ], + "angle": 0, + "content": "then with \\(\\| \\cdot \\|_{\\mathrm{spec}}\\) denoting the Spectral Norm and \\(\\gamma\\) being any micro value, given \\(\\| J_{h_\\theta}(\\mathbf{x} + \\alpha \\cdot \\Delta) - J_{h_\\theta}(\\mathbf{x})\\|_{\\mathrm{spec}} \\leq \\gamma\\), the function \\(h_\\theta(\\mathbf{x})\\) can be referred to as \\(K\\)-Lipschitz with the Lipschitz constant \\(K\\) upper-bounded by:" + }, + { + "type": "equation", + "bbox": [ + 0.611, + 0.662, + 0.808, + 0.686 + ], + "angle": 0, + "content": "\\[\nK \\leq \\sup _ {\\mathbf {x}} \\| J _ {h _ {\\theta}} (\\mathbf {x}) \\| _ {\\operatorname {s p e c}} + \\gamma .\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.697, + 0.804, + 0.713 + ], + "angle": 0, + "content": "Proof. We defer the proof to Appendix B.2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.718, + 0.906, + 0.825 + ], + "angle": 0, + "content": "This theorem tells us that Robust Perception also regularizes the change \\(\\gamma\\) in Jacobian along with the adversarial perturbation, which can be interpreted as limiting the increase of the global Lipschitz constant by \\(\\gamma\\) under perturbation. Previous works have demonstrated that this is expected to help achieve smoother decision boundaries and improve the accuracy-robustness trade-off [1, 11, 21, 33]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.859, + 0.85 + ], + "angle": 0, + "content": "4.2. Robust Perception Adversarial Training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Based on the above considerations, in this section, we propose a novel AT method named Robust Perception Adversarial Training (RPAT). It utilizes a new additional" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4695" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.246 + ], + "angle": 0, + "content": "regularization term appended to the conventional AT objective to achieve the Robust Perception. Formally, given a \\(C\\)-class classification task \\((C \\geq 2)\\) and a dataset \\(\\mathcal{D} = \\{(\\mathbf{x}_i, y_i)\\}_{i=1,\\dots,n}\\) with benign sample \\(\\mathbf{x}_i \\in \\mathbb{R}^d\\) and supervised label \\(y_i \\in \\{1,\\dots,C\\}\\), let \\(\\hat{\\mathbf{x}}_i'\\) denotes the adversarial sample corresponding to \\(\\mathbf{x}_i\\), which is generated as \\(\\hat{\\mathbf{x}}_i' = \\arg \\max_{\\|\\mathbf{x}_i' - \\mathbf{x}_i\\|_p \\leq \\epsilon} \\mathbb{1}(\\arg \\max_{k=1,\\dots,C} h_\\theta(\\mathbf{x}_i'))^{(k)} \\neq y_i)\\), then based on the Definition 1 in Section 4.1, the adversarial risk to be minimized for the proposed RPAT w.r.t. \\(\\mathbf{x}_i\\) can be formulated with the 0-1 loss [58] as:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.255, + 0.482, + 0.307 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {R} (\\boldsymbol {\\theta}, \\mathbf {x} _ {i}) := \\mathbb {1} \\left(f _ {\\boldsymbol {\\theta}} \\left(\\hat {\\mathbf {x}} _ {i} ^ {\\prime}\\right) \\neq y _ {i}\\right) \\\\ + \\mathbb {1} \\left(\\frac {h _ {\\boldsymbol {\\theta}} \\left(\\tilde {\\mathbf {x}} _ {i}\\right) - h _ {\\boldsymbol {\\theta}} \\left(\\mathbf {x} _ {i}\\right)}{\\alpha} \\neq \\frac {h _ {\\boldsymbol {\\theta}} \\left(\\hat {\\mathbf {x}} _ {i} ^ {\\prime}\\right) - h _ {\\boldsymbol {\\theta}} \\left(\\tilde {\\mathbf {x}} _ {i}\\right)}{1 - \\alpha}\\right), \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.318, + 0.303, + 0.334 + ], + "angle": 0, + "content": "where \\(\\tilde{\\mathbf{x}}_i = \\mathbf{x}_i + \\alpha \\cdot (\\hat{\\mathbf{x}}_i' - \\mathbf{x}_i)\\)" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.334, + 0.483, + 0.499 + ], + "angle": 0, + "content": "While the proposed Robust Perception aims to improve learning efficacy for the failure cases, the above risk does not explicitly distinguish success and failure cases. This is because Robust Perception is designed to mitigate the gap of the failure cases in the effectiveness of learning compared with the success ones, rather than require any new characteristic beyond the success cases. In other words, the success cases are expected to also satisfy Robust Perception at the beginning. Therefore, considering the computational efficiency, it is not necessary to specifically exclude the success cases for the Robust Perception term." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.501, + 0.484, + 0.756 + ], + "angle": 0, + "content": "Finally, by introducing practical surrogate losses to optimize the adversarial risk in Equation (2), we end up with our new RPAT method. To be specific, for the supervised learning term for adversarial samples, we adopt cross-entropy (CE) just as the conventional AT by default. Then, for the Robust Perception term, as it is formulated in a format of similarity regularization, various criterion that measures the similarity distance between the two input items are feasible options, such as mean squared error (MSE), Kullback-Leibler divergence (KL), Jensen-Shannon divergence (JS), and Cosine similarity. We simply adopt MSE loss for the basic version of RPAT here, and would also demonstrate the effectiveness of KL loss later in Section 5.3. With the logits \\(\\mathbf{z}(\\mathbf{x}_i,\\boldsymbol {\\theta})\\) utilized as the hidden representation \\(h_\\theta (\\cdot)\\) to denote the model perception, and \\(\\mathbf{p}(\\mathbf{x}_i,\\boldsymbol {\\theta})\\) being the softmax output as given in Equation (1), the overall objective of RPAT can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.767, + 0.483, + 0.83 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} ^ {R P A T} (\\boldsymbol {\\theta}, \\mathcal {D}, \\lambda , \\alpha) := \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left(\\mathcal {L} ^ {C E} \\left(\\mathbf {p} \\left(\\hat {\\mathbf {x}} _ {i} ^ {\\prime}, \\boldsymbol {\\theta}\\right), y _ {i}\\right) \\right. \\tag {3} \\\\ + \\lambda \\cdot \\mathcal {L} ^ {M S E} \\left( \\right.\\frac {\\mathbf {z} (\\tilde {\\mathbf {x}} _ {i} , \\boldsymbol {\\theta}) - \\mathbf {z} (\\mathbf {x} _ {i} , \\boldsymbol {\\theta})}{\\alpha} \\left| \\right.\\left| \\right. \\frac {\\mathbf {z} (\\hat {\\mathbf {x}} _ {i} ^ {\\prime} , \\boldsymbol {\\theta}) - \\mathbf {z} (\\tilde {\\mathbf {x}} _ {i} , \\boldsymbol {\\theta})}{1 - \\alpha}\\left. \\right)\\left. \\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a weight hyper-parameter to balance the conventional AT objective and our new Robust Perception objective. An intuitive illustration of the proposed RPAT method is provided in Figure 3." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.646, + 0.108 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.115, + 0.907, + 0.373 + ], + "angle": 0, + "content": "Following previous AT works, the effectiveness of the proposed RPAT is evaluated on CIFAR-10, CIFAR-100 [22] and Tiny-ImageNet [23] datasets. The main experiments can be divided into two parts. First, based on Tack et al. [40], we implement RPAT upon three common AT baselines, namely PGD-AT, TRADES, and MART, as well an advanced method, Consistency-AT, with ResNet-18 [15] architecture. This is to show the general effectiveness of the proposed Robust Perception in improving the current trade-off. Second, by integrating RPAT into one of the current SOTA methods, ReBAT [47], with PreActResNet-18 [16] and WideResNet-34-10 [56] architectures, we achieve new SOTA performance beyond 12 SOTA methods on the accuracy-robustness trade-off problem. This is to demonstrate the advancement achieved by the proposed RPAT in the AT area. The detailed experimental setup is provided in Appendix A." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.384, + 0.714, + 0.399 + ], + "angle": 0, + "content": "5.1. Evaluation Measures" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.407, + 0.907, + 0.603 + ], + "angle": 0, + "content": "There are five measures in total adopted to evaluate the experimental methods from three different perspectives. Firstly, we report clean accuracy on benign test data, which reflects the ability of the robust models to maintain performance in benign cases. Secondly, we evaluate adversarial robustness under two adversarial attack methods, namely PGD-20 [29] and Auto-Attack (AA) [7], with default settings aligning with Appendix A and random starts. Finally, to test the performance w.r.t. the current accuracy-robustness trade-off problem in AT, on the one hand, we report the mean of the clean accuracy and the robust score under Auto-Attack acquired above, and on the other hand, we introduce the Natural-Robustness Ratio (NRR) [13]:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.613, + 0.897, + 0.646 + ], + "angle": 0, + "content": "\\[\nN R R = \\frac {2 \\times C l e a n A c c u r a c y \\times A d v e r s a r i a l R o b u s t n e s s}{C l e a n A c c u r a c y + A d v e r s a r i a l R o b u s t n e s s},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.656, + 0.907, + 0.779 + ], + "angle": 0, + "content": "which emphasizes how well the robust model strikes a balance between clean accuracy and robustness. A lower NRR value suggests that the model might prioritize one side over the other. All the results reported are averages of three runs, with the \"best\" performance of each run recorded on the best checkpoint achieving the highest PGD-20 accuracy, which ensures that the training process is isolated from AA, the test-time adversary to evaluate the robustness." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.788, + 0.831, + 0.804 + ], + "angle": 0, + "content": "5.2. Effectiveness in Improving Baselines" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.906, + 0.902 + ], + "angle": 0, + "content": "The proposed RPAT can be easily implemented with existing AT methods by simply appending the Robust Perception term. In this section, we illustrate the improvements that RPAT can bring to our four experimental benchmarks, namely PGD-AT, TRADES, MART, and Consistency-AT, on CIFAR-10, CIFAR-100, and Tiny-ImageNet with" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4696" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.097, + 0.09, + 0.9, + 0.104 + ], + "angle": 0, + "content": "Table 1. Improvements that the proposed RPAT achieved upon four benchmarks. The results are acquired with ResNet-18 and \\( {\\ell }_{\\infty } \\) norm." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.115, + 0.904, + 0.317 + ], + "angle": 0, + "content": "
DatasetMethodBenchmarks+RPAT (Ours)
CleanPGD-20AAMeanNRRCleanPGD-20AAMeanNRR
bestbestbestbestfinalbestfinalbestfinal
CIFAR-10PGD-AT82.9250.6146.7464.83059.78283.2082.9951.2950.9748.0047.6065.60060.878
TRADES79.6751.9147.6263.64559.61080.0280.0251.2151.2147.4747.4563.74559.590
MART77.9353.0046.7062.31558.40275.4475.9552.8852.8747.0347.3061.23557.940
Consistency-AT83.4251.9647.7265.57060.71184.1284.0452.3352.1448.9849.0666.55061.911
CIFAR-100PGD-AT56.5628.8025.0240.79034.69358.2258.5829.1628.5724.8824.2341.55034.862
TRADES55.3929.3624.5139.95033.98357.5057.2529.4229.0025.0524.4541.27534.897
MART49.8330.3825.0037.41533.29550.7250.7230.3330.3325.3425.0638.03033.796
Consistency-AT58.5329.2825.3941.96035.41760.3360.5129.9729.6426.3125.9743.32036.641
Tiny-ImageNetPGD-AT46.3221.5817.0731.69524.94747.6848.3222.1322.0017.7717.9132.72525.891
TRADES46.7521.5216.6031.67524.50048.7748.8521.9821.9316.9216.7132.84525.124
MART39.7022.9017.1828.44023.98241.7641.6623.4023.2917.7917.7829.77524.951
Consistency-AT46.5421.8917.6032.07025.54149.7449.8523.1622.8418.8418.7434.29027.329
" + }, + { + "type": "table_caption", + "bbox": [ + 0.099, + 0.33, + 0.898, + 0.344 + ], + "angle": 0, + "content": "Table 2. Improvements that the proposed RPAT achieved upon four benchmarks. The results are acquired with ResNet-18 and \\( {\\ell }_{2} \\) norm." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.355, + 0.903, + 0.558 + ], + "angle": 0, + "content": "
DatasetMethodBenchmarks+RPAT (Ours)
CleanPGD-20AAMeanNRRCleanPGD-20AAMeanNRR
bestbestbestbestfinalbestfinalbestfinal
CIFAR-10PGD-AT87.7667.9266.3677.06075.57488.2088.4568.5668.5267.6367.6277.91576.557
TRADES83.9968.6065.9374.96073.87285.1784.8068.9568.9067.6767.3776.42075.418
MART84.0968.3266.2875.18574.13084.6584.6568.6068.6066.6966.7075.67074.604
Consistency-AT88.7669.3567.4678.11076.65889.3889.5270.4170.3669.4469.5579.41078.158
CIFAR-100PGD-AT65.0041.5439.2752.13548.96065.1465.4441.6141.2839.2339.2052.18548.969
TRADES61.2543.0540.1550.7048.50562.4162.6443.5643.1040.3640.1251.38549.020
MART60.0843.4239.8549.96547.9260.5760.5743.9243.9040.3640.3650.46548.442
Consistency-AT65.1442.2839.9252.53049.50365.5466.8142.8142.5140.2640.6152.90049.880
Tiny-ImageNetPGD-AT59.3643.1240.9050.13048.43160.2860.2844.0644.0641.8941.8751.08549.430
TRADES57.6144.7042.3349.97048.80258.4658.5044.8744.6542.4242.4450.44049.165
MART56.4744.8442.2249.34548.31657.1756.9545.0244.6742.5142.1649.84048.762
Consistency-AT61.4944.9242.5952.04050.32462.1962.6445.2844.8743.1742.8552.68050.963
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.482, + 0.839 + ], + "angle": 0, + "content": "ResNet-18, demonstrating the wide effectiveness of our method. The results under \\(\\ell_{\\infty}\\) and \\(\\ell_{2}\\) threat models are shown in Table 1 and Table 2, respectively, where the best clean accuracy, AA robustness, as well mean and NRR for their trade-off on each of the datasets are marked in bold, while individual scores not improved are marked with underline. Note that PGD-20 is the indicator for selecting the \"best\" checkpoint as explained in Section 5.1, and the \"final\" scores are additionally provided to show that RPAT does not suffer from serious robust-overfitting, which is also evidence for appropriate algorithm design. Thus, these fields are not involved in the comparison. It can be found that RPAT achieves significant improvement for most scores among the tables, where Consistency-AT+RPAT is basically the most competitive for all the experimental groups, which demonstrates the general effectiveness of our method in mitigating the current trade-off problem." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.373, + 0.864 + ], + "angle": 0, + "content": "5.3. Advancement beyond 12 SOTAs" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To demonstrate the advancement that the proposed RPAT contributes, we further integrate RPAT into ReBAT [47]," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.582, + 0.907, + 0.901 + ], + "angle": 0, + "content": "one of the current SOTA AT methods on the accuracy-robustness trade-off problem. It implements WA [18], averaging multiple points along the SGD trajectory for better generalization, through the exponential moving average (EMA) [37] strategy with a decay rate of 0.999. Thus, we also involve WA itself (i.e., with PGD-AT), as well as \\(\\mathrm{KD + SWA}\\) [5] which shares a similar idea, for comparison. Other differences in ReBAT include \\(KL\\) loss and a reduced learning rate decay factor from 0.1 to 0.5. We refer to our method upon ReBAT as \\(\\mathrm{RPAT}^{++}\\), and compare it with 12 previous SOTAs introduced in Section 2.2, respectively with PreActResNet-18 and WideResNet-34-10 as shown in Table 3 and Table 4. The clean and AA scores of the SOTAs are from the best checkpoint among their original papers, Kanai et al. [19], Wang et al. [47], and our reproduction if any (so due to the lack of previous record, ADR and CURE are excluded for PreActResNet-18). Beyond the basic CIFAR-10 results under \\(\\ell_{\\infty}\\) norm, Table 3 involves CIFAR-100 and \\(\\ell_2\\) norm, and Table 4 further consider the Cutmix [55] data augmentation strategy as marked by “*”, which ensures the comprehensiveness of our comparison." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4697" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.09, + 0.584, + 0.119 + ], + "angle": 0, + "content": "Table 3. Comparison of the proposed \\(\\mathrm{RPAT^{+ + }}\\) with current SOTAs on accuracy-robustness trade-off with PreActResNet-18. The best results are marked in bold." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.128, + 0.596, + 0.332 + ], + "angle": 0, + "content": "
NormMethodCIFAR-10CIFAR-100
CleanAAMeanNRRCleanAAMeanNRR
\\( \\ell_{\\infty} \\)WAUAI'1883.5049.8966.69562.46157.2625.8341.54535.601
MMAICLR'2085.5037.2061.35051.84460.6018.4039.50028.229
AWPNeurIPS'2081.1150.0965.60061.93354.1025.1639.63034.347
GAIRATICLR'2178.7037.7058.20050.97952.0019.8035.90028.680
KD+SWAICLR'2184.0649.8266.94062.56257.1725.6641.41535.422
EWATICML'2182.8048.2065.50060.93154.2023.5238.86032.805
MAILNeurIPS'2179.5039.6059.55052.86746.5016.7031.60024.574
TEICLR'2282.0450.1266.08062.22556.4125.8441.12535.444
SOVRICML'2381.9049.4065.65061.62852.1024.3038.20033.142
ReBATNeurIPS'2382.0950.7266.40562.70056.1327.6041.86537.004
RPAT++Ours82.6351.0066.81563.07256.8427.6842.26037.230
\\( \\ell_{2} \\)ReBATNeurIPS'2388.7971.0079.89578.90565.5842.6754.12551.701
RPAT++Ours89.0671.2680.16079.17265.6342.8554.24051.848
" + }, + { + "type": "table_caption", + "bbox": [ + 0.604, + 0.09, + 0.903, + 0.118 + ], + "angle": 0, + "content": "Table 4. Comparison of \\( {\\mathrm{{RPAT}}}^{+ + } \\) with SOTAs on CIFAR-10 with WideResNet-34-10 and \\( {\\ell }_{\\infty } \\) norm." + }, + { + "type": "table", + "bbox": [ + 0.609, + 0.129, + 0.899, + 0.331 + ], + "angle": 0, + "content": "
MethodCleanAAMeanNRR
WAUAI'1887.6652.6570.15565.787
MMAICLR'2087.8043.1065.45057.818
AWPNeurIPS'2085.6353.3269.47565.718
GAIRATICLR'2183.0041.8062.40055.599
KD+SWAICLR'2187.4553.5970.52066.456
EWATICML'2186.0051.6068.80064.500
MAILNeurIPS'2182.2043.3062.75056.721
TEICLR'2285.9752.8869.42565.482
SOVRICML'2385.0053.1069.05065.366
ReBATNeurIPS'2385.2554.7870.01566.700
ADRICLR'2484.6753.2568.96065.381
CUREICLR'2487.0552.1069.57565.186
\\(RPAT^{++}\\)Ours86.7654.9770.86567.300
ReBAT*NeurIPS'2386.6655.6471.15067.769
\\(RPAT^{++*}\\)Ours87.5755.7971.68068.158
" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.354, + 0.277, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.106, + 0.467, + 0.266, + 0.48 + ], + "angle": 0, + "content": "(a) Impact of loss function." + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.354, + 0.479, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.28, + 0.467, + 0.474, + 0.481 + ], + "angle": 0, + "content": "(b) Impact of hyper-parameter \\(\\alpha\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.492, + 0.484, + 0.548 + ], + "angle": 0, + "content": "Figure 4. Two ablation studies of the proposed RPAT on CIFAR-10 with ResNet-18 under \\(\\ell_{\\infty}\\) norm. The MSE loss in (a) and the points with \\(\\alpha = 0.5\\) in (b) correspond to the record of \"PGD-AT+RPAT\" on CIFAR-10 in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.577, + 0.484, + 0.639 + ], + "angle": 0, + "content": "The results show that the proposed \\(\\mathrm{RPAT^{+ + }}\\) achieves new SOTA performance for most of the measures beyond previous SOTAs on the accuracy-robustness trade-off problem, which strongly supports the contribution of this work." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.65, + 0.245, + 0.667 + ], + "angle": 0, + "content": "5.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.854 + ], + "angle": 0, + "content": "As the proposed RPAT can be implemented upon existing AT methods by simply appending the Robust Perception term into their original objectives, the comparison in performance between and after adding RPAT in Section 5.2 and Section 5.3 have achieved a comprehensive ablation study on the effectiveness of RPAT. Therefore, in this section, we mainly investigate the impact of two setting details, respectively the loss function adopted for the Robust Perception term (Figure 4 (a)) and the specific value for the hyperparameter \\(\\alpha\\) (Figure 4 (b)). Besides, we further discuss an optional trick suggested by ReBAT but excluded in the above experiments in Appendix 5.3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Specifically, just as we expected, different loss functions to calculate the similarity between model perceptions, including MSE, KL, JS and Cosine, would not significantly" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.356, + 0.907, + 0.583 + ], + "angle": 0, + "content": "impact the final performance of RPAT, which suggests that the effectiveness of our method is from the successful conceptual design of Robust Perception, instead of relying on specific implementation of the regularization term. Then, for \\(\\alpha\\), we try 0.2 and 0.8 besides the default value 0.5 on the one hand, and also introduce beta distribution \\(\\beta \\sim Beta(0.75, 0.75)\\) as suggested by Zhang et al. [57] on the other hand. Specifically, we try \\(\\alpha = \\min(\\beta, (1 - \\beta))\\) and \\(\\alpha = \\max(\\beta, (1 - \\beta))\\), namely indicated as \"Beta-\" and \"Beta+\" in (Figure 4 (b)), as they respectively approach adversarial and benign samples more. The results show that the default \\(\\alpha = 0.5\\) is an appropriate choice, which is reasonable because the midpoint is expected to be more representative for the ideal target that Robust Perception holds for all \\(\\alpha \\in [0, 1]\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.602, + 0.634, + 0.618 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.901 + ], + "angle": 0, + "content": "This work studies the current accuracy-robustness trade-off problem in AT. To the best of our knowledge, this is the first work that explicitly attributes the trade-off problem to the over-sufficient (rather than insufficient) learning of hard adversarial samples. We surprisingly reveal that such hard samples are already learned better than those successfully defended in model perception, which means the conventional AT objective is no longer effective to further optimize the model w.r.t. these hard samples. Instead, it forces the model to ignore the perturbations, in turn degrading their effectiveness in supporting the establishment of decision boundary to an appropriate location. To mitigate this, our new AT objective, Robust Perception, encourages smoother perception transition towards the decision boundary along with perturbations, based on which the proposed RPAT method achieves new SOTA performance. Still, exploring better methods to approximate this ideal objective would be a promising future direction." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4698" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.114, + 0.484, + 0.173 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China (No. 62471420), GuangDong Basic and Applied Basic Research Foundation (2025A15150122 96), and CCF-Tencent Rhino-Bird Open Research Fund." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.196, + 0.188, + 0.212 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.221, + 0.484, + 0.277 + ], + "angle": 0, + "content": "[1] Abulikemu Abuduweili and Changliu Liu. Estimating neural network robustness via lipschitz constant and architecture sensitivity. In Safe and Robust Robot Learning for Operation in the Real World Workshop @ CoRL, 2024. 2, 5, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.278, + 0.484, + 0.318 + ], + "angle": 0, + "content": "[2] Anish Athalye, Nicholas Carlini, and David Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In ICML, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.319, + 0.483, + 0.359 + ], + "angle": 0, + "content": "[3] Tao Bai, Jinqi Luo, Jun Zhao, Bihan Wen, and Qian Wang. Recent advances in adversarial training for adversarial robustness. In IJCAI, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.361, + 0.483, + 0.415 + ], + "angle": 0, + "content": "[4] Battista Biggio, Igino Corona, Davide Maiorca, Blaine Nelson, Nedim Šrndić, Pavel Laskov, Giorgio Giacinto, and Fabio Roli. Evasion attacks against machine learning at test time. In ECML-PKDD, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.416, + 0.482, + 0.457 + ], + "angle": 0, + "content": "[5] Tianlong Chen, Zhenyu Zhang, Sijia Liu, Shiyu Chang, and Zhangyang Wang. Robust overfitting may be mitigated by properly learned smoothening. In ICLR, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.458, + 0.482, + 0.499 + ], + "angle": 0, + "content": "[6] Minhao Cheng, Qi Lei, Pin-Yu Chen, Inderjit Dhillon, and Cho-Jui Hsieh. Cat: Customized adversarial training for improved robustness. In *IJCAI*, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.5, + 0.482, + 0.54 + ], + "angle": 0, + "content": "[7] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In ICML, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.541, + 0.482, + 0.595 + ], + "angle": 0, + "content": "[8] Gavin Weiguang Ding, Yash Sharma, Kry Yik Chau Lui, and Ruitong Huang. Mma training: Direct input space margin maximization through adversarial training. In ICLR, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.597, + 0.482, + 0.637 + ], + "angle": 0, + "content": "[9] Yinpeng Dong, Qi-An Fu, Xiao Yang, Tianyu Pang, Hang Su, Zihao Xiao, and Jun Zhu. Benchmarking adversarial robustness on image classification. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.638, + 0.482, + 0.679 + ], + "angle": 0, + "content": "[10] Yinpeng Dong, Ke Xu, Xiao Yang, Tianyu Pang, Zhijie Deng, Hang Su, and Jun Zhu. Exploring memorization in adversarial training. In ICLR, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.68, + 0.482, + 0.734 + ], + "angle": 0, + "content": "[11] Mahyar Fazlyab, Taha Entesari, Aniket Roy, and Rama Chellappa. Certified robustness via dynamic margin maximization and improved lipschitz regularization. In NeurIPS, 2023. 2, 5, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.735, + 0.482, + 0.775 + ], + "angle": 0, + "content": "[12] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In ICLR, 2015. 1, 2, 5, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.777, + 0.482, + 0.817 + ], + "angle": 0, + "content": "[13] Shruthi Gowda, Bahram Zonooz, and Elahe Arani. Conserve-update-revise to cure generalization and robustness trade-off in adversarial training. In ICLR, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.482, + 0.859 + ], + "angle": 0, + "content": "[14] Chuan Guo, Mayank Rana, Moustapha Cisse, and Laurens van der Maaten. Countering adversarial images using input transformations. In ICLR, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.221, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In ECCV, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[17] Xiaowei Huang, Daniel Kroening, Wenjie Ruan, James Sharp, Youcheng Sun, Emese Thamo, Min Wu, and Xinping Yi. A survey of safety and trustworthiness of deep neural networks: Verification, testing, adversarial attack and defence, and interpretability. Computer Science Review, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[18] Pavel Izmailov, Dmitrii Podoprikhin, Timur Garipov, Dmitry Vetrov, and Andrew Gordon Wilson. Averaging weights leads to wider optima and better generalization. In UAI, 2018. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[19] Sekitoshi Kanai, Shin'ya Yamaguchi, Masanori Yamada, Hiroshi Takahashi, Kentaro Ohno, and Yasutoshi Ida. One-vs-the-rest loss to focus on important samples in adversarial training. In ICML, 2023. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.319, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[20] Minseon Kim, Jihoon Tack, Jinwoo Shin, and Sung Ju Hwang. Entropy weighted adversarial training. In Adversarial Machine Learning Workshop @ ICML, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[21] Vishaal Krishnan, Al Makedah, Abed AlRahman, and Fabio Pasqualetti. Lipschitz bounds and provably robust training by laplacian smoothing. In NeurIPS, 2020. 2, 5, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.905, + 0.433 + ], + "angle": 0, + "content": "[22] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[23] Fei-Fei Li, Andrej Karpathy, and Justin Johnson. Tiny image dataset. 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[24] Zi Liang, Pinghui Wang, Ruofei Zhang, Nuo Xu, Shuo Zhang, Lifeng Xing, Haitao Bai, and Ziyang Zhou. Merge: Fast private text generation. In AAAI, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.505, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[25] Fangzhou Liao, Ming Liang, Yinpeng Dong, Tianyu Pang, Xiaolin Hu, and Jun Zhu. Defense against adversarial attacks using high-level representation guided denoiser. In CVPR, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.562, + 0.905, + 0.617 + ], + "angle": 0, + "content": "[26] Weilin Lin, Nanjun Zhou, Yanyun Wang, Jianze Li, Hui Xiong, and Li Liu. Backdoor: A comprehensive benchmark for backdoor learning in diffusion model. arXiv preprint arXiv:2502.11798, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.619, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[27] Feng Liu, Bo Han, Tongliang Liu, Chen Gong, Gang Niu, Mingyuan Zhou, Masashi Sugiyama, et al. Probabilistic margins for instance reweighting in adversarial training. In NeurIPS, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[28] Xin Liu, Yichen Yang, Kun He, and John E Hopcroft. Parameter interpolation adversarial training for robust image classification. IEEE Transactions on Information Forensics and Security (TIFS), 2025. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.732, + 0.905, + 0.785 + ], + "angle": 0, + "content": "[29] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In ICLR, 2018. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.905, + 0.828 + ], + "angle": 0, + "content": "[30] Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. In ICLR, 2018. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[31] Tianyu Pang, Xiao Yang, Yinpeng Dong, Hang Su, and Jun Zhu. Bag of tricks for adversarial training. In ICLR, 2021. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[32] Nicolas Papernot, Patrick McDaniel, Ian Goodfellow, Somesh Jha, Z Berkay Celik, and Ananthram Swami. Practi" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4699" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "cal black-box attacks against machine learning. In AsiaCCS, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.176 + ], + "angle": 0, + "content": "[33] Patricia Pauli, Anne Koch, Julian Berberich, Paul Kohler, and Frank Allgower. Training robust neural networks using lipschitz bounds. IEEE Control Systems Letters, 2021. 2, 5, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[34] Rahul Rade and Seyed-Mohsen Moosavi-Dezfooli. Reducing excessive margin to achieve a better accuracy vs. robustness trade-off. In ICLR, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.222, + 0.482, + 0.276 + ], + "angle": 0, + "content": "[35] Aditi Raghunathan, Sang Michael Xie, Fanny Yang, John Duchi, and Percy Liang. Adversarial training can hurt generalization. In Identifying and Understanding Deep Learning Phenomena Workshop @ ICML, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.331 + ], + "angle": 0, + "content": "[36] Aditi Raghunathan, Sang Michael Xie, Fanny Yang, John Duchi, and Percy Liang. Understanding and mitigating the tradeoff between robustness and accuracy. In ICML, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.376 + ], + "angle": 0, + "content": "[37] Sylvestre-Alvise Rebuffi, Sven Gowal, Dan Andrei Calian, Florian Stimberg, Olivia Wiles, and Timothy A Mann. Data augmentation can improve robustness. In NeurIPS, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[38] David Stutz, Matthias Hein, and Bernt Schiele. Confidence-calibrated adversarial training: Generalizing to unseen attacks. In ICML, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[39] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In ICLR, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[40] Jihoon Tack, Sihyun Yu, Jongheon Jeong, Minseon Kim, Sung Ju Hwang, and Jinwoo Shin. Consistency regularization for adversarial robustness. In AAAI, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.56 + ], + "angle": 0, + "content": "[41] Lue Tao, Lei Feng, Jinfeng Yi, Sheng-Jun Huang, and Songcan Chen. Better safe than sorry: Preventing delusive adversaries with adversarial training. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.603 + ], + "angle": 0, + "content": "[42] Florian Tramer, Nicholas Carlini, Wieland Brendel, and Aleksander Madry. On adaptive attacks to adversarial example defenses. In NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.605, + 0.482, + 0.646 + ], + "angle": 0, + "content": "[43] Dimitris Tsipras, Shibani Santurkar, Logan Engstrom, Alexander Turner, and Aleksander Madry. Robustness may be at odds with accuracy. In ICLR, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.648, + 0.482, + 0.701 + ], + "angle": 0, + "content": "[44] Haotao Wang, Tianlong Chen, Shupeng Gui, TingKuei Hu, Ji Liu, and Zhangyang Wang. Once-for-all adversarial training: In-situ tradeoff between robustness and accuracy for free. In NeurIPS, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[45] Yisen Wang, Difan Zou, Jinfeng Yi, James Bailey, Xingjun Ma, and Quanquan Gu. Improving adversarial robustness requires revisiting misclassified examples. In ICLR, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[46] Yanyun Wang, Dehui Du, Haibo Hu, Zi Liang, and Yuan-hao Liu. Tsfool: Crafting highly-imperceptible adversarial time series through multi-objective attack. arXiv preprint arXiv:2209.06388, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[47] Yifei Wang, Liangchen Li, Jiansheng Yang, Zhouchen Lin, and Yisen Wang. Balance, imbalance, and rebalance: Understanding robust overfitting from a minimax game perspective. In NeurIPS, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[48] Yanyun Wang, Li Liu, Zi Liang, Qingqing Ye, and Haibo Hu. New paradigm of adversarial training: Releasing accuracy" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "robustness trade-off via dummy class. arXiv preprint arXiv:2410.12671, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[49] Dongxian Wu, Shu-Tao Xia, and Yisen Wang. Adversarial weight perturbation helps robust generalization. In NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[50] Yu-Yu Wu, Hung-Jui Wang, and Shang-Tse Chen. Annealing self-distillation rectification improves adversarial training. In ICLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.208, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[51] Cihang Xie, Jianyu Wang, Zhishuai Zhang, Zhou Ren, and Alan Yuille. Mitigating adversarial effects through randomization. In ICLR, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.251, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[52] Weilin Xu, David Evans, and Yanjun Qi. Feature squeezing: Detecting adversarial examples in deep neural networks. In NDSS, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.294, + 0.905, + 0.32 + ], + "angle": 0, + "content": "[53] Shuo Yang and Chang Xu. One size does not fit all: Data-adaptive adversarial training. In ECCV, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.322, + 0.905, + 0.39 + ], + "angle": 0, + "content": "[54] Jia-Li Yin, Bin Chen, Wanqing Zhu, Bo-Hao Chen, and Ximeng Liu. Push stricter to decide better: A class-conditional feature adaptive framework for improving adversarial robustness. IEEE Transactions on Information Forensics and Security (TIFS), 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.393, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[55] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In ICCV, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.905, + 0.475 + ], + "angle": 0, + "content": "[56] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In BMVC, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.478, + 0.905, + 0.518 + ], + "angle": 0, + "content": "[57] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. Mixup: Beyond empirical risk minimization. In ICLR, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.521, + 0.905, + 0.575 + ], + "angle": 0, + "content": "[58] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In ICML, 2019, 2, 3, 6, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.578, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[59] Jingfeng Zhang, Jianing Zhu, Gang Niu, Bo Han, Masashi Sugiyama, and Mohan Kankanhalli. Geometry-aware instance-reweighted adversarial training. In ICLR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.634, + 0.905, + 0.689 + ], + "angle": 0, + "content": "[60] Ziyu Zhou, Gengyu Lyu, Yiming Huang, Zihao Wang, Ziyu Jia, and Zhen Yang. Sdformer: Transformer with spectral filter and dynamic attention for multivariate time series long-term forecasting. In IJCAI, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.689 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "4700" + } + ] +] \ No newline at end of file diff --git a/2025/Failure Cases Are Better Learned But Boundary Says Sorry_ Facilitating Smooth Perception Change for Accuracy-Robustness Trade-Off in Adversarial Training/9886c220-8af0-438e-8e12-4de3189d4e86_origin.pdf b/2025/Failure Cases Are Better Learned But Boundary Says Sorry_ Facilitating Smooth Perception Change for Accuracy-Robustness Trade-Off in Adversarial Training/9886c220-8af0-438e-8e12-4de3189d4e86_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..78870587edb0015303d75308ac1f87e89b3e3d9b --- /dev/null +++ b/2025/Failure Cases Are Better Learned But Boundary Says Sorry_ Facilitating Smooth Perception Change for Accuracy-Robustness Trade-Off in Adversarial Training/9886c220-8af0-438e-8e12-4de3189d4e86_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5d5eaaf120bd1f7bacc32318dc56c9b2aa1c1e24c776440f178f9603e84775b8 +size 1765707 diff --git a/2025/Failure Cases Are Better Learned But Boundary Says Sorry_ Facilitating Smooth Perception Change for Accuracy-Robustness Trade-Off in Adversarial Training/full.md b/2025/Failure Cases Are Better Learned But Boundary Says Sorry_ Facilitating Smooth Perception Change for Accuracy-Robustness Trade-Off in Adversarial Training/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9d947d84442cdc0b04c3d6b6cb29617f344ef9ad --- /dev/null +++ b/2025/Failure Cases Are Better Learned But Boundary Says Sorry_ Facilitating Smooth Perception Change for Accuracy-Robustness Trade-Off in Adversarial Training/full.md @@ -0,0 +1,286 @@ +# Failure Cases Are Better Learned But Boundary Says Sorry: Facilitating Smooth Perception Change for Accuracy-Robustness Trade-Off in Adversarial Training + +Yanyun Wang Li Liu* + +The Hong Kong University of Science and Technology (Guangzhou) + +# Abstract + +Adversarial Training (AT) is one of the most effective methods to train robust Deep Neural Networks (DNNs). However, AT creates an inherent trade-off between clean accuracy and adversarial robustness, which is commonly attributed to the more complicated decision boundary caused by the insufficient learning of hard adversarial samples. In this work, we reveal a counterintuitive fact for the first time: From the perspective of perception consistency, hard adversarial samples that can still attack the robust model after AT are already learned better than those successfully defended. Thus, different from previous views, we argue that it is rather the over-sufficient learning of hard adversarial samples that degrades the decision boundary and contributes to the trade-off problem. Specifically, the excessive pursuit of perception consistency would force the model to view the perturbations as noise and ignore the information within them, which should have been utilized to induce a smoother perception transition towards the decision boundary to support its establishment to an appropriate location. In response, we define a new AT objective named Robust Perception, encouraging the model perception to change smoothly with input perturbations, based on which we propose a novel Robust Perception Adversarial Training (RPAT) method, effectively mitigating the current accuracy-robustness trade-off. Experiments on CIFAR-10, CIFAR-100, and Tiny-ImageNet with ResNet-18, PreActResNet-18, and WideResNet-34-10 demonstrate the effectiveness of our method beyond four common baselines and 12 state-of-the-art (SOTA) works. The code is available at https://github.com/FlaAI/RPAT. + +# 1. Introduction + +Deep neural networks (DNNs) have demonstrated impressive performance across various real-world applications [24, 26, 60]. However, they are inherently vulnerable to adversarial attacks [4, 39]. Specifically, even impercep + +![](images/3af3a29bf8b909c59ece5b66e79e02c7ee88bde2635c70c29200b013605ee149.jpg) + +![](images/be6a139caa4fa6f60e810e855d5df80c8bbfbb95dd96fc195ab83bf5118e09d4.jpg) + +![](images/5a8f87284bdeaa37efaf2340baceebfc4f83998107bbda98f0f3e654e80ec287.jpg) +Figure 1. Success and failure cases of AT defense from the perspectives of prediction and perception, respectively. (a) and (b): AT works through building a robust ball with radius $d$ such that the model prediction remains unchanged within it. Then if the robust $d$ -ball fails to contain the $\epsilon$ -ball, there would still be failure cases beating the robust model under perturbation $\delta$ within the $\epsilon$ -ball. (c) and (d): In this work, we reveal that the smaller $d$ -ball in the failure cases is not a result of more vulnerable model perception as previous intuition. In fact, as demonstrated by the contour line, which shows the perturbation ball to the same degree of perception change, the failure cases already achieve less perception change under similar $\delta$ within the $\epsilon$ -ball. This implies that the issue of these failure cases is not failing to learn perception consistency following the current AT objective, but failing to support the establishment of decision boundary to an appropriate location. + +![](images/9c677d01bc3c8323c7274ecfa24a41d02b945ed5d31859ccbbddafa4d943183e.jpg) +(c) (d) Contour line of model perception + +tible perturbations to inputs can cause substantial changes in the outputs [12, 17]. As this problem raises significant security concerns [45], previously, several defense strategies, such as Defense Distillation [32], Feature Squeezing [52], Randomization [51], and Input Denoising [14, 25], have been proposed to improve the adversarial robustness of DNNs. However, most of these approaches have proven + +ineffective against more advanced adaptive attacks [2, 42], such as Auto-Attack [7]. Currently, Adversarial Training (AT) [12, 29] is regarded as one of the most effective methods for training robust DNNs [2, 9]. + +However, an accuracy-robustness trade-off problem have been widely identified for AT in recent studies [3, 28, 34-36, 43, 44, 48, 54, 58]. Compared with standard training, the improvement in adversarial robustness acquired by AT is usually at the expense of the reduction in clean accuracy, which is afraid to significantly degrade the user experience in clean context and hinder the application of AT in real-world practices. The current consensus is that AT aggravates the accuracy-robustness trade-off by creating overly complex decision boundaries, primarily because models struggle to sufficiently learn from hard adversarial samples during training [6, 8, 10, 19, 27, 28, 34, 48, 53, 54]. Specifically, as the conventional AT objective encourages the model to perceive any adversarial inputs within the $\epsilon$ -ball similar to the corresponding benign one to produce consistent predictions for them, previous works empirically assume that the wrong predictions of the hard adversarial samples are attributed to the insufficiently optimized perception of them under the current objective. Therefore, although defining "hard sample" from different perspectives (e.g., classification result and distance to decision boundary), previous works agree to further optimize the perception consistency on them, ending up with various learning strategies specific for them, such as data reweighting [27, 59], adaptive perturbation radius [8, 34], and non-one-hot supervised signal [10, 19]. + +However, in this work, we reveal that for the robust model trained by convention AT, the statistical difference in model perception from benign to hard adversarial samples causing failure defense is already smaller than to others (i.e., those successfully defended). It implies that the model has even more sufficiently learned them under the current AT objective, and should have successfully defended them if the above assumption holds. Thus, as demonstrated in Figure 1, we suggest that the key factor here is no longer the optimization of model perception but the location of the decision boundary. In that case, keeping on pursuing over-sufficient small perception change would not only help little in further reducing the prediction error, but also force the model to view the perturbations as noise and ignore the information within them, which impairs their ability in claiming robust balls with reasonable gradients to push the decision boundary to an appropriate location. As a consequence, such sample-level imbalance in boundary establishment would naturally increase its complexity, harming generalization ability and degrading the trade-off between accuracy and robustness. + +To mitigate the over-sufficient learning of model perception on the failure cases and retain effective gradient sig + +nals w.r.t. them to facilitate the establishment of the decision boundary to an appropriate location, firstly, we define a new AT objective named Robust Perception in addition to the conventional one, encouraging smoother perception changes towards the decision boundary along with the input perturbations. Rigorous theoretical derivations also demonstrate its effectiveness in smoothening the decision boundary and releasing the current trade-off problem from the perspectives of local linearity [12] and Lipschitz regularization [1, 11, 21, 33]. Secondly, based on our new objective, we propose a novel Robust Perception Adversarial Training (RPAT) method, concurrently improving clean accuracy and adversarial robustness in AT. Extensive experiments on CIFAR-10, CIFAR-100, and Tiny-ImageNet with ResNet-18, PreActResNet-18, and WideResNet-34-10 demonstrate the advancement of RPAT beyond four common benchmarks and 12 state-of-the-art (SOTA) works. + +# 2. Background and Related Works + +# 2.1. Robustness and Adversarial Training + +Adversarial robustness reflects model vulnerability to malicious perturbations, which is commonly quantified by test accuracy under adversarial attacks [3, 39, 46]. Representatively, the Fast Gradient Sign Method (FGSM) [12] perturbs input samples along the gradient of the loss function, whereas the Projected Gradient Descent (PGD) attack [29] applies FGSM iteratively and projects the perturbation onto the $\epsilon$ -ball. Auto-Attack [7] introduces and combines several parameter-free attacks to mitigate practical issues such as hyper-parameter tuning and gradient masking. Due to their effectiveness, PGD and Auto-Attack are widely used to assess adversarial robustness, and so does this work. + +Previously, although many defense strategies including Defense Distillation [32], Feature Squeezing [52], Input Denoising [14, 25] and Randomization [51] have been developed, most rely on obfuscated gradients [2] and can be evaded by advanced adaptive attacks [42]. Nowadays, Adversarial Training [12, 29] is recognized as the most effective method to enhance adversarial robustness [2, 9, 41]. Unlike clean training on benign samples, AT directly learns on adversarily augmented data. Its objective can be formulated as the following min-max problem [29]: + +$$ +\min _ {\boldsymbol {\theta}} \frac {1}{n} \sum_ {i = 1} ^ {n} \max _ {\| \mathbf {x} _ {i} ^ {\prime} - \mathbf {x} _ {i} \| _ {P} \leq \epsilon} \mathcal {L} \left(f _ {\boldsymbol {\theta}} \left(\mathbf {x} _ {i} ^ {\prime}\right), y _ {i}\right), \tag {1} +$$ + +where $f_{\pmb{\theta}}(\mathbf{x}_i) = \arg \max_{c=1,\dots,C} \mathbf{p}_c(\mathbf{x}_i, \pmb{\theta})$ with $\mathbf{p}_c$ denoting the softmax probability of class $c$ within the prediction of $\mathbf{x}_i$ , $n$ denotes the number of training samples, and $\mathbf{x}_i'$ is generated by adding the strongest $L_p$ -norm perturbation within the $\epsilon$ -ball to the original sample $\mathbf{x}_i$ . This objective aims at prediction consistency within the $\epsilon$ -ball, encouraging the model to also produce similar perception there. + +PGD-AT [29] is the first attempt at solving this min-max optimization iteratively by approximating the inner maximization via PGD to craft adversarial samples and then performing stochastic gradient descent (SGD) on these samples for the outer minimization. TRADES [58] first considers balancing clean accuracy and adversarial robustness in AT, which decomposes robust error into natural error and boundary error, with the latter occurring only if specific data points are sufficiently close to the decision boundary. MART [45] further refines AT by distinguishing misclassified from correctly classified samples, which suggests appending boundary error only for misclassified samples to the basic AT objective in PGD-AT. Consistency-AT [40] encourages the prediction between adversarial samples derived from different augmentations of the same benign sample to be similar, which explicitly enhances the prediction consistency further. + +# 2.2. Accuracy-Robustness Trade-Off in AT + +An important problem identified for AT is the trade-off between clean accuracy and adversarial robustness [3, 28, 34-36, 43, 44, 54, 58], which means the robustness gained by AT often comes at the cost of reduced model accuracy compared with clean training. Consequently, AT can degrade user experience under benign conditions, which may seriously delay its real-world deployment. A widely acknowledged cause of this problem is that AT tends to learn more complicated decision boundaries than clean training [6, 10, 28, 34, 53, 54], thereby harming the ability of the robust model to generalize to unseen data. A number of previous works attribute this to the insufficient learning of hard adversarial samples [8, 10, 19, 27, 34, 59]. Though defining "hard adversarial samples" from different perspectives, such as the classification result and the distance to decision boundary, they basically reached a consensus that the model perception w.r.t. hard adversarial samples is not well-optimized, and as a consequence, fails to produce the consistent final prediction. In response, these works propose various solutions to help the model learn better prediction consistency for those hard adversarial samples. + +For instance, GAIRAT [59] proposes that hard samples nearer to the decision boundary are more critical but less robust and should be assigned a larger weight. MAIL [27] further proposes three types of probabilistic margins to quantify the closeness for better reweighting adversarial data accordingly. In contrast, MMA [8] suggests learning hard adversarial samples with reduced perturbations, which employs an adaptive $\epsilon$ for perturbations to maximize the margin between data and the decision boundary. EWAT [20] weighs the loss for each adversarial sample proportionally to the entropy of its prediction distribution during AT to focus on those with more uncertain labels. TE [10] reveals that the model may attempt to memorize certain hard ad + +versarial samples during AT, as it is difficult to assign high-confidence one-hot labels for them [6, 38]. SOVR [19] suggests increasing the margins of logits for certain hard samples by switching to a one-vs-the-rest loss. + +Additionally, several works also study the trade-off problem from other perspectives. AWP [49] proposes a double perturbation mechanism that flattens the loss landscape by weight perturbation to improve robust generalization. KD+SWA [5] injects more smoothening during AT by leveraging knowledge distillation to smooth the logits and performing stochastic weight averaging [18] to smooth the weights. ADR [50] employs soft labels as a guidance mechanism, accurately reflecting the distribution shift under attack during AT. CURE [13] selectively updates specific layers while preserving others to enhance the learning capacity of the model. ReBAT [47] views AT as a dynamic min-max game between the model trainer and the attacker, rebalancing the two players by either constraining the trainer's capacity or intensifying the attack. + +# 3. Motivation: AT Over-Learns Failure Cases + +It is basically common sense for DNN-based classification that the model learns to perceive input better and is thus able to predict it more precisely (especially when the model perception is indicated by its logits). Therefore, it seems reasonable at first glance that the previous works assume that the wrong prediction of the hard adversarial samples is due to the insufficient model perception of them, and propose various solutions upon this assumption in response. + +However, to the best of our knowledge, there is no previous discussion on a simple but crucial question: What if the assumption does not hold? If the model perception on the hard adversarial samples is already relatively close to the corresponding benign ones (e.g., at a similar or even better level than non-hard ones), then it would be no longer effective enough to further optimize the model on such samples through an objective simply pursuing prediction consistency as the conventional AT. Also, in turn, this can no longer effectively benefit the further reduction of the prediction error. Surprisingly, our following proof-of-concept experiments reveal that this is exactly the case. + +The proof-of-concept experiments are conducted on CIFAR-10 dataset with ResNet-18 architecture, with the results illustrated in Figure 2. As suggested by the x-axis, we involve three cases with different pairs of training and evaluation methods, which are respectively 1) trained on clean data and evaluated with random perturbations, 2) trained and evaluated with randomly perturbed data, and 3) trained with PGD-AT and evaluated with PGD-20 adversary. The budget of the random perturbations $\epsilon = 8 / 255$ is assigned the same as PGD, following our experimental setup detailed in Appendix A, so do other training and evaluation settings. Given that the conventional AT objective in Equation (1) + +encourages the model to produce a similar perception for benign and adversarial samples, we utilize the similarity in model perception between the corresponding benign and perturbed samples to estimate the learning sufficiency of the perturbed samples in AT. Specifically, the model perception is represented by the logits, and the similarity is measured through the mean squared error (MSE) score. We calculate the score respectively for success and failure cases of defense, which are divided according to the correctness of the final classification results of the perturbed samples. In addition, the model accuracy on clean and perturbed data is also recorded and illustrated together. + +![](images/4c76dec9f1f0d9dc7d8a43f98f770561211e18dd756e5efa172bbaae57571c0b.jpg) +Figure 2. Proof-of-concept experiments for the different learning sufficiency in model perception for success and failure cases on CIFAR-10 with ResNet-18. In the x-axis, the "Random" for "Training Algorithm" means the training data is randomly perturbed within the budget $\epsilon$ , while that for "Perturbation Method" indicates the attack used for evaluation on test data is random perturbation. The bars demonstrate the MSE score on logits between each pair of benign and perturbed samples, with the lines additionally showing the model accuracy on clean and perturbed data. Our results reveal a surprising phenomenon that, different from the clean training (with or without random noise), where failure cases are worse learned as we intuitively expected, AT already learns a better model perception for failure cases. + +Surprisingly, we experimentally found that the failure cases already learn better in model perception than the success cases. First, for the clean trained model, the perception on failure cases is significantly worse than on the success ones, just as we expected. Although the difference between success and failure cases has decreased for the randomly trained model, the failure cases remain worse, which suggests that non-adversarial perturbations are influential but not enough to change the trend. However, unlike the previous intuition, the model trained by PGD-AT can already perceive failure cases better than success ones. Our finding updates the current knowledge that, unlike those in clean training, the failure cases in AT do not fail simply because they are not sufficiently learned. On the contrary, + +the model even learns them harder to try to achieve the prediction and perception consistency. The smaller perception change under similar perturbations is direct evidence for this point of view. + +Then, as a straightforward deduction, now that we have success cases with worse perception but failure cases with better perception, there should be a statistical difference between them in the maximum perception change they can tolerate while still maintaining the consistent prediction, just as demonstrated in Figure 1. In other words, the decision boundary is established relatively closer to the failure cases, which is instead the issue we should focus on, because such sample-level imbalance in boundary establishment can naturally complicate it and thus degrade the accuracy-robustness trade-off as mentioned in Section 2.2. In that case, the over-consistent model perception between benign and adversarial samples in the failure cases rather harms the final robustness, because this is expected to force the model to view the perturbations as noise and ignore the information within them, which should have been utilized to induce a smoother perception transition towards the decision boundary to support its establishment, thus impairing the ability of such cases to claim a reasonable robust distance during AT compared with success cases. Therefore, in this work, we argue that it is not insufficient but oversufficient learning of failure cases that contributes to the more complicated decision boundary and finally results in the current trade-off problem. + +# 4. Proposed Method + +# 4.1. Robust Perception: A New AT Objective + +In Section 3, the too-small perception change under perturbations caused by the over-sufficient learning of failure cases has been identified as an important factor contributing to the current trade-off problem, which in turn questions the effectiveness of the conventional AT objective on such failure cases. + +Formally, under Equation (1) which implicitly optimize the perception distance $\| h_{\theta}(\mathbf{x}^{\prime}) - h_{\theta}(\mathbf{x})\| \to 0$ between corresponding benign sample $\mathbf{x}$ and adversarial sample $\mathbf{x}^{\prime}$ , the gradient of the conventional AT loss w.r.t. the perturbation, $\nabla_{\| \mathbf{x}^{\prime} - \mathbf{x}\|}\mathcal{L}$ , would also approach 0. In that case, the model cannot further optimize the decision boundary using the perturbation information. Thus, a straightforward idea to mitigate this is to encourage smoother perception change towards the decision boundary along with the input perturbation, such that $\| h_{\theta}(\mathbf{x}^{\prime}) - h_{\theta}(\mathbf{x})\| \propto \| \mathbf{x}^{\prime} - \mathbf{x}\|$ , as long as the final prediction remains unchanged. This is expected to retain an effective gradient signal in the direction of the perturbation, which enables the model to utilize this information to guide the decision boundary to a more reasonable location. Accordingly, in this section, we propose Robust + +![](images/e7a9880cec0e57341631f7022e1a3db81a4ef2c5b4bd8502106ff36fd843a4cf.jpg) +Figure 3. Illustration of the proposed RPAT method. In addition to the conventional AT objective, which learns adversarial samples with the original supervised labels, RPAT introduces a novel regularization term to achieve the newly defined Robust Perception objective. Given benign sample $\mathbf{x}$ and adversarial samples $\mathbf{x}' = \mathbf{x} + \Delta$ , for any interpolation samples $\mathbf{x} + \lambda \cdot \Delta$ with $\lambda \in [0,1]$ , it encourages the ratio of the perception changes, $z_{\theta}(\mathbf{x} + \lambda \cdot \Delta) - z_{\theta}(\mathbf{x})$ and $z_{\theta}(\mathbf{x}') - z_{\theta}(\mathbf{x} + \lambda \cdot \Delta)$ , respectively from $\mathbf{x}$ to $\mathbf{x} + \lambda \cdot \Delta$ and from $\mathbf{x} + \lambda \cdot \Delta$ to $\mathbf{x}'$ , to approach the ratio indicated by $\lambda$ , such that the model perception would change smoothly along with the input perturbations from $\mathbf{x}$ to $\mathbf{x}'$ . + +Perception as a new AT objective in addition to the conventional one to realize this idea. + +Definition 1 (Robust Perception). Provided an AT task with target model $\theta$ , let $\mathbf{x}$ and $\mathbf{x}'$ be any pairs of corresponding benign and adversarial samples with $\Delta = \mathbf{x}' - \mathbf{x}$ , for any hidden representation $h_{\theta}(\mathbf{x})$ denoting the model perception, our additional AT objective can be formulated as: + +$$ +\forall \alpha \in [ 0, 1 ], \| h _ {\pmb {\theta}} (\mathbf {x} + \alpha \cdot \Delta) - h _ {\pmb {\theta}} (\mathbf {x}) \| = \alpha \cdot \| h _ {\pmb {\theta}} (\mathbf {x} ^ {\prime}) - h _ {\pmb {\theta}} (\mathbf {x}) \|. +$$ + +To facilitate the understanding of the proposed Robust Perception, in addition to the most straightforward perspective that it encourages the change in model perception along with the adversarial perturbations, the effectiveness of this new objective can also be demonstrated from other two perspectives, local linearity [12] and Lipschitz regularization [1, 11, 21, 33], both of which also help smooth the decision boundary and mitigate the current trade-off problem. Below, we further detail them theoretically. + +Theorem 1. Let $H$ be the Hessian Matrix such that $H_{h_\theta}(\mathbf{x}) = \nabla_{\mathbf{x}}^2 h_\theta(\mathbf{x})$ , then with the new optimization objective of Robust Perception, we have: + +$$ +\forall \Delta , \Delta^ {\top} \cdot H _ {h _ {\theta}} (\mathbf {x}) \cdot \Delta \rightarrow 0. +$$ + +Proof. We defer the proof to Appendix B.1. + +This means Robust Perception limits the second-order (and higher-order as well) nonlinear effects within the adversarial perturbation to the model perception, such that the perception would change mainly along with the linear + +term of the perturbation, and thus be smoother. This also aligns with the local linearity hypothesis of adversarial samples [12], which assumes that linear approximations dominate the model behavior under adversarial perturbations. + +Theorem 2. Let $J$ be the Jacobian Matrix such that $J_{h_{\theta}}(\mathbf{x}) = \nabla_{\mathbf{x}} h_{\theta}(\mathbf{x})$ , then with the new optimization objective of Robust Perception, we have: + +$$ +\forall \alpha \in [ 0, 1 ], J _ {h _ {\boldsymbol {\theta}}} (\mathbf {x} + \boldsymbol {\alpha} \cdot \Delta) \rightarrow J _ {h _ {\boldsymbol {\theta}}} (\mathbf {x}), +$$ + +then with $\| \cdot \|_{\mathrm{spec}}$ denoting the Spectral Norm and $\gamma$ being any micro value, given $\| J_{h_\theta}(\mathbf{x} + \alpha \cdot \Delta) - J_{h_\theta}(\mathbf{x})\|_{\mathrm{spec}} \leq \gamma$ , the function $h_\theta(\mathbf{x})$ can be referred to as $K$ -Lipschitz with the Lipschitz constant $K$ upper-bounded by: + +$$ +K \leq \sup _ {\mathbf {x}} \| J _ {h _ {\theta}} (\mathbf {x}) \| _ {\operatorname {s p e c}} + \gamma . +$$ + +Proof. We defer the proof to Appendix B.2. + +This theorem tells us that Robust Perception also regularizes the change $\gamma$ in Jacobian along with the adversarial perturbation, which can be interpreted as limiting the increase of the global Lipschitz constant by $\gamma$ under perturbation. Previous works have demonstrated that this is expected to help achieve smoother decision boundaries and improve the accuracy-robustness trade-off [1, 11, 21, 33]. + +# 4.2. Robust Perception Adversarial Training + +Based on the above considerations, in this section, we propose a novel AT method named Robust Perception Adversarial Training (RPAT). It utilizes a new additional + +regularization term appended to the conventional AT objective to achieve the Robust Perception. Formally, given a $C$ -class classification task $(C \geq 2)$ and a dataset $\mathcal{D} = \{(\mathbf{x}_i, y_i)\}_{i=1,\dots,n}$ with benign sample $\mathbf{x}_i \in \mathbb{R}^d$ and supervised label $y_i \in \{1,\dots,C\}$ , let $\hat{\mathbf{x}}_i'$ denotes the adversarial sample corresponding to $\mathbf{x}_i$ , which is generated as $\hat{\mathbf{x}}_i' = \arg \max_{\|\mathbf{x}_i' - \mathbf{x}_i\|_p \leq \epsilon} \mathbb{1}(\arg \max_{k=1,\dots,C} h_\theta(\mathbf{x}_i'))^{(k)} \neq y_i)$ , then based on the Definition 1 in Section 4.1, the adversarial risk to be minimized for the proposed RPAT w.r.t. $\mathbf{x}_i$ can be formulated with the 0-1 loss [58] as: + +$$ +\begin{array}{l} \mathcal {R} (\boldsymbol {\theta}, \mathbf {x} _ {i}) := \mathbb {1} \left(f _ {\boldsymbol {\theta}} \left(\hat {\mathbf {x}} _ {i} ^ {\prime}\right) \neq y _ {i}\right) \\ + \mathbb {1} \left(\frac {h _ {\boldsymbol {\theta}} \left(\tilde {\mathbf {x}} _ {i}\right) - h _ {\boldsymbol {\theta}} \left(\mathbf {x} _ {i}\right)}{\alpha} \neq \frac {h _ {\boldsymbol {\theta}} \left(\hat {\mathbf {x}} _ {i} ^ {\prime}\right) - h _ {\boldsymbol {\theta}} \left(\tilde {\mathbf {x}} _ {i}\right)}{1 - \alpha}\right), \tag {2} \\ \end{array} +$$ + +where $\tilde{\mathbf{x}}_i = \mathbf{x}_i + \alpha \cdot (\hat{\mathbf{x}}_i' - \mathbf{x}_i)$ + +While the proposed Robust Perception aims to improve learning efficacy for the failure cases, the above risk does not explicitly distinguish success and failure cases. This is because Robust Perception is designed to mitigate the gap of the failure cases in the effectiveness of learning compared with the success ones, rather than require any new characteristic beyond the success cases. In other words, the success cases are expected to also satisfy Robust Perception at the beginning. Therefore, considering the computational efficiency, it is not necessary to specifically exclude the success cases for the Robust Perception term. + +Finally, by introducing practical surrogate losses to optimize the adversarial risk in Equation (2), we end up with our new RPAT method. To be specific, for the supervised learning term for adversarial samples, we adopt cross-entropy (CE) just as the conventional AT by default. Then, for the Robust Perception term, as it is formulated in a format of similarity regularization, various criterion that measures the similarity distance between the two input items are feasible options, such as mean squared error (MSE), Kullback-Leibler divergence (KL), Jensen-Shannon divergence (JS), and Cosine similarity. We simply adopt MSE loss for the basic version of RPAT here, and would also demonstrate the effectiveness of KL loss later in Section 5.3. With the logits $\mathbf{z}(\mathbf{x}_i,\boldsymbol {\theta})$ utilized as the hidden representation $h_\theta (\cdot)$ to denote the model perception, and $\mathbf{p}(\mathbf{x}_i,\boldsymbol {\theta})$ being the softmax output as given in Equation (1), the overall objective of RPAT can be formulated as: + +$$ +\begin{array}{l} \mathcal {L} ^ {R P A T} (\boldsymbol {\theta}, \mathcal {D}, \lambda , \alpha) := \frac {1}{n} \sum_ {i = 1} ^ {n} \left(\mathcal {L} ^ {C E} \left(\mathbf {p} \left(\hat {\mathbf {x}} _ {i} ^ {\prime}, \boldsymbol {\theta}\right), y _ {i}\right) \right. \tag {3} \\ + \lambda \cdot \mathcal {L} ^ {M S E} \left( \right.\frac {\mathbf {z} (\tilde {\mathbf {x}} _ {i} , \boldsymbol {\theta}) - \mathbf {z} (\mathbf {x} _ {i} , \boldsymbol {\theta})}{\alpha} \left| \right.\left| \right. \frac {\mathbf {z} (\hat {\mathbf {x}} _ {i} ^ {\prime} , \boldsymbol {\theta}) - \mathbf {z} (\tilde {\mathbf {x}} _ {i} , \boldsymbol {\theta})}{1 - \alpha}\left. \right)\left. \right), \\ \end{array} +$$ + +where $\lambda$ is a weight hyper-parameter to balance the conventional AT objective and our new Robust Perception objective. An intuitive illustration of the proposed RPAT method is provided in Figure 3. + +# 5. Experiments + +Following previous AT works, the effectiveness of the proposed RPAT is evaluated on CIFAR-10, CIFAR-100 [22] and Tiny-ImageNet [23] datasets. The main experiments can be divided into two parts. First, based on Tack et al. [40], we implement RPAT upon three common AT baselines, namely PGD-AT, TRADES, and MART, as well an advanced method, Consistency-AT, with ResNet-18 [15] architecture. This is to show the general effectiveness of the proposed Robust Perception in improving the current trade-off. Second, by integrating RPAT into one of the current SOTA methods, ReBAT [47], with PreActResNet-18 [16] and WideResNet-34-10 [56] architectures, we achieve new SOTA performance beyond 12 SOTA methods on the accuracy-robustness trade-off problem. This is to demonstrate the advancement achieved by the proposed RPAT in the AT area. The detailed experimental setup is provided in Appendix A. + +# 5.1. Evaluation Measures + +There are five measures in total adopted to evaluate the experimental methods from three different perspectives. Firstly, we report clean accuracy on benign test data, which reflects the ability of the robust models to maintain performance in benign cases. Secondly, we evaluate adversarial robustness under two adversarial attack methods, namely PGD-20 [29] and Auto-Attack (AA) [7], with default settings aligning with Appendix A and random starts. Finally, to test the performance w.r.t. the current accuracy-robustness trade-off problem in AT, on the one hand, we report the mean of the clean accuracy and the robust score under Auto-Attack acquired above, and on the other hand, we introduce the Natural-Robustness Ratio (NRR) [13]: + +$$ +N R R = \frac {2 \times C l e a n A c c u r a c y \times A d v e r s a r i a l R o b u s t n e s s}{C l e a n A c c u r a c y + A d v e r s a r i a l R o b u s t n e s s}, +$$ + +which emphasizes how well the robust model strikes a balance between clean accuracy and robustness. A lower NRR value suggests that the model might prioritize one side over the other. All the results reported are averages of three runs, with the "best" performance of each run recorded on the best checkpoint achieving the highest PGD-20 accuracy, which ensures that the training process is isolated from AA, the test-time adversary to evaluate the robustness. + +# 5.2. Effectiveness in Improving Baselines + +The proposed RPAT can be easily implemented with existing AT methods by simply appending the Robust Perception term. In this section, we illustrate the improvements that RPAT can bring to our four experimental benchmarks, namely PGD-AT, TRADES, MART, and Consistency-AT, on CIFAR-10, CIFAR-100, and Tiny-ImageNet with + +Table 1. Improvements that the proposed RPAT achieved upon four benchmarks. The results are acquired with ResNet-18 and ${\ell }_{\infty }$ norm. + +
DatasetMethodBenchmarks+RPAT (Ours)
CleanPGD-20AAMeanNRRCleanPGD-20AAMeanNRR
bestbestbestbestfinalbestfinalbestfinal
CIFAR-10PGD-AT82.9250.6146.7464.83059.78283.2082.9951.2950.9748.0047.6065.60060.878
TRADES79.6751.9147.6263.64559.61080.0280.0251.2151.2147.4747.4563.74559.590
MART77.9353.0046.7062.31558.40275.4475.9552.8852.8747.0347.3061.23557.940
Consistency-AT83.4251.9647.7265.57060.71184.1284.0452.3352.1448.9849.0666.55061.911
CIFAR-100PGD-AT56.5628.8025.0240.79034.69358.2258.5829.1628.5724.8824.2341.55034.862
TRADES55.3929.3624.5139.95033.98357.5057.2529.4229.0025.0524.4541.27534.897
MART49.8330.3825.0037.41533.29550.7250.7230.3330.3325.3425.0638.03033.796
Consistency-AT58.5329.2825.3941.96035.41760.3360.5129.9729.6426.3125.9743.32036.641
Tiny-ImageNetPGD-AT46.3221.5817.0731.69524.94747.6848.3222.1322.0017.7717.9132.72525.891
TRADES46.7521.5216.6031.67524.50048.7748.8521.9821.9316.9216.7132.84525.124
MART39.7022.9017.1828.44023.98241.7641.6623.4023.2917.7917.7829.77524.951
Consistency-AT46.5421.8917.6032.07025.54149.7449.8523.1622.8418.8418.7434.29027.329
+ +Table 2. Improvements that the proposed RPAT achieved upon four benchmarks. The results are acquired with ResNet-18 and ${\ell }_{2}$ norm. + +
DatasetMethodBenchmarks+RPAT (Ours)
CleanPGD-20AAMeanNRRCleanPGD-20AAMeanNRR
bestbestbestbestfinalbestfinalbestfinal
CIFAR-10PGD-AT87.7667.9266.3677.06075.57488.2088.4568.5668.5267.6367.6277.91576.557
TRADES83.9968.6065.9374.96073.87285.1784.8068.9568.9067.6767.3776.42075.418
MART84.0968.3266.2875.18574.13084.6584.6568.6068.6066.6966.7075.67074.604
Consistency-AT88.7669.3567.4678.11076.65889.3889.5270.4170.3669.4469.5579.41078.158
CIFAR-100PGD-AT65.0041.5439.2752.13548.96065.1465.4441.6141.2839.2339.2052.18548.969
TRADES61.2543.0540.1550.7048.50562.4162.6443.5643.1040.3640.1251.38549.020
MART60.0843.4239.8549.96547.9260.5760.5743.9243.9040.3640.3650.46548.442
Consistency-AT65.1442.2839.9252.53049.50365.5466.8142.8142.5140.2640.6152.90049.880
Tiny-ImageNetPGD-AT59.3643.1240.9050.13048.43160.2860.2844.0644.0641.8941.8751.08549.430
TRADES57.6144.7042.3349.97048.80258.4658.5044.8744.6542.4242.4450.44049.165
MART56.4744.8442.2249.34548.31657.1756.9545.0244.6742.5142.1649.84048.762
Consistency-AT61.4944.9242.5952.04050.32462.1962.6445.2844.8743.1742.8552.68050.963
+ +ResNet-18, demonstrating the wide effectiveness of our method. The results under $\ell_{\infty}$ and $\ell_{2}$ threat models are shown in Table 1 and Table 2, respectively, where the best clean accuracy, AA robustness, as well mean and NRR for their trade-off on each of the datasets are marked in bold, while individual scores not improved are marked with underline. Note that PGD-20 is the indicator for selecting the "best" checkpoint as explained in Section 5.1, and the "final" scores are additionally provided to show that RPAT does not suffer from serious robust-overfitting, which is also evidence for appropriate algorithm design. Thus, these fields are not involved in the comparison. It can be found that RPAT achieves significant improvement for most scores among the tables, where Consistency-AT+RPAT is basically the most competitive for all the experimental groups, which demonstrates the general effectiveness of our method in mitigating the current trade-off problem. + +# 5.3. Advancement beyond 12 SOTAs + +To demonstrate the advancement that the proposed RPAT contributes, we further integrate RPAT into ReBAT [47], + +one of the current SOTA AT methods on the accuracy-robustness trade-off problem. It implements WA [18], averaging multiple points along the SGD trajectory for better generalization, through the exponential moving average (EMA) [37] strategy with a decay rate of 0.999. Thus, we also involve WA itself (i.e., with PGD-AT), as well as $\mathrm{KD + SWA}$ [5] which shares a similar idea, for comparison. Other differences in ReBAT include $KL$ loss and a reduced learning rate decay factor from 0.1 to 0.5. We refer to our method upon ReBAT as $\mathrm{RPAT}^{++}$ , and compare it with 12 previous SOTAs introduced in Section 2.2, respectively with PreActResNet-18 and WideResNet-34-10 as shown in Table 3 and Table 4. The clean and AA scores of the SOTAs are from the best checkpoint among their original papers, Kanai et al. [19], Wang et al. [47], and our reproduction if any (so due to the lack of previous record, ADR and CURE are excluded for PreActResNet-18). Beyond the basic CIFAR-10 results under $\ell_{\infty}$ norm, Table 3 involves CIFAR-100 and $\ell_2$ norm, and Table 4 further consider the Cutmix [55] data augmentation strategy as marked by “*”, which ensures the comprehensiveness of our comparison. + +Table 3. Comparison of the proposed $\mathrm{RPAT^{+ + }}$ with current SOTAs on accuracy-robustness trade-off with PreActResNet-18. The best results are marked in bold. + +
NormMethodCIFAR-10CIFAR-100
CleanAAMeanNRRCleanAAMeanNRR
\( \ell_{\infty} \)WAUAI'1883.5049.8966.69562.46157.2625.8341.54535.601
MMAICLR'2085.5037.2061.35051.84460.6018.4039.50028.229
AWPNeurIPS'2081.1150.0965.60061.93354.1025.1639.63034.347
GAIRATICLR'2178.7037.7058.20050.97952.0019.8035.90028.680
KD+SWAICLR'2184.0649.8266.94062.56257.1725.6641.41535.422
EWATICML'2182.8048.2065.50060.93154.2023.5238.86032.805
MAILNeurIPS'2179.5039.6059.55052.86746.5016.7031.60024.574
TEICLR'2282.0450.1266.08062.22556.4125.8441.12535.444
SOVRICML'2381.9049.4065.65061.62852.1024.3038.20033.142
ReBATNeurIPS'2382.0950.7266.40562.70056.1327.6041.86537.004
RPAT++Ours82.6351.0066.81563.07256.8427.6842.26037.230
\( \ell_{2} \)ReBATNeurIPS'2388.7971.0079.89578.90565.5842.6754.12551.701
RPAT++Ours89.0671.2680.16079.17265.6342.8554.24051.848
+ +Table 4. Comparison of ${\mathrm{{RPAT}}}^{+ + }$ with SOTAs on CIFAR-10 with WideResNet-34-10 and ${\ell }_{\infty }$ norm. + +
MethodCleanAAMeanNRR
WAUAI'1887.6652.6570.15565.787
MMAICLR'2087.8043.1065.45057.818
AWPNeurIPS'2085.6353.3269.47565.718
GAIRATICLR'2183.0041.8062.40055.599
KD+SWAICLR'2187.4553.5970.52066.456
EWATICML'2186.0051.6068.80064.500
MAILNeurIPS'2182.2043.3062.75056.721
TEICLR'2285.9752.8869.42565.482
SOVRICML'2385.0053.1069.05065.366
ReBATNeurIPS'2385.2554.7870.01566.700
ADRICLR'2484.6753.2568.96065.381
CUREICLR'2487.0552.1069.57565.186
\(RPAT^{++}\)Ours86.7654.9770.86567.300
ReBAT*NeurIPS'2386.6655.6471.15067.769
\(RPAT^{++*}\)Ours87.5755.7971.68068.158
+ +![](images/feec7bf49567c597de2bfb94966fcd39f9b246c998a5ab54408b2f62b07b554c.jpg) +(a) Impact of loss function. + +![](images/a1d5beb1720d50f0b83416088093087387fc3e74d0ea117a517ad85dff5164c0.jpg) +(b) Impact of hyper-parameter $\alpha$ +Figure 4. Two ablation studies of the proposed RPAT on CIFAR-10 with ResNet-18 under $\ell_{\infty}$ norm. The MSE loss in (a) and the points with $\alpha = 0.5$ in (b) correspond to the record of "PGD-AT+RPAT" on CIFAR-10 in Table 1. + +The results show that the proposed $\mathrm{RPAT^{+ + }}$ achieves new SOTA performance for most of the measures beyond previous SOTAs on the accuracy-robustness trade-off problem, which strongly supports the contribution of this work. + +# 5.4. Ablation Study + +As the proposed RPAT can be implemented upon existing AT methods by simply appending the Robust Perception term into their original objectives, the comparison in performance between and after adding RPAT in Section 5.2 and Section 5.3 have achieved a comprehensive ablation study on the effectiveness of RPAT. Therefore, in this section, we mainly investigate the impact of two setting details, respectively the loss function adopted for the Robust Perception term (Figure 4 (a)) and the specific value for the hyperparameter $\alpha$ (Figure 4 (b)). Besides, we further discuss an optional trick suggested by ReBAT but excluded in the above experiments in Appendix 5.3. + +Specifically, just as we expected, different loss functions to calculate the similarity between model perceptions, including MSE, KL, JS and Cosine, would not significantly + +impact the final performance of RPAT, which suggests that the effectiveness of our method is from the successful conceptual design of Robust Perception, instead of relying on specific implementation of the regularization term. Then, for $\alpha$ , we try 0.2 and 0.8 besides the default value 0.5 on the one hand, and also introduce beta distribution $\beta \sim Beta(0.75, 0.75)$ as suggested by Zhang et al. [57] on the other hand. Specifically, we try $\alpha = \min(\beta, (1 - \beta))$ and $\alpha = \max(\beta, (1 - \beta))$ , namely indicated as "Beta-" and "Beta+" in (Figure 4 (b)), as they respectively approach adversarial and benign samples more. The results show that the default $\alpha = 0.5$ is an appropriate choice, which is reasonable because the midpoint is expected to be more representative for the ideal target that Robust Perception holds for all $\alpha \in [0, 1]$ . + +# 6. Conclusion + +This work studies the current accuracy-robustness trade-off problem in AT. To the best of our knowledge, this is the first work that explicitly attributes the trade-off problem to the over-sufficient (rather than insufficient) learning of hard adversarial samples. We surprisingly reveal that such hard samples are already learned better than those successfully defended in model perception, which means the conventional AT objective is no longer effective to further optimize the model w.r.t. these hard samples. Instead, it forces the model to ignore the perturbations, in turn degrading their effectiveness in supporting the establishment of decision boundary to an appropriate location. To mitigate this, our new AT objective, Robust Perception, encourages smoother perception transition towards the decision boundary along with perturbations, based on which the proposed RPAT method achieves new SOTA performance. Still, exploring better methods to approximate this ideal objective would be a promising future direction. + +# Acknowledgements + +This work was supported by the National Natural Science Foundation of China (No. 62471420), GuangDong Basic and Applied Basic Research Foundation (2025A15150122 96), and CCF-Tencent Rhino-Bird Open Research Fund. + +# References + +[1] Abulikemu Abuduweili and Changliu Liu. Estimating neural network robustness via lipschitz constant and architecture sensitivity. In Safe and Robust Robot Learning for Operation in the Real World Workshop @ CoRL, 2024. 2, 5, 11 +[2] Anish Athalye, Nicholas Carlini, and David Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In ICML, 2018. 2 +[3] Tao Bai, Jinqi Luo, Jun Zhao, Bihan Wen, and Qian Wang. Recent advances in adversarial training for adversarial robustness. In IJCAI, 2021. 2, 3 +[4] Battista Biggio, Igino Corona, Davide Maiorca, Blaine Nelson, Nedim Šrndić, Pavel Laskov, Giorgio Giacinto, and Fabio Roli. Evasion attacks against machine learning at test time. In ECML-PKDD, 2013. 1 +[5] Tianlong Chen, Zhenyu Zhang, Sijia Liu, Shiyu Chang, and Zhangyang Wang. Robust overfitting may be mitigated by properly learned smoothening. In ICLR, 2021. 3, 7 +[6] Minhao Cheng, Qi Lei, Pin-Yu Chen, Inderjit Dhillon, and Cho-Jui Hsieh. Cat: Customized adversarial training for improved robustness. In *IJCAI*, 2022. 2, 3 +[7] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In ICML, 2020. 2, 6 +[8] Gavin Weiguang Ding, Yash Sharma, Kry Yik Chau Lui, and Ruitong Huang. Mma training: Direct input space margin maximization through adversarial training. In ICLR, 2020. 2, 3 +[9] Yinpeng Dong, Qi-An Fu, Xiao Yang, Tianyu Pang, Hang Su, Zihao Xiao, and Jun Zhu. Benchmarking adversarial robustness on image classification. In CVPR, 2020. 2 +[10] Yinpeng Dong, Ke Xu, Xiao Yang, Tianyu Pang, Zhijie Deng, Hang Su, and Jun Zhu. Exploring memorization in adversarial training. In ICLR, 2022. 2, 3 +[11] Mahyar Fazlyab, Taha Entesari, Aniket Roy, and Rama Chellappa. Certified robustness via dynamic margin maximization and improved lipschitz regularization. In NeurIPS, 2023. 2, 5, 11 +[12] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In ICLR, 2015. 1, 2, 5, 11 +[13] Shruthi Gowda, Bahram Zonooz, and Elahe Arani. Conserve-update-revise to cure generalization and robustness trade-off in adversarial training. In ICLR, 2024. 3, 6 +[14] Chuan Guo, Mayank Rana, Moustapha Cisse, and Laurens van der Maaten. Countering adversarial images using input transformations. In ICLR, 2018. 1, 2 +[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6 + +[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In ECCV, 2016. 6 +[17] Xiaowei Huang, Daniel Kroening, Wenjie Ruan, James Sharp, Youcheng Sun, Emese Thamo, Min Wu, and Xinping Yi. A survey of safety and trustworthiness of deep neural networks: Verification, testing, adversarial attack and defence, and interpretability. Computer Science Review, 2020. 1 +[18] Pavel Izmailov, Dmitrii Podoprikhin, Timur Garipov, Dmitry Vetrov, and Andrew Gordon Wilson. Averaging weights leads to wider optima and better generalization. In UAI, 2018. 3, 7 +[19] Sekitoshi Kanai, Shin'ya Yamaguchi, Masanori Yamada, Hiroshi Takahashi, Kentaro Ohno, and Yasutoshi Ida. One-vs-the-rest loss to focus on important samples in adversarial training. In ICML, 2023. 2, 3, 7 +[20] Minseon Kim, Jihoon Tack, Jinwoo Shin, and Sung Ju Hwang. Entropy weighted adversarial training. In Adversarial Machine Learning Workshop @ ICML, 2021. 3 +[21] Vishaal Krishnan, Al Makedah, Abed AlRahman, and Fabio Pasqualetti. Lipschitz bounds and provably robust training by laplacian smoothing. In NeurIPS, 2020. 2, 5, 11 +[22] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. 2009. 6 +[23] Fei-Fei Li, Andrej Karpathy, and Justin Johnson. Tiny image dataset. 2015. 6 +[24] Zi Liang, Pinghui Wang, Ruofei Zhang, Nuo Xu, Shuo Zhang, Lifeng Xing, Haitao Bai, and Ziyang Zhou. Merge: Fast private text generation. In AAAI, 2024. 1 +[25] Fangzhou Liao, Ming Liang, Yinpeng Dong, Tianyu Pang, Xiaolin Hu, and Jun Zhu. Defense against adversarial attacks using high-level representation guided denoiser. In CVPR, 2018. 1, 2 +[26] Weilin Lin, Nanjun Zhou, Yanyun Wang, Jianze Li, Hui Xiong, and Li Liu. Backdoor: A comprehensive benchmark for backdoor learning in diffusion model. arXiv preprint arXiv:2502.11798, 2025. 1 +[27] Feng Liu, Bo Han, Tongliang Liu, Chen Gong, Gang Niu, Mingyuan Zhou, Masashi Sugiyama, et al. Probabilistic margins for instance reweighting in adversarial training. In NeurIPS, 2021. 2, 3 +[28] Xin Liu, Yichen Yang, Kun He, and John E Hopcroft. Parameter interpolation adversarial training for robust image classification. IEEE Transactions on Information Forensics and Security (TIFS), 2025. 2, 3 +[29] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In ICLR, 2018. 2, 3, 6 +[30] Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. In ICLR, 2018. 12 +[31] Tianyu Pang, Xiao Yang, Yinpeng Dong, Hang Su, and Jun Zhu. Bag of tricks for adversarial training. In ICLR, 2021. 11 +[32] Nicolas Papernot, Patrick McDaniel, Ian Goodfellow, Somesh Jha, Z Berkay Celik, and Ananthram Swami. Practi + +cal black-box attacks against machine learning. In AsiaCCS, 2017. 1, 2 +[33] Patricia Pauli, Anne Koch, Julian Berberich, Paul Kohler, and Frank Allgower. Training robust neural networks using lipschitz bounds. IEEE Control Systems Letters, 2021. 2, 5, 11 +[34] Rahul Rade and Seyed-Mohsen Moosavi-Dezfooli. Reducing excessive margin to achieve a better accuracy vs. robustness trade-off. In ICLR, 2022. 2, 3 +[35] Aditi Raghunathan, Sang Michael Xie, Fanny Yang, John Duchi, and Percy Liang. Adversarial training can hurt generalization. In Identifying and Understanding Deep Learning Phenomena Workshop @ ICML, 2019. +[36] Aditi Raghunathan, Sang Michael Xie, Fanny Yang, John Duchi, and Percy Liang. Understanding and mitigating the tradeoff between robustness and accuracy. In ICML, 2020. 2, 3 +[37] Sylvestre-Alvise Rebuffi, Sven Gowal, Dan Andrei Calian, Florian Stimberg, Olivia Wiles, and Timothy A Mann. Data augmentation can improve robustness. In NeurIPS, 2021. 7 +[38] David Stutz, Matthias Hein, and Bernt Schiele. Confidence-calibrated adversarial training: Generalizing to unseen attacks. In ICML, 2020. 3 +[39] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In ICLR, 2014. 1, 2 +[40] Jihoon Tack, Sihyun Yu, Jongheon Jeong, Minseon Kim, Sung Ju Hwang, and Jinwoo Shin. Consistency regularization for adversarial robustness. In AAAI, 2022. 3, 6 +[41] Lue Tao, Lei Feng, Jinfeng Yi, Sheng-Jun Huang, and Songcan Chen. Better safe than sorry: Preventing delusive adversaries with adversarial training. In NeurIPS, 2021. 2 +[42] Florian Tramer, Nicholas Carlini, Wieland Brendel, and Aleksander Madry. On adaptive attacks to adversarial example defenses. In NeurIPS, 2020. 2 +[43] Dimitris Tsipras, Shibani Santurkar, Logan Engstrom, Alexander Turner, and Aleksander Madry. Robustness may be at odds with accuracy. In ICLR, 2019. 2, 3 +[44] Haotao Wang, Tianlong Chen, Shupeng Gui, TingKuei Hu, Ji Liu, and Zhangyang Wang. Once-for-all adversarial training: In-situ tradeoff between robustness and accuracy for free. In NeurIPS, 2020. 2, 3 +[45] Yisen Wang, Difan Zou, Jinfeng Yi, James Bailey, Xingjun Ma, and Quanquan Gu. Improving adversarial robustness requires revisiting misclassified examples. In ICLR, 2020. 1, 3 +[46] Yanyun Wang, Dehui Du, Haibo Hu, Zi Liang, and Yuan-hao Liu. Tsfool: Crafting highly-imperceptible adversarial time series through multi-objective attack. arXiv preprint arXiv:2209.06388, 2022. 2 +[47] Yifei Wang, Liangchen Li, Jiansheng Yang, Zhouchen Lin, and Yisen Wang. Balance, imbalance, and rebalance: Understanding robust overfitting from a minimax game perspective. In NeurIPS, 2023. 3, 6, 7 +[48] Yanyun Wang, Li Liu, Zi Liang, Qingqing Ye, and Haibo Hu. New paradigm of adversarial training: Releasing accuracy + +robustness trade-off via dummy class. arXiv preprint arXiv:2410.12671, 2024. 2 +[49] Dongxian Wu, Shu-Tao Xia, and Yisen Wang. Adversarial weight perturbation helps robust generalization. In NeurIPS, 2020. 3 +[50] Yu-Yu Wu, Hung-Jui Wang, and Shang-Tse Chen. Annealing self-distillation rectification improves adversarial training. In ICLR, 2024. 3 +[51] Cihang Xie, Jianyu Wang, Zhishuai Zhang, Zhou Ren, and Alan Yuille. Mitigating adversarial effects through randomization. In ICLR, 2018. 1, 2 +[52] Weilin Xu, David Evans, and Yanjun Qi. Feature squeezing: Detecting adversarial examples in deep neural networks. In NDSS, 2018. 1, 2 +[53] Shuo Yang and Chang Xu. One size does not fit all: Data-adaptive adversarial training. In ECCV, 2022. 2, 3 +[54] Jia-Li Yin, Bin Chen, Wanqing Zhu, Bo-Hao Chen, and Ximeng Liu. Push stricter to decide better: A class-conditional feature adaptive framework for improving adversarial robustness. IEEE Transactions on Information Forensics and Security (TIFS), 2023. 2, 3 +[55] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In ICCV, 2019. 7 +[56] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In BMVC, 2016. 6 +[57] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. Mixup: Beyond empirical risk minimization. In ICLR, 2019. 8 +[58] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In ICML, 2019, 2, 3, 6, 11 +[59] Jingfeng Zhang, Jianing Zhu, Gang Niu, Bo Han, Masashi Sugiyama, and Mohan Kankanhalli. Geometry-aware instance-reweighted adversarial training. In ICLR, 2021. 2, 3 +[60] Ziyu Zhou, Gengyu Lyu, Yiming Huang, Zihao Wang, Ziyu Jia, and Zhen Yang. Sdformer: Transformer with spectral filter and dynamic attention for multivariate time series long-term forecasting. In IJCAI, 2024. 1 \ No newline at end of file diff --git a/2025/Failure Cases Are Better Learned But Boundary Says Sorry_ Facilitating Smooth Perception Change for Accuracy-Robustness Trade-Off in Adversarial Training/images.zip b/2025/Failure Cases Are Better Learned But Boundary Says Sorry_ Facilitating Smooth Perception Change for Accuracy-Robustness Trade-Off in Adversarial Training/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cfc643bd42e025fcdfe6407fb92a917cf3506907 --- /dev/null +++ b/2025/Failure Cases Are Better Learned But Boundary Says Sorry_ Facilitating Smooth Perception Change for Accuracy-Robustness Trade-Off in Adversarial Training/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c94d5c56495dd11289c4965a5791b7d73ce869ff176d76b11177e4fae44b6051 +size 575117 diff --git a/2025/Failure Cases Are Better Learned But Boundary Says Sorry_ Facilitating Smooth Perception Change for Accuracy-Robustness Trade-Off in Adversarial Training/layout.json b/2025/Failure Cases Are Better Learned But Boundary Says Sorry_ Facilitating Smooth Perception Change for Accuracy-Robustness Trade-Off in Adversarial Training/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..43162df7ad913571efa2249533775072b47c68df --- /dev/null +++ b/2025/Failure Cases Are Better Learned But Boundary Says Sorry_ Facilitating Smooth Perception Change for Accuracy-Robustness Trade-Off in Adversarial Training/layout.json @@ -0,0 +1,8181 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 103, + 555, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 103, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 103, + 555, + 140 + ], + "type": "text", + "content": "Failure Cases Are Better Learned But Boundary Says Sorry: Facilitating Smooth Perception Change for Accuracy-Robustness Trade-Off in Adversarial Training" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 244, + 161, + 367, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 161, + 367, + 175 + ], + "spans": [ + { + "bbox": [ + 244, + 161, + 367, + 175 + ], + "type": "text", + "content": "Yanyun Wang Li Liu*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 179, + 470, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 179, + 470, + 194 + ], + "spans": [ + { + "bbox": [ + 137, + 179, + 470, + 194 + ], + "type": "text", + "content": "The Hong Kong University of Science and Technology (Guangzhou)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 220, + 200, + 234 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 220, + 200, + 234 + ], + "spans": [ + { + "bbox": [ + 151, + 220, + 200, + 234 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 246, + 297, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 246, + 297, + 605 + ], + "spans": [ + { + "bbox": [ + 54, + 246, + 297, + 605 + ], + "type": "text", + "content": "Adversarial Training (AT) is one of the most effective methods to train robust Deep Neural Networks (DNNs). However, AT creates an inherent trade-off between clean accuracy and adversarial robustness, which is commonly attributed to the more complicated decision boundary caused by the insufficient learning of hard adversarial samples. In this work, we reveal a counterintuitive fact for the first time: From the perspective of perception consistency, hard adversarial samples that can still attack the robust model after AT are already learned better than those successfully defended. Thus, different from previous views, we argue that it is rather the over-sufficient learning of hard adversarial samples that degrades the decision boundary and contributes to the trade-off problem. Specifically, the excessive pursuit of perception consistency would force the model to view the perturbations as noise and ignore the information within them, which should have been utilized to induce a smoother perception transition towards the decision boundary to support its establishment to an appropriate location. In response, we define a new AT objective named Robust Perception, encouraging the model perception to change smoothly with input perturbations, based on which we propose a novel Robust Perception Adversarial Training (RPAT) method, effectively mitigating the current accuracy-robustness trade-off. Experiments on CIFAR-10, CIFAR-100, and Tiny-ImageNet with ResNet-18, PreActResNet-18, and WideResNet-34-10 demonstrate the effectiveness of our method beyond four common baselines and 12 state-of-the-art (SOTA) works. The code is available at https://github.com/FlaAI/RPAT." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 627, + 135, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 627, + 135, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 627, + 135, + 639 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "type": "text", + "content": "Deep neural networks (DNNs) have demonstrated impressive performance across various real-world applications [24, 26, 60]. However, they are inherently vulnerable to adversarial attacks [4, 39]. Specifically, even impercep" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 219, + 434, + 316 + ], + "blocks": [ + { + "bbox": [ + 317, + 219, + 434, + 316 + ], + "lines": [ + { + "bbox": [ + 317, + 219, + 434, + 316 + ], + "spans": [ + { + "bbox": [ + 317, + 219, + 434, + 316 + ], + "type": "image", + "image_path": "3af3a29bf8b909c59ece5b66e79e02c7ee88bde2635c70c29200b013605ee149.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 443, + 219, + 553, + 316 + ], + "blocks": [ + { + "bbox": [ + 443, + 219, + 553, + 316 + ], + "lines": [ + { + "bbox": [ + 443, + 219, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 443, + 219, + 553, + 316 + ], + "type": "image", + "image_path": "be6a139caa4fa6f60e810e855d5df80c8bbfbb95dd96fc195ab83bf5118e09d4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 318, + 317, + 434, + 408 + ], + "blocks": [ + { + "bbox": [ + 318, + 317, + 434, + 408 + ], + "lines": [ + { + "bbox": [ + 318, + 317, + 434, + 408 + ], + "spans": [ + { + "bbox": [ + 318, + 317, + 434, + 408 + ], + "type": "image", + "image_path": "5a8f87284bdeaa37efaf2340baceebfc4f83998107bbda98f0f3e654e80ec287.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 443, + 556, + 609 + ], + "lines": [ + { + "bbox": [ + 313, + 443, + 556, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 556, + 609 + ], + "type": "text", + "content": "Figure 1. Success and failure cases of AT defense from the perspectives of prediction and perception, respectively. (a) and (b): AT works through building a robust ball with radius " + }, + { + "bbox": [ + 313, + 443, + 556, + 609 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 443, + 556, + 609 + ], + "type": "text", + "content": " such that the model prediction remains unchanged within it. Then if the robust " + }, + { + "bbox": [ + 313, + 443, + 556, + 609 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 443, + 556, + 609 + ], + "type": "text", + "content": "-ball fails to contain the " + }, + { + "bbox": [ + 313, + 443, + 556, + 609 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 443, + 556, + 609 + ], + "type": "text", + "content": "-ball, there would still be failure cases beating the robust model under perturbation " + }, + { + "bbox": [ + 313, + 443, + 556, + 609 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 443, + 556, + 609 + ], + "type": "text", + "content": " within the " + }, + { + "bbox": [ + 313, + 443, + 556, + 609 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 443, + 556, + 609 + ], + "type": "text", + "content": "-ball. (c) and (d): In this work, we reveal that the smaller " + }, + { + "bbox": [ + 313, + 443, + 556, + 609 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 443, + 556, + 609 + ], + "type": "text", + "content": "-ball in the failure cases is not a result of more vulnerable model perception as previous intuition. In fact, as demonstrated by the contour line, which shows the perturbation ball to the same degree of perception change, the failure cases already achieve less perception change under similar " + }, + { + "bbox": [ + 313, + 443, + 556, + 609 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 443, + 556, + 609 + ], + "type": "text", + "content": " within the " + }, + { + "bbox": [ + 313, + 443, + 556, + 609 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 443, + 556, + 609 + ], + "type": "text", + "content": "-ball. This implies that the issue of these failure cases is not failing to learn perception consistency following the current AT objective, but failing to support the establishment of decision boundary to an appropriate location." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 436, + 318, + 553, + 417 + ], + "blocks": [ + { + "bbox": [ + 436, + 318, + 553, + 417 + ], + "lines": [ + { + "bbox": [ + 436, + 318, + 553, + 417 + ], + "spans": [ + { + "bbox": [ + 436, + 318, + 553, + 417 + ], + "type": "image", + "image_path": "9c677d01bc3c8323c7274ecfa24a41d02b945ed5d31859ccbbddafa4d943183e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 319, + 411, + 550, + 434 + ], + "lines": [ + { + "bbox": [ + 319, + 411, + 550, + 434 + ], + "spans": [ + { + "bbox": [ + 319, + 411, + 550, + 434 + ], + "type": "text", + "content": "(c) (d) Contour line of model perception" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "tible perturbations to inputs can cause substantial changes in the outputs [12, 17]. As this problem raises significant security concerns [45], previously, several defense strategies, such as Defense Distillation [32], Feature Squeezing [52], Randomization [51], and Input Denoising [14, 25], have been proposed to improve the adversarial robustness of DNNs. However, most of these approaches have proven" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 285, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 285, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 285, + 713 + ], + "type": "text", + "content": "*Corresponding Author: avrillliu@hkust-gz.edu.cn" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "text", + "content": "4691" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "ineffective against more advanced adaptive attacks [2, 42], such as Auto-Attack [7]. Currently, Adversarial Training (AT) [12, 29] is regarded as one of the most effective methods for training robust DNNs [2, 9]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 122, + 296, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 122, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 56, + 122, + 296, + 434 + ], + "type": "text", + "content": "However, an accuracy-robustness trade-off problem have been widely identified for AT in recent studies [3, 28, 34-36, 43, 44, 48, 54, 58]. Compared with standard training, the improvement in adversarial robustness acquired by AT is usually at the expense of the reduction in clean accuracy, which is afraid to significantly degrade the user experience in clean context and hinder the application of AT in real-world practices. The current consensus is that AT aggravates the accuracy-robustness trade-off by creating overly complex decision boundaries, primarily because models struggle to sufficiently learn from hard adversarial samples during training [6, 8, 10, 19, 27, 28, 34, 48, 53, 54]. Specifically, as the conventional AT objective encourages the model to perceive any adversarial inputs within the " + }, + { + "bbox": [ + 56, + 122, + 296, + 434 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 56, + 122, + 296, + 434 + ], + "type": "text", + "content": "-ball similar to the corresponding benign one to produce consistent predictions for them, previous works empirically assume that the wrong predictions of the hard adversarial samples are attributed to the insufficiently optimized perception of them under the current objective. Therefore, although defining \"hard sample\" from different perspectives (e.g., classification result and distance to decision boundary), previous works agree to further optimize the perception consistency on them, ending up with various learning strategies specific for them, such as data reweighting [27, 59], adaptive perturbation radius [8, 34], and non-one-hot supervised signal [10, 19]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 436, + 295, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 295, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 295, + 686 + ], + "type": "text", + "content": "However, in this work, we reveal that for the robust model trained by convention AT, the statistical difference in model perception from benign to hard adversarial samples causing failure defense is already smaller than to others (i.e., those successfully defended). It implies that the model has even more sufficiently learned them under the current AT objective, and should have successfully defended them if the above assumption holds. Thus, as demonstrated in Figure 1, we suggest that the key factor here is no longer the optimization of model perception but the location of the decision boundary. In that case, keeping on pursuing over-sufficient small perception change would not only help little in further reducing the prediction error, but also force the model to view the perturbations as noise and ignore the information within them, which impairs their ability in claiming robust balls with reasonable gradients to push the decision boundary to an appropriate location. As a consequence, such sample-level imbalance in boundary establishment would naturally increase its complexity, harming generalization ability and degrading the trade-off between accuracy and robustness." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "To mitigate the over-sufficient learning of model perception on the failure cases and retain effective gradient sig" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "type": "text", + "content": "nals w.r.t. them to facilitate the establishment of the decision boundary to an appropriate location, firstly, we define a new AT objective named Robust Perception in addition to the conventional one, encouraging smoother perception changes towards the decision boundary along with the input perturbations. Rigorous theoretical derivations also demonstrate its effectiveness in smoothening the decision boundary and releasing the current trade-off problem from the perspectives of local linearity [12] and Lipschitz regularization [1, 11, 21, 33]. Secondly, based on our new objective, we propose a novel Robust Perception Adversarial Training (RPAT) method, concurrently improving clean accuracy and adversarial robustness in AT. Extensive experiments on CIFAR-10, CIFAR-100, and Tiny-ImageNet with ResNet-18, PreActResNet-18, and WideResNet-34-10 demonstrate the advancement of RPAT beyond four common benchmarks and 12 state-of-the-art (SOTA) works." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 285, + 493, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 493, + 298 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 493, + 298 + ], + "type": "text", + "content": "2. Background and Related Works" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 304, + 511, + 317 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 511, + 317 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 511, + 317 + ], + "type": "text", + "content": "2.1. Robustness and Adversarial Training" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 322, + 553, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 322, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 313, + 322, + 553, + 465 + ], + "type": "text", + "content": "Adversarial robustness reflects model vulnerability to malicious perturbations, which is commonly quantified by test accuracy under adversarial attacks [3, 39, 46]. Representatively, the Fast Gradient Sign Method (FGSM) [12] perturbs input samples along the gradient of the loss function, whereas the Projected Gradient Descent (PGD) attack [29] applies FGSM iteratively and projects the perturbation onto the " + }, + { + "bbox": [ + 313, + 322, + 553, + 465 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 322, + 553, + 465 + ], + "type": "text", + "content": "-ball. Auto-Attack [7] introduces and combines several parameter-free attacks to mitigate practical issues such as hyper-parameter tuning and gradient masking. Due to their effectiveness, PGD and Auto-Attack are widely used to assess adversarial robustness, and so does this work." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 466, + 553, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 466, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 466, + 553, + 586 + ], + "type": "text", + "content": "Previously, although many defense strategies including Defense Distillation [32], Feature Squeezing [52], Input Denoising [14, 25] and Randomization [51] have been developed, most rely on obfuscated gradients [2] and can be evaded by advanced adaptive attacks [42]. Nowadays, Adversarial Training [12, 29] is recognized as the most effective method to enhance adversarial robustness [2, 9, 41]. Unlike clean training on benign samples, AT directly learns on adversarily augmented data. Its objective can be formulated as the following min-max problem [29]:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 356, + 592, + 553, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 592, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 356, + 592, + 553, + 622 + ], + "type": "interline_equation", + "content": "\\min _ {\\boldsymbol {\\theta}} \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\max _ {\\| \\mathbf {x} _ {i} ^ {\\prime} - \\mathbf {x} _ {i} \\| _ {P} \\leq \\epsilon} \\mathcal {L} \\left(f _ {\\boldsymbol {\\theta}} \\left(\\mathbf {x} _ {i} ^ {\\prime}\\right), y _ {i}\\right), \\tag {1}", + "image_path": "41b6e70c5006715c59f5c8fe937752637337ccdc12d8100486cf4fe70016586a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "inline_equation", + "content": "f_{\\pmb{\\theta}}(\\mathbf{x}_i) = \\arg \\max_{c=1,\\dots,C} \\mathbf{p}_c(\\mathbf{x}_i, \\pmb{\\theta})" + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_c" + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "content": " denoting the softmax probability of class " + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "content": " within the prediction of " + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "content": " denotes the number of training samples, and " + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i'" + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "content": " is generated by adding the strongest " + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "inline_equation", + "content": "L_p" + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "content": "-norm perturbation within the " + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "content": "-ball to the original sample " + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "content": ". This objective aims at prediction consistency within the " + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "content": "-ball, encouraging the model to also produce similar perception there." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4692" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "type": "text", + "content": "PGD-AT [29] is the first attempt at solving this min-max optimization iteratively by approximating the inner maximization via PGD to craft adversarial samples and then performing stochastic gradient descent (SGD) on these samples for the outer minimization. TRADES [58] first considers balancing clean accuracy and adversarial robustness in AT, which decomposes robust error into natural error and boundary error, with the latter occurring only if specific data points are sufficiently close to the decision boundary. MART [45] further refines AT by distinguishing misclassified from correctly classified samples, which suggests appending boundary error only for misclassified samples to the basic AT objective in PGD-AT. Consistency-AT [40] encourages the prediction between adversarial samples derived from different augmentations of the same benign sample to be similar, which explicitly enhances the prediction consistency further." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 287, + 258, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 287, + 258, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 287, + 258, + 300 + ], + "type": "text", + "content": "2.2. Accuracy-Robustness Trade-Off in AT" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 305, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 305, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 305, + 295, + 556 + ], + "type": "text", + "content": "An important problem identified for AT is the trade-off between clean accuracy and adversarial robustness [3, 28, 34-36, 43, 44, 54, 58], which means the robustness gained by AT often comes at the cost of reduced model accuracy compared with clean training. Consequently, AT can degrade user experience under benign conditions, which may seriously delay its real-world deployment. A widely acknowledged cause of this problem is that AT tends to learn more complicated decision boundaries than clean training [6, 10, 28, 34, 53, 54], thereby harming the ability of the robust model to generalize to unseen data. A number of previous works attribute this to the insufficient learning of hard adversarial samples [8, 10, 19, 27, 34, 59]. Though defining \"hard adversarial samples\" from different perspectives, such as the classification result and the distance to decision boundary, they basically reached a consensus that the model perception w.r.t. hard adversarial samples is not well-optimized, and as a consequence, fails to produce the consistent final prediction. In response, these works propose various solutions to help the model learn better prediction consistency for those hard adversarial samples." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": "For instance, GAIRAT [59] proposes that hard samples nearer to the decision boundary are more critical but less robust and should be assigned a larger weight. MAIL [27] further proposes three types of probabilistic margins to quantify the closeness for better reweighting adversarial data accordingly. In contrast, MMA [8] suggests learning hard adversarial samples with reduced perturbations, which employs an adaptive " + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": " for perturbations to maximize the margin between data and the decision boundary. EWAT [20] weighs the loss for each adversarial sample proportionally to the entropy of its prediction distribution during AT to focus on those with more uncertain labels. TE [10] reveals that the model may attempt to memorize certain hard ad" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "type": "text", + "content": "versarial samples during AT, as it is difficult to assign high-confidence one-hot labels for them [6, 38]. SOVR [19] suggests increasing the margins of logits for certain hard samples by switching to a one-vs-the-rest loss." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 121, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 553, + 300 + ], + "type": "text", + "content": "Additionally, several works also study the trade-off problem from other perspectives. AWP [49] proposes a double perturbation mechanism that flattens the loss landscape by weight perturbation to improve robust generalization. KD+SWA [5] injects more smoothening during AT by leveraging knowledge distillation to smooth the logits and performing stochastic weight averaging [18] to smooth the weights. ADR [50] employs soft labels as a guidance mechanism, accurately reflecting the distribution shift under attack during AT. CURE [13] selectively updates specific layers while preserving others to enhance the learning capacity of the model. ReBAT [47] views AT as a dynamic min-max game between the model trainer and the attacker, rebalancing the two players by either constraining the trainer's capacity or intensifying the attack." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 310, + 549, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 310, + 549, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 310, + 549, + 323 + ], + "type": "text", + "content": "3. Motivation: AT Over-Learns Failure Cases" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 331, + 553, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 553, + 426 + ], + "type": "text", + "content": "It is basically common sense for DNN-based classification that the model learns to perceive input better and is thus able to predict it more precisely (especially when the model perception is indicated by its logits). Therefore, it seems reasonable at first glance that the previous works assume that the wrong prediction of the hard adversarial samples is due to the insufficient model perception of them, and propose various solutions upon this assumption in response." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 426, + 553, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 553, + 569 + ], + "type": "text", + "content": "However, to the best of our knowledge, there is no previous discussion on a simple but crucial question: What if the assumption does not hold? If the model perception on the hard adversarial samples is already relatively close to the corresponding benign ones (e.g., at a similar or even better level than non-hard ones), then it would be no longer effective enough to further optimize the model on such samples through an objective simply pursuing prediction consistency as the conventional AT. Also, in turn, this can no longer effectively benefit the further reduction of the prediction error. Surprisingly, our following proof-of-concept experiments reveal that this is exactly the case." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "type": "text", + "content": "The proof-of-concept experiments are conducted on CIFAR-10 dataset with ResNet-18 architecture, with the results illustrated in Figure 2. As suggested by the x-axis, we involve three cases with different pairs of training and evaluation methods, which are respectively 1) trained on clean data and evaluated with random perturbations, 2) trained and evaluated with randomly perturbed data, and 3) trained with PGD-AT and evaluated with PGD-20 adversary. The budget of the random perturbations " + }, + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon = 8 / 255" + }, + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "type": "text", + "content": " is assigned the same as PGD, following our experimental setup detailed in Appendix A, so do other training and evaluation settings. Given that the conventional AT objective in Equation (1)" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4693" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "content": "encourages the model to produce a similar perception for benign and adversarial samples, we utilize the similarity in model perception between the corresponding benign and perturbed samples to estimate the learning sufficiency of the perturbed samples in AT. Specifically, the model perception is represented by the logits, and the similarity is measured through the mean squared error (MSE) score. We calculate the score respectively for success and failure cases of defense, which are divided according to the correctness of the final classification results of the perturbed samples. In addition, the model accuracy on clean and perturbed data is also recorded and illustrated together." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 67, + 225, + 282, + 381 + ], + "blocks": [ + { + "bbox": [ + 67, + 225, + 282, + 381 + ], + "lines": [ + { + "bbox": [ + 67, + 225, + 282, + 381 + ], + "spans": [ + { + "bbox": [ + 67, + 225, + 282, + 381 + ], + "type": "image", + "image_path": "4c76dec9f1f0d9dc7d8a43f98f770561211e18dd756e5efa172bbaae57571c0b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 392, + 295, + 535 + ], + "lines": [ + { + "bbox": [ + 55, + 392, + 295, + 535 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 295, + 535 + ], + "type": "text", + "content": "Figure 2. Proof-of-concept experiments for the different learning sufficiency in model perception for success and failure cases on CIFAR-10 with ResNet-18. In the x-axis, the \"Random\" for \"Training Algorithm\" means the training data is randomly perturbed within the budget " + }, + { + "bbox": [ + 55, + 392, + 295, + 535 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 392, + 295, + 535 + ], + "type": "text", + "content": ", while that for \"Perturbation Method\" indicates the attack used for evaluation on test data is random perturbation. The bars demonstrate the MSE score on logits between each pair of benign and perturbed samples, with the lines additionally showing the model accuracy on clean and perturbed data. Our results reveal a surprising phenomenon that, different from the clean training (with or without random noise), where failure cases are worse learned as we intuitively expected, AT already learns a better model perception for failure cases." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "Surprisingly, we experimentally found that the failure cases already learn better in model perception than the success cases. First, for the clean trained model, the perception on failure cases is significantly worse than on the success ones, just as we expected. Although the difference between success and failure cases has decreased for the randomly trained model, the failure cases remain worse, which suggests that non-adversarial perturbations are influential but not enough to change the trend. However, unlike the previous intuition, the model trained by PGD-AT can already perceive failure cases better than success ones. Our finding updates the current knowledge that, unlike those in clean training, the failure cases in AT do not fail simply because they are not sufficiently learned. On the contrary," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "the model even learns them harder to try to achieve the prediction and perception consistency. The smaller perception change under similar perturbations is direct evidence for this point of view." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 121, + 555, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 555, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 555, + 408 + ], + "type": "text", + "content": "Then, as a straightforward deduction, now that we have success cases with worse perception but failure cases with better perception, there should be a statistical difference between them in the maximum perception change they can tolerate while still maintaining the consistent prediction, just as demonstrated in Figure 1. In other words, the decision boundary is established relatively closer to the failure cases, which is instead the issue we should focus on, because such sample-level imbalance in boundary establishment can naturally complicate it and thus degrade the accuracy-robustness trade-off as mentioned in Section 2.2. In that case, the over-consistent model perception between benign and adversarial samples in the failure cases rather harms the final robustness, because this is expected to force the model to view the perturbations as noise and ignore the information within them, which should have been utilized to induce a smoother perception transition towards the decision boundary to support its establishment, thus impairing the ability of such cases to claim a reasonable robust distance during AT compared with success cases. Therefore, in this work, we argue that it is not insufficient but oversufficient learning of failure cases that contributes to the more complicated decision boundary and finally results in the current trade-off problem." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 422, + 421, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 421, + 436 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 421, + 436 + ], + "type": "text", + "content": "4. Proposed Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 443, + 527, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 443, + 527, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 527, + 455 + ], + "type": "text", + "content": "4.1. Robust Perception: A New AT Objective" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 461, + 553, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 553, + 532 + ], + "type": "text", + "content": "In Section 3, the too-small perception change under perturbations caused by the over-sufficient learning of failure cases has been identified as an important factor contributing to the current trade-off problem, which in turn questions the effectiveness of the conventional AT objective on such failure cases." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "Formally, under Equation (1) which implicitly optimize the perception distance " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\| h_{\\theta}(\\mathbf{x}^{\\prime}) - h_{\\theta}(\\mathbf{x})\\| \\to 0" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " between corresponding benign sample " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " and adversarial sample " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\prime}" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": ", the gradient of the conventional AT loss w.r.t. the perturbation, " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\nabla_{\\| \\mathbf{x}^{\\prime} - \\mathbf{x}\\|}\\mathcal{L}" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": ", would also approach 0. In that case, the model cannot further optimize the decision boundary using the perturbation information. Thus, a straightforward idea to mitigate this is to encourage smoother perception change towards the decision boundary along with the input perturbation, such that " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\| h_{\\theta}(\\mathbf{x}^{\\prime}) - h_{\\theta}(\\mathbf{x})\\| \\propto \\| \\mathbf{x}^{\\prime} - \\mathbf{x}\\|" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": ", as long as the final prediction remains unchanged. This is expected to retain an effective gradient signal in the direction of the perturbation, which enables the model to utilize this information to guide the decision boundary to a more reasonable location. Accordingly, in this section, we propose Robust" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4694" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 70, + 547, + 262 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 547, + 262 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 547, + 262 + ], + "type": "image", + "image_path": "e7a9880cec0e57341631f7022e1a3db81a4ef2c5b4bd8502106ff36fd843a4cf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 269, + 555, + 326 + ], + "lines": [ + { + "bbox": [ + 54, + 269, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 54, + 269, + 555, + 326 + ], + "type": "text", + "content": "Figure 3. Illustration of the proposed RPAT method. In addition to the conventional AT objective, which learns adversarial samples with the original supervised labels, RPAT introduces a novel regularization term to achieve the newly defined Robust Perception objective. Given benign sample " + }, + { + "bbox": [ + 54, + 269, + 555, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 54, + 269, + 555, + 326 + ], + "type": "text", + "content": " and adversarial samples " + }, + { + "bbox": [ + 54, + 269, + 555, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{x}' = \\mathbf{x} + \\Delta" + }, + { + "bbox": [ + 54, + 269, + 555, + 326 + ], + "type": "text", + "content": ", for any interpolation samples " + }, + { + "bbox": [ + 54, + 269, + 555, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{x} + \\lambda \\cdot \\Delta" + }, + { + "bbox": [ + 54, + 269, + 555, + 326 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 54, + 269, + 555, + 326 + ], + "type": "inline_equation", + "content": "\\lambda \\in [0,1]" + }, + { + "bbox": [ + 54, + 269, + 555, + 326 + ], + "type": "text", + "content": ", it encourages the ratio of the perception changes, " + }, + { + "bbox": [ + 54, + 269, + 555, + 326 + ], + "type": "inline_equation", + "content": "z_{\\theta}(\\mathbf{x} + \\lambda \\cdot \\Delta) - z_{\\theta}(\\mathbf{x})" + }, + { + "bbox": [ + 54, + 269, + 555, + 326 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 269, + 555, + 326 + ], + "type": "inline_equation", + "content": "z_{\\theta}(\\mathbf{x}') - z_{\\theta}(\\mathbf{x} + \\lambda \\cdot \\Delta)" + }, + { + "bbox": [ + 54, + 269, + 555, + 326 + ], + "type": "text", + "content": ", respectively from " + }, + { + "bbox": [ + 54, + 269, + 555, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 54, + 269, + 555, + 326 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 54, + 269, + 555, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{x} + \\lambda \\cdot \\Delta" + }, + { + "bbox": [ + 54, + 269, + 555, + 326 + ], + "type": "text", + "content": " and from " + }, + { + "bbox": [ + 54, + 269, + 555, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{x} + \\lambda \\cdot \\Delta" + }, + { + "bbox": [ + 54, + 269, + 555, + 326 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 54, + 269, + 555, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{x}'" + }, + { + "bbox": [ + 54, + 269, + 555, + 326 + ], + "type": "text", + "content": ", to approach the ratio indicated by " + }, + { + "bbox": [ + 54, + 269, + 555, + 326 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 54, + 269, + 555, + 326 + ], + "type": "text", + "content": ", such that the model perception would change smoothly along with the input perturbations from " + }, + { + "bbox": [ + 54, + 269, + 555, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 54, + 269, + 555, + 326 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 54, + 269, + 555, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{x}'" + }, + { + "bbox": [ + 54, + 269, + 555, + 326 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 346, + 295, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 295, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 295, + 369 + ], + "type": "text", + "content": "Perception as a new AT objective in addition to the conventional one to realize this idea." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 374, + 296, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 374, + 296, + 436 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 296, + 436 + ], + "type": "text", + "content": "Definition 1 (Robust Perception). Provided an AT task with target model " + }, + { + "bbox": [ + 55, + 374, + 296, + 436 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 374, + 296, + 436 + ], + "type": "text", + "content": ", let " + }, + { + "bbox": [ + 55, + 374, + 296, + 436 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 374, + 296, + 436 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 374, + 296, + 436 + ], + "type": "inline_equation", + "content": "\\mathbf{x}'" + }, + { + "bbox": [ + 55, + 374, + 296, + 436 + ], + "type": "text", + "content": " be any pairs of corresponding benign and adversarial samples with " + }, + { + "bbox": [ + 55, + 374, + 296, + 436 + ], + "type": "inline_equation", + "content": "\\Delta = \\mathbf{x}' - \\mathbf{x}" + }, + { + "bbox": [ + 55, + 374, + 296, + 436 + ], + "type": "text", + "content": ", for any hidden representation " + }, + { + "bbox": [ + 55, + 374, + 296, + 436 + ], + "type": "inline_equation", + "content": "h_{\\theta}(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 374, + 296, + 436 + ], + "type": "text", + "content": " denoting the model perception, our additional AT objective can be formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 444, + 296, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 444, + 296, + 459 + ], + "spans": [ + { + "bbox": [ + 55, + 444, + 296, + 459 + ], + "type": "interline_equation", + "content": "\\forall \\alpha \\in [ 0, 1 ], \\| h _ {\\pmb {\\theta}} (\\mathbf {x} + \\alpha \\cdot \\Delta) - h _ {\\pmb {\\theta}} (\\mathbf {x}) \\| = \\alpha \\cdot \\| h _ {\\pmb {\\theta}} (\\mathbf {x} ^ {\\prime}) - h _ {\\pmb {\\theta}} (\\mathbf {x}) \\|.", + "image_path": "9c9e9aac56b8c25aff909963355038fa34972207202419e07d3ac3cf66593700.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 468, + 296, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 296, + 576 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 296, + 576 + ], + "type": "text", + "content": "To facilitate the understanding of the proposed Robust Perception, in addition to the most straightforward perspective that it encourages the change in model perception along with the adversarial perturbations, the effectiveness of this new objective can also be demonstrated from other two perspectives, local linearity [12] and Lipschitz regularization [1, 11, 21, 33], both of which also help smooth the decision boundary and mitigate the current trade-off problem. Below, we further detail them theoretically." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 580, + 297, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 297, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 297, + 616 + ], + "type": "text", + "content": "Theorem 1. Let " + }, + { + "bbox": [ + 55, + 580, + 297, + 616 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 580, + 297, + 616 + ], + "type": "text", + "content": " be the Hessian Matrix such that " + }, + { + "bbox": [ + 55, + 580, + 297, + 616 + ], + "type": "inline_equation", + "content": "H_{h_\\theta}(\\mathbf{x}) = \\nabla_{\\mathbf{x}}^2 h_\\theta(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 580, + 297, + 616 + ], + "type": "text", + "content": ", then with the new optimization objective of Robust Perception, we have:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 624, + 234, + 639 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 624, + 234, + 639 + ], + "spans": [ + { + "bbox": [ + 116, + 624, + 234, + 639 + ], + "type": "interline_equation", + "content": "\\forall \\Delta , \\Delta^ {\\top} \\cdot H _ {h _ {\\theta}} (\\mathbf {x}) \\cdot \\Delta \\rightarrow 0.", + "image_path": "a07e79021ab39c131d5617013871ae88ed181c0a41d7bc9d827265954aa42d1b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 648, + 233, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 648, + 233, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 648, + 233, + 661 + ], + "type": "text", + "content": "Proof. We defer the proof to Appendix B.1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "This means Robust Perception limits the second-order (and higher-order as well) nonlinear effects within the adversarial perturbation to the model perception, such that the perception would change mainly along with the linear" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 346, + 555, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 555, + 394 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 555, + 394 + ], + "type": "text", + "content": "term of the perturbation, and thus be smoother. This also aligns with the local linearity hypothesis of adversarial samples [12], which assumes that linear approximations dominate the model behavior under adversarial perturbations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 398, + 556, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 398, + 556, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 556, + 434 + ], + "type": "text", + "content": "Theorem 2. Let " + }, + { + "bbox": [ + 313, + 398, + 556, + 434 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 313, + 398, + 556, + 434 + ], + "type": "text", + "content": " be the Jacobian Matrix such that " + }, + { + "bbox": [ + 313, + 398, + 556, + 434 + ], + "type": "inline_equation", + "content": "J_{h_{\\theta}}(\\mathbf{x}) = \\nabla_{\\mathbf{x}} h_{\\theta}(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 398, + 556, + 434 + ], + "type": "text", + "content": ", then with the new optimization objective of Robust Perception, we have:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 350, + 444, + 517, + 458 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 444, + 517, + 458 + ], + "spans": [ + { + "bbox": [ + 350, + 444, + 517, + 458 + ], + "type": "interline_equation", + "content": "\\forall \\alpha \\in [ 0, 1 ], J _ {h _ {\\boldsymbol {\\theta}}} (\\mathbf {x} + \\boldsymbol {\\alpha} \\cdot \\Delta) \\rightarrow J _ {h _ {\\boldsymbol {\\theta}}} (\\mathbf {x}),", + "image_path": "3d7e918afc4fbdd109e48fb01f485bdd30f202d518d626162d289ed377710bb6.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 467, + 555, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 555, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 555, + 515 + ], + "type": "text", + "content": "then with " + }, + { + "bbox": [ + 313, + 467, + 555, + 515 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_{\\mathrm{spec}}" + }, + { + "bbox": [ + 313, + 467, + 555, + 515 + ], + "type": "text", + "content": " denoting the Spectral Norm and " + }, + { + "bbox": [ + 313, + 467, + 555, + 515 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 467, + 555, + 515 + ], + "type": "text", + "content": " being any micro value, given " + }, + { + "bbox": [ + 313, + 467, + 555, + 515 + ], + "type": "inline_equation", + "content": "\\| J_{h_\\theta}(\\mathbf{x} + \\alpha \\cdot \\Delta) - J_{h_\\theta}(\\mathbf{x})\\|_{\\mathrm{spec}} \\leq \\gamma" + }, + { + "bbox": [ + 313, + 467, + 555, + 515 + ], + "type": "text", + "content": ", the function " + }, + { + "bbox": [ + 313, + 467, + 555, + 515 + ], + "type": "inline_equation", + "content": "h_\\theta(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 467, + 555, + 515 + ], + "type": "text", + "content": " can be referred to as " + }, + { + "bbox": [ + 313, + 467, + 555, + 515 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 467, + 555, + 515 + ], + "type": "text", + "content": "-Lipschitz with the Lipschitz constant " + }, + { + "bbox": [ + 313, + 467, + 555, + 515 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 467, + 555, + 515 + ], + "type": "text", + "content": " upper-bounded by:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 373, + 524, + 494, + 543 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 524, + 494, + 543 + ], + "spans": [ + { + "bbox": [ + 373, + 524, + 494, + 543 + ], + "type": "interline_equation", + "content": "K \\leq \\sup _ {\\mathbf {x}} \\| J _ {h _ {\\theta}} (\\mathbf {x}) \\| _ {\\operatorname {s p e c}} + \\gamma .", + "image_path": "fbf489b27f23a0f5cabda9b7892df688dfa80cf115fc364b2048fb465fc9b6cd.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 552, + 492, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 492, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 492, + 564 + ], + "type": "text", + "content": "Proof. We defer the proof to Appendix B.2." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 568, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 554, + 653 + ], + "type": "text", + "content": "This theorem tells us that Robust Perception also regularizes the change " + }, + { + "bbox": [ + 313, + 568, + 554, + 653 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 568, + 554, + 653 + ], + "type": "text", + "content": " in Jacobian along with the adversarial perturbation, which can be interpreted as limiting the increase of the global Lipschitz constant by " + }, + { + "bbox": [ + 313, + 568, + 554, + 653 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 568, + 554, + 653 + ], + "type": "text", + "content": " under perturbation. Previous works have demonstrated that this is expected to help achieve smoother decision boundaries and improve the accuracy-robustness trade-off [1, 11, 21, 33]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 659, + 525, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 525, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 525, + 673 + ], + "type": "text", + "content": "4.2. Robust Perception Adversarial Training" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Based on the above considerations, in this section, we propose a novel AT method named Robust Perception Adversarial Training (RPAT). It utilizes a new additional" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4695" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 194 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 194 + ], + "type": "text", + "content": "regularization term appended to the conventional AT objective to achieve the Robust Perception. Formally, given a " + }, + { + "bbox": [ + 55, + 72, + 296, + 194 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 72, + 296, + 194 + ], + "type": "text", + "content": "-class classification task " + }, + { + "bbox": [ + 55, + 72, + 296, + 194 + ], + "type": "inline_equation", + "content": "(C \\geq 2)" + }, + { + "bbox": [ + 55, + 72, + 296, + 194 + ], + "type": "text", + "content": " and a dataset " + }, + { + "bbox": [ + 55, + 72, + 296, + 194 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(\\mathbf{x}_i, y_i)\\}_{i=1,\\dots,n}" + }, + { + "bbox": [ + 55, + 72, + 296, + 194 + ], + "type": "text", + "content": " with benign sample " + }, + { + "bbox": [ + 55, + 72, + 296, + 194 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 72, + 296, + 194 + ], + "type": "text", + "content": " and supervised label " + }, + { + "bbox": [ + 55, + 72, + 296, + 194 + ], + "type": "inline_equation", + "content": "y_i \\in \\{1,\\dots,C\\}" + }, + { + "bbox": [ + 55, + 72, + 296, + 194 + ], + "type": "text", + "content": ", let " + }, + { + "bbox": [ + 55, + 72, + 296, + 194 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_i'" + }, + { + "bbox": [ + 55, + 72, + 296, + 194 + ], + "type": "text", + "content": " denotes the adversarial sample corresponding to " + }, + { + "bbox": [ + 55, + 72, + 296, + 194 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 55, + 72, + 296, + 194 + ], + "type": "text", + "content": ", which is generated as " + }, + { + "bbox": [ + 55, + 72, + 296, + 194 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_i' = \\arg \\max_{\\|\\mathbf{x}_i' - \\mathbf{x}_i\\|_p \\leq \\epsilon} \\mathbb{1}(\\arg \\max_{k=1,\\dots,C} h_\\theta(\\mathbf{x}_i'))^{(k)} \\neq y_i)" + }, + { + "bbox": [ + 55, + 72, + 296, + 194 + ], + "type": "text", + "content": ", then based on the Definition 1 in Section 4.1, the adversarial risk to be minimized for the proposed RPAT w.r.t. " + }, + { + "bbox": [ + 55, + 72, + 296, + 194 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 55, + 72, + 296, + 194 + ], + "type": "text", + "content": " can be formulated with the 0-1 loss [58] as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 201, + 294, + 243 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 201, + 294, + 243 + ], + "spans": [ + { + "bbox": [ + 61, + 201, + 294, + 243 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {R} (\\boldsymbol {\\theta}, \\mathbf {x} _ {i}) := \\mathbb {1} \\left(f _ {\\boldsymbol {\\theta}} \\left(\\hat {\\mathbf {x}} _ {i} ^ {\\prime}\\right) \\neq y _ {i}\\right) \\\\ + \\mathbb {1} \\left(\\frac {h _ {\\boldsymbol {\\theta}} \\left(\\tilde {\\mathbf {x}} _ {i}\\right) - h _ {\\boldsymbol {\\theta}} \\left(\\mathbf {x} _ {i}\\right)}{\\alpha} \\neq \\frac {h _ {\\boldsymbol {\\theta}} \\left(\\hat {\\mathbf {x}} _ {i} ^ {\\prime}\\right) - h _ {\\boldsymbol {\\theta}} \\left(\\tilde {\\mathbf {x}} _ {i}\\right)}{1 - \\alpha}\\right), \\tag {2} \\\\ \\end{array}", + "image_path": "ba6ade6712db3a505822286c85e9751b669909f76efcde5ff6d13da06509fad2.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 251, + 185, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 185, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 185, + 264 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 251, + 185, + 264 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_i = \\mathbf{x}_i + \\alpha \\cdot (\\hat{\\mathbf{x}}_i' - \\mathbf{x}_i)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 264, + 295, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 264, + 295, + 395 + ], + "spans": [ + { + "bbox": [ + 54, + 264, + 295, + 395 + ], + "type": "text", + "content": "While the proposed Robust Perception aims to improve learning efficacy for the failure cases, the above risk does not explicitly distinguish success and failure cases. This is because Robust Perception is designed to mitigate the gap of the failure cases in the effectiveness of learning compared with the success ones, rather than require any new characteristic beyond the success cases. In other words, the success cases are expected to also satisfy Robust Perception at the beginning. Therefore, considering the computational efficiency, it is not necessary to specifically exclude the success cases for the Robust Perception term." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 396, + 296, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 296, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 296, + 598 + ], + "type": "text", + "content": "Finally, by introducing practical surrogate losses to optimize the adversarial risk in Equation (2), we end up with our new RPAT method. To be specific, for the supervised learning term for adversarial samples, we adopt cross-entropy (CE) just as the conventional AT by default. Then, for the Robust Perception term, as it is formulated in a format of similarity regularization, various criterion that measures the similarity distance between the two input items are feasible options, such as mean squared error (MSE), Kullback-Leibler divergence (KL), Jensen-Shannon divergence (JS), and Cosine similarity. We simply adopt MSE loss for the basic version of RPAT here, and would also demonstrate the effectiveness of KL loss later in Section 5.3. With the logits " + }, + { + "bbox": [ + 55, + 396, + 296, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{z}(\\mathbf{x}_i,\\boldsymbol {\\theta})" + }, + { + "bbox": [ + 55, + 396, + 296, + 598 + ], + "type": "text", + "content": " utilized as the hidden representation " + }, + { + "bbox": [ + 55, + 396, + 296, + 598 + ], + "type": "inline_equation", + "content": "h_\\theta (\\cdot)" + }, + { + "bbox": [ + 55, + 396, + 296, + 598 + ], + "type": "text", + "content": " to denote the model perception, and " + }, + { + "bbox": [ + 55, + 396, + 296, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{p}(\\mathbf{x}_i,\\boldsymbol {\\theta})" + }, + { + "bbox": [ + 55, + 396, + 296, + 598 + ], + "type": "text", + "content": " being the softmax output as given in Equation (1), the overall objective of RPAT can be formulated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 607, + 295, + 657 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 607, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 61, + 607, + 295, + 657 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} ^ {R P A T} (\\boldsymbol {\\theta}, \\mathcal {D}, \\lambda , \\alpha) := \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left(\\mathcal {L} ^ {C E} \\left(\\mathbf {p} \\left(\\hat {\\mathbf {x}} _ {i} ^ {\\prime}, \\boldsymbol {\\theta}\\right), y _ {i}\\right) \\right. \\tag {3} \\\\ + \\lambda \\cdot \\mathcal {L} ^ {M S E} \\left( \\right.\\frac {\\mathbf {z} (\\tilde {\\mathbf {x}} _ {i} , \\boldsymbol {\\theta}) - \\mathbf {z} (\\mathbf {x} _ {i} , \\boldsymbol {\\theta})}{\\alpha} \\left| \\right.\\left| \\right. \\frac {\\mathbf {z} (\\hat {\\mathbf {x}} _ {i} ^ {\\prime} , \\boldsymbol {\\theta}) - \\mathbf {z} (\\tilde {\\mathbf {x}} _ {i} , \\boldsymbol {\\theta})}{1 - \\alpha}\\left. \\right)\\left. \\right), \\\\ \\end{array}", + "image_path": "0b39296b402709cb77050da205aa55ace2924f3076595584a9845dc6cbf3e6ee.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": " is a weight hyper-parameter to balance the conventional AT objective and our new Robust Perception objective. An intuitive illustration of the proposed RPAT method is provided in Figure 3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 72, + 395, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 395, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 395, + 85 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 91, + 555, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 295 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 295 + ], + "type": "text", + "content": "Following previous AT works, the effectiveness of the proposed RPAT is evaluated on CIFAR-10, CIFAR-100 [22] and Tiny-ImageNet [23] datasets. The main experiments can be divided into two parts. First, based on Tack et al. [40], we implement RPAT upon three common AT baselines, namely PGD-AT, TRADES, and MART, as well an advanced method, Consistency-AT, with ResNet-18 [15] architecture. This is to show the general effectiveness of the proposed Robust Perception in improving the current trade-off. Second, by integrating RPAT into one of the current SOTA methods, ReBAT [47], with PreActResNet-18 [16] and WideResNet-34-10 [56] architectures, we achieve new SOTA performance beyond 12 SOTA methods on the accuracy-robustness trade-off problem. This is to demonstrate the advancement achieved by the proposed RPAT in the AT area. The detailed experimental setup is provided in Appendix A." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 304, + 436, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 436, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 436, + 316 + ], + "type": "text", + "content": "5.1. Evaluation Measures" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 322, + 555, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 322, + 555, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 322, + 555, + 477 + ], + "type": "text", + "content": "There are five measures in total adopted to evaluate the experimental methods from three different perspectives. Firstly, we report clean accuracy on benign test data, which reflects the ability of the robust models to maintain performance in benign cases. Secondly, we evaluate adversarial robustness under two adversarial attack methods, namely PGD-20 [29] and Auto-Attack (AA) [7], with default settings aligning with Appendix A and random starts. Finally, to test the performance w.r.t. the current accuracy-robustness trade-off problem in AT, on the one hand, we report the mean of the clean accuracy and the robust score under Auto-Attack acquired above, and on the other hand, we introduce the Natural-Robustness Ratio (NRR) [13]:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 319, + 485, + 548, + 511 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 485, + 548, + 511 + ], + "spans": [ + { + "bbox": [ + 319, + 485, + 548, + 511 + ], + "type": "interline_equation", + "content": "N R R = \\frac {2 \\times C l e a n A c c u r a c y \\times A d v e r s a r i a l R o b u s t n e s s}{C l e a n A c c u r a c y + A d v e r s a r i a l R o b u s t n e s s},", + "image_path": "a283da45ea37ecd2f86af480e82c78ee9a33b6f0a8b74baf1b498c88e499c1cf.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 519, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 519, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 519, + 555, + 616 + ], + "type": "text", + "content": "which emphasizes how well the robust model strikes a balance between clean accuracy and robustness. A lower NRR value suggests that the model might prioritize one side over the other. All the results reported are averages of three runs, with the \"best\" performance of each run recorded on the best checkpoint achieving the highest PGD-20 accuracy, which ensures that the training process is isolated from AA, the test-time adversary to evaluate the robustness." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 624, + 508, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 508, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 508, + 636 + ], + "type": "text", + "content": "5.2. Effectiveness in Improving Baselines" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "content": "The proposed RPAT can be easily implemented with existing AT methods by simply appending the Robust Perception term. In this section, we illustrate the improvements that RPAT can bring to our four experimental benchmarks, namely PGD-AT, TRADES, MART, and Consistency-AT, on CIFAR-10, CIFAR-100, and Tiny-ImageNet with" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4696" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 91, + 553, + 251 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 550, + 82 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 550, + 82 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 550, + 82 + ], + "type": "text", + "content": "Table 1. Improvements that the proposed RPAT achieved upon four benchmarks. The results are acquired with ResNet-18 and " + }, + { + "bbox": [ + 59, + 71, + 550, + 82 + ], + "type": "inline_equation", + "content": "{\\ell }_{\\infty }" + }, + { + "bbox": [ + 59, + 71, + 550, + 82 + ], + "type": "text", + "content": " norm." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 91, + 553, + 251 + ], + "lines": [ + { + "bbox": [ + 58, + 91, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 58, + 91, + 553, + 251 + ], + "type": "table", + "html": "
DatasetMethodBenchmarks+RPAT (Ours)
CleanPGD-20AAMeanNRRCleanPGD-20AAMeanNRR
bestbestbestbestfinalbestfinalbestfinal
CIFAR-10PGD-AT82.9250.6146.7464.83059.78283.2082.9951.2950.9748.0047.6065.60060.878
TRADES79.6751.9147.6263.64559.61080.0280.0251.2151.2147.4747.4563.74559.590
MART77.9353.0046.7062.31558.40275.4475.9552.8852.8747.0347.3061.23557.940
Consistency-AT83.4251.9647.7265.57060.71184.1284.0452.3352.1448.9849.0666.55061.911
CIFAR-100PGD-AT56.5628.8025.0240.79034.69358.2258.5829.1628.5724.8824.2341.55034.862
TRADES55.3929.3624.5139.95033.98357.5057.2529.4229.0025.0524.4541.27534.897
MART49.8330.3825.0037.41533.29550.7250.7230.3330.3325.3425.0638.03033.796
Consistency-AT58.5329.2825.3941.96035.41760.3360.5129.9729.6426.3125.9743.32036.641
Tiny-ImageNetPGD-AT46.3221.5817.0731.69524.94747.6848.3222.1322.0017.7717.9132.72525.891
TRADES46.7521.5216.6031.67524.50048.7748.8521.9821.9316.9216.7132.84525.124
MART39.7022.9017.1828.44023.98241.7641.6623.4023.2917.7917.7829.77524.951
Consistency-AT46.5421.8917.6032.07025.54149.7449.8523.1622.8418.8418.7434.29027.329
", + "image_path": "c3e6018b054b8a83a43decac3987908935d105e7e267d5588c66ea58ba3925d4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 58, + 281, + 552, + 441 + ], + "blocks": [ + { + "bbox": [ + 60, + 261, + 549, + 272 + ], + "lines": [ + { + "bbox": [ + 60, + 261, + 549, + 272 + ], + "spans": [ + { + "bbox": [ + 60, + 261, + 549, + 272 + ], + "type": "text", + "content": "Table 2. Improvements that the proposed RPAT achieved upon four benchmarks. The results are acquired with ResNet-18 and " + }, + { + "bbox": [ + 60, + 261, + 549, + 272 + ], + "type": "inline_equation", + "content": "{\\ell }_{2}" + }, + { + "bbox": [ + 60, + 261, + 549, + 272 + ], + "type": "text", + "content": " norm." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 281, + 552, + 441 + ], + "lines": [ + { + "bbox": [ + 58, + 281, + 552, + 441 + ], + "spans": [ + { + "bbox": [ + 58, + 281, + 552, + 441 + ], + "type": "table", + "html": "
DatasetMethodBenchmarks+RPAT (Ours)
CleanPGD-20AAMeanNRRCleanPGD-20AAMeanNRR
bestbestbestbestfinalbestfinalbestfinal
CIFAR-10PGD-AT87.7667.9266.3677.06075.57488.2088.4568.5668.5267.6367.6277.91576.557
TRADES83.9968.6065.9374.96073.87285.1784.8068.9568.9067.6767.3776.42075.418
MART84.0968.3266.2875.18574.13084.6584.6568.6068.6066.6966.7075.67074.604
Consistency-AT88.7669.3567.4678.11076.65889.3889.5270.4170.3669.4469.5579.41078.158
CIFAR-100PGD-AT65.0041.5439.2752.13548.96065.1465.4441.6141.2839.2339.2052.18548.969
TRADES61.2543.0540.1550.7048.50562.4162.6443.5643.1040.3640.1251.38549.020
MART60.0843.4239.8549.96547.9260.5760.5743.9243.9040.3640.3650.46548.442
Consistency-AT65.1442.2839.9252.53049.50365.5466.8142.8142.5140.2640.6152.90049.880
Tiny-ImageNetPGD-AT59.3643.1240.9050.13048.43160.2860.2844.0644.0641.8941.8751.08549.430
TRADES57.6144.7042.3349.97048.80258.4658.5044.8744.6542.4242.4450.44049.165
MART56.4744.8442.2249.34548.31657.1756.9545.0244.6742.5142.1649.84048.762
Consistency-AT61.4944.9242.5952.04050.32462.1962.6445.2844.8743.1742.8552.68050.963
", + "image_path": "40a2fbae0b216eace616d02a012d3508daf7058c29acafee77646090d21a6e77.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 460, + 294, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 294, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 294, + 664 + ], + "type": "text", + "content": "ResNet-18, demonstrating the wide effectiveness of our method. The results under " + }, + { + "bbox": [ + 55, + 460, + 294, + 664 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}" + }, + { + "bbox": [ + 55, + 460, + 294, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 460, + 294, + 664 + ], + "type": "inline_equation", + "content": "\\ell_{2}" + }, + { + "bbox": [ + 55, + 460, + 294, + 664 + ], + "type": "text", + "content": " threat models are shown in Table 1 and Table 2, respectively, where the best clean accuracy, AA robustness, as well mean and NRR for their trade-off on each of the datasets are marked in bold, while individual scores not improved are marked with underline. Note that PGD-20 is the indicator for selecting the \"best\" checkpoint as explained in Section 5.1, and the \"final\" scores are additionally provided to show that RPAT does not suffer from serious robust-overfitting, which is also evidence for appropriate algorithm design. Thus, these fields are not involved in the comparison. It can be found that RPAT achieves significant improvement for most scores among the tables, where Consistency-AT+RPAT is basically the most competitive for all the experimental groups, which demonstrates the general effectiveness of our method in mitigating the current trade-off problem." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 672, + 228, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 228, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 228, + 684 + ], + "type": "text", + "content": "5.3. Advancement beyond 12 SOTAs" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "To demonstrate the advancement that the proposed RPAT contributes, we further integrate RPAT into ReBAT [47]," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 460, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 713 + ], + "type": "text", + "content": "one of the current SOTA AT methods on the accuracy-robustness trade-off problem. It implements WA [18], averaging multiple points along the SGD trajectory for better generalization, through the exponential moving average (EMA) [37] strategy with a decay rate of 0.999. Thus, we also involve WA itself (i.e., with PGD-AT), as well as " + }, + { + "bbox": [ + 313, + 460, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{KD + SWA}" + }, + { + "bbox": [ + 313, + 460, + 555, + 713 + ], + "type": "text", + "content": " [5] which shares a similar idea, for comparison. Other differences in ReBAT include " + }, + { + "bbox": [ + 313, + 460, + 555, + 713 + ], + "type": "inline_equation", + "content": "KL" + }, + { + "bbox": [ + 313, + 460, + 555, + 713 + ], + "type": "text", + "content": " loss and a reduced learning rate decay factor from 0.1 to 0.5. We refer to our method upon ReBAT as " + }, + { + "bbox": [ + 313, + 460, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{RPAT}^{++}" + }, + { + "bbox": [ + 313, + 460, + 555, + 713 + ], + "type": "text", + "content": ", and compare it with 12 previous SOTAs introduced in Section 2.2, respectively with PreActResNet-18 and WideResNet-34-10 as shown in Table 3 and Table 4. The clean and AA scores of the SOTAs are from the best checkpoint among their original papers, Kanai et al. [19], Wang et al. [47], and our reproduction if any (so due to the lack of previous record, ADR and CURE are excluded for PreActResNet-18). Beyond the basic CIFAR-10 results under " + }, + { + "bbox": [ + 313, + 460, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}" + }, + { + "bbox": [ + 313, + 460, + 555, + 713 + ], + "type": "text", + "content": " norm, Table 3 involves CIFAR-100 and " + }, + { + "bbox": [ + 313, + 460, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 313, + 460, + 555, + 713 + ], + "type": "text", + "content": " norm, and Table 4 further consider the Cutmix [55] data augmentation strategy as marked by “*”, which ensures the comprehensiveness of our comparison." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4697" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 101, + 364, + 262 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 357, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 357, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 357, + 94 + ], + "type": "text", + "content": "Table 3. Comparison of the proposed " + }, + { + "bbox": [ + 55, + 71, + 357, + 94 + ], + "type": "inline_equation", + "content": "\\mathrm{RPAT^{+ + }}" + }, + { + "bbox": [ + 55, + 71, + 357, + 94 + ], + "type": "text", + "content": " with current SOTAs on accuracy-robustness trade-off with PreActResNet-18. The best results are marked in bold." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 101, + 364, + 262 + ], + "lines": [ + { + "bbox": [ + 58, + 101, + 364, + 262 + ], + "spans": [ + { + "bbox": [ + 58, + 101, + 364, + 262 + ], + "type": "table", + "html": "
NormMethodCIFAR-10CIFAR-100
CleanAAMeanNRRCleanAAMeanNRR
\\( \\ell_{\\infty} \\)WAUAI'1883.5049.8966.69562.46157.2625.8341.54535.601
MMAICLR'2085.5037.2061.35051.84460.6018.4039.50028.229
AWPNeurIPS'2081.1150.0965.60061.93354.1025.1639.63034.347
GAIRATICLR'2178.7037.7058.20050.97952.0019.8035.90028.680
KD+SWAICLR'2184.0649.8266.94062.56257.1725.6641.41535.422
EWATICML'2182.8048.2065.50060.93154.2023.5238.86032.805
MAILNeurIPS'2179.5039.6059.55052.86746.5016.7031.60024.574
TEICLR'2282.0450.1266.08062.22556.4125.8441.12535.444
SOVRICML'2381.9049.4065.65061.62852.1024.3038.20033.142
ReBATNeurIPS'2382.0950.7266.40562.70056.1327.6041.86537.004
RPAT++Ours82.6351.0066.81563.07256.8427.6842.26037.230
\\( \\ell_{2} \\)ReBATNeurIPS'2388.7971.0079.89578.90565.5842.6754.12551.701
RPAT++Ours89.0671.2680.16079.17265.6342.8554.24051.848
", + "image_path": "8e6de9a6b1bcf620105e396f5a7a89d55e46a412280ed19cbd3c9a8ee1a8182c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 372, + 102, + 550, + 262 + ], + "blocks": [ + { + "bbox": [ + 369, + 71, + 552, + 93 + ], + "lines": [ + { + "bbox": [ + 369, + 71, + 552, + 93 + ], + "spans": [ + { + "bbox": [ + 369, + 71, + 552, + 93 + ], + "type": "text", + "content": "Table 4. Comparison of " + }, + { + "bbox": [ + 369, + 71, + 552, + 93 + ], + "type": "inline_equation", + "content": "{\\mathrm{{RPAT}}}^{+ + }" + }, + { + "bbox": [ + 369, + 71, + 552, + 93 + ], + "type": "text", + "content": " with SOTAs on CIFAR-10 with WideResNet-34-10 and " + }, + { + "bbox": [ + 369, + 71, + 552, + 93 + ], + "type": "inline_equation", + "content": "{\\ell }_{\\infty }" + }, + { + "bbox": [ + 369, + 71, + 552, + 93 + ], + "type": "text", + "content": " norm." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 372, + 102, + 550, + 262 + ], + "lines": [ + { + "bbox": [ + 372, + 102, + 550, + 262 + ], + "spans": [ + { + "bbox": [ + 372, + 102, + 550, + 262 + ], + "type": "table", + "html": "
MethodCleanAAMeanNRR
WAUAI'1887.6652.6570.15565.787
MMAICLR'2087.8043.1065.45057.818
AWPNeurIPS'2085.6353.3269.47565.718
GAIRATICLR'2183.0041.8062.40055.599
KD+SWAICLR'2187.4553.5970.52066.456
EWATICML'2186.0051.6068.80064.500
MAILNeurIPS'2182.2043.3062.75056.721
TEICLR'2285.9752.8869.42565.482
SOVRICML'2385.0053.1069.05065.366
ReBATNeurIPS'2385.2554.7870.01566.700
ADRICLR'2484.6753.2568.96065.381
CUREICLR'2487.0552.1069.57565.186
\\(RPAT^{++}\\)Ours86.7654.9770.86567.300
ReBAT*NeurIPS'2386.6655.6471.15067.769
\\(RPAT^{++*}\\)Ours87.5755.7971.68068.158
", + "image_path": "1ca8e9d1d96a6a2979fa47d40176a820aea878f852909e21e84c5a72b163e855.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 59, + 280, + 169, + 366 + ], + "blocks": [ + { + "bbox": [ + 59, + 280, + 169, + 366 + ], + "lines": [ + { + "bbox": [ + 59, + 280, + 169, + 366 + ], + "spans": [ + { + "bbox": [ + 59, + 280, + 169, + 366 + ], + "type": "image", + "image_path": "feec7bf49567c597de2bfb94966fcd39f9b246c998a5ab54408b2f62b07b554c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 369, + 162, + 380 + ], + "lines": [ + { + "bbox": [ + 64, + 369, + 162, + 380 + ], + "spans": [ + { + "bbox": [ + 64, + 369, + 162, + 380 + ], + "type": "text", + "content": "(a) Impact of loss function." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 169, + 280, + 293, + 367 + ], + "blocks": [ + { + "bbox": [ + 169, + 280, + 293, + 367 + ], + "lines": [ + { + "bbox": [ + 169, + 280, + 293, + 367 + ], + "spans": [ + { + "bbox": [ + 169, + 280, + 293, + 367 + ], + "type": "image", + "image_path": "a1d5beb1720d50f0b83416088093087387fc3e74d0ea117a517ad85dff5164c0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 171, + 369, + 290, + 380 + ], + "lines": [ + { + "bbox": [ + 171, + 369, + 290, + 380 + ], + "spans": [ + { + "bbox": [ + 171, + 369, + 290, + 380 + ], + "type": "text", + "content": "(b) Impact of hyper-parameter " + }, + { + "bbox": [ + 171, + 369, + 290, + 380 + ], + "type": "inline_equation", + "content": "\\alpha" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 389, + 296, + 434 + ], + "lines": [ + { + "bbox": [ + 55, + 389, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 296, + 434 + ], + "type": "text", + "content": "Figure 4. Two ablation studies of the proposed RPAT on CIFAR-10 with ResNet-18 under " + }, + { + "bbox": [ + 55, + 389, + 296, + 434 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}" + }, + { + "bbox": [ + 55, + 389, + 296, + 434 + ], + "type": "text", + "content": " norm. The MSE loss in (a) and the points with " + }, + { + "bbox": [ + 55, + 389, + 296, + 434 + ], + "type": "inline_equation", + "content": "\\alpha = 0.5" + }, + { + "bbox": [ + 55, + 389, + 296, + 434 + ], + "type": "text", + "content": " in (b) correspond to the record of \"PGD-AT+RPAT\" on CIFAR-10 in Table 1." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 456, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 296, + 506 + ], + "type": "text", + "content": "The results show that the proposed " + }, + { + "bbox": [ + 55, + 456, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\mathrm{RPAT^{+ + }}" + }, + { + "bbox": [ + 55, + 456, + 296, + 506 + ], + "type": "text", + "content": " achieves new SOTA performance for most of the measures beyond previous SOTAs on the accuracy-robustness trade-off problem, which strongly supports the contribution of this work." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 514, + 149, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 149, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 149, + 528 + ], + "type": "text", + "content": "5.4. Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 533, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 676 + ], + "type": "text", + "content": "As the proposed RPAT can be implemented upon existing AT methods by simply appending the Robust Perception term into their original objectives, the comparison in performance between and after adding RPAT in Section 5.2 and Section 5.3 have achieved a comprehensive ablation study on the effectiveness of RPAT. Therefore, in this section, we mainly investigate the impact of two setting details, respectively the loss function adopted for the Robust Perception term (Figure 4 (a)) and the specific value for the hyperparameter " + }, + { + "bbox": [ + 55, + 533, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 533, + 295, + 676 + ], + "type": "text", + "content": " (Figure 4 (b)). Besides, we further discuss an optional trick suggested by ReBAT but excluded in the above experiments in Appendix 5.3." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Specifically, just as we expected, different loss functions to calculate the similarity between model perceptions, including MSE, KL, JS and Cosine, would not significantly" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 281, + 555, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 461 + ], + "type": "text", + "content": "impact the final performance of RPAT, which suggests that the effectiveness of our method is from the successful conceptual design of Robust Perception, instead of relying on specific implementation of the regularization term. Then, for " + }, + { + "bbox": [ + 313, + 281, + 555, + 461 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 281, + 555, + 461 + ], + "type": "text", + "content": ", we try 0.2 and 0.8 besides the default value 0.5 on the one hand, and also introduce beta distribution " + }, + { + "bbox": [ + 313, + 281, + 555, + 461 + ], + "type": "inline_equation", + "content": "\\beta \\sim Beta(0.75, 0.75)" + }, + { + "bbox": [ + 313, + 281, + 555, + 461 + ], + "type": "text", + "content": " as suggested by Zhang et al. [57] on the other hand. Specifically, we try " + }, + { + "bbox": [ + 313, + 281, + 555, + 461 + ], + "type": "inline_equation", + "content": "\\alpha = \\min(\\beta, (1 - \\beta))" + }, + { + "bbox": [ + 313, + 281, + 555, + 461 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 281, + 555, + 461 + ], + "type": "inline_equation", + "content": "\\alpha = \\max(\\beta, (1 - \\beta))" + }, + { + "bbox": [ + 313, + 281, + 555, + 461 + ], + "type": "text", + "content": ", namely indicated as \"Beta-\" and \"Beta+\" in (Figure 4 (b)), as they respectively approach adversarial and benign samples more. The results show that the default " + }, + { + "bbox": [ + 313, + 281, + 555, + 461 + ], + "type": "inline_equation", + "content": "\\alpha = 0.5" + }, + { + "bbox": [ + 313, + 281, + 555, + 461 + ], + "type": "text", + "content": " is an appropriate choice, which is reasonable because the midpoint is expected to be more representative for the ideal target that Robust Perception holds for all " + }, + { + "bbox": [ + 313, + 281, + 555, + 461 + ], + "type": "inline_equation", + "content": "\\alpha \\in [0, 1]" + }, + { + "bbox": [ + 313, + 281, + 555, + 461 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 476, + 388, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 476, + 388, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 476, + 388, + 489 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": "This work studies the current accuracy-robustness trade-off problem in AT. To the best of our knowledge, this is the first work that explicitly attributes the trade-off problem to the over-sufficient (rather than insufficient) learning of hard adversarial samples. We surprisingly reveal that such hard samples are already learned better than those successfully defended in model perception, which means the conventional AT objective is no longer effective to further optimize the model w.r.t. these hard samples. Instead, it forces the model to ignore the perturbations, in turn degrading their effectiveness in supporting the establishment of decision boundary to an appropriate location. To mitigate this, our new AT objective, Robust Perception, encourages smoother perception transition towards the decision boundary along with perturbations, based on which the proposed RPAT method achieves new SOTA performance. Still, exploring better methods to approximate this ideal objective would be a promising future direction." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4698" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 137 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 137 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China (No. 62471420), GuangDong Basic and Applied Basic Research Foundation (2025A15150122 96), and CCF-Tencent Rhino-Bird Open Research Fund." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 155, + 115, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 155, + 115, + 167 + ], + "spans": [ + { + "bbox": [ + 56, + 155, + 115, + 167 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 175, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 61, + 175, + 296, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 175, + 296, + 219 + ], + "spans": [ + { + "bbox": [ + 61, + 175, + 296, + 219 + ], + "type": "text", + "content": "[1] Abulikemu Abuduweili and Changliu Liu. Estimating neural network robustness via lipschitz constant and architecture sensitivity. In Safe and Robust Robot Learning for Operation in the Real World Workshop @ CoRL, 2024. 2, 5, 11" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 220, + 296, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 220, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 61, + 220, + 296, + 251 + ], + "type": "text", + "content": "[2] Anish Athalye, Nicholas Carlini, and David Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In ICML, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 252, + 295, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 252, + 295, + 284 + ], + "spans": [ + { + "bbox": [ + 62, + 252, + 295, + 284 + ], + "type": "text", + "content": "[3] Tao Bai, Jinqi Luo, Jun Zhao, Bihan Wen, and Qian Wang. Recent advances in adversarial training for adversarial robustness. In IJCAI, 2021. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 285, + 295, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 295, + 328 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 295, + 328 + ], + "type": "text", + "content": "[4] Battista Biggio, Igino Corona, Davide Maiorca, Blaine Nelson, Nedim Šrndić, Pavel Laskov, Giorgio Giacinto, and Fabio Roli. Evasion attacks against machine learning at test time. In ECML-PKDD, 2013. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 329, + 294, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 329, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 62, + 329, + 294, + 361 + ], + "type": "text", + "content": "[5] Tianlong Chen, Zhenyu Zhang, Sijia Liu, Shiyu Chang, and Zhangyang Wang. Robust overfitting may be mitigated by properly learned smoothening. In ICLR, 2021. 3, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 362, + 294, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 362, + 294, + 395 + ], + "spans": [ + { + "bbox": [ + 62, + 362, + 294, + 395 + ], + "type": "text", + "content": "[6] Minhao Cheng, Qi Lei, Pin-Yu Chen, Inderjit Dhillon, and Cho-Jui Hsieh. Cat: Customized adversarial training for improved robustness. In *IJCAI*, 2022. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 396, + 294, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 396, + 294, + 427 + ], + "spans": [ + { + "bbox": [ + 62, + 396, + 294, + 427 + ], + "type": "text", + "content": "[7] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In ICML, 2020. 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 428, + 294, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 428, + 294, + 471 + ], + "spans": [ + { + "bbox": [ + 62, + 428, + 294, + 471 + ], + "type": "text", + "content": "[8] Gavin Weiguang Ding, Yash Sharma, Kry Yik Chau Lui, and Ruitong Huang. Mma training: Direct input space margin maximization through adversarial training. In ICLR, 2020. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 472, + 294, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 472, + 294, + 504 + ], + "spans": [ + { + "bbox": [ + 62, + 472, + 294, + 504 + ], + "type": "text", + "content": "[9] Yinpeng Dong, Qi-An Fu, Xiao Yang, Tianyu Pang, Hang Su, Zihao Xiao, and Jun Zhu. Benchmarking adversarial robustness on image classification. In CVPR, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 505, + 294, + 537 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 505, + 294, + 537 + ], + "spans": [ + { + "bbox": [ + 57, + 505, + 294, + 537 + ], + "type": "text", + "content": "[10] Yinpeng Dong, Ke Xu, Xiao Yang, Tianyu Pang, Zhijie Deng, Hang Su, and Jun Zhu. Exploring memorization in adversarial training. In ICLR, 2022. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 538, + 294, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 538, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 57, + 538, + 294, + 581 + ], + "type": "text", + "content": "[11] Mahyar Fazlyab, Taha Entesari, Aniket Roy, and Rama Chellappa. Certified robustness via dynamic margin maximization and improved lipschitz regularization. In NeurIPS, 2023. 2, 5, 11" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 582, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 582, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 57, + 582, + 294, + 613 + ], + "type": "text", + "content": "[12] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In ICLR, 2015. 1, 2, 5, 11" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 615, + 294, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 615, + 294, + 647 + ], + "spans": [ + { + "bbox": [ + 57, + 615, + 294, + 647 + ], + "type": "text", + "content": "[13] Shruthi Gowda, Bahram Zonooz, and Elahe Arani. Conserve-update-revise to cure generalization and robustness trade-off in adversarial training. In ICLR, 2024. 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 647, + 294, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 680 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 680 + ], + "type": "text", + "content": "[14] Chuan Guo, Mayank Rana, Moustapha Cisse, and Laurens van der Maaten. Countering adversarial images using input transformations. In ICLR, 2018. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "type": "text", + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "type": "text", + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In ECCV, 2016. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 107, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 162 + ], + "type": "text", + "content": "[17] Xiaowei Huang, Daniel Kroening, Wenjie Ruan, James Sharp, Youcheng Sun, Emese Thamo, Min Wu, and Xinping Yi. A survey of safety and trustworthiness of deep neural networks: Verification, testing, adversarial attack and defence, and interpretability. Computer Science Review, 2020. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 163, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 206 + ], + "type": "text", + "content": "[18] Pavel Izmailov, Dmitrii Podoprikhin, Timur Garipov, Dmitry Vetrov, and Andrew Gordon Wilson. Averaging weights leads to wider optima and better generalization. In UAI, 2018. 3, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "text", + "content": "[19] Sekitoshi Kanai, Shin'ya Yamaguchi, Masanori Yamada, Hiroshi Takahashi, Kentaro Ohno, and Yasutoshi Ida. One-vs-the-rest loss to focus on important samples in adversarial training. In ICML, 2023. 2, 3, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 252, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 252, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 553, + 285 + ], + "type": "text", + "content": "[20] Minseon Kim, Jihoon Tack, Jinwoo Shin, and Sung Ju Hwang. Entropy weighted adversarial training. In Adversarial Machine Learning Workshop @ ICML, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 286, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 319 + ], + "type": "text", + "content": "[21] Vishaal Krishnan, Al Makedah, Abed AlRahman, and Fabio Pasqualetti. Lipschitz bounds and provably robust training by laplacian smoothing. In NeurIPS, 2020. 2, 5, 11" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 320, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 553, + 342 + ], + "type": "text", + "content": "[22] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. 2009. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 343, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 364 + ], + "type": "text", + "content": "[23] Fei-Fei Li, Andrej Karpathy, and Justin Johnson. Tiny image dataset. 2015. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 365, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 398 + ], + "type": "text", + "content": "[24] Zi Liang, Pinghui Wang, Ruofei Zhang, Nuo Xu, Shuo Zhang, Lifeng Xing, Haitao Bai, and Ziyang Zhou. Merge: Fast private text generation. In AAAI, 2024. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 399, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 553, + 443 + ], + "type": "text", + "content": "[25] Fangzhou Liao, Ming Liang, Yinpeng Dong, Tianyu Pang, Xiaolin Hu, and Jun Zhu. Defense against adversarial attacks using high-level representation guided denoiser. In CVPR, 2018. 1, 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 445, + 553, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 553, + 488 + ], + "type": "text", + "content": "[26] Weilin Lin, Nanjun Zhou, Yanyun Wang, Jianze Li, Hui Xiong, and Li Liu. Backdoor: A comprehensive benchmark for backdoor learning in diffusion model. arXiv preprint arXiv:2502.11798, 2025. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 490, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 490, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 490, + 553, + 533 + ], + "type": "text", + "content": "[27] Feng Liu, Bo Han, Tongliang Liu, Chen Gong, Gang Niu, Mingyuan Zhou, Masashi Sugiyama, et al. Probabilistic margins for instance reweighting in adversarial training. In NeurIPS, 2021. 2, 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 535, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 553, + 578 + ], + "type": "text", + "content": "[28] Xin Liu, Yichen Yang, Kun He, and John E Hopcroft. Parameter interpolation adversarial training for robust image classification. IEEE Transactions on Information Forensics and Security (TIFS), 2025. 2, 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 579, + 553, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 579, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 317, + 579, + 553, + 621 + ], + "type": "text", + "content": "[29] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In ICLR, 2018. 2, 3, 6" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 624, + 553, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 655 + ], + "type": "text", + "content": "[30] Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. In ICLR, 2018. 12" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "type": "text", + "content": "[31] Tianyu Pang, Xiao Yang, Yinpeng Dong, Hang Su, and Jun Zhu. Bag of tricks for adversarial training. In ICLR, 2021. 11" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[32] Nicolas Papernot, Patrick McDaniel, Ian Goodfellow, Somesh Jha, Z Berkay Celik, and Ananthram Swami. Practi" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4699" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "cal black-box attacks against machine learning. In AsiaCCS, 2017. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 139 + ], + "type": "text", + "content": "[33] Patricia Pauli, Anne Koch, Julian Berberich, Paul Kohler, and Frank Allgower. Training robust neural networks using lipschitz bounds. IEEE Control Systems Letters, 2021. 2, 5, 11" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 173 + ], + "type": "text", + "content": "[34] Rahul Rade and Seyed-Mohsen Moosavi-Dezfooli. Reducing excessive margin to achieve a better accuracy vs. robustness trade-off. In ICLR, 2022. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 294, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 218 + ], + "type": "text", + "content": "[35] Aditi Raghunathan, Sang Michael Xie, Fanny Yang, John Duchi, and Percy Liang. Adversarial training can hurt generalization. In Identifying and Understanding Deep Learning Phenomena Workshop @ ICML, 2019." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "type": "text", + "content": "[36] Aditi Raghunathan, Sang Michael Xie, Fanny Yang, John Duchi, and Percy Liang. Understanding and mitigating the tradeoff between robustness and accuracy. In ICML, 2020. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "text", + "content": "[37] Sylvestre-Alvise Rebuffi, Sven Gowal, Dan Andrei Calian, Florian Stimberg, Olivia Wiles, and Timothy A Mann. Data augmentation can improve robustness. In NeurIPS, 2021. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 298, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 294, + 330 + ], + "type": "text", + "content": "[38] David Stutz, Matthias Hein, and Bernt Schiele. Confidence-calibrated adversarial training: Generalizing to unseen attacks. In ICML, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 332, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 374 + ], + "type": "text", + "content": "[39] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In ICLR, 2014. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 376, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 409 + ], + "type": "text", + "content": "[40] Jihoon Tack, Sihyun Yu, Jongheon Jeong, Minseon Kim, Sung Ju Hwang, and Jinwoo Shin. Consistency regularization for adversarial robustness. In AAAI, 2022. 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 411, + 294, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 443 + ], + "type": "text", + "content": "[41] Lue Tao, Lei Feng, Jinfeng Yi, Sheng-Jun Huang, and Songcan Chen. Better safe than sorry: Preventing delusive adversaries with adversarial training. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 445, + 294, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 477 + ], + "type": "text", + "content": "[42] Florian Tramer, Nicholas Carlini, Wieland Brendel, and Aleksander Madry. On adaptive attacks to adversarial example defenses. In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 479, + 294, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 294, + 511 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 294, + 511 + ], + "type": "text", + "content": "[43] Dimitris Tsipras, Shibani Santurkar, Logan Engstrom, Alexander Turner, and Aleksander Madry. Robustness may be at odds with accuracy. In ICLR, 2019. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 513, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 513, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 513, + 294, + 555 + ], + "type": "text", + "content": "[44] Haotao Wang, Tianlong Chen, Shupeng Gui, TingKuei Hu, Ji Liu, and Zhangyang Wang. Once-for-all adversarial training: In-situ tradeoff between robustness and accuracy for free. In NeurIPS, 2020. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 557, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 600 + ], + "type": "text", + "content": "[45] Yisen Wang, Difan Zou, Jinfeng Yi, James Bailey, Xingjun Ma, and Quanquan Gu. Improving adversarial robustness requires revisiting misclassified examples. In ICLR, 2020. 1, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 602, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 644 + ], + "type": "text", + "content": "[46] Yanyun Wang, Dehui Du, Haibo Hu, Zi Liang, and Yuan-hao Liu. Tsfool: Crafting highly-imperceptible adversarial time series through multi-objective attack. arXiv preprint arXiv:2209.06388, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "text", + "content": "[47] Yifei Wang, Liangchen Li, Jiansheng Yang, Zhouchen Lin, and Yisen Wang. Balance, imbalance, and rebalance: Understanding robust overfitting from a minimax game perspective. In NeurIPS, 2023. 3, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "text", + "content": "[48] Yanyun Wang, Li Liu, Zi Liang, Qingqing Ye, and Haibo Hu. New paradigm of adversarial training: Releasing accuracy" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 545 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "robustness trade-off via dummy class. arXiv preprint arXiv:2410.12671, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "type": "text", + "content": "[49] Dongxian Wu, Shu-Tao Xia, and Yisen Wang. Adversarial weight perturbation helps robust generalization. In NeurIPS, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 130, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 162 + ], + "type": "text", + "content": "[50] Yu-Yu Wu, Hung-Jui Wang, and Shang-Tse Chen. Annealing self-distillation rectification improves adversarial training. In ICLR, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 164, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 164, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 164, + 553, + 196 + ], + "type": "text", + "content": "[51] Cihang Xie, Jianyu Wang, Zhishuai Zhang, Zhou Ren, and Alan Yuille. Mitigating adversarial effects through randomization. In ICLR, 2018. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 198, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 229 + ], + "type": "text", + "content": "[52] Weilin Xu, David Evans, and Yanjun Qi. Feature squeezing: Detecting adversarial examples in deep neural networks. In NDSS, 2018. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 232, + 553, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 232, + 553, + 253 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 553, + 253 + ], + "type": "text", + "content": "[53] Shuo Yang and Chang Xu. One size does not fit all: Data-adaptive adversarial training. In ECCV, 2022. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 255, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 255, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 255, + 553, + 308 + ], + "type": "text", + "content": "[54] Jia-Li Yin, Bin Chen, Wanqing Zhu, Bo-Hao Chen, and Ximeng Liu. Push stricter to decide better: A class-conditional feature adaptive framework for improving adversarial robustness. IEEE Transactions on Information Forensics and Security (TIFS), 2023. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 311, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 311, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 311, + 553, + 354 + ], + "type": "text", + "content": "[55] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In ICCV, 2019. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 355, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 553, + 376 + ], + "type": "text", + "content": "[56] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In BMVC, 2016. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 378, + 553, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 378, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 316, + 378, + 553, + 410 + ], + "type": "text", + "content": "[57] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. Mixup: Beyond empirical risk minimization. In ICLR, 2019. 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 412, + 553, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 412, + 553, + 455 + ], + "spans": [ + { + "bbox": [ + 316, + 412, + 553, + 455 + ], + "type": "text", + "content": "[58] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In ICML, 2019, 2, 3, 6, 11" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 457, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 457, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 457, + 553, + 498 + ], + "type": "text", + "content": "[59] Jingfeng Zhang, Jianing Zhu, Gang Niu, Bo Han, Masashi Sugiyama, and Mohan Kankanhalli. Geometry-aware instance-reweighted adversarial training. In ICLR, 2021. 2, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 502, + 553, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 502, + 553, + 545 + ], + "spans": [ + { + "bbox": [ + 316, + 502, + 553, + 545 + ], + "type": "text", + "content": "[60] Ziyu Zhou, Gengyu Lyu, Yiming Huang, Zihao Wang, Ziyu Jia, and Zhen Yang. Sdformer: Transformer with spectral filter and dynamic attention for multivariate time series long-term forecasting. In IJCAI, 2024. 1" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "4700" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Fair Generation without Unfair Distortions_ Debiasing Text-to-Image Generation with Entanglement-Free Attention/6ef80803-cc14-446f-a75b-f2593f51cc3f_content_list.json b/2025/Fair Generation without Unfair Distortions_ Debiasing Text-to-Image Generation with Entanglement-Free Attention/6ef80803-cc14-446f-a75b-f2593f51cc3f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a86c54bbd1a835a1e1dc5e541c87c6abf9ba02df --- /dev/null +++ b/2025/Fair Generation without Unfair Distortions_ Debiasing Text-to-Image Generation with Entanglement-Free Attention/6ef80803-cc14-446f-a75b-f2593f51cc3f_content_list.json @@ -0,0 +1,1597 @@ +[ + { + "type": "text", + "text": "Fair Generation without Unfair Distortions: Debiasing Text-to-Image Generation with Entanglement-Free Attention", + "text_level": 1, + "bbox": [ + 137, + 130, + 859, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jeonghoon Park* KAIST", + "bbox": [ + 209, + 204, + 346, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jeonghoon_park@kaist.ac.kr", + "bbox": [ + 165, + 242, + 393, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jaeseong Lee \nYonsei University", + "bbox": [ + 199, + 265, + 341, + 300 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jasonlee1995@yonsei.ac.kr", + "bbox": [ + 161, + 301, + 382, + 316 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Juyoung Lee\\* Kakao Corp.", + "bbox": [ + 467, + 204, + 581, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zealota11@gmail.com", + "bbox": [ + 439, + 242, + 609, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jaegul Choo KAIST", + "bbox": [ + 454, + 265, + 555, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jchoo@kaist.ac.kr", + "bbox": [ + 429, + 301, + 581, + 316 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chaeyeon Chung KAIST", + "bbox": [ + 674, + 204, + 813, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cy_chung@kaist.ac.kr", + "bbox": [ + 656, + 242, + 831, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jindong Gu University of Oxford", + "bbox": [ + 648, + 265, + 815, + 299 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jindong.gu@eng.ox.ac.uk", + "bbox": [ + 627, + 301, + 831, + 316 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "\"face of successful man with his house\"", + "text_level": 1, + "bbox": [ + 351, + 338, + 666, + 354 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fbdcdfa0ee8a26eaf8ee496c218844d4a23628980a15b10eccd0da4d276fbc20.jpg", + "image_caption": [ + "Stable Diffusion" + ], + "image_footnote": [], + "bbox": [ + 96, + 373, + 292, + 526 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f38018fff929b9db7940ef993dc65c926a2a2b202c1abf3235fe7208f2bc55e2.jpg", + "image_caption": [ + "UCE", + "Racial attributes changed Non-racial attributes preserved" + ], + "image_footnote": [], + "bbox": [ + 299, + 373, + 493, + 525 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b696fb8b08e0855bfa572c64def597237035baf501c58fd3c69ab917cc3e5745.jpg", + "image_caption": [ + "Interpret Diffusion" + ], + "image_footnote": [], + "bbox": [ + 503, + 375, + 694, + 523 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2db09a8ebc6f2a0d703ea0f190a6c708d472b1930d23fd51e25b697e478401c2.jpg", + "image_caption": [ + "EFA (Ours)", + "Racial attributes changed Non-racial attributes preserved", + "Figure 1. Generated images from the pretrained model (e.g., Stable Diffusion [1]), by previous debiasing methods (e.g., UCE [11] and Interpret Diffusion [15]), and ours. Previous debiasing methods often alter the non-target attributes (e.g., house) while modifying target bias attributes (e.g., racial attributes). In contrast, our approach preserves non-target attributes during bias mitigation (i.e., changing the racial attributes without modifying the house in the background), thereby preventing the introduction of unintended biases. The corresponding aligned images were generated using the same random seed." + ], + "image_footnote": [], + "bbox": [ + 702, + 375, + 900, + 525 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 650, + 326, + 666 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in diffusion-based text-to-image (T2I) models have enabled the generation of high-quality and photorealistic images from text. However, they often exhibit societal biases related to gender, race, and socioeconomic status, thereby potentially reinforcing harmful stereotypes and shaping public perception in unintended ways. While existing bias mitigation methods demonstrate effectiveness, they often encounter attribute entanglement, where adjustments to attributes relevant to the bias (i.e., target attributes) unintentionally alter attributes unassociated with the bias (i.e., non-target attributes), causing undesirable distribution shifts. To address this challenge, we introduce Entanglement-Free Attention (EFA), a method that accurately incorporates target attributes (e.g., White, Black, and", + "bbox": [ + 89, + 689, + 485, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Asian) while preserving non-target attributes (e.g., background) during bias mitigation. At inference time, EFA randomly samples a target attribute with equal probability and adjusts the cross-attention in selected layers to incorporate the sampled attribute, achieving a fair distribution of target attributes. Extensive experiments demonstrate that EFA outperforms existing methods in mitigating bias while preserving non-target attributes, thereby maintaining the original model's output distribution and generative capacity.", + "bbox": [ + 511, + 651, + 908, + 787 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 814, + 645, + 830 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion-based text-to-image (T2I) models [2, 18, 21, 23, 24] have gained significant attention due to their ability to reflect user intent through text input, making them highly applicable across various domains. However, despite their", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17567", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "advancements, these T2I models often perpetuate socially sensitive biases and stereotypes such as gender, race, or socioeconomic status [3, 7, 10]. Such societal biases can reinforce discriminatory perceptions of certain social groups and pose ethical and practical risks in real-world environments, where biased outputs can influence public perception and decision making.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous works attempt to mitigate such biases via two primary approaches. One widely explored strategy is finetuning, where pretrained models are adjusted using additional training on curated datasets [11, 19, 26]. Despite its effectiveness, this method often entails substantial computational costs [26]. Also, it may result in a decline in image quality and diversity due to the relatively small size of the finetuning dataset compared to the original pretraining dataset [25]. Since this approach inherently modifies the parameters of the model, it may lead to unintended changes in the output distribution. An alternative line of research explores inference-time steering, which avoids modifying the pretrained model and instead applies adjustments only during inference [8, 10, 13, 15, 20, 27]. Although this approach effectively reduces bias while preserving the parameters of the original model, we observe that it still introduces unintended alterations in the generated outputs.", + "bbox": [ + 91, + 200, + 482, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fig. 1 illustrates this issue by presenting images generated using the prompt \"face of successful man with his house\". The pretrained model correctly interprets the prompt by generating images of a man with a house but predominantly produces individuals from a limited racial background. When existing bias mitigation methods are applied, the representation of underrepresented racial attributes, such as Black and Indian individuals, is successfully improved. However, they also introduce unintended shifts in non-racial attributes. Specifically, they alter the material and architectural style of the house, associating certain racial groups with impoverished or deteriorated living conditions, thus reinforcing misleading associations between social groups and socioeconomic status. This finding suggests that previous methods may involve attribute entanglement between attributes relevant to the target bias (i.e., target attribute, such as racial attributes) and attributes unassociated with the target bias (i.e., non-target attribute, such as the appearance of the house), potentially leading to unintended alterations in the non-target attributes during the bias mitigation process. Without careful disentanglement, bias mitigation can unintentionally alter contextual details, degrade output diversity, and even introduce new biases.", + "bbox": [ + 91, + 459, + 482, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this issue, we propose Entanglement-Free Attention (EFA), a method that accurately incorporates target attributes while preserving non-target attributes. In this paper, we focus on human-centric bias, which plays a crucial role in shaping societal perceptions. Our EFA operates by leveraging predefined target attributes (e.g., White, Black,", + "bbox": [ + 89, + 810, + 480, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Asian, and Indian) that contribute to bias (e.g., race). To guide the model to focus on semantically relevant regions (e.g., face and body), EFA is trained to enhance target attributes in regions that exhibit high semantic relevance to the counterfactual prompts (i.e., prompts with alternative attributes other than the target), while human segmentation masks further constrain its effect to human regions. By employing both target and counterfactual prompts in training, EFA learns to dynamically adjust the enhancement strength based on the presence of the target attribute, applying stronger modifications when it is underrepresented. During inference, to ensure a fair distribution of target attributes in the generated output, one attribute is randomly sampled with equal probability. EFA then enhances the corresponding features in bias-relevant regions without relying on human-provided masks, effectively disentangling target and non-target attributes while preserving the original model's generative capacity and contextual consistency.", + "bbox": [ + 511, + 90, + 903, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our key contributions are as follows:", + "bbox": [ + 532, + 364, + 859, + 377 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We identify the issue of attribute entanglement, where bias mitigation unintentionally alters non-target attributes, leading to unintended distribution shifts in diffusion-based T2I models.", + "- We propose EFA, a novel debiasing approach that decouples target and non-target attributes, ensuring that bias mitigation is applied only to the intended target attributes while preserving the integrity of non-target attributes.", + "- Quantitative and qualitative experiments demonstrate that our approach successfully mitigates bias while maximally preserving the generation capability of the original model, outperforming previous methods." + ], + "bbox": [ + 513, + 378, + 903, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related works", + "text_level": 1, + "bbox": [ + 513, + 573, + 658, + 588 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing studies on mitigating bias in T2I generative models have mainly followed two approaches. The first approach involves directly modifying the parameters of diffusion-based T2I models. For instance, Orgad et al. [19] and Gandikota et al. [11] proposed a method to adjust the weights of the cross-attention module to alleviate the bias. Similarly, Shen et al. [26] finetune the text encoder to alleviate face-related biases (e.g., gender, race). However, since finetuning is conducted on a significantly smaller dataset compared to the pretraining dataset, such approach may potentially degrade original output quality or diversity. Furthermore, the approach of Shen et al. [26] requires substantial computational costs for training the backbone model.", + "bbox": [ + 511, + 598, + 903, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another line of research has explored inference-time steering approaches, where the pretrained model remains unchanged while adjustments are applied during inference. Friedrich et al. [10] proposed a bias mitigation framework that employs semantic guidance [4] to manipulate the target attributes by referring to registered information of a concept (e.g., nurse) and its bias (e.g., gender). However, this cannot", + "bbox": [ + 511, + 795, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17568", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d76059e9a7a449ec9996a39c21cb33d42d1fea70edb5c08f9100cdc71f2b9dd9.jpg", + "image_caption": [ + "Figure 2. Overview of our method. (a): EFA is applied to selected layers and trained to reconstruct the image $\\mathbf{x}_0$ (a-1) when given $p_a$ , the prompt containing the target attribute $a$ , or (a-2) its counterfactual prompt, $p_a^{\\mathrm{cf}}$ . We leverage a human segmentation mask $M$ to minimize the effect of EFA on the attention map of non-target attribute regions. (b): During the inference, our method generates outputs with a fair distribution of target attributes using the EFA of the randomly sampled target attribute." + ], + "image_footnote": [], + "bbox": [ + 89, + 85, + 911, + 328 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "mitigate bias for unregistered concepts, limiting its practicality. Chuang et al. [8] have developed a method to remove bias from text embeddings, and Kim et al. [13] trains the soft prompts for mitigating bias. Beyond employing the text embeddings, Shrestha et al. [27] introduced a technique that projects a reference image into text-compatible tokens, allowing the model to leverage the features of external images, such as gender or race for image generation. Additionally, Parihar et al. [20] and Li et al. [15] proposed methods to manipulate the h-space [14] of the model during inference. While these methods aim to mitigate bias without directly altering the backbone parameters, they often struggle to preserve the original context of the generated images.", + "bbox": [ + 88, + 412, + 480, + 608 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, Zhou et al. [31] addressed association-engendered stereotypes arising from co-occurrence of multiple objects in an image. They rely on a CLIP model trained to predict biases from prompts, requiring prior knowledge of the biases to be mitigated. In contrast, our approach aims to address subtle biases induced by contextual variations beyond the predefined association bias definitions. For instance, if a particular social group is predominantly depicted in low-resource work environments, while another is shown in corporate office settings, such context variations can reinforce cognitive biases, subtly implying disparities in professional status. Furthermore, it is impractical to anticipate all possible forms of such bias in advance.", + "bbox": [ + 89, + 609, + 482, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Entanglement-free attention", + "text_level": 1, + "bbox": [ + 89, + 821, + 352, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 89, + 847, + 200, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Without loss of generality, we assume a given target bias $C$ (e.g., gender) along with its corresponding set of target", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "attributes $\\mathcal{A}_C = \\{a_1, a_2, \\ldots, a_n\\}$ (e.g., 'female', 'male') whose elements should be uniformly distributed in the outputs. Then, our approach aims to alleviate the bias $C$ by adjusting the distribution of target attributes in $\\mathcal{A}_C$ while preserving the non-target attributes such as scene context and surrounding objects.", + "bbox": [ + 511, + 412, + 903, + 502 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To achieve this, we introduce EFA, which leverages the principle of cross-attention mechanism that assigns desired features to appropriate locations. Each target attribute $a_{i} \\in \\mathcal{A}_{C}$ is handled by a dedicated EFA module (i.e., $\\mathrm{EFA}_{C}^{a_{i}}$ ), ensuring faithful representation of $a_{i}$ in the outputs while minimizing its influence on non-target attribute regions. During inference, to enforce a fair distribution of target attributes in the generated output, we randomly sample an attribute $a_{i}$ from $\\mathcal{A}_{C}$ with equal probability and incorporate its representation using the corresponding EFA (Fig. 2(b)).", + "bbox": [ + 511, + 503, + 905, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To enhance the fidelity of $a_i$ , $\\mathrm{EFA}_C^{a_i}$ is trained to reconstruct images depicting $a_i$ when its representation is subtle by introducing counterfactual attributes $a_i^{\\mathrm{cf}} \\in \\mathcal{A}_C \\setminus \\{a_i\\}$ through counterfactual prompts. EFA is guided to enhance the target attribute in areas exhibiting strong semantic relevance to the counterfactual prompt through the cross-attention module. Additionally, human segmentation masks are used during training to constrain EFA's influence to the human region. This ensures that $\\mathrm{EFA}_C^{a_i}$ enhances the fidelity of $a_i$ while preventing unintended modifications to non-target attributes.", + "bbox": [ + 511, + 655, + 905, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Architecture of EFA", + "text_level": 1, + "bbox": [ + 511, + 832, + 705, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The original cross-attention module takes an intermediate spatial feature $\\mathbf{z}_t$ , which is derived by from a noised image $\\mathbf{x}_t$ and a prompt $p$ using the denoising model $\\phi$ (i.e.,", + "bbox": [ + 511, + 854, + 905, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17569", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\mathbf{z}_t = \\phi (\\mathbf{x}_t,p)$ . $\\mathbf{z}_t$ is transformed into a query representation using the mapping layer $W_{q}$ . The text prompt embeddings $\\pi (p)$ are projected into key and value representations through $W_{k}$ and $W_{v}$ , respectively. The output of the cross-attention layer is formulated as:", + "bbox": [ + 89, + 90, + 483, + 167 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {C r o s s - a t t e n t i o n} (\\mathbf {z} _ {t}) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {\\top}}{\\sqrt {d}}\\right) V,\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 176, + 424, + 212 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $d$ indicates the dimension of the query and key vectors, and $Q = \\mathbf{z}_tW_q$ , $K = \\pi(p)W_k$ , and $V = \\pi(p)W_v$ denote the query, key, and value matrices, respectively.", + "bbox": [ + 89, + 220, + 482, + 267 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As illustrated in Fig. 3, EFA modifies the cross-attention to integrate the desired attribute $a_{i} \\in \\mathcal{A}_{C}$ into the output while preserving the original attention structure. Specifically, EFA takes the intermediate feature $\\mathbf{z}_t$ and predicts the additional attention values that determine where and to what extent $a_{i}$ should be enhanced. This is achieved using the attention value predictor (AP), a lightweight module composed of three convolutional layers and two SiLU [9] activation functions. Further details of the architecture are provided in the Supplementary.", + "bbox": [ + 88, + 267, + 483, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The predicted attention value (i.e., $\\mathrm{AP}_{a_i}(\\mathbf{z}_t)$ ) are concatenated with the original cross-attention's attention values (i.e., $\\frac{QK^{\\top}}{\\sqrt{d}}$ ), and the softmax function is applied. After applying softmax, the original value vectors from the user prompt (i.e., $V$ ) and the value vectors of $a_i$ are weighted accordingly, as in the standard cross-attention operation. Formally, EFA incorporates $a_i$ into the cross-attention as follows:", + "bbox": [ + 89, + 417, + 483, + 542 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {E F A} _ {C} ^ {a _ {i}} (\\mathbf {z} _ {t}) = \\operatorname {s o f t m a x} \\left(\\left[ \\frac {Q K ^ {\\top}}{\\sqrt {d}}, \\mathrm {A P} _ {a _ {i}} (\\mathbf {z} _ {t}) \\right]\\right) [ V, V _ {a _ {i}} ],\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 551, + 454, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $V_{a_i} = \\pi(p_{a_i}) W_v$ , $p_{a_i}$ indicates the text prompt representing $a_i$ , and $[\\cdot, \\cdot]$ indicates the concatenation.", + "bbox": [ + 89, + 595, + 482, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When building an $\\mathrm{EFA}_C$ for a target bias $C$ , all AP modules share the same convolutional backbone while maintaining distinct output channels designated for each attribute in $\\mathcal{A}_C$ . This design allows AP models to be effectively trained together, ensuring that EFAs of target attributes for a given bias are jointly learned while avoiding the inefficiency of training multiple independent APs.", + "bbox": [ + 89, + 627, + 483, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Training", + "text_level": 1, + "bbox": [ + 89, + 742, + 194, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dataset construction. To train $\\mathrm{EFA}_C^{a_i}$ to represent a target attribute $a_{i}$ (e.g., female) while preserving non-target attributes, we create a dataset $\\mathcal{D}_{a_i}$ that captures $a_{i}$ and provides pseudo supervision on the human regions that contain $a_{i}$ . Specifically, we generate images using the original model with prompts explicitly describing $a_{i}$ (e.g., \"a female person\"). For each image, we extract a human segmentation mask $M$ using Grounded SAM2 [17, 22], assigning 1 to human regions and 0 elsewhere. Since this study focuses on", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/02f98c40ffc1e6cebfb3e755dad271f7b93be770eb596d9a6aa54ab44172384c.jpg", + "image_caption": [ + "Figure 3. Architecture of EFA. We denote $a_i \\in \\mathcal{A}_C$ as $a$ in this figure for notational simplicity." + ], + "image_footnote": [], + "bbox": [ + 514, + 87, + 906, + 220 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "human-centric bias, we utilize a human segmentation mask to serve as the basis for regulating the influence of $a_{i}$ during training. The resulting $\\mathcal{D}_{a_i}$ consists of pairs of images and their segmentation masks.", + "bbox": [ + 511, + 268, + 906, + 329 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training EFA. At inference, the extent of the features that need to be adjusted by EFA to faithfully reflect $a_{i}$ may vary across inputs. For instance, with the prompt \"nurse\", intermediate features may exhibit a high level of femininity, requiring minimal adjustment by $\\mathrm{EFA}_{\\text{gender}}^{\\text{female}}$ . In contrast, intermediate features of the prompt \"doctor\" may have stronger masculinity, requiring enhanced representation of female by $\\mathrm{EFA}_{\\text{gender}}^{\\text{female}}$ to incorporate 'female' into the output. To handle this variability, $\\mathrm{EFA}_{C}^{a_{i}}$ is trained in two different scenarios: (1) when intermediate features contain the attribute $a_{i}$ (e.g., female), and (2) when intermediate features contain its counterfactual attribute $a_{i}^{\\mathrm{cf}} \\in \\mathcal{A}_{C} \\setminus \\{a_{i}\\}$ (e.g., male). We denote $a_{i}$ as $a$ throughout this section for simplicity.", + "bbox": [ + 511, + 329, + 905, + 525 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let $\\mathbf{x}_0$ be an image and $M$ be its corresponding segmentation mask, sampled from $\\mathcal{D}_a$ , where $a$ denotes a target attribute (e.g., female). In the first scenario, where the intermediate features already encode $a$ , $\\mathrm{EFA}_C^a$ does not need to enhance $a$ but should instead preserve the original cross-attention output to maintain the non-target attributes. To enforce this behavior, we provide the model with a paired input: a prompt $p_a$ explicitly specifying $a$ (e.g., \"female person\") and a noised image $\\mathbf{x}_t$ that inherently contains $a$ , as depicted in Fig. 2 (a-1).", + "bbox": [ + 511, + 526, + 905, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To prevent unnecessary alterations, we introduce an L1 regularization term that encourages the attention values associated with $V_{a}$ (e.g., attention values associated with the value vectors of \"female\") to be zero. This regularization loss is defined as:", + "bbox": [ + 511, + 676, + 905, + 751 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e g}} (\\mathbf {z} _ {t}, W) = \\left\\| W \\odot A _ {a} \\left(\\mathbf {z} _ {t}\\right) \\right\\| _ {1},\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 762, + 816, + 780 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{z}_t = \\phi (\\mathbf{x}_t,p_a)$ $W$ is a spatial mask used to constrain the effect of the loss, and $\\odot$ represents the element-wise multiplication. Also, $A_{a}(\\mathbf{z}_{t})$ denotes the attention values (after softmax) associated with the value vectors $V_{a}$ given $\\mathbf{z}_t$ . More precisely, when $A(\\mathbf{z}_t) =$ softmax $([\\frac{QK^{\\top}}{\\sqrt{d}},\\mathrm{AP}_a(\\mathbf{z}_t)])$ , we define $A_{a}(\\mathbf{z}_{t})$ as the submatrix of $A(\\mathbf{z}_t)$ obtained by selecting values indexed by $I_{V_a}$", + "bbox": [ + 511, + 789, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17570", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $I_{V_a}$ denotes the set of indices corresponding to $V_a$ .", + "bbox": [ + 89, + 90, + 467, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this scenario, we set $W = \\mathbf{1}$ , a mask of ones, to apply the loss over all spatial positions. This regularization loss minimizes unnecessary modifications to the original cross-attention output introduced by $\\mathrm{EFA}_C^a$ . Hereafter, we denote $\\mathcal{L}_{\\mathrm{reg}}(\\mathbf{z}_t,\\mathbf{1})$ as $\\mathcal{L}_{\\mathrm{reg}}^{\\mathrm{trg}}$ for brevity.", + "bbox": [ + 89, + 106, + 480, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the second scenario, the intermediate features contain a counterfactual attribute $a^{\\mathrm{cf}}$ (e.g., male) induced by bias of prompt (e.g., \"photo of a doctor\" implicitly favoring male representation). Here, $\\mathrm{EFA}_C^a$ should suppress $a^{\\mathrm{cf}}$ and enhance $a$ (e.g., female) while preserving non-target attributes. To simulate this, we train $\\mathrm{EFA}_C^a$ to counteract counterfactual features introduced during the denoising process of a diffusion-based T2I model, which generates clean images by iteratively denoising from pure noise.", + "bbox": [ + 88, + 181, + 480, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To encourage representation of $a$ in human-related regions, we define a reconstruction loss leveraging the segmentation mask $M$ to constrain its effect on human areas. The objective is to predict the added noise $\\epsilon$ given $\\mathbf{x}_t$ and a counterfactual prompt $p_a^{\\mathrm{cf}}$ , formulated as follows:", + "bbox": [ + 89, + 316, + 480, + 393 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e c o n}} = \\mathbb {E} _ {\\mathbf {x} _ {0}, t, p _ {a} ^ {\\mathrm {c f}}, \\epsilon \\sim N (0, 1)} \\left[ \\| M \\odot (\\epsilon - \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t, p _ {a} ^ {\\mathrm {c f}})) \\| _ {2} ^ {2} \\right],\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 398, + 477, + 419 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\epsilon_{\\theta}$ denotes the model with $\\mathrm{EFA}_C^a$ . Through the reconstruction process, $\\mathrm{EFA}_C^a$ is encouraged to enhance $a$ in regions where cross-attention assigns high similarity to $a^{\\mathrm{cf}}$ , reflecting their semantic relevance. Using human masks ensures that $a$ is accurately reconstructed in human regions while mitigating the influence of $a^{\\mathrm{cf}}$ .", + "bbox": [ + 89, + 422, + 480, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although $\\mathcal{L}_{\\mathrm{reg}}^{\\mathrm{trg}}$ reduces unintended modifications, it does not explicitly confine its impact on human regions, potentially affecting non-target attributes. To address this, we impose an additional L1 regularization on the attention values corresponding to $V_{a}$ , applied outside the human region, as follows:", + "bbox": [ + 89, + 513, + 480, + 602 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {r e g}} ^ {\\mathrm {c f}} = \\mathcal {L} _ {\\mathrm {r e g}} (\\mathbf {z} _ {t} ^ {\\mathrm {c f}}, \\mathbf {1} - \\bar {M}),\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 603, + 370, + 622 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{z}_t^{\\mathrm{cf}} = \\phi (\\mathbf{x}_t,p_a^{\\mathrm{cf}})$ , and $\\bar{M}$ indicates the binary mask resized to match the resolution of the input feature of $\\mathrm{EFA}_C^a$ .", + "bbox": [ + 89, + 625, + 480, + 656 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, the overall training loss is formulated as:", + "bbox": [ + 109, + 657, + 434, + 672 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {t o t}} = \\mathcal {L} _ {\\mathrm {r e c o n}} + \\mathcal {L} _ {\\mathrm {r e g}} ^ {*},\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 679, + 356, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{reg}}^{*} = \\lambda_{1}\\mathcal{L}_{\\mathrm{reg}}^{\\mathrm{trg}} + \\lambda_{2}\\mathcal{L}_{\\mathrm{reg}}^{\\mathrm{cf}}$ and $\\lambda_{1}$ and $\\lambda_{2}$ are hyperparameters to control the relative importance between the losses. During training, only the parameters of $\\mathrm{AP}_a$ in $\\mathrm{EFA}_C^a$ are updated. Note that EFA is applied to the upblock layers with an input resolution of $16\\times 16$ , as lower-resolution features capture higher-level semantic information. The choice of applied layers is discussed in Section 5.2. The method of extending EFA to handle multiple biases is provided in the Supplementary.", + "bbox": [ + 89, + 704, + 480, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Inference", + "text_level": 1, + "bbox": [ + 89, + 848, + 200, + 862 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To mitigate the bias $C$ and ensure fair generation, we randomly sample an attribute $a_{i}$ with an equal probability from", + "bbox": [ + 89, + 869, + 480, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathcal{A}_C$ and apply the corresponding $\\mathrm{EFA}_C^{a_i}$ to the original model. This approach enables users to easily control the attribute distribution by adjusting the sampling frequency, without additional training. Note that our method requires no masks at inference and generates debiased images directly from noise, offering a practical and efficient solution.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 199, + 645, + 215 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental settings", + "text_level": 1, + "bbox": [ + 511, + 224, + 718, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. Following the previous studies [10, 11, 15], we mainly conducted experiments with gender and racial bias. Experiments with age bias are provided in Supplementary. The target attributes set of gender and racial bias is set as $\\mathcal{A}_{\\mathrm{gender}} = \\{\\mathrm{female}, \\mathrm{male}\\}$ and $\\mathcal{A}_{\\mathrm{race}} = \\{\\mathrm{White}, \\mathrm{Black}, \\mathrm{Asian}, \\mathrm{Indian}\\}$ , respectively. For multiple biases, the set of target attributes is defined as $\\mathcal{A}_{\\mathrm{gender} \\times \\mathrm{race}} = \\mathcal{A}_{\\mathrm{gender}} \\times \\mathcal{A}_{\\mathrm{race}}$ . We adopt a binary gender perspective to evaluate our method, following prior works [11, 15], due to the limitations of existing datasets and evaluation protocols. Further details of the dataset construction are included in the Supplementary.", + "bbox": [ + 511, + 247, + 903, + 429 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation prompt. To assess bias mitigation performance, we utilized 36 occupations from the WinoBias [30] dataset, which includes stereotypical occupations based on US Department of Labor statistics by following previous studies [11, 15, 19]. We employed two types of prompt templates for image generation. The first set (i.e., $\\mathcal{T}_{\\mathrm{basic}}$ ) follows the prompts used in prior research [11, 15], formatted as “{\\style} of {\\occupation}.” The style category includes photo, image, picture, headshot, and portrait. The second set (i.e., $\\mathcal{T}_{\\mathrm{complex}}$ ) includes background components to enable evaluation with more complex non-target attributes. Specifically, prompts are structured as “{\\style} of {\\occupation} in {\\background},” where the background is contextually relevant to the occupation.", + "bbox": [ + 511, + 430, + 905, + 641 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To evaluate the overall model preservation performance while ensuring non-target attributes remain unaffected, we assessed whether the model retained its original generation capability when generating images without human presence. Specifically, we filtered image-prompt pairs from the COCO-30K [16] dataset and constructed a new dataset, COCO-no-person, consisting of 14,926 pairs where no person is present in the corresponding images.", + "bbox": [ + 511, + 643, + 905, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation metrics. Our evaluation focuses on (1) bias mitigation, ensuring balanced attribute representation; (2) non-target attribute preservation, maintaining non-target attributes while adjusting target attributes; and (3) model preservation, maintaining the overall generative capability of the original model in non-human contexts. While model preservation also evaluates the preservation of non-target attributes, we use a distinct term to differentiate it from cases where target attributes should be enhanced.", + "bbox": [ + 511, + 763, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17571", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ea4c9bffff25d724cce4ff4335e2f09bdd8bc62b135c93ffe170d1e3a0d15abc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
\\(T_{\\text{basic}}\\)\\(T_{\\text{complex}}\\)
BiasNon-target attribute P.BiasNon-target attribute P.
DR ↓PSNR ↑LPIPS ↓DINO ↑DR ↓PSNR ↑LPIPS ↓DINO ↑
GenderOriginal SD0.71---0.71---
Concept Algebra [28]0.5921.100.11690.8230.6916.690.18520.834
UCE [11]0.3421.040.13740.7570.4916.780.22020.747
Finetuning Diffusion [26]0.4822.620.11660.8140.5619.270.16110.857
Interpret Diffusion [15]0.2617.180.22900.6160.3114.010.32110.618
EFA (Ours)0.0632.520.04110.9160.0629.700.04920.941
RaceOriginal SD0.60---0.55---
Concept Algebra [28]0.6421.470.11640.8390.5816.620.18880.841
UCE [11]0.2721.550.12610.7870.4117.810.18910.801
Finetuning Diffusion [26]0.2918.960.17950.7090.3716.460.22630.771
Interpret Diffusion [15]0.1616.870.24160.5840.2313.730.33380.584
EFA (Ours)0.0430.930.03530.9380.0628.550.04210.958
Gender × RaceOriginal SD0.56---0.50---
Concept Algebra [28]0.5120.120.13250.8050.4715.880.20310.818
UCE [11]0.1619.840.15750.7020.3216.790.21980.746
Finetuning Diffusion [26]0.2419.970.15700.7460.2817.240.20640.795
Interpret Diffusion [15]0.1616.910.24490.4660.2013.660.35290.475
EFA (Ours)0.0325.580.06840.8530.0523.780.07950.903
", + "bbox": [ + 99, + 88, + 903, + 409 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison to baselines. P. is the abbreviation for preservation. The best values are in bold. Our method achieves superior bias mitigation performance while effectively preserving non-target attributes, significantly outperforming previous approaches.", + "bbox": [ + 89, + 417, + 906, + 448 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3bd84c11dd5af4ab6eff29155185078b65e8724f06573fb65eafe4fcf2dc240f.jpg", + "image_caption": [ + "Figure 4. Generated images from the original Stable Diffusion and ours. Ours successfully generates images with a fair distribution of the target bias (i.e., gender and race). (a) and (b) represent the images generated from the prompts describing a nurse in a hospital and a carpenter in a workshop, respectively. The corresponding aligned images were generated using the same random seed. Further qualitative results, including those from the model mitigating gender and race simultaneously, are provided in the Supplementary." + ], + "image_footnote": [], + "bbox": [ + 91, + 460, + 493, + 720 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f17e94095042218fd2ba9ade9e55a707081fdbc36383b0ad1dc484e406d53b43.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 521, + 460, + 905, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For evaluating bias mitigation, we use a CLIP classifier to categorize each generated image into attribute-defined groups in $\\mathcal{A}_C$ and measure attribute balance using modified deviation ratio [15], which is designed to handle an arbitrary number of attributes. The deviation ratio (DR) is calculated as $\\max_{a_i\\in \\mathcal{A}_C}\\frac{|N_{a_i} / N - 1 / |\\mathcal{A}_C||}{1 - 1 / |\\mathcal{A}_C|}$ , where $N$ and $N_{a_i}$ indicate", + "bbox": [ + 89, + 801, + 485, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the total number of generated images and images classified as belonging to class $a_{i}$ , respectively. The lower DR indicates a better balance of attributes. We compute scores for each occupation and average them to obtain the final score.", + "bbox": [ + 511, + 801, + 906, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For non-target attribute preservation scores, we generate image pairs using the original Stable Diffusion (SD) [1] and", + "bbox": [ + 511, + 869, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17572", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0cf7aea8de1de8f39a4d49efaf67b64cf16b2c93c67b04a3577000077ef4ab4a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFID ↓CLIP-T ↑
-Original SD-26.17
GenderConcept Algebra [28]2.4126.02
UCE [11]11.6525.17
Finetuning Diffusion [26]1.9225.79
Interpret Diffusion [15]15.7824.80
EFA (Ours)0.2326.03
RaceConcept Algebra [28]2.4826.04
UCE [11]5.1626.13
Finetuning Diffusion [26]4.1225.78
Interpret Diffusion [15]21.6623.82
EFA (Ours)0.1326.21
G.R.Concept Algebra [28]2.5326.03
UCE [11]6.5725.71
Finetuning Diffusion [26]4.2625.71
Interpret Diffusion [15]41.6522.45
EFA (Ours)0.4526.20
", + "bbox": [ + 99, + 88, + 480, + 357 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Evaluation of model preservation using COCO-no-person. EFA outperforms the baselines in preserving the generation capability of the original model.", + "bbox": [ + 89, + 366, + 482, + 409 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "our approach with the same random seed to evaluate how well non-target attributes are preserved while modifying target attributes. First, we segment the non-human regions in the original image and measure how well they are retained in the image generated by ours using Peak Signal-to-Noise Ratio (PSNR) and Learned Perceptual Image Patch Similarity (LPIPS) [29]. To assess overall image preservation, including contextual elements within human regions (e.g., a doctor's stethoscope or white coat), we compute the DINO score, employing DINO's ability to capture low-level and semantically meaningful visual details. The score measures the cosine similarity between the DINO [5] embeddings of images generated by SD and our method.", + "bbox": [ + 88, + 431, + 483, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Lastly, for assessing model preservation, we use COCO-no-person dataset. To measure text fidelity, we use the CLIP-T score, which computes the cosine similarity between CLIP embeddings of the input prompt and the generated images. Additionally, Fréchet Inception Distance (FID) [12] is used to evaluate image quality and diversity. The FID scores are computed against the generation outputs of the original SD. Unlike non-target attribute preservation scores, this evaluation ensures that our method does not degrade the generation capability of the pretrained model when bias mitigation is not required. Note that bold text in all tables indicates the best score.", + "bbox": [ + 89, + 636, + 483, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation details. The experiments were conducted based on SD v1.5 followed by previous works [10, 20, 26, 31]. For evaluation with WinoBias, we generated 160 images per occupation, using random seeds where the original SD has a detectable face.", + "bbox": [ + 89, + 824, + 483, + 898 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/46ccff1e1fafee023e5745c964c638a4fdd3f91a87d0a55789f7171846f88136.jpg", + "image_caption": [ + "Figure 5. Attention map visualization when mitigating gender bias using EFA. (a) and (b) show the results of applying $\\mathrm{EFA}_{\\mathrm{gender}}^{\\mathrm{female}}$ and $\\mathrm{EFA}_{\\mathrm{gender}}^{\\mathrm{male}}$ , respectively. $<\\mathrm{sos}>$ and $<\\mathrm{eos}>$ denote special tokens that mark the beginning and end of a text sequence, respectively." + ], + "image_footnote": [], + "bbox": [ + 516, + 89, + 906, + 224 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Experimental results", + "text_level": 1, + "bbox": [ + 511, + 303, + 709, + 319 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Bias mitigation and non-target attribute preservation. Table 1 presents the comparative experimental results of our method and baseline models using the WinoBias dataset. Overall, models struggle to maintain non-target attributes more in $\\mathcal{T}_{\\mathrm{complex}}$ than $\\mathcal{T}_{\\mathrm{basic}}$ . Concept Algebra exhibits low performance in bias mitigation. Both UCE [11] and Fine-tuning Diffusion [26] modify model parameters directly, making it challenging to preserve the non-target attributes consistent with the original SD. Interpret Diffusion [15] demonstrates better bias mitigation performance compared to other baselines. However, it exhibits low non-target attribute preservation scores due to its uniform application of fixed semantic vectors without spatial awareness. In contrast, our input-aware architecture adaptively enhances target attributes based on spatial context, enabling effective preservation of non-target attributes.", + "bbox": [ + 511, + 325, + 906, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our proposed method outperforms existing baselines by achieving a lower DR while having substantially higher non-target attribute preservation scores. The high preservation scores validate the effectiveness of our approach, showing that the model successfully learns to preserve regions corresponding to non-target attributes.", + "bbox": [ + 511, + 566, + 905, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We hypothesize that the effectiveness in bias mitigation stems from the combination of two key factors: (1) the rich semantic information encoded in text embeddings, which have been pretrained on large-scale datasets, and (2) the capability of U-Net to faithfully reflect these embeddings in the generated output. By utilizing pretrained text embeddings without altering their inherent semantics, our approach ensures that target attributes are seamlessly integrated into the generation process, while U-Net's capacity to translate these embeddings into visual features enables precise control over attribute distributions. The extended experiments on Concept Algebra, provided in the supplementary material, corroborate our hypothesis. Consequently, our method successfully steers the distribution of target attributes while preserving non-target attributes.", + "bbox": [ + 511, + 657, + 906, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 4 demonstrates that our method effectively miti-", + "bbox": [ + 531, + 885, + 903, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17573", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f9c12192b67d865ec4ab958c87c191d8962b282ff26a48073b6e1569c14e10fa.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BiasNon-target attribute P.
DR ↓PSNR ↑LPIPS ↓DINO ↑
Original SD0.71--
EFA w/o Ltrg0.0727.760.05270.890
EFA w/o seg. mask0.0429.310.05620.888
EFA (Ours)0.0632.520.04110.916
", + "bbox": [ + 99, + 88, + 478, + 189 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "gates societal bias, generating outputs that represent a diverse range of gender and race. Furthermore, by leveraging regularization terms and utilizing text embeddings disentangled from other attributes, our method successfully preserves non-target attributes, such as overall layout, pose, and background, while effectively mitigating bias.", + "bbox": [ + 88, + 260, + 482, + 351 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Model preservation. The experimental results on model preservation are presented in Table 2. Our method achieves a significantly lower FID compared to baseline approaches, indicating superior preservation of both the quality and diversity of SD's outputs. Furthermore, our approach yields CLIP-T scores comparable to those of SD, effectively maintaining the text fidelity of the original model by preserving non-target attributes. Qualitative results on model preservation with COCO-no-person and additional experiments using COCO-30K are provided in the Supplementary.", + "bbox": [ + 89, + 352, + 482, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Analysis", + "text_level": 1, + "bbox": [ + 89, + 518, + 187, + 535 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1. Effect of regularization and segmentation mask", + "text_level": 1, + "bbox": [ + 89, + 544, + 482, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Table 3, we conduct an ablation study to examine the effects of regularization losses and segmentation masks on non-target attribute preservation. For the experiments, we train models to mitigate gender bias, and the evaluation is performed using $\\mathcal{T}_{\\mathrm{basic}}$ .", + "bbox": [ + 89, + 566, + 482, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "When training without segmentation masks, we exclude $\\mathcal{L}_{\\mathrm{reg}}^{\\mathrm{cf}}$ in training and apply $\\mathcal{L}_{\\mathrm{recon}}$ to the entire image. Since there is no constraint limiting target attribute enhancement, we observe a slight improvement in bias mitigation performance. However, without segmentation masks to distinguish human and non-human regions during training, EFA struggles to localize them precisely, leading to reduced preservation of non-target attributes at inference. Nonetheless, EFA still outperforms baselines even without segmentation masks. We conjecture that this is because counterfactual attributes are primarily activated in bias-relevant regions due to their semantic similarity in cross-attention, which encourages EFA to enhance the target attribute in those regions while ignoring the counterfactual prompt's influence. This allows EFA to approximate the localization of bias-relevant regions without relying on fine-grained masks of the target attribute. Also, when $\\mathcal{L}_{\\mathrm{reg}}^{\\mathrm{trg}}$ is not applied, the", + "bbox": [ + 89, + 643, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/433d238b35e002dfcbb70fa2400b353bd3771dcbbbab75a36271d809ce1df7bc.jpg", + "table_caption": [ + "Table 3. Effectiveness of the regularization loss and the segmentation mask in bias mitigation and non-target attribute preservation. P. and seg. indicate preservation and segmentation, respectively." + ], + "table_footnote": [], + "table_body": "
resolution of inputBiasNon-target attribute P.
DR ↓PSNR ↑LPIPS ↓DINO ↑
16×16 (Ours)0.0632.520.04110.916
16×16 - 32×320.0631.790.04070.915
16×16 - 64×640.0630.650.04380.911
", + "bbox": [ + 521, + 88, + 903, + 176 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Impact of input resolution employed by EFA. Extending EFA to higher-resolution layers compromises the preservation of non-target attributes. P. denotes preservation.", + "bbox": [ + 511, + 186, + 906, + 228 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "non-target attribute preservation performance deteriorates, even though the bias mitigation performance remains similar. These results highlight the necessity of the regularization mechanisms to preserve non-target attributes across diverse scenarios where the target attribute appears at different extents in the input.", + "bbox": [ + 511, + 237, + 903, + 328 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Analyzing the influence of resolution in EFA", + "text_level": 1, + "bbox": [ + 511, + 337, + 885, + 353 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To investigate the impact of input resolution, we integrated EFA into additional layers within the up-blocks while training a model for gender bias mitigation. The performance on $\\mathcal{T}_{\\mathrm{basic}}$ is reported in Table 4, where the row highlighted in gray represents our proposed method. When EFA is applied to higher-resolution features, the deviation ratio remains nearly unchanged, while PSNR and DINO scores exhibit a slight decline. This suggests that lower-resolution layers contain richer semantic information, which is crucial for attribute modification. Moreover, extending EFA to multiple layers inevitably introduces additional noise, potentially compromising the preservation of non-target attributes.", + "bbox": [ + 511, + 359, + 906, + 541 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Attention map visualization", + "text_level": 1, + "bbox": [ + 511, + 549, + 763, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Fig. 5, we visualize the attention maps predicted by EFA when the target attribute is enhanced. EFA assigns high values to the human region, effectively enabling gender transformation while preserving the generation capability of the original model, as the impact on accessories, clothing style, and background details remains minimal.", + "bbox": [ + 511, + 571, + 905, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 676, + 633, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we addressed attribute entanglement in diffusion-based T2I models, where bias mitigation often distorts non-target attributes, leading to unintended distribution shifts. We propose EFA, which mitigates human-centric biases while preserving non-target attributes. During training, EFA is guided to attend regions with high semantic relevance to counterfactual prompts, constrained by human segmentation masks. At inference, EFA randomly samples a target attribute and alters the cross-attention in selected layers to ensure a fair distribution of target attributes in the outputs. Experimental results verify that EFA outperforms prior methods in mitigating bias while preserving the original model's output distribution and generative capacity.", + "bbox": [ + 511, + 700, + 906, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17574", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by Institute for Information & communications Technology Planning & Evaluation(IITP) grant funded by the Korea government(MSIT) (RS-2019-II190075, Artificial Intelligence Graduate School Program(KAIST)), Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.RS-2021-II212068, Artificial Intelligence Innovation Hub), and the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT)(No. 2022R1A5A7083908).", + "bbox": [ + 89, + 113, + 485, + 255 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 281, + 187, + 297 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Stability AI. Stable diffusion. https://huggingface.co/runwayml/stable-diffusion-v1-5, 2022. [Online; accessed 3-November-2023]. 1, 6", + "[2] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. https://cdn.openai.com/papers/dall-e-3.pdf, 2023. [Online; accessed 4-December-2024]. 1", + "[3] Federico Bianchi, Pratyusha Kalluri, Esin Durmus, Faisal Ladhak, Myra Cheng, Debora Nozza, Tatsunori Hashimoto, Dan Jurafsky, James Zou, and Aylin Caliskan. Easily accessible text-to-image generation amplifies demographic stereotypes at large scale. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, pages 1493-1504, 2023. 2", + "[4] Manuel Brack, Felix Friedrich, Dominik Hintersdorf, Lukas Struppek, Patrick Schramowski, and Kristian Kersting. Sega: Instructing text-to-image models using semantic guidance. Advances in Neural Information Processing Systems, 36: 25365-25389, 2023. 2", + "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9650-9660, 2021. 7", + "[6] Aditya Chinchure, Pushkar Shukla, Gaurav Bhatt, Kiri Salij, Kartik Hosanagar, Leonid Sigal, and Matthew Turk. Tibet: Identifying and evaluating biases in text-to-image generative models. In European Conference on Computer Vision, pages 429-446. Springer, 2025. 5", + "[7] Jaemin Cho, Abhay Zala, and Mohit Bansal. Dall-eval: Probing the reasoning skills and social biases of text-to-image generation models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3043-3054, 2023. 2", + "[8] Ching-Yao Chuang, Varun Jampani, Yuanzhen Li, Antonio Torralba, and Stefanie Jegelka. Debiasing vision-language models via biased prompts. arXiv preprint arXiv:2302.00070, 2023. 2, 3", + "[9] Stefan Elfwing, Eiji Uchibe, and Kenji Doya. Sigmoid-weighted linear units for neural network function approxi" + ], + "bbox": [ + 99, + 306, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "mation in reinforcement learning. CoRR, abs/1702.03118, 2017.4", + "[10] Felix Friedrich, Manuel Brack, Lukas Struppek, Dominik Hintersdorf, Patrick Schramowski, Sasha Luccioni, and Kristian Kersting. Fair diffusion: Instructing text-to-image generation models on fairness. arXiv preprint at arXiv:2302.10893, 2023. 2, 5, 7", + "[11] Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzyńska, and David Bau. Unified concept editing in diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5111-5120, 2024. 1, 2, 5, 6, 7, 4", + "[12] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7", + "[13] Eunji Kim, Siwon Kim, Chaehun Shin, and Sungroh Yoon. De-stereotyping text-to-image models through prompt tuning. 2023. 2, 3", + "[14] Mingi Kwon, Jaeseok Jeong, and Youngjung Uh. Diffusion models already have a semantic latent space. In The Eleventh International Conference on Learning Representations, 2023. 3", + "[15] Hang Li, Chengzhi Shen, Philip Torr, Volker Tresp, and Jindong Gu. Self-discovering interpretable diffusion latent directions for responsible text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 3, 5, 6, 7, 4", + "[16] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer vision-ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part v 13, pages 740-755. Springer, 2014. 5", + "[17] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 4", + "[18] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1", + "[19] Hadas Orgad, Bahjat Kawar, and Yonatan Belinkov. Editing implicit assumptions in text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7053-7061, 2023. 2, 5", + "[20] Rishubh Parihar, Abhijnya Bhat, Abhipsa Basu, Saswat Mallick, Jogendra Nath Kundu, and R. Venkatesh Babu. Balancing act: Distribution-guided debiasing in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6668-6678, 2024. 2, 3, 7", + "[21] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image gen" + ], + "bbox": [ + 516, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17575", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "eration with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1", + "[22] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dólar, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos, 2024. 4", + "[23] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1", + "[24] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. 35:36479-36494, 2022. 1", + "[25] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. 35:25278-25294, 2022. 2", + "[26] Xudong Shen, Chao Du, Tianyu Pang, Min Lin, Yongkang Wong, and Mohan Kankanhalli. Finetuning text-to-image diffusion models for fairness. In The Twelfth International Conference on Learning Representations, 2024. 2, 6, 7, 4", + "[27] Robik Shrestha, Yang Zou, Qiuyu Chen, Zhiheng Li, Yusheng Xie, and Siqi Deng. Fairrag: Fair human generation via fair retrieval augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11996-12005, 2024. 2, 3", + "[28] Zihao Wang, Lin Gui, Jeffrey Negrea, and Victor Veitch. Concept algebra for (score-based) text-controlled generative models. In Proceedings of the 37th International Conference on Neural Information Processing Systems, pages 35331-35349, 2023. 6, 7, 3, 4", + "[29] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7", + "[30] Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, and Kai-Wei Chang. Gender bias in coreference resolution: Evaluation and debiasing methods. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 15–20, 2018. 5", + "[31] Junlei Zhou, Jiashi Gao, Xiangyu Zhao, Xin Yao, and Xuetao Wei. Association of objects may engender stereotypes: Mitigating association-engendered stereotypes in text-to-image generation. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 3, 7" + ], + "bbox": [ + 91, + 90, + 482, + 864 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17576", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Fair Generation without Unfair Distortions_ Debiasing Text-to-Image Generation with Entanglement-Free Attention/6ef80803-cc14-446f-a75b-f2593f51cc3f_model.json b/2025/Fair Generation without Unfair Distortions_ Debiasing Text-to-Image Generation with Entanglement-Free Attention/6ef80803-cc14-446f-a75b-f2593f51cc3f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..56b384ef2d5d5022323779b7b1803112bcef3a59 --- /dev/null +++ b/2025/Fair Generation without Unfair Distortions_ Debiasing Text-to-Image Generation with Entanglement-Free Attention/6ef80803-cc14-446f-a75b-f2593f51cc3f_model.json @@ -0,0 +1,2024 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.138, + 0.131, + 0.861, + 0.177 + ], + "angle": 0, + "content": "Fair Generation without Unfair Distortions: Debiasing Text-to-Image Generation with Entanglement-Free Attention" + }, + { + "type": "text", + "bbox": [ + 0.21, + 0.205, + 0.348, + 0.238 + ], + "angle": 0, + "content": "Jeonghoon Park* KAIST" + }, + { + "type": "text", + "bbox": [ + 0.166, + 0.243, + 0.394, + 0.257 + ], + "angle": 0, + "content": "jeonghoon_park@kaist.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.2, + 0.266, + 0.343, + 0.301 + ], + "angle": 0, + "content": "Jaeseong Lee \nYonsei University" + }, + { + "type": "text", + "bbox": [ + 0.162, + 0.303, + 0.383, + 0.317 + ], + "angle": 0, + "content": "jasonlee1995@yonsei.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.468, + 0.205, + 0.582, + 0.24 + ], + "angle": 0, + "content": "Juyoung Lee\\* Kakao Corp." + }, + { + "type": "text", + "bbox": [ + 0.441, + 0.243, + 0.611, + 0.256 + ], + "angle": 0, + "content": "zealota11@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.455, + 0.266, + 0.556, + 0.299 + ], + "angle": 0, + "content": "Jaegul Choo KAIST" + }, + { + "type": "text", + "bbox": [ + 0.43, + 0.303, + 0.582, + 0.317 + ], + "angle": 0, + "content": "jchoo@kaist.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.675, + 0.205, + 0.814, + 0.238 + ], + "angle": 0, + "content": "Chaeyeon Chung KAIST" + }, + { + "type": "text", + "bbox": [ + 0.657, + 0.243, + 0.832, + 0.257 + ], + "angle": 0, + "content": "cy_chung@kaist.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.649, + 0.266, + 0.816, + 0.3 + ], + "angle": 0, + "content": "Jindong Gu University of Oxford" + }, + { + "type": "text", + "bbox": [ + 0.629, + 0.303, + 0.832, + 0.317 + ], + "angle": 0, + "content": "jindong.gu@eng.ox.ac.uk" + }, + { + "type": "title", + "bbox": [ + 0.352, + 0.339, + 0.667, + 0.356 + ], + "angle": 0, + "content": "\"face of successful man with his house\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.139, + 0.359, + 0.25, + 0.372 + ], + "angle": 0, + "content": "Stable Diffusion" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.374, + 0.293, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.379, + 0.359, + 0.415, + 0.372 + ], + "angle": 0, + "content": "UCE" + }, + { + "type": "image", + "bbox": [ + 0.3, + 0.374, + 0.495, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.359, + 0.662, + 0.374 + ], + "angle": 0, + "content": "Interpret Diffusion" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.376, + 0.696, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.761, + 0.359, + 0.841, + 0.373 + ], + "angle": 0, + "content": "EFA (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.375, + 0.901, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.393, + 0.529, + 0.601, + 0.556 + ], + "angle": 0, + "content": "Racial attributes changed Non-racial attributes preserved" + }, + { + "type": "image_caption", + "bbox": [ + 0.696, + 0.529, + 0.904, + 0.557 + ], + "angle": 0, + "content": "Racial attributes changed Non-racial attributes preserved" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.568, + 0.907, + 0.639 + ], + "angle": 0, + "content": "Figure 1. Generated images from the pretrained model (e.g., Stable Diffusion [1]), by previous debiasing methods (e.g., UCE [11] and Interpret Diffusion [15]), and ours. Previous debiasing methods often alter the non-target attributes (e.g., house) while modifying target bias attributes (e.g., racial attributes). In contrast, our approach preserves non-target attributes during bias mitigation (i.e., changing the racial attributes without modifying the house in the background), thereby preventing the introduction of unintended biases. The corresponding aligned images were generated using the same random seed." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.651, + 0.327, + 0.667 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Recent advancements in diffusion-based text-to-image (T2I) models have enabled the generation of high-quality and photorealistic images from text. However, they often exhibit societal biases related to gender, race, and socioeconomic status, thereby potentially reinforcing harmful stereotypes and shaping public perception in unintended ways. While existing bias mitigation methods demonstrate effectiveness, they often encounter attribute entanglement, where adjustments to attributes relevant to the bias (i.e., target attributes) unintentionally alter attributes unassociated with the bias (i.e., non-target attributes), causing undesirable distribution shifts. To address this challenge, we introduce Entanglement-Free Attention (EFA), a method that accurately incorporates target attributes (e.g., White, Black, and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.652, + 0.909, + 0.789 + ], + "angle": 0, + "content": "Asian) while preserving non-target attributes (e.g., background) during bias mitigation. At inference time, EFA randomly samples a target attribute with equal probability and adjusts the cross-attention in selected layers to incorporate the sampled attribute, achieving a fair distribution of target attributes. Extensive experiments demonstrate that EFA outperforms existing methods in mitigating bias while preserving non-target attributes, thereby maintaining the original model's output distribution and generative capacity." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.815, + 0.646, + 0.831 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Diffusion-based text-to-image (T2I) models [2, 18, 21, 23, 24] have gained significant attention due to their ability to reflect user intent through text input, making them highly applicable across various domains. However, despite their" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17567" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "advancements, these T2I models often perpetuate socially sensitive biases and stereotypes such as gender, race, or socioeconomic status [3, 7, 10]. Such societal biases can reinforce discriminatory perceptions of certain social groups and pose ethical and practical risks in real-world environments, where biased outputs can influence public perception and decision making." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.201, + 0.483, + 0.457 + ], + "angle": 0, + "content": "Previous works attempt to mitigate such biases via two primary approaches. One widely explored strategy is finetuning, where pretrained models are adjusted using additional training on curated datasets [11, 19, 26]. Despite its effectiveness, this method often entails substantial computational costs [26]. Also, it may result in a decline in image quality and diversity due to the relatively small size of the finetuning dataset compared to the original pretraining dataset [25]. Since this approach inherently modifies the parameters of the model, it may lead to unintended changes in the output distribution. An alternative line of research explores inference-time steering, which avoids modifying the pretrained model and instead applies adjustments only during inference [8, 10, 13, 15, 20, 27]. Although this approach effectively reduces bias while preserving the parameters of the original model, we observe that it still introduces unintended alterations in the generated outputs." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.46, + 0.483, + 0.807 + ], + "angle": 0, + "content": "Fig. 1 illustrates this issue by presenting images generated using the prompt \"face of successful man with his house\". The pretrained model correctly interprets the prompt by generating images of a man with a house but predominantly produces individuals from a limited racial background. When existing bias mitigation methods are applied, the representation of underrepresented racial attributes, such as Black and Indian individuals, is successfully improved. However, they also introduce unintended shifts in non-racial attributes. Specifically, they alter the material and architectural style of the house, associating certain racial groups with impoverished or deteriorated living conditions, thus reinforcing misleading associations between social groups and socioeconomic status. This finding suggests that previous methods may involve attribute entanglement between attributes relevant to the target bias (i.e., target attribute, such as racial attributes) and attributes unassociated with the target bias (i.e., non-target attribute, such as the appearance of the house), potentially leading to unintended alterations in the non-target attributes during the bias mitigation process. Without careful disentanglement, bias mitigation can unintentionally alter contextual details, degrade output diversity, and even introduce new biases." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.482, + 0.902 + ], + "angle": 0, + "content": "To address this issue, we propose Entanglement-Free Attention (EFA), a method that accurately incorporates target attributes while preserving non-target attributes. In this paper, we focus on human-centric bias, which plays a crucial role in shaping societal perceptions. Our EFA operates by leveraging predefined target attributes (e.g., White, Black," + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.364 + ], + "angle": 0, + "content": "Asian, and Indian) that contribute to bias (e.g., race). To guide the model to focus on semantically relevant regions (e.g., face and body), EFA is trained to enhance target attributes in regions that exhibit high semantic relevance to the counterfactual prompts (i.e., prompts with alternative attributes other than the target), while human segmentation masks further constrain its effect to human regions. By employing both target and counterfactual prompts in training, EFA learns to dynamically adjust the enhancement strength based on the presence of the target attribute, applying stronger modifications when it is underrepresented. During inference, to ensure a fair distribution of target attributes in the generated output, one attribute is randomly sampled with equal probability. EFA then enhances the corresponding features in bias-relevant regions without relying on human-provided masks, effectively disentangling target and non-target attributes while preserving the original model's generative capacity and contextual consistency." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.365, + 0.86, + 0.378 + ], + "angle": 0, + "content": "In summary, our key contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.379, + 0.905, + 0.437 + ], + "angle": 0, + "content": "- We identify the issue of attribute entanglement, where bias mitigation unintentionally alters non-target attributes, leading to unintended distribution shifts in diffusion-based T2I models." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.44, + 0.905, + 0.499 + ], + "angle": 0, + "content": "- We propose EFA, a novel debiasing approach that decouples target and non-target attributes, ensuring that bias mitigation is applied only to the intended target attributes while preserving the integrity of non-target attributes." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.5, + 0.905, + 0.56 + ], + "angle": 0, + "content": "- Quantitative and qualitative experiments demonstrate that our approach successfully mitigates bias while maximally preserving the generation capability of the original model, outperforming previous methods." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.379, + 0.905, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.574, + 0.659, + 0.589 + ], + "angle": 0, + "content": "2. Related works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.905, + 0.794 + ], + "angle": 0, + "content": "Existing studies on mitigating bias in T2I generative models have mainly followed two approaches. The first approach involves directly modifying the parameters of diffusion-based T2I models. For instance, Orgad et al. [19] and Gandikota et al. [11] proposed a method to adjust the weights of the cross-attention module to alleviate the bias. Similarly, Shen et al. [26] finetune the text encoder to alleviate face-related biases (e.g., gender, race). However, since finetuning is conducted on a significantly smaller dataset compared to the pretraining dataset, such approach may potentially degrade original output quality or diversity. Furthermore, the approach of Shen et al. [26] requires substantial computational costs for training the backbone model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Another line of research has explored inference-time steering approaches, where the pretrained model remains unchanged while adjustments are applied during inference. Friedrich et al. [10] proposed a bias mitigation framework that employs semantic guidance [4] to manipulate the target attributes by referring to registered information of a concept (e.g., nurse) and its bias (e.g., gender). However, this cannot" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "17568" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.087, + 0.912, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.337, + 0.908, + 0.396 + ], + "angle": 0, + "content": "Figure 2. Overview of our method. (a): EFA is applied to selected layers and trained to reconstruct the image \\(\\mathbf{x}_0\\) (a-1) when given \\(p_a\\), the prompt containing the target attribute \\(a\\), or (a-2) its counterfactual prompt, \\(p_a^{\\mathrm{cf}}\\). We leverage a human segmentation mask \\(M\\) to minimize the effect of EFA on the attention map of non-target attribute regions. (b): During the inference, our method generates outputs with a fair distribution of target attributes using the EFA of the randomly sampled target attribute." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.413, + 0.482, + 0.609 + ], + "angle": 0, + "content": "mitigate bias for unregistered concepts, limiting its practicality. Chuang et al. [8] have developed a method to remove bias from text embeddings, and Kim et al. [13] trains the soft prompts for mitigating bias. Beyond employing the text embeddings, Shrestha et al. [27] introduced a technique that projects a reference image into text-compatible tokens, allowing the model to leverage the features of external images, such as gender or race for image generation. Additionally, Parihar et al. [20] and Li et al. [15] proposed methods to manipulate the h-space [14] of the model during inference. While these methods aim to mitigate bias without directly altering the backbone parameters, they often struggle to preserve the original context of the generated images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.61, + 0.483, + 0.807 + ], + "angle": 0, + "content": "Recently, Zhou et al. [31] addressed association-engendered stereotypes arising from co-occurrence of multiple objects in an image. They rely on a CLIP model trained to predict biases from prompts, requiring prior knowledge of the biases to be mitigated. In contrast, our approach aims to address subtle biases induced by contextual variations beyond the predefined association bias definitions. For instance, if a particular social group is predominantly depicted in low-resource work environments, while another is shown in corporate office settings, such context variations can reinforce cognitive biases, subtly implying disparities in professional status. Furthermore, it is impractical to anticipate all possible forms of such bias in advance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.823, + 0.353, + 0.839 + ], + "angle": 0, + "content": "3. Entanglement-free attention" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.201, + 0.863 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Without loss of generality, we assume a given target bias \\(C\\) (e.g., gender) along with its corresponding set of target" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.413, + 0.905, + 0.503 + ], + "angle": 0, + "content": "attributes \\(\\mathcal{A}_C = \\{a_1, a_2, \\ldots, a_n\\}\\) (e.g., 'female', 'male') whose elements should be uniformly distributed in the outputs. Then, our approach aims to alleviate the bias \\(C\\) by adjusting the distribution of target attributes in \\(\\mathcal{A}_C\\) while preserving the non-target attributes such as scene context and surrounding objects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.504, + 0.906, + 0.655 + ], + "angle": 0, + "content": "To achieve this, we introduce EFA, which leverages the principle of cross-attention mechanism that assigns desired features to appropriate locations. Each target attribute \\(a_{i} \\in \\mathcal{A}_{C}\\) is handled by a dedicated EFA module (i.e., \\(\\mathrm{EFA}_{C}^{a_{i}}\\)), ensuring faithful representation of \\(a_{i}\\) in the outputs while minimizing its influence on non-target attribute regions. During inference, to enforce a fair distribution of target attributes in the generated output, we randomly sample an attribute \\(a_{i}\\) from \\(\\mathcal{A}_{C}\\) with equal probability and incorporate its representation using the corresponding EFA (Fig. 2(b))." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.656, + 0.906, + 0.822 + ], + "angle": 0, + "content": "To enhance the fidelity of \\( a_i \\), \\( \\mathrm{EFA}_C^{a_i} \\) is trained to reconstruct images depicting \\( a_i \\) when its representation is subtle by introducing counterfactual attributes \\( a_i^{\\mathrm{cf}} \\in \\mathcal{A}_C \\setminus \\{a_i\\} \\) through counterfactual prompts. EFA is guided to enhance the target attribute in areas exhibiting strong semantic relevance to the counterfactual prompt through the cross-attention module. Additionally, human segmentation masks are used during training to constrain EFA's influence to the human region. This ensures that \\( \\mathrm{EFA}_C^{a_i} \\) enhances the fidelity of \\( a_i \\) while preventing unintended modifications to non-target attributes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.707, + 0.848 + ], + "angle": 0, + "content": "3.2. Architecture of EFA" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.903 + ], + "angle": 0, + "content": "The original cross-attention module takes an intermediate spatial feature \\(\\mathbf{z}_t\\), which is derived by from a noised image \\(\\mathbf{x}_t\\) and a prompt \\(p\\) using the denoising model \\(\\phi\\) (i.e.," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17569" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.168 + ], + "angle": 0, + "content": "\\(\\mathbf{z}_t = \\phi (\\mathbf{x}_t,p)\\). \\(\\mathbf{z}_t\\) is transformed into a query representation using the mapping layer \\(W_{q}\\). The text prompt embeddings \\(\\pi (p)\\) are projected into key and value representations through \\(W_{k}\\) and \\(W_{v}\\), respectively. The output of the cross-attention layer is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.178, + 0.426, + 0.213 + ], + "angle": 0, + "content": "\\[\n\\operatorname {C r o s s - a t t e n t i o n} (\\mathbf {z} _ {t}) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {\\top}}{\\sqrt {d}}\\right) V,\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.222, + 0.483, + 0.268 + ], + "angle": 0, + "content": "where \\(d\\) indicates the dimension of the query and key vectors, and \\(Q = \\mathbf{z}_tW_q\\), \\(K = \\pi(p)W_k\\), and \\(V = \\pi(p)W_v\\) denote the query, key, and value matrices, respectively." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.268, + 0.484, + 0.418 + ], + "angle": 0, + "content": "As illustrated in Fig. 3, EFA modifies the cross-attention to integrate the desired attribute \\( a_{i} \\in \\mathcal{A}_{C} \\) into the output while preserving the original attention structure. Specifically, EFA takes the intermediate feature \\( \\mathbf{z}_t \\) and predicts the additional attention values that determine where and to what extent \\( a_{i} \\) should be enhanced. This is achieved using the attention value predictor (AP), a lightweight module composed of three convolutional layers and two SiLU [9] activation functions. Further details of the architecture are provided in the Supplementary." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.419, + 0.484, + 0.543 + ], + "angle": 0, + "content": "The predicted attention value (i.e., \\(\\mathrm{AP}_{a_i}(\\mathbf{z}_t)\\)) are concatenated with the original cross-attention's attention values (i.e., \\(\\frac{QK^{\\top}}{\\sqrt{d}}\\)), and the softmax function is applied. After applying softmax, the original value vectors from the user prompt (i.e., \\(V\\)) and the value vectors of \\(a_i\\) are weighted accordingly, as in the standard cross-attention operation. Formally, EFA incorporates \\(a_i\\) into the cross-attention as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.118, + 0.552, + 0.455, + 0.587 + ], + "angle": 0, + "content": "\\[\n\\operatorname {E F A} _ {C} ^ {a _ {i}} (\\mathbf {z} _ {t}) = \\operatorname {s o f t m a x} \\left(\\left[ \\frac {Q K ^ {\\top}}{\\sqrt {d}}, \\mathrm {A P} _ {a _ {i}} (\\mathbf {z} _ {t}) \\right]\\right) [ V, V _ {a _ {i}} ],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.483, + 0.627 + ], + "angle": 0, + "content": "where \\( V_{a_i} = \\pi(p_{a_i}) W_v \\), \\( p_{a_i} \\) indicates the text prompt representing \\( a_i \\), and \\( [\\cdot, \\cdot] \\) indicates the concatenation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.484, + 0.734 + ], + "angle": 0, + "content": "When building an \\(\\mathrm{EFA}_C\\) for a target bias \\(C\\), all AP modules share the same convolutional backbone while maintaining distinct output channels designated for each attribute in \\(\\mathcal{A}_C\\). This design allows AP models to be effectively trained together, ensuring that EFAs of target attributes for a given bias are jointly learned while avoiding the inefficiency of training multiple independent APs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.196, + 0.759 + ], + "angle": 0, + "content": "3.3. Training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Dataset construction. To train \\(\\mathrm{EFA}_C^{a_i}\\) to represent a target attribute \\(a_{i}\\) (e.g., female) while preserving non-target attributes, we create a dataset \\(\\mathcal{D}_{a_i}\\) that captures \\(a_{i}\\) and provides pseudo supervision on the human regions that contain \\(a_{i}\\). Specifically, we generate images using the original model with prompts explicitly describing \\(a_{i}\\) (e.g., \"a female person\"). For each image, we extract a human segmentation mask \\(M\\) using Grounded SAM2 [17, 22], assigning 1 to human regions and 0 elsewhere. Since this study focuses on" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.088, + 0.907, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.232, + 0.907, + 0.261 + ], + "angle": 0, + "content": "Figure 3. Architecture of EFA. We denote \\( a_i \\in \\mathcal{A}_C \\) as \\( a \\) in this figure for notational simplicity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.269, + 0.907, + 0.33 + ], + "angle": 0, + "content": "human-centric bias, we utilize a human segmentation mask to serve as the basis for regulating the influence of \\(a_{i}\\) during training. The resulting \\(\\mathcal{D}_{a_i}\\) consists of pairs of images and their segmentation masks." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.33, + 0.906, + 0.526 + ], + "angle": 0, + "content": "Training EFA. At inference, the extent of the features that need to be adjusted by EFA to faithfully reflect \\( a_{i} \\) may vary across inputs. For instance, with the prompt \"nurse\", intermediate features may exhibit a high level of femininity, requiring minimal adjustment by \\( \\mathrm{EFA}_{\\text{gender}}^{\\text{female}} \\). In contrast, intermediate features of the prompt \"doctor\" may have stronger masculinity, requiring enhanced representation of female by \\( \\mathrm{EFA}_{\\text{gender}}^{\\text{female}} \\) to incorporate 'female' into the output. To handle this variability, \\( \\mathrm{EFA}_{C}^{a_{i}} \\) is trained in two different scenarios: (1) when intermediate features contain the attribute \\( a_{i} \\) (e.g., female), and (2) when intermediate features contain its counterfactual attribute \\( a_{i}^{\\mathrm{cf}} \\in \\mathcal{A}_{C} \\setminus \\{a_{i}\\} \\) (e.g., male). We denote \\( a_{i} \\) as \\( a \\) throughout this section for simplicity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.527, + 0.906, + 0.678 + ], + "angle": 0, + "content": "Let \\(\\mathbf{x}_0\\) be an image and \\(M\\) be its corresponding segmentation mask, sampled from \\(\\mathcal{D}_a\\), where \\(a\\) denotes a target attribute (e.g., female). In the first scenario, where the intermediate features already encode \\(a\\), \\(\\mathrm{EFA}_C^a\\) does not need to enhance \\(a\\) but should instead preserve the original cross-attention output to maintain the non-target attributes. To enforce this behavior, we provide the model with a paired input: a prompt \\(p_a\\) explicitly specifying \\(a\\) (e.g., \"female person\") and a noised image \\(\\mathbf{x}_t\\) that inherently contains \\(a\\), as depicted in Fig. 2 (a-1)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.678, + 0.906, + 0.752 + ], + "angle": 0, + "content": "To prevent unnecessary alterations, we introduce an L1 regularization term that encourages the attention values associated with \\( V_{a} \\) (e.g., attention values associated with the value vectors of \"female\") to be zero. This regularization loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.763, + 0.818, + 0.781 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e g}} (\\mathbf {z} _ {t}, W) = \\left\\| W \\odot A _ {a} \\left(\\mathbf {z} _ {t}\\right) \\right\\| _ {1},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.79, + 0.907, + 0.903 + ], + "angle": 0, + "content": "where \\(\\mathbf{z}_t = \\phi (\\mathbf{x}_t,p_a)\\) \\(W\\) is a spatial mask used to constrain the effect of the loss, and \\(\\odot\\) represents the element-wise multiplication. Also, \\(A_{a}(\\mathbf{z}_{t})\\) denotes the attention values (after softmax) associated with the value vectors \\(V_{a}\\) given \\(\\mathbf{z}_t\\). More precisely, when \\(A(\\mathbf{z}_t) =\\) softmax \\(([\\frac{QK^{\\top}}{\\sqrt{d}},\\mathrm{AP}_a(\\mathbf{z}_t)])\\) , we define \\(A_{a}(\\mathbf{z}_{t})\\) as the submatrix of \\(A(\\mathbf{z}_t)\\) obtained by selecting values indexed by \\(I_{V_a}\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17570" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.468, + 0.107 + ], + "angle": 0, + "content": "where \\( I_{V_a} \\) denotes the set of indices corresponding to \\( V_a \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.107, + 0.482, + 0.183 + ], + "angle": 0, + "content": "In this scenario, we set \\( W = \\mathbf{1} \\), a mask of ones, to apply the loss over all spatial positions. This regularization loss minimizes unnecessary modifications to the original cross-attention output introduced by \\( \\mathrm{EFA}_C^a \\). Hereafter, we denote \\( \\mathcal{L}_{\\mathrm{reg}}(\\mathbf{z}_t,\\mathbf{1}) \\) as \\( \\mathcal{L}_{\\mathrm{reg}}^{\\mathrm{trg}} \\) for brevity." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.183, + 0.482, + 0.318 + ], + "angle": 0, + "content": "In the second scenario, the intermediate features contain a counterfactual attribute \\(a^{\\mathrm{cf}}\\) (e.g., male) induced by bias of prompt (e.g., \"photo of a doctor\" implicitly favoring male representation). Here, \\(\\mathrm{EFA}_C^a\\) should suppress \\(a^{\\mathrm{cf}}\\) and enhance \\(a\\) (e.g., female) while preserving non-target attributes. To simulate this, we train \\(\\mathrm{EFA}_C^a\\) to counteract counterfactual features introduced during the denoising process of a diffusion-based T2I model, which generates clean images by iteratively denoising from pure noise." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.482, + 0.395 + ], + "angle": 0, + "content": "To encourage representation of \\(a\\) in human-related regions, we define a reconstruction loss leveraging the segmentation mask \\(M\\) to constrain its effect on human areas. The objective is to predict the added noise \\(\\epsilon\\) given \\(\\mathbf{x}_t\\) and a counterfactual prompt \\(p_a^{\\mathrm{cf}}\\), formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.4, + 0.478, + 0.42 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e c o n}} = \\mathbb {E} _ {\\mathbf {x} _ {0}, t, p _ {a} ^ {\\mathrm {c f}}, \\epsilon \\sim N (0, 1)} \\left[ \\| M \\odot (\\epsilon - \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t, p _ {a} ^ {\\mathrm {c f}})) \\| _ {2} ^ {2} \\right],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.424, + 0.482, + 0.514 + ], + "angle": 0, + "content": "where \\(\\epsilon_{\\theta}\\) denotes the model with \\(\\mathrm{EFA}_C^a\\). Through the reconstruction process, \\(\\mathrm{EFA}_C^a\\) is encouraged to enhance \\(a\\) in regions where cross-attention assigns high similarity to \\(a^{\\mathrm{cf}}\\), reflecting their semantic relevance. Using human masks ensures that \\(a\\) is accurately reconstructed in human regions while mitigating the influence of \\(a^{\\mathrm{cf}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.515, + 0.482, + 0.603 + ], + "angle": 0, + "content": "Although \\(\\mathcal{L}_{\\mathrm{reg}}^{\\mathrm{trg}}\\) reduces unintended modifications, it does not explicitly confine its impact on human regions, potentially affecting non-target attributes. To address this, we impose an additional L1 regularization on the attention values corresponding to \\(V_{a}\\), applied outside the human region, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.604, + 0.371, + 0.623 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {r e g}} ^ {\\mathrm {c f}} = \\mathcal {L} _ {\\mathrm {r e g}} (\\mathbf {z} _ {t} ^ {\\mathrm {c f}}, \\mathbf {1} - \\bar {M}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.626, + 0.482, + 0.657 + ], + "angle": 0, + "content": "where \\(\\mathbf{z}_t^{\\mathrm{cf}} = \\phi (\\mathbf{x}_t,p_a^{\\mathrm{cf}})\\), and \\(\\bar{M}\\) indicates the binary mask resized to match the resolution of the input feature of \\(\\mathrm{EFA}_C^a\\)." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.658, + 0.435, + 0.673 + ], + "angle": 0, + "content": "Finally, the overall training loss is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.68, + 0.357, + 0.699 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {t o t}} = \\mathcal {L} _ {\\mathrm {r e c o n}} + \\mathcal {L} _ {\\mathrm {r e g}} ^ {*},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.482, + 0.842 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{reg}}^{*} = \\lambda_{1}\\mathcal{L}_{\\mathrm{reg}}^{\\mathrm{trg}} + \\lambda_{2}\\mathcal{L}_{\\mathrm{reg}}^{\\mathrm{cf}}\\) and \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) are hyperparameters to control the relative importance between the losses. During training, only the parameters of \\(\\mathrm{AP}_a\\) in \\(\\mathrm{EFA}_C^a\\) are updated. Note that EFA is applied to the upblock layers with an input resolution of \\(16\\times 16\\), as lower-resolution features capture higher-level semantic information. The choice of applied layers is discussed in Section 5.2. The method of extending EFA to handle multiple biases is provided in the Supplementary." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.202, + 0.863 + ], + "angle": 0, + "content": "3.4. Inference" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.482, + 0.902 + ], + "angle": 0, + "content": "To mitigate the bias \\( C \\) and ensure fair generation, we randomly sample an attribute \\( a_{i} \\) with an equal probability from" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.183 + ], + "angle": 0, + "content": "\\(\\mathcal{A}_C\\) and apply the corresponding \\(\\mathrm{EFA}_C^{a_i}\\) to the original model. This approach enables users to easily control the attribute distribution by adjusting the sampling frequency, without additional training. Note that our method requires no masks at inference and generates debiased images directly from noise, offering a practical and efficient solution." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.2, + 0.646, + 0.216 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.225, + 0.719, + 0.241 + ], + "angle": 0, + "content": "4.1. Experimental settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.248, + 0.905, + 0.43 + ], + "angle": 0, + "content": "Datasets. Following the previous studies [10, 11, 15], we mainly conducted experiments with gender and racial bias. Experiments with age bias are provided in Supplementary. The target attributes set of gender and racial bias is set as \\(\\mathcal{A}_{\\mathrm{gender}} = \\{\\mathrm{female}, \\mathrm{male}\\}\\) and \\(\\mathcal{A}_{\\mathrm{race}} = \\{\\mathrm{White}, \\mathrm{Black}, \\mathrm{Asian}, \\mathrm{Indian}\\}\\), respectively. For multiple biases, the set of target attributes is defined as \\(\\mathcal{A}_{\\mathrm{gender} \\times \\mathrm{race}} = \\mathcal{A}_{\\mathrm{gender}} \\times \\mathcal{A}_{\\mathrm{race}}\\). We adopt a binary gender perspective to evaluate our method, following prior works [11, 15], due to the limitations of existing datasets and evaluation protocols. Further details of the dataset construction are included in the Supplementary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.906, + 0.642 + ], + "angle": 0, + "content": "Evaluation prompt. To assess bias mitigation performance, we utilized 36 occupations from the WinoBias [30] dataset, which includes stereotypical occupations based on US Department of Labor statistics by following previous studies [11, 15, 19]. We employed two types of prompt templates for image generation. The first set (i.e., \\(\\mathcal{T}_{\\mathrm{basic}}\\)) follows the prompts used in prior research [11, 15], formatted as “{\\style} of {\\occupation}.” The style category includes photo, image, picture, headshot, and portrait. The second set (i.e., \\(\\mathcal{T}_{\\mathrm{complex}}\\)) includes background components to enable evaluation with more complex non-target attributes. Specifically, prompts are structured as “{\\style} of {\\occupation} in {\\background},” where the background is contextually relevant to the occupation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.906, + 0.764 + ], + "angle": 0, + "content": "To evaluate the overall model preservation performance while ensuring non-target attributes remain unaffected, we assessed whether the model retained its original generation capability when generating images without human presence. Specifically, we filtered image-prompt pairs from the COCO-30K [16] dataset and constructed a new dataset, COCO-no-person, consisting of 14,926 pairs where no person is present in the corresponding images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Evaluation metrics. Our evaluation focuses on (1) bias mitigation, ensuring balanced attribute representation; (2) non-target attribute preservation, maintaining non-target attributes while adjusting target attributes; and (3) model preservation, maintaining the overall generative capability of the original model in non-human contexts. While model preservation also evaluates the preservation of non-target attributes, we use a distinct term to differentiate it from cases where target attributes should be enhanced." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17571" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.089, + 0.905, + 0.41 + ], + "angle": 0, + "content": "
\\(T_{\\text{basic}}\\)\\(T_{\\text{complex}}\\)
BiasNon-target attribute P.BiasNon-target attribute P.
DR ↓PSNR ↑LPIPS ↓DINO ↑DR ↓PSNR ↑LPIPS ↓DINO ↑
GenderOriginal SD0.71---0.71---
Concept Algebra [28]0.5921.100.11690.8230.6916.690.18520.834
UCE [11]0.3421.040.13740.7570.4916.780.22020.747
Finetuning Diffusion [26]0.4822.620.11660.8140.5619.270.16110.857
Interpret Diffusion [15]0.2617.180.22900.6160.3114.010.32110.618
EFA (Ours)0.0632.520.04110.9160.0629.700.04920.941
RaceOriginal SD0.60---0.55---
Concept Algebra [28]0.6421.470.11640.8390.5816.620.18880.841
UCE [11]0.2721.550.12610.7870.4117.810.18910.801
Finetuning Diffusion [26]0.2918.960.17950.7090.3716.460.22630.771
Interpret Diffusion [15]0.1616.870.24160.5840.2313.730.33380.584
EFA (Ours)0.0430.930.03530.9380.0628.550.04210.958
Gender × RaceOriginal SD0.56---0.50---
Concept Algebra [28]0.5120.120.13250.8050.4715.880.20310.818
UCE [11]0.1619.840.15750.7020.3216.790.21980.746
Finetuning Diffusion [26]0.2419.970.15700.7460.2817.240.20640.795
Interpret Diffusion [15]0.1616.910.24490.4660.2013.660.35290.475
EFA (Ours)0.0325.580.06840.8530.0523.780.07950.903
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.419, + 0.907, + 0.449 + ], + "angle": 0, + "content": "Table 1. Comparison to baselines. P. is the abbreviation for preservation. The best values are in bold. Our method achieves superior bias mitigation performance while effectively preserving non-target attributes, significantly outperforming previous approaches." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.462, + 0.495, + 0.721 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.462, + 0.906, + 0.721 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.731, + 0.908, + 0.789 + ], + "angle": 0, + "content": "Figure 4. Generated images from the original Stable Diffusion and ours. Ours successfully generates images with a fair distribution of the target bias (i.e., gender and race). (a) and (b) represent the images generated from the prompts describing a nurse in a hospital and a carpenter in a workshop, respectively. The corresponding aligned images were generated using the same random seed. Further qualitative results, including those from the model mitigating gender and race simultaneously, are provided in the Supplementary." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.803, + 0.486, + 0.9 + ], + "angle": 0, + "content": "For evaluating bias mitigation, we use a CLIP classifier to categorize each generated image into attribute-defined groups in \\(\\mathcal{A}_C\\) and measure attribute balance using modified deviation ratio [15], which is designed to handle an arbitrary number of attributes. The deviation ratio (DR) is calculated as \\(\\max_{a_i\\in \\mathcal{A}_C}\\frac{|N_{a_i} / N - 1 / |\\mathcal{A}_C||}{1 - 1 / |\\mathcal{A}_C|}\\), where \\(N\\) and \\(N_{a_i}\\) indicate" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.803, + 0.907, + 0.865 + ], + "angle": 0, + "content": "the total number of generated images and images classified as belonging to class \\( a_{i} \\), respectively. The lower DR indicates a better balance of attributes. We compute scores for each occupation and average them to obtain the final score." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.909, + 0.903 + ], + "angle": 0, + "content": "For non-target attribute preservation scores, we generate image pairs using the original Stable Diffusion (SD) [1] and" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "17572" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.089, + 0.482, + 0.358 + ], + "angle": 0, + "content": "
MethodFID ↓CLIP-T ↑
-Original SD-26.17
GenderConcept Algebra [28]2.4126.02
UCE [11]11.6525.17
Finetuning Diffusion [26]1.9225.79
Interpret Diffusion [15]15.7824.80
EFA (Ours)0.2326.03
RaceConcept Algebra [28]2.4826.04
UCE [11]5.1626.13
Finetuning Diffusion [26]4.1225.78
Interpret Diffusion [15]21.6623.82
EFA (Ours)0.1326.21
G.R.Concept Algebra [28]2.5326.03
UCE [11]6.5725.71
Finetuning Diffusion [26]4.2625.71
Interpret Diffusion [15]41.6522.45
EFA (Ours)0.4526.20
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.367, + 0.483, + 0.41 + ], + "angle": 0, + "content": "Table 2. Evaluation of model preservation using COCO-no-person. EFA outperforms the baselines in preserving the generation capability of the original model." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.433, + 0.484, + 0.63 + ], + "angle": 0, + "content": "our approach with the same random seed to evaluate how well non-target attributes are preserved while modifying target attributes. First, we segment the non-human regions in the original image and measure how well they are retained in the image generated by ours using Peak Signal-to-Noise Ratio (PSNR) and Learned Perceptual Image Patch Similarity (LPIPS) [29]. To assess overall image preservation, including contextual elements within human regions (e.g., a doctor's stethoscope or white coat), we compute the DINO score, employing DINO's ability to capture low-level and semantically meaningful visual details. The score measures the cosine similarity between the DINO [5] embeddings of images generated by SD and our method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.637, + 0.484, + 0.818 + ], + "angle": 0, + "content": "Lastly, for assessing model preservation, we use COCO-no-person dataset. To measure text fidelity, we use the CLIP-T score, which computes the cosine similarity between CLIP embeddings of the input prompt and the generated images. Additionally, Fréchet Inception Distance (FID) [12] is used to evaluate image quality and diversity. The FID scores are computed against the generation outputs of the original SD. Unlike non-target attribute preservation scores, this evaluation ensures that our method does not degrade the generation capability of the pretrained model when bias mitigation is not required. Note that bold text in all tables indicates the best score." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.9 + ], + "angle": 0, + "content": "Implementation details. The experiments were conducted based on SD v1.5 followed by previous works [10, 20, 26, 31]. For evaluation with WinoBias, we generated 160 images per occupation, using random seeds where the original SD has a detectable face." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.09, + 0.907, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.234, + 0.907, + 0.293 + ], + "angle": 0, + "content": "Figure 5. Attention map visualization when mitigating gender bias using EFA. (a) and (b) show the results of applying \\(\\mathrm{EFA}_{\\mathrm{gender}}^{\\mathrm{female}}\\) and \\(\\mathrm{EFA}_{\\mathrm{gender}}^{\\mathrm{male}}\\), respectively. \\(<\\mathrm{sos}>\\) and \\(<\\mathrm{eos}>\\) denote special tokens that mark the beginning and end of a text sequence, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.304, + 0.71, + 0.32 + ], + "angle": 0, + "content": "4.2. Experimental results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.326, + 0.907, + 0.568 + ], + "angle": 0, + "content": "Bias mitigation and non-target attribute preservation. Table 1 presents the comparative experimental results of our method and baseline models using the WinoBias dataset. Overall, models struggle to maintain non-target attributes more in \\(\\mathcal{T}_{\\mathrm{complex}}\\) than \\(\\mathcal{T}_{\\mathrm{basic}}\\). Concept Algebra exhibits low performance in bias mitigation. Both UCE [11] and Fine-tuning Diffusion [26] modify model parameters directly, making it challenging to preserve the non-target attributes consistent with the original SD. Interpret Diffusion [15] demonstrates better bias mitigation performance compared to other baselines. However, it exhibits low non-target attribute preservation scores due to its uniform application of fixed semantic vectors without spatial awareness. In contrast, our input-aware architecture adaptively enhances target attributes based on spatial context, enabling effective preservation of non-target attributes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.906, + 0.658 + ], + "angle": 0, + "content": "Our proposed method outperforms existing baselines by achieving a lower DR while having substantially higher non-target attribute preservation scores. The high preservation scores validate the effectiveness of our approach, showing that the model successfully learns to preserve regions corresponding to non-target attributes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.885 + ], + "angle": 0, + "content": "We hypothesize that the effectiveness in bias mitigation stems from the combination of two key factors: (1) the rich semantic information encoded in text embeddings, which have been pretrained on large-scale datasets, and (2) the capability of U-Net to faithfully reflect these embeddings in the generated output. By utilizing pretrained text embeddings without altering their inherent semantics, our approach ensures that target attributes are seamlessly integrated into the generation process, while U-Net's capacity to translate these embeddings into visual features enables precise control over attribute distributions. The extended experiments on Concept Algebra, provided in the supplementary material, corroborate our hypothesis. Consequently, our method successfully steers the distribution of target attributes while preserving non-target attributes." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Fig. 4 demonstrates that our method effectively miti-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17573" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.089, + 0.48, + 0.19 + ], + "angle": 0, + "content": "
BiasNon-target attribute P.
DR ↓PSNR ↑LPIPS ↓DINO ↑
Original SD0.71--
EFA w/o Ltrg0.0727.760.05270.890
EFA w/o seg. mask0.0429.310.05620.888
EFA (Ours)0.0632.520.04110.916
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.201, + 0.483, + 0.245 + ], + "angle": 0, + "content": "Table 3. Effectiveness of the regularization loss and the segmentation mask in bias mitigation and non-target attribute preservation. P. and seg. indicate preservation and segmentation, respectively." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.261, + 0.483, + 0.352 + ], + "angle": 0, + "content": "gates societal bias, generating outputs that represent a diverse range of gender and race. Furthermore, by leveraging regularization terms and utilizing text embeddings disentangled from other attributes, our method successfully preserves non-target attributes, such as overall layout, pose, and background, while effectively mitigating bias." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.353, + 0.483, + 0.504 + ], + "angle": 0, + "content": "Model preservation. The experimental results on model preservation are presented in Table 2. Our method achieves a significantly lower FID compared to baseline approaches, indicating superior preservation of both the quality and diversity of SD's outputs. Furthermore, our approach yields CLIP-T scores comparable to those of SD, effectively maintaining the text fidelity of the original model by preserving non-target attributes. Qualitative results on model preservation with COCO-no-person and additional experiments using COCO-30K are provided in the Supplementary." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.52, + 0.188, + 0.536 + ], + "angle": 0, + "content": "5. Analysis" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.545, + 0.483, + 0.562 + ], + "angle": 0, + "content": "5.1. Effect of regularization and segmentation mask" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.483, + 0.645 + ], + "angle": 0, + "content": "As shown in Table 3, we conduct an ablation study to examine the effects of regularization losses and segmentation masks on non-target attribute preservation. For the experiments, we train models to mitigate gender bias, and the evaluation is performed using \\(\\mathcal{T}_{\\mathrm{basic}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.483, + 0.902 + ], + "angle": 0, + "content": "When training without segmentation masks, we exclude \\(\\mathcal{L}_{\\mathrm{reg}}^{\\mathrm{cf}}\\) in training and apply \\(\\mathcal{L}_{\\mathrm{recon}}\\) to the entire image. Since there is no constraint limiting target attribute enhancement, we observe a slight improvement in bias mitigation performance. However, without segmentation masks to distinguish human and non-human regions during training, EFA struggles to localize them precisely, leading to reduced preservation of non-target attributes at inference. Nonetheless, EFA still outperforms baselines even without segmentation masks. We conjecture that this is because counterfactual attributes are primarily activated in bias-relevant regions due to their semantic similarity in cross-attention, which encourages EFA to enhance the target attribute in those regions while ignoring the counterfactual prompt's influence. This allows EFA to approximate the localization of bias-relevant regions without relying on fine-grained masks of the target attribute. Also, when \\(\\mathcal{L}_{\\mathrm{reg}}^{\\mathrm{trg}}\\) is not applied, the" + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.089, + 0.905, + 0.178 + ], + "angle": 0, + "content": "
resolution of inputBiasNon-target attribute P.
DR ↓PSNR ↑LPIPS ↓DINO ↑
16×16 (Ours)0.0632.520.04110.916
16×16 - 32×320.0631.790.04070.915
16×16 - 64×640.0630.650.04380.911
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.187, + 0.907, + 0.229 + ], + "angle": 0, + "content": "Table 4. Impact of input resolution employed by EFA. Extending EFA to higher-resolution layers compromises the preservation of non-target attributes. P. denotes preservation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.238, + 0.905, + 0.329 + ], + "angle": 0, + "content": "non-target attribute preservation performance deteriorates, even though the bias mitigation performance remains similar. These results highlight the necessity of the regularization mechanisms to preserve non-target attributes across diverse scenarios where the target attribute appears at different extents in the input." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.338, + 0.887, + 0.354 + ], + "angle": 0, + "content": "5.2. Analyzing the influence of resolution in EFA" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.36, + 0.907, + 0.542 + ], + "angle": 0, + "content": "To investigate the impact of input resolution, we integrated EFA into additional layers within the up-blocks while training a model for gender bias mitigation. The performance on \\(\\mathcal{T}_{\\mathrm{basic}}\\) is reported in Table 4, where the row highlighted in gray represents our proposed method. When EFA is applied to higher-resolution features, the deviation ratio remains nearly unchanged, while PSNR and DINO scores exhibit a slight decline. This suggests that lower-resolution layers contain richer semantic information, which is crucial for attribute modification. Moreover, extending EFA to multiple layers inevitably introduces additional noise, potentially compromising the preservation of non-target attributes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.55, + 0.764, + 0.567 + ], + "angle": 0, + "content": "5.3. Attention map visualization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.572, + 0.906, + 0.664 + ], + "angle": 0, + "content": "In Fig. 5, we visualize the attention maps predicted by EFA when the target attribute is enhanced. EFA assigns high values to the human region, effectively enabling gender transformation while preserving the generation capability of the original model, as the impact on accessories, clothing style, and background details remains minimal." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.677, + 0.634, + 0.693 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.702, + 0.907, + 0.9 + ], + "angle": 0, + "content": "In this work, we addressed attribute entanglement in diffusion-based T2I models, where bias mitigation often distorts non-target attributes, leading to unintended distribution shifts. We propose EFA, which mitigates human-centric biases while preserving non-target attributes. During training, EFA is guided to attend regions with high semantic relevance to counterfactual prompts, constrained by human segmentation masks. At inference, EFA randomly samples a target attribute and alters the cross-attention in selected layers to ensure a fair distribution of target attributes in the outputs. Experimental results verify that EFA outperforms prior methods in mitigating bias while preserving the original model's output distribution and generative capacity." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "17574" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.486, + 0.256 + ], + "angle": 0, + "content": "This work was supported by Institute for Information & communications Technology Planning & Evaluation(IITP) grant funded by the Korea government(MSIT) (RS-2019-II190075, Artificial Intelligence Graduate School Program(KAIST)), Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.RS-2021-II212068, Artificial Intelligence Innovation Hub), and the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT)(No. 2022R1A5A7083908)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.282, + 0.188, + 0.298 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.307, + 0.482, + 0.348 + ], + "angle": 0, + "content": "[1] Stability AI. Stable diffusion. https://huggingface.co/runwayml/stable-diffusion-v1-5, 2022. [Online; accessed 3-November-2023]. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.351, + 0.484, + 0.434 + ], + "angle": 0, + "content": "[2] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. https://cdn.openai.com/papers/dall-e-3.pdf, 2023. [Online; accessed 4-December-2024]. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.435, + 0.484, + 0.532 + ], + "angle": 0, + "content": "[3] Federico Bianchi, Pratyusha Kalluri, Esin Durmus, Faisal Ladhak, Myra Cheng, Debora Nozza, Tatsunori Hashimoto, Dan Jurafsky, James Zou, and Aylin Caliskan. Easily accessible text-to-image generation amplifies demographic stereotypes at large scale. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, pages 1493-1504, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.534, + 0.484, + 0.602 + ], + "angle": 0, + "content": "[4] Manuel Brack, Felix Friedrich, Dominik Hintersdorf, Lukas Struppek, Patrick Schramowski, and Kristian Kersting. Sega: Instructing text-to-image models using semantic guidance. Advances in Neural Information Processing Systems, 36: 25365-25389, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.604, + 0.484, + 0.674 + ], + "angle": 0, + "content": "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9650-9660, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.675, + 0.484, + 0.745 + ], + "angle": 0, + "content": "[6] Aditya Chinchure, Pushkar Shukla, Gaurav Bhatt, Kiri Salij, Kartik Hosanagar, Leonid Sigal, and Matthew Turk. Tibet: Identifying and evaluating biases in text-to-image generative models. In European Conference on Computer Vision, pages 429-446. Springer, 2025. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.746, + 0.484, + 0.815 + ], + "angle": 0, + "content": "[7] Jaemin Cho, Abhay Zala, and Mohit Bansal. Dall-eval: Probing the reasoning skills and social biases of text-to-image generation models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3043-3054, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.817, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[8] Ching-Yao Chuang, Varun Jampani, Yuanzhen Li, Antonio Torralba, and Stefanie Jegelka. Debiasing vision-language models via biased prompts. arXiv preprint arXiv:2302.00070, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.873, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[9] Stefan Elfwing, Eiji Uchibe, and Kenji Doya. Sigmoid-weighted linear units for neural network function approxi" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.307, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "mation in reinforcement learning. CoRR, abs/1702.03118, 2017.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.191 + ], + "angle": 0, + "content": "[10] Felix Friedrich, Manuel Brack, Lukas Struppek, Dominik Hintersdorf, Patrick Schramowski, Sasha Luccioni, and Kristian Kersting. Fair diffusion: Instructing text-to-image generation models on fairness. arXiv preprint at arXiv:2302.10893, 2023. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.906, + 0.262 + ], + "angle": 0, + "content": "[11] Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzyńska, and David Bau. Unified concept editing in diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5111-5120, 2024. 1, 2, 5, 6, 7, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.264, + 0.906, + 0.332 + ], + "angle": 0, + "content": "[12] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[13] Eunji Kim, Siwon Kim, Chaehun Shin, and Sungroh Yoon. De-stereotyping text-to-image models through prompt tuning. 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.378, + 0.906, + 0.433 + ], + "angle": 0, + "content": "[14] Mingi Kwon, Jaeseok Jeong, and Youngjung Uh. Diffusion models already have a semantic latent space. In The Eleventh International Conference on Learning Representations, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.435, + 0.906, + 0.504 + ], + "angle": 0, + "content": "[15] Hang Li, Chengzhi Shen, Philip Torr, Volker Tresp, and Jindong Gu. Self-discovering interpretable diffusion latent directions for responsible text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 3, 5, 6, 7, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.505, + 0.906, + 0.589 + ], + "angle": 0, + "content": "[16] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer vision-ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part v 13, pages 740-755. Springer, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.591, + 0.906, + 0.658 + ], + "angle": 0, + "content": "[17] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.661, + 0.906, + 0.73 + ], + "angle": 0, + "content": "[18] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.906, + 0.787 + ], + "angle": 0, + "content": "[19] Hadas Orgad, Bahjat Kawar, and Yonatan Belinkov. Editing implicit assumptions in text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7053-7061, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.906, + 0.871 + ], + "angle": 0, + "content": "[20] Rishubh Parihar, Abhijnya Bhat, Abhipsa Basu, Saswat Mallick, Jogendra Nath Kundu, and R. Venkatesh Babu. Balancing act: Distribution-guided debiasing in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6668-6678, 2024. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.906, + 0.902 + ], + "angle": 0, + "content": "[21] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image gen" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17575" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "eration with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.217 + ], + "angle": 0, + "content": "[22] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dólar, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.483, + 0.274 + ], + "angle": 0, + "content": "[23] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[24] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. 35:36479-36494, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.482, + 0.428 + ], + "angle": 0, + "content": "[25] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. 35:25278-25294, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.431, + 0.482, + 0.487 + ], + "angle": 0, + "content": "[26] Xudong Shen, Chao Du, Tianyu Pang, Min Lin, Yongkang Wong, and Mohan Kankanhalli. Finetuning text-to-image diffusion models for fairness. In The Twelfth International Conference on Learning Representations, 2024. 2, 6, 7, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.488, + 0.483, + 0.557 + ], + "angle": 0, + "content": "[27] Robik Shrestha, Yang Zou, Qiuyu Chen, Zhiheng Li, Yusheng Xie, and Siqi Deng. Fairrag: Fair human generation via fair retrieval augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11996-12005, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.559, + 0.482, + 0.627 + ], + "angle": 0, + "content": "[28] Zihao Wang, Lin Gui, Jeffrey Negrea, and Victor Veitch. Concept algebra for (score-based) text-controlled generative models. In Proceedings of the 37th International Conference on Neural Information Processing Systems, pages 35331-35349, 2023. 6, 7, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.629, + 0.482, + 0.697 + ], + "angle": 0, + "content": "[29] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.7, + 0.482, + 0.794 + ], + "angle": 0, + "content": "[30] Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, and Kai-Wei Chang. Gender bias in coreference resolution: Evaluation and debiasing methods. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 15–20, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.798, + 0.482, + 0.866 + ], + "angle": 0, + "content": "[31] Junlei Zhou, Jiashi Gao, Xiangyu Zhao, Xin Yao, and Xuetao Wei. Association of objects may engender stereotypes: Mitigating association-engendered stereotypes in text-to-image generation. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17576" + } + ] +] \ No newline at end of file diff --git a/2025/Fair Generation without Unfair Distortions_ Debiasing Text-to-Image Generation with Entanglement-Free Attention/6ef80803-cc14-446f-a75b-f2593f51cc3f_origin.pdf b/2025/Fair Generation without Unfair Distortions_ Debiasing Text-to-Image Generation with Entanglement-Free Attention/6ef80803-cc14-446f-a75b-f2593f51cc3f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7c56296199d27e0213ec6dec366240ee0c0f15b8 --- /dev/null +++ b/2025/Fair Generation without Unfair Distortions_ Debiasing Text-to-Image Generation with Entanglement-Free Attention/6ef80803-cc14-446f-a75b-f2593f51cc3f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9862fc8f8bd0365986aca86340fc8a012d96b9f1bd6958698825a5bd4b05ccb7 +size 7873589 diff --git a/2025/Fair Generation without Unfair Distortions_ Debiasing Text-to-Image Generation with Entanglement-Free Attention/full.md b/2025/Fair Generation without Unfair Distortions_ Debiasing Text-to-Image Generation with Entanglement-Free Attention/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3db1053b2704f22ebd129d576e12395468129ed2 --- /dev/null +++ b/2025/Fair Generation without Unfair Distortions_ Debiasing Text-to-Image Generation with Entanglement-Free Attention/full.md @@ -0,0 +1,301 @@ +# Fair Generation without Unfair Distortions: Debiasing Text-to-Image Generation with Entanglement-Free Attention + +Jeonghoon Park* KAIST + +jeonghoon_park@kaist.ac.kr + +Jaeseong Lee +Yonsei University + +jasonlee1995@yonsei.ac.kr + +Juyoung Lee\* Kakao Corp. + +zealota11@gmail.com + +Jaegul Choo KAIST + +jchoo@kaist.ac.kr + +Chaeyeon Chung KAIST + +cy_chung@kaist.ac.kr + +Jindong Gu University of Oxford + +jindong.gu@eng.ox.ac.uk + +# "face of successful man with his house" + +![](images/fbdcdfa0ee8a26eaf8ee496c218844d4a23628980a15b10eccd0da4d276fbc20.jpg) +Stable Diffusion + +![](images/f38018fff929b9db7940ef993dc65c926a2a2b202c1abf3235fe7208f2bc55e2.jpg) +UCE +Racial attributes changed Non-racial attributes preserved + +![](images/b696fb8b08e0855bfa572c64def597237035baf501c58fd3c69ab917cc3e5745.jpg) +Interpret Diffusion + +![](images/2db09a8ebc6f2a0d703ea0f190a6c708d472b1930d23fd51e25b697e478401c2.jpg) +EFA (Ours) +Racial attributes changed Non-racial attributes preserved +Figure 1. Generated images from the pretrained model (e.g., Stable Diffusion [1]), by previous debiasing methods (e.g., UCE [11] and Interpret Diffusion [15]), and ours. Previous debiasing methods often alter the non-target attributes (e.g., house) while modifying target bias attributes (e.g., racial attributes). In contrast, our approach preserves non-target attributes during bias mitigation (i.e., changing the racial attributes without modifying the house in the background), thereby preventing the introduction of unintended biases. The corresponding aligned images were generated using the same random seed. + +# Abstract + +Recent advancements in diffusion-based text-to-image (T2I) models have enabled the generation of high-quality and photorealistic images from text. However, they often exhibit societal biases related to gender, race, and socioeconomic status, thereby potentially reinforcing harmful stereotypes and shaping public perception in unintended ways. While existing bias mitigation methods demonstrate effectiveness, they often encounter attribute entanglement, where adjustments to attributes relevant to the bias (i.e., target attributes) unintentionally alter attributes unassociated with the bias (i.e., non-target attributes), causing undesirable distribution shifts. To address this challenge, we introduce Entanglement-Free Attention (EFA), a method that accurately incorporates target attributes (e.g., White, Black, and + +Asian) while preserving non-target attributes (e.g., background) during bias mitigation. At inference time, EFA randomly samples a target attribute with equal probability and adjusts the cross-attention in selected layers to incorporate the sampled attribute, achieving a fair distribution of target attributes. Extensive experiments demonstrate that EFA outperforms existing methods in mitigating bias while preserving non-target attributes, thereby maintaining the original model's output distribution and generative capacity. + +# 1. Introduction + +Diffusion-based text-to-image (T2I) models [2, 18, 21, 23, 24] have gained significant attention due to their ability to reflect user intent through text input, making them highly applicable across various domains. However, despite their + +advancements, these T2I models often perpetuate socially sensitive biases and stereotypes such as gender, race, or socioeconomic status [3, 7, 10]. Such societal biases can reinforce discriminatory perceptions of certain social groups and pose ethical and practical risks in real-world environments, where biased outputs can influence public perception and decision making. + +Previous works attempt to mitigate such biases via two primary approaches. One widely explored strategy is finetuning, where pretrained models are adjusted using additional training on curated datasets [11, 19, 26]. Despite its effectiveness, this method often entails substantial computational costs [26]. Also, it may result in a decline in image quality and diversity due to the relatively small size of the finetuning dataset compared to the original pretraining dataset [25]. Since this approach inherently modifies the parameters of the model, it may lead to unintended changes in the output distribution. An alternative line of research explores inference-time steering, which avoids modifying the pretrained model and instead applies adjustments only during inference [8, 10, 13, 15, 20, 27]. Although this approach effectively reduces bias while preserving the parameters of the original model, we observe that it still introduces unintended alterations in the generated outputs. + +Fig. 1 illustrates this issue by presenting images generated using the prompt "face of successful man with his house". The pretrained model correctly interprets the prompt by generating images of a man with a house but predominantly produces individuals from a limited racial background. When existing bias mitigation methods are applied, the representation of underrepresented racial attributes, such as Black and Indian individuals, is successfully improved. However, they also introduce unintended shifts in non-racial attributes. Specifically, they alter the material and architectural style of the house, associating certain racial groups with impoverished or deteriorated living conditions, thus reinforcing misleading associations between social groups and socioeconomic status. This finding suggests that previous methods may involve attribute entanglement between attributes relevant to the target bias (i.e., target attribute, such as racial attributes) and attributes unassociated with the target bias (i.e., non-target attribute, such as the appearance of the house), potentially leading to unintended alterations in the non-target attributes during the bias mitigation process. Without careful disentanglement, bias mitigation can unintentionally alter contextual details, degrade output diversity, and even introduce new biases. + +To address this issue, we propose Entanglement-Free Attention (EFA), a method that accurately incorporates target attributes while preserving non-target attributes. In this paper, we focus on human-centric bias, which plays a crucial role in shaping societal perceptions. Our EFA operates by leveraging predefined target attributes (e.g., White, Black, + +Asian, and Indian) that contribute to bias (e.g., race). To guide the model to focus on semantically relevant regions (e.g., face and body), EFA is trained to enhance target attributes in regions that exhibit high semantic relevance to the counterfactual prompts (i.e., prompts with alternative attributes other than the target), while human segmentation masks further constrain its effect to human regions. By employing both target and counterfactual prompts in training, EFA learns to dynamically adjust the enhancement strength based on the presence of the target attribute, applying stronger modifications when it is underrepresented. During inference, to ensure a fair distribution of target attributes in the generated output, one attribute is randomly sampled with equal probability. EFA then enhances the corresponding features in bias-relevant regions without relying on human-provided masks, effectively disentangling target and non-target attributes while preserving the original model's generative capacity and contextual consistency. + +In summary, our key contributions are as follows: + +- We identify the issue of attribute entanglement, where bias mitigation unintentionally alters non-target attributes, leading to unintended distribution shifts in diffusion-based T2I models. +- We propose EFA, a novel debiasing approach that decouples target and non-target attributes, ensuring that bias mitigation is applied only to the intended target attributes while preserving the integrity of non-target attributes. +- Quantitative and qualitative experiments demonstrate that our approach successfully mitigates bias while maximally preserving the generation capability of the original model, outperforming previous methods. + +# 2. Related works + +Existing studies on mitigating bias in T2I generative models have mainly followed two approaches. The first approach involves directly modifying the parameters of diffusion-based T2I models. For instance, Orgad et al. [19] and Gandikota et al. [11] proposed a method to adjust the weights of the cross-attention module to alleviate the bias. Similarly, Shen et al. [26] finetune the text encoder to alleviate face-related biases (e.g., gender, race). However, since finetuning is conducted on a significantly smaller dataset compared to the pretraining dataset, such approach may potentially degrade original output quality or diversity. Furthermore, the approach of Shen et al. [26] requires substantial computational costs for training the backbone model. + +Another line of research has explored inference-time steering approaches, where the pretrained model remains unchanged while adjustments are applied during inference. Friedrich et al. [10] proposed a bias mitigation framework that employs semantic guidance [4] to manipulate the target attributes by referring to registered information of a concept (e.g., nurse) and its bias (e.g., gender). However, this cannot + +![](images/d76059e9a7a449ec9996a39c21cb33d42d1fea70edb5c08f9100cdc71f2b9dd9.jpg) +Figure 2. Overview of our method. (a): EFA is applied to selected layers and trained to reconstruct the image $\mathbf{x}_0$ (a-1) when given $p_a$ , the prompt containing the target attribute $a$ , or (a-2) its counterfactual prompt, $p_a^{\mathrm{cf}}$ . We leverage a human segmentation mask $M$ to minimize the effect of EFA on the attention map of non-target attribute regions. (b): During the inference, our method generates outputs with a fair distribution of target attributes using the EFA of the randomly sampled target attribute. + +mitigate bias for unregistered concepts, limiting its practicality. Chuang et al. [8] have developed a method to remove bias from text embeddings, and Kim et al. [13] trains the soft prompts for mitigating bias. Beyond employing the text embeddings, Shrestha et al. [27] introduced a technique that projects a reference image into text-compatible tokens, allowing the model to leverage the features of external images, such as gender or race for image generation. Additionally, Parihar et al. [20] and Li et al. [15] proposed methods to manipulate the h-space [14] of the model during inference. While these methods aim to mitigate bias without directly altering the backbone parameters, they often struggle to preserve the original context of the generated images. + +Recently, Zhou et al. [31] addressed association-engendered stereotypes arising from co-occurrence of multiple objects in an image. They rely on a CLIP model trained to predict biases from prompts, requiring prior knowledge of the biases to be mitigated. In contrast, our approach aims to address subtle biases induced by contextual variations beyond the predefined association bias definitions. For instance, if a particular social group is predominantly depicted in low-resource work environments, while another is shown in corporate office settings, such context variations can reinforce cognitive biases, subtly implying disparities in professional status. Furthermore, it is impractical to anticipate all possible forms of such bias in advance. + +# 3. Entanglement-free attention + +# 3.1. Overview + +Without loss of generality, we assume a given target bias $C$ (e.g., gender) along with its corresponding set of target + +attributes $\mathcal{A}_C = \{a_1, a_2, \ldots, a_n\}$ (e.g., 'female', 'male') whose elements should be uniformly distributed in the outputs. Then, our approach aims to alleviate the bias $C$ by adjusting the distribution of target attributes in $\mathcal{A}_C$ while preserving the non-target attributes such as scene context and surrounding objects. + +To achieve this, we introduce EFA, which leverages the principle of cross-attention mechanism that assigns desired features to appropriate locations. Each target attribute $a_{i} \in \mathcal{A}_{C}$ is handled by a dedicated EFA module (i.e., $\mathrm{EFA}_{C}^{a_{i}}$ ), ensuring faithful representation of $a_{i}$ in the outputs while minimizing its influence on non-target attribute regions. During inference, to enforce a fair distribution of target attributes in the generated output, we randomly sample an attribute $a_{i}$ from $\mathcal{A}_{C}$ with equal probability and incorporate its representation using the corresponding EFA (Fig. 2(b)). + +To enhance the fidelity of $a_i$ , $\mathrm{EFA}_C^{a_i}$ is trained to reconstruct images depicting $a_i$ when its representation is subtle by introducing counterfactual attributes $a_i^{\mathrm{cf}} \in \mathcal{A}_C \setminus \{a_i\}$ through counterfactual prompts. EFA is guided to enhance the target attribute in areas exhibiting strong semantic relevance to the counterfactual prompt through the cross-attention module. Additionally, human segmentation masks are used during training to constrain EFA's influence to the human region. This ensures that $\mathrm{EFA}_C^{a_i}$ enhances the fidelity of $a_i$ while preventing unintended modifications to non-target attributes. + +# 3.2. Architecture of EFA + +The original cross-attention module takes an intermediate spatial feature $\mathbf{z}_t$ , which is derived by from a noised image $\mathbf{x}_t$ and a prompt $p$ using the denoising model $\phi$ (i.e., + +$\mathbf{z}_t = \phi (\mathbf{x}_t,p)$ . $\mathbf{z}_t$ is transformed into a query representation using the mapping layer $W_{q}$ . The text prompt embeddings $\pi (p)$ are projected into key and value representations through $W_{k}$ and $W_{v}$ , respectively. The output of the cross-attention layer is formulated as: + +$$ +\operatorname {C r o s s - a t t e n t i o n} (\mathbf {z} _ {t}) = \operatorname {s o f t m a x} \left(\frac {Q K ^ {\top}}{\sqrt {d}}\right) V, +$$ + +where $d$ indicates the dimension of the query and key vectors, and $Q = \mathbf{z}_tW_q$ , $K = \pi(p)W_k$ , and $V = \pi(p)W_v$ denote the query, key, and value matrices, respectively. + +As illustrated in Fig. 3, EFA modifies the cross-attention to integrate the desired attribute $a_{i} \in \mathcal{A}_{C}$ into the output while preserving the original attention structure. Specifically, EFA takes the intermediate feature $\mathbf{z}_t$ and predicts the additional attention values that determine where and to what extent $a_{i}$ should be enhanced. This is achieved using the attention value predictor (AP), a lightweight module composed of three convolutional layers and two SiLU [9] activation functions. Further details of the architecture are provided in the Supplementary. + +The predicted attention value (i.e., $\mathrm{AP}_{a_i}(\mathbf{z}_t)$ ) are concatenated with the original cross-attention's attention values (i.e., $\frac{QK^{\top}}{\sqrt{d}}$ ), and the softmax function is applied. After applying softmax, the original value vectors from the user prompt (i.e., $V$ ) and the value vectors of $a_i$ are weighted accordingly, as in the standard cross-attention operation. Formally, EFA incorporates $a_i$ into the cross-attention as follows: + +$$ +\operatorname {E F A} _ {C} ^ {a _ {i}} (\mathbf {z} _ {t}) = \operatorname {s o f t m a x} \left(\left[ \frac {Q K ^ {\top}}{\sqrt {d}}, \mathrm {A P} _ {a _ {i}} (\mathbf {z} _ {t}) \right]\right) [ V, V _ {a _ {i}} ], +$$ + +where $V_{a_i} = \pi(p_{a_i}) W_v$ , $p_{a_i}$ indicates the text prompt representing $a_i$ , and $[\cdot, \cdot]$ indicates the concatenation. + +When building an $\mathrm{EFA}_C$ for a target bias $C$ , all AP modules share the same convolutional backbone while maintaining distinct output channels designated for each attribute in $\mathcal{A}_C$ . This design allows AP models to be effectively trained together, ensuring that EFAs of target attributes for a given bias are jointly learned while avoiding the inefficiency of training multiple independent APs. + +# 3.3. Training + +Dataset construction. To train $\mathrm{EFA}_C^{a_i}$ to represent a target attribute $a_{i}$ (e.g., female) while preserving non-target attributes, we create a dataset $\mathcal{D}_{a_i}$ that captures $a_{i}$ and provides pseudo supervision on the human regions that contain $a_{i}$ . Specifically, we generate images using the original model with prompts explicitly describing $a_{i}$ (e.g., "a female person"). For each image, we extract a human segmentation mask $M$ using Grounded SAM2 [17, 22], assigning 1 to human regions and 0 elsewhere. Since this study focuses on + +![](images/02f98c40ffc1e6cebfb3e755dad271f7b93be770eb596d9a6aa54ab44172384c.jpg) +Figure 3. Architecture of EFA. We denote $a_i \in \mathcal{A}_C$ as $a$ in this figure for notational simplicity. + +human-centric bias, we utilize a human segmentation mask to serve as the basis for regulating the influence of $a_{i}$ during training. The resulting $\mathcal{D}_{a_i}$ consists of pairs of images and their segmentation masks. + +Training EFA. At inference, the extent of the features that need to be adjusted by EFA to faithfully reflect $a_{i}$ may vary across inputs. For instance, with the prompt "nurse", intermediate features may exhibit a high level of femininity, requiring minimal adjustment by $\mathrm{EFA}_{\text{gender}}^{\text{female}}$ . In contrast, intermediate features of the prompt "doctor" may have stronger masculinity, requiring enhanced representation of female by $\mathrm{EFA}_{\text{gender}}^{\text{female}}$ to incorporate 'female' into the output. To handle this variability, $\mathrm{EFA}_{C}^{a_{i}}$ is trained in two different scenarios: (1) when intermediate features contain the attribute $a_{i}$ (e.g., female), and (2) when intermediate features contain its counterfactual attribute $a_{i}^{\mathrm{cf}} \in \mathcal{A}_{C} \setminus \{a_{i}\}$ (e.g., male). We denote $a_{i}$ as $a$ throughout this section for simplicity. + +Let $\mathbf{x}_0$ be an image and $M$ be its corresponding segmentation mask, sampled from $\mathcal{D}_a$ , where $a$ denotes a target attribute (e.g., female). In the first scenario, where the intermediate features already encode $a$ , $\mathrm{EFA}_C^a$ does not need to enhance $a$ but should instead preserve the original cross-attention output to maintain the non-target attributes. To enforce this behavior, we provide the model with a paired input: a prompt $p_a$ explicitly specifying $a$ (e.g., "female person") and a noised image $\mathbf{x}_t$ that inherently contains $a$ , as depicted in Fig. 2 (a-1). + +To prevent unnecessary alterations, we introduce an L1 regularization term that encourages the attention values associated with $V_{a}$ (e.g., attention values associated with the value vectors of "female") to be zero. This regularization loss is defined as: + +$$ +\mathcal {L} _ {\text {r e g}} (\mathbf {z} _ {t}, W) = \left\| W \odot A _ {a} \left(\mathbf {z} _ {t}\right) \right\| _ {1}, +$$ + +where $\mathbf{z}_t = \phi (\mathbf{x}_t,p_a)$ $W$ is a spatial mask used to constrain the effect of the loss, and $\odot$ represents the element-wise multiplication. Also, $A_{a}(\mathbf{z}_{t})$ denotes the attention values (after softmax) associated with the value vectors $V_{a}$ given $\mathbf{z}_t$ . More precisely, when $A(\mathbf{z}_t) =$ softmax $([\frac{QK^{\top}}{\sqrt{d}},\mathrm{AP}_a(\mathbf{z}_t)])$ , we define $A_{a}(\mathbf{z}_{t})$ as the submatrix of $A(\mathbf{z}_t)$ obtained by selecting values indexed by $I_{V_a}$ + +where $I_{V_a}$ denotes the set of indices corresponding to $V_a$ . + +In this scenario, we set $W = \mathbf{1}$ , a mask of ones, to apply the loss over all spatial positions. This regularization loss minimizes unnecessary modifications to the original cross-attention output introduced by $\mathrm{EFA}_C^a$ . Hereafter, we denote $\mathcal{L}_{\mathrm{reg}}(\mathbf{z}_t,\mathbf{1})$ as $\mathcal{L}_{\mathrm{reg}}^{\mathrm{trg}}$ for brevity. + +In the second scenario, the intermediate features contain a counterfactual attribute $a^{\mathrm{cf}}$ (e.g., male) induced by bias of prompt (e.g., "photo of a doctor" implicitly favoring male representation). Here, $\mathrm{EFA}_C^a$ should suppress $a^{\mathrm{cf}}$ and enhance $a$ (e.g., female) while preserving non-target attributes. To simulate this, we train $\mathrm{EFA}_C^a$ to counteract counterfactual features introduced during the denoising process of a diffusion-based T2I model, which generates clean images by iteratively denoising from pure noise. + +To encourage representation of $a$ in human-related regions, we define a reconstruction loss leveraging the segmentation mask $M$ to constrain its effect on human areas. The objective is to predict the added noise $\epsilon$ given $\mathbf{x}_t$ and a counterfactual prompt $p_a^{\mathrm{cf}}$ , formulated as follows: + +$$ +\mathcal {L} _ {\text {r e c o n}} = \mathbb {E} _ {\mathbf {x} _ {0}, t, p _ {a} ^ {\mathrm {c f}}, \epsilon \sim N (0, 1)} \left[ \| M \odot (\epsilon - \epsilon_ {\theta} (\mathbf {x} _ {t}, t, p _ {a} ^ {\mathrm {c f}})) \| _ {2} ^ {2} \right], +$$ + +where $\epsilon_{\theta}$ denotes the model with $\mathrm{EFA}_C^a$ . Through the reconstruction process, $\mathrm{EFA}_C^a$ is encouraged to enhance $a$ in regions where cross-attention assigns high similarity to $a^{\mathrm{cf}}$ , reflecting their semantic relevance. Using human masks ensures that $a$ is accurately reconstructed in human regions while mitigating the influence of $a^{\mathrm{cf}}$ . + +Although $\mathcal{L}_{\mathrm{reg}}^{\mathrm{trg}}$ reduces unintended modifications, it does not explicitly confine its impact on human regions, potentially affecting non-target attributes. To address this, we impose an additional L1 regularization on the attention values corresponding to $V_{a}$ , applied outside the human region, as follows: + +$$ +\mathcal {L} _ {\mathrm {r e g}} ^ {\mathrm {c f}} = \mathcal {L} _ {\mathrm {r e g}} (\mathbf {z} _ {t} ^ {\mathrm {c f}}, \mathbf {1} - \bar {M}), +$$ + +where $\mathbf{z}_t^{\mathrm{cf}} = \phi (\mathbf{x}_t,p_a^{\mathrm{cf}})$ , and $\bar{M}$ indicates the binary mask resized to match the resolution of the input feature of $\mathrm{EFA}_C^a$ . + +Finally, the overall training loss is formulated as: + +$$ +\mathcal {L} _ {\mathrm {t o t}} = \mathcal {L} _ {\mathrm {r e c o n}} + \mathcal {L} _ {\mathrm {r e g}} ^ {*}, +$$ + +where $\mathcal{L}_{\mathrm{reg}}^{*} = \lambda_{1}\mathcal{L}_{\mathrm{reg}}^{\mathrm{trg}} + \lambda_{2}\mathcal{L}_{\mathrm{reg}}^{\mathrm{cf}}$ and $\lambda_{1}$ and $\lambda_{2}$ are hyperparameters to control the relative importance between the losses. During training, only the parameters of $\mathrm{AP}_a$ in $\mathrm{EFA}_C^a$ are updated. Note that EFA is applied to the upblock layers with an input resolution of $16\times 16$ , as lower-resolution features capture higher-level semantic information. The choice of applied layers is discussed in Section 5.2. The method of extending EFA to handle multiple biases is provided in the Supplementary. + +# 3.4. Inference + +To mitigate the bias $C$ and ensure fair generation, we randomly sample an attribute $a_{i}$ with an equal probability from + +$\mathcal{A}_C$ and apply the corresponding $\mathrm{EFA}_C^{a_i}$ to the original model. This approach enables users to easily control the attribute distribution by adjusting the sampling frequency, without additional training. Note that our method requires no masks at inference and generates debiased images directly from noise, offering a practical and efficient solution. + +# 4. Experiments + +# 4.1. Experimental settings + +Datasets. Following the previous studies [10, 11, 15], we mainly conducted experiments with gender and racial bias. Experiments with age bias are provided in Supplementary. The target attributes set of gender and racial bias is set as $\mathcal{A}_{\mathrm{gender}} = \{\mathrm{female}, \mathrm{male}\}$ and $\mathcal{A}_{\mathrm{race}} = \{\mathrm{White}, \mathrm{Black}, \mathrm{Asian}, \mathrm{Indian}\}$ , respectively. For multiple biases, the set of target attributes is defined as $\mathcal{A}_{\mathrm{gender} \times \mathrm{race}} = \mathcal{A}_{\mathrm{gender}} \times \mathcal{A}_{\mathrm{race}}$ . We adopt a binary gender perspective to evaluate our method, following prior works [11, 15], due to the limitations of existing datasets and evaluation protocols. Further details of the dataset construction are included in the Supplementary. + +Evaluation prompt. To assess bias mitigation performance, we utilized 36 occupations from the WinoBias [30] dataset, which includes stereotypical occupations based on US Department of Labor statistics by following previous studies [11, 15, 19]. We employed two types of prompt templates for image generation. The first set (i.e., $\mathcal{T}_{\mathrm{basic}}$ ) follows the prompts used in prior research [11, 15], formatted as “{\style} of {\occupation}.” The style category includes photo, image, picture, headshot, and portrait. The second set (i.e., $\mathcal{T}_{\mathrm{complex}}$ ) includes background components to enable evaluation with more complex non-target attributes. Specifically, prompts are structured as “{\style} of {\occupation} in {\background},” where the background is contextually relevant to the occupation. + +To evaluate the overall model preservation performance while ensuring non-target attributes remain unaffected, we assessed whether the model retained its original generation capability when generating images without human presence. Specifically, we filtered image-prompt pairs from the COCO-30K [16] dataset and constructed a new dataset, COCO-no-person, consisting of 14,926 pairs where no person is present in the corresponding images. + +Evaluation metrics. Our evaluation focuses on (1) bias mitigation, ensuring balanced attribute representation; (2) non-target attribute preservation, maintaining non-target attributes while adjusting target attributes; and (3) model preservation, maintaining the overall generative capability of the original model in non-human contexts. While model preservation also evaluates the preservation of non-target attributes, we use a distinct term to differentiate it from cases where target attributes should be enhanced. + +
\(T_{\text{basic}}\)\(T_{\text{complex}}\)
BiasNon-target attribute P.BiasNon-target attribute P.
DR ↓PSNR ↑LPIPS ↓DINO ↑DR ↓PSNR ↑LPIPS ↓DINO ↑
GenderOriginal SD0.71---0.71---
Concept Algebra [28]0.5921.100.11690.8230.6916.690.18520.834
UCE [11]0.3421.040.13740.7570.4916.780.22020.747
Finetuning Diffusion [26]0.4822.620.11660.8140.5619.270.16110.857
Interpret Diffusion [15]0.2617.180.22900.6160.3114.010.32110.618
EFA (Ours)0.0632.520.04110.9160.0629.700.04920.941
RaceOriginal SD0.60---0.55---
Concept Algebra [28]0.6421.470.11640.8390.5816.620.18880.841
UCE [11]0.2721.550.12610.7870.4117.810.18910.801
Finetuning Diffusion [26]0.2918.960.17950.7090.3716.460.22630.771
Interpret Diffusion [15]0.1616.870.24160.5840.2313.730.33380.584
EFA (Ours)0.0430.930.03530.9380.0628.550.04210.958
Gender × RaceOriginal SD0.56---0.50---
Concept Algebra [28]0.5120.120.13250.8050.4715.880.20310.818
UCE [11]0.1619.840.15750.7020.3216.790.21980.746
Finetuning Diffusion [26]0.2419.970.15700.7460.2817.240.20640.795
Interpret Diffusion [15]0.1616.910.24490.4660.2013.660.35290.475
EFA (Ours)0.0325.580.06840.8530.0523.780.07950.903
+ +Table 1. Comparison to baselines. P. is the abbreviation for preservation. The best values are in bold. Our method achieves superior bias mitigation performance while effectively preserving non-target attributes, significantly outperforming previous approaches. + +![](images/3bd84c11dd5af4ab6eff29155185078b65e8724f06573fb65eafe4fcf2dc240f.jpg) +Figure 4. Generated images from the original Stable Diffusion and ours. Ours successfully generates images with a fair distribution of the target bias (i.e., gender and race). (a) and (b) represent the images generated from the prompts describing a nurse in a hospital and a carpenter in a workshop, respectively. The corresponding aligned images were generated using the same random seed. Further qualitative results, including those from the model mitigating gender and race simultaneously, are provided in the Supplementary. + +![](images/f17e94095042218fd2ba9ade9e55a707081fdbc36383b0ad1dc484e406d53b43.jpg) + +For evaluating bias mitigation, we use a CLIP classifier to categorize each generated image into attribute-defined groups in $\mathcal{A}_C$ and measure attribute balance using modified deviation ratio [15], which is designed to handle an arbitrary number of attributes. The deviation ratio (DR) is calculated as $\max_{a_i\in \mathcal{A}_C}\frac{|N_{a_i} / N - 1 / |\mathcal{A}_C||}{1 - 1 / |\mathcal{A}_C|}$ , where $N$ and $N_{a_i}$ indicate + +the total number of generated images and images classified as belonging to class $a_{i}$ , respectively. The lower DR indicates a better balance of attributes. We compute scores for each occupation and average them to obtain the final score. + +For non-target attribute preservation scores, we generate image pairs using the original Stable Diffusion (SD) [1] and + +
MethodFID ↓CLIP-T ↑
-Original SD-26.17
GenderConcept Algebra [28]2.4126.02
UCE [11]11.6525.17
Finetuning Diffusion [26]1.9225.79
Interpret Diffusion [15]15.7824.80
EFA (Ours)0.2326.03
RaceConcept Algebra [28]2.4826.04
UCE [11]5.1626.13
Finetuning Diffusion [26]4.1225.78
Interpret Diffusion [15]21.6623.82
EFA (Ours)0.1326.21
G.R.Concept Algebra [28]2.5326.03
UCE [11]6.5725.71
Finetuning Diffusion [26]4.2625.71
Interpret Diffusion [15]41.6522.45
EFA (Ours)0.4526.20
+ +Table 2. Evaluation of model preservation using COCO-no-person. EFA outperforms the baselines in preserving the generation capability of the original model. + +our approach with the same random seed to evaluate how well non-target attributes are preserved while modifying target attributes. First, we segment the non-human regions in the original image and measure how well they are retained in the image generated by ours using Peak Signal-to-Noise Ratio (PSNR) and Learned Perceptual Image Patch Similarity (LPIPS) [29]. To assess overall image preservation, including contextual elements within human regions (e.g., a doctor's stethoscope or white coat), we compute the DINO score, employing DINO's ability to capture low-level and semantically meaningful visual details. The score measures the cosine similarity between the DINO [5] embeddings of images generated by SD and our method. + +Lastly, for assessing model preservation, we use COCO-no-person dataset. To measure text fidelity, we use the CLIP-T score, which computes the cosine similarity between CLIP embeddings of the input prompt and the generated images. Additionally, Fréchet Inception Distance (FID) [12] is used to evaluate image quality and diversity. The FID scores are computed against the generation outputs of the original SD. Unlike non-target attribute preservation scores, this evaluation ensures that our method does not degrade the generation capability of the pretrained model when bias mitigation is not required. Note that bold text in all tables indicates the best score. + +Implementation details. The experiments were conducted based on SD v1.5 followed by previous works [10, 20, 26, 31]. For evaluation with WinoBias, we generated 160 images per occupation, using random seeds where the original SD has a detectable face. + +![](images/46ccff1e1fafee023e5745c964c638a4fdd3f91a87d0a55789f7171846f88136.jpg) +Figure 5. Attention map visualization when mitigating gender bias using EFA. (a) and (b) show the results of applying $\mathrm{EFA}_{\mathrm{gender}}^{\mathrm{female}}$ and $\mathrm{EFA}_{\mathrm{gender}}^{\mathrm{male}}$ , respectively. $<\mathrm{sos}>$ and $<\mathrm{eos}>$ denote special tokens that mark the beginning and end of a text sequence, respectively. + +# 4.2. Experimental results + +Bias mitigation and non-target attribute preservation. Table 1 presents the comparative experimental results of our method and baseline models using the WinoBias dataset. Overall, models struggle to maintain non-target attributes more in $\mathcal{T}_{\mathrm{complex}}$ than $\mathcal{T}_{\mathrm{basic}}$ . Concept Algebra exhibits low performance in bias mitigation. Both UCE [11] and Fine-tuning Diffusion [26] modify model parameters directly, making it challenging to preserve the non-target attributes consistent with the original SD. Interpret Diffusion [15] demonstrates better bias mitigation performance compared to other baselines. However, it exhibits low non-target attribute preservation scores due to its uniform application of fixed semantic vectors without spatial awareness. In contrast, our input-aware architecture adaptively enhances target attributes based on spatial context, enabling effective preservation of non-target attributes. + +Our proposed method outperforms existing baselines by achieving a lower DR while having substantially higher non-target attribute preservation scores. The high preservation scores validate the effectiveness of our approach, showing that the model successfully learns to preserve regions corresponding to non-target attributes. + +We hypothesize that the effectiveness in bias mitigation stems from the combination of two key factors: (1) the rich semantic information encoded in text embeddings, which have been pretrained on large-scale datasets, and (2) the capability of U-Net to faithfully reflect these embeddings in the generated output. By utilizing pretrained text embeddings without altering their inherent semantics, our approach ensures that target attributes are seamlessly integrated into the generation process, while U-Net's capacity to translate these embeddings into visual features enables precise control over attribute distributions. The extended experiments on Concept Algebra, provided in the supplementary material, corroborate our hypothesis. Consequently, our method successfully steers the distribution of target attributes while preserving non-target attributes. + +Fig. 4 demonstrates that our method effectively miti- + +
BiasNon-target attribute P.
DR ↓PSNR ↑LPIPS ↓DINO ↑
Original SD0.71--
EFA w/o Ltrg0.0727.760.05270.890
EFA w/o seg. mask0.0429.310.05620.888
EFA (Ours)0.0632.520.04110.916
+ +gates societal bias, generating outputs that represent a diverse range of gender and race. Furthermore, by leveraging regularization terms and utilizing text embeddings disentangled from other attributes, our method successfully preserves non-target attributes, such as overall layout, pose, and background, while effectively mitigating bias. + +Model preservation. The experimental results on model preservation are presented in Table 2. Our method achieves a significantly lower FID compared to baseline approaches, indicating superior preservation of both the quality and diversity of SD's outputs. Furthermore, our approach yields CLIP-T scores comparable to those of SD, effectively maintaining the text fidelity of the original model by preserving non-target attributes. Qualitative results on model preservation with COCO-no-person and additional experiments using COCO-30K are provided in the Supplementary. + +# 5. Analysis + +# 5.1. Effect of regularization and segmentation mask + +As shown in Table 3, we conduct an ablation study to examine the effects of regularization losses and segmentation masks on non-target attribute preservation. For the experiments, we train models to mitigate gender bias, and the evaluation is performed using $\mathcal{T}_{\mathrm{basic}}$ . + +When training without segmentation masks, we exclude $\mathcal{L}_{\mathrm{reg}}^{\mathrm{cf}}$ in training and apply $\mathcal{L}_{\mathrm{recon}}$ to the entire image. Since there is no constraint limiting target attribute enhancement, we observe a slight improvement in bias mitigation performance. However, without segmentation masks to distinguish human and non-human regions during training, EFA struggles to localize them precisely, leading to reduced preservation of non-target attributes at inference. Nonetheless, EFA still outperforms baselines even without segmentation masks. We conjecture that this is because counterfactual attributes are primarily activated in bias-relevant regions due to their semantic similarity in cross-attention, which encourages EFA to enhance the target attribute in those regions while ignoring the counterfactual prompt's influence. This allows EFA to approximate the localization of bias-relevant regions without relying on fine-grained masks of the target attribute. Also, when $\mathcal{L}_{\mathrm{reg}}^{\mathrm{trg}}$ is not applied, the + +Table 3. Effectiveness of the regularization loss and the segmentation mask in bias mitigation and non-target attribute preservation. P. and seg. indicate preservation and segmentation, respectively. + +
resolution of inputBiasNon-target attribute P.
DR ↓PSNR ↑LPIPS ↓DINO ↑
16×16 (Ours)0.0632.520.04110.916
16×16 - 32×320.0631.790.04070.915
16×16 - 64×640.0630.650.04380.911
+ +Table 4. Impact of input resolution employed by EFA. Extending EFA to higher-resolution layers compromises the preservation of non-target attributes. P. denotes preservation. + +non-target attribute preservation performance deteriorates, even though the bias mitigation performance remains similar. These results highlight the necessity of the regularization mechanisms to preserve non-target attributes across diverse scenarios where the target attribute appears at different extents in the input. + +# 5.2. Analyzing the influence of resolution in EFA + +To investigate the impact of input resolution, we integrated EFA into additional layers within the up-blocks while training a model for gender bias mitigation. The performance on $\mathcal{T}_{\mathrm{basic}}$ is reported in Table 4, where the row highlighted in gray represents our proposed method. When EFA is applied to higher-resolution features, the deviation ratio remains nearly unchanged, while PSNR and DINO scores exhibit a slight decline. This suggests that lower-resolution layers contain richer semantic information, which is crucial for attribute modification. Moreover, extending EFA to multiple layers inevitably introduces additional noise, potentially compromising the preservation of non-target attributes. + +# 5.3. Attention map visualization + +In Fig. 5, we visualize the attention maps predicted by EFA when the target attribute is enhanced. EFA assigns high values to the human region, effectively enabling gender transformation while preserving the generation capability of the original model, as the impact on accessories, clothing style, and background details remains minimal. + +# 6. Conclusion + +In this work, we addressed attribute entanglement in diffusion-based T2I models, where bias mitigation often distorts non-target attributes, leading to unintended distribution shifts. We propose EFA, which mitigates human-centric biases while preserving non-target attributes. During training, EFA is guided to attend regions with high semantic relevance to counterfactual prompts, constrained by human segmentation masks. At inference, EFA randomly samples a target attribute and alters the cross-attention in selected layers to ensure a fair distribution of target attributes in the outputs. Experimental results verify that EFA outperforms prior methods in mitigating bias while preserving the original model's output distribution and generative capacity. + +# Acknowledgments + +This work was supported by Institute for Information & communications Technology Planning & Evaluation(IITP) grant funded by the Korea government(MSIT) (RS-2019-II190075, Artificial Intelligence Graduate School Program(KAIST)), Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.RS-2021-II212068, Artificial Intelligence Innovation Hub), and the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT)(No. 2022R1A5A7083908). + +# References + +[1] Stability AI. Stable diffusion. https://huggingface.co/runwayml/stable-diffusion-v1-5, 2022. [Online; accessed 3-November-2023]. 1, 6 +[2] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. https://cdn.openai.com/papers/dall-e-3.pdf, 2023. [Online; accessed 4-December-2024]. 1 +[3] Federico Bianchi, Pratyusha Kalluri, Esin Durmus, Faisal Ladhak, Myra Cheng, Debora Nozza, Tatsunori Hashimoto, Dan Jurafsky, James Zou, and Aylin Caliskan. Easily accessible text-to-image generation amplifies demographic stereotypes at large scale. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, pages 1493-1504, 2023. 2 +[4] Manuel Brack, Felix Friedrich, Dominik Hintersdorf, Lukas Struppek, Patrick Schramowski, and Kristian Kersting. Sega: Instructing text-to-image models using semantic guidance. Advances in Neural Information Processing Systems, 36: 25365-25389, 2023. 2 +[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9650-9660, 2021. 7 +[6] Aditya Chinchure, Pushkar Shukla, Gaurav Bhatt, Kiri Salij, Kartik Hosanagar, Leonid Sigal, and Matthew Turk. Tibet: Identifying and evaluating biases in text-to-image generative models. In European Conference on Computer Vision, pages 429-446. Springer, 2025. 5 +[7] Jaemin Cho, Abhay Zala, and Mohit Bansal. Dall-eval: Probing the reasoning skills and social biases of text-to-image generation models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3043-3054, 2023. 2 +[8] Ching-Yao Chuang, Varun Jampani, Yuanzhen Li, Antonio Torralba, and Stefanie Jegelka. Debiasing vision-language models via biased prompts. arXiv preprint arXiv:2302.00070, 2023. 2, 3 +[9] Stefan Elfwing, Eiji Uchibe, and Kenji Doya. Sigmoid-weighted linear units for neural network function approxi + +mation in reinforcement learning. CoRR, abs/1702.03118, 2017.4 +[10] Felix Friedrich, Manuel Brack, Lukas Struppek, Dominik Hintersdorf, Patrick Schramowski, Sasha Luccioni, and Kristian Kersting. Fair diffusion: Instructing text-to-image generation models on fairness. arXiv preprint at arXiv:2302.10893, 2023. 2, 5, 7 +[11] Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzyńska, and David Bau. Unified concept editing in diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5111-5120, 2024. 1, 2, 5, 6, 7, 4 +[12] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7 +[13] Eunji Kim, Siwon Kim, Chaehun Shin, and Sungroh Yoon. De-stereotyping text-to-image models through prompt tuning. 2023. 2, 3 +[14] Mingi Kwon, Jaeseok Jeong, and Youngjung Uh. Diffusion models already have a semantic latent space. In The Eleventh International Conference on Learning Representations, 2023. 3 +[15] Hang Li, Chengzhi Shen, Philip Torr, Volker Tresp, and Jindong Gu. Self-discovering interpretable diffusion latent directions for responsible text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 3, 5, 6, 7, 4 +[16] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer vision-ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part v 13, pages 740-755. Springer, 2014. 5 +[17] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 4 +[18] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1 +[19] Hadas Orgad, Bahjat Kawar, and Yonatan Belinkov. Editing implicit assumptions in text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7053-7061, 2023. 2, 5 +[20] Rishubh Parihar, Abhijnya Bhat, Abhipsa Basu, Saswat Mallick, Jogendra Nath Kundu, and R. Venkatesh Babu. Balancing act: Distribution-guided debiasing in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6668-6678, 2024. 2, 3, 7 +[21] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image gen + +eration with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1 +[22] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dólar, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos, 2024. 4 +[23] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1 +[24] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. 35:36479-36494, 2022. 1 +[25] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. 35:25278-25294, 2022. 2 +[26] Xudong Shen, Chao Du, Tianyu Pang, Min Lin, Yongkang Wong, and Mohan Kankanhalli. Finetuning text-to-image diffusion models for fairness. In The Twelfth International Conference on Learning Representations, 2024. 2, 6, 7, 4 +[27] Robik Shrestha, Yang Zou, Qiuyu Chen, Zhiheng Li, Yusheng Xie, and Siqi Deng. Fairrag: Fair human generation via fair retrieval augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11996-12005, 2024. 2, 3 +[28] Zihao Wang, Lin Gui, Jeffrey Negrea, and Victor Veitch. Concept algebra for (score-based) text-controlled generative models. In Proceedings of the 37th International Conference on Neural Information Processing Systems, pages 35331-35349, 2023. 6, 7, 3, 4 +[29] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7 +[30] Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, and Kai-Wei Chang. Gender bias in coreference resolution: Evaluation and debiasing methods. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 15–20, 2018. 5 +[31] Junlei Zhou, Jiashi Gao, Xiangyu Zhao, Xin Yao, and Xuetao Wei. Association of objects may engender stereotypes: Mitigating association-engendered stereotypes in text-to-image generation. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 3, 7 \ No newline at end of file diff --git a/2025/Fair Generation without Unfair Distortions_ Debiasing Text-to-Image Generation with Entanglement-Free Attention/images.zip b/2025/Fair Generation without Unfair Distortions_ Debiasing Text-to-Image Generation with Entanglement-Free Attention/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1d541caada02ecf0cc4a215ee6ec6ee70d1e3205 --- /dev/null +++ b/2025/Fair Generation without Unfair Distortions_ Debiasing Text-to-Image Generation with Entanglement-Free Attention/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e79ac2ae5247459fdebea9ce37127a39e8b82af1e2a830b30fced2f509e6437d +size 743366 diff --git a/2025/Fair Generation without Unfair Distortions_ Debiasing Text-to-Image Generation with Entanglement-Free Attention/layout.json b/2025/Fair Generation without Unfair Distortions_ Debiasing Text-to-Image Generation with Entanglement-Free Attention/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f652ceb16f42f756cf2db660306d7f5f73e2e894 --- /dev/null +++ b/2025/Fair Generation without Unfair Distortions_ Debiasing Text-to-Image Generation with Entanglement-Free Attention/layout.json @@ -0,0 +1,9172 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 84, + 103, + 526, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 103, + 526, + 140 + ], + "spans": [ + { + "bbox": [ + 84, + 103, + 526, + 140 + ], + "type": "text", + "content": "Fair Generation without Unfair Distortions: Debiasing Text-to-Image Generation with Entanglement-Free Attention" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 128, + 162, + 212, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 162, + 212, + 188 + ], + "spans": [ + { + "bbox": [ + 128, + 162, + 212, + 188 + ], + "type": "text", + "content": "Jeonghoon Park* KAIST" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 101, + 192, + 241, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 192, + 241, + 203 + ], + "spans": [ + { + "bbox": [ + 101, + 192, + 241, + 203 + ], + "type": "text", + "content": "jeonghoon_park@kaist.ac.kr" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 122, + 210, + 209, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 210, + 209, + 238 + ], + "spans": [ + { + "bbox": [ + 122, + 210, + 209, + 238 + ], + "type": "text", + "content": "Jaeseong Lee \nYonsei University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 99, + 239, + 234, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 239, + 234, + 251 + ], + "spans": [ + { + "bbox": [ + 99, + 239, + 234, + 251 + ], + "type": "text", + "content": "jasonlee1995@yonsei.ac.kr" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 286, + 162, + 356, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 162, + 356, + 190 + ], + "spans": [ + { + "bbox": [ + 286, + 162, + 356, + 190 + ], + "type": "text", + "content": "Juyoung Lee\\* Kakao Corp." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 269, + 192, + 373, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 269, + 192, + 373, + 202 + ], + "spans": [ + { + "bbox": [ + 269, + 192, + 373, + 202 + ], + "type": "text", + "content": "zealota11@gmail.com" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 278, + 210, + 340, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 278, + 210, + 340, + 236 + ], + "spans": [ + { + "bbox": [ + 278, + 210, + 340, + 236 + ], + "type": "text", + "content": "Jaegul Choo KAIST" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 263, + 239, + 356, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 263, + 239, + 356, + 251 + ], + "spans": [ + { + "bbox": [ + 263, + 239, + 356, + 251 + ], + "type": "text", + "content": "jchoo@kaist.ac.kr" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 413, + 162, + 498, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 413, + 162, + 498, + 188 + ], + "spans": [ + { + "bbox": [ + 413, + 162, + 498, + 188 + ], + "type": "text", + "content": "Chaeyeon Chung KAIST" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 402, + 192, + 509, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 192, + 509, + 203 + ], + "spans": [ + { + "bbox": [ + 402, + 192, + 509, + 203 + ], + "type": "text", + "content": "cy_chung@kaist.ac.kr" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 397, + 210, + 499, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 210, + 499, + 237 + ], + "spans": [ + { + "bbox": [ + 397, + 210, + 499, + 237 + ], + "type": "text", + "content": "Jindong Gu University of Oxford" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 384, + 239, + 509, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 239, + 509, + 251 + ], + "spans": [ + { + "bbox": [ + 384, + 239, + 509, + 251 + ], + "type": "text", + "content": "jindong.gu@eng.ox.ac.uk" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 215, + 268, + 408, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 268, + 408, + 281 + ], + "spans": [ + { + "bbox": [ + 215, + 268, + 408, + 281 + ], + "type": "text", + "content": "\"face of successful man with his house\"" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 59, + 296, + 179, + 417 + ], + "blocks": [ + { + "bbox": [ + 85, + 284, + 153, + 294 + ], + "lines": [ + { + "bbox": [ + 85, + 284, + 153, + 294 + ], + "spans": [ + { + "bbox": [ + 85, + 284, + 153, + 294 + ], + "type": "text", + "content": "Stable Diffusion" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 296, + 179, + 417 + ], + "lines": [ + { + "bbox": [ + 59, + 296, + 179, + 417 + ], + "spans": [ + { + "bbox": [ + 59, + 296, + 179, + 417 + ], + "type": "image", + "image_path": "fbdcdfa0ee8a26eaf8ee496c218844d4a23628980a15b10eccd0da4d276fbc20.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 183, + 296, + 302, + 416 + ], + "blocks": [ + { + "bbox": [ + 231, + 284, + 253, + 294 + ], + "lines": [ + { + "bbox": [ + 231, + 284, + 253, + 294 + ], + "spans": [ + { + "bbox": [ + 231, + 284, + 253, + 294 + ], + "type": "text", + "content": "UCE" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 183, + 296, + 302, + 416 + ], + "lines": [ + { + "bbox": [ + 183, + 296, + 302, + 416 + ], + "spans": [ + { + "bbox": [ + 183, + 296, + 302, + 416 + ], + "type": "image", + "image_path": "f38018fff929b9db7940ef993dc65c926a2a2b202c1abf3235fe7208f2bc55e2.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 240, + 418, + 367, + 440 + ], + "lines": [ + { + "bbox": [ + 240, + 418, + 367, + 440 + ], + "spans": [ + { + "bbox": [ + 240, + 418, + 367, + 440 + ], + "type": "text", + "content": "Racial attributes changed Non-racial attributes preserved" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 308, + 297, + 425, + 415 + ], + "blocks": [ + { + "bbox": [ + 328, + 284, + 405, + 296 + ], + "lines": [ + { + "bbox": [ + 328, + 284, + 405, + 296 + ], + "spans": [ + { + "bbox": [ + 328, + 284, + 405, + 296 + ], + "type": "text", + "content": "Interpret Diffusion" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 308, + 297, + 425, + 415 + ], + "lines": [ + { + "bbox": [ + 308, + 297, + 425, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 297, + 425, + 415 + ], + "type": "image", + "image_path": "b696fb8b08e0855bfa572c64def597237035baf501c58fd3c69ab917cc3e5745.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 430, + 297, + 551, + 416 + ], + "blocks": [ + { + "bbox": [ + 465, + 284, + 514, + 295 + ], + "lines": [ + { + "bbox": [ + 465, + 284, + 514, + 295 + ], + "spans": [ + { + "bbox": [ + 465, + 284, + 514, + 295 + ], + "type": "text", + "content": "EFA (Ours)" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 430, + 297, + 551, + 416 + ], + "lines": [ + { + "bbox": [ + 430, + 297, + 551, + 416 + ], + "spans": [ + { + "bbox": [ + 430, + 297, + 551, + 416 + ], + "type": "image", + "image_path": "2db09a8ebc6f2a0d703ea0f190a6c708d472b1930d23fd51e25b697e478401c2.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 425, + 418, + 553, + 441 + ], + "lines": [ + { + "bbox": [ + 425, + 418, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 425, + 418, + 553, + 441 + ], + "type": "text", + "content": "Racial attributes changed Non-racial attributes preserved" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 449, + 555, + 506 + ], + "lines": [ + { + "bbox": [ + 55, + 449, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 555, + 506 + ], + "type": "text", + "content": "Figure 1. Generated images from the pretrained model (e.g., Stable Diffusion [1]), by previous debiasing methods (e.g., UCE [11] and Interpret Diffusion [15]), and ours. Previous debiasing methods often alter the non-target attributes (e.g., house) while modifying target bias attributes (e.g., racial attributes). In contrast, our approach preserves non-target attributes during bias mitigation (i.e., changing the racial attributes without modifying the house in the background), thereby preventing the introduction of unintended biases. The corresponding aligned images were generated using the same random seed." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "bbox": [ + 152, + 515, + 200, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 515, + 200, + 528 + ], + "spans": [ + { + "bbox": [ + 152, + 515, + 200, + 528 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 55, + 546, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 297, + 714 + ], + "type": "text", + "content": "Recent advancements in diffusion-based text-to-image (T2I) models have enabled the generation of high-quality and photorealistic images from text. However, they often exhibit societal biases related to gender, race, and socioeconomic status, thereby potentially reinforcing harmful stereotypes and shaping public perception in unintended ways. While existing bias mitigation methods demonstrate effectiveness, they often encounter attribute entanglement, where adjustments to attributes relevant to the bias (i.e., target attributes) unintentionally alter attributes unassociated with the bias (i.e., non-target attributes), causing undesirable distribution shifts. To address this challenge, we introduce Entanglement-Free Attention (EFA), a method that accurately incorporates target attributes (e.g., White, Black, and" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 516, + 556, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 516, + 556, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 516, + 556, + 624 + ], + "type": "text", + "content": "Asian) while preserving non-target attributes (e.g., background) during bias mitigation. At inference time, EFA randomly samples a target attribute with equal probability and adjusts the cross-attention in selected layers to incorporate the sampled attribute, achieving a fair distribution of target attributes. Extensive experiments demonstrate that EFA outperforms existing methods in mitigating bias while preserving non-target attributes, thereby maintaining the original model's output distribution and generative capacity." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 314, + 645, + 395, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 645, + 395, + 658 + ], + "spans": [ + { + "bbox": [ + 314, + 645, + 395, + 658 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Diffusion-based text-to-image (T2I) models [2, 18, 21, 23, 24] have gained significant attention due to their ability to reflect user intent through text input, making them highly applicable across various domains. However, despite their" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17567" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "advancements, these T2I models often perpetuate socially sensitive biases and stereotypes such as gender, race, or socioeconomic status [3, 7, 10]. Such societal biases can reinforce discriminatory perceptions of certain social groups and pose ethical and practical risks in real-world environments, where biased outputs can influence public perception and decision making." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 159, + 295, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 159, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 56, + 159, + 295, + 361 + ], + "type": "text", + "content": "Previous works attempt to mitigate such biases via two primary approaches. One widely explored strategy is finetuning, where pretrained models are adjusted using additional training on curated datasets [11, 19, 26]. Despite its effectiveness, this method often entails substantial computational costs [26]. Also, it may result in a decline in image quality and diversity due to the relatively small size of the finetuning dataset compared to the original pretraining dataset [25]. Since this approach inherently modifies the parameters of the model, it may lead to unintended changes in the output distribution. An alternative line of research explores inference-time steering, which avoids modifying the pretrained model and instead applies adjustments only during inference [8, 10, 13, 15, 20, 27]. Although this approach effectively reduces bias while preserving the parameters of the original model, we observe that it still introduces unintended alterations in the generated outputs." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 364, + 295, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 364, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 56, + 364, + 295, + 639 + ], + "type": "text", + "content": "Fig. 1 illustrates this issue by presenting images generated using the prompt \"face of successful man with his house\". The pretrained model correctly interprets the prompt by generating images of a man with a house but predominantly produces individuals from a limited racial background. When existing bias mitigation methods are applied, the representation of underrepresented racial attributes, such as Black and Indian individuals, is successfully improved. However, they also introduce unintended shifts in non-racial attributes. Specifically, they alter the material and architectural style of the house, associating certain racial groups with impoverished or deteriorated living conditions, thus reinforcing misleading associations between social groups and socioeconomic status. This finding suggests that previous methods may involve attribute entanglement between attributes relevant to the target bias (i.e., target attribute, such as racial attributes) and attributes unassociated with the target bias (i.e., non-target attribute, such as the appearance of the house), potentially leading to unintended alterations in the non-target attributes during the bias mitigation process. Without careful disentanglement, bias mitigation can unintentionally alter contextual details, degrade output diversity, and even introduce new biases." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 642, + 294, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 294, + 714 + ], + "type": "text", + "content": "To address this issue, we propose Entanglement-Free Attention (EFA), a method that accurately incorporates target attributes while preserving non-target attributes. In this paper, we focus on human-centric bias, which plays a crucial role in shaping societal perceptions. Our EFA operates by leveraging predefined target attributes (e.g., White, Black," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 288 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 288 + ], + "type": "text", + "content": "Asian, and Indian) that contribute to bias (e.g., race). To guide the model to focus on semantically relevant regions (e.g., face and body), EFA is trained to enhance target attributes in regions that exhibit high semantic relevance to the counterfactual prompts (i.e., prompts with alternative attributes other than the target), while human segmentation masks further constrain its effect to human regions. By employing both target and counterfactual prompts in training, EFA learns to dynamically adjust the enhancement strength based on the presence of the target attribute, applying stronger modifications when it is underrepresented. During inference, to ensure a fair distribution of target attributes in the generated output, one attribute is randomly sampled with equal probability. EFA then enhances the corresponding features in bias-relevant regions without relying on human-provided masks, effectively disentangling target and non-target attributes while preserving the original model's generative capacity and contextual consistency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 326, + 289, + 526, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 289, + 526, + 299 + ], + "spans": [ + { + "bbox": [ + 326, + 289, + 526, + 299 + ], + "type": "text", + "content": "In summary, our key contributions are as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 300, + 553, + 443 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 300, + 553, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 300, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 314, + 300, + 553, + 346 + ], + "type": "text", + "content": "- We identify the issue of attribute entanglement, where bias mitigation unintentionally alters non-target attributes, leading to unintended distribution shifts in diffusion-based T2I models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 348, + 553, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 348, + 553, + 395 + ], + "spans": [ + { + "bbox": [ + 314, + 348, + 553, + 395 + ], + "type": "text", + "content": "- We propose EFA, a novel debiasing approach that decouples target and non-target attributes, ensuring that bias mitigation is applied only to the intended target attributes while preserving the integrity of non-target attributes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 396, + 553, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 396, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 314, + 396, + 553, + 443 + ], + "type": "text", + "content": "- Quantitative and qualitative experiments demonstrate that our approach successfully mitigates bias while maximally preserving the generation capability of the original model, outperforming previous methods." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 454, + 403, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 454, + 403, + 466 + ], + "spans": [ + { + "bbox": [ + 314, + 454, + 403, + 466 + ], + "type": "text", + "content": "2. Related works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 474, + 553, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 553, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 553, + 628 + ], + "type": "text", + "content": "Existing studies on mitigating bias in T2I generative models have mainly followed two approaches. The first approach involves directly modifying the parameters of diffusion-based T2I models. For instance, Orgad et al. [19] and Gandikota et al. [11] proposed a method to adjust the weights of the cross-attention module to alleviate the bias. Similarly, Shen et al. [26] finetune the text encoder to alleviate face-related biases (e.g., gender, race). However, since finetuning is conducted on a significantly smaller dataset compared to the pretraining dataset, such approach may potentially degrade original output quality or diversity. Furthermore, the approach of Shen et al. [26] requires substantial computational costs for training the backbone model." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": "Another line of research has explored inference-time steering approaches, where the pretrained model remains unchanged while adjustments are applied during inference. Friedrich et al. [10] proposed a bias mitigation framework that employs semantic guidance [4] to manipulate the target attributes by referring to registered information of a concept (e.g., nurse) and its bias (e.g., gender). However, this cannot" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17568" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 68, + 558, + 260 + ], + "blocks": [ + { + "bbox": [ + 55, + 68, + 558, + 260 + ], + "lines": [ + { + "bbox": [ + 55, + 68, + 558, + 260 + ], + "spans": [ + { + "bbox": [ + 55, + 68, + 558, + 260 + ], + "type": "image", + "image_path": "d76059e9a7a449ec9996a39c21cb33d42d1fea70edb5c08f9100cdc71f2b9dd9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 266, + 555, + 313 + ], + "lines": [ + { + "bbox": [ + 54, + 266, + 555, + 313 + ], + "spans": [ + { + "bbox": [ + 54, + 266, + 555, + 313 + ], + "type": "text", + "content": "Figure 2. Overview of our method. (a): EFA is applied to selected layers and trained to reconstruct the image " + }, + { + "bbox": [ + 54, + 266, + 555, + 313 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 54, + 266, + 555, + 313 + ], + "type": "text", + "content": " (a-1) when given " + }, + { + "bbox": [ + 54, + 266, + 555, + 313 + ], + "type": "inline_equation", + "content": "p_a" + }, + { + "bbox": [ + 54, + 266, + 555, + 313 + ], + "type": "text", + "content": ", the prompt containing the target attribute " + }, + { + "bbox": [ + 54, + 266, + 555, + 313 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 54, + 266, + 555, + 313 + ], + "type": "text", + "content": ", or (a-2) its counterfactual prompt, " + }, + { + "bbox": [ + 54, + 266, + 555, + 313 + ], + "type": "inline_equation", + "content": "p_a^{\\mathrm{cf}}" + }, + { + "bbox": [ + 54, + 266, + 555, + 313 + ], + "type": "text", + "content": ". We leverage a human segmentation mask " + }, + { + "bbox": [ + 54, + 266, + 555, + 313 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 54, + 266, + 555, + 313 + ], + "type": "text", + "content": " to minimize the effect of EFA on the attention map of non-target attribute regions. (b): During the inference, our method generates outputs with a fair distribution of target attributes using the EFA of the randomly sampled target attribute." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 327, + 294, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 327, + 294, + 482 + ], + "spans": [ + { + "bbox": [ + 54, + 327, + 294, + 482 + ], + "type": "text", + "content": "mitigate bias for unregistered concepts, limiting its practicality. Chuang et al. [8] have developed a method to remove bias from text embeddings, and Kim et al. [13] trains the soft prompts for mitigating bias. Beyond employing the text embeddings, Shrestha et al. [27] introduced a technique that projects a reference image into text-compatible tokens, allowing the model to leverage the features of external images, such as gender or race for image generation. Additionally, Parihar et al. [20] and Li et al. [15] proposed methods to manipulate the h-space [14] of the model during inference. While these methods aim to mitigate bias without directly altering the backbone parameters, they often struggle to preserve the original context of the generated images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 483, + 295, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 295, + 639 + ], + "type": "text", + "content": "Recently, Zhou et al. [31] addressed association-engendered stereotypes arising from co-occurrence of multiple objects in an image. They rely on a CLIP model trained to predict biases from prompts, requiring prior knowledge of the biases to be mitigated. In contrast, our approach aims to address subtle biases induced by contextual variations beyond the predefined association bias definitions. For instance, if a particular social group is predominantly depicted in low-resource work environments, while another is shown in corporate office settings, such context variations can reinforce cognitive biases, subtly implying disparities in professional status. Furthermore, it is impractical to anticipate all possible forms of such bias in advance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 651, + 216, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 216, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 216, + 664 + ], + "type": "text", + "content": "3. Entanglement-free attention" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 671, + 123, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 123, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 123, + 683 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "Without loss of generality, we assume a given target bias " + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": " (e.g., gender) along with its corresponding set of target" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 327, + 553, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 553, + 398 + ], + "type": "text", + "content": "attributes " + }, + { + "bbox": [ + 313, + 327, + 553, + 398 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_C = \\{a_1, a_2, \\ldots, a_n\\}" + }, + { + "bbox": [ + 313, + 327, + 553, + 398 + ], + "type": "text", + "content": " (e.g., 'female', 'male') whose elements should be uniformly distributed in the outputs. Then, our approach aims to alleviate the bias " + }, + { + "bbox": [ + 313, + 327, + 553, + 398 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 327, + 553, + 398 + ], + "type": "text", + "content": " by adjusting the distribution of target attributes in " + }, + { + "bbox": [ + 313, + 327, + 553, + 398 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_C" + }, + { + "bbox": [ + 313, + 327, + 553, + 398 + ], + "type": "text", + "content": " while preserving the non-target attributes such as scene context and surrounding objects." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 399, + 554, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 554, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 554, + 518 + ], + "type": "text", + "content": "To achieve this, we introduce EFA, which leverages the principle of cross-attention mechanism that assigns desired features to appropriate locations. Each target attribute " + }, + { + "bbox": [ + 313, + 399, + 554, + 518 + ], + "type": "inline_equation", + "content": "a_{i} \\in \\mathcal{A}_{C}" + }, + { + "bbox": [ + 313, + 399, + 554, + 518 + ], + "type": "text", + "content": " is handled by a dedicated EFA module (i.e., " + }, + { + "bbox": [ + 313, + 399, + 554, + 518 + ], + "type": "inline_equation", + "content": "\\mathrm{EFA}_{C}^{a_{i}}" + }, + { + "bbox": [ + 313, + 399, + 554, + 518 + ], + "type": "text", + "content": "), ensuring faithful representation of " + }, + { + "bbox": [ + 313, + 399, + 554, + 518 + ], + "type": "inline_equation", + "content": "a_{i}" + }, + { + "bbox": [ + 313, + 399, + 554, + 518 + ], + "type": "text", + "content": " in the outputs while minimizing its influence on non-target attribute regions. During inference, to enforce a fair distribution of target attributes in the generated output, we randomly sample an attribute " + }, + { + "bbox": [ + 313, + 399, + 554, + 518 + ], + "type": "inline_equation", + "content": "a_{i}" + }, + { + "bbox": [ + 313, + 399, + 554, + 518 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 399, + 554, + 518 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{C}" + }, + { + "bbox": [ + 313, + 399, + 554, + 518 + ], + "type": "text", + "content": " with equal probability and incorporate its representation using the corresponding EFA (Fig. 2(b))." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 519, + 554, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 519, + 554, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 519, + 554, + 651 + ], + "type": "text", + "content": "To enhance the fidelity of " + }, + { + "bbox": [ + 313, + 519, + 554, + 651 + ], + "type": "inline_equation", + "content": "a_i" + }, + { + "bbox": [ + 313, + 519, + 554, + 651 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 519, + 554, + 651 + ], + "type": "inline_equation", + "content": "\\mathrm{EFA}_C^{a_i}" + }, + { + "bbox": [ + 313, + 519, + 554, + 651 + ], + "type": "text", + "content": " is trained to reconstruct images depicting " + }, + { + "bbox": [ + 313, + 519, + 554, + 651 + ], + "type": "inline_equation", + "content": "a_i" + }, + { + "bbox": [ + 313, + 519, + 554, + 651 + ], + "type": "text", + "content": " when its representation is subtle by introducing counterfactual attributes " + }, + { + "bbox": [ + 313, + 519, + 554, + 651 + ], + "type": "inline_equation", + "content": "a_i^{\\mathrm{cf}} \\in \\mathcal{A}_C \\setminus \\{a_i\\}" + }, + { + "bbox": [ + 313, + 519, + 554, + 651 + ], + "type": "text", + "content": " through counterfactual prompts. EFA is guided to enhance the target attribute in areas exhibiting strong semantic relevance to the counterfactual prompt through the cross-attention module. Additionally, human segmentation masks are used during training to constrain EFA's influence to the human region. This ensures that " + }, + { + "bbox": [ + 313, + 519, + 554, + 651 + ], + "type": "inline_equation", + "content": "\\mathrm{EFA}_C^{a_i}" + }, + { + "bbox": [ + 313, + 519, + 554, + 651 + ], + "type": "text", + "content": " enhances the fidelity of " + }, + { + "bbox": [ + 313, + 519, + 554, + 651 + ], + "type": "inline_equation", + "content": "a_i" + }, + { + "bbox": [ + 313, + 519, + 554, + 651 + ], + "type": "text", + "content": " while preventing unintended modifications to non-target attributes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 659, + 432, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 432, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 432, + 671 + ], + "type": "text", + "content": "3.2. Architecture of EFA" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "content": "The original cross-attention module takes an intermediate spatial feature " + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "content": ", which is derived by from a noised image " + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "content": " and a prompt " + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "content": " using the denoising model " + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "content": " (i.e.," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17569" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t = \\phi (\\mathbf{x}_t,p)" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": " is transformed into a query representation using the mapping layer " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "W_{q}" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": ". The text prompt embeddings " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "\\pi (p)" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": " are projected into key and value representations through " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "W_{k}" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "W_{v}" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": ", respectively. The output of the cross-attention layer is formulated as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 90, + 140, + 260, + 168 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 140, + 260, + 168 + ], + "spans": [ + { + "bbox": [ + 90, + 140, + 260, + 168 + ], + "type": "interline_equation", + "content": "\\operatorname {C r o s s - a t t e n t i o n} (\\mathbf {z} _ {t}) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {\\top}}{\\sqrt {d}}\\right) V,", + "image_path": "9dd6f068913ba54c9a7cdd863b633ad1a8d3ec4c839866dea9839193d4220687.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 175, + 295, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 175, + 295, + 212 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 295, + 212 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 175, + 295, + 212 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 175, + 295, + 212 + ], + "type": "text", + "content": " indicates the dimension of the query and key vectors, and " + }, + { + "bbox": [ + 55, + 175, + 295, + 212 + ], + "type": "inline_equation", + "content": "Q = \\mathbf{z}_tW_q" + }, + { + "bbox": [ + 55, + 175, + 295, + 212 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 175, + 295, + 212 + ], + "type": "inline_equation", + "content": "K = \\pi(p)W_k" + }, + { + "bbox": [ + 55, + 175, + 295, + 212 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 175, + 295, + 212 + ], + "type": "inline_equation", + "content": "V = \\pi(p)W_v" + }, + { + "bbox": [ + 55, + 175, + 295, + 212 + ], + "type": "text", + "content": " denote the query, key, and value matrices, respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 212, + 296, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 212, + 296, + 331 + ], + "spans": [ + { + "bbox": [ + 54, + 212, + 296, + 331 + ], + "type": "text", + "content": "As illustrated in Fig. 3, EFA modifies the cross-attention to integrate the desired attribute " + }, + { + "bbox": [ + 54, + 212, + 296, + 331 + ], + "type": "inline_equation", + "content": "a_{i} \\in \\mathcal{A}_{C}" + }, + { + "bbox": [ + 54, + 212, + 296, + 331 + ], + "type": "text", + "content": " into the output while preserving the original attention structure. Specifically, EFA takes the intermediate feature " + }, + { + "bbox": [ + 54, + 212, + 296, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 54, + 212, + 296, + 331 + ], + "type": "text", + "content": " and predicts the additional attention values that determine where and to what extent " + }, + { + "bbox": [ + 54, + 212, + 296, + 331 + ], + "type": "inline_equation", + "content": "a_{i}" + }, + { + "bbox": [ + 54, + 212, + 296, + 331 + ], + "type": "text", + "content": " should be enhanced. This is achieved using the attention value predictor (AP), a lightweight module composed of three convolutional layers and two SiLU [9] activation functions. Further details of the architecture are provided in the Supplementary." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 331, + 296, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 296, + 430 + ], + "type": "text", + "content": "The predicted attention value (i.e., " + }, + { + "bbox": [ + 55, + 331, + 296, + 430 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{a_i}(\\mathbf{z}_t)" + }, + { + "bbox": [ + 55, + 331, + 296, + 430 + ], + "type": "text", + "content": ") are concatenated with the original cross-attention's attention values (i.e., " + }, + { + "bbox": [ + 55, + 331, + 296, + 430 + ], + "type": "inline_equation", + "content": "\\frac{QK^{\\top}}{\\sqrt{d}}" + }, + { + "bbox": [ + 55, + 331, + 296, + 430 + ], + "type": "text", + "content": "), and the softmax function is applied. After applying softmax, the original value vectors from the user prompt (i.e., " + }, + { + "bbox": [ + 55, + 331, + 296, + 430 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 331, + 296, + 430 + ], + "type": "text", + "content": ") and the value vectors of " + }, + { + "bbox": [ + 55, + 331, + 296, + 430 + ], + "type": "inline_equation", + "content": "a_i" + }, + { + "bbox": [ + 55, + 331, + 296, + 430 + ], + "type": "text", + "content": " are weighted accordingly, as in the standard cross-attention operation. Formally, EFA incorporates " + }, + { + "bbox": [ + 55, + 331, + 296, + 430 + ], + "type": "inline_equation", + "content": "a_i" + }, + { + "bbox": [ + 55, + 331, + 296, + 430 + ], + "type": "text", + "content": " into the cross-attention as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 72, + 437, + 278, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 437, + 278, + 464 + ], + "spans": [ + { + "bbox": [ + 72, + 437, + 278, + 464 + ], + "type": "interline_equation", + "content": "\\operatorname {E F A} _ {C} ^ {a _ {i}} (\\mathbf {z} _ {t}) = \\operatorname {s o f t m a x} \\left(\\left[ \\frac {Q K ^ {\\top}}{\\sqrt {d}}, \\mathrm {A P} _ {a _ {i}} (\\mathbf {z} _ {t}) \\right]\\right) [ V, V _ {a _ {i}} ],", + "image_path": "87afbf3ab4889b42a832e1c55571fde37755449dbea96ce7ba6bf72a536b5ba6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 472, + 295, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 496 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 472, + 295, + 496 + ], + "type": "inline_equation", + "content": "V_{a_i} = \\pi(p_{a_i}) W_v" + }, + { + "bbox": [ + 55, + 472, + 295, + 496 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 472, + 295, + 496 + ], + "type": "inline_equation", + "content": "p_{a_i}" + }, + { + "bbox": [ + 55, + 472, + 295, + 496 + ], + "type": "text", + "content": " indicates the text prompt representing " + }, + { + "bbox": [ + 55, + 472, + 295, + 496 + ], + "type": "inline_equation", + "content": "a_i" + }, + { + "bbox": [ + 55, + 472, + 295, + 496 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 472, + 295, + 496 + ], + "type": "inline_equation", + "content": "[\\cdot, \\cdot]" + }, + { + "bbox": [ + 55, + 472, + 295, + 496 + ], + "type": "text", + "content": " indicates the concatenation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 497, + 296, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 296, + 581 + ], + "type": "text", + "content": "When building an " + }, + { + "bbox": [ + 55, + 497, + 296, + 581 + ], + "type": "inline_equation", + "content": "\\mathrm{EFA}_C" + }, + { + "bbox": [ + 55, + 497, + 296, + 581 + ], + "type": "text", + "content": " for a target bias " + }, + { + "bbox": [ + 55, + 497, + 296, + 581 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 497, + 296, + 581 + ], + "type": "text", + "content": ", all AP modules share the same convolutional backbone while maintaining distinct output channels designated for each attribute in " + }, + { + "bbox": [ + 55, + 497, + 296, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_C" + }, + { + "bbox": [ + 55, + 497, + 296, + 581 + ], + "type": "text", + "content": ". This design allows AP models to be effectively trained together, ensuring that EFAs of target attributes for a given bias are jointly learned while avoiding the inefficiency of training multiple independent APs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 588, + 119, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 119, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 119, + 601 + ], + "type": "text", + "content": "3.3. Training" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "Dataset construction. To train " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{EFA}_C^{a_i}" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " to represent a target attribute " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "a_{i}" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " (e.g., female) while preserving non-target attributes, we create a dataset " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{a_i}" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " that captures " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "a_{i}" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " and provides pseudo supervision on the human regions that contain " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "a_{i}" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": ". Specifically, we generate images using the original model with prompts explicitly describing " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "a_{i}" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " (e.g., \"a female person\"). For each image, we extract a human segmentation mask " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " using Grounded SAM2 [17, 22], assigning 1 to human regions and 0 elsewhere. Since this study focuses on" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 315, + 69, + 555, + 175 + ], + "blocks": [ + { + "bbox": [ + 315, + 69, + 555, + 175 + ], + "lines": [ + { + "bbox": [ + 315, + 69, + 555, + 175 + ], + "spans": [ + { + "bbox": [ + 315, + 69, + 555, + 175 + ], + "type": "image", + "image_path": "02f98c40ffc1e6cebfb3e755dad271f7b93be770eb596d9a6aa54ab44172384c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 183, + 555, + 206 + ], + "lines": [ + { + "bbox": [ + 313, + 183, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 313, + 183, + 555, + 206 + ], + "type": "text", + "content": "Figure 3. Architecture of EFA. We denote " + }, + { + "bbox": [ + 313, + 183, + 555, + 206 + ], + "type": "inline_equation", + "content": "a_i \\in \\mathcal{A}_C" + }, + { + "bbox": [ + 313, + 183, + 555, + 206 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 183, + 555, + 206 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 313, + 183, + 555, + 206 + ], + "type": "text", + "content": " in this figure for notational simplicity." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 213, + 555, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 261 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 261 + ], + "type": "text", + "content": "human-centric bias, we utilize a human segmentation mask to serve as the basis for regulating the influence of " + }, + { + "bbox": [ + 313, + 213, + 555, + 261 + ], + "type": "inline_equation", + "content": "a_{i}" + }, + { + "bbox": [ + 313, + 213, + 555, + 261 + ], + "type": "text", + "content": " during training. The resulting " + }, + { + "bbox": [ + 313, + 213, + 555, + 261 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{a_i}" + }, + { + "bbox": [ + 313, + 213, + 555, + 261 + ], + "type": "text", + "content": " consists of pairs of images and their segmentation masks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 261, + 554, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 261, + 554, + 416 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 554, + 416 + ], + "type": "text", + "content": "Training EFA. At inference, the extent of the features that need to be adjusted by EFA to faithfully reflect " + }, + { + "bbox": [ + 313, + 261, + 554, + 416 + ], + "type": "inline_equation", + "content": "a_{i}" + }, + { + "bbox": [ + 313, + 261, + 554, + 416 + ], + "type": "text", + "content": " may vary across inputs. For instance, with the prompt \"nurse\", intermediate features may exhibit a high level of femininity, requiring minimal adjustment by " + }, + { + "bbox": [ + 313, + 261, + 554, + 416 + ], + "type": "inline_equation", + "content": "\\mathrm{EFA}_{\\text{gender}}^{\\text{female}}" + }, + { + "bbox": [ + 313, + 261, + 554, + 416 + ], + "type": "text", + "content": ". In contrast, intermediate features of the prompt \"doctor\" may have stronger masculinity, requiring enhanced representation of female by " + }, + { + "bbox": [ + 313, + 261, + 554, + 416 + ], + "type": "inline_equation", + "content": "\\mathrm{EFA}_{\\text{gender}}^{\\text{female}}" + }, + { + "bbox": [ + 313, + 261, + 554, + 416 + ], + "type": "text", + "content": " to incorporate 'female' into the output. To handle this variability, " + }, + { + "bbox": [ + 313, + 261, + 554, + 416 + ], + "type": "inline_equation", + "content": "\\mathrm{EFA}_{C}^{a_{i}}" + }, + { + "bbox": [ + 313, + 261, + 554, + 416 + ], + "type": "text", + "content": " is trained in two different scenarios: (1) when intermediate features contain the attribute " + }, + { + "bbox": [ + 313, + 261, + 554, + 416 + ], + "type": "inline_equation", + "content": "a_{i}" + }, + { + "bbox": [ + 313, + 261, + 554, + 416 + ], + "type": "text", + "content": " (e.g., female), and (2) when intermediate features contain its counterfactual attribute " + }, + { + "bbox": [ + 313, + 261, + 554, + 416 + ], + "type": "inline_equation", + "content": "a_{i}^{\\mathrm{cf}} \\in \\mathcal{A}_{C} \\setminus \\{a_{i}\\}" + }, + { + "bbox": [ + 313, + 261, + 554, + 416 + ], + "type": "text", + "content": " (e.g., male). We denote " + }, + { + "bbox": [ + 313, + 261, + 554, + 416 + ], + "type": "inline_equation", + "content": "a_{i}" + }, + { + "bbox": [ + 313, + 261, + 554, + 416 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 261, + 554, + 416 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 313, + 261, + 554, + 416 + ], + "type": "text", + "content": " throughout this section for simplicity." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 417, + 554, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 417, + 554, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 417, + 554, + 536 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 313, + 417, + 554, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 313, + 417, + 554, + 536 + ], + "type": "text", + "content": " be an image and " + }, + { + "bbox": [ + 313, + 417, + 554, + 536 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 417, + 554, + 536 + ], + "type": "text", + "content": " be its corresponding segmentation mask, sampled from " + }, + { + "bbox": [ + 313, + 417, + 554, + 536 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_a" + }, + { + "bbox": [ + 313, + 417, + 554, + 536 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 417, + 554, + 536 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 313, + 417, + 554, + 536 + ], + "type": "text", + "content": " denotes a target attribute (e.g., female). In the first scenario, where the intermediate features already encode " + }, + { + "bbox": [ + 313, + 417, + 554, + 536 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 313, + 417, + 554, + 536 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 417, + 554, + 536 + ], + "type": "inline_equation", + "content": "\\mathrm{EFA}_C^a" + }, + { + "bbox": [ + 313, + 417, + 554, + 536 + ], + "type": "text", + "content": " does not need to enhance " + }, + { + "bbox": [ + 313, + 417, + 554, + 536 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 313, + 417, + 554, + 536 + ], + "type": "text", + "content": " but should instead preserve the original cross-attention output to maintain the non-target attributes. To enforce this behavior, we provide the model with a paired input: a prompt " + }, + { + "bbox": [ + 313, + 417, + 554, + 536 + ], + "type": "inline_equation", + "content": "p_a" + }, + { + "bbox": [ + 313, + 417, + 554, + 536 + ], + "type": "text", + "content": " explicitly specifying " + }, + { + "bbox": [ + 313, + 417, + 554, + 536 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 313, + 417, + 554, + 536 + ], + "type": "text", + "content": " (e.g., \"female person\") and a noised image " + }, + { + "bbox": [ + 313, + 417, + 554, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 313, + 417, + 554, + 536 + ], + "type": "text", + "content": " that inherently contains " + }, + { + "bbox": [ + 313, + 417, + 554, + 536 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 313, + 417, + 554, + 536 + ], + "type": "text", + "content": ", as depicted in Fig. 2 (a-1)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 536, + 554, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 554, + 595 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 554, + 595 + ], + "type": "text", + "content": "To prevent unnecessary alterations, we introduce an L1 regularization term that encourages the attention values associated with " + }, + { + "bbox": [ + 313, + 536, + 554, + 595 + ], + "type": "inline_equation", + "content": "V_{a}" + }, + { + "bbox": [ + 313, + 536, + 554, + 595 + ], + "type": "text", + "content": " (e.g., attention values associated with the value vectors of \"female\") to be zero. This regularization loss is defined as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 367, + 604, + 500, + 618 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 604, + 500, + 618 + ], + "spans": [ + { + "bbox": [ + 367, + 604, + 500, + 618 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e g}} (\\mathbf {z} _ {t}, W) = \\left\\| W \\odot A _ {a} \\left(\\mathbf {z} _ {t}\\right) \\right\\| _ {1},", + "image_path": "707ac3e4b0fc84aa54506e3dbdc015337fb6b88330c0f8546eec778bf36f5b0d.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 625, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 625, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 625, + 555, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 625, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t = \\phi (\\mathbf{x}_t,p_a)" + }, + { + "bbox": [ + 313, + 625, + 555, + 715 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 625, + 555, + 715 + ], + "type": "text", + "content": " is a spatial mask used to constrain the effect of the loss, and " + }, + { + "bbox": [ + 313, + 625, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 313, + 625, + 555, + 715 + ], + "type": "text", + "content": " represents the element-wise multiplication. Also, " + }, + { + "bbox": [ + 313, + 625, + 555, + 715 + ], + "type": "inline_equation", + "content": "A_{a}(\\mathbf{z}_{t})" + }, + { + "bbox": [ + 313, + 625, + 555, + 715 + ], + "type": "text", + "content": " denotes the attention values (after softmax) associated with the value vectors " + }, + { + "bbox": [ + 313, + 625, + 555, + 715 + ], + "type": "inline_equation", + "content": "V_{a}" + }, + { + "bbox": [ + 313, + 625, + 555, + 715 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 313, + 625, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 313, + 625, + 555, + 715 + ], + "type": "text", + "content": ". More precisely, when " + }, + { + "bbox": [ + 313, + 625, + 555, + 715 + ], + "type": "inline_equation", + "content": "A(\\mathbf{z}_t) =" + }, + { + "bbox": [ + 313, + 625, + 555, + 715 + ], + "type": "text", + "content": " softmax " + }, + { + "bbox": [ + 313, + 625, + 555, + 715 + ], + "type": "inline_equation", + "content": "([\\frac{QK^{\\top}}{\\sqrt{d}},\\mathrm{AP}_a(\\mathbf{z}_t)])" + }, + { + "bbox": [ + 313, + 625, + 555, + 715 + ], + "type": "text", + "content": " , we define " + }, + { + "bbox": [ + 313, + 625, + 555, + 715 + ], + "type": "inline_equation", + "content": "A_{a}(\\mathbf{z}_{t})" + }, + { + "bbox": [ + 313, + 625, + 555, + 715 + ], + "type": "text", + "content": " as the submatrix of " + }, + { + "bbox": [ + 313, + 625, + 555, + 715 + ], + "type": "inline_equation", + "content": "A(\\mathbf{z}_t)" + }, + { + "bbox": [ + 313, + 625, + 555, + 715 + ], + "type": "text", + "content": " obtained by selecting values indexed by " + }, + { + "bbox": [ + 313, + 625, + 555, + 715 + ], + "type": "inline_equation", + "content": "I_{V_a}" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17570" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 286, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 286, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 286, + 84 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 286, + 84 + ], + "type": "inline_equation", + "content": "I_{V_a}" + }, + { + "bbox": [ + 55, + 72, + 286, + 84 + ], + "type": "text", + "content": " denotes the set of indices corresponding to " + }, + { + "bbox": [ + 55, + 72, + 286, + 84 + ], + "type": "inline_equation", + "content": "V_a" + }, + { + "bbox": [ + 55, + 72, + 286, + 84 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 84, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 84, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 84, + 294, + 144 + ], + "type": "text", + "content": "In this scenario, we set " + }, + { + "bbox": [ + 55, + 84, + 294, + 144 + ], + "type": "inline_equation", + "content": "W = \\mathbf{1}" + }, + { + "bbox": [ + 55, + 84, + 294, + 144 + ], + "type": "text", + "content": ", a mask of ones, to apply the loss over all spatial positions. This regularization loss minimizes unnecessary modifications to the original cross-attention output introduced by " + }, + { + "bbox": [ + 55, + 84, + 294, + 144 + ], + "type": "inline_equation", + "content": "\\mathrm{EFA}_C^a" + }, + { + "bbox": [ + 55, + 84, + 294, + 144 + ], + "type": "text", + "content": ". Hereafter, we denote " + }, + { + "bbox": [ + 55, + 84, + 294, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}(\\mathbf{z}_t,\\mathbf{1})" + }, + { + "bbox": [ + 55, + 84, + 294, + 144 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 84, + 294, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}^{\\mathrm{trg}}" + }, + { + "bbox": [ + 55, + 84, + 294, + 144 + ], + "type": "text", + "content": " for brevity." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 144, + 294, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 144, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 54, + 144, + 294, + 251 + ], + "type": "text", + "content": "In the second scenario, the intermediate features contain a counterfactual attribute " + }, + { + "bbox": [ + 54, + 144, + 294, + 251 + ], + "type": "inline_equation", + "content": "a^{\\mathrm{cf}}" + }, + { + "bbox": [ + 54, + 144, + 294, + 251 + ], + "type": "text", + "content": " (e.g., male) induced by bias of prompt (e.g., \"photo of a doctor\" implicitly favoring male representation). Here, " + }, + { + "bbox": [ + 54, + 144, + 294, + 251 + ], + "type": "inline_equation", + "content": "\\mathrm{EFA}_C^a" + }, + { + "bbox": [ + 54, + 144, + 294, + 251 + ], + "type": "text", + "content": " should suppress " + }, + { + "bbox": [ + 54, + 144, + 294, + 251 + ], + "type": "inline_equation", + "content": "a^{\\mathrm{cf}}" + }, + { + "bbox": [ + 54, + 144, + 294, + 251 + ], + "type": "text", + "content": " and enhance " + }, + { + "bbox": [ + 54, + 144, + 294, + 251 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 54, + 144, + 294, + 251 + ], + "type": "text", + "content": " (e.g., female) while preserving non-target attributes. To simulate this, we train " + }, + { + "bbox": [ + 54, + 144, + 294, + 251 + ], + "type": "inline_equation", + "content": "\\mathrm{EFA}_C^a" + }, + { + "bbox": [ + 54, + 144, + 294, + 251 + ], + "type": "text", + "content": " to counteract counterfactual features introduced during the denoising process of a diffusion-based T2I model, which generates clean images by iteratively denoising from pure noise." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 251, + 294, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 294, + 312 + ], + "type": "text", + "content": "To encourage representation of " + }, + { + "bbox": [ + 55, + 251, + 294, + 312 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 55, + 251, + 294, + 312 + ], + "type": "text", + "content": " in human-related regions, we define a reconstruction loss leveraging the segmentation mask " + }, + { + "bbox": [ + 55, + 251, + 294, + 312 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 251, + 294, + 312 + ], + "type": "text", + "content": " to constrain its effect on human areas. The objective is to predict the added noise " + }, + { + "bbox": [ + 55, + 251, + 294, + 312 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 251, + 294, + 312 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 55, + 251, + 294, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 55, + 251, + 294, + 312 + ], + "type": "text", + "content": " and a counterfactual prompt " + }, + { + "bbox": [ + 55, + 251, + 294, + 312 + ], + "type": "inline_equation", + "content": "p_a^{\\mathrm{cf}}" + }, + { + "bbox": [ + 55, + 251, + 294, + 312 + ], + "type": "text", + "content": ", formulated as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 316, + 292, + 332 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 316, + 292, + 332 + ], + "spans": [ + { + "bbox": [ + 58, + 316, + 292, + 332 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e c o n}} = \\mathbb {E} _ {\\mathbf {x} _ {0}, t, p _ {a} ^ {\\mathrm {c f}}, \\epsilon \\sim N (0, 1)} \\left[ \\| M \\odot (\\epsilon - \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t, p _ {a} ^ {\\mathrm {c f}})) \\| _ {2} ^ {2} \\right],", + "image_path": "8a10a8b6e9ca8473109c1c06869e4866106499cf8d64d356e79c633562b19cac.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 335, + 294, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 294, + 407 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 335, + 294, + 407 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 55, + 335, + 294, + 407 + ], + "type": "text", + "content": " denotes the model with " + }, + { + "bbox": [ + 55, + 335, + 294, + 407 + ], + "type": "inline_equation", + "content": "\\mathrm{EFA}_C^a" + }, + { + "bbox": [ + 55, + 335, + 294, + 407 + ], + "type": "text", + "content": ". Through the reconstruction process, " + }, + { + "bbox": [ + 55, + 335, + 294, + 407 + ], + "type": "inline_equation", + "content": "\\mathrm{EFA}_C^a" + }, + { + "bbox": [ + 55, + 335, + 294, + 407 + ], + "type": "text", + "content": " is encouraged to enhance " + }, + { + "bbox": [ + 55, + 335, + 294, + 407 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 55, + 335, + 294, + 407 + ], + "type": "text", + "content": " in regions where cross-attention assigns high similarity to " + }, + { + "bbox": [ + 55, + 335, + 294, + 407 + ], + "type": "inline_equation", + "content": "a^{\\mathrm{cf}}" + }, + { + "bbox": [ + 55, + 335, + 294, + 407 + ], + "type": "text", + "content": ", reflecting their semantic relevance. Using human masks ensures that " + }, + { + "bbox": [ + 55, + 335, + 294, + 407 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 55, + 335, + 294, + 407 + ], + "type": "text", + "content": " is accurately reconstructed in human regions while mitigating the influence of " + }, + { + "bbox": [ + 55, + 335, + 294, + 407 + ], + "type": "inline_equation", + "content": "a^{\\mathrm{cf}}" + }, + { + "bbox": [ + 55, + 335, + 294, + 407 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 407, + 294, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 294, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 294, + 477 + ], + "type": "text", + "content": "Although " + }, + { + "bbox": [ + 55, + 407, + 294, + 477 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}^{\\mathrm{trg}}" + }, + { + "bbox": [ + 55, + 407, + 294, + 477 + ], + "type": "text", + "content": " reduces unintended modifications, it does not explicitly confine its impact on human regions, potentially affecting non-target attributes. To address this, we impose an additional L1 regularization on the attention values corresponding to " + }, + { + "bbox": [ + 55, + 407, + 294, + 477 + ], + "type": "inline_equation", + "content": "V_{a}" + }, + { + "bbox": [ + 55, + 407, + 294, + 477 + ], + "type": "text", + "content": ", applied outside the human region, as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 123, + 478, + 227, + 493 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 478, + 227, + 493 + ], + "spans": [ + { + "bbox": [ + 123, + 478, + 227, + 493 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {r e g}} ^ {\\mathrm {c f}} = \\mathcal {L} _ {\\mathrm {r e g}} (\\mathbf {z} _ {t} ^ {\\mathrm {c f}}, \\mathbf {1} - \\bar {M}),", + "image_path": "6891db548a399e734bef6059f230587d60472ad89fa5c827b62f9e2e847fb758.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 495, + 294, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 294, + 520 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 495, + 294, + 520 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t^{\\mathrm{cf}} = \\phi (\\mathbf{x}_t,p_a^{\\mathrm{cf}})" + }, + { + "bbox": [ + 55, + 495, + 294, + 520 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 495, + 294, + 520 + ], + "type": "inline_equation", + "content": "\\bar{M}" + }, + { + "bbox": [ + 55, + 495, + 294, + 520 + ], + "type": "text", + "content": " indicates the binary mask resized to match the resolution of the input feature of " + }, + { + "bbox": [ + 55, + 495, + 294, + 520 + ], + "type": "inline_equation", + "content": "\\mathrm{EFA}_C^a" + }, + { + "bbox": [ + 55, + 495, + 294, + 520 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 521, + 266, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 521, + 266, + 533 + ], + "spans": [ + { + "bbox": [ + 67, + 521, + 266, + 533 + ], + "type": "text", + "content": "Finally, the overall training loss is formulated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 538, + 218, + 553 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 538, + 218, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 538, + 218, + 553 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {t o t}} = \\mathcal {L} _ {\\mathrm {r e c o n}} + \\mathcal {L} _ {\\mathrm {r e g}} ^ {*},", + "image_path": "a5e120d945b8b5a2ab509e56a0d4fbc71f36657709b871a10ad1cab9fab7371c.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 558, + 294, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 294, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 558, + 294, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}^{*} = \\lambda_{1}\\mathcal{L}_{\\mathrm{reg}}^{\\mathrm{trg}} + \\lambda_{2}\\mathcal{L}_{\\mathrm{reg}}^{\\mathrm{cf}}" + }, + { + "bbox": [ + 55, + 558, + 294, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 558, + 294, + 666 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 55, + 558, + 294, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 558, + 294, + 666 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 55, + 558, + 294, + 666 + ], + "type": "text", + "content": " are hyperparameters to control the relative importance between the losses. During training, only the parameters of " + }, + { + "bbox": [ + 55, + 558, + 294, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_a" + }, + { + "bbox": [ + 55, + 558, + 294, + 666 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 558, + 294, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{EFA}_C^a" + }, + { + "bbox": [ + 55, + 558, + 294, + 666 + ], + "type": "text", + "content": " are updated. Note that EFA is applied to the upblock layers with an input resolution of " + }, + { + "bbox": [ + 55, + 558, + 294, + 666 + ], + "type": "inline_equation", + "content": "16\\times 16" + }, + { + "bbox": [ + 55, + 558, + 294, + 666 + ], + "type": "text", + "content": ", as lower-resolution features capture higher-level semantic information. The choice of applied layers is discussed in Section 5.2. The method of extending EFA to handle multiple biases is provided in the Supplementary." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 672, + 123, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 123, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 123, + 683 + ], + "type": "text", + "content": "3.4. Inference" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 689, + 294, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 294, + 714 + ], + "type": "text", + "content": "To mitigate the bias " + }, + { + "bbox": [ + 55, + 689, + 294, + 714 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 689, + 294, + 714 + ], + "type": "text", + "content": " and ensure fair generation, we randomly sample an attribute " + }, + { + "bbox": [ + 55, + 689, + 294, + 714 + ], + "type": "inline_equation", + "content": "a_{i}" + }, + { + "bbox": [ + 55, + 689, + 294, + 714 + ], + "type": "text", + "content": " with an equal probability from" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_C" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " and apply the corresponding " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\mathrm{EFA}_C^{a_i}" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " to the original model. This approach enables users to easily control the attribute distribution by adjusting the sampling frequency, without additional training. Note that our method requires no masks at inference and generates debiased images directly from noise, offering a practical and efficient solution." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 158, + 395, + 171 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 158, + 395, + 171 + ], + "spans": [ + { + "bbox": [ + 313, + 158, + 395, + 171 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 178, + 440, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 178, + 440, + 190 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 440, + 190 + ], + "type": "text", + "content": "4.1. Experimental settings" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 196, + 553, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 196, + 553, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 196, + 553, + 340 + ], + "type": "text", + "content": "Datasets. Following the previous studies [10, 11, 15], we mainly conducted experiments with gender and racial bias. Experiments with age bias are provided in Supplementary. The target attributes set of gender and racial bias is set as " + }, + { + "bbox": [ + 313, + 196, + 553, + 340 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\mathrm{gender}} = \\{\\mathrm{female}, \\mathrm{male}\\}" + }, + { + "bbox": [ + 313, + 196, + 553, + 340 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 196, + 553, + 340 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\mathrm{race}} = \\{\\mathrm{White}, \\mathrm{Black}, \\mathrm{Asian}, \\mathrm{Indian}\\}" + }, + { + "bbox": [ + 313, + 196, + 553, + 340 + ], + "type": "text", + "content": ", respectively. For multiple biases, the set of target attributes is defined as " + }, + { + "bbox": [ + 313, + 196, + 553, + 340 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\mathrm{gender} \\times \\mathrm{race}} = \\mathcal{A}_{\\mathrm{gender}} \\times \\mathcal{A}_{\\mathrm{race}}" + }, + { + "bbox": [ + 313, + 196, + 553, + 340 + ], + "type": "text", + "content": ". We adopt a binary gender perspective to evaluate our method, following prior works [11, 15], due to the limitations of existing datasets and evaluation protocols. Further details of the dataset construction are included in the Supplementary." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 341, + 554, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 554, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 554, + 508 + ], + "type": "text", + "content": "Evaluation prompt. To assess bias mitigation performance, we utilized 36 occupations from the WinoBias [30] dataset, which includes stereotypical occupations based on US Department of Labor statistics by following previous studies [11, 15, 19]. We employed two types of prompt templates for image generation. The first set (i.e., " + }, + { + "bbox": [ + 313, + 341, + 554, + 508 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{basic}}" + }, + { + "bbox": [ + 313, + 341, + 554, + 508 + ], + "type": "text", + "content": ") follows the prompts used in prior research [11, 15], formatted as “{\\style} of {\\occupation}.” The style category includes photo, image, picture, headshot, and portrait. The second set (i.e., " + }, + { + "bbox": [ + 313, + 341, + 554, + 508 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{complex}}" + }, + { + "bbox": [ + 313, + 341, + 554, + 508 + ], + "type": "text", + "content": ") includes background components to enable evaluation with more complex non-target attributes. Specifically, prompts are structured as “{\\style} of {\\occupation} in {\\background},” where the background is contextually relevant to the occupation." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 510, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 554, + 605 + ], + "type": "text", + "content": "To evaluate the overall model preservation performance while ensuring non-target attributes remain unaffected, we assessed whether the model retained its original generation capability when generating images without human presence. Specifically, we filtered image-prompt pairs from the COCO-30K [16] dataset and constructed a new dataset, COCO-no-person, consisting of 14,926 pairs where no person is present in the corresponding images." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": "Evaluation metrics. Our evaluation focuses on (1) bias mitigation, ensuring balanced attribute representation; (2) non-target attribute preservation, maintaining non-target attributes while adjusting target attributes; and (3) model preservation, maintaining the overall generative capability of the original model in non-human contexts. While model preservation also evaluates the preservation of non-target attributes, we use a distinct term to differentiate it from cases where target attributes should be enhanced." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17571" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 553, + 324 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 553, + 324 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 553, + 324 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 553, + 324 + ], + "type": "table", + "html": "
\\(T_{\\text{basic}}\\)\\(T_{\\text{complex}}\\)
BiasNon-target attribute P.BiasNon-target attribute P.
DR ↓PSNR ↑LPIPS ↓DINO ↑DR ↓PSNR ↑LPIPS ↓DINO ↑
GenderOriginal SD0.71---0.71---
Concept Algebra [28]0.5921.100.11690.8230.6916.690.18520.834
UCE [11]0.3421.040.13740.7570.4916.780.22020.747
Finetuning Diffusion [26]0.4822.620.11660.8140.5619.270.16110.857
Interpret Diffusion [15]0.2617.180.22900.6160.3114.010.32110.618
EFA (Ours)0.0632.520.04110.9160.0629.700.04920.941
RaceOriginal SD0.60---0.55---
Concept Algebra [28]0.6421.470.11640.8390.5816.620.18880.841
UCE [11]0.2721.550.12610.7870.4117.810.18910.801
Finetuning Diffusion [26]0.2918.960.17950.7090.3716.460.22630.771
Interpret Diffusion [15]0.1616.870.24160.5840.2313.730.33380.584
EFA (Ours)0.0430.930.03530.9380.0628.550.04210.958
Gender × RaceOriginal SD0.56---0.50---
Concept Algebra [28]0.5120.120.13250.8050.4715.880.20310.818
UCE [11]0.1619.840.15750.7020.3216.790.21980.746
Finetuning Diffusion [26]0.2419.970.15700.7460.2817.240.20640.795
Interpret Diffusion [15]0.1616.910.24490.4660.2013.660.35290.475
EFA (Ours)0.0325.580.06840.8530.0523.780.07950.903
", + "image_path": "ea4c9bffff25d724cce4ff4335e2f09bdd8bc62b135c93ffe170d1e3a0d15abc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 331, + 555, + 355 + ], + "lines": [ + { + "bbox": [ + 55, + 331, + 555, + 355 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 555, + 355 + ], + "type": "text", + "content": "Table 1. Comparison to baselines. P. is the abbreviation for preservation. The best values are in bold. Our method achieves superior bias mitigation performance while effectively preserving non-target attributes, significantly outperforming previous approaches." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 56, + 365, + 302, + 571 + ], + "blocks": [ + { + "bbox": [ + 56, + 365, + 302, + 571 + ], + "lines": [ + { + "bbox": [ + 56, + 365, + 302, + 571 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 302, + 571 + ], + "type": "image", + "image_path": "3bd84c11dd5af4ab6eff29155185078b65e8724f06573fb65eafe4fcf2dc240f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 578, + 555, + 624 + ], + "lines": [ + { + "bbox": [ + 54, + 578, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 54, + 578, + 555, + 624 + ], + "type": "text", + "content": "Figure 4. Generated images from the original Stable Diffusion and ours. Ours successfully generates images with a fair distribution of the target bias (i.e., gender and race). (a) and (b) represent the images generated from the prompts describing a nurse in a hospital and a carpenter in a workshop, respectively. The corresponding aligned images were generated using the same random seed. Further qualitative results, including those from the model mitigating gender and race simultaneously, are provided in the Supplementary." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 319, + 365, + 554, + 571 + ], + "blocks": [ + { + "bbox": [ + 319, + 365, + 554, + 571 + ], + "lines": [ + { + "bbox": [ + 319, + 365, + 554, + 571 + ], + "spans": [ + { + "bbox": [ + 319, + 365, + 554, + 571 + ], + "type": "image", + "image_path": "f17e94095042218fd2ba9ade9e55a707081fdbc36383b0ad1dc484e406d53b43.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 635, + 297, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 297, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 297, + 712 + ], + "type": "text", + "content": "For evaluating bias mitigation, we use a CLIP classifier to categorize each generated image into attribute-defined groups in " + }, + { + "bbox": [ + 55, + 635, + 297, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_C" + }, + { + "bbox": [ + 55, + 635, + 297, + 712 + ], + "type": "text", + "content": " and measure attribute balance using modified deviation ratio [15], which is designed to handle an arbitrary number of attributes. The deviation ratio (DR) is calculated as " + }, + { + "bbox": [ + 55, + 635, + 297, + 712 + ], + "type": "inline_equation", + "content": "\\max_{a_i\\in \\mathcal{A}_C}\\frac{|N_{a_i} / N - 1 / |\\mathcal{A}_C||}{1 - 1 / |\\mathcal{A}_C|}" + }, + { + "bbox": [ + 55, + 635, + 297, + 712 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 635, + 297, + 712 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 635, + 297, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 635, + 297, + 712 + ], + "type": "inline_equation", + "content": "N_{a_i}" + }, + { + "bbox": [ + 55, + 635, + 297, + 712 + ], + "type": "text", + "content": " indicate" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 635, + 555, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 555, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 555, + 685 + ], + "type": "text", + "content": "the total number of generated images and images classified as belonging to class " + }, + { + "bbox": [ + 313, + 635, + 555, + 685 + ], + "type": "inline_equation", + "content": "a_{i}" + }, + { + "bbox": [ + 313, + 635, + 555, + 685 + ], + "type": "text", + "content": ", respectively. The lower DR indicates a better balance of attributes. We compute scores for each occupation and average them to obtain the final score." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "content": "For non-target attribute preservation scores, we generate image pairs using the original Stable Diffusion (SD) [1] and" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "17572" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 294, + 283 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 294, + 283 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 294, + 283 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 294, + 283 + ], + "type": "table", + "html": "
MethodFID ↓CLIP-T ↑
-Original SD-26.17
GenderConcept Algebra [28]2.4126.02
UCE [11]11.6525.17
Finetuning Diffusion [26]1.9225.79
Interpret Diffusion [15]15.7824.80
EFA (Ours)0.2326.03
RaceConcept Algebra [28]2.4826.04
UCE [11]5.1626.13
Finetuning Diffusion [26]4.1225.78
Interpret Diffusion [15]21.6623.82
EFA (Ours)0.1326.21
G.R.Concept Algebra [28]2.5326.03
UCE [11]6.5725.71
Finetuning Diffusion [26]4.2625.71
Interpret Diffusion [15]41.6522.45
EFA (Ours)0.4526.20
", + "image_path": "0cf7aea8de1de8f39a4d49efaf67b64cf16b2c93c67b04a3577000077ef4ab4a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 290, + 295, + 324 + ], + "lines": [ + { + "bbox": [ + 55, + 290, + 295, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 290, + 295, + 324 + ], + "type": "text", + "content": "Table 2. Evaluation of model preservation using COCO-no-person. EFA outperforms the baselines in preserving the generation capability of the original model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 342, + 296, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 342, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 54, + 342, + 296, + 498 + ], + "type": "text", + "content": "our approach with the same random seed to evaluate how well non-target attributes are preserved while modifying target attributes. First, we segment the non-human regions in the original image and measure how well they are retained in the image generated by ours using Peak Signal-to-Noise Ratio (PSNR) and Learned Perceptual Image Patch Similarity (LPIPS) [29]. To assess overall image preservation, including contextual elements within human regions (e.g., a doctor's stethoscope or white coat), we compute the DINO score, employing DINO's ability to capture low-level and semantically meaningful visual details. The score measures the cosine similarity between the DINO [5] embeddings of images generated by SD and our method." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 504, + 296, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 504, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 504, + 296, + 647 + ], + "type": "text", + "content": "Lastly, for assessing model preservation, we use COCO-no-person dataset. To measure text fidelity, we use the CLIP-T score, which computes the cosine similarity between CLIP embeddings of the input prompt and the generated images. Additionally, Fréchet Inception Distance (FID) [12] is used to evaluate image quality and diversity. The FID scores are computed against the generation outputs of the original SD. Unlike non-target attribute preservation scores, this evaluation ensures that our method does not degrade the generation capability of the pretrained model when bias mitigation is not required. Note that bold text in all tables indicates the best score." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 653, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 712 + ], + "type": "text", + "content": "Implementation details. The experiments were conducted based on SD v1.5 followed by previous works [10, 20, 26, 31]. For evaluation with WinoBias, we generated 160 images per occupation, using random seeds where the original SD has a detectable face." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 316, + 71, + 555, + 178 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 555, + 178 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 555, + 178 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 555, + 178 + ], + "type": "image", + "image_path": "46ccff1e1fafee023e5745c964c638a4fdd3f91a87d0a55789f7171846f88136.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 185, + 555, + 232 + ], + "lines": [ + { + "bbox": [ + 313, + 185, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 185, + 555, + 232 + ], + "type": "text", + "content": "Figure 5. Attention map visualization when mitigating gender bias using EFA. (a) and (b) show the results of applying " + }, + { + "bbox": [ + 313, + 185, + 555, + 232 + ], + "type": "inline_equation", + "content": "\\mathrm{EFA}_{\\mathrm{gender}}^{\\mathrm{female}}" + }, + { + "bbox": [ + 313, + 185, + 555, + 232 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 185, + 555, + 232 + ], + "type": "inline_equation", + "content": "\\mathrm{EFA}_{\\mathrm{gender}}^{\\mathrm{male}}" + }, + { + "bbox": [ + 313, + 185, + 555, + 232 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 313, + 185, + 555, + 232 + ], + "type": "inline_equation", + "content": "<\\mathrm{sos}>" + }, + { + "bbox": [ + 313, + 185, + 555, + 232 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 185, + 555, + 232 + ], + "type": "inline_equation", + "content": "<\\mathrm{eos}>" + }, + { + "bbox": [ + 313, + 185, + 555, + 232 + ], + "type": "text", + "content": " denote special tokens that mark the beginning and end of a text sequence, respectively." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 240, + 434, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 240, + 434, + 253 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 434, + 253 + ], + "type": "text", + "content": "4.2. Experimental results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 258, + 555, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 555, + 449 + ], + "type": "text", + "content": "Bias mitigation and non-target attribute preservation. Table 1 presents the comparative experimental results of our method and baseline models using the WinoBias dataset. Overall, models struggle to maintain non-target attributes more in " + }, + { + "bbox": [ + 313, + 258, + 555, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{complex}}" + }, + { + "bbox": [ + 313, + 258, + 555, + 449 + ], + "type": "text", + "content": " than " + }, + { + "bbox": [ + 313, + 258, + 555, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{basic}}" + }, + { + "bbox": [ + 313, + 258, + 555, + 449 + ], + "type": "text", + "content": ". Concept Algebra exhibits low performance in bias mitigation. Both UCE [11] and Fine-tuning Diffusion [26] modify model parameters directly, making it challenging to preserve the non-target attributes consistent with the original SD. Interpret Diffusion [15] demonstrates better bias mitigation performance compared to other baselines. However, it exhibits low non-target attribute preservation scores due to its uniform application of fixed semantic vectors without spatial awareness. In contrast, our input-aware architecture adaptively enhances target attributes based on spatial context, enabling effective preservation of non-target attributes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 449, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 554, + 521 + ], + "type": "text", + "content": "Our proposed method outperforms existing baselines by achieving a lower DR while having substantially higher non-target attribute preservation scores. The high preservation scores validate the effectiveness of our approach, showing that the model successfully learns to preserve regions corresponding to non-target attributes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 521, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 700 + ], + "type": "text", + "content": "We hypothesize that the effectiveness in bias mitigation stems from the combination of two key factors: (1) the rich semantic information encoded in text embeddings, which have been pretrained on large-scale datasets, and (2) the capability of U-Net to faithfully reflect these embeddings in the generated output. By utilizing pretrained text embeddings without altering their inherent semantics, our approach ensures that target attributes are seamlessly integrated into the generation process, while U-Net's capacity to translate these embeddings into visual features enables precise control over attribute distributions. The extended experiments on Concept Algebra, provided in the supplementary material, corroborate our hypothesis. Consequently, our method successfully steers the distribution of target attributes while preserving non-target attributes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "type": "text", + "content": "Fig. 4 demonstrates that our method effectively miti-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17573" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 293, + 150 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 293, + 150 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 293, + 150 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 293, + 150 + ], + "type": "table", + "html": "
BiasNon-target attribute P.
DR ↓PSNR ↑LPIPS ↓DINO ↑
Original SD0.71--
EFA w/o Ltrg0.0727.760.05270.890
EFA w/o seg. mask0.0429.310.05620.888
EFA (Ours)0.0632.520.04110.916
", + "image_path": "f9c12192b67d865ec4ab958c87c191d8962b282ff26a48073b6e1569c14e10fa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 206, + 295, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 206, + 295, + 278 + ], + "spans": [ + { + "bbox": [ + 54, + 206, + 295, + 278 + ], + "type": "text", + "content": "gates societal bias, generating outputs that represent a diverse range of gender and race. Furthermore, by leveraging regularization terms and utilizing text embeddings disentangled from other attributes, our method successfully preserves non-target attributes, such as overall layout, pose, and background, while effectively mitigating bias." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 279, + 295, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 279, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 295, + 399 + ], + "type": "text", + "content": "Model preservation. The experimental results on model preservation are presented in Table 2. Our method achieves a significantly lower FID compared to baseline approaches, indicating superior preservation of both the quality and diversity of SD's outputs. Furthermore, our approach yields CLIP-T scores comparable to those of SD, effectively maintaining the text fidelity of the original model by preserving non-target attributes. Qualitative results on model preservation with COCO-no-person and additional experiments using COCO-30K are provided in the Supplementary." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 411, + 115, + 424 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 115, + 424 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 115, + 424 + ], + "type": "text", + "content": "5. Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 431, + 295, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 431, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 431, + 295, + 445 + ], + "type": "text", + "content": "5.1. Effect of regularization and segmentation mask" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 449, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 510 + ], + "type": "text", + "content": "As shown in Table 3, we conduct an ablation study to examine the effects of regularization losses and segmentation masks on non-target attribute preservation. For the experiments, we train models to mitigate gender bias, and the evaluation is performed using " + }, + { + "bbox": [ + 55, + 449, + 295, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{basic}}" + }, + { + "bbox": [ + 55, + 449, + 295, + 510 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "text", + "content": "When training without segmentation masks, we exclude " + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}^{\\mathrm{cf}}" + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "text", + "content": " in training and apply " + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{recon}}" + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "text", + "content": " to the entire image. Since there is no constraint limiting target attribute enhancement, we observe a slight improvement in bias mitigation performance. However, without segmentation masks to distinguish human and non-human regions during training, EFA struggles to localize them precisely, leading to reduced preservation of non-target attributes at inference. Nonetheless, EFA still outperforms baselines even without segmentation masks. We conjecture that this is because counterfactual attributes are primarily activated in bias-relevant regions due to their semantic similarity in cross-attention, which encourages EFA to enhance the target attribute in those regions while ignoring the counterfactual prompt's influence. This allows EFA to approximate the localization of bias-relevant regions without relying on fine-grained masks of the target attribute. Also, when " + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}^{\\mathrm{trg}}" + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "text", + "content": " is not applied, the" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 319, + 70, + 553, + 140 + ], + "blocks": [ + { + "bbox": [ + 55, + 159, + 295, + 194 + ], + "lines": [ + { + "bbox": [ + 55, + 159, + 295, + 194 + ], + "spans": [ + { + "bbox": [ + 55, + 159, + 295, + 194 + ], + "type": "text", + "content": "Table 3. Effectiveness of the regularization loss and the segmentation mask in bias mitigation and non-target attribute preservation. P. and seg. indicate preservation and segmentation, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 70, + 553, + 140 + ], + "lines": [ + { + "bbox": [ + 319, + 70, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 319, + 70, + 553, + 140 + ], + "type": "table", + "html": "
resolution of inputBiasNon-target attribute P.
DR ↓PSNR ↑LPIPS ↓DINO ↑
16×16 (Ours)0.0632.520.04110.916
16×16 - 32×320.0631.790.04070.915
16×16 - 64×640.0630.650.04380.911
", + "image_path": "433d238b35e002dfcbb70fa2400b353bd3771dcbbbab75a36271d809ce1df7bc.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 148, + 555, + 181 + ], + "lines": [ + { + "bbox": [ + 313, + 148, + 555, + 181 + ], + "spans": [ + { + "bbox": [ + 313, + 148, + 555, + 181 + ], + "type": "text", + "content": "Table 4. Impact of input resolution employed by EFA. Extending EFA to higher-resolution layers compromises the preservation of non-target attributes. P. denotes preservation." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 188, + 553, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 188, + 553, + 260 + ], + "spans": [ + { + "bbox": [ + 313, + 188, + 553, + 260 + ], + "type": "text", + "content": "non-target attribute preservation performance deteriorates, even though the bias mitigation performance remains similar. These results highlight the necessity of the regularization mechanisms to preserve non-target attributes across diverse scenarios where the target attribute appears at different extents in the input." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 267, + 542, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 267, + 542, + 280 + ], + "spans": [ + { + "bbox": [ + 313, + 267, + 542, + 280 + ], + "type": "text", + "content": "5.2. Analyzing the influence of resolution in EFA" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 285, + 555, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 555, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 555, + 429 + ], + "type": "text", + "content": "To investigate the impact of input resolution, we integrated EFA into additional layers within the up-blocks while training a model for gender bias mitigation. The performance on " + }, + { + "bbox": [ + 313, + 285, + 555, + 429 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{basic}}" + }, + { + "bbox": [ + 313, + 285, + 555, + 429 + ], + "type": "text", + "content": " is reported in Table 4, where the row highlighted in gray represents our proposed method. When EFA is applied to higher-resolution features, the deviation ratio remains nearly unchanged, while PSNR and DINO scores exhibit a slight decline. This suggests that lower-resolution layers contain richer semantic information, which is crucial for attribute modification. Moreover, extending EFA to multiple layers inevitably introduces additional noise, potentially compromising the preservation of non-target attributes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 435, + 467, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 467, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 467, + 449 + ], + "type": "text", + "content": "5.3. Attention map visualization" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 453, + 554, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 554, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 554, + 525 + ], + "type": "text", + "content": "In Fig. 5, we visualize the attention maps predicted by EFA when the target attribute is enhanced. EFA assigns high values to the human region, effectively enabling gender transformation while preserving the generation capability of the original model, as the impact on accessories, clothing style, and background details remains minimal." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 536, + 388, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 388, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 388, + 548 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 555, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 555, + 712 + ], + "type": "text", + "content": "In this work, we addressed attribute entanglement in diffusion-based T2I models, where bias mitigation often distorts non-target attributes, leading to unintended distribution shifts. We propose EFA, which mitigates human-centric biases while preserving non-target attributes. During training, EFA is guided to attend regions with high semantic relevance to counterfactual prompts, constrained by human segmentation masks. At inference, EFA randomly samples a target attribute and alters the cross-attention in selected layers to ensure a fair distribution of target attributes in the outputs. Experimental results verify that EFA outperforms prior methods in mitigating bias while preserving the original model's output distribution and generative capacity." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "17574" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 297, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 202 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 202 + ], + "type": "text", + "content": "This work was supported by Institute for Information & communications Technology Planning & Evaluation(IITP) grant funded by the Korea government(MSIT) (RS-2019-II190075, Artificial Intelligence Graduate School Program(KAIST)), Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.RS-2021-II212068, Artificial Intelligence Innovation Hub), and the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT)(No. 2022R1A5A7083908)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 223, + 115, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 223, + 115, + 236 + ], + "spans": [ + { + "bbox": [ + 56, + 223, + 115, + 236 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 243, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 243, + 294, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 243, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 61, + 243, + 294, + 275 + ], + "type": "text", + "content": "[1] Stability AI. Stable diffusion. https://huggingface.co/runwayml/stable-diffusion-v1-5, 2022. [Online; accessed 3-November-2023]. 1, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 277, + 296, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 277, + 296, + 343 + ], + "spans": [ + { + "bbox": [ + 61, + 277, + 296, + 343 + ], + "type": "text", + "content": "[2] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. https://cdn.openai.com/papers/dall-e-3.pdf, 2023. [Online; accessed 4-December-2024]. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 344, + 296, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 344, + 296, + 421 + ], + "spans": [ + { + "bbox": [ + 61, + 344, + 296, + 421 + ], + "type": "text", + "content": "[3] Federico Bianchi, Pratyusha Kalluri, Esin Durmus, Faisal Ladhak, Myra Cheng, Debora Nozza, Tatsunori Hashimoto, Dan Jurafsky, James Zou, and Aylin Caliskan. Easily accessible text-to-image generation amplifies demographic stereotypes at large scale. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, pages 1493-1504, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 422, + 296, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 422, + 296, + 476 + ], + "spans": [ + { + "bbox": [ + 62, + 422, + 296, + 476 + ], + "type": "text", + "content": "[4] Manuel Brack, Felix Friedrich, Dominik Hintersdorf, Lukas Struppek, Patrick Schramowski, and Kristian Kersting. Sega: Instructing text-to-image models using semantic guidance. Advances in Neural Information Processing Systems, 36: 25365-25389, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 478, + 296, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 478, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 62, + 478, + 296, + 533 + ], + "type": "text", + "content": "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9650-9660, 2021. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 534, + 296, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 534, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 62, + 534, + 296, + 590 + ], + "type": "text", + "content": "[6] Aditya Chinchure, Pushkar Shukla, Gaurav Bhatt, Kiri Salij, Kartik Hosanagar, Leonid Sigal, and Matthew Turk. Tibet: Identifying and evaluating biases in text-to-image generative models. In European Conference on Computer Vision, pages 429-446. Springer, 2025. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 590, + 296, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 590, + 296, + 645 + ], + "spans": [ + { + "bbox": [ + 62, + 590, + 296, + 645 + ], + "type": "text", + "content": "[7] Jaemin Cho, Abhay Zala, and Mohit Bansal. Dall-eval: Probing the reasoning skills and social biases of text-to-image generation models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3043-3054, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 647, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 296, + 689 + ], + "type": "text", + "content": "[8] Ching-Yao Chuang, Varun Jampani, Yuanzhen Li, Antonio Torralba, and Stefanie Jegelka. Debiasing vision-language models via biased prompts. arXiv preprint arXiv:2302.00070, 2023. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 691, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 691, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 62, + 691, + 296, + 714 + ], + "type": "text", + "content": "[9] Stefan Elfwing, Eiji Uchibe, and Kenji Doya. Sigmoid-weighted linear units for neural network function approxi" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "text", + "content": "mation in reinforcement learning. CoRR, abs/1702.03118, 2017.4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 96, + 555, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 151 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 151 + ], + "type": "text", + "content": "[10] Felix Friedrich, Manuel Brack, Lukas Struppek, Dominik Hintersdorf, Patrick Schramowski, Sasha Luccioni, and Kristian Kersting. Fair diffusion: Instructing text-to-image generation models on fairness. arXiv preprint at arXiv:2302.10893, 2023. 2, 5, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 152, + 554, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 554, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 554, + 207 + ], + "type": "text", + "content": "[11] Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzyńska, and David Bau. Unified concept editing in diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5111-5120, 2024. 1, 2, 5, 6, 7, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 209, + 554, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 554, + 262 + ], + "type": "text", + "content": "[12] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 264, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 297 + ], + "type": "text", + "content": "[13] Eunji Kim, Siwon Kim, Chaehun Shin, and Sungroh Yoon. De-stereotyping text-to-image models through prompt tuning. 2023. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 299, + 554, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 299, + 554, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 299, + 554, + 342 + ], + "type": "text", + "content": "[14] Mingi Kwon, Jaeseok Jeong, and Youngjung Uh. Diffusion models already have a semantic latent space. In The Eleventh International Conference on Learning Representations, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 344, + 554, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 344, + 554, + 399 + ], + "spans": [ + { + "bbox": [ + 316, + 344, + 554, + 399 + ], + "type": "text", + "content": "[15] Hang Li, Chengzhi Shen, Philip Torr, Volker Tresp, and Jindong Gu. Self-discovering interpretable diffusion latent directions for responsible text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 3, 5, 6, 7, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 399, + 554, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 554, + 466 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 554, + 466 + ], + "type": "text", + "content": "[16] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer vision-ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part v 13, pages 740-755. Springer, 2014. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 468, + 554, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 468, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 317, + 468, + 554, + 521 + ], + "type": "text", + "content": "[17] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 523, + 554, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 523, + 554, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 523, + 554, + 578 + ], + "type": "text", + "content": "[18] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 579, + 554, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 554, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 554, + 623 + ], + "type": "text", + "content": "[19] Hadas Orgad, Bahjat Kawar, and Yonatan Belinkov. Editing implicit assumptions in text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7053-7061, 2023. 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 624, + 554, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 554, + 689 + ], + "type": "text", + "content": "[20] Rishubh Parihar, Abhijnya Bhat, Abhipsa Basu, Saswat Mallick, Jogendra Nath Kundu, and R. Venkatesh Babu. Balancing act: Distribution-guided debiasing in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6668-6678, 2024. 2, 3, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 691, + 554, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 554, + 714 + ], + "type": "text", + "content": "[21] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image gen" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17575" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 685 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "eration with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 171 + ], + "type": "text", + "content": "[22] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dólar, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos, 2024. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 295, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 295, + 217 + ], + "type": "text", + "content": "[23] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 274 + ], + "type": "text", + "content": "[24] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. 35:36479-36494, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 275, + 294, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 294, + 338 + ], + "type": "text", + "content": "[25] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. 35:25278-25294, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 341, + 294, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 341, + 294, + 385 + ], + "spans": [ + { + "bbox": [ + 56, + 341, + 294, + 385 + ], + "type": "text", + "content": "[26] Xudong Shen, Chao Du, Tianyu Pang, Min Lin, Yongkang Wong, and Mohan Kankanhalli. Finetuning text-to-image diffusion models for fairness. In The Twelfth International Conference on Learning Representations, 2024. 2, 6, 7, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 386, + 295, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 386, + 295, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 386, + 295, + 441 + ], + "type": "text", + "content": "[27] Robik Shrestha, Yang Zou, Qiuyu Chen, Zhiheng Li, Yusheng Xie, and Siqi Deng. Fairrag: Fair human generation via fair retrieval augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11996-12005, 2024. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 442, + 294, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 442, + 294, + 496 + ], + "spans": [ + { + "bbox": [ + 56, + 442, + 294, + 496 + ], + "type": "text", + "content": "[28] Zihao Wang, Lin Gui, Jeffrey Negrea, and Victor Veitch. Concept algebra for (score-based) text-controlled generative models. In Proceedings of the 37th International Conference on Neural Information Processing Systems, pages 35331-35349, 2023. 6, 7, 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 498, + 294, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 498, + 294, + 552 + ], + "spans": [ + { + "bbox": [ + 56, + 498, + 294, + 552 + ], + "type": "text", + "content": "[29] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 554, + 294, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 554, + 294, + 628 + ], + "spans": [ + { + "bbox": [ + 56, + 554, + 294, + 628 + ], + "type": "text", + "content": "[30] Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, and Kai-Wei Chang. Gender bias in coreference resolution: Evaluation and debiasing methods. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 15–20, 2018. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 632, + 294, + 685 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 632, + 294, + 685 + ], + "spans": [ + { + "bbox": [ + 56, + 632, + 294, + 685 + ], + "type": "text", + "content": "[31] Junlei Zhou, Jiashi Gao, Xiangyu Zhao, Xin Yao, and Xuetao Wei. Association of objects may engender stereotypes: Mitigating association-engendered stereotypes in text-to-image generation. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 3, 7" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "17576" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/FairGen_ Enhancing Fairness in Text-to-Image Diffusion Models via Self-Discovering Latent Directions/207183e6-db8a-436f-8e4f-faec1fb9d3a5_content_list.json b/2025/FairGen_ Enhancing Fairness in Text-to-Image Diffusion Models via Self-Discovering Latent Directions/207183e6-db8a-436f-8e4f-faec1fb9d3a5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..51322e53de0915a7e18ba3c634186b93b8ddfe44 --- /dev/null +++ b/2025/FairGen_ Enhancing Fairness in Text-to-Image Diffusion Models via Self-Discovering Latent Directions/207183e6-db8a-436f-8e4f-faec1fb9d3a5_content_list.json @@ -0,0 +1,1523 @@ +[ + { + "type": "text", + "text": "FairGen: Enhancing Fairness in Text-to-Image Diffusion Models via Self-Discovering Latent Directions", + "text_level": 1, + "bbox": [ + 155, + 128, + 843, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yilei Jiang $^{1}$ Wei-Hong Li $^{1,2*}$ Yiyuan Zhang $^{1}$ Minghong Cai $^{1}$ Xiangyu Yue $^{1,3*}$ $^{1}$ MMLab, CUHK $^{2}$ University of Bristol $^{3}$ SHIAE, CUHK", + "bbox": [ + 158, + 202, + 844, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 273, + 326, + 289 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While Diffusion Models (DM) exhibit remarkable performance across various image generative tasks, they nonetheless reflect the inherent bias presented in the training set. As DMs are now widely used in real-world applications, these biases could perpetuate a distorted worldview and hinder opportunities for minority groups. Existing methods on debiasing DMs usually requires model retraining with a human-crafted reference dataset or additional classifiers, which suffer from two major limitations: (1) collecting reference datasets causes expensive annotation cost; (2) the debiasing performance is heavily constrained by the quality of the reference dataset or the additional classifier. To address the above limitations, we propose FairGen, a plug-and-play method that learns attribute latent directions in a self-discovering manner, thus eliminating the reliance on such reference dataset. Specifically, FairGen consists of two parts: a set of attribute adapters and a distribution indicator. Each adapter in the set aims to learn an attribute latent direction, and is optimized via noise composition through a self-discovering process. Then, the distribution indicator is multiplied by the set of adapters to guide the generation process towards the prescribed distribution. Our method enables debiasing multiple attributes in DMs simultaneously, while remaining lightweight and easily integrable with other DMs, eliminating the need for retraining. Extensive experiments on debiasing gender, racial, and their intersectional biases show that our method outperforms previous SOTA by a large margin.", + "bbox": [ + 89, + 305, + 483, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 758, + 220, + 773 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "State-of-the-art Text-to-Image Diffusion Models (DMs) such as Stable Diffusion [26], DALL-E 3 [20] and Imagegen [27] have demonstrated remarkable performance in generating high-quality images. With the rapid development of DMs, an increasing number of individuals and corporations are choosing to utilize them to serve their own", + "bbox": [ + 89, + 784, + 483, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c666aa8c76200144f15fcef6ff46f095b4eca8a2e185d8b6cb4cec26163707dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 526, + 273, + 895, + 377 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2fe86488056ac03a8517bbb81ec4901cd2a0c8f1e4840b7f47a039e1db4f4f57.jpg", + "image_caption": [ + "Figure 1. Illustration of gender bias associated with different occupations in Stable Diffusion v2.1 and the debiased results for gender by our method. The leadership roles such as CEOs are biased towards male figures, whereas supportive roles such as nurses are biased towards female figures. Our method can generate debiased results regarding the gender bias." + ], + "image_footnote": [], + "bbox": [ + 524, + 382, + 893, + 486 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "purposes. For instance, Stable Diffusion v1.5 has been downloaded over 8 million times on Huggingface, and Midjourney is used by over a million users [5]. However, existing DMs have been found to generate biased content across various demographic factors, such as gender and race [13], which could have harmful effects on society when these models are implemented in real-world applications.", + "bbox": [ + 511, + 598, + 906, + 703 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In Figure 1, we randomly generate several images of two occupations using Stable Diffusion v2.1. Given the prompt of \"A photo of a CEO\", the generated images predominantly depict male figures, reinforcing the stereotype that leadership roles, such as CEOs and doctors, are male-dominated. On the contrary, when the prompt is \"A photo of a nurse\", the majority of generated images depict female figures, reflecting the bias that supportive roles are traditionally associated with women. Regarding racial bias, we randomly generate 1000 images using Stable Diffusion v2.1 with the prompt \"A photo of a worker\". The statistic of 1000 images depicted in Figure 2 shows a strong bias in racial representation, with White individuals making up $71\\%$ of the", + "bbox": [ + 511, + 704, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18411", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/98aa4a433620d2f35c65c4f44639e1828b7f119e834e55e194586642246eee2f.jpg", + "image_caption": [ + "Figure 2. Racial bias in randomly generated 1000 images with the prompt \"A photo of a worker\" using Stable Diffusion v2.1." + ], + "image_footnote": [], + "bbox": [ + 114, + 90, + 250, + 198 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/323da653ce762f0e6445cdba1deada4e466da9adaad83e93f0ce80c356568657.jpg", + "image_caption": [ + "Figure 3. Gender bias in randomly generated 1000 images with the prompt \"A photo of a CEO\" using Stable Diffusion v1.3, v1.4, v1.5, v2.0, and v2.1." + ], + "image_footnote": [], + "bbox": [ + 282, + 93, + 470, + 195 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "total, while minority groups like Middle Eastern, Latino, Black, and Indian each account for only $3 - 4\\%$ . This bias in DM, produce less accurate or fair results for underrepresented populations. We further investigate such bias situation across different versions of DMs. We randomly generate 1000 images with the prompt of \"A photo of a CEO\" using five versions of DMs: Stable Diffusion v1.3 [21], v1.4 [22], v.1.5 [23], v2.0 [24] and v2.1 [25]. Figure 3 demonstrates that gender bias exists across all versions of DMs.", + "bbox": [ + 89, + 305, + 483, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This bias arises from two primary sources. First, the training data, predominantly sourced from the web, contains inherent biases. Second, bias is partially inherited from the CLIP [19] model used in the generation process. In large text-to-image DMs, these biases are even more pronounced, often manifesting in gendered associations with specific professions. Several studies have attempted to mitigate bias in DMs through retraining [30, 36] or approaches that do not require retraining the DM [8, 17, 18]. In training-based methods, Shen et al. [30] propose curating a reference dataset and aligning the distribution of generated images with this dataset. Parihar et al. [18] avoid retraining the DM by leveraging demographic information embedded in the latent features of the denoising U-Net to guide generation. However, their method requires training an additional MLP-based Attribute Distribution Predictor using pseudo labels from existing attribute classifiers. This reliance on accurate attribute classifiers limits its debiasing performance. In contrast, Gandikota et al. [8] and Orgad et al. [17] employ closed-form editing to adjust concepts within DMs without retraining. While training-based methods depend on costly and quality-constrained annotated reference datasets, methods that do not require retraining the DM offer greater implementation efficiency but often exhibit lower debiasing effectiveness.", + "bbox": [ + 91, + 460, + 482, + 837 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these limitations, we propose FairGen, a plug-and-play method that achieves debiased image generation (Figure 1) by automatically learning attribute latent directions, mitigating the reliance on reference datasets.", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "FairGen is composed of two components: a set of attribute adapters and a distribution indicator. Each adapter is trained to learn an attribute-specific latent direction, optimized through noise composition in a self-discovering manner, i.e., our method automatically learns attribute latent directions without relying on a labeled reference dataset. Through noise composition, our method explores and optimizes attribute directions directly from the model's latent space, uncovering patterns without external supervision. At inference stage, the distribution indicator is then applied to select attribute-specific adapters, guiding the generative process towards the desired distribution. We comprehensively evaluate the effectiveness of our approach in debiasing gender, racial, and intersectional biases using occupational prompts. Experimental results demonstrate that our method not only achieves state-of-the-art performance in single and multiple attribute debiasing tasks but also preserves the generation quality of DM. Furthermore, we show that once FairGen is trained on one diffusion model, it can be seamlessly integrated into other models without retuning. Thanks to its strong transferability and plug-and-play functionality, our method offers a practical solution for both individual users and organizations, facilitating the responsible use of diffusion models in future applications.", + "bbox": [ + 511, + 90, + 906, + 453 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our main contributions are as following:", + "bbox": [ + 532, + 455, + 901, + 469 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose FairGen, a novel method for debiasing DMs by learning attribute latent directions in a self-discovering manner, eliminating the reliance on the reference dataset or classifier, and thus significantly reduce the cost.", + "- FairGen is lightweight, plug-and-play and shows good transferability across different DMs, making it more convenient to deploy in the real world.", + "- Extensive experiments show that our method achieves SOTA performance across diverse debiasing tasks while retaining the image generation quality." + ], + "bbox": [ + 532, + 470, + 903, + 636 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 652, + 656, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Bias in Diffusion Models. Diffusion models for text-to-image generation (T2I) have been observed to produce biased and stereotypical images, even when given neutral prompts. Cho et al. [2] found that Stable Diffusion (SD) tends to generate images of males when prompted with occupations, with skin tones predominantly centered around a few shades from the Monk Skin Tone Scale [15]. Seshadri et al. [29] noted that SD reinforces gender-occupation biases presented in its training data. In addition to occupations, Bianchi et al. [1] discovered that simple prompts involving character traits and other descriptors also result in stereotypical images. Luccioni et al. [13] created a tool to compare generated image collections across different genders and ethnicities. Moreover, Wang et al. [34] introduced a text-to-image association test and found that SD tends to", + "bbox": [ + 511, + 674, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18412", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "associate females more with family roles and males more with career-related roles.", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Debiasing Diffusion Models by retraining. Before DMs, previous approaches mainly focus on debiasing GANs by assuming access to the labels of sensitive attributes and ensure no correlation exists between the decision and sensitive attributes [3, 16, 28, 31-33, 35, 36]. More recently, regarding DMs, Shen et al. [30] propose a distributional alignment loss to guide the characteristics of the generated images towards target distribution and use adjusted finetuning to directly optimize losses on the generated images. This method requires a reference training dataset to complete the retraining process, whereas our method does not need such reference dataset, which largely reduces annotation costs. Yu et al. [36] enhance fairness in text-to-image synthesis by incorporating reference images. and leverage visual exemplars to more effectively represent attributes that are difficult to describe in words, such as nuanced variations in skin tones. Li et al. [11] find a interpretable latent vector for a given concept, and achieve debiasing by manipulating the h-space inside DMs during inference stage.", + "bbox": [ + 89, + 125, + 483, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Debiasing Diffusion Models without retraining. Parihar et al. [18] propose Distribution Guidance (DG), which guides the generated images to follow the prescribed attribute distribution. Although DG does not require retraining DMs, it requires training an Attribute Distribution Predictor (ADP) that maps the latent features to the distribution of attributes. Since ADP is trained with pseudo labels generated from existing attribute classifiers, the performance of DG is largely constrained by the accuracy of attribute classifiers. Gandikota et al. [8] and Orgad et al. [17] use closed-form editing approach to edits concepts inside DM without training. Despite the efficiency, these methods are less effective than training-based approaches.", + "bbox": [ + 89, + 414, + 483, + 611 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "**Controlling concepts in Diffusion Models.** Our work is inspired by recent work on controlling concepts [6, 7] in diffusion models. Gandikota et al. [6] propose to fine-tune the DM to erase an undesired concept while maintaining the model's behavior and capabilities, such as removing the NSFW (not safe for work) contents. ConceptSlider [7] also fine-tunes a LoRA adapter in a diffusion model to control the concept in the generated images. Unlike ESD that focuses on erasing concepts and ConceptSlider aims at precisely adjusting the extent of a concept (e.g., controlling the size of eyes), this work aims to debias the Diffusion Models, i.e., controlling the portion of generated images for different attributes such as male and female. Additionally, ConceptSlider learns an adapter per concept and use a scaling factor to control the extent of a concept (e.g., size of eye) while we learn an adapter per direction (e.g., female or male) in an attribute/concept (e.g., gender) and design an inference pipeline with an indicator to automatically select adapters based on a given distribution.", + "bbox": [ + 89, + 613, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminary", + "text_level": 1, + "bbox": [ + 511, + 90, + 638, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Latent Diffusion Models (LDMs) perform the diffusion process within the latent space [26]. During training, noise is added to the encoded latent representation of the input image $x$ , resulting in a noisy latent code $z_{t}$ at each time step $t$ . In the pretraining stage, an autoencoder framework is employed to map images into a lower-dimensional latent space via an encoder: $z = \\mathcal{E}(x)$ . The decoder then reconstructs images from these latent codes: $x \\approx \\mathcal{D}(\\mathcal{E}(x))$ . This process ensures that the latent space retains the essential semantic information of the image. The training objective of the diffusion model in the latent space is given by:", + "bbox": [ + 511, + 109, + 906, + 276 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {L D M}} = \\mathbb {E} _ {z \\sim \\mathcal {E} (x), c, \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (z _ {t}, c, t) \\| _ {2} ^ {2} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 282, + 906, + 310 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon$ is Gaussian noise sampled from a normal distribution $\\mathcal{N}(0,1)$ , $\\epsilon_{\\theta}$ is the denoising network, and $c$ represents any conditioning embeddings (e.g., text or class labels). At the inference stage, a latent code $z_{T}$ is sampled from Gaussian noise at the initial timestep $T$ . The denoising network $\\epsilon_{\\theta}$ is then applied iteratively to remove the noise over several steps, generating a denoised latent representation $z_{0}$ . Finally, the pretrained decoder reconstructs the image from the denoised latent code: $\\hat{x}_0 \\approx \\mathcal{D}(z_0)$ , where $\\hat{x}_0$ is the generated output image.", + "bbox": [ + 511, + 316, + 905, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Classifier-free Guidance aims to modulate image generation by steering the probability distribution towards data that is more probable according to an implicit classifier $p(c \\mid z_t)$ [9]. It operates at the inference phase and the model is jointly trained on both conditional and unconditional denoising tasks. During inference, both the conditional and unconditional denoising scores are derived from the model. The final score $\\tilde{\\epsilon}_{\\theta}(z_t, c, t)$ is then adjusted by weighting the conditioned score more heavily relative to the unconditioned score using a guidance scale $\\alpha > 1$ .", + "bbox": [ + 511, + 467, + 905, + 617 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\epsilon} _ {\\theta} \\left(z _ {t}, c, t\\right) = \\epsilon_ {\\theta} \\left(z _ {t}, t\\right) + \\alpha \\left(\\epsilon_ {\\theta} \\left(z _ {t}, c, t\\right) - \\epsilon_ {\\theta} \\left(z _ {t}, t\\right)\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 627, + 903, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The inference process begins with sampling a latent variable $z_{T} \\sim \\mathcal{N}(0,1)$ , which is subsequently denoised using $\\tilde{\\epsilon}_{\\theta}(z_t,c,t)$ to obtain $z_{t - 1}$ . The denoising is performed iteratively until obtaining $z_{0}$ . Finally, the decoder transforms the latent representation $z_{0}$ back into image space: $x_0 \\gets \\mathcal{D}(z_0)$ .", + "bbox": [ + 511, + 651, + 905, + 743 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 511, + 753, + 604, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by recent work that fine-tune diffusion models for erasing concepts [6] and controlling the level of concepts [7] (e.g., ages, size of eyes), we aim at reducing the bias in the DM by attaching and learning a set of lightweight adapters, each of which represents a category of an attribute (e.g., female of gender), guiding the DM towards an attribute latent direction. Unlike previous work that relies on additional reference datasets and has to finetune the", + "bbox": [ + 511, + 779, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18413", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e53d837e7dee188ed02e842596334fd404f0d6da60cc337a55385adc2773e2a0.jpg", + "image_caption": [ + "Figure 4. Overview of FairGen's training pipeline. Attribute-specific adapters $(M)$ are attached to the cross-attention layers in the denoising UNet. Target group and attribution direction are fed into the DM for composing the noise predictions (Eq. 5), which is used as self-discovering attribute direction guidance to optimize the adapters." + ], + "image_footnote": [], + "bbox": [ + 133, + 83, + 872, + 314 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "whole DM [30], we instead optimize the attached adapters via noise composition through a self-discovering process (detailed in Section 4.2). This significantly reduces the cost in computation and data. During the inference stage (Section 4.3), given a predefined target distribution (e.g., uniform), we introduce a distribution indicator implemented by a gating function to select one corresponding adapter which will be attached to the DM for generating image. In this way, the set of generated images will follow the predefined target distribution, and thus are not biased to some categories of an attribute (if the predefined target distribution is uniform). Our training and inference diagrams are illustrated in Figure 4 and Figure 5. In the following sections, we start elaborating our method in single attribute settings and then extend it to more general ones.", + "bbox": [ + 88, + 388, + 483, + 617 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Attaching Light-weight attribute-specific Adapters", + "text_level": 1, + "bbox": [ + 89, + 632, + 483, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To debias a DM for generating images of a given attribute that contains several categories (e.g., male and female are two categories of gender attribute), we first aim at equip the DM with skills of generating images for each category. To achieve this, we attach a light-weight adapter per category in each layer of the DM inspired by parameter-efficient fine-tuning (PEFT) instead of finetuning the whole model to achieve a good trade-off between performance and computational cost. In this work, we use the 1-dim adapter [14] and only add the adapter to each cross-attention layers of the denoising U-Net, as shown in Figure 4 as we find that attaching adapters to all layers does not help and will also increase the computational cost.", + "bbox": [ + 89, + 671, + 482, + 867 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, for the $i$ -th cross-attention layer parameterized by $\\mathbf{W}_i \\in \\mathbb{R}^{m \\times n}$ in the denoising U-Net, we attach", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "an adapter to the layer to guide the attribute towards a certain category. The adapter consists of two 1-dim vectors: $\\pmb{p} \\in \\mathbb{R}^{m}$ and $\\pmb{q} \\in \\mathbb{R}^{n}$ . The forward process of the $i$ -th cross-attention layer would be updated from $\\pmb{y}_{i} = \\pmb{W}_{i}\\pmb{x}_{i}$ to $\\pmb{y}_{i} = \\pmb{W}_{i}\\pmb{x}_{i} + (\\pmb{q}_{i}^{T}\\pmb{x}_{i})\\cdot \\pmb{p}_{i}$ . $\\pmb{x}_{i} \\in \\mathbb{R}^{n}$ and $\\pmb{y}_{i} \\in \\mathbb{R}^{m}$ represent the input and output of the layer, and superscript $T$ indicates transposition. Thus, adapters in all $(r)$ cross-attention layers for an attribute category $d$ are:", + "bbox": [ + 511, + 390, + 906, + 511 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM _ {d} = \\boldsymbol {Q} ^ {T} \\boldsymbol {P}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 661, + 521, + 906, + 541 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Q = [q_1, q_2, \\ldots, q_r], P = [p_1, p_2, \\ldots, p_r]$ . Each column vector in $Q$ and $P$ , we pad 0 to their end if their dimensions are not the same.", + "bbox": [ + 511, + 551, + 906, + 598 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Optimizing Adapters via self-discovering process", + "text_level": 1, + "bbox": [ + 511, + 609, + 905, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "One straightforward way to optimize the attached adapters is to collect a unbiased reference dataset as in prior work [30]. However, it is expensive to collect such dataset and the quality of the dataset would also limit the performance. To this end, we propose to train the adapters in a self-discovering manner.", + "bbox": [ + 511, + 646, + 905, + 736 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a target group $g_{t}$ (for example, \"CEO\") and model $\\theta$ , we want to optimize the adapters such that the model attached with the adapters generate image $X$ towards certain attribute category $d$ (for example, \"male\" or \"female\") when conditioned on $g_{t}$ :", + "bbox": [ + 511, + 737, + 905, + 813 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP _ {\\theta^ {*}} (X \\mid g _ {t}) \\leftarrow P _ {\\theta} (X \\mid g _ {t}) \\left(P _ {\\theta} (d | X)\\right) ^ {\\eta}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 825, + 903, + 843 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $P_{\\theta}(X|g_t)$ represents the distribution generated by the original model when conditioned on $g_{t}$ , and $\\theta^{*}$ represents the new model equipped with the adapters of $d$ . Note that", + "bbox": [ + 511, + 854, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18414", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/906182312817e1b41230e975d41b8c65ff6a3fa052dc1d8e8d4c4807b699f6ac.jpg", + "image_caption": [ + "Figure 5. Overview of FairGen's inference pipeline. The distribution indicator is generated according the prescribed distribution. Then, it is multiplied by the set of attribute adapters matrices to select attribute matrix adapter. The selected adapter is integrated to the DM with no overhead, guiding the generation towards the prescribed distribution." + ], + "image_footnote": [], + "bbox": [ + 125, + 90, + 875, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$g_{t}$ can be set to an empty string \" \" so that the model will be debiased for all possible groups.", + "bbox": [ + 88, + 364, + 482, + 395 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Applying the Bayes Formula, $P(d|X) = \\frac{P(X|d)P(d)}{P(X)}$ to Eq. 4, taking logarithm on both sides, we are able to derive that the gradient of the log probability $\\nabla \\log P_{\\theta^*}(X|g_t)$ would be proportional to:", + "bbox": [ + 89, + 393, + 483, + 458 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla \\log P _ {\\theta} (X | g _ {t}) + \\eta (\\nabla \\log P _ {\\theta} (X | d) - \\nabla \\log P _ {\\theta} (X | g _ {t})) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 468, + 483, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on Tweedie's formula [4] and the reparametrization trick of Classifier-free guidance [9], we introduce a time-varying noising process and represent each score (gradient of log probability) as a denoising prediction $\\epsilon(X_{t}, c_{t}, t)$ , which leads to our learning objective for adapters of category $d$ of an attribute:", + "bbox": [ + 89, + 510, + 483, + 602 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {\\theta^ {*}} \\left(X _ {t}, g _ {t}, t\\right) \\leftarrow \\epsilon_ {\\theta} \\left(X _ {t}, g _ {t}, t\\right) + \\eta \\left(\\epsilon_ {\\theta} \\left(X _ {t}, d, t\\right) - \\epsilon_ {\\theta} \\left(X _ {t}, g _ {t}, t\\right)\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 612, + 511, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Therefore, the guidance loss for optimizing adapters of category $d$ can be defined as follows:", + "bbox": [ + 89, + 652, + 482, + 683 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {G u i d a n c e}} = \\mathbb {L} _ {x _ {t}, t} \\left[ \\| \\epsilon_ {\\theta^ {*}} (X _ {t}, g _ {t}, t) - \\epsilon_ {L} \\| ^ {2} \\right],\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 693, + 452, + 722 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {L} = \\epsilon_ {\\theta} (X _ {t}, g _ {t}, t) + \\eta (\\epsilon_ {\\theta} (X _ {t}, d, t) - \\epsilon_ {\\theta} (X _ {t}, g _ {t}, t)),\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 722, + 452, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the goal is to align the noise $\\epsilon_{\\theta^{*}}(X_{t},g_{t},t)$ of the $\\theta^{*}$ (DM with adapters of category $d$ ) and the noise composition $\\epsilon_{L}$ of group $g_{t}$ and category $d$ from the frozen DM $\\theta$ And hence, our adapters optimization does not require any additional data.", + "bbox": [ + 89, + 748, + 483, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Inference with distribution indicator.", + "text_level": 1, + "bbox": [ + 89, + 833, + 410, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After the optimization, we obtain a set of adapters $\\mathcal{M} = \\{M_1,M_2,\\dots ,M_t\\}$ for $t$ categories of a given attribute. For instance, $t = 2$ for gender bias and $t = 4$ for racial bias", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "in our case. As shown in Figure 5, at the inference stage, we introduce a distribution indicator $h$ . Given a prescribed distribution $f_{\\theta}^{a}$ (e.g., uniform distribution), we define its Probabilistic Mass Function (PMF) as follows:", + "bbox": [ + 511, + 364, + 906, + 425 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP (X = x) = f _ {\\theta} ^ {a} (x) = \\left\\{ \\begin{array}{l l} \\frac {1}{t} & \\text {f o r} x \\in \\{1, 2, \\dots , t \\}, \\\\ 0 & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 441, + 906, + 482 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $t$ is the number of possible categories for each attribute, and $\\{1,2,\\dots ,t\\}$ represents the set of all possible values for the random variable $X$ . The parameter $\\theta$ controls the shape of the prescribed distribution, and in the case of the uniform distribution, $\\theta$ implies that all outcomes in the set have equal probability, i.e., $\\frac{1}{t}$ .", + "bbox": [ + 511, + 497, + 905, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, we randomly sample an index $k$ from the prescribed distribution $f_{\\theta}^{a}$ . The distribution indicator $\\pmb{h} \\in \\mathbb{R}^t$ is formulated to reflect the chosen index $k$ as follows:", + "bbox": [ + 511, + 590, + 905, + 635 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {h} _ {i} := \\left\\{ \\begin{array}{l l} 1 & \\text {i f} i = k, \\\\ 0 & \\text {i f} i \\neq k, \\end{array} \\right. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 637, + 657, + 906, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $i \\in \\{1,2,\\dots,t\\}$ , and $k$ is the sampled index based on the distribution $f_{\\theta}^{a}$ . The indicator $h$ is a one-hot vector, where the $k$ -th element is 1, indicating the sampled outcome, and all other elements are 0. After obtaining the distribution indicator, it is multiplied by the set of trained attribute matrix adapters. Subsequently, the final weight change $\\Delta W$ is given by:", + "bbox": [ + 511, + 713, + 905, + 820 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\boldsymbol {W} = \\boldsymbol {h} \\cdot \\mathcal {M}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 655, + 837, + 906, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and the model is updated as $\\mathbf{W} \\gets \\mathbf{W} + \\alpha \\Delta \\mathbf{W}$ , where $\\alpha$ is a scaling factor controlling the strength of the guidance.", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18415", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b978c236a569da3c60a08d20df040305b0ac45295bf23a9ad236928d06ae5e4a.jpg", + "image_caption": [ + "Figure 6. Images generated from the original Stable Diffusion model (left) and FairGen (right) for gender debiasing with the prompt \"A photo of a worker\". The gender distribution shifts from a male-to-female ratio of 13:2 in the original model to 8:7 with FairGen." + ], + "image_footnote": [], + "bbox": [ + 125, + 89, + 869, + 273 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Debiasing Multiple Attributes (intersectional debiasing)", + "text_level": 1, + "bbox": [ + 89, + 328, + 480, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our method can be inherently extended to debiasing multiple attributes in diffusion models (DM). Specifically, in the case of multiple attribute debiasing, the adapter for each attribute should not interfere with the others. Otherwise, the most recently trained adapter could degrade the performance of previously learned adapters. For each attribute to be debiased, we denote a set of adapter parameters as $\\{P_t, Q_t\\}$ . We have $P_t = [p_t^1, p_t^2, \\dots, p_t^r]$ , $Q_t = [q_t^1, q_t^2, \\dots, q_t^r]$ .", + "bbox": [ + 89, + 366, + 483, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To avoid interference between attribute adapters, we extend Eq. 7 by introducing an orthogonal regularization loss that regularizes the vector subspace spanned by each $P_{t}$ and $Q_{t}$ to be orthogonal to each other:", + "bbox": [ + 89, + 503, + 483, + 565 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {o r t h}} = \\sum_ {i = 1} ^ {t - 1} \\left(\\boldsymbol {P} _ {i} \\times \\boldsymbol {P} _ {t} + \\boldsymbol {Q} _ {i} \\times \\boldsymbol {Q} _ {t}\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 569, + 482, + 609 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {G u i d a n c e}} + \\gamma \\cdot \\mathcal {L} _ {\\text {o r t h}}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 621, + 480, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 89, + 667, + 215, + 684 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we first describe the experimental details and evaluation metrics, then present a quantitative and qualitative analysis of our method. Please refer to Appendix for more results.", + "bbox": [ + 89, + 693, + 483, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental Details", + "text_level": 1, + "bbox": [ + 89, + 761, + 290, + 776 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. We use Stable Diffusion v2.1 for all methods. We employ the prompt template \"a photo of the face of a {occupation}, a person\". At inference time, for each bias, we generate 100 images per occupation across 100 occupations, resulting in a total of 10,000 images. We set $\\eta = \\alpha = 0.3$ , and train for 1000 iterations with a learning rate of 1e-5. For gender bias, we use the CelebA [12] dataset to train a binary classifier with two", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "categories: {male, female}. For racial bias, we use the Fair-Face [10] dataset to train a classifier with the following four categories: WMELH = {White, Middle Eastern, Latino Hispanic}, Asian = {East Asian, Southeast Asian}, Black, and Indian. Please refer to the supplementary for more details.", + "bbox": [ + 511, + 329, + 905, + 405 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared methods. In this work, we compare our method with recent state-of-the-art (SOTA) methods, including two retraining-based methods, Finetuning for Fairness (F4Fair) [30] and Inclusive Text-to-Image Generation (ITI-GEN) [37], an approach that does not require retraining DM, H-Distribution Guidance (H Guidance) [18], a closed-form editing approach, Unified Concept Editing (UCE) [8] and an h-space manipulation approach, Interpretable Diffusion [11]. Please refer to Appendix for more detailed introduction of these methods.", + "bbox": [ + 511, + 426, + 906, + 577 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 513, + 592, + 696, + 606 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate all methods in three metrics:", + "bbox": [ + 511, + 614, + 787, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fairness Discrepancy (FD). Following prior work[18], we adopt the Fairness Discrepancy (FD) metric. For an attribute $a$ and target distribution $p_{\\theta}^{a}$ , we use a high-accuracy classifier $\\mathcal{C}_a$ to compute the fairness performance: $||p_{\\theta}^{a} - \\mathbb{E}_{\\mathbf{x}\\sim p_{\\theta}(\\mathbf{x})}(\\mathbf{y})||_2$ where $\\mathbf{y}$ is the softmax output of $\\mathcal{C}_a(\\mathbf{x})$ . The target distribution $p_{\\theta}^{a}$ can be any user-defined vector, typically uniform. A lower FD score indicates a closer match to the target distribution.", + "bbox": [ + 511, + 633, + 905, + 755 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\mathbf{CLIP}_{\\mathrm{sim}}$ . Besides fairness, the debiased model should maintain the ability to generate images that are semantically close to their text prompts. Therefore, following [30], we report the CLIP similarity score $\\mathbf{CLIP}_{\\mathrm{sim}}$ between the generated image and its prompt.", + "bbox": [ + 511, + 758, + 905, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "BRISQUE. The generated image quality is also important as the debiasing process should not influence the image generation ability. We use the BRISQUE metric for evaluate the quality of the generated images.", + "bbox": [ + 511, + 839, + 905, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18416", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f2805a736552659f68ca2a66baf6ac651c7b92f3064bb42ec86d667c73cff5a6.jpg", + "image_caption": [ + "(a) original" + ], + "image_footnote": [], + "bbox": [ + 124, + 89, + 478, + 263 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ad12b6be35af0bdd06443189bd11d335fd8d62b67208a9c383b9fca1f5116666.jpg", + "image_caption": [ + "(b) debiased (Ours)", + "Figure 7. Images generated from the original Stable Diffusion model (left) and FairGen (right) for racial debiasing with the prompt \"A photo of a sportsman\". The racial group distribution shifts from WMELH : Asian : Black : Indian = 7:2:5:1 in the original model to 4:4:4:3 with FairGen." + ], + "image_footnote": [], + "bbox": [ + 519, + 90, + 874, + 263 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c0dbcb810c4c4ffd5a8065c6fda40a940dd595d6a418621bd2367e7383e33292.jpg", + "table_caption": [ + "Table 1. Comparisons of our method to the SOTA methods in gender bias over two predefined distributions: $f_{\\theta}^{1} = (0.5, 0.5)$ and $f_{\\theta}^{2} = (0.2, 0.8)$ , representing the probability of male and female respectively." + ], + "table_footnote": [], + "table_body": "
MethodFD ↓CLIPsim ↑BRISQUE ↑
f1θf2θf1θf2θf1θf2θ
Original SD0.4240.8470.380.3838.6538.69
F4Fair0.1650.3870.360.3538.2137.64
ITI-GEN0.0930.4720.320.3337.6237.96
H Guidance0.1180.3980.310.3238.5438.65
UCE0.2840.5360.360.2937.1236.54
InterDiff0.1770.3860.350.3438.3238.04
Ours0.0030.0050.380.3838.4638.72
", + "bbox": [ + 94, + 383, + 450, + 522 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9c6e847e259b8b56458b4d9c4ff228cf17b8c1f2b2b417d216f1aad281c0aace.jpg", + "table_caption": [ + "Table 2. Comparisons of our method to the SOTA methods in racial bias over two distributions: $f_{\\theta}^{1} = (0.25, 0.25, 0.25, 0.25)$ and $f_{\\theta}^{2} = (0.4, 0.3, 0.2, 0.1)$ , representing the probability of WMELH, Asian, Black, and Indian respectively." + ], + "table_footnote": [], + "table_body": "
MethodFD ↓CLIPsim ↑BRISQUE ↑
f1θf2θf1θf2θf1θf2θ
Original SD0.3840.4970.460.4138.9438.65
F4Fair0.2200.3050.430.4038.8038.45
ITI-GEN0.1310.5420.310.3136.6137.09
H Guidance0.1500.2850.410.3738.6538.60
UCE0.2900.4600.450.3937.9037.80
InterDiff0.1500.3910.400.3837.6437.82
Ours0.0950.1500.460.4138.9638.66
", + "bbox": [ + 517, + 383, + 874, + 522 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Results", + "text_level": 1, + "bbox": [ + 89, + 546, + 184, + 560 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparisons in gender debiasing. Table 1 demonstrates that our method outperforms others in mitigating gender bias over two predefined distributions: $f_{\\theta}^{1} = (0.5,0.5)$ , representing equal likelihood of male and female, and $f_{\\theta}^{2} = (0.2,0.8)$ , where male and female have a $20\\%$ and $80\\%$ probability, respectively. Original SD model exhibits high FD scores (0.424 at $f_{\\theta}^{1}$ and 0.847 at $f_{\\theta}^{2}$ ), indicating significant bias towards one gender. While previous methods like F4Fair, ITI-GEN, H Guidance, and UCE reduce bias to some extent, our method achieves the most reduction, with FD scores of 0.003 at $f_{\\theta}^{1}$ and 0.005 at $f_{\\theta}^{2}$ . Additionally, our method preserves semantic similarity (CLIPsim of 0.38) and image quality (BRISQUE scores of 38.46 and 38.72), matching or slightly surpassing the Original SD model. The qualitative results in Figure 6 further demonstrate that our approach effectively mitigates gender bias without compromising image quality or semantic coherence.", + "bbox": [ + 88, + 565, + 482, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparisons in racial debiasing. Table 2 compares methods over two distributions: $f_{\\theta}^{1} = (0.25, 0.25, 0.25, 0.25)$ , representing equal probabilities for WMELH, Asian, Black, and Indian, and $f_{\\theta}^{2} = (0.4, 0.3, 0.2, 0.1)$ , with higher probabilities for", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "WMELH and Asian, and lower probabilities for Black and Indian. Original SD model shows significant racial bias, with FD scores of 0.384 for $f_{\\theta}^{1}$ and 0.497 for $f_{\\theta}^{2}$ , indicating poor calibration to either distribution.", + "bbox": [ + 511, + 547, + 906, + 607 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Debiasing methods such as F4Fair, H Guidance, and UCE reduce this bias, with H Guidance achieving relatively lower FD scores. However, our method performs the best, reducing bias to 0.095 for $f_{\\theta}^{1}$ and 0.150 for $f_{\\theta}^{2}$ . In terms of semantic similarity, the Original SD model sets a strong baseline (0.46 for $f_{\\theta}^{1}$ and 0.41 for $f_{\\theta}^{2}$ ), and our method maintains this high alignment, ensuring that debiasing does not impair semantic accuracy. Regarding image quality, our approach slightly improves upon the Original SD model, achieving the highest scores (39.10 for $f_{\\theta}^{1}$ and 38.90 for $f_{\\theta}^{2}$ ). Again, the qualitative results in Figure 7 further verify that our method outperforms others in reducing racial bias while preserving both semantic similarity and image quality.", + "bbox": [ + 511, + 608, + 906, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparisons in intersectional debiasing We also consider a more complex challenge of intersectional debiasing, i.e., jointly debiasing both gender and racial biases. The target distribution $f_{\\theta}$ is set to be uniform for both gender and racial bias. Table 4 shows that Original SD model exhibits significant bias with an FD score of 0.214. While F4Fair", + "bbox": [ + 511, + 809, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18417", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/183813f32b6d7ae60153c55868940445229fb2bca8d2059f99d478f17b28494f.jpg", + "table_caption": [ + "Table 3. Transferability of proposed method across different DMs" + ], + "table_footnote": [], + "table_body": "
TrainTestGenderRacial
FD ↓CLIPsim ↑BRISQUE ↑FD ↓CLIPsim ↑BRISQUE ↑
v2.1v1.40.0060.3738.400.0970.4538.90
v2.1v1.50.0060.3738.420.0960.4538.92
v2.1v2.00.0080.3838.450.0950.4638.95
v2.1v2.10.0030.3838.460.0950.4638.96
", + "bbox": [ + 256, + 104, + 741, + 196 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d358588df785b2a6ad696597e66a87279ca8e2d6a8b84eebbf02ac3344ca6ebf.jpg", + "table_caption": [ + "Table 4. Evaluation of mitigating intersectional bias." + ], + "table_footnote": [], + "table_body": "
MethodFD ↓CLIPsim↑BRISQUE ↑
Original SD0.2140.3539.24
F4Fair0.1450.3638.90
ITI-GEN0.0980.3238.01
H Guidance0.1300.3338.75
UCE0.1800.3438.60
InterDiff0.1580.3538.67
Ours0.0470.3639.50
", + "bbox": [ + 135, + 234, + 439, + 361 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and H Guidance reduce this bias to 0.145 and 0.130, respectively, our method achieves a much lower FD score of 0.047, reflecting a substantial improvement in fairness.", + "bbox": [ + 89, + 386, + 482, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For semantic similarity, the Original SD scores 0.35, and F4Fair slightly improves it to 0.36. Our method matches this performance, maintaining semantic coherence while reducing bias. Regarding image quality, the Original SD has a BRISQUE score of 39.24, while our method improves it to 39.50, indicating enhanced perceptual quality. These results demonstrate that our method excels at jointly debiasing both gender and racial biases, significantly reducing bias without sacrificing semantic accuracy or image quality. This highlights its robustness and practicality in real-world settings where multiple biases are present.", + "bbox": [ + 88, + 431, + 482, + 598 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Transferability across different DMs. We further evaluate the transferability of our method by training it with Stable Diffusion v2.1 and testing it over other versions, and results are reported in Table 3. From the table we can see that the performance of in all metrics are close to the optimal results achieved when training and testing are performed using the same model version (v2.1). While there is a slight increase in FD when tested on v1.4, v1.5, and v2.0, the differences are minor, indicating that our method can effectively generalize across different model versions. This robustness highlights the method's capability to transfer learned features across varying model conditions.", + "bbox": [ + 89, + 599, + 482, + 781 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 791, + 243, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Attaching adapters in U-Net. The results in Table 5 illustrate the impact of attaching adapters to different layers of the U-Net in the DM. When adapters are attached to all layers, the fairness score (FD) is 0.165, with a CLIP similarity score of 0.33 and a BRISQUE score of 35.00. The best results are obtained when adapters are attached only to", + "bbox": [ + 89, + 809, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1988f87967e91b293fe802a5d007d7b0fe2874892afd4098c8d50c0ffb9a463f.jpg", + "table_caption": [ + "Table 5. Ablation of layers that are attached adapters." + ], + "table_footnote": [], + "table_body": "
LocationFD ↓CLIPsim ↑BRISQUE ↑
All layers0.1650.3335.00
Non-CA layers0.1580.3437.00
CA layers0.0470.3639.50
", + "bbox": [ + 550, + 234, + 870, + 303 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the CA layers. This configuration significantly reduces the FD score to 0.047, increases semantic similarity $(\\mathrm{CLIP}_{\\mathrm{sim}} = 0.36)$ , and enhances image quality with a BRISQUE score of 39.50. These highlights that focusing the adapters on the cross-attention layers leads to the most substantial improvements in both fairness and image quality.", + "bbox": [ + 511, + 330, + 906, + 421 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of orthogonal regularization (OR). Table 6 demonstrates the impact of orthogonal regularization (OR). Without OR, the FD score is 0.143, semantic similarity $(\\mathrm{CLIP}_{\\mathrm{sim}})$ drops to 0.29, and image quality (BRISQUE) is 37.92. When OR is applied, performance improves significantly across all metrics, with a much lower FD score of 0.047, higher semantic similarity of 0.36, and improved image quality (BRISQUE = 39.50). This highlights the effectiveness of orthogonal regularization in reducing bias and improving overall performance.", + "bbox": [ + 511, + 425, + 908, + 578 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5bb5d3ab70c05f3e35f407642149a4953e1d9581778cccdf3a4aa4b3494646f0.jpg", + "table_caption": [ + "Table 6. Orthogonal Ablation" + ], + "table_footnote": [], + "table_body": "
LocationFD ↓CLIPsim ↑BRISQUE ↑
Ours w\\o OR0.1430.2937.92
Ours w\\OR0.0470.3639.50
", + "bbox": [ + 550, + 607, + 870, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 696, + 633, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose FairGen, a plug-and-play method that learns attribute latent directions in a self-discovering manner. Our method can not only jointly debias multiple attributes in DMs, but also enables the generated images to follow a prescribed attribute distribution. It is lightweight and can be integrated with other DMs without retraining. Extensive experiments on debiasing gender, racial, and their intersectional biases show that our method outperforms previous SOTA by a large margin. We believe that our work marks a critical advancement in addressing harmful societal stereotypes within diffusion models, and it contributes to the ethical real-world applications of text-to-image DMs.", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18418", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the National Natural Science Foundation of China under Grant No. 62306261, and the Shun Hing Institute of Advanced Engineering (SHIAE) Fund under Grant No. 8115074.", + "bbox": [ + 89, + 113, + 483, + 183 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 196, + 187, + 213 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Federico Bianchi, Pratyusha Kalluri, Esin Durmus, Faisal Ladhak, Myra Cheng, Debora Nozza, Tatsunori Hashimoto, Dan Jurafsky, James Zou, and Aylin Caliskan. Easily accessible text-to-image generation amplifies demographic stereotypes at large scale. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, pages 1493-1504, Chicago, Illinois, USA, 2023. ACM. 2", + "[2] Jaemin Cho, Abhay Zala, and Mohit Bansal. Dall-eval: Probing the reasoning skills and social biases of text-to-image generative transformers. In Proceedings of the International Conference on Computer Vision (ICCV), Paris, France, 2023. IEEE. 2", + "[3] Kristy Choi, Aditya Grover, Trisha Singh, Rui Shu, and Stefano Ermon. Fair generative modeling via weak supervision. In International Conference on Machine Learning, pages 1887-1898. PMLR, 2020. 3", + "[4] Bradley Efron. Tweedie's formula and selection bias, 2011. 5", + "[5] Mureji Fatunde and Crystal Tse. Digital media company stability ai raises funds at $1 billion value. https://www.bloomberg.com/news/articles/2022-10-17/digital-media-firm-stability-ai-raises-funds-at-1-billion-value, 2022.1", + "[6] Rohit Gandikota, Joanna Materzynska, Jaden Fiitto-Kaufman, and David Bau. Erasing concepts from diffusion models, 2023. 3", + "[7] Rohit Gandikota, Joanna Materzynska, Tingrui Zhou, Antonio Torralba, and David Bau. Concept sliders: Lora adaptors for precise control in diffusion models, 2023. 3", + "[8] Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzyńska, and David Bau. Unified concept editing in diffusion models. arXiv preprint arXiv:2308.14761, 2023. 2, 3, 6, 1", + "[9] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance, 2022. 3, 5", + "[10] Jungseock Joo. Fairface attribute model. https://github.com/jojoa/fairface, 2021. 6, 1", + "[11] Hang Li, Chengzhi Shen, Philip Torr, Volker Tresp, and Jindong Gu. Self-discovering interpretable diffusion latent directions for responsible text-to-image generation, 2024. 3, 6, 1", + "[12] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 3730-3738, Santiago, Chile, 2015. IEEE. 6, 1", + "[13] Alexandra Sasha Luccioni, Christopher Akiki, Margaret Mitchell, and Yacine Jernite. Stable bias: Analyzing so" + ], + "bbox": [ + 93, + 222, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "cietal representations in diffusion models. arXiv preprint arXiv:2303.11408, 2023. 1, 2", + "[14] Mengyao Lyu, Yuhong Yang, Haiwen Hong, Hui Chen, Xuan Jin, Yuan He, Hui Xue, Jungong Han, and Guiguang Ding. One-dimensional adapter to rule them all: Concepts, diffusion models and erasing applications, 2024. 4", + "[15] Ellis Monk. The monk skin tone scale, 2023. 2", + "[16] Junhyun Nam, Sangwoo Mo, Jaeho Lee, and Jinwoo Shin. Breaking the spurious causality of conditional generation via fairness intervention with corrective sampling. Transactions on Machine Learning Research, 2023. 3", + "[17] Hadas Orgad, Bahjat Kawar, and Yonatan Belinkov. Editing implicit assumptions in text-to-image diffusion models. 2023. 2, 3", + "[18] Rishubh Parihar, Abhijnya Bhat, Abhipsa Basu, Saswat Mallick, Jogendra Nath Kundu, and R. Venkatesh Babu. Balancing act: Distribution-guided debiasing in diffusion models, 2024. 2, 3, 6, 1", + "[19] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2", + "[20] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1", + "[21] Robin Rombach. Stable diffusion v1-3 model card, 2022. 2", + "[22] Robin Rombach. Stable diffusion v1-4 model card, 2022. 2", + "[23] Robin Rombach. Stable diffusion v1-5 model card, 2022. 2", + "[24] Robin Rombach. Stable diffusion v2-0 model card, 2022. 2", + "[25] Robin Rombach. Stable diffusion v2-1 model card, 2022. 2", + "[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3", + "[27] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1", + "[28] Prasanna Sattigeri, Samuel C Hoffman, Vijil Chenthamarakshan, and Kush R Varshney. Fairness gan: Generating datasets with fairness properties using a generative adversarial network. IBM Journal of Research and Development, 63 (4/5):3-1, 2019. 3", + "[29] Preethi Seshadri, Sameer Singh, and Yanai Elazar. The bias amplification paradox in text-to-image generation. arXiv preprint arXiv:2308.00755, 2023. 2", + "[30] Xudong Shen, Chao Du, Tianyu Pang, Min Lin, Yongkang Wong, and Mohan Kankanhalli. Finetuning text-to-image diffusion models for fairness, 2024. 2, 3, 4, 6, 1", + "[31] Christopher TH Teo, Milad Abdollahzadeh, and Ngai-Man Cheung. Fair generative models via transfer learning. In Pro" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18419", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ceedings of the AAAI Conference on Artificial Intelligence, pages 2429-2437, 2023. 3", + "[32] Soobin Um and Changho Suh. A fair generative model using lecam divergence. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10034-10042, 2023.", + "[33] Boris van Breugel, Trent Kyono, Jeroen Berrevoets, and Michaela van der Schaar. Decaf: Generating fair synthetic data using causally-aware generative networks. Advances in Neural Information Processing Systems, 34:22221-22233, 2021. 3", + "[34] Zihao Wang, Lin Gui, Jeffrey Negrea, and Victor Veitch. Concept algebra for score-based conditional model. In ICML 2023 Workshop on Structured Probabilistic Inference & Generative Modeling, Honolulu, HI, 2023. PMLR. 2", + "[35] Depeng Xu, Shuhan Yuan, Lu Zhang, and Xintao Wu. Fairgan: Fairness-aware generative adversarial networks. In 2018 IEEE International Conference on Big Data (Big Data), pages 570-575. IEEE, 2018. 3", + "[36] Ning Yu, Ke Li, Peng Zhou, Jitendra Malik, Larry Davis, and Mario Fritz. Inclusive gan: Improving data and minority coverage in generative models. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 377-393. Springer, 2020. 2, 3, 1", + "[37] Cheng Zhang, Xuanbai Chen, Siqi Chai, Chen Henry Wu, Dmitry Lagun, Thabo Beeler, and Fernando De la Torre. Itigen: Inclusive text-to-image generation, 2023. 6" + ], + "bbox": [ + 91, + 90, + 482, + 474 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18420", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/FairGen_ Enhancing Fairness in Text-to-Image Diffusion Models via Self-Discovering Latent Directions/207183e6-db8a-436f-8e4f-faec1fb9d3a5_model.json b/2025/FairGen_ Enhancing Fairness in Text-to-Image Diffusion Models via Self-Discovering Latent Directions/207183e6-db8a-436f-8e4f-faec1fb9d3a5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e53a6c466ec9a4791a18ef864b3517deb4ef2088 --- /dev/null +++ b/2025/FairGen_ Enhancing Fairness in Text-to-Image Diffusion Models via Self-Discovering Latent Directions/207183e6-db8a-436f-8e4f-faec1fb9d3a5_model.json @@ -0,0 +1,2024 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.156, + 0.13, + 0.844, + 0.177 + ], + "angle": 0, + "content": "FairGen: Enhancing Fairness in Text-to-Image Diffusion Models via Self-Discovering Latent Directions" + }, + { + "type": "text", + "bbox": [ + 0.16, + 0.203, + 0.846, + 0.24 + ], + "angle": 0, + "content": "Yilei Jiang\\(^{1}\\) Wei-Hong Li\\(^{1,2*}\\) Yiyuan Zhang\\(^{1}\\) Minghong Cai\\(^{1}\\) Xiangyu Yue\\(^{1,3*}\\) \n\\(^{1}\\)MMLab, CUHK \\(^{2}\\)University of Bristol \\(^{3}\\)SHIAE, CUHK" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.274, + 0.327, + 0.29 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.306, + 0.485, + 0.732 + ], + "angle": 0, + "content": "While Diffusion Models (DM) exhibit remarkable performance across various image generative tasks, they nonetheless reflect the inherent bias presented in the training set. As DMs are now widely used in real-world applications, these biases could perpetuate a distorted worldview and hinder opportunities for minority groups. Existing methods on debiasing DMs usually requires model retraining with a human-crafted reference dataset or additional classifiers, which suffer from two major limitations: (1) collecting reference datasets causes expensive annotation cost; (2) the debiasing performance is heavily constrained by the quality of the reference dataset or the additional classifier. To address the above limitations, we propose FairGen, a plug-and-play method that learns attribute latent directions in a self-discovering manner, thus eliminating the reliance on such reference dataset. Specifically, FairGen consists of two parts: a set of attribute adapters and a distribution indicator. Each adapter in the set aims to learn an attribute latent direction, and is optimized via noise composition through a self-discovering process. Then, the distribution indicator is multiplied by the set of adapters to guide the generation process towards the prescribed distribution. Our method enables debiasing multiple attributes in DMs simultaneously, while remaining lightweight and easily integrable with other DMs, eliminating the need for retraining. Extensive experiments on debiasing gender, racial, and their intersectional biases show that our method outperforms previous SOTA by a large margin." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.759, + 0.222, + 0.775 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.785, + 0.484, + 0.877 + ], + "angle": 0, + "content": "State-of-the-art Text-to-Image Diffusion Models (DMs) such as Stable Diffusion [26], DALL-E 3 [20] and Imagegen [27] have demonstrated remarkable performance in generating high-quality images. With the rapid development of DMs, an increasing number of individuals and corporations are choosing to utilize them to serve their own" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.274, + 0.896, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.383, + 0.895, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.489, + 0.907, + 0.572 + ], + "angle": 0, + "content": "Figure 1. Illustration of gender bias associated with different occupations in Stable Diffusion v2.1 and the debiased results for gender by our method. The leadership roles such as CEOs are biased towards male figures, whereas supportive roles such as nurses are biased towards female figures. Our method can generate debiased results regarding the gender bias." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.704 + ], + "angle": 0, + "content": "purposes. For instance, Stable Diffusion v1.5 has been downloaded over 8 million times on Huggingface, and Midjourney is used by over a million users [5]. However, existing DMs have been found to generate biased content across various demographic factors, such as gender and race [13], which could have harmful effects on society when these models are implemented in real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In Figure 1, we randomly generate several images of two occupations using Stable Diffusion v2.1. Given the prompt of \"A photo of a CEO\", the generated images predominantly depict male figures, reinforcing the stereotype that leadership roles, such as CEOs and doctors, are male-dominated. On the contrary, when the prompt is \"A photo of a nurse\", the majority of generated images depict female figures, reflecting the bias that supportive roles are traditionally associated with women. Regarding racial bias, we randomly generate 1000 images using Stable Diffusion v2.1 with the prompt \"A photo of a worker\". The statistic of 1000 images depicted in Figure 2 shows a strong bias in racial representation, with White individuals making up \\(71\\%\\) of the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18411" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.116, + 0.091, + 0.25, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.203, + 0.265, + 0.273 + ], + "angle": 0, + "content": "Figure 2. Racial bias in randomly generated 1000 images with the prompt \"A photo of a worker\" using Stable Diffusion v2.1." + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.094, + 0.472, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.276, + 0.203, + 0.483, + 0.273 + ], + "angle": 0, + "content": "Figure 3. Gender bias in randomly generated 1000 images with the prompt \"A photo of a CEO\" using Stable Diffusion v1.3, v1.4, v1.5, v2.0, and v2.1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.306, + 0.484, + 0.456 + ], + "angle": 0, + "content": "total, while minority groups like Middle Eastern, Latino, Black, and Indian each account for only \\(3 - 4\\%\\). This bias in DM, produce less accurate or fair results for underrepresented populations. We further investigate such bias situation across different versions of DMs. We randomly generate 1000 images with the prompt of \"A photo of a CEO\" using five versions of DMs: Stable Diffusion v1.3 [21], v1.4 [22], v.1.5 [23], v2.0 [24] and v2.1 [25]. Figure 3 demonstrates that gender bias exists across all versions of DMs." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.461, + 0.483, + 0.838 + ], + "angle": 0, + "content": "This bias arises from two primary sources. First, the training data, predominantly sourced from the web, contains inherent biases. Second, bias is partially inherited from the CLIP [19] model used in the generation process. In large text-to-image DMs, these biases are even more pronounced, often manifesting in gendered associations with specific professions. Several studies have attempted to mitigate bias in DMs through retraining [30, 36] or approaches that do not require retraining the DM [8, 17, 18]. In training-based methods, Shen et al. [30] propose curating a reference dataset and aligning the distribution of generated images with this dataset. Parihar et al. [18] avoid retraining the DM by leveraging demographic information embedded in the latent features of the denoising U-Net to guide generation. However, their method requires training an additional MLP-based Attribute Distribution Predictor using pseudo labels from existing attribute classifiers. This reliance on accurate attribute classifiers limits its debiasing performance. In contrast, Gandikota et al. [8] and Orgad et al. [17] employ closed-form editing to adjust concepts within DMs without retraining. While training-based methods depend on costly and quality-constrained annotated reference datasets, methods that do not require retraining the DM offer greater implementation efficiency but often exhibit lower debiasing effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To address these limitations, we propose FairGen, a plug-and-play method that achieves debiased image generation (Figure 1) by automatically learning attribute latent directions, mitigating the reliance on reference datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.454 + ], + "angle": 0, + "content": "FairGen is composed of two components: a set of attribute adapters and a distribution indicator. Each adapter is trained to learn an attribute-specific latent direction, optimized through noise composition in a self-discovering manner, i.e., our method automatically learns attribute latent directions without relying on a labeled reference dataset. Through noise composition, our method explores and optimizes attribute directions directly from the model's latent space, uncovering patterns without external supervision. At inference stage, the distribution indicator is then applied to select attribute-specific adapters, guiding the generative process towards the desired distribution. We comprehensively evaluate the effectiveness of our approach in debiasing gender, racial, and intersectional biases using occupational prompts. Experimental results demonstrate that our method not only achieves state-of-the-art performance in single and multiple attribute debiasing tasks but also preserves the generation quality of DM. Furthermore, we show that once FairGen is trained on one diffusion model, it can be seamlessly integrated into other models without retuning. Thanks to its strong transferability and plug-and-play functionality, our method offers a practical solution for both individual users and organizations, facilitating the responsible use of diffusion models in future applications." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.456, + 0.902, + 0.47 + ], + "angle": 0, + "content": "To summarize, our main contributions are as following:" + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.471, + 0.905, + 0.546 + ], + "angle": 0, + "content": "- We propose FairGen, a novel method for debiasing DMs by learning attribute latent directions in a self-discovering manner, eliminating the reliance on the reference dataset or classifier, and thus significantly reduce the cost." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.547, + 0.905, + 0.592 + ], + "angle": 0, + "content": "- FairGen is lightweight, plug-and-play and shows good transferability across different DMs, making it more convenient to deploy in the real world." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.592, + 0.905, + 0.637 + ], + "angle": 0, + "content": "- Extensive experiments show that our method achieves SOTA performance across diverse debiasing tasks while retaining the image generation quality." + }, + { + "type": "list", + "bbox": [ + 0.533, + 0.471, + 0.905, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.654, + 0.657, + 0.669 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Bias in Diffusion Models. Diffusion models for text-to-image generation (T2I) have been observed to produce biased and stereotypical images, even when given neutral prompts. Cho et al. [2] found that Stable Diffusion (SD) tends to generate images of males when prompted with occupations, with skin tones predominantly centered around a few shades from the Monk Skin Tone Scale [15]. Seshadri et al. [29] noted that SD reinforces gender-occupation biases presented in its training data. In addition to occupations, Bianchi et al. [1] discovered that simple prompts involving character traits and other descriptors also result in stereotypical images. Luccioni et al. [13] created a tool to compare generated image collections across different genders and ethnicities. Moreover, Wang et al. [34] introduced a text-to-image association test and found that SD tends to" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18412" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "associate females more with family roles and males more with career-related roles." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.125, + 0.484, + 0.412 + ], + "angle": 0, + "content": "Debiasing Diffusion Models by retraining. Before DMs, previous approaches mainly focus on debiasing GANs by assuming access to the labels of sensitive attributes and ensure no correlation exists between the decision and sensitive attributes [3, 16, 28, 31-33, 35, 36]. More recently, regarding DMs, Shen et al. [30] propose a distributional alignment loss to guide the characteristics of the generated images towards target distribution and use adjusted finetuning to directly optimize losses on the generated images. This method requires a reference training dataset to complete the retraining process, whereas our method does not need such reference dataset, which largely reduces annotation costs. Yu et al. [36] enhance fairness in text-to-image synthesis by incorporating reference images. and leverage visual exemplars to more effectively represent attributes that are difficult to describe in words, such as nuanced variations in skin tones. Li et al. [11] find a interpretable latent vector for a given concept, and achieve debiasing by manipulating the h-space inside DMs during inference stage." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.415, + 0.484, + 0.612 + ], + "angle": 0, + "content": "Debiasing Diffusion Models without retraining. Parihar et al. [18] propose Distribution Guidance (DG), which guides the generated images to follow the prescribed attribute distribution. Although DG does not require retraining DMs, it requires training an Attribute Distribution Predictor (ADP) that maps the latent features to the distribution of attributes. Since ADP is trained with pseudo labels generated from existing attribute classifiers, the performance of DG is largely constrained by the accuracy of attribute classifiers. Gandikota et al. [8] and Orgad et al. [17] use closed-form editing approach to edits concepts inside DM without training. Despite the efficiency, these methods are less effective than training-based approaches." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.484, + 0.901 + ], + "angle": 0, + "content": "**Controlling concepts in Diffusion Models.** Our work is inspired by recent work on controlling concepts [6, 7] in diffusion models. Gandikota et al. [6] propose to fine-tune the DM to erase an undesired concept while maintaining the model's behavior and capabilities, such as removing the NSFW (not safe for work) contents. ConceptSlider [7] also fine-tunes a LoRA adapter in a diffusion model to control the concept in the generated images. Unlike ESD that focuses on erasing concepts and ConceptSlider aims at precisely adjusting the extent of a concept (e.g., controlling the size of eyes), this work aims to debias the Diffusion Models, i.e., controlling the portion of generated images for different attributes such as male and female. Additionally, ConceptSlider learns an adapter per concept and use a scaling factor to control the extent of a concept (e.g., size of eye) while we learn an adapter per direction (e.g., female or male) in an attribute/concept (e.g., gender) and design an inference pipeline with an indicator to automatically select adapters based on a given distribution." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.091, + 0.64, + 0.108 + ], + "angle": 0, + "content": "3. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.11, + 0.907, + 0.277 + ], + "angle": 0, + "content": "Latent Diffusion Models (LDMs) perform the diffusion process within the latent space [26]. During training, noise is added to the encoded latent representation of the input image \\( x \\), resulting in a noisy latent code \\( z_{t} \\) at each time step \\( t \\). In the pretraining stage, an autoencoder framework is employed to map images into a lower-dimensional latent space via an encoder: \\( z = \\mathcal{E}(x) \\). The decoder then reconstructs images from these latent codes: \\( x \\approx \\mathcal{D}(\\mathcal{E}(x)) \\). This process ensures that the latent space retains the essential semantic information of the image. The training objective of the diffusion model in the latent space is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.284, + 0.907, + 0.311 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {L D M}} = \\mathbb {E} _ {z \\sim \\mathcal {E} (x), c, \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (z _ {t}, c, t) \\| _ {2} ^ {2} \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.317, + 0.906, + 0.466 + ], + "angle": 0, + "content": "where \\(\\epsilon\\) is Gaussian noise sampled from a normal distribution \\(\\mathcal{N}(0,1)\\), \\(\\epsilon_{\\theta}\\) is the denoising network, and \\(c\\) represents any conditioning embeddings (e.g., text or class labels). At the inference stage, a latent code \\(z_{T}\\) is sampled from Gaussian noise at the initial timestep \\(T\\). The denoising network \\(\\epsilon_{\\theta}\\) is then applied iteratively to remove the noise over several steps, generating a denoised latent representation \\(z_{0}\\). Finally, the pretrained decoder reconstructs the image from the denoised latent code: \\(\\hat{x}_0 \\approx \\mathcal{D}(z_0)\\), where \\(\\hat{x}_0\\) is the generated output image." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.468, + 0.906, + 0.618 + ], + "angle": 0, + "content": "Classifier-free Guidance aims to modulate image generation by steering the probability distribution towards data that is more probable according to an implicit classifier \\( p(c \\mid z_t) \\) [9]. It operates at the inference phase and the model is jointly trained on both conditional and unconditional denoising tasks. During inference, both the conditional and unconditional denoising scores are derived from the model. The final score \\( \\tilde{\\epsilon}_{\\theta}(z_t, c, t) \\) is then adjusted by weighting the conditioned score more heavily relative to the unconditioned score using a guidance scale \\( \\alpha > 1 \\)." + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.628, + 0.905, + 0.644 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\epsilon} _ {\\theta} \\left(z _ {t}, c, t\\right) = \\epsilon_ {\\theta} \\left(z _ {t}, t\\right) + \\alpha \\left(\\epsilon_ {\\theta} \\left(z _ {t}, c, t\\right) - \\epsilon_ {\\theta} \\left(z _ {t}, t\\right)\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.652, + 0.906, + 0.744 + ], + "angle": 0, + "content": "The inference process begins with sampling a latent variable \\( z_{T} \\sim \\mathcal{N}(0,1) \\), which is subsequently denoised using \\( \\tilde{\\epsilon}_{\\theta}(z_t,c,t) \\) to obtain \\( z_{t - 1} \\). The denoising is performed iteratively until obtaining \\( z_{0} \\). Finally, the decoder transforms the latent representation \\( z_{0} \\) back into image space: \\( x_0 \\gets \\mathcal{D}(z_0) \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.755, + 0.605, + 0.77 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Inspired by recent work that fine-tune diffusion models for erasing concepts [6] and controlling the level of concepts [7] (e.g., ages, size of eyes), we aim at reducing the bias in the DM by attaching and learning a set of lightweight adapters, each of which represents a category of an attribute (e.g., female of gender), guiding the DM towards an attribute latent direction. Unlike previous work that relies on additional reference datasets and has to finetune the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18413" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.084, + 0.874, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.322, + 0.908, + 0.365 + ], + "angle": 0, + "content": "Figure 4. Overview of FairGen's training pipeline. Attribute-specific adapters \\((M)\\) are attached to the cross-attention layers in the denoising UNet. Target group and attribution direction are fed into the DM for composing the noise predictions (Eq. 5), which is used as self-discovering attribute direction guidance to optimize the adapters." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.39, + 0.484, + 0.618 + ], + "angle": 0, + "content": "whole DM [30], we instead optimize the attached adapters via noise composition through a self-discovering process (detailed in Section 4.2). This significantly reduces the cost in computation and data. During the inference stage (Section 4.3), given a predefined target distribution (e.g., uniform), we introduce a distribution indicator implemented by a gating function to select one corresponding adapter which will be attached to the DM for generating image. In this way, the set of generated images will follow the predefined target distribution, and thus are not biased to some categories of an attribute (if the predefined target distribution is uniform). Our training and inference diagrams are illustrated in Figure 4 and Figure 5. In the following sections, we start elaborating our method in single attribute settings and then extend it to more general ones." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.633, + 0.484, + 0.665 + ], + "angle": 0, + "content": "4.1. Attaching Light-weight attribute-specific Adapters" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.672, + 0.483, + 0.868 + ], + "angle": 0, + "content": "To debias a DM for generating images of a given attribute that contains several categories (e.g., male and female are two categories of gender attribute), we first aim at equip the DM with skills of generating images for each category. To achieve this, we attach a light-weight adapter per category in each layer of the DM inspired by parameter-efficient fine-tuning (PEFT) instead of finetuning the whole model to achieve a good trade-off between performance and computational cost. In this work, we use the 1-dim adapter [14] and only add the adapter to each cross-attention layers of the denoising U-Net, as shown in Figure 4 as we find that attaching adapters to all layers does not help and will also increase the computational cost." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Specifically, for the \\(i\\)-th cross-attention layer parameterized by \\(\\mathbf{W}_i \\in \\mathbb{R}^{m \\times n}\\) in the denoising U-Net, we attach" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.391, + 0.907, + 0.512 + ], + "angle": 0, + "content": "an adapter to the layer to guide the attribute towards a certain category. The adapter consists of two 1-dim vectors: \\(\\pmb{p} \\in \\mathbb{R}^{m}\\) and \\(\\pmb{q} \\in \\mathbb{R}^{n}\\). The forward process of the \\(i\\)-th cross-attention layer would be updated from \\(\\pmb{y}_{i} = \\pmb{W}_{i}\\pmb{x}_{i}\\) to \\(\\pmb{y}_{i} = \\pmb{W}_{i}\\pmb{x}_{i} + (\\pmb{q}_{i}^{T}\\pmb{x}_{i})\\cdot \\pmb{p}_{i}\\). \\(\\pmb{x}_{i} \\in \\mathbb{R}^{n}\\) and \\(\\pmb{y}_{i} \\in \\mathbb{R}^{m}\\) represent the input and output of the layer, and superscript \\(T\\) indicates transposition. Thus, adapters in all \\((r)\\) cross-attention layers for an attribute category \\(d\\) are:" + }, + { + "type": "equation", + "bbox": [ + 0.663, + 0.522, + 0.907, + 0.542 + ], + "angle": 0, + "content": "\\[\nM _ {d} = \\boldsymbol {Q} ^ {T} \\boldsymbol {P}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.907, + 0.599 + ], + "angle": 0, + "content": "where \\(Q = [q_1, q_2, \\ldots, q_r], P = [p_1, p_2, \\ldots, p_r]\\). Each column vector in \\(Q\\) and \\(P\\), we pad 0 to their end if their dimensions are not the same." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.61, + 0.906, + 0.639 + ], + "angle": 0, + "content": "4.2. Optimizing Adapters via self-discovering process" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.647, + 0.906, + 0.737 + ], + "angle": 0, + "content": "One straightforward way to optimize the attached adapters is to collect a unbiased reference dataset as in prior work [30]. However, it is expensive to collect such dataset and the quality of the dataset would also limit the performance. To this end, we propose to train the adapters in a self-discovering manner." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.738, + 0.906, + 0.814 + ], + "angle": 0, + "content": "Given a target group \\( g_{t} \\) (for example, \"CEO\") and model \\( \\theta \\), we want to optimize the adapters such that the model attached with the adapters generate image \\( X \\) towards certain attribute category \\( d \\) (for example, \"male\" or \"female\") when conditioned on \\( g_{t} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.826, + 0.905, + 0.844 + ], + "angle": 0, + "content": "\\[\nP _ {\\theta^ {*}} (X \\mid g _ {t}) \\leftarrow P _ {\\theta} (X \\mid g _ {t}) \\left(P _ {\\theta} (d | X)\\right) ^ {\\eta}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\(P_{\\theta}(X|g_t)\\) represents the distribution generated by the original model when conditioned on \\(g_{t}\\), and \\(\\theta^{*}\\) represents the new model equipped with the adapters of \\(d\\). Note that" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18414" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.126, + 0.092, + 0.877, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.297, + 0.908, + 0.34 + ], + "angle": 0, + "content": "Figure 5. Overview of FairGen's inference pipeline. The distribution indicator is generated according the prescribed distribution. Then, it is multiplied by the set of attribute adapters matrices to select attribute matrix adapter. The selected adapter is integrated to the DM with no overhead, guiding the generation towards the prescribed distribution." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.366, + 0.483, + 0.396 + ], + "angle": 0, + "content": "\\(g_{t}\\) can be set to an empty string \" \" so that the model will be debiased for all possible groups." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.395, + 0.484, + 0.459 + ], + "angle": 0, + "content": "Applying the Bayes Formula, \\( P(d|X) = \\frac{P(X|d)P(d)}{P(X)} \\) to Eq. 4, taking logarithm on both sides, we are able to derive that the gradient of the log probability \\( \\nabla \\log P_{\\theta^*}(X|g_t) \\) would be proportional to:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.469, + 0.484, + 0.502 + ], + "angle": 0, + "content": "\\[\n\\nabla \\log P _ {\\theta} (X | g _ {t}) + \\eta (\\nabla \\log P _ {\\theta} (X | d) - \\nabla \\log P _ {\\theta} (X | g _ {t})) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.511, + 0.484, + 0.603 + ], + "angle": 0, + "content": "Based on Tweedie's formula [4] and the reparametrization trick of Classifier-free guidance [9], we introduce a time-varying noising process and represent each score (gradient of log probability) as a denoising prediction \\(\\epsilon(X_{t}, c_{t}, t)\\), which leads to our learning objective for adapters of category \\(d\\) of an attribute:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.613, + 0.512, + 0.645 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {\\theta^ {*}} \\left(X _ {t}, g _ {t}, t\\right) \\leftarrow \\epsilon_ {\\theta} \\left(X _ {t}, g _ {t}, t\\right) + \\eta \\left(\\epsilon_ {\\theta} \\left(X _ {t}, d, t\\right) - \\epsilon_ {\\theta} \\left(X _ {t}, g _ {t}, t\\right)\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.654, + 0.483, + 0.684 + ], + "angle": 0, + "content": "Therefore, the guidance loss for optimizing adapters of category \\(d\\) can be defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.694, + 0.454, + 0.723 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {G u i d a n c e}} = \\mathbb {L} _ {x _ {t}, t} \\left[ \\| \\epsilon_ {\\theta^ {*}} (X _ {t}, g _ {t}, t) - \\epsilon_ {L} \\| ^ {2} \\right],\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.723, + 0.454, + 0.74 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {L} = \\epsilon_ {\\theta} (X _ {t}, g _ {t}, t) + \\eta (\\epsilon_ {\\theta} (X _ {t}, d, t) - \\epsilon_ {\\theta} (X _ {t}, g _ {t}, t)),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.749, + 0.484, + 0.824 + ], + "angle": 0, + "content": "where the goal is to align the noise \\(\\epsilon_{\\theta^{*}}(X_{t},g_{t},t)\\) of the \\(\\theta^{*}\\) (DM with adapters of category \\(d\\) ) and the noise composition \\(\\epsilon_{L}\\) of group \\(g_{t}\\) and category \\(d\\) from the frozen DM \\(\\theta\\) And hence, our adapters optimization does not require any additional data." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.834, + 0.411, + 0.849 + ], + "angle": 0, + "content": "4.3. Inference with distribution indicator." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "After the optimization, we obtain a set of adapters \\(\\mathcal{M} = \\{M_1,M_2,\\dots ,M_t\\}\\) for \\(t\\) categories of a given attribute. For instance, \\(t = 2\\) for gender bias and \\(t = 4\\) for racial bias" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.366, + 0.907, + 0.426 + ], + "angle": 0, + "content": "in our case. As shown in Figure 5, at the inference stage, we introduce a distribution indicator \\( h \\). Given a prescribed distribution \\( f_{\\theta}^{a} \\) (e.g., uniform distribution), we define its Probabilistic Mass Function (PMF) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.442, + 0.907, + 0.483 + ], + "angle": 0, + "content": "\\[\nP (X = x) = f _ {\\theta} ^ {a} (x) = \\left\\{ \\begin{array}{l l} \\frac {1}{t} & \\text {f o r} x \\in \\{1, 2, \\dots , t \\}, \\\\ 0 & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.498, + 0.906, + 0.589 + ], + "angle": 0, + "content": "where \\(t\\) is the number of possible categories for each attribute, and \\(\\{1,2,\\dots ,t\\}\\) represents the set of all possible values for the random variable \\(X\\). The parameter \\(\\theta\\) controls the shape of the prescribed distribution, and in the case of the uniform distribution, \\(\\theta\\) implies that all outcomes in the set have equal probability, i.e., \\(\\frac{1}{t}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.591, + 0.906, + 0.636 + ], + "angle": 0, + "content": "Then, we randomly sample an index \\(k\\) from the prescribed distribution \\(f_{\\theta}^{a}\\). The distribution indicator \\(\\pmb{h} \\in \\mathbb{R}^t\\) is formulated to reflect the chosen index \\(k\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.638, + 0.658, + 0.907, + 0.699 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {h} _ {i} := \\left\\{ \\begin{array}{l l} 1 & \\text {i f} i = k, \\\\ 0 & \\text {i f} i \\neq k, \\end{array} \\right. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.714, + 0.906, + 0.821 + ], + "angle": 0, + "content": "where \\(i \\in \\{1,2,\\dots,t\\}\\), and \\(k\\) is the sampled index based on the distribution \\(f_{\\theta}^{a}\\). The indicator \\(h\\) is a one-hot vector, where the \\(k\\)-th element is 1, indicating the sampled outcome, and all other elements are 0. After obtaining the distribution indicator, it is multiplied by the set of trained attribute matrix adapters. Subsequently, the final weight change \\(\\Delta W\\) is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.656, + 0.838, + 0.907, + 0.854 + ], + "angle": 0, + "content": "\\[\n\\Delta \\boldsymbol {W} = \\boldsymbol {h} \\cdot \\mathcal {M}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "and the model is updated as \\(\\mathbf{W} \\gets \\mathbf{W} + \\alpha \\Delta \\mathbf{W}\\), where \\(\\alpha\\) is a scaling factor controlling the strength of the guidance." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18415" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.127, + 0.09, + 0.87, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.276, + 0.905, + 0.303 + ], + "angle": 0, + "content": "Figure 6. Images generated from the original Stable Diffusion model (left) and FairGen (right) for gender debiasing with the prompt \"A photo of a worker\". The gender distribution shifts from a male-to-female ratio of 13:2 in the original model to 8:7 with FairGen." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.329, + 0.482, + 0.361 + ], + "angle": 0, + "content": "4.4. Debiasing Multiple Attributes (intersectional debiasing)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.367, + 0.484, + 0.505 + ], + "angle": 0, + "content": "Our method can be inherently extended to debiasing multiple attributes in diffusion models (DM). Specifically, in the case of multiple attribute debiasing, the adapter for each attribute should not interfere with the others. Otherwise, the most recently trained adapter could degrade the performance of previously learned adapters. For each attribute to be debiased, we denote a set of adapter parameters as \\(\\{P_t, Q_t\\}\\). We have \\(P_t = [p_t^1, p_t^2, \\dots, p_t^r]\\), \\(Q_t = [q_t^1, q_t^2, \\dots, q_t^r]\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.505, + 0.484, + 0.566 + ], + "angle": 0, + "content": "To avoid interference between attribute adapters, we extend Eq. 7 by introducing an orthogonal regularization loss that regularizes the vector subspace spanned by each \\( P_{t} \\) and \\( Q_{t} \\) to be orthogonal to each other:" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.57, + 0.483, + 0.611 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {o r t h}} = \\sum_ {i = 1} ^ {t - 1} \\left(\\boldsymbol {P} _ {i} \\times \\boldsymbol {P} _ {t} + \\boldsymbol {Q} _ {i} \\times \\boldsymbol {Q} _ {t}\\right), \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.622, + 0.482, + 0.639 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {G u i d a n c e}} + \\gamma \\cdot \\mathcal {L} _ {\\text {o r t h}}. \\tag {12}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.668, + 0.216, + 0.685 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.694, + 0.484, + 0.755 + ], + "angle": 0, + "content": "In this section, we first describe the experimental details and evaluation metrics, then present a quantitative and qualitative analysis of our method. Please refer to Appendix for more results." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.762, + 0.291, + 0.777 + ], + "angle": 0, + "content": "5.1. Experimental Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Implementation details. We use Stable Diffusion v2.1 for all methods. We employ the prompt template \"a photo of the face of a {occupation}, a person\". At inference time, for each bias, we generate 100 images per occupation across 100 occupations, resulting in a total of 10,000 images. We set \\(\\eta = \\alpha = 0.3\\), and train for 1000 iterations with a learning rate of 1e-5. For gender bias, we use the CelebA [12] dataset to train a binary classifier with two" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.33, + 0.906, + 0.406 + ], + "angle": 0, + "content": "categories: {male, female}. For racial bias, we use the Fair-Face [10] dataset to train a classifier with the following four categories: WMELH = {White, Middle Eastern, Latino Hispanic}, Asian = {East Asian, Southeast Asian}, Black, and Indian. Please refer to the supplementary for more details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.427, + 0.907, + 0.578 + ], + "angle": 0, + "content": "Compared methods. In this work, we compare our method with recent state-of-the-art (SOTA) methods, including two retraining-based methods, Finetuning for Fairness (F4Fair) [30] and Inclusive Text-to-Image Generation (ITI-GEN) [37], an approach that does not require retraining DM, H-Distribution Guidance (H Guidance) [18], a closed-form editing approach, Unified Concept Editing (UCE) [8] and an h-space manipulation approach, Interpretable Diffusion [11]. Please refer to Appendix for more detailed introduction of these methods." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.593, + 0.697, + 0.607 + ], + "angle": 0, + "content": "5.2. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.615, + 0.788, + 0.629 + ], + "angle": 0, + "content": "We evaluate all methods in three metrics:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.635, + 0.906, + 0.756 + ], + "angle": 0, + "content": "Fairness Discrepancy (FD). Following prior work[18], we adopt the Fairness Discrepancy (FD) metric. For an attribute \\( a \\) and target distribution \\( p_{\\theta}^{a} \\), we use a high-accuracy classifier \\( \\mathcal{C}_a \\) to compute the fairness performance: \\( ||p_{\\theta}^{a} - \\mathbb{E}_{\\mathbf{x}\\sim p_{\\theta}(\\mathbf{x})}(\\mathbf{y})||_2 \\) where \\( \\mathbf{y} \\) is the softmax output of \\( \\mathcal{C}_a(\\mathbf{x}) \\). The target distribution \\( p_{\\theta}^{a} \\) can be any user-defined vector, typically uniform. A lower FD score indicates a closer match to the target distribution." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.76, + 0.906, + 0.837 + ], + "angle": 0, + "content": "\\(\\mathbf{CLIP}_{\\mathrm{sim}}\\). Besides fairness, the debiased model should maintain the ability to generate images that are semantically close to their text prompts. Therefore, following [30], we report the CLIP similarity score \\(\\mathbf{CLIP}_{\\mathrm{sim}}\\) between the generated image and its prompt." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.902 + ], + "angle": 0, + "content": "BRISQUE. The generated image quality is also important as the debiasing process should not influence the image generation ability. We use the BRISQUE metric for evaluate the quality of the generated images." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18416" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.125, + 0.09, + 0.48, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.267, + 0.266, + 0.343, + 0.279 + ], + "angle": 0, + "content": "(a) original" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.091, + 0.875, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.634, + 0.266, + 0.762, + 0.279 + ], + "angle": 0, + "content": "(b) debiased (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.28, + 0.908, + 0.321 + ], + "angle": 0, + "content": "Figure 7. Images generated from the original Stable Diffusion model (left) and FairGen (right) for racial debiasing with the prompt \"A photo of a sportsman\". The racial group distribution shifts from WMELH : Asian : Black : Indian = 7:2:5:1 in the original model to 4:4:4:3 with FairGen." + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.328, + 0.484, + 0.384 + ], + "angle": 0, + "content": "Table 1. Comparisons of our method to the SOTA methods in gender bias over two predefined distributions: \\( f_{\\theta}^{1} = (0.5, 0.5) \\) and \\( f_{\\theta}^{2} = (0.2, 0.8) \\), representing the probability of male and female respectively." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.385, + 0.451, + 0.523 + ], + "angle": 0, + "content": "
MethodFD ↓CLIPsim ↑BRISQUE ↑
f1θf2θf1θf2θf1θf2θ
Original SD0.4240.8470.380.3838.6538.69
F4Fair0.1650.3870.360.3538.2137.64
ITI-GEN0.0930.4720.320.3337.6237.96
H Guidance0.1180.3980.310.3238.5438.65
UCE0.2840.5360.360.2937.1236.54
InterDiff0.1770.3860.350.3438.3238.04
Ours0.0030.0050.380.3838.4638.72
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.328, + 0.907, + 0.384 + ], + "angle": 0, + "content": "Table 2. Comparisons of our method to the SOTA methods in racial bias over two distributions: \\( f_{\\theta}^{1} = (0.25, 0.25, 0.25, 0.25) \\) and \\( f_{\\theta}^{2} = (0.4, 0.3, 0.2, 0.1) \\), representing the probability of WMELH, Asian, Black, and Indian respectively." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.385, + 0.875, + 0.523 + ], + "angle": 0, + "content": "
MethodFD ↓CLIPsim ↑BRISQUE ↑
f1θf2θf1θf2θf1θf2θ
Original SD0.3840.4970.460.4138.9438.65
F4Fair0.2200.3050.430.4038.8038.45
ITI-GEN0.1310.5420.310.3136.6137.09
H Guidance0.1500.2850.410.3738.6538.60
UCE0.2900.4600.450.3937.9037.80
InterDiff0.1500.3910.400.3837.6437.82
Ours0.0950.1500.460.4138.9638.66
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.547, + 0.185, + 0.561 + ], + "angle": 0, + "content": "5.3. Results" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.566, + 0.483, + 0.822 + ], + "angle": 0, + "content": "Comparisons in gender debiasing. Table 1 demonstrates that our method outperforms others in mitigating gender bias over two predefined distributions: \\( f_{\\theta}^{1} = (0.5,0.5) \\), representing equal likelihood of male and female, and \\( f_{\\theta}^{2} = (0.2,0.8) \\), where male and female have a \\( 20\\% \\) and \\( 80\\% \\) probability, respectively. Original SD model exhibits high FD scores (0.424 at \\( f_{\\theta}^{1} \\) and 0.847 at \\( f_{\\theta}^{2} \\)), indicating significant bias towards one gender. While previous methods like F4Fair, ITI-GEN, H Guidance, and UCE reduce bias to some extent, our method achieves the most reduction, with FD scores of 0.003 at \\( f_{\\theta}^{1} \\) and 0.005 at \\( f_{\\theta}^{2} \\). Additionally, our method preserves semantic similarity (CLIPsim of 0.38) and image quality (BRISQUE scores of 38.46 and 38.72), matching or slightly surpassing the Original SD model. The qualitative results in Figure 6 further demonstrate that our approach effectively mitigates gender bias without compromising image quality or semantic coherence." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Comparisons in racial debiasing. Table 2 compares methods over two distributions: \\( f_{\\theta}^{1} = (0.25, 0.25, 0.25, 0.25) \\), representing equal probabilities for WMELH, Asian, Black, and Indian, and \\( f_{\\theta}^{2} = (0.4, 0.3, 0.2, 0.1) \\), with higher probabilities for" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.548, + 0.907, + 0.608 + ], + "angle": 0, + "content": "WMELH and Asian, and lower probabilities for Black and Indian. Original SD model shows significant racial bias, with FD scores of 0.384 for \\( f_{\\theta}^{1} \\) and 0.497 for \\( f_{\\theta}^{2} \\), indicating poor calibration to either distribution." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.609, + 0.907, + 0.807 + ], + "angle": 0, + "content": "Debiasing methods such as F4Fair, H Guidance, and UCE reduce this bias, with H Guidance achieving relatively lower FD scores. However, our method performs the best, reducing bias to 0.095 for \\( f_{\\theta}^{1} \\) and 0.150 for \\( f_{\\theta}^{2} \\). In terms of semantic similarity, the Original SD model sets a strong baseline (0.46 for \\( f_{\\theta}^{1} \\) and 0.41 for \\( f_{\\theta}^{2} \\)), and our method maintains this high alignment, ensuring that debiasing does not impair semantic accuracy. Regarding image quality, our approach slightly improves upon the Original SD model, achieving the highest scores (39.10 for \\( f_{\\theta}^{1} \\) and 38.90 for \\( f_{\\theta}^{2} \\)). Again, the qualitative results in Figure 7 further verify that our method outperforms others in reducing racial bias while preserving both semantic similarity and image quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Comparisons in intersectional debiasing We also consider a more complex challenge of intersectional debiasing, i.e., jointly debiasing both gender and racial biases. The target distribution \\( f_{\\theta} \\) is set to be uniform for both gender and racial bias. Table 4 shows that Original SD model exhibits significant bias with an FD score of 0.214. While F4Fair" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18417" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.303, + 0.089, + 0.694, + 0.103 + ], + "angle": 0, + "content": "Table 3. Transferability of proposed method across different DMs" + }, + { + "type": "table", + "bbox": [ + 0.258, + 0.105, + 0.742, + 0.198 + ], + "angle": 0, + "content": "
TrainTestGenderRacial
FD ↓CLIPsim ↑BRISQUE ↑FD ↓CLIPsim ↑BRISQUE ↑
v2.1v1.40.0060.3738.400.0970.4538.90
v2.1v1.50.0060.3738.420.0960.4538.92
v2.1v2.00.0080.3838.450.0950.4638.95
v2.1v2.10.0030.3838.460.0950.4638.96
" + }, + { + "type": "table_caption", + "bbox": [ + 0.131, + 0.22, + 0.443, + 0.234 + ], + "angle": 0, + "content": "Table 4. Evaluation of mitigating intersectional bias." + }, + { + "type": "table", + "bbox": [ + 0.136, + 0.235, + 0.44, + 0.362 + ], + "angle": 0, + "content": "
MethodFD ↓CLIPsim↑BRISQUE ↑
Original SD0.2140.3539.24
F4Fair0.1450.3638.90
ITI-GEN0.0980.3238.01
H Guidance0.1300.3338.75
UCE0.1800.3438.60
InterDiff0.1580.3538.67
Ours0.0470.3639.50
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.387, + 0.483, + 0.432 + ], + "angle": 0, + "content": "and H Guidance reduce this bias to 0.145 and 0.130, respectively, our method achieves a much lower FD score of 0.047, reflecting a substantial improvement in fairness." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.432, + 0.483, + 0.599 + ], + "angle": 0, + "content": "For semantic similarity, the Original SD scores 0.35, and F4Fair slightly improves it to 0.36. Our method matches this performance, maintaining semantic coherence while reducing bias. Regarding image quality, the Original SD has a BRISQUE score of 39.24, while our method improves it to 39.50, indicating enhanced perceptual quality. These results demonstrate that our method excels at jointly debiasing both gender and racial biases, significantly reducing bias without sacrificing semantic accuracy or image quality. This highlights its robustness and practicality in real-world settings where multiple biases are present." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.6, + 0.483, + 0.782 + ], + "angle": 0, + "content": "Transferability across different DMs. We further evaluate the transferability of our method by training it with Stable Diffusion v2.1 and testing it over other versions, and results are reported in Table 3. From the table we can see that the performance of in all metrics are close to the optimal results achieved when training and testing are performed using the same model version (v2.1). While there is a slight increase in FD when tested on v1.4, v1.5, and v2.0, the differences are minor, indicating that our method can effectively generalize across different model versions. This robustness highlights the method's capability to transfer learned features across varying model conditions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.792, + 0.245, + 0.807 + ], + "angle": 0, + "content": "5.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Attaching adapters in U-Net. The results in Table 5 illustrate the impact of attaching adapters to different layers of the U-Net in the DM. When adapters are attached to all layers, the fairness score (FD) is 0.165, with a CLIP similarity score of 0.33 and a BRISQUE score of 35.00. The best results are obtained when adapters are attached only to" + }, + { + "type": "table_caption", + "bbox": [ + 0.55, + 0.22, + 0.868, + 0.235 + ], + "angle": 0, + "content": "Table 5. Ablation of layers that are attached adapters." + }, + { + "type": "table", + "bbox": [ + 0.551, + 0.235, + 0.871, + 0.304 + ], + "angle": 0, + "content": "
LocationFD ↓CLIPsim ↑BRISQUE ↑
All layers0.1650.3335.00
Non-CA layers0.1580.3437.00
CA layers0.0470.3639.50
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.331, + 0.907, + 0.422 + ], + "angle": 0, + "content": "the CA layers. This configuration significantly reduces the FD score to 0.047, increases semantic similarity \\((\\mathrm{CLIP}_{\\mathrm{sim}} = 0.36)\\), and enhances image quality with a BRISQUE score of 39.50. These highlights that focusing the adapters on the cross-attention layers leads to the most substantial improvements in both fairness and image quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.426, + 0.909, + 0.579 + ], + "angle": 0, + "content": "Impact of orthogonal regularization (OR). Table 6 demonstrates the impact of orthogonal regularization (OR). Without OR, the FD score is 0.143, semantic similarity \\((\\mathrm{CLIP}_{\\mathrm{sim}})\\) drops to 0.29, and image quality (BRISQUE) is 37.92. When OR is applied, performance improves significantly across all metrics, with a much lower FD score of 0.047, higher semantic similarity of 0.36, and improved image quality (BRISQUE = 39.50). This highlights the effectiveness of orthogonal regularization in reducing bias and improving overall performance." + }, + { + "type": "table_caption", + "bbox": [ + 0.62, + 0.593, + 0.799, + 0.607 + ], + "angle": 0, + "content": "Table 6. Orthogonal Ablation" + }, + { + "type": "table", + "bbox": [ + 0.551, + 0.608, + 0.871, + 0.664 + ], + "angle": 0, + "content": "
LocationFD ↓CLIPsim ↑BRISQUE ↑
Ours w\\o OR0.1430.2937.92
Ours w\\OR0.0470.3639.50
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.697, + 0.634, + 0.712 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose FairGen, a plug-and-play method that learns attribute latent directions in a self-discovering manner. Our method can not only jointly debias multiple attributes in DMs, but also enables the generated images to follow a prescribed attribute distribution. It is lightweight and can be integrated with other DMs without retraining. Extensive experiments on debiasing gender, racial, and their intersectional biases show that our method outperforms previous SOTA by a large margin. We believe that our work marks a critical advancement in addressing harmful societal stereotypes within diffusion models, and it contributes to the ethical real-world applications of text-to-image DMs." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18418" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.484, + 0.184 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant No. 62306261, and the Shun Hing Institute of Advanced Engineering (SHIAE) Fund under Grant No. 8115074." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.198, + 0.188, + 0.214 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.223, + 0.484, + 0.32 + ], + "angle": 0, + "content": "[1] Federico Bianchi, Pratyusha Kalluri, Esin Durmus, Faisal Ladhak, Myra Cheng, Debora Nozza, Tatsunori Hashimoto, Dan Jurafsky, James Zou, and Aylin Caliskan. Easily accessible text-to-image generation amplifies demographic stereotypes at large scale. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, pages 1493-1504, Chicago, Illinois, USA, 2023. ACM. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.322, + 0.484, + 0.39 + ], + "angle": 0, + "content": "[2] Jaemin Cho, Abhay Zala, and Mohit Bansal. Dall-eval: Probing the reasoning skills and social biases of text-to-image generative transformers. In Proceedings of the International Conference on Computer Vision (ICCV), Paris, France, 2023. IEEE. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.392, + 0.483, + 0.447 + ], + "angle": 0, + "content": "[3] Kristy Choi, Aditya Grover, Trisha Singh, Rui Shu, and Stefano Ermon. Fair generative modeling via weak supervision. In International Conference on Machine Learning, pages 1887-1898. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.448, + 0.482, + 0.475 + ], + "angle": 0, + "content": "[4] Bradley Efron. Tweedie's formula and selection bias, 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.477, + 0.483, + 0.546 + ], + "angle": 0, + "content": "[5] Mureji Fatunde and Crystal Tse. Digital media company stability ai raises funds at $1 billion value. https://www.bloomberg.com/news/articles/2022-10-17/digital-media-firm-stability-ai-raises-funds-at-1-billion-value, 2022.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.548, + 0.483, + 0.588 + ], + "angle": 0, + "content": "[6] Rohit Gandikota, Joanna Materzynska, Jaden Fiitto-Kaufman, and David Bau. Erasing concepts from diffusion models, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.59, + 0.483, + 0.631 + ], + "angle": 0, + "content": "[7] Rohit Gandikota, Joanna Materzynska, Tingrui Zhou, Antonio Torralba, and David Bau. Concept sliders: Lora adaptors for precise control in diffusion models, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.633, + 0.483, + 0.687 + ], + "angle": 0, + "content": "[8] Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzyńska, and David Bau. Unified concept editing in diffusion models. arXiv preprint arXiv:2308.14761, 2023. 2, 3, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.689, + 0.483, + 0.716 + ], + "angle": 0, + "content": "[9] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.718, + 0.482, + 0.746 + ], + "angle": 0, + "content": "[10] Jungseock Joo. Fairface attribute model. https://github.com/jojoa/fairface, 2021. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.483, + 0.801 + ], + "angle": 0, + "content": "[11] Hang Li, Chengzhi Shen, Philip Torr, Volker Tresp, and Jindong Gu. Self-discovering interpretable diffusion latent directions for responsible text-to-image generation, 2024. 3, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[12] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 3730-3738, Santiago, Chile, 2015. IEEE. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Alexandra Sasha Luccioni, Christopher Akiki, Margaret Mitchell, and Yacine Jernite. Stable bias: Analyzing so" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.223, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.906, + 0.12 + ], + "angle": 0, + "content": "cietal representations in diffusion models. arXiv preprint arXiv:2303.11408, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.906, + 0.177 + ], + "angle": 0, + "content": "[14] Mengyao Lyu, Yuhong Yang, Haiwen Hong, Hui Chen, Xuan Jin, Yuan He, Hui Xue, Jungong Han, and Guiguang Ding. One-dimensional adapter to rule them all: Concepts, diffusion models and erasing applications, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.83, + 0.191 + ], + "angle": 0, + "content": "[15] Ellis Monk. The monk skin tone scale, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.193, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[16] Junhyun Nam, Sangwoo Mo, Jaeho Lee, and Jinwoo Shin. Breaking the spurious causality of conditional generation via fairness intervention with corrective sampling. Transactions on Machine Learning Research, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[17] Hadas Orgad, Bahjat Kawar, and Yonatan Belinkov. Editing implicit assumptions in text-to-image diffusion models. 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.292, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[18] Rishubh Parihar, Abhijnya Bhat, Abhipsa Basu, Saswat Mallick, Jogendra Nath Kundu, and R. Venkatesh Babu. Balancing act: Distribution-guided debiasing in diffusion models, 2024. 2, 3, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[19] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.905, + 0.487 + ], + "angle": 0, + "content": "[20] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.489, + 0.902, + 0.502 + ], + "angle": 0, + "content": "[21] Robin Rombach. Stable diffusion v1-3 model card, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.902, + 0.517 + ], + "angle": 0, + "content": "[22] Robin Rombach. Stable diffusion v1-4 model card, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.902, + 0.532 + ], + "angle": 0, + "content": "[23] Robin Rombach. Stable diffusion v1-5 model card, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.534, + 0.902, + 0.546 + ], + "angle": 0, + "content": "[24] Robin Rombach. Stable diffusion v2-0 model card, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.549, + 0.902, + 0.562 + ], + "angle": 0, + "content": "[25] Robin Rombach. Stable diffusion v2-1 model card, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.564, + 0.907, + 0.632 + ], + "angle": 0, + "content": "[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.634, + 0.906, + 0.716 + ], + "angle": 0, + "content": "[27] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.718, + 0.906, + 0.786 + ], + "angle": 0, + "content": "[28] Prasanna Sattigeri, Samuel C Hoffman, Vijil Chenthamarakshan, and Kush R Varshney. Fairness gan: Generating datasets with fairness properties using a generative adversarial network. IBM Journal of Research and Development, 63 (4/5):3-1, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[29] Preethi Seshadri, Sameer Singh, and Yanai Elazar. The bias amplification paradox in text-to-image generation. arXiv preprint arXiv:2308.00755, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[30] Xudong Shen, Chao Du, Tianyu Pang, Min Lin, Yongkang Wong, and Mohan Kankanhalli. Finetuning text-to-image diffusion models for fairness, 2024. 2, 3, 4, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[31] Christopher TH Teo, Milad Abdollahzadeh, and Ngai-Man Cheung. Fair generative models via transfer learning. In Pro" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18419" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "ceedings of the AAAI Conference on Artificial Intelligence, pages 2429-2437, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.163 + ], + "angle": 0, + "content": "[32] Soobin Um and Changho Suh. A fair generative model using lecam divergence. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10034-10042, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[33] Boris van Breugel, Trent Kyono, Jeroen Berrevoets, and Michaela van der Schaar. Decaf: Generating fair synthetic data using causally-aware generative networks. Advances in Neural Information Processing Systems, 34:22221-22233, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.29 + ], + "angle": 0, + "content": "[34] Zihao Wang, Lin Gui, Jeffrey Negrea, and Victor Veitch. Concept algebra for score-based conditional model. In ICML 2023 Workshop on Structured Probabilistic Inference & Generative Modeling, Honolulu, HI, 2023. PMLR. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.482, + 0.347 + ], + "angle": 0, + "content": "[35] Depeng Xu, Shuhan Yuan, Lu Zhang, and Xintao Wu. Fairgan: Fairness-aware generative adversarial networks. In 2018 IEEE International Conference on Big Data (Big Data), pages 570-575. IEEE, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.431 + ], + "angle": 0, + "content": "[36] Ning Yu, Ke Li, Peng Zhou, Jitendra Malik, Larry Davis, and Mario Fritz. Inclusive gan: Improving data and minority coverage in generative models. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 377-393. Springer, 2020. 2, 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.433, + 0.482, + 0.475 + ], + "angle": 0, + "content": "[37] Cheng Zhang, Xuanbai Chen, Siqi Chai, Chen Henry Wu, Dmitry Lagun, Thabo Beeler, and Fernando De la Torre. Itigen: Inclusive text-to-image generation, 2023. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18420" + } + ] +] \ No newline at end of file diff --git a/2025/FairGen_ Enhancing Fairness in Text-to-Image Diffusion Models via Self-Discovering Latent Directions/207183e6-db8a-436f-8e4f-faec1fb9d3a5_origin.pdf b/2025/FairGen_ Enhancing Fairness in Text-to-Image Diffusion Models via Self-Discovering Latent Directions/207183e6-db8a-436f-8e4f-faec1fb9d3a5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e1ab49b9a2174b5a0ef782e498cc9801878f0550 --- /dev/null +++ b/2025/FairGen_ Enhancing Fairness in Text-to-Image Diffusion Models via Self-Discovering Latent Directions/207183e6-db8a-436f-8e4f-faec1fb9d3a5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0d346cc050718c23145aca94a80fbbaa8f9c7af4a9fe00bef6612b671551643d +size 2002397 diff --git a/2025/FairGen_ Enhancing Fairness in Text-to-Image Diffusion Models via Self-Discovering Latent Directions/full.md b/2025/FairGen_ Enhancing Fairness in Text-to-Image Diffusion Models via Self-Discovering Latent Directions/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8ccb73930c0858831a4b738ae946e3a1da44141d --- /dev/null +++ b/2025/FairGen_ Enhancing Fairness in Text-to-Image Diffusion Models via Self-Discovering Latent Directions/full.md @@ -0,0 +1,308 @@ +# FairGen: Enhancing Fairness in Text-to-Image Diffusion Models via Self-Discovering Latent Directions + +Yilei Jiang $^{1}$ Wei-Hong Li $^{1,2*}$ Yiyuan Zhang $^{1}$ Minghong Cai $^{1}$ Xiangyu Yue $^{1,3*}$ $^{1}$ MMLab, CUHK $^{2}$ University of Bristol $^{3}$ SHIAE, CUHK + +# Abstract + +While Diffusion Models (DM) exhibit remarkable performance across various image generative tasks, they nonetheless reflect the inherent bias presented in the training set. As DMs are now widely used in real-world applications, these biases could perpetuate a distorted worldview and hinder opportunities for minority groups. Existing methods on debiasing DMs usually requires model retraining with a human-crafted reference dataset or additional classifiers, which suffer from two major limitations: (1) collecting reference datasets causes expensive annotation cost; (2) the debiasing performance is heavily constrained by the quality of the reference dataset or the additional classifier. To address the above limitations, we propose FairGen, a plug-and-play method that learns attribute latent directions in a self-discovering manner, thus eliminating the reliance on such reference dataset. Specifically, FairGen consists of two parts: a set of attribute adapters and a distribution indicator. Each adapter in the set aims to learn an attribute latent direction, and is optimized via noise composition through a self-discovering process. Then, the distribution indicator is multiplied by the set of adapters to guide the generation process towards the prescribed distribution. Our method enables debiasing multiple attributes in DMs simultaneously, while remaining lightweight and easily integrable with other DMs, eliminating the need for retraining. Extensive experiments on debiasing gender, racial, and their intersectional biases show that our method outperforms previous SOTA by a large margin. + +# 1. Introduction + +State-of-the-art Text-to-Image Diffusion Models (DMs) such as Stable Diffusion [26], DALL-E 3 [20] and Imagegen [27] have demonstrated remarkable performance in generating high-quality images. With the rapid development of DMs, an increasing number of individuals and corporations are choosing to utilize them to serve their own + +![](images/c666aa8c76200144f15fcef6ff46f095b4eca8a2e185d8b6cb4cec26163707dd.jpg) + +![](images/2fe86488056ac03a8517bbb81ec4901cd2a0c8f1e4840b7f47a039e1db4f4f57.jpg) +Figure 1. Illustration of gender bias associated with different occupations in Stable Diffusion v2.1 and the debiased results for gender by our method. The leadership roles such as CEOs are biased towards male figures, whereas supportive roles such as nurses are biased towards female figures. Our method can generate debiased results regarding the gender bias. + +purposes. For instance, Stable Diffusion v1.5 has been downloaded over 8 million times on Huggingface, and Midjourney is used by over a million users [5]. However, existing DMs have been found to generate biased content across various demographic factors, such as gender and race [13], which could have harmful effects on society when these models are implemented in real-world applications. + +In Figure 1, we randomly generate several images of two occupations using Stable Diffusion v2.1. Given the prompt of "A photo of a CEO", the generated images predominantly depict male figures, reinforcing the stereotype that leadership roles, such as CEOs and doctors, are male-dominated. On the contrary, when the prompt is "A photo of a nurse", the majority of generated images depict female figures, reflecting the bias that supportive roles are traditionally associated with women. Regarding racial bias, we randomly generate 1000 images using Stable Diffusion v2.1 with the prompt "A photo of a worker". The statistic of 1000 images depicted in Figure 2 shows a strong bias in racial representation, with White individuals making up $71\%$ of the + +![](images/98aa4a433620d2f35c65c4f44639e1828b7f119e834e55e194586642246eee2f.jpg) +Figure 2. Racial bias in randomly generated 1000 images with the prompt "A photo of a worker" using Stable Diffusion v2.1. + +![](images/323da653ce762f0e6445cdba1deada4e466da9adaad83e93f0ce80c356568657.jpg) +Figure 3. Gender bias in randomly generated 1000 images with the prompt "A photo of a CEO" using Stable Diffusion v1.3, v1.4, v1.5, v2.0, and v2.1. + +total, while minority groups like Middle Eastern, Latino, Black, and Indian each account for only $3 - 4\%$ . This bias in DM, produce less accurate or fair results for underrepresented populations. We further investigate such bias situation across different versions of DMs. We randomly generate 1000 images with the prompt of "A photo of a CEO" using five versions of DMs: Stable Diffusion v1.3 [21], v1.4 [22], v.1.5 [23], v2.0 [24] and v2.1 [25]. Figure 3 demonstrates that gender bias exists across all versions of DMs. + +This bias arises from two primary sources. First, the training data, predominantly sourced from the web, contains inherent biases. Second, bias is partially inherited from the CLIP [19] model used in the generation process. In large text-to-image DMs, these biases are even more pronounced, often manifesting in gendered associations with specific professions. Several studies have attempted to mitigate bias in DMs through retraining [30, 36] or approaches that do not require retraining the DM [8, 17, 18]. In training-based methods, Shen et al. [30] propose curating a reference dataset and aligning the distribution of generated images with this dataset. Parihar et al. [18] avoid retraining the DM by leveraging demographic information embedded in the latent features of the denoising U-Net to guide generation. However, their method requires training an additional MLP-based Attribute Distribution Predictor using pseudo labels from existing attribute classifiers. This reliance on accurate attribute classifiers limits its debiasing performance. In contrast, Gandikota et al. [8] and Orgad et al. [17] employ closed-form editing to adjust concepts within DMs without retraining. While training-based methods depend on costly and quality-constrained annotated reference datasets, methods that do not require retraining the DM offer greater implementation efficiency but often exhibit lower debiasing effectiveness. + +To address these limitations, we propose FairGen, a plug-and-play method that achieves debiased image generation (Figure 1) by automatically learning attribute latent directions, mitigating the reliance on reference datasets. + +FairGen is composed of two components: a set of attribute adapters and a distribution indicator. Each adapter is trained to learn an attribute-specific latent direction, optimized through noise composition in a self-discovering manner, i.e., our method automatically learns attribute latent directions without relying on a labeled reference dataset. Through noise composition, our method explores and optimizes attribute directions directly from the model's latent space, uncovering patterns without external supervision. At inference stage, the distribution indicator is then applied to select attribute-specific adapters, guiding the generative process towards the desired distribution. We comprehensively evaluate the effectiveness of our approach in debiasing gender, racial, and intersectional biases using occupational prompts. Experimental results demonstrate that our method not only achieves state-of-the-art performance in single and multiple attribute debiasing tasks but also preserves the generation quality of DM. Furthermore, we show that once FairGen is trained on one diffusion model, it can be seamlessly integrated into other models without retuning. Thanks to its strong transferability and plug-and-play functionality, our method offers a practical solution for both individual users and organizations, facilitating the responsible use of diffusion models in future applications. + +To summarize, our main contributions are as following: + +- We propose FairGen, a novel method for debiasing DMs by learning attribute latent directions in a self-discovering manner, eliminating the reliance on the reference dataset or classifier, and thus significantly reduce the cost. +- FairGen is lightweight, plug-and-play and shows good transferability across different DMs, making it more convenient to deploy in the real world. +- Extensive experiments show that our method achieves SOTA performance across diverse debiasing tasks while retaining the image generation quality. + +# 2. Related Work + +Bias in Diffusion Models. Diffusion models for text-to-image generation (T2I) have been observed to produce biased and stereotypical images, even when given neutral prompts. Cho et al. [2] found that Stable Diffusion (SD) tends to generate images of males when prompted with occupations, with skin tones predominantly centered around a few shades from the Monk Skin Tone Scale [15]. Seshadri et al. [29] noted that SD reinforces gender-occupation biases presented in its training data. In addition to occupations, Bianchi et al. [1] discovered that simple prompts involving character traits and other descriptors also result in stereotypical images. Luccioni et al. [13] created a tool to compare generated image collections across different genders and ethnicities. Moreover, Wang et al. [34] introduced a text-to-image association test and found that SD tends to + +associate females more with family roles and males more with career-related roles. + +Debiasing Diffusion Models by retraining. Before DMs, previous approaches mainly focus on debiasing GANs by assuming access to the labels of sensitive attributes and ensure no correlation exists between the decision and sensitive attributes [3, 16, 28, 31-33, 35, 36]. More recently, regarding DMs, Shen et al. [30] propose a distributional alignment loss to guide the characteristics of the generated images towards target distribution and use adjusted finetuning to directly optimize losses on the generated images. This method requires a reference training dataset to complete the retraining process, whereas our method does not need such reference dataset, which largely reduces annotation costs. Yu et al. [36] enhance fairness in text-to-image synthesis by incorporating reference images. and leverage visual exemplars to more effectively represent attributes that are difficult to describe in words, such as nuanced variations in skin tones. Li et al. [11] find a interpretable latent vector for a given concept, and achieve debiasing by manipulating the h-space inside DMs during inference stage. + +Debiasing Diffusion Models without retraining. Parihar et al. [18] propose Distribution Guidance (DG), which guides the generated images to follow the prescribed attribute distribution. Although DG does not require retraining DMs, it requires training an Attribute Distribution Predictor (ADP) that maps the latent features to the distribution of attributes. Since ADP is trained with pseudo labels generated from existing attribute classifiers, the performance of DG is largely constrained by the accuracy of attribute classifiers. Gandikota et al. [8] and Orgad et al. [17] use closed-form editing approach to edits concepts inside DM without training. Despite the efficiency, these methods are less effective than training-based approaches. + +**Controlling concepts in Diffusion Models.** Our work is inspired by recent work on controlling concepts [6, 7] in diffusion models. Gandikota et al. [6] propose to fine-tune the DM to erase an undesired concept while maintaining the model's behavior and capabilities, such as removing the NSFW (not safe for work) contents. ConceptSlider [7] also fine-tunes a LoRA adapter in a diffusion model to control the concept in the generated images. Unlike ESD that focuses on erasing concepts and ConceptSlider aims at precisely adjusting the extent of a concept (e.g., controlling the size of eyes), this work aims to debias the Diffusion Models, i.e., controlling the portion of generated images for different attributes such as male and female. Additionally, ConceptSlider learns an adapter per concept and use a scaling factor to control the extent of a concept (e.g., size of eye) while we learn an adapter per direction (e.g., female or male) in an attribute/concept (e.g., gender) and design an inference pipeline with an indicator to automatically select adapters based on a given distribution. + +# 3. Preliminary + +Latent Diffusion Models (LDMs) perform the diffusion process within the latent space [26]. During training, noise is added to the encoded latent representation of the input image $x$ , resulting in a noisy latent code $z_{t}$ at each time step $t$ . In the pretraining stage, an autoencoder framework is employed to map images into a lower-dimensional latent space via an encoder: $z = \mathcal{E}(x)$ . The decoder then reconstructs images from these latent codes: $x \approx \mathcal{D}(\mathcal{E}(x))$ . This process ensures that the latent space retains the essential semantic information of the image. The training objective of the diffusion model in the latent space is given by: + +$$ +\mathcal {L} _ {\mathrm {L D M}} = \mathbb {E} _ {z \sim \mathcal {E} (x), c, \epsilon \sim \mathcal {N} (0, 1), t} \left[ \| \epsilon - \epsilon_ {\theta} (z _ {t}, c, t) \| _ {2} ^ {2} \right], \tag {1} +$$ + +where $\epsilon$ is Gaussian noise sampled from a normal distribution $\mathcal{N}(0,1)$ , $\epsilon_{\theta}$ is the denoising network, and $c$ represents any conditioning embeddings (e.g., text or class labels). At the inference stage, a latent code $z_{T}$ is sampled from Gaussian noise at the initial timestep $T$ . The denoising network $\epsilon_{\theta}$ is then applied iteratively to remove the noise over several steps, generating a denoised latent representation $z_{0}$ . Finally, the pretrained decoder reconstructs the image from the denoised latent code: $\hat{x}_0 \approx \mathcal{D}(z_0)$ , where $\hat{x}_0$ is the generated output image. + +Classifier-free Guidance aims to modulate image generation by steering the probability distribution towards data that is more probable according to an implicit classifier $p(c \mid z_t)$ [9]. It operates at the inference phase and the model is jointly trained on both conditional and unconditional denoising tasks. During inference, both the conditional and unconditional denoising scores are derived from the model. The final score $\tilde{\epsilon}_{\theta}(z_t, c, t)$ is then adjusted by weighting the conditioned score more heavily relative to the unconditioned score using a guidance scale $\alpha > 1$ . + +$$ +\tilde {\epsilon} _ {\theta} \left(z _ {t}, c, t\right) = \epsilon_ {\theta} \left(z _ {t}, t\right) + \alpha \left(\epsilon_ {\theta} \left(z _ {t}, c, t\right) - \epsilon_ {\theta} \left(z _ {t}, t\right)\right) \tag {2} +$$ + +The inference process begins with sampling a latent variable $z_{T} \sim \mathcal{N}(0,1)$ , which is subsequently denoised using $\tilde{\epsilon}_{\theta}(z_t,c,t)$ to obtain $z_{t - 1}$ . The denoising is performed iteratively until obtaining $z_{0}$ . Finally, the decoder transforms the latent representation $z_{0}$ back into image space: $x_0 \gets \mathcal{D}(z_0)$ . + +# 4. Method + +Inspired by recent work that fine-tune diffusion models for erasing concepts [6] and controlling the level of concepts [7] (e.g., ages, size of eyes), we aim at reducing the bias in the DM by attaching and learning a set of lightweight adapters, each of which represents a category of an attribute (e.g., female of gender), guiding the DM towards an attribute latent direction. Unlike previous work that relies on additional reference datasets and has to finetune the + +![](images/e53d837e7dee188ed02e842596334fd404f0d6da60cc337a55385adc2773e2a0.jpg) +Figure 4. Overview of FairGen's training pipeline. Attribute-specific adapters $(M)$ are attached to the cross-attention layers in the denoising UNet. Target group and attribution direction are fed into the DM for composing the noise predictions (Eq. 5), which is used as self-discovering attribute direction guidance to optimize the adapters. + +whole DM [30], we instead optimize the attached adapters via noise composition through a self-discovering process (detailed in Section 4.2). This significantly reduces the cost in computation and data. During the inference stage (Section 4.3), given a predefined target distribution (e.g., uniform), we introduce a distribution indicator implemented by a gating function to select one corresponding adapter which will be attached to the DM for generating image. In this way, the set of generated images will follow the predefined target distribution, and thus are not biased to some categories of an attribute (if the predefined target distribution is uniform). Our training and inference diagrams are illustrated in Figure 4 and Figure 5. In the following sections, we start elaborating our method in single attribute settings and then extend it to more general ones. + +# 4.1. Attaching Light-weight attribute-specific Adapters + +To debias a DM for generating images of a given attribute that contains several categories (e.g., male and female are two categories of gender attribute), we first aim at equip the DM with skills of generating images for each category. To achieve this, we attach a light-weight adapter per category in each layer of the DM inspired by parameter-efficient fine-tuning (PEFT) instead of finetuning the whole model to achieve a good trade-off between performance and computational cost. In this work, we use the 1-dim adapter [14] and only add the adapter to each cross-attention layers of the denoising U-Net, as shown in Figure 4 as we find that attaching adapters to all layers does not help and will also increase the computational cost. + +Specifically, for the $i$ -th cross-attention layer parameterized by $\mathbf{W}_i \in \mathbb{R}^{m \times n}$ in the denoising U-Net, we attach + +an adapter to the layer to guide the attribute towards a certain category. The adapter consists of two 1-dim vectors: $\pmb{p} \in \mathbb{R}^{m}$ and $\pmb{q} \in \mathbb{R}^{n}$ . The forward process of the $i$ -th cross-attention layer would be updated from $\pmb{y}_{i} = \pmb{W}_{i}\pmb{x}_{i}$ to $\pmb{y}_{i} = \pmb{W}_{i}\pmb{x}_{i} + (\pmb{q}_{i}^{T}\pmb{x}_{i})\cdot \pmb{p}_{i}$ . $\pmb{x}_{i} \in \mathbb{R}^{n}$ and $\pmb{y}_{i} \in \mathbb{R}^{m}$ represent the input and output of the layer, and superscript $T$ indicates transposition. Thus, adapters in all $(r)$ cross-attention layers for an attribute category $d$ are: + +$$ +M _ {d} = \boldsymbol {Q} ^ {T} \boldsymbol {P}, \tag {3} +$$ + +where $Q = [q_1, q_2, \ldots, q_r], P = [p_1, p_2, \ldots, p_r]$ . Each column vector in $Q$ and $P$ , we pad 0 to their end if their dimensions are not the same. + +# 4.2. Optimizing Adapters via self-discovering process + +One straightforward way to optimize the attached adapters is to collect a unbiased reference dataset as in prior work [30]. However, it is expensive to collect such dataset and the quality of the dataset would also limit the performance. To this end, we propose to train the adapters in a self-discovering manner. + +Given a target group $g_{t}$ (for example, "CEO") and model $\theta$ , we want to optimize the adapters such that the model attached with the adapters generate image $X$ towards certain attribute category $d$ (for example, "male" or "female") when conditioned on $g_{t}$ : + +$$ +P _ {\theta^ {*}} (X \mid g _ {t}) \leftarrow P _ {\theta} (X \mid g _ {t}) \left(P _ {\theta} (d | X)\right) ^ {\eta}, \tag {4} +$$ + +where $P_{\theta}(X|g_t)$ represents the distribution generated by the original model when conditioned on $g_{t}$ , and $\theta^{*}$ represents the new model equipped with the adapters of $d$ . Note that + +![](images/906182312817e1b41230e975d41b8c65ff6a3fa052dc1d8e8d4c4807b699f6ac.jpg) +Figure 5. Overview of FairGen's inference pipeline. The distribution indicator is generated according the prescribed distribution. Then, it is multiplied by the set of attribute adapters matrices to select attribute matrix adapter. The selected adapter is integrated to the DM with no overhead, guiding the generation towards the prescribed distribution. + +$g_{t}$ can be set to an empty string " " so that the model will be debiased for all possible groups. + +Applying the Bayes Formula, $P(d|X) = \frac{P(X|d)P(d)}{P(X)}$ to Eq. 4, taking logarithm on both sides, we are able to derive that the gradient of the log probability $\nabla \log P_{\theta^*}(X|g_t)$ would be proportional to: + +$$ +\nabla \log P _ {\theta} (X | g _ {t}) + \eta (\nabla \log P _ {\theta} (X | d) - \nabla \log P _ {\theta} (X | g _ {t})) \tag {5} +$$ + +Based on Tweedie's formula [4] and the reparametrization trick of Classifier-free guidance [9], we introduce a time-varying noising process and represent each score (gradient of log probability) as a denoising prediction $\epsilon(X_{t}, c_{t}, t)$ , which leads to our learning objective for adapters of category $d$ of an attribute: + +$$ +\epsilon_ {\theta^ {*}} \left(X _ {t}, g _ {t}, t\right) \leftarrow \epsilon_ {\theta} \left(X _ {t}, g _ {t}, t\right) + \eta \left(\epsilon_ {\theta} \left(X _ {t}, d, t\right) - \epsilon_ {\theta} \left(X _ {t}, g _ {t}, t\right)\right). \tag {6} +$$ + +Therefore, the guidance loss for optimizing adapters of category $d$ can be defined as follows: + +$$ +\mathcal {L} _ {\mathrm {G u i d a n c e}} = \mathbb {L} _ {x _ {t}, t} \left[ \| \epsilon_ {\theta^ {*}} (X _ {t}, g _ {t}, t) - \epsilon_ {L} \| ^ {2} \right], +$$ + +$$ +\epsilon_ {L} = \epsilon_ {\theta} (X _ {t}, g _ {t}, t) + \eta (\epsilon_ {\theta} (X _ {t}, d, t) - \epsilon_ {\theta} (X _ {t}, g _ {t}, t)), +$$ + +where the goal is to align the noise $\epsilon_{\theta^{*}}(X_{t},g_{t},t)$ of the $\theta^{*}$ (DM with adapters of category $d$ ) and the noise composition $\epsilon_{L}$ of group $g_{t}$ and category $d$ from the frozen DM $\theta$ And hence, our adapters optimization does not require any additional data. + +# 4.3. Inference with distribution indicator. + +After the optimization, we obtain a set of adapters $\mathcal{M} = \{M_1,M_2,\dots ,M_t\}$ for $t$ categories of a given attribute. For instance, $t = 2$ for gender bias and $t = 4$ for racial bias + +in our case. As shown in Figure 5, at the inference stage, we introduce a distribution indicator $h$ . Given a prescribed distribution $f_{\theta}^{a}$ (e.g., uniform distribution), we define its Probabilistic Mass Function (PMF) as follows: + +$$ +P (X = x) = f _ {\theta} ^ {a} (x) = \left\{ \begin{array}{l l} \frac {1}{t} & \text {f o r} x \in \{1, 2, \dots , t \}, \\ 0 & \text {o t h e r w i s e}, \end{array} \right. \tag {8} +$$ + +where $t$ is the number of possible categories for each attribute, and $\{1,2,\dots ,t\}$ represents the set of all possible values for the random variable $X$ . The parameter $\theta$ controls the shape of the prescribed distribution, and in the case of the uniform distribution, $\theta$ implies that all outcomes in the set have equal probability, i.e., $\frac{1}{t}$ . + +Then, we randomly sample an index $k$ from the prescribed distribution $f_{\theta}^{a}$ . The distribution indicator $\pmb{h} \in \mathbb{R}^t$ is formulated to reflect the chosen index $k$ as follows: + +$$ +\boldsymbol {h} _ {i} := \left\{ \begin{array}{l l} 1 & \text {i f} i = k, \\ 0 & \text {i f} i \neq k, \end{array} \right. \tag {9} +$$ + +where $i \in \{1,2,\dots,t\}$ , and $k$ is the sampled index based on the distribution $f_{\theta}^{a}$ . The indicator $h$ is a one-hot vector, where the $k$ -th element is 1, indicating the sampled outcome, and all other elements are 0. After obtaining the distribution indicator, it is multiplied by the set of trained attribute matrix adapters. Subsequently, the final weight change $\Delta W$ is given by: + +$$ +\Delta \boldsymbol {W} = \boldsymbol {h} \cdot \mathcal {M}, \tag {10} +$$ + +and the model is updated as $\mathbf{W} \gets \mathbf{W} + \alpha \Delta \mathbf{W}$ , where $\alpha$ is a scaling factor controlling the strength of the guidance. + +![](images/b978c236a569da3c60a08d20df040305b0ac45295bf23a9ad236928d06ae5e4a.jpg) +Figure 6. Images generated from the original Stable Diffusion model (left) and FairGen (right) for gender debiasing with the prompt "A photo of a worker". The gender distribution shifts from a male-to-female ratio of 13:2 in the original model to 8:7 with FairGen. + +# 4.4. Debiasing Multiple Attributes (intersectional debiasing) + +Our method can be inherently extended to debiasing multiple attributes in diffusion models (DM). Specifically, in the case of multiple attribute debiasing, the adapter for each attribute should not interfere with the others. Otherwise, the most recently trained adapter could degrade the performance of previously learned adapters. For each attribute to be debiased, we denote a set of adapter parameters as $\{P_t, Q_t\}$ . We have $P_t = [p_t^1, p_t^2, \dots, p_t^r]$ , $Q_t = [q_t^1, q_t^2, \dots, q_t^r]$ . + +To avoid interference between attribute adapters, we extend Eq. 7 by introducing an orthogonal regularization loss that regularizes the vector subspace spanned by each $P_{t}$ and $Q_{t}$ to be orthogonal to each other: + +$$ +\mathcal {L} _ {\text {o r t h}} = \sum_ {i = 1} ^ {t - 1} \left(\boldsymbol {P} _ {i} \times \boldsymbol {P} _ {t} + \boldsymbol {Q} _ {i} \times \boldsymbol {Q} _ {t}\right), \tag {11} +$$ + +$$ +\mathcal {L} = \mathcal {L} _ {\text {G u i d a n c e}} + \gamma \cdot \mathcal {L} _ {\text {o r t h}}. \tag {12} +$$ + +# 5. Experiment + +In this section, we first describe the experimental details and evaluation metrics, then present a quantitative and qualitative analysis of our method. Please refer to Appendix for more results. + +# 5.1. Experimental Details + +Implementation details. We use Stable Diffusion v2.1 for all methods. We employ the prompt template "a photo of the face of a {occupation}, a person". At inference time, for each bias, we generate 100 images per occupation across 100 occupations, resulting in a total of 10,000 images. We set $\eta = \alpha = 0.3$ , and train for 1000 iterations with a learning rate of 1e-5. For gender bias, we use the CelebA [12] dataset to train a binary classifier with two + +categories: {male, female}. For racial bias, we use the Fair-Face [10] dataset to train a classifier with the following four categories: WMELH = {White, Middle Eastern, Latino Hispanic}, Asian = {East Asian, Southeast Asian}, Black, and Indian. Please refer to the supplementary for more details. + +Compared methods. In this work, we compare our method with recent state-of-the-art (SOTA) methods, including two retraining-based methods, Finetuning for Fairness (F4Fair) [30] and Inclusive Text-to-Image Generation (ITI-GEN) [37], an approach that does not require retraining DM, H-Distribution Guidance (H Guidance) [18], a closed-form editing approach, Unified Concept Editing (UCE) [8] and an h-space manipulation approach, Interpretable Diffusion [11]. Please refer to Appendix for more detailed introduction of these methods. + +# 5.2. Evaluation Metrics + +We evaluate all methods in three metrics: + +Fairness Discrepancy (FD). Following prior work[18], we adopt the Fairness Discrepancy (FD) metric. For an attribute $a$ and target distribution $p_{\theta}^{a}$ , we use a high-accuracy classifier $\mathcal{C}_a$ to compute the fairness performance: $||p_{\theta}^{a} - \mathbb{E}_{\mathbf{x}\sim p_{\theta}(\mathbf{x})}(\mathbf{y})||_2$ where $\mathbf{y}$ is the softmax output of $\mathcal{C}_a(\mathbf{x})$ . The target distribution $p_{\theta}^{a}$ can be any user-defined vector, typically uniform. A lower FD score indicates a closer match to the target distribution. + +$\mathbf{CLIP}_{\mathrm{sim}}$ . Besides fairness, the debiased model should maintain the ability to generate images that are semantically close to their text prompts. Therefore, following [30], we report the CLIP similarity score $\mathbf{CLIP}_{\mathrm{sim}}$ between the generated image and its prompt. + +BRISQUE. The generated image quality is also important as the debiasing process should not influence the image generation ability. We use the BRISQUE metric for evaluate the quality of the generated images. + +![](images/f2805a736552659f68ca2a66baf6ac651c7b92f3064bb42ec86d667c73cff5a6.jpg) +(a) original + +![](images/ad12b6be35af0bdd06443189bd11d335fd8d62b67208a9c383b9fca1f5116666.jpg) +(b) debiased (Ours) +Figure 7. Images generated from the original Stable Diffusion model (left) and FairGen (right) for racial debiasing with the prompt "A photo of a sportsman". The racial group distribution shifts from WMELH : Asian : Black : Indian = 7:2:5:1 in the original model to 4:4:4:3 with FairGen. + +Table 1. Comparisons of our method to the SOTA methods in gender bias over two predefined distributions: $f_{\theta}^{1} = (0.5, 0.5)$ and $f_{\theta}^{2} = (0.2, 0.8)$ , representing the probability of male and female respectively. + +
MethodFD ↓CLIPsim ↑BRISQUE ↑
f1θf2θf1θf2θf1θf2θ
Original SD0.4240.8470.380.3838.6538.69
F4Fair0.1650.3870.360.3538.2137.64
ITI-GEN0.0930.4720.320.3337.6237.96
H Guidance0.1180.3980.310.3238.5438.65
UCE0.2840.5360.360.2937.1236.54
InterDiff0.1770.3860.350.3438.3238.04
Ours0.0030.0050.380.3838.4638.72
+ +Table 2. Comparisons of our method to the SOTA methods in racial bias over two distributions: $f_{\theta}^{1} = (0.25, 0.25, 0.25, 0.25)$ and $f_{\theta}^{2} = (0.4, 0.3, 0.2, 0.1)$ , representing the probability of WMELH, Asian, Black, and Indian respectively. + +
MethodFD ↓CLIPsim ↑BRISQUE ↑
f1θf2θf1θf2θf1θf2θ
Original SD0.3840.4970.460.4138.9438.65
F4Fair0.2200.3050.430.4038.8038.45
ITI-GEN0.1310.5420.310.3136.6137.09
H Guidance0.1500.2850.410.3738.6538.60
UCE0.2900.4600.450.3937.9037.80
InterDiff0.1500.3910.400.3837.6437.82
Ours0.0950.1500.460.4138.9638.66
+ +# 5.3. Results + +Comparisons in gender debiasing. Table 1 demonstrates that our method outperforms others in mitigating gender bias over two predefined distributions: $f_{\theta}^{1} = (0.5,0.5)$ , representing equal likelihood of male and female, and $f_{\theta}^{2} = (0.2,0.8)$ , where male and female have a $20\%$ and $80\%$ probability, respectively. Original SD model exhibits high FD scores (0.424 at $f_{\theta}^{1}$ and 0.847 at $f_{\theta}^{2}$ ), indicating significant bias towards one gender. While previous methods like F4Fair, ITI-GEN, H Guidance, and UCE reduce bias to some extent, our method achieves the most reduction, with FD scores of 0.003 at $f_{\theta}^{1}$ and 0.005 at $f_{\theta}^{2}$ . Additionally, our method preserves semantic similarity (CLIPsim of 0.38) and image quality (BRISQUE scores of 38.46 and 38.72), matching or slightly surpassing the Original SD model. The qualitative results in Figure 6 further demonstrate that our approach effectively mitigates gender bias without compromising image quality or semantic coherence. + +Comparisons in racial debiasing. Table 2 compares methods over two distributions: $f_{\theta}^{1} = (0.25, 0.25, 0.25, 0.25)$ , representing equal probabilities for WMELH, Asian, Black, and Indian, and $f_{\theta}^{2} = (0.4, 0.3, 0.2, 0.1)$ , with higher probabilities for + +WMELH and Asian, and lower probabilities for Black and Indian. Original SD model shows significant racial bias, with FD scores of 0.384 for $f_{\theta}^{1}$ and 0.497 for $f_{\theta}^{2}$ , indicating poor calibration to either distribution. + +Debiasing methods such as F4Fair, H Guidance, and UCE reduce this bias, with H Guidance achieving relatively lower FD scores. However, our method performs the best, reducing bias to 0.095 for $f_{\theta}^{1}$ and 0.150 for $f_{\theta}^{2}$ . In terms of semantic similarity, the Original SD model sets a strong baseline (0.46 for $f_{\theta}^{1}$ and 0.41 for $f_{\theta}^{2}$ ), and our method maintains this high alignment, ensuring that debiasing does not impair semantic accuracy. Regarding image quality, our approach slightly improves upon the Original SD model, achieving the highest scores (39.10 for $f_{\theta}^{1}$ and 38.90 for $f_{\theta}^{2}$ ). Again, the qualitative results in Figure 7 further verify that our method outperforms others in reducing racial bias while preserving both semantic similarity and image quality. + +Comparisons in intersectional debiasing We also consider a more complex challenge of intersectional debiasing, i.e., jointly debiasing both gender and racial biases. The target distribution $f_{\theta}$ is set to be uniform for both gender and racial bias. Table 4 shows that Original SD model exhibits significant bias with an FD score of 0.214. While F4Fair + +Table 3. Transferability of proposed method across different DMs + +
TrainTestGenderRacial
FD ↓CLIPsim ↑BRISQUE ↑FD ↓CLIPsim ↑BRISQUE ↑
v2.1v1.40.0060.3738.400.0970.4538.90
v2.1v1.50.0060.3738.420.0960.4538.92
v2.1v2.00.0080.3838.450.0950.4638.95
v2.1v2.10.0030.3838.460.0950.4638.96
+ +Table 4. Evaluation of mitigating intersectional bias. + +
MethodFD ↓CLIPsim↑BRISQUE ↑
Original SD0.2140.3539.24
F4Fair0.1450.3638.90
ITI-GEN0.0980.3238.01
H Guidance0.1300.3338.75
UCE0.1800.3438.60
InterDiff0.1580.3538.67
Ours0.0470.3639.50
+ +and H Guidance reduce this bias to 0.145 and 0.130, respectively, our method achieves a much lower FD score of 0.047, reflecting a substantial improvement in fairness. + +For semantic similarity, the Original SD scores 0.35, and F4Fair slightly improves it to 0.36. Our method matches this performance, maintaining semantic coherence while reducing bias. Regarding image quality, the Original SD has a BRISQUE score of 39.24, while our method improves it to 39.50, indicating enhanced perceptual quality. These results demonstrate that our method excels at jointly debiasing both gender and racial biases, significantly reducing bias without sacrificing semantic accuracy or image quality. This highlights its robustness and practicality in real-world settings where multiple biases are present. + +Transferability across different DMs. We further evaluate the transferability of our method by training it with Stable Diffusion v2.1 and testing it over other versions, and results are reported in Table 3. From the table we can see that the performance of in all metrics are close to the optimal results achieved when training and testing are performed using the same model version (v2.1). While there is a slight increase in FD when tested on v1.4, v1.5, and v2.0, the differences are minor, indicating that our method can effectively generalize across different model versions. This robustness highlights the method's capability to transfer learned features across varying model conditions. + +# 5.4. Ablation Study + +Attaching adapters in U-Net. The results in Table 5 illustrate the impact of attaching adapters to different layers of the U-Net in the DM. When adapters are attached to all layers, the fairness score (FD) is 0.165, with a CLIP similarity score of 0.33 and a BRISQUE score of 35.00. The best results are obtained when adapters are attached only to + +Table 5. Ablation of layers that are attached adapters. + +
LocationFD ↓CLIPsim ↑BRISQUE ↑
All layers0.1650.3335.00
Non-CA layers0.1580.3437.00
CA layers0.0470.3639.50
+ +the CA layers. This configuration significantly reduces the FD score to 0.047, increases semantic similarity $(\mathrm{CLIP}_{\mathrm{sim}} = 0.36)$ , and enhances image quality with a BRISQUE score of 39.50. These highlights that focusing the adapters on the cross-attention layers leads to the most substantial improvements in both fairness and image quality. + +Impact of orthogonal regularization (OR). Table 6 demonstrates the impact of orthogonal regularization (OR). Without OR, the FD score is 0.143, semantic similarity $(\mathrm{CLIP}_{\mathrm{sim}})$ drops to 0.29, and image quality (BRISQUE) is 37.92. When OR is applied, performance improves significantly across all metrics, with a much lower FD score of 0.047, higher semantic similarity of 0.36, and improved image quality (BRISQUE = 39.50). This highlights the effectiveness of orthogonal regularization in reducing bias and improving overall performance. + +Table 6. Orthogonal Ablation + +
LocationFD ↓CLIPsim ↑BRISQUE ↑
Ours w\o OR0.1430.2937.92
Ours w\OR0.0470.3639.50
+ +# 6. Conclusion + +In this paper, we propose FairGen, a plug-and-play method that learns attribute latent directions in a self-discovering manner. Our method can not only jointly debias multiple attributes in DMs, but also enables the generated images to follow a prescribed attribute distribution. It is lightweight and can be integrated with other DMs without retraining. Extensive experiments on debiasing gender, racial, and their intersectional biases show that our method outperforms previous SOTA by a large margin. We believe that our work marks a critical advancement in addressing harmful societal stereotypes within diffusion models, and it contributes to the ethical real-world applications of text-to-image DMs. + +# Acknowledgements + +This work was supported in part by the National Natural Science Foundation of China under Grant No. 62306261, and the Shun Hing Institute of Advanced Engineering (SHIAE) Fund under Grant No. 8115074. + +# References + +[1] Federico Bianchi, Pratyusha Kalluri, Esin Durmus, Faisal Ladhak, Myra Cheng, Debora Nozza, Tatsunori Hashimoto, Dan Jurafsky, James Zou, and Aylin Caliskan. Easily accessible text-to-image generation amplifies demographic stereotypes at large scale. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, pages 1493-1504, Chicago, Illinois, USA, 2023. ACM. 2 +[2] Jaemin Cho, Abhay Zala, and Mohit Bansal. Dall-eval: Probing the reasoning skills and social biases of text-to-image generative transformers. In Proceedings of the International Conference on Computer Vision (ICCV), Paris, France, 2023. IEEE. 2 +[3] Kristy Choi, Aditya Grover, Trisha Singh, Rui Shu, and Stefano Ermon. Fair generative modeling via weak supervision. In International Conference on Machine Learning, pages 1887-1898. PMLR, 2020. 3 +[4] Bradley Efron. Tweedie's formula and selection bias, 2011. 5 +[5] Mureji Fatunde and Crystal Tse. Digital media company stability ai raises funds at $1 billion value. https://www.bloomberg.com/news/articles/2022-10-17/digital-media-firm-stability-ai-raises-funds-at-1-billion-value, 2022.1 +[6] Rohit Gandikota, Joanna Materzynska, Jaden Fiitto-Kaufman, and David Bau. Erasing concepts from diffusion models, 2023. 3 +[7] Rohit Gandikota, Joanna Materzynska, Tingrui Zhou, Antonio Torralba, and David Bau. Concept sliders: Lora adaptors for precise control in diffusion models, 2023. 3 +[8] Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzyńska, and David Bau. Unified concept editing in diffusion models. arXiv preprint arXiv:2308.14761, 2023. 2, 3, 6, 1 +[9] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance, 2022. 3, 5 +[10] Jungseock Joo. Fairface attribute model. https://github.com/jojoa/fairface, 2021. 6, 1 +[11] Hang Li, Chengzhi Shen, Philip Torr, Volker Tresp, and Jindong Gu. Self-discovering interpretable diffusion latent directions for responsible text-to-image generation, 2024. 3, 6, 1 +[12] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 3730-3738, Santiago, Chile, 2015. IEEE. 6, 1 +[13] Alexandra Sasha Luccioni, Christopher Akiki, Margaret Mitchell, and Yacine Jernite. Stable bias: Analyzing so + +cietal representations in diffusion models. arXiv preprint arXiv:2303.11408, 2023. 1, 2 +[14] Mengyao Lyu, Yuhong Yang, Haiwen Hong, Hui Chen, Xuan Jin, Yuan He, Hui Xue, Jungong Han, and Guiguang Ding. One-dimensional adapter to rule them all: Concepts, diffusion models and erasing applications, 2024. 4 +[15] Ellis Monk. The monk skin tone scale, 2023. 2 +[16] Junhyun Nam, Sangwoo Mo, Jaeho Lee, and Jinwoo Shin. Breaking the spurious causality of conditional generation via fairness intervention with corrective sampling. Transactions on Machine Learning Research, 2023. 3 +[17] Hadas Orgad, Bahjat Kawar, and Yonatan Belinkov. Editing implicit assumptions in text-to-image diffusion models. 2023. 2, 3 +[18] Rishubh Parihar, Abhijnya Bhat, Abhipsa Basu, Saswat Mallick, Jogendra Nath Kundu, and R. Venkatesh Babu. Balancing act: Distribution-guided debiasing in diffusion models, 2024. 2, 3, 6, 1 +[19] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2 +[20] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1 +[21] Robin Rombach. Stable diffusion v1-3 model card, 2022. 2 +[22] Robin Rombach. Stable diffusion v1-4 model card, 2022. 2 +[23] Robin Rombach. Stable diffusion v1-5 model card, 2022. 2 +[24] Robin Rombach. Stable diffusion v2-0 model card, 2022. 2 +[25] Robin Rombach. Stable diffusion v2-1 model card, 2022. 2 +[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3 +[27] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1 +[28] Prasanna Sattigeri, Samuel C Hoffman, Vijil Chenthamarakshan, and Kush R Varshney. Fairness gan: Generating datasets with fairness properties using a generative adversarial network. IBM Journal of Research and Development, 63 (4/5):3-1, 2019. 3 +[29] Preethi Seshadri, Sameer Singh, and Yanai Elazar. The bias amplification paradox in text-to-image generation. arXiv preprint arXiv:2308.00755, 2023. 2 +[30] Xudong Shen, Chao Du, Tianyu Pang, Min Lin, Yongkang Wong, and Mohan Kankanhalli. Finetuning text-to-image diffusion models for fairness, 2024. 2, 3, 4, 6, 1 +[31] Christopher TH Teo, Milad Abdollahzadeh, and Ngai-Man Cheung. Fair generative models via transfer learning. In Pro + +ceedings of the AAAI Conference on Artificial Intelligence, pages 2429-2437, 2023. 3 +[32] Soobin Um and Changho Suh. A fair generative model using lecam divergence. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10034-10042, 2023. +[33] Boris van Breugel, Trent Kyono, Jeroen Berrevoets, and Michaela van der Schaar. Decaf: Generating fair synthetic data using causally-aware generative networks. Advances in Neural Information Processing Systems, 34:22221-22233, 2021. 3 +[34] Zihao Wang, Lin Gui, Jeffrey Negrea, and Victor Veitch. Concept algebra for score-based conditional model. In ICML 2023 Workshop on Structured Probabilistic Inference & Generative Modeling, Honolulu, HI, 2023. PMLR. 2 +[35] Depeng Xu, Shuhan Yuan, Lu Zhang, and Xintao Wu. Fairgan: Fairness-aware generative adversarial networks. In 2018 IEEE International Conference on Big Data (Big Data), pages 570-575. IEEE, 2018. 3 +[36] Ning Yu, Ke Li, Peng Zhou, Jitendra Malik, Larry Davis, and Mario Fritz. Inclusive gan: Improving data and minority coverage in generative models. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 377-393. Springer, 2020. 2, 3, 1 +[37] Cheng Zhang, Xuanbai Chen, Siqi Chai, Chen Henry Wu, Dmitry Lagun, Thabo Beeler, and Fernando De la Torre. Itigen: Inclusive text-to-image generation, 2023. 6 \ No newline at end of file diff --git a/2025/FairGen_ Enhancing Fairness in Text-to-Image Diffusion Models via Self-Discovering Latent Directions/images.zip b/2025/FairGen_ Enhancing Fairness in Text-to-Image Diffusion Models via Self-Discovering Latent Directions/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..220948606c318d8831c85c659c622c15fe2036ce --- /dev/null +++ b/2025/FairGen_ Enhancing Fairness in Text-to-Image Diffusion Models via Self-Discovering Latent Directions/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f4ab26b1aa9040f8b2fec1f6cbbb929c635d0822b8d2fc1e5113279b1879547d +size 656176 diff --git a/2025/FairGen_ Enhancing Fairness in Text-to-Image Diffusion Models via Self-Discovering Latent Directions/layout.json b/2025/FairGen_ Enhancing Fairness in Text-to-Image Diffusion Models via Self-Discovering Latent Directions/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..272a443087ee3b3894ce381fef6434042c21b271 --- /dev/null +++ b/2025/FairGen_ Enhancing Fairness in Text-to-Image Diffusion Models via Self-Discovering Latent Directions/layout.json @@ -0,0 +1,8935 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 95, + 102, + 516, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 102, + 516, + 140 + ], + "spans": [ + { + "bbox": [ + 95, + 102, + 516, + 140 + ], + "type": "text", + "content": "FairGen: Enhancing Fairness in Text-to-Image Diffusion Models via Self-Discovering Latent Directions" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 97, + 160, + 517, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 160, + 517, + 190 + ], + "spans": [ + { + "bbox": [ + 97, + 160, + 517, + 190 + ], + "type": "text", + "content": "Yilei Jiang" + }, + { + "bbox": [ + 97, + 160, + 517, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 160, + 517, + 190 + ], + "type": "text", + "content": " Wei-Hong Li" + }, + { + "bbox": [ + 97, + 160, + 517, + 190 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 97, + 160, + 517, + 190 + ], + "type": "text", + "content": " Yiyuan Zhang" + }, + { + "bbox": [ + 97, + 160, + 517, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 160, + 517, + 190 + ], + "type": "text", + "content": " Minghong Cai" + }, + { + "bbox": [ + 97, + 160, + 517, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 160, + 517, + 190 + ], + "type": "text", + "content": " Xiangyu Yue" + }, + { + "bbox": [ + 97, + 160, + 517, + 190 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 97, + 160, + 517, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 160, + 517, + 190 + ], + "type": "text", + "content": "MMLab, CUHK " + }, + { + "bbox": [ + 97, + 160, + 517, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 97, + 160, + 517, + 190 + ], + "type": "text", + "content": "University of Bristol " + }, + { + "bbox": [ + 97, + 160, + 517, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 97, + 160, + 517, + 190 + ], + "type": "text", + "content": "SHIAE, CUHK" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 217, + 200, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 217, + 200, + 229 + ], + "spans": [ + { + "bbox": [ + 151, + 217, + 200, + 229 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 242, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 242, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 242, + 296, + 579 + ], + "type": "text", + "content": "While Diffusion Models (DM) exhibit remarkable performance across various image generative tasks, they nonetheless reflect the inherent bias presented in the training set. As DMs are now widely used in real-world applications, these biases could perpetuate a distorted worldview and hinder opportunities for minority groups. Existing methods on debiasing DMs usually requires model retraining with a human-crafted reference dataset or additional classifiers, which suffer from two major limitations: (1) collecting reference datasets causes expensive annotation cost; (2) the debiasing performance is heavily constrained by the quality of the reference dataset or the additional classifier. To address the above limitations, we propose FairGen, a plug-and-play method that learns attribute latent directions in a self-discovering manner, thus eliminating the reliance on such reference dataset. Specifically, FairGen consists of two parts: a set of attribute adapters and a distribution indicator. Each adapter in the set aims to learn an attribute latent direction, and is optimized via noise composition through a self-discovering process. Then, the distribution indicator is multiplied by the set of adapters to guide the generation process towards the prescribed distribution. Our method enables debiasing multiple attributes in DMs simultaneously, while remaining lightweight and easily integrable with other DMs, eliminating the need for retraining. Extensive experiments on debiasing gender, racial, and their intersectional biases show that our method outperforms previous SOTA by a large margin." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 601, + 135, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 135, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 135, + 613 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 621, + 296, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 621, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 621, + 296, + 694 + ], + "type": "text", + "content": "State-of-the-art Text-to-Image Diffusion Models (DMs) such as Stable Diffusion [26], DALL-E 3 [20] and Imagegen [27] have demonstrated remarkable performance in generating high-quality images. With the rapid development of DMs, an increasing number of individuals and corporations are choosing to utilize them to serve their own" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 322, + 217, + 548, + 299 + ], + "blocks": [ + { + "bbox": [ + 322, + 217, + 548, + 299 + ], + "lines": [ + { + "bbox": [ + 322, + 217, + 548, + 299 + ], + "spans": [ + { + "bbox": [ + 322, + 217, + 548, + 299 + ], + "type": "image", + "image_path": "c666aa8c76200144f15fcef6ff46f095b4eca8a2e185d8b6cb4cec26163707dd.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 321, + 303, + 547, + 385 + ], + "blocks": [ + { + "bbox": [ + 321, + 303, + 547, + 385 + ], + "lines": [ + { + "bbox": [ + 321, + 303, + 547, + 385 + ], + "spans": [ + { + "bbox": [ + 321, + 303, + 547, + 385 + ], + "type": "image", + "image_path": "2fe86488056ac03a8517bbb81ec4901cd2a0c8f1e4840b7f47a039e1db4f4f57.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 387, + 555, + 453 + ], + "lines": [ + { + "bbox": [ + 313, + 387, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 555, + 453 + ], + "type": "text", + "content": "Figure 1. Illustration of gender bias associated with different occupations in Stable Diffusion v2.1 and the debiased results for gender by our method. The leadership roles such as CEOs are biased towards male figures, whereas supportive roles such as nurses are biased towards female figures. Our method can generate debiased results regarding the gender bias." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 474, + 555, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 557 + ], + "type": "text", + "content": "purposes. For instance, Stable Diffusion v1.5 has been downloaded over 8 million times on Huggingface, and Midjourney is used by over a million users [5]. However, existing DMs have been found to generate biased content across various demographic factors, such as gender and race [13], which could have harmful effects on society when these models are implemented in real-world applications." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": "In Figure 1, we randomly generate several images of two occupations using Stable Diffusion v2.1. Given the prompt of \"A photo of a CEO\", the generated images predominantly depict male figures, reinforcing the stereotype that leadership roles, such as CEOs and doctors, are male-dominated. On the contrary, when the prompt is \"A photo of a nurse\", the majority of generated images depict female figures, reflecting the bias that supportive roles are traditionally associated with women. Regarding racial bias, we randomly generate 1000 images using Stable Diffusion v2.1 with the prompt \"A photo of a worker\". The statistic of 1000 images depicted in Figure 2 shows a strong bias in racial representation, with White individuals making up " + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "inline_equation", + "content": "71\\%" + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": " of the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18411" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 72, + 153, + 157 + ], + "blocks": [ + { + "bbox": [ + 70, + 72, + 153, + 157 + ], + "lines": [ + { + "bbox": [ + 70, + 72, + 153, + 157 + ], + "spans": [ + { + "bbox": [ + 70, + 72, + 153, + 157 + ], + "type": "image", + "image_path": "98aa4a433620d2f35c65c4f44639e1828b7f119e834e55e194586642246eee2f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 160, + 162, + 216 + ], + "lines": [ + { + "bbox": [ + 55, + 160, + 162, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 160, + 162, + 216 + ], + "type": "text", + "content": "Figure 2. Racial bias in randomly generated 1000 images with the prompt \"A photo of a worker\" using Stable Diffusion v2.1." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 74, + 288, + 155 + ], + "blocks": [ + { + "bbox": [ + 173, + 74, + 288, + 155 + ], + "lines": [ + { + "bbox": [ + 173, + 74, + 288, + 155 + ], + "spans": [ + { + "bbox": [ + 173, + 74, + 288, + 155 + ], + "type": "image", + "image_path": "323da653ce762f0e6445cdba1deada4e466da9adaad83e93f0ce80c356568657.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 168, + 160, + 295, + 216 + ], + "lines": [ + { + "bbox": [ + 168, + 160, + 295, + 216 + ], + "spans": [ + { + "bbox": [ + 168, + 160, + 295, + 216 + ], + "type": "text", + "content": "Figure 3. Gender bias in randomly generated 1000 images with the prompt \"A photo of a CEO\" using Stable Diffusion v1.3, v1.4, v1.5, v2.0, and v2.1." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 242, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 242, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 242, + 296, + 361 + ], + "type": "text", + "content": "total, while minority groups like Middle Eastern, Latino, Black, and Indian each account for only " + }, + { + "bbox": [ + 55, + 242, + 296, + 361 + ], + "type": "inline_equation", + "content": "3 - 4\\%" + }, + { + "bbox": [ + 55, + 242, + 296, + 361 + ], + "type": "text", + "content": ". This bias in DM, produce less accurate or fair results for underrepresented populations. We further investigate such bias situation across different versions of DMs. We randomly generate 1000 images with the prompt of \"A photo of a CEO\" using five versions of DMs: Stable Diffusion v1.3 [21], v1.4 [22], v.1.5 [23], v2.0 [24] and v2.1 [25]. Figure 3 demonstrates that gender bias exists across all versions of DMs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 365, + 295, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 295, + 663 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 295, + 663 + ], + "type": "text", + "content": "This bias arises from two primary sources. First, the training data, predominantly sourced from the web, contains inherent biases. Second, bias is partially inherited from the CLIP [19] model used in the generation process. In large text-to-image DMs, these biases are even more pronounced, often manifesting in gendered associations with specific professions. Several studies have attempted to mitigate bias in DMs through retraining [30, 36] or approaches that do not require retraining the DM [8, 17, 18]. In training-based methods, Shen et al. [30] propose curating a reference dataset and aligning the distribution of generated images with this dataset. Parihar et al. [18] avoid retraining the DM by leveraging demographic information embedded in the latent features of the denoising U-Net to guide generation. However, their method requires training an additional MLP-based Attribute Distribution Predictor using pseudo labels from existing attribute classifiers. This reliance on accurate attribute classifiers limits its debiasing performance. In contrast, Gandikota et al. [8] and Orgad et al. [17] employ closed-form editing to adjust concepts within DMs without retraining. While training-based methods depend on costly and quality-constrained annotated reference datasets, methods that do not require retraining the DM offer greater implementation efficiency but often exhibit lower debiasing effectiveness." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "To address these limitations, we propose FairGen, a plug-and-play method that achieves debiased image generation (Figure 1) by automatically learning attribute latent directions, mitigating the reliance on reference datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 359 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 359 + ], + "type": "text", + "content": "FairGen is composed of two components: a set of attribute adapters and a distribution indicator. Each adapter is trained to learn an attribute-specific latent direction, optimized through noise composition in a self-discovering manner, i.e., our method automatically learns attribute latent directions without relying on a labeled reference dataset. Through noise composition, our method explores and optimizes attribute directions directly from the model's latent space, uncovering patterns without external supervision. At inference stage, the distribution indicator is then applied to select attribute-specific adapters, guiding the generative process towards the desired distribution. We comprehensively evaluate the effectiveness of our approach in debiasing gender, racial, and intersectional biases using occupational prompts. Experimental results demonstrate that our method not only achieves state-of-the-art performance in single and multiple attribute debiasing tasks but also preserves the generation quality of DM. Furthermore, we show that once FairGen is trained on one diffusion model, it can be seamlessly integrated into other models without retuning. Thanks to its strong transferability and plug-and-play functionality, our method offers a practical solution for both individual users and organizations, facilitating the responsible use of diffusion models in future applications." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 326, + 361, + 552, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 361, + 552, + 372 + ], + "spans": [ + { + "bbox": [ + 326, + 361, + 552, + 372 + ], + "type": "text", + "content": "To summarize, our main contributions are as following:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 326, + 373, + 553, + 504 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 326, + 373, + 553, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 373, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 326, + 373, + 553, + 432 + ], + "type": "text", + "content": "- We propose FairGen, a novel method for debiasing DMs by learning attribute latent directions in a self-discovering manner, eliminating the reliance on the reference dataset or classifier, and thus significantly reduce the cost." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 326, + 433, + 553, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 433, + 553, + 468 + ], + "spans": [ + { + "bbox": [ + 326, + 433, + 553, + 468 + ], + "type": "text", + "content": "- FairGen is lightweight, plug-and-play and shows good transferability across different DMs, making it more convenient to deploy in the real world." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 326, + 468, + 553, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 468, + 553, + 504 + ], + "spans": [ + { + "bbox": [ + 326, + 468, + 553, + 504 + ], + "type": "text", + "content": "- Extensive experiments show that our method achieves SOTA performance across diverse debiasing tasks while retaining the image generation quality." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 517, + 402, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 402, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 402, + 529 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "Bias in Diffusion Models. Diffusion models for text-to-image generation (T2I) have been observed to produce biased and stereotypical images, even when given neutral prompts. Cho et al. [2] found that Stable Diffusion (SD) tends to generate images of males when prompted with occupations, with skin tones predominantly centered around a few shades from the Monk Skin Tone Scale [15]. Seshadri et al. [29] noted that SD reinforces gender-occupation biases presented in its training data. In addition to occupations, Bianchi et al. [1] discovered that simple prompts involving character traits and other descriptors also result in stereotypical images. Luccioni et al. [13] created a tool to compare generated image collections across different genders and ethnicities. Moreover, Wang et al. [34] introduced a text-to-image association test and found that SD tends to" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18412" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "associate females more with family roles and males more with career-related roles." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 99, + 296, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 99, + 296, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 99, + 296, + 326 + ], + "type": "text", + "content": "Debiasing Diffusion Models by retraining. Before DMs, previous approaches mainly focus on debiasing GANs by assuming access to the labels of sensitive attributes and ensure no correlation exists between the decision and sensitive attributes [3, 16, 28, 31-33, 35, 36]. More recently, regarding DMs, Shen et al. [30] propose a distributional alignment loss to guide the characteristics of the generated images towards target distribution and use adjusted finetuning to directly optimize losses on the generated images. This method requires a reference training dataset to complete the retraining process, whereas our method does not need such reference dataset, which largely reduces annotation costs. Yu et al. [36] enhance fairness in text-to-image synthesis by incorporating reference images. and leverage visual exemplars to more effectively represent attributes that are difficult to describe in words, such as nuanced variations in skin tones. Li et al. [11] find a interpretable latent vector for a given concept, and achieve debiasing by manipulating the h-space inside DMs during inference stage." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 328, + 296, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 328, + 296, + 484 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 296, + 484 + ], + "type": "text", + "content": "Debiasing Diffusion Models without retraining. Parihar et al. [18] propose Distribution Guidance (DG), which guides the generated images to follow the prescribed attribute distribution. Although DG does not require retraining DMs, it requires training an Attribute Distribution Predictor (ADP) that maps the latent features to the distribution of attributes. Since ADP is trained with pseudo labels generated from existing attribute classifiers, the performance of DG is largely constrained by the accuracy of attribute classifiers. Gandikota et al. [8] and Orgad et al. [17] use closed-form editing approach to edits concepts inside DM without training. Despite the efficiency, these methods are less effective than training-based approaches." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "text", + "content": "**Controlling concepts in Diffusion Models.** Our work is inspired by recent work on controlling concepts [6, 7] in diffusion models. Gandikota et al. [6] propose to fine-tune the DM to erase an undesired concept while maintaining the model's behavior and capabilities, such as removing the NSFW (not safe for work) contents. ConceptSlider [7] also fine-tunes a LoRA adapter in a diffusion model to control the concept in the generated images. Unlike ESD that focuses on erasing concepts and ConceptSlider aims at precisely adjusting the extent of a concept (e.g., controlling the size of eyes), this work aims to debias the Diffusion Models, i.e., controlling the portion of generated images for different attributes such as male and female. Additionally, ConceptSlider learns an adapter per concept and use a scaling factor to control the extent of a concept (e.g., size of eye) while we learn an adapter per direction (e.g., female or male) in an attribute/concept (e.g., gender) and design an inference pipeline with an indicator to automatically select adapters based on a given distribution." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 391, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 391, + 85 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 391, + 85 + ], + "type": "text", + "content": "3. Preliminary" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 87, + 555, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 87, + 555, + 219 + ], + "spans": [ + { + "bbox": [ + 313, + 87, + 555, + 219 + ], + "type": "text", + "content": "Latent Diffusion Models (LDMs) perform the diffusion process within the latent space [26]. During training, noise is added to the encoded latent representation of the input image " + }, + { + "bbox": [ + 313, + 87, + 555, + 219 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 87, + 555, + 219 + ], + "type": "text", + "content": ", resulting in a noisy latent code " + }, + { + "bbox": [ + 313, + 87, + 555, + 219 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 313, + 87, + 555, + 219 + ], + "type": "text", + "content": " at each time step " + }, + { + "bbox": [ + 313, + 87, + 555, + 219 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 87, + 555, + 219 + ], + "type": "text", + "content": ". In the pretraining stage, an autoencoder framework is employed to map images into a lower-dimensional latent space via an encoder: " + }, + { + "bbox": [ + 313, + 87, + 555, + 219 + ], + "type": "inline_equation", + "content": "z = \\mathcal{E}(x)" + }, + { + "bbox": [ + 313, + 87, + 555, + 219 + ], + "type": "text", + "content": ". The decoder then reconstructs images from these latent codes: " + }, + { + "bbox": [ + 313, + 87, + 555, + 219 + ], + "type": "inline_equation", + "content": "x \\approx \\mathcal{D}(\\mathcal{E}(x))" + }, + { + "bbox": [ + 313, + 87, + 555, + 219 + ], + "type": "text", + "content": ". This process ensures that the latent space retains the essential semantic information of the image. The training objective of the diffusion model in the latent space is given by:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 322, + 224, + 555, + 246 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 224, + 555, + 246 + ], + "spans": [ + { + "bbox": [ + 322, + 224, + 555, + 246 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {L D M}} = \\mathbb {E} _ {z \\sim \\mathcal {E} (x), c, \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (z _ {t}, c, t) \\| _ {2} ^ {2} \\right], \\tag {1}", + "image_path": "e41e6f66e2cd75fc6e7fd3f3a8e7ca51ed699ebbd94efbfb1419f5232e683a16.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 251, + 554, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 554, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 554, + 369 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 251, + 554, + 369 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 251, + 554, + 369 + ], + "type": "text", + "content": " is Gaussian noise sampled from a normal distribution " + }, + { + "bbox": [ + 313, + 251, + 554, + 369 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,1)" + }, + { + "bbox": [ + 313, + 251, + 554, + 369 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 251, + 554, + 369 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 251, + 554, + 369 + ], + "type": "text", + "content": " is the denoising network, and " + }, + { + "bbox": [ + 313, + 251, + 554, + 369 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 251, + 554, + 369 + ], + "type": "text", + "content": " represents any conditioning embeddings (e.g., text or class labels). At the inference stage, a latent code " + }, + { + "bbox": [ + 313, + 251, + 554, + 369 + ], + "type": "inline_equation", + "content": "z_{T}" + }, + { + "bbox": [ + 313, + 251, + 554, + 369 + ], + "type": "text", + "content": " is sampled from Gaussian noise at the initial timestep " + }, + { + "bbox": [ + 313, + 251, + 554, + 369 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 251, + 554, + 369 + ], + "type": "text", + "content": ". The denoising network " + }, + { + "bbox": [ + 313, + 251, + 554, + 369 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 251, + 554, + 369 + ], + "type": "text", + "content": " is then applied iteratively to remove the noise over several steps, generating a denoised latent representation " + }, + { + "bbox": [ + 313, + 251, + 554, + 369 + ], + "type": "inline_equation", + "content": "z_{0}" + }, + { + "bbox": [ + 313, + 251, + 554, + 369 + ], + "type": "text", + "content": ". Finally, the pretrained decoder reconstructs the image from the denoised latent code: " + }, + { + "bbox": [ + 313, + 251, + 554, + 369 + ], + "type": "inline_equation", + "content": "\\hat{x}_0 \\approx \\mathcal{D}(z_0)" + }, + { + "bbox": [ + 313, + 251, + 554, + 369 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 251, + 554, + 369 + ], + "type": "inline_equation", + "content": "\\hat{x}_0" + }, + { + "bbox": [ + 313, + 251, + 554, + 369 + ], + "type": "text", + "content": " is the generated output image." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 370, + 554, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 370, + 554, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 370, + 554, + 489 + ], + "type": "text", + "content": "Classifier-free Guidance aims to modulate image generation by steering the probability distribution towards data that is more probable according to an implicit classifier " + }, + { + "bbox": [ + 313, + 370, + 554, + 489 + ], + "type": "inline_equation", + "content": "p(c \\mid z_t)" + }, + { + "bbox": [ + 313, + 370, + 554, + 489 + ], + "type": "text", + "content": " [9]. It operates at the inference phase and the model is jointly trained on both conditional and unconditional denoising tasks. During inference, both the conditional and unconditional denoising scores are derived from the model. The final score " + }, + { + "bbox": [ + 313, + 370, + 554, + 489 + ], + "type": "inline_equation", + "content": "\\tilde{\\epsilon}_{\\theta}(z_t, c, t)" + }, + { + "bbox": [ + 313, + 370, + 554, + 489 + ], + "type": "text", + "content": " is then adjusted by weighting the conditioned score more heavily relative to the unconditioned score using a guidance scale " + }, + { + "bbox": [ + 313, + 370, + 554, + 489 + ], + "type": "inline_equation", + "content": "\\alpha > 1" + }, + { + "bbox": [ + 313, + 370, + 554, + 489 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 326, + 497, + 553, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 497, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 326, + 497, + 553, + 510 + ], + "type": "interline_equation", + "content": "\\tilde {\\epsilon} _ {\\theta} \\left(z _ {t}, c, t\\right) = \\epsilon_ {\\theta} \\left(z _ {t}, t\\right) + \\alpha \\left(\\epsilon_ {\\theta} \\left(z _ {t}, c, t\\right) - \\epsilon_ {\\theta} \\left(z _ {t}, t\\right)\\right) \\tag {2}", + "image_path": "afb9502b8e6793b622679ffdd9369be2caf386ee1a67073366377227cf1a7842.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 516, + 554, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 516, + 554, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 516, + 554, + 589 + ], + "type": "text", + "content": "The inference process begins with sampling a latent variable " + }, + { + "bbox": [ + 313, + 516, + 554, + 589 + ], + "type": "inline_equation", + "content": "z_{T} \\sim \\mathcal{N}(0,1)" + }, + { + "bbox": [ + 313, + 516, + 554, + 589 + ], + "type": "text", + "content": ", which is subsequently denoised using " + }, + { + "bbox": [ + 313, + 516, + 554, + 589 + ], + "type": "inline_equation", + "content": "\\tilde{\\epsilon}_{\\theta}(z_t,c,t)" + }, + { + "bbox": [ + 313, + 516, + 554, + 589 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 313, + 516, + 554, + 589 + ], + "type": "inline_equation", + "content": "z_{t - 1}" + }, + { + "bbox": [ + 313, + 516, + 554, + 589 + ], + "type": "text", + "content": ". The denoising is performed iteratively until obtaining " + }, + { + "bbox": [ + 313, + 516, + 554, + 589 + ], + "type": "inline_equation", + "content": "z_{0}" + }, + { + "bbox": [ + 313, + 516, + 554, + 589 + ], + "type": "text", + "content": ". Finally, the decoder transforms the latent representation " + }, + { + "bbox": [ + 313, + 516, + 554, + 589 + ], + "type": "inline_equation", + "content": "z_{0}" + }, + { + "bbox": [ + 313, + 516, + 554, + 589 + ], + "type": "text", + "content": " back into image space: " + }, + { + "bbox": [ + 313, + 516, + 554, + 589 + ], + "type": "inline_equation", + "content": "x_0 \\gets \\mathcal{D}(z_0)" + }, + { + "bbox": [ + 313, + 516, + 554, + 589 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 597, + 370, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 370, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 370, + 609 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": "Inspired by recent work that fine-tune diffusion models for erasing concepts [6] and controlling the level of concepts [7] (e.g., ages, size of eyes), we aim at reducing the bias in the DM by attaching and learning a set of lightweight adapters, each of which represents a category of an attribute (e.g., female of gender), guiding the DM towards an attribute latent direction. Unlike previous work that relies on additional reference datasets and has to finetune the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18413" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 66, + 534, + 249 + ], + "blocks": [ + { + "bbox": [ + 82, + 66, + 534, + 249 + ], + "lines": [ + { + "bbox": [ + 82, + 66, + 534, + 249 + ], + "spans": [ + { + "bbox": [ + 82, + 66, + 534, + 249 + ], + "type": "image", + "image_path": "e53d837e7dee188ed02e842596334fd404f0d6da60cc337a55385adc2773e2a0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 255, + 555, + 289 + ], + "lines": [ + { + "bbox": [ + 54, + 255, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 54, + 255, + 555, + 289 + ], + "type": "text", + "content": "Figure 4. Overview of FairGen's training pipeline. Attribute-specific adapters " + }, + { + "bbox": [ + 54, + 255, + 555, + 289 + ], + "type": "inline_equation", + "content": "(M)" + }, + { + "bbox": [ + 54, + 255, + 555, + 289 + ], + "type": "text", + "content": " are attached to the cross-attention layers in the denoising UNet. Target group and attribution direction are fed into the DM for composing the noise predictions (Eq. 5), which is used as self-discovering attribute direction guidance to optimize the adapters." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 308, + 296, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 308, + 296, + 489 + ], + "spans": [ + { + "bbox": [ + 54, + 308, + 296, + 489 + ], + "type": "text", + "content": "whole DM [30], we instead optimize the attached adapters via noise composition through a self-discovering process (detailed in Section 4.2). This significantly reduces the cost in computation and data. During the inference stage (Section 4.3), given a predefined target distribution (e.g., uniform), we introduce a distribution indicator implemented by a gating function to select one corresponding adapter which will be attached to the DM for generating image. In this way, the set of generated images will follow the predefined target distribution, and thus are not biased to some categories of an attribute (if the predefined target distribution is uniform). Our training and inference diagrams are illustrated in Figure 4 and Figure 5. In the following sections, we start elaborating our method in single attribute settings and then extend it to more general ones." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 501, + 296, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 501, + 296, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 501, + 296, + 526 + ], + "type": "text", + "content": "4.1. Attaching Light-weight attribute-specific Adapters" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 532, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 295, + 687 + ], + "type": "text", + "content": "To debias a DM for generating images of a given attribute that contains several categories (e.g., male and female are two categories of gender attribute), we first aim at equip the DM with skills of generating images for each category. To achieve this, we attach a light-weight adapter per category in each layer of the DM inspired by parameter-efficient fine-tuning (PEFT) instead of finetuning the whole model to achieve a good trade-off between performance and computational cost. In this work, we use the 1-dim adapter [14] and only add the adapter to each cross-attention layers of the denoising U-Net, as shown in Figure 4 as we find that attaching adapters to all layers does not help and will also increase the computational cost." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Specifically, for the " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "-th cross-attention layer parameterized by " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_i \\in \\mathbb{R}^{m \\times n}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " in the denoising U-Net, we attach" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 309, + 555, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 309, + 555, + 405 + ], + "spans": [ + { + "bbox": [ + 313, + 309, + 555, + 405 + ], + "type": "text", + "content": "an adapter to the layer to guide the attribute towards a certain category. The adapter consists of two 1-dim vectors: " + }, + { + "bbox": [ + 313, + 309, + 555, + 405 + ], + "type": "inline_equation", + "content": "\\pmb{p} \\in \\mathbb{R}^{m}" + }, + { + "bbox": [ + 313, + 309, + 555, + 405 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 309, + 555, + 405 + ], + "type": "inline_equation", + "content": "\\pmb{q} \\in \\mathbb{R}^{n}" + }, + { + "bbox": [ + 313, + 309, + 555, + 405 + ], + "type": "text", + "content": ". The forward process of the " + }, + { + "bbox": [ + 313, + 309, + 555, + 405 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 309, + 555, + 405 + ], + "type": "text", + "content": "-th cross-attention layer would be updated from " + }, + { + "bbox": [ + 313, + 309, + 555, + 405 + ], + "type": "inline_equation", + "content": "\\pmb{y}_{i} = \\pmb{W}_{i}\\pmb{x}_{i}" + }, + { + "bbox": [ + 313, + 309, + 555, + 405 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 309, + 555, + 405 + ], + "type": "inline_equation", + "content": "\\pmb{y}_{i} = \\pmb{W}_{i}\\pmb{x}_{i} + (\\pmb{q}_{i}^{T}\\pmb{x}_{i})\\cdot \\pmb{p}_{i}" + }, + { + "bbox": [ + 313, + 309, + 555, + 405 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 309, + 555, + 405 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{i} \\in \\mathbb{R}^{n}" + }, + { + "bbox": [ + 313, + 309, + 555, + 405 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 309, + 555, + 405 + ], + "type": "inline_equation", + "content": "\\pmb{y}_{i} \\in \\mathbb{R}^{m}" + }, + { + "bbox": [ + 313, + 309, + 555, + 405 + ], + "type": "text", + "content": " represent the input and output of the layer, and superscript " + }, + { + "bbox": [ + 313, + 309, + 555, + 405 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 309, + 555, + 405 + ], + "type": "text", + "content": " indicates transposition. Thus, adapters in all " + }, + { + "bbox": [ + 313, + 309, + 555, + 405 + ], + "type": "inline_equation", + "content": "(r)" + }, + { + "bbox": [ + 313, + 309, + 555, + 405 + ], + "type": "text", + "content": " cross-attention layers for an attribute category " + }, + { + "bbox": [ + 313, + 309, + 555, + 405 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 309, + 555, + 405 + ], + "type": "text", + "content": " are:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 405, + 413, + 555, + 429 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 413, + 555, + 429 + ], + "spans": [ + { + "bbox": [ + 405, + 413, + 555, + 429 + ], + "type": "interline_equation", + "content": "M _ {d} = \\boldsymbol {Q} ^ {T} \\boldsymbol {P}, \\tag {3}", + "image_path": "730a5a7faf652f343a03f809eb3da1713a5435c1e03fbcdf4e11d5f3434b00c2.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 437, + 555, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 474 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 437, + 555, + 474 + ], + "type": "inline_equation", + "content": "Q = [q_1, q_2, \\ldots, q_r], P = [p_1, p_2, \\ldots, p_r]" + }, + { + "bbox": [ + 313, + 437, + 555, + 474 + ], + "type": "text", + "content": ". Each column vector in " + }, + { + "bbox": [ + 313, + 437, + 555, + 474 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 437, + 555, + 474 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 437, + 555, + 474 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 437, + 555, + 474 + ], + "type": "text", + "content": ", we pad 0 to their end if their dimensions are not the same." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 483, + 554, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 554, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 554, + 506 + ], + "type": "text", + "content": "4.2. Optimizing Adapters via self-discovering process" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 512, + 554, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 512, + 554, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 512, + 554, + 583 + ], + "type": "text", + "content": "One straightforward way to optimize the attached adapters is to collect a unbiased reference dataset as in prior work [30]. However, it is expensive to collect such dataset and the quality of the dataset would also limit the performance. To this end, we propose to train the adapters in a self-discovering manner." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 584, + 554, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 584, + 554, + 644 + ], + "spans": [ + { + "bbox": [ + 313, + 584, + 554, + 644 + ], + "type": "text", + "content": "Given a target group " + }, + { + "bbox": [ + 313, + 584, + 554, + 644 + ], + "type": "inline_equation", + "content": "g_{t}" + }, + { + "bbox": [ + 313, + 584, + 554, + 644 + ], + "type": "text", + "content": " (for example, \"CEO\") and model " + }, + { + "bbox": [ + 313, + 584, + 554, + 644 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 584, + 554, + 644 + ], + "type": "text", + "content": ", we want to optimize the adapters such that the model attached with the adapters generate image " + }, + { + "bbox": [ + 313, + 584, + 554, + 644 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 584, + 554, + 644 + ], + "type": "text", + "content": " towards certain attribute category " + }, + { + "bbox": [ + 313, + 584, + 554, + 644 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 584, + 554, + 644 + ], + "type": "text", + "content": " (for example, \"male\" or \"female\") when conditioned on " + }, + { + "bbox": [ + 313, + 584, + 554, + 644 + ], + "type": "inline_equation", + "content": "g_{t}" + }, + { + "bbox": [ + 313, + 584, + 554, + 644 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 356, + 654, + 553, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 654, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 356, + 654, + 553, + 668 + ], + "type": "interline_equation", + "content": "P _ {\\theta^ {*}} (X \\mid g _ {t}) \\leftarrow P _ {\\theta} (X \\mid g _ {t}) \\left(P _ {\\theta} (d | X)\\right) ^ {\\eta}, \\tag {4}", + "image_path": "91b2c33326a7f4d9c51b5a016661c09ed116495916fc1046842332e829c28300.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "inline_equation", + "content": "P_{\\theta}(X|g_t)" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": " represents the distribution generated by the original model when conditioned on " + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "inline_equation", + "content": "g_{t}" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\theta^{*}" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": " represents the new model equipped with the adapters of " + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": ". Note that" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18414" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 72, + 536, + 232 + ], + "blocks": [ + { + "bbox": [ + 77, + 72, + 536, + 232 + ], + "lines": [ + { + "bbox": [ + 77, + 72, + 536, + 232 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 536, + 232 + ], + "type": "image", + "image_path": "906182312817e1b41230e975d41b8c65ff6a3fa052dc1d8e8d4c4807b699f6ac.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 235, + 555, + 269 + ], + "lines": [ + { + "bbox": [ + 54, + 235, + 555, + 269 + ], + "spans": [ + { + "bbox": [ + 54, + 235, + 555, + 269 + ], + "type": "text", + "content": "Figure 5. Overview of FairGen's inference pipeline. The distribution indicator is generated according the prescribed distribution. Then, it is multiplied by the set of attribute adapters matrices to select attribute matrix adapter. The selected adapter is integrated to the DM with no overhead, guiding the generation towards the prescribed distribution." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 289, + 295, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 289, + 295, + 313 + ], + "spans": [ + { + "bbox": [ + 54, + 289, + 295, + 313 + ], + "type": "inline_equation", + "content": "g_{t}" + }, + { + "bbox": [ + 54, + 289, + 295, + 313 + ], + "type": "text", + "content": " can be set to an empty string \" \" so that the model will be debiased for all possible groups." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 312, + 296, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 296, + 363 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 296, + 363 + ], + "type": "text", + "content": "Applying the Bayes Formula, " + }, + { + "bbox": [ + 55, + 312, + 296, + 363 + ], + "type": "inline_equation", + "content": "P(d|X) = \\frac{P(X|d)P(d)}{P(X)}" + }, + { + "bbox": [ + 55, + 312, + 296, + 363 + ], + "type": "text", + "content": " to Eq. 4, taking logarithm on both sides, we are able to derive that the gradient of the log probability " + }, + { + "bbox": [ + 55, + 312, + 296, + 363 + ], + "type": "inline_equation", + "content": "\\nabla \\log P_{\\theta^*}(X|g_t)" + }, + { + "bbox": [ + 55, + 312, + 296, + 363 + ], + "type": "text", + "content": " would be proportional to:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 371, + 296, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 371, + 296, + 397 + ], + "spans": [ + { + "bbox": [ + 61, + 371, + 296, + 397 + ], + "type": "interline_equation", + "content": "\\nabla \\log P _ {\\theta} (X | g _ {t}) + \\eta (\\nabla \\log P _ {\\theta} (X | d) - \\nabla \\log P _ {\\theta} (X | g _ {t})) \\tag {5}", + "image_path": "a7d586b9c24bd47f3ee47bb1e0f4f0d46e01ae2598ac554fb75dde61d1ff89bf.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 404, + 296, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 404, + 296, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 404, + 296, + 477 + ], + "type": "text", + "content": "Based on Tweedie's formula [4] and the reparametrization trick of Classifier-free guidance [9], we introduce a time-varying noising process and represent each score (gradient of log probability) as a denoising prediction " + }, + { + "bbox": [ + 55, + 404, + 296, + 477 + ], + "type": "inline_equation", + "content": "\\epsilon(X_{t}, c_{t}, t)" + }, + { + "bbox": [ + 55, + 404, + 296, + 477 + ], + "type": "text", + "content": ", which leads to our learning objective for adapters of category " + }, + { + "bbox": [ + 55, + 404, + 296, + 477 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 404, + 296, + 477 + ], + "type": "text", + "content": " of an attribute:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 485, + 313, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 313, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 313, + 510 + ], + "type": "interline_equation", + "content": "\\epsilon_ {\\theta^ {*}} \\left(X _ {t}, g _ {t}, t\\right) \\leftarrow \\epsilon_ {\\theta} \\left(X _ {t}, g _ {t}, t\\right) + \\eta \\left(\\epsilon_ {\\theta} \\left(X _ {t}, d, t\\right) - \\epsilon_ {\\theta} \\left(X _ {t}, g _ {t}, t\\right)\\right). \\tag {6}", + "image_path": "aef57f2ea71770bc7279397e22ae424b13e925630a882fb9b7726565b2824f55.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 517, + 295, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 295, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 295, + 541 + ], + "type": "text", + "content": "Therefore, the guidance loss for optimizing adapters of category " + }, + { + "bbox": [ + 55, + 517, + 295, + 541 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 517, + 295, + 541 + ], + "type": "text", + "content": " can be defined as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 99, + 549, + 277, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 549, + 277, + 572 + ], + "spans": [ + { + "bbox": [ + 99, + 549, + 277, + 572 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {G u i d a n c e}} = \\mathbb {L} _ {x _ {t}, t} \\left[ \\| \\epsilon_ {\\theta^ {*}} (X _ {t}, g _ {t}, t) - \\epsilon_ {L} \\| ^ {2} \\right],", + "image_path": "c9f702e338ae8a6381e436abe68ad11a6113a05a2477520cd4f00e9b8ef2db4f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 572, + 277, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 572, + 277, + 586 + ], + "spans": [ + { + "bbox": [ + 62, + 572, + 277, + 586 + ], + "type": "interline_equation", + "content": "\\epsilon_ {L} = \\epsilon_ {\\theta} (X _ {t}, g _ {t}, t) + \\eta (\\epsilon_ {\\theta} (X _ {t}, d, t) - \\epsilon_ {\\theta} (X _ {t}, g _ {t}, t)),", + "image_path": "528f9e7d56f3b5728d2d9c93c6be3413c268c1a226902a313890ceefddfbe9c3.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 593, + 296, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 593, + 296, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 593, + 296, + 652 + ], + "type": "text", + "content": "where the goal is to align the noise " + }, + { + "bbox": [ + 55, + 593, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta^{*}}(X_{t},g_{t},t)" + }, + { + "bbox": [ + 55, + 593, + 296, + 652 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 55, + 593, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\theta^{*}" + }, + { + "bbox": [ + 55, + 593, + 296, + 652 + ], + "type": "text", + "content": " (DM with adapters of category " + }, + { + "bbox": [ + 55, + 593, + 296, + 652 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 593, + 296, + 652 + ], + "type": "text", + "content": " ) and the noise composition " + }, + { + "bbox": [ + 55, + 593, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\epsilon_{L}" + }, + { + "bbox": [ + 55, + 593, + 296, + 652 + ], + "type": "text", + "content": " of group " + }, + { + "bbox": [ + 55, + 593, + 296, + 652 + ], + "type": "inline_equation", + "content": "g_{t}" + }, + { + "bbox": [ + 55, + 593, + 296, + 652 + ], + "type": "text", + "content": " and category " + }, + { + "bbox": [ + 55, + 593, + 296, + 652 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 593, + 296, + 652 + ], + "type": "text", + "content": " from the frozen DM " + }, + { + "bbox": [ + 55, + 593, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 593, + 296, + 652 + ], + "type": "text", + "content": " And hence, our adapters optimization does not require any additional data." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 660, + 251, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 251, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 251, + 672 + ], + "type": "text", + "content": "4.3. Inference with distribution indicator." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "After the optimization, we obtain a set of adapters " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = \\{M_1,M_2,\\dots ,M_t\\}" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " categories of a given attribute. For instance, " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "t = 2" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " for gender bias and " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "t = 4" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " for racial bias" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 289, + 555, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 555, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 555, + 337 + ], + "type": "text", + "content": "in our case. As shown in Figure 5, at the inference stage, we introduce a distribution indicator " + }, + { + "bbox": [ + 313, + 289, + 555, + 337 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 289, + 555, + 337 + ], + "type": "text", + "content": ". Given a prescribed distribution " + }, + { + "bbox": [ + 313, + 289, + 555, + 337 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{a}" + }, + { + "bbox": [ + 313, + 289, + 555, + 337 + ], + "type": "text", + "content": " (e.g., uniform distribution), we define its Probabilistic Mass Function (PMF) as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 324, + 350, + 555, + 382 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 350, + 555, + 382 + ], + "spans": [ + { + "bbox": [ + 324, + 350, + 555, + 382 + ], + "type": "interline_equation", + "content": "P (X = x) = f _ {\\theta} ^ {a} (x) = \\left\\{ \\begin{array}{l l} \\frac {1}{t} & \\text {f o r} x \\in \\{1, 2, \\dots , t \\}, \\\\ 0 & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {8}", + "image_path": "3cae7ce7aeb9ffd06bc6353e131143f5bba8b3aa8e8cc0ba42b8e89f6e339700.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 394, + 554, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 394, + 554, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 554, + 466 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 394, + 554, + 466 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 394, + 554, + 466 + ], + "type": "text", + "content": " is the number of possible categories for each attribute, and " + }, + { + "bbox": [ + 313, + 394, + 554, + 466 + ], + "type": "inline_equation", + "content": "\\{1,2,\\dots ,t\\}" + }, + { + "bbox": [ + 313, + 394, + 554, + 466 + ], + "type": "text", + "content": " represents the set of all possible values for the random variable " + }, + { + "bbox": [ + 313, + 394, + 554, + 466 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 394, + 554, + 466 + ], + "type": "text", + "content": ". The parameter " + }, + { + "bbox": [ + 313, + 394, + 554, + 466 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 394, + 554, + 466 + ], + "type": "text", + "content": " controls the shape of the prescribed distribution, and in the case of the uniform distribution, " + }, + { + "bbox": [ + 313, + 394, + 554, + 466 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 394, + 554, + 466 + ], + "type": "text", + "content": " implies that all outcomes in the set have equal probability, i.e., " + }, + { + "bbox": [ + 313, + 394, + 554, + 466 + ], + "type": "inline_equation", + "content": "\\frac{1}{t}" + }, + { + "bbox": [ + 313, + 394, + 554, + 466 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 468, + 554, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 554, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 554, + 503 + ], + "type": "text", + "content": "Then, we randomly sample an index " + }, + { + "bbox": [ + 313, + 468, + 554, + 503 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 468, + 554, + 503 + ], + "type": "text", + "content": " from the prescribed distribution " + }, + { + "bbox": [ + 313, + 468, + 554, + 503 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{a}" + }, + { + "bbox": [ + 313, + 468, + 554, + 503 + ], + "type": "text", + "content": ". The distribution indicator " + }, + { + "bbox": [ + 313, + 468, + 554, + 503 + ], + "type": "inline_equation", + "content": "\\pmb{h} \\in \\mathbb{R}^t" + }, + { + "bbox": [ + 313, + 468, + 554, + 503 + ], + "type": "text", + "content": " is formulated to reflect the chosen index " + }, + { + "bbox": [ + 313, + 468, + 554, + 503 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 468, + 554, + 503 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 390, + 521, + 555, + 553 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 521, + 555, + 553 + ], + "spans": [ + { + "bbox": [ + 390, + 521, + 555, + 553 + ], + "type": "interline_equation", + "content": "\\boldsymbol {h} _ {i} := \\left\\{ \\begin{array}{l l} 1 & \\text {i f} i = k, \\\\ 0 & \\text {i f} i \\neq k, \\end{array} \\right. \\tag {9}", + "image_path": "3955dd6d26d22688328bc1d2410eb6b90e1553e43f45e6bd36b12dae5957edce.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 565, + 554, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 565, + 554, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 565, + 554, + 650 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 565, + 554, + 650 + ], + "type": "inline_equation", + "content": "i \\in \\{1,2,\\dots,t\\}" + }, + { + "bbox": [ + 313, + 565, + 554, + 650 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 565, + 554, + 650 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 565, + 554, + 650 + ], + "type": "text", + "content": " is the sampled index based on the distribution " + }, + { + "bbox": [ + 313, + 565, + 554, + 650 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{a}" + }, + { + "bbox": [ + 313, + 565, + 554, + 650 + ], + "type": "text", + "content": ". The indicator " + }, + { + "bbox": [ + 313, + 565, + 554, + 650 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 565, + 554, + 650 + ], + "type": "text", + "content": " is a one-hot vector, where the " + }, + { + "bbox": [ + 313, + 565, + 554, + 650 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 565, + 554, + 650 + ], + "type": "text", + "content": "-th element is 1, indicating the sampled outcome, and all other elements are 0. After obtaining the distribution indicator, it is multiplied by the set of trained attribute matrix adapters. Subsequently, the final weight change " + }, + { + "bbox": [ + 313, + 565, + 554, + 650 + ], + "type": "inline_equation", + "content": "\\Delta W" + }, + { + "bbox": [ + 313, + 565, + 554, + 650 + ], + "type": "text", + "content": " is given by:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 401, + 663, + 555, + 676 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 401, + 663, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 401, + 663, + 555, + 676 + ], + "type": "interline_equation", + "content": "\\Delta \\boldsymbol {W} = \\boldsymbol {h} \\cdot \\mathcal {M}, \\tag {10}", + "image_path": "c74eadeacb4971579a9c4d6cf201a9a0fcb8c321ae629833d924b1a644d5e51c.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "and the model is updated as " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\gets \\mathbf{W} + \\alpha \\Delta \\mathbf{W}" + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": " is a scaling factor controlling the strength of the guidance." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18415" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 71, + 532, + 217 + ], + "blocks": [ + { + "bbox": [ + 77, + 71, + 532, + 217 + ], + "lines": [ + { + "bbox": [ + 77, + 71, + 532, + 217 + ], + "spans": [ + { + "bbox": [ + 77, + 71, + 532, + 217 + ], + "type": "image", + "image_path": "b978c236a569da3c60a08d20df040305b0ac45295bf23a9ad236928d06ae5e4a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 218, + 553, + 239 + ], + "lines": [ + { + "bbox": [ + 55, + 218, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 218, + 553, + 239 + ], + "type": "text", + "content": "Figure 6. Images generated from the original Stable Diffusion model (left) and FairGen (right) for gender debiasing with the prompt \"A photo of a worker\". The gender distribution shifts from a male-to-female ratio of 13:2 in the original model to 8:7 with FairGen." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 260, + 294, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 260, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 260, + 294, + 285 + ], + "type": "text", + "content": "4.4. Debiasing Multiple Attributes (intersectional debiasing)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 290, + 296, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 290, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 290, + 296, + 399 + ], + "type": "text", + "content": "Our method can be inherently extended to debiasing multiple attributes in diffusion models (DM). Specifically, in the case of multiple attribute debiasing, the adapter for each attribute should not interfere with the others. Otherwise, the most recently trained adapter could degrade the performance of previously learned adapters. For each attribute to be debiased, we denote a set of adapter parameters as " + }, + { + "bbox": [ + 55, + 290, + 296, + 399 + ], + "type": "inline_equation", + "content": "\\{P_t, Q_t\\}" + }, + { + "bbox": [ + 55, + 290, + 296, + 399 + ], + "type": "text", + "content": ". We have " + }, + { + "bbox": [ + 55, + 290, + 296, + 399 + ], + "type": "inline_equation", + "content": "P_t = [p_t^1, p_t^2, \\dots, p_t^r]" + }, + { + "bbox": [ + 55, + 290, + 296, + 399 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 290, + 296, + 399 + ], + "type": "inline_equation", + "content": "Q_t = [q_t^1, q_t^2, \\dots, q_t^r]" + }, + { + "bbox": [ + 55, + 290, + 296, + 399 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 399, + 296, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 296, + 448 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 296, + 448 + ], + "type": "text", + "content": "To avoid interference between attribute adapters, we extend Eq. 7 by introducing an orthogonal regularization loss that regularizes the vector subspace spanned by each " + }, + { + "bbox": [ + 55, + 399, + 296, + 448 + ], + "type": "inline_equation", + "content": "P_{t}" + }, + { + "bbox": [ + 55, + 399, + 296, + 448 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 399, + 296, + 448 + ], + "type": "inline_equation", + "content": "Q_{t}" + }, + { + "bbox": [ + 55, + 399, + 296, + 448 + ], + "type": "text", + "content": " to be orthogonal to each other:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 99, + 451, + 295, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 451, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 99, + 451, + 295, + 483 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {o r t h}} = \\sum_ {i = 1} ^ {t - 1} \\left(\\boldsymbol {P} _ {i} \\times \\boldsymbol {P} _ {t} + \\boldsymbol {Q} _ {i} \\times \\boldsymbol {Q} _ {t}\\right), \\tag {11}", + "image_path": "4f97d4e1380df76b8631c9f68389f88a8d88fdb2648e643fe27e34b9c5c5971d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 123, + 492, + 294, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 492, + 294, + 506 + ], + "spans": [ + { + "bbox": [ + 123, + 492, + 294, + 506 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {G u i d a n c e}} + \\gamma \\cdot \\mathcal {L} _ {\\text {o r t h}}. \\tag {12}", + "image_path": "d5f2f6d009c2d41af2c91616bce1c601d9059869581293c95aceb1b27858f468.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 529, + 132, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 132, + 542 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 132, + 542 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 549, + 296, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 549, + 296, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 549, + 296, + 597 + ], + "type": "text", + "content": "In this section, we first describe the experimental details and evaluation metrics, then present a quantitative and qualitative analysis of our method. Please refer to Appendix for more results." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 603, + 178, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 603, + 178, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 603, + 178, + 615 + ], + "type": "text", + "content": "5.1. Experimental Details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "Implementation details. We use Stable Diffusion v2.1 for all methods. We employ the prompt template \"a photo of the face of a {occupation}, a person\". At inference time, for each bias, we generate 100 images per occupation across 100 occupations, resulting in a total of 10,000 images. We set " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\eta = \\alpha = 0.3" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": ", and train for 1000 iterations with a learning rate of 1e-5. For gender bias, we use the CelebA [12] dataset to train a binary classifier with two" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 261, + 554, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 261, + 554, + 321 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 554, + 321 + ], + "type": "text", + "content": "categories: {male, female}. For racial bias, we use the Fair-Face [10] dataset to train a classifier with the following four categories: WMELH = {White, Middle Eastern, Latino Hispanic}, Asian = {East Asian, Southeast Asian}, Black, and Indian. Please refer to the supplementary for more details." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 338, + 555, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 555, + 457 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 555, + 457 + ], + "type": "text", + "content": "Compared methods. In this work, we compare our method with recent state-of-the-art (SOTA) methods, including two retraining-based methods, Finetuning for Fairness (F4Fair) [30] and Inclusive Text-to-Image Generation (ITI-GEN) [37], an approach that does not require retraining DM, H-Distribution Guidance (H Guidance) [18], a closed-form editing approach, Unified Concept Editing (UCE) [8] and an h-space manipulation approach, Interpretable Diffusion [11]. Please refer to Appendix for more detailed introduction of these methods." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 469, + 426, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 469, + 426, + 480 + ], + "spans": [ + { + "bbox": [ + 314, + 469, + 426, + 480 + ], + "type": "text", + "content": "5.2. Evaluation Metrics" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 487, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 482, + 498 + ], + "type": "text", + "content": "We evaluate all methods in three metrics:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 502, + 554, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 554, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 554, + 598 + ], + "type": "text", + "content": "Fairness Discrepancy (FD). Following prior work[18], we adopt the Fairness Discrepancy (FD) metric. For an attribute " + }, + { + "bbox": [ + 313, + 502, + 554, + 598 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 313, + 502, + 554, + 598 + ], + "type": "text", + "content": " and target distribution " + }, + { + "bbox": [ + 313, + 502, + 554, + 598 + ], + "type": "inline_equation", + "content": "p_{\\theta}^{a}" + }, + { + "bbox": [ + 313, + 502, + 554, + 598 + ], + "type": "text", + "content": ", we use a high-accuracy classifier " + }, + { + "bbox": [ + 313, + 502, + 554, + 598 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_a" + }, + { + "bbox": [ + 313, + 502, + 554, + 598 + ], + "type": "text", + "content": " to compute the fairness performance: " + }, + { + "bbox": [ + 313, + 502, + 554, + 598 + ], + "type": "inline_equation", + "content": "||p_{\\theta}^{a} - \\mathbb{E}_{\\mathbf{x}\\sim p_{\\theta}(\\mathbf{x})}(\\mathbf{y})||_2" + }, + { + "bbox": [ + 313, + 502, + 554, + 598 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 502, + 554, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 313, + 502, + 554, + 598 + ], + "type": "text", + "content": " is the softmax output of " + }, + { + "bbox": [ + 313, + 502, + 554, + 598 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_a(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 502, + 554, + 598 + ], + "type": "text", + "content": ". The target distribution " + }, + { + "bbox": [ + 313, + 502, + 554, + 598 + ], + "type": "inline_equation", + "content": "p_{\\theta}^{a}" + }, + { + "bbox": [ + 313, + 502, + 554, + 598 + ], + "type": "text", + "content": " can be any user-defined vector, typically uniform. A lower FD score indicates a closer match to the target distribution." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 601, + 554, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 601, + 554, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 601, + 554, + 662 + ], + "type": "inline_equation", + "content": "\\mathbf{CLIP}_{\\mathrm{sim}}" + }, + { + "bbox": [ + 313, + 601, + 554, + 662 + ], + "type": "text", + "content": ". Besides fairness, the debiased model should maintain the ability to generate images that are semantically close to their text prompts. Therefore, following [30], we report the CLIP similarity score " + }, + { + "bbox": [ + 313, + 601, + 554, + 662 + ], + "type": "inline_equation", + "content": "\\mathbf{CLIP}_{\\mathrm{sim}}" + }, + { + "bbox": [ + 313, + 601, + 554, + 662 + ], + "type": "text", + "content": " between the generated image and its prompt." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": "BRISQUE. The generated image quality is also important as the debiasing process should not influence the image generation ability. We use the BRISQUE metric for evaluate the quality of the generated images." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18416" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 76, + 71, + 293, + 209 + ], + "blocks": [ + { + "bbox": [ + 76, + 71, + 293, + 209 + ], + "lines": [ + { + "bbox": [ + 76, + 71, + 293, + 209 + ], + "spans": [ + { + "bbox": [ + 76, + 71, + 293, + 209 + ], + "type": "image", + "image_path": "f2805a736552659f68ca2a66baf6ac651c7b92f3064bb42ec86d667c73cff5a6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 163, + 210, + 209, + 220 + ], + "lines": [ + { + "bbox": [ + 163, + 210, + 209, + 220 + ], + "spans": [ + { + "bbox": [ + 163, + 210, + 209, + 220 + ], + "type": "text", + "content": "(a) original" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 318, + 72, + 535, + 209 + ], + "blocks": [ + { + "bbox": [ + 318, + 72, + 535, + 209 + ], + "lines": [ + { + "bbox": [ + 318, + 72, + 535, + 209 + ], + "spans": [ + { + "bbox": [ + 318, + 72, + 535, + 209 + ], + "type": "image", + "image_path": "ad12b6be35af0bdd06443189bd11d335fd8d62b67208a9c383b9fca1f5116666.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 210, + 466, + 220 + ], + "lines": [ + { + "bbox": [ + 388, + 210, + 466, + 220 + ], + "spans": [ + { + "bbox": [ + 388, + 210, + 466, + 220 + ], + "type": "text", + "content": "(b) debiased (Ours)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 221, + 555, + 254 + ], + "lines": [ + { + "bbox": [ + 54, + 221, + 555, + 254 + ], + "spans": [ + { + "bbox": [ + 54, + 221, + 555, + 254 + ], + "type": "text", + "content": "Figure 7. Images generated from the original Stable Diffusion model (left) and FairGen (right) for racial debiasing with the prompt \"A photo of a sportsman\". The racial group distribution shifts from WMELH : Asian : Black : Indian = 7:2:5:1 in the original model to 4:4:4:3 with FairGen." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 304, + 276, + 414 + ], + "blocks": [ + { + "bbox": [ + 55, + 259, + 296, + 304 + ], + "lines": [ + { + "bbox": [ + 55, + 259, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 259, + 296, + 304 + ], + "type": "text", + "content": "Table 1. Comparisons of our method to the SOTA methods in gender bias over two predefined distributions: " + }, + { + "bbox": [ + 55, + 259, + 296, + 304 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{1} = (0.5, 0.5)" + }, + { + "bbox": [ + 55, + 259, + 296, + 304 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 259, + 296, + 304 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{2} = (0.2, 0.8)" + }, + { + "bbox": [ + 55, + 259, + 296, + 304 + ], + "type": "text", + "content": ", representing the probability of male and female respectively." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 304, + 276, + 414 + ], + "lines": [ + { + "bbox": [ + 58, + 304, + 276, + 414 + ], + "spans": [ + { + "bbox": [ + 58, + 304, + 276, + 414 + ], + "type": "table", + "html": "
MethodFD ↓CLIPsim ↑BRISQUE ↑
f1θf2θf1θf2θf1θf2θ
Original SD0.4240.8470.380.3838.6538.69
F4Fair0.1650.3870.360.3538.2137.64
ITI-GEN0.0930.4720.320.3337.6237.96
H Guidance0.1180.3980.310.3238.5438.65
UCE0.2840.5360.360.2937.1236.54
InterDiff0.1770.3860.350.3438.3238.04
Ours0.0030.0050.380.3838.4638.72
", + "image_path": "c0dbcb810c4c4ffd5a8065c6fda40a940dd595d6a418621bd2367e7383e33292.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 304, + 535, + 414 + ], + "blocks": [ + { + "bbox": [ + 313, + 259, + 555, + 304 + ], + "lines": [ + { + "bbox": [ + 313, + 259, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 259, + 555, + 304 + ], + "type": "text", + "content": "Table 2. Comparisons of our method to the SOTA methods in racial bias over two distributions: " + }, + { + "bbox": [ + 313, + 259, + 555, + 304 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{1} = (0.25, 0.25, 0.25, 0.25)" + }, + { + "bbox": [ + 313, + 259, + 555, + 304 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 259, + 555, + 304 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{2} = (0.4, 0.3, 0.2, 0.1)" + }, + { + "bbox": [ + 313, + 259, + 555, + 304 + ], + "type": "text", + "content": ", representing the probability of WMELH, Asian, Black, and Indian respectively." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 304, + 535, + 414 + ], + "lines": [ + { + "bbox": [ + 317, + 304, + 535, + 414 + ], + "spans": [ + { + "bbox": [ + 317, + 304, + 535, + 414 + ], + "type": "table", + "html": "
MethodFD ↓CLIPsim ↑BRISQUE ↑
f1θf2θf1θf2θf1θf2θ
Original SD0.3840.4970.460.4138.9438.65
F4Fair0.2200.3050.430.4038.8038.45
ITI-GEN0.1310.5420.310.3136.6137.09
H Guidance0.1500.2850.410.3738.6538.60
UCE0.2900.4600.450.3937.9037.80
InterDiff0.1500.3910.400.3837.6437.82
Ours0.0950.1500.460.4138.9638.66
", + "image_path": "9c6e847e259b8b56458b4d9c4ff228cf17b8c1f2b2b417d216f1aad281c0aace.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 433, + 113, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 113, + 444 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 113, + 444 + ], + "type": "text", + "content": "5.3. Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 448, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 448, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 54, + 448, + 295, + 651 + ], + "type": "text", + "content": "Comparisons in gender debiasing. Table 1 demonstrates that our method outperforms others in mitigating gender bias over two predefined distributions: " + }, + { + "bbox": [ + 54, + 448, + 295, + 651 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{1} = (0.5,0.5)" + }, + { + "bbox": [ + 54, + 448, + 295, + 651 + ], + "type": "text", + "content": ", representing equal likelihood of male and female, and " + }, + { + "bbox": [ + 54, + 448, + 295, + 651 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{2} = (0.2,0.8)" + }, + { + "bbox": [ + 54, + 448, + 295, + 651 + ], + "type": "text", + "content": ", where male and female have a " + }, + { + "bbox": [ + 54, + 448, + 295, + 651 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 54, + 448, + 295, + 651 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 448, + 295, + 651 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 54, + 448, + 295, + 651 + ], + "type": "text", + "content": " probability, respectively. Original SD model exhibits high FD scores (0.424 at " + }, + { + "bbox": [ + 54, + 448, + 295, + 651 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{1}" + }, + { + "bbox": [ + 54, + 448, + 295, + 651 + ], + "type": "text", + "content": " and 0.847 at " + }, + { + "bbox": [ + 54, + 448, + 295, + 651 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{2}" + }, + { + "bbox": [ + 54, + 448, + 295, + 651 + ], + "type": "text", + "content": "), indicating significant bias towards one gender. While previous methods like F4Fair, ITI-GEN, H Guidance, and UCE reduce bias to some extent, our method achieves the most reduction, with FD scores of 0.003 at " + }, + { + "bbox": [ + 54, + 448, + 295, + 651 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{1}" + }, + { + "bbox": [ + 54, + 448, + 295, + 651 + ], + "type": "text", + "content": " and 0.005 at " + }, + { + "bbox": [ + 54, + 448, + 295, + 651 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{2}" + }, + { + "bbox": [ + 54, + 448, + 295, + 651 + ], + "type": "text", + "content": ". Additionally, our method preserves semantic similarity (CLIPsim of 0.38) and image quality (BRISQUE scores of 38.46 and 38.72), matching or slightly surpassing the Original SD model. The qualitative results in Figure 6 further demonstrate that our approach effectively mitigates gender bias without compromising image quality or semantic coherence." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "Comparisons in racial debiasing. Table 2 compares methods over two distributions: " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{1} = (0.25, 0.25, 0.25, 0.25)" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": ", representing equal probabilities for WMELH, Asian, Black, and Indian, and " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{2} = (0.4, 0.3, 0.2, 0.1)" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": ", with higher probabilities for" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 434, + 555, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 555, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 555, + 481 + ], + "type": "text", + "content": "WMELH and Asian, and lower probabilities for Black and Indian. Original SD model shows significant racial bias, with FD scores of 0.384 for " + }, + { + "bbox": [ + 313, + 434, + 555, + 481 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{1}" + }, + { + "bbox": [ + 313, + 434, + 555, + 481 + ], + "type": "text", + "content": " and 0.497 for " + }, + { + "bbox": [ + 313, + 434, + 555, + 481 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{2}" + }, + { + "bbox": [ + 313, + 434, + 555, + 481 + ], + "type": "text", + "content": ", indicating poor calibration to either distribution." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 482, + 555, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 555, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 555, + 639 + ], + "type": "text", + "content": "Debiasing methods such as F4Fair, H Guidance, and UCE reduce this bias, with H Guidance achieving relatively lower FD scores. However, our method performs the best, reducing bias to 0.095 for " + }, + { + "bbox": [ + 313, + 482, + 555, + 639 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{1}" + }, + { + "bbox": [ + 313, + 482, + 555, + 639 + ], + "type": "text", + "content": " and 0.150 for " + }, + { + "bbox": [ + 313, + 482, + 555, + 639 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{2}" + }, + { + "bbox": [ + 313, + 482, + 555, + 639 + ], + "type": "text", + "content": ". In terms of semantic similarity, the Original SD model sets a strong baseline (0.46 for " + }, + { + "bbox": [ + 313, + 482, + 555, + 639 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{1}" + }, + { + "bbox": [ + 313, + 482, + 555, + 639 + ], + "type": "text", + "content": " and 0.41 for " + }, + { + "bbox": [ + 313, + 482, + 555, + 639 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{2}" + }, + { + "bbox": [ + 313, + 482, + 555, + 639 + ], + "type": "text", + "content": "), and our method maintains this high alignment, ensuring that debiasing does not impair semantic accuracy. Regarding image quality, our approach slightly improves upon the Original SD model, achieving the highest scores (39.10 for " + }, + { + "bbox": [ + 313, + 482, + 555, + 639 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{1}" + }, + { + "bbox": [ + 313, + 482, + 555, + 639 + ], + "type": "text", + "content": " and 38.90 for " + }, + { + "bbox": [ + 313, + 482, + 555, + 639 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{2}" + }, + { + "bbox": [ + 313, + 482, + 555, + 639 + ], + "type": "text", + "content": "). Again, the qualitative results in Figure 7 further verify that our method outperforms others in reducing racial bias while preserving both semantic similarity and image quality." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": "Comparisons in intersectional debiasing We also consider a more complex challenge of intersectional debiasing, i.e., jointly debiasing both gender and racial biases. The target distribution " + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": " is set to be uniform for both gender and racial bias. Table 4 shows that Original SD model exhibits significant bias with an FD score of 0.214. While F4Fair" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18417" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 157, + 83, + 454, + 156 + ], + "blocks": [ + { + "bbox": [ + 185, + 70, + 424, + 81 + ], + "lines": [ + { + "bbox": [ + 185, + 70, + 424, + 81 + ], + "spans": [ + { + "bbox": [ + 185, + 70, + 424, + 81 + ], + "type": "text", + "content": "Table 3. Transferability of proposed method across different DMs" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 157, + 83, + 454, + 156 + ], + "lines": [ + { + "bbox": [ + 157, + 83, + 454, + 156 + ], + "spans": [ + { + "bbox": [ + 157, + 83, + 454, + 156 + ], + "type": "table", + "html": "
TrainTestGenderRacial
FD ↓CLIPsim ↑BRISQUE ↑FD ↓CLIPsim ↑BRISQUE ↑
v2.1v1.40.0060.3738.400.0970.4538.90
v2.1v1.50.0060.3738.420.0960.4538.92
v2.1v2.00.0080.3838.450.0950.4638.95
v2.1v2.10.0030.3838.460.0950.4638.96
", + "image_path": "183813f32b6d7ae60153c55868940445229fb2bca8d2059f99d478f17b28494f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 83, + 186, + 269, + 286 + ], + "blocks": [ + { + "bbox": [ + 80, + 174, + 271, + 185 + ], + "lines": [ + { + "bbox": [ + 80, + 174, + 271, + 185 + ], + "spans": [ + { + "bbox": [ + 80, + 174, + 271, + 185 + ], + "type": "text", + "content": "Table 4. Evaluation of mitigating intersectional bias." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 83, + 186, + 269, + 286 + ], + "lines": [ + { + "bbox": [ + 83, + 186, + 269, + 286 + ], + "spans": [ + { + "bbox": [ + 83, + 186, + 269, + 286 + ], + "type": "table", + "html": "
MethodFD ↓CLIPsim↑BRISQUE ↑
Original SD0.2140.3539.24
F4Fair0.1450.3638.90
ITI-GEN0.0980.3238.01
H Guidance0.1300.3338.75
UCE0.1800.3438.60
InterDiff0.1580.3538.67
Ours0.0470.3639.50
", + "image_path": "d358588df785b2a6ad696597e66a87279ca8e2d6a8b84eebbf02ac3344ca6ebf.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 306, + 295, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 306, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 306, + 295, + 342 + ], + "type": "text", + "content": "and H Guidance reduce this bias to 0.145 and 0.130, respectively, our method achieves a much lower FD score of 0.047, reflecting a substantial improvement in fairness." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 342, + 295, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 342, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 54, + 342, + 295, + 474 + ], + "type": "text", + "content": "For semantic similarity, the Original SD scores 0.35, and F4Fair slightly improves it to 0.36. Our method matches this performance, maintaining semantic coherence while reducing bias. Regarding image quality, the Original SD has a BRISQUE score of 39.24, while our method improves it to 39.50, indicating enhanced perceptual quality. These results demonstrate that our method excels at jointly debiasing both gender and racial biases, significantly reducing bias without sacrificing semantic accuracy or image quality. This highlights its robustness and practicality in real-world settings where multiple biases are present." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 475, + 295, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 295, + 619 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 295, + 619 + ], + "type": "text", + "content": "Transferability across different DMs. We further evaluate the transferability of our method by training it with Stable Diffusion v2.1 and testing it over other versions, and results are reported in Table 3. From the table we can see that the performance of in all metrics are close to the optimal results achieved when training and testing are performed using the same model version (v2.1). While there is a slight increase in FD when tested on v1.4, v1.5, and v2.0, the differences are minor, indicating that our method can effectively generalize across different model versions. This robustness highlights the method's capability to transfer learned features across varying model conditions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 627, + 149, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 627, + 149, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 627, + 149, + 639 + ], + "type": "text", + "content": "5.4. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "content": "Attaching adapters in U-Net. The results in Table 5 illustrate the impact of attaching adapters to different layers of the U-Net in the DM. When adapters are attached to all layers, the fairness score (FD) is 0.165, with a CLIP similarity score of 0.33 and a BRISQUE score of 35.00. The best results are obtained when adapters are attached only to" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 337, + 186, + 533, + 240 + ], + "blocks": [ + { + "bbox": [ + 336, + 174, + 531, + 186 + ], + "lines": [ + { + "bbox": [ + 336, + 174, + 531, + 186 + ], + "spans": [ + { + "bbox": [ + 336, + 174, + 531, + 186 + ], + "type": "text", + "content": "Table 5. Ablation of layers that are attached adapters." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 337, + 186, + 533, + 240 + ], + "lines": [ + { + "bbox": [ + 337, + 186, + 533, + 240 + ], + "spans": [ + { + "bbox": [ + 337, + 186, + 533, + 240 + ], + "type": "table", + "html": "
LocationFD ↓CLIPsim ↑BRISQUE ↑
All layers0.1650.3335.00
Non-CA layers0.1580.3437.00
CA layers0.0470.3639.50
", + "image_path": "1988f87967e91b293fe802a5d007d7b0fe2874892afd4098c8d50c0ffb9a463f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 262, + 555, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 555, + 334 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 555, + 334 + ], + "type": "text", + "content": "the CA layers. This configuration significantly reduces the FD score to 0.047, increases semantic similarity " + }, + { + "bbox": [ + 313, + 262, + 555, + 334 + ], + "type": "inline_equation", + "content": "(\\mathrm{CLIP}_{\\mathrm{sim}} = 0.36)" + }, + { + "bbox": [ + 313, + 262, + 555, + 334 + ], + "type": "text", + "content": ", and enhances image quality with a BRISQUE score of 39.50. These highlights that focusing the adapters on the cross-attention layers leads to the most substantial improvements in both fairness and image quality." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 337, + 556, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 337, + 556, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 556, + 458 + ], + "type": "text", + "content": "Impact of orthogonal regularization (OR). Table 6 demonstrates the impact of orthogonal regularization (OR). Without OR, the FD score is 0.143, semantic similarity " + }, + { + "bbox": [ + 313, + 337, + 556, + 458 + ], + "type": "inline_equation", + "content": "(\\mathrm{CLIP}_{\\mathrm{sim}})" + }, + { + "bbox": [ + 313, + 337, + 556, + 458 + ], + "type": "text", + "content": " drops to 0.29, and image quality (BRISQUE) is 37.92. When OR is applied, performance improves significantly across all metrics, with a much lower FD score of 0.047, higher semantic similarity of 0.36, and improved image quality (BRISQUE = 39.50). This highlights the effectiveness of orthogonal regularization in reducing bias and improving overall performance." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 337, + 481, + 533, + 525 + ], + "blocks": [ + { + "bbox": [ + 379, + 469, + 488, + 480 + ], + "lines": [ + { + "bbox": [ + 379, + 469, + 488, + 480 + ], + "spans": [ + { + "bbox": [ + 379, + 469, + 488, + 480 + ], + "type": "text", + "content": "Table 6. Orthogonal Ablation" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 337, + 481, + 533, + 525 + ], + "lines": [ + { + "bbox": [ + 337, + 481, + 533, + 525 + ], + "spans": [ + { + "bbox": [ + 337, + 481, + 533, + 525 + ], + "type": "table", + "html": "
LocationFD ↓CLIPsim ↑BRISQUE ↑
Ours w\\o OR0.1430.2937.92
Ours w\\OR0.0470.3639.50
", + "image_path": "5bb5d3ab70c05f3e35f407642149a4953e1d9581778cccdf3a4aa4b3494646f0.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 552, + 388, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 388, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 388, + 563 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we propose FairGen, a plug-and-play method that learns attribute latent directions in a self-discovering manner. Our method can not only jointly debias multiple attributes in DMs, but also enables the generated images to follow a prescribed attribute distribution. It is lightweight and can be integrated with other DMs without retraining. Extensive experiments on debiasing gender, racial, and their intersectional biases show that our method outperforms previous SOTA by a large margin. We believe that our work marks a critical advancement in addressing harmful societal stereotypes within diffusion models, and it contributes to the ethical real-world applications of text-to-image DMs." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18418" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 145 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant No. 62306261, and the Shun Hing Institute of Advanced Engineering (SHIAE) Fund under Grant No. 8115074." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 156, + 115, + 169 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 156, + 115, + 169 + ], + "spans": [ + { + "bbox": [ + 56, + 156, + 115, + 169 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 176, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 176, + 296, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 176, + 296, + 253 + ], + "spans": [ + { + "bbox": [ + 61, + 176, + 296, + 253 + ], + "type": "text", + "content": "[1] Federico Bianchi, Pratyusha Kalluri, Esin Durmus, Faisal Ladhak, Myra Cheng, Debora Nozza, Tatsunori Hashimoto, Dan Jurafsky, James Zou, and Aylin Caliskan. Easily accessible text-to-image generation amplifies demographic stereotypes at large scale. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, pages 1493-1504, Chicago, Illinois, USA, 2023. ACM. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 255, + 296, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 255, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 61, + 255, + 296, + 308 + ], + "type": "text", + "content": "[2] Jaemin Cho, Abhay Zala, and Mohit Bansal. Dall-eval: Probing the reasoning skills and social biases of text-to-image generative transformers. In Proceedings of the International Conference on Computer Vision (ICCV), Paris, France, 2023. IEEE. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 310, + 295, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 310, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 62, + 310, + 295, + 354 + ], + "type": "text", + "content": "[3] Kristy Choi, Aditya Grover, Trisha Singh, Rui Shu, and Stefano Ermon. Fair generative modeling via weak supervision. In International Conference on Machine Learning, pages 1887-1898. PMLR, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 354, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 354, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 62, + 354, + 294, + 376 + ], + "type": "text", + "content": "[4] Bradley Efron. Tweedie's formula and selection bias, 2011. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 377, + 295, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 377, + 295, + 432 + ], + "spans": [ + { + "bbox": [ + 62, + 377, + 295, + 432 + ], + "type": "text", + "content": "[5] Mureji Fatunde and Crystal Tse. Digital media company stability ai raises funds at $1 billion value. https://www.bloomberg.com/news/articles/2022-10-17/digital-media-firm-stability-ai-raises-funds-at-1-billion-value, 2022.1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 434, + 295, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 434, + 295, + 465 + ], + "spans": [ + { + "bbox": [ + 62, + 434, + 295, + 465 + ], + "type": "text", + "content": "[6] Rohit Gandikota, Joanna Materzynska, Jaden Fiitto-Kaufman, and David Bau. Erasing concepts from diffusion models, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 467, + 295, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 467, + 295, + 499 + ], + "spans": [ + { + "bbox": [ + 62, + 467, + 295, + 499 + ], + "type": "text", + "content": "[7] Rohit Gandikota, Joanna Materzynska, Tingrui Zhou, Antonio Torralba, and David Bau. Concept sliders: Lora adaptors for precise control in diffusion models, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 501, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 501, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 501, + 295, + 544 + ], + "type": "text", + "content": "[8] Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzyńska, and David Bau. Unified concept editing in diffusion models. arXiv preprint arXiv:2308.14761, 2023. 2, 3, 6, 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 545, + 295, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 545, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 62, + 545, + 295, + 567 + ], + "type": "text", + "content": "[9] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance, 2022. 3, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 568, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 568, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 568, + 294, + 590 + ], + "type": "text", + "content": "[10] Jungseock Joo. Fairface attribute model. https://github.com/jojoa/fairface, 2021. 6, 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 591, + 295, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 295, + 634 + ], + "type": "text", + "content": "[11] Hang Li, Chengzhi Shen, Philip Torr, Volker Tresp, and Jindong Gu. Self-discovering interpretable diffusion latent directions for responsible text-to-image generation, 2024. 3, 6, 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "text", + "content": "[12] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 3730-3738, Santiago, Chile, 2015. IEEE. 6, 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[13] Alexandra Sasha Luccioni, Christopher Akiki, Margaret Mitchell, and Yacine Jernite. Stable bias: Analyzing so" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "type": "text", + "content": "cietal representations in diffusion models. arXiv preprint arXiv:2303.11408, 2023. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 96, + 554, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 554, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 554, + 140 + ], + "type": "text", + "content": "[14] Mengyao Lyu, Yuhong Yang, Haiwen Hong, Hui Chen, Xuan Jin, Yuan He, Hui Xue, Jungong Han, and Guiguang Ding. One-dimensional adapter to rule them all: Concepts, diffusion models and erasing applications, 2024. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 141, + 507, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 507, + 151 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 507, + 151 + ], + "type": "text", + "content": "[15] Ellis Monk. The monk skin tone scale, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 152, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 553, + 196 + ], + "type": "text", + "content": "[16] Junhyun Nam, Sangwoo Mo, Jaeho Lee, and Jinwoo Shin. Breaking the spurious causality of conditional generation via fairness intervention with corrective sampling. Transactions on Machine Learning Research, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 198, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 229 + ], + "type": "text", + "content": "[17] Hadas Orgad, Bahjat Kawar, and Yonatan Belinkov. Editing implicit assumptions in text-to-image diffusion models. 2023. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 231, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 231, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 231, + 553, + 274 + ], + "type": "text", + "content": "[18] Rishubh Parihar, Abhijnya Bhat, Abhipsa Basu, Saswat Mallick, Jogendra Nath Kundu, and R. Venkatesh Babu. Balancing act: Distribution-guided debiasing in diffusion models, 2024. 2, 3, 6, 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 276, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 341 + ], + "type": "text", + "content": "[19] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 342, + 553, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 385 + ], + "type": "text", + "content": "[20] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 387, + 552, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 387, + 552, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 387, + 552, + 397 + ], + "type": "text", + "content": "[21] Robin Rombach. Stable diffusion v1-3 model card, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 399, + 552, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 552, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 552, + 409 + ], + "type": "text", + "content": "[22] Robin Rombach. Stable diffusion v1-4 model card, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 411, + 552, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 552, + 421 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 552, + 421 + ], + "type": "text", + "content": "[23] Robin Rombach. Stable diffusion v1-5 model card, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 422, + 552, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 552, + 432 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 552, + 432 + ], + "type": "text", + "content": "[24] Robin Rombach. Stable diffusion v2-0 model card, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 434, + 552, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 552, + 445 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 552, + 445 + ], + "type": "text", + "content": "[25] Robin Rombach. Stable diffusion v2-1 model card, 2022. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 446, + 555, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 446, + 555, + 500 + ], + "spans": [ + { + "bbox": [ + 317, + 446, + 555, + 500 + ], + "type": "text", + "content": "[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 502, + 554, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 502, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 502, + 554, + 567 + ], + "type": "text", + "content": "[27] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 568, + 554, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 568, + 554, + 622 + ], + "spans": [ + { + "bbox": [ + 317, + 568, + 554, + 622 + ], + "type": "text", + "content": "[28] Prasanna Sattigeri, Samuel C Hoffman, Vijil Chenthamarakshan, and Kush R Varshney. Fairness gan: Generating datasets with fairness properties using a generative adversarial network. IBM Journal of Research and Development, 63 (4/5):3-1, 2019. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 624, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 656 + ], + "type": "text", + "content": "[29] Preethi Seshadri, Sameer Singh, and Yanai Elazar. The bias amplification paradox in text-to-image generation. arXiv preprint arXiv:2308.00755, 2023. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "type": "text", + "content": "[30] Xudong Shen, Chao Du, Tianyu Pang, Min Lin, Yongkang Wong, and Mohan Kankanhalli. Finetuning text-to-image diffusion models for fairness, 2024. 2, 3, 4, 6, 1" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[31] Christopher TH Teo, Milad Abdollahzadeh, and Ngai-Man Cheung. Fair generative models via transfer learning. In Pro" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18419" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 376 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "ceedings of the AAAI Conference on Artificial Intelligence, pages 2429-2437, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "type": "text", + "content": "[32] Soobin Um and Changho Suh. A fair generative model using lecam divergence. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10034-10042, 2023." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 130, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 294, + 184 + ], + "type": "text", + "content": "[33] Boris van Breugel, Trent Kyono, Jeroen Berrevoets, and Michaela van der Schaar. Decaf: Generating fair synthetic data using causally-aware generative networks. Advances in Neural Information Processing Systems, 34:22221-22233, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 229 + ], + "type": "text", + "content": "[34] Zihao Wang, Lin Gui, Jeffrey Negrea, and Victor Veitch. Concept algebra for score-based conditional model. In ICML 2023 Workshop on Structured Probabilistic Inference & Generative Modeling, Honolulu, HI, 2023. PMLR. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "type": "text", + "content": "[35] Depeng Xu, Shuhan Yuan, Lu Zhang, and Xintao Wu. Fairgan: Fairness-aware generative adversarial networks. In 2018 IEEE International Conference on Big Data (Big Data), pages 570-575. IEEE, 2018. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 294, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 341 + ], + "type": "text", + "content": "[36] Ning Yu, Ke Li, Peng Zhou, Jitendra Malik, Larry Davis, and Mario Fritz. Inclusive gan: Improving data and minority coverage in generative models. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 377-393. Springer, 2020. 2, 3, 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 342, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 294, + 376 + ], + "type": "text", + "content": "[37] Cheng Zhang, Xuanbai Chen, Siqi Chai, Chen Henry Wu, Dmitry Lagun, Thabo Beeler, and Fernando De la Torre. Itigen: Inclusive text-to-image generation, 2023. 6" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18420" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/FairHuman_ Boosting Hand and Face Quality in Human Image Generation with Minimum Potential Delay Fairness in Diffusion Models/c1cadd0b-ab93-4bff-a36b-6e24c8c2c52f_content_list.json b/2025/FairHuman_ Boosting Hand and Face Quality in Human Image Generation with Minimum Potential Delay Fairness in Diffusion Models/c1cadd0b-ab93-4bff-a36b-6e24c8c2c52f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c5f3b7044a1e30c63649b4b9b587a35490d0e54c --- /dev/null +++ b/2025/FairHuman_ Boosting Hand and Face Quality in Human Image Generation with Minimum Potential Delay Fairness in Diffusion Models/c1cadd0b-ab93-4bff-a36b-6e24c8c2c52f_content_list.json @@ -0,0 +1,1646 @@ +[ + { + "type": "text", + "text": "FairHuman: Boosting Hand and Face Quality in Human Image Generation with Minimum Potential Delay Fairness in Diffusion Models", + "text_level": 1, + "bbox": [ + 119, + 92, + 879, + 137 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuxuan Wang1", + "bbox": [ + 186, + 165, + 307, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "wangyuxuan2019@bupt.edu.cn", + "bbox": [ + 130, + 186, + 364, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhongjiang He²", + "bbox": [ + 184, + 208, + 313, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "hezhongj_1@163.com", + "bbox": [ + 169, + 229, + 330, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tianwei $\\mathrm{Cao}^{1\\dagger}$", + "bbox": [ + 444, + 165, + 560, + 181 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "caotianwei@bupt.edu.cn", + "bbox": [ + 403, + 186, + 601, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kongming Liang $^{1\\dagger}$", + "bbox": [ + 411, + 207, + 563, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "liangkongming@bupt.edu.cn", + "bbox": [ + 377, + 229, + 599, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Huayu Zhang", + "bbox": [ + 694, + 165, + 810, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zhanghy56@chinatelecom.cn", + "bbox": [ + 642, + 186, + 864, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhanyu Ma", + "bbox": [ + 684, + 208, + 785, + 226 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mazhanyu@bupt.edu.cn", + "bbox": [ + 645, + 229, + 825, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Beijing University of Posts and Telecommunications 2 Institute of Artificial Intelligence, China Telecom", + "bbox": [ + 99, + 268, + 934, + 287 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6976644bf0064501270d002d22d9fbb7ac54aa0848538f72e93f2a56ead76646.jpg", + "image_caption": [ + "Figure 1. Compare with existing methods under two scenarios: 1) General T2I Generation (in the left part): from left to right are SDXL-base[31], MoLE[49] and Ours. 2) Controllable Generation (in the right part): from left to right are input condition, ControlNet-SDXL[47] and Ours. Zoom in for a better view of details." + ], + "image_footnote": [], + "bbox": [ + 96, + 318, + 563, + 598 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d46b8957a8f09948832c8fadda36083bd181dc07be6976380e379060a3a97b6f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 566, + 321, + 905, + 598 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 662, + 326, + 679 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image generation has achieved remarkable progress with the development of large-scale text-to-image models, especially diffusion-based models. However, generating human images with plausible details, such as faces or hands, remains challenging due to insufficient supervision of local regions during training. To address this issue, we propose FairHuman, a multi-objective fine-tuning approach designed to enhance both global and local generation quality fairly. Specifically, we first construct three learning objectives: a global objective derived from the default diffusion objective function and two local objectives for hands and faces based on pre-annotated positional priors. Subse", + "bbox": [ + 88, + 695, + 485, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sequently, we derive the optimal parameter updating strategy under the guidance of the Minimum Potential Delay (MPD) criterion, thereby attaining fairness-aware optimization for this multi-objective problem. Based on this, our proposed method can achieve significant improvements in generating challenging local details while maintaining overall quality. Extensive experiments showcase the effectiveness of our method in improving the performance of human image generation under different scenarios. Our code is available at https://github.com/PRIS-CV/FairHuman.", + "bbox": [ + 511, + 665, + 908, + 816 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 844, + 645, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion-based models have revolutionized image generation by achieving unprecedented high quality and realism", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 238, + 0, + 807, + 18 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 323, + 16, + 722, + 30 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 31, + 754, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding Author", + "bbox": [ + 114, + 887, + 246, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17046", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ad596aa3bc8fe4b1dc448f34482efd31afb58e227a3c1043b387e1af26554b5e.jpg", + "image_caption": [ + "Figure 2. Examples of human-related images generated by SDXL. The first row shows the close-up view images with prompts like \"a close-up human face\" and \"a detailed hand gesture\". The second row shows the human images under different scenes. Zoom in for a better view of details." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 480, + 239 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "through iterative denoising processes. However, their capability of human image generation remains challenged by persistent artifacts in anatomically complex regions, particularly faces and hands. These contents often suffer from generating degradation such as distorted facial expressions and implausible hand poses.", + "bbox": [ + 89, + 352, + 482, + 443 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this issue, previous work, including ControlNet[47], HumanSD[14], and HyperHuman[22], leverages anatomical information like pose or depth to guide human image generation. Moreover, a leading approach [49] fine-tunes two low-rank modules w.r.t hand and face generation, respectively. Then it uses the soft Mixture of Experts (MoE) mechanism[32] to fuse these modules. Essentially, these methods aim to enhance human image generation by injecting specific knowledge about hands and faces into the generative model.", + "bbox": [ + 89, + 445, + 482, + 595 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite the success of these methods, the balance between coarse appearance (global quality) and local details (quality of faces and hands) has not been addressed. Through the visual feedback of the human-related images generated by SDXL-base[31] in Fig.2, we have noticed an interesting phenomenon: When the model generates faces or hands separately as the close-up view images, it usually ensures sufficient detail and high quality in terms of structure and texture. However, when generating whole-body images especially when hand or face regions occupy a relatively small portion, noticeable distortions often occur. It suggests that the model inherently possesses the capability to generate plausible faces and hands. Nonetheless, the existence of other contents such as the background, reduces the focus on local regions, thus leading to a decline in generation quality. In addition, a post-processing method called ADetailer[3] repairs the faces and hands by cropping out the distorted regions from the original image and repainting them using super-resolution techniques. Its effectiveness further supports our hypothesis. In this way, the level of", + "bbox": [ + 89, + 598, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "attention to global and local aspects is indeed a significant factor influencing the generation quality. Relevant details will be elaborated in Sec.3.1.", + "bbox": [ + 511, + 90, + 903, + 135 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by this observation, we decide to construct specific objective functions for global and local generation, respectively. We notice that the default objective function in Eq.(1) for the diffusion model is based on the global mean squared error (MSE), which pays more attention to the overall performance. For complex tasks like human image generation, a single-objective mechanism may lead to the model excelling in certain aspects while underperforming in others. Therefore, here we transform the human image generation task into a multi-objective optimization problem and divide it into three different sub-objectives. Specifically, our optimization objectives include:", + "bbox": [ + 511, + 138, + 906, + 319 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Facial generation quality: the naturalness of facial structure and expressions.", + "- Hand generation quality: the plausibility of hand poses and shapes.", + "- Global generation quality: the coherence and perceptual quality of the overall image." + ], + "bbox": [ + 511, + 321, + 903, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As for the construction of objective functions, we retain the default MSE loss function for global features. Then, we create masks using pre-annotated positional information from the dataset and construct mask-based loss functions for local features, namely faces and hands.", + "bbox": [ + 511, + 412, + 903, + 488 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Considering the optimization method, we need to identify an appropriate strategy to maximize the improvement of each objective. Since faces and hands are more difficult to learn compared to global features during training, directly deploying a fixed or equal weight allocation, i.e., the degree of importance, is unreasonable. To achieve the Pareto optimality of human image generation performance, inspired by the idea of fairness criteria utilized in communication networks[26], we incorporate the minimum potential delay (MPD) fairness principle[25] into our optimization process for parameter updates. Specifically, it dynamically allocates the weight for each objective-specific gradient following the fairness constraints. In this way, the three objective functions mentioned above can be optimized simultaneously within a single diffusion model while mitigating potential conflicts between them. Technically, it can be deployed on fine-tuning different modules like LoRA[9] and ControlNet[47] for certain generation tasks.", + "bbox": [ + 511, + 489, + 905, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we propose a framework called FairHuman to generate human images with higher fidelity and quality under different scenarios. The key insight is that human image generation is a complex task that involves both global and local optimization. The previous single-objective methodology cannot improve the performance of different aspects at the same time. Therefore, we regard the learning for human image generation as a multi-objective optimization process and implement the MPD fairness prin", + "bbox": [ + 511, + 763, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17047", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ciple to achieve a more effective and balanced performance.", + "bbox": [ + 89, + 90, + 480, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method contains three procedures (see Fig.3): Firstly, we construct separate objective functions for global and regional generation by utilizing the default MSE loss function and regional masks obtained from the preannotated positional information. Secondly, we implement the minimum potential delay fairness principle for the joint optimization of all objectives. Eventually, we fine-tune LoRA[9] and ControlNet[47] based on the mentioned strategy for two scenarios of human image generation.", + "bbox": [ + 88, + 106, + 480, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, the contributions of our paper are as follows:", + "bbox": [ + 89, + 242, + 482, + 270 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We emphasize that the poor performance in generating local details of human images stems from insufficient attention during the training process. Based on this factor, we construct different objective functions to focus on global and local generation separately.", + "- We propose FairHuman, a general multi-objective optimization framework for diffusion model in improving the performance of human image generation. Here, we optimize different objective functions simultaneously within a single model based on the minimum potential delay fairness principle. To our knowledge, it has not been applied in the diffusion model before.", + "- We implement a widely adopted diffusion model SDXL[31] as the backbone and deploy our method on different modules, including LoRA[9] and ControlNet[47]. Extensive experimental results demonstrate that our method yields a superior performance in enhancing the generation of human images under diverse scenarios." + ], + "bbox": [ + 89, + 273, + 483, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 558, + 232, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Human Image Generation. Human image generation refers to generating high-quality human images complying with the given conditions (e.g., text, image). Among image synthesis tasks, it has garnered significant attention due to its considerable academic research values and application potentials[12]. Driven by advancements in scalability and training techniques, existing text-to-image models have demonstrated the ability to generate images with increasing resolution and fidelity. Following the success of GANs[37, 45, 46], diffusion models[8, 39] has emerged as a powerful alternative and become the predominant architecture currently. Models like Stable Diffusion[36] and DALL-E 2[34] leverage latent diffusion processes to generate highly detailed and semantically accurate images from textual descriptions. However, when creating human-centric images, the latest models[2, 5, 31] still struggle to generate fine-grained human parts, especially faces and hands. Previous approaches such as[14, 22, 23, 40] incorporate supplementary conditions like human pose to achieve better generation in a controllable pattern but have major limitations in consistency. A concurrent work[49] deals", + "bbox": [ + 89, + 583, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with the issue of faces and hands by utilizing two low-rank modules[9] with the Mixture of Experts(MoE)[11] mechanism. However, it requires fine-tuning each expert module separately, which brings additional training costs. Meanwhile, stacking of multiple modules also affects the generation efficiency. In our work, we implement joint optimization for both global and local generation performance following the minimum potential delay fairness principle.", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Datasets for Human Image Generation. The availability of high-quality datasets has significantly propelled the development of human image generation models. These datasets[4, 6, 10, 13, 48] provide the necessary diversity and richness required for training models in generating realistic and varied human images. However, most of them have two main drawbacks: 1) low resolution and poor image quality, which are inadequate to provide valid prior knowledge for diffusion models. 2) lack of annotations on morphological and spatial information of the human body, which cannot effectively guide the model in generating delicate parts like human faces and hands. To alleviate the above limitations, we collect images from available high-quality human-centric datasets[19, 27] that contain high-resolution real-world images. To further guarantee the image quality, we apply detectors[30, 35, 43] to filter out undesirable images while maintaining as much valuable information as possible. In addition, we annotate the body information for controllable generation and positional information for constructing mask-based objective functions. Implementing the procedures mentioned above, we can achieve enhanced training performance aligned with our objectives.", + "bbox": [ + 511, + 212, + 903, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 513, + 559, + 648, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Statement", + "text_level": 1, + "bbox": [ + 511, + 585, + 697, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The objective function of diffusion models is founded on the mean square error associated with the denoising task. It enhances the model's performance by minimizing the discrepancy between the predicted noise and the actual noise. In the case of Latent Diffusion Models (LDM)[36], this denoising process is executed within the latent space, facilitated by a VAE encoder[16]. This approach allows for more efficient and effective denoising by operating in a compressed representation of the data. Specifically, its optimization objective can be formulated as:", + "bbox": [ + 511, + 607, + 903, + 758 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} E _ {z _ {0}, t, c, \\epsilon} \\left[ \\| \\epsilon - \\hat {\\epsilon} _ {\\theta} \\left(z _ {t}, t, c\\right) \\| ^ {2} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 771, + 903, + 797 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $z_{t}$ is the noisy latent at timestep $t$ obtained from $z_{0}$ through the diffusion process by adding Gaussian noise $\\epsilon \\sim N(0,1)$ ; the initial representation $z_{0}$ is defined as $z_{0} = \\varepsilon(x_{0})$ , with $\\varepsilon$ denoting the encoder and $x_{0}$ being the ground truth image; $\\hat{\\epsilon}_{\\theta}(z_{t},t,c)$ is the noise predicted by the diffusion model; $c$ represents the embedding of input", + "bbox": [ + 511, + 810, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17048", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0c11849a7c4bab38bb60d8e770aacdea8c9b60c8c9cc3e9d8b97d8babe418075.jpg", + "image_caption": [ + "Figure 3. An overview of FairHuman framework. The entire pipeline is divided into three stages: 1) Construct multi objectives for human image generation; 2) Employ fairness-aware optimization strategy for parameter updates; 3) Apply the fine-tuned module into the backbone diffusion model. Here, we take LoRA[9] and ControlNet[47] as examples of different generation scenarios." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 903, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "conditions (e.g., textual prompt or image); $\\theta$ represents the parameters to be updated.", + "bbox": [ + 88, + 458, + 482, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By pre-training on large-scale datasets with this objective function, models such as SDXL[31] have shown the ability to generate high-quality human content. This is evident in the first row of Fig.2, where close-up images are generated accurately. However, when generating full-body human images within a scene (see the second row of Fig.2), the model demonstrates inferior performance in rendering local details (faces and hands). In other words, when face or hand regions occupy a larger proportion of the image, their generation results are acceptable. On the contrary, when they occupy a smaller proportion of the image, their corresponding generation quality would be worse.", + "bbox": [ + 88, + 498, + 482, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "What accounts for the occurrence of this issue? As indicated by Eq.(1), LDM calculates the loss based on the representation $z_{t}$ encoded from the entire image without distinguishing the optimization difficulty of different regions. For faces and hands with larger areas in close-up images, the default objective function can be easily optimized since these regions contribute more significantly to the loss. Conversely, for human-in-the-scene images, the representation $z_{t}$ comprises not only the hands and faces but also other body details and the background scene. Therefore, face and hand regions with smaller areas contribute less to the loss, thus leading to insufficient attention and resulting in poorer generation quality. In this way, different regions may compete during the generation process.", + "bbox": [ + 89, + 689, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Multi-objectives", + "text_level": 1, + "bbox": [ + 511, + 455, + 674, + 473 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As we discussed above, the competition between different regions may lead to insufficient attention to some local areas, such as hands or faces. To address this issue, the generative model requires a fair allocation of resources to both local and global regions, thus enhancing the overall generation performance. To this end, we first construct the following optimization objectives for both local and global generation tasks, respectively:", + "bbox": [ + 511, + 479, + 906, + 601 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nl _ {\\text {g l o b a l}} (\\theta) = \\mathbb {E} _ {z _ {0}, t, c, \\epsilon} \\left[ \\| \\epsilon - \\hat {\\epsilon} _ {\\theta} (z _ {t}, t, c) \\| ^ {2} \\right];\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 611, + 812, + 637 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nl _ {\\text {f a c e}} (\\theta) = \\mathbb {E} _ {z _ {0}, t, c, \\epsilon} \\left[ \\| m _ {f} \\odot \\left(\\epsilon - \\hat {\\epsilon} _ {\\theta} (z _ {t}, t, c)\\right) \\| ^ {2} \\right]; \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 638, + 906, + 662 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nl _ {\\mathrm {h a n d}} (\\theta) = \\mathbb {E} _ {z _ {0}, t, c, \\epsilon} \\left[ \\| m _ {h} \\odot (\\epsilon - \\hat {\\epsilon} _ {\\theta} (z _ {t}, t, c)) \\| ^ {2} \\right],\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 666, + 866, + 693 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\theta$ represents the learnable parameters; $l_{global}(\\theta)$ is the global loss for entire image generation; $l_{face}(\\theta)$ and $l_{hand}(\\theta)$ are local losses focusing on faces and hands; $m_f$ and $m_h$ are the masks used to select the face and hand regions respectively. Here, the masks $m_f$ and $m_h$ are obtained by reshaping the mask image to the latent shape, following inpainting mask $m$ denoted in Eq.(4) of [23]. For brevity, we also denote $l_{global}(\\theta), l_{face}(\\theta), l_{hand}(\\theta)$ as $l_1(\\theta), l_2(\\theta), l_3(\\theta)$ in the subsequent sections.", + "bbox": [ + 511, + 704, + 906, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on this, we aim to fine-tune a diffusion model to simultaneously minimize $l_{1}(\\theta)$ , $l_{2}(\\theta)$ , and $l_{3}(\\theta)$ . Meanwhile, the fine-tuning approach should enhance the generation quality of hands and faces while maintaining its global", + "bbox": [ + 511, + 840, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17049", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "performance. In this way, the fine-tuning task could be formulated as the following multi-objective optimization problem:", + "bbox": [ + 89, + 90, + 482, + 133 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\mathcal {L} (\\theta) = \\left[ l _ {1} (\\theta), l _ {2} (\\theta), l _ {3} (\\theta) \\right] ^ {\\top}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 133, + 482, + 157 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Notably, our primary target comprises two key aspects:", + "bbox": [ + 89, + 161, + 455, + 175 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Regional Fairness: We should ensure that the solution $\\theta^{*}$ to Eq.(3) yields a fairly substantial improvement across all three objectives, especially for $l_{2}(\\theta)$ and $l_{3}(\\theta)$ . As long as there is at least one insignificant improvement among the three objectives, it implies that our fine-tuning is ineffective. This aligns with our purpose of fair enhancement on the quality w.r.t face, hand, and global regions.", + "- Pareto Optimality: The solution $\\theta^{*}$ to Eq.(3) should be Pareto optimal[29, 38]. Here we say a solution $\\theta^{*}$ is Pareto optimal if there exists no other solution $\\theta^{\\prime}$ satisfying $l_{i}(\\theta^{\\prime})\\leq l_{i}(\\theta^{*})$ for all $i\\in \\{1,2,3\\}$ . In other words, updating $\\theta^{*}$ in any direction will disrupt at least one of the three objectives. Thus Pareto optimality here implies a trade-off between face, hand, and global quality." + ], + "bbox": [ + 89, + 176, + 483, + 404 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Fairness-aware Optimization", + "text_level": 1, + "bbox": [ + 89, + 410, + 349, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We aim to ensure the post fine-tuning values of $l_{1}(\\theta), l_{2}(\\theta), l_{3}(\\theta)$ are all fairly minimized to a sufficient degree. To this end, a straightforward approach is to enable each $l_{i}(\\theta)$ could be maximally decreased at each training step of $\\theta$ . Formally, we seek a stepwise updating direction $d$ that maximizes the contributions of $d$ to all of $l_{1}(\\theta), l_{2}(\\theta)$ and $l_{3}(\\theta)$ . From another perspective, this is equivalent to minimizing the negative effect of $d$ to all these objectives.", + "bbox": [ + 89, + 431, + 483, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, we employ the concept of Potential Delay [25, 26] to quantitatively describe the negative effect. The term \"Potential Delay\" originally refers to the expected time required for a user to complete a specific task in the context of resource allocation. For example, if an internet user transmits a unit-sized file at transmission rate $v$ , then the potential delay is $1 / v$ . In this paper, the common updating direction $d$ can be considered a shared resource among all objectives. The \"transmission rate\" of this resource to each $l_{i}(\\theta)$ is represented by the projection magnitude of $d$ onto the steepest descent direction $\\nabla l_{i}(\\theta)$ , denoted as $\\| \\mathrm{proj}_{\\nabla l_i(\\theta)}(d)\\|$ . Consequently, the corresponding potential delay is $1 / \\| \\mathrm{proj}_{\\nabla l_i(\\theta)}(d)\\|$ . Normally, projection $\\| \\mathrm{proj}_{\\nabla l_i(\\theta)}(d)\\|$ can serve as a metric for the contribution of $d$ to the optimization process of $l_{i}(\\theta)$ [44]. In this way, the potential delay $1 / \\| \\mathrm{proj}_{\\nabla l_i(\\theta)}(d)\\|$ is inversely proportional to the contribution from $d$ and thus directly proportional to the negative effect. In this way, we can determine the optimal direction $d$ by resolving the subsequent minimization problem:", + "bbox": [ + 89, + 551, + 483, + 854 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {d} \\mathcal {F} (d) = \\sum_ {i = 1} ^ {3} \\frac {1}{\\| \\operatorname {p r o j} _ {\\nabla l _ {i} (\\theta)} (d) \\|}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 168, + 863, + 482, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By optimizing Eq.(4), the potential delay of $l_{1}(\\theta)$ , $l_{2}(\\theta)$ , $l_{3}(\\theta)$ can be minimized simultaneously, thus it is termed as Minimal Potential Delay (MPD) Fairness [25, 26].", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Prior to solving this optimization, we assume that $d$ lies within a ball with radius $r$ centered at the origin. Based on this, the optimization problem w.r.t MPD fairness can be transformed into the following form:", + "bbox": [ + 511, + 136, + 903, + 198 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {d} \\mathcal {F} ^ {\\prime} (d) = \\sum_ {i = 1} ^ {3} \\frac {1}{\\nabla l _ {i} (\\theta) ^ {\\top} d}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 209, + 903, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "since that $\\mathcal{F}'(d)$ is proportional to an upper bound of $\\mathcal{F}(d)$ in Eq.(4), and the proof is provided in the Sec.1.1 of the supplementary material. Here, $\\mathcal{F}'(d)$ can be regarded as a special case of $\\alpha$ -fairness[26]. If we set $\\alpha = 2$ for Eq.(2) in [1], it will be equivalent to Eq.(5) in this paper. Thus, we can adopt a similar procedure to solve Eq.(5) following [1].", + "bbox": [ + 511, + 261, + 903, + 352 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, we first apply Lagrange multiplier method on Eq.(5) to obtain the following equation:", + "bbox": [ + 511, + 353, + 903, + 383 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nd = \\sum_ {i = 1} ^ {3} \\frac {\\nabla l _ {i} (\\theta)}{(\\nabla l _ {i} (\\theta) ^ {\\top} d) ^ {2}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 395, + 903, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The derivation process is described in Sec.1.2 of the supplementary material. To further simplify the problem, we follow [1] and express $d$ as a linear combination:", + "bbox": [ + 511, + 446, + 903, + 492 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nd = \\sum_ {i = 1} ^ {3} w _ {i} \\nabla l _ {i} (\\theta), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 503, + 903, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $w_{i}$ represents the weight of the $i$ -th objective. Given that each gradient $\\nabla l_{i}(\\theta)$ is readily available, the problem of determining $d$ is reformulated as determining every $w_{i}$ . By substituting Eq.(7) into Eq.(6), we can get that:", + "bbox": [ + 511, + 556, + 903, + 617 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nG (\\theta) ^ {\\top} G (\\theta) W = W ^ {- 1 / 2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 627, + 903, + 646 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $G(\\theta) = [\\nabla l_1(\\theta), \\nabla l_2(\\theta), \\nabla l_3(\\theta)]^\\top$ and $W = [w_1, w_2, w_3]^\\top$ . According to the relation in Eq.(8), we can further construct the following convex optimization problem:", + "bbox": [ + 511, + 656, + 903, + 715 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {W} \\left\\| G (\\theta) ^ {\\top} G (\\theta) W - W ^ {- 1 / 2} \\right\\| ^ {2}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 715, + 903, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By applying the standard least square method, we can obtain the following close-form solution of Eq.(9):", + "bbox": [ + 511, + 750, + 903, + 781 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nW = \\left(G (\\theta) ^ {\\top} G (\\theta)\\right) ^ {- \\frac {2}{3}} \\mathbf {1}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 792, + 903, + 815 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{1}$ represents an all-ones vector. With these weights, we can finally obtain the value of $d$ by using Eq.(7) and then conduct an updating step as:", + "bbox": [ + 511, + 825, + 903, + 872 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {\\text {n e w}} = \\theta - \\eta d, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 653, + 885, + 903, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17050", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\theta_{new}$ is the updated parameter and $\\eta$ is the learning rate. By adopting this approach, each training step is guided by MPD fairness, thereby ensuring the efficacy of the finetuning results across all objectives. Meanwhile, Thm.(7.3) in [1] also indicates that such a training process can converge to a Pareto optimal solution. And our constructed objective function decouples regional and global optimization, satisfying MPD's prerequisite assumption of linearly independent task gradients.", + "bbox": [ + 88, + 90, + 480, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Application", + "text_level": 1, + "bbox": [ + 89, + 236, + 217, + 253 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Considering the high cost of full fine-tuning for diffusion models, we opt to deploy our method in LoRA[9] and ControlNet[47] while freezing the parameters of the backbone model. They represent two types of generation patterns, namely, general T2I generation and controllable generation with control signals. Specifically, LoRA primarily enhances the quality of human images from the perspective of visual concepts and pixel information, while ControlNet focuses on more refined and controllable generation over the structural aspects. Both have significant influence and play important roles in the field of human image generation. We construct a high-quality dataset for their fine-tuning. Apart from the images and text descriptions, it also includes the positional information for creating mask-based objective functions and the pose-depth information for controllable generation. Their new outputs can be formulated as:", + "bbox": [ + 91, + 257, + 483, + 513 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} _ {\\text {n e w}} = N _ {\\text {b a s e}} (z; \\theta_ {\\text {b a s e}}) + \\beta N _ {\\text {n e w}} (c, z; \\theta_ {\\text {u p d a t e}}), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 526, + 482, + 544 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $z$ is the input latent and $N_{base}$ is the backbone network with the frozen parameters $\\theta_{base}$ . $N_{new}$ represents the fine-tuned module (LoRA or ControlNet here) with the updated parameters $\\theta_{update}$ . As for condition $c$ , it contains textual and timestep embeddings for the LoRA module and has additional control signals for the ControlNet module. $\\beta$ is the hyperparameter used for scaling.", + "bbox": [ + 89, + 553, + 482, + 660 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 672, + 223, + 690 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 89, + 696, + 297, + 714 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We collect human-centric images from two public datasets called CosmicMan[19] and BodyHands[27]. To enable effective training, we implement several preprocessing methods, including filtering, captioning, and labeling. Ultimately, we construct a high-quality human image dataset that contains approximately 50k images for training. As for the validation, we randomly sample about 5k in-the-wild human images under different scenes with well-processed captions provided by [49]. The details regarding the construction of our dataset will be elaborated in Sec.2 of the supplementary material.", + "bbox": [ + 89, + 719, + 482, + 885 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline Methods. We compare our approach with two", + "bbox": [ + 89, + 885, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "categories of baseline methods. 1) General T2I methods without extra conditions: SDXL-base[31] and MoLE[49]. Notably, MoLE is a recent work based on the collaboration of multiple LoRA modules. 2) Controllable methods with extra conditions: ControlNet-SDXL[47] and ControlNet Union[18]. Here, we compare performance under the same given conditions (depth and pose) using different training strategies.", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. We adopt evaluation metrics based on the optimization objectives outlined in the paper, focusing on two dimensions. 1) Overall image quality: Human Preference Score (HPS)[41] and ImageReward (IR)[42]. They are built on the CLIP[33] and BLIP[17] models respectively to reflect human visual preferences on the generated images from multiple aspects including image alignment, aesthetic, and plausibility. Moreover, CLIP score is applied to evaluate text-image alignment, and pose alignment is used for assessing the effectiveness of controllable generation. 2) Regional generation quality: FID[7] and detection confidence of mediapipe[24] are utilized to evaluate the quality of hands and faces in terms of realism and plausibility. The implementation details about our experiments are provided in Sec.3 of the supplementary material.", + "bbox": [ + 511, + 212, + 906, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with Existing Methods", + "text_level": 1, + "bbox": [ + 511, + 452, + 821, + 468 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Analysis. The evaluation results compared with existing methods are shown in Tab.1. We perform evaluations under two classic scenarios. Specifically, for general T2I methods, including SDXL-Base[31], MoLE[49], and our LoRA-based method, we randomly sample 5k prompts about human-in-the-scene images from the validation dataset and feed them to each model for image generation. As for the setting of hyperparameters, we keep the default CFG scale of 7.5, set the inference step to 30, and fix the seed at 0 for all methods. In particular, for our LoRA-based method, we implement three different weights (0.3, 0.4, and 0.5) in the experiments and use their average metric values as the final result. To more comprehensively evaluate the quality of generated images, we utilize the metrics in terms of overall image quality, text-image alignment, and regional quality. The results demonstrate that our method outperforms existing approaches in all these three aspects, especially in the generation quality of local hand and face details. Since our LoRA-based method and MoLE are deployed on SDXL sharing the same backbone model, we also conduct experiments on their memory usage and inference speed. Results suggest that our method can achieve better performance while requiring less memory size and inference time. Related details are demonstrated in Sec.4.1 of the supplementary material.", + "bbox": [ + 511, + 476, + 906, + 852 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Furthermore, to verify the effectiveness of our method across different task scenarios, we employ it in ControlNet[47] for the controllable generation. As shown", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17051", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/61b1028b70f62162287e8910ef512e59d3b0cc564a7f92596592d6770b74079f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Application ScenarioModelImage QualityText-Image AlignmentRegional QualityPose Alignment
HPS(%) ↑IR(%) ↑CLIP ↑FID ↓Hand Confi. ↑Face Confi. ↑Mean Diff ↓
General T2I MethodsSD-XL Base [31]30.91140.8526.8452.6993.1883.67/
MoLE [49]31.87152.3627.1748.9793.3284.48/
Ours32.73154.6027.2046.1194.0385.31/
Controllable MethodsControlNet-SDXL [47]32.20152.4927.1046.1893.5085.133.45
ControlNet-Union [18]32.80155.2227.2146.4394.2885.362.67
ControlNet-Ours32.72154.7427.2445.4694.4885.612.35
", + "bbox": [ + 94, + 93, + 903, + 205 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. The evaluation results for ours and other existing methods. We conduct comparisons under two different image generation scenarios separately. As for the general T2I methods, we compare our LoRA-based method with a widely adopted diffusion model SDXL[31] and a recent work MoLE[49]. For controllable methods, we compare our ControlNet-based method with the default ControlNet-SDXL[47] and the latest ControlNet-Union[18]. Notably, the mean difference of pose condition is employed here to evaluate the effectiveness of controllable generation. The best results are indicated in bold.", + "bbox": [ + 89, + 215, + 906, + 286 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in the lower part of Tab.1, we compare with existing controllable methods, including the default ControlNet-SDXL and the latest variant ControlNet-Union[18]. The evaluation is conducted using the same metrics as described above. Similar to our LoRA-based method, we set three different control scales (0.4, 0.5, and 0.6) here for testing and choose their average metric values. Besides, considering the characteristics of controllable generation, we also evaluate the model's performance based on pose alignment, i.e., the consistency between the input conditions and the actual generated content. Here, we use the mean difference between the input and extracted conditions as the metric. The corresponding results are shown in the last column of Tab.1. In comparison, our method still demonstrates superior performance in the controllable generation scenario. Although it is slightly inferior to ControlNet-Union in terms of HPS and IR metrics, the primary reason is that ControlNet-Union incorporates aesthetic and preference feedback learning during training, whereas our method focuses more on the realism and plausibility of the image, and significantly outperforms in regional details.", + "bbox": [ + 88, + 311, + 485, + 630 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d0e9f0d34f94e39244f2e7493307ecc663a019bc01f79d6e51601148bb29ae36.jpg", + "image_caption": [ + "Figure 4. User study results in terms of overall quality and regional quality. We present the ratio of users' preference for each method." + ], + "image_footnote": [], + "bbox": [ + 91, + 643, + 480, + 753 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Analysis. Fig.1 shows qualitative results. Compared with baseline methods, our method can generate human images with higher quality in both global and local aspects. Meanwhile, it achieves better controllability and plausibility in terms of controllable generation. We also", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "conduct a user study to provide a more intuitive evaluation. Volunteers are invited to select images that better align with their visual preferences based on overall and regional quality. The result is shown in Fig.4. It indicates that our method is more preferred by users in both two aspects. More details about the user study will be demonstrated in Sec.4.2 of the supplementary material.", + "bbox": [ + 511, + 311, + 906, + 419 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2e31b412e8f2e3796a6fc85c89f8928707202cbc0a41022163bbc6d270f4be89.jpg", + "image_caption": [ + "Figure 5. Examples of the images generated by different settings with the same prompt. First row: generated by LoRA. Second row: generated by ControlNet. Zoom in for a better view of details." + ], + "image_footnote": [], + "bbox": [ + 517, + 433, + 903, + 661 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 756, + 679, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We present the key ablation studies based on the construction of our objective functions and the strategy for optimization. Here, we analyze their effects according to the overall and local quality of the generated images, deploying on LoRA and ControlNet, respectively. We randomly select 1k human-related prompts and follow the same hyperparameter setting as mentioned above for testing. Their results are shown in Tab.2. In addition, some visualization results are", + "bbox": [ + 511, + 779, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17052", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/01973f37e222c4afca9619c4cb9a47031ffcd4a06d5ae708f32451fdafb3b4b1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelAblation SettingsImage QualityRegional QualityPose Alignment
HPS(%)↑IR(%)↑FID ↓Hand Confi. ↑Face Confi. ↑Mean Diff ↓
LoRAl1(θ)32.40151.0849.4892.8183.84/
l1(θ) + l2(θ) + l3(θ) (w/o fairness)32.60153.2347.2293.7084.82/
l2(θ) + l2(θ) + l3(θ) (w fairness)32.73154.6046.1194.0385.31/
ControlNetl1(θ)32.20152.4946.1893.5085.133.45
l1(θ) + l2(θ) + l3(θ) (w/o fairness)32.56154.3046.7993.9985.242.78
l1(θ) + l2(θ) + l3(θ) (w fairness)32.72154.7445.3694.4885.612.35
", + "bbox": [ + 94, + 93, + 903, + 224 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/164bea3458fcf8ab803b5b3d3ba084ab348b09caeb882b37b9c83a6c67e23f69.jpg", + "table_caption": [ + "Table 2. Ablation results deployed on LoRA and ControlNet. We primarily explore the effect of two aspects: additional objective functions for regional generation and the learning strategy. $l_{1}(\\theta)$ , $l_{2}(\\theta)$ , and $l_{3}(\\theta)$ are defined in Eq.2. The best results are indicated in bold." + ], + "table_footnote": [], + "table_body": "
LoRA RankImage QualityT2I AlignmentRegional Quality
HPS(%) ↑IR(%) ↑CLIP↑FID ↓Hand Confi. ↑Face Confi. ↑
r=6432.50148.5227.1147.8793.2584.69
r=12832.73154.827.1546.7993.7985.11
r=25632.73154.6027.2046.1194.0385.31
", + "bbox": [ + 93, + 305, + 480, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Comparison results for the choice of LoRA rank. The best results are indicated in bold.", + "bbox": [ + 89, + 375, + 482, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "provided in Fig.5.", + "bbox": [ + 89, + 430, + 212, + 445 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Regional Optimization Objective. To explore how regional loss functions including $l_{\\text{face}}(\\theta)$ and $l_{\\text{hand}}(\\theta)$ in Eq.(2) influence generation performance, we fine-tuned the corresponding LoRA and ControlNet by adopting the default loss function from Eq.(1) and the joint objective functions in Eq.(2), respectively. For the latter, we employ a fixed-weight linear scalarization method here, transforming the multi-objective optimization into a single-objective optimization. The first two rows for each model in Tab.2 present the evaluation results, validating that the inclusion of regional loss functions helps the model improve its generation capabilities for details such as hands and faces, thereby enhancing the overall quality of the images.", + "bbox": [ + 88, + 446, + 482, + 641 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Fairness Principle. In our work, we utilize the minimum potential delay (MPD) fairness to achieve better optimization based on the joint of objective functions for global and regional generation. Here, we compare two gradient update strategies: one without considering fairness (i.e., a fixed weight allocation based on linear scalarization method) and one considering fairness (i.e., a dynamical weight allocation based on MPD fairness). As shown in the last two rows for each model in Tab.2, the optimization strategy based on the MPD fairness can better focus on the more challenging objectives of generating faces and hands without compromising the overall quality. It achieves a more balanced performance among different objectives.", + "bbox": [ + 88, + 642, + 482, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. More Analysis", + "text_level": 1, + "bbox": [ + 89, + 848, + 238, + 864 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Considering that the rank setting significantly impacts the LoRA performance, we additionally conduct a hyperparam-", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "eter study to figure out the optimal rank configuration. The results are presented in Tab.3, which indicates that setting the rank value to 256 maximizes the model's ability to generate regional details without excessively jeopardizing the overall image quality. Besides, we also conduct comparison studies on some other multi-objective learning strategies, e.g., Dynamic Weight Average (DwA)[21], Uncertainty Weighting (UW)[15], and Random Loss Weighting (RLW)[20]. The results suggest that applying the MPD fairness principle to the optimization process can achieve a more balanced and superior performance in terms of the human image generation task. Relevant results and details are presented in Sec.4.3 of the supplementary material.", + "bbox": [ + 511, + 303, + 906, + 501 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Visualizations", + "text_level": 1, + "bbox": [ + 511, + 518, + 658, + 535 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We put more visualization results compared with existing methods in Sec.5 of the supplementary material, including other close-source methods like HyperHuman[22] and HanDiffuser[28]. Our method still demonstrates comparable or superior performance.", + "bbox": [ + 511, + 545, + 906, + 621 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 645, + 633, + 660 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose FairHuman to enhance the quality of human images generated by diffusion models. To address the distortion issue in local details like faces and hands, we specifically construct corresponding objective functions for regional generation. Then, we combine the default global objective function to achieve joint training. The key of our method is regarding the improvement of human image generation as a multi-objective optimization task. By utilizing the minimum potential delay principle and the gradient manipulation process, our method enables more balanced and effective performance. Extensive experiments demonstrate the superiority of our method under different scenarios of human image generation. For future work, we plan to further design and optimize corresponding objective functions for more attributes related to humans, such as feet and eyes.", + "bbox": [ + 511, + 674, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17053", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by National Natural Science Foundation of China(NSFC) No.62225601, 62476029, U23B2052, and sponsored by Beijing Nova Program.", + "bbox": [ + 89, + 114, + 485, + 157 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 186, + 187, + 203 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hao Ban and Kaiyi Ji. Fair resource allocation in multi-task learning. arXiv preprint arXiv:2402.15638, 2024. 5, 6", + "[2] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf, 2(3):8, 2023. 3", + "[3] Bing-su. Adeditor. https://github.com/Bing-su/adeditor.git, 2024.2", + "[4] Seunghwan Choi, Sunghyun Park, Minsoo Lee, and Jaegul Choo. Viton-hd: High-resolution virtual try-on via misalignment-aware normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14131-14140, 2021. 3", + "[5] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 3", + "[6] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen Change Loy, Wayne Wu, and Ziwei Liu. Stylegan-human: A data-centric odyssey of human generation. In European Conference on Computer Vision, pages 1-19. Springer, 2022. 3", + "[7] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6", + "[8] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3", + "[9] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 3, 4, 6", + "[10] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013. 3", + "[11] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. Neural computation, 3(1):79-87, 1991. 3", + "[12] Zhen Jia, Zhang Zhang, Liang Wang, and Tieniu Tan. Human image generation: A comprehensive survey. ACM Computing Surveys, 56(11):1-39, 2024. 3" + ], + "bbox": [ + 93, + 212, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Xuan Ju, Ailing Zeng, Jianan Wang, Qiang Xu, and Lei Zhang. Human-art: A versatile human-centric dataset bridging natural and artificial scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 618-629, 2023. 3", + "[14] Xuan Ju, Ailing Zeng, Chenchen Zhao, Jianan Wang, Lei Zhang, and Qiang Xu. Humansd: A native skeleton-guided diffusion model for human image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15988-15998, 2023. 2, 3", + "[15] Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7482-7491, 2018. 8", + "[16] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3", + "[17] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 6", + "[18] Ming Li, Taojiannan Yang, Huafeng Kuang, Jie Wu, Zhaoning Wang, Xuefeng Xiao, and Chen Chen. Controlnet++: Improving conditional controls with efficient consistency feedback: Project page: liming-ai. github.io/controlnet_plus_plus. In European Conference on Computer Vision, pages 129–147. Springer, 2024. 6, 7", + "[19] Shikai Li, Jianglin Fu, Kaiyuan Liu, Wentao Wang, KwanYee Lin, and Wayne Wu. Cosmicman: A text-to-image foundation model for humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6955-6965, 2024. 3, 6", + "[20] Baijiong Lin, Feiyang Ye, Yu Zhang, and Ivor W Tsang. Reasonable effectiveness of random weighting: A litmus test for multi-task learning. arXiv preprint arXiv:2111.10603, 2021.8", + "[21] Shikun Liu, Edward Johns, and Andrew J Davison. End-to-end multi-task learning with attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1871-1880, 2019. 8", + "[22] Xian Liu, Jian Ren, Aliaksandr Siarohin, Ivan Skorokhodov, Yanyu Li, Dahua Lin, Xihui Liu, Ziwei Liu, and Sergey Tulyakov. Hyperhuman: Hyper-realistic human generation with latent structural diffusion. arXiv preprint arXiv:2310.08579, 2023. 2, 3, 8", + "[23] Wenquan Lu, Yufei Xu, Jing Zhang, Chaoyue Wang, and Dacheng Tao. Handrefiner: Refining malformed hands in generated images by diffusion-based conditional inpainting. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 7085-7093, 2024. 3, 4", + "[24] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Guang Yong, Juhyun Lee, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, 2019. 6", + "[25] Laurent Massoulié and James Roberts. Bandwidth sharing: objectives and algorithms. In IEEE INFOCOM'99. Con-" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17054", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ference on Computer Communications. Proceedings. Eighteenth Annual Joint Conference of the IEEE Computer and Communications Societies. The Future is Now (Cat. No. 99CH36320), pages 1395-1403. IEEE, 1999. 2, 5", + "[26] Jeonghoon Mo and Jean Walrand. Fair end-to-end window-based congestion control. IEEE/ACM Transactions on networking, 8(5):556-567, 2000. 2, 5", + "[27] Supreeth Narasimhaswamy, Thanh Nguyen, Mingzhen Huang, and Minh Hoai. Whose hands are these? hand detection and hand-body association in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4889-4899, 2022. 3, 6", + "[28] Supreeth Narasimhaswamy, Uttaran Bhattacharya, Xiang Chen, Ishita Dasgupta, Saayan Mitra, and Minh Hoai. Handiffuser: Text-to-image generation with realistic hand appearances. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2468-2479, 2024. 8", + "[29] Patrick Ngatchou, Anahita Zarei, and A El-Sharkawi. Pareto multi objective optimization. In Proceedings of the 13th international conference on, intelligent systems application to power systems, pages 84-91. IEEE, 2005. 5", + "[30] Georgios Pavlakos, Dandan Shan, Ilija Radosavovic, Angjoo Kanazawa, David Fouhey, and Jitendra Malik. Reconstructing hands in 3d with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9826-9836, 2024. 3", + "[31] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 2, 3, 4, 6, 7", + "[32] Joan Puigcerver, Carlos Riquelme, Basil Mustafa, and Neil Houlsby. From sparse to soft mixtures of experts. arXiv preprint arXiv:2308.00951, 2023. 2", + "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 6", + "[34] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 3", + "[35] J Redmon. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2016. 3", + "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3", + "[37] Axel Sauer, Tero Karras, Samuli Laine, Andreas Geiger, and Timo Aila. Stylegan-t: Unlocking the power of gans for fast large-scale text-to-image synthesis. In International conference on machine learning, pages 30105-30118. PMLR, 2023. 3" + ], + "bbox": [ + 91, + 90, + 483, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Ozan Sener and Vladlen Koltun. Multi-task learning as multi-objective optimization. Advances in neural information processing systems, 31, 2018. 5", + "[39] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3", + "[40] Yuxuan Wang, Tianwei Cao, Kongming Liang, Zhongjiang He, Hao Sun, Yongxiang Li, and Zhanyu Ma. Mixture-of-hand-experts: Repainting the deformed hand images generated by diffusion models. In Chinese Conference on Pattern Recognition and Computer Vision (PRCV), pages 143-157. Springer, 2024. 3", + "[41] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023. 6", + "[42] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36:15903-15935, 2023. 6", + "[43] Zhendong Yang, Ailing Zeng, Chun Yuan, and Yu Li. Effective whole-body pose estimation with two-stages distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4210-4220, 2023. 3", + "[44] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. In Advances in Neural Information Processing Systems, pages 5824-5836, 2020. 5", + "[45] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 5907-5915, 2017. 3", + "[46] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stack-gan++: Realistic image synthesis with stacked generative adversarial networks. IEEE transactions on pattern analysis and machine intelligence, 41(8):1947-1962, 2018. 3", + "[47] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 2, 3, 4, 6, 7", + "[48] Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jingdong Wang, and Qi Tian. Scalable person re-identification: A benchmark. In Proceedings of the IEEE international conference on computer vision, pages 1116–1124, 2015. 3", + "[49] Jie Zhu, Yixiong Chen, Mingyu Ding, Ping Luo, Leye Wang, and Jingdong Wang. Mole: Enhancing human-centric text-to-image diffusion via mixture of low-rank experts. arXiv preprint arXiv:2410.23332, 2024. 1, 2, 3, 6, 7" + ], + "bbox": [ + 516, + 92, + 903, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17055", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/FairHuman_ Boosting Hand and Face Quality in Human Image Generation with Minimum Potential Delay Fairness in Diffusion Models/c1cadd0b-ab93-4bff-a36b-6e24c8c2c52f_model.json b/2025/FairHuman_ Boosting Hand and Face Quality in Human Image Generation with Minimum Potential Delay Fairness in Diffusion Models/c1cadd0b-ab93-4bff-a36b-6e24c8c2c52f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..743a6d9f10ffa64aec06a32ab9ebebd2741dfbf7 --- /dev/null +++ b/2025/FairHuman_ Boosting Hand and Face Quality in Human Image Generation with Minimum Potential Delay Fairness in Diffusion Models/c1cadd0b-ab93-4bff-a36b-6e24c8c2c52f_model.json @@ -0,0 +1,2233 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.02 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.325, + 0.017, + 0.723, + 0.031 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.032, + 0.755, + 0.047 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.12, + 0.093, + 0.88, + 0.138 + ], + "angle": 0, + "content": "FairHuman: Boosting Hand and Face Quality in Human Image Generation with Minimum Potential Delay Fairness in Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.187, + 0.166, + 0.308, + 0.185 + ], + "angle": 0, + "content": "Yuxuan Wang1" + }, + { + "type": "text", + "bbox": [ + 0.132, + 0.188, + 0.365, + 0.202 + ], + "angle": 0, + "content": "wangyuxuan2019@bupt.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.185, + 0.209, + 0.315, + 0.228 + ], + "angle": 0, + "content": "Zhongjiang He²" + }, + { + "type": "text", + "bbox": [ + 0.171, + 0.23, + 0.331, + 0.244 + ], + "angle": 0, + "content": "hezhongj_1@163.com" + }, + { + "type": "text", + "bbox": [ + 0.445, + 0.166, + 0.561, + 0.183 + ], + "angle": 0, + "content": "Tianwei \\(\\mathrm{Cao}^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.405, + 0.188, + 0.602, + 0.201 + ], + "angle": 0, + "content": "caotianwei@bupt.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.413, + 0.208, + 0.564, + 0.228 + ], + "angle": 0, + "content": "Kongming Liang\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.378, + 0.23, + 0.6, + 0.244 + ], + "angle": 0, + "content": "liangkongming@bupt.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.695, + 0.166, + 0.812, + 0.185 + ], + "angle": 0, + "content": "Huayu Zhang" + }, + { + "type": "text", + "bbox": [ + 0.643, + 0.188, + 0.866, + 0.201 + ], + "angle": 0, + "content": "zhanghy56@chinatelecom.cn" + }, + { + "type": "text", + "bbox": [ + 0.686, + 0.209, + 0.786, + 0.227 + ], + "angle": 0, + "content": "Zhanyu Ma" + }, + { + "type": "text", + "bbox": [ + 0.647, + 0.23, + 0.826, + 0.244 + ], + "angle": 0, + "content": "mazhanyu@bupt.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.269, + 0.936, + 0.289 + ], + "angle": 0, + "content": "1 Beijing University of Posts and Telecommunications 2 Institute of Artificial Intelligence, China Telecom" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.319, + 0.565, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.567, + 0.322, + 0.906, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.609, + 0.908, + 0.651 + ], + "angle": 0, + "content": "Figure 1. Compare with existing methods under two scenarios: 1) General T2I Generation (in the left part): from left to right are SDXL-base[31], MoLE[49] and Ours. 2) Controllable Generation (in the right part): from left to right are input condition, ControlNet-SDXL[47] and Ours. Zoom in for a better view of details." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.664, + 0.327, + 0.68 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.696, + 0.486, + 0.879 + ], + "angle": 0, + "content": "Image generation has achieved remarkable progress with the development of large-scale text-to-image models, especially diffusion-based models. However, generating human images with plausible details, such as faces or hands, remains challenging due to insufficient supervision of local regions during training. To address this issue, we propose FairHuman, a multi-objective fine-tuning approach designed to enhance both global and local generation quality fairly. Specifically, we first construct three learning objectives: a global objective derived from the default diffusion objective function and two local objectives for hands and faces based on pre-annotated positional priors. Subse" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.666, + 0.909, + 0.817 + ], + "angle": 0, + "content": "sequently, we derive the optimal parameter updating strategy under the guidance of the Minimum Potential Delay (MPD) criterion, thereby attaining fairness-aware optimization for this multi-objective problem. Based on this, our proposed method can achieve significant improvements in generating challenging local details while maintaining overall quality. Extensive experiments showcase the effectiveness of our method in improving the performance of human image generation under different scenarios. Our code is available at https://github.com/PRIS-CV/FairHuman." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.845, + 0.646, + 0.861 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.903 + ], + "angle": 0, + "content": "Diffusion-based models have revolutionized image generation by achieving unprecedented high quality and realism" + }, + { + "type": "page_footnote", + "bbox": [ + 0.116, + 0.888, + 0.248, + 0.901 + ], + "angle": 0, + "content": "† Corresponding Author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17046" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.481, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.252, + 0.483, + 0.321 + ], + "angle": 0, + "content": "Figure 2. Examples of human-related images generated by SDXL. The first row shows the close-up view images with prompts like \"a close-up human face\" and \"a detailed hand gesture\". The second row shows the human images under different scenes. Zoom in for a better view of details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.353, + 0.483, + 0.444 + ], + "angle": 0, + "content": "through iterative denoising processes. However, their capability of human image generation remains challenged by persistent artifacts in anatomically complex regions, particularly faces and hands. These contents often suffer from generating degradation such as distorted facial expressions and implausible hand poses." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.446, + 0.483, + 0.597 + ], + "angle": 0, + "content": "To address this issue, previous work, including ControlNet[47], HumanSD[14], and HyperHuman[22], leverages anatomical information like pose or depth to guide human image generation. Moreover, a leading approach [49] fine-tunes two low-rank modules w.r.t hand and face generation, respectively. Then it uses the soft Mixture of Experts (MoE) mechanism[32] to fuse these modules. Essentially, these methods aim to enhance human image generation by injecting specific knowledge about hands and faces into the generative model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Despite the success of these methods, the balance between coarse appearance (global quality) and local details (quality of faces and hands) has not been addressed. Through the visual feedback of the human-related images generated by SDXL-base[31] in Fig.2, we have noticed an interesting phenomenon: When the model generates faces or hands separately as the close-up view images, it usually ensures sufficient detail and high quality in terms of structure and texture. However, when generating whole-body images especially when hand or face regions occupy a relatively small portion, noticeable distortions often occur. It suggests that the model inherently possesses the capability to generate plausible faces and hands. Nonetheless, the existence of other contents such as the background, reduces the focus on local regions, thus leading to a decline in generation quality. In addition, a post-processing method called ADetailer[3] repairs the faces and hands by cropping out the distorted regions from the original image and repainting them using super-resolution techniques. Its effectiveness further supports our hypothesis. In this way, the level of" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.136 + ], + "angle": 0, + "content": "attention to global and local aspects is indeed a significant factor influencing the generation quality. Relevant details will be elaborated in Sec.3.1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.139, + 0.907, + 0.32 + ], + "angle": 0, + "content": "Motivated by this observation, we decide to construct specific objective functions for global and local generation, respectively. We notice that the default objective function in Eq.(1) for the diffusion model is based on the global mean squared error (MSE), which pays more attention to the overall performance. For complex tasks like human image generation, a single-objective mechanism may lead to the model excelling in certain aspects while underperforming in others. Therefore, here we transform the human image generation task into a multi-objective optimization problem and divide it into three different sub-objectives. Specifically, our optimization objectives include:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.322, + 0.905, + 0.351 + ], + "angle": 0, + "content": "- Facial generation quality: the naturalness of facial structure and expressions." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.352, + 0.905, + 0.382 + ], + "angle": 0, + "content": "- Hand generation quality: the plausibility of hand poses and shapes." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.383, + 0.905, + 0.412 + ], + "angle": 0, + "content": "- Global generation quality: the coherence and perceptual quality of the overall image." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.322, + 0.905, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.414, + 0.905, + 0.489 + ], + "angle": 0, + "content": "As for the construction of objective functions, we retain the default MSE loss function for global features. Then, we create masks using pre-annotated positional information from the dataset and construct mask-based loss functions for local features, namely faces and hands." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.491, + 0.906, + 0.763 + ], + "angle": 0, + "content": "Considering the optimization method, we need to identify an appropriate strategy to maximize the improvement of each objective. Since faces and hands are more difficult to learn compared to global features during training, directly deploying a fixed or equal weight allocation, i.e., the degree of importance, is unreasonable. To achieve the Pareto optimality of human image generation performance, inspired by the idea of fairness criteria utilized in communication networks[26], we incorporate the minimum potential delay (MPD) fairness principle[25] into our optimization process for parameter updates. Specifically, it dynamically allocates the weight for each objective-specific gradient following the fairness constraints. In this way, the three objective functions mentioned above can be optimized simultaneously within a single diffusion model while mitigating potential conflicts between them. Technically, it can be deployed on fine-tuning different modules like LoRA[9] and ControlNet[47] for certain generation tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.901 + ], + "angle": 0, + "content": "To this end, we propose a framework called FairHuman to generate human images with higher fidelity and quality under different scenarios. The key insight is that human image generation is a complex task that involves both global and local optimization. The previous single-objective methodology cannot improve the performance of different aspects at the same time. Therefore, we regard the learning for human image generation as a multi-objective optimization process and implement the MPD fairness prin" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17047" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.106 + ], + "angle": 0, + "content": "ciple to achieve a more effective and balanced performance." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.107, + 0.482, + 0.242 + ], + "angle": 0, + "content": "Our method contains three procedures (see Fig.3): Firstly, we construct separate objective functions for global and regional generation by utilizing the default MSE loss function and regional masks obtained from the preannotated positional information. Secondly, we implement the minimum potential delay fairness principle for the joint optimization of all objectives. Eventually, we fine-tune LoRA[9] and ControlNet[47] based on the mentioned strategy for two scenarios of human image generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.243, + 0.483, + 0.271 + ], + "angle": 0, + "content": "In summary, the contributions of our paper are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.274, + 0.483, + 0.348 + ], + "angle": 0, + "content": "- We emphasize that the poor performance in generating local details of human images stems from insufficient attention during the training process. Based on this factor, we construct different objective functions to focus on global and local generation separately." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.349, + 0.484, + 0.452 + ], + "angle": 0, + "content": "- We propose FairHuman, a general multi-objective optimization framework for diffusion model in improving the performance of human image generation. Here, we optimize different objective functions simultaneously within a single model based on the minimum potential delay fairness principle. To our knowledge, it has not been applied in the diffusion model before." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.454, + 0.484, + 0.543 + ], + "angle": 0, + "content": "- We implement a widely adopted diffusion model SDXL[31] as the backbone and deploy our method on different modules, including LoRA[9] and ControlNet[47]. Extensive experimental results demonstrate that our method yields a superior performance in enhancing the generation of human images under diverse scenarios." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.274, + 0.484, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.559, + 0.233, + 0.573 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Human Image Generation. Human image generation refers to generating high-quality human images complying with the given conditions (e.g., text, image). Among image synthesis tasks, it has garnered significant attention due to its considerable academic research values and application potentials[12]. Driven by advancements in scalability and training techniques, existing text-to-image models have demonstrated the ability to generate images with increasing resolution and fidelity. Following the success of GANs[37, 45, 46], diffusion models[8, 39] has emerged as a powerful alternative and become the predominant architecture currently. Models like Stable Diffusion[36] and DALL-E 2[34] leverage latent diffusion processes to generate highly detailed and semantically accurate images from textual descriptions. However, when creating human-centric images, the latest models[2, 5, 31] still struggle to generate fine-grained human parts, especially faces and hands. Previous approaches such as[14, 22, 23, 40] incorporate supplementary conditions like human pose to achieve better generation in a controllable pattern but have major limitations in consistency. A concurrent work[49] deals" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.212 + ], + "angle": 0, + "content": "with the issue of faces and hands by utilizing two low-rank modules[9] with the Mixture of Experts(MoE)[11] mechanism. However, it requires fine-tuning each expert module separately, which brings additional training costs. Meanwhile, stacking of multiple modules also affects the generation efficiency. In our work, we implement joint optimization for both global and local generation performance following the minimum potential delay fairness principle." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.213, + 0.905, + 0.545 + ], + "angle": 0, + "content": "Datasets for Human Image Generation. The availability of high-quality datasets has significantly propelled the development of human image generation models. These datasets[4, 6, 10, 13, 48] provide the necessary diversity and richness required for training models in generating realistic and varied human images. However, most of them have two main drawbacks: 1) low resolution and poor image quality, which are inadequate to provide valid prior knowledge for diffusion models. 2) lack of annotations on morphological and spatial information of the human body, which cannot effectively guide the model in generating delicate parts like human faces and hands. To alleviate the above limitations, we collect images from available high-quality human-centric datasets[19, 27] that contain high-resolution real-world images. To further guarantee the image quality, we apply detectors[30, 35, 43] to filter out undesirable images while maintaining as much valuable information as possible. In addition, we annotate the body information for controllable generation and positional information for constructing mask-based objective functions. Implementing the procedures mentioned above, we can achieve enhanced training performance aligned with our objectives." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.56, + 0.649, + 0.577 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.586, + 0.698, + 0.601 + ], + "angle": 0, + "content": "3.1. Problem Statement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.608, + 0.905, + 0.759 + ], + "angle": 0, + "content": "The objective function of diffusion models is founded on the mean square error associated with the denoising task. It enhances the model's performance by minimizing the discrepancy between the predicted noise and the actual noise. In the case of Latent Diffusion Models (LDM)[36], this denoising process is executed within the latent space, facilitated by a VAE encoder[16]. This approach allows for more efficient and effective denoising by operating in a compressed representation of the data. Specifically, its optimization objective can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.772, + 0.905, + 0.798 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} E _ {z _ {0}, t, c, \\epsilon} \\left[ \\| \\epsilon - \\hat {\\epsilon} _ {\\theta} \\left(z _ {t}, t, c\\right) \\| ^ {2} \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.811, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\( z_{t} \\) is the noisy latent at timestep \\( t \\) obtained from \\( z_{0} \\) through the diffusion process by adding Gaussian noise \\( \\epsilon \\sim N(0,1) \\); the initial representation \\( z_{0} \\) is defined as \\( z_{0} = \\varepsilon(x_{0}) \\), with \\( \\varepsilon \\) denoting the encoder and \\( x_{0} \\) being the ground truth image; \\( \\hat{\\epsilon}_{\\theta}(z_{t},t,c) \\) is the noise predicted by the diffusion model; \\( c \\) represents the embedding of input" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17048" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.905, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.389, + 0.908, + 0.434 + ], + "angle": 0, + "content": "Figure 3. An overview of FairHuman framework. The entire pipeline is divided into three stages: 1) Construct multi objectives for human image generation; 2) Employ fairness-aware optimization strategy for parameter updates; 3) Apply the fine-tuned module into the backbone diffusion model. Here, we take LoRA[9] and ControlNet[47] as examples of different generation scenarios." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.459, + 0.483, + 0.492 + ], + "angle": 0, + "content": "conditions (e.g., textual prompt or image); \\(\\theta\\) represents the parameters to be updated." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.499, + 0.483, + 0.681 + ], + "angle": 0, + "content": "By pre-training on large-scale datasets with this objective function, models such as SDXL[31] have shown the ability to generate high-quality human content. This is evident in the first row of Fig.2, where close-up images are generated accurately. However, when generating full-body human images within a scene (see the second row of Fig.2), the model demonstrates inferior performance in rendering local details (faces and hands). In other words, when face or hand regions occupy a larger proportion of the image, their generation results are acceptable. On the contrary, when they occupy a smaller proportion of the image, their corresponding generation quality would be worse." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.902 + ], + "angle": 0, + "content": "What accounts for the occurrence of this issue? As indicated by Eq.(1), LDM calculates the loss based on the representation \\( z_{t} \\) encoded from the entire image without distinguishing the optimization difficulty of different regions. For faces and hands with larger areas in close-up images, the default objective function can be easily optimized since these regions contribute more significantly to the loss. Conversely, for human-in-the-scene images, the representation \\( z_{t} \\) comprises not only the hands and faces but also other body details and the background scene. Therefore, face and hand regions with smaller areas contribute less to the loss, thus leading to insufficient attention and resulting in poorer generation quality. In this way, different regions may compete during the generation process." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.457, + 0.676, + 0.474 + ], + "angle": 0, + "content": "3.2. Multi-objectives" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.481, + 0.907, + 0.602 + ], + "angle": 0, + "content": "As we discussed above, the competition between different regions may lead to insufficient attention to some local areas, such as hands or faces. To address this issue, the generative model requires a fair allocation of resources to both local and global regions, thus enhancing the overall generation performance. To this end, we first construct the following optimization objectives for both local and global generation tasks, respectively:" + }, + { + "type": "equation", + "bbox": [ + 0.533, + 0.612, + 0.813, + 0.638 + ], + "angle": 0, + "content": "\\[\nl _ {\\text {g l o b a l}} (\\theta) = \\mathbb {E} _ {z _ {0}, t, c, \\epsilon} \\left[ \\| \\epsilon - \\hat {\\epsilon} _ {\\theta} (z _ {t}, t, c) \\| ^ {2} \\right];\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.639, + 0.907, + 0.664 + ], + "angle": 0, + "content": "\\[\nl _ {\\text {f a c e}} (\\theta) = \\mathbb {E} _ {z _ {0}, t, c, \\epsilon} \\left[ \\| m _ {f} \\odot \\left(\\epsilon - \\hat {\\epsilon} _ {\\theta} (z _ {t}, t, c)\\right) \\| ^ {2} \\right]; \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.539, + 0.667, + 0.867, + 0.694 + ], + "angle": 0, + "content": "\\[\nl _ {\\mathrm {h a n d}} (\\theta) = \\mathbb {E} _ {z _ {0}, t, c, \\epsilon} \\left[ \\| m _ {h} \\odot (\\epsilon - \\hat {\\epsilon} _ {\\theta} (z _ {t}, t, c)) \\| ^ {2} \\right],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.841 + ], + "angle": 0, + "content": "where \\(\\theta\\) represents the learnable parameters; \\(l_{global}(\\theta)\\) is the global loss for entire image generation; \\(l_{face}(\\theta)\\) and \\(l_{hand}(\\theta)\\) are local losses focusing on faces and hands; \\(m_f\\) and \\(m_h\\) are the masks used to select the face and hand regions respectively. Here, the masks \\(m_f\\) and \\(m_h\\) are obtained by reshaping the mask image to the latent shape, following inpainting mask \\(m\\) denoted in Eq.(4) of [23]. For brevity, we also denote \\(l_{global}(\\theta), l_{face}(\\theta), l_{hand}(\\theta)\\) as \\(l_1(\\theta), l_2(\\theta), l_3(\\theta)\\) in the subsequent sections." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Based on this, we aim to fine-tune a diffusion model to simultaneously minimize \\( l_{1}(\\theta) \\), \\( l_{2}(\\theta) \\), and \\( l_{3}(\\theta) \\). Meanwhile, the fine-tuning approach should enhance the generation quality of hands and faces while maintaining its global" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17049" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.135 + ], + "angle": 0, + "content": "performance. In this way, the fine-tuning task could be formulated as the following multi-objective optimization problem:" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.135, + 0.483, + 0.158 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\mathcal {L} (\\theta) = \\left[ l _ {1} (\\theta), l _ {2} (\\theta), l _ {3} (\\theta) \\right] ^ {\\top}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.162, + 0.457, + 0.176 + ], + "angle": 0, + "content": "Notably, our primary target comprises two key aspects:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.177, + 0.483, + 0.297 + ], + "angle": 0, + "content": "- Regional Fairness: We should ensure that the solution \\(\\theta^{*}\\) to Eq.(3) yields a fairly substantial improvement across all three objectives, especially for \\(l_{2}(\\theta)\\) and \\(l_{3}(\\theta)\\). As long as there is at least one insignificant improvement among the three objectives, it implies that our fine-tuning is ineffective. This aligns with our purpose of fair enhancement on the quality w.r.t face, hand, and global regions." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.297, + 0.484, + 0.405 + ], + "angle": 0, + "content": "- Pareto Optimality: The solution \\(\\theta^{*}\\) to Eq.(3) should be Pareto optimal[29, 38]. Here we say a solution \\(\\theta^{*}\\) is Pareto optimal if there exists no other solution \\(\\theta^{\\prime}\\) satisfying \\(l_{i}(\\theta^{\\prime})\\leq l_{i}(\\theta^{*})\\) for all \\(i\\in \\{1,2,3\\}\\). In other words, updating \\(\\theta^{*}\\) in any direction will disrupt at least one of the three objectives. Thus Pareto optimality here implies a trade-off between face, hand, and global quality." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.177, + 0.484, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.411, + 0.351, + 0.427 + ], + "angle": 0, + "content": "3.3. Fairness-aware Optimization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.433, + 0.484, + 0.553 + ], + "angle": 0, + "content": "We aim to ensure the post fine-tuning values of \\( l_{1}(\\theta), l_{2}(\\theta), l_{3}(\\theta) \\) are all fairly minimized to a sufficient degree. To this end, a straightforward approach is to enable each \\( l_{i}(\\theta) \\) could be maximally decreased at each training step of \\( \\theta \\). Formally, we seek a stepwise updating direction \\( d \\) that maximizes the contributions of \\( d \\) to all of \\( l_{1}(\\theta), l_{2}(\\theta) \\) and \\( l_{3}(\\theta) \\). From another perspective, this is equivalent to minimizing the negative effect of \\( d \\) to all these objectives." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.484, + 0.856 + ], + "angle": 0, + "content": "Here, we employ the concept of Potential Delay [25, 26] to quantitatively describe the negative effect. The term \"Potential Delay\" originally refers to the expected time required for a user to complete a specific task in the context of resource allocation. For example, if an internet user transmits a unit-sized file at transmission rate \\( v \\), then the potential delay is \\( 1 / v \\). In this paper, the common updating direction \\( d \\) can be considered a shared resource among all objectives. The \"transmission rate\" of this resource to each \\( l_{i}(\\theta) \\) is represented by the projection magnitude of \\( d \\) onto the steepest descent direction \\( \\nabla l_{i}(\\theta) \\), denoted as \\( \\| \\mathrm{proj}_{\\nabla l_i(\\theta)}(d)\\| \\). Consequently, the corresponding potential delay is \\( 1 / \\| \\mathrm{proj}_{\\nabla l_i(\\theta)}(d)\\| \\). Normally, projection \\( \\| \\mathrm{proj}_{\\nabla l_i(\\theta)}(d)\\| \\) can serve as a metric for the contribution of \\( d \\) to the optimization process of \\( l_{i}(\\theta) \\) [44]. In this way, the potential delay \\( 1 / \\| \\mathrm{proj}_{\\nabla l_i(\\theta)}(d)\\| \\) is inversely proportional to the contribution from \\( d \\) and thus directly proportional to the negative effect. In this way, we can determine the optimal direction \\( d \\) by resolving the subsequent minimization problem:" + }, + { + "type": "equation", + "bbox": [ + 0.169, + 0.864, + 0.483, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\min _ {d} \\mathcal {F} (d) = \\sum_ {i = 1} ^ {3} \\frac {1}{\\| \\operatorname {p r o j} _ {\\nabla l _ {i} (\\theta)} (d) \\|}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "By optimizing Eq.(4), the potential delay of \\( l_{1}(\\theta) \\), \\( l_{2}(\\theta) \\), \\( l_{3}(\\theta) \\) can be minimized simultaneously, thus it is termed as Minimal Potential Delay (MPD) Fairness [25, 26]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.137, + 0.905, + 0.199 + ], + "angle": 0, + "content": "Prior to solving this optimization, we assume that \\(d\\) lies within a ball with radius \\(r\\) centered at the origin. Based on this, the optimization problem w.r.t MPD fairness can be transformed into the following form:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.21, + 0.905, + 0.251 + ], + "angle": 0, + "content": "\\[\n\\min _ {d} \\mathcal {F} ^ {\\prime} (d) = \\sum_ {i = 1} ^ {3} \\frac {1}{\\nabla l _ {i} (\\theta) ^ {\\top} d}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.262, + 0.905, + 0.353 + ], + "angle": 0, + "content": "since that \\(\\mathcal{F}'(d)\\) is proportional to an upper bound of \\(\\mathcal{F}(d)\\) in Eq.(4), and the proof is provided in the Sec.1.1 of the supplementary material. Here, \\(\\mathcal{F}'(d)\\) can be regarded as a special case of \\(\\alpha\\)-fairness[26]. If we set \\(\\alpha = 2\\) for Eq.(2) in [1], it will be equivalent to Eq.(5) in this paper. Thus, we can adopt a similar procedure to solve Eq.(5) following [1]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.354, + 0.905, + 0.385 + ], + "angle": 0, + "content": "Specifically, we first apply Lagrange multiplier method on Eq.(5) to obtain the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.632, + 0.396, + 0.905, + 0.437 + ], + "angle": 0, + "content": "\\[\nd = \\sum_ {i = 1} ^ {3} \\frac {\\nabla l _ {i} (\\theta)}{(\\nabla l _ {i} (\\theta) ^ {\\top} d) ^ {2}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.448, + 0.905, + 0.493 + ], + "angle": 0, + "content": "The derivation process is described in Sec.1.2 of the supplementary material. To further simplify the problem, we follow [1] and express \\(d\\) as a linear combination:" + }, + { + "type": "equation", + "bbox": [ + 0.645, + 0.505, + 0.905, + 0.545 + ], + "angle": 0, + "content": "\\[\nd = \\sum_ {i = 1} ^ {3} w _ {i} \\nabla l _ {i} (\\theta), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.557, + 0.905, + 0.618 + ], + "angle": 0, + "content": "where \\(w_{i}\\) represents the weight of the \\(i\\)-th objective. Given that each gradient \\(\\nabla l_{i}(\\theta)\\) is readily available, the problem of determining \\(d\\) is reformulated as determining every \\(w_{i}\\). By substituting Eq.(7) into Eq.(6), we can get that:" + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.628, + 0.905, + 0.647 + ], + "angle": 0, + "content": "\\[\nG (\\theta) ^ {\\top} G (\\theta) W = W ^ {- 1 / 2}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.657, + 0.905, + 0.717 + ], + "angle": 0, + "content": "where \\(G(\\theta) = [\\nabla l_1(\\theta), \\nabla l_2(\\theta), \\nabla l_3(\\theta)]^\\top\\) and \\(W = [w_1, w_2, w_3]^\\top\\). According to the relation in Eq.(8), we can further construct the following convex optimization problem:" + }, + { + "type": "equation", + "bbox": [ + 0.593, + 0.717, + 0.905, + 0.745 + ], + "angle": 0, + "content": "\\[\n\\min _ {W} \\left\\| G (\\theta) ^ {\\top} G (\\theta) W - W ^ {- 1 / 2} \\right\\| ^ {2}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.751, + 0.905, + 0.782 + ], + "angle": 0, + "content": "By applying the standard least square method, we can obtain the following close-form solution of Eq.(9):" + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.793, + 0.905, + 0.816 + ], + "angle": 0, + "content": "\\[\nW = \\left(G (\\theta) ^ {\\top} G (\\theta)\\right) ^ {- \\frac {2}{3}} \\mathbf {1}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.827, + 0.905, + 0.873 + ], + "angle": 0, + "content": "where \\(\\mathbf{1}\\) represents an all-ones vector. With these weights, we can finally obtain the value of \\(d\\) by using Eq.(7) and then conduct an updating step as:" + }, + { + "type": "equation", + "bbox": [ + 0.655, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\theta_ {\\text {n e w}} = \\theta - \\eta d, \\tag {11}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17050" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.482, + 0.228 + ], + "angle": 0, + "content": "where \\(\\theta_{new}\\) is the updated parameter and \\(\\eta\\) is the learning rate. By adopting this approach, each training step is guided by MPD fairness, thereby ensuring the efficacy of the finetuning results across all objectives. Meanwhile, Thm.(7.3) in [1] also indicates that such a training process can converge to a Pareto optimal solution. And our constructed objective function decouples regional and global optimization, satisfying MPD's prerequisite assumption of linearly independent task gradients." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.237, + 0.218, + 0.254 + ], + "angle": 0, + "content": "3.4. Application" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.258, + 0.485, + 0.514 + ], + "angle": 0, + "content": "Considering the high cost of full fine-tuning for diffusion models, we opt to deploy our method in LoRA[9] and ControlNet[47] while freezing the parameters of the backbone model. They represent two types of generation patterns, namely, general T2I generation and controllable generation with control signals. Specifically, LoRA primarily enhances the quality of human images from the perspective of visual concepts and pixel information, while ControlNet focuses on more refined and controllable generation over the structural aspects. Both have significant influence and play important roles in the field of human image generation. We construct a high-quality dataset for their fine-tuning. Apart from the images and text descriptions, it also includes the positional information for creating mask-based objective functions and the pose-depth information for controllable generation. Their new outputs can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.527, + 0.483, + 0.545 + ], + "angle": 0, + "content": "\\[\n\\hat {y} _ {\\text {n e w}} = N _ {\\text {b a s e}} (z; \\theta_ {\\text {b a s e}}) + \\beta N _ {\\text {n e w}} (c, z; \\theta_ {\\text {u p d a t e}}), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.483, + 0.661 + ], + "angle": 0, + "content": "where \\(z\\) is the input latent and \\(N_{base}\\) is the backbone network with the frozen parameters \\(\\theta_{base}\\). \\(N_{new}\\) represents the fine-tuned module (LoRA or ControlNet here) with the updated parameters \\(\\theta_{update}\\). As for condition \\(c\\), it contains textual and timestep embeddings for the LoRA module and has additional control signals for the ControlNet module. \\(\\beta\\) is the hyperparameter used for scaling." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.673, + 0.224, + 0.691 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.698, + 0.298, + 0.715 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.886 + ], + "angle": 0, + "content": "Datasets. We collect human-centric images from two public datasets called CosmicMan[19] and BodyHands[27]. To enable effective training, we implement several preprocessing methods, including filtering, captioning, and labeling. Ultimately, we construct a high-quality human image dataset that contains approximately 50k images for training. As for the validation, we randomly sample about 5k in-the-wild human images under different scenes with well-processed captions provided by [49]. The details regarding the construction of our dataset will be elaborated in Sec.2 of the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Baseline Methods. We compare our approach with two" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.212 + ], + "angle": 0, + "content": "categories of baseline methods. 1) General T2I methods without extra conditions: SDXL-base[31] and MoLE[49]. Notably, MoLE is a recent work based on the collaboration of multiple LoRA modules. 2) Controllable methods with extra conditions: ControlNet-SDXL[47] and ControlNet Union[18]. Here, we compare performance under the same given conditions (depth and pose) using different training strategies." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.213, + 0.907, + 0.44 + ], + "angle": 0, + "content": "Metrics. We adopt evaluation metrics based on the optimization objectives outlined in the paper, focusing on two dimensions. 1) Overall image quality: Human Preference Score (HPS)[41] and ImageReward (IR)[42]. They are built on the CLIP[33] and BLIP[17] models respectively to reflect human visual preferences on the generated images from multiple aspects including image alignment, aesthetic, and plausibility. Moreover, CLIP score is applied to evaluate text-image alignment, and pose alignment is used for assessing the effectiveness of controllable generation. 2) Regional generation quality: FID[7] and detection confidence of mediapipe[24] are utilized to evaluate the quality of hands and faces in terms of realism and plausibility. The implementation details about our experiments are provided in Sec.3 of the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.453, + 0.822, + 0.469 + ], + "angle": 0, + "content": "4.2. Comparison with Existing Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.907, + 0.853 + ], + "angle": 0, + "content": "Quantitative Analysis. The evaluation results compared with existing methods are shown in Tab.1. We perform evaluations under two classic scenarios. Specifically, for general T2I methods, including SDXL-Base[31], MoLE[49], and our LoRA-based method, we randomly sample 5k prompts about human-in-the-scene images from the validation dataset and feed them to each model for image generation. As for the setting of hyperparameters, we keep the default CFG scale of 7.5, set the inference step to 30, and fix the seed at 0 for all methods. In particular, for our LoRA-based method, we implement three different weights (0.3, 0.4, and 0.5) in the experiments and use their average metric values as the final result. To more comprehensively evaluate the quality of generated images, we utilize the metrics in terms of overall image quality, text-image alignment, and regional quality. The results demonstrate that our method outperforms existing approaches in all these three aspects, especially in the generation quality of local hand and face details. Since our LoRA-based method and MoLE are deployed on SDXL sharing the same backbone model, we also conduct experiments on their memory usage and inference speed. Results suggest that our method can achieve better performance while requiring less memory size and inference time. Related details are demonstrated in Sec.4.1 of the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Furthermore, to verify the effectiveness of our method across different task scenarios, we employ it in ControlNet[47] for the controllable generation. As shown" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17051" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.094, + 0.904, + 0.206 + ], + "angle": 0, + "content": "
Application ScenarioModelImage QualityText-Image AlignmentRegional QualityPose Alignment
HPS(%) ↑IR(%) ↑CLIP ↑FID ↓Hand Confi. ↑Face Confi. ↑Mean Diff ↓
General T2I MethodsSD-XL Base [31]30.91140.8526.8452.6993.1883.67/
MoLE [49]31.87152.3627.1748.9793.3284.48/
Ours32.73154.6027.2046.1194.0385.31/
Controllable MethodsControlNet-SDXL [47]32.20152.4927.1046.1893.5085.133.45
ControlNet-Union [18]32.80155.2227.2146.4394.2885.362.67
ControlNet-Ours32.72154.7427.2445.4694.4885.612.35
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.217, + 0.907, + 0.287 + ], + "angle": 0, + "content": "Table 1. The evaluation results for ours and other existing methods. We conduct comparisons under two different image generation scenarios separately. As for the general T2I methods, we compare our LoRA-based method with a widely adopted diffusion model SDXL[31] and a recent work MoLE[49]. For controllable methods, we compare our ControlNet-based method with the default ControlNet-SDXL[47] and the latest ControlNet-Union[18]. Notably, the mean difference of pose condition is employed here to evaluate the effectiveness of controllable generation. The best results are indicated in bold." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.313, + 0.486, + 0.631 + ], + "angle": 0, + "content": "in the lower part of Tab.1, we compare with existing controllable methods, including the default ControlNet-SDXL and the latest variant ControlNet-Union[18]. The evaluation is conducted using the same metrics as described above. Similar to our LoRA-based method, we set three different control scales (0.4, 0.5, and 0.6) here for testing and choose their average metric values. Besides, considering the characteristics of controllable generation, we also evaluate the model's performance based on pose alignment, i.e., the consistency between the input conditions and the actual generated content. Here, we use the mean difference between the input and extracted conditions as the metric. The corresponding results are shown in the last column of Tab.1. In comparison, our method still demonstrates superior performance in the controllable generation scenario. Although it is slightly inferior to ControlNet-Union in terms of HPS and IR metrics, the primary reason is that ControlNet-Union incorporates aesthetic and preference feedback learning during training, whereas our method focuses more on the realism and plausibility of the image, and significantly outperforms in regional details." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.645, + 0.482, + 0.755 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.807 + ], + "angle": 0, + "content": "Figure 4. User study results in terms of overall quality and regional quality. We present the ratio of users' preference for each method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Qualitative Analysis. Fig.1 shows qualitative results. Compared with baseline methods, our method can generate human images with higher quality in both global and local aspects. Meanwhile, it achieves better controllability and plausibility in terms of controllable generation. We also" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.313, + 0.907, + 0.42 + ], + "angle": 0, + "content": "conduct a user study to provide a more intuitive evaluation. Volunteers are invited to select images that better align with their visual preferences based on overall and regional quality. The result is shown in Fig.4. It indicates that our method is more preferred by users in both two aspects. More details about the user study will be demonstrated in Sec.4.2 of the supplementary material." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.434, + 0.905, + 0.662 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.672, + 0.909, + 0.726 + ], + "angle": 0, + "content": "Figure 5. Examples of the images generated by different settings with the same prompt. First row: generated by LoRA. Second row: generated by ControlNet. Zoom in for a better view of details." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.757, + 0.681, + 0.773 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We present the key ablation studies based on the construction of our objective functions and the strategy for optimization. Here, we analyze their effects according to the overall and local quality of the generated images, deploying on LoRA and ControlNet, respectively. We randomly select 1k human-related prompts and follow the same hyperparameter setting as mentioned above for testing. Their results are shown in Tab.2. In addition, some visualization results are" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17052" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.094, + 0.904, + 0.226 + ], + "angle": 0, + "content": "
ModelAblation SettingsImage QualityRegional QualityPose Alignment
HPS(%)↑IR(%)↑FID ↓Hand Confi. ↑Face Confi. ↑Mean Diff ↓
LoRAl1(θ)32.40151.0849.4892.8183.84/
l1(θ) + l2(θ) + l3(θ) (w/o fairness)32.60153.2347.2293.7084.82/
l2(θ) + l2(θ) + l3(θ) (w fairness)32.73154.6046.1194.0385.31/
ControlNetl1(θ)32.20152.4946.1893.5085.133.45
l1(θ) + l2(θ) + l3(θ) (w/o fairness)32.56154.3046.7993.9985.242.78
l1(θ) + l2(θ) + l3(θ) (w fairness)32.72154.7445.3694.4885.612.35
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.234, + 0.907, + 0.276 + ], + "angle": 0, + "content": "Table 2. Ablation results deployed on LoRA and ControlNet. We primarily explore the effect of two aspects: additional objective functions for regional generation and the learning strategy. \\( l_{1}(\\theta) \\), \\( l_{2}(\\theta) \\), and \\( l_{3}(\\theta) \\) are defined in Eq.2. The best results are indicated in bold." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.306, + 0.481, + 0.366 + ], + "angle": 0, + "content": "
LoRA RankImage QualityT2I AlignmentRegional Quality
HPS(%) ↑IR(%) ↑CLIP↑FID ↓Hand Confi. ↑Face Confi. ↑
r=6432.50148.5227.1147.8793.2584.69
r=12832.73154.827.1546.7993.7985.11
r=25632.73154.6027.2046.1194.0385.31
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.375, + 0.483, + 0.404 + ], + "angle": 0, + "content": "Table 3. Comparison results for the choice of LoRA rank. The best results are indicated in bold." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.431, + 0.214, + 0.446 + ], + "angle": 0, + "content": "provided in Fig.5." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.447, + 0.483, + 0.642 + ], + "angle": 0, + "content": "Effect of Regional Optimization Objective. To explore how regional loss functions including \\( l_{\\text{face}}(\\theta) \\) and \\( l_{\\text{hand}}(\\theta) \\) in Eq.(2) influence generation performance, we fine-tuned the corresponding LoRA and ControlNet by adopting the default loss function from Eq.(1) and the joint objective functions in Eq.(2), respectively. For the latter, we employ a fixed-weight linear scalarization method here, transforming the multi-objective optimization into a single-objective optimization. The first two rows for each model in Tab.2 present the evaluation results, validating that the inclusion of regional loss functions helps the model improve its generation capabilities for details such as hands and faces, thereby enhancing the overall quality of the images." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.643, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Effect of Fairness Principle. In our work, we utilize the minimum potential delay (MPD) fairness to achieve better optimization based on the joint of objective functions for global and regional generation. Here, we compare two gradient update strategies: one without considering fairness (i.e., a fixed weight allocation based on linear scalarization method) and one considering fairness (i.e., a dynamical weight allocation based on MPD fairness). As shown in the last two rows for each model in Tab.2, the optimization strategy based on the MPD fairness can better focus on the more challenging objectives of generating faces and hands without compromising the overall quality. It achieves a more balanced performance among different objectives." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.24, + 0.866 + ], + "angle": 0, + "content": "4.4. More Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Considering that the rank setting significantly impacts the LoRA performance, we additionally conduct a hyperparam-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.304, + 0.907, + 0.502 + ], + "angle": 0, + "content": "eter study to figure out the optimal rank configuration. The results are presented in Tab.3, which indicates that setting the rank value to 256 maximizes the model's ability to generate regional details without excessively jeopardizing the overall image quality. Besides, we also conduct comparison studies on some other multi-objective learning strategies, e.g., Dynamic Weight Average (DwA)[21], Uncertainty Weighting (UW)[15], and Random Loss Weighting (RLW)[20]. The results suggest that applying the MPD fairness principle to the optimization process can achieve a more balanced and superior performance in terms of the human image generation task. Relevant results and details are presented in Sec.4.3 of the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.52, + 0.659, + 0.536 + ], + "angle": 0, + "content": "4.5. Visualizations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.546, + 0.907, + 0.622 + ], + "angle": 0, + "content": "We put more visualization results compared with existing methods in Sec.5 of the supplementary material, including other close-source methods like HyperHuman[22] and HanDiffuser[28]. Our method still demonstrates comparable or superior performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.646, + 0.634, + 0.661 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In this work, we propose FairHuman to enhance the quality of human images generated by diffusion models. To address the distortion issue in local details like faces and hands, we specifically construct corresponding objective functions for regional generation. Then, we combine the default global objective function to achieve joint training. The key of our method is regarding the improvement of human image generation as a multi-objective optimization task. By utilizing the minimum potential delay principle and the gradient manipulation process, our method enables more balanced and effective performance. Extensive experiments demonstrate the superiority of our method under different scenarios of human image generation. For future work, we plan to further design and optimize corresponding objective functions for more attributes related to humans, such as feet and eyes." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17053" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.486, + 0.158 + ], + "angle": 0, + "content": "This work was supported in part by National Natural Science Foundation of China(NSFC) No.62225601, 62476029, U23B2052, and sponsored by Beijing Nova Program." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.188, + 0.188, + 0.204 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.213, + 0.482, + 0.241 + ], + "angle": 0, + "content": "[1] Hao Ban and Kaiyi Ji. Fair resource allocation in multi-task learning. arXiv preprint arXiv:2402.15638, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.243, + 0.482, + 0.312 + ], + "angle": 0, + "content": "[2] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf, 2(3):8, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.314, + 0.482, + 0.342 + ], + "angle": 0, + "content": "[3] Bing-su. Adeditor. https://github.com/Bing-su/adeditor.git, 2024.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.345, + 0.483, + 0.413 + ], + "angle": 0, + "content": "[4] Seunghwan Choi, Sunghyun Park, Minsoo Lee, and Jaegul Choo. Viton-hd: High-resolution virtual try-on via misalignment-aware normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14131-14140, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.416, + 0.482, + 0.497 + ], + "angle": 0, + "content": "[5] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.501, + 0.482, + 0.57 + ], + "angle": 0, + "content": "[6] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen Change Loy, Wayne Wu, and Ziwei Liu. Stylegan-human: A data-centric odyssey of human generation. In European Conference on Computer Vision, pages 1-19. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.572, + 0.482, + 0.641 + ], + "angle": 0, + "content": "[7] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.644, + 0.482, + 0.685 + ], + "angle": 0, + "content": "[8] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.687, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[9] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.745, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[10] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.816, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[11] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. Neural computation, 3(1):79-87, 1991. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[12] Zhen Jia, Zhang Zhang, Liang Wang, and Tieniu Tan. Human image generation: A comprehensive survey. ACM Computing Surveys, 56(11):1-39, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.213, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[13] Xuan Ju, Ailing Zeng, Jianan Wang, Qiang Xu, and Lei Zhang. Human-art: A versatile human-centric dataset bridging natural and artificial scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 618-629, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.163, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[14] Xuan Ju, Ailing Zeng, Chenchen Zhao, Jianan Wang, Lei Zhang, and Qiang Xu. Humansd: A native skeleton-guided diffusion model for human image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15988-15998, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.233, + 0.905, + 0.3 + ], + "angle": 0, + "content": "[15] Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7482-7491, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.329 + ], + "angle": 0, + "content": "[16] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.331, + 0.905, + 0.397 + ], + "angle": 0, + "content": "[17] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.4, + 0.905, + 0.483 + ], + "angle": 0, + "content": "[18] Ming Li, Taojiannan Yang, Huafeng Kuang, Jie Wu, Zhaoning Wang, Xuefeng Xiao, and Chen Chen. Controlnet++: Improving conditional controls with efficient consistency feedback: Project page: liming-ai. github.io/controlnet_plus_plus. In European Conference on Computer Vision, pages 129–147. Springer, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.484, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[19] Shikai Li, Jianglin Fu, Kaiyuan Liu, Wentao Wang, KwanYee Lin, and Wayne Wu. Cosmicman: A text-to-image foundation model for humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6955-6965, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[20] Baijiong Lin, Feiyang Ye, Yu Zhang, and Ivor W Tsang. Reasonable effectiveness of random weighting: A litmus test for multi-task learning. arXiv preprint arXiv:2111.10603, 2021.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.609, + 0.907, + 0.664 + ], + "angle": 0, + "content": "[21] Shikun Liu, Edward Johns, and Andrew J Davison. End-to-end multi-task learning with attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1871-1880, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[22] Xian Liu, Jian Ren, Aliaksandr Siarohin, Ivan Skorokhodov, Yanyu Li, Dahua Lin, Xihui Liu, Ziwei Liu, and Sergey Tulyakov. Hyperhuman: Hyper-realistic human generation with latent structural diffusion. arXiv preprint arXiv:2310.08579, 2023. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.735, + 0.905, + 0.803 + ], + "angle": 0, + "content": "[23] Wenquan Lu, Yufei Xu, Jing Zhang, Chaoyue Wang, and Dacheng Tao. Handrefiner: Refining malformed hands in generated images by diffusion-based conditional inpainting. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 7085-7093, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[24] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Guang Yong, Juhyun Lee, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Laurent Massoulié and James Roberts. Bandwidth sharing: objectives and algorithms. In IEEE INFOCOM'99. Con-" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17054" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.123, + 0.092, + 0.484, + 0.147 + ], + "angle": 0, + "content": "ference on Computer Communications. Proceedings. Eighteenth Annual Joint Conference of the IEEE Computer and Communications Societies. The Future is Now (Cat. No. 99CH36320), pages 1395-1403. IEEE, 1999. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[26] Jeonghoon Mo and Jean Walrand. Fair end-to-end window-based congestion control. IEEE/ACM Transactions on networking, 8(5):556-567, 2000. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.191, + 0.484, + 0.259 + ], + "angle": 0, + "content": "[27] Supreeth Narasimhaswamy, Thanh Nguyen, Mingzhen Huang, and Minh Hoai. Whose hands are these? hand detection and hand-body association in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4889-4899, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.483, + 0.342 + ], + "angle": 0, + "content": "[28] Supreeth Narasimhaswamy, Uttaran Bhattacharya, Xiang Chen, Ishita Dasgupta, Saayan Mitra, and Minh Hoai. Handiffuser: Text-to-image generation with realistic hand appearances. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2468-2479, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.483, + 0.399 + ], + "angle": 0, + "content": "[29] Patrick Ngatchou, Anahita Zarei, and A El-Sharkawi. Pareto multi objective optimization. In Proceedings of the 13th international conference on, intelligent systems application to power systems, pages 84-91. IEEE, 2005. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.401, + 0.484, + 0.468 + ], + "angle": 0, + "content": "[30] Georgios Pavlakos, Dandan Shan, Ilija Radosavovic, Angjoo Kanazawa, David Fouhey, and Jitendra Malik. Reconstructing hands in 3d with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9826-9836, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.47, + 0.483, + 0.537 + ], + "angle": 0, + "content": "[31] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.483, + 0.58 + ], + "angle": 0, + "content": "[32] Joan Puigcerver, Carlos Riquelme, Basil Mustafa, and Neil Houlsby. From sparse to soft mixtures of experts. arXiv preprint arXiv:2308.00951, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.483, + 0.662 + ], + "angle": 0, + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.483, + 0.718 + ], + "angle": 0, + "content": "[34] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.721, + 0.483, + 0.762 + ], + "angle": 0, + "content": "[35] J Redmon. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.763, + 0.483, + 0.831 + ], + "angle": 0, + "content": "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.483, + 0.899 + ], + "angle": 0, + "content": "[37] Axel Sauer, Tero Karras, Samuli Laine, Andreas Geiger, and Timo Aila. Stylegan-t: Unlocking the power of gans for fast large-scale text-to-image synthesis. In International conference on machine learning, pages 30105-30118. PMLR, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[38] Ozan Sener and Vladlen Koltun. Multi-task learning as multi-objective optimization. Advances in neural information processing systems, 31, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[39] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[40] Yuxuan Wang, Tianwei Cao, Kongming Liang, Zhongjiang He, Hao Sun, Yongxiang Li, and Zhanyu Ma. Mixture-of-hand-experts: Repainting the deformed hand images generated by diffusion models. In Chinese Conference on Pattern Recognition and Computer Vision (PRCV), pages 143-157. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.33 + ], + "angle": 0, + "content": "[41] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.333, + 0.905, + 0.402 + ], + "angle": 0, + "content": "[42] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36:15903-15935, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.404, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[43] Zhendong Yang, Ailing Zeng, Chun Yuan, and Yu Li. Effective whole-body pose estimation with two-stages distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4210-4220, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.905, + 0.515 + ], + "angle": 0, + "content": "[44] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. In Advances in Neural Information Processing Systems, pages 5824-5836, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.517, + 0.905, + 0.598 + ], + "angle": 0, + "content": "[45] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 5907-5915, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.602, + 0.905, + 0.669 + ], + "angle": 0, + "content": "[46] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stack-gan++: Realistic image synthesis with stacked generative adversarial networks. IEEE transactions on pattern analysis and machine intelligence, 41(8):1947-1962, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.672, + 0.905, + 0.727 + ], + "angle": 0, + "content": "[47] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.729, + 0.905, + 0.784 + ], + "angle": 0, + "content": "[48] Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jingdong Wang, and Qi Tian. Scalable person re-identification: A benchmark. In Proceedings of the IEEE international conference on computer vision, pages 1116–1124, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.786, + 0.905, + 0.84 + ], + "angle": 0, + "content": "[49] Jie Zhu, Yixiong Chen, Mingyu Ding, Ping Luo, Leye Wang, and Jingdong Wang. Mole: Enhancing human-centric text-to-image diffusion via mixture of low-rank experts. arXiv preprint arXiv:2410.23332, 2024. 1, 2, 3, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.84 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17055" + } + ] +] \ No newline at end of file diff --git a/2025/FairHuman_ Boosting Hand and Face Quality in Human Image Generation with Minimum Potential Delay Fairness in Diffusion Models/c1cadd0b-ab93-4bff-a36b-6e24c8c2c52f_origin.pdf b/2025/FairHuman_ Boosting Hand and Face Quality in Human Image Generation with Minimum Potential Delay Fairness in Diffusion Models/c1cadd0b-ab93-4bff-a36b-6e24c8c2c52f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f4112f0938314f23927a6a8dfddb751f14e6114e --- /dev/null +++ b/2025/FairHuman_ Boosting Hand and Face Quality in Human Image Generation with Minimum Potential Delay Fairness in Diffusion Models/c1cadd0b-ab93-4bff-a36b-6e24c8c2c52f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:93241c9acd7f04d2f49356ca80edac51d550f77c113bf222ea6a2d4c76770ad2 +size 7248038 diff --git a/2025/FairHuman_ Boosting Hand and Face Quality in Human Image Generation with Minimum Potential Delay Fairness in Diffusion Models/full.md b/2025/FairHuman_ Boosting Hand and Face Quality in Human Image Generation with Minimum Potential Delay Fairness in Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f6c911310936f8712364807e90407c20c8f0f960 --- /dev/null +++ b/2025/FairHuman_ Boosting Hand and Face Quality in Human Image Generation with Minimum Potential Delay Fairness in Diffusion Models/full.md @@ -0,0 +1,329 @@ +# FairHuman: Boosting Hand and Face Quality in Human Image Generation with Minimum Potential Delay Fairness in Diffusion Models + +Yuxuan Wang1 + +wangyuxuan2019@bupt.edu.cn + +Zhongjiang He² + +hezhongj_1@163.com + +Tianwei $\mathrm{Cao}^{1\dagger}$ + +caotianwei@bupt.edu.cn + +Kongming Liang $^{1\dagger}$ + +liangkongming@bupt.edu.cn + +Huayu Zhang + +zhanghy56@chinatelecom.cn + +Zhanyu Ma + +mazhanyu@bupt.edu.cn + +1 Beijing University of Posts and Telecommunications 2 Institute of Artificial Intelligence, China Telecom + +![](images/6976644bf0064501270d002d22d9fbb7ac54aa0848538f72e93f2a56ead76646.jpg) +Figure 1. Compare with existing methods under two scenarios: 1) General T2I Generation (in the left part): from left to right are SDXL-base[31], MoLE[49] and Ours. 2) Controllable Generation (in the right part): from left to right are input condition, ControlNet-SDXL[47] and Ours. Zoom in for a better view of details. + +![](images/d46b8957a8f09948832c8fadda36083bd181dc07be6976380e379060a3a97b6f.jpg) + +# Abstract + +Image generation has achieved remarkable progress with the development of large-scale text-to-image models, especially diffusion-based models. However, generating human images with plausible details, such as faces or hands, remains challenging due to insufficient supervision of local regions during training. To address this issue, we propose FairHuman, a multi-objective fine-tuning approach designed to enhance both global and local generation quality fairly. Specifically, we first construct three learning objectives: a global objective derived from the default diffusion objective function and two local objectives for hands and faces based on pre-annotated positional priors. Subse + +sequently, we derive the optimal parameter updating strategy under the guidance of the Minimum Potential Delay (MPD) criterion, thereby attaining fairness-aware optimization for this multi-objective problem. Based on this, our proposed method can achieve significant improvements in generating challenging local details while maintaining overall quality. Extensive experiments showcase the effectiveness of our method in improving the performance of human image generation under different scenarios. Our code is available at https://github.com/PRIS-CV/FairHuman. + +# 1. Introduction + +Diffusion-based models have revolutionized image generation by achieving unprecedented high quality and realism + +![](images/ad596aa3bc8fe4b1dc448f34482efd31afb58e227a3c1043b387e1af26554b5e.jpg) +Figure 2. Examples of human-related images generated by SDXL. The first row shows the close-up view images with prompts like "a close-up human face" and "a detailed hand gesture". The second row shows the human images under different scenes. Zoom in for a better view of details. + +through iterative denoising processes. However, their capability of human image generation remains challenged by persistent artifacts in anatomically complex regions, particularly faces and hands. These contents often suffer from generating degradation such as distorted facial expressions and implausible hand poses. + +To address this issue, previous work, including ControlNet[47], HumanSD[14], and HyperHuman[22], leverages anatomical information like pose or depth to guide human image generation. Moreover, a leading approach [49] fine-tunes two low-rank modules w.r.t hand and face generation, respectively. Then it uses the soft Mixture of Experts (MoE) mechanism[32] to fuse these modules. Essentially, these methods aim to enhance human image generation by injecting specific knowledge about hands and faces into the generative model. + +Despite the success of these methods, the balance between coarse appearance (global quality) and local details (quality of faces and hands) has not been addressed. Through the visual feedback of the human-related images generated by SDXL-base[31] in Fig.2, we have noticed an interesting phenomenon: When the model generates faces or hands separately as the close-up view images, it usually ensures sufficient detail and high quality in terms of structure and texture. However, when generating whole-body images especially when hand or face regions occupy a relatively small portion, noticeable distortions often occur. It suggests that the model inherently possesses the capability to generate plausible faces and hands. Nonetheless, the existence of other contents such as the background, reduces the focus on local regions, thus leading to a decline in generation quality. In addition, a post-processing method called ADetailer[3] repairs the faces and hands by cropping out the distorted regions from the original image and repainting them using super-resolution techniques. Its effectiveness further supports our hypothesis. In this way, the level of + +attention to global and local aspects is indeed a significant factor influencing the generation quality. Relevant details will be elaborated in Sec.3.1. + +Motivated by this observation, we decide to construct specific objective functions for global and local generation, respectively. We notice that the default objective function in Eq.(1) for the diffusion model is based on the global mean squared error (MSE), which pays more attention to the overall performance. For complex tasks like human image generation, a single-objective mechanism may lead to the model excelling in certain aspects while underperforming in others. Therefore, here we transform the human image generation task into a multi-objective optimization problem and divide it into three different sub-objectives. Specifically, our optimization objectives include: + +- Facial generation quality: the naturalness of facial structure and expressions. +- Hand generation quality: the plausibility of hand poses and shapes. +- Global generation quality: the coherence and perceptual quality of the overall image. + +As for the construction of objective functions, we retain the default MSE loss function for global features. Then, we create masks using pre-annotated positional information from the dataset and construct mask-based loss functions for local features, namely faces and hands. + +Considering the optimization method, we need to identify an appropriate strategy to maximize the improvement of each objective. Since faces and hands are more difficult to learn compared to global features during training, directly deploying a fixed or equal weight allocation, i.e., the degree of importance, is unreasonable. To achieve the Pareto optimality of human image generation performance, inspired by the idea of fairness criteria utilized in communication networks[26], we incorporate the minimum potential delay (MPD) fairness principle[25] into our optimization process for parameter updates. Specifically, it dynamically allocates the weight for each objective-specific gradient following the fairness constraints. In this way, the three objective functions mentioned above can be optimized simultaneously within a single diffusion model while mitigating potential conflicts between them. Technically, it can be deployed on fine-tuning different modules like LoRA[9] and ControlNet[47] for certain generation tasks. + +To this end, we propose a framework called FairHuman to generate human images with higher fidelity and quality under different scenarios. The key insight is that human image generation is a complex task that involves both global and local optimization. The previous single-objective methodology cannot improve the performance of different aspects at the same time. Therefore, we regard the learning for human image generation as a multi-objective optimization process and implement the MPD fairness prin + +ciple to achieve a more effective and balanced performance. + +Our method contains three procedures (see Fig.3): Firstly, we construct separate objective functions for global and regional generation by utilizing the default MSE loss function and regional masks obtained from the preannotated positional information. Secondly, we implement the minimum potential delay fairness principle for the joint optimization of all objectives. Eventually, we fine-tune LoRA[9] and ControlNet[47] based on the mentioned strategy for two scenarios of human image generation. + +In summary, the contributions of our paper are as follows: + +- We emphasize that the poor performance in generating local details of human images stems from insufficient attention during the training process. Based on this factor, we construct different objective functions to focus on global and local generation separately. +- We propose FairHuman, a general multi-objective optimization framework for diffusion model in improving the performance of human image generation. Here, we optimize different objective functions simultaneously within a single model based on the minimum potential delay fairness principle. To our knowledge, it has not been applied in the diffusion model before. +- We implement a widely adopted diffusion model SDXL[31] as the backbone and deploy our method on different modules, including LoRA[9] and ControlNet[47]. Extensive experimental results demonstrate that our method yields a superior performance in enhancing the generation of human images under diverse scenarios. + +# 2. Related Work + +Human Image Generation. Human image generation refers to generating high-quality human images complying with the given conditions (e.g., text, image). Among image synthesis tasks, it has garnered significant attention due to its considerable academic research values and application potentials[12]. Driven by advancements in scalability and training techniques, existing text-to-image models have demonstrated the ability to generate images with increasing resolution and fidelity. Following the success of GANs[37, 45, 46], diffusion models[8, 39] has emerged as a powerful alternative and become the predominant architecture currently. Models like Stable Diffusion[36] and DALL-E 2[34] leverage latent diffusion processes to generate highly detailed and semantically accurate images from textual descriptions. However, when creating human-centric images, the latest models[2, 5, 31] still struggle to generate fine-grained human parts, especially faces and hands. Previous approaches such as[14, 22, 23, 40] incorporate supplementary conditions like human pose to achieve better generation in a controllable pattern but have major limitations in consistency. A concurrent work[49] deals + +with the issue of faces and hands by utilizing two low-rank modules[9] with the Mixture of Experts(MoE)[11] mechanism. However, it requires fine-tuning each expert module separately, which brings additional training costs. Meanwhile, stacking of multiple modules also affects the generation efficiency. In our work, we implement joint optimization for both global and local generation performance following the minimum potential delay fairness principle. + +Datasets for Human Image Generation. The availability of high-quality datasets has significantly propelled the development of human image generation models. These datasets[4, 6, 10, 13, 48] provide the necessary diversity and richness required for training models in generating realistic and varied human images. However, most of them have two main drawbacks: 1) low resolution and poor image quality, which are inadequate to provide valid prior knowledge for diffusion models. 2) lack of annotations on morphological and spatial information of the human body, which cannot effectively guide the model in generating delicate parts like human faces and hands. To alleviate the above limitations, we collect images from available high-quality human-centric datasets[19, 27] that contain high-resolution real-world images. To further guarantee the image quality, we apply detectors[30, 35, 43] to filter out undesirable images while maintaining as much valuable information as possible. In addition, we annotate the body information for controllable generation and positional information for constructing mask-based objective functions. Implementing the procedures mentioned above, we can achieve enhanced training performance aligned with our objectives. + +# 3. Methodology + +# 3.1. Problem Statement + +The objective function of diffusion models is founded on the mean square error associated with the denoising task. It enhances the model's performance by minimizing the discrepancy between the predicted noise and the actual noise. In the case of Latent Diffusion Models (LDM)[36], this denoising process is executed within the latent space, facilitated by a VAE encoder[16]. This approach allows for more efficient and effective denoising by operating in a compressed representation of the data. Specifically, its optimization objective can be formulated as: + +$$ +\min _ {\theta} E _ {z _ {0}, t, c, \epsilon} \left[ \| \epsilon - \hat {\epsilon} _ {\theta} \left(z _ {t}, t, c\right) \| ^ {2} \right], \tag {1} +$$ + +where $z_{t}$ is the noisy latent at timestep $t$ obtained from $z_{0}$ through the diffusion process by adding Gaussian noise $\epsilon \sim N(0,1)$ ; the initial representation $z_{0}$ is defined as $z_{0} = \varepsilon(x_{0})$ , with $\varepsilon$ denoting the encoder and $x_{0}$ being the ground truth image; $\hat{\epsilon}_{\theta}(z_{t},t,c)$ is the noise predicted by the diffusion model; $c$ represents the embedding of input + +![](images/0c11849a7c4bab38bb60d8e770aacdea8c9b60c8c9cc3e9d8b97d8babe418075.jpg) +Figure 3. An overview of FairHuman framework. The entire pipeline is divided into three stages: 1) Construct multi objectives for human image generation; 2) Employ fairness-aware optimization strategy for parameter updates; 3) Apply the fine-tuned module into the backbone diffusion model. Here, we take LoRA[9] and ControlNet[47] as examples of different generation scenarios. + +conditions (e.g., textual prompt or image); $\theta$ represents the parameters to be updated. + +By pre-training on large-scale datasets with this objective function, models such as SDXL[31] have shown the ability to generate high-quality human content. This is evident in the first row of Fig.2, where close-up images are generated accurately. However, when generating full-body human images within a scene (see the second row of Fig.2), the model demonstrates inferior performance in rendering local details (faces and hands). In other words, when face or hand regions occupy a larger proportion of the image, their generation results are acceptable. On the contrary, when they occupy a smaller proportion of the image, their corresponding generation quality would be worse. + +What accounts for the occurrence of this issue? As indicated by Eq.(1), LDM calculates the loss based on the representation $z_{t}$ encoded from the entire image without distinguishing the optimization difficulty of different regions. For faces and hands with larger areas in close-up images, the default objective function can be easily optimized since these regions contribute more significantly to the loss. Conversely, for human-in-the-scene images, the representation $z_{t}$ comprises not only the hands and faces but also other body details and the background scene. Therefore, face and hand regions with smaller areas contribute less to the loss, thus leading to insufficient attention and resulting in poorer generation quality. In this way, different regions may compete during the generation process. + +# 3.2. Multi-objectives + +As we discussed above, the competition between different regions may lead to insufficient attention to some local areas, such as hands or faces. To address this issue, the generative model requires a fair allocation of resources to both local and global regions, thus enhancing the overall generation performance. To this end, we first construct the following optimization objectives for both local and global generation tasks, respectively: + +$$ +l _ {\text {g l o b a l}} (\theta) = \mathbb {E} _ {z _ {0}, t, c, \epsilon} \left[ \| \epsilon - \hat {\epsilon} _ {\theta} (z _ {t}, t, c) \| ^ {2} \right]; +$$ + +$$ +l _ {\text {f a c e}} (\theta) = \mathbb {E} _ {z _ {0}, t, c, \epsilon} \left[ \| m _ {f} \odot \left(\epsilon - \hat {\epsilon} _ {\theta} (z _ {t}, t, c)\right) \| ^ {2} \right]; \tag {2} +$$ + +$$ +l _ {\mathrm {h a n d}} (\theta) = \mathbb {E} _ {z _ {0}, t, c, \epsilon} \left[ \| m _ {h} \odot (\epsilon - \hat {\epsilon} _ {\theta} (z _ {t}, t, c)) \| ^ {2} \right], +$$ + +where $\theta$ represents the learnable parameters; $l_{global}(\theta)$ is the global loss for entire image generation; $l_{face}(\theta)$ and $l_{hand}(\theta)$ are local losses focusing on faces and hands; $m_f$ and $m_h$ are the masks used to select the face and hand regions respectively. Here, the masks $m_f$ and $m_h$ are obtained by reshaping the mask image to the latent shape, following inpainting mask $m$ denoted in Eq.(4) of [23]. For brevity, we also denote $l_{global}(\theta), l_{face}(\theta), l_{hand}(\theta)$ as $l_1(\theta), l_2(\theta), l_3(\theta)$ in the subsequent sections. + +Based on this, we aim to fine-tune a diffusion model to simultaneously minimize $l_{1}(\theta)$ , $l_{2}(\theta)$ , and $l_{3}(\theta)$ . Meanwhile, the fine-tuning approach should enhance the generation quality of hands and faces while maintaining its global + +performance. In this way, the fine-tuning task could be formulated as the following multi-objective optimization problem: + +$$ +\min _ {\theta} \mathcal {L} (\theta) = \left[ l _ {1} (\theta), l _ {2} (\theta), l _ {3} (\theta) \right] ^ {\top}. \tag {3} +$$ + +Notably, our primary target comprises two key aspects: + +- Regional Fairness: We should ensure that the solution $\theta^{*}$ to Eq.(3) yields a fairly substantial improvement across all three objectives, especially for $l_{2}(\theta)$ and $l_{3}(\theta)$ . As long as there is at least one insignificant improvement among the three objectives, it implies that our fine-tuning is ineffective. This aligns with our purpose of fair enhancement on the quality w.r.t face, hand, and global regions. +- Pareto Optimality: The solution $\theta^{*}$ to Eq.(3) should be Pareto optimal[29, 38]. Here we say a solution $\theta^{*}$ is Pareto optimal if there exists no other solution $\theta^{\prime}$ satisfying $l_{i}(\theta^{\prime})\leq l_{i}(\theta^{*})$ for all $i\in \{1,2,3\}$ . In other words, updating $\theta^{*}$ in any direction will disrupt at least one of the three objectives. Thus Pareto optimality here implies a trade-off between face, hand, and global quality. + +# 3.3. Fairness-aware Optimization + +We aim to ensure the post fine-tuning values of $l_{1}(\theta), l_{2}(\theta), l_{3}(\theta)$ are all fairly minimized to a sufficient degree. To this end, a straightforward approach is to enable each $l_{i}(\theta)$ could be maximally decreased at each training step of $\theta$ . Formally, we seek a stepwise updating direction $d$ that maximizes the contributions of $d$ to all of $l_{1}(\theta), l_{2}(\theta)$ and $l_{3}(\theta)$ . From another perspective, this is equivalent to minimizing the negative effect of $d$ to all these objectives. + +Here, we employ the concept of Potential Delay [25, 26] to quantitatively describe the negative effect. The term "Potential Delay" originally refers to the expected time required for a user to complete a specific task in the context of resource allocation. For example, if an internet user transmits a unit-sized file at transmission rate $v$ , then the potential delay is $1 / v$ . In this paper, the common updating direction $d$ can be considered a shared resource among all objectives. The "transmission rate" of this resource to each $l_{i}(\theta)$ is represented by the projection magnitude of $d$ onto the steepest descent direction $\nabla l_{i}(\theta)$ , denoted as $\| \mathrm{proj}_{\nabla l_i(\theta)}(d)\|$ . Consequently, the corresponding potential delay is $1 / \| \mathrm{proj}_{\nabla l_i(\theta)}(d)\|$ . Normally, projection $\| \mathrm{proj}_{\nabla l_i(\theta)}(d)\|$ can serve as a metric for the contribution of $d$ to the optimization process of $l_{i}(\theta)$ [44]. In this way, the potential delay $1 / \| \mathrm{proj}_{\nabla l_i(\theta)}(d)\|$ is inversely proportional to the contribution from $d$ and thus directly proportional to the negative effect. In this way, we can determine the optimal direction $d$ by resolving the subsequent minimization problem: + +$$ +\min _ {d} \mathcal {F} (d) = \sum_ {i = 1} ^ {3} \frac {1}{\| \operatorname {p r o j} _ {\nabla l _ {i} (\theta)} (d) \|}, \tag {4} +$$ + +By optimizing Eq.(4), the potential delay of $l_{1}(\theta)$ , $l_{2}(\theta)$ , $l_{3}(\theta)$ can be minimized simultaneously, thus it is termed as Minimal Potential Delay (MPD) Fairness [25, 26]. + +Prior to solving this optimization, we assume that $d$ lies within a ball with radius $r$ centered at the origin. Based on this, the optimization problem w.r.t MPD fairness can be transformed into the following form: + +$$ +\min _ {d} \mathcal {F} ^ {\prime} (d) = \sum_ {i = 1} ^ {3} \frac {1}{\nabla l _ {i} (\theta) ^ {\top} d}, \tag {5} +$$ + +since that $\mathcal{F}'(d)$ is proportional to an upper bound of $\mathcal{F}(d)$ in Eq.(4), and the proof is provided in the Sec.1.1 of the supplementary material. Here, $\mathcal{F}'(d)$ can be regarded as a special case of $\alpha$ -fairness[26]. If we set $\alpha = 2$ for Eq.(2) in [1], it will be equivalent to Eq.(5) in this paper. Thus, we can adopt a similar procedure to solve Eq.(5) following [1]. + +Specifically, we first apply Lagrange multiplier method on Eq.(5) to obtain the following equation: + +$$ +d = \sum_ {i = 1} ^ {3} \frac {\nabla l _ {i} (\theta)}{(\nabla l _ {i} (\theta) ^ {\top} d) ^ {2}}, \tag {6} +$$ + +The derivation process is described in Sec.1.2 of the supplementary material. To further simplify the problem, we follow [1] and express $d$ as a linear combination: + +$$ +d = \sum_ {i = 1} ^ {3} w _ {i} \nabla l _ {i} (\theta), \tag {7} +$$ + +where $w_{i}$ represents the weight of the $i$ -th objective. Given that each gradient $\nabla l_{i}(\theta)$ is readily available, the problem of determining $d$ is reformulated as determining every $w_{i}$ . By substituting Eq.(7) into Eq.(6), we can get that: + +$$ +G (\theta) ^ {\top} G (\theta) W = W ^ {- 1 / 2}, \tag {8} +$$ + +where $G(\theta) = [\nabla l_1(\theta), \nabla l_2(\theta), \nabla l_3(\theta)]^\top$ and $W = [w_1, w_2, w_3]^\top$ . According to the relation in Eq.(8), we can further construct the following convex optimization problem: + +$$ +\min _ {W} \left\| G (\theta) ^ {\top} G (\theta) W - W ^ {- 1 / 2} \right\| ^ {2}, \tag {9} +$$ + +By applying the standard least square method, we can obtain the following close-form solution of Eq.(9): + +$$ +W = \left(G (\theta) ^ {\top} G (\theta)\right) ^ {- \frac {2}{3}} \mathbf {1}, \tag {10} +$$ + +where $\mathbf{1}$ represents an all-ones vector. With these weights, we can finally obtain the value of $d$ by using Eq.(7) and then conduct an updating step as: + +$$ +\theta_ {\text {n e w}} = \theta - \eta d, \tag {11} +$$ + +where $\theta_{new}$ is the updated parameter and $\eta$ is the learning rate. By adopting this approach, each training step is guided by MPD fairness, thereby ensuring the efficacy of the finetuning results across all objectives. Meanwhile, Thm.(7.3) in [1] also indicates that such a training process can converge to a Pareto optimal solution. And our constructed objective function decouples regional and global optimization, satisfying MPD's prerequisite assumption of linearly independent task gradients. + +# 3.4. Application + +Considering the high cost of full fine-tuning for diffusion models, we opt to deploy our method in LoRA[9] and ControlNet[47] while freezing the parameters of the backbone model. They represent two types of generation patterns, namely, general T2I generation and controllable generation with control signals. Specifically, LoRA primarily enhances the quality of human images from the perspective of visual concepts and pixel information, while ControlNet focuses on more refined and controllable generation over the structural aspects. Both have significant influence and play important roles in the field of human image generation. We construct a high-quality dataset for their fine-tuning. Apart from the images and text descriptions, it also includes the positional information for creating mask-based objective functions and the pose-depth information for controllable generation. Their new outputs can be formulated as: + +$$ +\hat {y} _ {\text {n e w}} = N _ {\text {b a s e}} (z; \theta_ {\text {b a s e}}) + \beta N _ {\text {n e w}} (c, z; \theta_ {\text {u p d a t e}}), \tag {12} +$$ + +where $z$ is the input latent and $N_{base}$ is the backbone network with the frozen parameters $\theta_{base}$ . $N_{new}$ represents the fine-tuned module (LoRA or ControlNet here) with the updated parameters $\theta_{update}$ . As for condition $c$ , it contains textual and timestep embeddings for the LoRA module and has additional control signals for the ControlNet module. $\beta$ is the hyperparameter used for scaling. + +# 4. Experiments + +# 4.1. Experimental Settings + +Datasets. We collect human-centric images from two public datasets called CosmicMan[19] and BodyHands[27]. To enable effective training, we implement several preprocessing methods, including filtering, captioning, and labeling. Ultimately, we construct a high-quality human image dataset that contains approximately 50k images for training. As for the validation, we randomly sample about 5k in-the-wild human images under different scenes with well-processed captions provided by [49]. The details regarding the construction of our dataset will be elaborated in Sec.2 of the supplementary material. + +Baseline Methods. We compare our approach with two + +categories of baseline methods. 1) General T2I methods without extra conditions: SDXL-base[31] and MoLE[49]. Notably, MoLE is a recent work based on the collaboration of multiple LoRA modules. 2) Controllable methods with extra conditions: ControlNet-SDXL[47] and ControlNet Union[18]. Here, we compare performance under the same given conditions (depth and pose) using different training strategies. + +Metrics. We adopt evaluation metrics based on the optimization objectives outlined in the paper, focusing on two dimensions. 1) Overall image quality: Human Preference Score (HPS)[41] and ImageReward (IR)[42]. They are built on the CLIP[33] and BLIP[17] models respectively to reflect human visual preferences on the generated images from multiple aspects including image alignment, aesthetic, and plausibility. Moreover, CLIP score is applied to evaluate text-image alignment, and pose alignment is used for assessing the effectiveness of controllable generation. 2) Regional generation quality: FID[7] and detection confidence of mediapipe[24] are utilized to evaluate the quality of hands and faces in terms of realism and plausibility. The implementation details about our experiments are provided in Sec.3 of the supplementary material. + +# 4.2. Comparison with Existing Methods + +Quantitative Analysis. The evaluation results compared with existing methods are shown in Tab.1. We perform evaluations under two classic scenarios. Specifically, for general T2I methods, including SDXL-Base[31], MoLE[49], and our LoRA-based method, we randomly sample 5k prompts about human-in-the-scene images from the validation dataset and feed them to each model for image generation. As for the setting of hyperparameters, we keep the default CFG scale of 7.5, set the inference step to 30, and fix the seed at 0 for all methods. In particular, for our LoRA-based method, we implement three different weights (0.3, 0.4, and 0.5) in the experiments and use their average metric values as the final result. To more comprehensively evaluate the quality of generated images, we utilize the metrics in terms of overall image quality, text-image alignment, and regional quality. The results demonstrate that our method outperforms existing approaches in all these three aspects, especially in the generation quality of local hand and face details. Since our LoRA-based method and MoLE are deployed on SDXL sharing the same backbone model, we also conduct experiments on their memory usage and inference speed. Results suggest that our method can achieve better performance while requiring less memory size and inference time. Related details are demonstrated in Sec.4.1 of the supplementary material. + +Furthermore, to verify the effectiveness of our method across different task scenarios, we employ it in ControlNet[47] for the controllable generation. As shown + +
Application ScenarioModelImage QualityText-Image AlignmentRegional QualityPose Alignment
HPS(%) ↑IR(%) ↑CLIP ↑FID ↓Hand Confi. ↑Face Confi. ↑Mean Diff ↓
General T2I MethodsSD-XL Base [31]30.91140.8526.8452.6993.1883.67/
MoLE [49]31.87152.3627.1748.9793.3284.48/
Ours32.73154.6027.2046.1194.0385.31/
Controllable MethodsControlNet-SDXL [47]32.20152.4927.1046.1893.5085.133.45
ControlNet-Union [18]32.80155.2227.2146.4394.2885.362.67
ControlNet-Ours32.72154.7427.2445.4694.4885.612.35
+ +Table 1. The evaluation results for ours and other existing methods. We conduct comparisons under two different image generation scenarios separately. As for the general T2I methods, we compare our LoRA-based method with a widely adopted diffusion model SDXL[31] and a recent work MoLE[49]. For controllable methods, we compare our ControlNet-based method with the default ControlNet-SDXL[47] and the latest ControlNet-Union[18]. Notably, the mean difference of pose condition is employed here to evaluate the effectiveness of controllable generation. The best results are indicated in bold. + +in the lower part of Tab.1, we compare with existing controllable methods, including the default ControlNet-SDXL and the latest variant ControlNet-Union[18]. The evaluation is conducted using the same metrics as described above. Similar to our LoRA-based method, we set three different control scales (0.4, 0.5, and 0.6) here for testing and choose their average metric values. Besides, considering the characteristics of controllable generation, we also evaluate the model's performance based on pose alignment, i.e., the consistency between the input conditions and the actual generated content. Here, we use the mean difference between the input and extracted conditions as the metric. The corresponding results are shown in the last column of Tab.1. In comparison, our method still demonstrates superior performance in the controllable generation scenario. Although it is slightly inferior to ControlNet-Union in terms of HPS and IR metrics, the primary reason is that ControlNet-Union incorporates aesthetic and preference feedback learning during training, whereas our method focuses more on the realism and plausibility of the image, and significantly outperforms in regional details. + +![](images/d0e9f0d34f94e39244f2e7493307ecc663a019bc01f79d6e51601148bb29ae36.jpg) +Figure 4. User study results in terms of overall quality and regional quality. We present the ratio of users' preference for each method. + +Qualitative Analysis. Fig.1 shows qualitative results. Compared with baseline methods, our method can generate human images with higher quality in both global and local aspects. Meanwhile, it achieves better controllability and plausibility in terms of controllable generation. We also + +conduct a user study to provide a more intuitive evaluation. Volunteers are invited to select images that better align with their visual preferences based on overall and regional quality. The result is shown in Fig.4. It indicates that our method is more preferred by users in both two aspects. More details about the user study will be demonstrated in Sec.4.2 of the supplementary material. + +![](images/2e31b412e8f2e3796a6fc85c89f8928707202cbc0a41022163bbc6d270f4be89.jpg) +Figure 5. Examples of the images generated by different settings with the same prompt. First row: generated by LoRA. Second row: generated by ControlNet. Zoom in for a better view of details. + +# 4.3. Ablation Studies + +We present the key ablation studies based on the construction of our objective functions and the strategy for optimization. Here, we analyze their effects according to the overall and local quality of the generated images, deploying on LoRA and ControlNet, respectively. We randomly select 1k human-related prompts and follow the same hyperparameter setting as mentioned above for testing. Their results are shown in Tab.2. In addition, some visualization results are + +
ModelAblation SettingsImage QualityRegional QualityPose Alignment
HPS(%)↑IR(%)↑FID ↓Hand Confi. ↑Face Confi. ↑Mean Diff ↓
LoRAl1(θ)32.40151.0849.4892.8183.84/
l1(θ) + l2(θ) + l3(θ) (w/o fairness)32.60153.2347.2293.7084.82/
l2(θ) + l2(θ) + l3(θ) (w fairness)32.73154.6046.1194.0385.31/
ControlNetl1(θ)32.20152.4946.1893.5085.133.45
l1(θ) + l2(θ) + l3(θ) (w/o fairness)32.56154.3046.7993.9985.242.78
l1(θ) + l2(θ) + l3(θ) (w fairness)32.72154.7445.3694.4885.612.35
+ +Table 2. Ablation results deployed on LoRA and ControlNet. We primarily explore the effect of two aspects: additional objective functions for regional generation and the learning strategy. $l_{1}(\theta)$ , $l_{2}(\theta)$ , and $l_{3}(\theta)$ are defined in Eq.2. The best results are indicated in bold. + +
LoRA RankImage QualityT2I AlignmentRegional Quality
HPS(%) ↑IR(%) ↑CLIP↑FID ↓Hand Confi. ↑Face Confi. ↑
r=6432.50148.5227.1147.8793.2584.69
r=12832.73154.827.1546.7993.7985.11
r=25632.73154.6027.2046.1194.0385.31
+ +Table 3. Comparison results for the choice of LoRA rank. The best results are indicated in bold. + +provided in Fig.5. + +Effect of Regional Optimization Objective. To explore how regional loss functions including $l_{\text{face}}(\theta)$ and $l_{\text{hand}}(\theta)$ in Eq.(2) influence generation performance, we fine-tuned the corresponding LoRA and ControlNet by adopting the default loss function from Eq.(1) and the joint objective functions in Eq.(2), respectively. For the latter, we employ a fixed-weight linear scalarization method here, transforming the multi-objective optimization into a single-objective optimization. The first two rows for each model in Tab.2 present the evaluation results, validating that the inclusion of regional loss functions helps the model improve its generation capabilities for details such as hands and faces, thereby enhancing the overall quality of the images. + +Effect of Fairness Principle. In our work, we utilize the minimum potential delay (MPD) fairness to achieve better optimization based on the joint of objective functions for global and regional generation. Here, we compare two gradient update strategies: one without considering fairness (i.e., a fixed weight allocation based on linear scalarization method) and one considering fairness (i.e., a dynamical weight allocation based on MPD fairness). As shown in the last two rows for each model in Tab.2, the optimization strategy based on the MPD fairness can better focus on the more challenging objectives of generating faces and hands without compromising the overall quality. It achieves a more balanced performance among different objectives. + +# 4.4. More Analysis + +Considering that the rank setting significantly impacts the LoRA performance, we additionally conduct a hyperparam- + +eter study to figure out the optimal rank configuration. The results are presented in Tab.3, which indicates that setting the rank value to 256 maximizes the model's ability to generate regional details without excessively jeopardizing the overall image quality. Besides, we also conduct comparison studies on some other multi-objective learning strategies, e.g., Dynamic Weight Average (DwA)[21], Uncertainty Weighting (UW)[15], and Random Loss Weighting (RLW)[20]. The results suggest that applying the MPD fairness principle to the optimization process can achieve a more balanced and superior performance in terms of the human image generation task. Relevant results and details are presented in Sec.4.3 of the supplementary material. + +# 4.5. Visualizations + +We put more visualization results compared with existing methods in Sec.5 of the supplementary material, including other close-source methods like HyperHuman[22] and HanDiffuser[28]. Our method still demonstrates comparable or superior performance. + +# 5. Conclusion + +In this work, we propose FairHuman to enhance the quality of human images generated by diffusion models. To address the distortion issue in local details like faces and hands, we specifically construct corresponding objective functions for regional generation. Then, we combine the default global objective function to achieve joint training. The key of our method is regarding the improvement of human image generation as a multi-objective optimization task. By utilizing the minimum potential delay principle and the gradient manipulation process, our method enables more balanced and effective performance. Extensive experiments demonstrate the superiority of our method under different scenarios of human image generation. For future work, we plan to further design and optimize corresponding objective functions for more attributes related to humans, such as feet and eyes. + +# Acknowledgements + +This work was supported in part by National Natural Science Foundation of China(NSFC) No.62225601, 62476029, U23B2052, and sponsored by Beijing Nova Program. + +# References + +[1] Hao Ban and Kaiyi Ji. Fair resource allocation in multi-task learning. arXiv preprint arXiv:2402.15638, 2024. 5, 6 +[2] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf, 2(3):8, 2023. 3 +[3] Bing-su. Adeditor. https://github.com/Bing-su/adeditor.git, 2024.2 +[4] Seunghwan Choi, Sunghyun Park, Minsoo Lee, and Jaegul Choo. Viton-hd: High-resolution virtual try-on via misalignment-aware normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14131-14140, 2021. 3 +[5] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 3 +[6] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen Change Loy, Wayne Wu, and Ziwei Liu. Stylegan-human: A data-centric odyssey of human generation. In European Conference on Computer Vision, pages 1-19. Springer, 2022. 3 +[7] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6 +[8] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3 +[9] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 3, 4, 6 +[10] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013. 3 +[11] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. Neural computation, 3(1):79-87, 1991. 3 +[12] Zhen Jia, Zhang Zhang, Liang Wang, and Tieniu Tan. Human image generation: A comprehensive survey. ACM Computing Surveys, 56(11):1-39, 2024. 3 + +[13] Xuan Ju, Ailing Zeng, Jianan Wang, Qiang Xu, and Lei Zhang. Human-art: A versatile human-centric dataset bridging natural and artificial scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 618-629, 2023. 3 +[14] Xuan Ju, Ailing Zeng, Chenchen Zhao, Jianan Wang, Lei Zhang, and Qiang Xu. Humansd: A native skeleton-guided diffusion model for human image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15988-15998, 2023. 2, 3 +[15] Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7482-7491, 2018. 8 +[16] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3 +[17] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 6 +[18] Ming Li, Taojiannan Yang, Huafeng Kuang, Jie Wu, Zhaoning Wang, Xuefeng Xiao, and Chen Chen. Controlnet++: Improving conditional controls with efficient consistency feedback: Project page: liming-ai. github.io/controlnet_plus_plus. In European Conference on Computer Vision, pages 129–147. Springer, 2024. 6, 7 +[19] Shikai Li, Jianglin Fu, Kaiyuan Liu, Wentao Wang, KwanYee Lin, and Wayne Wu. Cosmicman: A text-to-image foundation model for humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6955-6965, 2024. 3, 6 +[20] Baijiong Lin, Feiyang Ye, Yu Zhang, and Ivor W Tsang. Reasonable effectiveness of random weighting: A litmus test for multi-task learning. arXiv preprint arXiv:2111.10603, 2021.8 +[21] Shikun Liu, Edward Johns, and Andrew J Davison. End-to-end multi-task learning with attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1871-1880, 2019. 8 +[22] Xian Liu, Jian Ren, Aliaksandr Siarohin, Ivan Skorokhodov, Yanyu Li, Dahua Lin, Xihui Liu, Ziwei Liu, and Sergey Tulyakov. Hyperhuman: Hyper-realistic human generation with latent structural diffusion. arXiv preprint arXiv:2310.08579, 2023. 2, 3, 8 +[23] Wenquan Lu, Yufei Xu, Jing Zhang, Chaoyue Wang, and Dacheng Tao. Handrefiner: Refining malformed hands in generated images by diffusion-based conditional inpainting. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 7085-7093, 2024. 3, 4 +[24] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Guang Yong, Juhyun Lee, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, 2019. 6 +[25] Laurent Massoulié and James Roberts. Bandwidth sharing: objectives and algorithms. In IEEE INFOCOM'99. Con- + +ference on Computer Communications. Proceedings. Eighteenth Annual Joint Conference of the IEEE Computer and Communications Societies. The Future is Now (Cat. No. 99CH36320), pages 1395-1403. IEEE, 1999. 2, 5 +[26] Jeonghoon Mo and Jean Walrand. Fair end-to-end window-based congestion control. IEEE/ACM Transactions on networking, 8(5):556-567, 2000. 2, 5 +[27] Supreeth Narasimhaswamy, Thanh Nguyen, Mingzhen Huang, and Minh Hoai. Whose hands are these? hand detection and hand-body association in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4889-4899, 2022. 3, 6 +[28] Supreeth Narasimhaswamy, Uttaran Bhattacharya, Xiang Chen, Ishita Dasgupta, Saayan Mitra, and Minh Hoai. Handiffuser: Text-to-image generation with realistic hand appearances. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2468-2479, 2024. 8 +[29] Patrick Ngatchou, Anahita Zarei, and A El-Sharkawi. Pareto multi objective optimization. In Proceedings of the 13th international conference on, intelligent systems application to power systems, pages 84-91. IEEE, 2005. 5 +[30] Georgios Pavlakos, Dandan Shan, Ilija Radosavovic, Angjoo Kanazawa, David Fouhey, and Jitendra Malik. Reconstructing hands in 3d with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9826-9836, 2024. 3 +[31] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 2, 3, 4, 6, 7 +[32] Joan Puigcerver, Carlos Riquelme, Basil Mustafa, and Neil Houlsby. From sparse to soft mixtures of experts. arXiv preprint arXiv:2308.00951, 2023. 2 +[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 6 +[34] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 3 +[35] J Redmon. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2016. 3 +[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3 +[37] Axel Sauer, Tero Karras, Samuli Laine, Andreas Geiger, and Timo Aila. Stylegan-t: Unlocking the power of gans for fast large-scale text-to-image synthesis. In International conference on machine learning, pages 30105-30118. PMLR, 2023. 3 + +[38] Ozan Sener and Vladlen Koltun. Multi-task learning as multi-objective optimization. Advances in neural information processing systems, 31, 2018. 5 +[39] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3 +[40] Yuxuan Wang, Tianwei Cao, Kongming Liang, Zhongjiang He, Hao Sun, Yongxiang Li, and Zhanyu Ma. Mixture-of-hand-experts: Repainting the deformed hand images generated by diffusion models. In Chinese Conference on Pattern Recognition and Computer Vision (PRCV), pages 143-157. Springer, 2024. 3 +[41] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023. 6 +[42] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36:15903-15935, 2023. 6 +[43] Zhendong Yang, Ailing Zeng, Chun Yuan, and Yu Li. Effective whole-body pose estimation with two-stages distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4210-4220, 2023. 3 +[44] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. In Advances in Neural Information Processing Systems, pages 5824-5836, 2020. 5 +[45] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 5907-5915, 2017. 3 +[46] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stack-gan++: Realistic image synthesis with stacked generative adversarial networks. IEEE transactions on pattern analysis and machine intelligence, 41(8):1947-1962, 2018. 3 +[47] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 2, 3, 4, 6, 7 +[48] Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jingdong Wang, and Qi Tian. Scalable person re-identification: A benchmark. In Proceedings of the IEEE international conference on computer vision, pages 1116–1124, 2015. 3 +[49] Jie Zhu, Yixiong Chen, Mingyu Ding, Ping Luo, Leye Wang, and Jingdong Wang. Mole: Enhancing human-centric text-to-image diffusion via mixture of low-rank experts. arXiv preprint arXiv:2410.23332, 2024. 1, 2, 3, 6, 7 \ No newline at end of file diff --git a/2025/FairHuman_ Boosting Hand and Face Quality in Human Image Generation with Minimum Potential Delay Fairness in Diffusion Models/images.zip b/2025/FairHuman_ Boosting Hand and Face Quality in Human Image Generation with Minimum Potential Delay Fairness in Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..32f017f832ef5cf4538099d33e9199219d842891 --- /dev/null +++ b/2025/FairHuman_ Boosting Hand and Face Quality in Human Image Generation with Minimum Potential Delay Fairness in Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ee68875b2f2a89c345478a0b537ce33c00534b75f7672b8db8836197678623ff +size 614418 diff --git a/2025/FairHuman_ Boosting Hand and Face Quality in Human Image Generation with Minimum Potential Delay Fairness in Diffusion Models/layout.json b/2025/FairHuman_ Boosting Hand and Face Quality in Human Image Generation with Minimum Potential Delay Fairness in Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3b49c1e54c0948679dd4f1e2859ba3e797986333 --- /dev/null +++ b/2025/FairHuman_ Boosting Hand and Face Quality in Human Image Generation with Minimum Potential Delay Fairness in Diffusion Models/layout.json @@ -0,0 +1,8601 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 73, + 73, + 538, + 109 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 73, + 538, + 109 + ], + "spans": [ + { + "bbox": [ + 73, + 73, + 538, + 109 + ], + "type": "text", + "content": "FairHuman: Boosting Hand and Face Quality in Human Image Generation with Minimum Potential Delay Fairness in Diffusion Models" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 114, + 131, + 188, + 146 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 131, + 188, + 146 + ], + "spans": [ + { + "bbox": [ + 114, + 131, + 188, + 146 + ], + "type": "text", + "content": "Yuxuan Wang1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 80, + 148, + 223, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 148, + 223, + 159 + ], + "spans": [ + { + "bbox": [ + 80, + 148, + 223, + 159 + ], + "type": "text", + "content": "wangyuxuan2019@bupt.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 113, + 165, + 192, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 165, + 192, + 180 + ], + "spans": [ + { + "bbox": [ + 113, + 165, + 192, + 180 + ], + "type": "text", + "content": "Zhongjiang He²" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 182, + 202, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 182, + 202, + 193 + ], + "spans": [ + { + "bbox": [ + 104, + 182, + 202, + 193 + ], + "type": "text", + "content": "hezhongj_1@163.com" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 272, + 131, + 343, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 131, + 343, + 144 + ], + "spans": [ + { + "bbox": [ + 272, + 131, + 343, + 144 + ], + "type": "text", + "content": "Tianwei " + }, + { + "bbox": [ + 272, + 131, + 343, + 144 + ], + "type": "inline_equation", + "content": "\\mathrm{Cao}^{1\\dagger}" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 247, + 148, + 368, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 148, + 368, + 159 + ], + "spans": [ + { + "bbox": [ + 247, + 148, + 368, + 159 + ], + "type": "text", + "content": "caotianwei@bupt.edu.cn" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 252, + 164, + 345, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 164, + 345, + 180 + ], + "spans": [ + { + "bbox": [ + 252, + 164, + 345, + 180 + ], + "type": "text", + "content": "Kongming Liang" + }, + { + "bbox": [ + 252, + 164, + 345, + 180 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 231, + 182, + 367, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 182, + 367, + 193 + ], + "spans": [ + { + "bbox": [ + 231, + 182, + 367, + 193 + ], + "type": "text", + "content": "liangkongming@bupt.edu.cn" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 425, + 131, + 496, + 146 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 131, + 496, + 146 + ], + "spans": [ + { + "bbox": [ + 425, + 131, + 496, + 146 + ], + "type": "text", + "content": "Huayu Zhang" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 393, + 148, + 529, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 148, + 529, + 159 + ], + "spans": [ + { + "bbox": [ + 393, + 148, + 529, + 159 + ], + "type": "text", + "content": "zhanghy56@chinatelecom.cn" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 419, + 165, + 481, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 165, + 481, + 179 + ], + "spans": [ + { + "bbox": [ + 419, + 165, + 481, + 179 + ], + "type": "text", + "content": "Zhanyu Ma" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 395, + 182, + 505, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 182, + 505, + 193 + ], + "spans": [ + { + "bbox": [ + 395, + 182, + 505, + 193 + ], + "type": "text", + "content": "mazhanyu@bupt.edu.cn" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 61, + 213, + 572, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 213, + 572, + 228 + ], + "spans": [ + { + "bbox": [ + 61, + 213, + 572, + 228 + ], + "type": "text", + "content": "1 Beijing University of Posts and Telecommunications 2 Institute of Artificial Intelligence, China Telecom" + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 59, + 252, + 345, + 474 + ], + "blocks": [ + { + "bbox": [ + 59, + 252, + 345, + 474 + ], + "lines": [ + { + "bbox": [ + 59, + 252, + 345, + 474 + ], + "spans": [ + { + "bbox": [ + 59, + 252, + 345, + 474 + ], + "type": "image", + "image_path": "6976644bf0064501270d002d22d9fbb7ac54aa0848538f72e93f2a56ead76646.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 482, + 555, + 515 + ], + "lines": [ + { + "bbox": [ + 54, + 482, + 555, + 515 + ], + "spans": [ + { + "bbox": [ + 54, + 482, + 555, + 515 + ], + "type": "text", + "content": "Figure 1. Compare with existing methods under two scenarios: 1) General T2I Generation (in the left part): from left to right are SDXL-base[31], MoLE[49] and Ours. 2) Controllable Generation (in the right part): from left to right are input condition, ControlNet-SDXL[47] and Ours. Zoom in for a better view of details." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 347, + 255, + 554, + 474 + ], + "blocks": [ + { + "bbox": [ + 347, + 255, + 554, + 474 + ], + "lines": [ + { + "bbox": [ + 347, + 255, + 554, + 474 + ], + "spans": [ + { + "bbox": [ + 347, + 255, + 554, + 474 + ], + "type": "image", + "image_path": "d46b8957a8f09948832c8fadda36083bd181dc07be6976380e379060a3a97b6f.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 151, + 525, + 200, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 525, + 200, + 538 + ], + "spans": [ + { + "bbox": [ + 151, + 525, + 200, + 538 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 54, + 551, + 297, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 551, + 297, + 696 + ], + "spans": [ + { + "bbox": [ + 54, + 551, + 297, + 696 + ], + "type": "text", + "content": "Image generation has achieved remarkable progress with the development of large-scale text-to-image models, especially diffusion-based models. However, generating human images with plausible details, such as faces or hands, remains challenging due to insufficient supervision of local regions during training. To address this issue, we propose FairHuman, a multi-objective fine-tuning approach designed to enhance both global and local generation quality fairly. Specifically, we first construct three learning objectives: a global objective derived from the default diffusion objective function and two local objectives for hands and faces based on pre-annotated positional priors. Subse" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 527, + 556, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 527, + 556, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 527, + 556, + 647 + ], + "type": "text", + "content": "sequently, we derive the optimal parameter updating strategy under the guidance of the Minimum Potential Delay (MPD) criterion, thereby attaining fairness-aware optimization for this multi-objective problem. Based on this, our proposed method can achieve significant improvements in generating challenging local details while maintaining overall quality. Extensive experiments showcase the effectiveness of our method in improving the performance of human image generation under different scenarios. Our code is available at https://github.com/PRIS-CV/FairHuman." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 314, + 669, + 395, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 669, + 395, + 681 + ], + "spans": [ + { + "bbox": [ + 314, + 669, + 395, + 681 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "Diffusion-based models have revolutionized image generation by achieving unprecedented high quality and realism" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 15 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 15 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 15 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "spans": [ + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "spans": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 70, + 703, + 151, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 703, + 151, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 703, + 151, + 713 + ], + "type": "text", + "content": "† Corresponding Author" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17046" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 294, + 190 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 294, + 190 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 294, + 190 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 294, + 190 + ], + "type": "image", + "image_path": "ad596aa3bc8fe4b1dc448f34482efd31afb58e227a3c1043b387e1af26554b5e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 199, + 295, + 254 + ], + "lines": [ + { + "bbox": [ + 55, + 199, + 295, + 254 + ], + "spans": [ + { + "bbox": [ + 55, + 199, + 295, + 254 + ], + "type": "text", + "content": "Figure 2. Examples of human-related images generated by SDXL. The first row shows the close-up view images with prompts like \"a close-up human face\" and \"a detailed hand gesture\". The second row shows the human images under different scenes. Zoom in for a better view of details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 279, + 295, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 279, + 295, + 351 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 295, + 351 + ], + "type": "text", + "content": "through iterative denoising processes. However, their capability of human image generation remains challenged by persistent artifacts in anatomically complex regions, particularly faces and hands. These contents often suffer from generating degradation such as distorted facial expressions and implausible hand poses." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 353, + 295, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 353, + 295, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 353, + 295, + 472 + ], + "type": "text", + "content": "To address this issue, previous work, including ControlNet[47], HumanSD[14], and HyperHuman[22], leverages anatomical information like pose or depth to guide human image generation. Moreover, a leading approach [49] fine-tunes two low-rank modules w.r.t hand and face generation, respectively. Then it uses the soft Mixture of Experts (MoE) mechanism[32] to fuse these modules. Essentially, these methods aim to enhance human image generation by injecting specific knowledge about hands and faces into the generative model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "type": "text", + "content": "Despite the success of these methods, the balance between coarse appearance (global quality) and local details (quality of faces and hands) has not been addressed. Through the visual feedback of the human-related images generated by SDXL-base[31] in Fig.2, we have noticed an interesting phenomenon: When the model generates faces or hands separately as the close-up view images, it usually ensures sufficient detail and high quality in terms of structure and texture. However, when generating whole-body images especially when hand or face regions occupy a relatively small portion, noticeable distortions often occur. It suggests that the model inherently possesses the capability to generate plausible faces and hands. Nonetheless, the existence of other contents such as the background, reduces the focus on local regions, thus leading to a decline in generation quality. In addition, a post-processing method called ADetailer[3] repairs the faces and hands by cropping out the distorted regions from the original image and repainting them using super-resolution techniques. Its effectiveness further supports our hypothesis. In this way, the level of" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "content": "attention to global and local aspects is indeed a significant factor influencing the generation quality. Relevant details will be elaborated in Sec.3.1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 110, + 555, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 110, + 555, + 253 + ], + "spans": [ + { + "bbox": [ + 313, + 110, + 555, + 253 + ], + "type": "text", + "content": "Motivated by this observation, we decide to construct specific objective functions for global and local generation, respectively. We notice that the default objective function in Eq.(1) for the diffusion model is based on the global mean squared error (MSE), which pays more attention to the overall performance. For complex tasks like human image generation, a single-objective mechanism may lead to the model excelling in certain aspects while underperforming in others. Therefore, here we transform the human image generation task into a multi-objective optimization problem and divide it into three different sub-objectives. Specifically, our optimization objectives include:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 255, + 553, + 326 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 255, + 553, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 255, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 314, + 255, + 553, + 277 + ], + "type": "text", + "content": "- Facial generation quality: the naturalness of facial structure and expressions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 278, + 553, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 278, + 553, + 302 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 553, + 302 + ], + "type": "text", + "content": "- Hand generation quality: the plausibility of hand poses and shapes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 303, + 553, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 303, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 553, + 326 + ], + "type": "text", + "content": "- Global generation quality: the coherence and perceptual quality of the overall image." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 327, + 553, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 553, + 387 + ], + "type": "text", + "content": "As for the construction of objective functions, we retain the default MSE loss function for global features. Then, we create masks using pre-annotated positional information from the dataset and construct mask-based loss functions for local features, namely faces and hands." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 388, + 554, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 554, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 554, + 604 + ], + "type": "text", + "content": "Considering the optimization method, we need to identify an appropriate strategy to maximize the improvement of each objective. Since faces and hands are more difficult to learn compared to global features during training, directly deploying a fixed or equal weight allocation, i.e., the degree of importance, is unreasonable. To achieve the Pareto optimality of human image generation performance, inspired by the idea of fairness criteria utilized in communication networks[26], we incorporate the minimum potential delay (MPD) fairness principle[25] into our optimization process for parameter updates. Specifically, it dynamically allocates the weight for each objective-specific gradient following the fairness constraints. In this way, the three objective functions mentioned above can be optimized simultaneously within a single diffusion model while mitigating potential conflicts between them. Technically, it can be deployed on fine-tuning different modules like LoRA[9] and ControlNet[47] for certain generation tasks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": "To this end, we propose a framework called FairHuman to generate human images with higher fidelity and quality under different scenarios. The key insight is that human image generation is a complex task that involves both global and local optimization. The previous single-objective methodology cannot improve the performance of different aspects at the same time. Therefore, we regard the learning for human image generation as a multi-objective optimization process and implement the MPD fairness prin" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17047" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 83 + ], + "type": "text", + "content": "ciple to achieve a more effective and balanced performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 84, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 84, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 54, + 84, + 294, + 191 + ], + "type": "text", + "content": "Our method contains three procedures (see Fig.3): Firstly, we construct separate objective functions for global and regional generation by utilizing the default MSE loss function and regional masks obtained from the preannotated positional information. Secondly, we implement the minimum potential delay fairness principle for the joint optimization of all objectives. Eventually, we fine-tune LoRA[9] and ControlNet[47] based on the mentioned strategy for two scenarios of human image generation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 192, + 295, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 192, + 295, + 214 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 295, + 214 + ], + "type": "text", + "content": "In summary, the contributions of our paper are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 217, + 296, + 430 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 217, + 295, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 295, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 295, + 275 + ], + "type": "text", + "content": "- We emphasize that the poor performance in generating local details of human images stems from insufficient attention during the training process. Based on this factor, we construct different objective functions to focus on global and local generation separately." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 276, + 296, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 276, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 276, + 296, + 357 + ], + "type": "text", + "content": "- We propose FairHuman, a general multi-objective optimization framework for diffusion model in improving the performance of human image generation. Here, we optimize different objective functions simultaneously within a single model based on the minimum potential delay fairness principle. To our knowledge, it has not been applied in the diffusion model before." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 359, + 296, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 359, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 359, + 296, + 430 + ], + "type": "text", + "content": "- We implement a widely adopted diffusion model SDXL[31] as the backbone and deploy our method on different modules, including LoRA[9] and ControlNet[47]. Extensive experimental results demonstrate that our method yields a superior performance in enhancing the generation of human images under diverse scenarios." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 442, + 142, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 442, + 142, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 442, + 142, + 453 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "text", + "content": "Human Image Generation. Human image generation refers to generating high-quality human images complying with the given conditions (e.g., text, image). Among image synthesis tasks, it has garnered significant attention due to its considerable academic research values and application potentials[12]. Driven by advancements in scalability and training techniques, existing text-to-image models have demonstrated the ability to generate images with increasing resolution and fidelity. Following the success of GANs[37, 45, 46], diffusion models[8, 39] has emerged as a powerful alternative and become the predominant architecture currently. Models like Stable Diffusion[36] and DALL-E 2[34] leverage latent diffusion processes to generate highly detailed and semantically accurate images from textual descriptions. However, when creating human-centric images, the latest models[2, 5, 31] still struggle to generate fine-grained human parts, especially faces and hands. Previous approaches such as[14, 22, 23, 40] incorporate supplementary conditions like human pose to achieve better generation in a controllable pattern but have major limitations in consistency. A concurrent work[49] deals" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "with the issue of faces and hands by utilizing two low-rank modules[9] with the Mixture of Experts(MoE)[11] mechanism. However, it requires fine-tuning each expert module separately, which brings additional training costs. Meanwhile, stacking of multiple modules also affects the generation efficiency. In our work, we implement joint optimization for both global and local generation performance following the minimum potential delay fairness principle." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 168, + 553, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 553, + 431 + ], + "type": "text", + "content": "Datasets for Human Image Generation. The availability of high-quality datasets has significantly propelled the development of human image generation models. These datasets[4, 6, 10, 13, 48] provide the necessary diversity and richness required for training models in generating realistic and varied human images. However, most of them have two main drawbacks: 1) low resolution and poor image quality, which are inadequate to provide valid prior knowledge for diffusion models. 2) lack of annotations on morphological and spatial information of the human body, which cannot effectively guide the model in generating delicate parts like human faces and hands. To alleviate the above limitations, we collect images from available high-quality human-centric datasets[19, 27] that contain high-resolution real-world images. To further guarantee the image quality, we apply detectors[30, 35, 43] to filter out undesirable images while maintaining as much valuable information as possible. In addition, we annotate the body information for controllable generation and positional information for constructing mask-based objective functions. Implementing the procedures mentioned above, we can achieve enhanced training performance aligned with our objectives." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 443, + 397, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 443, + 397, + 456 + ], + "spans": [ + { + "bbox": [ + 314, + 443, + 397, + 456 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 464, + 427, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 427, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 427, + 475 + ], + "type": "text", + "content": "3.1. Problem Statement" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 481, + 553, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 481, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 481, + 553, + 601 + ], + "type": "text", + "content": "The objective function of diffusion models is founded on the mean square error associated with the denoising task. It enhances the model's performance by minimizing the discrepancy between the predicted noise and the actual noise. In the case of Latent Diffusion Models (LDM)[36], this denoising process is executed within the latent space, facilitated by a VAE encoder[16]. This approach allows for more efficient and effective denoising by operating in a compressed representation of the data. Specifically, its optimization objective can be formulated as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 361, + 611, + 553, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 611, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 361, + 611, + 553, + 632 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} E _ {z _ {0}, t, c, \\epsilon} \\left[ \\| \\epsilon - \\hat {\\epsilon} _ {\\theta} \\left(z _ {t}, t, c\\right) \\| ^ {2} \\right], \\tag {1}", + "image_path": "485b049e0212721b9cdbf1db5796f84424809bd126c762e096c3140913c4c6e0.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": " is the noisy latent at timestep " + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": " obtained from " + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "inline_equation", + "content": "z_{0}" + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": " through the diffusion process by adding Gaussian noise " + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim N(0,1)" + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": "; the initial representation " + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "inline_equation", + "content": "z_{0}" + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": " is defined as " + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "inline_equation", + "content": "z_{0} = \\varepsilon(x_{0})" + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": " denoting the encoder and " + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "inline_equation", + "content": "x_{0}" + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": " being the ground truth image; " + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_{\\theta}(z_{t},t,c)" + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": " is the noise predicted by the diffusion model; " + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": " represents the embedding of input" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17048" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 553, + 300 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 553, + 300 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 553, + 300 + ], + "type": "image", + "image_path": "0c11849a7c4bab38bb60d8e770aacdea8c9b60c8c9cc3e9d8b97d8babe418075.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 308, + 555, + 343 + ], + "lines": [ + { + "bbox": [ + 54, + 308, + 555, + 343 + ], + "spans": [ + { + "bbox": [ + 54, + 308, + 555, + 343 + ], + "type": "text", + "content": "Figure 3. An overview of FairHuman framework. The entire pipeline is divided into three stages: 1) Construct multi objectives for human image generation; 2) Employ fairness-aware optimization strategy for parameter updates; 3) Apply the fine-tuned module into the backbone diffusion model. Here, we take LoRA[9] and ControlNet[47] as examples of different generation scenarios." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 363, + 295, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 363, + 295, + 389 + ], + "spans": [ + { + "bbox": [ + 54, + 363, + 295, + 389 + ], + "type": "text", + "content": "conditions (e.g., textual prompt or image); " + }, + { + "bbox": [ + 54, + 363, + 295, + 389 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 54, + 363, + 295, + 389 + ], + "type": "text", + "content": " represents the parameters to be updated." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 395, + 295, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 395, + 295, + 539 + ], + "spans": [ + { + "bbox": [ + 54, + 395, + 295, + 539 + ], + "type": "text", + "content": "By pre-training on large-scale datasets with this objective function, models such as SDXL[31] have shown the ability to generate high-quality human content. This is evident in the first row of Fig.2, where close-up images are generated accurately. However, when generating full-body human images within a scene (see the second row of Fig.2), the model demonstrates inferior performance in rendering local details (faces and hands). In other words, when face or hand regions occupy a larger proportion of the image, their generation results are acceptable. On the contrary, when they occupy a smaller proportion of the image, their corresponding generation quality would be worse." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 546, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 714 + ], + "type": "text", + "content": "What accounts for the occurrence of this issue? As indicated by Eq.(1), LDM calculates the loss based on the representation " + }, + { + "bbox": [ + 55, + 546, + 295, + 714 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 55, + 546, + 295, + 714 + ], + "type": "text", + "content": " encoded from the entire image without distinguishing the optimization difficulty of different regions. For faces and hands with larger areas in close-up images, the default objective function can be easily optimized since these regions contribute more significantly to the loss. Conversely, for human-in-the-scene images, the representation " + }, + { + "bbox": [ + 55, + 546, + 295, + 714 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 55, + 546, + 295, + 714 + ], + "type": "text", + "content": " comprises not only the hands and faces but also other body details and the background scene. Therefore, face and hand regions with smaller areas contribute less to the loss, thus leading to insufficient attention and resulting in poorer generation quality. In this way, different regions may compete during the generation process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 361, + 413, + 375 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 413, + 375 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 413, + 375 + ], + "type": "text", + "content": "3.2. Multi-objectives" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 380, + 555, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 555, + 476 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 555, + 476 + ], + "type": "text", + "content": "As we discussed above, the competition between different regions may lead to insufficient attention to some local areas, such as hands or faces. To address this issue, the generative model requires a fair allocation of resources to both local and global regions, thus enhancing the overall generation performance. To this end, we first construct the following optimization objectives for both local and global generation tasks, respectively:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 326, + 484, + 497, + 505 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 484, + 497, + 505 + ], + "spans": [ + { + "bbox": [ + 326, + 484, + 497, + 505 + ], + "type": "interline_equation", + "content": "l _ {\\text {g l o b a l}} (\\theta) = \\mathbb {E} _ {z _ {0}, t, c, \\epsilon} \\left[ \\| \\epsilon - \\hat {\\epsilon} _ {\\theta} (z _ {t}, t, c) \\| ^ {2} \\right];", + "image_path": "4d28c7f31d2ad20d0946ae269a73a1a78af90c1708adfd55b8203b89039330ef.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 328, + 506, + 555, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 506, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 328, + 506, + 555, + 525 + ], + "type": "interline_equation", + "content": "l _ {\\text {f a c e}} (\\theta) = \\mathbb {E} _ {z _ {0}, t, c, \\epsilon} \\left[ \\| m _ {f} \\odot \\left(\\epsilon - \\hat {\\epsilon} _ {\\theta} (z _ {t}, t, c)\\right) \\| ^ {2} \\right]; \\tag {2}", + "image_path": "0eb3c62321b70abcd1255c4de9d2421daeb23a7cfb0b911c5544b0bfcf6cf32b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 329, + 528, + 530, + 549 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 528, + 530, + 549 + ], + "spans": [ + { + "bbox": [ + 329, + 528, + 530, + 549 + ], + "type": "interline_equation", + "content": "l _ {\\mathrm {h a n d}} (\\theta) = \\mathbb {E} _ {z _ {0}, t, c, \\epsilon} \\left[ \\| m _ {h} \\odot (\\epsilon - \\hat {\\epsilon} _ {\\theta} (z _ {t}, t, c)) \\| ^ {2} \\right],", + "image_path": "e88b91b7442b529b634c6f4f802e69897ddaf09c6091fc9ebcc72daa95adf906.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 558, + 555, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 558, + 555, + 666 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 558, + 555, + 666 + ], + "type": "text", + "content": " represents the learnable parameters; " + }, + { + "bbox": [ + 313, + 558, + 555, + 666 + ], + "type": "inline_equation", + "content": "l_{global}(\\theta)" + }, + { + "bbox": [ + 313, + 558, + 555, + 666 + ], + "type": "text", + "content": " is the global loss for entire image generation; " + }, + { + "bbox": [ + 313, + 558, + 555, + 666 + ], + "type": "inline_equation", + "content": "l_{face}(\\theta)" + }, + { + "bbox": [ + 313, + 558, + 555, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 558, + 555, + 666 + ], + "type": "inline_equation", + "content": "l_{hand}(\\theta)" + }, + { + "bbox": [ + 313, + 558, + 555, + 666 + ], + "type": "text", + "content": " are local losses focusing on faces and hands; " + }, + { + "bbox": [ + 313, + 558, + 555, + 666 + ], + "type": "inline_equation", + "content": "m_f" + }, + { + "bbox": [ + 313, + 558, + 555, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 558, + 555, + 666 + ], + "type": "inline_equation", + "content": "m_h" + }, + { + "bbox": [ + 313, + 558, + 555, + 666 + ], + "type": "text", + "content": " are the masks used to select the face and hand regions respectively. Here, the masks " + }, + { + "bbox": [ + 313, + 558, + 555, + 666 + ], + "type": "inline_equation", + "content": "m_f" + }, + { + "bbox": [ + 313, + 558, + 555, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 558, + 555, + 666 + ], + "type": "inline_equation", + "content": "m_h" + }, + { + "bbox": [ + 313, + 558, + 555, + 666 + ], + "type": "text", + "content": " are obtained by reshaping the mask image to the latent shape, following inpainting mask " + }, + { + "bbox": [ + 313, + 558, + 555, + 666 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 558, + 555, + 666 + ], + "type": "text", + "content": " denoted in Eq.(4) of [23]. For brevity, we also denote " + }, + { + "bbox": [ + 313, + 558, + 555, + 666 + ], + "type": "inline_equation", + "content": "l_{global}(\\theta), l_{face}(\\theta), l_{hand}(\\theta)" + }, + { + "bbox": [ + 313, + 558, + 555, + 666 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 558, + 555, + 666 + ], + "type": "inline_equation", + "content": "l_1(\\theta), l_2(\\theta), l_3(\\theta)" + }, + { + "bbox": [ + 313, + 558, + 555, + 666 + ], + "type": "text", + "content": " in the subsequent sections." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "content": "Based on this, we aim to fine-tune a diffusion model to simultaneously minimize " + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "inline_equation", + "content": "l_{1}(\\theta)" + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "inline_equation", + "content": "l_{2}(\\theta)" + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "inline_equation", + "content": "l_{3}(\\theta)" + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "content": ". Meanwhile, the fine-tuning approach should enhance the generation quality of hands and faces while maintaining its global" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17049" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 106 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 106 + ], + "type": "text", + "content": "performance. In this way, the fine-tuning task could be formulated as the following multi-objective optimization problem:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 106, + 295, + 125 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 106, + 295, + 125 + ], + "spans": [ + { + "bbox": [ + 105, + 106, + 295, + 125 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\mathcal {L} (\\theta) = \\left[ l _ {1} (\\theta), l _ {2} (\\theta), l _ {3} (\\theta) \\right] ^ {\\top}. \\tag {3}", + "image_path": "1731b26cb3cee92d4786296a77622410581f9cde3bb366dc44d140cc8a1d7cd7.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 128, + 279, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 128, + 279, + 139 + ], + "spans": [ + { + "bbox": [ + 55, + 128, + 279, + 139 + ], + "type": "text", + "content": "Notably, our primary target comprises two key aspects:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 140, + 296, + 320 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 140, + 295, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 140, + 295, + 235 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 295, + 235 + ], + "type": "text", + "content": "- Regional Fairness: We should ensure that the solution " + }, + { + "bbox": [ + 55, + 140, + 295, + 235 + ], + "type": "inline_equation", + "content": "\\theta^{*}" + }, + { + "bbox": [ + 55, + 140, + 295, + 235 + ], + "type": "text", + "content": " to Eq.(3) yields a fairly substantial improvement across all three objectives, especially for " + }, + { + "bbox": [ + 55, + 140, + 295, + 235 + ], + "type": "inline_equation", + "content": "l_{2}(\\theta)" + }, + { + "bbox": [ + 55, + 140, + 295, + 235 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 140, + 295, + 235 + ], + "type": "inline_equation", + "content": "l_{3}(\\theta)" + }, + { + "bbox": [ + 55, + 140, + 295, + 235 + ], + "type": "text", + "content": ". As long as there is at least one insignificant improvement among the three objectives, it implies that our fine-tuning is ineffective. This aligns with our purpose of fair enhancement on the quality w.r.t face, hand, and global regions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 235, + 296, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 235, + 296, + 320 + ], + "spans": [ + { + "bbox": [ + 55, + 235, + 296, + 320 + ], + "type": "text", + "content": "- Pareto Optimality: The solution " + }, + { + "bbox": [ + 55, + 235, + 296, + 320 + ], + "type": "inline_equation", + "content": "\\theta^{*}" + }, + { + "bbox": [ + 55, + 235, + 296, + 320 + ], + "type": "text", + "content": " to Eq.(3) should be Pareto optimal[29, 38]. Here we say a solution " + }, + { + "bbox": [ + 55, + 235, + 296, + 320 + ], + "type": "inline_equation", + "content": "\\theta^{*}" + }, + { + "bbox": [ + 55, + 235, + 296, + 320 + ], + "type": "text", + "content": " is Pareto optimal if there exists no other solution " + }, + { + "bbox": [ + 55, + 235, + 296, + 320 + ], + "type": "inline_equation", + "content": "\\theta^{\\prime}" + }, + { + "bbox": [ + 55, + 235, + 296, + 320 + ], + "type": "text", + "content": " satisfying " + }, + { + "bbox": [ + 55, + 235, + 296, + 320 + ], + "type": "inline_equation", + "content": "l_{i}(\\theta^{\\prime})\\leq l_{i}(\\theta^{*})" + }, + { + "bbox": [ + 55, + 235, + 296, + 320 + ], + "type": "text", + "content": " for all " + }, + { + "bbox": [ + 55, + 235, + 296, + 320 + ], + "type": "inline_equation", + "content": "i\\in \\{1,2,3\\}" + }, + { + "bbox": [ + 55, + 235, + 296, + 320 + ], + "type": "text", + "content": ". In other words, updating " + }, + { + "bbox": [ + 55, + 235, + 296, + 320 + ], + "type": "inline_equation", + "content": "\\theta^{*}" + }, + { + "bbox": [ + 55, + 235, + 296, + 320 + ], + "type": "text", + "content": " in any direction will disrupt at least one of the three objectives. Thus Pareto optimality here implies a trade-off between face, hand, and global quality." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 325, + 214, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 325, + 214, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 325, + 214, + 338 + ], + "type": "text", + "content": "3.3. Fairness-aware Optimization" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 342, + 296, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 296, + 437 + ], + "type": "text", + "content": "We aim to ensure the post fine-tuning values of " + }, + { + "bbox": [ + 55, + 342, + 296, + 437 + ], + "type": "inline_equation", + "content": "l_{1}(\\theta), l_{2}(\\theta), l_{3}(\\theta)" + }, + { + "bbox": [ + 55, + 342, + 296, + 437 + ], + "type": "text", + "content": " are all fairly minimized to a sufficient degree. To this end, a straightforward approach is to enable each " + }, + { + "bbox": [ + 55, + 342, + 296, + 437 + ], + "type": "inline_equation", + "content": "l_{i}(\\theta)" + }, + { + "bbox": [ + 55, + 342, + 296, + 437 + ], + "type": "text", + "content": " could be maximally decreased at each training step of " + }, + { + "bbox": [ + 55, + 342, + 296, + 437 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 342, + 296, + 437 + ], + "type": "text", + "content": ". Formally, we seek a stepwise updating direction " + }, + { + "bbox": [ + 55, + 342, + 296, + 437 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 342, + 296, + 437 + ], + "type": "text", + "content": " that maximizes the contributions of " + }, + { + "bbox": [ + 55, + 342, + 296, + 437 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 342, + 296, + 437 + ], + "type": "text", + "content": " to all of " + }, + { + "bbox": [ + 55, + 342, + 296, + 437 + ], + "type": "inline_equation", + "content": "l_{1}(\\theta), l_{2}(\\theta)" + }, + { + "bbox": [ + 55, + 342, + 296, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 342, + 296, + 437 + ], + "type": "inline_equation", + "content": "l_{3}(\\theta)" + }, + { + "bbox": [ + 55, + 342, + 296, + 437 + ], + "type": "text", + "content": ". From another perspective, this is equivalent to minimizing the negative effect of " + }, + { + "bbox": [ + 55, + 342, + 296, + 437 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 342, + 296, + 437 + ], + "type": "text", + "content": " to all these objectives." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 437, + 296, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 296, + 677 + ], + "type": "text", + "content": "Here, we employ the concept of Potential Delay [25, 26] to quantitatively describe the negative effect. The term \"Potential Delay\" originally refers to the expected time required for a user to complete a specific task in the context of resource allocation. For example, if an internet user transmits a unit-sized file at transmission rate " + }, + { + "bbox": [ + 55, + 437, + 296, + 677 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 437, + 296, + 677 + ], + "type": "text", + "content": ", then the potential delay is " + }, + { + "bbox": [ + 55, + 437, + 296, + 677 + ], + "type": "inline_equation", + "content": "1 / v" + }, + { + "bbox": [ + 55, + 437, + 296, + 677 + ], + "type": "text", + "content": ". In this paper, the common updating direction " + }, + { + "bbox": [ + 55, + 437, + 296, + 677 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 437, + 296, + 677 + ], + "type": "text", + "content": " can be considered a shared resource among all objectives. The \"transmission rate\" of this resource to each " + }, + { + "bbox": [ + 55, + 437, + 296, + 677 + ], + "type": "inline_equation", + "content": "l_{i}(\\theta)" + }, + { + "bbox": [ + 55, + 437, + 296, + 677 + ], + "type": "text", + "content": " is represented by the projection magnitude of " + }, + { + "bbox": [ + 55, + 437, + 296, + 677 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 437, + 296, + 677 + ], + "type": "text", + "content": " onto the steepest descent direction " + }, + { + "bbox": [ + 55, + 437, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\nabla l_{i}(\\theta)" + }, + { + "bbox": [ + 55, + 437, + 296, + 677 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 55, + 437, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\| \\mathrm{proj}_{\\nabla l_i(\\theta)}(d)\\|" + }, + { + "bbox": [ + 55, + 437, + 296, + 677 + ], + "type": "text", + "content": ". Consequently, the corresponding potential delay is " + }, + { + "bbox": [ + 55, + 437, + 296, + 677 + ], + "type": "inline_equation", + "content": "1 / \\| \\mathrm{proj}_{\\nabla l_i(\\theta)}(d)\\|" + }, + { + "bbox": [ + 55, + 437, + 296, + 677 + ], + "type": "text", + "content": ". Normally, projection " + }, + { + "bbox": [ + 55, + 437, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\| \\mathrm{proj}_{\\nabla l_i(\\theta)}(d)\\|" + }, + { + "bbox": [ + 55, + 437, + 296, + 677 + ], + "type": "text", + "content": " can serve as a metric for the contribution of " + }, + { + "bbox": [ + 55, + 437, + 296, + 677 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 437, + 296, + 677 + ], + "type": "text", + "content": " to the optimization process of " + }, + { + "bbox": [ + 55, + 437, + 296, + 677 + ], + "type": "inline_equation", + "content": "l_{i}(\\theta)" + }, + { + "bbox": [ + 55, + 437, + 296, + 677 + ], + "type": "text", + "content": " [44]. In this way, the potential delay " + }, + { + "bbox": [ + 55, + 437, + 296, + 677 + ], + "type": "inline_equation", + "content": "1 / \\| \\mathrm{proj}_{\\nabla l_i(\\theta)}(d)\\|" + }, + { + "bbox": [ + 55, + 437, + 296, + 677 + ], + "type": "text", + "content": " is inversely proportional to the contribution from " + }, + { + "bbox": [ + 55, + 437, + 296, + 677 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 437, + 296, + 677 + ], + "type": "text", + "content": " and thus directly proportional to the negative effect. In this way, we can determine the optimal direction " + }, + { + "bbox": [ + 55, + 437, + 296, + 677 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 437, + 296, + 677 + ], + "type": "text", + "content": " by resolving the subsequent minimization problem:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 103, + 684, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 684, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 103, + 684, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\min _ {d} \\mathcal {F} (d) = \\sum_ {i = 1} ^ {3} \\frac {1}{\\| \\operatorname {p r o j} _ {\\nabla l _ {i} (\\theta)} (d) \\|}, \\tag {4}", + "image_path": "3a5c057ecc7afc4b238773b214db61f869d35d32e1379e424b58461a06b535b2.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "By optimizing Eq.(4), the potential delay of " + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "inline_equation", + "content": "l_{1}(\\theta)" + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "inline_equation", + "content": "l_{2}(\\theta)" + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "inline_equation", + "content": "l_{3}(\\theta)" + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": " can be minimized simultaneously, thus it is termed as Minimal Potential Delay (MPD) Fairness [25, 26]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 108, + 553, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 553, + 157 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 553, + 157 + ], + "type": "text", + "content": "Prior to solving this optimization, we assume that " + }, + { + "bbox": [ + 313, + 108, + 553, + 157 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 108, + 553, + 157 + ], + "type": "text", + "content": " lies within a ball with radius " + }, + { + "bbox": [ + 313, + 108, + 553, + 157 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 108, + 553, + 157 + ], + "type": "text", + "content": " centered at the origin. Based on this, the optimization problem w.r.t MPD fairness can be transformed into the following form:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 372, + 166, + 553, + 198 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 166, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 372, + 166, + 553, + 198 + ], + "type": "interline_equation", + "content": "\\min _ {d} \\mathcal {F} ^ {\\prime} (d) = \\sum_ {i = 1} ^ {3} \\frac {1}{\\nabla l _ {i} (\\theta) ^ {\\top} d}, \\tag {5}", + "image_path": "c007a278189e6969cd78b07cb18afe6c3d9e4f8b6259d7fe38c05091c46e36fb.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 207, + 553, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 207, + 553, + 279 + ], + "spans": [ + { + "bbox": [ + 313, + 207, + 553, + 279 + ], + "type": "text", + "content": "since that " + }, + { + "bbox": [ + 313, + 207, + 553, + 279 + ], + "type": "inline_equation", + "content": "\\mathcal{F}'(d)" + }, + { + "bbox": [ + 313, + 207, + 553, + 279 + ], + "type": "text", + "content": " is proportional to an upper bound of " + }, + { + "bbox": [ + 313, + 207, + 553, + 279 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(d)" + }, + { + "bbox": [ + 313, + 207, + 553, + 279 + ], + "type": "text", + "content": " in Eq.(4), and the proof is provided in the Sec.1.1 of the supplementary material. Here, " + }, + { + "bbox": [ + 313, + 207, + 553, + 279 + ], + "type": "inline_equation", + "content": "\\mathcal{F}'(d)" + }, + { + "bbox": [ + 313, + 207, + 553, + 279 + ], + "type": "text", + "content": " can be regarded as a special case of " + }, + { + "bbox": [ + 313, + 207, + 553, + 279 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 207, + 553, + 279 + ], + "type": "text", + "content": "-fairness[26]. If we set " + }, + { + "bbox": [ + 313, + 207, + 553, + 279 + ], + "type": "inline_equation", + "content": "\\alpha = 2" + }, + { + "bbox": [ + 313, + 207, + 553, + 279 + ], + "type": "text", + "content": " for Eq.(2) in [1], it will be equivalent to Eq.(5) in this paper. Thus, we can adopt a similar procedure to solve Eq.(5) following [1]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 280, + 553, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 280, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 280, + 553, + 304 + ], + "type": "text", + "content": "Specifically, we first apply Lagrange multiplier method on Eq.(5) to obtain the following equation:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 386, + 313, + 553, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 313, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 386, + 313, + 553, + 346 + ], + "type": "interline_equation", + "content": "d = \\sum_ {i = 1} ^ {3} \\frac {\\nabla l _ {i} (\\theta)}{(\\nabla l _ {i} (\\theta) ^ {\\top} d) ^ {2}}, \\tag {6}", + "image_path": "885c9808f45bb1a8974892e6fd2dab417c5499ed4f97eb60395a915ee922b98f.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 354, + 553, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 553, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 553, + 390 + ], + "type": "text", + "content": "The derivation process is described in Sec.1.2 of the supplementary material. To further simplify the problem, we follow [1] and express " + }, + { + "bbox": [ + 313, + 354, + 553, + 390 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 354, + 553, + 390 + ], + "type": "text", + "content": " as a linear combination:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 394, + 399, + 553, + 431 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 399, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 394, + 399, + 553, + 431 + ], + "type": "interline_equation", + "content": "d = \\sum_ {i = 1} ^ {3} w _ {i} \\nabla l _ {i} (\\theta), \\tag {7}", + "image_path": "77aae1d8345fc42bbb378e2948ea13bcf3f7a5e91cfb168310189eabe557d9f0.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 441, + 553, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 553, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 553, + 489 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 441, + 553, + 489 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 313, + 441, + 553, + 489 + ], + "type": "text", + "content": " represents the weight of the " + }, + { + "bbox": [ + 313, + 441, + 553, + 489 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 441, + 553, + 489 + ], + "type": "text", + "content": "-th objective. Given that each gradient " + }, + { + "bbox": [ + 313, + 441, + 553, + 489 + ], + "type": "inline_equation", + "content": "\\nabla l_{i}(\\theta)" + }, + { + "bbox": [ + 313, + 441, + 553, + 489 + ], + "type": "text", + "content": " is readily available, the problem of determining " + }, + { + "bbox": [ + 313, + 441, + 553, + 489 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 441, + 553, + 489 + ], + "type": "text", + "content": " is reformulated as determining every " + }, + { + "bbox": [ + 313, + 441, + 553, + 489 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 313, + 441, + 553, + 489 + ], + "type": "text", + "content": ". By substituting Eq.(7) into Eq.(6), we can get that:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 380, + 497, + 553, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 497, + 553, + 512 + ], + "spans": [ + { + "bbox": [ + 380, + 497, + 553, + 512 + ], + "type": "interline_equation", + "content": "G (\\theta) ^ {\\top} G (\\theta) W = W ^ {- 1 / 2}, \\tag {8}", + "image_path": "722e36517af83cd9db5249f12f3ebc45d64cb6d13f4a54faffc6855bd8c8e6c6.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 520, + 553, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 553, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 520, + 553, + 567 + ], + "type": "inline_equation", + "content": "G(\\theta) = [\\nabla l_1(\\theta), \\nabla l_2(\\theta), \\nabla l_3(\\theta)]^\\top" + }, + { + "bbox": [ + 313, + 520, + 553, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 520, + 553, + 567 + ], + "type": "inline_equation", + "content": "W = [w_1, w_2, w_3]^\\top" + }, + { + "bbox": [ + 313, + 520, + 553, + 567 + ], + "type": "text", + "content": ". According to the relation in Eq.(8), we can further construct the following convex optimization problem:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 362, + 567, + 553, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 567, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 362, + 567, + 553, + 590 + ], + "type": "interline_equation", + "content": "\\min _ {W} \\left\\| G (\\theta) ^ {\\top} G (\\theta) W - W ^ {- 1 / 2} \\right\\| ^ {2}, \\tag {9}", + "image_path": "240eef74631df277aa88025c1ac1655b41e0691f68fce3c23e2fc83bf8e96c97.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 594, + 553, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 553, + 619 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 553, + 619 + ], + "type": "text", + "content": "By applying the standard least square method, we can obtain the following close-form solution of Eq.(9):" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 380, + 628, + 553, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 628, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 380, + 628, + 553, + 646 + ], + "type": "interline_equation", + "content": "W = \\left(G (\\theta) ^ {\\top} G (\\theta)\\right) ^ {- \\frac {2}{3}} \\mathbf {1}, \\tag {10}", + "image_path": "236b3ded5ff0c4f2a455a478760c9101241b0ddac1a9d7f05130e8440bff4b6c.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 654, + 553, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 691 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 691 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 654, + 553, + 691 + ], + "type": "inline_equation", + "content": "\\mathbf{1}" + }, + { + "bbox": [ + 313, + 654, + 553, + 691 + ], + "type": "text", + "content": " represents an all-ones vector. With these weights, we can finally obtain the value of " + }, + { + "bbox": [ + 313, + 654, + 553, + 691 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 654, + 553, + 691 + ], + "type": "text", + "content": " by using Eq.(7) and then conduct an updating step as:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 400, + 701, + 553, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 400, + 701, + 553, + 714 + ], + "type": "interline_equation", + "content": "\\theta_ {\\text {n e w}} = \\theta - \\eta d, \\tag {11}", + "image_path": "aa2a822ae0fb84080efcba652ae326fbac4db5b0c367d0925d45f85e9529a425.jpg" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17050" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 180 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 54, + 72, + 294, + 180 + ], + "type": "inline_equation", + "content": "\\theta_{new}" + }, + { + "bbox": [ + 54, + 72, + 294, + 180 + ], + "type": "text", + "content": " is the updated parameter and " + }, + { + "bbox": [ + 54, + 72, + 294, + 180 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 54, + 72, + 294, + 180 + ], + "type": "text", + "content": " is the learning rate. By adopting this approach, each training step is guided by MPD fairness, thereby ensuring the efficacy of the finetuning results across all objectives. Meanwhile, Thm.(7.3) in [1] also indicates that such a training process can converge to a Pareto optimal solution. And our constructed objective function decouples regional and global optimization, satisfying MPD's prerequisite assumption of linearly independent task gradients." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 187, + 133, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 187, + 133, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 187, + 133, + 201 + ], + "type": "text", + "content": "3.4. Application" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 204, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 204, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 204, + 296, + 407 + ], + "type": "text", + "content": "Considering the high cost of full fine-tuning for diffusion models, we opt to deploy our method in LoRA[9] and ControlNet[47] while freezing the parameters of the backbone model. They represent two types of generation patterns, namely, general T2I generation and controllable generation with control signals. Specifically, LoRA primarily enhances the quality of human images from the perspective of visual concepts and pixel information, while ControlNet focuses on more refined and controllable generation over the structural aspects. Both have significant influence and play important roles in the field of human image generation. We construct a high-quality dataset for their fine-tuning. Apart from the images and text descriptions, it also includes the positional information for creating mask-based objective functions and the pose-depth information for controllable generation. Their new outputs can be formulated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 65, + 417, + 295, + 431 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 417, + 295, + 431 + ], + "spans": [ + { + "bbox": [ + 65, + 417, + 295, + 431 + ], + "type": "interline_equation", + "content": "\\hat {y} _ {\\text {n e w}} = N _ {\\text {b a s e}} (z; \\theta_ {\\text {b a s e}}) + \\beta N _ {\\text {n e w}} (c, z; \\theta_ {\\text {u p d a t e}}), \\tag {12}", + "image_path": "be28830e2d52d7f6bd1000f471793ccb4607d4187063a9d5db6fe101be458099.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 438, + 295, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 295, + 523 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 295, + 523 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 438, + 295, + 523 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 438, + 295, + 523 + ], + "type": "text", + "content": " is the input latent and " + }, + { + "bbox": [ + 55, + 438, + 295, + 523 + ], + "type": "inline_equation", + "content": "N_{base}" + }, + { + "bbox": [ + 55, + 438, + 295, + 523 + ], + "type": "text", + "content": " is the backbone network with the frozen parameters " + }, + { + "bbox": [ + 55, + 438, + 295, + 523 + ], + "type": "inline_equation", + "content": "\\theta_{base}" + }, + { + "bbox": [ + 55, + 438, + 295, + 523 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 438, + 295, + 523 + ], + "type": "inline_equation", + "content": "N_{new}" + }, + { + "bbox": [ + 55, + 438, + 295, + 523 + ], + "type": "text", + "content": " represents the fine-tuned module (LoRA or ControlNet here) with the updated parameters " + }, + { + "bbox": [ + 55, + 438, + 295, + 523 + ], + "type": "inline_equation", + "content": "\\theta_{update}" + }, + { + "bbox": [ + 55, + 438, + 295, + 523 + ], + "type": "text", + "content": ". As for condition " + }, + { + "bbox": [ + 55, + 438, + 295, + 523 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 438, + 295, + 523 + ], + "type": "text", + "content": ", it contains textual and timestep embeddings for the LoRA module and has additional control signals for the ControlNet module. " + }, + { + "bbox": [ + 55, + 438, + 295, + 523 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 438, + 295, + 523 + ], + "type": "text", + "content": " is the hyperparameter used for scaling." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 533, + 137, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 137, + 547 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 137, + 547 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 552, + 182, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 182, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 182, + 566 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 570, + 295, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 701 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 701 + ], + "type": "text", + "content": "Datasets. We collect human-centric images from two public datasets called CosmicMan[19] and BodyHands[27]. To enable effective training, we implement several preprocessing methods, including filtering, captioning, and labeling. Ultimately, we construct a high-quality human image dataset that contains approximately 50k images for training. As for the validation, we randomly sample about 5k in-the-wild human images under different scenes with well-processed captions provided by [49]. The details regarding the construction of our dataset will be elaborated in Sec.2 of the supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "content": "Baseline Methods. We compare our approach with two" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "categories of baseline methods. 1) General T2I methods without extra conditions: SDXL-base[31] and MoLE[49]. Notably, MoLE is a recent work based on the collaboration of multiple LoRA modules. 2) Controllable methods with extra conditions: ControlNet-SDXL[47] and ControlNet Union[18]. Here, we compare performance under the same given conditions (depth and pose) using different training strategies." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 168, + 555, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 555, + 348 + ], + "type": "text", + "content": "Metrics. We adopt evaluation metrics based on the optimization objectives outlined in the paper, focusing on two dimensions. 1) Overall image quality: Human Preference Score (HPS)[41] and ImageReward (IR)[42]. They are built on the CLIP[33] and BLIP[17] models respectively to reflect human visual preferences on the generated images from multiple aspects including image alignment, aesthetic, and plausibility. Moreover, CLIP score is applied to evaluate text-image alignment, and pose alignment is used for assessing the effectiveness of controllable generation. 2) Regional generation quality: FID[7] and detection confidence of mediapipe[24] are utilized to evaluate the quality of hands and faces in terms of realism and plausibility. The implementation details about our experiments are provided in Sec.3 of the supplementary material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 358, + 503, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 503, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 503, + 371 + ], + "type": "text", + "content": "4.2. Comparison with Existing Methods" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 377, + 555, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 555, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 555, + 675 + ], + "type": "text", + "content": "Quantitative Analysis. The evaluation results compared with existing methods are shown in Tab.1. We perform evaluations under two classic scenarios. Specifically, for general T2I methods, including SDXL-Base[31], MoLE[49], and our LoRA-based method, we randomly sample 5k prompts about human-in-the-scene images from the validation dataset and feed them to each model for image generation. As for the setting of hyperparameters, we keep the default CFG scale of 7.5, set the inference step to 30, and fix the seed at 0 for all methods. In particular, for our LoRA-based method, we implement three different weights (0.3, 0.4, and 0.5) in the experiments and use their average metric values as the final result. To more comprehensively evaluate the quality of generated images, we utilize the metrics in terms of overall image quality, text-image alignment, and regional quality. The results demonstrate that our method outperforms existing approaches in all these three aspects, especially in the generation quality of local hand and face details. Since our LoRA-based method and MoLE are deployed on SDXL sharing the same backbone model, we also conduct experiments on their memory usage and inference speed. Results suggest that our method can achieve better performance while requiring less memory size and inference time. Related details are demonstrated in Sec.4.1 of the supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Furthermore, to verify the effectiveness of our method across different task scenarios, we employ it in ControlNet[47] for the controllable generation. As shown" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17051" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 74, + 553, + 163 + ], + "blocks": [ + { + "bbox": [ + 58, + 74, + 553, + 163 + ], + "lines": [ + { + "bbox": [ + 58, + 74, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 58, + 74, + 553, + 163 + ], + "type": "table", + "html": "
Application ScenarioModelImage QualityText-Image AlignmentRegional QualityPose Alignment
HPS(%) ↑IR(%) ↑CLIP ↑FID ↓Hand Confi. ↑Face Confi. ↑Mean Diff ↓
General T2I MethodsSD-XL Base [31]30.91140.8526.8452.6993.1883.67/
MoLE [49]31.87152.3627.1748.9793.3284.48/
Ours32.73154.6027.2046.1194.0385.31/
Controllable MethodsControlNet-SDXL [47]32.20152.4927.1046.1893.5085.133.45
ControlNet-Union [18]32.80155.2227.2146.4394.2885.362.67
ControlNet-Ours32.72154.7427.2445.4694.4885.612.35
", + "image_path": "61b1028b70f62162287e8910ef512e59d3b0cc564a7f92596592d6770b74079f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 171, + 555, + 227 + ], + "lines": [ + { + "bbox": [ + 55, + 171, + 555, + 227 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 555, + 227 + ], + "type": "text", + "content": "Table 1. The evaluation results for ours and other existing methods. We conduct comparisons under two different image generation scenarios separately. As for the general T2I methods, we compare our LoRA-based method with a widely adopted diffusion model SDXL[31] and a recent work MoLE[49]. For controllable methods, we compare our ControlNet-based method with the default ControlNet-SDXL[47] and the latest ControlNet-Union[18]. Notably, the mean difference of pose condition is employed here to evaluate the effectiveness of controllable generation. The best results are indicated in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 247, + 297, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 247, + 297, + 499 + ], + "spans": [ + { + "bbox": [ + 54, + 247, + 297, + 499 + ], + "type": "text", + "content": "in the lower part of Tab.1, we compare with existing controllable methods, including the default ControlNet-SDXL and the latest variant ControlNet-Union[18]. The evaluation is conducted using the same metrics as described above. Similar to our LoRA-based method, we set three different control scales (0.4, 0.5, and 0.6) here for testing and choose their average metric values. Besides, considering the characteristics of controllable generation, we also evaluate the model's performance based on pose alignment, i.e., the consistency between the input conditions and the actual generated content. Here, we use the mean difference between the input and extracted conditions as the metric. The corresponding results are shown in the last column of Tab.1. In comparison, our method still demonstrates superior performance in the controllable generation scenario. Although it is slightly inferior to ControlNet-Union in terms of HPS and IR metrics, the primary reason is that ControlNet-Union incorporates aesthetic and preference feedback learning during training, whereas our method focuses more on the realism and plausibility of the image, and significantly outperforms in regional details." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 56, + 510, + 294, + 597 + ], + "blocks": [ + { + "bbox": [ + 56, + 510, + 294, + 597 + ], + "lines": [ + { + "bbox": [ + 56, + 510, + 294, + 597 + ], + "spans": [ + { + "bbox": [ + 56, + 510, + 294, + 597 + ], + "type": "image", + "image_path": "d0e9f0d34f94e39244f2e7493307ecc663a019bc01f79d6e51601148bb29ae36.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 605, + 295, + 639 + ], + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 639 + ], + "type": "text", + "content": "Figure 4. User study results in terms of overall quality and regional quality. We present the ratio of users' preference for each method." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "Qualitative Analysis. Fig.1 shows qualitative results. Compared with baseline methods, our method can generate human images with higher quality in both global and local aspects. Meanwhile, it achieves better controllability and plausibility in terms of controllable generation. We also" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 247, + 555, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 555, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 555, + 332 + ], + "type": "text", + "content": "conduct a user study to provide a more intuitive evaluation. Volunteers are invited to select images that better align with their visual preferences based on overall and regional quality. The result is shown in Fig.4. It indicates that our method is more preferred by users in both two aspects. More details about the user study will be demonstrated in Sec.4.2 of the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 317, + 343, + 553, + 524 + ], + "blocks": [ + { + "bbox": [ + 317, + 343, + 553, + 524 + ], + "lines": [ + { + "bbox": [ + 317, + 343, + 553, + 524 + ], + "spans": [ + { + "bbox": [ + 317, + 343, + 553, + 524 + ], + "type": "image", + "image_path": "2e31b412e8f2e3796a6fc85c89f8928707202cbc0a41022163bbc6d270f4be89.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 532, + 556, + 574 + ], + "lines": [ + { + "bbox": [ + 313, + 532, + 556, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 556, + 574 + ], + "type": "text", + "content": "Figure 5. Examples of the images generated by different settings with the same prompt. First row: generated by LoRA. Second row: generated by ControlNet. Zoom in for a better view of details." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 599, + 416, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 599, + 416, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 599, + 416, + 612 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": "We present the key ablation studies based on the construction of our objective functions and the strategy for optimization. Here, we analyze their effects according to the overall and local quality of the generated images, deploying on LoRA and ControlNet, respectively. We randomly select 1k human-related prompts and follow the same hyperparameter setting as mentioned above for testing. Their results are shown in Tab.2. In addition, some visualization results are" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17052" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 74, + 553, + 178 + ], + "blocks": [ + { + "bbox": [ + 58, + 74, + 553, + 178 + ], + "lines": [ + { + "bbox": [ + 58, + 74, + 553, + 178 + ], + "spans": [ + { + "bbox": [ + 58, + 74, + 553, + 178 + ], + "type": "table", + "html": "
ModelAblation SettingsImage QualityRegional QualityPose Alignment
HPS(%)↑IR(%)↑FID ↓Hand Confi. ↑Face Confi. ↑Mean Diff ↓
LoRAl1(θ)32.40151.0849.4892.8183.84/
l1(θ) + l2(θ) + l3(θ) (w/o fairness)32.60153.2347.2293.7084.82/
l2(θ) + l2(θ) + l3(θ) (w fairness)32.73154.6046.1194.0385.31/
ControlNetl1(θ)32.20152.4946.1893.5085.133.45
l1(θ) + l2(θ) + l3(θ) (w/o fairness)32.56154.3046.7993.9985.242.78
l1(θ) + l2(θ) + l3(θ) (w fairness)32.72154.7445.3694.4885.612.35
", + "image_path": "01973f37e222c4afca9619c4cb9a47031ffcd4a06d5ae708f32451fdafb3b4b1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 242, + 294, + 289 + ], + "blocks": [ + { + "bbox": [ + 55, + 185, + 555, + 218 + ], + "lines": [ + { + "bbox": [ + 55, + 185, + 555, + 218 + ], + "spans": [ + { + "bbox": [ + 55, + 185, + 555, + 218 + ], + "type": "text", + "content": "Table 2. Ablation results deployed on LoRA and ControlNet. We primarily explore the effect of two aspects: additional objective functions for regional generation and the learning strategy. " + }, + { + "bbox": [ + 55, + 185, + 555, + 218 + ], + "type": "inline_equation", + "content": "l_{1}(\\theta)" + }, + { + "bbox": [ + 55, + 185, + 555, + 218 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 185, + 555, + 218 + ], + "type": "inline_equation", + "content": "l_{2}(\\theta)" + }, + { + "bbox": [ + 55, + 185, + 555, + 218 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 185, + 555, + 218 + ], + "type": "inline_equation", + "content": "l_{3}(\\theta)" + }, + { + "bbox": [ + 55, + 185, + 555, + 218 + ], + "type": "text", + "content": " are defined in Eq.2. The best results are indicated in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 242, + 294, + 289 + ], + "lines": [ + { + "bbox": [ + 57, + 242, + 294, + 289 + ], + "spans": [ + { + "bbox": [ + 57, + 242, + 294, + 289 + ], + "type": "table", + "html": "
LoRA RankImage QualityT2I AlignmentRegional Quality
HPS(%) ↑IR(%) ↑CLIP↑FID ↓Hand Confi. ↑Face Confi. ↑
r=6432.50148.5227.1147.8793.2584.69
r=12832.73154.827.1546.7993.7985.11
r=25632.73154.6027.2046.1194.0385.31
", + "image_path": "164bea3458fcf8ab803b5b3d3ba084ab348b09caeb882b37b9c83a6c67e23f69.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 297, + 295, + 319 + ], + "lines": [ + { + "bbox": [ + 55, + 297, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 295, + 319 + ], + "type": "text", + "content": "Table 3. Comparison results for the choice of LoRA rank. The best results are indicated in bold." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 341, + 130, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 341, + 130, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 341, + 130, + 353 + ], + "type": "text", + "content": "provided in Fig.5." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 354, + 295, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 354, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 54, + 354, + 295, + 508 + ], + "type": "text", + "content": "Effect of Regional Optimization Objective. To explore how regional loss functions including " + }, + { + "bbox": [ + 54, + 354, + 295, + 508 + ], + "type": "inline_equation", + "content": "l_{\\text{face}}(\\theta)" + }, + { + "bbox": [ + 54, + 354, + 295, + 508 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 354, + 295, + 508 + ], + "type": "inline_equation", + "content": "l_{\\text{hand}}(\\theta)" + }, + { + "bbox": [ + 54, + 354, + 295, + 508 + ], + "type": "text", + "content": " in Eq.(2) influence generation performance, we fine-tuned the corresponding LoRA and ControlNet by adopting the default loss function from Eq.(1) and the joint objective functions in Eq.(2), respectively. For the latter, we employ a fixed-weight linear scalarization method here, transforming the multi-objective optimization into a single-objective optimization. The first two rows for each model in Tab.2 present the evaluation results, validating that the inclusion of regional loss functions helps the model improve its generation capabilities for details such as hands and faces, thereby enhancing the overall quality of the images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 509, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 509, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 54, + 509, + 295, + 664 + ], + "type": "text", + "content": "Effect of Fairness Principle. In our work, we utilize the minimum potential delay (MPD) fairness to achieve better optimization based on the joint of objective functions for global and regional generation. Here, we compare two gradient update strategies: one without considering fairness (i.e., a fixed weight allocation based on linear scalarization method) and one considering fairness (i.e., a dynamical weight allocation based on MPD fairness). As shown in the last two rows for each model in Tab.2, the optimization strategy based on the MPD fairness can better focus on the more challenging objectives of generating faces and hands without compromising the overall quality. It achieves a more balanced performance among different objectives." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 672, + 146, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 146, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 146, + 685 + ], + "type": "text", + "content": "4.4. More Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "Considering that the rank setting significantly impacts the LoRA performance, we additionally conduct a hyperparam-" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 240, + 555, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 240, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 555, + 397 + ], + "type": "text", + "content": "eter study to figure out the optimal rank configuration. The results are presented in Tab.3, which indicates that setting the rank value to 256 maximizes the model's ability to generate regional details without excessively jeopardizing the overall image quality. Besides, we also conduct comparison studies on some other multi-objective learning strategies, e.g., Dynamic Weight Average (DwA)[21], Uncertainty Weighting (UW)[15], and Random Loss Weighting (RLW)[20]. The results suggest that applying the MPD fairness principle to the optimization process can achieve a more balanced and superior performance in terms of the human image generation task. Relevant results and details are presented in Sec.4.3 of the supplementary material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 411, + 403, + 424 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 403, + 424 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 403, + 424 + ], + "type": "text", + "content": "4.5. Visualizations" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 432, + 555, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 555, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 555, + 492 + ], + "type": "text", + "content": "We put more visualization results compared with existing methods in Sec.5 of the supplementary material, including other close-source methods like HyperHuman[22] and HanDiffuser[28]. Our method still demonstrates comparable or superior performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 511, + 388, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 511, + 388, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 511, + 388, + 523 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 534, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 715 + ], + "type": "text", + "content": "In this work, we propose FairHuman to enhance the quality of human images generated by diffusion models. To address the distortion issue in local details like faces and hands, we specifically construct corresponding objective functions for regional generation. Then, we combine the default global objective function to achieve joint training. The key of our method is regarding the improvement of human image generation as a multi-objective optimization task. By utilizing the minimum potential delay principle and the gradient manipulation process, our method enables more balanced and effective performance. Extensive experiments demonstrate the superiority of our method under different scenarios of human image generation. For future work, we plan to further design and optimize corresponding objective functions for more attributes related to humans, such as feet and eyes." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17053" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 125 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 125 + ], + "type": "text", + "content": "This work was supported in part by National Natural Science Foundation of China(NSFC) No.62225601, 62476029, U23B2052, and sponsored by Beijing Nova Program." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 148, + 115, + 161 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 148, + 115, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 148, + 115, + 161 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 168, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 168, + 294, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 168, + 294, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 168, + 294, + 190 + ], + "type": "text", + "content": "[1] Hao Ban and Kaiyi Ji. Fair resource allocation in multi-task learning. arXiv preprint arXiv:2402.15638, 2024. 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 192, + 294, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 192, + 294, + 247 + ], + "spans": [ + { + "bbox": [ + 62, + 192, + 294, + 247 + ], + "type": "text", + "content": "[2] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf, 2(3):8, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 248, + 294, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 248, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 62, + 248, + 294, + 270 + ], + "type": "text", + "content": "[3] Bing-su. Adeditor. https://github.com/Bing-su/adeditor.git, 2024.2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 273, + 295, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 273, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 62, + 273, + 295, + 327 + ], + "type": "text", + "content": "[4] Seunghwan Choi, Sunghyun Park, Minsoo Lee, and Jaegul Choo. Viton-hd: High-resolution virtual try-on via misalignment-aware normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14131-14140, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 329, + 294, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 329, + 294, + 393 + ], + "spans": [ + { + "bbox": [ + 62, + 329, + 294, + 393 + ], + "type": "text", + "content": "[5] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 396, + 294, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 396, + 294, + 451 + ], + "spans": [ + { + "bbox": [ + 62, + 396, + 294, + 451 + ], + "type": "text", + "content": "[6] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen Change Loy, Wayne Wu, and Ziwei Liu. Stylegan-human: A data-centric odyssey of human generation. In European Conference on Computer Vision, pages 1-19. Springer, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 453, + 294, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 453, + 294, + 507 + ], + "spans": [ + { + "bbox": [ + 62, + 453, + 294, + 507 + ], + "type": "text", + "content": "[7] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 510, + 294, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 510, + 294, + 542 + ], + "spans": [ + { + "bbox": [ + 62, + 510, + 294, + 542 + ], + "type": "text", + "content": "[8] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 544, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 544, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 62, + 544, + 294, + 588 + ], + "type": "text", + "content": "[9] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 3, 4, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 590, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 294, + 644 + ], + "type": "text", + "content": "[10] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 646, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 646, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 646, + 294, + 678 + ], + "type": "text", + "content": "[11] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. Neural computation, 3(1):79-87, 1991. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "type": "text", + "content": "[12] Zhen Jia, Zhang Zhang, Liang Wang, and Tieniu Tan. Human image generation: A comprehensive survey. ACM Computing Surveys, 56(11):1-39, 2024. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[13] Xuan Ju, Ailing Zeng, Jianan Wang, Qiang Xu, and Lei Zhang. Human-art: A versatile human-centric dataset bridging natural and artificial scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 618-629, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 129, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 183 + ], + "type": "text", + "content": "[14] Xuan Ju, Ailing Zeng, Chenchen Zhao, Jianan Wang, Lei Zhang, and Qiang Xu. Humansd: A native skeleton-guided diffusion model for human image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15988-15998, 2023. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 184, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 184, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 317, + 184, + 553, + 237 + ], + "type": "text", + "content": "[15] Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7482-7491, 2018. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 239, + 553, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 260 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 260 + ], + "type": "text", + "content": "[16] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 262, + 553, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 262, + 553, + 314 + ], + "spans": [ + { + "bbox": [ + 317, + 262, + 553, + 314 + ], + "type": "text", + "content": "[17] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 316, + 553, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 316, + 553, + 382 + ], + "spans": [ + { + "bbox": [ + 317, + 316, + 553, + 382 + ], + "type": "text", + "content": "[18] Ming Li, Taojiannan Yang, Huafeng Kuang, Jie Wu, Zhaoning Wang, Xuefeng Xiao, and Chen Chen. Controlnet++: Improving conditional controls with efficient consistency feedback: Project page: liming-ai. github.io/controlnet_plus_plus. In European Conference on Computer Vision, pages 129–147. Springer, 2024. 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 383, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 383, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 317, + 383, + 553, + 437 + ], + "type": "text", + "content": "[19] Shikai Li, Jianglin Fu, Kaiyuan Liu, Wentao Wang, KwanYee Lin, and Wayne Wu. Cosmicman: A text-to-image foundation model for humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6955-6965, 2024. 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 438, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 553, + 480 + ], + "type": "text", + "content": "[20] Baijiong Lin, Feiyang Ye, Yu Zhang, and Ivor W Tsang. Reasonable effectiveness of random weighting: A litmus test for multi-task learning. arXiv preprint arXiv:2111.10603, 2021.8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 482, + 555, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 482, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 482, + 555, + 525 + ], + "type": "text", + "content": "[21] Shikun Liu, Edward Johns, and Andrew J Davison. End-to-end multi-task learning with attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1871-1880, 2019. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 526, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 553, + 581 + ], + "type": "text", + "content": "[22] Xian Liu, Jian Ren, Aliaksandr Siarohin, Ivan Skorokhodov, Yanyu Li, Dahua Lin, Xihui Liu, Ziwei Liu, and Sergey Tulyakov. Hyperhuman: Hyper-realistic human generation with latent structural diffusion. arXiv preprint arXiv:2310.08579, 2023. 2, 3, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 582, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 582, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 582, + 553, + 635 + ], + "type": "text", + "content": "[23] Wenquan Lu, Yufei Xu, Jing Zhang, Chaoyue Wang, and Dacheng Tao. Handrefiner: Refining malformed hands in generated images by diffusion-based conditional inpainting. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 7085-7093, 2024. 3, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 636, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 553, + 690 + ], + "type": "text", + "content": "[24] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Guang Yong, Juhyun Lee, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, 2019. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[25] Laurent Massoulié and James Roberts. Bandwidth sharing: objectives and algorithms. In IEEE INFOCOM'99. Con-" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17054" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 75, + 72, + 296, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 296, + 116 + ], + "type": "text", + "content": "ference on Computer Communications. Proceedings. Eighteenth Annual Joint Conference of the IEEE Computer and Communications Societies. The Future is Now (Cat. No. 99CH36320), pages 1395-1403. IEEE, 1999. 2, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 149 + ], + "type": "text", + "content": "[26] Jeonghoon Mo and Jean Walrand. Fair end-to-end window-based congestion control. IEEE/ACM Transactions on networking, 8(5):556-567, 2000. 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 151, + 296, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 151, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 151, + 296, + 205 + ], + "type": "text", + "content": "[27] Supreeth Narasimhaswamy, Thanh Nguyen, Mingzhen Huang, and Minh Hoai. Whose hands are these? hand detection and hand-body association in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4889-4899, 2022. 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 295, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 295, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 295, + 270 + ], + "type": "text", + "content": "[28] Supreeth Narasimhaswamy, Uttaran Bhattacharya, Xiang Chen, Ishita Dasgupta, Saayan Mitra, and Minh Hoai. Handiffuser: Text-to-image generation with realistic hand appearances. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2468-2479, 2024. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 272, + 295, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 295, + 316 + ], + "type": "text", + "content": "[29] Patrick Ngatchou, Anahita Zarei, and A El-Sharkawi. Pareto multi objective optimization. In Proceedings of the 13th international conference on, intelligent systems application to power systems, pages 84-91. IEEE, 2005. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 317, + 296, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 317, + 296, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 317, + 296, + 370 + ], + "type": "text", + "content": "[30] Georgios Pavlakos, Dandan Shan, Ilija Radosavovic, Angjoo Kanazawa, David Fouhey, and Jitendra Malik. Reconstructing hands in 3d with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9826-9836, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 372, + 295, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 295, + 425 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 295, + 425 + ], + "type": "text", + "content": "[31] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 426, + 295, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 295, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 295, + 459 + ], + "type": "text", + "content": "[32] Joan Puigcerver, Carlos Riquelme, Basil Mustafa, and Neil Houlsby. From sparse to soft mixtures of experts. arXiv preprint arXiv:2308.00951, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 460, + 295, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 295, + 524 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 295, + 524 + ], + "type": "text", + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 526, + 295, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 295, + 568 + ], + "type": "text", + "content": "[34] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 571, + 295, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 571, + 295, + 603 + ], + "spans": [ + { + "bbox": [ + 56, + 571, + 295, + 603 + ], + "type": "text", + "content": "[35] J Redmon. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2016. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 604, + 295, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 604, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 604, + 295, + 658 + ], + "type": "text", + "content": "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "type": "text", + "content": "[37] Axel Sauer, Tero Karras, Samuli Laine, Andreas Geiger, and Timo Aila. Stylegan-t: Unlocking the power of gans for fast large-scale text-to-image synthesis. In International conference on machine learning, pages 30105-30118. PMLR, 2023. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 665 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[38] Ozan Sener and Vladlen Koltun. Multi-task learning as multi-objective optimization. Advances in neural information processing systems, 31, 2018. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 139 + ], + "type": "text", + "content": "[39] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 141, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 206 + ], + "type": "text", + "content": "[40] Yuxuan Wang, Tianwei Cao, Kongming Liang, Zhongjiang He, Hao Sun, Yongxiang Li, and Zhanyu Ma. Mixture-of-hand-experts: Repainting the deformed hand images generated by diffusion models. In Chinese Conference on Pattern Recognition and Computer Vision (PRCV), pages 143-157. Springer, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 208, + 553, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 261 + ], + "type": "text", + "content": "[41] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 263, + 553, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 263, + 553, + 318 + ], + "spans": [ + { + "bbox": [ + 316, + 263, + 553, + 318 + ], + "type": "text", + "content": "[42] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36:15903-15935, 2023. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 319, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 363 + ], + "type": "text", + "content": "[43] Zhendong Yang, Ailing Zeng, Chun Yuan, and Yu Li. Effective whole-body pose estimation with two-stages distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4210-4220, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 365, + 553, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 407 + ], + "type": "text", + "content": "[44] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. In Advances in Neural Information Processing Systems, pages 5824-5836, 2020. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 409, + 553, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 409, + 553, + 473 + ], + "spans": [ + { + "bbox": [ + 316, + 409, + 553, + 473 + ], + "type": "text", + "content": "[45] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 5907-5915, 2017. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 476, + 553, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 476, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 316, + 476, + 553, + 529 + ], + "type": "text", + "content": "[46] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stack-gan++: Realistic image synthesis with stacked generative adversarial networks. IEEE transactions on pattern analysis and machine intelligence, 41(8):1947-1962, 2018. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 532, + 553, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 532, + 553, + 575 + ], + "spans": [ + { + "bbox": [ + 316, + 532, + 553, + 575 + ], + "type": "text", + "content": "[47] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 577, + 553, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 577, + 553, + 620 + ], + "spans": [ + { + "bbox": [ + 316, + 577, + 553, + 620 + ], + "type": "text", + "content": "[48] Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jingdong Wang, and Qi Tian. Scalable person re-identification: A benchmark. In Proceedings of the IEEE international conference on computer vision, pages 1116–1124, 2015. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 622, + 553, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 622, + 553, + 665 + ], + "spans": [ + { + "bbox": [ + 316, + 622, + 553, + 665 + ], + "type": "text", + "content": "[49] Jie Zhu, Yixiong Chen, Mingyu Ding, Ping Luo, Leye Wang, and Jingdong Wang. Mole: Enhancing human-centric text-to-image diffusion via mixture of low-rank experts. arXiv preprint arXiv:2410.23332, 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17055" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/FakeRadar_ Probing Forgery Outliers to Detect Unknown Deepfake Videos/28a1e3bc-e2b9-40da-b7fe-fcc0d6369340_content_list.json b/2025/FakeRadar_ Probing Forgery Outliers to Detect Unknown Deepfake Videos/28a1e3bc-e2b9-40da-b7fe-fcc0d6369340_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..39cfbf9fd054cf3906c23e7ea1090e217cb0e343 --- /dev/null +++ b/2025/FakeRadar_ Probing Forgery Outliers to Detect Unknown Deepfake Videos/28a1e3bc-e2b9-40da-b7fe-fcc0d6369340_content_list.json @@ -0,0 +1,1510 @@ +[ + { + "type": "text", + "text": "FakeRadar: Probing Forgery Outliers to Detect Unknown Deepfake Videos", + "text_level": 1, + "bbox": [ + 119, + 130, + 879, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhaolun Li $^{1}$ Jichang Li $^{2*}$ Yinqi Cai $^{3}$ Junye Chen $^{3}$ Xiaonan Luo $^{1}$ Guanbin Li $^{3,2,5}$ Rushi Lan $^{1,4*}$", + "bbox": [ + 233, + 179, + 761, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Guilin University of Electronic Technology 2Pengcheng Laboratory 3Sun Yat-sen University 4Guangxi Key Laboratory of Image and Graphic Intelligent Processing", + "bbox": [ + 116, + 215, + 880, + 248 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{5}$ Guangdong Key Laboratory of Big Data Analysis and Processing", + "bbox": [ + 233, + 250, + 759, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zhaolunli@mails.guet.edu.cn, li.jichang@pcl.ac.cn, rslan2016@163.com", + "bbox": [ + 194, + 272, + 797, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 321, + 326, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose FakeRadar, a novel deepfake video detection framework designed to address the challenges of cross-domain generalization in real-world scenarios. Existing detection methods typically rely on manipulation-specific cues, performing well on known forgery types but exhibiting severe limitations against emerging manipulation techniques. This poor generalization stems from their inability to adapt effectively to unseen forgery patterns. To overcome this, we leverage large-scale pretrained models (e.g. CLIP) to proactively probe the feature space, explicitly highlighting distributional gaps between real videos, known forgeries, and unseen manipulations. Specifically, FakeRadar introduces Forgery Outlier Probing, which employs dynamic subcluster modeling and cluster-conditional outlier generation to synthesize outlier samples near boundaries of estimated subclusters, simulating novel forgery artifacts beyond known manipulation types. Additionally, we design Outlier-Guided Tri-Training, which optimizes the detector to distinguish real, fake, and outlier samples using proposed outlier-driven contrastive learning and outlier-conditioned cross-entropy losses. Experiments show that FakeRadar outperforms existing methods across various benchmark datasets for deepfake video detection, particularly in cross-domain evaluations, by handling the variety of emerging manipulation techniques.", + "bbox": [ + 89, + 353, + 483, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 760, + 220, + 776 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, generative techniques [28, 58, 62] for creating deepfake videos have advanced rapidly, producing highly realistic facial forgeries that challenge human perceptual limits [8, 23]. However, the malicious use of such technology, such as fabricated news and identity fraud, poses a significant threat to digital media credibility [7, 9]. Ear", + "bbox": [ + 89, + 785, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d94756103d0736965c074e381e3861971875cef335c582a411d06cbb35a5cddb.jpg", + "image_caption": [ + "Figure 1. Conceptual pipeline of FakeRadar. (a)-(b) Forgery Outlier Probing first performs dynamic subcluster modeling on these features to partition finer-grained subclusters, then generates cluster-conditional outliers for each estimated subcluster to simulate unseen forgeries, emulating unknown manipulations. (c)-(d) The model is optimized using outlier-driven tri-training, allowing the detector to establish optimal boundaries for the real, fake, and outlier classes, with the latter two merged into a new \"Fake\" class during testing." + ], + "image_footnote": [], + "bbox": [ + 516, + 319, + 906, + 526 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lier deepfake detection methods primarily rely on the passive learning of specific forgery artifacts, such as boundary inconsistencies [27, 54], blinking anomalies [32], and texture irregularities [17, 35]. While these approaches perform well against conventional deepfake manipulations, including Face2Face [46] and NeuralTextures [47], they exhibit critical vulnerabilities when confronted with emerging generative paradigms, such as diffusion models [55]. For instance, boundary-based detection methods [30, 44] fail to identify faces generated by advanced GANs (e.g., StyleGAN [25]). Similarly, blink analysis-based techniques [32] are rendered ineffective when processing temporally smoothed fake videos [57]. These limitations stem from a fundamental shortcoming: existing methods learn a constrained decision boundary in the feature space, where", + "bbox": [ + 511, + 672, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Authors.", + "bbox": [ + 109, + 887, + 243, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13382", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "known forgery traces occupy only a localized region, while novel manipulation techniques introduce artifacts in arbitrary locations. As a result, they struggle to generalize to new manipulation techniques that introduce artifacts in previously unseen ways, leading to significant performance degradation in cross-domain scenarios [11, 12].", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To better understand its underlying cause, [2, 36] have analyzed the feature distributions of real and synthetic data. These studies reveal a fundamental embedding-space disparity: real data exhibit compact, dense distributions, whereas deepfake forgeries form scattered, sparse clusters due to their generation with diverse manipulation types, as demonstrated in Figure 1. This difference in distribution poses a significant challenge for existing detection methods, which are typically optimized to classify previously observed forgeries. Since deepfake forgeries do not adhere to a fixed pattern but rather introduce an infinite variety of modifications, classifiers trained on known forgery patterns struggle to recognize novel deepfakes that deviate from past observations. Therefore, a robust detection approach must move beyond merely learning pre-defined forgery patterns and instead develop the capability to identify anomalies that fall outside of previously known distributions. Given the limitations of existing methods in handling unseen forgery types, we propose a paradigm shift in deepfake detection. Instead of passively fitting to known forgery patterns, our approach proactively explores potential anomaly regions in the feature space. Inspired by radar systems that scan for unknown targets through frequency-spectrum probing, our method enhances detection robustness against novel deep-fake manipulations.", + "bbox": [ + 91, + 185, + 483, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose a new deepfake video detection framework, called FakeRadar, which leverages the deep feature priors of large-scale pre-trained models (i.e. CLIP [40]) to proactively \"probe\" unknown forgeries and enhance cross-domain generalization. Unlike conventional detectors [2, 5, 51] that perform only binary classification, FakeRadar explicitly introduces \"Forgery outliers\" during training to simulate potential unknown forgeries and thus \"pre-explore and cover\" the latent space. As shown in Figure 1 (a) and (b), we first employ Forgery Outlier Probing, which conducts dynamic subcluster modeling on the features of real and known deepfake videos. By approximating the embedding space via fine-grained Gaussian mixture distributions, this algorithm automatically partitions training samples into distinct subclusters. Next, through cluster-conditional outlier generation, Forgery Outlier Probing synthesizes outlier samples near the boundaries of these estimated subclusters with a certain \"novel offset\", thereby covering a broader range of potential forgery regions.", + "bbox": [ + 89, + 565, + 482, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Additionally, FakeRadar adopts a model optimization strategy called Outlier-Guided Tri-Training, enabling the model to discriminate among \"Real\", \"Fake\", and \"Out", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "lier” samples, with its details as illustrated in Figure 1 (c) and (d). Specifically, Outlier-Guided Tri-Training proposes an outlier-driven contrastive loss and an outlier-conditioned cross-entropy loss for end-to-end model optimization. The former loss enhances the distance between real, fake, and outlier features via contrastive learning, emphasizing their differences, while the latter one ensures that the model has clear decision boundaries for the three categories, especially ensuring that outliers are not incorrectly classified as real. In this triplet-class (Real/Fake/Outlier) setting, the model can independently label unknown forgeries during training. During inference, both “Fake” and “Outlier” are classified as forgeries, empowering the detector to adapt effectively to unknown deepfakes.", + "bbox": [ + 511, + 90, + 903, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We summarize FakeRadar's principal contributions as follows, with its conceptual pipeline illustrated in Figure 1:", + "bbox": [ + 511, + 301, + 903, + 332 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose FakeRadar, a novel framework for deepfake video detection, designed to address the critical challenge of cross-domain generalization in real-world scenarios.", + "- We introduce Forgery Outlier Probing for synthesizing outlier samples to cover a broader range of potential forger patterns, expanding the detector's understanding of forgery distribution.", + "- We design Outlier-Guided Tri-Training for model optimization, which enables end-to-end training for distinguishing real, fake, and outlier samples, improving detection of unknown forgeries.", + "- Numerous experiments have revealed that FakeRadar outperforms current leading algorithms across multiple deepfake benchmarks, particularly excelling in cross-domain evaluations where existing methods struggle." + ], + "bbox": [ + 513, + 333, + 903, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 513, + 573, + 651, + 588 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deepfake Detection. In response to the malicious use of deepfake techniques, numerous approaches have been explored to address this issue. Generally, deepfake detection is formulated as a binary classification problem (real or fake), leading researchers to develop a variety of end-to-end detectors [45, 54] that directly distinguish reals from manipulated content. Existing detection methods can be broadly categorized into two groups: image-level and video-level detection. Specifically, image-level detectors primarily focus on identifying spatial artifacts caused by forgeries. Examples include frequency anomalies [3, 16, 29, 39], identity-based discrepancies [13, 14, 24], blending artifacts [30, 44], and localized distortions [35, 49]. In contrast, video-level detectors employ a wider range of strategies. For example, some algorithms, such as [20, 32], concentrate on temporal consistency, typically reflected in correlations across consecutive frames. Other video detection studies [18, 38, 61] incorporate both temporal and spatiotemporal features to establish a more comprehensive detection framework. Another line of research [22, 56] examines audio-visual con", + "bbox": [ + 511, + 598, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13383", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/586c8307ffd116b23c397ed996a152fff5e4080e49e37e8fbc482682fa729c41.jpg", + "image_caption": [ + "Figure 2. An overview of FakeRadar for deepfake video detection. FakeRadar consists of two key components: Forgery Outlier Probing and Outlier-Guided Tri-Training. (i) Forgery Outlier Probing performs dynamic subcluster modeling to model the feature-space distributions of real and fake samples using GMMs, followed by dynamic subcluster splitting and merging to produce finer-grained subcluster partitions. Afterwards, FakeRadar conducts cluster-conditioned outlier generation to simulate unseen forgeries for novel manipulation techniques. (ii) Outlier-Guided Tri-Training jointly optimizes the CLIP model's ViT-B/16 backbone, inserted with ST-Adapter [37] for parameter-efficient fine tuning, and a triplet-class classifier to distinguish between \"Real\", \"Fake\", and \"Outlier\" classes, with \"Fake\" and \"Outlier\" merged into a single new \"Fake\" class during evaluation. The training set includes real videos, as well as deepfake videos generated using four different manipulations, i.e. \"DF\" [41], \"FS\" [26], \"NT\" [47] and \"F2F\" [46]." + ], + "image_footnote": [], + "bbox": [ + 101, + 87, + 903, + 277 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sistency by analyzing the synchronization and correlation between video and audio signals to detect forgeries. In summary, these methods rely heavily on existing artifacts in training set, limiting their generalization to unseen manipulation types. In contrast, our approach explores feature-space anomalies, enhancing the generalization capacity of our proposed detector to diverse deepfake manipulations.", + "bbox": [ + 89, + 409, + 483, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Towards Generalization in Deepfake Detection. Rapid advance in deepfakes have significantly improved the quality and authenticity of generated content, posing substantial challenges to the generalization capability of detection systems. Existing detectors [6, 43, 50] often suffer from severe performance degradation when encountering novel forgery techniques or previously unseen manipulations. To address this issue, researchers have proposed various strategies to enhance detector generalization. For instance, FTCN [61] improves generalization by capturing temporal coherence features in video sequences. Similarly, AltFreezing [51] employs an alternating weight-freezing training strategy to facilitate the joint detection of spatial and temporal forgery artifacts. Another algorithm, StyleFlow [5], leverages the temporal dynamics of style latent vectors to identify anomalous patterns in generated videos.", + "bbox": [ + 89, + 537, + 482, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "More recently, data synthesis has emerged as an effective strategy for improving detector generalization. For example, FWA [31] enhances training data by introducing facial warping artifacts, while Face X-ray [30] detects forgeries by identifying blending boundaries, independent of specific facial manipulation techniques. Building on this concept, SBI [44] synthesizes training samples by blending different views of the same pristine image to simulate generic forgery", + "bbox": [ + 89, + 780, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "artifacts. Additionally, a series of recent studies [1, 27, 34] have further advanced this direction by employing various data synthesis techniques. Unlike existing methods that primarily focus on image- or video-level data synthesis for deepfake detection, our approach proposes to explore unseen forgeries in feature space.", + "bbox": [ + 511, + 409, + 906, + 501 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 513, + 513, + 648, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 539, + 653, + 554 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview. In this paper, we introduce FakeRadar, to overcome the challenges of cross-domain generalization in deepfake video detection by exploring the feature space to identify emerging manipulation patterns. Specifically, FakeRadar is built upon two key components: (i) Forgery Outlier Probing, which models the feature-space distributions of subclusters for training samples of both real and fake categories, and then generates cluster-conditioned outlier samples to simulate unseen forgeries, and (ii) Outlier-Guided Tri-Training, which jointly optimizes an adapter-inserted backbone and a triplet-class classifier to distinguish among samples from the \"Real\", \"Fake\", and \"Outlier\" classes. An overview of FakeRadar can be found in Figure 2.", + "bbox": [ + 511, + 561, + 906, + 760 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Model Adaptation with a Frozen Backbone. The FakeRadar framework employs a pre-trained Vision Transformer (ViT-B) from CLIP [40] as a frozen backbone to preserve its learned semantic features. This model serves as the video encoder for input clips. To adapt it for deepfake detection, we integrate lightweight ST-Adapter layers [37] after each Transformer block, enabling parameter-efficient fine-tuning. This significantly enhances the model's ability to", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13384", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "capture spatial manipulations and temporal inconsistencies, which are crucial for detecting deepfake videos. Given a feature vector $x$ from the video encoder, each adapter transforms it as:", + "bbox": [ + 89, + 90, + 483, + 151 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S T - A d p a t e r} (x) = x + \\operatorname {R e L U} (\\operatorname {C o n v 3 D} (x W _ {\\text {d o w n}})) W _ {\\text {u p}}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 164, + 482, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\mathrm{Conv3D}(\\cdot)$ denotes a 3D convolution operation employed for capturing spatio-temporal features from videos. As well, $W_{\\mathrm{down}}$ and $W_{\\mathrm{up}}$ are weight matrices that perform dimensionality reduction and expansion, respectively, and $\\mathrm{ReLU}(\\cdot)$ is the activation function. By introducing only a small number of additional parameters, these adapters allow efficient fine-tuning while retaining the original backbone's representational strength. After feature extraction, each video clip is represented as a $d$ -dimensional vector for subsequent analysis, with $d = 768$ .", + "bbox": [ + 89, + 193, + 483, + 345 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inference. During inference, the model classifies input samples as either real or deepfake, with deepfakes categorized as \"Fake\" or \"Outlier\". The \"Fake\" class includes fake samples with known manipulation patterns, while the \"Outlier\" class detects testing samples with novel forgeries that differ from established manipulations. By unifying these categories, the model enhances its robustness against evolving manipulation techniques, enabling accurate detection of both familiar and emerging deepfakes. This improves its generalization performance in cross-domain evaluations.", + "bbox": [ + 89, + 364, + 483, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Forgery Outlier Probing", + "text_level": 1, + "bbox": [ + 89, + 525, + 316, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we introduce Forgery Outlier Probing (FOP) to generate outlier samples with unseen forgeries beyond the existing real and fake videos, simulating previously unobserved manipulation patterns. Specifically, FOP first presents dynamic subcluster modeling, which performs category-specific clustering on the training samples of real and fake categories. Then, it models the distributions of these categories using a strategy based on Gaussian Mixture Models (GMMs), and follows by applying dynamic subcluster splitting and merging, thereby achieving finer-grained subcluster partitioning. Finally, we generate cluster-conditioned outliers based on the resulting subclusters to stimulate unseen forgeries.", + "bbox": [ + 89, + 547, + 483, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dynamic Subcluster Modeling. To capture subtle intra-class variations within real and deepfake videos, we apply category-specific clustering on the FaceForensics++ (FF++) dataset. Specifically, we treat the real class and each of the four fake classes (i.e. four types of deepfake videos created by different manipulation techniques [26, 41, 46, 47]) as five separate categories. For each category $c$ , let $\\mathcal{X}^{(c)} = \\{\\mathbf{x}_1, \\mathbf{x}_2, \\dots, \\mathbf{x}_N\\}$ denote the set of feature embeddings (i.e. extracted from the video encoder). Inspired by [42], we then", + "bbox": [ + 89, + 763, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "define a main clustering network:", + "bbox": [ + 513, + 90, + 736, + 106 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\operatorname {m a i n}} ^ {(c)} \\left(\\mathbf {x} _ {i}\\right) \\mapsto \\mathbf {r} _ {i} = \\left(r _ {i, 1}, r _ {i, 2}, \\dots , r _ {i, K}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 114, + 906, + 135 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which initially assumes $K$ clusters, empirically with $K = 5$ , providing a \"soft\" assignment of each embedding $\\mathbf{x}_i \\in \\mathcal{X}^{(c)}$ to these $K$ clusters. Here, $r_{i,k} \\in [0,1]$ and $\\sum_{k=1}^{K} r_{i,k} = 1$ .", + "bbox": [ + 511, + 142, + 906, + 205 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Each cluster can be modeled as a Gaussian distribution using a GMM, allowing us to represent the feature distribution of each category in a probabilistic manner. Specifically, for each cluster $k$ in category $c$ , we define a Gaussian component characterized by $\\{\\pi_k,\\pmb {\\mu}_k,\\pmb {\\Sigma}_k\\}$ , where $\\pi_{k}\\in (0,1)$ is the mixture weight for cluster $k$ , with $\\sum_{k = 1}^{K}\\pi_{k} = 1$ , while $\\pmb{\\mu}_{k}$ and $\\pmb{\\Sigma}_{k}$ represent its corresponding mean and covariance matrix. These parameters enable the clustering network to estimate the likelihood of each data sample belonging to different clusters, aligning our model with a GMM-based formulation.", + "bbox": [ + 511, + 205, + 906, + 371 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After that, the main clustering network needs to learn these parameters by aligning its soft assignments with GMM responsibilities:", + "bbox": [ + 511, + 373, + 905, + 417 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nr _ {i, k} ^ {E} = \\frac {\\pi_ {k} \\mathcal {N} \\left(\\mathbf {x} _ {i} ; \\boldsymbol {\\mu} _ {k} , \\boldsymbol {\\Sigma} _ {k}\\right)}{\\sum_ {k ^ {\\prime} = 1} ^ {K} \\pi_ {k ^ {\\prime}} \\mathcal {N} \\left(\\mathbf {x} _ {i} ; \\boldsymbol {\\mu} _ {k ^ {\\prime}} , \\boldsymbol {\\Sigma} _ {k ^ {\\prime}}\\right)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 426, + 906, + 467 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{N}(\\cdot ;\\pmb {\\mu}_k,\\pmb {\\Sigma}_k)$ is the multivariate Gaussian density function. The network is trained using the KL-divergence loss:", + "bbox": [ + 511, + 472, + 905, + 515 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m a i n}} = \\sum_ {i = 1} ^ {N} \\mathrm {K L} \\left(\\mathbf {r} _ {i} \\| \\mathbf {r} _ {i} ^ {E}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 515, + 905, + 555 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which ensures that the learned cluster assignments match the inferred GMM responsibilities.", + "bbox": [ + 511, + 559, + 905, + 588 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To further refine cluster structures, we introduce a subclustering network for each cluster $k$ , which attempts to split the cluster into two subclusters if multi-modal structures are detected. A compactness loss is applied to encourage subclusters to be well-separated. This loss function penalizes subclusters that are too close to each other:", + "bbox": [ + 511, + 589, + 905, + 679 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {s u b}} = \\sum_ {k = 1} ^ {K} \\sum_ {j = 1} ^ {2} \\sum_ {\\mathbf {x} _ {i} \\in X _ {k}} \\tilde {r} _ {i, j} \\| \\mathbf {x} _ {i} - \\tilde {\\boldsymbol {\\mu}} _ {k, j} \\| _ {2} ^ {2}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 688, + 905, + 732 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tilde{r}_{i,j}$ represents the probability of $\\mathbf{x}_i$ belonging to subcluster $j$ , and $\\tilde{\\mu}_{k,j}$ is its mean.", + "bbox": [ + 511, + 738, + 905, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the predefined cluster number $K$ may not be optimal, we propose to dynamically adjust it through split/merge proposals, inspired by [42]. Suppose a cluster $k$ contains $N_{k}$ points, $X_{k}$ . We split if subclustering indicates that $X_{k}$ is better viewed as two subgroups, $\\{X_{k,1}, X_{k,2}\\}$ . The Hastings ratio is given by", + "bbox": [ + 511, + 768, + 905, + 859 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nH _ {s} = \\frac {\\alpha \\Gamma (N _ {k , 1}) f _ {x} (X _ {k , 1}) \\Gamma (N _ {k , 2}) f _ {x} (X _ {k , 2})}{\\Gamma (N _ {k}) f _ {x} (X _ {k})}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 867, + 905, + 905 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13385", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{x}(\\cdot)$ is the marginal likelihood under a weak NIw prior, $\\alpha$ is the Dirichlet-process concentration parameter, and $\\Gamma (\\cdot)$ is the Gamma function.", + "bbox": [ + 89, + 90, + 480, + 135 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "If $H_{s}$ exceeds a threshold, cluster $k$ is split into two new clusters. The parameters for these new clusters are initialized from $\\tilde{\\mu}_{k,1},\\tilde{\\Sigma}_{k,1}$ and $\\tilde{\\mu}_{k,2},\\tilde{\\Sigma}_{k,2}$ , respectively. Merging is similarly proposed if two clusters significantly overlap, with acceptance ratio $H_{m} = 1 / H_{s}$ . Repeating these steps yields an effective number of clusters $K^{(c)}$ , plus subclusters for each category $c$ .", + "bbox": [ + 89, + 136, + 480, + 241 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After these dynamic adjustments, we obtain the final subcluster set $\\hat{\\mathcal{C}} = \\{\\hat{C}_1,\\hat{C}_2,\\dots,\\hat{C}_{K^{(c)}}\\}$ , where $K^{(c)}$ is the refined number of clusters. Each adjusted subcluster $\\hat{C}_k$ is characterized by its updated mean and covariance:", + "bbox": [ + 89, + 242, + 480, + 301 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {\\mu}} _ {k} = \\frac {1}{| \\hat {C} _ {k} |} \\sum_ {\\mathbf {x} _ {i} \\in \\hat {C} _ {k}} \\mathbf {x} _ {i}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 311, + 482, + 351 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {\\Sigma}} = \\frac {1}{| \\hat {C} _ {k} |} \\sum_ {\\mathbf {x} _ {i} \\in \\hat {C} _ {k}} \\left(\\mathbf {x} _ {i} - \\hat {\\boldsymbol {\\mu}} _ {k}\\right) \\left(\\mathbf {x} _ {i} - \\hat {\\boldsymbol {\\mu}} _ {k}\\right) ^ {\\top}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 362, + 482, + 401 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use online estimation for efficient training by maintaining a cluster-conditional queue with $|\\hat{C}_k|$ sample instances from each subcluster. In each iteration, we enqueue the embeddings of objects to their corresponding cluster-conditional queues and dequeue the same number of object embeddings.", + "bbox": [ + 89, + 406, + 483, + 497 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Cluster-Conditional Outlier Generation. After deriving finer-grained cluster partitions for each category, we generate feature-space outliers to probe potential unknown forgeries. Motivated by [15], in order to ensure that the sampled outliers lie near the cluster boundary, we maintain a sufficiently small value $\\epsilon$ for outlier sampling from the $\\epsilon$ -likelihood region of the estimated cluster-conditional distribution. Therefore, for a subcluster from $\\hat{C}$ indexed by $k$ , the outlier samples it generates can be formulated as follows:", + "bbox": [ + 89, + 515, + 483, + 650 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {V} _ {k} = \\left\\{\\mathbf {v} _ {k} \\mid \\frac {\\exp \\left(- \\frac {1}{2} \\left(\\mathbf {v} _ {k} - \\hat {\\boldsymbol {\\mu}} _ {k}\\right) ^ {\\top} \\hat {\\boldsymbol {\\Sigma}} ^ {- 1} \\left(\\mathbf {v} _ {k} - \\hat {\\boldsymbol {\\mu}} _ {k}\\right)\\right)}{(2 \\pi) ^ {d / 2} | \\hat {\\boldsymbol {\\Sigma}} | ^ {1 / 2}} < \\epsilon \\right\\}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 661, + 486, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $v_{k}\\sim \\mathcal{N}(\\hat{\\mu}_{k},\\hat{\\Sigma}_{k})$ represents the virtual outliers sampled from the estimated Gaussian distribution of subcluster $k$ , with $d$ denoting its embedding dimensionality. It is worth noting that during model optimization, all outliers generated here, i.e. $\\overline{\\mathcal{V}} = \\mathcal{V}_1\\cup \\dots \\cup \\mathcal{V}_k\\cup \\dots \\mathcal{V}_{|\\hat{\\mathcal{O}} |}$ are classified into the \"Outlier\" class, separate from the real and fake classes.", + "bbox": [ + 89, + 718, + 483, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Outlier-Guided Tri-Training", + "text_level": 1, + "bbox": [ + 89, + 833, + 346, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To enable better performance of FakeRadar for tasks of deepfake video detection, we design a strategy called Outlier-Guided Tri-Training for model optimization jointly", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "combining proposed outlier-driven contrastive loss and outlier-conditioned cross-entropy loss. With this approach, a triplet-class classifier (integrated with the pre-trained backbone and adapters) is trained to explicitly distinguish among three categories: \"Real\", \"Fake\", and \"Outlier\". This design not only improves the model's ability to detect known forgeries but also sensitizes it to emerging manipulation patterns.", + "bbox": [ + 511, + 90, + 903, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Outlier-Driven Contrastive Loss. In this work, we propose an outlier-driven contrastive loss to enhance the model's ability to delineate class boundaries within the feature space, thereby improving feature discriminability. This loss function is designed to maximize the similarity between samples and their respective subcluster centers while minimizing their similarity to other subcluster centers. Additionally, it treats generated outliers as negative examples, enforcing low similarity between samples and outliers to strengthen boundary discrimination.", + "bbox": [ + 511, + 231, + 903, + 381 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To this end, the loss is formulated using an InfoNCE-like contrastive loss given an input sample $i$ :", + "bbox": [ + 511, + 382, + 903, + 412 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o n}} = - \\log \\frac {\\exp \\left(\\frac {s \\left(h _ {i} , \\mu_ {k _ {i} ^ {+}}\\right)}{\\tau}\\right)}{\\sum_ {k ^ {-}} \\exp \\left(\\frac {s \\left(h _ {i} , \\mu_ {k ^ {-}}\\right)}{\\tau}\\right) + \\sum_ {v _ {j} \\in \\overline {{\\nu}}} \\exp \\left(\\frac {s \\left(h _ {i} , v _ {j}\\right)}{\\tau}\\right)}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 424, + 916, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $h_i$ denotes the feature representation of sample $i$ , while $\\mathcal{V}$ represents the outlier cache, providing additional hard negative samples to improve the model's ability to distinguish boundary samples. Additionally, $k^{-}$ represents all subclusters except the one to which sample $i$ belongs, with its corresponding $u_{k^{-}}$ denoting the centers of these subclusters, used as negative samples. The function $s(\\cdot ,\\cdot)$ denotes cosine similarity, which quantifies the similarity between samples, and $\\tau$ is a temperature hyperparameter that controls the smoothness of similarity scaling.", + "bbox": [ + 511, + 492, + 905, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Outlier-Conditioned Cross-Entropy Loss. While the proposed outlier-driven contrastive loss enforces stronger feature-space separation, we still require a concrete classification objective to handle a three-category task: real, fake, and outlier. In particular, the model must not only determine whether a video is fake or real but also assign a separate label for those samples situated near decision boundaries (i.e., outliers). This triplet-class cross-entropy loss is crucial, as it enables the model to explicitly differentiate uncertain boundary samples and label them as outliers, thereby avoiding confusion.", + "bbox": [ + 511, + 662, + 905, + 828 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Concretely, the triplet-class cross-entropy loss can be defined as follows:", + "bbox": [ + 511, + 829, + 903, + 858 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c l s}} = - \\sum_ {c \\in \\{\\text {R e a l}, \\text {F a k e}, \\text {O u t l i e r} \\}} y _ {c} \\log p _ {c}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 869, + 903, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13386", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/6788ca724ad03870e4dc24d83dfcd3a16c6be15bf3b90673dd91f1878b67ed37.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBackboneInput TypeTesting Set AUC (%)
FF++CDFv2DFDCPDFDCDFD
Xception [6]XceptionFrame99.373.7-70.9-
Face X-Ray [30]HRNetFrame97.879.5-65.595.4
PCL+i2G [60]ResnetFrame99.890.074.467.5-
SLADD [4]XceptionFrame98.479.776.0--
UIA-ViT [63]Vision TransformerFrame99.382.475.8-94.7
UCF [53]XceptionFrame-82.4-80.594.5
SeeABLE [27]EfficientNetFrame-87.386.375.9-
LSDA [54]EfficientNetFrame-91.181.277.095.6
FTCN [61]3D ResNetVideo99.886.9-74.0-
RealForensics [21]3D ResNetVideo99.886.9-75.9-
LTTD [19]Vision TransformerVideo99.589.3-80.4-
AltFreezing [51]3D ResNetVideo99.789.5--93.7
TALL [52]Swin TransformerVideo99.990.8-76.8-
StyleFlow [5]3D ResNetVideo99.189.0--96.1
NACO [59]Vision TransformerVideo99.889.5-76.7-
FakeRadar (Ours)Vision TransformerVideo99.191.788.584.196.2
", + "bbox": [ + 236, + 88, + 759, + 297 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $y_{c}$ is the ground-truth label for samples belonging to class $c$ , and $p_{c}$ is the model-predicted probability of samples from that class, with $c =$ \"Outlier\" denoting the outlier label. This loss compels the model to accurately partition each category in the feature space, especially for samples near class boundaries.", + "bbox": [ + 89, + 357, + 482, + 448 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Total Loss. Finally, we combine both losses into the total objective for model optimization:", + "bbox": [ + 89, + 465, + 482, + 498 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {c o n}} + \\lambda \\mathcal {L} _ {\\text {c l s}}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 508, + 480, + 525 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda = 0.5$ is a hyperparameter balancing the relative importance of classification and contrastive goals. In practice, each mini-batch comprises 16 training samples $(\\mathrm{FF}++)$ , and 16 outlier samples during model training.", + "bbox": [ + 89, + 535, + 483, + 598 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 609, + 223, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experiment Setups", + "text_level": 1, + "bbox": [ + 89, + 633, + 274, + 650 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. Following established experimental practices [5, 51, 52, 54], this study primarily utilizes the high-quality (HQ) version of the FaceForensics++ dataset [43], denoted as $\\mathrm{FF} + + (\\mathrm{HQ})$ , for FakeRadar's evaluation. This dataset consists of 1000 pristine videos and 4000 deepfakes videos generated using four different manipulation techniques: DFD [41], F2F [46], FS [26], and NT [47]. The dataset is split into training, validation, and testing sets according to the official protocol. This study primarily focuses on evaluating the cross-domain performance of our method. To this end, we incorporate four additional datasets for evaluation, including CDFv2 [33], DFDCP [11], DFDC [12], and DFD [41].", + "bbox": [ + 89, + 656, + 482, + 852 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. To ensure fair comparisons, we maintain most implementation details consistent with exist", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e97d01c9ae6b6f12c97ef8a97aac64da06b080237996b1455720a0af81b4ced0.jpg", + "table_caption": [ + "Table 1. Generalization comparisons of FakeRadar and state-of-the-art algorithms on cross-dataset evaluation, using video-level AUC (%) as the metric." + ], + "table_footnote": [], + "table_body": "
Training SetMethodDFF2FFSNTavg.
DFDCL [45]99.9877.1361.0175.0178.28
UIA-ViT [63]99.2974.4453.8970.9274.64
TALL [52]99.3569.3354.3874.7474.45
AltFreezing [51]98.8069.3372.0777.4179.40
FakeRadar (Ours)99.9586.4788.1266.3682.23
F2FDCL [45]91.9199.2159.5866.6779.34
UIA-ViT [63]83.3998.3268.3767.1779.31
TALL [52]85.5298.6762.1570.2479.15
AltFreezing [51]81.1298.8062.9770.2678.29
FakeRadar (Ours)96.9499.0982.3368.4686.75
FSDCL [45]74.8069.7599.9052.6074.26
UIA-ViT [63]81.0266.3099.0449.2673.91
TALL [52]77.8367.3098.7752.5774.12
AltFreezing [51]80.6765.9499.7857.2375.90
FakeRadar (Ours)96.3982.0699.8352.6482.73
NTDCL [45]91.2352.1379.3198.9780.41
UIA-ViT [63]79.3767.8945.9494.5971.95
TALL [52]84.0272.8551.6695.1675.92
AltFreezing [51]88.8176.7857.9295.7179.80
FakeRadar (Ours)96.7183.3467.5098.4186.49
", + "bbox": [ + 514, + 354, + 906, + 601 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Generalization comparisons of FakeRadar and state-of-the-art algorithms on cross-manipulation evaluation, using video-level AUC (\\%) as the metric.", + "bbox": [ + 511, + 607, + 906, + 650 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ing methods [5, 51, 59]. However, our FakeRadar framework employs ViT-B/16 from CLIP [40] as the backbone, into which we insert an ST-Adapter [37] with a 384-dimensional bottleneck for parameter-efficient fine-tuning. Additionally, the classification head of our triplet-class classifier consists of a single linear layer with softmax activation. During model training, we sample four temporal clips from each video, each containing 12 consecutive frames. We use Adam as the optimizer and apply a cosine learning rate scheduler with an initial learning rate of $1 \\times 10^{-4}$ . The model is trained for 60 epochs with a batch size of 16 and a weight decay of $5 \\times 10^{-4}$ . Prior to model optimization, we perform data preprocessing by detecting faces with RetinaFace [10] and then standardize all face images", + "bbox": [ + 511, + 688, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13387", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/922b9eeaef948d8603a75b38ca1f8e6234094f1e1a4be5d0b342ec5031c10e7e.jpg", + "image_caption": [ + "Figure 3. Evaluation of (a) Forgery Outlier Probing and (b)/(c) Outlier-Guided Tri-Training by FakeRadar. (a) \"FakeRadar w/o Dyn. Sub. M.\" and \"FakeRadar w/o Out. Gen.\" indicate that dynamic subcluster modeling and cluster-conditional outlier generation have been removed. The model is trained on FF++(HQ) and evaluated on DFDC [12], CDFv2 [33], and DFD [41]. (b) and (c) show, respectively, the count of real and fake test samples classified as \"fake\" across predicted probability intervals. The blue solid line denotes the binary classifier's distribution (distinguishing only real vs. fake), while the red dashed line denotes that of our proposed triplet classifier, which treats both fake and outlier samples as the new \"fake\". In this evaluation, the model is trained on FF++(HQ) and tested on CDFv2 [33]." + ], + "image_footnote": [], + "bbox": [ + 89, + 85, + 356, + 236 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/31311a0a131bb374e262660ed95c596955f78aaafd597515461b720503e075eb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 85, + 630, + 236 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/68738ff670573c43f6792a7e60db13307fc720d2a45c0edf65643c50fc6dc79b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 85, + 901, + 236 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "to $224\\times 224$ pixels.", + "bbox": [ + 89, + 345, + 217, + 361 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation Metrics. In this work, consistent with previous studies [5, 51, 52] employing video-level inputs, we use the area under the ROC curve (AUC) as the performance metric during evaluations. For methods that use images as model inputs, e.g. [27, 54], their video-level results are derived by averaging the predictions across all video frames.", + "bbox": [ + 89, + 380, + 483, + 472 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Comparisons with the State-of-the-Arts", + "text_level": 1, + "bbox": [ + 89, + 481, + 429, + 498 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this study, we focus on evaluating the capabilities of cross-domain generalization for our detector to validate its effectiveness in detecting deepfake videos manipulated through various techniques. To achieve this, we employ two strategies for evaluation: cross-dataset evaluation and cross-manipulation evaluation, both of which demonstrate that our results predominantly surpass existing state-of-the-art (SoTA) algorithms for deepfake video detection, as illustrated by Table 1 and Table 2.", + "bbox": [ + 88, + 503, + 483, + 640 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Cross-Dataset Evaluation. In this situation, we train our FakeRadar on the FF++(HQ) and test its performance on multiple benchmark datasets, including CDFv2 [33], DFD [41], DFDC [12], and DFDCP [11]. As reported in Table 1, FakeRadar demonstrates its superior generalization capability in cross-dataset evaluation, achieving the best performance across all cross-dataset cases while maintaining comparable performance on the FF++ test dataset against SoTA algorithms. Specifically, our algorithm significantly improves over other methods with images and videos as inputs, particularly outperforming the current best methods, UCF [53] and LTTD [19], by $3.6\\%$ and $3.7\\%$ in the AUC metric, respectively, on the dataset DFDC [12]. These results robustly showcase the proposed method's enhanced ability of generalization to handle unseen manipulation techniques in cross-dataset evaluation scenarios.", + "bbox": [ + 89, + 659, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4e7cdb2f535cd0c34a83f91c44d61b391575912f41095b67e457561f01c269d5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
M-(#)FOPOGTTCDFDFDCDFDCPDFDavg.
ODCLOCCE
1✓✓✓91.684.188.596.290.1
2✓✓89.981.285.694.987.9
3✓✓88.680.987.794.788.0
4✓88.880.286.794.587.6
5✓88.478.485.194.386.7
688.278.384.894.286.4
", + "bbox": [ + 516, + 343, + 903, + 431 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Ablation study results of FakeRadar by removing its one or more proposed components, using video-level AUC (\\%) as the metric. For clarity, \"FOP\", \"OGTT\", \"ODCL\", and \"OCCE\" refer to Forgery Outlier Probing, Outlier-Guided Tri-Training, Outlier-Driven Contrastive Loss, and Outlier-Conditioned Cross-Entropy Loss, respectively.", + "bbox": [ + 511, + 441, + 906, + 525 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Cross-Manipulation Evaluation. To further assess our model's generalization across different manipulation techniques, we conduct cross-manipulation evaluation on the dataset FF++(HQ). In this case, we train the model using only real samples from FF++(HQ) and one manipulation type of deepfake samples, then test its performance on the remaining three manipulation types within FF++(HQ). As shown in Table 2, the proposed FakeRadar achieves the best average performance in all testing cases, particularly excelling in scenarios trained with F2F as the manipulation type, surpassing the second-best performing method, DCL [45], by $7.41\\%$ in AUC. In evaluations focused on a single manipulation type, our method also achieves the best performance in most cases and the second-best in a few. This indicates that our approach, in cross-manipulation evaluation, demonstrates a performance advantage in cross-domain generalization compared to current SoTA algorithms of detecting deepfake videos.", + "bbox": [ + 511, + 534, + 906, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Analysis", + "text_level": 1, + "bbox": [ + 511, + 816, + 687, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct an ablation study to assess the contribution of each component in FakeRadar, with results presented in Table 3, Figure 3, and Figure 4. Experiments are performed in a cross-dataset setting, where the model is trained on", + "bbox": [ + 511, + 839, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13388", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$\\mathrm{FF}++(\\mathrm{HQ})$ and tested on CDF, DFDC, DFDCP, and DFD, with average performance reported. As shown in Table 3 (Rows M-(1) to M-(6)), removing individual or multiple components leads to a progressive decline in AUC, from $90.1\\%$ to $86.4\\%$ , demonstrating the significance of each component. Below, we analyze the impact of FakeRadar's key components in detail.", + "bbox": [ + 89, + 90, + 483, + 196 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of Forgery Outlier Probing. To evaluate the impact of FOP, we remove or modify its submodules: dynamic subcluster modeling and cluster-conditional outlier generation. In \"FakeRadar w/o Dyn. Sub. M.\", we disable dynamic subcluster modeling, generating outliers only from clusters of samples of real and fake categories in training set. In \"FakeRadar w/o Out. Gen.\", we replace outlier samples with Gaussian noise. As shown in Figure 3(a), removing dynamic subcluster modeling reduces AUC on CDFv2, DFDC, and DFD to $90.2\\%$ , $82.3\\%$ , and $94.7\\%$ , respectively. While removing outlier generation has a smaller impact, performance remains below the complete FakeRadar framework. On CDFv2, FakeRadar $(91.7\\%)$ outperforms \"FakeRadar w/o Out. Gen.\" by $2.0\\%$ , demonstrating that FOP enhances generalization by generating effective outlier samples that improve the detection of novel deep-fake manipulations.", + "bbox": [ + 89, + 223, + 483, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of Outlier-Guided Tri-Training. Comparing Table 3 (Rows M-(1) to M-(3)), removing Outlier-Driven Contrastive Loss or Outlier-Conditioned Cross-Entropy Loss leads to an AUC drop from $90.1\\%$ to $87.9\\%$ and $88.9\\%$ , respectively, highlighting their importance. Since both losses refine the use of outlier samples, removing them would render FOP ineffective.", + "bbox": [ + 89, + 506, + 483, + 611 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The core of our proposed optimization strategy is the integration of a triplet classifier to handle the outlier samples generated by FOP. To assess its impact, we replace the triplet classifier, which performs three-class classification (real, fake, and outlier), with a binary classifier that only distinguishes between real and fake. We then analyze the prediction probability distributions for cross-dataset evaluation, as shown in Figure 3 (b) and (c). As illustrated, for real samples, both the binary and triplet classifiers exhibit similar prediction patterns with high confidence in the low-probability region. However, for fake samples, the triplet classifier demonstrates a clear advantage. Compared to the binary classifier, it assigns higher forgery probabilities to a greater number of fake samples, indicating that the proposed outlier-guided tri-training increases the model's confidence in detecting unseen fake samples. This result validates the effectiveness of treating outlier samples as fake samples during inference and highlights its contribution to improving the model's generalization capability.", + "bbox": [ + 89, + 613, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/226d93e674b4791eef2686dd56434c9ff217a868a209b633ea98a8efe1423524.jpg", + "image_caption": [ + "Figure 4. Feature visualization using t-SNE. We visualize the features of (a) the pre-trained CLIP ViT-B model and our (b) FakeRadar here. We train both models using the training set FF++(HQ), while testing on the FF++ test set." + ], + "image_footnote": [], + "bbox": [ + 526, + 90, + 903, + 195 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Feature Visualization. We use t-SNE [48] to visualize the features of our FakeRadar and the pre-trained CLIP ViT-B model. As shown in Figure 4, the feature distribution of FakeRadar exhibits a more compact clustering structure and clearer boundaries between real and fake videos, compared to the pre-trained CLIP ViT-B model. Furthermore, FakeRadar effectively distinguishes real samples from the four types of forgery samples. This suggests that FakeRadar not only enables the model to learn discriminative features between categories but also establishes more robust decision boundaries with the aid of outlier samples. These feature distribution patterns underscore the advantages of FakeRadar in improving the model's generalization ability. The boundaries between categories are sharper, and sufficient tolerance space is provided, which is essential for addressing unseen forgery types in real-world scenarios.", + "bbox": [ + 511, + 277, + 908, + 521 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 513, + 551, + 640, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduced FakeRadar, a novel deepfake video detection framework designed to address the limitations of existing methods in cross-domain generalization. Unlike traditional detectors that passively learn specific forgery patterns, FakeRadar proactively probes unknown forgeries by incorporating Forgery Outlier Probing to simulate diverse unseen manipulation patterns. This approach enables the model to explore a broader range of potential forgeries in feature space, significantly enhancing its adaptability to novel manipulations. To further improve detection robustness, we proposed Outlier-Guided Tri-Training, allowing the model to distinguish between real, fake, and outlier samples. This training paradigm, introduced outlier-driven contrastive loss and outlier-conditioned cross-entropy loss, sharpens the decision boundaries and enhances the detector's capacity to generalize across different manipulation types. Extensive experiments on benchmark datasets confirm that FakeRadar consistently outperforms state-of-the-art detection algorithms, particularly in cross-domain evaluations where conventional detectors struggle.", + "bbox": [ + 511, + 583, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13389", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 203, + 187, + 218 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Weiming Bai, Yufan Liu, Zhipeng Zhang, Bing Li, and Weiming Hu. Aunet: Learning relations between action units for face forgery detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24709-24719, 2023. 3", + "[2] Junyi Cao, Chao Ma, Taiping Yao, Shen Chen, Shouhong Ding, and Xiaokang Yang. End-to-end reconstruction-classification learning for face forgery detection. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4103-4112, 2022. 2", + "[3] Keshigeyan Chandrasegaran, Ngoc-Trung Tran, and Ngai-Man Cheung. A closer look at fourier spectrum discrepancies for cnn-generated images detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7200-7209, 2021. 2", + "[4] Liang Chen, Yong Zhang, Yibing Song, Lingqiao Liu, and Jue Wang. Self-supervised learning of adversarial example: Towards good generalizations for deepfake detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18710-18719, 2022. 6", + "[5] Jongwook Choi, Taehoon Kim, Yonghyun Jeong, Seungryul Baek, and Jongwon Choi. Exploiting style latent flows for generalizing deepfake video detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1133-1143, 2024. 2, 3, 6, 7", + "[6] François Chollet. Xception: Deep learning with depthwise separable convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1251-1258, 2017. 3, 6", + "[7] Danielle Keats Citron and Robert M. Chesney. Deepfakes and the new disinformation war. Foreign Affairs, 2018. 1", + "[8] Di Cooke, Abigail Edwards, Sophia Barkoff, and Kathryn Kelly. As good as a coin toss: Human detection of a-generated images, videos, audio, and audiovisual stimuli. ArXiv, abs/2403.16760, 2024. 1", + "[9] Yunshu Dai, Jianwei Fei, and Fangjun Huang. Idguard: Robust, general, identity-centric poi proactive defense against face editing abuse. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11934-11943, 2024. 1", + "[10] Jiankang Deng, Jia Guo, Evangelos Ververas, Irene Kotsia, and Stefanos Zafeiriou. Retinaface: Single-shot multi-level face localisation in the wild. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5202-5211, 2020. 6", + "[11] B Dolhansky. The dee pfake detection challenge (dfdc) pre view dataset. arXiv preprint arXiv:1910.08854, 2019. 2, 6, 7" + ], + "bbox": [ + 93, + 227, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "This work was supported in part by the National Key R&D Program of China (2024YFB3908503), in part by the Guangxi Science and Technology Project (AB25069496), and in part by the National Natural Science Foundation of China (62172120 and 62322608).", + "[12] Brian Dolhansky, Joanna Bitton, Ben Pflaum, Jikuo Lu, Russ Howes, Menglin Wang, and Cristian Canton Ferrer. The deepfake detection challenge (dfdc) dataset. arXiv preprint arXiv:2006.07397, 2020. 2, 6, 7", + "[13] Chengdong Dong, Ajay Kumar, and Eryun Liu. Think twice before detecting gan-generated fake images from their spectral domain imprints. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7855-7864, 2022. 2", + "[14] Shichao Dong, Jin Wang, Renhe Ji, Jiajun Liang, Haoqiang Fan, and Zheng Ge. Implicit identity leakage: The stumbling block to improving deepfake detection generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3994-4004, 2023. 2", + "[15] Xuefeng Du, Zhaoning Wang, Mu Cai, and Sharon Li. Towards unknown-aware learning with virtual outlier synthesis. In International Conference on Learning Representations, 2022. 5", + "[16] Joel Frank, Thorsten Eisenhofer, Lea Schonherr, Asja Fischer, Dorothea Kolossa, and Thorsten Holz. Leveraging frequency analysis for deep fake image recognition. In International conference on machine learning, pages 3247-3258. PMLR, 2020. 2", + "[17] Qiqi Gu, Shen Chen, Taiping Yao, Yang Chen, Shouhong Ding, and Ran Yi. Exploiting fine-grained face forgery clues via progressive enhancement learning. In AAAI Conference on Artificial Intelligence, 2021. 1", + "[18] Zhihao Gu, Yang Chen, Taiping Yao, Shouhong Ding, Jilin Li, Feiyue Huang, and Lizhuang Ma. Spatiotemporal inconsistency learning for deepfake video detection. In Proceedings of the 29th ACM international conference on multimedia, pages 3473-3481, 2021. 2", + "[19] Jiazhi Guan, Hang Zhou, Zhibin Hong, Errui Ding, Jingdong Wang, Chengbin Quan, and Youjian Zhao. Delving into sequential patches for deepfake detection. Advances in Neural Information Processing Systems, 35:4517-4530, 2022. 6, 7", + "[20] Alexandros Haliassos, Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. Lips don't lie: A generalisable and robust approach to face forgery detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5039-5049, 2021. 2", + "[21] Alexandros Haliassos, Rodrigo Mira, Stavros Petridis, and Maja Pantic. Leveraging real talking faces via self-supervision for robust forgery detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14950-14962, 2022. 6", + "[22] Alexandros Haliassos, Rodrigo Mira, Stavros Petridis, and Maja Pantic. Leveraging real talking faces via self-supervision for robust forgery detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14950-14962, 2022. 2", + "[23] Ammarah Hashmi, Sahibzada Adil Shahzad, Chia-Wen Lin, Yu Tsao, and Hsin-Min Wang. Unmasking illusions: Understanding human perception of audiovisual deepfakes. ArXiv, abs/2405.04097, 2024. 1", + "[24] Baojin Huang, Zhongyuan Wang, Jifan Yang, Jiaxin Ai, Qin Zou, Qian Wang, and Dengpan Ye. Implicit identity" + ], + "bbox": [ + 93, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13390", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "driven deepfake face swapping detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4490-4499, 2023. 2", + "[25] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1", + "[26] Marek Kowalski. Faceswap. https://github.com/MarekKowalski/FaceSwap, 2018. GitHub repository, accessed on March 3, 2025. 3, 4, 6", + "[27] Nicolas Larue, Ngoc-Son Vu, Vitomir Struc, Peter Peer, and Vassilis Christophides. Seeable: Soft discrepancies and bounded contrastive learning for exposing deepfakes. In Proceedings of the IEEE/CVF international conference on computer vision, pages 21011-21021, 2023. 1, 3, 6, 7", + "[28] Hong Li, Yutang Feng, Song Xue, Xuhui Liu, Bohan Zeng, Shanglin Li, Boyu Liu, Jianzhuang Liu, Shumin Han, and Baochang Zhang. Uv-idm: Identity-conditioned latent diffusion model for face uv-texture generation. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10585-10595, 2024. 1", + "[29] Hanzhe Li, Jiaran Zhou, Yuezun Li, Baoyuan Wu, Bin Li, and Junyu Dong. Freqblender: Enhancing deepfake detection by blending frequency knowledge. arXiv preprint arXiv:2404.13872, 2024. 2", + "[30] Lingzhi Li, Jianmin Bao, Ting Zhang, Hao Yang, Dong Chen, Fang Wen, and Baining Guo. Face x-ray for more general face forgery detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5001-5010, 2020. 1, 2, 3, 6", + "[31] Yuezun Li and Siwei Lyu. Exposing deepfake videos by detecting face warping artifacts. arXiv preprint arXiv:1811.00656, 2018. 3", + "[32] Yuezun Li, Ming-Ching Chang, and Siwei Lyu. In ictu oculi: Exposing ai created fake videos by detecting eye blinking. In 2018 IEEE International Workshop on Information Forensics and Security (WIFS), pages 1-7, 2018. 1, 2", + "[33] Yuezun Li, Xin Yang, Pu Sun, Honggang Qi, and Siwei Lyu. Celeb-df: A large-scale challenging dataset for deepfake forensics. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3204-3213, 2020. 6, 7", + "[34] Yuzhen Lin, Wentang Song, Bin Li, Yuezun Li, Jiangqun Ni, Han Chen, and Qiushi Li. Fake it till you make it: Curricular dynamic forgery augmentations towards general deepfake detection. In European Conference on Computer Vision, pages 104-122. Springer, 2024. 3", + "[35] Dat NGUYEN, Nesryne MEJRI, Inder Pal SINGH, Polina KULESHOVA, Marcella ASTRID, Anis KACEM, Enjie GHORBEL, and Djamila AOUADA. Laa-net: Localized artifact attention network for quality-agnostic and generalizable deepfake detection. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17395–17405, 2024. 1, 2", + "[36] Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards universal fake image detectors that generalize across generative models. In Proceedings of the IEEE/CVF Conference" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Computer Vision and Pattern Recognition, pages 24480-24489, 2023. 2", + "[37] Junting Pan, Ziyi Lin, Xiatian Zhu, Jing Shao, and Hongsheng Li. St-adapter: Parameter-efficient image-to-video transfer learning. Advances in Neural Information Processing Systems, 35:26462-26477, 2022. 3, 6", + "[38] Guilin Pang, Baopeng Zhang, Zhu Teng, Zige Qi, and Jianping Fan. Mre-net: Multi-rate excitation network for deepfake video detection. IEEE Transactions on Circuits and Systems for Video Technology, 33(8):3663-3676, 2023. 2", + "[39] Yuyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. Thinking in frequency: Face forgery detection by mining frequency-aware clues. In European conference on computer vision, pages 86-103. Springer, 2020. 2", + "[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 2, 3, 6", + "[41] Md Shohel Rana, Mohammad Nur Nobi, Bedduhu Murali, and Andrew H Sung. Deepfake detection: A systematic literature review. IEEE access, 10:25494-25513, 2022. 3, 4, 6, 7", + "[42] Meitar Ronen, Shahaf E. Finder, and Oren Freifeld. Deep-dpm: Deep clustering with an unknown number of clusters. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9861-9870, 2022. 4", + "[43] Andreas Rössler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Niessner. Faceforensics++: Learning to detect manipulated facial images. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1-11, 2019. 3, 6", + "[44] Kaede Shiohara and Toshihiko Yamasaki. Detecting deepfakes with self-blended images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18720-18729, 2022. 1, 2, 3", + "[45] Ke Sun, Taiping Yao, Shen Chen, Shouhong Ding, Jilin Li, and Rongrong Ji. Dual contrastive learning for general face forgery detection. In Proceedings of the AAAI conference on artificial intelligence, pages 2316-2324, 2022. 2, 6, 7", + "[46] Justus Thies, Michael Zollhofer, Marc Stamminger, Christian Theobalt, and Matthias Nießner. Face2face: Real-time face capture and reenactment of rgb videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2387-2395, 2016. 1, 3, 4, 6", + "[47] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. Acm Transactions on Graphics (TOG), 38(4):1-12, 2019. 1, 3, 4, 6", + "[48] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9 (86):2579-2605, 2008. 8", + "[49] Junke Wang, Zuxuan Wu, Wenhao Ouyang, Xintong Han, Jingjing Chen, Yu-Gang Jiang, and Ser-Nam Li. M2tr: Multi-modal multi-scale transformers for deepfake detec" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13391", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tion. In Proceedings of the 2022 international conference on multimedia retrieval, pages 615-623, 2022. 2", + "[50] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnn-generated images are surprisingly easy to spot... for now. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8695-8704, 2020. 3", + "[51] Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, and Houqiang Li. Altfreezing for more general video face forgery detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4129-4138, 2023. 2, 3, 6, 7", + "[52] Yuting Xu, Jian Liang, Gengyun Jia, Ziming Yang, Yanhao Zhang, and Ran He. Tall: Thumbnail layout for deepfake video detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22658-22668, 2023. 6, 7", + "[53] Zhiyuan Yan, Yong Zhang, Yanbo Fan, and Baoyuan Wu. Ucf: Uncovering common features for generalizable deepfake detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22412-22423, 2023. 6, 7", + "[54] Zhiyuan Yan, Yuhao Luo, Siwei Lyu, Qingshan Liu, and Baoyuan Wu. Transcending forgery specificity with latent space augmentation for generalizable deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8984-8994, 2024. 1, 2, 6, 7", + "[55] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications. ACM Computing Surveys, 56(4): 1-39, 2023. 1", + "[56] Wenyuan Yang, Xiaoyu Zhou, Zhikai Chen, Bofei Guo, Zhongjie Ba, Zhihua Xia, Xiaochun Cao, and Kui Ren. Avoid-df: Audio-visual joint learning for detecting deepfake. IEEE Transactions on Information Forensics and Security, 18:2015–2029, 2023. 2", + "[57] Shenghai Yuan, Jina Huang, Xianyi He, Yunyuan Ge, Yu-jun Shi, Liuhan Chen, Jiebo Luo, and Li Yuan. Identity-preserving text-to-video generation by frequency decomposition, 2024. 1", + "[58] Yu Zeng, Zhe L. Lin, and Vishal M. Patel. Sketched: Mask-free local image manipulation with partial sketches. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5941-5951, 2021. 1", + "[59] Daichi Zhang, Zihao Xiao, Shikun Li, Fanzhao Lin, Jianmin Li, and Shiming Ge. Learning natural consistency representation for face forgery video detection. In European Conference on Computer Vision, pages 407-424. Springer, 2024. 6", + "[60] Tianchen Zhao, Xiang Xu, Mingze Xu, Hui Ding, Yuanjun Xiong, and Wei Xia. Learning self-consistency for deepfake detection. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 15003-15013, 2021. 6", + "[61] Yinglin Zheng, Jianmin Bao, Dong Chen, Ming Zeng, and Fang Wen. Exploring temporal coherence for more general video face forgery detection. In Proceedings of the" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF international conference on computer vision, pages 15044-15054, 2021. 2, 3, 6", + "[62] Yang Zhou, Zichong Chen, and Hui Huang. Deformable oneshot face stylization via dino semantic guidance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7787-7796, 2024. 1", + "[63] Wanyi Zhuang, Qi Chu, Zhentao Tan, Qiankun Liu, Haojie Yuan, Changtao Miao, Zixiang Luo, and Nenghai Yu. Uiavit: Unsupervised inconsistency-aware method based on vision transformer for face forgery detection. In European conference on computer vision, pages 391-407. Springer, 2022. 6" + ], + "bbox": [ + 516, + 92, + 903, + 258 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "13392", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/FakeRadar_ Probing Forgery Outliers to Detect Unknown Deepfake Videos/28a1e3bc-e2b9-40da-b7fe-fcc0d6369340_model.json b/2025/FakeRadar_ Probing Forgery Outliers to Detect Unknown Deepfake Videos/28a1e3bc-e2b9-40da-b7fe-fcc0d6369340_model.json new file mode 100644 index 0000000000000000000000000000000000000000..348ad2fced327d604b7212e2bfa82143aaac2300 --- /dev/null +++ b/2025/FakeRadar_ Probing Forgery Outliers to Detect Unknown Deepfake Videos/28a1e3bc-e2b9-40da-b7fe-fcc0d6369340_model.json @@ -0,0 +1,2235 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.12, + 0.131, + 0.88, + 0.154 + ], + "angle": 0, + "content": "FakeRadar: Probing Forgery Outliers to Detect Unknown Deepfake Videos" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.18, + 0.762, + 0.216 + ], + "angle": 0, + "content": "Zhaolun Li\\(^{1}\\) Jichang Li\\(^{2*}\\) Yinqi Cai\\(^{3}\\) Junye Chen\\(^{3}\\) Xiaonan Luo\\(^{1}\\) Guanbin Li\\(^{3,2,5}\\) Rushi Lan\\(^{1,4*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.216, + 0.882, + 0.249 + ], + "angle": 0, + "content": "1Guilin University of Electronic Technology 2Pengcheng Laboratory 3Sun Yat-sen University 4Guangxi Key Laboratory of Image and Graphic Intelligent Processing" + }, + { + "type": "text", + "bbox": [ + 0.234, + 0.251, + 0.76, + 0.27 + ], + "angle": 0, + "content": "\\(^{5}\\)Guangdong Key Laboratory of Big Data Analysis and Processing" + }, + { + "type": "text", + "bbox": [ + 0.195, + 0.273, + 0.798, + 0.287 + ], + "angle": 0, + "content": "zhaolunli@mails.guet.edu.cn, li.jichang@pcl.ac.cn, rslan2016@163.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.322, + 0.327, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.354, + 0.485, + 0.732 + ], + "angle": 0, + "content": "In this paper, we propose FakeRadar, a novel deepfake video detection framework designed to address the challenges of cross-domain generalization in real-world scenarios. Existing detection methods typically rely on manipulation-specific cues, performing well on known forgery types but exhibiting severe limitations against emerging manipulation techniques. This poor generalization stems from their inability to adapt effectively to unseen forgery patterns. To overcome this, we leverage large-scale pretrained models (e.g. CLIP) to proactively probe the feature space, explicitly highlighting distributional gaps between real videos, known forgeries, and unseen manipulations. Specifically, FakeRadar introduces Forgery Outlier Probing, which employs dynamic subcluster modeling and cluster-conditional outlier generation to synthesize outlier samples near boundaries of estimated subclusters, simulating novel forgery artifacts beyond known manipulation types. Additionally, we design Outlier-Guided Tri-Training, which optimizes the detector to distinguish real, fake, and outlier samples using proposed outlier-driven contrastive learning and outlier-conditioned cross-entropy losses. Experiments show that FakeRadar outperforms existing methods across various benchmark datasets for deepfake video detection, particularly in cross-domain evaluations, by handling the variety of emerging manipulation techniques." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.761, + 0.222, + 0.777 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.786, + 0.484, + 0.879 + ], + "angle": 0, + "content": "Recently, generative techniques [28, 58, 62] for creating deepfake videos have advanced rapidly, producing highly realistic facial forgeries that challenge human perceptual limits [8, 23]. However, the malicious use of such technology, such as fabricated news and identity fraud, poses a significant threat to digital media credibility [7, 9]. Ear" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.32, + 0.907, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.534, + 0.907, + 0.659 + ], + "angle": 0, + "content": "Figure 1. Conceptual pipeline of FakeRadar. (a)-(b) Forgery Outlier Probing first performs dynamic subcluster modeling on these features to partition finer-grained subclusters, then generates cluster-conditional outliers for each estimated subcluster to simulate unseen forgeries, emulating unknown manipulations. (c)-(d) The model is optimized using outlier-driven tri-training, allowing the detector to establish optimal boundaries for the real, fake, and outlier classes, with the latter two merged into a new \"Fake\" class during testing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.909, + 0.903 + ], + "angle": 0, + "content": "lier deepfake detection methods primarily rely on the passive learning of specific forgery artifacts, such as boundary inconsistencies [27, 54], blinking anomalies [32], and texture irregularities [17, 35]. While these approaches perform well against conventional deepfake manipulations, including Face2Face [46] and NeuralTextures [47], they exhibit critical vulnerabilities when confronted with emerging generative paradigms, such as diffusion models [55]. For instance, boundary-based detection methods [30, 44] fail to identify faces generated by advanced GANs (e.g., StyleGAN [25]). Similarly, blink analysis-based techniques [32] are rendered ineffective when processing temporally smoothed fake videos [57]. These limitations stem from a fundamental shortcoming: existing methods learn a constrained decision boundary in the feature space, where" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.245, + 0.9 + ], + "angle": 0, + "content": "*Corresponding Authors." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13382" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "known forgery traces occupy only a localized region, while novel manipulation techniques introduce artifacts in arbitrary locations. As a result, they struggle to generalize to new manipulation techniques that introduce artifacts in previously unseen ways, leading to significant performance degradation in cross-domain scenarios [11, 12]." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.186, + 0.485, + 0.562 + ], + "angle": 0, + "content": "To better understand its underlying cause, [2, 36] have analyzed the feature distributions of real and synthetic data. These studies reveal a fundamental embedding-space disparity: real data exhibit compact, dense distributions, whereas deepfake forgeries form scattered, sparse clusters due to their generation with diverse manipulation types, as demonstrated in Figure 1. This difference in distribution poses a significant challenge for existing detection methods, which are typically optimized to classify previously observed forgeries. Since deepfake forgeries do not adhere to a fixed pattern but rather introduce an infinite variety of modifications, classifiers trained on known forgery patterns struggle to recognize novel deepfakes that deviate from past observations. Therefore, a robust detection approach must move beyond merely learning pre-defined forgery patterns and instead develop the capability to identify anomalies that fall outside of previously known distributions. Given the limitations of existing methods in handling unseen forgery types, we propose a paradigm shift in deepfake detection. Instead of passively fitting to known forgery patterns, our approach proactively explores potential anomaly regions in the feature space. Inspired by radar systems that scan for unknown targets through frequency-spectrum probing, our method enhances detection robustness against novel deep-fake manipulations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.566, + 0.483, + 0.853 + ], + "angle": 0, + "content": "In this work, we propose a new deepfake video detection framework, called FakeRadar, which leverages the deep feature priors of large-scale pre-trained models (i.e. CLIP [40]) to proactively \"probe\" unknown forgeries and enhance cross-domain generalization. Unlike conventional detectors [2, 5, 51] that perform only binary classification, FakeRadar explicitly introduces \"Forgery outliers\" during training to simulate potential unknown forgeries and thus \"pre-explore and cover\" the latent space. As shown in Figure 1 (a) and (b), we first employ Forgery Outlier Probing, which conducts dynamic subcluster modeling on the features of real and known deepfake videos. By approximating the embedding space via fine-grained Gaussian mixture distributions, this algorithm automatically partitions training samples into distinct subclusters. Next, through cluster-conditional outlier generation, Forgery Outlier Probing synthesizes outlier samples near the boundaries of these estimated subclusters with a certain \"novel offset\", thereby covering a broader range of potential forgery regions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Additionally, FakeRadar adopts a model optimization strategy called Outlier-Guided Tri-Training, enabling the model to discriminate among \"Real\", \"Fake\", and \"Out" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.302 + ], + "angle": 0, + "content": "lier” samples, with its details as illustrated in Figure 1 (c) and (d). Specifically, Outlier-Guided Tri-Training proposes an outlier-driven contrastive loss and an outlier-conditioned cross-entropy loss for end-to-end model optimization. The former loss enhances the distance between real, fake, and outlier features via contrastive learning, emphasizing their differences, while the latter one ensures that the model has clear decision boundaries for the three categories, especially ensuring that outliers are not incorrectly classified as real. In this triplet-class (Real/Fake/Outlier) setting, the model can independently label unknown forgeries during training. During inference, both “Fake” and “Outlier” are classified as forgeries, empowering the detector to adapt effectively to unknown deepfakes." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.303, + 0.905, + 0.333 + ], + "angle": 0, + "content": "We summarize FakeRadar's principal contributions as follows, with its conceptual pipeline illustrated in Figure 1:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.334, + 0.905, + 0.378 + ], + "angle": 0, + "content": "- We propose FakeRadar, a novel framework for deepfake video detection, designed to address the critical challenge of cross-domain generalization in real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.379, + 0.905, + 0.438 + ], + "angle": 0, + "content": "- We introduce Forgery Outlier Probing for synthesizing outlier samples to cover a broader range of potential forger patterns, expanding the detector's understanding of forgery distribution." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.44, + 0.904, + 0.498 + ], + "angle": 0, + "content": "- We design Outlier-Guided Tri-Training for model optimization, which enables end-to-end training for distinguishing real, fake, and outlier samples, improving detection of unknown forgeries." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.5, + 0.905, + 0.56 + ], + "angle": 0, + "content": "- Numerous experiments have revealed that FakeRadar outperforms current leading algorithms across multiple deepfake benchmarks, particularly excelling in cross-domain evaluations where existing methods struggle." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.334, + 0.905, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.574, + 0.652, + 0.589 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Deepfake Detection. In response to the malicious use of deepfake techniques, numerous approaches have been explored to address this issue. Generally, deepfake detection is formulated as a binary classification problem (real or fake), leading researchers to develop a variety of end-to-end detectors [45, 54] that directly distinguish reals from manipulated content. Existing detection methods can be broadly categorized into two groups: image-level and video-level detection. Specifically, image-level detectors primarily focus on identifying spatial artifacts caused by forgeries. Examples include frequency anomalies [3, 16, 29, 39], identity-based discrepancies [13, 14, 24], blending artifacts [30, 44], and localized distortions [35, 49]. In contrast, video-level detectors employ a wider range of strategies. For example, some algorithms, such as [20, 32], concentrate on temporal consistency, typically reflected in correlations across consecutive frames. Other video detection studies [18, 38, 61] incorporate both temporal and spatiotemporal features to establish a more comprehensive detection framework. Another line of research [22, 56] examines audio-visual con" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "13383" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.088, + 0.905, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.285, + 0.907, + 0.397 + ], + "angle": 0, + "content": "Figure 2. An overview of FakeRadar for deepfake video detection. FakeRadar consists of two key components: Forgery Outlier Probing and Outlier-Guided Tri-Training. (i) Forgery Outlier Probing performs dynamic subcluster modeling to model the feature-space distributions of real and fake samples using GMMs, followed by dynamic subcluster splitting and merging to produce finer-grained subcluster partitions. Afterwards, FakeRadar conducts cluster-conditioned outlier generation to simulate unseen forgeries for novel manipulation techniques. (ii) Outlier-Guided Tri-Training jointly optimizes the CLIP model's ViT-B/16 backbone, inserted with ST-Adapter [37] for parameter-efficient fine tuning, and a triplet-class classifier to distinguish between \"Real\", \"Fake\", and \"Outlier\" classes, with \"Fake\" and \"Outlier\" merged into a single new \"Fake\" class during evaluation. The training set includes real videos, as well as deepfake videos generated using four different manipulations, i.e. \"DF\" [41], \"FS\" [26], \"NT\" [47] and \"F2F\" [46]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.41, + 0.484, + 0.515 + ], + "angle": 0, + "content": "sistency by analyzing the synchronization and correlation between video and audio signals to detect forgeries. In summary, these methods rely heavily on existing artifacts in training set, limiting their generalization to unseen manipulation types. In contrast, our approach explores feature-space anomalies, enhancing the generalization capacity of our proposed detector to diverse deepfake manipulations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.483, + 0.779 + ], + "angle": 0, + "content": "Towards Generalization in Deepfake Detection. Rapid advance in deepfakes have significantly improved the quality and authenticity of generated content, posing substantial challenges to the generalization capability of detection systems. Existing detectors [6, 43, 50] often suffer from severe performance degradation when encountering novel forgery techniques or previously unseen manipulations. To address this issue, researchers have proposed various strategies to enhance detector generalization. For instance, FTCN [61] improves generalization by capturing temporal coherence features in video sequences. Similarly, AltFreezing [51] employs an alternating weight-freezing training strategy to facilitate the joint detection of spatial and temporal forgery artifacts. Another algorithm, StyleFlow [5], leverages the temporal dynamics of style latent vectors to identify anomalous patterns in generated videos." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.902 + ], + "angle": 0, + "content": "More recently, data synthesis has emerged as an effective strategy for improving detector generalization. For example, FWA [31] enhances training data by introducing facial warping artifacts, while Face X-ray [30] detects forgeries by identifying blending boundaries, independent of specific facial manipulation techniques. Building on this concept, SBI [44] synthesizes training samples by blending different views of the same pristine image to simulate generic forgery" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.907, + 0.502 + ], + "angle": 0, + "content": "artifacts. Additionally, a series of recent studies [1, 27, 34] have further advanced this direction by employing various data synthesis techniques. Unlike existing methods that primarily focus on image- or video-level data synthesis for deepfake detection, our approach proposes to explore unseen forgeries in feature space." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.515, + 0.649, + 0.533 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.54, + 0.655, + 0.555 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.563, + 0.907, + 0.761 + ], + "angle": 0, + "content": "Overview. In this paper, we introduce FakeRadar, to overcome the challenges of cross-domain generalization in deepfake video detection by exploring the feature space to identify emerging manipulation patterns. Specifically, FakeRadar is built upon two key components: (i) Forgery Outlier Probing, which models the feature-space distributions of subclusters for training samples of both real and fake categories, and then generates cluster-conditioned outlier samples to simulate unseen forgeries, and (ii) Outlier-Guided Tri-Training, which jointly optimizes an adapter-inserted backbone and a triplet-class classifier to distinguish among samples from the \"Real\", \"Fake\", and \"Outlier\" classes. An overview of FakeRadar can be found in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Model Adaptation with a Frozen Backbone. The FakeRadar framework employs a pre-trained Vision Transformer (ViT-B) from CLIP [40] as a frozen backbone to preserve its learned semantic features. This model serves as the video encoder for input clips. To adapt it for deepfake detection, we integrate lightweight ST-Adapter layers [37] after each Transformer block, enabling parameter-efficient fine-tuning. This significantly enhances the model's ability to" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13384" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.152 + ], + "angle": 0, + "content": "capture spatial manipulations and temporal inconsistencies, which are crucial for detecting deepfake videos. Given a feature vector \\( x \\) from the video encoder, each adapter transforms it as:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.165, + 0.483, + 0.183 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S T - A d p a t e r} (x) = x + \\operatorname {R e L U} (\\operatorname {C o n v 3 D} (x W _ {\\text {d o w n}})) W _ {\\text {u p}}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.194, + 0.484, + 0.346 + ], + "angle": 0, + "content": "Here, \\(\\mathrm{Conv3D}(\\cdot)\\) denotes a 3D convolution operation employed for capturing spatio-temporal features from videos. As well, \\(W_{\\mathrm{down}}\\) and \\(W_{\\mathrm{up}}\\) are weight matrices that perform dimensionality reduction and expansion, respectively, and \\(\\mathrm{ReLU}(\\cdot)\\) is the activation function. By introducing only a small number of additional parameters, these adapters allow efficient fine-tuning while retaining the original backbone's representational strength. After feature extraction, each video clip is represented as a \\(d\\)-dimensional vector for subsequent analysis, with \\(d = 768\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.365, + 0.484, + 0.518 + ], + "angle": 0, + "content": "Inference. During inference, the model classifies input samples as either real or deepfake, with deepfakes categorized as \"Fake\" or \"Outlier\". The \"Fake\" class includes fake samples with known manipulation patterns, while the \"Outlier\" class detects testing samples with novel forgeries that differ from established manipulations. By unifying these categories, the model enhances its robustness against evolving manipulation techniques, enabling accurate detection of both familiar and emerging deepfakes. This improves its generalization performance in cross-domain evaluations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.526, + 0.317, + 0.543 + ], + "angle": 0, + "content": "3.2. Forgery Outlier Probing" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.549, + 0.484, + 0.746 + ], + "angle": 0, + "content": "In this section, we introduce Forgery Outlier Probing (FOP) to generate outlier samples with unseen forgeries beyond the existing real and fake videos, simulating previously unobserved manipulation patterns. Specifically, FOP first presents dynamic subcluster modeling, which performs category-specific clustering on the training samples of real and fake categories. Then, it models the distributions of these categories using a strategy based on Gaussian Mixture Models (GMMs), and follows by applying dynamic subcluster splitting and merging, thereby achieving finer-grained subcluster partitioning. Finally, we generate cluster-conditioned outliers based on the resulting subclusters to stimulate unseen forgeries." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Dynamic Subcluster Modeling. To capture subtle intra-class variations within real and deepfake videos, we apply category-specific clustering on the FaceForensics++ (FF++) dataset. Specifically, we treat the real class and each of the four fake classes (i.e. four types of deepfake videos created by different manipulation techniques [26, 41, 46, 47]) as five separate categories. For each category \\( c \\), let \\( \\mathcal{X}^{(c)} = \\{\\mathbf{x}_1, \\mathbf{x}_2, \\dots, \\mathbf{x}_N\\} \\) denote the set of feature embeddings (i.e. extracted from the video encoder). Inspired by [42], we then" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.738, + 0.107 + ], + "angle": 0, + "content": "define a main clustering network:" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.115, + 0.907, + 0.136 + ], + "angle": 0, + "content": "\\[\nf _ {\\operatorname {m a i n}} ^ {(c)} \\left(\\mathbf {x} _ {i}\\right) \\mapsto \\mathbf {r} _ {i} = \\left(r _ {i, 1}, r _ {i, 2}, \\dots , r _ {i, K}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.143, + 0.907, + 0.207 + ], + "angle": 0, + "content": "which initially assumes \\(K\\) clusters, empirically with \\(K = 5\\), providing a \"soft\" assignment of each embedding \\(\\mathbf{x}_i \\in \\mathcal{X}^{(c)}\\) to these \\(K\\) clusters. Here, \\(r_{i,k} \\in [0,1]\\) and \\(\\sum_{k=1}^{K} r_{i,k} = 1\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.207, + 0.907, + 0.372 + ], + "angle": 0, + "content": "Each cluster can be modeled as a Gaussian distribution using a GMM, allowing us to represent the feature distribution of each category in a probabilistic manner. Specifically, for each cluster \\(k\\) in category \\(c\\), we define a Gaussian component characterized by \\(\\{\\pi_k,\\pmb {\\mu}_k,\\pmb {\\Sigma}_k\\}\\), where \\(\\pi_{k}\\in (0,1)\\) is the mixture weight for cluster \\(k\\), with \\(\\sum_{k = 1}^{K}\\pi_{k} = 1\\), while \\(\\pmb{\\mu}_{k}\\) and \\(\\pmb{\\Sigma}_{k}\\) represent its corresponding mean and covariance matrix. These parameters enable the clustering network to estimate the likelihood of each data sample belonging to different clusters, aligning our model with a GMM-based formulation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.374, + 0.906, + 0.419 + ], + "angle": 0, + "content": "After that, the main clustering network needs to learn these parameters by aligning its soft assignments with GMM responsibilities:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.427, + 0.907, + 0.468 + ], + "angle": 0, + "content": "\\[\nr _ {i, k} ^ {E} = \\frac {\\pi_ {k} \\mathcal {N} \\left(\\mathbf {x} _ {i} ; \\boldsymbol {\\mu} _ {k} , \\boldsymbol {\\Sigma} _ {k}\\right)}{\\sum_ {k ^ {\\prime} = 1} ^ {K} \\pi_ {k ^ {\\prime}} \\mathcal {N} \\left(\\mathbf {x} _ {i} ; \\boldsymbol {\\mu} _ {k ^ {\\prime}} , \\boldsymbol {\\Sigma} _ {k ^ {\\prime}}\\right)}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.473, + 0.906, + 0.516 + ], + "angle": 0, + "content": "where \\(\\mathcal{N}(\\cdot ;\\pmb {\\mu}_k,\\pmb {\\Sigma}_k)\\) is the multivariate Gaussian density function. The network is trained using the KL-divergence loss:" + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.516, + 0.906, + 0.556 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m a i n}} = \\sum_ {i = 1} ^ {N} \\mathrm {K L} \\left(\\mathbf {r} _ {i} \\| \\mathbf {r} _ {i} ^ {E}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.56, + 0.906, + 0.589 + ], + "angle": 0, + "content": "which ensures that the learned cluster assignments match the inferred GMM responsibilities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.59, + 0.906, + 0.68 + ], + "angle": 0, + "content": "To further refine cluster structures, we introduce a subclustering network for each cluster \\( k \\), which attempts to split the cluster into two subclusters if multi-modal structures are detected. A compactness loss is applied to encourage subclusters to be well-separated. This loss function penalizes subclusters that are too close to each other:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.689, + 0.906, + 0.733 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {s u b}} = \\sum_ {k = 1} ^ {K} \\sum_ {j = 1} ^ {2} \\sum_ {\\mathbf {x} _ {i} \\in X _ {k}} \\tilde {r} _ {i, j} \\| \\mathbf {x} _ {i} - \\tilde {\\boldsymbol {\\mu}} _ {k, j} \\| _ {2} ^ {2}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.739, + 0.906, + 0.769 + ], + "angle": 0, + "content": "where \\(\\tilde{r}_{i,j}\\) represents the probability of \\(\\mathbf{x}_i\\) belonging to subcluster \\(j\\), and \\(\\tilde{\\mu}_{k,j}\\) is its mean." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.77, + 0.906, + 0.861 + ], + "angle": 0, + "content": "Since the predefined cluster number \\(K\\) may not be optimal, we propose to dynamically adjust it through split/merge proposals, inspired by [42]. Suppose a cluster \\(k\\) contains \\(N_{k}\\) points, \\(X_{k}\\). We split if subclustering indicates that \\(X_{k}\\) is better viewed as two subgroups, \\(\\{X_{k,1}, X_{k,2}\\}\\). The Hastings ratio is given by" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.868, + 0.906, + 0.906 + ], + "angle": 0, + "content": "\\[\nH _ {s} = \\frac {\\alpha \\Gamma (N _ {k , 1}) f _ {x} (X _ {k , 1}) \\Gamma (N _ {k , 2}) f _ {x} (X _ {k , 2})}{\\Gamma (N _ {k}) f _ {x} (X _ {k})}, \\tag {6}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13385" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.136 + ], + "angle": 0, + "content": "where \\( f_{x}(\\cdot) \\) is the marginal likelihood under a weak NIw prior, \\( \\alpha \\) is the Dirichlet-process concentration parameter, and \\( \\Gamma (\\cdot) \\) is the Gamma function." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.137, + 0.482, + 0.242 + ], + "angle": 0, + "content": "If \\(H_{s}\\) exceeds a threshold, cluster \\(k\\) is split into two new clusters. The parameters for these new clusters are initialized from \\(\\tilde{\\mu}_{k,1},\\tilde{\\Sigma}_{k,1}\\) and \\(\\tilde{\\mu}_{k,2},\\tilde{\\Sigma}_{k,2}\\), respectively. Merging is similarly proposed if two clusters significantly overlap, with acceptance ratio \\(H_{m} = 1 / H_{s}\\). Repeating these steps yields an effective number of clusters \\(K^{(c)}\\), plus subclusters for each category \\(c\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.243, + 0.482, + 0.303 + ], + "angle": 0, + "content": "After these dynamic adjustments, we obtain the final subcluster set \\(\\hat{\\mathcal{C}} = \\{\\hat{C}_1,\\hat{C}_2,\\dots,\\hat{C}_{K^{(c)}}\\}\\), where \\(K^{(c)}\\) is the refined number of clusters. Each adjusted subcluster \\(\\hat{C}_k\\) is characterized by its updated mean and covariance:" + }, + { + "type": "equation", + "bbox": [ + 0.216, + 0.313, + 0.483, + 0.352 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {\\mu}} _ {k} = \\frac {1}{| \\hat {C} _ {k} |} \\sum_ {\\mathbf {x} _ {i} \\in \\hat {C} _ {k}} \\mathbf {x} _ {i}, \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.363, + 0.483, + 0.402 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {\\Sigma}} = \\frac {1}{| \\hat {C} _ {k} |} \\sum_ {\\mathbf {x} _ {i} \\in \\hat {C} _ {k}} \\left(\\mathbf {x} _ {i} - \\hat {\\boldsymbol {\\mu}} _ {k}\\right) \\left(\\mathbf {x} _ {i} - \\hat {\\boldsymbol {\\mu}} _ {k}\\right) ^ {\\top}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.407, + 0.484, + 0.498 + ], + "angle": 0, + "content": "We use online estimation for efficient training by maintaining a cluster-conditional queue with \\( |\\hat{C}_k| \\) sample instances from each subcluster. In each iteration, we enqueue the embeddings of objects to their corresponding cluster-conditional queues and dequeue the same number of object embeddings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.516, + 0.484, + 0.651 + ], + "angle": 0, + "content": "Cluster-Conditional Outlier Generation. After deriving finer-grained cluster partitions for each category, we generate feature-space outliers to probe potential unknown forgeries. Motivated by [15], in order to ensure that the sampled outliers lie near the cluster boundary, we maintain a sufficiently small value \\(\\epsilon\\) for outlier sampling from the \\(\\epsilon\\)-likelihood region of the estimated cluster-conditional distribution. Therefore, for a subcluster from \\(\\hat{C}\\) indexed by \\(k\\), the outlier samples it generates can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.662, + 0.488, + 0.718 + ], + "angle": 0, + "content": "\\[\n\\mathcal {V} _ {k} = \\left\\{\\mathbf {v} _ {k} \\mid \\frac {\\exp \\left(- \\frac {1}{2} \\left(\\mathbf {v} _ {k} - \\hat {\\boldsymbol {\\mu}} _ {k}\\right) ^ {\\top} \\hat {\\boldsymbol {\\Sigma}} ^ {- 1} \\left(\\mathbf {v} _ {k} - \\hat {\\boldsymbol {\\mu}} _ {k}\\right)\\right)}{(2 \\pi) ^ {d / 2} | \\hat {\\boldsymbol {\\Sigma}} | ^ {1 / 2}} < \\epsilon \\right\\}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.484, + 0.824 + ], + "angle": 0, + "content": "where \\(v_{k}\\sim \\mathcal{N}(\\hat{\\mu}_{k},\\hat{\\Sigma}_{k})\\) represents the virtual outliers sampled from the estimated Gaussian distribution of subcluster \\(k\\), with \\(d\\) denoting its embedding dimensionality. It is worth noting that during model optimization, all outliers generated here, i.e. \\(\\overline{\\mathcal{V}} = \\mathcal{V}_1\\cup \\dots \\cup \\mathcal{V}_k\\cup \\dots \\mathcal{V}_{|\\hat{\\mathcal{O}} |}\\) are classified into the \"Outlier\" class, separate from the real and fake classes." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.834, + 0.347, + 0.85 + ], + "angle": 0, + "content": "3.3. Outlier-Guided Tri-Training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To enable better performance of FakeRadar for tasks of deepfake video detection, we design a strategy called Outlier-Guided Tri-Training for model optimization jointly" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.213 + ], + "angle": 0, + "content": "combining proposed outlier-driven contrastive loss and outlier-conditioned cross-entropy loss. With this approach, a triplet-class classifier (integrated with the pre-trained backbone and adapters) is trained to explicitly distinguish among three categories: \"Real\", \"Fake\", and \"Outlier\". This design not only improves the model's ability to detect known forgeries but also sensitizes it to emerging manipulation patterns." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.232, + 0.905, + 0.382 + ], + "angle": 0, + "content": "Outlier-Driven Contrastive Loss. In this work, we propose an outlier-driven contrastive loss to enhance the model's ability to delineate class boundaries within the feature space, thereby improving feature discriminability. This loss function is designed to maximize the similarity between samples and their respective subcluster centers while minimizing their similarity to other subcluster centers. Additionally, it treats generated outliers as negative examples, enforcing low similarity between samples and outliers to strengthen boundary discrimination." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.383, + 0.905, + 0.414 + ], + "angle": 0, + "content": "To this end, the loss is formulated using an InfoNCE-like contrastive loss given an input sample \\( i \\):" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.425, + 0.918, + 0.492 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o n}} = - \\log \\frac {\\exp \\left(\\frac {s \\left(h _ {i} , \\mu_ {k _ {i} ^ {+}}\\right)}{\\tau}\\right)}{\\sum_ {k ^ {-}} \\exp \\left(\\frac {s \\left(h _ {i} , \\mu_ {k ^ {-}}\\right)}{\\tau}\\right) + \\sum_ {v _ {j} \\in \\overline {{\\nu}}} \\exp \\left(\\frac {s \\left(h _ {i} , v _ {j}\\right)}{\\tau}\\right)}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.906, + 0.645 + ], + "angle": 0, + "content": "where \\( h_i \\) denotes the feature representation of sample \\( i \\), while \\( \\mathcal{V} \\) represents the outlier cache, providing additional hard negative samples to improve the model's ability to distinguish boundary samples. Additionally, \\( k^{-} \\) represents all subclusters except the one to which sample \\( i \\) belongs, with its corresponding \\( u_{k^{-}} \\) denoting the centers of these subclusters, used as negative samples. The function \\( s(\\cdot ,\\cdot) \\) denotes cosine similarity, which quantifies the similarity between samples, and \\( \\tau \\) is a temperature hyperparameter that controls the smoothness of similarity scaling." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.664, + 0.906, + 0.829 + ], + "angle": 0, + "content": "Outlier-Conditioned Cross-Entropy Loss. While the proposed outlier-driven contrastive loss enforces stronger feature-space separation, we still require a concrete classification objective to handle a three-category task: real, fake, and outlier. In particular, the model must not only determine whether a video is fake or real but also assign a separate label for those samples situated near decision boundaries (i.e., outliers). This triplet-class cross-entropy loss is crucial, as it enables the model to explicitly differentiate uncertain boundary samples and label them as outliers, thereby avoiding confusion." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.83, + 0.905, + 0.859 + ], + "angle": 0, + "content": "Concretely, the triplet-class cross-entropy loss can be defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.87, + 0.905, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c l s}} = - \\sum_ {c \\in \\{\\text {R e a l}, \\text {F a k e}, \\text {O u t l i e r} \\}} y _ {c} \\log p _ {c}, \\tag {11}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13386" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.237, + 0.089, + 0.761, + 0.299 + ], + "angle": 0, + "content": "
MethodBackboneInput TypeTesting Set AUC (%)
FF++CDFv2DFDCPDFDCDFD
Xception [6]XceptionFrame99.373.7-70.9-
Face X-Ray [30]HRNetFrame97.879.5-65.595.4
PCL+i2G [60]ResnetFrame99.890.074.467.5-
SLADD [4]XceptionFrame98.479.776.0--
UIA-ViT [63]Vision TransformerFrame99.382.475.8-94.7
UCF [53]XceptionFrame-82.4-80.594.5
SeeABLE [27]EfficientNetFrame-87.386.375.9-
LSDA [54]EfficientNetFrame-91.181.277.095.6
FTCN [61]3D ResNetVideo99.886.9-74.0-
RealForensics [21]3D ResNetVideo99.886.9-75.9-
LTTD [19]Vision TransformerVideo99.589.3-80.4-
AltFreezing [51]3D ResNetVideo99.789.5--93.7
TALL [52]Swin TransformerVideo99.990.8-76.8-
StyleFlow [5]3D ResNetVideo99.189.0--96.1
NACO [59]Vision TransformerVideo99.889.5-76.7-
FakeRadar (Ours)Vision TransformerVideo99.191.788.584.196.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.303, + 0.907, + 0.334 + ], + "angle": 0, + "content": "Table 1. Generalization comparisons of FakeRadar and state-of-the-art algorithms on cross-dataset evaluation, using video-level AUC (%) as the metric." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.358, + 0.483, + 0.449 + ], + "angle": 0, + "content": "where \\( y_{c} \\) is the ground-truth label for samples belonging to class \\( c \\), and \\( p_{c} \\) is the model-predicted probability of samples from that class, with \\( c = \\) \"Outlier\" denoting the outlier label. This loss compels the model to accurately partition each category in the feature space, especially for samples near class boundaries." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.467, + 0.483, + 0.499 + ], + "angle": 0, + "content": "Total Loss. Finally, we combine both losses into the total objective for model optimization:" + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.509, + 0.482, + 0.526 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {c o n}} + \\lambda \\mathcal {L} _ {\\text {c l s}}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.536, + 0.484, + 0.599 + ], + "angle": 0, + "content": "where \\(\\lambda = 0.5\\) is a hyperparameter balancing the relative importance of classification and contrastive goals. In practice, each mini-batch comprises 16 training samples \\((\\mathrm{FF}++)\\), and 16 outlier samples during model training." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.61, + 0.224, + 0.627 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.634, + 0.275, + 0.651 + ], + "angle": 0, + "content": "4.1. Experiment Setups" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.657, + 0.483, + 0.853 + ], + "angle": 0, + "content": "Datasets. Following established experimental practices [5, 51, 52, 54], this study primarily utilizes the high-quality (HQ) version of the FaceForensics++ dataset [43], denoted as \\(\\mathrm{FF} + + (\\mathrm{HQ})\\), for FakeRadar's evaluation. This dataset consists of 1000 pristine videos and 4000 deepfakes videos generated using four different manipulation techniques: DFD [41], F2F [46], FS [26], and NT [47]. The dataset is split into training, validation, and testing sets according to the official protocol. This study primarily focuses on evaluating the cross-domain performance of our method. To this end, we incorporate four additional datasets for evaluation, including CDFv2 [33], DFDCP [11], DFDC [12], and DFD [41]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Implementation Details. To ensure fair comparisons, we maintain most implementation details consistent with exist" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.356, + 0.907, + 0.602 + ], + "angle": 0, + "content": "
Training SetMethodDFF2FFSNTavg.
DFDCL [45]99.9877.1361.0175.0178.28
UIA-ViT [63]99.2974.4453.8970.9274.64
TALL [52]99.3569.3354.3874.7474.45
AltFreezing [51]98.8069.3372.0777.4179.40
FakeRadar (Ours)99.9586.4788.1266.3682.23
F2FDCL [45]91.9199.2159.5866.6779.34
UIA-ViT [63]83.3998.3268.3767.1779.31
TALL [52]85.5298.6762.1570.2479.15
AltFreezing [51]81.1298.8062.9770.2678.29
FakeRadar (Ours)96.9499.0982.3368.4686.75
FSDCL [45]74.8069.7599.9052.6074.26
UIA-ViT [63]81.0266.3099.0449.2673.91
TALL [52]77.8367.3098.7752.5774.12
AltFreezing [51]80.6765.9499.7857.2375.90
FakeRadar (Ours)96.3982.0699.8352.6482.73
NTDCL [45]91.2352.1379.3198.9780.41
UIA-ViT [63]79.3767.8945.9494.5971.95
TALL [52]84.0272.8551.6695.1675.92
AltFreezing [51]88.8176.7857.9295.7179.80
FakeRadar (Ours)96.7183.3467.5098.4186.49
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.608, + 0.907, + 0.651 + ], + "angle": 0, + "content": "Table 2. Generalization comparisons of FakeRadar and state-of-the-art algorithms on cross-manipulation evaluation, using video-level AUC (\\%) as the metric." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.903 + ], + "angle": 0, + "content": "ing methods [5, 51, 59]. However, our FakeRadar framework employs ViT-B/16 from CLIP [40] as the backbone, into which we insert an ST-Adapter [37] with a 384-dimensional bottleneck for parameter-efficient fine-tuning. Additionally, the classification head of our triplet-class classifier consists of a single linear layer with softmax activation. During model training, we sample four temporal clips from each video, each containing 12 consecutive frames. We use Adam as the optimizer and apply a cosine learning rate scheduler with an initial learning rate of \\(1 \\times 10^{-4}\\). The model is trained for 60 epochs with a batch size of 16 and a weight decay of \\(5 \\times 10^{-4}\\). Prior to model optimization, we perform data preprocessing by detecting faces with RetinaFace [10] and then standardize all face images" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13387" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.087, + 0.357, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.087, + 0.631, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.087, + 0.903, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.238, + 0.907, + 0.321 + ], + "angle": 0, + "content": "Figure 3. Evaluation of (a) Forgery Outlier Probing and (b)/(c) Outlier-Guided Tri-Training by FakeRadar. (a) \"FakeRadar w/o Dyn. Sub. M.\" and \"FakeRadar w/o Out. Gen.\" indicate that dynamic subcluster modeling and cluster-conditional outlier generation have been removed. The model is trained on FF++(HQ) and evaluated on DFDC [12], CDFv2 [33], and DFD [41]. (b) and (c) show, respectively, the count of real and fake test samples classified as \"fake\" across predicted probability intervals. The blue solid line denotes the binary classifier's distribution (distinguishing only real vs. fake), while the red dashed line denotes that of our proposed triplet classifier, which treats both fake and outlier samples as the new \"fake\". In this evaluation, the model is trained on FF++(HQ) and tested on CDFv2 [33]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.347, + 0.218, + 0.362 + ], + "angle": 0, + "content": "to \\(224\\times 224\\) pixels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.381, + 0.484, + 0.473 + ], + "angle": 0, + "content": "Evaluation Metrics. In this work, consistent with previous studies [5, 51, 52] employing video-level inputs, we use the area under the ROC curve (AUC) as the performance metric during evaluations. For methods that use images as model inputs, e.g. [27, 54], their video-level results are derived by averaging the predictions across all video frames." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.482, + 0.431, + 0.499 + ], + "angle": 0, + "content": "4.2. Comparisons with the State-of-the-Arts" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.504, + 0.484, + 0.641 + ], + "angle": 0, + "content": "In this study, we focus on evaluating the capabilities of cross-domain generalization for our detector to validate its effectiveness in detecting deepfake videos manipulated through various techniques. To achieve this, we employ two strategies for evaluation: cross-dataset evaluation and cross-manipulation evaluation, both of which demonstrate that our results predominantly surpass existing state-of-the-art (SoTA) algorithms for deepfake video detection, as illustrated by Table 1 and Table 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Cross-Dataset Evaluation. In this situation, we train our FakeRadar on the FF++(HQ) and test its performance on multiple benchmark datasets, including CDFv2 [33], DFD [41], DFDC [12], and DFDCP [11]. As reported in Table 1, FakeRadar demonstrates its superior generalization capability in cross-dataset evaluation, achieving the best performance across all cross-dataset cases while maintaining comparable performance on the FF++ test dataset against SoTA algorithms. Specifically, our algorithm significantly improves over other methods with images and videos as inputs, particularly outperforming the current best methods, UCF [53] and LTTD [19], by \\(3.6\\%\\) and \\(3.7\\%\\) in the AUC metric, respectively, on the dataset DFDC [12]. These results robustly showcase the proposed method's enhanced ability of generalization to handle unseen manipulation techniques in cross-dataset evaluation scenarios." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.433 + ], + "angle": 0, + "content": "
M-(#)FOPOGTTCDFDFDCDFDCPDFDavg.
ODCLOCCE
1✓✓✓91.684.188.596.290.1
2✓✓89.981.285.694.987.9
3✓✓88.680.987.794.788.0
4✓88.880.286.794.587.6
5✓88.478.485.194.386.7
688.278.384.894.286.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.442, + 0.907, + 0.526 + ], + "angle": 0, + "content": "Table 3. Ablation study results of FakeRadar by removing its one or more proposed components, using video-level AUC (\\%) as the metric. For clarity, \"FOP\", \"OGTT\", \"ODCL\", and \"OCCE\" refer to Forgery Outlier Probing, Outlier-Guided Tri-Training, Outlier-Driven Contrastive Loss, and Outlier-Conditioned Cross-Entropy Loss, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.907, + 0.809 + ], + "angle": 0, + "content": "Cross-Manipulation Evaluation. To further assess our model's generalization across different manipulation techniques, we conduct cross-manipulation evaluation on the dataset FF++(HQ). In this case, we train the model using only real samples from FF++(HQ) and one manipulation type of deepfake samples, then test its performance on the remaining three manipulation types within FF++(HQ). As shown in Table 2, the proposed FakeRadar achieves the best average performance in all testing cases, particularly excelling in scenarios trained with F2F as the manipulation type, surpassing the second-best performing method, DCL [45], by \\(7.41\\%\\) in AUC. In evaluations focused on a single manipulation type, our method also achieves the best performance in most cases and the second-best in a few. This indicates that our approach, in cross-manipulation evaluation, demonstrates a performance advantage in cross-domain generalization compared to current SoTA algorithms of detecting deepfake videos." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.818, + 0.688, + 0.835 + ], + "angle": 0, + "content": "4.3. Ablation Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We conduct an ablation study to assess the contribution of each component in FakeRadar, with results presented in Table 3, Figure 3, and Figure 4. Experiments are performed in a cross-dataset setting, where the model is trained on" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13388" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.198 + ], + "angle": 0, + "content": "\\(\\mathrm{FF}++(\\mathrm{HQ})\\) and tested on CDF, DFDC, DFDCP, and DFD, with average performance reported. As shown in Table 3 (Rows M-(1) to M-(6)), removing individual or multiple components leads to a progressive decline in AUC, from \\(90.1\\%\\) to \\(86.4\\%\\), demonstrating the significance of each component. Below, we analyze the impact of FakeRadar's key components in detail." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.224, + 0.485, + 0.481 + ], + "angle": 0, + "content": "Effectiveness of Forgery Outlier Probing. To evaluate the impact of FOP, we remove or modify its submodules: dynamic subcluster modeling and cluster-conditional outlier generation. In \"FakeRadar w/o Dyn. Sub. M.\", we disable dynamic subcluster modeling, generating outliers only from clusters of samples of real and fake categories in training set. In \"FakeRadar w/o Out. Gen.\", we replace outlier samples with Gaussian noise. As shown in Figure 3(a), removing dynamic subcluster modeling reduces AUC on CDFv2, DFDC, and DFD to \\(90.2\\%\\), \\(82.3\\%\\), and \\(94.7\\%\\), respectively. While removing outlier generation has a smaller impact, performance remains below the complete FakeRadar framework. On CDFv2, FakeRadar \\((91.7\\%)\\) outperforms \"FakeRadar w/o Out. Gen.\" by \\(2.0\\%\\), demonstrating that FOP enhances generalization by generating effective outlier samples that improve the detection of novel deep-fake manipulations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.507, + 0.484, + 0.612 + ], + "angle": 0, + "content": "Effectiveness of Outlier-Guided Tri-Training. Comparing Table 3 (Rows M-(1) to M-(3)), removing Outlier-Driven Contrastive Loss or Outlier-Conditioned Cross-Entropy Loss leads to an AUC drop from \\(90.1\\%\\) to \\(87.9\\%\\) and \\(88.9\\%\\), respectively, highlighting their importance. Since both losses refine the use of outlier samples, removing them would render FOP ineffective." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The core of our proposed optimization strategy is the integration of a triplet classifier to handle the outlier samples generated by FOP. To assess its impact, we replace the triplet classifier, which performs three-class classification (real, fake, and outlier), with a binary classifier that only distinguishes between real and fake. We then analyze the prediction probability distributions for cross-dataset evaluation, as shown in Figure 3 (b) and (c). As illustrated, for real samples, both the binary and triplet classifiers exhibit similar prediction patterns with high confidence in the low-probability region. However, for fake samples, the triplet classifier demonstrates a clear advantage. Compared to the binary classifier, it assigns higher forgery probabilities to a greater number of fake samples, indicating that the proposed outlier-guided tri-training increases the model's confidence in detecting unseen fake samples. This result validates the effectiveness of treating outlier samples as fake samples during inference and highlights its contribution to improving the model's generalization capability." + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.091, + 0.904, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.202, + 0.907, + 0.258 + ], + "angle": 0, + "content": "Figure 4. Feature visualization using t-SNE. We visualize the features of (a) the pre-trained CLIP ViT-B model and our (b) FakeRadar here. We train both models using the training set FF++(HQ), while testing on the FF++ test set." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.278, + 0.909, + 0.522 + ], + "angle": 0, + "content": "Feature Visualization. We use t-SNE [48] to visualize the features of our FakeRadar and the pre-trained CLIP ViT-B model. As shown in Figure 4, the feature distribution of FakeRadar exhibits a more compact clustering structure and clearer boundaries between real and fake videos, compared to the pre-trained CLIP ViT-B model. Furthermore, FakeRadar effectively distinguishes real samples from the four types of forgery samples. This suggests that FakeRadar not only enables the model to learn discriminative features between categories but also establishes more robust decision boundaries with the aid of outlier samples. These feature distribution patterns underscore the advantages of FakeRadar in improving the model's generalization ability. The boundaries between categories are sharper, and sufficient tolerance space is provided, which is essential for addressing unseen forgery types in real-world scenarios." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.552, + 0.642, + 0.568 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this work, we introduced FakeRadar, a novel deepfake video detection framework designed to address the limitations of existing methods in cross-domain generalization. Unlike traditional detectors that passively learn specific forgery patterns, FakeRadar proactively probes unknown forgeries by incorporating Forgery Outlier Probing to simulate diverse unseen manipulation patterns. This approach enables the model to explore a broader range of potential forgeries in feature space, significantly enhancing its adaptability to novel manipulations. To further improve detection robustness, we proposed Outlier-Guided Tri-Training, allowing the model to distinguish between real, fake, and outlier samples. This training paradigm, introduced outlier-driven contrastive loss and outlier-conditioned cross-entropy loss, sharpens the decision boundaries and enhances the detector's capacity to generalize across different manipulation types. Extensive experiments on benchmark datasets confirm that FakeRadar consistently outperforms state-of-the-art detection algorithms, particularly in cross-domain evaluations where conventional detectors struggle." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13389" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.192 + ], + "angle": 0, + "content": "This work was supported in part by the National Key R&D Program of China (2024YFB3908503), in part by the Guangxi Science and Technology Project (AB25069496), and in part by the National Natural Science Foundation of China (62172120 and 62322608)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.204, + 0.188, + 0.219 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.228, + 0.482, + 0.298 + ], + "angle": 0, + "content": "[1] Weiming Bai, Yufan Liu, Zhipeng Zhang, Bing Li, and Weiming Hu. Aunet: Learning relations between action units for face forgery detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24709-24719, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.299, + 0.482, + 0.368 + ], + "angle": 0, + "content": "[2] Junyi Cao, Chao Ma, Taiping Yao, Shen Chen, Shouhong Ding, and Xiaokang Yang. End-to-end reconstruction-classification learning for face forgery detection. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4103-4112, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.369, + 0.483, + 0.438 + ], + "angle": 0, + "content": "[3] Keshigeyan Chandrasegaran, Ngoc-Trung Tran, and Ngai-Man Cheung. A closer look at fourier spectrum discrepancies for cnn-generated images detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7200-7209, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.439, + 0.483, + 0.508 + ], + "angle": 0, + "content": "[4] Liang Chen, Yong Zhang, Yibing Song, Lingqiao Liu, and Jue Wang. Self-supervised learning of adversarial example: Towards good generalizations for deepfake detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18710-18719, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.509, + 0.484, + 0.579 + ], + "angle": 0, + "content": "[5] Jongwook Choi, Taehoon Kim, Yonghyun Jeong, Seungryul Baek, and Jongwon Choi. Exploiting style latent flows for generalizing deepfake video detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1133-1143, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.58, + 0.482, + 0.633 + ], + "angle": 0, + "content": "[6] François Chollet. Xception: Deep learning with depthwise separable convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1251-1258, 2017. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.635, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[7] Danielle Keats Citron and Robert M. Chesney. Deepfakes and the new disinformation war. Foreign Affairs, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.664, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[8] Di Cooke, Abigail Edwards, Sophia Barkoff, and Kathryn Kelly. As good as a coin toss: Human detection of a-generated images, videos, audio, and audiovisual stimuli. ArXiv, abs/2403.16760, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.72, + 0.483, + 0.788 + ], + "angle": 0, + "content": "[9] Yunshu Dai, Jianwei Fei, and Fangjun Huang. Idguard: Robust, general, identity-centric poi proactive defense against face editing abuse. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11934-11943, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.79, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[10] Jiankang Deng, Jia Guo, Evangelos Ververas, Irene Kotsia, and Stefanos Zafeiriou. Retinaface: Single-shot multi-level face localisation in the wild. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5202-5211, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[11] B Dolhansky. The dee pfake detection challenge (dfdc) pre view dataset. arXiv preprint arXiv:1910.08854, 2019. 2, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.228, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[12] Brian Dolhansky, Joanna Bitton, Ben Pflaum, Jikuo Lu, Russ Howes, Menglin Wang, and Cristian Canton Ferrer. The deepfake detection challenge (dfdc) dataset. arXiv preprint arXiv:2006.07397, 2020. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[13] Chengdong Dong, Ajay Kumar, and Eryun Liu. Think twice before detecting gan-generated fake images from their spectral domain imprints. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7855-7864, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[14] Shichao Dong, Jin Wang, Renhe Ji, Jiajun Liang, Haoqiang Fan, and Zheng Ge. Implicit identity leakage: The stumbling block to improving deepfake detection generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3994-4004, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.292, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[15] Xuefeng Du, Zhaoning Wang, Mu Cai, and Sharon Li. Towards unknown-aware learning with virtual outlier synthesis. In International Conference on Learning Representations, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.349, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[16] Joel Frank, Thorsten Eisenhofer, Lea Schonherr, Asja Fischer, Dorothea Kolossa, and Thorsten Holz. Leveraging frequency analysis for deep fake image recognition. In International conference on machine learning, pages 3247-3258. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[17] Qiqi Gu, Shen Chen, Taiping Yao, Yang Chen, Shouhong Ding, and Ran Yi. Exploiting fine-grained face forgery clues via progressive enhancement learning. In AAAI Conference on Artificial Intelligence, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[18] Zhihao Gu, Yang Chen, Taiping Yao, Shouhong Ding, Jilin Li, Feiyue Huang, and Lizhuang Ma. Spatiotemporal inconsistency learning for deepfake video detection. In Proceedings of the 29th ACM international conference on multimedia, pages 3473-3481, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[19] Jiazhi Guan, Hang Zhou, Zhibin Hong, Errui Ding, Jingdong Wang, Chengbin Quan, and Youjian Zhao. Delving into sequential patches for deepfake detection. Advances in Neural Information Processing Systems, 35:4517-4530, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.604, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[20] Alexandros Haliassos, Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. Lips don't lie: A generalisable and robust approach to face forgery detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5039-5049, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.907, + 0.744 + ], + "angle": 0, + "content": "[21] Alexandros Haliassos, Rodrigo Mira, Stavros Petridis, and Maja Pantic. Leveraging real talking faces via self-supervision for robust forgery detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14950-14962, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.746, + 0.907, + 0.815 + ], + "angle": 0, + "content": "[22] Alexandros Haliassos, Rodrigo Mira, Stavros Petridis, and Maja Pantic. Leveraging real talking faces via self-supervision for robust forgery detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14950-14962, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[23] Ammarah Hashmi, Sahibzada Adil Shahzad, Chia-Wen Lin, Yu Tsao, and Hsin-Min Wang. Unmasking illusions: Understanding human perception of audiovisual deepfakes. ArXiv, abs/2405.04097, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[24] Baojin Huang, Zhongyuan Wang, Jifan Yang, Jiaxin Ai, Qin Zou, Qian Wang, and Dengpan Ye. Implicit identity" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13390" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "driven deepfake face swapping detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4490-4499, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.192 + ], + "angle": 0, + "content": "[25] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.234 + ], + "angle": 0, + "content": "[26] Marek Kowalski. Faceswap. https://github.com/MarekKowalski/FaceSwap, 2018. GitHub repository, accessed on March 3, 2025. 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.236, + 0.483, + 0.305 + ], + "angle": 0, + "content": "[27] Nicolas Larue, Ngoc-Son Vu, Vitomir Struc, Peter Peer, and Vassilis Christophides. Seeable: Soft discrepancies and bounded contrastive learning for exposing deepfakes. In Proceedings of the IEEE/CVF international conference on computer vision, pages 21011-21021, 2023. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.482, + 0.39 + ], + "angle": 0, + "content": "[28] Hong Li, Yutang Feng, Song Xue, Xuhui Liu, Bohan Zeng, Shanglin Li, Boyu Liu, Jianzhuang Liu, Shumin Han, and Baochang Zhang. Uv-idm: Identity-conditioned latent diffusion model for face uv-texture generation. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10585-10595, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.446 + ], + "angle": 0, + "content": "[29] Hanzhe Li, Jiaran Zhou, Yuezun Li, Baoyuan Wu, Bin Li, and Junyu Dong. Freqblender: Enhancing deepfake detection by blending frequency knowledge. arXiv preprint arXiv:2404.13872, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[30] Lingzhi Li, Jianmin Bao, Ting Zhang, Hao Yang, Dong Chen, Fang Wen, and Baining Guo. Face x-ray for more general face forgery detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5001-5010, 2020. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.56 + ], + "angle": 0, + "content": "[31] Yuezun Li and Siwei Lyu. Exposing deepfake videos by detecting face warping artifacts. arXiv preprint arXiv:1811.00656, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.618 + ], + "angle": 0, + "content": "[32] Yuezun Li, Ming-Ching Chang, and Siwei Lyu. In ictu oculi: Exposing ai created fake videos by detecting eye blinking. In 2018 IEEE International Workshop on Information Forensics and Security (WIFS), pages 1-7, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[33] Yuezun Li, Xin Yang, Pu Sun, Honggang Qi, and Siwei Lyu. Celeb-df: A large-scale challenging dataset for deepfake forensics. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3204-3213, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.69, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[34] Yuzhen Lin, Wentang Song, Bin Li, Yuezun Li, Jiangqun Ni, Han Chen, and Qiushi Li. Fake it till you make it: Curricular dynamic forgery augmentations towards general deepfake detection. In European Conference on Computer Vision, pages 104-122. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[35] Dat NGUYEN, Nesryne MEJRI, Inder Pal SINGH, Polina KULESHOVA, Marcella ASTRID, Anis KACEM, Enjie GHORBEL, and Djamila AOUADA. Laa-net: Localized artifact attention network for quality-agnostic and generalizable deepfake detection. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17395–17405, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[36] Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards universal fake image detectors that generalize across generative models. In Proceedings of the IEEE/CVF Conference" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "on Computer Vision and Pattern Recognition, pages 24480-24489, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[37] Junting Pan, Ziyi Lin, Xiatian Zhu, Jing Shao, and Hongsheng Li. St-adapter: Parameter-efficient image-to-video transfer learning. Advances in Neural Information Processing Systems, 35:26462-26477, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[38] Guilin Pang, Baopeng Zhang, Zhu Teng, Zige Qi, and Jianping Fan. Mre-net: Multi-rate excitation network for deepfake video detection. IEEE Transactions on Circuits and Systems for Video Technology, 33(8):3663-3676, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.236, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[39] Yuyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. Thinking in frequency: Face forgery detection by mining frequency-aware clues. In European conference on computer vision, pages 86-103. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.293, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[41] Md Shohel Rana, Mohammad Nur Nobi, Bedduhu Murali, and Andrew H Sung. Deepfake detection: A systematic literature review. IEEE access, 10:25494-25513, 2022. 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.501 + ], + "angle": 0, + "content": "[42] Meitar Ronen, Shahaf E. Finder, and Oren Freifeld. Deep-dpm: Deep clustering with an unknown number of clusters. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9861-9870, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[43] Andreas Rössler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Niessner. Faceforensics++: Learning to detect manipulated facial images. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1-11, 2019. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[44] Kaede Shiohara and Toshihiko Yamasaki. Detecting deepfakes with self-blended images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18720-18729, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[45] Ke Sun, Taiping Yao, Shen Chen, Shouhong Ding, Jilin Li, and Rongrong Ji. Dual contrastive learning for general face forgery detection. In Proceedings of the AAAI conference on artificial intelligence, pages 2316-2324, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.759 + ], + "angle": 0, + "content": "[46] Justus Thies, Michael Zollhofer, Marc Stamminger, Christian Theobalt, and Matthias Nießner. Face2face: Real-time face capture and reenactment of rgb videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2387-2395, 2016. 1, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[47] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. Acm Transactions on Graphics (TOG), 38(4):1-12, 2019. 1, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[48] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9 (86):2579-2605, 2008. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[49] Junke Wang, Zuxuan Wu, Wenhao Ouyang, Xintong Han, Jingjing Chen, Yu-Gang Jiang, and Ser-Nam Li. M2tr: Multi-modal multi-scale transformers for deepfake detec" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13391" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "tion. In Proceedings of the 2022 international conference on multimedia retrieval, pages 615-623, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.189 + ], + "angle": 0, + "content": "[50] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnn-generated images are surprisingly easy to spot... for now. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8695-8704, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.191, + 0.483, + 0.259 + ], + "angle": 0, + "content": "[51] Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, and Houqiang Li. Altfreezing for more general video face forgery detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4129-4138, 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.483, + 0.328 + ], + "angle": 0, + "content": "[52] Yuting Xu, Jian Liang, Gengyun Jia, Ziming Yang, Yanhao Zhang, and Ran He. Tall: Thumbnail layout for deepfake video detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22658-22668, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.483, + 0.398 + ], + "angle": 0, + "content": "[53] Zhiyuan Yan, Yong Zhang, Yanbo Fan, and Baoyuan Wu. Ucf: Uncovering common features for generalizable deepfake detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22412-22423, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.483, + 0.481 + ], + "angle": 0, + "content": "[54] Zhiyuan Yan, Yuhao Luo, Siwei Lyu, Qingshan Liu, and Baoyuan Wu. Transcending forgery specificity with latent space augmentation for generalizable deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8984-8994, 2024. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.483, + 0.55 + ], + "angle": 0, + "content": "[55] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications. ACM Computing Surveys, 56(4): 1-39, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.483, + 0.62 + ], + "angle": 0, + "content": "[56] Wenyuan Yang, Xiaoyu Zhou, Zhikai Chen, Bofei Guo, Zhongjie Ba, Zhihua Xia, Xiaochun Cao, and Kui Ren. Avoid-df: Audio-visual joint learning for detecting deepfake. IEEE Transactions on Information Forensics and Security, 18:2015–2029, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.483, + 0.676 + ], + "angle": 0, + "content": "[57] Shenghai Yuan, Jina Huang, Xianyi He, Yunyuan Ge, Yu-jun Shi, Liuhan Chen, Jiebo Luo, and Li Yuan. Identity-preserving text-to-video generation by frequency decomposition, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.483, + 0.734 + ], + "angle": 0, + "content": "[58] Yu Zeng, Zhe L. Lin, and Vishal M. Patel. Sketched: Mask-free local image manipulation with partial sketches. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5941-5951, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.483, + 0.801 + ], + "angle": 0, + "content": "[59] Daichi Zhang, Zihao Xiao, Shikun Li, Fanzhao Lin, Jianmin Li, and Shiming Ge. Learning natural consistency representation for face forgery video detection. In European Conference on Computer Vision, pages 407-424. Springer, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.483, + 0.859 + ], + "angle": 0, + "content": "[60] Tianchen Zhao, Xiang Xu, Mingze Xu, Hui Ding, Yuanjun Xiong, and Wei Xia. Learning self-consistency for deepfake detection. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 15003-15013, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[61] Yinglin Zheng, Jianmin Bao, Dong Chen, Ming Zeng, and Fang Wen. Exploring temporal coherence for more general video face forgery detection. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "IEEE/CVF international conference on computer vision, pages 15044-15054, 2021. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[62] Yang Zhou, Zichong Chen, and Hui Huang. Deformable oneshot face stylization via dino semantic guidance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7787-7796, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[63] Wanyi Zhuang, Qi Chu, Zhentao Tan, Qiankun Liu, Haojie Yuan, Changtao Miao, Zixiang Luo, and Nenghai Yu. Uiavit: Unsupervised inconsistency-aware method based on vision transformer for face forgery detection. In European conference on computer vision, pages 391-407. Springer, 2022. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13392" + } + ] +] \ No newline at end of file diff --git a/2025/FakeRadar_ Probing Forgery Outliers to Detect Unknown Deepfake Videos/28a1e3bc-e2b9-40da-b7fe-fcc0d6369340_origin.pdf b/2025/FakeRadar_ Probing Forgery Outliers to Detect Unknown Deepfake Videos/28a1e3bc-e2b9-40da-b7fe-fcc0d6369340_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9bdbd785fc20b02878a1cc6ce90af08d96976634 --- /dev/null +++ b/2025/FakeRadar_ Probing Forgery Outliers to Detect Unknown Deepfake Videos/28a1e3bc-e2b9-40da-b7fe-fcc0d6369340_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e774b98cf3a99bd7e4e4d9b78e8c4be58d9cfdf29f6d3fff22ab0fb6cc7f99b6 +size 16793597 diff --git a/2025/FakeRadar_ Probing Forgery Outliers to Detect Unknown Deepfake Videos/full.md b/2025/FakeRadar_ Probing Forgery Outliers to Detect Unknown Deepfake Videos/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a5611681a140927676c657df67ce3054ab96101a --- /dev/null +++ b/2025/FakeRadar_ Probing Forgery Outliers to Detect Unknown Deepfake Videos/full.md @@ -0,0 +1,314 @@ +# FakeRadar: Probing Forgery Outliers to Detect Unknown Deepfake Videos + +Zhaolun Li $^{1}$ Jichang Li $^{2*}$ Yinqi Cai $^{3}$ Junye Chen $^{3}$ Xiaonan Luo $^{1}$ Guanbin Li $^{3,2,5}$ Rushi Lan $^{1,4*}$ + +1Guilin University of Electronic Technology 2Pengcheng Laboratory 3Sun Yat-sen University 4Guangxi Key Laboratory of Image and Graphic Intelligent Processing + +$^{5}$ Guangdong Key Laboratory of Big Data Analysis and Processing + +zhaolunli@mails.guet.edu.cn, li.jichang@pcl.ac.cn, rslan2016@163.com + +# Abstract + +In this paper, we propose FakeRadar, a novel deepfake video detection framework designed to address the challenges of cross-domain generalization in real-world scenarios. Existing detection methods typically rely on manipulation-specific cues, performing well on known forgery types but exhibiting severe limitations against emerging manipulation techniques. This poor generalization stems from their inability to adapt effectively to unseen forgery patterns. To overcome this, we leverage large-scale pretrained models (e.g. CLIP) to proactively probe the feature space, explicitly highlighting distributional gaps between real videos, known forgeries, and unseen manipulations. Specifically, FakeRadar introduces Forgery Outlier Probing, which employs dynamic subcluster modeling and cluster-conditional outlier generation to synthesize outlier samples near boundaries of estimated subclusters, simulating novel forgery artifacts beyond known manipulation types. Additionally, we design Outlier-Guided Tri-Training, which optimizes the detector to distinguish real, fake, and outlier samples using proposed outlier-driven contrastive learning and outlier-conditioned cross-entropy losses. Experiments show that FakeRadar outperforms existing methods across various benchmark datasets for deepfake video detection, particularly in cross-domain evaluations, by handling the variety of emerging manipulation techniques. + +# 1. Introduction + +Recently, generative techniques [28, 58, 62] for creating deepfake videos have advanced rapidly, producing highly realistic facial forgeries that challenge human perceptual limits [8, 23]. However, the malicious use of such technology, such as fabricated news and identity fraud, poses a significant threat to digital media credibility [7, 9]. Ear + +![](images/d94756103d0736965c074e381e3861971875cef335c582a411d06cbb35a5cddb.jpg) +Figure 1. Conceptual pipeline of FakeRadar. (a)-(b) Forgery Outlier Probing first performs dynamic subcluster modeling on these features to partition finer-grained subclusters, then generates cluster-conditional outliers for each estimated subcluster to simulate unseen forgeries, emulating unknown manipulations. (c)-(d) The model is optimized using outlier-driven tri-training, allowing the detector to establish optimal boundaries for the real, fake, and outlier classes, with the latter two merged into a new "Fake" class during testing. + +lier deepfake detection methods primarily rely on the passive learning of specific forgery artifacts, such as boundary inconsistencies [27, 54], blinking anomalies [32], and texture irregularities [17, 35]. While these approaches perform well against conventional deepfake manipulations, including Face2Face [46] and NeuralTextures [47], they exhibit critical vulnerabilities when confronted with emerging generative paradigms, such as diffusion models [55]. For instance, boundary-based detection methods [30, 44] fail to identify faces generated by advanced GANs (e.g., StyleGAN [25]). Similarly, blink analysis-based techniques [32] are rendered ineffective when processing temporally smoothed fake videos [57]. These limitations stem from a fundamental shortcoming: existing methods learn a constrained decision boundary in the feature space, where + +known forgery traces occupy only a localized region, while novel manipulation techniques introduce artifacts in arbitrary locations. As a result, they struggle to generalize to new manipulation techniques that introduce artifacts in previously unseen ways, leading to significant performance degradation in cross-domain scenarios [11, 12]. + +To better understand its underlying cause, [2, 36] have analyzed the feature distributions of real and synthetic data. These studies reveal a fundamental embedding-space disparity: real data exhibit compact, dense distributions, whereas deepfake forgeries form scattered, sparse clusters due to their generation with diverse manipulation types, as demonstrated in Figure 1. This difference in distribution poses a significant challenge for existing detection methods, which are typically optimized to classify previously observed forgeries. Since deepfake forgeries do not adhere to a fixed pattern but rather introduce an infinite variety of modifications, classifiers trained on known forgery patterns struggle to recognize novel deepfakes that deviate from past observations. Therefore, a robust detection approach must move beyond merely learning pre-defined forgery patterns and instead develop the capability to identify anomalies that fall outside of previously known distributions. Given the limitations of existing methods in handling unseen forgery types, we propose a paradigm shift in deepfake detection. Instead of passively fitting to known forgery patterns, our approach proactively explores potential anomaly regions in the feature space. Inspired by radar systems that scan for unknown targets through frequency-spectrum probing, our method enhances detection robustness against novel deep-fake manipulations. + +In this work, we propose a new deepfake video detection framework, called FakeRadar, which leverages the deep feature priors of large-scale pre-trained models (i.e. CLIP [40]) to proactively "probe" unknown forgeries and enhance cross-domain generalization. Unlike conventional detectors [2, 5, 51] that perform only binary classification, FakeRadar explicitly introduces "Forgery outliers" during training to simulate potential unknown forgeries and thus "pre-explore and cover" the latent space. As shown in Figure 1 (a) and (b), we first employ Forgery Outlier Probing, which conducts dynamic subcluster modeling on the features of real and known deepfake videos. By approximating the embedding space via fine-grained Gaussian mixture distributions, this algorithm automatically partitions training samples into distinct subclusters. Next, through cluster-conditional outlier generation, Forgery Outlier Probing synthesizes outlier samples near the boundaries of these estimated subclusters with a certain "novel offset", thereby covering a broader range of potential forgery regions. + +Additionally, FakeRadar adopts a model optimization strategy called Outlier-Guided Tri-Training, enabling the model to discriminate among "Real", "Fake", and "Out + +lier” samples, with its details as illustrated in Figure 1 (c) and (d). Specifically, Outlier-Guided Tri-Training proposes an outlier-driven contrastive loss and an outlier-conditioned cross-entropy loss for end-to-end model optimization. The former loss enhances the distance between real, fake, and outlier features via contrastive learning, emphasizing their differences, while the latter one ensures that the model has clear decision boundaries for the three categories, especially ensuring that outliers are not incorrectly classified as real. In this triplet-class (Real/Fake/Outlier) setting, the model can independently label unknown forgeries during training. During inference, both “Fake” and “Outlier” are classified as forgeries, empowering the detector to adapt effectively to unknown deepfakes. + +We summarize FakeRadar's principal contributions as follows, with its conceptual pipeline illustrated in Figure 1: + +- We propose FakeRadar, a novel framework for deepfake video detection, designed to address the critical challenge of cross-domain generalization in real-world scenarios. +- We introduce Forgery Outlier Probing for synthesizing outlier samples to cover a broader range of potential forger patterns, expanding the detector's understanding of forgery distribution. +- We design Outlier-Guided Tri-Training for model optimization, which enables end-to-end training for distinguishing real, fake, and outlier samples, improving detection of unknown forgeries. +- Numerous experiments have revealed that FakeRadar outperforms current leading algorithms across multiple deepfake benchmarks, particularly excelling in cross-domain evaluations where existing methods struggle. + +# 2. Related work + +Deepfake Detection. In response to the malicious use of deepfake techniques, numerous approaches have been explored to address this issue. Generally, deepfake detection is formulated as a binary classification problem (real or fake), leading researchers to develop a variety of end-to-end detectors [45, 54] that directly distinguish reals from manipulated content. Existing detection methods can be broadly categorized into two groups: image-level and video-level detection. Specifically, image-level detectors primarily focus on identifying spatial artifacts caused by forgeries. Examples include frequency anomalies [3, 16, 29, 39], identity-based discrepancies [13, 14, 24], blending artifacts [30, 44], and localized distortions [35, 49]. In contrast, video-level detectors employ a wider range of strategies. For example, some algorithms, such as [20, 32], concentrate on temporal consistency, typically reflected in correlations across consecutive frames. Other video detection studies [18, 38, 61] incorporate both temporal and spatiotemporal features to establish a more comprehensive detection framework. Another line of research [22, 56] examines audio-visual con + +![](images/586c8307ffd116b23c397ed996a152fff5e4080e49e37e8fbc482682fa729c41.jpg) +Figure 2. An overview of FakeRadar for deepfake video detection. FakeRadar consists of two key components: Forgery Outlier Probing and Outlier-Guided Tri-Training. (i) Forgery Outlier Probing performs dynamic subcluster modeling to model the feature-space distributions of real and fake samples using GMMs, followed by dynamic subcluster splitting and merging to produce finer-grained subcluster partitions. Afterwards, FakeRadar conducts cluster-conditioned outlier generation to simulate unseen forgeries for novel manipulation techniques. (ii) Outlier-Guided Tri-Training jointly optimizes the CLIP model's ViT-B/16 backbone, inserted with ST-Adapter [37] for parameter-efficient fine tuning, and a triplet-class classifier to distinguish between "Real", "Fake", and "Outlier" classes, with "Fake" and "Outlier" merged into a single new "Fake" class during evaluation. The training set includes real videos, as well as deepfake videos generated using four different manipulations, i.e. "DF" [41], "FS" [26], "NT" [47] and "F2F" [46]. + +sistency by analyzing the synchronization and correlation between video and audio signals to detect forgeries. In summary, these methods rely heavily on existing artifacts in training set, limiting their generalization to unseen manipulation types. In contrast, our approach explores feature-space anomalies, enhancing the generalization capacity of our proposed detector to diverse deepfake manipulations. + +Towards Generalization in Deepfake Detection. Rapid advance in deepfakes have significantly improved the quality and authenticity of generated content, posing substantial challenges to the generalization capability of detection systems. Existing detectors [6, 43, 50] often suffer from severe performance degradation when encountering novel forgery techniques or previously unseen manipulations. To address this issue, researchers have proposed various strategies to enhance detector generalization. For instance, FTCN [61] improves generalization by capturing temporal coherence features in video sequences. Similarly, AltFreezing [51] employs an alternating weight-freezing training strategy to facilitate the joint detection of spatial and temporal forgery artifacts. Another algorithm, StyleFlow [5], leverages the temporal dynamics of style latent vectors to identify anomalous patterns in generated videos. + +More recently, data synthesis has emerged as an effective strategy for improving detector generalization. For example, FWA [31] enhances training data by introducing facial warping artifacts, while Face X-ray [30] detects forgeries by identifying blending boundaries, independent of specific facial manipulation techniques. Building on this concept, SBI [44] synthesizes training samples by blending different views of the same pristine image to simulate generic forgery + +artifacts. Additionally, a series of recent studies [1, 27, 34] have further advanced this direction by employing various data synthesis techniques. Unlike existing methods that primarily focus on image- or video-level data synthesis for deepfake detection, our approach proposes to explore unseen forgeries in feature space. + +# 3. Methodology + +# 3.1. Preliminaries + +Overview. In this paper, we introduce FakeRadar, to overcome the challenges of cross-domain generalization in deepfake video detection by exploring the feature space to identify emerging manipulation patterns. Specifically, FakeRadar is built upon two key components: (i) Forgery Outlier Probing, which models the feature-space distributions of subclusters for training samples of both real and fake categories, and then generates cluster-conditioned outlier samples to simulate unseen forgeries, and (ii) Outlier-Guided Tri-Training, which jointly optimizes an adapter-inserted backbone and a triplet-class classifier to distinguish among samples from the "Real", "Fake", and "Outlier" classes. An overview of FakeRadar can be found in Figure 2. + +Model Adaptation with a Frozen Backbone. The FakeRadar framework employs a pre-trained Vision Transformer (ViT-B) from CLIP [40] as a frozen backbone to preserve its learned semantic features. This model serves as the video encoder for input clips. To adapt it for deepfake detection, we integrate lightweight ST-Adapter layers [37] after each Transformer block, enabling parameter-efficient fine-tuning. This significantly enhances the model's ability to + +capture spatial manipulations and temporal inconsistencies, which are crucial for detecting deepfake videos. Given a feature vector $x$ from the video encoder, each adapter transforms it as: + +$$ +\operatorname {S T - A d p a t e r} (x) = x + \operatorname {R e L U} (\operatorname {C o n v 3 D} (x W _ {\text {d o w n}})) W _ {\text {u p}}. \tag {1} +$$ + +Here, $\mathrm{Conv3D}(\cdot)$ denotes a 3D convolution operation employed for capturing spatio-temporal features from videos. As well, $W_{\mathrm{down}}$ and $W_{\mathrm{up}}$ are weight matrices that perform dimensionality reduction and expansion, respectively, and $\mathrm{ReLU}(\cdot)$ is the activation function. By introducing only a small number of additional parameters, these adapters allow efficient fine-tuning while retaining the original backbone's representational strength. After feature extraction, each video clip is represented as a $d$ -dimensional vector for subsequent analysis, with $d = 768$ . + +Inference. During inference, the model classifies input samples as either real or deepfake, with deepfakes categorized as "Fake" or "Outlier". The "Fake" class includes fake samples with known manipulation patterns, while the "Outlier" class detects testing samples with novel forgeries that differ from established manipulations. By unifying these categories, the model enhances its robustness against evolving manipulation techniques, enabling accurate detection of both familiar and emerging deepfakes. This improves its generalization performance in cross-domain evaluations. + +# 3.2. Forgery Outlier Probing + +In this section, we introduce Forgery Outlier Probing (FOP) to generate outlier samples with unseen forgeries beyond the existing real and fake videos, simulating previously unobserved manipulation patterns. Specifically, FOP first presents dynamic subcluster modeling, which performs category-specific clustering on the training samples of real and fake categories. Then, it models the distributions of these categories using a strategy based on Gaussian Mixture Models (GMMs), and follows by applying dynamic subcluster splitting and merging, thereby achieving finer-grained subcluster partitioning. Finally, we generate cluster-conditioned outliers based on the resulting subclusters to stimulate unseen forgeries. + +Dynamic Subcluster Modeling. To capture subtle intra-class variations within real and deepfake videos, we apply category-specific clustering on the FaceForensics++ (FF++) dataset. Specifically, we treat the real class and each of the four fake classes (i.e. four types of deepfake videos created by different manipulation techniques [26, 41, 46, 47]) as five separate categories. For each category $c$ , let $\mathcal{X}^{(c)} = \{\mathbf{x}_1, \mathbf{x}_2, \dots, \mathbf{x}_N\}$ denote the set of feature embeddings (i.e. extracted from the video encoder). Inspired by [42], we then + +define a main clustering network: + +$$ +f _ {\operatorname {m a i n}} ^ {(c)} \left(\mathbf {x} _ {i}\right) \mapsto \mathbf {r} _ {i} = \left(r _ {i, 1}, r _ {i, 2}, \dots , r _ {i, K}\right), \tag {2} +$$ + +which initially assumes $K$ clusters, empirically with $K = 5$ , providing a "soft" assignment of each embedding $\mathbf{x}_i \in \mathcal{X}^{(c)}$ to these $K$ clusters. Here, $r_{i,k} \in [0,1]$ and $\sum_{k=1}^{K} r_{i,k} = 1$ . + +Each cluster can be modeled as a Gaussian distribution using a GMM, allowing us to represent the feature distribution of each category in a probabilistic manner. Specifically, for each cluster $k$ in category $c$ , we define a Gaussian component characterized by $\{\pi_k,\pmb {\mu}_k,\pmb {\Sigma}_k\}$ , where $\pi_{k}\in (0,1)$ is the mixture weight for cluster $k$ , with $\sum_{k = 1}^{K}\pi_{k} = 1$ , while $\pmb{\mu}_{k}$ and $\pmb{\Sigma}_{k}$ represent its corresponding mean and covariance matrix. These parameters enable the clustering network to estimate the likelihood of each data sample belonging to different clusters, aligning our model with a GMM-based formulation. + +After that, the main clustering network needs to learn these parameters by aligning its soft assignments with GMM responsibilities: + +$$ +r _ {i, k} ^ {E} = \frac {\pi_ {k} \mathcal {N} \left(\mathbf {x} _ {i} ; \boldsymbol {\mu} _ {k} , \boldsymbol {\Sigma} _ {k}\right)}{\sum_ {k ^ {\prime} = 1} ^ {K} \pi_ {k ^ {\prime}} \mathcal {N} \left(\mathbf {x} _ {i} ; \boldsymbol {\mu} _ {k ^ {\prime}} , \boldsymbol {\Sigma} _ {k ^ {\prime}}\right)}, \tag {3} +$$ + +where $\mathcal{N}(\cdot ;\pmb {\mu}_k,\pmb {\Sigma}_k)$ is the multivariate Gaussian density function. The network is trained using the KL-divergence loss: + +$$ +\mathcal {L} _ {\text {m a i n}} = \sum_ {i = 1} ^ {N} \mathrm {K L} \left(\mathbf {r} _ {i} \| \mathbf {r} _ {i} ^ {E}\right), \tag {4} +$$ + +which ensures that the learned cluster assignments match the inferred GMM responsibilities. + +To further refine cluster structures, we introduce a subclustering network for each cluster $k$ , which attempts to split the cluster into two subclusters if multi-modal structures are detected. A compactness loss is applied to encourage subclusters to be well-separated. This loss function penalizes subclusters that are too close to each other: + +$$ +\mathcal {L} _ {\mathrm {s u b}} = \sum_ {k = 1} ^ {K} \sum_ {j = 1} ^ {2} \sum_ {\mathbf {x} _ {i} \in X _ {k}} \tilde {r} _ {i, j} \| \mathbf {x} _ {i} - \tilde {\boldsymbol {\mu}} _ {k, j} \| _ {2} ^ {2}, \tag {5} +$$ + +where $\tilde{r}_{i,j}$ represents the probability of $\mathbf{x}_i$ belonging to subcluster $j$ , and $\tilde{\mu}_{k,j}$ is its mean. + +Since the predefined cluster number $K$ may not be optimal, we propose to dynamically adjust it through split/merge proposals, inspired by [42]. Suppose a cluster $k$ contains $N_{k}$ points, $X_{k}$ . We split if subclustering indicates that $X_{k}$ is better viewed as two subgroups, $\{X_{k,1}, X_{k,2}\}$ . The Hastings ratio is given by + +$$ +H _ {s} = \frac {\alpha \Gamma (N _ {k , 1}) f _ {x} (X _ {k , 1}) \Gamma (N _ {k , 2}) f _ {x} (X _ {k , 2})}{\Gamma (N _ {k}) f _ {x} (X _ {k})}, \tag {6} +$$ + +where $f_{x}(\cdot)$ is the marginal likelihood under a weak NIw prior, $\alpha$ is the Dirichlet-process concentration parameter, and $\Gamma (\cdot)$ is the Gamma function. + +If $H_{s}$ exceeds a threshold, cluster $k$ is split into two new clusters. The parameters for these new clusters are initialized from $\tilde{\mu}_{k,1},\tilde{\Sigma}_{k,1}$ and $\tilde{\mu}_{k,2},\tilde{\Sigma}_{k,2}$ , respectively. Merging is similarly proposed if two clusters significantly overlap, with acceptance ratio $H_{m} = 1 / H_{s}$ . Repeating these steps yields an effective number of clusters $K^{(c)}$ , plus subclusters for each category $c$ . + +After these dynamic adjustments, we obtain the final subcluster set $\hat{\mathcal{C}} = \{\hat{C}_1,\hat{C}_2,\dots,\hat{C}_{K^{(c)}}\}$ , where $K^{(c)}$ is the refined number of clusters. Each adjusted subcluster $\hat{C}_k$ is characterized by its updated mean and covariance: + +$$ +\hat {\boldsymbol {\mu}} _ {k} = \frac {1}{| \hat {C} _ {k} |} \sum_ {\mathbf {x} _ {i} \in \hat {C} _ {k}} \mathbf {x} _ {i}, \tag {7} +$$ + +$$ +\hat {\boldsymbol {\Sigma}} = \frac {1}{| \hat {C} _ {k} |} \sum_ {\mathbf {x} _ {i} \in \hat {C} _ {k}} \left(\mathbf {x} _ {i} - \hat {\boldsymbol {\mu}} _ {k}\right) \left(\mathbf {x} _ {i} - \hat {\boldsymbol {\mu}} _ {k}\right) ^ {\top}. \tag {8} +$$ + +We use online estimation for efficient training by maintaining a cluster-conditional queue with $|\hat{C}_k|$ sample instances from each subcluster. In each iteration, we enqueue the embeddings of objects to their corresponding cluster-conditional queues and dequeue the same number of object embeddings. + +Cluster-Conditional Outlier Generation. After deriving finer-grained cluster partitions for each category, we generate feature-space outliers to probe potential unknown forgeries. Motivated by [15], in order to ensure that the sampled outliers lie near the cluster boundary, we maintain a sufficiently small value $\epsilon$ for outlier sampling from the $\epsilon$ -likelihood region of the estimated cluster-conditional distribution. Therefore, for a subcluster from $\hat{C}$ indexed by $k$ , the outlier samples it generates can be formulated as follows: + +$$ +\mathcal {V} _ {k} = \left\{\mathbf {v} _ {k} \mid \frac {\exp \left(- \frac {1}{2} \left(\mathbf {v} _ {k} - \hat {\boldsymbol {\mu}} _ {k}\right) ^ {\top} \hat {\boldsymbol {\Sigma}} ^ {- 1} \left(\mathbf {v} _ {k} - \hat {\boldsymbol {\mu}} _ {k}\right)\right)}{(2 \pi) ^ {d / 2} | \hat {\boldsymbol {\Sigma}} | ^ {1 / 2}} < \epsilon \right\}, \tag {9} +$$ + +where $v_{k}\sim \mathcal{N}(\hat{\mu}_{k},\hat{\Sigma}_{k})$ represents the virtual outliers sampled from the estimated Gaussian distribution of subcluster $k$ , with $d$ denoting its embedding dimensionality. It is worth noting that during model optimization, all outliers generated here, i.e. $\overline{\mathcal{V}} = \mathcal{V}_1\cup \dots \cup \mathcal{V}_k\cup \dots \mathcal{V}_{|\hat{\mathcal{O}} |}$ are classified into the "Outlier" class, separate from the real and fake classes. + +# 3.3. Outlier-Guided Tri-Training + +To enable better performance of FakeRadar for tasks of deepfake video detection, we design a strategy called Outlier-Guided Tri-Training for model optimization jointly + +combining proposed outlier-driven contrastive loss and outlier-conditioned cross-entropy loss. With this approach, a triplet-class classifier (integrated with the pre-trained backbone and adapters) is trained to explicitly distinguish among three categories: "Real", "Fake", and "Outlier". This design not only improves the model's ability to detect known forgeries but also sensitizes it to emerging manipulation patterns. + +Outlier-Driven Contrastive Loss. In this work, we propose an outlier-driven contrastive loss to enhance the model's ability to delineate class boundaries within the feature space, thereby improving feature discriminability. This loss function is designed to maximize the similarity between samples and their respective subcluster centers while minimizing their similarity to other subcluster centers. Additionally, it treats generated outliers as negative examples, enforcing low similarity between samples and outliers to strengthen boundary discrimination. + +To this end, the loss is formulated using an InfoNCE-like contrastive loss given an input sample $i$ : + +$$ +\mathcal {L} _ {\text {c o n}} = - \log \frac {\exp \left(\frac {s \left(h _ {i} , \mu_ {k _ {i} ^ {+}}\right)}{\tau}\right)}{\sum_ {k ^ {-}} \exp \left(\frac {s \left(h _ {i} , \mu_ {k ^ {-}}\right)}{\tau}\right) + \sum_ {v _ {j} \in \overline {{\nu}}} \exp \left(\frac {s \left(h _ {i} , v _ {j}\right)}{\tau}\right)}, \tag {10} +$$ + +where $h_i$ denotes the feature representation of sample $i$ , while $\mathcal{V}$ represents the outlier cache, providing additional hard negative samples to improve the model's ability to distinguish boundary samples. Additionally, $k^{-}$ represents all subclusters except the one to which sample $i$ belongs, with its corresponding $u_{k^{-}}$ denoting the centers of these subclusters, used as negative samples. The function $s(\cdot ,\cdot)$ denotes cosine similarity, which quantifies the similarity between samples, and $\tau$ is a temperature hyperparameter that controls the smoothness of similarity scaling. + +Outlier-Conditioned Cross-Entropy Loss. While the proposed outlier-driven contrastive loss enforces stronger feature-space separation, we still require a concrete classification objective to handle a three-category task: real, fake, and outlier. In particular, the model must not only determine whether a video is fake or real but also assign a separate label for those samples situated near decision boundaries (i.e., outliers). This triplet-class cross-entropy loss is crucial, as it enables the model to explicitly differentiate uncertain boundary samples and label them as outliers, thereby avoiding confusion. + +Concretely, the triplet-class cross-entropy loss can be defined as follows: + +$$ +\mathcal {L} _ {\mathrm {c l s}} = - \sum_ {c \in \{\text {R e a l}, \text {F a k e}, \text {O u t l i e r} \}} y _ {c} \log p _ {c}, \tag {11} +$$ + +
MethodBackboneInput TypeTesting Set AUC (%)
FF++CDFv2DFDCPDFDCDFD
Xception [6]XceptionFrame99.373.7-70.9-
Face X-Ray [30]HRNetFrame97.879.5-65.595.4
PCL+i2G [60]ResnetFrame99.890.074.467.5-
SLADD [4]XceptionFrame98.479.776.0--
UIA-ViT [63]Vision TransformerFrame99.382.475.8-94.7
UCF [53]XceptionFrame-82.4-80.594.5
SeeABLE [27]EfficientNetFrame-87.386.375.9-
LSDA [54]EfficientNetFrame-91.181.277.095.6
FTCN [61]3D ResNetVideo99.886.9-74.0-
RealForensics [21]3D ResNetVideo99.886.9-75.9-
LTTD [19]Vision TransformerVideo99.589.3-80.4-
AltFreezing [51]3D ResNetVideo99.789.5--93.7
TALL [52]Swin TransformerVideo99.990.8-76.8-
StyleFlow [5]3D ResNetVideo99.189.0--96.1
NACO [59]Vision TransformerVideo99.889.5-76.7-
FakeRadar (Ours)Vision TransformerVideo99.191.788.584.196.2
+ +where $y_{c}$ is the ground-truth label for samples belonging to class $c$ , and $p_{c}$ is the model-predicted probability of samples from that class, with $c =$ "Outlier" denoting the outlier label. This loss compels the model to accurately partition each category in the feature space, especially for samples near class boundaries. + +Total Loss. Finally, we combine both losses into the total objective for model optimization: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {c o n}} + \lambda \mathcal {L} _ {\text {c l s}}, \tag {12} +$$ + +where $\lambda = 0.5$ is a hyperparameter balancing the relative importance of classification and contrastive goals. In practice, each mini-batch comprises 16 training samples $(\mathrm{FF}++)$ , and 16 outlier samples during model training. + +# 4. Experiments + +# 4.1. Experiment Setups + +Datasets. Following established experimental practices [5, 51, 52, 54], this study primarily utilizes the high-quality (HQ) version of the FaceForensics++ dataset [43], denoted as $\mathrm{FF} + + (\mathrm{HQ})$ , for FakeRadar's evaluation. This dataset consists of 1000 pristine videos and 4000 deepfakes videos generated using four different manipulation techniques: DFD [41], F2F [46], FS [26], and NT [47]. The dataset is split into training, validation, and testing sets according to the official protocol. This study primarily focuses on evaluating the cross-domain performance of our method. To this end, we incorporate four additional datasets for evaluation, including CDFv2 [33], DFDCP [11], DFDC [12], and DFD [41]. + +Implementation Details. To ensure fair comparisons, we maintain most implementation details consistent with exist + +Table 1. Generalization comparisons of FakeRadar and state-of-the-art algorithms on cross-dataset evaluation, using video-level AUC (%) as the metric. + +
Training SetMethodDFF2FFSNTavg.
DFDCL [45]99.9877.1361.0175.0178.28
UIA-ViT [63]99.2974.4453.8970.9274.64
TALL [52]99.3569.3354.3874.7474.45
AltFreezing [51]98.8069.3372.0777.4179.40
FakeRadar (Ours)99.9586.4788.1266.3682.23
F2FDCL [45]91.9199.2159.5866.6779.34
UIA-ViT [63]83.3998.3268.3767.1779.31
TALL [52]85.5298.6762.1570.2479.15
AltFreezing [51]81.1298.8062.9770.2678.29
FakeRadar (Ours)96.9499.0982.3368.4686.75
FSDCL [45]74.8069.7599.9052.6074.26
UIA-ViT [63]81.0266.3099.0449.2673.91
TALL [52]77.8367.3098.7752.5774.12
AltFreezing [51]80.6765.9499.7857.2375.90
FakeRadar (Ours)96.3982.0699.8352.6482.73
NTDCL [45]91.2352.1379.3198.9780.41
UIA-ViT [63]79.3767.8945.9494.5971.95
TALL [52]84.0272.8551.6695.1675.92
AltFreezing [51]88.8176.7857.9295.7179.80
FakeRadar (Ours)96.7183.3467.5098.4186.49
+ +Table 2. Generalization comparisons of FakeRadar and state-of-the-art algorithms on cross-manipulation evaluation, using video-level AUC (\%) as the metric. + +ing methods [5, 51, 59]. However, our FakeRadar framework employs ViT-B/16 from CLIP [40] as the backbone, into which we insert an ST-Adapter [37] with a 384-dimensional bottleneck for parameter-efficient fine-tuning. Additionally, the classification head of our triplet-class classifier consists of a single linear layer with softmax activation. During model training, we sample four temporal clips from each video, each containing 12 consecutive frames. We use Adam as the optimizer and apply a cosine learning rate scheduler with an initial learning rate of $1 \times 10^{-4}$ . The model is trained for 60 epochs with a batch size of 16 and a weight decay of $5 \times 10^{-4}$ . Prior to model optimization, we perform data preprocessing by detecting faces with RetinaFace [10] and then standardize all face images + +![](images/922b9eeaef948d8603a75b38ca1f8e6234094f1e1a4be5d0b342ec5031c10e7e.jpg) +Figure 3. Evaluation of (a) Forgery Outlier Probing and (b)/(c) Outlier-Guided Tri-Training by FakeRadar. (a) "FakeRadar w/o Dyn. Sub. M." and "FakeRadar w/o Out. Gen." indicate that dynamic subcluster modeling and cluster-conditional outlier generation have been removed. The model is trained on FF++(HQ) and evaluated on DFDC [12], CDFv2 [33], and DFD [41]. (b) and (c) show, respectively, the count of real and fake test samples classified as "fake" across predicted probability intervals. The blue solid line denotes the binary classifier's distribution (distinguishing only real vs. fake), while the red dashed line denotes that of our proposed triplet classifier, which treats both fake and outlier samples as the new "fake". In this evaluation, the model is trained on FF++(HQ) and tested on CDFv2 [33]. + +![](images/31311a0a131bb374e262660ed95c596955f78aaafd597515461b720503e075eb.jpg) + +![](images/68738ff670573c43f6792a7e60db13307fc720d2a45c0edf65643c50fc6dc79b.jpg) + +to $224\times 224$ pixels. + +Evaluation Metrics. In this work, consistent with previous studies [5, 51, 52] employing video-level inputs, we use the area under the ROC curve (AUC) as the performance metric during evaluations. For methods that use images as model inputs, e.g. [27, 54], their video-level results are derived by averaging the predictions across all video frames. + +# 4.2. Comparisons with the State-of-the-Arts + +In this study, we focus on evaluating the capabilities of cross-domain generalization for our detector to validate its effectiveness in detecting deepfake videos manipulated through various techniques. To achieve this, we employ two strategies for evaluation: cross-dataset evaluation and cross-manipulation evaluation, both of which demonstrate that our results predominantly surpass existing state-of-the-art (SoTA) algorithms for deepfake video detection, as illustrated by Table 1 and Table 2. + +Cross-Dataset Evaluation. In this situation, we train our FakeRadar on the FF++(HQ) and test its performance on multiple benchmark datasets, including CDFv2 [33], DFD [41], DFDC [12], and DFDCP [11]. As reported in Table 1, FakeRadar demonstrates its superior generalization capability in cross-dataset evaluation, achieving the best performance across all cross-dataset cases while maintaining comparable performance on the FF++ test dataset against SoTA algorithms. Specifically, our algorithm significantly improves over other methods with images and videos as inputs, particularly outperforming the current best methods, UCF [53] and LTTD [19], by $3.6\%$ and $3.7\%$ in the AUC metric, respectively, on the dataset DFDC [12]. These results robustly showcase the proposed method's enhanced ability of generalization to handle unseen manipulation techniques in cross-dataset evaluation scenarios. + +
M-(#)FOPOGTTCDFDFDCDFDCPDFDavg.
ODCLOCCE
1✓✓✓91.684.188.596.290.1
2✓✓89.981.285.694.987.9
3✓✓88.680.987.794.788.0
4✓88.880.286.794.587.6
5✓88.478.485.194.386.7
688.278.384.894.286.4
+ +Table 3. Ablation study results of FakeRadar by removing its one or more proposed components, using video-level AUC (\%) as the metric. For clarity, "FOP", "OGTT", "ODCL", and "OCCE" refer to Forgery Outlier Probing, Outlier-Guided Tri-Training, Outlier-Driven Contrastive Loss, and Outlier-Conditioned Cross-Entropy Loss, respectively. + +Cross-Manipulation Evaluation. To further assess our model's generalization across different manipulation techniques, we conduct cross-manipulation evaluation on the dataset FF++(HQ). In this case, we train the model using only real samples from FF++(HQ) and one manipulation type of deepfake samples, then test its performance on the remaining three manipulation types within FF++(HQ). As shown in Table 2, the proposed FakeRadar achieves the best average performance in all testing cases, particularly excelling in scenarios trained with F2F as the manipulation type, surpassing the second-best performing method, DCL [45], by $7.41\%$ in AUC. In evaluations focused on a single manipulation type, our method also achieves the best performance in most cases and the second-best in a few. This indicates that our approach, in cross-manipulation evaluation, demonstrates a performance advantage in cross-domain generalization compared to current SoTA algorithms of detecting deepfake videos. + +# 4.3. Ablation Analysis + +We conduct an ablation study to assess the contribution of each component in FakeRadar, with results presented in Table 3, Figure 3, and Figure 4. Experiments are performed in a cross-dataset setting, where the model is trained on + +$\mathrm{FF}++(\mathrm{HQ})$ and tested on CDF, DFDC, DFDCP, and DFD, with average performance reported. As shown in Table 3 (Rows M-(1) to M-(6)), removing individual or multiple components leads to a progressive decline in AUC, from $90.1\%$ to $86.4\%$ , demonstrating the significance of each component. Below, we analyze the impact of FakeRadar's key components in detail. + +Effectiveness of Forgery Outlier Probing. To evaluate the impact of FOP, we remove or modify its submodules: dynamic subcluster modeling and cluster-conditional outlier generation. In "FakeRadar w/o Dyn. Sub. M.", we disable dynamic subcluster modeling, generating outliers only from clusters of samples of real and fake categories in training set. In "FakeRadar w/o Out. Gen.", we replace outlier samples with Gaussian noise. As shown in Figure 3(a), removing dynamic subcluster modeling reduces AUC on CDFv2, DFDC, and DFD to $90.2\%$ , $82.3\%$ , and $94.7\%$ , respectively. While removing outlier generation has a smaller impact, performance remains below the complete FakeRadar framework. On CDFv2, FakeRadar $(91.7\%)$ outperforms "FakeRadar w/o Out. Gen." by $2.0\%$ , demonstrating that FOP enhances generalization by generating effective outlier samples that improve the detection of novel deep-fake manipulations. + +Effectiveness of Outlier-Guided Tri-Training. Comparing Table 3 (Rows M-(1) to M-(3)), removing Outlier-Driven Contrastive Loss or Outlier-Conditioned Cross-Entropy Loss leads to an AUC drop from $90.1\%$ to $87.9\%$ and $88.9\%$ , respectively, highlighting their importance. Since both losses refine the use of outlier samples, removing them would render FOP ineffective. + +The core of our proposed optimization strategy is the integration of a triplet classifier to handle the outlier samples generated by FOP. To assess its impact, we replace the triplet classifier, which performs three-class classification (real, fake, and outlier), with a binary classifier that only distinguishes between real and fake. We then analyze the prediction probability distributions for cross-dataset evaluation, as shown in Figure 3 (b) and (c). As illustrated, for real samples, both the binary and triplet classifiers exhibit similar prediction patterns with high confidence in the low-probability region. However, for fake samples, the triplet classifier demonstrates a clear advantage. Compared to the binary classifier, it assigns higher forgery probabilities to a greater number of fake samples, indicating that the proposed outlier-guided tri-training increases the model's confidence in detecting unseen fake samples. This result validates the effectiveness of treating outlier samples as fake samples during inference and highlights its contribution to improving the model's generalization capability. + +![](images/226d93e674b4791eef2686dd56434c9ff217a868a209b633ea98a8efe1423524.jpg) +Figure 4. Feature visualization using t-SNE. We visualize the features of (a) the pre-trained CLIP ViT-B model and our (b) FakeRadar here. We train both models using the training set FF++(HQ), while testing on the FF++ test set. + +Feature Visualization. We use t-SNE [48] to visualize the features of our FakeRadar and the pre-trained CLIP ViT-B model. As shown in Figure 4, the feature distribution of FakeRadar exhibits a more compact clustering structure and clearer boundaries between real and fake videos, compared to the pre-trained CLIP ViT-B model. Furthermore, FakeRadar effectively distinguishes real samples from the four types of forgery samples. This suggests that FakeRadar not only enables the model to learn discriminative features between categories but also establishes more robust decision boundaries with the aid of outlier samples. These feature distribution patterns underscore the advantages of FakeRadar in improving the model's generalization ability. The boundaries between categories are sharper, and sufficient tolerance space is provided, which is essential for addressing unseen forgery types in real-world scenarios. + +# 5. Conclusions + +In this work, we introduced FakeRadar, a novel deepfake video detection framework designed to address the limitations of existing methods in cross-domain generalization. Unlike traditional detectors that passively learn specific forgery patterns, FakeRadar proactively probes unknown forgeries by incorporating Forgery Outlier Probing to simulate diverse unseen manipulation patterns. This approach enables the model to explore a broader range of potential forgeries in feature space, significantly enhancing its adaptability to novel manipulations. To further improve detection robustness, we proposed Outlier-Guided Tri-Training, allowing the model to distinguish between real, fake, and outlier samples. This training paradigm, introduced outlier-driven contrastive loss and outlier-conditioned cross-entropy loss, sharpens the decision boundaries and enhances the detector's capacity to generalize across different manipulation types. Extensive experiments on benchmark datasets confirm that FakeRadar consistently outperforms state-of-the-art detection algorithms, particularly in cross-domain evaluations where conventional detectors struggle. + +# Acknowledgments + +# References + +[1] Weiming Bai, Yufan Liu, Zhipeng Zhang, Bing Li, and Weiming Hu. Aunet: Learning relations between action units for face forgery detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24709-24719, 2023. 3 +[2] Junyi Cao, Chao Ma, Taiping Yao, Shen Chen, Shouhong Ding, and Xiaokang Yang. End-to-end reconstruction-classification learning for face forgery detection. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4103-4112, 2022. 2 +[3] Keshigeyan Chandrasegaran, Ngoc-Trung Tran, and Ngai-Man Cheung. A closer look at fourier spectrum discrepancies for cnn-generated images detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7200-7209, 2021. 2 +[4] Liang Chen, Yong Zhang, Yibing Song, Lingqiao Liu, and Jue Wang. Self-supervised learning of adversarial example: Towards good generalizations for deepfake detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18710-18719, 2022. 6 +[5] Jongwook Choi, Taehoon Kim, Yonghyun Jeong, Seungryul Baek, and Jongwon Choi. Exploiting style latent flows for generalizing deepfake video detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1133-1143, 2024. 2, 3, 6, 7 +[6] François Chollet. Xception: Deep learning with depthwise separable convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1251-1258, 2017. 3, 6 +[7] Danielle Keats Citron and Robert M. Chesney. Deepfakes and the new disinformation war. Foreign Affairs, 2018. 1 +[8] Di Cooke, Abigail Edwards, Sophia Barkoff, and Kathryn Kelly. As good as a coin toss: Human detection of a-generated images, videos, audio, and audiovisual stimuli. ArXiv, abs/2403.16760, 2024. 1 +[9] Yunshu Dai, Jianwei Fei, and Fangjun Huang. Idguard: Robust, general, identity-centric poi proactive defense against face editing abuse. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11934-11943, 2024. 1 +[10] Jiankang Deng, Jia Guo, Evangelos Ververas, Irene Kotsia, and Stefanos Zafeiriou. Retinaface: Single-shot multi-level face localisation in the wild. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5202-5211, 2020. 6 +[11] B Dolhansky. The dee pfake detection challenge (dfdc) pre view dataset. arXiv preprint arXiv:1910.08854, 2019. 2, 6, 7 + +This work was supported in part by the National Key R&D Program of China (2024YFB3908503), in part by the Guangxi Science and Technology Project (AB25069496), and in part by the National Natural Science Foundation of China (62172120 and 62322608). +[12] Brian Dolhansky, Joanna Bitton, Ben Pflaum, Jikuo Lu, Russ Howes, Menglin Wang, and Cristian Canton Ferrer. The deepfake detection challenge (dfdc) dataset. arXiv preprint arXiv:2006.07397, 2020. 2, 6, 7 +[13] Chengdong Dong, Ajay Kumar, and Eryun Liu. Think twice before detecting gan-generated fake images from their spectral domain imprints. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7855-7864, 2022. 2 +[14] Shichao Dong, Jin Wang, Renhe Ji, Jiajun Liang, Haoqiang Fan, and Zheng Ge. Implicit identity leakage: The stumbling block to improving deepfake detection generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3994-4004, 2023. 2 +[15] Xuefeng Du, Zhaoning Wang, Mu Cai, and Sharon Li. Towards unknown-aware learning with virtual outlier synthesis. In International Conference on Learning Representations, 2022. 5 +[16] Joel Frank, Thorsten Eisenhofer, Lea Schonherr, Asja Fischer, Dorothea Kolossa, and Thorsten Holz. Leveraging frequency analysis for deep fake image recognition. In International conference on machine learning, pages 3247-3258. PMLR, 2020. 2 +[17] Qiqi Gu, Shen Chen, Taiping Yao, Yang Chen, Shouhong Ding, and Ran Yi. Exploiting fine-grained face forgery clues via progressive enhancement learning. In AAAI Conference on Artificial Intelligence, 2021. 1 +[18] Zhihao Gu, Yang Chen, Taiping Yao, Shouhong Ding, Jilin Li, Feiyue Huang, and Lizhuang Ma. Spatiotemporal inconsistency learning for deepfake video detection. In Proceedings of the 29th ACM international conference on multimedia, pages 3473-3481, 2021. 2 +[19] Jiazhi Guan, Hang Zhou, Zhibin Hong, Errui Ding, Jingdong Wang, Chengbin Quan, and Youjian Zhao. Delving into sequential patches for deepfake detection. Advances in Neural Information Processing Systems, 35:4517-4530, 2022. 6, 7 +[20] Alexandros Haliassos, Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. Lips don't lie: A generalisable and robust approach to face forgery detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5039-5049, 2021. 2 +[21] Alexandros Haliassos, Rodrigo Mira, Stavros Petridis, and Maja Pantic. Leveraging real talking faces via self-supervision for robust forgery detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14950-14962, 2022. 6 +[22] Alexandros Haliassos, Rodrigo Mira, Stavros Petridis, and Maja Pantic. Leveraging real talking faces via self-supervision for robust forgery detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14950-14962, 2022. 2 +[23] Ammarah Hashmi, Sahibzada Adil Shahzad, Chia-Wen Lin, Yu Tsao, and Hsin-Min Wang. Unmasking illusions: Understanding human perception of audiovisual deepfakes. ArXiv, abs/2405.04097, 2024. 1 +[24] Baojin Huang, Zhongyuan Wang, Jifan Yang, Jiaxin Ai, Qin Zou, Qian Wang, and Dengpan Ye. Implicit identity + +driven deepfake face swapping detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4490-4499, 2023. 2 +[25] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1 +[26] Marek Kowalski. Faceswap. https://github.com/MarekKowalski/FaceSwap, 2018. GitHub repository, accessed on March 3, 2025. 3, 4, 6 +[27] Nicolas Larue, Ngoc-Son Vu, Vitomir Struc, Peter Peer, and Vassilis Christophides. Seeable: Soft discrepancies and bounded contrastive learning for exposing deepfakes. In Proceedings of the IEEE/CVF international conference on computer vision, pages 21011-21021, 2023. 1, 3, 6, 7 +[28] Hong Li, Yutang Feng, Song Xue, Xuhui Liu, Bohan Zeng, Shanglin Li, Boyu Liu, Jianzhuang Liu, Shumin Han, and Baochang Zhang. Uv-idm: Identity-conditioned latent diffusion model for face uv-texture generation. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10585-10595, 2024. 1 +[29] Hanzhe Li, Jiaran Zhou, Yuezun Li, Baoyuan Wu, Bin Li, and Junyu Dong. Freqblender: Enhancing deepfake detection by blending frequency knowledge. arXiv preprint arXiv:2404.13872, 2024. 2 +[30] Lingzhi Li, Jianmin Bao, Ting Zhang, Hao Yang, Dong Chen, Fang Wen, and Baining Guo. Face x-ray for more general face forgery detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5001-5010, 2020. 1, 2, 3, 6 +[31] Yuezun Li and Siwei Lyu. Exposing deepfake videos by detecting face warping artifacts. arXiv preprint arXiv:1811.00656, 2018. 3 +[32] Yuezun Li, Ming-Ching Chang, and Siwei Lyu. In ictu oculi: Exposing ai created fake videos by detecting eye blinking. In 2018 IEEE International Workshop on Information Forensics and Security (WIFS), pages 1-7, 2018. 1, 2 +[33] Yuezun Li, Xin Yang, Pu Sun, Honggang Qi, and Siwei Lyu. Celeb-df: A large-scale challenging dataset for deepfake forensics. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3204-3213, 2020. 6, 7 +[34] Yuzhen Lin, Wentang Song, Bin Li, Yuezun Li, Jiangqun Ni, Han Chen, and Qiushi Li. Fake it till you make it: Curricular dynamic forgery augmentations towards general deepfake detection. In European Conference on Computer Vision, pages 104-122. Springer, 2024. 3 +[35] Dat NGUYEN, Nesryne MEJRI, Inder Pal SINGH, Polina KULESHOVA, Marcella ASTRID, Anis KACEM, Enjie GHORBEL, and Djamila AOUADA. Laa-net: Localized artifact attention network for quality-agnostic and generalizable deepfake detection. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17395–17405, 2024. 1, 2 +[36] Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards universal fake image detectors that generalize across generative models. In Proceedings of the IEEE/CVF Conference + +on Computer Vision and Pattern Recognition, pages 24480-24489, 2023. 2 +[37] Junting Pan, Ziyi Lin, Xiatian Zhu, Jing Shao, and Hongsheng Li. St-adapter: Parameter-efficient image-to-video transfer learning. Advances in Neural Information Processing Systems, 35:26462-26477, 2022. 3, 6 +[38] Guilin Pang, Baopeng Zhang, Zhu Teng, Zige Qi, and Jianping Fan. Mre-net: Multi-rate excitation network for deepfake video detection. IEEE Transactions on Circuits and Systems for Video Technology, 33(8):3663-3676, 2023. 2 +[39] Yuyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. Thinking in frequency: Face forgery detection by mining frequency-aware clues. In European conference on computer vision, pages 86-103. Springer, 2020. 2 +[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 2, 3, 6 +[41] Md Shohel Rana, Mohammad Nur Nobi, Bedduhu Murali, and Andrew H Sung. Deepfake detection: A systematic literature review. IEEE access, 10:25494-25513, 2022. 3, 4, 6, 7 +[42] Meitar Ronen, Shahaf E. Finder, and Oren Freifeld. Deep-dpm: Deep clustering with an unknown number of clusters. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9861-9870, 2022. 4 +[43] Andreas Rössler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Niessner. Faceforensics++: Learning to detect manipulated facial images. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1-11, 2019. 3, 6 +[44] Kaede Shiohara and Toshihiko Yamasaki. Detecting deepfakes with self-blended images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18720-18729, 2022. 1, 2, 3 +[45] Ke Sun, Taiping Yao, Shen Chen, Shouhong Ding, Jilin Li, and Rongrong Ji. Dual contrastive learning for general face forgery detection. In Proceedings of the AAAI conference on artificial intelligence, pages 2316-2324, 2022. 2, 6, 7 +[46] Justus Thies, Michael Zollhofer, Marc Stamminger, Christian Theobalt, and Matthias Nießner. Face2face: Real-time face capture and reenactment of rgb videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2387-2395, 2016. 1, 3, 4, 6 +[47] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. Acm Transactions on Graphics (TOG), 38(4):1-12, 2019. 1, 3, 4, 6 +[48] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9 (86):2579-2605, 2008. 8 +[49] Junke Wang, Zuxuan Wu, Wenhao Ouyang, Xintong Han, Jingjing Chen, Yu-Gang Jiang, and Ser-Nam Li. M2tr: Multi-modal multi-scale transformers for deepfake detec + +tion. In Proceedings of the 2022 international conference on multimedia retrieval, pages 615-623, 2022. 2 +[50] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnn-generated images are surprisingly easy to spot... for now. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8695-8704, 2020. 3 +[51] Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, and Houqiang Li. Altfreezing for more general video face forgery detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4129-4138, 2023. 2, 3, 6, 7 +[52] Yuting Xu, Jian Liang, Gengyun Jia, Ziming Yang, Yanhao Zhang, and Ran He. Tall: Thumbnail layout for deepfake video detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22658-22668, 2023. 6, 7 +[53] Zhiyuan Yan, Yong Zhang, Yanbo Fan, and Baoyuan Wu. Ucf: Uncovering common features for generalizable deepfake detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22412-22423, 2023. 6, 7 +[54] Zhiyuan Yan, Yuhao Luo, Siwei Lyu, Qingshan Liu, and Baoyuan Wu. Transcending forgery specificity with latent space augmentation for generalizable deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8984-8994, 2024. 1, 2, 6, 7 +[55] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications. ACM Computing Surveys, 56(4): 1-39, 2023. 1 +[56] Wenyuan Yang, Xiaoyu Zhou, Zhikai Chen, Bofei Guo, Zhongjie Ba, Zhihua Xia, Xiaochun Cao, and Kui Ren. Avoid-df: Audio-visual joint learning for detecting deepfake. IEEE Transactions on Information Forensics and Security, 18:2015–2029, 2023. 2 +[57] Shenghai Yuan, Jina Huang, Xianyi He, Yunyuan Ge, Yu-jun Shi, Liuhan Chen, Jiebo Luo, and Li Yuan. Identity-preserving text-to-video generation by frequency decomposition, 2024. 1 +[58] Yu Zeng, Zhe L. Lin, and Vishal M. Patel. Sketched: Mask-free local image manipulation with partial sketches. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5941-5951, 2021. 1 +[59] Daichi Zhang, Zihao Xiao, Shikun Li, Fanzhao Lin, Jianmin Li, and Shiming Ge. Learning natural consistency representation for face forgery video detection. In European Conference on Computer Vision, pages 407-424. Springer, 2024. 6 +[60] Tianchen Zhao, Xiang Xu, Mingze Xu, Hui Ding, Yuanjun Xiong, and Wei Xia. Learning self-consistency for deepfake detection. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 15003-15013, 2021. 6 +[61] Yinglin Zheng, Jianmin Bao, Dong Chen, Ming Zeng, and Fang Wen. Exploring temporal coherence for more general video face forgery detection. In Proceedings of the + +IEEE/CVF international conference on computer vision, pages 15044-15054, 2021. 2, 3, 6 +[62] Yang Zhou, Zichong Chen, and Hui Huang. Deformable oneshot face stylization via dino semantic guidance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7787-7796, 2024. 1 +[63] Wanyi Zhuang, Qi Chu, Zhentao Tan, Qiankun Liu, Haojie Yuan, Changtao Miao, Zixiang Luo, and Nenghai Yu. Uiavit: Unsupervised inconsistency-aware method based on vision transformer for face forgery detection. In European conference on computer vision, pages 391-407. Springer, 2022. 6 \ No newline at end of file diff --git a/2025/FakeRadar_ Probing Forgery Outliers to Detect Unknown Deepfake Videos/images.zip b/2025/FakeRadar_ Probing Forgery Outliers to Detect Unknown Deepfake Videos/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cfa6d5b158918d255ae9b44ce0e28eee10d2dd1e --- /dev/null +++ b/2025/FakeRadar_ Probing Forgery Outliers to Detect Unknown Deepfake Videos/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:759d550b89957b2eb3d049ce792134b631c106c3b5bf1fb56cb4ff80e164c33f +size 447567 diff --git a/2025/FakeRadar_ Probing Forgery Outliers to Detect Unknown Deepfake Videos/layout.json b/2025/FakeRadar_ Probing Forgery Outliers to Detect Unknown Deepfake Videos/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..faca8e072dbc2ebf59fc1bff918b6d6b07954220 --- /dev/null +++ b/2025/FakeRadar_ Probing Forgery Outliers to Detect Unknown Deepfake Videos/layout.json @@ -0,0 +1,8680 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 73, + 103, + 538, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 103, + 538, + 121 + ], + "spans": [ + { + "bbox": [ + 73, + 103, + 538, + 121 + ], + "type": "text", + "content": "FakeRadar: Probing Forgery Outliers to Detect Unknown Deepfake Videos" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 143, + 142, + 466, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 142, + 466, + 171 + ], + "spans": [ + { + "bbox": [ + 143, + 142, + 466, + 171 + ], + "type": "text", + "content": "Zhaolun Li" + }, + { + "bbox": [ + 143, + 142, + 466, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 143, + 142, + 466, + 171 + ], + "type": "text", + "content": " Jichang Li" + }, + { + "bbox": [ + 143, + 142, + 466, + 171 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 143, + 142, + 466, + 171 + ], + "type": "text", + "content": " Yinqi Cai" + }, + { + "bbox": [ + 143, + 142, + 466, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 143, + 142, + 466, + 171 + ], + "type": "text", + "content": " Junye Chen" + }, + { + "bbox": [ + 143, + 142, + 466, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 143, + 142, + 466, + 171 + ], + "type": "text", + "content": " Xiaonan Luo" + }, + { + "bbox": [ + 143, + 142, + 466, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 143, + 142, + 466, + 171 + ], + "type": "text", + "content": " Guanbin Li" + }, + { + "bbox": [ + 143, + 142, + 466, + 171 + ], + "type": "inline_equation", + "content": "^{3,2,5}" + }, + { + "bbox": [ + 143, + 142, + 466, + 171 + ], + "type": "text", + "content": " Rushi Lan" + }, + { + "bbox": [ + 143, + 142, + 466, + 171 + ], + "type": "inline_equation", + "content": "^{1,4*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 71, + 171, + 539, + 197 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 171, + 539, + 197 + ], + "spans": [ + { + "bbox": [ + 71, + 171, + 539, + 197 + ], + "type": "text", + "content": "1Guilin University of Electronic Technology 2Pengcheng Laboratory 3Sun Yat-sen University 4Guangxi Key Laboratory of Image and Graphic Intelligent Processing" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 198, + 465, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 198, + 465, + 213 + ], + "spans": [ + { + "bbox": [ + 143, + 198, + 465, + 213 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 143, + 198, + 465, + 213 + ], + "type": "text", + "content": "Guangdong Key Laboratory of Big Data Analysis and Processing" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 119, + 216, + 488, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 216, + 488, + 227 + ], + "spans": [ + { + "bbox": [ + 119, + 216, + 488, + 227 + ], + "type": "text", + "content": "zhaolunli@mails.guet.edu.cn, li.jichang@pcl.ac.cn, rslan2016@163.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "spans": [ + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 280, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 280, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 280, + 296, + 579 + ], + "type": "text", + "content": "In this paper, we propose FakeRadar, a novel deepfake video detection framework designed to address the challenges of cross-domain generalization in real-world scenarios. Existing detection methods typically rely on manipulation-specific cues, performing well on known forgery types but exhibiting severe limitations against emerging manipulation techniques. This poor generalization stems from their inability to adapt effectively to unseen forgery patterns. To overcome this, we leverage large-scale pretrained models (e.g. CLIP) to proactively probe the feature space, explicitly highlighting distributional gaps between real videos, known forgeries, and unseen manipulations. Specifically, FakeRadar introduces Forgery Outlier Probing, which employs dynamic subcluster modeling and cluster-conditional outlier generation to synthesize outlier samples near boundaries of estimated subclusters, simulating novel forgery artifacts beyond known manipulation types. Additionally, we design Outlier-Guided Tri-Training, which optimizes the detector to distinguish real, fake, and outlier samples using proposed outlier-driven contrastive learning and outlier-conditioned cross-entropy losses. Experiments show that FakeRadar outperforms existing methods across various benchmark datasets for deepfake video detection, particularly in cross-domain evaluations, by handling the variety of emerging manipulation techniques." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 602, + 135, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 602, + 135, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 602, + 135, + 615 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 622, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 622, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 622, + 296, + 696 + ], + "type": "text", + "content": "Recently, generative techniques [28, 58, 62] for creating deepfake videos have advanced rapidly, producing highly realistic facial forgeries that challenge human perceptual limits [8, 23]. However, the malicious use of such technology, such as fabricated news and identity fraud, poses a significant threat to digital media credibility [7, 9]. Ear" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 253, + 555, + 417 + ], + "blocks": [ + { + "bbox": [ + 316, + 253, + 555, + 417 + ], + "lines": [ + { + "bbox": [ + 316, + 253, + 555, + 417 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 555, + 417 + ], + "type": "image", + "image_path": "d94756103d0736965c074e381e3861971875cef335c582a411d06cbb35a5cddb.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 422, + 555, + 521 + ], + "lines": [ + { + "bbox": [ + 313, + 422, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 555, + 521 + ], + "type": "text", + "content": "Figure 1. Conceptual pipeline of FakeRadar. (a)-(b) Forgery Outlier Probing first performs dynamic subcluster modeling on these features to partition finer-grained subclusters, then generates cluster-conditional outliers for each estimated subcluster to simulate unseen forgeries, emulating unknown manipulations. (c)-(d) The model is optimized using outlier-driven tri-training, allowing the detector to establish optimal boundaries for the real, fake, and outlier classes, with the latter two merged into a new \"Fake\" class during testing." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "text", + "content": "lier deepfake detection methods primarily rely on the passive learning of specific forgery artifacts, such as boundary inconsistencies [27, 54], blinking anomalies [32], and texture irregularities [17, 35]. While these approaches perform well against conventional deepfake manipulations, including Face2Face [46] and NeuralTextures [47], they exhibit critical vulnerabilities when confronted with emerging generative paradigms, such as diffusion models [55]. For instance, boundary-based detection methods [30, 44] fail to identify faces generated by advanced GANs (e.g., StyleGAN [25]). Similarly, blink analysis-based techniques [32] are rendered ineffective when processing temporally smoothed fake videos [57]. These limitations stem from a fundamental shortcoming: existing methods learn a constrained decision boundary in the feature space, where" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 149, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 149, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 149, + 712 + ], + "type": "text", + "content": "*Corresponding Authors." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13382" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "known forgery traces occupy only a localized region, while novel manipulation techniques introduce artifacts in arbitrary locations. As a result, they struggle to generalize to new manipulation techniques that introduce artifacts in previously unseen ways, leading to significant performance degradation in cross-domain scenarios [11, 12]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 147, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 147, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 56, + 147, + 296, + 445 + ], + "type": "text", + "content": "To better understand its underlying cause, [2, 36] have analyzed the feature distributions of real and synthetic data. These studies reveal a fundamental embedding-space disparity: real data exhibit compact, dense distributions, whereas deepfake forgeries form scattered, sparse clusters due to their generation with diverse manipulation types, as demonstrated in Figure 1. This difference in distribution poses a significant challenge for existing detection methods, which are typically optimized to classify previously observed forgeries. Since deepfake forgeries do not adhere to a fixed pattern but rather introduce an infinite variety of modifications, classifiers trained on known forgery patterns struggle to recognize novel deepfakes that deviate from past observations. Therefore, a robust detection approach must move beyond merely learning pre-defined forgery patterns and instead develop the capability to identify anomalies that fall outside of previously known distributions. Given the limitations of existing methods in handling unseen forgery types, we propose a paradigm shift in deepfake detection. Instead of passively fitting to known forgery patterns, our approach proactively explores potential anomaly regions in the feature space. Inspired by radar systems that scan for unknown targets through frequency-spectrum probing, our method enhances detection robustness against novel deep-fake manipulations." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 448, + 295, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 448, + 295, + 675 + ], + "spans": [ + { + "bbox": [ + 55, + 448, + 295, + 675 + ], + "type": "text", + "content": "In this work, we propose a new deepfake video detection framework, called FakeRadar, which leverages the deep feature priors of large-scale pre-trained models (i.e. CLIP [40]) to proactively \"probe\" unknown forgeries and enhance cross-domain generalization. Unlike conventional detectors [2, 5, 51] that perform only binary classification, FakeRadar explicitly introduces \"Forgery outliers\" during training to simulate potential unknown forgeries and thus \"pre-explore and cover\" the latent space. As shown in Figure 1 (a) and (b), we first employ Forgery Outlier Probing, which conducts dynamic subcluster modeling on the features of real and known deepfake videos. By approximating the embedding space via fine-grained Gaussian mixture distributions, this algorithm automatically partitions training samples into distinct subclusters. Next, through cluster-conditional outlier generation, Forgery Outlier Probing synthesizes outlier samples near the boundaries of these estimated subclusters with a certain \"novel offset\", thereby covering a broader range of potential forgery regions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "Additionally, FakeRadar adopts a model optimization strategy called Outlier-Guided Tri-Training, enabling the model to discriminate among \"Real\", \"Fake\", and \"Out" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "content": "lier” samples, with its details as illustrated in Figure 1 (c) and (d). Specifically, Outlier-Guided Tri-Training proposes an outlier-driven contrastive loss and an outlier-conditioned cross-entropy loss for end-to-end model optimization. The former loss enhances the distance between real, fake, and outlier features via contrastive learning, emphasizing their differences, while the latter one ensures that the model has clear decision boundaries for the three categories, especially ensuring that outliers are not incorrectly classified as real. In this triplet-class (Real/Fake/Outlier) setting, the model can independently label unknown forgeries during training. During inference, both “Fake” and “Outlier” are classified as forgeries, empowering the detector to adapt effectively to unknown deepfakes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 239, + 553, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 553, + 263 + ], + "type": "text", + "content": "We summarize FakeRadar's principal contributions as follows, with its conceptual pipeline illustrated in Figure 1:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 264, + 553, + 443 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 264, + 553, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 264, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 314, + 264, + 553, + 299 + ], + "type": "text", + "content": "- We propose FakeRadar, a novel framework for deepfake video detection, designed to address the critical challenge of cross-domain generalization in real-world scenarios." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 300, + 553, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 300, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 314, + 300, + 553, + 346 + ], + "type": "text", + "content": "- We introduce Forgery Outlier Probing for synthesizing outlier samples to cover a broader range of potential forger patterns, expanding the detector's understanding of forgery distribution." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 348, + 553, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 348, + 553, + 394 + ], + "spans": [ + { + "bbox": [ + 314, + 348, + 553, + 394 + ], + "type": "text", + "content": "- We design Outlier-Guided Tri-Training for model optimization, which enables end-to-end training for distinguishing real, fake, and outlier samples, improving detection of unknown forgeries." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 396, + 553, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 396, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 314, + 396, + 553, + 443 + ], + "type": "text", + "content": "- Numerous experiments have revealed that FakeRadar outperforms current leading algorithms across multiple deepfake benchmarks, particularly excelling in cross-domain evaluations where existing methods struggle." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 454, + 399, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 454, + 399, + 466 + ], + "spans": [ + { + "bbox": [ + 314, + 454, + 399, + 466 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 474, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 553, + 713 + ], + "type": "text", + "content": "Deepfake Detection. In response to the malicious use of deepfake techniques, numerous approaches have been explored to address this issue. Generally, deepfake detection is formulated as a binary classification problem (real or fake), leading researchers to develop a variety of end-to-end detectors [45, 54] that directly distinguish reals from manipulated content. Existing detection methods can be broadly categorized into two groups: image-level and video-level detection. Specifically, image-level detectors primarily focus on identifying spatial artifacts caused by forgeries. Examples include frequency anomalies [3, 16, 29, 39], identity-based discrepancies [13, 14, 24], blending artifacts [30, 44], and localized distortions [35, 49]. In contrast, video-level detectors employ a wider range of strategies. For example, some algorithms, such as [20, 32], concentrate on temporal consistency, typically reflected in correlations across consecutive frames. Other video detection studies [18, 38, 61] incorporate both temporal and spatiotemporal features to establish a more comprehensive detection framework. Another line of research [22, 56] examines audio-visual con" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13383" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 69, + 553, + 220 + ], + "blocks": [ + { + "bbox": [ + 62, + 69, + 553, + 220 + ], + "lines": [ + { + "bbox": [ + 62, + 69, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 62, + 69, + 553, + 220 + ], + "type": "image", + "image_path": "586c8307ffd116b23c397ed996a152fff5e4080e49e37e8fbc482682fa729c41.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 225, + 555, + 314 + ], + "lines": [ + { + "bbox": [ + 55, + 225, + 555, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 225, + 555, + 314 + ], + "type": "text", + "content": "Figure 2. An overview of FakeRadar for deepfake video detection. FakeRadar consists of two key components: Forgery Outlier Probing and Outlier-Guided Tri-Training. (i) Forgery Outlier Probing performs dynamic subcluster modeling to model the feature-space distributions of real and fake samples using GMMs, followed by dynamic subcluster splitting and merging to produce finer-grained subcluster partitions. Afterwards, FakeRadar conducts cluster-conditioned outlier generation to simulate unseen forgeries for novel manipulation techniques. (ii) Outlier-Guided Tri-Training jointly optimizes the CLIP model's ViT-B/16 backbone, inserted with ST-Adapter [37] for parameter-efficient fine tuning, and a triplet-class classifier to distinguish between \"Real\", \"Fake\", and \"Outlier\" classes, with \"Fake\" and \"Outlier\" merged into a single new \"Fake\" class during evaluation. The training set includes real videos, as well as deepfake videos generated using four different manipulations, i.e. \"DF\" [41], \"FS\" [26], \"NT\" [47] and \"F2F\" [46]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 324, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 296, + 407 + ], + "type": "text", + "content": "sistency by analyzing the synchronization and correlation between video and audio signals to detect forgeries. In summary, these methods rely heavily on existing artifacts in training set, limiting their generalization to unseen manipulation types. In contrast, our approach explores feature-space anomalies, enhancing the generalization capacity of our proposed detector to diverse deepfake manipulations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 426, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 616 + ], + "type": "text", + "content": "Towards Generalization in Deepfake Detection. Rapid advance in deepfakes have significantly improved the quality and authenticity of generated content, posing substantial challenges to the generalization capability of detection systems. Existing detectors [6, 43, 50] often suffer from severe performance degradation when encountering novel forgery techniques or previously unseen manipulations. To address this issue, researchers have proposed various strategies to enhance detector generalization. For instance, FTCN [61] improves generalization by capturing temporal coherence features in video sequences. Similarly, AltFreezing [51] employs an alternating weight-freezing training strategy to facilitate the joint detection of spatial and temporal forgery artifacts. Another algorithm, StyleFlow [5], leverages the temporal dynamics of style latent vectors to identify anomalous patterns in generated videos." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": "More recently, data synthesis has emerged as an effective strategy for improving detector generalization. For example, FWA [31] enhances training data by introducing facial warping artifacts, while Face X-ray [30] detects forgeries by identifying blending boundaries, independent of specific facial manipulation techniques. Building on this concept, SBI [44] synthesizes training samples by blending different views of the same pristine image to simulate generic forgery" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 324, + 555, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 555, + 397 + ], + "type": "text", + "content": "artifacts. Additionally, a series of recent studies [1, 27, 34] have further advanced this direction by employing various data synthesis techniques. Unlike existing methods that primarily focus on image- or video-level data synthesis for deepfake detection, our approach proposes to explore unseen forgeries in feature space." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 407, + 397, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 407, + 397, + 422 + ], + "spans": [ + { + "bbox": [ + 314, + 407, + 397, + 422 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 427, + 400, + 439 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 427, + 400, + 439 + ], + "spans": [ + { + "bbox": [ + 313, + 427, + 400, + 439 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 445, + 555, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 555, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 555, + 602 + ], + "type": "text", + "content": "Overview. In this paper, we introduce FakeRadar, to overcome the challenges of cross-domain generalization in deepfake video detection by exploring the feature space to identify emerging manipulation patterns. Specifically, FakeRadar is built upon two key components: (i) Forgery Outlier Probing, which models the feature-space distributions of subclusters for training samples of both real and fake categories, and then generates cluster-conditioned outlier samples to simulate unseen forgeries, and (ii) Outlier-Guided Tri-Training, which jointly optimizes an adapter-inserted backbone and a triplet-class classifier to distinguish among samples from the \"Real\", \"Fake\", and \"Outlier\" classes. An overview of FakeRadar can be found in Figure 2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "Model Adaptation with a Frozen Backbone. The FakeRadar framework employs a pre-trained Vision Transformer (ViT-B) from CLIP [40] as a frozen backbone to preserve its learned semantic features. This model serves as the video encoder for input clips. To adapt it for deepfake detection, we integrate lightweight ST-Adapter layers [37] after each Transformer block, enabling parameter-efficient fine-tuning. This significantly enhances the model's ability to" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13384" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": "capture spatial manipulations and temporal inconsistencies, which are crucial for detecting deepfake videos. Given a feature vector " + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": " from the video encoder, each adapter transforms it as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 130, + 295, + 144 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 130, + 295, + 144 + ], + "spans": [ + { + "bbox": [ + 61, + 130, + 295, + 144 + ], + "type": "interline_equation", + "content": "\\operatorname {S T - A d p a t e r} (x) = x + \\operatorname {R e L U} (\\operatorname {C o n v 3 D} (x W _ {\\text {d o w n}})) W _ {\\text {u p}}. \\tag {1}", + "image_path": "d953b042e3b9faf6fbb7005c3c1280c70e9c643d962d503a953ba3d6e25f2bcd.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 153, + 296, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 153, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 153, + 296, + 274 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 153, + 296, + 274 + ], + "type": "inline_equation", + "content": "\\mathrm{Conv3D}(\\cdot)" + }, + { + "bbox": [ + 55, + 153, + 296, + 274 + ], + "type": "text", + "content": " denotes a 3D convolution operation employed for capturing spatio-temporal features from videos. As well, " + }, + { + "bbox": [ + 55, + 153, + 296, + 274 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{down}}" + }, + { + "bbox": [ + 55, + 153, + 296, + 274 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 153, + 296, + 274 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{up}}" + }, + { + "bbox": [ + 55, + 153, + 296, + 274 + ], + "type": "text", + "content": " are weight matrices that perform dimensionality reduction and expansion, respectively, and " + }, + { + "bbox": [ + 55, + 153, + 296, + 274 + ], + "type": "inline_equation", + "content": "\\mathrm{ReLU}(\\cdot)" + }, + { + "bbox": [ + 55, + 153, + 296, + 274 + ], + "type": "text", + "content": " is the activation function. By introducing only a small number of additional parameters, these adapters allow efficient fine-tuning while retaining the original backbone's representational strength. After feature extraction, each video clip is represented as a " + }, + { + "bbox": [ + 55, + 153, + 296, + 274 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 153, + 296, + 274 + ], + "type": "text", + "content": "-dimensional vector for subsequent analysis, with " + }, + { + "bbox": [ + 55, + 153, + 296, + 274 + ], + "type": "inline_equation", + "content": "d = 768" + }, + { + "bbox": [ + 55, + 153, + 296, + 274 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 289, + 296, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 296, + 410 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 296, + 410 + ], + "type": "text", + "content": "Inference. During inference, the model classifies input samples as either real or deepfake, with deepfakes categorized as \"Fake\" or \"Outlier\". The \"Fake\" class includes fake samples with known manipulation patterns, while the \"Outlier\" class detects testing samples with novel forgeries that differ from established manipulations. By unifying these categories, the model enhances its robustness against evolving manipulation techniques, enabling accurate detection of both familiar and emerging deepfakes. This improves its generalization performance in cross-domain evaluations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 416, + 194, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 416, + 194, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 416, + 194, + 430 + ], + "type": "text", + "content": "3.2. Forgery Outlier Probing" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 434, + 296, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 296, + 590 + ], + "type": "text", + "content": "In this section, we introduce Forgery Outlier Probing (FOP) to generate outlier samples with unseen forgeries beyond the existing real and fake videos, simulating previously unobserved manipulation patterns. Specifically, FOP first presents dynamic subcluster modeling, which performs category-specific clustering on the training samples of real and fake categories. Then, it models the distributions of these categories using a strategy based on Gaussian Mixture Models (GMMs), and follows by applying dynamic subcluster splitting and merging, thereby achieving finer-grained subcluster partitioning. Finally, we generate cluster-conditioned outliers based on the resulting subclusters to stimulate unseen forgeries." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": "Dynamic Subcluster Modeling. To capture subtle intra-class variations within real and deepfake videos, we apply category-specific clustering on the FaceForensics++ (FF++) dataset. Specifically, we treat the real class and each of the four fake classes (i.e. four types of deepfake videos created by different manipulation techniques [26, 41, 46, 47]) as five separate categories. For each category " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": ", let " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{X}^{(c)} = \\{\\mathbf{x}_1, \\mathbf{x}_2, \\dots, \\mathbf{x}_N\\}" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": " denote the set of feature embeddings (i.e. extracted from the video encoder). Inspired by [42], we then" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 72, + 451, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 451, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 451, + 84 + ], + "type": "text", + "content": "define a main clustering network:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 352, + 91, + 555, + 107 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 91, + 555, + 107 + ], + "spans": [ + { + "bbox": [ + 352, + 91, + 555, + 107 + ], + "type": "interline_equation", + "content": "f _ {\\operatorname {m a i n}} ^ {(c)} \\left(\\mathbf {x} _ {i}\\right) \\mapsto \\mathbf {r} _ {i} = \\left(r _ {i, 1}, r _ {i, 2}, \\dots , r _ {i, K}\\right), \\tag {2}", + "image_path": "a8e40497ec950d356d21d0144ace802619ea918586767222e30da58a338088ea.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 113, + 555, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 113, + 555, + 163 + ], + "spans": [ + { + "bbox": [ + 313, + 113, + 555, + 163 + ], + "type": "text", + "content": "which initially assumes " + }, + { + "bbox": [ + 313, + 113, + 555, + 163 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 113, + 555, + 163 + ], + "type": "text", + "content": " clusters, empirically with " + }, + { + "bbox": [ + 313, + 113, + 555, + 163 + ], + "type": "inline_equation", + "content": "K = 5" + }, + { + "bbox": [ + 313, + 113, + 555, + 163 + ], + "type": "text", + "content": ", providing a \"soft\" assignment of each embedding " + }, + { + "bbox": [ + 313, + 113, + 555, + 163 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i \\in \\mathcal{X}^{(c)}" + }, + { + "bbox": [ + 313, + 113, + 555, + 163 + ], + "type": "text", + "content": " to these " + }, + { + "bbox": [ + 313, + 113, + 555, + 163 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 113, + 555, + 163 + ], + "type": "text", + "content": " clusters. Here, " + }, + { + "bbox": [ + 313, + 113, + 555, + 163 + ], + "type": "inline_equation", + "content": "r_{i,k} \\in [0,1]" + }, + { + "bbox": [ + 313, + 113, + 555, + 163 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 113, + 555, + 163 + ], + "type": "inline_equation", + "content": "\\sum_{k=1}^{K} r_{i,k} = 1" + }, + { + "bbox": [ + 313, + 113, + 555, + 163 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 163, + 555, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 163, + 555, + 294 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 555, + 294 + ], + "type": "text", + "content": "Each cluster can be modeled as a Gaussian distribution using a GMM, allowing us to represent the feature distribution of each category in a probabilistic manner. Specifically, for each cluster " + }, + { + "bbox": [ + 313, + 163, + 555, + 294 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 163, + 555, + 294 + ], + "type": "text", + "content": " in category " + }, + { + "bbox": [ + 313, + 163, + 555, + 294 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 163, + 555, + 294 + ], + "type": "text", + "content": ", we define a Gaussian component characterized by " + }, + { + "bbox": [ + 313, + 163, + 555, + 294 + ], + "type": "inline_equation", + "content": "\\{\\pi_k,\\pmb {\\mu}_k,\\pmb {\\Sigma}_k\\}" + }, + { + "bbox": [ + 313, + 163, + 555, + 294 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 163, + 555, + 294 + ], + "type": "inline_equation", + "content": "\\pi_{k}\\in (0,1)" + }, + { + "bbox": [ + 313, + 163, + 555, + 294 + ], + "type": "text", + "content": " is the mixture weight for cluster " + }, + { + "bbox": [ + 313, + 163, + 555, + 294 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 163, + 555, + 294 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 313, + 163, + 555, + 294 + ], + "type": "inline_equation", + "content": "\\sum_{k = 1}^{K}\\pi_{k} = 1" + }, + { + "bbox": [ + 313, + 163, + 555, + 294 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 313, + 163, + 555, + 294 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_{k}" + }, + { + "bbox": [ + 313, + 163, + 555, + 294 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 163, + 555, + 294 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}_{k}" + }, + { + "bbox": [ + 313, + 163, + 555, + 294 + ], + "type": "text", + "content": " represent its corresponding mean and covariance matrix. These parameters enable the clustering network to estimate the likelihood of each data sample belonging to different clusters, aligning our model with a GMM-based formulation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 296, + 554, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 296, + 554, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 296, + 554, + 331 + ], + "type": "text", + "content": "After that, the main clustering network needs to learn these parameters by aligning its soft assignments with GMM responsibilities:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 359, + 338, + 555, + 370 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 338, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 359, + 338, + 555, + 370 + ], + "type": "interline_equation", + "content": "r _ {i, k} ^ {E} = \\frac {\\pi_ {k} \\mathcal {N} \\left(\\mathbf {x} _ {i} ; \\boldsymbol {\\mu} _ {k} , \\boldsymbol {\\Sigma} _ {k}\\right)}{\\sum_ {k ^ {\\prime} = 1} ^ {K} \\pi_ {k ^ {\\prime}} \\mathcal {N} \\left(\\mathbf {x} _ {i} ; \\boldsymbol {\\mu} _ {k ^ {\\prime}} , \\boldsymbol {\\Sigma} _ {k ^ {\\prime}}\\right)}, \\tag {3}", + "image_path": "cffae684957c0578dbc45f58c93a42ac6045e3d39864f0445e1c25ffb5812a10.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 374, + 554, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 554, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 554, + 408 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 374, + 554, + 408 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\cdot ;\\pmb {\\mu}_k,\\pmb {\\Sigma}_k)" + }, + { + "bbox": [ + 313, + 374, + 554, + 408 + ], + "type": "text", + "content": " is the multivariate Gaussian density function. The network is trained using the KL-divergence loss:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 378, + 408, + 554, + 440 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 408, + 554, + 440 + ], + "spans": [ + { + "bbox": [ + 378, + 408, + 554, + 440 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m a i n}} = \\sum_ {i = 1} ^ {N} \\mathrm {K L} \\left(\\mathbf {r} _ {i} \\| \\mathbf {r} _ {i} ^ {E}\\right), \\tag {4}", + "image_path": "939c9242f98e5f9ca447f5d70a8973054c2749cc8207842428ed0ef6abf29e85.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 443, + 554, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 443, + 554, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 554, + 466 + ], + "type": "text", + "content": "which ensures that the learned cluster assignments match the inferred GMM responsibilities." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 467, + 554, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 554, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 554, + 538 + ], + "type": "text", + "content": "To further refine cluster structures, we introduce a subclustering network for each cluster " + }, + { + "bbox": [ + 313, + 467, + 554, + 538 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 467, + 554, + 538 + ], + "type": "text", + "content": ", which attempts to split the cluster into two subclusters if multi-modal structures are detected. A compactness loss is applied to encourage subclusters to be well-separated. This loss function penalizes subclusters that are too close to each other:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 349, + 545, + 554, + 580 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 545, + 554, + 580 + ], + "spans": [ + { + "bbox": [ + 349, + 545, + 554, + 580 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {s u b}} = \\sum_ {k = 1} ^ {K} \\sum_ {j = 1} ^ {2} \\sum_ {\\mathbf {x} _ {i} \\in X _ {k}} \\tilde {r} _ {i, j} \\| \\mathbf {x} _ {i} - \\tilde {\\boldsymbol {\\mu}} _ {k, j} \\| _ {2} ^ {2}, \\tag {5}", + "image_path": "74897c06dff9ec9d89932a4f422971a6d43e497f5601dc156819baa72ad8d1d0.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 585, + 554, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 585, + 554, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 585, + 554, + 609 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 585, + 554, + 609 + ], + "type": "inline_equation", + "content": "\\tilde{r}_{i,j}" + }, + { + "bbox": [ + 313, + 585, + 554, + 609 + ], + "type": "text", + "content": " represents the probability of " + }, + { + "bbox": [ + 313, + 585, + 554, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 313, + 585, + 554, + 609 + ], + "type": "text", + "content": " belonging to subcluster " + }, + { + "bbox": [ + 313, + 585, + 554, + 609 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 585, + 554, + 609 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 585, + 554, + 609 + ], + "type": "inline_equation", + "content": "\\tilde{\\mu}_{k,j}" + }, + { + "bbox": [ + 313, + 585, + 554, + 609 + ], + "type": "text", + "content": " is its mean." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 609, + 554, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 554, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 554, + 681 + ], + "type": "text", + "content": "Since the predefined cluster number " + }, + { + "bbox": [ + 313, + 609, + 554, + 681 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 609, + 554, + 681 + ], + "type": "text", + "content": " may not be optimal, we propose to dynamically adjust it through split/merge proposals, inspired by [42]. Suppose a cluster " + }, + { + "bbox": [ + 313, + 609, + 554, + 681 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 609, + 554, + 681 + ], + "type": "text", + "content": " contains " + }, + { + "bbox": [ + 313, + 609, + 554, + 681 + ], + "type": "inline_equation", + "content": "N_{k}" + }, + { + "bbox": [ + 313, + 609, + 554, + 681 + ], + "type": "text", + "content": " points, " + }, + { + "bbox": [ + 313, + 609, + 554, + 681 + ], + "type": "inline_equation", + "content": "X_{k}" + }, + { + "bbox": [ + 313, + 609, + 554, + 681 + ], + "type": "text", + "content": ". We split if subclustering indicates that " + }, + { + "bbox": [ + 313, + 609, + 554, + 681 + ], + "type": "inline_equation", + "content": "X_{k}" + }, + { + "bbox": [ + 313, + 609, + 554, + 681 + ], + "type": "text", + "content": " is better viewed as two subgroups, " + }, + { + "bbox": [ + 313, + 609, + 554, + 681 + ], + "type": "inline_equation", + "content": "\\{X_{k,1}, X_{k,2}\\}" + }, + { + "bbox": [ + 313, + 609, + 554, + 681 + ], + "type": "text", + "content": ". The Hastings ratio is given by" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 331, + 687, + 554, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 687, + 554, + 717 + ], + "spans": [ + { + "bbox": [ + 331, + 687, + 554, + 717 + ], + "type": "interline_equation", + "content": "H _ {s} = \\frac {\\alpha \\Gamma (N _ {k , 1}) f _ {x} (X _ {k , 1}) \\Gamma (N _ {k , 2}) f _ {x} (X _ {k , 2})}{\\Gamma (N _ {k}) f _ {x} (X _ {k})}, \\tag {6}", + "image_path": "60e21b97494c541ea2da036b83dbdc4d71d7dc33bc102591633e4328d4a70dfe.jpg" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13385" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 107 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 107 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 294, + 107 + ], + "type": "inline_equation", + "content": "f_{x}(\\cdot)" + }, + { + "bbox": [ + 55, + 72, + 294, + 107 + ], + "type": "text", + "content": " is the marginal likelihood under a weak NIw prior, " + }, + { + "bbox": [ + 55, + 72, + 294, + 107 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 72, + 294, + 107 + ], + "type": "text", + "content": " is the Dirichlet-process concentration parameter, and " + }, + { + "bbox": [ + 55, + 72, + 294, + 107 + ], + "type": "inline_equation", + "content": "\\Gamma (\\cdot)" + }, + { + "bbox": [ + 55, + 72, + 294, + 107 + ], + "type": "text", + "content": " is the Gamma function." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 108, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 108, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 108, + 294, + 191 + ], + "type": "text", + "content": "If " + }, + { + "bbox": [ + 55, + 108, + 294, + 191 + ], + "type": "inline_equation", + "content": "H_{s}" + }, + { + "bbox": [ + 55, + 108, + 294, + 191 + ], + "type": "text", + "content": " exceeds a threshold, cluster " + }, + { + "bbox": [ + 55, + 108, + 294, + 191 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 108, + 294, + 191 + ], + "type": "text", + "content": " is split into two new clusters. The parameters for these new clusters are initialized from " + }, + { + "bbox": [ + 55, + 108, + 294, + 191 + ], + "type": "inline_equation", + "content": "\\tilde{\\mu}_{k,1},\\tilde{\\Sigma}_{k,1}" + }, + { + "bbox": [ + 55, + 108, + 294, + 191 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 108, + 294, + 191 + ], + "type": "inline_equation", + "content": "\\tilde{\\mu}_{k,2},\\tilde{\\Sigma}_{k,2}" + }, + { + "bbox": [ + 55, + 108, + 294, + 191 + ], + "type": "text", + "content": ", respectively. Merging is similarly proposed if two clusters significantly overlap, with acceptance ratio " + }, + { + "bbox": [ + 55, + 108, + 294, + 191 + ], + "type": "inline_equation", + "content": "H_{m} = 1 / H_{s}" + }, + { + "bbox": [ + 55, + 108, + 294, + 191 + ], + "type": "text", + "content": ". Repeating these steps yields an effective number of clusters " + }, + { + "bbox": [ + 55, + 108, + 294, + 191 + ], + "type": "inline_equation", + "content": "K^{(c)}" + }, + { + "bbox": [ + 55, + 108, + 294, + 191 + ], + "type": "text", + "content": ", plus subclusters for each category " + }, + { + "bbox": [ + 55, + 108, + 294, + 191 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 108, + 294, + 191 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 192, + 294, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 192, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 294, + 239 + ], + "type": "text", + "content": "After these dynamic adjustments, we obtain the final subcluster set " + }, + { + "bbox": [ + 55, + 192, + 294, + 239 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{C}} = \\{\\hat{C}_1,\\hat{C}_2,\\dots,\\hat{C}_{K^{(c)}}\\}" + }, + { + "bbox": [ + 55, + 192, + 294, + 239 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 192, + 294, + 239 + ], + "type": "inline_equation", + "content": "K^{(c)}" + }, + { + "bbox": [ + 55, + 192, + 294, + 239 + ], + "type": "text", + "content": " is the refined number of clusters. Each adjusted subcluster " + }, + { + "bbox": [ + 55, + 192, + 294, + 239 + ], + "type": "inline_equation", + "content": "\\hat{C}_k" + }, + { + "bbox": [ + 55, + 192, + 294, + 239 + ], + "type": "text", + "content": " is characterized by its updated mean and covariance:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 247, + 295, + 278 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 247, + 295, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 247, + 295, + 278 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {\\mu}} _ {k} = \\frac {1}{| \\hat {C} _ {k} |} \\sum_ {\\mathbf {x} _ {i} \\in \\hat {C} _ {k}} \\mathbf {x} _ {i}, \\tag {7}", + "image_path": "d9ffbda5d90b052211ddc05b46b2db73c932dfdb68d4e2d1291a246ee69b9b31.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 94, + 287, + 295, + 318 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 287, + 295, + 318 + ], + "spans": [ + { + "bbox": [ + 94, + 287, + 295, + 318 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {\\Sigma}} = \\frac {1}{| \\hat {C} _ {k} |} \\sum_ {\\mathbf {x} _ {i} \\in \\hat {C} _ {k}} \\left(\\mathbf {x} _ {i} - \\hat {\\boldsymbol {\\mu}} _ {k}\\right) \\left(\\mathbf {x} _ {i} - \\hat {\\boldsymbol {\\mu}} _ {k}\\right) ^ {\\top}. \\tag {8}", + "image_path": "f49c9c0d2996ab5acac63b806f59e1ccfaf51a023cbdd800db5dc19a2f6f34de.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 322, + 296, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 322, + 296, + 394 + ], + "spans": [ + { + "bbox": [ + 55, + 322, + 296, + 394 + ], + "type": "text", + "content": "We use online estimation for efficient training by maintaining a cluster-conditional queue with " + }, + { + "bbox": [ + 55, + 322, + 296, + 394 + ], + "type": "inline_equation", + "content": "|\\hat{C}_k|" + }, + { + "bbox": [ + 55, + 322, + 296, + 394 + ], + "type": "text", + "content": " sample instances from each subcluster. In each iteration, we enqueue the embeddings of objects to their corresponding cluster-conditional queues and dequeue the same number of object embeddings." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 408, + 296, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 408, + 296, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 296, + 515 + ], + "type": "text", + "content": "Cluster-Conditional Outlier Generation. After deriving finer-grained cluster partitions for each category, we generate feature-space outliers to probe potential unknown forgeries. Motivated by [15], in order to ensure that the sampled outliers lie near the cluster boundary, we maintain a sufficiently small value " + }, + { + "bbox": [ + 55, + 408, + 296, + 515 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 408, + 296, + 515 + ], + "type": "text", + "content": " for outlier sampling from the " + }, + { + "bbox": [ + 55, + 408, + 296, + 515 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 408, + 296, + 515 + ], + "type": "text", + "content": "-likelihood region of the estimated cluster-conditional distribution. Therefore, for a subcluster from " + }, + { + "bbox": [ + 55, + 408, + 296, + 515 + ], + "type": "inline_equation", + "content": "\\hat{C}" + }, + { + "bbox": [ + 55, + 408, + 296, + 515 + ], + "type": "text", + "content": " indexed by " + }, + { + "bbox": [ + 55, + 408, + 296, + 515 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 408, + 296, + 515 + ], + "type": "text", + "content": ", the outlier samples it generates can be formulated as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 524, + 298, + 568 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 524, + 298, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 524, + 298, + 568 + ], + "type": "interline_equation", + "content": "\\mathcal {V} _ {k} = \\left\\{\\mathbf {v} _ {k} \\mid \\frac {\\exp \\left(- \\frac {1}{2} \\left(\\mathbf {v} _ {k} - \\hat {\\boldsymbol {\\mu}} _ {k}\\right) ^ {\\top} \\hat {\\boldsymbol {\\Sigma}} ^ {- 1} \\left(\\mathbf {v} _ {k} - \\hat {\\boldsymbol {\\mu}} _ {k}\\right)\\right)}{(2 \\pi) ^ {d / 2} | \\hat {\\boldsymbol {\\Sigma}} | ^ {1 / 2}} < \\epsilon \\right\\}, \\tag {9}", + "image_path": "f0463babdac33f44ad1c84ed8bb17123e2ef292b2aee2e5efbc1e92383cd68fa.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 569, + 296, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 296, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 296, + 652 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 569, + 296, + 652 + ], + "type": "inline_equation", + "content": "v_{k}\\sim \\mathcal{N}(\\hat{\\mu}_{k},\\hat{\\Sigma}_{k})" + }, + { + "bbox": [ + 55, + 569, + 296, + 652 + ], + "type": "text", + "content": " represents the virtual outliers sampled from the estimated Gaussian distribution of subcluster " + }, + { + "bbox": [ + 55, + 569, + 296, + 652 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 569, + 296, + 652 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 55, + 569, + 296, + 652 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 569, + 296, + 652 + ], + "type": "text", + "content": " denoting its embedding dimensionality. It is worth noting that during model optimization, all outliers generated here, i.e. " + }, + { + "bbox": [ + 55, + 569, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\overline{\\mathcal{V}} = \\mathcal{V}_1\\cup \\dots \\cup \\mathcal{V}_k\\cup \\dots \\mathcal{V}_{|\\hat{\\mathcal{O}} |}" + }, + { + "bbox": [ + 55, + 569, + 296, + 652 + ], + "type": "text", + "content": " are classified into the \"Outlier\" class, separate from the real and fake classes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 660, + 212, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 212, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 212, + 673 + ], + "type": "text", + "content": "3.3. Outlier-Guided Tri-Training" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "To enable better performance of FakeRadar for tasks of deepfake video detection, we design a strategy called Outlier-Guided Tri-Training for model optimization jointly" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "content": "combining proposed outlier-driven contrastive loss and outlier-conditioned cross-entropy loss. With this approach, a triplet-class classifier (integrated with the pre-trained backbone and adapters) is trained to explicitly distinguish among three categories: \"Real\", \"Fake\", and \"Outlier\". This design not only improves the model's ability to detect known forgeries but also sensitizes it to emerging manipulation patterns." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 183, + 553, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 183, + 553, + 302 + ], + "spans": [ + { + "bbox": [ + 313, + 183, + 553, + 302 + ], + "type": "text", + "content": "Outlier-Driven Contrastive Loss. In this work, we propose an outlier-driven contrastive loss to enhance the model's ability to delineate class boundaries within the feature space, thereby improving feature discriminability. This loss function is designed to maximize the similarity between samples and their respective subcluster centers while minimizing their similarity to other subcluster centers. Additionally, it treats generated outliers as negative examples, enforcing low similarity between samples and outliers to strengthen boundary discrimination." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 303, + 553, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 303, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 553, + 327 + ], + "type": "text", + "content": "To this end, the loss is formulated using an InfoNCE-like contrastive loss given an input sample " + }, + { + "bbox": [ + 313, + 303, + 553, + 327 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 303, + 553, + 327 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 336, + 561, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 336, + 561, + 389 + ], + "spans": [ + { + "bbox": [ + 314, + 336, + 561, + 389 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o n}} = - \\log \\frac {\\exp \\left(\\frac {s \\left(h _ {i} , \\mu_ {k _ {i} ^ {+}}\\right)}{\\tau}\\right)}{\\sum_ {k ^ {-}} \\exp \\left(\\frac {s \\left(h _ {i} , \\mu_ {k ^ {-}}\\right)}{\\tau}\\right) + \\sum_ {v _ {j} \\in \\overline {{\\nu}}} \\exp \\left(\\frac {s \\left(h _ {i} , v _ {j}\\right)}{\\tau}\\right)}, \\tag {10}", + "image_path": "689e2be9773a5373fc35056470f039d1bb114f732880acd26689ed1cd0357039.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 390, + 554, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 554, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 390, + 554, + 510 + ], + "type": "inline_equation", + "content": "h_i" + }, + { + "bbox": [ + 313, + 390, + 554, + 510 + ], + "type": "text", + "content": " denotes the feature representation of sample " + }, + { + "bbox": [ + 313, + 390, + 554, + 510 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 390, + 554, + 510 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 313, + 390, + 554, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 390, + 554, + 510 + ], + "type": "text", + "content": " represents the outlier cache, providing additional hard negative samples to improve the model's ability to distinguish boundary samples. Additionally, " + }, + { + "bbox": [ + 313, + 390, + 554, + 510 + ], + "type": "inline_equation", + "content": "k^{-}" + }, + { + "bbox": [ + 313, + 390, + 554, + 510 + ], + "type": "text", + "content": " represents all subclusters except the one to which sample " + }, + { + "bbox": [ + 313, + 390, + 554, + 510 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 390, + 554, + 510 + ], + "type": "text", + "content": " belongs, with its corresponding " + }, + { + "bbox": [ + 313, + 390, + 554, + 510 + ], + "type": "inline_equation", + "content": "u_{k^{-}}" + }, + { + "bbox": [ + 313, + 390, + 554, + 510 + ], + "type": "text", + "content": " denoting the centers of these subclusters, used as negative samples. The function " + }, + { + "bbox": [ + 313, + 390, + 554, + 510 + ], + "type": "inline_equation", + "content": "s(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 313, + 390, + 554, + 510 + ], + "type": "text", + "content": " denotes cosine similarity, which quantifies the similarity between samples, and " + }, + { + "bbox": [ + 313, + 390, + 554, + 510 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 390, + 554, + 510 + ], + "type": "text", + "content": " is a temperature hyperparameter that controls the smoothness of similarity scaling." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 525, + 554, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 554, + 656 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 554, + 656 + ], + "type": "text", + "content": "Outlier-Conditioned Cross-Entropy Loss. While the proposed outlier-driven contrastive loss enforces stronger feature-space separation, we still require a concrete classification objective to handle a three-category task: real, fake, and outlier. In particular, the model must not only determine whether a video is fake or real but also assign a separate label for those samples situated near decision boundaries (i.e., outliers). This triplet-class cross-entropy loss is crucial, as it enables the model to explicitly differentiate uncertain boundary samples and label them as outliers, thereby avoiding confusion." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 657, + 553, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 657, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 313, + 657, + 553, + 680 + ], + "type": "text", + "content": "Concretely, the triplet-class cross-entropy loss can be defined as follows:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 357, + 689, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 689, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 357, + 689, + 553, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c l s}} = - \\sum_ {c \\in \\{\\text {R e a l}, \\text {F a k e}, \\text {O u t l i e r} \\}} y _ {c} \\log p _ {c}, \\tag {11}", + "image_path": "1f1625bc6a30b9b1961a8a057aa39201ad6f3cf276c855ee166562502848e675.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13386" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 145, + 70, + 465, + 236 + ], + "blocks": [ + { + "bbox": [ + 145, + 70, + 465, + 236 + ], + "lines": [ + { + "bbox": [ + 145, + 70, + 465, + 236 + ], + "spans": [ + { + "bbox": [ + 145, + 70, + 465, + 236 + ], + "type": "table", + "html": "
MethodBackboneInput TypeTesting Set AUC (%)
FF++CDFv2DFDCPDFDCDFD
Xception [6]XceptionFrame99.373.7-70.9-
Face X-Ray [30]HRNetFrame97.879.5-65.595.4
PCL+i2G [60]ResnetFrame99.890.074.467.5-
SLADD [4]XceptionFrame98.479.776.0--
UIA-ViT [63]Vision TransformerFrame99.382.475.8-94.7
UCF [53]XceptionFrame-82.4-80.594.5
SeeABLE [27]EfficientNetFrame-87.386.375.9-
LSDA [54]EfficientNetFrame-91.181.277.095.6
FTCN [61]3D ResNetVideo99.886.9-74.0-
RealForensics [21]3D ResNetVideo99.886.9-75.9-
LTTD [19]Vision TransformerVideo99.589.3-80.4-
AltFreezing [51]3D ResNetVideo99.789.5--93.7
TALL [52]Swin TransformerVideo99.990.8-76.8-
StyleFlow [5]3D ResNetVideo99.189.0--96.1
NACO [59]Vision TransformerVideo99.889.5-76.7-
FakeRadar (Ours)Vision TransformerVideo99.191.788.584.196.2
", + "image_path": "6788ca724ad03870e4dc24d83dfcd3a16c6be15bf3b90673dd91f1878b67ed37.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 283, + 295, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 295, + 355 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 295, + 355 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 283, + 295, + 355 + ], + "type": "inline_equation", + "content": "y_{c}" + }, + { + "bbox": [ + 55, + 283, + 295, + 355 + ], + "type": "text", + "content": " is the ground-truth label for samples belonging to class " + }, + { + "bbox": [ + 55, + 283, + 295, + 355 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 283, + 295, + 355 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 283, + 295, + 355 + ], + "type": "inline_equation", + "content": "p_{c}" + }, + { + "bbox": [ + 55, + 283, + 295, + 355 + ], + "type": "text", + "content": " is the model-predicted probability of samples from that class, with " + }, + { + "bbox": [ + 55, + 283, + 295, + 355 + ], + "type": "inline_equation", + "content": "c =" + }, + { + "bbox": [ + 55, + 283, + 295, + 355 + ], + "type": "text", + "content": " \"Outlier\" denoting the outlier label. This loss compels the model to accurately partition each category in the feature space, especially for samples near class boundaries." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 369, + 295, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 295, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 295, + 395 + ], + "type": "text", + "content": "Total Loss. Finally, we combine both losses into the total objective for model optimization:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 129, + 403, + 294, + 416 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 403, + 294, + 416 + ], + "spans": [ + { + "bbox": [ + 129, + 403, + 294, + 416 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {c o n}} + \\lambda \\mathcal {L} _ {\\text {c l s}}, \\tag {12}", + "image_path": "cbcc78652c7818be66fe6a39ee4d078d84e1b1b2c4167ab9dc23dfa00d6d2b83.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 424, + 296, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 424, + 296, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 424, + 296, + 474 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 424, + 296, + 474 + ], + "type": "inline_equation", + "content": "\\lambda = 0.5" + }, + { + "bbox": [ + 55, + 424, + 296, + 474 + ], + "type": "text", + "content": " is a hyperparameter balancing the relative importance of classification and contrastive goals. In practice, each mini-batch comprises 16 training samples " + }, + { + "bbox": [ + 55, + 424, + 296, + 474 + ], + "type": "inline_equation", + "content": "(\\mathrm{FF}++)" + }, + { + "bbox": [ + 55, + 424, + 296, + 474 + ], + "type": "text", + "content": ", and 16 outlier samples during model training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 483, + 137, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 137, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 137, + 496 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 502, + 168, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 168, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 168, + 515 + ], + "type": "text", + "content": "4.1. Experiment Setups" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 520, + 295, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 295, + 675 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 295, + 675 + ], + "type": "text", + "content": "Datasets. Following established experimental practices [5, 51, 52, 54], this study primarily utilizes the high-quality (HQ) version of the FaceForensics++ dataset [43], denoted as " + }, + { + "bbox": [ + 55, + 520, + 295, + 675 + ], + "type": "inline_equation", + "content": "\\mathrm{FF} + + (\\mathrm{HQ})" + }, + { + "bbox": [ + 55, + 520, + 295, + 675 + ], + "type": "text", + "content": ", for FakeRadar's evaluation. This dataset consists of 1000 pristine videos and 4000 deepfakes videos generated using four different manipulation techniques: DFD [41], F2F [46], FS [26], and NT [47]. The dataset is split into training, validation, and testing sets according to the official protocol. This study primarily focuses on evaluating the cross-domain performance of our method. To this end, we incorporate four additional datasets for evaluation, including CDFv2 [33], DFDCP [11], DFDC [12], and DFD [41]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Implementation Details. To ensure fair comparisons, we maintain most implementation details consistent with exist" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 315, + 281, + 555, + 476 + ], + "blocks": [ + { + "bbox": [ + 55, + 239, + 555, + 264 + ], + "lines": [ + { + "bbox": [ + 55, + 239, + 555, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 555, + 264 + ], + "type": "text", + "content": "Table 1. Generalization comparisons of FakeRadar and state-of-the-art algorithms on cross-dataset evaluation, using video-level AUC (%) as the metric." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 281, + 555, + 476 + ], + "lines": [ + { + "bbox": [ + 315, + 281, + 555, + 476 + ], + "spans": [ + { + "bbox": [ + 315, + 281, + 555, + 476 + ], + "type": "table", + "html": "
Training SetMethodDFF2FFSNTavg.
DFDCL [45]99.9877.1361.0175.0178.28
UIA-ViT [63]99.2974.4453.8970.9274.64
TALL [52]99.3569.3354.3874.7474.45
AltFreezing [51]98.8069.3372.0777.4179.40
FakeRadar (Ours)99.9586.4788.1266.3682.23
F2FDCL [45]91.9199.2159.5866.6779.34
UIA-ViT [63]83.3998.3268.3767.1779.31
TALL [52]85.5298.6762.1570.2479.15
AltFreezing [51]81.1298.8062.9770.2678.29
FakeRadar (Ours)96.9499.0982.3368.4686.75
FSDCL [45]74.8069.7599.9052.6074.26
UIA-ViT [63]81.0266.3099.0449.2673.91
TALL [52]77.8367.3098.7752.5774.12
AltFreezing [51]80.6765.9499.7857.2375.90
FakeRadar (Ours)96.3982.0699.8352.6482.73
NTDCL [45]91.2352.1379.3198.9780.41
UIA-ViT [63]79.3767.8945.9494.5971.95
TALL [52]84.0272.8551.6695.1675.92
AltFreezing [51]88.8176.7857.9295.7179.80
FakeRadar (Ours)96.7183.3467.5098.4186.49
", + "image_path": "e97d01c9ae6b6f12c97ef8a97aac64da06b080237996b1455720a0af81b4ced0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 481, + 555, + 515 + ], + "lines": [ + { + "bbox": [ + 313, + 481, + 555, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 481, + 555, + 515 + ], + "type": "text", + "content": "Table 2. Generalization comparisons of FakeRadar and state-of-the-art algorithms on cross-manipulation evaluation, using video-level AUC (\\%) as the metric." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 545, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 715 + ], + "type": "text", + "content": "ing methods [5, 51, 59]. However, our FakeRadar framework employs ViT-B/16 from CLIP [40] as the backbone, into which we insert an ST-Adapter [37] with a 384-dimensional bottleneck for parameter-efficient fine-tuning. Additionally, the classification head of our triplet-class classifier consists of a single linear layer with softmax activation. During model training, we sample four temporal clips from each video, each containing 12 consecutive frames. We use Adam as the optimizer and apply a cosine learning rate scheduler with an initial learning rate of " + }, + { + "bbox": [ + 313, + 545, + 555, + 715 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 313, + 545, + 555, + 715 + ], + "type": "text", + "content": ". The model is trained for 60 epochs with a batch size of 16 and a weight decay of " + }, + { + "bbox": [ + 313, + 545, + 555, + 715 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-4}" + }, + { + "bbox": [ + 313, + 545, + 555, + 715 + ], + "type": "text", + "content": ". Prior to model optimization, we perform data preprocessing by detecting faces with RetinaFace [10] and then standardize all face images" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13387" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 68, + 218, + 187 + ], + "blocks": [ + { + "bbox": [ + 55, + 68, + 218, + 187 + ], + "lines": [ + { + "bbox": [ + 55, + 68, + 218, + 187 + ], + "spans": [ + { + "bbox": [ + 55, + 68, + 218, + 187 + ], + "type": "image", + "image_path": "922b9eeaef948d8603a75b38ca1f8e6234094f1e1a4be5d0b342ec5031c10e7e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 188, + 555, + 254 + ], + "lines": [ + { + "bbox": [ + 54, + 188, + 555, + 254 + ], + "spans": [ + { + "bbox": [ + 54, + 188, + 555, + 254 + ], + "type": "text", + "content": "Figure 3. Evaluation of (a) Forgery Outlier Probing and (b)/(c) Outlier-Guided Tri-Training by FakeRadar. (a) \"FakeRadar w/o Dyn. Sub. M.\" and \"FakeRadar w/o Out. Gen.\" indicate that dynamic subcluster modeling and cluster-conditional outlier generation have been removed. The model is trained on FF++(HQ) and evaluated on DFDC [12], CDFv2 [33], and DFD [41]. (b) and (c) show, respectively, the count of real and fake test samples classified as \"fake\" across predicted probability intervals. The blue solid line denotes the binary classifier's distribution (distinguishing only real vs. fake), while the red dashed line denotes that of our proposed triplet classifier, which treats both fake and outlier samples as the new \"fake\". In this evaluation, the model is trained on FF++(HQ) and tested on CDFv2 [33]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 223, + 68, + 386, + 187 + ], + "blocks": [ + { + "bbox": [ + 223, + 68, + 386, + 187 + ], + "lines": [ + { + "bbox": [ + 223, + 68, + 386, + 187 + ], + "spans": [ + { + "bbox": [ + 223, + 68, + 386, + 187 + ], + "type": "image", + "image_path": "31311a0a131bb374e262660ed95c596955f78aaafd597515461b720503e075eb.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 391, + 68, + 552, + 187 + ], + "blocks": [ + { + "bbox": [ + 391, + 68, + 552, + 187 + ], + "lines": [ + { + "bbox": [ + 391, + 68, + 552, + 187 + ], + "spans": [ + { + "bbox": [ + 391, + 68, + 552, + 187 + ], + "type": "image", + "image_path": "68738ff670573c43f6792a7e60db13307fc720d2a45c0edf65643c50fc6dc79b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 274, + 133, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 274, + 133, + 286 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 133, + 286 + ], + "type": "text", + "content": "to " + }, + { + "bbox": [ + 55, + 274, + 133, + 286 + ], + "type": "inline_equation", + "content": "224\\times 224" + }, + { + "bbox": [ + 55, + 274, + 133, + 286 + ], + "type": "text", + "content": " pixels." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 301, + 296, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 301, + 296, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 301, + 296, + 374 + ], + "type": "text", + "content": "Evaluation Metrics. In this work, consistent with previous studies [5, 51, 52] employing video-level inputs, we use the area under the ROC curve (AUC) as the performance metric during evaluations. For methods that use images as model inputs, e.g. [27, 54], their video-level results are derived by averaging the predictions across all video frames." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 381, + 263, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 381, + 263, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 381, + 263, + 395 + ], + "type": "text", + "content": "4.2. Comparisons with the State-of-the-Arts" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 399, + 296, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 399, + 296, + 507 + ], + "spans": [ + { + "bbox": [ + 54, + 399, + 296, + 507 + ], + "type": "text", + "content": "In this study, we focus on evaluating the capabilities of cross-domain generalization for our detector to validate its effectiveness in detecting deepfake videos manipulated through various techniques. To achieve this, we employ two strategies for evaluation: cross-dataset evaluation and cross-manipulation evaluation, both of which demonstrate that our results predominantly surpass existing state-of-the-art (SoTA) algorithms for deepfake video detection, as illustrated by Table 1 and Table 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "text", + "content": "Cross-Dataset Evaluation. In this situation, we train our FakeRadar on the FF++(HQ) and test its performance on multiple benchmark datasets, including CDFv2 [33], DFD [41], DFDC [12], and DFDCP [11]. As reported in Table 1, FakeRadar demonstrates its superior generalization capability in cross-dataset evaluation, achieving the best performance across all cross-dataset cases while maintaining comparable performance on the FF++ test dataset against SoTA algorithms. Specifically, our algorithm significantly improves over other methods with images and videos as inputs, particularly outperforming the current best methods, UCF [53] and LTTD [19], by " + }, + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "inline_equation", + "content": "3.6\\%" + }, + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "inline_equation", + "content": "3.7\\%" + }, + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "text", + "content": " in the AUC metric, respectively, on the dataset DFDC [12]. These results robustly showcase the proposed method's enhanced ability of generalization to handle unseen manipulation techniques in cross-dataset evaluation scenarios." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 316, + 272, + 553, + 342 + ], + "blocks": [ + { + "bbox": [ + 316, + 272, + 553, + 342 + ], + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 342 + ], + "type": "table", + "html": "
M-(#)FOPOGTTCDFDFDCDFDCPDFDavg.
ODCLOCCE
1✓✓✓91.684.188.596.290.1
2✓✓89.981.285.694.987.9
3✓✓88.680.987.794.788.0
4✓88.880.286.794.587.6
5✓88.478.485.194.386.7
688.278.384.894.286.4
", + "image_path": "4e7cdb2f535cd0c34a83f91c44d61b391575912f41095b67e457561f01c269d5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 350, + 555, + 416 + ], + "lines": [ + { + "bbox": [ + 313, + 350, + 555, + 416 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 555, + 416 + ], + "type": "text", + "content": "Table 3. Ablation study results of FakeRadar by removing its one or more proposed components, using video-level AUC (\\%) as the metric. For clarity, \"FOP\", \"OGTT\", \"ODCL\", and \"OCCE\" refer to Forgery Outlier Probing, Outlier-Guided Tri-Training, Outlier-Driven Contrastive Loss, and Outlier-Conditioned Cross-Entropy Loss, respectively." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 423, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 555, + 640 + ], + "type": "text", + "content": "Cross-Manipulation Evaluation. To further assess our model's generalization across different manipulation techniques, we conduct cross-manipulation evaluation on the dataset FF++(HQ). In this case, we train the model using only real samples from FF++(HQ) and one manipulation type of deepfake samples, then test its performance on the remaining three manipulation types within FF++(HQ). As shown in Table 2, the proposed FakeRadar achieves the best average performance in all testing cases, particularly excelling in scenarios trained with F2F as the manipulation type, surpassing the second-best performing method, DCL [45], by " + }, + { + "bbox": [ + 313, + 423, + 555, + 640 + ], + "type": "inline_equation", + "content": "7.41\\%" + }, + { + "bbox": [ + 313, + 423, + 555, + 640 + ], + "type": "text", + "content": " in AUC. In evaluations focused on a single manipulation type, our method also achieves the best performance in most cases and the second-best in a few. This indicates that our approach, in cross-manipulation evaluation, demonstrates a performance advantage in cross-domain generalization compared to current SoTA algorithms of detecting deepfake videos." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 647, + 421, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 421, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 421, + 661 + ], + "type": "text", + "content": "4.3. Ablation Analysis" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "We conduct an ablation study to assess the contribution of each component in FakeRadar, with results presented in Table 3, Figure 3, and Figure 4. Experiments are performed in a cross-dataset setting, where the model is trained on" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13388" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\mathrm{FF}++(\\mathrm{HQ})" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " and tested on CDF, DFDC, DFDCP, and DFD, with average performance reported. As shown in Table 3 (Rows M-(1) to M-(6)), removing individual or multiple components leads to a progressive decline in AUC, from " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "90.1\\%" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "86.4\\%" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": ", demonstrating the significance of each component. Below, we analyze the impact of FakeRadar's key components in detail." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 177, + 296, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 177, + 296, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 177, + 296, + 380 + ], + "type": "text", + "content": "Effectiveness of Forgery Outlier Probing. To evaluate the impact of FOP, we remove or modify its submodules: dynamic subcluster modeling and cluster-conditional outlier generation. In \"FakeRadar w/o Dyn. Sub. M.\", we disable dynamic subcluster modeling, generating outliers only from clusters of samples of real and fake categories in training set. In \"FakeRadar w/o Out. Gen.\", we replace outlier samples with Gaussian noise. As shown in Figure 3(a), removing dynamic subcluster modeling reduces AUC on CDFv2, DFDC, and DFD to " + }, + { + "bbox": [ + 55, + 177, + 296, + 380 + ], + "type": "inline_equation", + "content": "90.2\\%" + }, + { + "bbox": [ + 55, + 177, + 296, + 380 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 177, + 296, + 380 + ], + "type": "inline_equation", + "content": "82.3\\%" + }, + { + "bbox": [ + 55, + 177, + 296, + 380 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 177, + 296, + 380 + ], + "type": "inline_equation", + "content": "94.7\\%" + }, + { + "bbox": [ + 55, + 177, + 296, + 380 + ], + "type": "text", + "content": ", respectively. While removing outlier generation has a smaller impact, performance remains below the complete FakeRadar framework. On CDFv2, FakeRadar " + }, + { + "bbox": [ + 55, + 177, + 296, + 380 + ], + "type": "inline_equation", + "content": "(91.7\\%)" + }, + { + "bbox": [ + 55, + 177, + 296, + 380 + ], + "type": "text", + "content": " outperforms \"FakeRadar w/o Out. Gen.\" by " + }, + { + "bbox": [ + 55, + 177, + 296, + 380 + ], + "type": "inline_equation", + "content": "2.0\\%" + }, + { + "bbox": [ + 55, + 177, + 296, + 380 + ], + "type": "text", + "content": ", demonstrating that FOP enhances generalization by generating effective outlier samples that improve the detection of novel deep-fake manipulations." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 401, + 296, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 296, + 484 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 296, + 484 + ], + "type": "text", + "content": "Effectiveness of Outlier-Guided Tri-Training. Comparing Table 3 (Rows M-(1) to M-(3)), removing Outlier-Driven Contrastive Loss or Outlier-Conditioned Cross-Entropy Loss leads to an AUC drop from " + }, + { + "bbox": [ + 55, + 401, + 296, + 484 + ], + "type": "inline_equation", + "content": "90.1\\%" + }, + { + "bbox": [ + 55, + 401, + 296, + 484 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 401, + 296, + 484 + ], + "type": "inline_equation", + "content": "87.9\\%" + }, + { + "bbox": [ + 55, + 401, + 296, + 484 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 401, + 296, + 484 + ], + "type": "inline_equation", + "content": "88.9\\%" + }, + { + "bbox": [ + 55, + 401, + 296, + 484 + ], + "type": "text", + "content": ", respectively, highlighting their importance. Since both losses refine the use of outlier samples, removing them would render FOP ineffective." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 486, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 296, + 714 + ], + "type": "text", + "content": "The core of our proposed optimization strategy is the integration of a triplet classifier to handle the outlier samples generated by FOP. To assess its impact, we replace the triplet classifier, which performs three-class classification (real, fake, and outlier), with a binary classifier that only distinguishes between real and fake. We then analyze the prediction probability distributions for cross-dataset evaluation, as shown in Figure 3 (b) and (c). As illustrated, for real samples, both the binary and triplet classifiers exhibit similar prediction patterns with high confidence in the low-probability region. However, for fake samples, the triplet classifier demonstrates a clear advantage. Compared to the binary classifier, it assigns higher forgery probabilities to a greater number of fake samples, indicating that the proposed outlier-guided tri-training increases the model's confidence in detecting unseen fake samples. This result validates the effectiveness of treating outlier samples as fake samples during inference and highlights its contribution to improving the model's generalization capability." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 322, + 72, + 553, + 155 + ], + "blocks": [ + { + "bbox": [ + 322, + 72, + 553, + 155 + ], + "lines": [ + { + "bbox": [ + 322, + 72, + 553, + 155 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 553, + 155 + ], + "type": "image", + "image_path": "226d93e674b4791eef2686dd56434c9ff217a868a209b633ea98a8efe1423524.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 159, + 555, + 204 + ], + "lines": [ + { + "bbox": [ + 313, + 159, + 555, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 159, + 555, + 204 + ], + "type": "text", + "content": "Figure 4. Feature visualization using t-SNE. We visualize the features of (a) the pre-trained CLIP ViT-B model and our (b) FakeRadar here. We train both models using the training set FF++(HQ), while testing on the FF++ test set." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 220, + 556, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 556, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 556, + 413 + ], + "type": "text", + "content": "Feature Visualization. We use t-SNE [48] to visualize the features of our FakeRadar and the pre-trained CLIP ViT-B model. As shown in Figure 4, the feature distribution of FakeRadar exhibits a more compact clustering structure and clearer boundaries between real and fake videos, compared to the pre-trained CLIP ViT-B model. Furthermore, FakeRadar effectively distinguishes real samples from the four types of forgery samples. This suggests that FakeRadar not only enables the model to learn discriminative features between categories but also establishes more robust decision boundaries with the aid of outlier samples. These feature distribution patterns underscore the advantages of FakeRadar in improving the model's generalization ability. The boundaries between categories are sharper, and sufficient tolerance space is provided, which is essential for addressing unseen forgery types in real-world scenarios." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 437, + 392, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 437, + 392, + 449 + ], + "spans": [ + { + "bbox": [ + 314, + 437, + 392, + 449 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 462, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 556, + 715 + ], + "type": "text", + "content": "In this work, we introduced FakeRadar, a novel deepfake video detection framework designed to address the limitations of existing methods in cross-domain generalization. Unlike traditional detectors that passively learn specific forgery patterns, FakeRadar proactively probes unknown forgeries by incorporating Forgery Outlier Probing to simulate diverse unseen manipulation patterns. This approach enables the model to explore a broader range of potential forgeries in feature space, significantly enhancing its adaptability to novel manipulations. To further improve detection robustness, we proposed Outlier-Guided Tri-Training, allowing the model to distinguish between real, fake, and outlier samples. This training paradigm, introduced outlier-driven contrastive loss and outlier-conditioned cross-entropy loss, sharpens the decision boundaries and enhances the detector's capacity to generalize across different manipulation types. Extensive experiments on benchmark datasets confirm that FakeRadar consistently outperforms state-of-the-art detection algorithms, particularly in cross-domain evaluations where conventional detectors struggle." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13389" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 161, + 115, + 173 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 161, + 115, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 161, + 115, + 173 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 180, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 180, + 294, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 180, + 294, + 236 + ], + "spans": [ + { + "bbox": [ + 61, + 180, + 294, + 236 + ], + "type": "text", + "content": "[1] Weiming Bai, Yufan Liu, Zhipeng Zhang, Bing Li, and Weiming Hu. Aunet: Learning relations between action units for face forgery detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24709-24719, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 236, + 294, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 236, + 294, + 291 + ], + "spans": [ + { + "bbox": [ + 61, + 236, + 294, + 291 + ], + "type": "text", + "content": "[2] Junyi Cao, Chao Ma, Taiping Yao, Shen Chen, Shouhong Ding, and Xiaokang Yang. End-to-end reconstruction-classification learning for face forgery detection. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4103-4112, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 292, + 295, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 292, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 61, + 292, + 295, + 346 + ], + "type": "text", + "content": "[3] Keshigeyan Chandrasegaran, Ngoc-Trung Tran, and Ngai-Man Cheung. A closer look at fourier spectrum discrepancies for cnn-generated images detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7200-7209, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 347, + 295, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 347, + 295, + 402 + ], + "spans": [ + { + "bbox": [ + 62, + 347, + 295, + 402 + ], + "type": "text", + "content": "[4] Liang Chen, Yong Zhang, Yibing Song, Lingqiao Liu, and Jue Wang. Self-supervised learning of adversarial example: Towards good generalizations for deepfake detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18710-18719, 2022. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 403, + 296, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 296, + 458 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 296, + 458 + ], + "type": "text", + "content": "[5] Jongwook Choi, Taehoon Kim, Yonghyun Jeong, Seungryul Baek, and Jongwon Choi. Exploiting style latent flows for generalizing deepfake video detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1133-1143, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 459, + 294, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 459, + 294, + 501 + ], + "spans": [ + { + "bbox": [ + 62, + 459, + 294, + 501 + ], + "type": "text", + "content": "[6] François Chollet. Xception: Deep learning with depthwise separable convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1251-1258, 2017. 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 502, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 294, + 525 + ], + "type": "text", + "content": "[7] Danielle Keats Citron and Robert M. Chesney. Deepfakes and the new disinformation war. Foreign Affairs, 2018. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 525, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 525, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 62, + 525, + 294, + 568 + ], + "type": "text", + "content": "[8] Di Cooke, Abigail Edwards, Sophia Barkoff, and Kathryn Kelly. As good as a coin toss: Human detection of a-generated images, videos, audio, and audiovisual stimuli. ArXiv, abs/2403.16760, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 570, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 570, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 62, + 570, + 295, + 624 + ], + "type": "text", + "content": "[9] Yunshu Dai, Jianwei Fei, and Fangjun Huang. Idguard: Robust, general, identity-centric poi proactive defense against face editing abuse. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11934-11943, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 58, + 625, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 625, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 58, + 625, + 294, + 678 + ], + "type": "text", + "content": "[10] Jiankang Deng, Jia Guo, Evangelos Ververas, Irene Kotsia, and Stefanos Zafeiriou. Retinaface: Single-shot multi-level face localisation in the wild. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5202-5211, 2020. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 680, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 294, + 712 + ], + "type": "text", + "content": "[11] B Dolhansky. The dee pfake detection challenge (dfdc) pre view dataset. arXiv preprint arXiv:1910.08854, 2019. 2, 6, 7" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 55, + 91, + 297, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 152 + ], + "type": "text", + "content": "This work was supported in part by the National Key R&D Program of China (2024YFB3908503), in part by the Guangxi Science and Technology Project (AB25069496), and in part by the National Natural Science Foundation of China (62172120 and 62322608)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[12] Brian Dolhansky, Joanna Bitton, Ben Pflaum, Jikuo Lu, Russ Howes, Menglin Wang, and Cristian Canton Ferrer. The deepfake detection challenge (dfdc) dataset. arXiv preprint arXiv:2006.07397, 2020. 2, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 118, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 172 + ], + "type": "text", + "content": "[13] Chengdong Dong, Ajay Kumar, and Eryun Liu. Think twice before detecting gan-generated fake images from their spectral domain imprints. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7855-7864, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 175, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 229 + ], + "type": "text", + "content": "[14] Shichao Dong, Jin Wang, Renhe Ji, Jiajun Liang, Haoqiang Fan, and Zheng Ge. Implicit identity leakage: The stumbling block to improving deepfake detection generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3994-4004, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 231, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 231, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 231, + 553, + 274 + ], + "type": "text", + "content": "[15] Xuefeng Du, Zhaoning Wang, Mu Cai, and Sharon Li. Towards unknown-aware learning with virtual outlier synthesis. In International Conference on Learning Representations, 2022. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 276, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 276, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 317, + 276, + 553, + 330 + ], + "type": "text", + "content": "[16] Joel Frank, Thorsten Eisenhofer, Lea Schonherr, Asja Fischer, Dorothea Kolossa, and Thorsten Holz. Leveraging frequency analysis for deep fake image recognition. In International conference on machine learning, pages 3247-3258. PMLR, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 332, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 374 + ], + "type": "text", + "content": "[17] Qiqi Gu, Shen Chen, Taiping Yao, Yang Chen, Shouhong Ding, and Ran Yi. Exploiting fine-grained face forgery clues via progressive enhancement learning. In AAAI Conference on Artificial Intelligence, 2021. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 376, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 431 + ], + "type": "text", + "content": "[18] Zhihao Gu, Yang Chen, Taiping Yao, Shouhong Ding, Jilin Li, Feiyue Huang, and Lizhuang Ma. Spatiotemporal inconsistency learning for deepfake video detection. In Proceedings of the 29th ACM international conference on multimedia, pages 3473-3481, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 434, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 476 + ], + "type": "text", + "content": "[19] Jiazhi Guan, Hang Zhou, Zhibin Hong, Errui Ding, Jingdong Wang, Chengbin Quan, and Youjian Zhao. Delving into sequential patches for deepfake detection. Advances in Neural Information Processing Systems, 35:4517-4530, 2022. 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 478, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 478, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 478, + 553, + 533 + ], + "type": "text", + "content": "[20] Alexandros Haliassos, Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. Lips don't lie: A generalisable and robust approach to face forgery detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5039-5049, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 535, + 555, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 555, + 589 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 555, + 589 + ], + "type": "text", + "content": "[21] Alexandros Haliassos, Rodrigo Mira, Stavros Petridis, and Maja Pantic. Leveraging real talking faces via self-supervision for robust forgery detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14950-14962, 2022. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 590, + 555, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 555, + 645 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 555, + 645 + ], + "type": "text", + "content": "[22] Alexandros Haliassos, Rodrigo Mira, Stavros Petridis, and Maja Pantic. Leveraging real talking faces via self-supervision for robust forgery detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14950-14962, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[23] Ammarah Hashmi, Sahibzada Adil Shahzad, Chia-Wen Lin, Yu Tsao, and Hsin-Min Wang. Unmasking illusions: Understanding human perception of audiovisual deepfakes. ArXiv, abs/2405.04097, 2024. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "text", + "content": "[24] Baojin Huang, Zhongyuan Wang, Jifan Yang, Jiaxin Ai, Qin Zou, Qian Wang, and Dengpan Ye. Implicit identity" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13390" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "text", + "content": "driven deepfake face swapping detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4490-4499, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 152 + ], + "type": "text", + "content": "[25] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "type": "text", + "content": "[26] Marek Kowalski. Faceswap. https://github.com/MarekKowalski/FaceSwap, 2018. GitHub repository, accessed on March 3, 2025. 3, 4, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 295, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 295, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 295, + 241 + ], + "type": "text", + "content": "[27] Nicolas Larue, Ngoc-Son Vu, Vitomir Struc, Peter Peer, and Vassilis Christophides. Seeable: Soft discrepancies and bounded contrastive learning for exposing deepfakes. In Proceedings of the IEEE/CVF international conference on computer vision, pages 21011-21021, 2023. 1, 3, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 243, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 294, + 308 + ], + "type": "text", + "content": "[28] Hong Li, Yutang Feng, Song Xue, Xuhui Liu, Bohan Zeng, Shanglin Li, Boyu Liu, Jianzhuang Liu, Shumin Han, and Baochang Zhang. Uv-idm: Identity-conditioned latent diffusion model for face uv-texture generation. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10585-10595, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 309, + 294, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 353 + ], + "type": "text", + "content": "[29] Hanzhe Li, Jiaran Zhou, Yuezun Li, Baoyuan Wu, Bin Li, and Junyu Dong. Freqblender: Enhancing deepfake detection by blending frequency knowledge. arXiv preprint arXiv:2404.13872, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 409 + ], + "type": "text", + "content": "[30] Lingzhi Li, Jianmin Bao, Ting Zhang, Hao Yang, Dong Chen, Fang Wen, and Baining Guo. Face x-ray for more general face forgery detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5001-5010, 2020. 1, 2, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 294, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 443 + ], + "type": "text", + "content": "[31] Yuezun Li and Siwei Lyu. Exposing deepfake videos by detecting face warping artifacts. arXiv preprint arXiv:1811.00656, 2018. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 445, + 294, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 489 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 489 + ], + "type": "text", + "content": "[32] Yuezun Li, Ming-Ching Chang, and Siwei Lyu. In ictu oculi: Exposing ai created fake videos by detecting eye blinking. In 2018 IEEE International Workshop on Information Forensics and Security (WIFS), pages 1-7, 2018. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 490, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 294, + 544 + ], + "type": "text", + "content": "[33] Yuezun Li, Xin Yang, Pu Sun, Honggang Qi, and Siwei Lyu. Celeb-df: A large-scale challenging dataset for deepfake forensics. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3204-3213, 2020. 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 546, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 546, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 546, + 294, + 601 + ], + "type": "text", + "content": "[34] Yuzhen Lin, Wentang Song, Bin Li, Yuezun Li, Jiangqun Ni, Han Chen, and Qiushi Li. Fake it till you make it: Curricular dynamic forgery augmentations towards general deepfake detection. In European Conference on Computer Vision, pages 104-122. Springer, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 602, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 678 + ], + "type": "text", + "content": "[35] Dat NGUYEN, Nesryne MEJRI, Inder Pal SINGH, Polina KULESHOVA, Marcella ASTRID, Anis KACEM, Enjie GHORBEL, and Djamila AOUADA. Laa-net: Localized artifact attention network for quality-agnostic and generalizable deepfake detection. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17395–17405, 2024. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "text", + "content": "[36] Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards universal fake image detectors that generalize across generative models. In Proceedings of the IEEE/CVF Conference" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "on Computer Vision and Pattern Recognition, pages 24480-24489, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[37] Junting Pan, Ziyi Lin, Xiatian Zhu, Jing Shao, and Hongsheng Li. St-adapter: Parameter-efficient image-to-video transfer learning. Advances in Neural Information Processing Systems, 35:26462-26477, 2022. 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "type": "text", + "content": "[38] Guilin Pang, Baopeng Zhang, Zhu Teng, Zige Qi, and Jianping Fan. Mre-net: Multi-rate excitation network for deepfake video detection. IEEE Transactions on Circuits and Systems for Video Technology, 33(8):3663-3676, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 186, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 230 + ], + "type": "text", + "content": "[39] Yuyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. Thinking in frequency: Face forgery detection by mining frequency-aware clues. In European conference on computer vision, pages 86-103. Springer, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 232, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 232, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 553, + 297 + ], + "type": "text", + "content": "[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 2, 3, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 298, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 553, + 342 + ], + "type": "text", + "content": "[41] Md Shohel Rana, Mohammad Nur Nobi, Bedduhu Murali, and Andrew H Sung. Deepfake detection: A systematic literature review. IEEE access, 10:25494-25513, 2022. 3, 4, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 343, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 396 + ], + "type": "text", + "content": "[42] Meitar Ronen, Shahaf E. Finder, and Oren Freifeld. Deep-dpm: Deep clustering with an unknown number of clusters. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9861-9870, 2022. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 399, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 454 + ], + "type": "text", + "content": "[43] Andreas Rössler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Niessner. Faceforensics++: Learning to detect manipulated facial images. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1-11, 2019. 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 456, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 499 + ], + "type": "text", + "content": "[44] Kaede Shiohara and Toshihiko Yamasaki. Detecting deepfakes with self-blended images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18720-18729, 2022. 1, 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 500, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 553, + 544 + ], + "type": "text", + "content": "[45] Ke Sun, Taiping Yao, Shen Chen, Shouhong Ding, Jilin Li, and Rongrong Ji. Dual contrastive learning for general face forgery detection. In Proceedings of the AAAI conference on artificial intelligence, pages 2316-2324, 2022. 2, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 545, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 601 + ], + "type": "text", + "content": "[46] Justus Thies, Michael Zollhofer, Marc Stamminger, Christian Theobalt, and Matthias Nießner. Face2face: Real-time face capture and reenactment of rgb videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2387-2395, 2016. 1, 3, 4, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 601, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 644 + ], + "type": "text", + "content": "[47] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. Acm Transactions on Graphics (TOG), 38(4):1-12, 2019. 1, 3, 4, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 647, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 678 + ], + "type": "text", + "content": "[48] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9 (86):2579-2605, 2008. 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[49] Junke Wang, Zuxuan Wu, Wenhao Ouyang, Xintong Han, Jingjing Chen, Yu-Gang Jiang, and Ser-Nam Li. M2tr: Multi-modal multi-scale transformers for deepfake detec" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13391" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "text", + "content": "tion. In Proceedings of the 2022 international conference on multimedia retrieval, pages 615-623, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 149 + ], + "type": "text", + "content": "[50] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnn-generated images are surprisingly easy to spot... for now. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8695-8704, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 151, + 295, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 151, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 151, + 295, + 205 + ], + "type": "text", + "content": "[51] Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, and Houqiang Li. Altfreezing for more general video face forgery detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4129-4138, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 295, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 295, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 295, + 259 + ], + "type": "text", + "content": "[52] Yuting Xu, Jian Liang, Gengyun Jia, Ziming Yang, Yanhao Zhang, and Ran He. Tall: Thumbnail layout for deepfake video detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22658-22668, 2023. 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 295, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 295, + 315 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 295, + 315 + ], + "type": "text", + "content": "[53] Zhiyuan Yan, Yong Zhang, Yanbo Fan, and Baoyuan Wu. Ucf: Uncovering common features for generalizable deepfake detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22412-22423, 2023. 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 316, + 295, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 295, + 380 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 295, + 380 + ], + "type": "text", + "content": "[54] Zhiyuan Yan, Yuhao Luo, Siwei Lyu, Qingshan Liu, and Baoyuan Wu. Transcending forgery specificity with latent space augmentation for generalizable deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8984-8994, 2024. 1, 2, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 383, + 295, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 295, + 435 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 295, + 435 + ], + "type": "text", + "content": "[55] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications. ACM Computing Surveys, 56(4): 1-39, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 437, + 295, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 295, + 491 + ], + "type": "text", + "content": "[56] Wenyuan Yang, Xiaoyu Zhou, Zhikai Chen, Bofei Guo, Zhongjie Ba, Zhihua Xia, Xiaochun Cao, and Kui Ren. Avoid-df: Audio-visual joint learning for detecting deepfake. IEEE Transactions on Information Forensics and Security, 18:2015–2029, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 493, + 295, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 295, + 535 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 295, + 535 + ], + "type": "text", + "content": "[57] Shenghai Yuan, Jina Huang, Xianyi He, Yunyuan Ge, Yu-jun Shi, Liuhan Chen, Jiebo Luo, and Li Yuan. Identity-preserving text-to-video generation by frequency decomposition, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 537, + 295, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 295, + 581 + ], + "type": "text", + "content": "[58] Yu Zeng, Zhe L. Lin, and Vishal M. Patel. Sketched: Mask-free local image manipulation with partial sketches. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5941-5951, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 582, + 295, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 295, + 634 + ], + "type": "text", + "content": "[59] Daichi Zhang, Zihao Xiao, Shikun Li, Fanzhao Lin, Jianmin Li, and Shiming Ge. Learning natural consistency representation for face forgery video detection. In European Conference on Computer Vision, pages 407-424. Springer, 2024. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 636, + 295, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 295, + 680 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 295, + 680 + ], + "type": "text", + "content": "[60] Tianchen Zhao, Xiang Xu, Mingze Xu, Hui Ding, Yuanjun Xiong, and Wei Xia. Learning self-consistency for deepfake detection. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 15003-15013, 2021. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "type": "text", + "content": "[61] Yinglin Zheng, Jianmin Bao, Dong Chen, Ming Zeng, and Fang Wen. Exploring temporal coherence for more general video face forgery detection. In Proceedings of the" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 205 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "IEEE/CVF international conference on computer vision, pages 15044-15054, 2021. 2, 3, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[62] Yang Zhou, Zichong Chen, and Hui Huang. Deformable oneshot face stylization via dino semantic guidance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7787-7796, 2024. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 141, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 205 + ], + "type": "text", + "content": "[63] Wanyi Zhuang, Qi Chu, Zhentao Tan, Qiankun Liu, Haojie Yuan, Changtao Miao, Zixiang Luo, and Nenghai Yu. Uiavit: Unsupervised inconsistency-aware method based on vision transformer for face forgery detection. In European conference on computer vision, pages 391-407. Springer, 2022. 6" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13392" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Fast Globally Optimal and Geometrically Consistent 3D Shape Matching/cbbb1f39-340b-4752-9244-371644b15edf_content_list.json b/2025/Fast Globally Optimal and Geometrically Consistent 3D Shape Matching/cbbb1f39-340b-4752-9244-371644b15edf_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..610ee70e77540291313275cb31fa7f658803abef --- /dev/null +++ b/2025/Fast Globally Optimal and Geometrically Consistent 3D Shape Matching/cbbb1f39-340b-4752-9244-371644b15edf_content_list.json @@ -0,0 +1,1805 @@ +[ + { + "type": "text", + "text": "Fast Globally Optimal and Geometrically Consistent 3D Shape Matching", + "text_level": 1, + "bbox": [ + 130, + 130, + 867, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Paul Roetzer1,2 \n1 University of Bonn", + "bbox": [ + 321, + 179, + 491, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Florian Bernard $^{1,2}$ $^{2}$ Lamarr Institute", + "bbox": [ + 501, + 179, + 669, + 215 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0a98acca2030be1129c670887f981b56749839a31efa968a436fe3aad142706c.jpg", + "image_caption": [ + "Key Idea of Our Approach" + ], + "image_footnote": [], + "bbox": [ + 76, + 289, + 166, + 359 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/da7792113543c129534f8b3b3f8ae27c7ca047d050c5a4f5d44f4a5ca4dc2900.jpg", + "image_caption": [ + "Figure 1. Left: Illustration of our approach for geometrically consistent matching of 3D shape $\\mathcal{X}$ to 3D shape $\\mathcal{Y}$ . We represent shape $\\mathcal{X}$ using $n$ surface cycles $\\mathcal{C}_1, \\ldots, \\mathcal{C}_n$ and then consistently match these to shape $\\mathcal{Y}$ while preserving neighbourhood relations between the cycles. Middle: The triangulation- and colour-transfer between shapes using computed matchings of Cao et al. [13] (not geometrically consistent) and ours (geometrically consistent) illustrates the importance of geometric consistency. Right: Our method is the first 3D shape matching approach that yields globally geometrically consistent matchings, that is globally optimal, and scalable in practice, opposed to all competing methods (OOM stands for 'out of memory' and the dashed line indicates a weak notion of geometric consistency, see explanation of methods in Sec. 5)." + ], + "image_footnote": [], + "bbox": [ + 171, + 297, + 243, + 354 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7859227a1ef7048dba4adb49f6a26daca995307b19acecec2690ee0441d4cc5e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 243, + 297, + 333, + 354 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/241bd2df8f0e62117ef665127831755f94d96f1844cb21e8d2c0ea8b38d81f07.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 247, + 424, + 363 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7e9e9f9740cdb1b9eb4df97b173b15c91cd4fa72228294bb8c5f79454d001282.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 433, + 247, + 504, + 363 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cf3ff8894f376bfda227b0ff94dfd96b1bcb13c2271ce3d1100c715e5322205b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 246, + 581, + 363 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/759b602e79dc3967063d7b58bbd065350cf5f611a2f65b8f1de1ded492b331af.jpg", + "image_caption": [ + "Source Cao et al. Ours Triangulation Transfer via Computed Matchings", + "Runtime Comparison on FAUST Dataset" + ], + "image_footnote": [], + "bbox": [ + 602, + 252, + 897, + 382 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 520, + 326, + 535 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Geometric consistency, i.e. the preservation of neighbourhoods, is a natural and strong prior in 3D shape matching. Geometrically consistent matchings are crucial for many downstream applications, such as texture transfer or statistical shape modelling. Yet, in practice, geometric consistency is often overlooked, or only achieved under severely limiting assumptions (e.g. a good initialisation). In this work, we propose a novel formalism for computing globally optimal and geometrically consistent matchings between 3D shapes which is scalable in practice. Our key idea is to represent the surface of the source shape as a collection of cyclic graphs, which are then consistently matched to the target shape. Mathematically, we construct a hyper product graph (between source and target shape), and then cast 3D shape matching as a minimum-cost circulation flow problem in this hyper graph, which yields global geometrically consistent matchings between both shapes. We empirically show that our formalism is efficiently solvable and that it leads to high-quality results. Our code is publicly available.1", + "bbox": [ + 89, + 554, + 483, + 854 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 520, + 643, + 535 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The availability of correspondences between visual data are a key prerequisite for many visual computing tasks, including 3D reconstruction [50], loss computation for deep learning [14], protein alignment [45], anomaly detection [86], object recognition [44], shape modelling [24] and others. In many practical cases, it is desirable that correspondences preserve neighbourhood relations. For example, in 3D shape matching this could mean that when bringing a pair of neighbouring points on a source shape into correspondence with a pair of points on a target shape, the pair of points must remain neighbours on the target. Such a neighbourhood preservation can serve as strong topological prior, for example to resolve ambiguities, achieve robustness, or ensure well-posedness of an otherwise ill-posed problem. However, despite its crucial importance, neighbourhood preservation is in practice often-times overlooked. This is because many of the well-known formalisms are not efficiently solvable (e.g. the quadratic assignment problem [64], or general integer programming [76], which have been used to tackle image keypoint matching [55], graph matching [39, 91], or 3D shape matching [8, 34, 58, 89]).", + "bbox": [ + 511, + 549, + 906, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work we specifically focus on the task of non-rigid geometrically consistent 3D shape matching, in which", + "bbox": [ + 511, + 871, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1https://github.com/paul0noah/geco", + "bbox": [ + 107, + 887, + 377, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "912", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/c5fed2dade15da79455c5fefca16af65aaa75c7030a7c42cb55f0dcf2b8c6c87.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodGlobally OptimalGeometr. ConsistentScalable
MINA [8](✓)✗✗
SIGMA [34](✓)✗✗
PMSDP [58]✗✗✓
Windheuser et al. [89]✗✓✗
SpiderMatch [67]✓(✓)(✓)
Ours✓✓✓
", + "bbox": [ + 120, + 88, + 452, + 204 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Comparison of axiomatic 3D shape matching methods.", + "bbox": [ + 93, + 215, + 478, + 229 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "neighbourhood-preserving correspondences between two given 2D manifolds (embedded in 3D space) are sought for. So far, there does not exist any 3D shape matching approach that combines the following three desirable properties: (i) global optimality, (ii) neighbourhood preservation, and (iii) scalability, see Tab. 1. While the recent approach SpiderMatch [67] fulfils (i)-(iii) to some extent, its scalability relies heavily on the choice of the cycle that is used to represent one of the shapes (cf. Fig. 1 right: solid vs. dashed purple line), thereby trading off neighbourhood preservation with scalability. Furthermore, proper geometric consistency (Definition 4) can only be enforced at intersection points of the cycle and thus also the quality of geometric consistency depends on the choice of the cycle. Inspired by this idea of representing a shape using a cycle, we represent one shape using a collection of cycles and show that this leads to an (in practice) efficiently solvable formalism which ensures proper geometric consistency. We summarise our main contributions as follows:", + "bbox": [ + 91, + 248, + 482, + 534 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- For the first time, we present a globally geometrically consistent formalism for 3D shape matching that is efficiently solvable to global optimality in practice.", + "- To achieve this, we introduce a novel 3D shape representation in which we represent a 3D shape using a collection of surface cycles, so that 3D shape matching can be cast as finding a minimum-cost flow circulation in a hyper product graph.", + "- We experimentally show that our formalism leads to high quality and geometrically consistent matchings between two 3D shapes.", + "- In addition to 3D shape matching, we also show, in a proof-of-concept manner, that our formalism can generalise to specific instances of other matching problems, such as graph matching when the source graph is planar." + ], + "bbox": [ + 93, + 536, + 482, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 781, + 232, + 797 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the following we discuss works that are most relevant to our approach. We start by discussing efficiently solvable matching problems, continue with 3D shape matching, and conclude with geometrically consistent 3D shape matching.", + "bbox": [ + 89, + 808, + 482, + 868 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Efficiently Solvable Matching Problems. There are various instances of correspondence problems that can be", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "solved efficiently. Among them is the linear assignment problem (LAP) [46], which matches points without considering their neighbourhood relations. For matching time series data, sequences, or (open) contours, the popular dynamic time warping algorithm can efficiently compute solutions while preserving neighbourhood relations [70]. Closed contours can be matched analogously via graph cuts in product graphs [71]. Certain tracking problems can be formulated as efficiently solvable flow problems [49], and model-based image segmentation can also be solved efficiently by matching a 2D contour to an image [18, 32, 74]. The matching of a 2D contour to a 3D shape can be solved using variants of Dijkstra's algorithm for finding minimum cost cycles in product graphs [48, 68]. The mentioned approaches show that there are formalisms to efficiently solve a diverse range of matching problems. Yet, these approaches do not directly generalise to 3D shape matching. In this work, we propose a novel geometrically consistent 3D shape matching formalism that is (in practice) efficiently solvable to global optimality.", + "bbox": [ + 511, + 90, + 903, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Shape Matching is the task of finding correspondences between two non-rigidly deformed surfaces. For an in-depth overview on 3D shape matching we refer the reader to survey papers [19, 82, 85]. Many axiomatic shape matching approaches [11, 60, 65] build on the functional maps framework, efficiently solving shape matching in the spectral domain [28, 35, 56, 58, 59, 62]. In addition, functional maps have been adopted in several deep shape matching variants, either trained in a supervised [36, 52, 54, 83, 88] or in a unsupervised manner [2, 4, 13, 20, 22, 38, 51, 78]. There are also alternative approaches based on consensus maximisation [61], or using convex relaxations [15, 58] or mixed integer programming [8, 34]. Yet, many of the existing 3D shape matching methods neglect geometric consistency, as it leads to hard-to-solve formalisms.", + "bbox": [ + 511, + 397, + 903, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Geometrically Consistent 3D Shape Matching. There are several approaches that have recognised the importance of geometric consistency in 3D shape matching. Some of them incorporate neighbourhood information by using formalisms based on the quadratic assignment problem (QAP) - yet, the QAP is NP-hard [64], so that respective approaches consider relaxations [12, 23, 47] or heuristics [7, 40, 79]. Opposed to such discrete formulations, there are also approaches that tackle 3D shape matching by deforming a continuous shape parametrisation using local optimisation. However, due to the severe nonconvexity of resulting problems, they rely on a good initialisation, e.g. in the form of sparse sets of landmark correspondences [72, 73, 75, 77, 81], or in the form of dense correspondences [30, 31, 87]. Windheuser et al. [89, 90] have modelled geometrically consistent shape matching in the discrete domain by matching triangles to triangles using an (expensive to solve) integer linear program. Re", + "bbox": [ + 511, + 628, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "913", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/69b5e98be273fdc8f74d031b2df2ed0b889e0a3fb5d92e9ecc0932e085e8139f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SymbolDescription
X=(Vx,FX)3D shape
Gx=(Vx,EX)3D shape graph of X
C=(VC,EC)surface cycle (cycle in Gx)
{C1,...,Cn}Representation of X with n surface cycles
nNumber of surface cycles
Y=(Vy, fy)3D shape Y
Gy=(Vy, Fy)3D shape graph of Y
Pi=(VPi, EPi)Product graph (of C_i and mesh Y)
{P1,...,Pn}Product graph collection (of {C1,...,Cn} and mesh Y)
H=(VH, EH)Hyper product graph (coupled product graphs {P,...,Pn})
mNumber of hyper edges |EH|
HVertex edge incidence matrices of H
", + "bbox": [ + 91, + 88, + 480, + 291 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 2. Summary of the notation used in this paper.", + "bbox": [ + 127, + 295, + 442, + 309 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "cently, approximative solvers [66, 69] and extensions for partial shapes [26, 27] have been proposed. Furthermore, 3D shape matching has recently been formulated as a shortest path problem with additional constraints [67]. In this work, the surface mesh of the source 3D shape is represented using a long self-intersecting curve that traces the shape surface, and is then matched to the other shape while preserving intersections of the curve. While their idea of considering alternative 3D shape representations is promising, in the presented framework proper geometric consistency is only enforced at intersection points of the curve (see Tab. 4). In addition, runtimes grow drastically with an increasing number of intersection points, which can be seen in Fig. 1 right when comparing SpiderMatch [67] (using a curve with few intersections) to SpiderMatch (geo.) (using a curve with many intersections). Inspired by the idea of an alternative path-based 3D shape representation, we represent a 3D shape as a collection of cyclic graphs, and then tackle 3D shape matching by solving coupled matching problems of the individual cyclic paths. Based on the couplings, we are able to ensure global geometry consistency. Furthermore, we empirically observe that the resulting formalism is efficiently solvable to global optimality.", + "bbox": [ + 89, + 335, + 483, + 684 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Background", + "text_level": 1, + "bbox": [ + 89, + 696, + 218, + 713 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We consider the task of finding a matching between a source shape $\\mathcal{X}$ and a target shape $\\mathcal{Y}$ (Sec. 3.1) such that the matching is geometrically consistent (Sec. 3.2). Our main notation is summarised in Tab. 2.", + "bbox": [ + 89, + 722, + 482, + 781 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Shapes and Graphs", + "text_level": 1, + "bbox": [ + 89, + 792, + 277, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the following, we define shapes and other relevant concepts, which are also illustrated in Fig. 2. We consider $3D$ shapes represented as triangular surface mesh:", + "bbox": [ + 89, + 814, + 482, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 1 (3D shape). A 3D shape $\\mathcal{X}$ is defined as a tuple $(\\mathcal{V}_{\\mathcal{X}},\\mathcal{F}_{\\mathcal{X}})$ of vertices $\\mathcal{V}_{\\mathcal{X}}$ and consistently oriented", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/955846ec6a4e5bd46dfbf3f4b8d7e861f28a79d13f04de3bb40351afdd4e5f47.jpg", + "image_caption": [ + "Figure 2. We represent a 3D shape (left) as a (directed) shape graph $\\mathcal{G}_{\\mathcal{X}}$ (middle). We call a cycle $\\mathcal{C}$ in $\\mathcal{G}_{\\mathcal{X}}$ a surface cycle (right)." + ], + "image_footnote": [], + "bbox": [ + 552, + 95, + 867, + 178 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(e.g. clock-wise) triangles $\\mathcal{F}_{\\mathcal{X}} \\subset \\mathcal{V}_{\\mathcal{X}} \\times \\mathcal{V}_{\\mathcal{X}} \\times \\mathcal{V}_{\\mathcal{X}}$ , such that $\\mathcal{X}$ forms an orientable continuous $2D$ manifold (possibly with boundary) embedded in $3D$ space.", + "bbox": [ + 511, + 244, + 903, + 290 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For our formalism it is advantageous to interpret a 3D shape as a graph, which we denote the shape graph:", + "bbox": [ + 511, + 301, + 903, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 2 (Shape graph). The shape graph (of 3D shape $\\mathcal{X}$ ) is a tuple $\\mathcal{G}_{\\mathcal{X}} = (\\mathcal{V}_{\\mathcal{X}}, \\mathcal{E}_{\\mathcal{X}})$ of vertices $\\mathcal{V}_{\\mathcal{X}}$ and directed edges $\\mathcal{E}_{\\mathcal{X}} \\subset \\mathcal{V}_{\\mathcal{X}} \\times \\mathcal{V}_{\\mathcal{X}}$ , such that each triangle in $\\mathcal{F}_{\\mathcal{X}}$ is represented by three unique and consistently directed edges.", + "bbox": [ + 511, + 340, + 903, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In addition to the shape graph (representing the whole 3D shape $\\mathcal{X}$ ), we consider certain subgraphs of $\\mathcal{G}_{\\mathcal{X}}$ :", + "bbox": [ + 511, + 412, + 903, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 3 (Surface cycle). A surface cycle (on 3D shape $\\mathcal{X}$ ) is a tuple $\\mathcal{C} = (\\mathcal{V}_{\\mathcal{C}}, \\mathcal{E}_{\\mathcal{C}})$ with (non-empty) vertices $\\mathcal{V}_{\\mathcal{C}} \\subset \\mathcal{V}_{\\mathcal{X}}$ and (non-empty) directed edges $\\mathcal{E}_{\\mathcal{C}} \\subset \\mathcal{V}_{\\mathcal{C}} \\times \\mathcal{V}_{\\mathcal{C}}$ , such that (i) $\\mathcal{E}_{\\mathcal{C}} \\subset \\mathcal{E}_{\\mathcal{X}}$ , and (ii) each vertex $v \\in \\mathcal{V}_{\\mathcal{C}}$ has exactly one incoming and one outgoing edge in $\\mathcal{E}_{\\mathcal{C}}$ .", + "bbox": [ + 511, + 452, + 903, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Geometrically Consistent 3D Shape Matching", + "text_level": 1, + "bbox": [ + 511, + 547, + 898, + 564 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We define geometrically consistent 3D shape matching in terms of the respective shape graphs:", + "bbox": [ + 511, + 569, + 903, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 4 (Geometrically Consistent Matching). The mapping $\\phi : \\mathcal{V}_{\\mathcal{X}} \\to \\mathcal{V}_{\\mathcal{Y}}$ is a geometrically consistent matching from the source shape $\\mathcal{X}$ to the target shape $\\mathcal{Y}$ if $\\phi$ preserves neighbourhoods as follows: whenever two vertices $x, \\bar{x} \\in \\mathcal{V}_{\\mathcal{X}}$ are connected (i.e. $(x, \\bar{x}) \\in \\mathcal{E}_{\\mathcal{X}}$ or $(\\bar{x}, x) \\in \\mathcal{E}_{\\mathcal{X}}$ ), then their corresponding vertices $\\phi(x), \\phi(\\bar{x}) \\in \\mathcal{V}_{\\mathcal{Y}}$ must either be (i) connected (i.e. $(\\phi(x), \\phi(\\bar{x})) \\in \\mathcal{E}_{\\mathcal{Y}}$ or $(\\phi(\\bar{x}), \\phi(x)) \\in \\mathcal{E}_{\\mathcal{Y}}$ ), or (ii) the same vertex (i.e. $\\phi(x) = \\phi(\\bar{x})$ ).", + "bbox": [ + 511, + 609, + 905, + 746 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The intuition is that neighbouring elements of $\\mathcal{X}$ must be matched to neighbouring elements of $\\mathcal{Y}$ . We note that we require both shapes to have equal genus such that our definition of geometric consistency makes sense.", + "bbox": [ + 511, + 755, + 905, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Our 3D Shape Matching Approach", + "text_level": 1, + "bbox": [ + 511, + 830, + 828, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section we develop our geometrically consistent 3D shape matching formulation which ensures global geometric consistency. Our approach is based on representing the", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "914", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/08f95eed8fca9222ec95803c635bf89d5cd3957a8923d6682e657d4f0f26eb72.jpg", + "image_caption": [ + "Figure 3. Surface cycle collection (middle) for representing the 3D shape of an icosahedron (left). Individual surface cycles are glued together via shared edges (neighbouring surface cycles share opposite edges, cf. pairs of thickened edges on the right: $\\parallel$ and $\\parallel$ )." + ], + "image_footnote": [], + "bbox": [ + 145, + 89, + 433, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "source 3D shape using a collection of surface cycles, which we then match to the target shape (Sec. 4.1). To this end, we formulate an individual subproblem for each surface cycle (Sec. 4.2.1). Further, we couple individual subproblems (Secs. 4.2.2 and 4.2.3) so that our matchings are geometrically consistent according to Definition 4. Finally, we show that our overall formalism can be interpreted as a minimum-cost circulation flow problem on a hyper product graph (Sec. 4.5).", + "bbox": [ + 89, + 263, + 483, + 400 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Surface Cycle-based 3D Shape Representation", + "text_level": 1, + "bbox": [ + 89, + 409, + 482, + 426 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We aim to represent the surface of 3D shape $\\mathcal{X}$ with a collection of surface cycles which are glued together at opposite edges:", + "bbox": [ + 89, + 431, + 483, + 477 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Definition 5 (Opposite edge). The opposite edge to an edge $(v, w)$ is defined as $-(v, w) := (w, v)$ .", + "bbox": [ + 89, + 488, + 483, + 520 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We note that for every non-boundary edge $e \\in \\mathcal{E}_{\\mathcal{X}}$ of shape $\\mathcal{X}$ , the opposite edge $-e \\in \\mathcal{E}_{\\mathcal{X}}$ is also part of the shape graph $\\mathcal{G}_{\\mathcal{X}}$ (cf. Fig. 3 right) because the pair of opposite edges $e$ and $-e$ belong to two neighbouring triangles.", + "bbox": [ + 89, + 529, + 483, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Definition 6 (Shape as collection of surface cycles). We represent a 3D shape $\\mathcal{X}$ with a collection of $n \\in \\mathbb{N}^+$ surface cycles $\\mathcal{C}_1, \\ldots, \\mathcal{C}_n$ (on shape $\\mathcal{X}$ ) that partition the surface of $\\mathcal{X}$ into $n$ polygonal patches, such that:", + "bbox": [ + 89, + 599, + 483, + 660 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(i) $\\mathcal{E}_{\\mathcal{C}_i}\\cap \\mathcal{E}_{\\mathcal{C}_j} = \\emptyset$ for all $i,j\\in \\{1,\\ldots ,n\\} ,i\\neq j$ , and (ii) for every non-boundary edge $e\\in \\mathcal{E}_{\\mathcal{C}_i}$ of shape $\\mathcal{X}$ there exists $j$ such that $-e\\in \\mathcal{E}_{\\mathcal{C}_j}$", + "bbox": [ + 89, + 660, + 483, + 708 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Condition (i) ensures that each edge of $\\mathcal{X}$ can be part of at most one surface cycle, and (ii) ensures that neighbouring surface cycles cover opposite edges, see Fig. 3. We note that from here on we assume that each surface cycle represents an individual triangle of shape $\\mathcal{X}$ . We discuss more general polygonal surface cycles in Sec. C.3 in the supplementary.", + "bbox": [ + 89, + 717, + 483, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Our Coupled Product Graph Formalism", + "text_level": 1, + "bbox": [ + 89, + 818, + 437, + 834 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "On a high-level, we use the product graph formalism introduced in [48] to address the matching between individual surface cycles and shape $\\mathcal{V}$ , which results in geometric consistency along the path within each individual cycle.", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Further, we introduce coupling constraints that have the effect of glueing neighbouring product graphs together. Combined with additional injectivity constraints, our matching is globally geometrically consistent, i.e. it preserves neighbourhoods between surface cycles, see Fig. 4 for an overview.", + "bbox": [ + 511, + 90, + 905, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2.1. Individual Surface Cycle Matching Subproblems", + "text_level": 1, + "bbox": [ + 511, + 194, + 898, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Each individual surface cycle $\\mathcal{C}_i$ of $\\mathcal{X}$ can be matched to shape $\\mathcal{V}$ by finding a cyclic path in their product graph $\\mathcal{P}_i = (\\mathcal{V}_{\\mathcal{P}_i},\\mathcal{E}_{\\mathcal{P}_i})$ . It has been introduced in [48] for the problem of matching a 2D shape to a 3D shape and reads:", + "bbox": [ + 511, + 215, + 905, + 276 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Definition 7 (Product graph $\\mathcal{P}_i$ ). The product graph $\\mathcal{P}_i = (\\mathcal{V}_{\\mathcal{P}_i}, \\mathcal{E}_{\\mathcal{P}_i})$ between the $i$ -th surface cycle $\\mathcal{C}_i$ and the 3D shape $\\mathcal{Y}$ is a directed graph defined as", + "bbox": [ + 511, + 291, + 906, + 338 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {V} _ {\\mathcal {P} _ {i}} = \\mathcal {V} _ {\\mathcal {C} _ {i}} \\times \\mathcal {V} _ {\\mathcal {Y}}, \\\\ \\mathcal {E} _ {\\mathcal {P} _ {i}} = \\left\\{\\left(v, \\bar {v}\\right) \\in \\mathcal {V} _ {\\mathcal {P} _ {i}} \\times \\mathcal {V} _ {\\mathcal {P} _ {i}} \\mid v = \\left( \\begin{array}{l} x \\\\ y \\end{array} \\right), \\bar {v} = \\left( \\begin{array}{l} \\bar {x} \\\\ \\bar {y} \\end{array} \\right) \\right. \\tag {1} \\\\ (x, \\bar {x}) \\in \\mathcal {E} _ {\\mathcal {C} _ {i}}, (y, \\bar {y}) \\in \\mathcal {E} _ {\\mathcal {Y}} ^ {+} \\}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 352, + 905, + 411 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with the extended edge set $\\mathcal{E}_{\\mathcal{V}}^{+} := \\mathcal{E}_{\\mathcal{V}} \\cup \\{(y, y) \\mid y \\in \\mathcal{V}_{\\mathcal{V}}\\}$ .", + "bbox": [ + 511, + 425, + 898, + 444 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The key idea is that each edge $(v,\\bar{v})\\in \\mathcal{E}_{\\mathcal{P}_i}$ in the product graph can be interpreted as a (potential) matching between an edge of $C_i$ and an edge (or a vertex to account for stretching/compression) of shape $\\mathcal{V}$ . As shown in [48], matching surface cycle $C_i$ to shape $\\mathcal{V}$ amounts to solving a cyclic shortest path problem in the product graph $\\mathcal{P}_i$ .", + "bbox": [ + 511, + 457, + 905, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, for our setting of 3D-to-3D shape matching, the neighbourhood between pairs of surface cycles cannot be ensured when considering vanilla shortest path algorithms, since they would solve the $n$ individual surface cycle matching subproblems independently. To tackle this, we consider the linear programming (LP) formalism of cyclic shortest path problems and add additional constraints to couple the individual product graphs $\\mathcal{P}_1,\\ldots ,\\mathcal{P}_n$ , as explained next.", + "bbox": [ + 511, + 549, + 905, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2.2. Subproblem Coupling", + "text_level": 1, + "bbox": [ + 511, + 698, + 714, + 714 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We couple the individual surface cycle matching subproblems $\\mathcal{P}_1, \\ldots, \\mathcal{P}_n$ by glueing them together at opposite edges. To this end, we introduce coupling constraints $\\mathcal{V}_{\\mathcal{L}}$ which serve the purpose of ensuring that matchings of opposite edges are consistent, i.e. resulting matchings (or rather resulting shortest cyclic paths of neighbouring surface cycles) must go through opposite edges, see Fig. 4 (iii). Defacto, this enforces the glueing of neighbouring product graphs, and thus ensures geometric consistency of neighbouring surface cycles. In addition to the coupling, we want to ensure injectivity for each surface cycle edge, i.e. each such edge is matched exactly once.", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "915", + "bbox": [ + 485, + 944, + 511, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5abe4fc4747faf4a0e3de82d6ad423af0d73645707f5a72b3135d03496466f81.jpg", + "image_caption": [ + "Figure 4. Conceptual summary of our formalism. Left: Illustration of product graph $\\mathcal{P}_i$ between the surface cycle $\\mathcal{C}_i$ and shape graph $\\mathcal{V}$ . Each black dot inside the big yellow circle represents a product vertex. Overall, the product graph can be thought of having one copy of $\\mathcal{V}$ for each vertex of $\\mathcal{C}_i$ , which are appropriately connected via edges (only a subset of edges are drawn to reduce visual clutter). Middle: Two neighbouring product graphs $\\mathcal{P}_i$ and $\\mathcal{P}_j$ that arise from two neighbouring surface cycles $\\mathcal{C}_i$ and $\\mathcal{C}_j$ . Right: Illustration of constrained neighbouring product graphs. We add additional coupling constraints so that opposite edges of the product graphs (i.e. edges with same colour, see coloured lines ——) are coupled via coupling constraints $q_{ij}^{e_1}, q_{ij}^{e_2}, q_{ij}^{e_3} \\in \\mathcal{V}_{\\mathcal{L}}$ (see green dots $\\bigcirc$ ). Furthermore, we bundle edges of the product graphs so that each edge of the surface cycles is matched exactly once. To this end, we add injectivity constraints $s_{x\\bar{x}}$ , $s_{\\bar{x} x} \\in \\mathcal{V}_{\\mathcal{S}}$ (see blue dots $\\bigcirc$ within bundles of edges)." + ], + "image_footnote": [], + "bbox": [ + 140, + 87, + 330, + 243 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7a4898c7890f4c8b9df7ba737963678f48c438f1634aed4a3f583595cb63b303.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 344, + 85, + 612, + 244 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9dd4d70409dd43ce884b230e381a50e9f4d825b3d49f2a698d96a2483e8b94ef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 85, + 883, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.3. Surface Cycle Matching Injectivity", + "text_level": 1, + "bbox": [ + 89, + 390, + 379, + 406 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To enforce that each surface cycle edge is matched exactly once, i.e. it is matched to exactly one edge (or vertex) of $\\mathcal{V}$ , we introduce injectivity constraints $\\mathcal{V}_{\\mathcal{S}}$ . For that, let us denote the set of all edges (in the product graph) that are potential matching candidates of the edge $(x,\\bar{x})\\in \\mathcal{E}_{\\mathcal{C}_i}$ as the edge bundle of $(x,\\bar{x})$ . For each such edge bundle, we introduce one injectivity constraint (see Fig. 4 (iii)), which has the purpose to ensure that only a single edge of each edge bundle is part of the final matching.", + "bbox": [ + 89, + 409, + 483, + 546 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Constraint Matrices", + "text_level": 1, + "bbox": [ + 89, + 556, + 282, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As mentioned, we use the LP formalism of the independent cyclic shortest path problems and add additional constraints for coupling an matching injectivity. To this end, we consider the constraint matrix $H = \\left[ \\begin{array}{c} P \\\\ L \\\\ S \\end{array} \\right]$ consisting of submatrices $P, L$ and $S$ . Here, submatrix $P$ represents the collection of incidence matrices of individual product graphs, submatrix $L$ the couplings, and submatrix $S$ the injectivity components.", + "bbox": [ + 89, + 578, + 483, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The matrix $P$ . The vertex edge incidence matrix $P \\in \\{-1,0,1\\}^{c \\times m}$ (with $m = |\\mathcal{E}_{\\mathcal{H}}|$ columns, i.e. one column for each product edge, and $c = \\sum_{i=1}^{n} |\\mathcal{V}_{\\mathcal{C}_i}| |\\mathcal{V}_{\\mathcal{Y}}|$ many rows) represents all $n$ product graphs and is defined as $P := \\mathrm{diag}(P_1, \\ldots, P_n)$ , i.e. it contains the vertex edge incidence matrices $P_i$ of individual product graphs $\\mathcal{P}_i$ as diagonal blocks. To illustrate the structure of the matrix representation of one $\\mathcal{P}_i$ , we now discuss the simplified case of non-degenerate matchings (i.e. we do not allow edge to vertex matchings, see Sec. A in the supplementary for more details). In this case, the incidence matrix $P_i$ reads", + "bbox": [ + 89, + 705, + 483, + 872 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP _ {i} := C _ {i} ^ {+} \\otimes Y ^ {+} - C _ {i} ^ {-} \\otimes Y ^ {-}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 883, + 482, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with $\\otimes$ being the Kroenecker product and with $C_i \\in \\{-1,0,1\\}^{|\\mathcal{V}_{c_i}| \\times |\\mathcal{E}_{c_i}^{\\prime}|\\}$ and $Y \\in \\{-1,0,1\\}^{|\\mathcal{V}_{\\mathcal{Y}}| \\times |\\mathcal{E}_{\\mathcal{Y}}^{\\prime}|\\}$ being the incidence matrices of surface cycle $C_i$ and shape graph $\\mathcal{G}_{\\mathcal{Y}}$ respectively (we note that $C_i$ has exactly two non-zero of opposite sign in each row and column since it is the incidence matrix of a cycle). We use the notation $+$ and $-$ to split the incidence matrix into incoming and outgoing incidence matrix, respectively. In other words, $C_i^+$ and $C_i^-$ as well as $Y^+$ and $Y^-$ are the non-negative and non-positive entries of the incidence matrices $C_i$ and $Y$ , respectively. We note that the sign splitting of incidence matrices $C_i$ and $Y$ is necessary to account for proper edge directions in the resulting product graph. By definition, their edge directions induce edge directions of product edges of respective product graph $\\mathcal{P}_i$ . Each $P_i$ has a block structure with $|\\mathcal{V}_{\\mathcal{Y}}| \\times |\\mathcal{E}_{\\mathcal{Y}}|$ sized blocks and each block contains either only $+1$ or $-1$ entries. Further, each block contains exactly one non-zero per column, see Fig. A.1 in the supplementary.", + "bbox": [ + 511, + 390, + 906, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The matrix $L$ . The coupling of opposite edges of neighbouring product graphs is captured in the matrix", + "bbox": [ + 511, + 662, + 905, + 691 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL := K \\otimes I _ {\\left| \\varepsilon_ {y} ^ {+} \\right|}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 651, + 703, + 903, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here $K \\in \\{-1,0,1\\}^{p \\times |\\mathcal{E}_{\\mathcal{X}}|}$ ( $p$ is the number of undirected non-boundary edges of $\\mathcal{X}$ ) is an incidence matrix which represents the incidence of opposite edges across all $n$ surface cycles. $K$ has at most one non-zero element per column and exactly two non-zeros per row (with opposite sign) and thus $L$ has identical block structure with non-negative and non-positive blocks respectively. We note that the definition of $P$ requires columns of the non-positive blocks of $L$ to be permuted (so that it maps opposite edges of neighbouring product graphs), see Sec. A in the supplementary.", + "bbox": [ + 511, + 733, + 906, + 885 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The matrix $S$ . The bundling of edges of all $n$ product", + "bbox": [ + 532, + 885, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "916", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4acadaf178973b318df2a9ad57d35170319dfecec43d7f341715c10441167fc2.jpg", + "image_caption": [ + "Figure 5. Visualisation of edge cost $c((v, \\bar{v}))$ computation for the product edge $(v, \\bar{v})$ . The product edge defines a potential matching between edges $(x, \\bar{x}) \\in \\mathcal{E}_{\\mathcal{X}}$ of shape $\\mathcal{X}$ and $(y, \\bar{y}) \\in \\mathcal{E}_{\\mathcal{Y}}^{+}$ of shape $\\mathcal{Y}$ . Thus, we define the matching cost $c((v, \\bar{v}))$ as feature difference between features $\\| f_x - f_y \\|$ and $\\| f_{\\bar{x}} - f_{\\bar{y}} \\|$ at respective source vertices $x, y$ and target vertices $\\bar{x}, \\bar{y}$ ." + ], + "image_footnote": [], + "bbox": [ + 133, + 88, + 444, + 189 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "graphs can be described using the matrix", + "bbox": [ + 89, + 310, + 362, + 325 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nS := \\boldsymbol {I} _ {\\left| \\mathcal {E} _ {x} \\right|} \\otimes 1 _ {\\left| \\mathcal {E} _ {y} ^ {+} \\right|} ^ {T}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 334, + 482, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, $S$ has block diagonal structure consisting of $|\\mathcal{E}_{\\mathcal{X}}|$ many non-negative blocks (i.e. $1_{|\\mathcal{E}_{\\mathcal{Y}}^{+}|}^{T}$ blocks). We note that $Sx = 1$ enforces matching injectivity by enumerating product edges such that all $|\\mathcal{E}_{\\mathcal{Y}}^{+}|$ -many matching candidates of a single surface cycle are grouped.", + "bbox": [ + 89, + 369, + 483, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Resulting Integer Linear Program", + "text_level": 1, + "bbox": [ + 89, + 458, + 390, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use the previously described constraint matrix $H$ to cast geometrically consistent shape matching as a linear program which essentially represents $n$ many coupled cyclic shortest path problems.", + "bbox": [ + 89, + 479, + 483, + 541 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Matching costs. The matching cost $c(e)$ of product edge $e = \\left(\\binom{x}{y},\\binom{\\bar{x}}{\\bar{y}}\\right)$ measures how well the edge $(x,\\bar{x})\\in \\mathcal{E}_{\\mathcal{X}}$ of shape $\\mathcal{X}$ and the edge $(y,\\bar{y})\\in \\mathcal{E}_{\\mathcal{Y}}^{+}$ of shape $\\mathcal{V}$ fit together, e.g. by comparing geometric properties or feature descriptors, see also Fig. 5.", + "bbox": [ + 89, + 541, + 483, + 617 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Optimisation problem. We collect matching costs of all $m$ edges of the $n$ product graphs in the vector $c \\in \\mathbb{R}_+^m$ . With that, we can use an indicator representation $x$ for each edge ( $x_k = 1$ means that $k$ -th edge is part of the final matching) so that our matching formalism reads", + "bbox": [ + 89, + 618, + 483, + 694 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {x \\in \\{0, 1 \\} ^ {m}} c ^ {T} x \\quad \\text {s . t .} \\quad H x = b. \\qquad \\mathrm {(G e C o 3 D)}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 703, + 482, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, $H$ is the previously introduced constraint matrix and $b \\coloneqq \\left[0_{|\\mathcal{V}_{\\mathcal{H}}| - |\\mathcal{V}_{S}|}; 1_{|\\mathcal{V}_{S}|}\\right]$ is a column vector with all zeros except for ones in rows belonging to rows of submatrix $S$ of $H$ . We solve (GeCo3D) by considering its LP-relaxation, i.e. we replace $x \\in \\{0,1\\}^m$ with $x \\in [0,1]^m$ . We empirically observe in our experiments, that all solutions are integral and consequently globally optimal.", + "bbox": [ + 89, + 739, + 483, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Lemma 8. Matchings between shapes $\\mathcal{X}$ and $\\mathcal{Y}$ obtained by solving (GeCo3D) are globally geometrically consistent according to Definition 4.", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Proof. We represent each triangle of the source shape $\\mathcal{X}$ using a surface cycle. Furthermore, matchings obtained by solving (GeCo3D) preserve neighbourhood relations along the path of a matched surface cycle as well as between the surface cycles. Overall, this ensures that connected vertices on $\\mathcal{X}$ are only matched to connected vertices (or the same vertex) on shape $\\mathcal{Y}$ .", + "bbox": [ + 511, + 90, + 906, + 198 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.5. Hyper Product Graph Interpretation", + "text_level": 1, + "bbox": [ + 511, + 213, + 834, + 229 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our optimisation problem (GeCo3D) can be interpreted as finding a minimum-cost flow circulation in a directed hyper product graph $\\mathcal{H}$ , see [6] for definitions of flows in directed hyper graphs. Here, $\\mathcal{H}$ can be obtained by interpreting $H$ as a vertex hyper edge incidence matrix, i.e. by interpreting rows of $H$ as vertices and columns of $H$ as directed hyper edges, see Fig. 6 for an illustration. For a formal definition of $\\mathcal{H}$ we refer to Sec. B in the supplementary.", + "bbox": [ + 511, + 236, + 906, + 357 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In general, such flow circulation problems on hypergraphs are NP-hard [5]. However, there are certain subclasses that are known to be solvable in polynomial time [10, 29, 42, 80, 84]. Unfortunately, the specific structure of our hyper graph is not listed among the known polynomial-time solvable subclasses [10, 29, 42, 80, 84]. Nevertheless, we empirically observe that all instances considered in our experiments yield an integral (and thus globally optimal) solution when solved using LP-relaxations.", + "bbox": [ + 511, + 358, + 908, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 510, + 645, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we experimentally evaluate our method's performance for 3D shape matching. Further, we show as proof of concept that our method is applicable to planar graph matching.", + "bbox": [ + 511, + 537, + 906, + 598 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Setup. We conduct runtime experiments on an Intel Core i9 12900K with 128 GB DDR5 RAM. For solving our linear programm (GeCo3D), we use off-the-shelf solver Gurobi [37] (version 10). We explicitly resolve coupling constraints during problem construction to obtain smaller constraint matrices, see Sec. C.2 in the supplementary. To better account for shrinking and stretching, we integrate a distortion bound, see Sec. C.1 in the supplementary. For", + "bbox": [ + 511, + 598, + 908, + 720 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/451545cb9391a4aeb472c1cf66df41e21d7605e9f77eab77a8b3f891ca91a69d.jpg", + "image_caption": [ + "Figure 6. Illustration of interpreting our constraint matrix $H$ as the vertex hyper edge incidence matrix of a hyper product graph $\\mathcal{H}$ , see Sec. B in the supplementary. To obtain $\\mathcal{H}$ , rows of $H$ are interpreted as vertices, and columns as directed hyper edges of $\\mathcal{H}$ ." + ], + "image_footnote": [], + "bbox": [ + 557, + 739, + 870, + 830 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "917", + "bbox": [ + 485, + 944, + 511, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "quantitative comparison, we decimate shapes to 1000 triangles using algorithms provided in [41]. We use the deep features [13] for all methods except for SmoothShells [28] and DiscrOpt [62] for which we use the original (axiomatic) features as reported in respective papers.", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metrics. We evaluate matching accuracy using geodesic errors, i.e. geodesic distance to ground-truth matchings. We follow the Princeton protocol [43] and normalise errors by the square-root of the shape area (see [43, Sec. 8.2]). Furthermore, we evaluate geometric consistency of matchings using Dirichlet energies, i.e. the deformation energy induced by the matching (see [67, Sec. 8]).", + "bbox": [ + 89, + 169, + 480, + 275 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Methods. We compare the matching quality of the methods ULRSSM [13], SmoothShells [28], DiscrOpt [62] and SpiderMatch [67]. ULRSSM [13] is an unsupervised deep shape matching method achieving best matchings results on numerous benchmarks. Because of its good performance and its extensive evaluation, we consider it to be representative for deep shape matching methods. SmoothShells [28] is a functional map based shape alignment method which iteratively adds geometric information to compute matchings. DiscrOpt [62] is a functional map based method which relates functional maps to point-wise maps. SpiderMatch [67] is the only other method that is scalable and at the same time considers geometric consistency. Furthermore, for runtime evaluations, we only compare to other geometrically consistent methods since these aim to solve a much harder to solve problem (yet most of these methods do not scale well to relevant shape resolutions, see also Fig. 1 right). In this sense, we consider Windheuser et al. [89], SM-Comb [69], DiscomMatch [66] and SpiderMatch (geo.). Windheuser et al. [89] propose a geometrically consistent 3D shape matching formalism which they solve using LP-relaxations and iterative variable fixations. The methods SM-Comb [69] and DiscomMatch [66] aim to solve this formalism approximately on CPU and GPU, respectively. We use SpiderMatch (geo.) to indicate the SpiderMatch [67] method when using a different curve to represent the source shape: we consider a curve that covers all edges of the source shape and thus, contains intersections at every vertex of the source shape. This is necessary to guarantee global geometric consistency since SpiderMatch [67] can only ensure geometric consistency (according to Definition 4) at intersection points of the curve. Thus, SpiderMatch (geo.) yields a stronger notion of geometric consistency compared to SpiderMatch [67].", + "bbox": [ + 91, + 277, + 483, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Datasets. We evaluate shape matching on four different datasets: remeshed FAUST [9, 21, 63] (100 near-isometric deformed human shapes from which we sample 100 test set pairs), SMAL [92] (49 non-isometric deformed animal shapes of eight species from which we sample 100 test set pairs), DT4D-H [57] (9 different classes of humanoid/game character shapes in different poses taken from DeformingThings4D [53] from which we sample 100 intra class test", + "bbox": [ + 89, + 780, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ef86d4f8b8007f5d1bfb3f184cc29d9cb0b726be1008d25f54d86bef6ea63963.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFAUSTSMALDT4D IntraDT4D InterBeCoS
ULRSSM [13]0.0310.0480.0330.0410.057
SmoothShells [28]0.3790.3760.3730.4200.370
DiscrOpt [62]0.1100.2680.0750.1700.270
SpiderMatch [67]0.0290.0440.0270.0410.057
Ours0.0270.0440.0240.0390.056
", + "bbox": [ + 516, + 89, + 897, + 170 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Comparison of mean geodesic errors (↓) of various shape matching methods. We can see that our method consistently outperforms other methods on all five datasets.", + "bbox": [ + 511, + 176, + 903, + 218 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "set pairs and 100 inter class test set pairs) and BeCoS [25] (2543 animal and humanoid shapes from various datasets of which we consider the 141 full-to-full test set pairs).", + "bbox": [ + 511, + 253, + 903, + 297 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1.3D Shape Matching", + "text_level": 1, + "bbox": [ + 511, + 319, + 702, + 335 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the following, we evaluate our method's shape matching performance w.r.t. runtime and matching quality.", + "bbox": [ + 511, + 345, + 903, + 376 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Runtime. In Fig. 1 right, we show runtime comparison to formalisms which consider geometric consistency. Curves show median runtimes over five instances of FAUST dataset. Among all methods, ours and SpiderMatch [67] scale best. They can handle 3D shapes with approximately twice as many vertices as their fastest competitors DiscoMatch [66], which tackles the problem proposed by Windheuser et al. [89]. However, we note that when considering SpiderMatch (geo.) and consequently curves with more intersections (which is necessary for global geometric consistency) the scalability of SpiderMatch (geo.) degrades drastically. This stems from the branch and bound algorithm (having exponential worst-case runtime in general) which is used to solve SpiderMatch [67]. In contrast, we empirically observe (in all of our tested instances) that the linear programming relaxation of (GeCo3D) is always tight (i.e. yields an integral optimal solution) which explains the scalability of our approach. We emphasise that matchings computed from (GeCo3D) are provably geometrically consistent according to Definition 4, while the matchings of SpiderMatch [67] are not (cf. also Tab. 4).", + "bbox": [ + 511, + 378, + 906, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Full Shape Matching. In Tab. 3 and Tab. 4, we quantitatively compare the shape matching performance of various methods. Our method consistently produces best results across all datasets w.r.t. mean geodesic errors and smoothest results w.r.t Dirichlet energies. This showcases the beneficial effect of strong priors induced by geometric consistency. In addition, in Fig. 7, we show qualitative results, which further showcases high-quality matching results computed with our method. We provide more results as well as ablation studies in Sec. D.1 in the supplementary.", + "bbox": [ + 511, + 700, + 903, + 851 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Partial-to-Full Shape Matching. As proof of concept we show qualitative results of matchings computed with our method in the partial-to-full setting, see Fig. 8.", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "918", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a60528368ca34bf380f251a12f7dfcad670f6d748799a9684fd87e443cd610f1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFAUSTSMALDT4D IntraDT4D InterBeCoS
ULRSSM [13]2.12.63.42.64.1
SmoothShells [28]1.92.81.62.35.7
DiscrOpt [62]8.913.67.510.71.1
SpiderMatch [67]1.71.81.72.10.81
Ours0.460.530.480.620.77
", + "bbox": [ + 94, + 88, + 477, + 170 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Comparison of mean Dirichlet energies $(\\downarrow)$ of various shape matching methods to measure smoothness of the matching. Our method consistently yields best results, which shows the effect of strong priors induced by global geometric consistency.", + "bbox": [ + 89, + 180, + 482, + 236 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a016ec165a49cb5a38e9690d8799ed00c5aec061fc538666c804d7e4ee4b0c79.jpg", + "image_caption": [ + "Figure 7. Qualitative shape matching results computed with ULRSSM [13], SpiderMatch [67] and ours. We visualise matchings using colour and triangulation transfer from source to target shape. At legs and arms of shapes we can see stronger geometric consistency of ours compared to SpiderMatch [67] (see red and green arrows and distorted triangles)." + ], + "image_footnote": [], + "bbox": [ + 89, + 241, + 473, + 526 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Application to Planar Graph Matching", + "text_level": 1, + "bbox": [ + 89, + 641, + 426, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As a proof of concept we show that our formalism directly applies to planar graph matching. To this end, we use keypoints on images of the WILLOW [16] dataset and obtain graphs by computing Delaunay triangulations of keypoints. Furthermore, we represent one of the resulting graphs using surface cycles to match it to the other graph, see Fig. 9 and in Sec. D.2 in the supplementary.", + "bbox": [ + 89, + 662, + 483, + 770 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d8a786580d2b7c194417389b964a2002c231d1c82fc7774f4fa0741053f02419.jpg", + "image_caption": [ + "Figure 8. Qualitative results computed with ours for partial-to-full shape matching of test set shapes [26] from SHREC'16 [17]." + ], + "image_footnote": [], + "bbox": [ + 94, + 789, + 480, + 863 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b7ac21ad303a65684758c20e553fbc6dc0c300a439d0d3eab932387da6fead76.jpg", + "image_caption": [ + "Figure 9. Examples of planar graph matching results using our method on ducks and cars from WILLOW [16] dataset." + ], + "image_footnote": [], + "bbox": [ + 514, + 87, + 903, + 167 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Discussion & Limitations", + "text_level": 1, + "bbox": [ + 511, + 232, + 750, + 247 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our algorithm enforces geometric consistency and thus introduces a strong prior to resolve ambiguities in 3D shape matching problems. Yet, in theory our formalism contains undesirable matchings in the solution space: On the one hand, multiple surface cycles could be matched to the same vertex on the target shape. While this ensures our optimisation problem is always feasible (since it contains the extreme case of matching all surface cycles to a single vertex), this is an undesirable solution. In the future, this could be resolved by generalising our approach towards a symmetric formalism. On the other hand, our formalism allows for inside-out flips, as each individual matching element (i.e. an edge) cannot disambiguate extrinsic orientations.", + "bbox": [ + 511, + 258, + 906, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have empirically shown that results can be computed efficiently using off-the-shelf LP-solvers [37] and that in all considered instances the linear the programming relaxations are tight, i.e. we find a globally optimal solution for all instances. Yet, we observe that our constraint matrix is not totally unimodular. We leave an in-depth analysis of our formalism, along with answering the question whether there exists a polynomial time algorithm for geometrically consistent 3D shape matching, for future works.", + "bbox": [ + 511, + 458, + 908, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 614, + 633, + 631 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have presented a novel formalism for non-rigid 3D shape matching which globally enforces geometric consistency. Our key idea is to consider a shape representation that allows to cast 3D shape matching as an integer linear program, which is efficiently solvable in practice. In addition, we illustrate that the resulting problem can be interpreted as a minimum-cost flow circulation problem in a hyper graph. Overall, we consider our work to be important for the 3D shape analysis community. Geometric consistency is crucial in most practical 3D shape matching settings, which is often times ignored (especially by learning-based methods). Our work allows to obtain such matchings efficiently and with that poses a step towards integrating geometric consistency into learning-based methods. Furthermore, we hope to inspire follow-up works on neighbourhood preserving matching formalisms within the broader field of visual computing and beyond.", + "bbox": [ + 511, + 643, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "919", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 243, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We thank Johan Thunberg, Tobias Weißberg and Zorah Lähner for insightful discussions about our formalism. This work is supported by the ERC starting grant no. 101160648 (Harmony).", + "bbox": [ + 89, + 112, + 483, + 174 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 189, + 187, + 205 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] MOSEK ApS. MOSEK Optimizer API for Python 11.0.9, 2025. 5, 6", + "[2] Souhaib Attaiki and Maks Ovsjanikov. Understanding and improving features learned in deep functional maps. In CVPR, 2023. 2", + "[3] Mathieu Aubry, Ulrich Schlickewei, and Daniel Cremers. The wave kernel signature: A quantum mechanical approach to shape analysis. In ICCV workshops. IEEE, 2011. 4, 5", + "[4] Lennart Bastian, Yizheng Xie, Nassir Navab, and Zorah Lahner. Hybrid functional maps for crease-aware nonisometric shape matching. In CVPR, 2024. 2", + "[5] Isabel Beckenbach. A hypergraph network simplex algorithm. In GOR, 2018. 6", + "[6] Isabel Leonie Beckenbach. Matchings and flows in hypergraphs. PhD thesis, 2019. 6", + "[7] Marcel Seelbach Benkner, Zorah Lähner, Vladislav Golyanik, Christof Wunderlich, Christian Theobalt, and Michael Moeller. Q-match: Iterative shape matching via quantum annealing. In ICCV, 2021. 2", + "[8] Florian Bernard, Zeeshan Khan Suri, and Christian Theobalt. Mina: Convex mixed-integer programming for non-rigid shape alignment. In CVPR, 2020. 1, 2", + "[9] Federica Bogo, Javier Romero, Matthew Loper, and Michael J Black. Faust: Dataset and evaluation for 3d mesh registration. In CVPR, 2014. 7", + "[10] Kateřina Böhmová, Jérémie Chalopin, Matúš Mihalák, Guido Proietti, and Peter Widmayer. Sequence hypergraphs: paths, flows, and cuts. Adventures Between Lower Bounds and Higher Altitudes: Essays Dedicated to Juraj Hromković on the Occasion of His 60th Birthday, 2018. 6", + "[11] Alexander M Bronstein, Michael M Bronstein, Ron Kimmel, Mona Mahmoudi, and Guillermo Sapiro. A gromov-hausdorff framework with diffusion geometry for topologically-robust non-rigid shape matching. International Journal of Computer Vision, 2010. 2", + "[12] Oliver Burghard and Reinhard Klein. Efficient lifted relaxations of the quadratic assignment problem. In CVMV, 2017. 2", + "[13] Dongliang Cao, Paul Roetzer, and Florian Bernard. Unsupervised learning of robust spectral shape matching. TOG, 2023. 1, 2, 7, 8, 4, 5, 6", + "[14] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 1", + "[15] Qifeng Chen and Vladlen Koltun. Robust nonrigid registration by convex optimization. In ICCV, 2015. 2", + "[16] Minsu Cho, Karteek Alahari, and Jean Ponce. Learning graphs to match. In ICCV, 2013. 8, 6", + "[17] Luca Cosmo, Emanuele Rodola, Michael M Bronstein, Andrea Torsello, Daniel Cremers, and Y Sahillioglu. Shrec'16: Partial matching of deformable shapes. 3DOR, 2016. 8, 6, 9" + ], + "bbox": [ + 93, + 214, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] James Coughlan, Alan Yuille, Camper English, and Dan Snow. Efficient deformable template detection and localization without user initialization. CVIU, 2000. 2", + "[19] Bailin Deng, Yuxin Yao, Roberto M Dyke, and Juyong Zhang. A survey of non-rigid 3d registration. In CGF, 2022. 2", + "[20] Nicolas Donati, Etienne Corman, and Maks Ovsjanikov. Deep orientation-aware functional maps: Tackling symmetry issues in shape matching. In CVPR, 2022. 2", + "[21] Nicolas Donati, Abhishek Sharma, and Maks Ovsjanikov. Deep geometric functional maps: Robust feature learning for shape correspondence. In CVPR, 2020. 7", + "[22] Niladri Shekhar Dutt, Sanjeev Muralikrishnan, and Niloy J. Mitra. Diffusion 3d features (diff3f): Decorating textured shapes with distilled semantic features. In CVPR, 2024. 2, 4, 5", + "[23] Nadav Dym, Haggai Maron, and Yaron Lipman. DS++ - A Flexible, Scalable and Provably Tight Relaxation for Matching Problems. TOG, 2017. 2", + "[24] Bernhard Egger, William A. P. Smith, Ayush Tewari, Stefanie Wuhrer, Michael Zollhoefer, Thabo Beeler, Florian Bernard, Timo Bolkart, Adam Kortylewski, Sami Romdhani, Christian Theobalt, Volker Blanz, and Thomas Vetter. 3d morphable face models—past, present, and future. TOG, 2020. 1", + "[25] Viktoria Ehm, Nafie El Amrani, Yizheng Xie, Lennart Bastian, Maolin Gao, Weikang Wang, Lu Sang, Dongliang Cao, Tobias Weiberg, Zorah Lahner, Daniel Cremers, and Florian Bernard. Beyond Complete Shapes: A Benchmark for Quantitative Evaluation of 3D Shape Matching Algorithms. CGF, 2025. 7", + "[26] Viktoria Ehm, Maolin Gao, Paul Roetzer, Marvin Eisenberger, Daniel Cremers, and Florian Bernard. Partial-to-partial shape matching with geometric consistency. In CVPR, 2024. 3, 8", + "[27] Viktoria Ehm, Paul Roetzer, Marvin Eisenberger, Maolin Gao, Florian Bernard, and Daniel Cremers. Geometrically consistent partial shape matching. In 3DV, 2024. 3, 9", + "[28] Marvin Eisenberger, Zorah Lahner, and Daniel Cremers. Smooth shells: Multi-scale shape registration with functional maps. In CVPR, 2020. 2, 7, 8, 6", + "[29] James R Evans and John J Jarvis. Network topology and integral multicommodity flow problems. Networks, 1978. 6", + "[30] Danielle Ezuz and Mirela Ben-Chen. Deblurring and denoising of maps between shapes. In CGF, 2017. 2", + "[31] Danielle Ezuz, Behrend Heeren, Omri Azencot, Martin Rumpf, and Mirela Ben-Chen. Elastic correspondence between triangle meshes. In CGF, 2019. 2", + "[32] Pedro F Felzenszwalb. Representation and detection of deformable shapes. TPAMI, 2005. 2", + "[33] Giorgio Gallo, Giustino Longo, Stefano Pallottino, and Sang Nguyen. Directed hypergraphs and applications. Discrete applied mathematics, 1993. 2", + "[34] Maolin Gao, Paul Roetzer, Marvin Eisenberger, Zorah Lahner, Michael Moeller, Daniel Cremers, and Florian Bernard. Sigma: Scale-invariant global sparse shape matching. In ICCV, 2023. 1, 2", + "[35] Andrea Gasparetto, Luca Cosmo, Emanuele Rodola, Michael Bronstein, and Andrea Torsello. Spatial maps: From" + ], + "bbox": [ + 516, + 92, + 903, + 887 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "920", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "low rank spectral to sparse spatial functional representations. In 3DV, 2017. 2", + "[36] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. 3d-coded: 3d correspondences by deep deformation. In ECCV, 2018. 2", + "[37] Gurobi Optimization, LLC. Gurobi Optimizer Reference Manual, 2023. 6, 8, 5", + "[38] Oshri Halimi, Or Litany, Emanuele Rodola, Alex M Bronstein, and Ron Kimmel. Unsupervised learning of dense shape correspondence. In CVPR, 2019. 2", + "[39] Stefan Haller, Lorenz Feineis, Lisa Hutschenreiter, Florian Bernard, Carsten Rother, Dagmar Kainmüller, Paul Swoboda, and Bogdan Savchynskyy. A comparative study of graph matching algorithms in computer vision. In ECCV, 2022. 1", + "[40] Benjamin Holzschuh, Zorah Lähner, and Daniel Cremers. Simulated annealing for 3d shape correspondence. In 3DV, 2020. 2", + "[41] Alec Jacobson, Daniele Panozzo, et al. libigl: A simple C++ geometry processing library, 2018. https://libigl.github.io/. 7", + "[42] Robert G Jeroslow, Kipp Martin, Ronald L Rardin, and Jinchang Wang. Gainfree leontief substitution flow problems. Mathematical Programming, 1992. 6", + "[43] Vladimir G Kim, Yaron Lipman, and Thomas Funkhouser. Blended intrinsic maps. TOG, 2011. 7", + "[44] Whoi-Yul Kim and Avinash C. Kak. 3-d object recognition using bipartite matching embedded in discrete relaxation. IEEE Transactions on Pattern Analysis & Machine Intelligence, 1991. 1", + "[45] Eugene Krissinel and Kim Henrick. Secondary-structure matching (ssm), a new tool for fast protein structure alignment in three dimensions. Biological Crystallography, 2004. 1", + "[46] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 1955. 2", + "[47] Yam Kushinsky, Haggai Maron, Nadav Dym, and Yaron Lipman. Sinkhorn Algorithm for Lifted Assignment Problems. SIAM Journal on Imaging Sciences, 2019. 2", + "[48] Zorah Lähner, Emanuele Rodola, Frank R Schmidt, Michael M Bronstein, and Daniel Cremers. Efficient globally optimal 2d-to-3d deformable shape matching. In CVPR, 2016. 2, 4, 1, 3", + "[49] Laura Leal-Taixe and Bodo Rosenhahn. Pedestrian interaction in tracking: the social force model and global optimization methods. In Modeling, Simulation and Visual Analysis of Crowds: A Multidisciplinary Perspective. 2013. 2", + "[50] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. In ECCV, 2024. 1", + "[51] Lei Li, Nicolas Donati, and Maks Ovsjanikov. Learning multi-resolution functional maps with spectral attention for robust shape matching. NIPS, 2022. 2", + "[52] Qinsong Li, Shengjun Liu, Ling Hu, and Xinru Liu. Shape correspondence using anisotropic chebyshev spectral cnns. In CVPR, 2020. 2", + "[53] Yang Li, Hikari Takehara, Takafumi Taketomi, Bo Zheng, and Matthias Nießner. 4dcomplete: Non-rigid motion estimation beyond the observable surface. In ICCV, 2021. 7", + "[54] Or Litany, Tal Remez, Emanuele Rodola, Alex Bronstein, and Michael Bronstein. Deep functional maps: Structured prediction for dense shape correspondence. In ICCV, 2017." + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "2", + "[55] Jiayi Ma, Xingyu Jiang, Aoxiang Fan, Junjun Jiang, and Junchi Yan. Image matching from handcrafted to deep features: A survey. IJCV, 2021. 1", + "[56] Filippo Maggioli, Daniele Baieri, Emanuele Rodola, and Simone Melzi. Rematching: Low-resolution representations for scalable shape correspondence. In ECCV, 2024. 2", + "[57] Robin Magnet, Jing Ren, Olga Sorkine-Hornung, and Maks Ovsjanikov. Smooth non-rigid shape matching via effective dirichlet energy optimization. In 3DV, 2022. 7", + "[58] Haggai Maron, Nadav Dym, Itay Kezurer, Shahar Kovalsky, and Yaron Lipman. Point registration via efficient convex relaxation. TOG, 2016. 1, 2", + "[59] Maks Ovsjanikov, Mirela Ben-Chen, Justin Solomon, Adrian Butscher, and Leonidas Guibas. Functional maps: a flexible representation of maps between shapes. TOG, 2012. 2", + "[60] Maks Ovsjanikov, Quentin Merigot, Facundo Memoli, and Leonidas Guibas. One point isometric matching with the heat kernel. In CGF, 2010. 2", + "[61] Thomas Probst, Danda Pani Paudel, Ajad Chhatkuli, and Luc Van Gool. Unsupervised learning of consensus maximization for 3d vision problems. In CVPR, 2019. 2", + "[62] Jing Ren, Simone Melzi, Peter Wonka, and Maks Ovsjanikov. Discrete optimization for shape matching. In CGF, 2021. 2, 7, 8, 6", + "[63] Jing Ren, Adrien Poulenard, Peter Wonka, and Maks Ovsjanikov. Continuous and orientation-preserving correspondences via functional maps. TOG, 2018. 7", + "[64] F Rendl, P Pardalos, and H Wolkowicz. The Quadratic Assignment Problem: A Survey and Recent Developments. In DIMACS workshop, 1994. 1, 2", + "[65] Emanuele Rodola, Alex M Bronstein, Andrea Albarelli, Filippo Bergamasco, and Andrea Torsello. A game-theoretic approach to deformable shape matching. In CVPR, 2012. 2", + "[66] Paul Roetzer, Ahmed Abbas, Dongliang Cao, Florian Bernard, and Paul Swoboda. Discomatch: Fast discrete optimisation for geometrically consistent 3d shape matching. In ECCV, 2024. 3, 7", + "[67] Paul Roetzer and Florian Bernard. Spidermatch: 3d shape matching with global optimality and geometric consistency. In CVPR, 2024. 2, 3, 7, 8, 6", + "[68] Paul Roetzer, Zorah Lähner, and Florian Bernard. Conjugate product graphs for globally optimal 2d-3d shape matching. In CVPR, 2023. 2", + "[69] Paul Roetzer, Paul Swoboda, Daniel Cremers, and Florian Bernard. A scalable combinatorial solver for elastic geometrically consistent 3d shape matching. In CVPR, 2022. 3, 7", + "[70] David Sankoff and Joseph B Kruskal. Time warps, string edits, and macromolecules: the theory and practice of sequence comparison. 1983. 2", + "[71] Frank R. Schmidt, Eno Töppe, and Daniel Cremers. Efficient planar graph cuts with applications in computer vision. In CVPR, 2009. 2", + "[72] Patrick Schmidt, Marcel Campen, Janis Born, and Leif Kobbelt. Inter-surface maps via constant-curvature metrics. TOG, 39(4):119-1, 2020. 2", + "[73] Patrick Schmidt, Dörte Pieper, and Leif Kobbelt. Surface maps via adaptive triangulations. CGF, 2023. 2", + "[74] Thomas Schoenemann and Daniel Cremers. A combinatorial" + ], + "bbox": [ + 516, + 93, + 903, + 896 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "921", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "solution for model-based image segmentation and real-time tracking. TPAMI, 2009. 2, 3", + "[75] John Schreiner, Arul Asirvatham, Emil Praun, and Hugues Hoppe. Inter-surface mapping. In SIGGRAPH. 2004. 2", + "[76] Alexander Schrijver et al. Combinatorial optimization: polyhedra and efficiency. Springer, 2003. 1", + "[77] Avinash Sharma, Radu Horaud, Jan Cech, and Edmond Boyer. Topologically-robust 3d shape matching based on diffusion geometry and seed growing. In CVPR, 2011. 2", + "[78] Nicholas Sharp, Souhaib Attaiki, Keenan Crane, and Maks Ovsjanikov. Diffusionnet: Discretization agnostic learning on surfaces. TOG, 2022. 2", + "[79] Justin Solomon, Gabriel Peyré, Vladimir G Kim, and Suvrit Sra. Entropic metric alignment for correspondence problems. TOG, 2016. 2", + "[80] Y Soun and Klaus Truemper. Single commodity representation of multicommodity networks. SIAM Journal on Algebraic Discrete Methods, 1980. 6", + "[81] Kenshi Takayama. Compatible intrinsic triangulations. TOG, 2022. 2", + "[82] Gary KL Tam, Zhi-Quan Cheng, Yu-Kun Lai, Frank C Langbein, Yonghuai Liu, David Marshall, Ralph R Martin, Xian-Fang Sun, and Paul L Rosin. Registration of 3d point clouds and meshes: A survey from rigid to nonrigid. TVCG, 2012. 2", + "[83] Giovanni Trappolini, Luca Cosmo, Luca Moschella, Ricardo Marin, Simone Melzi, and Emanuele Rodola. Shape registration in the time of transformers. NeurIPS, 2021. 2", + "[84] Klaus Truemper. Unimodular matrices of flow problems with additional constraints. Networks, 1977. 6", + "[85] Oliver Van Kaick, Hao Zhang, Ghassan Hamarneh, and Daniel Cohen-Or. A survey on shape correspondence. In CGF, 2011. 2", + "[86] Matthias Van Onsem, Dieter De Paepe, S Vanden Hautte, Pieter Bonte, Veerle Ledoux, Annelies Lejon, Femke Ongenae, Dennis Dreesen, and Sofie Van Hoecke. Hierarchical pattern matching for anomaly detection in time series. Computer Communications, 2022. 1", + "[87] Matthias Vestner, Roee Litman, Emanuele Rodola, Alex Bronstein, and Daniel Cremers. Product manifold filter: Non-rigid shape correspondence via kernel density estimation in the product space. In CVPR, 2017. 2", + "[88] Ruben Wiersma, Elmar Eisemann, and Klaus Hildebrandt. Cnns on surfaces using rotation-equivariant features. TOG, 2020. 2", + "[89] Thomas Windheuser, Ulrich Schlickewei, Frank R Schmidt, and Daniel Cremers. Geometrically consistent elastic matching of 3d shapes: A linear programming solution. In ICCV, 2011. 1, 2, 7, 6", + "[90] Thomas Windheuser, Ulrich Schlickwei, Frank R Schimdt, and Daniel Cremers. Large-scale integer linear programming for orientation preserving 3d shape matching. In CGF, 2011. 2", + "[91] Junchi Yan, Xu-Cheng Yin, Weiyao Lin, Cheng Deng, Hongyuan Zha, and Xiaokang Yang. A short survey of recent advances in graph matching. In ICMR, 2016. 1", + "[92] Silvia Zuffi, Angjoo Kanazawa, David W Jacobs, and Michael J Black. 3d menagerie: Modeling the 3d shape and pose of animals. In CVPR, 2017. 7" + ], + "bbox": [ + 93, + 92, + 480, + 885 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "922", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Fast Globally Optimal and Geometrically Consistent 3D Shape Matching/cbbb1f39-340b-4752-9244-371644b15edf_model.json b/2025/Fast Globally Optimal and Geometrically Consistent 3D Shape Matching/cbbb1f39-340b-4752-9244-371644b15edf_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6260371cc5dbd19b0f9128fa88ec585d33a51337 --- /dev/null +++ b/2025/Fast Globally Optimal and Geometrically Consistent 3D Shape Matching/cbbb1f39-340b-4752-9244-371644b15edf_model.json @@ -0,0 +1,2840 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.131, + 0.131, + 0.868, + 0.154 + ], + "angle": 0, + "content": "Fast Globally Optimal and Geometrically Consistent 3D Shape Matching" + }, + { + "type": "text", + "bbox": [ + 0.322, + 0.18, + 0.493, + 0.217 + ], + "angle": 0, + "content": "Paul Roetzer1,2 \n1 University of Bonn" + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.18, + 0.671, + 0.216 + ], + "angle": 0, + "content": "Florian Bernard\\(^{1,2}\\) \n\\(^{2}\\) Lamarr Institute" + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.29, + 0.168, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.173, + 0.299, + 0.245, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.299, + 0.334, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.156, + 0.381, + 0.3, + 0.395 + ], + "angle": 0, + "content": "Key Idea of Our Approach" + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.248, + 0.425, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.434, + 0.248, + 0.505, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.247, + 0.582, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.35, + 0.371, + 0.605, + 0.395 + ], + "angle": 0, + "content": "Source Cao et al. Ours Triangulation Transfer via Computed Matchings" + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.253, + 0.898, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.66, + 0.383, + 0.875, + 0.394 + ], + "angle": 0, + "content": "Runtime Comparison on FAUST Dataset" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.403, + 0.907, + 0.501 + ], + "angle": 0, + "content": "Figure 1. Left: Illustration of our approach for geometrically consistent matching of 3D shape \\(\\mathcal{X}\\) to 3D shape \\(\\mathcal{Y}\\). We represent shape \\(\\mathcal{X}\\) using \\(n\\) surface cycles \\(\\mathcal{C}_1, \\ldots, \\mathcal{C}_n\\) and then consistently match these to shape \\(\\mathcal{Y}\\) while preserving neighbourhood relations between the cycles. Middle: The triangulation- and colour-transfer between shapes using computed matchings of Cao et al. [13] (not geometrically consistent) and ours (geometrically consistent) illustrates the importance of geometric consistency. Right: Our method is the first 3D shape matching approach that yields globally geometrically consistent matchings, that is globally optimal, and scalable in practice, opposed to all competing methods (OOM stands for 'out of memory' and the dashed line indicates a weak notion of geometric consistency, see explanation of methods in Sec. 5)." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.521, + 0.327, + 0.536 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.555, + 0.485, + 0.856 + ], + "angle": 0, + "content": "Geometric consistency, i.e. the preservation of neighbourhoods, is a natural and strong prior in 3D shape matching. Geometrically consistent matchings are crucial for many downstream applications, such as texture transfer or statistical shape modelling. Yet, in practice, geometric consistency is often overlooked, or only achieved under severely limiting assumptions (e.g. a good initialisation). In this work, we propose a novel formalism for computing globally optimal and geometrically consistent matchings between 3D shapes which is scalable in practice. Our key idea is to represent the surface of the source shape as a collection of cyclic graphs, which are then consistently matched to the target shape. Mathematically, we construct a hyper product graph (between source and target shape), and then cast 3D shape matching as a minimum-cost circulation flow problem in this hyper graph, which yields global geometrically consistent matchings between both shapes. We empirically show that our formalism is efficiently solvable and that it leads to high-quality results. Our code is publicly available.1" + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.521, + 0.645, + 0.536 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.55, + 0.907, + 0.869 + ], + "angle": 0, + "content": "The availability of correspondences between visual data are a key prerequisite for many visual computing tasks, including 3D reconstruction [50], loss computation for deep learning [14], protein alignment [45], anomaly detection [86], object recognition [44], shape modelling [24] and others. In many practical cases, it is desirable that correspondences preserve neighbourhood relations. For example, in 3D shape matching this could mean that when bringing a pair of neighbouring points on a source shape into correspondence with a pair of points on a target shape, the pair of points must remain neighbours on the target. Such a neighbourhood preservation can serve as strong topological prior, for example to resolve ambiguities, achieve robustness, or ensure well-posedness of an otherwise ill-posed problem. However, despite its crucial importance, neighbourhood preservation is in practice often-times overlooked. This is because many of the well-known formalisms are not efficiently solvable (e.g. the quadratic assignment problem [64], or general integer programming [76], which have been used to tackle image keypoint matching [55], graph matching [39, 91], or 3D shape matching [8, 34, 58, 89])." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.872, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In this work we specifically focus on the task of non-rigid geometrically consistent 3D shape matching, in which" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.888, + 0.378, + 0.902 + ], + "angle": 0, + "content": "1https://github.com/paul0noah/geco" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "912" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.122, + 0.089, + 0.454, + 0.205 + ], + "angle": 0, + "content": "
MethodGlobally OptimalGeometr. ConsistentScalable
MINA [8](✓)✗✗
SIGMA [34](✓)✗✗
PMSDP [58]✗✗✓
Windheuser et al. [89]✗✓✗
SpiderMatch [67]✓(✓)(✓)
Ours✓✓✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.094, + 0.216, + 0.479, + 0.23 + ], + "angle": 0, + "content": "Table 1. Comparison of axiomatic 3D shape matching methods." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.249, + 0.483, + 0.535 + ], + "angle": 0, + "content": "neighbourhood-preserving correspondences between two given 2D manifolds (embedded in 3D space) are sought for. So far, there does not exist any 3D shape matching approach that combines the following three desirable properties: (i) global optimality, (ii) neighbourhood preservation, and (iii) scalability, see Tab. 1. While the recent approach SpiderMatch [67] fulfils (i)-(iii) to some extent, its scalability relies heavily on the choice of the cycle that is used to represent one of the shapes (cf. Fig. 1 right: solid vs. dashed purple line), thereby trading off neighbourhood preservation with scalability. Furthermore, proper geometric consistency (Definition 4) can only be enforced at intersection points of the cycle and thus also the quality of geometric consistency depends on the choice of the cycle. Inspired by this idea of representing a shape using a cycle, we represent one shape using a collection of cycles and show that this leads to an (in practice) efficiently solvable formalism which ensures proper geometric consistency. We summarise our main contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.537, + 0.483, + 0.582 + ], + "angle": 0, + "content": "- For the first time, we present a globally geometrically consistent formalism for 3D shape matching that is efficiently solvable to global optimality in practice." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.582, + 0.483, + 0.658 + ], + "angle": 0, + "content": "- To achieve this, we introduce a novel 3D shape representation in which we represent a 3D shape using a collection of surface cycles, so that 3D shape matching can be cast as finding a minimum-cost flow circulation in a hyper product graph." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.658, + 0.483, + 0.703 + ], + "angle": 0, + "content": "- We experimentally show that our formalism leads to high quality and geometrically consistent matchings between two 3D shapes." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.704, + 0.483, + 0.763 + ], + "angle": 0, + "content": "- In addition to 3D shape matching, we also show, in a proof-of-concept manner, that our formalism can generalise to specific instances of other matching problems, such as graph matching when the source graph is planar." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.537, + 0.483, + 0.763 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.782, + 0.233, + 0.798 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.809, + 0.483, + 0.869 + ], + "angle": 0, + "content": "In the following we discuss works that are most relevant to our approach. We start by discussing efficiently solvable matching problems, continue with 3D shape matching, and conclude with geometrically consistent 3D shape matching." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Efficiently Solvable Matching Problems. There are various instances of correspondence problems that can be" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.394 + ], + "angle": 0, + "content": "solved efficiently. Among them is the linear assignment problem (LAP) [46], which matches points without considering their neighbourhood relations. For matching time series data, sequences, or (open) contours, the popular dynamic time warping algorithm can efficiently compute solutions while preserving neighbourhood relations [70]. Closed contours can be matched analogously via graph cuts in product graphs [71]. Certain tracking problems can be formulated as efficiently solvable flow problems [49], and model-based image segmentation can also be solved efficiently by matching a 2D contour to an image [18, 32, 74]. The matching of a 2D contour to a 3D shape can be solved using variants of Dijkstra's algorithm for finding minimum cost cycles in product graphs [48, 68]. The mentioned approaches show that there are formalisms to efficiently solve a diverse range of matching problems. Yet, these approaches do not directly generalise to 3D shape matching. In this work, we propose a novel geometrically consistent 3D shape matching formalism that is (in practice) efficiently solvable to global optimality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.398, + 0.905, + 0.625 + ], + "angle": 0, + "content": "3D Shape Matching is the task of finding correspondences between two non-rigidly deformed surfaces. For an in-depth overview on 3D shape matching we refer the reader to survey papers [19, 82, 85]. Many axiomatic shape matching approaches [11, 60, 65] build on the functional maps framework, efficiently solving shape matching in the spectral domain [28, 35, 56, 58, 59, 62]. In addition, functional maps have been adopted in several deep shape matching variants, either trained in a supervised [36, 52, 54, 83, 88] or in a unsupervised manner [2, 4, 13, 20, 22, 38, 51, 78]. There are also alternative approaches based on consensus maximisation [61], or using convex relaxations [15, 58] or mixed integer programming [8, 34]. Yet, many of the existing 3D shape matching methods neglect geometric consistency, as it leads to hard-to-solve formalisms." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Geometrically Consistent 3D Shape Matching. There are several approaches that have recognised the importance of geometric consistency in 3D shape matching. Some of them incorporate neighbourhood information by using formalisms based on the quadratic assignment problem (QAP) - yet, the QAP is NP-hard [64], so that respective approaches consider relaxations [12, 23, 47] or heuristics [7, 40, 79]. Opposed to such discrete formulations, there are also approaches that tackle 3D shape matching by deforming a continuous shape parametrisation using local optimisation. However, due to the severe nonconvexity of resulting problems, they rely on a good initialisation, e.g. in the form of sparse sets of landmark correspondences [72, 73, 75, 77, 81], or in the form of dense correspondences [30, 31, 87]. Windheuser et al. [89, 90] have modelled geometrically consistent shape matching in the discrete domain by matching triangles to triangles using an (expensive to solve) integer linear program. Re" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "913" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.482, + 0.292 + ], + "angle": 0, + "content": "
SymbolDescription
X=(Vx,FX)3D shape
Gx=(Vx,EX)3D shape graph of X
C=(VC,EC)surface cycle (cycle in Gx)
{C1,...,Cn}Representation of X with n surface cycles
nNumber of surface cycles
Y=(Vy, fy)3D shape Y
Gy=(Vy, Fy)3D shape graph of Y
Pi=(VPi, EPi)Product graph (of C_i and mesh Y)
{P1,...,Pn}Product graph collection (of {C1,...,Cn} and mesh Y)
H=(VH, EH)Hyper product graph (coupled product graphs {P,...,Pn})
mNumber of hyper edges |EH|
HVertex edge incidence matrices of H
" + }, + { + "type": "table_caption", + "bbox": [ + 0.129, + 0.296, + 0.444, + 0.31 + ], + "angle": 0, + "content": "Table 2. Summary of the notation used in this paper." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.336, + 0.485, + 0.685 + ], + "angle": 0, + "content": "cently, approximative solvers [66, 69] and extensions for partial shapes [26, 27] have been proposed. Furthermore, 3D shape matching has recently been formulated as a shortest path problem with additional constraints [67]. In this work, the surface mesh of the source 3D shape is represented using a long self-intersecting curve that traces the shape surface, and is then matched to the other shape while preserving intersections of the curve. While their idea of considering alternative 3D shape representations is promising, in the presented framework proper geometric consistency is only enforced at intersection points of the curve (see Tab. 4). In addition, runtimes grow drastically with an increasing number of intersection points, which can be seen in Fig. 1 right when comparing SpiderMatch [67] (using a curve with few intersections) to SpiderMatch (geo.) (using a curve with many intersections). Inspired by the idea of an alternative path-based 3D shape representation, we represent a 3D shape as a collection of cyclic graphs, and then tackle 3D shape matching by solving coupled matching problems of the individual cyclic paths. Based on the couplings, we are able to ensure global geometry consistency. Furthermore, we empirically observe that the resulting formalism is efficiently solvable to global optimality." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.698, + 0.219, + 0.714 + ], + "angle": 0, + "content": "3. Background" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.723, + 0.483, + 0.782 + ], + "angle": 0, + "content": "We consider the task of finding a matching between a source shape \\(\\mathcal{X}\\) and a target shape \\(\\mathcal{Y}\\) (Sec. 3.1) such that the matching is geometrically consistent (Sec. 3.2). Our main notation is summarised in Tab. 2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.793, + 0.279, + 0.809 + ], + "angle": 0, + "content": "3.1. Shapes and Graphs" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.815, + 0.483, + 0.86 + ], + "angle": 0, + "content": "In the following, we define shapes and other relevant concepts, which are also illustrated in Fig. 2. We consider \\(3D\\) shapes represented as triangular surface mesh:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Definition 1 (3D shape). A 3D shape \\(\\mathcal{X}\\) is defined as a tuple \\((\\mathcal{V}_{\\mathcal{X}},\\mathcal{F}_{\\mathcal{X}})\\) of vertices \\(\\mathcal{V}_{\\mathcal{X}}\\) and consistently oriented" + }, + { + "type": "image", + "bbox": [ + 0.553, + 0.096, + 0.868, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.191, + 0.905, + 0.22 + ], + "angle": 0, + "content": "Figure 2. We represent a 3D shape (left) as a (directed) shape graph \\(\\mathcal{G}_{\\mathcal{X}}\\) (middle). We call a cycle \\(\\mathcal{C}\\) in \\(\\mathcal{G}_{\\mathcal{X}}\\) a surface cycle (right)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.246, + 0.905, + 0.291 + ], + "angle": 0, + "content": "(e.g. clock-wise) triangles \\(\\mathcal{F}_{\\mathcal{X}} \\subset \\mathcal{V}_{\\mathcal{X}} \\times \\mathcal{V}_{\\mathcal{X}} \\times \\mathcal{V}_{\\mathcal{X}}\\), such that \\(\\mathcal{X}\\) forms an orientable continuous \\(2D\\) manifold (possibly with boundary) embedded in \\(3D\\) space." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.302, + 0.905, + 0.332 + ], + "angle": 0, + "content": "For our formalism it is advantageous to interpret a 3D shape as a graph, which we denote the shape graph:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.342, + 0.905, + 0.403 + ], + "angle": 0, + "content": "Definition 2 (Shape graph). The shape graph (of 3D shape \\(\\mathcal{X}\\)) is a tuple \\(\\mathcal{G}_{\\mathcal{X}} = (\\mathcal{V}_{\\mathcal{X}}, \\mathcal{E}_{\\mathcal{X}})\\) of vertices \\(\\mathcal{V}_{\\mathcal{X}}\\) and directed edges \\(\\mathcal{E}_{\\mathcal{X}} \\subset \\mathcal{V}_{\\mathcal{X}} \\times \\mathcal{V}_{\\mathcal{X}}\\), such that each triangle in \\(\\mathcal{F}_{\\mathcal{X}}\\) is represented by three unique and consistently directed edges." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.413, + 0.905, + 0.443 + ], + "angle": 0, + "content": "In addition to the shape graph (representing the whole 3D shape \\(\\mathcal{X}\\)), we consider certain subgraphs of \\(\\mathcal{G}_{\\mathcal{X}}\\):" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.453, + 0.905, + 0.528 + ], + "angle": 0, + "content": "Definition 3 (Surface cycle). A surface cycle (on 3D shape \\(\\mathcal{X}\\)) is a tuple \\(\\mathcal{C} = (\\mathcal{V}_{\\mathcal{C}}, \\mathcal{E}_{\\mathcal{C}})\\) with (non-empty) vertices \\(\\mathcal{V}_{\\mathcal{C}} \\subset \\mathcal{V}_{\\mathcal{X}}\\) and (non-empty) directed edges \\(\\mathcal{E}_{\\mathcal{C}} \\subset \\mathcal{V}_{\\mathcal{C}} \\times \\mathcal{V}_{\\mathcal{C}}\\), such that (i) \\(\\mathcal{E}_{\\mathcal{C}} \\subset \\mathcal{E}_{\\mathcal{X}}\\), and (ii) each vertex \\(v \\in \\mathcal{V}_{\\mathcal{C}}\\) has exactly one incoming and one outgoing edge in \\(\\mathcal{E}_{\\mathcal{C}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.548, + 0.9, + 0.565 + ], + "angle": 0, + "content": "3.2. Geometrically Consistent 3D Shape Matching" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.57, + 0.905, + 0.601 + ], + "angle": 0, + "content": "We define geometrically consistent 3D shape matching in terms of the respective shape graphs:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.61, + 0.906, + 0.747 + ], + "angle": 0, + "content": "Definition 4 (Geometrically Consistent Matching). The mapping \\(\\phi : \\mathcal{V}_{\\mathcal{X}} \\to \\mathcal{V}_{\\mathcal{Y}}\\) is a geometrically consistent matching from the source shape \\(\\mathcal{X}\\) to the target shape \\(\\mathcal{Y}\\) if \\(\\phi\\) preserves neighbourhoods as follows: whenever two vertices \\(x, \\bar{x} \\in \\mathcal{V}_{\\mathcal{X}}\\) are connected (i.e. \\((x, \\bar{x}) \\in \\mathcal{E}_{\\mathcal{X}}\\) or \\((\\bar{x}, x) \\in \\mathcal{E}_{\\mathcal{X}}\\)), then their corresponding vertices \\(\\phi(x), \\phi(\\bar{x}) \\in \\mathcal{V}_{\\mathcal{Y}}\\) must either be (i) connected (i.e. \\((\\phi(x), \\phi(\\bar{x})) \\in \\mathcal{E}_{\\mathcal{Y}}\\) or \\((\\phi(\\bar{x}), \\phi(x)) \\in \\mathcal{E}_{\\mathcal{Y}}\\)), or (ii) the same vertex (i.e. \\(\\phi(x) = \\phi(\\bar{x})\\))." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.756, + 0.906, + 0.817 + ], + "angle": 0, + "content": "The intuition is that neighbouring elements of \\(\\mathcal{X}\\) must be matched to neighbouring elements of \\(\\mathcal{Y}\\). We note that we require both shapes to have equal genus such that our definition of geometric consistency makes sense." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.831, + 0.83, + 0.848 + ], + "angle": 0, + "content": "4. Our 3D Shape Matching Approach" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "In this section we develop our geometrically consistent 3D shape matching formulation which ensures global geometric consistency. Our approach is based on representing the" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "914" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.147, + 0.09, + 0.434, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.181, + 0.486, + 0.239 + ], + "angle": 0, + "content": "Figure 3. Surface cycle collection (middle) for representing the 3D shape of an icosahedron (left). Individual surface cycles are glued together via shared edges (neighbouring surface cycles share opposite edges, cf. pairs of thickened edges on the right: \\(\\parallel\\) and \\(\\parallel\\))." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.264, + 0.485, + 0.401 + ], + "angle": 0, + "content": "source 3D shape using a collection of surface cycles, which we then match to the target shape (Sec. 4.1). To this end, we formulate an individual subproblem for each surface cycle (Sec. 4.2.1). Further, we couple individual subproblems (Secs. 4.2.2 and 4.2.3) so that our matchings are geometrically consistent according to Definition 4. Finally, we show that our overall formalism can be interpreted as a minimum-cost circulation flow problem on a hyper product graph (Sec. 4.5)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.41, + 0.483, + 0.427 + ], + "angle": 0, + "content": "4.1. Surface Cycle-based 3D Shape Representation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.432, + 0.484, + 0.478 + ], + "angle": 0, + "content": "We aim to represent the surface of 3D shape \\(\\mathcal{X}\\) with a collection of surface cycles which are glued together at opposite edges:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.489, + 0.484, + 0.521 + ], + "angle": 0, + "content": "Definition 5 (Opposite edge). The opposite edge to an edge \\((v, w)\\) is defined as \\(-(v, w) := (w, v)\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.53, + 0.484, + 0.591 + ], + "angle": 0, + "content": "We note that for every non-boundary edge \\( e \\in \\mathcal{E}_{\\mathcal{X}} \\) of shape \\( \\mathcal{X} \\), the opposite edge \\( -e \\in \\mathcal{E}_{\\mathcal{X}} \\) is also part of the shape graph \\( \\mathcal{G}_{\\mathcal{X}} \\) (cf. Fig. 3 right) because the pair of opposite edges \\( e \\) and \\( -e \\) belong to two neighbouring triangles." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.601, + 0.485, + 0.661 + ], + "angle": 0, + "content": "Definition 6 (Shape as collection of surface cycles). We represent a 3D shape \\(\\mathcal{X}\\) with a collection of \\(n \\in \\mathbb{N}^+\\) surface cycles \\(\\mathcal{C}_1, \\ldots, \\mathcal{C}_n\\) (on shape \\(\\mathcal{X}\\)) that partition the surface of \\(\\mathcal{X}\\) into \\(n\\) polygonal patches, such that:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.661, + 0.484, + 0.709 + ], + "angle": 0, + "content": "(i) \\(\\mathcal{E}_{\\mathcal{C}_i}\\cap \\mathcal{E}_{\\mathcal{C}_j} = \\emptyset\\) for all \\(i,j\\in \\{1,\\ldots ,n\\} ,i\\neq j\\) , and (ii) for every non-boundary edge \\(e\\in \\mathcal{E}_{\\mathcal{C}_i}\\) of shape \\(\\mathcal{X}\\) there exists \\(j\\) such that \\(-e\\in \\mathcal{E}_{\\mathcal{C}_j}\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.718, + 0.484, + 0.81 + ], + "angle": 0, + "content": "Condition (i) ensures that each edge of \\(\\mathcal{X}\\) can be part of at most one surface cycle, and (ii) ensures that neighbouring surface cycles cover opposite edges, see Fig. 3. We note that from here on we assume that each surface cycle represents an individual triangle of shape \\(\\mathcal{X}\\). We discuss more general polygonal surface cycles in Sec. C.3 in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.819, + 0.439, + 0.835 + ], + "angle": 0, + "content": "4.2. Our Coupled Product Graph Formalism" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "On a high-level, we use the product graph formalism introduced in [48] to address the matching between individual surface cycles and shape \\(\\mathcal{V}\\), which results in geometric consistency along the path within each individual cycle." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.182 + ], + "angle": 0, + "content": "Further, we introduce coupling constraints that have the effect of glueing neighbouring product graphs together. Combined with additional injectivity constraints, our matching is globally geometrically consistent, i.e. it preserves neighbourhoods between surface cycles, see Fig. 4 for an overview." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.195, + 0.9, + 0.211 + ], + "angle": 0, + "content": "4.2.1. Individual Surface Cycle Matching Subproblems" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.216, + 0.906, + 0.277 + ], + "angle": 0, + "content": "Each individual surface cycle \\(\\mathcal{C}_i\\) of \\(\\mathcal{X}\\) can be matched to shape \\(\\mathcal{V}\\) by finding a cyclic path in their product graph \\(\\mathcal{P}_i = (\\mathcal{V}_{\\mathcal{P}_i},\\mathcal{E}_{\\mathcal{P}_i})\\). It has been introduced in [48] for the problem of matching a 2D shape to a 3D shape and reads:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.292, + 0.908, + 0.339 + ], + "angle": 0, + "content": "Definition 7 (Product graph \\(\\mathcal{P}_i\\)). The product graph \\(\\mathcal{P}_i = (\\mathcal{V}_{\\mathcal{P}_i}, \\mathcal{E}_{\\mathcal{P}_i})\\) between the \\(i\\)-th surface cycle \\(\\mathcal{C}_i\\) and the 3D shape \\(\\mathcal{Y}\\) is a directed graph defined as" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.353, + 0.906, + 0.412 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {V} _ {\\mathcal {P} _ {i}} = \\mathcal {V} _ {\\mathcal {C} _ {i}} \\times \\mathcal {V} _ {\\mathcal {Y}}, \\\\ \\mathcal {E} _ {\\mathcal {P} _ {i}} = \\left\\{\\left(v, \\bar {v}\\right) \\in \\mathcal {V} _ {\\mathcal {P} _ {i}} \\times \\mathcal {V} _ {\\mathcal {P} _ {i}} \\mid v = \\left( \\begin{array}{l} x \\\\ y \\end{array} \\right), \\bar {v} = \\left( \\begin{array}{l} \\bar {x} \\\\ \\bar {y} \\end{array} \\right) \\right. \\tag {1} \\\\ (x, \\bar {x}) \\in \\mathcal {E} _ {\\mathcal {C} _ {i}}, (y, \\bar {y}) \\in \\mathcal {E} _ {\\mathcal {Y}} ^ {+} \\}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.426, + 0.899, + 0.445 + ], + "angle": 0, + "content": "with the extended edge set \\(\\mathcal{E}_{\\mathcal{V}}^{+} := \\mathcal{E}_{\\mathcal{V}} \\cup \\{(y, y) \\mid y \\in \\mathcal{V}_{\\mathcal{V}}\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.458, + 0.906, + 0.548 + ], + "angle": 0, + "content": "The key idea is that each edge \\((v,\\bar{v})\\in \\mathcal{E}_{\\mathcal{P}_i}\\) in the product graph can be interpreted as a (potential) matching between an edge of \\(C_i\\) and an edge (or a vertex to account for stretching/compression) of shape \\(\\mathcal{V}\\). As shown in [48], matching surface cycle \\(C_i\\) to shape \\(\\mathcal{V}\\) amounts to solving a cyclic shortest path problem in the product graph \\(\\mathcal{P}_i\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.55, + 0.906, + 0.687 + ], + "angle": 0, + "content": "However, for our setting of 3D-to-3D shape matching, the neighbourhood between pairs of surface cycles cannot be ensured when considering vanilla shortest path algorithms, since they would solve the \\(n\\) individual surface cycle matching subproblems independently. To tackle this, we consider the linear programming (LP) formalism of cyclic shortest path problems and add additional constraints to couple the individual product graphs \\(\\mathcal{P}_1,\\ldots ,\\mathcal{P}_n\\), as explained next." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.699, + 0.715, + 0.715 + ], + "angle": 0, + "content": "4.2.2. Subproblem Coupling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We couple the individual surface cycle matching subproblems \\(\\mathcal{P}_1, \\ldots, \\mathcal{P}_n\\) by glueing them together at opposite edges. To this end, we introduce coupling constraints \\(\\mathcal{V}_{\\mathcal{L}}\\) which serve the purpose of ensuring that matchings of opposite edges are consistent, i.e. resulting matchings (or rather resulting shortest cyclic paths of neighbouring surface cycles) must go through opposite edges, see Fig. 4 (iii). Defacto, this enforces the glueing of neighbouring product graphs, and thus ensures geometric consistency of neighbouring surface cycles. In addition to the coupling, we want to ensure injectivity for each surface cycle edge, i.e. each such edge is matched exactly once." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.512, + 0.957 + ], + "angle": 0, + "content": "915" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.141, + 0.088, + 0.331, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.087, + 0.613, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.087, + 0.884, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.253, + 0.908, + 0.366 + ], + "angle": 0, + "content": "Figure 4. Conceptual summary of our formalism. Left: Illustration of product graph \\(\\mathcal{P}_i\\) between the surface cycle \\(\\mathcal{C}_i\\) and shape graph \\(\\mathcal{V}\\). Each black dot inside the big yellow circle represents a product vertex. Overall, the product graph can be thought of having one copy of \\(\\mathcal{V}\\) for each vertex of \\(\\mathcal{C}_i\\), which are appropriately connected via edges (only a subset of edges are drawn to reduce visual clutter). Middle: Two neighbouring product graphs \\(\\mathcal{P}_i\\) and \\(\\mathcal{P}_j\\) that arise from two neighbouring surface cycles \\(\\mathcal{C}_i\\) and \\(\\mathcal{C}_j\\). Right: Illustration of constrained neighbouring product graphs. We add additional coupling constraints so that opposite edges of the product graphs (i.e. edges with same colour, see coloured lines ——) are coupled via coupling constraints \\(q_{ij}^{e_1}, q_{ij}^{e_2}, q_{ij}^{e_3} \\in \\mathcal{V}_{\\mathcal{L}}\\) (see green dots \\(\\bigcirc\\)). Furthermore, we bundle edges of the product graphs so that each edge of the surface cycles is matched exactly once. To this end, we add injectivity constraints \\(s_{x\\bar{x}}\\), \\(s_{\\bar{x} x} \\in \\mathcal{V}_{\\mathcal{S}}\\) (see blue dots \\(\\bigcirc\\) within bundles of edges)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.391, + 0.38, + 0.407 + ], + "angle": 0, + "content": "4.2.3. Surface Cycle Matching Injectivity" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.41, + 0.484, + 0.547 + ], + "angle": 0, + "content": "To enforce that each surface cycle edge is matched exactly once, i.e. it is matched to exactly one edge (or vertex) of \\(\\mathcal{V}\\), we introduce injectivity constraints \\(\\mathcal{V}_{\\mathcal{S}}\\). For that, let us denote the set of all edges (in the product graph) that are potential matching candidates of the edge \\((x,\\bar{x})\\in \\mathcal{E}_{\\mathcal{C}_i}\\) as the edge bundle of \\((x,\\bar{x})\\). For each such edge bundle, we introduce one injectivity constraint (see Fig. 4 (iii)), which has the purpose to ensure that only a single edge of each edge bundle is part of the final matching." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.557, + 0.283, + 0.572 + ], + "angle": 0, + "content": "4.3. Constraint Matrices" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.579, + 0.484, + 0.706 + ], + "angle": 0, + "content": "As mentioned, we use the LP formalism of the independent cyclic shortest path problems and add additional constraints for coupling an matching injectivity. To this end, we consider the constraint matrix \\( H = \\left[ \\begin{array}{c} P \\\\ L \\\\ S \\end{array} \\right] \\) consisting of submatrices \\( P, L \\) and \\( S \\). Here, submatrix \\( P \\) represents the collection of incidence matrices of individual product graphs, submatrix \\( L \\) the couplings, and submatrix \\( S \\) the injectivity components." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.707, + 0.484, + 0.873 + ], + "angle": 0, + "content": "The matrix \\(P\\). The vertex edge incidence matrix \\(P \\in \\{-1,0,1\\}^{c \\times m}\\) (with \\(m = |\\mathcal{E}_{\\mathcal{H}}|\\) columns, i.e. one column for each product edge, and \\(c = \\sum_{i=1}^{n} |\\mathcal{V}_{\\mathcal{C}_i}| |\\mathcal{V}_{\\mathcal{Y}}|\\) many rows) represents all \\(n\\) product graphs and is defined as \\(P := \\mathrm{diag}(P_1, \\ldots, P_n)\\), i.e. it contains the vertex edge incidence matrices \\(P_i\\) of individual product graphs \\(\\mathcal{P}_i\\) as diagonal blocks. To illustrate the structure of the matrix representation of one \\(\\mathcal{P}_i\\), we now discuss the simplified case of non-degenerate matchings (i.e. we do not allow edge to vertex matchings, see Sec. A in the supplementary for more details). In this case, the incidence matrix \\(P_i\\) reads" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.884, + 0.483, + 0.903 + ], + "angle": 0, + "content": "\\[\nP _ {i} := C _ {i} ^ {+} \\otimes Y ^ {+} - C _ {i} ^ {-} \\otimes Y ^ {-}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.391, + 0.907, + 0.663 + ], + "angle": 0, + "content": "with \\(\\otimes\\) being the Kroenecker product and with \\(C_i \\in \\{-1,0,1\\}^{|\\mathcal{V}_{c_i}| \\times |\\mathcal{E}_{c_i}^{\\prime}|\\}\\) and \\(Y \\in \\{-1,0,1\\}^{|\\mathcal{V}_{\\mathcal{Y}}| \\times |\\mathcal{E}_{\\mathcal{Y}}^{\\prime}|\\}\\) being the incidence matrices of surface cycle \\(C_i\\) and shape graph \\(\\mathcal{G}_{\\mathcal{Y}}\\) respectively (we note that \\(C_i\\) has exactly two non-zero of opposite sign in each row and column since it is the incidence matrix of a cycle). We use the notation \\(+\\) and \\(-\\) to split the incidence matrix into incoming and outgoing incidence matrix, respectively. In other words, \\(C_i^+\\) and \\(C_i^-\\) as well as \\(Y^+\\) and \\(Y^-\\) are the non-negative and non-positive entries of the incidence matrices \\(C_i\\) and \\(Y\\), respectively. We note that the sign splitting of incidence matrices \\(C_i\\) and \\(Y\\) is necessary to account for proper edge directions in the resulting product graph. By definition, their edge directions induce edge directions of product edges of respective product graph \\(\\mathcal{P}_i\\). Each \\(P_i\\) has a block structure with \\(|\\mathcal{V}_{\\mathcal{Y}}| \\times |\\mathcal{E}_{\\mathcal{Y}}|\\) sized blocks and each block contains either only \\(+1\\) or \\(-1\\) entries. Further, each block contains exactly one non-zero per column, see Fig. A.1 in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.663, + 0.906, + 0.693 + ], + "angle": 0, + "content": "The matrix \\(L\\). The coupling of opposite edges of neighbouring product graphs is captured in the matrix" + }, + { + "type": "equation", + "bbox": [ + 0.652, + 0.704, + 0.905, + 0.725 + ], + "angle": 0, + "content": "\\[\nL := K \\otimes I _ {\\left| \\varepsilon_ {y} ^ {+} \\right|}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.734, + 0.907, + 0.886 + ], + "angle": 0, + "content": "Here \\( K \\in \\{-1,0,1\\}^{p \\times |\\mathcal{E}_{\\mathcal{X}}|} \\) (\\( p \\) is the number of undirected non-boundary edges of \\( \\mathcal{X} \\)) is an incidence matrix which represents the incidence of opposite edges across all \\( n \\) surface cycles. \\( K \\) has at most one non-zero element per column and exactly two non-zeros per row (with opposite sign) and thus \\( L \\) has identical block structure with non-negative and non-positive blocks respectively. We note that the definition of \\( P \\) requires columns of the non-positive blocks of \\( L \\) to be permuted (so that it maps opposite edges of neighbouring product graphs), see Sec. A in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "The matrix \\(S\\). The bundling of edges of all \\(n\\) product" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "916" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.135, + 0.089, + 0.445, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.2, + 0.485, + 0.287 + ], + "angle": 0, + "content": "Figure 5. Visualisation of edge cost \\( c((v, \\bar{v})) \\) computation for the product edge \\( (v, \\bar{v}) \\). The product edge defines a potential matching between edges \\( (x, \\bar{x}) \\in \\mathcal{E}_{\\mathcal{X}} \\) of shape \\( \\mathcal{X} \\) and \\( (y, \\bar{y}) \\in \\mathcal{E}_{\\mathcal{Y}}^{+} \\) of shape \\( \\mathcal{Y} \\). Thus, we define the matching cost \\( c((v, \\bar{v})) \\) as feature difference between features \\( \\| f_x - f_y \\| \\) and \\( \\| f_{\\bar{x}} - f_{\\bar{y}} \\| \\) at respective source vertices \\( x, y \\) and target vertices \\( \\bar{x}, \\bar{y} \\)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.311, + 0.364, + 0.327 + ], + "angle": 0, + "content": "graphs can be described using the matrix" + }, + { + "type": "equation", + "bbox": [ + 0.22, + 0.335, + 0.483, + 0.361 + ], + "angle": 0, + "content": "\\[\nS := \\boldsymbol {I} _ {\\left| \\mathcal {E} _ {x} \\right|} \\otimes 1 _ {\\left| \\mathcal {E} _ {y} ^ {+} \\right|} ^ {T}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.37, + 0.484, + 0.451 + ], + "angle": 0, + "content": "Here, \\(S\\) has block diagonal structure consisting of \\(|\\mathcal{E}_{\\mathcal{X}}|\\) many non-negative blocks (i.e. \\(1_{|\\mathcal{E}_{\\mathcal{Y}}^{+}|}^{T}\\) blocks). We note that \\(Sx = 1\\) enforces matching injectivity by enumerating product edges such that all \\(|\\mathcal{E}_{\\mathcal{Y}}^{+}|\\)-many matching candidates of a single surface cycle are grouped." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.459, + 0.391, + 0.475 + ], + "angle": 0, + "content": "4.4. Resulting Integer Linear Program" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.481, + 0.484, + 0.542 + ], + "angle": 0, + "content": "We use the previously described constraint matrix \\( H \\) to cast geometrically consistent shape matching as a linear program which essentially represents \\( n \\) many coupled cyclic shortest path problems." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.542, + 0.484, + 0.618 + ], + "angle": 0, + "content": "Matching costs. The matching cost \\( c(e) \\) of product edge \\( e = \\left(\\binom{x}{y},\\binom{\\bar{x}}{\\bar{y}}\\right) \\) measures how well the edge \\( (x,\\bar{x})\\in \\mathcal{E}_{\\mathcal{X}} \\) of shape \\( \\mathcal{X} \\) and the edge \\( (y,\\bar{y})\\in \\mathcal{E}_{\\mathcal{Y}}^{+} \\) of shape \\( \\mathcal{V} \\) fit together, e.g. by comparing geometric properties or feature descriptors, see also Fig. 5." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.619, + 0.484, + 0.695 + ], + "angle": 0, + "content": "Optimisation problem. We collect matching costs of all \\(m\\) edges of the \\(n\\) product graphs in the vector \\(c \\in \\mathbb{R}_+^m\\). With that, we can use an indicator representation \\(x\\) for each edge (\\(x_k = 1\\) means that \\(k\\)-th edge is part of the final matching) so that our matching formalism reads" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.704, + 0.483, + 0.731 + ], + "angle": 0, + "content": "\\[\n\\min _ {x \\in \\{0, 1 \\} ^ {m}} c ^ {T} x \\quad \\text {s . t .} \\quad H x = b. \\qquad \\mathrm {(G e C o 3 D)}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.74, + 0.484, + 0.848 + ], + "angle": 0, + "content": "Here, \\(H\\) is the previously introduced constraint matrix and \\(b \\coloneqq \\left[0_{|\\mathcal{V}_{\\mathcal{H}}| - |\\mathcal{V}_{S}|}; 1_{|\\mathcal{V}_{S}|}\\right]\\) is a column vector with all zeros except for ones in rows belonging to rows of submatrix \\(S\\) of \\(H\\). We solve (GeCo3D) by considering its LP-relaxation, i.e. we replace \\(x \\in \\{0,1\\}^m\\) with \\(x \\in [0,1]^m\\). We empirically observe in our experiments, that all solutions are integral and consequently globally optimal." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Lemma 8. Matchings between shapes \\(\\mathcal{X}\\) and \\(\\mathcal{Y}\\) obtained by solving (GeCo3D) are globally geometrically consistent according to Definition 4." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.199 + ], + "angle": 0, + "content": "Proof. We represent each triangle of the source shape \\(\\mathcal{X}\\) using a surface cycle. Furthermore, matchings obtained by solving (GeCo3D) preserve neighbourhood relations along the path of a matched surface cycle as well as between the surface cycles. Overall, this ensures that connected vertices on \\(\\mathcal{X}\\) are only matched to connected vertices (or the same vertex) on shape \\(\\mathcal{Y}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.214, + 0.836, + 0.231 + ], + "angle": 0, + "content": "4.5. Hyper Product Graph Interpretation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.237, + 0.907, + 0.358 + ], + "angle": 0, + "content": "Our optimisation problem (GeCo3D) can be interpreted as finding a minimum-cost flow circulation in a directed hyper product graph \\(\\mathcal{H}\\), see [6] for definitions of flows in directed hyper graphs. Here, \\(\\mathcal{H}\\) can be obtained by interpreting \\(H\\) as a vertex hyper edge incidence matrix, i.e. by interpreting rows of \\(H\\) as vertices and columns of \\(H\\) as directed hyper edges, see Fig. 6 for an illustration. For a formal definition of \\(\\mathcal{H}\\) we refer to Sec. B in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.359, + 0.909, + 0.496 + ], + "angle": 0, + "content": "In general, such flow circulation problems on hypergraphs are NP-hard [5]. However, there are certain subclasses that are known to be solvable in polynomial time [10, 29, 42, 80, 84]. Unfortunately, the specific structure of our hyper graph is not listed among the known polynomial-time solvable subclasses [10, 29, 42, 80, 84]. Nevertheless, we empirically observe that all instances considered in our experiments yield an integral (and thus globally optimal) solution when solved using LP-relaxations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.511, + 0.646, + 0.529 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.907, + 0.599 + ], + "angle": 0, + "content": "In this section, we experimentally evaluate our method's performance for 3D shape matching. Further, we show as proof of concept that our method is applicable to planar graph matching." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.909, + 0.721 + ], + "angle": 0, + "content": "Setup. We conduct runtime experiments on an Intel Core i9 12900K with 128 GB DDR5 RAM. For solving our linear programm (GeCo3D), we use off-the-shelf solver Gurobi [37] (version 10). We explicitly resolve coupling constraints during problem construction to obtain smaller constraint matrices, see Sec. C.2 in the supplementary. To better account for shrinking and stretching, we integrate a distortion bound, see Sec. C.1 in the supplementary. For" + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.741, + 0.871, + 0.832 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.842, + 0.909, + 0.898 + ], + "angle": 0, + "content": "Figure 6. Illustration of interpreting our constraint matrix \\( H \\) as the vertex hyper edge incidence matrix of a hyper product graph \\( \\mathcal{H} \\), see Sec. B in the supplementary. To obtain \\( \\mathcal{H} \\), rows of \\( H \\) are interpreted as vertices, and columns as directed hyper edges of \\( \\mathcal{H} \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.512, + 0.957 + ], + "angle": 0, + "content": "917" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "quantitative comparison, we decimate shapes to 1000 triangles using algorithms provided in [41]. We use the deep features [13] for all methods except for SmoothShells [28] and DiscrOpt [62] for which we use the original (axiomatic) features as reported in respective papers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.17, + 0.482, + 0.276 + ], + "angle": 0, + "content": "Metrics. We evaluate matching accuracy using geodesic errors, i.e. geodesic distance to ground-truth matchings. We follow the Princeton protocol [43] and normalise errors by the square-root of the shape area (see [43, Sec. 8.2]). Furthermore, we evaluate geometric consistency of matchings using Dirichlet energies, i.e. the deformation energy induced by the matching (see [67, Sec. 8])." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.279, + 0.485, + 0.779 + ], + "angle": 0, + "content": "Methods. We compare the matching quality of the methods ULRSSM [13], SmoothShells [28], DiscrOpt [62] and SpiderMatch [67]. ULRSSM [13] is an unsupervised deep shape matching method achieving best matchings results on numerous benchmarks. Because of its good performance and its extensive evaluation, we consider it to be representative for deep shape matching methods. SmoothShells [28] is a functional map based shape alignment method which iteratively adds geometric information to compute matchings. DiscrOpt [62] is a functional map based method which relates functional maps to point-wise maps. SpiderMatch [67] is the only other method that is scalable and at the same time considers geometric consistency. Furthermore, for runtime evaluations, we only compare to other geometrically consistent methods since these aim to solve a much harder to solve problem (yet most of these methods do not scale well to relevant shape resolutions, see also Fig. 1 right). In this sense, we consider Windheuser et al. [89], SM-Comb [69], DiscomMatch [66] and SpiderMatch (geo.). Windheuser et al. [89] propose a geometrically consistent 3D shape matching formalism which they solve using LP-relaxations and iterative variable fixations. The methods SM-Comb [69] and DiscomMatch [66] aim to solve this formalism approximately on CPU and GPU, respectively. We use SpiderMatch (geo.) to indicate the SpiderMatch [67] method when using a different curve to represent the source shape: we consider a curve that covers all edges of the source shape and thus, contains intersections at every vertex of the source shape. This is necessary to guarantee global geometric consistency since SpiderMatch [67] can only ensure geometric consistency (according to Definition 4) at intersection points of the curve. Thus, SpiderMatch (geo.) yields a stronger notion of geometric consistency compared to SpiderMatch [67]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Datasets. We evaluate shape matching on four different datasets: remeshed FAUST [9, 21, 63] (100 near-isometric deformed human shapes from which we sample 100 test set pairs), SMAL [92] (49 non-isometric deformed animal shapes of eight species from which we sample 100 test set pairs), DT4D-H [57] (9 different classes of humanoid/game character shapes in different poses taken from DeformingThings4D [53] from which we sample 100 intra class test" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.09, + 0.898, + 0.171 + ], + "angle": 0, + "content": "
MethodFAUSTSMALDT4D IntraDT4D InterBeCoS
ULRSSM [13]0.0310.0480.0330.0410.057
SmoothShells [28]0.3790.3760.3730.4200.370
DiscrOpt [62]0.1100.2680.0750.1700.270
SpiderMatch [67]0.0290.0440.0270.0410.057
Ours0.0270.0440.0240.0390.056
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.177, + 0.905, + 0.219 + ], + "angle": 0, + "content": "Table 3. Comparison of mean geodesic errors (↓) of various shape matching methods. We can see that our method consistently outperforms other methods on all five datasets." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.254, + 0.905, + 0.299 + ], + "angle": 0, + "content": "set pairs and 100 inter class test set pairs) and BeCoS [25] (2543 animal and humanoid shapes from various datasets of which we consider the 141 full-to-full test set pairs)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.32, + 0.704, + 0.336 + ], + "angle": 0, + "content": "5.1.3D Shape Matching" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.346, + 0.905, + 0.377 + ], + "angle": 0, + "content": "In the following, we evaluate our method's shape matching performance w.r.t. runtime and matching quality." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.38, + 0.907, + 0.697 + ], + "angle": 0, + "content": "Runtime. In Fig. 1 right, we show runtime comparison to formalisms which consider geometric consistency. Curves show median runtimes over five instances of FAUST dataset. Among all methods, ours and SpiderMatch [67] scale best. They can handle 3D shapes with approximately twice as many vertices as their fastest competitors DiscoMatch [66], which tackles the problem proposed by Windheuser et al. [89]. However, we note that when considering SpiderMatch (geo.) and consequently curves with more intersections (which is necessary for global geometric consistency) the scalability of SpiderMatch (geo.) degrades drastically. This stems from the branch and bound algorithm (having exponential worst-case runtime in general) which is used to solve SpiderMatch [67]. In contrast, we empirically observe (in all of our tested instances) that the linear programming relaxation of (GeCo3D) is always tight (i.e. yields an integral optimal solution) which explains the scalability of our approach. We emphasise that matchings computed from (GeCo3D) are provably geometrically consistent according to Definition 4, while the matchings of SpiderMatch [67] are not (cf. also Tab. 4)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.701, + 0.905, + 0.852 + ], + "angle": 0, + "content": "Full Shape Matching. In Tab. 3 and Tab. 4, we quantitatively compare the shape matching performance of various methods. Our method consistently produces best results across all datasets w.r.t. mean geodesic errors and smoothest results w.r.t Dirichlet energies. This showcases the beneficial effect of strong priors induced by geometric consistency. In addition, in Fig. 7, we show qualitative results, which further showcases high-quality matching results computed with our method. We provide more results as well as ablation studies in Sec. D.1 in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Partial-to-Full Shape Matching. As proof of concept we show qualitative results of matchings computed with our method in the partial-to-full setting, see Fig. 8." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "918" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.478, + 0.171 + ], + "angle": 0, + "content": "
MethodFAUSTSMALDT4D IntraDT4D InterBeCoS
ULRSSM [13]2.12.63.42.64.1
SmoothShells [28]1.92.81.62.35.7
DiscrOpt [62]8.913.67.510.71.1
SpiderMatch [67]1.71.81.72.10.81
Ours0.460.530.480.620.77
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.181, + 0.483, + 0.237 + ], + "angle": 0, + "content": "Table 4. Comparison of mean Dirichlet energies \\((\\downarrow)\\) of various shape matching methods to measure smoothness of the matching. Our method consistently yields best results, which shows the effect of strong priors induced by global geometric consistency." + }, + { + "type": "image", + "bbox": [ + 0.09, + 0.242, + 0.475, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.535, + 0.484, + 0.62 + ], + "angle": 0, + "content": "Figure 7. Qualitative shape matching results computed with ULRSSM [13], SpiderMatch [67] and ours. We visualise matchings using colour and triangulation transfer from source to target shape. At legs and arms of shapes we can see stronger geometric consistency of ours compared to SpiderMatch [67] (see red and green arrows and distorted triangles)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.642, + 0.427, + 0.659 + ], + "angle": 0, + "content": "5.2. Application to Planar Graph Matching" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.664, + 0.484, + 0.771 + ], + "angle": 0, + "content": "As a proof of concept we show that our formalism directly applies to planar graph matching. To this end, we use keypoints on images of the WILLOW [16] dataset and obtain graphs by computing Delaunay triangulations of keypoints. Furthermore, we represent one of the resulting graphs using surface cycles to match it to the other graph, see Fig. 9 and in Sec. D.2 in the supplementary." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.79, + 0.482, + 0.864 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.869, + 0.483, + 0.898 + ], + "angle": 0, + "content": "Figure 8. Qualitative results computed with ours for partial-to-full shape matching of test set shapes [26] from SHREC'16 [17]." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.088, + 0.905, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.181, + 0.907, + 0.21 + ], + "angle": 0, + "content": "Figure 9. Examples of planar graph matching results using our method on ducks and cars from WILLOW [16] dataset." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.233, + 0.75, + 0.248 + ], + "angle": 0, + "content": "6. Discussion & Limitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.26, + 0.907, + 0.457 + ], + "angle": 0, + "content": "Our algorithm enforces geometric consistency and thus introduces a strong prior to resolve ambiguities in 3D shape matching problems. Yet, in theory our formalism contains undesirable matchings in the solution space: On the one hand, multiple surface cycles could be matched to the same vertex on the target shape. While this ensures our optimisation problem is always feasible (since it contains the extreme case of matching all surface cycles to a single vertex), this is an undesirable solution. In the future, this could be resolved by generalising our approach towards a symmetric formalism. On the other hand, our formalism allows for inside-out flips, as each individual matching element (i.e. an edge) cannot disambiguate extrinsic orientations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.459, + 0.909, + 0.596 + ], + "angle": 0, + "content": "We have empirically shown that results can be computed efficiently using off-the-shelf LP-solvers [37] and that in all considered instances the linear the programming relaxations are tight, i.e. we find a globally optimal solution for all instances. Yet, we observe that our constraint matrix is not totally unimodular. We leave an in-depth analysis of our formalism, along with answering the question whether there exists a polynomial time algorithm for geometrically consistent 3D shape matching, for future works." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.616, + 0.634, + 0.632 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We have presented a novel formalism for non-rigid 3D shape matching which globally enforces geometric consistency. Our key idea is to consider a shape representation that allows to cast 3D shape matching as an integer linear program, which is efficiently solvable in practice. In addition, we illustrate that the resulting problem can be interpreted as a minimum-cost flow circulation problem in a hyper graph. Overall, we consider our work to be important for the 3D shape analysis community. Geometric consistency is crucial in most practical 3D shape matching settings, which is often times ignored (especially by learning-based methods). Our work allows to obtain such matchings efficiently and with that poses a step towards integrating geometric consistency into learning-based methods. Furthermore, we hope to inspire follow-up works on neighbourhood preserving matching formalisms within the broader field of visual computing and beyond." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "919" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.092, + 0.245, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.113, + 0.484, + 0.175 + ], + "angle": 0, + "content": "We thank Johan Thunberg, Tobias Weißberg and Zorah Lähner for insightful discussions about our formalism. This work is supported by the ERC starting grant no. 101160648 (Harmony)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.19, + 0.188, + 0.206 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.215, + 0.482, + 0.242 + ], + "angle": 0, + "content": "[1] MOSEK ApS. MOSEK Optimizer API for Python 11.0.9, 2025. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.243, + 0.483, + 0.282 + ], + "angle": 0, + "content": "[2] Souhaib Attaiki and Maks Ovsjanikov. Understanding and improving features learned in deep functional maps. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.283, + 0.482, + 0.323 + ], + "angle": 0, + "content": "[3] Mathieu Aubry, Ulrich Schlickewei, and Daniel Cremers. The wave kernel signature: A quantum mechanical approach to shape analysis. In ICCV workshops. IEEE, 2011. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.323, + 0.482, + 0.363 + ], + "angle": 0, + "content": "[4] Lennart Bastian, Yizheng Xie, Nassir Navab, and Zorah Lahner. Hybrid functional maps for crease-aware nonisometric shape matching. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.363, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[5] Isabel Beckenbach. A hypergraph network simplex algorithm. In GOR, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.39, + 0.482, + 0.416 + ], + "angle": 0, + "content": "[6] Isabel Leonie Beckenbach. Matchings and flows in hypergraphs. PhD thesis, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.416, + 0.482, + 0.47 + ], + "angle": 0, + "content": "[7] Marcel Seelbach Benkner, Zorah Lähner, Vladislav Golyanik, Christof Wunderlich, Christian Theobalt, and Michael Moeller. Q-match: Iterative shape matching via quantum annealing. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.471, + 0.482, + 0.51 + ], + "angle": 0, + "content": "[8] Florian Bernard, Zeeshan Khan Suri, and Christian Theobalt. Mina: Convex mixed-integer programming for non-rigid shape alignment. In CVPR, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.511, + 0.482, + 0.55 + ], + "angle": 0, + "content": "[9] Federica Bogo, Javier Romero, Matthew Loper, and Michael J Black. Faust: Dataset and evaluation for 3d mesh registration. In CVPR, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.55, + 0.483, + 0.619 + ], + "angle": 0, + "content": "[10] Kateřina Böhmová, Jérémie Chalopin, Matúš Mihalák, Guido Proietti, and Peter Widmayer. Sequence hypergraphs: paths, flows, and cuts. Adventures Between Lower Bounds and Higher Altitudes: Essays Dedicated to Juraj Hromković on the Occasion of His 60th Birthday, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.619, + 0.483, + 0.687 + ], + "angle": 0, + "content": "[11] Alexander M Bronstein, Michael M Bronstein, Ron Kimmel, Mona Mahmoudi, and Guillermo Sapiro. A gromov-hausdorff framework with diffusion geometry for topologically-robust non-rigid shape matching. International Journal of Computer Vision, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.687, + 0.483, + 0.724 + ], + "angle": 0, + "content": "[12] Oliver Burghard and Reinhard Klein. Efficient lifted relaxations of the quadratic assignment problem. In CVMV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.726, + 0.483, + 0.766 + ], + "angle": 0, + "content": "[13] Dongliang Cao, Paul Roetzer, and Florian Bernard. Unsupervised learning of robust spectral shape matching. TOG, 2023. 1, 2, 7, 8, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.767, + 0.483, + 0.807 + ], + "angle": 0, + "content": "[14] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.808, + 0.483, + 0.834 + ], + "angle": 0, + "content": "[15] Qifeng Chen and Vladlen Koltun. Robust nonrigid registration by convex optimization. In ICCV, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.835, + 0.483, + 0.86 + ], + "angle": 0, + "content": "[16] Minsu Cho, Karteek Alahari, and Jean Ponce. Learning graphs to match. In ICCV, 2013. 8, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[17] Luca Cosmo, Emanuele Rodola, Michael M Bronstein, Andrea Torsello, Daniel Cremers, and Y Sahillioglu. Shrec'16: Partial matching of deformable shapes. 3DOR, 2016. 8, 6, 9" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.215, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[18] James Coughlan, Alan Yuille, Camper English, and Dan Snow. Efficient deformable template detection and localization without user initialization. CVIU, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.134, + 0.905, + 0.172 + ], + "angle": 0, + "content": "[19] Bailin Deng, Yuxin Yao, Roberto M Dyke, and Juyong Zhang. A survey of non-rigid 3d registration. In CGF, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.173, + 0.905, + 0.214 + ], + "angle": 0, + "content": "[20] Nicolas Donati, Etienne Corman, and Maks Ovsjanikov. Deep orientation-aware functional maps: Tackling symmetry issues in shape matching. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.215, + 0.905, + 0.254 + ], + "angle": 0, + "content": "[21] Nicolas Donati, Abhishek Sharma, and Maks Ovsjanikov. Deep geometric functional maps: Robust feature learning for shape correspondence. In CVPR, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.254, + 0.905, + 0.307 + ], + "angle": 0, + "content": "[22] Niladri Shekhar Dutt, Sanjeev Muralikrishnan, and Niloy J. Mitra. Diffusion 3d features (diff3f): Decorating textured shapes with distilled semantic features. In CVPR, 2024. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.308, + 0.905, + 0.349 + ], + "angle": 0, + "content": "[23] Nadav Dym, Haggai Maron, and Yaron Lipman. DS++ - A Flexible, Scalable and Provably Tight Relaxation for Matching Problems. TOG, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.43 + ], + "angle": 0, + "content": "[24] Bernhard Egger, William A. P. Smith, Ayush Tewari, Stefanie Wuhrer, Michael Zollhoefer, Thabo Beeler, Florian Bernard, Timo Bolkart, Adam Kortylewski, Sami Romdhani, Christian Theobalt, Volker Blanz, and Thomas Vetter. 3d morphable face models—past, present, and future. TOG, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.431, + 0.905, + 0.512 + ], + "angle": 0, + "content": "[25] Viktoria Ehm, Nafie El Amrani, Yizheng Xie, Lennart Bastian, Maolin Gao, Weikang Wang, Lu Sang, Dongliang Cao, Tobias Weiberg, Zorah Lahner, Daniel Cremers, and Florian Bernard. Beyond Complete Shapes: A Benchmark for Quantitative Evaluation of 3D Shape Matching Algorithms. CGF, 2025. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.512, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[26] Viktoria Ehm, Maolin Gao, Paul Roetzer, Marvin Eisenberger, Daniel Cremers, and Florian Bernard. Partial-to-partial shape matching with geometric consistency. In CVPR, 2024. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.566, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[27] Viktoria Ehm, Paul Roetzer, Marvin Eisenberger, Maolin Gao, Florian Bernard, and Daniel Cremers. Geometrically consistent partial shape matching. In 3DV, 2024. 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.608, + 0.905, + 0.647 + ], + "angle": 0, + "content": "[28] Marvin Eisenberger, Zorah Lahner, and Daniel Cremers. Smooth shells: Multi-scale shape registration with functional maps. In CVPR, 2020. 2, 7, 8, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[29] James R Evans and John J Jarvis. Network topology and integral multicommodity flow problems. Networks, 1978. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.673, + 0.905, + 0.699 + ], + "angle": 0, + "content": "[30] Danielle Ezuz and Mirela Ben-Chen. Deblurring and denoising of maps between shapes. In CGF, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.7, + 0.905, + 0.74 + ], + "angle": 0, + "content": "[31] Danielle Ezuz, Behrend Heeren, Omri Azencot, Martin Rumpf, and Mirela Ben-Chen. Elastic correspondence between triangle meshes. In CGF, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.74, + 0.905, + 0.766 + ], + "angle": 0, + "content": "[32] Pedro F Felzenszwalb. Representation and detection of deformable shapes. TPAMI, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.766, + 0.905, + 0.807 + ], + "angle": 0, + "content": "[33] Giorgio Gallo, Giustino Longo, Stefano Pallottino, and Sang Nguyen. Directed hypergraphs and applications. Discrete applied mathematics, 1993. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.807, + 0.905, + 0.86 + ], + "angle": 0, + "content": "[34] Maolin Gao, Paul Roetzer, Marvin Eisenberger, Zorah Lahner, Michael Moeller, Daniel Cremers, and Florian Bernard. Sigma: Scale-invariant global sparse shape matching. In ICCV, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.888 + ], + "angle": 0, + "content": "[35] Andrea Gasparetto, Luca Cosmo, Emanuele Rodola, Michael Bronstein, and Andrea Torsello. Spatial maps: From" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.888 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "920" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.119 + ], + "angle": 0, + "content": "low rank spectral to sparse spatial functional representations. In 3DV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.121, + 0.483, + 0.16 + ], + "angle": 0, + "content": "[36] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. 3d-coded: 3d correspondences by deep deformation. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.16, + 0.482, + 0.186 + ], + "angle": 0, + "content": "[37] Gurobi Optimization, LLC. Gurobi Optimizer Reference Manual, 2023. 6, 8, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.186, + 0.482, + 0.226 + ], + "angle": 0, + "content": "[38] Oshri Halimi, Or Litany, Emanuele Rodola, Alex M Bronstein, and Ron Kimmel. Unsupervised learning of dense shape correspondence. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.226, + 0.483, + 0.293 + ], + "angle": 0, + "content": "[39] Stefan Haller, Lorenz Feineis, Lisa Hutschenreiter, Florian Bernard, Carsten Rother, Dagmar Kainmüller, Paul Swoboda, and Bogdan Savchynskyy. A comparative study of graph matching algorithms in computer vision. In ECCV, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.293, + 0.482, + 0.333 + ], + "angle": 0, + "content": "[40] Benjamin Holzschuh, Zorah Lähner, and Daniel Cremers. Simulated annealing for 3d shape correspondence. In 3DV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.362 + ], + "angle": 0, + "content": "[41] Alec Jacobson, Daniele Panozzo, et al. libigl: A simple C++ geometry processing library, 2018. https://libigl.github.io/. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.402 + ], + "angle": 0, + "content": "[42] Robert G Jeroslow, Kipp Martin, Ronald L Rardin, and Jinchang Wang. Gainfree leontief substitution flow problems. Mathematical Programming, 1992. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.402, + 0.482, + 0.428 + ], + "angle": 0, + "content": "[43] Vladimir G Kim, Yaron Lipman, and Thomas Funkhouser. Blended intrinsic maps. TOG, 2011. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.429, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[44] Whoi-Yul Kim and Avinash C. Kak. 3-d object recognition using bipartite matching embedded in discrete relaxation. IEEE Transactions on Pattern Analysis & Machine Intelligence, 1991. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.482, + 0.482, + 0.535 + ], + "angle": 0, + "content": "[45] Eugene Krissinel and Kim Henrick. Secondary-structure matching (ssm), a new tool for fast protein structure alignment in three dimensions. Biological Crystallography, 2004. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.536, + 0.482, + 0.563 + ], + "angle": 0, + "content": "[46] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 1955. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.563, + 0.482, + 0.603 + ], + "angle": 0, + "content": "[47] Yam Kushinsky, Haggai Maron, Nadav Dym, and Yaron Lipman. Sinkhorn Algorithm for Lifted Assignment Problems. SIAM Journal on Imaging Sciences, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.603, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[48] Zorah Lähner, Emanuele Rodola, Frank R Schmidt, Michael M Bronstein, and Daniel Cremers. Efficient globally optimal 2d-to-3d deformable shape matching. In CVPR, 2016. 2, 4, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.657, + 0.482, + 0.711 + ], + "angle": 0, + "content": "[49] Laura Leal-Taixe and Bodo Rosenhahn. Pedestrian interaction in tracking: the social force model and global optimization methods. In Modeling, Simulation and Visual Analysis of Crowds: A Multidisciplinary Perspective. 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.711, + 0.482, + 0.738 + ], + "angle": 0, + "content": "[50] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. In ECCV, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.738, + 0.482, + 0.777 + ], + "angle": 0, + "content": "[51] Lei Li, Nicolas Donati, and Maks Ovsjanikov. Learning multi-resolution functional maps with spectral attention for robust shape matching. NIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.778, + 0.482, + 0.817 + ], + "angle": 0, + "content": "[52] Qinsong Li, Shengjun Liu, Ling Hu, and Xinru Liu. Shape correspondence using anisotropic chebyshev spectral cnns. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[53] Yang Li, Hikari Takehara, Takafumi Taketomi, Bo Zheng, and Matthias Nießner. 4dcomplete: Non-rigid motion estimation beyond the observable surface. In ICCV, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.858, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[54] Or Litany, Tal Remez, Emanuele Rodola, Alex Bronstein, and Michael Bronstein. Deep functional maps: Structured prediction for dense shape correspondence. In ICCV, 2017." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.548, + 0.094, + 0.559, + 0.105 + ], + "angle": 0, + "content": "2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.106, + 0.905, + 0.146 + ], + "angle": 0, + "content": "[55] Jiayi Ma, Xingyu Jiang, Aoxiang Fan, Junjun Jiang, and Junchi Yan. Image matching from handcrafted to deep features: A survey. IJCV, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.147, + 0.905, + 0.186 + ], + "angle": 0, + "content": "[56] Filippo Maggioli, Daniele Baieri, Emanuele Rodola, and Simone Melzi. Rematching: Low-resolution representations for scalable shape correspondence. In ECCV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.186, + 0.905, + 0.226 + ], + "angle": 0, + "content": "[57] Robin Magnet, Jing Ren, Olga Sorkine-Hornung, and Maks Ovsjanikov. Smooth non-rigid shape matching via effective dirichlet energy optimization. In 3DV, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.226, + 0.905, + 0.267 + ], + "angle": 0, + "content": "[58] Haggai Maron, Nadav Dym, Itay Kezurer, Shahar Kovalsky, and Yaron Lipman. Point registration via efficient convex relaxation. TOG, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.267, + 0.905, + 0.307 + ], + "angle": 0, + "content": "[59] Maks Ovsjanikov, Mirela Ben-Chen, Justin Solomon, Adrian Butscher, and Leonidas Guibas. Functional maps: a flexible representation of maps between shapes. TOG, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[60] Maks Ovsjanikov, Quentin Merigot, Facundo Memoli, and Leonidas Guibas. One point isometric matching with the heat kernel. In CGF, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.347, + 0.905, + 0.387 + ], + "angle": 0, + "content": "[61] Thomas Probst, Danda Pani Paudel, Ajad Chhatkuli, and Luc Van Gool. Unsupervised learning of consensus maximization for 3d vision problems. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.387, + 0.905, + 0.428 + ], + "angle": 0, + "content": "[62] Jing Ren, Simone Melzi, Peter Wonka, and Maks Ovsjanikov. Discrete optimization for shape matching. In CGF, 2021. 2, 7, 8, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.429, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[63] Jing Ren, Adrien Poulenard, Peter Wonka, and Maks Ovsjanikov. Continuous and orientation-preserving correspondences via functional maps. TOG, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.468, + 0.905, + 0.508 + ], + "angle": 0, + "content": "[64] F Rendl, P Pardalos, and H Wolkowicz. The Quadratic Assignment Problem: A Survey and Recent Developments. In DIMACS workshop, 1994. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.508, + 0.905, + 0.55 + ], + "angle": 0, + "content": "[65] Emanuele Rodola, Alex M Bronstein, Andrea Albarelli, Filippo Bergamasco, and Andrea Torsello. A game-theoretic approach to deformable shape matching. In CVPR, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.55, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[66] Paul Roetzer, Ahmed Abbas, Dongliang Cao, Florian Bernard, and Paul Swoboda. Discomatch: Fast discrete optimisation for geometrically consistent 3d shape matching. In ECCV, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.602, + 0.905, + 0.642 + ], + "angle": 0, + "content": "[67] Paul Roetzer and Florian Bernard. Spidermatch: 3d shape matching with global optimality and geometric consistency. In CVPR, 2024. 2, 3, 7, 8, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.642, + 0.905, + 0.682 + ], + "angle": 0, + "content": "[68] Paul Roetzer, Zorah Lähner, and Florian Bernard. Conjugate product graphs for globally optimal 2d-3d shape matching. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.682, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[69] Paul Roetzer, Paul Swoboda, Daniel Cremers, and Florian Bernard. A scalable combinatorial solver for elastic geometrically consistent 3d shape matching. In CVPR, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.737, + 0.905, + 0.777 + ], + "angle": 0, + "content": "[70] David Sankoff and Joseph B Kruskal. Time warps, string edits, and macromolecules: the theory and practice of sequence comparison. 1983. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.778, + 0.905, + 0.817 + ], + "angle": 0, + "content": "[71] Frank R. Schmidt, Eno Töppe, and Daniel Cremers. Efficient planar graph cuts with applications in computer vision. In CVPR, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[72] Patrick Schmidt, Marcel Campen, Janis Born, and Leif Kobbelt. Inter-surface maps via constant-curvature metrics. TOG, 39(4):119-1, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.858, + 0.905, + 0.884 + ], + "angle": 0, + "content": "[73] Patrick Schmidt, Dörte Pieper, and Leif Kobbelt. Surface maps via adaptive triangulations. CGF, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.884, + 0.905, + 0.897 + ], + "angle": 0, + "content": "[74] Thomas Schoenemann and Daniel Cremers. A combinatorial" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.094, + 0.905, + 0.897 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "921" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "solution for model-based image segmentation and real-time tracking. TPAMI, 2009. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.12, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[75] John Schreiner, Arul Asirvatham, Emil Praun, and Hugues Hoppe. Inter-surface mapping. In SIGGRAPH. 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.147, + 0.482, + 0.173 + ], + "angle": 0, + "content": "[76] Alexander Schrijver et al. Combinatorial optimization: polyhedra and efficiency. Springer, 2003. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.173, + 0.482, + 0.213 + ], + "angle": 0, + "content": "[77] Avinash Sharma, Radu Horaud, Jan Cech, and Edmond Boyer. Topologically-robust 3d shape matching based on diffusion geometry and seed growing. In CVPR, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.213, + 0.482, + 0.252 + ], + "angle": 0, + "content": "[78] Nicholas Sharp, Souhaib Attaiki, Keenan Crane, and Maks Ovsjanikov. Diffusionnet: Discretization agnostic learning on surfaces. TOG, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.253, + 0.482, + 0.293 + ], + "angle": 0, + "content": "[79] Justin Solomon, Gabriel Peyré, Vladimir G Kim, and Suvrit Sra. Entropic metric alignment for correspondence problems. TOG, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.293, + 0.482, + 0.333 + ], + "angle": 0, + "content": "[80] Y Soun and Klaus Truemper. Single commodity representation of multicommodity networks. SIAM Journal on Algebraic Discrete Methods, 1980. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.334, + 0.482, + 0.359 + ], + "angle": 0, + "content": "[81] Kenshi Takayama. Compatible intrinsic triangulations. TOG, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.36, + 0.482, + 0.426 + ], + "angle": 0, + "content": "[82] Gary KL Tam, Zhi-Quan Cheng, Yu-Kun Lai, Frank C Langbein, Yonghuai Liu, David Marshall, Ralph R Martin, Xian-Fang Sun, and Paul L Rosin. Registration of 3d point clouds and meshes: A survey from rigid to nonrigid. TVCG, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.428, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[83] Giovanni Trappolini, Luca Cosmo, Luca Moschella, Ricardo Marin, Simone Melzi, and Emanuele Rodola. Shape registration in the time of transformers. NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.469, + 0.482, + 0.494 + ], + "angle": 0, + "content": "[84] Klaus Truemper. Unimodular matrices of flow problems with additional constraints. Networks, 1977. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.495, + 0.482, + 0.534 + ], + "angle": 0, + "content": "[85] Oliver Van Kaick, Hao Zhang, Ghassan Hamarneh, and Daniel Cohen-Or. A survey on shape correspondence. In CGF, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.535, + 0.482, + 0.603 + ], + "angle": 0, + "content": "[86] Matthias Van Onsem, Dieter De Paepe, S Vanden Hautte, Pieter Bonte, Veerle Ledoux, Annelies Lejon, Femke Ongenae, Dennis Dreesen, and Sofie Van Hoecke. Hierarchical pattern matching for anomaly detection in time series. Computer Communications, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.603, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[87] Matthias Vestner, Roee Litman, Emanuele Rodola, Alex Bronstein, and Daniel Cremers. Product manifold filter: Non-rigid shape correspondence via kernel density estimation in the product space. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.657, + 0.482, + 0.696 + ], + "angle": 0, + "content": "[88] Ruben Wiersma, Elmar Eisemann, and Klaus Hildebrandt. Cnns on surfaces using rotation-equivariant features. TOG, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.697, + 0.482, + 0.751 + ], + "angle": 0, + "content": "[89] Thomas Windheuser, Ulrich Schlickewei, Frank R Schmidt, and Daniel Cremers. Geometrically consistent elastic matching of 3d shapes: A linear programming solution. In ICCV, 2011. 1, 2, 7, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.752, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[90] Thomas Windheuser, Ulrich Schlickwei, Frank R Schimdt, and Daniel Cremers. Large-scale integer linear programming for orientation preserving 3d shape matching. In CGF, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.805, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[91] Junchi Yan, Xu-Cheng Yin, Weiyao Lin, Cheng Deng, Hongyuan Zha, and Xiaokang Yang. A short survey of recent advances in graph matching. In ICMR, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.886 + ], + "angle": 0, + "content": "[92] Silvia Zuffi, Angjoo Kanazawa, David W Jacobs, and Michael J Black. 3d menagerie: Modeling the 3d shape and pose of animals. In CVPR, 2017. 7" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.482, + 0.886 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "922" + } + ] +] \ No newline at end of file diff --git a/2025/Fast Globally Optimal and Geometrically Consistent 3D Shape Matching/cbbb1f39-340b-4752-9244-371644b15edf_origin.pdf b/2025/Fast Globally Optimal and Geometrically Consistent 3D Shape Matching/cbbb1f39-340b-4752-9244-371644b15edf_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fa581886e22c34649a64647c9da960fa4b08b773 --- /dev/null +++ b/2025/Fast Globally Optimal and Geometrically Consistent 3D Shape Matching/cbbb1f39-340b-4752-9244-371644b15edf_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9f65b7935d570bd517ed4f864facc7cf185a9aed6aab3cb8dafe0e3525a31950 +size 3043175 diff --git a/2025/Fast Globally Optimal and Geometrically Consistent 3D Shape Matching/full.md b/2025/Fast Globally Optimal and Geometrically Consistent 3D Shape Matching/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7dbf37ca412f43bd2f8e2b2fb7b897cca5863825 --- /dev/null +++ b/2025/Fast Globally Optimal and Geometrically Consistent 3D Shape Matching/full.md @@ -0,0 +1,377 @@ +# Fast Globally Optimal and Geometrically Consistent 3D Shape Matching + +Paul Roetzer1,2 +1 University of Bonn + +Florian Bernard $^{1,2}$ $^{2}$ Lamarr Institute + +![](images/0a98acca2030be1129c670887f981b56749839a31efa968a436fe3aad142706c.jpg) +Key Idea of Our Approach + +![](images/da7792113543c129534f8b3b3f8ae27c7ca047d050c5a4f5d44f4a5ca4dc2900.jpg) +Figure 1. Left: Illustration of our approach for geometrically consistent matching of 3D shape $\mathcal{X}$ to 3D shape $\mathcal{Y}$ . We represent shape $\mathcal{X}$ using $n$ surface cycles $\mathcal{C}_1, \ldots, \mathcal{C}_n$ and then consistently match these to shape $\mathcal{Y}$ while preserving neighbourhood relations between the cycles. Middle: The triangulation- and colour-transfer between shapes using computed matchings of Cao et al. [13] (not geometrically consistent) and ours (geometrically consistent) illustrates the importance of geometric consistency. Right: Our method is the first 3D shape matching approach that yields globally geometrically consistent matchings, that is globally optimal, and scalable in practice, opposed to all competing methods (OOM stands for 'out of memory' and the dashed line indicates a weak notion of geometric consistency, see explanation of methods in Sec. 5). + +![](images/7859227a1ef7048dba4adb49f6a26daca995307b19acecec2690ee0441d4cc5e.jpg) + +![](images/241bd2df8f0e62117ef665127831755f94d96f1844cb21e8d2c0ea8b38d81f07.jpg) + +![](images/7e9e9f9740cdb1b9eb4df97b173b15c91cd4fa72228294bb8c5f79454d001282.jpg) + +![](images/cf3ff8894f376bfda227b0ff94dfd96b1bcb13c2271ce3d1100c715e5322205b.jpg) + +![](images/759b602e79dc3967063d7b58bbd065350cf5f611a2f65b8f1de1ded492b331af.jpg) +Source Cao et al. Ours Triangulation Transfer via Computed Matchings +Runtime Comparison on FAUST Dataset + +# Abstract + +Geometric consistency, i.e. the preservation of neighbourhoods, is a natural and strong prior in 3D shape matching. Geometrically consistent matchings are crucial for many downstream applications, such as texture transfer or statistical shape modelling. Yet, in practice, geometric consistency is often overlooked, or only achieved under severely limiting assumptions (e.g. a good initialisation). In this work, we propose a novel formalism for computing globally optimal and geometrically consistent matchings between 3D shapes which is scalable in practice. Our key idea is to represent the surface of the source shape as a collection of cyclic graphs, which are then consistently matched to the target shape. Mathematically, we construct a hyper product graph (between source and target shape), and then cast 3D shape matching as a minimum-cost circulation flow problem in this hyper graph, which yields global geometrically consistent matchings between both shapes. We empirically show that our formalism is efficiently solvable and that it leads to high-quality results. Our code is publicly available.1 + +# 1. Introduction + +The availability of correspondences between visual data are a key prerequisite for many visual computing tasks, including 3D reconstruction [50], loss computation for deep learning [14], protein alignment [45], anomaly detection [86], object recognition [44], shape modelling [24] and others. In many practical cases, it is desirable that correspondences preserve neighbourhood relations. For example, in 3D shape matching this could mean that when bringing a pair of neighbouring points on a source shape into correspondence with a pair of points on a target shape, the pair of points must remain neighbours on the target. Such a neighbourhood preservation can serve as strong topological prior, for example to resolve ambiguities, achieve robustness, or ensure well-posedness of an otherwise ill-posed problem. However, despite its crucial importance, neighbourhood preservation is in practice often-times overlooked. This is because many of the well-known formalisms are not efficiently solvable (e.g. the quadratic assignment problem [64], or general integer programming [76], which have been used to tackle image keypoint matching [55], graph matching [39, 91], or 3D shape matching [8, 34, 58, 89]). + +In this work we specifically focus on the task of non-rigid geometrically consistent 3D shape matching, in which + +
MethodGlobally OptimalGeometr. ConsistentScalable
MINA [8](✓)✗✗
SIGMA [34](✓)✗✗
PMSDP [58]✗✗✓
Windheuser et al. [89]✗✓✗
SpiderMatch [67]✓(✓)(✓)
Ours✓✓✓
+ +Table 1. Comparison of axiomatic 3D shape matching methods. + +neighbourhood-preserving correspondences between two given 2D manifolds (embedded in 3D space) are sought for. So far, there does not exist any 3D shape matching approach that combines the following three desirable properties: (i) global optimality, (ii) neighbourhood preservation, and (iii) scalability, see Tab. 1. While the recent approach SpiderMatch [67] fulfils (i)-(iii) to some extent, its scalability relies heavily on the choice of the cycle that is used to represent one of the shapes (cf. Fig. 1 right: solid vs. dashed purple line), thereby trading off neighbourhood preservation with scalability. Furthermore, proper geometric consistency (Definition 4) can only be enforced at intersection points of the cycle and thus also the quality of geometric consistency depends on the choice of the cycle. Inspired by this idea of representing a shape using a cycle, we represent one shape using a collection of cycles and show that this leads to an (in practice) efficiently solvable formalism which ensures proper geometric consistency. We summarise our main contributions as follows: + +- For the first time, we present a globally geometrically consistent formalism for 3D shape matching that is efficiently solvable to global optimality in practice. +- To achieve this, we introduce a novel 3D shape representation in which we represent a 3D shape using a collection of surface cycles, so that 3D shape matching can be cast as finding a minimum-cost flow circulation in a hyper product graph. +- We experimentally show that our formalism leads to high quality and geometrically consistent matchings between two 3D shapes. +- In addition to 3D shape matching, we also show, in a proof-of-concept manner, that our formalism can generalise to specific instances of other matching problems, such as graph matching when the source graph is planar. + +# 2. Related Work + +In the following we discuss works that are most relevant to our approach. We start by discussing efficiently solvable matching problems, continue with 3D shape matching, and conclude with geometrically consistent 3D shape matching. + +Efficiently Solvable Matching Problems. There are various instances of correspondence problems that can be + +solved efficiently. Among them is the linear assignment problem (LAP) [46], which matches points without considering their neighbourhood relations. For matching time series data, sequences, or (open) contours, the popular dynamic time warping algorithm can efficiently compute solutions while preserving neighbourhood relations [70]. Closed contours can be matched analogously via graph cuts in product graphs [71]. Certain tracking problems can be formulated as efficiently solvable flow problems [49], and model-based image segmentation can also be solved efficiently by matching a 2D contour to an image [18, 32, 74]. The matching of a 2D contour to a 3D shape can be solved using variants of Dijkstra's algorithm for finding minimum cost cycles in product graphs [48, 68]. The mentioned approaches show that there are formalisms to efficiently solve a diverse range of matching problems. Yet, these approaches do not directly generalise to 3D shape matching. In this work, we propose a novel geometrically consistent 3D shape matching formalism that is (in practice) efficiently solvable to global optimality. + +3D Shape Matching is the task of finding correspondences between two non-rigidly deformed surfaces. For an in-depth overview on 3D shape matching we refer the reader to survey papers [19, 82, 85]. Many axiomatic shape matching approaches [11, 60, 65] build on the functional maps framework, efficiently solving shape matching in the spectral domain [28, 35, 56, 58, 59, 62]. In addition, functional maps have been adopted in several deep shape matching variants, either trained in a supervised [36, 52, 54, 83, 88] or in a unsupervised manner [2, 4, 13, 20, 22, 38, 51, 78]. There are also alternative approaches based on consensus maximisation [61], or using convex relaxations [15, 58] or mixed integer programming [8, 34]. Yet, many of the existing 3D shape matching methods neglect geometric consistency, as it leads to hard-to-solve formalisms. + +Geometrically Consistent 3D Shape Matching. There are several approaches that have recognised the importance of geometric consistency in 3D shape matching. Some of them incorporate neighbourhood information by using formalisms based on the quadratic assignment problem (QAP) - yet, the QAP is NP-hard [64], so that respective approaches consider relaxations [12, 23, 47] or heuristics [7, 40, 79]. Opposed to such discrete formulations, there are also approaches that tackle 3D shape matching by deforming a continuous shape parametrisation using local optimisation. However, due to the severe nonconvexity of resulting problems, they rely on a good initialisation, e.g. in the form of sparse sets of landmark correspondences [72, 73, 75, 77, 81], or in the form of dense correspondences [30, 31, 87]. Windheuser et al. [89, 90] have modelled geometrically consistent shape matching in the discrete domain by matching triangles to triangles using an (expensive to solve) integer linear program. Re + +
SymbolDescription
X=(Vx,FX)3D shape
Gx=(Vx,EX)3D shape graph of X
C=(VC,EC)surface cycle (cycle in Gx)
{C1,...,Cn}Representation of X with n surface cycles
nNumber of surface cycles
Y=(Vy, fy)3D shape Y
Gy=(Vy, Fy)3D shape graph of Y
Pi=(VPi, EPi)Product graph (of C_i and mesh Y)
{P1,...,Pn}Product graph collection (of {C1,...,Cn} and mesh Y)
H=(VH, EH)Hyper product graph (coupled product graphs {P,...,Pn})
mNumber of hyper edges |EH|
HVertex edge incidence matrices of H
+ +Table 2. Summary of the notation used in this paper. + +cently, approximative solvers [66, 69] and extensions for partial shapes [26, 27] have been proposed. Furthermore, 3D shape matching has recently been formulated as a shortest path problem with additional constraints [67]. In this work, the surface mesh of the source 3D shape is represented using a long self-intersecting curve that traces the shape surface, and is then matched to the other shape while preserving intersections of the curve. While their idea of considering alternative 3D shape representations is promising, in the presented framework proper geometric consistency is only enforced at intersection points of the curve (see Tab. 4). In addition, runtimes grow drastically with an increasing number of intersection points, which can be seen in Fig. 1 right when comparing SpiderMatch [67] (using a curve with few intersections) to SpiderMatch (geo.) (using a curve with many intersections). Inspired by the idea of an alternative path-based 3D shape representation, we represent a 3D shape as a collection of cyclic graphs, and then tackle 3D shape matching by solving coupled matching problems of the individual cyclic paths. Based on the couplings, we are able to ensure global geometry consistency. Furthermore, we empirically observe that the resulting formalism is efficiently solvable to global optimality. + +# 3. Background + +We consider the task of finding a matching between a source shape $\mathcal{X}$ and a target shape $\mathcal{Y}$ (Sec. 3.1) such that the matching is geometrically consistent (Sec. 3.2). Our main notation is summarised in Tab. 2. + +# 3.1. Shapes and Graphs + +In the following, we define shapes and other relevant concepts, which are also illustrated in Fig. 2. We consider $3D$ shapes represented as triangular surface mesh: + +Definition 1 (3D shape). A 3D shape $\mathcal{X}$ is defined as a tuple $(\mathcal{V}_{\mathcal{X}},\mathcal{F}_{\mathcal{X}})$ of vertices $\mathcal{V}_{\mathcal{X}}$ and consistently oriented + +![](images/955846ec6a4e5bd46dfbf3f4b8d7e861f28a79d13f04de3bb40351afdd4e5f47.jpg) +Figure 2. We represent a 3D shape (left) as a (directed) shape graph $\mathcal{G}_{\mathcal{X}}$ (middle). We call a cycle $\mathcal{C}$ in $\mathcal{G}_{\mathcal{X}}$ a surface cycle (right). + +(e.g. clock-wise) triangles $\mathcal{F}_{\mathcal{X}} \subset \mathcal{V}_{\mathcal{X}} \times \mathcal{V}_{\mathcal{X}} \times \mathcal{V}_{\mathcal{X}}$ , such that $\mathcal{X}$ forms an orientable continuous $2D$ manifold (possibly with boundary) embedded in $3D$ space. + +For our formalism it is advantageous to interpret a 3D shape as a graph, which we denote the shape graph: + +Definition 2 (Shape graph). The shape graph (of 3D shape $\mathcal{X}$ ) is a tuple $\mathcal{G}_{\mathcal{X}} = (\mathcal{V}_{\mathcal{X}}, \mathcal{E}_{\mathcal{X}})$ of vertices $\mathcal{V}_{\mathcal{X}}$ and directed edges $\mathcal{E}_{\mathcal{X}} \subset \mathcal{V}_{\mathcal{X}} \times \mathcal{V}_{\mathcal{X}}$ , such that each triangle in $\mathcal{F}_{\mathcal{X}}$ is represented by three unique and consistently directed edges. + +In addition to the shape graph (representing the whole 3D shape $\mathcal{X}$ ), we consider certain subgraphs of $\mathcal{G}_{\mathcal{X}}$ : + +Definition 3 (Surface cycle). A surface cycle (on 3D shape $\mathcal{X}$ ) is a tuple $\mathcal{C} = (\mathcal{V}_{\mathcal{C}}, \mathcal{E}_{\mathcal{C}})$ with (non-empty) vertices $\mathcal{V}_{\mathcal{C}} \subset \mathcal{V}_{\mathcal{X}}$ and (non-empty) directed edges $\mathcal{E}_{\mathcal{C}} \subset \mathcal{V}_{\mathcal{C}} \times \mathcal{V}_{\mathcal{C}}$ , such that (i) $\mathcal{E}_{\mathcal{C}} \subset \mathcal{E}_{\mathcal{X}}$ , and (ii) each vertex $v \in \mathcal{V}_{\mathcal{C}}$ has exactly one incoming and one outgoing edge in $\mathcal{E}_{\mathcal{C}}$ . + +# 3.2. Geometrically Consistent 3D Shape Matching + +We define geometrically consistent 3D shape matching in terms of the respective shape graphs: + +Definition 4 (Geometrically Consistent Matching). The mapping $\phi : \mathcal{V}_{\mathcal{X}} \to \mathcal{V}_{\mathcal{Y}}$ is a geometrically consistent matching from the source shape $\mathcal{X}$ to the target shape $\mathcal{Y}$ if $\phi$ preserves neighbourhoods as follows: whenever two vertices $x, \bar{x} \in \mathcal{V}_{\mathcal{X}}$ are connected (i.e. $(x, \bar{x}) \in \mathcal{E}_{\mathcal{X}}$ or $(\bar{x}, x) \in \mathcal{E}_{\mathcal{X}}$ ), then their corresponding vertices $\phi(x), \phi(\bar{x}) \in \mathcal{V}_{\mathcal{Y}}$ must either be (i) connected (i.e. $(\phi(x), \phi(\bar{x})) \in \mathcal{E}_{\mathcal{Y}}$ or $(\phi(\bar{x}), \phi(x)) \in \mathcal{E}_{\mathcal{Y}}$ ), or (ii) the same vertex (i.e. $\phi(x) = \phi(\bar{x})$ ). + +The intuition is that neighbouring elements of $\mathcal{X}$ must be matched to neighbouring elements of $\mathcal{Y}$ . We note that we require both shapes to have equal genus such that our definition of geometric consistency makes sense. + +# 4. Our 3D Shape Matching Approach + +In this section we develop our geometrically consistent 3D shape matching formulation which ensures global geometric consistency. Our approach is based on representing the + +![](images/08f95eed8fca9222ec95803c635bf89d5cd3957a8923d6682e657d4f0f26eb72.jpg) +Figure 3. Surface cycle collection (middle) for representing the 3D shape of an icosahedron (left). Individual surface cycles are glued together via shared edges (neighbouring surface cycles share opposite edges, cf. pairs of thickened edges on the right: $\parallel$ and $\parallel$ ). + +source 3D shape using a collection of surface cycles, which we then match to the target shape (Sec. 4.1). To this end, we formulate an individual subproblem for each surface cycle (Sec. 4.2.1). Further, we couple individual subproblems (Secs. 4.2.2 and 4.2.3) so that our matchings are geometrically consistent according to Definition 4. Finally, we show that our overall formalism can be interpreted as a minimum-cost circulation flow problem on a hyper product graph (Sec. 4.5). + +# 4.1. Surface Cycle-based 3D Shape Representation + +We aim to represent the surface of 3D shape $\mathcal{X}$ with a collection of surface cycles which are glued together at opposite edges: + +Definition 5 (Opposite edge). The opposite edge to an edge $(v, w)$ is defined as $-(v, w) := (w, v)$ . + +We note that for every non-boundary edge $e \in \mathcal{E}_{\mathcal{X}}$ of shape $\mathcal{X}$ , the opposite edge $-e \in \mathcal{E}_{\mathcal{X}}$ is also part of the shape graph $\mathcal{G}_{\mathcal{X}}$ (cf. Fig. 3 right) because the pair of opposite edges $e$ and $-e$ belong to two neighbouring triangles. + +Definition 6 (Shape as collection of surface cycles). We represent a 3D shape $\mathcal{X}$ with a collection of $n \in \mathbb{N}^+$ surface cycles $\mathcal{C}_1, \ldots, \mathcal{C}_n$ (on shape $\mathcal{X}$ ) that partition the surface of $\mathcal{X}$ into $n$ polygonal patches, such that: + +(i) $\mathcal{E}_{\mathcal{C}_i}\cap \mathcal{E}_{\mathcal{C}_j} = \emptyset$ for all $i,j\in \{1,\ldots ,n\} ,i\neq j$ , and (ii) for every non-boundary edge $e\in \mathcal{E}_{\mathcal{C}_i}$ of shape $\mathcal{X}$ there exists $j$ such that $-e\in \mathcal{E}_{\mathcal{C}_j}$ + +Condition (i) ensures that each edge of $\mathcal{X}$ can be part of at most one surface cycle, and (ii) ensures that neighbouring surface cycles cover opposite edges, see Fig. 3. We note that from here on we assume that each surface cycle represents an individual triangle of shape $\mathcal{X}$ . We discuss more general polygonal surface cycles in Sec. C.3 in the supplementary. + +# 4.2. Our Coupled Product Graph Formalism + +On a high-level, we use the product graph formalism introduced in [48] to address the matching between individual surface cycles and shape $\mathcal{V}$ , which results in geometric consistency along the path within each individual cycle. + +Further, we introduce coupling constraints that have the effect of glueing neighbouring product graphs together. Combined with additional injectivity constraints, our matching is globally geometrically consistent, i.e. it preserves neighbourhoods between surface cycles, see Fig. 4 for an overview. + +# 4.2.1. Individual Surface Cycle Matching Subproblems + +Each individual surface cycle $\mathcal{C}_i$ of $\mathcal{X}$ can be matched to shape $\mathcal{V}$ by finding a cyclic path in their product graph $\mathcal{P}_i = (\mathcal{V}_{\mathcal{P}_i},\mathcal{E}_{\mathcal{P}_i})$ . It has been introduced in [48] for the problem of matching a 2D shape to a 3D shape and reads: + +Definition 7 (Product graph $\mathcal{P}_i$ ). The product graph $\mathcal{P}_i = (\mathcal{V}_{\mathcal{P}_i}, \mathcal{E}_{\mathcal{P}_i})$ between the $i$ -th surface cycle $\mathcal{C}_i$ and the 3D shape $\mathcal{Y}$ is a directed graph defined as + +$$ +\begin{array}{l} \mathcal {V} _ {\mathcal {P} _ {i}} = \mathcal {V} _ {\mathcal {C} _ {i}} \times \mathcal {V} _ {\mathcal {Y}}, \\ \mathcal {E} _ {\mathcal {P} _ {i}} = \left\{\left(v, \bar {v}\right) \in \mathcal {V} _ {\mathcal {P} _ {i}} \times \mathcal {V} _ {\mathcal {P} _ {i}} \mid v = \left( \begin{array}{l} x \\ y \end{array} \right), \bar {v} = \left( \begin{array}{l} \bar {x} \\ \bar {y} \end{array} \right) \right. \tag {1} \\ (x, \bar {x}) \in \mathcal {E} _ {\mathcal {C} _ {i}}, (y, \bar {y}) \in \mathcal {E} _ {\mathcal {Y}} ^ {+} \}, \\ \end{array} +$$ + +with the extended edge set $\mathcal{E}_{\mathcal{V}}^{+} := \mathcal{E}_{\mathcal{V}} \cup \{(y, y) \mid y \in \mathcal{V}_{\mathcal{V}}\}$ . + +The key idea is that each edge $(v,\bar{v})\in \mathcal{E}_{\mathcal{P}_i}$ in the product graph can be interpreted as a (potential) matching between an edge of $C_i$ and an edge (or a vertex to account for stretching/compression) of shape $\mathcal{V}$ . As shown in [48], matching surface cycle $C_i$ to shape $\mathcal{V}$ amounts to solving a cyclic shortest path problem in the product graph $\mathcal{P}_i$ . + +However, for our setting of 3D-to-3D shape matching, the neighbourhood between pairs of surface cycles cannot be ensured when considering vanilla shortest path algorithms, since they would solve the $n$ individual surface cycle matching subproblems independently. To tackle this, we consider the linear programming (LP) formalism of cyclic shortest path problems and add additional constraints to couple the individual product graphs $\mathcal{P}_1,\ldots ,\mathcal{P}_n$ , as explained next. + +# 4.2.2. Subproblem Coupling + +We couple the individual surface cycle matching subproblems $\mathcal{P}_1, \ldots, \mathcal{P}_n$ by glueing them together at opposite edges. To this end, we introduce coupling constraints $\mathcal{V}_{\mathcal{L}}$ which serve the purpose of ensuring that matchings of opposite edges are consistent, i.e. resulting matchings (or rather resulting shortest cyclic paths of neighbouring surface cycles) must go through opposite edges, see Fig. 4 (iii). Defacto, this enforces the glueing of neighbouring product graphs, and thus ensures geometric consistency of neighbouring surface cycles. In addition to the coupling, we want to ensure injectivity for each surface cycle edge, i.e. each such edge is matched exactly once. + +![](images/5abe4fc4747faf4a0e3de82d6ad423af0d73645707f5a72b3135d03496466f81.jpg) +Figure 4. Conceptual summary of our formalism. Left: Illustration of product graph $\mathcal{P}_i$ between the surface cycle $\mathcal{C}_i$ and shape graph $\mathcal{V}$ . Each black dot inside the big yellow circle represents a product vertex. Overall, the product graph can be thought of having one copy of $\mathcal{V}$ for each vertex of $\mathcal{C}_i$ , which are appropriately connected via edges (only a subset of edges are drawn to reduce visual clutter). Middle: Two neighbouring product graphs $\mathcal{P}_i$ and $\mathcal{P}_j$ that arise from two neighbouring surface cycles $\mathcal{C}_i$ and $\mathcal{C}_j$ . Right: Illustration of constrained neighbouring product graphs. We add additional coupling constraints so that opposite edges of the product graphs (i.e. edges with same colour, see coloured lines ——) are coupled via coupling constraints $q_{ij}^{e_1}, q_{ij}^{e_2}, q_{ij}^{e_3} \in \mathcal{V}_{\mathcal{L}}$ (see green dots $\bigcirc$ ). Furthermore, we bundle edges of the product graphs so that each edge of the surface cycles is matched exactly once. To this end, we add injectivity constraints $s_{x\bar{x}}$ , $s_{\bar{x} x} \in \mathcal{V}_{\mathcal{S}}$ (see blue dots $\bigcirc$ within bundles of edges). + +![](images/7a4898c7890f4c8b9df7ba737963678f48c438f1634aed4a3f583595cb63b303.jpg) + +![](images/9dd4d70409dd43ce884b230e381a50e9f4d825b3d49f2a698d96a2483e8b94ef.jpg) + +# 4.2.3. Surface Cycle Matching Injectivity + +To enforce that each surface cycle edge is matched exactly once, i.e. it is matched to exactly one edge (or vertex) of $\mathcal{V}$ , we introduce injectivity constraints $\mathcal{V}_{\mathcal{S}}$ . For that, let us denote the set of all edges (in the product graph) that are potential matching candidates of the edge $(x,\bar{x})\in \mathcal{E}_{\mathcal{C}_i}$ as the edge bundle of $(x,\bar{x})$ . For each such edge bundle, we introduce one injectivity constraint (see Fig. 4 (iii)), which has the purpose to ensure that only a single edge of each edge bundle is part of the final matching. + +# 4.3. Constraint Matrices + +As mentioned, we use the LP formalism of the independent cyclic shortest path problems and add additional constraints for coupling an matching injectivity. To this end, we consider the constraint matrix $H = \left[ \begin{array}{c} P \\ L \\ S \end{array} \right]$ consisting of submatrices $P, L$ and $S$ . Here, submatrix $P$ represents the collection of incidence matrices of individual product graphs, submatrix $L$ the couplings, and submatrix $S$ the injectivity components. + +The matrix $P$ . The vertex edge incidence matrix $P \in \{-1,0,1\}^{c \times m}$ (with $m = |\mathcal{E}_{\mathcal{H}}|$ columns, i.e. one column for each product edge, and $c = \sum_{i=1}^{n} |\mathcal{V}_{\mathcal{C}_i}| |\mathcal{V}_{\mathcal{Y}}|$ many rows) represents all $n$ product graphs and is defined as $P := \mathrm{diag}(P_1, \ldots, P_n)$ , i.e. it contains the vertex edge incidence matrices $P_i$ of individual product graphs $\mathcal{P}_i$ as diagonal blocks. To illustrate the structure of the matrix representation of one $\mathcal{P}_i$ , we now discuss the simplified case of non-degenerate matchings (i.e. we do not allow edge to vertex matchings, see Sec. A in the supplementary for more details). In this case, the incidence matrix $P_i$ reads + +$$ +P _ {i} := C _ {i} ^ {+} \otimes Y ^ {+} - C _ {i} ^ {-} \otimes Y ^ {-}, \tag {2} +$$ + +with $\otimes$ being the Kroenecker product and with $C_i \in \{-1,0,1\}^{|\mathcal{V}_{c_i}| \times |\mathcal{E}_{c_i}^{\prime}|\}$ and $Y \in \{-1,0,1\}^{|\mathcal{V}_{\mathcal{Y}}| \times |\mathcal{E}_{\mathcal{Y}}^{\prime}|\}$ being the incidence matrices of surface cycle $C_i$ and shape graph $\mathcal{G}_{\mathcal{Y}}$ respectively (we note that $C_i$ has exactly two non-zero of opposite sign in each row and column since it is the incidence matrix of a cycle). We use the notation $+$ and $-$ to split the incidence matrix into incoming and outgoing incidence matrix, respectively. In other words, $C_i^+$ and $C_i^-$ as well as $Y^+$ and $Y^-$ are the non-negative and non-positive entries of the incidence matrices $C_i$ and $Y$ , respectively. We note that the sign splitting of incidence matrices $C_i$ and $Y$ is necessary to account for proper edge directions in the resulting product graph. By definition, their edge directions induce edge directions of product edges of respective product graph $\mathcal{P}_i$ . Each $P_i$ has a block structure with $|\mathcal{V}_{\mathcal{Y}}| \times |\mathcal{E}_{\mathcal{Y}}|$ sized blocks and each block contains either only $+1$ or $-1$ entries. Further, each block contains exactly one non-zero per column, see Fig. A.1 in the supplementary. + +The matrix $L$ . The coupling of opposite edges of neighbouring product graphs is captured in the matrix + +$$ +L := K \otimes I _ {\left| \varepsilon_ {y} ^ {+} \right|}. \tag {3} +$$ + +Here $K \in \{-1,0,1\}^{p \times |\mathcal{E}_{\mathcal{X}}|}$ ( $p$ is the number of undirected non-boundary edges of $\mathcal{X}$ ) is an incidence matrix which represents the incidence of opposite edges across all $n$ surface cycles. $K$ has at most one non-zero element per column and exactly two non-zeros per row (with opposite sign) and thus $L$ has identical block structure with non-negative and non-positive blocks respectively. We note that the definition of $P$ requires columns of the non-positive blocks of $L$ to be permuted (so that it maps opposite edges of neighbouring product graphs), see Sec. A in the supplementary. + +The matrix $S$ . The bundling of edges of all $n$ product + +![](images/4acadaf178973b318df2a9ad57d35170319dfecec43d7f341715c10441167fc2.jpg) +Figure 5. Visualisation of edge cost $c((v, \bar{v}))$ computation for the product edge $(v, \bar{v})$ . The product edge defines a potential matching between edges $(x, \bar{x}) \in \mathcal{E}_{\mathcal{X}}$ of shape $\mathcal{X}$ and $(y, \bar{y}) \in \mathcal{E}_{\mathcal{Y}}^{+}$ of shape $\mathcal{Y}$ . Thus, we define the matching cost $c((v, \bar{v}))$ as feature difference between features $\| f_x - f_y \|$ and $\| f_{\bar{x}} - f_{\bar{y}} \|$ at respective source vertices $x, y$ and target vertices $\bar{x}, \bar{y}$ . + +graphs can be described using the matrix + +$$ +S := \boldsymbol {I} _ {\left| \mathcal {E} _ {x} \right|} \otimes 1 _ {\left| \mathcal {E} _ {y} ^ {+} \right|} ^ {T}. \tag {4} +$$ + +Here, $S$ has block diagonal structure consisting of $|\mathcal{E}_{\mathcal{X}}|$ many non-negative blocks (i.e. $1_{|\mathcal{E}_{\mathcal{Y}}^{+}|}^{T}$ blocks). We note that $Sx = 1$ enforces matching injectivity by enumerating product edges such that all $|\mathcal{E}_{\mathcal{Y}}^{+}|$ -many matching candidates of a single surface cycle are grouped. + +# 4.4. Resulting Integer Linear Program + +We use the previously described constraint matrix $H$ to cast geometrically consistent shape matching as a linear program which essentially represents $n$ many coupled cyclic shortest path problems. + +Matching costs. The matching cost $c(e)$ of product edge $e = \left(\binom{x}{y},\binom{\bar{x}}{\bar{y}}\right)$ measures how well the edge $(x,\bar{x})\in \mathcal{E}_{\mathcal{X}}$ of shape $\mathcal{X}$ and the edge $(y,\bar{y})\in \mathcal{E}_{\mathcal{Y}}^{+}$ of shape $\mathcal{V}$ fit together, e.g. by comparing geometric properties or feature descriptors, see also Fig. 5. + +Optimisation problem. We collect matching costs of all $m$ edges of the $n$ product graphs in the vector $c \in \mathbb{R}_+^m$ . With that, we can use an indicator representation $x$ for each edge ( $x_k = 1$ means that $k$ -th edge is part of the final matching) so that our matching formalism reads + +$$ +\min _ {x \in \{0, 1 \} ^ {m}} c ^ {T} x \quad \text {s . t .} \quad H x = b. \qquad \mathrm {(G e C o 3 D)} +$$ + +Here, $H$ is the previously introduced constraint matrix and $b \coloneqq \left[0_{|\mathcal{V}_{\mathcal{H}}| - |\mathcal{V}_{S}|}; 1_{|\mathcal{V}_{S}|}\right]$ is a column vector with all zeros except for ones in rows belonging to rows of submatrix $S$ of $H$ . We solve (GeCo3D) by considering its LP-relaxation, i.e. we replace $x \in \{0,1\}^m$ with $x \in [0,1]^m$ . We empirically observe in our experiments, that all solutions are integral and consequently globally optimal. + +Lemma 8. Matchings between shapes $\mathcal{X}$ and $\mathcal{Y}$ obtained by solving (GeCo3D) are globally geometrically consistent according to Definition 4. + +Proof. We represent each triangle of the source shape $\mathcal{X}$ using a surface cycle. Furthermore, matchings obtained by solving (GeCo3D) preserve neighbourhood relations along the path of a matched surface cycle as well as between the surface cycles. Overall, this ensures that connected vertices on $\mathcal{X}$ are only matched to connected vertices (or the same vertex) on shape $\mathcal{Y}$ . + +# 4.5. Hyper Product Graph Interpretation + +Our optimisation problem (GeCo3D) can be interpreted as finding a minimum-cost flow circulation in a directed hyper product graph $\mathcal{H}$ , see [6] for definitions of flows in directed hyper graphs. Here, $\mathcal{H}$ can be obtained by interpreting $H$ as a vertex hyper edge incidence matrix, i.e. by interpreting rows of $H$ as vertices and columns of $H$ as directed hyper edges, see Fig. 6 for an illustration. For a formal definition of $\mathcal{H}$ we refer to Sec. B in the supplementary. + +In general, such flow circulation problems on hypergraphs are NP-hard [5]. However, there are certain subclasses that are known to be solvable in polynomial time [10, 29, 42, 80, 84]. Unfortunately, the specific structure of our hyper graph is not listed among the known polynomial-time solvable subclasses [10, 29, 42, 80, 84]. Nevertheless, we empirically observe that all instances considered in our experiments yield an integral (and thus globally optimal) solution when solved using LP-relaxations. + +# 5. Experiments + +In this section, we experimentally evaluate our method's performance for 3D shape matching. Further, we show as proof of concept that our method is applicable to planar graph matching. + +Setup. We conduct runtime experiments on an Intel Core i9 12900K with 128 GB DDR5 RAM. For solving our linear programm (GeCo3D), we use off-the-shelf solver Gurobi [37] (version 10). We explicitly resolve coupling constraints during problem construction to obtain smaller constraint matrices, see Sec. C.2 in the supplementary. To better account for shrinking and stretching, we integrate a distortion bound, see Sec. C.1 in the supplementary. For + +![](images/451545cb9391a4aeb472c1cf66df41e21d7605e9f77eab77a8b3f891ca91a69d.jpg) +Figure 6. Illustration of interpreting our constraint matrix $H$ as the vertex hyper edge incidence matrix of a hyper product graph $\mathcal{H}$ , see Sec. B in the supplementary. To obtain $\mathcal{H}$ , rows of $H$ are interpreted as vertices, and columns as directed hyper edges of $\mathcal{H}$ . + +quantitative comparison, we decimate shapes to 1000 triangles using algorithms provided in [41]. We use the deep features [13] for all methods except for SmoothShells [28] and DiscrOpt [62] for which we use the original (axiomatic) features as reported in respective papers. + +Metrics. We evaluate matching accuracy using geodesic errors, i.e. geodesic distance to ground-truth matchings. We follow the Princeton protocol [43] and normalise errors by the square-root of the shape area (see [43, Sec. 8.2]). Furthermore, we evaluate geometric consistency of matchings using Dirichlet energies, i.e. the deformation energy induced by the matching (see [67, Sec. 8]). + +Methods. We compare the matching quality of the methods ULRSSM [13], SmoothShells [28], DiscrOpt [62] and SpiderMatch [67]. ULRSSM [13] is an unsupervised deep shape matching method achieving best matchings results on numerous benchmarks. Because of its good performance and its extensive evaluation, we consider it to be representative for deep shape matching methods. SmoothShells [28] is a functional map based shape alignment method which iteratively adds geometric information to compute matchings. DiscrOpt [62] is a functional map based method which relates functional maps to point-wise maps. SpiderMatch [67] is the only other method that is scalable and at the same time considers geometric consistency. Furthermore, for runtime evaluations, we only compare to other geometrically consistent methods since these aim to solve a much harder to solve problem (yet most of these methods do not scale well to relevant shape resolutions, see also Fig. 1 right). In this sense, we consider Windheuser et al. [89], SM-Comb [69], DiscomMatch [66] and SpiderMatch (geo.). Windheuser et al. [89] propose a geometrically consistent 3D shape matching formalism which they solve using LP-relaxations and iterative variable fixations. The methods SM-Comb [69] and DiscomMatch [66] aim to solve this formalism approximately on CPU and GPU, respectively. We use SpiderMatch (geo.) to indicate the SpiderMatch [67] method when using a different curve to represent the source shape: we consider a curve that covers all edges of the source shape and thus, contains intersections at every vertex of the source shape. This is necessary to guarantee global geometric consistency since SpiderMatch [67] can only ensure geometric consistency (according to Definition 4) at intersection points of the curve. Thus, SpiderMatch (geo.) yields a stronger notion of geometric consistency compared to SpiderMatch [67]. + +Datasets. We evaluate shape matching on four different datasets: remeshed FAUST [9, 21, 63] (100 near-isometric deformed human shapes from which we sample 100 test set pairs), SMAL [92] (49 non-isometric deformed animal shapes of eight species from which we sample 100 test set pairs), DT4D-H [57] (9 different classes of humanoid/game character shapes in different poses taken from DeformingThings4D [53] from which we sample 100 intra class test + +
MethodFAUSTSMALDT4D IntraDT4D InterBeCoS
ULRSSM [13]0.0310.0480.0330.0410.057
SmoothShells [28]0.3790.3760.3730.4200.370
DiscrOpt [62]0.1100.2680.0750.1700.270
SpiderMatch [67]0.0290.0440.0270.0410.057
Ours0.0270.0440.0240.0390.056
+ +Table 3. Comparison of mean geodesic errors (↓) of various shape matching methods. We can see that our method consistently outperforms other methods on all five datasets. + +set pairs and 100 inter class test set pairs) and BeCoS [25] (2543 animal and humanoid shapes from various datasets of which we consider the 141 full-to-full test set pairs). + +# 5.1.3D Shape Matching + +In the following, we evaluate our method's shape matching performance w.r.t. runtime and matching quality. + +Runtime. In Fig. 1 right, we show runtime comparison to formalisms which consider geometric consistency. Curves show median runtimes over five instances of FAUST dataset. Among all methods, ours and SpiderMatch [67] scale best. They can handle 3D shapes with approximately twice as many vertices as their fastest competitors DiscoMatch [66], which tackles the problem proposed by Windheuser et al. [89]. However, we note that when considering SpiderMatch (geo.) and consequently curves with more intersections (which is necessary for global geometric consistency) the scalability of SpiderMatch (geo.) degrades drastically. This stems from the branch and bound algorithm (having exponential worst-case runtime in general) which is used to solve SpiderMatch [67]. In contrast, we empirically observe (in all of our tested instances) that the linear programming relaxation of (GeCo3D) is always tight (i.e. yields an integral optimal solution) which explains the scalability of our approach. We emphasise that matchings computed from (GeCo3D) are provably geometrically consistent according to Definition 4, while the matchings of SpiderMatch [67] are not (cf. also Tab. 4). + +Full Shape Matching. In Tab. 3 and Tab. 4, we quantitatively compare the shape matching performance of various methods. Our method consistently produces best results across all datasets w.r.t. mean geodesic errors and smoothest results w.r.t Dirichlet energies. This showcases the beneficial effect of strong priors induced by geometric consistency. In addition, in Fig. 7, we show qualitative results, which further showcases high-quality matching results computed with our method. We provide more results as well as ablation studies in Sec. D.1 in the supplementary. + +Partial-to-Full Shape Matching. As proof of concept we show qualitative results of matchings computed with our method in the partial-to-full setting, see Fig. 8. + +
MethodFAUSTSMALDT4D IntraDT4D InterBeCoS
ULRSSM [13]2.12.63.42.64.1
SmoothShells [28]1.92.81.62.35.7
DiscrOpt [62]8.913.67.510.71.1
SpiderMatch [67]1.71.81.72.10.81
Ours0.460.530.480.620.77
+ +Table 4. Comparison of mean Dirichlet energies $(\downarrow)$ of various shape matching methods to measure smoothness of the matching. Our method consistently yields best results, which shows the effect of strong priors induced by global geometric consistency. + +![](images/a016ec165a49cb5a38e9690d8799ed00c5aec061fc538666c804d7e4ee4b0c79.jpg) +Figure 7. Qualitative shape matching results computed with ULRSSM [13], SpiderMatch [67] and ours. We visualise matchings using colour and triangulation transfer from source to target shape. At legs and arms of shapes we can see stronger geometric consistency of ours compared to SpiderMatch [67] (see red and green arrows and distorted triangles). + +# 5.2. Application to Planar Graph Matching + +As a proof of concept we show that our formalism directly applies to planar graph matching. To this end, we use keypoints on images of the WILLOW [16] dataset and obtain graphs by computing Delaunay triangulations of keypoints. Furthermore, we represent one of the resulting graphs using surface cycles to match it to the other graph, see Fig. 9 and in Sec. D.2 in the supplementary. + +![](images/d8a786580d2b7c194417389b964a2002c231d1c82fc7774f4fa0741053f02419.jpg) +Figure 8. Qualitative results computed with ours for partial-to-full shape matching of test set shapes [26] from SHREC'16 [17]. + +![](images/b7ac21ad303a65684758c20e553fbc6dc0c300a439d0d3eab932387da6fead76.jpg) +Figure 9. Examples of planar graph matching results using our method on ducks and cars from WILLOW [16] dataset. + +# 6. Discussion & Limitations + +Our algorithm enforces geometric consistency and thus introduces a strong prior to resolve ambiguities in 3D shape matching problems. Yet, in theory our formalism contains undesirable matchings in the solution space: On the one hand, multiple surface cycles could be matched to the same vertex on the target shape. While this ensures our optimisation problem is always feasible (since it contains the extreme case of matching all surface cycles to a single vertex), this is an undesirable solution. In the future, this could be resolved by generalising our approach towards a symmetric formalism. On the other hand, our formalism allows for inside-out flips, as each individual matching element (i.e. an edge) cannot disambiguate extrinsic orientations. + +We have empirically shown that results can be computed efficiently using off-the-shelf LP-solvers [37] and that in all considered instances the linear the programming relaxations are tight, i.e. we find a globally optimal solution for all instances. Yet, we observe that our constraint matrix is not totally unimodular. We leave an in-depth analysis of our formalism, along with answering the question whether there exists a polynomial time algorithm for geometrically consistent 3D shape matching, for future works. + +# 7. Conclusion + +We have presented a novel formalism for non-rigid 3D shape matching which globally enforces geometric consistency. Our key idea is to consider a shape representation that allows to cast 3D shape matching as an integer linear program, which is efficiently solvable in practice. In addition, we illustrate that the resulting problem can be interpreted as a minimum-cost flow circulation problem in a hyper graph. Overall, we consider our work to be important for the 3D shape analysis community. Geometric consistency is crucial in most practical 3D shape matching settings, which is often times ignored (especially by learning-based methods). Our work allows to obtain such matchings efficiently and with that poses a step towards integrating geometric consistency into learning-based methods. Furthermore, we hope to inspire follow-up works on neighbourhood preserving matching formalisms within the broader field of visual computing and beyond. + +# Acknowledgements + +We thank Johan Thunberg, Tobias Weißberg and Zorah Lähner for insightful discussions about our formalism. This work is supported by the ERC starting grant no. 101160648 (Harmony). + +# References + +[1] MOSEK ApS. MOSEK Optimizer API for Python 11.0.9, 2025. 5, 6 +[2] Souhaib Attaiki and Maks Ovsjanikov. Understanding and improving features learned in deep functional maps. In CVPR, 2023. 2 +[3] Mathieu Aubry, Ulrich Schlickewei, and Daniel Cremers. The wave kernel signature: A quantum mechanical approach to shape analysis. In ICCV workshops. IEEE, 2011. 4, 5 +[4] Lennart Bastian, Yizheng Xie, Nassir Navab, and Zorah Lahner. Hybrid functional maps for crease-aware nonisometric shape matching. In CVPR, 2024. 2 +[5] Isabel Beckenbach. A hypergraph network simplex algorithm. In GOR, 2018. 6 +[6] Isabel Leonie Beckenbach. Matchings and flows in hypergraphs. PhD thesis, 2019. 6 +[7] Marcel Seelbach Benkner, Zorah Lähner, Vladislav Golyanik, Christof Wunderlich, Christian Theobalt, and Michael Moeller. Q-match: Iterative shape matching via quantum annealing. In ICCV, 2021. 2 +[8] Florian Bernard, Zeeshan Khan Suri, and Christian Theobalt. Mina: Convex mixed-integer programming for non-rigid shape alignment. In CVPR, 2020. 1, 2 +[9] Federica Bogo, Javier Romero, Matthew Loper, and Michael J Black. Faust: Dataset and evaluation for 3d mesh registration. In CVPR, 2014. 7 +[10] Kateřina Böhmová, Jérémie Chalopin, Matúš Mihalák, Guido Proietti, and Peter Widmayer. Sequence hypergraphs: paths, flows, and cuts. Adventures Between Lower Bounds and Higher Altitudes: Essays Dedicated to Juraj Hromković on the Occasion of His 60th Birthday, 2018. 6 +[11] Alexander M Bronstein, Michael M Bronstein, Ron Kimmel, Mona Mahmoudi, and Guillermo Sapiro. A gromov-hausdorff framework with diffusion geometry for topologically-robust non-rigid shape matching. International Journal of Computer Vision, 2010. 2 +[12] Oliver Burghard and Reinhard Klein. Efficient lifted relaxations of the quadratic assignment problem. In CVMV, 2017. 2 +[13] Dongliang Cao, Paul Roetzer, and Florian Bernard. Unsupervised learning of robust spectral shape matching. TOG, 2023. 1, 2, 7, 8, 4, 5, 6 +[14] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 1 +[15] Qifeng Chen and Vladlen Koltun. Robust nonrigid registration by convex optimization. In ICCV, 2015. 2 +[16] Minsu Cho, Karteek Alahari, and Jean Ponce. Learning graphs to match. In ICCV, 2013. 8, 6 +[17] Luca Cosmo, Emanuele Rodola, Michael M Bronstein, Andrea Torsello, Daniel Cremers, and Y Sahillioglu. Shrec'16: Partial matching of deformable shapes. 3DOR, 2016. 8, 6, 9 + +[18] James Coughlan, Alan Yuille, Camper English, and Dan Snow. Efficient deformable template detection and localization without user initialization. CVIU, 2000. 2 +[19] Bailin Deng, Yuxin Yao, Roberto M Dyke, and Juyong Zhang. A survey of non-rigid 3d registration. In CGF, 2022. 2 +[20] Nicolas Donati, Etienne Corman, and Maks Ovsjanikov. Deep orientation-aware functional maps: Tackling symmetry issues in shape matching. In CVPR, 2022. 2 +[21] Nicolas Donati, Abhishek Sharma, and Maks Ovsjanikov. Deep geometric functional maps: Robust feature learning for shape correspondence. In CVPR, 2020. 7 +[22] Niladri Shekhar Dutt, Sanjeev Muralikrishnan, and Niloy J. Mitra. Diffusion 3d features (diff3f): Decorating textured shapes with distilled semantic features. In CVPR, 2024. 2, 4, 5 +[23] Nadav Dym, Haggai Maron, and Yaron Lipman. DS++ - A Flexible, Scalable and Provably Tight Relaxation for Matching Problems. TOG, 2017. 2 +[24] Bernhard Egger, William A. P. Smith, Ayush Tewari, Stefanie Wuhrer, Michael Zollhoefer, Thabo Beeler, Florian Bernard, Timo Bolkart, Adam Kortylewski, Sami Romdhani, Christian Theobalt, Volker Blanz, and Thomas Vetter. 3d morphable face models—past, present, and future. TOG, 2020. 1 +[25] Viktoria Ehm, Nafie El Amrani, Yizheng Xie, Lennart Bastian, Maolin Gao, Weikang Wang, Lu Sang, Dongliang Cao, Tobias Weiberg, Zorah Lahner, Daniel Cremers, and Florian Bernard. Beyond Complete Shapes: A Benchmark for Quantitative Evaluation of 3D Shape Matching Algorithms. CGF, 2025. 7 +[26] Viktoria Ehm, Maolin Gao, Paul Roetzer, Marvin Eisenberger, Daniel Cremers, and Florian Bernard. Partial-to-partial shape matching with geometric consistency. In CVPR, 2024. 3, 8 +[27] Viktoria Ehm, Paul Roetzer, Marvin Eisenberger, Maolin Gao, Florian Bernard, and Daniel Cremers. Geometrically consistent partial shape matching. In 3DV, 2024. 3, 9 +[28] Marvin Eisenberger, Zorah Lahner, and Daniel Cremers. Smooth shells: Multi-scale shape registration with functional maps. In CVPR, 2020. 2, 7, 8, 6 +[29] James R Evans and John J Jarvis. Network topology and integral multicommodity flow problems. Networks, 1978. 6 +[30] Danielle Ezuz and Mirela Ben-Chen. Deblurring and denoising of maps between shapes. In CGF, 2017. 2 +[31] Danielle Ezuz, Behrend Heeren, Omri Azencot, Martin Rumpf, and Mirela Ben-Chen. Elastic correspondence between triangle meshes. In CGF, 2019. 2 +[32] Pedro F Felzenszwalb. Representation and detection of deformable shapes. TPAMI, 2005. 2 +[33] Giorgio Gallo, Giustino Longo, Stefano Pallottino, and Sang Nguyen. Directed hypergraphs and applications. Discrete applied mathematics, 1993. 2 +[34] Maolin Gao, Paul Roetzer, Marvin Eisenberger, Zorah Lahner, Michael Moeller, Daniel Cremers, and Florian Bernard. Sigma: Scale-invariant global sparse shape matching. In ICCV, 2023. 1, 2 +[35] Andrea Gasparetto, Luca Cosmo, Emanuele Rodola, Michael Bronstein, and Andrea Torsello. Spatial maps: From + +low rank spectral to sparse spatial functional representations. In 3DV, 2017. 2 +[36] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. 3d-coded: 3d correspondences by deep deformation. In ECCV, 2018. 2 +[37] Gurobi Optimization, LLC. Gurobi Optimizer Reference Manual, 2023. 6, 8, 5 +[38] Oshri Halimi, Or Litany, Emanuele Rodola, Alex M Bronstein, and Ron Kimmel. Unsupervised learning of dense shape correspondence. In CVPR, 2019. 2 +[39] Stefan Haller, Lorenz Feineis, Lisa Hutschenreiter, Florian Bernard, Carsten Rother, Dagmar Kainmüller, Paul Swoboda, and Bogdan Savchynskyy. A comparative study of graph matching algorithms in computer vision. In ECCV, 2022. 1 +[40] Benjamin Holzschuh, Zorah Lähner, and Daniel Cremers. Simulated annealing for 3d shape correspondence. In 3DV, 2020. 2 +[41] Alec Jacobson, Daniele Panozzo, et al. libigl: A simple C++ geometry processing library, 2018. https://libigl.github.io/. 7 +[42] Robert G Jeroslow, Kipp Martin, Ronald L Rardin, and Jinchang Wang. Gainfree leontief substitution flow problems. Mathematical Programming, 1992. 6 +[43] Vladimir G Kim, Yaron Lipman, and Thomas Funkhouser. Blended intrinsic maps. TOG, 2011. 7 +[44] Whoi-Yul Kim and Avinash C. Kak. 3-d object recognition using bipartite matching embedded in discrete relaxation. IEEE Transactions on Pattern Analysis & Machine Intelligence, 1991. 1 +[45] Eugene Krissinel and Kim Henrick. Secondary-structure matching (ssm), a new tool for fast protein structure alignment in three dimensions. Biological Crystallography, 2004. 1 +[46] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 1955. 2 +[47] Yam Kushinsky, Haggai Maron, Nadav Dym, and Yaron Lipman. Sinkhorn Algorithm for Lifted Assignment Problems. SIAM Journal on Imaging Sciences, 2019. 2 +[48] Zorah Lähner, Emanuele Rodola, Frank R Schmidt, Michael M Bronstein, and Daniel Cremers. Efficient globally optimal 2d-to-3d deformable shape matching. In CVPR, 2016. 2, 4, 1, 3 +[49] Laura Leal-Taixe and Bodo Rosenhahn. Pedestrian interaction in tracking: the social force model and global optimization methods. In Modeling, Simulation and Visual Analysis of Crowds: A Multidisciplinary Perspective. 2013. 2 +[50] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. In ECCV, 2024. 1 +[51] Lei Li, Nicolas Donati, and Maks Ovsjanikov. Learning multi-resolution functional maps with spectral attention for robust shape matching. NIPS, 2022. 2 +[52] Qinsong Li, Shengjun Liu, Ling Hu, and Xinru Liu. Shape correspondence using anisotropic chebyshev spectral cnns. In CVPR, 2020. 2 +[53] Yang Li, Hikari Takehara, Takafumi Taketomi, Bo Zheng, and Matthias Nießner. 4dcomplete: Non-rigid motion estimation beyond the observable surface. In ICCV, 2021. 7 +[54] Or Litany, Tal Remez, Emanuele Rodola, Alex Bronstein, and Michael Bronstein. Deep functional maps: Structured prediction for dense shape correspondence. In ICCV, 2017. + +2 +[55] Jiayi Ma, Xingyu Jiang, Aoxiang Fan, Junjun Jiang, and Junchi Yan. Image matching from handcrafted to deep features: A survey. IJCV, 2021. 1 +[56] Filippo Maggioli, Daniele Baieri, Emanuele Rodola, and Simone Melzi. Rematching: Low-resolution representations for scalable shape correspondence. In ECCV, 2024. 2 +[57] Robin Magnet, Jing Ren, Olga Sorkine-Hornung, and Maks Ovsjanikov. Smooth non-rigid shape matching via effective dirichlet energy optimization. In 3DV, 2022. 7 +[58] Haggai Maron, Nadav Dym, Itay Kezurer, Shahar Kovalsky, and Yaron Lipman. Point registration via efficient convex relaxation. TOG, 2016. 1, 2 +[59] Maks Ovsjanikov, Mirela Ben-Chen, Justin Solomon, Adrian Butscher, and Leonidas Guibas. Functional maps: a flexible representation of maps between shapes. TOG, 2012. 2 +[60] Maks Ovsjanikov, Quentin Merigot, Facundo Memoli, and Leonidas Guibas. One point isometric matching with the heat kernel. In CGF, 2010. 2 +[61] Thomas Probst, Danda Pani Paudel, Ajad Chhatkuli, and Luc Van Gool. Unsupervised learning of consensus maximization for 3d vision problems. In CVPR, 2019. 2 +[62] Jing Ren, Simone Melzi, Peter Wonka, and Maks Ovsjanikov. Discrete optimization for shape matching. In CGF, 2021. 2, 7, 8, 6 +[63] Jing Ren, Adrien Poulenard, Peter Wonka, and Maks Ovsjanikov. Continuous and orientation-preserving correspondences via functional maps. TOG, 2018. 7 +[64] F Rendl, P Pardalos, and H Wolkowicz. The Quadratic Assignment Problem: A Survey and Recent Developments. In DIMACS workshop, 1994. 1, 2 +[65] Emanuele Rodola, Alex M Bronstein, Andrea Albarelli, Filippo Bergamasco, and Andrea Torsello. A game-theoretic approach to deformable shape matching. In CVPR, 2012. 2 +[66] Paul Roetzer, Ahmed Abbas, Dongliang Cao, Florian Bernard, and Paul Swoboda. Discomatch: Fast discrete optimisation for geometrically consistent 3d shape matching. In ECCV, 2024. 3, 7 +[67] Paul Roetzer and Florian Bernard. Spidermatch: 3d shape matching with global optimality and geometric consistency. In CVPR, 2024. 2, 3, 7, 8, 6 +[68] Paul Roetzer, Zorah Lähner, and Florian Bernard. Conjugate product graphs for globally optimal 2d-3d shape matching. In CVPR, 2023. 2 +[69] Paul Roetzer, Paul Swoboda, Daniel Cremers, and Florian Bernard. A scalable combinatorial solver for elastic geometrically consistent 3d shape matching. In CVPR, 2022. 3, 7 +[70] David Sankoff and Joseph B Kruskal. Time warps, string edits, and macromolecules: the theory and practice of sequence comparison. 1983. 2 +[71] Frank R. Schmidt, Eno Töppe, and Daniel Cremers. Efficient planar graph cuts with applications in computer vision. In CVPR, 2009. 2 +[72] Patrick Schmidt, Marcel Campen, Janis Born, and Leif Kobbelt. Inter-surface maps via constant-curvature metrics. TOG, 39(4):119-1, 2020. 2 +[73] Patrick Schmidt, Dörte Pieper, and Leif Kobbelt. Surface maps via adaptive triangulations. CGF, 2023. 2 +[74] Thomas Schoenemann and Daniel Cremers. A combinatorial + +solution for model-based image segmentation and real-time tracking. TPAMI, 2009. 2, 3 +[75] John Schreiner, Arul Asirvatham, Emil Praun, and Hugues Hoppe. Inter-surface mapping. In SIGGRAPH. 2004. 2 +[76] Alexander Schrijver et al. Combinatorial optimization: polyhedra and efficiency. Springer, 2003. 1 +[77] Avinash Sharma, Radu Horaud, Jan Cech, and Edmond Boyer. Topologically-robust 3d shape matching based on diffusion geometry and seed growing. In CVPR, 2011. 2 +[78] Nicholas Sharp, Souhaib Attaiki, Keenan Crane, and Maks Ovsjanikov. Diffusionnet: Discretization agnostic learning on surfaces. TOG, 2022. 2 +[79] Justin Solomon, Gabriel Peyré, Vladimir G Kim, and Suvrit Sra. Entropic metric alignment for correspondence problems. TOG, 2016. 2 +[80] Y Soun and Klaus Truemper. Single commodity representation of multicommodity networks. SIAM Journal on Algebraic Discrete Methods, 1980. 6 +[81] Kenshi Takayama. Compatible intrinsic triangulations. TOG, 2022. 2 +[82] Gary KL Tam, Zhi-Quan Cheng, Yu-Kun Lai, Frank C Langbein, Yonghuai Liu, David Marshall, Ralph R Martin, Xian-Fang Sun, and Paul L Rosin. Registration of 3d point clouds and meshes: A survey from rigid to nonrigid. TVCG, 2012. 2 +[83] Giovanni Trappolini, Luca Cosmo, Luca Moschella, Ricardo Marin, Simone Melzi, and Emanuele Rodola. Shape registration in the time of transformers. NeurIPS, 2021. 2 +[84] Klaus Truemper. Unimodular matrices of flow problems with additional constraints. Networks, 1977. 6 +[85] Oliver Van Kaick, Hao Zhang, Ghassan Hamarneh, and Daniel Cohen-Or. A survey on shape correspondence. In CGF, 2011. 2 +[86] Matthias Van Onsem, Dieter De Paepe, S Vanden Hautte, Pieter Bonte, Veerle Ledoux, Annelies Lejon, Femke Ongenae, Dennis Dreesen, and Sofie Van Hoecke. Hierarchical pattern matching for anomaly detection in time series. Computer Communications, 2022. 1 +[87] Matthias Vestner, Roee Litman, Emanuele Rodola, Alex Bronstein, and Daniel Cremers. Product manifold filter: Non-rigid shape correspondence via kernel density estimation in the product space. In CVPR, 2017. 2 +[88] Ruben Wiersma, Elmar Eisemann, and Klaus Hildebrandt. Cnns on surfaces using rotation-equivariant features. TOG, 2020. 2 +[89] Thomas Windheuser, Ulrich Schlickewei, Frank R Schmidt, and Daniel Cremers. Geometrically consistent elastic matching of 3d shapes: A linear programming solution. In ICCV, 2011. 1, 2, 7, 6 +[90] Thomas Windheuser, Ulrich Schlickwei, Frank R Schimdt, and Daniel Cremers. Large-scale integer linear programming for orientation preserving 3d shape matching. In CGF, 2011. 2 +[91] Junchi Yan, Xu-Cheng Yin, Weiyao Lin, Cheng Deng, Hongyuan Zha, and Xiaokang Yang. A short survey of recent advances in graph matching. In ICMR, 2016. 1 +[92] Silvia Zuffi, Angjoo Kanazawa, David W Jacobs, and Michael J Black. 3d menagerie: Modeling the 3d shape and pose of animals. In CVPR, 2017. 7 \ No newline at end of file diff --git a/2025/Fast Globally Optimal and Geometrically Consistent 3D Shape Matching/images.zip b/2025/Fast Globally Optimal and Geometrically Consistent 3D Shape Matching/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..bb0da949913ac3fe4c9dc34ce5aa9d555fcc02c6 --- /dev/null +++ b/2025/Fast Globally Optimal and Geometrically Consistent 3D Shape Matching/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ce6fd2443101e8851bc2b9c503c61cb750cfd822f4b13102e07c2ce34bc456c5 +size 411635 diff --git a/2025/Fast Globally Optimal and Geometrically Consistent 3D Shape Matching/layout.json b/2025/Fast Globally Optimal and Geometrically Consistent 3D Shape Matching/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..24a322ae6d98e252221aed4d91d63fe9305b985a --- /dev/null +++ b/2025/Fast Globally Optimal and Geometrically Consistent 3D Shape Matching/layout.json @@ -0,0 +1,12962 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 80, + 103, + 531, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 103, + 531, + 121 + ], + "spans": [ + { + "bbox": [ + 80, + 103, + 531, + 121 + ], + "type": "text", + "content": "Fast Globally Optimal and Geometrically Consistent 3D Shape Matching" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 197, + 142, + 301, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 142, + 301, + 171 + ], + "spans": [ + { + "bbox": [ + 197, + 142, + 301, + 171 + ], + "type": "text", + "content": "Paul Roetzer1,2 \n1 University of Bonn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 307, + 142, + 410, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 142, + 410, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 142, + 410, + 171 + ], + "type": "text", + "content": "Florian Bernard" + }, + { + "bbox": [ + 307, + 142, + 410, + 171 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 307, + 142, + 410, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 307, + 142, + 410, + 171 + ], + "type": "text", + "content": " Lamarr Institute" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 47, + 229, + 102, + 285 + ], + "blocks": [ + { + "bbox": [ + 47, + 229, + 102, + 285 + ], + "lines": [ + { + "bbox": [ + 47, + 229, + 102, + 285 + ], + "spans": [ + { + "bbox": [ + 47, + 229, + 102, + 285 + ], + "type": "image", + "image_path": "0a98acca2030be1129c670887f981b56749839a31efa968a436fe3aad142706c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 95, + 301, + 183, + 312 + ], + "lines": [ + { + "bbox": [ + 95, + 301, + 183, + 312 + ], + "spans": [ + { + "bbox": [ + 95, + 301, + 183, + 312 + ], + "type": "text", + "content": "Key Idea of Our Approach" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 105, + 236, + 149, + 281 + ], + "blocks": [ + { + "bbox": [ + 105, + 236, + 149, + 281 + ], + "lines": [ + { + "bbox": [ + 105, + 236, + 149, + 281 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 149, + 281 + ], + "type": "image", + "image_path": "da7792113543c129534f8b3b3f8ae27c7ca047d050c5a4f5d44f4a5ca4dc2900.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 319, + 555, + 396 + ], + "lines": [ + { + "bbox": [ + 54, + 319, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 54, + 319, + 555, + 396 + ], + "type": "text", + "content": "Figure 1. Left: Illustration of our approach for geometrically consistent matching of 3D shape " + }, + { + "bbox": [ + 54, + 319, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 54, + 319, + 555, + 396 + ], + "type": "text", + "content": " to 3D shape " + }, + { + "bbox": [ + 54, + 319, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}" + }, + { + "bbox": [ + 54, + 319, + 555, + 396 + ], + "type": "text", + "content": ". We represent shape " + }, + { + "bbox": [ + 54, + 319, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 54, + 319, + 555, + 396 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 54, + 319, + 555, + 396 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 54, + 319, + 555, + 396 + ], + "type": "text", + "content": " surface cycles " + }, + { + "bbox": [ + 54, + 319, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_1, \\ldots, \\mathcal{C}_n" + }, + { + "bbox": [ + 54, + 319, + 555, + 396 + ], + "type": "text", + "content": " and then consistently match these to shape " + }, + { + "bbox": [ + 54, + 319, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}" + }, + { + "bbox": [ + 54, + 319, + 555, + 396 + ], + "type": "text", + "content": " while preserving neighbourhood relations between the cycles. Middle: The triangulation- and colour-transfer between shapes using computed matchings of Cao et al. [13] (not geometrically consistent) and ours (geometrically consistent) illustrates the importance of geometric consistency. Right: Our method is the first 3D shape matching approach that yields globally geometrically consistent matchings, that is globally optimal, and scalable in practice, opposed to all competing methods (OOM stands for 'out of memory' and the dashed line indicates a weak notion of geometric consistency, see explanation of methods in Sec. 5)." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 149, + 236, + 204, + 281 + ], + "blocks": [ + { + "bbox": [ + 149, + 236, + 204, + 281 + ], + "lines": [ + { + "bbox": [ + 149, + 236, + 204, + 281 + ], + "spans": [ + { + "bbox": [ + 149, + 236, + 204, + 281 + ], + "type": "image", + "image_path": "7859227a1ef7048dba4adb49f6a26daca995307b19acecec2690ee0441d4cc5e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 225, + 196, + 260, + 288 + ], + "blocks": [ + { + "bbox": [ + 225, + 196, + 260, + 288 + ], + "lines": [ + { + "bbox": [ + 225, + 196, + 260, + 288 + ], + "spans": [ + { + "bbox": [ + 225, + 196, + 260, + 288 + ], + "type": "image", + "image_path": "241bd2df8f0e62117ef665127831755f94d96f1844cb21e8d2c0ea8b38d81f07.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 265, + 196, + 309, + 288 + ], + "blocks": [ + { + "bbox": [ + 265, + 196, + 309, + 288 + ], + "lines": [ + { + "bbox": [ + 265, + 196, + 309, + 288 + ], + "spans": [ + { + "bbox": [ + 265, + 196, + 309, + 288 + ], + "type": "image", + "image_path": "7e9e9f9740cdb1b9eb4df97b173b15c91cd4fa72228294bb8c5f79454d001282.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 312, + 195, + 356, + 288 + ], + "blocks": [ + { + "bbox": [ + 312, + 195, + 356, + 288 + ], + "lines": [ + { + "bbox": [ + 312, + 195, + 356, + 288 + ], + "spans": [ + { + "bbox": [ + 312, + 195, + 356, + 288 + ], + "type": "image", + "image_path": "cf3ff8894f376bfda227b0ff94dfd96b1bcb13c2271ce3d1100c715e5322205b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 369, + 200, + 549, + 303 + ], + "blocks": [ + { + "bbox": [ + 214, + 293, + 370, + 312 + ], + "lines": [ + { + "bbox": [ + 214, + 293, + 370, + 312 + ], + "spans": [ + { + "bbox": [ + 214, + 293, + 370, + 312 + ], + "type": "text", + "content": "Source Cao et al. Ours Triangulation Transfer via Computed Matchings" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 369, + 200, + 549, + 303 + ], + "lines": [ + { + "bbox": [ + 369, + 200, + 549, + 303 + ], + "spans": [ + { + "bbox": [ + 369, + 200, + 549, + 303 + ], + "type": "image", + "image_path": "759b602e79dc3967063d7b58bbd065350cf5f611a2f65b8f1de1ded492b331af.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 303, + 535, + 312 + ], + "lines": [ + { + "bbox": [ + 403, + 303, + 535, + 312 + ], + "spans": [ + { + "bbox": [ + 403, + 303, + 535, + 312 + ], + "type": "text", + "content": "Runtime Comparison on FAUST Dataset" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 151, + 412, + 200, + 424 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 412, + 200, + 424 + ], + "spans": [ + { + "bbox": [ + 151, + 412, + 200, + 424 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 439, + 296, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 439, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 439, + 296, + 677 + ], + "type": "text", + "content": "Geometric consistency, i.e. the preservation of neighbourhoods, is a natural and strong prior in 3D shape matching. Geometrically consistent matchings are crucial for many downstream applications, such as texture transfer or statistical shape modelling. Yet, in practice, geometric consistency is often overlooked, or only achieved under severely limiting assumptions (e.g. a good initialisation). In this work, we propose a novel formalism for computing globally optimal and geometrically consistent matchings between 3D shapes which is scalable in practice. Our key idea is to represent the surface of the source shape as a collection of cyclic graphs, which are then consistently matched to the target shape. Mathematically, we construct a hyper product graph (between source and target shape), and then cast 3D shape matching as a minimum-cost circulation flow problem in this hyper graph, which yields global geometrically consistent matchings between both shapes. We empirically show that our formalism is efficiently solvable and that it leads to high-quality results. Our code is publicly available.1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 412, + 394, + 424 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 412, + 394, + 424 + ], + "spans": [ + { + "bbox": [ + 315, + 412, + 394, + 424 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 435, + 555, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 555, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 555, + 688 + ], + "type": "text", + "content": "The availability of correspondences between visual data are a key prerequisite for many visual computing tasks, including 3D reconstruction [50], loss computation for deep learning [14], protein alignment [45], anomaly detection [86], object recognition [44], shape modelling [24] and others. In many practical cases, it is desirable that correspondences preserve neighbourhood relations. For example, in 3D shape matching this could mean that when bringing a pair of neighbouring points on a source shape into correspondence with a pair of points on a target shape, the pair of points must remain neighbours on the target. Such a neighbourhood preservation can serve as strong topological prior, for example to resolve ambiguities, achieve robustness, or ensure well-posedness of an otherwise ill-posed problem. However, despite its crucial importance, neighbourhood preservation is in practice often-times overlooked. This is because many of the well-known formalisms are not efficiently solvable (e.g. the quadratic assignment problem [64], or general integer programming [76], which have been used to tackle image keypoint matching [55], graph matching [39, 91], or 3D shape matching [8, 34, 58, 89])." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 690, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 690, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 690, + 555, + 715 + ], + "type": "text", + "content": "In this work we specifically focus on the task of non-rigid geometrically consistent 3D shape matching, in which" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 703, + 231, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 703, + 231, + 714 + ], + "spans": [ + { + "bbox": [ + 66, + 703, + 231, + 714 + ], + "type": "text", + "content": "1https://github.com/paul0noah/geco" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "912" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 74, + 70, + 277, + 162 + ], + "blocks": [ + { + "bbox": [ + 74, + 70, + 277, + 162 + ], + "lines": [ + { + "bbox": [ + 74, + 70, + 277, + 162 + ], + "spans": [ + { + "bbox": [ + 74, + 70, + 277, + 162 + ], + "type": "table", + "html": "
MethodGlobally OptimalGeometr. ConsistentScalable
MINA [8](✓)✗✗
SIGMA [34](✓)✗✗
PMSDP [58]✗✗✓
Windheuser et al. [89]✗✓✗
SpiderMatch [67]✓(✓)(✓)
Ours✓✓✓
", + "image_path": "c5fed2dade15da79455c5fefca16af65aaa75c7030a7c42cb55f0dcf2b8c6c87.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 171, + 293, + 182 + ], + "lines": [ + { + "bbox": [ + 57, + 171, + 293, + 182 + ], + "spans": [ + { + "bbox": [ + 57, + 171, + 293, + 182 + ], + "type": "text", + "content": "Table 1. Comparison of axiomatic 3D shape matching methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 56, + 197, + 295, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 295, + 423 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 295, + 423 + ], + "type": "text", + "content": "neighbourhood-preserving correspondences between two given 2D manifolds (embedded in 3D space) are sought for. So far, there does not exist any 3D shape matching approach that combines the following three desirable properties: (i) global optimality, (ii) neighbourhood preservation, and (iii) scalability, see Tab. 1. While the recent approach SpiderMatch [67] fulfils (i)-(iii) to some extent, its scalability relies heavily on the choice of the cycle that is used to represent one of the shapes (cf. Fig. 1 right: solid vs. dashed purple line), thereby trading off neighbourhood preservation with scalability. Furthermore, proper geometric consistency (Definition 4) can only be enforced at intersection points of the cycle and thus also the quality of geometric consistency depends on the choice of the cycle. Inspired by this idea of representing a shape using a cycle, we represent one shape using a collection of cycles and show that this leads to an (in practice) efficiently solvable formalism which ensures proper geometric consistency. We summarise our main contributions as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 425, + 295, + 604 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 57, + 425, + 295, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 425, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 57, + 425, + 295, + 460 + ], + "type": "text", + "content": "- For the first time, we present a globally geometrically consistent formalism for 3D shape matching that is efficiently solvable to global optimality in practice." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 460, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 460, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 57, + 460, + 295, + 521 + ], + "type": "text", + "content": "- To achieve this, we introduce a novel 3D shape representation in which we represent a 3D shape using a collection of surface cycles, so that 3D shape matching can be cast as finding a minimum-cost flow circulation in a hyper product graph." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 521, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 521, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 57, + 521, + 295, + 556 + ], + "type": "text", + "content": "- We experimentally show that our formalism leads to high quality and geometrically consistent matchings between two 3D shapes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 557, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 557, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 57, + 557, + 295, + 604 + ], + "type": "text", + "content": "- In addition to 3D shape matching, we also show, in a proof-of-concept manner, that our formalism can generalise to specific instances of other matching problems, such as graph matching when the source graph is planar." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 619, + 142, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 619, + 142, + 632 + ], + "spans": [ + { + "bbox": [ + 55, + 619, + 142, + 632 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 640, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 640, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 640, + 295, + 688 + ], + "type": "text", + "content": "In the following we discuss works that are most relevant to our approach. We start by discussing efficiently solvable matching problems, continue with 3D shape matching, and conclude with geometrically consistent 3D shape matching." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Efficiently Solvable Matching Problems. There are various instances of correspondence problems that can be" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "text", + "content": "solved efficiently. Among them is the linear assignment problem (LAP) [46], which matches points without considering their neighbourhood relations. For matching time series data, sequences, or (open) contours, the popular dynamic time warping algorithm can efficiently compute solutions while preserving neighbourhood relations [70]. Closed contours can be matched analogously via graph cuts in product graphs [71]. Certain tracking problems can be formulated as efficiently solvable flow problems [49], and model-based image segmentation can also be solved efficiently by matching a 2D contour to an image [18, 32, 74]. The matching of a 2D contour to a 3D shape can be solved using variants of Dijkstra's algorithm for finding minimum cost cycles in product graphs [48, 68]. The mentioned approaches show that there are formalisms to efficiently solve a diverse range of matching problems. Yet, these approaches do not directly generalise to 3D shape matching. In this work, we propose a novel geometrically consistent 3D shape matching formalism that is (in practice) efficiently solvable to global optimality." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 315, + 553, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 553, + 495 + ], + "type": "text", + "content": "3D Shape Matching is the task of finding correspondences between two non-rigidly deformed surfaces. For an in-depth overview on 3D shape matching we refer the reader to survey papers [19, 82, 85]. Many axiomatic shape matching approaches [11, 60, 65] build on the functional maps framework, efficiently solving shape matching in the spectral domain [28, 35, 56, 58, 59, 62]. In addition, functional maps have been adopted in several deep shape matching variants, either trained in a supervised [36, 52, 54, 83, 88] or in a unsupervised manner [2, 4, 13, 20, 22, 38, 51, 78]. There are also alternative approaches based on consensus maximisation [61], or using convex relaxations [15, 58] or mixed integer programming [8, 34]. Yet, many of the existing 3D shape matching methods neglect geometric consistency, as it leads to hard-to-solve formalisms." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 498, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 553, + 713 + ], + "type": "text", + "content": "Geometrically Consistent 3D Shape Matching. There are several approaches that have recognised the importance of geometric consistency in 3D shape matching. Some of them incorporate neighbourhood information by using formalisms based on the quadratic assignment problem (QAP) - yet, the QAP is NP-hard [64], so that respective approaches consider relaxations [12, 23, 47] or heuristics [7, 40, 79]. Opposed to such discrete formulations, there are also approaches that tackle 3D shape matching by deforming a continuous shape parametrisation using local optimisation. However, due to the severe nonconvexity of resulting problems, they rely on a good initialisation, e.g. in the form of sparse sets of landmark correspondences [72, 73, 75, 77, 81], or in the form of dense correspondences [30, 31, 87]. Windheuser et al. [89, 90] have modelled geometrically consistent shape matching in the discrete domain by matching triangles to triangles using an (expensive to solve) integer linear program. Re" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "913" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 294, + 231 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 294, + 231 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 294, + 231 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 294, + 231 + ], + "type": "table", + "html": "
SymbolDescription
X=(Vx,FX)3D shape
Gx=(Vx,EX)3D shape graph of X
C=(VC,EC)surface cycle (cycle in Gx)
{C1,...,Cn}Representation of X with n surface cycles
nNumber of surface cycles
Y=(Vy, fy)3D shape Y
Gy=(Vy, Fy)3D shape graph of Y
Pi=(VPi, EPi)Product graph (of C_i and mesh Y)
{P1,...,Pn}Product graph collection (of {C1,...,Cn} and mesh Y)
H=(VH, EH)Hyper product graph (coupled product graphs {P,...,Pn})
mNumber of hyper edges |EH|
HVertex edge incidence matrices of H
", + "image_path": "69b5e98be273fdc8f74d031b2df2ed0b889e0a3fb5d92e9ecc0932e085e8139f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 78, + 234, + 271, + 245 + ], + "lines": [ + { + "bbox": [ + 78, + 234, + 271, + 245 + ], + "spans": [ + { + "bbox": [ + 78, + 234, + 271, + 245 + ], + "type": "text", + "content": "Table 2. Summary of the notation used in this paper." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 266, + 296, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 296, + 542 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 296, + 542 + ], + "type": "text", + "content": "cently, approximative solvers [66, 69] and extensions for partial shapes [26, 27] have been proposed. Furthermore, 3D shape matching has recently been formulated as a shortest path problem with additional constraints [67]. In this work, the surface mesh of the source 3D shape is represented using a long self-intersecting curve that traces the shape surface, and is then matched to the other shape while preserving intersections of the curve. While their idea of considering alternative 3D shape representations is promising, in the presented framework proper geometric consistency is only enforced at intersection points of the curve (see Tab. 4). In addition, runtimes grow drastically with an increasing number of intersection points, which can be seen in Fig. 1 right when comparing SpiderMatch [67] (using a curve with few intersections) to SpiderMatch (geo.) (using a curve with many intersections). Inspired by the idea of an alternative path-based 3D shape representation, we represent a 3D shape as a collection of cyclic graphs, and then tackle 3D shape matching by solving coupled matching problems of the individual cyclic paths. Based on the couplings, we are able to ensure global geometry consistency. Furthermore, we empirically observe that the resulting formalism is efficiently solvable to global optimality." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 552, + 134, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 134, + 565 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 134, + 565 + ], + "type": "text", + "content": "3. Background" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 572, + 295, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 572, + 295, + 619 + ], + "spans": [ + { + "bbox": [ + 55, + 572, + 295, + 619 + ], + "type": "text", + "content": "We consider the task of finding a matching between a source shape " + }, + { + "bbox": [ + 55, + 572, + 295, + 619 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 55, + 572, + 295, + 619 + ], + "type": "text", + "content": " and a target shape " + }, + { + "bbox": [ + 55, + 572, + 295, + 619 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}" + }, + { + "bbox": [ + 55, + 572, + 295, + 619 + ], + "type": "text", + "content": " (Sec. 3.1) such that the matching is geometrically consistent (Sec. 3.2). Our main notation is summarised in Tab. 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 628, + 170, + 640 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 170, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 170, + 640 + ], + "type": "text", + "content": "3.1. Shapes and Graphs" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 645, + 295, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 645, + 295, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 645, + 295, + 681 + ], + "type": "text", + "content": "In the following, we define shapes and other relevant concepts, which are also illustrated in Fig. 2. We consider " + }, + { + "bbox": [ + 55, + 645, + 295, + 681 + ], + "type": "inline_equation", + "content": "3D" + }, + { + "bbox": [ + 55, + 645, + 295, + 681 + ], + "type": "text", + "content": " shapes represented as triangular surface mesh:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Definition 1 (3D shape). A 3D shape " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": " is defined as a tuple " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "(\\mathcal{V}_{\\mathcal{X}},\\mathcal{F}_{\\mathcal{X}})" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": " of vertices " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{\\mathcal{X}}" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": " and consistently oriented" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 338, + 76, + 531, + 141 + ], + "blocks": [ + { + "bbox": [ + 338, + 76, + 531, + 141 + ], + "lines": [ + { + "bbox": [ + 338, + 76, + 531, + 141 + ], + "spans": [ + { + "bbox": [ + 338, + 76, + 531, + 141 + ], + "type": "image", + "image_path": "955846ec6a4e5bd46dfbf3f4b8d7e861f28a79d13f04de3bb40351afdd4e5f47.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 151, + 553, + 174 + ], + "lines": [ + { + "bbox": [ + 313, + 151, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 313, + 151, + 553, + 174 + ], + "type": "text", + "content": "Figure 2. We represent a 3D shape (left) as a (directed) shape graph " + }, + { + "bbox": [ + 313, + 151, + 553, + 174 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\mathcal{X}}" + }, + { + "bbox": [ + 313, + 151, + 553, + 174 + ], + "type": "text", + "content": " (middle). We call a cycle " + }, + { + "bbox": [ + 313, + 151, + 553, + 174 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 151, + 553, + 174 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 151, + 553, + 174 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\mathcal{X}}" + }, + { + "bbox": [ + 313, + 151, + 553, + 174 + ], + "type": "text", + "content": " a surface cycle (right)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 194, + 553, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 553, + 230 + ], + "type": "text", + "content": "(e.g. clock-wise) triangles " + }, + { + "bbox": [ + 313, + 194, + 553, + 230 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathcal{X}} \\subset \\mathcal{V}_{\\mathcal{X}} \\times \\mathcal{V}_{\\mathcal{X}} \\times \\mathcal{V}_{\\mathcal{X}}" + }, + { + "bbox": [ + 313, + 194, + 553, + 230 + ], + "type": "text", + "content": ", such that " + }, + { + "bbox": [ + 313, + 194, + 553, + 230 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 313, + 194, + 553, + 230 + ], + "type": "text", + "content": " forms an orientable continuous " + }, + { + "bbox": [ + 313, + 194, + 553, + 230 + ], + "type": "inline_equation", + "content": "2D" + }, + { + "bbox": [ + 313, + 194, + 553, + 230 + ], + "type": "text", + "content": " manifold (possibly with boundary) embedded in " + }, + { + "bbox": [ + 313, + 194, + 553, + 230 + ], + "type": "inline_equation", + "content": "3D" + }, + { + "bbox": [ + 313, + 194, + 553, + 230 + ], + "type": "text", + "content": " space." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 239, + 553, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 553, + 262 + ], + "type": "text", + "content": "For our formalism it is advantageous to interpret a 3D shape as a graph, which we denote the shape graph:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 270, + 553, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 553, + 319 + ], + "type": "text", + "content": "Definition 2 (Shape graph). The shape graph (of 3D shape " + }, + { + "bbox": [ + 313, + 270, + 553, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 313, + 270, + 553, + 319 + ], + "type": "text", + "content": ") is a tuple " + }, + { + "bbox": [ + 313, + 270, + 553, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\mathcal{X}} = (\\mathcal{V}_{\\mathcal{X}}, \\mathcal{E}_{\\mathcal{X}})" + }, + { + "bbox": [ + 313, + 270, + 553, + 319 + ], + "type": "text", + "content": " of vertices " + }, + { + "bbox": [ + 313, + 270, + 553, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{\\mathcal{X}}" + }, + { + "bbox": [ + 313, + 270, + 553, + 319 + ], + "type": "text", + "content": " and directed edges " + }, + { + "bbox": [ + 313, + 270, + 553, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{X}} \\subset \\mathcal{V}_{\\mathcal{X}} \\times \\mathcal{V}_{\\mathcal{X}}" + }, + { + "bbox": [ + 313, + 270, + 553, + 319 + ], + "type": "text", + "content": ", such that each triangle in " + }, + { + "bbox": [ + 313, + 270, + 553, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathcal{X}}" + }, + { + "bbox": [ + 313, + 270, + 553, + 319 + ], + "type": "text", + "content": " is represented by three unique and consistently directed edges." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 327, + 553, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 553, + 350 + ], + "type": "text", + "content": "In addition to the shape graph (representing the whole 3D shape " + }, + { + "bbox": [ + 313, + 327, + 553, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 313, + 327, + 553, + 350 + ], + "type": "text", + "content": "), we consider certain subgraphs of " + }, + { + "bbox": [ + 313, + 327, + 553, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\mathcal{X}}" + }, + { + "bbox": [ + 313, + 327, + 553, + 350 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 358, + 553, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 553, + 418 + ], + "type": "text", + "content": "Definition 3 (Surface cycle). A surface cycle (on 3D shape " + }, + { + "bbox": [ + 313, + 358, + 553, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 313, + 358, + 553, + 418 + ], + "type": "text", + "content": ") is a tuple " + }, + { + "bbox": [ + 313, + 358, + 553, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = (\\mathcal{V}_{\\mathcal{C}}, \\mathcal{E}_{\\mathcal{C}})" + }, + { + "bbox": [ + 313, + 358, + 553, + 418 + ], + "type": "text", + "content": " with (non-empty) vertices " + }, + { + "bbox": [ + 313, + 358, + 553, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{\\mathcal{C}} \\subset \\mathcal{V}_{\\mathcal{X}}" + }, + { + "bbox": [ + 313, + 358, + 553, + 418 + ], + "type": "text", + "content": " and (non-empty) directed edges " + }, + { + "bbox": [ + 313, + 358, + 553, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{C}} \\subset \\mathcal{V}_{\\mathcal{C}} \\times \\mathcal{V}_{\\mathcal{C}}" + }, + { + "bbox": [ + 313, + 358, + 553, + 418 + ], + "type": "text", + "content": ", such that (i) " + }, + { + "bbox": [ + 313, + 358, + 553, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{C}} \\subset \\mathcal{E}_{\\mathcal{X}}" + }, + { + "bbox": [ + 313, + 358, + 553, + 418 + ], + "type": "text", + "content": ", and (ii) each vertex " + }, + { + "bbox": [ + 313, + 358, + 553, + 418 + ], + "type": "inline_equation", + "content": "v \\in \\mathcal{V}_{\\mathcal{C}}" + }, + { + "bbox": [ + 313, + 358, + 553, + 418 + ], + "type": "text", + "content": " has exactly one incoming and one outgoing edge in " + }, + { + "bbox": [ + 313, + 358, + 553, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{C}}" + }, + { + "bbox": [ + 313, + 358, + 553, + 418 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 434, + 550, + 447 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 550, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 550, + 447 + ], + "type": "text", + "content": "3.2. Geometrically Consistent 3D Shape Matching" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 451, + 553, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 553, + 475 + ], + "type": "text", + "content": "We define geometrically consistent 3D shape matching in terms of the respective shape graphs:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 483, + 554, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 554, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 554, + 591 + ], + "type": "text", + "content": "Definition 4 (Geometrically Consistent Matching). The mapping " + }, + { + "bbox": [ + 313, + 483, + 554, + 591 + ], + "type": "inline_equation", + "content": "\\phi : \\mathcal{V}_{\\mathcal{X}} \\to \\mathcal{V}_{\\mathcal{Y}}" + }, + { + "bbox": [ + 313, + 483, + 554, + 591 + ], + "type": "text", + "content": " is a geometrically consistent matching from the source shape " + }, + { + "bbox": [ + 313, + 483, + 554, + 591 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 313, + 483, + 554, + 591 + ], + "type": "text", + "content": " to the target shape " + }, + { + "bbox": [ + 313, + 483, + 554, + 591 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}" + }, + { + "bbox": [ + 313, + 483, + 554, + 591 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 313, + 483, + 554, + 591 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 483, + 554, + 591 + ], + "type": "text", + "content": " preserves neighbourhoods as follows: whenever two vertices " + }, + { + "bbox": [ + 313, + 483, + 554, + 591 + ], + "type": "inline_equation", + "content": "x, \\bar{x} \\in \\mathcal{V}_{\\mathcal{X}}" + }, + { + "bbox": [ + 313, + 483, + 554, + 591 + ], + "type": "text", + "content": " are connected (i.e. " + }, + { + "bbox": [ + 313, + 483, + 554, + 591 + ], + "type": "inline_equation", + "content": "(x, \\bar{x}) \\in \\mathcal{E}_{\\mathcal{X}}" + }, + { + "bbox": [ + 313, + 483, + 554, + 591 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 483, + 554, + 591 + ], + "type": "inline_equation", + "content": "(\\bar{x}, x) \\in \\mathcal{E}_{\\mathcal{X}}" + }, + { + "bbox": [ + 313, + 483, + 554, + 591 + ], + "type": "text", + "content": "), then their corresponding vertices " + }, + { + "bbox": [ + 313, + 483, + 554, + 591 + ], + "type": "inline_equation", + "content": "\\phi(x), \\phi(\\bar{x}) \\in \\mathcal{V}_{\\mathcal{Y}}" + }, + { + "bbox": [ + 313, + 483, + 554, + 591 + ], + "type": "text", + "content": " must either be (i) connected (i.e. " + }, + { + "bbox": [ + 313, + 483, + 554, + 591 + ], + "type": "inline_equation", + "content": "(\\phi(x), \\phi(\\bar{x})) \\in \\mathcal{E}_{\\mathcal{Y}}" + }, + { + "bbox": [ + 313, + 483, + 554, + 591 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 483, + 554, + 591 + ], + "type": "inline_equation", + "content": "(\\phi(\\bar{x}), \\phi(x)) \\in \\mathcal{E}_{\\mathcal{Y}}" + }, + { + "bbox": [ + 313, + 483, + 554, + 591 + ], + "type": "text", + "content": "), or (ii) the same vertex (i.e. " + }, + { + "bbox": [ + 313, + 483, + 554, + 591 + ], + "type": "inline_equation", + "content": "\\phi(x) = \\phi(\\bar{x})" + }, + { + "bbox": [ + 313, + 483, + 554, + 591 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 598, + 554, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 598, + 554, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 598, + 554, + 647 + ], + "type": "text", + "content": "The intuition is that neighbouring elements of " + }, + { + "bbox": [ + 313, + 598, + 554, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 313, + 598, + 554, + 647 + ], + "type": "text", + "content": " must be matched to neighbouring elements of " + }, + { + "bbox": [ + 313, + 598, + 554, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}" + }, + { + "bbox": [ + 313, + 598, + 554, + 647 + ], + "type": "text", + "content": ". We note that we require both shapes to have equal genus such that our definition of geometric consistency makes sense." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 658, + 507, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 507, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 507, + 671 + ], + "type": "text", + "content": "4. Our 3D Shape Matching Approach" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "In this section we develop our geometrically consistent 3D shape matching formulation which ensures global geometric consistency. Our approach is based on representing the" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "914" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 89, + 71, + 265, + 133 + ], + "blocks": [ + { + "bbox": [ + 89, + 71, + 265, + 133 + ], + "lines": [ + { + "bbox": [ + 89, + 71, + 265, + 133 + ], + "spans": [ + { + "bbox": [ + 89, + 71, + 265, + 133 + ], + "type": "image", + "image_path": "08f95eed8fca9222ec95803c635bf89d5cd3957a8923d6682e657d4f0f26eb72.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 143, + 297, + 189 + ], + "lines": [ + { + "bbox": [ + 55, + 143, + 297, + 189 + ], + "spans": [ + { + "bbox": [ + 55, + 143, + 297, + 189 + ], + "type": "text", + "content": "Figure 3. Surface cycle collection (middle) for representing the 3D shape of an icosahedron (left). Individual surface cycles are glued together via shared edges (neighbouring surface cycles share opposite edges, cf. pairs of thickened edges on the right: " + }, + { + "bbox": [ + 55, + 143, + 297, + 189 + ], + "type": "inline_equation", + "content": "\\parallel" + }, + { + "bbox": [ + 55, + 143, + 297, + 189 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 143, + 297, + 189 + ], + "type": "inline_equation", + "content": "\\parallel" + }, + { + "bbox": [ + 55, + 143, + 297, + 189 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 209, + 296, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 209, + 296, + 317 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 296, + 317 + ], + "type": "text", + "content": "source 3D shape using a collection of surface cycles, which we then match to the target shape (Sec. 4.1). To this end, we formulate an individual subproblem for each surface cycle (Sec. 4.2.1). Further, we couple individual subproblems (Secs. 4.2.2 and 4.2.3) so that our matchings are geometrically consistent according to Definition 4. Finally, we show that our overall formalism can be interpreted as a minimum-cost circulation flow problem on a hyper product graph (Sec. 4.5)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 324, + 295, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 295, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 295, + 338 + ], + "type": "text", + "content": "4.1. Surface Cycle-based 3D Shape Representation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 342, + 296, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 296, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 296, + 378 + ], + "type": "text", + "content": "We aim to represent the surface of 3D shape " + }, + { + "bbox": [ + 55, + 342, + 296, + 378 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 55, + 342, + 296, + 378 + ], + "type": "text", + "content": " with a collection of surface cycles which are glued together at opposite edges:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 387, + 296, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 387, + 296, + 412 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 296, + 412 + ], + "type": "text", + "content": "Definition 5 (Opposite edge). The opposite edge to an edge " + }, + { + "bbox": [ + 55, + 387, + 296, + 412 + ], + "type": "inline_equation", + "content": "(v, w)" + }, + { + "bbox": [ + 55, + 387, + 296, + 412 + ], + "type": "text", + "content": " is defined as " + }, + { + "bbox": [ + 55, + 387, + 296, + 412 + ], + "type": "inline_equation", + "content": "-(v, w) := (w, v)" + }, + { + "bbox": [ + 55, + 387, + 296, + 412 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 419, + 296, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 296, + 468 + ], + "type": "text", + "content": "We note that for every non-boundary edge " + }, + { + "bbox": [ + 55, + 419, + 296, + 468 + ], + "type": "inline_equation", + "content": "e \\in \\mathcal{E}_{\\mathcal{X}}" + }, + { + "bbox": [ + 55, + 419, + 296, + 468 + ], + "type": "text", + "content": " of shape " + }, + { + "bbox": [ + 55, + 419, + 296, + 468 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 55, + 419, + 296, + 468 + ], + "type": "text", + "content": ", the opposite edge " + }, + { + "bbox": [ + 55, + 419, + 296, + 468 + ], + "type": "inline_equation", + "content": "-e \\in \\mathcal{E}_{\\mathcal{X}}" + }, + { + "bbox": [ + 55, + 419, + 296, + 468 + ], + "type": "text", + "content": " is also part of the shape graph " + }, + { + "bbox": [ + 55, + 419, + 296, + 468 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\mathcal{X}}" + }, + { + "bbox": [ + 55, + 419, + 296, + 468 + ], + "type": "text", + "content": " (cf. Fig. 3 right) because the pair of opposite edges " + }, + { + "bbox": [ + 55, + 419, + 296, + 468 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 55, + 419, + 296, + 468 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 419, + 296, + 468 + ], + "type": "inline_equation", + "content": "-e" + }, + { + "bbox": [ + 55, + 419, + 296, + 468 + ], + "type": "text", + "content": " belong to two neighbouring triangles." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 475, + 296, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 296, + 523 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 296, + 523 + ], + "type": "text", + "content": "Definition 6 (Shape as collection of surface cycles). We represent a 3D shape " + }, + { + "bbox": [ + 55, + 475, + 296, + 523 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 55, + 475, + 296, + 523 + ], + "type": "text", + "content": " with a collection of " + }, + { + "bbox": [ + 55, + 475, + 296, + 523 + ], + "type": "inline_equation", + "content": "n \\in \\mathbb{N}^+" + }, + { + "bbox": [ + 55, + 475, + 296, + 523 + ], + "type": "text", + "content": " surface cycles " + }, + { + "bbox": [ + 55, + 475, + 296, + 523 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_1, \\ldots, \\mathcal{C}_n" + }, + { + "bbox": [ + 55, + 475, + 296, + 523 + ], + "type": "text", + "content": " (on shape " + }, + { + "bbox": [ + 55, + 475, + 296, + 523 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 55, + 475, + 296, + 523 + ], + "type": "text", + "content": ") that partition the surface of " + }, + { + "bbox": [ + 55, + 475, + 296, + 523 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 55, + 475, + 296, + 523 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 55, + 475, + 296, + 523 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 475, + 296, + 523 + ], + "type": "text", + "content": " polygonal patches, such that:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 523, + 296, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 523, + 296, + 561 + ], + "spans": [ + { + "bbox": [ + 55, + 523, + 296, + 561 + ], + "type": "text", + "content": "(i) " + }, + { + "bbox": [ + 55, + 523, + 296, + 561 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{C}_i}\\cap \\mathcal{E}_{\\mathcal{C}_j} = \\emptyset" + }, + { + "bbox": [ + 55, + 523, + 296, + 561 + ], + "type": "text", + "content": " for all " + }, + { + "bbox": [ + 55, + 523, + 296, + 561 + ], + "type": "inline_equation", + "content": "i,j\\in \\{1,\\ldots ,n\\} ,i\\neq j" + }, + { + "bbox": [ + 55, + 523, + 296, + 561 + ], + "type": "text", + "content": " , and (ii) for every non-boundary edge " + }, + { + "bbox": [ + 55, + 523, + 296, + 561 + ], + "type": "inline_equation", + "content": "e\\in \\mathcal{E}_{\\mathcal{C}_i}" + }, + { + "bbox": [ + 55, + 523, + 296, + 561 + ], + "type": "text", + "content": " of shape " + }, + { + "bbox": [ + 55, + 523, + 296, + 561 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 55, + 523, + 296, + 561 + ], + "type": "text", + "content": " there exists " + }, + { + "bbox": [ + 55, + 523, + 296, + 561 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 523, + 296, + 561 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 55, + 523, + 296, + 561 + ], + "type": "inline_equation", + "content": "-e\\in \\mathcal{E}_{\\mathcal{C}_j}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 568, + 296, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 296, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 296, + 641 + ], + "type": "text", + "content": "Condition (i) ensures that each edge of " + }, + { + "bbox": [ + 55, + 568, + 296, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 55, + 568, + 296, + 641 + ], + "type": "text", + "content": " can be part of at most one surface cycle, and (ii) ensures that neighbouring surface cycles cover opposite edges, see Fig. 3. We note that from here on we assume that each surface cycle represents an individual triangle of shape " + }, + { + "bbox": [ + 55, + 568, + 296, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 55, + 568, + 296, + 641 + ], + "type": "text", + "content": ". We discuss more general polygonal surface cycles in Sec. C.3 in the supplementary." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 648, + 268, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 648, + 268, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 648, + 268, + 661 + ], + "type": "text", + "content": "4.2. Our Coupled Product Graph Formalism" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "On a high-level, we use the product graph formalism introduced in [48] to address the matching between individual surface cycles and shape " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ", which results in geometric consistency along the path within each individual cycle." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": "Further, we introduce coupling constraints that have the effect of glueing neighbouring product graphs together. Combined with additional injectivity constraints, our matching is globally geometrically consistent, i.e. it preserves neighbourhoods between surface cycles, see Fig. 4 for an overview." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 154, + 550, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 154, + 550, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 154, + 550, + 167 + ], + "type": "text", + "content": "4.2.1. Individual Surface Cycle Matching Subproblems" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 171, + 554, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 171, + 554, + 219 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 554, + 219 + ], + "type": "text", + "content": "Each individual surface cycle " + }, + { + "bbox": [ + 313, + 171, + 554, + 219 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i" + }, + { + "bbox": [ + 313, + 171, + 554, + 219 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 171, + 554, + 219 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 313, + 171, + 554, + 219 + ], + "type": "text", + "content": " can be matched to shape " + }, + { + "bbox": [ + 313, + 171, + 554, + 219 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 171, + 554, + 219 + ], + "type": "text", + "content": " by finding a cyclic path in their product graph " + }, + { + "bbox": [ + 313, + 171, + 554, + 219 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i = (\\mathcal{V}_{\\mathcal{P}_i},\\mathcal{E}_{\\mathcal{P}_i})" + }, + { + "bbox": [ + 313, + 171, + 554, + 219 + ], + "type": "text", + "content": ". It has been introduced in [48] for the problem of matching a 2D shape to a 3D shape and reads:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 231, + 555, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 231, + 555, + 268 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 555, + 268 + ], + "type": "text", + "content": "Definition 7 (Product graph " + }, + { + "bbox": [ + 313, + 231, + 555, + 268 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 313, + 231, + 555, + 268 + ], + "type": "text", + "content": "). The product graph " + }, + { + "bbox": [ + 313, + 231, + 555, + 268 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i = (\\mathcal{V}_{\\mathcal{P}_i}, \\mathcal{E}_{\\mathcal{P}_i})" + }, + { + "bbox": [ + 313, + 231, + 555, + 268 + ], + "type": "text", + "content": " between the " + }, + { + "bbox": [ + 313, + 231, + 555, + 268 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 231, + 555, + 268 + ], + "type": "text", + "content": "-th surface cycle " + }, + { + "bbox": [ + 313, + 231, + 555, + 268 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i" + }, + { + "bbox": [ + 313, + 231, + 555, + 268 + ], + "type": "text", + "content": " and the 3D shape " + }, + { + "bbox": [ + 313, + 231, + 555, + 268 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}" + }, + { + "bbox": [ + 313, + 231, + 555, + 268 + ], + "type": "text", + "content": " is a directed graph defined as" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 329, + 279, + 554, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 279, + 554, + 326 + ], + "spans": [ + { + "bbox": [ + 329, + 279, + 554, + 326 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {V} _ {\\mathcal {P} _ {i}} = \\mathcal {V} _ {\\mathcal {C} _ {i}} \\times \\mathcal {V} _ {\\mathcal {Y}}, \\\\ \\mathcal {E} _ {\\mathcal {P} _ {i}} = \\left\\{\\left(v, \\bar {v}\\right) \\in \\mathcal {V} _ {\\mathcal {P} _ {i}} \\times \\mathcal {V} _ {\\mathcal {P} _ {i}} \\mid v = \\left( \\begin{array}{l} x \\\\ y \\end{array} \\right), \\bar {v} = \\left( \\begin{array}{l} \\bar {x} \\\\ \\bar {y} \\end{array} \\right) \\right. \\tag {1} \\\\ (x, \\bar {x}) \\in \\mathcal {E} _ {\\mathcal {C} _ {i}}, (y, \\bar {y}) \\in \\mathcal {E} _ {\\mathcal {Y}} ^ {+} \\}, \\\\ \\end{array}", + "image_path": "07bce47f9890e9ff16880ad8ca761c1d2eb90969feef6fbff5d43110e2049bd4.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 337, + 550, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 337, + 550, + 352 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 550, + 352 + ], + "type": "text", + "content": "with the extended edge set " + }, + { + "bbox": [ + 313, + 337, + 550, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{V}}^{+} := \\mathcal{E}_{\\mathcal{V}} \\cup \\{(y, y) \\mid y \\in \\mathcal{V}_{\\mathcal{V}}\\}" + }, + { + "bbox": [ + 313, + 337, + 550, + 352 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 362, + 554, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 362, + 554, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 362, + 554, + 434 + ], + "type": "text", + "content": "The key idea is that each edge " + }, + { + "bbox": [ + 313, + 362, + 554, + 434 + ], + "type": "inline_equation", + "content": "(v,\\bar{v})\\in \\mathcal{E}_{\\mathcal{P}_i}" + }, + { + "bbox": [ + 313, + 362, + 554, + 434 + ], + "type": "text", + "content": " in the product graph can be interpreted as a (potential) matching between an edge of " + }, + { + "bbox": [ + 313, + 362, + 554, + 434 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 313, + 362, + 554, + 434 + ], + "type": "text", + "content": " and an edge (or a vertex to account for stretching/compression) of shape " + }, + { + "bbox": [ + 313, + 362, + 554, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 362, + 554, + 434 + ], + "type": "text", + "content": ". As shown in [48], matching surface cycle " + }, + { + "bbox": [ + 313, + 362, + 554, + 434 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 313, + 362, + 554, + 434 + ], + "type": "text", + "content": " to shape " + }, + { + "bbox": [ + 313, + 362, + 554, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 362, + 554, + 434 + ], + "type": "text", + "content": " amounts to solving a cyclic shortest path problem in the product graph " + }, + { + "bbox": [ + 313, + 362, + 554, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 313, + 362, + 554, + 434 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 435, + 554, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 554, + 544 + ], + "type": "text", + "content": "However, for our setting of 3D-to-3D shape matching, the neighbourhood between pairs of surface cycles cannot be ensured when considering vanilla shortest path algorithms, since they would solve the " + }, + { + "bbox": [ + 313, + 435, + 554, + 544 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 435, + 554, + 544 + ], + "type": "text", + "content": " individual surface cycle matching subproblems independently. To tackle this, we consider the linear programming (LP) formalism of cyclic shortest path problems and add additional constraints to couple the individual product graphs " + }, + { + "bbox": [ + 313, + 435, + 554, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_1,\\ldots ,\\mathcal{P}_n" + }, + { + "bbox": [ + 313, + 435, + 554, + 544 + ], + "type": "text", + "content": ", as explained next." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 553, + 437, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 553, + 437, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 553, + 437, + 566 + ], + "type": "text", + "content": "4.2.2. Subproblem Coupling" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "We couple the individual surface cycle matching subproblems " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_1, \\ldots, \\mathcal{P}_n" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " by glueing them together at opposite edges. To this end, we introduce coupling constraints " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{\\mathcal{L}}" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " which serve the purpose of ensuring that matchings of opposite edges are consistent, i.e. resulting matchings (or rather resulting shortest cyclic paths of neighbouring surface cycles) must go through opposite edges, see Fig. 4 (iii). Defacto, this enforces the glueing of neighbouring product graphs, and thus ensures geometric consistency of neighbouring surface cycles. In addition to the coupling, we want to ensure injectivity for each surface cycle edge, i.e. each such edge is matched exactly once." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "text", + "content": "915" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 69, + 202, + 193 + ], + "blocks": [ + { + "bbox": [ + 86, + 69, + 202, + 193 + ], + "lines": [ + { + "bbox": [ + 86, + 69, + 202, + 193 + ], + "spans": [ + { + "bbox": [ + 86, + 69, + 202, + 193 + ], + "type": "image", + "image_path": "5abe4fc4747faf4a0e3de82d6ad423af0d73645707f5a72b3135d03496466f81.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 200, + 555, + 289 + ], + "lines": [ + { + "bbox": [ + 55, + 200, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 200, + 555, + 289 + ], + "type": "text", + "content": "Figure 4. Conceptual summary of our formalism. Left: Illustration of product graph " + }, + { + "bbox": [ + 55, + 200, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 55, + 200, + 555, + 289 + ], + "type": "text", + "content": " between the surface cycle " + }, + { + "bbox": [ + 55, + 200, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i" + }, + { + "bbox": [ + 55, + 200, + 555, + 289 + ], + "type": "text", + "content": " and shape graph " + }, + { + "bbox": [ + 55, + 200, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 55, + 200, + 555, + 289 + ], + "type": "text", + "content": ". Each black dot inside the big yellow circle represents a product vertex. Overall, the product graph can be thought of having one copy of " + }, + { + "bbox": [ + 55, + 200, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 55, + 200, + 555, + 289 + ], + "type": "text", + "content": " for each vertex of " + }, + { + "bbox": [ + 55, + 200, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i" + }, + { + "bbox": [ + 55, + 200, + 555, + 289 + ], + "type": "text", + "content": ", which are appropriately connected via edges (only a subset of edges are drawn to reduce visual clutter). Middle: Two neighbouring product graphs " + }, + { + "bbox": [ + 55, + 200, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 55, + 200, + 555, + 289 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 200, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_j" + }, + { + "bbox": [ + 55, + 200, + 555, + 289 + ], + "type": "text", + "content": " that arise from two neighbouring surface cycles " + }, + { + "bbox": [ + 55, + 200, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i" + }, + { + "bbox": [ + 55, + 200, + 555, + 289 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 200, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_j" + }, + { + "bbox": [ + 55, + 200, + 555, + 289 + ], + "type": "text", + "content": ". Right: Illustration of constrained neighbouring product graphs. We add additional coupling constraints so that opposite edges of the product graphs (i.e. edges with same colour, see coloured lines ——) are coupled via coupling constraints " + }, + { + "bbox": [ + 55, + 200, + 555, + 289 + ], + "type": "inline_equation", + "content": "q_{ij}^{e_1}, q_{ij}^{e_2}, q_{ij}^{e_3} \\in \\mathcal{V}_{\\mathcal{L}}" + }, + { + "bbox": [ + 55, + 200, + 555, + 289 + ], + "type": "text", + "content": " (see green dots " + }, + { + "bbox": [ + 55, + 200, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\bigcirc" + }, + { + "bbox": [ + 55, + 200, + 555, + 289 + ], + "type": "text", + "content": "). Furthermore, we bundle edges of the product graphs so that each edge of the surface cycles is matched exactly once. To this end, we add injectivity constraints " + }, + { + "bbox": [ + 55, + 200, + 555, + 289 + ], + "type": "inline_equation", + "content": "s_{x\\bar{x}}" + }, + { + "bbox": [ + 55, + 200, + 555, + 289 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 200, + 555, + 289 + ], + "type": "inline_equation", + "content": "s_{\\bar{x} x} \\in \\mathcal{V}_{\\mathcal{S}}" + }, + { + "bbox": [ + 55, + 200, + 555, + 289 + ], + "type": "text", + "content": " (see blue dots " + }, + { + "bbox": [ + 55, + 200, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\bigcirc" + }, + { + "bbox": [ + 55, + 200, + 555, + 289 + ], + "type": "text", + "content": " within bundles of edges)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 211, + 68, + 375, + 194 + ], + "blocks": [ + { + "bbox": [ + 211, + 68, + 375, + 194 + ], + "lines": [ + { + "bbox": [ + 211, + 68, + 375, + 194 + ], + "spans": [ + { + "bbox": [ + 211, + 68, + 375, + 194 + ], + "type": "image", + "image_path": "7a4898c7890f4c8b9df7ba737963678f48c438f1634aed4a3f583595cb63b303.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 378, + 68, + 541, + 194 + ], + "blocks": [ + { + "bbox": [ + 378, + 68, + 541, + 194 + ], + "lines": [ + { + "bbox": [ + 378, + 68, + 541, + 194 + ], + "spans": [ + { + "bbox": [ + 378, + 68, + 541, + 194 + ], + "type": "image", + "image_path": "9dd4d70409dd43ce884b230e381a50e9f4d825b3d49f2a698d96a2483e8b94ef.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 309, + 232, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 309, + 232, + 322 + ], + "spans": [ + { + "bbox": [ + 55, + 309, + 232, + 322 + ], + "type": "text", + "content": "4.2.3. Surface Cycle Matching Injectivity" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 324, + 296, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 296, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 296, + 433 + ], + "type": "text", + "content": "To enforce that each surface cycle edge is matched exactly once, i.e. it is matched to exactly one edge (or vertex) of " + }, + { + "bbox": [ + 55, + 324, + 296, + 433 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 55, + 324, + 296, + 433 + ], + "type": "text", + "content": ", we introduce injectivity constraints " + }, + { + "bbox": [ + 55, + 324, + 296, + 433 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{\\mathcal{S}}" + }, + { + "bbox": [ + 55, + 324, + 296, + 433 + ], + "type": "text", + "content": ". For that, let us denote the set of all edges (in the product graph) that are potential matching candidates of the edge " + }, + { + "bbox": [ + 55, + 324, + 296, + 433 + ], + "type": "inline_equation", + "content": "(x,\\bar{x})\\in \\mathcal{E}_{\\mathcal{C}_i}" + }, + { + "bbox": [ + 55, + 324, + 296, + 433 + ], + "type": "text", + "content": " as the edge bundle of " + }, + { + "bbox": [ + 55, + 324, + 296, + 433 + ], + "type": "inline_equation", + "content": "(x,\\bar{x})" + }, + { + "bbox": [ + 55, + 324, + 296, + 433 + ], + "type": "text", + "content": ". For each such edge bundle, we introduce one injectivity constraint (see Fig. 4 (iii)), which has the purpose to ensure that only a single edge of each edge bundle is part of the final matching." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 441, + 173, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 173, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 173, + 453 + ], + "type": "text", + "content": "4.3. Constraint Matrices" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 458, + 296, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 458, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 458, + 296, + 559 + ], + "type": "text", + "content": "As mentioned, we use the LP formalism of the independent cyclic shortest path problems and add additional constraints for coupling an matching injectivity. To this end, we consider the constraint matrix " + }, + { + "bbox": [ + 55, + 458, + 296, + 559 + ], + "type": "inline_equation", + "content": "H = \\left[ \\begin{array}{c} P \\\\ L \\\\ S \\end{array} \\right]" + }, + { + "bbox": [ + 55, + 458, + 296, + 559 + ], + "type": "text", + "content": " consisting of submatrices " + }, + { + "bbox": [ + 55, + 458, + 296, + 559 + ], + "type": "inline_equation", + "content": "P, L" + }, + { + "bbox": [ + 55, + 458, + 296, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 458, + 296, + 559 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 458, + 296, + 559 + ], + "type": "text", + "content": ". Here, submatrix " + }, + { + "bbox": [ + 55, + 458, + 296, + 559 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 458, + 296, + 559 + ], + "type": "text", + "content": " represents the collection of incidence matrices of individual product graphs, submatrix " + }, + { + "bbox": [ + 55, + 458, + 296, + 559 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 458, + 296, + 559 + ], + "type": "text", + "content": " the couplings, and submatrix " + }, + { + "bbox": [ + 55, + 458, + 296, + 559 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 458, + 296, + 559 + ], + "type": "text", + "content": " the injectivity components." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 559, + 296, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 296, + 691 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 296, + 691 + ], + "type": "text", + "content": "The matrix " + }, + { + "bbox": [ + 55, + 559, + 296, + 691 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 559, + 296, + 691 + ], + "type": "text", + "content": ". The vertex edge incidence matrix " + }, + { + "bbox": [ + 55, + 559, + 296, + 691 + ], + "type": "inline_equation", + "content": "P \\in \\{-1,0,1\\}^{c \\times m}" + }, + { + "bbox": [ + 55, + 559, + 296, + 691 + ], + "type": "text", + "content": " (with " + }, + { + "bbox": [ + 55, + 559, + 296, + 691 + ], + "type": "inline_equation", + "content": "m = |\\mathcal{E}_{\\mathcal{H}}|" + }, + { + "bbox": [ + 55, + 559, + 296, + 691 + ], + "type": "text", + "content": " columns, i.e. one column for each product edge, and " + }, + { + "bbox": [ + 55, + 559, + 296, + 691 + ], + "type": "inline_equation", + "content": "c = \\sum_{i=1}^{n} |\\mathcal{V}_{\\mathcal{C}_i}| |\\mathcal{V}_{\\mathcal{Y}}|" + }, + { + "bbox": [ + 55, + 559, + 296, + 691 + ], + "type": "text", + "content": " many rows) represents all " + }, + { + "bbox": [ + 55, + 559, + 296, + 691 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 559, + 296, + 691 + ], + "type": "text", + "content": " product graphs and is defined as " + }, + { + "bbox": [ + 55, + 559, + 296, + 691 + ], + "type": "inline_equation", + "content": "P := \\mathrm{diag}(P_1, \\ldots, P_n)" + }, + { + "bbox": [ + 55, + 559, + 296, + 691 + ], + "type": "text", + "content": ", i.e. it contains the vertex edge incidence matrices " + }, + { + "bbox": [ + 55, + 559, + 296, + 691 + ], + "type": "inline_equation", + "content": "P_i" + }, + { + "bbox": [ + 55, + 559, + 296, + 691 + ], + "type": "text", + "content": " of individual product graphs " + }, + { + "bbox": [ + 55, + 559, + 296, + 691 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 55, + 559, + 296, + 691 + ], + "type": "text", + "content": " as diagonal blocks. To illustrate the structure of the matrix representation of one " + }, + { + "bbox": [ + 55, + 559, + 296, + 691 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 55, + 559, + 296, + 691 + ], + "type": "text", + "content": ", we now discuss the simplified case of non-degenerate matchings (i.e. we do not allow edge to vertex matchings, see Sec. A in the supplementary for more details). In this case, the incidence matrix " + }, + { + "bbox": [ + 55, + 559, + 296, + 691 + ], + "type": "inline_equation", + "content": "P_i" + }, + { + "bbox": [ + 55, + 559, + 296, + 691 + ], + "type": "text", + "content": " reads" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 112, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 112, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "P _ {i} := C _ {i} ^ {+} \\otimes Y ^ {+} - C _ {i} ^ {-} \\otimes Y ^ {-}, \\tag {2}", + "image_path": "21f68346915514442151641588d60d0fb0b9a1dad6398e62c9c785bb39a076d9.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 309, + 555, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 309, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 309, + 555, + 525 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 313, + 309, + 555, + 525 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 313, + 309, + 555, + 525 + ], + "type": "text", + "content": " being the Kroenecker product and with " + }, + { + "bbox": [ + 313, + 309, + 555, + 525 + ], + "type": "inline_equation", + "content": "C_i \\in \\{-1,0,1\\}^{|\\mathcal{V}_{c_i}| \\times |\\mathcal{E}_{c_i}^{\\prime}|\\}" + }, + { + "bbox": [ + 313, + 309, + 555, + 525 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 309, + 555, + 525 + ], + "type": "inline_equation", + "content": "Y \\in \\{-1,0,1\\}^{|\\mathcal{V}_{\\mathcal{Y}}| \\times |\\mathcal{E}_{\\mathcal{Y}}^{\\prime}|\\}" + }, + { + "bbox": [ + 313, + 309, + 555, + 525 + ], + "type": "text", + "content": " being the incidence matrices of surface cycle " + }, + { + "bbox": [ + 313, + 309, + 555, + 525 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 313, + 309, + 555, + 525 + ], + "type": "text", + "content": " and shape graph " + }, + { + "bbox": [ + 313, + 309, + 555, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\mathcal{Y}}" + }, + { + "bbox": [ + 313, + 309, + 555, + 525 + ], + "type": "text", + "content": " respectively (we note that " + }, + { + "bbox": [ + 313, + 309, + 555, + 525 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 313, + 309, + 555, + 525 + ], + "type": "text", + "content": " has exactly two non-zero of opposite sign in each row and column since it is the incidence matrix of a cycle). We use the notation " + }, + { + "bbox": [ + 313, + 309, + 555, + 525 + ], + "type": "inline_equation", + "content": "+" + }, + { + "bbox": [ + 313, + 309, + 555, + 525 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 309, + 555, + 525 + ], + "type": "inline_equation", + "content": "-" + }, + { + "bbox": [ + 313, + 309, + 555, + 525 + ], + "type": "text", + "content": " to split the incidence matrix into incoming and outgoing incidence matrix, respectively. In other words, " + }, + { + "bbox": [ + 313, + 309, + 555, + 525 + ], + "type": "inline_equation", + "content": "C_i^+" + }, + { + "bbox": [ + 313, + 309, + 555, + 525 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 309, + 555, + 525 + ], + "type": "inline_equation", + "content": "C_i^-" + }, + { + "bbox": [ + 313, + 309, + 555, + 525 + ], + "type": "text", + "content": " as well as " + }, + { + "bbox": [ + 313, + 309, + 555, + 525 + ], + "type": "inline_equation", + "content": "Y^+" + }, + { + "bbox": [ + 313, + 309, + 555, + 525 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 309, + 555, + 525 + ], + "type": "inline_equation", + "content": "Y^-" + }, + { + "bbox": [ + 313, + 309, + 555, + 525 + ], + "type": "text", + "content": " are the non-negative and non-positive entries of the incidence matrices " + }, + { + "bbox": [ + 313, + 309, + 555, + 525 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 313, + 309, + 555, + 525 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 309, + 555, + 525 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 313, + 309, + 555, + 525 + ], + "type": "text", + "content": ", respectively. We note that the sign splitting of incidence matrices " + }, + { + "bbox": [ + 313, + 309, + 555, + 525 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 313, + 309, + 555, + 525 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 309, + 555, + 525 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 313, + 309, + 555, + 525 + ], + "type": "text", + "content": " is necessary to account for proper edge directions in the resulting product graph. By definition, their edge directions induce edge directions of product edges of respective product graph " + }, + { + "bbox": [ + 313, + 309, + 555, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 313, + 309, + 555, + 525 + ], + "type": "text", + "content": ". Each " + }, + { + "bbox": [ + 313, + 309, + 555, + 525 + ], + "type": "inline_equation", + "content": "P_i" + }, + { + "bbox": [ + 313, + 309, + 555, + 525 + ], + "type": "text", + "content": " has a block structure with " + }, + { + "bbox": [ + 313, + 309, + 555, + 525 + ], + "type": "inline_equation", + "content": "|\\mathcal{V}_{\\mathcal{Y}}| \\times |\\mathcal{E}_{\\mathcal{Y}}|" + }, + { + "bbox": [ + 313, + 309, + 555, + 525 + ], + "type": "text", + "content": " sized blocks and each block contains either only " + }, + { + "bbox": [ + 313, + 309, + 555, + 525 + ], + "type": "inline_equation", + "content": "+1" + }, + { + "bbox": [ + 313, + 309, + 555, + 525 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 309, + 555, + 525 + ], + "type": "inline_equation", + "content": "-1" + }, + { + "bbox": [ + 313, + 309, + 555, + 525 + ], + "type": "text", + "content": " entries. Further, each block contains exactly one non-zero per column, see Fig. A.1 in the supplementary." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 525, + 554, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 554, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 554, + 548 + ], + "type": "text", + "content": "The matrix " + }, + { + "bbox": [ + 313, + 525, + 554, + 548 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 525, + 554, + 548 + ], + "type": "text", + "content": ". The coupling of opposite edges of neighbouring product graphs is captured in the matrix" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 399, + 557, + 553, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 557, + 553, + 574 + ], + "spans": [ + { + "bbox": [ + 399, + 557, + 553, + 574 + ], + "type": "interline_equation", + "content": "L := K \\otimes I _ {\\left| \\varepsilon_ {y} ^ {+} \\right|}. \\tag {3}", + "image_path": "3650a339733a9ba52c9e49f105932fd082cb244dd3159ce9d01811d6b4f67b39.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 581, + 555, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 555, + 701 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 555, + 701 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 313, + 581, + 555, + 701 + ], + "type": "inline_equation", + "content": "K \\in \\{-1,0,1\\}^{p \\times |\\mathcal{E}_{\\mathcal{X}}|}" + }, + { + "bbox": [ + 313, + 581, + 555, + 701 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 581, + 555, + 701 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 581, + 555, + 701 + ], + "type": "text", + "content": " is the number of undirected non-boundary edges of " + }, + { + "bbox": [ + 313, + 581, + 555, + 701 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 313, + 581, + 555, + 701 + ], + "type": "text", + "content": ") is an incidence matrix which represents the incidence of opposite edges across all " + }, + { + "bbox": [ + 313, + 581, + 555, + 701 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 581, + 555, + 701 + ], + "type": "text", + "content": " surface cycles. " + }, + { + "bbox": [ + 313, + 581, + 555, + 701 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 581, + 555, + 701 + ], + "type": "text", + "content": " has at most one non-zero element per column and exactly two non-zeros per row (with opposite sign) and thus " + }, + { + "bbox": [ + 313, + 581, + 555, + 701 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 581, + 555, + 701 + ], + "type": "text", + "content": " has identical block structure with non-negative and non-positive blocks respectively. We note that the definition of " + }, + { + "bbox": [ + 313, + 581, + 555, + 701 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 581, + 555, + 701 + ], + "type": "text", + "content": " requires columns of the non-positive blocks of " + }, + { + "bbox": [ + 313, + 581, + 555, + 701 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 581, + 555, + 701 + ], + "type": "text", + "content": " to be permuted (so that it maps opposite edges of neighbouring product graphs), see Sec. A in the supplementary." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "content": "The matrix " + }, + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "content": ". The bundling of edges of all " + }, + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "content": " product" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "916" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 70, + 272, + 150 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 272, + 150 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 272, + 150 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 272, + 150 + ], + "type": "image", + "image_path": "4acadaf178973b318df2a9ad57d35170319dfecec43d7f341715c10441167fc2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 158, + 296, + 227 + ], + "lines": [ + { + "bbox": [ + 55, + 158, + 296, + 227 + ], + "spans": [ + { + "bbox": [ + 55, + 158, + 296, + 227 + ], + "type": "text", + "content": "Figure 5. Visualisation of edge cost " + }, + { + "bbox": [ + 55, + 158, + 296, + 227 + ], + "type": "inline_equation", + "content": "c((v, \\bar{v}))" + }, + { + "bbox": [ + 55, + 158, + 296, + 227 + ], + "type": "text", + "content": " computation for the product edge " + }, + { + "bbox": [ + 55, + 158, + 296, + 227 + ], + "type": "inline_equation", + "content": "(v, \\bar{v})" + }, + { + "bbox": [ + 55, + 158, + 296, + 227 + ], + "type": "text", + "content": ". The product edge defines a potential matching between edges " + }, + { + "bbox": [ + 55, + 158, + 296, + 227 + ], + "type": "inline_equation", + "content": "(x, \\bar{x}) \\in \\mathcal{E}_{\\mathcal{X}}" + }, + { + "bbox": [ + 55, + 158, + 296, + 227 + ], + "type": "text", + "content": " of shape " + }, + { + "bbox": [ + 55, + 158, + 296, + 227 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 55, + 158, + 296, + 227 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 158, + 296, + 227 + ], + "type": "inline_equation", + "content": "(y, \\bar{y}) \\in \\mathcal{E}_{\\mathcal{Y}}^{+}" + }, + { + "bbox": [ + 55, + 158, + 296, + 227 + ], + "type": "text", + "content": " of shape " + }, + { + "bbox": [ + 55, + 158, + 296, + 227 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}" + }, + { + "bbox": [ + 55, + 158, + 296, + 227 + ], + "type": "text", + "content": ". Thus, we define the matching cost " + }, + { + "bbox": [ + 55, + 158, + 296, + 227 + ], + "type": "inline_equation", + "content": "c((v, \\bar{v}))" + }, + { + "bbox": [ + 55, + 158, + 296, + 227 + ], + "type": "text", + "content": " as feature difference between features " + }, + { + "bbox": [ + 55, + 158, + 296, + 227 + ], + "type": "inline_equation", + "content": "\\| f_x - f_y \\|" + }, + { + "bbox": [ + 55, + 158, + 296, + 227 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 158, + 296, + 227 + ], + "type": "inline_equation", + "content": "\\| f_{\\bar{x}} - f_{\\bar{y}} \\|" + }, + { + "bbox": [ + 55, + 158, + 296, + 227 + ], + "type": "text", + "content": " at respective source vertices " + }, + { + "bbox": [ + 55, + 158, + 296, + 227 + ], + "type": "inline_equation", + "content": "x, y" + }, + { + "bbox": [ + 55, + 158, + 296, + 227 + ], + "type": "text", + "content": " and target vertices " + }, + { + "bbox": [ + 55, + 158, + 296, + 227 + ], + "type": "inline_equation", + "content": "\\bar{x}, \\bar{y}" + }, + { + "bbox": [ + 55, + 158, + 296, + 227 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 246, + 222, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 246, + 222, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 246, + 222, + 258 + ], + "type": "text", + "content": "graphs can be described using the matrix" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 265, + 295, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 265, + 295, + 285 + ], + "spans": [ + { + "bbox": [ + 134, + 265, + 295, + 285 + ], + "type": "interline_equation", + "content": "S := \\boldsymbol {I} _ {\\left| \\mathcal {E} _ {x} \\right|} \\otimes 1 _ {\\left| \\mathcal {E} _ {y} ^ {+} \\right|} ^ {T}. \\tag {4}", + "image_path": "acf87c9df402d95227b7a99b4bdaa1b773f2b8177be20c5e15194b15f4aca6cd.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 293, + 296, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 296, + 357 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 293, + 296, + 357 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 293, + 296, + 357 + ], + "type": "text", + "content": " has block diagonal structure consisting of " + }, + { + "bbox": [ + 55, + 293, + 296, + 357 + ], + "type": "inline_equation", + "content": "|\\mathcal{E}_{\\mathcal{X}}|" + }, + { + "bbox": [ + 55, + 293, + 296, + 357 + ], + "type": "text", + "content": " many non-negative blocks (i.e. " + }, + { + "bbox": [ + 55, + 293, + 296, + 357 + ], + "type": "inline_equation", + "content": "1_{|\\mathcal{E}_{\\mathcal{Y}}^{+}|}^{T}" + }, + { + "bbox": [ + 55, + 293, + 296, + 357 + ], + "type": "text", + "content": " blocks). We note that " + }, + { + "bbox": [ + 55, + 293, + 296, + 357 + ], + "type": "inline_equation", + "content": "Sx = 1" + }, + { + "bbox": [ + 55, + 293, + 296, + 357 + ], + "type": "text", + "content": " enforces matching injectivity by enumerating product edges such that all " + }, + { + "bbox": [ + 55, + 293, + 296, + 357 + ], + "type": "inline_equation", + "content": "|\\mathcal{E}_{\\mathcal{Y}}^{+}|" + }, + { + "bbox": [ + 55, + 293, + 296, + 357 + ], + "type": "text", + "content": "-many matching candidates of a single surface cycle are grouped." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 363, + 239, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 363, + 239, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 239, + 376 + ], + "type": "text", + "content": "4.4. Resulting Integer Linear Program" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 380, + 296, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 296, + 429 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 296, + 429 + ], + "type": "text", + "content": "We use the previously described constraint matrix " + }, + { + "bbox": [ + 55, + 380, + 296, + 429 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 380, + 296, + 429 + ], + "type": "text", + "content": " to cast geometrically consistent shape matching as a linear program which essentially represents " + }, + { + "bbox": [ + 55, + 380, + 296, + 429 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 380, + 296, + 429 + ], + "type": "text", + "content": " many coupled cyclic shortest path problems." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 429, + 296, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 429, + 296, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 429, + 296, + 489 + ], + "type": "text", + "content": "Matching costs. The matching cost " + }, + { + "bbox": [ + 55, + 429, + 296, + 489 + ], + "type": "inline_equation", + "content": "c(e)" + }, + { + "bbox": [ + 55, + 429, + 296, + 489 + ], + "type": "text", + "content": " of product edge " + }, + { + "bbox": [ + 55, + 429, + 296, + 489 + ], + "type": "inline_equation", + "content": "e = \\left(\\binom{x}{y},\\binom{\\bar{x}}{\\bar{y}}\\right)" + }, + { + "bbox": [ + 55, + 429, + 296, + 489 + ], + "type": "text", + "content": " measures how well the edge " + }, + { + "bbox": [ + 55, + 429, + 296, + 489 + ], + "type": "inline_equation", + "content": "(x,\\bar{x})\\in \\mathcal{E}_{\\mathcal{X}}" + }, + { + "bbox": [ + 55, + 429, + 296, + 489 + ], + "type": "text", + "content": " of shape " + }, + { + "bbox": [ + 55, + 429, + 296, + 489 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 55, + 429, + 296, + 489 + ], + "type": "text", + "content": " and the edge " + }, + { + "bbox": [ + 55, + 429, + 296, + 489 + ], + "type": "inline_equation", + "content": "(y,\\bar{y})\\in \\mathcal{E}_{\\mathcal{Y}}^{+}" + }, + { + "bbox": [ + 55, + 429, + 296, + 489 + ], + "type": "text", + "content": " of shape " + }, + { + "bbox": [ + 55, + 429, + 296, + 489 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 55, + 429, + 296, + 489 + ], + "type": "text", + "content": " fit together, e.g. by comparing geometric properties or feature descriptors, see also Fig. 5." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 490, + 296, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 490, + 296, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 490, + 296, + 550 + ], + "type": "text", + "content": "Optimisation problem. We collect matching costs of all " + }, + { + "bbox": [ + 55, + 490, + 296, + 550 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 490, + 296, + 550 + ], + "type": "text", + "content": " edges of the " + }, + { + "bbox": [ + 55, + 490, + 296, + 550 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 490, + 296, + 550 + ], + "type": "text", + "content": " product graphs in the vector " + }, + { + "bbox": [ + 55, + 490, + 296, + 550 + ], + "type": "inline_equation", + "content": "c \\in \\mathbb{R}_+^m" + }, + { + "bbox": [ + 55, + 490, + 296, + 550 + ], + "type": "text", + "content": ". With that, we can use an indicator representation " + }, + { + "bbox": [ + 55, + 490, + 296, + 550 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 490, + 296, + 550 + ], + "type": "text", + "content": " for each edge (" + }, + { + "bbox": [ + 55, + 490, + 296, + 550 + ], + "type": "inline_equation", + "content": "x_k = 1" + }, + { + "bbox": [ + 55, + 490, + 296, + 550 + ], + "type": "text", + "content": " means that " + }, + { + "bbox": [ + 55, + 490, + 296, + 550 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 490, + 296, + 550 + ], + "type": "text", + "content": "-th edge is part of the final matching) so that our matching formalism reads" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 93, + 557, + 295, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 557, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 93, + 557, + 295, + 578 + ], + "type": "interline_equation", + "content": "\\min _ {x \\in \\{0, 1 \\} ^ {m}} c ^ {T} x \\quad \\text {s . t .} \\quad H x = b. \\qquad \\mathrm {(G e C o 3 D)}", + "image_path": "36f3b29460b0a82d696b72bbd75c20323b412ee2571e855064f2f9c51245bce9.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 586, + 296, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 296, + 671 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 296, + 671 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 586, + 296, + 671 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 586, + 296, + 671 + ], + "type": "text", + "content": " is the previously introduced constraint matrix and " + }, + { + "bbox": [ + 55, + 586, + 296, + 671 + ], + "type": "inline_equation", + "content": "b \\coloneqq \\left[0_{|\\mathcal{V}_{\\mathcal{H}}| - |\\mathcal{V}_{S}|}; 1_{|\\mathcal{V}_{S}|}\\right]" + }, + { + "bbox": [ + 55, + 586, + 296, + 671 + ], + "type": "text", + "content": " is a column vector with all zeros except for ones in rows belonging to rows of submatrix " + }, + { + "bbox": [ + 55, + 586, + 296, + 671 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 586, + 296, + 671 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 586, + 296, + 671 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 586, + 296, + 671 + ], + "type": "text", + "content": ". We solve (GeCo3D) by considering its LP-relaxation, i.e. we replace " + }, + { + "bbox": [ + 55, + 586, + 296, + 671 + ], + "type": "inline_equation", + "content": "x \\in \\{0,1\\}^m" + }, + { + "bbox": [ + 55, + 586, + 296, + 671 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 586, + 296, + 671 + ], + "type": "inline_equation", + "content": "x \\in [0,1]^m" + }, + { + "bbox": [ + 55, + 586, + 296, + 671 + ], + "type": "text", + "content": ". We empirically observe in our experiments, that all solutions are integral and consequently globally optimal." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "Lemma 8. Matchings between shapes " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " obtained by solving (GeCo3D) are globally geometrically consistent according to Definition 4." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 555, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 157 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 157 + ], + "type": "text", + "content": "Proof. We represent each triangle of the source shape " + }, + { + "bbox": [ + 313, + 72, + 555, + 157 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 313, + 72, + 555, + 157 + ], + "type": "text", + "content": " using a surface cycle. Furthermore, matchings obtained by solving (GeCo3D) preserve neighbourhood relations along the path of a matched surface cycle as well as between the surface cycles. Overall, this ensures that connected vertices on " + }, + { + "bbox": [ + 313, + 72, + 555, + 157 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 313, + 72, + 555, + 157 + ], + "type": "text", + "content": " are only matched to connected vertices (or the same vertex) on shape " + }, + { + "bbox": [ + 313, + 72, + 555, + 157 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}" + }, + { + "bbox": [ + 313, + 72, + 555, + 157 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 169, + 511, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 169, + 511, + 182 + ], + "spans": [ + { + "bbox": [ + 313, + 169, + 511, + 182 + ], + "type": "text", + "content": "4.5. Hyper Product Graph Interpretation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 187, + 555, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 187, + 555, + 283 + ], + "spans": [ + { + "bbox": [ + 313, + 187, + 555, + 283 + ], + "type": "text", + "content": "Our optimisation problem (GeCo3D) can be interpreted as finding a minimum-cost flow circulation in a directed hyper product graph " + }, + { + "bbox": [ + 313, + 187, + 555, + 283 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 187, + 555, + 283 + ], + "type": "text", + "content": ", see [6] for definitions of flows in directed hyper graphs. Here, " + }, + { + "bbox": [ + 313, + 187, + 555, + 283 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 187, + 555, + 283 + ], + "type": "text", + "content": " can be obtained by interpreting " + }, + { + "bbox": [ + 313, + 187, + 555, + 283 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 187, + 555, + 283 + ], + "type": "text", + "content": " as a vertex hyper edge incidence matrix, i.e. by interpreting rows of " + }, + { + "bbox": [ + 313, + 187, + 555, + 283 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 187, + 555, + 283 + ], + "type": "text", + "content": " as vertices and columns of " + }, + { + "bbox": [ + 313, + 187, + 555, + 283 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 187, + 555, + 283 + ], + "type": "text", + "content": " as directed hyper edges, see Fig. 6 for an illustration. For a formal definition of " + }, + { + "bbox": [ + 313, + 187, + 555, + 283 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 187, + 555, + 283 + ], + "type": "text", + "content": " we refer to Sec. B in the supplementary." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 284, + 556, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 284, + 556, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 284, + 556, + 392 + ], + "type": "text", + "content": "In general, such flow circulation problems on hypergraphs are NP-hard [5]. However, there are certain subclasses that are known to be solvable in polynomial time [10, 29, 42, 80, 84]. Unfortunately, the specific structure of our hyper graph is not listed among the known polynomial-time solvable subclasses [10, 29, 42, 80, 84]. Nevertheless, we empirically observe that all instances considered in our experiments yield an integral (and thus globally optimal) solution when solved using LP-relaxations." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 404, + 395, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 404, + 395, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 404, + 395, + 418 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 426, + 555, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 474 + ], + "type": "text", + "content": "In this section, we experimentally evaluate our method's performance for 3D shape matching. Further, we show as proof of concept that our method is applicable to planar graph matching." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 474, + 556, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 556, + 571 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 556, + 571 + ], + "type": "text", + "content": "Setup. We conduct runtime experiments on an Intel Core i9 12900K with 128 GB DDR5 RAM. For solving our linear programm (GeCo3D), we use off-the-shelf solver Gurobi [37] (version 10). We explicitly resolve coupling constraints during problem construction to obtain smaller constraint matrices, see Sec. C.2 in the supplementary. To better account for shrinking and stretching, we integrate a distortion bound, see Sec. C.1 in the supplementary. For" + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 341, + 586, + 533, + 658 + ], + "blocks": [ + { + "bbox": [ + 341, + 586, + 533, + 658 + ], + "lines": [ + { + "bbox": [ + 341, + 586, + 533, + 658 + ], + "spans": [ + { + "bbox": [ + 341, + 586, + 533, + 658 + ], + "type": "image", + "image_path": "451545cb9391a4aeb472c1cf66df41e21d7605e9f77eab77a8b3f891ca91a69d.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 666, + 556, + 711 + ], + "lines": [ + { + "bbox": [ + 313, + 666, + 556, + 711 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 556, + 711 + ], + "type": "text", + "content": "Figure 6. Illustration of interpreting our constraint matrix " + }, + { + "bbox": [ + 313, + 666, + 556, + 711 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 666, + 556, + 711 + ], + "type": "text", + "content": " as the vertex hyper edge incidence matrix of a hyper product graph " + }, + { + "bbox": [ + 313, + 666, + 556, + 711 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 666, + 556, + 711 + ], + "type": "text", + "content": ", see Sec. B in the supplementary. To obtain " + }, + { + "bbox": [ + 313, + 666, + 556, + 711 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 666, + 556, + 711 + ], + "type": "text", + "content": ", rows of " + }, + { + "bbox": [ + 313, + 666, + 556, + 711 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 666, + 556, + 711 + ], + "type": "text", + "content": " are interpreted as vertices, and columns as directed hyper edges of " + }, + { + "bbox": [ + 313, + 666, + 556, + 711 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 666, + 556, + 711 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "text", + "content": "917" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "quantitative comparison, we decimate shapes to 1000 triangles using algorithms provided in [41]. We use the deep features [13] for all methods except for SmoothShells [28] and DiscrOpt [62] for which we use the original (axiomatic) features as reported in respective papers." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 134, + 294, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 134, + 294, + 218 + ], + "spans": [ + { + "bbox": [ + 55, + 134, + 294, + 218 + ], + "type": "text", + "content": "Metrics. We evaluate matching accuracy using geodesic errors, i.e. geodesic distance to ground-truth matchings. We follow the Princeton protocol [43] and normalise errors by the square-root of the shape area (see [43, Sec. 8.2]). Furthermore, we evaluate geometric consistency of matchings using Dirichlet energies, i.e. the deformation energy induced by the matching (see [67, Sec. 8])." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 220, + 296, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 296, + 616 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 296, + 616 + ], + "type": "text", + "content": "Methods. We compare the matching quality of the methods ULRSSM [13], SmoothShells [28], DiscrOpt [62] and SpiderMatch [67]. ULRSSM [13] is an unsupervised deep shape matching method achieving best matchings results on numerous benchmarks. Because of its good performance and its extensive evaluation, we consider it to be representative for deep shape matching methods. SmoothShells [28] is a functional map based shape alignment method which iteratively adds geometric information to compute matchings. DiscrOpt [62] is a functional map based method which relates functional maps to point-wise maps. SpiderMatch [67] is the only other method that is scalable and at the same time considers geometric consistency. Furthermore, for runtime evaluations, we only compare to other geometrically consistent methods since these aim to solve a much harder to solve problem (yet most of these methods do not scale well to relevant shape resolutions, see also Fig. 1 right). In this sense, we consider Windheuser et al. [89], SM-Comb [69], DiscomMatch [66] and SpiderMatch (geo.). Windheuser et al. [89] propose a geometrically consistent 3D shape matching formalism which they solve using LP-relaxations and iterative variable fixations. The methods SM-Comb [69] and DiscomMatch [66] aim to solve this formalism approximately on CPU and GPU, respectively. We use SpiderMatch (geo.) to indicate the SpiderMatch [67] method when using a different curve to represent the source shape: we consider a curve that covers all edges of the source shape and thus, contains intersections at every vertex of the source shape. This is necessary to guarantee global geometric consistency since SpiderMatch [67] can only ensure geometric consistency (according to Definition 4) at intersection points of the curve. Thus, SpiderMatch (geo.) yields a stronger notion of geometric consistency compared to SpiderMatch [67]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": "Datasets. We evaluate shape matching on four different datasets: remeshed FAUST [9, 21, 63] (100 near-isometric deformed human shapes from which we sample 100 test set pairs), SMAL [92] (49 non-isometric deformed animal shapes of eight species from which we sample 100 test set pairs), DT4D-H [57] (9 different classes of humanoid/game character shapes in different poses taken from DeformingThings4D [53] from which we sample 100 intra class test" + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 316, + 71, + 549, + 135 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 549, + 135 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 549, + 135 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 549, + 135 + ], + "type": "table", + "html": "
MethodFAUSTSMALDT4D IntraDT4D InterBeCoS
ULRSSM [13]0.0310.0480.0330.0410.057
SmoothShells [28]0.3790.3760.3730.4200.370
DiscrOpt [62]0.1100.2680.0750.1700.270
SpiderMatch [67]0.0290.0440.0270.0410.057
Ours0.0270.0440.0240.0390.056
", + "image_path": "ef86d4f8b8007f5d1bfb3f184cc29d9cb0b726be1008d25f54d86bef6ea63963.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 140, + 553, + 173 + ], + "lines": [ + { + "bbox": [ + 313, + 140, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 313, + 140, + 553, + 173 + ], + "type": "text", + "content": "Table 3. Comparison of mean geodesic errors (↓) of various shape matching methods. We can see that our method consistently outperforms other methods on all five datasets." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 201, + 553, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 201, + 553, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 553, + 236 + ], + "type": "text", + "content": "set pairs and 100 inter class test set pairs) and BeCoS [25] (2543 animal and humanoid shapes from various datasets of which we consider the 141 full-to-full test set pairs)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 253, + 430, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 430, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 430, + 266 + ], + "type": "text", + "content": "5.1.3D Shape Matching" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 274, + 553, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 553, + 298 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 553, + 298 + ], + "type": "text", + "content": "In the following, we evaluate our method's shape matching performance w.r.t. runtime and matching quality." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 300, + 555, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 555, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 555, + 552 + ], + "type": "text", + "content": "Runtime. In Fig. 1 right, we show runtime comparison to formalisms which consider geometric consistency. Curves show median runtimes over five instances of FAUST dataset. Among all methods, ours and SpiderMatch [67] scale best. They can handle 3D shapes with approximately twice as many vertices as their fastest competitors DiscoMatch [66], which tackles the problem proposed by Windheuser et al. [89]. However, we note that when considering SpiderMatch (geo.) and consequently curves with more intersections (which is necessary for global geometric consistency) the scalability of SpiderMatch (geo.) degrades drastically. This stems from the branch and bound algorithm (having exponential worst-case runtime in general) which is used to solve SpiderMatch [67]. In contrast, we empirically observe (in all of our tested instances) that the linear programming relaxation of (GeCo3D) is always tight (i.e. yields an integral optimal solution) which explains the scalability of our approach. We emphasise that matchings computed from (GeCo3D) are provably geometrically consistent according to Definition 4, while the matchings of SpiderMatch [67] are not (cf. also Tab. 4)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 555, + 553, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 553, + 674 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 553, + 674 + ], + "type": "text", + "content": "Full Shape Matching. In Tab. 3 and Tab. 4, we quantitatively compare the shape matching performance of various methods. Our method consistently produces best results across all datasets w.r.t. mean geodesic errors and smoothest results w.r.t Dirichlet energies. This showcases the beneficial effect of strong priors induced by geometric consistency. In addition, in Fig. 7, we show qualitative results, which further showcases high-quality matching results computed with our method. We provide more results as well as ablation studies in Sec. D.1 in the supplementary." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Partial-to-Full Shape Matching. As proof of concept we show qualitative results of matchings computed with our method in the partial-to-full setting, see Fig. 8." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "918" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 292, + 135 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 292, + 135 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 292, + 135 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 292, + 135 + ], + "type": "table", + "html": "
MethodFAUSTSMALDT4D IntraDT4D InterBeCoS
ULRSSM [13]2.12.63.42.64.1
SmoothShells [28]1.92.81.62.35.7
DiscrOpt [62]8.913.67.510.71.1
SpiderMatch [67]1.71.81.72.10.81
Ours0.460.530.480.620.77
", + "image_path": "a60528368ca34bf380f251a12f7dfcad670f6d748799a9684fd87e443cd610f1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 143, + 295, + 187 + ], + "lines": [ + { + "bbox": [ + 55, + 143, + 295, + 187 + ], + "spans": [ + { + "bbox": [ + 55, + 143, + 295, + 187 + ], + "type": "text", + "content": "Table 4. Comparison of mean Dirichlet energies " + }, + { + "bbox": [ + 55, + 143, + 295, + 187 + ], + "type": "inline_equation", + "content": "(\\downarrow)" + }, + { + "bbox": [ + 55, + 143, + 295, + 187 + ], + "type": "text", + "content": " of various shape matching methods to measure smoothness of the matching. Our method consistently yields best results, which shows the effect of strong priors induced by global geometric consistency." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 55, + 191, + 290, + 417 + ], + "blocks": [ + { + "bbox": [ + 55, + 191, + 290, + 417 + ], + "lines": [ + { + "bbox": [ + 55, + 191, + 290, + 417 + ], + "spans": [ + { + "bbox": [ + 55, + 191, + 290, + 417 + ], + "type": "image", + "image_path": "a016ec165a49cb5a38e9690d8799ed00c5aec061fc538666c804d7e4ee4b0c79.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 423, + 296, + 491 + ], + "lines": [ + { + "bbox": [ + 55, + 423, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 296, + 491 + ], + "type": "text", + "content": "Figure 7. Qualitative shape matching results computed with ULRSSM [13], SpiderMatch [67] and ours. We visualise matchings using colour and triangulation transfer from source to target shape. At legs and arms of shapes we can see stronger geometric consistency of ours compared to SpiderMatch [67] (see red and green arrows and distorted triangles)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 508, + 261, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 261, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 261, + 521 + ], + "type": "text", + "content": "5.2. Application to Planar Graph Matching" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 525, + 296, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 296, + 610 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 296, + 610 + ], + "type": "text", + "content": "As a proof of concept we show that our formalism directly applies to planar graph matching. To this end, we use keypoints on images of the WILLOW [16] dataset and obtain graphs by computing Delaunay triangulations of keypoints. Furthermore, we represent one of the resulting graphs using surface cycles to match it to the other graph, see Fig. 9 and in Sec. D.2 in the supplementary." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 58, + 625, + 294, + 684 + ], + "blocks": [ + { + "bbox": [ + 58, + 625, + 294, + 684 + ], + "lines": [ + { + "bbox": [ + 58, + 625, + 294, + 684 + ], + "spans": [ + { + "bbox": [ + 58, + 625, + 294, + 684 + ], + "type": "image", + "image_path": "d8a786580d2b7c194417389b964a2002c231d1c82fc7774f4fa0741053f02419.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 688, + 295, + 711 + ], + "lines": [ + { + "bbox": [ + 55, + 688, + 295, + 711 + ], + "spans": [ + { + "bbox": [ + 55, + 688, + 295, + 711 + ], + "type": "text", + "content": "Figure 8. Qualitative results computed with ours for partial-to-full shape matching of test set shapes [26] from SHREC'16 [17]." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 315, + 69, + 553, + 133 + ], + "blocks": [ + { + "bbox": [ + 315, + 69, + 553, + 133 + ], + "lines": [ + { + "bbox": [ + 315, + 69, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 315, + 69, + 553, + 133 + ], + "type": "image", + "image_path": "b7ac21ad303a65684758c20e553fbc6dc0c300a439d0d3eab932387da6fead76.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 143, + 555, + 166 + ], + "lines": [ + { + "bbox": [ + 313, + 143, + 555, + 166 + ], + "spans": [ + { + "bbox": [ + 313, + 143, + 555, + 166 + ], + "type": "text", + "content": "Figure 9. Examples of planar graph matching results using our method on ducks and cars from WILLOW [16] dataset." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 184, + 459, + 196 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 184, + 459, + 196 + ], + "spans": [ + { + "bbox": [ + 313, + 184, + 459, + 196 + ], + "type": "text", + "content": "6. Discussion & Limitations" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 205, + 555, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 555, + 361 + ], + "type": "text", + "content": "Our algorithm enforces geometric consistency and thus introduces a strong prior to resolve ambiguities in 3D shape matching problems. Yet, in theory our formalism contains undesirable matchings in the solution space: On the one hand, multiple surface cycles could be matched to the same vertex on the target shape. While this ensures our optimisation problem is always feasible (since it contains the extreme case of matching all surface cycles to a single vertex), this is an undesirable solution. In the future, this could be resolved by generalising our approach towards a symmetric formalism. On the other hand, our formalism allows for inside-out flips, as each individual matching element (i.e. an edge) cannot disambiguate extrinsic orientations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 363, + 556, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 556, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 556, + 472 + ], + "type": "text", + "content": "We have empirically shown that results can be computed efficiently using off-the-shelf LP-solvers [37] and that in all considered instances the linear the programming relaxations are tight, i.e. we find a globally optimal solution for all instances. Yet, we observe that our constraint matrix is not totally unimodular. We leave an in-depth analysis of our formalism, along with answering the question whether there exists a polynomial time algorithm for geometrically consistent 3D shape matching, for future works." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 487, + 388, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 487, + 388, + 500 + ], + "spans": [ + { + "bbox": [ + 314, + 487, + 388, + 500 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": "We have presented a novel formalism for non-rigid 3D shape matching which globally enforces geometric consistency. Our key idea is to consider a shape representation that allows to cast 3D shape matching as an integer linear program, which is efficiently solvable in practice. In addition, we illustrate that the resulting problem can be interpreted as a minimum-cost flow circulation problem in a hyper graph. Overall, we consider our work to be important for the 3D shape analysis community. Geometric consistency is crucial in most practical 3D shape matching settings, which is often times ignored (especially by learning-based methods). Our work allows to obtain such matchings efficiently and with that poses a step towards integrating geometric consistency into learning-based methods. Furthermore, we hope to inspire follow-up works on neighbourhood preserving matching formalisms within the broader field of visual computing and beyond." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "919" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 149, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 149, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 149, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 138 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 138 + ], + "type": "text", + "content": "We thank Johan Thunberg, Tobias Weißberg and Zorah Lähner for insightful discussions about our formalism. This work is supported by the ERC starting grant no. 101160648 (Harmony)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 115, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 115, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 115, + 163 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 170, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 61, + 170, + 294, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 170, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 61, + 170, + 294, + 191 + ], + "type": "text", + "content": "[1] MOSEK ApS. MOSEK Optimizer API for Python 11.0.9, 2025. 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 192, + 295, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 192, + 295, + 223 + ], + "spans": [ + { + "bbox": [ + 62, + 192, + 295, + 223 + ], + "type": "text", + "content": "[2] Souhaib Attaiki and Maks Ovsjanikov. Understanding and improving features learned in deep functional maps. In CVPR, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 224, + 294, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 224, + 294, + 255 + ], + "spans": [ + { + "bbox": [ + 62, + 224, + 294, + 255 + ], + "type": "text", + "content": "[3] Mathieu Aubry, Ulrich Schlickewei, and Daniel Cremers. The wave kernel signature: A quantum mechanical approach to shape analysis. In ICCV workshops. IEEE, 2011. 4, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 255, + 294, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 255, + 294, + 287 + ], + "spans": [ + { + "bbox": [ + 62, + 255, + 294, + 287 + ], + "type": "text", + "content": "[4] Lennart Bastian, Yizheng Xie, Nassir Navab, and Zorah Lahner. Hybrid functional maps for crease-aware nonisometric shape matching. In CVPR, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 287, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 287, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 62, + 287, + 294, + 308 + ], + "type": "text", + "content": "[5] Isabel Beckenbach. A hypergraph network simplex algorithm. In GOR, 2018. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 308, + 294, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 308, + 294, + 329 + ], + "spans": [ + { + "bbox": [ + 62, + 308, + 294, + 329 + ], + "type": "text", + "content": "[6] Isabel Leonie Beckenbach. Matchings and flows in hypergraphs. PhD thesis, 2019. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 329, + 294, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 329, + 294, + 372 + ], + "spans": [ + { + "bbox": [ + 62, + 329, + 294, + 372 + ], + "type": "text", + "content": "[7] Marcel Seelbach Benkner, Zorah Lähner, Vladislav Golyanik, Christof Wunderlich, Christian Theobalt, and Michael Moeller. Q-match: Iterative shape matching via quantum annealing. In ICCV, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 373, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 373, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 62, + 373, + 294, + 403 + ], + "type": "text", + "content": "[8] Florian Bernard, Zeeshan Khan Suri, and Christian Theobalt. Mina: Convex mixed-integer programming for non-rigid shape alignment. In CVPR, 2020. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 404, + 294, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 404, + 294, + 435 + ], + "spans": [ + { + "bbox": [ + 62, + 404, + 294, + 435 + ], + "type": "text", + "content": "[9] Federica Bogo, Javier Romero, Matthew Loper, and Michael J Black. Faust: Dataset and evaluation for 3d mesh registration. In CVPR, 2014. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 435, + 295, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 435, + 295, + 490 + ], + "spans": [ + { + "bbox": [ + 57, + 435, + 295, + 490 + ], + "type": "text", + "content": "[10] Kateřina Böhmová, Jérémie Chalopin, Matúš Mihalák, Guido Proietti, and Peter Widmayer. Sequence hypergraphs: paths, flows, and cuts. Adventures Between Lower Bounds and Higher Altitudes: Essays Dedicated to Juraj Hromković on the Occasion of His 60th Birthday, 2018. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 490, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 490, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 57, + 490, + 295, + 544 + ], + "type": "text", + "content": "[11] Alexander M Bronstein, Michael M Bronstein, Ron Kimmel, Mona Mahmoudi, and Guillermo Sapiro. A gromov-hausdorff framework with diffusion geometry for topologically-robust non-rigid shape matching. International Journal of Computer Vision, 2010. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 544, + 295, + 573 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 544, + 295, + 573 + ], + "spans": [ + { + "bbox": [ + 57, + 544, + 295, + 573 + ], + "type": "text", + "content": "[12] Oliver Burghard and Reinhard Klein. Efficient lifted relaxations of the quadratic assignment problem. In CVMV, 2017. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 574, + 295, + 606 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 574, + 295, + 606 + ], + "spans": [ + { + "bbox": [ + 57, + 574, + 295, + 606 + ], + "type": "text", + "content": "[13] Dongliang Cao, Paul Roetzer, and Florian Bernard. Unsupervised learning of robust spectral shape matching. TOG, 2023. 1, 2, 7, 8, 4, 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 607, + 295, + 639 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 607, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 57, + 607, + 295, + 639 + ], + "type": "text", + "content": "[14] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 57, + 639, + 295, + 660 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 639, + 295, + 660 + ], + "spans": [ + { + "bbox": [ + 57, + 639, + 295, + 660 + ], + "type": "text", + "content": "[15] Qifeng Chen and Vladlen Koltun. Robust nonrigid registration by convex optimization. In ICCV, 2015. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 57, + 661, + 295, + 681 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 661, + 295, + 681 + ], + "spans": [ + { + "bbox": [ + 57, + 661, + 295, + 681 + ], + "type": "text", + "content": "[16] Minsu Cho, Karteek Alahari, and Jean Ponce. Learning graphs to match. In ICCV, 2013. 8, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[17] Luca Cosmo, Emanuele Rodola, Michael M Bronstein, Andrea Torsello, Daniel Cremers, and Y Sahillioglu. Shrec'16: Partial matching of deformable shapes. 3DOR, 2016. 8, 6, 9" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 703 + ], + "type": "list", + "angle": 0, + "index": 39, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[18] James Coughlan, Alan Yuille, Camper English, and Dan Snow. Efficient deformable template detection and localization without user initialization. CVIU, 2000. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 106, + 553, + 136 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 106, + 553, + 136 + ], + "spans": [ + { + "bbox": [ + 316, + 106, + 553, + 136 + ], + "type": "text", + "content": "[19] Bailin Deng, Yuxin Yao, Roberto M Dyke, and Juyong Zhang. A survey of non-rigid 3d registration. In CGF, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 137, + 553, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 137, + 553, + 169 + ], + "spans": [ + { + "bbox": [ + 316, + 137, + 553, + 169 + ], + "type": "text", + "content": "[20] Nicolas Donati, Etienne Corman, and Maks Ovsjanikov. Deep orientation-aware functional maps: Tackling symmetry issues in shape matching. In CVPR, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 170, + 553, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 170, + 553, + 201 + ], + "spans": [ + { + "bbox": [ + 316, + 170, + 553, + 201 + ], + "type": "text", + "content": "[21] Nicolas Donati, Abhishek Sharma, and Maks Ovsjanikov. Deep geometric functional maps: Robust feature learning for shape correspondence. In CVPR, 2020. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 201, + 553, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 201, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 316, + 201, + 553, + 243 + ], + "type": "text", + "content": "[22] Niladri Shekhar Dutt, Sanjeev Muralikrishnan, and Niloy J. Mitra. Diffusion 3d features (diff3f): Decorating textured shapes with distilled semantic features. In CVPR, 2024. 2, 4, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 243, + 553, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 276 + ], + "type": "text", + "content": "[23] Nadav Dym, Haggai Maron, and Yaron Lipman. DS++ - A Flexible, Scalable and Provably Tight Relaxation for Matching Problems. TOG, 2017. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 276, + 553, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 340 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 340 + ], + "type": "text", + "content": "[24] Bernhard Egger, William A. P. Smith, Ayush Tewari, Stefanie Wuhrer, Michael Zollhoefer, Thabo Beeler, Florian Bernard, Timo Bolkart, Adam Kortylewski, Sami Romdhani, Christian Theobalt, Volker Blanz, and Thomas Vetter. 3d morphable face models—past, present, and future. TOG, 2020. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 341, + 553, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 341, + 553, + 405 + ], + "spans": [ + { + "bbox": [ + 316, + 341, + 553, + 405 + ], + "type": "text", + "content": "[25] Viktoria Ehm, Nafie El Amrani, Yizheng Xie, Lennart Bastian, Maolin Gao, Weikang Wang, Lu Sang, Dongliang Cao, Tobias Weiberg, Zorah Lahner, Daniel Cremers, and Florian Bernard. Beyond Complete Shapes: A Benchmark for Quantitative Evaluation of 3D Shape Matching Algorithms. CGF, 2025. 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 405, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 405, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 405, + 553, + 448 + ], + "type": "text", + "content": "[26] Viktoria Ehm, Maolin Gao, Paul Roetzer, Marvin Eisenberger, Daniel Cremers, and Florian Bernard. Partial-to-partial shape matching with geometric consistency. In CVPR, 2024. 3, 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 448, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 448, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 448, + 553, + 480 + ], + "type": "text", + "content": "[27] Viktoria Ehm, Paul Roetzer, Marvin Eisenberger, Maolin Gao, Florian Bernard, and Daniel Cremers. Geometrically consistent partial shape matching. In 3DV, 2024. 3, 9" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 481, + 553, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 481, + 553, + 512 + ], + "spans": [ + { + "bbox": [ + 316, + 481, + 553, + 512 + ], + "type": "text", + "content": "[28] Marvin Eisenberger, Zorah Lahner, and Daniel Cremers. Smooth shells: Multi-scale shape registration with functional maps. In CVPR, 2020. 2, 7, 8, 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 512, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 553, + 533 + ], + "type": "text", + "content": "[29] James R Evans and John J Jarvis. Network topology and integral multicommodity flow problems. Networks, 1978. 6" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 533, + 553, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 553, + 553 + ], + "type": "text", + "content": "[30] Danielle Ezuz and Mirela Ben-Chen. Deblurring and denoising of maps between shapes. In CGF, 2017. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 554, + 553, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 554, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 316, + 554, + 553, + 586 + ], + "type": "text", + "content": "[31] Danielle Ezuz, Behrend Heeren, Omri Azencot, Martin Rumpf, and Mirela Ben-Chen. Elastic correspondence between triangle meshes. In CGF, 2019. 2" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 316, + 586, + 553, + 606 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 586, + 553, + 606 + ], + "spans": [ + { + "bbox": [ + 316, + 586, + 553, + 606 + ], + "type": "text", + "content": "[32] Pedro F Felzenszwalb. Representation and detection of deformable shapes. TPAMI, 2005. 2" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 316, + 606, + 553, + 639 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 606, + 553, + 639 + ], + "spans": [ + { + "bbox": [ + 316, + 606, + 553, + 639 + ], + "type": "text", + "content": "[33] Giorgio Gallo, Giustino Longo, Stefano Pallottino, and Sang Nguyen. Directed hypergraphs and applications. Discrete applied mathematics, 1993. 2" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 316, + 639, + 553, + 681 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 639, + 553, + 681 + ], + "spans": [ + { + "bbox": [ + 316, + 639, + 553, + 681 + ], + "type": "text", + "content": "[34] Maolin Gao, Paul Roetzer, Marvin Eisenberger, Zorah Lahner, Michael Moeller, Daniel Cremers, and Florian Bernard. Sigma: Scale-invariant global sparse shape matching. In ICCV, 2023. 1, 2" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 316, + 681, + 553, + 703 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 703 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 703 + ], + "type": "text", + "content": "[35] Andrea Gasparetto, Luca Cosmo, Emanuele Rodola, Michael Bronstein, and Andrea Torsello. Spatial maps: From" + } + ] + } + ], + "index": 38 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "920" + } + ] + } + ], + "index": 40 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 94 + ], + "type": "text", + "content": "low rank spectral to sparse spatial functional representations. In 3DV, 2017. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 95, + 295, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 95, + 295, + 126 + ], + "spans": [ + { + "bbox": [ + 57, + 95, + 295, + 126 + ], + "type": "text", + "content": "[36] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. 3d-coded: 3d correspondences by deep deformation. In ECCV, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 126, + 294, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 126, + 294, + 147 + ], + "spans": [ + { + "bbox": [ + 57, + 126, + 294, + 147 + ], + "type": "text", + "content": "[37] Gurobi Optimization, LLC. Gurobi Optimizer Reference Manual, 2023. 6, 8, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 147, + 294, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 147, + 294, + 178 + ], + "spans": [ + { + "bbox": [ + 56, + 147, + 294, + 178 + ], + "type": "text", + "content": "[38] Oshri Halimi, Or Litany, Emanuele Rodola, Alex M Bronstein, and Ron Kimmel. Unsupervised learning of dense shape correspondence. In CVPR, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 178, + 295, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 178, + 295, + 232 + ], + "spans": [ + { + "bbox": [ + 56, + 178, + 295, + 232 + ], + "type": "text", + "content": "[39] Stefan Haller, Lorenz Feineis, Lisa Hutschenreiter, Florian Bernard, Carsten Rother, Dagmar Kainmüller, Paul Swoboda, and Bogdan Savchynskyy. A comparative study of graph matching algorithms in computer vision. In ECCV, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 232, + 294, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 294, + 263 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 294, + 263 + ], + "type": "text", + "content": "[40] Benjamin Holzschuh, Zorah Lähner, and Daniel Cremers. Simulated annealing for 3d shape correspondence. In 3DV, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 264, + 294, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 286 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 286 + ], + "type": "text", + "content": "[41] Alec Jacobson, Daniele Panozzo, et al. libigl: A simple C++ geometry processing library, 2018. https://libigl.github.io/. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 286, + 294, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 318 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 318 + ], + "type": "text", + "content": "[42] Robert G Jeroslow, Kipp Martin, Ronald L Rardin, and Jinchang Wang. Gainfree leontief substitution flow problems. Mathematical Programming, 1992. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 318, + 294, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 318, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 56, + 318, + 294, + 338 + ], + "type": "text", + "content": "[43] Vladimir G Kim, Yaron Lipman, and Thomas Funkhouser. Blended intrinsic maps. TOG, 2011. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 339, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 339, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 339, + 294, + 381 + ], + "type": "text", + "content": "[44] Whoi-Yul Kim and Avinash C. Kak. 3-d object recognition using bipartite matching embedded in discrete relaxation. IEEE Transactions on Pattern Analysis & Machine Intelligence, 1991. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 381, + 294, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 381, + 294, + 423 + ], + "spans": [ + { + "bbox": [ + 56, + 381, + 294, + 423 + ], + "type": "text", + "content": "[45] Eugene Krissinel and Kim Henrick. Secondary-structure matching (ssm), a new tool for fast protein structure alignment in three dimensions. Biological Crystallography, 2004. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 424, + 294, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 424, + 294, + 445 + ], + "spans": [ + { + "bbox": [ + 56, + 424, + 294, + 445 + ], + "type": "text", + "content": "[46] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 1955. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 445, + 294, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 477 + ], + "type": "text", + "content": "[47] Yam Kushinsky, Haggai Maron, Nadav Dym, and Yaron Lipman. Sinkhorn Algorithm for Lifted Assignment Problems. SIAM Journal on Imaging Sciences, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 477, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 477, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 477, + 294, + 520 + ], + "type": "text", + "content": "[48] Zorah Lähner, Emanuele Rodola, Frank R Schmidt, Michael M Bronstein, and Daniel Cremers. Efficient globally optimal 2d-to-3d deformable shape matching. In CVPR, 2016. 2, 4, 1, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 520, + 294, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 520, + 294, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 520, + 294, + 563 + ], + "type": "text", + "content": "[49] Laura Leal-Taixe and Bodo Rosenhahn. Pedestrian interaction in tracking: the social force model and global optimization methods. In Modeling, Simulation and Visual Analysis of Crowds: A Multidisciplinary Perspective. 2013. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 563, + 294, + 584 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 563, + 294, + 584 + ], + "spans": [ + { + "bbox": [ + 56, + 563, + 294, + 584 + ], + "type": "text", + "content": "[50] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. In ECCV, 2024. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 584, + 294, + 615 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 584, + 294, + 615 + ], + "spans": [ + { + "bbox": [ + 56, + 584, + 294, + 615 + ], + "type": "text", + "content": "[51] Lei Li, Nicolas Donati, and Maks Ovsjanikov. Learning multi-resolution functional maps with spectral attention for robust shape matching. NIPS, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 56, + 616, + 294, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 616, + 294, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 616, + 294, + 647 + ], + "type": "text", + "content": "[52] Qinsong Li, Shengjun Liu, Ling Hu, and Xinru Liu. Shape correspondence using anisotropic chebyshev spectral cnns. In CVPR, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 56, + 647, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 679 + ], + "type": "text", + "content": "[53] Yang Li, Hikari Takehara, Takafumi Taketomi, Bo Zheng, and Matthias Nießner. 4dcomplete: Non-rigid motion estimation beyond the observable surface. In ICCV, 2021. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 56, + 679, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 679, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 679, + 294, + 712 + ], + "type": "text", + "content": "[54] Or Litany, Tal Remez, Emanuele Rodola, Alex Bronstein, and Michael Bronstein. Deep functional maps: Structured prediction for dense shape correspondence. In ICCV, 2017." + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 74, + 553, + 710 + ], + "type": "list", + "angle": 0, + "index": 42, + "blocks": [ + { + "bbox": [ + 335, + 74, + 342, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 342, + 83 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 342, + 83 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 83, + 553, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 83, + 553, + 115 + ], + "spans": [ + { + "bbox": [ + 316, + 83, + 553, + 115 + ], + "type": "text", + "content": "[55] Jiayi Ma, Xingyu Jiang, Aoxiang Fan, Junjun Jiang, and Junchi Yan. Image matching from handcrafted to deep features: A survey. IJCV, 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 116, + 553, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 116, + 553, + 147 + ], + "spans": [ + { + "bbox": [ + 316, + 116, + 553, + 147 + ], + "type": "text", + "content": "[56] Filippo Maggioli, Daniele Baieri, Emanuele Rodola, and Simone Melzi. Rematching: Low-resolution representations for scalable shape correspondence. In ECCV, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 147, + 553, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 147, + 553, + 178 + ], + "spans": [ + { + "bbox": [ + 316, + 147, + 553, + 178 + ], + "type": "text", + "content": "[57] Robin Magnet, Jing Ren, Olga Sorkine-Hornung, and Maks Ovsjanikov. Smooth non-rigid shape matching via effective dirichlet energy optimization. In 3DV, 2022. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 178, + 553, + 211 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 178, + 553, + 211 + ], + "spans": [ + { + "bbox": [ + 316, + 178, + 553, + 211 + ], + "type": "text", + "content": "[58] Haggai Maron, Nadav Dym, Itay Kezurer, Shahar Kovalsky, and Yaron Lipman. Point registration via efficient convex relaxation. TOG, 2016. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 211, + 553, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 211, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 316, + 211, + 553, + 243 + ], + "type": "text", + "content": "[59] Maks Ovsjanikov, Mirela Ben-Chen, Justin Solomon, Adrian Butscher, and Leonidas Guibas. Functional maps: a flexible representation of maps between shapes. TOG, 2012. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 243, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 274 + ], + "type": "text", + "content": "[60] Maks Ovsjanikov, Quentin Merigot, Facundo Memoli, and Leonidas Guibas. One point isometric matching with the heat kernel. In CGF, 2010. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 274, + 553, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 553, + 306 + ], + "type": "text", + "content": "[61] Thomas Probst, Danda Pani Paudel, Ajad Chhatkuli, and Luc Van Gool. Unsupervised learning of consensus maximization for 3d vision problems. In CVPR, 2019. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 306, + 553, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 306, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 306, + 553, + 338 + ], + "type": "text", + "content": "[62] Jing Ren, Simone Melzi, Peter Wonka, and Maks Ovsjanikov. Discrete optimization for shape matching. In CGF, 2021. 2, 7, 8, 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 339, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 339, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 339, + 553, + 370 + ], + "type": "text", + "content": "[63] Jing Ren, Adrien Poulenard, Peter Wonka, and Maks Ovsjanikov. Continuous and orientation-preserving correspondences via functional maps. TOG, 2018. 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 370, + 553, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 370, + 553, + 402 + ], + "spans": [ + { + "bbox": [ + 316, + 370, + 553, + 402 + ], + "type": "text", + "content": "[64] F Rendl, P Pardalos, and H Wolkowicz. The Quadratic Assignment Problem: A Survey and Recent Developments. In DIMACS workshop, 1994. 1, 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 402, + 553, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 402, + 553, + 435 + ], + "spans": [ + { + "bbox": [ + 316, + 402, + 553, + 435 + ], + "type": "text", + "content": "[65] Emanuele Rodola, Alex M Bronstein, Andrea Albarelli, Filippo Bergamasco, and Andrea Torsello. A game-theoretic approach to deformable shape matching. In CVPR, 2012. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 435, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 435, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 435, + 553, + 476 + ], + "type": "text", + "content": "[66] Paul Roetzer, Ahmed Abbas, Dongliang Cao, Florian Bernard, and Paul Swoboda. Discomatch: Fast discrete optimisation for geometrically consistent 3d shape matching. In ECCV, 2024. 3, 7" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 476, + 553, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 476, + 553, + 508 + ], + "spans": [ + { + "bbox": [ + 316, + 476, + 553, + 508 + ], + "type": "text", + "content": "[67] Paul Roetzer and Florian Bernard. Spidermatch: 3d shape matching with global optimality and geometric consistency. In CVPR, 2024. 2, 3, 7, 8, 6" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 316, + 508, + 553, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 508, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 316, + 508, + 553, + 540 + ], + "type": "text", + "content": "[68] Paul Roetzer, Zorah Lähner, and Florian Bernard. Conjugate product graphs for globally optimal 2d-3d shape matching. In CVPR, 2023. 2" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 316, + 540, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 540, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 316, + 540, + 553, + 581 + ], + "type": "text", + "content": "[69] Paul Roetzer, Paul Swoboda, Daniel Cremers, and Florian Bernard. A scalable combinatorial solver for elastic geometrically consistent 3d shape matching. In CVPR, 2022. 3, 7" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 316, + 583, + 553, + 615 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 583, + 553, + 615 + ], + "spans": [ + { + "bbox": [ + 316, + 583, + 553, + 615 + ], + "type": "text", + "content": "[70] David Sankoff and Joseph B Kruskal. Time warps, string edits, and macromolecules: the theory and practice of sequence comparison. 1983. 2" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 316, + 616, + 553, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 616, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 316, + 616, + 553, + 647 + ], + "type": "text", + "content": "[71] Frank R. Schmidt, Eno Töppe, and Daniel Cremers. Efficient planar graph cuts with applications in computer vision. In CVPR, 2009. 2" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 316, + 647, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 679 + ], + "type": "text", + "content": "[72] Patrick Schmidt, Marcel Campen, Janis Born, and Leif Kobbelt. Inter-surface maps via constant-curvature metrics. TOG, 39(4):119-1, 2020. 2" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 316, + 679, + 553, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 679, + 553, + 700 + ], + "spans": [ + { + "bbox": [ + 316, + 679, + 553, + 700 + ], + "type": "text", + "content": "[73] Patrick Schmidt, Dörte Pieper, and Leif Kobbelt. Surface maps via adaptive triangulations. CGF, 2023. 2" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 316, + 700, + 553, + 710 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 700, + 553, + 710 + ], + "spans": [ + { + "bbox": [ + 316, + 700, + 553, + 710 + ], + "type": "text", + "content": "[74] Thomas Schoenemann and Daniel Cremers. A combinatorial" + } + ] + } + ], + "index": 41 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "921" + } + ] + } + ], + "index": 43 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 73, + 294, + 701 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "text", + "content": "solution for model-based image segmentation and real-time tracking. TPAMI, 2009. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 95, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 95, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 57, + 95, + 294, + 115 + ], + "type": "text", + "content": "[75] John Schreiner, Arul Asirvatham, Emil Praun, and Hugues Hoppe. Inter-surface mapping. In SIGGRAPH. 2004. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 116, + 294, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 116, + 294, + 137 + ], + "spans": [ + { + "bbox": [ + 57, + 116, + 294, + 137 + ], + "type": "text", + "content": "[76] Alexander Schrijver et al. Combinatorial optimization: polyhedra and efficiency. Springer, 2003. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 137, + 294, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 137, + 294, + 168 + ], + "spans": [ + { + "bbox": [ + 57, + 137, + 294, + 168 + ], + "type": "text", + "content": "[77] Avinash Sharma, Radu Horaud, Jan Cech, and Edmond Boyer. Topologically-robust 3d shape matching based on diffusion geometry and seed growing. In CVPR, 2011. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 168, + 294, + 199 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 168, + 294, + 199 + ], + "spans": [ + { + "bbox": [ + 57, + 168, + 294, + 199 + ], + "type": "text", + "content": "[78] Nicholas Sharp, Souhaib Attaiki, Keenan Crane, and Maks Ovsjanikov. Diffusionnet: Discretization agnostic learning on surfaces. TOG, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 200, + 294, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 200, + 294, + 232 + ], + "spans": [ + { + "bbox": [ + 57, + 200, + 294, + 232 + ], + "type": "text", + "content": "[79] Justin Solomon, Gabriel Peyré, Vladimir G Kim, and Suvrit Sra. Entropic metric alignment for correspondence problems. TOG, 2016. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 232, + 294, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 232, + 294, + 263 + ], + "spans": [ + { + "bbox": [ + 57, + 232, + 294, + 263 + ], + "type": "text", + "content": "[80] Y Soun and Klaus Truemper. Single commodity representation of multicommodity networks. SIAM Journal on Algebraic Discrete Methods, 1980. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 264, + 294, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 264, + 294, + 284 + ], + "spans": [ + { + "bbox": [ + 57, + 264, + 294, + 284 + ], + "type": "text", + "content": "[81] Kenshi Takayama. Compatible intrinsic triangulations. TOG, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 285, + 294, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 285, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 57, + 285, + 294, + 337 + ], + "type": "text", + "content": "[82] Gary KL Tam, Zhi-Quan Cheng, Yu-Kun Lai, Frank C Langbein, Yonghuai Liu, David Marshall, Ralph R Martin, Xian-Fang Sun, and Paul L Rosin. Registration of 3d point clouds and meshes: A survey from rigid to nonrigid. TVCG, 2012. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 338, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 338, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 57, + 338, + 294, + 370 + ], + "type": "text", + "content": "[83] Giovanni Trappolini, Luca Cosmo, Luca Moschella, Ricardo Marin, Simone Melzi, and Emanuele Rodola. Shape registration in the time of transformers. NeurIPS, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 371, + 294, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 371, + 294, + 391 + ], + "spans": [ + { + "bbox": [ + 57, + 371, + 294, + 391 + ], + "type": "text", + "content": "[84] Klaus Truemper. Unimodular matrices of flow problems with additional constraints. Networks, 1977. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 392, + 294, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 392, + 294, + 422 + ], + "spans": [ + { + "bbox": [ + 57, + 392, + 294, + 422 + ], + "type": "text", + "content": "[85] Oliver Van Kaick, Hao Zhang, Ghassan Hamarneh, and Daniel Cohen-Or. A survey on shape correspondence. In CGF, 2011. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 423, + 294, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 423, + 294, + 477 + ], + "spans": [ + { + "bbox": [ + 57, + 423, + 294, + 477 + ], + "type": "text", + "content": "[86] Matthias Van Onsem, Dieter De Paepe, S Vanden Hautte, Pieter Bonte, Veerle Ledoux, Annelies Lejon, Femke Ongenae, Dennis Dreesen, and Sofie Van Hoecke. Hierarchical pattern matching for anomaly detection in time series. Computer Communications, 2022. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 477, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 477, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 57, + 477, + 294, + 520 + ], + "type": "text", + "content": "[87] Matthias Vestner, Roee Litman, Emanuele Rodola, Alex Bronstein, and Daniel Cremers. Product manifold filter: Non-rigid shape correspondence via kernel density estimation in the product space. In CVPR, 2017. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 520, + 294, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 520, + 294, + 551 + ], + "spans": [ + { + "bbox": [ + 57, + 520, + 294, + 551 + ], + "type": "text", + "content": "[88] Ruben Wiersma, Elmar Eisemann, and Klaus Hildebrandt. Cnns on surfaces using rotation-equivariant features. TOG, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 552, + 294, + 594 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 552, + 294, + 594 + ], + "spans": [ + { + "bbox": [ + 57, + 552, + 294, + 594 + ], + "type": "text", + "content": "[89] Thomas Windheuser, Ulrich Schlickewei, Frank R Schmidt, and Daniel Cremers. Geometrically consistent elastic matching of 3d shapes: A linear programming solution. In ICCV, 2011. 1, 2, 7, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 595, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 595, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 595, + 294, + 635 + ], + "type": "text", + "content": "[90] Thomas Windheuser, Ulrich Schlickwei, Frank R Schimdt, and Daniel Cremers. Large-scale integer linear programming for orientation preserving 3d shape matching. In CGF, 2011. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 57, + 637, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 637, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 57, + 637, + 294, + 669 + ], + "type": "text", + "content": "[91] Junchi Yan, Xu-Cheng Yin, Weiyao Lin, Cheng Deng, Hongyuan Zha, and Xiaokang Yang. A short survey of recent advances in graph matching. In ICMR, 2016. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 57, + 670, + 294, + 701 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 701 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 701 + ], + "type": "text", + "content": "[92] Silvia Zuffi, Angjoo Kanazawa, David W Jacobs, and Michael J Black. 3d menagerie: Modeling the 3d shape and pose of animals. In CVPR, 2017. 7" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "922" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Fast Image Super-Resolution via Consistency Rectified Flow/5e42c446-52f8-4769-b738-a2b096f3a7a5_content_list.json b/2025/Fast Image Super-Resolution via Consistency Rectified Flow/5e42c446-52f8-4769-b738-a2b096f3a7a5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d3f043fa3839348aae04c2647d6e683b4c7e6e01 --- /dev/null +++ b/2025/Fast Image Super-Resolution via Consistency Rectified Flow/5e42c446-52f8-4769-b738-a2b096f3a7a5_content_list.json @@ -0,0 +1,2220 @@ +[ + { + "type": "text", + "text": "Fast Image Super-Resolution via Consistency Rectified Flow", + "text_level": 1, + "bbox": [ + 194, + 130, + 803, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiaqi $\\mathrm{Xu}^{1,2*}$ , Wenbo $\\mathrm{Li}^{2\\dagger}$ , Haoze $\\mathrm{Sun}^2$ , Fan $\\mathrm{Li}^2$ , Zhixin $\\mathrm{Wang}^2$ , Long $\\mathrm{Peng}^2$ , Jingjing $\\mathrm{Ren}^3$ , Haoran $\\mathrm{Yang}^1$ , Xiaowei $\\mathrm{Hu}^4$ , Renjing $\\mathrm{Pei}^{2\\dagger}$ , and Pheng-Ann $\\mathrm{Heng}^1$", + "bbox": [ + 142, + 179, + 852, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1The Chinese University of Hong Kong 2Huawei Noah's Ark Lab 3HKUST (GZ) 4South China University of Technology", + "bbox": [ + 238, + 219, + 756, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 289, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion models (DMs) have demonstrated remarkable success in real-world image super-resolution (SR), yet their reliance on time-consuming multi-step sampling largely hinders their practical applications. While recent efforts have introduced few- or single-step solutions, existing methods either inefficiently model the process from noisy input or fail to fully exploit iterative generative priors, compromising the fidelity and quality of the reconstructed images. To address this issue, we propose FlowSR, a novel approach that reformulates the SR problem as a rectified flow from low-resolution (LR) to high-resolution (HR) images. Our method leverages an improved consistency learning strategy to enable high-quality SR in a single step. Specifically, we refine the original consistency distillation process by incorporating HR regularization, ensuring that the learned SR flow not only enforces self-consistency but also converges precisely to the ground-truth HR target. Furthermore, we introduce a fast-slow scheduling strategy, where adjacent timesteps for consistency learning are sampled from two distinct schedulers: a fast scheduler with fewer timesteps to improve efficiency, and a slow scheduler with more timesteps to capture fine-grained texture details. Extensive experiments demonstrate that FlowSR achieves outstanding performance in both efficiency and image quality.", + "bbox": [ + 88, + 321, + 483, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 715, + 220, + 732 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Real-world image super-resolution (SR) aims to reconstruct high-resolution (HR) images from their low-resolution (LR) counterparts while simultaneously removing unknown degradations. With the remarkable success of diffusion models (DMs) [10, 37], SR methods leveraging these models—particularly those built on powerful text-to-image (T2I) models like Stable Diffusion (SD) [31]—have demonstrated outstanding performance [42, 49, 58]. However,", + "bbox": [ + 89, + 741, + 483, + 863 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9e40a1bbe783c249056bf060a0e4fe9007e9106282587e07ec741ee5b370fc60.jpg", + "image_caption": [ + "LR" + ], + "image_footnote": [], + "bbox": [ + 517, + 287, + 614, + 417 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/04211c3713bb2157bb9098e89025f7dd7483308b81d3e26e1207556f740e36a2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 289, + 707, + 362 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5c730eb00a9e8e280a5248bfc3ef3d4093360c5c8a0b254d7b8dd5b43568c0a1.jpg", + "image_caption": [ + "1-step" + ], + "image_footnote": [], + "bbox": [ + 614, + 363, + 707, + 417 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/61e635f8088d6668e4b0bcc9fdf3e3510f5d9aaa03f499ca25f2baa5a738310a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 289, + 803, + 362 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/07628966f945812b770dca43a957ffd6e3d71db2bcd5402911110f120610080f.jpg", + "image_caption": [ + "4-step" + ], + "image_footnote": [], + "bbox": [ + 709, + 363, + 802, + 417 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a1f8bc0202f25b9aac971afa920984f5bdc92d07c1552ed072191f50fdf81acf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 803, + 289, + 898, + 362 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/aa55a7b9cbf5adeee59ea0a73ab91d51d81f58bb0a06fa8d5e89535c57d1e3cd.jpg", + "image_caption": [ + "distilled, 1-step" + ], + "image_footnote": [], + "bbox": [ + 803, + 363, + 898, + 417 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/98c21e30f8e91eb87338b126ec47bf7680cb0d4b8cab350af7b327b72aa46e84.jpg", + "image_caption": [ + "Figure 1. Our consistency SR flow model achieves high-quality single-step inference (distilled, 1-step) by distilling from multi-step higher-quality sampling process (top). We achieve this by formulating SR as a rectified flow that bridges LR and HR images, combined with improved consistency learning (bottom)." + ], + "image_footnote": [], + "bbox": [ + 517, + 441, + 898, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "these diffusion-based SR approaches require an iterative reverse sampling process that gradually refines noisy inputs into HR outputs. Despite advancements such as efficient ordinary differential equation (ODE) solvers like DDIM [36], the slow inference speed remains a bottleneck, limiting their practicality in real-world applications.", + "bbox": [ + 511, + 641, + 906, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, few-step or single-step SR methods derived from diffusion models are designed [3, 45, 48, 59, 61]. Several studies focus on designing more efficient diffusion processes for SR. ResShift [59] speeds up diffusion by shifting the LR-HR residual into the Markov chain and reduces sampling to 15 steps, while DoSSR [3] introduces a diffusion process more compatible with pre-trained DMs. However, they rely on DDPM [10], which uses a more curved transition trajectory and starts with a noise-perturbed LR image that loses crucial information, ultimately limiting performance. In parallel, another line of work tar", + "bbox": [ + 511, + 734, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Part of work done as an intern at Huawei.", + "bbox": [ + 107, + 875, + 334, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11755", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "gets single-step SR by learning the LR-HR mapping directly. For example, SinSR [45] learns one-step SR prediction by leveraging the teacher's output from ResShift [59] and ground-truth HR as target, whereas OSEDiff adapts pre-trained SD for SR and refines one-step prediction with the VSD loss [46]. However, these methods fail to fully harness the advantages of iterative generative modeling, which naturally facilitates high-quality texture synthesis.", + "bbox": [ + 89, + 90, + 480, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we present FlowSR, which unifies rectified flow with consistency models (CMs) to enable efficient, single-step image super-resolution. As illustrated in Fig. 1, we reformulate SR as a rectified flow [21], which establishes a simple and straight ODE-based mapping between LR and HR images. By explicitly modeling this trajectory, FlowSR learns fine-grained LR-to-HR transformations for better SR quality and facilitates fast sampling. Building upon this formulation, we further leverage consistency models [38] to enhance single-step inference. Enforcing consistency across points on the same SR flow trajectory distills multi-step higher-quality restoration into fewer steps that reach to the same HR result.", + "bbox": [ + 89, + 214, + 482, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, the naive consistency distillation (CD) objective in CMs is suboptimal for SR flow, where the SR task demands both high quality and high fidelity. While CD enforces self-consistency across discretization steps, there is no guarantee that the final distillation target aligns well with the ground-truth HR. To address this, we propose HR-regularized consistency learning, which explicitly requires the model's predictions to match real HR images. This additional constraint mitigates teacher-induced errors in the distillation target and enhances the reconstruction of fine-grained details. To further improve efficiency and robustness of the consistency SR flow model, we introduce a fast-slow time scheduling strategy. Rather than sampling distillation timestep pairs from the two boundaries of a discretized interval, we sample adjacent timesteps from distinct \"fast\" and \"slow\" schedulers. The fast scheduler uses fewer timesteps to facilitate efficient inference, while the slow scheduler employs more granular steps to maintain alignment with the SR flow objectives. This mixed sampling introduces large and flexible jumps with mild perturbations into the HR regularization steps, allowing for a broader coverage of SR flow trajectories.", + "bbox": [ + 89, + 412, + 482, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To train our consistency SR flow model, we first finetune a pre-trained SD model to align with the SR flow objective, followed by consistency SR flow distillation. Additionally, we incorporate a GAN [7] loss and an image quality alignment loss, with the latter promoting desirable text-described attributes in the restored images, thereby improving the SR quality.", + "bbox": [ + 89, + 747, + 480, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As illustrated in Fig. 1, our model leverages the proposed consistency SR flow distillation to transfer the quality improvement typically achieved with multiple sampling", + "bbox": [ + 89, + 854, + 480, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "steps into a single step. Experiments on real-world datasets demonstrate the superiority of FlowSR. Our core contribution lie in a new paradigm to solve the one-step SR problem. First, we explore efficient flow modeling for SR and identify potential challenges when incorporating consistency models. Second, we introduce several techniques, including HR regularization in consistency learning and a fast-slow time scheduling to enhance one-step SR performance.", + "bbox": [ + 511, + 90, + 903, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 224, + 653, + 239 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Image Super-Resolution", + "text_level": 1, + "bbox": [ + 511, + 250, + 733, + 266 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Real-world image super-resolution (SR) aims to reconstruct high-resolution (HR) images from low-resolution (LR) inputs, tackling complex and unknown degradations like noise, blur, and compression artifacts. Since Dong et al. [5], numerous deep learning-based methods have been proposed to address the SR problem from different perspectives, including network design [18], model training [15], and degradation simulation [28, 44, 62]. Recently, diffusion models [10, 37] have achieved remarkable success in image generation [30, 31], which inspires their application to SR and demonstrate significant advancements [32, 42].", + "bbox": [ + 511, + 272, + 903, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion Model-Based SR Conditioned on the LR image, diffusion model-based SR iteratively denoises towards the target HR image. The LR condition can be leveraged either by input concatenation [32] or via adapters [26, 63]. StableSR [42] based on Stable Diffusion [31], uses a trainable time-aware encoder to incorporates LR. Subsequent improvements are achieved through enhanced LR conditioning in DiffBIR [19], semantics-aware prompts in SeeSR [49], reference image generation in CoSeR [39], and scaled-up training in SUPIR [58]. However, these DM-based SR approaches, starting from noise, suffer from low inference speed, typically requiring 50-200 sampling steps. In contrast, ResShift [59] constructs an efficient diffusion model by shifting the residual between HR and LR. Denoising from the noised LR reduces the sampling steps to 15 but it still suffers from limited restoration quality.", + "bbox": [ + 511, + 441, + 903, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Single-Step / Few-Step Image SR Fast diffusion model-based SR methods require only one or a few sampling steps during inference, typically achieved through knowledge distillation [23] or adversarial training [15, 50]. SinSR [45] derives a deterministic sampling process for ResShift and reduces its number of sampling steps to one using distillation. OSEDiff [48] applies variational score distillation [46, 57] to enhance one-step restoration quality by minimizing the KL-divergence between the distribution of its generated outputs and that of the pre-trained T2I model. AddSR [51] achieves 4-step inference by tailoring adversarial diffusion distillation [34] for SR. Similarly, other recent methods, accelerate the inference by using distillation [6, 8, 27] or GAN [2, 16, 60, 61]. On the other hand, more", + "bbox": [ + 511, + 688, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11756", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f92d7b59b0a007023af77f66d56f8ecc8f168c539bb3637331196eca26a6f6f2.jpg", + "image_caption": [ + "Figure 2. Overview of the training process. The consistency SR flow distills multi-step, high-quality SR capability into single-step inference, while HR regularization ensures the model converges to the high-resolution target. Times $t$ , $t'$ are sampled alternately from fast and slow schedulers. Note that the flow loss $\\mathcal{L}_{flow}$ and the consistency loss $\\mathcal{L}_{hrcd}$ are computed for different samples within each batch." + ], + "image_footnote": [], + "bbox": [ + 112, + 85, + 883, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "efficient SR diffusion processes are constructed using domain shift [3] or flow matching [35], which allow few-step sampling. Yet, they still suffer from inefficient modeling designs and suboptimal performance.", + "bbox": [ + 88, + 363, + 480, + 425 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Consistency Models", + "text_level": 1, + "bbox": [ + 89, + 433, + 279, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Consistency Models (CMs) [38] learn to map any point on the ODE trajectory to its origin, which enable single-step generation and allow trade-offs between quality and computation through multi-step sampling. CMs can be implemented for modern T2I models in latent space [24]. Consistency Trajectory Models (CTMs) [14] mitigate the multi-step sampling issues in CMs by learning an any-to-any mapping from initial points to final points on the ODE trajectory. The Phased Consistency Model [40] partitions the entire ODE trajectory into multiple sub-trajectory phases and ensures consistency within each phase, while PeRFlow [52] straightens the sub-trajectories using the Reflow [22] operation. Additionally, Consistency Flow Matching [53] further enforces velocity field consistency.", + "bbox": [ + 88, + 455, + 482, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 681, + 230, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Rectified Flow", + "text_level": 1, + "bbox": [ + 89, + 705, + 236, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Rectified flow [21] is an ODE-based generative model that constructs a simple, straight trajectory to transform samples between two distributions: $\\pi_0$ (e.g., data) and $\\pi_1$ (e.g., noise). Rectified flow defines a linear interpolation path between observed samples $X_0\\sim \\pi_0$ and $X_{1}\\sim \\pi_{1}$ as: $X_{t} = tX_{1} + (1 - t)X_{0}$ with time $t\\in [0,1]$ . The core idea is to learn a velocity field $v_{\\theta}(X_{t},t)$ , parameterized by a neural network with weights $\\theta$ , that matches the derivative of this trajectory. This is achieved by optimizing the following objective: $\\min_{\\theta}\\int_{0}^{1}\\mathbb{E}_{X_0\\sim \\pi_0,X_1\\sim \\pi_1}[\\| v_{\\theta}(X_t,t) - (X_1 - X_0)\\| ^2 ]dt,$ which encourages the learned velocity field to align with", + "bbox": [ + 89, + 728, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the direction of the straight path $X_{1} - X_{0}$ . Once trained, samples from $\\pi_{1}$ can be transformed into $\\pi_{0}$ by solving the ODE: $dX_{t} = v_{\\theta}(X_{t}, t)dt$ , $X_{1} \\sim \\pi_{1}$ . A key advantage of rectified flow lies in its straight trajectories. By design, the optimal velocity field corresponds to a constant-speed flow between $X_{0}$ and $X_{1}$ , enabling efficient and stable sampling with only a few ODE steps.", + "bbox": [ + 511, + 363, + 906, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Consistency Models", + "text_level": 1, + "bbox": [ + 511, + 483, + 702, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Consistency models [38] are a class of generative models that enable high-quality sample generation with few computational steps. These models learn to map any point $(X_{t}, t)$ along a probability flow ODE trajectory directly to its origin $X_{\\epsilon}$ , where $\\epsilon$ is a fixed small positive number denoting the trajectory's start. This capability is formalized through the self-consistency property: for any pair of points $(X_{t}, t)$ and $(X_{t'}, t')$ on the same ODE trajectory, the model $f_{\\theta}(\\cdot, t)$ is trained to satisfy: $f_{\\theta}(X_{t}, t) = f_{\\theta}(X_{t}', t') = X_{\\epsilon}$ for all $t, t' \\in [\\epsilon, 1]$ . To ensure consistency during training, the boundary condition $f_{\\theta}(X_{\\epsilon}, \\epsilon) = X_{\\epsilon}$ is enforced. Consistency models eliminate trajectory drift by mapping all intermediate states to a consistent origin, avoiding errors from iterative denoising steps. This enables them to generate high-fidelity samples even with large step sizes, such as directly mapping from $t = 1$ to the origin in a single step.", + "bbox": [ + 511, + 507, + 906, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 511, + 767, + 648, + 785 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We leverage the generative capability of rectified flow for high-quality SR and explore consistency models to enable fast inference with fewer sampling steps (i.e., one step), without compromising quality. We define a rectified flow, termed SR flow, for image super-resolution in Sec. 4.1. We then introduce the improved consistency SR flow learning method in Sec. 4.2 and detail the training in Sec. 4.3.", + "bbox": [ + 511, + 794, + 905, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11757", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Rectified Flow for SR", + "text_level": 1, + "bbox": [ + 89, + 90, + 292, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SR Flow We consider Rectified Flow naturally aligns with image super-resolution, where high-resolution (HR) images $X_{\\mathrm{HR}} \\sim \\pi_0$ and their low-resolution (LR) counterparts $X_{\\mathrm{LR}} \\sim \\pi_1$ . We define the forward process as a straight path via a linear interpolation between them:", + "bbox": [ + 89, + 112, + 483, + 188 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nX _ {t} = (1 - t) X _ {\\mathrm {H R}} + t X _ {\\mathrm {L R}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 194, + 482, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "here time $t\\in [0,1]$ with $t = 0$ corresponding to $X_{\\mathrm{HR}}\\in \\mathbb{E}^{H\\times W\\times 3}$ and $t = 1$ corresponding to $X_{\\mathrm{LR}}\\in \\mathbb{E}^{H\\times W\\times 3}$ , where $X_{\\mathrm{LR}}$ is upscaled to match the spatial dimensions of $X_{\\mathrm{HR}}$ . The learning objective of the SR neural network $v_{\\theta}$ is to regress the conditional vector fields [20, 21] by following the direction $X_{\\mathrm{LR}} - X_{\\mathrm{HR}}$ :", + "bbox": [ + 89, + 215, + 483, + 306 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {t, X _ {t}} \\| v _ {\\theta} \\left(X _ {t}, t\\right) - \\left(X _ {\\mathrm {L R}} - X _ {\\mathrm {H R}}\\right) \\| _ {2} ^ {2}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 313, + 482, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Intuitively, this SR flow establishes a smooth transition between HR and LR. During training, the SR flow model $v_{\\theta}$ implicitly learns to invert degradations (e.g., blur, noise) and recover high-frequency details through intermediate refinements such as edge sharpening and texture synthesis. At inference, HR images are reconstructed via reverse sampling along the learned trajectory. Starting from $X_{\\mathrm{LR}}$ , we solve the reverse ODE using numerical methods like Euler solver: $X_{t'} = X_t + \\Delta t \\cdot v_{\\theta}(X_t, t)$ from $t = 1$ to $t = 0$ . Notably, unlike diffusion-based SR methods that corrupt LR inputs with noise [3, 35, 59], our transition process directly starts from LR, which preserves most of the structural information in the LR image, enabling stable and efficient sampling.", + "bbox": [ + 89, + 335, + 483, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Faster Inference with SR Flow A key advantage of SR flow is its flexibility in sampling steps. It supports faster inference by reducing the number of sampling steps, requiring as few as a single step that directly maps the LR image to its corresponding HR output: $\\hat{X}_{\\mathrm{HR}} = X_{\\mathrm{LR}} - 1 \\cdot v_{\\theta}(X_{\\mathrm{LR}}, 1)$ . This is more easily achievable than the noise-to-image mapping or initializing from noisy inputs in ResShift [59], thanks to the strong correlation between LR and HR images. Additionally, SR flow also supports any number of sampling steps and produces better visual quality when using iterative multi-step sampling, as illustrated in Fig. 1.", + "bbox": [ + 89, + 536, + 483, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Consistency SR Flow", + "text_level": 1, + "bbox": [ + 89, + 709, + 289, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We enhance single-step super-resolution in SR flow models by distilling multi-step sampling capability (e.g., four steps) into one via consistency learning (Fig. 1). To ensure accurate HR reconstruction, we incorporate the HR target explicitly into consistency learning. Additionally, a fast-slow time sampling schedule is introduced to improve efficiency and robustness. Fig. 2 shows the overall training process.", + "bbox": [ + 89, + 729, + 483, + 837 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Consistency Distillation in SR Flow We start with consistency distillation (CD) to distill the capability of a (pre-trained) teacher SR flow model $v_{\\phi}(X,t)$ , enabling high-quality SR in fewer inference steps. Given timesteps $t$ and", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/32c0695adbb690cffb3c861d73585f794967f8771716d86b4ab48c68d3ad15e5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 90, + 903, + 207 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ece13ce9fb4a39df413b03a6811fd38bff15f2cc7b47fff6208b1db6651bae20.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 208, + 594, + 267 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4a1e98cf55e51444b16d07dc7af45db010b60ca8e4b635d85576d712a0eed804.jpg", + "image_caption": [ + "$X_{t^{\\prime}}$" + ], + "image_footnote": [], + "bbox": [ + 517, + 267, + 594, + 325 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/201cc9c752b509fd9869df838e4bc1d998c1c7f57d1f0c00ed8347e3c49e5dba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 594, + 208, + 671, + 267 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6e2de18bc63fa8a34c7c4892b50c23112c56d457376c770854b4b5d93c171ef4.jpg", + "image_caption": [ + "$X_{t}$" + ], + "image_footnote": [], + "bbox": [ + 594, + 267, + 671, + 325 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/56f16afb06b14c6d00c844a92b850f7a2f48c3ee491ae4b7d2882856f2d8d813.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 208, + 750, + 267 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1cd07dc6e4e4c566461dbf8e6ccb4ae43c2984e0da7520a0371a91d4f75db23b.jpg", + "image_caption": [ + "$\\hat{X}_t^\\phi$" + ], + "image_footnote": [], + "bbox": [ + 671, + 267, + 750, + 325 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/27edee5717286b1868303ddbc9b29dba251c32cf8da0420421ba49e585e1ad81.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 751, + 208, + 826, + 267 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/98fffe65e49318c271c277df25ae717964e996b77729f039ce3686a0b94dbf69.jpg", + "image_caption": [ + "$\\hat{X}_{t\\rightarrow 0}^{\\theta^{-}}$", + "Figure 3. Illustration of HR-regularized consistency learning. Approximation errors may cause the distillation target $\\hat{X}_{t\\rightarrow 0}^{\\theta^{-}}$ to deviate from the true high-quality HR target. HR regularization corrects this by enforcing HR alignment under mild perturbations." + ], + "image_footnote": [], + "bbox": [ + 751, + 267, + 826, + 325 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e3ac8bcb1b75a8ecfa075cfad810a7a01e0e93619f25a5feac9290a779ca8787.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 828, + 208, + 903, + 267 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/89bfce9f9f4e0b8fc38173cc8d81f9cefc07450ae8b495c63c6e7ea4598c994d.jpg", + "image_caption": [ + "$X_0$" + ], + "image_footnote": [], + "bbox": [ + 828, + 267, + 903, + 325 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$t^{\\prime} = t + \\Delta t$ (where $\\Delta t > 0$ ), we first sample $X_{t'}$ from the forward degradation process (Eq. (1)). The teacher model then estimates $\\hat{X}_t^\\phi$ from its degraded counterpart $X_{t'}$ :", + "bbox": [ + 511, + 420, + 905, + 467 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {X} _ {t} ^ {\\phi} = X _ {t ^ {\\prime}} - \\Delta t \\cdot v _ {\\phi} \\left(X _ {t ^ {\\prime}}, t ^ {\\prime}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 476, + 903, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where Euler solver is used. The consistency function $f_{\\theta}(X_t, t)$ for deriving the origin is defined by:", + "bbox": [ + 511, + 503, + 905, + 535 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\theta} \\left(X _ {t}, t\\right) = X _ {t} - t \\cdot v _ {\\theta} \\left(X _ {t}, t\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 544, + 903, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which maps any degraded input $X_{t}$ at timestep $t$ to the HR image domain. The consistency distillation in SR flow learns a vector field $v_{\\theta}$ by minimizing the distance between the prediction of the network $\\hat{X}_{t^{\\prime}\\to 0}^{\\theta} = f_{\\theta}(X_{t^{\\prime}},t^{\\prime})$ and the distillation target $\\hat{X}_{t\\rightarrow 0}^{\\theta^{-}} = f_{\\theta^{-}}(\\hat{X}_{t}^{\\phi},t)$ :", + "bbox": [ + 511, + 570, + 905, + 648 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c d} = \\mathbb {E} \\left[ d \\left(f _ {\\theta} \\left(X _ {t ^ {\\prime}}, t ^ {\\prime}\\right), f _ {\\theta^ {-}} \\left(\\hat {X} _ {t} ^ {\\phi}, t\\right)\\right) \\right], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 657, + 903, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $d(\\cdot, \\cdot)$ is a distance metric, and $\\theta^{-}$ denotes the target model parameters (e.g., exponential moving average (EMA) of $\\theta$ ). This distillation loss enables the newly trained model, $f_{\\theta}$ , to generate high-quality samples in a single step.", + "bbox": [ + 511, + 685, + 905, + 746 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "HR-Regularized Consistency Learning Although the original CD objective theoretically ensures that all trajectory points map to the same origin, it does not explicitly restrict alignment between the generated samples and the corresponding HR targets. This is because velocity field approximation errors from the teacher model $v_{\\phi}$ and the target model $v_{\\theta^{-}}$ can propagate into the distillation target $f_{\\theta^{-}}(\\hat{X}_{t}^{\\phi}, t)$ , as shown in Fig. 3.", + "bbox": [ + 511, + 750, + 905, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this, we introduce an extra HR regularization into Eq. (5), which directly enforces the student's", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11758", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "predictions $\\hat{X}_{t\\rightarrow 0}^{\\theta} = f_{\\theta}(\\hat{X}_t^\\phi ,t)$ (thus distillation target $f_{\\theta^{-}}(\\hat{X}_t^\\phi ,t))$ to match the ground-truth HR images $X_0$ :", + "bbox": [ + 89, + 89, + 485, + 126 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {h r c d} = \\mathbb {E} \\left[ d \\left(f _ {\\theta} \\left(X _ {t ^ {\\prime}}, t ^ {\\prime}\\right), f _ {\\theta^ {-}} \\left(\\hat {X} _ {t} ^ {\\phi}, t\\right)\\right) + d \\left(f _ {\\theta} \\left(\\hat {X} _ {t} ^ {\\phi}, t\\right), X _ {0}\\right) \\right]. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 133, + 483, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By training with real HR targets, consistency learning reduces reliance on the imperfect teacher and target model. This dual consistency objective ensures that $f_{\\theta}$ maintains both trajectory consistency and fidelity to HR data. Note that the input $(\\hat{X}_t^\\phi ,t)$ can be viewed as a perturbed version of sampled $(X_{t},t)$ from the forward process in Eq. (1), as visualized in Fig. 3. This perturbation resembles signal noise and distribution shifts, thereby improving the robustness of the trained model.", + "bbox": [ + 89, + 167, + 483, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fast-Slow Time Scheduling To further improve the efficiency and robustness of consistency learning, we propose sampling pairs of timesteps, $t$ and $t'$ , from two distinct time schedulers: a \"fast\" scheduler with relatively few timesteps (e.g., 4) and a \"slow\" scheduler with more granular timesteps (e.g., 1000). The slow scheduler, a standard practice for diffusion or flow models, provides fine-grained learning signals to keep intermediate velocity field predictions well-aligned with the SR flow objectives (Eq. (2)). In contrast, the fast scheduler emphasizes larger jumps along the trajectory, reflecting more desired one-step inference.", + "bbox": [ + 89, + 309, + 483, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During training, we randomly select one scheduler (fast or slow) to first sample $t + \\Delta t$ , then sample $t$ in its adjacent region from the other scheduler, and compute $(\\hat{X}_t^\\phi ,t)$ accordingly using Eq. (3), as illustrated in Fig. 3. This procedure allows flexible and larger jumps $\\Delta t$ through the fast sampler and relaxes the requirement for accurately estimating of $X_{t}$ from $X_{t + \\Delta t}$ by running one discretization step of a numerical ODE solver [38]. We hypothesize that introducing mild perturbations and diversity to $\\hat{X}_t^\\phi$ , deviating it from the sampled point $X_{t}$ , enhances the SR model's robustness and mitigates distribution shifts, particularly when combined with HR regularization.", + "bbox": [ + 89, + 476, + 483, + 656 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Training Consistency SR Flow", + "text_level": 1, + "bbox": [ + 89, + 666, + 361, + 683 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SR Flow in Image Space We use Stable Diffusion [31] as the backbone for our SR model, comprising a VAE to map images $X$ into latent representations $\\mathbf{x}$ , along with a UNet $\\theta$ adapted for velocity field learning. While the conditional flow matching loss in Eq. (2) can be formulated in latent space, our early experiments observed that such a latent-space objective does not consistently maintain a favorable balance between fidelity and visual quality over varying numbers of inference steps (e.g., from one to four).", + "bbox": [ + 89, + 688, + 483, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In practice, we find that loss in the image space is more effective. Specifically, during SR flow training, we first predict the velocity prediction $v_{\\theta}(\\mathrm{x}_t,t)$ in the latent space and compute the HR latent representation $\\hat{\\mathrm{x}}_{\\mathrm{HR}} = f_{\\theta}(\\mathrm{x}_t,t)$ (see Eq. (4)). Next, we decode $\\hat{\\mathrm{x}}_{\\mathrm{HR}}$ into the SR image $\\hat{X}_{\\mathrm{HR}}$", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "using the VAE's decoder. Last, we compare $\\hat{X}_{\\mathrm{HR}}$ to the ground-truth HR image $X_{\\mathrm{HR}}$ using an $l_{2}$ loss and a perceptual (LPIPS [64]) loss:", + "bbox": [ + 511, + 90, + 905, + 137 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {f l o w} = \\left\\| \\hat {X} _ {\\mathrm {H R}} - X _ {\\mathrm {H R}} \\right\\| _ {2} ^ {2} + \\lambda_ {p} \\operatorname {L P I P S} \\left(\\hat {X} _ {\\mathrm {H R}}, X _ {\\mathrm {H R}}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 151, + 906, + 170 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{p} = 2$ is set to weight the LPIPS term.", + "bbox": [ + 511, + 184, + 821, + 200 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adversarial Loss We incorporate an adversarial GAN objective to further enhance SR quality, particularly for texture synthesis. Specifically, a discriminator $\\mathcal{D}$ is trained to distinguish between real HR images and those restored by our model. We use the pre-trained diffusion model as a feature extractor in the latent space and attach several additional discriminator heads, following [33, 40, 56]. The hinge loss is adopted for training:", + "bbox": [ + 511, + 213, + 906, + 335 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {a d v} = \\mathbb {E} \\left[ \\max (0, 1 - \\mathcal {D} \\left(\\mathrm {x} _ {\\mathrm {H R}}\\right)) + \\max (0, 1 + \\mathcal {D} \\left(\\hat {\\mathrm {x}} _ {\\mathrm {H R}}\\right)) \\right]. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 348, + 905, + 380 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that the input to $\\mathcal{D}$ is the noised latent with diffusion time $t$ (omitted here for simplicity). The SR model is then trained to minimize this adversarial objective.", + "bbox": [ + 511, + 381, + 905, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image Quality Alignment Loss To further align the SR outputs with desirable image quality attributes (e.g., high-resolution, sharp, detailed), we propose an image quality alignment loss based on a text-image contrastive loss. Specifically, we use large vision-language models to generate a positive quality caption $c_{pos}$ (e.g., \"image quality is good, with clear details and vibrant colors\") from the HR image and a negative quality caption $c_{neg}$ (e.g., \"image quality is poor, blurry, and noisy\") from its LR counterpart. We then encode these captions using the text encoder $\\mathcal{E}_T$ of CLIP [29], while the SR result $\\hat{X}$ is encoded via CLIP's image encoder $\\mathcal{E}_I$ (ViT-L/14 is used). Image quality alignment loss encourages $\\hat{X}$ to be close to $c_{pos}$ and far from $c_{neg}$ :", + "bbox": [ + 511, + 440, + 906, + 638 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i q a} = - \\log \\frac {\\exp \\left(\\sin \\left(\\hat {X} , c _ {p o s}\\right)\\right)}{\\exp \\left(\\sin \\left(\\hat {X} , c _ {p o s}\\right)\\right) + \\exp \\left(\\sin \\left(\\hat {X} , c _ {n e g}\\right)\\right)}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 652, + 903, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{sim}(X,c) = \\cos (\\mathcal{E}_I(X),\\mathcal{E}_T(c))$ . Intuitively, this loss implicitly guides the network to improve perceptual fidelity and clarity by bridging the gap between human-perceived \"good\" and \"bad\" image attributes.", + "bbox": [ + 511, + 702, + 905, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Putting Things Together We first pre-train the SR flow model using the basic conditional flow loss in Eq. (7). Then we continue with the consistency SR flow training by finetuning it. The overall loss function is defined as:", + "bbox": [ + 511, + 776, + 905, + 838 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {f l o w}} + \\lambda_ {c d} \\mathcal {L} _ {h r c d} + \\lambda_ {a d v} \\mathcal {L} _ {a d v} + \\lambda_ {i q a} \\mathcal {L} _ {i q a}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 853, + 906, + 871 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{cd}$ , $\\lambda_{adv}$ , and $\\lambda_{iqa}$ are hyperparameters.", + "bbox": [ + 511, + 885, + 830, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11759", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/90ac527278377cc6439f4cb64cc6ffdb820cff736236d3f6ce33ad7db990f84d.jpg", + "table_caption": [ + "Table 1. Quantitative comparison with state-of-the-art DM-based SR methods on RealSR [1] and DRealSR [47]. The best and second-best results are highlighted in bold and underlined." + ], + "table_footnote": [], + "table_body": "
DatasetsMethods#StepsPSNR ↑SSIM ↑LPIPS ↓DISTS ↓FID ↓NIQE ↓MUSIQ ↑MANIQA ↑CLIPIQA ↑
RealSRStableSR [42]20024.700.70850.30180.2288128.515.9165.780.62210.6178
DiffBIR [19]5024.750.65670.36360.2312128.995.5364.980.62460.6463
SeeSR [49]5025.180.72160.30090.2223125.555.4169.770.64420.6612
PASD [55]2025.210.67980.33800.2260124.295.4168.750.64870.6620
ResShift [59]1526.310.74210.34600.2498141.717.2658.430.52850.5444
SinSR [45]126.280.73470.31880.2353135.936.2960.800.53850.6122
OSEDiff [48]125.150.73410.29210.2128123.495.6569.090.63260.6693
DoSSR [3]125.670.73870.33560.2741153.2511.1962.690.52430.6259
FlowSR125.540.74340.27280.2013112.605.2869.220.64860.6701
DRealSRStableSR [42]20028.030.75360.32840.2269148.986.5258.510.56010.6356
DiffBIR [19]5026.710.65710.45570.2748166.796.3161.070.59300.6395
SeeSR [49]5028.170.76910.31890.2315147.396.4064.930.60420.6804
PASD [55]2027.360.70730.37600.2531156.135.5564.870.61690.6808
ResShift [59]1528.460.76730.40060.2656172.268.1250.600.45860.5342
SinSR [45]128.360.75150.36650.2485170.576.9955.330.48840.6383
OSEDiff [48]127.920.78350.29680.2165135.306.4964.650.58990.6963
DoSSR [3]128.550.79910.33530.2801166.1912.2556.720.46230.5739
FlowSR128.500.78590.29750.2115130.306.1365.460.61720.7074
", + "bbox": [ + 91, + 121, + 906, + 382 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 392, + 223, + 409 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 89, + 417, + 297, + 434 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details Our model is based on the pretrained SD 2.1-base [31]. During training, we fine-tune only the U-Net using a LoRA [11] rank of 32. The learning rate is set to 2e-5, with a training patch size of $512 \\times 512$ and a batch size of 16. The training process consists of two stages: we firstly train the SR flow model for 10k iterations, followed by consistency learning for an additional 20k iterations. The loss weights $\\lambda_{cd}, \\lambda_{adv}, \\lambda_{iqa}$ are set to 0.1, 0.05, and 0.1, respectively.", + "bbox": [ + 89, + 440, + 483, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Data We train our model using LSDIR [17] and the first 10K face images from FFHQ [12]. The LR-HR training pairs are synthesized using the degradation pipeline of RealESRGAN [44]. To generate image quality captions for HR and LR images, we use Qwen2-VL [43]. We adopt the real-world test set of StableSR [42] for evaluation and comparison. The test sets include RealSR [1] and DRealSR [47].", + "bbox": [ + 89, + 584, + 483, + 690 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared Methods We compare our method against several state-of-the-art diffusion-based SR approaches, including multi-step StableSR [42], DiffBIR [19], SeeSR [49], PASD [55], ResShift [59], and one-step SinSR [45], OSEDiff [48], and DoSSR [3].", + "bbox": [ + 89, + 696, + 483, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics We evaluate our model using both reference-based and no-reference metrics. For fidelity, we report PSNR and SSIM (on the Y channel in YCbCr space), and for perceptual quality, we use LPIPS [64] and DISTS [4]. FID [9] is reported to compare the distribution of restored images with the ground truth. For no-reference evaluation, we use NIQE [25], MUSIQ [13], MANIQA [54], and CLIPIQA [41].", + "bbox": [ + 89, + 779, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Comparison with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 511, + 392, + 880, + 409 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Comparisons Table 1 shows that our approach achieves superior or at least competitive performance compared to recent state-of-the-art methods. Compared with multi-step diffusion-based methods such as SeeSR [49] and PASD [55], our FlowSR obtains equivalent or better no-reference metrics, such as MUSIQ, MANIQA, and CLIPIQA, and surpasses these methods in fidelity-based scores, e.g. PSNR and LPIPS, while reducing inference to a single step. Compared to efficient diffusion-based SR methods such as ResShift [59], our method demonstrates a clear advantage. Against the strong DDPM-based one-step competitor OSEDiff [48], FlowSR achieves superior performance across nearly all metrics. Overall, these results demonstrate that our flow-based SR method effectively balances fidelity and perceived quality while showing superiority over compared methods.", + "bbox": [ + 511, + 416, + 906, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Comparisons Fig. 4 shows visual comparisons of two real-world examples. Multi-step DM-based methods, such as DiffBIR and PASD, often generate rich but inaccurate textures, while one-step methods like SinSR and OSEDiff tend to produce blurry or less detailed results. In contrast, our method generates faithful SR results, effectively recovering structures such as the stone breakwater and the textures of the cloth.", + "bbox": [ + 511, + 667, + 908, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 801, + 679, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effects of SR Flow We first study the flow formulation for image super-resolution. Our SR flow is a straight path between $\\pi_0 = X_{\\mathrm{HR}}$ and $\\pi_1 = X_{\\mathrm{LR}}$ . We consider alternative formulations starting from noise, where $\\pi_1 = \\mathcal{N}(0,1)$ , as used in pre-trained T2I models, or starting from noised", + "bbox": [ + 511, + 824, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11760", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/54ca3c2cdf6d193f30d80467308a569a4733ccf75ab8c171d543842c8cb8d9c3.jpg", + "image_caption": [ + "LR Image" + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 330, + 273 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/042271f964a8bd2b41f4561931764176e55ae68a0891e43be10a2913e7c6f0c2.jpg", + "image_caption": [ + "HR" + ], + "image_footnote": [], + "bbox": [ + 331, + 88, + 446, + 176 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2335d0a9e3c20b22e292a413ea24cb7147093ad51b816d13569cf946a1786b7e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 447, + 89, + 560, + 175 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7ae6ec87e9e13671c5fbce08ae17b8a88ef065515fb7dece036c9e3937d4c532.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 560, + 90, + 674, + 175 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d3e7a7995566469f3751fd5e13b66d6fa97308963b5fa3a548ea274d28db309e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 676, + 90, + 789, + 175 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ffdcf532b63640a46e3b1e4fa965394367c0771e270acfdcf069a07cd58587b6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 790, + 90, + 903, + 175 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7a7f5580fd114f25da788d007a33db727b9e5492426ffd1eb15b384d4758a628.jpg", + "image_caption": [ + "LR Image" + ], + "image_footnote": [], + "bbox": [ + 91, + 286, + 330, + 470 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2556f83fe362a993be4e8e05fc131e3e97c2963e723ff608dd33c231645d68ab.jpg", + "image_caption": [ + "ResShift (15)" + ], + "image_footnote": [], + "bbox": [ + 331, + 186, + 446, + 272 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d8b83d480f026a3c3978bf04f45a8a228ff1814ed5b2690573e027986acc6dbd.jpg", + "image_caption": [ + "StableSR (200)", + "SinSR (1)" + ], + "image_footnote": [], + "bbox": [ + 447, + 186, + 560, + 273 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8d4acf7a89ffd5d63f9b6b453a71c2cb7bd162e6a781f1951492ef65b6d1701d.jpg", + "image_caption": [ + "DiffBIR (50)", + "OSEDiff (1)" + ], + "image_footnote": [], + "bbox": [ + 560, + 186, + 674, + 273 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/691fad6514142eff3400fd73ed44afd439a33039b84f3ad3f50ab54fc88ab68d.jpg", + "image_caption": [ + "SeeSR (50)", + "DoSSR (1)" + ], + "image_footnote": [], + "bbox": [ + 676, + 186, + 789, + 273 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/16e31545a1e4d897f2c1fac6ccebd5e1c196f90722bc9327e26486bdc8eb3ad9.jpg", + "image_caption": [ + "PASD (20)", + "FlowSR (1)" + ], + "image_footnote": [], + "bbox": [ + 790, + 186, + 903, + 273 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c2d787425c24e1cba5784055449b89c0f1c381a2372324f824b1548e6f8fc6cb.jpg", + "image_caption": [ + "HR" + ], + "image_footnote": [], + "bbox": [ + 331, + 286, + 446, + 375 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dc15a0f907d55c49679a14f74c7379fa049a309bb89fec07fa1de9dff1997bb9.jpg", + "image_caption": [ + "StableSR (200)" + ], + "image_footnote": [], + "bbox": [ + 447, + 286, + 560, + 373 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a28445e04f5a94e8668ca63da78effc189ef7e9c61b94d022716ce79f67954e3.jpg", + "image_caption": [ + "DiffBIR (50)" + ], + "image_footnote": [], + "bbox": [ + 560, + 286, + 674, + 373 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e295b648acc4829cc109a3ff5a6370d5ca565261faf9a2ef4588a3cef1cc92d0.jpg", + "image_caption": [ + "SeeSR (50)" + ], + "image_footnote": [], + "bbox": [ + 676, + 286, + 789, + 373 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d0d7bb8ee3403385ef0c9915a543dbb8f5ffc51d52ecae11c8059975e27c9ff6.jpg", + "image_caption": [ + "PASD (20)" + ], + "image_footnote": [], + "bbox": [ + 790, + 286, + 903, + 373 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/090d29280d1f2335183e731a9ddc23d15db14a0cae759e3b8a535e58f6cea11e.jpg", + "image_caption": [ + "Figure 4. Visual comparisons of different SR methods on real-world examples. The number of sampling steps are indicated in parentheses. Please zoom in for a better view." + ], + "image_footnote": [], + "bbox": [ + 93, + 523, + 189, + 598 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ccc115ca9f373cd25e40fd871f8ad73df503f27d3221ce3d8bed5f35a00f147a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 191, + 523, + 285, + 598 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2ca2b5d3d20e3cc93e4197ce3c50626f7de236eee48d730b394483801f1fc3ac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 523, + 383, + 598 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bb1236b864ecc899729334b5e6d586f2cc471b7adfafe3a9a44e358954face1a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 523, + 480, + 598 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f3ec8a1aa5d57dac5a0b60ac2f4842d085d8e77e096aac29dc8c8545b095e662.jpg", + "image_caption": [ + "LR Image", + "Zoomed LR", + "Figure 5. Ablation of SR flow. Our SR flow model produces more accurate SR results with improved structure with more steps." + ], + "image_footnote": [], + "bbox": [ + 94, + 611, + 189, + 685 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a0bc5dea370fb5836135c149f48b0e2030a0629a96662408f1b49d4a3840fc29.jpg", + "image_caption": [ + "$\\mathrm{Noise}\\rightarrow \\mathrm{HR}$ (4)", + "$\\mathrm{Noise}\\rightarrow \\mathrm{HR}$ (1)" + ], + "image_footnote": [], + "bbox": [ + 191, + 611, + 285, + 685 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/74e8bb9ff6cb2de759e636de029967c0d69683835d28d475909dfa4072001db5.jpg", + "image_caption": [ + "noised $\\mathrm{LR}\\rightarrow \\mathrm{HR}$ (4)", + "noised $\\mathrm{LR}\\rightarrow \\mathrm{HR}$ 1" + ], + "image_footnote": [], + "bbox": [ + 287, + 611, + 383, + 685 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7f0998d4de7ad064534623635986ef9f1dcfa5b2a3cee1255d1361d1e7dc7ca8.jpg", + "image_caption": [ + "SR Flow (4)", + "SR Flow (1)" + ], + "image_footnote": [], + "bbox": [ + 385, + 611, + 480, + 685 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "LR, where $\\pi_1 = X_{\\mathrm{LR}} + \\mu \\epsilon$ with $\\epsilon \\sim \\mathcal{N}(0,1)$ , as proposed in ResShift [59]. For these alternatives, we concatenate $X_{t}$ with $X_{\\mathrm{LR}}$ as the input to provide the LR condition. The trained models are evaluated on the RealSR [1] and DRealSR [47] datasets, as shown in Table 2. We observe that with more inference steps, the image quality (MUSIQ) generally improves, while the fidelity (PSNR) slightly decreases. Fig. 5 presents the visual comparisons, where our SR flow model produces more accurate predictions, with further structure and overall quality enhancement as addi", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0908201fbf7810f12ffaaee344a2fbe1c83112325476546fde6fda96d9698f14.jpg", + "table_caption": [ + "Table 2. Ablation of SR flow. SR flow mapping LR to HR outperforms alternative flow formulations that start from noise or ResShift [59]-style noised LR on RealSR and DRealSR." + ], + "table_footnote": [], + "table_body": "
#StepsMethodRealSRDRealSR
PSNR ↑MUSIQ ↑PSNR ↑MUSIQ ↑
4Noise→HR23.8366.0126.7562.76
noised LR→HR24.4765.1527.7461.76
SR Flow (ours)25.0068.0927.9865.05
1Noise→HR24.8265.1728.0759.34
noised LR→HR24.6765.2528.0360.64
SR Flow (ours)25.5167.4028.6562.53
", + "bbox": [ + 514, + 571, + 906, + 683 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tional sampling steps (i.e., 4 steps) are applied. Notably, these results demonstrate that the model trained using our SR flow clearly outperforms other transition variants in both single-step and multi-step inference settings.", + "bbox": [ + 511, + 696, + 906, + 757 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effects of consistency learning Next, we fine-tune the pre-trained SR flow model using the consistency objective described in Sec. 4.2. As shown in Table 3, enforcing self-consistency $\\mathcal{L}_{cd}$ across neighboring time steps stabilizes intermediate representations and often reduces LPIPS. However, when the teacher's predictions deviate from the true HR manifold, these errors may propagate to the student model, leading to diminished IQA scores. We address this by introducing the HR regularization term in Eq. (6), which", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11761", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7e6582e6f7b95003933045da3b103124ac87bbaab71469fc16e0ec12e894b6ad.jpg", + "image_caption": [ + "LR Image" + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 189, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/595550517b40f3904fd2e631ea29f4c27835dcfb6e430484035ab7510409ef81.jpg", + "image_caption": [ + "SR Flow" + ], + "image_footnote": [], + "bbox": [ + 191, + 88, + 287, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e60e498a9a0c3ab24db64f162977c46ba25056f10c7c04433d8c8417136ac7fc.jpg", + "image_caption": [ + "w/ $\\mathcal{L}_{cd}$" + ], + "image_footnote": [], + "bbox": [ + 287, + 88, + 383, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/67a8ea22fe8df4e90e5ea80f6cd078fbd28bbe51347f61ccc630a60fdc348873.jpg", + "image_caption": [ + "w/ $\\mathcal{L}_{hrcd}$", + "Figure 6. Ablation of consistency learning. Our HR-regularized consistency learning effectively reduces distortions in SR outputs while producing high-quality and sharper results (zoom in)." + ], + "image_footnote": [], + "bbox": [ + 385, + 88, + 480, + 234 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8f6410ee2e1d4add35fb48a4fdefe0a35b594908d3ec5cc21625b11accabf735.jpg", + "table_caption": [ + "Table 3. Ablation of consistency learning. Fine-tuned SR flow models with consistency objectives are evaluated on DRealSR." + ], + "table_footnote": [], + "table_body": "
#StepsMethodPSNR ↑LPIPS ↓MUSIQ ↑MANIQA ↑CLIPQA ↑
4SR Flow27.980.276365.050.62960.6681
w/ Lcd28.640.265461.820.60420.6180
w/ Lhrcd28.440.279965.710.63740.6709
1SR Flow28.650.282162.530.58950.6545
w/ Lcd28.590.260961.690.59490.6391
w/ Lhr28.850.280562.460.58710.6589
w/ Lhrcd28.620.283064.770.61560.6748
", + "bbox": [ + 91, + 337, + 483, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "directly aligns student predictions $f_{\\theta}(\\hat{X}_t^\\phi, t)$ with real HR images, mitigating the teacher's approximation errors. This extra ground-truth constraint helps the model recover fine textural details, which boosts IQA metrics (e.g., MUSIQ and MANIQA) while preserves perceptual fidelity. Consistency learning enhances the robustness of flow-based inference by distilling multi-step capabilities into a single step, allowing for faster sampling without sacrificing quality. Additionally, HR regularization alone (w/ $\\mathcal{L}_{hr}$ ) provides minimal IQA gains. This highlights the importance of consistency learning and its synergy with HR regularization for acceleration in SR flow. Fig. 6 presents visual comparisons, showing that consistency learning helps resolve distorted textures in the baseline SR flow model. Notably, HR-regularized consistency learning generates more realistic and sharper results compared to its baseline.", + "bbox": [ + 88, + 449, + 482, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis of Fast-Slow Time Sampling We evaluate the effectiveness of fast-slow time sampling by comparing it to the $N$ -interval scheduler [38] and by varying the number of fast-scheduler timesteps, while fixing the slow scheduler fixed at 1000 timesteps. When the fast scheduler uses only one timestep, we set $t + \\Delta t = 1$ and then sample $t$ from the slow scheduler (see Eq. (3)). As shown in Table 4, increasing $\\Delta t$ (thus reducing the number of intervals) within a reasonable range improves IQA metrics. A plausible explanation is that the slightly larger perturbations introduced at intermediate predictions $\\hat{X}_t^\\phi$ from $X_{t}$ encourage the model to be more robust to distribution shifts; see Fig. 3. This, in turn, reduces variance in its estimates of", + "bbox": [ + 89, + 703, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/59329c91986bef0bac8120c2cafb9aeb6f6b964a5bfc1a3082f28ab5456606aa.jpg", + "image_caption": [ + "LR Image" + ], + "image_footnote": [], + "bbox": [ + 516, + 87, + 591, + 148 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/780bcaa054749caf97071e120c0425b2e0cd49781c665e6aaa0aae205ee8932e.jpg", + "image_caption": [ + "w/o $\\mathcal{L}_{hrcd}$", + "Figure 7. Ablation of training loss." + ], + "image_footnote": [], + "bbox": [ + 593, + 88, + 669, + 148 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/81cb7bb39ac06e52da70564e0982e7baa84256e20ab80b984a683471f6a04282.jpg", + "image_caption": [ + "w/o $\\mathcal{L}_{adv}$" + ], + "image_footnote": [], + "bbox": [ + 671, + 88, + 750, + 148 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eaa2d2c2b0495c668cf91ce3d829cc23eb13afe83f9c0838f74b81af34fb7ffc.jpg", + "image_caption": [ + "w/o $\\mathcal{L}_{iqa}$" + ], + "image_footnote": [], + "bbox": [ + 750, + 88, + 826, + 148 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/de9f49c534a1ad7105b677ceb26274e9349c8f5b32862bb613657d2966bb3ea7.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 828, + 88, + 903, + 148 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a8a6833ad93af13a8fbe017ee77f82872c5a36a0f2d9aadd329607078a7a7afe.jpg", + "table_caption": [ + "Table 4. Ablation of fast-slow time sampling. We compare our fast-slow sampling approach with the $N$ -interval method on DRealSR. The number of $N$ and fast-scheduler timesteps are indicated in parentheses." + ], + "table_footnote": [], + "table_body": "
MethodPSNR ↑LPIPS ↓MUSIQ ↑MANIQA ↑CLIPiQA ↑
N-Interval (50)28.610.280762.480.59140.6599
N-Interval (18)28.740.279262.900.59720.6605
N-Interval (4)28.660.281863.030.59720.6617
Slow Only (1000)28.780.280162.720.59160.6586
Fast-Slow (8)28.540.276064.010.60980.6634
Fast-Slow (4)28.620.283064.770.61560.6748
Fast-Slow (1)28.790.283563.010.59480.6605
", + "bbox": [ + 514, + 250, + 903, + 344 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c161b7dde2254ee6deb1ba68b28d1481f3d196f82409e58ba6cbb812fa663cfb.jpg", + "table_caption": [ + "Table 5. Ablation of training loss on DRealSR." + ], + "table_footnote": [], + "table_body": "
MethodPSNR ↑LPIPS ↓MUSIQ ↑MANIQA ↑CLIPiQA ↑
w/o Lhred28.520.301364.420.60610.6954
w/o Ladv28.450.296165.290.61540.7099
w/o Liaq28.640.295864.830.61110.7012
Ours28.500.297565.460.61720.7074
", + "bbox": [ + 514, + 371, + 903, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the ODE trajectory's starting point, i.e., HR, resulting in sharper restorations. Moreover, compared to the $N$ -interval strategy or slow-only scheduling, our fast-slow scheduler further boosts IQA metrics while maintaining comparable fidelity. This is because the slow scheduler ensures fine-grained SR flow estimation, whereas the fast scheduler enables efficient sampling. Based on these findings, we set the fast scheduler to 4 timesteps throughout our experiments.", + "bbox": [ + 511, + 455, + 906, + 577 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Influence of loss function In addition to flow and consistency learning, we incorporate a GAN loss and an image quality alignment loss to further enhance SR performance, as detailed in Eq. (10). As demonstrated in Table 5 and Fig. 7, these losses all improve no-reference metrics and visual quality. The ablation highlights that eliminating any of these losses results in reduced SR quality, with consistency learning contributing a substantial improvement.", + "bbox": [ + 511, + 585, + 908, + 708 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 723, + 633, + 738 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presents FlowSR, a new approach for efficient one-step image super-resolution. FlowSR reformulates SR as a rectified flow, leveraging the strengths of iterative generative modeling. To enable high-quality single-step inference, we incorporate consistency learning and devise HR regularization to address its distillation target drifting issue. Additionally, a fast-slow time scheduling strategy is designed to enhance the efficiency and robustness of the consistency SR flow model. FlowSR contributes to the advancement of efficient real-world SR applications.", + "bbox": [ + 511, + 750, + 906, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11762", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The work described in this paper was supported in part by the National Key R&D Program of China (Grant No. 2023YFE0202700), Research Grants Council of the Hong Kong Special Administrative Region, China, under Project 14200824; and by the Hong Kong Innovation and Technology Fund, under Project MHP/092/22.", + "bbox": [ + 89, + 114, + 483, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 220, + 187, + 236 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang. Toward real-world single image super-resolution: A new benchmark and a new model. In ICCV, 2019. 6, 7", + "[2] Bin Chen, Gehui Li, Rongyuan Wu, Xindong Zhang, Jie Chen, Jian Zhang, and Lei Zhang. Adversarial diffusion compression for real-world image super-resolution. In CVPR, 2025. 2", + "[3] Qinpeng Cui, Yixuan Liu, Xinyi Zhang, Qiqi Bao, Zhongdao Wang, Qingmin Liao, Li Wang, Tian Lu, and Emad Barsoum. Taming diffusion prior for image super-resolution with domain shift sdes. NeurIPS, 2024. 1, 3, 4, 6", + "[4] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. TPAMI, 2020. 6", + "[5] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image super-resolution. In ECCV, 2014. 2", + "[6] Linwei Dong, Qingnan Fan, Yihong Guo, Zhonghao Wang, Qi Zhang, Jinwei Chen, Yawei Luo, and Changqing Zou. Tsd-sr: One-step diffusion with target score distillation for real-world image super-resolution. In CVPR, 2025. 2", + "[7] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. NeurIPS, 2014. 2", + "[8] Xiao He, Huaao Tang, Zhijun Tu, Junchao Zhang, Kun Cheng, Hanting Chen, Yong Guo, Mingrui Zhu, Nannan Wang, Xinbo Gao, et al. One step diffusion-based super-resolution with time-aware distillation. arXiv preprint arXiv:2408.07476, 2024. 2", + "[9] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. *NeruIPS*, 2017. 6", + "[10] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. *NeruIPS*, 2020. 1, 2", + "[11] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 6", + "[12] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 6", + "[13] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In ICCV, 2021. 6" + ], + "bbox": [ + 93, + 244, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Dongjun Kim, Chieh-Hsin Lai, Wei-Hsiang Liao, Naoki Murata, Yuhta Takida, Toshimitsu Uesaka, Yutong He, Yuki Mitsufuji, and Stefano Ermon. Consistency trajectory models: Learning probability flow ode trajectory of diffusion. In ICLR, 2024. 3", + "[15] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In CVPR, 2017. 2", + "[16] Jianze Li, Jiezhang Cao, Zichen Zou, Xiongfei Su, Xin Yuan, Yulun Zhang, Yong Guo, and Xiaokang Yang. Distillation-free one-step diffusion for real-world image super-resolution. arXiv preprint arXiv:2410.04224, 2024. 2", + "[17] Yawei Li, Kai Zhang, Jingyun Liang, Jiezhang Cao, Ce Liu, Rui Gong, Yulun Zhang, Hao Tang, Yun Liu, Denis Demandolx, et al. Lsdir: A large scale dataset for image restoration. In CVPR, 2023. 6", + "[18] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swin transformer. In ICCV Workshop, 2021. 2", + "[19] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Yu Qiao, Wanli Ouyang, and Chao Dong. Diffir: Toward blind image restoration with generative diffusion prior. In ECCV, 2024. 2, 6", + "[20] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. In ICLR, 2023. 4", + "[21] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. In ICLR, 2023. 2, 3, 4", + "[22] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, et al. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. In ICLR, 2024. 3", + "[23] Eric Luhman and Troy Luhman. Knowledge distillation in iterative generative models for improved sampling speed. arXiv preprint arXiv:2101.02388, 2021. 2", + "[24] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 3", + "[25] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal processing letters, 2012. 6", + "[26] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In AAAI, 2024. 2", + "[27] Mehdi Noroozi, Isma Hadji, Brais Martinez, Adrian Bulat, and Georgios Tzimiropoulos. You only need one step: Fast super-resolution with stable diffusion via scale distillation. In ECCV, 2024. 2", + "[28] Long Peng, Wenbo Li, Renjing Pei, Jingjing Ren, Jiaqi Xu, Yang Wang, Yang Cao, and Zheng-Jun Zha. Towards realistic data generation for real-world super-resolution. In ICLR, 2025. 2" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11763", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 5", + "[30] Jingjing Ren, Wenbo Li, Haoyu Chen, Renjing Pei, Bin Shao, Yong Guo, Long Peng, Fenglong Song, and Lei Zhu. Ultrapixel: Advancing ultra high-resolution image synthesis to new peaks. NeurIPS, 2024. 2", + "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2, 5, 6", + "[32] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. TPAMI, 2022. 2", + "[33] Axel Sauer, Frederic Boesel, Tim Dockhorn, Andreas Blattmann, Patrick Esser, and Robin Rombach. Fast high-resolution image synthesis with latent adversarial diffusion distillation. In SIGGRAPH Asia, 2024. 5", + "[34] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In ECCV, 2024. 2", + "[35] Johannes Schusterbauer, Ming Gui, Pingchuan Ma, Nick Stracke, Stefan Andreas Baumann, Vincent Tao Hu, and Björn Ommer. Boosting latent diffusion with flow matching. In ECCV, 2024. 3, 4", + "[36] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 1", + "[37] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 1, 2", + "[38] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In ICML, 2023. 2, 3, 5, 8", + "[39] Haoze Sun, Wenbo Li, Jianzhuang Liu, Haoyu Chen, Renjing Pei, Xueyi Zou, Youliang Yan, and Yujiu Yang. Coser: Bridging image and language for cognitive super-resolution. In CVPR, 2024. 2", + "[40] Fu-Yun Wang, Zhaoyang Huang, Alexander Bergman, Dazhong Shen, Peng Gao, Michael Lingelbach, Keqiang Sun, Weikang Bian, Guanglu Song, Yu Liu, et al. Phased consistency models. NeurIPS, 2024. 3, 5", + "[41] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In AAAI, 2023. 6", + "[42] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. IJCV, 2024. 1, 2, 6", + "[43] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 6", + "[44] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In ICCV Workshop, 2021. 2, 6" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[45] Yufei Wang, Wenhan Yang, Xinyuan Chen, Yaohui Wang, Lanqing Guo, Lap-Pui Chau, Ziwei Liu, Yu Qiao, Alex C Kot, and Bihan Wen. Sinsr: diffusion-based image superresolution in a single step. In CVPR, 2024. 1, 2, 6", + "[46] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. NeurIPS, 2023. 2", + "[47] Pengxu Wei, Ziwei Xie, Hannan Lu, Zongyuan Zhan, Qixiang Ye, Wangmeng Zuo, and Liang Lin. Component divide-and-conquer for real-world image super-resolution. In ECCV, 2020. 6, 7", + "[48] Rongyuan Wu, Lingchen Sun, Zhiyuan Ma, and Lei Zhang. One-step effective diffusion network for real-world image super-resolution. NeurIPS, 2024. 1, 2, 6", + "[49] Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. Seesr: Towards semantics-aware real-world image super-resolution. In CVPR, 2024. 1, 2, 6", + "[50] Zhisheng Xiao, Karsten Kreis, and Arash Vahdat. Tackling the generative learning trilemma with denoising diffusion gans. In ICLR, 2022. 2", + "[51] Rui Xie, Ying Tai, Chen Zhao, Kai Zhang, Zhenyu Zhang, Jun Zhou, Xiaogian Ye, Qian Wang, and Jian Yang. Addsr: Accelerating diffusion-based blind superresolution with adversarial diffusion distillation. arXiv preprint arXiv:2404.01717, 2024. 2", + "[52] Hanshu Yan, Xingchao Liu, Jiachun Pan, Jun Hao Liew, Qiang Liu, and Jiashi Feng. Perflow: Piecewise rectified flow as universal plug-and-play accelerator. NeurIPS, 2024. 3", + "[53] Ling Yang, Zixiang Zhang, Zhilong Zhang, Xingchao Liu, Minkai Xu, Wentao Zhang, Chenlin Meng, Stefano Ermon, and Bin Cui. Consistency flow matching: Defining straight flows with velocity consistency. arXiv preprint arXiv:2407.02398, 2024. 3", + "[54] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Maniaq: Multi-dimension attention network for no-reference image quality assessment. In CVPR, 2022. 6", + "[55] Tao Yang, Rongyuan Wu, Peiran Ren, Xuansong Xie, and Lei Zhang. Pixel-aware stable diffusion for realistic image super-resolution and personalized stylization. In ECCV, 2024.6", + "[56] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and Bill Freeman. Improved distribution matching distillation for fast image synthesis. NeurIPS, 2024. 5", + "[57] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In CVPR, 2024. 2", + "[58] Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. Scaling up to excellence: Practicing model scaling for photorealistic image restoration in the wild. In CVPR, 2024. 1, 2", + "[59] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Reshift: Efficient diffusion model for image super" + ], + "bbox": [ + 514, + 90, + 905, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11764", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "resolution by residual shifting. NeurIPS, 2023. 1, 2, 4, 6, 7", + "[60] Zongsheng Yue, Kang Liao, and Chen Change Loy. Arbitrary-steps image super-resolution via diffusion inversion. In CVPR, 2025. 2", + "[61] Aiping Zhang, Zongsheng Yue, Renjing Pei, Wenqi Ren, and Xiaochun Cao. Degradation-guided one-step image super-resolution with diffusion priors. arXiv preprint arXiv:2409.17058, 2024. 1, 2", + "[62] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In ICCV, 2021. 2", + "[63] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2", + "[64] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5, 6" + ], + "bbox": [ + 91, + 90, + 482, + 347 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11765", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Fast Image Super-Resolution via Consistency Rectified Flow/5e42c446-52f8-4769-b738-a2b096f3a7a5_model.json b/2025/Fast Image Super-Resolution via Consistency Rectified Flow/5e42c446-52f8-4769-b738-a2b096f3a7a5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2f6be962f7ef7ae7e2dd7e4c7f35c6307f7757f3 --- /dev/null +++ b/2025/Fast Image Super-Resolution via Consistency Rectified Flow/5e42c446-52f8-4769-b738-a2b096f3a7a5_model.json @@ -0,0 +1,3247 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.196, + 0.131, + 0.804, + 0.154 + ], + "angle": 0, + "content": "Fast Image Super-Resolution via Consistency Rectified Flow" + }, + { + "type": "text", + "bbox": [ + 0.143, + 0.18, + 0.854, + 0.217 + ], + "angle": 0, + "content": "Jiaqi \\(\\mathrm{Xu}^{1,2*}\\), Wenbo \\(\\mathrm{Li}^{2\\dagger}\\), Haoze \\(\\mathrm{Sun}^2\\), Fan \\(\\mathrm{Li}^2\\), Zhixin \\(\\mathrm{Wang}^2\\), Long \\(\\mathrm{Peng}^2\\), Jingjing \\(\\mathrm{Ren}^3\\), Haoran \\(\\mathrm{Yang}^1\\), Xiaowei \\(\\mathrm{Hu}^4\\), Renjing \\(\\mathrm{Pei}^{2\\dagger}\\), and Pheng-Ann \\(\\mathrm{Heng}^1\\)" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.22, + 0.758, + 0.256 + ], + "angle": 0, + "content": "1The Chinese University of Hong Kong 2Huawei Noah's Ark Lab 3HKUST (GZ) 4South China University of Technology" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.29, + 0.327, + 0.307 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.323, + 0.485, + 0.687 + ], + "angle": 0, + "content": "Diffusion models (DMs) have demonstrated remarkable success in real-world image super-resolution (SR), yet their reliance on time-consuming multi-step sampling largely hinders their practical applications. While recent efforts have introduced few- or single-step solutions, existing methods either inefficiently model the process from noisy input or fail to fully exploit iterative generative priors, compromising the fidelity and quality of the reconstructed images. To address this issue, we propose FlowSR, a novel approach that reformulates the SR problem as a rectified flow from low-resolution (LR) to high-resolution (HR) images. Our method leverages an improved consistency learning strategy to enable high-quality SR in a single step. Specifically, we refine the original consistency distillation process by incorporating HR regularization, ensuring that the learned SR flow not only enforces self-consistency but also converges precisely to the ground-truth HR target. Furthermore, we introduce a fast-slow scheduling strategy, where adjacent timesteps for consistency learning are sampled from two distinct schedulers: a fast scheduler with fewer timesteps to improve efficiency, and a slow scheduler with more timesteps to capture fine-grained texture details. Extensive experiments demonstrate that FlowSR achieves outstanding performance in both efficiency and image quality." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.717, + 0.222, + 0.733 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.742, + 0.484, + 0.864 + ], + "angle": 0, + "content": "Real-world image super-resolution (SR) aims to reconstruct high-resolution (HR) images from their low-resolution (LR) counterparts while simultaneously removing unknown degradations. With the remarkable success of diffusion models (DMs) [10, 37], SR methods leveraging these models—particularly those built on powerful text-to-image (T2I) models like Stable Diffusion (SD) [31]—have demonstrated outstanding performance [42, 49, 58]. However," + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.289, + 0.616, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.42, + 0.571, + 0.432 + ], + "angle": 0, + "content": "LR" + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.29, + 0.709, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.364, + 0.708, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.645, + 0.42, + 0.68, + 0.433 + ], + "angle": 0, + "content": "1-step" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.29, + 0.804, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.364, + 0.803, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.738, + 0.42, + 0.775, + 0.433 + ], + "angle": 0, + "content": "4-step" + }, + { + "type": "image", + "bbox": [ + 0.805, + 0.29, + 0.899, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.805, + 0.364, + 0.899, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.806, + 0.42, + 0.896, + 0.433 + ], + "angle": 0, + "content": "distilled, 1-step" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.442, + 0.9, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.556, + 0.907, + 0.627 + ], + "angle": 0, + "content": "Figure 1. Our consistency SR flow model achieves high-quality single-step inference (distilled, 1-step) by distilling from multi-step higher-quality sampling process (top). We achieve this by formulating SR as a rectified flow that bridges LR and HR images, combined with improved consistency learning (bottom)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.907, + 0.734 + ], + "angle": 0, + "content": "these diffusion-based SR approaches require an iterative reverse sampling process that gradually refines noisy inputs into HR outputs. Despite advancements such as efficient ordinary differential equation (ODE) solvers like DDIM [36], the slow inference speed remains a bottleneck, limiting their practicality in real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Recently, few-step or single-step SR methods derived from diffusion models are designed [3, 45, 48, 59, 61]. Several studies focus on designing more efficient diffusion processes for SR. ResShift [59] speeds up diffusion by shifting the LR-HR residual into the Markov chain and reduces sampling to 15 steps, while DoSSR [3] introduces a diffusion process more compatible with pre-trained DMs. However, they rely on DDPM [10], which uses a more curved transition trajectory and starts with a noise-perturbed LR image that loses crucial information, ultimately limiting performance. In parallel, another line of work tar" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.336, + 0.888 + ], + "angle": 0, + "content": "*Part of work done as an intern at Huawei." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.336, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11755" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.213 + ], + "angle": 0, + "content": "gets single-step SR by learning the LR-HR mapping directly. For example, SinSR [45] learns one-step SR prediction by leveraging the teacher's output from ResShift [59] and ground-truth HR as target, whereas OSEDiff adapts pre-trained SD for SR and refines one-step prediction with the VSD loss [46]. However, these methods fail to fully harness the advantages of iterative generative modeling, which naturally facilitates high-quality texture synthesis." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.215, + 0.483, + 0.41 + ], + "angle": 0, + "content": "In this work, we present FlowSR, which unifies rectified flow with consistency models (CMs) to enable efficient, single-step image super-resolution. As illustrated in Fig. 1, we reformulate SR as a rectified flow [21], which establishes a simple and straight ODE-based mapping between LR and HR images. By explicitly modeling this trajectory, FlowSR learns fine-grained LR-to-HR transformations for better SR quality and facilitates fast sampling. Building upon this formulation, we further leverage consistency models [38] to enhance single-step inference. Enforcing consistency across points on the same SR flow trajectory distills multi-step higher-quality restoration into fewer steps that reach to the same HR result." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.413, + 0.483, + 0.746 + ], + "angle": 0, + "content": "However, the naive consistency distillation (CD) objective in CMs is suboptimal for SR flow, where the SR task demands both high quality and high fidelity. While CD enforces self-consistency across discretization steps, there is no guarantee that the final distillation target aligns well with the ground-truth HR. To address this, we propose HR-regularized consistency learning, which explicitly requires the model's predictions to match real HR images. This additional constraint mitigates teacher-induced errors in the distillation target and enhances the reconstruction of fine-grained details. To further improve efficiency and robustness of the consistency SR flow model, we introduce a fast-slow time scheduling strategy. Rather than sampling distillation timestep pairs from the two boundaries of a discretized interval, we sample adjacent timesteps from distinct \"fast\" and \"slow\" schedulers. The fast scheduler uses fewer timesteps to facilitate efficient inference, while the slow scheduler employs more granular steps to maintain alignment with the SR flow objectives. This mixed sampling introduces large and flexible jumps with mild perturbations into the HR regularization steps, allowing for a broader coverage of SR flow trajectories." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.748, + 0.482, + 0.853 + ], + "angle": 0, + "content": "To train our consistency SR flow model, we first finetune a pre-trained SD model to align with the SR flow objective, followed by consistency SR flow distillation. Additionally, we incorporate a GAN [7] loss and an image quality alignment loss, with the latter promoting desirable text-described attributes in the restored images, thereby improving the SR quality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.482, + 0.902 + ], + "angle": 0, + "content": "As illustrated in Fig. 1, our model leverages the proposed consistency SR flow distillation to transfer the quality improvement typically achieved with multiple sampling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.213 + ], + "angle": 0, + "content": "steps into a single step. Experiments on real-world datasets demonstrate the superiority of FlowSR. Our core contribution lie in a new paradigm to solve the one-step SR problem. First, we explore efficient flow modeling for SR and identify potential challenges when incorporating consistency models. Second, we introduce several techniques, including HR regularization in consistency learning and a fast-slow time scheduling to enhance one-step SR performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.226, + 0.655, + 0.241 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.251, + 0.735, + 0.267 + ], + "angle": 0, + "content": "2.1. Image Super-Resolution" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.273, + 0.905, + 0.44 + ], + "angle": 0, + "content": "Real-world image super-resolution (SR) aims to reconstruct high-resolution (HR) images from low-resolution (LR) inputs, tackling complex and unknown degradations like noise, blur, and compression artifacts. Since Dong et al. [5], numerous deep learning-based methods have been proposed to address the SR problem from different perspectives, including network design [18], model training [15], and degradation simulation [28, 44, 62]. Recently, diffusion models [10, 37] have achieved remarkable success in image generation [30, 31], which inspires their application to SR and demonstrate significant advancements [32, 42]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.443, + 0.905, + 0.685 + ], + "angle": 0, + "content": "Diffusion Model-Based SR Conditioned on the LR image, diffusion model-based SR iteratively denoises towards the target HR image. The LR condition can be leveraged either by input concatenation [32] or via adapters [26, 63]. StableSR [42] based on Stable Diffusion [31], uses a trainable time-aware encoder to incorporates LR. Subsequent improvements are achieved through enhanced LR conditioning in DiffBIR [19], semantics-aware prompts in SeeSR [49], reference image generation in CoSeR [39], and scaled-up training in SUPIR [58]. However, these DM-based SR approaches, starting from noise, suffer from low inference speed, typically requiring 50-200 sampling steps. In contrast, ResShift [59] constructs an efficient diffusion model by shifting the residual between HR and LR. Denoising from the noised LR reduces the sampling steps to 15 but it still suffers from limited restoration quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Single-Step / Few-Step Image SR Fast diffusion model-based SR methods require only one or a few sampling steps during inference, typically achieved through knowledge distillation [23] or adversarial training [15, 50]. SinSR [45] derives a deterministic sampling process for ResShift and reduces its number of sampling steps to one using distillation. OSEDiff [48] applies variational score distillation [46, 57] to enhance one-step restoration quality by minimizing the KL-divergence between the distribution of its generated outputs and that of the pre-trained T2I model. AddSR [51] achieves 4-step inference by tailoring adversarial diffusion distillation [34] for SR. Similarly, other recent methods, accelerate the inference by using distillation [6, 8, 27] or GAN [2, 16, 60, 61]. On the other hand, more" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11756" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.087, + 0.885, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.309, + 0.908, + 0.355 + ], + "angle": 0, + "content": "Figure 2. Overview of the training process. The consistency SR flow distills multi-step, high-quality SR capability into single-step inference, while HR regularization ensures the model converges to the high-resolution target. Times \\( t \\), \\( t' \\) are sampled alternately from fast and slow schedulers. Note that the flow loss \\( \\mathcal{L}_{flow} \\) and the consistency loss \\( \\mathcal{L}_{hrcd} \\) are computed for different samples within each batch." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.364, + 0.482, + 0.426 + ], + "angle": 0, + "content": "efficient SR diffusion processes are constructed using domain shift [3] or flow matching [35], which allow few-step sampling. Yet, they still suffer from inefficient modeling designs and suboptimal performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.434, + 0.281, + 0.451 + ], + "angle": 0, + "content": "2.2. Consistency Models" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.457, + 0.483, + 0.669 + ], + "angle": 0, + "content": "Consistency Models (CMs) [38] learn to map any point on the ODE trajectory to its origin, which enable single-step generation and allow trade-offs between quality and computation through multi-step sampling. CMs can be implemented for modern T2I models in latent space [24]. Consistency Trajectory Models (CTMs) [14] mitigate the multi-step sampling issues in CMs by learning an any-to-any mapping from initial points to final points on the ODE trajectory. The Phased Consistency Model [40] partitions the entire ODE trajectory into multiple sub-trajectory phases and ensures consistency within each phase, while PeRFlow [52] straightens the sub-trajectories using the Reflow [22] operation. Additionally, Consistency Flow Matching [53] further enforces velocity field consistency." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.231, + 0.699 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.707, + 0.238, + 0.722 + ], + "angle": 0, + "content": "3.1. Rectified Flow" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.729, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Rectified flow [21] is an ODE-based generative model that constructs a simple, straight trajectory to transform samples between two distributions: \\(\\pi_0\\) (e.g., data) and \\(\\pi_1\\) (e.g., noise). Rectified flow defines a linear interpolation path between observed samples \\(X_0\\sim \\pi_0\\) and \\(X_{1}\\sim \\pi_{1}\\) as: \\(X_{t} = tX_{1} + (1 - t)X_{0}\\) with time \\(t\\in [0,1]\\). The core idea is to learn a velocity field \\(v_{\\theta}(X_{t},t)\\), parameterized by a neural network with weights \\(\\theta\\), that matches the derivative of this trajectory. This is achieved by optimizing the following objective: \\(\\min_{\\theta}\\int_{0}^{1}\\mathbb{E}_{X_0\\sim \\pi_0,X_1\\sim \\pi_1}[\\| v_{\\theta}(X_t,t) - (X_1 - X_0)\\| ^2 ]dt,\\) which encourages the learned velocity field to align with" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.907, + 0.472 + ], + "angle": 0, + "content": "the direction of the straight path \\( X_{1} - X_{0} \\). Once trained, samples from \\( \\pi_{1} \\) can be transformed into \\( \\pi_{0} \\) by solving the ODE: \\( dX_{t} = v_{\\theta}(X_{t}, t)dt \\), \\( X_{1} \\sim \\pi_{1} \\). A key advantage of rectified flow lies in its straight trajectories. By design, the optimal velocity field corresponds to a constant-speed flow between \\( X_{0} \\) and \\( X_{1} \\), enabling efficient and stable sampling with only a few ODE steps." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.484, + 0.704, + 0.5 + ], + "angle": 0, + "content": "3.2. Consistency Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.907, + 0.751 + ], + "angle": 0, + "content": "Consistency models [38] are a class of generative models that enable high-quality sample generation with few computational steps. These models learn to map any point \\((X_{t}, t)\\) along a probability flow ODE trajectory directly to its origin \\(X_{\\epsilon}\\), where \\(\\epsilon\\) is a fixed small positive number denoting the trajectory's start. This capability is formalized through the self-consistency property: for any pair of points \\((X_{t}, t)\\) and \\((X_{t'}, t')\\) on the same ODE trajectory, the model \\(f_{\\theta}(\\cdot, t)\\) is trained to satisfy: \\(f_{\\theta}(X_{t}, t) = f_{\\theta}(X_{t}', t') = X_{\\epsilon}\\) for all \\(t, t' \\in [\\epsilon, 1]\\). To ensure consistency during training, the boundary condition \\(f_{\\theta}(X_{\\epsilon}, \\epsilon) = X_{\\epsilon}\\) is enforced. Consistency models eliminate trajectory drift by mapping all intermediate states to a consistent origin, avoiding errors from iterative denoising steps. This enables them to generate high-fidelity samples even with large step sizes, such as directly mapping from \\(t = 1\\) to the origin in a single step." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.768, + 0.65, + 0.786 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.903 + ], + "angle": 0, + "content": "We leverage the generative capability of rectified flow for high-quality SR and explore consistency models to enable fast inference with fewer sampling steps (i.e., one step), without compromising quality. We define a rectified flow, termed SR flow, for image super-resolution in Sec. 4.1. We then introduce the improved consistency SR flow learning method in Sec. 4.2 and detail the training in Sec. 4.3." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "11757" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.294, + 0.106 + ], + "angle": 0, + "content": "4.1. Rectified Flow for SR" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.485, + 0.189 + ], + "angle": 0, + "content": "SR Flow We consider Rectified Flow naturally aligns with image super-resolution, where high-resolution (HR) images \\(X_{\\mathrm{HR}} \\sim \\pi_0\\) and their low-resolution (LR) counterparts \\(X_{\\mathrm{LR}} \\sim \\pi_1\\). We define the forward process as a straight path via a linear interpolation between them:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.195, + 0.483, + 0.212 + ], + "angle": 0, + "content": "\\[\nX _ {t} = (1 - t) X _ {\\mathrm {H R}} + t X _ {\\mathrm {L R}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.217, + 0.484, + 0.308 + ], + "angle": 0, + "content": "here time \\(t\\in [0,1]\\) with \\(t = 0\\) corresponding to \\(X_{\\mathrm{HR}}\\in \\mathbb{E}^{H\\times W\\times 3}\\) and \\(t = 1\\) corresponding to \\(X_{\\mathrm{LR}}\\in \\mathbb{E}^{H\\times W\\times 3}\\), where \\(X_{\\mathrm{LR}}\\) is upscaled to match the spatial dimensions of \\(X_{\\mathrm{HR}}\\). The learning objective of the SR neural network \\(v_{\\theta}\\) is to regress the conditional vector fields [20, 21] by following the direction \\(X_{\\mathrm{LR}} - X_{\\mathrm{HR}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.314, + 0.483, + 0.332 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {t, X _ {t}} \\| v _ {\\theta} \\left(X _ {t}, t\\right) - \\left(X _ {\\mathrm {L R}} - X _ {\\mathrm {H R}}\\right) \\| _ {2} ^ {2}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.337, + 0.484, + 0.534 + ], + "angle": 0, + "content": "Intuitively, this SR flow establishes a smooth transition between HR and LR. During training, the SR flow model \\( v_{\\theta} \\) implicitly learns to invert degradations (e.g., blur, noise) and recover high-frequency details through intermediate refinements such as edge sharpening and texture synthesis. At inference, HR images are reconstructed via reverse sampling along the learned trajectory. Starting from \\( X_{\\mathrm{LR}} \\), we solve the reverse ODE using numerical methods like Euler solver: \\( X_{t'} = X_t + \\Delta t \\cdot v_{\\theta}(X_t, t) \\) from \\( t = 1 \\) to \\( t = 0 \\). Notably, unlike diffusion-based SR methods that corrupt LR inputs with noise [3, 35, 59], our transition process directly starts from LR, which preserves most of the structural information in the LR image, enabling stable and efficient sampling." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.537, + 0.484, + 0.704 + ], + "angle": 0, + "content": "Faster Inference with SR Flow A key advantage of SR flow is its flexibility in sampling steps. It supports faster inference by reducing the number of sampling steps, requiring as few as a single step that directly maps the LR image to its corresponding HR output: \\(\\hat{X}_{\\mathrm{HR}} = X_{\\mathrm{LR}} - 1 \\cdot v_{\\theta}(X_{\\mathrm{LR}}, 1)\\). This is more easily achievable than the noise-to-image mapping or initializing from noisy inputs in ResShift [59], thanks to the strong correlation between LR and HR images. Additionally, SR flow also supports any number of sampling steps and produces better visual quality when using iterative multi-step sampling, as illustrated in Fig. 1." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.71, + 0.29, + 0.725 + ], + "angle": 0, + "content": "4.2. Consistency SR Flow" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.731, + 0.484, + 0.838 + ], + "angle": 0, + "content": "We enhance single-step super-resolution in SR flow models by distilling multi-step sampling capability (e.g., four steps) into one via consistency learning (Fig. 1). To ensure accurate HR reconstruction, we incorporate the HR target explicitly into consistency learning. Additionally, a fast-slow time sampling schedule is introduced to improve efficiency and robustness. Fig. 2 shows the overall training process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Consistency Distillation in SR Flow We start with consistency distillation (CD) to distill the capability of a (pre-trained) teacher SR flow model \\( v_{\\phi}(X,t) \\), enabling high-quality SR in fewer inference steps. Given timesteps \\( t \\) and" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.209, + 0.596, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.268, + 0.595, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.544, + 0.329, + 0.564, + 0.345 + ], + "angle": 0, + "content": "\\(X_{t^{\\prime}}\\)" + }, + { + "type": "image", + "bbox": [ + 0.596, + 0.209, + 0.672, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.596, + 0.268, + 0.672, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.624, + 0.329, + 0.641, + 0.345 + ], + "angle": 0, + "content": "\\(X_{t}\\)" + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.209, + 0.75, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.268, + 0.75, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.701, + 0.329, + 0.72, + 0.345 + ], + "angle": 0, + "content": "\\(\\hat{X}_t^\\phi\\)" + }, + { + "type": "image", + "bbox": [ + 0.752, + 0.209, + 0.828, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.752, + 0.268, + 0.827, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.774, + 0.329, + 0.802, + 0.345 + ], + "angle": 0, + "content": "\\(\\hat{X}_{t\\rightarrow 0}^{\\theta^{-}}\\)" + }, + { + "type": "image", + "bbox": [ + 0.829, + 0.209, + 0.905, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.83, + 0.268, + 0.905, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.858, + 0.329, + 0.875, + 0.345 + ], + "angle": 0, + "content": "\\(X_0\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.353, + 0.906, + 0.412 + ], + "angle": 0, + "content": "Figure 3. Illustration of HR-regularized consistency learning. Approximation errors may cause the distillation target \\(\\hat{X}_{t\\rightarrow 0}^{\\theta^{-}}\\) to deviate from the true high-quality HR target. HR regularization corrects this by enforcing HR alignment under mild perturbations." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.421, + 0.906, + 0.468 + ], + "angle": 0, + "content": "\\(t^{\\prime} = t + \\Delta t\\) (where \\(\\Delta t > 0\\)), we first sample \\(X_{t'}\\) from the forward degradation process (Eq. (1)). The teacher model then estimates \\(\\hat{X}_t^\\phi\\) from its degraded counterpart \\(X_{t'}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.477, + 0.905, + 0.496 + ], + "angle": 0, + "content": "\\[\n\\hat {X} _ {t} ^ {\\phi} = X _ {t ^ {\\prime}} - \\Delta t \\cdot v _ {\\phi} \\left(X _ {t ^ {\\prime}}, t ^ {\\prime}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.505, + 0.906, + 0.536 + ], + "angle": 0, + "content": "where Euler solver is used. The consistency function \\( f_{\\theta}(X_t, t) \\) for deriving the origin is defined by:" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.545, + 0.905, + 0.562 + ], + "angle": 0, + "content": "\\[\nf _ {\\theta} \\left(X _ {t}, t\\right) = X _ {t} - t \\cdot v _ {\\theta} \\left(X _ {t}, t\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.571, + 0.906, + 0.65 + ], + "angle": 0, + "content": "which maps any degraded input \\( X_{t} \\) at timestep \\( t \\) to the HR image domain. The consistency distillation in SR flow learns a vector field \\( v_{\\theta} \\) by minimizing the distance between the prediction of the network \\( \\hat{X}_{t^{\\prime}\\to 0}^{\\theta} = f_{\\theta}(X_{t^{\\prime}},t^{\\prime}) \\) and the distillation target \\( \\hat{X}_{t\\rightarrow 0}^{\\theta^{-}} = f_{\\theta^{-}}(\\hat{X}_{t}^{\\phi},t) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.659, + 0.905, + 0.677 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c d} = \\mathbb {E} \\left[ d \\left(f _ {\\theta} \\left(X _ {t ^ {\\prime}}, t ^ {\\prime}\\right), f _ {\\theta^ {-}} \\left(\\hat {X} _ {t} ^ {\\phi}, t\\right)\\right) \\right], \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.686, + 0.906, + 0.747 + ], + "angle": 0, + "content": "where \\(d(\\cdot, \\cdot)\\) is a distance metric, and \\(\\theta^{-}\\) denotes the target model parameters (e.g., exponential moving average (EMA) of \\(\\theta\\)). This distillation loss enables the newly trained model, \\(f_{\\theta}\\), to generate high-quality samples in a single step." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.906, + 0.871 + ], + "angle": 0, + "content": "HR-Regularized Consistency Learning Although the original CD objective theoretically ensures that all trajectory points map to the same origin, it does not explicitly restrict alignment between the generated samples and the corresponding HR targets. This is because velocity field approximation errors from the teacher model \\( v_{\\phi} \\) and the target model \\( v_{\\theta^{-}} \\) can propagate into the distillation target \\( f_{\\theta^{-}}(\\hat{X}_{t}^{\\phi}, t) \\), as shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "To address this, we introduce an extra HR regularization into Eq. (5), which directly enforces the student's" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11758" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.09, + 0.486, + 0.127 + ], + "angle": 0, + "content": "predictions \\(\\hat{X}_{t\\rightarrow 0}^{\\theta} = f_{\\theta}(\\hat{X}_t^\\phi ,t)\\) (thus distillation target \\(f_{\\theta^{-}}(\\hat{X}_t^\\phi ,t))\\) to match the ground-truth HR images \\(X_0\\):" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.135, + 0.484, + 0.168 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {h r c d} = \\mathbb {E} \\left[ d \\left(f _ {\\theta} \\left(X _ {t ^ {\\prime}}, t ^ {\\prime}\\right), f _ {\\theta^ {-}} \\left(\\hat {X} _ {t} ^ {\\phi}, t\\right)\\right) + d \\left(f _ {\\theta} \\left(\\hat {X} _ {t} ^ {\\phi}, t\\right), X _ {0}\\right) \\right]. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.169, + 0.485, + 0.304 + ], + "angle": 0, + "content": "By training with real HR targets, consistency learning reduces reliance on the imperfect teacher and target model. This dual consistency objective ensures that \\( f_{\\theta} \\) maintains both trajectory consistency and fidelity to HR data. Note that the input \\( (\\hat{X}_t^\\phi ,t) \\) can be viewed as a perturbed version of sampled \\( (X_{t},t) \\) from the forward process in Eq. (1), as visualized in Fig. 3. This perturbation resembles signal noise and distribution shifts, thereby improving the robustness of the trained model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.31, + 0.484, + 0.476 + ], + "angle": 0, + "content": "Fast-Slow Time Scheduling To further improve the efficiency and robustness of consistency learning, we propose sampling pairs of timesteps, \\(t\\) and \\(t'\\), from two distinct time schedulers: a \"fast\" scheduler with relatively few timesteps (e.g., 4) and a \"slow\" scheduler with more granular timesteps (e.g., 1000). The slow scheduler, a standard practice for diffusion or flow models, provides fine-grained learning signals to keep intermediate velocity field predictions well-aligned with the SR flow objectives (Eq. (2)). In contrast, the fast scheduler emphasizes larger jumps along the trajectory, reflecting more desired one-step inference." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.477, + 0.484, + 0.657 + ], + "angle": 0, + "content": "During training, we randomly select one scheduler (fast or slow) to first sample \\( t + \\Delta t \\), then sample \\( t \\) in its adjacent region from the other scheduler, and compute \\( (\\hat{X}_t^\\phi ,t) \\) accordingly using Eq. (3), as illustrated in Fig. 3. This procedure allows flexible and larger jumps \\( \\Delta t \\) through the fast sampler and relaxes the requirement for accurately estimating of \\( X_{t} \\) from \\( X_{t + \\Delta t} \\) by running one discretization step of a numerical ODE solver [38]. We hypothesize that introducing mild perturbations and diversity to \\( \\hat{X}_t^\\phi \\), deviating it from the sampled point \\( X_{t} \\), enhances the SR model's robustness and mitigates distribution shifts, particularly when combined with HR regularization." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.667, + 0.362, + 0.684 + ], + "angle": 0, + "content": "4.3. Training Consistency SR Flow" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.825 + ], + "angle": 0, + "content": "SR Flow in Image Space We use Stable Diffusion [31] as the backbone for our SR model, comprising a VAE to map images \\(X\\) into latent representations \\(\\mathbf{x}\\), along with a UNet \\(\\theta\\) adapted for velocity field learning. While the conditional flow matching loss in Eq. (2) can be formulated in latent space, our early experiments observed that such a latent-space objective does not consistently maintain a favorable balance between fidelity and visual quality over varying numbers of inference steps (e.g., from one to four)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In practice, we find that loss in the image space is more effective. Specifically, during SR flow training, we first predict the velocity prediction \\( v_{\\theta}(\\mathrm{x}_t,t) \\) in the latent space and compute the HR latent representation \\( \\hat{\\mathrm{x}}_{\\mathrm{HR}} = f_{\\theta}(\\mathrm{x}_t,t) \\) (see Eq. (4)). Next, we decode \\( \\hat{\\mathrm{x}}_{\\mathrm{HR}} \\) into the SR image \\( \\hat{X}_{\\mathrm{HR}} \\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.091, + 0.906, + 0.138 + ], + "angle": 0, + "content": "using the VAE's decoder. Last, we compare \\(\\hat{X}_{\\mathrm{HR}}\\) to the ground-truth HR image \\(X_{\\mathrm{HR}}\\) using an \\(l_{2}\\) loss and a perceptual (LPIPS [64]) loss:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.152, + 0.907, + 0.171 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {f l o w} = \\left\\| \\hat {X} _ {\\mathrm {H R}} - X _ {\\mathrm {H R}} \\right\\| _ {2} ^ {2} + \\lambda_ {p} \\operatorname {L P I P S} \\left(\\hat {X} _ {\\mathrm {H R}}, X _ {\\mathrm {H R}}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.185, + 0.822, + 0.201 + ], + "angle": 0, + "content": "where \\(\\lambda_{p} = 2\\) is set to weight the LPIPS term." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.214, + 0.907, + 0.336 + ], + "angle": 0, + "content": "Adversarial Loss We incorporate an adversarial GAN objective to further enhance SR quality, particularly for texture synthesis. Specifically, a discriminator \\(\\mathcal{D}\\) is trained to distinguish between real HR images and those restored by our model. We use the pre-trained diffusion model as a feature extractor in the latent space and attach several additional discriminator heads, following [33, 40, 56]. The hinge loss is adopted for training:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.349, + 0.906, + 0.381 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {a d v} = \\mathbb {E} \\left[ \\max (0, 1 - \\mathcal {D} \\left(\\mathrm {x} _ {\\mathrm {H R}}\\right)) + \\max (0, 1 + \\mathcal {D} \\left(\\hat {\\mathrm {x}} _ {\\mathrm {H R}}\\right)) \\right]. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.382, + 0.906, + 0.427 + ], + "angle": 0, + "content": "Note that the input to \\(\\mathcal{D}\\) is the noised latent with diffusion time \\(t\\) (omitted here for simplicity). The SR model is then trained to minimize this adversarial objective." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.441, + 0.907, + 0.64 + ], + "angle": 0, + "content": "Image Quality Alignment Loss To further align the SR outputs with desirable image quality attributes (e.g., high-resolution, sharp, detailed), we propose an image quality alignment loss based on a text-image contrastive loss. Specifically, we use large vision-language models to generate a positive quality caption \\( c_{pos} \\) (e.g., \"image quality is good, with clear details and vibrant colors\") from the HR image and a negative quality caption \\( c_{neg} \\) (e.g., \"image quality is poor, blurry, and noisy\") from its LR counterpart. We then encode these captions using the text encoder \\( \\mathcal{E}_T \\) of CLIP [29], while the SR result \\( \\hat{X} \\) is encoded via CLIP's image encoder \\( \\mathcal{E}_I \\) (ViT-L/14 is used). Image quality alignment loss encourages \\( \\hat{X} \\) to be close to \\( c_{pos} \\) and far from \\( c_{neg} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.654, + 0.905, + 0.702 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i q a} = - \\log \\frac {\\exp \\left(\\sin \\left(\\hat {X} , c _ {p o s}\\right)\\right)}{\\exp \\left(\\sin \\left(\\hat {X} , c _ {p o s}\\right)\\right) + \\exp \\left(\\sin \\left(\\hat {X} , c _ {n e g}\\right)\\right)}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.703, + 0.906, + 0.764 + ], + "angle": 0, + "content": "where \\(\\mathrm{sim}(X,c) = \\cos (\\mathcal{E}_I(X),\\mathcal{E}_T(c))\\). Intuitively, this loss implicitly guides the network to improve perceptual fidelity and clarity by bridging the gap between human-perceived \"good\" and \"bad\" image attributes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.777, + 0.906, + 0.839 + ], + "angle": 0, + "content": "Putting Things Together We first pre-train the SR flow model using the basic conditional flow loss in Eq. (7). Then we continue with the consistency SR flow training by finetuning it. The overall loss function is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.854, + 0.907, + 0.872 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {f l o w}} + \\lambda_ {c d} \\mathcal {L} _ {h r c d} + \\lambda_ {a d v} \\mathcal {L} _ {a d v} + \\lambda_ {i q a} \\mathcal {L} _ {i q a}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.831, + 0.902 + ], + "angle": 0, + "content": "where \\(\\lambda_{cd}\\), \\(\\lambda_{adv}\\), and \\(\\lambda_{iqa}\\) are hyperparameters." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11759" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison with state-of-the-art DM-based SR methods on RealSR [1] and DRealSR [47]. The best and second-best results are highlighted in bold and underlined." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.122, + 0.907, + 0.383 + ], + "angle": 0, + "content": "
DatasetsMethods#StepsPSNR ↑SSIM ↑LPIPS ↓DISTS ↓FID ↓NIQE ↓MUSIQ ↑MANIQA ↑CLIPIQA ↑
RealSRStableSR [42]20024.700.70850.30180.2288128.515.9165.780.62210.6178
DiffBIR [19]5024.750.65670.36360.2312128.995.5364.980.62460.6463
SeeSR [49]5025.180.72160.30090.2223125.555.4169.770.64420.6612
PASD [55]2025.210.67980.33800.2260124.295.4168.750.64870.6620
ResShift [59]1526.310.74210.34600.2498141.717.2658.430.52850.5444
SinSR [45]126.280.73470.31880.2353135.936.2960.800.53850.6122
OSEDiff [48]125.150.73410.29210.2128123.495.6569.090.63260.6693
DoSSR [3]125.670.73870.33560.2741153.2511.1962.690.52430.6259
FlowSR125.540.74340.27280.2013112.605.2869.220.64860.6701
DRealSRStableSR [42]20028.030.75360.32840.2269148.986.5258.510.56010.6356
DiffBIR [19]5026.710.65710.45570.2748166.796.3161.070.59300.6395
SeeSR [49]5028.170.76910.31890.2315147.396.4064.930.60420.6804
PASD [55]2027.360.70730.37600.2531156.135.5564.870.61690.6808
ResShift [59]1528.460.76730.40060.2656172.268.1250.600.45860.5342
SinSR [45]128.360.75150.36650.2485170.576.9955.330.48840.6383
OSEDiff [48]127.920.78350.29680.2165135.306.4964.650.58990.6963
DoSSR [3]128.550.79910.33530.2801166.1912.2556.720.46230.5739
FlowSR128.500.78590.29750.2115130.306.1365.460.61720.7074
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.393, + 0.224, + 0.41 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.418, + 0.299, + 0.435 + ], + "angle": 0, + "content": "5.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.441, + 0.484, + 0.577 + ], + "angle": 0, + "content": "Implementation Details Our model is based on the pretrained SD 2.1-base [31]. During training, we fine-tune only the U-Net using a LoRA [11] rank of 32. The learning rate is set to 2e-5, with a training patch size of \\(512 \\times 512\\) and a batch size of 16. The training process consists of two stages: we firstly train the SR flow model for 10k iterations, followed by consistency learning for an additional 20k iterations. The loss weights \\(\\lambda_{cd}, \\lambda_{adv}, \\lambda_{iqa}\\) are set to 0.1, 0.05, and 0.1, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.585, + 0.484, + 0.691 + ], + "angle": 0, + "content": "Data We train our model using LSDIR [17] and the first 10K face images from FFHQ [12]. The LR-HR training pairs are synthesized using the degradation pipeline of RealESRGAN [44]. To generate image quality captions for HR and LR images, we use Qwen2-VL [43]. We adopt the real-world test set of StableSR [42] for evaluation and comparison. The test sets include RealSR [1] and DRealSR [47]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.697, + 0.484, + 0.773 + ], + "angle": 0, + "content": "Compared Methods We compare our method against several state-of-the-art diffusion-based SR approaches, including multi-step StableSR [42], DiffBIR [19], SeeSR [49], PASD [55], ResShift [59], and one-step SinSR [45], OSEDiff [48], and DoSSR [3]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Evaluation Metrics We evaluate our model using both reference-based and no-reference metrics. For fidelity, we report PSNR and SSIM (on the Y channel in YCbCr space), and for perceptual quality, we use LPIPS [64] and DISTS [4]. FID [9] is reported to compare the distribution of restored images with the ground truth. For no-reference evaluation, we use NIQE [25], MUSIQ [13], MANIQA [54], and CLIPIQA [41]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.393, + 0.882, + 0.41 + ], + "angle": 0, + "content": "5.2. Comparison with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.417, + 0.907, + 0.659 + ], + "angle": 0, + "content": "Quantitative Comparisons Table 1 shows that our approach achieves superior or at least competitive performance compared to recent state-of-the-art methods. Compared with multi-step diffusion-based methods such as SeeSR [49] and PASD [55], our FlowSR obtains equivalent or better no-reference metrics, such as MUSIQ, MANIQA, and CLIPIQA, and surpasses these methods in fidelity-based scores, e.g. PSNR and LPIPS, while reducing inference to a single step. Compared to efficient diffusion-based SR methods such as ResShift [59], our method demonstrates a clear advantage. Against the strong DDPM-based one-step competitor OSEDiff [48], FlowSR achieves superior performance across nearly all metrics. Overall, these results demonstrate that our flow-based SR method effectively balances fidelity and perceived quality while showing superiority over compared methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.669, + 0.909, + 0.79 + ], + "angle": 0, + "content": "Qualitative Comparisons Fig. 4 shows visual comparisons of two real-world examples. Multi-step DM-based methods, such as DiffBIR and PASD, often generate rich but inaccurate textures, while one-step methods like SinSR and OSEDiff tend to produce blurry or less detailed results. In contrast, our method generates faithful SR results, effectively recovering structures such as the stone breakwater and the textures of the cloth." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.802, + 0.681, + 0.818 + ], + "angle": 0, + "content": "5.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Effects of SR Flow We first study the flow formulation for image super-resolution. Our SR flow is a straight path between \\(\\pi_0 = X_{\\mathrm{HR}}\\) and \\(\\pi_1 = X_{\\mathrm{LR}}\\). We consider alternative formulations starting from noise, where \\(\\pi_1 = \\mathcal{N}(0,1)\\), as used in pre-trained T2I models, or starting from noised" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11760" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.331, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.185, + 0.274, + 0.241, + 0.285 + ], + "angle": 0, + "content": "LR Image" + }, + { + "type": "image", + "bbox": [ + 0.333, + 0.089, + 0.447, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.38, + 0.177, + 0.4, + 0.186 + ], + "angle": 0, + "content": "HR" + }, + { + "type": "image", + "bbox": [ + 0.449, + 0.09, + 0.561, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.463, + 0.177, + 0.546, + 0.187 + ], + "angle": 0, + "content": "StableSR (200)" + }, + { + "type": "image", + "bbox": [ + 0.562, + 0.091, + 0.675, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.584, + 0.177, + 0.655, + 0.187 + ], + "angle": 0, + "content": "DiffBIR (50)" + }, + { + "type": "image", + "bbox": [ + 0.677, + 0.091, + 0.79, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.702, + 0.177, + 0.765, + 0.187 + ], + "angle": 0, + "content": "SeeSR (50)" + }, + { + "type": "image", + "bbox": [ + 0.792, + 0.091, + 0.904, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.817, + 0.177, + 0.877, + 0.187 + ], + "angle": 0, + "content": "PASD (20)" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.287, + 0.331, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.185, + 0.472, + 0.241, + 0.483 + ], + "angle": 0, + "content": "LR Image" + }, + { + "type": "image", + "bbox": [ + 0.333, + 0.187, + 0.447, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.354, + 0.274, + 0.426, + 0.285 + ], + "angle": 0, + "content": "ResShift (15)" + }, + { + "type": "image", + "bbox": [ + 0.449, + 0.187, + 0.561, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.477, + 0.275, + 0.532, + 0.285 + ], + "angle": 0, + "content": "SinSR (1)" + }, + { + "type": "image", + "bbox": [ + 0.562, + 0.187, + 0.675, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.585, + 0.275, + 0.653, + 0.285 + ], + "angle": 0, + "content": "OSEDiff (1)" + }, + { + "type": "image", + "bbox": [ + 0.677, + 0.187, + 0.79, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.702, + 0.275, + 0.763, + 0.285 + ], + "angle": 0, + "content": "DoSSR (1)" + }, + { + "type": "image", + "bbox": [ + 0.792, + 0.187, + 0.904, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.815, + 0.275, + 0.879, + 0.285 + ], + "angle": 0, + "content": "FlowSR (1)" + }, + { + "type": "image", + "bbox": [ + 0.333, + 0.287, + 0.447, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.381, + 0.376, + 0.4, + 0.384 + ], + "angle": 0, + "content": "HR" + }, + { + "type": "image", + "bbox": [ + 0.449, + 0.287, + 0.561, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.464, + 0.375, + 0.545, + 0.385 + ], + "angle": 0, + "content": "StableSR (200)" + }, + { + "type": "image", + "bbox": [ + 0.562, + 0.287, + 0.675, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.584, + 0.375, + 0.654, + 0.385 + ], + "angle": 0, + "content": "DiffBIR (50)" + }, + { + "type": "image", + "bbox": [ + 0.677, + 0.287, + 0.79, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.702, + 0.375, + 0.764, + 0.385 + ], + "angle": 0, + "content": "SeeSR (50)" + }, + { + "type": "image", + "bbox": [ + 0.792, + 0.287, + 0.904, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.817, + 0.375, + 0.877, + 0.385 + ], + "angle": 0, + "content": "PASD (20)" + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.488, + 0.904, + 0.515 + ], + "angle": 0, + "content": "Figure 4. Visual comparisons of different SR methods on real-world examples. The number of sampling steps are indicated in parentheses. Please zoom in for a better view." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.525, + 0.191, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.602, + 0.166, + 0.611 + ], + "angle": 0, + "content": "LR Image" + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.525, + 0.287, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.202, + 0.602, + 0.276, + 0.611 + ], + "angle": 0, + "content": "\\(\\mathrm{Noise}\\rightarrow \\mathrm{HR}\\) (4)" + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.525, + 0.384, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.29, + 0.602, + 0.383, + 0.611 + ], + "angle": 0, + "content": "noised \\(\\mathrm{LR}\\rightarrow \\mathrm{HR}\\) (4)" + }, + { + "type": "image", + "bbox": [ + 0.385, + 0.525, + 0.481, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.403, + 0.602, + 0.462, + 0.611 + ], + "angle": 0, + "content": "SR Flow (4)" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.612, + 0.19, + 0.686 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.688, + 0.172, + 0.697 + ], + "angle": 0, + "content": "Zoomed LR" + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.612, + 0.287, + 0.686 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.203, + 0.688, + 0.276, + 0.698 + ], + "angle": 0, + "content": "\\(\\mathrm{Noise}\\rightarrow \\mathrm{HR}\\) (1)" + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.612, + 0.384, + 0.686 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.29, + 0.688, + 0.383, + 0.698 + ], + "angle": 0, + "content": "noised \\(\\mathrm{LR}\\rightarrow \\mathrm{HR}\\) 1" + }, + { + "type": "image", + "bbox": [ + 0.386, + 0.612, + 0.481, + 0.686 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.404, + 0.688, + 0.462, + 0.698 + ], + "angle": 0, + "content": "SR Flow (1)" + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.704, + 0.483, + 0.732 + ], + "angle": 0, + "content": "Figure 5. Ablation of SR flow. Our SR flow model produces more accurate SR results with improved structure with more steps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.901 + ], + "angle": 0, + "content": "LR, where \\(\\pi_1 = X_{\\mathrm{LR}} + \\mu \\epsilon\\) with \\(\\epsilon \\sim \\mathcal{N}(0,1)\\), as proposed in ResShift [59]. For these alternatives, we concatenate \\(X_{t}\\) with \\(X_{\\mathrm{LR}}\\) as the input to provide the LR condition. The trained models are evaluated on the RealSR [1] and DRealSR [47] datasets, as shown in Table 2. We observe that with more inference steps, the image quality (MUSIQ) generally improves, while the fidelity (PSNR) slightly decreases. Fig. 5 presents the visual comparisons, where our SR flow model produces more accurate predictions, with further structure and overall quality enhancement as addi" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.519, + 0.907, + 0.561 + ], + "angle": 0, + "content": "Table 2. Ablation of SR flow. SR flow mapping LR to HR outperforms alternative flow formulations that start from noise or ResShift [59]-style noised LR on RealSR and DRealSR." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.572, + 0.907, + 0.684 + ], + "angle": 0, + "content": "
#StepsMethodRealSRDRealSR
PSNR ↑MUSIQ ↑PSNR ↑MUSIQ ↑
4Noise→HR23.8366.0126.7562.76
noised LR→HR24.4765.1527.7461.76
SR Flow (ours)25.0068.0927.9865.05
1Noise→HR24.8265.1728.0759.34
noised LR→HR24.6765.2528.0360.64
SR Flow (ours)25.5167.4028.6562.53
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.698, + 0.907, + 0.758 + ], + "angle": 0, + "content": "tional sampling steps (i.e., 4 steps) are applied. Notably, these results demonstrate that the model trained using our SR flow clearly outperforms other transition variants in both single-step and multi-step inference settings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Effects of consistency learning Next, we fine-tune the pre-trained SR flow model using the consistency objective described in Sec. 4.2. As shown in Table 3, enforcing self-consistency \\(\\mathcal{L}_{cd}\\) across neighboring time steps stabilizes intermediate representations and often reduces LPIPS. However, when the teacher's predictions deviate from the true HR manifold, these errors may propagate to the student model, leading to diminished IQA scores. We address this by introducing the HR regularization term in Eq. (6), which" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11761" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.19, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.238, + 0.166, + 0.248 + ], + "angle": 0, + "content": "LR Image" + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.089, + 0.288, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.238, + 0.261, + 0.248 + ], + "angle": 0, + "content": "SR Flow" + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.089, + 0.385, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.318, + 0.238, + 0.354, + 0.248 + ], + "angle": 0, + "content": "w/ \\(\\mathcal{L}_{cd}\\)" + }, + { + "type": "image", + "bbox": [ + 0.386, + 0.089, + 0.482, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.407, + 0.238, + 0.458, + 0.248 + ], + "angle": 0, + "content": "w/ \\(\\mathcal{L}_{hrcd}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.253, + 0.483, + 0.295 + ], + "angle": 0, + "content": "Figure 6. Ablation of consistency learning. Our HR-regularized consistency learning effectively reduces distortions in SR outputs while producing high-quality and sharper results (zoom in)." + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.305, + 0.483, + 0.334 + ], + "angle": 0, + "content": "Table 3. Ablation of consistency learning. Fine-tuned SR flow models with consistency objectives are evaluated on DRealSR." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.338, + 0.485, + 0.438 + ], + "angle": 0, + "content": "
#StepsMethodPSNR ↑LPIPS ↓MUSIQ ↑MANIQA ↑CLIPQA ↑
4SR Flow27.980.276365.050.62960.6681
w/ Lcd28.640.265461.820.60420.6180
w/ Lhrcd28.440.279965.710.63740.6709
1SR Flow28.650.282162.530.58950.6545
w/ Lcd28.590.260961.690.59490.6391
w/ Lhr28.850.280562.460.58710.6589
w/ Lhrcd28.620.283064.770.61560.6748
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.45, + 0.483, + 0.693 + ], + "angle": 0, + "content": "directly aligns student predictions \\( f_{\\theta}(\\hat{X}_t^\\phi, t) \\) with real HR images, mitigating the teacher's approximation errors. This extra ground-truth constraint helps the model recover fine textural details, which boosts IQA metrics (e.g., MUSIQ and MANIQA) while preserves perceptual fidelity. Consistency learning enhances the robustness of flow-based inference by distilling multi-step capabilities into a single step, allowing for faster sampling without sacrificing quality. Additionally, HR regularization alone (w/ \\( \\mathcal{L}_{hr} \\)) provides minimal IQA gains. This highlights the importance of consistency learning and its synergy with HR regularization for acceleration in SR flow. Fig. 6 presents visual comparisons, showing that consistency learning helps resolve distorted textures in the baseline SR flow model. Notably, HR-regularized consistency learning generates more realistic and sharper results compared to its baseline." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Analysis of Fast-Slow Time Sampling We evaluate the effectiveness of fast-slow time sampling by comparing it to the \\(N\\)-interval scheduler [38] and by varying the number of fast-scheduler timesteps, while fixing the slow scheduler fixed at 1000 timesteps. When the fast scheduler uses only one timestep, we set \\(t + \\Delta t = 1\\) and then sample \\(t\\) from the slow scheduler (see Eq. (3)). As shown in Table 4, increasing \\(\\Delta t\\) (thus reducing the number of intervals) within a reasonable range improves IQA metrics. A plausible explanation is that the slightly larger perturbations introduced at intermediate predictions \\(\\hat{X}_t^\\phi\\) from \\(X_{t}\\) encourage the model to be more robust to distribution shifts; see Fig. 3. This, in turn, reduces variance in its estimates of" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.088, + 0.593, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.15, + 0.58, + 0.161 + ], + "angle": 0, + "content": "LR Image" + }, + { + "type": "image", + "bbox": [ + 0.594, + 0.089, + 0.671, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.604, + 0.15, + 0.661, + 0.161 + ], + "angle": 0, + "content": "w/o \\(\\mathcal{L}_{hrcd}\\)" + }, + { + "type": "image", + "bbox": [ + 0.672, + 0.089, + 0.75, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.685, + 0.15, + 0.735, + 0.161 + ], + "angle": 0, + "content": "w/o \\(\\mathcal{L}_{adv}\\)" + }, + { + "type": "image", + "bbox": [ + 0.751, + 0.089, + 0.827, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.764, + 0.15, + 0.812, + 0.161 + ], + "angle": 0, + "content": "w/o \\(\\mathcal{L}_{iqa}\\)" + }, + { + "type": "image", + "bbox": [ + 0.829, + 0.089, + 0.904, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.853, + 0.15, + 0.878, + 0.16 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.605, + 0.165, + 0.814, + 0.18 + ], + "angle": 0, + "content": "Figure 7. Ablation of training loss." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.189, + 0.907, + 0.245 + ], + "angle": 0, + "content": "Table 4. Ablation of fast-slow time sampling. We compare our fast-slow sampling approach with the \\(N\\)-interval method on DRealSR. The number of \\(N\\) and fast-scheduler timesteps are indicated in parentheses." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.25, + 0.905, + 0.345 + ], + "angle": 0, + "content": "
MethodPSNR ↑LPIPS ↓MUSIQ ↑MANIQA ↑CLIPiQA ↑
N-Interval (50)28.610.280762.480.59140.6599
N-Interval (18)28.740.279262.900.59720.6605
N-Interval (4)28.660.281863.030.59720.6617
Slow Only (1000)28.780.280162.720.59160.6586
Fast-Slow (8)28.540.276064.010.60980.6634
Fast-Slow (4)28.620.283064.770.61560.6748
Fast-Slow (1)28.790.283563.010.59480.6605
" + }, + { + "type": "table_caption", + "bbox": [ + 0.57, + 0.354, + 0.849, + 0.368 + ], + "angle": 0, + "content": "Table 5. Ablation of training loss on DRealSR." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.372, + 0.905, + 0.443 + ], + "angle": 0, + "content": "
MethodPSNR ↑LPIPS ↓MUSIQ ↑MANIQA ↑CLIPiQA ↑
w/o Lhred28.520.301364.420.60610.6954
w/o Ladv28.450.296165.290.61540.7099
w/o Liaq28.640.295864.830.61110.7012
Ours28.500.297565.460.61720.7074
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.456, + 0.907, + 0.578 + ], + "angle": 0, + "content": "the ODE trajectory's starting point, i.e., HR, resulting in sharper restorations. Moreover, compared to the \\(N\\)-interval strategy or slow-only scheduling, our fast-slow scheduler further boosts IQA metrics while maintaining comparable fidelity. This is because the slow scheduler ensures fine-grained SR flow estimation, whereas the fast scheduler enables efficient sampling. Based on these findings, we set the fast scheduler to 4 timesteps throughout our experiments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.587, + 0.909, + 0.709 + ], + "angle": 0, + "content": "Influence of loss function In addition to flow and consistency learning, we incorporate a GAN loss and an image quality alignment loss to further enhance SR performance, as detailed in Eq. (10). As demonstrated in Table 5 and Fig. 7, these losses all improve no-reference metrics and visual quality. The ablation highlights that eliminating any of these losses results in reduced SR quality, with consistency learning contributing a substantial improvement." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.724, + 0.634, + 0.739 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.902 + ], + "angle": 0, + "content": "This paper presents FlowSR, a new approach for efficient one-step image super-resolution. FlowSR reformulates SR as a rectified flow, leveraging the strengths of iterative generative modeling. To enable high-quality single-step inference, we incorporate consistency learning and devise HR regularization to address its distillation target drifting issue. Additionally, a fast-slow time scheduling strategy is designed to enhance the efficiency and robustness of the consistency SR flow model. FlowSR contributes to the advancement of efficient real-world SR applications." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11762" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.484, + 0.207 + ], + "angle": 0, + "content": "The work described in this paper was supported in part by the National Key R&D Program of China (Grant No. 2023YFE0202700), Research Grants Council of the Hong Kong Special Administrative Region, China, under Project 14200824; and by the Hong Kong Innovation and Technology Fund, under Project MHP/092/22." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.221, + 0.188, + 0.237 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.246, + 0.483, + 0.288 + ], + "angle": 0, + "content": "[1] Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang. Toward real-world single image super-resolution: A new benchmark and a new model. In ICCV, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.29, + 0.484, + 0.344 + ], + "angle": 0, + "content": "[2] Bin Chen, Gehui Li, Rongyuan Wu, Xindong Zhang, Jie Chen, Jian Zhang, and Lei Zhang. Adversarial diffusion compression for real-world image super-resolution. In CVPR, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.346, + 0.483, + 0.401 + ], + "angle": 0, + "content": "[3] Qinpeng Cui, Yixuan Liu, Xinyi Zhang, Qiqi Bao, Zhongdao Wang, Qingmin Liao, Li Wang, Tian Lu, and Emad Barsoum. Taming diffusion prior for image super-resolution with domain shift sdes. NeurIPS, 2024. 1, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.403, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[4] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. TPAMI, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.446, + 0.482, + 0.488 + ], + "angle": 0, + "content": "[5] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image super-resolution. In ECCV, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.489, + 0.483, + 0.544 + ], + "angle": 0, + "content": "[6] Linwei Dong, Qingnan Fan, Yihong Guo, Zhonghao Wang, Qi Zhang, Jinwei Chen, Yawei Luo, and Changqing Zou. Tsd-sr: One-step diffusion with target score distillation for real-world image super-resolution. In CVPR, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.546, + 0.483, + 0.599 + ], + "angle": 0, + "content": "[7] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. NeurIPS, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.603, + 0.483, + 0.671 + ], + "angle": 0, + "content": "[8] Xiao He, Huaao Tang, Zhijun Tu, Junchao Zhang, Kun Cheng, Hanting Chen, Yong Guo, Mingrui Zhu, Nannan Wang, Xinbo Gao, et al. One step diffusion-based super-resolution with time-aware distillation. arXiv preprint arXiv:2408.07476, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.673, + 0.483, + 0.729 + ], + "angle": 0, + "content": "[9] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. *NeruIPS*, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.731, + 0.483, + 0.758 + ], + "angle": 0, + "content": "[10] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. *NeruIPS*, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.76, + 0.483, + 0.814 + ], + "angle": 0, + "content": "[11] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[12] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[13] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In ICCV, 2021. 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.246, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[14] Dongjun Kim, Chieh-Hsin Lai, Wei-Hsiang Liao, Naoki Murata, Yuhta Takida, Toshimitsu Uesaka, Yutong He, Yuki Mitsufuji, and Stefano Ermon. Consistency trajectory models: Learning probability flow ode trajectory of diffusion. In ICLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[15] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.236, + 0.905, + 0.292 + ], + "angle": 0, + "content": "[16] Jianze Li, Jiezhang Cao, Zichen Zou, Xiongfei Su, Xin Yuan, Yulun Zhang, Yong Guo, and Xiaokang Yang. Distillation-free one-step diffusion for real-world image super-resolution. arXiv preprint arXiv:2410.04224, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.294, + 0.905, + 0.348 + ], + "angle": 0, + "content": "[17] Yawei Li, Kai Zhang, Jingyun Liang, Jiezhang Cao, Ce Liu, Rui Gong, Yulun Zhang, Hao Tang, Yun Liu, Denis Demandolx, et al. Lsdir: A large scale dataset for image restoration. In CVPR, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.351, + 0.905, + 0.393 + ], + "angle": 0, + "content": "[18] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swin transformer. In ICCV Workshop, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.396, + 0.905, + 0.45 + ], + "angle": 0, + "content": "[19] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Yu Qiao, Wanli Ouyang, and Chao Dong. Diffir: Toward blind image restoration with generative diffusion prior. In ECCV, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.453, + 0.905, + 0.494 + ], + "angle": 0, + "content": "[20] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. In ICLR, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[21] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. In ICLR, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.541, + 0.905, + 0.582 + ], + "angle": 0, + "content": "[22] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, et al. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. In ICLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.585, + 0.905, + 0.626 + ], + "angle": 0, + "content": "[23] Eric Luhman and Troy Luhman. Knowledge distillation in iterative generative models for improved sampling speed. arXiv preprint arXiv:2101.02388, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.629, + 0.905, + 0.683 + ], + "angle": 0, + "content": "[24] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.686, + 0.905, + 0.728 + ], + "angle": 0, + "content": "[25] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal processing letters, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.731, + 0.905, + 0.784 + ], + "angle": 0, + "content": "[26] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In AAAI, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.905, + 0.842 + ], + "angle": 0, + "content": "[27] Mehdi Noroozi, Isma Hadji, Brais Martinez, Adrian Bulat, and Georgios Tzimiropoulos. You only need one step: Fast super-resolution with stable diffusion via scale distillation. In ECCV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[28] Long Peng, Wenbo Li, Renjing Pei, Jingjing Ren, Jiaqi Xu, Yang Wang, Yang Cao, and Zheng-Jun Zha. Towards realistic data generation for real-world super-resolution. In ICLR, 2025. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11763" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[30] Jingjing Ren, Wenbo Li, Haoyu Chen, Renjing Pei, Bin Shao, Yong Guo, Long Peng, Fenglong Song, and Lei Zhu. Ultrapixel: Advancing ultra high-resolution image synthesis to new peaks. NeurIPS, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.482, + 0.272 + ], + "angle": 0, + "content": "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[32] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. TPAMI, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.37 + ], + "angle": 0, + "content": "[33] Axel Sauer, Frederic Boesel, Tim Dockhorn, Andreas Blattmann, Patrick Esser, and Robin Rombach. Fast high-resolution image synthesis with latent adversarial diffusion distillation. In SIGGRAPH Asia, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.371, + 0.482, + 0.411 + ], + "angle": 0, + "content": "[34] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In ECCV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[35] Johannes Schusterbauer, Ming Gui, Pingchuan Ma, Nick Stracke, Stefan Andreas Baumann, Vincent Tao Hu, and Björn Ommer. Boosting latent diffusion with flow matching. In ECCV, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.497 + ], + "angle": 0, + "content": "[36] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[37] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.593 + ], + "angle": 0, + "content": "[38] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In ICML, 2023. 2, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.649 + ], + "angle": 0, + "content": "[39] Haoze Sun, Wenbo Li, Jianzhuang Liu, Haoyu Chen, Renjing Pei, Xueyi Zou, Youliang Yan, and Yujiu Yang. Coser: Bridging image and language for cognitive super-resolution. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.65, + 0.482, + 0.705 + ], + "angle": 0, + "content": "[40] Fu-Yun Wang, Zhaoyang Huang, Alexander Bergman, Dazhong Shen, Peng Gao, Michael Lingelbach, Keqiang Sun, Weikang Bian, Guanglu Song, Yu Liu, et al. Phased consistency models. NeurIPS, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.706, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[41] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In AAAI, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.789 + ], + "angle": 0, + "content": "[42] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. IJCV, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[43] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[44] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In ICCV Workshop, 2021. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.906, + 0.148 + ], + "angle": 0, + "content": "[45] Yufei Wang, Wenhan Yang, Xinyuan Chen, Yaohui Wang, Lanqing Guo, Lap-Pui Chau, Ziwei Liu, Yu Qiao, Alex C Kot, and Bihan Wen. Sinsr: diffusion-based image superresolution in a single step. In CVPR, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.149, + 0.906, + 0.204 + ], + "angle": 0, + "content": "[46] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. NeurIPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.205, + 0.906, + 0.261 + ], + "angle": 0, + "content": "[47] Pengxu Wei, Ziwei Xie, Hannan Lu, Zongyuan Zhan, Qixiang Ye, Wangmeng Zuo, and Liang Lin. Component divide-and-conquer for real-world image super-resolution. In ECCV, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.262, + 0.906, + 0.304 + ], + "angle": 0, + "content": "[48] Rongyuan Wu, Lingchen Sun, Zhiyuan Ma, and Lei Zhang. One-step effective diffusion network for real-world image super-resolution. NeurIPS, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.305, + 0.906, + 0.347 + ], + "angle": 0, + "content": "[49] Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. Seesr: Towards semantics-aware real-world image super-resolution. In CVPR, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.349, + 0.906, + 0.39 + ], + "angle": 0, + "content": "[50] Zhisheng Xiao, Karsten Kreis, and Arash Vahdat. Tackling the generative learning trilemma with denoising diffusion gans. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.391, + 0.906, + 0.461 + ], + "angle": 0, + "content": "[51] Rui Xie, Ying Tai, Chen Zhao, Kai Zhang, Zhenyu Zhang, Jun Zhou, Xiaogian Ye, Qian Wang, and Jian Yang. Addsr: Accelerating diffusion-based blind superresolution with adversarial diffusion distillation. arXiv preprint arXiv:2404.01717, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.462, + 0.906, + 0.516 + ], + "angle": 0, + "content": "[52] Hanshu Yan, Xingchao Liu, Jiachun Pan, Jun Hao Liew, Qiang Liu, and Jiashi Feng. Perflow: Piecewise rectified flow as universal plug-and-play accelerator. NeurIPS, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.518, + 0.906, + 0.587 + ], + "angle": 0, + "content": "[53] Ling Yang, Zixiang Zhang, Zhilong Zhang, Xingchao Liu, Minkai Xu, Wentao Zhang, Chenlin Meng, Stefano Ermon, and Bin Cui. Consistency flow matching: Defining straight flows with velocity consistency. arXiv preprint arXiv:2407.02398, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.589, + 0.906, + 0.645 + ], + "angle": 0, + "content": "[54] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Maniaq: Multi-dimension attention network for no-reference image quality assessment. In CVPR, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.646, + 0.906, + 0.7 + ], + "angle": 0, + "content": "[55] Tao Yang, Rongyuan Wu, Peiran Ren, Xuansong Xie, and Lei Zhang. Pixel-aware stable diffusion for realistic image super-resolution and personalized stylization. In ECCV, 2024.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.702, + 0.906, + 0.757 + ], + "angle": 0, + "content": "[56] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and Bill Freeman. Improved distribution matching distillation for fast image synthesis. NeurIPS, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.758, + 0.906, + 0.814 + ], + "angle": 0, + "content": "[57] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.816, + 0.906, + 0.871 + ], + "angle": 0, + "content": "[58] Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. Scaling up to excellence: Practicing model scaling for photorealistic image restoration in the wild. In CVPR, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.872, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[59] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Reshift: Efficient diffusion model for image super" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11764" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "resolution by residual shifting. NeurIPS, 2023. 1, 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.162 + ], + "angle": 0, + "content": "[60] Zongsheng Yue, Kang Liao, and Chen Change Loy. Arbitrary-steps image super-resolution via diffusion inversion. In CVPR, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[61] Aiping Zhang, Zongsheng Yue, Renjing Pei, Wenqi Ren, and Xiaochun Cao. Degradation-guided one-step image super-resolution with diffusion priors. arXiv preprint arXiv:2409.17058, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[62] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.304 + ], + "angle": 0, + "content": "[63] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.482, + 0.348 + ], + "angle": 0, + "content": "[64] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "11765" + } + ] +] \ No newline at end of file diff --git a/2025/Fast Image Super-Resolution via Consistency Rectified Flow/5e42c446-52f8-4769-b738-a2b096f3a7a5_origin.pdf b/2025/Fast Image Super-Resolution via Consistency Rectified Flow/5e42c446-52f8-4769-b738-a2b096f3a7a5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f75bb6f5d154b4a1bcf573e2ee56c1c2f0e2020e --- /dev/null +++ b/2025/Fast Image Super-Resolution via Consistency Rectified Flow/5e42c446-52f8-4769-b738-a2b096f3a7a5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:db2fa9b7756ee81f85926cf5bb96716b016999243eef951504f16f357de45638 +size 3650783 diff --git a/2025/Fast Image Super-Resolution via Consistency Rectified Flow/full.md b/2025/Fast Image Super-Resolution via Consistency Rectified Flow/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f8bc4496dea34710af238bd73ae6ae529dce138c --- /dev/null +++ b/2025/Fast Image Super-Resolution via Consistency Rectified Flow/full.md @@ -0,0 +1,453 @@ +# Fast Image Super-Resolution via Consistency Rectified Flow + +Jiaqi $\mathrm{Xu}^{1,2*}$ , Wenbo $\mathrm{Li}^{2\dagger}$ , Haoze $\mathrm{Sun}^2$ , Fan $\mathrm{Li}^2$ , Zhixin $\mathrm{Wang}^2$ , Long $\mathrm{Peng}^2$ , Jingjing $\mathrm{Ren}^3$ , Haoran $\mathrm{Yang}^1$ , Xiaowei $\mathrm{Hu}^4$ , Renjing $\mathrm{Pei}^{2\dagger}$ , and Pheng-Ann $\mathrm{Heng}^1$ + +1The Chinese University of Hong Kong 2Huawei Noah's Ark Lab 3HKUST (GZ) 4South China University of Technology + +# Abstract + +Diffusion models (DMs) have demonstrated remarkable success in real-world image super-resolution (SR), yet their reliance on time-consuming multi-step sampling largely hinders their practical applications. While recent efforts have introduced few- or single-step solutions, existing methods either inefficiently model the process from noisy input or fail to fully exploit iterative generative priors, compromising the fidelity and quality of the reconstructed images. To address this issue, we propose FlowSR, a novel approach that reformulates the SR problem as a rectified flow from low-resolution (LR) to high-resolution (HR) images. Our method leverages an improved consistency learning strategy to enable high-quality SR in a single step. Specifically, we refine the original consistency distillation process by incorporating HR regularization, ensuring that the learned SR flow not only enforces self-consistency but also converges precisely to the ground-truth HR target. Furthermore, we introduce a fast-slow scheduling strategy, where adjacent timesteps for consistency learning are sampled from two distinct schedulers: a fast scheduler with fewer timesteps to improve efficiency, and a slow scheduler with more timesteps to capture fine-grained texture details. Extensive experiments demonstrate that FlowSR achieves outstanding performance in both efficiency and image quality. + +# 1. Introduction + +Real-world image super-resolution (SR) aims to reconstruct high-resolution (HR) images from their low-resolution (LR) counterparts while simultaneously removing unknown degradations. With the remarkable success of diffusion models (DMs) [10, 37], SR methods leveraging these models—particularly those built on powerful text-to-image (T2I) models like Stable Diffusion (SD) [31]—have demonstrated outstanding performance [42, 49, 58]. However, + +![](images/9e40a1bbe783c249056bf060a0e4fe9007e9106282587e07ec741ee5b370fc60.jpg) +LR + +![](images/04211c3713bb2157bb9098e89025f7dd7483308b81d3e26e1207556f740e36a2.jpg) + +![](images/5c730eb00a9e8e280a5248bfc3ef3d4093360c5c8a0b254d7b8dd5b43568c0a1.jpg) +1-step + +![](images/61e635f8088d6668e4b0bcc9fdf3e3510f5d9aaa03f499ca25f2baa5a738310a.jpg) + +![](images/07628966f945812b770dca43a957ffd6e3d71db2bcd5402911110f120610080f.jpg) +4-step + +![](images/a1f8bc0202f25b9aac971afa920984f5bdc92d07c1552ed072191f50fdf81acf.jpg) + +![](images/aa55a7b9cbf5adeee59ea0a73ab91d51d81f58bb0a06fa8d5e89535c57d1e3cd.jpg) +distilled, 1-step + +![](images/98c21e30f8e91eb87338b126ec47bf7680cb0d4b8cab350af7b327b72aa46e84.jpg) +Figure 1. Our consistency SR flow model achieves high-quality single-step inference (distilled, 1-step) by distilling from multi-step higher-quality sampling process (top). We achieve this by formulating SR as a rectified flow that bridges LR and HR images, combined with improved consistency learning (bottom). + +these diffusion-based SR approaches require an iterative reverse sampling process that gradually refines noisy inputs into HR outputs. Despite advancements such as efficient ordinary differential equation (ODE) solvers like DDIM [36], the slow inference speed remains a bottleneck, limiting their practicality in real-world applications. + +Recently, few-step or single-step SR methods derived from diffusion models are designed [3, 45, 48, 59, 61]. Several studies focus on designing more efficient diffusion processes for SR. ResShift [59] speeds up diffusion by shifting the LR-HR residual into the Markov chain and reduces sampling to 15 steps, while DoSSR [3] introduces a diffusion process more compatible with pre-trained DMs. However, they rely on DDPM [10], which uses a more curved transition trajectory and starts with a noise-perturbed LR image that loses crucial information, ultimately limiting performance. In parallel, another line of work tar + +gets single-step SR by learning the LR-HR mapping directly. For example, SinSR [45] learns one-step SR prediction by leveraging the teacher's output from ResShift [59] and ground-truth HR as target, whereas OSEDiff adapts pre-trained SD for SR and refines one-step prediction with the VSD loss [46]. However, these methods fail to fully harness the advantages of iterative generative modeling, which naturally facilitates high-quality texture synthesis. + +In this work, we present FlowSR, which unifies rectified flow with consistency models (CMs) to enable efficient, single-step image super-resolution. As illustrated in Fig. 1, we reformulate SR as a rectified flow [21], which establishes a simple and straight ODE-based mapping between LR and HR images. By explicitly modeling this trajectory, FlowSR learns fine-grained LR-to-HR transformations for better SR quality and facilitates fast sampling. Building upon this formulation, we further leverage consistency models [38] to enhance single-step inference. Enforcing consistency across points on the same SR flow trajectory distills multi-step higher-quality restoration into fewer steps that reach to the same HR result. + +However, the naive consistency distillation (CD) objective in CMs is suboptimal for SR flow, where the SR task demands both high quality and high fidelity. While CD enforces self-consistency across discretization steps, there is no guarantee that the final distillation target aligns well with the ground-truth HR. To address this, we propose HR-regularized consistency learning, which explicitly requires the model's predictions to match real HR images. This additional constraint mitigates teacher-induced errors in the distillation target and enhances the reconstruction of fine-grained details. To further improve efficiency and robustness of the consistency SR flow model, we introduce a fast-slow time scheduling strategy. Rather than sampling distillation timestep pairs from the two boundaries of a discretized interval, we sample adjacent timesteps from distinct "fast" and "slow" schedulers. The fast scheduler uses fewer timesteps to facilitate efficient inference, while the slow scheduler employs more granular steps to maintain alignment with the SR flow objectives. This mixed sampling introduces large and flexible jumps with mild perturbations into the HR regularization steps, allowing for a broader coverage of SR flow trajectories. + +To train our consistency SR flow model, we first finetune a pre-trained SD model to align with the SR flow objective, followed by consistency SR flow distillation. Additionally, we incorporate a GAN [7] loss and an image quality alignment loss, with the latter promoting desirable text-described attributes in the restored images, thereby improving the SR quality. + +As illustrated in Fig. 1, our model leverages the proposed consistency SR flow distillation to transfer the quality improvement typically achieved with multiple sampling + +steps into a single step. Experiments on real-world datasets demonstrate the superiority of FlowSR. Our core contribution lie in a new paradigm to solve the one-step SR problem. First, we explore efficient flow modeling for SR and identify potential challenges when incorporating consistency models. Second, we introduce several techniques, including HR regularization in consistency learning and a fast-slow time scheduling to enhance one-step SR performance. + +# 2. Related Work + +# 2.1. Image Super-Resolution + +Real-world image super-resolution (SR) aims to reconstruct high-resolution (HR) images from low-resolution (LR) inputs, tackling complex and unknown degradations like noise, blur, and compression artifacts. Since Dong et al. [5], numerous deep learning-based methods have been proposed to address the SR problem from different perspectives, including network design [18], model training [15], and degradation simulation [28, 44, 62]. Recently, diffusion models [10, 37] have achieved remarkable success in image generation [30, 31], which inspires their application to SR and demonstrate significant advancements [32, 42]. + +Diffusion Model-Based SR Conditioned on the LR image, diffusion model-based SR iteratively denoises towards the target HR image. The LR condition can be leveraged either by input concatenation [32] or via adapters [26, 63]. StableSR [42] based on Stable Diffusion [31], uses a trainable time-aware encoder to incorporates LR. Subsequent improvements are achieved through enhanced LR conditioning in DiffBIR [19], semantics-aware prompts in SeeSR [49], reference image generation in CoSeR [39], and scaled-up training in SUPIR [58]. However, these DM-based SR approaches, starting from noise, suffer from low inference speed, typically requiring 50-200 sampling steps. In contrast, ResShift [59] constructs an efficient diffusion model by shifting the residual between HR and LR. Denoising from the noised LR reduces the sampling steps to 15 but it still suffers from limited restoration quality. + +Single-Step / Few-Step Image SR Fast diffusion model-based SR methods require only one or a few sampling steps during inference, typically achieved through knowledge distillation [23] or adversarial training [15, 50]. SinSR [45] derives a deterministic sampling process for ResShift and reduces its number of sampling steps to one using distillation. OSEDiff [48] applies variational score distillation [46, 57] to enhance one-step restoration quality by minimizing the KL-divergence between the distribution of its generated outputs and that of the pre-trained T2I model. AddSR [51] achieves 4-step inference by tailoring adversarial diffusion distillation [34] for SR. Similarly, other recent methods, accelerate the inference by using distillation [6, 8, 27] or GAN [2, 16, 60, 61]. On the other hand, more + +![](images/f92d7b59b0a007023af77f66d56f8ecc8f168c539bb3637331196eca26a6f6f2.jpg) +Figure 2. Overview of the training process. The consistency SR flow distills multi-step, high-quality SR capability into single-step inference, while HR regularization ensures the model converges to the high-resolution target. Times $t$ , $t'$ are sampled alternately from fast and slow schedulers. Note that the flow loss $\mathcal{L}_{flow}$ and the consistency loss $\mathcal{L}_{hrcd}$ are computed for different samples within each batch. + +efficient SR diffusion processes are constructed using domain shift [3] or flow matching [35], which allow few-step sampling. Yet, they still suffer from inefficient modeling designs and suboptimal performance. + +# 2.2. Consistency Models + +Consistency Models (CMs) [38] learn to map any point on the ODE trajectory to its origin, which enable single-step generation and allow trade-offs between quality and computation through multi-step sampling. CMs can be implemented for modern T2I models in latent space [24]. Consistency Trajectory Models (CTMs) [14] mitigate the multi-step sampling issues in CMs by learning an any-to-any mapping from initial points to final points on the ODE trajectory. The Phased Consistency Model [40] partitions the entire ODE trajectory into multiple sub-trajectory phases and ensures consistency within each phase, while PeRFlow [52] straightens the sub-trajectories using the Reflow [22] operation. Additionally, Consistency Flow Matching [53] further enforces velocity field consistency. + +# 3. Preliminaries + +# 3.1. Rectified Flow + +Rectified flow [21] is an ODE-based generative model that constructs a simple, straight trajectory to transform samples between two distributions: $\pi_0$ (e.g., data) and $\pi_1$ (e.g., noise). Rectified flow defines a linear interpolation path between observed samples $X_0\sim \pi_0$ and $X_{1}\sim \pi_{1}$ as: $X_{t} = tX_{1} + (1 - t)X_{0}$ with time $t\in [0,1]$ . The core idea is to learn a velocity field $v_{\theta}(X_{t},t)$ , parameterized by a neural network with weights $\theta$ , that matches the derivative of this trajectory. This is achieved by optimizing the following objective: $\min_{\theta}\int_{0}^{1}\mathbb{E}_{X_0\sim \pi_0,X_1\sim \pi_1}[\| v_{\theta}(X_t,t) - (X_1 - X_0)\| ^2 ]dt,$ which encourages the learned velocity field to align with + +the direction of the straight path $X_{1} - X_{0}$ . Once trained, samples from $\pi_{1}$ can be transformed into $\pi_{0}$ by solving the ODE: $dX_{t} = v_{\theta}(X_{t}, t)dt$ , $X_{1} \sim \pi_{1}$ . A key advantage of rectified flow lies in its straight trajectories. By design, the optimal velocity field corresponds to a constant-speed flow between $X_{0}$ and $X_{1}$ , enabling efficient and stable sampling with only a few ODE steps. + +# 3.2. Consistency Models + +Consistency models [38] are a class of generative models that enable high-quality sample generation with few computational steps. These models learn to map any point $(X_{t}, t)$ along a probability flow ODE trajectory directly to its origin $X_{\epsilon}$ , where $\epsilon$ is a fixed small positive number denoting the trajectory's start. This capability is formalized through the self-consistency property: for any pair of points $(X_{t}, t)$ and $(X_{t'}, t')$ on the same ODE trajectory, the model $f_{\theta}(\cdot, t)$ is trained to satisfy: $f_{\theta}(X_{t}, t) = f_{\theta}(X_{t}', t') = X_{\epsilon}$ for all $t, t' \in [\epsilon, 1]$ . To ensure consistency during training, the boundary condition $f_{\theta}(X_{\epsilon}, \epsilon) = X_{\epsilon}$ is enforced. Consistency models eliminate trajectory drift by mapping all intermediate states to a consistent origin, avoiding errors from iterative denoising steps. This enables them to generate high-fidelity samples even with large step sizes, such as directly mapping from $t = 1$ to the origin in a single step. + +# 4. Methodology + +We leverage the generative capability of rectified flow for high-quality SR and explore consistency models to enable fast inference with fewer sampling steps (i.e., one step), without compromising quality. We define a rectified flow, termed SR flow, for image super-resolution in Sec. 4.1. We then introduce the improved consistency SR flow learning method in Sec. 4.2 and detail the training in Sec. 4.3. + +# 4.1. Rectified Flow for SR + +SR Flow We consider Rectified Flow naturally aligns with image super-resolution, where high-resolution (HR) images $X_{\mathrm{HR}} \sim \pi_0$ and their low-resolution (LR) counterparts $X_{\mathrm{LR}} \sim \pi_1$ . We define the forward process as a straight path via a linear interpolation between them: + +$$ +X _ {t} = (1 - t) X _ {\mathrm {H R}} + t X _ {\mathrm {L R}}, \tag {1} +$$ + +here time $t\in [0,1]$ with $t = 0$ corresponding to $X_{\mathrm{HR}}\in \mathbb{E}^{H\times W\times 3}$ and $t = 1$ corresponding to $X_{\mathrm{LR}}\in \mathbb{E}^{H\times W\times 3}$ , where $X_{\mathrm{LR}}$ is upscaled to match the spatial dimensions of $X_{\mathrm{HR}}$ . The learning objective of the SR neural network $v_{\theta}$ is to regress the conditional vector fields [20, 21] by following the direction $X_{\mathrm{LR}} - X_{\mathrm{HR}}$ : + +$$ +\mathbb {E} _ {t, X _ {t}} \| v _ {\theta} \left(X _ {t}, t\right) - \left(X _ {\mathrm {L R}} - X _ {\mathrm {H R}}\right) \| _ {2} ^ {2}. \tag {2} +$$ + +Intuitively, this SR flow establishes a smooth transition between HR and LR. During training, the SR flow model $v_{\theta}$ implicitly learns to invert degradations (e.g., blur, noise) and recover high-frequency details through intermediate refinements such as edge sharpening and texture synthesis. At inference, HR images are reconstructed via reverse sampling along the learned trajectory. Starting from $X_{\mathrm{LR}}$ , we solve the reverse ODE using numerical methods like Euler solver: $X_{t'} = X_t + \Delta t \cdot v_{\theta}(X_t, t)$ from $t = 1$ to $t = 0$ . Notably, unlike diffusion-based SR methods that corrupt LR inputs with noise [3, 35, 59], our transition process directly starts from LR, which preserves most of the structural information in the LR image, enabling stable and efficient sampling. + +Faster Inference with SR Flow A key advantage of SR flow is its flexibility in sampling steps. It supports faster inference by reducing the number of sampling steps, requiring as few as a single step that directly maps the LR image to its corresponding HR output: $\hat{X}_{\mathrm{HR}} = X_{\mathrm{LR}} - 1 \cdot v_{\theta}(X_{\mathrm{LR}}, 1)$ . This is more easily achievable than the noise-to-image mapping or initializing from noisy inputs in ResShift [59], thanks to the strong correlation between LR and HR images. Additionally, SR flow also supports any number of sampling steps and produces better visual quality when using iterative multi-step sampling, as illustrated in Fig. 1. + +# 4.2. Consistency SR Flow + +We enhance single-step super-resolution in SR flow models by distilling multi-step sampling capability (e.g., four steps) into one via consistency learning (Fig. 1). To ensure accurate HR reconstruction, we incorporate the HR target explicitly into consistency learning. Additionally, a fast-slow time sampling schedule is introduced to improve efficiency and robustness. Fig. 2 shows the overall training process. + +Consistency Distillation in SR Flow We start with consistency distillation (CD) to distill the capability of a (pre-trained) teacher SR flow model $v_{\phi}(X,t)$ , enabling high-quality SR in fewer inference steps. Given timesteps $t$ and + +![](images/32c0695adbb690cffb3c861d73585f794967f8771716d86b4ab48c68d3ad15e5.jpg) + +![](images/ece13ce9fb4a39df413b03a6811fd38bff15f2cc7b47fff6208b1db6651bae20.jpg) + +![](images/4a1e98cf55e51444b16d07dc7af45db010b60ca8e4b635d85576d712a0eed804.jpg) +$X_{t^{\prime}}$ + +![](images/201cc9c752b509fd9869df838e4bc1d998c1c7f57d1f0c00ed8347e3c49e5dba.jpg) + +![](images/6e2de18bc63fa8a34c7c4892b50c23112c56d457376c770854b4b5d93c171ef4.jpg) +$X_{t}$ + +![](images/56f16afb06b14c6d00c844a92b850f7a2f48c3ee491ae4b7d2882856f2d8d813.jpg) + +![](images/1cd07dc6e4e4c566461dbf8e6ccb4ae43c2984e0da7520a0371a91d4f75db23b.jpg) +$\hat{X}_t^\phi$ + +![](images/27edee5717286b1868303ddbc9b29dba251c32cf8da0420421ba49e585e1ad81.jpg) + +![](images/98fffe65e49318c271c277df25ae717964e996b77729f039ce3686a0b94dbf69.jpg) +$\hat{X}_{t\rightarrow 0}^{\theta^{-}}$ +Figure 3. Illustration of HR-regularized consistency learning. Approximation errors may cause the distillation target $\hat{X}_{t\rightarrow 0}^{\theta^{-}}$ to deviate from the true high-quality HR target. HR regularization corrects this by enforcing HR alignment under mild perturbations. + +![](images/e3ac8bcb1b75a8ecfa075cfad810a7a01e0e93619f25a5feac9290a779ca8787.jpg) + +![](images/89bfce9f9f4e0b8fc38173cc8d81f9cefc07450ae8b495c63c6e7ea4598c994d.jpg) +$X_0$ + +$t^{\prime} = t + \Delta t$ (where $\Delta t > 0$ ), we first sample $X_{t'}$ from the forward degradation process (Eq. (1)). The teacher model then estimates $\hat{X}_t^\phi$ from its degraded counterpart $X_{t'}$ : + +$$ +\hat {X} _ {t} ^ {\phi} = X _ {t ^ {\prime}} - \Delta t \cdot v _ {\phi} \left(X _ {t ^ {\prime}}, t ^ {\prime}\right), \tag {3} +$$ + +where Euler solver is used. The consistency function $f_{\theta}(X_t, t)$ for deriving the origin is defined by: + +$$ +f _ {\theta} \left(X _ {t}, t\right) = X _ {t} - t \cdot v _ {\theta} \left(X _ {t}, t\right), \tag {4} +$$ + +which maps any degraded input $X_{t}$ at timestep $t$ to the HR image domain. The consistency distillation in SR flow learns a vector field $v_{\theta}$ by minimizing the distance between the prediction of the network $\hat{X}_{t^{\prime}\to 0}^{\theta} = f_{\theta}(X_{t^{\prime}},t^{\prime})$ and the distillation target $\hat{X}_{t\rightarrow 0}^{\theta^{-}} = f_{\theta^{-}}(\hat{X}_{t}^{\phi},t)$ : + +$$ +\mathcal {L} _ {c d} = \mathbb {E} \left[ d \left(f _ {\theta} \left(X _ {t ^ {\prime}}, t ^ {\prime}\right), f _ {\theta^ {-}} \left(\hat {X} _ {t} ^ {\phi}, t\right)\right) \right], \tag {5} +$$ + +where $d(\cdot, \cdot)$ is a distance metric, and $\theta^{-}$ denotes the target model parameters (e.g., exponential moving average (EMA) of $\theta$ ). This distillation loss enables the newly trained model, $f_{\theta}$ , to generate high-quality samples in a single step. + +HR-Regularized Consistency Learning Although the original CD objective theoretically ensures that all trajectory points map to the same origin, it does not explicitly restrict alignment between the generated samples and the corresponding HR targets. This is because velocity field approximation errors from the teacher model $v_{\phi}$ and the target model $v_{\theta^{-}}$ can propagate into the distillation target $f_{\theta^{-}}(\hat{X}_{t}^{\phi}, t)$ , as shown in Fig. 3. + +To address this, we introduce an extra HR regularization into Eq. (5), which directly enforces the student's + +predictions $\hat{X}_{t\rightarrow 0}^{\theta} = f_{\theta}(\hat{X}_t^\phi ,t)$ (thus distillation target $f_{\theta^{-}}(\hat{X}_t^\phi ,t))$ to match the ground-truth HR images $X_0$ : + +$$ +\mathcal {L} _ {h r c d} = \mathbb {E} \left[ d \left(f _ {\theta} \left(X _ {t ^ {\prime}}, t ^ {\prime}\right), f _ {\theta^ {-}} \left(\hat {X} _ {t} ^ {\phi}, t\right)\right) + d \left(f _ {\theta} \left(\hat {X} _ {t} ^ {\phi}, t\right), X _ {0}\right) \right]. \tag {6} +$$ + +By training with real HR targets, consistency learning reduces reliance on the imperfect teacher and target model. This dual consistency objective ensures that $f_{\theta}$ maintains both trajectory consistency and fidelity to HR data. Note that the input $(\hat{X}_t^\phi ,t)$ can be viewed as a perturbed version of sampled $(X_{t},t)$ from the forward process in Eq. (1), as visualized in Fig. 3. This perturbation resembles signal noise and distribution shifts, thereby improving the robustness of the trained model. + +Fast-Slow Time Scheduling To further improve the efficiency and robustness of consistency learning, we propose sampling pairs of timesteps, $t$ and $t'$ , from two distinct time schedulers: a "fast" scheduler with relatively few timesteps (e.g., 4) and a "slow" scheduler with more granular timesteps (e.g., 1000). The slow scheduler, a standard practice for diffusion or flow models, provides fine-grained learning signals to keep intermediate velocity field predictions well-aligned with the SR flow objectives (Eq. (2)). In contrast, the fast scheduler emphasizes larger jumps along the trajectory, reflecting more desired one-step inference. + +During training, we randomly select one scheduler (fast or slow) to first sample $t + \Delta t$ , then sample $t$ in its adjacent region from the other scheduler, and compute $(\hat{X}_t^\phi ,t)$ accordingly using Eq. (3), as illustrated in Fig. 3. This procedure allows flexible and larger jumps $\Delta t$ through the fast sampler and relaxes the requirement for accurately estimating of $X_{t}$ from $X_{t + \Delta t}$ by running one discretization step of a numerical ODE solver [38]. We hypothesize that introducing mild perturbations and diversity to $\hat{X}_t^\phi$ , deviating it from the sampled point $X_{t}$ , enhances the SR model's robustness and mitigates distribution shifts, particularly when combined with HR regularization. + +# 4.3. Training Consistency SR Flow + +SR Flow in Image Space We use Stable Diffusion [31] as the backbone for our SR model, comprising a VAE to map images $X$ into latent representations $\mathbf{x}$ , along with a UNet $\theta$ adapted for velocity field learning. While the conditional flow matching loss in Eq. (2) can be formulated in latent space, our early experiments observed that such a latent-space objective does not consistently maintain a favorable balance between fidelity and visual quality over varying numbers of inference steps (e.g., from one to four). + +In practice, we find that loss in the image space is more effective. Specifically, during SR flow training, we first predict the velocity prediction $v_{\theta}(\mathrm{x}_t,t)$ in the latent space and compute the HR latent representation $\hat{\mathrm{x}}_{\mathrm{HR}} = f_{\theta}(\mathrm{x}_t,t)$ (see Eq. (4)). Next, we decode $\hat{\mathrm{x}}_{\mathrm{HR}}$ into the SR image $\hat{X}_{\mathrm{HR}}$ + +using the VAE's decoder. Last, we compare $\hat{X}_{\mathrm{HR}}$ to the ground-truth HR image $X_{\mathrm{HR}}$ using an $l_{2}$ loss and a perceptual (LPIPS [64]) loss: + +$$ +\mathcal {L} _ {f l o w} = \left\| \hat {X} _ {\mathrm {H R}} - X _ {\mathrm {H R}} \right\| _ {2} ^ {2} + \lambda_ {p} \operatorname {L P I P S} \left(\hat {X} _ {\mathrm {H R}}, X _ {\mathrm {H R}}\right), \tag {7} +$$ + +where $\lambda_{p} = 2$ is set to weight the LPIPS term. + +Adversarial Loss We incorporate an adversarial GAN objective to further enhance SR quality, particularly for texture synthesis. Specifically, a discriminator $\mathcal{D}$ is trained to distinguish between real HR images and those restored by our model. We use the pre-trained diffusion model as a feature extractor in the latent space and attach several additional discriminator heads, following [33, 40, 56]. The hinge loss is adopted for training: + +$$ +\mathcal {L} _ {a d v} = \mathbb {E} \left[ \max (0, 1 - \mathcal {D} \left(\mathrm {x} _ {\mathrm {H R}}\right)) + \max (0, 1 + \mathcal {D} \left(\hat {\mathrm {x}} _ {\mathrm {H R}}\right)) \right]. \tag {8} +$$ + +Note that the input to $\mathcal{D}$ is the noised latent with diffusion time $t$ (omitted here for simplicity). The SR model is then trained to minimize this adversarial objective. + +Image Quality Alignment Loss To further align the SR outputs with desirable image quality attributes (e.g., high-resolution, sharp, detailed), we propose an image quality alignment loss based on a text-image contrastive loss. Specifically, we use large vision-language models to generate a positive quality caption $c_{pos}$ (e.g., "image quality is good, with clear details and vibrant colors") from the HR image and a negative quality caption $c_{neg}$ (e.g., "image quality is poor, blurry, and noisy") from its LR counterpart. We then encode these captions using the text encoder $\mathcal{E}_T$ of CLIP [29], while the SR result $\hat{X}$ is encoded via CLIP's image encoder $\mathcal{E}_I$ (ViT-L/14 is used). Image quality alignment loss encourages $\hat{X}$ to be close to $c_{pos}$ and far from $c_{neg}$ : + +$$ +\mathcal {L} _ {i q a} = - \log \frac {\exp \left(\sin \left(\hat {X} , c _ {p o s}\right)\right)}{\exp \left(\sin \left(\hat {X} , c _ {p o s}\right)\right) + \exp \left(\sin \left(\hat {X} , c _ {n e g}\right)\right)}, \tag {9} +$$ + +where $\mathrm{sim}(X,c) = \cos (\mathcal{E}_I(X),\mathcal{E}_T(c))$ . Intuitively, this loss implicitly guides the network to improve perceptual fidelity and clarity by bridging the gap between human-perceived "good" and "bad" image attributes. + +Putting Things Together We first pre-train the SR flow model using the basic conditional flow loss in Eq. (7). Then we continue with the consistency SR flow training by finetuning it. The overall loss function is defined as: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {f l o w}} + \lambda_ {c d} \mathcal {L} _ {h r c d} + \lambda_ {a d v} \mathcal {L} _ {a d v} + \lambda_ {i q a} \mathcal {L} _ {i q a}, \tag {10} +$$ + +where $\lambda_{cd}$ , $\lambda_{adv}$ , and $\lambda_{iqa}$ are hyperparameters. + +Table 1. Quantitative comparison with state-of-the-art DM-based SR methods on RealSR [1] and DRealSR [47]. The best and second-best results are highlighted in bold and underlined. + +
DatasetsMethods#StepsPSNR ↑SSIM ↑LPIPS ↓DISTS ↓FID ↓NIQE ↓MUSIQ ↑MANIQA ↑CLIPIQA ↑
RealSRStableSR [42]20024.700.70850.30180.2288128.515.9165.780.62210.6178
DiffBIR [19]5024.750.65670.36360.2312128.995.5364.980.62460.6463
SeeSR [49]5025.180.72160.30090.2223125.555.4169.770.64420.6612
PASD [55]2025.210.67980.33800.2260124.295.4168.750.64870.6620
ResShift [59]1526.310.74210.34600.2498141.717.2658.430.52850.5444
SinSR [45]126.280.73470.31880.2353135.936.2960.800.53850.6122
OSEDiff [48]125.150.73410.29210.2128123.495.6569.090.63260.6693
DoSSR [3]125.670.73870.33560.2741153.2511.1962.690.52430.6259
FlowSR125.540.74340.27280.2013112.605.2869.220.64860.6701
DRealSRStableSR [42]20028.030.75360.32840.2269148.986.5258.510.56010.6356
DiffBIR [19]5026.710.65710.45570.2748166.796.3161.070.59300.6395
SeeSR [49]5028.170.76910.31890.2315147.396.4064.930.60420.6804
PASD [55]2027.360.70730.37600.2531156.135.5564.870.61690.6808
ResShift [59]1528.460.76730.40060.2656172.268.1250.600.45860.5342
SinSR [45]128.360.75150.36650.2485170.576.9955.330.48840.6383
OSEDiff [48]127.920.78350.29680.2165135.306.4964.650.58990.6963
DoSSR [3]128.550.79910.33530.2801166.1912.2556.720.46230.5739
FlowSR128.500.78590.29750.2115130.306.1365.460.61720.7074
+ +# 5. Experiments + +# 5.1. Experimental Settings + +Implementation Details Our model is based on the pretrained SD 2.1-base [31]. During training, we fine-tune only the U-Net using a LoRA [11] rank of 32. The learning rate is set to 2e-5, with a training patch size of $512 \times 512$ and a batch size of 16. The training process consists of two stages: we firstly train the SR flow model for 10k iterations, followed by consistency learning for an additional 20k iterations. The loss weights $\lambda_{cd}, \lambda_{adv}, \lambda_{iqa}$ are set to 0.1, 0.05, and 0.1, respectively. + +Data We train our model using LSDIR [17] and the first 10K face images from FFHQ [12]. The LR-HR training pairs are synthesized using the degradation pipeline of RealESRGAN [44]. To generate image quality captions for HR and LR images, we use Qwen2-VL [43]. We adopt the real-world test set of StableSR [42] for evaluation and comparison. The test sets include RealSR [1] and DRealSR [47]. + +Compared Methods We compare our method against several state-of-the-art diffusion-based SR approaches, including multi-step StableSR [42], DiffBIR [19], SeeSR [49], PASD [55], ResShift [59], and one-step SinSR [45], OSEDiff [48], and DoSSR [3]. + +Evaluation Metrics We evaluate our model using both reference-based and no-reference metrics. For fidelity, we report PSNR and SSIM (on the Y channel in YCbCr space), and for perceptual quality, we use LPIPS [64] and DISTS [4]. FID [9] is reported to compare the distribution of restored images with the ground truth. For no-reference evaluation, we use NIQE [25], MUSIQ [13], MANIQA [54], and CLIPIQA [41]. + +# 5.2. Comparison with State-of-the-Art Methods + +Quantitative Comparisons Table 1 shows that our approach achieves superior or at least competitive performance compared to recent state-of-the-art methods. Compared with multi-step diffusion-based methods such as SeeSR [49] and PASD [55], our FlowSR obtains equivalent or better no-reference metrics, such as MUSIQ, MANIQA, and CLIPIQA, and surpasses these methods in fidelity-based scores, e.g. PSNR and LPIPS, while reducing inference to a single step. Compared to efficient diffusion-based SR methods such as ResShift [59], our method demonstrates a clear advantage. Against the strong DDPM-based one-step competitor OSEDiff [48], FlowSR achieves superior performance across nearly all metrics. Overall, these results demonstrate that our flow-based SR method effectively balances fidelity and perceived quality while showing superiority over compared methods. + +Qualitative Comparisons Fig. 4 shows visual comparisons of two real-world examples. Multi-step DM-based methods, such as DiffBIR and PASD, often generate rich but inaccurate textures, while one-step methods like SinSR and OSEDiff tend to produce blurry or less detailed results. In contrast, our method generates faithful SR results, effectively recovering structures such as the stone breakwater and the textures of the cloth. + +# 5.3. Ablation Studies + +Effects of SR Flow We first study the flow formulation for image super-resolution. Our SR flow is a straight path between $\pi_0 = X_{\mathrm{HR}}$ and $\pi_1 = X_{\mathrm{LR}}$ . We consider alternative formulations starting from noise, where $\pi_1 = \mathcal{N}(0,1)$ , as used in pre-trained T2I models, or starting from noised + +![](images/54ca3c2cdf6d193f30d80467308a569a4733ccf75ab8c171d543842c8cb8d9c3.jpg) +LR Image + +![](images/042271f964a8bd2b41f4561931764176e55ae68a0891e43be10a2913e7c6f0c2.jpg) +HR + +![](images/2335d0a9e3c20b22e292a413ea24cb7147093ad51b816d13569cf946a1786b7e.jpg) + +![](images/7ae6ec87e9e13671c5fbce08ae17b8a88ef065515fb7dece036c9e3937d4c532.jpg) + +![](images/d3e7a7995566469f3751fd5e13b66d6fa97308963b5fa3a548ea274d28db309e.jpg) + +![](images/ffdcf532b63640a46e3b1e4fa965394367c0771e270acfdcf069a07cd58587b6.jpg) + +![](images/7a7f5580fd114f25da788d007a33db727b9e5492426ffd1eb15b384d4758a628.jpg) +LR Image + +![](images/2556f83fe362a993be4e8e05fc131e3e97c2963e723ff608dd33c231645d68ab.jpg) +ResShift (15) + +![](images/d8b83d480f026a3c3978bf04f45a8a228ff1814ed5b2690573e027986acc6dbd.jpg) +StableSR (200) +SinSR (1) + +![](images/8d4acf7a89ffd5d63f9b6b453a71c2cb7bd162e6a781f1951492ef65b6d1701d.jpg) +DiffBIR (50) +OSEDiff (1) + +![](images/691fad6514142eff3400fd73ed44afd439a33039b84f3ad3f50ab54fc88ab68d.jpg) +SeeSR (50) +DoSSR (1) + +![](images/16e31545a1e4d897f2c1fac6ccebd5e1c196f90722bc9327e26486bdc8eb3ad9.jpg) +PASD (20) +FlowSR (1) + +![](images/c2d787425c24e1cba5784055449b89c0f1c381a2372324f824b1548e6f8fc6cb.jpg) +HR + +![](images/dc15a0f907d55c49679a14f74c7379fa049a309bb89fec07fa1de9dff1997bb9.jpg) +StableSR (200) + +![](images/a28445e04f5a94e8668ca63da78effc189ef7e9c61b94d022716ce79f67954e3.jpg) +DiffBIR (50) + +![](images/e295b648acc4829cc109a3ff5a6370d5ca565261faf9a2ef4588a3cef1cc92d0.jpg) +SeeSR (50) + +![](images/d0d7bb8ee3403385ef0c9915a543dbb8f5ffc51d52ecae11c8059975e27c9ff6.jpg) +PASD (20) + +![](images/090d29280d1f2335183e731a9ddc23d15db14a0cae759e3b8a535e58f6cea11e.jpg) +Figure 4. Visual comparisons of different SR methods on real-world examples. The number of sampling steps are indicated in parentheses. Please zoom in for a better view. + +![](images/ccc115ca9f373cd25e40fd871f8ad73df503f27d3221ce3d8bed5f35a00f147a.jpg) + +![](images/2ca2b5d3d20e3cc93e4197ce3c50626f7de236eee48d730b394483801f1fc3ac.jpg) + +![](images/bb1236b864ecc899729334b5e6d586f2cc471b7adfafe3a9a44e358954face1a.jpg) + +![](images/f3ec8a1aa5d57dac5a0b60ac2f4842d085d8e77e096aac29dc8c8545b095e662.jpg) +LR Image +Zoomed LR +Figure 5. Ablation of SR flow. Our SR flow model produces more accurate SR results with improved structure with more steps. + +![](images/a0bc5dea370fb5836135c149f48b0e2030a0629a96662408f1b49d4a3840fc29.jpg) +$\mathrm{Noise}\rightarrow \mathrm{HR}$ (4) +$\mathrm{Noise}\rightarrow \mathrm{HR}$ (1) + +![](images/74e8bb9ff6cb2de759e636de029967c0d69683835d28d475909dfa4072001db5.jpg) +noised $\mathrm{LR}\rightarrow \mathrm{HR}$ (4) +noised $\mathrm{LR}\rightarrow \mathrm{HR}$ 1 + +![](images/7f0998d4de7ad064534623635986ef9f1dcfa5b2a3cee1255d1361d1e7dc7ca8.jpg) +SR Flow (4) +SR Flow (1) + +LR, where $\pi_1 = X_{\mathrm{LR}} + \mu \epsilon$ with $\epsilon \sim \mathcal{N}(0,1)$ , as proposed in ResShift [59]. For these alternatives, we concatenate $X_{t}$ with $X_{\mathrm{LR}}$ as the input to provide the LR condition. The trained models are evaluated on the RealSR [1] and DRealSR [47] datasets, as shown in Table 2. We observe that with more inference steps, the image quality (MUSIQ) generally improves, while the fidelity (PSNR) slightly decreases. Fig. 5 presents the visual comparisons, where our SR flow model produces more accurate predictions, with further structure and overall quality enhancement as addi + +Table 2. Ablation of SR flow. SR flow mapping LR to HR outperforms alternative flow formulations that start from noise or ResShift [59]-style noised LR on RealSR and DRealSR. + +
#StepsMethodRealSRDRealSR
PSNR ↑MUSIQ ↑PSNR ↑MUSIQ ↑
4Noise→HR23.8366.0126.7562.76
noised LR→HR24.4765.1527.7461.76
SR Flow (ours)25.0068.0927.9865.05
1Noise→HR24.8265.1728.0759.34
noised LR→HR24.6765.2528.0360.64
SR Flow (ours)25.5167.4028.6562.53
+ +tional sampling steps (i.e., 4 steps) are applied. Notably, these results demonstrate that the model trained using our SR flow clearly outperforms other transition variants in both single-step and multi-step inference settings. + +Effects of consistency learning Next, we fine-tune the pre-trained SR flow model using the consistency objective described in Sec. 4.2. As shown in Table 3, enforcing self-consistency $\mathcal{L}_{cd}$ across neighboring time steps stabilizes intermediate representations and often reduces LPIPS. However, when the teacher's predictions deviate from the true HR manifold, these errors may propagate to the student model, leading to diminished IQA scores. We address this by introducing the HR regularization term in Eq. (6), which + +![](images/7e6582e6f7b95003933045da3b103124ac87bbaab71469fc16e0ec12e894b6ad.jpg) +LR Image + +![](images/595550517b40f3904fd2e631ea29f4c27835dcfb6e430484035ab7510409ef81.jpg) +SR Flow + +![](images/e60e498a9a0c3ab24db64f162977c46ba25056f10c7c04433d8c8417136ac7fc.jpg) +w/ $\mathcal{L}_{cd}$ + +![](images/67a8ea22fe8df4e90e5ea80f6cd078fbd28bbe51347f61ccc630a60fdc348873.jpg) +w/ $\mathcal{L}_{hrcd}$ +Figure 6. Ablation of consistency learning. Our HR-regularized consistency learning effectively reduces distortions in SR outputs while producing high-quality and sharper results (zoom in). + +Table 3. Ablation of consistency learning. Fine-tuned SR flow models with consistency objectives are evaluated on DRealSR. + +
#StepsMethodPSNR ↑LPIPS ↓MUSIQ ↑MANIQA ↑CLIPQA ↑
4SR Flow27.980.276365.050.62960.6681
w/ Lcd28.640.265461.820.60420.6180
w/ Lhrcd28.440.279965.710.63740.6709
1SR Flow28.650.282162.530.58950.6545
w/ Lcd28.590.260961.690.59490.6391
w/ Lhr28.850.280562.460.58710.6589
w/ Lhrcd28.620.283064.770.61560.6748
+ +directly aligns student predictions $f_{\theta}(\hat{X}_t^\phi, t)$ with real HR images, mitigating the teacher's approximation errors. This extra ground-truth constraint helps the model recover fine textural details, which boosts IQA metrics (e.g., MUSIQ and MANIQA) while preserves perceptual fidelity. Consistency learning enhances the robustness of flow-based inference by distilling multi-step capabilities into a single step, allowing for faster sampling without sacrificing quality. Additionally, HR regularization alone (w/ $\mathcal{L}_{hr}$ ) provides minimal IQA gains. This highlights the importance of consistency learning and its synergy with HR regularization for acceleration in SR flow. Fig. 6 presents visual comparisons, showing that consistency learning helps resolve distorted textures in the baseline SR flow model. Notably, HR-regularized consistency learning generates more realistic and sharper results compared to its baseline. + +Analysis of Fast-Slow Time Sampling We evaluate the effectiveness of fast-slow time sampling by comparing it to the $N$ -interval scheduler [38] and by varying the number of fast-scheduler timesteps, while fixing the slow scheduler fixed at 1000 timesteps. When the fast scheduler uses only one timestep, we set $t + \Delta t = 1$ and then sample $t$ from the slow scheduler (see Eq. (3)). As shown in Table 4, increasing $\Delta t$ (thus reducing the number of intervals) within a reasonable range improves IQA metrics. A plausible explanation is that the slightly larger perturbations introduced at intermediate predictions $\hat{X}_t^\phi$ from $X_{t}$ encourage the model to be more robust to distribution shifts; see Fig. 3. This, in turn, reduces variance in its estimates of + +![](images/59329c91986bef0bac8120c2cafb9aeb6f6b964a5bfc1a3082f28ab5456606aa.jpg) +LR Image + +![](images/780bcaa054749caf97071e120c0425b2e0cd49781c665e6aaa0aae205ee8932e.jpg) +w/o $\mathcal{L}_{hrcd}$ +Figure 7. Ablation of training loss. + +![](images/81cb7bb39ac06e52da70564e0982e7baa84256e20ab80b984a683471f6a04282.jpg) +w/o $\mathcal{L}_{adv}$ + +![](images/eaa2d2c2b0495c668cf91ce3d829cc23eb13afe83f9c0838f74b81af34fb7ffc.jpg) +w/o $\mathcal{L}_{iqa}$ + +![](images/de9f49c534a1ad7105b677ceb26274e9349c8f5b32862bb613657d2966bb3ea7.jpg) +Ours + +Table 4. Ablation of fast-slow time sampling. We compare our fast-slow sampling approach with the $N$ -interval method on DRealSR. The number of $N$ and fast-scheduler timesteps are indicated in parentheses. + +
MethodPSNR ↑LPIPS ↓MUSIQ ↑MANIQA ↑CLIPiQA ↑
N-Interval (50)28.610.280762.480.59140.6599
N-Interval (18)28.740.279262.900.59720.6605
N-Interval (4)28.660.281863.030.59720.6617
Slow Only (1000)28.780.280162.720.59160.6586
Fast-Slow (8)28.540.276064.010.60980.6634
Fast-Slow (4)28.620.283064.770.61560.6748
Fast-Slow (1)28.790.283563.010.59480.6605
+ +Table 5. Ablation of training loss on DRealSR. + +
MethodPSNR ↑LPIPS ↓MUSIQ ↑MANIQA ↑CLIPiQA ↑
w/o Lhred28.520.301364.420.60610.6954
w/o Ladv28.450.296165.290.61540.7099
w/o Liaq28.640.295864.830.61110.7012
Ours28.500.297565.460.61720.7074
+ +the ODE trajectory's starting point, i.e., HR, resulting in sharper restorations. Moreover, compared to the $N$ -interval strategy or slow-only scheduling, our fast-slow scheduler further boosts IQA metrics while maintaining comparable fidelity. This is because the slow scheduler ensures fine-grained SR flow estimation, whereas the fast scheduler enables efficient sampling. Based on these findings, we set the fast scheduler to 4 timesteps throughout our experiments. + +Influence of loss function In addition to flow and consistency learning, we incorporate a GAN loss and an image quality alignment loss to further enhance SR performance, as detailed in Eq. (10). As demonstrated in Table 5 and Fig. 7, these losses all improve no-reference metrics and visual quality. The ablation highlights that eliminating any of these losses results in reduced SR quality, with consistency learning contributing a substantial improvement. + +# 6. Conclusion + +This paper presents FlowSR, a new approach for efficient one-step image super-resolution. FlowSR reformulates SR as a rectified flow, leveraging the strengths of iterative generative modeling. To enable high-quality single-step inference, we incorporate consistency learning and devise HR regularization to address its distillation target drifting issue. Additionally, a fast-slow time scheduling strategy is designed to enhance the efficiency and robustness of the consistency SR flow model. FlowSR contributes to the advancement of efficient real-world SR applications. + +# Acknowledgement + +The work described in this paper was supported in part by the National Key R&D Program of China (Grant No. 2023YFE0202700), Research Grants Council of the Hong Kong Special Administrative Region, China, under Project 14200824; and by the Hong Kong Innovation and Technology Fund, under Project MHP/092/22. + +# References + +[1] Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang. Toward real-world single image super-resolution: A new benchmark and a new model. In ICCV, 2019. 6, 7 +[2] Bin Chen, Gehui Li, Rongyuan Wu, Xindong Zhang, Jie Chen, Jian Zhang, and Lei Zhang. Adversarial diffusion compression for real-world image super-resolution. In CVPR, 2025. 2 +[3] Qinpeng Cui, Yixuan Liu, Xinyi Zhang, Qiqi Bao, Zhongdao Wang, Qingmin Liao, Li Wang, Tian Lu, and Emad Barsoum. Taming diffusion prior for image super-resolution with domain shift sdes. NeurIPS, 2024. 1, 3, 4, 6 +[4] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. TPAMI, 2020. 6 +[5] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image super-resolution. In ECCV, 2014. 2 +[6] Linwei Dong, Qingnan Fan, Yihong Guo, Zhonghao Wang, Qi Zhang, Jinwei Chen, Yawei Luo, and Changqing Zou. Tsd-sr: One-step diffusion with target score distillation for real-world image super-resolution. In CVPR, 2025. 2 +[7] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. NeurIPS, 2014. 2 +[8] Xiao He, Huaao Tang, Zhijun Tu, Junchao Zhang, Kun Cheng, Hanting Chen, Yong Guo, Mingrui Zhu, Nannan Wang, Xinbo Gao, et al. One step diffusion-based super-resolution with time-aware distillation. arXiv preprint arXiv:2408.07476, 2024. 2 +[9] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. *NeruIPS*, 2017. 6 +[10] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. *NeruIPS*, 2020. 1, 2 +[11] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 6 +[12] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 6 +[13] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In ICCV, 2021. 6 + +[14] Dongjun Kim, Chieh-Hsin Lai, Wei-Hsiang Liao, Naoki Murata, Yuhta Takida, Toshimitsu Uesaka, Yutong He, Yuki Mitsufuji, and Stefano Ermon. Consistency trajectory models: Learning probability flow ode trajectory of diffusion. In ICLR, 2024. 3 +[15] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In CVPR, 2017. 2 +[16] Jianze Li, Jiezhang Cao, Zichen Zou, Xiongfei Su, Xin Yuan, Yulun Zhang, Yong Guo, and Xiaokang Yang. Distillation-free one-step diffusion for real-world image super-resolution. arXiv preprint arXiv:2410.04224, 2024. 2 +[17] Yawei Li, Kai Zhang, Jingyun Liang, Jiezhang Cao, Ce Liu, Rui Gong, Yulun Zhang, Hao Tang, Yun Liu, Denis Demandolx, et al. Lsdir: A large scale dataset for image restoration. In CVPR, 2023. 6 +[18] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swin transformer. In ICCV Workshop, 2021. 2 +[19] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Yu Qiao, Wanli Ouyang, and Chao Dong. Diffir: Toward blind image restoration with generative diffusion prior. In ECCV, 2024. 2, 6 +[20] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. In ICLR, 2023. 4 +[21] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. In ICLR, 2023. 2, 3, 4 +[22] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, et al. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. In ICLR, 2024. 3 +[23] Eric Luhman and Troy Luhman. Knowledge distillation in iterative generative models for improved sampling speed. arXiv preprint arXiv:2101.02388, 2021. 2 +[24] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 3 +[25] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a "completely blind" image quality analyzer. IEEE Signal processing letters, 2012. 6 +[26] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In AAAI, 2024. 2 +[27] Mehdi Noroozi, Isma Hadji, Brais Martinez, Adrian Bulat, and Georgios Tzimiropoulos. You only need one step: Fast super-resolution with stable diffusion via scale distillation. In ECCV, 2024. 2 +[28] Long Peng, Wenbo Li, Renjing Pei, Jingjing Ren, Jiaqi Xu, Yang Wang, Yang Cao, and Zheng-Jun Zha. Towards realistic data generation for real-world super-resolution. In ICLR, 2025. 2 + +[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 5 +[30] Jingjing Ren, Wenbo Li, Haoyu Chen, Renjing Pei, Bin Shao, Yong Guo, Long Peng, Fenglong Song, and Lei Zhu. Ultrapixel: Advancing ultra high-resolution image synthesis to new peaks. NeurIPS, 2024. 2 +[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2, 5, 6 +[32] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. TPAMI, 2022. 2 +[33] Axel Sauer, Frederic Boesel, Tim Dockhorn, Andreas Blattmann, Patrick Esser, and Robin Rombach. Fast high-resolution image synthesis with latent adversarial diffusion distillation. In SIGGRAPH Asia, 2024. 5 +[34] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In ECCV, 2024. 2 +[35] Johannes Schusterbauer, Ming Gui, Pingchuan Ma, Nick Stracke, Stefan Andreas Baumann, Vincent Tao Hu, and Björn Ommer. Boosting latent diffusion with flow matching. In ECCV, 2024. 3, 4 +[36] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 1 +[37] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 1, 2 +[38] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In ICML, 2023. 2, 3, 5, 8 +[39] Haoze Sun, Wenbo Li, Jianzhuang Liu, Haoyu Chen, Renjing Pei, Xueyi Zou, Youliang Yan, and Yujiu Yang. Coser: Bridging image and language for cognitive super-resolution. In CVPR, 2024. 2 +[40] Fu-Yun Wang, Zhaoyang Huang, Alexander Bergman, Dazhong Shen, Peng Gao, Michael Lingelbach, Keqiang Sun, Weikang Bian, Guanglu Song, Yu Liu, et al. Phased consistency models. NeurIPS, 2024. 3, 5 +[41] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In AAAI, 2023. 6 +[42] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. IJCV, 2024. 1, 2, 6 +[43] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 6 +[44] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In ICCV Workshop, 2021. 2, 6 + +[45] Yufei Wang, Wenhan Yang, Xinyuan Chen, Yaohui Wang, Lanqing Guo, Lap-Pui Chau, Ziwei Liu, Yu Qiao, Alex C Kot, and Bihan Wen. Sinsr: diffusion-based image superresolution in a single step. In CVPR, 2024. 1, 2, 6 +[46] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. NeurIPS, 2023. 2 +[47] Pengxu Wei, Ziwei Xie, Hannan Lu, Zongyuan Zhan, Qixiang Ye, Wangmeng Zuo, and Liang Lin. Component divide-and-conquer for real-world image super-resolution. In ECCV, 2020. 6, 7 +[48] Rongyuan Wu, Lingchen Sun, Zhiyuan Ma, and Lei Zhang. One-step effective diffusion network for real-world image super-resolution. NeurIPS, 2024. 1, 2, 6 +[49] Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. Seesr: Towards semantics-aware real-world image super-resolution. In CVPR, 2024. 1, 2, 6 +[50] Zhisheng Xiao, Karsten Kreis, and Arash Vahdat. Tackling the generative learning trilemma with denoising diffusion gans. In ICLR, 2022. 2 +[51] Rui Xie, Ying Tai, Chen Zhao, Kai Zhang, Zhenyu Zhang, Jun Zhou, Xiaogian Ye, Qian Wang, and Jian Yang. Addsr: Accelerating diffusion-based blind superresolution with adversarial diffusion distillation. arXiv preprint arXiv:2404.01717, 2024. 2 +[52] Hanshu Yan, Xingchao Liu, Jiachun Pan, Jun Hao Liew, Qiang Liu, and Jiashi Feng. Perflow: Piecewise rectified flow as universal plug-and-play accelerator. NeurIPS, 2024. 3 +[53] Ling Yang, Zixiang Zhang, Zhilong Zhang, Xingchao Liu, Minkai Xu, Wentao Zhang, Chenlin Meng, Stefano Ermon, and Bin Cui. Consistency flow matching: Defining straight flows with velocity consistency. arXiv preprint arXiv:2407.02398, 2024. 3 +[54] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Maniaq: Multi-dimension attention network for no-reference image quality assessment. In CVPR, 2022. 6 +[55] Tao Yang, Rongyuan Wu, Peiran Ren, Xuansong Xie, and Lei Zhang. Pixel-aware stable diffusion for realistic image super-resolution and personalized stylization. In ECCV, 2024.6 +[56] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and Bill Freeman. Improved distribution matching distillation for fast image synthesis. NeurIPS, 2024. 5 +[57] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In CVPR, 2024. 2 +[58] Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. Scaling up to excellence: Practicing model scaling for photorealistic image restoration in the wild. In CVPR, 2024. 1, 2 +[59] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Reshift: Efficient diffusion model for image super + +resolution by residual shifting. NeurIPS, 2023. 1, 2, 4, 6, 7 +[60] Zongsheng Yue, Kang Liao, and Chen Change Loy. Arbitrary-steps image super-resolution via diffusion inversion. In CVPR, 2025. 2 +[61] Aiping Zhang, Zongsheng Yue, Renjing Pei, Wenqi Ren, and Xiaochun Cao. Degradation-guided one-step image super-resolution with diffusion priors. arXiv preprint arXiv:2409.17058, 2024. 1, 2 +[62] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In ICCV, 2021. 2 +[63] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2 +[64] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5, 6 \ No newline at end of file diff --git a/2025/Fast Image Super-Resolution via Consistency Rectified Flow/images.zip b/2025/Fast Image Super-Resolution via Consistency Rectified Flow/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4db11aedf3aba0a2a9fcf6ef9c5a72c37589527e --- /dev/null +++ b/2025/Fast Image Super-Resolution via Consistency Rectified Flow/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f91de3c38159cfc5066908828c4a2dc4b884b7d283d102e05245d8994ba9eb00 +size 775582 diff --git a/2025/Fast Image Super-Resolution via Consistency Rectified Flow/layout.json b/2025/Fast Image Super-Resolution via Consistency Rectified Flow/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ac6cab0b3be9d99053d15b98f2a8bc25631ec1c2 --- /dev/null +++ b/2025/Fast Image Super-Resolution via Consistency Rectified Flow/layout.json @@ -0,0 +1,13311 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 119, + 103, + 492, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 103, + 492, + 121 + ], + "spans": [ + { + "bbox": [ + 119, + 103, + 492, + 121 + ], + "type": "text", + "content": "Fast Image Super-Resolution via Consistency Rectified Flow" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 87, + 142, + 522, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 142, + 522, + 171 + ], + "spans": [ + { + "bbox": [ + 87, + 142, + 522, + 171 + ], + "type": "text", + "content": "Jiaqi " + }, + { + "bbox": [ + 87, + 142, + 522, + 171 + ], + "type": "inline_equation", + "content": "\\mathrm{Xu}^{1,2*}" + }, + { + "bbox": [ + 87, + 142, + 522, + 171 + ], + "type": "text", + "content": ", Wenbo " + }, + { + "bbox": [ + 87, + 142, + 522, + 171 + ], + "type": "inline_equation", + "content": "\\mathrm{Li}^{2\\dagger}" + }, + { + "bbox": [ + 87, + 142, + 522, + 171 + ], + "type": "text", + "content": ", Haoze " + }, + { + "bbox": [ + 87, + 142, + 522, + 171 + ], + "type": "inline_equation", + "content": "\\mathrm{Sun}^2" + }, + { + "bbox": [ + 87, + 142, + 522, + 171 + ], + "type": "text", + "content": ", Fan " + }, + { + "bbox": [ + 87, + 142, + 522, + 171 + ], + "type": "inline_equation", + "content": "\\mathrm{Li}^2" + }, + { + "bbox": [ + 87, + 142, + 522, + 171 + ], + "type": "text", + "content": ", Zhixin " + }, + { + "bbox": [ + 87, + 142, + 522, + 171 + ], + "type": "inline_equation", + "content": "\\mathrm{Wang}^2" + }, + { + "bbox": [ + 87, + 142, + 522, + 171 + ], + "type": "text", + "content": ", Long " + }, + { + "bbox": [ + 87, + 142, + 522, + 171 + ], + "type": "inline_equation", + "content": "\\mathrm{Peng}^2" + }, + { + "bbox": [ + 87, + 142, + 522, + 171 + ], + "type": "text", + "content": ", Jingjing " + }, + { + "bbox": [ + 87, + 142, + 522, + 171 + ], + "type": "inline_equation", + "content": "\\mathrm{Ren}^3" + }, + { + "bbox": [ + 87, + 142, + 522, + 171 + ], + "type": "text", + "content": ", Haoran " + }, + { + "bbox": [ + 87, + 142, + 522, + 171 + ], + "type": "inline_equation", + "content": "\\mathrm{Yang}^1" + }, + { + "bbox": [ + 87, + 142, + 522, + 171 + ], + "type": "text", + "content": ", Xiaowei " + }, + { + "bbox": [ + 87, + 142, + 522, + 171 + ], + "type": "inline_equation", + "content": "\\mathrm{Hu}^4" + }, + { + "bbox": [ + 87, + 142, + 522, + 171 + ], + "type": "text", + "content": ", Renjing " + }, + { + "bbox": [ + 87, + 142, + 522, + 171 + ], + "type": "inline_equation", + "content": "\\mathrm{Pei}^{2\\dagger}" + }, + { + "bbox": [ + 87, + 142, + 522, + 171 + ], + "type": "text", + "content": ", and Pheng-Ann " + }, + { + "bbox": [ + 87, + 142, + 522, + 171 + ], + "type": "inline_equation", + "content": "\\mathrm{Heng}^1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 174, + 463, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 174, + 463, + 202 + ], + "spans": [ + { + "bbox": [ + 146, + 174, + 463, + 202 + ], + "type": "text", + "content": "1The Chinese University of Hong Kong 2Huawei Noah's Ark Lab 3HKUST (GZ) 4South China University of Technology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 229, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 229, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 229, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 255, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 255, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 54, + 255, + 296, + 544 + ], + "type": "text", + "content": "Diffusion models (DMs) have demonstrated remarkable success in real-world image super-resolution (SR), yet their reliance on time-consuming multi-step sampling largely hinders their practical applications. While recent efforts have introduced few- or single-step solutions, existing methods either inefficiently model the process from noisy input or fail to fully exploit iterative generative priors, compromising the fidelity and quality of the reconstructed images. To address this issue, we propose FlowSR, a novel approach that reformulates the SR problem as a rectified flow from low-resolution (LR) to high-resolution (HR) images. Our method leverages an improved consistency learning strategy to enable high-quality SR in a single step. Specifically, we refine the original consistency distillation process by incorporating HR regularization, ensuring that the learned SR flow not only enforces self-consistency but also converges precisely to the ground-truth HR target. Furthermore, we introduce a fast-slow scheduling strategy, where adjacent timesteps for consistency learning are sampled from two distinct schedulers: a fast scheduler with fewer timesteps to improve efficiency, and a slow scheduler with more timesteps to capture fine-grained texture details. Extensive experiments demonstrate that FlowSR achieves outstanding performance in both efficiency and image quality." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 567, + 135, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 135, + 580 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 135, + 580 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 587, + 296, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 296, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 296, + 684 + ], + "type": "text", + "content": "Real-world image super-resolution (SR) aims to reconstruct high-resolution (HR) images from their low-resolution (LR) counterparts while simultaneously removing unknown degradations. With the remarkable success of diffusion models (DMs) [10, 37], SR methods leveraging these models—particularly those built on powerful text-to-image (T2I) models like Stable Diffusion (SD) [31]—have demonstrated outstanding performance [42, 49, 58]. However," + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 228, + 376, + 331 + ], + "blocks": [ + { + "bbox": [ + 317, + 228, + 376, + 331 + ], + "lines": [ + { + "bbox": [ + 317, + 228, + 376, + 331 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 376, + 331 + ], + "type": "image", + "image_path": "9e40a1bbe783c249056bf060a0e4fe9007e9106282587e07ec741ee5b370fc60.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 336, + 332, + 349, + 342 + ], + "lines": [ + { + "bbox": [ + 336, + 332, + 349, + 342 + ], + "spans": [ + { + "bbox": [ + 336, + 332, + 349, + 342 + ], + "type": "text", + "content": "LR" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 376, + 229, + 433, + 287 + ], + "blocks": [ + { + "bbox": [ + 376, + 229, + 433, + 287 + ], + "lines": [ + { + "bbox": [ + 376, + 229, + 433, + 287 + ], + "spans": [ + { + "bbox": [ + 376, + 229, + 433, + 287 + ], + "type": "image", + "image_path": "04211c3713bb2157bb9098e89025f7dd7483308b81d3e26e1207556f740e36a2.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 376, + 288, + 433, + 331 + ], + "blocks": [ + { + "bbox": [ + 376, + 288, + 433, + 331 + ], + "lines": [ + { + "bbox": [ + 376, + 288, + 433, + 331 + ], + "spans": [ + { + "bbox": [ + 376, + 288, + 433, + 331 + ], + "type": "image", + "image_path": "5c730eb00a9e8e280a5248bfc3ef3d4093360c5c8a0b254d7b8dd5b43568c0a1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 394, + 332, + 416, + 342 + ], + "lines": [ + { + "bbox": [ + 394, + 332, + 416, + 342 + ], + "spans": [ + { + "bbox": [ + 394, + 332, + 416, + 342 + ], + "type": "text", + "content": "1-step" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 434, + 229, + 492, + 287 + ], + "blocks": [ + { + "bbox": [ + 434, + 229, + 492, + 287 + ], + "lines": [ + { + "bbox": [ + 434, + 229, + 492, + 287 + ], + "spans": [ + { + "bbox": [ + 434, + 229, + 492, + 287 + ], + "type": "image", + "image_path": "61e635f8088d6668e4b0bcc9fdf3e3510f5d9aaa03f499ca25f2baa5a738310a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 434, + 288, + 491, + 331 + ], + "blocks": [ + { + "bbox": [ + 434, + 288, + 491, + 331 + ], + "lines": [ + { + "bbox": [ + 434, + 288, + 491, + 331 + ], + "spans": [ + { + "bbox": [ + 434, + 288, + 491, + 331 + ], + "type": "image", + "image_path": "07628966f945812b770dca43a957ffd6e3d71db2bcd5402911110f120610080f.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 451, + 332, + 474, + 342 + ], + "lines": [ + { + "bbox": [ + 451, + 332, + 474, + 342 + ], + "spans": [ + { + "bbox": [ + 451, + 332, + 474, + 342 + ], + "type": "text", + "content": "4-step" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 492, + 229, + 550, + 287 + ], + "blocks": [ + { + "bbox": [ + 492, + 229, + 550, + 287 + ], + "lines": [ + { + "bbox": [ + 492, + 229, + 550, + 287 + ], + "spans": [ + { + "bbox": [ + 492, + 229, + 550, + 287 + ], + "type": "image", + "image_path": "a1f8bc0202f25b9aac971afa920984f5bdc92d07c1552ed072191f50fdf81acf.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 492, + 288, + 550, + 331 + ], + "blocks": [ + { + "bbox": [ + 492, + 288, + 550, + 331 + ], + "lines": [ + { + "bbox": [ + 492, + 288, + 550, + 331 + ], + "spans": [ + { + "bbox": [ + 492, + 288, + 550, + 331 + ], + "type": "image", + "image_path": "aa55a7b9cbf5adeee59ea0a73ab91d51d81f58bb0a06fa8d5e89535c57d1e3cd.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 493, + 332, + 548, + 342 + ], + "lines": [ + { + "bbox": [ + 493, + 332, + 548, + 342 + ], + "spans": [ + { + "bbox": [ + 493, + 332, + 548, + 342 + ], + "type": "text", + "content": "distilled, 1-step" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 317, + 350, + 550, + 430 + ], + "blocks": [ + { + "bbox": [ + 317, + 350, + 550, + 430 + ], + "lines": [ + { + "bbox": [ + 317, + 350, + 550, + 430 + ], + "spans": [ + { + "bbox": [ + 317, + 350, + 550, + 430 + ], + "type": "image", + "image_path": "98c21e30f8e91eb87338b126ec47bf7680cb0d4b8cab350af7b327b72aa46e84.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 440, + 555, + 496 + ], + "lines": [ + { + "bbox": [ + 313, + 440, + 555, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 440, + 555, + 496 + ], + "type": "text", + "content": "Figure 1. Our consistency SR flow model achieves high-quality single-step inference (distilled, 1-step) by distilling from multi-step higher-quality sampling process (top). We achieve this by formulating SR as a rectified flow that bridges LR and HR images, combined with improved consistency learning (bottom)." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 508, + 555, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 555, + 581 + ], + "type": "text", + "content": "these diffusion-based SR approaches require an iterative reverse sampling process that gradually refines noisy inputs into HR outputs. Despite advancements such as efficient ordinary differential equation (ODE) solvers like DDIM [36], the slow inference speed remains a bottleneck, limiting their practicality in real-world applications." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "content": "Recently, few-step or single-step SR methods derived from diffusion models are designed [3, 45, 48, 59, 61]. Several studies focus on designing more efficient diffusion processes for SR. ResShift [59] speeds up diffusion by shifting the LR-HR residual into the Markov chain and reduces sampling to 15 steps, while DoSSR [3] introduces a diffusion process more compatible with pre-trained DMs. However, they rely on DDPM [10], which uses a more curved transition trajectory and starts with a noise-perturbed LR image that loses crucial information, ultimately limiting performance. In parallel, another line of work tar" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 205, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 205, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 205, + 703 + ], + "type": "text", + "content": "*Part of work done as an intern at Huawei." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11755" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "content": "gets single-step SR by learning the LR-HR mapping directly. For example, SinSR [45] learns one-step SR prediction by leveraging the teacher's output from ResShift [59] and ground-truth HR as target, whereas OSEDiff adapts pre-trained SD for SR and refines one-step prediction with the VSD loss [46]. However, these methods fail to fully harness the advantages of iterative generative modeling, which naturally facilitates high-quality texture synthesis." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 170, + 295, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 170, + 295, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 170, + 295, + 324 + ], + "type": "text", + "content": "In this work, we present FlowSR, which unifies rectified flow with consistency models (CMs) to enable efficient, single-step image super-resolution. As illustrated in Fig. 1, we reformulate SR as a rectified flow [21], which establishes a simple and straight ODE-based mapping between LR and HR images. By explicitly modeling this trajectory, FlowSR learns fine-grained LR-to-HR transformations for better SR quality and facilitates fast sampling. Building upon this formulation, we further leverage consistency models [38] to enhance single-step inference. Enforcing consistency across points on the same SR flow trajectory distills multi-step higher-quality restoration into fewer steps that reach to the same HR result." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 327, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 295, + 590 + ], + "type": "text", + "content": "However, the naive consistency distillation (CD) objective in CMs is suboptimal for SR flow, where the SR task demands both high quality and high fidelity. While CD enforces self-consistency across discretization steps, there is no guarantee that the final distillation target aligns well with the ground-truth HR. To address this, we propose HR-regularized consistency learning, which explicitly requires the model's predictions to match real HR images. This additional constraint mitigates teacher-induced errors in the distillation target and enhances the reconstruction of fine-grained details. To further improve efficiency and robustness of the consistency SR flow model, we introduce a fast-slow time scheduling strategy. Rather than sampling distillation timestep pairs from the two boundaries of a discretized interval, we sample adjacent timesteps from distinct \"fast\" and \"slow\" schedulers. The fast scheduler uses fewer timesteps to facilitate efficient inference, while the slow scheduler employs more granular steps to maintain alignment with the SR flow objectives. This mixed sampling introduces large and flexible jumps with mild perturbations into the HR regularization steps, allowing for a broader coverage of SR flow trajectories." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 592, + 294, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 592, + 294, + 675 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 294, + 675 + ], + "type": "text", + "content": "To train our consistency SR flow model, we first finetune a pre-trained SD model to align with the SR flow objective, followed by consistency SR flow distillation. Additionally, we incorporate a GAN [7] loss and an image quality alignment loss, with the latter promoting desirable text-described attributes in the restored images, thereby improving the SR quality." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 677, + 294, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 294, + 714 + ], + "type": "text", + "content": "As illustrated in Fig. 1, our model leverages the proposed consistency SR flow distillation to transfer the quality improvement typically achieved with multiple sampling" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "content": "steps into a single step. Experiments on real-world datasets demonstrate the superiority of FlowSR. Our core contribution lie in a new paradigm to solve the one-step SR problem. First, we explore efficient flow modeling for SR and identify potential challenges when incorporating consistency models. Second, we introduce several techniques, including HR regularization in consistency learning and a fast-slow time scheduling to enhance one-step SR performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 178, + 400, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 178, + 400, + 190 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 400, + 190 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 198, + 449, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 449, + 211 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 449, + 211 + ], + "type": "text", + "content": "2.1. Image Super-Resolution" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 216, + 553, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 553, + 348 + ], + "type": "text", + "content": "Real-world image super-resolution (SR) aims to reconstruct high-resolution (HR) images from low-resolution (LR) inputs, tackling complex and unknown degradations like noise, blur, and compression artifacts. Since Dong et al. [5], numerous deep learning-based methods have been proposed to address the SR problem from different perspectives, including network design [18], model training [15], and degradation simulation [28, 44, 62]. Recently, diffusion models [10, 37] have achieved remarkable success in image generation [30, 31], which inspires their application to SR and demonstrate significant advancements [32, 42]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 350, + 553, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 553, + 542 + ], + "type": "text", + "content": "Diffusion Model-Based SR Conditioned on the LR image, diffusion model-based SR iteratively denoises towards the target HR image. The LR condition can be leveraged either by input concatenation [32] or via adapters [26, 63]. StableSR [42] based on Stable Diffusion [31], uses a trainable time-aware encoder to incorporates LR. Subsequent improvements are achieved through enhanced LR conditioning in DiffBIR [19], semantics-aware prompts in SeeSR [49], reference image generation in CoSeR [39], and scaled-up training in SUPIR [58]. However, these DM-based SR approaches, starting from noise, suffer from low inference speed, typically requiring 50-200 sampling steps. In contrast, ResShift [59] constructs an efficient diffusion model by shifting the residual between HR and LR. Denoising from the noised LR reduces the sampling steps to 15 but it still suffers from limited restoration quality." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 545, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 553, + 713 + ], + "type": "text", + "content": "Single-Step / Few-Step Image SR Fast diffusion model-based SR methods require only one or a few sampling steps during inference, typically achieved through knowledge distillation [23] or adversarial training [15, 50]. SinSR [45] derives a deterministic sampling process for ResShift and reduces its number of sampling steps to one using distillation. OSEDiff [48] applies variational score distillation [46, 57] to enhance one-step restoration quality by minimizing the KL-divergence between the distribution of its generated outputs and that of the pre-trained T2I model. AddSR [51] achieves 4-step inference by tailoring adversarial diffusion distillation [34] for SR. Similarly, other recent methods, accelerate the inference by using distillation [6, 8, 27] or GAN [2, 16, 60, 61]. On the other hand, more" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11756" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 68, + 541, + 240 + ], + "blocks": [ + { + "bbox": [ + 69, + 68, + 541, + 240 + ], + "lines": [ + { + "bbox": [ + 69, + 68, + 541, + 240 + ], + "spans": [ + { + "bbox": [ + 69, + 68, + 541, + 240 + ], + "type": "image", + "image_path": "f92d7b59b0a007023af77f66d56f8ecc8f168c539bb3637331196eca26a6f6f2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 244, + 555, + 281 + ], + "lines": [ + { + "bbox": [ + 54, + 244, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 54, + 244, + 555, + 281 + ], + "type": "text", + "content": "Figure 2. Overview of the training process. The consistency SR flow distills multi-step, high-quality SR capability into single-step inference, while HR regularization ensures the model converges to the high-resolution target. Times " + }, + { + "bbox": [ + 54, + 244, + 555, + 281 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 244, + 555, + 281 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 244, + 555, + 281 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 54, + 244, + 555, + 281 + ], + "type": "text", + "content": " are sampled alternately from fast and slow schedulers. Note that the flow loss " + }, + { + "bbox": [ + 54, + 244, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{flow}" + }, + { + "bbox": [ + 54, + 244, + 555, + 281 + ], + "type": "text", + "content": " and the consistency loss " + }, + { + "bbox": [ + 54, + 244, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{hrcd}" + }, + { + "bbox": [ + 54, + 244, + 555, + 281 + ], + "type": "text", + "content": " are computed for different samples within each batch." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 288, + 294, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 288, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 54, + 288, + 294, + 337 + ], + "type": "text", + "content": "efficient SR diffusion processes are constructed using domain shift [3] or flow matching [35], which allow few-step sampling. Yet, they still suffer from inefficient modeling designs and suboptimal performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 343, + 171, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 343, + 171, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 343, + 171, + 357 + ], + "type": "text", + "content": "2.2. Consistency Models" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 361, + 295, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 361, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 54, + 361, + 295, + 529 + ], + "type": "text", + "content": "Consistency Models (CMs) [38] learn to map any point on the ODE trajectory to its origin, which enable single-step generation and allow trade-offs between quality and computation through multi-step sampling. CMs can be implemented for modern T2I models in latent space [24]. Consistency Trajectory Models (CTMs) [14] mitigate the multi-step sampling issues in CMs by learning an any-to-any mapping from initial points to final points on the ODE trajectory. The Phased Consistency Model [40] partitions the entire ODE trajectory into multiple sub-trajectory phases and ensures consistency within each phase, while PeRFlow [52] straightens the sub-trajectories using the Reflow [22] operation. Additionally, Consistency Flow Matching [53] further enforces velocity field consistency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 540, + 141, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 141, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 141, + 553 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 559, + 145, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 145, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 145, + 571 + ], + "type": "text", + "content": "3.1. Rectified Flow" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 577, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 577, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 577, + 296, + 714 + ], + "type": "text", + "content": "Rectified flow [21] is an ODE-based generative model that constructs a simple, straight trajectory to transform samples between two distributions: " + }, + { + "bbox": [ + 55, + 577, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\pi_0" + }, + { + "bbox": [ + 55, + 577, + 296, + 714 + ], + "type": "text", + "content": " (e.g., data) and " + }, + { + "bbox": [ + 55, + 577, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\pi_1" + }, + { + "bbox": [ + 55, + 577, + 296, + 714 + ], + "type": "text", + "content": " (e.g., noise). Rectified flow defines a linear interpolation path between observed samples " + }, + { + "bbox": [ + 55, + 577, + 296, + 714 + ], + "type": "inline_equation", + "content": "X_0\\sim \\pi_0" + }, + { + "bbox": [ + 55, + 577, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 577, + 296, + 714 + ], + "type": "inline_equation", + "content": "X_{1}\\sim \\pi_{1}" + }, + { + "bbox": [ + 55, + 577, + 296, + 714 + ], + "type": "text", + "content": " as: " + }, + { + "bbox": [ + 55, + 577, + 296, + 714 + ], + "type": "inline_equation", + "content": "X_{t} = tX_{1} + (1 - t)X_{0}" + }, + { + "bbox": [ + 55, + 577, + 296, + 714 + ], + "type": "text", + "content": " with time " + }, + { + "bbox": [ + 55, + 577, + 296, + 714 + ], + "type": "inline_equation", + "content": "t\\in [0,1]" + }, + { + "bbox": [ + 55, + 577, + 296, + 714 + ], + "type": "text", + "content": ". The core idea is to learn a velocity field " + }, + { + "bbox": [ + 55, + 577, + 296, + 714 + ], + "type": "inline_equation", + "content": "v_{\\theta}(X_{t},t)" + }, + { + "bbox": [ + 55, + 577, + 296, + 714 + ], + "type": "text", + "content": ", parameterized by a neural network with weights " + }, + { + "bbox": [ + 55, + 577, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 577, + 296, + 714 + ], + "type": "text", + "content": ", that matches the derivative of this trajectory. This is achieved by optimizing the following objective: " + }, + { + "bbox": [ + 55, + 577, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\min_{\\theta}\\int_{0}^{1}\\mathbb{E}_{X_0\\sim \\pi_0,X_1\\sim \\pi_1}[\\| v_{\\theta}(X_t,t) - (X_1 - X_0)\\| ^2 ]dt," + }, + { + "bbox": [ + 55, + 577, + 296, + 714 + ], + "type": "text", + "content": " which encourages the learned velocity field to align with" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 288, + 555, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 555, + 373 + ], + "type": "text", + "content": "the direction of the straight path " + }, + { + "bbox": [ + 313, + 288, + 555, + 373 + ], + "type": "inline_equation", + "content": "X_{1} - X_{0}" + }, + { + "bbox": [ + 313, + 288, + 555, + 373 + ], + "type": "text", + "content": ". Once trained, samples from " + }, + { + "bbox": [ + 313, + 288, + 555, + 373 + ], + "type": "inline_equation", + "content": "\\pi_{1}" + }, + { + "bbox": [ + 313, + 288, + 555, + 373 + ], + "type": "text", + "content": " can be transformed into " + }, + { + "bbox": [ + 313, + 288, + 555, + 373 + ], + "type": "inline_equation", + "content": "\\pi_{0}" + }, + { + "bbox": [ + 313, + 288, + 555, + 373 + ], + "type": "text", + "content": " by solving the ODE: " + }, + { + "bbox": [ + 313, + 288, + 555, + 373 + ], + "type": "inline_equation", + "content": "dX_{t} = v_{\\theta}(X_{t}, t)dt" + }, + { + "bbox": [ + 313, + 288, + 555, + 373 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 288, + 555, + 373 + ], + "type": "inline_equation", + "content": "X_{1} \\sim \\pi_{1}" + }, + { + "bbox": [ + 313, + 288, + 555, + 373 + ], + "type": "text", + "content": ". A key advantage of rectified flow lies in its straight trajectories. By design, the optimal velocity field corresponds to a constant-speed flow between " + }, + { + "bbox": [ + 313, + 288, + 555, + 373 + ], + "type": "inline_equation", + "content": "X_{0}" + }, + { + "bbox": [ + 313, + 288, + 555, + 373 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 288, + 555, + 373 + ], + "type": "inline_equation", + "content": "X_{1}" + }, + { + "bbox": [ + 313, + 288, + 555, + 373 + ], + "type": "text", + "content": ", enabling efficient and stable sampling with only a few ODE steps." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 383, + 430, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 383, + 430, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 383, + 430, + 396 + ], + "type": "text", + "content": "3.2. Consistency Models" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 402, + 555, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 555, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 555, + 594 + ], + "type": "text", + "content": "Consistency models [38] are a class of generative models that enable high-quality sample generation with few computational steps. These models learn to map any point " + }, + { + "bbox": [ + 313, + 402, + 555, + 594 + ], + "type": "inline_equation", + "content": "(X_{t}, t)" + }, + { + "bbox": [ + 313, + 402, + 555, + 594 + ], + "type": "text", + "content": " along a probability flow ODE trajectory directly to its origin " + }, + { + "bbox": [ + 313, + 402, + 555, + 594 + ], + "type": "inline_equation", + "content": "X_{\\epsilon}" + }, + { + "bbox": [ + 313, + 402, + 555, + 594 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 402, + 555, + 594 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 402, + 555, + 594 + ], + "type": "text", + "content": " is a fixed small positive number denoting the trajectory's start. This capability is formalized through the self-consistency property: for any pair of points " + }, + { + "bbox": [ + 313, + 402, + 555, + 594 + ], + "type": "inline_equation", + "content": "(X_{t}, t)" + }, + { + "bbox": [ + 313, + 402, + 555, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 402, + 555, + 594 + ], + "type": "inline_equation", + "content": "(X_{t'}, t')" + }, + { + "bbox": [ + 313, + 402, + 555, + 594 + ], + "type": "text", + "content": " on the same ODE trajectory, the model " + }, + { + "bbox": [ + 313, + 402, + 555, + 594 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\cdot, t)" + }, + { + "bbox": [ + 313, + 402, + 555, + 594 + ], + "type": "text", + "content": " is trained to satisfy: " + }, + { + "bbox": [ + 313, + 402, + 555, + 594 + ], + "type": "inline_equation", + "content": "f_{\\theta}(X_{t}, t) = f_{\\theta}(X_{t}', t') = X_{\\epsilon}" + }, + { + "bbox": [ + 313, + 402, + 555, + 594 + ], + "type": "text", + "content": " for all " + }, + { + "bbox": [ + 313, + 402, + 555, + 594 + ], + "type": "inline_equation", + "content": "t, t' \\in [\\epsilon, 1]" + }, + { + "bbox": [ + 313, + 402, + 555, + 594 + ], + "type": "text", + "content": ". To ensure consistency during training, the boundary condition " + }, + { + "bbox": [ + 313, + 402, + 555, + 594 + ], + "type": "inline_equation", + "content": "f_{\\theta}(X_{\\epsilon}, \\epsilon) = X_{\\epsilon}" + }, + { + "bbox": [ + 313, + 402, + 555, + 594 + ], + "type": "text", + "content": " is enforced. Consistency models eliminate trajectory drift by mapping all intermediate states to a consistent origin, avoiding errors from iterative denoising steps. This enables them to generate high-fidelity samples even with large step sizes, such as directly mapping from " + }, + { + "bbox": [ + 313, + 402, + 555, + 594 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 313, + 402, + 555, + 594 + ], + "type": "text", + "content": " to the origin in a single step." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 608, + 397, + 622 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 608, + 397, + 622 + ], + "spans": [ + { + "bbox": [ + 313, + 608, + 397, + 622 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "content": "We leverage the generative capability of rectified flow for high-quality SR and explore consistency models to enable fast inference with fewer sampling steps (i.e., one step), without compromising quality. We define a rectified flow, termed SR flow, for image super-resolution in Sec. 4.1. We then introduce the improved consistency SR flow learning method in Sec. 4.2 and detail the training in Sec. 4.3." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "11757" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 179, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 179, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 179, + 83 + ], + "type": "text", + "content": "4.1. Rectified Flow for SR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 149 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 149 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 149 + ], + "type": "text", + "content": "SR Flow We consider Rectified Flow naturally aligns with image super-resolution, where high-resolution (HR) images " + }, + { + "bbox": [ + 55, + 89, + 296, + 149 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{HR}} \\sim \\pi_0" + }, + { + "bbox": [ + 55, + 89, + 296, + 149 + ], + "type": "text", + "content": " and their low-resolution (LR) counterparts " + }, + { + "bbox": [ + 55, + 89, + 296, + 149 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{LR}} \\sim \\pi_1" + }, + { + "bbox": [ + 55, + 89, + 296, + 149 + ], + "type": "text", + "content": ". We define the forward process as a straight path via a linear interpolation between them:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 117, + 154, + 295, + 167 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 154, + 295, + 167 + ], + "spans": [ + { + "bbox": [ + 117, + 154, + 295, + 167 + ], + "type": "interline_equation", + "content": "X _ {t} = (1 - t) X _ {\\mathrm {H R}} + t X _ {\\mathrm {L R}}, \\tag {1}", + "image_path": "8b118908da9a5f48d5d1e317e2e40720445474aabe19cbd8807e5426f83b4126.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 171, + 296, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 171, + 296, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 296, + 243 + ], + "type": "text", + "content": "here time " + }, + { + "bbox": [ + 55, + 171, + 296, + 243 + ], + "type": "inline_equation", + "content": "t\\in [0,1]" + }, + { + "bbox": [ + 55, + 171, + 296, + 243 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 171, + 296, + 243 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 55, + 171, + 296, + 243 + ], + "type": "text", + "content": " corresponding to " + }, + { + "bbox": [ + 55, + 171, + 296, + 243 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{HR}}\\in \\mathbb{E}^{H\\times W\\times 3}" + }, + { + "bbox": [ + 55, + 171, + 296, + 243 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 171, + 296, + 243 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 55, + 171, + 296, + 243 + ], + "type": "text", + "content": " corresponding to " + }, + { + "bbox": [ + 55, + 171, + 296, + 243 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{LR}}\\in \\mathbb{E}^{H\\times W\\times 3}" + }, + { + "bbox": [ + 55, + 171, + 296, + 243 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 171, + 296, + 243 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{LR}}" + }, + { + "bbox": [ + 55, + 171, + 296, + 243 + ], + "type": "text", + "content": " is upscaled to match the spatial dimensions of " + }, + { + "bbox": [ + 55, + 171, + 296, + 243 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{HR}}" + }, + { + "bbox": [ + 55, + 171, + 296, + 243 + ], + "type": "text", + "content": ". The learning objective of the SR neural network " + }, + { + "bbox": [ + 55, + 171, + 296, + 243 + ], + "type": "inline_equation", + "content": "v_{\\theta}" + }, + { + "bbox": [ + 55, + 171, + 296, + 243 + ], + "type": "text", + "content": " is to regress the conditional vector fields [20, 21] by following the direction " + }, + { + "bbox": [ + 55, + 171, + 296, + 243 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{LR}} - X_{\\mathrm{HR}}" + }, + { + "bbox": [ + 55, + 171, + 296, + 243 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 99, + 248, + 295, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 248, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 99, + 248, + 295, + 262 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {t, X _ {t}} \\| v _ {\\theta} \\left(X _ {t}, t\\right) - \\left(X _ {\\mathrm {L R}} - X _ {\\mathrm {H R}}\\right) \\| _ {2} ^ {2}. \\tag {2}", + "image_path": "9cf5d098652321e0b11f2c770854e42743a59fc3978f1c0bac7dd12495d255f0.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 266, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 296, + 422 + ], + "type": "text", + "content": "Intuitively, this SR flow establishes a smooth transition between HR and LR. During training, the SR flow model " + }, + { + "bbox": [ + 55, + 266, + 296, + 422 + ], + "type": "inline_equation", + "content": "v_{\\theta}" + }, + { + "bbox": [ + 55, + 266, + 296, + 422 + ], + "type": "text", + "content": " implicitly learns to invert degradations (e.g., blur, noise) and recover high-frequency details through intermediate refinements such as edge sharpening and texture synthesis. At inference, HR images are reconstructed via reverse sampling along the learned trajectory. Starting from " + }, + { + "bbox": [ + 55, + 266, + 296, + 422 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{LR}}" + }, + { + "bbox": [ + 55, + 266, + 296, + 422 + ], + "type": "text", + "content": ", we solve the reverse ODE using numerical methods like Euler solver: " + }, + { + "bbox": [ + 55, + 266, + 296, + 422 + ], + "type": "inline_equation", + "content": "X_{t'} = X_t + \\Delta t \\cdot v_{\\theta}(X_t, t)" + }, + { + "bbox": [ + 55, + 266, + 296, + 422 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 266, + 296, + 422 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 55, + 266, + 296, + 422 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 266, + 296, + 422 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 55, + 266, + 296, + 422 + ], + "type": "text", + "content": ". Notably, unlike diffusion-based SR methods that corrupt LR inputs with noise [3, 35, 59], our transition process directly starts from LR, which preserves most of the structural information in the LR image, enabling stable and efficient sampling." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 425, + 296, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 296, + 557 + ], + "type": "text", + "content": "Faster Inference with SR Flow A key advantage of SR flow is its flexibility in sampling steps. It supports faster inference by reducing the number of sampling steps, requiring as few as a single step that directly maps the LR image to its corresponding HR output: " + }, + { + "bbox": [ + 55, + 425, + 296, + 557 + ], + "type": "inline_equation", + "content": "\\hat{X}_{\\mathrm{HR}} = X_{\\mathrm{LR}} - 1 \\cdot v_{\\theta}(X_{\\mathrm{LR}}, 1)" + }, + { + "bbox": [ + 55, + 425, + 296, + 557 + ], + "type": "text", + "content": ". This is more easily achievable than the noise-to-image mapping or initializing from noisy inputs in ResShift [59], thanks to the strong correlation between LR and HR images. Additionally, SR flow also supports any number of sampling steps and produces better visual quality when using iterative multi-step sampling, as illustrated in Fig. 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 562, + 177, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 562, + 177, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 562, + 177, + 574 + ], + "type": "text", + "content": "4.2. Consistency SR Flow" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 578, + 296, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 296, + 663 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 296, + 663 + ], + "type": "text", + "content": "We enhance single-step super-resolution in SR flow models by distilling multi-step sampling capability (e.g., four steps) into one via consistency learning (Fig. 1). To ensure accurate HR reconstruction, we incorporate the HR target explicitly into consistency learning. Additionally, a fast-slow time sampling schedule is introduced to improve efficiency and robustness. Fig. 2 shows the overall training process." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Consistency Distillation in SR Flow We start with consistency distillation (CD) to distill the capability of a (pre-trained) teacher SR flow model " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "v_{\\phi}(X,t)" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ", enabling high-quality SR in fewer inference steps. Given timesteps " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 72, + 553, + 164 + ], + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 164 + ], + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 164 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 164 + ], + "type": "image", + "image_path": "32c0695adbb690cffb3c861d73585f794967f8771716d86b4ab48c68d3ad15e5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 165, + 364, + 212 + ], + "blocks": [ + { + "bbox": [ + 317, + 165, + 364, + 212 + ], + "lines": [ + { + "bbox": [ + 317, + 165, + 364, + 212 + ], + "spans": [ + { + "bbox": [ + 317, + 165, + 364, + 212 + ], + "type": "image", + "image_path": "ece13ce9fb4a39df413b03a6811fd38bff15f2cc7b47fff6208b1db6651bae20.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 317, + 212, + 364, + 258 + ], + "blocks": [ + { + "bbox": [ + 317, + 212, + 364, + 258 + ], + "lines": [ + { + "bbox": [ + 317, + 212, + 364, + 258 + ], + "spans": [ + { + "bbox": [ + 317, + 212, + 364, + 258 + ], + "type": "image", + "image_path": "4a1e98cf55e51444b16d07dc7af45db010b60ca8e4b635d85576d712a0eed804.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 332, + 260, + 345, + 273 + ], + "lines": [ + { + "bbox": [ + 332, + 260, + 345, + 273 + ], + "spans": [ + { + "bbox": [ + 332, + 260, + 345, + 273 + ], + "type": "inline_equation", + "content": "X_{t^{\\prime}}" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 364, + 165, + 411, + 212 + ], + "blocks": [ + { + "bbox": [ + 364, + 165, + 411, + 212 + ], + "lines": [ + { + "bbox": [ + 364, + 165, + 411, + 212 + ], + "spans": [ + { + "bbox": [ + 364, + 165, + 411, + 212 + ], + "type": "image", + "image_path": "201cc9c752b509fd9869df838e4bc1d998c1c7f57d1f0c00ed8347e3c49e5dba.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 364, + 212, + 411, + 258 + ], + "blocks": [ + { + "bbox": [ + 364, + 212, + 411, + 258 + ], + "lines": [ + { + "bbox": [ + 364, + 212, + 411, + 258 + ], + "spans": [ + { + "bbox": [ + 364, + 212, + 411, + 258 + ], + "type": "image", + "image_path": "6e2de18bc63fa8a34c7c4892b50c23112c56d457376c770854b4b5d93c171ef4.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 381, + 260, + 392, + 273 + ], + "lines": [ + { + "bbox": [ + 381, + 260, + 392, + 273 + ], + "spans": [ + { + "bbox": [ + 381, + 260, + 392, + 273 + ], + "type": "inline_equation", + "content": "X_{t}" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 411, + 165, + 459, + 212 + ], + "blocks": [ + { + "bbox": [ + 411, + 165, + 459, + 212 + ], + "lines": [ + { + "bbox": [ + 411, + 165, + 459, + 212 + ], + "spans": [ + { + "bbox": [ + 411, + 165, + 459, + 212 + ], + "type": "image", + "image_path": "56f16afb06b14c6d00c844a92b850f7a2f48c3ee491ae4b7d2882856f2d8d813.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 411, + 212, + 459, + 258 + ], + "blocks": [ + { + "bbox": [ + 411, + 212, + 459, + 258 + ], + "lines": [ + { + "bbox": [ + 411, + 212, + 459, + 258 + ], + "spans": [ + { + "bbox": [ + 411, + 212, + 459, + 258 + ], + "type": "image", + "image_path": "1cd07dc6e4e4c566461dbf8e6ccb4ae43c2984e0da7520a0371a91d4f75db23b.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 260, + 440, + 273 + ], + "lines": [ + { + "bbox": [ + 429, + 260, + 440, + 273 + ], + "spans": [ + { + "bbox": [ + 429, + 260, + 440, + 273 + ], + "type": "inline_equation", + "content": "\\hat{X}_t^\\phi" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 460, + 165, + 506, + 212 + ], + "blocks": [ + { + "bbox": [ + 460, + 165, + 506, + 212 + ], + "lines": [ + { + "bbox": [ + 460, + 165, + 506, + 212 + ], + "spans": [ + { + "bbox": [ + 460, + 165, + 506, + 212 + ], + "type": "image", + "image_path": "27edee5717286b1868303ddbc9b29dba251c32cf8da0420421ba49e585e1ad81.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 460, + 212, + 506, + 258 + ], + "blocks": [ + { + "bbox": [ + 460, + 212, + 506, + 258 + ], + "lines": [ + { + "bbox": [ + 460, + 212, + 506, + 258 + ], + "spans": [ + { + "bbox": [ + 460, + 212, + 506, + 258 + ], + "type": "image", + "image_path": "98fffe65e49318c271c277df25ae717964e996b77729f039ce3686a0b94dbf69.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 260, + 490, + 273 + ], + "lines": [ + { + "bbox": [ + 473, + 260, + 490, + 273 + ], + "spans": [ + { + "bbox": [ + 473, + 260, + 490, + 273 + ], + "type": "inline_equation", + "content": "\\hat{X}_{t\\rightarrow 0}^{\\theta^{-}}" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 279, + 554, + 326 + ], + "lines": [ + { + "bbox": [ + 313, + 279, + 554, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 554, + 326 + ], + "type": "text", + "content": "Figure 3. Illustration of HR-regularized consistency learning. Approximation errors may cause the distillation target " + }, + { + "bbox": [ + 313, + 279, + 554, + 326 + ], + "type": "inline_equation", + "content": "\\hat{X}_{t\\rightarrow 0}^{\\theta^{-}}" + }, + { + "bbox": [ + 313, + 279, + 554, + 326 + ], + "type": "text", + "content": " to deviate from the true high-quality HR target. HR regularization corrects this by enforcing HR alignment under mild perturbations." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 507, + 165, + 553, + 212 + ], + "blocks": [ + { + "bbox": [ + 507, + 165, + 553, + 212 + ], + "lines": [ + { + "bbox": [ + 507, + 165, + 553, + 212 + ], + "spans": [ + { + "bbox": [ + 507, + 165, + 553, + 212 + ], + "type": "image", + "image_path": "e3ac8bcb1b75a8ecfa075cfad810a7a01e0e93619f25a5feac9290a779ca8787.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 507, + 212, + 553, + 258 + ], + "blocks": [ + { + "bbox": [ + 507, + 212, + 553, + 258 + ], + "lines": [ + { + "bbox": [ + 507, + 212, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 507, + 212, + 553, + 258 + ], + "type": "image", + "image_path": "89bfce9f9f4e0b8fc38173cc8d81f9cefc07450ae8b495c63c6e7ea4598c994d.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 525, + 260, + 535, + 273 + ], + "lines": [ + { + "bbox": [ + 525, + 260, + 535, + 273 + ], + "spans": [ + { + "bbox": [ + 525, + 260, + 535, + 273 + ], + "type": "inline_equation", + "content": "X_0" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 333, + 554, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 333, + 554, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 333, + 554, + 370 + ], + "type": "inline_equation", + "content": "t^{\\prime} = t + \\Delta t" + }, + { + "bbox": [ + 313, + 333, + 554, + 370 + ], + "type": "text", + "content": " (where " + }, + { + "bbox": [ + 313, + 333, + 554, + 370 + ], + "type": "inline_equation", + "content": "\\Delta t > 0" + }, + { + "bbox": [ + 313, + 333, + 554, + 370 + ], + "type": "text", + "content": "), we first sample " + }, + { + "bbox": [ + 313, + 333, + 554, + 370 + ], + "type": "inline_equation", + "content": "X_{t'}" + }, + { + "bbox": [ + 313, + 333, + 554, + 370 + ], + "type": "text", + "content": " from the forward degradation process (Eq. (1)). The teacher model then estimates " + }, + { + "bbox": [ + 313, + 333, + 554, + 370 + ], + "type": "inline_equation", + "content": "\\hat{X}_t^\\phi" + }, + { + "bbox": [ + 313, + 333, + 554, + 370 + ], + "type": "text", + "content": " from its degraded counterpart " + }, + { + "bbox": [ + 313, + 333, + 554, + 370 + ], + "type": "inline_equation", + "content": "X_{t'}" + }, + { + "bbox": [ + 313, + 333, + 554, + 370 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 373, + 377, + 553, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 377, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 373, + 377, + 553, + 392 + ], + "type": "interline_equation", + "content": "\\hat {X} _ {t} ^ {\\phi} = X _ {t ^ {\\prime}} - \\Delta t \\cdot v _ {\\phi} \\left(X _ {t ^ {\\prime}}, t ^ {\\prime}\\right), \\tag {3}", + "image_path": "06eb0549025c01614536a4211ecc58e4837dec44413c368fc88a1fae36ca26b7.jpg" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 399, + 554, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 554, + 424 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 554, + 424 + ], + "type": "text", + "content": "where Euler solver is used. The consistency function " + }, + { + "bbox": [ + 313, + 399, + 554, + 424 + ], + "type": "inline_equation", + "content": "f_{\\theta}(X_t, t)" + }, + { + "bbox": [ + 313, + 399, + 554, + 424 + ], + "type": "text", + "content": " for deriving the origin is defined by:" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 370, + 431, + 553, + 445 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 431, + 553, + 445 + ], + "spans": [ + { + "bbox": [ + 370, + 431, + 553, + 445 + ], + "type": "interline_equation", + "content": "f _ {\\theta} \\left(X _ {t}, t\\right) = X _ {t} - t \\cdot v _ {\\theta} \\left(X _ {t}, t\\right), \\tag {4}", + "image_path": "06da801ea395e1939d48b8293c53dfa2a9a52a4c92e639e9656eb39d203dcd77.jpg" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 452, + 554, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 452, + 554, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 452, + 554, + 514 + ], + "type": "text", + "content": "which maps any degraded input " + }, + { + "bbox": [ + 313, + 452, + 554, + 514 + ], + "type": "inline_equation", + "content": "X_{t}" + }, + { + "bbox": [ + 313, + 452, + 554, + 514 + ], + "type": "text", + "content": " at timestep " + }, + { + "bbox": [ + 313, + 452, + 554, + 514 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 452, + 554, + 514 + ], + "type": "text", + "content": " to the HR image domain. The consistency distillation in SR flow learns a vector field " + }, + { + "bbox": [ + 313, + 452, + 554, + 514 + ], + "type": "inline_equation", + "content": "v_{\\theta}" + }, + { + "bbox": [ + 313, + 452, + 554, + 514 + ], + "type": "text", + "content": " by minimizing the distance between the prediction of the network " + }, + { + "bbox": [ + 313, + 452, + 554, + 514 + ], + "type": "inline_equation", + "content": "\\hat{X}_{t^{\\prime}\\to 0}^{\\theta} = f_{\\theta}(X_{t^{\\prime}},t^{\\prime})" + }, + { + "bbox": [ + 313, + 452, + 554, + 514 + ], + "type": "text", + "content": " and the distillation target " + }, + { + "bbox": [ + 313, + 452, + 554, + 514 + ], + "type": "inline_equation", + "content": "\\hat{X}_{t\\rightarrow 0}^{\\theta^{-}} = f_{\\theta^{-}}(\\hat{X}_{t}^{\\phi},t)" + }, + { + "bbox": [ + 313, + 452, + 554, + 514 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 358, + 521, + 553, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 521, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 358, + 521, + 553, + 536 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c d} = \\mathbb {E} \\left[ d \\left(f _ {\\theta} \\left(X _ {t ^ {\\prime}}, t ^ {\\prime}\\right), f _ {\\theta^ {-}} \\left(\\hat {X} _ {t} ^ {\\phi}, t\\right)\\right) \\right], \\tag {5}", + "image_path": "c938582d8c630c54fce4aa2afeb2f0ee7a9913aa8621ebec04880b80fc1cc49e.jpg" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 543, + 554, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 543, + 554, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 543, + 554, + 591 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 543, + 554, + 591 + ], + "type": "inline_equation", + "content": "d(\\cdot, \\cdot)" + }, + { + "bbox": [ + 313, + 543, + 554, + 591 + ], + "type": "text", + "content": " is a distance metric, and " + }, + { + "bbox": [ + 313, + 543, + 554, + 591 + ], + "type": "inline_equation", + "content": "\\theta^{-}" + }, + { + "bbox": [ + 313, + 543, + 554, + 591 + ], + "type": "text", + "content": " denotes the target model parameters (e.g., exponential moving average (EMA) of " + }, + { + "bbox": [ + 313, + 543, + 554, + 591 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 543, + 554, + 591 + ], + "type": "text", + "content": "). This distillation loss enables the newly trained model, " + }, + { + "bbox": [ + 313, + 543, + 554, + 591 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 313, + 543, + 554, + 591 + ], + "type": "text", + "content": ", to generate high-quality samples in a single step." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 594, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 554, + 689 + ], + "type": "text", + "content": "HR-Regularized Consistency Learning Although the original CD objective theoretically ensures that all trajectory points map to the same origin, it does not explicitly restrict alignment between the generated samples and the corresponding HR targets. This is because velocity field approximation errors from the teacher model " + }, + { + "bbox": [ + 313, + 594, + 554, + 689 + ], + "type": "inline_equation", + "content": "v_{\\phi}" + }, + { + "bbox": [ + 313, + 594, + 554, + 689 + ], + "type": "text", + "content": " and the target model " + }, + { + "bbox": [ + 313, + 594, + 554, + 689 + ], + "type": "inline_equation", + "content": "v_{\\theta^{-}}" + }, + { + "bbox": [ + 313, + 594, + 554, + 689 + ], + "type": "text", + "content": " can propagate into the distillation target " + }, + { + "bbox": [ + 313, + 594, + 554, + 689 + ], + "type": "inline_equation", + "content": "f_{\\theta^{-}}(\\hat{X}_{t}^{\\phi}, t)" + }, + { + "bbox": [ + 313, + 594, + 554, + 689 + ], + "type": "text", + "content": ", as shown in Fig. 3." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "To address this, we introduce an extra HR regularization into Eq. (5), which directly enforces the student's" + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11758" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 297, + 100 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 297, + 100 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 297, + 100 + ], + "type": "text", + "content": "predictions " + }, + { + "bbox": [ + 55, + 71, + 297, + 100 + ], + "type": "inline_equation", + "content": "\\hat{X}_{t\\rightarrow 0}^{\\theta} = f_{\\theta}(\\hat{X}_t^\\phi ,t)" + }, + { + "bbox": [ + 55, + 71, + 297, + 100 + ], + "type": "text", + "content": " (thus distillation target " + }, + { + "bbox": [ + 55, + 71, + 297, + 100 + ], + "type": "inline_equation", + "content": "f_{\\theta^{-}}(\\hat{X}_t^\\phi ,t))" + }, + { + "bbox": [ + 55, + 71, + 297, + 100 + ], + "type": "text", + "content": " to match the ground-truth HR images " + }, + { + "bbox": [ + 55, + 71, + 297, + 100 + ], + "type": "inline_equation", + "content": "X_0" + }, + { + "bbox": [ + 55, + 71, + 297, + 100 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 296, + 133 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 296, + 133 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {h r c d} = \\mathbb {E} \\left[ d \\left(f _ {\\theta} \\left(X _ {t ^ {\\prime}}, t ^ {\\prime}\\right), f _ {\\theta^ {-}} \\left(\\hat {X} _ {t} ^ {\\phi}, t\\right)\\right) + d \\left(f _ {\\theta} \\left(\\hat {X} _ {t} ^ {\\phi}, t\\right), X _ {0}\\right) \\right]. \\tag {6}", + "image_path": "bddc419e82bac8f117f4e655d9af988f6982aaed01c835c372a2c88a244fed5d.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 133, + 296, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 296, + 240 + ], + "type": "text", + "content": "By training with real HR targets, consistency learning reduces reliance on the imperfect teacher and target model. This dual consistency objective ensures that " + }, + { + "bbox": [ + 55, + 133, + 296, + 240 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 55, + 133, + 296, + 240 + ], + "type": "text", + "content": " maintains both trajectory consistency and fidelity to HR data. Note that the input " + }, + { + "bbox": [ + 55, + 133, + 296, + 240 + ], + "type": "inline_equation", + "content": "(\\hat{X}_t^\\phi ,t)" + }, + { + "bbox": [ + 55, + 133, + 296, + 240 + ], + "type": "text", + "content": " can be viewed as a perturbed version of sampled " + }, + { + "bbox": [ + 55, + 133, + 296, + 240 + ], + "type": "inline_equation", + "content": "(X_{t},t)" + }, + { + "bbox": [ + 55, + 133, + 296, + 240 + ], + "type": "text", + "content": " from the forward process in Eq. (1), as visualized in Fig. 3. This perturbation resembles signal noise and distribution shifts, thereby improving the robustness of the trained model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 245, + 296, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 245, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 245, + 296, + 376 + ], + "type": "text", + "content": "Fast-Slow Time Scheduling To further improve the efficiency and robustness of consistency learning, we propose sampling pairs of timesteps, " + }, + { + "bbox": [ + 55, + 245, + 296, + 376 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 245, + 296, + 376 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 245, + 296, + 376 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 55, + 245, + 296, + 376 + ], + "type": "text", + "content": ", from two distinct time schedulers: a \"fast\" scheduler with relatively few timesteps (e.g., 4) and a \"slow\" scheduler with more granular timesteps (e.g., 1000). The slow scheduler, a standard practice for diffusion or flow models, provides fine-grained learning signals to keep intermediate velocity field predictions well-aligned with the SR flow objectives (Eq. (2)). In contrast, the fast scheduler emphasizes larger jumps along the trajectory, reflecting more desired one-step inference." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 377, + 296, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 377, + 296, + 520 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 296, + 520 + ], + "type": "text", + "content": "During training, we randomly select one scheduler (fast or slow) to first sample " + }, + { + "bbox": [ + 55, + 377, + 296, + 520 + ], + "type": "inline_equation", + "content": "t + \\Delta t" + }, + { + "bbox": [ + 55, + 377, + 296, + 520 + ], + "type": "text", + "content": ", then sample " + }, + { + "bbox": [ + 55, + 377, + 296, + 520 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 377, + 296, + 520 + ], + "type": "text", + "content": " in its adjacent region from the other scheduler, and compute " + }, + { + "bbox": [ + 55, + 377, + 296, + 520 + ], + "type": "inline_equation", + "content": "(\\hat{X}_t^\\phi ,t)" + }, + { + "bbox": [ + 55, + 377, + 296, + 520 + ], + "type": "text", + "content": " accordingly using Eq. (3), as illustrated in Fig. 3. This procedure allows flexible and larger jumps " + }, + { + "bbox": [ + 55, + 377, + 296, + 520 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 55, + 377, + 296, + 520 + ], + "type": "text", + "content": " through the fast sampler and relaxes the requirement for accurately estimating of " + }, + { + "bbox": [ + 55, + 377, + 296, + 520 + ], + "type": "inline_equation", + "content": "X_{t}" + }, + { + "bbox": [ + 55, + 377, + 296, + 520 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 377, + 296, + 520 + ], + "type": "inline_equation", + "content": "X_{t + \\Delta t}" + }, + { + "bbox": [ + 55, + 377, + 296, + 520 + ], + "type": "text", + "content": " by running one discretization step of a numerical ODE solver [38]. We hypothesize that introducing mild perturbations and diversity to " + }, + { + "bbox": [ + 55, + 377, + 296, + 520 + ], + "type": "inline_equation", + "content": "\\hat{X}_t^\\phi" + }, + { + "bbox": [ + 55, + 377, + 296, + 520 + ], + "type": "text", + "content": ", deviating it from the sampled point " + }, + { + "bbox": [ + 55, + 377, + 296, + 520 + ], + "type": "inline_equation", + "content": "X_{t}" + }, + { + "bbox": [ + 55, + 377, + 296, + 520 + ], + "type": "text", + "content": ", enhances the SR model's robustness and mitigates distribution shifts, particularly when combined with HR regularization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 528, + 221, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 221, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 221, + 541 + ], + "type": "text", + "content": "4.3. Training Consistency SR Flow" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 545, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 653 + ], + "type": "text", + "content": "SR Flow in Image Space We use Stable Diffusion [31] as the backbone for our SR model, comprising a VAE to map images " + }, + { + "bbox": [ + 55, + 545, + 296, + 653 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 545, + 296, + 653 + ], + "type": "text", + "content": " into latent representations " + }, + { + "bbox": [ + 55, + 545, + 296, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 545, + 296, + 653 + ], + "type": "text", + "content": ", along with a UNet " + }, + { + "bbox": [ + 55, + 545, + 296, + 653 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 545, + 296, + 653 + ], + "type": "text", + "content": " adapted for velocity field learning. While the conditional flow matching loss in Eq. (2) can be formulated in latent space, our early experiments observed that such a latent-space objective does not consistently maintain a favorable balance between fidelity and visual quality over varying numbers of inference steps (e.g., from one to four)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "In practice, we find that loss in the image space is more effective. Specifically, during SR flow training, we first predict the velocity prediction " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "v_{\\theta}(\\mathrm{x}_t,t)" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " in the latent space and compute the HR latent representation " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathrm{x}}_{\\mathrm{HR}} = f_{\\theta}(\\mathrm{x}_t,t)" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " (see Eq. (4)). Next, we decode " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathrm{x}}_{\\mathrm{HR}}" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " into the SR image " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{X}_{\\mathrm{HR}}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": "using the VAE's decoder. Last, we compare " + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "inline_equation", + "content": "\\hat{X}_{\\mathrm{HR}}" + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": " to the ground-truth HR image " + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{HR}}" + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": " using an " + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": " loss and a perceptual (LPIPS [64]) loss:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 321, + 120, + 555, + 135 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 120, + 555, + 135 + ], + "spans": [ + { + "bbox": [ + 321, + 120, + 555, + 135 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {f l o w} = \\left\\| \\hat {X} _ {\\mathrm {H R}} - X _ {\\mathrm {H R}} \\right\\| _ {2} ^ {2} + \\lambda_ {p} \\operatorname {L P I P S} \\left(\\hat {X} _ {\\mathrm {H R}}, X _ {\\mathrm {H R}}\\right), \\tag {7}", + "image_path": "bfe4d96080fa83521269998673f211f46b1be77d87c576b6f4caf36b0cdab10c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 146, + 503, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 146, + 503, + 159 + ], + "spans": [ + { + "bbox": [ + 313, + 146, + 503, + 159 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 146, + 503, + 159 + ], + "type": "inline_equation", + "content": "\\lambda_{p} = 2" + }, + { + "bbox": [ + 313, + 146, + 503, + 159 + ], + "type": "text", + "content": " is set to weight the LPIPS term." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 169, + 555, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 169, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 169, + 555, + 266 + ], + "type": "text", + "content": "Adversarial Loss We incorporate an adversarial GAN objective to further enhance SR quality, particularly for texture synthesis. Specifically, a discriminator " + }, + { + "bbox": [ + 313, + 169, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 169, + 555, + 266 + ], + "type": "text", + "content": " is trained to distinguish between real HR images and those restored by our model. We use the pre-trained diffusion model as a feature extractor in the latent space and attach several additional discriminator heads, following [33, 40, 56]. The hinge loss is adopted for training:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 276, + 554, + 301 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 276, + 554, + 301 + ], + "spans": [ + { + "bbox": [ + 315, + 276, + 554, + 301 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {a d v} = \\mathbb {E} \\left[ \\max (0, 1 - \\mathcal {D} \\left(\\mathrm {x} _ {\\mathrm {H R}}\\right)) + \\max (0, 1 + \\mathcal {D} \\left(\\hat {\\mathrm {x}} _ {\\mathrm {H R}}\\right)) \\right]. \\tag {8}", + "image_path": "0b21faf3f9f4290234f9b2c12864dd3f2f13f74a1b714e7ae302c0d184fcebe8.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 302, + 554, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 302, + 554, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 302, + 554, + 338 + ], + "type": "text", + "content": "Note that the input to " + }, + { + "bbox": [ + 313, + 302, + 554, + 338 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 302, + 554, + 338 + ], + "type": "text", + "content": " is the noised latent with diffusion time " + }, + { + "bbox": [ + 313, + 302, + 554, + 338 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 302, + 554, + 338 + ], + "type": "text", + "content": " (omitted here for simplicity). The SR model is then trained to minimize this adversarial objective." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 349, + 555, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 555, + 506 + ], + "type": "text", + "content": "Image Quality Alignment Loss To further align the SR outputs with desirable image quality attributes (e.g., high-resolution, sharp, detailed), we propose an image quality alignment loss based on a text-image contrastive loss. Specifically, we use large vision-language models to generate a positive quality caption " + }, + { + "bbox": [ + 313, + 349, + 555, + 506 + ], + "type": "inline_equation", + "content": "c_{pos}" + }, + { + "bbox": [ + 313, + 349, + 555, + 506 + ], + "type": "text", + "content": " (e.g., \"image quality is good, with clear details and vibrant colors\") from the HR image and a negative quality caption " + }, + { + "bbox": [ + 313, + 349, + 555, + 506 + ], + "type": "inline_equation", + "content": "c_{neg}" + }, + { + "bbox": [ + 313, + 349, + 555, + 506 + ], + "type": "text", + "content": " (e.g., \"image quality is poor, blurry, and noisy\") from its LR counterpart. We then encode these captions using the text encoder " + }, + { + "bbox": [ + 313, + 349, + 555, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_T" + }, + { + "bbox": [ + 313, + 349, + 555, + 506 + ], + "type": "text", + "content": " of CLIP [29], while the SR result " + }, + { + "bbox": [ + 313, + 349, + 555, + 506 + ], + "type": "inline_equation", + "content": "\\hat{X}" + }, + { + "bbox": [ + 313, + 349, + 555, + 506 + ], + "type": "text", + "content": " is encoded via CLIP's image encoder " + }, + { + "bbox": [ + 313, + 349, + 555, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_I" + }, + { + "bbox": [ + 313, + 349, + 555, + 506 + ], + "type": "text", + "content": " (ViT-L/14 is used). Image quality alignment loss encourages " + }, + { + "bbox": [ + 313, + 349, + 555, + 506 + ], + "type": "inline_equation", + "content": "\\hat{X}" + }, + { + "bbox": [ + 313, + 349, + 555, + 506 + ], + "type": "text", + "content": " to be close to " + }, + { + "bbox": [ + 313, + 349, + 555, + 506 + ], + "type": "inline_equation", + "content": "c_{pos}" + }, + { + "bbox": [ + 313, + 349, + 555, + 506 + ], + "type": "text", + "content": " and far from " + }, + { + "bbox": [ + 313, + 349, + 555, + 506 + ], + "type": "inline_equation", + "content": "c_{neg}" + }, + { + "bbox": [ + 313, + 349, + 555, + 506 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 517, + 553, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 517, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 321, + 517, + 553, + 555 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i q a} = - \\log \\frac {\\exp \\left(\\sin \\left(\\hat {X} , c _ {p o s}\\right)\\right)}{\\exp \\left(\\sin \\left(\\hat {X} , c _ {p o s}\\right)\\right) + \\exp \\left(\\sin \\left(\\hat {X} , c _ {n e g}\\right)\\right)}, \\tag {9}", + "image_path": "9e78bb281e206f953a0939fc80ec71f7744b9a9a9768b3809ada8c78f17c7011.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 556, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 554, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 556, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\mathrm{sim}(X,c) = \\cos (\\mathcal{E}_I(X),\\mathcal{E}_T(c))" + }, + { + "bbox": [ + 313, + 556, + 554, + 605 + ], + "type": "text", + "content": ". Intuitively, this loss implicitly guides the network to improve perceptual fidelity and clarity by bridging the gap between human-perceived \"good\" and \"bad\" image attributes." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 615, + 554, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 615, + 554, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 615, + 554, + 664 + ], + "type": "text", + "content": "Putting Things Together We first pre-train the SR flow model using the basic conditional flow loss in Eq. (7). Then we continue with the consistency SR flow training by finetuning it. The overall loss function is defined as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 325, + 676, + 555, + 690 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 676, + 555, + 690 + ], + "spans": [ + { + "bbox": [ + 325, + 676, + 555, + 690 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {f l o w}} + \\lambda_ {c d} \\mathcal {L} _ {h r c d} + \\lambda_ {a d v} \\mathcal {L} _ {a d v} + \\lambda_ {i q a} \\mathcal {L} _ {i q a}, \\tag {10}", + "image_path": "a895f82530f2c69c9ebeb7d3084b30d4ba00edaf0f335e3e843a2302e5136e2e.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 701, + 508, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 508, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 508, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 701, + 508, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{cd}" + }, + { + "bbox": [ + 313, + 701, + 508, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 701, + 508, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{adv}" + }, + { + "bbox": [ + 313, + 701, + 508, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 701, + 508, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{iqa}" + }, + { + "bbox": [ + 313, + 701, + 508, + 714 + ], + "type": "text", + "content": " are hyperparameters." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11759" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 96, + 555, + 303 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Quantitative comparison with state-of-the-art DM-based SR methods on RealSR [1] and DRealSR [47]. The best and second-best results are highlighted in bold and underlined." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 96, + 555, + 303 + ], + "lines": [ + { + "bbox": [ + 56, + 96, + 555, + 303 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 555, + 303 + ], + "type": "table", + "html": "
DatasetsMethods#StepsPSNR ↑SSIM ↑LPIPS ↓DISTS ↓FID ↓NIQE ↓MUSIQ ↑MANIQA ↑CLIPIQA ↑
RealSRStableSR [42]20024.700.70850.30180.2288128.515.9165.780.62210.6178
DiffBIR [19]5024.750.65670.36360.2312128.995.5364.980.62460.6463
SeeSR [49]5025.180.72160.30090.2223125.555.4169.770.64420.6612
PASD [55]2025.210.67980.33800.2260124.295.4168.750.64870.6620
ResShift [59]1526.310.74210.34600.2498141.717.2658.430.52850.5444
SinSR [45]126.280.73470.31880.2353135.936.2960.800.53850.6122
OSEDiff [48]125.150.73410.29210.2128123.495.6569.090.63260.6693
DoSSR [3]125.670.73870.33560.2741153.2511.1962.690.52430.6259
FlowSR125.540.74340.27280.2013112.605.2869.220.64860.6701
DRealSRStableSR [42]20028.030.75360.32840.2269148.986.5258.510.56010.6356
DiffBIR [19]5026.710.65710.45570.2748166.796.3161.070.59300.6395
SeeSR [49]5028.170.76910.31890.2315147.396.4064.930.60420.6804
PASD [55]2027.360.70730.37600.2531156.135.5564.870.61690.6808
ResShift [59]1528.460.76730.40060.2656172.268.1250.600.45860.5342
SinSR [45]128.360.75150.36650.2485170.576.9955.330.48840.6383
OSEDiff [48]127.920.78350.29680.2165135.306.4964.650.58990.6963
DoSSR [3]128.550.79910.33530.2801166.1912.2556.720.46230.5739
FlowSR128.500.78590.29750.2115130.306.1365.460.61720.7074
", + "image_path": "90ac527278377cc6439f4cb64cc6ffdb820cff736236d3f6ce33ad7db990f84d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 311, + 137, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 311, + 137, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 311, + 137, + 324 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 331, + 182, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 182, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 182, + 344 + ], + "type": "text", + "content": "5.1. Experimental Settings" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 349, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 349, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 296, + 456 + ], + "type": "text", + "content": "Implementation Details Our model is based on the pretrained SD 2.1-base [31]. During training, we fine-tune only the U-Net using a LoRA [11] rank of 32. The learning rate is set to 2e-5, with a training patch size of " + }, + { + "bbox": [ + 55, + 349, + 296, + 456 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 349, + 296, + 456 + ], + "type": "text", + "content": " and a batch size of 16. The training process consists of two stages: we firstly train the SR flow model for 10k iterations, followed by consistency learning for an additional 20k iterations. The loss weights " + }, + { + "bbox": [ + 55, + 349, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\lambda_{cd}, \\lambda_{adv}, \\lambda_{iqa}" + }, + { + "bbox": [ + 55, + 349, + 296, + 456 + ], + "type": "text", + "content": " are set to 0.1, 0.05, and 0.1, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 463, + 296, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 296, + 547 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 296, + 547 + ], + "type": "text", + "content": "Data We train our model using LSDIR [17] and the first 10K face images from FFHQ [12]. The LR-HR training pairs are synthesized using the degradation pipeline of RealESRGAN [44]. To generate image quality captions for HR and LR images, we use Qwen2-VL [43]. We adopt the real-world test set of StableSR [42] for evaluation and comparison. The test sets include RealSR [1] and DRealSR [47]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 552, + 296, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 296, + 612 + ], + "type": "text", + "content": "Compared Methods We compare our method against several state-of-the-art diffusion-based SR approaches, including multi-step StableSR [42], DiffBIR [19], SeeSR [49], PASD [55], ResShift [59], and one-step SinSR [45], OSEDiff [48], and DoSSR [3]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "text", + "content": "Evaluation Metrics We evaluate our model using both reference-based and no-reference metrics. For fidelity, we report PSNR and SSIM (on the Y channel in YCbCr space), and for perceptual quality, we use LPIPS [64] and DISTS [4]. FID [9] is reported to compare the distribution of restored images with the ground truth. For no-reference evaluation, we use NIQE [25], MUSIQ [13], MANIQA [54], and CLIPIQA [41]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 311, + 539, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 539, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 539, + 324 + ], + "type": "text", + "content": "5.2. Comparison with State-of-the-Art Methods" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 330, + 555, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 330, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 555, + 521 + ], + "type": "text", + "content": "Quantitative Comparisons Table 1 shows that our approach achieves superior or at least competitive performance compared to recent state-of-the-art methods. Compared with multi-step diffusion-based methods such as SeeSR [49] and PASD [55], our FlowSR obtains equivalent or better no-reference metrics, such as MUSIQ, MANIQA, and CLIPIQA, and surpasses these methods in fidelity-based scores, e.g. PSNR and LPIPS, while reducing inference to a single step. Compared to efficient diffusion-based SR methods such as ResShift [59], our method demonstrates a clear advantage. Against the strong DDPM-based one-step competitor OSEDiff [48], FlowSR achieves superior performance across nearly all metrics. Overall, these results demonstrate that our flow-based SR method effectively balances fidelity and perceived quality while showing superiority over compared methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 529, + 556, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 556, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 556, + 625 + ], + "type": "text", + "content": "Qualitative Comparisons Fig. 4 shows visual comparisons of two real-world examples. Multi-step DM-based methods, such as DiffBIR and PASD, often generate rich but inaccurate textures, while one-step methods like SinSR and OSEDiff tend to produce blurry or less detailed results. In contrast, our method generates faithful SR results, effectively recovering structures such as the stone breakwater and the textures of the cloth." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 635, + 416, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 416, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 416, + 647 + ], + "type": "text", + "content": "5.3. Ablation Studies" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": "Effects of SR Flow We first study the flow formulation for image super-resolution. Our SR flow is a straight path between " + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\pi_0 = X_{\\mathrm{HR}}" + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\pi_1 = X_{\\mathrm{LR}}" + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": ". We consider alternative formulations starting from noise, where " + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\pi_1 = \\mathcal{N}(0,1)" + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": ", as used in pre-trained T2I models, or starting from noised" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11760" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 202, + 217 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 202, + 217 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 202, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 202, + 217 + ], + "type": "image", + "image_path": "54ca3c2cdf6d193f30d80467308a569a4733ccf75ab8c171d543842c8cb8d9c3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 113, + 217, + 147, + 225 + ], + "lines": [ + { + "bbox": [ + 113, + 217, + 147, + 225 + ], + "spans": [ + { + "bbox": [ + 113, + 217, + 147, + 225 + ], + "type": "text", + "content": "LR Image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 203, + 70, + 273, + 140 + ], + "blocks": [ + { + "bbox": [ + 203, + 70, + 273, + 140 + ], + "lines": [ + { + "bbox": [ + 203, + 70, + 273, + 140 + ], + "spans": [ + { + "bbox": [ + 203, + 70, + 273, + 140 + ], + "type": "image", + "image_path": "042271f964a8bd2b41f4561931764176e55ae68a0891e43be10a2913e7c6f0c2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 232, + 140, + 244, + 147 + ], + "lines": [ + { + "bbox": [ + 232, + 140, + 244, + 147 + ], + "spans": [ + { + "bbox": [ + 232, + 140, + 244, + 147 + ], + "type": "text", + "content": "HR" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 274, + 71, + 343, + 139 + ], + "blocks": [ + { + "bbox": [ + 274, + 71, + 343, + 139 + ], + "lines": [ + { + "bbox": [ + 274, + 71, + 343, + 139 + ], + "spans": [ + { + "bbox": [ + 274, + 71, + 343, + 139 + ], + "type": "image", + "image_path": "2335d0a9e3c20b22e292a413ea24cb7147093ad51b816d13569cf946a1786b7e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 343, + 72, + 413, + 139 + ], + "blocks": [ + { + "bbox": [ + 343, + 72, + 413, + 139 + ], + "lines": [ + { + "bbox": [ + 343, + 72, + 413, + 139 + ], + "spans": [ + { + "bbox": [ + 343, + 72, + 413, + 139 + ], + "type": "image", + "image_path": "7ae6ec87e9e13671c5fbce08ae17b8a88ef065515fb7dece036c9e3937d4c532.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 414, + 72, + 483, + 139 + ], + "blocks": [ + { + "bbox": [ + 414, + 72, + 483, + 139 + ], + "lines": [ + { + "bbox": [ + 414, + 72, + 483, + 139 + ], + "spans": [ + { + "bbox": [ + 414, + 72, + 483, + 139 + ], + "type": "image", + "image_path": "d3e7a7995566469f3751fd5e13b66d6fa97308963b5fa3a548ea274d28db309e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 484, + 72, + 553, + 139 + ], + "blocks": [ + { + "bbox": [ + 484, + 72, + 553, + 139 + ], + "lines": [ + { + "bbox": [ + 484, + 72, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 484, + 72, + 553, + 139 + ], + "type": "image", + "image_path": "ffdcf532b63640a46e3b1e4fa965394367c0771e270acfdcf069a07cd58587b6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 56, + 227, + 202, + 373 + ], + "blocks": [ + { + "bbox": [ + 56, + 227, + 202, + 373 + ], + "lines": [ + { + "bbox": [ + 56, + 227, + 202, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 227, + 202, + 373 + ], + "type": "image", + "image_path": "7a7f5580fd114f25da788d007a33db727b9e5492426ffd1eb15b384d4758a628.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 113, + 373, + 147, + 382 + ], + "lines": [ + { + "bbox": [ + 113, + 373, + 147, + 382 + ], + "spans": [ + { + "bbox": [ + 113, + 373, + 147, + 382 + ], + "type": "text", + "content": "LR Image" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 203, + 148, + 273, + 216 + ], + "blocks": [ + { + "bbox": [ + 203, + 148, + 273, + 216 + ], + "lines": [ + { + "bbox": [ + 203, + 148, + 273, + 216 + ], + "spans": [ + { + "bbox": [ + 203, + 148, + 273, + 216 + ], + "type": "image", + "image_path": "2556f83fe362a993be4e8e05fc131e3e97c2963e723ff608dd33c231645d68ab.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 216, + 217, + 260, + 225 + ], + "lines": [ + { + "bbox": [ + 216, + 217, + 260, + 225 + ], + "spans": [ + { + "bbox": [ + 216, + 217, + 260, + 225 + ], + "type": "text", + "content": "ResShift (15)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 274, + 148, + 343, + 217 + ], + "blocks": [ + { + "bbox": [ + 283, + 140, + 334, + 148 + ], + "lines": [ + { + "bbox": [ + 283, + 140, + 334, + 148 + ], + "spans": [ + { + "bbox": [ + 283, + 140, + 334, + 148 + ], + "type": "text", + "content": "StableSR (200)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 274, + 148, + 343, + 217 + ], + "lines": [ + { + "bbox": [ + 274, + 148, + 343, + 217 + ], + "spans": [ + { + "bbox": [ + 274, + 148, + 343, + 217 + ], + "type": "image", + "image_path": "d8b83d480f026a3c3978bf04f45a8a228ff1814ed5b2690573e027986acc6dbd.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 291, + 217, + 325, + 225 + ], + "lines": [ + { + "bbox": [ + 291, + 217, + 325, + 225 + ], + "spans": [ + { + "bbox": [ + 291, + 217, + 325, + 225 + ], + "type": "text", + "content": "SinSR (1)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 343, + 148, + 413, + 217 + ], + "blocks": [ + { + "bbox": [ + 357, + 140, + 400, + 148 + ], + "lines": [ + { + "bbox": [ + 357, + 140, + 400, + 148 + ], + "spans": [ + { + "bbox": [ + 357, + 140, + 400, + 148 + ], + "type": "text", + "content": "DiffBIR (50)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 343, + 148, + 413, + 217 + ], + "lines": [ + { + "bbox": [ + 343, + 148, + 413, + 217 + ], + "spans": [ + { + "bbox": [ + 343, + 148, + 413, + 217 + ], + "type": "image", + "image_path": "8d4acf7a89ffd5d63f9b6b453a71c2cb7bd162e6a781f1951492ef65b6d1701d.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 358, + 217, + 399, + 225 + ], + "lines": [ + { + "bbox": [ + 358, + 217, + 399, + 225 + ], + "spans": [ + { + "bbox": [ + 358, + 217, + 399, + 225 + ], + "type": "text", + "content": "OSEDiff (1)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 414, + 148, + 483, + 217 + ], + "blocks": [ + { + "bbox": [ + 429, + 140, + 468, + 148 + ], + "lines": [ + { + "bbox": [ + 429, + 140, + 468, + 148 + ], + "spans": [ + { + "bbox": [ + 429, + 140, + 468, + 148 + ], + "type": "text", + "content": "SeeSR (50)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 414, + 148, + 483, + 217 + ], + "lines": [ + { + "bbox": [ + 414, + 148, + 483, + 217 + ], + "spans": [ + { + "bbox": [ + 414, + 148, + 483, + 217 + ], + "type": "image", + "image_path": "691fad6514142eff3400fd73ed44afd439a33039b84f3ad3f50ab54fc88ab68d.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 217, + 466, + 225 + ], + "lines": [ + { + "bbox": [ + 429, + 217, + 466, + 225 + ], + "spans": [ + { + "bbox": [ + 429, + 217, + 466, + 225 + ], + "type": "text", + "content": "DoSSR (1)" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 484, + 148, + 553, + 217 + ], + "blocks": [ + { + "bbox": [ + 500, + 140, + 536, + 148 + ], + "lines": [ + { + "bbox": [ + 500, + 140, + 536, + 148 + ], + "spans": [ + { + "bbox": [ + 500, + 140, + 536, + 148 + ], + "type": "text", + "content": "PASD (20)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 484, + 148, + 553, + 217 + ], + "lines": [ + { + "bbox": [ + 484, + 148, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 484, + 148, + 553, + 217 + ], + "type": "image", + "image_path": "16e31545a1e4d897f2c1fac6ccebd5e1c196f90722bc9327e26486bdc8eb3ad9.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 498, + 217, + 537, + 225 + ], + "lines": [ + { + "bbox": [ + 498, + 217, + 537, + 225 + ], + "spans": [ + { + "bbox": [ + 498, + 217, + 537, + 225 + ], + "type": "text", + "content": "FlowSR (1)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 203, + 227, + 273, + 297 + ], + "blocks": [ + { + "bbox": [ + 203, + 227, + 273, + 297 + ], + "lines": [ + { + "bbox": [ + 203, + 227, + 273, + 297 + ], + "spans": [ + { + "bbox": [ + 203, + 227, + 273, + 297 + ], + "type": "image", + "image_path": "c2d787425c24e1cba5784055449b89c0f1c381a2372324f824b1548e6f8fc6cb.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 233, + 297, + 244, + 304 + ], + "lines": [ + { + "bbox": [ + 233, + 297, + 244, + 304 + ], + "spans": [ + { + "bbox": [ + 233, + 297, + 244, + 304 + ], + "type": "text", + "content": "HR" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 274, + 227, + 343, + 296 + ], + "blocks": [ + { + "bbox": [ + 274, + 227, + 343, + 296 + ], + "lines": [ + { + "bbox": [ + 274, + 227, + 343, + 296 + ], + "spans": [ + { + "bbox": [ + 274, + 227, + 343, + 296 + ], + "type": "image", + "image_path": "dc15a0f907d55c49679a14f74c7379fa049a309bb89fec07fa1de9dff1997bb9.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 283, + 297, + 333, + 304 + ], + "lines": [ + { + "bbox": [ + 283, + 297, + 333, + 304 + ], + "spans": [ + { + "bbox": [ + 283, + 297, + 333, + 304 + ], + "type": "text", + "content": "StableSR (200)" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 343, + 227, + 413, + 296 + ], + "blocks": [ + { + "bbox": [ + 343, + 227, + 413, + 296 + ], + "lines": [ + { + "bbox": [ + 343, + 227, + 413, + 296 + ], + "spans": [ + { + "bbox": [ + 343, + 227, + 413, + 296 + ], + "type": "image", + "image_path": "a28445e04f5a94e8668ca63da78effc189ef7e9c61b94d022716ce79f67954e3.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 357, + 297, + 400, + 304 + ], + "lines": [ + { + "bbox": [ + 357, + 297, + 400, + 304 + ], + "spans": [ + { + "bbox": [ + 357, + 297, + 400, + 304 + ], + "type": "text", + "content": "DiffBIR (50)" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 414, + 227, + 483, + 296 + ], + "blocks": [ + { + "bbox": [ + 414, + 227, + 483, + 296 + ], + "lines": [ + { + "bbox": [ + 414, + 227, + 483, + 296 + ], + "spans": [ + { + "bbox": [ + 414, + 227, + 483, + 296 + ], + "type": "image", + "image_path": "e295b648acc4829cc109a3ff5a6370d5ca565261faf9a2ef4588a3cef1cc92d0.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 297, + 467, + 304 + ], + "lines": [ + { + "bbox": [ + 429, + 297, + 467, + 304 + ], + "spans": [ + { + "bbox": [ + 429, + 297, + 467, + 304 + ], + "type": "text", + "content": "SeeSR (50)" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 484, + 227, + 553, + 296 + ], + "blocks": [ + { + "bbox": [ + 484, + 227, + 553, + 296 + ], + "lines": [ + { + "bbox": [ + 484, + 227, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 484, + 227, + 553, + 296 + ], + "type": "image", + "image_path": "d0d7bb8ee3403385ef0c9915a543dbb8f5ffc51d52ecae11c8059975e27c9ff6.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 500, + 297, + 536, + 304 + ], + "lines": [ + { + "bbox": [ + 500, + 297, + 536, + 304 + ], + "spans": [ + { + "bbox": [ + 500, + 297, + 536, + 304 + ], + "type": "text", + "content": "PASD (20)" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 57, + 415, + 116, + 474 + ], + "blocks": [ + { + "bbox": [ + 55, + 386, + 553, + 407 + ], + "lines": [ + { + "bbox": [ + 55, + 386, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 386, + 553, + 407 + ], + "type": "text", + "content": "Figure 4. Visual comparisons of different SR methods on real-world examples. The number of sampling steps are indicated in parentheses. Please zoom in for a better view." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 415, + 116, + 474 + ], + "lines": [ + { + "bbox": [ + 57, + 415, + 116, + 474 + ], + "spans": [ + { + "bbox": [ + 57, + 415, + 116, + 474 + ], + "type": "image", + "image_path": "090d29280d1f2335183e731a9ddc23d15db14a0cae759e3b8a535e58f6cea11e.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 117, + 415, + 175, + 474 + ], + "blocks": [ + { + "bbox": [ + 117, + 415, + 175, + 474 + ], + "lines": [ + { + "bbox": [ + 117, + 415, + 175, + 474 + ], + "spans": [ + { + "bbox": [ + 117, + 415, + 175, + 474 + ], + "type": "image", + "image_path": "ccc115ca9f373cd25e40fd871f8ad73df503f27d3221ce3d8bed5f35a00f147a.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 176, + 415, + 235, + 474 + ], + "blocks": [ + { + "bbox": [ + 176, + 415, + 235, + 474 + ], + "lines": [ + { + "bbox": [ + 176, + 415, + 235, + 474 + ], + "spans": [ + { + "bbox": [ + 176, + 415, + 235, + 474 + ], + "type": "image", + "image_path": "2ca2b5d3d20e3cc93e4197ce3c50626f7de236eee48d730b394483801f1fc3ac.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 235, + 415, + 294, + 474 + ], + "blocks": [ + { + "bbox": [ + 235, + 415, + 294, + 474 + ], + "lines": [ + { + "bbox": [ + 235, + 415, + 294, + 474 + ], + "spans": [ + { + "bbox": [ + 235, + 415, + 294, + 474 + ], + "type": "image", + "image_path": "bb1236b864ecc899729334b5e6d586f2cc471b7adfafe3a9a44e358954face1a.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 58, + 484, + 116, + 543 + ], + "blocks": [ + { + "bbox": [ + 72, + 476, + 101, + 483 + ], + "lines": [ + { + "bbox": [ + 72, + 476, + 101, + 483 + ], + "spans": [ + { + "bbox": [ + 72, + 476, + 101, + 483 + ], + "type": "text", + "content": "LR Image" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 484, + 116, + 543 + ], + "lines": [ + { + "bbox": [ + 58, + 484, + 116, + 543 + ], + "spans": [ + { + "bbox": [ + 58, + 484, + 116, + 543 + ], + "type": "image", + "image_path": "f3ec8a1aa5d57dac5a0b60ac2f4842d085d8e77e096aac29dc8c8545b095e662.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 544, + 105, + 552 + ], + "lines": [ + { + "bbox": [ + 69, + 544, + 105, + 552 + ], + "spans": [ + { + "bbox": [ + 69, + 544, + 105, + 552 + ], + "type": "text", + "content": "Zoomed LR" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 557, + 295, + 579 + ], + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 579 + ], + "type": "text", + "content": "Figure 5. Ablation of SR flow. Our SR flow model produces more accurate SR results with improved structure with more steps." + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 117, + 484, + 175, + 543 + ], + "blocks": [ + { + "bbox": [ + 123, + 476, + 168, + 483 + ], + "lines": [ + { + "bbox": [ + 123, + 476, + 168, + 483 + ], + "spans": [ + { + "bbox": [ + 123, + 476, + 168, + 483 + ], + "type": "inline_equation", + "content": "\\mathrm{Noise}\\rightarrow \\mathrm{HR}" + }, + { + "bbox": [ + 123, + 476, + 168, + 483 + ], + "type": "text", + "content": " (4)" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 117, + 484, + 175, + 543 + ], + "lines": [ + { + "bbox": [ + 117, + 484, + 175, + 543 + ], + "spans": [ + { + "bbox": [ + 117, + 484, + 175, + 543 + ], + "type": "image", + "image_path": "a0bc5dea370fb5836135c149f48b0e2030a0629a96662408f1b49d4a3840fc29.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 124, + 544, + 168, + 552 + ], + "lines": [ + { + "bbox": [ + 124, + 544, + 168, + 552 + ], + "spans": [ + { + "bbox": [ + 124, + 544, + 168, + 552 + ], + "type": "inline_equation", + "content": "\\mathrm{Noise}\\rightarrow \\mathrm{HR}" + }, + { + "bbox": [ + 124, + 544, + 168, + 552 + ], + "type": "text", + "content": " (1)" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 176, + 484, + 235, + 543 + ], + "blocks": [ + { + "bbox": [ + 177, + 476, + 234, + 483 + ], + "lines": [ + { + "bbox": [ + 177, + 476, + 234, + 483 + ], + "spans": [ + { + "bbox": [ + 177, + 476, + 234, + 483 + ], + "type": "text", + "content": "noised " + }, + { + "bbox": [ + 177, + 476, + 234, + 483 + ], + "type": "inline_equation", + "content": "\\mathrm{LR}\\rightarrow \\mathrm{HR}" + }, + { + "bbox": [ + 177, + 476, + 234, + 483 + ], + "type": "text", + "content": " (4)" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 176, + 484, + 235, + 543 + ], + "lines": [ + { + "bbox": [ + 176, + 484, + 235, + 543 + ], + "spans": [ + { + "bbox": [ + 176, + 484, + 235, + 543 + ], + "type": "image", + "image_path": "74e8bb9ff6cb2de759e636de029967c0d69683835d28d475909dfa4072001db5.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 177, + 544, + 234, + 552 + ], + "lines": [ + { + "bbox": [ + 177, + 544, + 234, + 552 + ], + "spans": [ + { + "bbox": [ + 177, + 544, + 234, + 552 + ], + "type": "text", + "content": "noised " + }, + { + "bbox": [ + 177, + 544, + 234, + 552 + ], + "type": "inline_equation", + "content": "\\mathrm{LR}\\rightarrow \\mathrm{HR}" + }, + { + "bbox": [ + 177, + 544, + 234, + 552 + ], + "type": "text", + "content": " 1" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 236, + 484, + 294, + 543 + ], + "blocks": [ + { + "bbox": [ + 246, + 476, + 282, + 483 + ], + "lines": [ + { + "bbox": [ + 246, + 476, + 282, + 483 + ], + "spans": [ + { + "bbox": [ + 246, + 476, + 282, + 483 + ], + "type": "text", + "content": "SR Flow (4)" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 236, + 484, + 294, + 543 + ], + "lines": [ + { + "bbox": [ + 236, + 484, + 294, + 543 + ], + "spans": [ + { + "bbox": [ + 236, + 484, + 294, + 543 + ], + "type": "image", + "image_path": "7f0998d4de7ad064534623635986ef9f1dcfa5b2a3cee1255d1361d1e7dc7ca8.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 247, + 544, + 282, + 552 + ], + "lines": [ + { + "bbox": [ + 247, + 544, + 282, + 552 + ], + "spans": [ + { + "bbox": [ + 247, + 544, + 282, + 552 + ], + "type": "text", + "content": "SR Flow (1)" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_caption" + } + ], + "index": 49 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "LR, where " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\pi_1 = X_{\\mathrm{LR}} + \\mu \\epsilon" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,1)" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": ", as proposed in ResShift [59]. For these alternatives, we concatenate " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "X_{t}" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{LR}}" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " as the input to provide the LR condition. The trained models are evaluated on the RealSR [1] and DRealSR [47] datasets, as shown in Table 2. We observe that with more inference steps, the image quality (MUSIQ) generally improves, while the fidelity (PSNR) slightly decreases. Fig. 5 presents the visual comparisons, where our SR flow model produces more accurate predictions, with further structure and overall quality enhancement as addi" + } + ] + } + ], + "index": 52 + }, + { + "type": "table", + "bbox": [ + 315, + 453, + 555, + 541 + ], + "blocks": [ + { + "bbox": [ + 313, + 411, + 555, + 444 + ], + "lines": [ + { + "bbox": [ + 313, + 411, + 555, + 444 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 555, + 444 + ], + "type": "text", + "content": "Table 2. Ablation of SR flow. SR flow mapping LR to HR outperforms alternative flow formulations that start from noise or ResShift [59]-style noised LR on RealSR and DRealSR." + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 453, + 555, + 541 + ], + "lines": [ + { + "bbox": [ + 315, + 453, + 555, + 541 + ], + "spans": [ + { + "bbox": [ + 315, + 453, + 555, + 541 + ], + "type": "table", + "html": "
#StepsMethodRealSRDRealSR
PSNR ↑MUSIQ ↑PSNR ↑MUSIQ ↑
4Noise→HR23.8366.0126.7562.76
noised LR→HR24.4765.1527.7461.76
SR Flow (ours)25.0068.0927.9865.05
1Noise→HR24.8265.1728.0759.34
noised LR→HR24.6765.2528.0360.64
SR Flow (ours)25.5167.4028.6562.53
", + "image_path": "0908201fbf7810f12ffaaee344a2fbe1c83112325476546fde6fda96d9698f14.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "table_body" + } + ], + "index": 54 + }, + { + "bbox": [ + 313, + 552, + 555, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 555, + 600 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 555, + 600 + ], + "type": "text", + "content": "tional sampling steps (i.e., 4 steps) are applied. Notably, these results demonstrate that the model trained using our SR flow clearly outperforms other transition variants in both single-step and multi-step inference settings." + } + ] + } + ], + "index": 55 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "Effects of consistency learning Next, we fine-tune the pre-trained SR flow model using the consistency objective described in Sec. 4.2. As shown in Table 3, enforcing self-consistency " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cd}" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " across neighboring time steps stabilizes intermediate representations and often reduces LPIPS. However, when the teacher's predictions deviate from the true HR manifold, these errors may propagate to the student model, leading to diminished IQA scores. We address this by introducing the HR regularization term in Eq. (6), which" + } + ] + } + ], + "index": 56 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11761" + } + ] + } + ], + "index": 57 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 116, + 186 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 116, + 186 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 116, + 186 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 116, + 186 + ], + "type": "image", + "image_path": "7e6582e6f7b95003933045da3b103124ac87bbaab71469fc16e0ec12e894b6ad.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 71, + 188, + 101, + 196 + ], + "lines": [ + { + "bbox": [ + 71, + 188, + 101, + 196 + ], + "spans": [ + { + "bbox": [ + 71, + 188, + 101, + 196 + ], + "type": "text", + "content": "LR Image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 117, + 70, + 176, + 186 + ], + "blocks": [ + { + "bbox": [ + 117, + 70, + 176, + 186 + ], + "lines": [ + { + "bbox": [ + 117, + 70, + 176, + 186 + ], + "spans": [ + { + "bbox": [ + 117, + 70, + 176, + 186 + ], + "type": "image", + "image_path": "595550517b40f3904fd2e631ea29f4c27835dcfb6e430484035ab7510409ef81.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 188, + 159, + 196 + ], + "lines": [ + { + "bbox": [ + 133, + 188, + 159, + 196 + ], + "spans": [ + { + "bbox": [ + 133, + 188, + 159, + 196 + ], + "type": "text", + "content": "SR Flow" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 176, + 70, + 235, + 186 + ], + "blocks": [ + { + "bbox": [ + 176, + 70, + 235, + 186 + ], + "lines": [ + { + "bbox": [ + 176, + 70, + 235, + 186 + ], + "spans": [ + { + "bbox": [ + 176, + 70, + 235, + 186 + ], + "type": "image", + "image_path": "e60e498a9a0c3ab24db64f162977c46ba25056f10c7c04433d8c8417136ac7fc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 194, + 188, + 216, + 196 + ], + "lines": [ + { + "bbox": [ + 194, + 188, + 216, + 196 + ], + "spans": [ + { + "bbox": [ + 194, + 188, + 216, + 196 + ], + "type": "text", + "content": "w/ " + }, + { + "bbox": [ + 194, + 188, + 216, + 196 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cd}" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 236, + 70, + 294, + 186 + ], + "blocks": [ + { + "bbox": [ + 236, + 70, + 294, + 186 + ], + "lines": [ + { + "bbox": [ + 236, + 70, + 294, + 186 + ], + "spans": [ + { + "bbox": [ + 236, + 70, + 294, + 186 + ], + "type": "image", + "image_path": "67a8ea22fe8df4e90e5ea80f6cd078fbd28bbe51347f61ccc630a60fdc348873.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 188, + 280, + 196 + ], + "lines": [ + { + "bbox": [ + 249, + 188, + 280, + 196 + ], + "spans": [ + { + "bbox": [ + 249, + 188, + 280, + 196 + ], + "type": "text", + "content": "w/ " + }, + { + "bbox": [ + 249, + 188, + 280, + 196 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{hrcd}" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 200, + 295, + 233 + ], + "lines": [ + { + "bbox": [ + 55, + 200, + 295, + 233 + ], + "spans": [ + { + "bbox": [ + 55, + 200, + 295, + 233 + ], + "type": "text", + "content": "Figure 6. Ablation of consistency learning. Our HR-regularized consistency learning effectively reduces distortions in SR outputs while producing high-quality and sharper results (zoom in)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 56, + 267, + 296, + 346 + ], + "blocks": [ + { + "bbox": [ + 55, + 241, + 295, + 264 + ], + "lines": [ + { + "bbox": [ + 55, + 241, + 295, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 241, + 295, + 264 + ], + "type": "text", + "content": "Table 3. Ablation of consistency learning. Fine-tuned SR flow models with consistency objectives are evaluated on DRealSR." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 267, + 296, + 346 + ], + "lines": [ + { + "bbox": [ + 56, + 267, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 56, + 267, + 296, + 346 + ], + "type": "table", + "html": "
#StepsMethodPSNR ↑LPIPS ↓MUSIQ ↑MANIQA ↑CLIPQA ↑
4SR Flow27.980.276365.050.62960.6681
w/ Lcd28.640.265461.820.60420.6180
w/ Lhrcd28.440.279965.710.63740.6709
1SR Flow28.650.282162.530.58950.6545
w/ Lcd28.590.260961.690.59490.6391
w/ Lhr28.850.280562.460.58710.6589
w/ Lhrcd28.620.283064.770.61560.6748
", + "image_path": "8f6410ee2e1d4add35fb48a4fdefe0a35b594908d3ec5cc21625b11accabf735.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 54, + 356, + 295, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 356, + 295, + 548 + ], + "spans": [ + { + "bbox": [ + 54, + 356, + 295, + 548 + ], + "type": "text", + "content": "directly aligns student predictions " + }, + { + "bbox": [ + 54, + 356, + 295, + 548 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\hat{X}_t^\\phi, t)" + }, + { + "bbox": [ + 54, + 356, + 295, + 548 + ], + "type": "text", + "content": " with real HR images, mitigating the teacher's approximation errors. This extra ground-truth constraint helps the model recover fine textural details, which boosts IQA metrics (e.g., MUSIQ and MANIQA) while preserves perceptual fidelity. Consistency learning enhances the robustness of flow-based inference by distilling multi-step capabilities into a single step, allowing for faster sampling without sacrificing quality. Additionally, HR regularization alone (w/ " + }, + { + "bbox": [ + 54, + 356, + 295, + 548 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{hr}" + }, + { + "bbox": [ + 54, + 356, + 295, + 548 + ], + "type": "text", + "content": ") provides minimal IQA gains. This highlights the importance of consistency learning and its synergy with HR regularization for acceleration in SR flow. Fig. 6 presents visual comparisons, showing that consistency learning helps resolve distorted textures in the baseline SR flow model. Notably, HR-regularized consistency learning generates more realistic and sharper results compared to its baseline." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": "Analysis of Fast-Slow Time Sampling We evaluate the effectiveness of fast-slow time sampling by comparing it to the " + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": "-interval scheduler [38] and by varying the number of fast-scheduler timesteps, while fixing the slow scheduler fixed at 1000 timesteps. When the fast scheduler uses only one timestep, we set " + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "inline_equation", + "content": "t + \\Delta t = 1" + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": " and then sample " + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": " from the slow scheduler (see Eq. (3)). As shown in Table 4, increasing " + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": " (thus reducing the number of intervals) within a reasonable range improves IQA metrics. A plausible explanation is that the slightly larger perturbations introduced at intermediate predictions " + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\hat{X}_t^\\phi" + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "inline_equation", + "content": "X_{t}" + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": " encourage the model to be more robust to distribution shifts; see Fig. 3. This, in turn, reduces variance in its estimates of" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 316, + 69, + 362, + 118 + ], + "blocks": [ + { + "bbox": [ + 316, + 69, + 362, + 118 + ], + "lines": [ + { + "bbox": [ + 316, + 69, + 362, + 118 + ], + "spans": [ + { + "bbox": [ + 316, + 69, + 362, + 118 + ], + "type": "image", + "image_path": "59329c91986bef0bac8120c2cafb9aeb6f6b964a5bfc1a3082f28ab5456606aa.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 325, + 118, + 354, + 127 + ], + "lines": [ + { + "bbox": [ + 325, + 118, + 354, + 127 + ], + "spans": [ + { + "bbox": [ + 325, + 118, + 354, + 127 + ], + "type": "text", + "content": "LR Image" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 363, + 70, + 410, + 118 + ], + "blocks": [ + { + "bbox": [ + 363, + 70, + 410, + 118 + ], + "lines": [ + { + "bbox": [ + 363, + 70, + 410, + 118 + ], + "spans": [ + { + "bbox": [ + 363, + 70, + 410, + 118 + ], + "type": "image", + "image_path": "780bcaa054749caf97071e120c0425b2e0cd49781c665e6aaa0aae205ee8932e.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 118, + 404, + 127 + ], + "lines": [ + { + "bbox": [ + 369, + 118, + 404, + 127 + ], + "spans": [ + { + "bbox": [ + 369, + 118, + 404, + 127 + ], + "type": "text", + "content": "w/o " + }, + { + "bbox": [ + 369, + 118, + 404, + 127 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{hrcd}" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 370, + 130, + 498, + 142 + ], + "lines": [ + { + "bbox": [ + 370, + 130, + 498, + 142 + ], + "spans": [ + { + "bbox": [ + 370, + 130, + 498, + 142 + ], + "type": "text", + "content": "Figure 7. Ablation of training loss." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 411, + 70, + 459, + 118 + ], + "blocks": [ + { + "bbox": [ + 411, + 70, + 459, + 118 + ], + "lines": [ + { + "bbox": [ + 411, + 70, + 459, + 118 + ], + "spans": [ + { + "bbox": [ + 411, + 70, + 459, + 118 + ], + "type": "image", + "image_path": "81cb7bb39ac06e52da70564e0982e7baa84256e20ab80b984a683471f6a04282.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 118, + 449, + 127 + ], + "lines": [ + { + "bbox": [ + 419, + 118, + 449, + 127 + ], + "spans": [ + { + "bbox": [ + 419, + 118, + 449, + 127 + ], + "type": "text", + "content": "w/o " + }, + { + "bbox": [ + 419, + 118, + 449, + 127 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{adv}" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 459, + 70, + 506, + 118 + ], + "blocks": [ + { + "bbox": [ + 459, + 70, + 506, + 118 + ], + "lines": [ + { + "bbox": [ + 459, + 70, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 459, + 70, + 506, + 118 + ], + "type": "image", + "image_path": "eaa2d2c2b0495c668cf91ce3d829cc23eb13afe83f9c0838f74b81af34fb7ffc.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 467, + 118, + 496, + 127 + ], + "lines": [ + { + "bbox": [ + 467, + 118, + 496, + 127 + ], + "spans": [ + { + "bbox": [ + 467, + 118, + 496, + 127 + ], + "type": "text", + "content": "w/o " + }, + { + "bbox": [ + 467, + 118, + 496, + 127 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{iqa}" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 507, + 70, + 553, + 118 + ], + "blocks": [ + { + "bbox": [ + 507, + 70, + 553, + 118 + ], + "lines": [ + { + "bbox": [ + 507, + 70, + 553, + 118 + ], + "spans": [ + { + "bbox": [ + 507, + 70, + 553, + 118 + ], + "type": "image", + "image_path": "de9f49c534a1ad7105b677ceb26274e9349c8f5b32862bb613657d2966bb3ea7.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 522, + 118, + 537, + 126 + ], + "lines": [ + { + "bbox": [ + 522, + 118, + 537, + 126 + ], + "spans": [ + { + "bbox": [ + 522, + 118, + 537, + 126 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "table", + "bbox": [ + 315, + 198, + 553, + 273 + ], + "blocks": [ + { + "bbox": [ + 313, + 149, + 555, + 194 + ], + "lines": [ + { + "bbox": [ + 313, + 149, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 313, + 149, + 555, + 194 + ], + "type": "text", + "content": "Table 4. Ablation of fast-slow time sampling. We compare our fast-slow sampling approach with the " + }, + { + "bbox": [ + 313, + 149, + 555, + 194 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 149, + 555, + 194 + ], + "type": "text", + "content": "-interval method on DRealSR. The number of " + }, + { + "bbox": [ + 313, + 149, + 555, + 194 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 149, + 555, + 194 + ], + "type": "text", + "content": " and fast-scheduler timesteps are indicated in parentheses." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 198, + 553, + 273 + ], + "lines": [ + { + "bbox": [ + 315, + 198, + 553, + 273 + ], + "spans": [ + { + "bbox": [ + 315, + 198, + 553, + 273 + ], + "type": "table", + "html": "
MethodPSNR ↑LPIPS ↓MUSIQ ↑MANIQA ↑CLIPiQA ↑
N-Interval (50)28.610.280762.480.59140.6599
N-Interval (18)28.740.279262.900.59720.6605
N-Interval (4)28.660.281863.030.59720.6617
Slow Only (1000)28.780.280162.720.59160.6586
Fast-Slow (8)28.540.276064.010.60980.6634
Fast-Slow (4)28.620.283064.770.61560.6748
Fast-Slow (1)28.790.283563.010.59480.6605
", + "image_path": "a8a6833ad93af13a8fbe017ee77f82872c5a36a0f2d9aadd329607078a7a7afe.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "table_body" + } + ], + "index": 25 + }, + { + "type": "table", + "bbox": [ + 315, + 294, + 553, + 350 + ], + "blocks": [ + { + "bbox": [ + 348, + 280, + 519, + 291 + ], + "lines": [ + { + "bbox": [ + 348, + 280, + 519, + 291 + ], + "spans": [ + { + "bbox": [ + 348, + 280, + 519, + 291 + ], + "type": "text", + "content": "Table 5. Ablation of training loss on DRealSR." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 294, + 553, + 350 + ], + "lines": [ + { + "bbox": [ + 315, + 294, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 315, + 294, + 553, + 350 + ], + "type": "table", + "html": "
MethodPSNR ↑LPIPS ↓MUSIQ ↑MANIQA ↑CLIPiQA ↑
w/o Lhred28.520.301364.420.60610.6954
w/o Ladv28.450.296165.290.61540.7099
w/o Liaq28.640.295864.830.61110.7012
Ours28.500.297565.460.61720.7074
", + "image_path": "c161b7dde2254ee6deb1ba68b28d1481f3d196f82409e58ba6cbb812fa663cfb.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "table_body" + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 361, + 555, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 555, + 457 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 555, + 457 + ], + "type": "text", + "content": "the ODE trajectory's starting point, i.e., HR, resulting in sharper restorations. Moreover, compared to the " + }, + { + "bbox": [ + 313, + 361, + 555, + 457 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 361, + 555, + 457 + ], + "type": "text", + "content": "-interval strategy or slow-only scheduling, our fast-slow scheduler further boosts IQA metrics while maintaining comparable fidelity. This is because the slow scheduler ensures fine-grained SR flow estimation, whereas the fast scheduler enables efficient sampling. Based on these findings, we set the fast scheduler to 4 timesteps throughout our experiments." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 464, + 556, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 556, + 561 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 556, + 561 + ], + "type": "text", + "content": "Influence of loss function In addition to flow and consistency learning, we incorporate a GAN loss and an image quality alignment loss to further enhance SR performance, as detailed in Eq. (10). As demonstrated in Table 5 and Fig. 7, these losses all improve no-reference metrics and visual quality. The ablation highlights that eliminating any of these losses results in reduced SR quality, with consistency learning contributing a substantial improvement." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 573, + 388, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 573, + 388, + 585 + ], + "spans": [ + { + "bbox": [ + 313, + 573, + 388, + 585 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": "This paper presents FlowSR, a new approach for efficient one-step image super-resolution. FlowSR reformulates SR as a rectified flow, leveraging the strengths of iterative generative modeling. To enable high-quality single-step inference, we incorporate consistency learning and devise HR regularization to address its distillation target drifting issue. Additionally, a fast-slow time scheduling strategy is designed to enhance the efficiency and robustness of the consistency SR flow model. FlowSR contributes to the advancement of efficient real-world SR applications." + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11762" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": "The work described in this paper was supported in part by the National Key R&D Program of China (Grant No. 2023YFE0202700), Research Grants Council of the Hong Kong Special Administrative Region, China, under Project 14200824; and by the Hong Kong Innovation and Technology Fund, under Project MHP/092/22." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 194, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 194, + 295, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 194, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 61, + 194, + 295, + 228 + ], + "type": "text", + "content": "[1] Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang. Toward real-world single image super-resolution: A new benchmark and a new model. In ICCV, 2019. 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 229, + 296, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 229, + 296, + 272 + ], + "spans": [ + { + "bbox": [ + 61, + 229, + 296, + 272 + ], + "type": "text", + "content": "[2] Bin Chen, Gehui Li, Rongyuan Wu, Xindong Zhang, Jie Chen, Jian Zhang, and Lei Zhang. Adversarial diffusion compression for real-world image super-resolution. In CVPR, 2025. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 274, + 295, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 274, + 295, + 317 + ], + "spans": [ + { + "bbox": [ + 62, + 274, + 295, + 317 + ], + "type": "text", + "content": "[3] Qinpeng Cui, Yixuan Liu, Xinyi Zhang, Qiqi Bao, Zhongdao Wang, Qingmin Liao, Li Wang, Tian Lu, and Emad Barsoum. Taming diffusion prior for image super-resolution with domain shift sdes. NeurIPS, 2024. 1, 3, 4, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 319, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 319, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 62, + 319, + 294, + 352 + ], + "type": "text", + "content": "[4] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. TPAMI, 2020. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 353, + 294, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 353, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 62, + 353, + 294, + 386 + ], + "type": "text", + "content": "[5] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image super-resolution. In ECCV, 2014. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 387, + 295, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 387, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 387, + 295, + 430 + ], + "type": "text", + "content": "[6] Linwei Dong, Qingnan Fan, Yihong Guo, Zhonghao Wang, Qi Zhang, Jinwei Chen, Yawei Luo, and Changqing Zou. Tsd-sr: One-step diffusion with target score distillation for real-world image super-resolution. In CVPR, 2025. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 432, + 295, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 432, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 62, + 432, + 295, + 474 + ], + "type": "text", + "content": "[7] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. NeurIPS, 2014. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 477, + 295, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 477, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 62, + 477, + 295, + 531 + ], + "type": "text", + "content": "[8] Xiao He, Huaao Tang, Zhijun Tu, Junchao Zhang, Kun Cheng, Hanting Chen, Yong Guo, Mingrui Zhu, Nannan Wang, Xinbo Gao, et al. One step diffusion-based super-resolution with time-aware distillation. arXiv preprint arXiv:2408.07476, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 533, + 295, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 533, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 62, + 533, + 295, + 577 + ], + "type": "text", + "content": "[9] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. *NeruIPS*, 2017. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 578, + 295, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 578, + 295, + 600 + ], + "spans": [ + { + "bbox": [ + 57, + 578, + 295, + 600 + ], + "type": "text", + "content": "[10] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. *NeruIPS*, 2020. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 601, + 295, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 295, + 644 + ], + "type": "text", + "content": "[11] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 647, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 678 + ], + "type": "text", + "content": "[12] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 680, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 295, + 712 + ], + "type": "text", + "content": "[13] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In ICCV, 2021. 6" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[14] Dongjun Kim, Chieh-Hsin Lai, Wei-Hsiang Liao, Naoki Murata, Yuhta Takida, Toshimitsu Uesaka, Yutong He, Yuki Mitsufuji, and Stefano Ermon. Consistency trajectory models: Learning probability flow ode trajectory of diffusion. In ICLR, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 130, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 184 + ], + "type": "text", + "content": "[15] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In CVPR, 2017. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 186, + 553, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 231 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 231 + ], + "type": "text", + "content": "[16] Jianze Li, Jiezhang Cao, Zichen Zou, Xiongfei Su, Xin Yuan, Yulun Zhang, Yong Guo, and Xiaokang Yang. Distillation-free one-step diffusion for real-world image super-resolution. arXiv preprint arXiv:2410.04224, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 232, + 553, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 232, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 553, + 275 + ], + "type": "text", + "content": "[17] Yawei Li, Kai Zhang, Jingyun Liang, Jiezhang Cao, Ce Liu, Rui Gong, Yulun Zhang, Hao Tang, Yun Liu, Denis Demandolx, et al. Lsdir: A large scale dataset for image restoration. In CVPR, 2023. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 277, + 553, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 277, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 317, + 277, + 553, + 311 + ], + "type": "text", + "content": "[18] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swin transformer. In ICCV Workshop, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 313, + 553, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 313, + 553, + 356 + ], + "spans": [ + { + "bbox": [ + 317, + 313, + 553, + 356 + ], + "type": "text", + "content": "[19] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Yu Qiao, Wanli Ouyang, and Chao Dong. Diffir: Toward blind image restoration with generative diffusion prior. In ECCV, 2024. 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 358, + 553, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 358, + 553, + 391 + ], + "spans": [ + { + "bbox": [ + 317, + 358, + 553, + 391 + ], + "type": "text", + "content": "[20] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. In ICLR, 2023. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 393, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 553, + 425 + ], + "type": "text", + "content": "[21] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. In ICLR, 2023. 2, 3, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 428, + 553, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 428, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 316, + 428, + 553, + 460 + ], + "type": "text", + "content": "[22] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, et al. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. In ICLR, 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 463, + 553, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 463, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 317, + 463, + 553, + 495 + ], + "type": "text", + "content": "[23] Eric Luhman and Troy Luhman. Knowledge distillation in iterative generative models for improved sampling speed. arXiv preprint arXiv:2101.02388, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 498, + 553, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 498, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 317, + 498, + 553, + 540 + ], + "type": "text", + "content": "[24] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 543, + 553, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 543, + 553, + 576 + ], + "spans": [ + { + "bbox": [ + 316, + 543, + 553, + 576 + ], + "type": "text", + "content": "[25] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal processing letters, 2012. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 578, + 553, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 553, + 620 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 553, + 620 + ], + "type": "text", + "content": "[26] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In AAAI, 2024. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 624, + 553, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 666 + ], + "type": "text", + "content": "[27] Mehdi Noroozi, Isma Hadji, Brais Martinez, Adrian Bulat, and Georgios Tzimiropoulos. You only need one step: Fast super-resolution with stable diffusion via scale distillation. In ECCV, 2024. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[28] Long Peng, Wenbo Li, Renjing Pei, Jingjing Ren, Jiaqi Xu, Yang Wang, Yang Cao, and Zheng-Jun Zha. Towards realistic data generation for real-world super-resolution. In ICLR, 2025. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11763" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 294, + 172 + ], + "type": "text", + "content": "[30] Jingjing Ren, Wenbo Li, Haoyu Chen, Renjing Pei, Bin Shao, Yong Guo, Long Peng, Fenglong Song, and Lei Zhu. Ultrapixel: Advancing ultra high-resolution image synthesis to new peaks. NeurIPS, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 294, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 294, + 215 + ], + "type": "text", + "content": "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 249 + ], + "type": "text", + "content": "[32] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. TPAMI, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "type": "text", + "content": "[33] Axel Sauer, Frederic Boesel, Tim Dockhorn, Andreas Blattmann, Patrick Esser, and Robin Rombach. Fast high-resolution image synthesis with latent adversarial diffusion distillation. In SIGGRAPH Asia, 2024. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 293, + 294, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 294, + 325 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 294, + 325 + ], + "type": "text", + "content": "[34] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In ECCV, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 327, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 370 + ], + "type": "text", + "content": "[35] Johannes Schusterbauer, Ming Gui, Pingchuan Ma, Nick Stracke, Stefan Andreas Baumann, Vincent Tao Hu, and Björn Ommer. Boosting latent diffusion with flow matching. In ECCV, 2024. 3, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 371, + 294, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 393 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 393 + ], + "type": "text", + "content": "[36] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 393, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 294, + 437 + ], + "type": "text", + "content": "[37] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 437, + 294, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 469 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 469 + ], + "type": "text", + "content": "[38] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In ICML, 2023. 2, 3, 5, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 471, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 514 + ], + "type": "text", + "content": "[39] Haoze Sun, Wenbo Li, Jianzhuang Liu, Haoyu Chen, Renjing Pei, Xueyi Zou, Youliang Yan, and Yujiu Yang. Coser: Bridging image and language for cognitive super-resolution. In CVPR, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 514, + 294, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 294, + 558 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 294, + 558 + ], + "type": "text", + "content": "[40] Fu-Yun Wang, Zhaoyang Huang, Alexander Bergman, Dazhong Shen, Peng Gao, Michael Lingelbach, Keqiang Sun, Weikang Bian, Guanglu Song, Yu Liu, et al. Phased consistency models. NeurIPS, 2024. 3, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 559, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 591 + ], + "type": "text", + "content": "[41] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In AAAI, 2023. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 592, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 624 + ], + "type": "text", + "content": "[42] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. IJCV, 2024. 1, 2, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 625, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 679 + ], + "type": "text", + "content": "[43] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "text", + "content": "[44] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In ICCV Workshop, 2021. 2, 6" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 315, + 72, + 554, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 554, + 117 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 554, + 117 + ], + "type": "text", + "content": "[45] Yufei Wang, Wenhan Yang, Xinyuan Chen, Yaohui Wang, Lanqing Guo, Lap-Pui Chau, Ziwei Liu, Yu Qiao, Alex C Kot, and Bihan Wen. Sinsr: diffusion-based image superresolution in a single step. In CVPR, 2024. 1, 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 118, + 554, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 118, + 554, + 161 + ], + "spans": [ + { + "bbox": [ + 315, + 118, + 554, + 161 + ], + "type": "text", + "content": "[46] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. NeurIPS, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 162, + 554, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 162, + 554, + 206 + ], + "spans": [ + { + "bbox": [ + 315, + 162, + 554, + 206 + ], + "type": "text", + "content": "[47] Pengxu Wei, Ziwei Xie, Hannan Lu, Zongyuan Zhan, Qixiang Ye, Wangmeng Zuo, and Liang Lin. Component divide-and-conquer for real-world image super-resolution. In ECCV, 2020. 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 207, + 554, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 207, + 554, + 240 + ], + "spans": [ + { + "bbox": [ + 315, + 207, + 554, + 240 + ], + "type": "text", + "content": "[48] Rongyuan Wu, Lingchen Sun, Zhiyuan Ma, and Lei Zhang. One-step effective diffusion network for real-world image super-resolution. NeurIPS, 2024. 1, 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 241, + 554, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 241, + 554, + 274 + ], + "spans": [ + { + "bbox": [ + 315, + 241, + 554, + 274 + ], + "type": "text", + "content": "[49] Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. Seesr: Towards semantics-aware real-world image super-resolution. In CVPR, 2024. 1, 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 276, + 554, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 276, + 554, + 308 + ], + "spans": [ + { + "bbox": [ + 315, + 276, + 554, + 308 + ], + "type": "text", + "content": "[50] Zhisheng Xiao, Karsten Kreis, and Arash Vahdat. Tackling the generative learning trilemma with denoising diffusion gans. In ICLR, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 309, + 554, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 309, + 554, + 365 + ], + "spans": [ + { + "bbox": [ + 315, + 309, + 554, + 365 + ], + "type": "text", + "content": "[51] Rui Xie, Ying Tai, Chen Zhao, Kai Zhang, Zhenyu Zhang, Jun Zhou, Xiaogian Ye, Qian Wang, and Jian Yang. Addsr: Accelerating diffusion-based blind superresolution with adversarial diffusion distillation. arXiv preprint arXiv:2404.01717, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 365, + 554, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 365, + 554, + 408 + ], + "spans": [ + { + "bbox": [ + 315, + 365, + 554, + 408 + ], + "type": "text", + "content": "[52] Hanshu Yan, Xingchao Liu, Jiachun Pan, Jun Hao Liew, Qiang Liu, and Jiashi Feng. Perflow: Piecewise rectified flow as universal plug-and-play accelerator. NeurIPS, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 410, + 554, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 410, + 554, + 464 + ], + "spans": [ + { + "bbox": [ + 315, + 410, + 554, + 464 + ], + "type": "text", + "content": "[53] Ling Yang, Zixiang Zhang, Zhilong Zhang, Xingchao Liu, Minkai Xu, Wentao Zhang, Chenlin Meng, Stefano Ermon, and Bin Cui. Consistency flow matching: Defining straight flows with velocity consistency. arXiv preprint arXiv:2407.02398, 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 466, + 554, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 466, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 315, + 466, + 554, + 510 + ], + "type": "text", + "content": "[54] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Maniaq: Multi-dimension attention network for no-reference image quality assessment. In CVPR, 2022. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 511, + 554, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 511, + 554, + 554 + ], + "spans": [ + { + "bbox": [ + 315, + 511, + 554, + 554 + ], + "type": "text", + "content": "[55] Tao Yang, Rongyuan Wu, Peiran Ren, Xuansong Xie, and Lei Zhang. Pixel-aware stable diffusion for realistic image super-resolution and personalized stylization. In ECCV, 2024.6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 315, + 555, + 554, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 555, + 554, + 599 + ], + "spans": [ + { + "bbox": [ + 315, + 555, + 554, + 599 + ], + "type": "text", + "content": "[56] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and Bill Freeman. Improved distribution matching distillation for fast image synthesis. NeurIPS, 2024. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 315, + 600, + 554, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 600, + 554, + 644 + ], + "spans": [ + { + "bbox": [ + 315, + 600, + 554, + 644 + ], + "type": "text", + "content": "[57] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In CVPR, 2024. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 315, + 646, + 554, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 646, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 315, + 646, + 554, + 689 + ], + "type": "text", + "content": "[58] Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. Scaling up to excellence: Practicing model scaling for photorealistic image restoration in the wild. In CVPR, 2024. 1, 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 315, + 690, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 690, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 690, + 554, + 713 + ], + "type": "text", + "content": "[59] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Reshift: Efficient diffusion model for image super" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11764" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 275 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "resolution by residual shifting. NeurIPS, 2023. 1, 2, 4, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 128 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 128 + ], + "type": "text", + "content": "[60] Zongsheng Yue, Kang Liao, and Chen Change Loy. Arbitrary-steps image super-resolution via diffusion inversion. In CVPR, 2025. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 130, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 294, + 173 + ], + "type": "text", + "content": "[61] Aiping Zhang, Zongsheng Yue, Renjing Pei, Wenqi Ren, and Xiaochun Cao. Degradation-guided one-step image super-resolution with diffusion priors. arXiv preprint arXiv:2409.17058, 2024. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 207 + ], + "type": "text", + "content": "[62] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In ICCV, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 209, + 294, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 240 + ], + "type": "text", + "content": "[63] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 243, + 294, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 294, + 275 + ], + "type": "text", + "content": "[64] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5, 6" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "11765" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/FastJSMA_ Accelerating Jacobian-based Saliency Map Attacks through Gradient Decoupling/caddbdfc-9052-49a6-8b20-7e4e162ec5a3_content_list.json b/2025/FastJSMA_ Accelerating Jacobian-based Saliency Map Attacks through Gradient Decoupling/caddbdfc-9052-49a6-8b20-7e4e162ec5a3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..806aaec8a1f901c55290b3bcc2eb0381eabced4e --- /dev/null +++ b/2025/FastJSMA_ Accelerating Jacobian-based Saliency Map Attacks through Gradient Decoupling/caddbdfc-9052-49a6-8b20-7e4e162ec5a3_content_list.json @@ -0,0 +1,1778 @@ +[ + { + "type": "text", + "text": "FastJSMA: Accelerating Jacobian-based Saliency Map Attacks through Gradient Decoupling", + "text_level": 1, + "bbox": [ + 138, + 128, + 859, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhenghao $\\mathrm{Gao}^{1,*}$ Shengjie $\\mathrm{Xu}^{2,*}$ Zijing Li $^{2}$ Meixi Chen $^{3}$ Chaojian $\\mathrm{Yu}^{4}$ Yuanjie $\\mathrm{Shao}^{4}$ Changxin $\\mathrm{Gao}^{1,\\dagger}$", + "bbox": [ + 153, + 204, + 841, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1School of Artificial Intelligence and Automation, Huazhong University of Science and Technology", + "bbox": [ + 171, + 222, + 826, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ School of Software Engineering, Huazhong University of Science and Technology", + "bbox": [ + 250, + 239, + 745, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Journalism and Information Communication School, Huazhong University of Science and Technology", + "bbox": [ + 191, + 257, + 803, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ School of Electronic Information and Communications, Huazhong University of Science and Technology", + "bbox": [ + 184, + 276, + 812, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{u202215226,u202317280,u202317273,u202317218,yuchaojian,shaoyuanjie,cgao}@hust.edu.cn", + "bbox": [ + 202, + 296, + 787, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 344, + 326, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Adversarial attack plays a critical role in evaluating the robustness of deep learning models. Jacobian-based Saliency Map Attack (JSMA) is an interpretable adversarial method that offers excellent pixel-level control and provides valuable insights into model vulnerabilities. However, its quadratic computational complexity $O(M^2 \\times N)$ renders it impractical for large-scale datasets, limiting its application despite its inherent value. This paper proposes FastJSMA, an efficient attack method that addresses these computational limitations. Our approach introduces a gradient decoupling mechanism that decomposes the Jacobian calculation into complementary class suppression ( $g^{-}$ ) and class excitation ( $g^{+}$ ) gradients, reducing complexity to $O(M\\sqrt{N})$ . Additionally, we implement a class probing mechanism and an adaptive saliency threshold to further optimize the process. Experimental results across multiple datasets demonstrate that FastJSMA maintains comparable attack success rates while dramatically reducing computation time—requiring only 2.9% of JSMA's processing time on CIFAR-10 and 1.2% on CIFAR-100, and successfully operating on ImageNet where traditional JSMA fails due to memory constraints. This advancement enables the practical application of interpretable saliency map-based attacks on large-scale datasets, balancing effectiveness with computational efficiency.", + "bbox": [ + 89, + 376, + 485, + 755 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 784, + 220, + 799 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1.1. Research Background", + "text_level": 1, + "bbox": [ + 91, + 809, + 295, + 825 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning has achieved breakthrough progress in computer vision, natural language processing, and other fields,", + "bbox": [ + 89, + 832, + 483, + 863 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "but its security issues have increasingly attracted attention. In 2013, Szegedy et al. [1] first discovered the sensitivity of deep neural networks to small perturbations, i.e., by adding imperceptible perturbations to inputs, models can be induced to produce incorrect predictions. This discovery revealed potential vulnerabilities in deep learning systems and drove rapid development in adversarial attack research.", + "bbox": [ + 511, + 345, + 906, + 450 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Adversarial attack techniques have evolved from simple approaches like the Fast Gradient Sign Method (FGSM) [2] to complex methods such as PGD [3] and C&W attacks [4], each offering different trade-offs between effectiveness, efficiency, and perturbation magnitude.", + "bbox": [ + 511, + 457, + 908, + 534 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1.2. Research Motivation", + "text_level": 1, + "bbox": [ + 513, + 561, + 709, + 578 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The Jacobian-based Saliency Map Attack (JSMA) [5] is an adversarial method that identifies and modifies the most influential pixels through Jacobian matrix analysis. It offers unique advantages including intuitive visualization through saliency maps that enhance interpretability, precise control over modified pixels, and strong effectiveness across various datasets. Despite these merits and attempts to improve it through weighted approaches like WJSMA [6], JSMA remains largely underutilized due to prohibitive computational costs. Its full Jacobian matrix calculation requires $O(M^2 \\times N)$ complexity (where $M$ is the number of pixels and $N$ is the number of classes), creating an insurmountable barrier for large-scale applications like ImageNet [7].", + "bbox": [ + 511, + 590, + 908, + 789 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This limitation stems from three factors: (a) separate backward propagation for each output node; (b) saliency evaluations across the entire Jacobian matrix; and (c) repeated calculations across iterations. Consequently, the adversarial machine learning community has effectively abandoned JSMA's valuable properties due to computational constraints rather than methodological limitations.", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "0* Equal contribution", + "bbox": [ + 107, + 875, + 223, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{0\\dagger}$ Corresponding author", + "bbox": [ + 109, + 887, + 238, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1506", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1.3. Contributions", + "text_level": 1, + "bbox": [ + 91, + 90, + 235, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To revitalize this valuable but computationally constrained method, we propose FastJSMA (Fast Jacobian-based Saliency Map Attack), an efficient adversarial attack based on gradient decoupling approximation:", + "bbox": [ + 89, + 112, + 482, + 172 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Gradient Decoupling Mechanism: We decompose the Jacobian matrix into complementary class suppression $(g^{-})$ and excitation $(g^{+})$ gradients, reducing computational complexity from $O(M^2 \\times N)$ to $O(M\\sqrt{N})$ while maintaining attack effectiveness through dynamic weight allocation.", + "- Class Probing and Adaptive Threshold: We introduce a square-root sampling heuristic to efficiently identify vulnerable decision boundaries and an adaptive threshold mechanism to prioritize influential pixels, optimizing both efficiency and effectiveness.", + "- Comprehensive Evaluation: Our experiments demonstrate dramatic efficiency gains—requiring only $2.9\\%$ of JSMANT's processing time on CIFAR-10 and achieving up to $81.4 \\times$ speedup on CIFAR-100 while maintaining comparable attack success rates. Most notably, FastJSMA successfully processes ImageNet samples, where traditional JSMA and WJSMA fail entirely due to memory constraints." + ], + "bbox": [ + 89, + 175, + 483, + 463 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 476, + 233, + 491 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Overview of Adversarial Attack Methods", + "text_level": 1, + "bbox": [ + 89, + 501, + 442, + 516 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Adversarial attacks can be categorized as white-box or black-box attacks based on the attacker's knowledge of the model. White-box attacks assume the attacker has complete access to the model structure and parameters, with representative methods including FGSM [2] and PGD [3]. Black-box attacks can only access the model's input-output interface, primarily relying on transferability [8] or query optimization [9]. Recent work has also focused on improving adversarial training defenses [10-12].", + "bbox": [ + 89, + 523, + 482, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "From the perspective of attack objectives, they can be divided into targeted and non-targeted attacks. Targeted attacks aim to misclassify input samples as specified target classes, such as DeepFool [13]; non-targeted attacks only need to cause incorrect predictions, regardless of the specific error class, such as FGSM. These two types of attack methods have their respective advantages in practical applications: targeted attacks enable more precise control, while non-targeted attacks typically have higher success rates and lower computational costs.", + "bbox": [ + 89, + 659, + 482, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Gradient-based adversarial attacks have evolved significantly. After FGSM was proposed, researchers developed improved variants: (1) MI-FGSM [14] introduces momentum terms; (2) NI-FGSM [15] utilizes Nesterov accelerated gradients; (3) TI-FGSM [16] enhances transferability through translation invariance; (4) DI-FGSM [17] adopts", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "diversified input strategies. Recent research by Mao et al. [18] discovered that adversarial attacks are reversible with natural supervision, providing new insights into the fundamental properties of adversarial perturbations. These methods optimize gradient-based attack strategies from different perspectives but have not fully addressed the issue of interpretability.", + "bbox": [ + 511, + 90, + 906, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Detailed Analysis of JSMA", + "text_level": 1, + "bbox": [ + 511, + 205, + 756, + 220 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As an interpretable attack method, JSMA was proposed by Papernot et al. [5], with its core idea being the use of the Jacobian matrix to identify input features that have the greatest impact on model output. This approach builds upon fundamental research in saliency detection, such as the boolean map approach [19], which provides theoretical foundations for identifying visually significant regions. Specifically, for input $x$ and target class $t$ , JSMA first calculates the Jacobian matrix:", + "bbox": [ + 511, + 227, + 906, + 362 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nJ _ {F} (x) = \\frac {\\partial F (x)}{\\partial x} = \\left[ \\frac {\\partial F _ {j} (x)}{\\partial x _ {i}} \\right] _ {i, j} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 372, + 905, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $F_{j}(x)$ represents the model's prediction probability for class $j$ . Based on the Jacobian matrix, JSMA defines the saliency map $S$ :", + "bbox": [ + 511, + 412, + 905, + 459 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nS (x, t) [ i ] = \\left\\{ \\begin{array}{l} 0, \\text {i f} \\frac {\\partial F _ {t} (x)}{\\partial x _ {i}} < 0 \\text {o r} \\sum_ {j \\neq t} \\frac {\\partial F _ {j} (x)}{\\partial x _ {i}} > 0 \\\\ \\left| \\frac {\\partial F _ {t} (x)}{\\partial x _ {i}} \\right| \\cdot \\left| \\sum_ {j \\neq t} \\frac {\\partial F _ {j} (x)}{\\partial x _ {i}} \\right|, \\text {o t h e r w i s e} \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 465, + 903, + 523 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Subsequently, researchers have made various improvements to JSMA: (1) JSMA++ [8] enhances attack efficiency by optimizing saliency calculations; (2) One-pixel JSMA [20] restricts modifications to a single pixel, achieving more precise control; (3) Adaptive JSMA [21] introduces adaptive threshold strategies, improving attack success rates. The concept of saliency maps has also been extended to other domains, such as 3D point clouds [22], demonstrating the versatility of saliency-based approaches across different data modalities.", + "bbox": [ + 511, + 523, + 905, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "An important advancement is the Weighted JSMA (WJSMA) [6], which addresses limitations in the original JSMA's saliency calculation. WJSMA penalizes gradients associated with small probabilities to create more balanced saliency maps. WJSMA incorporates class probabilities as weights in the saliency calculation, setting the saliency to zero when specific gradient conditions are not met. This weighting mechanism makes WJSMA faster than the original JSMA while maintaining good balance between attack success rate and computational cost. However, even with these improvements, WJSMA still inherits the fundamental computational inefficiency of JSMA for large-scale datasets, as it retains the $O(M^2 \\times N)$ complexity that creates memory and processing bottlenecks on datasets like ImageNet.", + "bbox": [ + 511, + 674, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1507", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Related Optimization Techniques", + "text_level": 1, + "bbox": [ + 89, + 90, + 382, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the efficiency problems of adversarial attacks in large-scale scenarios, researchers have proposed various optimization strategies. One class of methods improves efficiency by enhancing gradient computation, such as OnePixel Attack [20], which reduces search space through evolutionary algorithms, and SparseFool [23], which optimizes computational efficiency by leveraging sparsity.", + "bbox": [ + 89, + 113, + 483, + 220 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Another class of methods proposes specialized optimization strategies for large-scale datasets. For example, Fast Gradient Method [24] enhances attack efficiency through ensemble techniques, while AutoAttack [25] adopts automated strategies to select optimal attack parameters.", + "bbox": [ + 89, + 220, + 483, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent research has also explored algorithmic optimizations specifically targeting the computational complexity of gradient-based attacks. These approaches include gradient approximation techniques [26], adaptive norm constraints for efficient minimum-norm attacks [27], and decomposition strategies that reduce the dimensionality of the optimization space [21].", + "bbox": [ + 89, + 297, + 483, + 404 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "These optimization methods have their respective advantages and disadvantages: gradient-based methods offer high computational efficiency but lower interpretability; optimization-based methods provide good results but incur high computational costs; heuristic-based methods are simple and fast but have unstable success rates. In comparison, our proposed FastJSMA maintains JSMA's interpretability while significantly improving computational efficiency through dual-gradient approximation, effectively reducing the complexity from $O(M^2 \\times N)$ to $O(M\\sqrt{N})$ and making saliency map-based attacks practical for large-scale datasets for the first time.", + "bbox": [ + 89, + 405, + 483, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 604, + 225, + 622 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section details the technical principles of FastJSMA. We first establish the mathematical model of the problem, then elaborate on the design rationale of the dual-gradient approximation strategy, followed by a detailed explanation of the saliency map generation process, and finally theoretically analyze the computational complexity of the algorithm.", + "bbox": [ + 89, + 630, + 483, + 737 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Modeling and Attack Objectives", + "text_level": 1, + "bbox": [ + 89, + 750, + 444, + 767 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For an input sample $x \\in \\mathbb{R}^{C \\times H \\times W}$ and a target classification model $f: \\mathbb{R}^{C \\times H \\times W} \\to \\mathbb{R}^N$ , where $C, H, W$ represent the number of channels, height, and width respectively, and $N$ is the total number of classes. Given the original predicted class $t = \\arg \\max f(x)$ , the goal of the adversarial attack is to find a perturbation $\\delta$ such that:", + "bbox": [ + 89, + 772, + 483, + 864 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\arg \\max f (x + \\delta) \\neq t \\quad \\text {s . t .} \\quad \\| \\delta \\| _ {0} \\leq \\epsilon \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 883, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon$ represents the allowed number of pixels to modify, and $\\| \\cdot \\|_0$ is the $L_0$ norm. The reasons for choosing the $L_0$ norm constraint rather than other norms (such as $L_2$ or $\\mathrm{L\\infty}$ ) are: (1) $L_0$ norm directly limits the number of modified pixels, conforming to the principle of minimal intervention; (2) a limited number of modifications makes the attack process more interpretable and understandable; (3) maintains consistent constraint forms with traditional JSMA for fair comparison.", + "bbox": [ + 511, + 90, + 903, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Traditional JSMA computes the complete Jacobian matrix $J = \\nabla_{x}f(x)$ with $O(M^2\\times N)$ complexity, where $M$ represents the total number of pixels and $N$ denotes the number of classification classes. This substantial computational burden and memory requirements render JSMA impractical for large-scale datasets and real-world applications.", + "bbox": [ + 511, + 228, + 903, + 332 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "FastJSMA addresses this limitation through a dual-gradient approximation strategy that reduces complexity to $O(M\\sqrt{N})$ while maintaining attack effectiveness.", + "bbox": [ + 511, + 335, + 903, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Target Function Construction", + "text_level": 1, + "bbox": [ + 511, + 393, + 777, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The core mathematical insight that enables FastJSMA's efficiency breakthrough lies in our recognition that the Jacobian-based saliency computation in traditional JSMA can be decomposed into two complementary objectives: first, decreasing the probability of the current class, and second, increasing the probability of alternative classes. This decomposition forms the theoretical foundation of our dual-gradient approximation strategy.", + "bbox": [ + 511, + 416, + 903, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Analyzing the mathematical structure of traditional JSMA reveals that its saliency map computation essentially measures two gradient components for each pixel: how modifying that pixel decreases the current class score and how it increases other class scores. While traditional JSMA explicitly computes these gradients through the full Jacobian matrix, we observed that these two objectives could be approximated through carefully designed loss functions without requiring the prohibitively expensive Jacobian computation.", + "bbox": [ + 511, + 539, + 905, + 690 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To avoid computing the complete Jacobian matrix, we propose a dual-gradient approximation strategy. The core idea of this strategy is to construct two complementary objective functions that simulate the enhancement effect on the target class and the suppression effect on other classes in JSMA. The specific definitions are as follows:", + "bbox": [ + 511, + 691, + 905, + 782 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "1. Class Suppression Gradient: Construct a virtual label $\\hat{y}^{-} \\in \\mathbb{R}^{N}$ , where", + "bbox": [ + 513, + 787, + 905, + 816 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} _ {i} ^ {-} = \\left\\{ \\begin{array}{l l} f (x) _ {i} - 1 & i = t \\\\ f (x) _ {i} + \\sqrt {N} & i = a (\\text {a t t e m p t c l a s s , i f s p e c i f i e d}) \\\\ f (x) _ {i} & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 830, + 911, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1508", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e8442bc3756db7a5fe074877cc37b4f65160e8e9bd2315d3cd688ade07e7cf9e.jpg", + "image_caption": [ + "Figure 1. Architectural comparison between traditional JSMA and our proposed FastJSMA. Left: Traditional JSMA calculates the full Jacobian matrix with computational complexity $O(M^2 \\times N)$ and relies on an iterative process to achieve misclassification, making it prohibitively expensive for large-scale datasets. Right: FastJSMA employs a class probing mechanism to select $\\sqrt{N}$ candidate classes and a gradient decoupling strategy that approximates the saliency computation through class suppression gradient $(g^-)$ and class excitation gradient $(g^+)$ . The class suppression gradient reduces the current class probability while boosting an attempt class by $\\sqrt{N}$ , and the significance threshold efficiently filters the most influential pixels. This approach reduces complexity to $O(M\\sqrt{N})$ while maintaining comparable effectiveness." + ], + "image_footnote": [], + "bbox": [ + 130, + 88, + 872, + 565 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Calculate the gradient:", + "bbox": [ + 109, + 699, + 264, + 714 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ng ^ {-} = \\nabla_ {x} \\mathcal {L} (f (x), \\hat {y} ^ {-}) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 728, + 482, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This gradient reflects the direction of reducing the prediction probability of the current class and optionally enhancing a specific alternative class if an attempt parameter is provided.", + "bbox": [ + 109, + 758, + 482, + 819 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2. Class Excitation Gradient: Construct an enhancement label $\\hat{y}^{+}\\in \\mathbb{R}^{N}$ , where", + "bbox": [ + 89, + 820, + 482, + 849 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} _ {i} ^ {+} = \\left\\{ \\begin{array}{l l} f (x) _ {i} - 1 & i = t \\\\ f (x) _ {i} + 1 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 863, + 482, + 904 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Calculate the gradient:", + "bbox": [ + 532, + 699, + 687, + 714 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ng ^ {+} = \\nabla_ {x} \\mathcal {L} (f (x), \\hat {y} ^ {+}) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 722, + 903, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This gradient indicates the direction of enhancing the prediction probability of other classes while suppressing the current class.", + "bbox": [ + 532, + 747, + 906, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This formulation represents a significant mathematical innovation over traditional JSMA. Rather than explicitly computing $N \\times M$ partial derivatives as required by the Jacobian matrix, our approach leverages automatic differentiation to compute just two gradient vectors—reducing the computational complexity by a factor proportional to the number of classes $N$ and the number of pixels $M$ . Import-", + "bbox": [ + 511, + 795, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1509", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tantly, this efficiency gain comes without compromising the underlying attack methodology; the dual-gradient approach preserves the core mathematical principles of JSMA while dramatically reducing its computational footprint.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The design of these two gradients cleverly avoids directly computing the Jacobian matrix while retaining the core idea of JSMA: simultaneously considering the suppression of the target class and the excitation of other classes.", + "bbox": [ + 89, + 151, + 482, + 224 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Saliency Map Generation", + "text_level": 1, + "bbox": [ + 89, + 236, + 323, + 253 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Traditional JSMA generates saliency maps through exhaustive computation of the Jacobian matrix followed by pairwise evaluation of pixel modifications—a process that scales quadratically with the number of pixels. FastJSMA fundamentally reconceptualizes this approach through an elegant mathematical transformation that yields equivalent results with substantially reduced computation.", + "bbox": [ + 89, + 258, + 482, + 363 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on the dual-gradient approximation, we apply gradient normalization to ensure balanced contribution while strategically emphasizing the class suppression component:", + "bbox": [ + 89, + 364, + 482, + 409 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ng _ {\\text {n o r m a l i z e d}} ^ {-} = \\frac {g ^ {-}}{\\operatorname {m e a n} (| g ^ {-} |)} \\times N \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 421, + 482, + 454 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ng _ {\\text {n o r m a l i z e d}} ^ {+} = \\frac {g ^ {+}}{\\operatorname {m e a n} (| g ^ {+} |)} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 463, + 480, + 496 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This normalization procedure includes a deliberate engineering optimization: the class suppression gradient is scaled by a factor of $N$ (the number of classes) to amplify its influence in the overall saliency calculation. Our empirical investigation indicates that this scaling enhances attack effectiveness by emphasizing the suppression of the original class prediction.", + "bbox": [ + 89, + 500, + 482, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We then integrate these normalized gradients to obtain a comprehensive gain matrix:", + "bbox": [ + 89, + 606, + 482, + 636 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nG = g _ {\\text {n o r m a l i z e d}} ^ {+} + g _ {\\text {n o r m a l i z e d}} ^ {-} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 650, + 480, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This matrix captures the bidirectional influence of modifying each pixel on both decreasing the target class probability and increasing alternative class probabilities—precisely the information encapsulated in traditional JSMA's saliency computation, but obtained through a mathematically equivalent yet computationally superior formulation.", + "bbox": [ + 89, + 674, + 482, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After obtaining the gain matrix, we apply a significance threshold to filter out pixels with minimal impact. This threshold operation is visually represented as a funnel-like process in Figure 1, focusing computational resources only on the most influential pixels. By implementing this significance threshold, we effectively reduce the search space while ensuring that the selected pixels have meaningful contributions to the adversarial perturbation.", + "bbox": [ + 89, + 780, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This selective modification approach, governed by both the saliency ranking and the significance threshold, ensures efficient resource allocation by concentrating perturbations on the most influential and impactful pixels.", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Attack Success Rate Optimization via Class Probing", + "text_level": 1, + "bbox": [ + 511, + 160, + 903, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A significant enhancement in our implementation is the introduction of a class-specific probing mechanism, visually illustrated in Figure 1 as \"Class Probing Mechanism\". This component substantially improves attack success rates while maintaining the untargeted nature of FastJSMA by intelligently selecting candidate classes for the attempt parameter. The optimization addresses a fundamental challenge in untargeted adversarial attacks: determining which gradient directions most efficiently induce misclassification.", + "bbox": [ + 511, + 196, + 903, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our implementation employs an intelligent candidate class exploration strategy governed by a square-root sampling heuristic:", + "bbox": [ + 511, + 348, + 903, + 393 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\text {n u m} _ {\\text {c a n d i d a t e s}} = \\sqrt {N} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 629, + 407, + 903, + 425 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N$ represents the number of classes in the classification problem. This square-root relationship establishes a principled balance between computational efficiency and exploration thoroughness—scaling sublinearly with the class space dimensionality to maintain tractability even for large-scale classification problems.", + "bbox": [ + 511, + 431, + 903, + 521 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For each candidate class $c$ among the top numCandidates predicted classes (excluding the original class), the algorithm:", + "bbox": [ + 511, + 522, + 903, + 566 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Applies FastJSMA with class $c$ as an \"attempt\" parameter, which specifically modifies the class suppression gradient $(g^{-})$ in two ways: - Decreases the current class probability (by setting $\\hat{y}_i^- = f(x)_i - 1$ for $i = t$ ) - Boosts the attempt class $c$ by a factor of $\\sqrt{N}$ (by setting $\\hat{y}_i^- = f(x)_i + \\sqrt{N}$ for $i = c$ )", + "2. Evaluates whether this modified gradient direction successfully induces misclassification", + "3. Returns the first successful perturbation, terminating further exploration" + ], + "bbox": [ + 511, + 566, + 903, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This dual action of decreasing the current class while simultaneously boosting a specific attempt class efficiently guides the gradient toward vulnerable regions of the decision boundary, substantially improving attack success rates without requiring exhaustive exploration of all possible class transitions.", + "bbox": [ + 511, + 719, + 903, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It is crucial to emphasize that this approach does not constitute a targeted attack in the conventional sense. Rather, the attempt parameter merely directs the gradient toward regions where transitions away from the original class are most probable, while maintaining an untargeted classification objective. This gradient steering mechanism represents", + "bbox": [ + 511, + 810, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1510", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "a methodological innovation particularly effective in high-dimensional classification spaces where traditional methods might struggle to efficiently locate vulnerable decision boundaries.", + "bbox": [ + 89, + 90, + 480, + 150 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Only if all candidate-based attempts fail does the algorithm revert to the standard untargeted FastJSMA approach without class-specific gradient modification. This hierarchical strategy enhances attack success rates while maintaining the untargeted objective—an engineering optimization that improves practical performance without altering the fundamental attack paradigm.", + "bbox": [ + 89, + 151, + 480, + 257 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5. Complexity Analysis", + "text_level": 1, + "bbox": [ + 89, + 266, + 285, + 282 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To theoretically analyze algorithm efficiency, we compare the computational complexity of FastJSMA with traditional JSMA. Assuming a dataset with $N$ classes, where each image has $M$ pixels. The computational complexity of traditional JSMA is $O(M^2 \\times N)$ . This is because, when constructing the saliency map, it needs to calculate the joint saliency of each pixel pair, considering $M^2$ order of magnitude pixel pairs; the saliency calculation for each pixel pair involves $N$ classes. Moreover, as illustrated in Figure 1, traditional JSMA typically requires multiple iterations of this expensive computation process to achieve successful misclassification, further compounding the computational burden.", + "bbox": [ + 89, + 287, + 482, + 482 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In contrast, our FastJSMA has a computational complexity of $O(M\\sqrt{N})$ . This is because we only perform 2 gradient calculations. The combination of gradient decoupling, class probing mechanism, and significance threshold filtering together dramatically reduces the computational requirements while maintaining attack effectiveness.", + "bbox": [ + 89, + 484, + 482, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This significant reduction in complexity enables FastJSMA to efficiently process large-scale datasets. Due to the dual-gradient approximation strategy retaining JSMA's core ideas, the attack effectiveness is preserved despite substantial complexity reduction, as verified in our experiments.", + "bbox": [ + 89, + 575, + 482, + 665 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our additional optimizations—including gradient normalization with class-specific scaling, dynamic significance thresholds, and the square-root probing strategy—further enhance efficiency without compromising mathematical foundations. These refinements demonstrate the importance of implementation details in translating theoretical innovations into practical adversarial methodologies that scale to contemporary deep learning applications.", + "bbox": [ + 89, + 666, + 482, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 800, + 220, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We present a comprehensive evaluation of FastJSMA, examining its effectiveness, efficiency, and transferability across multiple datasets and model architectures. Our experiments are designed to answer three key questions: (1) How does FastJSMA compare to traditional JSMA and its", + "bbox": [ + 89, + 825, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "variants in terms of attack success and computational efficiency? (2) How do different components of FastJSMA contribute to its overall performance? (3) How sensitive is FastJSMA to key parameter settings?", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 162, + 702, + 179 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.1. Implementation Environment", + "text_level": 1, + "bbox": [ + 511, + 185, + 766, + 200 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "All experiments were conducted on a single NVIDIA RTX 3090 GPU with 24GB memory without virtual memory allocation. We implemented all methods using PyTorch framework and will release our code upon publication to ensure reproducibility.", + "bbox": [ + 511, + 204, + 903, + 280 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.2. Datasets and Models", + "text_level": 1, + "bbox": [ + 511, + 290, + 702, + 304 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate our method across varying scales of complexity, we selected three representative datasets: CIFAR10/100 [28] (natural color images sized $32 \\times 32 \\times 3$ , containing 10 and 100 classes respectively, each with 10,000 test samples) and ImageNet [7] (large-scale classification dataset with 1,000 classes, sized $224 \\times 224 \\times 3$ , with 50,000 validation samples).", + "bbox": [ + 511, + 309, + 903, + 414 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For model architectures, we employed widely-used ResNet variants (primarily ResNet18) [29], using pretrained checkpoints from HuggingFace1. These models achieve competitive performance on their respective datasets, providing a reliable foundation for our adversarial attack evaluation.", + "bbox": [ + 511, + 415, + 903, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.3. Attack Configuration and Baselines", + "text_level": 1, + "bbox": [ + 511, + 515, + 805, + 530 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare FastJSMA with several baseline methods:", + "bbox": [ + 511, + 534, + 875, + 549 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- JSMANT: The traditional Jacobian-based Saliency Map Attack [5] in its non-targeted form, which computes the full Jacobian matrix.", + "- WJSMA: The Weighted JSMA variant proposed by Combey et al. [6], which introduces probabilistic weights to enhance attack effectiveness.", + "- Random: A control baseline that randomly selects pixels for modification." + ], + "bbox": [ + 513, + 553, + 903, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To ensure fair comparison, we maintain consistent attack parameters across all methods. Since different datasets have different image sizes, we use a proportional approach to pixel modification: 8 pixels for $32 \\times 32$ images (CIFAR-10/100) and 196 pixels for $224 \\times 224$ images (ImageNet), representing approximately $0.78\\%$ of the total pixels in each case.", + "bbox": [ + 511, + 678, + 903, + 781 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "All experiments for CIFAR-10/100 were conducted on their complete test sets (10,000 samples each). For ImageNet, we used the full ImageNet1K validation set (50,000 samples across 1,000 classes), where \"1K\" refers to the number of classes rather than sample count. Attack Success Rate (ASR) was calculated as the percentage of sam", + "bbox": [ + 511, + 784, + 903, + 875 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "", + "bbox": [ + 531, + 886, + 799, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1511", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9dada6ac53716799aa6099470b6d5ca3758d3824b23d28d8dc5e81e9a64e2ca7.jpg", + "table_caption": [ + "Table 1. Performance comparison between FastJSMA and JSMA variants with ResNet18" + ], + "table_footnote": [ + "Note: JSMANT and WJSMA could not be tested on ImageNet due to memory constraints" + ], + "table_body": "
Model/DatasetMethodASR(%)Time(μs)L2Norm
CIFAR-10Random7.5531.42
FastJSMA43.983391.53
JSMANT43.7911,7691.57
WJSMA51.568,6411.29
CIFAR-100Random15.3131.44
FastJSMA45.307511.75
JSMANT43.8761,1661.55
WJSMA49.7851,0651.24
ImageNetRandom2.9787.78
FastJSMA54.897,9089.19
", + "bbox": [ + 99, + 128, + 477, + 280 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "plies that were successfully misclassified after perturbation out of all tested samples.", + "bbox": [ + 89, + 319, + 482, + 349 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Notably, since JSMA-based methods operate iteratively (modifying two pixels per iteration), JSMANT and WJSMA use half the number of iterations compared to the total pixel modification count. FastJSMA modifies one pixel per iteration for fair comparison. The attempt parameter controls the number of candidate classes explored during class probing: we use attempt $= 10$ for CIFAR-10, attempt $= \\sqrt{N}$ for larger datasets (e.g., attempt $= 10$ for CIFAR-100's 100 classes, attempt $= 32$ for ImageNet's 1000 classes).", + "bbox": [ + 88, + 349, + 482, + 501 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Performance Analysis", + "text_level": 1, + "bbox": [ + 89, + 511, + 294, + 526 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.1. Comparison with JSMA Variants", + "text_level": 1, + "bbox": [ + 89, + 534, + 370, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1 presents our main experimental results, comparing FastJSMA against JSMANT, WJSMA, and the random baseline across different datasets and model architectures. We report three key metrics: Attack Success Rate (ASR), average generation time per sample, and average perturbation magnitude ( $L_{2}$ norm).", + "bbox": [ + 89, + 551, + 482, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results empirically validate our theoretical complexity reduction from $O(M^2 \\times N)$ to $O(M\\sqrt{N})$ proposed in Section 3.4. FastJSMA achieves attack success rates comparable to or exceeding traditional JSMANT across all datasets while demonstrating remarkable computational efficiency—requiring just $2.9\\%$ of JSMANT's processing time on CIFAR-10 (339 $\\mu$ s vs. 11,769 $\\mu$ s) and merely $1.2\\%$ on CIFAR-100 (751 $\\mu$ s vs. 61,166 $\\mu$ s). These speed-ups of $34.7\\times$ and $81.4\\times$ respectively confirm that the gradient decoupling mechanism theorized in our mathematical formulation successfully transforms the computational scaling properties of saliency-based attacks.", + "bbox": [ + 88, + 643, + 482, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The efficiency advantage increases with dataset complexity, closely matching our theoretical prediction that traditional JSMA's complexity worsens quadratically with increasing image size and class count. Most significantly, FastJSMA successfully executes on ImageNet with a rea", + "bbox": [ + 89, + 825, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/463bf61f22276ff4d7d74253113bd2fdca98876925f71e96ddc03fb4351dcc81.jpg", + "table_caption": [ + "Table 2. Ablation study on gradient components" + ], + "table_footnote": [], + "table_body": "
Gradient ComponentASR(%)Time(μs)L2Norm
FastJSMA (Complete)43.983391.66
G+30.664321.71
G-35.754431.66
", + "bbox": [ + 550, + 114, + 874, + 181 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sonable computation time of $7,908\\mu s$ per sample, while both JSMANT and WJSMA fail due to memory constraints. This empirical finding not only validates our theoretical analysis but extends its implications—demonstrating that our mathematical reformulation enables saliency-based attacks on previously inaccessible large-scale datasets.", + "bbox": [ + 511, + 207, + 903, + 297 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "WJSMA achieves somewhat higher ASRs on CIFAR datasets, which aligns with our theoretical understanding of its specialized optimization mechanisms. However, its computational demands confirm our analysis that approaches relying on the full Jacobian matrix face fundamental limitations for larger-scale applications. These results demonstrate the theoretical-practical trade-off that motivated our work: FastJSMA sacrifices minimal attack effectiveness for transformative computational efficiency gains.", + "bbox": [ + 511, + 297, + 905, + 434 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.2. Ablation Study on Gradient Components", + "text_level": 1, + "bbox": [ + 511, + 443, + 844, + 458 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To investigate the contribution of different gradient components in our dual-gradient approximation strategy, we conducted an ablation study comparing three variants of FastJSMA using the CIFAR-10 dataset with ResNet18, 8 pixels modification, and attempt=10.", + "bbox": [ + 511, + 460, + 905, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "These results validate and refine our theoretical decomposition of the Jacobian matrix described in Section 3.2. The complete FastJSMA incorporating both gradients achieves the highest ASR (43.98%), empirically confirming that our dual-gradient formulation successfully captures the bidirectional influence encoded in traditional JSMA's saliency computation. The class suppression gradient $(g^{-})$ alone yields a $35.75\\%$ ASR, outperforming the class excitation gradient $(g^{+})$ alone (30.66%), revealing a key theoretical insight: reducing the current class probability has a more significant impact than increasing alternative class probabilities.", + "bbox": [ + 511, + 537, + 905, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This empirical finding extends our initial theoretical understanding, revealing an asymmetric contribution pattern not initially anticipated in our mathematical formulation. This discovery explains the effectiveness of our normalization procedure in Section 3.3, which amplifies the class suppression gradient by a factor of $N$ —the ablation empirically validates this engineering optimization while simultaneously deepening our theoretical understanding. The comparable computation times across variants (339-443 $\\mu$ s) further validate our complexity analysis that dual gradient computation maintains $O(M\\sqrt{N})$ efficiency regardless of variant configuration.", + "bbox": [ + 511, + 719, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1512", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d2d6803a5666c1e24d0a1dc0b433e4cc1ce61fe9ec414e897fdc68eb24897430.jpg", + "image_caption": [ + "(a) Fast-JSMA Sensitivity to Max Attempt Fast-JSMA Parameter Sensitivity Analysis (b) Fast-JSMA Sensitivity to Pixels Perturbed" + ], + "image_footnote": [], + "bbox": [ + 94, + 108, + 496, + 262 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/529430c6687121ff635434c1cc1500c63bdaf29e978dbfa5809953de1e1906a8.jpg", + "image_caption": [ + "Figure 2. Parameter sensitivity analysis of FastJSMA on CIFAR-10. (a) Sensitivity to Max Attempt parameter showing ASR and Time curves. (b) Sensitivity to pixels perturbed showing ASR, Time, and Noise curves. All metrics are normalized using min-max scaling for visual comparison; actual ASR percentages shown on right y-axis. Experiments conducted on the complete CIFAR-10 test set (10,000 samples) with ResNet18." + ], + "image_footnote": [], + "bbox": [ + 496, + 109, + 901, + 262 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2.3. Parameter Sensitivity Analysis", + "text_level": 1, + "bbox": [ + 89, + 363, + 351, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To evaluate FastJSMA's robustness to parameter settings, we analyzed two key parameters: pixel modification count and the attempt parameter. Figure 2 presents these analyses conducted on the complete CIFAR-10 test set (10,000 samples).", + "bbox": [ + 89, + 383, + 482, + 458 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our Max Attempt parameter analysis (Figure 2a) shows that attack success rate follows a non-linear pattern: rapid initial growth from attempt $= 1$ to attempt $= 4$ , followed by gradual increase thereafter. This validates our square-root sampling heuristic described in Section 3.4, confirming that smaller attempt values provide higher marginal benefits. For scalability, we empirically determined that $\\sqrt{N}$ sampling works well for larger datasets. Computational cost scales linearly with attempt value, supporting our $O(M\\sqrt{N})$ complexity analysis. The data ranges from $35.2\\%$ ASR at attempt $= 1$ to $44.7\\%$ ASR at attempt $= 10$ , with corresponding time costs increasing from $314\\mu s$ to $417\\mu s$ .", + "bbox": [ + 89, + 460, + 482, + 641 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The pixel modification analysis (Figure 2b) reveals that FastJSMA achieves substantial attack success even with minimal pixel modifications, with performance following an S-curve: initial rapid growth $(2\\rightarrow 16$ pixels), moderate middle phase growth $(16\\rightarrow 36$ pixels), and final diminishing returns $(36\\rightarrow 50$ pixels). This validates our significance threshold mechanism in Section 3.3. Runtime increases only marginally despite significant increases in modified pixels, while L2 norm scales approximately linearly with pixel count. The analysis shows ASR ranging from $26.0\\%$ with 2 pixels to $69.5\\%$ with 50 pixels, with average generation time remaining stable around 0.0003-0.0004 seconds.", + "bbox": [ + 89, + 642, + 482, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "These findings both validate and extend our theoretical model. The consistent computational efficiency across different pixel modification counts demonstrates FastJSMA's scalability, while the sharp initial gains in both analyses highlight the method's effectiveness at identifying influen", + "bbox": [ + 89, + 824, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tial pixels with minimal computational overhead, making it particularly suitable for constrained attack scenarios.", + "bbox": [ + 511, + 363, + 906, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 460, + 632, + 477 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper introduces FastJSMA, a computationally efficient reformulation of saliency-based adversarial attacks. By decomposing the traditional Jacobian calculation into complementary gradient components—class suppression $(g^{-})$ and class excitation $(g^{+})$ —we reduce computational complexity from $O(M^2 \\times N)$ to $O(M\\sqrt{N})$ while maintaining attack effectiveness. Our class probing mechanism and significance threshold further optimize performance by systematically identifying vulnerable decision boundaries and prioritizing influential pixels.", + "bbox": [ + 511, + 503, + 906, + 656 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Experiments validate that FastJSMA achieves dramatic computational efficiency gains—up to $81.4 \\times$ speedup on CIFAR-100 and successful execution on ImageNet where traditional JSMA fails entirely—while maintaining comparable attack success rates. The ablation study reveals the asymmetric contribution of gradient components, providing new insights that extend our theoretical understanding. These findings demonstrate that our mathematical reformulation successfully overcomes the computational barriers that have limited JSMA's applicability in large-scale settings. Beyond immediate technical contributions, FastJSMA's gradient decoupling principles may inform future research in computationally efficient adversarial attacks and explainable AI, where feature importance analysis at scale remains a significant challenge.", + "bbox": [ + 511, + 674, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1513", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 250, + 187, + 266 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 1", + "[2] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 1, 2", + "[3] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. 1, 2", + "[4] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In 2017 IEEE Symposium on Security and Privacy (SP), pages 39-57. IEEE, 2017. 1", + "[5] Nicolas Papernot, Patrick McDaniel, Somesh Jha, Matt Fredrikson, Z. Berkay Celik, and Ananthram Swami. The limitations of deep learning in adversarial settings. In 2016 IEEE European Symposium on Security and Privacy (EuroS&P), pages 372-387. IEEE, 2016. 1, 2, 6", + "[6] Théo Combey, António Loison, Maxime Faucher, and Hatem Hajri. Probabilistic jacobian-based saliency maps attacks. Machine Learning and Knowledge Extraction, 2(4): 558-578, 2020. doi: 10.3390/make2040030. 1, 2, 6", + "[7] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In CVPR09, 2009. 1, 6", + "[8] Nicolas Papernot, Patrick McDaniel, and Ian Goodfellow. Transferability in machine learning: from phenomena to black-box attacks using adversarial samples. arXiv preprint arXiv:1605.07277, 2016. 2", + "[9] Pin-Yu Chen, Huan Zhang, Yash Sharma, Jinfeng Yi, and Cho-Jui Hsieh. Zoo: Zeroth order optimization based blackbox attacks to deep neural networks without training substitute models. In Proceedings of the 10th ACM Workshop on Artificial Intelligence and Security, pages 15-26. ACM, 2017. 2", + "[10] Dawei Zhou, Nannan Wang, Tongliang Liu, and Xinbo Gao. Improving adversarial training from the perspective of class-flipping distribution. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 2", + "[11] Dawei Zhou, Nannan Wang, Bo Han, Tongliang Liu, and Xinbo Gao. Defending against adversarial examples via modeling adversarial noise. International Journal of Computer Vision, pages 1-18, 2025." + ], + "bbox": [ + 93, + 276, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "We would like to thank the HPC Platform of Huazhong University of Science and Technology for providing computational resources. We also express our gratitude to the School of Artificial Intelligence and Automation, the School of Software Engineering and the School of Electronic Information and Communications at Huazhong University of Science and Technology for their strong support throughout this research.", + "[12] Chaojian Yu, Bo Han, Li Shen, Jun Yu, Chen Gong, Mingming Gong, and Tongliang Liu. Understanding robust overfitting of adversarial training and beyond. In International Conference on Machine Learning, pages 25595-25610. PMLR, 2022. 2", + "[13] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and accurate method to fool deep neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2574–2582, 2016. 2", + "[14] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9185-9193, 2018. 2", + "[15] Jiadong Lin, Chuanbiao Song, Kun He, Liwei Wang, and John E. Hopcroft. Nesterov accelerated gradient and scale invariance for adversarial attacks. In International Conference on Learning Representations, 2020. 2", + "[16] Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Evading defenses to transferable adversarial examples by translation-invariant attacks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4312-4321, 2019. 2", + "[17] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L. Yuille. Improving transferability of adversarial examples with input diversity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2730-2739, 2019. 2", + "[18] Chengzhi Mao, Mia Chiquier, Hao Wang, Junfeng Yang, and Carl Vondrick. Adversarial attacks are reversible with natural supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 661-671, October 2021. 2", + "[19] Jianming Zhang and Stan Sclaroff. Saliency detection: A boolean map approach. In 2013 IEEE International Conference on Computer Vision, pages 153-160, 2013. doi: 10.1109/ICCV.2013.26.2", + "[20] Jiawei Su, Danilo Vasconcellos Vargas, and Kouichi Sakurai. One pixel attack for fooling deep neural networks. IEEE Transactions on Evolutionary Computation, 23(5):828-841, 2019. 2, 3", + "[21] Jianbo Chen, Michael I. Jordan, and Martin J. Wainwright. Hopskipjumpattack: A query-efficient decision-based attack. In 2020 IEEE Symposium on Security and Privacy (SP), pages 1277-1294. IEEE, 2020. 2, 3", + "[22] Tianhang Zheng, Changyou Chen, Junsong Yuan, Bo Li, and Kui Ren. Pointcloud saliency maps. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019. 2", + "[23] Apostolos Modas, Seyed-Mohsen Moosavi-Dezfooli, and Pascal Frossard. Sparsefool: a few pixels make a big difference. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9087–9096, 2019. 3", + "[24] Florian Tramér, Alexey Kurakin, Nicolas Papernot, Ian Goodfellow, Dan Boneh, and Patrick McDaniel. Ensemble" + ], + "bbox": [ + 93, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1514", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "adversarial training: Attacks and defenses. In International Conference on Learning Representations, 2018. 3", + "[25] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International Conference on Machine Learning, pages 2206-2216. PMLR, 2020. 3", + "[26] Maksym Andriushchenko, Francesco Croce, Nicolas Flammarion, and Matthias Hein. Square attack: a query-efficient black-box adversarial attack via random search. In European Conference on Computer Vision, pages 484-501. Springer, 2020. 3", + "[27] Maura Pintor, Fabio Roli, Wieland Brendel, and Battista Biggio. Fast minimum-norm adversarial attacks through adaptive norm constraints. arXiv preprint arXiv:2102.12827, 2021.3", + "[28] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. Technical Report 4, University of Toronto, 2009. 6", + "[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + ], + "bbox": [ + 91, + 90, + 483, + 402 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1515", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/FastJSMA_ Accelerating Jacobian-based Saliency Map Attacks through Gradient Decoupling/caddbdfc-9052-49a6-8b20-7e4e162ec5a3_model.json b/2025/FastJSMA_ Accelerating Jacobian-based Saliency Map Attacks through Gradient Decoupling/caddbdfc-9052-49a6-8b20-7e4e162ec5a3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8e936434421738b4c2969aea2c49b1e2a2cc4884 --- /dev/null +++ b/2025/FastJSMA_ Accelerating Jacobian-based Saliency Map Attacks through Gradient Decoupling/caddbdfc-9052-49a6-8b20-7e4e162ec5a3_model.json @@ -0,0 +1,2200 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.139, + 0.13, + 0.861, + 0.178 + ], + "angle": 0, + "content": "FastJSMA: Accelerating Jacobian-based Saliency Map Attacks through Gradient Decoupling" + }, + { + "type": "text", + "bbox": [ + 0.154, + 0.205, + 0.842, + 0.221 + ], + "angle": 0, + "content": "Zhenghao \\(\\mathrm{Gao}^{1,*}\\) Shengjie \\(\\mathrm{Xu}^{2,*}\\) Zijing Li\\(^{2}\\) Meixi Chen\\(^{3}\\) Chaojian \\(\\mathrm{Yu}^{4}\\) Yuanjie \\(\\mathrm{Shao}^{4}\\) Changxin \\(\\mathrm{Gao}^{1,\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.173, + 0.223, + 0.827, + 0.239 + ], + "angle": 0, + "content": "1School of Artificial Intelligence and Automation, Huazhong University of Science and Technology" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.241, + 0.746, + 0.257 + ], + "angle": 0, + "content": "\\(^{2}\\)School of Software Engineering, Huazhong University of Science and Technology" + }, + { + "type": "text", + "bbox": [ + 0.192, + 0.258, + 0.805, + 0.275 + ], + "angle": 0, + "content": "3Journalism and Information Communication School, Huazhong University of Science and Technology" + }, + { + "type": "text", + "bbox": [ + 0.185, + 0.277, + 0.813, + 0.292 + ], + "angle": 0, + "content": "\\(^{4}\\)School of Electronic Information and Communications, Huazhong University of Science and Technology" + }, + { + "type": "text", + "bbox": [ + 0.204, + 0.297, + 0.789, + 0.31 + ], + "angle": 0, + "content": "{u202215226,u202317280,u202317273,u202317218,yuchaojian,shaoyuanjie,cgao}@hust.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.345, + 0.327, + 0.361 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.377, + 0.486, + 0.756 + ], + "angle": 0, + "content": "Adversarial attack plays a critical role in evaluating the robustness of deep learning models. Jacobian-based Saliency Map Attack (JSMA) is an interpretable adversarial method that offers excellent pixel-level control and provides valuable insights into model vulnerabilities. However, its quadratic computational complexity \\( O(M^2 \\times N) \\) renders it impractical for large-scale datasets, limiting its application despite its inherent value. This paper proposes FastJSMA, an efficient attack method that addresses these computational limitations. Our approach introduces a gradient decoupling mechanism that decomposes the Jacobian calculation into complementary class suppression (\\( g^{-} \\)) and class excitation (\\( g^{+} \\)) gradients, reducing complexity to \\( O(M\\sqrt{N}) \\). Additionally, we implement a class probing mechanism and an adaptive saliency threshold to further optimize the process. Experimental results across multiple datasets demonstrate that FastJSMA maintains comparable attack success rates while dramatically reducing computation time—requiring only 2.9% of JSMA's processing time on CIFAR-10 and 1.2% on CIFAR-100, and successfully operating on ImageNet where traditional JSMA fails due to memory constraints. This advancement enables the practical application of interpretable saliency map-based attacks on large-scale datasets, balancing effectiveness with computational efficiency." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.785, + 0.222, + 0.8 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.81, + 0.296, + 0.827 + ], + "angle": 0, + "content": "1.1. Research Background" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.833, + 0.484, + 0.864 + ], + "angle": 0, + "content": "Deep learning has achieved breakthrough progress in computer vision, natural language processing, and other fields," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.346, + 0.907, + 0.452 + ], + "angle": 0, + "content": "but its security issues have increasingly attracted attention. In 2013, Szegedy et al. [1] first discovered the sensitivity of deep neural networks to small perturbations, i.e., by adding imperceptible perturbations to inputs, models can be induced to produce incorrect predictions. This discovery revealed potential vulnerabilities in deep learning systems and drove rapid development in adversarial attack research." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.458, + 0.909, + 0.535 + ], + "angle": 0, + "content": "Adversarial attack techniques have evolved from simple approaches like the Fast Gradient Sign Method (FGSM) [2] to complex methods such as PGD [3] and C&W attacks [4], each offering different trade-offs between effectiveness, efficiency, and perturbation magnitude." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.563, + 0.71, + 0.579 + ], + "angle": 0, + "content": "1.2. Research Motivation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.592, + 0.909, + 0.79 + ], + "angle": 0, + "content": "The Jacobian-based Saliency Map Attack (JSMA) [5] is an adversarial method that identifies and modifies the most influential pixels through Jacobian matrix analysis. It offers unique advantages including intuitive visualization through saliency maps that enhance interpretability, precise control over modified pixels, and strong effectiveness across various datasets. Despite these merits and attempts to improve it through weighted approaches like WJSMA [6], JSMA remains largely underutilized due to prohibitive computational costs. Its full Jacobian matrix calculation requires \\( O(M^2 \\times N) \\) complexity (where \\( M \\) is the number of pixels and \\( N \\) is the number of classes), creating an insurmountable barrier for large-scale applications like ImageNet [7]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "This limitation stems from three factors: (a) separate backward propagation for each output node; (b) saliency evaluations across the entire Jacobian matrix; and (c) repeated calculations across iterations. Consequently, the adversarial machine learning community has effectively abandoned JSMA's valuable properties due to computational constraints rather than methodological limitations." + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.224, + 0.888 + ], + "angle": 0, + "content": "0* Equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.24, + 0.9 + ], + "angle": 0, + "content": "\\(^{0\\dagger}\\) Corresponding author" + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.24, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1506" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.236, + 0.105 + ], + "angle": 0, + "content": "1.3. Contributions" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.483, + 0.174 + ], + "angle": 0, + "content": "To revitalize this valuable but computationally constrained method, we propose FastJSMA (Fast Jacobian-based Saliency Map Attack), an efficient adversarial attack based on gradient decoupling approximation:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.176, + 0.483, + 0.265 + ], + "angle": 0, + "content": "- Gradient Decoupling Mechanism: We decompose the Jacobian matrix into complementary class suppression \\((g^{-})\\) and excitation \\((g^{+})\\) gradients, reducing computational complexity from \\(O(M^2 \\times N)\\) to \\(O(M\\sqrt{N})\\) while maintaining attack effectiveness through dynamic weight allocation." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.267, + 0.484, + 0.342 + ], + "angle": 0, + "content": "- Class Probing and Adaptive Threshold: We introduce a square-root sampling heuristic to efficiently identify vulnerable decision boundaries and an adaptive threshold mechanism to prioritize influential pixels, optimizing both efficiency and effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.343, + 0.484, + 0.464 + ], + "angle": 0, + "content": "- Comprehensive Evaluation: Our experiments demonstrate dramatic efficiency gains—requiring only \\(2.9\\%\\) of JSMANT's processing time on CIFAR-10 and achieving up to \\(81.4 \\times\\) speedup on CIFAR-100 while maintaining comparable attack success rates. Most notably, FastJSMA successfully processes ImageNet samples, where traditional JSMA and WJSMA fail entirely due to memory constraints." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.176, + 0.484, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.477, + 0.234, + 0.492 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.502, + 0.443, + 0.517 + ], + "angle": 0, + "content": "2.1. Overview of Adversarial Attack Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.524, + 0.483, + 0.659 + ], + "angle": 0, + "content": "Adversarial attacks can be categorized as white-box or black-box attacks based on the attacker's knowledge of the model. White-box attacks assume the attacker has complete access to the model structure and parameters, with representative methods including FGSM [2] and PGD [3]. Black-box attacks can only access the model's input-output interface, primarily relying on transferability [8] or query optimization [9]. Recent work has also focused on improving adversarial training defenses [10-12]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.81 + ], + "angle": 0, + "content": "From the perspective of attack objectives, they can be divided into targeted and non-targeted attacks. Targeted attacks aim to misclassify input samples as specified target classes, such as DeepFool [13]; non-targeted attacks only need to cause incorrect predictions, regardless of the specific error class, such as FGSM. These two types of attack methods have their respective advantages in practical applications: targeted attacks enable more precise control, while non-targeted attacks typically have higher success rates and lower computational costs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Gradient-based adversarial attacks have evolved significantly. After FGSM was proposed, researchers developed improved variants: (1) MI-FGSM [14] introduces momentum terms; (2) NI-FGSM [15] utilizes Nesterov accelerated gradients; (3) TI-FGSM [16] enhances transferability through translation invariance; (4) DI-FGSM [17] adopts" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.198 + ], + "angle": 0, + "content": "diversified input strategies. Recent research by Mao et al. [18] discovered that adversarial attacks are reversible with natural supervision, providing new insights into the fundamental properties of adversarial perturbations. These methods optimize gradient-based attack strategies from different perspectives but have not fully addressed the issue of interpretability." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.206, + 0.757, + 0.221 + ], + "angle": 0, + "content": "2.2. Detailed Analysis of JSMA" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.228, + 0.907, + 0.363 + ], + "angle": 0, + "content": "As an interpretable attack method, JSMA was proposed by Papernot et al. [5], with its core idea being the use of the Jacobian matrix to identify input features that have the greatest impact on model output. This approach builds upon fundamental research in saliency detection, such as the boolean map approach [19], which provides theoretical foundations for identifying visually significant regions. Specifically, for input \\( x \\) and target class \\( t \\), JSMA first calculates the Jacobian matrix:" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.373, + 0.906, + 0.41 + ], + "angle": 0, + "content": "\\[\nJ _ {F} (x) = \\frac {\\partial F (x)}{\\partial x} = \\left[ \\frac {\\partial F _ {j} (x)}{\\partial x _ {i}} \\right] _ {i, j} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.414, + 0.906, + 0.46 + ], + "angle": 0, + "content": "where \\( F_{j}(x) \\) represents the model's prediction probability for class \\( j \\). Based on the Jacobian matrix, JSMA defines the saliency map \\( S \\):" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.466, + 0.905, + 0.524 + ], + "angle": 0, + "content": "\\[\nS (x, t) [ i ] = \\left\\{ \\begin{array}{l} 0, \\text {i f} \\frac {\\partial F _ {t} (x)}{\\partial x _ {i}} < 0 \\text {o r} \\sum_ {j \\neq t} \\frac {\\partial F _ {j} (x)}{\\partial x _ {i}} > 0 \\\\ \\left| \\frac {\\partial F _ {t} (x)}{\\partial x _ {i}} \\right| \\cdot \\left| \\sum_ {j \\neq t} \\frac {\\partial F _ {j} (x)}{\\partial x _ {i}} \\right|, \\text {o t h e r w i s e} \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.906, + 0.674 + ], + "angle": 0, + "content": "Subsequently, researchers have made various improvements to JSMA: (1) JSMA++ [8] enhances attack efficiency by optimizing saliency calculations; (2) One-pixel JSMA [20] restricts modifications to a single pixel, achieving more precise control; (3) Adaptive JSMA [21] introduces adaptive threshold strategies, improving attack success rates. The concept of saliency maps has also been extended to other domains, such as 3D point clouds [22], demonstrating the versatility of saliency-based approaches across different data modalities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.902 + ], + "angle": 0, + "content": "An important advancement is the Weighted JSMA (WJSMA) [6], which addresses limitations in the original JSMA's saliency calculation. WJSMA penalizes gradients associated with small probabilities to create more balanced saliency maps. WJSMA incorporates class probabilities as weights in the saliency calculation, setting the saliency to zero when specific gradient conditions are not met. This weighting mechanism makes WJSMA faster than the original JSMA while maintaining good balance between attack success rate and computational cost. However, even with these improvements, WJSMA still inherits the fundamental computational inefficiency of JSMA for large-scale datasets, as it retains the \\(O(M^2 \\times N)\\) complexity that creates memory and processing bottlenecks on datasets like ImageNet." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1507" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.383, + 0.108 + ], + "angle": 0, + "content": "2.3. Related Optimization Techniques" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.484, + 0.221 + ], + "angle": 0, + "content": "To address the efficiency problems of adversarial attacks in large-scale scenarios, researchers have proposed various optimization strategies. One class of methods improves efficiency by enhancing gradient computation, such as OnePixel Attack [20], which reduces search space through evolutionary algorithms, and SparseFool [23], which optimizes computational efficiency by leveraging sparsity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.222, + 0.484, + 0.298 + ], + "angle": 0, + "content": "Another class of methods proposes specialized optimization strategies for large-scale datasets. For example, Fast Gradient Method [24] enhances attack efficiency through ensemble techniques, while AutoAttack [25] adopts automated strategies to select optimal attack parameters." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.299, + 0.484, + 0.405 + ], + "angle": 0, + "content": "Recent research has also explored algorithmic optimizations specifically targeting the computational complexity of gradient-based attacks. These approaches include gradient approximation techniques [26], adaptive norm constraints for efficient minimum-norm attacks [27], and decomposition strategies that reduce the dimensionality of the optimization space [21]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.406, + 0.484, + 0.587 + ], + "angle": 0, + "content": "These optimization methods have their respective advantages and disadvantages: gradient-based methods offer high computational efficiency but lower interpretability; optimization-based methods provide good results but incur high computational costs; heuristic-based methods are simple and fast but have unstable success rates. In comparison, our proposed FastJSMA maintains JSMA's interpretability while significantly improving computational efficiency through dual-gradient approximation, effectively reducing the complexity from \\( O(M^2 \\times N) \\) to \\( O(M\\sqrt{N}) \\) and making saliency map-based attacks practical for large-scale datasets for the first time." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.605, + 0.226, + 0.623 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.631, + 0.484, + 0.738 + ], + "angle": 0, + "content": "This section details the technical principles of FastJSMA. We first establish the mathematical model of the problem, then elaborate on the design rationale of the dual-gradient approximation strategy, followed by a detailed explanation of the saliency map generation process, and finally theoretically analyze the computational complexity of the algorithm." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.751, + 0.446, + 0.768 + ], + "angle": 0, + "content": "3.1. Problem Modeling and Attack Objectives" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.773, + 0.484, + 0.866 + ], + "angle": 0, + "content": "For an input sample \\( x \\in \\mathbb{R}^{C \\times H \\times W} \\) and a target classification model \\( f: \\mathbb{R}^{C \\times H \\times W} \\to \\mathbb{R}^N \\), where \\( C, H, W \\) represent the number of channels, height, and width respectively, and \\( N \\) is the total number of classes. Given the original predicted class \\( t = \\arg \\max f(x) \\), the goal of the adversarial attack is to find a perturbation \\( \\delta \\) such that:" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.885, + 0.484, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\arg \\max f (x + \\delta) \\neq t \\quad \\text {s . t .} \\quad \\| \\delta \\| _ {0} \\leq \\epsilon \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.228 + ], + "angle": 0, + "content": "where \\(\\epsilon\\) represents the allowed number of pixels to modify, and \\(\\| \\cdot \\|_0\\) is the \\(L_0\\) norm. The reasons for choosing the \\(L_0\\) norm constraint rather than other norms (such as \\(L_2\\) or \\(\\mathrm{L\\infty}\\)) are: (1) \\(L_0\\) norm directly limits the number of modified pixels, conforming to the principle of minimal intervention; (2) a limited number of modifications makes the attack process more interpretable and understandable; (3) maintains consistent constraint forms with traditional JSMA for fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.229, + 0.905, + 0.333 + ], + "angle": 0, + "content": "Traditional JSMA computes the complete Jacobian matrix \\( J = \\nabla_{x}f(x) \\) with \\( O(M^2\\times N) \\) complexity, where \\( M \\) represents the total number of pixels and \\( N \\) denotes the number of classification classes. This substantial computational burden and memory requirements render JSMA impractical for large-scale datasets and real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.336, + 0.905, + 0.381 + ], + "angle": 0, + "content": "FastJSMA addresses this limitation through a dual-gradient approximation strategy that reduces complexity to \\( O(M\\sqrt{N}) \\) while maintaining attack effectiveness." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.395, + 0.779, + 0.41 + ], + "angle": 0, + "content": "3.2. Target Function Construction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.417, + 0.905, + 0.538 + ], + "angle": 0, + "content": "The core mathematical insight that enables FastJSMA's efficiency breakthrough lies in our recognition that the Jacobian-based saliency computation in traditional JSMA can be decomposed into two complementary objectives: first, decreasing the probability of the current class, and second, increasing the probability of alternative classes. This decomposition forms the theoretical foundation of our dual-gradient approximation strategy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.54, + 0.906, + 0.691 + ], + "angle": 0, + "content": "Analyzing the mathematical structure of traditional JSMA reveals that its saliency map computation essentially measures two gradient components for each pixel: how modifying that pixel decreases the current class score and how it increases other class scores. While traditional JSMA explicitly computes these gradients through the full Jacobian matrix, we observed that these two objectives could be approximated through carefully designed loss functions without requiring the prohibitively expensive Jacobian computation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.692, + 0.906, + 0.783 + ], + "angle": 0, + "content": "To avoid computing the complete Jacobian matrix, we propose a dual-gradient approximation strategy. The core idea of this strategy is to construct two complementary objective functions that simulate the enhancement effect on the target class and the suppression effect on other classes in JSMA. The specific definitions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.788, + 0.906, + 0.818 + ], + "angle": 0, + "content": "1. Class Suppression Gradient: Construct a virtual label \\(\\hat{y}^{-} \\in \\mathbb{R}^{N}\\), where" + }, + { + "type": "equation", + "bbox": [ + 0.533, + 0.832, + 0.912, + 0.901 + ], + "angle": 0, + "content": "\\[\n\\hat {y} _ {i} ^ {-} = \\left\\{ \\begin{array}{l l} f (x) _ {i} - 1 & i = t \\\\ f (x) _ {i} + \\sqrt {N} & i = a (\\text {a t t e m p t c l a s s , i f s p e c i f i e d}) \\\\ f (x) _ {i} & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1508" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.131, + 0.089, + 0.874, + 0.566 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.575, + 0.907, + 0.675 + ], + "angle": 0, + "content": "Figure 1. Architectural comparison between traditional JSMA and our proposed FastJSMA. Left: Traditional JSMA calculates the full Jacobian matrix with computational complexity \\( O(M^2 \\times N) \\) and relies on an iterative process to achieve misclassification, making it prohibitively expensive for large-scale datasets. Right: FastJSMA employs a class probing mechanism to select \\( \\sqrt{N} \\) candidate classes and a gradient decoupling strategy that approximates the saliency computation through class suppression gradient \\( (g^-) \\) and class excitation gradient \\( (g^+) \\). The class suppression gradient reduces the current class probability while boosting an attempt class by \\( \\sqrt{N} \\), and the significance threshold efficiently filters the most influential pixels. This approach reduces complexity to \\( O(M\\sqrt{N}) \\) while maintaining comparable effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.7, + 0.265, + 0.715 + ], + "angle": 0, + "content": "Calculate the gradient:" + }, + { + "type": "equation", + "bbox": [ + 0.221, + 0.729, + 0.483, + 0.746 + ], + "angle": 0, + "content": "\\[\ng ^ {-} = \\nabla_ {x} \\mathcal {L} (f (x), \\hat {y} ^ {-}) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.76, + 0.483, + 0.82 + ], + "angle": 0, + "content": "This gradient reflects the direction of reducing the prediction probability of the current class and optionally enhancing a specific alternative class if an attempt parameter is provided." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.821, + 0.483, + 0.851 + ], + "angle": 0, + "content": "2. Class Excitation Gradient: Construct an enhancement label \\(\\hat{y}^{+}\\in \\mathbb{R}^{N}\\), where" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.864, + 0.483, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\hat {y} _ {i} ^ {+} = \\left\\{ \\begin{array}{l l} f (x) _ {i} - 1 & i = t \\\\ f (x) _ {i} + 1 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.7, + 0.688, + 0.715 + ], + "angle": 0, + "content": "Calculate the gradient:" + }, + { + "type": "equation", + "bbox": [ + 0.644, + 0.723, + 0.905, + 0.74 + ], + "angle": 0, + "content": "\\[\ng ^ {+} = \\nabla_ {x} \\mathcal {L} (f (x), \\hat {y} ^ {+}) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.748, + 0.907, + 0.793 + ], + "angle": 0, + "content": "This gradient indicates the direction of enhancing the prediction probability of other classes while suppressing the current class." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.907, + 0.903 + ], + "angle": 0, + "content": "This formulation represents a significant mathematical innovation over traditional JSMA. Rather than explicitly computing \\( N \\times M \\) partial derivatives as required by the Jacobian matrix, our approach leverages automatic differentiation to compute just two gradient vectors—reducing the computational complexity by a factor proportional to the number of classes \\( N \\) and the number of pixels \\( M \\). Import-" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1509" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "tantly, this efficiency gain comes without compromising the underlying attack methodology; the dual-gradient approach preserves the core mathematical principles of JSMA while dramatically reducing its computational footprint." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.152, + 0.483, + 0.226 + ], + "angle": 0, + "content": "The design of these two gradients cleverly avoids directly computing the Jacobian matrix while retaining the core idea of JSMA: simultaneously considering the suppression of the target class and the excitation of other classes." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.237, + 0.325, + 0.254 + ], + "angle": 0, + "content": "3.3. Saliency Map Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.259, + 0.483, + 0.364 + ], + "angle": 0, + "content": "Traditional JSMA generates saliency maps through exhaustive computation of the Jacobian matrix followed by pairwise evaluation of pixel modifications—a process that scales quadratically with the number of pixels. FastJSMA fundamentally reconceptualizes this approach through an elegant mathematical transformation that yields equivalent results with substantially reduced computation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.365, + 0.483, + 0.41 + ], + "angle": 0, + "content": "Based on the dual-gradient approximation, we apply gradient normalization to ensure balanced contribution while strategically emphasizing the class suppression component:" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.422, + 0.483, + 0.455 + ], + "angle": 0, + "content": "\\[\ng _ {\\text {n o r m a l i z e d}} ^ {-} = \\frac {g ^ {-}}{\\operatorname {m e a n} (| g ^ {-} |)} \\times N \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.203, + 0.464, + 0.482, + 0.497 + ], + "angle": 0, + "content": "\\[\ng _ {\\text {n o r m a l i z e d}} ^ {+} = \\frac {g ^ {+}}{\\operatorname {m e a n} (| g ^ {+} |)} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.501, + 0.483, + 0.606 + ], + "angle": 0, + "content": "This normalization procedure includes a deliberate engineering optimization: the class suppression gradient is scaled by a factor of \\( N \\) (the number of classes) to amplify its influence in the overall saliency calculation. Our empirical investigation indicates that this scaling enhances attack effectiveness by emphasizing the suppression of the original class prediction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.607, + 0.483, + 0.637 + ], + "angle": 0, + "content": "We then integrate these normalized gradients to obtain a comprehensive gain matrix:" + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.651, + 0.482, + 0.669 + ], + "angle": 0, + "content": "\\[\nG = g _ {\\text {n o r m a l i z e d}} ^ {+} + g _ {\\text {n o r m a l i z e d}} ^ {-} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.779 + ], + "angle": 0, + "content": "This matrix captures the bidirectional influence of modifying each pixel on both decreasing the target class probability and increasing alternative class probabilities—precisely the information encapsulated in traditional JSMA's saliency computation, but obtained through a mathematically equivalent yet computationally superior formulation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.901 + ], + "angle": 0, + "content": "After obtaining the gain matrix, we apply a significance threshold to filter out pixels with minimal impact. This threshold operation is visually represented as a funnel-like process in Figure 1, focusing computational resources only on the most influential pixels. By implementing this significance threshold, we effectively reduce the search space while ensuring that the selected pixels have meaningful contributions to the adversarial perturbation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "This selective modification approach, governed by both the saliency ranking and the significance threshold, ensures efficient resource allocation by concentrating perturbations on the most influential and impactful pixels." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.161, + 0.905, + 0.192 + ], + "angle": 0, + "content": "3.4. Attack Success Rate Optimization via Class Probing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.198, + 0.905, + 0.348 + ], + "angle": 0, + "content": "A significant enhancement in our implementation is the introduction of a class-specific probing mechanism, visually illustrated in Figure 1 as \"Class Probing Mechanism\". This component substantially improves attack success rates while maintaining the untargeted nature of FastJSMA by intelligently selecting candidate classes for the attempt parameter. The optimization addresses a fundamental challenge in untargeted adversarial attacks: determining which gradient directions most efficiently induce misclassification." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.349, + 0.905, + 0.395 + ], + "angle": 0, + "content": "Our implementation employs an intelligent candidate class exploration strategy governed by a square-root sampling heuristic:" + }, + { + "type": "equation", + "bbox": [ + 0.63, + 0.408, + 0.905, + 0.426 + ], + "angle": 0, + "content": "\\[\n\\text {n u m} _ {\\text {c a n d i d a t e s}} = \\sqrt {N} \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.432, + 0.905, + 0.522 + ], + "angle": 0, + "content": "where \\(N\\) represents the number of classes in the classification problem. This square-root relationship establishes a principled balance between computational efficiency and exploration thoroughness—scaling sublinearly with the class space dimensionality to maintain tractability even for large-scale classification problems." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.523, + 0.905, + 0.567 + ], + "angle": 0, + "content": "For each candidate class \\(c\\) among the top numCandidates predicted classes (excluding the original class), the algorithm:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.568, + 0.905, + 0.659 + ], + "angle": 0, + "content": "1. Applies FastJSMA with class \\( c \\) as an \"attempt\" parameter, which specifically modifies the class suppression gradient \\( (g^{-}) \\) in two ways: - Decreases the current class probability (by setting \\( \\hat{y}_i^- = f(x)_i - 1 \\) for \\( i = t \\)) - Boosts the attempt class \\( c \\) by a factor of \\( \\sqrt{N} \\) (by setting \\( \\hat{y}_i^- = f(x)_i + \\sqrt{N} \\) for \\( i = c \\))" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.66, + 0.904, + 0.689 + ], + "angle": 0, + "content": "2. Evaluates whether this modified gradient direction successfully induces misclassification" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.69, + 0.905, + 0.719 + ], + "angle": 0, + "content": "3. Returns the first successful perturbation, terminating further exploration" + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.568, + 0.905, + 0.719 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.905, + 0.809 + ], + "angle": 0, + "content": "This dual action of decreasing the current class while simultaneously boosting a specific attempt class efficiently guides the gradient toward vulnerable regions of the decision boundary, substantially improving attack success rates without requiring exhaustive exploration of all possible class transitions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.905, + 0.901 + ], + "angle": 0, + "content": "It is crucial to emphasize that this approach does not constitute a targeted attack in the conventional sense. Rather, the attempt parameter merely directs the gradient toward regions where transitions away from the original class are most probable, while maintaining an untargeted classification objective. This gradient steering mechanism represents" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1510" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.151 + ], + "angle": 0, + "content": "a methodological innovation particularly effective in high-dimensional classification spaces where traditional methods might struggle to efficiently locate vulnerable decision boundaries." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.152, + 0.482, + 0.258 + ], + "angle": 0, + "content": "Only if all candidate-based attempts fail does the algorithm revert to the standard untargeted FastJSMA approach without class-specific gradient modification. This hierarchical strategy enhances attack success rates while maintaining the untargeted objective—an engineering optimization that improves practical performance without altering the fundamental attack paradigm." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.267, + 0.286, + 0.283 + ], + "angle": 0, + "content": "3.5. Complexity Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.289, + 0.483, + 0.483 + ], + "angle": 0, + "content": "To theoretically analyze algorithm efficiency, we compare the computational complexity of FastJSMA with traditional JSMA. Assuming a dataset with \\(N\\) classes, where each image has \\(M\\) pixels. The computational complexity of traditional JSMA is \\(O(M^2 \\times N)\\). This is because, when constructing the saliency map, it needs to calculate the joint saliency of each pixel pair, considering \\(M^2\\) order of magnitude pixel pairs; the saliency calculation for each pixel pair involves \\(N\\) classes. Moreover, as illustrated in Figure 1, traditional JSMA typically requires multiple iterations of this expensive computation process to achieve successful misclassification, further compounding the computational burden." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.485, + 0.483, + 0.576 + ], + "angle": 0, + "content": "In contrast, our FastJSMA has a computational complexity of \\( O(M\\sqrt{N}) \\). This is because we only perform 2 gradient calculations. The combination of gradient decoupling, class probing mechanism, and significance threshold filtering together dramatically reduces the computational requirements while maintaining attack effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.576, + 0.483, + 0.666 + ], + "angle": 0, + "content": "This significant reduction in complexity enables FastJSMA to efficiently process large-scale datasets. Due to the dual-gradient approximation strategy retaining JSMA's core ideas, the attack effectiveness is preserved despite substantial complexity reduction, as verified in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.667, + 0.483, + 0.788 + ], + "angle": 0, + "content": "Our additional optimizations—including gradient normalization with class-specific scaling, dynamic significance thresholds, and the square-root probing strategy—further enhance efficiency without compromising mathematical foundations. These refinements demonstrate the importance of implementation details in translating theoretical innovations into practical adversarial methodologies that scale to contemporary deep learning applications." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.801, + 0.222, + 0.817 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We present a comprehensive evaluation of FastJSMA, examining its effectiveness, efficiency, and transferability across multiple datasets and model architectures. Our experiments are designed to answer three key questions: (1) How does FastJSMA compare to traditional JSMA and its" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "variants in terms of attack success and computational efficiency? (2) How do different components of FastJSMA contribute to its overall performance? (3) How sensitive is FastJSMA to key parameter settings?" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.164, + 0.704, + 0.18 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.186, + 0.767, + 0.201 + ], + "angle": 0, + "content": "4.1.1. Implementation Environment" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.205, + 0.905, + 0.281 + ], + "angle": 0, + "content": "All experiments were conducted on a single NVIDIA RTX 3090 GPU with 24GB memory without virtual memory allocation. We implemented all methods using PyTorch framework and will release our code upon publication to ensure reproducibility." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.291, + 0.704, + 0.305 + ], + "angle": 0, + "content": "4.1.2. Datasets and Models" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.31, + 0.905, + 0.415 + ], + "angle": 0, + "content": "To evaluate our method across varying scales of complexity, we selected three representative datasets: CIFAR10/100 [28] (natural color images sized \\(32 \\times 32 \\times 3\\), containing 10 and 100 classes respectively, each with 10,000 test samples) and ImageNet [7] (large-scale classification dataset with 1,000 classes, sized \\(224 \\times 224 \\times 3\\), with 50,000 validation samples)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.416, + 0.905, + 0.505 + ], + "angle": 0, + "content": "For model architectures, we employed widely-used ResNet variants (primarily ResNet18) [29], using pretrained checkpoints from HuggingFace1. These models achieve competitive performance on their respective datasets, providing a reliable foundation for our adversarial attack evaluation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.516, + 0.807, + 0.531 + ], + "angle": 0, + "content": "4.1.3. Attack Configuration and Baselines" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.535, + 0.876, + 0.55 + ], + "angle": 0, + "content": "We compare FastJSMA with several baseline methods:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.554, + 0.905, + 0.598 + ], + "angle": 0, + "content": "- JSMANT: The traditional Jacobian-based Saliency Map Attack [5] in its non-targeted form, which computes the full Jacobian matrix." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.6, + 0.905, + 0.643 + ], + "angle": 0, + "content": "- WJSMA: The Weighted JSMA variant proposed by Combey et al. [6], which introduces probabilistic weights to enhance attack effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.645, + 0.905, + 0.674 + ], + "angle": 0, + "content": "- Random: A control baseline that randomly selects pixels for modification." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.554, + 0.905, + 0.674 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.679, + 0.905, + 0.782 + ], + "angle": 0, + "content": "To ensure fair comparison, we maintain consistent attack parameters across all methods. Since different datasets have different image sizes, we use a proportional approach to pixel modification: 8 pixels for \\(32 \\times 32\\) images (CIFAR-10/100) and 196 pixels for \\(224 \\times 224\\) images (ImageNet), representing approximately \\(0.78\\%\\) of the total pixels in each case." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.785, + 0.905, + 0.876 + ], + "angle": 0, + "content": "All experiments for CIFAR-10/100 were conducted on their complete test sets (10,000 samples each). For ImageNet, we used the full ImageNet1K validation set (50,000 samples across 1,000 classes), where \"1K\" refers to the number of classes rather than sample count. Attack Success Rate (ASR) was calculated as the percentage of sam" + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.887, + 0.8, + 0.9 + ], + "angle": 0, + "content": "" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1511" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.089, + 0.486, + 0.119 + ], + "angle": 0, + "content": "Table 1. Performance comparison between FastJSMA and JSMA variants with ResNet18" + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.129, + 0.478, + 0.281 + ], + "angle": 0, + "content": "
Model/DatasetMethodASR(%)Time(μs)L2Norm
CIFAR-10Random7.5531.42
FastJSMA43.983391.53
JSMANT43.7911,7691.57
WJSMA51.568,6411.29
CIFAR-100Random15.3131.44
FastJSMA45.307511.75
JSMANT43.8761,1661.55
WJSMA49.7851,0651.24
ImageNetRandom2.9787.78
FastJSMA54.897,9089.19
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.103, + 0.282, + 0.474, + 0.292 + ], + "angle": 0, + "content": "Note: JSMANT and WJSMA could not be tested on ImageNet due to memory constraints" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.32, + 0.483, + 0.35 + ], + "angle": 0, + "content": "plies that were successfully misclassified after perturbation out of all tested samples." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.351, + 0.483, + 0.502 + ], + "angle": 0, + "content": "Notably, since JSMA-based methods operate iteratively (modifying two pixels per iteration), JSMANT and WJSMA use half the number of iterations compared to the total pixel modification count. FastJSMA modifies one pixel per iteration for fair comparison. The attempt parameter controls the number of candidate classes explored during class probing: we use attempt \\(= 10\\) for CIFAR-10, attempt \\(= \\sqrt{N}\\) for larger datasets (e.g., attempt \\(= 10\\) for CIFAR-100's 100 classes, attempt \\(= 32\\) for ImageNet's 1000 classes)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.512, + 0.295, + 0.527 + ], + "angle": 0, + "content": "4.2. Performance Analysis" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.535, + 0.371, + 0.549 + ], + "angle": 0, + "content": "4.2.1. Comparison with JSMA Variants" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.483, + 0.644 + ], + "angle": 0, + "content": "Table 1 presents our main experimental results, comparing FastJSMA against JSMANT, WJSMA, and the random baseline across different datasets and model architectures. We report three key metrics: Attack Success Rate (ASR), average generation time per sample, and average perturbation magnitude (\\( L_{2} \\) norm)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.644, + 0.483, + 0.825 + ], + "angle": 0, + "content": "The results empirically validate our theoretical complexity reduction from \\( O(M^2 \\times N) \\) to \\( O(M\\sqrt{N}) \\) proposed in Section 3.4. FastJSMA achieves attack success rates comparable to or exceeding traditional JSMANT across all datasets while demonstrating remarkable computational efficiency—requiring just \\( 2.9\\% \\) of JSMANT's processing time on CIFAR-10 (339\\( \\mu \\)s vs. 11,769\\( \\mu \\)s) and merely \\( 1.2\\% \\) on CIFAR-100 (751\\( \\mu \\)s vs. 61,166\\( \\mu \\)s). These speed-ups of \\( 34.7\\times \\) and \\( 81.4\\times \\) respectively confirm that the gradient decoupling mechanism theorized in our mathematical formulation successfully transforms the computational scaling properties of saliency-based attacks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.903 + ], + "angle": 0, + "content": "The efficiency advantage increases with dataset complexity, closely matching our theoretical prediction that traditional JSMA's complexity worsens quadratically with increasing image size and class count. Most significantly, FastJSMA successfully executes on ImageNet with a rea" + }, + { + "type": "table_caption", + "bbox": [ + 0.566, + 0.09, + 0.855, + 0.105 + ], + "angle": 0, + "content": "Table 2. Ablation study on gradient components" + }, + { + "type": "table", + "bbox": [ + 0.551, + 0.116, + 0.875, + 0.182 + ], + "angle": 0, + "content": "
Gradient ComponentASR(%)Time(μs)L2Norm
FastJSMA (Complete)43.983391.66
G+30.664321.71
G-35.754431.66
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.208, + 0.905, + 0.298 + ], + "angle": 0, + "content": "sonable computation time of \\(7,908\\mu s\\) per sample, while both JSMANT and WJSMA fail due to memory constraints. This empirical finding not only validates our theoretical analysis but extends its implications—demonstrating that our mathematical reformulation enables saliency-based attacks on previously inaccessible large-scale datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.298, + 0.906, + 0.435 + ], + "angle": 0, + "content": "WJSMA achieves somewhat higher ASRs on CIFAR datasets, which aligns with our theoretical understanding of its specialized optimization mechanisms. However, its computational demands confirm our analysis that approaches relying on the full Jacobian matrix face fundamental limitations for larger-scale applications. These results demonstrate the theoretical-practical trade-off that motivated our work: FastJSMA sacrifices minimal attack effectiveness for transformative computational efficiency gains." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.444, + 0.846, + 0.459 + ], + "angle": 0, + "content": "4.2.2. Ablation Study on Gradient Components" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.906, + 0.538 + ], + "angle": 0, + "content": "To investigate the contribution of different gradient components in our dual-gradient approximation strategy, we conducted an ablation study comparing three variants of FastJSMA using the CIFAR-10 dataset with ResNet18, 8 pixels modification, and attempt=10." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.906, + 0.719 + ], + "angle": 0, + "content": "These results validate and refine our theoretical decomposition of the Jacobian matrix described in Section 3.2. The complete FastJSMA incorporating both gradients achieves the highest ASR (43.98%), empirically confirming that our dual-gradient formulation successfully captures the bidirectional influence encoded in traditional JSMA's saliency computation. The class suppression gradient \\((g^{-})\\) alone yields a \\(35.75\\%\\) ASR, outperforming the class excitation gradient \\((g^{+})\\) alone (30.66%), revealing a key theoretical insight: reducing the current class probability has a more significant impact than increasing alternative class probabilities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.901 + ], + "angle": 0, + "content": "This empirical finding extends our initial theoretical understanding, revealing an asymmetric contribution pattern not initially anticipated in our mathematical formulation. This discovery explains the effectiveness of our normalization procedure in Section 3.3, which amplifies the class suppression gradient by a factor of \\( N \\) —the ablation empirically validates this engineering optimization while simultaneously deepening our theoretical understanding. The comparable computation times across variants (339-443\\(\\mu\\)s) further validate our complexity analysis that dual gradient computation maintains \\( O(M\\sqrt{N}) \\) efficiency regardless of variant configuration." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1512" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.207, + 0.093, + 0.804, + 0.107 + ], + "angle": 0, + "content": "(a) Fast-JSMA Sensitivity to Max Attempt Fast-JSMA Parameter Sensitivity Analysis (b) Fast-JSMA Sensitivity to Pixels Perturbed" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.109, + 0.498, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.11, + 0.902, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.281, + 0.908, + 0.337 + ], + "angle": 0, + "content": "Figure 2. Parameter sensitivity analysis of FastJSMA on CIFAR-10. (a) Sensitivity to Max Attempt parameter showing ASR and Time curves. (b) Sensitivity to pixels perturbed showing ASR, Time, and Noise curves. All metrics are normalized using min-max scaling for visual comparison; actual ASR percentages shown on right y-axis. Experiments conducted on the complete CIFAR-10 test set (10,000 samples) with ResNet18." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.364, + 0.352, + 0.379 + ], + "angle": 0, + "content": "4.2.3. Parameter Sensitivity Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.384, + 0.483, + 0.459 + ], + "angle": 0, + "content": "To evaluate FastJSMA's robustness to parameter settings, we analyzed two key parameters: pixel modification count and the attempt parameter. Figure 2 presents these analyses conducted on the complete CIFAR-10 test set (10,000 samples)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.461, + 0.483, + 0.642 + ], + "angle": 0, + "content": "Our Max Attempt parameter analysis (Figure 2a) shows that attack success rate follows a non-linear pattern: rapid initial growth from attempt \\(= 1\\) to attempt \\(= 4\\), followed by gradual increase thereafter. This validates our square-root sampling heuristic described in Section 3.4, confirming that smaller attempt values provide higher marginal benefits. For scalability, we empirically determined that \\(\\sqrt{N}\\) sampling works well for larger datasets. Computational cost scales linearly with attempt value, supporting our \\(O(M\\sqrt{N})\\) complexity analysis. The data ranges from \\(35.2\\%\\) ASR at attempt \\(= 1\\) to \\(44.7\\%\\) ASR at attempt \\(= 10\\), with corresponding time costs increasing from \\(314\\mu s\\) to \\(417\\mu s\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.824 + ], + "angle": 0, + "content": "The pixel modification analysis (Figure 2b) reveals that FastJSMA achieves substantial attack success even with minimal pixel modifications, with performance following an S-curve: initial rapid growth \\((2\\rightarrow 16\\) pixels), moderate middle phase growth \\((16\\rightarrow 36\\) pixels), and final diminishing returns \\((36\\rightarrow 50\\) pixels). This validates our significance threshold mechanism in Section 3.3. Runtime increases only marginally despite significant increases in modified pixels, while L2 norm scales approximately linearly with pixel count. The analysis shows ASR ranging from \\(26.0\\%\\) with 2 pixels to \\(69.5\\%\\) with 50 pixels, with average generation time remaining stable around 0.0003-0.0004 seconds." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.902 + ], + "angle": 0, + "content": "These findings both validate and extend our theoretical model. The consistent computational efficiency across different pixel modification counts demonstrates FastJSMA's scalability, while the sharp initial gains in both analyses highlight the method's effectiveness at identifying influen" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.364, + 0.907, + 0.394 + ], + "angle": 0, + "content": "tial pixels with minimal computational overhead, making it particularly suitable for constrained attack scenarios." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.462, + 0.633, + 0.478 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.505, + 0.907, + 0.657 + ], + "angle": 0, + "content": "This paper introduces FastJSMA, a computationally efficient reformulation of saliency-based adversarial attacks. By decomposing the traditional Jacobian calculation into complementary gradient components—class suppression \\((g^{-})\\) and class excitation \\((g^{+})\\)—we reduce computational complexity from \\(O(M^2 \\times N)\\) to \\(O(M\\sqrt{N})\\) while maintaining attack effectiveness. Our class probing mechanism and significance threshold further optimize performance by systematically identifying vulnerable decision boundaries and prioritizing influential pixels." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Experiments validate that FastJSMA achieves dramatic computational efficiency gains—up to \\(81.4 \\times\\) speedup on CIFAR-100 and successful execution on ImageNet where traditional JSMA fails entirely—while maintaining comparable attack success rates. The ablation study reveals the asymmetric contribution of gradient components, providing new insights that extend our theoretical understanding. These findings demonstrate that our mathematical reformulation successfully overcomes the computational barriers that have limited JSMA's applicability in large-scale settings. Beyond immediate technical contributions, FastJSMA's gradient decoupling principles may inform future research in computationally efficient adversarial attacks and explainable AI, where feature importance analysis at scale remains a significant challenge." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1513" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.237 + ], + "angle": 0, + "content": "We would like to thank the HPC Platform of Huazhong University of Science and Technology for providing computational resources. We also express our gratitude to the School of Artificial Intelligence and Automation, the School of Software Engineering and the School of Electronic Information and Communications at Huazhong University of Science and Technology for their strong support throughout this research." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.251, + 0.188, + 0.267 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.277, + 0.484, + 0.331 + ], + "angle": 0, + "content": "[1] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.333, + 0.484, + 0.375 + ], + "angle": 0, + "content": "[2] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.378, + 0.484, + 0.432 + ], + "angle": 0, + "content": "[3] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.434, + 0.484, + 0.475 + ], + "angle": 0, + "content": "[4] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In 2017 IEEE Symposium on Security and Privacy (SP), pages 39-57. IEEE, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.477, + 0.484, + 0.546 + ], + "angle": 0, + "content": "[5] Nicolas Papernot, Patrick McDaniel, Somesh Jha, Matt Fredrikson, Z. Berkay Celik, and Ananthram Swami. The limitations of deep learning in adversarial settings. In 2016 IEEE European Symposium on Security and Privacy (EuroS&P), pages 372-387. IEEE, 2016. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.548, + 0.484, + 0.603 + ], + "angle": 0, + "content": "[6] Théo Combey, António Loison, Maxime Faucher, and Hatem Hajri. Probabilistic jacobian-based saliency maps attacks. Machine Learning and Knowledge Extraction, 2(4): 558-578, 2020. doi: 10.3390/make2040030. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.604, + 0.484, + 0.645 + ], + "angle": 0, + "content": "[7] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In CVPR09, 2009. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.647, + 0.484, + 0.702 + ], + "angle": 0, + "content": "[8] Nicolas Papernot, Patrick McDaniel, and Ian Goodfellow. Transferability in machine learning: from phenomena to black-box attacks using adversarial samples. arXiv preprint arXiv:1605.07277, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.704, + 0.484, + 0.786 + ], + "angle": 0, + "content": "[9] Pin-Yu Chen, Huan Zhang, Yash Sharma, Jinfeng Yi, and Cho-Jui Hsieh. Zoo: Zeroth order optimization based blackbox attacks to deep neural networks without training substitute models. In Proceedings of the 10th ACM Workshop on Artificial Intelligence and Security, pages 15-26. ACM, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.484, + 0.844 + ], + "angle": 0, + "content": "[10] Dawei Zhou, Nannan Wang, Tongliang Liu, and Xinbo Gao. Improving adversarial training from the perspective of class-flipping distribution. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[11] Dawei Zhou, Nannan Wang, Bo Han, Tongliang Liu, and Xinbo Gao. Defending against adversarial examples via modeling adversarial noise. International Journal of Computer Vision, pages 1-18, 2025." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.277, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.162 + ], + "angle": 0, + "content": "[12] Chaojian Yu, Bo Han, Li Shen, Jun Yu, Chen Gong, Mingming Gong, and Tongliang Liu. Understanding robust overfitting of adversarial training and beyond. In International Conference on Machine Learning, pages 25595-25610. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.907, + 0.233 + ], + "angle": 0, + "content": "[13] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and accurate method to fool deep neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2574–2582, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.235, + 0.906, + 0.303 + ], + "angle": 0, + "content": "[14] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9185-9193, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.906, + 0.36 + ], + "angle": 0, + "content": "[15] Jiadong Lin, Chuanbiao Song, Kun He, Liwei Wang, and John E. Hopcroft. Nesterov accelerated gradient and scale invariance for adversarial attacks. In International Conference on Learning Representations, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.906, + 0.432 + ], + "angle": 0, + "content": "[16] Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Evading defenses to transferable adversarial examples by translation-invariant attacks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4312-4321, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.906, + 0.503 + ], + "angle": 0, + "content": "[17] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L. Yuille. Improving transferability of adversarial examples with input diversity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2730-2739, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.906, + 0.572 + ], + "angle": 0, + "content": "[18] Chengzhi Mao, Mia Chiquier, Hao Wang, Junfeng Yang, and Carl Vondrick. Adversarial attacks are reversible with natural supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 661-671, October 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.575, + 0.906, + 0.63 + ], + "angle": 0, + "content": "[19] Jianming Zhang and Stan Sclaroff. Saliency detection: A boolean map approach. In 2013 IEEE International Conference on Computer Vision, pages 153-160, 2013. doi: 10.1109/ICCV.2013.26.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.906, + 0.686 + ], + "angle": 0, + "content": "[20] Jiawei Su, Danilo Vasconcellos Vargas, and Kouichi Sakurai. One pixel attack for fooling deep neural networks. IEEE Transactions on Evolutionary Computation, 23(5):828-841, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.688, + 0.906, + 0.745 + ], + "angle": 0, + "content": "[21] Jianbo Chen, Michael I. Jordan, and Martin J. Wainwright. Hopskipjumpattack: A query-efficient decision-based attack. In 2020 IEEE Symposium on Security and Privacy (SP), pages 1277-1294. IEEE, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.907, + 0.801 + ], + "angle": 0, + "content": "[22] Tianhang Zheng, Changyou Chen, Junsong Yuan, Bo Li, and Kui Ren. Pointcloud saliency maps. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.906, + 0.87 + ], + "angle": 0, + "content": "[23] Apostolos Modas, Seyed-Mohsen Moosavi-Dezfooli, and Pascal Frossard. Sparsefool: a few pixels make a big difference. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9087–9096, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.906, + 0.902 + ], + "angle": 0, + "content": "[24] Florian Tramér, Alexey Kurakin, Nicolas Papernot, Ian Goodfellow, Dan Boneh, and Patrick McDaniel. Ensemble" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1514" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "adversarial training: Attacks and defenses. In International Conference on Learning Representations, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.178 + ], + "angle": 0, + "content": "[25] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International Conference on Machine Learning, pages 2206-2216. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.483, + 0.247 + ], + "angle": 0, + "content": "[26] Maksym Andriushchenko, Francesco Croce, Nicolas Flammarion, and Matthias Hein. Square attack: a query-efficient black-box adversarial attack via random search. In European Conference on Computer Vision, pages 484-501. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.483, + 0.303 + ], + "angle": 0, + "content": "[27] Maura Pintor, Fabio Roli, Wieland Brendel, and Battista Biggio. Fast minimum-norm adversarial attacks through adaptive norm constraints. arXiv preprint arXiv:2102.12827, 2021.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.483, + 0.347 + ], + "angle": 0, + "content": "[28] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. Technical Report 4, University of Toronto, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.483, + 0.404 + ], + "angle": 0, + "content": "[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1515" + } + ] +] \ No newline at end of file diff --git a/2025/FastJSMA_ Accelerating Jacobian-based Saliency Map Attacks through Gradient Decoupling/caddbdfc-9052-49a6-8b20-7e4e162ec5a3_origin.pdf b/2025/FastJSMA_ Accelerating Jacobian-based Saliency Map Attacks through Gradient Decoupling/caddbdfc-9052-49a6-8b20-7e4e162ec5a3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4476aa45d1705779f83dd5067945e4e9de724b99 --- /dev/null +++ b/2025/FastJSMA_ Accelerating Jacobian-based Saliency Map Attacks through Gradient Decoupling/caddbdfc-9052-49a6-8b20-7e4e162ec5a3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aaa594212270a0a09ae004c4f7e83ef4aa674d165778947377162e75d211f690 +size 604399 diff --git a/2025/FastJSMA_ Accelerating Jacobian-based Saliency Map Attacks through Gradient Decoupling/full.md b/2025/FastJSMA_ Accelerating Jacobian-based Saliency Map Attacks through Gradient Decoupling/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b6830d9a55b75edf524d1b5d11078a9ae8fe9d38 --- /dev/null +++ b/2025/FastJSMA_ Accelerating Jacobian-based Saliency Map Attacks through Gradient Decoupling/full.md @@ -0,0 +1,336 @@ +# FastJSMA: Accelerating Jacobian-based Saliency Map Attacks through Gradient Decoupling + +Zhenghao $\mathrm{Gao}^{1,*}$ Shengjie $\mathrm{Xu}^{2,*}$ Zijing Li $^{2}$ Meixi Chen $^{3}$ Chaojian $\mathrm{Yu}^{4}$ Yuanjie $\mathrm{Shao}^{4}$ Changxin $\mathrm{Gao}^{1,\dagger}$ + +1School of Artificial Intelligence and Automation, Huazhong University of Science and Technology + +$^{2}$ School of Software Engineering, Huazhong University of Science and Technology + +3Journalism and Information Communication School, Huazhong University of Science and Technology + +$^{4}$ School of Electronic Information and Communications, Huazhong University of Science and Technology + +{u202215226,u202317280,u202317273,u202317218,yuchaojian,shaoyuanjie,cgao}@hust.edu.cn + +# Abstract + +Adversarial attack plays a critical role in evaluating the robustness of deep learning models. Jacobian-based Saliency Map Attack (JSMA) is an interpretable adversarial method that offers excellent pixel-level control and provides valuable insights into model vulnerabilities. However, its quadratic computational complexity $O(M^2 \times N)$ renders it impractical for large-scale datasets, limiting its application despite its inherent value. This paper proposes FastJSMA, an efficient attack method that addresses these computational limitations. Our approach introduces a gradient decoupling mechanism that decomposes the Jacobian calculation into complementary class suppression ( $g^{-}$ ) and class excitation ( $g^{+}$ ) gradients, reducing complexity to $O(M\sqrt{N})$ . Additionally, we implement a class probing mechanism and an adaptive saliency threshold to further optimize the process. Experimental results across multiple datasets demonstrate that FastJSMA maintains comparable attack success rates while dramatically reducing computation time—requiring only 2.9% of JSMA's processing time on CIFAR-10 and 1.2% on CIFAR-100, and successfully operating on ImageNet where traditional JSMA fails due to memory constraints. This advancement enables the practical application of interpretable saliency map-based attacks on large-scale datasets, balancing effectiveness with computational efficiency. + +# 1. Introduction + +# 1.1. Research Background + +Deep learning has achieved breakthrough progress in computer vision, natural language processing, and other fields, + +but its security issues have increasingly attracted attention. In 2013, Szegedy et al. [1] first discovered the sensitivity of deep neural networks to small perturbations, i.e., by adding imperceptible perturbations to inputs, models can be induced to produce incorrect predictions. This discovery revealed potential vulnerabilities in deep learning systems and drove rapid development in adversarial attack research. + +Adversarial attack techniques have evolved from simple approaches like the Fast Gradient Sign Method (FGSM) [2] to complex methods such as PGD [3] and C&W attacks [4], each offering different trade-offs between effectiveness, efficiency, and perturbation magnitude. + +# 1.2. Research Motivation + +The Jacobian-based Saliency Map Attack (JSMA) [5] is an adversarial method that identifies and modifies the most influential pixels through Jacobian matrix analysis. It offers unique advantages including intuitive visualization through saliency maps that enhance interpretability, precise control over modified pixels, and strong effectiveness across various datasets. Despite these merits and attempts to improve it through weighted approaches like WJSMA [6], JSMA remains largely underutilized due to prohibitive computational costs. Its full Jacobian matrix calculation requires $O(M^2 \times N)$ complexity (where $M$ is the number of pixels and $N$ is the number of classes), creating an insurmountable barrier for large-scale applications like ImageNet [7]. + +This limitation stems from three factors: (a) separate backward propagation for each output node; (b) saliency evaluations across the entire Jacobian matrix; and (c) repeated calculations across iterations. Consequently, the adversarial machine learning community has effectively abandoned JSMA's valuable properties due to computational constraints rather than methodological limitations. + +# 1.3. Contributions + +To revitalize this valuable but computationally constrained method, we propose FastJSMA (Fast Jacobian-based Saliency Map Attack), an efficient adversarial attack based on gradient decoupling approximation: + +- Gradient Decoupling Mechanism: We decompose the Jacobian matrix into complementary class suppression $(g^{-})$ and excitation $(g^{+})$ gradients, reducing computational complexity from $O(M^2 \times N)$ to $O(M\sqrt{N})$ while maintaining attack effectiveness through dynamic weight allocation. +- Class Probing and Adaptive Threshold: We introduce a square-root sampling heuristic to efficiently identify vulnerable decision boundaries and an adaptive threshold mechanism to prioritize influential pixels, optimizing both efficiency and effectiveness. +- Comprehensive Evaluation: Our experiments demonstrate dramatic efficiency gains—requiring only $2.9\%$ of JSMANT's processing time on CIFAR-10 and achieving up to $81.4 \times$ speedup on CIFAR-100 while maintaining comparable attack success rates. Most notably, FastJSMA successfully processes ImageNet samples, where traditional JSMA and WJSMA fail entirely due to memory constraints. + +# 2. Related Work + +# 2.1. Overview of Adversarial Attack Methods + +Adversarial attacks can be categorized as white-box or black-box attacks based on the attacker's knowledge of the model. White-box attacks assume the attacker has complete access to the model structure and parameters, with representative methods including FGSM [2] and PGD [3]. Black-box attacks can only access the model's input-output interface, primarily relying on transferability [8] or query optimization [9]. Recent work has also focused on improving adversarial training defenses [10-12]. + +From the perspective of attack objectives, they can be divided into targeted and non-targeted attacks. Targeted attacks aim to misclassify input samples as specified target classes, such as DeepFool [13]; non-targeted attacks only need to cause incorrect predictions, regardless of the specific error class, such as FGSM. These two types of attack methods have their respective advantages in practical applications: targeted attacks enable more precise control, while non-targeted attacks typically have higher success rates and lower computational costs. + +Gradient-based adversarial attacks have evolved significantly. After FGSM was proposed, researchers developed improved variants: (1) MI-FGSM [14] introduces momentum terms; (2) NI-FGSM [15] utilizes Nesterov accelerated gradients; (3) TI-FGSM [16] enhances transferability through translation invariance; (4) DI-FGSM [17] adopts + +diversified input strategies. Recent research by Mao et al. [18] discovered that adversarial attacks are reversible with natural supervision, providing new insights into the fundamental properties of adversarial perturbations. These methods optimize gradient-based attack strategies from different perspectives but have not fully addressed the issue of interpretability. + +# 2.2. Detailed Analysis of JSMA + +As an interpretable attack method, JSMA was proposed by Papernot et al. [5], with its core idea being the use of the Jacobian matrix to identify input features that have the greatest impact on model output. This approach builds upon fundamental research in saliency detection, such as the boolean map approach [19], which provides theoretical foundations for identifying visually significant regions. Specifically, for input $x$ and target class $t$ , JSMA first calculates the Jacobian matrix: + +$$ +J _ {F} (x) = \frac {\partial F (x)}{\partial x} = \left[ \frac {\partial F _ {j} (x)}{\partial x _ {i}} \right] _ {i, j} \tag {1} +$$ + +where $F_{j}(x)$ represents the model's prediction probability for class $j$ . Based on the Jacobian matrix, JSMA defines the saliency map $S$ : + +$$ +S (x, t) [ i ] = \left\{ \begin{array}{l} 0, \text {i f} \frac {\partial F _ {t} (x)}{\partial x _ {i}} < 0 \text {o r} \sum_ {j \neq t} \frac {\partial F _ {j} (x)}{\partial x _ {i}} > 0 \\ \left| \frac {\partial F _ {t} (x)}{\partial x _ {i}} \right| \cdot \left| \sum_ {j \neq t} \frac {\partial F _ {j} (x)}{\partial x _ {i}} \right|, \text {o t h e r w i s e} \end{array} \right. \tag {2} +$$ + +Subsequently, researchers have made various improvements to JSMA: (1) JSMA++ [8] enhances attack efficiency by optimizing saliency calculations; (2) One-pixel JSMA [20] restricts modifications to a single pixel, achieving more precise control; (3) Adaptive JSMA [21] introduces adaptive threshold strategies, improving attack success rates. The concept of saliency maps has also been extended to other domains, such as 3D point clouds [22], demonstrating the versatility of saliency-based approaches across different data modalities. + +An important advancement is the Weighted JSMA (WJSMA) [6], which addresses limitations in the original JSMA's saliency calculation. WJSMA penalizes gradients associated with small probabilities to create more balanced saliency maps. WJSMA incorporates class probabilities as weights in the saliency calculation, setting the saliency to zero when specific gradient conditions are not met. This weighting mechanism makes WJSMA faster than the original JSMA while maintaining good balance between attack success rate and computational cost. However, even with these improvements, WJSMA still inherits the fundamental computational inefficiency of JSMA for large-scale datasets, as it retains the $O(M^2 \times N)$ complexity that creates memory and processing bottlenecks on datasets like ImageNet. + +# 2.3. Related Optimization Techniques + +To address the efficiency problems of adversarial attacks in large-scale scenarios, researchers have proposed various optimization strategies. One class of methods improves efficiency by enhancing gradient computation, such as OnePixel Attack [20], which reduces search space through evolutionary algorithms, and SparseFool [23], which optimizes computational efficiency by leveraging sparsity. + +Another class of methods proposes specialized optimization strategies for large-scale datasets. For example, Fast Gradient Method [24] enhances attack efficiency through ensemble techniques, while AutoAttack [25] adopts automated strategies to select optimal attack parameters. + +Recent research has also explored algorithmic optimizations specifically targeting the computational complexity of gradient-based attacks. These approaches include gradient approximation techniques [26], adaptive norm constraints for efficient minimum-norm attacks [27], and decomposition strategies that reduce the dimensionality of the optimization space [21]. + +These optimization methods have their respective advantages and disadvantages: gradient-based methods offer high computational efficiency but lower interpretability; optimization-based methods provide good results but incur high computational costs; heuristic-based methods are simple and fast but have unstable success rates. In comparison, our proposed FastJSMA maintains JSMA's interpretability while significantly improving computational efficiency through dual-gradient approximation, effectively reducing the complexity from $O(M^2 \times N)$ to $O(M\sqrt{N})$ and making saliency map-based attacks practical for large-scale datasets for the first time. + +# 3. Methodology + +This section details the technical principles of FastJSMA. We first establish the mathematical model of the problem, then elaborate on the design rationale of the dual-gradient approximation strategy, followed by a detailed explanation of the saliency map generation process, and finally theoretically analyze the computational complexity of the algorithm. + +# 3.1. Problem Modeling and Attack Objectives + +For an input sample $x \in \mathbb{R}^{C \times H \times W}$ and a target classification model $f: \mathbb{R}^{C \times H \times W} \to \mathbb{R}^N$ , where $C, H, W$ represent the number of channels, height, and width respectively, and $N$ is the total number of classes. Given the original predicted class $t = \arg \max f(x)$ , the goal of the adversarial attack is to find a perturbation $\delta$ such that: + +$$ +\arg \max f (x + \delta) \neq t \quad \text {s . t .} \quad \| \delta \| _ {0} \leq \epsilon \tag {3} +$$ + +where $\epsilon$ represents the allowed number of pixels to modify, and $\| \cdot \|_0$ is the $L_0$ norm. The reasons for choosing the $L_0$ norm constraint rather than other norms (such as $L_2$ or $\mathrm{L\infty}$ ) are: (1) $L_0$ norm directly limits the number of modified pixels, conforming to the principle of minimal intervention; (2) a limited number of modifications makes the attack process more interpretable and understandable; (3) maintains consistent constraint forms with traditional JSMA for fair comparison. + +Traditional JSMA computes the complete Jacobian matrix $J = \nabla_{x}f(x)$ with $O(M^2\times N)$ complexity, where $M$ represents the total number of pixels and $N$ denotes the number of classification classes. This substantial computational burden and memory requirements render JSMA impractical for large-scale datasets and real-world applications. + +FastJSMA addresses this limitation through a dual-gradient approximation strategy that reduces complexity to $O(M\sqrt{N})$ while maintaining attack effectiveness. + +# 3.2. Target Function Construction + +The core mathematical insight that enables FastJSMA's efficiency breakthrough lies in our recognition that the Jacobian-based saliency computation in traditional JSMA can be decomposed into two complementary objectives: first, decreasing the probability of the current class, and second, increasing the probability of alternative classes. This decomposition forms the theoretical foundation of our dual-gradient approximation strategy. + +Analyzing the mathematical structure of traditional JSMA reveals that its saliency map computation essentially measures two gradient components for each pixel: how modifying that pixel decreases the current class score and how it increases other class scores. While traditional JSMA explicitly computes these gradients through the full Jacobian matrix, we observed that these two objectives could be approximated through carefully designed loss functions without requiring the prohibitively expensive Jacobian computation. + +To avoid computing the complete Jacobian matrix, we propose a dual-gradient approximation strategy. The core idea of this strategy is to construct two complementary objective functions that simulate the enhancement effect on the target class and the suppression effect on other classes in JSMA. The specific definitions are as follows: + +1. Class Suppression Gradient: Construct a virtual label $\hat{y}^{-} \in \mathbb{R}^{N}$ , where + +$$ +\hat {y} _ {i} ^ {-} = \left\{ \begin{array}{l l} f (x) _ {i} - 1 & i = t \\ f (x) _ {i} + \sqrt {N} & i = a (\text {a t t e m p t c l a s s , i f s p e c i f i e d}) \\ f (x) _ {i} & \text {o t h e r w i s e} \end{array} \right. \tag {4} +$$ + +![](images/e8442bc3756db7a5fe074877cc37b4f65160e8e9bd2315d3cd688ade07e7cf9e.jpg) +Figure 1. Architectural comparison between traditional JSMA and our proposed FastJSMA. Left: Traditional JSMA calculates the full Jacobian matrix with computational complexity $O(M^2 \times N)$ and relies on an iterative process to achieve misclassification, making it prohibitively expensive for large-scale datasets. Right: FastJSMA employs a class probing mechanism to select $\sqrt{N}$ candidate classes and a gradient decoupling strategy that approximates the saliency computation through class suppression gradient $(g^-)$ and class excitation gradient $(g^+)$ . The class suppression gradient reduces the current class probability while boosting an attempt class by $\sqrt{N}$ , and the significance threshold efficiently filters the most influential pixels. This approach reduces complexity to $O(M\sqrt{N})$ while maintaining comparable effectiveness. + +Calculate the gradient: + +$$ +g ^ {-} = \nabla_ {x} \mathcal {L} (f (x), \hat {y} ^ {-}) \tag {5} +$$ + +This gradient reflects the direction of reducing the prediction probability of the current class and optionally enhancing a specific alternative class if an attempt parameter is provided. + +2. Class Excitation Gradient: Construct an enhancement label $\hat{y}^{+}\in \mathbb{R}^{N}$ , where + +$$ +\hat {y} _ {i} ^ {+} = \left\{ \begin{array}{l l} f (x) _ {i} - 1 & i = t \\ f (x) _ {i} + 1 & \text {o t h e r w i s e} \end{array} \right. \tag {6} +$$ + +Calculate the gradient: + +$$ +g ^ {+} = \nabla_ {x} \mathcal {L} (f (x), \hat {y} ^ {+}) \tag {7} +$$ + +This gradient indicates the direction of enhancing the prediction probability of other classes while suppressing the current class. + +This formulation represents a significant mathematical innovation over traditional JSMA. Rather than explicitly computing $N \times M$ partial derivatives as required by the Jacobian matrix, our approach leverages automatic differentiation to compute just two gradient vectors—reducing the computational complexity by a factor proportional to the number of classes $N$ and the number of pixels $M$ . Import- + +tantly, this efficiency gain comes without compromising the underlying attack methodology; the dual-gradient approach preserves the core mathematical principles of JSMA while dramatically reducing its computational footprint. + +The design of these two gradients cleverly avoids directly computing the Jacobian matrix while retaining the core idea of JSMA: simultaneously considering the suppression of the target class and the excitation of other classes. + +# 3.3. Saliency Map Generation + +Traditional JSMA generates saliency maps through exhaustive computation of the Jacobian matrix followed by pairwise evaluation of pixel modifications—a process that scales quadratically with the number of pixels. FastJSMA fundamentally reconceptualizes this approach through an elegant mathematical transformation that yields equivalent results with substantially reduced computation. + +Based on the dual-gradient approximation, we apply gradient normalization to ensure balanced contribution while strategically emphasizing the class suppression component: + +$$ +g _ {\text {n o r m a l i z e d}} ^ {-} = \frac {g ^ {-}}{\operatorname {m e a n} (| g ^ {-} |)} \times N \tag {8} +$$ + +$$ +g _ {\text {n o r m a l i z e d}} ^ {+} = \frac {g ^ {+}}{\operatorname {m e a n} (| g ^ {+} |)} \tag {9} +$$ + +This normalization procedure includes a deliberate engineering optimization: the class suppression gradient is scaled by a factor of $N$ (the number of classes) to amplify its influence in the overall saliency calculation. Our empirical investigation indicates that this scaling enhances attack effectiveness by emphasizing the suppression of the original class prediction. + +We then integrate these normalized gradients to obtain a comprehensive gain matrix: + +$$ +G = g _ {\text {n o r m a l i z e d}} ^ {+} + g _ {\text {n o r m a l i z e d}} ^ {-} \tag {10} +$$ + +This matrix captures the bidirectional influence of modifying each pixel on both decreasing the target class probability and increasing alternative class probabilities—precisely the information encapsulated in traditional JSMA's saliency computation, but obtained through a mathematically equivalent yet computationally superior formulation. + +After obtaining the gain matrix, we apply a significance threshold to filter out pixels with minimal impact. This threshold operation is visually represented as a funnel-like process in Figure 1, focusing computational resources only on the most influential pixels. By implementing this significance threshold, we effectively reduce the search space while ensuring that the selected pixels have meaningful contributions to the adversarial perturbation. + +This selective modification approach, governed by both the saliency ranking and the significance threshold, ensures efficient resource allocation by concentrating perturbations on the most influential and impactful pixels. + +# 3.4. Attack Success Rate Optimization via Class Probing + +A significant enhancement in our implementation is the introduction of a class-specific probing mechanism, visually illustrated in Figure 1 as "Class Probing Mechanism". This component substantially improves attack success rates while maintaining the untargeted nature of FastJSMA by intelligently selecting candidate classes for the attempt parameter. The optimization addresses a fundamental challenge in untargeted adversarial attacks: determining which gradient directions most efficiently induce misclassification. + +Our implementation employs an intelligent candidate class exploration strategy governed by a square-root sampling heuristic: + +$$ +\text {n u m} _ {\text {c a n d i d a t e s}} = \sqrt {N} \tag {11} +$$ + +where $N$ represents the number of classes in the classification problem. This square-root relationship establishes a principled balance between computational efficiency and exploration thoroughness—scaling sublinearly with the class space dimensionality to maintain tractability even for large-scale classification problems. + +For each candidate class $c$ among the top numCandidates predicted classes (excluding the original class), the algorithm: + +1. Applies FastJSMA with class $c$ as an "attempt" parameter, which specifically modifies the class suppression gradient $(g^{-})$ in two ways: - Decreases the current class probability (by setting $\hat{y}_i^- = f(x)_i - 1$ for $i = t$ ) - Boosts the attempt class $c$ by a factor of $\sqrt{N}$ (by setting $\hat{y}_i^- = f(x)_i + \sqrt{N}$ for $i = c$ ) +2. Evaluates whether this modified gradient direction successfully induces misclassification +3. Returns the first successful perturbation, terminating further exploration + +This dual action of decreasing the current class while simultaneously boosting a specific attempt class efficiently guides the gradient toward vulnerable regions of the decision boundary, substantially improving attack success rates without requiring exhaustive exploration of all possible class transitions. + +It is crucial to emphasize that this approach does not constitute a targeted attack in the conventional sense. Rather, the attempt parameter merely directs the gradient toward regions where transitions away from the original class are most probable, while maintaining an untargeted classification objective. This gradient steering mechanism represents + +a methodological innovation particularly effective in high-dimensional classification spaces where traditional methods might struggle to efficiently locate vulnerable decision boundaries. + +Only if all candidate-based attempts fail does the algorithm revert to the standard untargeted FastJSMA approach without class-specific gradient modification. This hierarchical strategy enhances attack success rates while maintaining the untargeted objective—an engineering optimization that improves practical performance without altering the fundamental attack paradigm. + +# 3.5. Complexity Analysis + +To theoretically analyze algorithm efficiency, we compare the computational complexity of FastJSMA with traditional JSMA. Assuming a dataset with $N$ classes, where each image has $M$ pixels. The computational complexity of traditional JSMA is $O(M^2 \times N)$ . This is because, when constructing the saliency map, it needs to calculate the joint saliency of each pixel pair, considering $M^2$ order of magnitude pixel pairs; the saliency calculation for each pixel pair involves $N$ classes. Moreover, as illustrated in Figure 1, traditional JSMA typically requires multiple iterations of this expensive computation process to achieve successful misclassification, further compounding the computational burden. + +In contrast, our FastJSMA has a computational complexity of $O(M\sqrt{N})$ . This is because we only perform 2 gradient calculations. The combination of gradient decoupling, class probing mechanism, and significance threshold filtering together dramatically reduces the computational requirements while maintaining attack effectiveness. + +This significant reduction in complexity enables FastJSMA to efficiently process large-scale datasets. Due to the dual-gradient approximation strategy retaining JSMA's core ideas, the attack effectiveness is preserved despite substantial complexity reduction, as verified in our experiments. + +Our additional optimizations—including gradient normalization with class-specific scaling, dynamic significance thresholds, and the square-root probing strategy—further enhance efficiency without compromising mathematical foundations. These refinements demonstrate the importance of implementation details in translating theoretical innovations into practical adversarial methodologies that scale to contemporary deep learning applications. + +# 4. Experiments + +We present a comprehensive evaluation of FastJSMA, examining its effectiveness, efficiency, and transferability across multiple datasets and model architectures. Our experiments are designed to answer three key questions: (1) How does FastJSMA compare to traditional JSMA and its + +variants in terms of attack success and computational efficiency? (2) How do different components of FastJSMA contribute to its overall performance? (3) How sensitive is FastJSMA to key parameter settings? + +# 4.1. Experimental Setup + +# 4.1.1. Implementation Environment + +All experiments were conducted on a single NVIDIA RTX 3090 GPU with 24GB memory without virtual memory allocation. We implemented all methods using PyTorch framework and will release our code upon publication to ensure reproducibility. + +# 4.1.2. Datasets and Models + +To evaluate our method across varying scales of complexity, we selected three representative datasets: CIFAR10/100 [28] (natural color images sized $32 \times 32 \times 3$ , containing 10 and 100 classes respectively, each with 10,000 test samples) and ImageNet [7] (large-scale classification dataset with 1,000 classes, sized $224 \times 224 \times 3$ , with 50,000 validation samples). + +For model architectures, we employed widely-used ResNet variants (primarily ResNet18) [29], using pretrained checkpoints from HuggingFace1. These models achieve competitive performance on their respective datasets, providing a reliable foundation for our adversarial attack evaluation. + +# 4.1.3. Attack Configuration and Baselines + +We compare FastJSMA with several baseline methods: + +- JSMANT: The traditional Jacobian-based Saliency Map Attack [5] in its non-targeted form, which computes the full Jacobian matrix. +- WJSMA: The Weighted JSMA variant proposed by Combey et al. [6], which introduces probabilistic weights to enhance attack effectiveness. +- Random: A control baseline that randomly selects pixels for modification. + +To ensure fair comparison, we maintain consistent attack parameters across all methods. Since different datasets have different image sizes, we use a proportional approach to pixel modification: 8 pixels for $32 \times 32$ images (CIFAR-10/100) and 196 pixels for $224 \times 224$ images (ImageNet), representing approximately $0.78\%$ of the total pixels in each case. + +All experiments for CIFAR-10/100 were conducted on their complete test sets (10,000 samples each). For ImageNet, we used the full ImageNet1K validation set (50,000 samples across 1,000 classes), where "1K" refers to the number of classes rather than sample count. Attack Success Rate (ASR) was calculated as the percentage of sam + +Table 1. Performance comparison between FastJSMA and JSMA variants with ResNet18 + +
Model/DatasetMethodASR(%)Time(μs)L2Norm
CIFAR-10Random7.5531.42
FastJSMA43.983391.53
JSMANT43.7911,7691.57
WJSMA51.568,6411.29
CIFAR-100Random15.3131.44
FastJSMA45.307511.75
JSMANT43.8761,1661.55
WJSMA49.7851,0651.24
ImageNetRandom2.9787.78
FastJSMA54.897,9089.19
+ +Note: JSMANT and WJSMA could not be tested on ImageNet due to memory constraints + +plies that were successfully misclassified after perturbation out of all tested samples. + +Notably, since JSMA-based methods operate iteratively (modifying two pixels per iteration), JSMANT and WJSMA use half the number of iterations compared to the total pixel modification count. FastJSMA modifies one pixel per iteration for fair comparison. The attempt parameter controls the number of candidate classes explored during class probing: we use attempt $= 10$ for CIFAR-10, attempt $= \sqrt{N}$ for larger datasets (e.g., attempt $= 10$ for CIFAR-100's 100 classes, attempt $= 32$ for ImageNet's 1000 classes). + +# 4.2. Performance Analysis + +# 4.2.1. Comparison with JSMA Variants + +Table 1 presents our main experimental results, comparing FastJSMA against JSMANT, WJSMA, and the random baseline across different datasets and model architectures. We report three key metrics: Attack Success Rate (ASR), average generation time per sample, and average perturbation magnitude ( $L_{2}$ norm). + +The results empirically validate our theoretical complexity reduction from $O(M^2 \times N)$ to $O(M\sqrt{N})$ proposed in Section 3.4. FastJSMA achieves attack success rates comparable to or exceeding traditional JSMANT across all datasets while demonstrating remarkable computational efficiency—requiring just $2.9\%$ of JSMANT's processing time on CIFAR-10 (339 $\mu$ s vs. 11,769 $\mu$ s) and merely $1.2\%$ on CIFAR-100 (751 $\mu$ s vs. 61,166 $\mu$ s). These speed-ups of $34.7\times$ and $81.4\times$ respectively confirm that the gradient decoupling mechanism theorized in our mathematical formulation successfully transforms the computational scaling properties of saliency-based attacks. + +The efficiency advantage increases with dataset complexity, closely matching our theoretical prediction that traditional JSMA's complexity worsens quadratically with increasing image size and class count. Most significantly, FastJSMA successfully executes on ImageNet with a rea + +Table 2. Ablation study on gradient components + +
Gradient ComponentASR(%)Time(μs)L2Norm
FastJSMA (Complete)43.983391.66
G+30.664321.71
G-35.754431.66
+ +sonable computation time of $7,908\mu s$ per sample, while both JSMANT and WJSMA fail due to memory constraints. This empirical finding not only validates our theoretical analysis but extends its implications—demonstrating that our mathematical reformulation enables saliency-based attacks on previously inaccessible large-scale datasets. + +WJSMA achieves somewhat higher ASRs on CIFAR datasets, which aligns with our theoretical understanding of its specialized optimization mechanisms. However, its computational demands confirm our analysis that approaches relying on the full Jacobian matrix face fundamental limitations for larger-scale applications. These results demonstrate the theoretical-practical trade-off that motivated our work: FastJSMA sacrifices minimal attack effectiveness for transformative computational efficiency gains. + +# 4.2.2. Ablation Study on Gradient Components + +To investigate the contribution of different gradient components in our dual-gradient approximation strategy, we conducted an ablation study comparing three variants of FastJSMA using the CIFAR-10 dataset with ResNet18, 8 pixels modification, and attempt=10. + +These results validate and refine our theoretical decomposition of the Jacobian matrix described in Section 3.2. The complete FastJSMA incorporating both gradients achieves the highest ASR (43.98%), empirically confirming that our dual-gradient formulation successfully captures the bidirectional influence encoded in traditional JSMA's saliency computation. The class suppression gradient $(g^{-})$ alone yields a $35.75\%$ ASR, outperforming the class excitation gradient $(g^{+})$ alone (30.66%), revealing a key theoretical insight: reducing the current class probability has a more significant impact than increasing alternative class probabilities. + +This empirical finding extends our initial theoretical understanding, revealing an asymmetric contribution pattern not initially anticipated in our mathematical formulation. This discovery explains the effectiveness of our normalization procedure in Section 3.3, which amplifies the class suppression gradient by a factor of $N$ —the ablation empirically validates this engineering optimization while simultaneously deepening our theoretical understanding. The comparable computation times across variants (339-443 $\mu$ s) further validate our complexity analysis that dual gradient computation maintains $O(M\sqrt{N})$ efficiency regardless of variant configuration. + +![](images/d2d6803a5666c1e24d0a1dc0b433e4cc1ce61fe9ec414e897fdc68eb24897430.jpg) +(a) Fast-JSMA Sensitivity to Max Attempt Fast-JSMA Parameter Sensitivity Analysis (b) Fast-JSMA Sensitivity to Pixels Perturbed + +![](images/529430c6687121ff635434c1cc1500c63bdaf29e978dbfa5809953de1e1906a8.jpg) +Figure 2. Parameter sensitivity analysis of FastJSMA on CIFAR-10. (a) Sensitivity to Max Attempt parameter showing ASR and Time curves. (b) Sensitivity to pixels perturbed showing ASR, Time, and Noise curves. All metrics are normalized using min-max scaling for visual comparison; actual ASR percentages shown on right y-axis. Experiments conducted on the complete CIFAR-10 test set (10,000 samples) with ResNet18. + +# 4.2.3. Parameter Sensitivity Analysis + +To evaluate FastJSMA's robustness to parameter settings, we analyzed two key parameters: pixel modification count and the attempt parameter. Figure 2 presents these analyses conducted on the complete CIFAR-10 test set (10,000 samples). + +Our Max Attempt parameter analysis (Figure 2a) shows that attack success rate follows a non-linear pattern: rapid initial growth from attempt $= 1$ to attempt $= 4$ , followed by gradual increase thereafter. This validates our square-root sampling heuristic described in Section 3.4, confirming that smaller attempt values provide higher marginal benefits. For scalability, we empirically determined that $\sqrt{N}$ sampling works well for larger datasets. Computational cost scales linearly with attempt value, supporting our $O(M\sqrt{N})$ complexity analysis. The data ranges from $35.2\%$ ASR at attempt $= 1$ to $44.7\%$ ASR at attempt $= 10$ , with corresponding time costs increasing from $314\mu s$ to $417\mu s$ . + +The pixel modification analysis (Figure 2b) reveals that FastJSMA achieves substantial attack success even with minimal pixel modifications, with performance following an S-curve: initial rapid growth $(2\rightarrow 16$ pixels), moderate middle phase growth $(16\rightarrow 36$ pixels), and final diminishing returns $(36\rightarrow 50$ pixels). This validates our significance threshold mechanism in Section 3.3. Runtime increases only marginally despite significant increases in modified pixels, while L2 norm scales approximately linearly with pixel count. The analysis shows ASR ranging from $26.0\%$ with 2 pixels to $69.5\%$ with 50 pixels, with average generation time remaining stable around 0.0003-0.0004 seconds. + +These findings both validate and extend our theoretical model. The consistent computational efficiency across different pixel modification counts demonstrates FastJSMA's scalability, while the sharp initial gains in both analyses highlight the method's effectiveness at identifying influen + +tial pixels with minimal computational overhead, making it particularly suitable for constrained attack scenarios. + +# 5. Conclusion + +This paper introduces FastJSMA, a computationally efficient reformulation of saliency-based adversarial attacks. By decomposing the traditional Jacobian calculation into complementary gradient components—class suppression $(g^{-})$ and class excitation $(g^{+})$ —we reduce computational complexity from $O(M^2 \times N)$ to $O(M\sqrt{N})$ while maintaining attack effectiveness. Our class probing mechanism and significance threshold further optimize performance by systematically identifying vulnerable decision boundaries and prioritizing influential pixels. + +Experiments validate that FastJSMA achieves dramatic computational efficiency gains—up to $81.4 \times$ speedup on CIFAR-100 and successful execution on ImageNet where traditional JSMA fails entirely—while maintaining comparable attack success rates. The ablation study reveals the asymmetric contribution of gradient components, providing new insights that extend our theoretical understanding. These findings demonstrate that our mathematical reformulation successfully overcomes the computational barriers that have limited JSMA's applicability in large-scale settings. Beyond immediate technical contributions, FastJSMA's gradient decoupling principles may inform future research in computationally efficient adversarial attacks and explainable AI, where feature importance analysis at scale remains a significant challenge. + +# Acknowledgments + +# References + +[1] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 1 +[2] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 1, 2 +[3] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. 1, 2 +[4] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In 2017 IEEE Symposium on Security and Privacy (SP), pages 39-57. IEEE, 2017. 1 +[5] Nicolas Papernot, Patrick McDaniel, Somesh Jha, Matt Fredrikson, Z. Berkay Celik, and Ananthram Swami. The limitations of deep learning in adversarial settings. In 2016 IEEE European Symposium on Security and Privacy (EuroS&P), pages 372-387. IEEE, 2016. 1, 2, 6 +[6] Théo Combey, António Loison, Maxime Faucher, and Hatem Hajri. Probabilistic jacobian-based saliency maps attacks. Machine Learning and Knowledge Extraction, 2(4): 558-578, 2020. doi: 10.3390/make2040030. 1, 2, 6 +[7] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In CVPR09, 2009. 1, 6 +[8] Nicolas Papernot, Patrick McDaniel, and Ian Goodfellow. Transferability in machine learning: from phenomena to black-box attacks using adversarial samples. arXiv preprint arXiv:1605.07277, 2016. 2 +[9] Pin-Yu Chen, Huan Zhang, Yash Sharma, Jinfeng Yi, and Cho-Jui Hsieh. Zoo: Zeroth order optimization based blackbox attacks to deep neural networks without training substitute models. In Proceedings of the 10th ACM Workshop on Artificial Intelligence and Security, pages 15-26. ACM, 2017. 2 +[10] Dawei Zhou, Nannan Wang, Tongliang Liu, and Xinbo Gao. Improving adversarial training from the perspective of class-flipping distribution. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 2 +[11] Dawei Zhou, Nannan Wang, Bo Han, Tongliang Liu, and Xinbo Gao. Defending against adversarial examples via modeling adversarial noise. International Journal of Computer Vision, pages 1-18, 2025. + +We would like to thank the HPC Platform of Huazhong University of Science and Technology for providing computational resources. We also express our gratitude to the School of Artificial Intelligence and Automation, the School of Software Engineering and the School of Electronic Information and Communications at Huazhong University of Science and Technology for their strong support throughout this research. +[12] Chaojian Yu, Bo Han, Li Shen, Jun Yu, Chen Gong, Mingming Gong, and Tongliang Liu. Understanding robust overfitting of adversarial training and beyond. In International Conference on Machine Learning, pages 25595-25610. PMLR, 2022. 2 +[13] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and accurate method to fool deep neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2574–2582, 2016. 2 +[14] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9185-9193, 2018. 2 +[15] Jiadong Lin, Chuanbiao Song, Kun He, Liwei Wang, and John E. Hopcroft. Nesterov accelerated gradient and scale invariance for adversarial attacks. In International Conference on Learning Representations, 2020. 2 +[16] Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Evading defenses to transferable adversarial examples by translation-invariant attacks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4312-4321, 2019. 2 +[17] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L. Yuille. Improving transferability of adversarial examples with input diversity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2730-2739, 2019. 2 +[18] Chengzhi Mao, Mia Chiquier, Hao Wang, Junfeng Yang, and Carl Vondrick. Adversarial attacks are reversible with natural supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 661-671, October 2021. 2 +[19] Jianming Zhang and Stan Sclaroff. Saliency detection: A boolean map approach. In 2013 IEEE International Conference on Computer Vision, pages 153-160, 2013. doi: 10.1109/ICCV.2013.26.2 +[20] Jiawei Su, Danilo Vasconcellos Vargas, and Kouichi Sakurai. One pixel attack for fooling deep neural networks. IEEE Transactions on Evolutionary Computation, 23(5):828-841, 2019. 2, 3 +[21] Jianbo Chen, Michael I. Jordan, and Martin J. Wainwright. Hopskipjumpattack: A query-efficient decision-based attack. In 2020 IEEE Symposium on Security and Privacy (SP), pages 1277-1294. IEEE, 2020. 2, 3 +[22] Tianhang Zheng, Changyou Chen, Junsong Yuan, Bo Li, and Kui Ren. Pointcloud saliency maps. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019. 2 +[23] Apostolos Modas, Seyed-Mohsen Moosavi-Dezfooli, and Pascal Frossard. Sparsefool: a few pixels make a big difference. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9087–9096, 2019. 3 +[24] Florian Tramér, Alexey Kurakin, Nicolas Papernot, Ian Goodfellow, Dan Boneh, and Patrick McDaniel. Ensemble + +adversarial training: Attacks and defenses. In International Conference on Learning Representations, 2018. 3 +[25] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International Conference on Machine Learning, pages 2206-2216. PMLR, 2020. 3 +[26] Maksym Andriushchenko, Francesco Croce, Nicolas Flammarion, and Matthias Hein. Square attack: a query-efficient black-box adversarial attack via random search. In European Conference on Computer Vision, pages 484-501. Springer, 2020. 3 +[27] Maura Pintor, Fabio Roli, Wieland Brendel, and Battista Biggio. Fast minimum-norm adversarial attacks through adaptive norm constraints. arXiv preprint arXiv:2102.12827, 2021.3 +[28] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. Technical Report 4, University of Toronto, 2009. 6 +[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6 \ No newline at end of file diff --git a/2025/FastJSMA_ Accelerating Jacobian-based Saliency Map Attacks through Gradient Decoupling/images.zip b/2025/FastJSMA_ Accelerating Jacobian-based Saliency Map Attacks through Gradient Decoupling/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7229fa457de2f4689e3874298013f9466cea261b --- /dev/null +++ b/2025/FastJSMA_ Accelerating Jacobian-based Saliency Map Attacks through Gradient Decoupling/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7070e969e88dc0249955f1e7deddcfb1ab2d3b26dcf083c7e16ab8e5405d4c1b +size 283781 diff --git a/2025/FastJSMA_ Accelerating Jacobian-based Saliency Map Attacks through Gradient Decoupling/layout.json b/2025/FastJSMA_ Accelerating Jacobian-based Saliency Map Attacks through Gradient Decoupling/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..11ece6b85d37a2287fbe2c7418b4886b407f32be --- /dev/null +++ b/2025/FastJSMA_ Accelerating Jacobian-based Saliency Map Attacks through Gradient Decoupling/layout.json @@ -0,0 +1,9016 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 85, + 102, + 526, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 102, + 526, + 140 + ], + "spans": [ + { + "bbox": [ + 85, + 102, + 526, + 140 + ], + "type": "text", + "content": "FastJSMA: Accelerating Jacobian-based Saliency Map Attacks through Gradient Decoupling" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 94, + 162, + 515, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 162, + 515, + 175 + ], + "spans": [ + { + "bbox": [ + 94, + 162, + 515, + 175 + ], + "type": "text", + "content": "Zhenghao " + }, + { + "bbox": [ + 94, + 162, + 515, + 175 + ], + "type": "inline_equation", + "content": "\\mathrm{Gao}^{1,*}" + }, + { + "bbox": [ + 94, + 162, + 515, + 175 + ], + "type": "text", + "content": " Shengjie " + }, + { + "bbox": [ + 94, + 162, + 515, + 175 + ], + "type": "inline_equation", + "content": "\\mathrm{Xu}^{2,*}" + }, + { + "bbox": [ + 94, + 162, + 515, + 175 + ], + "type": "text", + "content": " Zijing Li" + }, + { + "bbox": [ + 94, + 162, + 515, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 94, + 162, + 515, + 175 + ], + "type": "text", + "content": " Meixi Chen" + }, + { + "bbox": [ + 94, + 162, + 515, + 175 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 94, + 162, + 515, + 175 + ], + "type": "text", + "content": " Chaojian " + }, + { + "bbox": [ + 94, + 162, + 515, + 175 + ], + "type": "inline_equation", + "content": "\\mathrm{Yu}^{4}" + }, + { + "bbox": [ + 94, + 162, + 515, + 175 + ], + "type": "text", + "content": " Yuanjie " + }, + { + "bbox": [ + 94, + 162, + 515, + 175 + ], + "type": "inline_equation", + "content": "\\mathrm{Shao}^{4}" + }, + { + "bbox": [ + 94, + 162, + 515, + 175 + ], + "type": "text", + "content": " Changxin " + }, + { + "bbox": [ + 94, + 162, + 515, + 175 + ], + "type": "inline_equation", + "content": "\\mathrm{Gao}^{1,\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 176, + 506, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 176, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 506, + 189 + ], + "type": "text", + "content": "1School of Artificial Intelligence and Automation, Huazhong University of Science and Technology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 153, + 190, + 456, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 190, + 456, + 203 + ], + "spans": [ + { + "bbox": [ + 153, + 190, + 456, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 153, + 190, + 456, + 203 + ], + "type": "text", + "content": "School of Software Engineering, Huazhong University of Science and Technology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 117, + 204, + 492, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 204, + 492, + 217 + ], + "spans": [ + { + "bbox": [ + 117, + 204, + 492, + 217 + ], + "type": "text", + "content": "3Journalism and Information Communication School, Huazhong University of Science and Technology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 113, + 219, + 497, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 219, + 497, + 231 + ], + "spans": [ + { + "bbox": [ + 113, + 219, + 497, + 231 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 113, + 219, + 497, + 231 + ], + "type": "text", + "content": "School of Electronic Information and Communications, Huazhong University of Science and Technology" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 124, + 235, + 482, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 235, + 482, + 245 + ], + "spans": [ + { + "bbox": [ + 124, + 235, + 482, + 245 + ], + "type": "text", + "content": "{u202215226,u202317280,u202317273,u202317218,yuchaojian,shaoyuanjie,cgao}@hust.edu.cn" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "spans": [ + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 298, + 297, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 298, + 297, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 298, + 297, + 598 + ], + "type": "text", + "content": "Adversarial attack plays a critical role in evaluating the robustness of deep learning models. Jacobian-based Saliency Map Attack (JSMA) is an interpretable adversarial method that offers excellent pixel-level control and provides valuable insights into model vulnerabilities. However, its quadratic computational complexity " + }, + { + "bbox": [ + 55, + 298, + 297, + 598 + ], + "type": "inline_equation", + "content": "O(M^2 \\times N)" + }, + { + "bbox": [ + 55, + 298, + 297, + 598 + ], + "type": "text", + "content": " renders it impractical for large-scale datasets, limiting its application despite its inherent value. This paper proposes FastJSMA, an efficient attack method that addresses these computational limitations. Our approach introduces a gradient decoupling mechanism that decomposes the Jacobian calculation into complementary class suppression (" + }, + { + "bbox": [ + 55, + 298, + 297, + 598 + ], + "type": "inline_equation", + "content": "g^{-}" + }, + { + "bbox": [ + 55, + 298, + 297, + 598 + ], + "type": "text", + "content": ") and class excitation (" + }, + { + "bbox": [ + 55, + 298, + 297, + 598 + ], + "type": "inline_equation", + "content": "g^{+}" + }, + { + "bbox": [ + 55, + 298, + 297, + 598 + ], + "type": "text", + "content": ") gradients, reducing complexity to " + }, + { + "bbox": [ + 55, + 298, + 297, + 598 + ], + "type": "inline_equation", + "content": "O(M\\sqrt{N})" + }, + { + "bbox": [ + 55, + 298, + 297, + 598 + ], + "type": "text", + "content": ". Additionally, we implement a class probing mechanism and an adaptive saliency threshold to further optimize the process. Experimental results across multiple datasets demonstrate that FastJSMA maintains comparable attack success rates while dramatically reducing computation time—requiring only 2.9% of JSMA's processing time on CIFAR-10 and 1.2% on CIFAR-100, and successfully operating on ImageNet where traditional JSMA fails due to memory constraints. This advancement enables the practical application of interpretable saliency map-based attacks on large-scale datasets, balancing effectiveness with computational efficiency." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 621, + 135, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 621, + 135, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 621, + 135, + 633 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 641, + 181, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 641, + 181, + 654 + ], + "spans": [ + { + "bbox": [ + 56, + 641, + 181, + 654 + ], + "type": "text", + "content": "1.1. Research Background" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 659, + 296, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 296, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 296, + 684 + ], + "type": "text", + "content": "Deep learning has achieved breakthrough progress in computer vision, natural language processing, and other fields," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 274, + 555, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 555, + 357 + ], + "type": "text", + "content": "but its security issues have increasingly attracted attention. In 2013, Szegedy et al. [1] first discovered the sensitivity of deep neural networks to small perturbations, i.e., by adding imperceptible perturbations to inputs, models can be induced to produce incorrect predictions. This discovery revealed potential vulnerabilities in deep learning systems and drove rapid development in adversarial attack research." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 362, + 556, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 362, + 556, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 362, + 556, + 423 + ], + "type": "text", + "content": "Adversarial attack techniques have evolved from simple approaches like the Fast Gradient Sign Method (FGSM) [2] to complex methods such as PGD [3] and C&W attacks [4], each offering different trade-offs between effectiveness, efficiency, and perturbation magnitude." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 445, + 434, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 445, + 434, + 458 + ], + "spans": [ + { + "bbox": [ + 314, + 445, + 434, + 458 + ], + "type": "text", + "content": "1.2. Research Motivation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 468, + 556, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 556, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 556, + 625 + ], + "type": "text", + "content": "The Jacobian-based Saliency Map Attack (JSMA) [5] is an adversarial method that identifies and modifies the most influential pixels through Jacobian matrix analysis. It offers unique advantages including intuitive visualization through saliency maps that enhance interpretability, precise control over modified pixels, and strong effectiveness across various datasets. Despite these merits and attempts to improve it through weighted approaches like WJSMA [6], JSMA remains largely underutilized due to prohibitive computational costs. Its full Jacobian matrix calculation requires " + }, + { + "bbox": [ + 313, + 468, + 556, + 625 + ], + "type": "inline_equation", + "content": "O(M^2 \\times N)" + }, + { + "bbox": [ + 313, + 468, + 556, + 625 + ], + "type": "text", + "content": " complexity (where " + }, + { + "bbox": [ + 313, + 468, + 556, + 625 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 468, + 556, + 625 + ], + "type": "text", + "content": " is the number of pixels and " + }, + { + "bbox": [ + 313, + 468, + 556, + 625 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 468, + 556, + 625 + ], + "type": "text", + "content": " is the number of classes), creating an insurmountable barrier for large-scale applications like ImageNet [7]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "This limitation stems from three factors: (a) separate backward propagation for each output node; (b) saliency evaluations across the entire Jacobian matrix; and (c) repeated calculations across iterations. Consequently, the adversarial machine learning community has effectively abandoned JSMA's valuable properties due to computational constraints rather than methodological limitations." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 137, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 137, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 137, + 703 + ], + "type": "text", + "content": "0* Equal contribution" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "type": "inline_equation", + "content": "^{0\\dagger}" + }, + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "type": "text", + "content": " Corresponding author" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1506" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 144, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 144, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 144, + 83 + ], + "type": "text", + "content": "1.3. Contributions" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 295, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 295, + 137 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 295, + 137 + ], + "type": "text", + "content": "To revitalize this valuable but computationally constrained method, we propose FastJSMA (Fast Jacobian-based Saliency Map Attack), an efficient adversarial attack based on gradient decoupling approximation:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 139, + 296, + 367 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 139, + 295, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 139, + 295, + 209 + ], + "spans": [ + { + "bbox": [ + 55, + 139, + 295, + 209 + ], + "type": "text", + "content": "- Gradient Decoupling Mechanism: We decompose the Jacobian matrix into complementary class suppression " + }, + { + "bbox": [ + 55, + 139, + 295, + 209 + ], + "type": "inline_equation", + "content": "(g^{-})" + }, + { + "bbox": [ + 55, + 139, + 295, + 209 + ], + "type": "text", + "content": " and excitation " + }, + { + "bbox": [ + 55, + 139, + 295, + 209 + ], + "type": "inline_equation", + "content": "(g^{+})" + }, + { + "bbox": [ + 55, + 139, + 295, + 209 + ], + "type": "text", + "content": " gradients, reducing computational complexity from " + }, + { + "bbox": [ + 55, + 139, + 295, + 209 + ], + "type": "inline_equation", + "content": "O(M^2 \\times N)" + }, + { + "bbox": [ + 55, + 139, + 295, + 209 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 139, + 295, + 209 + ], + "type": "inline_equation", + "content": "O(M\\sqrt{N})" + }, + { + "bbox": [ + 55, + 139, + 295, + 209 + ], + "type": "text", + "content": " while maintaining attack effectiveness through dynamic weight allocation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 211, + 296, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 211, + 296, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 296, + 270 + ], + "type": "text", + "content": "- Class Probing and Adaptive Threshold: We introduce a square-root sampling heuristic to efficiently identify vulnerable decision boundaries and an adaptive threshold mechanism to prioritize influential pixels, optimizing both efficiency and effectiveness." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 271, + 296, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 271, + 296, + 367 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 296, + 367 + ], + "type": "text", + "content": "- Comprehensive Evaluation: Our experiments demonstrate dramatic efficiency gains—requiring only " + }, + { + "bbox": [ + 55, + 271, + 296, + 367 + ], + "type": "inline_equation", + "content": "2.9\\%" + }, + { + "bbox": [ + 55, + 271, + 296, + 367 + ], + "type": "text", + "content": " of JSMANT's processing time on CIFAR-10 and achieving up to " + }, + { + "bbox": [ + 55, + 271, + 296, + 367 + ], + "type": "inline_equation", + "content": "81.4 \\times" + }, + { + "bbox": [ + 55, + 271, + 296, + 367 + ], + "type": "text", + "content": " speedup on CIFAR-100 while maintaining comparable attack success rates. Most notably, FastJSMA successfully processes ImageNet samples, where traditional JSMA and WJSMA fail entirely due to memory constraints." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 377, + 143, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 377, + 143, + 389 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 143, + 389 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 397, + 271, + 409 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 397, + 271, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 397, + 271, + 409 + ], + "type": "text", + "content": "2.1. Overview of Adversarial Attack Methods" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 415, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 295, + 521 + ], + "type": "text", + "content": "Adversarial attacks can be categorized as white-box or black-box attacks based on the attacker's knowledge of the model. White-box attacks assume the attacker has complete access to the model structure and parameters, with representative methods including FGSM [2] and PGD [3]. Black-box attacks can only access the model's input-output interface, primarily relying on transferability [8] or query optimization [9]. Recent work has also focused on improving adversarial training defenses [10-12]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 522, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 641 + ], + "type": "text", + "content": "From the perspective of attack objectives, they can be divided into targeted and non-targeted attacks. Targeted attacks aim to misclassify input samples as specified target classes, such as DeepFool [13]; non-targeted attacks only need to cause incorrect predictions, regardless of the specific error class, such as FGSM. These two types of attack methods have their respective advantages in practical applications: targeted attacks enable more precise control, while non-targeted attacks typically have higher success rates and lower computational costs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Gradient-based adversarial attacks have evolved significantly. After FGSM was proposed, researchers developed improved variants: (1) MI-FGSM [14] introduces momentum terms; (2) NI-FGSM [15] utilizes Nesterov accelerated gradients; (3) TI-FGSM [16] enhances transferability through translation invariance; (4) DI-FGSM [17] adopts" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": "diversified input strategies. Recent research by Mao et al. [18] discovered that adversarial attacks are reversible with natural supervision, providing new insights into the fundamental properties of adversarial perturbations. These methods optimize gradient-based attack strategies from different perspectives but have not fully addressed the issue of interpretability." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 163, + 463, + 175 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 163, + 463, + 175 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 463, + 175 + ], + "type": "text", + "content": "2.2. Detailed Analysis of JSMA" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 180, + 555, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 555, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 555, + 287 + ], + "type": "text", + "content": "As an interpretable attack method, JSMA was proposed by Papernot et al. [5], with its core idea being the use of the Jacobian matrix to identify input features that have the greatest impact on model output. This approach builds upon fundamental research in saliency detection, such as the boolean map approach [19], which provides theoretical foundations for identifying visually significant regions. Specifically, for input " + }, + { + "bbox": [ + 313, + 180, + 555, + 287 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 180, + 555, + 287 + ], + "type": "text", + "content": " and target class " + }, + { + "bbox": [ + 313, + 180, + 555, + 287 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 180, + 555, + 287 + ], + "type": "text", + "content": ", JSMA first calculates the Jacobian matrix:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 365, + 295, + 554, + 324 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 295, + 554, + 324 + ], + "spans": [ + { + "bbox": [ + 365, + 295, + 554, + 324 + ], + "type": "interline_equation", + "content": "J _ {F} (x) = \\frac {\\partial F (x)}{\\partial x} = \\left[ \\frac {\\partial F _ {j} (x)}{\\partial x _ {i}} \\right] _ {i, j} \\tag {1}", + "image_path": "2758785e585e980e6dedc38f3c72271d7e866cbc87475449f88cb869090fcced.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 327, + 554, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 554, + 364 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 554, + 364 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 327, + 554, + 364 + ], + "type": "inline_equation", + "content": "F_{j}(x)" + }, + { + "bbox": [ + 313, + 327, + 554, + 364 + ], + "type": "text", + "content": " represents the model's prediction probability for class " + }, + { + "bbox": [ + 313, + 327, + 554, + 364 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 327, + 554, + 364 + ], + "type": "text", + "content": ". Based on the Jacobian matrix, JSMA defines the saliency map " + }, + { + "bbox": [ + 313, + 327, + 554, + 364 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 327, + 554, + 364 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 323, + 369, + 553, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 369, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 323, + 369, + 553, + 415 + ], + "type": "interline_equation", + "content": "S (x, t) [ i ] = \\left\\{ \\begin{array}{l} 0, \\text {i f} \\frac {\\partial F _ {t} (x)}{\\partial x _ {i}} < 0 \\text {o r} \\sum_ {j \\neq t} \\frac {\\partial F _ {j} (x)}{\\partial x _ {i}} > 0 \\\\ \\left| \\frac {\\partial F _ {t} (x)}{\\partial x _ {i}} \\right| \\cdot \\left| \\sum_ {j \\neq t} \\frac {\\partial F _ {j} (x)}{\\partial x _ {i}} \\right|, \\text {o t h e r w i s e} \\end{array} \\right. \\tag {2}", + "image_path": "e1f357ef99d167510f4eca4b0b378c83bd9b197ea4c5a19ae0b99725449cdc45.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 415, + 554, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 554, + 533 + ], + "type": "text", + "content": "Subsequently, researchers have made various improvements to JSMA: (1) JSMA++ [8] enhances attack efficiency by optimizing saliency calculations; (2) One-pixel JSMA [20] restricts modifications to a single pixel, achieving more precise control; (3) Adaptive JSMA [21] introduces adaptive threshold strategies, improving attack success rates. The concept of saliency maps has also been extended to other domains, such as 3D point clouds [22], demonstrating the versatility of saliency-based approaches across different data modalities." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 534, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 714 + ], + "type": "text", + "content": "An important advancement is the Weighted JSMA (WJSMA) [6], which addresses limitations in the original JSMA's saliency calculation. WJSMA penalizes gradients associated with small probabilities to create more balanced saliency maps. WJSMA incorporates class probabilities as weights in the saliency calculation, setting the saliency to zero when specific gradient conditions are not met. This weighting mechanism makes WJSMA faster than the original JSMA while maintaining good balance between attack success rate and computational cost. However, even with these improvements, WJSMA still inherits the fundamental computational inefficiency of JSMA for large-scale datasets, as it retains the " + }, + { + "bbox": [ + 313, + 534, + 555, + 714 + ], + "type": "inline_equation", + "content": "O(M^2 \\times N)" + }, + { + "bbox": [ + 313, + 534, + 555, + 714 + ], + "type": "text", + "content": " complexity that creates memory and processing bottlenecks on datasets like ImageNet." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1507" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 234, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 234, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 234, + 85 + ], + "type": "text", + "content": "2.3. Related Optimization Techniques" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 175 + ], + "type": "text", + "content": "To address the efficiency problems of adversarial attacks in large-scale scenarios, researchers have proposed various optimization strategies. One class of methods improves efficiency by enhancing gradient computation, such as OnePixel Attack [20], which reduces search space through evolutionary algorithms, and SparseFool [23], which optimizes computational efficiency by leveraging sparsity." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 175, + 296, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 175, + 296, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 296, + 236 + ], + "type": "text", + "content": "Another class of methods proposes specialized optimization strategies for large-scale datasets. For example, Fast Gradient Method [24] enhances attack efficiency through ensemble techniques, while AutoAttack [25] adopts automated strategies to select optimal attack parameters." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 236, + 296, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 236, + 296, + 320 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 296, + 320 + ], + "type": "text", + "content": "Recent research has also explored algorithmic optimizations specifically targeting the computational complexity of gradient-based attacks. These approaches include gradient approximation techniques [26], adaptive norm constraints for efficient minimum-norm attacks [27], and decomposition strategies that reduce the dimensionality of the optimization space [21]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 321, + 296, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 321, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 321, + 296, + 464 + ], + "type": "text", + "content": "These optimization methods have their respective advantages and disadvantages: gradient-based methods offer high computational efficiency but lower interpretability; optimization-based methods provide good results but incur high computational costs; heuristic-based methods are simple and fast but have unstable success rates. In comparison, our proposed FastJSMA maintains JSMA's interpretability while significantly improving computational efficiency through dual-gradient approximation, effectively reducing the complexity from " + }, + { + "bbox": [ + 55, + 321, + 296, + 464 + ], + "type": "inline_equation", + "content": "O(M^2 \\times N)" + }, + { + "bbox": [ + 55, + 321, + 296, + 464 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 321, + 296, + 464 + ], + "type": "inline_equation", + "content": "O(M\\sqrt{N})" + }, + { + "bbox": [ + 55, + 321, + 296, + 464 + ], + "type": "text", + "content": " and making saliency map-based attacks practical for large-scale datasets for the first time." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 479, + 138, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 138, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 138, + 493 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 499, + 296, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 499, + 296, + 584 + ], + "spans": [ + { + "bbox": [ + 55, + 499, + 296, + 584 + ], + "type": "text", + "content": "This section details the technical principles of FastJSMA. We first establish the mathematical model of the problem, then elaborate on the design rationale of the dual-gradient approximation strategy, followed by a detailed explanation of the saliency map generation process, and finally theoretically analyze the computational complexity of the algorithm." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 594, + 272, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 272, + 608 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 272, + 608 + ], + "type": "text", + "content": "3.1. Problem Modeling and Attack Objectives" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 612, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 296, + 685 + ], + "type": "text", + "content": "For an input sample " + }, + { + "bbox": [ + 55, + 612, + 296, + 685 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 55, + 612, + 296, + 685 + ], + "type": "text", + "content": " and a target classification model " + }, + { + "bbox": [ + 55, + 612, + 296, + 685 + ], + "type": "inline_equation", + "content": "f: \\mathbb{R}^{C \\times H \\times W} \\to \\mathbb{R}^N" + }, + { + "bbox": [ + 55, + 612, + 296, + 685 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 612, + 296, + 685 + ], + "type": "inline_equation", + "content": "C, H, W" + }, + { + "bbox": [ + 55, + 612, + 296, + 685 + ], + "type": "text", + "content": " represent the number of channels, height, and width respectively, and " + }, + { + "bbox": [ + 55, + 612, + 296, + 685 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 612, + 296, + 685 + ], + "type": "text", + "content": " is the total number of classes. Given the original predicted class " + }, + { + "bbox": [ + 55, + 612, + 296, + 685 + ], + "type": "inline_equation", + "content": "t = \\arg \\max f(x)" + }, + { + "bbox": [ + 55, + 612, + 296, + 685 + ], + "type": "text", + "content": ", the goal of the adversarial attack is to find a perturbation " + }, + { + "bbox": [ + 55, + 612, + 296, + 685 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 612, + 296, + 685 + ], + "type": "text", + "content": " such that:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 94, + 700, + 296, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 700, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 94, + 700, + 296, + 715 + ], + "type": "interline_equation", + "content": "\\arg \\max f (x + \\delta) \\neq t \\quad \\text {s . t .} \\quad \\| \\delta \\| _ {0} \\leq \\epsilon \\tag {3}", + "image_path": "68d69e92355d295f3068d86ee135fac4a7295da86ccf726bb231d6d959697c4e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": " represents the allowed number of pixels to modify, and " + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_0" + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "inline_equation", + "content": "L_0" + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": " norm. The reasons for choosing the " + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "inline_equation", + "content": "L_0" + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": " norm constraint rather than other norms (such as " + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "inline_equation", + "content": "L_2" + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "inline_equation", + "content": "\\mathrm{L\\infty}" + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": ") are: (1) " + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "inline_equation", + "content": "L_0" + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": " norm directly limits the number of modified pixels, conforming to the principle of minimal intervention; (2) a limited number of modifications makes the attack process more interpretable and understandable; (3) maintains consistent constraint forms with traditional JSMA for fair comparison." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 181, + 553, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 181, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 181, + 553, + 263 + ], + "type": "text", + "content": "Traditional JSMA computes the complete Jacobian matrix " + }, + { + "bbox": [ + 313, + 181, + 553, + 263 + ], + "type": "inline_equation", + "content": "J = \\nabla_{x}f(x)" + }, + { + "bbox": [ + 313, + 181, + 553, + 263 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 181, + 553, + 263 + ], + "type": "inline_equation", + "content": "O(M^2\\times N)" + }, + { + "bbox": [ + 313, + 181, + 553, + 263 + ], + "type": "text", + "content": " complexity, where " + }, + { + "bbox": [ + 313, + 181, + 553, + 263 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 181, + 553, + 263 + ], + "type": "text", + "content": " represents the total number of pixels and " + }, + { + "bbox": [ + 313, + 181, + 553, + 263 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 181, + 553, + 263 + ], + "type": "text", + "content": " denotes the number of classification classes. This substantial computational burden and memory requirements render JSMA impractical for large-scale datasets and real-world applications." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 266, + 553, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 553, + 301 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 553, + 301 + ], + "type": "text", + "content": "FastJSMA addresses this limitation through a dual-gradient approximation strategy that reduces complexity to " + }, + { + "bbox": [ + 313, + 266, + 553, + 301 + ], + "type": "inline_equation", + "content": "O(M\\sqrt{N})" + }, + { + "bbox": [ + 313, + 266, + 553, + 301 + ], + "type": "text", + "content": " while maintaining attack effectiveness." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 312, + 476, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 476, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 476, + 324 + ], + "type": "text", + "content": "3.2. Target Function Construction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 330, + 553, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 330, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 553, + 426 + ], + "type": "text", + "content": "The core mathematical insight that enables FastJSMA's efficiency breakthrough lies in our recognition that the Jacobian-based saliency computation in traditional JSMA can be decomposed into two complementary objectives: first, decreasing the probability of the current class, and second, increasing the probability of alternative classes. This decomposition forms the theoretical foundation of our dual-gradient approximation strategy." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 427, + 554, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 427, + 554, + 547 + ], + "spans": [ + { + "bbox": [ + 313, + 427, + 554, + 547 + ], + "type": "text", + "content": "Analyzing the mathematical structure of traditional JSMA reveals that its saliency map computation essentially measures two gradient components for each pixel: how modifying that pixel decreases the current class score and how it increases other class scores. While traditional JSMA explicitly computes these gradients through the full Jacobian matrix, we observed that these two objectives could be approximated through carefully designed loss functions without requiring the prohibitively expensive Jacobian computation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 548, + 554, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 554, + 620 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 554, + 620 + ], + "type": "text", + "content": "To avoid computing the complete Jacobian matrix, we propose a dual-gradient approximation strategy. The core idea of this strategy is to construct two complementary objective functions that simulate the enhancement effect on the target class and the suppression effect on other classes in JSMA. The specific definitions are as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 624, + 554, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 624, + 554, + 647 + ], + "spans": [ + { + "bbox": [ + 314, + 624, + 554, + 647 + ], + "type": "text", + "content": "1. Class Suppression Gradient: Construct a virtual label " + }, + { + "bbox": [ + 314, + 624, + 554, + 647 + ], + "type": "inline_equation", + "content": "\\hat{y}^{-} \\in \\mathbb{R}^{N}" + }, + { + "bbox": [ + 314, + 624, + 554, + 647 + ], + "type": "text", + "content": ", where" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 326, + 658, + 558, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 658, + 558, + 713 + ], + "spans": [ + { + "bbox": [ + 326, + 658, + 558, + 713 + ], + "type": "interline_equation", + "content": "\\hat {y} _ {i} ^ {-} = \\left\\{ \\begin{array}{l l} f (x) _ {i} - 1 & i = t \\\\ f (x) _ {i} + \\sqrt {N} & i = a (\\text {a t t e m p t c l a s s , i f s p e c i f i e d}) \\\\ f (x) _ {i} & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {4}", + "image_path": "aa7a2d77568dde6880a70c3578cd54506c649b9796b1d87a5fb4e437b36a1477.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1508" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 70, + 534, + 448 + ], + "blocks": [ + { + "bbox": [ + 80, + 70, + 534, + 448 + ], + "lines": [ + { + "bbox": [ + 80, + 70, + 534, + 448 + ], + "spans": [ + { + "bbox": [ + 80, + 70, + 534, + 448 + ], + "type": "image", + "image_path": "e8442bc3756db7a5fe074877cc37b4f65160e8e9bd2315d3cd688ade07e7cf9e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 455, + 555, + 534 + ], + "lines": [ + { + "bbox": [ + 55, + 455, + 555, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 555, + 534 + ], + "type": "text", + "content": "Figure 1. Architectural comparison between traditional JSMA and our proposed FastJSMA. Left: Traditional JSMA calculates the full Jacobian matrix with computational complexity " + }, + { + "bbox": [ + 55, + 455, + 555, + 534 + ], + "type": "inline_equation", + "content": "O(M^2 \\times N)" + }, + { + "bbox": [ + 55, + 455, + 555, + 534 + ], + "type": "text", + "content": " and relies on an iterative process to achieve misclassification, making it prohibitively expensive for large-scale datasets. Right: FastJSMA employs a class probing mechanism to select " + }, + { + "bbox": [ + 55, + 455, + 555, + 534 + ], + "type": "inline_equation", + "content": "\\sqrt{N}" + }, + { + "bbox": [ + 55, + 455, + 555, + 534 + ], + "type": "text", + "content": " candidate classes and a gradient decoupling strategy that approximates the saliency computation through class suppression gradient " + }, + { + "bbox": [ + 55, + 455, + 555, + 534 + ], + "type": "inline_equation", + "content": "(g^-)" + }, + { + "bbox": [ + 55, + 455, + 555, + 534 + ], + "type": "text", + "content": " and class excitation gradient " + }, + { + "bbox": [ + 55, + 455, + 555, + 534 + ], + "type": "inline_equation", + "content": "(g^+)" + }, + { + "bbox": [ + 55, + 455, + 555, + 534 + ], + "type": "text", + "content": ". The class suppression gradient reduces the current class probability while boosting an attempt class by " + }, + { + "bbox": [ + 55, + 455, + 555, + 534 + ], + "type": "inline_equation", + "content": "\\sqrt{N}" + }, + { + "bbox": [ + 55, + 455, + 555, + 534 + ], + "type": "text", + "content": ", and the significance threshold efficiently filters the most influential pixels. This approach reduces complexity to " + }, + { + "bbox": [ + 55, + 455, + 555, + 534 + ], + "type": "inline_equation", + "content": "O(M\\sqrt{N})" + }, + { + "bbox": [ + 55, + 455, + 555, + 534 + ], + "type": "text", + "content": " while maintaining comparable effectiveness." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 554, + 162, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 554, + 162, + 566 + ], + "spans": [ + { + "bbox": [ + 67, + 554, + 162, + 566 + ], + "type": "text", + "content": "Calculate the gradient:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 135, + 577, + 295, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 577, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 135, + 577, + 295, + 590 + ], + "type": "interline_equation", + "content": "g ^ {-} = \\nabla_ {x} \\mathcal {L} (f (x), \\hat {y} ^ {-}) \\tag {5}", + "image_path": "b85313f1e2118660188b967054e3c075c3fcfc73d6bb22398e74e47df212320c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 601, + 295, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 601, + 295, + 649 + ], + "spans": [ + { + "bbox": [ + 67, + 601, + 295, + 649 + ], + "type": "text", + "content": "This gradient reflects the direction of reducing the prediction probability of the current class and optionally enhancing a specific alternative class if an attempt parameter is provided." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 650, + 295, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 650, + 295, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 650, + 295, + 673 + ], + "type": "text", + "content": "2. Class Excitation Gradient: Construct an enhancement label " + }, + { + "bbox": [ + 55, + 650, + 295, + 673 + ], + "type": "inline_equation", + "content": "\\hat{y}^{+}\\in \\mathbb{R}^{N}" + }, + { + "bbox": [ + 55, + 650, + 295, + 673 + ], + "type": "text", + "content": ", where" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 117, + 684, + 295, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 684, + 295, + 716 + ], + "spans": [ + { + "bbox": [ + 117, + 684, + 295, + 716 + ], + "type": "interline_equation", + "content": "\\hat {y} _ {i} ^ {+} = \\left\\{ \\begin{array}{l l} f (x) _ {i} - 1 & i = t \\\\ f (x) _ {i} + 1 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {6}", + "image_path": "55dc7712173f02e534fd26dc03272e67c7ff454c10c6c41ed8a5b10f069b2a09.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 326, + 554, + 421, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 554, + 421, + 566 + ], + "spans": [ + { + "bbox": [ + 326, + 554, + 421, + 566 + ], + "type": "text", + "content": "Calculate the gradient:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 394, + 572, + 553, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 572, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 394, + 572, + 553, + 586 + ], + "type": "interline_equation", + "content": "g ^ {+} = \\nabla_ {x} \\mathcal {L} (f (x), \\hat {y} ^ {+}) \\tag {7}", + "image_path": "aa68ed319115a5d7e2f68bdc6a88d7a945304eac265cc283db7ada16f56c3cd4.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 326, + 592, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 592, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 326, + 592, + 555, + 628 + ], + "type": "text", + "content": "This gradient indicates the direction of enhancing the prediction probability of other classes while suppressing the current class." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "type": "text", + "content": "This formulation represents a significant mathematical innovation over traditional JSMA. Rather than explicitly computing " + }, + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "type": "inline_equation", + "content": "N \\times M" + }, + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "type": "text", + "content": " partial derivatives as required by the Jacobian matrix, our approach leverages automatic differentiation to compute just two gradient vectors—reducing the computational complexity by a factor proportional to the number of classes " + }, + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "type": "text", + "content": " and the number of pixels " + }, + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "type": "text", + "content": ". Import-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1509" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "tantly, this efficiency gain comes without compromising the underlying attack methodology; the dual-gradient approach preserves the core mathematical principles of JSMA while dramatically reducing its computational footprint." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 120, + 295, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 120, + 295, + 178 + ], + "spans": [ + { + "bbox": [ + 55, + 120, + 295, + 178 + ], + "type": "text", + "content": "The design of these two gradients cleverly avoids directly computing the Jacobian matrix while retaining the core idea of JSMA: simultaneously considering the suppression of the target class and the excitation of other classes." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 187, + 198, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 187, + 198, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 187, + 198, + 201 + ], + "type": "text", + "content": "3.3. Saliency Map Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 205, + 295, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 295, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 295, + 288 + ], + "type": "text", + "content": "Traditional JSMA generates saliency maps through exhaustive computation of the Jacobian matrix followed by pairwise evaluation of pixel modifications—a process that scales quadratically with the number of pixels. FastJSMA fundamentally reconceptualizes this approach through an elegant mathematical transformation that yields equivalent results with substantially reduced computation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 289, + 295, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 295, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 295, + 324 + ], + "type": "text", + "content": "Based on the dual-gradient approximation, we apply gradient normalization to ensure balanced contribution while strategically emphasizing the class suppression component:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 113, + 334, + 295, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 334, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 113, + 334, + 295, + 360 + ], + "type": "interline_equation", + "content": "g _ {\\text {n o r m a l i z e d}} ^ {-} = \\frac {g ^ {-}}{\\operatorname {m e a n} (| g ^ {-} |)} \\times N \\tag {8}", + "image_path": "563ea5c119d7a4cfcc19669e8e8c6075b9c377545acdaa36ac4e96df34889fdf.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 124, + 367, + 294, + 393 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 367, + 294, + 393 + ], + "spans": [ + { + "bbox": [ + 124, + 367, + 294, + 393 + ], + "type": "interline_equation", + "content": "g _ {\\text {n o r m a l i z e d}} ^ {+} = \\frac {g ^ {+}}{\\operatorname {m e a n} (| g ^ {+} |)} \\tag {9}", + "image_path": "8b2bc051278e52a5c460215a13a84a728d3f399eb9c19596230f8a9b25ba9d43.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 396, + 295, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 295, + 479 + ], + "type": "text", + "content": "This normalization procedure includes a deliberate engineering optimization: the class suppression gradient is scaled by a factor of " + }, + { + "bbox": [ + 55, + 396, + 295, + 479 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 396, + 295, + 479 + ], + "type": "text", + "content": " (the number of classes) to amplify its influence in the overall saliency calculation. Our empirical investigation indicates that this scaling enhances attack effectiveness by emphasizing the suppression of the original class prediction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 480, + 295, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 295, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 295, + 504 + ], + "type": "text", + "content": "We then integrate these normalized gradients to obtain a comprehensive gain matrix:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 120, + 515, + 294, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 515, + 294, + 529 + ], + "spans": [ + { + "bbox": [ + 120, + 515, + 294, + 529 + ], + "type": "interline_equation", + "content": "G = g _ {\\text {n o r m a l i z e d}} ^ {+} + g _ {\\text {n o r m a l i z e d}} ^ {-} \\tag {10}", + "image_path": "5dcc03783098845f56067987df5b1a082b7c7bc73abc9a46844880c548a405a6.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 534, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 616 + ], + "type": "text", + "content": "This matrix captures the bidirectional influence of modifying each pixel on both decreasing the target class probability and increasing alternative class probabilities—precisely the information encapsulated in traditional JSMA's saliency computation, but obtained through a mathematically equivalent yet computationally superior formulation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "content": "After obtaining the gain matrix, we apply a significance threshold to filter out pixels with minimal impact. This threshold operation is visually represented as a funnel-like process in Figure 1, focusing computational resources only on the most influential pixels. By implementing this significance threshold, we effectively reduce the search space while ensuring that the selected pixels have meaningful contributions to the adversarial perturbation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "This selective modification approach, governed by both the saliency ranking and the significance threshold, ensures efficient resource allocation by concentrating perturbations on the most influential and impactful pixels." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 127, + 553, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 127, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 313, + 127, + 553, + 152 + ], + "type": "text", + "content": "3.4. Attack Success Rate Optimization via Class Probing" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 156, + 553, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 553, + 275 + ], + "type": "text", + "content": "A significant enhancement in our implementation is the introduction of a class-specific probing mechanism, visually illustrated in Figure 1 as \"Class Probing Mechanism\". This component substantially improves attack success rates while maintaining the untargeted nature of FastJSMA by intelligently selecting candidate classes for the attempt parameter. The optimization addresses a fundamental challenge in untargeted adversarial attacks: determining which gradient directions most efficiently induce misclassification." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 276, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 276, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 553, + 312 + ], + "type": "text", + "content": "Our implementation employs an intelligent candidate class exploration strategy governed by a square-root sampling heuristic:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 385, + 323, + 553, + 337 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 323, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 385, + 323, + 553, + 337 + ], + "type": "interline_equation", + "content": "\\text {n u m} _ {\\text {c a n d i d a t e s}} = \\sqrt {N} \\tag {11}", + "image_path": "57b6c4cd01605c782ce4dbfbd6ac75c2e7323b1a87c59469c600e912992863f6.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 342, + 553, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 553, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 553, + 413 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 342, + 553, + 413 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 342, + 553, + 413 + ], + "type": "text", + "content": " represents the number of classes in the classification problem. This square-root relationship establishes a principled balance between computational efficiency and exploration thoroughness—scaling sublinearly with the class space dimensionality to maintain tractability even for large-scale classification problems." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 414, + 553, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 553, + 449 + ], + "type": "text", + "content": "For each candidate class " + }, + { + "bbox": [ + 313, + 414, + 553, + 449 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 414, + 553, + 449 + ], + "type": "text", + "content": " among the top numCandidates predicted classes (excluding the original class), the algorithm:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 449, + 553, + 569 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 313, + 449, + 553, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 553, + 521 + ], + "type": "text", + "content": "1. Applies FastJSMA with class " + }, + { + "bbox": [ + 313, + 449, + 553, + 521 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 449, + 553, + 521 + ], + "type": "text", + "content": " as an \"attempt\" parameter, which specifically modifies the class suppression gradient " + }, + { + "bbox": [ + 313, + 449, + 553, + 521 + ], + "type": "inline_equation", + "content": "(g^{-})" + }, + { + "bbox": [ + 313, + 449, + 553, + 521 + ], + "type": "text", + "content": " in two ways: - Decreases the current class probability (by setting " + }, + { + "bbox": [ + 313, + 449, + 553, + 521 + ], + "type": "inline_equation", + "content": "\\hat{y}_i^- = f(x)_i - 1" + }, + { + "bbox": [ + 313, + 449, + 553, + 521 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 449, + 553, + 521 + ], + "type": "inline_equation", + "content": "i = t" + }, + { + "bbox": [ + 313, + 449, + 553, + 521 + ], + "type": "text", + "content": ") - Boosts the attempt class " + }, + { + "bbox": [ + 313, + 449, + 553, + 521 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 449, + 553, + 521 + ], + "type": "text", + "content": " by a factor of " + }, + { + "bbox": [ + 313, + 449, + 553, + 521 + ], + "type": "inline_equation", + "content": "\\sqrt{N}" + }, + { + "bbox": [ + 313, + 449, + 553, + 521 + ], + "type": "text", + "content": " (by setting " + }, + { + "bbox": [ + 313, + 449, + 553, + 521 + ], + "type": "inline_equation", + "content": "\\hat{y}_i^- = f(x)_i + \\sqrt{N}" + }, + { + "bbox": [ + 313, + 449, + 553, + 521 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 449, + 553, + 521 + ], + "type": "inline_equation", + "content": "i = c" + }, + { + "bbox": [ + 313, + 449, + 553, + 521 + ], + "type": "text", + "content": ")" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 522, + 553, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 553, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 553, + 545 + ], + "type": "text", + "content": "2. Evaluates whether this modified gradient direction successfully induces misclassification" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 546, + 553, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 553, + 569 + ], + "type": "text", + "content": "3. Returns the first successful perturbation, terminating further exploration" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 570, + 553, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 553, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 553, + 640 + ], + "type": "text", + "content": "This dual action of decreasing the current class while simultaneously boosting a specific attempt class efficiently guides the gradient toward vulnerable regions of the decision boundary, substantially improving attack success rates without requiring exhaustive exploration of all possible class transitions." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "content": "It is crucial to emphasize that this approach does not constitute a targeted attack in the conventional sense. Rather, the attempt parameter merely directs the gradient toward regions where transitions away from the original class are most probable, while maintaining an untargeted classification objective. This gradient steering mechanism represents" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1510" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "content": "a methodological innovation particularly effective in high-dimensional classification spaces where traditional methods might struggle to efficiently locate vulnerable decision boundaries." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 120, + 294, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 120, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 120, + 294, + 204 + ], + "type": "text", + "content": "Only if all candidate-based attempts fail does the algorithm revert to the standard untargeted FastJSMA approach without class-specific gradient modification. This hierarchical strategy enhances attack success rates while maintaining the untargeted objective—an engineering optimization that improves practical performance without altering the fundamental attack paradigm." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 211, + 175, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 211, + 175, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 175, + 224 + ], + "type": "text", + "content": "3.5. Complexity Analysis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 228, + 295, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 295, + 382 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 295, + 382 + ], + "type": "text", + "content": "To theoretically analyze algorithm efficiency, we compare the computational complexity of FastJSMA with traditional JSMA. Assuming a dataset with " + }, + { + "bbox": [ + 55, + 228, + 295, + 382 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 228, + 295, + 382 + ], + "type": "text", + "content": " classes, where each image has " + }, + { + "bbox": [ + 55, + 228, + 295, + 382 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 228, + 295, + 382 + ], + "type": "text", + "content": " pixels. The computational complexity of traditional JSMA is " + }, + { + "bbox": [ + 55, + 228, + 295, + 382 + ], + "type": "inline_equation", + "content": "O(M^2 \\times N)" + }, + { + "bbox": [ + 55, + 228, + 295, + 382 + ], + "type": "text", + "content": ". This is because, when constructing the saliency map, it needs to calculate the joint saliency of each pixel pair, considering " + }, + { + "bbox": [ + 55, + 228, + 295, + 382 + ], + "type": "inline_equation", + "content": "M^2" + }, + { + "bbox": [ + 55, + 228, + 295, + 382 + ], + "type": "text", + "content": " order of magnitude pixel pairs; the saliency calculation for each pixel pair involves " + }, + { + "bbox": [ + 55, + 228, + 295, + 382 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 228, + 295, + 382 + ], + "type": "text", + "content": " classes. Moreover, as illustrated in Figure 1, traditional JSMA typically requires multiple iterations of this expensive computation process to achieve successful misclassification, further compounding the computational burden." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 384, + 295, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 295, + 456 + ], + "type": "text", + "content": "In contrast, our FastJSMA has a computational complexity of " + }, + { + "bbox": [ + 55, + 384, + 295, + 456 + ], + "type": "inline_equation", + "content": "O(M\\sqrt{N})" + }, + { + "bbox": [ + 55, + 384, + 295, + 456 + ], + "type": "text", + "content": ". This is because we only perform 2 gradient calculations. The combination of gradient decoupling, class probing mechanism, and significance threshold filtering together dramatically reduces the computational requirements while maintaining attack effectiveness." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 456, + 295, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 295, + 527 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 295, + 527 + ], + "type": "text", + "content": "This significant reduction in complexity enables FastJSMA to efficiently process large-scale datasets. Due to the dual-gradient approximation strategy retaining JSMA's core ideas, the attack effectiveness is preserved despite substantial complexity reduction, as verified in our experiments." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 528, + 295, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 295, + 624 + ], + "type": "text", + "content": "Our additional optimizations—including gradient normalization with class-specific scaling, dynamic significance thresholds, and the square-root probing strategy—further enhance efficiency without compromising mathematical foundations. These refinements demonstrate the importance of implementation details in translating theoretical innovations into practical adversarial methodologies that scale to contemporary deep learning applications." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 634, + 135, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 634, + 135, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 634, + 135, + 647 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "content": "We present a comprehensive evaluation of FastJSMA, examining its effectiveness, efficiency, and transferability across multiple datasets and model architectures. Our experiments are designed to answer three key questions: (1) How does FastJSMA compare to traditional JSMA and its" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "variants in terms of attack success and computational efficiency? (2) How do different components of FastJSMA contribute to its overall performance? (3) How sensitive is FastJSMA to key parameter settings?" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 129, + 430, + 142 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 129, + 430, + 142 + ], + "spans": [ + { + "bbox": [ + 313, + 129, + 430, + 142 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 147, + 469, + 159 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 147, + 469, + 159 + ], + "spans": [ + { + "bbox": [ + 313, + 147, + 469, + 159 + ], + "type": "text", + "content": "4.1.1. Implementation Environment" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 162, + 553, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 162, + 553, + 222 + ], + "spans": [ + { + "bbox": [ + 313, + 162, + 553, + 222 + ], + "type": "text", + "content": "All experiments were conducted on a single NVIDIA RTX 3090 GPU with 24GB memory without virtual memory allocation. We implemented all methods using PyTorch framework and will release our code upon publication to ensure reproducibility." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 230, + 430, + 241 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 230, + 430, + 241 + ], + "spans": [ + { + "bbox": [ + 313, + 230, + 430, + 241 + ], + "type": "text", + "content": "4.1.2. Datasets and Models" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 245, + 553, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 245, + 553, + 328 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 553, + 328 + ], + "type": "text", + "content": "To evaluate our method across varying scales of complexity, we selected three representative datasets: CIFAR10/100 [28] (natural color images sized " + }, + { + "bbox": [ + 313, + 245, + 553, + 328 + ], + "type": "inline_equation", + "content": "32 \\times 32 \\times 3" + }, + { + "bbox": [ + 313, + 245, + 553, + 328 + ], + "type": "text", + "content": ", containing 10 and 100 classes respectively, each with 10,000 test samples) and ImageNet [7] (large-scale classification dataset with 1,000 classes, sized " + }, + { + "bbox": [ + 313, + 245, + 553, + 328 + ], + "type": "inline_equation", + "content": "224 \\times 224 \\times 3" + }, + { + "bbox": [ + 313, + 245, + 553, + 328 + ], + "type": "text", + "content": ", with 50,000 validation samples)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 329, + 553, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 553, + 399 + ], + "type": "text", + "content": "For model architectures, we employed widely-used ResNet variants (primarily ResNet18) [29], using pretrained checkpoints from HuggingFace1. These models achieve competitive performance on their respective datasets, providing a reliable foundation for our adversarial attack evaluation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 408, + 493, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 493, + 420 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 493, + 420 + ], + "type": "text", + "content": "4.1.3. Attack Configuration and Baselines" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 423, + 536, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 536, + 435 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 536, + 435 + ], + "type": "text", + "content": "We compare FastJSMA with several baseline methods:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 438, + 553, + 533 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 314, + 438, + 553, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 438, + 553, + 473 + ], + "spans": [ + { + "bbox": [ + 314, + 438, + 553, + 473 + ], + "type": "text", + "content": "- JSMANT: The traditional Jacobian-based Saliency Map Attack [5] in its non-targeted form, which computes the full Jacobian matrix." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 475, + 553, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 475, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 314, + 475, + 553, + 509 + ], + "type": "text", + "content": "- WJSMA: The Weighted JSMA variant proposed by Combey et al. [6], which introduces probabilistic weights to enhance attack effectiveness." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 510, + 553, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 510, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 314, + 510, + 553, + 533 + ], + "type": "text", + "content": "- Random: A control baseline that randomly selects pixels for modification." + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 537, + 553, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 537, + 553, + 619 + ], + "spans": [ + { + "bbox": [ + 313, + 537, + 553, + 619 + ], + "type": "text", + "content": "To ensure fair comparison, we maintain consistent attack parameters across all methods. Since different datasets have different image sizes, we use a proportional approach to pixel modification: 8 pixels for " + }, + { + "bbox": [ + 313, + 537, + 553, + 619 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 313, + 537, + 553, + 619 + ], + "type": "text", + "content": " images (CIFAR-10/100) and 196 pixels for " + }, + { + "bbox": [ + 313, + 537, + 553, + 619 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 313, + 537, + 553, + 619 + ], + "type": "text", + "content": " images (ImageNet), representing approximately " + }, + { + "bbox": [ + 313, + 537, + 553, + 619 + ], + "type": "inline_equation", + "content": "0.78\\%" + }, + { + "bbox": [ + 313, + 537, + 553, + 619 + ], + "type": "text", + "content": " of the total pixels in each case." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 621, + 553, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 621, + 553, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 621, + 553, + 693 + ], + "type": "text", + "content": "All experiments for CIFAR-10/100 were conducted on their complete test sets (10,000 samples each). For ImageNet, we used the full ImageNet1K validation set (50,000 samples across 1,000 classes), where \"1K\" refers to the number of classes rather than sample count. Attack Success Rate (ASR) was calculated as the percentage of sam" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 702, + 489, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 702, + 489, + 712 + ], + "spans": [ + { + "bbox": [ + 325, + 702, + 489, + 712 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1511" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 102, + 292, + 222 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 297, + 94 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 297, + 94 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 297, + 94 + ], + "type": "text", + "content": "Table 1. Performance comparison between FastJSMA and JSMA variants with ResNet18" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 102, + 292, + 222 + ], + "lines": [ + { + "bbox": [ + 61, + 102, + 292, + 222 + ], + "spans": [ + { + "bbox": [ + 61, + 102, + 292, + 222 + ], + "type": "table", + "html": "
Model/DatasetMethodASR(%)Time(μs)L2Norm
CIFAR-10Random7.5531.42
FastJSMA43.983391.53
JSMANT43.7911,7691.57
WJSMA51.568,6411.29
CIFAR-100Random15.3131.44
FastJSMA45.307511.75
JSMANT43.8761,1661.55
WJSMA49.7851,0651.24
ImageNetRandom2.9787.78
FastJSMA54.897,9089.19
", + "image_path": "9dada6ac53716799aa6099470b6d5ca3758d3824b23d28d8dc5e81e9a64e2ca7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 63, + 223, + 290, + 231 + ], + "lines": [ + { + "bbox": [ + 63, + 223, + 290, + 231 + ], + "spans": [ + { + "bbox": [ + 63, + 223, + 290, + 231 + ], + "type": "text", + "content": "Note: JSMANT and WJSMA could not be tested on ImageNet due to memory constraints" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 253, + 295, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 253, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 295, + 277 + ], + "type": "text", + "content": "plies that were successfully misclassified after perturbation out of all tested samples." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 277, + 295, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 277, + 295, + 397 + ], + "spans": [ + { + "bbox": [ + 54, + 277, + 295, + 397 + ], + "type": "text", + "content": "Notably, since JSMA-based methods operate iteratively (modifying two pixels per iteration), JSMANT and WJSMA use half the number of iterations compared to the total pixel modification count. FastJSMA modifies one pixel per iteration for fair comparison. The attempt parameter controls the number of candidate classes explored during class probing: we use attempt " + }, + { + "bbox": [ + 54, + 277, + 295, + 397 + ], + "type": "inline_equation", + "content": "= 10" + }, + { + "bbox": [ + 54, + 277, + 295, + 397 + ], + "type": "text", + "content": " for CIFAR-10, attempt " + }, + { + "bbox": [ + 54, + 277, + 295, + 397 + ], + "type": "inline_equation", + "content": "= \\sqrt{N}" + }, + { + "bbox": [ + 54, + 277, + 295, + 397 + ], + "type": "text", + "content": " for larger datasets (e.g., attempt " + }, + { + "bbox": [ + 54, + 277, + 295, + 397 + ], + "type": "inline_equation", + "content": "= 10" + }, + { + "bbox": [ + 54, + 277, + 295, + 397 + ], + "type": "text", + "content": " for CIFAR-100's 100 classes, attempt " + }, + { + "bbox": [ + 54, + 277, + 295, + 397 + ], + "type": "inline_equation", + "content": "= 32" + }, + { + "bbox": [ + 54, + 277, + 295, + 397 + ], + "type": "text", + "content": " for ImageNet's 1000 classes)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 405, + 180, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 405, + 180, + 417 + ], + "spans": [ + { + "bbox": [ + 55, + 405, + 180, + 417 + ], + "type": "text", + "content": "4.2. Performance Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 423, + 227, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 227, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 227, + 434 + ], + "type": "text", + "content": "4.2.1. Comparison with JSMA Variants" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 437, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 510 + ], + "type": "text", + "content": "Table 1 presents our main experimental results, comparing FastJSMA against JSMANT, WJSMA, and the random baseline across different datasets and model architectures. We report three key metrics: Attack Success Rate (ASR), average generation time per sample, and average perturbation magnitude (" + }, + { + "bbox": [ + 55, + 437, + 295, + 510 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 55, + 437, + 295, + 510 + ], + "type": "text", + "content": " norm)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 510, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 510, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 54, + 510, + 295, + 653 + ], + "type": "text", + "content": "The results empirically validate our theoretical complexity reduction from " + }, + { + "bbox": [ + 54, + 510, + 295, + 653 + ], + "type": "inline_equation", + "content": "O(M^2 \\times N)" + }, + { + "bbox": [ + 54, + 510, + 295, + 653 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 54, + 510, + 295, + 653 + ], + "type": "inline_equation", + "content": "O(M\\sqrt{N})" + }, + { + "bbox": [ + 54, + 510, + 295, + 653 + ], + "type": "text", + "content": " proposed in Section 3.4. FastJSMA achieves attack success rates comparable to or exceeding traditional JSMANT across all datasets while demonstrating remarkable computational efficiency—requiring just " + }, + { + "bbox": [ + 54, + 510, + 295, + 653 + ], + "type": "inline_equation", + "content": "2.9\\%" + }, + { + "bbox": [ + 54, + 510, + 295, + 653 + ], + "type": "text", + "content": " of JSMANT's processing time on CIFAR-10 (339" + }, + { + "bbox": [ + 54, + 510, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 54, + 510, + 295, + 653 + ], + "type": "text", + "content": "s vs. 11,769" + }, + { + "bbox": [ + 54, + 510, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 54, + 510, + 295, + 653 + ], + "type": "text", + "content": "s) and merely " + }, + { + "bbox": [ + 54, + 510, + 295, + 653 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 54, + 510, + 295, + 653 + ], + "type": "text", + "content": " on CIFAR-100 (751" + }, + { + "bbox": [ + 54, + 510, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 54, + 510, + 295, + 653 + ], + "type": "text", + "content": "s vs. 61,166" + }, + { + "bbox": [ + 54, + 510, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 54, + 510, + 295, + 653 + ], + "type": "text", + "content": "s). These speed-ups of " + }, + { + "bbox": [ + 54, + 510, + 295, + 653 + ], + "type": "inline_equation", + "content": "34.7\\times" + }, + { + "bbox": [ + 54, + 510, + 295, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 510, + 295, + 653 + ], + "type": "inline_equation", + "content": "81.4\\times" + }, + { + "bbox": [ + 54, + 510, + 295, + 653 + ], + "type": "text", + "content": " respectively confirm that the gradient decoupling mechanism theorized in our mathematical formulation successfully transforms the computational scaling properties of saliency-based attacks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": "The efficiency advantage increases with dataset complexity, closely matching our theoretical prediction that traditional JSMA's complexity worsens quadratically with increasing image size and class count. Most significantly, FastJSMA successfully executes on ImageNet with a rea" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 337, + 91, + 535, + 144 + ], + "blocks": [ + { + "bbox": [ + 346, + 71, + 523, + 83 + ], + "lines": [ + { + "bbox": [ + 346, + 71, + 523, + 83 + ], + "spans": [ + { + "bbox": [ + 346, + 71, + 523, + 83 + ], + "type": "text", + "content": "Table 2. Ablation study on gradient components" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 337, + 91, + 535, + 144 + ], + "lines": [ + { + "bbox": [ + 337, + 91, + 535, + 144 + ], + "spans": [ + { + "bbox": [ + 337, + 91, + 535, + 144 + ], + "type": "table", + "html": "
Gradient ComponentASR(%)Time(μs)L2Norm
FastJSMA (Complete)43.983391.66
G+30.664321.71
G-35.754431.66
", + "image_path": "463bf61f22276ff4d7d74253113bd2fdca98876925f71e96ddc03fb4351dcc81.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 164, + 553, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 164, + 553, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 164, + 553, + 236 + ], + "type": "text", + "content": "sonable computation time of " + }, + { + "bbox": [ + 313, + 164, + 553, + 236 + ], + "type": "inline_equation", + "content": "7,908\\mu s" + }, + { + "bbox": [ + 313, + 164, + 553, + 236 + ], + "type": "text", + "content": " per sample, while both JSMANT and WJSMA fail due to memory constraints. This empirical finding not only validates our theoretical analysis but extends its implications—demonstrating that our mathematical reformulation enables saliency-based attacks on previously inaccessible large-scale datasets." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 236, + 554, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 554, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 554, + 344 + ], + "type": "text", + "content": "WJSMA achieves somewhat higher ASRs on CIFAR datasets, which aligns with our theoretical understanding of its specialized optimization mechanisms. However, its computational demands confirm our analysis that approaches relying on the full Jacobian matrix face fundamental limitations for larger-scale applications. These results demonstrate the theoretical-practical trade-off that motivated our work: FastJSMA sacrifices minimal attack effectiveness for transformative computational efficiency gains." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 351, + 517, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 351, + 517, + 363 + ], + "spans": [ + { + "bbox": [ + 313, + 351, + 517, + 363 + ], + "type": "text", + "content": "4.2.2. Ablation Study on Gradient Components" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 365, + 554, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 554, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 554, + 426 + ], + "type": "text", + "content": "To investigate the contribution of different gradient components in our dual-gradient approximation strategy, we conducted an ablation study comparing three variants of FastJSMA using the CIFAR-10 dataset with ResNet18, 8 pixels modification, and attempt=10." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 426, + 554, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 554, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 554, + 569 + ], + "type": "text", + "content": "These results validate and refine our theoretical decomposition of the Jacobian matrix described in Section 3.2. The complete FastJSMA incorporating both gradients achieves the highest ASR (43.98%), empirically confirming that our dual-gradient formulation successfully captures the bidirectional influence encoded in traditional JSMA's saliency computation. The class suppression gradient " + }, + { + "bbox": [ + 313, + 426, + 554, + 569 + ], + "type": "inline_equation", + "content": "(g^{-})" + }, + { + "bbox": [ + 313, + 426, + 554, + 569 + ], + "type": "text", + "content": " alone yields a " + }, + { + "bbox": [ + 313, + 426, + 554, + 569 + ], + "type": "inline_equation", + "content": "35.75\\%" + }, + { + "bbox": [ + 313, + 426, + 554, + 569 + ], + "type": "text", + "content": " ASR, outperforming the class excitation gradient " + }, + { + "bbox": [ + 313, + 426, + 554, + 569 + ], + "type": "inline_equation", + "content": "(g^{+})" + }, + { + "bbox": [ + 313, + 426, + 554, + 569 + ], + "type": "text", + "content": " alone (30.66%), revealing a key theoretical insight: reducing the current class probability has a more significant impact than increasing alternative class probabilities." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "content": "This empirical finding extends our initial theoretical understanding, revealing an asymmetric contribution pattern not initially anticipated in our mathematical formulation. This discovery explains the effectiveness of our normalization procedure in Section 3.3, which amplifies the class suppression gradient by a factor of " + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "content": " —the ablation empirically validates this engineering optimization while simultaneously deepening our theoretical understanding. The comparable computation times across variants (339-443" + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "content": "s) further validate our complexity analysis that dual gradient computation maintains " + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "inline_equation", + "content": "O(M\\sqrt{N})" + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "content": " efficiency regardless of variant configuration." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1512" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 86, + 304, + 208 + ], + "blocks": [ + { + "bbox": [ + 126, + 73, + 492, + 84 + ], + "lines": [ + { + "bbox": [ + 126, + 73, + 492, + 84 + ], + "spans": [ + { + "bbox": [ + 126, + 73, + 492, + 84 + ], + "type": "text", + "content": "(a) Fast-JSMA Sensitivity to Max Attempt Fast-JSMA Parameter Sensitivity Analysis (b) Fast-JSMA Sensitivity to Pixels Perturbed" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 86, + 304, + 208 + ], + "lines": [ + { + "bbox": [ + 58, + 86, + 304, + 208 + ], + "spans": [ + { + "bbox": [ + 58, + 86, + 304, + 208 + ], + "type": "image", + "image_path": "d2d6803a5666c1e24d0a1dc0b433e4cc1ce61fe9ec414e897fdc68eb24897430.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 304, + 87, + 552, + 208 + ], + "blocks": [ + { + "bbox": [ + 304, + 87, + 552, + 208 + ], + "lines": [ + { + "bbox": [ + 304, + 87, + 552, + 208 + ], + "spans": [ + { + "bbox": [ + 304, + 87, + 552, + 208 + ], + "type": "image", + "image_path": "529430c6687121ff635434c1cc1500c63bdaf29e978dbfa5809953de1e1906a8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 222, + 555, + 266 + ], + "lines": [ + { + "bbox": [ + 55, + 222, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 222, + 555, + 266 + ], + "type": "text", + "content": "Figure 2. Parameter sensitivity analysis of FastJSMA on CIFAR-10. (a) Sensitivity to Max Attempt parameter showing ASR and Time curves. (b) Sensitivity to pixels perturbed showing ASR, Time, and Noise curves. All metrics are normalized using min-max scaling for visual comparison; actual ASR percentages shown on right y-axis. Experiments conducted on the complete CIFAR-10 test set (10,000 samples) with ResNet18." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 288, + 215, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 215, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 215, + 300 + ], + "type": "text", + "content": "4.2.3. Parameter Sensitivity Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 304, + 295, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 295, + 363 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 295, + 363 + ], + "type": "text", + "content": "To evaluate FastJSMA's robustness to parameter settings, we analyzed two key parameters: pixel modification count and the attempt parameter. Figure 2 presents these analyses conducted on the complete CIFAR-10 test set (10,000 samples)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 365, + 295, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 508 + ], + "type": "text", + "content": "Our Max Attempt parameter analysis (Figure 2a) shows that attack success rate follows a non-linear pattern: rapid initial growth from attempt " + }, + { + "bbox": [ + 55, + 365, + 295, + 508 + ], + "type": "inline_equation", + "content": "= 1" + }, + { + "bbox": [ + 55, + 365, + 295, + 508 + ], + "type": "text", + "content": " to attempt " + }, + { + "bbox": [ + 55, + 365, + 295, + 508 + ], + "type": "inline_equation", + "content": "= 4" + }, + { + "bbox": [ + 55, + 365, + 295, + 508 + ], + "type": "text", + "content": ", followed by gradual increase thereafter. This validates our square-root sampling heuristic described in Section 3.4, confirming that smaller attempt values provide higher marginal benefits. For scalability, we empirically determined that " + }, + { + "bbox": [ + 55, + 365, + 295, + 508 + ], + "type": "inline_equation", + "content": "\\sqrt{N}" + }, + { + "bbox": [ + 55, + 365, + 295, + 508 + ], + "type": "text", + "content": " sampling works well for larger datasets. Computational cost scales linearly with attempt value, supporting our " + }, + { + "bbox": [ + 55, + 365, + 295, + 508 + ], + "type": "inline_equation", + "content": "O(M\\sqrt{N})" + }, + { + "bbox": [ + 55, + 365, + 295, + 508 + ], + "type": "text", + "content": " complexity analysis. The data ranges from " + }, + { + "bbox": [ + 55, + 365, + 295, + 508 + ], + "type": "inline_equation", + "content": "35.2\\%" + }, + { + "bbox": [ + 55, + 365, + 295, + 508 + ], + "type": "text", + "content": " ASR at attempt " + }, + { + "bbox": [ + 55, + 365, + 295, + 508 + ], + "type": "inline_equation", + "content": "= 1" + }, + { + "bbox": [ + 55, + 365, + 295, + 508 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 365, + 295, + 508 + ], + "type": "inline_equation", + "content": "44.7\\%" + }, + { + "bbox": [ + 55, + 365, + 295, + 508 + ], + "type": "text", + "content": " ASR at attempt " + }, + { + "bbox": [ + 55, + 365, + 295, + 508 + ], + "type": "inline_equation", + "content": "= 10" + }, + { + "bbox": [ + 55, + 365, + 295, + 508 + ], + "type": "text", + "content": ", with corresponding time costs increasing from " + }, + { + "bbox": [ + 55, + 365, + 295, + 508 + ], + "type": "inline_equation", + "content": "314\\mu s" + }, + { + "bbox": [ + 55, + 365, + 295, + 508 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 365, + 295, + 508 + ], + "type": "inline_equation", + "content": "417\\mu s" + }, + { + "bbox": [ + 55, + 365, + 295, + 508 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 509, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 652 + ], + "type": "text", + "content": "The pixel modification analysis (Figure 2b) reveals that FastJSMA achieves substantial attack success even with minimal pixel modifications, with performance following an S-curve: initial rapid growth " + }, + { + "bbox": [ + 55, + 509, + 295, + 652 + ], + "type": "inline_equation", + "content": "(2\\rightarrow 16" + }, + { + "bbox": [ + 55, + 509, + 295, + 652 + ], + "type": "text", + "content": " pixels), moderate middle phase growth " + }, + { + "bbox": [ + 55, + 509, + 295, + 652 + ], + "type": "inline_equation", + "content": "(16\\rightarrow 36" + }, + { + "bbox": [ + 55, + 509, + 295, + 652 + ], + "type": "text", + "content": " pixels), and final diminishing returns " + }, + { + "bbox": [ + 55, + 509, + 295, + 652 + ], + "type": "inline_equation", + "content": "(36\\rightarrow 50" + }, + { + "bbox": [ + 55, + 509, + 295, + 652 + ], + "type": "text", + "content": " pixels). This validates our significance threshold mechanism in Section 3.3. Runtime increases only marginally despite significant increases in modified pixels, while L2 norm scales approximately linearly with pixel count. The analysis shows ASR ranging from " + }, + { + "bbox": [ + 55, + 509, + 295, + 652 + ], + "type": "inline_equation", + "content": "26.0\\%" + }, + { + "bbox": [ + 55, + 509, + 295, + 652 + ], + "type": "text", + "content": " with 2 pixels to " + }, + { + "bbox": [ + 55, + 509, + 295, + 652 + ], + "type": "inline_equation", + "content": "69.5\\%" + }, + { + "bbox": [ + 55, + 509, + 295, + 652 + ], + "type": "text", + "content": " with 50 pixels, with average generation time remaining stable around 0.0003-0.0004 seconds." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "text", + "content": "These findings both validate and extend our theoretical model. The consistent computational efficiency across different pixel modification counts demonstrates FastJSMA's scalability, while the sharp initial gains in both analyses highlight the method's effectiveness at identifying influen" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 288, + 555, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 555, + 312 + ], + "type": "text", + "content": "tial pixels with minimal computational overhead, making it particularly suitable for constrained attack scenarios." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 365, + 387, + 378 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 365, + 387, + 378 + ], + "spans": [ + { + "bbox": [ + 314, + 365, + 387, + 378 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 399, + 555, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 555, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 555, + 520 + ], + "type": "text", + "content": "This paper introduces FastJSMA, a computationally efficient reformulation of saliency-based adversarial attacks. By decomposing the traditional Jacobian calculation into complementary gradient components—class suppression " + }, + { + "bbox": [ + 313, + 399, + 555, + 520 + ], + "type": "inline_equation", + "content": "(g^{-})" + }, + { + "bbox": [ + 313, + 399, + 555, + 520 + ], + "type": "text", + "content": " and class excitation " + }, + { + "bbox": [ + 313, + 399, + 555, + 520 + ], + "type": "inline_equation", + "content": "(g^{+})" + }, + { + "bbox": [ + 313, + 399, + 555, + 520 + ], + "type": "text", + "content": "—we reduce computational complexity from " + }, + { + "bbox": [ + 313, + 399, + 555, + 520 + ], + "type": "inline_equation", + "content": "O(M^2 \\times N)" + }, + { + "bbox": [ + 313, + 399, + 555, + 520 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 399, + 555, + 520 + ], + "type": "inline_equation", + "content": "O(M\\sqrt{N})" + }, + { + "bbox": [ + 313, + 399, + 555, + 520 + ], + "type": "text", + "content": " while maintaining attack effectiveness. Our class probing mechanism and significance threshold further optimize performance by systematically identifying vulnerable decision boundaries and prioritizing influential pixels." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "content": "Experiments validate that FastJSMA achieves dramatic computational efficiency gains—up to " + }, + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "inline_equation", + "content": "81.4 \\times" + }, + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "content": " speedup on CIFAR-100 and successful execution on ImageNet where traditional JSMA fails entirely—while maintaining comparable attack success rates. The ablation study reveals the asymmetric contribution of gradient components, providing new insights that extend our theoretical understanding. These findings demonstrate that our mathematical reformulation successfully overcomes the computational barriers that have limited JSMA's applicability in large-scale settings. Beyond immediate technical contributions, FastJSMA's gradient decoupling principles may inform future research in computationally efficient adversarial attacks and explainable AI, where feature importance analysis at scale remains a significant challenge." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1513" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 198, + 115, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 115, + 211 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 115, + 211 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 219, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 219, + 296, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 219, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 61, + 219, + 296, + 262 + ], + "type": "text", + "content": "[1] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 263, + 296, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 263, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 61, + 263, + 296, + 297 + ], + "type": "text", + "content": "[2] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 299, + 296, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 299, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 62, + 299, + 296, + 342 + ], + "type": "text", + "content": "[3] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 343, + 296, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 343, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 62, + 343, + 296, + 376 + ], + "type": "text", + "content": "[4] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In 2017 IEEE Symposium on Security and Privacy (SP), pages 39-57. IEEE, 2017. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 377, + 296, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 377, + 296, + 432 + ], + "spans": [ + { + "bbox": [ + 62, + 377, + 296, + 432 + ], + "type": "text", + "content": "[5] Nicolas Papernot, Patrick McDaniel, Somesh Jha, Matt Fredrikson, Z. Berkay Celik, and Ananthram Swami. The limitations of deep learning in adversarial settings. In 2016 IEEE European Symposium on Security and Privacy (EuroS&P), pages 372-387. IEEE, 2016. 1, 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 434, + 296, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 434, + 296, + 477 + ], + "spans": [ + { + "bbox": [ + 62, + 434, + 296, + 477 + ], + "type": "text", + "content": "[6] Théo Combey, António Loison, Maxime Faucher, and Hatem Hajri. Probabilistic jacobian-based saliency maps attacks. Machine Learning and Knowledge Extraction, 2(4): 558-578, 2020. doi: 10.3390/make2040030. 1, 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 478, + 296, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 478, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 62, + 478, + 296, + 510 + ], + "type": "text", + "content": "[7] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In CVPR09, 2009. 1, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 512, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 512, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 512, + 296, + 555 + ], + "type": "text", + "content": "[8] Nicolas Papernot, Patrick McDaniel, and Ian Goodfellow. Transferability in machine learning: from phenomena to black-box attacks using adversarial samples. arXiv preprint arXiv:1605.07277, 2016. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 557, + 296, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 557, + 296, + 622 + ], + "spans": [ + { + "bbox": [ + 62, + 557, + 296, + 622 + ], + "type": "text", + "content": "[9] Pin-Yu Chen, Huan Zhang, Yash Sharma, Jinfeng Yi, and Cho-Jui Hsieh. Zoo: Zeroth order optimization based blackbox attacks to deep neural networks without training substitute models. In Proceedings of the 10th ACM Workshop on Artificial Intelligence and Security, pages 15-26. ACM, 2017. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 624, + 296, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 296, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 296, + 668 + ], + "type": "text", + "content": "[10] Dawei Zhou, Nannan Wang, Tongliang Liu, and Xinbo Gao. Improving adversarial training from the perspective of class-flipping distribution. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 714 + ], + "type": "text", + "content": "[11] Dawei Zhou, Nannan Wang, Bo Han, Tongliang Liu, and Xinbo Gao. Defending against adversarial examples via modeling adversarial noise. International Journal of Computer Vision, pages 1-18, 2025." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 55, + 91, + 297, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 187 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 187 + ], + "type": "text", + "content": "We would like to thank the HPC Platform of Huazhong University of Science and Technology for providing computational resources. We also express our gratitude to the School of Artificial Intelligence and Automation, the School of Software Engineering and the School of Electronic Information and Communications at Huazhong University of Science and Technology for their strong support throughout this research." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 316, + 73, + 555, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 128 + ], + "type": "text", + "content": "[12] Chaojian Yu, Bo Han, Li Shen, Jun Yu, Chen Gong, Mingming Gong, and Tongliang Liu. Understanding robust overfitting of adversarial training and beyond. In International Conference on Machine Learning, pages 25595-25610. PMLR, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 129, + 555, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 555, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 555, + 184 + ], + "type": "text", + "content": "[13] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and accurate method to fool deep neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2574–2582, 2016. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 186, + 554, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 554, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 554, + 239 + ], + "type": "text", + "content": "[14] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9185-9193, 2018. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 241, + 554, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 554, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 554, + 285 + ], + "type": "text", + "content": "[15] Jiadong Lin, Chuanbiao Song, Kun He, Liwei Wang, and John E. Hopcroft. Nesterov accelerated gradient and scale invariance for adversarial attacks. In International Conference on Learning Representations, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 286, + 554, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 554, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 554, + 342 + ], + "type": "text", + "content": "[16] Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Evading defenses to transferable adversarial examples by translation-invariant attacks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4312-4321, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 343, + 554, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 554, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 554, + 398 + ], + "type": "text", + "content": "[17] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L. Yuille. Improving transferability of adversarial examples with input diversity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2730-2739, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 399, + 554, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 554, + 453 + ], + "type": "text", + "content": "[18] Chengzhi Mao, Mia Chiquier, Hao Wang, Junfeng Yang, and Carl Vondrick. Adversarial attacks are reversible with natural supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 661-671, October 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 455, + 554, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 455, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 554, + 498 + ], + "type": "text", + "content": "[19] Jianming Zhang and Stan Sclaroff. Saliency detection: A boolean map approach. In 2013 IEEE International Conference on Computer Vision, pages 153-160, 2013. doi: 10.1109/ICCV.2013.26.2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 500, + 554, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 554, + 543 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 554, + 543 + ], + "type": "text", + "content": "[20] Jiawei Su, Danilo Vasconcellos Vargas, and Kouichi Sakurai. One pixel attack for fooling deep neural networks. IEEE Transactions on Evolutionary Computation, 23(5):828-841, 2019. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 544, + 554, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 554, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 554, + 590 + ], + "type": "text", + "content": "[21] Jianbo Chen, Michael I. Jordan, and Martin J. Wainwright. Hopskipjumpattack: A query-efficient decision-based attack. In 2020 IEEE Symposium on Security and Privacy (SP), pages 1277-1294. IEEE, 2020. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 590, + 555, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 555, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 555, + 634 + ], + "type": "text", + "content": "[22] Tianhang Zheng, Changyou Chen, Junsong Yuan, Bo Li, and Kui Ren. Pointcloud saliency maps. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 635, + 554, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 554, + 689 + ], + "type": "text", + "content": "[23] Apostolos Modas, Seyed-Mohsen Moosavi-Dezfooli, and Pascal Frossard. Sparsefool: a few pixels make a big difference. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9087–9096, 2019. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 554, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 554, + 714 + ], + "type": "text", + "content": "[24] Florian Tramér, Alexey Kurakin, Nicolas Papernot, Ian Goodfellow, Dan Boneh, and Patrick McDaniel. Ensemble" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1514" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 319 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "text", + "content": "adversarial training: Attacks and defenses. In International Conference on Learning Representations, 2018. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "text", + "content": "[25] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International Conference on Machine Learning, pages 2206-2216. PMLR, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 295, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 295, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 295, + 195 + ], + "type": "text", + "content": "[26] Maksym Andriushchenko, Francesco Croce, Nicolas Flammarion, and Matthias Hein. Square attack: a query-efficient black-box adversarial attack via random search. In European Conference on Computer Vision, pages 484-501. Springer, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 295, + 239 + ], + "type": "text", + "content": "[27] Maura Pintor, Fabio Roli, Wieland Brendel, and Battista Biggio. Fast minimum-norm adversarial attacks through adaptive norm constraints. arXiv preprint arXiv:2102.12827, 2021.3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 295, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 295, + 274 + ], + "type": "text", + "content": "[28] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. Technical Report 4, University of Toronto, 2009. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 295, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 295, + 319 + ], + "type": "text", + "content": "[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1515" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/FastPoint_ Accelerating 3D Point Cloud Model Inference via Sample Point Distance Prediction/670cb46f-786b-455d-998f-328c26d7e47d_content_list.json b/2025/FastPoint_ Accelerating 3D Point Cloud Model Inference via Sample Point Distance Prediction/670cb46f-786b-455d-998f-328c26d7e47d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c3fddb89350eebb3237918328a3d32cb5e972a8b --- /dev/null +++ b/2025/FastPoint_ Accelerating 3D Point Cloud Model Inference via Sample Point Distance Prediction/670cb46f-786b-455d-998f-328c26d7e47d_content_list.json @@ -0,0 +1,1683 @@ +[ + { + "type": "text", + "text": "FastPoint: Accelerating 3D Point Cloud Model Inference via Sample Point Distance Prediction", + "text_level": 1, + "bbox": [ + 196, + 130, + 803, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Donghyun Lee $^{1}$ Dawoon Jeong $^{1}$ Jae W. Lee $^{1}$ Hongil Yoon $^{1,2}$ Seoul National University ${ }^{2}$ Google", + "bbox": [ + 225, + 202, + 769, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{eudh1206, daun20211, jaewlee}@snu.ac.kr, hongilyoon@google.com", + "bbox": [ + 217, + 241, + 774, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep neural networks have revolutionized 3D point cloud processing, yet efficiently handling large and irregular point clouds remains challenging. To tackle this problem, we introduce FastPoint, a novel software-based acceleration technique that leverages the predictable distance trend between sampled points during farthest point sampling. By predicting the distance curve, we can efficiently identify subsequent sample points without exhaustively computing all pairwise distances. Our proposal substantially accelerates farthest point sampling and neighbor search operations while preserving sampling quality and model performance. By integrating FastPoint into state-of-the-art 3D point cloud models, we achieve $2.55 \\times$ end-to-end speedup on NVIDIA RTX 3090 GPU without sacrificing accuracy.", + "bbox": [ + 89, + 324, + 485, + 537 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 569, + 220, + 585 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D point clouds have become increasingly important for representing and understanding 3D scenes, driving advancements in fields like robotics and autonomous driving. Deep neural networks [4, 6, 7, 13, 17, 20, 23-26, 30, 32, 33, 37, 42, 43] have emerged as powerful tools for processing this type of data. While PointNet++ [24] and its successors have improved efficiency and performance of these point cloud models, the irregular nature and growing size of point cloud data continue to impose significant computational challenges, especially for real time use cases. Operations such as Farthest Point Sampling (FPS) and neighbor search remain major bottlenecks.", + "bbox": [ + 89, + 595, + 482, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While various hardware accelerators [8, 9, 12, 16, 18, 36, 40] have been explored to accelerate the models, software-only solutions are less common. Given the high cost and complexity of developing specialized accelerators, software-based acceleration is a more practical approach. Recently, several software-based techniques [15, 39] have been proposed to accelerate these models through approximation. However, such approaches often come at the cost", + "bbox": [ + 89, + 779, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of significant accuracy loss, limiting their practical use.", + "bbox": [ + 511, + 292, + 877, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address this challenge, we propose FastPoint, a novel software-based acceleration technique with sample point distance prediction. We observe that the inherent nature of FPS, which iteratively selects the point farthest from the current set of sampled points, leads to two predictable trends in the minimum distances between these points:", + "bbox": [ + 511, + 308, + 905, + 398 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Decreasing minimum distance: As FPS makes progress, the number of remaining points decreases, forcing the algorithm to select points closer to those already sampled. This results in a smoothly decreasing curve of minimum distances between sampled points.", + "bbox": [ + 514, + 398, + 905, + 474 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2. Early structure capture: The initial sample points tend to be the extremities of the point cloud, effectively capturing its overall shape and boundaries.", + "bbox": [ + 511, + 474, + 903, + 518 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our proposal leverages these observations to accurately estimate the distance curve using only a few initial FPS iterations. By predicting the distance curve, we can efficiently identify subsequent sample points without computing and comparing pairwise distances at each FPS iteration. This leads to significant latency savings while maintaining sampling quality comparable to that of the original FPS. Furthermore, predicting the distance curve enables the sampling process to be decoupled from distance computation. This decoupling exposes new opportunities for parallelizing distance calculations in the proposed sampling, further enhancing efficiency. This approach not only accelerates FPS itself but also benefits subsequent operations such as neighbor search by leveraging pre-computed distances.", + "bbox": [ + 511, + 520, + 905, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our key contributions are summarized as follows:", + "bbox": [ + 532, + 732, + 862, + 744 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We identify two key latency bottlenecks in point cloud models: FPS and neighbor search. Limited parallelism in FPS and redundant distance computations across both operations are the primary sources of inefficiency.", + "- We empirically analyze 3D point cloud models and identify the decreasing trend and early structure capture trend.", + "- Based on the observations, we introduce a novel sampling approach with sample point distance prediction. With the estimated distance curve, this approach not only accelerates FPS itself but also benefits subsequent neighbor" + ], + "bbox": [ + 511, + 750, + 903, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25114", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1cb09cf346bf4144a51258661e69dc8af82c06e8da911b5af9dcbc417c035d59.jpg", + "image_caption": [ + "Figure 1. Overview of PointNet++ Based Model Architecture" + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 911, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "search, which relies on pre-computed distances.", + "bbox": [ + 104, + 311, + 421, + 325 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We integrate FastPoint into OpenPoints library and experiment on state-of-the-art PointNet++ based models, achieving geomean $2.55 \\times$ end-to-end speedup on NVIDIA RTX 3090 GPU while preserving accuracy.", + "bbox": [ + 91, + 327, + 483, + 388 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Background and Related Work", + "text_level": 1, + "bbox": [ + 89, + 398, + 377, + 415 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Deep Learning on 3D Point Clouds", + "text_level": 1, + "bbox": [ + 89, + 422, + 395, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deep Neural Networks on Raw Point Cloud Since the release of PointNet [23], which is the first to apply deep neural networks directly to the raw point clouds, numerous successors [6, 13, 17, 19, 24-27, 30, 31, 34, 41] have been developed, improving both model performance and efficiency. PointNet++ [24] introduces a hierarchical structure to PointNet through sampling and grouping, which has become the core architecture of subsequent PointNet++ based models. PointNeXt [27] revisits PointNet++, achieving substantial improvements in model performance, while PointVector [6] and PointMetaBase [17] push the model performance and efficiency even further. Although the architectural details of each model differ, all models follow the PointNet++ structure, utilizing Set Abstraction and Feature Propagation layers as core components. This work focuses on identifying and accelerating the primary latency bottlenecks in PointNet++ based models.", + "bbox": [ + 89, + 445, + 483, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deep Neural Networks on Voxelized Point Cloud Several works [3, 4, 10, 22, 28, 43] propose to voxelize point clouds, using voxels as input rather than raw points. The advent of 3D sparse convolution [4] has made this voxel-based approach increasingly popular due to its efficiency and high performance. Recently, transformer-based models [7, 21, 37, 38] operating on voxelized point clouds are gaining significant attention. Although voxelization alleviates the computational cost of mapping operations (i.e., downsampling and neighbor search), it has a notable limitation due to the loss of position information during the voxelization process.", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 310, + 658, + 325 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Numerous solutions have been proposed to accelerate sampling and neighbor search operations in PointNet++ based models. We categorize them into hardware- and purely software-based approaches.", + "bbox": [ + 511, + 333, + 906, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Hardware-Assisted Acceleration QuickFPS [12] introduces a k-d tree based FPS algorithm which reduces computation and memory access in each FPS iteration, along with specialized hardware designed for this method. MARS [36] and PTrAcc [16] both propose hardware accelerators that employ a distance filtering technique which skips unnecessary distance computations in each iteration of FPS. Although these techniques achieve substantial speedups in FPS without any loss in sampling quality, they require specialized hardware to fully harness their algorithms, limiting their effectiveness on commodity hardware like GPU. For more thorough evaluation, we compare the speedup of our proposal with pure software version [11] of QuickFPS in Section 5.4, evaluating its effectiveness on GPU.", + "bbox": [ + 511, + 412, + 908, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Software-Only Acceleration To mitigate the high latency associated with FPS, various alternative sampling methods have been explored. RandLA-Net [13] replaces FPS with random sampling and introduces a local feature aggregation module that compensates for the low sampling quality of random sampling. Grid sampling, the most commonly used alternative, is adopted by Grid-GCN [35] and KPCov [29], offering faster processing with reasonable sampling quality compared to random sampling. Despite the latency advantages of these methods, none have consistently outperformed FPS for overall performance. This is illustrated by the fact that most PointNet++-based models [6, 17, 24-27] continue to rely on FPS, highlighting its broad applicability and superior sampling quality. The experimental results in Appendix B.1 corroborate this claim.", + "bbox": [ + 511, + 643, + 908, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "EdgePC [39] structures the point cloud with morton codes to accelerate FPS and neighbor search on edge GPUs,", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25115", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ae2db72dee1c17f01cd52410b02dd70628795cce94c8d8d03b0d8080d45789cf.jpg", + "image_caption": [ + "Figure 2. Latency Breakdown of PointNet++ Based Models" + ], + "image_footnote": [], + "bbox": [ + 112, + 90, + 460, + 276 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "while Adjustable FPS [15] accelerates FPS by leveraging the intrinsic locality of the point cloud data. However, both methods demonstrate significant accuracy loss compared to the baseline (i.e., up to $2\\%$ mIoU loss) due to their aggressive approximations, limiting their effectiveness. Prior work [14] accelerates the training of PointNet++ based models by precomputing the minimum point spacing of farthest-sampled points before training and reusing this value across epochs for fast sampling. While this approach effectively speeds up training, it is not suitable for inference, as the minimum spacing value cannot be determined in advance in inference scenarios. Detailed comparisons with our work are provided in Appendix A.4.", + "bbox": [ + 89, + 320, + 483, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Challenges of Point Cloud Models", + "text_level": 1, + "bbox": [ + 89, + 534, + 398, + 550 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Model Architecture Overview", + "text_level": 1, + "bbox": [ + 89, + 559, + 356, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 1 illustrates the detailed process of PointNet++-based models. Similar to 2D convolutional neural networks, these models consist of both downsampling and upsampling layers. Commonly referred to as the set abstraction layer, the downsampling layer in PointNet++-based models comprises three main stages, as outlined below:", + "bbox": [ + 89, + 582, + 482, + 672 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1 Farthest Point Sampling (FPS): Downsampling begins by selecting $N$ /stride output points from $N$ input points, where stride represents the downsampling rate. Typically, the FPS is used to maintain the shape and structure of the input point cloud.", + "$\\bullet$ Ball Query: Next, a ball query with radius $R$ is performed around each sampled point to gather neighboring points for feature aggregation.", + "Aggregation: The features of neighboring points identified in the ball query are then processed. A multi-layer perceptron (MLP) is applied to these features, followed by max pooling to aggregate the information for each centroid. The details of this process, e.g., positional embedding [17], can vary depending on the model architectures." + ], + "bbox": [ + 101, + 674, + 482, + 898 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Algorithm 1 Farthest Point Sampling", + "text_level": 1, + "bbox": [ + 514, + 90, + 764, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Input $P$ : input point cloud, $n$ : number of points to sample, $N$ : number of original points", + "bbox": [ + 514, + 108, + 893, + 135 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Output sampledidx: indices of sampled points", + "bbox": [ + 514, + 136, + 799, + 148 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: sampled_idx[0] $\\leftarrow$ seed_idx // Starts with a random point", + "2: min_dists[i] $\\leftarrow \\infty$ for $i = 0,\\dots,N - 1$", + "3: for $i \\gets 1$ to $n - 1$ do // Not parallelized", + "4: for $j\\gets 0$ to $N - 1$ do // Parallelized", + "5: $d_{\\text{new}} \\gets \\text{dist}(P[\\text{sampled\\_idx}[i - 1]], P[j])$", + "6: if min_dists[j] > $d_{new}$ then", + "7: min_dists[j] $\\leftarrow d_{new}$", + "8: sampledidx[i] $\\leftarrow$ argmax(min_dists)" + ], + "bbox": [ + 521, + 151, + 895, + 263 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The upsampling layer, often referred to as a feature propagation layer, takes the downsampled points from the corresponding downsampling layer as inputs and propagates the features back to the original point cloud. This process is done by following steps.", + "bbox": [ + 511, + 310, + 903, + 383 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$\\bullet$ k-NN: For each upsampled point, k-NN is applied to identify its k nearest neighbors among the input points.", + "$\\Theta$ Interpolation: Based on the k-NN results, neighboring input points are interpolated, concatenated with skip-connected features from the corresponding downsampling layer, and the information is passed through an MLP to propagate the features to the subsequent layer." + ], + "bbox": [ + 524, + 385, + 903, + 491 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Latency Breakdown", + "text_level": 1, + "bbox": [ + 513, + 500, + 705, + 515 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 2 presents the latency breakdown of PointMetaBase-L model evaluated on both indoor (S3DIS) and outdoor (SemanticKITTI) scenes. The results show that coordinate-based operations (i.e., FPS, k-NN, and Ball Query) dominate the execution time, consuming significantly more time than feature-based operations (i.e., Aggregation and MLP). These coordinate-based operations account for approximately $88.1\\%$ of the total inference time for indoor scenes and $94.1\\%$ for the outdoor scenes. Among these, FPS emerges as the primary performance bottleneck, requiring the most execution time. Additionally, neighbor search operations, including k-NN and Ball Query, contribute significantly to the overall latency. To address these challenges, we first analyze the root causes of these inefficiencies and propose novel optimization techniques in subsequent sections.", + "bbox": [ + 511, + 522, + 906, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Inefficiencies in Farthest Point Sampling", + "text_level": 1, + "bbox": [ + 511, + 772, + 859, + 789 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Farthest point sampling operates iteratively by selecting one point at a time. In each iteration, the sampling algorithm identifies the point farthest from the set of the previously sampled points. The detailed process is outlined in Algorithm 1. Initially, a random point in $P$ is chosen as the seed point (Line 1) and added to the sampled_idx. Then, a distance matrix (i.e., min_dists) is created to store the distance", + "bbox": [ + 511, + 795, + 905, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25116", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "between each point in $P$ and the set of the previously sampled points, where the distance from a point to a point set is defined as minimum distance between the point and any point in the set. This matrix is initialized to $\\infty$ (Line 2). In each iteration, the distance between the most recently sampled point and every other point in $P$ is calculated (Lines 3-7). If a newly calculated distance is smaller than the corresponding entry in the min_dist matrix, the matrix is updated with the smaller value. Finally, the index of the point with the maximum distance in the min_dist is added to the sampled_idx (Line 8). This process is repeated until the specified number of points $(n)$ is sampled.", + "bbox": [ + 89, + 90, + 480, + 271 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The sequential nature of the FPS algorithm is the primary reason of the inefficiency. Because only one point is sampled per iteration, the $O(N)$ distance calculations required in each iteration must be sequentially executed $n$ times. This tight coupling between sampling and distance calculation significantly limits the potential for parallelism in distance computations.", + "bbox": [ + 89, + 272, + 480, + 377 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To unlock this parallelism potential, we introduce Minimum Distance Prediction Sampling (MDPS) in Section 4.1, a novel sampling strategy that fully decouples sampling from distance computation, enabling greater parallelism and faster processing speeds.", + "bbox": [ + 89, + 378, + 480, + 453 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Inefficiencies in Neighbor Search", + "text_level": 1, + "bbox": [ + 89, + 460, + 379, + 477 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As explained in Section 3.1, PointNet++ based models employ two types of neighbor search algorithms: Ball Query and k-NN. These operations play a critical role in capturing local geometric relationships in the point cloud. Ball Query identifies neighbor points within a fixed radius around the downsampled points. k-NN identifies k nearest neighbors among the downsampled points for each original point.", + "bbox": [ + 89, + 483, + 480, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The inefficiencies in these neighbor search operations stem from redundant distance computations in both the FPS and subsequent neighbor search operations. For instance, as shown in Figure 1, the distance between P0 and P2 is computed while updating P2's minimum distance in FPS. However, the same distance is redundantly recalculated in both Ball Query and k-NN. Section 4.2 introduces redundancy-free neighbor search techniques to address this issue.", + "bbox": [ + 88, + 589, + 480, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. FastPoint: Fast Point Cloud Inference", + "text_level": 1, + "bbox": [ + 89, + 722, + 431, + 738 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Trends in Minimum Distance of Sample Points The FPS iteratively selects the point farthest from the current set of sampled points, which leads to predictable trends in the distance between the sampled points.", + "bbox": [ + 89, + 747, + 480, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "1. Decreasing minimum distance: As sampling proceeds, the pool of the remaining points shrinks, the FPS algorithm selects points closer to those already sampled over time. Thus, the maximized minimum distance (i.e., $\\max (min\\_dist)$ in Algorithm 1) decreases, which we define as minimum distance for brevity. This is particu", + "bbox": [ + 91, + 810, + 480, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/55827e7bb3e7aff332dc1de514fd528fb813171139164c5a918de76c6fce7757.jpg", + "image_caption": [ + "(a) Minimum Distance Curve of ScanNet scene" + ], + "image_footnote": [], + "bbox": [ + 513, + 95, + 648, + 214 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e9ddf982be8815f4919e16673089ab17a2b4ffbcf22e1f86d439a2176cd5a2d7.jpg", + "image_caption": [ + "(b) Visualization of ScanNet scene sampled at 1/10 of the original iterations", + "Figure 3. Motivation of Minimum Distance Curve Estimation" + ], + "image_footnote": [], + "bbox": [ + 656, + 98, + 776, + 215 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/53faa4f40d59824438cfe9fd48e63d59c0fc5b24b0ba8a64c09b93452036366c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 779, + 95, + 893, + 215 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "larly evident in Figure 3a, showing a smooth, decreasing curve for various input point clouds.", + "bbox": [ + 531, + 285, + 903, + 315 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2. Early structure capture: Importantly, we observe that a few initial sample points are sufficient to capture the overall shape and boundaries of the input point cloud (Figure 3b). This is because the initial sample points tend to be the extremities of the point cloud.", + "bbox": [ + 513, + 315, + 903, + 390 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Optimization Opportunities The trend of early structure capture suggests that the initial iterations of FPS are important in accurately representing the point cloud's structure. The Decreasing trend suggests that the later portion of the minimum distance curve becomes more predictable as sampling progresses. Building upon these trends, if we could accurately estimate this curve using only a few initial FPS iterations, we could maintain a comparable sampling quality while achieving substantial latency reduction by selecting points guided by the predicted curve.", + "bbox": [ + 511, + 409, + 906, + 560 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proposed Techniques We introduce two optimizations by leveraging these opportunities: Minimum Distance Prediction Sampling and Redundancy-Free Neighbor Search. Minimum Distance Prediction Sampling (MDPS) is a novel sampling strategy designed to approximate the sampling quality of FPS while significantly reducing latency. This approach estimates the distance curve using only a few initial non-parallelizable FPS iterations. Subsequent sample points are selected based on the predicted distances, eliminating the need for pairwise distance comparisons in later iterations. Furthermore, MDPS enables Redundancy-Free Neighbor Search by reusing precomputed distance information, thereby eliminating redundant computations during subsequent Ball Query and k-NN operations. This optimization further enhances the overall efficiency of PointNet++-based models during inference.", + "bbox": [ + 511, + 566, + 908, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Minimum Distance Prediction Sampling", + "text_level": 1, + "bbox": [ + 511, + 818, + 857, + 834 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "1 Minimum Distance Curve Estimation To build a low-cost estimator for the minimum distance curve (as shown in Figure 4), we aim to estimate the curve with minimal error using the results from as few FPS iterations as possible. We", + "bbox": [ + 511, + 839, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25117", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/29f6522d3450a20ddef2d4220cb9359d6f788d9dedb8a0cdcb5970978ad7f18b.jpg", + "image_caption": [ + "Figure 4. Overall Flow of Minimum Distance Prediction Sampling" + ], + "image_footnote": [], + "bbox": [ + 89, + 85, + 908, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "formulate the estimator $\\mathcal{E}$ as a function which gets initial $p$ segment (where $p$ is the ratio of initial FPS iterations to total iterations, $0 < p < 1$ ) of the curve $C$ as input and outputs the rest of the estimated curve $\\hat{C}$ (i.e., Equation 1). Note that $C(1)$ represents the minimum distance among the sampled points after all sampling iterations are finished.", + "bbox": [ + 88, + 387, + 482, + 479 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {C} (t) = \\mathcal {E} (C (t) \\mid t \\in [ 0, p ]) \\quad \\text {f o r} t \\in (p, 1 ] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 488, + 482, + 506 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We have experimented with various models, including polynomial functions and MLPs, and find that MLPs yield the best performance (Appendix A.3). We train a 3-layer MLP on minimum distance curves extracted from the training split of each point cloud dataset, setting $p = 0.1$ to balance error and efficiency. Our estimator achieves a low Mean Absolute Percentage Error (MAPE) of $1.02\\%$ , $0.77\\%$ , and $1.93\\%$ on S3DIS, ScanNet, and SemanticKITTI datasets, respectively, performing better on indoor datasets (S3DIS, ScanNet) due to their more consistent point density and distribution. Appendix B.6 and B.8 discuss ablation study for $p$ and cross-dataset applicability of estimators.", + "bbox": [ + 88, + 515, + 482, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\Theta$ Distance Curve Segmentation Ideally, what we want is to sample points whose minimum distance matches the value on the curve at each iteration. However, directly finding the point that yields the exact minimum distance in every iteration is computationally expensive.", + "bbox": [ + 88, + 704, + 482, + 780 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To efficiently sample points along the estimated minimum distance curve, we first divide the curve into multiple segments. Note that each segment is associated with a group of consecutive iterations. Then, the corresponding predicted minimum distances are employed to identify sample point at the segment-level granularity in the sampling process.", + "bbox": [ + 88, + 779, + 482, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As illustrated in Figure 4, the curve is segmented at iterations $d_{1}$ , $d_{2}$ , and $d_{3}$ , resulting in three segments. For", + "bbox": [ + 88, + 869, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "each input point, we identify points within a specified radius of each segment boundary (i.e., $R_{1}$ , $R_{2}$ , and $R_{3}$ ) and store them in the exclusion list for three segments. This list prevents sampling closely located points by considering the sampled points.", + "bbox": [ + 511, + 387, + 903, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To optimize computational efficiency, building the exclusion list across all segments is fused into a single GPU kernel. By calculating the distance between the points only once, regardless of the number of segments, we minimize the latency overhead. For example, the distance between P0 and P1 is computed only once. If P1 is within $R_{3}$ (i.e., $\\mathrm{dist(P0,P1)} < R_3 < R_2 < R_1$ ), P1 is added to the exclusion list of P0 at all segments. Refer to Table 3 for latency overhead with an increase in the segment count.", + "bbox": [ + 511, + 465, + 905, + 601 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Building the exclusion list still requires computing pairwise distances; however, this process is much faster than the original FPS algorithm. This improvement is achieved because all computations are fully parallelizable, as sampling and distance calculations are completely decoupled. For details on the algorithm, refer to Appendix A.1.", + "bbox": [ + 511, + 602, + 905, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sampling with Predicted Distance The sampling process employs the exclusion list. Starting from Segment 1, a seed point (P0 in Figure 4) is chosen. In each sampling iteration, points within a radius $R_{1}$ from the most recently sampled point are filtered out based on the exclusion list. We employ bitmaps to record which points are available for sampling. For instance, in the first iteration of Figure 4, Point P0, P1, P4, P5, P6, P7, and P11 are excluded based on the exclusion list of Segment 1. Point P3 is randomly selected among the eight available points for the second iteration in the same segment. Excluded points remain unavailable until the current segment is complete.", + "bbox": [ + 511, + 702, + 905, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As the sampling progresses and the segment changes", + "bbox": [ + 532, + 885, + 903, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25118", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/33c964742bb34613ffa0c942e141c996bba899b16190410740caa8d5d6692ef1.jpg", + "image_caption": [ + "Figure 5. Minimum spacing distribution comparison with increasing segment counts." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 486, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(e.g., from Segment 1 to Segment 2), the radius threshold is adjusted (e.g., from $R_{1}$ to $R_{2}$ ). This transition changes the set of points that are available for sampling. By iteratively applying this process, we ensure that the distance between sampled points meets or exceeds the threshold for each segment. This results in a minimum distance that lies above the blue step function in the segmented distance curve (i.e., $R_{i}$ serves as the lower bound of the minimum distance in Segment $i$ ). Increasing the number of segments allows the step function to approach the red minimum distance curve, improving sampling quality and converging towards the quality of FPS. Figure 5 demonstrates this convergence. With more segments, the distribution of minimum distance between sampled points becomes closer to that of FPS. Details on implementation are provided in Appendix A.1.", + "bbox": [ + 88, + 357, + 483, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\bullet$ Early Termination While the estimated minimum distance curve exhibits low error, overestimation can still negatively impact sampling quality. This is because it can lead to an excessive number of point exclusions, limiting the availability of points in later iterations. To mitigate this issue, we introduce Early Termination. This technique monitors the availability of points at each iteration. If no points remain, it advances to the next segment earlier than planned. When the final segment is exhausted, Sampling with Predicted Distance is terminated, and the remaining iterations are completed using the original FPS algorithm. For seamless transition to FPS, we efficiently update the minimum distance matrix (dists in Algorithm 1) by leveraging the exclusion list. This approach avoids recalculating all-to-all distances between points, significantly reducing the search space and minimizing the impact on sampling time. Appendix A.1 describes the details of Early Termination.", + "bbox": [ + 89, + 590, + 483, + 848 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Factors Contributing to Speedup The primary factors contributing to the latency of MDPS are:", + "bbox": [ + 89, + 854, + 482, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "1. Minimum Distance Curve Estimation: This step, requir", + "bbox": [ + 89, + 885, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b6010c9a380db52d887551a50b09d33c6045f13ae9f5532d0eaa12e43f001d09.jpg", + "image_caption": [ + "Figure 6. Redundancy Free Neighbor Search" + ], + "image_footnote": [], + "bbox": [ + 516, + 88, + 905, + 284 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ing 1/10 of the original FPS iterations, adds only a fraction of the original FPS latency.", + "bbox": [ + 532, + 325, + 903, + 356 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "2. Exclusion List Construction: The required all-to-all distance calculations are fully parallelizable, resulting in minimal latency overhead. Detailed analysis on this factor is provided in Appendix A.2.", + "3. Sampling Process: The use of exclusion list eliminates the need for additional distance calculations. Early termination requires additional FPS iterations. However, the impact is minimal: the overhead of $2.02\\%$ , $0.58\\%$ , and $1.18\\%$ of original FPS iterations in S3DIS, ScanNet, and SemanticKITTI dataset, respectively." + ], + "bbox": [ + 511, + 356, + 906, + 507 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Due to these optimizations, MDPS can achieve a $4 - 5 \\times$ speedup relative to the original FPS algorithm. Detailed latency breakdown of MDPS is provided in Appendix B.2.", + "bbox": [ + 511, + 507, + 905, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Redundancy-Free Neighbor Search", + "text_level": 1, + "bbox": [ + 511, + 561, + 821, + 578 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The exclusion list in MDPS captures spatial relationships between points, providing an opportunity to optimize subsequent Ball Query and k-NN operations. We introduce Redundancy Free Ball Query and k-NN, which fully leverage the exclusion list from the sampling stage.", + "bbox": [ + 511, + 584, + 905, + 660 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Redundancy-Free Ball Query Ball query, which identifies neighbors within a radius $R$ , is performed after sampling, before aggregation (refer to Figure 1). The exclusion list is well-suited for this task, as it contains points within a specified radius of each point. By adding an extra list to the exclusion list for the radius $R$ , as illustrated in Figure 6, we can directly use it for ball query. As discussed in Section 4.1, this approach minimizes computational overhead; distances between points are calculated only once, regardless of the number of segments. After sampling, we extract the relevant entries from the corresponding exclusion list for the ball query, which serve as centroids for aggregation.", + "bbox": [ + 511, + 666, + 906, + 849 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Redundancy-Free k-NN While directly reusing the exclusion list for k-NN is challenging, we can leverage it to reduce the search space. By using the exclusion list of Seg", + "bbox": [ + 511, + 854, + 905, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25119", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e0bc3bb64e6e077a2a4a4f3e867a8a159b5041e378f103e14ee67c98ccea8bd0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetAverage Minimum Distance (Ratio to Baseline)
Baseline (FPS)Random SamplingGrid SamplingMDPS (Ours)
S3DIS0.063950.03478 (54.38%)0.04971 (77.72%)0.06359 (99.45%)
ScanNet0.037010.02035 (54.99%)0.02400 (64.84%)0.03677 (99.35%)
SemanticKITTI0.255050.12775 (50.09%)0.21519 (84.37%)0.25087 (98.36%)
", + "bbox": [ + 99, + 88, + 470, + 203 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ment 1, we can limit the search space by eliminating points that are too far away to be potential k-NN neighbors. For example, to find the three nearest neighbors of point P7 in Figure 6, we calculate distances between P7 and the points from the corresponding exclusion list to identify the three nearest neighbors. Because k-NN upsampling relies on the downsampled set, points that were not downsampled are excluded from the search space. This approach significantly reduces the number of distance calculations, leading to substantial latency reduction.", + "bbox": [ + 89, + 250, + 483, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Evaluation", + "text_level": 1, + "bbox": [ + 89, + 414, + 207, + 430 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Methodology", + "text_level": 1, + "bbox": [ + 89, + 439, + 228, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate FastPoint on PointMetaBase [17] and PointVector [6], two recent models based on PointNet++. We evaluate our proposal on both indoor and outdoor datasets to demonstrate its scalability. We use S3DIS [1] and ScanNet [5] for indoor datasets and use SemanticKITTI [2] for outdoor datasets.", + "bbox": [ + 89, + 460, + 482, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The end-to-end latency is measured for processing all scenes in each dataset's validation split. We use mean Intersection over Union (mIoU) as a model accuracy evaluation metric. Detailed evaluation methodologies follow the same approach in PointMetaBase [17], and PointVector [6]. We apply FastPoint and the other baseline methods only to the first layer of each model as the first layer's FPS and neighbor search dominates the total computation time ( $>90\\%$ ).", + "bbox": [ + 89, + 551, + 482, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We implement FastPoint with custom CUDA kernels and integrate them into OpenPoints library. We also integrate the software implementation [11] of QuickFPS [12] opensourced by the authors to OpenPoints for comparison. The experiments are conducted using an NVIDIA Geforce RTX 3090 GPU with 24GB of memory. We use CUDA 11.8 and PyTorch 1.10.1 for software setup.", + "bbox": [ + 89, + 672, + 482, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Sampling Quality of MDPS", + "text_level": 1, + "bbox": [ + 89, + 787, + 336, + 804 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate sampling quality, we compute the average minimum distance between neighboring sampled points. A larger average distance indicates better coverage of sample points from the perspectives of point distribution and structure preservation. As shown in Table 1, MDPS achieves sampling quality comparable to FPS, with an average min", + "bbox": [ + 89, + 810, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/64f00d78e8eb052365a71882b47fdbeff936ad04d3aab4fa9668fe40f99b571b.jpg", + "table_caption": [ + "Table 1. Comparison of Sampling Quality among FPS, Random Sampling, Grid Sampling and MDPS." + ], + "table_footnote": [], + "table_body": "
ModelDatasetmIoU (Diff. to Baseline FPS)
Baseline (FPS)Random SamplingGrid SamplingMDPS (Ours)
PV-LS3DIS71.3364.83 (-6.5)70.97 (-0.36)71.37 (+0.04)
ScanNet70.7059.94 (-10.76)69.61 (-1.09)70.63 (-0.07)
Semantic KITTI50.9138.46 (-12.45)50.62 (-0.29)50.79 (-0.12)
PMB-LS3DIS69.7268.21 (-1.51)69.54 (-0.18)69.74 (+0.02)
ScanNet70.8667.66 (-3.2)70.36 (-0.50)70.89 (+0.03)
Semantic KITTI52.1947.26 (-4.93)52.27 (+0.08)52.09 (-0.10)
", + "bbox": [ + 517, + 88, + 895, + 275 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Accuracy comparison. PV and PMB stand for PointVector and PointMetaBase.", + "bbox": [ + 511, + 280, + 903, + 308 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "imum distance reaching over $99\\%$ of the FPS baseline for S3DIS and ScanNet, and over $98\\%$ for SemanticKITTI. In contrast, Random Sampling and Grid Sampling [29] exhibit significantly lower quality, with their average minimum distances deviating substantially from the FPS baseline. We also compare the minimum spacing distribution of each sampling method via visualization in Appendix B.9 and evaluate the robustness of MDPS in Appendix B.7.", + "bbox": [ + 511, + 321, + 906, + 444 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Accuracy Impact of MDPS", + "text_level": 1, + "bbox": [ + 511, + 452, + 756, + 469 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We analyze the impact of MDPS on accuracy by applying it to the inference stage of PointNet++ models initially trained with FPS. Since MDPS is designed to closely replicate the FPS sampling pattern, it is compatible with FPS-trained models, thereby minimizing accuracy degradation. Note that Redundancy-Free Neighbor Search does not introduce any approximation, thus has no impact on accuracy. To compare the effects of different sampling methods, we also evaluate random sampling and grid sampling1.", + "bbox": [ + 511, + 474, + 906, + 611 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2 shows that MDPS excels in maintaining accuracy, with negligible differences observed across all datasets and models. This highlights its effectiveness in replicating FPS behavior. However, the grid sampling exhibits larger and less consistent accuracy drops. When applied only during inference, it leads to declines of up to $1.09\\%$ (PointVector-L, ScanNet). While it outperforms FastPoint in a single instance (PointMetaBase-L, SemanticKITTI), its performance is generally inferior, with significant accuracy reductions in other scenarios. The random sampling shows considerable accuracy drops across all data points.", + "bbox": [ + 511, + 611, + 905, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Overall, FastPoint effectively preserves accuracy by closely approximating the sampling pattern of FPS, achieving accuracy levels near-identical to FPS without retraining. Latency-accuracy comparisons (Appendix B.5) show that FastPoint resides above the pareto front, emphasizing the effectiveness of FastPoint.", + "bbox": [ + 511, + 777, + 905, + 867 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "1Grid sampling is applied only during inference. Appendix B.1 shows results for cases where both training and inference utilize grid sampling.", + "bbox": [ + 511, + 875, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25120", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/57815b22f7209fe1dbda940f928fc4136274f42c75d81735e2b475863bf9448a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodNumber of SegmentsEvaluation Metrics
Accuracy (mIoU)Sampling Time (ms)Sampling Quality
FPS-70.7085.830.03701
MDPS (Ours)1 Segment70.3418.620.03532
2 Segments70.6019.430.03598
3 Segments70.4619.790.03636
4 Segments70.5421.010.03662
5 Segments70.5221.410.03665
6 Segments70.6322.060.03662
7 Segments70.6322.750.03664
", + "bbox": [ + 114, + 88, + 459, + 227 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8a4407dc43eaa96414579591ba55d0eb1d1540328aa6b50592e9a3a806c6e316.jpg", + "image_caption": [ + "Figure 7. End-to-end speedup of FastPoint and QuickFPS." + ], + "image_footnote": [], + "bbox": [ + 91, + 301, + 478, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Latency Reduction of FastPoint", + "text_level": 1, + "bbox": [ + 89, + 513, + 370, + 529 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To evaluate the impact of FastPoint on end-to-end latency, we compare its latency to baseline FPS and QuickFPS. Note that FastPoint includes both MDPS and Redundancy-Free Neighbor Search. As explained in Section 2.2, QuickFPS accelerates FPS using k-d tree without approximation. Since MDPS also contains FPS steps when predicting the minimum distance curve, we can achieve further latency reductions by replacing them with QuickFPS, meaning that FastPoint and QuickFPS can be synergetic.", + "bbox": [ + 89, + 537, + 483, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 7 presents the results of this comparison. Across S3DIS, ScanNet, and SemanticKITTI datasets, FastPoint achieves substantial latency reductions, with a geomean end-to-end speedup of $2.55 \\times$ compared to baseline FPS. Integrating QuickFPS with FastPoint further reduces the execution latency achieving a geomean end-to-end speedup of $2.76 \\times$ compared to baseline FPS and $1.41 \\times$ compared to standalone QuickFPS. Speedup specific to the sampling and neighbor search are provided in Appendix B.3 and B.4.", + "bbox": [ + 89, + 672, + 482, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "These findings substantiate the effectiveness of Fast-Point in reducing latency, surpassing existing methods like QuickFPS. Moreover, the successful combination of Fast-Point with QuickFPS demonstrates its compatibility with existing FPS optimizations. Scalability to Non-PointNet++ based models are discussed in Appendix B.10.", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ba498ebc9c248fbc0d55b074d03ba960e8c375fbec16f2000683e9a26e2bbcaf.jpg", + "table_caption": [ + "Table 3. Comparison of mIoU, sampling time, and sampling quality for FPS and MDPS across segments 1 to 7. Sampling time is measured with a single scene of ScanNet dataset. PointVector-L model is used for mIoU comparison." + ], + "table_footnote": [ + "(a) Comparison of component-wise end-to-end speedup of FastPoint." + ], + "table_body": "
ModelDatasetMethodSpeedup
PV-LS3DISMDPS1.90×
All1.98×
ScanNetMDPS2.33×
All2.52×
Semantic KITTIMDPS2.07×
All2.20×
PMB-LS3DISMDPS2.50×
All2.82×
ScanNetMDPS2.89×
All3.38×
Semantic KITTIMDPS2.31×
All2.62×
", + "bbox": [ + 514, + 87, + 717, + 224 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f021a8229ed5295838f67bbc80a2b5332176aa25625a0787552ac22c2c323c8a.jpg", + "table_caption": [], + "table_footnote": [ + "(b) Comparison of speedup as the number of points increases." + ], + "table_body": "
Model# of PointsSpeedup
PV-L16,0001.33×
32,0002.41×
48,0002.68×
64,0003.01×
80,0003.25×
96,0003.40×
PMB-L16,0001.58×
32,0002.64×
48,0002.91×
64,0003.06×
80,0003.30×
96,0003.44×
", + "bbox": [ + 717, + 87, + 895, + 224 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation study for speedup breakdown and scalability.", + "bbox": [ + 519, + 256, + 895, + 270 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Ablation Study", + "text_level": 1, + "bbox": [ + 513, + 289, + 666, + 305 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Segment Count We evaluate the impact of varying the segment count on accuracy, sampling time, and sampling quality (i.e., average minimum distance). Table 3 shows that while sampling time increases gradually with more segments due to the exclusion list construction (refer to Section 4.1), both sampling quality and accuracy improve. However, they get saturated beyond a certain number of segments. Based on the empirical analysis, we use 6 segments by default, where accuracy saturates for all workloads.", + "bbox": [ + 511, + 310, + 903, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Component-wise Speedup To understand the individual contributions of each component in FastPoint, we evaluate the end-to-end speedup by sequentially applying MDPS and Redundancy Free Neighbor Search. Table 4a shows that MDPS alone provides significant speedup. Incorporating Redundancy Free Neighbor Search further provides additional improvement, demonstrating that each component makes a distinct contribution to the overall speedup.", + "bbox": [ + 511, + 450, + 905, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.6. Scalability Study", + "text_level": 1, + "bbox": [ + 511, + 580, + 679, + 598 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To assess the scalability of FastPoint, we evaluate the end-to-end speedup compared to FPS by varying point cloud sizes. Table 4b shows that FastPoint consistently achieves higher speedups as the number of points increases. This trend highlights the efficiency of our approach, particularly when processing larger point clouds where the computational cost of the FPS becomes a significant bottleneck.", + "bbox": [ + 511, + 603, + 905, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 724, + 633, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3D point clouds are critical for representing and understanding 3D scenes, and deep learning models like PointNet++ have shown great promise in processing this data. However, computational efficiency remains a key challenge. We propose FastPoint, a software acceleration technique that leverages predictable minimum distances between sampled points to efficiently identify subsequent points without exhaustive computations. Our evaluation shows a $2.55 \\times$ speedup over baseline FPS while maintaining accuracy, enabling low-cost deployment of PointNet++ based models.", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25121", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 259, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the Ministry of Science and ICT (MSIT), Korea, under the following grants: the Global Scholars Invitation Program (RS-2024-00456287), the Graduate School of Artificial Intelligence Semiconductor grant (IITP-2025-RS-2023-00256081), and the Artificial Intelligence Innovation Hub (No. 2021-0-02068), all supervised by the Institute for Information & Communications Technology Planning & Evaluation (IITP). Additional support was provided by the National Research Foundation of Korea (NRF) grant (RS-2024-00405857) funded by the Korea government (MSIT), and the Google Faculty Research Award. The source code is available at https://github.com/SNU-ARC/FastPoint.git. Jae W. Lee and Hongil Yoon are the corresponding authors.", + "bbox": [ + 89, + 114, + 485, + 328 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 343, + 187, + 359 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Iro Armeni, Ozan Sener, Amir R. Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In CVPR, 2016. 7", + "[2] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Juergen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In ICCV, 2019. 7", + "[3] Yukang Chen, Jianhui Liu, Xiangyu Zhang, Xiaojuan Qi, and Jiaya Jia. Largekernel3d: Scaling up kernels in 3d sparse cnns. In CVPR, 2023. 2", + "[4] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, 2019. 1, 2", + "[5] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 7", + "[6] Xin Deng, WenYu Zhang, Qing Ding, and XinMing Zhang. Pointvector: A vector representation in point cloud analysis. In CVPR, 2023. 1, 2, 7", + "[7] Lue Fan, Ziqi Pang, Tianyuan Zhang, Yu-Xiong Wang, Hang Zhao, Feng Wang, Naiyan Wang, and Zhaoxiang Zhang. Embracing Single Stride 3D Object Detector with Sparse Transformer. In CVPR, 2022. 1, 2", + "[8] Yu Feng, Boyuan Tian, Tiancheng Xu, Paul Whatmough, and Yuhao Zhu. Mesorasi: Architecture support for point cloud analytics via delayed-aggregation. In Proceedings of the 53th International Symposium on Microarchitecture (MICRO), 2020. 1", + "[9] Yu Feng, Gunnar Hammonds, Yiming Gan, and Yuhao Zhu. Crescent: Taming memory irregularities for accelerating deep point cloud analytics. In Proceedings of the 49th Annual International Symposium on Computer Architecture (ISCA), 2022. 1", + "[10] Lei Han, Tian Zheng, Lan Xu, and Lu Fang. Occuseg: Occupancy-aware 3d instance segmentation. In CVPR, 2020. 2" + ], + "bbox": [ + 93, + 368, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Meng Han, Liang Wang, Limin Xiao, Hao Zhang, Chenhao Zhang, Xiangrong Xu, and Jianfeng Zhu. Quickfps. http://github.com/hanm2019/bucket-based_farthest-point-sampling_GPU.2,7", + "[12] Meng Han, Liang Wang, Limin Xiao, Hao Zhang, Chenhao Zhang, Xiangrong Xu, and Jianfeng Zhu. Quickfps: Architecture and algorithm co-design for farthest point sampling in large-scale point clouds. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 2023. 1, 2, 7", + "[13] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. In CVPR, 2020. 1, 2", + "[14] Donghyun Lee, Yejin Lee, Jae W. Lee, and Hongil Yoon. Frugal 3d point cloud model training via progressive near point filtering and fused aggregation. In ECCV, 2024. 3", + "[15] Jingtao Li, Jian Zhou, Yan Xiong, Xing Chen, and Chaitali Chakrabarti. An adjustable farthest point sampling method for approximately-sorted point cloud data. In 2022 IEEE Workshop on Signal Processing Systems (SiPS), 2022. 1, 3", + "[16] Yaoxiu Lian, Xinhao Yang, Ke Hong, Yu Wang, Guohao Dai, and Ningyi Xu. A point transformer accelerator with fine-grained pipelines and distribution-aware dynamic fps. In IEEE/ACM International Conference on Computer-Aided Design (ICCAD), 2023. 1, 2", + "[17] Haojia Lin, Xiawu Zheng, Lijiang Li, Fei Chao, Shanshan Wang, Yan Wang, Yonghong Tian, and Rongrong Ji. Meta architecture for point cloud analysis. In CVPR, 2023. 1, 2, 3, 7", + "[18] Xueyuan Liu, Zhuoran Song, Guohao Dai, Gang Li, Can Xiao, Yan Xiang, Dehui Kong, Ke Xu, and Xiaoyao Liang. Fusionarch: A fusion-based accelerator for point-based point cloud neural networks. In ACM/IEEE Design, Automation and Test in Europe Conference (DATE), 2024. 1", + "[19] Yongcheng Liu, Bin Fan, Gaofeng Meng, Jiwen Lu, Shiming Xiang, and Chunhong Pan. Densepoint: Learning densely contextual representation for efficient point cloud processing. In ICCV, 2019. 2", + "[20] Zhijian Liu, Xinyu Yang, Haotian Tang, Shang Yang, and Song Han. Flatformer: Flattened window attention for efficient point cloud transformer. In CVPR, 2023. 1", + "[21] Jiageng Mao, Yujing Xue, Minzhe Niu, Haoyue Bai, Jiashi Feng, Xiaodan Liang, Hang Xu, and Chunjing Xu. Voxel transformer for 3d object detection. In ICCV, 2021. 2", + "[22] Alexey Nekrasov, Jonas Schult, Or Litany, Bastian Leibe, and Francis Engelmann. Mix3D: Out-of-Context Data Augmentation for 3D Scenes. In International Conference on 3D Vision (3DV), 2021. 2", + "[23] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. arXiv preprint arXiv:1612.00593, 2016. 1, 2", + "[24] Charles R Qi, Li Yi, Hao Su, and Leonidas J Guibas. Point-net++: Deep hierarchical feature learning on point sets in a metric space. arXiv preprint arXiv:1706.02413, 2017. 1, 2" + ], + "bbox": [ + 514, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25122", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Charles R. Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas J. Guibas. Frustum pointnets for 3d object detection from rgb-d data. In CVPR, 2018.", + "[26] Guocheng Qian, Hasan Hammoud, Guohao Li, Ali Thabet, and Bernard Ghanem. Assanet: An anisotropical separable set abstraction for efficient point cloud representation learning. In NeurIPS, 2021. 1", + "[27] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. In NeurIPS, 2022. 2", + "[28] Danila Rukhovich, Anna Vorontsova, and Anton Konushin. Imvoxelnet: Image to voxels projection for monocular and multi-view general-purpose 3d object detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2022. 2", + "[29] Hugues Thomas, Charles R. Qi, Jean-Emmanuel Deschaud, Beatriz Marcotegui, François Goulette, and Leonidas J. Guibas. Kpconv: Flexible and deformable convolution for point clouds. In ICCV, 2019. 2, 7", + "[30] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E. Sarma, Michael M. Bronstein, and Justin M. Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (TOG), 2019. 1, 2", + "[31] Wenxuan Wu, Zhongang Qi, and Li Fuxin. Pointconv: Deep convolutional networks on 3d point clouds. In CVPR, 2019. 2", + "[32] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. In NeurIPS, 2022. 1", + "[33] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler, faster, stronger. In CVPR, 2024. 1", + "[34] Mutian Xu, Runyu Ding, Hengshuang Zhao, and Xiaojuan Qi. Paconv: Position adaptive convolution with dynamic kernel assembling on point clouds. In CVPR, 2021. 2", + "[35] Qiangeng Xu, Xudong Sun, Cho-Ying Wu, Panqu Wang, and Ulrich Neumann. Grid-gen for fast and scalable point cloud learning, 2020. 2", + "[36] Xinhao Yang, Tianyu Fu, Guohao Dai, Shulin Zeng, Kai Zhong, Ke Hong, and Yu Wang. An efficient accelerator for point-based and voxel-based point cloud neural networks. In ACM/IEEE Design Automation Conference (DAC), 2023. 1, 2", + "[37] Yu-Qi Yang, Yu-Xiao Guo, Jian-Yu Xiong, Yang Liu, Hao Pan, Peng-Shuai Wang, Xin Tong, and Baining Guo. Swin3d: A pretrained transformer backbone for 3d indoor scene understanding. arXiv preprint arXiv:2304.06906, 2023.1,2", + "[38] Yu-Qi Yang, Yu-Xiao Guo, and Yang Liu. Swin3d++: Effective multi-source pretraining for 3d indoor scene understanding. arXiv preprint arXiv:2304.06906, 2024. 2", + "[39] Ziyu Ying, Sandeepa Bhuyan, Yan Kang, Yingtian Zhang, Mahmut T. Kandemir, and Chita R. Das. Edgepc: Efficient deep learning analytics for point clouds on edge devices. In Proceedings of the 50th Annual International Symposium on Computer Architecture (ISCA), 2023. 1, 2" + ], + "bbox": [ + 89, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Hyunsung Yoon and Jae-Joon Kim. Efficient sampling and grouping acceleration for point cloud deep learning via single coordinate comparison. In IEEE/ACM International Conference on Computer-Aided Design (ICCAD), 2023. 1", + "[41] Hengshuang Zhao, Li Jiang, Chi-Wing Fu, and Jiaya Jia. PointWeb: Enhancing local neighborhood features for point cloud processing. In CVPR, 2019. 2", + "[42] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip Torr, and Vladlen Koltun. Point transformer. In ICCV, 2021. 1", + "[43] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. arXiv preprint arXiv:2011.10033, 2020. 1, 2" + ], + "bbox": [ + 514, + 90, + 905, + 276 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25123", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/FastPoint_ Accelerating 3D Point Cloud Model Inference via Sample Point Distance Prediction/670cb46f-786b-455d-998f-328c26d7e47d_model.json b/2025/FastPoint_ Accelerating 3D Point Cloud Model Inference via Sample Point Distance Prediction/670cb46f-786b-455d-998f-328c26d7e47d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..88abc1871e062d92d73dbcbd280bd4e223413798 --- /dev/null +++ b/2025/FastPoint_ Accelerating 3D Point Cloud Model Inference via Sample Point Distance Prediction/670cb46f-786b-455d-998f-328c26d7e47d_model.json @@ -0,0 +1,2354 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.197, + 0.131, + 0.804, + 0.175 + ], + "angle": 0, + "content": "FastPoint: Accelerating 3D Point Cloud Model Inference via Sample Point Distance Prediction" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.203, + 0.77, + 0.24 + ], + "angle": 0, + "content": "Donghyun Lee\\(^{1}\\) Dawoon Jeong\\(^{1}\\) Jae W. Lee\\(^{1}\\) Hongil Yoon\\(^{1,2}\\) Seoul National University \\({ }^{2}\\)Google" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.242, + 0.776, + 0.258 + ], + "angle": 0, + "content": "{eudh1206, daun20211, jaewlee}@snu.ac.kr, hongilyoon@google.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.325, + 0.486, + 0.538 + ], + "angle": 0, + "content": "Deep neural networks have revolutionized 3D point cloud processing, yet efficiently handling large and irregular point clouds remains challenging. To tackle this problem, we introduce FastPoint, a novel software-based acceleration technique that leverages the predictable distance trend between sampled points during farthest point sampling. By predicting the distance curve, we can efficiently identify subsequent sample points without exhaustively computing all pairwise distances. Our proposal substantially accelerates farthest point sampling and neighbor search operations while preserving sampling quality and model performance. By integrating FastPoint into state-of-the-art 3D point cloud models, we achieve \\(2.55 \\times\\) end-to-end speedup on NVIDIA RTX 3090 GPU without sacrificing accuracy." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.57, + 0.222, + 0.587 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.483, + 0.779 + ], + "angle": 0, + "content": "3D point clouds have become increasingly important for representing and understanding 3D scenes, driving advancements in fields like robotics and autonomous driving. Deep neural networks [4, 6, 7, 13, 17, 20, 23-26, 30, 32, 33, 37, 42, 43] have emerged as powerful tools for processing this type of data. While PointNet++ [24] and its successors have improved efficiency and performance of these point cloud models, the irregular nature and growing size of point cloud data continue to impose significant computational challenges, especially for real time use cases. Operations such as Farthest Point Sampling (FPS) and neighbor search remain major bottlenecks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "While various hardware accelerators [8, 9, 12, 16, 18, 36, 40] have been explored to accelerate the models, software-only solutions are less common. Given the high cost and complexity of developing specialized accelerators, software-based acceleration is a more practical approach. Recently, several software-based techniques [15, 39] have been proposed to accelerate these models through approximation. However, such approaches often come at the cost" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.293, + 0.879, + 0.308 + ], + "angle": 0, + "content": "of significant accuracy loss, limiting their practical use." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.309, + 0.906, + 0.399 + ], + "angle": 0, + "content": "To address this challenge, we propose FastPoint, a novel software-based acceleration technique with sample point distance prediction. We observe that the inherent nature of FPS, which iteratively selects the point farthest from the current set of sampled points, leads to two predictable trends in the minimum distances between these points:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.4, + 0.906, + 0.476 + ], + "angle": 0, + "content": "1. Decreasing minimum distance: As FPS makes progress, the number of remaining points decreases, forcing the algorithm to select points closer to those already sampled. This results in a smoothly decreasing curve of minimum distances between sampled points." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.476, + 0.905, + 0.52 + ], + "angle": 0, + "content": "2. Early structure capture: The initial sample points tend to be the extremities of the point cloud, effectively capturing its overall shape and boundaries." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.521, + 0.906, + 0.732 + ], + "angle": 0, + "content": "Our proposal leverages these observations to accurately estimate the distance curve using only a few initial FPS iterations. By predicting the distance curve, we can efficiently identify subsequent sample points without computing and comparing pairwise distances at each FPS iteration. This leads to significant latency savings while maintaining sampling quality comparable to that of the original FPS. Furthermore, predicting the distance curve enables the sampling process to be decoupled from distance computation. This decoupling exposes new opportunities for parallelizing distance calculations in the proposed sampling, further enhancing efficiency. This approach not only accelerates FPS itself but also benefits subsequent operations such as neighbor search by leveraging pre-computed distances." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.733, + 0.864, + 0.746 + ], + "angle": 0, + "content": "Our key contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.75, + 0.905, + 0.81 + ], + "angle": 0, + "content": "- We identify two key latency bottlenecks in point cloud models: FPS and neighbor search. Limited parallelism in FPS and redundant distance computations across both operations are the primary sources of inefficiency." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.811, + 0.905, + 0.841 + ], + "angle": 0, + "content": "- We empirically analyze 3D point cloud models and identify the decreasing trend and early structure capture trend." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.841, + 0.905, + 0.901 + ], + "angle": 0, + "content": "- Based on the observations, we introduce a novel sampling approach with sample point distance prediction. With the estimated distance curve, this approach not only accelerates FPS itself but also benefits subsequent neighbor" + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.75, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25114" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.912, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.314, + 0.28, + 0.685, + 0.296 + ], + "angle": 0, + "content": "Figure 1. Overview of PointNet++ Based Model Architecture" + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.313, + 0.423, + 0.327 + ], + "angle": 0, + "content": "search, which relies on pre-computed distances." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.328, + 0.484, + 0.389 + ], + "angle": 0, + "content": "- We integrate FastPoint into OpenPoints library and experiment on state-of-the-art PointNet++ based models, achieving geomean \\(2.55 \\times\\) end-to-end speedup on NVIDIA RTX 3090 GPU while preserving accuracy." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.399, + 0.379, + 0.416 + ], + "angle": 0, + "content": "2. Background and Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.424, + 0.396, + 0.44 + ], + "angle": 0, + "content": "2.1. Deep Learning on 3D Point Clouds" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.446, + 0.484, + 0.702 + ], + "angle": 0, + "content": "Deep Neural Networks on Raw Point Cloud Since the release of PointNet [23], which is the first to apply deep neural networks directly to the raw point clouds, numerous successors [6, 13, 17, 19, 24-27, 30, 31, 34, 41] have been developed, improving both model performance and efficiency. PointNet++ [24] introduces a hierarchical structure to PointNet through sampling and grouping, which has become the core architecture of subsequent PointNet++ based models. PointNeXt [27] revisits PointNet++, achieving substantial improvements in model performance, while PointVector [6] and PointMetaBase [17] push the model performance and efficiency even further. Although the architectural details of each model differ, all models follow the PointNet++ structure, utilizing Set Abstraction and Feature Propagation layers as core components. This work focuses on identifying and accelerating the primary latency bottlenecks in PointNet++ based models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Deep Neural Networks on Voxelized Point Cloud Several works [3, 4, 10, 22, 28, 43] propose to voxelize point clouds, using voxels as input rather than raw points. The advent of 3D sparse convolution [4] has made this voxel-based approach increasingly popular due to its efficiency and high performance. Recently, transformer-based models [7, 21, 37, 38] operating on voxelized point clouds are gaining significant attention. Although voxelization alleviates the computational cost of mapping operations (i.e., downsampling and neighbor search), it has a notable limitation due to the loss of position information during the voxelization process." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.311, + 0.659, + 0.327 + ], + "angle": 0, + "content": "2.2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.334, + 0.907, + 0.395 + ], + "angle": 0, + "content": "Numerous solutions have been proposed to accelerate sampling and neighbor search operations in PointNet++ based models. We categorize them into hardware- and purely software-based approaches." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.414, + 0.909, + 0.626 + ], + "angle": 0, + "content": "Hardware-Assisted Acceleration QuickFPS [12] introduces a k-d tree based FPS algorithm which reduces computation and memory access in each FPS iteration, along with specialized hardware designed for this method. MARS [36] and PTrAcc [16] both propose hardware accelerators that employ a distance filtering technique which skips unnecessary distance computations in each iteration of FPS. Although these techniques achieve substantial speedups in FPS without any loss in sampling quality, they require specialized hardware to fully harness their algorithms, limiting their effectiveness on commodity hardware like GPU. For more thorough evaluation, we compare the speedup of our proposal with pure software version [11] of QuickFPS in Section 5.4, evaluating its effectiveness on GPU." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.87 + ], + "angle": 0, + "content": "Software-Only Acceleration To mitigate the high latency associated with FPS, various alternative sampling methods have been explored. RandLA-Net [13] replaces FPS with random sampling and introduces a local feature aggregation module that compensates for the low sampling quality of random sampling. Grid sampling, the most commonly used alternative, is adopted by Grid-GCN [35] and KPCov [29], offering faster processing with reasonable sampling quality compared to random sampling. Despite the latency advantages of these methods, none have consistently outperformed FPS for overall performance. This is illustrated by the fact that most PointNet++-based models [6, 17, 24-27] continue to rely on FPS, highlighting its broad applicability and superior sampling quality. The experimental results in Appendix B.1 corroborate this claim." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.908, + 0.901 + ], + "angle": 0, + "content": "EdgePC [39] structures the point cloud with morton codes to accelerate FPS and neighbor search on edge GPUs," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25115" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.092, + 0.462, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.109, + 0.288, + 0.466, + 0.303 + ], + "angle": 0, + "content": "Figure 2. Latency Breakdown of PointNet++ Based Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.321, + 0.484, + 0.519 + ], + "angle": 0, + "content": "while Adjustable FPS [15] accelerates FPS by leveraging the intrinsic locality of the point cloud data. However, both methods demonstrate significant accuracy loss compared to the baseline (i.e., up to \\(2\\%\\) mIoU loss) due to their aggressive approximations, limiting their effectiveness. Prior work [14] accelerates the training of PointNet++ based models by precomputing the minimum point spacing of farthest-sampled points before training and reusing this value across epochs for fast sampling. While this approach effectively speeds up training, it is not suitable for inference, as the minimum spacing value cannot be determined in advance in inference scenarios. Detailed comparisons with our work are provided in Appendix A.4." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.535, + 0.4, + 0.551 + ], + "angle": 0, + "content": "3. Challenges of Point Cloud Models" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.56, + 0.357, + 0.575 + ], + "angle": 0, + "content": "3.1. Model Architecture Overview" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.583, + 0.483, + 0.674 + ], + "angle": 0, + "content": "Figure 1 illustrates the detailed process of PointNet++-based models. Similar to 2D convolutional neural networks, these models consist of both downsampling and upsampling layers. Commonly referred to as the set abstraction layer, the downsampling layer in PointNet++-based models comprises three main stages, as outlined below:" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.675, + 0.483, + 0.75 + ], + "angle": 0, + "content": "1 Farthest Point Sampling (FPS): Downsampling begins by selecting \\( N \\)/stride output points from \\( N \\) input points, where stride represents the downsampling rate. Typically, the FPS is used to maintain the shape and structure of the input point cloud." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.75, + 0.483, + 0.796 + ], + "angle": 0, + "content": "\\(\\bullet\\) Ball Query: Next, a ball query with radius \\(R\\) is performed around each sampled point to gather neighboring points for feature aggregation." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.796, + 0.483, + 0.9 + ], + "angle": 0, + "content": "Aggregation: The features of neighboring points identified in the ball query are then processed. A multi-layer perceptron (MLP) is applied to these features, followed by max pooling to aggregate the information for each centroid. The details of this process, e.g., positional embedding [17], can vary depending on the model architectures." + }, + { + "type": "list", + "bbox": [ + 0.102, + 0.675, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.091, + 0.766, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Farthest Point Sampling" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.109, + 0.895, + 0.136 + ], + "angle": 0, + "content": "Input \\(P\\) : input point cloud, \\(n\\) : number of points to sample, \\(N\\) : number of original points" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.137, + 0.8, + 0.15 + ], + "angle": 0, + "content": "Output sampledidx: indices of sampled points" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.152, + 0.896, + 0.166 + ], + "angle": 0, + "content": "1: sampled_idx[0] \\(\\leftarrow\\) seed_idx // Starts with a random point" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.167, + 0.773, + 0.179 + ], + "angle": 0, + "content": "2: min_dists[i] \\(\\leftarrow \\infty\\) for \\(i = 0,\\dots,N - 1\\)" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.181, + 0.798, + 0.193 + ], + "angle": 0, + "content": "3: for \\( i \\gets 1 \\) to \\( n - 1 \\) do // Not parallelized" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.194, + 0.801, + 0.207 + ], + "angle": 0, + "content": "4: for \\(j\\gets 0\\) to \\(N - 1\\) do // Parallelized" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.208, + 0.857, + 0.222 + ], + "angle": 0, + "content": "5: \\(d_{\\text{new}} \\gets \\text{dist}(P[\\text{sampled\\_idx}[i - 1]], P[j])\\)" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.223, + 0.754, + 0.235 + ], + "angle": 0, + "content": "6: if min_dists[j] > \\(d_{new}\\) then" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.236, + 0.734, + 0.249 + ], + "angle": 0, + "content": "7: min_dists[j] \\(\\leftarrow d_{new}\\)" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.251, + 0.793, + 0.265 + ], + "angle": 0, + "content": "8: sampledidx[i] \\(\\leftarrow\\) argmax(min_dists)" + }, + { + "type": "list", + "bbox": [ + 0.522, + 0.152, + 0.896, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.311, + 0.905, + 0.385 + ], + "angle": 0, + "content": "The upsampling layer, often referred to as a feature propagation layer, takes the downsampled points from the corresponding downsampling layer as inputs and propagates the features back to the original point cloud. This process is done by following steps." + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.386, + 0.905, + 0.416 + ], + "angle": 0, + "content": "\\(\\bullet\\) k-NN: For each upsampled point, k-NN is applied to identify its k nearest neighbors among the input points." + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.416, + 0.905, + 0.492 + ], + "angle": 0, + "content": "\\(\\Theta\\) Interpolation: Based on the k-NN results, neighboring input points are interpolated, concatenated with skip-connected features from the corresponding downsampling layer, and the information is passed through an MLP to propagate the features to the subsequent layer." + }, + { + "type": "list", + "bbox": [ + 0.525, + 0.386, + 0.905, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.501, + 0.706, + 0.516 + ], + "angle": 0, + "content": "3.2. Latency Breakdown" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.523, + 0.907, + 0.765 + ], + "angle": 0, + "content": "Figure 2 presents the latency breakdown of PointMetaBase-L model evaluated on both indoor (S3DIS) and outdoor (SemanticKITTI) scenes. The results show that coordinate-based operations (i.e., FPS, k-NN, and Ball Query) dominate the execution time, consuming significantly more time than feature-based operations (i.e., Aggregation and MLP). These coordinate-based operations account for approximately \\(88.1\\%\\) of the total inference time for indoor scenes and \\(94.1\\%\\) for the outdoor scenes. Among these, FPS emerges as the primary performance bottleneck, requiring the most execution time. Additionally, neighbor search operations, including k-NN and Ball Query, contribute significantly to the overall latency. To address these challenges, we first analyze the root causes of these inefficiencies and propose novel optimization techniques in subsequent sections." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.773, + 0.861, + 0.79 + ], + "angle": 0, + "content": "3.3. Inefficiencies in Farthest Point Sampling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.9 + ], + "angle": 0, + "content": "Farthest point sampling operates iteratively by selecting one point at a time. In each iteration, the sampling algorithm identifies the point farthest from the set of the previously sampled points. The detailed process is outlined in Algorithm 1. Initially, a random point in \\( P \\) is chosen as the seed point (Line 1) and added to the sampled_idx. Then, a distance matrix (i.e., min_dists) is created to store the distance" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "25116" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.272 + ], + "angle": 0, + "content": "between each point in \\( P \\) and the set of the previously sampled points, where the distance from a point to a point set is defined as minimum distance between the point and any point in the set. This matrix is initialized to \\( \\infty \\) (Line 2). In each iteration, the distance between the most recently sampled point and every other point in \\( P \\) is calculated (Lines 3-7). If a newly calculated distance is smaller than the corresponding entry in the min_dist matrix, the matrix is updated with the smaller value. Finally, the index of the point with the maximum distance in the min_dist is added to the sampled_idx (Line 8). This process is repeated until the specified number of points \\( (n) \\) is sampled." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.273, + 0.482, + 0.378 + ], + "angle": 0, + "content": "The sequential nature of the FPS algorithm is the primary reason of the inefficiency. Because only one point is sampled per iteration, the \\( O(N) \\) distance calculations required in each iteration must be sequentially executed \\( n \\) times. This tight coupling between sampling and distance calculation significantly limits the potential for parallelism in distance computations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.379, + 0.482, + 0.454 + ], + "angle": 0, + "content": "To unlock this parallelism potential, we introduce Minimum Distance Prediction Sampling (MDPS) in Section 4.1, a novel sampling strategy that fully decouples sampling from distance computation, enabling greater parallelism and faster processing speeds." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.462, + 0.38, + 0.478 + ], + "angle": 0, + "content": "3.4. Inefficiencies in Neighbor Search" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.484, + 0.482, + 0.59 + ], + "angle": 0, + "content": "As explained in Section 3.1, PointNet++ based models employ two types of neighbor search algorithms: Ball Query and k-NN. These operations play a critical role in capturing local geometric relationships in the point cloud. Ball Query identifies neighbor points within a fixed radius around the downsampled points. k-NN identifies k nearest neighbors among the downsampled points for each original point." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.59, + 0.482, + 0.711 + ], + "angle": 0, + "content": "The inefficiencies in these neighbor search operations stem from redundant distance computations in both the FPS and subsequent neighbor search operations. For instance, as shown in Figure 1, the distance between P0 and P2 is computed while updating P2's minimum distance in FPS. However, the same distance is redundantly recalculated in both Ball Query and k-NN. Section 4.2 introduces redundancy-free neighbor search techniques to address this issue." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.723, + 0.433, + 0.739 + ], + "angle": 0, + "content": "4. FastPoint: Fast Point Cloud Inference" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.748, + 0.482, + 0.809 + ], + "angle": 0, + "content": "Trends in Minimum Distance of Sample Points The FPS iteratively selects the point farthest from the current set of sampled points, which leads to predictable trends in the distance between the sampled points." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.811, + 0.482, + 0.901 + ], + "angle": 0, + "content": "1. Decreasing minimum distance: As sampling proceeds, the pool of the remaining points shrinks, the FPS algorithm selects points closer to those already sampled over time. Thus, the maximized minimum distance (i.e., \\(\\max (min\\_dist)\\) in Algorithm 1) decreases, which we define as minimum distance for brevity. This is particu" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.097, + 0.65, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.217, + 0.648, + 0.242 + ], + "angle": 0, + "content": "(a) Minimum Distance Curve of ScanNet scene" + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.099, + 0.777, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.651, + 0.217, + 0.906, + 0.242 + ], + "angle": 0, + "content": "(b) Visualization of ScanNet scene sampled at 1/10 of the original iterations" + }, + { + "type": "image", + "bbox": [ + 0.78, + 0.097, + 0.895, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.254, + 0.893, + 0.268 + ], + "angle": 0, + "content": "Figure 3. Motivation of Minimum Distance Curve Estimation" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.286, + 0.905, + 0.316 + ], + "angle": 0, + "content": "larly evident in Figure 3a, showing a smooth, decreasing curve for various input point clouds." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.316, + 0.905, + 0.391 + ], + "angle": 0, + "content": "2. Early structure capture: Importantly, we observe that a few initial sample points are sufficient to capture the overall shape and boundaries of the input point cloud (Figure 3b). This is because the initial sample points tend to be the extremities of the point cloud." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.907, + 0.561 + ], + "angle": 0, + "content": "Optimization Opportunities The trend of early structure capture suggests that the initial iterations of FPS are important in accurately representing the point cloud's structure. The Decreasing trend suggests that the later portion of the minimum distance curve becomes more predictable as sampling progresses. Building upon these trends, if we could accurately estimate this curve using only a few initial FPS iterations, we could maintain a comparable sampling quality while achieving substantial latency reduction by selecting points guided by the predicted curve." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.909, + 0.81 + ], + "angle": 0, + "content": "Proposed Techniques We introduce two optimizations by leveraging these opportunities: Minimum Distance Prediction Sampling and Redundancy-Free Neighbor Search. Minimum Distance Prediction Sampling (MDPS) is a novel sampling strategy designed to approximate the sampling quality of FPS while significantly reducing latency. This approach estimates the distance curve using only a few initial non-parallelizable FPS iterations. Subsequent sample points are selected based on the predicted distances, eliminating the need for pairwise distance comparisons in later iterations. Furthermore, MDPS enables Redundancy-Free Neighbor Search by reusing precomputed distance information, thereby eliminating redundant computations during subsequent Ball Query and k-NN operations. This optimization further enhances the overall efficiency of PointNet++-based models during inference." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.819, + 0.859, + 0.835 + ], + "angle": 0, + "content": "4.1. Minimum Distance Prediction Sampling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.901 + ], + "angle": 0, + "content": "1 Minimum Distance Curve Estimation To build a low-cost estimator for the minimum distance curve (as shown in Figure 4), we aim to estimate the curve with minimal error using the results from as few FPS iterations as possible. We" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25117" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.087, + 0.91, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.299, + 0.356, + 0.696, + 0.371 + ], + "angle": 0, + "content": "Figure 4. Overall Flow of Minimum Distance Prediction Sampling" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.388, + 0.483, + 0.48 + ], + "angle": 0, + "content": "formulate the estimator \\(\\mathcal{E}\\) as a function which gets initial \\(p\\) segment (where \\(p\\) is the ratio of initial FPS iterations to total iterations, \\(0 < p < 1\\)) of the curve \\(C\\) as input and outputs the rest of the estimated curve \\(\\hat{C}\\) (i.e., Equation 1). Note that \\(C(1)\\) represents the minimum distance among the sampled points after all sampling iterations are finished." + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.489, + 0.483, + 0.507 + ], + "angle": 0, + "content": "\\[\n\\hat {C} (t) = \\mathcal {E} (C (t) \\mid t \\in [ 0, p ]) \\quad \\text {f o r} t \\in (p, 1 ] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.516, + 0.483, + 0.699 + ], + "angle": 0, + "content": "We have experimented with various models, including polynomial functions and MLPs, and find that MLPs yield the best performance (Appendix A.3). We train a 3-layer MLP on minimum distance curves extracted from the training split of each point cloud dataset, setting \\( p = 0.1 \\) to balance error and efficiency. Our estimator achieves a low Mean Absolute Percentage Error (MAPE) of \\( 1.02\\% \\), \\( 0.77\\% \\), and \\( 1.93\\% \\) on S3DIS, ScanNet, and SemanticKITTI datasets, respectively, performing better on indoor datasets (S3DIS, ScanNet) due to their more consistent point density and distribution. Appendix B.6 and B.8 discuss ablation study for \\( p \\) and cross-dataset applicability of estimators." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.705, + 0.483, + 0.781 + ], + "angle": 0, + "content": "\\(\\Theta\\) Distance Curve Segmentation Ideally, what we want is to sample points whose minimum distance matches the value on the curve at each iteration. However, directly finding the point that yields the exact minimum distance in every iteration is computationally expensive." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.78, + 0.483, + 0.871 + ], + "angle": 0, + "content": "To efficiently sample points along the estimated minimum distance curve, we first divide the curve into multiple segments. Note that each segment is associated with a group of consecutive iterations. Then, the corresponding predicted minimum distances are employed to identify sample point at the segment-level granularity in the sampling process." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "As illustrated in Figure 4, the curve is segmented at iterations \\(d_{1}\\), \\(d_{2}\\), and \\(d_{3}\\), resulting in three segments. For" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.388, + 0.905, + 0.463 + ], + "angle": 0, + "content": "each input point, we identify points within a specified radius of each segment boundary (i.e., \\( R_{1} \\), \\( R_{2} \\), and \\( R_{3} \\)) and store them in the exclusion list for three segments. This list prevents sampling closely located points by considering the sampled points." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.466, + 0.906, + 0.602 + ], + "angle": 0, + "content": "To optimize computational efficiency, building the exclusion list across all segments is fused into a single GPU kernel. By calculating the distance between the points only once, regardless of the number of segments, we minimize the latency overhead. For example, the distance between P0 and P1 is computed only once. If P1 is within \\( R_{3} \\) (i.e., \\( \\mathrm{dist(P0,P1)} < R_3 < R_2 < R_1 \\)), P1 is added to the exclusion list of P0 at all segments. Refer to Table 3 for latency overhead with an increase in the segment count." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.603, + 0.906, + 0.695 + ], + "angle": 0, + "content": "Building the exclusion list still requires computing pairwise distances; however, this process is much faster than the original FPS algorithm. This improvement is achieved because all computations are fully parallelizable, as sampling and distance calculations are completely decoupled. For details on the algorithm, refer to Appendix A.1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.703, + 0.906, + 0.884 + ], + "angle": 0, + "content": "Sampling with Predicted Distance The sampling process employs the exclusion list. Starting from Segment 1, a seed point (P0 in Figure 4) is chosen. In each sampling iteration, points within a radius \\( R_{1} \\) from the most recently sampled point are filtered out based on the exclusion list. We employ bitmaps to record which points are available for sampling. For instance, in the first iteration of Figure 4, Point P0, P1, P4, P5, P6, P7, and P11 are excluded based on the exclusion list of Segment 1. Point P3 is randomly selected among the eight available points for the second iteration in the same segment. Excluded points remain unavailable until the current segment is complete." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "As the sampling progresses and the segment changes" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25118" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.488, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.311, + 0.486, + 0.342 + ], + "angle": 0, + "content": "Figure 5. Minimum spacing distribution comparison with increasing segment counts." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.358, + 0.485, + 0.584 + ], + "angle": 0, + "content": "(e.g., from Segment 1 to Segment 2), the radius threshold is adjusted (e.g., from \\( R_{1} \\) to \\( R_{2} \\)). This transition changes the set of points that are available for sampling. By iteratively applying this process, we ensure that the distance between sampled points meets or exceeds the threshold for each segment. This results in a minimum distance that lies above the blue step function in the segmented distance curve (i.e., \\( R_{i} \\) serves as the lower bound of the minimum distance in Segment \\( i \\)). Increasing the number of segments allows the step function to approach the red minimum distance curve, improving sampling quality and converging towards the quality of FPS. Figure 5 demonstrates this convergence. With more segments, the distribution of minimum distance between sampled points becomes closer to that of FPS. Details on implementation are provided in Appendix A.1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.591, + 0.484, + 0.849 + ], + "angle": 0, + "content": "\\(\\bullet\\) Early Termination While the estimated minimum distance curve exhibits low error, overestimation can still negatively impact sampling quality. This is because it can lead to an excessive number of point exclusions, limiting the availability of points in later iterations. To mitigate this issue, we introduce Early Termination. This technique monitors the availability of points at each iteration. If no points remain, it advances to the next segment earlier than planned. When the final segment is exhausted, Sampling with Predicted Distance is terminated, and the remaining iterations are completed using the original FPS algorithm. For seamless transition to FPS, we efficiently update the minimum distance matrix (dists in Algorithm 1) by leveraging the exclusion list. This approach avoids recalculating all-to-all distances between points, significantly reducing the search space and minimizing the impact on sampling time. Appendix A.1 describes the details of Early Termination." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.885 + ], + "angle": 0, + "content": "Factors Contributing to Speedup The primary factors contributing to the latency of MDPS are:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "1. Minimum Distance Curve Estimation: This step, requir" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.906, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.575, + 0.295, + 0.846, + 0.31 + ], + "angle": 0, + "content": "Figure 6. Redundancy Free Neighbor Search" + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.326, + 0.905, + 0.357 + ], + "angle": 0, + "content": "ing 1/10 of the original FPS iterations, adds only a fraction of the original FPS latency." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.357, + 0.907, + 0.416 + ], + "angle": 0, + "content": "2. Exclusion List Construction: The required all-to-all distance calculations are fully parallelizable, resulting in minimal latency overhead. Detailed analysis on this factor is provided in Appendix A.2." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.417, + 0.906, + 0.508 + ], + "angle": 0, + "content": "3. Sampling Process: The use of exclusion list eliminates the need for additional distance calculations. Early termination requires additional FPS iterations. However, the impact is minimal: the overhead of \\(2.02\\%\\), \\(0.58\\%\\), and \\(1.18\\%\\) of original FPS iterations in S3DIS, ScanNet, and SemanticKITTI dataset, respectively." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.357, + 0.907, + 0.508 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.906, + 0.553 + ], + "angle": 0, + "content": "Due to these optimizations, MDPS can achieve a \\(4 - 5 \\times\\) speedup relative to the original FPS algorithm. Detailed latency breakdown of MDPS is provided in Appendix B.2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.563, + 0.822, + 0.579 + ], + "angle": 0, + "content": "4.2. Redundancy-Free Neighbor Search" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.585, + 0.906, + 0.661 + ], + "angle": 0, + "content": "The exclusion list in MDPS captures spatial relationships between points, providing an opportunity to optimize subsequent Ball Query and k-NN operations. We introduce Redundancy Free Ball Query and k-NN, which fully leverage the exclusion list from the sampling stage." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.667, + 0.907, + 0.85 + ], + "angle": 0, + "content": "Redundancy-Free Ball Query Ball query, which identifies neighbors within a radius \\( R \\), is performed after sampling, before aggregation (refer to Figure 1). The exclusion list is well-suited for this task, as it contains points within a specified radius of each point. By adding an extra list to the exclusion list for the radius \\( R \\), as illustrated in Figure 6, we can directly use it for ball query. As discussed in Section 4.1, this approach minimizes computational overhead; distances between points are calculated only once, regardless of the number of segments. After sampling, we extract the relevant entries from the corresponding exclusion list for the ball query, which serve as centroids for aggregation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Redundancy-Free k-NN While directly reusing the exclusion list for k-NN is challenging, we can leverage it to reduce the search space. By using the exclusion list of Seg" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25119" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.089, + 0.472, + 0.204 + ], + "angle": 0, + "content": "
DatasetAverage Minimum Distance (Ratio to Baseline)
Baseline (FPS)Random SamplingGrid SamplingMDPS (Ours)
S3DIS0.063950.03478 (54.38%)0.04971 (77.72%)0.06359 (99.45%)
ScanNet0.037010.02035 (54.99%)0.02400 (64.84%)0.03677 (99.35%)
SemanticKITTI0.255050.12775 (50.09%)0.21519 (84.37%)0.25087 (98.36%)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.21, + 0.483, + 0.24 + ], + "angle": 0, + "content": "Table 1. Comparison of Sampling Quality among FPS, Random Sampling, Grid Sampling and MDPS." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.251, + 0.484, + 0.403 + ], + "angle": 0, + "content": "ment 1, we can limit the search space by eliminating points that are too far away to be potential k-NN neighbors. For example, to find the three nearest neighbors of point P7 in Figure 6, we calculate distances between P7 and the points from the corresponding exclusion list to identify the three nearest neighbors. Because k-NN upsampling relies on the downsampled set, points that were not downsampled are excluded from the search space. This approach significantly reduces the number of distance calculations, leading to substantial latency reduction." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.415, + 0.208, + 0.431 + ], + "angle": 0, + "content": "5. Evaluation" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.44, + 0.229, + 0.457 + ], + "angle": 0, + "content": "5.1. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.462, + 0.483, + 0.553 + ], + "angle": 0, + "content": "We evaluate FastPoint on PointMetaBase [17] and PointVector [6], two recent models based on PointNet++. We evaluate our proposal on both indoor and outdoor datasets to demonstrate its scalability. We use S3DIS [1] and ScanNet [5] for indoor datasets and use SemanticKITTI [2] for outdoor datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.483, + 0.674 + ], + "angle": 0, + "content": "The end-to-end latency is measured for processing all scenes in each dataset's validation split. We use mean Intersection over Union (mIoU) as a model accuracy evaluation metric. Detailed evaluation methodologies follow the same approach in PointMetaBase [17], and PointVector [6]. We apply FastPoint and the other baseline methods only to the first layer of each model as the first layer's FPS and neighbor search dominates the total computation time (\\(>90\\%\\))." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.78 + ], + "angle": 0, + "content": "We implement FastPoint with custom CUDA kernels and integrate them into OpenPoints library. We also integrate the software implementation [11] of QuickFPS [12] opensourced by the authors to OpenPoints for comparison. The experiments are conducted using an NVIDIA Geforce RTX 3090 GPU with 24GB of memory. We use CUDA 11.8 and PyTorch 1.10.1 for software setup." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.788, + 0.338, + 0.805 + ], + "angle": 0, + "content": "5.2. Sampling Quality of MDPS" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.903 + ], + "angle": 0, + "content": "To evaluate sampling quality, we compute the average minimum distance between neighboring sampled points. A larger average distance indicates better coverage of sample points from the perspectives of point distribution and structure preservation. As shown in Table 1, MDPS achieves sampling quality comparable to FPS, with an average min" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.897, + 0.276 + ], + "angle": 0, + "content": "
ModelDatasetmIoU (Diff. to Baseline FPS)
Baseline (FPS)Random SamplingGrid SamplingMDPS (Ours)
PV-LS3DIS71.3364.83 (-6.5)70.97 (-0.36)71.37 (+0.04)
ScanNet70.7059.94 (-10.76)69.61 (-1.09)70.63 (-0.07)
Semantic KITTI50.9138.46 (-12.45)50.62 (-0.29)50.79 (-0.12)
PMB-LS3DIS69.7268.21 (-1.51)69.54 (-0.18)69.74 (+0.02)
ScanNet70.8667.66 (-3.2)70.36 (-0.50)70.89 (+0.03)
Semantic KITTI52.1947.26 (-4.93)52.27 (+0.08)52.09 (-0.10)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.281, + 0.905, + 0.309 + ], + "angle": 0, + "content": "Table 2. Accuracy comparison. PV and PMB stand for PointVector and PointMetaBase." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.322, + 0.907, + 0.445 + ], + "angle": 0, + "content": "imum distance reaching over \\(99\\%\\) of the FPS baseline for S3DIS and ScanNet, and over \\(98\\%\\) for SemanticKITTI. In contrast, Random Sampling and Grid Sampling [29] exhibit significantly lower quality, with their average minimum distances deviating substantially from the FPS baseline. We also compare the minimum spacing distribution of each sampling method via visualization in Appendix B.9 and evaluate the robustness of MDPS in Appendix B.7." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.453, + 0.757, + 0.47 + ], + "angle": 0, + "content": "5.3. Accuracy Impact of MDPS" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.475, + 0.907, + 0.612 + ], + "angle": 0, + "content": "We analyze the impact of MDPS on accuracy by applying it to the inference stage of PointNet++ models initially trained with FPS. Since MDPS is designed to closely replicate the FPS sampling pattern, it is compatible with FPS-trained models, thereby minimizing accuracy degradation. Note that Redundancy-Free Neighbor Search does not introduce any approximation, thus has no impact on accuracy. To compare the effects of different sampling methods, we also evaluate random sampling and grid sampling1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.612, + 0.906, + 0.778 + ], + "angle": 0, + "content": "Table 2 shows that MDPS excels in maintaining accuracy, with negligible differences observed across all datasets and models. This highlights its effectiveness in replicating FPS behavior. However, the grid sampling exhibits larger and less consistent accuracy drops. When applied only during inference, it leads to declines of up to \\(1.09\\%\\) (PointVector-L, ScanNet). While it outperforms FastPoint in a single instance (PointMetaBase-L, SemanticKITTI), its performance is generally inferior, with significant accuracy reductions in other scenarios. The random sampling shows considerable accuracy drops across all data points." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.778, + 0.906, + 0.868 + ], + "angle": 0, + "content": "Overall, FastPoint effectively preserves accuracy by closely approximating the sampling pattern of FPS, achieving accuracy levels near-identical to FPS without retraining. Latency-accuracy comparisons (Appendix B.5) show that FastPoint resides above the pareto front, emphasizing the effectiveness of FastPoint." + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.905, + 0.901 + ], + "angle": 0, + "content": "1Grid sampling is applied only during inference. Appendix B.1 shows results for cases where both training and inference utilize grid sampling." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25120" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.115, + 0.089, + 0.46, + 0.228 + ], + "angle": 0, + "content": "
MethodNumber of SegmentsEvaluation Metrics
Accuracy (mIoU)Sampling Time (ms)Sampling Quality
FPS-70.7085.830.03701
MDPS (Ours)1 Segment70.3418.620.03532
2 Segments70.6019.430.03598
3 Segments70.4619.790.03636
4 Segments70.5421.010.03662
5 Segments70.5221.410.03665
6 Segments70.6322.060.03662
7 Segments70.6322.750.03664
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.238, + 0.483, + 0.295 + ], + "angle": 0, + "content": "Table 3. Comparison of mIoU, sampling time, and sampling quality for FPS and MDPS across segments 1 to 7. Sampling time is measured with a single scene of ScanNet dataset. PointVector-L model is used for mIoU comparison." + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.302, + 0.48, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.482, + 0.461, + 0.496 + ], + "angle": 0, + "content": "Figure 7. End-to-end speedup of FastPoint and QuickFPS." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.514, + 0.371, + 0.53 + ], + "angle": 0, + "content": "5.4. Latency Reduction of FastPoint" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.484, + 0.674 + ], + "angle": 0, + "content": "To evaluate the impact of FastPoint on end-to-end latency, we compare its latency to baseline FPS and QuickFPS. Note that FastPoint includes both MDPS and Redundancy-Free Neighbor Search. As explained in Section 2.2, QuickFPS accelerates FPS using k-d tree without approximation. Since MDPS also contains FPS steps when predicting the minimum distance curve, we can achieve further latency reductions by replacing them with QuickFPS, meaning that FastPoint and QuickFPS can be synergetic." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.81 + ], + "angle": 0, + "content": "Figure 7 presents the results of this comparison. Across S3DIS, ScanNet, and SemanticKITTI datasets, FastPoint achieves substantial latency reductions, with a geomean end-to-end speedup of \\(2.55 \\times\\) compared to baseline FPS. Integrating QuickFPS with FastPoint further reduces the execution latency achieving a geomean end-to-end speedup of \\(2.76 \\times\\) compared to baseline FPS and \\(1.41 \\times\\) compared to standalone QuickFPS. Speedup specific to the sampling and neighbor search are provided in Appendix B.3 and B.4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "These findings substantiate the effectiveness of Fast-Point in reducing latency, surpassing existing methods like QuickFPS. Moreover, the successful combination of Fast-Point with QuickFPS demonstrates its compatibility with existing FPS optimizations. Scalability to Non-PointNet++ based models are discussed in Appendix B.10." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.088, + 0.718, + 0.226 + ], + "angle": 0, + "content": "
ModelDatasetMethodSpeedup
PV-LS3DISMDPS1.90×
All1.98×
ScanNetMDPS2.33×
All2.52×
Semantic KITTIMDPS2.07×
All2.20×
PMB-LS3DISMDPS2.50×
All2.82×
ScanNetMDPS2.89×
All3.38×
Semantic KITTIMDPS2.31×
All2.62×
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.514, + 0.228, + 0.716, + 0.253 + ], + "angle": 0, + "content": "(a) Comparison of component-wise end-to-end speedup of FastPoint." + }, + { + "type": "table", + "bbox": [ + 0.718, + 0.088, + 0.897, + 0.226 + ], + "angle": 0, + "content": "
Model# of PointsSpeedup
PV-L16,0001.33×
32,0002.41×
48,0002.68×
64,0003.01×
80,0003.25×
96,0003.40×
PMB-L16,0001.58×
32,0002.64×
48,0002.91×
64,0003.06×
80,0003.30×
96,0003.44×
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.718, + 0.228, + 0.905, + 0.253 + ], + "angle": 0, + "content": "(b) Comparison of speedup as the number of points increases." + }, + { + "type": "table_caption", + "bbox": [ + 0.521, + 0.257, + 0.897, + 0.271 + ], + "angle": 0, + "content": "Table 4. Ablation study for speedup breakdown and scalability." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.29, + 0.667, + 0.306 + ], + "angle": 0, + "content": "5.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.311, + 0.905, + 0.448 + ], + "angle": 0, + "content": "Segment Count We evaluate the impact of varying the segment count on accuracy, sampling time, and sampling quality (i.e., average minimum distance). Table 3 shows that while sampling time increases gradually with more segments due to the exclusion list construction (refer to Section 4.1), both sampling quality and accuracy improve. However, they get saturated beyond a certain number of segments. Based on the empirical analysis, we use 6 segments by default, where accuracy saturates for all workloads." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.451, + 0.906, + 0.572 + ], + "angle": 0, + "content": "Component-wise Speedup To understand the individual contributions of each component in FastPoint, we evaluate the end-to-end speedup by sequentially applying MDPS and Redundancy Free Neighbor Search. Table 4a shows that MDPS alone provides significant speedup. Incorporating Redundancy Free Neighbor Search further provides additional improvement, demonstrating that each component makes a distinct contribution to the overall speedup." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.582, + 0.681, + 0.599 + ], + "angle": 0, + "content": "5.6. Scalability Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.604, + 0.906, + 0.71 + ], + "angle": 0, + "content": "To assess the scalability of FastPoint, we evaluate the end-to-end speedup compared to FPS by varying point cloud sizes. Table 4b shows that FastPoint consistently achieves higher speedups as the number of points increases. This trend highlights the efficiency of our approach, particularly when processing larger point clouds where the computational cost of the FPS becomes a significant bottleneck." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.725, + 0.634, + 0.74 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "3D point clouds are critical for representing and understanding 3D scenes, and deep learning models like PointNet++ have shown great promise in processing this data. However, computational efficiency remains a key challenge. We propose FastPoint, a software acceleration technique that leverages predictable minimum distances between sampled points to efficiently identify subsequent points without exhaustive computations. Our evaluation shows a \\(2.55 \\times\\) speedup over baseline FPS while maintaining accuracy, enabling low-cost deployment of PointNet++ based models." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25121" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.26, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.486, + 0.329 + ], + "angle": 0, + "content": "This work was supported by the Ministry of Science and ICT (MSIT), Korea, under the following grants: the Global Scholars Invitation Program (RS-2024-00456287), the Graduate School of Artificial Intelligence Semiconductor grant (IITP-2025-RS-2023-00256081), and the Artificial Intelligence Innovation Hub (No. 2021-0-02068), all supervised by the Institute for Information & Communications Technology Planning & Evaluation (IITP). Additional support was provided by the National Research Foundation of Korea (NRF) grant (RS-2024-00405857) funded by the Korea government (MSIT), and the Google Faculty Research Award. The source code is available at https://github.com/SNU-ARC/FastPoint.git. Jae W. Lee and Hongil Yoon are the corresponding authors." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.344, + 0.188, + 0.36 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.369, + 0.484, + 0.411 + ], + "angle": 0, + "content": "[1] Iro Armeni, Ozan Sener, Amir R. Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In CVPR, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.413, + 0.484, + 0.468 + ], + "angle": 0, + "content": "[2] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Juergen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In ICCV, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.47, + 0.483, + 0.512 + ], + "angle": 0, + "content": "[3] Yukang Chen, Jianhui Liu, Xiangyu Zhang, Xiaojuan Qi, and Jiaya Jia. Largekernel3d: Scaling up kernels in 3d sparse cnns. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.515, + 0.483, + 0.556 + ], + "angle": 0, + "content": "[4] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.558, + 0.484, + 0.613 + ], + "angle": 0, + "content": "[5] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.616, + 0.484, + 0.656 + ], + "angle": 0, + "content": "[6] Xin Deng, WenYu Zhang, Qing Ding, and XinMing Zhang. Pointvector: A vector representation in point cloud analysis. In CVPR, 2023. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.659, + 0.484, + 0.714 + ], + "angle": 0, + "content": "[7] Lue Fan, Ziqi Pang, Tianyuan Zhang, Yu-Xiong Wang, Hang Zhao, Feng Wang, Naiyan Wang, and Zhaoxiang Zhang. Embracing Single Stride 3D Object Detector with Sparse Transformer. In CVPR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.717, + 0.485, + 0.785 + ], + "angle": 0, + "content": "[8] Yu Feng, Boyuan Tian, Tiancheng Xu, Paul Whatmough, and Yuhao Zhu. Mesorasi: Architecture support for point cloud analytics via delayed-aggregation. In Proceedings of the 53th International Symposium on Microarchitecture (MICRO), 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.788, + 0.484, + 0.857 + ], + "angle": 0, + "content": "[9] Yu Feng, Gunnar Hammonds, Yiming Gan, and Yuhao Zhu. Crescent: Taming memory irregularities for accelerating deep point cloud analytics. In Proceedings of the 49th Annual International Symposium on Computer Architecture (ISCA), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.484, + 0.899 + ], + "angle": 0, + "content": "[10] Lei Han, Tian Zheng, Lan Xu, and Lu Fang. Occuseg: Occupancy-aware 3d instance segmentation. In CVPR, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.369, + 0.485, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.148 + ], + "angle": 0, + "content": "[11] Meng Han, Liang Wang, Limin Xiao, Hao Zhang, Chenhao Zhang, Xiangrong Xu, and Jianfeng Zhu. Quickfps. http://github.com/hanm2019/bucket-based_farthest-point-sampling_GPU.2,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.906, + 0.233 + ], + "angle": 0, + "content": "[12] Meng Han, Liang Wang, Limin Xiao, Hao Zhang, Chenhao Zhang, Xiangrong Xu, and Jianfeng Zhu. Quickfps: Architecture and algorithm co-design for farthest point sampling in large-scale point clouds. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 2023. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.236, + 0.906, + 0.291 + ], + "angle": 0, + "content": "[13] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. In CVPR, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.295, + 0.906, + 0.336 + ], + "angle": 0, + "content": "[14] Donghyun Lee, Yejin Lee, Jae W. Lee, and Hongil Yoon. Frugal 3d point cloud model training via progressive near point filtering and fused aggregation. In ECCV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.339, + 0.906, + 0.394 + ], + "angle": 0, + "content": "[15] Jingtao Li, Jian Zhou, Yan Xiong, Xing Chen, and Chaitali Chakrabarti. An adjustable farthest point sampling method for approximately-sorted point cloud data. In 2022 IEEE Workshop on Signal Processing Systems (SiPS), 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.397, + 0.906, + 0.465 + ], + "angle": 0, + "content": "[16] Yaoxiu Lian, Xinhao Yang, Ke Hong, Yu Wang, Guohao Dai, and Ningyi Xu. A point transformer accelerator with fine-grained pipelines and distribution-aware dynamic fps. In IEEE/ACM International Conference on Computer-Aided Design (ICCAD), 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.468, + 0.906, + 0.521 + ], + "angle": 0, + "content": "[17] Haojia Lin, Xiawu Zheng, Lijiang Li, Fei Chao, Shanshan Wang, Yan Wang, Yonghong Tian, and Rongrong Ji. Meta architecture for point cloud analysis. In CVPR, 2023. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.906, + 0.595 + ], + "angle": 0, + "content": "[18] Xueyuan Liu, Zhuoran Song, Guohao Dai, Gang Li, Can Xiao, Yan Xiang, Dehui Kong, Ke Xu, and Xiaoyao Liang. Fusionarch: A fusion-based accelerator for point-based point cloud neural networks. In ACM/IEEE Design, Automation and Test in Europe Conference (DATE), 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.598, + 0.906, + 0.652 + ], + "angle": 0, + "content": "[19] Yongcheng Liu, Bin Fan, Gaofeng Meng, Jiwen Lu, Shiming Xiang, and Chunhong Pan. Densepoint: Learning densely contextual representation for efficient point cloud processing. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.655, + 0.906, + 0.696 + ], + "angle": 0, + "content": "[20] Zhijian Liu, Xinyu Yang, Haotian Tang, Shang Yang, and Song Han. Flatformer: Flattened window attention for efficient point cloud transformer. In CVPR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.699, + 0.906, + 0.741 + ], + "angle": 0, + "content": "[21] Jiageng Mao, Yujing Xue, Minzhe Niu, Haoyue Bai, Jiashi Feng, Xiaodan Liang, Hang Xu, and Chunjing Xu. Voxel transformer for 3d object detection. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.744, + 0.906, + 0.799 + ], + "angle": 0, + "content": "[22] Alexey Nekrasov, Jonas Schult, Or Litany, Bastian Leibe, and Francis Engelmann. Mix3D: Out-of-Context Data Augmentation for 3D Scenes. In International Conference on 3D Vision (3DV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.906, + 0.856 + ], + "angle": 0, + "content": "[23] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. arXiv preprint arXiv:1612.00593, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[24] Charles R Qi, Li Yi, Hao Su, and Leonidas J Guibas. Point-net++: Deep hierarchical feature learning on point sets in a metric space. arXiv preprint arXiv:1706.02413, 2017. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25122" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[25] Charles R. Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas J. Guibas. Frustum pointnets for 3d object detection from rgb-d data. In CVPR, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.134, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[26] Guocheng Qian, Hasan Hammoud, Guohao Li, Ali Thabet, and Bernard Ghanem. Assanet: An anisotropical separable set abstraction for efficient point cloud representation learning. In NeurIPS, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.19, + 0.483, + 0.247 + ], + "angle": 0, + "content": "[27] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.247, + 0.485, + 0.316 + ], + "angle": 0, + "content": "[28] Danila Rukhovich, Anna Vorontsova, and Anton Konushin. Imvoxelnet: Image to voxels projection for monocular and multi-view general-purpose 3d object detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.316, + 0.483, + 0.372 + ], + "angle": 0, + "content": "[29] Hugues Thomas, Charles R. Qi, Jean-Emmanuel Deschaud, Beatriz Marcotegui, François Goulette, and Leonidas J. Guibas. Kpconv: Flexible and deformable convolution for point clouds. In ICCV, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.372, + 0.483, + 0.427 + ], + "angle": 0, + "content": "[30] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E. Sarma, Michael M. Bronstein, and Justin M. Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (TOG), 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.427, + 0.483, + 0.468 + ], + "angle": 0, + "content": "[31] Wenxuan Wu, Zhongang Qi, and Li Fuxin. Pointconv: Deep convolutional networks on 3d point clouds. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.469, + 0.483, + 0.512 + ], + "angle": 0, + "content": "[32] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. In NeurIPS, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.512, + 0.483, + 0.567 + ], + "angle": 0, + "content": "[33] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler, faster, stronger. In CVPR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.567, + 0.483, + 0.609 + ], + "angle": 0, + "content": "[34] Mutian Xu, Runyu Ding, Hengshuang Zhao, and Xiaojuan Qi. Paconv: Position adaptive convolution with dynamic kernel assembling on point clouds. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.609, + 0.483, + 0.651 + ], + "angle": 0, + "content": "[35] Qiangeng Xu, Xudong Sun, Cho-Ying Wu, Panqu Wang, and Ulrich Neumann. Grid-gen for fast and scalable point cloud learning, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.651, + 0.483, + 0.719 + ], + "angle": 0, + "content": "[36] Xinhao Yang, Tianyu Fu, Guohao Dai, Shulin Zeng, Kai Zhong, Ke Hong, and Yu Wang. An efficient accelerator for point-based and voxel-based point cloud neural networks. In ACM/IEEE Design Automation Conference (DAC), 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.72, + 0.483, + 0.789 + ], + "angle": 0, + "content": "[37] Yu-Qi Yang, Yu-Xiao Guo, Jian-Yu Xiong, Yang Liu, Hao Pan, Peng-Shuai Wang, Xin Tong, and Baining Guo. Swin3d: A pretrained transformer backbone for 3d indoor scene understanding. arXiv preprint arXiv:2304.06906, 2023.1,2" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.79, + 0.483, + 0.832 + ], + "angle": 0, + "content": "[38] Yu-Qi Yang, Yu-Xiao Guo, and Yang Liu. Swin3d++: Effective multi-source pretraining for 3d indoor scene understanding. arXiv preprint arXiv:2304.06906, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.832, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[39] Ziyu Ying, Sandeepa Bhuyan, Yan Kang, Yingtian Zhang, Mahmut T. Kandemir, and Chita R. Das. Edgepc: Efficient deep learning analytics for point clouds on edge devices. In Proceedings of the 50th Annual International Symposium on Computer Architecture (ISCA), 2023. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.092, + 0.485, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.092, + 0.906, + 0.149 + ], + "angle": 0, + "content": "[40] Hyunsung Yoon and Jae-Joon Kim. Efficient sampling and grouping acceleration for point cloud deep learning via single coordinate comparison. In IEEE/ACM International Conference on Computer-Aided Design (ICCAD), 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.149, + 0.906, + 0.192 + ], + "angle": 0, + "content": "[41] Hengshuang Zhao, Li Jiang, Chi-Wing Fu, and Jiaya Jia. PointWeb: Enhancing local neighborhood features for point cloud processing. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.192, + 0.906, + 0.221 + ], + "angle": 0, + "content": "[42] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip Torr, and Vladlen Koltun. Point transformer. In ICCV, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.221, + 0.906, + 0.277 + ], + "angle": 0, + "content": "[43] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. arXiv preprint arXiv:2011.10033, 2020. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.092, + 0.906, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25123" + } + ] +] \ No newline at end of file diff --git a/2025/FastPoint_ Accelerating 3D Point Cloud Model Inference via Sample Point Distance Prediction/670cb46f-786b-455d-998f-328c26d7e47d_origin.pdf b/2025/FastPoint_ Accelerating 3D Point Cloud Model Inference via Sample Point Distance Prediction/670cb46f-786b-455d-998f-328c26d7e47d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9e2f2e4efc7f1869c8514620744ef5379d5edad7 --- /dev/null +++ b/2025/FastPoint_ Accelerating 3D Point Cloud Model Inference via Sample Point Distance Prediction/670cb46f-786b-455d-998f-328c26d7e47d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2b58f90f87df6c69038c0f7d67e45c16a99ae365d16bf247e667bbb5ad072099 +size 1083510 diff --git a/2025/FastPoint_ Accelerating 3D Point Cloud Model Inference via Sample Point Distance Prediction/full.md b/2025/FastPoint_ Accelerating 3D Point Cloud Model Inference via Sample Point Distance Prediction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1dfcabbbb90eb53fdddc58e7b5fa1f6639459a41 --- /dev/null +++ b/2025/FastPoint_ Accelerating 3D Point Cloud Model Inference via Sample Point Distance Prediction/full.md @@ -0,0 +1,320 @@ +# FastPoint: Accelerating 3D Point Cloud Model Inference via Sample Point Distance Prediction + +Donghyun Lee $^{1}$ Dawoon Jeong $^{1}$ Jae W. Lee $^{1}$ Hongil Yoon $^{1,2}$ Seoul National University ${ }^{2}$ Google + +{eudh1206, daun20211, jaewlee}@snu.ac.kr, hongilyoon@google.com + +# Abstract + +Deep neural networks have revolutionized 3D point cloud processing, yet efficiently handling large and irregular point clouds remains challenging. To tackle this problem, we introduce FastPoint, a novel software-based acceleration technique that leverages the predictable distance trend between sampled points during farthest point sampling. By predicting the distance curve, we can efficiently identify subsequent sample points without exhaustively computing all pairwise distances. Our proposal substantially accelerates farthest point sampling and neighbor search operations while preserving sampling quality and model performance. By integrating FastPoint into state-of-the-art 3D point cloud models, we achieve $2.55 \times$ end-to-end speedup on NVIDIA RTX 3090 GPU without sacrificing accuracy. + +# 1. Introduction + +3D point clouds have become increasingly important for representing and understanding 3D scenes, driving advancements in fields like robotics and autonomous driving. Deep neural networks [4, 6, 7, 13, 17, 20, 23-26, 30, 32, 33, 37, 42, 43] have emerged as powerful tools for processing this type of data. While PointNet++ [24] and its successors have improved efficiency and performance of these point cloud models, the irregular nature and growing size of point cloud data continue to impose significant computational challenges, especially for real time use cases. Operations such as Farthest Point Sampling (FPS) and neighbor search remain major bottlenecks. + +While various hardware accelerators [8, 9, 12, 16, 18, 36, 40] have been explored to accelerate the models, software-only solutions are less common. Given the high cost and complexity of developing specialized accelerators, software-based acceleration is a more practical approach. Recently, several software-based techniques [15, 39] have been proposed to accelerate these models through approximation. However, such approaches often come at the cost + +of significant accuracy loss, limiting their practical use. + +To address this challenge, we propose FastPoint, a novel software-based acceleration technique with sample point distance prediction. We observe that the inherent nature of FPS, which iteratively selects the point farthest from the current set of sampled points, leads to two predictable trends in the minimum distances between these points: + +1. Decreasing minimum distance: As FPS makes progress, the number of remaining points decreases, forcing the algorithm to select points closer to those already sampled. This results in a smoothly decreasing curve of minimum distances between sampled points. + +2. Early structure capture: The initial sample points tend to be the extremities of the point cloud, effectively capturing its overall shape and boundaries. + +Our proposal leverages these observations to accurately estimate the distance curve using only a few initial FPS iterations. By predicting the distance curve, we can efficiently identify subsequent sample points without computing and comparing pairwise distances at each FPS iteration. This leads to significant latency savings while maintaining sampling quality comparable to that of the original FPS. Furthermore, predicting the distance curve enables the sampling process to be decoupled from distance computation. This decoupling exposes new opportunities for parallelizing distance calculations in the proposed sampling, further enhancing efficiency. This approach not only accelerates FPS itself but also benefits subsequent operations such as neighbor search by leveraging pre-computed distances. + +Our key contributions are summarized as follows: + +- We identify two key latency bottlenecks in point cloud models: FPS and neighbor search. Limited parallelism in FPS and redundant distance computations across both operations are the primary sources of inefficiency. +- We empirically analyze 3D point cloud models and identify the decreasing trend and early structure capture trend. +- Based on the observations, we introduce a novel sampling approach with sample point distance prediction. With the estimated distance curve, this approach not only accelerates FPS itself but also benefits subsequent neighbor + +![](images/1cb09cf346bf4144a51258661e69dc8af82c06e8da911b5af9dcbc417c035d59.jpg) +Figure 1. Overview of PointNet++ Based Model Architecture + +search, which relies on pre-computed distances. + +- We integrate FastPoint into OpenPoints library and experiment on state-of-the-art PointNet++ based models, achieving geomean $2.55 \times$ end-to-end speedup on NVIDIA RTX 3090 GPU while preserving accuracy. + +# 2. Background and Related Work + +# 2.1. Deep Learning on 3D Point Clouds + +Deep Neural Networks on Raw Point Cloud Since the release of PointNet [23], which is the first to apply deep neural networks directly to the raw point clouds, numerous successors [6, 13, 17, 19, 24-27, 30, 31, 34, 41] have been developed, improving both model performance and efficiency. PointNet++ [24] introduces a hierarchical structure to PointNet through sampling and grouping, which has become the core architecture of subsequent PointNet++ based models. PointNeXt [27] revisits PointNet++, achieving substantial improvements in model performance, while PointVector [6] and PointMetaBase [17] push the model performance and efficiency even further. Although the architectural details of each model differ, all models follow the PointNet++ structure, utilizing Set Abstraction and Feature Propagation layers as core components. This work focuses on identifying and accelerating the primary latency bottlenecks in PointNet++ based models. + +Deep Neural Networks on Voxelized Point Cloud Several works [3, 4, 10, 22, 28, 43] propose to voxelize point clouds, using voxels as input rather than raw points. The advent of 3D sparse convolution [4] has made this voxel-based approach increasingly popular due to its efficiency and high performance. Recently, transformer-based models [7, 21, 37, 38] operating on voxelized point clouds are gaining significant attention. Although voxelization alleviates the computational cost of mapping operations (i.e., downsampling and neighbor search), it has a notable limitation due to the loss of position information during the voxelization process. + +# 2.2. Related Work + +Numerous solutions have been proposed to accelerate sampling and neighbor search operations in PointNet++ based models. We categorize them into hardware- and purely software-based approaches. + +Hardware-Assisted Acceleration QuickFPS [12] introduces a k-d tree based FPS algorithm which reduces computation and memory access in each FPS iteration, along with specialized hardware designed for this method. MARS [36] and PTrAcc [16] both propose hardware accelerators that employ a distance filtering technique which skips unnecessary distance computations in each iteration of FPS. Although these techniques achieve substantial speedups in FPS without any loss in sampling quality, they require specialized hardware to fully harness their algorithms, limiting their effectiveness on commodity hardware like GPU. For more thorough evaluation, we compare the speedup of our proposal with pure software version [11] of QuickFPS in Section 5.4, evaluating its effectiveness on GPU. + +Software-Only Acceleration To mitigate the high latency associated with FPS, various alternative sampling methods have been explored. RandLA-Net [13] replaces FPS with random sampling and introduces a local feature aggregation module that compensates for the low sampling quality of random sampling. Grid sampling, the most commonly used alternative, is adopted by Grid-GCN [35] and KPCov [29], offering faster processing with reasonable sampling quality compared to random sampling. Despite the latency advantages of these methods, none have consistently outperformed FPS for overall performance. This is illustrated by the fact that most PointNet++-based models [6, 17, 24-27] continue to rely on FPS, highlighting its broad applicability and superior sampling quality. The experimental results in Appendix B.1 corroborate this claim. + +EdgePC [39] structures the point cloud with morton codes to accelerate FPS and neighbor search on edge GPUs, + +![](images/ae2db72dee1c17f01cd52410b02dd70628795cce94c8d8d03b0d8080d45789cf.jpg) +Figure 2. Latency Breakdown of PointNet++ Based Models + +while Adjustable FPS [15] accelerates FPS by leveraging the intrinsic locality of the point cloud data. However, both methods demonstrate significant accuracy loss compared to the baseline (i.e., up to $2\%$ mIoU loss) due to their aggressive approximations, limiting their effectiveness. Prior work [14] accelerates the training of PointNet++ based models by precomputing the minimum point spacing of farthest-sampled points before training and reusing this value across epochs for fast sampling. While this approach effectively speeds up training, it is not suitable for inference, as the minimum spacing value cannot be determined in advance in inference scenarios. Detailed comparisons with our work are provided in Appendix A.4. + +# 3. Challenges of Point Cloud Models + +# 3.1. Model Architecture Overview + +Figure 1 illustrates the detailed process of PointNet++-based models. Similar to 2D convolutional neural networks, these models consist of both downsampling and upsampling layers. Commonly referred to as the set abstraction layer, the downsampling layer in PointNet++-based models comprises three main stages, as outlined below: + +1 Farthest Point Sampling (FPS): Downsampling begins by selecting $N$ /stride output points from $N$ input points, where stride represents the downsampling rate. Typically, the FPS is used to maintain the shape and structure of the input point cloud. +$\bullet$ Ball Query: Next, a ball query with radius $R$ is performed around each sampled point to gather neighboring points for feature aggregation. +Aggregation: The features of neighboring points identified in the ball query are then processed. A multi-layer perceptron (MLP) is applied to these features, followed by max pooling to aggregate the information for each centroid. The details of this process, e.g., positional embedding [17], can vary depending on the model architectures. + +# Algorithm 1 Farthest Point Sampling + +Input $P$ : input point cloud, $n$ : number of points to sample, $N$ : number of original points + +Output sampledidx: indices of sampled points + +1: sampled_idx[0] $\leftarrow$ seed_idx // Starts with a random point +2: min_dists[i] $\leftarrow \infty$ for $i = 0,\dots,N - 1$ +3: for $i \gets 1$ to $n - 1$ do // Not parallelized +4: for $j\gets 0$ to $N - 1$ do // Parallelized +5: $d_{\text{new}} \gets \text{dist}(P[\text{sampled\_idx}[i - 1]], P[j])$ +6: if min_dists[j] > $d_{new}$ then +7: min_dists[j] $\leftarrow d_{new}$ +8: sampledidx[i] $\leftarrow$ argmax(min_dists) + +The upsampling layer, often referred to as a feature propagation layer, takes the downsampled points from the corresponding downsampling layer as inputs and propagates the features back to the original point cloud. This process is done by following steps. + +$\bullet$ k-NN: For each upsampled point, k-NN is applied to identify its k nearest neighbors among the input points. +$\Theta$ Interpolation: Based on the k-NN results, neighboring input points are interpolated, concatenated with skip-connected features from the corresponding downsampling layer, and the information is passed through an MLP to propagate the features to the subsequent layer. + +# 3.2. Latency Breakdown + +Figure 2 presents the latency breakdown of PointMetaBase-L model evaluated on both indoor (S3DIS) and outdoor (SemanticKITTI) scenes. The results show that coordinate-based operations (i.e., FPS, k-NN, and Ball Query) dominate the execution time, consuming significantly more time than feature-based operations (i.e., Aggregation and MLP). These coordinate-based operations account for approximately $88.1\%$ of the total inference time for indoor scenes and $94.1\%$ for the outdoor scenes. Among these, FPS emerges as the primary performance bottleneck, requiring the most execution time. Additionally, neighbor search operations, including k-NN and Ball Query, contribute significantly to the overall latency. To address these challenges, we first analyze the root causes of these inefficiencies and propose novel optimization techniques in subsequent sections. + +# 3.3. Inefficiencies in Farthest Point Sampling + +Farthest point sampling operates iteratively by selecting one point at a time. In each iteration, the sampling algorithm identifies the point farthest from the set of the previously sampled points. The detailed process is outlined in Algorithm 1. Initially, a random point in $P$ is chosen as the seed point (Line 1) and added to the sampled_idx. Then, a distance matrix (i.e., min_dists) is created to store the distance + +between each point in $P$ and the set of the previously sampled points, where the distance from a point to a point set is defined as minimum distance between the point and any point in the set. This matrix is initialized to $\infty$ (Line 2). In each iteration, the distance between the most recently sampled point and every other point in $P$ is calculated (Lines 3-7). If a newly calculated distance is smaller than the corresponding entry in the min_dist matrix, the matrix is updated with the smaller value. Finally, the index of the point with the maximum distance in the min_dist is added to the sampled_idx (Line 8). This process is repeated until the specified number of points $(n)$ is sampled. + +The sequential nature of the FPS algorithm is the primary reason of the inefficiency. Because only one point is sampled per iteration, the $O(N)$ distance calculations required in each iteration must be sequentially executed $n$ times. This tight coupling between sampling and distance calculation significantly limits the potential for parallelism in distance computations. + +To unlock this parallelism potential, we introduce Minimum Distance Prediction Sampling (MDPS) in Section 4.1, a novel sampling strategy that fully decouples sampling from distance computation, enabling greater parallelism and faster processing speeds. + +# 3.4. Inefficiencies in Neighbor Search + +As explained in Section 3.1, PointNet++ based models employ two types of neighbor search algorithms: Ball Query and k-NN. These operations play a critical role in capturing local geometric relationships in the point cloud. Ball Query identifies neighbor points within a fixed radius around the downsampled points. k-NN identifies k nearest neighbors among the downsampled points for each original point. + +The inefficiencies in these neighbor search operations stem from redundant distance computations in both the FPS and subsequent neighbor search operations. For instance, as shown in Figure 1, the distance between P0 and P2 is computed while updating P2's minimum distance in FPS. However, the same distance is redundantly recalculated in both Ball Query and k-NN. Section 4.2 introduces redundancy-free neighbor search techniques to address this issue. + +# 4. FastPoint: Fast Point Cloud Inference + +Trends in Minimum Distance of Sample Points The FPS iteratively selects the point farthest from the current set of sampled points, which leads to predictable trends in the distance between the sampled points. + +1. Decreasing minimum distance: As sampling proceeds, the pool of the remaining points shrinks, the FPS algorithm selects points closer to those already sampled over time. Thus, the maximized minimum distance (i.e., $\max (min\_dist)$ in Algorithm 1) decreases, which we define as minimum distance for brevity. This is particu + +![](images/55827e7bb3e7aff332dc1de514fd528fb813171139164c5a918de76c6fce7757.jpg) +(a) Minimum Distance Curve of ScanNet scene + +![](images/e9ddf982be8815f4919e16673089ab17a2b4ffbcf22e1f86d439a2176cd5a2d7.jpg) +(b) Visualization of ScanNet scene sampled at 1/10 of the original iterations +Figure 3. Motivation of Minimum Distance Curve Estimation + +![](images/53faa4f40d59824438cfe9fd48e63d59c0fc5b24b0ba8a64c09b93452036366c.jpg) + +larly evident in Figure 3a, showing a smooth, decreasing curve for various input point clouds. + +2. Early structure capture: Importantly, we observe that a few initial sample points are sufficient to capture the overall shape and boundaries of the input point cloud (Figure 3b). This is because the initial sample points tend to be the extremities of the point cloud. + +Optimization Opportunities The trend of early structure capture suggests that the initial iterations of FPS are important in accurately representing the point cloud's structure. The Decreasing trend suggests that the later portion of the minimum distance curve becomes more predictable as sampling progresses. Building upon these trends, if we could accurately estimate this curve using only a few initial FPS iterations, we could maintain a comparable sampling quality while achieving substantial latency reduction by selecting points guided by the predicted curve. + +Proposed Techniques We introduce two optimizations by leveraging these opportunities: Minimum Distance Prediction Sampling and Redundancy-Free Neighbor Search. Minimum Distance Prediction Sampling (MDPS) is a novel sampling strategy designed to approximate the sampling quality of FPS while significantly reducing latency. This approach estimates the distance curve using only a few initial non-parallelizable FPS iterations. Subsequent sample points are selected based on the predicted distances, eliminating the need for pairwise distance comparisons in later iterations. Furthermore, MDPS enables Redundancy-Free Neighbor Search by reusing precomputed distance information, thereby eliminating redundant computations during subsequent Ball Query and k-NN operations. This optimization further enhances the overall efficiency of PointNet++-based models during inference. + +# 4.1. Minimum Distance Prediction Sampling + +1 Minimum Distance Curve Estimation To build a low-cost estimator for the minimum distance curve (as shown in Figure 4), we aim to estimate the curve with minimal error using the results from as few FPS iterations as possible. We + +![](images/29f6522d3450a20ddef2d4220cb9359d6f788d9dedb8a0cdcb5970978ad7f18b.jpg) +Figure 4. Overall Flow of Minimum Distance Prediction Sampling + +formulate the estimator $\mathcal{E}$ as a function which gets initial $p$ segment (where $p$ is the ratio of initial FPS iterations to total iterations, $0 < p < 1$ ) of the curve $C$ as input and outputs the rest of the estimated curve $\hat{C}$ (i.e., Equation 1). Note that $C(1)$ represents the minimum distance among the sampled points after all sampling iterations are finished. + +$$ +\hat {C} (t) = \mathcal {E} (C (t) \mid t \in [ 0, p ]) \quad \text {f o r} t \in (p, 1 ] \tag {1} +$$ + +We have experimented with various models, including polynomial functions and MLPs, and find that MLPs yield the best performance (Appendix A.3). We train a 3-layer MLP on minimum distance curves extracted from the training split of each point cloud dataset, setting $p = 0.1$ to balance error and efficiency. Our estimator achieves a low Mean Absolute Percentage Error (MAPE) of $1.02\%$ , $0.77\%$ , and $1.93\%$ on S3DIS, ScanNet, and SemanticKITTI datasets, respectively, performing better on indoor datasets (S3DIS, ScanNet) due to their more consistent point density and distribution. Appendix B.6 and B.8 discuss ablation study for $p$ and cross-dataset applicability of estimators. + +$\Theta$ Distance Curve Segmentation Ideally, what we want is to sample points whose minimum distance matches the value on the curve at each iteration. However, directly finding the point that yields the exact minimum distance in every iteration is computationally expensive. + +To efficiently sample points along the estimated minimum distance curve, we first divide the curve into multiple segments. Note that each segment is associated with a group of consecutive iterations. Then, the corresponding predicted minimum distances are employed to identify sample point at the segment-level granularity in the sampling process. + +As illustrated in Figure 4, the curve is segmented at iterations $d_{1}$ , $d_{2}$ , and $d_{3}$ , resulting in three segments. For + +each input point, we identify points within a specified radius of each segment boundary (i.e., $R_{1}$ , $R_{2}$ , and $R_{3}$ ) and store them in the exclusion list for three segments. This list prevents sampling closely located points by considering the sampled points. + +To optimize computational efficiency, building the exclusion list across all segments is fused into a single GPU kernel. By calculating the distance between the points only once, regardless of the number of segments, we minimize the latency overhead. For example, the distance between P0 and P1 is computed only once. If P1 is within $R_{3}$ (i.e., $\mathrm{dist(P0,P1)} < R_3 < R_2 < R_1$ ), P1 is added to the exclusion list of P0 at all segments. Refer to Table 3 for latency overhead with an increase in the segment count. + +Building the exclusion list still requires computing pairwise distances; however, this process is much faster than the original FPS algorithm. This improvement is achieved because all computations are fully parallelizable, as sampling and distance calculations are completely decoupled. For details on the algorithm, refer to Appendix A.1. + +Sampling with Predicted Distance The sampling process employs the exclusion list. Starting from Segment 1, a seed point (P0 in Figure 4) is chosen. In each sampling iteration, points within a radius $R_{1}$ from the most recently sampled point are filtered out based on the exclusion list. We employ bitmaps to record which points are available for sampling. For instance, in the first iteration of Figure 4, Point P0, P1, P4, P5, P6, P7, and P11 are excluded based on the exclusion list of Segment 1. Point P3 is randomly selected among the eight available points for the second iteration in the same segment. Excluded points remain unavailable until the current segment is complete. + +As the sampling progresses and the segment changes + +![](images/33c964742bb34613ffa0c942e141c996bba899b16190410740caa8d5d6692ef1.jpg) +Figure 5. Minimum spacing distribution comparison with increasing segment counts. + +(e.g., from Segment 1 to Segment 2), the radius threshold is adjusted (e.g., from $R_{1}$ to $R_{2}$ ). This transition changes the set of points that are available for sampling. By iteratively applying this process, we ensure that the distance between sampled points meets or exceeds the threshold for each segment. This results in a minimum distance that lies above the blue step function in the segmented distance curve (i.e., $R_{i}$ serves as the lower bound of the minimum distance in Segment $i$ ). Increasing the number of segments allows the step function to approach the red minimum distance curve, improving sampling quality and converging towards the quality of FPS. Figure 5 demonstrates this convergence. With more segments, the distribution of minimum distance between sampled points becomes closer to that of FPS. Details on implementation are provided in Appendix A.1. + +$\bullet$ Early Termination While the estimated minimum distance curve exhibits low error, overestimation can still negatively impact sampling quality. This is because it can lead to an excessive number of point exclusions, limiting the availability of points in later iterations. To mitigate this issue, we introduce Early Termination. This technique monitors the availability of points at each iteration. If no points remain, it advances to the next segment earlier than planned. When the final segment is exhausted, Sampling with Predicted Distance is terminated, and the remaining iterations are completed using the original FPS algorithm. For seamless transition to FPS, we efficiently update the minimum distance matrix (dists in Algorithm 1) by leveraging the exclusion list. This approach avoids recalculating all-to-all distances between points, significantly reducing the search space and minimizing the impact on sampling time. Appendix A.1 describes the details of Early Termination. + +Factors Contributing to Speedup The primary factors contributing to the latency of MDPS are: + +1. Minimum Distance Curve Estimation: This step, requir + +![](images/b6010c9a380db52d887551a50b09d33c6045f13ae9f5532d0eaa12e43f001d09.jpg) +Figure 6. Redundancy Free Neighbor Search + +ing 1/10 of the original FPS iterations, adds only a fraction of the original FPS latency. + +2. Exclusion List Construction: The required all-to-all distance calculations are fully parallelizable, resulting in minimal latency overhead. Detailed analysis on this factor is provided in Appendix A.2. +3. Sampling Process: The use of exclusion list eliminates the need for additional distance calculations. Early termination requires additional FPS iterations. However, the impact is minimal: the overhead of $2.02\%$ , $0.58\%$ , and $1.18\%$ of original FPS iterations in S3DIS, ScanNet, and SemanticKITTI dataset, respectively. + +Due to these optimizations, MDPS can achieve a $4 - 5 \times$ speedup relative to the original FPS algorithm. Detailed latency breakdown of MDPS is provided in Appendix B.2. + +# 4.2. Redundancy-Free Neighbor Search + +The exclusion list in MDPS captures spatial relationships between points, providing an opportunity to optimize subsequent Ball Query and k-NN operations. We introduce Redundancy Free Ball Query and k-NN, which fully leverage the exclusion list from the sampling stage. + +Redundancy-Free Ball Query Ball query, which identifies neighbors within a radius $R$ , is performed after sampling, before aggregation (refer to Figure 1). The exclusion list is well-suited for this task, as it contains points within a specified radius of each point. By adding an extra list to the exclusion list for the radius $R$ , as illustrated in Figure 6, we can directly use it for ball query. As discussed in Section 4.1, this approach minimizes computational overhead; distances between points are calculated only once, regardless of the number of segments. After sampling, we extract the relevant entries from the corresponding exclusion list for the ball query, which serve as centroids for aggregation. + +Redundancy-Free k-NN While directly reusing the exclusion list for k-NN is challenging, we can leverage it to reduce the search space. By using the exclusion list of Seg + +
DatasetAverage Minimum Distance (Ratio to Baseline)
Baseline (FPS)Random SamplingGrid SamplingMDPS (Ours)
S3DIS0.063950.03478 (54.38%)0.04971 (77.72%)0.06359 (99.45%)
ScanNet0.037010.02035 (54.99%)0.02400 (64.84%)0.03677 (99.35%)
SemanticKITTI0.255050.12775 (50.09%)0.21519 (84.37%)0.25087 (98.36%)
+ +ment 1, we can limit the search space by eliminating points that are too far away to be potential k-NN neighbors. For example, to find the three nearest neighbors of point P7 in Figure 6, we calculate distances between P7 and the points from the corresponding exclusion list to identify the three nearest neighbors. Because k-NN upsampling relies on the downsampled set, points that were not downsampled are excluded from the search space. This approach significantly reduces the number of distance calculations, leading to substantial latency reduction. + +# 5. Evaluation + +# 5.1. Methodology + +We evaluate FastPoint on PointMetaBase [17] and PointVector [6], two recent models based on PointNet++. We evaluate our proposal on both indoor and outdoor datasets to demonstrate its scalability. We use S3DIS [1] and ScanNet [5] for indoor datasets and use SemanticKITTI [2] for outdoor datasets. + +The end-to-end latency is measured for processing all scenes in each dataset's validation split. We use mean Intersection over Union (mIoU) as a model accuracy evaluation metric. Detailed evaluation methodologies follow the same approach in PointMetaBase [17], and PointVector [6]. We apply FastPoint and the other baseline methods only to the first layer of each model as the first layer's FPS and neighbor search dominates the total computation time ( $>90\%$ ). + +We implement FastPoint with custom CUDA kernels and integrate them into OpenPoints library. We also integrate the software implementation [11] of QuickFPS [12] opensourced by the authors to OpenPoints for comparison. The experiments are conducted using an NVIDIA Geforce RTX 3090 GPU with 24GB of memory. We use CUDA 11.8 and PyTorch 1.10.1 for software setup. + +# 5.2. Sampling Quality of MDPS + +To evaluate sampling quality, we compute the average minimum distance between neighboring sampled points. A larger average distance indicates better coverage of sample points from the perspectives of point distribution and structure preservation. As shown in Table 1, MDPS achieves sampling quality comparable to FPS, with an average min + +Table 1. Comparison of Sampling Quality among FPS, Random Sampling, Grid Sampling and MDPS. + +
ModelDatasetmIoU (Diff. to Baseline FPS)
Baseline (FPS)Random SamplingGrid SamplingMDPS (Ours)
PV-LS3DIS71.3364.83 (-6.5)70.97 (-0.36)71.37 (+0.04)
ScanNet70.7059.94 (-10.76)69.61 (-1.09)70.63 (-0.07)
Semantic KITTI50.9138.46 (-12.45)50.62 (-0.29)50.79 (-0.12)
PMB-LS3DIS69.7268.21 (-1.51)69.54 (-0.18)69.74 (+0.02)
ScanNet70.8667.66 (-3.2)70.36 (-0.50)70.89 (+0.03)
Semantic KITTI52.1947.26 (-4.93)52.27 (+0.08)52.09 (-0.10)
+ +Table 2. Accuracy comparison. PV and PMB stand for PointVector and PointMetaBase. + +imum distance reaching over $99\%$ of the FPS baseline for S3DIS and ScanNet, and over $98\%$ for SemanticKITTI. In contrast, Random Sampling and Grid Sampling [29] exhibit significantly lower quality, with their average minimum distances deviating substantially from the FPS baseline. We also compare the minimum spacing distribution of each sampling method via visualization in Appendix B.9 and evaluate the robustness of MDPS in Appendix B.7. + +# 5.3. Accuracy Impact of MDPS + +We analyze the impact of MDPS on accuracy by applying it to the inference stage of PointNet++ models initially trained with FPS. Since MDPS is designed to closely replicate the FPS sampling pattern, it is compatible with FPS-trained models, thereby minimizing accuracy degradation. Note that Redundancy-Free Neighbor Search does not introduce any approximation, thus has no impact on accuracy. To compare the effects of different sampling methods, we also evaluate random sampling and grid sampling1. + +Table 2 shows that MDPS excels in maintaining accuracy, with negligible differences observed across all datasets and models. This highlights its effectiveness in replicating FPS behavior. However, the grid sampling exhibits larger and less consistent accuracy drops. When applied only during inference, it leads to declines of up to $1.09\%$ (PointVector-L, ScanNet). While it outperforms FastPoint in a single instance (PointMetaBase-L, SemanticKITTI), its performance is generally inferior, with significant accuracy reductions in other scenarios. The random sampling shows considerable accuracy drops across all data points. + +Overall, FastPoint effectively preserves accuracy by closely approximating the sampling pattern of FPS, achieving accuracy levels near-identical to FPS without retraining. Latency-accuracy comparisons (Appendix B.5) show that FastPoint resides above the pareto front, emphasizing the effectiveness of FastPoint. + +
MethodNumber of SegmentsEvaluation Metrics
Accuracy (mIoU)Sampling Time (ms)Sampling Quality
FPS-70.7085.830.03701
MDPS (Ours)1 Segment70.3418.620.03532
2 Segments70.6019.430.03598
3 Segments70.4619.790.03636
4 Segments70.5421.010.03662
5 Segments70.5221.410.03665
6 Segments70.6322.060.03662
7 Segments70.6322.750.03664
+ +![](images/8a4407dc43eaa96414579591ba55d0eb1d1540328aa6b50592e9a3a806c6e316.jpg) +Figure 7. End-to-end speedup of FastPoint and QuickFPS. + +# 5.4. Latency Reduction of FastPoint + +To evaluate the impact of FastPoint on end-to-end latency, we compare its latency to baseline FPS and QuickFPS. Note that FastPoint includes both MDPS and Redundancy-Free Neighbor Search. As explained in Section 2.2, QuickFPS accelerates FPS using k-d tree without approximation. Since MDPS also contains FPS steps when predicting the minimum distance curve, we can achieve further latency reductions by replacing them with QuickFPS, meaning that FastPoint and QuickFPS can be synergetic. + +Figure 7 presents the results of this comparison. Across S3DIS, ScanNet, and SemanticKITTI datasets, FastPoint achieves substantial latency reductions, with a geomean end-to-end speedup of $2.55 \times$ compared to baseline FPS. Integrating QuickFPS with FastPoint further reduces the execution latency achieving a geomean end-to-end speedup of $2.76 \times$ compared to baseline FPS and $1.41 \times$ compared to standalone QuickFPS. Speedup specific to the sampling and neighbor search are provided in Appendix B.3 and B.4. + +These findings substantiate the effectiveness of Fast-Point in reducing latency, surpassing existing methods like QuickFPS. Moreover, the successful combination of Fast-Point with QuickFPS demonstrates its compatibility with existing FPS optimizations. Scalability to Non-PointNet++ based models are discussed in Appendix B.10. + +Table 3. Comparison of mIoU, sampling time, and sampling quality for FPS and MDPS across segments 1 to 7. Sampling time is measured with a single scene of ScanNet dataset. PointVector-L model is used for mIoU comparison. + +
ModelDatasetMethodSpeedup
PV-LS3DISMDPS1.90×
All1.98×
ScanNetMDPS2.33×
All2.52×
Semantic KITTIMDPS2.07×
All2.20×
PMB-LS3DISMDPS2.50×
All2.82×
ScanNetMDPS2.89×
All3.38×
Semantic KITTIMDPS2.31×
All2.62×
+ +(a) Comparison of component-wise end-to-end speedup of FastPoint. + +
Model# of PointsSpeedup
PV-L16,0001.33×
32,0002.41×
48,0002.68×
64,0003.01×
80,0003.25×
96,0003.40×
PMB-L16,0001.58×
32,0002.64×
48,0002.91×
64,0003.06×
80,0003.30×
96,0003.44×
+ +(b) Comparison of speedup as the number of points increases. + +Table 4. Ablation study for speedup breakdown and scalability. + +# 5.5. Ablation Study + +Segment Count We evaluate the impact of varying the segment count on accuracy, sampling time, and sampling quality (i.e., average minimum distance). Table 3 shows that while sampling time increases gradually with more segments due to the exclusion list construction (refer to Section 4.1), both sampling quality and accuracy improve. However, they get saturated beyond a certain number of segments. Based on the empirical analysis, we use 6 segments by default, where accuracy saturates for all workloads. + +Component-wise Speedup To understand the individual contributions of each component in FastPoint, we evaluate the end-to-end speedup by sequentially applying MDPS and Redundancy Free Neighbor Search. Table 4a shows that MDPS alone provides significant speedup. Incorporating Redundancy Free Neighbor Search further provides additional improvement, demonstrating that each component makes a distinct contribution to the overall speedup. + +# 5.6. Scalability Study + +To assess the scalability of FastPoint, we evaluate the end-to-end speedup compared to FPS by varying point cloud sizes. Table 4b shows that FastPoint consistently achieves higher speedups as the number of points increases. This trend highlights the efficiency of our approach, particularly when processing larger point clouds where the computational cost of the FPS becomes a significant bottleneck. + +# 6. Conclusion + +3D point clouds are critical for representing and understanding 3D scenes, and deep learning models like PointNet++ have shown great promise in processing this data. However, computational efficiency remains a key challenge. We propose FastPoint, a software acceleration technique that leverages predictable minimum distances between sampled points to efficiently identify subsequent points without exhaustive computations. Our evaluation shows a $2.55 \times$ speedup over baseline FPS while maintaining accuracy, enabling low-cost deployment of PointNet++ based models. + +# Acknowledgements + +This work was supported by the Ministry of Science and ICT (MSIT), Korea, under the following grants: the Global Scholars Invitation Program (RS-2024-00456287), the Graduate School of Artificial Intelligence Semiconductor grant (IITP-2025-RS-2023-00256081), and the Artificial Intelligence Innovation Hub (No. 2021-0-02068), all supervised by the Institute for Information & Communications Technology Planning & Evaluation (IITP). Additional support was provided by the National Research Foundation of Korea (NRF) grant (RS-2024-00405857) funded by the Korea government (MSIT), and the Google Faculty Research Award. The source code is available at https://github.com/SNU-ARC/FastPoint.git. Jae W. Lee and Hongil Yoon are the corresponding authors. + +# References + +[1] Iro Armeni, Ozan Sener, Amir R. Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In CVPR, 2016. 7 +[2] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Juergen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In ICCV, 2019. 7 +[3] Yukang Chen, Jianhui Liu, Xiangyu Zhang, Xiaojuan Qi, and Jiaya Jia. Largekernel3d: Scaling up kernels in 3d sparse cnns. In CVPR, 2023. 2 +[4] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, 2019. 1, 2 +[5] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 7 +[6] Xin Deng, WenYu Zhang, Qing Ding, and XinMing Zhang. Pointvector: A vector representation in point cloud analysis. In CVPR, 2023. 1, 2, 7 +[7] Lue Fan, Ziqi Pang, Tianyuan Zhang, Yu-Xiong Wang, Hang Zhao, Feng Wang, Naiyan Wang, and Zhaoxiang Zhang. Embracing Single Stride 3D Object Detector with Sparse Transformer. In CVPR, 2022. 1, 2 +[8] Yu Feng, Boyuan Tian, Tiancheng Xu, Paul Whatmough, and Yuhao Zhu. Mesorasi: Architecture support for point cloud analytics via delayed-aggregation. In Proceedings of the 53th International Symposium on Microarchitecture (MICRO), 2020. 1 +[9] Yu Feng, Gunnar Hammonds, Yiming Gan, and Yuhao Zhu. Crescent: Taming memory irregularities for accelerating deep point cloud analytics. In Proceedings of the 49th Annual International Symposium on Computer Architecture (ISCA), 2022. 1 +[10] Lei Han, Tian Zheng, Lan Xu, and Lu Fang. Occuseg: Occupancy-aware 3d instance segmentation. In CVPR, 2020. 2 + +[11] Meng Han, Liang Wang, Limin Xiao, Hao Zhang, Chenhao Zhang, Xiangrong Xu, and Jianfeng Zhu. Quickfps. http://github.com/hanm2019/bucket-based_farthest-point-sampling_GPU.2,7 +[12] Meng Han, Liang Wang, Limin Xiao, Hao Zhang, Chenhao Zhang, Xiangrong Xu, and Jianfeng Zhu. Quickfps: Architecture and algorithm co-design for farthest point sampling in large-scale point clouds. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 2023. 1, 2, 7 +[13] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. In CVPR, 2020. 1, 2 +[14] Donghyun Lee, Yejin Lee, Jae W. Lee, and Hongil Yoon. Frugal 3d point cloud model training via progressive near point filtering and fused aggregation. In ECCV, 2024. 3 +[15] Jingtao Li, Jian Zhou, Yan Xiong, Xing Chen, and Chaitali Chakrabarti. An adjustable farthest point sampling method for approximately-sorted point cloud data. In 2022 IEEE Workshop on Signal Processing Systems (SiPS), 2022. 1, 3 +[16] Yaoxiu Lian, Xinhao Yang, Ke Hong, Yu Wang, Guohao Dai, and Ningyi Xu. A point transformer accelerator with fine-grained pipelines and distribution-aware dynamic fps. In IEEE/ACM International Conference on Computer-Aided Design (ICCAD), 2023. 1, 2 +[17] Haojia Lin, Xiawu Zheng, Lijiang Li, Fei Chao, Shanshan Wang, Yan Wang, Yonghong Tian, and Rongrong Ji. Meta architecture for point cloud analysis. In CVPR, 2023. 1, 2, 3, 7 +[18] Xueyuan Liu, Zhuoran Song, Guohao Dai, Gang Li, Can Xiao, Yan Xiang, Dehui Kong, Ke Xu, and Xiaoyao Liang. Fusionarch: A fusion-based accelerator for point-based point cloud neural networks. In ACM/IEEE Design, Automation and Test in Europe Conference (DATE), 2024. 1 +[19] Yongcheng Liu, Bin Fan, Gaofeng Meng, Jiwen Lu, Shiming Xiang, and Chunhong Pan. Densepoint: Learning densely contextual representation for efficient point cloud processing. In ICCV, 2019. 2 +[20] Zhijian Liu, Xinyu Yang, Haotian Tang, Shang Yang, and Song Han. Flatformer: Flattened window attention for efficient point cloud transformer. In CVPR, 2023. 1 +[21] Jiageng Mao, Yujing Xue, Minzhe Niu, Haoyue Bai, Jiashi Feng, Xiaodan Liang, Hang Xu, and Chunjing Xu. Voxel transformer for 3d object detection. In ICCV, 2021. 2 +[22] Alexey Nekrasov, Jonas Schult, Or Litany, Bastian Leibe, and Francis Engelmann. Mix3D: Out-of-Context Data Augmentation for 3D Scenes. In International Conference on 3D Vision (3DV), 2021. 2 +[23] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. arXiv preprint arXiv:1612.00593, 2016. 1, 2 +[24] Charles R Qi, Li Yi, Hao Su, and Leonidas J Guibas. Point-net++: Deep hierarchical feature learning on point sets in a metric space. arXiv preprint arXiv:1706.02413, 2017. 1, 2 + +[25] Charles R. Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas J. Guibas. Frustum pointnets for 3d object detection from rgb-d data. In CVPR, 2018. +[26] Guocheng Qian, Hasan Hammoud, Guohao Li, Ali Thabet, and Bernard Ghanem. Assanet: An anisotropical separable set abstraction for efficient point cloud representation learning. In NeurIPS, 2021. 1 +[27] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. In NeurIPS, 2022. 2 +[28] Danila Rukhovich, Anna Vorontsova, and Anton Konushin. Imvoxelnet: Image to voxels projection for monocular and multi-view general-purpose 3d object detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2022. 2 +[29] Hugues Thomas, Charles R. Qi, Jean-Emmanuel Deschaud, Beatriz Marcotegui, François Goulette, and Leonidas J. Guibas. Kpconv: Flexible and deformable convolution for point clouds. In ICCV, 2019. 2, 7 +[30] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E. Sarma, Michael M. Bronstein, and Justin M. Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (TOG), 2019. 1, 2 +[31] Wenxuan Wu, Zhongang Qi, and Li Fuxin. Pointconv: Deep convolutional networks on 3d point clouds. In CVPR, 2019. 2 +[32] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. In NeurIPS, 2022. 1 +[33] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler, faster, stronger. In CVPR, 2024. 1 +[34] Mutian Xu, Runyu Ding, Hengshuang Zhao, and Xiaojuan Qi. Paconv: Position adaptive convolution with dynamic kernel assembling on point clouds. In CVPR, 2021. 2 +[35] Qiangeng Xu, Xudong Sun, Cho-Ying Wu, Panqu Wang, and Ulrich Neumann. Grid-gen for fast and scalable point cloud learning, 2020. 2 +[36] Xinhao Yang, Tianyu Fu, Guohao Dai, Shulin Zeng, Kai Zhong, Ke Hong, and Yu Wang. An efficient accelerator for point-based and voxel-based point cloud neural networks. In ACM/IEEE Design Automation Conference (DAC), 2023. 1, 2 +[37] Yu-Qi Yang, Yu-Xiao Guo, Jian-Yu Xiong, Yang Liu, Hao Pan, Peng-Shuai Wang, Xin Tong, and Baining Guo. Swin3d: A pretrained transformer backbone for 3d indoor scene understanding. arXiv preprint arXiv:2304.06906, 2023.1,2 +[38] Yu-Qi Yang, Yu-Xiao Guo, and Yang Liu. Swin3d++: Effective multi-source pretraining for 3d indoor scene understanding. arXiv preprint arXiv:2304.06906, 2024. 2 +[39] Ziyu Ying, Sandeepa Bhuyan, Yan Kang, Yingtian Zhang, Mahmut T. Kandemir, and Chita R. Das. Edgepc: Efficient deep learning analytics for point clouds on edge devices. In Proceedings of the 50th Annual International Symposium on Computer Architecture (ISCA), 2023. 1, 2 + +[40] Hyunsung Yoon and Jae-Joon Kim. Efficient sampling and grouping acceleration for point cloud deep learning via single coordinate comparison. In IEEE/ACM International Conference on Computer-Aided Design (ICCAD), 2023. 1 +[41] Hengshuang Zhao, Li Jiang, Chi-Wing Fu, and Jiaya Jia. PointWeb: Enhancing local neighborhood features for point cloud processing. In CVPR, 2019. 2 +[42] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip Torr, and Vladlen Koltun. Point transformer. In ICCV, 2021. 1 +[43] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. arXiv preprint arXiv:2011.10033, 2020. 1, 2 \ No newline at end of file diff --git a/2025/FastPoint_ Accelerating 3D Point Cloud Model Inference via Sample Point Distance Prediction/images.zip b/2025/FastPoint_ Accelerating 3D Point Cloud Model Inference via Sample Point Distance Prediction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..94d426c1f16b54055427e4fd006cd0f4f0c1634a --- /dev/null +++ b/2025/FastPoint_ Accelerating 3D Point Cloud Model Inference via Sample Point Distance Prediction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e4e7d08d3b12f53b864d1505281b88f2e7715c6d23a26c4d0357ffe34b247b3 +size 512250 diff --git a/2025/FastPoint_ Accelerating 3D Point Cloud Model Inference via Sample Point Distance Prediction/layout.json b/2025/FastPoint_ Accelerating 3D Point Cloud Model Inference via Sample Point Distance Prediction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e3aa47b4e38f09e3eaa42bb27528b3d483148049 --- /dev/null +++ b/2025/FastPoint_ Accelerating 3D Point Cloud Model Inference via Sample Point Distance Prediction/layout.json @@ -0,0 +1,8615 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 120, + 103, + 492, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 103, + 492, + 138 + ], + "spans": [ + { + "bbox": [ + 120, + 103, + 492, + 138 + ], + "type": "text", + "content": "FastPoint: Accelerating 3D Point Cloud Model Inference via Sample Point Distance Prediction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 160, + 471, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 160, + 471, + 190 + ], + "spans": [ + { + "bbox": [ + 138, + 160, + 471, + 190 + ], + "type": "text", + "content": "Donghyun Lee" + }, + { + "bbox": [ + 138, + 160, + 471, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 160, + 471, + 190 + ], + "type": "text", + "content": " Dawoon Jeong" + }, + { + "bbox": [ + 138, + 160, + 471, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 160, + 471, + 190 + ], + "type": "text", + "content": " Jae W. Lee" + }, + { + "bbox": [ + 138, + 160, + 471, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 160, + 471, + 190 + ], + "type": "text", + "content": " Hongil Yoon" + }, + { + "bbox": [ + 138, + 160, + 471, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 138, + 160, + 471, + 190 + ], + "type": "text", + "content": " Seoul National University " + }, + { + "bbox": [ + 138, + 160, + 471, + 190 + ], + "type": "inline_equation", + "content": "{ }^{2}" + }, + { + "bbox": [ + 138, + 160, + 471, + 190 + ], + "type": "text", + "content": "Google" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 191, + 474, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 191, + 474, + 204 + ], + "spans": [ + { + "bbox": [ + 133, + 191, + 474, + 204 + ], + "type": "text", + "content": "{eudh1206, daun20211, jaewlee}@snu.ac.kr, hongilyoon@google.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 257, + 297, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 257, + 297, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 297, + 426 + ], + "type": "text", + "content": "Deep neural networks have revolutionized 3D point cloud processing, yet efficiently handling large and irregular point clouds remains challenging. To tackle this problem, we introduce FastPoint, a novel software-based acceleration technique that leverages the predictable distance trend between sampled points during farthest point sampling. By predicting the distance curve, we can efficiently identify subsequent sample points without exhaustively computing all pairwise distances. Our proposal substantially accelerates farthest point sampling and neighbor search operations while preserving sampling quality and model performance. By integrating FastPoint into state-of-the-art 3D point cloud models, we achieve " + }, + { + "bbox": [ + 55, + 257, + 297, + 426 + ], + "type": "inline_equation", + "content": "2.55 \\times" + }, + { + "bbox": [ + 55, + 257, + 297, + 426 + ], + "type": "text", + "content": " end-to-end speedup on NVIDIA RTX 3090 GPU without sacrificing accuracy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 451, + 135, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 135, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 135, + 464 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 472, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 616 + ], + "type": "text", + "content": "3D point clouds have become increasingly important for representing and understanding 3D scenes, driving advancements in fields like robotics and autonomous driving. Deep neural networks [4, 6, 7, 13, 17, 20, 23-26, 30, 32, 33, 37, 42, 43] have emerged as powerful tools for processing this type of data. While PointNet++ [24] and its successors have improved efficiency and performance of these point cloud models, the irregular nature and growing size of point cloud data continue to impose significant computational challenges, especially for real time use cases. Operations such as Farthest Point Sampling (FPS) and neighbor search remain major bottlenecks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": "While various hardware accelerators [8, 9, 12, 16, 18, 36, 40] have been explored to accelerate the models, software-only solutions are less common. Given the high cost and complexity of developing specialized accelerators, software-based acceleration is a more practical approach. Recently, several software-based techniques [15, 39] have been proposed to accelerate these models through approximation. However, such approaches often come at the cost" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 232, + 537, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 537, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 537, + 243 + ], + "type": "text", + "content": "of significant accuracy loss, limiting their practical use." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 244, + 554, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 244, + 554, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 244, + 554, + 316 + ], + "type": "text", + "content": "To address this challenge, we propose FastPoint, a novel software-based acceleration technique with sample point distance prediction. We observe that the inherent nature of FPS, which iteratively selects the point farthest from the current set of sampled points, leads to two predictable trends in the minimum distances between these points:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 316, + 554, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 316, + 554, + 376 + ], + "spans": [ + { + "bbox": [ + 315, + 316, + 554, + 376 + ], + "type": "text", + "content": "1. Decreasing minimum distance: As FPS makes progress, the number of remaining points decreases, forcing the algorithm to select points closer to those already sampled. This results in a smoothly decreasing curve of minimum distances between sampled points." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 376, + 553, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 553, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 553, + 411 + ], + "type": "text", + "content": "2. Early structure capture: The initial sample points tend to be the extremities of the point cloud, effectively capturing its overall shape and boundaries." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 412, + 554, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 554, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 554, + 579 + ], + "type": "text", + "content": "Our proposal leverages these observations to accurately estimate the distance curve using only a few initial FPS iterations. By predicting the distance curve, we can efficiently identify subsequent sample points without computing and comparing pairwise distances at each FPS iteration. This leads to significant latency savings while maintaining sampling quality comparable to that of the original FPS. Furthermore, predicting the distance curve enables the sampling process to be decoupled from distance computation. This decoupling exposes new opportunities for parallelizing distance calculations in the proposed sampling, further enhancing efficiency. This approach not only accelerates FPS itself but also benefits subsequent operations such as neighbor search by leveraging pre-computed distances." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 326, + 580, + 528, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 580, + 528, + 590 + ], + "spans": [ + { + "bbox": [ + 326, + 580, + 528, + 590 + ], + "type": "text", + "content": "Our key contributions are summarized as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 594, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 313, + 594, + 553, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 553, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 553, + 641 + ], + "type": "text", + "content": "- We identify two key latency bottlenecks in point cloud models: FPS and neighbor search. Limited parallelism in FPS and redundant distance computations across both operations are the primary sources of inefficiency." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 642, + 553, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 666 + ], + "type": "text", + "content": "- We empirically analyze 3D point cloud models and identify the decreasing trend and early structure capture trend." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "content": "- Based on the observations, we introduce a novel sampling approach with sample point distance prediction. With the estimated distance curve, this approach not only accelerates FPS itself but also benefits subsequent neighbor" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25114" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 558, + 216 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 558, + 216 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 558, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 558, + 216 + ], + "type": "image", + "image_path": "1cb09cf346bf4144a51258661e69dc8af82c06e8da911b5af9dcbc417c035d59.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 221, + 419, + 234 + ], + "lines": [ + { + "bbox": [ + 192, + 221, + 419, + 234 + ], + "spans": [ + { + "bbox": [ + 192, + 221, + 419, + 234 + ], + "type": "text", + "content": "Figure 1. Overview of PointNet++ Based Model Architecture" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 64, + 247, + 258, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 247, + 258, + 258 + ], + "spans": [ + { + "bbox": [ + 64, + 247, + 258, + 258 + ], + "type": "text", + "content": "search, which relies on pre-computed distances." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 259, + 296, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 259, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 259, + 296, + 308 + ], + "type": "text", + "content": "- We integrate FastPoint into OpenPoints library and experiment on state-of-the-art PointNet++ based models, achieving geomean " + }, + { + "bbox": [ + 56, + 259, + 296, + 308 + ], + "type": "inline_equation", + "content": "2.55 \\times" + }, + { + "bbox": [ + 56, + 259, + 296, + 308 + ], + "type": "text", + "content": " end-to-end speedup on NVIDIA RTX 3090 GPU while preserving accuracy." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 316, + 231, + 329 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 231, + 329 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 231, + 329 + ], + "type": "text", + "content": "2. Background and Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 335, + 242, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 242, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 242, + 348 + ], + "type": "text", + "content": "2.1. Deep Learning on 3D Point Clouds" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 353, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 353, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 353, + 296, + 555 + ], + "type": "text", + "content": "Deep Neural Networks on Raw Point Cloud Since the release of PointNet [23], which is the first to apply deep neural networks directly to the raw point clouds, numerous successors [6, 13, 17, 19, 24-27, 30, 31, 34, 41] have been developed, improving both model performance and efficiency. PointNet++ [24] introduces a hierarchical structure to PointNet through sampling and grouping, which has become the core architecture of subsequent PointNet++ based models. PointNeXt [27] revisits PointNet++, achieving substantial improvements in model performance, while PointVector [6] and PointMetaBase [17] push the model performance and efficiency even further. Although the architectural details of each model differ, all models follow the PointNet++ structure, utilizing Set Abstraction and Feature Propagation layers as core components. This work focuses on identifying and accelerating the primary latency bottlenecks in PointNet++ based models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "Deep Neural Networks on Voxelized Point Cloud Several works [3, 4, 10, 22, 28, 43] propose to voxelize point clouds, using voxels as input rather than raw points. The advent of 3D sparse convolution [4] has made this voxel-based approach increasingly popular due to its efficiency and high performance. Recently, transformer-based models [7, 21, 37, 38] operating on voxelized point clouds are gaining significant attention. Although voxelization alleviates the computational cost of mapping operations (i.e., downsampling and neighbor search), it has a notable limitation due to the loss of position information during the voxelization process." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 246, + 403, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 246, + 403, + 258 + ], + "spans": [ + { + "bbox": [ + 314, + 246, + 403, + 258 + ], + "type": "text", + "content": "2.2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 264, + 555, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 555, + 312 + ], + "type": "text", + "content": "Numerous solutions have been proposed to accelerate sampling and neighbor search operations in PointNet++ based models. We categorize them into hardware- and purely software-based approaches." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 327, + 556, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 556, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 556, + 495 + ], + "type": "text", + "content": "Hardware-Assisted Acceleration QuickFPS [12] introduces a k-d tree based FPS algorithm which reduces computation and memory access in each FPS iteration, along with specialized hardware designed for this method. MARS [36] and PTrAcc [16] both propose hardware accelerators that employ a distance filtering technique which skips unnecessary distance computations in each iteration of FPS. Although these techniques achieve substantial speedups in FPS without any loss in sampling quality, they require specialized hardware to fully harness their algorithms, limiting their effectiveness on commodity hardware like GPU. For more thorough evaluation, we compare the speedup of our proposal with pure software version [11] of QuickFPS in Section 5.4, evaluating its effectiveness on GPU." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 510, + 556, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 689 + ], + "type": "text", + "content": "Software-Only Acceleration To mitigate the high latency associated with FPS, various alternative sampling methods have been explored. RandLA-Net [13] replaces FPS with random sampling and introduces a local feature aggregation module that compensates for the low sampling quality of random sampling. Grid sampling, the most commonly used alternative, is adopted by Grid-GCN [35] and KPCov [29], offering faster processing with reasonable sampling quality compared to random sampling. Despite the latency advantages of these methods, none have consistently outperformed FPS for overall performance. This is illustrated by the fact that most PointNet++-based models [6, 17, 24-27] continue to rely on FPS, highlighting its broad applicability and superior sampling quality. The experimental results in Appendix B.1 corroborate this claim." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "EdgePC [39] structures the point cloud with morton codes to accelerate FPS and neighbor search on edge GPUs," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25115" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 72, + 282, + 219 + ], + "blocks": [ + { + "bbox": [ + 69, + 72, + 282, + 219 + ], + "lines": [ + { + "bbox": [ + 69, + 72, + 282, + 219 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 282, + 219 + ], + "type": "image", + "image_path": "ae2db72dee1c17f01cd52410b02dd70628795cce94c8d8d03b0d8080d45789cf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 228, + 285, + 239 + ], + "lines": [ + { + "bbox": [ + 66, + 228, + 285, + 239 + ], + "spans": [ + { + "bbox": [ + 66, + 228, + 285, + 239 + ], + "type": "text", + "content": "Figure 2. Latency Breakdown of PointNet++ Based Models" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 254, + 296, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 254, + 296, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 254, + 296, + 411 + ], + "type": "text", + "content": "while Adjustable FPS [15] accelerates FPS by leveraging the intrinsic locality of the point cloud data. However, both methods demonstrate significant accuracy loss compared to the baseline (i.e., up to " + }, + { + "bbox": [ + 55, + 254, + 296, + 411 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 55, + 254, + 296, + 411 + ], + "type": "text", + "content": " mIoU loss) due to their aggressive approximations, limiting their effectiveness. Prior work [14] accelerates the training of PointNet++ based models by precomputing the minimum point spacing of farthest-sampled points before training and reusing this value across epochs for fast sampling. While this approach effectively speeds up training, it is not suitable for inference, as the minimum spacing value cannot be determined in advance in inference scenarios. Detailed comparisons with our work are provided in Appendix A.4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 423, + 244, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 244, + 436 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 244, + 436 + ], + "type": "text", + "content": "3. Challenges of Point Cloud Models" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 443, + 218, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 443, + 218, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 218, + 455 + ], + "type": "text", + "content": "3.1. Model Architecture Overview" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 461, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 295, + 533 + ], + "type": "text", + "content": "Figure 1 illustrates the detailed process of PointNet++-based models. Similar to 2D convolutional neural networks, these models consist of both downsampling and upsampling layers. Commonly referred to as the set abstraction layer, the downsampling layer in PointNet++-based models comprises three main stages, as outlined below:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 534, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 62, + 534, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 534, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 62, + 534, + 295, + 594 + ], + "type": "text", + "content": "1 Farthest Point Sampling (FPS): Downsampling begins by selecting " + }, + { + "bbox": [ + 62, + 534, + 295, + 594 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 62, + 534, + 295, + 594 + ], + "type": "text", + "content": "/stride output points from " + }, + { + "bbox": [ + 62, + 534, + 295, + 594 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 62, + 534, + 295, + 594 + ], + "type": "text", + "content": " input points, where stride represents the downsampling rate. Typically, the FPS is used to maintain the shape and structure of the input point cloud." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 594, + 295, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 594, + 295, + 630 + ], + "spans": [ + { + "bbox": [ + 62, + 594, + 295, + 630 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 62, + 594, + 295, + 630 + ], + "type": "text", + "content": " Ball Query: Next, a ball query with radius " + }, + { + "bbox": [ + 62, + 594, + 295, + 630 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 62, + 594, + 295, + 630 + ], + "type": "text", + "content": " is performed around each sampled point to gather neighboring points for feature aggregation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 630, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 630, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 630, + 295, + 712 + ], + "type": "text", + "content": "Aggregation: The features of neighboring points identified in the ball query are then processed. A multi-layer perceptron (MLP) is applied to these features, followed by max pooling to aggregate the information for each centroid. The details of this process, e.g., positional embedding [17], can vary depending on the model architectures." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 315, + 72, + 468, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 468, + 84 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 468, + 84 + ], + "type": "text", + "content": "Algorithm 1 Farthest Point Sampling" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 86, + 547, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 86, + 547, + 107 + ], + "spans": [ + { + "bbox": [ + 315, + 86, + 547, + 107 + ], + "type": "text", + "content": "Input " + }, + { + "bbox": [ + 315, + 86, + 547, + 107 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 315, + 86, + 547, + 107 + ], + "type": "text", + "content": " : input point cloud, " + }, + { + "bbox": [ + 315, + 86, + 547, + 107 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 315, + 86, + 547, + 107 + ], + "type": "text", + "content": " : number of points to sample, " + }, + { + "bbox": [ + 315, + 86, + 547, + 107 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 315, + 86, + 547, + 107 + ], + "type": "text", + "content": " : number of original points" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 108, + 489, + 118 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 108, + 489, + 118 + ], + "spans": [ + { + "bbox": [ + 315, + 108, + 489, + 118 + ], + "type": "text", + "content": "Output sampledidx: indices of sampled points" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 319, + 120, + 548, + 209 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 319, + 120, + 548, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 120, + 548, + 131 + ], + "spans": [ + { + "bbox": [ + 319, + 120, + 548, + 131 + ], + "type": "text", + "content": "1: sampled_idx[0] " + }, + { + "bbox": [ + 319, + 120, + 548, + 131 + ], + "type": "inline_equation", + "content": "\\leftarrow" + }, + { + "bbox": [ + 319, + 120, + 548, + 131 + ], + "type": "text", + "content": " seed_idx // Starts with a random point" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 319, + 132, + 473, + 141 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 132, + 473, + 141 + ], + "spans": [ + { + "bbox": [ + 319, + 132, + 473, + 141 + ], + "type": "text", + "content": "2: min_dists[i] " + }, + { + "bbox": [ + 319, + 132, + 473, + 141 + ], + "type": "inline_equation", + "content": "\\leftarrow \\infty" + }, + { + "bbox": [ + 319, + 132, + 473, + 141 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 319, + 132, + 473, + 141 + ], + "type": "inline_equation", + "content": "i = 0,\\dots,N - 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 319, + 143, + 488, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 143, + 488, + 152 + ], + "spans": [ + { + "bbox": [ + 319, + 143, + 488, + 152 + ], + "type": "text", + "content": "3: for " + }, + { + "bbox": [ + 319, + 143, + 488, + 152 + ], + "type": "inline_equation", + "content": "i \\gets 1" + }, + { + "bbox": [ + 319, + 143, + 488, + 152 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 319, + 143, + 488, + 152 + ], + "type": "inline_equation", + "content": "n - 1" + }, + { + "bbox": [ + 319, + 143, + 488, + 152 + ], + "type": "text", + "content": " do // Not parallelized" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 319, + 153, + 490, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 153, + 490, + 163 + ], + "spans": [ + { + "bbox": [ + 319, + 153, + 490, + 163 + ], + "type": "text", + "content": "4: for " + }, + { + "bbox": [ + 319, + 153, + 490, + 163 + ], + "type": "inline_equation", + "content": "j\\gets 0" + }, + { + "bbox": [ + 319, + 153, + 490, + 163 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 319, + 153, + 490, + 163 + ], + "type": "inline_equation", + "content": "N - 1" + }, + { + "bbox": [ + 319, + 153, + 490, + 163 + ], + "type": "text", + "content": " do // Parallelized" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 319, + 164, + 524, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 164, + 524, + 175 + ], + "spans": [ + { + "bbox": [ + 319, + 164, + 524, + 175 + ], + "type": "text", + "content": "5: " + }, + { + "bbox": [ + 319, + 164, + 524, + 175 + ], + "type": "inline_equation", + "content": "d_{\\text{new}} \\gets \\text{dist}(P[\\text{sampled\\_idx}[i - 1]], P[j])" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 319, + 176, + 461, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 176, + 461, + 186 + ], + "spans": [ + { + "bbox": [ + 319, + 176, + 461, + 186 + ], + "type": "text", + "content": "6: if min_dists[j] > " + }, + { + "bbox": [ + 319, + 176, + 461, + 186 + ], + "type": "inline_equation", + "content": "d_{new}" + }, + { + "bbox": [ + 319, + 176, + 461, + 186 + ], + "type": "text", + "content": " then" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 319, + 186, + 449, + 197 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 186, + 449, + 197 + ], + "spans": [ + { + "bbox": [ + 319, + 186, + 449, + 197 + ], + "type": "text", + "content": "7: min_dists[j] " + }, + { + "bbox": [ + 319, + 186, + 449, + 197 + ], + "type": "inline_equation", + "content": "\\leftarrow d_{new}" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 319, + 198, + 485, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 198, + 485, + 209 + ], + "spans": [ + { + "bbox": [ + 319, + 198, + 485, + 209 + ], + "type": "text", + "content": "8: sampledidx[i] " + }, + { + "bbox": [ + 319, + 198, + 485, + 209 + ], + "type": "inline_equation", + "content": "\\leftarrow" + }, + { + "bbox": [ + 319, + 198, + 485, + 209 + ], + "type": "text", + "content": " argmax(min_dists)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 246, + 553, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 246, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 553, + 304 + ], + "type": "text", + "content": "The upsampling layer, often referred to as a feature propagation layer, takes the downsampled points from the corresponding downsampling layer as inputs and propagates the features back to the original point cloud. This process is done by following steps." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 305, + 553, + 389 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 321, + 305, + 553, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 305, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 321, + 305, + 553, + 329 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 321, + 305, + 553, + 329 + ], + "type": "text", + "content": " k-NN: For each upsampled point, k-NN is applied to identify its k nearest neighbors among the input points." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 329, + 553, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 329, + 553, + 389 + ], + "spans": [ + { + "bbox": [ + 321, + 329, + 553, + 389 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 321, + 329, + 553, + 389 + ], + "type": "text", + "content": " Interpolation: Based on the k-NN results, neighboring input points are interpolated, concatenated with skip-connected features from the corresponding downsampling layer, and the information is passed through an MLP to propagate the features to the subsequent layer." + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 396, + 432, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 396, + 432, + 408 + ], + "spans": [ + { + "bbox": [ + 314, + 396, + 432, + 408 + ], + "type": "text", + "content": "3.2. Latency Breakdown" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 414, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 555, + 605 + ], + "type": "text", + "content": "Figure 2 presents the latency breakdown of PointMetaBase-L model evaluated on both indoor (S3DIS) and outdoor (SemanticKITTI) scenes. The results show that coordinate-based operations (i.e., FPS, k-NN, and Ball Query) dominate the execution time, consuming significantly more time than feature-based operations (i.e., Aggregation and MLP). These coordinate-based operations account for approximately " + }, + { + "bbox": [ + 313, + 414, + 555, + 605 + ], + "type": "inline_equation", + "content": "88.1\\%" + }, + { + "bbox": [ + 313, + 414, + 555, + 605 + ], + "type": "text", + "content": " of the total inference time for indoor scenes and " + }, + { + "bbox": [ + 313, + 414, + 555, + 605 + ], + "type": "inline_equation", + "content": "94.1\\%" + }, + { + "bbox": [ + 313, + 414, + 555, + 605 + ], + "type": "text", + "content": " for the outdoor scenes. Among these, FPS emerges as the primary performance bottleneck, requiring the most execution time. Additionally, neighbor search operations, including k-NN and Ball Query, contribute significantly to the overall latency. To address these challenges, we first analyze the root causes of these inefficiencies and propose novel optimization techniques in subsequent sections." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 612, + 526, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 526, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 526, + 625 + ], + "type": "text", + "content": "3.3. Inefficiencies in Farthest Point Sampling" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 630, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 712 + ], + "type": "text", + "content": "Farthest point sampling operates iteratively by selecting one point at a time. In each iteration, the sampling algorithm identifies the point farthest from the set of the previously sampled points. The detailed process is outlined in Algorithm 1. Initially, a random point in " + }, + { + "bbox": [ + 313, + 630, + 554, + 712 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 630, + 554, + 712 + ], + "type": "text", + "content": " is chosen as the seed point (Line 1) and added to the sampled_idx. Then, a distance matrix (i.e., min_dists) is created to store the distance" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25116" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": "between each point in " + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": " and the set of the previously sampled points, where the distance from a point to a point set is defined as minimum distance between the point and any point in the set. This matrix is initialized to " + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "inline_equation", + "content": "\\infty" + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": " (Line 2). In each iteration, the distance between the most recently sampled point and every other point in " + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": " is calculated (Lines 3-7). If a newly calculated distance is smaller than the corresponding entry in the min_dist matrix, the matrix is updated with the smaller value. Finally, the index of the point with the maximum distance in the min_dist is added to the sampled_idx (Line 8). This process is repeated until the specified number of points " + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "inline_equation", + "content": "(n)" + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": " is sampled." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 216, + 294, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 216, + 294, + 299 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 294, + 299 + ], + "type": "text", + "content": "The sequential nature of the FPS algorithm is the primary reason of the inefficiency. Because only one point is sampled per iteration, the " + }, + { + "bbox": [ + 55, + 216, + 294, + 299 + ], + "type": "inline_equation", + "content": "O(N)" + }, + { + "bbox": [ + 55, + 216, + 294, + 299 + ], + "type": "text", + "content": " distance calculations required in each iteration must be sequentially executed " + }, + { + "bbox": [ + 55, + 216, + 294, + 299 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 216, + 294, + 299 + ], + "type": "text", + "content": " times. This tight coupling between sampling and distance calculation significantly limits the potential for parallelism in distance computations." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 300, + 294, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 294, + 359 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 294, + 359 + ], + "type": "text", + "content": "To unlock this parallelism potential, we introduce Minimum Distance Prediction Sampling (MDPS) in Section 4.1, a novel sampling strategy that fully decouples sampling from distance computation, enabling greater parallelism and faster processing speeds." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 365, + 232, + 378 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 232, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 232, + 378 + ], + "type": "text", + "content": "3.4. Inefficiencies in Neighbor Search" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 383, + 294, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 383, + 294, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 383, + 294, + 467 + ], + "type": "text", + "content": "As explained in Section 3.1, PointNet++ based models employ two types of neighbor search algorithms: Ball Query and k-NN. These operations play a critical role in capturing local geometric relationships in the point cloud. Ball Query identifies neighbor points within a fixed radius around the downsampled points. k-NN identifies k nearest neighbors among the downsampled points for each original point." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 467, + 294, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 467, + 294, + 563 + ], + "spans": [ + { + "bbox": [ + 54, + 467, + 294, + 563 + ], + "type": "text", + "content": "The inefficiencies in these neighbor search operations stem from redundant distance computations in both the FPS and subsequent neighbor search operations. For instance, as shown in Figure 1, the distance between P0 and P2 is computed while updating P2's minimum distance in FPS. However, the same distance is redundantly recalculated in both Ball Query and k-NN. Section 4.2 introduces redundancy-free neighbor search techniques to address this issue." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 572, + 264, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 572, + 264, + 585 + ], + "spans": [ + { + "bbox": [ + 55, + 572, + 264, + 585 + ], + "type": "text", + "content": "4. FastPoint: Fast Point Cloud Inference" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 592, + 294, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 592, + 294, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 294, + 640 + ], + "type": "text", + "content": "Trends in Minimum Distance of Sample Points The FPS iteratively selects the point farthest from the current set of sampled points, which leads to predictable trends in the distance between the sampled points." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 642, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 642, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 642, + 294, + 713 + ], + "type": "text", + "content": "1. Decreasing minimum distance: As sampling proceeds, the pool of the remaining points shrinks, the FPS algorithm selects points closer to those already sampled over time. Thus, the maximized minimum distance (i.e., " + }, + { + "bbox": [ + 56, + 642, + 294, + 713 + ], + "type": "inline_equation", + "content": "\\max (min\\_dist)" + }, + { + "bbox": [ + 56, + 642, + 294, + 713 + ], + "type": "text", + "content": " in Algorithm 1) decreases, which we define as minimum distance for brevity. This is particu" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 314, + 76, + 397, + 170 + ], + "blocks": [ + { + "bbox": [ + 314, + 76, + 397, + 170 + ], + "lines": [ + { + "bbox": [ + 314, + 76, + 397, + 170 + ], + "spans": [ + { + "bbox": [ + 314, + 76, + 397, + 170 + ], + "type": "image", + "image_path": "55827e7bb3e7aff332dc1de514fd528fb813171139164c5a918de76c6fce7757.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 171, + 396, + 191 + ], + "lines": [ + { + "bbox": [ + 315, + 171, + 396, + 191 + ], + "spans": [ + { + "bbox": [ + 315, + 171, + 396, + 191 + ], + "type": "text", + "content": "(a) Minimum Distance Curve of ScanNet scene" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 402, + 78, + 475, + 171 + ], + "blocks": [ + { + "bbox": [ + 402, + 78, + 475, + 171 + ], + "lines": [ + { + "bbox": [ + 402, + 78, + 475, + 171 + ], + "spans": [ + { + "bbox": [ + 402, + 78, + 475, + 171 + ], + "type": "image", + "image_path": "e9ddf982be8815f4919e16673089ab17a2b4ffbcf22e1f86d439a2176cd5a2d7.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 398, + 171, + 554, + 191 + ], + "lines": [ + { + "bbox": [ + 398, + 171, + 554, + 191 + ], + "spans": [ + { + "bbox": [ + 398, + 171, + 554, + 191 + ], + "type": "text", + "content": "(b) Visualization of ScanNet scene sampled at 1/10 of the original iterations" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 321, + 201, + 546, + 212 + ], + "lines": [ + { + "bbox": [ + 321, + 201, + 546, + 212 + ], + "spans": [ + { + "bbox": [ + 321, + 201, + 546, + 212 + ], + "type": "text", + "content": "Figure 3. Motivation of Minimum Distance Curve Estimation" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 477, + 76, + 547, + 171 + ], + "blocks": [ + { + "bbox": [ + 477, + 76, + 547, + 171 + ], + "lines": [ + { + "bbox": [ + 477, + 76, + 547, + 171 + ], + "spans": [ + { + "bbox": [ + 477, + 76, + 547, + 171 + ], + "type": "image", + "image_path": "53faa4f40d59824438cfe9fd48e63d59c0fc5b24b0ba8a64c09b93452036366c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 325, + 226, + 553, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 226, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 325, + 226, + 553, + 250 + ], + "type": "text", + "content": "larly evident in Figure 3a, showing a smooth, decreasing curve for various input point clouds." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 250, + 553, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 250, + 553, + 309 + ], + "spans": [ + { + "bbox": [ + 314, + 250, + 553, + 309 + ], + "type": "text", + "content": "2. Early structure capture: Importantly, we observe that a few initial sample points are sufficient to capture the overall shape and boundaries of the input point cloud (Figure 3b). This is because the initial sample points tend to be the extremities of the point cloud." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 324, + 555, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 555, + 444 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 555, + 444 + ], + "type": "text", + "content": "Optimization Opportunities The trend of early structure capture suggests that the initial iterations of FPS are important in accurately representing the point cloud's structure. The Decreasing trend suggests that the later portion of the minimum distance curve becomes more predictable as sampling progresses. Building upon these trends, if we could accurately estimate this curve using only a few initial FPS iterations, we could maintain a comparable sampling quality while achieving substantial latency reduction by selecting points guided by the predicted curve." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 449, + 556, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 556, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 556, + 641 + ], + "type": "text", + "content": "Proposed Techniques We introduce two optimizations by leveraging these opportunities: Minimum Distance Prediction Sampling and Redundancy-Free Neighbor Search. Minimum Distance Prediction Sampling (MDPS) is a novel sampling strategy designed to approximate the sampling quality of FPS while significantly reducing latency. This approach estimates the distance curve using only a few initial non-parallelizable FPS iterations. Subsequent sample points are selected based on the predicted distances, eliminating the need for pairwise distance comparisons in later iterations. Furthermore, MDPS enables Redundancy-Free Neighbor Search by reusing precomputed distance information, thereby eliminating redundant computations during subsequent Ball Query and k-NN operations. This optimization further enhances the overall efficiency of PointNet++-based models during inference." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 648, + 525, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 648, + 525, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 648, + 525, + 661 + ], + "type": "text", + "content": "4.1. Minimum Distance Prediction Sampling" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "1 Minimum Distance Curve Estimation To build a low-cost estimator for the minimum distance curve (as shown in Figure 4), we aim to estimate the curve with minimal error using the results from as few FPS iterations as possible. We" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25117" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 68, + 556, + 277 + ], + "blocks": [ + { + "bbox": [ + 55, + 68, + 556, + 277 + ], + "lines": [ + { + "bbox": [ + 55, + 68, + 556, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 68, + 556, + 277 + ], + "type": "image", + "image_path": "29f6522d3450a20ddef2d4220cb9359d6f788d9dedb8a0cdcb5970978ad7f18b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 182, + 281, + 425, + 293 + ], + "lines": [ + { + "bbox": [ + 182, + 281, + 425, + 293 + ], + "spans": [ + { + "bbox": [ + 182, + 281, + 425, + 293 + ], + "type": "text", + "content": "Figure 4. Overall Flow of Minimum Distance Prediction Sampling" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 307, + 295, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 307, + 295, + 380 + ], + "spans": [ + { + "bbox": [ + 54, + 307, + 295, + 380 + ], + "type": "text", + "content": "formulate the estimator " + }, + { + "bbox": [ + 54, + 307, + 295, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 54, + 307, + 295, + 380 + ], + "type": "text", + "content": " as a function which gets initial " + }, + { + "bbox": [ + 54, + 307, + 295, + 380 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 54, + 307, + 295, + 380 + ], + "type": "text", + "content": " segment (where " + }, + { + "bbox": [ + 54, + 307, + 295, + 380 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 54, + 307, + 295, + 380 + ], + "type": "text", + "content": " is the ratio of initial FPS iterations to total iterations, " + }, + { + "bbox": [ + 54, + 307, + 295, + 380 + ], + "type": "inline_equation", + "content": "0 < p < 1" + }, + { + "bbox": [ + 54, + 307, + 295, + 380 + ], + "type": "text", + "content": ") of the curve " + }, + { + "bbox": [ + 54, + 307, + 295, + 380 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 54, + 307, + 295, + 380 + ], + "type": "text", + "content": " as input and outputs the rest of the estimated curve " + }, + { + "bbox": [ + 54, + 307, + 295, + 380 + ], + "type": "inline_equation", + "content": "\\hat{C}" + }, + { + "bbox": [ + 54, + 307, + 295, + 380 + ], + "type": "text", + "content": " (i.e., Equation 1). Note that " + }, + { + "bbox": [ + 54, + 307, + 295, + 380 + ], + "type": "inline_equation", + "content": "C(1)" + }, + { + "bbox": [ + 54, + 307, + 295, + 380 + ], + "type": "text", + "content": " represents the minimum distance among the sampled points after all sampling iterations are finished." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 86, + 387, + 295, + 401 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 387, + 295, + 401 + ], + "spans": [ + { + "bbox": [ + 86, + 387, + 295, + 401 + ], + "type": "interline_equation", + "content": "\\hat {C} (t) = \\mathcal {E} (C (t) \\mid t \\in [ 0, p ]) \\quad \\text {f o r} t \\in (p, 1 ] \\tag {1}", + "image_path": "b26f9274c6425537aa19da49705e3b304839c47241d373082ad3ca7774b53e8d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 408, + 295, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 408, + 295, + 553 + ], + "spans": [ + { + "bbox": [ + 54, + 408, + 295, + 553 + ], + "type": "text", + "content": "We have experimented with various models, including polynomial functions and MLPs, and find that MLPs yield the best performance (Appendix A.3). We train a 3-layer MLP on minimum distance curves extracted from the training split of each point cloud dataset, setting " + }, + { + "bbox": [ + 54, + 408, + 295, + 553 + ], + "type": "inline_equation", + "content": "p = 0.1" + }, + { + "bbox": [ + 54, + 408, + 295, + 553 + ], + "type": "text", + "content": " to balance error and efficiency. Our estimator achieves a low Mean Absolute Percentage Error (MAPE) of " + }, + { + "bbox": [ + 54, + 408, + 295, + 553 + ], + "type": "inline_equation", + "content": "1.02\\%" + }, + { + "bbox": [ + 54, + 408, + 295, + 553 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 408, + 295, + 553 + ], + "type": "inline_equation", + "content": "0.77\\%" + }, + { + "bbox": [ + 54, + 408, + 295, + 553 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 408, + 295, + 553 + ], + "type": "inline_equation", + "content": "1.93\\%" + }, + { + "bbox": [ + 54, + 408, + 295, + 553 + ], + "type": "text", + "content": " on S3DIS, ScanNet, and SemanticKITTI datasets, respectively, performing better on indoor datasets (S3DIS, ScanNet) due to their more consistent point density and distribution. Appendix B.6 and B.8 discuss ablation study for " + }, + { + "bbox": [ + 54, + 408, + 295, + 553 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 54, + 408, + 295, + 553 + ], + "type": "text", + "content": " and cross-dataset applicability of estimators." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 558, + 295, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 558, + 295, + 618 + ], + "spans": [ + { + "bbox": [ + 54, + 558, + 295, + 618 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 54, + 558, + 295, + 618 + ], + "type": "text", + "content": " Distance Curve Segmentation Ideally, what we want is to sample points whose minimum distance matches the value on the curve at each iteration. However, directly finding the point that yields the exact minimum distance in every iteration is computationally expensive." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 617, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 617, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 54, + 617, + 295, + 689 + ], + "type": "text", + "content": "To efficiently sample points along the estimated minimum distance curve, we first divide the curve into multiple segments. Note that each segment is associated with a group of consecutive iterations. Then, the corresponding predicted minimum distances are employed to identify sample point at the segment-level granularity in the sampling process." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 689, + 295, + 713 + ], + "type": "text", + "content": "As illustrated in Figure 4, the curve is segmented at iterations " + }, + { + "bbox": [ + 54, + 689, + 295, + 713 + ], + "type": "inline_equation", + "content": "d_{1}" + }, + { + "bbox": [ + 54, + 689, + 295, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 689, + 295, + 713 + ], + "type": "inline_equation", + "content": "d_{2}" + }, + { + "bbox": [ + 54, + 689, + 295, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 689, + 295, + 713 + ], + "type": "inline_equation", + "content": "d_{3}" + }, + { + "bbox": [ + 54, + 689, + 295, + 713 + ], + "type": "text", + "content": ", resulting in three segments. For" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 307, + 553, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 553, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 553, + 366 + ], + "type": "text", + "content": "each input point, we identify points within a specified radius of each segment boundary (i.e., " + }, + { + "bbox": [ + 313, + 307, + 553, + 366 + ], + "type": "inline_equation", + "content": "R_{1}" + }, + { + "bbox": [ + 313, + 307, + 553, + 366 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 307, + 553, + 366 + ], + "type": "inline_equation", + "content": "R_{2}" + }, + { + "bbox": [ + 313, + 307, + 553, + 366 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 307, + 553, + 366 + ], + "type": "inline_equation", + "content": "R_{3}" + }, + { + "bbox": [ + 313, + 307, + 553, + 366 + ], + "type": "text", + "content": ") and store them in the exclusion list for three segments. This list prevents sampling closely located points by considering the sampled points." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 369, + 554, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 554, + 476 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 554, + 476 + ], + "type": "text", + "content": "To optimize computational efficiency, building the exclusion list across all segments is fused into a single GPU kernel. By calculating the distance between the points only once, regardless of the number of segments, we minimize the latency overhead. For example, the distance between P0 and P1 is computed only once. If P1 is within " + }, + { + "bbox": [ + 313, + 369, + 554, + 476 + ], + "type": "inline_equation", + "content": "R_{3}" + }, + { + "bbox": [ + 313, + 369, + 554, + 476 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 313, + 369, + 554, + 476 + ], + "type": "inline_equation", + "content": "\\mathrm{dist(P0,P1)} < R_3 < R_2 < R_1" + }, + { + "bbox": [ + 313, + 369, + 554, + 476 + ], + "type": "text", + "content": "), P1 is added to the exclusion list of P0 at all segments. Refer to Table 3 for latency overhead with an increase in the segment count." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 477, + 554, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 477, + 554, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 477, + 554, + 550 + ], + "type": "text", + "content": "Building the exclusion list still requires computing pairwise distances; however, this process is much faster than the original FPS algorithm. This improvement is achieved because all computations are fully parallelizable, as sampling and distance calculations are completely decoupled. For details on the algorithm, refer to Appendix A.1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 556, + 554, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 554, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 554, + 700 + ], + "type": "text", + "content": "Sampling with Predicted Distance The sampling process employs the exclusion list. Starting from Segment 1, a seed point (P0 in Figure 4) is chosen. In each sampling iteration, points within a radius " + }, + { + "bbox": [ + 313, + 556, + 554, + 700 + ], + "type": "inline_equation", + "content": "R_{1}" + }, + { + "bbox": [ + 313, + 556, + 554, + 700 + ], + "type": "text", + "content": " from the most recently sampled point are filtered out based on the exclusion list. We employ bitmaps to record which points are available for sampling. For instance, in the first iteration of Figure 4, Point P0, P1, P4, P5, P6, P7, and P11 are excluded based on the exclusion list of Segment 1. Point P3 is randomly selected among the eight available points for the second iteration in the same segment. Excluded points remain unavailable until the current segment is complete." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "type": "text", + "content": "As the sampling progresses and the segment changes" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25118" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 298, + 240 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 298, + 240 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 298, + 240 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 298, + 240 + ], + "type": "image", + "image_path": "33c964742bb34613ffa0c942e141c996bba899b16190410740caa8d5d6692ef1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 246, + 297, + 270 + ], + "lines": [ + { + "bbox": [ + 55, + 246, + 297, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 246, + 297, + 270 + ], + "type": "text", + "content": "Figure 5. Minimum spacing distribution comparison with increasing segment counts." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 283, + 296, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 283, + 296, + 462 + ], + "spans": [ + { + "bbox": [ + 54, + 283, + 296, + 462 + ], + "type": "text", + "content": "(e.g., from Segment 1 to Segment 2), the radius threshold is adjusted (e.g., from " + }, + { + "bbox": [ + 54, + 283, + 296, + 462 + ], + "type": "inline_equation", + "content": "R_{1}" + }, + { + "bbox": [ + 54, + 283, + 296, + 462 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 54, + 283, + 296, + 462 + ], + "type": "inline_equation", + "content": "R_{2}" + }, + { + "bbox": [ + 54, + 283, + 296, + 462 + ], + "type": "text", + "content": "). This transition changes the set of points that are available for sampling. By iteratively applying this process, we ensure that the distance between sampled points meets or exceeds the threshold for each segment. This results in a minimum distance that lies above the blue step function in the segmented distance curve (i.e., " + }, + { + "bbox": [ + 54, + 283, + 296, + 462 + ], + "type": "inline_equation", + "content": "R_{i}" + }, + { + "bbox": [ + 54, + 283, + 296, + 462 + ], + "type": "text", + "content": " serves as the lower bound of the minimum distance in Segment " + }, + { + "bbox": [ + 54, + 283, + 296, + 462 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 54, + 283, + 296, + 462 + ], + "type": "text", + "content": "). Increasing the number of segments allows the step function to approach the red minimum distance curve, improving sampling quality and converging towards the quality of FPS. Figure 5 demonstrates this convergence. With more segments, the distribution of minimum distance between sampled points becomes closer to that of FPS. Details on implementation are provided in Appendix A.1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 468, + 296, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 296, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 296, + 672 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 55, + 468, + 296, + 672 + ], + "type": "text", + "content": " Early Termination While the estimated minimum distance curve exhibits low error, overestimation can still negatively impact sampling quality. This is because it can lead to an excessive number of point exclusions, limiting the availability of points in later iterations. To mitigate this issue, we introduce Early Termination. This technique monitors the availability of points at each iteration. If no points remain, it advances to the next segment earlier than planned. When the final segment is exhausted, Sampling with Predicted Distance is terminated, and the remaining iterations are completed using the original FPS algorithm. For seamless transition to FPS, we efficiently update the minimum distance matrix (dists in Algorithm 1) by leveraging the exclusion list. This approach avoids recalculating all-to-all distances between points, significantly reducing the search space and minimizing the impact on sampling time. Appendix A.1 describes the details of Early Termination." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 677, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 700 + ], + "type": "text", + "content": "Factors Contributing to Speedup The primary factors contributing to the latency of MDPS are:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "content": "1. Minimum Distance Curve Estimation: This step, requir" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 554, + 225 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 554, + 225 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 554, + 225 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 554, + 225 + ], + "type": "image", + "image_path": "b6010c9a380db52d887551a50b09d33c6045f13ae9f5532d0eaa12e43f001d09.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 233, + 517, + 245 + ], + "lines": [ + { + "bbox": [ + 351, + 233, + 517, + 245 + ], + "spans": [ + { + "bbox": [ + 351, + 233, + 517, + 245 + ], + "type": "text", + "content": "Figure 6. Redundancy Free Neighbor Search" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 326, + 258, + 553, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 258, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 326, + 258, + 553, + 282 + ], + "type": "text", + "content": "ing 1/10 of the original FPS iterations, adds only a fraction of the original FPS latency." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 282, + 555, + 402 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 314, + 282, + 555, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 282, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 314, + 282, + 555, + 329 + ], + "type": "text", + "content": "2. Exclusion List Construction: The required all-to-all distance calculations are fully parallelizable, resulting in minimal latency overhead. Detailed analysis on this factor is provided in Appendix A.2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 330, + 554, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 330, + 554, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 554, + 402 + ], + "type": "text", + "content": "3. Sampling Process: The use of exclusion list eliminates the need for additional distance calculations. Early termination requires additional FPS iterations. However, the impact is minimal: the overhead of " + }, + { + "bbox": [ + 313, + 330, + 554, + 402 + ], + "type": "inline_equation", + "content": "2.02\\%" + }, + { + "bbox": [ + 313, + 330, + 554, + 402 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 330, + 554, + 402 + ], + "type": "inline_equation", + "content": "0.58\\%" + }, + { + "bbox": [ + 313, + 330, + 554, + 402 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 330, + 554, + 402 + ], + "type": "inline_equation", + "content": "1.18\\%" + }, + { + "bbox": [ + 313, + 330, + 554, + 402 + ], + "type": "text", + "content": " of original FPS iterations in S3DIS, ScanNet, and SemanticKITTI dataset, respectively." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 402, + 554, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 554, + 437 + ], + "type": "text", + "content": "Due to these optimizations, MDPS can achieve a " + }, + { + "bbox": [ + 313, + 402, + 554, + 437 + ], + "type": "inline_equation", + "content": "4 - 5 \\times" + }, + { + "bbox": [ + 313, + 402, + 554, + 437 + ], + "type": "text", + "content": " speedup relative to the original FPS algorithm. Detailed latency breakdown of MDPS is provided in Appendix B.2." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 445, + 503, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 503, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 503, + 458 + ], + "type": "text", + "content": "4.2. Redundancy-Free Neighbor Search" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 463, + 554, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 554, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 554, + 523 + ], + "type": "text", + "content": "The exclusion list in MDPS captures spatial relationships between points, providing an opportunity to optimize subsequent Ball Query and k-NN operations. We introduce Redundancy Free Ball Query and k-NN, which fully leverage the exclusion list from the sampling stage." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 528, + 555, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 555, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 555, + 673 + ], + "type": "text", + "content": "Redundancy-Free Ball Query Ball query, which identifies neighbors within a radius " + }, + { + "bbox": [ + 313, + 528, + 555, + 673 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 528, + 555, + 673 + ], + "type": "text", + "content": ", is performed after sampling, before aggregation (refer to Figure 1). The exclusion list is well-suited for this task, as it contains points within a specified radius of each point. By adding an extra list to the exclusion list for the radius " + }, + { + "bbox": [ + 313, + 528, + 555, + 673 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 528, + 555, + 673 + ], + "type": "text", + "content": ", as illustrated in Figure 6, we can directly use it for ball query. As discussed in Section 4.1, this approach minimizes computational overhead; distances between points are calculated only once, regardless of the number of segments. After sampling, we extract the relevant entries from the corresponding exclusion list for the ball query, which serve as centroids for aggregation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "content": "Redundancy-Free k-NN While directly reusing the exclusion list for k-NN is challenging, we can leverage it to reduce the search space. By using the exclusion list of Seg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25119" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 288, + 161 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 288, + 161 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 288, + 161 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 288, + 161 + ], + "type": "table", + "html": "
DatasetAverage Minimum Distance (Ratio to Baseline)
Baseline (FPS)Random SamplingGrid SamplingMDPS (Ours)
S3DIS0.063950.03478 (54.38%)0.04971 (77.72%)0.06359 (99.45%)
ScanNet0.037010.02035 (54.99%)0.02400 (64.84%)0.03677 (99.35%)
SemanticKITTI0.255050.12775 (50.09%)0.21519 (84.37%)0.25087 (98.36%)
", + "image_path": "e0bc3bb64e6e077a2a4a4f3e867a8a159b5041e378f103e14ee67c98ccea8bd0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 198, + 296, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 198, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 296, + 319 + ], + "type": "text", + "content": "ment 1, we can limit the search space by eliminating points that are too far away to be potential k-NN neighbors. For example, to find the three nearest neighbors of point P7 in Figure 6, we calculate distances between P7 and the points from the corresponding exclusion list to identify the three nearest neighbors. Because k-NN upsampling relies on the downsampled set, points that were not downsampled are excluded from the search space. This approach significantly reduces the number of distance calculations, leading to substantial latency reduction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 328, + 127, + 341 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 328, + 127, + 341 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 127, + 341 + ], + "type": "text", + "content": "5. Evaluation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 348, + 140, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 140, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 140, + 361 + ], + "type": "text", + "content": "5.1. Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 365, + 295, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 437 + ], + "type": "text", + "content": "We evaluate FastPoint on PointMetaBase [17] and PointVector [6], two recent models based on PointNet++. We evaluate our proposal on both indoor and outdoor datasets to demonstrate its scalability. We use S3DIS [1] and ScanNet [5] for indoor datasets and use SemanticKITTI [2] for outdoor datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 437, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 533 + ], + "type": "text", + "content": "The end-to-end latency is measured for processing all scenes in each dataset's validation split. We use mean Intersection over Union (mIoU) as a model accuracy evaluation metric. Detailed evaluation methodologies follow the same approach in PointMetaBase [17], and PointVector [6]. We apply FastPoint and the other baseline methods only to the first layer of each model as the first layer's FPS and neighbor search dominates the total computation time (" + }, + { + "bbox": [ + 55, + 437, + 295, + 533 + ], + "type": "inline_equation", + "content": ">90\\%" + }, + { + "bbox": [ + 55, + 437, + 295, + 533 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 533, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 617 + ], + "type": "text", + "content": "We implement FastPoint with custom CUDA kernels and integrate them into OpenPoints library. We also integrate the software implementation [11] of QuickFPS [12] opensourced by the authors to OpenPoints for comparison. The experiments are conducted using an NVIDIA Geforce RTX 3090 GPU with 24GB of memory. We use CUDA 11.8 and PyTorch 1.10.1 for software setup." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 624, + 206, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 206, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 206, + 637 + ], + "type": "text", + "content": "5.2. Sampling Quality of MDPS" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 642, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 715 + ], + "type": "text", + "content": "To evaluate sampling quality, we compute the average minimum distance between neighboring sampled points. A larger average distance indicates better coverage of sample points from the perspectives of point distribution and structure preservation. As shown in Table 1, MDPS achieves sampling quality comparable to FPS, with an average min" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 548, + 218 + ], + "blocks": [ + { + "bbox": [ + 55, + 166, + 295, + 190 + ], + "lines": [ + { + "bbox": [ + 55, + 166, + 295, + 190 + ], + "spans": [ + { + "bbox": [ + 55, + 166, + 295, + 190 + ], + "type": "text", + "content": "Table 1. Comparison of Sampling Quality among FPS, Random Sampling, Grid Sampling and MDPS." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 70, + 548, + 218 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 548, + 218 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 548, + 218 + ], + "type": "table", + "html": "
ModelDatasetmIoU (Diff. to Baseline FPS)
Baseline (FPS)Random SamplingGrid SamplingMDPS (Ours)
PV-LS3DIS71.3364.83 (-6.5)70.97 (-0.36)71.37 (+0.04)
ScanNet70.7059.94 (-10.76)69.61 (-1.09)70.63 (-0.07)
Semantic KITTI50.9138.46 (-12.45)50.62 (-0.29)50.79 (-0.12)
PMB-LS3DIS69.7268.21 (-1.51)69.54 (-0.18)69.74 (+0.02)
ScanNet70.8667.66 (-3.2)70.36 (-0.50)70.89 (+0.03)
Semantic KITTI52.1947.26 (-4.93)52.27 (+0.08)52.09 (-0.10)
", + "image_path": "64f00d78e8eb052365a71882b47fdbeff936ad04d3aab4fa9668fe40f99b571b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 222, + 553, + 244 + ], + "lines": [ + { + "bbox": [ + 313, + 222, + 553, + 244 + ], + "spans": [ + { + "bbox": [ + 313, + 222, + 553, + 244 + ], + "type": "text", + "content": "Table 2. Accuracy comparison. PV and PMB stand for PointVector and PointMetaBase." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 255, + 555, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 555, + 352 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 555, + 352 + ], + "type": "text", + "content": "imum distance reaching over " + }, + { + "bbox": [ + 313, + 255, + 555, + 352 + ], + "type": "inline_equation", + "content": "99\\%" + }, + { + "bbox": [ + 313, + 255, + 555, + 352 + ], + "type": "text", + "content": " of the FPS baseline for S3DIS and ScanNet, and over " + }, + { + "bbox": [ + 313, + 255, + 555, + 352 + ], + "type": "inline_equation", + "content": "98\\%" + }, + { + "bbox": [ + 313, + 255, + 555, + 352 + ], + "type": "text", + "content": " for SemanticKITTI. In contrast, Random Sampling and Grid Sampling [29] exhibit significantly lower quality, with their average minimum distances deviating substantially from the FPS baseline. We also compare the minimum spacing distribution of each sampling method via visualization in Appendix B.9 and evaluate the robustness of MDPS in Appendix B.7." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 358, + 463, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 463, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 463, + 372 + ], + "type": "text", + "content": "5.3. Accuracy Impact of MDPS" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 376, + 555, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 555, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 555, + 484 + ], + "type": "text", + "content": "We analyze the impact of MDPS on accuracy by applying it to the inference stage of PointNet++ models initially trained with FPS. Since MDPS is designed to closely replicate the FPS sampling pattern, it is compatible with FPS-trained models, thereby minimizing accuracy degradation. Note that Redundancy-Free Neighbor Search does not introduce any approximation, thus has no impact on accuracy. To compare the effects of different sampling methods, we also evaluate random sampling and grid sampling1." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 484, + 554, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 554, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 554, + 616 + ], + "type": "text", + "content": "Table 2 shows that MDPS excels in maintaining accuracy, with negligible differences observed across all datasets and models. This highlights its effectiveness in replicating FPS behavior. However, the grid sampling exhibits larger and less consistent accuracy drops. When applied only during inference, it leads to declines of up to " + }, + { + "bbox": [ + 313, + 484, + 554, + 616 + ], + "type": "inline_equation", + "content": "1.09\\%" + }, + { + "bbox": [ + 313, + 484, + 554, + 616 + ], + "type": "text", + "content": " (PointVector-L, ScanNet). While it outperforms FastPoint in a single instance (PointMetaBase-L, SemanticKITTI), its performance is generally inferior, with significant accuracy reductions in other scenarios. The random sampling shows considerable accuracy drops across all data points." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 616, + 554, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 554, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 554, + 687 + ], + "type": "text", + "content": "Overall, FastPoint effectively preserves accuracy by closely approximating the sampling pattern of FPS, achieving accuracy levels near-identical to FPS without retraining. Latency-accuracy comparisons (Appendix B.5) show that FastPoint resides above the pareto front, emphasizing the effectiveness of FastPoint." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "type": "text", + "content": "1Grid sampling is applied only during inference. Appendix B.1 shows results for cases where both training and inference utilize grid sampling." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25120" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 70, + 281, + 180 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 281, + 180 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 281, + 180 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 281, + 180 + ], + "type": "table", + "html": "
MethodNumber of SegmentsEvaluation Metrics
Accuracy (mIoU)Sampling Time (ms)Sampling Quality
FPS-70.7085.830.03701
MDPS (Ours)1 Segment70.3418.620.03532
2 Segments70.6019.430.03598
3 Segments70.4619.790.03636
4 Segments70.5421.010.03662
5 Segments70.5221.410.03665
6 Segments70.6322.060.03662
7 Segments70.6322.750.03664
", + "image_path": "57815b22f7209fe1dbda940f928fc4136274f42c75d81735e2b475863bf9448a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 239, + 293, + 373 + ], + "blocks": [ + { + "bbox": [ + 56, + 239, + 293, + 373 + ], + "lines": [ + { + "bbox": [ + 56, + 239, + 293, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 293, + 373 + ], + "type": "image", + "image_path": "8a4407dc43eaa96414579591ba55d0eb1d1540328aa6b50592e9a3a806c6e316.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 381, + 282, + 392 + ], + "lines": [ + { + "bbox": [ + 69, + 381, + 282, + 392 + ], + "spans": [ + { + "bbox": [ + 69, + 381, + 282, + 392 + ], + "type": "text", + "content": "Figure 7. End-to-end speedup of FastPoint and QuickFPS." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 407, + 227, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 227, + 419 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 227, + 419 + ], + "type": "text", + "content": "5.4. Latency Reduction of FastPoint" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 426, + 296, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 533 + ], + "type": "text", + "content": "To evaluate the impact of FastPoint on end-to-end latency, we compare its latency to baseline FPS and QuickFPS. Note that FastPoint includes both MDPS and Redundancy-Free Neighbor Search. As explained in Section 2.2, QuickFPS accelerates FPS using k-d tree without approximation. Since MDPS also contains FPS steps when predicting the minimum distance curve, we can achieve further latency reductions by replacing them with QuickFPS, meaning that FastPoint and QuickFPS can be synergetic." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 533, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 641 + ], + "type": "text", + "content": "Figure 7 presents the results of this comparison. Across S3DIS, ScanNet, and SemanticKITTI datasets, FastPoint achieves substantial latency reductions, with a geomean end-to-end speedup of " + }, + { + "bbox": [ + 55, + 533, + 295, + 641 + ], + "type": "inline_equation", + "content": "2.55 \\times" + }, + { + "bbox": [ + 55, + 533, + 295, + 641 + ], + "type": "text", + "content": " compared to baseline FPS. Integrating QuickFPS with FastPoint further reduces the execution latency achieving a geomean end-to-end speedup of " + }, + { + "bbox": [ + 55, + 533, + 295, + 641 + ], + "type": "inline_equation", + "content": "2.76 \\times" + }, + { + "bbox": [ + 55, + 533, + 295, + 641 + ], + "type": "text", + "content": " compared to baseline FPS and " + }, + { + "bbox": [ + 55, + 533, + 295, + 641 + ], + "type": "inline_equation", + "content": "1.41 \\times" + }, + { + "bbox": [ + 55, + 533, + 295, + 641 + ], + "type": "text", + "content": " compared to standalone QuickFPS. Speedup specific to the sampling and neighbor search are provided in Appendix B.3 and B.4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "These findings substantiate the effectiveness of Fast-Point in reducing latency, surpassing existing methods like QuickFPS. Moreover, the successful combination of Fast-Point with QuickFPS demonstrates its compatibility with existing FPS optimizations. Scalability to Non-PointNet++ based models are discussed in Appendix B.10." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 315, + 69, + 439, + 178 + ], + "blocks": [ + { + "bbox": [ + 55, + 188, + 295, + 233 + ], + "lines": [ + { + "bbox": [ + 55, + 188, + 295, + 233 + ], + "spans": [ + { + "bbox": [ + 55, + 188, + 295, + 233 + ], + "type": "text", + "content": "Table 3. Comparison of mIoU, sampling time, and sampling quality for FPS and MDPS across segments 1 to 7. Sampling time is measured with a single scene of ScanNet dataset. PointVector-L model is used for mIoU comparison." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 69, + 439, + 178 + ], + "lines": [ + { + "bbox": [ + 315, + 69, + 439, + 178 + ], + "spans": [ + { + "bbox": [ + 315, + 69, + 439, + 178 + ], + "type": "table", + "html": "
ModelDatasetMethodSpeedup
PV-LS3DISMDPS1.90×
All1.98×
ScanNetMDPS2.33×
All2.52×
Semantic KITTIMDPS2.07×
All2.20×
PMB-LS3DISMDPS2.50×
All2.82×
ScanNetMDPS2.89×
All3.38×
Semantic KITTIMDPS2.31×
All2.62×
", + "image_path": "ba498ebc9c248fbc0d55b074d03ba960e8c375fbec16f2000683e9a26e2bbcaf.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 314, + 180, + 438, + 200 + ], + "lines": [ + { + "bbox": [ + 314, + 180, + 438, + 200 + ], + "spans": [ + { + "bbox": [ + 314, + 180, + 438, + 200 + ], + "type": "text", + "content": "(a) Comparison of component-wise end-to-end speedup of FastPoint." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 439, + 69, + 548, + 178 + ], + "blocks": [ + { + "bbox": [ + 439, + 69, + 548, + 178 + ], + "lines": [ + { + "bbox": [ + 439, + 69, + 548, + 178 + ], + "spans": [ + { + "bbox": [ + 439, + 69, + 548, + 178 + ], + "type": "table", + "html": "
Model# of PointsSpeedup
PV-L16,0001.33×
32,0002.41×
48,0002.68×
64,0003.01×
80,0003.25×
96,0003.40×
PMB-L16,0001.58×
32,0002.64×
48,0002.91×
64,0003.06×
80,0003.30×
96,0003.44×
", + "image_path": "f021a8229ed5295838f67bbc80a2b5332176aa25625a0787552ac22c2c323c8a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 439, + 180, + 553, + 200 + ], + "lines": [ + { + "bbox": [ + 439, + 180, + 553, + 200 + ], + "spans": [ + { + "bbox": [ + 439, + 180, + 553, + 200 + ], + "type": "text", + "content": "(b) Comparison of speedup as the number of points increases." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 10 + }, + { + "bbox": [ + 318, + 203, + 548, + 214 + ], + "lines": [ + { + "bbox": [ + 318, + 203, + 548, + 214 + ], + "spans": [ + { + "bbox": [ + 318, + 203, + 548, + 214 + ], + "type": "text", + "content": "Table 4. Ablation study for speedup breakdown and scalability." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 314, + 229, + 408, + 242 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 229, + 408, + 242 + ], + "spans": [ + { + "bbox": [ + 314, + 229, + 408, + 242 + ], + "type": "text", + "content": "5.5. Ablation Study" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 246, + 553, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 246, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 553, + 354 + ], + "type": "text", + "content": "Segment Count We evaluate the impact of varying the segment count on accuracy, sampling time, and sampling quality (i.e., average minimum distance). Table 3 shows that while sampling time increases gradually with more segments due to the exclusion list construction (refer to Section 4.1), both sampling quality and accuracy improve. However, they get saturated beyond a certain number of segments. Based on the empirical analysis, we use 6 segments by default, where accuracy saturates for all workloads." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 357, + 554, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 554, + 453 + ], + "type": "text", + "content": "Component-wise Speedup To understand the individual contributions of each component in FastPoint, we evaluate the end-to-end speedup by sequentially applying MDPS and Redundancy Free Neighbor Search. Table 4a shows that MDPS alone provides significant speedup. Incorporating Redundancy Free Neighbor Search further provides additional improvement, demonstrating that each component makes a distinct contribution to the overall speedup." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 460, + 416, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 416, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 416, + 474 + ], + "type": "text", + "content": "5.6. Scalability Study" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 478, + 554, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 478, + 554, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 554, + 562 + ], + "type": "text", + "content": "To assess the scalability of FastPoint, we evaluate the end-to-end speedup compared to FPS by varying point cloud sizes. Table 4b shows that FastPoint consistently achieves higher speedups as the number of points increases. This trend highlights the efficiency of our approach, particularly when processing larger point clouds where the computational cost of the FPS becomes a significant bottleneck." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 574, + 388, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 388, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 388, + 586 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "3D point clouds are critical for representing and understanding 3D scenes, and deep learning models like PointNet++ have shown great promise in processing this data. However, computational efficiency remains a key challenge. We propose FastPoint, a software acceleration technique that leverages predictable minimum distances between sampled points to efficiently identify subsequent points without exhaustive computations. Our evaluation shows a " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "2.55 \\times" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " speedup over baseline FPS while maintaining accuracy, enabling low-cost deployment of PointNet++ based models." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25121" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 159, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 159, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 159, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 260 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 260 + ], + "type": "text", + "content": "This work was supported by the Ministry of Science and ICT (MSIT), Korea, under the following grants: the Global Scholars Invitation Program (RS-2024-00456287), the Graduate School of Artificial Intelligence Semiconductor grant (IITP-2025-RS-2023-00256081), and the Artificial Intelligence Innovation Hub (No. 2021-0-02068), all supervised by the Institute for Information & Communications Technology Planning & Evaluation (IITP). Additional support was provided by the National Research Foundation of Korea (NRF) grant (RS-2024-00405857) funded by the Korea government (MSIT), and the Google Faculty Research Award. The source code is available at https://github.com/SNU-ARC/FastPoint.git. Jae W. Lee and Hongil Yoon are the corresponding authors." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 272, + 115, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 115, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 115, + 285 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 292, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 292, + 296, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 292, + 296, + 325 + ], + "spans": [ + { + "bbox": [ + 61, + 292, + 296, + 325 + ], + "type": "text", + "content": "[1] Iro Armeni, Ozan Sener, Amir R. Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In CVPR, 2016. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 327, + 296, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 327, + 296, + 370 + ], + "spans": [ + { + "bbox": [ + 61, + 327, + 296, + 370 + ], + "type": "text", + "content": "[2] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Juergen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In ICCV, 2019. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 372, + 295, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 372, + 295, + 405 + ], + "spans": [ + { + "bbox": [ + 62, + 372, + 295, + 405 + ], + "type": "text", + "content": "[3] Yukang Chen, Jianhui Liu, Xiangyu Zhang, Xiaojuan Qi, and Jiaya Jia. Largekernel3d: Scaling up kernels in 3d sparse cnns. In CVPR, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 407, + 295, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 407, + 295, + 440 + ], + "spans": [ + { + "bbox": [ + 62, + 407, + 295, + 440 + ], + "type": "text", + "content": "[4] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, 2019. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 441, + 296, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 441, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 62, + 441, + 296, + 485 + ], + "type": "text", + "content": "[5] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 487, + 296, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 487, + 296, + 519 + ], + "spans": [ + { + "bbox": [ + 62, + 487, + 296, + 519 + ], + "type": "text", + "content": "[6] Xin Deng, WenYu Zhang, Qing Ding, and XinMing Zhang. Pointvector: A vector representation in point cloud analysis. In CVPR, 2023. 1, 2, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 521, + 296, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 521, + 296, + 565 + ], + "spans": [ + { + "bbox": [ + 62, + 521, + 296, + 565 + ], + "type": "text", + "content": "[7] Lue Fan, Ziqi Pang, Tianyuan Zhang, Yu-Xiong Wang, Hang Zhao, Feng Wang, Naiyan Wang, and Zhaoxiang Zhang. Embracing Single Stride 3D Object Detector with Sparse Transformer. In CVPR, 2022. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 567, + 296, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 567, + 296, + 621 + ], + "spans": [ + { + "bbox": [ + 62, + 567, + 296, + 621 + ], + "type": "text", + "content": "[8] Yu Feng, Boyuan Tian, Tiancheng Xu, Paul Whatmough, and Yuhao Zhu. Mesorasi: Architecture support for point cloud analytics via delayed-aggregation. In Proceedings of the 53th International Symposium on Microarchitecture (MICRO), 2020. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 624, + 296, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 624, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 62, + 624, + 296, + 678 + ], + "type": "text", + "content": "[9] Yu Feng, Gunnar Hammonds, Yiming Gan, and Yuhao Zhu. Crescent: Taming memory irregularities for accelerating deep point cloud analytics. In Proceedings of the 49th Annual International Symposium on Computer Architecture (ISCA), 2022. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 680, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 296, + 712 + ], + "type": "text", + "content": "[10] Lei Han, Tian Zheng, Lan Xu, and Lu Fang. Occuseg: Occupancy-aware 3d instance segmentation. In CVPR, 2020. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 117 + ], + "type": "text", + "content": "[11] Meng Han, Liang Wang, Limin Xiao, Hao Zhang, Chenhao Zhang, Xiangrong Xu, and Jianfeng Zhu. Quickfps. http://github.com/hanm2019/bucket-based_farthest-point-sampling_GPU.2,7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 119, + 554, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 554, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 554, + 184 + ], + "type": "text", + "content": "[12] Meng Han, Liang Wang, Limin Xiao, Hao Zhang, Chenhao Zhang, Xiangrong Xu, and Jianfeng Zhu. Quickfps: Architecture and algorithm co-design for farthest point sampling in large-scale point clouds. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 2023. 1, 2, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 186, + 554, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 554, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 554, + 230 + ], + "type": "text", + "content": "[13] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. In CVPR, 2020. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 233, + 554, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 233, + 554, + 266 + ], + "spans": [ + { + "bbox": [ + 316, + 233, + 554, + 266 + ], + "type": "text", + "content": "[14] Donghyun Lee, Yejin Lee, Jae W. Lee, and Hongil Yoon. Frugal 3d point cloud model training via progressive near point filtering and fused aggregation. In ECCV, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 268, + 554, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 268, + 554, + 312 + ], + "spans": [ + { + "bbox": [ + 316, + 268, + 554, + 312 + ], + "type": "text", + "content": "[15] Jingtao Li, Jian Zhou, Yan Xiong, Xing Chen, and Chaitali Chakrabarti. An adjustable farthest point sampling method for approximately-sorted point cloud data. In 2022 IEEE Workshop on Signal Processing Systems (SiPS), 2022. 1, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 314, + 554, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 314, + 554, + 368 + ], + "spans": [ + { + "bbox": [ + 316, + 314, + 554, + 368 + ], + "type": "text", + "content": "[16] Yaoxiu Lian, Xinhao Yang, Ke Hong, Yu Wang, Guohao Dai, and Ningyi Xu. A point transformer accelerator with fine-grained pipelines and distribution-aware dynamic fps. In IEEE/ACM International Conference on Computer-Aided Design (ICCAD), 2023. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 370, + 554, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 370, + 554, + 412 + ], + "spans": [ + { + "bbox": [ + 316, + 370, + 554, + 412 + ], + "type": "text", + "content": "[17] Haojia Lin, Xiawu Zheng, Lijiang Li, Fei Chao, Shanshan Wang, Yan Wang, Yonghong Tian, and Rongrong Ji. Meta architecture for point cloud analysis. In CVPR, 2023. 1, 2, 3, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 416, + 554, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 554, + 471 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 554, + 471 + ], + "type": "text", + "content": "[18] Xueyuan Liu, Zhuoran Song, Guohao Dai, Gang Li, Can Xiao, Yan Xiang, Dehui Kong, Ke Xu, and Xiaoyao Liang. Fusionarch: A fusion-based accelerator for point-based point cloud neural networks. In ACM/IEEE Design, Automation and Test in Europe Conference (DATE), 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 473, + 554, + 516 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 473, + 554, + 516 + ], + "spans": [ + { + "bbox": [ + 316, + 473, + 554, + 516 + ], + "type": "text", + "content": "[19] Yongcheng Liu, Bin Fan, Gaofeng Meng, Jiwen Lu, Shiming Xiang, and Chunhong Pan. Densepoint: Learning densely contextual representation for efficient point cloud processing. In ICCV, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 518, + 554, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 518, + 554, + 551 + ], + "spans": [ + { + "bbox": [ + 315, + 518, + 554, + 551 + ], + "type": "text", + "content": "[20] Zhijian Liu, Xinyu Yang, Haotian Tang, Shang Yang, and Song Han. Flatformer: Flattened window attention for efficient point cloud transformer. In CVPR, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 553, + 554, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 553, + 554, + 586 + ], + "spans": [ + { + "bbox": [ + 315, + 553, + 554, + 586 + ], + "type": "text", + "content": "[21] Jiageng Mao, Yujing Xue, Minzhe Niu, Haoyue Bai, Jiashi Feng, Xiaodan Liang, Hang Xu, and Chunjing Xu. Voxel transformer for 3d object detection. In ICCV, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 589, + 554, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 589, + 554, + 632 + ], + "spans": [ + { + "bbox": [ + 316, + 589, + 554, + 632 + ], + "type": "text", + "content": "[22] Alexey Nekrasov, Jonas Schult, Or Litany, Bastian Leibe, and Francis Engelmann. Mix3D: Out-of-Context Data Augmentation for 3D Scenes. In International Conference on 3D Vision (3DV), 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 635, + 554, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 554, + 677 + ], + "type": "text", + "content": "[23] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. arXiv preprint arXiv:1612.00593, 2016. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 680, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 554, + 713 + ], + "type": "text", + "content": "[24] Charles R Qi, Li Yi, Hao Su, and Leonidas J Guibas. Point-net++: Deep hierarchical feature learning on point sets in a metric space. arXiv preprint arXiv:1706.02413, 2017. 1, 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25122" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[25] Charles R. Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas J. Guibas. Frustum pointnets for 3d object detection from rgb-d data. In CVPR, 2018." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 106, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 106, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 55, + 106, + 295, + 150 + ], + "type": "text", + "content": "[26] Guocheng Qian, Hasan Hammoud, Guohao Li, Ali Thabet, and Bernard Ghanem. Assanet: An anisotropical separable set abstraction for efficient point cloud representation learning. In NeurIPS, 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 295, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 295, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 295, + 195 + ], + "type": "text", + "content": "[27] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 195, + 296, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 296, + 250 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 296, + 250 + ], + "type": "text", + "content": "[28] Danila Rukhovich, Anna Vorontsova, and Anton Konushin. Imvoxelnet: Image to voxels projection for monocular and multi-view general-purpose 3d object detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 250, + 295, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 295, + 294 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 295, + 294 + ], + "type": "text", + "content": "[29] Hugues Thomas, Charles R. Qi, Jean-Emmanuel Deschaud, Beatriz Marcotegui, François Goulette, and Leonidas J. Guibas. Kpconv: Flexible and deformable convolution for point clouds. In ICCV, 2019. 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 294, + 295, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 295, + 338 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 295, + 338 + ], + "type": "text", + "content": "[30] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E. Sarma, Michael M. Bronstein, and Justin M. Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (TOG), 2019. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 338, + 295, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 295, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 295, + 370 + ], + "type": "text", + "content": "[31] Wenxuan Wu, Zhongang Qi, and Li Fuxin. Pointconv: Deep convolutional networks on 3d point clouds. In CVPR, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 371, + 295, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 295, + 405 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 295, + 405 + ], + "type": "text", + "content": "[32] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. In NeurIPS, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 405, + 295, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 405, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 56, + 405, + 295, + 449 + ], + "type": "text", + "content": "[33] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler, faster, stronger. In CVPR, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 449, + 295, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 295, + 482 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 295, + 482 + ], + "type": "text", + "content": "[34] Mutian Xu, Runyu Ding, Hengshuang Zhao, and Xiaojuan Qi. Paconv: Position adaptive convolution with dynamic kernel assembling on point clouds. In CVPR, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 482, + 295, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 482, + 295, + 515 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 295, + 515 + ], + "type": "text", + "content": "[35] Qiangeng Xu, Xudong Sun, Cho-Ying Wu, Panqu Wang, and Ulrich Neumann. Grid-gen for fast and scalable point cloud learning, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 515, + 295, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 515, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 56, + 515, + 295, + 569 + ], + "type": "text", + "content": "[36] Xinhao Yang, Tianyu Fu, Guohao Dai, Shulin Zeng, Kai Zhong, Ke Hong, and Yu Wang. An efficient accelerator for point-based and voxel-based point cloud neural networks. In ACM/IEEE Design Automation Conference (DAC), 2023. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 570, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 295, + 624 + ], + "type": "text", + "content": "[37] Yu-Qi Yang, Yu-Xiao Guo, Jian-Yu Xiong, Yang Liu, Hao Pan, Peng-Shuai Wang, Xin Tong, and Baining Guo. Swin3d: A pretrained transformer backbone for 3d indoor scene understanding. arXiv preprint arXiv:2304.06906, 2023.1,2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 625, + 295, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 295, + 658 + ], + "type": "text", + "content": "[38] Yu-Qi Yang, Yu-Xiao Guo, and Yang Liu. Swin3d++: Effective multi-source pretraining for 3d indoor scene understanding. arXiv preprint arXiv:2304.06906, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "type": "text", + "content": "[39] Ziyu Ying, Sandeepa Bhuyan, Yan Kang, Yingtian Zhang, Mahmut T. Kandemir, and Chita R. Das. Edgepc: Efficient deep learning analytics for point clouds on edge devices. In Proceedings of the 50th Annual International Symposium on Computer Architecture (ISCA), 2023. 1, 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 554, + 219 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 315, + 72, + 554, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 554, + 118 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 554, + 118 + ], + "type": "text", + "content": "[40] Hyunsung Yoon and Jae-Joon Kim. Efficient sampling and grouping acceleration for point cloud deep learning via single coordinate comparison. In IEEE/ACM International Conference on Computer-Aided Design (ICCAD), 2023. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 118, + 554, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 118, + 554, + 152 + ], + "spans": [ + { + "bbox": [ + 315, + 118, + 554, + 152 + ], + "type": "text", + "content": "[41] Hengshuang Zhao, Li Jiang, Chi-Wing Fu, and Jiaya Jia. PointWeb: Enhancing local neighborhood features for point cloud processing. In CVPR, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 152, + 554, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 152, + 554, + 175 + ], + "spans": [ + { + "bbox": [ + 315, + 152, + 554, + 175 + ], + "type": "text", + "content": "[42] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip Torr, and Vladlen Koltun. Point transformer. In ICCV, 2021. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 175, + 554, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 175, + 554, + 219 + ], + "spans": [ + { + "bbox": [ + 315, + 175, + 554, + 219 + ], + "type": "text", + "content": "[43] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. arXiv preprint arXiv:2011.10033, 2020. 1, 2" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25123" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/FastVAR_ Linear Visual Autoregressive Modeling via Cached Token Pruning/70725db1-fcd5-4496-a95d-07038d174933_content_list.json b/2025/FastVAR_ Linear Visual Autoregressive Modeling via Cached Token Pruning/70725db1-fcd5-4496-a95d-07038d174933_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bf91287954cee2dfd0dbbfd048bb89cbcd2b8419 --- /dev/null +++ b/2025/FastVAR_ Linear Visual Autoregressive Modeling via Cached Token Pruning/70725db1-fcd5-4496-a95d-07038d174933_content_list.json @@ -0,0 +1,1479 @@ +[ + { + "type": "text", + "text": "FastVAR: Linear Visual Autoregressive Modeling via Cached Token Pruning", + "text_level": 1, + "bbox": [ + 112, + 130, + 883, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hang Guo $^{1}$ Yawei Li $^{2,*,†}$ Taolin Zhang $^{1}$ Jiangshan Wang $^{1}$", + "bbox": [ + 261, + 179, + 730, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tao Dai $^{3,*}$ Shu-Tao Xia $^{1,4}$ Luca Benini $^{2}$", + "bbox": [ + 334, + 198, + 661, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Tsinghua University $^{2}$ ETH Zürich $^{3}$ Shenzhen University $^{4}$ Peng Cheng Laboratory", + "bbox": [ + 163, + 215, + 833, + 234 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f7a8af838d0135eb4524fec8191bba62387690d99ec0844c15a7df93abc6edde.jpg", + "image_caption": [ + "Figure 1. Our FastVAR can generate one 2K image using one NVIDIA 3090 GPU, while existing baseline fails due to out of memory." + ], + "image_footnote": [], + "bbox": [ + 93, + 237, + 903, + 425 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 455, + 326, + 470 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual Autoregressive (VAR) modeling has gained popularity for its shift towards next-scale prediction. However, existing VAR paradigms process the entire token map at each scale step, leading to the complexity and runtime scaling dramatically with image resolution. To address this challenge, we propose FastVAR, a post-training acceleration method for efficient resolution scaling with VARs. Our key finding is that the majority of latency arises from the large-scale step where most tokens have already converged. Leveraging this observation, we develop the cached token pruning strategy that only forwards pivotal tokens for scale-specific modeling while using cached tokens from previous scale steps to restore the pruned slots. This significantly reduces the number of forwarded tokens and improves the efficiency at larger resolutions. Experiments show the proposed FastVAR can further speedup FlashAttention-accelerated VAR by $2.7 \\times$ with negligible performance drop of $< 1\\%$ . We further extend FastVAR to zero-shot generation of higher resolution images. In particular, FastVAR can generate one $2K$ image with 15GB memory footprints in 1.5s on a single NVIDIA 3090 GPU. Code is available at https://github.com/csguoh/FastVAR.", + "bbox": [ + 88, + 476, + 485, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 821, + 220, + 837 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The next-token prediction of Autoregressive (AR) models [31, 33, 47, 55, 59] has demonstrated performance com", + "bbox": [ + 89, + 848, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "petitive with diffusion models for visual generation [47] as well as the potential for unified vision understanding and generation [55-57, 59]. However, this token-by-token paradigm suffers from numerous decoding steps. Recently, Visual Autoregressive (VAR) modeling [50] has shifted the paradigm to next-scale prediction, enabling image generation in fewer steps. Under this new paradigm, some works [22, 48] have developed VAR-based models for text-to-image generation and obtained promising results.", + "bbox": [ + 511, + 455, + 905, + 592 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite their potential, existing VAR-based methods [22, 48, 50] face a critical challenge: the computational complexity and runtime latency scale dramatically with image resolution. Specifically, unlike next-token prediction which processes only one token per step, the next-scale prediction of VAR requires processing the entire token map at each decoding step. As a result, the number of tokens increases in $\\mathcal{O}(n^2)$ with the image resolution $n\\times n$ , and even leads to $\\mathcal{O}(n^4)$ complexity in the attention [53] layer. Empirically, as shown in Fig. 2(a), even when FlashAttention [13] is enabled, VAR models still exhibit super-linear runtime latency. Consequently, this significant computational complexity prevents existing VAR-based models from scaling to higher resolutions, such as 2K.", + "bbox": [ + 511, + 595, + 905, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we aim to address the resolution scaling challenge of VARs by pruning forwarded tokens. Through an in-depth analysis of the pre-trained VAR models, we identify the following key findings. 1) Large-scale steps are speed bottleneck but appear robustness. Runtime profiling in Fig. 2(a) finds that the last two large-scale steps account", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Authors, †Project Lead.", + "bbox": [ + 107, + 886, + 323, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19011", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1b3a005446838e9bb0555a4a0b5ba28a31911a56df4a362507c535f304c8b8f0.jpg", + "image_caption": [ + "(a) Runtime at different scale steps" + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 277, + 193 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9ff3b0530a79108f977ec72bf362dc257ced8ca70cbd26c0c4c984d31bacdb20.jpg", + "image_caption": [ + "(b) Efficiency-performance tradeoff", + "Figure 2. FastVAR exhibits promising resolution scalability, and can achieve noticeable speedup with negligible performance drop." + ], + "image_footnote": [], + "bbox": [ + 285, + 88, + 475, + 193 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "for $60\\%$ of the total runtime. Further investigation in Tab. 4 reveals that VAR is more resilient to token pruning at large-scale steps than at smaller scales. Thus, we direct our efforts to the large-scale steps. 2) High-frequency modeling matters at large-scale steps. Spectrum analysis in Fig. 3(b) reveals that the large-scale steps are optimized to model high-frequency tokens, such as texture details, and the remaining low-frequency tokens almost converge in these steps. Consequently, we can only forward the high-frequency tokens for pruning. 3) Tokens from different scales are related. Attention map analysis in Fig. 3(c) shows strong diagonal sparsity across scales, indicating that tokens attend not only to same-scale neighbors but also to those from the preceding scales. Thus, we cache tokens in previous steps to compensate for the pruned slots, preserving the 2D image lattice structure and maintaining information flow.", + "bbox": [ + 88, + 244, + 482, + 486 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on the above observations, we propose FastVAR, a post-training acceleration recipe for efficient resolution scaling with VARs. At its core, FastVAR employs \"cached token pruning\", which retains only pivotal tokens at large-scale steps to reduce computational overhead while using cached token maps from early-scale steps to compensate for information loss. To identify the pivotal tokens, we develop Pivotal Token Selection (PTS), a frequency-based scoring mechanism. Token importance is determined by filtering out the low-frequency component estimated via the direct current component, which enables efficient frequency-based token selection directly in the spatial domain. The selected Top-K pivotal tokens are then processed by the VAR model. To restore pruned tokens, we propose Cached Token Restoration (CTR), which first interpolates the token map from the cached scale and then reinstates the pruned tokens by indexing the interpolated token map to the pruned location. By integrating these strategies, we present FastVAR which enjoys the following benefits:", + "bbox": [ + 91, + 488, + 482, + 773 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The proposed method is training-free and plug-and-play for various VAR-based backbones.", + "- As shown in Fig. 2(b), FastVAR can be integrated with FlashAttention, with further $2.7 \\times$ speedup and $< 1\\%$ performance drop.", + "- In Fig. 1, FastVAR facilitates zero-shot scaling to larger-resolution and can produce one 2K image using only 15GB memory in 1.5s on a single NVIDIA 3090 GPU." + ], + "bbox": [ + 89, + 777, + 482, + 897 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 89, + 653, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Autoregressive Visual Generation. The previous autoregressive (AR) methods [27, 42, 61, 64] mostly adopt the next-token prediction paradigm, which treats each pixel as one token and generates pixels in a GPT or BERT style [1, 4, 6, 14, 35]. For instance, some pioneering works [9, 51] generate pixels in raster scan order with transformer models. Later, VQVAE [52] and VQGAN [15] propose to quantize image patches into discrete tokens to ease the training process. Benefiting from the scaling up [24, 26], recent works have shown promising results that the autoregressive models can achieve a more competitive performance than state-of-the-art Diffusion models [7, 8, 38, 39, 41, 43]. Despite this progress, existing next-token generation pipelines struggle to efficiently synthesize high-resolution images due to the numerous decoding steps [10]. More recently, the Visual Autoregressive (VAR) modeling [50], which shifts to next-scale generation fashion to allow only a few forward steps, further advances the generation quality as well as the model efficiency. Building on top of VAR, existing methods have scaled up the model for production-level text-to-image generation. For example, HART [48] adopts the continuous diffusion module to compensate for the quantization error. Infinity [22] employs a bitwise tokenizer for an extremely large vector vocabulary. Despite the promising ability of VAR-based methods, they struggle to scale to larger resolutions with the increasing token numbers.", + "bbox": [ + 511, + 114, + 903, + 507 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Efficient Visual Generation. To speed up diffusion models [25], many efforts have been made in recent years, including distillation [40, 44], quantization [21, 32, 46], pruning [2, 17, 54, 62, 65], and caching [30, 34, 36, 37], which either reduce the total denoising steps or to reduce the cost of single forward. Specifically, DeepCache [37] proposes to reuse the intermediate features of low-resolution layers in the U-Net. ToMeSD [2] merges similar tokens into one token with subsequent unmerging for acceleration. Unfortunately, these methods are specifically designed for diffusion, which can achieve only sub-optimal performance or cannot be used in VAR. In the efforts to accelerate AR modeling [31, 33, 47, 55, 59], one prevalent solution is to reduce the forward step using parallel decoding strategies [23, 45, 49]. For instance, speculative decoding [5, 28] utilizes a small draft model to generate candidate tokens, which are then verified in parallel by the larger model. However, the well-developed decoding strategy cannot be directly applied to the next-scale paradigm, whose each step involves multiple pixels [50]. As for fast VAR models, one related work is the CoDe [10], which uses model ensemble techniques to use a small model on the costly large resolution. However, CoDe highly depends on the availability of different-sized models. In this work, we attempt to speed up the VARs by pruning extra tokens, which is training-free and is generic for multiple VAR methods.", + "bbox": [ + 511, + 508, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19012", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9d8bdeccafa1ca64e9f1197238f8b3456a726305eeec8e8e90c1bb34b173930d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 88, + 890, + 138 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/493f32c14f90245c102a0a13a3f008567e5e08c10860e4d798a7f2eb4fa4e4e3.jpg", + "image_caption": [ + "(a) Visualization of Intermediate Outputs from Different Scale Steps" + ], + "image_footnote": [], + "bbox": [ + 99, + 143, + 888, + 202 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0fa08fb31825a077185edd39ad7f10f0df57ae52c4e69c777996cd656e04713f.jpg", + "image_caption": [ + "(b) Fourier Analysis of Large-Scale Steps", + "Figure 3. (a) We resize different $\\tilde{r}_k$ to the same size for better presentation. (b) Each curve represents the frequency characteristics of a certain-scale token map. (c) The row represents keys of scale steps $\\{1,2,\\dots ,k\\}$ , and the column is queries of the $k$ -th scale step." + ], + "image_footnote": [], + "bbox": [ + 98, + 210, + 377, + 353 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/35ccdaaf069943da7325975a09b06757f9d4c75129cd14066a203a9768ffcf96.jpg", + "image_caption": [ + "(c) Cross-scale Attention Map Visualization" + ], + "image_footnote": [], + "bbox": [ + 403, + 213, + 586, + 349 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5a5fa61cc374bd3fb16d4ad55eecf03d9eb4bc15bb293975f8120a80c36588cc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 588, + 213, + 723, + 349 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/492600e259c2361c9bd4d44256ebf080910d89623d55b7e588ec4836a513edc6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 727, + 213, + 893, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 409, + 181, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 89, + 434, + 218, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The Visual Autoregressive (VAR) modeling [50] redefines the traditional AR by shifting from a \"next-token prediction\" to a \"next-scale prediction\", where each autoregressive unit is a token map of varying scales instead of a single token. For a given image feature map, VAR first quantizes it into $K$ multi-scale token maps $\\mathcal{R} = \\{r_1, r_2, \\dots, r_K\\}$ with progressively larger resolutions using pre-defined scale schdual $\\{(h_1, w_1), (h_2, w_2), \\dots, (h_K, w_K)\\}$ . Then the $\\mathcal{R}$ is used to train a Transformer [53] to learn the joint distribution using the causal formulation:", + "bbox": [ + 89, + 455, + 483, + 608 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np \\left(r _ {1}, r _ {2}, \\dots , r _ {K}\\right) = \\prod_ {k = 1} ^ {K} p \\left(r _ {k} \\mid r _ {1}, r _ {2}, \\dots , r _ {k - 1}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 619, + 482, + 661 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the start token $r_1$ is the encoded text embedding from the LLMs [11, 60] under given user prompts. During training, existing methods [22, 48, 50] usually employ a residual strategy to reduce learning difficulty. In detail, the model outputs at the $k$ -th scale step $f_k$ is the residual of the intermediate prediction $\\tilde{r}_k$ :", + "bbox": [ + 89, + 672, + 483, + 763 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {r} _ {k} = \\operatorname {i n t e r p o l a t e} \\left(\\tilde {r} _ {k - 1}, \\left(h _ {k}, w _ {k}\\right)\\right) + f _ {k}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 776, + 482, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the \"interpolate $(\\cdot ,(h_k,w_k))$ denotes upsample a token map to the size of $(h_k,w_k)$ . Unfolding Eq. (2) derives the following cumulative form:", + "bbox": [ + 89, + 804, + 483, + 849 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {r} _ {k} = \\sum_ {i = 1} ^ {k} \\operatorname {i n t e r p o l a t e} \\left(f _ {i}, \\left(h _ {k}, w _ {k}\\right)\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 863, + 482, + 904 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the last scale step, the predicted token maps $\\tilde{r}_K$ is used to generate the final image. During inference, since the outputs at different scales are generated autoregressively, the KV-Cache can be adopted to avoid re-computing of previous steps. Although the VAR paradigm can generate images within only $K$ scale steps, the Transformer needs to process all $h_k \\times w_k$ tokens at the $k$ -th step, hindering scaling to higher-resolution image synthesis.", + "bbox": [ + 511, + 410, + 906, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Motivation", + "text_level": 1, + "bbox": [ + 511, + 541, + 635, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we aim to approach linear VAR models to advance high-resolution image generation. To this end, we investigate the scale-wise behavior of the pre-trained VAR model [22]. We examine from three angles: the latency profiling, the specific role of each step, and the cross-scale token dependency. Conclusions are summarized as follows.", + "bbox": [ + 511, + 564, + 906, + 655 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Observation 1: Large-scale steps are speed bottleneck but appear robustness. As shown in the runtime curve in Fig. 2(a), even the FlashAttention is enabled, the inference latency of the VAR exhibits a super-linear complexity with resolution steps. This is because the Transformer needs to take all the tokens at once. As a result, the tokens from the large-scale steps make up the majority of the total token length. For example, the last two steps even occupy $60\\%$ of the total time. On the other hand, the experiments in Sec. 4.3 suggest the large-scale steps are more robust to token drop than small-scale counterparts under the same pruning ratio. Therefore, it is accessible to prune some \"unimportant tokens\" at the large scale step to speed up token processing. However, challenges still remain about which tokens should be pruned and how to compensate for the information loss from these pruned tokens.", + "bbox": [ + 511, + 659, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19013", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/24e5cc3c3e7b46970a2922892723195dab6692937a52949043f0f5fe1c5ba1ae.jpg", + "image_caption": [ + "(a) Overall pipeline of the proposed FastVAR" + ], + "image_footnote": [], + "bbox": [ + 117, + 88, + 486, + 280 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8d36b4ffa8b04f161b459102072d2324e7ff1b95e97aa8a295b4acc60ecc1d20.jpg", + "image_caption": [ + "(b) Token-wise implementation for one VAR layer at stage $\\mathcal{T}$", + "Figure 4. (a) We keep the Structure Construction Stage as standard VAR and additionally store token maps at the $(K - N)$ -th step as the cache. Then the Texture Filling Stage applies FastVAR to process only pivotal tokens for fast inference. (b) For Attention or FFN layers using FastVAR, we first employ the Pivotal Token Selection (PTS) for token reduction. After the model forward, we develop the Cached Token Restoration (CTR) to restore the original token numbers using the caches from the previous scale step." + ], + "image_footnote": [], + "bbox": [ + 516, + 92, + 880, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Observation 2: High-frequency modeling matters at large-scale steps. To find which token should be pruned, we visualize the intermediate predictions $\\tilde{r}_k$ in Fig. 3(a). Interestingly, we find that the generation process of the pretrained VAR model can be decomposed into two stages. The Structure Construction Stage, which includes small-scale steps, is trained to generate the outline of a subject. After that, at the Texture Filling Stage, which is composed of large-scale steps, is responsible for adding details based on the previous sketch. Given that the outline is low-frequency while the texture is high-frequency, we therefore infer that the VAR mainly models the high-frequency texture in the large-scale stage while keeping the low-frequency structure almost intact. For further verification, we show in Fig. 3(b) that the low-frequency component almost converges in the large-scale steps, while the high-frequency modeling stays optimized with noticeable variations. Therefore, we can prune the redundant low-frequency tokens and forward only the pivotal high-frequency tokens at the large-scale steps for token number reduction.", + "bbox": [ + 88, + 380, + 483, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Observation 3: Tokens from different scales are related.", + "text_level": 1, + "bbox": [ + 89, + 689, + 482, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After reducing the input tokens for fast forward, we still need to restore the output tokens to the original numbers in order to maintain the 2D image structure. To this end, we analyze the cross-scale attention map of the unpruned VAR model, where the query at the current scale interacts with KV caches from all previous scales. As depicted in Fig. 3(c), we observe that the attention map not only exhibits a high diagonal score at self-scale, but also a high diagonal score at cross-scale. This behavior indicates that a token in the current scale not only attends to its same-scale locals but also exhibits strong correlation to tokens at the neighboring positions in previous scales. This diagonal attention sparsity thus facilitates us to use only a few", + "bbox": [ + 88, + 704, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tokens to estimate the original output of the pruned slots, and avoids weighting multiple tokens across the token map. Specifically, we can approximate the outputs of pruned tokens by indexing in the previous-scale token maps based on the pruning location. In this way, the information loss can be compensated thanks to this high cross-scale similarity.", + "bbox": [ + 511, + 380, + 906, + 472 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Efficient Resolution Scaling for VARs", + "text_level": 1, + "bbox": [ + 511, + 483, + 836, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We instantiate the above idea and present FastVAR to approximate linear VAR models. The core is the cached token pruning, which prunes tokens at large scale steps while using cache from early scales to maintain information flow.", + "bbox": [ + 511, + 506, + 905, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overview. As shown in Fig. 4(a), denote the scale step set of Structure Construction Stage as $S = \\{1,2,\\dots ,K - N\\}$ the step set of Texture Filling Stage as $\\mathcal{T} = \\{K - N + 1,\\dots ,K\\}$ . As stated in Sec. 3.2, the speedup of pruning token maps in $S$ is insignificant while producing a noticeable performance drop, we therefore keep $S$ intact as the standard VAR. In addition, we save the per-layer outputs at the last scale in $S$ , i.e., the $(K - N)$ -th step, as the cache for subsequent token restoration. For the set $\\mathcal{T}$ , we apply FastVAR to prune tokens for fast-forward, followed by the token number restoration. More details are given below.", + "bbox": [ + 511, + 566, + 906, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pivotal Token Selection. We first introduce the Pivotal Token Selection (PTS) to reduce the number of tokens. As observed in Fig. 3(b), the low-frequency tokens almost converge in the later scale steps. Therefore, we can feed only the high-frequency tokens and prune the remaining. However, the challenge lies in that common frequency operators, such as FFT, work in the frequency domain, making it difficult to recognize the frequency characteristics of certain tokens in the original token map. To this end, our proposed PTS adopts an approximation solution. Specifically, denote $x_{k}$ as the input of one layer, we estimate its low-frequency", + "bbox": [ + 511, + 734, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19014", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "component $\\bar{x}_k$ as the direct current component, which can be easily obtained via the spatially global average pooling:", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {x} _ {k} = \\text {g l o b a l - a v g - p o o l} \\left(x _ {k}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 132, + 482, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After that, the high-frequency component is computed as the difference between the $x_{k}$ and $\\bar{x}_k$ . The pivotal score $s_k$ is defined as the L2 norm of the high-frequency maps:", + "bbox": [ + 89, + 159, + 482, + 205 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns _ {k} = \\left\\| x _ {k} - \\bar {x} _ {k} \\right\\| _ {2}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 222, + 215, + 482, + 233 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Subsequently, we obtain the index set $\\mathcal{I}$ for pivotal token selection by keeping TopK scores in $s_k$ . Finally, we perform token pruning on $x_{k}$ through indexing with $\\mathcal{I}$ to allow the fast model forward. Notably, since the PTS reduces the number of input tokens in one transformer layer, as an additional benefit, the KV-cache is also reduced accordingly, thus optimizing the GPU footprints as well as the cross-scale attention for the subsequent scale steps.", + "bbox": [ + 89, + 242, + 483, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Cached Token Restoration. The proposed PTS can effectively reduce the number of tokens. However, the visual generation tasks need to restore the original token numbers to regain the structural 2D image. For this reason, we propose Cached Token Restoration (CTR). The main rationale stems from the diagonal attention sparsity that the token at the $k$ -th scale not only attends to itself, but also exhibits strong correlation with the corresponding slot from previous $\\{1,2,\\dots,k-1\\}$ steps. Therefore, our proposed CTR approximates the layer outputs in the pruned locations using the counterparts from the cached step. Formally, given the cached token map $y_{K-N}$ with size $(h_{K-N},w_{K-N})$ , which is the output of the corresponding layer at step $K-N$ , we first upsample it to the same size of the current scale:", + "bbox": [ + 89, + 364, + 483, + 575 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ny _ {k} ^ {\\text {c a c h e}} = \\operatorname {i n t e r p o l a t e} \\left(y _ {K - N}, \\left(h _ {k}, w _ {k}\\right)\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 585, + 482, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then the value from $y_{k}^{cache}$ is scattered into the pruned slots of $y_{k}$ using the index set $\\mathcal{I}$ , to generate the restored layer output $y_{k}^{\\prime}$ , which shares the same token numbers as the $x_{k}$ . We also provide an explanation for the choice of $K - N$ as the caching step, see Suppl. for more details.", + "bbox": [ + 89, + 613, + 482, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation of FastVAR. Thanks to the token number invariance of the $\\mathrm{PTS + CTR}$ , as shown in Fig. 4(b), we apply this operator pair for each Attention&FFN layer at the large-scale steps. Furthermore, we empirically find in Sec. 4.3 that the larger scale steps exhibit stronger robustness to pruning than that of early scale steps. Therefore, we develop a progressive pruning ratio schedule, in which we assign larger pruning ratios to larger scale steps in the set $\\mathcal{T}$ . We summarize the complete algorithm of FastVAR in the Suppl.. At last, it is worth noting that our FastVAR does not access the attention map, and we show in Sec. 4.4 that the proposed FastVAR can be used in combination with other acceleration techniques like FlashAttention [12, 13] for even faster VAR generation.", + "bbox": [ + 89, + 688, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 90, + 645, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 513, + 114, + 702, + 132 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Models and Evaluations. We apply our FastVAR on two VAR-based text-to-image models, namely HART [48] and Infinity [22]. Both models can generate images of up to $1024 \\times 1024$ resolution. For a fair comparison, we keep all the hyperparameters the same as their default settings. For evaluation metrics, we compare both in terms of generation quality and inference efficiency. For generation quality, we use two popular benchmarks, i.e., the GenEval [20] and MJHQ30K [29] to validate the high-level semantic consistency and the perceptual quality, respectively. For efficiency evaluation, we adopt metrics including running time, throughputs, speedup ratio, and GPU memory costs.", + "bbox": [ + 511, + 138, + 906, + 319 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. For the number of pruned scale steps $N$ , we set $N = 4$ for Infinity and $N = 2$ for HART, i.e., only token maps from the last 4 or 2 scale steps are applied with the proposed FastVAR, with other steps kept as standard VAR. For the progressive pruning ratio schedule, we set it to $\\{40\\%, 50\\%, 100\\%, 100\\}$ for Infinity, and $\\{50\\%, 75\\%\\}$ for HART. The $100\\%$ pruning ratio indicates all tokens are dropped, i.e., we skip the corresponding steps and interpolate the intermediate outputs to the target resolution as the final outputs. Note that this extreme ratio depends on the pruned backbones, and we provide further discussion in the Suppl.. Unless specified, the efficiency of the unpruned baselines are already accelerated by FlashAttention [13], and we apply the proposed FastVAR on top of it. Following the setup of existing methods [10, 47, 50], the inference speeds for all methods are measured without including the VAE time cost since this is a shared cost for all methods. All experiments are conducted on one single NVIDIA RTX 3090 GPU with 24GB memory.", + "bbox": [ + 511, + 321, + 908, + 608 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 513, + 619, + 653, + 635 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparison on GenEval. We first evaluate the quality-efficiency trade-off on $1024 \\times 1024$ text-to-image generation using the GenEval benchmark [20]. We compare our FastVAR against various state-of-the-art methods, including Diffusion, AR, and VAR models. The results are shown in Tab. 1. Compared with traditional AR models, our Infinity+FastVAR can achieve even $39.7 \\times$ speedup than LlamaGen [47], while achieving $125\\%$ performance boosts on the GenEval. Moreover, our FastVAR achieves acceleration almost without performance loss, e.g., $1.5 \\times$ speedup on the HART with the same GenEval score as the unpruned baseline. Our approach also reduces the inference GPU memory. For example, FastVAR achieves $22.2\\%$ reduction to 14.7GB on top of FlashAttention of 18.9GB, to produce a $1024 \\times 1024$ image in 0.95s, facilitating generation on consumer-level GPUs. The above results demonstrate the generality and effectiveness of our method.", + "bbox": [ + 511, + 643, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19015", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/303101b09a27c9c10096c0527fd1f9b91d178cccd6951b4c87c2c09797ecf125.jpg", + "table_caption": [ + "Table 1. Quantitative comparison on efficiency and quality on ${1024} \\times {1024}$ GenEval benchmarks. The marks $\\bullet$ , $\\circ$ ,and $\\diamond$ denote the Diffusion,AR,and VAR-based methods,respectively. Note that the efficiency of HART and Infinity baselines are tested under FlashAttention." + ], + "table_footnote": [], + "table_body": "
MethodsInference EfficiencyGeneration Quality
#Steps↓Speedup↑Latency↓Throughput↑#Param↓two_object↑position↑color_attr↑Overall↑
SDXL [41]40-4.3s0.23it/s2.6B0.740.150.230.55
PixArt-Sigma [8]20-2.7s2.50it/s0.6B0.620.140.270.55
SD3-medium [16]28-4.4s3.45it/s2.0B0.740.340.360.62
○LlamaGen [47]1024-37.7s0.03it/s0.8B0.340.070.040.32
○Show-o [59]1024-50.3s0.02it/s1.3B0.800.310.500.68
HART [48]141.0×0.95s1.05it/s0.7B0.620.130.180.51
+FastVAR141.5×0.63s1.59it/s0.7B0.570.160.240.51
Infinity [22]131.0×2.61s0.38it/s2.0B0.850.440.530.73
+FastVAR132.7×0.95s1.05it/s2.0B0.810.450.520.72
", + "bbox": [ + 106, + 121, + 898, + 287 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ca3246aafd90a0ded831663d6ab6ee845d1373f41b746f038973bc2887e5f6f5.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison between the original baselines and our proposed FastVAR on $1024 \\times 1024$ image generation. Our FastVAR achieves significant speedups, e.g., $2.7 \\times$ on Infinity [22], while keeping high-quality results similar to the original model." + ], + "image_footnote": [], + "bbox": [ + 99, + 297, + 897, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison on MJHQ30K. In Tab. 2, we further validate the perceptual quality on the MJHQ30K [29] benchmark. It can be seen that our FastVAR achieves a reasonable performance while maintaining a high speedup ratio. For instance, on the well-known challenging \"people\" category, our HART+FastVAR even achieves a FID reduction of 2.42 with $1.5 \\times$ acceleration. In other categories, our method also maintains good performance with significant acceleration. We also give a qualitative visualization in Fig. 5, and one can see that the generated images from FastVAR maintain exceptionally high quality and accurate semantic information, indicating the effectiveness of the proposed method.", + "bbox": [ + 89, + 652, + 482, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Difference from ToMe. Token Merging (ToMe) [3], which merges multiple tokens into one token for efficient forward followed by token unmerging to restore the 2D shape, appears applicable to accelerate VAR backbones due to its to-", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/76429357bd40d645a0983f2ba83f45456508211724e43a144241b55bf76dd7cf.jpg", + "table_caption": [ + "Table 2. Quantitative comparisons of FID and CLIP score with different generation categories on the MJHQ30K dataset." + ], + "table_footnote": [], + "table_body": "
MethodsSpeeduplandscapepeoplefood
FID↓CLIP↑FID↓CLIP↑FID↓CLIP↑
SDXL [41]-30.7826.3535.5628.0135.2627.98
HART [48]1.0×25.4326.8230.6128.4730.3728.03
+ FastVAR1.5×22.5226.5128.1928.3430.9728.25
Infinity [22]1.0×24.6826.6230.2727.8231.5526.66
+ FastVAR2.7×24.6826.6230.5528.2832.5427.08
", + "bbox": [ + 527, + 679, + 890, + 785 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ken number invariance. In Tab. 3, we set different FastVAR and ToMe variants using varying pruning ratios. One can see that ToMe fails to achieve high speedup. For example, even the $1.36 \\times$ speedup can lead to noticeable FID degradation since it is difficult to compress the whole token map into limited tokens. In contrast, our FastVAR can achieve $1.7 \\times$ higher speedup with better FID performance.", + "bbox": [ + 511, + 794, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19016", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/95d59ba43b5936222d43f5d4b119ef4b9fc8a663b998853ec6e0a833a5ad1c13.jpg", + "image_caption": [ + "Figure 6. Quantitative and qualitative ablation experiments on different pruning ratios. We set varying pruning ratios on the last scale step of the HART model while keeping the other scale steps as the original setup." + ], + "image_footnote": [], + "bbox": [ + 101, + 88, + 302, + 219 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8afb013a4243a68dbfbd5ea9c511585290e78de6618b0c219fbeb6bdc3c136b2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 310, + 88, + 522, + 219 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fe0218ae1536f8b1422222173a463b64a4ccfebc5c3c28d9f407643be9810b49.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 529, + 88, + 890, + 218 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e40f35637c23aeeaf20f9445cc4c4160f2897465b40b04a7f8accbfaeb31cbfd.jpg", + "table_caption": [ + "Table 3. Quantitative comparison results with ToMe [2] under different speedup settings." + ], + "table_footnote": [], + "table_body": "
MethodsSpeedup↑Latency↓Throughput↑FID↓CLIP↑GenEval↑
HART [48]1.00×0.95s1.05it/s30.6128.470.51
+ ToMe [2]1.19×0.80s1.25it/s29.0728.330.48
+ ToMe [2]1.36×0.70s1.43it/s35.2227.990.46
+ FastVAR1.51×0.63s1.59it/s28.1928.340.51
+ FastVAR1.70×0.56s1.79it/s28.9728.240.50
", + "bbox": [ + 109, + 291, + 467, + 375 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0b2713db4625cd506a8a1a5c75b88a12d0cf6327becca45eb0a113cad31a45c6.jpg", + "table_caption": [ + "Table 4. Ablation experiments on the scale-wise sensitivity. We focus on the scale range [16,21,27,36,48,64] with every two consecutive scales as a pruning group. The FastVAR is applied to the selected group while keeping the others unpruned." + ], + "table_footnote": [], + "table_body": "
scalesratioSpeedup↑Latency↓FID↓CLIP↑GenEval↑
baseline01.00×0.95s30.6128.470.51
[16, 21]75%1.01×0.94s34.2728.430.48
[27, 36]50%1.09×0.87s27.9228.490.51
[27, 36]75%1.14×0.83s29.3828.560.50
[48, 64]50%1.30×0.73s27.1328.430.51
[48, 64]75%1.64×0.58s28.5628.360.50
", + "bbox": [ + 107, + 448, + 465, + 540 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 559, + 254, + 574 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Different Pruning Ratios. The pruning ratio plays an important role in balancing between efficiency and performance. Here, we conduct ablation experiments to investigate the impact of different pruning ratios. Both the quantitative and qualitative results are given in Fig. 6. Intuitively, increasing the pruning ratio leads to a stable runtime reduction since the model only needs to process fewer tokens. However, an over-large pruning ratio can also bring performance degradation due to the information loss of pivotal high-frequency tokens. Furthermore, the quantitative visualization suggests that increasing the pruning ratio makes certain textures and details discontinuous since the cache from the previous scale is not always an optimal approximation of the pruned tokens. Therefore, a moderate pruning ratio, e.g., $40\\% -75\\%$ , is adopted for HART+FastVAR to strike a sweet spot between performance and efficiency.", + "bbox": [ + 88, + 582, + 482, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Scale-wise Sensitivity. As discussed in Sec. 3.2, we only focus on the token pruning of the last few scale steps. Here, we conduct ablation experiments in Tab. 4 to justify the rationality. It can be seen that the acceleration from pruning at an early scale steps is very limited due to the small token", + "bbox": [ + 88, + 824, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1b68d0fa6560028ce10e49ad4ae2ce9f7fbe612e2597aae4face0ac9d5b69d86.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 262, + 898, + 335 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ba02d2ee399a712ff23a0a09bbbbc5393e13c2ea3edc938e7c8d71cf3ee0c2c6.jpg", + "image_caption": [ + "Figure 7. Visualization of pruned tokens with different ratios." + ], + "image_footnote": [], + "bbox": [ + 517, + 335, + 898, + 397 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "map size. For example, even reducing $50\\%$ tokens at the 48 and 64 scale steps is significantly faster than that of $75\\%$ at 16 and 21 scale steps. In addition, pruning at early scales also leads to a noticeable performance drop on both perceptual quality and semantic consistency. For instance, pruning at 48 and 64 scale steps brings 5.71 lower FID compared to that of 16 and 21 scale steps. This is because token reduction at small-scale steps disrupts the structure construction stage, and the inaccurate subject structure can exacerbate errors in subsequent scale steps.", + "bbox": [ + 511, + 431, + 906, + 582 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visualization of Pruned Slots. In Fig. 7, we visualize the selected TopK pivotal tokens under different pruning ratios. For a given pruning ratio, e.g., $50\\%$ , our FastVAR prioritizes the retention of high-frequency edges or texture slots, and removes tokens from flat regions that have already converged at early scale steps. Furthermore, the importance order within the selected tokens is also meaningful. For example, in the first row, when increasing the pruning ratio from $50\\%$ to $90\\%$ , the human eyes, hair and mouth regions, which are the most detailed, are consistently preserved while the low-frequency cheek regions are gradually pruned. Given that our FastVAR is gradient-free, this quantitative result confirms that the proposed frequency-based pivotal token selection is a reasonable heuristic.", + "bbox": [ + 511, + 582, + 908, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Discussion", + "text_level": 1, + "bbox": [ + 511, + 803, + 633, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Conjunction with FlashAttention. Existing token pruning methods [18, 19, 58, 63, 65] often depend on the attention map for token importance, leading to difficulties in combination with FlashAttention [13] in which the attention map is inaccessible. In comparison, our FastVAR", + "bbox": [ + 511, + 824, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19017", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b321a9722cc7bb24a63b7df84e1787542d633788fb0fa5de757c3a20c69497d9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 89, + 313, + 311 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/96941d5de7ea757e77f7404198a6d0d3145848b84ade08f06c3418fa34a01af3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 88, + 602, + 311 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4cc7f8454fd293b35cfc53f41e94a8b905adbb4baf016501f225b4405feff53c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 604, + 88, + 748, + 311 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6635a24e83fcf8de2d4544eca05daa876b66e9a1d10467262244e5d3cb69d40e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 751, + 88, + 898, + 311 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2e3b1a0d29fff03fb5a775037489a2ba8061355577a03fbc6a62d2b8577df01a.jpg", + "image_caption": [ + "Figure 8. We extend the scale schedule of pre-trained VAR model [22] for zero-shot larger resolution generation. We chose a modest extended scale since a larger one is observed quality drop due to the resolution gap. The images are re-scaled for better presentation." + ], + "image_footnote": [], + "bbox": [ + 94, + 314, + 699, + 512 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6bff136e74124efb6b958946717949c6dcc432b26b39c279cb7df56138b7edb6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 704, + 314, + 897, + 512 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "can be seamlessly integrated with FlashAttention to facilitate more efficient inference. In Tab. 5, we show that using the FastVAR-only can obtain even $2.1 \\times$ speedup against FlashAttention only, with $20.4\\%$ GPU memory reduction. After combining both, the resulting version leads to $2.7 \\times$ speedup than FlashAttention only. This experiment validates that our FastVAR can be orthogonally combined with other methods for further acceleration.", + "bbox": [ + 89, + 560, + 482, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Towards Larger Resolution. As presented in Sec. 3.2, the original VAR struggles to scale to higher-resolution due to the increasing token numbers. Our FastVAR can facilitate efficient resolution scaling, offering the potential to generate images at a larger resolution. To this end, we apply our FastVAR to the Infinity [22] for zero-shot larger resolution synthesis by appending additional steps. The results are shown in Fig. 8. It is noteworthy that even the FlashAttention accelerated baseline is out of memory on a 24GB NVIDIA 3090 GPU. In contrast, our FastVAR costs only 15GB memory and 1.3s runtime to generate a $1344 \\times 1344$ image. Moreover, the generated images are of high quality, suggesting that our FastVAR can facilitate production-level image generation on consumer-level GPUs.", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/547754d69a4a32042a7511ba28e9be002ec18e4c6f512a9885914fde5b585b65.jpg", + "table_caption": [ + "Table 5. Efficiency comparison with FlashAttn on $1024\\times 1024$ image generation." + ], + "table_footnote": [], + "table_body": "
setupSpeedup↑Latency↓Throughput↑Memory↓
SlowAttn [53]---OOM
FlashAttn only [13]1.0×2.61s0.38it/s16.1GB
SlowAttn+FastVAR2.1×1.25s0.80it/s12.8GB
FlashAttn+FastVAR2.7×0.95s1.05it/s11.9GB
", + "bbox": [ + 521, + 588, + 897, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 676, + 633, + 693 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work presents FastVAR to advance high-resolution image synthesis with VARs by addressing the resolution scaling challenge. We reveal that the main latency bottleneck is the large-scale steps at which the low-frequency tokens have almost converged. To this end, we develop the cached token pruning technique to allow only process pivotal high-frequency tokens and use caches from the previous scale steps to restore the original token numbers. Thanks to the reduced forwarded tokens, our FastVAR approximates linear complexity, further enabling generation at a larger resolution such as 2K. Extensive experiments validate our FastVAR as an effective and generic solution for efficient resolution scaling of VAR models.", + "bbox": [ + 509, + 703, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19018", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported in part by the National Natural Science Foundation of China, under Grant (62302309, 62171248), Shenzhen Science and Technology Program (JCYJ20220818101014030, JCYJ20220818101012025).", + "bbox": [ + 89, + 114, + 483, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 191, + 187, + 208 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. GPT-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2", + "[2] Daniel Bolya and Judy Hoffman. Token merging for fast stable diffusion. In CVPR, pages 4599-4603, 2023. 2, 7", + "[3] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token Merging: Your ViT but faster. arXiv preprint arXiv:2210.09461, 2022. 6", + "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 33:1877-1901, 2020. 2", + "[5] Tianle Cai, Yuhong Li, Zhengyang Geng, Hongwu Peng, Jason D Lee, Deming Chen, and Tri Dao. Medusa: Simple LLM inference acceleration framework with multiple decoding heads. arXiv preprint arXiv:2401.10774, 2024. 2", + "[6] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 2", + "[7] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. PixArt-Alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 2", + "[8] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. PixArt-Sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. In ECCV, pages 74-91. Springer, 2024. 2, 6", + "[9] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In ICML, pages 1691-1703. PMLR, 2020. 2", + "[10] Zigeng Chen, Xinyin Ma, Gongfan Fang, and Xinchao Wang. Collaborative decoding makes visual auto-regressive modeling efficient. arXiv preprint arXiv:2411.17787, 2024. 2, 5", + "[11] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instructionfinetuned language models. Journal of Machine Learning Research, 25(70):1-53, 2024. 3" + ], + "bbox": [ + 93, + 215, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Tri Dao. FlashAttention-2: Faster attention with better parallelism and work partitioning. In ICLR, 2024. 5", + "[13] Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flash Attention: Fast and memory-efficient exact attention with IO-awareness. In NeurIPS, 2022. 1, 5, 7, 8", + "[14] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In ACL, pages 4171–4186, 2019. 2", + "[15] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, pages 12873-12883, 2021. 2", + "[16] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 6", + "[17] Gongfan Fang, Xinyin Ma, and Xinchao Wang. Structural pruning for diffusion models. In NeurIPS, 2023. 2", + "[18] Qichen Fu, Minsik Cho, Thomas Merth, Sachin Mehta, Mohammad Rastegari, and Mahyar Najibi. LazyLLM: Dynamic token pruning for efficient long context llm inference. arXiv preprint arXiv:2407.14057, 2024. 7", + "[19] Suyu Ge, Yunan Zhang, Liyuan Liu, Minjia Zhang, Jiawei Han, and Jianfeng Gao. Model tells you what to discard: Adaptive KV cache compression for LLMs. arXiv preprint arXiv:2310.01801, 2023. 7", + "[20] Dhruba Ghosh, Hannaneh Hajishirzi, and Ludwig Schmidt. GenEval: An object-focused framework for evaluating text-to-image alignment. NeurIPS, 36, 2024. 5", + "[21] Hang Guo, Yawei Li, Tao Dai, Shu-Tao Xia, and Luca Benini. Intlora: Integral low-rank adaptation of quantized diffusion models. arXiv preprint arXiv:2410.21759, 2024. 2", + "[22] Jian Han, Jinlai Liu, Yi Jiang, Bin Yan, Yuqi Zhang, Zehuan Yuan, Bingyue Peng, and Xiaobing Liu. Infinity: Scaling bitwise autoregressive modeling for high-resolution image synthesis. arXiv preprint arXiv:2412.04431, 2024. 1, 2, 3, 5, 6, 8", + "[23] Yefei He, Feng Chen, Yuanyu He, Shaoxuan He, Hong Zhou, Kaipeng Zhang, and Bohan Zhuang. ZIPAR: Accelerating autoregressive image generation through spatial locality. arXiv preprint arXiv:2412.04062, 2024. 2", + "[24] Tom Henighan, Jared Kaplan, Mor Katz, Mark Chen, Christopher Hesse, Jacob Jackson, Heewoo Jun, Tom B Brown, Prafulla Dhariwal, Scott Gray, et al. Scaling laws for autoregressive generative modeling. arXiv preprint arXiv:2010.14701, 2020. 2", + "[25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. *NeurIPS*, 33:6840-6851, 2020. 2", + "[26] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361, 2020. 2" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19019", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. In CVPR, pages 11523-11532, 2022. 2", + "[28] Yaniv Leviathan, Matan Kalman, and Yossi Matias. Fast inference from transformers via speculative decoding. In ICML, pages 19274-19286. PMLR, 2023. 2", + "[29] Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Linmiao Xu, and Suhail Doshi. Playground v2.5: Three insights towards enhancing aesthetic quality in text-to-image generation. arXiv preprint arXiv:2402.17245, 2024. 5, 6", + "[30] Senmao Li, Taihang Hu, Fahad Shahbaz Khan, Linxuan Li, Shiqi Yang, Yaxing Wang, Ming-Ming Cheng, and Jian Yang. Faster Diffusion: Rethinking the role of unet encoder in diffusion models. NeurIPS, 2023. 2", + "[31] Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization. arXiv preprint arXiv:2406.11838, 2024. 1, 2", + "[32] Xiuyu Li, Yijiang Liu, Long Lian, Huanrui Yang, Zhen Dong, Daniel Kang, Shanghang Zhang, and Kurt Keutzer. Q-Diffusion: Quantizing diffusion models. In ICCV, pages 17535-17545, 2023. 2", + "[33] Dongyang Liu, Shitian Zhao, Le Zhuo, Weifeng Lin, Yu Qiao, Hongsheng Li, and Peng Gao. Lumina-mgpt: Illuminate flexible photorealistic text-to-image generation with multimodal generative pretraining. arXiv preprint arXiv:2408.02657, 2024. 1, 2", + "[34] Haozhe Liu, Wentian Zhang, Jinheng Xie, Francesco Faccio, Mengmeng Xu, Tao Xiang, Mike Zheng Shou, Juan-Manuel Perez-Rua, and Jurgen Schmidhuber. Faster diffusion via temporal attention decomposition. arXiv preprint arXiv:2404.02747, 2024. 2", + "[35] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 2", + "[36] Xinyin Ma, Gongfan Fang, Michael Bi Mi, and Xinchao Wang. Learning-to-Cache: Accelerating diffusion transformer via layer caching. arXiv preprint arXiv:2406.01733, 2024. 2", + "[37] Xinyin Ma, Gongfan Fang, and Xinchao Wang. DeepCache: Accelerating diffusion models for free. In CVPR, pages 15762-15772, 2024. 2", + "[38] Yue Ma, Yingqing He, Xiaodong Cun, Xintao Wang, Siran Chen, Xiu Li, and Qifeng Chen. Follow your pose: Pose-guided text-to-video generation using pose-free videos. In AAAI, pages 4117-4125, 2024. 2", + "[39] Yue Ma, Hongyu Liu, Hongfa Wang, Heng Pan, Yingqing He, Junkun Yuan, Ailing Zeng, Chengfei Cai, Heung-Yeung Shum, Wei Liu, et al. Follow-your-emoji: Fine-controllable and expressive freestyle portrait animation. In SIGGRAPH Asia, pages 1–12, 2024. 2", + "[40] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. In CVPR, pages 14297-14306, 2023. 2" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 6", + "[42] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with VQ-VAE-2. NeurIPS, 32, 2019. 2", + "[43] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2", + "[44] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 2", + "[45] Andrea Santilli, Silvio Severino, Emilian Postolache, Valentino Maiorca, Michele Mancusi, Riccardo Marin, and Emanuele Rodola. Accelerating transformer inference for translation via parallel decoding. arXiv preprint arXiv:2305.10427, 2023. 2", + "[46] Yuzhang Shang, Zhihang Yuan, Bin Xie, Bingzhe Wu, and Yan Yan. Post-training quantization on diffusion models. In CVPR, pages 1972-1981, 2023. 2", + "[47] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 1, 2, 5, 6", + "[48] Haotian Tang, Yecheng Wu, Shang Yang, Enze Xie, Junsong Chen, Junyu Chen, Zhuoyang Zhang, Han Cai, Yao Lu, and Song Han. HART: Efficient visual generation with hybrid autoregressive transformer. arXiv preprint arXiv:2410.10812, 2024. 1, 2, 3, 5, 6, 7", + "[49] Yao Teng, Han Shi, Xian Liu, Xuefei Ning, Guohao Dai, Yu Wang, Zhenguo Li, and Xihui Liu. Accelerating autoregressive text-to-image generation with training-free speculative jacobi decoding. arXiv preprint arXiv:2410.01699, 2024. 2", + "[50] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. NeurIPS, 2024. 1, 2, 3, 5", + "[51] Aaron Van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Conditional image generation with pixelCNN decoders. NeurIPS, 29, 2016. 2", + "[52] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. NeurIPS, 30, 2017. 2", + "[53] A Vaswani. Attention is all you need. NeurIPS, 2017. 1, 3, 8", + "[54] Hongjie Wang, Difan Liu, Yan Kang, Yijun Li, Zhe Lin, Niraj K Jha, and Yuchen Liu. Attention-driven training-free efficiency enhancement of diffusion models. In CVPR, pages 16080-16089, 2024. 2", + "[55] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. EMU3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 1, 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19020", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[56] Chengyue Wu, Xiaokang Chen, Zhiyu Wu, Yiyang Ma, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, Chong Ruan, et al. Janus: Decoupling visual encoding for unified multimodal understanding and generation. arXiv preprint arXiv:2410.13848, 2024.", + "[57] Yecheng Wu, Zhuoyang Zhang, Junyu Chen, Haotian Tang, Dacheng Li, Yunhao Fang, Ligeng Zhu, Enze Xie, Hongxu Yin, Li Yi, et al. VILA-U: A unified foundation model integrating visual understanding and generation. arXiv preprint arXiv:2409.04429, 2024. 1", + "[58] Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. Efficient streaming language models with attention sinks. arXiv preprint arXiv:2309.17453, 2023. 7", + "[59] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 1, 2, 6", + "[60] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115, 2024. 3", + "[61] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved VQGAN. arXiv preprint arXiv:2110.04627, 2021. 2", + "[62] Evelyn Zhang, Bang Xiao, Jiayi Tang, Qianli Ma, Chang Zou, Xuefei Ning, Xuming Hu, and Linfeng Zhang. Token pruning for caching better: 9 times acceleration on stable diffusion for free. arXiv preprint arXiv:2501.00375, 2024. 2", + "[63] Zhenyu Zhang, Ying Sheng, Tianyi Zhou, Tianlong Chen, Lianmin Zheng, Ruisi Cai, Zhao Song, Yuandong Tian, Christopher Ré, Clark Barrett, et al. H2O: Heavy-hitter oracle for efficient generative inference of large language models. NeurIPS, 36:34661-34710, 2023. 7", + "[64] Chuanxia Zheng, Tung-Long Vuong, Jianfei Cai, and Dinh Phung. MOVQ: Modulating quantized vectors for high-fidelity image generation. NeurIPS, 35:23412-23425, 2022. 2", + "[65] Chang Zou, Xuyang Liu, Ting Liu, Siteng Huang, and Linfeng Zhang. Accelerating diffusion transformers with tokenwise feature caching. arXiv preprint arXiv:2410.05317, 2024. 2, 7" + ], + "bbox": [ + 91, + 90, + 480, + 724 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "19021", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/FastVAR_ Linear Visual Autoregressive Modeling via Cached Token Pruning/70725db1-fcd5-4496-a95d-07038d174933_model.json b/2025/FastVAR_ Linear Visual Autoregressive Modeling via Cached Token Pruning/70725db1-fcd5-4496-a95d-07038d174933_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6185d6cef0dbacf9dff43c3f3bd079432111586e --- /dev/null +++ b/2025/FastVAR_ Linear Visual Autoregressive Modeling via Cached Token Pruning/70725db1-fcd5-4496-a95d-07038d174933_model.json @@ -0,0 +1,2268 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.131, + 0.885, + 0.153 + ], + "angle": 0, + "content": "FastVAR: Linear Visual Autoregressive Modeling via Cached Token Pruning" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.18, + 0.732, + 0.199 + ], + "angle": 0, + "content": "Hang Guo\\(^{1}\\) Yawei Li\\(^{2,*,†}\\) Taolin Zhang\\(^{1}\\) Jiangshan Wang\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.336, + 0.199, + 0.662, + 0.216 + ], + "angle": 0, + "content": "Tao Dai\\(^{3,*}\\) Shu-Tao Xia\\(^{1,4}\\) Luca Benini\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.165, + 0.216, + 0.834, + 0.235 + ], + "angle": 0, + "content": "\\(^{1}\\)Tsinghua University \\(^{2}\\)ETH Zürich \\(^{3}\\)Shenzhen University \\(^{4}\\)Peng Cheng Laboratory" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.238, + 0.905, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.428, + 0.893, + 0.443 + ], + "angle": 0, + "content": "Figure 1. Our FastVAR can generate one 2K image using one NVIDIA 3090 GPU, while existing baseline fails due to out of memory." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.456, + 0.327, + 0.471 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.477, + 0.486, + 0.81 + ], + "angle": 0, + "content": "Visual Autoregressive (VAR) modeling has gained popularity for its shift towards next-scale prediction. However, existing VAR paradigms process the entire token map at each scale step, leading to the complexity and runtime scaling dramatically with image resolution. To address this challenge, we propose FastVAR, a post-training acceleration method for efficient resolution scaling with VARs. Our key finding is that the majority of latency arises from the large-scale step where most tokens have already converged. Leveraging this observation, we develop the cached token pruning strategy that only forwards pivotal tokens for scale-specific modeling while using cached tokens from previous scale steps to restore the pruned slots. This significantly reduces the number of forwarded tokens and improves the efficiency at larger resolutions. Experiments show the proposed FastVAR can further speedup FlashAttention-accelerated VAR by \\(2.7 \\times\\) with negligible performance drop of \\(< 1\\%\\). We further extend FastVAR to zero-shot generation of higher resolution images. In particular, FastVAR can generate one \\(2K\\) image with 15GB memory footprints in 1.5s on a single NVIDIA 3090 GPU. Code is available at https://github.com/csguoh/FastVAR." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.823, + 0.222, + 0.838 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.849, + 0.483, + 0.879 + ], + "angle": 0, + "content": "The next-token prediction of Autoregressive (AR) models [31, 33, 47, 55, 59] has demonstrated performance com" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.457, + 0.906, + 0.593 + ], + "angle": 0, + "content": "petitive with diffusion models for visual generation [47] as well as the potential for unified vision understanding and generation [55-57, 59]. However, this token-by-token paradigm suffers from numerous decoding steps. Recently, Visual Autoregressive (VAR) modeling [50] has shifted the paradigm to next-scale prediction, enabling image generation in fewer steps. Under this new paradigm, some works [22, 48] have developed VAR-based models for text-to-image generation and obtained promising results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.906, + 0.807 + ], + "angle": 0, + "content": "Despite their potential, existing VAR-based methods [22, 48, 50] face a critical challenge: the computational complexity and runtime latency scale dramatically with image resolution. Specifically, unlike next-token prediction which processes only one token per step, the next-scale prediction of VAR requires processing the entire token map at each decoding step. As a result, the number of tokens increases in \\(\\mathcal{O}(n^2)\\) with the image resolution \\(n\\times n\\), and even leads to \\(\\mathcal{O}(n^4)\\) complexity in the attention [53] layer. Empirically, as shown in Fig. 2(a), even when FlashAttention [13] is enabled, VAR models still exhibit super-linear runtime latency. Consequently, this significant computational complexity prevents existing VAR-based models from scaling to higher resolutions, such as 2K." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "In this work, we aim to address the resolution scaling challenge of VARs by pruning forwarded tokens. Through an in-depth analysis of the pre-trained VAR models, we identify the following key findings. 1) Large-scale steps are speed bottleneck but appear robustness. Runtime profiling in Fig. 2(a) finds that the last two large-scale steps account" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.325, + 0.901 + ], + "angle": 0, + "content": "*Corresponding Authors, †Project Lead." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19011" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.278, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.103, + 0.195, + 0.282, + 0.206 + ], + "angle": 0, + "content": "(a) Runtime at different scale steps" + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.089, + 0.476, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.296, + 0.195, + 0.479, + 0.206 + ], + "angle": 0, + "content": "(b) Efficiency-performance tradeoff" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.208, + 0.483, + 0.237 + ], + "angle": 0, + "content": "Figure 2. FastVAR exhibits promising resolution scalability, and can achieve noticeable speedup with negligible performance drop." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.245, + 0.483, + 0.487 + ], + "angle": 0, + "content": "for \\(60\\%\\) of the total runtime. Further investigation in Tab. 4 reveals that VAR is more resilient to token pruning at large-scale steps than at smaller scales. Thus, we direct our efforts to the large-scale steps. 2) High-frequency modeling matters at large-scale steps. Spectrum analysis in Fig. 3(b) reveals that the large-scale steps are optimized to model high-frequency tokens, such as texture details, and the remaining low-frequency tokens almost converge in these steps. Consequently, we can only forward the high-frequency tokens for pruning. 3) Tokens from different scales are related. Attention map analysis in Fig. 3(c) shows strong diagonal sparsity across scales, indicating that tokens attend not only to same-scale neighbors but also to those from the preceding scales. Thus, we cache tokens in previous steps to compensate for the pruned slots, preserving the 2D image lattice structure and maintaining information flow." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.489, + 0.483, + 0.775 + ], + "angle": 0, + "content": "Based on the above observations, we propose FastVAR, a post-training acceleration recipe for efficient resolution scaling with VARs. At its core, FastVAR employs \"cached token pruning\", which retains only pivotal tokens at large-scale steps to reduce computational overhead while using cached token maps from early-scale steps to compensate for information loss. To identify the pivotal tokens, we develop Pivotal Token Selection (PTS), a frequency-based scoring mechanism. Token importance is determined by filtering out the low-frequency component estimated via the direct current component, which enables efficient frequency-based token selection directly in the spatial domain. The selected Top-K pivotal tokens are then processed by the VAR model. To restore pruned tokens, we propose Cached Token Restoration (CTR), which first interpolates the token map from the cached scale and then reinstates the pruned tokens by indexing the interpolated token map to the pruned location. By integrating these strategies, we present FastVAR which enjoys the following benefits:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.778, + 0.483, + 0.806 + ], + "angle": 0, + "content": "- The proposed method is training-free and plug-and-play for various VAR-based backbones." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.808, + 0.483, + 0.852 + ], + "angle": 0, + "content": "- As shown in Fig. 2(b), FastVAR can be integrated with FlashAttention, with further \\(2.7 \\times\\) speedup and \\(< 1\\%\\) performance drop." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.853, + 0.483, + 0.898 + ], + "angle": 0, + "content": "- In Fig. 1, FastVAR facilitates zero-shot scaling to larger-resolution and can produce one 2K image using only 15GB memory in 1.5s on a single NVIDIA 3090 GPU." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.778, + 0.483, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.655, + 0.105 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.116, + 0.905, + 0.508 + ], + "angle": 0, + "content": "Autoregressive Visual Generation. The previous autoregressive (AR) methods [27, 42, 61, 64] mostly adopt the next-token prediction paradigm, which treats each pixel as one token and generates pixels in a GPT or BERT style [1, 4, 6, 14, 35]. For instance, some pioneering works [9, 51] generate pixels in raster scan order with transformer models. Later, VQVAE [52] and VQGAN [15] propose to quantize image patches into discrete tokens to ease the training process. Benefiting from the scaling up [24, 26], recent works have shown promising results that the autoregressive models can achieve a more competitive performance than state-of-the-art Diffusion models [7, 8, 38, 39, 41, 43]. Despite this progress, existing next-token generation pipelines struggle to efficiently synthesize high-resolution images due to the numerous decoding steps [10]. More recently, the Visual Autoregressive (VAR) modeling [50], which shifts to next-scale generation fashion to allow only a few forward steps, further advances the generation quality as well as the model efficiency. Building on top of VAR, existing methods have scaled up the model for production-level text-to-image generation. For example, HART [48] adopts the continuous diffusion module to compensate for the quantization error. Infinity [22] employs a bitwise tokenizer for an extremely large vector vocabulary. Despite the promising ability of VAR-based methods, they struggle to scale to larger resolutions with the increasing token numbers." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.509, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Efficient Visual Generation. To speed up diffusion models [25], many efforts have been made in recent years, including distillation [40, 44], quantization [21, 32, 46], pruning [2, 17, 54, 62, 65], and caching [30, 34, 36, 37], which either reduce the total denoising steps or to reduce the cost of single forward. Specifically, DeepCache [37] proposes to reuse the intermediate features of low-resolution layers in the U-Net. ToMeSD [2] merges similar tokens into one token with subsequent unmerging for acceleration. Unfortunately, these methods are specifically designed for diffusion, which can achieve only sub-optimal performance or cannot be used in VAR. In the efforts to accelerate AR modeling [31, 33, 47, 55, 59], one prevalent solution is to reduce the forward step using parallel decoding strategies [23, 45, 49]. For instance, speculative decoding [5, 28] utilizes a small draft model to generate candidate tokens, which are then verified in parallel by the larger model. However, the well-developed decoding strategy cannot be directly applied to the next-scale paradigm, whose each step involves multiple pixels [50]. As for fast VAR models, one related work is the CoDe [10], which uses model ensemble techniques to use a small model on the costly large resolution. However, CoDe highly depends on the availability of different-sized models. In this work, we attempt to speed up the VARs by pruning extra tokens, which is training-free and is generic for multiple VAR methods." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19012" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.089, + 0.891, + 0.14 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.144, + 0.89, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.202, + 0.692, + 0.214 + ], + "angle": 0, + "content": "(a) Visualization of Intermediate Outputs from Different Scale Steps" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.212, + 0.379, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.355, + 0.352, + 0.367 + ], + "angle": 0, + "content": "(b) Fourier Analysis of Large-Scale Steps" + }, + { + "type": "image", + "bbox": [ + 0.405, + 0.214, + 0.587, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.214, + 0.725, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.728, + 0.214, + 0.895, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.354, + 0.76, + 0.367 + ], + "angle": 0, + "content": "(c) Cross-scale Attention Map Visualization" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.368, + 0.908, + 0.397 + ], + "angle": 0, + "content": "Figure 3. (a) We resize different \\(\\tilde{r}_k\\) to the same size for better presentation. (b) Each curve represents the frequency characteristics of a certain-scale token map. (c) The row represents keys of scale steps \\(\\{1,2,\\dots ,k\\}\\), and the column is queries of the \\(k\\)-th scale step." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.41, + 0.182, + 0.424 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.435, + 0.22, + 0.451 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.457, + 0.484, + 0.609 + ], + "angle": 0, + "content": "The Visual Autoregressive (VAR) modeling [50] redefines the traditional AR by shifting from a \"next-token prediction\" to a \"next-scale prediction\", where each autoregressive unit is a token map of varying scales instead of a single token. For a given image feature map, VAR first quantizes it into \\( K \\) multi-scale token maps \\( \\mathcal{R} = \\{r_1, r_2, \\dots, r_K\\} \\) with progressively larger resolutions using pre-defined scale schdual \\( \\{(h_1, w_1), (h_2, w_2), \\dots, (h_K, w_K)\\} \\). Then the \\( \\mathcal{R} \\) is used to train a Transformer [53] to learn the joint distribution using the causal formulation:" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.621, + 0.483, + 0.662 + ], + "angle": 0, + "content": "\\[\np \\left(r _ {1}, r _ {2}, \\dots , r _ {K}\\right) = \\prod_ {k = 1} ^ {K} p \\left(r _ {k} \\mid r _ {1}, r _ {2}, \\dots , r _ {k - 1}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.484, + 0.764 + ], + "angle": 0, + "content": "where the start token \\( r_1 \\) is the encoded text embedding from the LLMs [11, 60] under given user prompts. During training, existing methods [22, 48, 50] usually employ a residual strategy to reduce learning difficulty. In detail, the model outputs at the \\( k \\)-th scale step \\( f_k \\) is the residual of the intermediate prediction \\( \\tilde{r}_k \\):" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.777, + 0.483, + 0.793 + ], + "angle": 0, + "content": "\\[\n\\tilde {r} _ {k} = \\operatorname {i n t e r p o l a t e} \\left(\\tilde {r} _ {k - 1}, \\left(h _ {k}, w _ {k}\\right)\\right) + f _ {k}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.805, + 0.484, + 0.851 + ], + "angle": 0, + "content": "where the \"interpolate \\((\\cdot ,(h_k,w_k))\\) denotes upsample a token map to the size of \\((h_k,w_k)\\). Unfolding Eq. (2) derives the following cumulative form:" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.864, + 0.483, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\tilde {r} _ {k} = \\sum_ {i = 1} ^ {k} \\operatorname {i n t e r p o l a t e} \\left(f _ {i}, \\left(h _ {k}, w _ {k}\\right)\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.411, + 0.907, + 0.533 + ], + "angle": 0, + "content": "In the last scale step, the predicted token maps \\(\\tilde{r}_K\\) is used to generate the final image. During inference, since the outputs at different scales are generated autoregressively, the KV-Cache can be adopted to avoid re-computing of previous steps. Although the VAR paradigm can generate images within only \\(K\\) scale steps, the Transformer needs to process all \\(h_k \\times w_k\\) tokens at the \\(k\\)-th step, hindering scaling to higher-resolution image synthesis." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.542, + 0.636, + 0.557 + ], + "angle": 0, + "content": "3.2. Motivation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.907, + 0.656 + ], + "angle": 0, + "content": "In this work, we aim to approach linear VAR models to advance high-resolution image generation. To this end, we investigate the scale-wise behavior of the pre-trained VAR model [22]. We examine from three angles: the latency profiling, the specific role of each step, and the cross-scale token dependency. Conclusions are summarized as follows." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Observation 1: Large-scale steps are speed bottleneck but appear robustness. As shown in the runtime curve in Fig. 2(a), even the FlashAttention is enabled, the inference latency of the VAR exhibits a super-linear complexity with resolution steps. This is because the Transformer needs to take all the tokens at once. As a result, the tokens from the large-scale steps make up the majority of the total token length. For example, the last two steps even occupy \\(60\\%\\) of the total time. On the other hand, the experiments in Sec. 4.3 suggest the large-scale steps are more robust to token drop than small-scale counterparts under the same pruning ratio. Therefore, it is accessible to prune some \"unimportant tokens\" at the large scale step to speed up token processing. However, challenges still remain about which tokens should be pruned and how to compensate for the information loss from these pruned tokens." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19013" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.089, + 0.488, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.171, + 0.282, + 0.422, + 0.294 + ], + "angle": 0, + "content": "(a) Overall pipeline of the proposed FastVAR" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.093, + 0.882, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.281, + 0.868, + 0.294 + ], + "angle": 0, + "content": "(b) Token-wise implementation for one VAR layer at stage \\(\\mathcal{T}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.299, + 0.907, + 0.355 + ], + "angle": 0, + "content": "Figure 4. (a) We keep the Structure Construction Stage as standard VAR and additionally store token maps at the \\((K - N)\\)-th step as the cache. Then the Texture Filling Stage applies FastVAR to process only pivotal tokens for fast inference. (b) For Attention or FFN layers using FastVAR, we first employ the Pivotal Token Selection (PTS) for token reduction. After the model forward, we develop the Cached Token Restoration (CTR) to restore the original token numbers using the caches from the previous scale step." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.381, + 0.485, + 0.681 + ], + "angle": 0, + "content": "Observation 2: High-frequency modeling matters at large-scale steps. To find which token should be pruned, we visualize the intermediate predictions \\(\\tilde{r}_k\\) in Fig. 3(a). Interestingly, we find that the generation process of the pretrained VAR model can be decomposed into two stages. The Structure Construction Stage, which includes small-scale steps, is trained to generate the outline of a subject. After that, at the Texture Filling Stage, which is composed of large-scale steps, is responsible for adding details based on the previous sketch. Given that the outline is low-frequency while the texture is high-frequency, we therefore infer that the VAR mainly models the high-frequency texture in the large-scale stage while keeping the low-frequency structure almost intact. For further verification, we show in Fig. 3(b) that the low-frequency component almost converges in the large-scale steps, while the high-frequency modeling stays optimized with noticeable variations. Therefore, we can prune the redundant low-frequency tokens and forward only the pivotal high-frequency tokens at the large-scale steps for token number reduction." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.703 + ], + "angle": 0, + "content": "Observation 3: Tokens from different scales are related." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.705, + 0.483, + 0.901 + ], + "angle": 0, + "content": "After reducing the input tokens for fast forward, we still need to restore the output tokens to the original numbers in order to maintain the 2D image structure. To this end, we analyze the cross-scale attention map of the unpruned VAR model, where the query at the current scale interacts with KV caches from all previous scales. As depicted in Fig. 3(c), we observe that the attention map not only exhibits a high diagonal score at self-scale, but also a high diagonal score at cross-scale. This behavior indicates that a token in the current scale not only attends to its same-scale locals but also exhibits strong correlation to tokens at the neighboring positions in previous scales. This diagonal attention sparsity thus facilitates us to use only a few" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.381, + 0.907, + 0.473 + ], + "angle": 0, + "content": "tokens to estimate the original output of the pruned slots, and avoids weighting multiple tokens across the token map. Specifically, we can approximate the outputs of pruned tokens by indexing in the previous-scale token maps based on the pruning location. In this way, the information loss can be compensated thanks to this high cross-scale similarity." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.484, + 0.837, + 0.5 + ], + "angle": 0, + "content": "3.3. Efficient Resolution Scaling for VARs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.906, + 0.567 + ], + "angle": 0, + "content": "We instantiate the above idea and present FastVAR to approximate linear VAR models. The core is the cached token pruning, which prunes tokens at large scale steps while using cache from early scales to maintain information flow." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.907, + 0.734 + ], + "angle": 0, + "content": "Overview. As shown in Fig. 4(a), denote the scale step set of Structure Construction Stage as \\( S = \\{1,2,\\dots ,K - N\\} \\) the step set of Texture Filling Stage as \\( \\mathcal{T} = \\{K - N + 1,\\dots ,K\\} \\). As stated in Sec. 3.2, the speedup of pruning token maps in \\( S \\) is insignificant while producing a noticeable performance drop, we therefore keep \\( S \\) intact as the standard VAR. In addition, we save the per-layer outputs at the last scale in \\( S \\), i.e., the \\( (K - N) \\)-th step, as the cache for subsequent token restoration. For the set \\( \\mathcal{T} \\), we apply FastVAR to prune tokens for fast-forward, followed by the token number restoration. More details are given below." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Pivotal Token Selection. We first introduce the Pivotal Token Selection (PTS) to reduce the number of tokens. As observed in Fig. 3(b), the low-frequency tokens almost converge in the later scale steps. Therefore, we can feed only the high-frequency tokens and prune the remaining. However, the challenge lies in that common frequency operators, such as FFT, work in the frequency domain, making it difficult to recognize the frequency characteristics of certain tokens in the original token map. To this end, our proposed PTS adopts an approximation solution. Specifically, denote \\( x_{k} \\) as the input of one layer, we estimate its low-frequency" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19014" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.123 + ], + "angle": 0, + "content": "component \\(\\bar{x}_k\\) as the direct current component, which can be easily obtained via the spatially global average pooling:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.133, + 0.483, + 0.151 + ], + "angle": 0, + "content": "\\[\n\\bar {x} _ {k} = \\text {g l o b a l - a v g - p o o l} \\left(x _ {k}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.16, + 0.483, + 0.207 + ], + "angle": 0, + "content": "After that, the high-frequency component is computed as the difference between the \\( x_{k} \\) and \\( \\bar{x}_k \\). The pivotal score \\( s_k \\) is defined as the L2 norm of the high-frequency maps:" + }, + { + "type": "equation", + "bbox": [ + 0.223, + 0.217, + 0.483, + 0.234 + ], + "angle": 0, + "content": "\\[\ns _ {k} = \\left\\| x _ {k} - \\bar {x} _ {k} \\right\\| _ {2}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.243, + 0.484, + 0.365 + ], + "angle": 0, + "content": "Subsequently, we obtain the index set \\(\\mathcal{I}\\) for pivotal token selection by keeping TopK scores in \\(s_k\\). Finally, we perform token pruning on \\(x_{k}\\) through indexing with \\(\\mathcal{I}\\) to allow the fast model forward. Notably, since the PTS reduces the number of input tokens in one transformer layer, as an additional benefit, the KV-cache is also reduced accordingly, thus optimizing the GPU footprints as well as the cross-scale attention for the subsequent scale steps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.365, + 0.484, + 0.577 + ], + "angle": 0, + "content": "Cached Token Restoration. The proposed PTS can effectively reduce the number of tokens. However, the visual generation tasks need to restore the original token numbers to regain the structural 2D image. For this reason, we propose Cached Token Restoration (CTR). The main rationale stems from the diagonal attention sparsity that the token at the \\(k\\)-th scale not only attends to itself, but also exhibits strong correlation with the corresponding slot from previous \\(\\{1,2,\\dots,k-1\\}\\) steps. Therefore, our proposed CTR approximates the layer outputs in the pruned locations using the counterparts from the cached step. Formally, given the cached token map \\(y_{K-N}\\) with size \\((h_{K-N},w_{K-N})\\), which is the output of the corresponding layer at step \\(K-N\\), we first upsample it to the same size of the current scale:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.586, + 0.483, + 0.605 + ], + "angle": 0, + "content": "\\[\ny _ {k} ^ {\\text {c a c h e}} = \\operatorname {i n t e r p o l a t e} \\left(y _ {K - N}, \\left(h _ {k}, w _ {k}\\right)\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.689 + ], + "angle": 0, + "content": "Then the value from \\( y_{k}^{cache} \\) is scattered into the pruned slots of \\( y_{k} \\) using the index set \\( \\mathcal{I} \\), to generate the restored layer output \\( y_{k}^{\\prime} \\), which shares the same token numbers as the \\( x_{k} \\). We also provide an explanation for the choice of \\( K - N \\) as the caching step, see Suppl. for more details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Implementation of FastVAR. Thanks to the token number invariance of the \\(\\mathrm{PTS + CTR}\\), as shown in Fig. 4(b), we apply this operator pair for each Attention&FFN layer at the large-scale steps. Furthermore, we empirically find in Sec. 4.3 that the larger scale steps exhibit stronger robustness to pruning than that of early scale steps. Therefore, we develop a progressive pruning ratio schedule, in which we assign larger pruning ratios to larger scale steps in the set \\(\\mathcal{T}\\). We summarize the complete algorithm of FastVAR in the Suppl.. At last, it is worth noting that our FastVAR does not access the attention map, and we show in Sec. 4.4 that the proposed FastVAR can be used in combination with other acceleration techniques like FlashAttention [12, 13] for even faster VAR generation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.646, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.116, + 0.704, + 0.133 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.139, + 0.907, + 0.32 + ], + "angle": 0, + "content": "Models and Evaluations. We apply our FastVAR on two VAR-based text-to-image models, namely HART [48] and Infinity [22]. Both models can generate images of up to \\(1024 \\times 1024\\) resolution. For a fair comparison, we keep all the hyperparameters the same as their default settings. For evaluation metrics, we compare both in terms of generation quality and inference efficiency. For generation quality, we use two popular benchmarks, i.e., the GenEval [20] and MJHQ30K [29] to validate the high-level semantic consistency and the perceptual quality, respectively. For efficiency evaluation, we adopt metrics including running time, throughputs, speedup ratio, and GPU memory costs." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.322, + 0.909, + 0.609 + ], + "angle": 0, + "content": "Implementation Details. For the number of pruned scale steps \\(N\\), we set \\(N = 4\\) for Infinity and \\(N = 2\\) for HART, i.e., only token maps from the last 4 or 2 scale steps are applied with the proposed FastVAR, with other steps kept as standard VAR. For the progressive pruning ratio schedule, we set it to \\(\\{40\\%, 50\\%, 100\\%, 100\\}\\) for Infinity, and \\(\\{50\\%, 75\\%\\}\\) for HART. The \\(100\\%\\) pruning ratio indicates all tokens are dropped, i.e., we skip the corresponding steps and interpolate the intermediate outputs to the target resolution as the final outputs. Note that this extreme ratio depends on the pruned backbones, and we provide further discussion in the Suppl.. Unless specified, the efficiency of the unpruned baselines are already accelerated by FlashAttention [13], and we apply the proposed FastVAR on top of it. Following the setup of existing methods [10, 47, 50], the inference speeds for all methods are measured without including the VAE time cost since this is a shared cost for all methods. All experiments are conducted on one single NVIDIA RTX 3090 GPU with 24GB memory." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.621, + 0.654, + 0.636 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Comparison on GenEval. We first evaluate the quality-efficiency trade-off on \\(1024 \\times 1024\\) text-to-image generation using the GenEval benchmark [20]. We compare our FastVAR against various state-of-the-art methods, including Diffusion, AR, and VAR models. The results are shown in Tab. 1. Compared with traditional AR models, our Infinity+FastVAR can achieve even \\(39.7 \\times\\) speedup than LlamaGen [47], while achieving \\(125\\%\\) performance boosts on the GenEval. Moreover, our FastVAR achieves acceleration almost without performance loss, e.g., \\(1.5 \\times\\) speedup on the HART with the same GenEval score as the unpruned baseline. Our approach also reduces the inference GPU memory. For example, FastVAR achieves \\(22.2\\%\\) reduction to 14.7GB on top of FlashAttention of 18.9GB, to produce a \\(1024 \\times 1024\\) image in 0.95s, facilitating generation on consumer-level GPUs. The above results demonstrate the generality and effectiveness of our method." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19015" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison on efficiency and quality on \\( {1024} \\times {1024} \\) GenEval benchmarks. The marks \\( \\bullet \\) , \\( \\circ \\) ,and \\( \\diamond \\) denote the Diffusion,AR,and VAR-based methods,respectively. Note that the efficiency of HART and Infinity baselines are tested under FlashAttention." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.122, + 0.9, + 0.289 + ], + "angle": 0, + "content": "
MethodsInference EfficiencyGeneration Quality
#Steps↓Speedup↑Latency↓Throughput↑#Param↓two_object↑position↑color_attr↑Overall↑
SDXL [41]40-4.3s0.23it/s2.6B0.740.150.230.55
PixArt-Sigma [8]20-2.7s2.50it/s0.6B0.620.140.270.55
SD3-medium [16]28-4.4s3.45it/s2.0B0.740.340.360.62
○LlamaGen [47]1024-37.7s0.03it/s0.8B0.340.070.040.32
○Show-o [59]1024-50.3s0.02it/s1.3B0.800.310.500.68
HART [48]141.0×0.95s1.05it/s0.7B0.620.130.180.51
+FastVAR141.5×0.63s1.59it/s0.7B0.570.160.240.51
Infinity [22]131.0×2.61s0.38it/s2.0B0.850.440.530.73
+FastVAR132.7×0.95s1.05it/s2.0B0.810.450.520.72
" + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.298, + 0.898, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.609, + 0.908, + 0.639 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison between the original baselines and our proposed FastVAR on \\(1024 \\times 1024\\) image generation. Our FastVAR achieves significant speedups, e.g., \\(2.7 \\times\\) on Infinity [22], while keeping high-quality results similar to the original model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.653, + 0.483, + 0.835 + ], + "angle": 0, + "content": "Comparison on MJHQ30K. In Tab. 2, we further validate the perceptual quality on the MJHQ30K [29] benchmark. It can be seen that our FastVAR achieves a reasonable performance while maintaining a high speedup ratio. For instance, on the well-known challenging \"people\" category, our HART+FastVAR even achieves a FID reduction of 2.42 with \\(1.5 \\times\\) acceleration. In other categories, our method also maintains good performance with significant acceleration. We also give a qualitative visualization in Fig. 5, and one can see that the generated images from FastVAR maintain exceptionally high quality and accurate semantic information, indicating the effectiveness of the proposed method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Difference from ToMe. Token Merging (ToMe) [3], which merges multiple tokens into one token for efficient forward followed by token unmerging to restore the 2D shape, appears applicable to accelerate VAR backbones due to its to-" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.651, + 0.907, + 0.679 + ], + "angle": 0, + "content": "Table 2. Quantitative comparisons of FID and CLIP score with different generation categories on the MJHQ30K dataset." + }, + { + "type": "table", + "bbox": [ + 0.529, + 0.68, + 0.892, + 0.786 + ], + "angle": 0, + "content": "
MethodsSpeeduplandscapepeoplefood
FID↓CLIP↑FID↓CLIP↑FID↓CLIP↑
SDXL [41]-30.7826.3535.5628.0135.2627.98
HART [48]1.0×25.4326.8230.6128.4730.3728.03
+ FastVAR1.5×22.5226.5128.1928.3430.9728.25
Infinity [22]1.0×24.6826.6230.2727.8231.5526.66
+ FastVAR2.7×24.6826.6230.5528.2832.5427.08
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.903 + ], + "angle": 0, + "content": "ken number invariance. In Tab. 3, we set different FastVAR and ToMe variants using varying pruning ratios. One can see that ToMe fails to achieve high speedup. For example, even the \\(1.36 \\times\\) speedup can lead to noticeable FID degradation since it is difficult to compress the whole token map into limited tokens. In contrast, our FastVAR can achieve \\(1.7 \\times\\) higher speedup with better FID performance." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19016" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.089, + 0.303, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.089, + 0.523, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.089, + 0.892, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.221, + 0.908, + 0.25 + ], + "angle": 0, + "content": "Figure 6. Quantitative and qualitative ablation experiments on different pruning ratios. We set varying pruning ratios on the last scale step of the HART model while keeping the other scale steps as the original setup." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.263, + 0.483, + 0.292 + ], + "angle": 0, + "content": "Table 3. Quantitative comparison results with ToMe [2] under different speedup settings." + }, + { + "type": "table", + "bbox": [ + 0.11, + 0.292, + 0.468, + 0.376 + ], + "angle": 0, + "content": "
MethodsSpeedup↑Latency↓Throughput↑FID↓CLIP↑GenEval↑
HART [48]1.00×0.95s1.05it/s30.6128.470.51
+ ToMe [2]1.19×0.80s1.25it/s29.0728.330.48
+ ToMe [2]1.36×0.70s1.43it/s35.2227.990.46
+ FastVAR1.51×0.63s1.59it/s28.1928.340.51
+ FastVAR1.70×0.56s1.79it/s28.9728.240.50
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.391, + 0.484, + 0.448 + ], + "angle": 0, + "content": "Table 4. Ablation experiments on the scale-wise sensitivity. We focus on the scale range [16,21,27,36,48,64] with every two consecutive scales as a pruning group. The FastVAR is applied to the selected group while keeping the others unpruned." + }, + { + "type": "table", + "bbox": [ + 0.108, + 0.449, + 0.466, + 0.541 + ], + "angle": 0, + "content": "
scalesratioSpeedup↑Latency↓FID↓CLIP↑GenEval↑
baseline01.00×0.95s30.6128.470.51
[16, 21]75%1.01×0.94s34.2728.430.48
[27, 36]50%1.09×0.87s27.9228.490.51
[27, 36]75%1.14×0.83s29.3828.560.50
[48, 64]50%1.30×0.73s27.1328.430.51
[48, 64]75%1.64×0.58s28.5628.360.50
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.56, + 0.255, + 0.575 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.583, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Different Pruning Ratios. The pruning ratio plays an important role in balancing between efficiency and performance. Here, we conduct ablation experiments to investigate the impact of different pruning ratios. Both the quantitative and qualitative results are given in Fig. 6. Intuitively, increasing the pruning ratio leads to a stable runtime reduction since the model only needs to process fewer tokens. However, an over-large pruning ratio can also bring performance degradation due to the information loss of pivotal high-frequency tokens. Furthermore, the quantitative visualization suggests that increasing the pruning ratio makes certain textures and details discontinuous since the cache from the previous scale is not always an optimal approximation of the pruned tokens. Therefore, a moderate pruning ratio, e.g., \\(40\\% -75\\%\\), is adopted for HART+FastVAR to strike a sweet spot between performance and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Scale-wise Sensitivity. As discussed in Sec. 3.2, we only focus on the token pruning of the last few scale steps. Here, we conduct ablation experiments in Tab. 4 to justify the rationality. It can be seen that the acceleration from pruning at an early scale steps is very limited due to the small token" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.263, + 0.9, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.337, + 0.9, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.401, + 0.892, + 0.416 + ], + "angle": 0, + "content": "Figure 7. Visualization of pruned tokens with different ratios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.432, + 0.907, + 0.583 + ], + "angle": 0, + "content": "map size. For example, even reducing \\(50\\%\\) tokens at the 48 and 64 scale steps is significantly faster than that of \\(75\\%\\) at 16 and 21 scale steps. In addition, pruning at early scales also leads to a noticeable performance drop on both perceptual quality and semantic consistency. For instance, pruning at 48 and 64 scale steps brings 5.71 lower FID compared to that of 16 and 21 scale steps. This is because token reduction at small-scale steps disrupts the structure construction stage, and the inaccurate subject structure can exacerbate errors in subsequent scale steps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.909, + 0.795 + ], + "angle": 0, + "content": "Visualization of Pruned Slots. In Fig. 7, we visualize the selected TopK pivotal tokens under different pruning ratios. For a given pruning ratio, e.g., \\(50\\%\\), our FastVAR prioritizes the retention of high-frequency edges or texture slots, and removes tokens from flat regions that have already converged at early scale steps. Furthermore, the importance order within the selected tokens is also meaningful. For example, in the first row, when increasing the pruning ratio from \\(50\\%\\) to \\(90\\%\\), the human eyes, hair and mouth regions, which are the most detailed, are consistently preserved while the low-frequency cheek regions are gradually pruned. Given that our FastVAR is gradient-free, this quantitative result confirms that the proposed frequency-based pivotal token selection is a reasonable heuristic." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.804, + 0.634, + 0.818 + ], + "angle": 0, + "content": "4.4. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In Conjunction with FlashAttention. Existing token pruning methods [18, 19, 58, 63, 65] often depend on the attention map for token importance, leading to difficulties in combination with FlashAttention [13] in which the attention map is inaccessible. In comparison, our FastVAR" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19017" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.09, + 0.315, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.317, + 0.089, + 0.603, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.605, + 0.089, + 0.749, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.752, + 0.089, + 0.899, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.315, + 0.7, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.315, + 0.898, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.518, + 0.908, + 0.547 + ], + "angle": 0, + "content": "Figure 8. We extend the scale schedule of pre-trained VAR model [22] for zero-shot larger resolution generation. We chose a modest extended scale since a larger one is observed quality drop due to the resolution gap. The images are re-scaled for better presentation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.561, + 0.483, + 0.682 + ], + "angle": 0, + "content": "can be seamlessly integrated with FlashAttention to facilitate more efficient inference. In Tab. 5, we show that using the FastVAR-only can obtain even \\(2.1 \\times\\) speedup against FlashAttention only, with \\(20.4\\%\\) GPU memory reduction. After combining both, the resulting version leads to \\(2.7 \\times\\) speedup than FlashAttention only. This experiment validates that our FastVAR can be orthogonally combined with other methods for further acceleration." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Towards Larger Resolution. As presented in Sec. 3.2, the original VAR struggles to scale to higher-resolution due to the increasing token numbers. Our FastVAR can facilitate efficient resolution scaling, offering the potential to generate images at a larger resolution. To this end, we apply our FastVAR to the Infinity [22] for zero-shot larger resolution synthesis by appending additional steps. The results are shown in Fig. 8. It is noteworthy that even the FlashAttention accelerated baseline is out of memory on a 24GB NVIDIA 3090 GPU. In contrast, our FastVAR costs only 15GB memory and 1.3s runtime to generate a \\(1344 \\times 1344\\) image. Moreover, the generated images are of high quality, suggesting that our FastVAR can facilitate production-level image generation on consumer-level GPUs." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.56, + 0.907, + 0.588 + ], + "angle": 0, + "content": "Table 5. Efficiency comparison with FlashAttn on \\(1024\\times 1024\\) image generation." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.589, + 0.898, + 0.663 + ], + "angle": 0, + "content": "
setupSpeedup↑Latency↓Throughput↑Memory↓
SlowAttn [53]---OOM
FlashAttn only [13]1.0×2.61s0.38it/s16.1GB
SlowAttn+FastVAR2.1×1.25s0.80it/s12.8GB
FlashAttn+FastVAR2.7×0.95s1.05it/s11.9GB
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.678, + 0.634, + 0.694 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.704, + 0.907, + 0.901 + ], + "angle": 0, + "content": "This work presents FastVAR to advance high-resolution image synthesis with VARs by addressing the resolution scaling challenge. We reveal that the main latency bottleneck is the large-scale steps at which the low-frequency tokens have almost converged. To this end, we develop the cached token pruning technique to allow only process pivotal high-frequency tokens and use caches from the previous scale steps to restore the original token numbers. Thanks to the reduced forwarded tokens, our FastVAR approximates linear complexity, further enabling generation at a larger resolution such as 2K. Extensive experiments validate our FastVAR as an effective and generic solution for efficient resolution scaling of VAR models." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19018" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.484, + 0.177 + ], + "angle": 0, + "content": "This work is supported in part by the National Natural Science Foundation of China, under Grant (62302309, 62171248), Shenzhen Science and Technology Program (JCYJ20220818101014030, JCYJ20220818101012025)." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.192, + 0.188, + 0.209 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.217, + 0.484, + 0.287 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. GPT-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.289, + 0.484, + 0.317 + ], + "angle": 0, + "content": "[2] Daniel Bolya and Judy Hoffman. Token merging for fast stable diffusion. In CVPR, pages 4599-4603, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.318, + 0.484, + 0.373 + ], + "angle": 0, + "content": "[3] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token Merging: Your ViT but faster. arXiv preprint arXiv:2210.09461, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.376, + 0.484, + 0.444 + ], + "angle": 0, + "content": "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 33:1877-1901, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.446, + 0.484, + 0.502 + ], + "angle": 0, + "content": "[5] Tianle Cai, Yuhong Li, Zhengyang Geng, Hongwu Peng, Jason D Lee, Deming Chen, and Tri Dao. Medusa: Simple LLM inference acceleration framework with multiple decoding heads. arXiv preprint arXiv:2401.10774, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.504, + 0.484, + 0.573 + ], + "angle": 0, + "content": "[6] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.575, + 0.484, + 0.645 + ], + "angle": 0, + "content": "[7] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. PixArt-Alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.647, + 0.484, + 0.716 + ], + "angle": 0, + "content": "[8] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. PixArt-Sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. In ECCV, pages 74-91. Springer, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.718, + 0.484, + 0.771 + ], + "angle": 0, + "content": "[9] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In ICML, pages 1691-1703. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.774, + 0.484, + 0.829 + ], + "angle": 0, + "content": "[10] Zigeng Chen, Xinyin Ma, Gongfan Fang, and Xinchao Wang. Collaborative decoding makes visual auto-regressive modeling efficient. arXiv preprint arXiv:2411.17787, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[11] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instructionfinetuned language models. Journal of Machine Learning Research, 25(70):1-53, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.217, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "[12] Tri Dao. FlashAttention-2: Faster attention with better parallelism and work partitioning. In ICLR, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.124, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[13] Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flash Attention: Fast and memory-efficient exact attention with IO-awareness. In NeurIPS, 2022. 1, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.181, + 0.905, + 0.235 + ], + "angle": 0, + "content": "[14] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In ACL, pages 4171–4186, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.239, + 0.905, + 0.28 + ], + "angle": 0, + "content": "[15] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, pages 12873-12883, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.282, + 0.905, + 0.351 + ], + "angle": 0, + "content": "[16] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.354, + 0.904, + 0.382 + ], + "angle": 0, + "content": "[17] Gongfan Fang, Xinyin Ma, and Xinchao Wang. Structural pruning for diffusion models. In NeurIPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.384, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[18] Qichen Fu, Minsik Cho, Thomas Merth, Sachin Mehta, Mohammad Rastegari, and Mahyar Najibi. LazyLLM: Dynamic token pruning for efficient long context llm inference. arXiv preprint arXiv:2407.14057, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[19] Suyu Ge, Yunan Zhang, Liyuan Liu, Minjia Zhang, Jiawei Han, and Jianfeng Gao. Model tells you what to discard: Adaptive KV cache compression for LLMs. arXiv preprint arXiv:2310.01801, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.499, + 0.905, + 0.541 + ], + "angle": 0, + "content": "[20] Dhruba Ghosh, Hannaneh Hajishirzi, and Ludwig Schmidt. GenEval: An object-focused framework for evaluating text-to-image alignment. NeurIPS, 36, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.543, + 0.905, + 0.585 + ], + "angle": 0, + "content": "[21] Hang Guo, Yawei Li, Tao Dai, Shu-Tao Xia, and Luca Benini. Intlora: Integral low-rank adaptation of quantized diffusion models. arXiv preprint arXiv:2410.21759, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.587, + 0.905, + 0.655 + ], + "angle": 0, + "content": "[22] Jian Han, Jinlai Liu, Yi Jiang, Bin Yan, Yuqi Zhang, Zehuan Yuan, Bingyue Peng, and Xiaobing Liu. Infinity: Scaling bitwise autoregressive modeling for high-resolution image synthesis. arXiv preprint arXiv:2412.04431, 2024. 1, 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.657, + 0.905, + 0.714 + ], + "angle": 0, + "content": "[23] Yefei He, Feng Chen, Yuanyu He, Shaoxuan He, Hong Zhou, Kaipeng Zhang, and Bohan Zhuang. ZIPAR: Accelerating autoregressive image generation through spatial locality. arXiv preprint arXiv:2412.04062, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.716, + 0.905, + 0.785 + ], + "angle": 0, + "content": "[24] Tom Henighan, Jared Kaplan, Mor Katz, Mark Chen, Christopher Hesse, Jacob Jackson, Heewoo Jun, Tom B Brown, Prafulla Dhariwal, Scott Gray, et al. Scaling laws for autoregressive generative modeling. arXiv preprint arXiv:2010.14701, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. *NeurIPS*, 33:6840-6851, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[26] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19019" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[27] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. In CVPR, pages 11523-11532, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.148, + 0.482, + 0.189 + ], + "angle": 0, + "content": "[28] Yaniv Leviathan, Matan Kalman, and Yossi Matias. Fast inference from transformers via speculative decoding. In ICML, pages 19274-19286. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[29] Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Linmiao Xu, and Suhail Doshi. Playground v2.5: Three insights towards enhancing aesthetic quality in text-to-image generation. arXiv preprint arXiv:2402.17245, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.482, + 0.3 + ], + "angle": 0, + "content": "[30] Senmao Li, Taihang Hu, Fahad Shahbaz Khan, Linxuan Li, Shiqi Yang, Yaxing Wang, Ming-Ming Cheng, and Jian Yang. Faster Diffusion: Rethinking the role of unet encoder in diffusion models. NeurIPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.482, + 0.355 + ], + "angle": 0, + "content": "[31] Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization. arXiv preprint arXiv:2406.11838, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.358, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[32] Xiuyu Li, Yijiang Liu, Long Lian, Huanrui Yang, Zhen Dong, Daniel Kang, Shanghang Zhang, and Kurt Keutzer. Q-Diffusion: Quantizing diffusion models. In ICCV, pages 17535-17545, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[33] Dongyang Liu, Shitian Zhao, Le Zhuo, Weifeng Lin, Yu Qiao, Hongsheng Li, and Peng Gao. Lumina-mgpt: Illuminate flexible photorealistic text-to-image generation with multimodal generative pretraining. arXiv preprint arXiv:2408.02657, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.551 + ], + "angle": 0, + "content": "[34] Haozhe Liu, Wentian Zhang, Jinheng Xie, Francesco Faccio, Mengmeng Xu, Tao Xiang, Mike Zheng Shou, Juan-Manuel Perez-Rua, and Jurgen Schmidhuber. Faster diffusion via temporal attention decomposition. arXiv preprint arXiv:2404.02747, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.62 + ], + "angle": 0, + "content": "[35] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.482, + 0.676 + ], + "angle": 0, + "content": "[36] Xinyin Ma, Gongfan Fang, Michael Bi Mi, and Xinchao Wang. Learning-to-Cache: Accelerating diffusion transformer via layer caching. arXiv preprint arXiv:2406.01733, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[37] Xinyin Ma, Gongfan Fang, and Xinchao Wang. DeepCache: Accelerating diffusion models for free. In CVPR, pages 15762-15772, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.482, + 0.775 + ], + "angle": 0, + "content": "[38] Yue Ma, Yingqing He, Xiaodong Cun, Xintao Wang, Siran Chen, Xiu Li, and Qifeng Chen. Follow your pose: Pose-guided text-to-video generation using pose-free videos. In AAAI, pages 4117-4125, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.777, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[39] Yue Ma, Hongyu Liu, Hongfa Wang, Heng Pan, Yingqing He, Junkun Yuan, Ailing Zeng, Chengfei Cai, Heung-Yeung Shum, Wei Liu, et al. Follow-your-emoji: Fine-controllable and expressive freestyle portrait animation. In SIGGRAPH Asia, pages 1–12, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[40] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. In CVPR, pages 14297-14306, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[41] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[42] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with VQ-VAE-2. NeurIPS, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.208, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[43] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.265, + 0.905, + 0.307 + ], + "angle": 0, + "content": "[44] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.31, + 0.905, + 0.378 + ], + "angle": 0, + "content": "[45] Andrea Santilli, Silvio Severino, Emilian Postolache, Valentino Maiorca, Michele Mancusi, Riccardo Marin, and Emanuele Rodola. Accelerating transformer inference for translation via parallel decoding. arXiv preprint arXiv:2305.10427, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.381, + 0.905, + 0.422 + ], + "angle": 0, + "content": "[46] Yuzhang Shang, Zhihang Yuan, Bin Xie, Bingzhe Wu, and Yan Yan. Post-training quantization on diffusion models. In CVPR, pages 1972-1981, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.424, + 0.905, + 0.481 + ], + "angle": 0, + "content": "[47] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.483, + 0.905, + 0.551 + ], + "angle": 0, + "content": "[48] Haotian Tang, Yecheng Wu, Shang Yang, Enze Xie, Junsong Chen, Junyu Chen, Zhuoyang Zhang, Han Cai, Yao Lu, and Song Han. HART: Efficient visual generation with hybrid autoregressive transformer. arXiv preprint arXiv:2410.10812, 2024. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.905, + 0.622 + ], + "angle": 0, + "content": "[49] Yao Teng, Han Shi, Xian Liu, Xuefei Ning, Guohao Dai, Yu Wang, Zhenguo Li, and Xihui Liu. Accelerating autoregressive text-to-image generation with training-free speculative jacobi decoding. arXiv preprint arXiv:2410.01699, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.625, + 0.905, + 0.68 + ], + "angle": 0, + "content": "[50] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. NeurIPS, 2024. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.683, + 0.905, + 0.725 + ], + "angle": 0, + "content": "[51] Aaron Van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Conditional image generation with pixelCNN decoders. NeurIPS, 29, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.727, + 0.905, + 0.755 + ], + "angle": 0, + "content": "[52] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. NeurIPS, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.758, + 0.905, + 0.784 + ], + "angle": 0, + "content": "[53] A Vaswani. Attention is all you need. NeurIPS, 2017. 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.787, + 0.905, + 0.842 + ], + "angle": 0, + "content": "[54] Hongjie Wang, Difan Liu, Yan Kang, Yijun Li, Zhe Lin, Niraj K Jha, and Yuchen Liu. Attention-driven training-free efficiency enhancement of diffusion models. In CVPR, pages 16080-16089, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[55] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. EMU3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19020" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[56] Chengyue Wu, Xiaokang Chen, Zhiyu Wu, Yiyang Ma, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, Chong Ruan, et al. Janus: Decoupling visual encoding for unified multimodal understanding and generation. arXiv preprint arXiv:2410.13848, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[57] Yecheng Wu, Zhuoyang Zhang, Junyu Chen, Haotian Tang, Dacheng Li, Yunhao Fang, Ligeng Zhu, Enze Xie, Hongxu Yin, Li Yi, et al. VILA-U: A unified foundation model integrating visual understanding and generation. arXiv preprint arXiv:2409.04429, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[58] Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. Efficient streaming language models with attention sinks. arXiv preprint arXiv:2309.17453, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[59] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.361, + 0.482, + 0.415 + ], + "angle": 0, + "content": "[60] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.418, + 0.482, + 0.485 + ], + "angle": 0, + "content": "[61] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved VQGAN. arXiv preprint arXiv:2110.04627, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.488, + 0.482, + 0.543 + ], + "angle": 0, + "content": "[62] Evelyn Zhang, Bang Xiao, Jiayi Tang, Qianli Ma, Chang Zou, Xuefei Ning, Xuming Hu, and Linfeng Zhang. Token pruning for caching better: 9 times acceleration on stable diffusion for free. arXiv preprint arXiv:2501.00375, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.545, + 0.482, + 0.613 + ], + "angle": 0, + "content": "[63] Zhenyu Zhang, Ying Sheng, Tianyi Zhou, Tianlong Chen, Lianmin Zheng, Ruisi Cai, Zhao Song, Yuandong Tian, Christopher Ré, Clark Barrett, et al. H2O: Heavy-hitter oracle for efficient generative inference of large language models. NeurIPS, 36:34661-34710, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.616, + 0.482, + 0.668 + ], + "angle": 0, + "content": "[64] Chuanxia Zheng, Tung-Long Vuong, Jianfei Cai, and Dinh Phung. MOVQ: Modulating quantized vectors for high-fidelity image generation. NeurIPS, 35:23412-23425, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.672, + 0.482, + 0.725 + ], + "angle": 0, + "content": "[65] Chang Zou, Xuyang Liu, Ting Liu, Siteng Huang, and Linfeng Zhang. Accelerating diffusion transformers with tokenwise feature caching. arXiv preprint arXiv:2410.05317, 2024. 2, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.725 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19021" + } + ] +] \ No newline at end of file diff --git a/2025/FastVAR_ Linear Visual Autoregressive Modeling via Cached Token Pruning/70725db1-fcd5-4496-a95d-07038d174933_origin.pdf b/2025/FastVAR_ Linear Visual Autoregressive Modeling via Cached Token Pruning/70725db1-fcd5-4496-a95d-07038d174933_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ab65a111bd42434a4625892ed61e625b4378bff6 --- /dev/null +++ b/2025/FastVAR_ Linear Visual Autoregressive Modeling via Cached Token Pruning/70725db1-fcd5-4496-a95d-07038d174933_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2404765172435515ae428ba43617defbe9869a79a73722d83dd1517615dc15be +size 4434409 diff --git a/2025/FastVAR_ Linear Visual Autoregressive Modeling via Cached Token Pruning/full.md b/2025/FastVAR_ Linear Visual Autoregressive Modeling via Cached Token Pruning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3bee4d8fbf1f9a308aef37b66e4772ef6001f774 --- /dev/null +++ b/2025/FastVAR_ Linear Visual Autoregressive Modeling via Cached Token Pruning/full.md @@ -0,0 +1,302 @@ +# FastVAR: Linear Visual Autoregressive Modeling via Cached Token Pruning + +Hang Guo $^{1}$ Yawei Li $^{2,*,†}$ Taolin Zhang $^{1}$ Jiangshan Wang $^{1}$ + +Tao Dai $^{3,*}$ Shu-Tao Xia $^{1,4}$ Luca Benini $^{2}$ + +$^{1}$ Tsinghua University $^{2}$ ETH Zürich $^{3}$ Shenzhen University $^{4}$ Peng Cheng Laboratory + +![](images/f7a8af838d0135eb4524fec8191bba62387690d99ec0844c15a7df93abc6edde.jpg) +Figure 1. Our FastVAR can generate one 2K image using one NVIDIA 3090 GPU, while existing baseline fails due to out of memory. + +# Abstract + +Visual Autoregressive (VAR) modeling has gained popularity for its shift towards next-scale prediction. However, existing VAR paradigms process the entire token map at each scale step, leading to the complexity and runtime scaling dramatically with image resolution. To address this challenge, we propose FastVAR, a post-training acceleration method for efficient resolution scaling with VARs. Our key finding is that the majority of latency arises from the large-scale step where most tokens have already converged. Leveraging this observation, we develop the cached token pruning strategy that only forwards pivotal tokens for scale-specific modeling while using cached tokens from previous scale steps to restore the pruned slots. This significantly reduces the number of forwarded tokens and improves the efficiency at larger resolutions. Experiments show the proposed FastVAR can further speedup FlashAttention-accelerated VAR by $2.7 \times$ with negligible performance drop of $< 1\%$ . We further extend FastVAR to zero-shot generation of higher resolution images. In particular, FastVAR can generate one $2K$ image with 15GB memory footprints in 1.5s on a single NVIDIA 3090 GPU. Code is available at https://github.com/csguoh/FastVAR. + +# 1. Introduction + +The next-token prediction of Autoregressive (AR) models [31, 33, 47, 55, 59] has demonstrated performance com + +petitive with diffusion models for visual generation [47] as well as the potential for unified vision understanding and generation [55-57, 59]. However, this token-by-token paradigm suffers from numerous decoding steps. Recently, Visual Autoregressive (VAR) modeling [50] has shifted the paradigm to next-scale prediction, enabling image generation in fewer steps. Under this new paradigm, some works [22, 48] have developed VAR-based models for text-to-image generation and obtained promising results. + +Despite their potential, existing VAR-based methods [22, 48, 50] face a critical challenge: the computational complexity and runtime latency scale dramatically with image resolution. Specifically, unlike next-token prediction which processes only one token per step, the next-scale prediction of VAR requires processing the entire token map at each decoding step. As a result, the number of tokens increases in $\mathcal{O}(n^2)$ with the image resolution $n\times n$ , and even leads to $\mathcal{O}(n^4)$ complexity in the attention [53] layer. Empirically, as shown in Fig. 2(a), even when FlashAttention [13] is enabled, VAR models still exhibit super-linear runtime latency. Consequently, this significant computational complexity prevents existing VAR-based models from scaling to higher resolutions, such as 2K. + +In this work, we aim to address the resolution scaling challenge of VARs by pruning forwarded tokens. Through an in-depth analysis of the pre-trained VAR models, we identify the following key findings. 1) Large-scale steps are speed bottleneck but appear robustness. Runtime profiling in Fig. 2(a) finds that the last two large-scale steps account + +![](images/1b3a005446838e9bb0555a4a0b5ba28a31911a56df4a362507c535f304c8b8f0.jpg) +(a) Runtime at different scale steps + +![](images/9ff3b0530a79108f977ec72bf362dc257ced8ca70cbd26c0c4c984d31bacdb20.jpg) +(b) Efficiency-performance tradeoff +Figure 2. FastVAR exhibits promising resolution scalability, and can achieve noticeable speedup with negligible performance drop. + +for $60\%$ of the total runtime. Further investigation in Tab. 4 reveals that VAR is more resilient to token pruning at large-scale steps than at smaller scales. Thus, we direct our efforts to the large-scale steps. 2) High-frequency modeling matters at large-scale steps. Spectrum analysis in Fig. 3(b) reveals that the large-scale steps are optimized to model high-frequency tokens, such as texture details, and the remaining low-frequency tokens almost converge in these steps. Consequently, we can only forward the high-frequency tokens for pruning. 3) Tokens from different scales are related. Attention map analysis in Fig. 3(c) shows strong diagonal sparsity across scales, indicating that tokens attend not only to same-scale neighbors but also to those from the preceding scales. Thus, we cache tokens in previous steps to compensate for the pruned slots, preserving the 2D image lattice structure and maintaining information flow. + +Based on the above observations, we propose FastVAR, a post-training acceleration recipe for efficient resolution scaling with VARs. At its core, FastVAR employs "cached token pruning", which retains only pivotal tokens at large-scale steps to reduce computational overhead while using cached token maps from early-scale steps to compensate for information loss. To identify the pivotal tokens, we develop Pivotal Token Selection (PTS), a frequency-based scoring mechanism. Token importance is determined by filtering out the low-frequency component estimated via the direct current component, which enables efficient frequency-based token selection directly in the spatial domain. The selected Top-K pivotal tokens are then processed by the VAR model. To restore pruned tokens, we propose Cached Token Restoration (CTR), which first interpolates the token map from the cached scale and then reinstates the pruned tokens by indexing the interpolated token map to the pruned location. By integrating these strategies, we present FastVAR which enjoys the following benefits: + +- The proposed method is training-free and plug-and-play for various VAR-based backbones. +- As shown in Fig. 2(b), FastVAR can be integrated with FlashAttention, with further $2.7 \times$ speedup and $< 1\%$ performance drop. +- In Fig. 1, FastVAR facilitates zero-shot scaling to larger-resolution and can produce one 2K image using only 15GB memory in 1.5s on a single NVIDIA 3090 GPU. + +# 2. Related Work + +Autoregressive Visual Generation. The previous autoregressive (AR) methods [27, 42, 61, 64] mostly adopt the next-token prediction paradigm, which treats each pixel as one token and generates pixels in a GPT or BERT style [1, 4, 6, 14, 35]. For instance, some pioneering works [9, 51] generate pixels in raster scan order with transformer models. Later, VQVAE [52] and VQGAN [15] propose to quantize image patches into discrete tokens to ease the training process. Benefiting from the scaling up [24, 26], recent works have shown promising results that the autoregressive models can achieve a more competitive performance than state-of-the-art Diffusion models [7, 8, 38, 39, 41, 43]. Despite this progress, existing next-token generation pipelines struggle to efficiently synthesize high-resolution images due to the numerous decoding steps [10]. More recently, the Visual Autoregressive (VAR) modeling [50], which shifts to next-scale generation fashion to allow only a few forward steps, further advances the generation quality as well as the model efficiency. Building on top of VAR, existing methods have scaled up the model for production-level text-to-image generation. For example, HART [48] adopts the continuous diffusion module to compensate for the quantization error. Infinity [22] employs a bitwise tokenizer for an extremely large vector vocabulary. Despite the promising ability of VAR-based methods, they struggle to scale to larger resolutions with the increasing token numbers. + +Efficient Visual Generation. To speed up diffusion models [25], many efforts have been made in recent years, including distillation [40, 44], quantization [21, 32, 46], pruning [2, 17, 54, 62, 65], and caching [30, 34, 36, 37], which either reduce the total denoising steps or to reduce the cost of single forward. Specifically, DeepCache [37] proposes to reuse the intermediate features of low-resolution layers in the U-Net. ToMeSD [2] merges similar tokens into one token with subsequent unmerging for acceleration. Unfortunately, these methods are specifically designed for diffusion, which can achieve only sub-optimal performance or cannot be used in VAR. In the efforts to accelerate AR modeling [31, 33, 47, 55, 59], one prevalent solution is to reduce the forward step using parallel decoding strategies [23, 45, 49]. For instance, speculative decoding [5, 28] utilizes a small draft model to generate candidate tokens, which are then verified in parallel by the larger model. However, the well-developed decoding strategy cannot be directly applied to the next-scale paradigm, whose each step involves multiple pixels [50]. As for fast VAR models, one related work is the CoDe [10], which uses model ensemble techniques to use a small model on the costly large resolution. However, CoDe highly depends on the availability of different-sized models. In this work, we attempt to speed up the VARs by pruning extra tokens, which is training-free and is generic for multiple VAR methods. + +![](images/9d8bdeccafa1ca64e9f1197238f8b3456a726305eeec8e8e90c1bb34b173930d.jpg) + +![](images/493f32c14f90245c102a0a13a3f008567e5e08c10860e4d798a7f2eb4fa4e4e3.jpg) +(a) Visualization of Intermediate Outputs from Different Scale Steps + +![](images/0fa08fb31825a077185edd39ad7f10f0df57ae52c4e69c777996cd656e04713f.jpg) +(b) Fourier Analysis of Large-Scale Steps +Figure 3. (a) We resize different $\tilde{r}_k$ to the same size for better presentation. (b) Each curve represents the frequency characteristics of a certain-scale token map. (c) The row represents keys of scale steps $\{1,2,\dots ,k\}$ , and the column is queries of the $k$ -th scale step. + +![](images/35ccdaaf069943da7325975a09b06757f9d4c75129cd14066a203a9768ffcf96.jpg) +(c) Cross-scale Attention Map Visualization + +![](images/5a5fa61cc374bd3fb16d4ad55eecf03d9eb4bc15bb293975f8120a80c36588cc.jpg) + +![](images/492600e259c2361c9bd4d44256ebf080910d89623d55b7e588ec4836a513edc6.jpg) + +# 3. Method + +# 3.1. Preliminary + +The Visual Autoregressive (VAR) modeling [50] redefines the traditional AR by shifting from a "next-token prediction" to a "next-scale prediction", where each autoregressive unit is a token map of varying scales instead of a single token. For a given image feature map, VAR first quantizes it into $K$ multi-scale token maps $\mathcal{R} = \{r_1, r_2, \dots, r_K\}$ with progressively larger resolutions using pre-defined scale schdual $\{(h_1, w_1), (h_2, w_2), \dots, (h_K, w_K)\}$ . Then the $\mathcal{R}$ is used to train a Transformer [53] to learn the joint distribution using the causal formulation: + +$$ +p \left(r _ {1}, r _ {2}, \dots , r _ {K}\right) = \prod_ {k = 1} ^ {K} p \left(r _ {k} \mid r _ {1}, r _ {2}, \dots , r _ {k - 1}\right), \tag {1} +$$ + +where the start token $r_1$ is the encoded text embedding from the LLMs [11, 60] under given user prompts. During training, existing methods [22, 48, 50] usually employ a residual strategy to reduce learning difficulty. In detail, the model outputs at the $k$ -th scale step $f_k$ is the residual of the intermediate prediction $\tilde{r}_k$ : + +$$ +\tilde {r} _ {k} = \operatorname {i n t e r p o l a t e} \left(\tilde {r} _ {k - 1}, \left(h _ {k}, w _ {k}\right)\right) + f _ {k}, \tag {2} +$$ + +where the "interpolate $(\cdot ,(h_k,w_k))$ denotes upsample a token map to the size of $(h_k,w_k)$ . Unfolding Eq. (2) derives the following cumulative form: + +$$ +\tilde {r} _ {k} = \sum_ {i = 1} ^ {k} \operatorname {i n t e r p o l a t e} \left(f _ {i}, \left(h _ {k}, w _ {k}\right)\right). \tag {3} +$$ + +In the last scale step, the predicted token maps $\tilde{r}_K$ is used to generate the final image. During inference, since the outputs at different scales are generated autoregressively, the KV-Cache can be adopted to avoid re-computing of previous steps. Although the VAR paradigm can generate images within only $K$ scale steps, the Transformer needs to process all $h_k \times w_k$ tokens at the $k$ -th step, hindering scaling to higher-resolution image synthesis. + +# 3.2. Motivation + +In this work, we aim to approach linear VAR models to advance high-resolution image generation. To this end, we investigate the scale-wise behavior of the pre-trained VAR model [22]. We examine from three angles: the latency profiling, the specific role of each step, and the cross-scale token dependency. Conclusions are summarized as follows. + +Observation 1: Large-scale steps are speed bottleneck but appear robustness. As shown in the runtime curve in Fig. 2(a), even the FlashAttention is enabled, the inference latency of the VAR exhibits a super-linear complexity with resolution steps. This is because the Transformer needs to take all the tokens at once. As a result, the tokens from the large-scale steps make up the majority of the total token length. For example, the last two steps even occupy $60\%$ of the total time. On the other hand, the experiments in Sec. 4.3 suggest the large-scale steps are more robust to token drop than small-scale counterparts under the same pruning ratio. Therefore, it is accessible to prune some "unimportant tokens" at the large scale step to speed up token processing. However, challenges still remain about which tokens should be pruned and how to compensate for the information loss from these pruned tokens. + +![](images/24e5cc3c3e7b46970a2922892723195dab6692937a52949043f0f5fe1c5ba1ae.jpg) +(a) Overall pipeline of the proposed FastVAR + +![](images/8d36b4ffa8b04f161b459102072d2324e7ff1b95e97aa8a295b4acc60ecc1d20.jpg) +(b) Token-wise implementation for one VAR layer at stage $\mathcal{T}$ +Figure 4. (a) We keep the Structure Construction Stage as standard VAR and additionally store token maps at the $(K - N)$ -th step as the cache. Then the Texture Filling Stage applies FastVAR to process only pivotal tokens for fast inference. (b) For Attention or FFN layers using FastVAR, we first employ the Pivotal Token Selection (PTS) for token reduction. After the model forward, we develop the Cached Token Restoration (CTR) to restore the original token numbers using the caches from the previous scale step. + +Observation 2: High-frequency modeling matters at large-scale steps. To find which token should be pruned, we visualize the intermediate predictions $\tilde{r}_k$ in Fig. 3(a). Interestingly, we find that the generation process of the pretrained VAR model can be decomposed into two stages. The Structure Construction Stage, which includes small-scale steps, is trained to generate the outline of a subject. After that, at the Texture Filling Stage, which is composed of large-scale steps, is responsible for adding details based on the previous sketch. Given that the outline is low-frequency while the texture is high-frequency, we therefore infer that the VAR mainly models the high-frequency texture in the large-scale stage while keeping the low-frequency structure almost intact. For further verification, we show in Fig. 3(b) that the low-frequency component almost converges in the large-scale steps, while the high-frequency modeling stays optimized with noticeable variations. Therefore, we can prune the redundant low-frequency tokens and forward only the pivotal high-frequency tokens at the large-scale steps for token number reduction. + +# Observation 3: Tokens from different scales are related. + +After reducing the input tokens for fast forward, we still need to restore the output tokens to the original numbers in order to maintain the 2D image structure. To this end, we analyze the cross-scale attention map of the unpruned VAR model, where the query at the current scale interacts with KV caches from all previous scales. As depicted in Fig. 3(c), we observe that the attention map not only exhibits a high diagonal score at self-scale, but also a high diagonal score at cross-scale. This behavior indicates that a token in the current scale not only attends to its same-scale locals but also exhibits strong correlation to tokens at the neighboring positions in previous scales. This diagonal attention sparsity thus facilitates us to use only a few + +tokens to estimate the original output of the pruned slots, and avoids weighting multiple tokens across the token map. Specifically, we can approximate the outputs of pruned tokens by indexing in the previous-scale token maps based on the pruning location. In this way, the information loss can be compensated thanks to this high cross-scale similarity. + +# 3.3. Efficient Resolution Scaling for VARs + +We instantiate the above idea and present FastVAR to approximate linear VAR models. The core is the cached token pruning, which prunes tokens at large scale steps while using cache from early scales to maintain information flow. + +Overview. As shown in Fig. 4(a), denote the scale step set of Structure Construction Stage as $S = \{1,2,\dots ,K - N\}$ the step set of Texture Filling Stage as $\mathcal{T} = \{K - N + 1,\dots ,K\}$ . As stated in Sec. 3.2, the speedup of pruning token maps in $S$ is insignificant while producing a noticeable performance drop, we therefore keep $S$ intact as the standard VAR. In addition, we save the per-layer outputs at the last scale in $S$ , i.e., the $(K - N)$ -th step, as the cache for subsequent token restoration. For the set $\mathcal{T}$ , we apply FastVAR to prune tokens for fast-forward, followed by the token number restoration. More details are given below. + +Pivotal Token Selection. We first introduce the Pivotal Token Selection (PTS) to reduce the number of tokens. As observed in Fig. 3(b), the low-frequency tokens almost converge in the later scale steps. Therefore, we can feed only the high-frequency tokens and prune the remaining. However, the challenge lies in that common frequency operators, such as FFT, work in the frequency domain, making it difficult to recognize the frequency characteristics of certain tokens in the original token map. To this end, our proposed PTS adopts an approximation solution. Specifically, denote $x_{k}$ as the input of one layer, we estimate its low-frequency + +component $\bar{x}_k$ as the direct current component, which can be easily obtained via the spatially global average pooling: + +$$ +\bar {x} _ {k} = \text {g l o b a l - a v g - p o o l} \left(x _ {k}\right). \tag {4} +$$ + +After that, the high-frequency component is computed as the difference between the $x_{k}$ and $\bar{x}_k$ . The pivotal score $s_k$ is defined as the L2 norm of the high-frequency maps: + +$$ +s _ {k} = \left\| x _ {k} - \bar {x} _ {k} \right\| _ {2}. \tag {5} +$$ + +Subsequently, we obtain the index set $\mathcal{I}$ for pivotal token selection by keeping TopK scores in $s_k$ . Finally, we perform token pruning on $x_{k}$ through indexing with $\mathcal{I}$ to allow the fast model forward. Notably, since the PTS reduces the number of input tokens in one transformer layer, as an additional benefit, the KV-cache is also reduced accordingly, thus optimizing the GPU footprints as well as the cross-scale attention for the subsequent scale steps. + +Cached Token Restoration. The proposed PTS can effectively reduce the number of tokens. However, the visual generation tasks need to restore the original token numbers to regain the structural 2D image. For this reason, we propose Cached Token Restoration (CTR). The main rationale stems from the diagonal attention sparsity that the token at the $k$ -th scale not only attends to itself, but also exhibits strong correlation with the corresponding slot from previous $\{1,2,\dots,k-1\}$ steps. Therefore, our proposed CTR approximates the layer outputs in the pruned locations using the counterparts from the cached step. Formally, given the cached token map $y_{K-N}$ with size $(h_{K-N},w_{K-N})$ , which is the output of the corresponding layer at step $K-N$ , we first upsample it to the same size of the current scale: + +$$ +y _ {k} ^ {\text {c a c h e}} = \operatorname {i n t e r p o l a t e} \left(y _ {K - N}, \left(h _ {k}, w _ {k}\right)\right). \tag {6} +$$ + +Then the value from $y_{k}^{cache}$ is scattered into the pruned slots of $y_{k}$ using the index set $\mathcal{I}$ , to generate the restored layer output $y_{k}^{\prime}$ , which shares the same token numbers as the $x_{k}$ . We also provide an explanation for the choice of $K - N$ as the caching step, see Suppl. for more details. + +Implementation of FastVAR. Thanks to the token number invariance of the $\mathrm{PTS + CTR}$ , as shown in Fig. 4(b), we apply this operator pair for each Attention&FFN layer at the large-scale steps. Furthermore, we empirically find in Sec. 4.3 that the larger scale steps exhibit stronger robustness to pruning than that of early scale steps. Therefore, we develop a progressive pruning ratio schedule, in which we assign larger pruning ratios to larger scale steps in the set $\mathcal{T}$ . We summarize the complete algorithm of FastVAR in the Suppl.. At last, it is worth noting that our FastVAR does not access the attention map, and we show in Sec. 4.4 that the proposed FastVAR can be used in combination with other acceleration techniques like FlashAttention [12, 13] for even faster VAR generation. + +# 4. Experiments + +# 4.1. Experimental Setup + +Models and Evaluations. We apply our FastVAR on two VAR-based text-to-image models, namely HART [48] and Infinity [22]. Both models can generate images of up to $1024 \times 1024$ resolution. For a fair comparison, we keep all the hyperparameters the same as their default settings. For evaluation metrics, we compare both in terms of generation quality and inference efficiency. For generation quality, we use two popular benchmarks, i.e., the GenEval [20] and MJHQ30K [29] to validate the high-level semantic consistency and the perceptual quality, respectively. For efficiency evaluation, we adopt metrics including running time, throughputs, speedup ratio, and GPU memory costs. + +Implementation Details. For the number of pruned scale steps $N$ , we set $N = 4$ for Infinity and $N = 2$ for HART, i.e., only token maps from the last 4 or 2 scale steps are applied with the proposed FastVAR, with other steps kept as standard VAR. For the progressive pruning ratio schedule, we set it to $\{40\%, 50\%, 100\%, 100\}$ for Infinity, and $\{50\%, 75\%\}$ for HART. The $100\%$ pruning ratio indicates all tokens are dropped, i.e., we skip the corresponding steps and interpolate the intermediate outputs to the target resolution as the final outputs. Note that this extreme ratio depends on the pruned backbones, and we provide further discussion in the Suppl.. Unless specified, the efficiency of the unpruned baselines are already accelerated by FlashAttention [13], and we apply the proposed FastVAR on top of it. Following the setup of existing methods [10, 47, 50], the inference speeds for all methods are measured without including the VAE time cost since this is a shared cost for all methods. All experiments are conducted on one single NVIDIA RTX 3090 GPU with 24GB memory. + +# 4.2. Main Results + +Comparison on GenEval. We first evaluate the quality-efficiency trade-off on $1024 \times 1024$ text-to-image generation using the GenEval benchmark [20]. We compare our FastVAR against various state-of-the-art methods, including Diffusion, AR, and VAR models. The results are shown in Tab. 1. Compared with traditional AR models, our Infinity+FastVAR can achieve even $39.7 \times$ speedup than LlamaGen [47], while achieving $125\%$ performance boosts on the GenEval. Moreover, our FastVAR achieves acceleration almost without performance loss, e.g., $1.5 \times$ speedup on the HART with the same GenEval score as the unpruned baseline. Our approach also reduces the inference GPU memory. For example, FastVAR achieves $22.2\%$ reduction to 14.7GB on top of FlashAttention of 18.9GB, to produce a $1024 \times 1024$ image in 0.95s, facilitating generation on consumer-level GPUs. The above results demonstrate the generality and effectiveness of our method. + +Table 1. Quantitative comparison on efficiency and quality on ${1024} \times {1024}$ GenEval benchmarks. The marks $\bullet$ , $\circ$ ,and $\diamond$ denote the Diffusion,AR,and VAR-based methods,respectively. Note that the efficiency of HART and Infinity baselines are tested under FlashAttention. + +
MethodsInference EfficiencyGeneration Quality
#Steps↓Speedup↑Latency↓Throughput↑#Param↓two_object↑position↑color_attr↑Overall↑
SDXL [41]40-4.3s0.23it/s2.6B0.740.150.230.55
PixArt-Sigma [8]20-2.7s2.50it/s0.6B0.620.140.270.55
SD3-medium [16]28-4.4s3.45it/s2.0B0.740.340.360.62
○LlamaGen [47]1024-37.7s0.03it/s0.8B0.340.070.040.32
○Show-o [59]1024-50.3s0.02it/s1.3B0.800.310.500.68
HART [48]141.0×0.95s1.05it/s0.7B0.620.130.180.51
+FastVAR141.5×0.63s1.59it/s0.7B0.570.160.240.51
Infinity [22]131.0×2.61s0.38it/s2.0B0.850.440.530.73
+FastVAR132.7×0.95s1.05it/s2.0B0.810.450.520.72
+ +![](images/ca3246aafd90a0ded831663d6ab6ee845d1373f41b746f038973bc2887e5f6f5.jpg) +Figure 5. Qualitative comparison between the original baselines and our proposed FastVAR on $1024 \times 1024$ image generation. Our FastVAR achieves significant speedups, e.g., $2.7 \times$ on Infinity [22], while keeping high-quality results similar to the original model. + +Comparison on MJHQ30K. In Tab. 2, we further validate the perceptual quality on the MJHQ30K [29] benchmark. It can be seen that our FastVAR achieves a reasonable performance while maintaining a high speedup ratio. For instance, on the well-known challenging "people" category, our HART+FastVAR even achieves a FID reduction of 2.42 with $1.5 \times$ acceleration. In other categories, our method also maintains good performance with significant acceleration. We also give a qualitative visualization in Fig. 5, and one can see that the generated images from FastVAR maintain exceptionally high quality and accurate semantic information, indicating the effectiveness of the proposed method. + +Difference from ToMe. Token Merging (ToMe) [3], which merges multiple tokens into one token for efficient forward followed by token unmerging to restore the 2D shape, appears applicable to accelerate VAR backbones due to its to- + +Table 2. Quantitative comparisons of FID and CLIP score with different generation categories on the MJHQ30K dataset. + +
MethodsSpeeduplandscapepeoplefood
FID↓CLIP↑FID↓CLIP↑FID↓CLIP↑
SDXL [41]-30.7826.3535.5628.0135.2627.98
HART [48]1.0×25.4326.8230.6128.4730.3728.03
+ FastVAR1.5×22.5226.5128.1928.3430.9728.25
Infinity [22]1.0×24.6826.6230.2727.8231.5526.66
+ FastVAR2.7×24.6826.6230.5528.2832.5427.08
+ +ken number invariance. In Tab. 3, we set different FastVAR and ToMe variants using varying pruning ratios. One can see that ToMe fails to achieve high speedup. For example, even the $1.36 \times$ speedup can lead to noticeable FID degradation since it is difficult to compress the whole token map into limited tokens. In contrast, our FastVAR can achieve $1.7 \times$ higher speedup with better FID performance. + +![](images/95d59ba43b5936222d43f5d4b119ef4b9fc8a663b998853ec6e0a833a5ad1c13.jpg) +Figure 6. Quantitative and qualitative ablation experiments on different pruning ratios. We set varying pruning ratios on the last scale step of the HART model while keeping the other scale steps as the original setup. + +![](images/8afb013a4243a68dbfbd5ea9c511585290e78de6618b0c219fbeb6bdc3c136b2.jpg) + +![](images/fe0218ae1536f8b1422222173a463b64a4ccfebc5c3c28d9f407643be9810b49.jpg) + +Table 3. Quantitative comparison results with ToMe [2] under different speedup settings. + +
MethodsSpeedup↑Latency↓Throughput↑FID↓CLIP↑GenEval↑
HART [48]1.00×0.95s1.05it/s30.6128.470.51
+ ToMe [2]1.19×0.80s1.25it/s29.0728.330.48
+ ToMe [2]1.36×0.70s1.43it/s35.2227.990.46
+ FastVAR1.51×0.63s1.59it/s28.1928.340.51
+ FastVAR1.70×0.56s1.79it/s28.9728.240.50
+ +Table 4. Ablation experiments on the scale-wise sensitivity. We focus on the scale range [16,21,27,36,48,64] with every two consecutive scales as a pruning group. The FastVAR is applied to the selected group while keeping the others unpruned. + +
scalesratioSpeedup↑Latency↓FID↓CLIP↑GenEval↑
baseline01.00×0.95s30.6128.470.51
[16, 21]75%1.01×0.94s34.2728.430.48
[27, 36]50%1.09×0.87s27.9228.490.51
[27, 36]75%1.14×0.83s29.3828.560.50
[48, 64]50%1.30×0.73s27.1328.430.51
[48, 64]75%1.64×0.58s28.5628.360.50
+ +# 4.3. Ablation Studies + +Different Pruning Ratios. The pruning ratio plays an important role in balancing between efficiency and performance. Here, we conduct ablation experiments to investigate the impact of different pruning ratios. Both the quantitative and qualitative results are given in Fig. 6. Intuitively, increasing the pruning ratio leads to a stable runtime reduction since the model only needs to process fewer tokens. However, an over-large pruning ratio can also bring performance degradation due to the information loss of pivotal high-frequency tokens. Furthermore, the quantitative visualization suggests that increasing the pruning ratio makes certain textures and details discontinuous since the cache from the previous scale is not always an optimal approximation of the pruned tokens. Therefore, a moderate pruning ratio, e.g., $40\% -75\%$ , is adopted for HART+FastVAR to strike a sweet spot between performance and efficiency. + +Scale-wise Sensitivity. As discussed in Sec. 3.2, we only focus on the token pruning of the last few scale steps. Here, we conduct ablation experiments in Tab. 4 to justify the rationality. It can be seen that the acceleration from pruning at an early scale steps is very limited due to the small token + +![](images/1b68d0fa6560028ce10e49ad4ae2ce9f7fbe612e2597aae4face0ac9d5b69d86.jpg) + +![](images/ba02d2ee399a712ff23a0a09bbbbc5393e13c2ea3edc938e7c8d71cf3ee0c2c6.jpg) +Figure 7. Visualization of pruned tokens with different ratios. + +map size. For example, even reducing $50\%$ tokens at the 48 and 64 scale steps is significantly faster than that of $75\%$ at 16 and 21 scale steps. In addition, pruning at early scales also leads to a noticeable performance drop on both perceptual quality and semantic consistency. For instance, pruning at 48 and 64 scale steps brings 5.71 lower FID compared to that of 16 and 21 scale steps. This is because token reduction at small-scale steps disrupts the structure construction stage, and the inaccurate subject structure can exacerbate errors in subsequent scale steps. + +Visualization of Pruned Slots. In Fig. 7, we visualize the selected TopK pivotal tokens under different pruning ratios. For a given pruning ratio, e.g., $50\%$ , our FastVAR prioritizes the retention of high-frequency edges or texture slots, and removes tokens from flat regions that have already converged at early scale steps. Furthermore, the importance order within the selected tokens is also meaningful. For example, in the first row, when increasing the pruning ratio from $50\%$ to $90\%$ , the human eyes, hair and mouth regions, which are the most detailed, are consistently preserved while the low-frequency cheek regions are gradually pruned. Given that our FastVAR is gradient-free, this quantitative result confirms that the proposed frequency-based pivotal token selection is a reasonable heuristic. + +# 4.4. Discussion + +In Conjunction with FlashAttention. Existing token pruning methods [18, 19, 58, 63, 65] often depend on the attention map for token importance, leading to difficulties in combination with FlashAttention [13] in which the attention map is inaccessible. In comparison, our FastVAR + +![](images/b321a9722cc7bb24a63b7df84e1787542d633788fb0fa5de757c3a20c69497d9.jpg) + +![](images/96941d5de7ea757e77f7404198a6d0d3145848b84ade08f06c3418fa34a01af3.jpg) + +![](images/4cc7f8454fd293b35cfc53f41e94a8b905adbb4baf016501f225b4405feff53c.jpg) + +![](images/6635a24e83fcf8de2d4544eca05daa876b66e9a1d10467262244e5d3cb69d40e.jpg) + +![](images/2e3b1a0d29fff03fb5a775037489a2ba8061355577a03fbc6a62d2b8577df01a.jpg) +Figure 8. We extend the scale schedule of pre-trained VAR model [22] for zero-shot larger resolution generation. We chose a modest extended scale since a larger one is observed quality drop due to the resolution gap. The images are re-scaled for better presentation. + +![](images/6bff136e74124efb6b958946717949c6dcc432b26b39c279cb7df56138b7edb6.jpg) + +can be seamlessly integrated with FlashAttention to facilitate more efficient inference. In Tab. 5, we show that using the FastVAR-only can obtain even $2.1 \times$ speedup against FlashAttention only, with $20.4\%$ GPU memory reduction. After combining both, the resulting version leads to $2.7 \times$ speedup than FlashAttention only. This experiment validates that our FastVAR can be orthogonally combined with other methods for further acceleration. + +Towards Larger Resolution. As presented in Sec. 3.2, the original VAR struggles to scale to higher-resolution due to the increasing token numbers. Our FastVAR can facilitate efficient resolution scaling, offering the potential to generate images at a larger resolution. To this end, we apply our FastVAR to the Infinity [22] for zero-shot larger resolution synthesis by appending additional steps. The results are shown in Fig. 8. It is noteworthy that even the FlashAttention accelerated baseline is out of memory on a 24GB NVIDIA 3090 GPU. In contrast, our FastVAR costs only 15GB memory and 1.3s runtime to generate a $1344 \times 1344$ image. Moreover, the generated images are of high quality, suggesting that our FastVAR can facilitate production-level image generation on consumer-level GPUs. + +Table 5. Efficiency comparison with FlashAttn on $1024\times 1024$ image generation. + +
setupSpeedup↑Latency↓Throughput↑Memory↓
SlowAttn [53]---OOM
FlashAttn only [13]1.0×2.61s0.38it/s16.1GB
SlowAttn+FastVAR2.1×1.25s0.80it/s12.8GB
FlashAttn+FastVAR2.7×0.95s1.05it/s11.9GB
+ +# 5. Conclusion + +This work presents FastVAR to advance high-resolution image synthesis with VARs by addressing the resolution scaling challenge. We reveal that the main latency bottleneck is the large-scale steps at which the low-frequency tokens have almost converged. To this end, we develop the cached token pruning technique to allow only process pivotal high-frequency tokens and use caches from the previous scale steps to restore the original token numbers. Thanks to the reduced forwarded tokens, our FastVAR approximates linear complexity, further enabling generation at a larger resolution such as 2K. Extensive experiments validate our FastVAR as an effective and generic solution for efficient resolution scaling of VAR models. + +# Acknowledgments + +This work is supported in part by the National Natural Science Foundation of China, under Grant (62302309, 62171248), Shenzhen Science and Technology Program (JCYJ20220818101014030, JCYJ20220818101012025). + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. GPT-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2 +[2] Daniel Bolya and Judy Hoffman. Token merging for fast stable diffusion. In CVPR, pages 4599-4603, 2023. 2, 7 +[3] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token Merging: Your ViT but faster. arXiv preprint arXiv:2210.09461, 2022. 6 +[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 33:1877-1901, 2020. 2 +[5] Tianle Cai, Yuhong Li, Zhengyang Geng, Hongwu Peng, Jason D Lee, Deming Chen, and Tri Dao. Medusa: Simple LLM inference acceleration framework with multiple decoding heads. arXiv preprint arXiv:2401.10774, 2024. 2 +[6] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 2 +[7] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. PixArt-Alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 2 +[8] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. PixArt-Sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. In ECCV, pages 74-91. Springer, 2024. 2, 6 +[9] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In ICML, pages 1691-1703. PMLR, 2020. 2 +[10] Zigeng Chen, Xinyin Ma, Gongfan Fang, and Xinchao Wang. Collaborative decoding makes visual auto-regressive modeling efficient. arXiv preprint arXiv:2411.17787, 2024. 2, 5 +[11] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instructionfinetuned language models. Journal of Machine Learning Research, 25(70):1-53, 2024. 3 + +[12] Tri Dao. FlashAttention-2: Faster attention with better parallelism and work partitioning. In ICLR, 2024. 5 +[13] Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flash Attention: Fast and memory-efficient exact attention with IO-awareness. In NeurIPS, 2022. 1, 5, 7, 8 +[14] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In ACL, pages 4171–4186, 2019. 2 +[15] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, pages 12873-12883, 2021. 2 +[16] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 6 +[17] Gongfan Fang, Xinyin Ma, and Xinchao Wang. Structural pruning for diffusion models. In NeurIPS, 2023. 2 +[18] Qichen Fu, Minsik Cho, Thomas Merth, Sachin Mehta, Mohammad Rastegari, and Mahyar Najibi. LazyLLM: Dynamic token pruning for efficient long context llm inference. arXiv preprint arXiv:2407.14057, 2024. 7 +[19] Suyu Ge, Yunan Zhang, Liyuan Liu, Minjia Zhang, Jiawei Han, and Jianfeng Gao. Model tells you what to discard: Adaptive KV cache compression for LLMs. arXiv preprint arXiv:2310.01801, 2023. 7 +[20] Dhruba Ghosh, Hannaneh Hajishirzi, and Ludwig Schmidt. GenEval: An object-focused framework for evaluating text-to-image alignment. NeurIPS, 36, 2024. 5 +[21] Hang Guo, Yawei Li, Tao Dai, Shu-Tao Xia, and Luca Benini. Intlora: Integral low-rank adaptation of quantized diffusion models. arXiv preprint arXiv:2410.21759, 2024. 2 +[22] Jian Han, Jinlai Liu, Yi Jiang, Bin Yan, Yuqi Zhang, Zehuan Yuan, Bingyue Peng, and Xiaobing Liu. Infinity: Scaling bitwise autoregressive modeling for high-resolution image synthesis. arXiv preprint arXiv:2412.04431, 2024. 1, 2, 3, 5, 6, 8 +[23] Yefei He, Feng Chen, Yuanyu He, Shaoxuan He, Hong Zhou, Kaipeng Zhang, and Bohan Zhuang. ZIPAR: Accelerating autoregressive image generation through spatial locality. arXiv preprint arXiv:2412.04062, 2024. 2 +[24] Tom Henighan, Jared Kaplan, Mor Katz, Mark Chen, Christopher Hesse, Jacob Jackson, Heewoo Jun, Tom B Brown, Prafulla Dhariwal, Scott Gray, et al. Scaling laws for autoregressive generative modeling. arXiv preprint arXiv:2010.14701, 2020. 2 +[25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. *NeurIPS*, 33:6840-6851, 2020. 2 +[26] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361, 2020. 2 + +[27] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. In CVPR, pages 11523-11532, 2022. 2 +[28] Yaniv Leviathan, Matan Kalman, and Yossi Matias. Fast inference from transformers via speculative decoding. In ICML, pages 19274-19286. PMLR, 2023. 2 +[29] Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Linmiao Xu, and Suhail Doshi. Playground v2.5: Three insights towards enhancing aesthetic quality in text-to-image generation. arXiv preprint arXiv:2402.17245, 2024. 5, 6 +[30] Senmao Li, Taihang Hu, Fahad Shahbaz Khan, Linxuan Li, Shiqi Yang, Yaxing Wang, Ming-Ming Cheng, and Jian Yang. Faster Diffusion: Rethinking the role of unet encoder in diffusion models. NeurIPS, 2023. 2 +[31] Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization. arXiv preprint arXiv:2406.11838, 2024. 1, 2 +[32] Xiuyu Li, Yijiang Liu, Long Lian, Huanrui Yang, Zhen Dong, Daniel Kang, Shanghang Zhang, and Kurt Keutzer. Q-Diffusion: Quantizing diffusion models. In ICCV, pages 17535-17545, 2023. 2 +[33] Dongyang Liu, Shitian Zhao, Le Zhuo, Weifeng Lin, Yu Qiao, Hongsheng Li, and Peng Gao. Lumina-mgpt: Illuminate flexible photorealistic text-to-image generation with multimodal generative pretraining. arXiv preprint arXiv:2408.02657, 2024. 1, 2 +[34] Haozhe Liu, Wentian Zhang, Jinheng Xie, Francesco Faccio, Mengmeng Xu, Tao Xiang, Mike Zheng Shou, Juan-Manuel Perez-Rua, and Jurgen Schmidhuber. Faster diffusion via temporal attention decomposition. arXiv preprint arXiv:2404.02747, 2024. 2 +[35] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 2 +[36] Xinyin Ma, Gongfan Fang, Michael Bi Mi, and Xinchao Wang. Learning-to-Cache: Accelerating diffusion transformer via layer caching. arXiv preprint arXiv:2406.01733, 2024. 2 +[37] Xinyin Ma, Gongfan Fang, and Xinchao Wang. DeepCache: Accelerating diffusion models for free. In CVPR, pages 15762-15772, 2024. 2 +[38] Yue Ma, Yingqing He, Xiaodong Cun, Xintao Wang, Siran Chen, Xiu Li, and Qifeng Chen. Follow your pose: Pose-guided text-to-video generation using pose-free videos. In AAAI, pages 4117-4125, 2024. 2 +[39] Yue Ma, Hongyu Liu, Hongfa Wang, Heng Pan, Yingqing He, Junkun Yuan, Ailing Zeng, Chengfei Cai, Heung-Yeung Shum, Wei Liu, et al. Follow-your-emoji: Fine-controllable and expressive freestyle portrait animation. In SIGGRAPH Asia, pages 1–12, 2024. 2 +[40] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. In CVPR, pages 14297-14306, 2023. 2 + +[41] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 6 +[42] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with VQ-VAE-2. NeurIPS, 32, 2019. 2 +[43] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2 +[44] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 2 +[45] Andrea Santilli, Silvio Severino, Emilian Postolache, Valentino Maiorca, Michele Mancusi, Riccardo Marin, and Emanuele Rodola. Accelerating transformer inference for translation via parallel decoding. arXiv preprint arXiv:2305.10427, 2023. 2 +[46] Yuzhang Shang, Zhihang Yuan, Bin Xie, Bingzhe Wu, and Yan Yan. Post-training quantization on diffusion models. In CVPR, pages 1972-1981, 2023. 2 +[47] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 1, 2, 5, 6 +[48] Haotian Tang, Yecheng Wu, Shang Yang, Enze Xie, Junsong Chen, Junyu Chen, Zhuoyang Zhang, Han Cai, Yao Lu, and Song Han. HART: Efficient visual generation with hybrid autoregressive transformer. arXiv preprint arXiv:2410.10812, 2024. 1, 2, 3, 5, 6, 7 +[49] Yao Teng, Han Shi, Xian Liu, Xuefei Ning, Guohao Dai, Yu Wang, Zhenguo Li, and Xihui Liu. Accelerating autoregressive text-to-image generation with training-free speculative jacobi decoding. arXiv preprint arXiv:2410.01699, 2024. 2 +[50] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. NeurIPS, 2024. 1, 2, 3, 5 +[51] Aaron Van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Conditional image generation with pixelCNN decoders. NeurIPS, 29, 2016. 2 +[52] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. NeurIPS, 30, 2017. 2 +[53] A Vaswani. Attention is all you need. NeurIPS, 2017. 1, 3, 8 +[54] Hongjie Wang, Difan Liu, Yan Kang, Yijun Li, Zhe Lin, Niraj K Jha, and Yuchen Liu. Attention-driven training-free efficiency enhancement of diffusion models. In CVPR, pages 16080-16089, 2024. 2 +[55] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. EMU3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 1, 2 + +[56] Chengyue Wu, Xiaokang Chen, Zhiyu Wu, Yiyang Ma, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, Chong Ruan, et al. Janus: Decoupling visual encoding for unified multimodal understanding and generation. arXiv preprint arXiv:2410.13848, 2024. +[57] Yecheng Wu, Zhuoyang Zhang, Junyu Chen, Haotian Tang, Dacheng Li, Yunhao Fang, Ligeng Zhu, Enze Xie, Hongxu Yin, Li Yi, et al. VILA-U: A unified foundation model integrating visual understanding and generation. arXiv preprint arXiv:2409.04429, 2024. 1 +[58] Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. Efficient streaming language models with attention sinks. arXiv preprint arXiv:2309.17453, 2023. 7 +[59] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 1, 2, 6 +[60] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115, 2024. 3 +[61] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved VQGAN. arXiv preprint arXiv:2110.04627, 2021. 2 +[62] Evelyn Zhang, Bang Xiao, Jiayi Tang, Qianli Ma, Chang Zou, Xuefei Ning, Xuming Hu, and Linfeng Zhang. Token pruning for caching better: 9 times acceleration on stable diffusion for free. arXiv preprint arXiv:2501.00375, 2024. 2 +[63] Zhenyu Zhang, Ying Sheng, Tianyi Zhou, Tianlong Chen, Lianmin Zheng, Ruisi Cai, Zhao Song, Yuandong Tian, Christopher Ré, Clark Barrett, et al. H2O: Heavy-hitter oracle for efficient generative inference of large language models. NeurIPS, 36:34661-34710, 2023. 7 +[64] Chuanxia Zheng, Tung-Long Vuong, Jianfei Cai, and Dinh Phung. MOVQ: Modulating quantized vectors for high-fidelity image generation. NeurIPS, 35:23412-23425, 2022. 2 +[65] Chang Zou, Xuyang Liu, Ting Liu, Siteng Huang, and Linfeng Zhang. Accelerating diffusion transformers with tokenwise feature caching. arXiv preprint arXiv:2410.05317, 2024. 2, 7 \ No newline at end of file diff --git a/2025/FastVAR_ Linear Visual Autoregressive Modeling via Cached Token Pruning/images.zip b/2025/FastVAR_ Linear Visual Autoregressive Modeling via Cached Token Pruning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..810469ee8bed6918db23815e8a0257ac602a8672 --- /dev/null +++ b/2025/FastVAR_ Linear Visual Autoregressive Modeling via Cached Token Pruning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f9c8dacb4a7f9952ff5b707386e8ddf0235f8f59efab275cf99450afc891ab36 +size 1019872 diff --git a/2025/FastVAR_ Linear Visual Autoregressive Modeling via Cached Token Pruning/layout.json b/2025/FastVAR_ Linear Visual Autoregressive Modeling via Cached Token Pruning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5986e051be020eabac2443d03122dabedde6065a --- /dev/null +++ b/2025/FastVAR_ Linear Visual Autoregressive Modeling via Cached Token Pruning/layout.json @@ -0,0 +1,9112 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 69, + 103, + 541, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 103, + 541, + 121 + ], + "spans": [ + { + "bbox": [ + 69, + 103, + 541, + 121 + ], + "type": "text", + "content": "FastVAR: Linear Visual Autoregressive Modeling via Cached Token Pruning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 142, + 447, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 142, + 447, + 157 + ], + "spans": [ + { + "bbox": [ + 160, + 142, + 447, + 157 + ], + "type": "text", + "content": "Hang Guo" + }, + { + "bbox": [ + 160, + 142, + 447, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 160, + 142, + 447, + 157 + ], + "type": "text", + "content": " Yawei Li" + }, + { + "bbox": [ + 160, + 142, + 447, + 157 + ], + "type": "inline_equation", + "content": "^{2,*,†}" + }, + { + "bbox": [ + 160, + 142, + 447, + 157 + ], + "type": "text", + "content": " Taolin Zhang" + }, + { + "bbox": [ + 160, + 142, + 447, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 160, + 142, + 447, + 157 + ], + "type": "text", + "content": " Jiangshan Wang" + }, + { + "bbox": [ + 160, + 142, + 447, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 205, + 157, + 405, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 157, + 405, + 171 + ], + "spans": [ + { + "bbox": [ + 205, + 157, + 405, + 171 + ], + "type": "text", + "content": "Tao Dai" + }, + { + "bbox": [ + 205, + 157, + 405, + 171 + ], + "type": "inline_equation", + "content": "^{3,*}" + }, + { + "bbox": [ + 205, + 157, + 405, + 171 + ], + "type": "text", + "content": " Shu-Tao Xia" + }, + { + "bbox": [ + 205, + 157, + 405, + 171 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 205, + 157, + 405, + 171 + ], + "type": "text", + "content": " Luca Benini" + }, + { + "bbox": [ + 205, + 157, + 405, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 100, + 171, + 510, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 171, + 510, + 186 + ], + "spans": [ + { + "bbox": [ + 100, + 171, + 510, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 100, + 171, + 510, + 186 + ], + "type": "text", + "content": "Tsinghua University " + }, + { + "bbox": [ + 100, + 171, + 510, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 100, + 171, + 510, + 186 + ], + "type": "text", + "content": "ETH Zürich " + }, + { + "bbox": [ + 100, + 171, + 510, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 100, + 171, + 510, + 186 + ], + "type": "text", + "content": "Shenzhen University " + }, + { + "bbox": [ + 100, + 171, + 510, + 186 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 100, + 171, + 510, + 186 + ], + "type": "text", + "content": "Peng Cheng Laboratory" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 57, + 188, + 553, + 337 + ], + "blocks": [ + { + "bbox": [ + 57, + 188, + 553, + 337 + ], + "lines": [ + { + "bbox": [ + 57, + 188, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 57, + 188, + 553, + 337 + ], + "type": "image", + "image_path": "f7a8af838d0135eb4524fec8191bba62387690d99ec0844c15a7df93abc6edde.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 338, + 546, + 350 + ], + "lines": [ + { + "bbox": [ + 62, + 338, + 546, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 338, + 546, + 350 + ], + "type": "text", + "content": "Figure 1. Our FastVAR can generate one 2K image using one NVIDIA 3090 GPU, while existing baseline fails due to out of memory." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 361, + 200, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 361, + 200, + 373 + ], + "spans": [ + { + "bbox": [ + 151, + 361, + 200, + 373 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 377, + 297, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 377, + 297, + 641 + ], + "spans": [ + { + "bbox": [ + 54, + 377, + 297, + 641 + ], + "type": "text", + "content": "Visual Autoregressive (VAR) modeling has gained popularity for its shift towards next-scale prediction. However, existing VAR paradigms process the entire token map at each scale step, leading to the complexity and runtime scaling dramatically with image resolution. To address this challenge, we propose FastVAR, a post-training acceleration method for efficient resolution scaling with VARs. Our key finding is that the majority of latency arises from the large-scale step where most tokens have already converged. Leveraging this observation, we develop the cached token pruning strategy that only forwards pivotal tokens for scale-specific modeling while using cached tokens from previous scale steps to restore the pruned slots. This significantly reduces the number of forwarded tokens and improves the efficiency at larger resolutions. Experiments show the proposed FastVAR can further speedup FlashAttention-accelerated VAR by " + }, + { + "bbox": [ + 54, + 377, + 297, + 641 + ], + "type": "inline_equation", + "content": "2.7 \\times" + }, + { + "bbox": [ + 54, + 377, + 297, + 641 + ], + "type": "text", + "content": " with negligible performance drop of " + }, + { + "bbox": [ + 54, + 377, + 297, + 641 + ], + "type": "inline_equation", + "content": "< 1\\%" + }, + { + "bbox": [ + 54, + 377, + 297, + 641 + ], + "type": "text", + "content": ". We further extend FastVAR to zero-shot generation of higher resolution images. In particular, FastVAR can generate one " + }, + { + "bbox": [ + 54, + 377, + 297, + 641 + ], + "type": "inline_equation", + "content": "2K" + }, + { + "bbox": [ + 54, + 377, + 297, + 641 + ], + "type": "text", + "content": " image with 15GB memory footprints in 1.5s on a single NVIDIA 3090 GPU. Code is available at https://github.com/csguoh/FastVAR." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 651, + 135, + 663 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 135, + 663 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 135, + 663 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 672, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 295, + 696 + ], + "type": "text", + "content": "The next-token prediction of Autoregressive (AR) models [31, 33, 47, 55, 59] has demonstrated performance com" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 361, + 554, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 554, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 554, + 469 + ], + "type": "text", + "content": "petitive with diffusion models for visual generation [47] as well as the potential for unified vision understanding and generation [55-57, 59]. However, this token-by-token paradigm suffers from numerous decoding steps. Recently, Visual Autoregressive (VAR) modeling [50] has shifted the paradigm to next-scale prediction, enabling image generation in fewer steps. Under this new paradigm, some works [22, 48] have developed VAR-based models for text-to-image generation and obtained promising results." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 472, + 554, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 554, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 554, + 639 + ], + "type": "text", + "content": "Despite their potential, existing VAR-based methods [22, 48, 50] face a critical challenge: the computational complexity and runtime latency scale dramatically with image resolution. Specifically, unlike next-token prediction which processes only one token per step, the next-scale prediction of VAR requires processing the entire token map at each decoding step. As a result, the number of tokens increases in " + }, + { + "bbox": [ + 313, + 472, + 554, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(n^2)" + }, + { + "bbox": [ + 313, + 472, + 554, + 639 + ], + "type": "text", + "content": " with the image resolution " + }, + { + "bbox": [ + 313, + 472, + 554, + 639 + ], + "type": "inline_equation", + "content": "n\\times n" + }, + { + "bbox": [ + 313, + 472, + 554, + 639 + ], + "type": "text", + "content": ", and even leads to " + }, + { + "bbox": [ + 313, + 472, + 554, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(n^4)" + }, + { + "bbox": [ + 313, + 472, + 554, + 639 + ], + "type": "text", + "content": " complexity in the attention [53] layer. Empirically, as shown in Fig. 2(a), even when FlashAttention [13] is enabled, VAR models still exhibit super-linear runtime latency. Consequently, this significant computational complexity prevents existing VAR-based models from scaling to higher resolutions, such as 2K." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "In this work, we aim to address the resolution scaling challenge of VARs by pruning forwarded tokens. Through an in-depth analysis of the pre-trained VAR models, we identify the following key findings. 1) Large-scale steps are speed bottleneck but appear robustness. Runtime profiling in Fig. 2(a) finds that the last two large-scale steps account" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 198, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 198, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 198, + 713 + ], + "type": "text", + "content": "*Corresponding Authors, †Project Lead." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19011" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 170, + 153 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 170, + 153 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 170, + 153 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 170, + 153 + ], + "type": "image", + "image_path": "1b3a005446838e9bb0555a4a0b5ba28a31911a56df4a362507c535f304c8b8f0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 154, + 172, + 163 + ], + "lines": [ + { + "bbox": [ + 63, + 154, + 172, + 163 + ], + "spans": [ + { + "bbox": [ + 63, + 154, + 172, + 163 + ], + "type": "text", + "content": "(a) Runtime at different scale steps" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 175, + 70, + 291, + 153 + ], + "blocks": [ + { + "bbox": [ + 175, + 70, + 291, + 153 + ], + "lines": [ + { + "bbox": [ + 175, + 70, + 291, + 153 + ], + "spans": [ + { + "bbox": [ + 175, + 70, + 291, + 153 + ], + "type": "image", + "image_path": "9ff3b0530a79108f977ec72bf362dc257ced8ca70cbd26c0c4c984d31bacdb20.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 154, + 293, + 163 + ], + "lines": [ + { + "bbox": [ + 181, + 154, + 293, + 163 + ], + "spans": [ + { + "bbox": [ + 181, + 154, + 293, + 163 + ], + "type": "text", + "content": "(b) Efficiency-performance tradeoff" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 164, + 295, + 187 + ], + "lines": [ + { + "bbox": [ + 55, + 164, + 295, + 187 + ], + "spans": [ + { + "bbox": [ + 55, + 164, + 295, + 187 + ], + "type": "text", + "content": "Figure 2. FastVAR exhibits promising resolution scalability, and can achieve noticeable speedup with negligible performance drop." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 194, + 295, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 194, + 295, + 385 + ], + "spans": [ + { + "bbox": [ + 54, + 194, + 295, + 385 + ], + "type": "text", + "content": "for " + }, + { + "bbox": [ + 54, + 194, + 295, + 385 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 54, + 194, + 295, + 385 + ], + "type": "text", + "content": " of the total runtime. Further investigation in Tab. 4 reveals that VAR is more resilient to token pruning at large-scale steps than at smaller scales. Thus, we direct our efforts to the large-scale steps. 2) High-frequency modeling matters at large-scale steps. Spectrum analysis in Fig. 3(b) reveals that the large-scale steps are optimized to model high-frequency tokens, such as texture details, and the remaining low-frequency tokens almost converge in these steps. Consequently, we can only forward the high-frequency tokens for pruning. 3) Tokens from different scales are related. Attention map analysis in Fig. 3(c) shows strong diagonal sparsity across scales, indicating that tokens attend not only to same-scale neighbors but also to those from the preceding scales. Thus, we cache tokens in previous steps to compensate for the pruned slots, preserving the 2D image lattice structure and maintaining information flow." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 387, + 295, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 387, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 387, + 295, + 613 + ], + "type": "text", + "content": "Based on the above observations, we propose FastVAR, a post-training acceleration recipe for efficient resolution scaling with VARs. At its core, FastVAR employs \"cached token pruning\", which retains only pivotal tokens at large-scale steps to reduce computational overhead while using cached token maps from early-scale steps to compensate for information loss. To identify the pivotal tokens, we develop Pivotal Token Selection (PTS), a frequency-based scoring mechanism. Token importance is determined by filtering out the low-frequency component estimated via the direct current component, which enables efficient frequency-based token selection directly in the spatial domain. The selected Top-K pivotal tokens are then processed by the VAR model. To restore pruned tokens, we propose Cached Token Restoration (CTR), which first interpolates the token map from the cached scale and then reinstates the pruned tokens by indexing the interpolated token map to the pruned location. By integrating these strategies, we present FastVAR which enjoys the following benefits:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 616, + 295, + 711 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 55, + 616, + 295, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 295, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 295, + 638 + ], + "type": "text", + "content": "- The proposed method is training-free and plug-and-play for various VAR-based backbones." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 639, + 295, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 639, + 295, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 639, + 295, + 674 + ], + "type": "text", + "content": "- As shown in Fig. 2(b), FastVAR can be integrated with FlashAttention, with further " + }, + { + "bbox": [ + 55, + 639, + 295, + 674 + ], + "type": "inline_equation", + "content": "2.7 \\times" + }, + { + "bbox": [ + 55, + 639, + 295, + 674 + ], + "type": "text", + "content": " speedup and " + }, + { + "bbox": [ + 55, + 639, + 295, + 674 + ], + "type": "inline_equation", + "content": "< 1\\%" + }, + { + "bbox": [ + 55, + 639, + 295, + 674 + ], + "type": "text", + "content": " performance drop." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 675, + 295, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 675, + 295, + 711 + ], + "spans": [ + { + "bbox": [ + 55, + 675, + 295, + 711 + ], + "type": "text", + "content": "- In Fig. 1, FastVAR facilitates zero-shot scaling to larger-resolution and can produce one 2K image using only 15GB memory in 1.5s on a single NVIDIA 3090 GPU." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 91, + 553, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 553, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 553, + 402 + ], + "type": "text", + "content": "Autoregressive Visual Generation. The previous autoregressive (AR) methods [27, 42, 61, 64] mostly adopt the next-token prediction paradigm, which treats each pixel as one token and generates pixels in a GPT or BERT style [1, 4, 6, 14, 35]. For instance, some pioneering works [9, 51] generate pixels in raster scan order with transformer models. Later, VQVAE [52] and VQGAN [15] propose to quantize image patches into discrete tokens to ease the training process. Benefiting from the scaling up [24, 26], recent works have shown promising results that the autoregressive models can achieve a more competitive performance than state-of-the-art Diffusion models [7, 8, 38, 39, 41, 43]. Despite this progress, existing next-token generation pipelines struggle to efficiently synthesize high-resolution images due to the numerous decoding steps [10]. More recently, the Visual Autoregressive (VAR) modeling [50], which shifts to next-scale generation fashion to allow only a few forward steps, further advances the generation quality as well as the model efficiency. Building on top of VAR, existing methods have scaled up the model for production-level text-to-image generation. For example, HART [48] adopts the continuous diffusion module to compensate for the quantization error. Infinity [22] employs a bitwise tokenizer for an extremely large vector vocabulary. Despite the promising ability of VAR-based methods, they struggle to scale to larger resolutions with the increasing token numbers." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 403, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 553, + 713 + ], + "type": "text", + "content": "Efficient Visual Generation. To speed up diffusion models [25], many efforts have been made in recent years, including distillation [40, 44], quantization [21, 32, 46], pruning [2, 17, 54, 62, 65], and caching [30, 34, 36, 37], which either reduce the total denoising steps or to reduce the cost of single forward. Specifically, DeepCache [37] proposes to reuse the intermediate features of low-resolution layers in the U-Net. ToMeSD [2] merges similar tokens into one token with subsequent unmerging for acceleration. Unfortunately, these methods are specifically designed for diffusion, which can achieve only sub-optimal performance or cannot be used in VAR. In the efforts to accelerate AR modeling [31, 33, 47, 55, 59], one prevalent solution is to reduce the forward step using parallel decoding strategies [23, 45, 49]. For instance, speculative decoding [5, 28] utilizes a small draft model to generate candidate tokens, which are then verified in parallel by the larger model. However, the well-developed decoding strategy cannot be directly applied to the next-scale paradigm, whose each step involves multiple pixels [50]. As for fast VAR models, one related work is the CoDe [10], which uses model ensemble techniques to use a small model on the costly large resolution. However, CoDe highly depends on the availability of different-sized models. In this work, we attempt to speed up the VARs by pruning extra tokens, which is training-free and is generic for multiple VAR methods." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19012" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 545, + 110 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 545, + 110 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 545, + 110 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 545, + 110 + ], + "type": "image", + "image_path": "9d8bdeccafa1ca64e9f1197238f8b3456a726305eeec8e8e90c1bb34b173930d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 61, + 114, + 544, + 160 + ], + "blocks": [ + { + "bbox": [ + 61, + 114, + 544, + 160 + ], + "lines": [ + { + "bbox": [ + 61, + 114, + 544, + 160 + ], + "spans": [ + { + "bbox": [ + 61, + 114, + 544, + 160 + ], + "type": "image", + "image_path": "493f32c14f90245c102a0a13a3f008567e5e08c10860e4d798a7f2eb4fa4e4e3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 159, + 423, + 169 + ], + "lines": [ + { + "bbox": [ + 184, + 159, + 423, + 169 + ], + "spans": [ + { + "bbox": [ + 184, + 159, + 423, + 169 + ], + "type": "text", + "content": "(a) Visualization of Intermediate Outputs from Different Scale Steps" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 60, + 167, + 231, + 280 + ], + "blocks": [ + { + "bbox": [ + 60, + 167, + 231, + 280 + ], + "lines": [ + { + "bbox": [ + 60, + 167, + 231, + 280 + ], + "spans": [ + { + "bbox": [ + 60, + 167, + 231, + 280 + ], + "type": "image", + "image_path": "0fa08fb31825a077185edd39ad7f10f0df57ae52c4e69c777996cd656e04713f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 281, + 215, + 290 + ], + "lines": [ + { + "bbox": [ + 69, + 281, + 215, + 290 + ], + "spans": [ + { + "bbox": [ + 69, + 281, + 215, + 290 + ], + "type": "text", + "content": "(b) Fourier Analysis of Large-Scale Steps" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 291, + 555, + 314 + ], + "lines": [ + { + "bbox": [ + 55, + 291, + 555, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 291, + 555, + 314 + ], + "type": "text", + "content": "Figure 3. (a) We resize different " + }, + { + "bbox": [ + 55, + 291, + 555, + 314 + ], + "type": "inline_equation", + "content": "\\tilde{r}_k" + }, + { + "bbox": [ + 55, + 291, + 555, + 314 + ], + "type": "text", + "content": " to the same size for better presentation. (b) Each curve represents the frequency characteristics of a certain-scale token map. (c) The row represents keys of scale steps " + }, + { + "bbox": [ + 55, + 291, + 555, + 314 + ], + "type": "inline_equation", + "content": "\\{1,2,\\dots ,k\\}" + }, + { + "bbox": [ + 55, + 291, + 555, + 314 + ], + "type": "text", + "content": ", and the column is queries of the " + }, + { + "bbox": [ + 55, + 291, + 555, + 314 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 291, + 555, + 314 + ], + "type": "text", + "content": "-th scale step." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 247, + 169, + 359, + 277 + ], + "blocks": [ + { + "bbox": [ + 247, + 169, + 359, + 277 + ], + "lines": [ + { + "bbox": [ + 247, + 169, + 359, + 277 + ], + "spans": [ + { + "bbox": [ + 247, + 169, + 359, + 277 + ], + "type": "image", + "image_path": "35ccdaaf069943da7325975a09b06757f9d4c75129cd14066a203a9768ffcf96.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 280, + 465, + 290 + ], + "lines": [ + { + "bbox": [ + 310, + 280, + 465, + 290 + ], + "spans": [ + { + "bbox": [ + 310, + 280, + 465, + 290 + ], + "type": "text", + "content": "(c) Cross-scale Attention Map Visualization" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 360, + 169, + 443, + 277 + ], + "blocks": [ + { + "bbox": [ + 360, + 169, + 443, + 277 + ], + "lines": [ + { + "bbox": [ + 360, + 169, + 443, + 277 + ], + "spans": [ + { + "bbox": [ + 360, + 169, + 443, + 277 + ], + "type": "image", + "image_path": "5a5fa61cc374bd3fb16d4ad55eecf03d9eb4bc15bb293975f8120a80c36588cc.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 445, + 169, + 547, + 277 + ], + "blocks": [ + { + "bbox": [ + 445, + 169, + 547, + 277 + ], + "lines": [ + { + "bbox": [ + 445, + 169, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 445, + 169, + 547, + 277 + ], + "type": "image", + "image_path": "492600e259c2361c9bd4d44256ebf080910d89623d55b7e588ec4836a513edc6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 324, + 111, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 111, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 111, + 335 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 344, + 134, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 344, + 134, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 344, + 134, + 357 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 361, + 296, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 296, + 482 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 296, + 482 + ], + "type": "text", + "content": "The Visual Autoregressive (VAR) modeling [50] redefines the traditional AR by shifting from a \"next-token prediction\" to a \"next-scale prediction\", where each autoregressive unit is a token map of varying scales instead of a single token. For a given image feature map, VAR first quantizes it into " + }, + { + "bbox": [ + 55, + 361, + 296, + 482 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 361, + 296, + 482 + ], + "type": "text", + "content": " multi-scale token maps " + }, + { + "bbox": [ + 55, + 361, + 296, + 482 + ], + "type": "inline_equation", + "content": "\\mathcal{R} = \\{r_1, r_2, \\dots, r_K\\}" + }, + { + "bbox": [ + 55, + 361, + 296, + 482 + ], + "type": "text", + "content": " with progressively larger resolutions using pre-defined scale schdual " + }, + { + "bbox": [ + 55, + 361, + 296, + 482 + ], + "type": "inline_equation", + "content": "\\{(h_1, w_1), (h_2, w_2), \\dots, (h_K, w_K)\\}" + }, + { + "bbox": [ + 55, + 361, + 296, + 482 + ], + "type": "text", + "content": ". Then the " + }, + { + "bbox": [ + 55, + 361, + 296, + 482 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 55, + 361, + 296, + 482 + ], + "type": "text", + "content": " is used to train a Transformer [53] to learn the joint distribution using the causal formulation:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 70, + 491, + 295, + 524 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 491, + 295, + 524 + ], + "spans": [ + { + "bbox": [ + 70, + 491, + 295, + 524 + ], + "type": "interline_equation", + "content": "p \\left(r _ {1}, r _ {2}, \\dots , r _ {K}\\right) = \\prod_ {k = 1} ^ {K} p \\left(r _ {k} \\mid r _ {1}, r _ {2}, \\dots , r _ {k - 1}\\right), \\tag {1}", + "image_path": "f5cd7b4cb17e5ef9e3b29b1d3e1a2d01334f3fc380b115a03c3abaeb371afad3.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 533, + 296, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 296, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 296, + 605 + ], + "type": "text", + "content": "where the start token " + }, + { + "bbox": [ + 55, + 533, + 296, + 605 + ], + "type": "inline_equation", + "content": "r_1" + }, + { + "bbox": [ + 55, + 533, + 296, + 605 + ], + "type": "text", + "content": " is the encoded text embedding from the LLMs [11, 60] under given user prompts. During training, existing methods [22, 48, 50] usually employ a residual strategy to reduce learning difficulty. In detail, the model outputs at the " + }, + { + "bbox": [ + 55, + 533, + 296, + 605 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 533, + 296, + 605 + ], + "type": "text", + "content": "-th scale step " + }, + { + "bbox": [ + 55, + 533, + 296, + 605 + ], + "type": "inline_equation", + "content": "f_k" + }, + { + "bbox": [ + 55, + 533, + 296, + 605 + ], + "type": "text", + "content": " is the residual of the intermediate prediction " + }, + { + "bbox": [ + 55, + 533, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\tilde{r}_k" + }, + { + "bbox": [ + 55, + 533, + 296, + 605 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 93, + 615, + 295, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 615, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 93, + 615, + 295, + 628 + ], + "type": "interline_equation", + "content": "\\tilde {r} _ {k} = \\operatorname {i n t e r p o l a t e} \\left(\\tilde {r} _ {k - 1}, \\left(h _ {k}, w _ {k}\\right)\\right) + f _ {k}, \\tag {2}", + "image_path": "56d94a3cc96952d3f07aac4049df77f90b52d0e4bce70ca328f5b63d8730e7fd.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 637, + 296, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 637, + 296, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 637, + 296, + 673 + ], + "type": "text", + "content": "where the \"interpolate " + }, + { + "bbox": [ + 55, + 637, + 296, + 673 + ], + "type": "inline_equation", + "content": "(\\cdot ,(h_k,w_k))" + }, + { + "bbox": [ + 55, + 637, + 296, + 673 + ], + "type": "text", + "content": " denotes upsample a token map to the size of " + }, + { + "bbox": [ + 55, + 637, + 296, + 673 + ], + "type": "inline_equation", + "content": "(h_k,w_k)" + }, + { + "bbox": [ + 55, + 637, + 296, + 673 + ], + "type": "text", + "content": ". Unfolding Eq. (2) derives the following cumulative form:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 100, + 684, + 295, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 684, + 295, + 716 + ], + "spans": [ + { + "bbox": [ + 100, + 684, + 295, + 716 + ], + "type": "interline_equation", + "content": "\\tilde {r} _ {k} = \\sum_ {i = 1} ^ {k} \\operatorname {i n t e r p o l a t e} \\left(f _ {i}, \\left(h _ {k}, w _ {k}\\right)\\right). \\tag {3}", + "image_path": "aa6131fe5758dbc9a645c3cc370b3e032c55ae27f024c89e3dc0e8073b695c0b.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 325, + 555, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 325, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 325, + 555, + 422 + ], + "type": "text", + "content": "In the last scale step, the predicted token maps " + }, + { + "bbox": [ + 313, + 325, + 555, + 422 + ], + "type": "inline_equation", + "content": "\\tilde{r}_K" + }, + { + "bbox": [ + 313, + 325, + 555, + 422 + ], + "type": "text", + "content": " is used to generate the final image. During inference, since the outputs at different scales are generated autoregressively, the KV-Cache can be adopted to avoid re-computing of previous steps. Although the VAR paradigm can generate images within only " + }, + { + "bbox": [ + 313, + 325, + 555, + 422 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 325, + 555, + 422 + ], + "type": "text", + "content": " scale steps, the Transformer needs to process all " + }, + { + "bbox": [ + 313, + 325, + 555, + 422 + ], + "type": "inline_equation", + "content": "h_k \\times w_k" + }, + { + "bbox": [ + 313, + 325, + 555, + 422 + ], + "type": "text", + "content": " tokens at the " + }, + { + "bbox": [ + 313, + 325, + 555, + 422 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 325, + 555, + 422 + ], + "type": "text", + "content": "-th step, hindering scaling to higher-resolution image synthesis." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 429, + 389, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 389, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 389, + 441 + ], + "type": "text", + "content": "3.2. Motivation" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 447, + 555, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 555, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 555, + 519 + ], + "type": "text", + "content": "In this work, we aim to approach linear VAR models to advance high-resolution image generation. To this end, we investigate the scale-wise behavior of the pre-trained VAR model [22]. We examine from three angles: the latency profiling, the specific role of each step, and the cross-scale token dependency. Conclusions are summarized as follows." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 522, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 556, + 714 + ], + "type": "text", + "content": "Observation 1: Large-scale steps are speed bottleneck but appear robustness. As shown in the runtime curve in Fig. 2(a), even the FlashAttention is enabled, the inference latency of the VAR exhibits a super-linear complexity with resolution steps. This is because the Transformer needs to take all the tokens at once. As a result, the tokens from the large-scale steps make up the majority of the total token length. For example, the last two steps even occupy " + }, + { + "bbox": [ + 313, + 522, + 556, + 714 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 313, + 522, + 556, + 714 + ], + "type": "text", + "content": " of the total time. On the other hand, the experiments in Sec. 4.3 suggest the large-scale steps are more robust to token drop than small-scale counterparts under the same pruning ratio. Therefore, it is accessible to prune some \"unimportant tokens\" at the large scale step to speed up token processing. However, challenges still remain about which tokens should be pruned and how to compensate for the information loss from these pruned tokens." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19013" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 70, + 298, + 222 + ], + "blocks": [ + { + "bbox": [ + 72, + 70, + 298, + 222 + ], + "lines": [ + { + "bbox": [ + 72, + 70, + 298, + 222 + ], + "spans": [ + { + "bbox": [ + 72, + 70, + 298, + 222 + ], + "type": "image", + "image_path": "24e5cc3c3e7b46970a2922892723195dab6692937a52949043f0f5fe1c5ba1ae.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 104, + 223, + 258, + 232 + ], + "lines": [ + { + "bbox": [ + 104, + 223, + 258, + 232 + ], + "spans": [ + { + "bbox": [ + 104, + 223, + 258, + 232 + ], + "type": "text", + "content": "(a) Overall pipeline of the proposed FastVAR" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 316, + 73, + 539, + 220 + ], + "blocks": [ + { + "bbox": [ + 316, + 73, + 539, + 220 + ], + "lines": [ + { + "bbox": [ + 316, + 73, + 539, + 220 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 539, + 220 + ], + "type": "image", + "image_path": "8d36b4ffa8b04f161b459102072d2324e7ff1b95e97aa8a295b4acc60ecc1d20.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 222, + 531, + 232 + ], + "lines": [ + { + "bbox": [ + 326, + 222, + 531, + 232 + ], + "spans": [ + { + "bbox": [ + 326, + 222, + 531, + 232 + ], + "type": "text", + "content": "(b) Token-wise implementation for one VAR layer at stage " + }, + { + "bbox": [ + 326, + 222, + 531, + 232 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 236, + 555, + 281 + ], + "lines": [ + { + "bbox": [ + 54, + 236, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 54, + 236, + 555, + 281 + ], + "type": "text", + "content": "Figure 4. (a) We keep the Structure Construction Stage as standard VAR and additionally store token maps at the " + }, + { + "bbox": [ + 54, + 236, + 555, + 281 + ], + "type": "inline_equation", + "content": "(K - N)" + }, + { + "bbox": [ + 54, + 236, + 555, + 281 + ], + "type": "text", + "content": "-th step as the cache. Then the Texture Filling Stage applies FastVAR to process only pivotal tokens for fast inference. (b) For Attention or FFN layers using FastVAR, we first employ the Pivotal Token Selection (PTS) for token reduction. After the model forward, we develop the Cached Token Restoration (CTR) to restore the original token numbers using the caches from the previous scale step." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 301, + 296, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 301, + 296, + 539 + ], + "spans": [ + { + "bbox": [ + 54, + 301, + 296, + 539 + ], + "type": "text", + "content": "Observation 2: High-frequency modeling matters at large-scale steps. To find which token should be pruned, we visualize the intermediate predictions " + }, + { + "bbox": [ + 54, + 301, + 296, + 539 + ], + "type": "inline_equation", + "content": "\\tilde{r}_k" + }, + { + "bbox": [ + 54, + 301, + 296, + 539 + ], + "type": "text", + "content": " in Fig. 3(a). Interestingly, we find that the generation process of the pretrained VAR model can be decomposed into two stages. The Structure Construction Stage, which includes small-scale steps, is trained to generate the outline of a subject. After that, at the Texture Filling Stage, which is composed of large-scale steps, is responsible for adding details based on the previous sketch. Given that the outline is low-frequency while the texture is high-frequency, we therefore infer that the VAR mainly models the high-frequency texture in the large-scale stage while keeping the low-frequency structure almost intact. For further verification, we show in Fig. 3(b) that the low-frequency component almost converges in the large-scale steps, while the high-frequency modeling stays optimized with noticeable variations. Therefore, we can prune the redundant low-frequency tokens and forward only the pivotal high-frequency tokens at the large-scale steps for token number reduction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 546, + 295, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 556 + ], + "type": "text", + "content": "Observation 3: Tokens from different scales are related." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 558, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 558, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 558, + 295, + 713 + ], + "type": "text", + "content": "After reducing the input tokens for fast forward, we still need to restore the output tokens to the original numbers in order to maintain the 2D image structure. To this end, we analyze the cross-scale attention map of the unpruned VAR model, where the query at the current scale interacts with KV caches from all previous scales. As depicted in Fig. 3(c), we observe that the attention map not only exhibits a high diagonal score at self-scale, but also a high diagonal score at cross-scale. This behavior indicates that a token in the current scale not only attends to its same-scale locals but also exhibits strong correlation to tokens at the neighboring positions in previous scales. This diagonal attention sparsity thus facilitates us to use only a few" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 301, + 555, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 555, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 555, + 374 + ], + "type": "text", + "content": "tokens to estimate the original output of the pruned slots, and avoids weighting multiple tokens across the token map. Specifically, we can approximate the outputs of pruned tokens by indexing in the previous-scale token maps based on the pruning location. In this way, the information loss can be compensated thanks to this high cross-scale similarity." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 383, + 512, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 383, + 512, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 383, + 512, + 396 + ], + "type": "text", + "content": "3.3. Efficient Resolution Scaling for VARs" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 401, + 554, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 554, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 554, + 449 + ], + "type": "text", + "content": "We instantiate the above idea and present FastVAR to approximate linear VAR models. The core is the cached token pruning, which prunes tokens at large scale steps while using cache from early scales to maintain information flow." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 449, + 555, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 581 + ], + "type": "text", + "content": "Overview. As shown in Fig. 4(a), denote the scale step set of Structure Construction Stage as " + }, + { + "bbox": [ + 313, + 449, + 555, + 581 + ], + "type": "inline_equation", + "content": "S = \\{1,2,\\dots ,K - N\\}" + }, + { + "bbox": [ + 313, + 449, + 555, + 581 + ], + "type": "text", + "content": " the step set of Texture Filling Stage as " + }, + { + "bbox": [ + 313, + 449, + 555, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{K - N + 1,\\dots ,K\\}" + }, + { + "bbox": [ + 313, + 449, + 555, + 581 + ], + "type": "text", + "content": ". As stated in Sec. 3.2, the speedup of pruning token maps in " + }, + { + "bbox": [ + 313, + 449, + 555, + 581 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 449, + 555, + 581 + ], + "type": "text", + "content": " is insignificant while producing a noticeable performance drop, we therefore keep " + }, + { + "bbox": [ + 313, + 449, + 555, + 581 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 449, + 555, + 581 + ], + "type": "text", + "content": " intact as the standard VAR. In addition, we save the per-layer outputs at the last scale in " + }, + { + "bbox": [ + 313, + 449, + 555, + 581 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 449, + 555, + 581 + ], + "type": "text", + "content": ", i.e., the " + }, + { + "bbox": [ + 313, + 449, + 555, + 581 + ], + "type": "inline_equation", + "content": "(K - N)" + }, + { + "bbox": [ + 313, + 449, + 555, + 581 + ], + "type": "text", + "content": "-th step, as the cache for subsequent token restoration. For the set " + }, + { + "bbox": [ + 313, + 449, + 555, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 449, + 555, + 581 + ], + "type": "text", + "content": ", we apply FastVAR to prune tokens for fast-forward, followed by the token number restoration. More details are given below." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "content": "Pivotal Token Selection. We first introduce the Pivotal Token Selection (PTS) to reduce the number of tokens. As observed in Fig. 3(b), the low-frequency tokens almost converge in the later scale steps. Therefore, we can feed only the high-frequency tokens and prune the remaining. However, the challenge lies in that common frequency operators, such as FFT, work in the frequency domain, making it difficult to recognize the frequency characteristics of certain tokens in the original token map. To this end, our proposed PTS adopts an approximation solution. Specifically, denote " + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "inline_equation", + "content": "x_{k}" + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "content": " as the input of one layer, we estimate its low-frequency" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19014" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "component " + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "\\bar{x}_k" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": " as the direct current component, which can be easily obtained via the spatially global average pooling:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 117, + 105, + 295, + 119 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 105, + 295, + 119 + ], + "spans": [ + { + "bbox": [ + 117, + 105, + 295, + 119 + ], + "type": "interline_equation", + "content": "\\bar {x} _ {k} = \\text {g l o b a l - a v g - p o o l} \\left(x _ {k}\\right). \\tag {4}", + "image_path": "c3afe3d9568d6ca9f87b0144565eff72f01700b5a29c8241dc55ce35fb1eada8.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 126, + 295, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 126, + 295, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 126, + 295, + 163 + ], + "type": "text", + "content": "After that, the high-frequency component is computed as the difference between the " + }, + { + "bbox": [ + 55, + 126, + 295, + 163 + ], + "type": "inline_equation", + "content": "x_{k}" + }, + { + "bbox": [ + 55, + 126, + 295, + 163 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 126, + 295, + 163 + ], + "type": "inline_equation", + "content": "\\bar{x}_k" + }, + { + "bbox": [ + 55, + 126, + 295, + 163 + ], + "type": "text", + "content": ". The pivotal score " + }, + { + "bbox": [ + 55, + 126, + 295, + 163 + ], + "type": "inline_equation", + "content": "s_k" + }, + { + "bbox": [ + 55, + 126, + 295, + 163 + ], + "type": "text", + "content": " is defined as the L2 norm of the high-frequency maps:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 136, + 171, + 295, + 185 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 171, + 295, + 185 + ], + "spans": [ + { + "bbox": [ + 136, + 171, + 295, + 185 + ], + "type": "interline_equation", + "content": "s _ {k} = \\left\\| x _ {k} - \\bar {x} _ {k} \\right\\| _ {2}. \\tag {5}", + "image_path": "bb522e84c963f384a5ae2174e8b5b9d78dec1ab52fe7b547e2b1d77d681ef262.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 192, + 296, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 192, + 296, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 296, + 289 + ], + "type": "text", + "content": "Subsequently, we obtain the index set " + }, + { + "bbox": [ + 55, + 192, + 296, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 55, + 192, + 296, + 289 + ], + "type": "text", + "content": " for pivotal token selection by keeping TopK scores in " + }, + { + "bbox": [ + 55, + 192, + 296, + 289 + ], + "type": "inline_equation", + "content": "s_k" + }, + { + "bbox": [ + 55, + 192, + 296, + 289 + ], + "type": "text", + "content": ". Finally, we perform token pruning on " + }, + { + "bbox": [ + 55, + 192, + 296, + 289 + ], + "type": "inline_equation", + "content": "x_{k}" + }, + { + "bbox": [ + 55, + 192, + 296, + 289 + ], + "type": "text", + "content": " through indexing with " + }, + { + "bbox": [ + 55, + 192, + 296, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 55, + 192, + 296, + 289 + ], + "type": "text", + "content": " to allow the fast model forward. Notably, since the PTS reduces the number of input tokens in one transformer layer, as an additional benefit, the KV-cache is also reduced accordingly, thus optimizing the GPU footprints as well as the cross-scale attention for the subsequent scale steps." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 289, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 296, + 456 + ], + "type": "text", + "content": "Cached Token Restoration. The proposed PTS can effectively reduce the number of tokens. However, the visual generation tasks need to restore the original token numbers to regain the structural 2D image. For this reason, we propose Cached Token Restoration (CTR). The main rationale stems from the diagonal attention sparsity that the token at the " + }, + { + "bbox": [ + 55, + 289, + 296, + 456 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 289, + 296, + 456 + ], + "type": "text", + "content": "-th scale not only attends to itself, but also exhibits strong correlation with the corresponding slot from previous " + }, + { + "bbox": [ + 55, + 289, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\{1,2,\\dots,k-1\\}" + }, + { + "bbox": [ + 55, + 289, + 296, + 456 + ], + "type": "text", + "content": " steps. Therefore, our proposed CTR approximates the layer outputs in the pruned locations using the counterparts from the cached step. Formally, given the cached token map " + }, + { + "bbox": [ + 55, + 289, + 296, + 456 + ], + "type": "inline_equation", + "content": "y_{K-N}" + }, + { + "bbox": [ + 55, + 289, + 296, + 456 + ], + "type": "text", + "content": " with size " + }, + { + "bbox": [ + 55, + 289, + 296, + 456 + ], + "type": "inline_equation", + "content": "(h_{K-N},w_{K-N})" + }, + { + "bbox": [ + 55, + 289, + 296, + 456 + ], + "type": "text", + "content": ", which is the output of the corresponding layer at step " + }, + { + "bbox": [ + 55, + 289, + 296, + 456 + ], + "type": "inline_equation", + "content": "K-N" + }, + { + "bbox": [ + 55, + 289, + 296, + 456 + ], + "type": "text", + "content": ", we first upsample it to the same size of the current scale:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 93, + 464, + 295, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 464, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 93, + 464, + 295, + 479 + ], + "type": "interline_equation", + "content": "y _ {k} ^ {\\text {c a c h e}} = \\operatorname {i n t e r p o l a t e} \\left(y _ {K - N}, \\left(h _ {k}, w _ {k}\\right)\\right). \\tag {6}", + "image_path": "00f3d42accb0d34860327e67e78d106ab12f976acd5b53a90822301ccce474b8.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 486, + 295, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 545 + ], + "type": "text", + "content": "Then the value from " + }, + { + "bbox": [ + 55, + 486, + 295, + 545 + ], + "type": "inline_equation", + "content": "y_{k}^{cache}" + }, + { + "bbox": [ + 55, + 486, + 295, + 545 + ], + "type": "text", + "content": " is scattered into the pruned slots of " + }, + { + "bbox": [ + 55, + 486, + 295, + 545 + ], + "type": "inline_equation", + "content": "y_{k}" + }, + { + "bbox": [ + 55, + 486, + 295, + 545 + ], + "type": "text", + "content": " using the index set " + }, + { + "bbox": [ + 55, + 486, + 295, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 55, + 486, + 295, + 545 + ], + "type": "text", + "content": ", to generate the restored layer output " + }, + { + "bbox": [ + 55, + 486, + 295, + 545 + ], + "type": "inline_equation", + "content": "y_{k}^{\\prime}" + }, + { + "bbox": [ + 55, + 486, + 295, + 545 + ], + "type": "text", + "content": ", which shares the same token numbers as the " + }, + { + "bbox": [ + 55, + 486, + 295, + 545 + ], + "type": "inline_equation", + "content": "x_{k}" + }, + { + "bbox": [ + 55, + 486, + 295, + 545 + ], + "type": "text", + "content": ". We also provide an explanation for the choice of " + }, + { + "bbox": [ + 55, + 486, + 295, + 545 + ], + "type": "inline_equation", + "content": "K - N" + }, + { + "bbox": [ + 55, + 486, + 295, + 545 + ], + "type": "text", + "content": " as the caching step, see Suppl. for more details." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": "Implementation of FastVAR. Thanks to the token number invariance of the " + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{PTS + CTR}" + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": ", as shown in Fig. 4(b), we apply this operator pair for each Attention&FFN layer at the large-scale steps. Furthermore, we empirically find in Sec. 4.3 that the larger scale steps exhibit stronger robustness to pruning than that of early scale steps. Therefore, we develop a progressive pruning ratio schedule, in which we assign larger pruning ratios to larger scale steps in the set " + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": ". We summarize the complete algorithm of FastVAR in the Suppl.. At last, it is worth noting that our FastVAR does not access the attention map, and we show in Sec. 4.4 that the proposed FastVAR can be used in combination with other acceleration techniques like FlashAttention [12, 13] for even faster VAR generation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 72, + 395, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 395, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 395, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 91, + 430, + 105 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 91, + 430, + 105 + ], + "spans": [ + { + "bbox": [ + 314, + 91, + 430, + 105 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 110, + 555, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 110, + 555, + 253 + ], + "spans": [ + { + "bbox": [ + 313, + 110, + 555, + 253 + ], + "type": "text", + "content": "Models and Evaluations. We apply our FastVAR on two VAR-based text-to-image models, namely HART [48] and Infinity [22]. Both models can generate images of up to " + }, + { + "bbox": [ + 313, + 110, + 555, + 253 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 313, + 110, + 555, + 253 + ], + "type": "text", + "content": " resolution. For a fair comparison, we keep all the hyperparameters the same as their default settings. For evaluation metrics, we compare both in terms of generation quality and inference efficiency. For generation quality, we use two popular benchmarks, i.e., the GenEval [20] and MJHQ30K [29] to validate the high-level semantic consistency and the perceptual quality, respectively. For efficiency evaluation, we adopt metrics including running time, throughputs, speedup ratio, and GPU memory costs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 255, + 556, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 556, + 482 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 556, + 482 + ], + "type": "text", + "content": "Implementation Details. For the number of pruned scale steps " + }, + { + "bbox": [ + 313, + 255, + 556, + 482 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 255, + 556, + 482 + ], + "type": "text", + "content": ", we set " + }, + { + "bbox": [ + 313, + 255, + 556, + 482 + ], + "type": "inline_equation", + "content": "N = 4" + }, + { + "bbox": [ + 313, + 255, + 556, + 482 + ], + "type": "text", + "content": " for Infinity and " + }, + { + "bbox": [ + 313, + 255, + 556, + 482 + ], + "type": "inline_equation", + "content": "N = 2" + }, + { + "bbox": [ + 313, + 255, + 556, + 482 + ], + "type": "text", + "content": " for HART, i.e., only token maps from the last 4 or 2 scale steps are applied with the proposed FastVAR, with other steps kept as standard VAR. For the progressive pruning ratio schedule, we set it to " + }, + { + "bbox": [ + 313, + 255, + 556, + 482 + ], + "type": "inline_equation", + "content": "\\{40\\%, 50\\%, 100\\%, 100\\}" + }, + { + "bbox": [ + 313, + 255, + 556, + 482 + ], + "type": "text", + "content": " for Infinity, and " + }, + { + "bbox": [ + 313, + 255, + 556, + 482 + ], + "type": "inline_equation", + "content": "\\{50\\%, 75\\%\\}" + }, + { + "bbox": [ + 313, + 255, + 556, + 482 + ], + "type": "text", + "content": " for HART. The " + }, + { + "bbox": [ + 313, + 255, + 556, + 482 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 313, + 255, + 556, + 482 + ], + "type": "text", + "content": " pruning ratio indicates all tokens are dropped, i.e., we skip the corresponding steps and interpolate the intermediate outputs to the target resolution as the final outputs. Note that this extreme ratio depends on the pruned backbones, and we provide further discussion in the Suppl.. Unless specified, the efficiency of the unpruned baselines are already accelerated by FlashAttention [13], and we apply the proposed FastVAR on top of it. Following the setup of existing methods [10, 47, 50], the inference speeds for all methods are measured without including the VAE time cost since this is a shared cost for all methods. All experiments are conducted on one single NVIDIA RTX 3090 GPU with 24GB memory." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 491, + 400, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 491, + 400, + 503 + ], + "spans": [ + { + "bbox": [ + 314, + 491, + 400, + 503 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": "Comparison on GenEval. We first evaluate the quality-efficiency trade-off on " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " text-to-image generation using the GenEval benchmark [20]. We compare our FastVAR against various state-of-the-art methods, including Diffusion, AR, and VAR models. The results are shown in Tab. 1. Compared with traditional AR models, our Infinity+FastVAR can achieve even " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "39.7 \\times" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " speedup than LlamaGen [47], while achieving " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "125\\%" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " performance boosts on the GenEval. Moreover, our FastVAR achieves acceleration almost without performance loss, e.g., " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "1.5 \\times" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " speedup on the HART with the same GenEval score as the unpruned baseline. Our approach also reduces the inference GPU memory. For example, FastVAR achieves " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "22.2\\%" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " reduction to 14.7GB on top of FlashAttention of 18.9GB, to produce a " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " image in 0.95s, facilitating generation on consumer-level GPUs. The above results demonstrate the generality and effectiveness of our method." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19015" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 96, + 550, + 228 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Quantitative comparison on efficiency and quality on " + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "inline_equation", + "content": "{1024} \\times {1024}" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": " GenEval benchmarks. The marks " + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": " , " + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": " ,and " + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "inline_equation", + "content": "\\diamond" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": " denote the Diffusion,AR,and VAR-based methods,respectively. Note that the efficiency of HART and Infinity baselines are tested under FlashAttention." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 96, + 550, + 228 + ], + "lines": [ + { + "bbox": [ + 65, + 96, + 550, + 228 + ], + "spans": [ + { + "bbox": [ + 65, + 96, + 550, + 228 + ], + "type": "table", + "html": "
MethodsInference EfficiencyGeneration Quality
#Steps↓Speedup↑Latency↓Throughput↑#Param↓two_object↑position↑color_attr↑Overall↑
SDXL [41]40-4.3s0.23it/s2.6B0.740.150.230.55
PixArt-Sigma [8]20-2.7s2.50it/s0.6B0.620.140.270.55
SD3-medium [16]28-4.4s3.45it/s2.0B0.740.340.360.62
○LlamaGen [47]1024-37.7s0.03it/s0.8B0.340.070.040.32
○Show-o [59]1024-50.3s0.02it/s1.3B0.800.310.500.68
HART [48]141.0×0.95s1.05it/s0.7B0.620.130.180.51
+FastVAR141.5×0.63s1.59it/s0.7B0.570.160.240.51
Infinity [22]131.0×2.61s0.38it/s2.0B0.850.440.530.73
+FastVAR132.7×0.95s1.05it/s2.0B0.810.450.520.72
", + "image_path": "303101b09a27c9c10096c0527fd1f9b91d178cccd6951b4c87c2c09797ecf125.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 61, + 236, + 549, + 479 + ], + "blocks": [ + { + "bbox": [ + 61, + 236, + 549, + 479 + ], + "lines": [ + { + "bbox": [ + 61, + 236, + 549, + 479 + ], + "spans": [ + { + "bbox": [ + 61, + 236, + 549, + 479 + ], + "type": "image", + "image_path": "ca3246aafd90a0ded831663d6ab6ee845d1373f41b746f038973bc2887e5f6f5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 482, + 555, + 506 + ], + "lines": [ + { + "bbox": [ + 55, + 482, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 482, + 555, + 506 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison between the original baselines and our proposed FastVAR on " + }, + { + "bbox": [ + 55, + 482, + 555, + 506 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 55, + 482, + 555, + 506 + ], + "type": "text", + "content": " image generation. Our FastVAR achieves significant speedups, e.g., " + }, + { + "bbox": [ + 55, + 482, + 555, + 506 + ], + "type": "inline_equation", + "content": "2.7 \\times" + }, + { + "bbox": [ + 55, + 482, + 555, + 506 + ], + "type": "text", + "content": " on Infinity [22], while keeping high-quality results similar to the original model." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 517, + 295, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 295, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 295, + 661 + ], + "type": "text", + "content": "Comparison on MJHQ30K. In Tab. 2, we further validate the perceptual quality on the MJHQ30K [29] benchmark. It can be seen that our FastVAR achieves a reasonable performance while maintaining a high speedup ratio. For instance, on the well-known challenging \"people\" category, our HART+FastVAR even achieves a FID reduction of 2.42 with " + }, + { + "bbox": [ + 55, + 517, + 295, + 661 + ], + "type": "inline_equation", + "content": "1.5 \\times" + }, + { + "bbox": [ + 55, + 517, + 295, + 661 + ], + "type": "text", + "content": " acceleration. In other categories, our method also maintains good performance with significant acceleration. We also give a qualitative visualization in Fig. 5, and one can see that the generated images from FastVAR maintain exceptionally high quality and accurate semantic information, indicating the effectiveness of the proposed method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Difference from ToMe. Token Merging (ToMe) [3], which merges multiple tokens into one token for efficient forward followed by token unmerging to restore the 2D shape, appears applicable to accelerate VAR backbones due to its to-" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 323, + 538, + 545, + 622 + ], + "blocks": [ + { + "bbox": [ + 313, + 515, + 555, + 537 + ], + "lines": [ + { + "bbox": [ + 313, + 515, + 555, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 515, + 555, + 537 + ], + "type": "text", + "content": "Table 2. Quantitative comparisons of FID and CLIP score with different generation categories on the MJHQ30K dataset." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 538, + 545, + 622 + ], + "lines": [ + { + "bbox": [ + 323, + 538, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 323, + 538, + 545, + 622 + ], + "type": "table", + "html": "
MethodsSpeeduplandscapepeoplefood
FID↓CLIP↑FID↓CLIP↑FID↓CLIP↑
SDXL [41]-30.7826.3535.5628.0135.2627.98
HART [48]1.0×25.4326.8230.6128.4730.3728.03
+ FastVAR1.5×22.5226.5128.1928.3430.9728.25
Infinity [22]1.0×24.6826.6230.2727.8231.5526.66
+ FastVAR2.7×24.6826.6230.5528.2832.5427.08
", + "image_path": "76429357bd40d645a0983f2ba83f45456508211724e43a144241b55bf76dd7cf.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": "ken number invariance. In Tab. 3, we set different FastVAR and ToMe variants using varying pruning ratios. One can see that ToMe fails to achieve high speedup. For example, even the " + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "inline_equation", + "content": "1.36 \\times" + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": " speedup can lead to noticeable FID degradation since it is difficult to compress the whole token map into limited tokens. In contrast, our FastVAR can achieve " + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "inline_equation", + "content": "1.7 \\times" + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": " higher speedup with better FID performance." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19016" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 70, + 185, + 174 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 185, + 174 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 185, + 174 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 185, + 174 + ], + "type": "image", + "image_path": "95d59ba43b5936222d43f5d4b119ef4b9fc8a663b998853ec6e0a833a5ad1c13.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 175, + 555, + 198 + ], + "lines": [ + { + "bbox": [ + 55, + 175, + 555, + 198 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 555, + 198 + ], + "type": "text", + "content": "Figure 6. Quantitative and qualitative ablation experiments on different pruning ratios. We set varying pruning ratios on the last scale step of the HART model while keeping the other scale steps as the original setup." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 190, + 70, + 320, + 174 + ], + "blocks": [ + { + "bbox": [ + 190, + 70, + 320, + 174 + ], + "lines": [ + { + "bbox": [ + 190, + 70, + 320, + 174 + ], + "spans": [ + { + "bbox": [ + 190, + 70, + 320, + 174 + ], + "type": "image", + "image_path": "8afb013a4243a68dbfbd5ea9c511585290e78de6618b0c219fbeb6bdc3c136b2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 324, + 70, + 545, + 173 + ], + "blocks": [ + { + "bbox": [ + 324, + 70, + 545, + 173 + ], + "lines": [ + { + "bbox": [ + 324, + 70, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 324, + 70, + 545, + 173 + ], + "type": "image", + "image_path": "fe0218ae1536f8b1422222173a463b64a4ccfebc5c3c28d9f407643be9810b49.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 67, + 231, + 286, + 297 + ], + "blocks": [ + { + "bbox": [ + 55, + 208, + 295, + 231 + ], + "lines": [ + { + "bbox": [ + 55, + 208, + 295, + 231 + ], + "spans": [ + { + "bbox": [ + 55, + 208, + 295, + 231 + ], + "type": "text", + "content": "Table 3. Quantitative comparison results with ToMe [2] under different speedup settings." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 231, + 286, + 297 + ], + "lines": [ + { + "bbox": [ + 67, + 231, + 286, + 297 + ], + "spans": [ + { + "bbox": [ + 67, + 231, + 286, + 297 + ], + "type": "table", + "html": "
MethodsSpeedup↑Latency↓Throughput↑FID↓CLIP↑GenEval↑
HART [48]1.00×0.95s1.05it/s30.6128.470.51
+ ToMe [2]1.19×0.80s1.25it/s29.0728.330.48
+ ToMe [2]1.36×0.70s1.43it/s35.2227.990.46
+ FastVAR1.51×0.63s1.59it/s28.1928.340.51
+ FastVAR1.70×0.56s1.79it/s28.9728.240.50
", + "image_path": "e40f35637c23aeeaf20f9445cc4c4160f2897465b40b04a7f8accbfaeb31cbfd.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 66, + 355, + 285, + 428 + ], + "blocks": [ + { + "bbox": [ + 55, + 309, + 296, + 354 + ], + "lines": [ + { + "bbox": [ + 55, + 309, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 309, + 296, + 354 + ], + "type": "text", + "content": "Table 4. Ablation experiments on the scale-wise sensitivity. We focus on the scale range [16,21,27,36,48,64] with every two consecutive scales as a pruning group. The FastVAR is applied to the selected group while keeping the others unpruned." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 355, + 285, + 428 + ], + "lines": [ + { + "bbox": [ + 66, + 355, + 285, + 428 + ], + "spans": [ + { + "bbox": [ + 66, + 355, + 285, + 428 + ], + "type": "table", + "html": "
scalesratioSpeedup↑Latency↓FID↓CLIP↑GenEval↑
baseline01.00×0.95s30.6128.470.51
[16, 21]75%1.01×0.94s34.2728.430.48
[27, 36]50%1.09×0.87s27.9228.490.51
[27, 36]75%1.14×0.83s29.3828.560.50
[48, 64]50%1.30×0.73s27.1328.430.51
[48, 64]75%1.64×0.58s28.5628.360.50
", + "image_path": "0b2713db4625cd506a8a1a5c75b88a12d0cf6327becca45eb0a113cad31a45c6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 443, + 156, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 443, + 156, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 156, + 455 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 461, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 461, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 54, + 461, + 295, + 653 + ], + "type": "text", + "content": "Different Pruning Ratios. The pruning ratio plays an important role in balancing between efficiency and performance. Here, we conduct ablation experiments to investigate the impact of different pruning ratios. Both the quantitative and qualitative results are given in Fig. 6. Intuitively, increasing the pruning ratio leads to a stable runtime reduction since the model only needs to process fewer tokens. However, an over-large pruning ratio can also bring performance degradation due to the information loss of pivotal high-frequency tokens. Furthermore, the quantitative visualization suggests that increasing the pruning ratio makes certain textures and details discontinuous since the cache from the previous scale is not always an optimal approximation of the pruned tokens. Therefore, a moderate pruning ratio, e.g., " + }, + { + "bbox": [ + 54, + 461, + 295, + 653 + ], + "type": "inline_equation", + "content": "40\\% -75\\%" + }, + { + "bbox": [ + 54, + 461, + 295, + 653 + ], + "type": "text", + "content": ", is adopted for HART+FastVAR to strike a sweet spot between performance and efficiency." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 653, + 296, + 714 + ], + "type": "text", + "content": "Scale-wise Sensitivity. As discussed in Sec. 3.2, we only focus on the token pruning of the last few scale steps. Here, we conduct ablation experiments in Tab. 4 to justify the rationality. It can be seen that the acceleration from pruning at an early scale steps is very limited due to the small token" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 208, + 550, + 266 + ], + "blocks": [ + { + "bbox": [ + 316, + 208, + 550, + 266 + ], + "lines": [ + { + "bbox": [ + 316, + 208, + 550, + 266 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 550, + 266 + ], + "type": "image", + "image_path": "1b68d0fa6560028ce10e49ad4ae2ce9f7fbe612e2597aae4face0ac9d5b69d86.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 317, + 266, + 550, + 315 + ], + "blocks": [ + { + "bbox": [ + 317, + 266, + 550, + 315 + ], + "lines": [ + { + "bbox": [ + 317, + 266, + 550, + 315 + ], + "spans": [ + { + "bbox": [ + 317, + 266, + 550, + 315 + ], + "type": "image", + "image_path": "ba02d2ee399a712ff23a0a09bbbbc5393e13c2ea3edc938e7c8d71cf3ee0c2c6.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 317, + 545, + 329 + ], + "lines": [ + { + "bbox": [ + 322, + 317, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 322, + 317, + 545, + 329 + ], + "type": "text", + "content": "Figure 7. Visualization of pruned tokens with different ratios." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 342, + 555, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 461 + ], + "type": "text", + "content": "map size. For example, even reducing " + }, + { + "bbox": [ + 313, + 342, + 555, + 461 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 342, + 555, + 461 + ], + "type": "text", + "content": " tokens at the 48 and 64 scale steps is significantly faster than that of " + }, + { + "bbox": [ + 313, + 342, + 555, + 461 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 313, + 342, + 555, + 461 + ], + "type": "text", + "content": " at 16 and 21 scale steps. In addition, pruning at early scales also leads to a noticeable performance drop on both perceptual quality and semantic consistency. For instance, pruning at 48 and 64 scale steps brings 5.71 lower FID compared to that of 16 and 21 scale steps. This is because token reduction at small-scale steps disrupts the structure construction stage, and the inaccurate subject structure can exacerbate errors in subsequent scale steps." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 461, + 556, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 556, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 556, + 629 + ], + "type": "text", + "content": "Visualization of Pruned Slots. In Fig. 7, we visualize the selected TopK pivotal tokens under different pruning ratios. For a given pruning ratio, e.g., " + }, + { + "bbox": [ + 313, + 461, + 556, + 629 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 461, + 556, + 629 + ], + "type": "text", + "content": ", our FastVAR prioritizes the retention of high-frequency edges or texture slots, and removes tokens from flat regions that have already converged at early scale steps. Furthermore, the importance order within the selected tokens is also meaningful. For example, in the first row, when increasing the pruning ratio from " + }, + { + "bbox": [ + 313, + 461, + 556, + 629 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 461, + 556, + 629 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 461, + 556, + 629 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 313, + 461, + 556, + 629 + ], + "type": "text", + "content": ", the human eyes, hair and mouth regions, which are the most detailed, are consistently preserved while the low-frequency cheek regions are gradually pruned. Given that our FastVAR is gradient-free, this quantitative result confirms that the proposed frequency-based pivotal token selection is a reasonable heuristic." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 636, + 388, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 636, + 388, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 636, + 388, + 647 + ], + "type": "text", + "content": "4.4. Discussion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "content": "In Conjunction with FlashAttention. Existing token pruning methods [18, 19, 58, 63, 65] often depend on the attention map for token importance, leading to difficulties in combination with FlashAttention [13] in which the attention map is inaccessible. In comparison, our FastVAR" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19017" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 192, + 247 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 192, + 247 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 192, + 247 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 192, + 247 + ], + "type": "image", + "image_path": "b321a9722cc7bb24a63b7df84e1787542d633788fb0fa5de757c3a20c69497d9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 194, + 70, + 369, + 247 + ], + "blocks": [ + { + "bbox": [ + 194, + 70, + 369, + 247 + ], + "lines": [ + { + "bbox": [ + 194, + 70, + 369, + 247 + ], + "spans": [ + { + "bbox": [ + 194, + 70, + 369, + 247 + ], + "type": "image", + "image_path": "96941d5de7ea757e77f7404198a6d0d3145848b84ade08f06c3418fa34a01af3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 370, + 70, + 458, + 247 + ], + "blocks": [ + { + "bbox": [ + 370, + 70, + 458, + 247 + ], + "lines": [ + { + "bbox": [ + 370, + 70, + 458, + 247 + ], + "spans": [ + { + "bbox": [ + 370, + 70, + 458, + 247 + ], + "type": "image", + "image_path": "4cc7f8454fd293b35cfc53f41e94a8b905adbb4baf016501f225b4405feff53c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 460, + 70, + 550, + 247 + ], + "blocks": [ + { + "bbox": [ + 460, + 70, + 550, + 247 + ], + "lines": [ + { + "bbox": [ + 460, + 70, + 550, + 247 + ], + "spans": [ + { + "bbox": [ + 460, + 70, + 550, + 247 + ], + "type": "image", + "image_path": "6635a24e83fcf8de2d4544eca05daa876b66e9a1d10467262244e5d3cb69d40e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 58, + 249, + 428, + 406 + ], + "blocks": [ + { + "bbox": [ + 58, + 249, + 428, + 406 + ], + "lines": [ + { + "bbox": [ + 58, + 249, + 428, + 406 + ], + "spans": [ + { + "bbox": [ + 58, + 249, + 428, + 406 + ], + "type": "image", + "image_path": "2e3b1a0d29fff03fb5a775037489a2ba8061355577a03fbc6a62d2b8577df01a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 410, + 555, + 433 + ], + "lines": [ + { + "bbox": [ + 55, + 410, + 555, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 555, + 433 + ], + "type": "text", + "content": "Figure 8. We extend the scale schedule of pre-trained VAR model [22] for zero-shot larger resolution generation. We chose a modest extended scale since a larger one is observed quality drop due to the resolution gap. The images are re-scaled for better presentation." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 431, + 249, + 549, + 406 + ], + "blocks": [ + { + "bbox": [ + 431, + 249, + 549, + 406 + ], + "lines": [ + { + "bbox": [ + 431, + 249, + 549, + 406 + ], + "spans": [ + { + "bbox": [ + 431, + 249, + 549, + 406 + ], + "type": "image", + "image_path": "6bff136e74124efb6b958946717949c6dcc432b26b39c279cb7df56138b7edb6.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 444, + 295, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 444, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 444, + 295, + 540 + ], + "type": "text", + "content": "can be seamlessly integrated with FlashAttention to facilitate more efficient inference. In Tab. 5, we show that using the FastVAR-only can obtain even " + }, + { + "bbox": [ + 55, + 444, + 295, + 540 + ], + "type": "inline_equation", + "content": "2.1 \\times" + }, + { + "bbox": [ + 55, + 444, + 295, + 540 + ], + "type": "text", + "content": " speedup against FlashAttention only, with " + }, + { + "bbox": [ + 55, + 444, + 295, + 540 + ], + "type": "inline_equation", + "content": "20.4\\%" + }, + { + "bbox": [ + 55, + 444, + 295, + 540 + ], + "type": "text", + "content": " GPU memory reduction. After combining both, the resulting version leads to " + }, + { + "bbox": [ + 55, + 444, + 295, + 540 + ], + "type": "inline_equation", + "content": "2.7 \\times" + }, + { + "bbox": [ + 55, + 444, + 295, + 540 + ], + "type": "text", + "content": " speedup than FlashAttention only. This experiment validates that our FastVAR can be orthogonally combined with other methods for further acceleration." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "Towards Larger Resolution. As presented in Sec. 3.2, the original VAR struggles to scale to higher-resolution due to the increasing token numbers. Our FastVAR can facilitate efficient resolution scaling, offering the potential to generate images at a larger resolution. To this end, we apply our FastVAR to the Infinity [22] for zero-shot larger resolution synthesis by appending additional steps. The results are shown in Fig. 8. It is noteworthy that even the FlashAttention accelerated baseline is out of memory on a 24GB NVIDIA 3090 GPU. In contrast, our FastVAR costs only 15GB memory and 1.3s runtime to generate a " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "1344 \\times 1344" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": " image. Moreover, the generated images are of high quality, suggesting that our FastVAR can facilitate production-level image generation on consumer-level GPUs." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 319, + 466, + 549, + 525 + ], + "blocks": [ + { + "bbox": [ + 313, + 443, + 555, + 465 + ], + "lines": [ + { + "bbox": [ + 313, + 443, + 555, + 465 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 555, + 465 + ], + "type": "text", + "content": "Table 5. Efficiency comparison with FlashAttn on " + }, + { + "bbox": [ + 313, + 443, + 555, + 465 + ], + "type": "inline_equation", + "content": "1024\\times 1024" + }, + { + "bbox": [ + 313, + 443, + 555, + 465 + ], + "type": "text", + "content": " image generation." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 466, + 549, + 525 + ], + "lines": [ + { + "bbox": [ + 319, + 466, + 549, + 525 + ], + "spans": [ + { + "bbox": [ + 319, + 466, + 549, + 525 + ], + "type": "table", + "html": "
setupSpeedup↑Latency↓Throughput↑Memory↓
SlowAttn [53]---OOM
FlashAttn only [13]1.0×2.61s0.38it/s16.1GB
SlowAttn+FastVAR2.1×1.25s0.80it/s12.8GB
FlashAttn+FastVAR2.7×0.95s1.05it/s11.9GB
", + "image_path": "547754d69a4a32042a7511ba28e9be002ec18e4c6f512a9885914fde5b585b65.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 536, + 388, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 388, + 549 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 388, + 549 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 557, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 557, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 557, + 555, + 713 + ], + "type": "text", + "content": "This work presents FastVAR to advance high-resolution image synthesis with VARs by addressing the resolution scaling challenge. We reveal that the main latency bottleneck is the large-scale steps at which the low-frequency tokens have almost converged. To this end, we develop the cached token pruning technique to allow only process pivotal high-frequency tokens and use caches from the previous scale steps to restore the original token numbers. Thanks to the reduced forwarded tokens, our FastVAR approximates linear complexity, further enabling generation at a larger resolution such as 2K. Extensive experiments validate our FastVAR as an effective and generic solution for efficient resolution scaling of VAR models." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19018" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "type": "text", + "content": "This work is supported in part by the National Natural Science Foundation of China, under Grant (62302309, 62171248), Shenzhen Science and Technology Program (JCYJ20220818101014030, JCYJ20220818101012025)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 115, + 165 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 115, + 165 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 115, + 165 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 171, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 171, + 296, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 171, + 296, + 227 + ], + "spans": [ + { + "bbox": [ + 61, + 171, + 296, + 227 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. GPT-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 228, + 296, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 228, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 61, + 228, + 296, + 251 + ], + "type": "text", + "content": "[2] Daniel Bolya and Judy Hoffman. Token merging for fast stable diffusion. In CVPR, pages 4599-4603, 2023. 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 251, + 296, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 251, + 296, + 295 + ], + "spans": [ + { + "bbox": [ + 61, + 251, + 296, + 295 + ], + "type": "text", + "content": "[3] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token Merging: Your ViT but faster. arXiv preprint arXiv:2210.09461, 2022. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 297, + 296, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 297, + 296, + 351 + ], + "spans": [ + { + "bbox": [ + 62, + 297, + 296, + 351 + ], + "type": "text", + "content": "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 33:1877-1901, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 353, + 296, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 353, + 296, + 397 + ], + "spans": [ + { + "bbox": [ + 62, + 353, + 296, + 397 + ], + "type": "text", + "content": "[5] Tianle Cai, Yuhong Li, Zhengyang Geng, Hongwu Peng, Jason D Lee, Deming Chen, and Tri Dao. Medusa: Simple LLM inference acceleration framework with multiple decoding heads. arXiv preprint arXiv:2401.10774, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 399, + 296, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 399, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 62, + 399, + 296, + 453 + ], + "type": "text", + "content": "[6] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 455, + 296, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 455, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 62, + 455, + 296, + 510 + ], + "type": "text", + "content": "[7] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. PixArt-Alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 512, + 296, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 512, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 62, + 512, + 296, + 567 + ], + "type": "text", + "content": "[8] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. PixArt-Sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. In ECCV, pages 74-91. Springer, 2024. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 568, + 296, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 568, + 296, + 610 + ], + "spans": [ + { + "bbox": [ + 62, + 568, + 296, + 610 + ], + "type": "text", + "content": "[9] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In ICML, pages 1691-1703. PMLR, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 613, + 296, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 296, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 296, + 656 + ], + "type": "text", + "content": "[10] Zigeng Chen, Xinyin Ma, Gongfan Fang, and Xinchao Wang. Collaborative decoding makes visual auto-regressive modeling efficient. arXiv preprint arXiv:2411.17787, 2024. 2, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "type": "text", + "content": "[11] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instructionfinetuned language models. Journal of Machine Learning Research, 25(70):1-53, 2024. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "text", + "content": "[12] Tri Dao. FlashAttention-2: Faster attention with better parallelism and work partitioning. In ICLR, 2024. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 98, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 98, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 98, + 553, + 140 + ], + "type": "text", + "content": "[13] Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flash Attention: Fast and memory-efficient exact attention with IO-awareness. In NeurIPS, 2022. 1, 5, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 143, + 553, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 143, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 316, + 143, + 553, + 186 + ], + "type": "text", + "content": "[14] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In ACL, pages 4171–4186, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 189, + 553, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 189, + 553, + 221 + ], + "spans": [ + { + "bbox": [ + 317, + 189, + 553, + 221 + ], + "type": "text", + "content": "[15] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, pages 12873-12883, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 223, + 553, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 223, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 316, + 223, + 553, + 277 + ], + "type": "text", + "content": "[16] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 280, + 553, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 280, + 553, + 302 + ], + "spans": [ + { + "bbox": [ + 316, + 280, + 553, + 302 + ], + "type": "text", + "content": "[17] Gongfan Fang, Xinyin Ma, and Xinchao Wang. Structural pruning for diffusion models. In NeurIPS, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 304, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 553, + 348 + ], + "type": "text", + "content": "[18] Qichen Fu, Minsik Cho, Thomas Merth, Sachin Mehta, Mohammad Rastegari, and Mahyar Najibi. LazyLLM: Dynamic token pruning for efficient long context llm inference. arXiv preprint arXiv:2407.14057, 2024. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "text", + "content": "[19] Suyu Ge, Yunan Zhang, Liyuan Liu, Minjia Zhang, Jiawei Han, and Jianfeng Gao. Model tells you what to discard: Adaptive KV cache compression for LLMs. arXiv preprint arXiv:2310.01801, 2023. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 395, + 553, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 395, + 553, + 428 + ], + "spans": [ + { + "bbox": [ + 316, + 395, + 553, + 428 + ], + "type": "text", + "content": "[20] Dhruba Ghosh, Hannaneh Hajishirzi, and Ludwig Schmidt. GenEval: An object-focused framework for evaluating text-to-image alignment. NeurIPS, 36, 2024. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 430, + 553, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 430, + 553, + 463 + ], + "spans": [ + { + "bbox": [ + 316, + 430, + 553, + 463 + ], + "type": "text", + "content": "[21] Hang Guo, Yawei Li, Tao Dai, Shu-Tao Xia, and Luca Benini. Intlora: Integral low-rank adaptation of quantized diffusion models. arXiv preprint arXiv:2410.21759, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 464, + 553, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 464, + 553, + 518 + ], + "spans": [ + { + "bbox": [ + 317, + 464, + 553, + 518 + ], + "type": "text", + "content": "[22] Jian Han, Jinlai Liu, Yi Jiang, Bin Yan, Yuqi Zhang, Zehuan Yuan, Bingyue Peng, and Xiaobing Liu. Infinity: Scaling bitwise autoregressive modeling for high-resolution image synthesis. arXiv preprint arXiv:2412.04431, 2024. 1, 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 520, + 553, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 520, + 553, + 565 + ], + "spans": [ + { + "bbox": [ + 316, + 520, + 553, + 565 + ], + "type": "text", + "content": "[23] Yefei He, Feng Chen, Yuanyu He, Shaoxuan He, Hong Zhou, Kaipeng Zhang, and Bohan Zhuang. ZIPAR: Accelerating autoregressive image generation through spatial locality. arXiv preprint arXiv:2412.04062, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 567, + 553, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 553, + 621 + ], + "type": "text", + "content": "[24] Tom Henighan, Jared Kaplan, Mor Katz, Mark Chen, Christopher Hesse, Jacob Jackson, Heewoo Jun, Tom B Brown, Prafulla Dhariwal, Scott Gray, et al. Scaling laws for autoregressive generative modeling. arXiv preprint arXiv:2010.14701, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 624, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 656 + ], + "type": "text", + "content": "[25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. *NeurIPS*, 33:6840-6851, 2020. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[26] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361, 2020. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19019" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "text", + "content": "[27] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. In CVPR, pages 11523-11532, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 117, + 294, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 117, + 294, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 117, + 294, + 149 + ], + "type": "text", + "content": "[28] Yaniv Leviathan, Matan Kalman, and Yossi Matias. Fast inference from transformers via speculative decoding. In ICML, pages 19274-19286. PMLR, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "type": "text", + "content": "[29] Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Linmiao Xu, and Suhail Doshi. Playground v2.5: Three insights towards enhancing aesthetic quality in text-to-image generation. arXiv preprint arXiv:2402.17245, 2024. 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 195, + 294, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 294, + 237 + ], + "type": "text", + "content": "[30] Senmao Li, Taihang Hu, Fahad Shahbaz Khan, Linxuan Li, Shiqi Yang, Yaxing Wang, Ming-Ming Cheng, and Jian Yang. Faster Diffusion: Rethinking the role of unet encoder in diffusion models. NeurIPS, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 294, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 281 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 281 + ], + "type": "text", + "content": "[31] Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization. arXiv preprint arXiv:2406.11838, 2024. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 283, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 294, + 326 + ], + "type": "text", + "content": "[32] Xiuyu Li, Yijiang Liu, Long Lian, Huanrui Yang, Zhen Dong, Daniel Kang, Shanghang Zhang, and Kurt Keutzer. Q-Diffusion: Quantizing diffusion models. In ICCV, pages 17535-17545, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 327, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 381 + ], + "type": "text", + "content": "[33] Dongyang Liu, Shitian Zhao, Le Zhuo, Weifeng Lin, Yu Qiao, Hongsheng Li, and Peng Gao. Lumina-mgpt: Illuminate flexible photorealistic text-to-image generation with multimodal generative pretraining. arXiv preprint arXiv:2408.02657, 2024. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 383, + 294, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 436 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 436 + ], + "type": "text", + "content": "[34] Haozhe Liu, Wentian Zhang, Jinheng Xie, Francesco Faccio, Mengmeng Xu, Tao Xiang, Mike Zheng Shou, Juan-Manuel Perez-Rua, and Jurgen Schmidhuber. Faster diffusion via temporal attention decomposition. arXiv preprint arXiv:2404.02747, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 437, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 491 + ], + "type": "text", + "content": "[35] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 493, + 294, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 294, + 535 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 294, + 535 + ], + "type": "text", + "content": "[36] Xinyin Ma, Gongfan Fang, Michael Bi Mi, and Xinchao Wang. Learning-to-Cache: Accelerating diffusion transformer via layer caching. arXiv preprint arXiv:2406.01733, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 537, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 568 + ], + "type": "text", + "content": "[37] Xinyin Ma, Gongfan Fang, and Xinchao Wang. DeepCache: Accelerating diffusion models for free. In CVPR, pages 15762-15772, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 570, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 294, + 613 + ], + "type": "text", + "content": "[38] Yue Ma, Yingqing He, Xiaodong Cun, Xintao Wang, Siran Chen, Xiu Li, and Qifeng Chen. Follow your pose: Pose-guided text-to-video generation using pose-free videos. In AAAI, pages 4117-4125, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 615, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 615, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 615, + 294, + 669 + ], + "type": "text", + "content": "[39] Yue Ma, Hongyu Liu, Hongfa Wang, Heng Pan, Yingqing He, Junkun Yuan, Ailing Zeng, Chengfei Cai, Heung-Yeung Shum, Wei Liu, et al. Follow-your-emoji: Fine-controllable and expressive freestyle portrait animation. In SIGGRAPH Asia, pages 1–12, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[40] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. In CVPR, pages 14297-14306, 2023. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[41] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 162 + ], + "type": "text", + "content": "[42] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with VQ-VAE-2. NeurIPS, 32, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 164, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 164, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 164, + 553, + 207 + ], + "type": "text", + "content": "[43] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 209, + 553, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 243 + ], + "type": "text", + "content": "[44] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 245, + 553, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 245, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 316, + 245, + 553, + 299 + ], + "type": "text", + "content": "[45] Andrea Santilli, Silvio Severino, Emilian Postolache, Valentino Maiorca, Michele Mancusi, Riccardo Marin, and Emanuele Rodola. Accelerating transformer inference for translation via parallel decoding. arXiv preprint arXiv:2305.10427, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 301, + 553, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 301, + 553, + 334 + ], + "spans": [ + { + "bbox": [ + 316, + 301, + 553, + 334 + ], + "type": "text", + "content": "[46] Yuzhang Shang, Zhihang Yuan, Bin Xie, Bingzhe Wu, and Yan Yan. Post-training quantization on diffusion models. In CVPR, pages 1972-1981, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 335, + 553, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 335, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 316, + 335, + 553, + 380 + ], + "type": "text", + "content": "[47] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 1, 2, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 382, + 553, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 382, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 316, + 382, + 553, + 436 + ], + "type": "text", + "content": "[48] Haotian Tang, Yecheng Wu, Shang Yang, Enze Xie, Junsong Chen, Junyu Chen, Zhuoyang Zhang, Han Cai, Yao Lu, and Song Han. HART: Efficient visual generation with hybrid autoregressive transformer. arXiv preprint arXiv:2410.10812, 2024. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 438, + 553, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 553, + 492 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 553, + 492 + ], + "type": "text", + "content": "[49] Yao Teng, Han Shi, Xian Liu, Xuefei Ning, Guohao Dai, Yu Wang, Zhenguo Li, and Xihui Liu. Accelerating autoregressive text-to-image generation with training-free speculative jacobi decoding. arXiv preprint arXiv:2410.01699, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 495, + 553, + 538 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 495, + 553, + 538 + ], + "spans": [ + { + "bbox": [ + 316, + 495, + 553, + 538 + ], + "type": "text", + "content": "[50] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. NeurIPS, 2024. 1, 2, 3, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 540, + 553, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 540, + 553, + 574 + ], + "spans": [ + { + "bbox": [ + 316, + 540, + 553, + 574 + ], + "type": "text", + "content": "[51] Aaron Van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Conditional image generation with pixelCNN decoders. NeurIPS, 29, 2016. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 575, + 553, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 575, + 553, + 597 + ], + "spans": [ + { + "bbox": [ + 316, + 575, + 553, + 597 + ], + "type": "text", + "content": "[52] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. NeurIPS, 30, 2017. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 600, + 553, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 600, + 553, + 620 + ], + "spans": [ + { + "bbox": [ + 316, + 600, + 553, + 620 + ], + "type": "text", + "content": "[53] A Vaswani. Attention is all you need. NeurIPS, 2017. 1, 3, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 623, + 553, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 623, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 316, + 623, + 553, + 666 + ], + "type": "text", + "content": "[54] Hongjie Wang, Difan Liu, Yan Kang, Yijun Li, Zhe Lin, Niraj K Jha, and Yuchen Liu. Attention-driven training-free efficiency enhancement of diffusion models. In CVPR, pages 16080-16089, 2024. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "text", + "content": "[55] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. EMU3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 1, 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19020" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 574 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[56] Chengyue Wu, Xiaokang Chen, Zhiyu Wu, Yiyang Ma, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, Chong Ruan, et al. Janus: Decoupling visual encoding for unified multimodal understanding and generation. arXiv preprint arXiv:2410.13848, 2024." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "type": "text", + "content": "[57] Yecheng Wu, Zhuoyang Zhang, Junyu Chen, Haotian Tang, Dacheng Li, Yunhao Fang, Ligeng Zhu, Enze Xie, Hongxu Yin, Li Yi, et al. VILA-U: A unified foundation model integrating visual understanding and generation. arXiv preprint arXiv:2409.04429, 2024. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 217 + ], + "type": "text", + "content": "[58] Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. Efficient streaming language models with attention sinks. arXiv preprint arXiv:2309.17453, 2023. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 282 + ], + "type": "text", + "content": "[59] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 1, 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 285, + 294, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 285, + 294, + 328 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 294, + 328 + ], + "type": "text", + "content": "[60] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 294, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 384 + ], + "type": "text", + "content": "[61] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved VQGAN. arXiv preprint arXiv:2110.04627, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 386, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 386, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 386, + 294, + 430 + ], + "type": "text", + "content": "[62] Evelyn Zhang, Bang Xiao, Jiayi Tang, Qianli Ma, Chang Zou, Xuefei Ning, Xuming Hu, and Linfeng Zhang. Token pruning for caching better: 9 times acceleration on stable diffusion for free. arXiv preprint arXiv:2501.00375, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 431, + 294, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 431, + 294, + 485 + ], + "spans": [ + { + "bbox": [ + 56, + 431, + 294, + 485 + ], + "type": "text", + "content": "[63] Zhenyu Zhang, Ying Sheng, Tianyi Zhou, Tianlong Chen, Lianmin Zheng, Ruisi Cai, Zhao Song, Yuandong Tian, Christopher Ré, Clark Barrett, et al. H2O: Heavy-hitter oracle for efficient generative inference of large language models. NeurIPS, 36:34661-34710, 2023. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 487, + 294, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 487, + 294, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 487, + 294, + 529 + ], + "type": "text", + "content": "[64] Chuanxia Zheng, Tung-Long Vuong, Jianfei Cai, and Dinh Phung. MOVQ: Modulating quantized vectors for high-fidelity image generation. NeurIPS, 35:23412-23425, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 532, + 294, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 532, + 294, + 574 + ], + "spans": [ + { + "bbox": [ + 56, + 532, + 294, + 574 + ], + "type": "text", + "content": "[65] Chang Zou, Xuyang Liu, Ting Liu, Siteng Huang, and Linfeng Zhang. Accelerating diffusion transformers with tokenwise feature caching. arXiv preprint arXiv:2410.05317, 2024. 2, 7" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19021" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Faster and Better 3D Splatting via Group Training/21033709-b705-4347-a1bb-c3e7a58591e2_content_list.json b/2025/Faster and Better 3D Splatting via Group Training/21033709-b705-4347-a1bb-c3e7a58591e2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9daed620508899e694f833c6374992f1d8f387a2 --- /dev/null +++ b/2025/Faster and Better 3D Splatting via Group Training/21033709-b705-4347-a1bb-c3e7a58591e2_content_list.json @@ -0,0 +1,1736 @@ +[ + { + "type": "text", + "text": "Faster and Better 3D Splatting via Group Training", + "text_level": 1, + "bbox": [ + 241, + 130, + 756, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chengbo Wang $^{1,2,4,5}$ Guozheng Ma $^{3}$ Yifei Xue $^{1,2,4,5}$ Yizhen Lao $^{1,4,5,\\dagger}$", + "bbox": [ + 169, + 178, + 823, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Design, Hunan University $^{2}$ DaHe.AI", + "bbox": [ + 300, + 198, + 697, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Nanyang Technological University $^{4}$ Lushan Innovation Lab", + "bbox": [ + 240, + 215, + 756, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "5 Key Laboratory of Digital Culture Smart Design Technology, Minstry of Culture and Tourism", + "bbox": [ + 124, + 233, + 874, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{wangchb, yizhenlao}@hnu.edu.cn, GUOZHENG001@e.ntu.edu.sg, iflyhsueh@gmail.com", + "bbox": [ + 150, + 253, + 841, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 303, + 326, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D Gaussian Splatting (3DGS) has emerged as a powerful technique for novel view synthesis, demonstrating remarkable capability in high-fidelity scene reconstruction through its Gaussian primitive representations. However, the computational overhead induced by the massive number of primitives poses a significant bottleneck to training efficiency. To overcome this challenge, we propose Group Training, a simple yet effective strategy that organizes Gaussian primitives into manageable groups, optimizing training efficiency and improving rendering quality. This approach shows universal compatibility with existing 3DGS frameworks, including vanilla 3DGS and MipSplatting, consistently achieving accelerated training while maintaining superior synthesis quality. Extensive experiments reveal that our straightforward Group Training strategy achieves up to $30\\%$ faster convergence and improved rendering quality across diverse scenarios. Project Website: https://chengbo-wang.github.io/3DGS-with-Group-Training/.", + "bbox": [ + 88, + 335, + 485, + 623 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 652, + 220, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Novel view synthesis (NVS) is a pivotal technology in various domains, such as virtual reality [11], augmented reality [34], and autonomous driving [36]. Recently, 3D Gaussian Splitting (3DGS) [12] achieved significant success in NVS due to its real-time [15, 29, 37], high-quality rendering [2, 20]. This high-quality rendering is made possible by using millions of Gaussians [3], each includes a complex set of attributes, namely, position, size, orientation, opacity, and color. These attributes are optimized via multi-view photometric losses.", + "bbox": [ + 89, + 678, + 483, + 830 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "- Motivation. However, 3DGS also introduces a potential risk during training, as the exponentially growing Gaussians", + "bbox": [ + 89, + 845, + 483, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e12c1d0c85d56fcb917d085b4e8d9aab3296458ada4367a1d0642994a03ee1ae.jpg", + "image_caption": [ + "(a) Reconstruction of the \"Stump\"[1] scene by 3DGS[12]." + ], + "image_footnote": [], + "bbox": [ + 526, + 313, + 743, + 523 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/040807b518210d2411686473d3735d33f8e24fe01daec70141c3a75e4cfc92d6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 746, + 316, + 883, + 521 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d2018b32da4126112037c08c7717a2bcd8c253eea18b944cd57237fe5cc2947f.jpg", + "image_caption": [ + "(b) Reconstruction of the \"Dr. Johnson\"[9] scene by Mip-Splatting[33]." + ], + "image_footnote": [], + "bbox": [ + 527, + 541, + 743, + 750 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/867ac9203a42177cd3afa097d1a40e833ac8f83acd33f602f97529f04d0a92f4.jpg", + "image_caption": [ + "Figure 1. Improvements and Illustrations of Applying Group Training in 3DGS[12] and Mip-Splatting[33]. Grouping Training (solid line) achieves a significant increase in reconstruction speed and superior scene quality compared to baseline methods (dashed line). In the \"Stump\"[1] scene, Grouping Training delivers a $28\\%$ faster reconstruction while rendering leaf details with greater clarity. In the \"Dr. Johnson\"[9] scene, Mip-Splatting with Group Training achieves a $33\\%$ reduction in reconstruction time and effectively suppresses the generation of floating artifacts." + ], + "image_footnote": [], + "bbox": [ + 746, + 541, + 882, + 744 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author: Yizhen Lao (yizhenlao@hnu.edu.cn)", + "bbox": [ + 114, + 886, + 478, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "27968", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "significantly increase the training burden. Consequently, a primary research focus is the development of methods that enable rapid and cost-effective 3D scene reconstruction without compromising the rendering performance.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "During the densification stage, the cyclic pruning of Gaussians with low opacity emerges as a straightforward yet potentially effective strategy [12, 28] for enhancing the training efficiency of 3DGS. Specifically, in each round, 3DGS prunes Gaussians whose opacity falls below a designated threshold $\\epsilon_{\\alpha}$ , which is a vital parameter that facilitates control over the number of generated Gaussians. However, this method has inherent limitations. If $\\epsilon_{\\alpha}$ is set too conservatively, the resultant acceleration will be minimal while an excessively lenient threshold defeats the quality of the NVS. Thus, achieving a balance between accelerating the training and preserving the rendering quality poses a considerable challenge.", + "bbox": [ + 89, + 152, + 482, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Intuition and contributions. Inspired by the limitations of the cyclic pruning strategy, we propose to cache a portion of the Gaussians instead of directly pruning Gaussians. This \"cache\" concept can retain the \"important\" Gaussians while reducing the total number of them. Therefore, we propose a novel Group Training strategy to accelerate the training of 3DGS. Specifically, (1) a portion of Gaussians is cyclically cached during training. These cached Gaussians are temporarily excluded from scene rendering and model training, which greatly reduces the number of Undertraining Gaussians and cuts down on training time. (2) Additionally, through cyclic resampling, the cached Gaussians are reintegrated into the 3DGS model, reducing the impact of direct pruning. (3) A crucial element of Group Training is determining how to sample Gaussians to distinguish between under-training and cached Gaussians. A naive sampling approach would be random sampling. Interestingly, we discovered that the distribution of Gaussian opacity values influences both the quantity of Gaussians generated during densification and the rendering speed of the scene. Therefore, we introduce an Opacity-based Prioritized Sampling method, which effectively reduces the generation of redundant Gaussians and improves the training speed of 3DGS. Specifically, the main technical contributions of our work are as follows:", + "bbox": [ + 91, + 363, + 482, + 738 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We propose a simple yet highly efficient training framework called Group Training for 3D Gaussian Splatting, which is seamlessly integrated into existing 3DGS frameworks, including both 3DGS and Mip-Splatting.", + "2. Our framework significantly improves the efficiency of scene reconstruction and the quality of NVS." + ], + "bbox": [ + 89, + 741, + 482, + 832 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 844, + 232, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we review relevant work in novel view synthesis and 3D Gaussian Splatting, followed by the distinct", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/20d53fa2901d726544c400c53d7c50e49d51ce9bc9539dcafdc29b85709d5880.jpg", + "image_caption": [ + "Figure 2. PSNR and Time Performance of Pruning [12, 28] and Group Training under Varying Hyperparameters. The reconstructions are performed on the \"Train\" scene [14] using 3DGS [12] as the baseline. Left: The pruning method exhibits substantial instability in optimizing the trade-off between reconstruction efficiency and quality. The hyperparameter sensitivity (Opacity Threshold) presents significant challenges for optimal parameter tuning. Right: Group Training demonstrates consistent improvements in both reconstruction speed and quality, with robust performance across a wide range of hyperparameter values (Caching Ratio), enabling straightforward parameter optimization." + ], + "image_footnote": [], + "bbox": [ + 516, + 92, + 709, + 284 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/18379785de7df5df1624c71a6500b7ffa871fb5b3af417abd41b1f7d782e2186.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 714, + 93, + 901, + 285 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tion of our approach.", + "bbox": [ + 511, + 468, + 653, + 483 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- NVS. NVS involves generating new views of a scene from different viewpoints than those in the original dataset. Neural Radiance Fields (NeRF) [21] have been pivotal in NVS for producing photo-realistic images through volume rendering. Neural Radiance Fields (NeRF) have significantly advanced NVS by producing photo-realistic images via volume rendering, but they require substantial computational resources and lengthy training due to costly MLP evaluations. To overcome these limitations, methods such as voxel grids [7], hash grids [23] and points [30] have been introduced to facilitate real-time NeRF rendering.", + "bbox": [ + 511, + 489, + 903, + 655 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, 3D Gaussian Splatting (3DGS) [12] has emerged as a promising alternative in the NVS field. Unlike the implicit structure of NeRF, 3DGS uses rasterization rather than ray tracing, enabling efficient rendering suitable for applications requiring both speed and image quality.", + "bbox": [ + 511, + 657, + 903, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- 3DGS. 3D Gaussian Splatting (3DGS) is a recent approach to high-quality, real-time novel view synthesis that represents scenes explicitly as 3D Gaussian primitives. By rasterizing these Gaussians into 2D splats and blending them using $\\alpha$ -blending [21], 3DGS produces detailed, high-resolution outputs, making it well-suited for applications that demand both speed and visual fidelity.", + "bbox": [ + 511, + 748, + 903, + 853 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The 3DGS training process involves two main stages: \"Densification\" and \"Optimization\". During the densification phase (0~15K iterations), Adaptive Density", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "27969", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/338431931c015b47ac68b48761db4d231d197858d06f616db538edde0aec3ac0.jpg", + "image_caption": [ + "Figure 3. The overall framework of Group at regular iteration intervals, Gaussians from are categorized into the Under-training Group the Under-training Group is utilized for Ga Group remains inactive and does not participate in any calculations." + ], + "image_footnote": [], + "bbox": [ + 117, + 89, + 996, + 292 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Control [12] is applied to increase the density of Gaussian primitives, thereby enhancing the model's capacity for detailed scene representation. In the optimization phase (15K~30K iterations), densification stops, and the existing Gaussians undergo global refinement, yielding a more precise scene depiction. Recent studies have driven significant advancements in 3DGS across several areas, notably improving rendering quality [2, 13, 33], memory efficiency [5, 25, 27], and scalability for large scenes [16]. Beyond these enhancements, additional research has focused on increasing the efficiency of 3DGS optimizers [10] and refining Gaussian Densification [18, 19, 31] and Pruning strategies [5, 6, 26]. These improvements yield more compact 3DGS models that support faster reconstruction but sacrificing detail.", + "bbox": [ + 88, + 325, + 483, + 553 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Distinction. Unlike previous optimization-focused approaches, we propose Group Training, a principled strategy that dynamically manages Gaussian primitives during training through controlled group sampling. This approach not only accelerates training but also improves reconstruction quality by maintaining effective primitive optimization, achieving a better balance between computational efficiency and visual fidelity.", + "bbox": [ + 89, + 568, + 483, + 689 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 694, + 225, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In Sec. 3.1, we will provide a comprehensive introduction to the Group Training method and its multifunctional advantages in 3D Gaussian Splatting (3DGS). Section 3.2 introduces two sampling paradigms: Random Sampling and Prioritized Sampling (incorporating importance score, opacity-based, and volume-based criteria), with comparative analysis of their computational efficiency and performance trade-offs. We mathematically demonstrate that Opacity-based Prioritized Sampling achieves effective 3DGS densification and efficient rendering. Finally, in Sec. 3.3, we summarize training characteristics of 3DGS with Group Training, and the Sampling Strategies are identified.", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7f2ded8c80e7931cf9c204c5b93eb9a9b38f142792d1b8cb7ccf609d1da010b8.jpg", + "image_caption": [ + "Figure 4. Schedule for activating Group Training during reconstruction. Group Training is enabled at regular intervals. Groups Merging (Grouping without resampling) occurring at iteration $I_{\\mathrm{D}}$ and $I_{\\mathrm{O}}$ before the completion of the densification and optimization processes for global densification and optimization." + ], + "image_footnote": [], + "bbox": [ + 526, + 324, + 895, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Group Training", + "text_level": 1, + "bbox": [ + 511, + 493, + 671, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The number of Gaussian primitives in the Gaussian set increases rapidly during the densification phase to achieve rich representation capabilities. This implies that the computational time to render a scene from a specific viewpoint will also increase, and this effect will extend to the global optimisation phase. To this end, we employ the Group Training method during training.", + "bbox": [ + 511, + 515, + 906, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For any given set of Gaussian $G$ , the Under-training Group $G_{\\text{Under-training}}$ is obtained through sampling, while the remaining primitives constitute the Cached Group $G_{\\text{Cached}}$ .", + "bbox": [ + 511, + 619, + 905, + 662 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nG _ {\\text {U n d e r - t r a i n i n g}} = \\left\\{g _ {i} \\mid g _ {i} \\in G, i \\in I, I \\subseteq \\{1, 2, 3, \\dots , | G | \\} \\right\\} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 671, + 906, + 688 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nG _ {\\text {C a c h e d}} = G \\backslash G _ {\\text {U n d e r - t r a i n i n g}} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 696, + 906, + 713 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Before the next grouping, the Cached Group is merged with the Under-training Group, after which the new $G_{\\text{Under-training}}$ and $G_{\\text{Cached}}$ groups are resampled. A visual representation of the Group Training process is provided in Fig. 3. This procedure ensures that each Gaussian primitive has an opportunity to contribute to the training process, and effectively reducing the number of Under-training Gaussians while retaining those deemed \"important.\"", + "bbox": [ + 511, + 719, + 905, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As outlined in Sec. 2, the 3DGS training process consists of two distinct phases: \"Densification\" and \"Optimization\". Group Training is facilitated in both phases and concludes with Groups Merging at each phases, as illustrated in Fig. 4.", + "bbox": [ + 511, + 840, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "27970", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Effective and Efficient Sampling Strategies", + "text_level": 1, + "bbox": [ + 89, + 90, + 455, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Within each grouping, a subset of Gaussians is sampled from the merged Gaussians. The simplest sampling strategy is Random Sampling (RS). We propose attribute-driven Prioritized Sampling to optimize 3DGS further.", + "bbox": [ + 89, + 112, + 483, + 172 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We evaluated various attributes, including opacity and volume-based approaches [5] and importance scores [6, 8, 26, 35] (see appendix for detailed comparisons). Among these, Opacity-based Prioritized Sampling demonstrated superior performance in both training speed and rendering quality. As mathematically proven in Sec. 3.2.1 and 3.2.2, opacity intrinsically governs two critical 3DGS properties, both experimentally validated.", + "bbox": [ + 89, + 172, + 483, + 294 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1. Sampling Strategy for Better Densification", + "text_level": 1, + "bbox": [ + 89, + 300, + 431, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For any Gaussian primitive $G_{m}$ projected onto the imaging plane with 2D center coordinates $[x_{\\mathrm{m}},y_{\\mathrm{m}}]$ , the partial derivatives of loss $L$ w.r.t the projected coordinates are $[\\frac{\\partial L}{\\partial x_{\\mathrm{m}}},\\frac{\\partial L}{\\partial y_{\\mathrm{m}}} ]$ , which govern the densification criterion by:", + "bbox": [ + 89, + 319, + 483, + 385 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\sqrt {\\left(\\frac {\\partial L}{\\partial x _ {\\mathrm {m}}}\\right) ^ {2} + \\left(\\frac {\\partial L}{\\partial y _ {\\mathrm {m}}}\\right) ^ {2}} > \\tau_ {\\text {g r a d}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 392, + 482, + 426 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tau_{\\mathrm{grad}}$ denotes the predefined gradient threshold.", + "bbox": [ + 89, + 429, + 450, + 445 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We formulate the first fundamental property of 3DGS:", + "bbox": [ + 109, + 445, + 470, + 459 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Proposition 1. Opacity-based Effective Gaussians Denomination: Gaussian primitives with higher opacity serve as the primary contributors to densification of 3DGS.", + "bbox": [ + 89, + 460, + 482, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proof. Under our assumptions of mutual independence between Gaussian attributes and within-primitive parameter independence, the partial derivatives are computed as:", + "bbox": [ + 89, + 505, + 483, + 550 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial L}{\\partial x _ {\\mathrm {m}}} = o _ {m} \\sum_ {\\text {p i x e l}} \\frac {\\partial L}{\\partial \\hat {C}} \\frac {\\partial \\hat {C}}{\\partial \\alpha_ {m}} \\frac {\\partial G _ {m} ^ {2 D}}{\\partial \\Delta x} \\frac {\\partial \\Delta x}{x _ {\\mathrm {m}}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 168, + 559, + 482, + 594 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial L}{\\partial y _ {\\mathrm {m}}} = o _ {m} \\sum_ {\\text {p i x e l}} \\frac {\\partial L}{\\partial \\hat {C}} \\frac {\\partial \\hat {C}}{\\partial \\alpha_ {m}} \\frac {\\partial G _ {m} ^ {2 D}}{\\partial \\Delta y} \\frac {\\partial \\Delta y}{y _ {\\mathrm {m}}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 168, + 604, + 482, + 640 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $o_m$ denotes the opacity of $G_{m}$ , $\\hat{C}$ represents the rendered pixel value, and $\\Delta x, \\Delta y$ indicate coordinate offsets between pixels and $[x_{\\mathrm{m}},y_{\\mathrm{m}}]$ . Under the parameter mutual independence assumptions, the $[\\frac{\\partial\\Delta x}{\\partial x_{\\mathrm{m}}},\\frac{\\partial\\Delta y}{\\partial y_{\\mathrm{m}}} ],[\\frac{\\partial G_m^{2\\mathrm{D}}}{\\partial\\Delta x},\\frac{\\partial G_m^{2\\mathrm{D}}}{\\partial\\Delta y} ]$ and $\\frac{\\partial L}{\\partial\\hat{C}}$ remain independent of $o_m$ . We prove that $\\frac{\\partial\\hat{C}}{\\partial\\alpha_m}$ increases with higher expected value of $o_m$ in appendix:", + "bbox": [ + 89, + 646, + 483, + 748 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} \\left[ \\frac {\\partial \\hat {C}}{\\partial \\alpha_ {m}} \\right] = \\frac {(c _ {0} - c _ {\\mathrm {b g}}) T _ {\\mathrm {s a t u r a t i o n}}}{1 - \\mathbb {E} [ o _ {m} ] \\cdot \\mathbb {E} [ G _ {m} ^ {\\mathrm {2 D}} ]}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 755, + 482, + 797 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $c_{0}$ and $\\alpha_0$ denote the expected color and opacity respectively, and $T_{\\mathrm{saturation}}$ controls $\\alpha$ saturation [12].", + "bbox": [ + 89, + 804, + 482, + 833 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With fixed $L$ , Eqs. (4) to (6) demonstrate that elevated opacity distributions yield greater gradient magnitudes in $\\left[\\frac{\\partial L}{\\partial x_{\\mathrm{m}}}, \\frac{\\partial L}{\\partial y_{\\mathrm{m}}}\\right]$ , thereby making such Gaussians statistically prioritized for densification.", + "bbox": [ + 89, + 834, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/15f31b0e10f5a87f4c7b7a46d85325f1188a34b27cfb393f1b2978f628c81a24.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 90, + 707, + 200 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3a1d709fe8a16b95741037f9a51203efff6972535d4b59cc56e962dcccc3cc7e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 90, + 903, + 202 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/40bbbeeea462a6d92ee1506a0560f5095bae56ca1cc91925b76b194a8f2d31be.jpg", + "image_caption": [ + "The distribution of opacity.", + "The distribution of volume." + ], + "image_footnote": [], + "bbox": [ + 517, + 218, + 707, + 342 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1844f67d302ca153d70be6c8c8f4d8ed4acc664250e623317b4120d972c51e07.jpg", + "image_caption": [ + "Figure 5. The distribution of Gaussian attributes. The distribution of all Gaussian attributes (left) and those contributing specifically to densification (right) in the \"Bicycle\"[1] with 3DGS [12]. Top row: While the opacities are primarily concentrated around 0 and 1, the Gaussians that contribute to densification are predominantly situated around 1. Bottom row: The distribution of Volume. As densification progresses, Gaussians with lower volume become increasingly involved in the densification process." + ], + "image_footnote": [], + "bbox": [ + 712, + 218, + 901, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Experimental Validation: Our experiments track Gaussian primitives contributing to densification, focusing on opacity and volumes values. Statistical analysis of attribute distributions during densification steps (Fig. 5) reveals that Gaussians with high opacity and small volume are the primary sources implementing Gaussian densification.", + "bbox": [ + 511, + 494, + 905, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Notably, insufficient high-opacity Gaussians increase the photometric loss, exacerbating both under-reconstruction and over-reconstruction [12], thereby forces redundant densification. While small volumes can be employed as a compression metric [5] in the model compression of 3D Gaussian Splatting, they simultaneously drive reasonable densification. Take into account all factors, it is advisable to cache Gaussians with low-opacity for better gaussian densification.", + "bbox": [ + 511, + 585, + 905, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2. Sampling Strategy for Faster Rendering", + "text_level": 1, + "bbox": [ + 511, + 729, + 839, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During viewpoint-specific image synthesis in 3DGS, the framework implements $\\alpha$ -blending [21] through recursive application of the compositing equation on a per-pixel.", + "bbox": [ + 511, + 747, + 905, + 794 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC _ {p} = \\sum_ {i = 1} ^ {N} c _ {i} \\alpha_ {i} T _ {i}, \\quad T _ {i} = \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\qquad (7)\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 803, + 906, + 848 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nN = \\min \\left\\{i \\in \\mathbb {N} ^ {+}, \\left| \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}) \\leq T _ {\\text {s a t u r a t i o n}} \\right. \\right\\}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 857, + 906, + 906 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "27971", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As established in 3DGS [12], the determination of $N$ hinges on achieving opacity saturation within a pixel's accumulated $\\alpha$ value (Transmittance $T_{i} < T_{\\mathrm{saturation}}$ ). This critical threshold implementation introduces a fundamental constraint on rendering efficiency: during forward rendering operations, the sequential traversal of Gaussian primitives inherently prevents parallel computation of $T_{i}$ values. Consequently, $N$ emerges as the principal determinant of rendering speed, where any reduction on $N$ directly translates to accelerated renderings.", + "bbox": [ + 89, + 90, + 480, + 241 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on the preceding analysis, we formulate the second fundamental property of the 3DGS framework:", + "bbox": [ + 89, + 242, + 482, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Proposition 2. Opacity-based Efficient Rendering Acceleration: Gaussian primitives with higher opacity enable faster rendering through faster achieving $\\alpha$ saturation.", + "bbox": [ + 89, + 273, + 482, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proof. Building upon the premise of mutually independent $\\alpha_{i}$ values between Gaussians, we posit that the mathematical expectation $\\mathbb{E}[T_N] \\approx T_{\\mathrm{saturation}}$ ,", + "bbox": [ + 89, + 316, + 482, + 364 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} [ T _ {N} ] = \\left(1 - \\mathbb {E} [ \\alpha_ {i} ]\\right) ^ {N} = \\left(1 - \\mathbb {E} [ o _ {i} ] \\cdot \\mathbb {E} [ G _ {i} ^ {\\mathrm {2 D}} ]\\right) ^ {N}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 373, + 482, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha_{i}$ combines its intrinsic opacity $o_i$ and $G_{i}^{2\\mathrm{D}}$ (dependent on view projection and covariance, independent of $o_i$ ). We decouple the expected value of $\\alpha_{i}$ .", + "bbox": [ + 89, + 402, + 482, + 448 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Experimental Validation: We generated a 3DGS model with Gaussian $o_i$ values sampled from $\\mathcal{N}(\\mu_o,0.1)$ , fixing the other parameters. The rendering time as a function of $\\mu_o$ is shown in Fig. 6. The results reveals that Gaussian primitives with higher-opacity induce faster transmittance saturation, thereby reducing the blending steps $N$ and rendering computational load.", + "bbox": [ + 89, + 449, + 483, + 555 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Analysis and Design Insights", + "text_level": 1, + "bbox": [ + 89, + 565, + 346, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Through a comprehensive testing program and a detailed examination of the densification and rendering formulas in 3DGS, we have identified the acceleration effect of the opacity attribute in the 3DGS training process. Based on these insights, we propose an Opacity-based Prioritised Sampling (OPS) strategy.", + "bbox": [ + 89, + 587, + 482, + 676 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Opacity-based sampling probability $p_i$ for selecting an Under-training Gaussian $G_i$ is defined as:", + "bbox": [ + 89, + 678, + 482, + 708 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {i} = \\frac {\\alpha_ {i}}{\\sum_ {i = 1} ^ {N} \\alpha_ {i}} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 233, + 712, + 480, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha_{i}$ represents the opacity of Gaussian primitive $G_{i}$ , and $N$ is the total number of Gaussian primitives.", + "bbox": [ + 89, + 755, + 482, + 786 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 799, + 222, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We undertake an experimental investigation of Group Training within the context of 3D Gaussian Splatting [12] and Mip-Splatting [33]. The experimental settings are detailed in Sec. 4.1. Subsequently, we present the experimental results in Sec. 4.2. Cross-baselines applicability validation", + "bbox": [ + 89, + 824, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ca078b0976ec7ffb4514a3c52ced74f8fcfb29b6ddc1f5665142e4276de50df0.jpg", + "image_caption": [ + "Figure 6. High-opacity Gaussians accelerate 3DGS rendering. Left: The rendering time decreases by $40\\%$ as $\\mu_{o}$ increases. Right: The mechanisms that highers-opacity Gaussians achieve faster $\\alpha$ saturation with less number of Gaussians." + ], + "image_footnote": [], + "bbox": [ + 519, + 93, + 898, + 186 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "across diverse scenarios is further presented in Sec. 4.3. The ablation experiments and their results are comprehensively discussed in Sec. 4.4. Finally, we analyse the results of different sampling strategies in Sec. 4.5.", + "bbox": [ + 511, + 256, + 903, + 318 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 511, + 328, + 720, + 344 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. The Group Training method was integrated as a plug-in to facilitate its utilisation in conjunction with 3DGS. We implemented our method within both 3D Gaussian Splatting [12] and Mip-Splatting [33].", + "bbox": [ + 511, + 349, + 903, + 411 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Group Training was implemented with a default grouping interval of 500 iterations throughout the training process. We employed both Random Sampling (RS) and Opacity-Based Prioritized Sampling (OPS) strategies with an Under-Training Ratio (UTR) of 0.6. To ensure computational stability, considering the importance of initial Gaussians [4], Group Training was activated at the 500th iteration after densification begins. The global densification $(I_{\\mathrm{D}})$ and optimization $(I_{\\mathrm{O}})$ were scheduled at $14.5\\mathrm{K}$ and $29\\mathrm{K}$ iterations, respectively. To manage the increased GPU memory consumption during group merging, we fixed the SH coefficients [25, 28] at iterations $I_{\\mathrm{D}}$ and $I_{\\mathrm{O}}$ . We also conducted experiments applying Group Training solely during the densification phase.", + "bbox": [ + 511, + 412, + 905, + 623 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset and Metrics. The efficacy of our method is evaluated on the Mip-NeRF360 [1], Tanks & Temples [14], Deep Blending [9], and NeRF-Synthetic [21] using a single RTX 3090. Subsequently, the model's reconstruction performance is evaluated using PSNR, SSIM, and LPIPS. Furthermore, the training time, peak memory usage, and model size for each scene are recorded.", + "bbox": [ + 511, + 623, + 905, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Experiments Results", + "text_level": 1, + "bbox": [ + 511, + 739, + 707, + 755 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This section presents the experimental results of integrating the Group-Training method into the 3D Gaussian Splat", + "bbox": [ + 511, + 762, + 903, + 792 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "https://github.com/graphdeco-inria/gaussian-splatting", + "bbox": [ + 514, + 803, + 903, + 827 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "https://github.com/autonomousvision/mip-splatting", + "bbox": [ + 514, + 828, + 903, + 852 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "https://jonbarron.info/mipnerf360/", + "bbox": [ + 537, + 852, + 807, + 863 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "https://www.tanksandtemples.org/", + "bbox": [ + 537, + 864, + 790, + 875 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "http://visual.cs.ucl.ac.uk/pubs/deepblending/", + "bbox": [ + 537, + 876, + 892, + 887 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "https://www.matthewancik.com/nerf", + "bbox": [ + 537, + 888, + 808, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "27972", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/60f6c6a18be386967e2e63235d98699ac776b368b335958ae9bb7b3770c876be.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Grouping IterationsMip-NeRF360 [1]Tanks & Temples [14]Deep Blending [9]Blender [21]
PM ↓Size ↓Time ↓PM ↓Size ↓Time ↓PM ↓Size ↓Time ↓PM ↓Size ↓Time ↓
3D Gaussian Splatting*-8.6979226.724.5943415.007.7266623.902.78696.06
Group Training w/ RS0~15K9.5190726.625.0849614.607.6464421.852.76635.66
0~30K9.4890222.585.0449512.257.6364318.452.76635.26
Group Training w/ OPS0~15K8.5767822.534.5438412.806.8148719.252.65435.01
0~30K8.5767919.564.5538410.956.8048716.852.64434.79
", + "bbox": [ + 93, + 88, + 903, + 205 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7cdadf300b636d5ef84e36699f727be72c49b1173550c6afce0dcf6b559a742c.jpg", + "table_caption": [ + "Table 1. Comparison of reconstruction efficiency for 3DGS [12]. Group Training significantly enhances reconstruction speed across all four datasets, with the OPS strategy achieving the highest acceleration. This effect is particularly notable in complex scenes. While Group Training with RS produces a more redundant model, the OPS strategy yields more compact models and reduced GPU peak memory usage. * indicates that the model was retrained. PM stands for GPU peak memory allocation, with Size in MB and Time in minutes." + ], + "table_footnote": [], + "table_body": "
Grouping IterationsMip-NeRF360 [1]Tanks&Temples [14]Deep Blending [9]Blender [21]
PSNR ↑SSIM ↑LPIPS ↓Time ↓PSNR ↑SSIM ↑LPIPS ↓Time ↓PSNR ↑SSIM ↑LPIPS ↓Time ↓PSNR ↑SSIM ↑LPIPS ↓Time ↓
3D GS [12]-27.2050.8150.2143-23.1420.8410.183-29.4050.9030.2425-33.330.9690.030-
3D GS*-27.4450.8160.215526.723.6970.8490.176415.029.5860.9040.243723.933.7720.9700.03066.1
Group Training0~15K27.6210.8220.207426.623.7730.8510.170314.629.1850.9010.247321.933.9590.9710.02845.7
w/ RS0~30K27.5370.8210.215722.623.7030.8480.182312.329.4170.9070.248118.533.8770.9710.02955.3
Group Training0~15K27.5820.8200.210322.523.8420.8510.172612.829.7130.9060.243519.333.8890.9710.02925.0
w/ OPS0~30K27.5640.8200.213319.623.8530.8500.176411.029.7460.9070.245016.933.8080.9700.02994.8
", + "bbox": [ + 93, + 284, + 903, + 381 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparison of reconstruction quality for 3DGS [12]. Group Training simultaneously accelerates the 3DGS scene reconstruction process and improves the accuracy of 3D reconstructions. Notably, Group Training with OPS achieves a notable improvement in reconstruction quality while maintaining the highest reconstruction speed.", + "bbox": [ + 88, + 391, + 906, + 434 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ting [12] and Mip-Splatting [33]. The test outcomes demonstrate that the Group Training with OPS offers a significant reduction in training time compared to the baseline in all scenes, while concurrently enhancing the rendering quality of the novel view, as presented in Fig. 7.", + "bbox": [ + 88, + 453, + 480, + 529 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3D Gaussian Splatting. We first tested the effectiveness of Group Training based on 3DGS. The reconstruction efficiency for all scenes is summarized in Tab. 1, and Tab. 2 presents the NVS rendering quality across all scenes.", + "bbox": [ + 88, + 536, + 482, + 597 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Across these scenes, Group Training consistently resulted in faster reconstruction, and this speedup was more pronounced in complex scenes. However, Group Training with RS could lead to an overabundance of redundant Gaussian primitives in some reconstruction tasks, which complicates subsequent operations, such as model compression [24, 25, 28]. Additionally, GPU memory usage significantly increased due to this redundancy [17], as shown in Tab. 1. In comparison to Group Training with RS, Group Training with OPS demonstrated a more pronounced improvement in reconstruction speed. For the reconstruction of all three complex scenes, the average speed increase was about $30\\%$ . Moreover, OPS greatly reduced the generation of redundant Gaussian primitives, leading to a smaller reconstruction model size than the original 3DGS reconstruction model in all tasks. Across all reconstruction tasks, the model size was reduced by between $10\\%$ and $40\\%$ , resulting in a more compact model overall.", + "bbox": [ + 89, + 598, + 482, + 868 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Mip-Splatting. We tested Mip-Splatting with Group Training, another significant variant of the Gaussian Splat", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ting model. The experimental results are presented in Tab. 3 and Tab. 4. Overall, the performance of Group Training with Mip-Splatting was similar to 3DGS. Group Training with RS consistently improved both reconstruction quality and speed for Mip-Splatting but produced the highest number of Gaussian primitives during the densification phase. Group Training with OPS achieved the fastest reconstruction speed, and its reconstruction quality was nearly optimal. The OPS reduced the proportion of densification and resulted in a smaller model size that was either the smallest among the tested methods or comparable to the baseline.", + "bbox": [ + 511, + 453, + 906, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Counterintuitive Findings. Our experiments reveal two notable findings that challenge conventional understanding:", + "bbox": [ + 511, + 633, + 903, + 664 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Despite producing larger model sizes, Group Training with RS consistently improves reconstruction speed, indicating that training dynamics rather than model size dominates reconstruction efficiency.", + "2. Group Training with OPS simultaneously achieves better reconstruction quality with reduced model size. This demonstrates that superior performance stem from its fundamental impact on the optimization process rather than from model size." + ], + "bbox": [ + 511, + 666, + 903, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Methodological Applicability", + "text_level": 1, + "bbox": [ + 511, + 816, + 772, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Group-Training framework fundamentally addresses accelerated training and rendering in 3DGS through its dataset-agnostic improvement. Designed as a pluggable training component, the method demonstrates dual efficacy in both", + "bbox": [ + 511, + 839, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "27973", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8cbd95e05e89d2c5f2f25c87cae54220153952039ce0b24adbb69b5a24bfde4d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Grouping IterationsTanks & Temples [14]Deep Blending [9]Blender [21]
PSNR ↑SSIM ↑LPIPS ↓Time ↓PSNR ↑SSIM ↑LPIPS ↓Time ↓PSNR ↑SSIM ↑LPIPS ↓Time ↓
Mip-Splatting*-23.7490.8600.156223.029.3580.9030.239035.133.9950.9700.02968.5
Group Training w/ RS0~15K23.9530.8630.150424.628.9290.9000.243131.734.2030.9710.02789.3
0~30K24.1460.8620.159820.429.3920.9080.242026.534.1230.9720.02858.3
Group Training w/ OPS0~15K23.9580.8630.151921.729.6650.9070.236428.134.1320.9710.02858.2
0~30K24.1560.8630.155918.229.7880.9080.238424.034.1100.9710.02867.5
", + "bbox": [ + 93, + 88, + 903, + 198 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ff90e21113ffdf76b29872915933befa834534f90c4dc58114fb6061b226b1d0.jpg", + "table_caption": [ + "Table 3. Comparison of reconstruction quality for Mip-Splatting [33]. Group Training achieves simultaneous improvements in both speed and reconstruction quality for Mip-Splatting across all tasks. Specifically, Group Training with OPS attains the fastest reconstruction speed across all three datasets and delivers the highest reconstruction quality in the Tanks & Temples and Deep Blending dataset." + ], + "table_footnote": [], + "table_body": "
Grouping IterationsOutdoorIndoorAverage
PSNR ↑SSIM ↑LPIPS ↓Size ↓Time ↓PSNR ↑SSIM ↑LPIPS ↓Size ↓Time ↓PSNR ↑SSIM ↑LPIPS ↓Size ↓Time ↓
Mip-Splatting*-24.8140.7480.1995143043.331.2650.9290.173844833.327.6810.8280.188199438.8
Group Training0~15K24.9690.7570.1897201745.931.2660.9310.169755634.927.7680.8340.1808137741.0
w/ RS0~30K25.2160.7630.1969201736.831.1640.9290.175755630.527.8600.8370.1875137734.0
Group Training0~15K24.9810.7540.1940157140.731.4170.9310.172537529.127.8410.8320.1844103935.5
w/ OPS0~30K24.9840.7540.1967157132.931.4070.9300.173937526.427.8390.8320.1865103930.0
", + "bbox": [ + 93, + 263, + 903, + 357 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2e35644c3e76d07ec5d4fa5909c413a698fbbbebcecabbd70640f1dd951b9049.jpg", + "table_caption": [ + "Table 4. Quantitative evaluation on the Mip-NeRF360 [1] reconstructed by Mip-Splatting [33]. For the reconstruction of outdoor scenes, Grouping Training with RS yields a larger model size, resulting in the best reconstruction quality. Grouping Training with OPS achieves the fastest reconstruction speed but provides sub-optimal reconstruction quality. For indoor scenes, Grouping Training with OPS continues to offer both the fastest reconstruction speed and the highest reconstruction quality." + ], + "table_footnote": [], + "table_body": "
Cyclic ResampleGlobal DenseifyGlobal OptimizeTanks &Temples
PSNR ↑SSIM ↑LPIPS ↓Size ↓Time ↓
3D GS*---23.6970.8490.176443415.0
Group Training✓23.8660.8500.174929211.8
✓✓23.7690.8490.176523111.0
✓✓23.8350.8500.175448511.8
✓✓✓23.8530.8500.176438411.0
", + "bbox": [ + 93, + 439, + 480, + 522 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Ablation Study of Group Training with OPS on Tanks & Temples [14]. Cyclic Resampling provides the most substantial acceleration for scene reconstruction, while Global Densification further improves efficiency by reducing model size. With Global Densification enabled, Global Optimization enhances the reconstruction quality of the model.", + "bbox": [ + 89, + 532, + 482, + 617 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3DGS acceleration model [6] and compression model [5], which is detailed in appendix.", + "bbox": [ + 89, + 633, + 482, + 665 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 679, + 254, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conducted an ablation study on the Tanks & Temples dataset [14] to evaluate the impact of three key components of the Group Training method. The detailed comparison results are provided in Tab. 5.", + "bbox": [ + 89, + 702, + 482, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results indicate that cyclic resampling contributes the most to improving training speed. Specifically, during iterations $I_{\\mathrm{D}} \\sim 15\\mathrm{K}$ , the Global Densification process adaptively performs cyclic pruning of low-opacity Gaussians within 3DGS, reducing redundant Gaussians, decreasing the model size, and thereby enhancing training efficiency. When Global Densification is enabled, the 3DGS model removes most redundant Gaussians during global densification, resulting in fewer redundant Gaussians in the Caching", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Group during the Optimization phase. In this scenario, Global Optimization can make full use of all Gaussians, further improving reconstruction performance.", + "bbox": [ + 511, + 441, + 906, + 488 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Discussion", + "text_level": 1, + "bbox": [ + 511, + 497, + 632, + 512 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our experimental results demonstrate that enabling Group Training and utilizing OPS improved both reconstruction speed and quality. To better clarify these improvements, two key questions arise from these findings:", + "bbox": [ + 511, + 521, + 905, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "- **Why faster?** Group Training significantly accelerates reconstruction primarily by reducing the number of Undertraining Gaussian primitives, thereby decreasing the rendering burden on the model. This speedup is observed even when the final model size is larger than the baseline. In alignment with our earlier analysis in Sec. 3.2, the OPS also contributes notably to this speedup effect. During the densification phase, Group Training with OPS preserves high-opacity Gaussians within the Under-Training Group, without which could lead to instances of both under-reconstruction and over-reconstruction [12, 32]. These issues can lead to redundant densification. As a result, models reconstructed with Group Training and OPS contain fewer redundant Gaussians, leading to smaller model sizes and faster reconstruction overall. Additionally, OPS prioritizes sampling high-opacity Gaussians, which rapidly drives $\\alpha$ saturation, which also discussed in Sec. 3.2. Focusing on these most impactful Gaussians reduces the number of Gaussians needed for splatting, further enhancing training and rendering efficiency.", + "bbox": [ + 511, + 582, + 906, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "- Why better? Group Training with the OPS prioritizes the", + "bbox": [ + 511, + 885, + 903, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "27974", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7ac07db189326658c53b3f138a974651370dfa348f08dbbe94f47087001e601f.jpg", + "image_caption": [ + "Figure 7. Visual Comparison of Grouping Training and Baseline on 3DGS [12] and Mip-Splatting [33]. The Grouping Training method accelerates reconstruction by approximately $30\\%$ while achieving more detailed and accurate scene rendering. Notable improvements can be observed in the picture frame in the \"Dr. Johnson\" scene and the bucket handle in the \"Garden\" scene, where finer details are captured. Furthermore, Grouping Training significantly reduces floaters in Mip-Splatting, enhancing overall reconstruction fidelity." + ], + "image_footnote": [], + "bbox": [ + 116, + 88, + 883, + 550 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "retention and optimization of high-opacity Gaussian primitives. This achieves comparable or superior fidelity to RS, while significantly reducing the model size. The Gaussian primitives with higher opacity inherently have a greater influence on the final rendered image [5, 19, 22, 28], so emphasizing their optimization leads to significant improvements in overall rendering quality. Besides, cyclic resampling temporarily excluding a randomly selected subset of Gaussian primitives, each remaining primitive is compelled to learn how to independently represent essential scene features leading to a more robust representation.", + "bbox": [ + 89, + 637, + 483, + 803 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In summary, by integrating these strategies, Group Training with OPS simultaneously enhances both reconstruction speed and rendering quality. By focusing on the most impactful primitives and encouraging independent feature representation, Group Training achieves efficient, high-quality scene reconstruction.", + "bbox": [ + 89, + 810, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 636, + 633, + 651 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presents Group Training, an effective strategy that addresses the computational challenges in 3D Gaussian Splitting by dynamically managing Gaussian primitives in organized groups. Through theoretical analysis and extensive experiments, we demonstrate that our method not only accelerates training by up to $30\\%$ but also improves rendering quality across diverse scenarios. Our opacity-based sampling strategy shows universal compatibility with existing 3DGS frameworks, achieving consistent improvements without compromising model complexity. The success of Group Training reveals the importance of dynamic primitive management in 3DGS optimization, suggesting promising directions for future research, such as adaptive grouping strategies and dynamic sampling mechanisms for more complex scenes. Our work provides valuable insights for future developments in efficient novel view synthesis.", + "bbox": [ + 511, + 655, + 906, + 897 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "27975", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported by the National Key R&D Program of China (No.2022ZD0119000), Hunan Provincial Key R&D Program of China (No.2024JK2020 and 2024JK2021), Hunan Provincial Natural Science Foundation of China (No.2024JJ10027), Young Talents of Huxiang (No.Z202433000575), and Changsha Science Fund for Distinguished Young Scholars (kq2306002).", + "bbox": [ + 89, + 114, + 483, + 220 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 234, + 187, + 250 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jonathan T. Barron. Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields. In CVPR, 2022. 1, 4, 5, 6, 7, 8", + "[2] Samuel Rota Bulò, Lorenzo Porzi, and Peter Kontschieder. Revising densification in gaussian splatting. arXiv preprint arXiv:2404.06109, 2024. 1, 3", + "[3] Guikun Chen and Wenguan Wang. A survey on 3d gaussian splatting. arXiv preprint arXiv:2401.03890, 2024. 1", + "[4] Jaeyoung Chung, Jeongtaek Oh, and Kyoung Mu Lee. Depth-regularized optimization for 3d gaussian splatting in few-shot images. arXiv preprint arXiv:2311.13398, 2023. 5", + "[5] Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, and Zhangyang Wang. Lightgaussian: Unbounded 3d gaussian compression with $15\\mathrm{x}$ reduction and $200+$ fps, 2023. 3, 4, 7, 8", + "[6] Guangchi Fang and Bing Wang. Mini-splatting: Representing scenes with a constrained number of gaussians, 2024. 3, 4, 7", + "[7] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 2", + "[8] Sharath Girish, Kamal Gupta, and Abhinav Shrivastava. Eagles: Efficient accelerated 3d gaussians with lightweight encodings. arXiv preprint arXiv:2312.04564, 2023. 4", + "[9] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep Blending for Free-Viewpoint Image-Based Rendering. In SIGGRAPH, 2018. 1, 5, 6, 7, 8", + "[10] Lukas Hollein, Aljaž Božić, Michael Zollhöfer, and Matthias Nießner. 3dgs-lm: Faster gaussian-splatting optimization with levenberg-marquardt. arXiV:2409.12892, 2024. 3", + "[11] Ying Jiang, Chang Yu, Tianyi Xie, Xuan Li, Yutao Feng, Huamin Wang, Minchen Li, Henry Lau, Feng Gao, Yin Yang, et al. Vr-gs: A physical dynamics-aware interactive gaussian splatting system in virtual reality. In ACM SIGGRAPH 2024 Conference Papers, pages 1-1, 2024. 1", + "[12] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 2, 3, 4, 5, 6, 7, 8", + "[13] Shakiba Kheradmand, Daniel Rebain, Gopal Sharma, Weiwei Sun, Yang-Che Tseng, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. 3d gaussian splatting as markov chain monte carlo. Advances in Neural Information Processing Systems, 37:80965-80986, 2024. 3" + ], + "bbox": [ + 93, + 258, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and Temples: Benchmarking Large-Scale Scene Reconstruction. In SIGGRAPH, 2017. 2, 5, 6, 7, 8", + "[15] Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaussian feature splatting for real-time dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8508-8520, 2024. 1", + "[16] Yang Liu, Chuanchen Luo, Lue Fan, Naiyan Wang, Junran Peng, and Zhaoxiang Zhang. Citygaussian: Real-time high-quality large-scale scene rendering with gaussians. In European Conference on Computer Vision, pages 265-282. Springer, 2025. 3", + "[17] Jiahao Lu, Yifan Zhang, Qiuhong Shen, Xinchao Wang, and Shuicheng Yan. Poisson-splat: Computation cost attack on 3d gaussian splatting. arXiv preprint arXiv:2410.08190, 2024. 6", + "[18] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 3", + "[19] Saswat Subhajyoti Mallick, Rahul Goel, Bernhard Kerbl, Markus Steinberger, Francisco Vicente Carrasco, and Fernando De La Torre. Taming 3dgs: High-quality radiance fields with limited resources. In SIGGRAPH Asia 2024 Conference Papers, New York, NY, USA, 2024. Association for Computing Machinery. 3, 8", + "[20] Jiarui Meng, Hajie Li, Yanmin Wu, Qiankun Gao, Shuzhou Yang, Jian Zhang, and Siwei Ma. Mirror-3dgs: Incorporating mirror reflections into 3d gaussian splatting. arXiv preprint arXiv:2404.01168, 2024. 1", + "[21] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 4, 5, 6, 7", + "[22] Wieland Morgenstern, Florian Barthel, Anna Hilsmann, and Peter Eisert. Compact 3d scene representation via self-organizing gaussian grids. arXiv preprint arXiv:2312.13299, 2023.8", + "[23] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 2", + "[24] KL Navaneet, Kossar Pourahmadi Meibodi, Soroush Abbasi Koohpayegani, and Hamed Pirsiavash. Compgs: Smaller and faster gaussian splatting with vector quantization. ECCV, 2024. 6", + "[25] Simon Niedermayr, Josef Stumpfegger, and Rüdiger Westermann. Compressed 3d gaussian splattering for accelerated novel view synthesis, 2023. 3, 5, 6", + "[26] Michael Niemeyer, Fabian Manhardt, Marie-Julie Rakoto-saona, Michael Oechsle, Daniel Duckworth, Rama Gosula, Keisuke Tateno, John Bates, Dominik Kaeser, and Federico Tombari. Radsplat: Radiance field-informed gaussian splatting for robust real-time rendering with $900+$ fps. arXiv.org, 2024. 3, 4" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "27976", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Panagiotis Papantonakis, Georgios Kopanas, Bernhard Kerbl, Alexandre Lanvin, and George Drettakis. Reducing the memory footprint of 3d gaussian splatting. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 7(1):1-17, 2024. 3", + "[28] Panagiotis Papantonakis, Georgios Kopanas, Bernhard Kerbl, Alexandre Lanvin, and George Drettakis. Reducing the memory footprint of 3d gaussian splatting. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 2024. 2, 5, 6, 8", + "[29] Zhexi Peng, Tianjia Shao, Yong Liu, Jingke Zhou, Yin Yang, Jingdong Wang, and Kun Zhou. Rtg-slam: Real-time 3d reconstruction at scale using gaussian splatting. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 1", + "[30] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In CVPR, 2022. 2", + "[31] Haosen Yang, Chenhao Zhang, Wenqing Wang, Marco Volino, Adrian Hilton, Li Zhang, and Xiatian Zhu. Localized gaussian point management. arXiv e-prints, pages arXiv-2406, 2024. 3", + "[32] Zongxin Ye, Wenyu Li, Sidun Liu, Peng Qiao, and Yong Dou. Absgs: Recovering fine details for 3d gaussian splatt ing, 2024. 7", + "[33] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In CVPR, 2024. 1, 3, 5, 6, 7, 8", + "[34] Hongjia Zhai, Xiyu Zhang, Boming Zhao, Hai Li, Yijia He, Zhaopeng Cui, Hujun Bao, and Guofeng Zhang. Splatloc: 3d gaussian splatting-based visual localization for augmented reality. arXiv preprint arXiv:2409.14067, 2024. 1", + "[35] Zhaoliang Zhang, Tianchen Song, Yongjae Lee, Li Yang, Cheng Peng, Rama Chellappa, and Deliang Fan. Lp-3dgs: Learning to prune 3d gaussian splatting. arXiv preprint arXiv:2405.18784, 2024. 4", + "[36] Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21634-21643, 2024. 1", + "[37] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. In European Conference on Computer Vision, pages 145-163. Springer, 2025. 1" + ], + "bbox": [ + 91, + 92, + 480, + 728 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "27977", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Faster and Better 3D Splatting via Group Training/21033709-b705-4347-a1bb-c3e7a58591e2_model.json b/2025/Faster and Better 3D Splatting via Group Training/21033709-b705-4347-a1bb-c3e7a58591e2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..955fdc781e6dd10fa6c206581471d83cb6dbbb28 --- /dev/null +++ b/2025/Faster and Better 3D Splatting via Group Training/21033709-b705-4347-a1bb-c3e7a58591e2_model.json @@ -0,0 +1,2222 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.243, + 0.131, + 0.757, + 0.154 + ], + "angle": 0, + "content": "Faster and Better 3D Splatting via Group Training" + }, + { + "type": "text", + "bbox": [ + 0.171, + 0.179, + 0.825, + 0.199 + ], + "angle": 0, + "content": "Chengbo Wang\\(^{1,2,4,5}\\) Guozheng Ma\\(^{3}\\) Yifei Xue\\(^{1,2,4,5}\\) Yizhen Lao\\(^{1,4,5,\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.301, + 0.199, + 0.699, + 0.216 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Design, Hunan University \\(^{2}\\)DaHe.AI" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.216, + 0.758, + 0.235 + ], + "angle": 0, + "content": "\\(^{3}\\)Nanyang Technological University \\(^{4}\\)Lushan Innovation Lab" + }, + { + "type": "text", + "bbox": [ + 0.125, + 0.234, + 0.875, + 0.253 + ], + "angle": 0, + "content": "5 Key Laboratory of Digital Culture Smart Design Technology, Minstry of Culture and Tourism" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.254, + 0.842, + 0.27 + ], + "angle": 0, + "content": "{wangchb, yizhenlao}@hnu.edu.cn, GUOZHENG001@e.ntu.edu.sg, iflyhsueh@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.304, + 0.327, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.336, + 0.486, + 0.624 + ], + "angle": 0, + "content": "3D Gaussian Splatting (3DGS) has emerged as a powerful technique for novel view synthesis, demonstrating remarkable capability in high-fidelity scene reconstruction through its Gaussian primitive representations. However, the computational overhead induced by the massive number of primitives poses a significant bottleneck to training efficiency. To overcome this challenge, we propose Group Training, a simple yet effective strategy that organizes Gaussian primitives into manageable groups, optimizing training efficiency and improving rendering quality. This approach shows universal compatibility with existing 3DGS frameworks, including vanilla 3DGS and MipSplatting, consistently achieving accelerated training while maintaining superior synthesis quality. Extensive experiments reveal that our straightforward Group Training strategy achieves up to \\(30\\%\\) faster convergence and improved rendering quality across diverse scenarios. Project Website: https://chengbo-wang.github.io/3DGS-with-Group-Training/." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.654, + 0.222, + 0.669 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.679, + 0.484, + 0.831 + ], + "angle": 0, + "content": "Novel view synthesis (NVS) is a pivotal technology in various domains, such as virtual reality [11], augmented reality [34], and autonomous driving [36]. Recently, 3D Gaussian Splitting (3DGS) [12] achieved significant success in NVS due to its real-time [15, 29, 37], high-quality rendering [2, 20]. This high-quality rendering is made possible by using millions of Gaussians [3], each includes a complex set of attributes, namely, position, size, orientation, opacity, and color. These attributes are optimized via multi-view photometric losses." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.846, + 0.484, + 0.877 + ], + "angle": 0, + "content": "- Motivation. However, 3DGS also introduces a potential risk during training, as the exponentially growing Gaussians" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.314, + 0.744, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.747, + 0.317, + 0.884, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.567, + 0.526, + 0.853, + 0.538 + ], + "angle": 0, + "content": "(a) Reconstruction of the \"Stump\"[1] scene by 3DGS[12]." + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.542, + 0.744, + 0.751 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.747, + 0.542, + 0.883, + 0.746 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.753, + 0.885, + 0.766 + ], + "angle": 0, + "content": "(b) Reconstruction of the \"Dr. Johnson\"[9] scene by Mip-Splatting[33]." + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.777, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Figure 1. Improvements and Illustrations of Applying Group Training in 3DGS[12] and Mip-Splatting[33]. Grouping Training (solid line) achieves a significant increase in reconstruction speed and superior scene quality compared to baseline methods (dashed line). In the \"Stump\"[1] scene, Grouping Training delivers a \\(28\\%\\) faster reconstruction while rendering leaf details with greater clarity. In the \"Dr. Johnson\"[9] scene, Mip-Splatting with Group Training achieves a \\(33\\%\\) reduction in reconstruction time and effectively suppresses the generation of floating artifacts." + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.887, + 0.48, + 0.901 + ], + "angle": 0, + "content": "† Corresponding author: Yizhen Lao (yizhenlao@hnu.edu.cn)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27968" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "significantly increase the training burden. Consequently, a primary research focus is the development of methods that enable rapid and cost-effective 3D scene reconstruction without compromising the rendering performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.153, + 0.483, + 0.349 + ], + "angle": 0, + "content": "During the densification stage, the cyclic pruning of Gaussians with low opacity emerges as a straightforward yet potentially effective strategy [12, 28] for enhancing the training efficiency of 3DGS. Specifically, in each round, 3DGS prunes Gaussians whose opacity falls below a designated threshold \\(\\epsilon_{\\alpha}\\), which is a vital parameter that facilitates control over the number of generated Gaussians. However, this method has inherent limitations. If \\(\\epsilon_{\\alpha}\\) is set too conservatively, the resultant acceleration will be minimal while an excessively lenient threshold defeats the quality of the NVS. Thus, achieving a balance between accelerating the training and preserving the rendering quality poses a considerable challenge." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.364, + 0.483, + 0.739 + ], + "angle": 0, + "content": "- Intuition and contributions. Inspired by the limitations of the cyclic pruning strategy, we propose to cache a portion of the Gaussians instead of directly pruning Gaussians. This \"cache\" concept can retain the \"important\" Gaussians while reducing the total number of them. Therefore, we propose a novel Group Training strategy to accelerate the training of 3DGS. Specifically, (1) a portion of Gaussians is cyclically cached during training. These cached Gaussians are temporarily excluded from scene rendering and model training, which greatly reduces the number of Undertraining Gaussians and cuts down on training time. (2) Additionally, through cyclic resampling, the cached Gaussians are reintegrated into the 3DGS model, reducing the impact of direct pruning. (3) A crucial element of Group Training is determining how to sample Gaussians to distinguish between under-training and cached Gaussians. A naive sampling approach would be random sampling. Interestingly, we discovered that the distribution of Gaussian opacity values influences both the quantity of Gaussians generated during densification and the rendering speed of the scene. Therefore, we introduce an Opacity-based Prioritized Sampling method, which effectively reduces the generation of redundant Gaussians and improves the training speed of 3DGS. Specifically, the main technical contributions of our work are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.742, + 0.483, + 0.802 + ], + "angle": 0, + "content": "1. We propose a simple yet highly efficient training framework called Group Training for 3D Gaussian Splatting, which is seamlessly integrated into existing 3DGS frameworks, including both 3DGS and Mip-Splatting." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.803, + 0.483, + 0.833 + ], + "angle": 0, + "content": "2. Our framework significantly improves the efficiency of scene reconstruction and the quality of NVS." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.742, + 0.483, + 0.833 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.845, + 0.233, + 0.86 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this section, we review relevant work in novel view synthesis and 3D Gaussian Splatting, followed by the distinct" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.093, + 0.71, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.094, + 0.902, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.298, + 0.907, + 0.451 + ], + "angle": 0, + "content": "Figure 2. PSNR and Time Performance of Pruning [12, 28] and Group Training under Varying Hyperparameters. The reconstructions are performed on the \"Train\" scene [14] using 3DGS [12] as the baseline. Left: The pruning method exhibits substantial instability in optimizing the trade-off between reconstruction efficiency and quality. The hyperparameter sensitivity (Opacity Threshold) presents significant challenges for optimal parameter tuning. Right: Group Training demonstrates consistent improvements in both reconstruction speed and quality, with robust performance across a wide range of hyperparameter values (Caching Ratio), enabling straightforward parameter optimization." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.469, + 0.655, + 0.484 + ], + "angle": 0, + "content": "tion of our approach." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.491, + 0.905, + 0.656 + ], + "angle": 0, + "content": "- NVS. NVS involves generating new views of a scene from different viewpoints than those in the original dataset. Neural Radiance Fields (NeRF) [21] have been pivotal in NVS for producing photo-realistic images through volume rendering. Neural Radiance Fields (NeRF) have significantly advanced NVS by producing photo-realistic images via volume rendering, but they require substantial computational resources and lengthy training due to costly MLP evaluations. To overcome these limitations, methods such as voxel grids [7], hash grids [23] and points [30] have been introduced to facilitate real-time NeRF rendering." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.905, + 0.735 + ], + "angle": 0, + "content": "Recently, 3D Gaussian Splatting (3DGS) [12] has emerged as a promising alternative in the NVS field. Unlike the implicit structure of NeRF, 3DGS uses rasterization rather than ray tracing, enabling efficient rendering suitable for applications requiring both speed and image quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.749, + 0.905, + 0.854 + ], + "angle": 0, + "content": "- 3DGS. 3D Gaussian Splatting (3DGS) is a recent approach to high-quality, real-time novel view synthesis that represents scenes explicitly as 3D Gaussian primitives. By rasterizing these Gaussians into 2D splats and blending them using \\(\\alpha\\)-blending [21], 3DGS produces detailed, high-resolution outputs, making it well-suited for applications that demand both speed and visual fidelity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "The 3DGS training process involves two main stages: \"Densification\" and \"Optimization\". During the densification phase (0~15K iterations), Adaptive Density" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27969" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.119, + 0.09, + 0.997, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.241, + 0.496, + 0.312 + ], + "angle": 0, + "content": "Figure 3. The overall framework of Group at regular iteration intervals, Gaussians from are categorized into the Under-training Group the Under-training Group is utilized for Ga Group remains inactive and does not participate in any calculations." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.327, + 0.485, + 0.554 + ], + "angle": 0, + "content": "Control [12] is applied to increase the density of Gaussian primitives, thereby enhancing the model's capacity for detailed scene representation. In the optimization phase (15K~30K iterations), densification stops, and the existing Gaussians undergo global refinement, yielding a more precise scene depiction. Recent studies have driven significant advancements in 3DGS across several areas, notably improving rendering quality [2, 13, 33], memory efficiency [5, 25, 27], and scalability for large scenes [16]. Beyond these enhancements, additional research has focused on increasing the efficiency of 3DGS optimizers [10] and refining Gaussian Densification [18, 19, 31] and Pruning strategies [5, 6, 26]. These improvements yield more compact 3DGS models that support faster reconstruction but sacrificing detail." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.484, + 0.69 + ], + "angle": 0, + "content": "- Distinction. Unlike previous optimization-focused approaches, we propose Group Training, a principled strategy that dynamically manages Gaussian primitives during training through controlled group sampling. This approach not only accelerates training but also improves reconstruction quality by maintaining effective primitive optimization, achieving a better balance between computational efficiency and visual fidelity." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.695, + 0.227, + 0.712 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In Sec. 3.1, we will provide a comprehensive introduction to the Group Training method and its multifunctional advantages in 3D Gaussian Splatting (3DGS). Section 3.2 introduces two sampling paradigms: Random Sampling and Prioritized Sampling (incorporating importance score, opacity-based, and volume-based criteria), with comparative analysis of their computational efficiency and performance trade-offs. We mathematically demonstrate that Opacity-based Prioritized Sampling achieves effective 3DGS densification and efficient rendering. Finally, in Sec. 3.3, we summarize training characteristics of 3DGS with Group Training, and the Sampling Strategies are identified." + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.325, + 0.897, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.411, + 0.907, + 0.482 + ], + "angle": 0, + "content": "Figure 4. Schedule for activating Group Training during reconstruction. Group Training is enabled at regular intervals. Groups Merging (Grouping without resampling) occurring at iteration \\( I_{\\mathrm{D}} \\) and \\( I_{\\mathrm{O}} \\) before the completion of the densification and optimization processes for global densification and optimization." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.494, + 0.673, + 0.51 + ], + "angle": 0, + "content": "3.1. Group Training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.516, + 0.907, + 0.621 + ], + "angle": 0, + "content": "The number of Gaussian primitives in the Gaussian set increases rapidly during the densification phase to achieve rich representation capabilities. This implies that the computational time to render a scene from a specific viewpoint will also increase, and this effect will extend to the global optimisation phase. To this end, we employ the Group Training method during training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.621, + 0.906, + 0.664 + ], + "angle": 0, + "content": "For any given set of Gaussian \\(G\\), the Under-training Group \\(G_{\\text{Under-training}}\\) is obtained through sampling, while the remaining primitives constitute the Cached Group \\(G_{\\text{Cached}}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.531, + 0.672, + 0.907, + 0.689 + ], + "angle": 0, + "content": "\\[\nG _ {\\text {U n d e r - t r a i n i n g}} = \\left\\{g _ {i} \\mid g _ {i} \\in G, i \\in I, I \\subseteq \\{1, 2, 3, \\dots , | G | \\} \\right\\} \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.618, + 0.698, + 0.907, + 0.714 + ], + "angle": 0, + "content": "\\[\nG _ {\\text {C a c h e d}} = G \\backslash G _ {\\text {U n d e r - t r a i n i n g}} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.84 + ], + "angle": 0, + "content": "Before the next grouping, the Cached Group is merged with the Under-training Group, after which the new \\( G_{\\text{Under-training}} \\) and \\( G_{\\text{Cached}} \\) groups are resampled. A visual representation of the Group Training process is provided in Fig. 3. This procedure ensures that each Gaussian primitive has an opportunity to contribute to the training process, and effectively reducing the number of Under-training Gaussians while retaining those deemed \"important.\"" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.906, + 0.902 + ], + "angle": 0, + "content": "As outlined in Sec. 2, the 3DGS training process consists of two distinct phases: \"Densification\" and \"Optimization\". Group Training is facilitated in both phases and concludes with Groups Merging at each phases, as illustrated in Fig. 4." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27970" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.457, + 0.108 + ], + "angle": 0, + "content": "3.2. Effective and Efficient Sampling Strategies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.485, + 0.174 + ], + "angle": 0, + "content": "Within each grouping, a subset of Gaussians is sampled from the merged Gaussians. The simplest sampling strategy is Random Sampling (RS). We propose attribute-driven Prioritized Sampling to optimize 3DGS further." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.174, + 0.484, + 0.295 + ], + "angle": 0, + "content": "We evaluated various attributes, including opacity and volume-based approaches [5] and importance scores [6, 8, 26, 35] (see appendix for detailed comparisons). Among these, Opacity-based Prioritized Sampling demonstrated superior performance in both training speed and rendering quality. As mathematically proven in Sec. 3.2.1 and 3.2.2, opacity intrinsically governs two critical 3DGS properties, both experimentally validated." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.301, + 0.433, + 0.317 + ], + "angle": 0, + "content": "3.2.1. Sampling Strategy for Better Densification" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.32, + 0.484, + 0.386 + ], + "angle": 0, + "content": "For any Gaussian primitive \\( G_{m} \\) projected onto the imaging plane with 2D center coordinates \\( [x_{\\mathrm{m}},y_{\\mathrm{m}}] \\), the partial derivatives of loss \\( L \\) w.r.t the projected coordinates are \\( [\\frac{\\partial L}{\\partial x_{\\mathrm{m}}},\\frac{\\partial L}{\\partial y_{\\mathrm{m}}} ] \\), which govern the densification criterion by:" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.393, + 0.483, + 0.428 + ], + "angle": 0, + "content": "\\[\n\\sqrt {\\left(\\frac {\\partial L}{\\partial x _ {\\mathrm {m}}}\\right) ^ {2} + \\left(\\frac {\\partial L}{\\partial y _ {\\mathrm {m}}}\\right) ^ {2}} > \\tau_ {\\text {g r a d}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.43, + 0.451, + 0.446 + ], + "angle": 0, + "content": "where \\(\\tau_{\\mathrm{grad}}\\) denotes the predefined gradient threshold." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.446, + 0.471, + 0.46 + ], + "angle": 0, + "content": "We formulate the first fundamental property of 3DGS:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.461, + 0.483, + 0.505 + ], + "angle": 0, + "content": "- Proposition 1. Opacity-based Effective Gaussians Denomination: Gaussian primitives with higher opacity serve as the primary contributors to densification of 3DGS." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.506, + 0.484, + 0.551 + ], + "angle": 0, + "content": "Proof. Under our assumptions of mutual independence between Gaussian attributes and within-primitive parameter independence, the partial derivatives are computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.169, + 0.56, + 0.483, + 0.595 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial L}{\\partial x _ {\\mathrm {m}}} = o _ {m} \\sum_ {\\text {p i x e l}} \\frac {\\partial L}{\\partial \\hat {C}} \\frac {\\partial \\hat {C}}{\\partial \\alpha_ {m}} \\frac {\\partial G _ {m} ^ {2 D}}{\\partial \\Delta x} \\frac {\\partial \\Delta x}{x _ {\\mathrm {m}}}, \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.169, + 0.605, + 0.483, + 0.641 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial L}{\\partial y _ {\\mathrm {m}}} = o _ {m} \\sum_ {\\text {p i x e l}} \\frac {\\partial L}{\\partial \\hat {C}} \\frac {\\partial \\hat {C}}{\\partial \\alpha_ {m}} \\frac {\\partial G _ {m} ^ {2 D}}{\\partial \\Delta y} \\frac {\\partial \\Delta y}{y _ {\\mathrm {m}}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.647, + 0.484, + 0.749 + ], + "angle": 0, + "content": "where \\(o_m\\) denotes the opacity of \\(G_{m}\\), \\(\\hat{C}\\) represents the rendered pixel value, and \\(\\Delta x, \\Delta y\\) indicate coordinate offsets between pixels and \\([x_{\\mathrm{m}},y_{\\mathrm{m}}]\\). Under the parameter mutual independence assumptions, the \\([\\frac{\\partial\\Delta x}{\\partial x_{\\mathrm{m}}},\\frac{\\partial\\Delta y}{\\partial y_{\\mathrm{m}}} ],[\\frac{\\partial G_m^{2\\mathrm{D}}}{\\partial\\Delta x},\\frac{\\partial G_m^{2\\mathrm{D}}}{\\partial\\Delta y} ]\\) and \\(\\frac{\\partial L}{\\partial\\hat{C}}\\) remain independent of \\(o_m\\). We prove that \\(\\frac{\\partial\\hat{C}}{\\partial\\alpha_m}\\) increases with higher expected value of \\(o_m\\) in appendix:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.756, + 0.483, + 0.798 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} \\left[ \\frac {\\partial \\hat {C}}{\\partial \\alpha_ {m}} \\right] = \\frac {(c _ {0} - c _ {\\mathrm {b g}}) T _ {\\mathrm {s a t u r a t i o n}}}{1 - \\mathbb {E} [ o _ {m} ] \\cdot \\mathbb {E} [ G _ {m} ^ {\\mathrm {2 D}} ]}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.805, + 0.483, + 0.834 + ], + "angle": 0, + "content": "where \\(c_{0}\\) and \\(\\alpha_0\\) denote the expected color and opacity respectively, and \\(T_{\\mathrm{saturation}}\\) controls \\(\\alpha\\) saturation [12]." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.835, + 0.484, + 0.902 + ], + "angle": 0, + "content": "With fixed \\( L \\), Eqs. (4) to (6) demonstrate that elevated opacity distributions yield greater gradient magnitudes in \\( \\left[\\frac{\\partial L}{\\partial x_{\\mathrm{m}}}, \\frac{\\partial L}{\\partial y_{\\mathrm{m}}}\\right] \\), thereby making such Gaussians statistically prioritized for densification." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.091, + 0.709, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.091, + 0.904, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.638, + 0.205, + 0.783, + 0.218 + ], + "angle": 0, + "content": "The distribution of opacity." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.219, + 0.709, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.219, + 0.902, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.638, + 0.346, + 0.783, + 0.358 + ], + "angle": 0, + "content": "The distribution of volume." + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.37, + 0.907, + 0.484 + ], + "angle": 0, + "content": "Figure 5. The distribution of Gaussian attributes. The distribution of all Gaussian attributes (left) and those contributing specifically to densification (right) in the \"Bicycle\"[1] with 3DGS [12]. Top row: While the opacities are primarily concentrated around 0 and 1, the Gaussians that contribute to densification are predominantly situated around 1. Bottom row: The distribution of Volume. As densification progresses, Gaussians with lower volume become increasingly involved in the densification process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.495, + 0.906, + 0.586 + ], + "angle": 0, + "content": "- Experimental Validation: Our experiments track Gaussian primitives contributing to densification, focusing on opacity and volumes values. Statistical analysis of attribute distributions during densification steps (Fig. 5) reveals that Gaussians with high opacity and small volume are the primary sources implementing Gaussian densification." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.587, + 0.906, + 0.721 + ], + "angle": 0, + "content": "Notably, insufficient high-opacity Gaussians increase the photometric loss, exacerbating both under-reconstruction and over-reconstruction [12], thereby forces redundant densification. While small volumes can be employed as a compression metric [5] in the model compression of 3D Gaussian Splatting, they simultaneously drive reasonable densification. Take into account all factors, it is advisable to cache Gaussians with low-opacity for better gaussian densification." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.73, + 0.84, + 0.746 + ], + "angle": 0, + "content": "3.2.2. Sampling Strategy for Faster Rendering" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.748, + 0.906, + 0.795 + ], + "angle": 0, + "content": "During viewpoint-specific image synthesis in 3DGS, the framework implements \\(\\alpha\\)-blending [21] through recursive application of the compositing equation on a per-pixel." + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.804, + 0.907, + 0.849 + ], + "angle": 0, + "content": "\\[\nC _ {p} = \\sum_ {i = 1} ^ {N} c _ {i} \\alpha_ {i} T _ {i}, \\quad T _ {i} = \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\qquad (7)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.858, + 0.907, + 0.907 + ], + "angle": 0, + "content": "\\[\nN = \\min \\left\\{i \\in \\mathbb {N} ^ {+}, \\left| \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}) \\leq T _ {\\text {s a t u r a t i o n}} \\right. \\right\\}, \\tag {8}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27971" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.242 + ], + "angle": 0, + "content": "As established in 3DGS [12], the determination of \\( N \\) hinges on achieving opacity saturation within a pixel's accumulated \\( \\alpha \\) value (Transmittance \\( T_{i} < T_{\\mathrm{saturation}} \\)). This critical threshold implementation introduces a fundamental constraint on rendering efficiency: during forward rendering operations, the sequential traversal of Gaussian primitives inherently prevents parallel computation of \\( T_{i} \\) values. Consequently, \\( N \\) emerges as the principal determinant of rendering speed, where any reduction on \\( N \\) directly translates to accelerated renderings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.243, + 0.483, + 0.273 + ], + "angle": 0, + "content": "Based on the preceding analysis, we formulate the second fundamental property of the 3DGS framework:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.274, + 0.483, + 0.317 + ], + "angle": 0, + "content": "- Proposition 2. Opacity-based Efficient Rendering Acceleration: Gaussian primitives with higher opacity enable faster rendering through faster achieving \\(\\alpha\\) saturation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.483, + 0.365 + ], + "angle": 0, + "content": "Proof. Building upon the premise of mutually independent \\(\\alpha_{i}\\) values between Gaussians, we posit that the mathematical expectation \\(\\mathbb{E}[T_N] \\approx T_{\\mathrm{saturation}}\\)," + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.374, + 0.483, + 0.394 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} [ T _ {N} ] = \\left(1 - \\mathbb {E} [ \\alpha_ {i} ]\\right) ^ {N} = \\left(1 - \\mathbb {E} [ o _ {i} ] \\cdot \\mathbb {E} [ G _ {i} ^ {\\mathrm {2 D}} ]\\right) ^ {N}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.404, + 0.483, + 0.449 + ], + "angle": 0, + "content": "where \\(\\alpha_{i}\\) combines its intrinsic opacity \\(o_i\\) and \\(G_{i}^{2\\mathrm{D}}\\) (dependent on view projection and covariance, independent of \\(o_i\\)). We decouple the expected value of \\(\\alpha_{i}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.45, + 0.484, + 0.556 + ], + "angle": 0, + "content": "- Experimental Validation: We generated a 3DGS model with Gaussian \\(o_i\\) values sampled from \\(\\mathcal{N}(\\mu_o,0.1)\\), fixing the other parameters. The rendering time as a function of \\(\\mu_o\\) is shown in Fig. 6. The results reveals that Gaussian primitives with higher-opacity induce faster transmittance saturation, thereby reducing the blending steps \\(N\\) and rendering computational load." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.566, + 0.347, + 0.582 + ], + "angle": 0, + "content": "3.3. Analysis and Design Insights" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.588, + 0.483, + 0.678 + ], + "angle": 0, + "content": "Through a comprehensive testing program and a detailed examination of the densification and rendering formulas in 3DGS, we have identified the acceleration effect of the opacity attribute in the 3DGS training process. Based on these insights, we propose an Opacity-based Prioritised Sampling (OPS) strategy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.679, + 0.483, + 0.709 + ], + "angle": 0, + "content": "The Opacity-based sampling probability \\( p_i \\) for selecting an Under-training Gaussian \\( G_i \\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.235, + 0.713, + 0.482, + 0.746 + ], + "angle": 0, + "content": "\\[\np _ {i} = \\frac {\\alpha_ {i}}{\\sum_ {i = 1} ^ {N} \\alpha_ {i}} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.756, + 0.483, + 0.787 + ], + "angle": 0, + "content": "where \\(\\alpha_{i}\\) represents the opacity of Gaussian primitive \\(G_{i}\\), and \\(N\\) is the total number of Gaussian primitives." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.8, + 0.223, + 0.817 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We undertake an experimental investigation of Group Training within the context of 3D Gaussian Splatting [12] and Mip-Splatting [33]. The experimental settings are detailed in Sec. 4.1. Subsequently, we present the experimental results in Sec. 4.2. Cross-baselines applicability validation" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.094, + 0.9, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.191, + 0.905, + 0.246 + ], + "angle": 0, + "content": "Figure 6. High-opacity Gaussians accelerate 3DGS rendering. Left: The rendering time decreases by \\(40\\%\\) as \\(\\mu_{o}\\) increases. Right: The mechanisms that highers-opacity Gaussians achieve faster \\(\\alpha\\) saturation with less number of Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.257, + 0.905, + 0.319 + ], + "angle": 0, + "content": "across diverse scenarios is further presented in Sec. 4.3. The ablation experiments and their results are comprehensively discussed in Sec. 4.4. Finally, we analyse the results of different sampling strategies in Sec. 4.5." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.329, + 0.722, + 0.345 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.351, + 0.905, + 0.412 + ], + "angle": 0, + "content": "Implementation Details. The Group Training method was integrated as a plug-in to facilitate its utilisation in conjunction with 3DGS. We implemented our method within both 3D Gaussian Splatting [12] and Mip-Splatting [33]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.413, + 0.906, + 0.624 + ], + "angle": 0, + "content": "Group Training was implemented with a default grouping interval of 500 iterations throughout the training process. We employed both Random Sampling (RS) and Opacity-Based Prioritized Sampling (OPS) strategies with an Under-Training Ratio (UTR) of 0.6. To ensure computational stability, considering the importance of initial Gaussians [4], Group Training was activated at the 500th iteration after densification begins. The global densification \\((I_{\\mathrm{D}})\\) and optimization \\((I_{\\mathrm{O}})\\) were scheduled at \\(14.5\\mathrm{K}\\) and \\(29\\mathrm{K}\\) iterations, respectively. To manage the increased GPU memory consumption during group merging, we fixed the SH coefficients [25, 28] at iterations \\(I_{\\mathrm{D}}\\) and \\(I_{\\mathrm{O}}\\). We also conducted experiments applying Group Training solely during the densification phase." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.624, + 0.906, + 0.729 + ], + "angle": 0, + "content": "Dataset and Metrics. The efficacy of our method is evaluated on the Mip-NeRF360 [1], Tanks & Temples [14], Deep Blending [9], and NeRF-Synthetic [21] using a single RTX 3090. Subsequently, the model's reconstruction performance is evaluated using PSNR, SSIM, and LPIPS. Furthermore, the training time, peak memory usage, and model size for each scene are recorded." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.74, + 0.709, + 0.756 + ], + "angle": 0, + "content": "4.2. Experiments Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.763, + 0.905, + 0.794 + ], + "angle": 0, + "content": "This section presents the experimental results of integrating the Group-Training method into the 3D Gaussian Splat" + }, + { + "type": "page_footnote", + "bbox": [ + 0.515, + 0.804, + 0.905, + 0.828 + ], + "angle": 0, + "content": "https://github.com/graphdeco-inria/gaussian-splatting" + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.829, + 0.904, + 0.853 + ], + "angle": 0, + "content": "https://github.com/autonomousvision/mip-splatting" + }, + { + "type": "page_footnote", + "bbox": [ + 0.539, + 0.853, + 0.808, + 0.864 + ], + "angle": 0, + "content": "https://jonbarron.info/mipnerf360/" + }, + { + "type": "page_footnote", + "bbox": [ + 0.539, + 0.865, + 0.792, + 0.876 + ], + "angle": 0, + "content": "https://www.tanksandtemples.org/" + }, + { + "type": "page_footnote", + "bbox": [ + 0.539, + 0.877, + 0.893, + 0.888 + ], + "angle": 0, + "content": "http://visual.cs.ucl.ac.uk/pubs/deepblending/" + }, + { + "type": "page_footnote", + "bbox": [ + 0.539, + 0.889, + 0.809, + 0.9 + ], + "angle": 0, + "content": "https://www.matthewancik.com/nerf" + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.804, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27972" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.905, + 0.206 + ], + "angle": 0, + "content": "
Grouping IterationsMip-NeRF360 [1]Tanks & Temples [14]Deep Blending [9]Blender [21]
PM ↓Size ↓Time ↓PM ↓Size ↓Time ↓PM ↓Size ↓Time ↓PM ↓Size ↓Time ↓
3D Gaussian Splatting*-8.6979226.724.5943415.007.7266623.902.78696.06
Group Training w/ RS0~15K9.5190726.625.0849614.607.6464421.852.76635.66
0~30K9.4890222.585.0449512.257.6364318.452.76635.26
Group Training w/ OPS0~15K8.5767822.534.5438412.806.8148719.252.65435.01
0~30K8.5767919.564.5538410.956.8048716.852.64434.79
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.216, + 0.907, + 0.273 + ], + "angle": 0, + "content": "Table 1. Comparison of reconstruction efficiency for 3DGS [12]. Group Training significantly enhances reconstruction speed across all four datasets, with the OPS strategy achieving the highest acceleration. This effect is particularly notable in complex scenes. While Group Training with RS produces a more redundant model, the OPS strategy yields more compact models and reduced GPU peak memory usage. * indicates that the model was retrained. PM stands for GPU peak memory allocation, with Size in MB and Time in minutes." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.285, + 0.905, + 0.382 + ], + "angle": 0, + "content": "
Grouping IterationsMip-NeRF360 [1]Tanks&Temples [14]Deep Blending [9]Blender [21]
PSNR ↑SSIM ↑LPIPS ↓Time ↓PSNR ↑SSIM ↑LPIPS ↓Time ↓PSNR ↑SSIM ↑LPIPS ↓Time ↓PSNR ↑SSIM ↑LPIPS ↓Time ↓
3D GS [12]-27.2050.8150.2143-23.1420.8410.183-29.4050.9030.2425-33.330.9690.030-
3D GS*-27.4450.8160.215526.723.6970.8490.176415.029.5860.9040.243723.933.7720.9700.03066.1
Group Training0~15K27.6210.8220.207426.623.7730.8510.170314.629.1850.9010.247321.933.9590.9710.02845.7
w/ RS0~30K27.5370.8210.215722.623.7030.8480.182312.329.4170.9070.248118.533.8770.9710.02955.3
Group Training0~15K27.5820.8200.210322.523.8420.8510.172612.829.7130.9060.243519.333.8890.9710.02925.0
w/ OPS0~30K27.5640.8200.213319.623.8530.8500.176411.029.7460.9070.245016.933.8080.9700.02994.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.392, + 0.908, + 0.435 + ], + "angle": 0, + "content": "Table 2. Comparison of reconstruction quality for 3DGS [12]. Group Training simultaneously accelerates the 3DGS scene reconstruction process and improves the accuracy of 3D reconstructions. Notably, Group Training with OPS achieves a notable improvement in reconstruction quality while maintaining the highest reconstruction speed." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.454, + 0.482, + 0.53 + ], + "angle": 0, + "content": "ting [12] and Mip-Splatting [33]. The test outcomes demonstrate that the Group Training with OPS offers a significant reduction in training time compared to the baseline in all scenes, while concurrently enhancing the rendering quality of the novel view, as presented in Fig. 7." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.537, + 0.483, + 0.598 + ], + "angle": 0, + "content": "3D Gaussian Splatting. We first tested the effectiveness of Group Training based on 3DGS. The reconstruction efficiency for all scenes is summarized in Tab. 1, and Tab. 2 presents the NVS rendering quality across all scenes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.869 + ], + "angle": 0, + "content": "Across these scenes, Group Training consistently resulted in faster reconstruction, and this speedup was more pronounced in complex scenes. However, Group Training with RS could lead to an overabundance of redundant Gaussian primitives in some reconstruction tasks, which complicates subsequent operations, such as model compression [24, 25, 28]. Additionally, GPU memory usage significantly increased due to this redundancy [17], as shown in Tab. 1. In comparison to Group Training with RS, Group Training with OPS demonstrated a more pronounced improvement in reconstruction speed. For the reconstruction of all three complex scenes, the average speed increase was about \\(30\\%\\). Moreover, OPS greatly reduced the generation of redundant Gaussian primitives, leading to a smaller reconstruction model size than the original 3DGS reconstruction model in all tasks. Across all reconstruction tasks, the model size was reduced by between \\(10\\%\\) and \\(40\\%\\), resulting in a more compact model overall." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Mip-Splatting. We tested Mip-Splatting with Group Training, another significant variant of the Gaussian Splat" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.454, + 0.907, + 0.621 + ], + "angle": 0, + "content": "ting model. The experimental results are presented in Tab. 3 and Tab. 4. Overall, the performance of Group Training with Mip-Splatting was similar to 3DGS. Group Training with RS consistently improved both reconstruction quality and speed for Mip-Splatting but produced the highest number of Gaussian primitives during the densification phase. Group Training with OPS achieved the fastest reconstruction speed, and its reconstruction quality was nearly optimal. The OPS reduced the proportion of densification and resulted in a smaller model size that was either the smallest among the tested methods or comparable to the baseline." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.635, + 0.905, + 0.665 + ], + "angle": 0, + "content": "Counterintuitive Findings. Our experiments reveal two notable findings that challenge conventional understanding:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.667, + 0.905, + 0.727 + ], + "angle": 0, + "content": "1. Despite producing larger model sizes, Group Training with RS consistently improves reconstruction speed, indicating that training dynamics rather than model size dominates reconstruction efficiency." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.727, + 0.905, + 0.801 + ], + "angle": 0, + "content": "2. Group Training with OPS simultaneously achieves better reconstruction quality with reduced model size. This demonstrates that superior performance stem from its fundamental impact on the optimization process rather than from model size." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.667, + 0.905, + 0.801 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.817, + 0.773, + 0.833 + ], + "angle": 0, + "content": "4.3. Methodological Applicability" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Group-Training framework fundamentally addresses accelerated training and rendering in 3DGS through its dataset-agnostic improvement. Designed as a pluggable training component, the method demonstrates dual efficacy in both" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27973" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.905, + 0.199 + ], + "angle": 0, + "content": "
Grouping IterationsTanks & Temples [14]Deep Blending [9]Blender [21]
PSNR ↑SSIM ↑LPIPS ↓Time ↓PSNR ↑SSIM ↑LPIPS ↓Time ↓PSNR ↑SSIM ↑LPIPS ↓Time ↓
Mip-Splatting*-23.7490.8600.156223.029.3580.9030.239035.133.9950.9700.02968.5
Group Training w/ RS0~15K23.9530.8630.150424.628.9290.9000.243131.734.2030.9710.02789.3
0~30K24.1460.8620.159820.429.3920.9080.242026.534.1230.9720.02858.3
Group Training w/ OPS0~15K23.9580.8630.151921.729.6650.9070.236428.134.1320.9710.02858.2
0~30K24.1560.8630.155918.229.7880.9080.238424.034.1100.9710.02867.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.208, + 0.907, + 0.251 + ], + "angle": 0, + "content": "Table 3. Comparison of reconstruction quality for Mip-Splatting [33]. Group Training achieves simultaneous improvements in both speed and reconstruction quality for Mip-Splatting across all tasks. Specifically, Group Training with OPS attains the fastest reconstruction speed across all three datasets and delivers the highest reconstruction quality in the Tanks & Temples and Deep Blending dataset." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.264, + 0.905, + 0.358 + ], + "angle": 0, + "content": "
Grouping IterationsOutdoorIndoorAverage
PSNR ↑SSIM ↑LPIPS ↓Size ↓Time ↓PSNR ↑SSIM ↑LPIPS ↓Size ↓Time ↓PSNR ↑SSIM ↑LPIPS ↓Size ↓Time ↓
Mip-Splatting*-24.8140.7480.1995143043.331.2650.9290.173844833.327.6810.8280.188199438.8
Group Training0~15K24.9690.7570.1897201745.931.2660.9310.169755634.927.7680.8340.1808137741.0
w/ RS0~30K25.2160.7630.1969201736.831.1640.9290.175755630.527.8600.8370.1875137734.0
Group Training0~15K24.9810.7540.1940157140.731.4170.9310.172537529.127.8410.8320.1844103935.5
w/ OPS0~30K24.9840.7540.1967157132.931.4070.9300.173937526.427.8390.8320.1865103930.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.368, + 0.907, + 0.424 + ], + "angle": 0, + "content": "Table 4. Quantitative evaluation on the Mip-NeRF360 [1] reconstructed by Mip-Splatting [33]. For the reconstruction of outdoor scenes, Grouping Training with RS yields a larger model size, resulting in the best reconstruction quality. Grouping Training with OPS achieves the fastest reconstruction speed but provides sub-optimal reconstruction quality. For indoor scenes, Grouping Training with OPS continues to offer both the fastest reconstruction speed and the highest reconstruction quality." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.44, + 0.481, + 0.523 + ], + "angle": 0, + "content": "
Cyclic ResampleGlobal DenseifyGlobal OptimizeTanks &Temples
PSNR ↑SSIM ↑LPIPS ↓Size ↓Time ↓
3D GS*---23.6970.8490.176443415.0
Group Training✓23.8660.8500.174929211.8
✓✓23.7690.8490.176523111.0
✓✓23.8350.8500.175448511.8
✓✓✓23.8530.8500.176438411.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.534, + 0.483, + 0.618 + ], + "angle": 0, + "content": "Table 5. Ablation Study of Group Training with OPS on Tanks & Temples [14]. Cyclic Resampling provides the most substantial acceleration for scene reconstruction, while Global Densification further improves efficiency by reducing model size. With Global Densification enabled, Global Optimization enhances the reconstruction quality of the model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.635, + 0.483, + 0.666 + ], + "angle": 0, + "content": "3DGS acceleration model [6] and compression model [5], which is detailed in appendix." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.68, + 0.255, + 0.695 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.703, + 0.483, + 0.763 + ], + "angle": 0, + "content": "We conducted an ablation study on the Tanks & Temples dataset [14] to evaluate the impact of three key components of the Group Training method. The detailed comparison results are provided in Tab. 5." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "The results indicate that cyclic resampling contributes the most to improving training speed. Specifically, during iterations \\( I_{\\mathrm{D}} \\sim 15\\mathrm{K} \\), the Global Densification process adaptively performs cyclic pruning of low-opacity Gaussians within 3DGS, reducing redundant Gaussians, decreasing the model size, and thereby enhancing training efficiency. When Global Densification is enabled, the 3DGS model removes most redundant Gaussians during global densification, resulting in fewer redundant Gaussians in the Caching" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.443, + 0.907, + 0.489 + ], + "angle": 0, + "content": "Group during the Optimization phase. In this scenario, Global Optimization can make full use of all Gaussians, further improving reconstruction performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.498, + 0.633, + 0.513 + ], + "angle": 0, + "content": "4.5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.906, + 0.581 + ], + "angle": 0, + "content": "Our experimental results demonstrate that enabling Group Training and utilizing OPS improved both reconstruction speed and quality. To better clarify these improvements, two key questions arise from these findings:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.907, + 0.884 + ], + "angle": 0, + "content": "- **Why faster?** Group Training significantly accelerates reconstruction primarily by reducing the number of Undertraining Gaussian primitives, thereby decreasing the rendering burden on the model. This speedup is observed even when the final model size is larger than the baseline. In alignment with our earlier analysis in Sec. 3.2, the OPS also contributes notably to this speedup effect. During the densification phase, Group Training with OPS preserves high-opacity Gaussians within the Under-Training Group, without which could lead to instances of both under-reconstruction and over-reconstruction [12, 32]. These issues can lead to redundant densification. As a result, models reconstructed with Group Training and OPS contain fewer redundant Gaussians, leading to smaller model sizes and faster reconstruction overall. Additionally, OPS prioritizes sampling high-opacity Gaussians, which rapidly drives \\(\\alpha\\) saturation, which also discussed in Sec. 3.2. Focusing on these most impactful Gaussians reduces the number of Gaussians needed for splatting, further enhancing training and rendering efficiency." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "- Why better? Group Training with the OPS prioritizes the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27974" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.089, + 0.885, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.561, + 0.908, + 0.619 + ], + "angle": 0, + "content": "Figure 7. Visual Comparison of Grouping Training and Baseline on 3DGS [12] and Mip-Splatting [33]. The Grouping Training method accelerates reconstruction by approximately \\(30\\%\\) while achieving more detailed and accurate scene rendering. Notable improvements can be observed in the picture frame in the \"Dr. Johnson\" scene and the bucket handle in the \"Garden\" scene, where finer details are captured. Furthermore, Grouping Training significantly reduces floaters in Mip-Splatting, enhancing overall reconstruction fidelity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.638, + 0.484, + 0.804 + ], + "angle": 0, + "content": "retention and optimization of high-opacity Gaussian primitives. This achieves comparable or superior fidelity to RS, while significantly reducing the model size. The Gaussian primitives with higher opacity inherently have a greater influence on the final rendered image [5, 19, 22, 28], so emphasizing their optimization leads to significant improvements in overall rendering quality. Besides, cyclic resampling temporarily excluding a randomly selected subset of Gaussian primitives, each remaining primitive is compelled to learn how to independently represent essential scene features leading to a more robust representation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.903 + ], + "angle": 0, + "content": "In summary, by integrating these strategies, Group Training with OPS simultaneously enhances both reconstruction speed and rendering quality. By focusing on the most impactful primitives and encouraging independent feature representation, Group Training achieves efficient, high-quality scene reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.637, + 0.634, + 0.652 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.656, + 0.907, + 0.898 + ], + "angle": 0, + "content": "This paper presents Group Training, an effective strategy that addresses the computational challenges in 3D Gaussian Splitting by dynamically managing Gaussian primitives in organized groups. Through theoretical analysis and extensive experiments, we demonstrate that our method not only accelerates training by up to \\(30\\%\\) but also improves rendering quality across diverse scenarios. Our opacity-based sampling strategy shows universal compatibility with existing 3DGS frameworks, achieving consistent improvements without compromising model complexity. The success of Group Training reveals the importance of dynamic primitive management in 3DGS optimization, suggesting promising directions for future research, such as adaptive grouping strategies and dynamic sampling mechanisms for more complex scenes. Our work provides valuable insights for future developments in efficient novel view synthesis." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27975" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.222 + ], + "angle": 0, + "content": "This work is supported by the National Key R&D Program of China (No.2022ZD0119000), Hunan Provincial Key R&D Program of China (No.2024JK2020 and 2024JK2021), Hunan Provincial Natural Science Foundation of China (No.2024JJ10027), Young Talents of Huxiang (No.Z202433000575), and Changsha Science Fund for Distinguished Young Scholars (kq2306002)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.236, + 0.188, + 0.251 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.26, + 0.484, + 0.302 + ], + "angle": 0, + "content": "[1] Jonathan T. Barron. Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields. In CVPR, 2022. 1, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.304, + 0.484, + 0.345 + ], + "angle": 0, + "content": "[2] Samuel Rota Bulò, Lorenzo Porzi, and Peter Kontschieder. Revising densification in gaussian splatting. arXiv preprint arXiv:2404.06109, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.347, + 0.483, + 0.375 + ], + "angle": 0, + "content": "[3] Guikun Chen and Wenguan Wang. A survey on 3d gaussian splatting. arXiv preprint arXiv:2401.03890, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.376, + 0.483, + 0.417 + ], + "angle": 0, + "content": "[4] Jaeyoung Chung, Jeongtaek Oh, and Kyoung Mu Lee. Depth-regularized optimization for 3d gaussian splatting in few-shot images. arXiv preprint arXiv:2311.13398, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.419, + 0.483, + 0.473 + ], + "angle": 0, + "content": "[5] Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, and Zhangyang Wang. Lightgaussian: Unbounded 3d gaussian compression with \\(15\\mathrm{x}\\) reduction and \\(200+\\) fps, 2023. 3, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.476, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[6] Guangchi Fang and Bing Wang. Mini-splatting: Representing scenes with a constrained number of gaussians, 2024. 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.518, + 0.483, + 0.56 + ], + "angle": 0, + "content": "[7] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.562, + 0.483, + 0.603 + ], + "angle": 0, + "content": "[8] Sharath Girish, Kamal Gupta, and Abhinav Shrivastava. Eagles: Efficient accelerated 3d gaussians with lightweight encodings. arXiv preprint arXiv:2312.04564, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.604, + 0.483, + 0.659 + ], + "angle": 0, + "content": "[9] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep Blending for Free-Viewpoint Image-Based Rendering. In SIGGRAPH, 2018. 1, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.661, + 0.483, + 0.702 + ], + "angle": 0, + "content": "[10] Lukas Hollein, Aljaž Božić, Michael Zollhöfer, and Matthias Nießner. 3dgs-lm: Faster gaussian-splatting optimization with levenberg-marquardt. arXiV:2409.12892, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.704, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[11] Ying Jiang, Chang Yu, Tianyi Xie, Xuan Li, Yutao Feng, Huamin Wang, Minchen Li, Henry Lau, Feng Gao, Yin Yang, et al. Vr-gs: A physical dynamics-aware interactive gaussian splatting system in virtual reality. In ACM SIGGRAPH 2024 Conference Papers, pages 1-1, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.483, + 0.83 + ], + "angle": 0, + "content": "[12] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Shakiba Kheradmand, Daniel Rebain, Gopal Sharma, Weiwei Sun, Yang-Che Tseng, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. 3d gaussian splatting as markov chain monte carlo. Advances in Neural Information Processing Systems, 37:80965-80986, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.26, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[14] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and Temples: Benchmarking Large-Scale Scene Reconstruction. In SIGGRAPH, 2017. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[15] Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaussian feature splatting for real-time dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8508-8520, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[16] Yang Liu, Chuanchen Luo, Lue Fan, Naiyan Wang, Junran Peng, and Zhaoxiang Zhang. Citygaussian: Real-time high-quality large-scale scene rendering with gaussians. In European Conference on Computer Vision, pages 265-282. Springer, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.33 + ], + "angle": 0, + "content": "[17] Jiahao Lu, Yifan Zhang, Qiuhong Shen, Xinchao Wang, and Shuicheng Yan. Poisson-splat: Computation cost attack on 3d gaussian splatting. arXiv preprint arXiv:2410.08190, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.907, + 0.403 + ], + "angle": 0, + "content": "[18] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.406, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[19] Saswat Subhajyoti Mallick, Rahul Goel, Bernhard Kerbl, Markus Steinberger, Francisco Vicente Carrasco, and Fernando De La Torre. Taming 3dgs: High-quality radiance fields with limited resources. In SIGGRAPH Asia 2024 Conference Papers, New York, NY, USA, 2024. Association for Computing Machinery. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.544 + ], + "angle": 0, + "content": "[20] Jiarui Meng, Hajie Li, Yanmin Wu, Qiankun Gao, Shuzhou Yang, Jian Zhang, and Siwei Ma. Mirror-3dgs: Incorporating mirror reflections into 3d gaussian splatting. arXiv preprint arXiv:2404.01168, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[21] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[22] Wieland Morgenstern, Florian Barthel, Anna Hilsmann, and Peter Eisert. Compact 3d scene representation via self-organizing gaussian grids. arXiv preprint arXiv:2312.13299, 2023.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[23] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.718, + 0.905, + 0.772 + ], + "angle": 0, + "content": "[24] KL Navaneet, Kossar Pourahmadi Meibodi, Soroush Abbasi Koohpayegani, and Hamed Pirsiavash. Compgs: Smaller and faster gaussian splatting with vector quantization. ECCV, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[25] Simon Niedermayr, Josef Stumpfegger, and Rüdiger Westermann. Compressed 3d gaussian splattering for accelerated novel view synthesis, 2023. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[26] Michael Niemeyer, Fabian Manhardt, Marie-Julie Rakoto-saona, Michael Oechsle, Daniel Duckworth, Rama Gosula, Keisuke Tateno, John Bates, Dominik Kaeser, and Federico Tombari. Radsplat: Radiance field-informed gaussian splatting for robust real-time rendering with \\(900+\\) fps. arXiv.org, 2024. 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27976" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[27] Panagiotis Papantonakis, Georgios Kopanas, Bernhard Kerbl, Alexandre Lanvin, and George Drettakis. Reducing the memory footprint of 3d gaussian splatting. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 7(1):1-17, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[28] Panagiotis Papantonakis, Georgios Kopanas, Bernhard Kerbl, Alexandre Lanvin, and George Drettakis. Reducing the memory footprint of 3d gaussian splatting. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 2024. 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[29] Zhexi Peng, Tianjia Shao, Yong Liu, Jingke Zhou, Yin Yang, Jingdong Wang, and Kun Zhou. Rtg-slam: Real-time 3d reconstruction at scale using gaussian splatting. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.331 + ], + "angle": 0, + "content": "[30] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.333, + 0.482, + 0.387 + ], + "angle": 0, + "content": "[31] Haosen Yang, Chenhao Zhang, Wenqing Wang, Marco Volino, Adrian Hilton, Li Zhang, and Xiatian Zhu. Localized gaussian point management. arXiv e-prints, pages arXiv-2406, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.39, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[32] Zongxin Ye, Wenyu Li, Sidun Liu, Peng Qiao, and Yong Dou. Absgs: Recovering fine details for 3d gaussian splatt ing, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[33] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In CVPR, 2024. 1, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.531 + ], + "angle": 0, + "content": "[34] Hongjia Zhai, Xiyu Zhang, Boming Zhao, Hai Li, Yijia He, Zhaopeng Cui, Hujun Bao, and Guofeng Zhang. Splatloc: 3d gaussian splatting-based visual localization for augmented reality. arXiv preprint arXiv:2409.14067, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.532, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[35] Zhaoliang Zhang, Tianchen Song, Yongjae Lee, Li Yang, Cheng Peng, Rama Chellappa, and Deliang Fan. Lp-3dgs: Learning to prune 3d gaussian splatting. arXiv preprint arXiv:2405.18784, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.589, + 0.482, + 0.671 + ], + "angle": 0, + "content": "[36] Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21634-21643, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.673, + 0.482, + 0.729 + ], + "angle": 0, + "content": "[37] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. In European Conference on Computer Vision, pages 145-163. Springer, 2025. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.729 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27977" + } + ] +] \ No newline at end of file diff --git a/2025/Faster and Better 3D Splatting via Group Training/21033709-b705-4347-a1bb-c3e7a58591e2_origin.pdf b/2025/Faster and Better 3D Splatting via Group Training/21033709-b705-4347-a1bb-c3e7a58591e2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7553c727f3d5745e882384913458402b226b55e0 --- /dev/null +++ b/2025/Faster and Better 3D Splatting via Group Training/21033709-b705-4347-a1bb-c3e7a58591e2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8c9771b47b731d87320f2a24fb7ae7103b300eb58b807e7f91eb45a93d5d4197 +size 2716280 diff --git a/2025/Faster and Better 3D Splatting via Group Training/full.md b/2025/Faster and Better 3D Splatting via Group Training/full.md new file mode 100644 index 0000000000000000000000000000000000000000..79c0c3e25f33cf8bbd16a1547cc46649197049a8 --- /dev/null +++ b/2025/Faster and Better 3D Splatting via Group Training/full.md @@ -0,0 +1,322 @@ +# Faster and Better 3D Splatting via Group Training + +Chengbo Wang $^{1,2,4,5}$ Guozheng Ma $^{3}$ Yifei Xue $^{1,2,4,5}$ Yizhen Lao $^{1,4,5,\dagger}$ + +$^{1}$ School of Design, Hunan University $^{2}$ DaHe.AI + +$^{3}$ Nanyang Technological University $^{4}$ Lushan Innovation Lab + +5 Key Laboratory of Digital Culture Smart Design Technology, Minstry of Culture and Tourism + +{wangchb, yizhenlao}@hnu.edu.cn, GUOZHENG001@e.ntu.edu.sg, iflyhsueh@gmail.com + +# Abstract + +3D Gaussian Splatting (3DGS) has emerged as a powerful technique for novel view synthesis, demonstrating remarkable capability in high-fidelity scene reconstruction through its Gaussian primitive representations. However, the computational overhead induced by the massive number of primitives poses a significant bottleneck to training efficiency. To overcome this challenge, we propose Group Training, a simple yet effective strategy that organizes Gaussian primitives into manageable groups, optimizing training efficiency and improving rendering quality. This approach shows universal compatibility with existing 3DGS frameworks, including vanilla 3DGS and MipSplatting, consistently achieving accelerated training while maintaining superior synthesis quality. Extensive experiments reveal that our straightforward Group Training strategy achieves up to $30\%$ faster convergence and improved rendering quality across diverse scenarios. Project Website: https://chengbo-wang.github.io/3DGS-with-Group-Training/. + +# 1. Introduction + +Novel view synthesis (NVS) is a pivotal technology in various domains, such as virtual reality [11], augmented reality [34], and autonomous driving [36]. Recently, 3D Gaussian Splitting (3DGS) [12] achieved significant success in NVS due to its real-time [15, 29, 37], high-quality rendering [2, 20]. This high-quality rendering is made possible by using millions of Gaussians [3], each includes a complex set of attributes, namely, position, size, orientation, opacity, and color. These attributes are optimized via multi-view photometric losses. + +- Motivation. However, 3DGS also introduces a potential risk during training, as the exponentially growing Gaussians + +![](images/e12c1d0c85d56fcb917d085b4e8d9aab3296458ada4367a1d0642994a03ee1ae.jpg) +(a) Reconstruction of the "Stump"[1] scene by 3DGS[12]. + +![](images/040807b518210d2411686473d3735d33f8e24fe01daec70141c3a75e4cfc92d6.jpg) + +![](images/d2018b32da4126112037c08c7717a2bcd8c253eea18b944cd57237fe5cc2947f.jpg) +(b) Reconstruction of the "Dr. Johnson"[9] scene by Mip-Splatting[33]. + +![](images/867ac9203a42177cd3afa097d1a40e833ac8f83acd33f602f97529f04d0a92f4.jpg) +Figure 1. Improvements and Illustrations of Applying Group Training in 3DGS[12] and Mip-Splatting[33]. Grouping Training (solid line) achieves a significant increase in reconstruction speed and superior scene quality compared to baseline methods (dashed line). In the "Stump"[1] scene, Grouping Training delivers a $28\%$ faster reconstruction while rendering leaf details with greater clarity. In the "Dr. Johnson"[9] scene, Mip-Splatting with Group Training achieves a $33\%$ reduction in reconstruction time and effectively suppresses the generation of floating artifacts. + +significantly increase the training burden. Consequently, a primary research focus is the development of methods that enable rapid and cost-effective 3D scene reconstruction without compromising the rendering performance. + +During the densification stage, the cyclic pruning of Gaussians with low opacity emerges as a straightforward yet potentially effective strategy [12, 28] for enhancing the training efficiency of 3DGS. Specifically, in each round, 3DGS prunes Gaussians whose opacity falls below a designated threshold $\epsilon_{\alpha}$ , which is a vital parameter that facilitates control over the number of generated Gaussians. However, this method has inherent limitations. If $\epsilon_{\alpha}$ is set too conservatively, the resultant acceleration will be minimal while an excessively lenient threshold defeats the quality of the NVS. Thus, achieving a balance between accelerating the training and preserving the rendering quality poses a considerable challenge. + +- Intuition and contributions. Inspired by the limitations of the cyclic pruning strategy, we propose to cache a portion of the Gaussians instead of directly pruning Gaussians. This "cache" concept can retain the "important" Gaussians while reducing the total number of them. Therefore, we propose a novel Group Training strategy to accelerate the training of 3DGS. Specifically, (1) a portion of Gaussians is cyclically cached during training. These cached Gaussians are temporarily excluded from scene rendering and model training, which greatly reduces the number of Undertraining Gaussians and cuts down on training time. (2) Additionally, through cyclic resampling, the cached Gaussians are reintegrated into the 3DGS model, reducing the impact of direct pruning. (3) A crucial element of Group Training is determining how to sample Gaussians to distinguish between under-training and cached Gaussians. A naive sampling approach would be random sampling. Interestingly, we discovered that the distribution of Gaussian opacity values influences both the quantity of Gaussians generated during densification and the rendering speed of the scene. Therefore, we introduce an Opacity-based Prioritized Sampling method, which effectively reduces the generation of redundant Gaussians and improves the training speed of 3DGS. Specifically, the main technical contributions of our work are as follows: + +1. We propose a simple yet highly efficient training framework called Group Training for 3D Gaussian Splatting, which is seamlessly integrated into existing 3DGS frameworks, including both 3DGS and Mip-Splatting. +2. Our framework significantly improves the efficiency of scene reconstruction and the quality of NVS. + +# 2. Related Work + +In this section, we review relevant work in novel view synthesis and 3D Gaussian Splatting, followed by the distinct + +![](images/20d53fa2901d726544c400c53d7c50e49d51ce9bc9539dcafdc29b85709d5880.jpg) +Figure 2. PSNR and Time Performance of Pruning [12, 28] and Group Training under Varying Hyperparameters. The reconstructions are performed on the "Train" scene [14] using 3DGS [12] as the baseline. Left: The pruning method exhibits substantial instability in optimizing the trade-off between reconstruction efficiency and quality. The hyperparameter sensitivity (Opacity Threshold) presents significant challenges for optimal parameter tuning. Right: Group Training demonstrates consistent improvements in both reconstruction speed and quality, with robust performance across a wide range of hyperparameter values (Caching Ratio), enabling straightforward parameter optimization. + +![](images/18379785de7df5df1624c71a6500b7ffa871fb5b3af417abd41b1f7d782e2186.jpg) + +tion of our approach. + +- NVS. NVS involves generating new views of a scene from different viewpoints than those in the original dataset. Neural Radiance Fields (NeRF) [21] have been pivotal in NVS for producing photo-realistic images through volume rendering. Neural Radiance Fields (NeRF) have significantly advanced NVS by producing photo-realistic images via volume rendering, but they require substantial computational resources and lengthy training due to costly MLP evaluations. To overcome these limitations, methods such as voxel grids [7], hash grids [23] and points [30] have been introduced to facilitate real-time NeRF rendering. + +Recently, 3D Gaussian Splatting (3DGS) [12] has emerged as a promising alternative in the NVS field. Unlike the implicit structure of NeRF, 3DGS uses rasterization rather than ray tracing, enabling efficient rendering suitable for applications requiring both speed and image quality. + +- 3DGS. 3D Gaussian Splatting (3DGS) is a recent approach to high-quality, real-time novel view synthesis that represents scenes explicitly as 3D Gaussian primitives. By rasterizing these Gaussians into 2D splats and blending them using $\alpha$ -blending [21], 3DGS produces detailed, high-resolution outputs, making it well-suited for applications that demand both speed and visual fidelity. + +The 3DGS training process involves two main stages: "Densification" and "Optimization". During the densification phase (0~15K iterations), Adaptive Density + +![](images/338431931c015b47ac68b48761db4d231d197858d06f616db538edde0aec3ac0.jpg) +Figure 3. The overall framework of Group at regular iteration intervals, Gaussians from are categorized into the Under-training Group the Under-training Group is utilized for Ga Group remains inactive and does not participate in any calculations. + +Control [12] is applied to increase the density of Gaussian primitives, thereby enhancing the model's capacity for detailed scene representation. In the optimization phase (15K~30K iterations), densification stops, and the existing Gaussians undergo global refinement, yielding a more precise scene depiction. Recent studies have driven significant advancements in 3DGS across several areas, notably improving rendering quality [2, 13, 33], memory efficiency [5, 25, 27], and scalability for large scenes [16]. Beyond these enhancements, additional research has focused on increasing the efficiency of 3DGS optimizers [10] and refining Gaussian Densification [18, 19, 31] and Pruning strategies [5, 6, 26]. These improvements yield more compact 3DGS models that support faster reconstruction but sacrificing detail. + +- Distinction. Unlike previous optimization-focused approaches, we propose Group Training, a principled strategy that dynamically manages Gaussian primitives during training through controlled group sampling. This approach not only accelerates training but also improves reconstruction quality by maintaining effective primitive optimization, achieving a better balance between computational efficiency and visual fidelity. + +# 3. Methodology + +In Sec. 3.1, we will provide a comprehensive introduction to the Group Training method and its multifunctional advantages in 3D Gaussian Splatting (3DGS). Section 3.2 introduces two sampling paradigms: Random Sampling and Prioritized Sampling (incorporating importance score, opacity-based, and volume-based criteria), with comparative analysis of their computational efficiency and performance trade-offs. We mathematically demonstrate that Opacity-based Prioritized Sampling achieves effective 3DGS densification and efficient rendering. Finally, in Sec. 3.3, we summarize training characteristics of 3DGS with Group Training, and the Sampling Strategies are identified. + +![](images/7f2ded8c80e7931cf9c204c5b93eb9a9b38f142792d1b8cb7ccf609d1da010b8.jpg) +Figure 4. Schedule for activating Group Training during reconstruction. Group Training is enabled at regular intervals. Groups Merging (Grouping without resampling) occurring at iteration $I_{\mathrm{D}}$ and $I_{\mathrm{O}}$ before the completion of the densification and optimization processes for global densification and optimization. + +# 3.1. Group Training + +The number of Gaussian primitives in the Gaussian set increases rapidly during the densification phase to achieve rich representation capabilities. This implies that the computational time to render a scene from a specific viewpoint will also increase, and this effect will extend to the global optimisation phase. To this end, we employ the Group Training method during training. + +For any given set of Gaussian $G$ , the Under-training Group $G_{\text{Under-training}}$ is obtained through sampling, while the remaining primitives constitute the Cached Group $G_{\text{Cached}}$ . + +$$ +G _ {\text {U n d e r - t r a i n i n g}} = \left\{g _ {i} \mid g _ {i} \in G, i \in I, I \subseteq \{1, 2, 3, \dots , | G | \} \right\} \tag {1} +$$ + +$$ +G _ {\text {C a c h e d}} = G \backslash G _ {\text {U n d e r - t r a i n i n g}} \tag {2} +$$ + +Before the next grouping, the Cached Group is merged with the Under-training Group, after which the new $G_{\text{Under-training}}$ and $G_{\text{Cached}}$ groups are resampled. A visual representation of the Group Training process is provided in Fig. 3. This procedure ensures that each Gaussian primitive has an opportunity to contribute to the training process, and effectively reducing the number of Under-training Gaussians while retaining those deemed "important." + +As outlined in Sec. 2, the 3DGS training process consists of two distinct phases: "Densification" and "Optimization". Group Training is facilitated in both phases and concludes with Groups Merging at each phases, as illustrated in Fig. 4. + +# 3.2. Effective and Efficient Sampling Strategies + +Within each grouping, a subset of Gaussians is sampled from the merged Gaussians. The simplest sampling strategy is Random Sampling (RS). We propose attribute-driven Prioritized Sampling to optimize 3DGS further. + +We evaluated various attributes, including opacity and volume-based approaches [5] and importance scores [6, 8, 26, 35] (see appendix for detailed comparisons). Among these, Opacity-based Prioritized Sampling demonstrated superior performance in both training speed and rendering quality. As mathematically proven in Sec. 3.2.1 and 3.2.2, opacity intrinsically governs two critical 3DGS properties, both experimentally validated. + +# 3.2.1. Sampling Strategy for Better Densification + +For any Gaussian primitive $G_{m}$ projected onto the imaging plane with 2D center coordinates $[x_{\mathrm{m}},y_{\mathrm{m}}]$ , the partial derivatives of loss $L$ w.r.t the projected coordinates are $[\frac{\partial L}{\partial x_{\mathrm{m}}},\frac{\partial L}{\partial y_{\mathrm{m}}} ]$ , which govern the densification criterion by: + +$$ +\sqrt {\left(\frac {\partial L}{\partial x _ {\mathrm {m}}}\right) ^ {2} + \left(\frac {\partial L}{\partial y _ {\mathrm {m}}}\right) ^ {2}} > \tau_ {\text {g r a d}}, \tag {3} +$$ + +where $\tau_{\mathrm{grad}}$ denotes the predefined gradient threshold. + +We formulate the first fundamental property of 3DGS: + +- Proposition 1. Opacity-based Effective Gaussians Denomination: Gaussian primitives with higher opacity serve as the primary contributors to densification of 3DGS. + +Proof. Under our assumptions of mutual independence between Gaussian attributes and within-primitive parameter independence, the partial derivatives are computed as: + +$$ +\frac {\partial L}{\partial x _ {\mathrm {m}}} = o _ {m} \sum_ {\text {p i x e l}} \frac {\partial L}{\partial \hat {C}} \frac {\partial \hat {C}}{\partial \alpha_ {m}} \frac {\partial G _ {m} ^ {2 D}}{\partial \Delta x} \frac {\partial \Delta x}{x _ {\mathrm {m}}}, \tag {4} +$$ + +$$ +\frac {\partial L}{\partial y _ {\mathrm {m}}} = o _ {m} \sum_ {\text {p i x e l}} \frac {\partial L}{\partial \hat {C}} \frac {\partial \hat {C}}{\partial \alpha_ {m}} \frac {\partial G _ {m} ^ {2 D}}{\partial \Delta y} \frac {\partial \Delta y}{y _ {\mathrm {m}}}, \tag {5} +$$ + +where $o_m$ denotes the opacity of $G_{m}$ , $\hat{C}$ represents the rendered pixel value, and $\Delta x, \Delta y$ indicate coordinate offsets between pixels and $[x_{\mathrm{m}},y_{\mathrm{m}}]$ . Under the parameter mutual independence assumptions, the $[\frac{\partial\Delta x}{\partial x_{\mathrm{m}}},\frac{\partial\Delta y}{\partial y_{\mathrm{m}}} ],[\frac{\partial G_m^{2\mathrm{D}}}{\partial\Delta x},\frac{\partial G_m^{2\mathrm{D}}}{\partial\Delta y} ]$ and $\frac{\partial L}{\partial\hat{C}}$ remain independent of $o_m$ . We prove that $\frac{\partial\hat{C}}{\partial\alpha_m}$ increases with higher expected value of $o_m$ in appendix: + +$$ +\mathbb {E} \left[ \frac {\partial \hat {C}}{\partial \alpha_ {m}} \right] = \frac {(c _ {0} - c _ {\mathrm {b g}}) T _ {\mathrm {s a t u r a t i o n}}}{1 - \mathbb {E} [ o _ {m} ] \cdot \mathbb {E} [ G _ {m} ^ {\mathrm {2 D}} ]}, \tag {6} +$$ + +where $c_{0}$ and $\alpha_0$ denote the expected color and opacity respectively, and $T_{\mathrm{saturation}}$ controls $\alpha$ saturation [12]. + +With fixed $L$ , Eqs. (4) to (6) demonstrate that elevated opacity distributions yield greater gradient magnitudes in $\left[\frac{\partial L}{\partial x_{\mathrm{m}}}, \frac{\partial L}{\partial y_{\mathrm{m}}}\right]$ , thereby making such Gaussians statistically prioritized for densification. + +![](images/15f31b0e10f5a87f4c7b7a46d85325f1188a34b27cfb393f1b2978f628c81a24.jpg) + +![](images/3a1d709fe8a16b95741037f9a51203efff6972535d4b59cc56e962dcccc3cc7e.jpg) + +![](images/40bbbeeea462a6d92ee1506a0560f5095bae56ca1cc91925b76b194a8f2d31be.jpg) +The distribution of opacity. +The distribution of volume. + +![](images/1844f67d302ca153d70be6c8c8f4d8ed4acc664250e623317b4120d972c51e07.jpg) +Figure 5. The distribution of Gaussian attributes. The distribution of all Gaussian attributes (left) and those contributing specifically to densification (right) in the "Bicycle"[1] with 3DGS [12]. Top row: While the opacities are primarily concentrated around 0 and 1, the Gaussians that contribute to densification are predominantly situated around 1. Bottom row: The distribution of Volume. As densification progresses, Gaussians with lower volume become increasingly involved in the densification process. + +- Experimental Validation: Our experiments track Gaussian primitives contributing to densification, focusing on opacity and volumes values. Statistical analysis of attribute distributions during densification steps (Fig. 5) reveals that Gaussians with high opacity and small volume are the primary sources implementing Gaussian densification. + +Notably, insufficient high-opacity Gaussians increase the photometric loss, exacerbating both under-reconstruction and over-reconstruction [12], thereby forces redundant densification. While small volumes can be employed as a compression metric [5] in the model compression of 3D Gaussian Splatting, they simultaneously drive reasonable densification. Take into account all factors, it is advisable to cache Gaussians with low-opacity for better gaussian densification. + +# 3.2.2. Sampling Strategy for Faster Rendering + +During viewpoint-specific image synthesis in 3DGS, the framework implements $\alpha$ -blending [21] through recursive application of the compositing equation on a per-pixel. + +$$ +C _ {p} = \sum_ {i = 1} ^ {N} c _ {i} \alpha_ {i} T _ {i}, \quad T _ {i} = \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j}), \qquad (7) +$$ + +$$ +N = \min \left\{i \in \mathbb {N} ^ {+}, \left| \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j}) \leq T _ {\text {s a t u r a t i o n}} \right. \right\}, \tag {8} +$$ + +As established in 3DGS [12], the determination of $N$ hinges on achieving opacity saturation within a pixel's accumulated $\alpha$ value (Transmittance $T_{i} < T_{\mathrm{saturation}}$ ). This critical threshold implementation introduces a fundamental constraint on rendering efficiency: during forward rendering operations, the sequential traversal of Gaussian primitives inherently prevents parallel computation of $T_{i}$ values. Consequently, $N$ emerges as the principal determinant of rendering speed, where any reduction on $N$ directly translates to accelerated renderings. + +Based on the preceding analysis, we formulate the second fundamental property of the 3DGS framework: + +- Proposition 2. Opacity-based Efficient Rendering Acceleration: Gaussian primitives with higher opacity enable faster rendering through faster achieving $\alpha$ saturation. + +Proof. Building upon the premise of mutually independent $\alpha_{i}$ values between Gaussians, we posit that the mathematical expectation $\mathbb{E}[T_N] \approx T_{\mathrm{saturation}}$ , + +$$ +\mathbb {E} [ T _ {N} ] = \left(1 - \mathbb {E} [ \alpha_ {i} ]\right) ^ {N} = \left(1 - \mathbb {E} [ o _ {i} ] \cdot \mathbb {E} [ G _ {i} ^ {\mathrm {2 D}} ]\right) ^ {N}, \tag {9} +$$ + +where $\alpha_{i}$ combines its intrinsic opacity $o_i$ and $G_{i}^{2\mathrm{D}}$ (dependent on view projection and covariance, independent of $o_i$ ). We decouple the expected value of $\alpha_{i}$ . + +- Experimental Validation: We generated a 3DGS model with Gaussian $o_i$ values sampled from $\mathcal{N}(\mu_o,0.1)$ , fixing the other parameters. The rendering time as a function of $\mu_o$ is shown in Fig. 6. The results reveals that Gaussian primitives with higher-opacity induce faster transmittance saturation, thereby reducing the blending steps $N$ and rendering computational load. + +# 3.3. Analysis and Design Insights + +Through a comprehensive testing program and a detailed examination of the densification and rendering formulas in 3DGS, we have identified the acceleration effect of the opacity attribute in the 3DGS training process. Based on these insights, we propose an Opacity-based Prioritised Sampling (OPS) strategy. + +The Opacity-based sampling probability $p_i$ for selecting an Under-training Gaussian $G_i$ is defined as: + +$$ +p _ {i} = \frac {\alpha_ {i}}{\sum_ {i = 1} ^ {N} \alpha_ {i}} \tag {10} +$$ + +where $\alpha_{i}$ represents the opacity of Gaussian primitive $G_{i}$ , and $N$ is the total number of Gaussian primitives. + +# 4. Experiments + +We undertake an experimental investigation of Group Training within the context of 3D Gaussian Splatting [12] and Mip-Splatting [33]. The experimental settings are detailed in Sec. 4.1. Subsequently, we present the experimental results in Sec. 4.2. Cross-baselines applicability validation + +![](images/ca078b0976ec7ffb4514a3c52ced74f8fcfb29b6ddc1f5665142e4276de50df0.jpg) +Figure 6. High-opacity Gaussians accelerate 3DGS rendering. Left: The rendering time decreases by $40\%$ as $\mu_{o}$ increases. Right: The mechanisms that highers-opacity Gaussians achieve faster $\alpha$ saturation with less number of Gaussians. + +across diverse scenarios is further presented in Sec. 4.3. The ablation experiments and their results are comprehensively discussed in Sec. 4.4. Finally, we analyse the results of different sampling strategies in Sec. 4.5. + +# 4.1. Experimental Settings + +Implementation Details. The Group Training method was integrated as a plug-in to facilitate its utilisation in conjunction with 3DGS. We implemented our method within both 3D Gaussian Splatting [12] and Mip-Splatting [33]. + +Group Training was implemented with a default grouping interval of 500 iterations throughout the training process. We employed both Random Sampling (RS) and Opacity-Based Prioritized Sampling (OPS) strategies with an Under-Training Ratio (UTR) of 0.6. To ensure computational stability, considering the importance of initial Gaussians [4], Group Training was activated at the 500th iteration after densification begins. The global densification $(I_{\mathrm{D}})$ and optimization $(I_{\mathrm{O}})$ were scheduled at $14.5\mathrm{K}$ and $29\mathrm{K}$ iterations, respectively. To manage the increased GPU memory consumption during group merging, we fixed the SH coefficients [25, 28] at iterations $I_{\mathrm{D}}$ and $I_{\mathrm{O}}$ . We also conducted experiments applying Group Training solely during the densification phase. + +Dataset and Metrics. The efficacy of our method is evaluated on the Mip-NeRF360 [1], Tanks & Temples [14], Deep Blending [9], and NeRF-Synthetic [21] using a single RTX 3090. Subsequently, the model's reconstruction performance is evaluated using PSNR, SSIM, and LPIPS. Furthermore, the training time, peak memory usage, and model size for each scene are recorded. + +# 4.2. Experiments Results + +This section presents the experimental results of integrating the Group-Training method into the 3D Gaussian Splat + +
Grouping IterationsMip-NeRF360 [1]Tanks & Temples [14]Deep Blending [9]Blender [21]
PM ↓Size ↓Time ↓PM ↓Size ↓Time ↓PM ↓Size ↓Time ↓PM ↓Size ↓Time ↓
3D Gaussian Splatting*-8.6979226.724.5943415.007.7266623.902.78696.06
Group Training w/ RS0~15K9.5190726.625.0849614.607.6464421.852.76635.66
0~30K9.4890222.585.0449512.257.6364318.452.76635.26
Group Training w/ OPS0~15K8.5767822.534.5438412.806.8148719.252.65435.01
0~30K8.5767919.564.5538410.956.8048716.852.64434.79
+ +Table 1. Comparison of reconstruction efficiency for 3DGS [12]. Group Training significantly enhances reconstruction speed across all four datasets, with the OPS strategy achieving the highest acceleration. This effect is particularly notable in complex scenes. While Group Training with RS produces a more redundant model, the OPS strategy yields more compact models and reduced GPU peak memory usage. * indicates that the model was retrained. PM stands for GPU peak memory allocation, with Size in MB and Time in minutes. + +
Grouping IterationsMip-NeRF360 [1]Tanks&Temples [14]Deep Blending [9]Blender [21]
PSNR ↑SSIM ↑LPIPS ↓Time ↓PSNR ↑SSIM ↑LPIPS ↓Time ↓PSNR ↑SSIM ↑LPIPS ↓Time ↓PSNR ↑SSIM ↑LPIPS ↓Time ↓
3D GS [12]-27.2050.8150.2143-23.1420.8410.183-29.4050.9030.2425-33.330.9690.030-
3D GS*-27.4450.8160.215526.723.6970.8490.176415.029.5860.9040.243723.933.7720.9700.03066.1
Group Training0~15K27.6210.8220.207426.623.7730.8510.170314.629.1850.9010.247321.933.9590.9710.02845.7
w/ RS0~30K27.5370.8210.215722.623.7030.8480.182312.329.4170.9070.248118.533.8770.9710.02955.3
Group Training0~15K27.5820.8200.210322.523.8420.8510.172612.829.7130.9060.243519.333.8890.9710.02925.0
w/ OPS0~30K27.5640.8200.213319.623.8530.8500.176411.029.7460.9070.245016.933.8080.9700.02994.8
+ +Table 2. Comparison of reconstruction quality for 3DGS [12]. Group Training simultaneously accelerates the 3DGS scene reconstruction process and improves the accuracy of 3D reconstructions. Notably, Group Training with OPS achieves a notable improvement in reconstruction quality while maintaining the highest reconstruction speed. + +ting [12] and Mip-Splatting [33]. The test outcomes demonstrate that the Group Training with OPS offers a significant reduction in training time compared to the baseline in all scenes, while concurrently enhancing the rendering quality of the novel view, as presented in Fig. 7. + +3D Gaussian Splatting. We first tested the effectiveness of Group Training based on 3DGS. The reconstruction efficiency for all scenes is summarized in Tab. 1, and Tab. 2 presents the NVS rendering quality across all scenes. + +Across these scenes, Group Training consistently resulted in faster reconstruction, and this speedup was more pronounced in complex scenes. However, Group Training with RS could lead to an overabundance of redundant Gaussian primitives in some reconstruction tasks, which complicates subsequent operations, such as model compression [24, 25, 28]. Additionally, GPU memory usage significantly increased due to this redundancy [17], as shown in Tab. 1. In comparison to Group Training with RS, Group Training with OPS demonstrated a more pronounced improvement in reconstruction speed. For the reconstruction of all three complex scenes, the average speed increase was about $30\%$ . Moreover, OPS greatly reduced the generation of redundant Gaussian primitives, leading to a smaller reconstruction model size than the original 3DGS reconstruction model in all tasks. Across all reconstruction tasks, the model size was reduced by between $10\%$ and $40\%$ , resulting in a more compact model overall. + +Mip-Splatting. We tested Mip-Splatting with Group Training, another significant variant of the Gaussian Splat + +ting model. The experimental results are presented in Tab. 3 and Tab. 4. Overall, the performance of Group Training with Mip-Splatting was similar to 3DGS. Group Training with RS consistently improved both reconstruction quality and speed for Mip-Splatting but produced the highest number of Gaussian primitives during the densification phase. Group Training with OPS achieved the fastest reconstruction speed, and its reconstruction quality was nearly optimal. The OPS reduced the proportion of densification and resulted in a smaller model size that was either the smallest among the tested methods or comparable to the baseline. + +Counterintuitive Findings. Our experiments reveal two notable findings that challenge conventional understanding: + +1. Despite producing larger model sizes, Group Training with RS consistently improves reconstruction speed, indicating that training dynamics rather than model size dominates reconstruction efficiency. +2. Group Training with OPS simultaneously achieves better reconstruction quality with reduced model size. This demonstrates that superior performance stem from its fundamental impact on the optimization process rather than from model size. + +# 4.3. Methodological Applicability + +Group-Training framework fundamentally addresses accelerated training and rendering in 3DGS through its dataset-agnostic improvement. Designed as a pluggable training component, the method demonstrates dual efficacy in both + +
Grouping IterationsTanks & Temples [14]Deep Blending [9]Blender [21]
PSNR ↑SSIM ↑LPIPS ↓Time ↓PSNR ↑SSIM ↑LPIPS ↓Time ↓PSNR ↑SSIM ↑LPIPS ↓Time ↓
Mip-Splatting*-23.7490.8600.156223.029.3580.9030.239035.133.9950.9700.02968.5
Group Training w/ RS0~15K23.9530.8630.150424.628.9290.9000.243131.734.2030.9710.02789.3
0~30K24.1460.8620.159820.429.3920.9080.242026.534.1230.9720.02858.3
Group Training w/ OPS0~15K23.9580.8630.151921.729.6650.9070.236428.134.1320.9710.02858.2
0~30K24.1560.8630.155918.229.7880.9080.238424.034.1100.9710.02867.5
+ +Table 3. Comparison of reconstruction quality for Mip-Splatting [33]. Group Training achieves simultaneous improvements in both speed and reconstruction quality for Mip-Splatting across all tasks. Specifically, Group Training with OPS attains the fastest reconstruction speed across all three datasets and delivers the highest reconstruction quality in the Tanks & Temples and Deep Blending dataset. + +
Grouping IterationsOutdoorIndoorAverage
PSNR ↑SSIM ↑LPIPS ↓Size ↓Time ↓PSNR ↑SSIM ↑LPIPS ↓Size ↓Time ↓PSNR ↑SSIM ↑LPIPS ↓Size ↓Time ↓
Mip-Splatting*-24.8140.7480.1995143043.331.2650.9290.173844833.327.6810.8280.188199438.8
Group Training0~15K24.9690.7570.1897201745.931.2660.9310.169755634.927.7680.8340.1808137741.0
w/ RS0~30K25.2160.7630.1969201736.831.1640.9290.175755630.527.8600.8370.1875137734.0
Group Training0~15K24.9810.7540.1940157140.731.4170.9310.172537529.127.8410.8320.1844103935.5
w/ OPS0~30K24.9840.7540.1967157132.931.4070.9300.173937526.427.8390.8320.1865103930.0
+ +Table 4. Quantitative evaluation on the Mip-NeRF360 [1] reconstructed by Mip-Splatting [33]. For the reconstruction of outdoor scenes, Grouping Training with RS yields a larger model size, resulting in the best reconstruction quality. Grouping Training with OPS achieves the fastest reconstruction speed but provides sub-optimal reconstruction quality. For indoor scenes, Grouping Training with OPS continues to offer both the fastest reconstruction speed and the highest reconstruction quality. + +
Cyclic ResampleGlobal DenseifyGlobal OptimizeTanks &Temples
PSNR ↑SSIM ↑LPIPS ↓Size ↓Time ↓
3D GS*---23.6970.8490.176443415.0
Group Training✓23.8660.8500.174929211.8
✓✓23.7690.8490.176523111.0
✓✓23.8350.8500.175448511.8
✓✓✓23.8530.8500.176438411.0
+ +Table 5. Ablation Study of Group Training with OPS on Tanks & Temples [14]. Cyclic Resampling provides the most substantial acceleration for scene reconstruction, while Global Densification further improves efficiency by reducing model size. With Global Densification enabled, Global Optimization enhances the reconstruction quality of the model. + +3DGS acceleration model [6] and compression model [5], which is detailed in appendix. + +# 4.4. Ablation Studies + +We conducted an ablation study on the Tanks & Temples dataset [14] to evaluate the impact of three key components of the Group Training method. The detailed comparison results are provided in Tab. 5. + +The results indicate that cyclic resampling contributes the most to improving training speed. Specifically, during iterations $I_{\mathrm{D}} \sim 15\mathrm{K}$ , the Global Densification process adaptively performs cyclic pruning of low-opacity Gaussians within 3DGS, reducing redundant Gaussians, decreasing the model size, and thereby enhancing training efficiency. When Global Densification is enabled, the 3DGS model removes most redundant Gaussians during global densification, resulting in fewer redundant Gaussians in the Caching + +Group during the Optimization phase. In this scenario, Global Optimization can make full use of all Gaussians, further improving reconstruction performance. + +# 4.5. Discussion + +Our experimental results demonstrate that enabling Group Training and utilizing OPS improved both reconstruction speed and quality. To better clarify these improvements, two key questions arise from these findings: + +- **Why faster?** Group Training significantly accelerates reconstruction primarily by reducing the number of Undertraining Gaussian primitives, thereby decreasing the rendering burden on the model. This speedup is observed even when the final model size is larger than the baseline. In alignment with our earlier analysis in Sec. 3.2, the OPS also contributes notably to this speedup effect. During the densification phase, Group Training with OPS preserves high-opacity Gaussians within the Under-Training Group, without which could lead to instances of both under-reconstruction and over-reconstruction [12, 32]. These issues can lead to redundant densification. As a result, models reconstructed with Group Training and OPS contain fewer redundant Gaussians, leading to smaller model sizes and faster reconstruction overall. Additionally, OPS prioritizes sampling high-opacity Gaussians, which rapidly drives $\alpha$ saturation, which also discussed in Sec. 3.2. Focusing on these most impactful Gaussians reduces the number of Gaussians needed for splatting, further enhancing training and rendering efficiency. + +- Why better? Group Training with the OPS prioritizes the + +![](images/7ac07db189326658c53b3f138a974651370dfa348f08dbbe94f47087001e601f.jpg) +Figure 7. Visual Comparison of Grouping Training and Baseline on 3DGS [12] and Mip-Splatting [33]. The Grouping Training method accelerates reconstruction by approximately $30\%$ while achieving more detailed and accurate scene rendering. Notable improvements can be observed in the picture frame in the "Dr. Johnson" scene and the bucket handle in the "Garden" scene, where finer details are captured. Furthermore, Grouping Training significantly reduces floaters in Mip-Splatting, enhancing overall reconstruction fidelity. + +retention and optimization of high-opacity Gaussian primitives. This achieves comparable or superior fidelity to RS, while significantly reducing the model size. The Gaussian primitives with higher opacity inherently have a greater influence on the final rendered image [5, 19, 22, 28], so emphasizing their optimization leads to significant improvements in overall rendering quality. Besides, cyclic resampling temporarily excluding a randomly selected subset of Gaussian primitives, each remaining primitive is compelled to learn how to independently represent essential scene features leading to a more robust representation. + +In summary, by integrating these strategies, Group Training with OPS simultaneously enhances both reconstruction speed and rendering quality. By focusing on the most impactful primitives and encouraging independent feature representation, Group Training achieves efficient, high-quality scene reconstruction. + +# 5. Conclusion + +This paper presents Group Training, an effective strategy that addresses the computational challenges in 3D Gaussian Splitting by dynamically managing Gaussian primitives in organized groups. Through theoretical analysis and extensive experiments, we demonstrate that our method not only accelerates training by up to $30\%$ but also improves rendering quality across diverse scenarios. Our opacity-based sampling strategy shows universal compatibility with existing 3DGS frameworks, achieving consistent improvements without compromising model complexity. The success of Group Training reveals the importance of dynamic primitive management in 3DGS optimization, suggesting promising directions for future research, such as adaptive grouping strategies and dynamic sampling mechanisms for more complex scenes. Our work provides valuable insights for future developments in efficient novel view synthesis. + +# Acknowledgements + +This work is supported by the National Key R&D Program of China (No.2022ZD0119000), Hunan Provincial Key R&D Program of China (No.2024JK2020 and 2024JK2021), Hunan Provincial Natural Science Foundation of China (No.2024JJ10027), Young Talents of Huxiang (No.Z202433000575), and Changsha Science Fund for Distinguished Young Scholars (kq2306002). + +# References + +[1] Jonathan T. Barron. Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields. In CVPR, 2022. 1, 4, 5, 6, 7, 8 +[2] Samuel Rota Bulò, Lorenzo Porzi, and Peter Kontschieder. Revising densification in gaussian splatting. arXiv preprint arXiv:2404.06109, 2024. 1, 3 +[3] Guikun Chen and Wenguan Wang. A survey on 3d gaussian splatting. arXiv preprint arXiv:2401.03890, 2024. 1 +[4] Jaeyoung Chung, Jeongtaek Oh, and Kyoung Mu Lee. Depth-regularized optimization for 3d gaussian splatting in few-shot images. arXiv preprint arXiv:2311.13398, 2023. 5 +[5] Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, and Zhangyang Wang. Lightgaussian: Unbounded 3d gaussian compression with $15\mathrm{x}$ reduction and $200+$ fps, 2023. 3, 4, 7, 8 +[6] Guangchi Fang and Bing Wang. Mini-splatting: Representing scenes with a constrained number of gaussians, 2024. 3, 4, 7 +[7] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 2 +[8] Sharath Girish, Kamal Gupta, and Abhinav Shrivastava. Eagles: Efficient accelerated 3d gaussians with lightweight encodings. arXiv preprint arXiv:2312.04564, 2023. 4 +[9] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep Blending for Free-Viewpoint Image-Based Rendering. In SIGGRAPH, 2018. 1, 5, 6, 7, 8 +[10] Lukas Hollein, Aljaž Božić, Michael Zollhöfer, and Matthias Nießner. 3dgs-lm: Faster gaussian-splatting optimization with levenberg-marquardt. arXiV:2409.12892, 2024. 3 +[11] Ying Jiang, Chang Yu, Tianyi Xie, Xuan Li, Yutao Feng, Huamin Wang, Minchen Li, Henry Lau, Feng Gao, Yin Yang, et al. Vr-gs: A physical dynamics-aware interactive gaussian splatting system in virtual reality. In ACM SIGGRAPH 2024 Conference Papers, pages 1-1, 2024. 1 +[12] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 2, 3, 4, 5, 6, 7, 8 +[13] Shakiba Kheradmand, Daniel Rebain, Gopal Sharma, Weiwei Sun, Yang-Che Tseng, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. 3d gaussian splatting as markov chain monte carlo. Advances in Neural Information Processing Systems, 37:80965-80986, 2024. 3 + +[14] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and Temples: Benchmarking Large-Scale Scene Reconstruction. In SIGGRAPH, 2017. 2, 5, 6, 7, 8 +[15] Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaussian feature splatting for real-time dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8508-8520, 2024. 1 +[16] Yang Liu, Chuanchen Luo, Lue Fan, Naiyan Wang, Junran Peng, and Zhaoxiang Zhang. Citygaussian: Real-time high-quality large-scale scene rendering with gaussians. In European Conference on Computer Vision, pages 265-282. Springer, 2025. 3 +[17] Jiahao Lu, Yifan Zhang, Qiuhong Shen, Xinchao Wang, and Shuicheng Yan. Poisson-splat: Computation cost attack on 3d gaussian splatting. arXiv preprint arXiv:2410.08190, 2024. 6 +[18] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 3 +[19] Saswat Subhajyoti Mallick, Rahul Goel, Bernhard Kerbl, Markus Steinberger, Francisco Vicente Carrasco, and Fernando De La Torre. Taming 3dgs: High-quality radiance fields with limited resources. In SIGGRAPH Asia 2024 Conference Papers, New York, NY, USA, 2024. Association for Computing Machinery. 3, 8 +[20] Jiarui Meng, Hajie Li, Yanmin Wu, Qiankun Gao, Shuzhou Yang, Jian Zhang, and Siwei Ma. Mirror-3dgs: Incorporating mirror reflections into 3d gaussian splatting. arXiv preprint arXiv:2404.01168, 2024. 1 +[21] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 4, 5, 6, 7 +[22] Wieland Morgenstern, Florian Barthel, Anna Hilsmann, and Peter Eisert. Compact 3d scene representation via self-organizing gaussian grids. arXiv preprint arXiv:2312.13299, 2023.8 +[23] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 2 +[24] KL Navaneet, Kossar Pourahmadi Meibodi, Soroush Abbasi Koohpayegani, and Hamed Pirsiavash. Compgs: Smaller and faster gaussian splatting with vector quantization. ECCV, 2024. 6 +[25] Simon Niedermayr, Josef Stumpfegger, and Rüdiger Westermann. Compressed 3d gaussian splattering for accelerated novel view synthesis, 2023. 3, 5, 6 +[26] Michael Niemeyer, Fabian Manhardt, Marie-Julie Rakoto-saona, Michael Oechsle, Daniel Duckworth, Rama Gosula, Keisuke Tateno, John Bates, Dominik Kaeser, and Federico Tombari. Radsplat: Radiance field-informed gaussian splatting for robust real-time rendering with $900+$ fps. arXiv.org, 2024. 3, 4 + +[27] Panagiotis Papantonakis, Georgios Kopanas, Bernhard Kerbl, Alexandre Lanvin, and George Drettakis. Reducing the memory footprint of 3d gaussian splatting. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 7(1):1-17, 2024. 3 +[28] Panagiotis Papantonakis, Georgios Kopanas, Bernhard Kerbl, Alexandre Lanvin, and George Drettakis. Reducing the memory footprint of 3d gaussian splatting. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 2024. 2, 5, 6, 8 +[29] Zhexi Peng, Tianjia Shao, Yong Liu, Jingke Zhou, Yin Yang, Jingdong Wang, and Kun Zhou. Rtg-slam: Real-time 3d reconstruction at scale using gaussian splatting. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 1 +[30] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In CVPR, 2022. 2 +[31] Haosen Yang, Chenhao Zhang, Wenqing Wang, Marco Volino, Adrian Hilton, Li Zhang, and Xiatian Zhu. Localized gaussian point management. arXiv e-prints, pages arXiv-2406, 2024. 3 +[32] Zongxin Ye, Wenyu Li, Sidun Liu, Peng Qiao, and Yong Dou. Absgs: Recovering fine details for 3d gaussian splatt ing, 2024. 7 +[33] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In CVPR, 2024. 1, 3, 5, 6, 7, 8 +[34] Hongjia Zhai, Xiyu Zhang, Boming Zhao, Hai Li, Yijia He, Zhaopeng Cui, Hujun Bao, and Guofeng Zhang. Splatloc: 3d gaussian splatting-based visual localization for augmented reality. arXiv preprint arXiv:2409.14067, 2024. 1 +[35] Zhaoliang Zhang, Tianchen Song, Yongjae Lee, Li Yang, Cheng Peng, Rama Chellappa, and Deliang Fan. Lp-3dgs: Learning to prune 3d gaussian splatting. arXiv preprint arXiv:2405.18784, 2024. 4 +[36] Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21634-21643, 2024. 1 +[37] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. In European Conference on Computer Vision, pages 145-163. Springer, 2025. 1 \ No newline at end of file diff --git a/2025/Faster and Better 3D Splatting via Group Training/images.zip b/2025/Faster and Better 3D Splatting via Group Training/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..23e40244e227cc3ae4bd166251b6beb2d0fab74d --- /dev/null +++ b/2025/Faster and Better 3D Splatting via Group Training/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6236f2021a49700c6b522ab0ca38e395b5094879e203e8f3e259c17408779745 +size 819054 diff --git a/2025/Faster and Better 3D Splatting via Group Training/layout.json b/2025/Faster and Better 3D Splatting via Group Training/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5879b0732541068c0964e361eef3c34dcb8b8f3a --- /dev/null +++ b/2025/Faster and Better 3D Splatting via Group Training/layout.json @@ -0,0 +1,8501 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 148, + 103, + 463, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 103, + 463, + 121 + ], + "spans": [ + { + "bbox": [ + 148, + 103, + 463, + 121 + ], + "type": "text", + "content": "Faster and Better 3D Splatting via Group Training" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 141, + 504, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 141, + 504, + 157 + ], + "spans": [ + { + "bbox": [ + 104, + 141, + 504, + 157 + ], + "type": "text", + "content": "Chengbo Wang" + }, + { + "bbox": [ + 104, + 141, + 504, + 157 + ], + "type": "inline_equation", + "content": "^{1,2,4,5}" + }, + { + "bbox": [ + 104, + 141, + 504, + 157 + ], + "type": "text", + "content": " Guozheng Ma" + }, + { + "bbox": [ + 104, + 141, + 504, + 157 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 104, + 141, + 504, + 157 + ], + "type": "text", + "content": " Yifei Xue" + }, + { + "bbox": [ + 104, + 141, + 504, + 157 + ], + "type": "inline_equation", + "content": "^{1,2,4,5}" + }, + { + "bbox": [ + 104, + 141, + 504, + 157 + ], + "type": "text", + "content": " Yizhen Lao" + }, + { + "bbox": [ + 104, + 141, + 504, + 157 + ], + "type": "inline_equation", + "content": "^{1,4,5,\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 184, + 157, + 427, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 157, + 427, + 171 + ], + "spans": [ + { + "bbox": [ + 184, + 157, + 427, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 184, + 157, + 427, + 171 + ], + "type": "text", + "content": "School of Design, Hunan University " + }, + { + "bbox": [ + 184, + 157, + 427, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 184, + 157, + 427, + 171 + ], + "type": "text", + "content": "DaHe.AI" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 147, + 171, + 463, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 171, + 463, + 186 + ], + "spans": [ + { + "bbox": [ + 147, + 171, + 463, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 147, + 171, + 463, + 186 + ], + "type": "text", + "content": "Nanyang Technological University " + }, + { + "bbox": [ + 147, + 171, + 463, + 186 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 147, + 171, + 463, + 186 + ], + "type": "text", + "content": "Lushan Innovation Lab" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 76, + 185, + 535, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 185, + 535, + 200 + ], + "spans": [ + { + "bbox": [ + 76, + 185, + 535, + 200 + ], + "type": "text", + "content": "5 Key Laboratory of Digital Culture Smart Design Technology, Minstry of Culture and Tourism" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 92, + 201, + 515, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 201, + 515, + 213 + ], + "spans": [ + { + "bbox": [ + 92, + 201, + 515, + 213 + ], + "type": "text", + "content": "{wangchb, yizhenlao}@hnu.edu.cn, GUOZHENG001@e.ntu.edu.sg, iflyhsueh@gmail.com" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 240, + 200, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 240, + 200, + 253 + ], + "spans": [ + { + "bbox": [ + 151, + 240, + 200, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 266, + 297, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 266, + 297, + 494 + ], + "spans": [ + { + "bbox": [ + 54, + 266, + 297, + 494 + ], + "type": "text", + "content": "3D Gaussian Splatting (3DGS) has emerged as a powerful technique for novel view synthesis, demonstrating remarkable capability in high-fidelity scene reconstruction through its Gaussian primitive representations. However, the computational overhead induced by the massive number of primitives poses a significant bottleneck to training efficiency. To overcome this challenge, we propose Group Training, a simple yet effective strategy that organizes Gaussian primitives into manageable groups, optimizing training efficiency and improving rendering quality. This approach shows universal compatibility with existing 3DGS frameworks, including vanilla 3DGS and MipSplatting, consistently achieving accelerated training while maintaining superior synthesis quality. Extensive experiments reveal that our straightforward Group Training strategy achieves up to " + }, + { + "bbox": [ + 54, + 266, + 297, + 494 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 54, + 266, + 297, + 494 + ], + "type": "text", + "content": " faster convergence and improved rendering quality across diverse scenarios. Project Website: https://chengbo-wang.github.io/3DGS-with-Group-Training/." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 517, + 135, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 135, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 135, + 529 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 537, + 296, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 537, + 296, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 537, + 296, + 658 + ], + "type": "text", + "content": "Novel view synthesis (NVS) is a pivotal technology in various domains, such as virtual reality [11], augmented reality [34], and autonomous driving [36]. Recently, 3D Gaussian Splitting (3DGS) [12] achieved significant success in NVS due to its real-time [15, 29, 37], high-quality rendering [2, 20]. This high-quality rendering is made possible by using millions of Gaussians [3], each includes a complex set of attributes, namely, position, size, orientation, opacity, and color. These attributes are optimized via multi-view photometric losses." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 670, + 296, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 670, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 670, + 296, + 694 + ], + "type": "text", + "content": "- Motivation. However, 3DGS also introduces a potential risk during training, as the exponentially growing Gaussians" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 322, + 248, + 455, + 415 + ], + "blocks": [ + { + "bbox": [ + 322, + 248, + 455, + 415 + ], + "lines": [ + { + "bbox": [ + 322, + 248, + 455, + 415 + ], + "spans": [ + { + "bbox": [ + 322, + 248, + 455, + 415 + ], + "type": "image", + "image_path": "e12c1d0c85d56fcb917d085b4e8d9aab3296458ada4367a1d0642994a03ee1ae.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 416, + 522, + 426 + ], + "lines": [ + { + "bbox": [ + 347, + 416, + 522, + 426 + ], + "spans": [ + { + "bbox": [ + 347, + 416, + 522, + 426 + ], + "type": "text", + "content": "(a) Reconstruction of the \"Stump\"[1] scene by 3DGS[12]." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 457, + 251, + 541, + 413 + ], + "blocks": [ + { + "bbox": [ + 457, + 251, + 541, + 413 + ], + "lines": [ + { + "bbox": [ + 457, + 251, + 541, + 413 + ], + "spans": [ + { + "bbox": [ + 457, + 251, + 541, + 413 + ], + "type": "image", + "image_path": "040807b518210d2411686473d3735d33f8e24fe01daec70141c3a75e4cfc92d6.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 323, + 429, + 455, + 594 + ], + "blocks": [ + { + "bbox": [ + 323, + 429, + 455, + 594 + ], + "lines": [ + { + "bbox": [ + 323, + 429, + 455, + 594 + ], + "spans": [ + { + "bbox": [ + 323, + 429, + 455, + 594 + ], + "type": "image", + "image_path": "d2018b32da4126112037c08c7717a2bcd8c253eea18b944cd57237fe5cc2947f.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 596, + 541, + 606 + ], + "lines": [ + { + "bbox": [ + 326, + 596, + 541, + 606 + ], + "spans": [ + { + "bbox": [ + 326, + 596, + 541, + 606 + ], + "type": "text", + "content": "(b) Reconstruction of the \"Dr. Johnson\"[9] scene by Mip-Splatting[33]." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 457, + 429, + 540, + 590 + ], + "blocks": [ + { + "bbox": [ + 457, + 429, + 540, + 590 + ], + "lines": [ + { + "bbox": [ + 457, + 429, + 540, + 590 + ], + "spans": [ + { + "bbox": [ + 457, + 429, + 540, + 590 + ], + "type": "image", + "image_path": "867ac9203a42177cd3afa097d1a40e833ac8f83acd33f602f97529f04d0a92f4.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 615, + 555, + 715 + ], + "lines": [ + { + "bbox": [ + 313, + 615, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 615, + 555, + 715 + ], + "type": "text", + "content": "Figure 1. Improvements and Illustrations of Applying Group Training in 3DGS[12] and Mip-Splatting[33]. Grouping Training (solid line) achieves a significant increase in reconstruction speed and superior scene quality compared to baseline methods (dashed line). In the \"Stump\"[1] scene, Grouping Training delivers a " + }, + { + "bbox": [ + 313, + 615, + 555, + 715 + ], + "type": "inline_equation", + "content": "28\\%" + }, + { + "bbox": [ + 313, + 615, + 555, + 715 + ], + "type": "text", + "content": " faster reconstruction while rendering leaf details with greater clarity. In the \"Dr. Johnson\"[9] scene, Mip-Splatting with Group Training achieves a " + }, + { + "bbox": [ + 313, + 615, + 555, + 715 + ], + "type": "inline_equation", + "content": "33\\%" + }, + { + "bbox": [ + 313, + 615, + 555, + 715 + ], + "type": "text", + "content": " reduction in reconstruction time and effectively suppresses the generation of floating artifacts." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 293, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 293, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 293, + 713 + ], + "type": "text", + "content": "† Corresponding author: Yizhen Lao (yizhenlao@hnu.edu.cn)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27968" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "significantly increase the training burden. Consequently, a primary research focus is the development of methods that enable rapid and cost-effective 3D scene reconstruction without compromising the rendering performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 295, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 295, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 295, + 276 + ], + "type": "text", + "content": "During the densification stage, the cyclic pruning of Gaussians with low opacity emerges as a straightforward yet potentially effective strategy [12, 28] for enhancing the training efficiency of 3DGS. Specifically, in each round, 3DGS prunes Gaussians whose opacity falls below a designated threshold " + }, + { + "bbox": [ + 55, + 121, + 295, + 276 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\alpha}" + }, + { + "bbox": [ + 55, + 121, + 295, + 276 + ], + "type": "text", + "content": ", which is a vital parameter that facilitates control over the number of generated Gaussians. However, this method has inherent limitations. If " + }, + { + "bbox": [ + 55, + 121, + 295, + 276 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\alpha}" + }, + { + "bbox": [ + 55, + 121, + 295, + 276 + ], + "type": "text", + "content": " is set too conservatively, the resultant acceleration will be minimal while an excessively lenient threshold defeats the quality of the NVS. Thus, achieving a balance between accelerating the training and preserving the rendering quality poses a considerable challenge." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 288, + 295, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 288, + 295, + 585 + ], + "spans": [ + { + "bbox": [ + 56, + 288, + 295, + 585 + ], + "type": "text", + "content": "- Intuition and contributions. Inspired by the limitations of the cyclic pruning strategy, we propose to cache a portion of the Gaussians instead of directly pruning Gaussians. This \"cache\" concept can retain the \"important\" Gaussians while reducing the total number of them. Therefore, we propose a novel Group Training strategy to accelerate the training of 3DGS. Specifically, (1) a portion of Gaussians is cyclically cached during training. These cached Gaussians are temporarily excluded from scene rendering and model training, which greatly reduces the number of Undertraining Gaussians and cuts down on training time. (2) Additionally, through cyclic resampling, the cached Gaussians are reintegrated into the 3DGS model, reducing the impact of direct pruning. (3) A crucial element of Group Training is determining how to sample Gaussians to distinguish between under-training and cached Gaussians. A naive sampling approach would be random sampling. Interestingly, we discovered that the distribution of Gaussian opacity values influences both the quantity of Gaussians generated during densification and the rendering speed of the scene. Therefore, we introduce an Opacity-based Prioritized Sampling method, which effectively reduces the generation of redundant Gaussians and improves the training speed of 3DGS. Specifically, the main technical contributions of our work are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 587, + 295, + 659 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 56, + 587, + 295, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 587, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 587, + 295, + 635 + ], + "type": "text", + "content": "1. We propose a simple yet highly efficient training framework called Group Training for 3D Gaussian Splatting, which is seamlessly integrated into existing 3DGS frameworks, including both 3DGS and Mip-Splatting." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 635, + 295, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 295, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 295, + 659 + ], + "type": "text", + "content": "2. Our framework significantly improves the efficiency of scene reconstruction and the quality of NVS." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 669, + 142, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 669, + 142, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 669, + 142, + 681 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "In this section, we review relevant work in novel view synthesis and 3D Gaussian Splatting, followed by the distinct" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 73, + 434, + 225 + ], + "blocks": [ + { + "bbox": [ + 316, + 73, + 434, + 225 + ], + "lines": [ + { + "bbox": [ + 316, + 73, + 434, + 225 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 434, + 225 + ], + "type": "image", + "image_path": "20d53fa2901d726544c400c53d7c50e49d51ce9bc9539dcafdc29b85709d5880.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 236, + 555, + 357 + ], + "lines": [ + { + "bbox": [ + 313, + 236, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 555, + 357 + ], + "type": "text", + "content": "Figure 2. PSNR and Time Performance of Pruning [12, 28] and Group Training under Varying Hyperparameters. The reconstructions are performed on the \"Train\" scene [14] using 3DGS [12] as the baseline. Left: The pruning method exhibits substantial instability in optimizing the trade-off between reconstruction efficiency and quality. The hyperparameter sensitivity (Opacity Threshold) presents significant challenges for optimal parameter tuning. Right: Group Training demonstrates consistent improvements in both reconstruction speed and quality, with robust performance across a wide range of hyperparameter values (Caching Ratio), enabling straightforward parameter optimization." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 437, + 74, + 552, + 226 + ], + "blocks": [ + { + "bbox": [ + 437, + 74, + 552, + 226 + ], + "lines": [ + { + "bbox": [ + 437, + 74, + 552, + 226 + ], + "spans": [ + { + "bbox": [ + 437, + 74, + 552, + 226 + ], + "type": "image", + "image_path": "18379785de7df5df1624c71a6500b7ffa871fb5b3af417abd41b1f7d782e2186.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 371, + 400, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 400, + 383 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 400, + 383 + ], + "type": "text", + "content": "tion of our approach." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 388, + 553, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 553, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 553, + 519 + ], + "type": "text", + "content": "- NVS. NVS involves generating new views of a scene from different viewpoints than those in the original dataset. Neural Radiance Fields (NeRF) [21] have been pivotal in NVS for producing photo-realistic images through volume rendering. Neural Radiance Fields (NeRF) have significantly advanced NVS by producing photo-realistic images via volume rendering, but they require substantial computational resources and lengthy training due to costly MLP evaluations. To overcome these limitations, methods such as voxel grids [7], hash grids [23] and points [30] have been introduced to facilitate real-time NeRF rendering." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 521, + 553, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 553, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 553, + 582 + ], + "type": "text", + "content": "Recently, 3D Gaussian Splatting (3DGS) [12] has emerged as a promising alternative in the NVS field. Unlike the implicit structure of NeRF, 3DGS uses rasterization rather than ray tracing, enabling efficient rendering suitable for applications requiring both speed and image quality." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 593, + 553, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 593, + 553, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 593, + 553, + 676 + ], + "type": "text", + "content": "- 3DGS. 3D Gaussian Splatting (3DGS) is a recent approach to high-quality, real-time novel view synthesis that represents scenes explicitly as 3D Gaussian primitives. By rasterizing these Gaussians into 2D splats and blending them using " + }, + { + "bbox": [ + 313, + 593, + 553, + 676 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 593, + 553, + 676 + ], + "type": "text", + "content": "-blending [21], 3DGS produces detailed, high-resolution outputs, making it well-suited for applications that demand both speed and visual fidelity." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "The 3DGS training process involves two main stages: \"Densification\" and \"Optimization\". During the densification phase (0~15K iterations), Adaptive Density" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27969" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 71, + 610, + 232 + ], + "blocks": [ + { + "bbox": [ + 72, + 71, + 610, + 232 + ], + "lines": [ + { + "bbox": [ + 72, + 71, + 610, + 232 + ], + "spans": [ + { + "bbox": [ + 72, + 71, + 610, + 232 + ], + "type": "image", + "image_path": "338431931c015b47ac68b48761db4d231d197858d06f616db538edde0aec3ac0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 190, + 303, + 247 + ], + "lines": [ + { + "bbox": [ + 55, + 190, + 303, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 303, + 247 + ], + "type": "text", + "content": "Figure 3. The overall framework of Group at regular iteration intervals, Gaussians from are categorized into the Under-training Group the Under-training Group is utilized for Ga Group remains inactive and does not participate in any calculations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 258, + 296, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 258, + 296, + 438 + ], + "spans": [ + { + "bbox": [ + 54, + 258, + 296, + 438 + ], + "type": "text", + "content": "Control [12] is applied to increase the density of Gaussian primitives, thereby enhancing the model's capacity for detailed scene representation. In the optimization phase (15K~30K iterations), densification stops, and the existing Gaussians undergo global refinement, yielding a more precise scene depiction. Recent studies have driven significant advancements in 3DGS across several areas, notably improving rendering quality [2, 13, 33], memory efficiency [5, 25, 27], and scalability for large scenes [16]. Beyond these enhancements, additional research has focused on increasing the efficiency of 3DGS optimizers [10] and refining Gaussian Densification [18, 19, 31] and Pruning strategies [5, 6, 26]. These improvements yield more compact 3DGS models that support faster reconstruction but sacrificing detail." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 450, + 296, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 296, + 546 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 296, + 546 + ], + "type": "text", + "content": "- Distinction. Unlike previous optimization-focused approaches, we propose Group Training, a principled strategy that dynamically manages Gaussian primitives during training through controlled group sampling. This approach not only accelerates training but also improves reconstruction quality by maintaining effective primitive optimization, achieving a better balance between computational efficiency and visual fidelity." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 550, + 138, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 550, + 138, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 550, + 138, + 563 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "In Sec. 3.1, we will provide a comprehensive introduction to the Group Training method and its multifunctional advantages in 3D Gaussian Splatting (3DGS). Section 3.2 introduces two sampling paradigms: Random Sampling and Prioritized Sampling (incorporating importance score, opacity-based, and volume-based criteria), with comparative analysis of their computational efficiency and performance trade-offs. We mathematically demonstrate that Opacity-based Prioritized Sampling achieves effective 3DGS densification and efficient rendering. Finally, in Sec. 3.3, we summarize training characteristics of 3DGS with Group Training, and the Sampling Strategies are identified." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 322, + 257, + 548, + 316 + ], + "blocks": [ + { + "bbox": [ + 322, + 257, + 548, + 316 + ], + "lines": [ + { + "bbox": [ + 322, + 257, + 548, + 316 + ], + "spans": [ + { + "bbox": [ + 322, + 257, + 548, + 316 + ], + "type": "image", + "image_path": "7f2ded8c80e7931cf9c204c5b93eb9a9b38f142792d1b8cb7ccf609d1da010b8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 325, + 555, + 381 + ], + "lines": [ + { + "bbox": [ + 313, + 325, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 313, + 325, + 555, + 381 + ], + "type": "text", + "content": "Figure 4. Schedule for activating Group Training during reconstruction. Group Training is enabled at regular intervals. Groups Merging (Grouping without resampling) occurring at iteration " + }, + { + "bbox": [ + 313, + 325, + 555, + 381 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{D}}" + }, + { + "bbox": [ + 313, + 325, + 555, + 381 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 325, + 555, + 381 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{O}}" + }, + { + "bbox": [ + 313, + 325, + 555, + 381 + ], + "type": "text", + "content": " before the completion of the densification and optimization processes for global densification and optimization." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 391, + 411, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 391, + 411, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 411, + 403 + ], + "type": "text", + "content": "3.1. Group Training" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 408, + 555, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 555, + 491 + ], + "type": "text", + "content": "The number of Gaussian primitives in the Gaussian set increases rapidly during the densification phase to achieve rich representation capabilities. This implies that the computational time to render a scene from a specific viewpoint will also increase, and this effect will extend to the global optimisation phase. To this end, we employ the Group Training method during training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 491, + 554, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 554, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 554, + 525 + ], + "type": "text", + "content": "For any given set of Gaussian " + }, + { + "bbox": [ + 313, + 491, + 554, + 525 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 491, + 554, + 525 + ], + "type": "text", + "content": ", the Under-training Group " + }, + { + "bbox": [ + 313, + 491, + 554, + 525 + ], + "type": "inline_equation", + "content": "G_{\\text{Under-training}}" + }, + { + "bbox": [ + 313, + 491, + 554, + 525 + ], + "type": "text", + "content": " is obtained through sampling, while the remaining primitives constitute the Cached Group " + }, + { + "bbox": [ + 313, + 491, + 554, + 525 + ], + "type": "inline_equation", + "content": "G_{\\text{Cached}}" + }, + { + "bbox": [ + 313, + 491, + 554, + 525 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 324, + 532, + 555, + 545 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 532, + 555, + 545 + ], + "spans": [ + { + "bbox": [ + 324, + 532, + 555, + 545 + ], + "type": "interline_equation", + "content": "G _ {\\text {U n d e r - t r a i n i n g}} = \\left\\{g _ {i} \\mid g _ {i} \\in G, i \\in I, I \\subseteq \\{1, 2, 3, \\dots , | G | \\} \\right\\} \\tag {1}", + "image_path": "a7d7202a0836482259460c3e7e1caef76f2a7ac5858a3d3985bf2093ecec13a5.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 378, + 552, + 555, + 565 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 552, + 555, + 565 + ], + "spans": [ + { + "bbox": [ + 378, + 552, + 555, + 565 + ], + "type": "interline_equation", + "content": "G _ {\\text {C a c h e d}} = G \\backslash G _ {\\text {U n d e r - t r a i n i n g}} \\tag {2}", + "image_path": "976bbd47759a2d21a5e8fe7d690c703a02d3b0f80b5b5a4f13b79b4796a18652.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 570, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 665 + ], + "type": "text", + "content": "Before the next grouping, the Cached Group is merged with the Under-training Group, after which the new " + }, + { + "bbox": [ + 313, + 570, + 554, + 665 + ], + "type": "inline_equation", + "content": "G_{\\text{Under-training}}" + }, + { + "bbox": [ + 313, + 570, + 554, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 570, + 554, + 665 + ], + "type": "inline_equation", + "content": "G_{\\text{Cached}}" + }, + { + "bbox": [ + 313, + 570, + 554, + 665 + ], + "type": "text", + "content": " groups are resampled. A visual representation of the Group Training process is provided in Fig. 3. This procedure ensures that each Gaussian primitive has an opportunity to contribute to the training process, and effectively reducing the number of Under-training Gaussians while retaining those deemed \"important.\"" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "type": "text", + "content": "As outlined in Sec. 2, the 3DGS training process consists of two distinct phases: \"Densification\" and \"Optimization\". Group Training is facilitated in both phases and concludes with Groups Merging at each phases, as illustrated in Fig. 4." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27970" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 279, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 279, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 279, + 85 + ], + "type": "text", + "content": "3.2. Effective and Efficient Sampling Strategies" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 137 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 137 + ], + "type": "text", + "content": "Within each grouping, a subset of Gaussians is sampled from the merged Gaussians. The simplest sampling strategy is Random Sampling (RS). We propose attribute-driven Prioritized Sampling to optimize 3DGS further." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 137, + 296, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 137, + 296, + 233 + ], + "spans": [ + { + "bbox": [ + 55, + 137, + 296, + 233 + ], + "type": "text", + "content": "We evaluated various attributes, including opacity and volume-based approaches [5] and importance scores [6, 8, 26, 35] (see appendix for detailed comparisons). Among these, Opacity-based Prioritized Sampling demonstrated superior performance in both training speed and rendering quality. As mathematically proven in Sec. 3.2.1 and 3.2.2, opacity intrinsically governs two critical 3DGS properties, both experimentally validated." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 238, + 264, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 264, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 264, + 251 + ], + "type": "text", + "content": "3.2.1. Sampling Strategy for Better Densification" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 253, + 296, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 253, + 296, + 305 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 296, + 305 + ], + "type": "text", + "content": "For any Gaussian primitive " + }, + { + "bbox": [ + 55, + 253, + 296, + 305 + ], + "type": "inline_equation", + "content": "G_{m}" + }, + { + "bbox": [ + 55, + 253, + 296, + 305 + ], + "type": "text", + "content": " projected onto the imaging plane with 2D center coordinates " + }, + { + "bbox": [ + 55, + 253, + 296, + 305 + ], + "type": "inline_equation", + "content": "[x_{\\mathrm{m}},y_{\\mathrm{m}}]" + }, + { + "bbox": [ + 55, + 253, + 296, + 305 + ], + "type": "text", + "content": ", the partial derivatives of loss " + }, + { + "bbox": [ + 55, + 253, + 296, + 305 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 253, + 296, + 305 + ], + "type": "text", + "content": " w.r.t the projected coordinates are " + }, + { + "bbox": [ + 55, + 253, + 296, + 305 + ], + "type": "inline_equation", + "content": "[\\frac{\\partial L}{\\partial x_{\\mathrm{m}}},\\frac{\\partial L}{\\partial y_{\\mathrm{m}}} ]" + }, + { + "bbox": [ + 55, + 253, + 296, + 305 + ], + "type": "text", + "content": ", which govern the densification criterion by:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 311, + 295, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 311, + 295, + 338 + ], + "spans": [ + { + "bbox": [ + 115, + 311, + 295, + 338 + ], + "type": "interline_equation", + "content": "\\sqrt {\\left(\\frac {\\partial L}{\\partial x _ {\\mathrm {m}}}\\right) ^ {2} + \\left(\\frac {\\partial L}{\\partial y _ {\\mathrm {m}}}\\right) ^ {2}} > \\tau_ {\\text {g r a d}}, \\tag {3}", + "image_path": "98711092f800cbb4febd15d3e20e861cf4e93093a0c193d769657a2d40394548.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 340, + 276, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 276, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 276, + 353 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 340, + 276, + 353 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{grad}}" + }, + { + "bbox": [ + 55, + 340, + 276, + 353 + ], + "type": "text", + "content": " denotes the predefined gradient threshold." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 353, + 288, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 353, + 288, + 364 + ], + "spans": [ + { + "bbox": [ + 67, + 353, + 288, + 364 + ], + "type": "text", + "content": "We formulate the first fundamental property of 3DGS:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 365, + 295, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 399 + ], + "type": "text", + "content": "- Proposition 1. Opacity-based Effective Gaussians Denomination: Gaussian primitives with higher opacity serve as the primary contributors to densification of 3DGS." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 400, + 296, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 400, + 296, + 436 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 296, + 436 + ], + "type": "text", + "content": "Proof. Under our assumptions of mutual independence between Gaussian attributes and within-primitive parameter independence, the partial derivatives are computed as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 103, + 443, + 295, + 471 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 443, + 295, + 471 + ], + "spans": [ + { + "bbox": [ + 103, + 443, + 295, + 471 + ], + "type": "interline_equation", + "content": "\\frac {\\partial L}{\\partial x _ {\\mathrm {m}}} = o _ {m} \\sum_ {\\text {p i x e l}} \\frac {\\partial L}{\\partial \\hat {C}} \\frac {\\partial \\hat {C}}{\\partial \\alpha_ {m}} \\frac {\\partial G _ {m} ^ {2 D}}{\\partial \\Delta x} \\frac {\\partial \\Delta x}{x _ {\\mathrm {m}}}, \\tag {4}", + "image_path": "fe0b07b9bc1b54cf1e634f5eec2b792fc713c1ef579a59029b613fdeab5f104f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 103, + 479, + 295, + 507 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 479, + 295, + 507 + ], + "spans": [ + { + "bbox": [ + 103, + 479, + 295, + 507 + ], + "type": "interline_equation", + "content": "\\frac {\\partial L}{\\partial y _ {\\mathrm {m}}} = o _ {m} \\sum_ {\\text {p i x e l}} \\frac {\\partial L}{\\partial \\hat {C}} \\frac {\\partial \\hat {C}}{\\partial \\alpha_ {m}} \\frac {\\partial G _ {m} ^ {2 D}}{\\partial \\Delta y} \\frac {\\partial \\Delta y}{y _ {\\mathrm {m}}}, \\tag {5}", + "image_path": "b7e4123bb4a45ba6185f153ff0135f18748429f84945bca0bc849f6ee6f1c9eb.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 512, + 296, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 512, + 296, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 512, + 296, + 593 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 512, + 296, + 593 + ], + "type": "inline_equation", + "content": "o_m" + }, + { + "bbox": [ + 55, + 512, + 296, + 593 + ], + "type": "text", + "content": " denotes the opacity of " + }, + { + "bbox": [ + 55, + 512, + 296, + 593 + ], + "type": "inline_equation", + "content": "G_{m}" + }, + { + "bbox": [ + 55, + 512, + 296, + 593 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 512, + 296, + 593 + ], + "type": "inline_equation", + "content": "\\hat{C}" + }, + { + "bbox": [ + 55, + 512, + 296, + 593 + ], + "type": "text", + "content": " represents the rendered pixel value, and " + }, + { + "bbox": [ + 55, + 512, + 296, + 593 + ], + "type": "inline_equation", + "content": "\\Delta x, \\Delta y" + }, + { + "bbox": [ + 55, + 512, + 296, + 593 + ], + "type": "text", + "content": " indicate coordinate offsets between pixels and " + }, + { + "bbox": [ + 55, + 512, + 296, + 593 + ], + "type": "inline_equation", + "content": "[x_{\\mathrm{m}},y_{\\mathrm{m}}]" + }, + { + "bbox": [ + 55, + 512, + 296, + 593 + ], + "type": "text", + "content": ". Under the parameter mutual independence assumptions, the " + }, + { + "bbox": [ + 55, + 512, + 296, + 593 + ], + "type": "inline_equation", + "content": "[\\frac{\\partial\\Delta x}{\\partial x_{\\mathrm{m}}},\\frac{\\partial\\Delta y}{\\partial y_{\\mathrm{m}}} ],[\\frac{\\partial G_m^{2\\mathrm{D}}}{\\partial\\Delta x},\\frac{\\partial G_m^{2\\mathrm{D}}}{\\partial\\Delta y} ]" + }, + { + "bbox": [ + 55, + 512, + 296, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 512, + 296, + 593 + ], + "type": "inline_equation", + "content": "\\frac{\\partial L}{\\partial\\hat{C}}" + }, + { + "bbox": [ + 55, + 512, + 296, + 593 + ], + "type": "text", + "content": " remain independent of " + }, + { + "bbox": [ + 55, + 512, + 296, + 593 + ], + "type": "inline_equation", + "content": "o_m" + }, + { + "bbox": [ + 55, + 512, + 296, + 593 + ], + "type": "text", + "content": ". We prove that " + }, + { + "bbox": [ + 55, + 512, + 296, + 593 + ], + "type": "inline_equation", + "content": "\\frac{\\partial\\hat{C}}{\\partial\\alpha_m}" + }, + { + "bbox": [ + 55, + 512, + 296, + 593 + ], + "type": "text", + "content": " increases with higher expected value of " + }, + { + "bbox": [ + 55, + 512, + 296, + 593 + ], + "type": "inline_equation", + "content": "o_m" + }, + { + "bbox": [ + 55, + 512, + 296, + 593 + ], + "type": "text", + "content": " in appendix:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 102, + 598, + 295, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 598, + 295, + 632 + ], + "spans": [ + { + "bbox": [ + 102, + 598, + 295, + 632 + ], + "type": "interline_equation", + "content": "\\mathbb {E} \\left[ \\frac {\\partial \\hat {C}}{\\partial \\alpha_ {m}} \\right] = \\frac {(c _ {0} - c _ {\\mathrm {b g}}) T _ {\\mathrm {s a t u r a t i o n}}}{1 - \\mathbb {E} [ o _ {m} ] \\cdot \\mathbb {E} [ G _ {m} ^ {\\mathrm {2 D}} ]}, \\tag {6}", + "image_path": "4fc16a32e64fb735969629d34bd6098c7ab0f716ac69fdb397b8f13734fd6657.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 637, + 295, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 637, + 295, + 660 + ], + "spans": [ + { + "bbox": [ + 55, + 637, + 295, + 660 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 637, + 295, + 660 + ], + "type": "inline_equation", + "content": "c_{0}" + }, + { + "bbox": [ + 55, + 637, + 295, + 660 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 637, + 295, + 660 + ], + "type": "inline_equation", + "content": "\\alpha_0" + }, + { + "bbox": [ + 55, + 637, + 295, + 660 + ], + "type": "text", + "content": " denote the expected color and opacity respectively, and " + }, + { + "bbox": [ + 55, + 637, + 295, + 660 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{saturation}}" + }, + { + "bbox": [ + 55, + 637, + 295, + 660 + ], + "type": "text", + "content": " controls " + }, + { + "bbox": [ + 55, + 637, + 295, + 660 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 637, + 295, + 660 + ], + "type": "text", + "content": " saturation [12]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 661, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 661, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 661, + 296, + 714 + ], + "type": "text", + "content": "With fixed " + }, + { + "bbox": [ + 55, + 661, + 296, + 714 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 661, + 296, + 714 + ], + "type": "text", + "content": ", Eqs. (4) to (6) demonstrate that elevated opacity distributions yield greater gradient magnitudes in " + }, + { + "bbox": [ + 55, + 661, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\left[\\frac{\\partial L}{\\partial x_{\\mathrm{m}}}, \\frac{\\partial L}{\\partial y_{\\mathrm{m}}}\\right]" + }, + { + "bbox": [ + 55, + 661, + 296, + 714 + ], + "type": "text", + "content": ", thereby making such Gaussians statistically prioritized for densification." + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 433, + 159 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 433, + 159 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 433, + 159 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 433, + 159 + ], + "type": "image", + "image_path": "15f31b0e10f5a87f4c7b7a46d85325f1188a34b27cfb393f1b2978f628c81a24.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 436, + 72, + 553, + 160 + ], + "blocks": [ + { + "bbox": [ + 436, + 72, + 553, + 160 + ], + "lines": [ + { + "bbox": [ + 436, + 72, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 436, + 72, + 553, + 160 + ], + "type": "image", + "image_path": "3a1d709fe8a16b95741037f9a51203efff6972535d4b59cc56e962dcccc3cc7e.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 317, + 173, + 433, + 271 + ], + "blocks": [ + { + "bbox": [ + 390, + 162, + 479, + 172 + ], + "lines": [ + { + "bbox": [ + 390, + 162, + 479, + 172 + ], + "spans": [ + { + "bbox": [ + 390, + 162, + 479, + 172 + ], + "type": "text", + "content": "The distribution of opacity." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 173, + 433, + 271 + ], + "lines": [ + { + "bbox": [ + 317, + 173, + 433, + 271 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 433, + 271 + ], + "type": "image", + "image_path": "40bbbeeea462a6d92ee1506a0560f5095bae56ca1cc91925b76b194a8f2d31be.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 390, + 274, + 479, + 283 + ], + "lines": [ + { + "bbox": [ + 390, + 274, + 479, + 283 + ], + "spans": [ + { + "bbox": [ + 390, + 274, + 479, + 283 + ], + "type": "text", + "content": "The distribution of volume." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 436, + 173, + 552, + 270 + ], + "blocks": [ + { + "bbox": [ + 436, + 173, + 552, + 270 + ], + "lines": [ + { + "bbox": [ + 436, + 173, + 552, + 270 + ], + "spans": [ + { + "bbox": [ + 436, + 173, + 552, + 270 + ], + "type": "image", + "image_path": "1844f67d302ca153d70be6c8c8f4d8ed4acc664250e623317b4120d972c51e07.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 293, + 555, + 383 + ], + "lines": [ + { + "bbox": [ + 313, + 293, + 555, + 383 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 555, + 383 + ], + "type": "text", + "content": "Figure 5. The distribution of Gaussian attributes. The distribution of all Gaussian attributes (left) and those contributing specifically to densification (right) in the \"Bicycle\"[1] with 3DGS [12]. Top row: While the opacities are primarily concentrated around 0 and 1, the Gaussians that contribute to densification are predominantly situated around 1. Bottom row: The distribution of Volume. As densification progresses, Gaussians with lower volume become increasingly involved in the densification process." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "type": "text", + "content": "- Experimental Validation: Our experiments track Gaussian primitives contributing to densification, focusing on opacity and volumes values. Statistical analysis of attribute distributions during densification steps (Fig. 5) reveals that Gaussians with high opacity and small volume are the primary sources implementing Gaussian densification." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 464, + 554, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 554, + 571 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 554, + 571 + ], + "type": "text", + "content": "Notably, insufficient high-opacity Gaussians increase the photometric loss, exacerbating both under-reconstruction and over-reconstruction [12], thereby forces redundant densification. While small volumes can be employed as a compression metric [5] in the model compression of 3D Gaussian Splatting, they simultaneously drive reasonable densification. Take into account all factors, it is advisable to cache Gaussians with low-opacity for better gaussian densification." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 578, + 514, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 514, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 514, + 590 + ], + "type": "text", + "content": "3.2.2. Sampling Strategy for Faster Rendering" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 592, + 554, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 592, + 554, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 592, + 554, + 629 + ], + "type": "text", + "content": "During viewpoint-specific image synthesis in 3DGS, the framework implements " + }, + { + "bbox": [ + 313, + 592, + 554, + 629 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 592, + 554, + 629 + ], + "type": "text", + "content": "-blending [21] through recursive application of the compositing equation on a per-pixel." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 355, + 636, + 555, + 672 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 636, + 555, + 672 + ], + "spans": [ + { + "bbox": [ + 355, + 636, + 555, + 672 + ], + "type": "interline_equation", + "content": "C _ {p} = \\sum_ {i = 1} ^ {N} c _ {i} \\alpha_ {i} T _ {i}, \\quad T _ {i} = \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\qquad (7)", + "image_path": "c9c53ac484a6b57cf09e3d00edabbc25b1145173861351b123fbbc139a0a3777.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 326, + 679, + 555, + 718 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 679, + 555, + 718 + ], + "spans": [ + { + "bbox": [ + 326, + 679, + 555, + 718 + ], + "type": "interline_equation", + "content": "N = \\min \\left\\{i \\in \\mathbb {N} ^ {+}, \\left| \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}) \\leq T _ {\\text {s a t u r a t i o n}} \\right. \\right\\}, \\tag {8}", + "image_path": "ab9a2b55e80f0abbb7336e0f9b0a3aafbbe46a4caf7aefd3682438f155f1ed97.jpg" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27971" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": "As established in 3DGS [12], the determination of " + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": " hinges on achieving opacity saturation within a pixel's accumulated " + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": " value (Transmittance " + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "inline_equation", + "content": "T_{i} < T_{\\mathrm{saturation}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": "). This critical threshold implementation introduces a fundamental constraint on rendering efficiency: during forward rendering operations, the sequential traversal of Gaussian primitives inherently prevents parallel computation of " + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": " values. Consequently, " + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": " emerges as the principal determinant of rendering speed, where any reduction on " + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": " directly translates to accelerated renderings." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 192, + 295, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 192, + 295, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 295, + 216 + ], + "type": "text", + "content": "Based on the preceding analysis, we formulate the second fundamental property of the 3DGS framework:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 217, + 295, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 295, + 251 + ], + "type": "text", + "content": "- Proposition 2. Opacity-based Efficient Rendering Acceleration: Gaussian primitives with higher opacity enable faster rendering through faster achieving " + }, + { + "bbox": [ + 55, + 217, + 295, + 251 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 217, + 295, + 251 + ], + "type": "text", + "content": " saturation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 251, + 295, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 295, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 295, + 289 + ], + "type": "text", + "content": "Proof. Building upon the premise of mutually independent " + }, + { + "bbox": [ + 55, + 251, + 295, + 289 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 55, + 251, + 295, + 289 + ], + "type": "text", + "content": " values between Gaussians, we posit that the mathematical expectation " + }, + { + "bbox": [ + 55, + 251, + 295, + 289 + ], + "type": "inline_equation", + "content": "\\mathbb{E}[T_N] \\approx T_{\\mathrm{saturation}}" + }, + { + "bbox": [ + 55, + 251, + 295, + 289 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 296, + 295, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 296, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 67, + 296, + 295, + 312 + ], + "type": "interline_equation", + "content": "\\mathbb {E} [ T _ {N} ] = \\left(1 - \\mathbb {E} [ \\alpha_ {i} ]\\right) ^ {N} = \\left(1 - \\mathbb {E} [ o _ {i} ] \\cdot \\mathbb {E} [ G _ {i} ^ {\\mathrm {2 D}} ]\\right) ^ {N}, \\tag {9}", + "image_path": "abcfcc8cfa1da860dfd83017a2aea297763e10bfa9c7b94efbfeda37d69d4965.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 319, + 295, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 295, + 355 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 295, + 355 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 319, + 295, + 355 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 55, + 319, + 295, + 355 + ], + "type": "text", + "content": " combines its intrinsic opacity " + }, + { + "bbox": [ + 55, + 319, + 295, + 355 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 55, + 319, + 295, + 355 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 319, + 295, + 355 + ], + "type": "inline_equation", + "content": "G_{i}^{2\\mathrm{D}}" + }, + { + "bbox": [ + 55, + 319, + 295, + 355 + ], + "type": "text", + "content": " (dependent on view projection and covariance, independent of " + }, + { + "bbox": [ + 55, + 319, + 295, + 355 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 55, + 319, + 295, + 355 + ], + "type": "text", + "content": "). We decouple the expected value of " + }, + { + "bbox": [ + 55, + 319, + 295, + 355 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 55, + 319, + 295, + 355 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 356, + 296, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 356, + 296, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 356, + 296, + 440 + ], + "type": "text", + "content": "- Experimental Validation: We generated a 3DGS model with Gaussian " + }, + { + "bbox": [ + 55, + 356, + 296, + 440 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 55, + 356, + 296, + 440 + ], + "type": "text", + "content": " values sampled from " + }, + { + "bbox": [ + 55, + 356, + 296, + 440 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mu_o,0.1)" + }, + { + "bbox": [ + 55, + 356, + 296, + 440 + ], + "type": "text", + "content": ", fixing the other parameters. The rendering time as a function of " + }, + { + "bbox": [ + 55, + 356, + 296, + 440 + ], + "type": "inline_equation", + "content": "\\mu_o" + }, + { + "bbox": [ + 55, + 356, + 296, + 440 + ], + "type": "text", + "content": " is shown in Fig. 6. The results reveals that Gaussian primitives with higher-opacity induce faster transmittance saturation, thereby reducing the blending steps " + }, + { + "bbox": [ + 55, + 356, + 296, + 440 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 356, + 296, + 440 + ], + "type": "text", + "content": " and rendering computational load." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 448, + 212, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 448, + 212, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 448, + 212, + 460 + ], + "type": "text", + "content": "3.3. Analysis and Design Insights" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 465, + 295, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 465, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 55, + 465, + 295, + 536 + ], + "type": "text", + "content": "Through a comprehensive testing program and a detailed examination of the densification and rendering formulas in 3DGS, we have identified the acceleration effect of the opacity attribute in the 3DGS training process. Based on these insights, we propose an Opacity-based Prioritised Sampling (OPS) strategy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 537, + 295, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 537, + 295, + 561 + ], + "spans": [ + { + "bbox": [ + 55, + 537, + 295, + 561 + ], + "type": "text", + "content": "The Opacity-based sampling probability " + }, + { + "bbox": [ + 55, + 537, + 295, + 561 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 55, + 537, + 295, + 561 + ], + "type": "text", + "content": " for selecting an Under-training Gaussian " + }, + { + "bbox": [ + 55, + 537, + 295, + 561 + ], + "type": "inline_equation", + "content": "G_i" + }, + { + "bbox": [ + 55, + 537, + 295, + 561 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 143, + 564, + 294, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 564, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 143, + 564, + 294, + 590 + ], + "type": "interline_equation", + "content": "p _ {i} = \\frac {\\alpha_ {i}}{\\sum_ {i = 1} ^ {N} \\alpha_ {i}} \\tag {10}", + "image_path": "3ff09d7b2dadc02ccc76e8a3d841efc4820224a3f12e96662a0bf06fdf9ac2e7.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 598, + 295, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 598, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 598, + 295, + 623 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 598, + 295, + 623 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 55, + 598, + 295, + 623 + ], + "type": "text", + "content": " represents the opacity of Gaussian primitive " + }, + { + "bbox": [ + 55, + 598, + 295, + 623 + ], + "type": "inline_equation", + "content": "G_{i}" + }, + { + "bbox": [ + 55, + 598, + 295, + 623 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 598, + 295, + 623 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 598, + 295, + 623 + ], + "type": "text", + "content": " is the total number of Gaussian primitives." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 633, + 136, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 633, + 136, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 633, + 136, + 647 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "content": "We undertake an experimental investigation of Group Training within the context of 3D Gaussian Splatting [12] and Mip-Splatting [33]. The experimental settings are detailed in Sec. 4.1. Subsequently, we present the experimental results in Sec. 4.2. Cross-baselines applicability validation" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 318, + 74, + 550, + 148 + ], + "blocks": [ + { + "bbox": [ + 318, + 74, + 550, + 148 + ], + "lines": [ + { + "bbox": [ + 318, + 74, + 550, + 148 + ], + "spans": [ + { + "bbox": [ + 318, + 74, + 550, + 148 + ], + "type": "image", + "image_path": "ca078b0976ec7ffb4514a3c52ced74f8fcfb29b6ddc1f5665142e4276de50df0.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 151, + 553, + 194 + ], + "lines": [ + { + "bbox": [ + 313, + 151, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 313, + 151, + 553, + 194 + ], + "type": "text", + "content": "Figure 6. High-opacity Gaussians accelerate 3DGS rendering. Left: The rendering time decreases by " + }, + { + "bbox": [ + 313, + 151, + 553, + 194 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 313, + 151, + 553, + 194 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 151, + 553, + 194 + ], + "type": "inline_equation", + "content": "\\mu_{o}" + }, + { + "bbox": [ + 313, + 151, + 553, + 194 + ], + "type": "text", + "content": " increases. Right: The mechanisms that highers-opacity Gaussians achieve faster " + }, + { + "bbox": [ + 313, + 151, + 553, + 194 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 151, + 553, + 194 + ], + "type": "text", + "content": " saturation with less number of Gaussians." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 203, + 553, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 203, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 203, + 553, + 252 + ], + "type": "text", + "content": "across diverse scenarios is further presented in Sec. 4.3. The ablation experiments and their results are comprehensively discussed in Sec. 4.4. Finally, we analyse the results of different sampling strategies in Sec. 4.5." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 260, + 441, + 273 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 260, + 441, + 273 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 441, + 273 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 277, + 553, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 553, + 326 + ], + "type": "text", + "content": "Implementation Details. The Group Training method was integrated as a plug-in to facilitate its utilisation in conjunction with 3DGS. We implemented our method within both 3D Gaussian Splatting [12] and Mip-Splatting [33]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 327, + 554, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 554, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 554, + 494 + ], + "type": "text", + "content": "Group Training was implemented with a default grouping interval of 500 iterations throughout the training process. We employed both Random Sampling (RS) and Opacity-Based Prioritized Sampling (OPS) strategies with an Under-Training Ratio (UTR) of 0.6. To ensure computational stability, considering the importance of initial Gaussians [4], Group Training was activated at the 500th iteration after densification begins. The global densification " + }, + { + "bbox": [ + 313, + 327, + 554, + 494 + ], + "type": "inline_equation", + "content": "(I_{\\mathrm{D}})" + }, + { + "bbox": [ + 313, + 327, + 554, + 494 + ], + "type": "text", + "content": " and optimization " + }, + { + "bbox": [ + 313, + 327, + 554, + 494 + ], + "type": "inline_equation", + "content": "(I_{\\mathrm{O}})" + }, + { + "bbox": [ + 313, + 327, + 554, + 494 + ], + "type": "text", + "content": " were scheduled at " + }, + { + "bbox": [ + 313, + 327, + 554, + 494 + ], + "type": "inline_equation", + "content": "14.5\\mathrm{K}" + }, + { + "bbox": [ + 313, + 327, + 554, + 494 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 327, + 554, + 494 + ], + "type": "inline_equation", + "content": "29\\mathrm{K}" + }, + { + "bbox": [ + 313, + 327, + 554, + 494 + ], + "type": "text", + "content": " iterations, respectively. To manage the increased GPU memory consumption during group merging, we fixed the SH coefficients [25, 28] at iterations " + }, + { + "bbox": [ + 313, + 327, + 554, + 494 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{D}}" + }, + { + "bbox": [ + 313, + 327, + 554, + 494 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 327, + 554, + 494 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{O}}" + }, + { + "bbox": [ + 313, + 327, + 554, + 494 + ], + "type": "text", + "content": ". We also conducted experiments applying Group Training solely during the densification phase." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 494, + 554, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 494, + 554, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 494, + 554, + 577 + ], + "type": "text", + "content": "Dataset and Metrics. The efficacy of our method is evaluated on the Mip-NeRF360 [1], Tanks & Temples [14], Deep Blending [9], and NeRF-Synthetic [21] using a single RTX 3090. Subsequently, the model's reconstruction performance is evaluated using PSNR, SSIM, and LPIPS. Furthermore, the training time, peak memory usage, and model size for each scene are recorded." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 586, + 433, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 433, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 433, + 598 + ], + "type": "text", + "content": "4.2. Experiments Results" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 604, + 553, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 604, + 553, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 604, + 553, + 628 + ], + "type": "text", + "content": "This section presents the experimental results of integrating the Group-Training method into the 3D Gaussian Splat" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 636, + 553, + 655 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 636, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 315, + 636, + 553, + 655 + ], + "type": "text", + "content": "https://github.com/graphdeco-inria/gaussian-splatting" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 656, + 553, + 675 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 656, + 553, + 675 + ], + "spans": [ + { + "bbox": [ + 315, + 656, + 553, + 675 + ], + "type": "text", + "content": "https://github.com/autonomousvision/mip-splatting" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 329, + 675, + 494, + 684 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 675, + 494, + 684 + ], + "spans": [ + { + "bbox": [ + 329, + 675, + 494, + 684 + ], + "type": "text", + "content": "https://jonbarron.info/mipnerf360/" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 329, + 685, + 484, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 685, + 484, + 693 + ], + "spans": [ + { + "bbox": [ + 329, + 685, + 484, + 693 + ], + "type": "text", + "content": "https://www.tanksandtemples.org/" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 329, + 694, + 546, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 694, + 546, + 703 + ], + "spans": [ + { + "bbox": [ + 329, + 694, + 546, + 703 + ], + "type": "text", + "content": "http://visual.cs.ucl.ac.uk/pubs/deepblending/" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 329, + 704, + 495, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 704, + 495, + 712 + ], + "spans": [ + { + "bbox": [ + 329, + 704, + 495, + 712 + ], + "type": "text", + "content": "https://www.matthewancik.com/nerf" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27972" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 553, + 163 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 553, + 163 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 553, + 163 + ], + "type": "table", + "html": "
Grouping IterationsMip-NeRF360 [1]Tanks & Temples [14]Deep Blending [9]Blender [21]
PM ↓Size ↓Time ↓PM ↓Size ↓Time ↓PM ↓Size ↓Time ↓PM ↓Size ↓Time ↓
3D Gaussian Splatting*-8.6979226.724.5943415.007.7266623.902.78696.06
Group Training w/ RS0~15K9.5190726.625.0849614.607.6464421.852.76635.66
0~30K9.4890222.585.0449512.257.6364318.452.76635.26
Group Training w/ OPS0~15K8.5767822.534.5438412.806.8148719.252.65435.01
0~30K8.5767919.564.5538410.956.8048716.852.64434.79
", + "image_path": "60f6c6a18be386967e2e63235d98699ac776b368b335958ae9bb7b3770c876be.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 225, + 553, + 302 + ], + "blocks": [ + { + "bbox": [ + 55, + 171, + 555, + 216 + ], + "lines": [ + { + "bbox": [ + 55, + 171, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 555, + 216 + ], + "type": "text", + "content": "Table 1. Comparison of reconstruction efficiency for 3DGS [12]. Group Training significantly enhances reconstruction speed across all four datasets, with the OPS strategy achieving the highest acceleration. This effect is particularly notable in complex scenes. While Group Training with RS produces a more redundant model, the OPS strategy yields more compact models and reduced GPU peak memory usage. * indicates that the model was retrained. PM stands for GPU peak memory allocation, with Size in MB and Time in minutes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 225, + 553, + 302 + ], + "lines": [ + { + "bbox": [ + 57, + 225, + 553, + 302 + ], + "spans": [ + { + "bbox": [ + 57, + 225, + 553, + 302 + ], + "type": "table", + "html": "
Grouping IterationsMip-NeRF360 [1]Tanks&Temples [14]Deep Blending [9]Blender [21]
PSNR ↑SSIM ↑LPIPS ↓Time ↓PSNR ↑SSIM ↑LPIPS ↓Time ↓PSNR ↑SSIM ↑LPIPS ↓Time ↓PSNR ↑SSIM ↑LPIPS ↓Time ↓
3D GS [12]-27.2050.8150.2143-23.1420.8410.183-29.4050.9030.2425-33.330.9690.030-
3D GS*-27.4450.8160.215526.723.6970.8490.176415.029.5860.9040.243723.933.7720.9700.03066.1
Group Training0~15K27.6210.8220.207426.623.7730.8510.170314.629.1850.9010.247321.933.9590.9710.02845.7
w/ RS0~30K27.5370.8210.215722.623.7030.8480.182312.329.4170.9070.248118.533.8770.9710.02955.3
Group Training0~15K27.5820.8200.210322.523.8420.8510.172612.829.7130.9060.243519.333.8890.9710.02925.0
w/ OPS0~30K27.5640.8200.213319.623.8530.8500.176411.029.7460.9070.245016.933.8080.9700.02994.8
", + "image_path": "7cdadf300b636d5ef84e36699f727be72c49b1173550c6afce0dcf6b559a742c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 310, + 555, + 344 + ], + "lines": [ + { + "bbox": [ + 54, + 310, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 54, + 310, + 555, + 344 + ], + "type": "text", + "content": "Table 2. Comparison of reconstruction quality for 3DGS [12]. Group Training simultaneously accelerates the 3DGS scene reconstruction process and improves the accuracy of 3D reconstructions. Notably, Group Training with OPS achieves a notable improvement in reconstruction quality while maintaining the highest reconstruction speed." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 359, + 294, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 359, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 54, + 359, + 294, + 419 + ], + "type": "text", + "content": "ting [12] and Mip-Splatting [33]. The test outcomes demonstrate that the Group Training with OPS offers a significant reduction in training time compared to the baseline in all scenes, while concurrently enhancing the rendering quality of the novel view, as presented in Fig. 7." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 425, + 295, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 425, + 295, + 473 + ], + "spans": [ + { + "bbox": [ + 54, + 425, + 295, + 473 + ], + "type": "text", + "content": "3D Gaussian Splatting. We first tested the effectiveness of Group Training based on 3DGS. The reconstruction efficiency for all scenes is summarized in Tab. 1, and Tab. 2 presents the NVS rendering quality across all scenes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 474, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 688 + ], + "type": "text", + "content": "Across these scenes, Group Training consistently resulted in faster reconstruction, and this speedup was more pronounced in complex scenes. However, Group Training with RS could lead to an overabundance of redundant Gaussian primitives in some reconstruction tasks, which complicates subsequent operations, such as model compression [24, 25, 28]. Additionally, GPU memory usage significantly increased due to this redundancy [17], as shown in Tab. 1. In comparison to Group Training with RS, Group Training with OPS demonstrated a more pronounced improvement in reconstruction speed. For the reconstruction of all three complex scenes, the average speed increase was about " + }, + { + "bbox": [ + 55, + 474, + 295, + 688 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 55, + 474, + 295, + 688 + ], + "type": "text", + "content": ". Moreover, OPS greatly reduced the generation of redundant Gaussian primitives, leading to a smaller reconstruction model size than the original 3DGS reconstruction model in all tasks. Across all reconstruction tasks, the model size was reduced by between " + }, + { + "bbox": [ + 55, + 474, + 295, + 688 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 474, + 295, + 688 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 474, + 295, + 688 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 55, + 474, + 295, + 688 + ], + "type": "text", + "content": ", resulting in a more compact model overall." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "Mip-Splatting. We tested Mip-Splatting with Group Training, another significant variant of the Gaussian Splat" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 359, + 555, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 359, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 359, + 555, + 491 + ], + "type": "text", + "content": "ting model. The experimental results are presented in Tab. 3 and Tab. 4. Overall, the performance of Group Training with Mip-Splatting was similar to 3DGS. Group Training with RS consistently improved both reconstruction quality and speed for Mip-Splatting but produced the highest number of Gaussian primitives during the densification phase. Group Training with OPS achieved the fastest reconstruction speed, and its reconstruction quality was nearly optimal. The OPS reduced the proportion of densification and resulted in a smaller model size that was either the smallest among the tested methods or comparable to the baseline." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 502, + 553, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 553, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 553, + 526 + ], + "type": "text", + "content": "Counterintuitive Findings. Our experiments reveal two notable findings that challenge conventional understanding:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 528, + 553, + 634 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 314, + 528, + 553, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 528, + 553, + 575 + ], + "spans": [ + { + "bbox": [ + 314, + 528, + 553, + 575 + ], + "type": "text", + "content": "1. Despite producing larger model sizes, Group Training with RS consistently improves reconstruction speed, indicating that training dynamics rather than model size dominates reconstruction efficiency." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 575, + 553, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 575, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 313, + 575, + 553, + 634 + ], + "type": "text", + "content": "2. Group Training with OPS simultaneously achieves better reconstruction quality with reduced model size. This demonstrates that superior performance stem from its fundamental impact on the optimization process rather than from model size." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 647, + 473, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 473, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 473, + 659 + ], + "type": "text", + "content": "4.3. Methodological Applicability" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": "Group-Training framework fundamentally addresses accelerated training and rendering in 3DGS through its dataset-agnostic improvement. Designed as a pluggable training component, the method demonstrates dual efficacy in both" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27973" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 553, + 157 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 553, + 157 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 553, + 157 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 553, + 157 + ], + "type": "table", + "html": "
Grouping IterationsTanks & Temples [14]Deep Blending [9]Blender [21]
PSNR ↑SSIM ↑LPIPS ↓Time ↓PSNR ↑SSIM ↑LPIPS ↓Time ↓PSNR ↑SSIM ↑LPIPS ↓Time ↓
Mip-Splatting*-23.7490.8600.156223.029.3580.9030.239035.133.9950.9700.02968.5
Group Training w/ RS0~15K23.9530.8630.150424.628.9290.9000.243131.734.2030.9710.02789.3
0~30K24.1460.8620.159820.429.3920.9080.242026.534.1230.9720.02858.3
Group Training w/ OPS0~15K23.9580.8630.151921.729.6650.9070.236428.134.1320.9710.02858.2
0~30K24.1560.8630.155918.229.7880.9080.238424.034.1100.9710.02867.5
", + "image_path": "8cbd95e05e89d2c5f2f25c87cae54220153952039ce0b24adbb69b5a24bfde4d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 209, + 553, + 283 + ], + "blocks": [ + { + "bbox": [ + 55, + 164, + 555, + 198 + ], + "lines": [ + { + "bbox": [ + 55, + 164, + 555, + 198 + ], + "spans": [ + { + "bbox": [ + 55, + 164, + 555, + 198 + ], + "type": "text", + "content": "Table 3. Comparison of reconstruction quality for Mip-Splatting [33]. Group Training achieves simultaneous improvements in both speed and reconstruction quality for Mip-Splatting across all tasks. Specifically, Group Training with OPS attains the fastest reconstruction speed across all three datasets and delivers the highest reconstruction quality in the Tanks & Temples and Deep Blending dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 209, + 553, + 283 + ], + "lines": [ + { + "bbox": [ + 57, + 209, + 553, + 283 + ], + "spans": [ + { + "bbox": [ + 57, + 209, + 553, + 283 + ], + "type": "table", + "html": "
Grouping IterationsOutdoorIndoorAverage
PSNR ↑SSIM ↑LPIPS ↓Size ↓Time ↓PSNR ↑SSIM ↑LPIPS ↓Size ↓Time ↓PSNR ↑SSIM ↑LPIPS ↓Size ↓Time ↓
Mip-Splatting*-24.8140.7480.1995143043.331.2650.9290.173844833.327.6810.8280.188199438.8
Group Training0~15K24.9690.7570.1897201745.931.2660.9310.169755634.927.7680.8340.1808137741.0
w/ RS0~30K25.2160.7630.1969201736.831.1640.9290.175755630.527.8600.8370.1875137734.0
Group Training0~15K24.9810.7540.1940157140.731.4170.9310.172537529.127.8410.8320.1844103935.5
w/ OPS0~30K24.9840.7540.1967157132.931.4070.9300.173937526.427.8390.8320.1865103930.0
", + "image_path": "ff90e21113ffdf76b29872915933befa834534f90c4dc58114fb6061b226b1d0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 57, + 348, + 294, + 414 + ], + "blocks": [ + { + "bbox": [ + 55, + 291, + 555, + 335 + ], + "lines": [ + { + "bbox": [ + 55, + 291, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 291, + 555, + 335 + ], + "type": "text", + "content": "Table 4. Quantitative evaluation on the Mip-NeRF360 [1] reconstructed by Mip-Splatting [33]. For the reconstruction of outdoor scenes, Grouping Training with RS yields a larger model size, resulting in the best reconstruction quality. Grouping Training with OPS achieves the fastest reconstruction speed but provides sub-optimal reconstruction quality. For indoor scenes, Grouping Training with OPS continues to offer both the fastest reconstruction speed and the highest reconstruction quality." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 348, + 294, + 414 + ], + "lines": [ + { + "bbox": [ + 57, + 348, + 294, + 414 + ], + "spans": [ + { + "bbox": [ + 57, + 348, + 294, + 414 + ], + "type": "table", + "html": "
Cyclic ResampleGlobal DenseifyGlobal OptimizeTanks &Temples
PSNR ↑SSIM ↑LPIPS ↓Size ↓Time ↓
3D GS*---23.6970.8490.176443415.0
Group Training✓23.8660.8500.174929211.8
✓✓23.7690.8490.176523111.0
✓✓23.8350.8500.175448511.8
✓✓✓23.8530.8500.176438411.0
", + "image_path": "2e35644c3e76d07ec5d4fa5909c413a698fbbbebcecabbd70640f1dd951b9049.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 422, + 295, + 489 + ], + "lines": [ + { + "bbox": [ + 55, + 422, + 295, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 295, + 489 + ], + "type": "text", + "content": "Table 5. Ablation Study of Group Training with OPS on Tanks & Temples [14]. Cyclic Resampling provides the most substantial acceleration for scene reconstruction, while Global Densification further improves efficiency by reducing model size. With Global Densification enabled, Global Optimization enhances the reconstruction quality of the model." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 502, + 295, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 295, + 527 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 295, + 527 + ], + "type": "text", + "content": "3DGS acceleration model [6] and compression model [5], which is detailed in appendix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 538, + 156, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 538, + 156, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 538, + 156, + 550 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 556, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 295, + 604 + ], + "type": "text", + "content": "We conducted an ablation study on the Tanks & Temples dataset [14] to evaluate the impact of three key components of the Group Training method. The detailed comparison results are provided in Tab. 5." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "The results indicate that cyclic resampling contributes the most to improving training speed. Specifically, during iterations " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{D}} \\sim 15\\mathrm{K}" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": ", the Global Densification process adaptively performs cyclic pruning of low-opacity Gaussians within 3DGS, reducing redundant Gaussians, decreasing the model size, and thereby enhancing training efficiency. When Global Densification is enabled, the 3DGS model removes most redundant Gaussians during global densification, resulting in fewer redundant Gaussians in the Caching" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 350, + 555, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 555, + 387 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 555, + 387 + ], + "type": "text", + "content": "Group during the Optimization phase. In this scenario, Global Optimization can make full use of all Gaussians, further improving reconstruction performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 394, + 387, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 394, + 387, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 387, + 406 + ], + "type": "text", + "content": "4.5. Discussion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 413, + 554, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 554, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 554, + 460 + ], + "type": "text", + "content": "Our experimental results demonstrate that enabling Group Training and utilizing OPS improved both reconstruction speed and quality. To better clarify these improvements, two key questions arise from these findings:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 461, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 555, + 700 + ], + "type": "text", + "content": "- **Why faster?** Group Training significantly accelerates reconstruction primarily by reducing the number of Undertraining Gaussian primitives, thereby decreasing the rendering burden on the model. This speedup is observed even when the final model size is larger than the baseline. In alignment with our earlier analysis in Sec. 3.2, the OPS also contributes notably to this speedup effect. During the densification phase, Group Training with OPS preserves high-opacity Gaussians within the Under-Training Group, without which could lead to instances of both under-reconstruction and over-reconstruction [12, 32]. These issues can lead to redundant densification. As a result, models reconstructed with Group Training and OPS contain fewer redundant Gaussians, leading to smaller model sizes and faster reconstruction overall. Additionally, OPS prioritizes sampling high-opacity Gaussians, which rapidly drives " + }, + { + "bbox": [ + 313, + 461, + 555, + 700 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 461, + 555, + 700 + ], + "type": "text", + "content": " saturation, which also discussed in Sec. 3.2. Focusing on these most impactful Gaussians reduces the number of Gaussians needed for splatting, further enhancing training and rendering efficiency." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "text", + "content": "- Why better? Group Training with the OPS prioritizes the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27974" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 70, + 541, + 436 + ], + "blocks": [ + { + "bbox": [ + 71, + 70, + 541, + 436 + ], + "lines": [ + { + "bbox": [ + 71, + 70, + 541, + 436 + ], + "spans": [ + { + "bbox": [ + 71, + 70, + 541, + 436 + ], + "type": "image", + "image_path": "7ac07db189326658c53b3f138a974651370dfa348f08dbbe94f47087001e601f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 444, + 555, + 490 + ], + "lines": [ + { + "bbox": [ + 55, + 444, + 555, + 490 + ], + "spans": [ + { + "bbox": [ + 55, + 444, + 555, + 490 + ], + "type": "text", + "content": "Figure 7. Visual Comparison of Grouping Training and Baseline on 3DGS [12] and Mip-Splatting [33]. The Grouping Training method accelerates reconstruction by approximately " + }, + { + "bbox": [ + 55, + 444, + 555, + 490 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 55, + 444, + 555, + 490 + ], + "type": "text", + "content": " while achieving more detailed and accurate scene rendering. Notable improvements can be observed in the picture frame in the \"Dr. Johnson\" scene and the bucket handle in the \"Garden\" scene, where finer details are captured. Furthermore, Grouping Training significantly reduces floaters in Mip-Splatting, enhancing overall reconstruction fidelity." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 505, + 296, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 505, + 296, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 505, + 296, + 636 + ], + "type": "text", + "content": "retention and optimization of high-opacity Gaussian primitives. This achieves comparable or superior fidelity to RS, while significantly reducing the model size. The Gaussian primitives with higher opacity inherently have a greater influence on the final rendered image [5, 19, 22, 28], so emphasizing their optimization leads to significant improvements in overall rendering quality. Besides, cyclic resampling temporarily excluding a randomly selected subset of Gaussian primitives, each remaining primitive is compelled to learn how to independently represent essential scene features leading to a more robust representation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": "In summary, by integrating these strategies, Group Training with OPS simultaneously enhances both reconstruction speed and rendering quality. By focusing on the most impactful primitives and encouraging independent feature representation, Group Training achieves efficient, high-quality scene reconstruction." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 504, + 388, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 388, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 388, + 516 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 519, + 555, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 519, + 555, + 711 + ], + "spans": [ + { + "bbox": [ + 313, + 519, + 555, + 711 + ], + "type": "text", + "content": "This paper presents Group Training, an effective strategy that addresses the computational challenges in 3D Gaussian Splitting by dynamically managing Gaussian primitives in organized groups. Through theoretical analysis and extensive experiments, we demonstrate that our method not only accelerates training by up to " + }, + { + "bbox": [ + 313, + 519, + 555, + 711 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 313, + 519, + 555, + 711 + ], + "type": "text", + "content": " but also improves rendering quality across diverse scenarios. Our opacity-based sampling strategy shows universal compatibility with existing 3DGS frameworks, achieving consistent improvements without compromising model complexity. The success of Group Training reveals the importance of dynamic primitive management in 3DGS optimization, suggesting promising directions for future research, such as adaptive grouping strategies and dynamic sampling mechanisms for more complex scenes. Our work provides valuable insights for future developments in efficient novel view synthesis." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27975" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "type": "text", + "content": "This work is supported by the National Key R&D Program of China (No.2022ZD0119000), Hunan Provincial Key R&D Program of China (No.2024JK2020 and 2024JK2021), Hunan Provincial Natural Science Foundation of China (No.2024JJ10027), Young Talents of Huxiang (No.Z202433000575), and Changsha Science Fund for Distinguished Young Scholars (kq2306002)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 186, + 115, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 115, + 198 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 115, + 198 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 205, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 205, + 296, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 205, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 61, + 205, + 296, + 239 + ], + "type": "text", + "content": "[1] Jonathan T. Barron. Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields. In CVPR, 2022. 1, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 240, + 296, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 240, + 296, + 273 + ], + "spans": [ + { + "bbox": [ + 61, + 240, + 296, + 273 + ], + "type": "text", + "content": "[2] Samuel Rota Bulò, Lorenzo Porzi, and Peter Kontschieder. Revising densification in gaussian splatting. arXiv preprint arXiv:2404.06109, 2024. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 274, + 295, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 274, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 61, + 274, + 295, + 297 + ], + "type": "text", + "content": "[3] Guikun Chen and Wenguan Wang. A survey on 3d gaussian splatting. arXiv preprint arXiv:2401.03890, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 297, + 295, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 297, + 295, + 330 + ], + "spans": [ + { + "bbox": [ + 62, + 297, + 295, + 330 + ], + "type": "text", + "content": "[4] Jaeyoung Chung, Jeongtaek Oh, and Kyoung Mu Lee. Depth-regularized optimization for 3d gaussian splatting in few-shot images. arXiv preprint arXiv:2311.13398, 2023. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 331, + 295, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 331, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 62, + 331, + 295, + 374 + ], + "type": "text", + "content": "[5] Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, and Zhangyang Wang. Lightgaussian: Unbounded 3d gaussian compression with " + }, + { + "bbox": [ + 62, + 331, + 295, + 374 + ], + "type": "inline_equation", + "content": "15\\mathrm{x}" + }, + { + "bbox": [ + 62, + 331, + 295, + 374 + ], + "type": "text", + "content": " reduction and " + }, + { + "bbox": [ + 62, + 331, + 295, + 374 + ], + "type": "inline_equation", + "content": "200+" + }, + { + "bbox": [ + 62, + 331, + 295, + 374 + ], + "type": "text", + "content": " fps, 2023. 3, 4, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 376, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 376, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 62, + 376, + 294, + 409 + ], + "type": "text", + "content": "[6] Guangchi Fang and Bing Wang. Mini-splatting: Representing scenes with a constrained number of gaussians, 2024. 3, 4, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 410, + 295, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 410, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 62, + 410, + 295, + 443 + ], + "type": "text", + "content": "[7] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 445, + 295, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 445, + 295, + 477 + ], + "spans": [ + { + "bbox": [ + 62, + 445, + 295, + 477 + ], + "type": "text", + "content": "[8] Sharath Girish, Kamal Gupta, and Abhinav Shrivastava. Eagles: Efficient accelerated 3d gaussians with lightweight encodings. arXiv preprint arXiv:2312.04564, 2023. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 478, + 295, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 478, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 62, + 478, + 295, + 521 + ], + "type": "text", + "content": "[9] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep Blending for Free-Viewpoint Image-Based Rendering. In SIGGRAPH, 2018. 1, 5, 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 523, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 523, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 57, + 523, + 295, + 555 + ], + "type": "text", + "content": "[10] Lukas Hollein, Aljaž Božić, Michael Zollhöfer, and Matthias Nießner. 3dgs-lm: Faster gaussian-splatting optimization with levenberg-marquardt. arXiV:2409.12892, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 557, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 557, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 57, + 557, + 295, + 612 + ], + "type": "text", + "content": "[11] Ying Jiang, Chang Yu, Tianyi Xie, Xuan Li, Yutao Feng, Huamin Wang, Minchen Li, Henry Lau, Feng Gao, Yin Yang, et al. Vr-gs: A physical dynamics-aware interactive gaussian splatting system in virtual reality. In ACM SIGGRAPH 2024 Conference Papers, pages 1-1, 2024. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 613, + 295, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 295, + 657 + ], + "type": "text", + "content": "[12] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "text", + "content": "[13] Shakiba Kheradmand, Daniel Rebain, Gopal Sharma, Weiwei Sun, Yang-Che Tseng, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. 3d gaussian splatting as markov chain monte carlo. Advances in Neural Information Processing Systems, 37:80965-80986, 2024. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[14] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and Temples: Benchmarking Large-Scale Scene Reconstruction. In SIGGRAPH, 2017. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 108, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 161 + ], + "type": "text", + "content": "[15] Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaussian feature splatting for real-time dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8508-8520, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 163, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 217 + ], + "type": "text", + "content": "[16] Yang Liu, Chuanchen Luo, Lue Fan, Naiyan Wang, Junran Peng, and Zhaoxiang Zhang. Citygaussian: Real-time high-quality large-scale scene rendering with gaussians. In European Conference on Computer Vision, pages 265-282. Springer, 2025. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 220, + 553, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 261 + ], + "type": "text", + "content": "[17] Jiahao Lu, Yifan Zhang, Qiuhong Shen, Xinchao Wang, and Shuicheng Yan. Poisson-splat: Computation cost attack on 3d gaussian splatting. arXiv preprint arXiv:2410.08190, 2024. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 264, + 555, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 555, + 319 + ], + "type": "text", + "content": "[18] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 321, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 321, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 317, + 321, + 553, + 387 + ], + "type": "text", + "content": "[19] Saswat Subhajyoti Mallick, Rahul Goel, Bernhard Kerbl, Markus Steinberger, Francisco Vicente Carrasco, and Fernando De La Torre. Taming 3dgs: High-quality radiance fields with limited resources. In SIGGRAPH Asia 2024 Conference Papers, New York, NY, USA, 2024. Association for Computing Machinery. 3, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "type": "text", + "content": "[20] Jiarui Meng, Hajie Li, Yanmin Wu, Qiankun Gao, Shuzhou Yang, Jian Zhang, and Siwei Ma. Mirror-3dgs: Incorporating mirror reflections into 3d gaussian splatting. arXiv preprint arXiv:2404.01168, 2024. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 433, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 553, + 476 + ], + "type": "text", + "content": "[21] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 478, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 520 + ], + "type": "text", + "content": "[22] Wieland Morgenstern, Florian Barthel, Anna Hilsmann, and Peter Eisert. Compact 3d scene representation via self-organizing gaussian grids. arXiv preprint arXiv:2312.13299, 2023.8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 523, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 566 + ], + "type": "text", + "content": "[23] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 568, + 553, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 568, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 317, + 568, + 553, + 611 + ], + "type": "text", + "content": "[24] KL Navaneet, Kossar Pourahmadi Meibodi, Soroush Abbasi Koohpayegani, and Hamed Pirsiavash. Compgs: Smaller and faster gaussian splatting with vector quantization. ECCV, 2024. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 613, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 646 + ], + "type": "text", + "content": "[25] Simon Niedermayr, Josef Stumpfegger, and Rüdiger Westermann. Compressed 3d gaussian splattering for accelerated novel view synthesis, 2023. 3, 5, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "text", + "content": "[26] Michael Niemeyer, Fabian Manhardt, Marie-Julie Rakoto-saona, Michael Oechsle, Daniel Duckworth, Rama Gosula, Keisuke Tateno, John Bates, Dominik Kaeser, and Federico Tombari. Radsplat: Radiance field-informed gaussian splatting for robust real-time rendering with " + }, + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "inline_equation", + "content": "900+" + }, + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "text", + "content": " fps. arXiv.org, 2024. 3, 4" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27976" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 577 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[27] Panagiotis Papantonakis, Georgios Kopanas, Bernhard Kerbl, Alexandre Lanvin, and George Drettakis. Reducing the memory footprint of 3d gaussian splatting. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 7(1):1-17, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "type": "text", + "content": "[28] Panagiotis Papantonakis, Georgios Kopanas, Bernhard Kerbl, Alexandre Lanvin, and George Drettakis. Reducing the memory footprint of 3d gaussian splatting. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 2024. 2, 5, 6, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "type": "text", + "content": "[29] Zhexi Peng, Tianjia Shao, Yong Liu, Jingke Zhou, Yin Yang, Jingdong Wang, and Kun Zhou. Rtg-slam: Real-time 3d reconstruction at scale using gaussian splatting. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 262 + ], + "type": "text", + "content": "[30] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In CVPR, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 263, + 294, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 294, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 294, + 306 + ], + "type": "text", + "content": "[31] Haosen Yang, Chenhao Zhang, Wenqing Wang, Marco Volino, Adrian Hilton, Li Zhang, and Xiatian Zhu. Localized gaussian point management. arXiv e-prints, pages arXiv-2406, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 308, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 294, + 342 + ], + "type": "text", + "content": "[32] Zongxin Ye, Wenyu Li, Sidun Liu, Peng Qiao, and Yong Dou. Absgs: Recovering fine details for 3d gaussian splatt ing, 2024. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 343, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 294, + 374 + ], + "type": "text", + "content": "[33] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In CVPR, 2024. 1, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 376, + 294, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 420 + ], + "type": "text", + "content": "[34] Hongjia Zhai, Xiyu Zhang, Boming Zhao, Hai Li, Yijia He, Zhaopeng Cui, Hujun Bao, and Guofeng Zhang. Splatloc: 3d gaussian splatting-based visual localization for augmented reality. arXiv preprint arXiv:2409.14067, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 421, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 421, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 421, + 294, + 464 + ], + "type": "text", + "content": "[35] Zhaoliang Zhang, Tianchen Song, Yongjae Lee, Li Yang, Cheng Peng, Rama Chellappa, and Deliang Fan. Lp-3dgs: Learning to prune 3d gaussian splatting. arXiv preprint arXiv:2405.18784, 2024. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 466, + 294, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 466, + 294, + 531 + ], + "spans": [ + { + "bbox": [ + 56, + 466, + 294, + 531 + ], + "type": "text", + "content": "[36] Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21634-21643, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 533, + 294, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 533, + 294, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 533, + 294, + 577 + ], + "type": "text", + "content": "[37] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. In European Conference on Computer Vision, pages 145-163. Springer, 2025. 1" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27977" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Feather the Throttle_ Revisiting Visual Token Pruning for Vision-Language Model Acceleration/eb2a11ee-50f9-4c43-8cc0-429a524795eb_content_list.json b/2025/Feather the Throttle_ Revisiting Visual Token Pruning for Vision-Language Model Acceleration/eb2a11ee-50f9-4c43-8cc0-429a524795eb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..934ff43bc58dc9635a41c3eedc90cfa7c95b3182 --- /dev/null +++ b/2025/Feather the Throttle_ Revisiting Visual Token Pruning for Vision-Language Model Acceleration/eb2a11ee-50f9-4c43-8cc0-429a524795eb_content_list.json @@ -0,0 +1,1419 @@ +[ + { + "type": "text", + "text": "Feather the Throttle: Revisiting Visual Token Pruning for Vision-Language Model Acceleration", + "text_level": 1, + "bbox": [ + 207, + 130, + 789, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mark Endo, Xiaohan Wang, Serena Yeung-Levy \nStanford University", + "bbox": [ + 305, + 203, + 691, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{markendo,xhanwang,syyeung}@stanford.edu", + "bbox": [ + 318, + 241, + 671, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent works on accelerating Vision-Language Models achieve strong performance across a variety of vision-language tasks despite highly compressing visual information. In this work, we examine the popular acceleration approach of early pruning of visual tokens inside the language model. Surprisingly, we find that while strong performance is maintained across many tasks, it exhibits drastically different behavior for a subset of vision-centric tasks such as localization. Upon further investigation, we uncover a core issue with the acceleration approach where most tokens towards the top of the image are pruned away. Yet, on many benchmarks aiming to evaluate vision-centric capabilities, strong performance persists with the flawed pruning strategy, highlighting these benchmarks' limited ability to assess fine-grained visual capabilities. Based on these findings, we propose FEATHER (Fast and Effective Acceleration with Ensemble cRicerca), a straightforward approach that resolves the discovered early-layer pruning issue and further enhances the preservation of relevant tokens via multistage pruning with early uniform sampling to ensure broad image coverage. With comparable computational savings, we find that FEATHER achieves more than $5 \\times$ performance improvement on the vision-centric localization benchmarks compared to the original acceleration approach.", + "bbox": [ + 86, + 325, + 483, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 722, + 220, + 738 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The exploration of Vision-Language Models (VLMs) is a critical area of computer vision and natural language processing research, centered on combining large language models (LLMs) with visual encoders to enable multimodal perception, reasoning, and understanding capabilities. While earlier works explored sophisticated schemes for conditioning language models with visual information [2, 22, 23], more recently the space has shifted to predominately using the simplistic approach of taking patch features from pre-trained visual encoders and projecting them", + "bbox": [ + 89, + 750, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4d43fa273c72025972152408e9df7d7b3aeef6ba8cfda68fe40c567bda3e264b.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 544, + 290, + 901, + 559 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d8fc2f3288d4e37f83c664b0544d201797876a8f1b938aefe429ed16c63a5d47.jpg", + "image_caption": [ + "(b)", + "Figure 1. (a) Although FastV prunes most visual tokens from the upper portion of the image, the approach still displays strong performance on a variety of evaluated vision-language tasks except for a small subset of vision-centric tasks like localization. (b) Based on our findings, we propose FEATHER (Fast and Effective Acceleration wiTH Ensemble cRicerca), a straightforward approach that resolves the existing issue of selecting bottom tokens, additionally maintains uniformly sampled tokens to ensure good coverage over the whole image, and prunes in two stages." + ], + "image_footnote": [], + "bbox": [ + 540, + 565, + 898, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to the input space of the language model with a light-weight adapter [4, 27, 29, 38]. Using image patches as tokens, however, comes with the drawback of being computationally inefficient. To achieve fine-grained resolution, the im", + "bbox": [ + 511, + 839, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22826", + "bbox": [ + 478, + 945, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "age is divided into many patches. This large number of patches significantly increases computational demands due to the quadratic complexity of the attention operation in Transformers. As a result, many recent works have focused on accelerating these methods by compressing visual information, demonstrating that heavy compression can still maintain strong performance across a wide variety of tasks [3, 7, 10, 31, 39]. For instance, FastV [10] prunes $50\\%$ of visual tokens after the shallow layers of the Language Model (LLM) while not compromising in performance across a range of image and video understanding tasks [10]. Another work reveals that for a fixed computational budget on visual reasoning tasks, optimal performance is achieved by using the largest LLM possible while sacrificing visual information, often reducing the visual token count to a single token [24]. With such compressed visual information, it is still unclear how these methods achieve high performance on tasks assessing vision capabilities such as visual reasoning and understanding, and whether there are more demanding visual tasks where these methods fail.", + "bbox": [ + 89, + 90, + 485, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The motivation of our study is to get a better understanding of the vision capabilities of accelerated VLMs given that they leverage highly compressed visual information, focusing specifically on FastV approach. Evaluating across a wide range of vision-language tasks, we find that while the approach maintains strong performance across many tasks, there is a substantial decrease in performance for TextVQA and a severe decrease for localization tasks. When analyzing the approach's poor performance, we uncover a fundamental issue with the pruning criteria in early layers where tokens towards the top of the image are disproportionally removed due to the long-term decay property of RoPE. Since this issue is not task-specific, we examine how performance is maintained across most evaluated tasks and find that even when tokens are entirely discarded using the flawed criteria, performance remains largely unchanged. This highlights how many benchmarks require minimal visual grounding. Our finding underscores a significant challenge in the field of multimodal learning, not only for measuring the effectiveness of VLM acceleration methods but also for benchmarking the visual capabilities of VLMs as a whole.", + "bbox": [ + 91, + 397, + 483, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given the discovered limitation of the studied approach, we next explore whether the attention-based criteria can be modified to enable robust visual token pruning, leading to our final method, FEATHER (Fast and Effective Acceleration with Ensemble Ricerca). Specifically, we propose a straightforward modification to the attention-based criteria by removing RoPE, demonstrating that this resolves the positional bias issue and greatly improves performance. Additionally, we incorporate uniform sampling in early token pruning to ensure adequate coverage of all image regions and apply more extensive pruning at a later layer, when the attention-based criteria better discerns im", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "important tokens. This strategy is analogous to how a race-car driver feathers the throttle by gradually pressing the accelerator at the beginning of a turn to maintain grip and then accelerating more aggressively once the car is past the apex. We show that FEATHER results in substantial performance gains compared to the original acceleration approach, improving localization performance more than $5 \\times 5$ with comparable computational savings. Strikingly, we find that our approach achieves this performance improvement while only retaining $3.3\\%$ of visual tokens for the second half of LLM layers. Overall, our work demonstrates that while visual compression can maintain strong performance even on challenging vision-centric tasks, its effectiveness depends on a well-designed strategy, which is currently difficult to assess due to many vision-language benchmarks not thoroughly evaluating vision capabilities.", + "bbox": [ + 511, + 90, + 906, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions include:", + "bbox": [ + 532, + 333, + 792, + 347 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Revealing inconsistent effect of early pruning: Our evaluations across 12 benchmarks show that early visual token pruning minimally affects performance on most tasks, but causes a substantial dropoff for TextVQA and a drastic decline for localization tasks.", + "- Analyzing poor performance: We uncover that the attention-based criteria for selecting important tokens exhibits a bias towards bottom image tokens in early layers due to the long-term decay property of RoPE.", + "- Identifying benchmark limitations in assessing token pruning: We find that performance remains strong even with suboptimal pruning criteria and removal of pruned tokens before transferring information in the LLM, demonstrating how many benchmarks fail to assess fine-grained visual capabilities.", + "- Improving attention-based criteria: With the straightforward modification of removing RoPE, we demonstrate that attention can be used as an effective criteria for selecting tokens and performance can be further improved with simple strategies like incorporating uniform sampling and pruning more extensively at a later layer, when the criteria better discerns important tokens." + ], + "bbox": [ + 535, + 349, + 903, + 696 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 709, + 656, + 724 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent efforts to accelerate VLMs can be broadly divided into two main categories: compressing visual information before it enters the LLM, and compressing visual information within the LLM itself. In the first category, ChatUniVi [17] dynamically merges visual tokens with similar semantic meanings. PruMerge [31], FasterVLM [44], and VisionZip [40] select important tokens based on ViT attention ([CLS] or average attention received). PruMerge then merges important tokens with unselected ones while VisionZip merges unselected tokens together. For the LLaVA-NeXT [28] approach of partitioning an image into", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22827", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sub-images where inefficiency is an even bigger problem, HiRED [3] selects tokens with top feature important on each sub-image with an allocated budget. Other methods argue that the input image alone does not include enough information to select important patches and thus use the textual input to recover visually meaningful tokens [11, 41]. Alternatively, some approaches reduce visual information within the ViT by merging or pruning tokens [6, 8, 30].", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the second category of approaches, where visual information is compressed within the LLM itself, LOOK-M reduces the multimodal KV cache size [36]. In our study, we focus on the popular FastV approach [10]. This work identifies that attention over image tokens is sparse in deeper layers of the LLM. Based on this observation, they propose to prune away unimportant vision tokens after the shallow layers of the LLM, achieving a $45\\%$ reduction in FLOPS with nearly no performance loss. Works since have proposed alterations to this setup such as adaptively determining the number of pruned tokens instead of using a fixed ratio [14] or pruning in multiple stages [39].", + "bbox": [ + 89, + 212, + 482, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Revising Visual Token Pruning", + "text_level": 1, + "bbox": [ + 89, + 406, + 375, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we aim to get a better understanding of the vision capabilities of the VLM acceleration approach of pruning visual tokens after shallow LLM layers. After outlining preliminaries (§3.1), we take a closer look at how the approach performs across a broad range of tasks. Upon inspection, we discover that while heavily pruning visual tokens after the shallow LLM layers has little effect on performance across a variety of tasks, this approach fails decisively on more vision-centric tasks, particularly localization (§3.2). Next, we examine why this method struggles with localization, uncovering that the poor performance is due to the ineffectiveness of the pruning criteria to select important tokens when applied at an early layer (§3.3). As this defect is not specific to localization, we explore what can be attributed to the method's high performance on numerous other tasks and find that these tasks require minimal visual grounding (§3.4).", + "bbox": [ + 89, + 431, + 482, + 689 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 696, + 232, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Before our analyses, we provide a background on the explored adapter-style VLM with visual token pruning, the evaluated benchmarks, and experimental settings.", + "bbox": [ + 89, + 719, + 482, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "VLM and token pruning. Formally, an adapter-style VLM takes as input an image $x_{\\mathrm{img}}$ and text prompt tokens $x_{\\mathrm{prompt}}$ . A pre-trained vision backbone $f$ first encodes visual features $z_{\\mathrm{img}} = f(x_{\\mathrm{img}}) \\in \\mathbb{R}^{n \\times d_{\\mathrm{vision}}}$ where $n$ is the number of image patches and $d_{\\mathrm{vision}}$ is the dimensionality of the vision encoder. Next, an adapter $p$ (either a simple linear layer or MLP) projects the vision features to embeddings $h_{\\mathrm{img}} = p(z_{\\mathrm{img}}) \\in \\mathbb{R}^{n \\times d_{\\mathrm{text}}}$ where $d_{\\mathrm{text}}$ is the dimensionality of the LLM. Lastly, $h_{\\mathrm{img}}$ is concatenated with text", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "prompt embeddings $h_{\\mathrm{prompt}} = \\mathrm{embed}(x_{\\mathrm{prompt}})$ and passed into the language model LM to generate the output text $y = \\mathrm{LM}([h_{\\mathrm{img}}; h_{\\mathrm{prompt}}])$ .", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we study the inference acceleration of VLMs where visual tokens are pruned within the attention mechanism of LM. We focus on the FastV [10] approach, where after layer $K$ in LM, $R\\%$ of visual tokens are pruned away based on a ranking function $g_{\\phi}$ which ranks tokens based on criteria $\\phi$ . In practice, the attention score received from the last text token is used as the criteria, referred to as $\\phi_{\\mathrm{original}}$ . Note that the positional information is preserved when performing the pruning.", + "bbox": [ + 511, + 136, + 903, + 271 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With this approach, we measure the acceleration using the theoretical FLOPS reduction ratio related to the image tokens. For one Transformer layer, the total FLOPS is estimated as $C = 4nd^2 + 2n^2d + 2nd_m$ where $d = d_{\\mathrm{text}}$ and $m$ is the intermediate size of FNN. Given that the Transformer has $T$ layers in total and after layer $K$ , we maintain $\\hat{n} = (1 - R\\%) * n$ visual tokens, we calculate the FLOPS reduction as", + "bbox": [ + 511, + 271, + 903, + 381 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n1 - \\frac {K * C + (T - K) * (4 \\hat {n} d ^ {2} + 2 \\hat {n} ^ {2} d + 2 \\hat {n} d m)}{C}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 385, + 903, + 414 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Benchmarking. We evaluate the accelerated VLMs on a large suite of benchmarks from [18] which includes evaluations spanning the areas of localization, open-ended visual question answering, and challenge sets. For all benchmarks, we follow the same evaluation protocol as [18].", + "bbox": [ + 511, + 417, + 903, + 492 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Localization. We evaluate localization performance using the RefCOCO, RefCOCO+, RefCOCOg [19, 42] and OCID-Ref [37] datasets. RefCOCO contains short, spatially grounded descriptions, while RefCOCO+ focuses on appearance-based descriptions. RefCOCOg, on the other hand, features longer and more detailed descriptions. OCID-Ref is a robotics dataset that tests generalization to out-of-distribution scenarios, particularly for object localization in cluttered environments.", + "bbox": [ + 511, + 493, + 903, + 627 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Open-Ended Visual Question Answering. We evaluate general visual reasoning using the VizWiz [5] and VQAv2 [13] datasets, spatial reasoning using the GQA [16] dataset, and reasoning around text using the TextVQA [32] dataset. For TextVQA, following [18], we exclude OCR-system parsed input tokens to better evaluate the effect of pruning on visual capabilities.", + "bbox": [ + 511, + 628, + 903, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Challenge Sets. We additionally evaluate on the VSR [26], TallyQA [1], POPE [25], and AI2D [20] closed-set prediction datasets. VSR includes binary spatial relationship questions, TallyQA consists of counting questions, POPE probes hallucination, and AI2D contains multiple-choice questions referring to scientific diagrams and charts.", + "bbox": [ + 511, + 734, + 903, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Experimental settings. In our experiments, we utilize a VLM with SigLIP ViT-SO400M [43] as $f$ , a one-layer MLP with GELU activation as $p$ , and Llama 2 7B [35] as LM. The model was trained on the multimodal instruction tuning dataset presented in [27] in a single-stage with $f$ frozen.", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22828", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9f876d33145c48301090e7ce73fff2709a0513bc1bcc71d3f92ce0bec797b033.jpg", + "image_caption": [ + "Figure 2. Contrasting the difference in performance dropoff on the challenging vision-centric localization task (Left) versus the other evaluated tasks (Middle) when pruning visual tokens after the shallow LLM layers. Whereas performance decrease is minimal for most tasks, localization exhibits roughly a linear decrease to zero as the ratio of pruned tokens increases. Right: Per-task performance breakdown across various setups of pruning ratios. Using $K = 3$ for all pruning setups." + ], + "image_footnote": [], + "bbox": [ + 91, + 97, + 344, + 251 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/286cdba0b0aa6f7e1371ab0a3b257a79e651aa61adad2d8e7c0d50598ed9a34b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 346, + 98, + 602, + 251 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4f4bf20589ae09615e518ba86872a1f80bc19be70d89303d0cb3e26acde805b6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 89, + 903, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Impact of early visual token pruning varies drastically by task", + "text_level": 1, + "bbox": [ + 89, + 349, + 482, + 380 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We examine the effect of pruning visual tokens after the shallow LLM layers $(K = 3)$ across a variety of tasks. We compare various pruning ratios $(R \\in \\{0.25, 0.5, 0.75, 0.9\\})$ and the baseline non-pruned model in Figure 2.", + "bbox": [ + 89, + 387, + 482, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the majority of evaluated tasks, heavily reducing the number of visual tokens results in minimal performance decrease (3.5% for AI2D, 0.1% for VSR, 3.7% for TallyQA, 4.8% for POPE, 5.1% for VizWiz, 7.9% for VQAv2 and 7.7% for GQA when dropping 75% of tokens). However, for a small subset of tasks—particularly TextVQA, and more severely, localization-performance drops substantially, with decreases of 42.0%, 88.9%, 88.9%, 91.0%, and 86.0% for TextVQA, RefCOCO, RefCOCO+, RefCOCOg, and OCID-Ref, respectively. Investigating further, Figure 2 shows that localization performance declines roughly linearly to zero as the pruning ratio progresses from 0 to 1, a pattern that strongly contrasts with most other tasks, where performance remains largely unaffected. While prior works report a much smaller decline in TextVQA performance [9, 39], they incorporate OCR-parsed input tokens into prompts, reducing the reliance on visual information.", + "bbox": [ + 91, + 449, + 483, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Discussion. Although localization is undoubtedly challenging for a model that discards visual tokens, as it requires fine-grained visual grounding and precise boundary identification, it is still surprising that (1) localization performance declines to the extent it does even when many tokens are maintained (e.g., dropping $50\\%$ of tokens reduces average performance by $45\\%$ ), and (2) most non-localization benchmarks, except for TextVQA, exhibit a drastically different pattern despite these benchmarks aiming to evaluate visual grounding abilities, such as counting for TallyQA, spatial reasoning for GQA, and chart understanding for AI2D. In the following sections, we seek to explain these unexpected findings and understand their implications.", + "bbox": [ + 89, + 704, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Interpreting poor vision-centric task performance", + "text_level": 1, + "bbox": [ + 511, + 349, + 906, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We next investigate why pruning visual tokens after the shallow LLM layers has such an adverse effect on localization performance. Intuitively, even localization should not require all image tokens, as there are background tokens irrelevant to identifying specific objects that can be pruned. Thus, it is surprising that performance begins to degrade even with minimal token pruning. This suggests that the pruning criteria, intended to retain important tokens, fails to effectively distinguish which tokens contain necessary visual information.", + "bbox": [ + 511, + 387, + 906, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To assess the efficacy of the pruning criteria, we first examine the distribution of retained tokens across all benchmark examples. As shown in Figure 3(b), we find that pruning visual tokens after the shallow LLM layers $(K = 3)$ retains tokens concentrated at the bottom of the image. For example, when $R = 0.75$ , the average sampled token position across all datasets is situated $80.7\\%$ of the way down the image. A Chi-Square test confirms the non-uniformity of selected tokens' y-positions ( $p$ -value $< 0.05$ ). We also examine the criteria behavior when pruning after later layers ( $K \\in (8,16,24)$ ) and observe that the bias of selecting bottom image tokens is reduced.", + "bbox": [ + 511, + 537, + 908, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition, we observe that as the pruning layer increases, not only is the positional bias reduced, but the criteria also begins to correctly select tokens relevant to the text instruction. For example, as shown in Figure 3(a), pruning after the later layers results in a distinct concentration of selected tokens around the area corresponding to the reference expression for localization. To quantify this change in criteria effectiveness across layers, we measure the performance when varying the pruning layer $(K\\in \\{3,8,16,24\\})$ As shown in Figure 3(c), performance remains stable when pruning is performed at deeper layers $K = 16$ and $K =$ 24), whereas pruning at earlier layers $K = 3$ and $K = 8$", + "bbox": [ + 511, + 719, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22829", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/72a40db4bc8d989578cc28510d68d7755ecdb5e3bb820e170b9fbd971f1e7047.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 120, + 90, + 236, + 183 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f8b3451ae6cb8d90bbd59c738920d5f151d2634ee388ec2d74cd64afe48a1763.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 254, + 90, + 369, + 184 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f4306611f75804fbbac25d2b58ad4a9d0cade5402eacdf25ba9177093104e7fc.jpg", + "image_caption": [ + "Reference expression: a bowl of blueberries" + ], + "image_footnote": [], + "bbox": [ + 122, + 193, + 235, + 287 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4b89b253336473434b20c3d8fb99788acf6767c1652f44b5df2bfe6c9704de6c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 254, + 193, + 369, + 286 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/38e0177f484f621655db8b2c810f33c3284f3c5c2be5e90b4bca068413632f90.jpg", + "image_caption": [ + "x position K-16" + ], + "image_footnote": [], + "bbox": [ + 401, + 90, + 535, + 183 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/80b6d7be38344bc1ce262474fdcc32da3b9c5dd1cd3409bdf9e8bd20c063a787.jpg", + "image_caption": [ + "x position" + ], + "image_footnote": [], + "bbox": [ + 401, + 198, + 535, + 285 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d85b30c96cfc96f8dd07fc317f4f7e2b89fd514cce55bb57f70c7e4e1dde6c89.jpg", + "image_caption": [ + "x position $K = 24$" + ], + "image_footnote": [], + "bbox": [ + 547, + 90, + 684, + 183 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8e979172f1352d135f6b68fc32a7be4d7be7bc91db119e4c6bdfb4c56c011131.jpg", + "image_caption": [ + "x position" + ], + "image_footnote": [], + "bbox": [ + 547, + 199, + 684, + 286 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f1c771f753620cb6569dbf5c32c855866fef6378fe644ea4594a4efa2992a3f3.jpg", + "image_caption": [ + "Figure 3. (a) Example demonstrating that when pruning in early layers, selected tokens are concentrated in the bottom of the image, whereas in later layers the selected tokens more precisely cover the region of the image important for answering the question. (b) Heatmap illustrating that averaged across all benchmark examples, as the pruning layer increases, the selection of bottom visual tokens by the criteria is reduced. (c) Visualizing the effect of the pruning layer on performance for both localization and non-localization tasks. We find that pruning after layer 16 or later results in little performance dropoff, whereas pruning earlier results in a performance decrease, particularly for localization. Using $R = 0.75$ for all setups." + ], + "image_footnote": [], + "bbox": [ + 464, + 297, + 638, + 323 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ee6aacba1a99a9860ed9dd5a5ced81005abc1f80fc85a078c3cd09f037284089.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 723, + 92, + 901, + 200 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f39e6d530bafc921bbca4dbb67e4aa6c80c528f3b91748e4391e3276b911ce4d.jpg", + "image_caption": [ + "Non-localization" + ], + "image_footnote": [], + "bbox": [ + 720, + 215, + 903, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "leads to performance degradation, particularly for localization.", + "bbox": [ + 88, + 446, + 482, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Discussion. While previous research attributes the poor performance of early token pruning to the low redundancy of image tokens in shallow layers [39], we reveal a critical insight: the variable performance across pruning layers is linked to the effectiveness of selecting important tokens, as early token pruning leads to suboptimal token selection biased toward bottom tokens. To understand why the criteria disproportionately retains bottom image tokens, we must consider the LM architecture. Specifically, token selection is based on the attention received from the last text token. Since the LM uses RoPE to encode positional information [33], it exhibits a long-term decay property. While this property is well-suited for language modeling, we uncover that it poses an issue for selecting visual tokens, where tokens appearing later in the flattened, raster-scan order have higher attention scores purely due to their position. In language modeling, prior work finds that there is greater emphasis on shorter-distance information in shallow layers [15]. We predict that the visually-conditioned LM exhibits the same behavior, explaining why the positional bias is most pronounced when pruning after shallow layers.", + "bbox": [ + 91, + 477, + 483, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Explaining strong performance on other tasks", + "text_level": 1, + "bbox": [ + 89, + 818, + 478, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given that early pruning of visual tokens leads to a suboptimal tokens selection where remaining tokens are concentrated towards the bottom of the image, a natural question arises: how does the approach still maintain high perfor", + "bbox": [ + 89, + 839, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f3dc8989efff27dd6c48a5ff61c10977b787a5d7dabe9faca6396a03a895e550.jpg", + "image_caption": [ + "Figure 4. Assessing whether the strong performance of early visual token pruning for many tasks can be attributed to visual information transfer before pruning or benchmarks' lack of assessing fine-grained visual capabilities. We observe that for all evaluated tasks, allowing information transfer before pruning (shown in green) does not result in substantial performance improvement over the setup without visual information transfer (shown in light green), highlighting a limitation of many benchmarks." + ], + "image_footnote": [], + "bbox": [ + 521, + 446, + 898, + 654 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "mance across a diverse range of tasks, including those aiming to evaluate visual grounding abilities? In this section, we test the following contrasting explanations:", + "bbox": [ + 511, + 806, + 905, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "1. Important visual information from top image tokens is transferred to later tokens before pruning with the unidirectional attention of the LLM.", + "bbox": [ + 513, + 854, + 906, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22830", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2. Many questions can still be inferred with access to only the suboptimal set of visual tokens primarily from the bottom of the image.", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate these hypotheses across tasks, we save the tokens retained by the studied pruning approach and, in a subsequent run, remove all other tokens before they reach the LLM, ensuring that no visual information from the pruned tokens is retained. As an additional comparison, we also evaluate a text-only model in which all visual tokens are removed before entering the LLM. As shown in Figure 4, we find that across all evaluated tasks, removing pruned tokens before the LLM results in little to no performance decrease compared to the method that allows information transfer in shallow layers prior to pruning.", + "bbox": [ + 89, + 138, + 482, + 304 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Discussion. This finding indicates that the high performance of early visual token pruning on many tasks does not stem from the effectiveness of visual information transfer in early layers but rather signifies that many benchmarks do not demand a detailed understanding of visual information to answer questions accurately. Note that for the majority of these benchmarks (except VizWiz and AI2D), performance substantially declines when not including any visual information, signifying that simply comparing vision-enabled and disabled setups is insufficient to determine a benchmark's ability to assess visual grounding.", + "bbox": [ + 89, + 305, + 482, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Improving Visual Token Pruning", + "text_level": 1, + "bbox": [ + 89, + 482, + 390, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Based on our findings from §3, we now seek to answer: can attention serve as an effective criteria for token pruning even in early layers? To this end, we propose a straightforward solution to removing the positional bias of the attention-based criteria and demonstrate that our modifications enable effective pruning even when applied after early LLM layers (§4.1). Guided by our insights, we present our final approach, FEATHER (Fast and Effective Acceleration with Ensemble cRicerca) (§4.2), and showcase its impressive visual capabilities while offering high computational efficiency (§4.3).", + "bbox": [ + 89, + 508, + 482, + 674 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Assessing potential of attention-based criteria", + "text_level": 1, + "bbox": [ + 89, + 681, + 475, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "An attention-based criteria offers the potential to naturally identify and retain only the most relevant visual tokens for answering a question. However, as shown in §3, positional bias strongly influences token selection, limiting its effectiveness. Thus, we aim to resolve the positional bias issue and explore the true promise of attention-based criteria. We compare against non-attention-based approaches that incorporate an explicit inductive bias toward reducing redundancy. Below, we outline the explored criteria.", + "bbox": [ + 89, + 704, + 482, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\phi_{-R}$ : Given the long-term decay property of RoPE, we propose the following straightforward adjustment to the pruning criteria in order to mitigate this bias and more effectively select tokens relevant to the text instruction. Namely,", + "bbox": [ + 89, + 840, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "we still compute the attention score received from the last text token, except we do not apply RoPE to the attention mechanism, thereby removing the long-term decay effect. Note that this lightweight calculation is only done once per pruning, and all other state computations are compatible with FlashAttention [12].", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\phi_{\\mathrm{uniform}}$ : In this simple criteria, we uniformly sample visual tokens in the image, using a set stride. This criteria ensures that there is good image coverage, but does so by sacrificing the ability to have more densely captured visual information for a particular image region. To use a similar computational saving to the attention-based criteria when $R = 0.75$ , we use a stride of two, resulting in 196 selected tokens.", + "bbox": [ + 511, + 181, + 903, + 300 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\phi_{\\mathrm{KNN}}$ : Inspired by dynamic visual tokenization [17], we select tokens based on local density. For each visual token $z_{img}[i]$ where $i \\in \\{0, 1, \\dots, n-1\\}$ , we compute its local density $\\rho_i$ using K-nearest neighbors and calculate distance index $\\delta_i$ as:", + "bbox": [ + 511, + 301, + 903, + 371 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\delta_ {i} = \\left\\{ \\begin{array}{l l} \\min _ {j: \\rho_ {j} > \\rho_ {i}} \\| z _ {i m g} [ i ] - z _ {i m g} [ j ] \\| ^ {2}, & \\text {i f} \\exists j \\text {s . t .} \\rho_ {j} > \\rho_ {i}. \\\\ \\max _ {j} \\| z _ {i m g} [ i ] - z _ {i m g} [ j ] \\| ^ {2}, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 378, + 903, + 424 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\delta_{i}$ represents how far away $z_{img}[i]$ is from other high-density tokens. We use $\\rho_{i}*\\delta_{i}$ as the token importance score for the criteria and select 196 tokens.", + "bbox": [ + 511, + 431, + 903, + 477 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\phi_{-R} + \\phi_{\\mathrm{uniform}}$ : Since $\\phi_{-R}$ focuses on selecting important tokens and $\\phi_{\\mathrm{uniform}}$ aims to reduce redundancy, we propose an ensemble of these two to leverage their respective strengths. Specifically, we apply $\\phi_{-R}$ while incorporating a small number of uniformly sampled tokens using a stride of three.", + "bbox": [ + 511, + 478, + 903, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Improvement of removing RoPE: In Table 1, we report results evaluating the effectiveness of these criteria. Comparing $\\phi_{-R}$ with $\\phi_{\\mathrm{original}}$ , we find that by removing RoPE, at $K = 3$ there is a $183\\%$ average improvement on localization tasks and at $K = 8$ , there is a $17\\%$ average improvement. These performance improvements demonstrate that once the impact of token position on attention is removed, attention score can more effectively be used for the criteria when applied after early LLM layers. Note that the narrowing gap in performance between $\\phi_{-R}$ and $\\phi_{\\mathrm{original}}$ from $K = 3$ to $K = 8$ is likely because $\\phi_{\\mathrm{original}}$ has less of a bias towards selecting bottom image tokens as the pruning layer increases.", + "bbox": [ + 511, + 568, + 903, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Token selection improves when pruning later: Comparing $\\phi_{-R}$ when $K = 8$ versus $K = 3$ , we see that the average localization performance improves by $63\\%$ . However, it remains unclear whether this performance increase is affected by factors other than the ability of the criteria to select important tokens. Therefore, to directly compare the criteria across different layers, we take the tokens selected by both criteria and evaluate how well the model can perform with only these tokens passed into the LLM (as is", + "bbox": [ + 511, + 763, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22831", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e23128864cf3cc686470aa0ab3d5fc22c58669e2d99905958f5beb449f71a20a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Pruning LayerCriteriaFLOPS RedLocalizationOpen-Ended VQAChallenge Sets
AvgOCID-RefRefCOCObRefCOCO+RefCOCOAvgTextVQAGQAVQAv2VizWizAvgPOPETallyQAVSRAI2D
K=3Attention-based
φoriginal68%5.95.75.16.16.754.831.858.472.756.364.083.257.163.352.4
φR(Ours)68%16.722.915.113.315.359.041.661.276.057.364.785.258.262.253.2
Δ+10.7+17.2+10.0+7.3+8.6+4.2+9.7+2.8+3.2+1.1+0.7+1.9+1.1-1.1+0.8
Non-attention-based
φKNN66%23.915.124.926.029.658.439.960.974.458.462.881.255.961.552.8
φuniform66%28.020.628.629.733.359.041.461.875.957.164.685.258.161.953.0
Ensemble
φR+φuniform(Ours)61%27.229.127.224.727.761.246.662.377.458.465.486.058.962.754.0
K=8Attention-based
φoriginal56%23.319.423.524.026.359.845.060.376.157.864.685.457.562.653.0
φR(Ours)56%27.327.126.726.429.261.449.061.577.457.865.586.758.663.053.7
Δ+4.0+7.6+3.2+2.5+2.9+1.6+4.0+1.2+1.3+0.0+0.8+1.3+1.1+0.4+0.6
Non-attention-based
φKNN55%23.615.424.425.229.458.640.261.174.558.562.981.456.260.953.0
φuniform55%30.324.631.030.934.859.342.261.876.057.464.485.357.961.053.2
Ensemble
φR+φuniform(Ours)50%35.632.035.935.438.862.751.762.478.158.666.087.459.163.654.0
", + "bbox": [ + 91, + 85, + 906, + 459 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4c27e9f0a13c1d9fa01d9fead10fe2d571a9a67486b5057f2bda53264bf9cd50.jpg", + "table_caption": [ + "Table 1. Evaluating alternative criteria for token pruning after the early LLM layers. For each task and pruning layer, we **bold** the best result and **underline** the second-best result. Our main findings include: (1) our proposed RoPE-free criteria $\\phi_{-R}$ substantially improves pruning performance compared to the original criteria $\\phi_{\\mathrm{original}}$ ; (2) pruning later ( $K = 8$ ) yields higher performance than pruning earlier ( $K = 3$ ); and (3) integrating uniform sampling into the attention-based criteria with $\\phi_{-R} + \\phi_{\\mathrm{uniform}}$ enhances effectiveness. Using $R = 0.75$ for $\\phi_{\\mathrm{original}}$ , $\\phi_{-R}$ , and $\\phi_{-R} + \\phi_{\\mathrm{uniform}}$ . See §4.1 for criteria definitions." + ], + "table_footnote": [], + "table_body": "
KOCID-RefRefCOCOgRefCOCO+RefCOCO
323.816.214.516.3
826.729.826.829.8
", + "bbox": [ + 94, + 560, + 475, + 621 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Assessing localization performance when using the selected tokens from $\\phi_{-R}$ for the entirety of the LLM. We find that $\\phi_{-R}$ applied at a later LLM layer results in a better token selection.", + "bbox": [ + 89, + 630, + 482, + 674 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "done in §3.4). As shown in Table 2, we find the selected tokens from $\\phi_{-R}$ applied after layer 8 are superior compared to the tokens from $\\phi_{-R}$ applied after layer 3.", + "bbox": [ + 89, + 702, + 482, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Integrating uniform sampling with attention-based criteria is beneficial in early layers: When comparing the attention-based criteria $\\phi_{-R}$ with the two non-attention based criteria $\\phi_{\\mathrm{KNN}}$ and $\\phi_{\\mathrm{uniform}}$ , we find that $\\phi_{-R}$ results in better performance on some tasks (e.g., OCID-Ref, TextVQA), while $\\phi_{\\mathrm{KNN}}$ and $\\phi_{\\mathrm{uniform}}$ outperform on other tasks (e.g., RefCOCO). This varying effectiveness of criteria types presumably comes from some tasks requiring a more detailed understanding of specific image regions while others benefit from a broader understanding with full image", + "bbox": [ + 88, + 750, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "coverage. When we combine these two criteria types with $\\phi_{-R} + \\phi_{\\mathrm{uniform}}$ , we find that when $K = 3$ this approach improves upon localization performance of $\\phi_{-R}$ by $63\\%$ but has a slight decrease compared to $\\phi_{\\mathrm{uniform}}$ by $2.9\\%$ . However, when $K = 8$ , this approach far outperforms both of the individual criteria, improving $\\phi_{-R}$ by $30\\%$ and $\\phi_{\\mathrm{uniform}}$ by $17\\%$ . Note that this setup results in slightly less computation efficiency than $\\phi_{-R}$ (7% drop in FLOPS reduction for $K = 3$ and $6\\%$ for $K = 16$ ), but this is outweighed by the performance improvements.", + "bbox": [ + 511, + 564, + 906, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We provide qualitative examples of various criteria token selection and additional results using a different model setup in the supplement.", + "bbox": [ + 511, + 715, + 905, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Distilling insights", + "text_level": 1, + "bbox": [ + 511, + 771, + 684, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Guided by our insights, we now present our final approach FEATHER (Fast and Effective Acceleration wiTH Ensemble cRicerca). In this approach, we first perform pruning after an early layer $(K = 8)$ , utilizing our proposed criteria $\\phi_{-R} + \\phi_{\\mathrm{uniform}}$ to retain $(1 - R)\\%$ of tokens. Given our finding that the criteria improves as the LLM layer increases, we additionally prune a second time at $K = 16$ .", + "bbox": [ + 511, + 795, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22832", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4f4a5029d3b97b52c56898e84e33a5b61a00f898caeb79e93e6a3149b6ddd97a.jpg", + "image_caption": [ + "Figure 5. Comparing FEATHER performance against FastV [10] and PyramidDrop [39]. We find that FEATHER far outperforms both compared methods, particularly for the vision-centric task of localization." + ], + "image_footnote": [], + "bbox": [ + 91, + 89, + 908, + 258 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "At this stage, we utilize $\\phi_{-R}$ as the criteria, as uniform sampling should not be necessary when the attention-based criteria reliably selects important tokens. For the reduction ratio, we choose to only retain $(1 - R)^{2\\%}$ of the remaining tokens since the attention-based criteria has proved highly effective when pruning at later layers, even when using the original criteria $\\phi_{\\mathrm{original}}$ (see Figure 3(c)).", + "bbox": [ + 88, + 325, + 480, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Comparison against FastV and PyramidDrop", + "text_level": 1, + "bbox": [ + 89, + 446, + 475, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We compare our approach against the one-stage early pruning approach of FastV ( $K = 3$ ) [10] and the multi-stage approach of PyramidDrop ( $K = [8, 16, 24]$ ) [39]. Note that both of these methods use $\\phi_{\\mathrm{original}}$ for the criteria.", + "bbox": [ + 88, + 472, + 482, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Figure 5, we find that FEATHER far outperforms the baselines, especially on the localization tasks. Specifically, for comparable computational costs (64% FLOPS reduction for our approach, 68% FLOPS reduction for FastV, and 65% FLOPS reduction for PyramidDrop), we observe that for localization tasks, FEATHER exhibits more than $5 \\times$ average performance improvement compared to FastV and a $36\\%$ average performance improvement compared to PyramidDrop. For non-localization tasks, FEATHER has a $7.8\\%$ improvement over FastV and $1.5\\%$ improvement over PyramidDrop. Note that PyramidDrop performs substantially better than FastV as it prunes fewer tokens in an early layer. However, it still suffers from an ineffective pruning strategy at this stage, though the impact is less pronounced since it predominantly prunes later.", + "bbox": [ + 88, + 535, + 482, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Remarkably, with our $64\\%$ FLOPS reduction setup, after layer 16 only $3.3\\%$ of tokens are retained, yet the average localization performance decrease compared to the baseline method with no token pruning is only $26\\%$ (Figure 1 includes an example of retained tokens after layer 16). This finding illustrates that even for vision-centric tasks, maintaining strong performance while gaining huge acceleration speedups with extensive pruning is possible, but it heavily relies on the effectiveness of the pruning criteria.", + "bbox": [ + 88, + 763, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 323, + 633, + 339 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we examine the visual capabilities of the VLM acceleration approach of pruning visual tokens after shallow LLM layers. While strong performance is maintained on most evaluated tasks, it fails on more vision-centric tasks like TextVQA and localization due to its flawed pruning criteria that predominately selects visual tokens from the bottom part of the image. Observing this same flawed criteria on other tasks, we show that strong performance is largely due to the benchmarks' inability to assess fine-grained visual capabilities. Next, we propose and evaluate several alternative criteria to improve visual capabilities, ultimately arriving at our final method, FEATHER. This approach refines the attention-based criteria to address the token selection bias while using uniform sampling for better image coverage. It then prunes more aggressively when the criteria more effectively identifies important tokens. We find that FEATHER has more than $5 \\times$ performance improvement on localization compared to the original acceleration approach.", + "bbox": [ + 511, + 348, + 906, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While we study a particular type of VLM acceleration, our work highlights a broad challenge in evaluating VLMs. Building on [34], which shows that text-only models can perform well on some vision-language benchmarks, we demonstrate that even benchmarks with substantial differences between vision-enabled and disabled setups may not assess fine-grained visual capabilities. We address this issue by focusing on more vision-centric localization tasks, though this limits the analysis to a few specific types of skills. To accurately assess a wider range of visual capabilities, future work may explore how to resolve current dataset biases that models can exploit [21]. Additionally, while removing RoPE for token pruning effectively eliminates positional bias, it may introduce unintended effects on attention weights influencing token selection. Future research could investigate more robust methods for encoding positional information in visually conditioned language models to prevent positional artifacts in cross-modal interactions.", + "bbox": [ + 511, + 619, + 908, + 892 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22833", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work is supported in part by the National Science Foundation (NSF) under Grant No. 2026498, the Stanford AIMI-HAI Partnership Grant, and the NSF Graduate Research Fellowship Program under Grant No. DGE-2146755 (for M.E.). Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of any other entity.", + "bbox": [ + 89, + 90, + 485, + 217 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 233, + 187, + 248 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Manoj Acharya, Kushal Kafle, and Christopher Kanan. Tallyqa: Answering complex counting questions. In Proceedings of the AAAI conference on artificial intelligence, pages 8076-8084, 2019. 3", + "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 1", + "[3] Kazi Hasan Ibn Arif, JinYi Yoon, Dimitrios S Nikolopoulos, Hans Vandierendonck, Deepu John, and Bo Ji. Hired: Attention-guided token dropping for efficient inference of high-resolution vision-language models in resource-constrained environments. arXiv preprint arXiv:2408.10945, 2024. 2, 3", + "[4] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 1", + "[5] Jeffrey P Bigham, Chandrika Jayant, Hanjie Ji, Greg Little, Andrew Miller, Robert C Miller, Robin Miller, Aubrey Tatarowicz, Brandyn White, Samual White, et al. Vizwiz: nearly real-time answers to visual questions. In Proceedings of the 23nd annual ACM symposium on User interface software and technology, pages 333-342, 2010. 3", + "[6] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. In The Eleventh International Conference on Learning Representations, 2023. 3", + "[7] Mu Cai, Jianwei Yang, Jianfeng Gao, and Yong Jae Lee. Matryoshka multimodal models. arXiv preprint arXiv:2405.17430, 2024. 2", + "[8] Wenhao Chai, Enxin Song, Yilun Du, Chenlin Meng, Vashisht Madhavan, Omer Bar-Tal, Jeng-Neng Hwang, Saining Xie, and Christopher D Manning. Auroracap: Efficient, performant video detailed captioning and a new benchmark. arXiv preprint arXiv:2410.03051, 2024. 3", + "[9] Jieneng Chen, Luoxin Ye, Ju He, Zhao-Yang Wang, Daniel Khashabi, and Alan Yuille. Efficient large multi-modal models via visual context compression. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 4", + "[10] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2" + ], + "bbox": [ + 93, + 258, + 480, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. ECCV, 2024. 2, 3, 8", + "[11] Yi Chen, Jian Xu, Xu-Yao Zhang, Wen-Zhuo Liu, YangYang Liu, and Cheng-Lin Liu. Recoverable compression: A multimodal vision token recovery mechanism guided by text information. arXiv preprint arXiv:2409.01179, 2024. 3", + "[12] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in neural information processing systems, 35:16344-16359, 2022. 6", + "[13] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 3", + "[14] Yefei He, Feng Chen, Jing Liu, Wenqi Shao, Hong Zhou, Kaipeng Zhang, and Bohan Zhuang. Zipvl: Efficient large vision-language models with dynamic token sparsification and kv cache compression. arXiv preprint arXiv:2410.08584, 2024.3", + "[15] Xiangyu Hong, Che Jiang, Biqing Qi, Fandong Meng, Mo Yu, Bowen Zhou, and Jie Zhou. On the token distance modeling ability of higher rope attention dimension. arXiv preprint arXiv:2410.08703, 2024. 5", + "[16] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 3", + "[17] Peng Jin, Ryuichi Takanobu, Wancai Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13700-13710, 2024. 2, 6", + "[18] Siddharth Karamcheti, Suraj Nair, Ashwin Balakrishna, Percy Liang, Thomas Kollar, and Dorsa Sadigh. Prismatic VLMs: Investigating the design space of visually-conditioned language models. In Proceedings of the 41st International Conference on Machine Learning, pages 23123-23144. PMLR, 2024. 3", + "[19] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014. 3", + "[20] Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Minjoon Seo, Hannaneh Hajishirzi, and Ali Farhadi. A diagram is worth a dozen images. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 235-251. Springer, 2016. 3", + "[21] Baiqi Li, Zhiqiu Lin, Wenxuan Peng, Jean de Dieu Nyandwi, Daniel Jiang, Zixian Ma, Simran Khanuja, Ranjay Krishna, Graham Neubig, and Deva Ramanan. Naturalbench: Evaluating vision-language models on natural adversarial samples. Advances in Neural Information Processing Systems, 37:17044-17068, 2025. 8" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22834", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[22] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 1", + "[23] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 1", + "[24] Kevin Y Li, Sachin Goyal, Joao D Semedo, and J Zico Kolter. Inference optimal vlms need only one visual token but larger models. arXiv preprint arXiv:2411.03312, 2024. 2", + "[25] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 3", + "[26] Fangyu Liu, Guy Emerson, and Nigel Collier. Visual spatial reasoning. Transactions of the Association for Computational Linguistics, 11:635-651, 2023. 3", + "[27] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 1, 3", + "[28] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024. 2", + "[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1", + "[30] Yifei Liu, Mathias Gehrig, Nico Messikommer, Marco Cannici, and Davide Scaramuzza. Revisiting token pruning for object detection and instance segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2658-2668, 2024. 3", + "[31] Yuzhang Shang, Mu Cai, Bingxin Xu, Yong Jae Lee, and Yan Yan. Llava-prumerge: Adaptive token reduction for efficient large multimodal models. arXiv preprint arXiv:2403.15388, 2024. 2", + "[32] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 3", + "[33] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 5", + "[34] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, et al. Cambrian1: A fully open, vision-centric exploration of multimodal llms. arXiv preprint arXiv:2406.16860, 2024. 8", + "[35] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al." + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3", + "[36] Zhongwei Wan, Ziang Wu, Che Liu, Jina Huang, Zhihong Zhu, Peng Jin, Longyue Wang, and Li Yuan. Look-m: Look-once optimization in kv cache for efficient multimodal long-context inference. arXiv preprint arXiv:2406.18139, 2024. 3", + "[37] Ke-Jyun Wang, Yun-Hsuan Liu, Hung-Ting Su, Jen-Wei Wang, Yu-Siang Wang, Winston Hsu, and Wen-Chin Chen. OCID-ref: A 3D robotic dataset with embodied language for clutter scene grounding. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5333–5338, Online, 2021. Association for Computational Linguistics. 3", + "[38] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1", + "[39] Long Xing, Qidong Huang, Xiaoyi Dong, Jiajie Lu, Pan Zhang, Yuhang Zang, Yuhang Cao, Conghui He, Jiaqi Wang, Feng Wu, et al. Pyramiddrop: Accelerating your large vision-language models via pyramid visual redundancy reduction. arXiv preprint arXiv:2410.17247, 2024. 2, 3, 4, 5, 8", + "[40] Senqiao Yang, Yukang Chen, Zhuotao Tian, Chengyao Wang, Jingyao Li, Bei Yu, and Jiaya Jia. Visionzip: Longer is better but not necessary in vision language models. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 19792-19802, 2025. 2", + "[41] Gaotong Yu, Yi Chen, and Jian Xu. Balancing performance and efficiency: A multimodal large language model pruning method based image text interaction. arXiv preprint arXiv:2409.01162, 2024.3", + "[42] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 3", + "[43] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 3", + "[44] Qizhe Zhang, Aosong Cheng, Ming Lu, Zhiyong Zhuo, Minqi Wang, Jiajun Cao, Shaobo Guo, Qi She, and Shang-hang Zhang. [cls] attention is all you need for training-free visual token pruning: Make vlm inference faster. arXiv eprints, pages arXiv-2412, 2024. 2" + ], + "bbox": [ + 516, + 92, + 903, + 796 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22835", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Feather the Throttle_ Revisiting Visual Token Pruning for Vision-Language Model Acceleration/eb2a11ee-50f9-4c43-8cc0-429a524795eb_model.json b/2025/Feather the Throttle_ Revisiting Visual Token Pruning for Vision-Language Model Acceleration/eb2a11ee-50f9-4c43-8cc0-429a524795eb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bff8c8afab3a1448c25c858c3802a38eb616de1e --- /dev/null +++ b/2025/Feather the Throttle_ Revisiting Visual Token Pruning for Vision-Language Model Acceleration/eb2a11ee-50f9-4c43-8cc0-429a524795eb_model.json @@ -0,0 +1,2024 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.208, + 0.131, + 0.79, + 0.177 + ], + "angle": 0, + "content": "Feather the Throttle: Revisiting Visual Token Pruning for Vision-Language Model Acceleration" + }, + { + "type": "text", + "bbox": [ + 0.307, + 0.204, + 0.692, + 0.24 + ], + "angle": 0, + "content": "Mark Endo, Xiaohan Wang, Serena Yeung-Levy \nStanford University" + }, + { + "type": "text", + "bbox": [ + 0.32, + 0.242, + 0.673, + 0.258 + ], + "angle": 0, + "content": "{markendo,xhanwang,syyeung}@stanford.edu" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.327, + 0.485, + 0.691 + ], + "angle": 0, + "content": "Recent works on accelerating Vision-Language Models achieve strong performance across a variety of vision-language tasks despite highly compressing visual information. In this work, we examine the popular acceleration approach of early pruning of visual tokens inside the language model. Surprisingly, we find that while strong performance is maintained across many tasks, it exhibits drastically different behavior for a subset of vision-centric tasks such as localization. Upon further investigation, we uncover a core issue with the acceleration approach where most tokens towards the top of the image are pruned away. Yet, on many benchmarks aiming to evaluate vision-centric capabilities, strong performance persists with the flawed pruning strategy, highlighting these benchmarks' limited ability to assess fine-grained visual capabilities. Based on these findings, we propose FEATHER (Fast and Effective Acceleration with Ensemble cRicerca), a straightforward approach that resolves the discovered early-layer pruning issue and further enhances the preservation of relevant tokens via multistage pruning with early uniform sampling to ensure broad image coverage. With comparable computational savings, we find that FEATHER achieves more than \\(5 \\times\\) performance improvement on the vision-centric localization benchmarks compared to the original acceleration approach." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.723, + 0.222, + 0.739 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The exploration of Vision-Language Models (VLMs) is a critical area of computer vision and natural language processing research, centered on combining large language models (LLMs) with visual encoders to enable multimodal perception, reasoning, and understanding capabilities. While earlier works explored sophisticated schemes for conditioning language models with visual information [2, 22, 23], more recently the space has shifted to predominately using the simplistic approach of taking patch features from pre-trained visual encoders and projecting them" + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.292, + 0.539, + 0.306 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.291, + 0.902, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.566, + 0.539, + 0.579 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.566, + 0.9, + 0.665 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.679, + 0.907, + 0.804 + ], + "angle": 0, + "content": "Figure 1. (a) Although FastV prunes most visual tokens from the upper portion of the image, the approach still displays strong performance on a variety of evaluated vision-language tasks except for a small subset of vision-centric tasks like localization. (b) Based on our findings, we propose FEATHER (Fast and Effective Acceleration wiTH Ensemble cRicerca), a straightforward approach that resolves the existing issue of selecting bottom tokens, additionally maintains uniformly sampled tokens to ensure good coverage over the whole image, and prunes in two stages." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.903 + ], + "angle": 0, + "content": "to the input space of the language model with a light-weight adapter [4, 27, 29, 38]. Using image patches as tokens, however, comes with the drawback of being computationally inefficient. To achieve fine-grained resolution, the im" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22826" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.394 + ], + "angle": 0, + "content": "age is divided into many patches. This large number of patches significantly increases computational demands due to the quadratic complexity of the attention operation in Transformers. As a result, many recent works have focused on accelerating these methods by compressing visual information, demonstrating that heavy compression can still maintain strong performance across a wide variety of tasks [3, 7, 10, 31, 39]. For instance, FastV [10] prunes \\(50\\%\\) of visual tokens after the shallow layers of the Language Model (LLM) while not compromising in performance across a range of image and video understanding tasks [10]. Another work reveals that for a fixed computational budget on visual reasoning tasks, optimal performance is achieved by using the largest LLM possible while sacrificing visual information, often reducing the visual token count to a single token [24]. With such compressed visual information, it is still unclear how these methods achieve high performance on tasks assessing vision capabilities such as visual reasoning and understanding, and whether there are more demanding visual tasks where these methods fail." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.398, + 0.485, + 0.716 + ], + "angle": 0, + "content": "The motivation of our study is to get a better understanding of the vision capabilities of accelerated VLMs given that they leverage highly compressed visual information, focusing specifically on FastV approach. Evaluating across a wide range of vision-language tasks, we find that while the approach maintains strong performance across many tasks, there is a substantial decrease in performance for TextVQA and a severe decrease for localization tasks. When analyzing the approach's poor performance, we uncover a fundamental issue with the pruning criteria in early layers where tokens towards the top of the image are disproportionally removed due to the long-term decay property of RoPE. Since this issue is not task-specific, we examine how performance is maintained across most evaluated tasks and find that even when tokens are entirely discarded using the flawed criteria, performance remains largely unchanged. This highlights how many benchmarks require minimal visual grounding. Our finding underscores a significant challenge in the field of multimodal learning, not only for measuring the effectiveness of VLM acceleration methods but also for benchmarking the visual capabilities of VLMs as a whole." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Given the discovered limitation of the studied approach, we next explore whether the attention-based criteria can be modified to enable robust visual token pruning, leading to our final method, FEATHER (Fast and Effective Acceleration with Ensemble Ricerca). Specifically, we propose a straightforward modification to the attention-based criteria by removing RoPE, demonstrating that this resolves the positional bias issue and greatly improves performance. Additionally, we incorporate uniform sampling in early token pruning to ensure adequate coverage of all image regions and apply more extensive pruning at a later layer, when the attention-based criteria better discerns im" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.333 + ], + "angle": 0, + "content": "important tokens. This strategy is analogous to how a race-car driver feathers the throttle by gradually pressing the accelerator at the beginning of a turn to maintain grip and then accelerating more aggressively once the car is past the apex. We show that FEATHER results in substantial performance gains compared to the original acceleration approach, improving localization performance more than \\(5 \\times 5\\) with comparable computational savings. Strikingly, we find that our approach achieves this performance improvement while only retaining \\(3.3\\%\\) of visual tokens for the second half of LLM layers. Overall, our work demonstrates that while visual compression can maintain strong performance even on challenging vision-centric tasks, its effectiveness depends on a well-designed strategy, which is currently difficult to assess due to many vision-language benchmarks not thoroughly evaluating vision capabilities." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.334, + 0.794, + 0.348 + ], + "angle": 0, + "content": "In summary, our contributions include:" + }, + { + "type": "text", + "bbox": [ + 0.537, + 0.351, + 0.905, + 0.425 + ], + "angle": 0, + "content": "- Revealing inconsistent effect of early pruning: Our evaluations across 12 benchmarks show that early visual token pruning minimally affects performance on most tasks, but causes a substantial dropoff for TextVQA and a drastic decline for localization tasks." + }, + { + "type": "text", + "bbox": [ + 0.538, + 0.426, + 0.905, + 0.486 + ], + "angle": 0, + "content": "- Analyzing poor performance: We uncover that the attention-based criteria for selecting important tokens exhibits a bias towards bottom image tokens in early layers due to the long-term decay property of RoPE." + }, + { + "type": "text", + "bbox": [ + 0.538, + 0.487, + 0.905, + 0.576 + ], + "angle": 0, + "content": "- Identifying benchmark limitations in assessing token pruning: We find that performance remains strong even with suboptimal pruning criteria and removal of pruned tokens before transferring information in the LLM, demonstrating how many benchmarks fail to assess fine-grained visual capabilities." + }, + { + "type": "text", + "bbox": [ + 0.538, + 0.577, + 0.905, + 0.697 + ], + "angle": 0, + "content": "- Improving attention-based criteria: With the straightforward modification of removing RoPE, we demonstrate that attention can be used as an effective criteria for selecting tokens and performance can be further improved with simple strategies like incorporating uniform sampling and pruning more extensively at a later layer, when the criteria better discerns important tokens." + }, + { + "type": "list", + "bbox": [ + 0.537, + 0.351, + 0.905, + 0.697 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.71, + 0.658, + 0.725 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Recent efforts to accelerate VLMs can be broadly divided into two main categories: compressing visual information before it enters the LLM, and compressing visual information within the LLM itself. In the first category, ChatUniVi [17] dynamically merges visual tokens with similar semantic meanings. PruMerge [31], FasterVLM [44], and VisionZip [40] select important tokens based on ViT attention ([CLS] or average attention received). PruMerge then merges important tokens with unselected ones while VisionZip merges unselected tokens together. For the LLaVA-NeXT [28] approach of partitioning an image into" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22827" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "sub-images where inefficiency is an even bigger problem, HiRED [3] selects tokens with top feature important on each sub-image with an allocated budget. Other methods argue that the input image alone does not include enough information to select important patches and thus use the textual input to recover visually meaningful tokens [11, 41]. Alternatively, some approaches reduce visual information within the ViT by merging or pruning tokens [6, 8, 30]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.213, + 0.483, + 0.394 + ], + "angle": 0, + "content": "For the second category of approaches, where visual information is compressed within the LLM itself, LOOK-M reduces the multimodal KV cache size [36]. In our study, we focus on the popular FastV approach [10]. This work identifies that attention over image tokens is sparse in deeper layers of the LLM. Based on this observation, they propose to prune away unimportant vision tokens after the shallow layers of the LLM, achieving a \\(45\\%\\) reduction in FLOPS with nearly no performance loss. Works since have proposed alterations to this setup such as adaptively determining the number of pruned tokens instead of using a fixed ratio [14] or pruning in multiple stages [39]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.407, + 0.376, + 0.424 + ], + "angle": 0, + "content": "3. Revising Visual Token Pruning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.432, + 0.483, + 0.69 + ], + "angle": 0, + "content": "In this section, we aim to get a better understanding of the vision capabilities of the VLM acceleration approach of pruning visual tokens after shallow LLM layers. After outlining preliminaries (§3.1), we take a closer look at how the approach performs across a broad range of tasks. Upon inspection, we discover that while heavily pruning visual tokens after the shallow LLM layers has little effect on performance across a variety of tasks, this approach fails decisively on more vision-centric tasks, particularly localization (§3.2). Next, we examine why this method struggles with localization, uncovering that the poor performance is due to the ineffectiveness of the pruning criteria to select important tokens when applied at an early layer (§3.3). As this defect is not specific to localization, we explore what can be attributed to the method's high performance on numerous other tasks and find that these tasks require minimal visual grounding (§3.4)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.698, + 0.233, + 0.713 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.765 + ], + "angle": 0, + "content": "Before our analyses, we provide a background on the explored adapter-style VLM with visual token pruning, the evaluated benchmarks, and experimental settings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "VLM and token pruning. Formally, an adapter-style VLM takes as input an image \\( x_{\\mathrm{img}} \\) and text prompt tokens \\( x_{\\mathrm{prompt}} \\). A pre-trained vision backbone \\( f \\) first encodes visual features \\( z_{\\mathrm{img}} = f(x_{\\mathrm{img}}) \\in \\mathbb{R}^{n \\times d_{\\mathrm{vision}}} \\) where \\( n \\) is the number of image patches and \\( d_{\\mathrm{vision}} \\) is the dimensionality of the vision encoder. Next, an adapter \\( p \\) (either a simple linear layer or MLP) projects the vision features to embeddings \\( h_{\\mathrm{img}} = p(z_{\\mathrm{img}}) \\in \\mathbb{R}^{n \\times d_{\\mathrm{text}}} \\) where \\( d_{\\mathrm{text}} \\) is the dimensionality of the LLM. Lastly, \\( h_{\\mathrm{img}} \\) is concatenated with text" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "prompt embeddings \\( h_{\\mathrm{prompt}} = \\mathrm{embed}(x_{\\mathrm{prompt}}) \\) and passed into the language model LM to generate the output text \\( y = \\mathrm{LM}([h_{\\mathrm{img}}; h_{\\mathrm{prompt}}]) \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.137, + 0.905, + 0.272 + ], + "angle": 0, + "content": "In this work, we study the inference acceleration of VLMs where visual tokens are pruned within the attention mechanism of LM. We focus on the FastV [10] approach, where after layer \\( K \\) in LM, \\( R\\% \\) of visual tokens are pruned away based on a ranking function \\( g_{\\phi} \\) which ranks tokens based on criteria \\( \\phi \\). In practice, the attention score received from the last text token is used as the criteria, referred to as \\( \\phi_{\\mathrm{original}} \\). Note that the positional information is preserved when performing the pruning." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.272, + 0.905, + 0.382 + ], + "angle": 0, + "content": "With this approach, we measure the acceleration using the theoretical FLOPS reduction ratio related to the image tokens. For one Transformer layer, the total FLOPS is estimated as \\( C = 4nd^2 + 2n^2d + 2nd_m \\) where \\( d = d_{\\mathrm{text}} \\) and \\( m \\) is the intermediate size of FNN. Given that the Transformer has \\( T \\) layers in total and after layer \\( K \\), we maintain \\( \\hat{n} = (1 - R\\%) * n \\) visual tokens, we calculate the FLOPS reduction as" + }, + { + "type": "equation", + "bbox": [ + 0.552, + 0.386, + 0.905, + 0.415 + ], + "angle": 0, + "content": "\\[\n1 - \\frac {K * C + (T - K) * (4 \\hat {n} d ^ {2} + 2 \\hat {n} ^ {2} d + 2 \\hat {n} d m)}{C}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.418, + 0.905, + 0.493 + ], + "angle": 0, + "content": "Benchmarking. We evaluate the accelerated VLMs on a large suite of benchmarks from [18] which includes evaluations spanning the areas of localization, open-ended visual question answering, and challenge sets. For all benchmarks, we follow the same evaluation protocol as [18]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.494, + 0.905, + 0.628 + ], + "angle": 0, + "content": "Localization. We evaluate localization performance using the RefCOCO, RefCOCO+, RefCOCOg [19, 42] and OCID-Ref [37] datasets. RefCOCO contains short, spatially grounded descriptions, while RefCOCO+ focuses on appearance-based descriptions. RefCOCOg, on the other hand, features longer and more detailed descriptions. OCID-Ref is a robotics dataset that tests generalization to out-of-distribution scenarios, particularly for object localization in cluttered environments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.905, + 0.734 + ], + "angle": 0, + "content": "Open-Ended Visual Question Answering. We evaluate general visual reasoning using the VizWiz [5] and VQAv2 [13] datasets, spatial reasoning using the GQA [16] dataset, and reasoning around text using the TextVQA [32] dataset. For TextVQA, following [18], we exclude OCR-system parsed input tokens to better evaluate the effect of pruning on visual capabilities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.905, + 0.825 + ], + "angle": 0, + "content": "Challenge Sets. We additionally evaluate on the VSR [26], TallyQA [1], POPE [25], and AI2D [20] closed-set prediction datasets. VSR includes binary spatial relationship questions, TallyQA consists of counting questions, POPE probes hallucination, and AI2D contains multiple-choice questions referring to scientific diagrams and charts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Experimental settings. In our experiments, we utilize a VLM with SigLIP ViT-SO400M [43] as \\( f \\), a one-layer MLP with GELU activation as \\( p \\), and Llama 2 7B [35] as LM. The model was trained on the multimodal instruction tuning dataset presented in [27] in a single-stage with \\( f \\) frozen." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22828" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.098, + 0.345, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.347, + 0.099, + 0.603, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.09, + 0.904, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.268, + 0.908, + 0.326 + ], + "angle": 0, + "content": "Figure 2. Contrasting the difference in performance dropoff on the challenging vision-centric localization task (Left) versus the other evaluated tasks (Middle) when pruning visual tokens after the shallow LLM layers. Whereas performance decrease is minimal for most tasks, localization exhibits roughly a linear decrease to zero as the ratio of pruned tokens increases. Right: Per-task performance breakdown across various setups of pruning ratios. Using \\( K = 3 \\) for all pruning setups." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.351, + 0.483, + 0.381 + ], + "angle": 0, + "content": "3.2. Impact of early visual token pruning varies drastically by task" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.388, + 0.483, + 0.448 + ], + "angle": 0, + "content": "We examine the effect of pruning visual tokens after the shallow LLM layers \\((K = 3)\\) across a variety of tasks. We compare various pruning ratios \\((R \\in \\{0.25, 0.5, 0.75, 0.9\\})\\) and the baseline non-pruned model in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.45, + 0.485, + 0.704 + ], + "angle": 0, + "content": "For the majority of evaluated tasks, heavily reducing the number of visual tokens results in minimal performance decrease (3.5% for AI2D, 0.1% for VSR, 3.7% for TallyQA, 4.8% for POPE, 5.1% for VizWiz, 7.9% for VQAv2 and 7.7% for GQA when dropping 75% of tokens). However, for a small subset of tasks—particularly TextVQA, and more severely, localization-performance drops substantially, with decreases of 42.0%, 88.9%, 88.9%, 91.0%, and 86.0% for TextVQA, RefCOCO, RefCOCO+, RefCOCOg, and OCID-Ref, respectively. Investigating further, Figure 2 shows that localization performance declines roughly linearly to zero as the pruning ratio progresses from 0 to 1, a pattern that strongly contrasts with most other tasks, where performance remains largely unaffected. While prior works report a much smaller decline in TextVQA performance [9, 39], they incorporate OCR-parsed input tokens into prompts, reducing the reliance on visual information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Discussion. Although localization is undoubtedly challenging for a model that discards visual tokens, as it requires fine-grained visual grounding and precise boundary identification, it is still surprising that (1) localization performance declines to the extent it does even when many tokens are maintained (e.g., dropping \\(50\\%\\) of tokens reduces average performance by \\(45\\%\\)), and (2) most non-localization benchmarks, except for TextVQA, exhibit a drastically different pattern despite these benchmarks aiming to evaluate visual grounding abilities, such as counting for TallyQA, spatial reasoning for GQA, and chart understanding for AI2D. In the following sections, we seek to explain these unexpected findings and understand their implications." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.351, + 0.907, + 0.38 + ], + "angle": 0, + "content": "3.3. Interpreting poor vision-centric task performance" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.388, + 0.907, + 0.538 + ], + "angle": 0, + "content": "We next investigate why pruning visual tokens after the shallow LLM layers has such an adverse effect on localization performance. Intuitively, even localization should not require all image tokens, as there are background tokens irrelevant to identifying specific objects that can be pruned. Thus, it is surprising that performance begins to degrade even with minimal token pruning. This suggests that the pruning criteria, intended to retain important tokens, fails to effectively distinguish which tokens contain necessary visual information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.909, + 0.719 + ], + "angle": 0, + "content": "To assess the efficacy of the pruning criteria, we first examine the distribution of retained tokens across all benchmark examples. As shown in Figure 3(b), we find that pruning visual tokens after the shallow LLM layers \\((K = 3)\\) retains tokens concentrated at the bottom of the image. For example, when \\(R = 0.75\\), the average sampled token position across all datasets is situated \\(80.7\\%\\) of the way down the image. A Chi-Square test confirms the non-uniformity of selected tokens' y-positions (\\(p\\)-value \\(< 0.05\\)). We also examine the criteria behavior when pruning after later layers (\\(K \\in (8,16,24)\\)) and observe that the bias of selecting bottom image tokens is reduced." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.901 + ], + "angle": 0, + "content": "In addition, we observe that as the pruning layer increases, not only is the positional bias reduced, but the criteria also begins to correctly select tokens relevant to the text instruction. For example, as shown in Figure 3(a), pruning after the later layers results in a distinct concentration of selected tokens around the area corresponding to the reference expression for localization. To quantify this change in criteria effectiveness across layers, we measure the performance when varying the pruning layer \\((K\\in \\{3,8,16,24\\})\\) As shown in Figure 3(c), performance remains stable when pruning is performed at deeper layers \\(K = 16\\) and \\(K =\\) 24), whereas pruning at earlier layers \\(K = 3\\) and \\(K = 8\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22829" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.097, + 0.114, + 0.112 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.122, + 0.091, + 0.237, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.091, + 0.37, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.123, + 0.194, + 0.236, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.194, + 0.37, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.14, + 0.291, + 0.345, + 0.301 + ], + "angle": 0, + "content": "Reference expression: a bowl of blueberries" + }, + { + "type": "image_caption", + "bbox": [ + 0.379, + 0.091, + 0.4, + 0.111 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.402, + 0.091, + 0.537, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.459, + 0.186, + 0.491, + 0.198 + ], + "angle": 0, + "content": "x position K-16" + }, + { + "type": "image", + "bbox": [ + 0.402, + 0.199, + 0.537, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.462, + 0.288, + 0.493, + 0.295 + ], + "angle": 0, + "content": "x position" + }, + { + "type": "image", + "bbox": [ + 0.548, + 0.091, + 0.686, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.186, + 0.64, + 0.199 + ], + "angle": 0, + "content": "x position \\(K = 24\\)" + }, + { + "type": "image", + "bbox": [ + 0.549, + 0.2, + 0.686, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.288, + 0.641, + 0.295 + ], + "angle": 0, + "content": "x position" + }, + { + "type": "image", + "bbox": [ + 0.465, + 0.298, + 0.639, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.705, + 0.098, + 0.722, + 0.11 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.724, + 0.093, + 0.903, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.791, + 0.208, + 0.859, + 0.216 + ], + "angle": 0, + "content": "Non-localization" + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.216, + 0.904, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.336, + 0.907, + 0.42 + ], + "angle": 0, + "content": "Figure 3. (a) Example demonstrating that when pruning in early layers, selected tokens are concentrated in the bottom of the image, whereas in later layers the selected tokens more precisely cover the region of the image important for answering the question. (b) Heatmap illustrating that averaged across all benchmark examples, as the pruning layer increases, the selection of bottom visual tokens by the criteria is reduced. (c) Visualizing the effect of the pruning layer on performance for both localization and non-localization tasks. We find that pruning after layer 16 or later results in little performance dropoff, whereas pruning earlier results in a performance decrease, particularly for localization. Using \\( R = 0.75 \\) for all setups." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.447, + 0.483, + 0.476 + ], + "angle": 0, + "content": "leads to performance degradation, particularly for localization." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.478, + 0.484, + 0.809 + ], + "angle": 0, + "content": "Discussion. While previous research attributes the poor performance of early token pruning to the low redundancy of image tokens in shallow layers [39], we reveal a critical insight: the variable performance across pruning layers is linked to the effectiveness of selecting important tokens, as early token pruning leads to suboptimal token selection biased toward bottom tokens. To understand why the criteria disproportionately retains bottom image tokens, we must consider the LM architecture. Specifically, token selection is based on the attention received from the last text token. Since the LM uses RoPE to encode positional information [33], it exhibits a long-term decay property. While this property is well-suited for language modeling, we uncover that it poses an issue for selecting visual tokens, where tokens appearing later in the flattened, raster-scan order have higher attention scores purely due to their position. In language modeling, prior work finds that there is greater emphasis on shorter-distance information in shallow layers [15]. We predict that the visually-conditioned LM exhibits the same behavior, explaining why the positional bias is most pronounced when pruning after shallow layers." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.819, + 0.48, + 0.835 + ], + "angle": 0, + "content": "3.4. Explaining strong performance on other tasks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Given that early pruning of visual tokens leads to a suboptimal tokens selection where remaining tokens are concentrated towards the bottom of the image, a natural question arises: how does the approach still maintain high perfor" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.447, + 0.9, + 0.655 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.67, + 0.907, + 0.782 + ], + "angle": 0, + "content": "Figure 4. Assessing whether the strong performance of early visual token pruning for many tasks can be attributed to visual information transfer before pruning or benchmarks' lack of assessing fine-grained visual capabilities. We observe that for all evaluated tasks, allowing information transfer before pruning (shown in green) does not result in substantial performance improvement over the setup without visual information transfer (shown in light green), highlighting a limitation of many benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.808, + 0.906, + 0.854 + ], + "angle": 0, + "content": "mance across a diverse range of tasks, including those aiming to evaluate visual grounding abilities? In this section, we test the following contrasting explanations:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.856, + 0.907, + 0.9 + ], + "angle": 0, + "content": "1. Important visual information from top image tokens is transferred to later tokens before pruning with the unidirectional attention of the LLM." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22830" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "2. Many questions can still be inferred with access to only the suboptimal set of visual tokens primarily from the bottom of the image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.139, + 0.483, + 0.305 + ], + "angle": 0, + "content": "To evaluate these hypotheses across tasks, we save the tokens retained by the studied pruning approach and, in a subsequent run, remove all other tokens before they reach the LLM, ensuring that no visual information from the pruned tokens is retained. As an additional comparison, we also evaluate a text-only model in which all visual tokens are removed before entering the LLM. As shown in Figure 4, we find that across all evaluated tasks, removing pruned tokens before the LLM results in little to no performance decrease compared to the method that allows information transfer in shallow layers prior to pruning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.306, + 0.483, + 0.471 + ], + "angle": 0, + "content": "Discussion. This finding indicates that the high performance of early visual token pruning on many tasks does not stem from the effectiveness of visual information transfer in early layers but rather signifies that many benchmarks do not demand a detailed understanding of visual information to answer questions accurately. Note that for the majority of these benchmarks (except VizWiz and AI2D), performance substantially declines when not including any visual information, signifying that simply comparing vision-enabled and disabled setups is insufficient to determine a benchmark's ability to assess visual grounding." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.483, + 0.392, + 0.501 + ], + "angle": 0, + "content": "4. Improving Visual Token Pruning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.509, + 0.483, + 0.675 + ], + "angle": 0, + "content": "Based on our findings from §3, we now seek to answer: can attention serve as an effective criteria for token pruning even in early layers? To this end, we propose a straightforward solution to removing the positional bias of the attention-based criteria and demonstrate that our modifications enable effective pruning even when applied after early LLM layers (§4.1). Guided by our insights, we present our final approach, FEATHER (Fast and Effective Acceleration with Ensemble cRicerca) (§4.2), and showcase its impressive visual capabilities while offering high computational efficiency (§4.3)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.683, + 0.476, + 0.699 + ], + "angle": 0, + "content": "4.1. Assessing potential of attention-based criteria" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.84 + ], + "angle": 0, + "content": "An attention-based criteria offers the potential to naturally identify and retain only the most relevant visual tokens for answering a question. However, as shown in §3, positional bias strongly influences token selection, limiting its effectiveness. Thus, we aim to resolve the positional bias issue and explore the true promise of attention-based criteria. We compare against non-attention-based approaches that incorporate an explicit inductive bias toward reducing redundancy. Below, we outline the explored criteria." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.483, + 0.902 + ], + "angle": 0, + "content": "\\(\\phi_{-R}\\): Given the long-term decay property of RoPE, we propose the following straightforward adjustment to the pruning criteria in order to mitigate this bias and more effectively select tokens relevant to the text instruction. Namely," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "we still compute the attention score received from the last text token, except we do not apply RoPE to the attention mechanism, thereby removing the long-term decay effect. Note that this lightweight calculation is only done once per pruning, and all other state computations are compatible with FlashAttention [12]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.183, + 0.905, + 0.301 + ], + "angle": 0, + "content": "\\(\\phi_{\\mathrm{uniform}}\\): In this simple criteria, we uniformly sample visual tokens in the image, using a set stride. This criteria ensures that there is good image coverage, but does so by sacrificing the ability to have more densely captured visual information for a particular image region. To use a similar computational saving to the attention-based criteria when \\(R = 0.75\\), we use a stride of two, resulting in 196 selected tokens." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.302, + 0.905, + 0.372 + ], + "angle": 0, + "content": "\\(\\phi_{\\mathrm{KNN}}\\): Inspired by dynamic visual tokenization [17], we select tokens based on local density. For each visual token \\(z_{img}[i]\\) where \\(i \\in \\{0, 1, \\dots, n-1\\}\\), we compute its local density \\(\\rho_i\\) using K-nearest neighbors and calculate distance index \\(\\delta_i\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.38, + 0.905, + 0.425 + ], + "angle": 0, + "content": "\\[\n\\delta_ {i} = \\left\\{ \\begin{array}{l l} \\min _ {j: \\rho_ {j} > \\rho_ {i}} \\| z _ {i m g} [ i ] - z _ {i m g} [ j ] \\| ^ {2}, & \\text {i f} \\exists j \\text {s . t .} \\rho_ {j} > \\rho_ {i}. \\\\ \\max _ {j} \\| z _ {i m g} [ i ] - z _ {i m g} [ j ] \\| ^ {2}, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.433, + 0.905, + 0.478 + ], + "angle": 0, + "content": "\\(\\delta_{i}\\) represents how far away \\(z_{img}[i]\\) is from other high-density tokens. We use \\(\\rho_{i}*\\delta_{i}\\) as the token importance score for the criteria and select 196 tokens." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.479, + 0.905, + 0.567 + ], + "angle": 0, + "content": "\\(\\phi_{-R} + \\phi_{\\mathrm{uniform}}\\): Since \\(\\phi_{-R}\\) focuses on selecting important tokens and \\(\\phi_{\\mathrm{uniform}}\\) aims to reduce redundancy, we propose an ensemble of these two to leverage their respective strengths. Specifically, we apply \\(\\phi_{-R}\\) while incorporating a small number of uniformly sampled tokens using a stride of three." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.569, + 0.905, + 0.764 + ], + "angle": 0, + "content": "Improvement of removing RoPE: In Table 1, we report results evaluating the effectiveness of these criteria. Comparing \\(\\phi_{-R}\\) with \\(\\phi_{\\mathrm{original}}\\), we find that by removing RoPE, at \\(K = 3\\) there is a \\(183\\%\\) average improvement on localization tasks and at \\(K = 8\\), there is a \\(17\\%\\) average improvement. These performance improvements demonstrate that once the impact of token position on attention is removed, attention score can more effectively be used for the criteria when applied after early LLM layers. Note that the narrowing gap in performance between \\(\\phi_{-R}\\) and \\(\\phi_{\\mathrm{original}}\\) from \\(K = 3\\) to \\(K = 8\\) is likely because \\(\\phi_{\\mathrm{original}}\\) has less of a bias towards selecting bottom image tokens as the pruning layer increases." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Token selection improves when pruning later: Comparing \\(\\phi_{-R}\\) when \\(K = 8\\) versus \\(K = 3\\), we see that the average localization performance improves by \\(63\\%\\). However, it remains unclear whether this performance increase is affected by factors other than the ability of the criteria to select important tokens. Therefore, to directly compare the criteria across different layers, we take the tokens selected by both criteria and evaluate how well the model can perform with only these tokens passed into the LLM (as is" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22831" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.087, + 0.907, + 0.46 + ], + "angle": 0, + "content": "
Pruning LayerCriteriaFLOPS RedLocalizationOpen-Ended VQAChallenge Sets
AvgOCID-RefRefCOCObRefCOCO+RefCOCOAvgTextVQAGQAVQAv2VizWizAvgPOPETallyQAVSRAI2D
K=3Attention-based
φoriginal68%5.95.75.16.16.754.831.858.472.756.364.083.257.163.352.4
φR(Ours)68%16.722.915.113.315.359.041.661.276.057.364.785.258.262.253.2
Δ+10.7+17.2+10.0+7.3+8.6+4.2+9.7+2.8+3.2+1.1+0.7+1.9+1.1-1.1+0.8
Non-attention-based
φKNN66%23.915.124.926.029.658.439.960.974.458.462.881.255.961.552.8
φuniform66%28.020.628.629.733.359.041.461.875.957.164.685.258.161.953.0
Ensemble
φR+φuniform(Ours)61%27.229.127.224.727.761.246.662.377.458.465.486.058.962.754.0
K=8Attention-based
φoriginal56%23.319.423.524.026.359.845.060.376.157.864.685.457.562.653.0
φR(Ours)56%27.327.126.726.429.261.449.061.577.457.865.586.758.663.053.7
Δ+4.0+7.6+3.2+2.5+2.9+1.6+4.0+1.2+1.3+0.0+0.8+1.3+1.1+0.4+0.6
Non-attention-based
φKNN55%23.615.424.425.229.458.640.261.174.558.562.981.456.260.953.0
φuniform55%30.324.631.030.934.859.342.261.876.057.464.485.357.961.053.2
Ensemble
φR+φuniform(Ours)50%35.632.035.935.438.862.751.762.478.158.666.087.459.163.654.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.468, + 0.907, + 0.54 + ], + "angle": 0, + "content": "Table 1. Evaluating alternative criteria for token pruning after the early LLM layers. For each task and pruning layer, we **bold** the best result and **underline** the second-best result. Our main findings include: (1) our proposed RoPE-free criteria \\(\\phi_{-R}\\) substantially improves pruning performance compared to the original criteria \\(\\phi_{\\mathrm{original}}\\); (2) pruning later (\\(K = 8\\)) yields higher performance than pruning earlier (\\(K = 3\\)); and (3) integrating uniform sampling into the attention-based criteria with \\(\\phi_{-R} + \\phi_{\\mathrm{uniform}}\\) enhances effectiveness. Using \\(R = 0.75\\) for \\(\\phi_{\\mathrm{original}}\\), \\(\\phi_{-R}\\), and \\(\\phi_{-R} + \\phi_{\\mathrm{uniform}}\\). See §4.1 for criteria definitions." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.561, + 0.477, + 0.622 + ], + "angle": 0, + "content": "
KOCID-RefRefCOCOgRefCOCO+RefCOCO
323.816.214.516.3
826.729.826.829.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.631, + 0.483, + 0.675 + ], + "angle": 0, + "content": "Table 2. Assessing localization performance when using the selected tokens from \\(\\phi_{-R}\\) for the entirety of the LLM. We find that \\(\\phi_{-R}\\) applied at a later LLM layer results in a better token selection." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.703, + 0.483, + 0.748 + ], + "angle": 0, + "content": "done in §3.4). As shown in Table 2, we find the selected tokens from \\(\\phi_{-R}\\) applied after layer 8 are superior compared to the tokens from \\(\\phi_{-R}\\) applied after layer 3." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Integrating uniform sampling with attention-based criteria is beneficial in early layers: When comparing the attention-based criteria \\(\\phi_{-R}\\) with the two non-attention based criteria \\(\\phi_{\\mathrm{KNN}}\\) and \\(\\phi_{\\mathrm{uniform}}\\), we find that \\(\\phi_{-R}\\) results in better performance on some tasks (e.g., OCID-Ref, TextVQA), while \\(\\phi_{\\mathrm{KNN}}\\) and \\(\\phi_{\\mathrm{uniform}}\\) outperform on other tasks (e.g., RefCOCO). This varying effectiveness of criteria types presumably comes from some tasks requiring a more detailed understanding of specific image regions while others benefit from a broader understanding with full image" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.907, + 0.716 + ], + "angle": 0, + "content": "coverage. When we combine these two criteria types with \\(\\phi_{-R} + \\phi_{\\mathrm{uniform}}\\), we find that when \\(K = 3\\) this approach improves upon localization performance of \\(\\phi_{-R}\\) by \\(63\\%\\) but has a slight decrease compared to \\(\\phi_{\\mathrm{uniform}}\\) by \\(2.9\\%\\). However, when \\(K = 8\\), this approach far outperforms both of the individual criteria, improving \\(\\phi_{-R}\\) by \\(30\\%\\) and \\(\\phi_{\\mathrm{uniform}}\\) by \\(17\\%\\). Note that this setup results in slightly less computation efficiency than \\(\\phi_{-R}\\) (7% drop in FLOPS reduction for \\(K = 3\\) and \\(6\\%\\) for \\(K = 16\\)), but this is outweighed by the performance improvements." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.717, + 0.906, + 0.763 + ], + "angle": 0, + "content": "We provide qualitative examples of various criteria token selection and additional results using a different model setup in the supplement." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.772, + 0.685, + 0.789 + ], + "angle": 0, + "content": "4.2. Distilling insights" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Guided by our insights, we now present our final approach FEATHER (Fast and Effective Acceleration wiTH Ensemble cRicerca). In this approach, we first perform pruning after an early layer \\((K = 8)\\), utilizing our proposed criteria \\(\\phi_{-R} + \\phi_{\\mathrm{uniform}}\\) to retain \\((1 - R)\\%\\) of tokens. Given our finding that the criteria improves as the LLM layer increases, we additionally prune a second time at \\(K = 16\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22832" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.09, + 0.91, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.27, + 0.908, + 0.3 + ], + "angle": 0, + "content": "Figure 5. Comparing FEATHER performance against FastV [10] and PyramidDrop [39]. We find that FEATHER far outperforms both compared methods, particularly for the vision-centric task of localization." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.326, + 0.482, + 0.432 + ], + "angle": 0, + "content": "At this stage, we utilize \\(\\phi_{-R}\\) as the criteria, as uniform sampling should not be necessary when the attention-based criteria reliably selects important tokens. For the reduction ratio, we choose to only retain \\((1 - R)^{2\\%}\\) of the remaining tokens since the attention-based criteria has proved highly effective when pruning at later layers, even when using the original criteria \\(\\phi_{\\mathrm{original}}\\) (see Figure 3(c))." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.448, + 0.476, + 0.466 + ], + "angle": 0, + "content": "4.3. Comparison against FastV and PyramidDrop" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.473, + 0.483, + 0.534 + ], + "angle": 0, + "content": "We compare our approach against the one-stage early pruning approach of FastV (\\(K = 3\\)) [10] and the multi-stage approach of PyramidDrop (\\(K = [8, 16, 24]\\)) [39]. Note that both of these methods use \\(\\phi_{\\mathrm{original}}\\) for the criteria." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.536, + 0.483, + 0.763 + ], + "angle": 0, + "content": "As shown in Figure 5, we find that FEATHER far outperforms the baselines, especially on the localization tasks. Specifically, for comparable computational costs (64% FLOPS reduction for our approach, 68% FLOPS reduction for FastV, and 65% FLOPS reduction for PyramidDrop), we observe that for localization tasks, FEATHER exhibits more than \\(5 \\times\\) average performance improvement compared to FastV and a \\(36\\%\\) average performance improvement compared to PyramidDrop. For non-localization tasks, FEATHER has a \\(7.8\\%\\) improvement over FastV and \\(1.5\\%\\) improvement over PyramidDrop. Note that PyramidDrop performs substantially better than FastV as it prunes fewer tokens in an early layer. However, it still suffers from an ineffective pruning strategy at this stage, though the impact is less pronounced since it predominantly prunes later." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Remarkably, with our \\(64\\%\\) FLOPS reduction setup, after layer 16 only \\(3.3\\%\\) of tokens are retained, yet the average localization performance decrease compared to the baseline method with no token pruning is only \\(26\\%\\) (Figure 1 includes an example of retained tokens after layer 16). This finding illustrates that even for vision-centric tasks, maintaining strong performance while gaining huge acceleration speedups with extensive pruning is possible, but it heavily relies on the effectiveness of the pruning criteria." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.324, + 0.634, + 0.34 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.349, + 0.907, + 0.621 + ], + "angle": 0, + "content": "In this work, we examine the visual capabilities of the VLM acceleration approach of pruning visual tokens after shallow LLM layers. While strong performance is maintained on most evaluated tasks, it fails on more vision-centric tasks like TextVQA and localization due to its flawed pruning criteria that predominately selects visual tokens from the bottom part of the image. Observing this same flawed criteria on other tasks, we show that strong performance is largely due to the benchmarks' inability to assess fine-grained visual capabilities. Next, we propose and evaluate several alternative criteria to improve visual capabilities, ultimately arriving at our final method, FEATHER. This approach refines the attention-based criteria to address the token selection bias while using uniform sampling for better image coverage. It then prunes more aggressively when the criteria more effectively identifies important tokens. We find that FEATHER has more than \\(5 \\times\\) performance improvement on localization compared to the original acceleration approach." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.621, + 0.909, + 0.893 + ], + "angle": 0, + "content": "While we study a particular type of VLM acceleration, our work highlights a broad challenge in evaluating VLMs. Building on [34], which shows that text-only models can perform well on some vision-language benchmarks, we demonstrate that even benchmarks with substantial differences between vision-enabled and disabled setups may not assess fine-grained visual capabilities. We address this issue by focusing on more vision-centric localization tasks, though this limits the analysis to a few specific types of skills. To accurately assess a wider range of visual capabilities, future work may explore how to resolve current dataset biases that models can exploit [21]. Additionally, while removing RoPE for token pruning effectively eliminates positional bias, it may introduce unintended effects on attention weights influencing token selection. Future research could investigate more robust methods for encoding positional information in visually conditioned language models to prevent positional artifacts in cross-modal interactions." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "22833" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.218 + ], + "angle": 0, + "content": "Acknowledgments. This work is supported in part by the National Science Foundation (NSF) under Grant No. 2026498, the Stanford AIMI-HAI Partnership Grant, and the NSF Graduate Research Fellowship Program under Grant No. DGE-2146755 (for M.E.). Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of any other entity." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.234, + 0.188, + 0.249 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.259, + 0.482, + 0.314 + ], + "angle": 0, + "content": "[1] Manoj Acharya, Kushal Kafle, and Christopher Kanan. Tallyqa: Answering complex counting questions. In Proceedings of the AAAI conference on artificial intelligence, pages 8076-8084, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.317, + 0.482, + 0.398 + ], + "angle": 0, + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.402, + 0.482, + 0.483 + ], + "angle": 0, + "content": "[3] Kazi Hasan Ibn Arif, JinYi Yoon, Dimitrios S Nikolopoulos, Hans Vandierendonck, Deepu John, and Bo Ji. Hired: Attention-guided token dropping for efficient inference of high-resolution vision-language models in resource-constrained environments. arXiv preprint arXiv:2408.10945, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.487, + 0.482, + 0.542 + ], + "angle": 0, + "content": "[4] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.545, + 0.482, + 0.627 + ], + "angle": 0, + "content": "[5] Jeffrey P Bigham, Chandrika Jayant, Hanjie Ji, Greg Little, Andrew Miller, Robert C Miller, Robin Miller, Aubrey Tatarowicz, Brandyn White, Samual White, et al. Vizwiz: nearly real-time answers to visual questions. In Proceedings of the 23nd annual ACM symposium on User interface software and technology, pages 333-342, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.63, + 0.482, + 0.685 + ], + "angle": 0, + "content": "[6] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. In The Eleventh International Conference on Learning Representations, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.688, + 0.482, + 0.728 + ], + "angle": 0, + "content": "[7] Mu Cai, Jianwei Yang, Jianfeng Gao, and Yong Jae Lee. Matryoshka multimodal models. arXiv preprint arXiv:2405.17430, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.731, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[8] Wenhao Chai, Enxin Song, Yilun Du, Chenlin Meng, Vashisht Madhavan, Omer Bar-Tal, Jeng-Neng Hwang, Saining Xie, and Christopher D Manning. Auroracap: Efficient, performant video detailed captioning and a new benchmark. arXiv preprint arXiv:2410.03051, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.803, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[9] Jieneng Chen, Luoxin Ye, Ju He, Zhao-Yang Wang, Daniel Khashabi, and Alan Yuille. Efficient large multi-modal models via visual context compression. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[10] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.259, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. ECCV, 2024. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[11] Yi Chen, Jian Xu, Xu-Yao Zhang, Wen-Zhuo Liu, YangYang Liu, and Cheng-Lin Liu. Recoverable compression: A multimodal vision token recovery mechanism guided by text information. arXiv preprint arXiv:2409.01179, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[12] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in neural information processing systems, 35:16344-16359, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.234, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[13] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[14] Yefei He, Feng Chen, Jing Liu, Wenqi Shao, Hong Zhou, Kaipeng Zhang, and Bohan Zhuang. Zipvl: Efficient large vision-language models with dynamic token sparsification and kv cache compression. arXiv preprint arXiv:2410.08584, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.373, + 0.905, + 0.427 + ], + "angle": 0, + "content": "[15] Xiangyu Hong, Che Jiang, Biqing Qi, Fandong Meng, Mo Yu, Bowen Zhou, and Jie Zhou. On the token distance modeling ability of higher rope attention dimension. arXiv preprint arXiv:2410.08703, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.429, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[16] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.498, + 0.905, + 0.579 + ], + "angle": 0, + "content": "[17] Peng Jin, Ryuichi Takanobu, Wancai Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13700-13710, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.582, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[18] Siddharth Karamcheti, Suraj Nair, Ashwin Balakrishna, Percy Liang, Thomas Kollar, and Dorsa Sadigh. Prismatic VLMs: Investigating the design space of visually-conditioned language models. In Proceedings of the 41st International Conference on Machine Learning, pages 23123-23144. PMLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.666, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[19] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.735, + 0.905, + 0.817 + ], + "angle": 0, + "content": "[20] Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Minjoon Seo, Hannaneh Hajishirzi, and Ali Farhadi. A diagram is worth a dozen images. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 235-251. Springer, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.818, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[21] Baiqi Li, Zhiqiu Lin, Wenxuan Peng, Jean de Dieu Nyandwi, Daniel Jiang, Zixian Ma, Simran Khanuja, Ranjay Krishna, Graham Neubig, and Deva Ramanan. Naturalbench: Evaluating vision-language models on natural adversarial samples. Advances in Neural Information Processing Systems, 37:17044-17068, 2025. 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22834" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[22] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[23] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.288 + ], + "angle": 0, + "content": "[24] Kevin Y Li, Sachin Goyal, Joao D Semedo, and J Zico Kolter. Inference optimal vlms need only one visual token but larger models. arXiv preprint arXiv:2411.03312, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[25] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[26] Fangyu Liu, Guy Emerson, and Nigel Collier. Visual spatial reasoning. Transactions of the Association for Computational Linguistics, 11:635-651, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.446 + ], + "angle": 0, + "content": "[27] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.482, + 0.49 + ], + "angle": 0, + "content": "[28] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.533 + ], + "angle": 0, + "content": "[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.534, + 0.482, + 0.603 + ], + "angle": 0, + "content": "[30] Yifei Liu, Mathias Gehrig, Nico Messikommer, Marco Cannici, and Davide Scaramuzza. Revisiting token pruning for object detection and instance segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2658-2668, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.658 + ], + "angle": 0, + "content": "[31] Yuzhang Shang, Mu Cai, Bingxin Xu, Yong Jae Lee, and Yan Yan. Llava-prumerge: Adaptive token reduction for efficient large multimodal models. arXiv preprint arXiv:2403.15388, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[32] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[33] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[34] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, et al. Cambrian1: A fully open, vision-centric exploration of multimodal llms. arXiv preprint arXiv:2406.16860, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[35] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[36] Zhongwei Wan, Ziang Wu, Che Liu, Jina Huang, Zhihong Zhu, Peng Jin, Longyue Wang, and Li Yuan. Look-m: Look-once optimization in kv cache for efficient multimodal long-context inference. arXiv preprint arXiv:2406.18139, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[37] Ke-Jyun Wang, Yun-Hsuan Liu, Hung-Ting Su, Jen-Wei Wang, Yu-Siang Wang, Winston Hsu, and Wen-Chin Chen. OCID-ref: A 3D robotic dataset with embodied language for clutter scene grounding. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5333–5338, Online, 2021. Association for Computational Linguistics. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.905, + 0.373 + ], + "angle": 0, + "content": "[38] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.374, + 0.905, + 0.457 + ], + "angle": 0, + "content": "[39] Long Xing, Qidong Huang, Xiaoyi Dong, Jiajie Lu, Pan Zhang, Yuhang Zang, Yuhang Cao, Conghui He, Jiaqi Wang, Feng Wu, et al. Pyramiddrop: Accelerating your large vision-language models via pyramid visual redundancy reduction. arXiv preprint arXiv:2410.17247, 2024. 2, 3, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.459, + 0.905, + 0.529 + ], + "angle": 0, + "content": "[40] Senqiao Yang, Yukang Chen, Zhuotao Tian, Chengyao Wang, Jingyao Li, Bei Yu, and Jiaya Jia. Visionzip: Longer is better but not necessary in vision language models. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 19792-19802, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.53, + 0.905, + 0.584 + ], + "angle": 0, + "content": "[41] Gaotong Yu, Yi Chen, and Jian Xu. Balancing performance and efficiency: A multimodal large language model pruning method based image text interaction. arXiv preprint arXiv:2409.01162, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.585, + 0.905, + 0.668 + ], + "angle": 0, + "content": "[42] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.67, + 0.905, + 0.726 + ], + "angle": 0, + "content": "[43] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.727, + 0.905, + 0.797 + ], + "angle": 0, + "content": "[44] Qizhe Zhang, Aosong Cheng, Ming Lu, Zhiyong Zhuo, Minqi Wang, Jiajun Cao, Shaobo Guo, Qi She, and Shang-hang Zhang. [cls] attention is all you need for training-free visual token pruning: Make vlm inference faster. arXiv eprints, pages arXiv-2412, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.797 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22835" + } + ] +] \ No newline at end of file diff --git a/2025/Feather the Throttle_ Revisiting Visual Token Pruning for Vision-Language Model Acceleration/eb2a11ee-50f9-4c43-8cc0-429a524795eb_origin.pdf b/2025/Feather the Throttle_ Revisiting Visual Token Pruning for Vision-Language Model Acceleration/eb2a11ee-50f9-4c43-8cc0-429a524795eb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..45c46720dc5dce368b980833c74a50c179104443 --- /dev/null +++ b/2025/Feather the Throttle_ Revisiting Visual Token Pruning for Vision-Language Model Acceleration/eb2a11ee-50f9-4c43-8cc0-429a524795eb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e642cc001aae9b91f01cb76619c8bc36c9dbcc714539aeab5869b4916bd91476 +size 1302013 diff --git a/2025/Feather the Throttle_ Revisiting Visual Token Pruning for Vision-Language Model Acceleration/full.md b/2025/Feather the Throttle_ Revisiting Visual Token Pruning for Vision-Language Model Acceleration/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cdb439c575bdfc5155ad86855963ee2489efb357 --- /dev/null +++ b/2025/Feather the Throttle_ Revisiting Visual Token Pruning for Vision-Language Model Acceleration/full.md @@ -0,0 +1,273 @@ +# Feather the Throttle: Revisiting Visual Token Pruning for Vision-Language Model Acceleration + +Mark Endo, Xiaohan Wang, Serena Yeung-Levy +Stanford University + +{markendo,xhanwang,syyeung}@stanford.edu + +# Abstract + +Recent works on accelerating Vision-Language Models achieve strong performance across a variety of vision-language tasks despite highly compressing visual information. In this work, we examine the popular acceleration approach of early pruning of visual tokens inside the language model. Surprisingly, we find that while strong performance is maintained across many tasks, it exhibits drastically different behavior for a subset of vision-centric tasks such as localization. Upon further investigation, we uncover a core issue with the acceleration approach where most tokens towards the top of the image are pruned away. Yet, on many benchmarks aiming to evaluate vision-centric capabilities, strong performance persists with the flawed pruning strategy, highlighting these benchmarks' limited ability to assess fine-grained visual capabilities. Based on these findings, we propose FEATHER (Fast and Effective Acceleration with Ensemble cRicerca), a straightforward approach that resolves the discovered early-layer pruning issue and further enhances the preservation of relevant tokens via multistage pruning with early uniform sampling to ensure broad image coverage. With comparable computational savings, we find that FEATHER achieves more than $5 \times$ performance improvement on the vision-centric localization benchmarks compared to the original acceleration approach. + +# 1. Introduction + +The exploration of Vision-Language Models (VLMs) is a critical area of computer vision and natural language processing research, centered on combining large language models (LLMs) with visual encoders to enable multimodal perception, reasoning, and understanding capabilities. While earlier works explored sophisticated schemes for conditioning language models with visual information [2, 22, 23], more recently the space has shifted to predominately using the simplistic approach of taking patch features from pre-trained visual encoders and projecting them + +![](images/4d43fa273c72025972152408e9df7d7b3aeef6ba8cfda68fe40c567bda3e264b.jpg) +(a) + +![](images/d8fc2f3288d4e37f83c664b0544d201797876a8f1b938aefe429ed16c63a5d47.jpg) +(b) +Figure 1. (a) Although FastV prunes most visual tokens from the upper portion of the image, the approach still displays strong performance on a variety of evaluated vision-language tasks except for a small subset of vision-centric tasks like localization. (b) Based on our findings, we propose FEATHER (Fast and Effective Acceleration wiTH Ensemble cRicerca), a straightforward approach that resolves the existing issue of selecting bottom tokens, additionally maintains uniformly sampled tokens to ensure good coverage over the whole image, and prunes in two stages. + +to the input space of the language model with a light-weight adapter [4, 27, 29, 38]. Using image patches as tokens, however, comes with the drawback of being computationally inefficient. To achieve fine-grained resolution, the im + +age is divided into many patches. This large number of patches significantly increases computational demands due to the quadratic complexity of the attention operation in Transformers. As a result, many recent works have focused on accelerating these methods by compressing visual information, demonstrating that heavy compression can still maintain strong performance across a wide variety of tasks [3, 7, 10, 31, 39]. For instance, FastV [10] prunes $50\%$ of visual tokens after the shallow layers of the Language Model (LLM) while not compromising in performance across a range of image and video understanding tasks [10]. Another work reveals that for a fixed computational budget on visual reasoning tasks, optimal performance is achieved by using the largest LLM possible while sacrificing visual information, often reducing the visual token count to a single token [24]. With such compressed visual information, it is still unclear how these methods achieve high performance on tasks assessing vision capabilities such as visual reasoning and understanding, and whether there are more demanding visual tasks where these methods fail. + +The motivation of our study is to get a better understanding of the vision capabilities of accelerated VLMs given that they leverage highly compressed visual information, focusing specifically on FastV approach. Evaluating across a wide range of vision-language tasks, we find that while the approach maintains strong performance across many tasks, there is a substantial decrease in performance for TextVQA and a severe decrease for localization tasks. When analyzing the approach's poor performance, we uncover a fundamental issue with the pruning criteria in early layers where tokens towards the top of the image are disproportionally removed due to the long-term decay property of RoPE. Since this issue is not task-specific, we examine how performance is maintained across most evaluated tasks and find that even when tokens are entirely discarded using the flawed criteria, performance remains largely unchanged. This highlights how many benchmarks require minimal visual grounding. Our finding underscores a significant challenge in the field of multimodal learning, not only for measuring the effectiveness of VLM acceleration methods but also for benchmarking the visual capabilities of VLMs as a whole. + +Given the discovered limitation of the studied approach, we next explore whether the attention-based criteria can be modified to enable robust visual token pruning, leading to our final method, FEATHER (Fast and Effective Acceleration with Ensemble Ricerca). Specifically, we propose a straightforward modification to the attention-based criteria by removing RoPE, demonstrating that this resolves the positional bias issue and greatly improves performance. Additionally, we incorporate uniform sampling in early token pruning to ensure adequate coverage of all image regions and apply more extensive pruning at a later layer, when the attention-based criteria better discerns im + +important tokens. This strategy is analogous to how a race-car driver feathers the throttle by gradually pressing the accelerator at the beginning of a turn to maintain grip and then accelerating more aggressively once the car is past the apex. We show that FEATHER results in substantial performance gains compared to the original acceleration approach, improving localization performance more than $5 \times 5$ with comparable computational savings. Strikingly, we find that our approach achieves this performance improvement while only retaining $3.3\%$ of visual tokens for the second half of LLM layers. Overall, our work demonstrates that while visual compression can maintain strong performance even on challenging vision-centric tasks, its effectiveness depends on a well-designed strategy, which is currently difficult to assess due to many vision-language benchmarks not thoroughly evaluating vision capabilities. + +In summary, our contributions include: + +- Revealing inconsistent effect of early pruning: Our evaluations across 12 benchmarks show that early visual token pruning minimally affects performance on most tasks, but causes a substantial dropoff for TextVQA and a drastic decline for localization tasks. +- Analyzing poor performance: We uncover that the attention-based criteria for selecting important tokens exhibits a bias towards bottom image tokens in early layers due to the long-term decay property of RoPE. +- Identifying benchmark limitations in assessing token pruning: We find that performance remains strong even with suboptimal pruning criteria and removal of pruned tokens before transferring information in the LLM, demonstrating how many benchmarks fail to assess fine-grained visual capabilities. +- Improving attention-based criteria: With the straightforward modification of removing RoPE, we demonstrate that attention can be used as an effective criteria for selecting tokens and performance can be further improved with simple strategies like incorporating uniform sampling and pruning more extensively at a later layer, when the criteria better discerns important tokens. + +# 2. Related Work + +Recent efforts to accelerate VLMs can be broadly divided into two main categories: compressing visual information before it enters the LLM, and compressing visual information within the LLM itself. In the first category, ChatUniVi [17] dynamically merges visual tokens with similar semantic meanings. PruMerge [31], FasterVLM [44], and VisionZip [40] select important tokens based on ViT attention ([CLS] or average attention received). PruMerge then merges important tokens with unselected ones while VisionZip merges unselected tokens together. For the LLaVA-NeXT [28] approach of partitioning an image into + +sub-images where inefficiency is an even bigger problem, HiRED [3] selects tokens with top feature important on each sub-image with an allocated budget. Other methods argue that the input image alone does not include enough information to select important patches and thus use the textual input to recover visually meaningful tokens [11, 41]. Alternatively, some approaches reduce visual information within the ViT by merging or pruning tokens [6, 8, 30]. + +For the second category of approaches, where visual information is compressed within the LLM itself, LOOK-M reduces the multimodal KV cache size [36]. In our study, we focus on the popular FastV approach [10]. This work identifies that attention over image tokens is sparse in deeper layers of the LLM. Based on this observation, they propose to prune away unimportant vision tokens after the shallow layers of the LLM, achieving a $45\%$ reduction in FLOPS with nearly no performance loss. Works since have proposed alterations to this setup such as adaptively determining the number of pruned tokens instead of using a fixed ratio [14] or pruning in multiple stages [39]. + +# 3. Revising Visual Token Pruning + +In this section, we aim to get a better understanding of the vision capabilities of the VLM acceleration approach of pruning visual tokens after shallow LLM layers. After outlining preliminaries (§3.1), we take a closer look at how the approach performs across a broad range of tasks. Upon inspection, we discover that while heavily pruning visual tokens after the shallow LLM layers has little effect on performance across a variety of tasks, this approach fails decisively on more vision-centric tasks, particularly localization (§3.2). Next, we examine why this method struggles with localization, uncovering that the poor performance is due to the ineffectiveness of the pruning criteria to select important tokens when applied at an early layer (§3.3). As this defect is not specific to localization, we explore what can be attributed to the method's high performance on numerous other tasks and find that these tasks require minimal visual grounding (§3.4). + +# 3.1. Preliminaries + +Before our analyses, we provide a background on the explored adapter-style VLM with visual token pruning, the evaluated benchmarks, and experimental settings. + +VLM and token pruning. Formally, an adapter-style VLM takes as input an image $x_{\mathrm{img}}$ and text prompt tokens $x_{\mathrm{prompt}}$ . A pre-trained vision backbone $f$ first encodes visual features $z_{\mathrm{img}} = f(x_{\mathrm{img}}) \in \mathbb{R}^{n \times d_{\mathrm{vision}}}$ where $n$ is the number of image patches and $d_{\mathrm{vision}}$ is the dimensionality of the vision encoder. Next, an adapter $p$ (either a simple linear layer or MLP) projects the vision features to embeddings $h_{\mathrm{img}} = p(z_{\mathrm{img}}) \in \mathbb{R}^{n \times d_{\mathrm{text}}}$ where $d_{\mathrm{text}}$ is the dimensionality of the LLM. Lastly, $h_{\mathrm{img}}$ is concatenated with text + +prompt embeddings $h_{\mathrm{prompt}} = \mathrm{embed}(x_{\mathrm{prompt}})$ and passed into the language model LM to generate the output text $y = \mathrm{LM}([h_{\mathrm{img}}; h_{\mathrm{prompt}}])$ . + +In this work, we study the inference acceleration of VLMs where visual tokens are pruned within the attention mechanism of LM. We focus on the FastV [10] approach, where after layer $K$ in LM, $R\%$ of visual tokens are pruned away based on a ranking function $g_{\phi}$ which ranks tokens based on criteria $\phi$ . In practice, the attention score received from the last text token is used as the criteria, referred to as $\phi_{\mathrm{original}}$ . Note that the positional information is preserved when performing the pruning. + +With this approach, we measure the acceleration using the theoretical FLOPS reduction ratio related to the image tokens. For one Transformer layer, the total FLOPS is estimated as $C = 4nd^2 + 2n^2d + 2nd_m$ where $d = d_{\mathrm{text}}$ and $m$ is the intermediate size of FNN. Given that the Transformer has $T$ layers in total and after layer $K$ , we maintain $\hat{n} = (1 - R\%) * n$ visual tokens, we calculate the FLOPS reduction as + +$$ +1 - \frac {K * C + (T - K) * (4 \hat {n} d ^ {2} + 2 \hat {n} ^ {2} d + 2 \hat {n} d m)}{C}. \tag {1} +$$ + +Benchmarking. We evaluate the accelerated VLMs on a large suite of benchmarks from [18] which includes evaluations spanning the areas of localization, open-ended visual question answering, and challenge sets. For all benchmarks, we follow the same evaluation protocol as [18]. + +Localization. We evaluate localization performance using the RefCOCO, RefCOCO+, RefCOCOg [19, 42] and OCID-Ref [37] datasets. RefCOCO contains short, spatially grounded descriptions, while RefCOCO+ focuses on appearance-based descriptions. RefCOCOg, on the other hand, features longer and more detailed descriptions. OCID-Ref is a robotics dataset that tests generalization to out-of-distribution scenarios, particularly for object localization in cluttered environments. + +Open-Ended Visual Question Answering. We evaluate general visual reasoning using the VizWiz [5] and VQAv2 [13] datasets, spatial reasoning using the GQA [16] dataset, and reasoning around text using the TextVQA [32] dataset. For TextVQA, following [18], we exclude OCR-system parsed input tokens to better evaluate the effect of pruning on visual capabilities. + +Challenge Sets. We additionally evaluate on the VSR [26], TallyQA [1], POPE [25], and AI2D [20] closed-set prediction datasets. VSR includes binary spatial relationship questions, TallyQA consists of counting questions, POPE probes hallucination, and AI2D contains multiple-choice questions referring to scientific diagrams and charts. + +Experimental settings. In our experiments, we utilize a VLM with SigLIP ViT-SO400M [43] as $f$ , a one-layer MLP with GELU activation as $p$ , and Llama 2 7B [35] as LM. The model was trained on the multimodal instruction tuning dataset presented in [27] in a single-stage with $f$ frozen. + +![](images/9f876d33145c48301090e7ce73fff2709a0513bc1bcc71d3f92ce0bec797b033.jpg) +Figure 2. Contrasting the difference in performance dropoff on the challenging vision-centric localization task (Left) versus the other evaluated tasks (Middle) when pruning visual tokens after the shallow LLM layers. Whereas performance decrease is minimal for most tasks, localization exhibits roughly a linear decrease to zero as the ratio of pruned tokens increases. Right: Per-task performance breakdown across various setups of pruning ratios. Using $K = 3$ for all pruning setups. + +![](images/286cdba0b0aa6f7e1371ab0a3b257a79e651aa61adad2d8e7c0d50598ed9a34b.jpg) + +![](images/4f4bf20589ae09615e518ba86872a1f80bc19be70d89303d0cb3e26acde805b6.jpg) + +# 3.2. Impact of early visual token pruning varies drastically by task + +We examine the effect of pruning visual tokens after the shallow LLM layers $(K = 3)$ across a variety of tasks. We compare various pruning ratios $(R \in \{0.25, 0.5, 0.75, 0.9\})$ and the baseline non-pruned model in Figure 2. + +For the majority of evaluated tasks, heavily reducing the number of visual tokens results in minimal performance decrease (3.5% for AI2D, 0.1% for VSR, 3.7% for TallyQA, 4.8% for POPE, 5.1% for VizWiz, 7.9% for VQAv2 and 7.7% for GQA when dropping 75% of tokens). However, for a small subset of tasks—particularly TextVQA, and more severely, localization-performance drops substantially, with decreases of 42.0%, 88.9%, 88.9%, 91.0%, and 86.0% for TextVQA, RefCOCO, RefCOCO+, RefCOCOg, and OCID-Ref, respectively. Investigating further, Figure 2 shows that localization performance declines roughly linearly to zero as the pruning ratio progresses from 0 to 1, a pattern that strongly contrasts with most other tasks, where performance remains largely unaffected. While prior works report a much smaller decline in TextVQA performance [9, 39], they incorporate OCR-parsed input tokens into prompts, reducing the reliance on visual information. + +Discussion. Although localization is undoubtedly challenging for a model that discards visual tokens, as it requires fine-grained visual grounding and precise boundary identification, it is still surprising that (1) localization performance declines to the extent it does even when many tokens are maintained (e.g., dropping $50\%$ of tokens reduces average performance by $45\%$ ), and (2) most non-localization benchmarks, except for TextVQA, exhibit a drastically different pattern despite these benchmarks aiming to evaluate visual grounding abilities, such as counting for TallyQA, spatial reasoning for GQA, and chart understanding for AI2D. In the following sections, we seek to explain these unexpected findings and understand their implications. + +# 3.3. Interpreting poor vision-centric task performance + +We next investigate why pruning visual tokens after the shallow LLM layers has such an adverse effect on localization performance. Intuitively, even localization should not require all image tokens, as there are background tokens irrelevant to identifying specific objects that can be pruned. Thus, it is surprising that performance begins to degrade even with minimal token pruning. This suggests that the pruning criteria, intended to retain important tokens, fails to effectively distinguish which tokens contain necessary visual information. + +To assess the efficacy of the pruning criteria, we first examine the distribution of retained tokens across all benchmark examples. As shown in Figure 3(b), we find that pruning visual tokens after the shallow LLM layers $(K = 3)$ retains tokens concentrated at the bottom of the image. For example, when $R = 0.75$ , the average sampled token position across all datasets is situated $80.7\%$ of the way down the image. A Chi-Square test confirms the non-uniformity of selected tokens' y-positions ( $p$ -value $< 0.05$ ). We also examine the criteria behavior when pruning after later layers ( $K \in (8,16,24)$ ) and observe that the bias of selecting bottom image tokens is reduced. + +In addition, we observe that as the pruning layer increases, not only is the positional bias reduced, but the criteria also begins to correctly select tokens relevant to the text instruction. For example, as shown in Figure 3(a), pruning after the later layers results in a distinct concentration of selected tokens around the area corresponding to the reference expression for localization. To quantify this change in criteria effectiveness across layers, we measure the performance when varying the pruning layer $(K\in \{3,8,16,24\})$ As shown in Figure 3(c), performance remains stable when pruning is performed at deeper layers $K = 16$ and $K =$ 24), whereas pruning at earlier layers $K = 3$ and $K = 8$ + +![](images/72a40db4bc8d989578cc28510d68d7755ecdb5e3bb820e170b9fbd971f1e7047.jpg) +(a) + +![](images/f8b3451ae6cb8d90bbd59c738920d5f151d2634ee388ec2d74cd64afe48a1763.jpg) +(b) + +![](images/f4306611f75804fbbac25d2b58ad4a9d0cade5402eacdf25ba9177093104e7fc.jpg) +Reference expression: a bowl of blueberries + +![](images/4b89b253336473434b20c3d8fb99788acf6767c1652f44b5df2bfe6c9704de6c.jpg) + +![](images/38e0177f484f621655db8b2c810f33c3284f3c5c2be5e90b4bca068413632f90.jpg) +x position K-16 + +![](images/80b6d7be38344bc1ce262474fdcc32da3b9c5dd1cd3409bdf9e8bd20c063a787.jpg) +x position + +![](images/d85b30c96cfc96f8dd07fc317f4f7e2b89fd514cce55bb57f70c7e4e1dde6c89.jpg) +x position $K = 24$ + +![](images/8e979172f1352d135f6b68fc32a7be4d7be7bc91db119e4c6bdfb4c56c011131.jpg) +x position + +![](images/f1c771f753620cb6569dbf5c32c855866fef6378fe644ea4594a4efa2992a3f3.jpg) +Figure 3. (a) Example demonstrating that when pruning in early layers, selected tokens are concentrated in the bottom of the image, whereas in later layers the selected tokens more precisely cover the region of the image important for answering the question. (b) Heatmap illustrating that averaged across all benchmark examples, as the pruning layer increases, the selection of bottom visual tokens by the criteria is reduced. (c) Visualizing the effect of the pruning layer on performance for both localization and non-localization tasks. We find that pruning after layer 16 or later results in little performance dropoff, whereas pruning earlier results in a performance decrease, particularly for localization. Using $R = 0.75$ for all setups. + +![](images/ee6aacba1a99a9860ed9dd5a5ced81005abc1f80fc85a078c3cd09f037284089.jpg) +(c) + +![](images/f39e6d530bafc921bbca4dbb67e4aa6c80c528f3b91748e4391e3276b911ce4d.jpg) +Non-localization + +leads to performance degradation, particularly for localization. + +Discussion. While previous research attributes the poor performance of early token pruning to the low redundancy of image tokens in shallow layers [39], we reveal a critical insight: the variable performance across pruning layers is linked to the effectiveness of selecting important tokens, as early token pruning leads to suboptimal token selection biased toward bottom tokens. To understand why the criteria disproportionately retains bottom image tokens, we must consider the LM architecture. Specifically, token selection is based on the attention received from the last text token. Since the LM uses RoPE to encode positional information [33], it exhibits a long-term decay property. While this property is well-suited for language modeling, we uncover that it poses an issue for selecting visual tokens, where tokens appearing later in the flattened, raster-scan order have higher attention scores purely due to their position. In language modeling, prior work finds that there is greater emphasis on shorter-distance information in shallow layers [15]. We predict that the visually-conditioned LM exhibits the same behavior, explaining why the positional bias is most pronounced when pruning after shallow layers. + +# 3.4. Explaining strong performance on other tasks + +Given that early pruning of visual tokens leads to a suboptimal tokens selection where remaining tokens are concentrated towards the bottom of the image, a natural question arises: how does the approach still maintain high perfor + +![](images/f3dc8989efff27dd6c48a5ff61c10977b787a5d7dabe9faca6396a03a895e550.jpg) +Figure 4. Assessing whether the strong performance of early visual token pruning for many tasks can be attributed to visual information transfer before pruning or benchmarks' lack of assessing fine-grained visual capabilities. We observe that for all evaluated tasks, allowing information transfer before pruning (shown in green) does not result in substantial performance improvement over the setup without visual information transfer (shown in light green), highlighting a limitation of many benchmarks. + +mance across a diverse range of tasks, including those aiming to evaluate visual grounding abilities? In this section, we test the following contrasting explanations: + +1. Important visual information from top image tokens is transferred to later tokens before pruning with the unidirectional attention of the LLM. + +2. Many questions can still be inferred with access to only the suboptimal set of visual tokens primarily from the bottom of the image. + +To evaluate these hypotheses across tasks, we save the tokens retained by the studied pruning approach and, in a subsequent run, remove all other tokens before they reach the LLM, ensuring that no visual information from the pruned tokens is retained. As an additional comparison, we also evaluate a text-only model in which all visual tokens are removed before entering the LLM. As shown in Figure 4, we find that across all evaluated tasks, removing pruned tokens before the LLM results in little to no performance decrease compared to the method that allows information transfer in shallow layers prior to pruning. + +Discussion. This finding indicates that the high performance of early visual token pruning on many tasks does not stem from the effectiveness of visual information transfer in early layers but rather signifies that many benchmarks do not demand a detailed understanding of visual information to answer questions accurately. Note that for the majority of these benchmarks (except VizWiz and AI2D), performance substantially declines when not including any visual information, signifying that simply comparing vision-enabled and disabled setups is insufficient to determine a benchmark's ability to assess visual grounding. + +# 4. Improving Visual Token Pruning + +Based on our findings from §3, we now seek to answer: can attention serve as an effective criteria for token pruning even in early layers? To this end, we propose a straightforward solution to removing the positional bias of the attention-based criteria and demonstrate that our modifications enable effective pruning even when applied after early LLM layers (§4.1). Guided by our insights, we present our final approach, FEATHER (Fast and Effective Acceleration with Ensemble cRicerca) (§4.2), and showcase its impressive visual capabilities while offering high computational efficiency (§4.3). + +# 4.1. Assessing potential of attention-based criteria + +An attention-based criteria offers the potential to naturally identify and retain only the most relevant visual tokens for answering a question. However, as shown in §3, positional bias strongly influences token selection, limiting its effectiveness. Thus, we aim to resolve the positional bias issue and explore the true promise of attention-based criteria. We compare against non-attention-based approaches that incorporate an explicit inductive bias toward reducing redundancy. Below, we outline the explored criteria. + +$\phi_{-R}$ : Given the long-term decay property of RoPE, we propose the following straightforward adjustment to the pruning criteria in order to mitigate this bias and more effectively select tokens relevant to the text instruction. Namely, + +we still compute the attention score received from the last text token, except we do not apply RoPE to the attention mechanism, thereby removing the long-term decay effect. Note that this lightweight calculation is only done once per pruning, and all other state computations are compatible with FlashAttention [12]. + +$\phi_{\mathrm{uniform}}$ : In this simple criteria, we uniformly sample visual tokens in the image, using a set stride. This criteria ensures that there is good image coverage, but does so by sacrificing the ability to have more densely captured visual information for a particular image region. To use a similar computational saving to the attention-based criteria when $R = 0.75$ , we use a stride of two, resulting in 196 selected tokens. + +$\phi_{\mathrm{KNN}}$ : Inspired by dynamic visual tokenization [17], we select tokens based on local density. For each visual token $z_{img}[i]$ where $i \in \{0, 1, \dots, n-1\}$ , we compute its local density $\rho_i$ using K-nearest neighbors and calculate distance index $\delta_i$ as: + +$$ +\delta_ {i} = \left\{ \begin{array}{l l} \min _ {j: \rho_ {j} > \rho_ {i}} \| z _ {i m g} [ i ] - z _ {i m g} [ j ] \| ^ {2}, & \text {i f} \exists j \text {s . t .} \rho_ {j} > \rho_ {i}. \\ \max _ {j} \| z _ {i m g} [ i ] - z _ {i m g} [ j ] \| ^ {2}, & \text {o t h e r w i s e .} \end{array} \right. \tag {2} +$$ + +$\delta_{i}$ represents how far away $z_{img}[i]$ is from other high-density tokens. We use $\rho_{i}*\delta_{i}$ as the token importance score for the criteria and select 196 tokens. + +$\phi_{-R} + \phi_{\mathrm{uniform}}$ : Since $\phi_{-R}$ focuses on selecting important tokens and $\phi_{\mathrm{uniform}}$ aims to reduce redundancy, we propose an ensemble of these two to leverage their respective strengths. Specifically, we apply $\phi_{-R}$ while incorporating a small number of uniformly sampled tokens using a stride of three. + +Improvement of removing RoPE: In Table 1, we report results evaluating the effectiveness of these criteria. Comparing $\phi_{-R}$ with $\phi_{\mathrm{original}}$ , we find that by removing RoPE, at $K = 3$ there is a $183\%$ average improvement on localization tasks and at $K = 8$ , there is a $17\%$ average improvement. These performance improvements demonstrate that once the impact of token position on attention is removed, attention score can more effectively be used for the criteria when applied after early LLM layers. Note that the narrowing gap in performance between $\phi_{-R}$ and $\phi_{\mathrm{original}}$ from $K = 3$ to $K = 8$ is likely because $\phi_{\mathrm{original}}$ has less of a bias towards selecting bottom image tokens as the pruning layer increases. + +Token selection improves when pruning later: Comparing $\phi_{-R}$ when $K = 8$ versus $K = 3$ , we see that the average localization performance improves by $63\%$ . However, it remains unclear whether this performance increase is affected by factors other than the ability of the criteria to select important tokens. Therefore, to directly compare the criteria across different layers, we take the tokens selected by both criteria and evaluate how well the model can perform with only these tokens passed into the LLM (as is + +
Pruning LayerCriteriaFLOPS RedLocalizationOpen-Ended VQAChallenge Sets
AvgOCID-RefRefCOCObRefCOCO+RefCOCOAvgTextVQAGQAVQAv2VizWizAvgPOPETallyQAVSRAI2D
K=3Attention-based
φoriginal68%5.95.75.16.16.754.831.858.472.756.364.083.257.163.352.4
φR(Ours)68%16.722.915.113.315.359.041.661.276.057.364.785.258.262.253.2
Δ+10.7+17.2+10.0+7.3+8.6+4.2+9.7+2.8+3.2+1.1+0.7+1.9+1.1-1.1+0.8
Non-attention-based
φKNN66%23.915.124.926.029.658.439.960.974.458.462.881.255.961.552.8
φuniform66%28.020.628.629.733.359.041.461.875.957.164.685.258.161.953.0
Ensemble
φR+φuniform(Ours)61%27.229.127.224.727.761.246.662.377.458.465.486.058.962.754.0
K=8Attention-based
φoriginal56%23.319.423.524.026.359.845.060.376.157.864.685.457.562.653.0
φR(Ours)56%27.327.126.726.429.261.449.061.577.457.865.586.758.663.053.7
Δ+4.0+7.6+3.2+2.5+2.9+1.6+4.0+1.2+1.3+0.0+0.8+1.3+1.1+0.4+0.6
Non-attention-based
φKNN55%23.615.424.425.229.458.640.261.174.558.562.981.456.260.953.0
φuniform55%30.324.631.030.934.859.342.261.876.057.464.485.357.961.053.2
Ensemble
φR+φuniform(Ours)50%35.632.035.935.438.862.751.762.478.158.666.087.459.163.654.0
+ +Table 1. Evaluating alternative criteria for token pruning after the early LLM layers. For each task and pruning layer, we **bold** the best result and **underline** the second-best result. Our main findings include: (1) our proposed RoPE-free criteria $\phi_{-R}$ substantially improves pruning performance compared to the original criteria $\phi_{\mathrm{original}}$ ; (2) pruning later ( $K = 8$ ) yields higher performance than pruning earlier ( $K = 3$ ); and (3) integrating uniform sampling into the attention-based criteria with $\phi_{-R} + \phi_{\mathrm{uniform}}$ enhances effectiveness. Using $R = 0.75$ for $\phi_{\mathrm{original}}$ , $\phi_{-R}$ , and $\phi_{-R} + \phi_{\mathrm{uniform}}$ . See §4.1 for criteria definitions. + +
KOCID-RefRefCOCOgRefCOCO+RefCOCO
323.816.214.516.3
826.729.826.829.8
+ +Table 2. Assessing localization performance when using the selected tokens from $\phi_{-R}$ for the entirety of the LLM. We find that $\phi_{-R}$ applied at a later LLM layer results in a better token selection. + +done in §3.4). As shown in Table 2, we find the selected tokens from $\phi_{-R}$ applied after layer 8 are superior compared to the tokens from $\phi_{-R}$ applied after layer 3. + +Integrating uniform sampling with attention-based criteria is beneficial in early layers: When comparing the attention-based criteria $\phi_{-R}$ with the two non-attention based criteria $\phi_{\mathrm{KNN}}$ and $\phi_{\mathrm{uniform}}$ , we find that $\phi_{-R}$ results in better performance on some tasks (e.g., OCID-Ref, TextVQA), while $\phi_{\mathrm{KNN}}$ and $\phi_{\mathrm{uniform}}$ outperform on other tasks (e.g., RefCOCO). This varying effectiveness of criteria types presumably comes from some tasks requiring a more detailed understanding of specific image regions while others benefit from a broader understanding with full image + +coverage. When we combine these two criteria types with $\phi_{-R} + \phi_{\mathrm{uniform}}$ , we find that when $K = 3$ this approach improves upon localization performance of $\phi_{-R}$ by $63\%$ but has a slight decrease compared to $\phi_{\mathrm{uniform}}$ by $2.9\%$ . However, when $K = 8$ , this approach far outperforms both of the individual criteria, improving $\phi_{-R}$ by $30\%$ and $\phi_{\mathrm{uniform}}$ by $17\%$ . Note that this setup results in slightly less computation efficiency than $\phi_{-R}$ (7% drop in FLOPS reduction for $K = 3$ and $6\%$ for $K = 16$ ), but this is outweighed by the performance improvements. + +We provide qualitative examples of various criteria token selection and additional results using a different model setup in the supplement. + +# 4.2. Distilling insights + +Guided by our insights, we now present our final approach FEATHER (Fast and Effective Acceleration wiTH Ensemble cRicerca). In this approach, we first perform pruning after an early layer $(K = 8)$ , utilizing our proposed criteria $\phi_{-R} + \phi_{\mathrm{uniform}}$ to retain $(1 - R)\%$ of tokens. Given our finding that the criteria improves as the LLM layer increases, we additionally prune a second time at $K = 16$ . + +![](images/4f4a5029d3b97b52c56898e84e33a5b61a00f898caeb79e93e6a3149b6ddd97a.jpg) +Figure 5. Comparing FEATHER performance against FastV [10] and PyramidDrop [39]. We find that FEATHER far outperforms both compared methods, particularly for the vision-centric task of localization. + +At this stage, we utilize $\phi_{-R}$ as the criteria, as uniform sampling should not be necessary when the attention-based criteria reliably selects important tokens. For the reduction ratio, we choose to only retain $(1 - R)^{2\%}$ of the remaining tokens since the attention-based criteria has proved highly effective when pruning at later layers, even when using the original criteria $\phi_{\mathrm{original}}$ (see Figure 3(c)). + +# 4.3. Comparison against FastV and PyramidDrop + +We compare our approach against the one-stage early pruning approach of FastV ( $K = 3$ ) [10] and the multi-stage approach of PyramidDrop ( $K = [8, 16, 24]$ ) [39]. Note that both of these methods use $\phi_{\mathrm{original}}$ for the criteria. + +As shown in Figure 5, we find that FEATHER far outperforms the baselines, especially on the localization tasks. Specifically, for comparable computational costs (64% FLOPS reduction for our approach, 68% FLOPS reduction for FastV, and 65% FLOPS reduction for PyramidDrop), we observe that for localization tasks, FEATHER exhibits more than $5 \times$ average performance improvement compared to FastV and a $36\%$ average performance improvement compared to PyramidDrop. For non-localization tasks, FEATHER has a $7.8\%$ improvement over FastV and $1.5\%$ improvement over PyramidDrop. Note that PyramidDrop performs substantially better than FastV as it prunes fewer tokens in an early layer. However, it still suffers from an ineffective pruning strategy at this stage, though the impact is less pronounced since it predominantly prunes later. + +Remarkably, with our $64\%$ FLOPS reduction setup, after layer 16 only $3.3\%$ of tokens are retained, yet the average localization performance decrease compared to the baseline method with no token pruning is only $26\%$ (Figure 1 includes an example of retained tokens after layer 16). This finding illustrates that even for vision-centric tasks, maintaining strong performance while gaining huge acceleration speedups with extensive pruning is possible, but it heavily relies on the effectiveness of the pruning criteria. + +# 5. Conclusion + +In this work, we examine the visual capabilities of the VLM acceleration approach of pruning visual tokens after shallow LLM layers. While strong performance is maintained on most evaluated tasks, it fails on more vision-centric tasks like TextVQA and localization due to its flawed pruning criteria that predominately selects visual tokens from the bottom part of the image. Observing this same flawed criteria on other tasks, we show that strong performance is largely due to the benchmarks' inability to assess fine-grained visual capabilities. Next, we propose and evaluate several alternative criteria to improve visual capabilities, ultimately arriving at our final method, FEATHER. This approach refines the attention-based criteria to address the token selection bias while using uniform sampling for better image coverage. It then prunes more aggressively when the criteria more effectively identifies important tokens. We find that FEATHER has more than $5 \times$ performance improvement on localization compared to the original acceleration approach. + +While we study a particular type of VLM acceleration, our work highlights a broad challenge in evaluating VLMs. Building on [34], which shows that text-only models can perform well on some vision-language benchmarks, we demonstrate that even benchmarks with substantial differences between vision-enabled and disabled setups may not assess fine-grained visual capabilities. We address this issue by focusing on more vision-centric localization tasks, though this limits the analysis to a few specific types of skills. To accurately assess a wider range of visual capabilities, future work may explore how to resolve current dataset biases that models can exploit [21]. Additionally, while removing RoPE for token pruning effectively eliminates positional bias, it may introduce unintended effects on attention weights influencing token selection. Future research could investigate more robust methods for encoding positional information in visually conditioned language models to prevent positional artifacts in cross-modal interactions. + +Acknowledgments. This work is supported in part by the National Science Foundation (NSF) under Grant No. 2026498, the Stanford AIMI-HAI Partnership Grant, and the NSF Graduate Research Fellowship Program under Grant No. DGE-2146755 (for M.E.). Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of any other entity. + +# References + +[1] Manoj Acharya, Kushal Kafle, and Christopher Kanan. Tallyqa: Answering complex counting questions. In Proceedings of the AAAI conference on artificial intelligence, pages 8076-8084, 2019. 3 +[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 1 +[3] Kazi Hasan Ibn Arif, JinYi Yoon, Dimitrios S Nikolopoulos, Hans Vandierendonck, Deepu John, and Bo Ji. Hired: Attention-guided token dropping for efficient inference of high-resolution vision-language models in resource-constrained environments. arXiv preprint arXiv:2408.10945, 2024. 2, 3 +[4] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 1 +[5] Jeffrey P Bigham, Chandrika Jayant, Hanjie Ji, Greg Little, Andrew Miller, Robert C Miller, Robin Miller, Aubrey Tatarowicz, Brandyn White, Samual White, et al. Vizwiz: nearly real-time answers to visual questions. In Proceedings of the 23nd annual ACM symposium on User interface software and technology, pages 333-342, 2010. 3 +[6] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. In The Eleventh International Conference on Learning Representations, 2023. 3 +[7] Mu Cai, Jianwei Yang, Jianfeng Gao, and Yong Jae Lee. Matryoshka multimodal models. arXiv preprint arXiv:2405.17430, 2024. 2 +[8] Wenhao Chai, Enxin Song, Yilun Du, Chenlin Meng, Vashisht Madhavan, Omer Bar-Tal, Jeng-Neng Hwang, Saining Xie, and Christopher D Manning. Auroracap: Efficient, performant video detailed captioning and a new benchmark. arXiv preprint arXiv:2410.03051, 2024. 3 +[9] Jieneng Chen, Luoxin Ye, Ju He, Zhao-Yang Wang, Daniel Khashabi, and Alan Yuille. Efficient large multi-modal models via visual context compression. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 4 +[10] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 + +tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. ECCV, 2024. 2, 3, 8 +[11] Yi Chen, Jian Xu, Xu-Yao Zhang, Wen-Zhuo Liu, YangYang Liu, and Cheng-Lin Liu. Recoverable compression: A multimodal vision token recovery mechanism guided by text information. arXiv preprint arXiv:2409.01179, 2024. 3 +[12] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in neural information processing systems, 35:16344-16359, 2022. 6 +[13] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 3 +[14] Yefei He, Feng Chen, Jing Liu, Wenqi Shao, Hong Zhou, Kaipeng Zhang, and Bohan Zhuang. Zipvl: Efficient large vision-language models with dynamic token sparsification and kv cache compression. arXiv preprint arXiv:2410.08584, 2024.3 +[15] Xiangyu Hong, Che Jiang, Biqing Qi, Fandong Meng, Mo Yu, Bowen Zhou, and Jie Zhou. On the token distance modeling ability of higher rope attention dimension. arXiv preprint arXiv:2410.08703, 2024. 5 +[16] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 3 +[17] Peng Jin, Ryuichi Takanobu, Wancai Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13700-13710, 2024. 2, 6 +[18] Siddharth Karamcheti, Suraj Nair, Ashwin Balakrishna, Percy Liang, Thomas Kollar, and Dorsa Sadigh. Prismatic VLMs: Investigating the design space of visually-conditioned language models. In Proceedings of the 41st International Conference on Machine Learning, pages 23123-23144. PMLR, 2024. 3 +[19] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014. 3 +[20] Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Minjoon Seo, Hannaneh Hajishirzi, and Ali Farhadi. A diagram is worth a dozen images. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 235-251. Springer, 2016. 3 +[21] Baiqi Li, Zhiqiu Lin, Wenxuan Peng, Jean de Dieu Nyandwi, Daniel Jiang, Zixian Ma, Simran Khanuja, Ranjay Krishna, Graham Neubig, and Deva Ramanan. Naturalbench: Evaluating vision-language models on natural adversarial samples. Advances in Neural Information Processing Systems, 37:17044-17068, 2025. 8 + +[22] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 1 +[23] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 1 +[24] Kevin Y Li, Sachin Goyal, Joao D Semedo, and J Zico Kolter. Inference optimal vlms need only one visual token but larger models. arXiv preprint arXiv:2411.03312, 2024. 2 +[25] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 3 +[26] Fangyu Liu, Guy Emerson, and Nigel Collier. Visual spatial reasoning. Transactions of the Association for Computational Linguistics, 11:635-651, 2023. 3 +[27] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 1, 3 +[28] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024. 2 +[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1 +[30] Yifei Liu, Mathias Gehrig, Nico Messikommer, Marco Cannici, and Davide Scaramuzza. Revisiting token pruning for object detection and instance segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2658-2668, 2024. 3 +[31] Yuzhang Shang, Mu Cai, Bingxin Xu, Yong Jae Lee, and Yan Yan. Llava-prumerge: Adaptive token reduction for efficient large multimodal models. arXiv preprint arXiv:2403.15388, 2024. 2 +[32] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 3 +[33] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 5 +[34] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, et al. Cambrian1: A fully open, vision-centric exploration of multimodal llms. arXiv preprint arXiv:2406.16860, 2024. 8 +[35] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. + +Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3 +[36] Zhongwei Wan, Ziang Wu, Che Liu, Jina Huang, Zhihong Zhu, Peng Jin, Longyue Wang, and Li Yuan. Look-m: Look-once optimization in kv cache for efficient multimodal long-context inference. arXiv preprint arXiv:2406.18139, 2024. 3 +[37] Ke-Jyun Wang, Yun-Hsuan Liu, Hung-Ting Su, Jen-Wei Wang, Yu-Siang Wang, Winston Hsu, and Wen-Chin Chen. OCID-ref: A 3D robotic dataset with embodied language for clutter scene grounding. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5333–5338, Online, 2021. Association for Computational Linguistics. 3 +[38] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1 +[39] Long Xing, Qidong Huang, Xiaoyi Dong, Jiajie Lu, Pan Zhang, Yuhang Zang, Yuhang Cao, Conghui He, Jiaqi Wang, Feng Wu, et al. Pyramiddrop: Accelerating your large vision-language models via pyramid visual redundancy reduction. arXiv preprint arXiv:2410.17247, 2024. 2, 3, 4, 5, 8 +[40] Senqiao Yang, Yukang Chen, Zhuotao Tian, Chengyao Wang, Jingyao Li, Bei Yu, and Jiaya Jia. Visionzip: Longer is better but not necessary in vision language models. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 19792-19802, 2025. 2 +[41] Gaotong Yu, Yi Chen, and Jian Xu. Balancing performance and efficiency: A multimodal large language model pruning method based image text interaction. arXiv preprint arXiv:2409.01162, 2024.3 +[42] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 3 +[43] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 3 +[44] Qizhe Zhang, Aosong Cheng, Ming Lu, Zhiyong Zhuo, Minqi Wang, Jiajun Cao, Shaobo Guo, Qi She, and Shang-hang Zhang. [cls] attention is all you need for training-free visual token pruning: Make vlm inference faster. arXiv eprints, pages arXiv-2412, 2024. 2 \ No newline at end of file diff --git a/2025/Feather the Throttle_ Revisiting Visual Token Pruning for Vision-Language Model Acceleration/images.zip b/2025/Feather the Throttle_ Revisiting Visual Token Pruning for Vision-Language Model Acceleration/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e29ed924f785c992f3b45aa639d5e504dd687eb8 --- /dev/null +++ b/2025/Feather the Throttle_ Revisiting Visual Token Pruning for Vision-Language Model Acceleration/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a94f9a9ace948713e59238f97f9b84d7df994d555956a29ab3e99bb5bb0babe7 +size 467824 diff --git a/2025/Feather the Throttle_ Revisiting Visual Token Pruning for Vision-Language Model Acceleration/layout.json b/2025/Feather the Throttle_ Revisiting Visual Token Pruning for Vision-Language Model Acceleration/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7bad47b3c4e59aa92875a45a13b19c708916832f --- /dev/null +++ b/2025/Feather the Throttle_ Revisiting Visual Token Pruning for Vision-Language Model Acceleration/layout.json @@ -0,0 +1,8821 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 127, + 103, + 483, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 103, + 483, + 140 + ], + "spans": [ + { + "bbox": [ + 127, + 103, + 483, + 140 + ], + "type": "text", + "content": "Feather the Throttle: Revisiting Visual Token Pruning for Vision-Language Model Acceleration" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 187, + 161, + 423, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 161, + 423, + 190 + ], + "spans": [ + { + "bbox": [ + 187, + 161, + 423, + 190 + ], + "type": "text", + "content": "Mark Endo, Xiaohan Wang, Serena Yeung-Levy \nStanford University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 195, + 191, + 411, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 191, + 411, + 204 + ], + "spans": [ + { + "bbox": [ + 195, + 191, + 411, + 204 + ], + "type": "text", + "content": "{markendo,xhanwang,syyeung}@stanford.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 152, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 152, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 258, + 296, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 296, + 547 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 296, + 547 + ], + "type": "text", + "content": "Recent works on accelerating Vision-Language Models achieve strong performance across a variety of vision-language tasks despite highly compressing visual information. In this work, we examine the popular acceleration approach of early pruning of visual tokens inside the language model. Surprisingly, we find that while strong performance is maintained across many tasks, it exhibits drastically different behavior for a subset of vision-centric tasks such as localization. Upon further investigation, we uncover a core issue with the acceleration approach where most tokens towards the top of the image are pruned away. Yet, on many benchmarks aiming to evaluate vision-centric capabilities, strong performance persists with the flawed pruning strategy, highlighting these benchmarks' limited ability to assess fine-grained visual capabilities. Based on these findings, we propose FEATHER (Fast and Effective Acceleration with Ensemble cRicerca), a straightforward approach that resolves the discovered early-layer pruning issue and further enhances the preservation of relevant tokens via multistage pruning with early uniform sampling to ensure broad image coverage. With comparable computational savings, we find that FEATHER achieves more than " + }, + { + "bbox": [ + 53, + 258, + 296, + 547 + ], + "type": "inline_equation", + "content": "5 \\times" + }, + { + "bbox": [ + 53, + 258, + 296, + 547 + ], + "type": "text", + "content": " performance improvement on the vision-centric localization benchmarks compared to the original acceleration approach." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 572, + 135, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 572, + 135, + 585 + ], + "spans": [ + { + "bbox": [ + 55, + 572, + 135, + 585 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": "The exploration of Vision-Language Models (VLMs) is a critical area of computer vision and natural language processing research, centered on combining large language models (LLMs) with visual encoders to enable multimodal perception, reasoning, and understanding capabilities. While earlier works explored sophisticated schemes for conditioning language models with visual information [2, 22, 23], more recently the space has shifted to predominately using the simplistic approach of taking patch features from pre-trained visual encoders and projecting them" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 333, + 230, + 552, + 443 + ], + "blocks": [ + { + "bbox": [ + 317, + 231, + 329, + 242 + ], + "lines": [ + { + "bbox": [ + 317, + 231, + 329, + 242 + ], + "spans": [ + { + "bbox": [ + 317, + 231, + 329, + 242 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 333, + 230, + 552, + 443 + ], + "lines": [ + { + "bbox": [ + 333, + 230, + 552, + 443 + ], + "spans": [ + { + "bbox": [ + 333, + 230, + 552, + 443 + ], + "type": "image", + "image_path": "4d43fa273c72025972152408e9df7d7b3aeef6ba8cfda68fe40c567bda3e264b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 331, + 448, + 550, + 526 + ], + "blocks": [ + { + "bbox": [ + 317, + 448, + 329, + 458 + ], + "lines": [ + { + "bbox": [ + 317, + 448, + 329, + 458 + ], + "spans": [ + { + "bbox": [ + 317, + 448, + 329, + 458 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 331, + 448, + 550, + 526 + ], + "lines": [ + { + "bbox": [ + 331, + 448, + 550, + 526 + ], + "spans": [ + { + "bbox": [ + 331, + 448, + 550, + 526 + ], + "type": "image", + "image_path": "d8fc2f3288d4e37f83c664b0544d201797876a8f1b938aefe429ed16c63a5d47.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 537, + 555, + 636 + ], + "lines": [ + { + "bbox": [ + 313, + 537, + 555, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 537, + 555, + 636 + ], + "type": "text", + "content": "Figure 1. (a) Although FastV prunes most visual tokens from the upper portion of the image, the approach still displays strong performance on a variety of evaluated vision-language tasks except for a small subset of vision-centric tasks like localization. (b) Based on our findings, we propose FEATHER (Fast and Effective Acceleration wiTH Ensemble cRicerca), a straightforward approach that resolves the existing issue of selecting bottom tokens, additionally maintains uniformly sampled tokens to ensure good coverage over the whole image, and prunes in two stages." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": "to the input space of the language model with a light-weight adapter [4, 27, 29, 38]. Using image patches as tokens, however, comes with the drawback of being computationally inefficient. To achieve fine-grained resolution, the im" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 749, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 758 + ], + "type": "text", + "content": "22826" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 312 + ], + "type": "text", + "content": "age is divided into many patches. This large number of patches significantly increases computational demands due to the quadratic complexity of the attention operation in Transformers. As a result, many recent works have focused on accelerating these methods by compressing visual information, demonstrating that heavy compression can still maintain strong performance across a wide variety of tasks [3, 7, 10, 31, 39]. For instance, FastV [10] prunes " + }, + { + "bbox": [ + 55, + 72, + 297, + 312 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 72, + 297, + 312 + ], + "type": "text", + "content": " of visual tokens after the shallow layers of the Language Model (LLM) while not compromising in performance across a range of image and video understanding tasks [10]. Another work reveals that for a fixed computational budget on visual reasoning tasks, optimal performance is achieved by using the largest LLM possible while sacrificing visual information, often reducing the visual token count to a single token [24]. With such compressed visual information, it is still unclear how these methods achieve high performance on tasks assessing vision capabilities such as visual reasoning and understanding, and whether there are more demanding visual tasks where these methods fail." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 315, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 315, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 315, + 296, + 567 + ], + "type": "text", + "content": "The motivation of our study is to get a better understanding of the vision capabilities of accelerated VLMs given that they leverage highly compressed visual information, focusing specifically on FastV approach. Evaluating across a wide range of vision-language tasks, we find that while the approach maintains strong performance across many tasks, there is a substantial decrease in performance for TextVQA and a severe decrease for localization tasks. When analyzing the approach's poor performance, we uncover a fundamental issue with the pruning criteria in early layers where tokens towards the top of the image are disproportionally removed due to the long-term decay property of RoPE. Since this issue is not task-specific, we examine how performance is maintained across most evaluated tasks and find that even when tokens are entirely discarded using the flawed criteria, performance remains largely unchanged. This highlights how many benchmarks require minimal visual grounding. Our finding underscores a significant challenge in the field of multimodal learning, not only for measuring the effectiveness of VLM acceleration methods but also for benchmarking the visual capabilities of VLMs as a whole." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "Given the discovered limitation of the studied approach, we next explore whether the attention-based criteria can be modified to enable robust visual token pruning, leading to our final method, FEATHER (Fast and Effective Acceleration with Ensemble Ricerca). Specifically, we propose a straightforward modification to the attention-based criteria by removing RoPE, demonstrating that this resolves the positional bias issue and greatly improves performance. Additionally, we incorporate uniform sampling in early token pruning to ensure adequate coverage of all image regions and apply more extensive pruning at a later layer, when the attention-based criteria better discerns im" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 555, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 263 + ], + "type": "text", + "content": "important tokens. This strategy is analogous to how a race-car driver feathers the throttle by gradually pressing the accelerator at the beginning of a turn to maintain grip and then accelerating more aggressively once the car is past the apex. We show that FEATHER results in substantial performance gains compared to the original acceleration approach, improving localization performance more than " + }, + { + "bbox": [ + 313, + 72, + 555, + 263 + ], + "type": "inline_equation", + "content": "5 \\times 5" + }, + { + "bbox": [ + 313, + 72, + 555, + 263 + ], + "type": "text", + "content": " with comparable computational savings. Strikingly, we find that our approach achieves this performance improvement while only retaining " + }, + { + "bbox": [ + 313, + 72, + 555, + 263 + ], + "type": "inline_equation", + "content": "3.3\\%" + }, + { + "bbox": [ + 313, + 72, + 555, + 263 + ], + "type": "text", + "content": " of visual tokens for the second half of LLM layers. Overall, our work demonstrates that while visual compression can maintain strong performance even on challenging vision-centric tasks, its effectiveness depends on a well-designed strategy, which is currently difficult to assess due to many vision-language benchmarks not thoroughly evaluating vision capabilities." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 326, + 264, + 485, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 264, + 485, + 275 + ], + "spans": [ + { + "bbox": [ + 326, + 264, + 485, + 275 + ], + "type": "text", + "content": "In summary, our contributions include:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 328, + 277, + 553, + 552 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 328, + 277, + 553, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 277, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 328, + 277, + 553, + 336 + ], + "type": "text", + "content": "- Revealing inconsistent effect of early pruning: Our evaluations across 12 benchmarks show that early visual token pruning minimally affects performance on most tasks, but causes a substantial dropoff for TextVQA and a drastic decline for localization tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 329, + 337, + 553, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 337, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 329, + 337, + 553, + 384 + ], + "type": "text", + "content": "- Analyzing poor performance: We uncover that the attention-based criteria for selecting important tokens exhibits a bias towards bottom image tokens in early layers due to the long-term decay property of RoPE." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 329, + 385, + 553, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 385, + 553, + 456 + ], + "spans": [ + { + "bbox": [ + 329, + 385, + 553, + 456 + ], + "type": "text", + "content": "- Identifying benchmark limitations in assessing token pruning: We find that performance remains strong even with suboptimal pruning criteria and removal of pruned tokens before transferring information in the LLM, demonstrating how many benchmarks fail to assess fine-grained visual capabilities." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 329, + 456, + 553, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 456, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 329, + 456, + 553, + 552 + ], + "type": "text", + "content": "- Improving attention-based criteria: With the straightforward modification of removing RoPE, we demonstrate that attention can be used as an effective criteria for selecting tokens and performance can be further improved with simple strategies like incorporating uniform sampling and pruning more extensively at a later layer, when the criteria better discerns important tokens." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 562, + 402, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 562, + 402, + 574 + ], + "spans": [ + { + "bbox": [ + 314, + 562, + 402, + 574 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "Recent efforts to accelerate VLMs can be broadly divided into two main categories: compressing visual information before it enters the LLM, and compressing visual information within the LLM itself. In the first category, ChatUniVi [17] dynamically merges visual tokens with similar semantic meanings. PruMerge [31], FasterVLM [44], and VisionZip [40] select important tokens based on ViT attention ([CLS] or average attention received). PruMerge then merges important tokens with unselected ones while VisionZip merges unselected tokens together. For the LLaVA-NeXT [28] approach of partitioning an image into" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22827" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "sub-images where inefficiency is an even bigger problem, HiRED [3] selects tokens with top feature important on each sub-image with an allocated budget. Other methods argue that the input image alone does not include enough information to select important patches and thus use the textual input to recover visually meaningful tokens [11, 41]. Alternatively, some approaches reduce visual information within the ViT by merging or pruning tokens [6, 8, 30]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 168, + 295, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 168, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 295, + 312 + ], + "type": "text", + "content": "For the second category of approaches, where visual information is compressed within the LLM itself, LOOK-M reduces the multimodal KV cache size [36]. In our study, we focus on the popular FastV approach [10]. This work identifies that attention over image tokens is sparse in deeper layers of the LLM. Based on this observation, they propose to prune away unimportant vision tokens after the shallow layers of the LLM, achieving a " + }, + { + "bbox": [ + 55, + 168, + 295, + 312 + ], + "type": "inline_equation", + "content": "45\\%" + }, + { + "bbox": [ + 55, + 168, + 295, + 312 + ], + "type": "text", + "content": " reduction in FLOPS with nearly no performance loss. Works since have proposed alterations to this setup such as adaptively determining the number of pruned tokens instead of using a fixed ratio [14] or pruning in multiple stages [39]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 322, + 230, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 322, + 230, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 322, + 230, + 335 + ], + "type": "text", + "content": "3. Revising Visual Token Pruning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 342, + 295, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 295, + 546 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 295, + 546 + ], + "type": "text", + "content": "In this section, we aim to get a better understanding of the vision capabilities of the VLM acceleration approach of pruning visual tokens after shallow LLM layers. After outlining preliminaries (§3.1), we take a closer look at how the approach performs across a broad range of tasks. Upon inspection, we discover that while heavily pruning visual tokens after the shallow LLM layers has little effect on performance across a variety of tasks, this approach fails decisively on more vision-centric tasks, particularly localization (§3.2). Next, we examine why this method struggles with localization, uncovering that the poor performance is due to the ineffectiveness of the pruning criteria to select important tokens when applied at an early layer (§3.3). As this defect is not specific to localization, we explore what can be attributed to the method's high performance on numerous other tasks and find that these tasks require minimal visual grounding (§3.4)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 552, + 142, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 142, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 142, + 564 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 605 + ], + "type": "text", + "content": "Before our analyses, we provide a background on the explored adapter-style VLM with visual token pruning, the evaluated benchmarks, and experimental settings." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "VLM and token pruning. Formally, an adapter-style VLM takes as input an image " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{img}}" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " and text prompt tokens " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{prompt}}" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": ". A pre-trained vision backbone " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " first encodes visual features " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{img}} = f(x_{\\mathrm{img}}) \\in \\mathbb{R}^{n \\times d_{\\mathrm{vision}}}" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " is the number of image patches and " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{vision}}" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " is the dimensionality of the vision encoder. Next, an adapter " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " (either a simple linear layer or MLP) projects the vision features to embeddings " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{img}} = p(z_{\\mathrm{img}}) \\in \\mathbb{R}^{n \\times d_{\\mathrm{text}}}" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{text}}" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " is the dimensionality of the LLM. Lastly, " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{img}}" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " is concatenated with text" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "prompt embeddings " + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{prompt}} = \\mathrm{embed}(x_{\\mathrm{prompt}})" + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": " and passed into the language model LM to generate the output text " + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "inline_equation", + "content": "y = \\mathrm{LM}([h_{\\mathrm{img}}; h_{\\mathrm{prompt}}])" + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 108, + 553, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 553, + 215 + ], + "type": "text", + "content": "In this work, we study the inference acceleration of VLMs where visual tokens are pruned within the attention mechanism of LM. We focus on the FastV [10] approach, where after layer " + }, + { + "bbox": [ + 313, + 108, + 553, + 215 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 108, + 553, + 215 + ], + "type": "text", + "content": " in LM, " + }, + { + "bbox": [ + 313, + 108, + 553, + 215 + ], + "type": "inline_equation", + "content": "R\\%" + }, + { + "bbox": [ + 313, + 108, + 553, + 215 + ], + "type": "text", + "content": " of visual tokens are pruned away based on a ranking function " + }, + { + "bbox": [ + 313, + 108, + 553, + 215 + ], + "type": "inline_equation", + "content": "g_{\\phi}" + }, + { + "bbox": [ + 313, + 108, + 553, + 215 + ], + "type": "text", + "content": " which ranks tokens based on criteria " + }, + { + "bbox": [ + 313, + 108, + 553, + 215 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 108, + 553, + 215 + ], + "type": "text", + "content": ". In practice, the attention score received from the last text token is used as the criteria, referred to as " + }, + { + "bbox": [ + 313, + 108, + 553, + 215 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{original}}" + }, + { + "bbox": [ + 313, + 108, + 553, + 215 + ], + "type": "text", + "content": ". Note that the positional information is preserved when performing the pruning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 215, + 553, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 215, + 553, + 302 + ], + "spans": [ + { + "bbox": [ + 313, + 215, + 553, + 302 + ], + "type": "text", + "content": "With this approach, we measure the acceleration using the theoretical FLOPS reduction ratio related to the image tokens. For one Transformer layer, the total FLOPS is estimated as " + }, + { + "bbox": [ + 313, + 215, + 553, + 302 + ], + "type": "inline_equation", + "content": "C = 4nd^2 + 2n^2d + 2nd_m" + }, + { + "bbox": [ + 313, + 215, + 553, + 302 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 215, + 553, + 302 + ], + "type": "inline_equation", + "content": "d = d_{\\mathrm{text}}" + }, + { + "bbox": [ + 313, + 215, + 553, + 302 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 215, + 553, + 302 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 215, + 553, + 302 + ], + "type": "text", + "content": " is the intermediate size of FNN. Given that the Transformer has " + }, + { + "bbox": [ + 313, + 215, + 553, + 302 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 215, + 553, + 302 + ], + "type": "text", + "content": " layers in total and after layer " + }, + { + "bbox": [ + 313, + 215, + 553, + 302 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 215, + 553, + 302 + ], + "type": "text", + "content": ", we maintain " + }, + { + "bbox": [ + 313, + 215, + 553, + 302 + ], + "type": "inline_equation", + "content": "\\hat{n} = (1 - R\\%) * n" + }, + { + "bbox": [ + 313, + 215, + 553, + 302 + ], + "type": "text", + "content": " visual tokens, we calculate the FLOPS reduction as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 337, + 305, + 553, + 328 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 305, + 553, + 328 + ], + "spans": [ + { + "bbox": [ + 337, + 305, + 553, + 328 + ], + "type": "interline_equation", + "content": "1 - \\frac {K * C + (T - K) * (4 \\hat {n} d ^ {2} + 2 \\hat {n} ^ {2} d + 2 \\hat {n} d m)}{C}. \\tag {1}", + "image_path": "4de912175a0cc9405cef2f0370e5548ef7d91b4167a7e6dd28f26838374ffda8.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 331, + 553, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 553, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 553, + 390 + ], + "type": "text", + "content": "Benchmarking. We evaluate the accelerated VLMs on a large suite of benchmarks from [18] which includes evaluations spanning the areas of localization, open-ended visual question answering, and challenge sets. For all benchmarks, we follow the same evaluation protocol as [18]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 391, + 553, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 391, + 553, + 497 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 553, + 497 + ], + "type": "text", + "content": "Localization. We evaluate localization performance using the RefCOCO, RefCOCO+, RefCOCOg [19, 42] and OCID-Ref [37] datasets. RefCOCO contains short, spatially grounded descriptions, while RefCOCO+ focuses on appearance-based descriptions. RefCOCOg, on the other hand, features longer and more detailed descriptions. OCID-Ref is a robotics dataset that tests generalization to out-of-distribution scenarios, particularly for object localization in cluttered environments." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 498, + 553, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 553, + 581 + ], + "type": "text", + "content": "Open-Ended Visual Question Answering. We evaluate general visual reasoning using the VizWiz [5] and VQAv2 [13] datasets, spatial reasoning using the GQA [16] dataset, and reasoning around text using the TextVQA [32] dataset. For TextVQA, following [18], we exclude OCR-system parsed input tokens to better evaluate the effect of pruning on visual capabilities." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 582, + 553, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 553, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 553, + 653 + ], + "type": "text", + "content": "Challenge Sets. We additionally evaluate on the VSR [26], TallyQA [1], POPE [25], and AI2D [20] closed-set prediction datasets. VSR includes binary spatial relationship questions, TallyQA consists of counting questions, POPE probes hallucination, and AI2D contains multiple-choice questions referring to scientific diagrams and charts." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "Experimental settings. In our experiments, we utilize a VLM with SigLIP ViT-SO400M [43] as " + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": ", a one-layer MLP with GELU activation as " + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": ", and Llama 2 7B [35] as LM. The model was trained on the multimodal instruction tuning dataset presented in [27] in a single-stage with " + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": " frozen." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22828" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 77, + 211, + 199 + ], + "blocks": [ + { + "bbox": [ + 56, + 77, + 211, + 199 + ], + "lines": [ + { + "bbox": [ + 56, + 77, + 211, + 199 + ], + "spans": [ + { + "bbox": [ + 56, + 77, + 211, + 199 + ], + "type": "image", + "image_path": "9f876d33145c48301090e7ce73fff2709a0513bc1bcc71d3f92ce0bec797b033.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 212, + 555, + 258 + ], + "lines": [ + { + "bbox": [ + 54, + 212, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 54, + 212, + 555, + 258 + ], + "type": "text", + "content": "Figure 2. Contrasting the difference in performance dropoff on the challenging vision-centric localization task (Left) versus the other evaluated tasks (Middle) when pruning visual tokens after the shallow LLM layers. Whereas performance decrease is minimal for most tasks, localization exhibits roughly a linear decrease to zero as the ratio of pruned tokens increases. Right: Per-task performance breakdown across various setups of pruning ratios. Using " + }, + { + "bbox": [ + 54, + 212, + 555, + 258 + ], + "type": "inline_equation", + "content": "K = 3" + }, + { + "bbox": [ + 54, + 212, + 555, + 258 + ], + "type": "text", + "content": " for all pruning setups." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 212, + 78, + 369, + 199 + ], + "blocks": [ + { + "bbox": [ + 212, + 78, + 369, + 199 + ], + "lines": [ + { + "bbox": [ + 212, + 78, + 369, + 199 + ], + "spans": [ + { + "bbox": [ + 212, + 78, + 369, + 199 + ], + "type": "image", + "image_path": "286cdba0b0aa6f7e1371ab0a3b257a79e651aa61adad2d8e7c0d50598ed9a34b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 391, + 71, + 553, + 204 + ], + "blocks": [ + { + "bbox": [ + 391, + 71, + 553, + 204 + ], + "lines": [ + { + "bbox": [ + 391, + 71, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 391, + 71, + 553, + 204 + ], + "type": "image", + "image_path": "4f4bf20589ae09615e518ba86872a1f80bc19be70d89303d0cb3e26acde805b6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 277, + 295, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 295, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 295, + 301 + ], + "type": "text", + "content": "3.2. Impact of early visual token pruning varies drastically by task" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 307, + 295, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 307, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 307, + 295, + 354 + ], + "type": "text", + "content": "We examine the effect of pruning visual tokens after the shallow LLM layers " + }, + { + "bbox": [ + 55, + 307, + 295, + 354 + ], + "type": "inline_equation", + "content": "(K = 3)" + }, + { + "bbox": [ + 55, + 307, + 295, + 354 + ], + "type": "text", + "content": " across a variety of tasks. We compare various pruning ratios " + }, + { + "bbox": [ + 55, + 307, + 295, + 354 + ], + "type": "inline_equation", + "content": "(R \\in \\{0.25, 0.5, 0.75, 0.9\\})" + }, + { + "bbox": [ + 55, + 307, + 295, + 354 + ], + "type": "text", + "content": " and the baseline non-pruned model in Figure 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 356, + 296, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 356, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 56, + 356, + 296, + 557 + ], + "type": "text", + "content": "For the majority of evaluated tasks, heavily reducing the number of visual tokens results in minimal performance decrease (3.5% for AI2D, 0.1% for VSR, 3.7% for TallyQA, 4.8% for POPE, 5.1% for VizWiz, 7.9% for VQAv2 and 7.7% for GQA when dropping 75% of tokens). However, for a small subset of tasks—particularly TextVQA, and more severely, localization-performance drops substantially, with decreases of 42.0%, 88.9%, 88.9%, 91.0%, and 86.0% for TextVQA, RefCOCO, RefCOCO+, RefCOCOg, and OCID-Ref, respectively. Investigating further, Figure 2 shows that localization performance declines roughly linearly to zero as the pruning ratio progresses from 0 to 1, a pattern that strongly contrasts with most other tasks, where performance remains largely unaffected. While prior works report a much smaller decline in TextVQA performance [9, 39], they incorporate OCR-parsed input tokens into prompts, reducing the reliance on visual information." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": "Discussion. Although localization is undoubtedly challenging for a model that discards visual tokens, as it requires fine-grained visual grounding and precise boundary identification, it is still surprising that (1) localization performance declines to the extent it does even when many tokens are maintained (e.g., dropping " + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": " of tokens reduces average performance by " + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "inline_equation", + "content": "45\\%" + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": "), and (2) most non-localization benchmarks, except for TextVQA, exhibit a drastically different pattern despite these benchmarks aiming to evaluate visual grounding abilities, such as counting for TallyQA, spatial reasoning for GQA, and chart understanding for AI2D. In the following sections, we seek to explain these unexpected findings and understand their implications." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 277, + 555, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 555, + 300 + ], + "type": "text", + "content": "3.3. Interpreting poor vision-centric task performance" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 307, + 555, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 555, + 426 + ], + "type": "text", + "content": "We next investigate why pruning visual tokens after the shallow LLM layers has such an adverse effect on localization performance. Intuitively, even localization should not require all image tokens, as there are background tokens irrelevant to identifying specific objects that can be pruned. Thus, it is surprising that performance begins to degrade even with minimal token pruning. This suggests that the pruning criteria, intended to retain important tokens, fails to effectively distinguish which tokens contain necessary visual information." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 426, + 556, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 556, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 556, + 569 + ], + "type": "text", + "content": "To assess the efficacy of the pruning criteria, we first examine the distribution of retained tokens across all benchmark examples. As shown in Figure 3(b), we find that pruning visual tokens after the shallow LLM layers " + }, + { + "bbox": [ + 313, + 426, + 556, + 569 + ], + "type": "inline_equation", + "content": "(K = 3)" + }, + { + "bbox": [ + 313, + 426, + 556, + 569 + ], + "type": "text", + "content": " retains tokens concentrated at the bottom of the image. For example, when " + }, + { + "bbox": [ + 313, + 426, + 556, + 569 + ], + "type": "inline_equation", + "content": "R = 0.75" + }, + { + "bbox": [ + 313, + 426, + 556, + 569 + ], + "type": "text", + "content": ", the average sampled token position across all datasets is situated " + }, + { + "bbox": [ + 313, + 426, + 556, + 569 + ], + "type": "inline_equation", + "content": "80.7\\%" + }, + { + "bbox": [ + 313, + 426, + 556, + 569 + ], + "type": "text", + "content": " of the way down the image. A Chi-Square test confirms the non-uniformity of selected tokens' y-positions (" + }, + { + "bbox": [ + 313, + 426, + 556, + 569 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 426, + 556, + 569 + ], + "type": "text", + "content": "-value " + }, + { + "bbox": [ + 313, + 426, + 556, + 569 + ], + "type": "inline_equation", + "content": "< 0.05" + }, + { + "bbox": [ + 313, + 426, + 556, + 569 + ], + "type": "text", + "content": "). We also examine the criteria behavior when pruning after later layers (" + }, + { + "bbox": [ + 313, + 426, + 556, + 569 + ], + "type": "inline_equation", + "content": "K \\in (8,16,24)" + }, + { + "bbox": [ + 313, + 426, + 556, + 569 + ], + "type": "text", + "content": ") and observe that the bias of selecting bottom image tokens is reduced." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": "In addition, we observe that as the pruning layer increases, not only is the positional bias reduced, but the criteria also begins to correctly select tokens relevant to the text instruction. For example, as shown in Figure 3(a), pruning after the later layers results in a distinct concentration of selected tokens around the area corresponding to the reference expression for localization. To quantify this change in criteria effectiveness across layers, we measure the performance when varying the pruning layer " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "(K\\in \\{3,8,16,24\\})" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": " As shown in Figure 3(c), performance remains stable when pruning is performed at deeper layers " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "K = 16" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "K =" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": " 24), whereas pruning at earlier layers " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "K = 3" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "K = 8" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "22829" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 74, + 72, + 145, + 145 + ], + "blocks": [ + { + "bbox": [ + 58, + 76, + 69, + 88 + ], + "lines": [ + { + "bbox": [ + 58, + 76, + 69, + 88 + ], + "spans": [ + { + "bbox": [ + 58, + 76, + 69, + 88 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 74, + 72, + 145, + 145 + ], + "lines": [ + { + "bbox": [ + 74, + 72, + 145, + 145 + ], + "spans": [ + { + "bbox": [ + 74, + 72, + 145, + 145 + ], + "type": "image", + "image_path": "72a40db4bc8d989578cc28510d68d7755ecdb5e3bb820e170b9fbd971f1e7047.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 156, + 72, + 226, + 146 + ], + "blocks": [ + { + "bbox": [ + 156, + 72, + 226, + 146 + ], + "lines": [ + { + "bbox": [ + 156, + 72, + 226, + 146 + ], + "spans": [ + { + "bbox": [ + 156, + 72, + 226, + 146 + ], + "type": "image", + "image_path": "f8b3451ae6cb8d90bbd59c738920d5f151d2634ee388ec2d74cd64afe48a1763.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 231, + 72, + 244, + 87 + ], + "lines": [ + { + "bbox": [ + 231, + 72, + 244, + 87 + ], + "spans": [ + { + "bbox": [ + 231, + 72, + 244, + 87 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 75, + 153, + 144, + 228 + ], + "blocks": [ + { + "bbox": [ + 75, + 153, + 144, + 228 + ], + "lines": [ + { + "bbox": [ + 75, + 153, + 144, + 228 + ], + "spans": [ + { + "bbox": [ + 75, + 153, + 144, + 228 + ], + "type": "image", + "image_path": "f4306611f75804fbbac25d2b58ad4a9d0cade5402eacdf25ba9177093104e7fc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 230, + 211, + 238 + ], + "lines": [ + { + "bbox": [ + 85, + 230, + 211, + 238 + ], + "spans": [ + { + "bbox": [ + 85, + 230, + 211, + 238 + ], + "type": "text", + "content": "Reference expression: a bowl of blueberries" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 156, + 153, + 226, + 227 + ], + "blocks": [ + { + "bbox": [ + 156, + 153, + 226, + 227 + ], + "lines": [ + { + "bbox": [ + 156, + 153, + 226, + 227 + ], + "spans": [ + { + "bbox": [ + 156, + 153, + 226, + 227 + ], + "type": "image", + "image_path": "4b89b253336473434b20c3d8fb99788acf6767c1652f44b5df2bfe6c9704de6c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 246, + 72, + 328, + 145 + ], + "blocks": [ + { + "bbox": [ + 246, + 72, + 328, + 145 + ], + "lines": [ + { + "bbox": [ + 246, + 72, + 328, + 145 + ], + "spans": [ + { + "bbox": [ + 246, + 72, + 328, + 145 + ], + "type": "image", + "image_path": "38e0177f484f621655db8b2c810f33c3284f3c5c2be5e90b4bca068413632f90.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 280, + 147, + 300, + 156 + ], + "lines": [ + { + "bbox": [ + 280, + 147, + 300, + 156 + ], + "spans": [ + { + "bbox": [ + 280, + 147, + 300, + 156 + ], + "type": "text", + "content": "x position K-16" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 246, + 157, + 328, + 226 + ], + "blocks": [ + { + "bbox": [ + 246, + 157, + 328, + 226 + ], + "lines": [ + { + "bbox": [ + 246, + 157, + 328, + 226 + ], + "spans": [ + { + "bbox": [ + 246, + 157, + 328, + 226 + ], + "type": "image", + "image_path": "80b6d7be38344bc1ce262474fdcc32da3b9c5dd1cd3409bdf9e8bd20c063a787.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 282, + 228, + 301, + 233 + ], + "lines": [ + { + "bbox": [ + 282, + 228, + 301, + 233 + ], + "spans": [ + { + "bbox": [ + 282, + 228, + 301, + 233 + ], + "type": "text", + "content": "x position" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 335, + 72, + 419, + 145 + ], + "blocks": [ + { + "bbox": [ + 335, + 72, + 419, + 145 + ], + "lines": [ + { + "bbox": [ + 335, + 72, + 419, + 145 + ], + "spans": [ + { + "bbox": [ + 335, + 72, + 419, + 145 + ], + "type": "image", + "image_path": "d85b30c96cfc96f8dd07fc317f4f7e2b89fd514cce55bb57f70c7e4e1dde6c89.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 371, + 147, + 391, + 157 + ], + "lines": [ + { + "bbox": [ + 371, + 147, + 391, + 157 + ], + "spans": [ + { + "bbox": [ + 371, + 147, + 391, + 157 + ], + "type": "text", + "content": "x position " + }, + { + "bbox": [ + 371, + 147, + 391, + 157 + ], + "type": "inline_equation", + "content": "K = 24" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 335, + 158, + 419, + 227 + ], + "blocks": [ + { + "bbox": [ + 335, + 158, + 419, + 227 + ], + "lines": [ + { + "bbox": [ + 335, + 158, + 419, + 227 + ], + "spans": [ + { + "bbox": [ + 335, + 158, + 419, + 227 + ], + "type": "image", + "image_path": "8e979172f1352d135f6b68fc32a7be4d7be7bc91db119e4c6bdfb4c56c011131.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 371, + 228, + 392, + 233 + ], + "lines": [ + { + "bbox": [ + 371, + 228, + 392, + 233 + ], + "spans": [ + { + "bbox": [ + 371, + 228, + 392, + 233 + ], + "type": "text", + "content": "x position" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 284, + 236, + 391, + 256 + ], + "blocks": [ + { + "bbox": [ + 284, + 236, + 391, + 256 + ], + "lines": [ + { + "bbox": [ + 284, + 236, + 391, + 256 + ], + "spans": [ + { + "bbox": [ + 284, + 236, + 391, + 256 + ], + "type": "image", + "image_path": "f1c771f753620cb6569dbf5c32c855866fef6378fe644ea4594a4efa2992a3f3.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 266, + 555, + 332 + ], + "lines": [ + { + "bbox": [ + 54, + 266, + 555, + 332 + ], + "spans": [ + { + "bbox": [ + 54, + 266, + 555, + 332 + ], + "type": "text", + "content": "Figure 3. (a) Example demonstrating that when pruning in early layers, selected tokens are concentrated in the bottom of the image, whereas in later layers the selected tokens more precisely cover the region of the image important for answering the question. (b) Heatmap illustrating that averaged across all benchmark examples, as the pruning layer increases, the selection of bottom visual tokens by the criteria is reduced. (c) Visualizing the effect of the pruning layer on performance for both localization and non-localization tasks. We find that pruning after layer 16 or later results in little performance dropoff, whereas pruning earlier results in a performance decrease, particularly for localization. Using " + }, + { + "bbox": [ + 54, + 266, + 555, + 332 + ], + "type": "inline_equation", + "content": "R = 0.75" + }, + { + "bbox": [ + 54, + 266, + 555, + 332 + ], + "type": "text", + "content": " for all setups." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 443, + 73, + 552, + 159 + ], + "blocks": [ + { + "bbox": [ + 431, + 77, + 441, + 87 + ], + "lines": [ + { + "bbox": [ + 431, + 77, + 441, + 87 + ], + "spans": [ + { + "bbox": [ + 431, + 77, + 441, + 87 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 443, + 73, + 552, + 159 + ], + "lines": [ + { + "bbox": [ + 443, + 73, + 552, + 159 + ], + "spans": [ + { + "bbox": [ + 443, + 73, + 552, + 159 + ], + "type": "image", + "image_path": "ee6aacba1a99a9860ed9dd5a5ced81005abc1f80fc85a078c3cd09f037284089.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 441, + 171, + 553, + 251 + ], + "blocks": [ + { + "bbox": [ + 484, + 164, + 525, + 171 + ], + "lines": [ + { + "bbox": [ + 484, + 164, + 525, + 171 + ], + "spans": [ + { + "bbox": [ + 484, + 164, + 525, + 171 + ], + "type": "text", + "content": "Non-localization" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 441, + 171, + 553, + 251 + ], + "lines": [ + { + "bbox": [ + 441, + 171, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 441, + 171, + 553, + 251 + ], + "type": "image", + "image_path": "f39e6d530bafc921bbca4dbb67e4aa6c80c528f3b91748e4391e3276b911ce4d.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 54, + 354, + 295, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 354, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 54, + 354, + 295, + 376 + ], + "type": "text", + "content": "leads to performance degradation, particularly for localization." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 56, + 378, + 296, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 378, + 296, + 640 + ], + "spans": [ + { + "bbox": [ + 56, + 378, + 296, + 640 + ], + "type": "text", + "content": "Discussion. While previous research attributes the poor performance of early token pruning to the low redundancy of image tokens in shallow layers [39], we reveal a critical insight: the variable performance across pruning layers is linked to the effectiveness of selecting important tokens, as early token pruning leads to suboptimal token selection biased toward bottom tokens. To understand why the criteria disproportionately retains bottom image tokens, we must consider the LM architecture. Specifically, token selection is based on the attention received from the last text token. Since the LM uses RoPE to encode positional information [33], it exhibits a long-term decay property. While this property is well-suited for language modeling, we uncover that it poses an issue for selecting visual tokens, where tokens appearing later in the flattened, raster-scan order have higher attention scores purely due to their position. In language modeling, prior work finds that there is greater emphasis on shorter-distance information in shallow layers [15]. We predict that the visually-conditioned LM exhibits the same behavior, explaining why the positional bias is most pronounced when pruning after shallow layers." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 648, + 293, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 648, + 293, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 648, + 293, + 661 + ], + "type": "text", + "content": "3.4. Explaining strong performance on other tasks" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": "Given that early pruning of visual tokens leads to a suboptimal tokens selection where remaining tokens are concentrated towards the bottom of the image, a natural question arises: how does the approach still maintain high perfor" + } + ] + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 319, + 354, + 550, + 518 + ], + "blocks": [ + { + "bbox": [ + 319, + 354, + 550, + 518 + ], + "lines": [ + { + "bbox": [ + 319, + 354, + 550, + 518 + ], + "spans": [ + { + "bbox": [ + 319, + 354, + 550, + 518 + ], + "type": "image", + "image_path": "f3dc8989efff27dd6c48a5ff61c10977b787a5d7dabe9faca6396a03a895e550.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 530, + 555, + 619 + ], + "lines": [ + { + "bbox": [ + 313, + 530, + 555, + 619 + ], + "spans": [ + { + "bbox": [ + 313, + 530, + 555, + 619 + ], + "type": "text", + "content": "Figure 4. Assessing whether the strong performance of early visual token pruning for many tasks can be attributed to visual information transfer before pruning or benchmarks' lack of assessing fine-grained visual capabilities. We observe that for all evaluated tasks, allowing information transfer before pruning (shown in green) does not result in substantial performance improvement over the setup without visual information transfer (shown in light green), highlighting a limitation of many benchmarks." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 639, + 554, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 554, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 554, + 676 + ], + "type": "text", + "content": "mance across a diverse range of tasks, including those aiming to evaluate visual grounding abilities? In this section, we test the following contrasting explanations:" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 314, + 677, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 677, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 314, + 677, + 555, + 712 + ], + "type": "text", + "content": "1. Important visual information from top image tokens is transferred to later tokens before pruning with the unidirectional attention of the LLM." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22830" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "2. Many questions can still be inferred with access to only the suboptimal set of visual tokens primarily from the bottom of the image." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 110, + 295, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 110, + 295, + 241 + ], + "spans": [ + { + "bbox": [ + 55, + 110, + 295, + 241 + ], + "type": "text", + "content": "To evaluate these hypotheses across tasks, we save the tokens retained by the studied pruning approach and, in a subsequent run, remove all other tokens before they reach the LLM, ensuring that no visual information from the pruned tokens is retained. As an additional comparison, we also evaluate a text-only model in which all visual tokens are removed before entering the LLM. As shown in Figure 4, we find that across all evaluated tasks, removing pruned tokens before the LLM results in little to no performance decrease compared to the method that allows information transfer in shallow layers prior to pruning." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 242, + 295, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 242, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 242, + 295, + 373 + ], + "type": "text", + "content": "Discussion. This finding indicates that the high performance of early visual token pruning on many tasks does not stem from the effectiveness of visual information transfer in early layers but rather signifies that many benchmarks do not demand a detailed understanding of visual information to answer questions accurately. Note that for the majority of these benchmarks (except VizWiz and AI2D), performance substantially declines when not including any visual information, signifying that simply comparing vision-enabled and disabled setups is insufficient to determine a benchmark's ability to assess visual grounding." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 382, + 239, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 382, + 239, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 382, + 239, + 396 + ], + "type": "text", + "content": "4. Improving Visual Token Pruning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 403, + 295, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 295, + 534 + ], + "type": "text", + "content": "Based on our findings from §3, we now seek to answer: can attention serve as an effective criteria for token pruning even in early layers? To this end, we propose a straightforward solution to removing the positional bias of the attention-based criteria and demonstrate that our modifications enable effective pruning even when applied after early LLM layers (§4.1). Guided by our insights, we present our final approach, FEATHER (Fast and Effective Acceleration with Ensemble cRicerca) (§4.2), and showcase its impressive visual capabilities while offering high computational efficiency (§4.3)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 540, + 291, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 291, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 291, + 553 + ], + "type": "text", + "content": "4.1. Assessing potential of attention-based criteria" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 558, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 665 + ], + "type": "text", + "content": "An attention-based criteria offers the potential to naturally identify and retain only the most relevant visual tokens for answering a question. However, as shown in §3, positional bias strongly influences token selection, limiting its effectiveness. Thus, we aim to resolve the positional bias issue and explore the true promise of attention-based criteria. We compare against non-attention-based approaches that incorporate an explicit inductive bias toward reducing redundancy. Below, we outline the explored criteria." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 666, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\phi_{-R}" + }, + { + "bbox": [ + 55, + 666, + 295, + 714 + ], + "type": "text", + "content": ": Given the long-term decay property of RoPE, we propose the following straightforward adjustment to the pruning criteria in order to mitigate this bias and more effectively select tokens relevant to the text instruction. Namely," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "we still compute the attention score received from the last text token, except we do not apply RoPE to the attention mechanism, thereby removing the long-term decay effect. Note that this lightweight calculation is only done once per pruning, and all other state computations are compatible with FlashAttention [12]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 144, + 553, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 553, + 238 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 553, + 238 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{uniform}}" + }, + { + "bbox": [ + 313, + 144, + 553, + 238 + ], + "type": "text", + "content": ": In this simple criteria, we uniformly sample visual tokens in the image, using a set stride. This criteria ensures that there is good image coverage, but does so by sacrificing the ability to have more densely captured visual information for a particular image region. To use a similar computational saving to the attention-based criteria when " + }, + { + "bbox": [ + 313, + 144, + 553, + 238 + ], + "type": "inline_equation", + "content": "R = 0.75" + }, + { + "bbox": [ + 313, + 144, + 553, + 238 + ], + "type": "text", + "content": ", we use a stride of two, resulting in 196 selected tokens." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 239, + 553, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 553, + 294 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 553, + 294 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{KNN}}" + }, + { + "bbox": [ + 313, + 239, + 553, + 294 + ], + "type": "text", + "content": ": Inspired by dynamic visual tokenization [17], we select tokens based on local density. For each visual token " + }, + { + "bbox": [ + 313, + 239, + 553, + 294 + ], + "type": "inline_equation", + "content": "z_{img}[i]" + }, + { + "bbox": [ + 313, + 239, + 553, + 294 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 239, + 553, + 294 + ], + "type": "inline_equation", + "content": "i \\in \\{0, 1, \\dots, n-1\\}" + }, + { + "bbox": [ + 313, + 239, + 553, + 294 + ], + "type": "text", + "content": ", we compute its local density " + }, + { + "bbox": [ + 313, + 239, + 553, + 294 + ], + "type": "inline_equation", + "content": "\\rho_i" + }, + { + "bbox": [ + 313, + 239, + 553, + 294 + ], + "type": "text", + "content": " using K-nearest neighbors and calculate distance index " + }, + { + "bbox": [ + 313, + 239, + 553, + 294 + ], + "type": "inline_equation", + "content": "\\delta_i" + }, + { + "bbox": [ + 313, + 239, + 553, + 294 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 322, + 300, + 553, + 336 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 300, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 322, + 300, + 553, + 336 + ], + "type": "interline_equation", + "content": "\\delta_ {i} = \\left\\{ \\begin{array}{l l} \\min _ {j: \\rho_ {j} > \\rho_ {i}} \\| z _ {i m g} [ i ] - z _ {i m g} [ j ] \\| ^ {2}, & \\text {i f} \\exists j \\text {s . t .} \\rho_ {j} > \\rho_ {i}. \\\\ \\max _ {j} \\| z _ {i m g} [ i ] - z _ {i m g} [ j ] \\| ^ {2}, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {2}", + "image_path": "70eccc676d4c4fa5dd0c9923f5d2c145c14957c7ef2d677f14bf94278ef7bc78.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 342, + 553, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 553, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 553, + 378 + ], + "type": "inline_equation", + "content": "\\delta_{i}" + }, + { + "bbox": [ + 313, + 342, + 553, + 378 + ], + "type": "text", + "content": " represents how far away " + }, + { + "bbox": [ + 313, + 342, + 553, + 378 + ], + "type": "inline_equation", + "content": "z_{img}[i]" + }, + { + "bbox": [ + 313, + 342, + 553, + 378 + ], + "type": "text", + "content": " is from other high-density tokens. We use " + }, + { + "bbox": [ + 313, + 342, + 553, + 378 + ], + "type": "inline_equation", + "content": "\\rho_{i}*\\delta_{i}" + }, + { + "bbox": [ + 313, + 342, + 553, + 378 + ], + "type": "text", + "content": " as the token importance score for the criteria and select 196 tokens." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 379, + 553, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 379, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 553, + 449 + ], + "type": "inline_equation", + "content": "\\phi_{-R} + \\phi_{\\mathrm{uniform}}" + }, + { + "bbox": [ + 313, + 379, + 553, + 449 + ], + "type": "text", + "content": ": Since " + }, + { + "bbox": [ + 313, + 379, + 553, + 449 + ], + "type": "inline_equation", + "content": "\\phi_{-R}" + }, + { + "bbox": [ + 313, + 379, + 553, + 449 + ], + "type": "text", + "content": " focuses on selecting important tokens and " + }, + { + "bbox": [ + 313, + 379, + 553, + 449 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{uniform}}" + }, + { + "bbox": [ + 313, + 379, + 553, + 449 + ], + "type": "text", + "content": " aims to reduce redundancy, we propose an ensemble of these two to leverage their respective strengths. Specifically, we apply " + }, + { + "bbox": [ + 313, + 379, + 553, + 449 + ], + "type": "inline_equation", + "content": "\\phi_{-R}" + }, + { + "bbox": [ + 313, + 379, + 553, + 449 + ], + "type": "text", + "content": " while incorporating a small number of uniformly sampled tokens using a stride of three." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 450, + 553, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 553, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 553, + 605 + ], + "type": "text", + "content": "Improvement of removing RoPE: In Table 1, we report results evaluating the effectiveness of these criteria. Comparing " + }, + { + "bbox": [ + 313, + 450, + 553, + 605 + ], + "type": "inline_equation", + "content": "\\phi_{-R}" + }, + { + "bbox": [ + 313, + 450, + 553, + 605 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 450, + 553, + 605 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{original}}" + }, + { + "bbox": [ + 313, + 450, + 553, + 605 + ], + "type": "text", + "content": ", we find that by removing RoPE, at " + }, + { + "bbox": [ + 313, + 450, + 553, + 605 + ], + "type": "inline_equation", + "content": "K = 3" + }, + { + "bbox": [ + 313, + 450, + 553, + 605 + ], + "type": "text", + "content": " there is a " + }, + { + "bbox": [ + 313, + 450, + 553, + 605 + ], + "type": "inline_equation", + "content": "183\\%" + }, + { + "bbox": [ + 313, + 450, + 553, + 605 + ], + "type": "text", + "content": " average improvement on localization tasks and at " + }, + { + "bbox": [ + 313, + 450, + 553, + 605 + ], + "type": "inline_equation", + "content": "K = 8" + }, + { + "bbox": [ + 313, + 450, + 553, + 605 + ], + "type": "text", + "content": ", there is a " + }, + { + "bbox": [ + 313, + 450, + 553, + 605 + ], + "type": "inline_equation", + "content": "17\\%" + }, + { + "bbox": [ + 313, + 450, + 553, + 605 + ], + "type": "text", + "content": " average improvement. These performance improvements demonstrate that once the impact of token position on attention is removed, attention score can more effectively be used for the criteria when applied after early LLM layers. Note that the narrowing gap in performance between " + }, + { + "bbox": [ + 313, + 450, + 553, + 605 + ], + "type": "inline_equation", + "content": "\\phi_{-R}" + }, + { + "bbox": [ + 313, + 450, + 553, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 450, + 553, + 605 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{original}}" + }, + { + "bbox": [ + 313, + 450, + 553, + 605 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 450, + 553, + 605 + ], + "type": "inline_equation", + "content": "K = 3" + }, + { + "bbox": [ + 313, + 450, + 553, + 605 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 450, + 553, + 605 + ], + "type": "inline_equation", + "content": "K = 8" + }, + { + "bbox": [ + 313, + 450, + 553, + 605 + ], + "type": "text", + "content": " is likely because " + }, + { + "bbox": [ + 313, + 450, + 553, + 605 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{original}}" + }, + { + "bbox": [ + 313, + 450, + 553, + 605 + ], + "type": "text", + "content": " has less of a bias towards selecting bottom image tokens as the pruning layer increases." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "text", + "content": "Token selection improves when pruning later: Comparing " + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\phi_{-R}" + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "inline_equation", + "content": "K = 8" + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "text", + "content": " versus " + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "inline_equation", + "content": "K = 3" + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "text", + "content": ", we see that the average localization performance improves by " + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "inline_equation", + "content": "63\\%" + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "text", + "content": ". However, it remains unclear whether this performance increase is affected by factors other than the ability of the criteria to select important tokens. Therefore, to directly compare the criteria across different layers, we take the tokens selected by both criteria and evaluate how well the model can perform with only these tokens passed into the LLM (as is" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22831" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 68, + 555, + 364 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 555, + 364 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 555, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 555, + 364 + ], + "type": "table", + "html": "
Pruning LayerCriteriaFLOPS RedLocalizationOpen-Ended VQAChallenge Sets
AvgOCID-RefRefCOCObRefCOCO+RefCOCOAvgTextVQAGQAVQAv2VizWizAvgPOPETallyQAVSRAI2D
K=3Attention-based
φoriginal68%5.95.75.16.16.754.831.858.472.756.364.083.257.163.352.4
φR(Ours)68%16.722.915.113.315.359.041.661.276.057.364.785.258.262.253.2
Δ+10.7+17.2+10.0+7.3+8.6+4.2+9.7+2.8+3.2+1.1+0.7+1.9+1.1-1.1+0.8
Non-attention-based
φKNN66%23.915.124.926.029.658.439.960.974.458.462.881.255.961.552.8
φuniform66%28.020.628.629.733.359.041.461.875.957.164.685.258.161.953.0
Ensemble
φR+φuniform(Ours)61%27.229.127.224.727.761.246.662.377.458.465.486.058.962.754.0
K=8Attention-based
φoriginal56%23.319.423.524.026.359.845.060.376.157.864.685.457.562.653.0
φR(Ours)56%27.327.126.726.429.261.449.061.577.457.865.586.758.663.053.7
Δ+4.0+7.6+3.2+2.5+2.9+1.6+4.0+1.2+1.3+0.0+0.8+1.3+1.1+0.4+0.6
Non-attention-based
φKNN55%23.615.424.425.229.458.640.261.174.558.562.981.456.260.953.0
φuniform55%30.324.631.030.934.859.342.261.876.057.464.485.357.961.053.2
Ensemble
φR+φuniform(Ours)50%35.632.035.935.438.862.751.762.478.158.666.087.459.163.654.0
", + "image_path": "e23128864cf3cc686470aa0ab3d5fc22c58669e2d99905958f5beb449f71a20a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 444, + 291, + 492 + ], + "blocks": [ + { + "bbox": [ + 55, + 370, + 555, + 427 + ], + "lines": [ + { + "bbox": [ + 55, + 370, + 555, + 427 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 555, + 427 + ], + "type": "text", + "content": "Table 1. Evaluating alternative criteria for token pruning after the early LLM layers. For each task and pruning layer, we **bold** the best result and **underline** the second-best result. Our main findings include: (1) our proposed RoPE-free criteria " + }, + { + "bbox": [ + 55, + 370, + 555, + 427 + ], + "type": "inline_equation", + "content": "\\phi_{-R}" + }, + { + "bbox": [ + 55, + 370, + 555, + 427 + ], + "type": "text", + "content": " substantially improves pruning performance compared to the original criteria " + }, + { + "bbox": [ + 55, + 370, + 555, + 427 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{original}}" + }, + { + "bbox": [ + 55, + 370, + 555, + 427 + ], + "type": "text", + "content": "; (2) pruning later (" + }, + { + "bbox": [ + 55, + 370, + 555, + 427 + ], + "type": "inline_equation", + "content": "K = 8" + }, + { + "bbox": [ + 55, + 370, + 555, + 427 + ], + "type": "text", + "content": ") yields higher performance than pruning earlier (" + }, + { + "bbox": [ + 55, + 370, + 555, + 427 + ], + "type": "inline_equation", + "content": "K = 3" + }, + { + "bbox": [ + 55, + 370, + 555, + 427 + ], + "type": "text", + "content": "); and (3) integrating uniform sampling into the attention-based criteria with " + }, + { + "bbox": [ + 55, + 370, + 555, + 427 + ], + "type": "inline_equation", + "content": "\\phi_{-R} + \\phi_{\\mathrm{uniform}}" + }, + { + "bbox": [ + 55, + 370, + 555, + 427 + ], + "type": "text", + "content": " enhances effectiveness. Using " + }, + { + "bbox": [ + 55, + 370, + 555, + 427 + ], + "type": "inline_equation", + "content": "R = 0.75" + }, + { + "bbox": [ + 55, + 370, + 555, + 427 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 370, + 555, + 427 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{original}}" + }, + { + "bbox": [ + 55, + 370, + 555, + 427 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 370, + 555, + 427 + ], + "type": "inline_equation", + "content": "\\phi_{-R}" + }, + { + "bbox": [ + 55, + 370, + 555, + 427 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 370, + 555, + 427 + ], + "type": "inline_equation", + "content": "\\phi_{-R} + \\phi_{\\mathrm{uniform}}" + }, + { + "bbox": [ + 55, + 370, + 555, + 427 + ], + "type": "text", + "content": ". See §4.1 for criteria definitions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 444, + 291, + 492 + ], + "lines": [ + { + "bbox": [ + 58, + 444, + 291, + 492 + ], + "spans": [ + { + "bbox": [ + 58, + 444, + 291, + 492 + ], + "type": "table", + "html": "
KOCID-RefRefCOCOgRefCOCO+RefCOCO
323.816.214.516.3
826.729.826.829.8
", + "image_path": "4c27e9f0a13c1d9fa01d9fead10fe2d571a9a67486b5057f2bda53264bf9cd50.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 499, + 295, + 534 + ], + "lines": [ + { + "bbox": [ + 55, + 499, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 499, + 295, + 534 + ], + "type": "text", + "content": "Table 2. Assessing localization performance when using the selected tokens from " + }, + { + "bbox": [ + 55, + 499, + 295, + 534 + ], + "type": "inline_equation", + "content": "\\phi_{-R}" + }, + { + "bbox": [ + 55, + 499, + 295, + 534 + ], + "type": "text", + "content": " for the entirety of the LLM. We find that " + }, + { + "bbox": [ + 55, + 499, + 295, + 534 + ], + "type": "inline_equation", + "content": "\\phi_{-R}" + }, + { + "bbox": [ + 55, + 499, + 295, + 534 + ], + "type": "text", + "content": " applied at a later LLM layer results in a better token selection." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 556, + 295, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 295, + 592 + ], + "type": "text", + "content": "done in §3.4). As shown in Table 2, we find the selected tokens from " + }, + { + "bbox": [ + 55, + 556, + 295, + 592 + ], + "type": "inline_equation", + "content": "\\phi_{-R}" + }, + { + "bbox": [ + 55, + 556, + 295, + 592 + ], + "type": "text", + "content": " applied after layer 8 are superior compared to the tokens from " + }, + { + "bbox": [ + 55, + 556, + 295, + 592 + ], + "type": "inline_equation", + "content": "\\phi_{-R}" + }, + { + "bbox": [ + 55, + 556, + 295, + 592 + ], + "type": "text", + "content": " applied after layer 3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": "Integrating uniform sampling with attention-based criteria is beneficial in early layers: When comparing the attention-based criteria " + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\phi_{-R}" + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": " with the two non-attention based criteria " + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{KNN}}" + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{uniform}}" + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": ", we find that " + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\phi_{-R}" + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": " results in better performance on some tasks (e.g., OCID-Ref, TextVQA), while " + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{KNN}}" + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{uniform}}" + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": " outperform on other tasks (e.g., RefCOCO). This varying effectiveness of criteria types presumably comes from some tasks requiring a more detailed understanding of specific image regions while others benefit from a broader understanding with full image" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 447, + 555, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 555, + 567 + ], + "type": "text", + "content": "coverage. When we combine these two criteria types with " + }, + { + "bbox": [ + 313, + 447, + 555, + 567 + ], + "type": "inline_equation", + "content": "\\phi_{-R} + \\phi_{\\mathrm{uniform}}" + }, + { + "bbox": [ + 313, + 447, + 555, + 567 + ], + "type": "text", + "content": ", we find that when " + }, + { + "bbox": [ + 313, + 447, + 555, + 567 + ], + "type": "inline_equation", + "content": "K = 3" + }, + { + "bbox": [ + 313, + 447, + 555, + 567 + ], + "type": "text", + "content": " this approach improves upon localization performance of " + }, + { + "bbox": [ + 313, + 447, + 555, + 567 + ], + "type": "inline_equation", + "content": "\\phi_{-R}" + }, + { + "bbox": [ + 313, + 447, + 555, + 567 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 313, + 447, + 555, + 567 + ], + "type": "inline_equation", + "content": "63\\%" + }, + { + "bbox": [ + 313, + 447, + 555, + 567 + ], + "type": "text", + "content": " but has a slight decrease compared to " + }, + { + "bbox": [ + 313, + 447, + 555, + 567 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{uniform}}" + }, + { + "bbox": [ + 313, + 447, + 555, + 567 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 313, + 447, + 555, + 567 + ], + "type": "inline_equation", + "content": "2.9\\%" + }, + { + "bbox": [ + 313, + 447, + 555, + 567 + ], + "type": "text", + "content": ". However, when " + }, + { + "bbox": [ + 313, + 447, + 555, + 567 + ], + "type": "inline_equation", + "content": "K = 8" + }, + { + "bbox": [ + 313, + 447, + 555, + 567 + ], + "type": "text", + "content": ", this approach far outperforms both of the individual criteria, improving " + }, + { + "bbox": [ + 313, + 447, + 555, + 567 + ], + "type": "inline_equation", + "content": "\\phi_{-R}" + }, + { + "bbox": [ + 313, + 447, + 555, + 567 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 313, + 447, + 555, + 567 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 313, + 447, + 555, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 447, + 555, + 567 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{uniform}}" + }, + { + "bbox": [ + 313, + 447, + 555, + 567 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 313, + 447, + 555, + 567 + ], + "type": "inline_equation", + "content": "17\\%" + }, + { + "bbox": [ + 313, + 447, + 555, + 567 + ], + "type": "text", + "content": ". Note that this setup results in slightly less computation efficiency than " + }, + { + "bbox": [ + 313, + 447, + 555, + 567 + ], + "type": "inline_equation", + "content": "\\phi_{-R}" + }, + { + "bbox": [ + 313, + 447, + 555, + 567 + ], + "type": "text", + "content": " (7% drop in FLOPS reduction for " + }, + { + "bbox": [ + 313, + 447, + 555, + 567 + ], + "type": "inline_equation", + "content": "K = 3" + }, + { + "bbox": [ + 313, + 447, + 555, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 447, + 555, + 567 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 313, + 447, + 555, + 567 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 447, + 555, + 567 + ], + "type": "inline_equation", + "content": "K = 16" + }, + { + "bbox": [ + 313, + 447, + 555, + 567 + ], + "type": "text", + "content": "), but this is outweighed by the performance improvements." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 567, + 554, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 554, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 554, + 604 + ], + "type": "text", + "content": "We provide qualitative examples of various criteria token selection and additional results using a different model setup in the supplement." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 611, + 419, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 419, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 419, + 624 + ], + "type": "text", + "content": "4.2. Distilling insights" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "type": "text", + "content": "Guided by our insights, we now present our final approach FEATHER (Fast and Effective Acceleration wiTH Ensemble cRicerca). In this approach, we first perform pruning after an early layer " + }, + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "type": "inline_equation", + "content": "(K = 8)" + }, + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "type": "text", + "content": ", utilizing our proposed criteria " + }, + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\phi_{-R} + \\phi_{\\mathrm{uniform}}" + }, + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "type": "text", + "content": " to retain " + }, + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "type": "inline_equation", + "content": "(1 - R)\\%" + }, + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "type": "text", + "content": " of tokens. Given our finding that the criteria improves as the LLM layer increases, we additionally prune a second time at " + }, + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "type": "inline_equation", + "content": "K = 16" + }, + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22832" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 71, + 556, + 205 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 556, + 205 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 556, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 556, + 205 + ], + "type": "image", + "image_path": "4f4a5029d3b97b52c56898e84e33a5b61a00f898caeb79e93e6a3149b6ddd97a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 213, + 555, + 237 + ], + "lines": [ + { + "bbox": [ + 54, + 213, + 555, + 237 + ], + "spans": [ + { + "bbox": [ + 54, + 213, + 555, + 237 + ], + "type": "text", + "content": "Figure 5. Comparing FEATHER performance against FastV [10] and PyramidDrop [39]. We find that FEATHER far outperforms both compared methods, particularly for the vision-centric task of localization." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 258, + 294, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 258, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 54, + 258, + 294, + 342 + ], + "type": "text", + "content": "At this stage, we utilize " + }, + { + "bbox": [ + 54, + 258, + 294, + 342 + ], + "type": "inline_equation", + "content": "\\phi_{-R}" + }, + { + "bbox": [ + 54, + 258, + 294, + 342 + ], + "type": "text", + "content": " as the criteria, as uniform sampling should not be necessary when the attention-based criteria reliably selects important tokens. For the reduction ratio, we choose to only retain " + }, + { + "bbox": [ + 54, + 258, + 294, + 342 + ], + "type": "inline_equation", + "content": "(1 - R)^{2\\%}" + }, + { + "bbox": [ + 54, + 258, + 294, + 342 + ], + "type": "text", + "content": " of the remaining tokens since the attention-based criteria has proved highly effective when pruning at later layers, even when using the original criteria " + }, + { + "bbox": [ + 54, + 258, + 294, + 342 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{original}}" + }, + { + "bbox": [ + 54, + 258, + 294, + 342 + ], + "type": "text", + "content": " (see Figure 3(c))." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 354, + 291, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 291, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 291, + 369 + ], + "type": "text", + "content": "4.3. Comparison against FastV and PyramidDrop" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 374, + 295, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 374, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 54, + 374, + 295, + 422 + ], + "type": "text", + "content": "We compare our approach against the one-stage early pruning approach of FastV (" + }, + { + "bbox": [ + 54, + 374, + 295, + 422 + ], + "type": "inline_equation", + "content": "K = 3" + }, + { + "bbox": [ + 54, + 374, + 295, + 422 + ], + "type": "text", + "content": ") [10] and the multi-stage approach of PyramidDrop (" + }, + { + "bbox": [ + 54, + 374, + 295, + 422 + ], + "type": "inline_equation", + "content": "K = [8, 16, 24]" + }, + { + "bbox": [ + 54, + 374, + 295, + 422 + ], + "type": "text", + "content": ") [39]. Note that both of these methods use " + }, + { + "bbox": [ + 54, + 374, + 295, + 422 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{original}}" + }, + { + "bbox": [ + 54, + 374, + 295, + 422 + ], + "type": "text", + "content": " for the criteria." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 424, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 424, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 54, + 424, + 295, + 604 + ], + "type": "text", + "content": "As shown in Figure 5, we find that FEATHER far outperforms the baselines, especially on the localization tasks. Specifically, for comparable computational costs (64% FLOPS reduction for our approach, 68% FLOPS reduction for FastV, and 65% FLOPS reduction for PyramidDrop), we observe that for localization tasks, FEATHER exhibits more than " + }, + { + "bbox": [ + 54, + 424, + 295, + 604 + ], + "type": "inline_equation", + "content": "5 \\times" + }, + { + "bbox": [ + 54, + 424, + 295, + 604 + ], + "type": "text", + "content": " average performance improvement compared to FastV and a " + }, + { + "bbox": [ + 54, + 424, + 295, + 604 + ], + "type": "inline_equation", + "content": "36\\%" + }, + { + "bbox": [ + 54, + 424, + 295, + 604 + ], + "type": "text", + "content": " average performance improvement compared to PyramidDrop. For non-localization tasks, FEATHER has a " + }, + { + "bbox": [ + 54, + 424, + 295, + 604 + ], + "type": "inline_equation", + "content": "7.8\\%" + }, + { + "bbox": [ + 54, + 424, + 295, + 604 + ], + "type": "text", + "content": " improvement over FastV and " + }, + { + "bbox": [ + 54, + 424, + 295, + 604 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 54, + 424, + 295, + 604 + ], + "type": "text", + "content": " improvement over PyramidDrop. Note that PyramidDrop performs substantially better than FastV as it prunes fewer tokens in an early layer. However, it still suffers from an ineffective pruning strategy at this stage, though the impact is less pronounced since it predominantly prunes later." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "type": "text", + "content": "Remarkably, with our " + }, + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "type": "inline_equation", + "content": "64\\%" + }, + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "type": "text", + "content": " FLOPS reduction setup, after layer 16 only " + }, + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "type": "inline_equation", + "content": "3.3\\%" + }, + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "type": "text", + "content": " of tokens are retained, yet the average localization performance decrease compared to the baseline method with no token pruning is only " + }, + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "type": "inline_equation", + "content": "26\\%" + }, + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "type": "text", + "content": " (Figure 1 includes an example of retained tokens after layer 16). This finding illustrates that even for vision-centric tasks, maintaining strong performance while gaining huge acceleration speedups with extensive pruning is possible, but it heavily relies on the effectiveness of the pruning criteria." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 256, + 388, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 256, + 388, + 269 + ], + "spans": [ + { + "bbox": [ + 313, + 256, + 388, + 269 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 276, + 555, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 276, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 555, + 491 + ], + "type": "text", + "content": "In this work, we examine the visual capabilities of the VLM acceleration approach of pruning visual tokens after shallow LLM layers. While strong performance is maintained on most evaluated tasks, it fails on more vision-centric tasks like TextVQA and localization due to its flawed pruning criteria that predominately selects visual tokens from the bottom part of the image. Observing this same flawed criteria on other tasks, we show that strong performance is largely due to the benchmarks' inability to assess fine-grained visual capabilities. Next, we propose and evaluate several alternative criteria to improve visual capabilities, ultimately arriving at our final method, FEATHER. This approach refines the attention-based criteria to address the token selection bias while using uniform sampling for better image coverage. It then prunes more aggressively when the criteria more effectively identifies important tokens. We find that FEATHER has more than " + }, + { + "bbox": [ + 313, + 276, + 555, + 491 + ], + "type": "inline_equation", + "content": "5 \\times" + }, + { + "bbox": [ + 313, + 276, + 555, + 491 + ], + "type": "text", + "content": " performance improvement on localization compared to the original acceleration approach." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 491, + 556, + 707 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 556, + 707 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 556, + 707 + ], + "type": "text", + "content": "While we study a particular type of VLM acceleration, our work highlights a broad challenge in evaluating VLMs. Building on [34], which shows that text-only models can perform well on some vision-language benchmarks, we demonstrate that even benchmarks with substantial differences between vision-enabled and disabled setups may not assess fine-grained visual capabilities. We address this issue by focusing on more vision-centric localization tasks, though this limits the analysis to a few specific types of skills. To accurately assess a wider range of visual capabilities, future work may explore how to resolve current dataset biases that models can exploit [21]. Additionally, while removing RoPE for token pruning effectively eliminates positional bias, it may introduce unintended effects on attention weights influencing token selection. Future research could investigate more robust methods for encoding positional information in visually conditioned language models to prevent positional artifacts in cross-modal interactions." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22833" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 172 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 172 + ], + "type": "text", + "content": "Acknowledgments. This work is supported in part by the National Science Foundation (NSF) under Grant No. 2026498, the Stanford AIMI-HAI Partnership Grant, and the NSF Graduate Research Fellowship Program under Grant No. DGE-2146755 (for M.E.). Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of any other entity." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 185, + 115, + 197 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 115, + 197 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 115, + 197 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 205, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 205, + 294, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 205, + 294, + 248 + ], + "spans": [ + { + "bbox": [ + 61, + 205, + 294, + 248 + ], + "type": "text", + "content": "[1] Manoj Acharya, Kushal Kafle, and Christopher Kanan. Tallyqa: Answering complex counting questions. In Proceedings of the AAAI conference on artificial intelligence, pages 8076-8084, 2019. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 251, + 294, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 251, + 294, + 315 + ], + "spans": [ + { + "bbox": [ + 61, + 251, + 294, + 315 + ], + "type": "text", + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 318, + 294, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 318, + 294, + 382 + ], + "spans": [ + { + "bbox": [ + 62, + 318, + 294, + 382 + ], + "type": "text", + "content": "[3] Kazi Hasan Ibn Arif, JinYi Yoon, Dimitrios S Nikolopoulos, Hans Vandierendonck, Deepu John, and Bo Ji. Hired: Attention-guided token dropping for efficient inference of high-resolution vision-language models in resource-constrained environments. arXiv preprint arXiv:2408.10945, 2024. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 385, + 294, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 385, + 294, + 429 + ], + "spans": [ + { + "bbox": [ + 62, + 385, + 294, + 429 + ], + "type": "text", + "content": "[4] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 431, + 294, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 431, + 294, + 496 + ], + "spans": [ + { + "bbox": [ + 62, + 431, + 294, + 496 + ], + "type": "text", + "content": "[5] Jeffrey P Bigham, Chandrika Jayant, Hanjie Ji, Greg Little, Andrew Miller, Robert C Miller, Robin Miller, Aubrey Tatarowicz, Brandyn White, Samual White, et al. Vizwiz: nearly real-time answers to visual questions. In Proceedings of the 23nd annual ACM symposium on User interface software and technology, pages 333-342, 2010. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 498, + 294, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 498, + 294, + 542 + ], + "spans": [ + { + "bbox": [ + 62, + 498, + 294, + 542 + ], + "type": "text", + "content": "[6] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. In The Eleventh International Conference on Learning Representations, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 544, + 294, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 544, + 294, + 576 + ], + "spans": [ + { + "bbox": [ + 62, + 544, + 294, + 576 + ], + "type": "text", + "content": "[7] Mu Cai, Jianwei Yang, Jianfeng Gao, and Yong Jae Lee. Matryoshka multimodal models. arXiv preprint arXiv:2405.17430, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 578, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 578, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 62, + 578, + 294, + 633 + ], + "type": "text", + "content": "[8] Wenhao Chai, Enxin Song, Yilun Du, Chenlin Meng, Vashisht Madhavan, Omer Bar-Tal, Jeng-Neng Hwang, Saining Xie, and Christopher D Manning. Auroracap: Efficient, performant video detailed captioning and a new benchmark. arXiv preprint arXiv:2410.03051, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 635, + 294, + 689 + ], + "type": "text", + "content": "[9] Jieneng Chen, Luoxin Ye, Ju He, Zhao-Yang Wang, Daniel Khashabi, and Alan Yuille. Efficient large multi-modal models via visual context compression. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "text", + "content": "[10] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. ECCV, 2024. 2, 3, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "text", + "content": "[11] Yi Chen, Jian Xu, Xu-Yao Zhang, Wen-Zhuo Liu, YangYang Liu, and Cheng-Lin Liu. Recoverable compression: A multimodal vision token recovery mechanism guided by text information. arXiv preprint arXiv:2409.01179, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 141, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 553, + 183 + ], + "type": "text", + "content": "[12] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in neural information processing systems, 35:16344-16359, 2022. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 185, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 185, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 185, + 553, + 239 + ], + "type": "text", + "content": "[13] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 240, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 553, + 293 + ], + "type": "text", + "content": "[14] Yefei He, Feng Chen, Jing Liu, Wenqi Shao, Hong Zhou, Kaipeng Zhang, and Bohan Zhuang. Zipvl: Efficient large vision-language models with dynamic token sparsification and kv cache compression. arXiv preprint arXiv:2410.08584, 2024.3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 295, + 553, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 295, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 295, + 553, + 338 + ], + "type": "text", + "content": "[15] Xiangyu Hong, Che Jiang, Biqing Qi, Fandong Meng, Mo Yu, Bowen Zhou, and Jie Zhou. On the token distance modeling ability of higher rope attention dimension. arXiv preprint arXiv:2410.08703, 2024. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 339, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 339, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 317, + 339, + 553, + 392 + ], + "type": "text", + "content": "[16] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 394, + 553, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 394, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 317, + 394, + 553, + 458 + ], + "type": "text", + "content": "[17] Peng Jin, Ryuichi Takanobu, Wancai Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13700-13710, 2024. 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 460, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 553, + 525 + ], + "type": "text", + "content": "[18] Siddharth Karamcheti, Suraj Nair, Ashwin Balakrishna, Percy Liang, Thomas Kollar, and Dorsa Sadigh. Prismatic VLMs: Investigating the design space of visually-conditioned language models. In Proceedings of the 41st International Conference on Machine Learning, pages 23123-23144. PMLR, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 527, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 527, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 317, + 527, + 553, + 581 + ], + "type": "text", + "content": "[19] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 582, + 553, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 582, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 317, + 582, + 553, + 647 + ], + "type": "text", + "content": "[20] Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Minjoon Seo, Hannaneh Hajishirzi, and Ali Farhadi. A diagram is worth a dozen images. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 235-251. Springer, 2016. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 647, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 712 + ], + "type": "text", + "content": "[21] Baiqi Li, Zhiqiu Lin, Wenxuan Peng, Jean de Dieu Nyandwi, Daniel Jiang, Zixian Ma, Simran Khanuja, Ranjay Krishna, Graham Neubig, and Deva Ramanan. Naturalbench: Evaluating vision-language models on natural adversarial samples. Advances in Neural Information Processing Systems, 37:17044-17068, 2025. 8" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22834" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[22] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "type": "text", + "content": "[23] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "type": "text", + "content": "[24] Kevin Y Li, Sachin Goyal, Joao D Semedo, and J Zico Kolter. Inference optimal vlms need only one visual token but larger models. arXiv preprint arXiv:2411.03312, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "type": "text", + "content": "[25] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 275, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 294, + 308 + ], + "type": "text", + "content": "[26] Fangyu Liu, Guy Emerson, and Nigel Collier. Visual spatial reasoning. Transactions of the Association for Computational Linguistics, 11:635-651, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 309, + 294, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 353 + ], + "type": "text", + "content": "[27] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 294, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 388 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 388 + ], + "type": "text", + "content": "[28] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 294, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 422 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 422 + ], + "type": "text", + "content": "[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 422, + 294, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 477 + ], + "type": "text", + "content": "[30] Yifei Liu, Mathias Gehrig, Nico Messikommer, Marco Cannici, and Davide Scaramuzza. Revisiting token pruning for object detection and instance segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2658-2668, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 478, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 521 + ], + "type": "text", + "content": "[31] Yuzhang Shang, Mu Cai, Bingxin Xu, Yong Jae Lee, and Yan Yan. Llava-prumerge: Adaptive token reduction for efficient large multimodal models. arXiv preprint arXiv:2403.15388, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "type": "text", + "content": "[32] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 579, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 622 + ], + "type": "text", + "content": "[33] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 679 + ], + "type": "text", + "content": "[34] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, et al. Cambrian1: A fully open, vision-centric exploration of multimodal llms. arXiv preprint arXiv:2406.16860, 2024. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[35] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 631 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "text", + "content": "[36] Zhongwei Wan, Ziang Wu, Che Liu, Jina Huang, Zhihong Zhu, Peng Jin, Longyue Wang, and Li Yuan. Look-m: Look-once optimization in kv cache for efficient multimodal long-context inference. arXiv preprint arXiv:2406.18139, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 239 + ], + "type": "text", + "content": "[37] Ke-Jyun Wang, Yun-Hsuan Liu, Hung-Ting Su, Jen-Wei Wang, Yu-Siang Wang, Winston Hsu, and Wen-Chin Chen. OCID-ref: A 3D robotic dataset with embodied language for clutter scene grounding. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5333–5338, Online, 2021. Association for Computational Linguistics. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 240, + 553, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 553, + 295 + ], + "type": "text", + "content": "[38] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 296, + 553, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 296, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 316, + 296, + 553, + 361 + ], + "type": "text", + "content": "[39] Long Xing, Qidong Huang, Xiaoyi Dong, Jiajie Lu, Pan Zhang, Yuhang Zang, Yuhang Cao, Conghui He, Jiaqi Wang, Feng Wu, et al. Pyramiddrop: Accelerating your large vision-language models via pyramid visual redundancy reduction. arXiv preprint arXiv:2410.17247, 2024. 2, 3, 4, 5, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 363, + 553, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 363, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 316, + 363, + 553, + 418 + ], + "type": "text", + "content": "[40] Senqiao Yang, Yukang Chen, Zhuotao Tian, Chengyao Wang, Jingyao Li, Bei Yu, and Jiaya Jia. Visionzip: Longer is better but not necessary in vision language models. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 19792-19802, 2025. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 419, + 553, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 419, + 553, + 462 + ], + "spans": [ + { + "bbox": [ + 316, + 419, + 553, + 462 + ], + "type": "text", + "content": "[41] Gaotong Yu, Yi Chen, and Jian Xu. Balancing performance and efficiency: A multimodal large language model pruning method based image text interaction. arXiv preprint arXiv:2409.01162, 2024.3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 463, + 553, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 463, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 316, + 463, + 553, + 529 + ], + "type": "text", + "content": "[42] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 530, + 553, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 530, + 553, + 574 + ], + "spans": [ + { + "bbox": [ + 316, + 530, + 553, + 574 + ], + "type": "text", + "content": "[43] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 575, + 553, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 575, + 553, + 631 + ], + "spans": [ + { + "bbox": [ + 316, + 575, + 553, + 631 + ], + "type": "text", + "content": "[44] Qizhe Zhang, Aosong Cheng, Ming Lu, Zhiyong Zhuo, Minqi Wang, Jiajun Cao, Shaobo Guo, Qi She, and Shang-hang Zhang. [cls] attention is all you need for training-free visual token pruning: Make vlm inference faster. arXiv eprints, pages arXiv-2412, 2024. 2" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22835" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Feature Coding in the Era of Large Models_ Dataset, Test Conditions, and Benchmark/83895aed-d59c-465f-afa3-c8ed37ae4a4b_content_list.json b/2025/Feature Coding in the Era of Large Models_ Dataset, Test Conditions, and Benchmark/83895aed-d59c-465f-afa3-c8ed37ae4a4b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..66044f2621e2db7055da5a1353c034da6eec0d6b --- /dev/null +++ b/2025/Feature Coding in the Era of Large Models_ Dataset, Test Conditions, and Benchmark/83895aed-d59c-465f-afa3-c8ed37ae4a4b_content_list.json @@ -0,0 +1,1778 @@ +[ + { + "type": "text", + "text": "Feature Coding in the Era of Large Models: Dataset, Test Conditions, and Benchmark", + "text_level": 1, + "bbox": [ + 276, + 130, + 722, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Changsheng Gao $^{1}$ , Yifan Ma $^{2}$ , Qiaoxi Chen $^{2}$ , Yenan Xu $^{2}$ , Dong Liu $^{2}$ , Weisi Lin $^{1\\dagger}$ $^{1}$ Media Interactive Computing Lab, Nanyang Technological University, Singapore 639798 \n $^{2}$ MOE Key Laboratory of Brain-Inspired Intelligent Perception and Cognition, University of Science and Technology of China, Hefei 230093, China \n{changsheng.gao, wslin} @ntu.edu.sg; {mayf,xxii,yenanxu} @mail.ustc.edu.cn; dongeliu@ustc.edu.cn", + "bbox": [ + 101, + 202, + 908, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large models have achieved remarkable performance across various tasks, yet they incur significant computational costs and privacy concerns during both training and inference. Distributed deployment has emerged as a potential solution, but it necessitates the exchange of intermediate information between model segments, with feature representations serving as crucial information carriers. To optimize information exchange, feature coding is required to reduce transmission and storage overhead. Despite its importance, feature coding for large models remains an under-explored area. In this paper, we draw attention to large model feature coding and make three fundamental contributions. First, we introduce a comprehensive dataset encompassing diverse features generated by three representative types of large models. Second, we establish unified test conditions, enabling standardized evaluation pipelines and fair comparisons across future feature coding studies. Third, we introduce two baseline methods derived from widely used image coding techniques and benchmark their performance on the proposed dataset. These contributions aim to provide a foundation for future research and inspire broader engagement in this field. To support a long-term study, all source code and the dataset are made available at https://github.com/chansongoal/LaMoFC.", + "bbox": [ + 89, + 358, + 485, + 720 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 752, + 220, + 768 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large models have revolutionized artificial intelligence (AI) with their impressive performance. However, they require substantial training data and extensive computational resources, posing significant challenges for practical de", + "bbox": [ + 89, + 777, + 483, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ployments [19]. To address these challenges, distributed deployment has been proposed as a promising solution [14, 23, 38, 43, 49, 52]. By distributing the computational workload across multiple platforms, this approach reduces the resource strain on individual systems and enables the integration of private client data into training while maintaining data security [4, 49]. As model scale continues to grow, distributed deployment is expected to become a mainstream strategy for large model deployments.", + "bbox": [ + 511, + 327, + 906, + 465 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In distributed deployments, effective information exchange between model segments is essential. During forward propagation, features generated by early segments are passed to later segments to complete the inference. Given the vast amount of data in large model applications, it is necessary to encode features to a smaller size to reduce storage and transmission burden. Analogous to image and video coding, feature coding has been proposed to address this problem [6, 16, 18, 44].", + "bbox": [ + 511, + 465, + 908, + 603 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing feature coding research faces three key limitations: narrow model types, limited task diversity, and constrained data modalities. Most existing feature coding efforts focus on small-scale convolution-based networks, such as ResNet, ignoring large models. Furthermore, existing research primarily targets visual features and neglects the investigation of textual features from large language models (LLMs). This gap indicates a need for broader exploration across both models and data modalities to advance feature coding.", + "bbox": [ + 511, + 606, + 908, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we shift the feature coding research from small-scale to large-scale models. To address the existing limitations, we construct a comprehensive test dataset with features from five tasks across three large models, covering both visual and textual data for increased modality diversity. Furthermore, we establish unified test conditions to enable fair comparisons, including formulated bitrate computation and task-specific evaluation pipelines.", + "bbox": [ + 511, + 761, + 908, + 882 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To lay the groundwork for large model feature coding,", + "bbox": [ + 532, + 885, + 903, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "This work was supported by the Ministry of Education of Singapore under Grant T2EP20123-0006. We acknowledge the support of GPU cluster built by MCC Lab of Information Science and Technology Institution, USTC.", + "bbox": [ + 89, + 851, + 483, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1068", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5cfbbc208d6e297dafd9fcdd8a637b69d8c1059ff25571f063af8341e5fa09d8.jpg", + "image_caption": [ + "Figure 1. Three application scenarios for feature coding in large model deployments. Cloud-centralized training: Features generated by foundation models are encoded and stored in a feature database. This database supports various downstream tasks, reducing computational demands by avoiding repeated inference of large foundation models. Cloud-edge distributed training: To protect user data privacy, raw data is first processed by an edge model to generate features, which are then encoded and transmitted to the cloud. The cloud decodes these features to complete the remaining forward propagation. Cloud-edge distributed inference: A large model is partitioned across multiple parts, with portions offloaded to edge devices to distribute the computational load. The transmitted features ensure privacy protection and enable customized downstream tasks. Please refer to Sec. 3 for detailed illustrations." + ], + "image_footnote": [], + "bbox": [ + 102, + 89, + 380, + 265 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ad6fad99f7d852a9319900b8a43a2a291e26eea226a6f9e5fbde9b202841cfc9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 88, + 607, + 267 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5ad02696a7e2311c91c90042b81a11cd74167bdf55ef20d157a5db745d534702.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 88, + 893, + 266 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "we propose two baseline methods and evaluate their applicability to the dataset. The baselines leverage two image coding methods: the handcrafted VVC Intra coding [2] and the learning-based Hyperprior method [1]. We introduce pre-processing and post-processing modules to adapt the baselines to feature characteristics. The resulting benchmark and analyses provide valuable insights into large model feature coding and identify promising directions for further exploration. In summary, our main contributions are as follows:", + "bbox": [ + 89, + 402, + 483, + 551 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We highlight the importance of feature coding in large model deployments, identify relevant application scenarios, and introduce a new research area: large model feature coding.", + "- We construct a comprehensive feature dataset encompassing three model types, five tasks, and two kinds of source data modalities, along with unified test conditions to support long-term large model feature coding research.", + "- We propose two baselines and evaluate their usability to the large model feature coding. We build a benchmark and suggest promising directions for further research." + ], + "bbox": [ + 89, + 553, + 482, + 719 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Background and motivation", + "text_level": 1, + "bbox": [ + 89, + 733, + 349, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Feature coding", + "text_level": 1, + "bbox": [ + 89, + 757, + 241, + 773 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Originally introduced within the Coding for Machines framework [10, 48], feature coding is one of two primary branches alongside visual coding. Unlike visual coding [15, 25, 26, 28, 31, 35, 39, 40, 50], which encodes and reconstructs the original visual data, feature coding involves encoding features extracted from source data into bitstreams. While visual coding has already been explored in the context of large models [20, 41], feature coding re", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "mains unexplored in this domain.", + "bbox": [ + 513, + 402, + 735, + 417 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite recent progress, feature coding research still faces three key limitations. First, there is a restricted variety of model types in use. The existing feature coding focuses on small-scale CNN features such as [3, 17, 21, 24, 27, 30, 37]. However, as Transformer architectures now dominate large model research [11, 12, 33, 42], it is crucial to investigate feature coding for Transformer-based models. Second, the field is constrained by the limited scope of task types. Existing studies primarily address discriminative tasks such as image classification and segmentation [5, 7, 13, 32, 47, 51], while generative tasks remain largely unexplored. Given the significance of generative models, extending feature coding to include these tasks, such as image synthesis, is essential. The third limitation is the constrained range of source modality. Current feature coding research predominantly focuses on features extracted from visual data, overlooking the increasingly important features generated from textual data, particularly in light of advancements in LLMs. Expanding to multiple modalities is crucial as the diversity of AI applications grows.", + "bbox": [ + 511, + 428, + 906, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, two challenges hinder progress in large model feature coding: the lack of a public test dataset and unified test conditions. Without a standardized dataset and evaluation conditions, fair comparisons across methods can not be conducted. A public dataset ensures consistent input data across studies, while unified test conditions establish a standard pipeline to evaluate performance.", + "bbox": [ + 511, + 739, + 905, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These limitations motivate us to construct a diverse test dataset and formulate unified test conditions to support large model feature coding research.", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1069", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/c73c35a121b76d32a87f31607e67d82154137c176732733ade1bdd3d58294ccf.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelParametersTaskSplit PointFeature ShapeSource Data
DINOv21.1BImage Classification (Cls)40th ViT Block257 × 1536ImageNet, 500 samples
Semantic Segmentation (Seg)40th ViT Block2 × 1370 × 1536VOC2012, 100 samples
Depth Estimation (Dpt)10th, 20th, 30th, 40th ViT Blocks2 × 4 × 1611 × 1536NYU-Depth-v2, 80 samples
Llama38BCommon Sense Reasoning (CSR)32th Decoder LayerN × 4096Arc-Challenge, 500 samples
SD38BText-to-Image Synthesis (TTI)Input Layer of VAE Decoder16 × 128 × 128COCO2017, 500 samples
", + "bbox": [ + 94, + 88, + 906, + 186 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Summary of the model and task selection, split point decision, and source data collection.", + "bbox": [ + 205, + 196, + 787, + 209 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Large model deployment", + "text_level": 1, + "bbox": [ + 89, + 236, + 318, + 253 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The impressive performance of large models is driven by scaling laws [34, 36], which emphasize the importance of vast model sizes and extensive training data. For example, GPT-3 was built with 175 billion parameters and trained on 499 trillion tokens, both contributing to substantial computational demands. To alleviate these resource requirements, distributed training and inference methods have been proposed in [14, 49, 52].", + "bbox": [ + 89, + 258, + 482, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "User privacy protection presents another challenge in client services [4, 29, 46]. Split learning has been proposed as an effective solution to protect data privacy by partitioning models between clients and the cloud [49].", + "bbox": [ + 89, + 380, + 482, + 439 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In distributed deployments, efficient information exchange between model segments is crucial. However, academic research has largely overlooked the storage and transmission costs associated with feature exchange. To address this gap, we propose to encode features into compact bitstreams to improve the efficiency of large model deployments in distributed environments.", + "bbox": [ + 89, + 440, + 482, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Feature coding in large models", + "text_level": 1, + "bbox": [ + 89, + 559, + 372, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In Fig. 1, we illustrate three application scenarios of feature coding in large model deployments, highlighting their roles in optimizing the deployment efficiency of large models.", + "bbox": [ + 89, + 585, + 482, + 631 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Large model training", + "text_level": 1, + "bbox": [ + 89, + 640, + 290, + 656 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Large models achieve their high performance through extensive training on large-scale datasets. Cloud-centralized training is commonly employed to leverage the vast computational resources and data in the cloud. However, in sensitive domains like healthcare, high-quality public data may be scarce, necessitating access to private data through cloud-edge collaborative training.", + "bbox": [ + 89, + 662, + 482, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1. Cloud-centralized training", + "text_level": 1, + "bbox": [ + 89, + 776, + 321, + 791 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In cloud-centralized training, source databases are maintained in the cloud, where intensive computational demands become the main challenge. Complex tasks are increasingly addressed through the collaborative use of multiple models. For instance, in vision-language tasks, visual data is first processed by a vision model, and its features are then passed to an LLM for advanced reasoning. Similarly, foundation", + "bbox": [ + 89, + 795, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "models often serve as core backbones for various tasks. For example, features extracted from DINOv2 are repurposed for downstream tasks like segmentation [9]. In such collaborative workflows, features generated by one model become inputs for others. However, during training, each data sample may undergo repeated inference, resulting in substantial and unnecessary computational costs.", + "bbox": [ + 511, + 237, + 903, + 342 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address this, an effective solution is to generate features once, store them, and reuse them across tasks. The workflow is outlined on the left of Fig. 1. First, a foundation model extracts features from the source database, which are then encoded and stored in a feature database. For subsequent task-specific training, only the precomputed feature database is needed, enabling efficient access to features without redundant feature extraction. This approach significantly reduces computational load and accelerates downstream task training. However, raw feature data consumes considerable storage space, particularly with large datasets. Therefore, encoding raw features into compact bitstreams becomes essential to reduce storage demands and support efficient large model training.", + "bbox": [ + 511, + 343, + 906, + 554 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.2. Cloud-edge distributed training", + "text_level": 1, + "bbox": [ + 511, + 563, + 781, + 579 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In addition to computational resources, large model training requires access to large-scale datasets, which are often only available to a few major high-tech companies. Furthermore, sensitive data, such as medical records and financial information, cannot be freely shared due to privacy concerns and regulatory constraints. To leverage distributed, privately owned data, split learning has emerged as a viable solution [43]. Split learning divides the training process into two phases, with the bulk of the computational load handled by the cloud, while edge devices, which are often resource-constrained, handle a smaller portion of the task. This approach minimizes the computational demand on the edge while enabling collaboration with powerful cloud resources.", + "bbox": [ + 511, + 583, + 905, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The cloud-edge distributed training pipeline is depicted in the middle of Fig. 1. The whole pipeline is divided into two parts: edge models and cloud models. During forward propagation, the edge model processes private data to generate features, which are then encoded into bitstreams and transmitted to the cloud. Upon receiving the bitstreams, the cloud decodes them back into features and feeds them into", + "bbox": [ + 511, + 795, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1070", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/76661b8b03c3bc3213842678c159134140ddff8c59cf6966f062cd93e1ee1105.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 143, + 85, + 318, + 183 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5a00698ce78813fe56b4e6ea13e3c7e94f635c41218e3122552d908ea572d8e7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 321, + 85, + 496, + 184 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2dc1cd5df87737ee76131aef07637878118bea7855ccd3e49293906fb1d0ba34.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 498, + 85, + 673, + 184 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c9602c75c816fb295c97bb07fcaa193d1eac5aca76d3e2975f165abaf9b58ec1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 676, + 85, + 852, + 184 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/267ce443f187ef595e12361b25cda2283d70ebef1745a44be56020658672aa4a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 143, + 185, + 316, + 281 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c599d8542e9e22d57c8bbc3ea612cc9fdcf60e17615447cc8435f265a2ba0e8f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 321, + 185, + 495, + 281 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cd11309c31c337a747dc838b16c574bc3c2cda1f8e1a8688ef4f6998c59bd093.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 498, + 185, + 673, + 281 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/22879c822d994a8b091d6d2ebaf337098868a49b854aa8aa5cb90e64408d38a4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 676, + 185, + 852, + 281 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/996d38b45cadd165743fb1b5413305207b37cf597af18146e705bde40f01f188.jpg", + "image_caption": [ + "Figure 2. Frequency and CDF comparisons between features in the proposed dataset and commonly used existing features. The proposed features exhibit distinct distributions compared to existing features, highlighting the necessity and value of the dataset. In addition, the proposed features demonstrate greater diversity, enhancing their representativeness and suitability for long-term use. Frequencies are scaled logarithmically for better visualization." + ], + "image_footnote": [], + "bbox": [ + 143, + 282, + 316, + 378 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/88f5b447d96c7fb03ba0baab467268b32bae698984944d13bdfe7cf05a3882cd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 321, + 282, + 495, + 378 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9a3488ba0353a68041d78a8d85f4daf6e13bdafe4294e86d6efde24099ed6aea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 282, + 673, + 378 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/19f40d0d6443436d59b6e7511b5620ff29af7c212c90ec73ff796d5e7b3d8a61.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 676, + 282, + 852, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the cloud model. In the backward pass, gradients are computed and backpropagated along the reverse path. To reduce bandwidth cost and minimize latency, it is critical to apply feature coding before transmission.", + "bbox": [ + 89, + 473, + 482, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Cloud-edge distributed inference", + "text_level": 1, + "bbox": [ + 89, + 545, + 377, + 560 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Once large models are trained, they are deployed in products to provide client services. For a product or service to succeed, two primary concerns must be addressed: supporting a high volume of simultaneous cloud server access and protecting user privacy. For the first concern, managing computational load is essential, particularly in generative applications that involve repetitive diffusion processes. A practical solution is to offload part of this computational demand to edge devices. For the second concern, the same strategy in the cloud-edge distributed training, as discussed in Sec. 3.1.2, can be applied.", + "bbox": [ + 89, + 566, + 482, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The cloud-edge distributed inference pipeline is illustrated on the right of Fig. 1. In the upload phase, source data is converted into features by an edge model on the client side. These features are then encoded and transmitted to the cloud, where they are further processed by large models. In the download phase, the processed features are sent back to the client to complete a specific task. Depending on the task head used, various functionalities can be performed with the returned features. For example, Stable Diffusion 3 (SD3) features derived from an image can be used to synthesize video or generate segmentation masks. This approach not", + "bbox": [ + 89, + 734, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "only reduces the computational load on the cloud server but also enhances data privacy by keeping raw user data on the edge side.", + "bbox": [ + 511, + 473, + 906, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Dataset construction and analysis", + "text_level": 1, + "bbox": [ + 511, + 535, + 816, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Dataset construction", + "text_level": 1, + "bbox": [ + 511, + 561, + 707, + 574 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To guarantee the dataset's representativeness and support the long-term study of feature coding research, we curate the test dataset with consideration of three key aspects. Table 1 outlines our selected models and tasks, split points, and source data.", + "bbox": [ + 511, + 583, + 905, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1.1. Model and task selection", + "text_level": 1, + "bbox": [ + 511, + 669, + 732, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the vast number of large models available, it is impractical to include all of them in a single research paper. Therefore, we focus on selecting representative models. To ensure the dataset's representativeness and long-term relevance, we choose one widely recognized model for each type of large model. Specifically, we select DINOv2 [33], Llama3 [11], and SD3 [12] as representatives of discriminative models, generative models, and hybrid models, respectively. The trend of adopting Transformer architectures in large models is expected to persist in the coming years. As a result, the selected models are anticipated to retain their research value and practical utility in the near future. DINOv2 and Llama3 handle visual and textual inputs, respectively, while SD3 processes textual inputs to generate visual out-", + "bbox": [ + 511, + 688, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1071", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/86cbfb6ed34c740157333ae45c8978f02c3bd4a0ba9f80ea410c63f34caeba56.jpg", + "image_caption": [ + "Figure 3. Visualization of the original feature blocks and their corresponding DCT blocks (absolute values). In the feature domain, the proposed features exhibit higher vertical redundancy, whereas the existing features vary smoothly in both directions. In the DCT domain, the proposed features display more dispersed energy distributions. These distinct redundancy characteristics indicate different coding properties. We use $7 \\times 7$ blocks for $C5$ and $C5$ -ReLU and $16 \\times 16$ blocks for other features." + ], + "image_footnote": [], + "bbox": [ + 158, + 79, + 836, + 401 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "puts. Together, these models provide comprehensive feature representations across visual, textual, and cross-modal applications.", + "bbox": [ + 88, + 494, + 480, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For DINOv2, we include image classification (Cls), semantic segmentation (Seg), and depth estimation (Dpt) tasks. Llama3 is applied to the common sense reasoning task (CSR), and SD3 is used for text-to-image synthesis (TTI). The selection covers three visual tasks, one textual task, and one text-to-visual task, forming a comprehensive foundation for analyzing feature coding across diverse tasks and model types.", + "bbox": [ + 88, + 539, + 482, + 660 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.2. Split point decision", + "text_level": 1, + "bbox": [ + 89, + 669, + 272, + 684 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The split points are decided to support various downstream tasks. For DINOv2, we define two types of split points: 1) $SP_{DS}$ , the output of the $40^{th}$ ViT layer, and 2) $SP_{DM}$ , which aggregates outputs from the $10^{th}$ , $20^{th}$ , $30^{th}$ , and $40^{th}$ ViT layers. $SP_{DS}$ is used for tasks utilizing single-layer features, while $SP_{DM}$ supports tasks that require multi-layer features. In this study, $SP_{DS}$ is applied to Cls and Seg, and $SP_{DM}$ to Dpt. The corresponding features are designated as $F_{Cls}$ , $F_{Seg}$ , and $F_{Dpt}$ , respectively.", + "bbox": [ + 89, + 688, + 482, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For Llama3, the split point $SP_{G}$ is set at the output of the $32^{nd}$ decoder layer. Features extracted at $SP_{G}$ can support both task-specific heads and integration with other large models [45]. In CSR, $SP_{G}$ produces $N \\times 4096$ features, where $N$ is the number of tokens.", + "bbox": [ + 89, + 825, + 483, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For SD3, we take the input layer of the VAE decoder as the split point $SP_{H}$ . This split point generates features of size $16 \\times 128 \\times 128$ , denoted as $F_{TTI}$ , which are then passed to the VAE decoder for image synthesis.", + "bbox": [ + 511, + 494, + 906, + 555 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.3. Source data collection", + "text_level": 1, + "bbox": [ + 511, + 561, + 714, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We organize our feature test dataset into five classes, each supported by a curated subset of publicly available datasets used as source data. For Cls, we select 500 images from ImageNet [8], each representing a unique class accurately classified by DINOv2. For Seg, we collect 100 images from VOC2012, ensuring coverage of all 20 object classes. For Dpt, we source 80 images from NYU-Depth-v2, adhering to the train-test split proposed in [22] and selecting 3-8 images for each scene category. For CSR, we utilize 500 samples from the Arc-Challenge dataset, chosen for their longest input prompts and correct prediction by Llama3. For TTI, we first select 500 images from COCO2017 and then collect their longest captions.", + "bbox": [ + 511, + 580, + 906, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We noticed that learning-based methods require a dedicated training set. Please follow the source code to generate training data and find our training data in the supplementary.", + "bbox": [ + 511, + 777, + 906, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Feature analysis", + "text_level": 1, + "bbox": [ + 511, + 832, + 674, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we compare the proposed features with those commonly used in existing research to demonstrate the necessity and importance of introducing a new dataset.", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1072", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/3e0941e7833f46f4b008ce20586747f1e4b6566992414853eb3b98d8545d947d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TaskHeadMetric
ClsNorm + Linear1Accuracy
SegNorm + Linear1mIoU
DptLinear4RMSE
CSRNorm + Linear + SoftmaxAccuracy
TTIVAE DecoderFID
", + "bbox": [ + 161, + 88, + 416, + 188 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Summary of task head and accuracy metric settings. Find definitions of Heads in the original papers [11, 12, 33].", + "bbox": [ + 89, + 198, + 482, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.1. Distribution analysis", + "text_level": 1, + "bbox": [ + 89, + 252, + 282, + 266 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first compare frequency distributions and cumulative distribution functions (CDFs) of the proposed and commonly used existing features in Fig. 2. $C5$ and $C5$ -ReLU are derived from the official ResNet50 pretrained on ImageNet, while $P$ -layer features are generated by the ResNeXT101-based MaskRCNN pre-trained on COCO2017.", + "bbox": [ + 89, + 271, + 482, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For multi-layer split points, we observe four key distinctions between $SP_{DM}$ and $P$ -layer features: (1) $SP_{DM}$ distribution range expands progressively with deeper layers; (2) $SP_{DM}$ distribution is more asymmetric; (3) $SP_{DM}$ feature values are more concentrated, as evidenced by CDFs; and (4) $SP_{DM}$ features contain more regions with sparse feature points. For single-layer split points, $SP_{GS}$ has a more concentrated distribution, while $SP_{H}$ shows more peaks. The proposed dataset omits C5-ReLU-like features, as ReLU is rarely used in large models.", + "bbox": [ + 89, + 377, + 482, + 526 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The proposed features exhibit different distributions compared to the existing features, highlighting the necessity of introducing the new dataset. Additionally, the high diversity of the proposed features enhances the dataset's representativeness and suitability for long-term research.", + "bbox": [ + 89, + 527, + 482, + 603 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.2. Redundancy analysis", + "text_level": 1, + "bbox": [ + 89, + 609, + 284, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since coding aims to remove redundancy in inputs, we conduct a redundancy comparison on feature blocks. We visualize the original feature blocks and their DCT blocks (absolute values) in Fig. 3. In the feature domain, clear spatial redundancy is observed in both $SP_{DM}$ and $SP_{H}$ features. $SP_{H}$ exhibits redundancy in both horizontal and vertical directions, while $SP_{DM}$ shows stronger vertical redundancy. In contrast, $SP_{GS}$ features vary rapidly in both directions, resulting in weak spatial redundancy. This is because Llama3 only processes textual data, which lacks inherent spatial correlation. Existing features, on the other hand, exhibit smoother spatial variation and higher spatial redundancy. We attribute this to the translation invariance of convolutional layers, which preserve the relative positions of pixels in the feature domain, resulting in consistent redundancy patterns. In contrast, transformer-based vision models partition images into patches before transforming them, shifting spatial redundancy primarily to the vertical", + "bbox": [ + 89, + 628, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b7a862637e6c6c4a12fc13308b6262bdc21f27ef8e7c4f271d0445a4704e5cea.jpg", + "image_caption": [ + "Figure 4. The pipeline of the proposed baseline methods." + ], + "image_footnote": [], + "bbox": [ + 517, + 87, + 903, + 157 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "direction, as seen in $SP_{DM}$ .", + "bbox": [ + 511, + 205, + 702, + 220 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the DCT domain, spatial redundancy is reflected in the energy distribution. The proposed features display more dispersed energy, whereas existing features concentrate energy in the top-left DC component. $SP_{H}$ has the highest energy concentration as it closely resembles images.", + "bbox": [ + 511, + 220, + 903, + 296 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "These distinct redundancy characteristics further demonstrate the necessity of the proposed dataset and suggest that future studies should explore feature coding methods tailored to these unique properties.", + "bbox": [ + 511, + 297, + 903, + 357 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Unified test conditions", + "text_level": 1, + "bbox": [ + 511, + 369, + 723, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Even with a test dataset, fair comparisons are still challenging without identical test conditions. To address this, we establish unified test conditions here.", + "bbox": [ + 511, + 393, + 903, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Bitrate computation", + "text_level": 1, + "bbox": [ + 511, + 448, + 705, + 464 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We introduce a new bitrate measurement, Bits Per Feature Point (BPPF), replacing the commonly used Bits Per Pixel (BPP). BPPF is calculated by dividing the total coding bits by the number of feature points. Our rationale for shifting from BPP to BPPF is twofold. First, BPP is unsuitable for features extracted from non-visual data where pixels do not exist. Second, we believe bitrate should reflect the actual encoded data (the feature itself) rather than the source data. Using BPP to calculate bitrate on source data introduces ambiguity and limits fair comparisons between coding methods. For instance, applying the same BPP to different features extracted from the same image can be misleading. In addition, BPP may incentivize downscaling source images to reduce feature size, yielding lower bitrates that do not represent true coding efficiency. In contrast, BPPF measures bitrate on the feature itself, enabling fair comparisons across coding methods as long as the same feature is used.", + "bbox": [ + 511, + 470, + 906, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Task accuracy evaluation", + "text_level": 1, + "bbox": [ + 511, + 734, + 743, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The decoded features undergo task heads to perform accuracy evaluation. To isolate the impact of task heads, we choose simple task heads and fix their parameters during evaluation. The task heads and accuracy metrics are presented in Table 2.", + "bbox": [ + 511, + 757, + 903, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Baselines and benchmark", + "text_level": 1, + "bbox": [ + 511, + 844, + 751, + 859 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As a starting point, we introduce two baselines and establish a benchmark. We aim to examine their applicability to", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1073", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/be80f77614e36e8da69a0f16d4691b14e9cbeb898a7255556dd970635c07e713.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TaskImage ClassificationSemantic SegmentationDepth EstimationCommon Sense ReasoningText-to-Image Synthesis
MetricBPFPAccuracyMSEBPFPmIoUMSEBPFPRMSEMSEBPFPAccuracyMSEBPFPFIDMSE
Original3210003283.390320.369503210003200
VTMBaseline1.9499.802.94991.7680.421.86682.210.50210.61082.6999.800.01091.416.140.0066
1.0397.403.16930.8879.312.08361.270.65110.68261.831000.02600.7419.940.0184
0.2374.603.69930.2373.472.54400.330.95300.92250.8998.800.08100.2958.420.0421
0.0425.804.06510.0456.532.95260.031.48501.06310.1682.200.18680.11109.230.0715
0.017.204.44000.0137.423.27620.0032.11741.10720.0423.800.24550.05171.630.1072
HyperpriorBaseline2.0193.203.45011.7177.962.20101.510.42560.66006.3491.400.07761.4425.250.0138
1.1388.803.69661.3077.272.27951.010.49650.68443.6087.800.08080.6652.120.0300
0.9183.603.83270.5474.822.58450.430.66990.78371.6882.400.16220.2895.030.0541
0.3729.004.29090.1262.583.01510.081.03511.08441.5057.800.13550.15124.840.0734
0.2314.604.78140.0337.113.54490.012.73021.18661.3534.800.16240.08170.010.1006
", + "bbox": [ + 114, + 88, + 885, + 256 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2935ad05a460ea7b0bbd9b2cec70c95e794f4c86eb390c7d9dd462eb9ee036c5.jpg", + "table_caption": [ + "Table 3. Rate-accuracy (R-A) performance evaluations on the two baseline methods (MSE is calculated between $F_{o}$ and $\\hat{F}_r$ )." + ], + "table_footnote": [], + "table_body": "
TaskSplit PointOrg. RegionTrun. R. (VTM)Trun. R. (Hyperprior)
Cls\\(SP_{DS}\\)[-542.30, 94.14][-20, 20][-5, 5]
Seg\\(SP_{DS}\\)[-506.97, 105.95][-20, 20][-5, 5]
Dpt\\(SP_{DM1}\\)[-2.38, 3.27][-1, 1][-1, 1]
\\(SP_{DM2}\\)[-26.44, 5.03][-2, 2][-2, 2]
\\(SP_{DM3}\\)[-323.30, 25.05][-10, 10][-10, 10]
\\(SP_{DM4}\\)[-504.43, 100.27][-20, 20][-10, 10]
CSR\\(SP_G\\)[-71.50, 47.75][-5, 5][-5, 5]
TTI\\(SP_H\\)[-5.79, 4.46][-5.79, 4.46][-5.79, 4.46]
", + "bbox": [ + 96, + 304, + 478, + 426 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Summary of the original and truncation region settings.", + "bbox": [ + 94, + 436, + 475, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "large model feature coding and provide insights for interested researchers.", + "bbox": [ + 89, + 476, + 482, + 506 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.1. The baseline methods", + "text_level": 1, + "bbox": [ + 89, + 516, + 294, + 531 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The pipeline for the proposed baselines, illustrated in Fig. 4, includes pre-processing, core codec, and post-processing modules. First, the original feature $F_{o}$ is truncated to a smaller range and then quantized into integers. Next, the quantized $F_{q}$ is packed into a 2D feature $F_{p}$ . The codec takes pre-processed $F_{p}$ as the input and outputs decoded feature $\\hat{F}_{p}$ . Two core codecs are used: a handcrafted codec VTM and a learning-based codec Hyperprior. In the post-processing module, $\\hat{F}_{p}$ is unpacked to $\\hat{F}_{q}$ and then converted back into a floating-point feature $\\hat{F}_{r}$ .", + "bbox": [ + 88, + 537, + 482, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.1.1. Pre-processing and post-processing", + "text_level": 1, + "bbox": [ + 89, + 700, + 382, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The truncation operation is proposed to remove feature points that deviate significantly from the central region. The specific truncated regions for each task are listed in Table 4. Different truncations are used for the two baselines for their distinct coding strategies. The quantization operation uniformly quantizes the truncated features to 10-bit integers. Given the input requirement of VTM, we pack the features into a 2D YUV-400 format. The same packing applies in the Hyperprior baseline. $F_{Seg}$ , $F_{Dpt}$ , and $F_{TTI}$ are packed into the shapes of $2740 \\times 1536$ , $3222 \\times 6144$ , and $512 \\times 512$ , respectively. Please refer to supplementary materials for the packing details. The post-processing performs unpacking", + "bbox": [ + 88, + 719, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/08c33940578f9d96b74a21eac8f66689400098e8b2b51485485bdd528ae602a7.jpg", + "image_caption": [ + "Figure 5. Rate-accuracy-drop (R-AD) comparisons among different tasks. Left: VTM baseline. Right: Hyperprior baseline." + ], + "image_footnote": [], + "bbox": [ + 517, + 305, + 709, + 411 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3b74e4b041fe53f57d9d431048afbbb86a5a2eff29e5313509193bbdceac2db8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 305, + 901, + 411 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and de-quantization, which are inverse processes of packing and quantization.", + "bbox": [ + 511, + 479, + 905, + 510 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.1.2. Encoding and decoding", + "text_level": 1, + "bbox": [ + 511, + 518, + 723, + 534 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "VTM baseline: We employ VTM-23.3 as the codec. The Intra coding is applied using the main configuration file encoder_intra_vtm.cfg. The InputChromaFormat is set to 400, and ConformanceWindowMode is enabled. The Internal-BitDepth, InputBitDepth, and OutputBitDepth are all set to 10. All other configurations are set to default. In our experiments, five QPs {22, 27, 32, 37, 42} are used.", + "bbox": [ + 511, + 537, + 905, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Hyperprior baseline: We modify the channel of the input and output of the original Hyperprior as 1 to accommodate 2D features. The loss function is defined as:", + "bbox": [ + 511, + 643, + 905, + 688 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nL = B P F P + \\lambda \\times \\left\\| \\left(F _ {o} - \\hat {F} _ {r}\\right) \\right\\| ^ {2} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 700, + 903, + 719 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\lambda$ is a scaling factor used to adjust the bitrate. Please find more details in the supplementary.", + "bbox": [ + 511, + 731, + 905, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2. Rate-accuracy (R-A) analysis", + "text_level": 1, + "bbox": [ + 511, + 772, + 774, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The R-A evaluation results are presented in Table 3. Both baselines can achieve significant bitrate savings, while the VTM baseline outperforms the Hyperprior baseline. To further assess the impact of bitrate on accuracy, we introduce a new metric, accuracy drop (AD), defined as the percentage decrease in accuracy. For Dpt, the reciprocal of RMSE is employed. The R-AD curves, shown in Fig. 5, reveal", + "bbox": [ + 511, + 794, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1074", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5d5ec203a17cfcbc15d47fa78173195c99761e6f7e84d0178a4a418160e40e2d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CodecClsSegDptCSRTTI
VTM Baseline0.94130.89550.85490.74900.9983
Hyperprior Baseline0.92960.92800.74370.47130.9996
", + "bbox": [ + 104, + 89, + 470, + 136 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "that most tasks exhibit an inflection point where accuracy starts to decrease significantly. However, for Dpt, this point is observed only in the Hyperprior baseline. In the Hyperprior baseline, inflection points are more distracted across tasks, which may caused by the fact that different models are trained for different tasks. TTI is excluded as images generated from the original features have an FID of 0.", + "bbox": [ + 89, + 188, + 480, + 294 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.3. Distortion-accuracy (D-A) analysis", + "text_level": 1, + "bbox": [ + 89, + 303, + 393, + 318 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluate the linear correlation between feature MSE (D) and task accuracy (A) and present coefficients of determination in Table 5. Among the baselines, only TTI exhibits high linear correlations. This is reasonable since TTI involves generating images, and the baselines employ image CODECs. In contrast, CSR fails to achieve a high linear correlation, as its features are derived from textual data. Overall, the VTM baseline outperforms the Hyperprior baseline, demonstrating its superior adaptability to large model features. Our analysis reveals that feature MSE struggles to characterize semantic distortion, particularly for textual data. This finding suggests that feature MSE is not an effective metric for measuring semantic distortion.", + "bbox": [ + 89, + 324, + 482, + 521 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.4. Complexity analysis", + "text_level": 1, + "bbox": [ + 89, + 530, + 282, + 546 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We run the VTM baseline and Hyperprior baseline on an Intel® Xeon® E5-2690 v4 CPU and a single NVIDIA GeForce RTX 4090 GPU, respectively. The VTM baseline exhibits encoding times ranging from a few seconds to several hours, while its decoding time remains within a few seconds. In contrast, the Hyperprior baseline exhibits comparable encoding and decoding times, most of which are on the order of milliseconds. Detailed results are provided in the supplementary. We note that runtime depends heavily on both hardware capabilities and implementation details. A well-optimized codec and high-performance hardware can significantly accelerate feature coding.", + "bbox": [ + 89, + 551, + 482, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.5. Generalizability analysis", + "text_level": 1, + "bbox": [ + 89, + 742, + 316, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We assess the generalizability of trained models by compressing features extracted from one task using models trained on a different task and evaluating their R-A performance. We select three tasks and present the results in Table 6. Overall, for a given task (along column), three kinds of models behave differently in both bitrate and accuracy. For example, models trained on CSR struggle to achieve low bitrates when applied to Seg and TTI, while models trained on Seg fail to achieve low FID in TTI. More analyses can", + "bbox": [ + 89, + 763, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/65c037153d537746857e86ad343a0fb211f23346206ab2abbaf1d71993a04d38.jpg", + "table_caption": [ + "Table 5. Distortion-accuracy (D-A) analysis on the two baselines." + ], + "table_footnote": [], + "table_body": "
TaskSegCSRTTI
ModelsBPFPmIoUBPFPAcc.BPFPFID
Trained on Seg1.7177.962.0883.800.97119.83
1.3077.271.3766.600.62208.50
0.5474.820.780.200.31333.75
0.1262.580.250.000.12363.32
0.0337.110.090.000.05354.22
Trained on TTI2.5878.302.2897.601.4425.25
1.9676.341.4677.800.6652.12
1.1273.370.6532.000.2895.03
0.5356.900.300.000.15124.84
0.2642.940.110.000.08170.01
Trained on CSR5.2978.206.3491.404.7758.19
3.2777.943.6087.802.5973.19
1.8076.771.6882.401.02196.56
1.7375.821.5057.801.02191.69
1.5475.801.3534.800.94193.08
", + "bbox": [ + 526, + 89, + 897, + 303 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Generalizability evaluation on Seg, CSR, and TTI tasks.", + "bbox": [ + 514, + 315, + 900, + 329 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "be found in the supplementary. In practical applications, an effective compression model should be capable of handling diverse inputs. Therefore, a compression method with high generalizability is highly desirable.", + "bbox": [ + 511, + 356, + 903, + 417 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.6. Discussion", + "text_level": 1, + "bbox": [ + 511, + 426, + 630, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Large model applications require a balance of high accuracy, low complexity, and strong generalizability. However, both baselines struggle to maintain high accuracy at low bitrates, underscoring the need for more efficient feature coding strategies tailored to the unique characteristics of large model features. In addition, semantic distortion measurement remains a critical challenge in feature coding. Advancing semantic distortion metrics could further optimize codec performance and enhance overall coding efficiency.", + "bbox": [ + 511, + 449, + 903, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The two baseline approaches present distinct advantages and drawbacks. The handcrafted approach aligns seamlessly with existing codecs but suffers from excessive runtime, limiting its real-time applicability. In contrast, the learning-based approach benefits from efficient GPU acceleration but lacks generalizability, as separate models must be trained for different features. Therefore, developing lightweight and generic codecs remains a promising research direction for feature coding.", + "bbox": [ + 511, + 585, + 903, + 722 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 736, + 633, + 752 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we highlight the importance of feature coding in large model deployments and introduce a new research area: large model feature coding. As the first step, we provide a test dataset, unified test conditions, baseline methods, and a benchmark. We examine the applicability of the two baselines and identify promising directions for future research. Our aim is to encourage collaboration between image/video coding and large model communities to drive the development of feature coding.", + "bbox": [ + 511, + 760, + 903, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1075", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Johannes Balle, David C. Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. ArXiv, abs/1802.01436, 2018. 2", + "[2] Benjamin Bross, Ye-Kui Wang, Yan Ye, Shan Liu, Jianle Chen, Gary J. Sullivan, and Jens-Rainer Ohm. Overview of the versatile video coding (VVC) standard and its applications. IEEE Transactions on Circuits and Systems for Video Technology, 31(10):3736-3764, 2021. 2", + "[3] Yangang Cai, Peiyin Xing, and Xuesong Gao. High efficient 3D convolution feature compression. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2022. 2", + "[4] Jingxue Chen, Hang Yan, Zhiyuan Liu, Min Zhang, Hu Xiong, and Shui Yu. When federated learning meets privacy-preserving computation. ACM Computing Surveys, 56(12): 1-36, 2024. 1, 3", + "[5] Qiaoxi Chen, Changsheng Gao, and Dong Liu. End-to-end learned scalable multilayer feature compression for machine vision tasks. In ICIP, pages 1781-1787, 2024. 2", + "[6] Zhuo Chen, Kui Fan, Shiqi Wang, Lingyu Duan, Weisi Lin, and Alex Chichung Kot. Toward intelligent sensing: Intermediate deep feature compression. IEEE Transactions on Image Processing, 29:2230-2243, 2020. 1", + "[7] Hyomin Choi and Ivan V. Bajic. Latent-space scalability for multi-task collaborative intelligence. In ICIP, pages 3562-3566, 2021. 2", + "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: a large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 5", + "[9] Ronan Docherty, Antonis Vamvakeros, and Samuel J Cooper. Upsampling DINoV2 features for unsupervised vision tasks and weakly supervised materials segmentation. arXiv preprint arXiv:2410.19836, 2024. 3", + "[10] Lingyu Duan, Jiaying Liu, Wenhan Yang, Tiejun Huang, and Wen Gao. Video coding for machines: A paradigm of collaborative compression and intelligent analytics. IEEE Transactions on Image Processing, 29:8680-8695, 2020. 2", + "[11] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The Llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2, 4, 6", + "[12] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 2, 4, 6", + "[13] Ruoyu Feng, Xin Jin, Zongyu Guo, Runsen Feng, Yixin Gao, Tianyu He, Zhizheng Zhang, Simeng Sun, and Zhibo Chen. Image coding for machines with omnipotent feature learning. In ECCV, pages 510-528. Springer, 2022. 2", + "[14] Othmane Friha, Mohamed Amine Ferrag, Burak Kantarci, Burak Cakmak, Arda Ozgun, and Nassira Ghoualmi-Zine. LLM-based edge intelligence: A comprehensive survey on architectures, applications, security and trustworthiness." + ], + "bbox": [ + 93, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE Open Journal of the Communications Society, 5:5799-5856, 2024. 1, 3", + "[15] Changsheng Gao, Dong Liu, Li Li, and Feng Wu. Towards task-generic image compression: A study of semantics-oriented metrics. IEEE Transactions on Multimedia, 25:721-735, 2023. 2", + "[16] Changsheng Gao, Yiheng Jiang, Li Li, Dong Liu, and Feng Wu. DMOFC: discrimination metric-optimized feature compression. In PCS, pages 1–5, 2024. 1", + "[17] Changsheng Gao, Zhuoyuan Li, Li Li, Dong Liu, and Feng Wu. Rethinking the joint optimization in video coding for machines: A case study. In DCC, pages 556-556, 2024. 2", + "[18] Changsheng Gao, Yiheng Jiang, Siqi Wu, Yifan Ma, Li Li, and Dong Liu. IMOFC: identity-level metric optimized feature compression for identification tasks. IEEE Transactions on Circuits and Systems for Video Technology, 35(2):1855-1869, 2025. 1", + "[19] Jean Kaddour, Joshua Harris, Maximilian Mozes, Herbie Bradley, Roberta Raileanu, and Robert McHardy. Challenges and applications of large language models. arXiv preprint arXiv:2307.10169, 2023. 1", + "[20] Chia-Hao Kao, Cheng Chien, Yu-Jen Tseng, Yi-Hsin Chen, Alessandro Gnutti, Shao-Yuan Lo, Wen-Hsiao Peng, and Riccardo Leonardi. Bridging compressed image latents and multimodal large language models. arXiv preprint arXiv:2407.19651, 2024. 2", + "[21] Yeongwoong Kim, Hyewon Jeong, Janghyun Yu, Younhee Kim, Jooyoung Lee, Se Yoon Jeong, and Hui Yong Kim. End-to-end learnable multi-scale feature compression for VCM. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2023. 2", + "[22] Jin Han Lee, Myung-Kyu Han, Dong Wook Ko, and Il Hong Suh. From big to small: Multi-scale local planar guidance for monocular depth estimation. arXiv preprint arXiv:1907.10326, 2019. 5", + "[23] Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan First, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. GShard: Scaling giant models with conditional computation and automatic sharding. arXiv preprint arXiv:2006.16668, 2020. 1", + "[24] Shibao Li, Chenxu Ma, Yunwu Zhang, Longfei Li, Chengzhi Wang, Xuerong Cui, and Jianhang Liu. Attention-based variable-size feature compression module for edge inference. The Journal of Supercomputing, 2023. 2", + "[25] Zhuoyuan Li, Junqi Liao, Chuanbo Tang, Haotian Zhang, Yuqi Li, Yifan Bian, Xihua Sheng, Xinmin Feng, Yao Li, Changsheng Gao, et al. Ustc-td: A test dataset and benchmark for image and video coding in 2020s. arXiv preprint arXiv:2409.08481, 2024. 2", + "[26] Zhuoyuan Li, Zikun Yuan, Li Li, Dong Liu, Xiaohu Tang, and Feng Wu. Object segmentation-assisted inter prediction for versatile video coding. IEEE Transactions on Broadcasting, 2024. 2", + "[27] Tie Liu, Mai Xu, Shengxi Li, Chaoran Chen, Li Yang, and Zhuoyi Lv. Learn mutual feature compression for machine vision. In ICASSP, pages 1-5, 2023. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1076", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Guo Lu, Xingtong Ge, Tianxiong Zhong, Qiang Hu, and Jing Geng. Preprocessing enhanced image compression for machine vision. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2024. 2", + "[29] Lingjuan Lyu, Han Yu, Xingjun Ma, Chen Chen, Lichao Sun, Jun Zhao, Qiang Yang, and Philip S. Yu. Privacy and robustness in federated learning: Attacks and defenses. IEEE Transactions on Neural Networks and Learning Systems, 35 (7):8726-8746, 2024. 3", + "[30] Yifan Ma, Changsheng Gao, Qiaoxi Chen, Li Li, Dong Liu, and Xiaoyan Sun. Feature compression with 3d sparse convolution. In VCIP, pages 1-5, 2024. 2", + "[31] Rui Mao, Xinmin Feng, Changsheng Gao, Li Li, Dong Liu, and Xiaoyan Sun. Perceptual image compression with conditional diffusion transformers. In VCIP, pages 1-5, 2024. 2", + "[32] Kiran Misra, Tianying Ji, Andrew Segall, and Frank Bossen. Video feature compression for machine tasks. In ICME, pages 1-6, 2022. 2", + "[33] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. DINOv2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 4, 6", + "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 3", + "[35] Xihua Sheng, Li Li, Dong Liu, and Houqiang Li. Vnvc: A versatile neural video coding framework for efficient human-machine vision. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(7):4579-4596, 2024. 2", + "[36] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-CLIP: improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 3", + "[37] Satoshi Suzuki, Shoichiro Takeda, Motohiro Takagi, Ryuichi Tanida, Hideaki Kimata, and Hayaru Shouno. Deep feature compression using spatio-temporal arrangement toward collaborative intelligent world. IEEE Transactions on Circuits and Systems for Video Technology, 32(6):3934-3946, 2022. 2", + "[38] Yuanyishu Tian, Yao Wan, Lingjuan Lyu, Dezhong Yao, Hai Jin, and Lichao Sun. FedBERT: when federated learning meets pre-training. ACM Transactions on Intelligent Systems and Technology, 13(4):1-26, 2022. 1", + "[39] Yuan Tian, Guo Lu, Guangtao Zhai, and Zhiyong Gao. Nonsemantics suppressed mask learning for unsupervised video semantic compression. In ICCV, pages 13564-13576, 2023. 2", + "[40] Yuan Tian, Guo Lu, and Guangtao Zhai. SMC++: masked learning of unsupervised video semantic compression. arXiv preprint arXiv:2406.04765, 2024. 2", + "[41] Yuan Tian, Guo Lu, and Guangtao Zhai. Free-VSC: free semantics from visual foundation models for unsupervised video semantic compression. In ECCV, pages 163-183, 2025. 2" + ], + "bbox": [ + 91, + 90, + 485, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2", + "[43] Praneeth Vepakomma, Otkrist Gupta, Tristan Swedish, and Ramesh Raskar. Split learning for health: Distributed deep learning without sharing raw patient data. arXiv preprint arXiv:1812.00564, 2018. 1, 3", + "[44] Shurun Wang, Shiqi Wang, Wenhan Yang, Xinfeng Zhang, Shanshe Wang, Siwei Ma, and Wen Gao. Towards analysis-friendly face representation with scalable feature and texture compression. IEEE Transactions on Multimedia, 24:3169-3181, 2022. 1", + "[45] Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng Chua. NExT-GPT: Any-to-any multimodal LLM. In ICML, pages 53366-53397, 2024. 5", + "[46] Biwei Yan, Kun Li, Minghui Xu, Yuean Dong, Yue Zhang, Zhaochun Ren, and Xiuzhen Cheng. On protecting the data privacy of large language models (LLMs): A survey. arXiv preprint arXiv:2403.05156, 2024. 3", + "[47] Ning Yan, Changsheng Gao, Dong Liu, Houqiang Li, Li Li, and Feng Wu. SSSIC: Semantics-to-signal scalable image coding with learned structural representations. IEEE Transactions on Image Processing, 30:8939-8954, 2021. 2", + "[48] Wenhan Yang, Haofeng Huang, Yueyu Hu, Ling-Yu Duan, and Jiaying Liu. Video coding for machines: Compact visual representation compression for intelligent collaborative analytics. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-18, 2024. 2", + "[49] Rui Ye, Wenhao Wang, Jingyi Chai, Dihan Li, Zexi Li, Yinda Xu, Yaxin Du, Yanfeng Wang, and Siheng Chen. OpenFedLLM: Training large language models on decentralized private data via federated learning. In ACM SIGKDD, pages 6137-6147, 2024. 1, 3", + "[50] Xu Zhang, Peiyao Guo, Ming Lu, and Zhan Ma. All-in-one image coding for joint human-machine vision with multipath aggregation. arXiv preprint arXiv:2409.19660, 2024. 2", + "[51] Zhicong Zhang, Mengyang Wang, Mengyao Ma, Jiahui Li, and Xiaopeng Fan. MSFC: Deep feature compression in multi-task network. In ICME, pages 1-6, 2021. 2", + "[52] Jiaying Zheng, Hainan Zhang, Lingxiang Wang, Wangjie Qiu, Hongwei Zheng, and Zhiming Zheng. Safely learning with private data: A federated learning framework for large language model. arXiv preprint arXiv:2406.14898, 2024. 1, 3" + ], + "bbox": [ + 514, + 90, + 905, + 753 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1077", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Feature Coding in the Era of Large Models_ Dataset, Test Conditions, and Benchmark/83895aed-d59c-465f-afa3-c8ed37ae4a4b_model.json b/2025/Feature Coding in the Era of Large Models_ Dataset, Test Conditions, and Benchmark/83895aed-d59c-465f-afa3-c8ed37ae4a4b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3f6f4f70a372f6cfa74765098901ab0f7f841fb9 --- /dev/null +++ b/2025/Feature Coding in the Era of Large Models_ Dataset, Test Conditions, and Benchmark/83895aed-d59c-465f-afa3-c8ed37ae4a4b_model.json @@ -0,0 +1,2321 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.277, + 0.131, + 0.723, + 0.175 + ], + "angle": 0, + "content": "Feature Coding in the Era of Large Models: Dataset, Test Conditions, and Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.203, + 0.91, + 0.294 + ], + "angle": 0, + "content": "Changsheng Gao\\(^{1}\\), Yifan Ma\\(^{2}\\), Qiaoxi Chen\\(^{2}\\), Yenan Xu\\(^{2}\\), Dong Liu\\(^{2}\\), Weisi Lin\\(^{1\\dagger}\\) \n\\(^{1}\\)Media Interactive Computing Lab, Nanyang Technological University, Singapore 639798 \n\\(^{2}\\)MOE Key Laboratory of Brain-Inspired Intelligent Perception and Cognition, University of Science and Technology of China, Hefei 230093, China \n{changsheng.gao, wslin} @ntu.edu.sg; {mayf,xxii,yenanxu} @mail.ustc.edu.cn; dongeliu@ustc.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.359, + 0.486, + 0.722 + ], + "angle": 0, + "content": "Large models have achieved remarkable performance across various tasks, yet they incur significant computational costs and privacy concerns during both training and inference. Distributed deployment has emerged as a potential solution, but it necessitates the exchange of intermediate information between model segments, with feature representations serving as crucial information carriers. To optimize information exchange, feature coding is required to reduce transmission and storage overhead. Despite its importance, feature coding for large models remains an under-explored area. In this paper, we draw attention to large model feature coding and make three fundamental contributions. First, we introduce a comprehensive dataset encompassing diverse features generated by three representative types of large models. Second, we establish unified test conditions, enabling standardized evaluation pipelines and fair comparisons across future feature coding studies. Third, we introduce two baseline methods derived from widely used image coding techniques and benchmark their performance on the proposed dataset. These contributions aim to provide a foundation for future research and inspire broader engagement in this field. To support a long-term study, all source code and the dataset are made available at https://github.com/chansongoal/LaMoFC." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.753, + 0.222, + 0.769 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.779, + 0.484, + 0.84 + ], + "angle": 0, + "content": "Large models have revolutionized artificial intelligence (AI) with their impressive performance. However, they require substantial training data and extensive computational resources, posing significant challenges for practical de" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.328, + 0.908, + 0.466 + ], + "angle": 0, + "content": "ployments [19]. To address these challenges, distributed deployment has been proposed as a promising solution [14, 23, 38, 43, 49, 52]. By distributing the computational workload across multiple platforms, this approach reduces the resource strain on individual systems and enables the integration of private client data into training while maintaining data security [4, 49]. As model scale continues to grow, distributed deployment is expected to become a mainstream strategy for large model deployments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.467, + 0.909, + 0.604 + ], + "angle": 0, + "content": "In distributed deployments, effective information exchange between model segments is essential. During forward propagation, features generated by early segments are passed to later segments to complete the inference. Given the vast amount of data in large model applications, it is necessary to encode features to a smaller size to reduce storage and transmission burden. Analogous to image and video coding, feature coding has been proposed to address this problem [6, 16, 18, 44]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.607, + 0.909, + 0.759 + ], + "angle": 0, + "content": "Existing feature coding research faces three key limitations: narrow model types, limited task diversity, and constrained data modalities. Most existing feature coding efforts focus on small-scale convolution-based networks, such as ResNet, ignoring large models. Furthermore, existing research primarily targets visual features and neglects the investigation of textual features from large language models (LLMs). This gap indicates a need for broader exploration across both models and data modalities to advance feature coding." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.762, + 0.909, + 0.883 + ], + "angle": 0, + "content": "In this work, we shift the feature coding research from small-scale to large-scale models. To address the existing limitations, we construct a comprehensive test dataset with features from five tasks across three large models, covering both visual and textual data for increased modality diversity. Furthermore, we establish unified test conditions to enable fair comparisons, including formulated bitrate computation and task-specific evaluation pipelines." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "To lay the groundwork for large model feature coding," + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.852, + 0.484, + 0.9 + ], + "angle": 0, + "content": "This work was supported by the Ministry of Education of Singapore under Grant T2EP20123-0006. We acknowledge the support of GPU cluster built by MCC Lab of Information Science and Technology Institution, USTC." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1068" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.09, + 0.382, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.089, + 0.608, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.089, + 0.895, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.279, + 0.907, + 0.377 + ], + "angle": 0, + "content": "Figure 1. Three application scenarios for feature coding in large model deployments. Cloud-centralized training: Features generated by foundation models are encoded and stored in a feature database. This database supports various downstream tasks, reducing computational demands by avoiding repeated inference of large foundation models. Cloud-edge distributed training: To protect user data privacy, raw data is first processed by an edge model to generate features, which are then encoded and transmitted to the cloud. The cloud decodes these features to complete the remaining forward propagation. Cloud-edge distributed inference: A large model is partitioned across multiple parts, with portions offloaded to edge devices to distribute the computational load. The transmitted features ensure privacy protection and enable customized downstream tasks. Please refer to Sec. 3 for detailed illustrations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.403, + 0.484, + 0.553 + ], + "angle": 0, + "content": "we propose two baseline methods and evaluate their applicability to the dataset. The baselines leverage two image coding methods: the handcrafted VVC Intra coding [2] and the learning-based Hyperprior method [1]. We introduce pre-processing and post-processing modules to adapt the baselines to feature characteristics. The resulting benchmark and analyses provide valuable insights into large model feature coding and identify promising directions for further exploration. In summary, our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.554, + 0.483, + 0.614 + ], + "angle": 0, + "content": "- We highlight the importance of feature coding in large model deployments, identify relevant application scenarios, and introduce a new research area: large model feature coding." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.615, + 0.483, + 0.675 + ], + "angle": 0, + "content": "- We construct a comprehensive feature dataset encompassing three model types, five tasks, and two kinds of source data modalities, along with unified test conditions to support long-term large model feature coding research." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.675, + 0.483, + 0.72 + ], + "angle": 0, + "content": "- We propose two baselines and evaluate their usability to the large model feature coding. We build a benchmark and suggest promising directions for further research." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.554, + 0.483, + 0.72 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.734, + 0.35, + 0.751 + ], + "angle": 0, + "content": "2. Background and motivation" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.758, + 0.243, + 0.774 + ], + "angle": 0, + "content": "2.1. Feature coding" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Originally introduced within the Coding for Machines framework [10, 48], feature coding is one of two primary branches alongside visual coding. Unlike visual coding [15, 25, 26, 28, 31, 35, 39, 40, 50], which encodes and reconstructs the original visual data, feature coding involves encoding features extracted from source data into bitstreams. While visual coding has already been explored in the context of large models [20, 41], feature coding re" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.404, + 0.736, + 0.418 + ], + "angle": 0, + "content": "mains unexplored in this domain." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.429, + 0.907, + 0.732 + ], + "angle": 0, + "content": "Despite recent progress, feature coding research still faces three key limitations. First, there is a restricted variety of model types in use. The existing feature coding focuses on small-scale CNN features such as [3, 17, 21, 24, 27, 30, 37]. However, as Transformer architectures now dominate large model research [11, 12, 33, 42], it is crucial to investigate feature coding for Transformer-based models. Second, the field is constrained by the limited scope of task types. Existing studies primarily address discriminative tasks such as image classification and segmentation [5, 7, 13, 32, 47, 51], while generative tasks remain largely unexplored. Given the significance of generative models, extending feature coding to include these tasks, such as image synthesis, is essential. The third limitation is the constrained range of source modality. Current feature coding research predominantly focuses on features extracted from visual data, overlooking the increasingly important features generated from textual data, particularly in light of advancements in LLMs. Expanding to multiple modalities is crucial as the diversity of AI applications grows." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.74, + 0.906, + 0.846 + ], + "angle": 0, + "content": "Furthermore, two challenges hinder progress in large model feature coding: the lack of a public test dataset and unified test conditions. Without a standardized dataset and evaluation conditions, fair comparisons across methods can not be conducted. A public dataset ensures consistent input data across studies, while unified test conditions establish a standard pipeline to evaluate performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "These limitations motivate us to construct a diverse test dataset and formulate unified test conditions to support large model feature coding research." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1069" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.908, + 0.187 + ], + "angle": 0, + "content": "
ModelParametersTaskSplit PointFeature ShapeSource Data
DINOv21.1BImage Classification (Cls)40th ViT Block257 × 1536ImageNet, 500 samples
Semantic Segmentation (Seg)40th ViT Block2 × 1370 × 1536VOC2012, 100 samples
Depth Estimation (Dpt)10th, 20th, 30th, 40th ViT Blocks2 × 4 × 1611 × 1536NYU-Depth-v2, 80 samples
Llama38BCommon Sense Reasoning (CSR)32th Decoder LayerN × 4096Arc-Challenge, 500 samples
SD38BText-to-Image Synthesis (TTI)Input Layer of VAE Decoder16 × 128 × 128COCO2017, 500 samples
" + }, + { + "type": "table_caption", + "bbox": [ + 0.207, + 0.197, + 0.788, + 0.21 + ], + "angle": 0, + "content": "Table 1. Summary of the model and task selection, split point decision, and source data collection." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.237, + 0.32, + 0.254 + ], + "angle": 0, + "content": "2.2. Large model deployment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.259, + 0.483, + 0.379 + ], + "angle": 0, + "content": "The impressive performance of large models is driven by scaling laws [34, 36], which emphasize the importance of vast model sizes and extensive training data. For example, GPT-3 was built with 175 billion parameters and trained on 499 trillion tokens, both contributing to substantial computational demands. To alleviate these resource requirements, distributed training and inference methods have been proposed in [14, 49, 52]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.381, + 0.483, + 0.44 + ], + "angle": 0, + "content": "User privacy protection presents another challenge in client services [4, 29, 46]. Split learning has been proposed as an effective solution to protect data privacy by partitioning models between clients and the cloud [49]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.441, + 0.483, + 0.545 + ], + "angle": 0, + "content": "In distributed deployments, efficient information exchange between model segments is crucial. However, academic research has largely overlooked the storage and transmission costs associated with feature exchange. To address this gap, we propose to encode features into compact bitstreams to improve the efficiency of large model deployments in distributed environments." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.56, + 0.373, + 0.577 + ], + "angle": 0, + "content": "3. Feature coding in large models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.586, + 0.483, + 0.632 + ], + "angle": 0, + "content": "In Fig. 1, we illustrate three application scenarios of feature coding in large model deployments, highlighting their roles in optimizing the deployment efficiency of large models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.641, + 0.292, + 0.657 + ], + "angle": 0, + "content": "3.1. Large model training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.663, + 0.483, + 0.769 + ], + "angle": 0, + "content": "Large models achieve their high performance through extensive training on large-scale datasets. Cloud-centralized training is commonly employed to leverage the vast computational resources and data in the cloud. However, in sensitive domains like healthcare, high-quality public data may be scarce, necessitating access to private data through cloud-edge collaborative training." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.777, + 0.323, + 0.792 + ], + "angle": 0, + "content": "3.1.1. Cloud-centralized training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In cloud-centralized training, source databases are maintained in the cloud, where intensive computational demands become the main challenge. Complex tasks are increasingly addressed through the collaborative use of multiple models. For instance, in vision-language tasks, visual data is first processed by a vision model, and its features are then passed to an LLM for advanced reasoning. Similarly, foundation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.238, + 0.905, + 0.343 + ], + "angle": 0, + "content": "models often serve as core backbones for various tasks. For example, features extracted from DINOv2 are repurposed for downstream tasks like segmentation [9]. In such collaborative workflows, features generated by one model become inputs for others. However, during training, each data sample may undergo repeated inference, resulting in substantial and unnecessary computational costs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.344, + 0.907, + 0.555 + ], + "angle": 0, + "content": "To address this, an effective solution is to generate features once, store them, and reuse them across tasks. The workflow is outlined on the left of Fig. 1. First, a foundation model extracts features from the source database, which are then encoded and stored in a feature database. For subsequent task-specific training, only the precomputed feature database is needed, enabling efficient access to features without redundant feature extraction. This approach significantly reduces computational load and accelerates downstream task training. However, raw feature data consumes considerable storage space, particularly with large datasets. Therefore, encoding raw features into compact bitstreams becomes essential to reduce storage demands and support efficient large model training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.564, + 0.782, + 0.58 + ], + "angle": 0, + "content": "3.1.2. Cloud-edge distributed training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.906, + 0.793 + ], + "angle": 0, + "content": "In addition to computational resources, large model training requires access to large-scale datasets, which are often only available to a few major high-tech companies. Furthermore, sensitive data, such as medical records and financial information, cannot be freely shared due to privacy concerns and regulatory constraints. To leverage distributed, privately owned data, split learning has emerged as a viable solution [43]. Split learning divides the training process into two phases, with the bulk of the computational load handled by the cloud, while edge devices, which are often resource-constrained, handle a smaller portion of the task. This approach minimizes the computational demand on the edge while enabling collaboration with powerful cloud resources." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.901 + ], + "angle": 0, + "content": "The cloud-edge distributed training pipeline is depicted in the middle of Fig. 1. The whole pipeline is divided into two parts: edge models and cloud models. During forward propagation, the edge model processes private data to generate features, which are then encoded into bitstreams and transmitted to the cloud. Upon receiving the bitstreams, the cloud decodes them back into features and feeds them into" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1070" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.145, + 0.087, + 0.319, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.323, + 0.087, + 0.497, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.087, + 0.674, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.678, + 0.087, + 0.853, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.144, + 0.186, + 0.318, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.322, + 0.186, + 0.496, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.186, + 0.674, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.678, + 0.186, + 0.853, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.144, + 0.283, + 0.318, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.322, + 0.283, + 0.496, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.283, + 0.674, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.678, + 0.283, + 0.853, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.392, + 0.908, + 0.448 + ], + "angle": 0, + "content": "Figure 2. Frequency and CDF comparisons between features in the proposed dataset and commonly used existing features. The proposed features exhibit distinct distributions compared to existing features, highlighting the necessity and value of the dataset. In addition, the proposed features demonstrate greater diversity, enhancing their representativeness and suitability for long-term use. Frequencies are scaled logarithmically for better visualization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.474, + 0.483, + 0.534 + ], + "angle": 0, + "content": "the cloud model. In the backward pass, gradients are computed and backpropagated along the reverse path. To reduce bandwidth cost and minimize latency, it is critical to apply feature coding before transmission." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.546, + 0.379, + 0.561 + ], + "angle": 0, + "content": "3.2. Cloud-edge distributed inference" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.483, + 0.734 + ], + "angle": 0, + "content": "Once large models are trained, they are deployed in products to provide client services. For a product or service to succeed, two primary concerns must be addressed: supporting a high volume of simultaneous cloud server access and protecting user privacy. For the first concern, managing computational load is essential, particularly in generative applications that involve repetitive diffusion processes. A practical solution is to offload part of this computational demand to edge devices. For the second concern, the same strategy in the cloud-edge distributed training, as discussed in Sec. 3.1.2, can be applied." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.902 + ], + "angle": 0, + "content": "The cloud-edge distributed inference pipeline is illustrated on the right of Fig. 1. In the upload phase, source data is converted into features by an edge model on the client side. These features are then encoded and transmitted to the cloud, where they are further processed by large models. In the download phase, the processed features are sent back to the client to complete a specific task. Depending on the task head used, various functionalities can be performed with the returned features. For example, Stable Diffusion 3 (SD3) features derived from an image can be used to synthesize video or generate segmentation masks. This approach not" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.474, + 0.907, + 0.52 + ], + "angle": 0, + "content": "only reduces the computational load on the cloud server but also enhances data privacy by keeping raw user data on the edge side." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.536, + 0.818, + 0.553 + ], + "angle": 0, + "content": "4. Dataset construction and analysis" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.562, + 0.708, + 0.575 + ], + "angle": 0, + "content": "4.1. Dataset construction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.906, + 0.659 + ], + "angle": 0, + "content": "To guarantee the dataset's representativeness and support the long-term study of feature coding research, we curate the test dataset with consideration of three key aspects. Table 1 outlines our selected models and tasks, split points, and source data." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.67, + 0.733, + 0.684 + ], + "angle": 0, + "content": "4.1.1. Model and task selection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Given the vast number of large models available, it is impractical to include all of them in a single research paper. Therefore, we focus on selecting representative models. To ensure the dataset's representativeness and long-term relevance, we choose one widely recognized model for each type of large model. Specifically, we select DINOv2 [33], Llama3 [11], and SD3 [12] as representatives of discriminative models, generative models, and hybrid models, respectively. The trend of adopting Transformer architectures in large models is expected to persist in the coming years. As a result, the selected models are anticipated to retain their research value and practical utility in the near future. DINOv2 and Llama3 handle visual and textual inputs, respectively, while SD3 processes textual inputs to generate visual out-" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1071" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.16, + 0.08, + 0.837, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.412, + 0.908, + 0.469 + ], + "angle": 0, + "content": "Figure 3. Visualization of the original feature blocks and their corresponding DCT blocks (absolute values). In the feature domain, the proposed features exhibit higher vertical redundancy, whereas the existing features vary smoothly in both directions. In the DCT domain, the proposed features display more dispersed energy distributions. These distinct redundancy characteristics indicate different coding properties. We use \\(7 \\times 7\\) blocks for \\(C5\\) and \\(C5\\)-ReLU and \\(16 \\times 16\\) blocks for other features." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.495, + 0.482, + 0.539 + ], + "angle": 0, + "content": "puts. Together, these models provide comprehensive feature representations across visual, textual, and cross-modal applications." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.54, + 0.483, + 0.661 + ], + "angle": 0, + "content": "For DINOv2, we include image classification (Cls), semantic segmentation (Seg), and depth estimation (Dpt) tasks. Llama3 is applied to the common sense reasoning task (CSR), and SD3 is used for text-to-image synthesis (TTI). The selection covers three visual tasks, one textual task, and one text-to-visual task, forming a comprehensive foundation for analyzing feature coding across diverse tasks and model types." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.67, + 0.273, + 0.685 + ], + "angle": 0, + "content": "4.1.2. Split point decision" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.825 + ], + "angle": 0, + "content": "The split points are decided to support various downstream tasks. For DINOv2, we define two types of split points: 1) \\(SP_{DS}\\), the output of the \\(40^{th}\\) ViT layer, and 2) \\(SP_{DM}\\), which aggregates outputs from the \\(10^{th}\\), \\(20^{th}\\), \\(30^{th}\\), and \\(40^{th}\\) ViT layers. \\(SP_{DS}\\) is used for tasks utilizing single-layer features, while \\(SP_{DM}\\) supports tasks that require multi-layer features. In this study, \\(SP_{DS}\\) is applied to Cls and Seg, and \\(SP_{DM}\\) to Dpt. The corresponding features are designated as \\(F_{Cls}\\), \\(F_{Seg}\\), and \\(F_{Dpt}\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.9 + ], + "angle": 0, + "content": "For Llama3, the split point \\(SP_{G}\\) is set at the output of the \\(32^{nd}\\) decoder layer. Features extracted at \\(SP_{G}\\) can support both task-specific heads and integration with other large models [45]. In CSR, \\(SP_{G}\\) produces \\(N \\times 4096\\) features, where \\(N\\) is the number of tokens." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.495, + 0.907, + 0.556 + ], + "angle": 0, + "content": "For SD3, we take the input layer of the VAE decoder as the split point \\( SP_{H} \\). This split point generates features of size \\( 16 \\times 128 \\times 128 \\), denoted as \\( F_{TTI} \\), which are then passed to the VAE decoder for image synthesis." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.563, + 0.715, + 0.577 + ], + "angle": 0, + "content": "4.1.3. Source data collection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.582, + 0.907, + 0.778 + ], + "angle": 0, + "content": "We organize our feature test dataset into five classes, each supported by a curated subset of publicly available datasets used as source data. For Cls, we select 500 images from ImageNet [8], each representing a unique class accurately classified by DINOv2. For Seg, we collect 100 images from VOC2012, ensuring coverage of all 20 object classes. For Dpt, we source 80 images from NYU-Depth-v2, adhering to the train-test split proposed in [22] and selecting 3-8 images for each scene category. For CSR, we utilize 500 samples from the Arc-Challenge dataset, chosen for their longest input prompts and correct prediction by Llama3. For TTI, we first select 500 images from COCO2017 and then collect their longest captions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.779, + 0.907, + 0.825 + ], + "angle": 0, + "content": "We noticed that learning-based methods require a dedicated training set. Please follow the source code to generate training data and find our training data in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.676, + 0.849 + ], + "angle": 0, + "content": "4.2. Feature analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "In this section, we compare the proposed features with those commonly used in existing research to demonstrate the necessity and importance of introducing a new dataset." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1072" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.162, + 0.089, + 0.418, + 0.189 + ], + "angle": 0, + "content": "
TaskHeadMetric
ClsNorm + Linear1Accuracy
SegNorm + Linear1mIoU
DptLinear4RMSE
CSRNorm + Linear + SoftmaxAccuracy
TTIVAE DecoderFID
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.199, + 0.483, + 0.228 + ], + "angle": 0, + "content": "Table 2. Summary of task head and accuracy metric settings. Find definitions of Heads in the original papers [11, 12, 33]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.253, + 0.283, + 0.267 + ], + "angle": 0, + "content": "4.2.1. Distribution analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.272, + 0.483, + 0.376 + ], + "angle": 0, + "content": "We first compare frequency distributions and cumulative distribution functions (CDFs) of the proposed and commonly used existing features in Fig. 2. \\(C5\\) and \\(C5\\)-ReLU are derived from the official ResNet50 pretrained on ImageNet, while \\(P\\)-layer features are generated by the ResNeXT101-based MaskRCNN pre-trained on COCO2017." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.378, + 0.483, + 0.527 + ], + "angle": 0, + "content": "For multi-layer split points, we observe four key distinctions between \\(SP_{DM}\\) and \\(P\\)-layer features: (1) \\(SP_{DM}\\) distribution range expands progressively with deeper layers; (2) \\(SP_{DM}\\) distribution is more asymmetric; (3) \\(SP_{DM}\\) feature values are more concentrated, as evidenced by CDFs; and (4) \\(SP_{DM}\\) features contain more regions with sparse feature points. For single-layer split points, \\(SP_{GS}\\) has a more concentrated distribution, while \\(SP_{H}\\) shows more peaks. The proposed dataset omits C5-ReLU-like features, as ReLU is rarely used in large models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.528, + 0.483, + 0.604 + ], + "angle": 0, + "content": "The proposed features exhibit different distributions compared to the existing features, highlighting the necessity of introducing the new dataset. Additionally, the high diversity of the proposed features enhances the dataset's representativeness and suitability for long-term research." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.611, + 0.285, + 0.626 + ], + "angle": 0, + "content": "4.2.2. Redundancy analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Since coding aims to remove redundancy in inputs, we conduct a redundancy comparison on feature blocks. We visualize the original feature blocks and their DCT blocks (absolute values) in Fig. 3. In the feature domain, clear spatial redundancy is observed in both \\(SP_{DM}\\) and \\(SP_{H}\\) features. \\(SP_{H}\\) exhibits redundancy in both horizontal and vertical directions, while \\(SP_{DM}\\) shows stronger vertical redundancy. In contrast, \\(SP_{GS}\\) features vary rapidly in both directions, resulting in weak spatial redundancy. This is because Llama3 only processes textual data, which lacks inherent spatial correlation. Existing features, on the other hand, exhibit smoother spatial variation and higher spatial redundancy. We attribute this to the translation invariance of convolutional layers, which preserve the relative positions of pixels in the feature domain, resulting in consistent redundancy patterns. In contrast, transformer-based vision models partition images into patches before transforming them, shifting spatial redundancy primarily to the vertical" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.088, + 0.905, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.169, + 0.88, + 0.183 + ], + "angle": 0, + "content": "Figure 4. The pipeline of the proposed baseline methods." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.207, + 0.704, + 0.221 + ], + "angle": 0, + "content": "direction, as seen in \\(SP_{DM}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.222, + 0.905, + 0.297 + ], + "angle": 0, + "content": "In the DCT domain, spatial redundancy is reflected in the energy distribution. The proposed features display more dispersed energy, whereas existing features concentrate energy in the top-left DC component. \\(SP_{H}\\) has the highest energy concentration as it closely resembles images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.298, + 0.905, + 0.358 + ], + "angle": 0, + "content": "These distinct redundancy characteristics further demonstrate the necessity of the proposed dataset and suggest that future studies should explore feature coding methods tailored to these unique properties." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.37, + 0.725, + 0.385 + ], + "angle": 0, + "content": "5. Unified test conditions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.395, + 0.905, + 0.44 + ], + "angle": 0, + "content": "Even with a test dataset, fair comparisons are still challenging without identical test conditions. To address this, we establish unified test conditions here." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.449, + 0.706, + 0.465 + ], + "angle": 0, + "content": "5.1. Bitrate computation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.471, + 0.907, + 0.728 + ], + "angle": 0, + "content": "We introduce a new bitrate measurement, Bits Per Feature Point (BPPF), replacing the commonly used Bits Per Pixel (BPP). BPPF is calculated by dividing the total coding bits by the number of feature points. Our rationale for shifting from BPP to BPPF is twofold. First, BPP is unsuitable for features extracted from non-visual data where pixels do not exist. Second, we believe bitrate should reflect the actual encoded data (the feature itself) rather than the source data. Using BPP to calculate bitrate on source data introduces ambiguity and limits fair comparisons between coding methods. For instance, applying the same BPP to different features extracted from the same image can be misleading. In addition, BPP may incentivize downscaling source images to reduce feature size, yielding lower bitrates that do not represent true coding efficiency. In contrast, BPPF measures bitrate on the feature itself, enabling fair comparisons across coding methods as long as the same feature is used." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.736, + 0.744, + 0.751 + ], + "angle": 0, + "content": "5.2. Task accuracy evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.758, + 0.905, + 0.832 + ], + "angle": 0, + "content": "The decoded features undergo task heads to perform accuracy evaluation. To isolate the impact of task heads, we choose simple task heads and fix their parameters during evaluation. The task heads and accuracy metrics are presented in Table 2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.845, + 0.753, + 0.861 + ], + "angle": 0, + "content": "6. Baselines and benchmark" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "As a starting point, we introduce two baselines and establish a benchmark. We aim to examine their applicability to" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1073" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.115, + 0.089, + 0.887, + 0.257 + ], + "angle": 0, + "content": "
TaskImage ClassificationSemantic SegmentationDepth EstimationCommon Sense ReasoningText-to-Image Synthesis
MetricBPFPAccuracyMSEBPFPmIoUMSEBPFPRMSEMSEBPFPAccuracyMSEBPFPFIDMSE
Original3210003283.390320.369503210003200
VTMBaseline1.9499.802.94991.7680.421.86682.210.50210.61082.6999.800.01091.416.140.0066
1.0397.403.16930.8879.312.08361.270.65110.68261.831000.02600.7419.940.0184
0.2374.603.69930.2373.472.54400.330.95300.92250.8998.800.08100.2958.420.0421
0.0425.804.06510.0456.532.95260.031.48501.06310.1682.200.18680.11109.230.0715
0.017.204.44000.0137.423.27620.0032.11741.10720.0423.800.24550.05171.630.1072
HyperpriorBaseline2.0193.203.45011.7177.962.20101.510.42560.66006.3491.400.07761.4425.250.0138
1.1388.803.69661.3077.272.27951.010.49650.68443.6087.800.08080.6652.120.0300
0.9183.603.83270.5474.822.58450.430.66990.78371.6882.400.16220.2895.030.0541
0.3729.004.29090.1262.583.01510.081.03511.08441.5057.800.13550.15124.840.0734
0.2314.604.78140.0337.113.54490.012.73021.18661.3534.800.16240.08170.010.1006
" + }, + { + "type": "table_caption", + "bbox": [ + 0.129, + 0.266, + 0.868, + 0.281 + ], + "angle": 0, + "content": "Table 3. Rate-accuracy (R-A) performance evaluations on the two baseline methods (MSE is calculated between \\(F_{o}\\) and \\(\\hat{F}_r\\))." + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.305, + 0.48, + 0.427 + ], + "angle": 0, + "content": "
TaskSplit PointOrg. RegionTrun. R. (VTM)Trun. R. (Hyperprior)
Cls\\(SP_{DS}\\)[-542.30, 94.14][-20, 20][-5, 5]
Seg\\(SP_{DS}\\)[-506.97, 105.95][-20, 20][-5, 5]
Dpt\\(SP_{DM1}\\)[-2.38, 3.27][-1, 1][-1, 1]
\\(SP_{DM2}\\)[-26.44, 5.03][-2, 2][-2, 2]
\\(SP_{DM3}\\)[-323.30, 25.05][-10, 10][-10, 10]
\\(SP_{DM4}\\)[-504.43, 100.27][-20, 20][-10, 10]
CSR\\(SP_G\\)[-71.50, 47.75][-5, 5][-5, 5]
TTI\\(SP_H\\)[-5.79, 4.46][-5.79, 4.46][-5.79, 4.46]
" + }, + { + "type": "table_caption", + "bbox": [ + 0.096, + 0.437, + 0.477, + 0.451 + ], + "angle": 0, + "content": "Table 4. Summary of the original and truncation region settings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.477, + 0.483, + 0.507 + ], + "angle": 0, + "content": "large model feature coding and provide insights for interested researchers." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.517, + 0.295, + 0.532 + ], + "angle": 0, + "content": "6.1. The baseline methods" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.539, + 0.483, + 0.693 + ], + "angle": 0, + "content": "The pipeline for the proposed baselines, illustrated in Fig. 4, includes pre-processing, core codec, and post-processing modules. First, the original feature \\( F_{o} \\) is truncated to a smaller range and then quantized into integers. Next, the quantized \\( F_{q} \\) is packed into a 2D feature \\( F_{p} \\). The codec takes pre-processed \\( F_{p} \\) as the input and outputs decoded feature \\( \\hat{F}_{p} \\). Two core codecs are used: a handcrafted codec VTM and a learning-based codec Hyperprior. In the post-processing module, \\( \\hat{F}_{p} \\) is unpacked to \\( \\hat{F}_{q} \\) and then converted back into a floating-point feature \\( \\hat{F}_{r} \\)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.701, + 0.383, + 0.717 + ], + "angle": 0, + "content": "6.1.1. Pre-processing and post-processing" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.72, + 0.483, + 0.902 + ], + "angle": 0, + "content": "The truncation operation is proposed to remove feature points that deviate significantly from the central region. The specific truncated regions for each task are listed in Table 4. Different truncations are used for the two baselines for their distinct coding strategies. The quantization operation uniformly quantizes the truncated features to 10-bit integers. Given the input requirement of VTM, we pack the features into a 2D YUV-400 format. The same packing applies in the Hyperprior baseline. \\( F_{Seg} \\), \\( F_{Dpt} \\), and \\( F_{TTI} \\) are packed into the shapes of \\( 2740 \\times 1536 \\), \\( 3222 \\times 6144 \\), and \\( 512 \\times 512 \\), respectively. Please refer to supplementary materials for the packing details. The post-processing performs unpacking" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.306, + 0.71, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.306, + 0.903, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.425, + 0.906, + 0.454 + ], + "angle": 0, + "content": "Figure 5. Rate-accuracy-drop (R-AD) comparisons among different tasks. Left: VTM baseline. Right: Hyperprior baseline." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.481, + 0.906, + 0.511 + ], + "angle": 0, + "content": "and de-quantization, which are inverse processes of packing and quantization." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.52, + 0.724, + 0.535 + ], + "angle": 0, + "content": "6.1.2. Encoding and decoding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.906, + 0.645 + ], + "angle": 0, + "content": "VTM baseline: We employ VTM-23.3 as the codec. The Intra coding is applied using the main configuration file encoder_intra_vtm.cfg. The InputChromaFormat is set to 400, and ConformanceWindowMode is enabled. The Internal-BitDepth, InputBitDepth, and OutputBitDepth are all set to 10. All other configurations are set to default. In our experiments, five QPs {22, 27, 32, 37, 42} are used." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.906, + 0.689 + ], + "angle": 0, + "content": "Hyperprior baseline: We modify the channel of the input and output of the original Hyperprior as 1 to accommodate 2D features. The loss function is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.702, + 0.905, + 0.72 + ], + "angle": 0, + "content": "\\[\nL = B P F P + \\lambda \\times \\left\\| \\left(F _ {o} - \\hat {F} _ {r}\\right) \\right\\| ^ {2} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.732, + 0.906, + 0.763 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a scaling factor used to adjust the bitrate. Please find more details in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.773, + 0.776, + 0.79 + ], + "angle": 0, + "content": "6.2. Rate-accuracy (R-A) analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.901 + ], + "angle": 0, + "content": "The R-A evaluation results are presented in Table 3. Both baselines can achieve significant bitrate savings, while the VTM baseline outperforms the Hyperprior baseline. To further assess the impact of bitrate on accuracy, we introduce a new metric, accuracy drop (AD), defined as the percentage decrease in accuracy. For Dpt, the reciprocal of RMSE is employed. The R-AD curves, shown in Fig. 5, reveal" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1074" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.106, + 0.09, + 0.472, + 0.137 + ], + "angle": 0, + "content": "
CodecClsSegDptCSRTTI
VTM Baseline0.94130.89550.85490.74900.9983
Hyperprior Baseline0.92960.92800.74370.47130.9996
" + }, + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.149, + 0.48, + 0.162 + ], + "angle": 0, + "content": "Table 5. Distortion-accuracy (D-A) analysis on the two baselines." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.189, + 0.482, + 0.295 + ], + "angle": 0, + "content": "that most tasks exhibit an inflection point where accuracy starts to decrease significantly. However, for Dpt, this point is observed only in the Hyperprior baseline. In the Hyperprior baseline, inflection points are more distracted across tasks, which may caused by the fact that different models are trained for different tasks. TTI is excluded as images generated from the original features have an FID of 0." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.304, + 0.395, + 0.319 + ], + "angle": 0, + "content": "6.3. Distortion-accuracy (D-A) analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.325, + 0.483, + 0.522 + ], + "angle": 0, + "content": "We evaluate the linear correlation between feature MSE (D) and task accuracy (A) and present coefficients of determination in Table 5. Among the baselines, only TTI exhibits high linear correlations. This is reasonable since TTI involves generating images, and the baselines employ image CODECs. In contrast, CSR fails to achieve a high linear correlation, as its features are derived from textual data. Overall, the VTM baseline outperforms the Hyperprior baseline, demonstrating its superior adaptability to large model features. Our analysis reveals that feature MSE struggles to characterize semantic distortion, particularly for textual data. This finding suggests that feature MSE is not an effective metric for measuring semantic distortion." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.531, + 0.283, + 0.547 + ], + "angle": 0, + "content": "6.4. Complexity analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.483, + 0.735 + ], + "angle": 0, + "content": "We run the VTM baseline and Hyperprior baseline on an Intel® Xeon® E5-2690 v4 CPU and a single NVIDIA GeForce RTX 4090 GPU, respectively. The VTM baseline exhibits encoding times ranging from a few seconds to several hours, while its decoding time remains within a few seconds. In contrast, the Hyperprior baseline exhibits comparable encoding and decoding times, most of which are on the order of milliseconds. Detailed results are provided in the supplementary. We note that runtime depends heavily on both hardware capabilities and implementation details. A well-optimized codec and high-performance hardware can significantly accelerate feature coding." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.317, + 0.759 + ], + "angle": 0, + "content": "6.5. Generalizability analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We assess the generalizability of trained models by compressing features extracted from one task using models trained on a different task and evaluating their R-A performance. We select three tasks and present the results in Table 6. Overall, for a given task (along column), three kinds of models behave differently in both bitrate and accuracy. For example, models trained on CSR struggle to achieve low bitrates when applied to Seg and TTI, while models trained on Seg fail to achieve low FID in TTI. More analyses can" + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.09, + 0.898, + 0.304 + ], + "angle": 0, + "content": "
TaskSegCSRTTI
ModelsBPFPmIoUBPFPAcc.BPFPFID
Trained on Seg1.7177.962.0883.800.97119.83
1.3077.271.3766.600.62208.50
0.5474.820.780.200.31333.75
0.1262.580.250.000.12363.32
0.0337.110.090.000.05354.22
Trained on TTI2.5878.302.2897.601.4425.25
1.9676.341.4677.800.6652.12
1.1273.370.6532.000.2895.03
0.5356.900.300.000.15124.84
0.2642.940.110.000.08170.01
Trained on CSR5.2978.206.3491.404.7758.19
3.2777.943.6087.802.5973.19
1.8076.771.6882.401.02196.56
1.7375.821.5057.801.02191.69
1.5475.801.3534.800.94193.08
" + }, + { + "type": "table_caption", + "bbox": [ + 0.516, + 0.316, + 0.901, + 0.33 + ], + "angle": 0, + "content": "Table 6. Generalizability evaluation on Seg, CSR, and TTI tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.357, + 0.905, + 0.418 + ], + "angle": 0, + "content": "be found in the supplementary. In practical applications, an effective compression model should be capable of handling diverse inputs. Therefore, a compression method with high generalizability is highly desirable." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.428, + 0.631, + 0.442 + ], + "angle": 0, + "content": "6.6. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.45, + 0.905, + 0.587 + ], + "angle": 0, + "content": "Large model applications require a balance of high accuracy, low complexity, and strong generalizability. However, both baselines struggle to maintain high accuracy at low bitrates, underscoring the need for more efficient feature coding strategies tailored to the unique characteristics of large model features. In addition, semantic distortion measurement remains a critical challenge in feature coding. Advancing semantic distortion metrics could further optimize codec performance and enhance overall coding efficiency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.587, + 0.905, + 0.723 + ], + "angle": 0, + "content": "The two baseline approaches present distinct advantages and drawbacks. The handcrafted approach aligns seamlessly with existing codecs but suffers from excessive runtime, limiting its real-time applicability. In contrast, the learning-based approach benefits from efficient GPU acceleration but lacks generalizability, as separate models must be trained for different features. Therefore, developing lightweight and generic codecs remains a promising research direction for feature coding." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.737, + 0.634, + 0.753 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.761, + 0.905, + 0.901 + ], + "angle": 0, + "content": "In this paper, we highlight the importance of feature coding in large model deployments and introduce a new research area: large model feature coding. As the first step, we provide a test dataset, unified test conditions, baseline methods, and a benchmark. We examine the applicability of the two baselines and identify promising directions for future research. Our aim is to encourage collaboration between image/video coding and large model communities to drive the development of feature coding." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1075" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.482, + 0.157 + ], + "angle": 0, + "content": "[1] Johannes Balle, David C. Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. ArXiv, abs/1802.01436, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.483, + 0.228 + ], + "angle": 0, + "content": "[2] Benjamin Bross, Ye-Kui Wang, Yan Ye, Shan Liu, Jianle Chen, Gary J. Sullivan, and Jens-Rainer Ohm. Overview of the versatile video coding (VVC) standard and its applications. IEEE Transactions on Circuits and Systems for Video Technology, 31(10):3736-3764, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.231, + 0.483, + 0.285 + ], + "angle": 0, + "content": "[3] Yangang Cai, Peiyin Xing, and Xuesong Gao. High efficient 3D convolution feature compression. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.288, + 0.482, + 0.342 + ], + "angle": 0, + "content": "[4] Jingxue Chen, Hang Yan, Zhiyuan Liu, Min Zhang, Hu Xiong, and Shui Yu. When federated learning meets privacy-preserving computation. ACM Computing Surveys, 56(12): 1-36, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.345, + 0.483, + 0.386 + ], + "angle": 0, + "content": "[5] Qiaoxi Chen, Changsheng Gao, and Dong Liu. End-to-end learned scalable multilayer feature compression for machine vision tasks. In ICIP, pages 1781-1787, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.388, + 0.482, + 0.443 + ], + "angle": 0, + "content": "[6] Zhuo Chen, Kui Fan, Shiqi Wang, Lingyu Duan, Weisi Lin, and Alex Chichung Kot. Toward intelligent sensing: Intermediate deep feature compression. IEEE Transactions on Image Processing, 29:2230-2243, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.445, + 0.482, + 0.487 + ], + "angle": 0, + "content": "[7] Hyomin Choi and Ivan V. Bajic. Latent-space scalability for multi-task collaborative intelligence. In ICIP, pages 3562-3566, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.489, + 0.482, + 0.531 + ], + "angle": 0, + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: a large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.532, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[9] Ronan Docherty, Antonis Vamvakeros, and Samuel J Cooper. Upsampling DINoV2 features for unsupervised vision tasks and weakly supervised materials segmentation. arXiv preprint arXiv:2410.19836, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.589, + 0.482, + 0.644 + ], + "angle": 0, + "content": "[10] Lingyu Duan, Jiaying Liu, Wenhan Yang, Tiejun Huang, and Wen Gao. Video coding for machines: A paradigm of collaborative compression and intelligent analytics. IEEE Transactions on Image Processing, 29:8680-8695, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.647, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[11] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The Llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.717, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[12] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[13] Ruoyu Feng, Xin Jin, Zongyu Guo, Runsen Feng, Yixin Gao, Tianyu He, Zhizheng Zhang, Simeng Sun, and Zhibo Chen. Image coding for machines with omnipotent feature learning. In ECCV, pages 510-528. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[14] Othmane Friha, Mohamed Amine Ferrag, Burak Kantarci, Burak Cakmak, Arda Ozgun, and Nassira Ghoualmi-Zine. LLM-based edge intelligence: A comprehensive survey on architectures, applications, security and trustworthiness." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "IEEE Open Journal of the Communications Society, 5:5799-5856, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.124, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[15] Changsheng Gao, Dong Liu, Li Li, and Feng Wu. Towards task-generic image compression: A study of semantics-oriented metrics. IEEE Transactions on Multimedia, 25:721-735, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.181, + 0.905, + 0.223 + ], + "angle": 0, + "content": "[16] Changsheng Gao, Yiheng Jiang, Li Li, Dong Liu, and Feng Wu. DMOFC: discrimination metric-optimized feature compression. In PCS, pages 1–5, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.226, + 0.905, + 0.267 + ], + "angle": 0, + "content": "[17] Changsheng Gao, Zhuoyuan Li, Li Li, Dong Liu, and Feng Wu. Rethinking the joint optimization in video coding for machines: A case study. In DCC, pages 556-556, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.269, + 0.905, + 0.338 + ], + "angle": 0, + "content": "[18] Changsheng Gao, Yiheng Jiang, Siqi Wu, Yifan Ma, Li Li, and Dong Liu. IMOFC: identity-level metric optimized feature compression for identification tasks. IEEE Transactions on Circuits and Systems for Video Technology, 35(2):1855-1869, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.341, + 0.905, + 0.397 + ], + "angle": 0, + "content": "[19] Jean Kaddour, Joshua Harris, Maximilian Mozes, Herbie Bradley, Roberta Raileanu, and Robert McHardy. Challenges and applications of large language models. arXiv preprint arXiv:2307.10169, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.399, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[20] Chia-Hao Kao, Cheng Chien, Yu-Jen Tseng, Yi-Hsin Chen, Alessandro Gnutti, Shao-Yuan Lo, Wen-Hsiao Peng, and Riccardo Leonardi. Bridging compressed image latents and multimodal large language models. arXiv preprint arXiv:2407.19651, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.471, + 0.905, + 0.54 + ], + "angle": 0, + "content": "[21] Yeongwoong Kim, Hyewon Jeong, Janghyun Yu, Younhee Kim, Jooyoung Lee, Se Yoon Jeong, and Hui Yong Kim. End-to-end learnable multi-scale feature compression for VCM. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.542, + 0.905, + 0.597 + ], + "angle": 0, + "content": "[22] Jin Han Lee, Myung-Kyu Han, Dong Wook Ko, and Il Hong Suh. From big to small: Multi-scale local planar guidance for monocular depth estimation. arXiv preprint arXiv:1907.10326, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.601, + 0.905, + 0.669 + ], + "angle": 0, + "content": "[23] Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan First, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. GShard: Scaling giant models with conditional computation and automatic sharding. arXiv preprint arXiv:2006.16668, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.672, + 0.905, + 0.728 + ], + "angle": 0, + "content": "[24] Shibao Li, Chenxu Ma, Yunwu Zhang, Longfei Li, Chengzhi Wang, Xuerong Cui, and Jianhang Liu. Attention-based variable-size feature compression module for edge inference. The Journal of Supercomputing, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.731, + 0.905, + 0.798 + ], + "angle": 0, + "content": "[25] Zhuoyuan Li, Junqi Liao, Chuanbo Tang, Haotian Zhang, Yuqi Li, Yifan Bian, Xihua Sheng, Xinmin Feng, Yao Li, Changsheng Gao, et al. Ustc-td: A test dataset and benchmark for image and video coding in 2020s. arXiv preprint arXiv:2409.08481, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.802, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[26] Zhuoyuan Li, Zikun Yuan, Li Li, Dong Liu, Xiaohu Tang, and Feng Wu. Object segmentation-assisted inter prediction for versatile video coding. IEEE Transactions on Broadcasting, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Tie Liu, Mai Xu, Shengxi Li, Chaoran Chen, Li Yang, and Zhuoyi Lv. Learn mutual feature compression for machine vision. In ICASSP, pages 1-5, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1076" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.148 + ], + "angle": 0, + "content": "[28] Guo Lu, Xingtong Ge, Tianxiong Zhong, Qiang Hu, and Jing Geng. Preprocessing enhanced image compression for machine vision. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.219 + ], + "angle": 0, + "content": "[29] Lingjuan Lyu, Han Yu, Xingjun Ma, Chen Chen, Lichao Sun, Jun Zhao, Qiang Yang, and Philip S. Yu. Privacy and robustness in federated learning: Attacks and defenses. IEEE Transactions on Neural Networks and Learning Systems, 35 (7):8726-8746, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.483, + 0.262 + ], + "angle": 0, + "content": "[30] Yifan Ma, Changsheng Gao, Qiaoxi Chen, Li Li, Dong Liu, and Xiaoyan Sun. Feature compression with 3d sparse convolution. In VCIP, pages 1-5, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[31] Rui Mao, Xinmin Feng, Changsheng Gao, Li Li, Dong Liu, and Xiaoyan Sun. Perceptual image compression with conditional diffusion transformers. In VCIP, pages 1-5, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.319, + 0.482, + 0.362 + ], + "angle": 0, + "content": "[32] Kiran Misra, Tianying Ji, Andrew Segall, and Frank Bossen. Video feature compression for machine tasks. In ICME, pages 1-6, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.363, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[33] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. DINOv2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.482, + 0.503 + ], + "angle": 0, + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.483, + 0.56 + ], + "angle": 0, + "content": "[35] Xihua Sheng, Li Li, Dong Liu, and Houqiang Li. Vnvc: A versatile neural video coding framework for efficient human-machine vision. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(7):4579-4596, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.561, + 0.483, + 0.603 + ], + "angle": 0, + "content": "[36] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-CLIP: improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.483, + 0.686 + ], + "angle": 0, + "content": "[37] Satoshi Suzuki, Shoichiro Takeda, Motohiro Takagi, Ryuichi Tanida, Hideaki Kimata, and Hayaru Shouno. Deep feature compression using spatio-temporal arrangement toward collaborative intelligent world. IEEE Transactions on Circuits and Systems for Video Technology, 32(6):3934-3946, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.688, + 0.483, + 0.744 + ], + "angle": 0, + "content": "[38] Yuanyishu Tian, Yao Wan, Lingjuan Lyu, Dezhong Yao, Hai Jin, and Lichao Sun. FedBERT: when federated learning meets pre-training. ACM Transactions on Intelligent Systems and Technology, 13(4):1-26, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.745, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[39] Yuan Tian, Guo Lu, Guangtao Zhai, and Zhiyong Gao. Nonsemantics suppressed mask learning for unsupervised video semantic compression. In ICCV, pages 13564-13576, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[40] Yuan Tian, Guo Lu, and Guangtao Zhai. SMC++: masked learning of unsupervised video semantic compression. arXiv preprint arXiv:2406.04765, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[41] Yuan Tian, Guo Lu, and Guangtao Zhai. Free-VSC: free semantics from visual foundation models for unsupervised video semantic compression. In ECCV, pages 163-183, 2025. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.906, + 0.162 + ], + "angle": 0, + "content": "[42] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.906, + 0.219 + ], + "angle": 0, + "content": "[43] Praneeth Vepakomma, Otkrist Gupta, Tristan Swedish, and Ramesh Raskar. Split learning for health: Distributed deep learning without sharing raw patient data. arXiv preprint arXiv:1812.00564, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.906, + 0.288 + ], + "angle": 0, + "content": "[44] Shurun Wang, Shiqi Wang, Wenhan Yang, Xinfeng Zhang, Shanshe Wang, Siwei Ma, and Wen Gao. Towards analysis-friendly face representation with scalable feature and texture compression. IEEE Transactions on Multimedia, 24:3169-3181, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.29, + 0.906, + 0.332 + ], + "angle": 0, + "content": "[45] Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng Chua. NExT-GPT: Any-to-any multimodal LLM. In ICML, pages 53366-53397, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.332, + 0.906, + 0.389 + ], + "angle": 0, + "content": "[46] Biwei Yan, Kun Li, Minghui Xu, Yuean Dong, Yue Zhang, Zhaochun Ren, and Xiuzhen Cheng. On protecting the data privacy of large language models (LLMs): A survey. arXiv preprint arXiv:2403.05156, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.906, + 0.445 + ], + "angle": 0, + "content": "[47] Ning Yan, Changsheng Gao, Dong Liu, Houqiang Li, Li Li, and Feng Wu. SSSIC: Semantics-to-signal scalable image coding with learned structural representations. IEEE Transactions on Image Processing, 30:8939-8954, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.446, + 0.906, + 0.516 + ], + "angle": 0, + "content": "[48] Wenhan Yang, Haofeng Huang, Yueyu Hu, Ling-Yu Duan, and Jiaying Liu. Video coding for machines: Compact visual representation compression for intelligent collaborative analytics. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-18, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.517, + 0.906, + 0.585 + ], + "angle": 0, + "content": "[49] Rui Ye, Wenhao Wang, Jingyi Chai, Dihan Li, Zexi Li, Yinda Xu, Yaxin Du, Yanfeng Wang, and Siheng Chen. OpenFedLLM: Training large language models on decentralized private data via federated learning. In ACM SIGKDD, pages 6137-6147, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.587, + 0.906, + 0.641 + ], + "angle": 0, + "content": "[50] Xu Zhang, Peiyao Guo, Ming Lu, and Zhan Ma. All-in-one image coding for joint human-machine vision with multipath aggregation. arXiv preprint arXiv:2409.19660, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.643, + 0.906, + 0.686 + ], + "angle": 0, + "content": "[51] Zhicong Zhang, Mengyang Wang, Mengyao Ma, Jiahui Li, and Xiaopeng Fan. MSFC: Deep feature compression in multi-task network. In ICME, pages 1-6, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.686, + 0.906, + 0.754 + ], + "angle": 0, + "content": "[52] Jiaying Zheng, Hainan Zhang, Lingxiang Wang, Wangjie Qiu, Hongwei Zheng, and Zhiming Zheng. Safely learning with private data: A federated learning framework for large language model. arXiv preprint arXiv:2406.14898, 2024. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.906, + 0.754 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1077" + } + ] +] \ No newline at end of file diff --git a/2025/Feature Coding in the Era of Large Models_ Dataset, Test Conditions, and Benchmark/83895aed-d59c-465f-afa3-c8ed37ae4a4b_origin.pdf b/2025/Feature Coding in the Era of Large Models_ Dataset, Test Conditions, and Benchmark/83895aed-d59c-465f-afa3-c8ed37ae4a4b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5f667f660dccbddd66742e329e46823983bfb1ea --- /dev/null +++ b/2025/Feature Coding in the Era of Large Models_ Dataset, Test Conditions, and Benchmark/83895aed-d59c-465f-afa3-c8ed37ae4a4b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:789a0e1ee6464cdc38d805f4d0c6d14ff0ae5ef19c474bb69d9836f3cb0caf69 +size 744501 diff --git a/2025/Feature Coding in the Era of Large Models_ Dataset, Test Conditions, and Benchmark/full.md b/2025/Feature Coding in the Era of Large Models_ Dataset, Test Conditions, and Benchmark/full.md new file mode 100644 index 0000000000000000000000000000000000000000..937cc43433f1af8d498c1c73e1ac4a44a04099d6 --- /dev/null +++ b/2025/Feature Coding in the Era of Large Models_ Dataset, Test Conditions, and Benchmark/full.md @@ -0,0 +1,329 @@ +# Feature Coding in the Era of Large Models: Dataset, Test Conditions, and Benchmark + +Changsheng Gao $^{1}$ , Yifan Ma $^{2}$ , Qiaoxi Chen $^{2}$ , Yenan Xu $^{2}$ , Dong Liu $^{2}$ , Weisi Lin $^{1\dagger}$ $^{1}$ Media Interactive Computing Lab, Nanyang Technological University, Singapore 639798 + $^{2}$ MOE Key Laboratory of Brain-Inspired Intelligent Perception and Cognition, University of Science and Technology of China, Hefei 230093, China +{changsheng.gao, wslin} @ntu.edu.sg; {mayf,xxii,yenanxu} @mail.ustc.edu.cn; dongeliu@ustc.edu.cn + +# Abstract + +Large models have achieved remarkable performance across various tasks, yet they incur significant computational costs and privacy concerns during both training and inference. Distributed deployment has emerged as a potential solution, but it necessitates the exchange of intermediate information between model segments, with feature representations serving as crucial information carriers. To optimize information exchange, feature coding is required to reduce transmission and storage overhead. Despite its importance, feature coding for large models remains an under-explored area. In this paper, we draw attention to large model feature coding and make three fundamental contributions. First, we introduce a comprehensive dataset encompassing diverse features generated by three representative types of large models. Second, we establish unified test conditions, enabling standardized evaluation pipelines and fair comparisons across future feature coding studies. Third, we introduce two baseline methods derived from widely used image coding techniques and benchmark their performance on the proposed dataset. These contributions aim to provide a foundation for future research and inspire broader engagement in this field. To support a long-term study, all source code and the dataset are made available at https://github.com/chansongoal/LaMoFC. + +# 1. Introduction + +Large models have revolutionized artificial intelligence (AI) with their impressive performance. However, they require substantial training data and extensive computational resources, posing significant challenges for practical de + +ployments [19]. To address these challenges, distributed deployment has been proposed as a promising solution [14, 23, 38, 43, 49, 52]. By distributing the computational workload across multiple platforms, this approach reduces the resource strain on individual systems and enables the integration of private client data into training while maintaining data security [4, 49]. As model scale continues to grow, distributed deployment is expected to become a mainstream strategy for large model deployments. + +In distributed deployments, effective information exchange between model segments is essential. During forward propagation, features generated by early segments are passed to later segments to complete the inference. Given the vast amount of data in large model applications, it is necessary to encode features to a smaller size to reduce storage and transmission burden. Analogous to image and video coding, feature coding has been proposed to address this problem [6, 16, 18, 44]. + +Existing feature coding research faces three key limitations: narrow model types, limited task diversity, and constrained data modalities. Most existing feature coding efforts focus on small-scale convolution-based networks, such as ResNet, ignoring large models. Furthermore, existing research primarily targets visual features and neglects the investigation of textual features from large language models (LLMs). This gap indicates a need for broader exploration across both models and data modalities to advance feature coding. + +In this work, we shift the feature coding research from small-scale to large-scale models. To address the existing limitations, we construct a comprehensive test dataset with features from five tasks across three large models, covering both visual and textual data for increased modality diversity. Furthermore, we establish unified test conditions to enable fair comparisons, including formulated bitrate computation and task-specific evaluation pipelines. + +To lay the groundwork for large model feature coding, + +![](images/5cfbbc208d6e297dafd9fcdd8a637b69d8c1059ff25571f063af8341e5fa09d8.jpg) +Figure 1. Three application scenarios for feature coding in large model deployments. Cloud-centralized training: Features generated by foundation models are encoded and stored in a feature database. This database supports various downstream tasks, reducing computational demands by avoiding repeated inference of large foundation models. Cloud-edge distributed training: To protect user data privacy, raw data is first processed by an edge model to generate features, which are then encoded and transmitted to the cloud. The cloud decodes these features to complete the remaining forward propagation. Cloud-edge distributed inference: A large model is partitioned across multiple parts, with portions offloaded to edge devices to distribute the computational load. The transmitted features ensure privacy protection and enable customized downstream tasks. Please refer to Sec. 3 for detailed illustrations. + +![](images/ad6fad99f7d852a9319900b8a43a2a291e26eea226a6f9e5fbde9b202841cfc9.jpg) + +![](images/5ad02696a7e2311c91c90042b81a11cd74167bdf55ef20d157a5db745d534702.jpg) + +we propose two baseline methods and evaluate their applicability to the dataset. The baselines leverage two image coding methods: the handcrafted VVC Intra coding [2] and the learning-based Hyperprior method [1]. We introduce pre-processing and post-processing modules to adapt the baselines to feature characteristics. The resulting benchmark and analyses provide valuable insights into large model feature coding and identify promising directions for further exploration. In summary, our main contributions are as follows: + +- We highlight the importance of feature coding in large model deployments, identify relevant application scenarios, and introduce a new research area: large model feature coding. +- We construct a comprehensive feature dataset encompassing three model types, five tasks, and two kinds of source data modalities, along with unified test conditions to support long-term large model feature coding research. +- We propose two baselines and evaluate their usability to the large model feature coding. We build a benchmark and suggest promising directions for further research. + +# 2. Background and motivation + +# 2.1. Feature coding + +Originally introduced within the Coding for Machines framework [10, 48], feature coding is one of two primary branches alongside visual coding. Unlike visual coding [15, 25, 26, 28, 31, 35, 39, 40, 50], which encodes and reconstructs the original visual data, feature coding involves encoding features extracted from source data into bitstreams. While visual coding has already been explored in the context of large models [20, 41], feature coding re + +mains unexplored in this domain. + +Despite recent progress, feature coding research still faces three key limitations. First, there is a restricted variety of model types in use. The existing feature coding focuses on small-scale CNN features such as [3, 17, 21, 24, 27, 30, 37]. However, as Transformer architectures now dominate large model research [11, 12, 33, 42], it is crucial to investigate feature coding for Transformer-based models. Second, the field is constrained by the limited scope of task types. Existing studies primarily address discriminative tasks such as image classification and segmentation [5, 7, 13, 32, 47, 51], while generative tasks remain largely unexplored. Given the significance of generative models, extending feature coding to include these tasks, such as image synthesis, is essential. The third limitation is the constrained range of source modality. Current feature coding research predominantly focuses on features extracted from visual data, overlooking the increasingly important features generated from textual data, particularly in light of advancements in LLMs. Expanding to multiple modalities is crucial as the diversity of AI applications grows. + +Furthermore, two challenges hinder progress in large model feature coding: the lack of a public test dataset and unified test conditions. Without a standardized dataset and evaluation conditions, fair comparisons across methods can not be conducted. A public dataset ensures consistent input data across studies, while unified test conditions establish a standard pipeline to evaluate performance. + +These limitations motivate us to construct a diverse test dataset and formulate unified test conditions to support large model feature coding research. + +
ModelParametersTaskSplit PointFeature ShapeSource Data
DINOv21.1BImage Classification (Cls)40th ViT Block257 × 1536ImageNet, 500 samples
Semantic Segmentation (Seg)40th ViT Block2 × 1370 × 1536VOC2012, 100 samples
Depth Estimation (Dpt)10th, 20th, 30th, 40th ViT Blocks2 × 4 × 1611 × 1536NYU-Depth-v2, 80 samples
Llama38BCommon Sense Reasoning (CSR)32th Decoder LayerN × 4096Arc-Challenge, 500 samples
SD38BText-to-Image Synthesis (TTI)Input Layer of VAE Decoder16 × 128 × 128COCO2017, 500 samples
+ +Table 1. Summary of the model and task selection, split point decision, and source data collection. + +# 2.2. Large model deployment + +The impressive performance of large models is driven by scaling laws [34, 36], which emphasize the importance of vast model sizes and extensive training data. For example, GPT-3 was built with 175 billion parameters and trained on 499 trillion tokens, both contributing to substantial computational demands. To alleviate these resource requirements, distributed training and inference methods have been proposed in [14, 49, 52]. + +User privacy protection presents another challenge in client services [4, 29, 46]. Split learning has been proposed as an effective solution to protect data privacy by partitioning models between clients and the cloud [49]. + +In distributed deployments, efficient information exchange between model segments is crucial. However, academic research has largely overlooked the storage and transmission costs associated with feature exchange. To address this gap, we propose to encode features into compact bitstreams to improve the efficiency of large model deployments in distributed environments. + +# 3. Feature coding in large models + +In Fig. 1, we illustrate three application scenarios of feature coding in large model deployments, highlighting their roles in optimizing the deployment efficiency of large models. + +# 3.1. Large model training + +Large models achieve their high performance through extensive training on large-scale datasets. Cloud-centralized training is commonly employed to leverage the vast computational resources and data in the cloud. However, in sensitive domains like healthcare, high-quality public data may be scarce, necessitating access to private data through cloud-edge collaborative training. + +# 3.1.1. Cloud-centralized training + +In cloud-centralized training, source databases are maintained in the cloud, where intensive computational demands become the main challenge. Complex tasks are increasingly addressed through the collaborative use of multiple models. For instance, in vision-language tasks, visual data is first processed by a vision model, and its features are then passed to an LLM for advanced reasoning. Similarly, foundation + +models often serve as core backbones for various tasks. For example, features extracted from DINOv2 are repurposed for downstream tasks like segmentation [9]. In such collaborative workflows, features generated by one model become inputs for others. However, during training, each data sample may undergo repeated inference, resulting in substantial and unnecessary computational costs. + +To address this, an effective solution is to generate features once, store them, and reuse them across tasks. The workflow is outlined on the left of Fig. 1. First, a foundation model extracts features from the source database, which are then encoded and stored in a feature database. For subsequent task-specific training, only the precomputed feature database is needed, enabling efficient access to features without redundant feature extraction. This approach significantly reduces computational load and accelerates downstream task training. However, raw feature data consumes considerable storage space, particularly with large datasets. Therefore, encoding raw features into compact bitstreams becomes essential to reduce storage demands and support efficient large model training. + +# 3.1.2. Cloud-edge distributed training + +In addition to computational resources, large model training requires access to large-scale datasets, which are often only available to a few major high-tech companies. Furthermore, sensitive data, such as medical records and financial information, cannot be freely shared due to privacy concerns and regulatory constraints. To leverage distributed, privately owned data, split learning has emerged as a viable solution [43]. Split learning divides the training process into two phases, with the bulk of the computational load handled by the cloud, while edge devices, which are often resource-constrained, handle a smaller portion of the task. This approach minimizes the computational demand on the edge while enabling collaboration with powerful cloud resources. + +The cloud-edge distributed training pipeline is depicted in the middle of Fig. 1. The whole pipeline is divided into two parts: edge models and cloud models. During forward propagation, the edge model processes private data to generate features, which are then encoded into bitstreams and transmitted to the cloud. Upon receiving the bitstreams, the cloud decodes them back into features and feeds them into + +![](images/76661b8b03c3bc3213842678c159134140ddff8c59cf6966f062cd93e1ee1105.jpg) + +![](images/5a00698ce78813fe56b4e6ea13e3c7e94f635c41218e3122552d908ea572d8e7.jpg) + +![](images/2dc1cd5df87737ee76131aef07637878118bea7855ccd3e49293906fb1d0ba34.jpg) + +![](images/c9602c75c816fb295c97bb07fcaa193d1eac5aca76d3e2975f165abaf9b58ec1.jpg) + +![](images/267ce443f187ef595e12361b25cda2283d70ebef1745a44be56020658672aa4a.jpg) + +![](images/c599d8542e9e22d57c8bbc3ea612cc9fdcf60e17615447cc8435f265a2ba0e8f.jpg) + +![](images/cd11309c31c337a747dc838b16c574bc3c2cda1f8e1a8688ef4f6998c59bd093.jpg) + +![](images/22879c822d994a8b091d6d2ebaf337098868a49b854aa8aa5cb90e64408d38a4.jpg) + +![](images/996d38b45cadd165743fb1b5413305207b37cf597af18146e705bde40f01f188.jpg) +Figure 2. Frequency and CDF comparisons between features in the proposed dataset and commonly used existing features. The proposed features exhibit distinct distributions compared to existing features, highlighting the necessity and value of the dataset. In addition, the proposed features demonstrate greater diversity, enhancing their representativeness and suitability for long-term use. Frequencies are scaled logarithmically for better visualization. + +![](images/88f5b447d96c7fb03ba0baab467268b32bae698984944d13bdfe7cf05a3882cd.jpg) + +![](images/9a3488ba0353a68041d78a8d85f4daf6e13bdafe4294e86d6efde24099ed6aea.jpg) + +![](images/19f40d0d6443436d59b6e7511b5620ff29af7c212c90ec73ff796d5e7b3d8a61.jpg) + +the cloud model. In the backward pass, gradients are computed and backpropagated along the reverse path. To reduce bandwidth cost and minimize latency, it is critical to apply feature coding before transmission. + +# 3.2. Cloud-edge distributed inference + +Once large models are trained, they are deployed in products to provide client services. For a product or service to succeed, two primary concerns must be addressed: supporting a high volume of simultaneous cloud server access and protecting user privacy. For the first concern, managing computational load is essential, particularly in generative applications that involve repetitive diffusion processes. A practical solution is to offload part of this computational demand to edge devices. For the second concern, the same strategy in the cloud-edge distributed training, as discussed in Sec. 3.1.2, can be applied. + +The cloud-edge distributed inference pipeline is illustrated on the right of Fig. 1. In the upload phase, source data is converted into features by an edge model on the client side. These features are then encoded and transmitted to the cloud, where they are further processed by large models. In the download phase, the processed features are sent back to the client to complete a specific task. Depending on the task head used, various functionalities can be performed with the returned features. For example, Stable Diffusion 3 (SD3) features derived from an image can be used to synthesize video or generate segmentation masks. This approach not + +only reduces the computational load on the cloud server but also enhances data privacy by keeping raw user data on the edge side. + +# 4. Dataset construction and analysis + +# 4.1. Dataset construction + +To guarantee the dataset's representativeness and support the long-term study of feature coding research, we curate the test dataset with consideration of three key aspects. Table 1 outlines our selected models and tasks, split points, and source data. + +# 4.1.1. Model and task selection + +Given the vast number of large models available, it is impractical to include all of them in a single research paper. Therefore, we focus on selecting representative models. To ensure the dataset's representativeness and long-term relevance, we choose one widely recognized model for each type of large model. Specifically, we select DINOv2 [33], Llama3 [11], and SD3 [12] as representatives of discriminative models, generative models, and hybrid models, respectively. The trend of adopting Transformer architectures in large models is expected to persist in the coming years. As a result, the selected models are anticipated to retain their research value and practical utility in the near future. DINOv2 and Llama3 handle visual and textual inputs, respectively, while SD3 processes textual inputs to generate visual out- + +![](images/86cbfb6ed34c740157333ae45c8978f02c3bd4a0ba9f80ea410c63f34caeba56.jpg) +Figure 3. Visualization of the original feature blocks and their corresponding DCT blocks (absolute values). In the feature domain, the proposed features exhibit higher vertical redundancy, whereas the existing features vary smoothly in both directions. In the DCT domain, the proposed features display more dispersed energy distributions. These distinct redundancy characteristics indicate different coding properties. We use $7 \times 7$ blocks for $C5$ and $C5$ -ReLU and $16 \times 16$ blocks for other features. + +puts. Together, these models provide comprehensive feature representations across visual, textual, and cross-modal applications. + +For DINOv2, we include image classification (Cls), semantic segmentation (Seg), and depth estimation (Dpt) tasks. Llama3 is applied to the common sense reasoning task (CSR), and SD3 is used for text-to-image synthesis (TTI). The selection covers three visual tasks, one textual task, and one text-to-visual task, forming a comprehensive foundation for analyzing feature coding across diverse tasks and model types. + +# 4.1.2. Split point decision + +The split points are decided to support various downstream tasks. For DINOv2, we define two types of split points: 1) $SP_{DS}$ , the output of the $40^{th}$ ViT layer, and 2) $SP_{DM}$ , which aggregates outputs from the $10^{th}$ , $20^{th}$ , $30^{th}$ , and $40^{th}$ ViT layers. $SP_{DS}$ is used for tasks utilizing single-layer features, while $SP_{DM}$ supports tasks that require multi-layer features. In this study, $SP_{DS}$ is applied to Cls and Seg, and $SP_{DM}$ to Dpt. The corresponding features are designated as $F_{Cls}$ , $F_{Seg}$ , and $F_{Dpt}$ , respectively. + +For Llama3, the split point $SP_{G}$ is set at the output of the $32^{nd}$ decoder layer. Features extracted at $SP_{G}$ can support both task-specific heads and integration with other large models [45]. In CSR, $SP_{G}$ produces $N \times 4096$ features, where $N$ is the number of tokens. + +For SD3, we take the input layer of the VAE decoder as the split point $SP_{H}$ . This split point generates features of size $16 \times 128 \times 128$ , denoted as $F_{TTI}$ , which are then passed to the VAE decoder for image synthesis. + +# 4.1.3. Source data collection + +We organize our feature test dataset into five classes, each supported by a curated subset of publicly available datasets used as source data. For Cls, we select 500 images from ImageNet [8], each representing a unique class accurately classified by DINOv2. For Seg, we collect 100 images from VOC2012, ensuring coverage of all 20 object classes. For Dpt, we source 80 images from NYU-Depth-v2, adhering to the train-test split proposed in [22] and selecting 3-8 images for each scene category. For CSR, we utilize 500 samples from the Arc-Challenge dataset, chosen for their longest input prompts and correct prediction by Llama3. For TTI, we first select 500 images from COCO2017 and then collect their longest captions. + +We noticed that learning-based methods require a dedicated training set. Please follow the source code to generate training data and find our training data in the supplementary. + +# 4.2. Feature analysis + +In this section, we compare the proposed features with those commonly used in existing research to demonstrate the necessity and importance of introducing a new dataset. + +
TaskHeadMetric
ClsNorm + Linear1Accuracy
SegNorm + Linear1mIoU
DptLinear4RMSE
CSRNorm + Linear + SoftmaxAccuracy
TTIVAE DecoderFID
+ +Table 2. Summary of task head and accuracy metric settings. Find definitions of Heads in the original papers [11, 12, 33]. + +# 4.2.1. Distribution analysis + +We first compare frequency distributions and cumulative distribution functions (CDFs) of the proposed and commonly used existing features in Fig. 2. $C5$ and $C5$ -ReLU are derived from the official ResNet50 pretrained on ImageNet, while $P$ -layer features are generated by the ResNeXT101-based MaskRCNN pre-trained on COCO2017. + +For multi-layer split points, we observe four key distinctions between $SP_{DM}$ and $P$ -layer features: (1) $SP_{DM}$ distribution range expands progressively with deeper layers; (2) $SP_{DM}$ distribution is more asymmetric; (3) $SP_{DM}$ feature values are more concentrated, as evidenced by CDFs; and (4) $SP_{DM}$ features contain more regions with sparse feature points. For single-layer split points, $SP_{GS}$ has a more concentrated distribution, while $SP_{H}$ shows more peaks. The proposed dataset omits C5-ReLU-like features, as ReLU is rarely used in large models. + +The proposed features exhibit different distributions compared to the existing features, highlighting the necessity of introducing the new dataset. Additionally, the high diversity of the proposed features enhances the dataset's representativeness and suitability for long-term research. + +# 4.2.2. Redundancy analysis + +Since coding aims to remove redundancy in inputs, we conduct a redundancy comparison on feature blocks. We visualize the original feature blocks and their DCT blocks (absolute values) in Fig. 3. In the feature domain, clear spatial redundancy is observed in both $SP_{DM}$ and $SP_{H}$ features. $SP_{H}$ exhibits redundancy in both horizontal and vertical directions, while $SP_{DM}$ shows stronger vertical redundancy. In contrast, $SP_{GS}$ features vary rapidly in both directions, resulting in weak spatial redundancy. This is because Llama3 only processes textual data, which lacks inherent spatial correlation. Existing features, on the other hand, exhibit smoother spatial variation and higher spatial redundancy. We attribute this to the translation invariance of convolutional layers, which preserve the relative positions of pixels in the feature domain, resulting in consistent redundancy patterns. In contrast, transformer-based vision models partition images into patches before transforming them, shifting spatial redundancy primarily to the vertical + +![](images/b7a862637e6c6c4a12fc13308b6262bdc21f27ef8e7c4f271d0445a4704e5cea.jpg) +Figure 4. The pipeline of the proposed baseline methods. + +direction, as seen in $SP_{DM}$ . + +In the DCT domain, spatial redundancy is reflected in the energy distribution. The proposed features display more dispersed energy, whereas existing features concentrate energy in the top-left DC component. $SP_{H}$ has the highest energy concentration as it closely resembles images. + +These distinct redundancy characteristics further demonstrate the necessity of the proposed dataset and suggest that future studies should explore feature coding methods tailored to these unique properties. + +# 5. Unified test conditions + +Even with a test dataset, fair comparisons are still challenging without identical test conditions. To address this, we establish unified test conditions here. + +# 5.1. Bitrate computation + +We introduce a new bitrate measurement, Bits Per Feature Point (BPPF), replacing the commonly used Bits Per Pixel (BPP). BPPF is calculated by dividing the total coding bits by the number of feature points. Our rationale for shifting from BPP to BPPF is twofold. First, BPP is unsuitable for features extracted from non-visual data where pixels do not exist. Second, we believe bitrate should reflect the actual encoded data (the feature itself) rather than the source data. Using BPP to calculate bitrate on source data introduces ambiguity and limits fair comparisons between coding methods. For instance, applying the same BPP to different features extracted from the same image can be misleading. In addition, BPP may incentivize downscaling source images to reduce feature size, yielding lower bitrates that do not represent true coding efficiency. In contrast, BPPF measures bitrate on the feature itself, enabling fair comparisons across coding methods as long as the same feature is used. + +# 5.2. Task accuracy evaluation + +The decoded features undergo task heads to perform accuracy evaluation. To isolate the impact of task heads, we choose simple task heads and fix their parameters during evaluation. The task heads and accuracy metrics are presented in Table 2. + +# 6. Baselines and benchmark + +As a starting point, we introduce two baselines and establish a benchmark. We aim to examine their applicability to + +
TaskImage ClassificationSemantic SegmentationDepth EstimationCommon Sense ReasoningText-to-Image Synthesis
MetricBPFPAccuracyMSEBPFPmIoUMSEBPFPRMSEMSEBPFPAccuracyMSEBPFPFIDMSE
Original3210003283.390320.369503210003200
VTMBaseline1.9499.802.94991.7680.421.86682.210.50210.61082.6999.800.01091.416.140.0066
1.0397.403.16930.8879.312.08361.270.65110.68261.831000.02600.7419.940.0184
0.2374.603.69930.2373.472.54400.330.95300.92250.8998.800.08100.2958.420.0421
0.0425.804.06510.0456.532.95260.031.48501.06310.1682.200.18680.11109.230.0715
0.017.204.44000.0137.423.27620.0032.11741.10720.0423.800.24550.05171.630.1072
HyperpriorBaseline2.0193.203.45011.7177.962.20101.510.42560.66006.3491.400.07761.4425.250.0138
1.1388.803.69661.3077.272.27951.010.49650.68443.6087.800.08080.6652.120.0300
0.9183.603.83270.5474.822.58450.430.66990.78371.6882.400.16220.2895.030.0541
0.3729.004.29090.1262.583.01510.081.03511.08441.5057.800.13550.15124.840.0734
0.2314.604.78140.0337.113.54490.012.73021.18661.3534.800.16240.08170.010.1006
+ +Table 3. Rate-accuracy (R-A) performance evaluations on the two baseline methods (MSE is calculated between $F_{o}$ and $\hat{F}_r$ ). + +
TaskSplit PointOrg. RegionTrun. R. (VTM)Trun. R. (Hyperprior)
Cls\(SP_{DS}\)[-542.30, 94.14][-20, 20][-5, 5]
Seg\(SP_{DS}\)[-506.97, 105.95][-20, 20][-5, 5]
Dpt\(SP_{DM1}\)[-2.38, 3.27][-1, 1][-1, 1]
\(SP_{DM2}\)[-26.44, 5.03][-2, 2][-2, 2]
\(SP_{DM3}\)[-323.30, 25.05][-10, 10][-10, 10]
\(SP_{DM4}\)[-504.43, 100.27][-20, 20][-10, 10]
CSR\(SP_G\)[-71.50, 47.75][-5, 5][-5, 5]
TTI\(SP_H\)[-5.79, 4.46][-5.79, 4.46][-5.79, 4.46]
+ +Table 4. Summary of the original and truncation region settings. + +large model feature coding and provide insights for interested researchers. + +# 6.1. The baseline methods + +The pipeline for the proposed baselines, illustrated in Fig. 4, includes pre-processing, core codec, and post-processing modules. First, the original feature $F_{o}$ is truncated to a smaller range and then quantized into integers. Next, the quantized $F_{q}$ is packed into a 2D feature $F_{p}$ . The codec takes pre-processed $F_{p}$ as the input and outputs decoded feature $\hat{F}_{p}$ . Two core codecs are used: a handcrafted codec VTM and a learning-based codec Hyperprior. In the post-processing module, $\hat{F}_{p}$ is unpacked to $\hat{F}_{q}$ and then converted back into a floating-point feature $\hat{F}_{r}$ . + +# 6.1.1. Pre-processing and post-processing + +The truncation operation is proposed to remove feature points that deviate significantly from the central region. The specific truncated regions for each task are listed in Table 4. Different truncations are used for the two baselines for their distinct coding strategies. The quantization operation uniformly quantizes the truncated features to 10-bit integers. Given the input requirement of VTM, we pack the features into a 2D YUV-400 format. The same packing applies in the Hyperprior baseline. $F_{Seg}$ , $F_{Dpt}$ , and $F_{TTI}$ are packed into the shapes of $2740 \times 1536$ , $3222 \times 6144$ , and $512 \times 512$ , respectively. Please refer to supplementary materials for the packing details. The post-processing performs unpacking + +![](images/08c33940578f9d96b74a21eac8f66689400098e8b2b51485485bdd528ae602a7.jpg) +Figure 5. Rate-accuracy-drop (R-AD) comparisons among different tasks. Left: VTM baseline. Right: Hyperprior baseline. + +![](images/3b74e4b041fe53f57d9d431048afbbb86a5a2eff29e5313509193bbdceac2db8.jpg) + +and de-quantization, which are inverse processes of packing and quantization. + +# 6.1.2. Encoding and decoding + +VTM baseline: We employ VTM-23.3 as the codec. The Intra coding is applied using the main configuration file encoder_intra_vtm.cfg. The InputChromaFormat is set to 400, and ConformanceWindowMode is enabled. The Internal-BitDepth, InputBitDepth, and OutputBitDepth are all set to 10. All other configurations are set to default. In our experiments, five QPs {22, 27, 32, 37, 42} are used. + +Hyperprior baseline: We modify the channel of the input and output of the original Hyperprior as 1 to accommodate 2D features. The loss function is defined as: + +$$ +L = B P F P + \lambda \times \left\| \left(F _ {o} - \hat {F} _ {r}\right) \right\| ^ {2} \tag {1} +$$ + +where $\lambda$ is a scaling factor used to adjust the bitrate. Please find more details in the supplementary. + +# 6.2. Rate-accuracy (R-A) analysis + +The R-A evaluation results are presented in Table 3. Both baselines can achieve significant bitrate savings, while the VTM baseline outperforms the Hyperprior baseline. To further assess the impact of bitrate on accuracy, we introduce a new metric, accuracy drop (AD), defined as the percentage decrease in accuracy. For Dpt, the reciprocal of RMSE is employed. The R-AD curves, shown in Fig. 5, reveal + +
CodecClsSegDptCSRTTI
VTM Baseline0.94130.89550.85490.74900.9983
Hyperprior Baseline0.92960.92800.74370.47130.9996
+ +that most tasks exhibit an inflection point where accuracy starts to decrease significantly. However, for Dpt, this point is observed only in the Hyperprior baseline. In the Hyperprior baseline, inflection points are more distracted across tasks, which may caused by the fact that different models are trained for different tasks. TTI is excluded as images generated from the original features have an FID of 0. + +# 6.3. Distortion-accuracy (D-A) analysis + +We evaluate the linear correlation between feature MSE (D) and task accuracy (A) and present coefficients of determination in Table 5. Among the baselines, only TTI exhibits high linear correlations. This is reasonable since TTI involves generating images, and the baselines employ image CODECs. In contrast, CSR fails to achieve a high linear correlation, as its features are derived from textual data. Overall, the VTM baseline outperforms the Hyperprior baseline, demonstrating its superior adaptability to large model features. Our analysis reveals that feature MSE struggles to characterize semantic distortion, particularly for textual data. This finding suggests that feature MSE is not an effective metric for measuring semantic distortion. + +# 6.4. Complexity analysis + +We run the VTM baseline and Hyperprior baseline on an Intel® Xeon® E5-2690 v4 CPU and a single NVIDIA GeForce RTX 4090 GPU, respectively. The VTM baseline exhibits encoding times ranging from a few seconds to several hours, while its decoding time remains within a few seconds. In contrast, the Hyperprior baseline exhibits comparable encoding and decoding times, most of which are on the order of milliseconds. Detailed results are provided in the supplementary. We note that runtime depends heavily on both hardware capabilities and implementation details. A well-optimized codec and high-performance hardware can significantly accelerate feature coding. + +# 6.5. Generalizability analysis + +We assess the generalizability of trained models by compressing features extracted from one task using models trained on a different task and evaluating their R-A performance. We select three tasks and present the results in Table 6. Overall, for a given task (along column), three kinds of models behave differently in both bitrate and accuracy. For example, models trained on CSR struggle to achieve low bitrates when applied to Seg and TTI, while models trained on Seg fail to achieve low FID in TTI. More analyses can + +Table 5. Distortion-accuracy (D-A) analysis on the two baselines. + +
TaskSegCSRTTI
ModelsBPFPmIoUBPFPAcc.BPFPFID
Trained on Seg1.7177.962.0883.800.97119.83
1.3077.271.3766.600.62208.50
0.5474.820.780.200.31333.75
0.1262.580.250.000.12363.32
0.0337.110.090.000.05354.22
Trained on TTI2.5878.302.2897.601.4425.25
1.9676.341.4677.800.6652.12
1.1273.370.6532.000.2895.03
0.5356.900.300.000.15124.84
0.2642.940.110.000.08170.01
Trained on CSR5.2978.206.3491.404.7758.19
3.2777.943.6087.802.5973.19
1.8076.771.6882.401.02196.56
1.7375.821.5057.801.02191.69
1.5475.801.3534.800.94193.08
+ +Table 6. Generalizability evaluation on Seg, CSR, and TTI tasks. + +be found in the supplementary. In practical applications, an effective compression model should be capable of handling diverse inputs. Therefore, a compression method with high generalizability is highly desirable. + +# 6.6. Discussion + +Large model applications require a balance of high accuracy, low complexity, and strong generalizability. However, both baselines struggle to maintain high accuracy at low bitrates, underscoring the need for more efficient feature coding strategies tailored to the unique characteristics of large model features. In addition, semantic distortion measurement remains a critical challenge in feature coding. Advancing semantic distortion metrics could further optimize codec performance and enhance overall coding efficiency. + +The two baseline approaches present distinct advantages and drawbacks. The handcrafted approach aligns seamlessly with existing codecs but suffers from excessive runtime, limiting its real-time applicability. In contrast, the learning-based approach benefits from efficient GPU acceleration but lacks generalizability, as separate models must be trained for different features. Therefore, developing lightweight and generic codecs remains a promising research direction for feature coding. + +# 7. Conclusion + +In this paper, we highlight the importance of feature coding in large model deployments and introduce a new research area: large model feature coding. As the first step, we provide a test dataset, unified test conditions, baseline methods, and a benchmark. We examine the applicability of the two baselines and identify promising directions for future research. Our aim is to encourage collaboration between image/video coding and large model communities to drive the development of feature coding. + +# References + +[1] Johannes Balle, David C. Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. ArXiv, abs/1802.01436, 2018. 2 +[2] Benjamin Bross, Ye-Kui Wang, Yan Ye, Shan Liu, Jianle Chen, Gary J. Sullivan, and Jens-Rainer Ohm. Overview of the versatile video coding (VVC) standard and its applications. IEEE Transactions on Circuits and Systems for Video Technology, 31(10):3736-3764, 2021. 2 +[3] Yangang Cai, Peiyin Xing, and Xuesong Gao. High efficient 3D convolution feature compression. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2022. 2 +[4] Jingxue Chen, Hang Yan, Zhiyuan Liu, Min Zhang, Hu Xiong, and Shui Yu. When federated learning meets privacy-preserving computation. ACM Computing Surveys, 56(12): 1-36, 2024. 1, 3 +[5] Qiaoxi Chen, Changsheng Gao, and Dong Liu. End-to-end learned scalable multilayer feature compression for machine vision tasks. In ICIP, pages 1781-1787, 2024. 2 +[6] Zhuo Chen, Kui Fan, Shiqi Wang, Lingyu Duan, Weisi Lin, and Alex Chichung Kot. Toward intelligent sensing: Intermediate deep feature compression. IEEE Transactions on Image Processing, 29:2230-2243, 2020. 1 +[7] Hyomin Choi and Ivan V. Bajic. Latent-space scalability for multi-task collaborative intelligence. In ICIP, pages 3562-3566, 2021. 2 +[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: a large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 5 +[9] Ronan Docherty, Antonis Vamvakeros, and Samuel J Cooper. Upsampling DINoV2 features for unsupervised vision tasks and weakly supervised materials segmentation. arXiv preprint arXiv:2410.19836, 2024. 3 +[10] Lingyu Duan, Jiaying Liu, Wenhan Yang, Tiejun Huang, and Wen Gao. Video coding for machines: A paradigm of collaborative compression and intelligent analytics. IEEE Transactions on Image Processing, 29:8680-8695, 2020. 2 +[11] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The Llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2, 4, 6 +[12] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 2, 4, 6 +[13] Ruoyu Feng, Xin Jin, Zongyu Guo, Runsen Feng, Yixin Gao, Tianyu He, Zhizheng Zhang, Simeng Sun, and Zhibo Chen. Image coding for machines with omnipotent feature learning. In ECCV, pages 510-528. Springer, 2022. 2 +[14] Othmane Friha, Mohamed Amine Ferrag, Burak Kantarci, Burak Cakmak, Arda Ozgun, and Nassira Ghoualmi-Zine. LLM-based edge intelligence: A comprehensive survey on architectures, applications, security and trustworthiness. + +IEEE Open Journal of the Communications Society, 5:5799-5856, 2024. 1, 3 +[15] Changsheng Gao, Dong Liu, Li Li, and Feng Wu. Towards task-generic image compression: A study of semantics-oriented metrics. IEEE Transactions on Multimedia, 25:721-735, 2023. 2 +[16] Changsheng Gao, Yiheng Jiang, Li Li, Dong Liu, and Feng Wu. DMOFC: discrimination metric-optimized feature compression. In PCS, pages 1–5, 2024. 1 +[17] Changsheng Gao, Zhuoyuan Li, Li Li, Dong Liu, and Feng Wu. Rethinking the joint optimization in video coding for machines: A case study. In DCC, pages 556-556, 2024. 2 +[18] Changsheng Gao, Yiheng Jiang, Siqi Wu, Yifan Ma, Li Li, and Dong Liu. IMOFC: identity-level metric optimized feature compression for identification tasks. IEEE Transactions on Circuits and Systems for Video Technology, 35(2):1855-1869, 2025. 1 +[19] Jean Kaddour, Joshua Harris, Maximilian Mozes, Herbie Bradley, Roberta Raileanu, and Robert McHardy. Challenges and applications of large language models. arXiv preprint arXiv:2307.10169, 2023. 1 +[20] Chia-Hao Kao, Cheng Chien, Yu-Jen Tseng, Yi-Hsin Chen, Alessandro Gnutti, Shao-Yuan Lo, Wen-Hsiao Peng, and Riccardo Leonardi. Bridging compressed image latents and multimodal large language models. arXiv preprint arXiv:2407.19651, 2024. 2 +[21] Yeongwoong Kim, Hyewon Jeong, Janghyun Yu, Younhee Kim, Jooyoung Lee, Se Yoon Jeong, and Hui Yong Kim. End-to-end learnable multi-scale feature compression for VCM. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2023. 2 +[22] Jin Han Lee, Myung-Kyu Han, Dong Wook Ko, and Il Hong Suh. From big to small: Multi-scale local planar guidance for monocular depth estimation. arXiv preprint arXiv:1907.10326, 2019. 5 +[23] Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan First, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. GShard: Scaling giant models with conditional computation and automatic sharding. arXiv preprint arXiv:2006.16668, 2020. 1 +[24] Shibao Li, Chenxu Ma, Yunwu Zhang, Longfei Li, Chengzhi Wang, Xuerong Cui, and Jianhang Liu. Attention-based variable-size feature compression module for edge inference. The Journal of Supercomputing, 2023. 2 +[25] Zhuoyuan Li, Junqi Liao, Chuanbo Tang, Haotian Zhang, Yuqi Li, Yifan Bian, Xihua Sheng, Xinmin Feng, Yao Li, Changsheng Gao, et al. Ustc-td: A test dataset and benchmark for image and video coding in 2020s. arXiv preprint arXiv:2409.08481, 2024. 2 +[26] Zhuoyuan Li, Zikun Yuan, Li Li, Dong Liu, Xiaohu Tang, and Feng Wu. Object segmentation-assisted inter prediction for versatile video coding. IEEE Transactions on Broadcasting, 2024. 2 +[27] Tie Liu, Mai Xu, Shengxi Li, Chaoran Chen, Li Yang, and Zhuoyi Lv. Learn mutual feature compression for machine vision. In ICASSP, pages 1-5, 2023. 2 + +[28] Guo Lu, Xingtong Ge, Tianxiong Zhong, Qiang Hu, and Jing Geng. Preprocessing enhanced image compression for machine vision. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2024. 2 +[29] Lingjuan Lyu, Han Yu, Xingjun Ma, Chen Chen, Lichao Sun, Jun Zhao, Qiang Yang, and Philip S. Yu. Privacy and robustness in federated learning: Attacks and defenses. IEEE Transactions on Neural Networks and Learning Systems, 35 (7):8726-8746, 2024. 3 +[30] Yifan Ma, Changsheng Gao, Qiaoxi Chen, Li Li, Dong Liu, and Xiaoyan Sun. Feature compression with 3d sparse convolution. In VCIP, pages 1-5, 2024. 2 +[31] Rui Mao, Xinmin Feng, Changsheng Gao, Li Li, Dong Liu, and Xiaoyan Sun. Perceptual image compression with conditional diffusion transformers. In VCIP, pages 1-5, 2024. 2 +[32] Kiran Misra, Tianying Ji, Andrew Segall, and Frank Bossen. Video feature compression for machine tasks. In ICME, pages 1-6, 2022. 2 +[33] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. DINOv2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 4, 6 +[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 3 +[35] Xihua Sheng, Li Li, Dong Liu, and Houqiang Li. Vnvc: A versatile neural video coding framework for efficient human-machine vision. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(7):4579-4596, 2024. 2 +[36] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-CLIP: improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 3 +[37] Satoshi Suzuki, Shoichiro Takeda, Motohiro Takagi, Ryuichi Tanida, Hideaki Kimata, and Hayaru Shouno. Deep feature compression using spatio-temporal arrangement toward collaborative intelligent world. IEEE Transactions on Circuits and Systems for Video Technology, 32(6):3934-3946, 2022. 2 +[38] Yuanyishu Tian, Yao Wan, Lingjuan Lyu, Dezhong Yao, Hai Jin, and Lichao Sun. FedBERT: when federated learning meets pre-training. ACM Transactions on Intelligent Systems and Technology, 13(4):1-26, 2022. 1 +[39] Yuan Tian, Guo Lu, Guangtao Zhai, and Zhiyong Gao. Nonsemantics suppressed mask learning for unsupervised video semantic compression. In ICCV, pages 13564-13576, 2023. 2 +[40] Yuan Tian, Guo Lu, and Guangtao Zhai. SMC++: masked learning of unsupervised video semantic compression. arXiv preprint arXiv:2406.04765, 2024. 2 +[41] Yuan Tian, Guo Lu, and Guangtao Zhai. Free-VSC: free semantics from visual foundation models for unsupervised video semantic compression. In ECCV, pages 163-183, 2025. 2 + +[42] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2 +[43] Praneeth Vepakomma, Otkrist Gupta, Tristan Swedish, and Ramesh Raskar. Split learning for health: Distributed deep learning without sharing raw patient data. arXiv preprint arXiv:1812.00564, 2018. 1, 3 +[44] Shurun Wang, Shiqi Wang, Wenhan Yang, Xinfeng Zhang, Shanshe Wang, Siwei Ma, and Wen Gao. Towards analysis-friendly face representation with scalable feature and texture compression. IEEE Transactions on Multimedia, 24:3169-3181, 2022. 1 +[45] Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng Chua. NExT-GPT: Any-to-any multimodal LLM. In ICML, pages 53366-53397, 2024. 5 +[46] Biwei Yan, Kun Li, Minghui Xu, Yuean Dong, Yue Zhang, Zhaochun Ren, and Xiuzhen Cheng. On protecting the data privacy of large language models (LLMs): A survey. arXiv preprint arXiv:2403.05156, 2024. 3 +[47] Ning Yan, Changsheng Gao, Dong Liu, Houqiang Li, Li Li, and Feng Wu. SSSIC: Semantics-to-signal scalable image coding with learned structural representations. IEEE Transactions on Image Processing, 30:8939-8954, 2021. 2 +[48] Wenhan Yang, Haofeng Huang, Yueyu Hu, Ling-Yu Duan, and Jiaying Liu. Video coding for machines: Compact visual representation compression for intelligent collaborative analytics. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-18, 2024. 2 +[49] Rui Ye, Wenhao Wang, Jingyi Chai, Dihan Li, Zexi Li, Yinda Xu, Yaxin Du, Yanfeng Wang, and Siheng Chen. OpenFedLLM: Training large language models on decentralized private data via federated learning. In ACM SIGKDD, pages 6137-6147, 2024. 1, 3 +[50] Xu Zhang, Peiyao Guo, Ming Lu, and Zhan Ma. All-in-one image coding for joint human-machine vision with multipath aggregation. arXiv preprint arXiv:2409.19660, 2024. 2 +[51] Zhicong Zhang, Mengyang Wang, Mengyao Ma, Jiahui Li, and Xiaopeng Fan. MSFC: Deep feature compression in multi-task network. In ICME, pages 1-6, 2021. 2 +[52] Jiaying Zheng, Hainan Zhang, Lingxiang Wang, Wangjie Qiu, Hongwei Zheng, and Zhiming Zheng. Safely learning with private data: A federated learning framework for large language model. arXiv preprint arXiv:2406.14898, 2024. 1, 3 \ No newline at end of file diff --git a/2025/Feature Coding in the Era of Large Models_ Dataset, Test Conditions, and Benchmark/images.zip b/2025/Feature Coding in the Era of Large Models_ Dataset, Test Conditions, and Benchmark/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5105fa770a95a25616e0a07a2297393cbdf217d9 --- /dev/null +++ b/2025/Feature Coding in the Era of Large Models_ Dataset, Test Conditions, and Benchmark/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2b5a026617cf7155aae155a3935e4bebe1fbcd6ae9389409130ac250c4837f44 +size 574866 diff --git a/2025/Feature Coding in the Era of Large Models_ Dataset, Test Conditions, and Benchmark/layout.json b/2025/Feature Coding in the Era of Large Models_ Dataset, Test Conditions, and Benchmark/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2f59ebb0b072ed8d42f60225f4e22b155e2af96a --- /dev/null +++ b/2025/Feature Coding in the Era of Large Models_ Dataset, Test Conditions, and Benchmark/layout.json @@ -0,0 +1,8615 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 169, + 103, + 442, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 103, + 442, + 138 + ], + "spans": [ + { + "bbox": [ + 169, + 103, + 442, + 138 + ], + "type": "text", + "content": "Feature Coding in the Era of Large Models: Dataset, Test Conditions, and Benchmark" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 160, + 556, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 160, + 556, + 232 + ], + "spans": [ + { + "bbox": [ + 62, + 160, + 556, + 232 + ], + "type": "text", + "content": "Changsheng Gao" + }, + { + "bbox": [ + 62, + 160, + 556, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 160, + 556, + 232 + ], + "type": "text", + "content": ", Yifan Ma" + }, + { + "bbox": [ + 62, + 160, + 556, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 62, + 160, + 556, + 232 + ], + "type": "text", + "content": ", Qiaoxi Chen" + }, + { + "bbox": [ + 62, + 160, + 556, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 62, + 160, + 556, + 232 + ], + "type": "text", + "content": ", Yenan Xu" + }, + { + "bbox": [ + 62, + 160, + 556, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 62, + 160, + 556, + 232 + ], + "type": "text", + "content": ", Dong Liu" + }, + { + "bbox": [ + 62, + 160, + 556, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 62, + 160, + 556, + 232 + ], + "type": "text", + "content": ", Weisi Lin" + }, + { + "bbox": [ + 62, + 160, + 556, + 232 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 62, + 160, + 556, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 160, + 556, + 232 + ], + "type": "text", + "content": "Media Interactive Computing Lab, Nanyang Technological University, Singapore 639798 \n" + }, + { + "bbox": [ + 62, + 160, + 556, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 62, + 160, + 556, + 232 + ], + "type": "text", + "content": "MOE Key Laboratory of Brain-Inspired Intelligent Perception and Cognition, University of Science and Technology of China, Hefei 230093, China \n{changsheng.gao, wslin} @ntu.edu.sg; {mayf,xxii,yenanxu} @mail.ustc.edu.cn; dongeliu@ustc.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 284, + 297, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 284, + 297, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 297, + 571 + ], + "type": "text", + "content": "Large models have achieved remarkable performance across various tasks, yet they incur significant computational costs and privacy concerns during both training and inference. Distributed deployment has emerged as a potential solution, but it necessitates the exchange of intermediate information between model segments, with feature representations serving as crucial information carriers. To optimize information exchange, feature coding is required to reduce transmission and storage overhead. Despite its importance, feature coding for large models remains an under-explored area. In this paper, we draw attention to large model feature coding and make three fundamental contributions. First, we introduce a comprehensive dataset encompassing diverse features generated by three representative types of large models. Second, we establish unified test conditions, enabling standardized evaluation pipelines and fair comparisons across future feature coding studies. Third, we introduce two baseline methods derived from widely used image coding techniques and benchmark their performance on the proposed dataset. These contributions aim to provide a foundation for future research and inspire broader engagement in this field. To support a long-term study, all source code and the dataset are made available at https://github.com/chansongoal/LaMoFC." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 596, + 135, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 596, + 135, + 609 + ], + "spans": [ + { + "bbox": [ + 56, + 596, + 135, + 609 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 616, + 296, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 296, + 665 + ], + "type": "text", + "content": "Large models have revolutionized artificial intelligence (AI) with their impressive performance. However, they require substantial training data and extensive computational resources, posing significant challenges for practical de" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 259, + 555, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 259, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 259, + 555, + 369 + ], + "type": "text", + "content": "ployments [19]. To address these challenges, distributed deployment has been proposed as a promising solution [14, 23, 38, 43, 49, 52]. By distributing the computational workload across multiple platforms, this approach reduces the resource strain on individual systems and enables the integration of private client data into training while maintaining data security [4, 49]. As model scale continues to grow, distributed deployment is expected to become a mainstream strategy for large model deployments." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 369, + 556, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 556, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 556, + 478 + ], + "type": "text", + "content": "In distributed deployments, effective information exchange between model segments is essential. During forward propagation, features generated by early segments are passed to later segments to complete the inference. Given the vast amount of data in large model applications, it is necessary to encode features to a smaller size to reduce storage and transmission burden. Analogous to image and video coding, feature coding has been proposed to address this problem [6, 16, 18, 44]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 480, + 556, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 556, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 556, + 601 + ], + "type": "text", + "content": "Existing feature coding research faces three key limitations: narrow model types, limited task diversity, and constrained data modalities. Most existing feature coding efforts focus on small-scale convolution-based networks, such as ResNet, ignoring large models. Furthermore, existing research primarily targets visual features and neglects the investigation of textual features from large language models (LLMs). This gap indicates a need for broader exploration across both models and data modalities to advance feature coding." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 603, + 556, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 603, + 556, + 699 + ], + "spans": [ + { + "bbox": [ + 313, + 603, + 556, + 699 + ], + "type": "text", + "content": "In this work, we shift the feature coding research from small-scale to large-scale models. To address the existing limitations, we construct a comprehensive test dataset with features from five tasks across three large models, covering both visual and textual data for increased modality diversity. Furthermore, we establish unified test conditions to enable fair comparisons, including formulated bitrate computation and task-specific evaluation pipelines." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "type": "text", + "content": "To lay the groundwork for large model feature coding," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 674, + 296, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 674, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 674, + 296, + 712 + ], + "type": "text", + "content": "This work was supported by the Ministry of Education of Singapore under Grant T2EP20123-0006. We acknowledge the support of GPU cluster built by MCC Lab of Information Science and Technology Institution, USTC." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1068" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 71, + 233, + 210 + ], + "blocks": [ + { + "bbox": [ + 63, + 71, + 233, + 210 + ], + "lines": [ + { + "bbox": [ + 63, + 71, + 233, + 210 + ], + "spans": [ + { + "bbox": [ + 63, + 71, + 233, + 210 + ], + "type": "image", + "image_path": "5cfbbc208d6e297dafd9fcdd8a637b69d8c1059ff25571f063af8341e5fa09d8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 220, + 555, + 298 + ], + "lines": [ + { + "bbox": [ + 55, + 220, + 555, + 298 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 555, + 298 + ], + "type": "text", + "content": "Figure 1. Three application scenarios for feature coding in large model deployments. Cloud-centralized training: Features generated by foundation models are encoded and stored in a feature database. This database supports various downstream tasks, reducing computational demands by avoiding repeated inference of large foundation models. Cloud-edge distributed training: To protect user data privacy, raw data is first processed by an edge model to generate features, which are then encoded and transmitted to the cloud. The cloud decodes these features to complete the remaining forward propagation. Cloud-edge distributed inference: A large model is partitioned across multiple parts, with portions offloaded to edge devices to distribute the computational load. The transmitted features ensure privacy protection and enable customized downstream tasks. Please refer to Sec. 3 for detailed illustrations." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 238, + 70, + 372, + 212 + ], + "blocks": [ + { + "bbox": [ + 238, + 70, + 372, + 212 + ], + "lines": [ + { + "bbox": [ + 238, + 70, + 372, + 212 + ], + "spans": [ + { + "bbox": [ + 238, + 70, + 372, + 212 + ], + "type": "image", + "image_path": "ad6fad99f7d852a9319900b8a43a2a291e26eea226a6f9e5fbde9b202841cfc9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 376, + 70, + 547, + 211 + ], + "blocks": [ + { + "bbox": [ + 376, + 70, + 547, + 211 + ], + "lines": [ + { + "bbox": [ + 376, + 70, + 547, + 211 + ], + "spans": [ + { + "bbox": [ + 376, + 70, + 547, + 211 + ], + "type": "image", + "image_path": "5ad02696a7e2311c91c90042b81a11cd74167bdf55ef20d157a5db745d534702.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 319, + 296, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 296, + 437 + ], + "type": "text", + "content": "we propose two baseline methods and evaluate their applicability to the dataset. The baselines leverage two image coding methods: the handcrafted VVC Intra coding [2] and the learning-based Hyperprior method [1]. We introduce pre-processing and post-processing modules to adapt the baselines to feature characteristics. The resulting benchmark and analyses provide valuable insights into large model feature coding and identify promising directions for further exploration. In summary, our main contributions are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 438, + 295, + 570 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 55, + 438, + 295, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 295, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 295, + 486 + ], + "type": "text", + "content": "- We highlight the importance of feature coding in large model deployments, identify relevant application scenarios, and introduce a new research area: large model feature coding." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 487, + 295, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 295, + 534 + ], + "type": "text", + "content": "- We construct a comprehensive feature dataset encompassing three model types, five tasks, and two kinds of source data modalities, along with unified test conditions to support long-term large model feature coding research." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 534, + 295, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 570 + ], + "type": "text", + "content": "- We propose two baselines and evaluate their usability to the large model feature coding. We build a benchmark and suggest promising directions for further research." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 581, + 214, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 214, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 214, + 594 + ], + "type": "text", + "content": "2. Background and motivation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 600, + 148, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 148, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 148, + 613 + ], + "type": "text", + "content": "2.1. Feature coding" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "Originally introduced within the Coding for Machines framework [10, 48], feature coding is one of two primary branches alongside visual coding. Unlike visual coding [15, 25, 26, 28, 31, 35, 39, 40, 50], which encodes and reconstructs the original visual data, feature coding involves encoding features extracted from source data into bitstreams. While visual coding has already been explored in the context of large models [20, 41], feature coding re" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 319, + 450, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 319, + 450, + 331 + ], + "spans": [ + { + "bbox": [ + 314, + 319, + 450, + 331 + ], + "type": "text", + "content": "mains unexplored in this domain." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 339, + 555, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 555, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 555, + 579 + ], + "type": "text", + "content": "Despite recent progress, feature coding research still faces three key limitations. First, there is a restricted variety of model types in use. The existing feature coding focuses on small-scale CNN features such as [3, 17, 21, 24, 27, 30, 37]. However, as Transformer architectures now dominate large model research [11, 12, 33, 42], it is crucial to investigate feature coding for Transformer-based models. Second, the field is constrained by the limited scope of task types. Existing studies primarily address discriminative tasks such as image classification and segmentation [5, 7, 13, 32, 47, 51], while generative tasks remain largely unexplored. Given the significance of generative models, extending feature coding to include these tasks, such as image synthesis, is essential. The third limitation is the constrained range of source modality. Current feature coding research predominantly focuses on features extracted from visual data, overlooking the increasingly important features generated from textual data, particularly in light of advancements in LLMs. Expanding to multiple modalities is crucial as the diversity of AI applications grows." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 586, + 554, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 554, + 670 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 554, + 670 + ], + "type": "text", + "content": "Furthermore, two challenges hinder progress in large model feature coding: the lack of a public test dataset and unified test conditions. Without a standardized dataset and evaluation conditions, fair comparisons across methods can not be conducted. A public dataset ensures consistent input data across studies, while unified test conditions establish a standard pipeline to evaluate performance." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "These limitations motivate us to construct a diverse test dataset and formulate unified test conditions to support large model feature coding research." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1069" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 555, + 148 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 555, + 148 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 555, + 148 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 555, + 148 + ], + "type": "table", + "html": "
ModelParametersTaskSplit PointFeature ShapeSource Data
DINOv21.1BImage Classification (Cls)40th ViT Block257 × 1536ImageNet, 500 samples
Semantic Segmentation (Seg)40th ViT Block2 × 1370 × 1536VOC2012, 100 samples
Depth Estimation (Dpt)10th, 20th, 30th, 40th ViT Blocks2 × 4 × 1611 × 1536NYU-Depth-v2, 80 samples
Llama38BCommon Sense Reasoning (CSR)32th Decoder LayerN × 4096Arc-Challenge, 500 samples
SD38BText-to-Image Synthesis (TTI)Input Layer of VAE Decoder16 × 128 × 128COCO2017, 500 samples
", + "image_path": "c73c35a121b76d32a87f31607e67d82154137c176732733ade1bdd3d58294ccf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 126, + 156, + 482, + 166 + ], + "lines": [ + { + "bbox": [ + 126, + 156, + 482, + 166 + ], + "spans": [ + { + "bbox": [ + 126, + 156, + 482, + 166 + ], + "type": "text", + "content": "Table 1. Summary of the model and task selection, split point decision, and source data collection." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 187, + 195, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 187, + 195, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 187, + 195, + 201 + ], + "type": "text", + "content": "2.2. Large model deployment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 205, + 295, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 295, + 300 + ], + "type": "text", + "content": "The impressive performance of large models is driven by scaling laws [34, 36], which emphasize the importance of vast model sizes and extensive training data. For example, GPT-3 was built with 175 billion parameters and trained on 499 trillion tokens, both contributing to substantial computational demands. To alleviate these resource requirements, distributed training and inference methods have been proposed in [14, 49, 52]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 301, + 295, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 301, + 295, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 301, + 295, + 348 + ], + "type": "text", + "content": "User privacy protection presents another challenge in client services [4, 29, 46]. Split learning has been proposed as an effective solution to protect data privacy by partitioning models between clients and the cloud [49]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 349, + 295, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 349, + 295, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 295, + 431 + ], + "type": "text", + "content": "In distributed deployments, efficient information exchange between model segments is crucial. However, academic research has largely overlooked the storage and transmission costs associated with feature exchange. To address this gap, we propose to encode features into compact bitstreams to improve the efficiency of large model deployments in distributed environments." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 443, + 228, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 443, + 228, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 228, + 456 + ], + "type": "text", + "content": "3. Feature coding in large models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 464, + 295, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 295, + 500 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 295, + 500 + ], + "type": "text", + "content": "In Fig. 1, we illustrate three application scenarios of feature coding in large model deployments, highlighting their roles in optimizing the deployment efficiency of large models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 507, + 178, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 507, + 178, + 520 + ], + "spans": [ + { + "bbox": [ + 55, + 507, + 178, + 520 + ], + "type": "text", + "content": "3.1. Large model training" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 525, + 295, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 295, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 295, + 609 + ], + "type": "text", + "content": "Large models achieve their high performance through extensive training on large-scale datasets. Cloud-centralized training is commonly employed to leverage the vast computational resources and data in the cloud. However, in sensitive domains like healthcare, high-quality public data may be scarce, necessitating access to private data through cloud-edge collaborative training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 615, + 197, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 615, + 197, + 627 + ], + "spans": [ + { + "bbox": [ + 55, + 615, + 197, + 627 + ], + "type": "text", + "content": "3.1.1. Cloud-centralized training" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": "In cloud-centralized training, source databases are maintained in the cloud, where intensive computational demands become the main challenge. Complex tasks are increasingly addressed through the collaborative use of multiple models. For instance, in vision-language tasks, visual data is first processed by a vision model, and its features are then passed to an LLM for advanced reasoning. Similarly, foundation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 188, + 553, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 188, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 313, + 188, + 553, + 271 + ], + "type": "text", + "content": "models often serve as core backbones for various tasks. For example, features extracted from DINOv2 are repurposed for downstream tasks like segmentation [9]. In such collaborative workflows, features generated by one model become inputs for others. However, during training, each data sample may undergo repeated inference, resulting in substantial and unnecessary computational costs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 272, + 555, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 272, + 555, + 439 + ], + "spans": [ + { + "bbox": [ + 313, + 272, + 555, + 439 + ], + "type": "text", + "content": "To address this, an effective solution is to generate features once, store them, and reuse them across tasks. The workflow is outlined on the left of Fig. 1. First, a foundation model extracts features from the source database, which are then encoded and stored in a feature database. For subsequent task-specific training, only the precomputed feature database is needed, enabling efficient access to features without redundant feature extraction. This approach significantly reduces computational load and accelerates downstream task training. However, raw feature data consumes considerable storage space, particularly with large datasets. Therefore, encoding raw features into compact bitstreams becomes essential to reduce storage demands and support efficient large model training." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 446, + 478, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 446, + 478, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 446, + 478, + 459 + ], + "type": "text", + "content": "3.1.2. Cloud-edge distributed training" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 462, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 554, + 628 + ], + "type": "text", + "content": "In addition to computational resources, large model training requires access to large-scale datasets, which are often only available to a few major high-tech companies. Furthermore, sensitive data, such as medical records and financial information, cannot be freely shared due to privacy concerns and regulatory constraints. To leverage distributed, privately owned data, split learning has emerged as a viable solution [43]. Split learning divides the training process into two phases, with the bulk of the computational load handled by the cloud, while edge devices, which are often resource-constrained, handle a smaller portion of the task. This approach minimizes the computational demand on the edge while enabling collaboration with powerful cloud resources." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "The cloud-edge distributed training pipeline is depicted in the middle of Fig. 1. The whole pipeline is divided into two parts: edge models and cloud models. During forward propagation, the edge model processes private data to generate features, which are then encoded into bitstreams and transmitted to the cloud. Upon receiving the bitstreams, the cloud decodes them back into features and feeds them into" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1070" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 88, + 68, + 195, + 145 + ], + "blocks": [ + { + "bbox": [ + 88, + 68, + 195, + 145 + ], + "lines": [ + { + "bbox": [ + 88, + 68, + 195, + 145 + ], + "spans": [ + { + "bbox": [ + 88, + 68, + 195, + 145 + ], + "type": "image", + "image_path": "76661b8b03c3bc3213842678c159134140ddff8c59cf6966f062cd93e1ee1105.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 197, + 68, + 304, + 146 + ], + "blocks": [ + { + "bbox": [ + 197, + 68, + 304, + 146 + ], + "lines": [ + { + "bbox": [ + 197, + 68, + 304, + 146 + ], + "spans": [ + { + "bbox": [ + 197, + 68, + 304, + 146 + ], + "type": "image", + "image_path": "5a00698ce78813fe56b4e6ea13e3c7e94f635c41218e3122552d908ea572d8e7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 305, + 68, + 412, + 146 + ], + "blocks": [ + { + "bbox": [ + 305, + 68, + 412, + 146 + ], + "lines": [ + { + "bbox": [ + 305, + 68, + 412, + 146 + ], + "spans": [ + { + "bbox": [ + 305, + 68, + 412, + 146 + ], + "type": "image", + "image_path": "2dc1cd5df87737ee76131aef07637878118bea7855ccd3e49293906fb1d0ba34.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 414, + 68, + 522, + 146 + ], + "blocks": [ + { + "bbox": [ + 414, + 68, + 522, + 146 + ], + "lines": [ + { + "bbox": [ + 414, + 68, + 522, + 146 + ], + "spans": [ + { + "bbox": [ + 414, + 68, + 522, + 146 + ], + "type": "image", + "image_path": "c9602c75c816fb295c97bb07fcaa193d1eac5aca76d3e2975f165abaf9b58ec1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 88, + 147, + 194, + 223 + ], + "blocks": [ + { + "bbox": [ + 88, + 147, + 194, + 223 + ], + "lines": [ + { + "bbox": [ + 88, + 147, + 194, + 223 + ], + "spans": [ + { + "bbox": [ + 88, + 147, + 194, + 223 + ], + "type": "image", + "image_path": "267ce443f187ef595e12361b25cda2283d70ebef1745a44be56020658672aa4a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 197, + 147, + 303, + 223 + ], + "blocks": [ + { + "bbox": [ + 197, + 147, + 303, + 223 + ], + "lines": [ + { + "bbox": [ + 197, + 147, + 303, + 223 + ], + "spans": [ + { + "bbox": [ + 197, + 147, + 303, + 223 + ], + "type": "image", + "image_path": "c599d8542e9e22d57c8bbc3ea612cc9fdcf60e17615447cc8435f265a2ba0e8f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 305, + 147, + 412, + 223 + ], + "blocks": [ + { + "bbox": [ + 305, + 147, + 412, + 223 + ], + "lines": [ + { + "bbox": [ + 305, + 147, + 412, + 223 + ], + "spans": [ + { + "bbox": [ + 305, + 147, + 412, + 223 + ], + "type": "image", + "image_path": "cd11309c31c337a747dc838b16c574bc3c2cda1f8e1a8688ef4f6998c59bd093.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 414, + 147, + 522, + 223 + ], + "blocks": [ + { + "bbox": [ + 414, + 147, + 522, + 223 + ], + "lines": [ + { + "bbox": [ + 414, + 147, + 522, + 223 + ], + "spans": [ + { + "bbox": [ + 414, + 147, + 522, + 223 + ], + "type": "image", + "image_path": "22879c822d994a8b091d6d2ebaf337098868a49b854aa8aa5cb90e64408d38a4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 88, + 224, + 194, + 300 + ], + "blocks": [ + { + "bbox": [ + 88, + 224, + 194, + 300 + ], + "lines": [ + { + "bbox": [ + 88, + 224, + 194, + 300 + ], + "spans": [ + { + "bbox": [ + 88, + 224, + 194, + 300 + ], + "type": "image", + "image_path": "996d38b45cadd165743fb1b5413305207b37cf597af18146e705bde40f01f188.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 310, + 555, + 354 + ], + "lines": [ + { + "bbox": [ + 55, + 310, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 310, + 555, + 354 + ], + "type": "text", + "content": "Figure 2. Frequency and CDF comparisons between features in the proposed dataset and commonly used existing features. The proposed features exhibit distinct distributions compared to existing features, highlighting the necessity and value of the dataset. In addition, the proposed features demonstrate greater diversity, enhancing their representativeness and suitability for long-term use. Frequencies are scaled logarithmically for better visualization." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 197, + 224, + 303, + 300 + ], + "blocks": [ + { + "bbox": [ + 197, + 224, + 303, + 300 + ], + "lines": [ + { + "bbox": [ + 197, + 224, + 303, + 300 + ], + "spans": [ + { + "bbox": [ + 197, + 224, + 303, + 300 + ], + "type": "image", + "image_path": "88f5b447d96c7fb03ba0baab467268b32bae698984944d13bdfe7cf05a3882cd.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 306, + 224, + 412, + 300 + ], + "blocks": [ + { + "bbox": [ + 306, + 224, + 412, + 300 + ], + "lines": [ + { + "bbox": [ + 306, + 224, + 412, + 300 + ], + "spans": [ + { + "bbox": [ + 306, + 224, + 412, + 300 + ], + "type": "image", + "image_path": "9a3488ba0353a68041d78a8d85f4daf6e13bdafe4294e86d6efde24099ed6aea.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 414, + 224, + 522, + 300 + ], + "blocks": [ + { + "bbox": [ + 414, + 224, + 522, + 300 + ], + "lines": [ + { + "bbox": [ + 414, + 224, + 522, + 300 + ], + "spans": [ + { + "bbox": [ + 414, + 224, + 522, + 300 + ], + "type": "image", + "image_path": "19f40d0d6443436d59b6e7511b5620ff29af7c212c90ec73ff796d5e7b3d8a61.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 375, + 295, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 375, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 375, + 295, + 422 + ], + "type": "text", + "content": "the cloud model. In the backward pass, gradients are computed and backpropagated along the reverse path. To reduce bandwidth cost and minimize latency, it is critical to apply feature coding before transmission." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 432, + 231, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 432, + 231, + 444 + ], + "spans": [ + { + "bbox": [ + 55, + 432, + 231, + 444 + ], + "type": "text", + "content": "3.2. Cloud-edge distributed inference" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 449, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 581 + ], + "type": "text", + "content": "Once large models are trained, they are deployed in products to provide client services. For a product or service to succeed, two primary concerns must be addressed: supporting a high volume of simultaneous cloud server access and protecting user privacy. For the first concern, managing computational load is essential, particularly in generative applications that involve repetitive diffusion processes. A practical solution is to offload part of this computational demand to edge devices. For the second concern, the same strategy in the cloud-edge distributed training, as discussed in Sec. 3.1.2, can be applied." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": "The cloud-edge distributed inference pipeline is illustrated on the right of Fig. 1. In the upload phase, source data is converted into features by an edge model on the client side. These features are then encoded and transmitted to the cloud, where they are further processed by large models. In the download phase, the processed features are sent back to the client to complete a specific task. Depending on the task head used, various functionalities can be performed with the returned features. For example, Stable Diffusion 3 (SD3) features derived from an image can be used to synthesize video or generate segmentation masks. This approach not" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 375, + 555, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 375, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 555, + 411 + ], + "type": "text", + "content": "only reduces the computational load on the cloud server but also enhances data privacy by keeping raw user data on the edge side." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 424, + 500, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 424, + 500, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 424, + 500, + 437 + ], + "type": "text", + "content": "4. Dataset construction and analysis" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 445, + 433, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 433, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 433, + 455 + ], + "type": "text", + "content": "4.1. Dataset construction" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 462, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 554, + 521 + ], + "type": "text", + "content": "To guarantee the dataset's representativeness and support the long-term study of feature coding research, we curate the test dataset with consideration of three key aspects. Table 1 outlines our selected models and tasks, split points, and source data." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 530, + 448, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 530, + 448, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 530, + 448, + 541 + ], + "type": "text", + "content": "4.1.1. Model and task selection" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": "Given the vast number of large models available, it is impractical to include all of them in a single research paper. Therefore, we focus on selecting representative models. To ensure the dataset's representativeness and long-term relevance, we choose one widely recognized model for each type of large model. Specifically, we select DINOv2 [33], Llama3 [11], and SD3 [12] as representatives of discriminative models, generative models, and hybrid models, respectively. The trend of adopting Transformer architectures in large models is expected to persist in the coming years. As a result, the selected models are anticipated to retain their research value and practical utility in the near future. DINOv2 and Llama3 handle visual and textual inputs, respectively, while SD3 processes textual inputs to generate visual out-" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1071" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 97, + 63, + 512, + 318 + ], + "blocks": [ + { + "bbox": [ + 97, + 63, + 512, + 318 + ], + "lines": [ + { + "bbox": [ + 97, + 63, + 512, + 318 + ], + "spans": [ + { + "bbox": [ + 97, + 63, + 512, + 318 + ], + "type": "image", + "image_path": "86cbfb6ed34c740157333ae45c8978f02c3bd4a0ba9f80ea410c63f34caeba56.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 326, + 555, + 371 + ], + "lines": [ + { + "bbox": [ + 54, + 326, + 555, + 371 + ], + "spans": [ + { + "bbox": [ + 54, + 326, + 555, + 371 + ], + "type": "text", + "content": "Figure 3. Visualization of the original feature blocks and their corresponding DCT blocks (absolute values). In the feature domain, the proposed features exhibit higher vertical redundancy, whereas the existing features vary smoothly in both directions. In the DCT domain, the proposed features display more dispersed energy distributions. These distinct redundancy characteristics indicate different coding properties. We use " + }, + { + "bbox": [ + 54, + 326, + 555, + 371 + ], + "type": "inline_equation", + "content": "7 \\times 7" + }, + { + "bbox": [ + 54, + 326, + 555, + 371 + ], + "type": "text", + "content": " blocks for " + }, + { + "bbox": [ + 54, + 326, + 555, + 371 + ], + "type": "inline_equation", + "content": "C5" + }, + { + "bbox": [ + 54, + 326, + 555, + 371 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 326, + 555, + 371 + ], + "type": "inline_equation", + "content": "C5" + }, + { + "bbox": [ + 54, + 326, + 555, + 371 + ], + "type": "text", + "content": "-ReLU and " + }, + { + "bbox": [ + 54, + 326, + 555, + 371 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 54, + 326, + 555, + 371 + ], + "type": "text", + "content": " blocks for other features." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 392, + 294, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 392, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 54, + 392, + 294, + 426 + ], + "type": "text", + "content": "puts. Together, these models provide comprehensive feature representations across visual, textual, and cross-modal applications." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 427, + 295, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 427, + 295, + 523 + ], + "spans": [ + { + "bbox": [ + 54, + 427, + 295, + 523 + ], + "type": "text", + "content": "For DINOv2, we include image classification (Cls), semantic segmentation (Seg), and depth estimation (Dpt) tasks. Llama3 is applied to the common sense reasoning task (CSR), and SD3 is used for text-to-image synthesis (TTI). The selection covers three visual tasks, one textual task, and one text-to-visual task, forming a comprehensive foundation for analyzing feature coding across diverse tasks and model types." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 530, + 167, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 530, + 167, + 542 + ], + "spans": [ + { + "bbox": [ + 55, + 530, + 167, + 542 + ], + "type": "text", + "content": "4.1.2. Split point decision" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "type": "text", + "content": "The split points are decided to support various downstream tasks. For DINOv2, we define two types of split points: 1) " + }, + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "type": "inline_equation", + "content": "SP_{DS}" + }, + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "type": "text", + "content": ", the output of the " + }, + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "type": "inline_equation", + "content": "40^{th}" + }, + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "type": "text", + "content": " ViT layer, and 2) " + }, + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "type": "inline_equation", + "content": "SP_{DM}" + }, + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "type": "text", + "content": ", which aggregates outputs from the " + }, + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "type": "inline_equation", + "content": "10^{th}" + }, + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "type": "inline_equation", + "content": "20^{th}" + }, + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "type": "inline_equation", + "content": "30^{th}" + }, + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "type": "inline_equation", + "content": "40^{th}" + }, + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "type": "text", + "content": " ViT layers. " + }, + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "type": "inline_equation", + "content": "SP_{DS}" + }, + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "type": "text", + "content": " is used for tasks utilizing single-layer features, while " + }, + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "type": "inline_equation", + "content": "SP_{DM}" + }, + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "type": "text", + "content": " supports tasks that require multi-layer features. In this study, " + }, + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "type": "inline_equation", + "content": "SP_{DS}" + }, + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "type": "text", + "content": " is applied to Cls and Seg, and " + }, + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "type": "inline_equation", + "content": "SP_{DM}" + }, + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "type": "text", + "content": " to Dpt. The corresponding features are designated as " + }, + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "type": "inline_equation", + "content": "F_{Cls}" + }, + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "type": "inline_equation", + "content": "F_{Seg}" + }, + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "type": "inline_equation", + "content": "F_{Dpt}" + }, + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "text", + "content": "For Llama3, the split point " + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "inline_equation", + "content": "SP_{G}" + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "text", + "content": " is set at the output of the " + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "inline_equation", + "content": "32^{nd}" + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "text", + "content": " decoder layer. Features extracted at " + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "inline_equation", + "content": "SP_{G}" + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "text", + "content": " can support both task-specific heads and integration with other large models [45]. In CSR, " + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "inline_equation", + "content": "SP_{G}" + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "text", + "content": " produces " + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "inline_equation", + "content": "N \\times 4096" + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "text", + "content": " features, where " + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "text", + "content": " is the number of tokens." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 392, + 555, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 555, + 440 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 555, + 440 + ], + "type": "text", + "content": "For SD3, we take the input layer of the VAE decoder as the split point " + }, + { + "bbox": [ + 313, + 392, + 555, + 440 + ], + "type": "inline_equation", + "content": "SP_{H}" + }, + { + "bbox": [ + 313, + 392, + 555, + 440 + ], + "type": "text", + "content": ". This split point generates features of size " + }, + { + "bbox": [ + 313, + 392, + 555, + 440 + ], + "type": "inline_equation", + "content": "16 \\times 128 \\times 128" + }, + { + "bbox": [ + 313, + 392, + 555, + 440 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 313, + 392, + 555, + 440 + ], + "type": "inline_equation", + "content": "F_{TTI}" + }, + { + "bbox": [ + 313, + 392, + 555, + 440 + ], + "type": "text", + "content": ", which are then passed to the VAE decoder for image synthesis." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 445, + 437, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 437, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 437, + 456 + ], + "type": "text", + "content": "4.1.3. Source data collection" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 460, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 616 + ], + "type": "text", + "content": "We organize our feature test dataset into five classes, each supported by a curated subset of publicly available datasets used as source data. For Cls, we select 500 images from ImageNet [8], each representing a unique class accurately classified by DINOv2. For Seg, we collect 100 images from VOC2012, ensuring coverage of all 20 object classes. For Dpt, we source 80 images from NYU-Depth-v2, adhering to the train-test split proposed in [22] and selecting 3-8 images for each scene category. For CSR, we utilize 500 samples from the Arc-Challenge dataset, chosen for their longest input prompts and correct prediction by Llama3. For TTI, we first select 500 images from COCO2017 and then collect their longest captions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 616, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 555, + 653 + ], + "type": "text", + "content": "We noticed that learning-based methods require a dedicated training set. Please follow the source code to generate training data and find our training data in the supplementary." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 659, + 413, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 413, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 413, + 672 + ], + "type": "text", + "content": "4.2. Feature analysis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "In this section, we compare the proposed features with those commonly used in existing research to demonstrate the necessity and importance of introducing a new dataset." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1072" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 99, + 70, + 255, + 149 + ], + "blocks": [ + { + "bbox": [ + 99, + 70, + 255, + 149 + ], + "lines": [ + { + "bbox": [ + 99, + 70, + 255, + 149 + ], + "spans": [ + { + "bbox": [ + 99, + 70, + 255, + 149 + ], + "type": "table", + "html": "
TaskHeadMetric
ClsNorm + Linear1Accuracy
SegNorm + Linear1mIoU
DptLinear4RMSE
CSRNorm + Linear + SoftmaxAccuracy
TTIVAE DecoderFID
", + "image_path": "3e0941e7833f46f4b008ce20586747f1e4b6566992414853eb3b98d8545d947d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 157, + 295, + 180 + ], + "lines": [ + { + "bbox": [ + 55, + 157, + 295, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 157, + 295, + 180 + ], + "type": "text", + "content": "Table 2. Summary of task head and accuracy metric settings. Find definitions of Heads in the original papers [11, 12, 33]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 200, + 173, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 200, + 173, + 211 + ], + "spans": [ + { + "bbox": [ + 55, + 200, + 173, + 211 + ], + "type": "text", + "content": "4.2.1. Distribution analysis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 215, + 295, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 215, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 215, + 295, + 297 + ], + "type": "text", + "content": "We first compare frequency distributions and cumulative distribution functions (CDFs) of the proposed and commonly used existing features in Fig. 2. " + }, + { + "bbox": [ + 55, + 215, + 295, + 297 + ], + "type": "inline_equation", + "content": "C5" + }, + { + "bbox": [ + 55, + 215, + 295, + 297 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 215, + 295, + 297 + ], + "type": "inline_equation", + "content": "C5" + }, + { + "bbox": [ + 55, + 215, + 295, + 297 + ], + "type": "text", + "content": "-ReLU are derived from the official ResNet50 pretrained on ImageNet, while " + }, + { + "bbox": [ + 55, + 215, + 295, + 297 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 215, + 295, + 297 + ], + "type": "text", + "content": "-layer features are generated by the ResNeXT101-based MaskRCNN pre-trained on COCO2017." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 299, + 295, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 299, + 295, + 417 + ], + "spans": [ + { + "bbox": [ + 55, + 299, + 295, + 417 + ], + "type": "text", + "content": "For multi-layer split points, we observe four key distinctions between " + }, + { + "bbox": [ + 55, + 299, + 295, + 417 + ], + "type": "inline_equation", + "content": "SP_{DM}" + }, + { + "bbox": [ + 55, + 299, + 295, + 417 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 299, + 295, + 417 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 299, + 295, + 417 + ], + "type": "text", + "content": "-layer features: (1) " + }, + { + "bbox": [ + 55, + 299, + 295, + 417 + ], + "type": "inline_equation", + "content": "SP_{DM}" + }, + { + "bbox": [ + 55, + 299, + 295, + 417 + ], + "type": "text", + "content": " distribution range expands progressively with deeper layers; (2) " + }, + { + "bbox": [ + 55, + 299, + 295, + 417 + ], + "type": "inline_equation", + "content": "SP_{DM}" + }, + { + "bbox": [ + 55, + 299, + 295, + 417 + ], + "type": "text", + "content": " distribution is more asymmetric; (3) " + }, + { + "bbox": [ + 55, + 299, + 295, + 417 + ], + "type": "inline_equation", + "content": "SP_{DM}" + }, + { + "bbox": [ + 55, + 299, + 295, + 417 + ], + "type": "text", + "content": " feature values are more concentrated, as evidenced by CDFs; and (4) " + }, + { + "bbox": [ + 55, + 299, + 295, + 417 + ], + "type": "inline_equation", + "content": "SP_{DM}" + }, + { + "bbox": [ + 55, + 299, + 295, + 417 + ], + "type": "text", + "content": " features contain more regions with sparse feature points. For single-layer split points, " + }, + { + "bbox": [ + 55, + 299, + 295, + 417 + ], + "type": "inline_equation", + "content": "SP_{GS}" + }, + { + "bbox": [ + 55, + 299, + 295, + 417 + ], + "type": "text", + "content": " has a more concentrated distribution, while " + }, + { + "bbox": [ + 55, + 299, + 295, + 417 + ], + "type": "inline_equation", + "content": "SP_{H}" + }, + { + "bbox": [ + 55, + 299, + 295, + 417 + ], + "type": "text", + "content": " shows more peaks. The proposed dataset omits C5-ReLU-like features, as ReLU is rarely used in large models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 418, + 295, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 295, + 478 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 295, + 478 + ], + "type": "text", + "content": "The proposed features exhibit different distributions compared to the existing features, highlighting the necessity of introducing the new dataset. Additionally, the high diversity of the proposed features enhances the dataset's representativeness and suitability for long-term research." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 483, + 174, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 174, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 174, + 495 + ], + "type": "text", + "content": "4.2.2. Redundancy analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "text", + "content": "Since coding aims to remove redundancy in inputs, we conduct a redundancy comparison on feature blocks. We visualize the original feature blocks and their DCT blocks (absolute values) in Fig. 3. In the feature domain, clear spatial redundancy is observed in both " + }, + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "inline_equation", + "content": "SP_{DM}" + }, + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "inline_equation", + "content": "SP_{H}" + }, + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "text", + "content": " features. " + }, + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "inline_equation", + "content": "SP_{H}" + }, + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "text", + "content": " exhibits redundancy in both horizontal and vertical directions, while " + }, + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "inline_equation", + "content": "SP_{DM}" + }, + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "text", + "content": " shows stronger vertical redundancy. In contrast, " + }, + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "inline_equation", + "content": "SP_{GS}" + }, + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "text", + "content": " features vary rapidly in both directions, resulting in weak spatial redundancy. This is because Llama3 only processes textual data, which lacks inherent spatial correlation. Existing features, on the other hand, exhibit smoother spatial variation and higher spatial redundancy. We attribute this to the translation invariance of convolutional layers, which preserve the relative positions of pixels in the feature domain, resulting in consistent redundancy patterns. In contrast, transformer-based vision models partition images into patches before transforming them, shifting spatial redundancy primarily to the vertical" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 69, + 553, + 125 + ], + "blocks": [ + { + "bbox": [ + 317, + 69, + 553, + 125 + ], + "lines": [ + { + "bbox": [ + 317, + 69, + 553, + 125 + ], + "spans": [ + { + "bbox": [ + 317, + 69, + 553, + 125 + ], + "type": "image", + "image_path": "b7a862637e6c6c4a12fc13308b6262bdc21f27ef8e7c4f271d0445a4704e5cea.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 133, + 538, + 144 + ], + "lines": [ + { + "bbox": [ + 329, + 133, + 538, + 144 + ], + "spans": [ + { + "bbox": [ + 329, + 133, + 538, + 144 + ], + "type": "text", + "content": "Figure 4. The pipeline of the proposed baseline methods." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 163, + 430, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 163, + 430, + 175 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 430, + 175 + ], + "type": "text", + "content": "direction, as seen in " + }, + { + "bbox": [ + 313, + 163, + 430, + 175 + ], + "type": "inline_equation", + "content": "SP_{DM}" + }, + { + "bbox": [ + 313, + 163, + 430, + 175 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 175, + 553, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 175, + 553, + 235 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 553, + 235 + ], + "type": "text", + "content": "In the DCT domain, spatial redundancy is reflected in the energy distribution. The proposed features display more dispersed energy, whereas existing features concentrate energy in the top-left DC component. " + }, + { + "bbox": [ + 313, + 175, + 553, + 235 + ], + "type": "inline_equation", + "content": "SP_{H}" + }, + { + "bbox": [ + 313, + 175, + 553, + 235 + ], + "type": "text", + "content": " has the highest energy concentration as it closely resembles images." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 236, + 553, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 553, + 283 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 553, + 283 + ], + "type": "text", + "content": "These distinct redundancy characteristics further demonstrate the necessity of the proposed dataset and suggest that future studies should explore feature coding methods tailored to these unique properties." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 293, + 443, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 443, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 443, + 304 + ], + "type": "text", + "content": "5. Unified test conditions" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 312, + 553, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 553, + 348 + ], + "type": "text", + "content": "Even with a test dataset, fair comparisons are still challenging without identical test conditions. To address this, we establish unified test conditions here." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 355, + 432, + 368 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 355, + 432, + 368 + ], + "spans": [ + { + "bbox": [ + 313, + 355, + 432, + 368 + ], + "type": "text", + "content": "5.1. Bitrate computation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 373, + 555, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 555, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 555, + 576 + ], + "type": "text", + "content": "We introduce a new bitrate measurement, Bits Per Feature Point (BPPF), replacing the commonly used Bits Per Pixel (BPP). BPPF is calculated by dividing the total coding bits by the number of feature points. Our rationale for shifting from BPP to BPPF is twofold. First, BPP is unsuitable for features extracted from non-visual data where pixels do not exist. Second, we believe bitrate should reflect the actual encoded data (the feature itself) rather than the source data. Using BPP to calculate bitrate on source data introduces ambiguity and limits fair comparisons between coding methods. For instance, applying the same BPP to different features extracted from the same image can be misleading. In addition, BPP may incentivize downscaling source images to reduce feature size, yielding lower bitrates that do not represent true coding efficiency. In contrast, BPPF measures bitrate on the feature itself, enabling fair comparisons across coding methods as long as the same feature is used." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 582, + 455, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 455, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 455, + 594 + ], + "type": "text", + "content": "5.2. Task accuracy evaluation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 600, + 553, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 553, + 658 + ], + "type": "text", + "content": "The decoded features undergo task heads to perform accuracy evaluation. To isolate the impact of task heads, we choose simple task heads and fix their parameters during evaluation. The task heads and accuracy metrics are presented in Table 2." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 669, + 460, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 669, + 460, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 669, + 460, + 681 + ], + "type": "text", + "content": "6. Baselines and benchmark" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "As a starting point, we introduce two baselines and establish a benchmark. We aim to examine their applicability to" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1073" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 70, + 542, + 203 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 542, + 203 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 542, + 203 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 542, + 203 + ], + "type": "table", + "html": "
TaskImage ClassificationSemantic SegmentationDepth EstimationCommon Sense ReasoningText-to-Image Synthesis
MetricBPFPAccuracyMSEBPFPmIoUMSEBPFPRMSEMSEBPFPAccuracyMSEBPFPFIDMSE
Original3210003283.390320.369503210003200
VTMBaseline1.9499.802.94991.7680.421.86682.210.50210.61082.6999.800.01091.416.140.0066
1.0397.403.16930.8879.312.08361.270.65110.68261.831000.02600.7419.940.0184
0.2374.603.69930.2373.472.54400.330.95300.92250.8998.800.08100.2958.420.0421
0.0425.804.06510.0456.532.95260.031.48501.06310.1682.200.18680.11109.230.0715
0.017.204.44000.0137.423.27620.0032.11741.10720.0423.800.24550.05171.630.1072
HyperpriorBaseline2.0193.203.45011.7177.962.20101.510.42560.66006.3491.400.07761.4425.250.0138
1.1388.803.69661.3077.272.27951.010.49650.68443.6087.800.08080.6652.120.0300
0.9183.603.83270.5474.822.58450.430.66990.78371.6882.400.16220.2895.030.0541
0.3729.004.29090.1262.583.01510.081.03511.08441.5057.800.13550.15124.840.0734
0.2314.604.78140.0337.113.54490.012.73021.18661.3534.800.16240.08170.010.1006
", + "image_path": "be80f77614e36e8da69a0f16d4691b14e9cbeb898a7255556dd970635c07e713.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 241, + 293, + 338 + ], + "blocks": [ + { + "bbox": [ + 78, + 210, + 531, + 222 + ], + "lines": [ + { + "bbox": [ + 78, + 210, + 531, + 222 + ], + "spans": [ + { + "bbox": [ + 78, + 210, + 531, + 222 + ], + "type": "text", + "content": "Table 3. Rate-accuracy (R-A) performance evaluations on the two baseline methods (MSE is calculated between " + }, + { + "bbox": [ + 78, + 210, + 531, + 222 + ], + "type": "inline_equation", + "content": "F_{o}" + }, + { + "bbox": [ + 78, + 210, + 531, + 222 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 78, + 210, + 531, + 222 + ], + "type": "inline_equation", + "content": "\\hat{F}_r" + }, + { + "bbox": [ + 78, + 210, + 531, + 222 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 241, + 293, + 338 + ], + "lines": [ + { + "bbox": [ + 59, + 241, + 293, + 338 + ], + "spans": [ + { + "bbox": [ + 59, + 241, + 293, + 338 + ], + "type": "table", + "html": "
TaskSplit PointOrg. RegionTrun. R. (VTM)Trun. R. (Hyperprior)
Cls\\(SP_{DS}\\)[-542.30, 94.14][-20, 20][-5, 5]
Seg\\(SP_{DS}\\)[-506.97, 105.95][-20, 20][-5, 5]
Dpt\\(SP_{DM1}\\)[-2.38, 3.27][-1, 1][-1, 1]
\\(SP_{DM2}\\)[-26.44, 5.03][-2, 2][-2, 2]
\\(SP_{DM3}\\)[-323.30, 25.05][-10, 10][-10, 10]
\\(SP_{DM4}\\)[-504.43, 100.27][-20, 20][-10, 10]
CSR\\(SP_G\\)[-71.50, 47.75][-5, 5][-5, 5]
TTI\\(SP_H\\)[-5.79, 4.46][-5.79, 4.46][-5.79, 4.46]
", + "image_path": "2935ad05a460ea7b0bbd9b2cec70c95e794f4c86eb390c7d9dd462eb9ee036c5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 346, + 291, + 357 + ], + "lines": [ + { + "bbox": [ + 58, + 346, + 291, + 357 + ], + "spans": [ + { + "bbox": [ + 58, + 346, + 291, + 357 + ], + "type": "text", + "content": "Table 4. Summary of the original and truncation region settings." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 377, + 295, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 377, + 295, + 401 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 295, + 401 + ], + "type": "text", + "content": "large model feature coding and provide insights for interested researchers." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 409, + 180, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 409, + 180, + 421 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 180, + 421 + ], + "type": "text", + "content": "6.1. The baseline methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 426, + 295, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 426, + 295, + 548 + ], + "spans": [ + { + "bbox": [ + 54, + 426, + 295, + 548 + ], + "type": "text", + "content": "The pipeline for the proposed baselines, illustrated in Fig. 4, includes pre-processing, core codec, and post-processing modules. First, the original feature " + }, + { + "bbox": [ + 54, + 426, + 295, + 548 + ], + "type": "inline_equation", + "content": "F_{o}" + }, + { + "bbox": [ + 54, + 426, + 295, + 548 + ], + "type": "text", + "content": " is truncated to a smaller range and then quantized into integers. Next, the quantized " + }, + { + "bbox": [ + 54, + 426, + 295, + 548 + ], + "type": "inline_equation", + "content": "F_{q}" + }, + { + "bbox": [ + 54, + 426, + 295, + 548 + ], + "type": "text", + "content": " is packed into a 2D feature " + }, + { + "bbox": [ + 54, + 426, + 295, + 548 + ], + "type": "inline_equation", + "content": "F_{p}" + }, + { + "bbox": [ + 54, + 426, + 295, + 548 + ], + "type": "text", + "content": ". The codec takes pre-processed " + }, + { + "bbox": [ + 54, + 426, + 295, + 548 + ], + "type": "inline_equation", + "content": "F_{p}" + }, + { + "bbox": [ + 54, + 426, + 295, + 548 + ], + "type": "text", + "content": " as the input and outputs decoded feature " + }, + { + "bbox": [ + 54, + 426, + 295, + 548 + ], + "type": "inline_equation", + "content": "\\hat{F}_{p}" + }, + { + "bbox": [ + 54, + 426, + 295, + 548 + ], + "type": "text", + "content": ". Two core codecs are used: a handcrafted codec VTM and a learning-based codec Hyperprior. In the post-processing module, " + }, + { + "bbox": [ + 54, + 426, + 295, + 548 + ], + "type": "inline_equation", + "content": "\\hat{F}_{p}" + }, + { + "bbox": [ + 54, + 426, + 295, + 548 + ], + "type": "text", + "content": " is unpacked to " + }, + { + "bbox": [ + 54, + 426, + 295, + 548 + ], + "type": "inline_equation", + "content": "\\hat{F}_{q}" + }, + { + "bbox": [ + 54, + 426, + 295, + 548 + ], + "type": "text", + "content": " and then converted back into a floating-point feature " + }, + { + "bbox": [ + 54, + 426, + 295, + 548 + ], + "type": "inline_equation", + "content": "\\hat{F}_{r}" + }, + { + "bbox": [ + 54, + 426, + 295, + 548 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 555, + 234, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 234, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 234, + 567 + ], + "type": "text", + "content": "6.1.1. Pre-processing and post-processing" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "text", + "content": "The truncation operation is proposed to remove feature points that deviate significantly from the central region. The specific truncated regions for each task are listed in Table 4. Different truncations are used for the two baselines for their distinct coding strategies. The quantization operation uniformly quantizes the truncated features to 10-bit integers. Given the input requirement of VTM, we pack the features into a 2D YUV-400 format. The same packing applies in the Hyperprior baseline. " + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "F_{Seg}" + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "F_{Dpt}" + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "F_{TTI}" + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "text", + "content": " are packed into the shapes of " + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "2740 \\times 1536" + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "3222 \\times 6144" + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "text", + "content": ", respectively. Please refer to supplementary materials for the packing details. The post-processing performs unpacking" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 242, + 434, + 326 + ], + "blocks": [ + { + "bbox": [ + 317, + 242, + 434, + 326 + ], + "lines": [ + { + "bbox": [ + 317, + 242, + 434, + 326 + ], + "spans": [ + { + "bbox": [ + 317, + 242, + 434, + 326 + ], + "type": "image", + "image_path": "08c33940578f9d96b74a21eac8f66689400098e8b2b51485485bdd528ae602a7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 336, + 554, + 359 + ], + "lines": [ + { + "bbox": [ + 313, + 336, + 554, + 359 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 554, + 359 + ], + "type": "text", + "content": "Figure 5. Rate-accuracy-drop (R-AD) comparisons among different tasks. Left: VTM baseline. Right: Hyperprior baseline." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 436, + 242, + 552, + 326 + ], + "blocks": [ + { + "bbox": [ + 436, + 242, + 552, + 326 + ], + "lines": [ + { + "bbox": [ + 436, + 242, + 552, + 326 + ], + "spans": [ + { + "bbox": [ + 436, + 242, + 552, + 326 + ], + "type": "image", + "image_path": "3b74e4b041fe53f57d9d431048afbbb86a5a2eff29e5313509193bbdceac2db8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 380, + 554, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 554, + 404 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 554, + 404 + ], + "type": "text", + "content": "and de-quantization, which are inverse processes of packing and quantization." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 411, + 443, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 443, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 443, + 423 + ], + "type": "text", + "content": "6.1.2. Encoding and decoding" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 426, + 554, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 554, + 510 + ], + "type": "text", + "content": "VTM baseline: We employ VTM-23.3 as the codec. The Intra coding is applied using the main configuration file encoder_intra_vtm.cfg. The InputChromaFormat is set to 400, and ConformanceWindowMode is enabled. The Internal-BitDepth, InputBitDepth, and OutputBitDepth are all set to 10. All other configurations are set to default. In our experiments, five QPs {22, 27, 32, 37, 42} are used." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 510, + 554, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 554, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 554, + 545 + ], + "type": "text", + "content": "Hyperprior baseline: We modify the channel of the input and output of the original Hyperprior as 1 to accommodate 2D features. The loss function is defined as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 363, + 555, + 553, + 570 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 555, + 553, + 570 + ], + "spans": [ + { + "bbox": [ + 363, + 555, + 553, + 570 + ], + "type": "interline_equation", + "content": "L = B P F P + \\lambda \\times \\left\\| \\left(F _ {o} - \\hat {F} _ {r}\\right) \\right\\| ^ {2} \\tag {1}", + "image_path": "822199ddd2d78a24b3e097ab390027a33dbd58d4126075c2f72c606bfa83b002.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 579, + 554, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 579, + 554, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 579, + 554, + 604 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 579, + 554, + 604 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 579, + 554, + 604 + ], + "type": "text", + "content": " is a scaling factor used to adjust the bitrate. Please find more details in the supplementary." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 612, + 474, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 474, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 474, + 625 + ], + "type": "text", + "content": "6.2. Rate-accuracy (R-A) analysis" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": "The R-A evaluation results are presented in Table 3. Both baselines can achieve significant bitrate savings, while the VTM baseline outperforms the Hyperprior baseline. To further assess the impact of bitrate on accuracy, we introduce a new metric, accuracy drop (AD), defined as the percentage decrease in accuracy. For Dpt, the reciprocal of RMSE is employed. The R-AD curves, shown in Fig. 5, reveal" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1074" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 64, + 71, + 288, + 108 + ], + "blocks": [ + { + "bbox": [ + 64, + 71, + 288, + 108 + ], + "lines": [ + { + "bbox": [ + 64, + 71, + 288, + 108 + ], + "spans": [ + { + "bbox": [ + 64, + 71, + 288, + 108 + ], + "type": "table", + "html": "
CodecClsSegDptCSRTTI
VTM Baseline0.94130.89550.85490.74900.9983
Hyperprior Baseline0.92960.92800.74370.47130.9996
", + "image_path": "5d5ec203a17cfcbc15d47fa78173195c99761e6f7e84d0178a4a418160e40e2d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 149, + 294, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 149, + 294, + 233 + ], + "spans": [ + { + "bbox": [ + 55, + 149, + 294, + 233 + ], + "type": "text", + "content": "that most tasks exhibit an inflection point where accuracy starts to decrease significantly. However, for Dpt, this point is observed only in the Hyperprior baseline. In the Hyperprior baseline, inflection points are more distracted across tasks, which may caused by the fact that different models are trained for different tasks. TTI is excluded as images generated from the original features have an FID of 0." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 240, + 241, + 252 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 240, + 241, + 252 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 241, + 252 + ], + "type": "text", + "content": "6.3. Distortion-accuracy (D-A) analysis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 257, + 295, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 257, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 295, + 413 + ], + "type": "text", + "content": "We evaluate the linear correlation between feature MSE (D) and task accuracy (A) and present coefficients of determination in Table 5. Among the baselines, only TTI exhibits high linear correlations. This is reasonable since TTI involves generating images, and the baselines employ image CODECs. In contrast, CSR fails to achieve a high linear correlation, as its features are derived from textual data. Overall, the VTM baseline outperforms the Hyperprior baseline, demonstrating its superior adaptability to large model features. Our analysis reveals that feature MSE struggles to characterize semantic distortion, particularly for textual data. This finding suggests that feature MSE is not an effective metric for measuring semantic distortion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 420, + 173, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 420, + 173, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 420, + 173, + 433 + ], + "type": "text", + "content": "6.4. Complexity analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 437, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 582 + ], + "type": "text", + "content": "We run the VTM baseline and Hyperprior baseline on an Intel® Xeon® E5-2690 v4 CPU and a single NVIDIA GeForce RTX 4090 GPU, respectively. The VTM baseline exhibits encoding times ranging from a few seconds to several hours, while its decoding time remains within a few seconds. In contrast, the Hyperprior baseline exhibits comparable encoding and decoding times, most of which are on the order of milliseconds. Detailed results are provided in the supplementary. We note that runtime depends heavily on both hardware capabilities and implementation details. A well-optimized codec and high-performance hardware can significantly accelerate feature coding." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 588, + 194, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 194, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 194, + 601 + ], + "type": "text", + "content": "6.5. Generalizability analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": "We assess the generalizability of trained models by compressing features extracted from one task using models trained on a different task and evaluating their R-A performance. We select three tasks and present the results in Table 6. Overall, for a given task (along column), three kinds of models behave differently in both bitrate and accuracy. For example, models trained on CSR struggle to achieve low bitrates when applied to Seg and TTI, while models trained on Seg fail to achieve low FID in TTI. More analyses can" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 322, + 71, + 549, + 240 + ], + "blocks": [ + { + "bbox": [ + 56, + 118, + 293, + 128 + ], + "lines": [ + { + "bbox": [ + 56, + 118, + 293, + 128 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 293, + 128 + ], + "type": "text", + "content": "Table 5. Distortion-accuracy (D-A) analysis on the two baselines." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 322, + 71, + 549, + 240 + ], + "lines": [ + { + "bbox": [ + 322, + 71, + 549, + 240 + ], + "spans": [ + { + "bbox": [ + 322, + 71, + 549, + 240 + ], + "type": "table", + "html": "
TaskSegCSRTTI
ModelsBPFPmIoUBPFPAcc.BPFPFID
Trained on Seg1.7177.962.0883.800.97119.83
1.3077.271.3766.600.62208.50
0.5474.820.780.200.31333.75
0.1262.580.250.000.12363.32
0.0337.110.090.000.05354.22
Trained on TTI2.5878.302.2897.601.4425.25
1.9676.341.4677.800.6652.12
1.1273.370.6532.000.2895.03
0.5356.900.300.000.15124.84
0.2642.940.110.000.08170.01
Trained on CSR5.2978.206.3491.404.7758.19
3.2777.943.6087.802.5973.19
1.8076.771.6882.401.02196.56
1.7375.821.5057.801.02191.69
1.5475.801.3534.800.94193.08
", + "image_path": "65c037153d537746857e86ad343a0fb211f23346206ab2abbaf1d71993a04d38.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 250, + 551, + 261 + ], + "lines": [ + { + "bbox": [ + 315, + 250, + 551, + 261 + ], + "spans": [ + { + "bbox": [ + 315, + 250, + 551, + 261 + ], + "type": "text", + "content": "Table 6. Generalizability evaluation on Seg, CSR, and TTI tasks." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 282, + 553, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 553, + 331 + ], + "type": "text", + "content": "be found in the supplementary. In practical applications, an effective compression model should be capable of handling diverse inputs. Therefore, a compression method with high generalizability is highly desirable." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 338, + 386, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 386, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 386, + 350 + ], + "type": "text", + "content": "6.6. Discussion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 356, + 553, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 356, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 553, + 464 + ], + "type": "text", + "content": "Large model applications require a balance of high accuracy, low complexity, and strong generalizability. However, both baselines struggle to maintain high accuracy at low bitrates, underscoring the need for more efficient feature coding strategies tailored to the unique characteristics of large model features. In addition, semantic distortion measurement remains a critical challenge in feature coding. Advancing semantic distortion metrics could further optimize codec performance and enhance overall coding efficiency." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 464, + 553, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 553, + 572 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 553, + 572 + ], + "type": "text", + "content": "The two baseline approaches present distinct advantages and drawbacks. The handcrafted approach aligns seamlessly with existing codecs but suffers from excessive runtime, limiting its real-time applicability. In contrast, the learning-based approach benefits from efficient GPU acceleration but lacks generalizability, as separate models must be trained for different features. Therefore, developing lightweight and generic codecs remains a promising research direction for feature coding." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 583, + 388, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 583, + 388, + 596 + ], + "spans": [ + { + "bbox": [ + 313, + 583, + 388, + 596 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 602, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 602, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 602, + 553, + 713 + ], + "type": "text", + "content": "In this paper, we highlight the importance of feature coding in large model deployments and introduce a new research area: large model feature coding. As the first step, we provide a test dataset, unified test conditions, baseline methods, and a benchmark. We examine the applicability of the two baselines and identify promising directions for future research. Our aim is to encourage collaboration between image/video coding and large model communities to drive the development of feature coding." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1075" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 124 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 124 + ], + "type": "text", + "content": "[1] Johannes Balle, David C. Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. ArXiv, abs/1802.01436, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 295, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 295, + 180 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 295, + 180 + ], + "type": "text", + "content": "[2] Benjamin Bross, Ye-Kui Wang, Yan Ye, Shan Liu, Jianle Chen, Gary J. Sullivan, and Jens-Rainer Ohm. Overview of the versatile video coding (VVC) standard and its applications. IEEE Transactions on Circuits and Systems for Video Technology, 31(10):3736-3764, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 182, + 295, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 295, + 225 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 295, + 225 + ], + "type": "text", + "content": "[3] Yangang Cai, Peiyin Xing, and Xuesong Gao. High efficient 3D convolution feature compression. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 228, + 294, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 228, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 62, + 228, + 294, + 270 + ], + "type": "text", + "content": "[4] Jingxue Chen, Hang Yan, Zhiyuan Liu, Min Zhang, Hu Xiong, and Shui Yu. When federated learning meets privacy-preserving computation. ACM Computing Surveys, 56(12): 1-36, 2024. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 273, + 295, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 273, + 295, + 305 + ], + "spans": [ + { + "bbox": [ + 62, + 273, + 295, + 305 + ], + "type": "text", + "content": "[5] Qiaoxi Chen, Changsheng Gao, and Dong Liu. End-to-end learned scalable multilayer feature compression for machine vision tasks. In ICIP, pages 1781-1787, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 307, + 294, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 307, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 307, + 294, + 350 + ], + "type": "text", + "content": "[6] Zhuo Chen, Kui Fan, Shiqi Wang, Lingyu Duan, Weisi Lin, and Alex Chichung Kot. Toward intelligent sensing: Intermediate deep feature compression. IEEE Transactions on Image Processing, 29:2230-2243, 2020. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 352, + 294, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 352, + 294, + 385 + ], + "spans": [ + { + "bbox": [ + 62, + 352, + 294, + 385 + ], + "type": "text", + "content": "[7] Hyomin Choi and Ivan V. Bajic. Latent-space scalability for multi-task collaborative intelligence. In ICIP, pages 3562-3566, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 387, + 294, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 387, + 294, + 420 + ], + "spans": [ + { + "bbox": [ + 62, + 387, + 294, + 420 + ], + "type": "text", + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: a large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 421, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 421, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 62, + 421, + 294, + 464 + ], + "type": "text", + "content": "[9] Ronan Docherty, Antonis Vamvakeros, and Samuel J Cooper. Upsampling DINoV2 features for unsupervised vision tasks and weakly supervised materials segmentation. arXiv preprint arXiv:2410.19836, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 466, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 466, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 57, + 466, + 294, + 510 + ], + "type": "text", + "content": "[10] Lingyu Duan, Jiaying Liu, Wenhan Yang, Tiejun Huang, and Wen Gao. Video coding for machines: A paradigm of collaborative compression and intelligent analytics. IEEE Transactions on Image Processing, 29:8680-8695, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 512, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 512, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 57, + 512, + 294, + 566 + ], + "type": "text", + "content": "[11] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The Llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2, 4, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 567, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 567, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 57, + 567, + 294, + 622 + ], + "type": "text", + "content": "[12] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 2, 4, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 624, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 294, + 668 + ], + "type": "text", + "content": "[13] Ruoyu Feng, Xin Jin, Zongyu Guo, Runsen Feng, Yixin Gao, Tianyu He, Zhizheng Zhang, Simeng Sun, and Zhibo Chen. Image coding for machines with omnipotent feature learning. In ECCV, pages 510-528. Springer, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "text", + "content": "[14] Othmane Friha, Mohamed Amine Ferrag, Burak Kantarci, Burak Cakmak, Arda Ozgun, and Nassira Ghoualmi-Zine. LLM-based edge intelligence: A comprehensive survey on architectures, applications, security and trustworthiness." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "IEEE Open Journal of the Communications Society, 5:5799-5856, 2024. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 98, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 98, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 98, + 553, + 140 + ], + "type": "text", + "content": "[15] Changsheng Gao, Dong Liu, Li Li, and Feng Wu. Towards task-generic image compression: A study of semantics-oriented metrics. IEEE Transactions on Multimedia, 25:721-735, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 143, + 553, + 176 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 143, + 553, + 176 + ], + "spans": [ + { + "bbox": [ + 316, + 143, + 553, + 176 + ], + "type": "text", + "content": "[16] Changsheng Gao, Yiheng Jiang, Li Li, Dong Liu, and Feng Wu. DMOFC: discrimination metric-optimized feature compression. In PCS, pages 1–5, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 178, + 553, + 211 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 178, + 553, + 211 + ], + "spans": [ + { + "bbox": [ + 316, + 178, + 553, + 211 + ], + "type": "text", + "content": "[17] Changsheng Gao, Zhuoyuan Li, Li Li, Dong Liu, and Feng Wu. Rethinking the joint optimization in video coding for machines: A case study. In DCC, pages 556-556, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 213, + 553, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 213, + 553, + 267 + ], + "spans": [ + { + "bbox": [ + 317, + 213, + 553, + 267 + ], + "type": "text", + "content": "[18] Changsheng Gao, Yiheng Jiang, Siqi Wu, Yifan Ma, Li Li, and Dong Liu. IMOFC: identity-level metric optimized feature compression for identification tasks. IEEE Transactions on Circuits and Systems for Video Technology, 35(2):1855-1869, 2025. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 270, + 553, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 270, + 553, + 314 + ], + "spans": [ + { + "bbox": [ + 316, + 270, + 553, + 314 + ], + "type": "text", + "content": "[19] Jean Kaddour, Joshua Harris, Maximilian Mozes, Herbie Bradley, Roberta Raileanu, and Robert McHardy. Challenges and applications of large language models. arXiv preprint arXiv:2307.10169, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 316, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 316, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 317, + 316, + 553, + 370 + ], + "type": "text", + "content": "[20] Chia-Hao Kao, Cheng Chien, Yu-Jen Tseng, Yi-Hsin Chen, Alessandro Gnutti, Shao-Yuan Lo, Wen-Hsiao Peng, and Riccardo Leonardi. Bridging compressed image latents and multimodal large language models. arXiv preprint arXiv:2407.19651, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 373, + 553, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 373, + 553, + 427 + ], + "spans": [ + { + "bbox": [ + 316, + 373, + 553, + 427 + ], + "type": "text", + "content": "[21] Yeongwoong Kim, Hyewon Jeong, Janghyun Yu, Younhee Kim, Jooyoung Lee, Se Yoon Jeong, and Hui Yong Kim. End-to-end learnable multi-scale feature compression for VCM. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 429, + 553, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 429, + 553, + 472 + ], + "spans": [ + { + "bbox": [ + 316, + 429, + 553, + 472 + ], + "type": "text", + "content": "[22] Jin Han Lee, Myung-Kyu Han, Dong Wook Ko, and Il Hong Suh. From big to small: Multi-scale local planar guidance for monocular depth estimation. arXiv preprint arXiv:1907.10326, 2019. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 475, + 553, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 475, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 317, + 475, + 553, + 529 + ], + "type": "text", + "content": "[23] Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan First, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. GShard: Scaling giant models with conditional computation and automatic sharding. arXiv preprint arXiv:2006.16668, 2020. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 532, + 553, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 532, + 553, + 576 + ], + "spans": [ + { + "bbox": [ + 317, + 532, + 553, + 576 + ], + "type": "text", + "content": "[24] Shibao Li, Chenxu Ma, Yunwu Zhang, Longfei Li, Chengzhi Wang, Xuerong Cui, and Jianhang Liu. Attention-based variable-size feature compression module for edge inference. The Journal of Supercomputing, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 578, + 553, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 578, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 317, + 578, + 553, + 632 + ], + "type": "text", + "content": "[25] Zhuoyuan Li, Junqi Liao, Chuanbo Tang, Haotian Zhang, Yuqi Li, Yifan Bian, Xihua Sheng, Xinmin Feng, Yao Li, Changsheng Gao, et al. Ustc-td: A test dataset and benchmark for image and video coding in 2020s. arXiv preprint arXiv:2409.08481, 2024. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 635, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 678 + ], + "type": "text", + "content": "[26] Zhuoyuan Li, Zikun Yuan, Li Li, Dong Liu, Xiaohu Tang, and Feng Wu. Object segmentation-assisted inter prediction for versatile video coding. IEEE Transactions on Broadcasting, 2024. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[27] Tie Liu, Mai Xu, Shengxi Li, Chaoran Chen, Li Yang, and Zhuoyi Lv. Learn mutual feature compression for machine vision. In ICASSP, pages 1-5, 2023. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1076" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 117 + ], + "type": "text", + "content": "[28] Guo Lu, Xingtong Ge, Tianxiong Zhong, Qiang Hu, and Jing Geng. Preprocessing enhanced image compression for machine vision. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 173 + ], + "type": "text", + "content": "[29] Lingjuan Lyu, Han Yu, Xingjun Ma, Chen Chen, Lichao Sun, Jun Zhao, Qiang Yang, and Philip S. Yu. Privacy and robustness in federated learning: Attacks and defenses. IEEE Transactions on Neural Networks and Learning Systems, 35 (7):8726-8746, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 295, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 295, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 295, + 207 + ], + "type": "text", + "content": "[30] Yifan Ma, Changsheng Gao, Qiaoxi Chen, Li Li, Dong Liu, and Xiaoyan Sun. Feature compression with 3d sparse convolution. In VCIP, pages 1-5, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "type": "text", + "content": "[31] Rui Mao, Xinmin Feng, Changsheng Gao, Li Li, Dong Liu, and Xiaoyan Sun. Perceptual image compression with conditional diffusion transformers. In VCIP, pages 1-5, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 252, + 294, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 252, + 294, + 286 + ], + "spans": [ + { + "bbox": [ + 56, + 252, + 294, + 286 + ], + "type": "text", + "content": "[32] Kiran Misra, Tianying Ji, Andrew Segall, and Frank Bossen. Video feature compression for machine tasks. In ICME, pages 1-6, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 287, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 287, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 287, + 294, + 342 + ], + "type": "text", + "content": "[33] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. DINOv2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 4, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 343, + 294, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 294, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 294, + 398 + ], + "type": "text", + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 295, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 295, + 443 + ], + "type": "text", + "content": "[35] Xihua Sheng, Li Li, Dong Liu, and Houqiang Li. Vnvc: A versatile neural video coding framework for efficient human-machine vision. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(7):4579-4596, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 444, + 295, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 444, + 295, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 444, + 295, + 477 + ], + "type": "text", + "content": "[36] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-CLIP: improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 478, + 295, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 295, + 543 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 295, + 543 + ], + "type": "text", + "content": "[37] Satoshi Suzuki, Shoichiro Takeda, Motohiro Takagi, Ryuichi Tanida, Hideaki Kimata, and Hayaru Shouno. Deep feature compression using spatio-temporal arrangement toward collaborative intelligent world. IEEE Transactions on Circuits and Systems for Video Technology, 32(6):3934-3946, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 544, + 295, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 295, + 589 + ], + "type": "text", + "content": "[38] Yuanyishu Tian, Yao Wan, Lingjuan Lyu, Dezhong Yao, Hai Jin, and Lichao Sun. FedBERT: when federated learning meets pre-training. ACM Transactions on Intelligent Systems and Technology, 13(4):1-26, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 590, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 295, + 633 + ], + "type": "text", + "content": "[39] Yuan Tian, Guo Lu, Guangtao Zhai, and Zhiyong Gao. Nonsemantics suppressed mask learning for unsupervised video semantic compression. In ICCV, pages 13564-13576, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 635, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 295, + 668 + ], + "type": "text", + "content": "[40] Yuan Tian, Guo Lu, and Guangtao Zhai. SMC++: masked learning of unsupervised video semantic compression. arXiv preprint arXiv:2406.04765, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 669, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 295, + 712 + ], + "type": "text", + "content": "[41] Yuan Tian, Guo Lu, and Guangtao Zhai. Free-VSC: free semantics from visual foundation models for unsupervised video semantic compression. In ECCV, pages 163-183, 2025. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 554, + 597 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 315, + 72, + 554, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 554, + 128 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 554, + 128 + ], + "type": "text", + "content": "[42] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 554, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 554, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 554, + 173 + ], + "type": "text", + "content": "[43] Praneeth Vepakomma, Otkrist Gupta, Tristan Swedish, and Ramesh Raskar. Split learning for health: Distributed deep learning without sharing raw patient data. arXiv preprint arXiv:1812.00564, 2018. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 174, + 554, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 554, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 554, + 228 + ], + "type": "text", + "content": "[44] Shurun Wang, Shiqi Wang, Wenhan Yang, Xinfeng Zhang, Shanshe Wang, Siwei Ma, and Wen Gao. Towards analysis-friendly face representation with scalable feature and texture compression. IEEE Transactions on Multimedia, 24:3169-3181, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 229, + 554, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 229, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 229, + 554, + 262 + ], + "type": "text", + "content": "[45] Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng Chua. NExT-GPT: Any-to-any multimodal LLM. In ICML, pages 53366-53397, 2024. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 262, + 554, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 554, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 554, + 308 + ], + "type": "text", + "content": "[46] Biwei Yan, Kun Li, Minghui Xu, Yuean Dong, Yue Zhang, Zhaochun Ren, and Xiuzhen Cheng. On protecting the data privacy of large language models (LLMs): A survey. arXiv preprint arXiv:2403.05156, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 308, + 554, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 554, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 554, + 352 + ], + "type": "text", + "content": "[47] Ning Yan, Changsheng Gao, Dong Liu, Houqiang Li, Li Li, and Feng Wu. SSSIC: Semantics-to-signal scalable image coding with learned structural representations. IEEE Transactions on Image Processing, 30:8939-8954, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 353, + 554, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 353, + 554, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 353, + 554, + 408 + ], + "type": "text", + "content": "[48] Wenhan Yang, Haofeng Huang, Yueyu Hu, Ling-Yu Duan, and Jiaying Liu. Video coding for machines: Compact visual representation compression for intelligent collaborative analytics. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-18, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 409, + 554, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 409, + 554, + 463 + ], + "spans": [ + { + "bbox": [ + 316, + 409, + 554, + 463 + ], + "type": "text", + "content": "[49] Rui Ye, Wenhao Wang, Jingyi Chai, Dihan Li, Zexi Li, Yinda Xu, Yaxin Du, Yanfeng Wang, and Siheng Chen. OpenFedLLM: Training large language models on decentralized private data via federated learning. In ACM SIGKDD, pages 6137-6147, 2024. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 464, + 554, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 464, + 554, + 507 + ], + "spans": [ + { + "bbox": [ + 316, + 464, + 554, + 507 + ], + "type": "text", + "content": "[50] Xu Zhang, Peiyao Guo, Ming Lu, and Zhan Ma. All-in-one image coding for joint human-machine vision with multipath aggregation. arXiv preprint arXiv:2409.19660, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 509, + 554, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 509, + 554, + 543 + ], + "spans": [ + { + "bbox": [ + 316, + 509, + 554, + 543 + ], + "type": "text", + "content": "[51] Zhicong Zhang, Mengyang Wang, Mengyao Ma, Jiahui Li, and Xiaopeng Fan. MSFC: Deep feature compression in multi-task network. In ICME, pages 1-6, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 543, + 554, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 543, + 554, + 597 + ], + "spans": [ + { + "bbox": [ + 316, + 543, + 554, + 597 + ], + "type": "text", + "content": "[52] Jiaying Zheng, Hainan Zhang, Lingxiang Wang, Wangjie Qiu, Hongwei Zheng, and Zhiming Zheng. Safely learning with private data: A federated learning framework for large language model. arXiv preprint arXiv:2406.14898, 2024. 1, 3" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1077" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Feature Decomposition-Recomposition in Large Vision-Language Model for Few-Shot Class-Incremental Learning/df51b4e9-12ac-40bf-8801-dbd6097351a5_content_list.json b/2025/Feature Decomposition-Recomposition in Large Vision-Language Model for Few-Shot Class-Incremental Learning/df51b4e9-12ac-40bf-8801-dbd6097351a5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7d78f44fb732bcc07de0889d8f312c7bc6583f12 --- /dev/null +++ b/2025/Feature Decomposition-Recomposition in Large Vision-Language Model for Few-Shot Class-Incremental Learning/df51b4e9-12ac-40bf-8801-dbd6097351a5_content_list.json @@ -0,0 +1,1461 @@ +[ + { + "type": "text", + "text": "Feature Decomposition-Recomposition in Large Vision-Language Model for Few-Shot Class-Incremental Learning", + "text_level": 1, + "bbox": [ + 117, + 128, + 880, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zongyao Xue $^{1,2}$ , Meina Kan $^{1,2}$ , Shiguang Shan $^{1,2,3}$ , Xin Chen $^{1,2}$", + "bbox": [ + 238, + 202, + 756, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Institute of Computing Technology, Chinese Academy of Sciences, Beijing, 100090, China", + "bbox": [ + 135, + 220, + 861, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ University of Chinese Academy of Sciences, Beijing, 100090, China", + "bbox": [ + 223, + 238, + 772, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Peng Cheng Laboratory, Shenzhen, 518055, China", + "bbox": [ + 292, + 256, + 699, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Few-Shot Class-Incremental Learning (FSCIL) focuses on incrementally learning novel classes using only a limited number of samples from novel classes, which faces dual challenges: catastrophic forgetting of previously learned classes and over-fitting to novel classes with few available samples. Recent advances in large pre-trained vision-language models (VLMs), such as CLIP, provide rich feature representations that generalize well across diverse classes. Therefore, freezing the pre-trained backbone and aggregating class features as prototypes becomes an intuitive and effective way to mitigate catastrophic forgetting. However, this strategy fails to address the overfitting challenge, and the prototypes of novel classes exhibit semantic bias due to the few samples per class. To address these limitations, we propose a semantic Feature Decomposition-Recomposition (FDR) method based on VLMs. Firstly, we decompose the CLIP features into semantically distinct segments guided by text keywords from base classes. Then, these segments are adaptively recomposed at the attribute level given text descriptions, forming calibrated prototypes for novel classes. The recomposition process operates linearly at the attribute level but induces nonlinear adjustments across the entire prototype. This fine-grained and non-linear recomposition inherits the generalization capabilities of VLMs and the adaptive recomposition ability of base classes, leading to enhanced performance in FSCIL. Extensive experiments demonstrate our method's effectiveness, particularly in 1-shot scenarios where it achieves improvements between $6.70\\%$ and $19.66\\%$ for novel classes over previous prototype-based methods on CUB200. Code is available at https://github.com/herb1999/FDR.", + "bbox": [ + 88, + 340, + 485, + 811 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 843, + 220, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The remarkable success of deep neural networks (DNNs) is largely based on training with large-scale static datasets", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "[23]. However, real-world scenarios such as autonomous driving and robotic field operations often involve continuously arriving data streams, where new categories emerge incrementally with only sparse samples available. Although humans excel at adapting to such dynamic environments, existing learning-based algorithms struggle under these constraints. To bridge this gap, Few-Shot Class-Incremental Learning (FSCIL)[22] is introduced as a more practical and challenging task, aiming to incrementally recognize novel classes from few-shot examples while preserving performance on base classes. The FSCIL task inherits dual challenges from few-shot learning and continual learning. First, models must overcome the unreliable empirical risk minimization problem [26] when generalizing with scarce novel class samples. Second, they must maintain stability against catastrophic forgetting [16] of base classes while adapting to incremental updates.", + "bbox": [ + 511, + 309, + 906, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Since its introduction in 2020 [22], FSCIL has attracted significant attention due to its potential in real-world applications, leading to a variety of early methodological developments, including meta learning[9, 31, 33], specialized model architectures[28, 29], optimization strategies [19] and replay-based methods[10, 15]. However, these approaches face a longstanding plasticity-stability dilemma, wherein models struggle to simultaneously preserve base class knowledge while effectively accommodating novel class acquisition.", + "bbox": [ + 511, + 566, + 908, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, leveraging the generalizable representations of pre-trained models such as CLIP, prototype-based metric learning approaches [1, 21, 30, 32] have emerged as a prominent solution to mitigate catastrophic forgetting. These methods typically freeze the pre-trained backbone networks and aggregate features from samples within each class to construct prototypes, enabling effective knowledge preservation and incremental learning. However, the scarce samples of novel classes hinder the full exploitation of features from large pre-trained models, yielding semantically biased novel prototypes that deviate from true class distributions. To address these biases, existing methods [1, 7, 10]", + "bbox": [ + 511, + 719, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3153", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "leverage supplementary information from base class prototypes to calibrate novel ones. Nevertheless, these approaches directly employ entire similar base prototypes to compensate for the biased novel ones, lacking mechanisms to identify which specific components are transferable. This oversight leads to reduced inter-class distances between novel and base classes, causing feature space entanglement.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these limitations, we propose a Feature Decomposition-Recomposition (FDR) framework that enables fine-grained prototype calibration through attribute-level knowledge transfer. Our key insight is that although the feature representations from large pre-trained models contain sufficient information for classifying diverse classes, their correct classification depends on strategically extracting generalizable features, particularly for few-shot classes. Specifically, given features from pre-trained VLMs like CLIP, we first decompose the visual prototypes of base classes into semantic-aware feature segments by establishing mappings between CLIP visual features and attribute keywords through a keyword grounding function. These semantic feature segments are transferable across classes. Next, for novel classes, given a set of keywords describing them, each corresponding feature segment of the biased prototype derived from few-shot samples is calibrated by recombining segments from similar base class prototypes (e.g., calibrating the \"cannon\" prototype by combining \"wheel\" segments from classes like bicycles and \"cylindrical\" segments from classes like lipsticks). The combinatorial possibilities for synthesizing novel prototypes grow exponentially as the number of base classes increases, effectively addressing the prototype entanglement problem.", + "bbox": [ + 91, + 198, + 486, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions of this work are threefold:", + "bbox": [ + 109, + 560, + 403, + 573 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We propose a semantic feature decomposition and recomposition method, FDR, that divides CLIP's frozen features into reusable attribute-level feature segments, enabling precise and easy knowledge transfer from base classes to novel ones for enhanced incremental learning.", + "2. We design a keyword grounding function to decompose CLIP's frozen features into semantic-aware segments, and introduce a segment-wise recomposition method to calibrate biased prototypes at the fine-grained attribute-level.", + "3. Our approach is training-free, simple yet highly effective, achieving improvements of $6.70\\% - 19.66\\%$ in novel class accuracy over state-of-the-art methods, showcasing its practical utility." + ], + "bbox": [ + 89, + 574, + 482, + 785 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 799, + 232, + 814 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Incremental learning (IL), also referred to as continual or lifelong learning, distinguishes itself from conventional machine learning by addressing dynamic data distributions where models must continually integrate new knowledge from sequential tasks while preserving performance on pre", + "bbox": [ + 89, + 825, + 482, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "viously learned ones [25]. The core challenge lies in balancing learning plasticity for new knowledge and memory stability for old knowledge [16]. An intuitive approach to ensure the stability of old knowledge is to retain representative samples from previous tasks and replay them during training on new tasks[8, 13, 18]. However, growing memory demands and privacy concerns limit their scalability. To enhance storage efficiency, retaining features of old samples and replaying them later is also an option[12]. Another method for preserving old knowledge is generative replay[20, 27], where a generative model learns the distribution of old data and replays generated samples or features. These methods primarily focus on preserving old knowledge rather than leveraging it to enhance new task learning.", + "bbox": [ + 511, + 90, + 903, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Few-Shot Learning (FSL) requires models to learn from a limited number of labeled samples, making it suitable for real-world scenarios where data is scarce due to safety, privacy, or ethical concerns[26]. Metric-based methods [14, 21, 24] have proven to be simple and effective solutions by learning transferable embedding spaces. Prototypical Networks[21], for instance, define a metric space where classification is performed by computing distances to prototype representations of each class. These prototypes, computed as class-mean feature vectors, naturally accommodate incremental updates by freezing the embedding backbone - a property that makes them particularly suitable for few-shot class-incremental learning [22]. However, prototype quality degrades significantly under extreme data scarcity (e.g., 1-shot settings), as limited samples fail to capture true class distributions, leading to biased prototypes that misalign with semantic realities.", + "bbox": [ + 511, + 306, + 903, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Few-Shot Class-Incremental Learning (FSCIL) aims to incrementally learn novel classes given only its few shots[22], which merges challenges from IL and FSL: preventing catastrophic forgetting of base classes while generalizing to novel classes with sparse samples. A common strategy to mitigate this is to use semantically similar base classes to correct biases in novel classes[1, 7, 10]. Additionally, incorporating side information from language has been explored to enhance model performance[7, 10]. The closest to our work is [1], which reduces novel prototype bias by globally calibrating them with similar base prototypes. However, such coarse-grained adjustments compress interclass distances, leading to prototype entanglement in later phases. Our method addresses these limitations through fine-grained calibration, decomposing prototypes into semantically meaningful sub-features for targeted composition. This enables precise inheritance of relevant visual attributes from multiple base classes while maintaining interclass separability. The integration of language semantics also resolves ambiguity in visual similarity measurements, particularly for classes sharing partial attributes rather than global appearances.", + "bbox": [ + 511, + 569, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3154", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Background", + "text_level": 1, + "bbox": [ + 89, + 89, + 220, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following the existing works [1, 7, 10, 22], the Few-Shot Class-Incremental Learning (FSCIL) task is defined as a process consisting of $(1 + T)$ sequential learning sessions, which are divided into two consecutive phases:", + "bbox": [ + 89, + 114, + 483, + 175 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Base phase (the first session $S_0$ ): during $S_0$ , the model is trained on $M$ base classes $\\mathcal{C}^0 = \\{C_1, \\dots, C_M\\}$ with abundant samples per class.", + "- Incremental phase (sessions $S_{1} \\sim S_{T}$ ): following the N-way K-shot protocol, at each session $S_{t}|_{t=1}^{T}$ , the model learns $N$ novel classes with $K$ samples per class (typically $K \\leq 5$ ). The novel classes are denoted as $\\mathcal{C}^t = \\{C_{M^t + 1}, \\dots, C_{M^t + N}\\}$ with $M^t = M + (t - 1) \\times N$ . The label spaces across distinct sessions remain strictly disjoint, i.e. $\\mathcal{C}^i \\cap \\mathcal{C}^j = \\emptyset \\quad \\forall i, j \\in \\{0, \\dots, T\\}, i \\neq j$ ." + ], + "bbox": [ + 89, + 176, + 483, + 328 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The model in few-shot class-incremental learning is expected to maintain classification accuracy across all base classes while achieving precise classification for novel classes. When leveraging a large pre-trained model as the feature extractor, it can extract rich and informative features for both base and novel classes attributing to its generalization capability. Consequently, obtaining a well-defined prototype is often sufficient for classification.", + "bbox": [ + 89, + 329, + 483, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For base classes, prototypes can be easily derived by aggregating abundant samples, as demonstrated by existing methods like [21]. Specifically, each class $C$ is represented by a prototype vector $\\mathbf{p}_c\\in \\mathbb{R}^d$ , computed as the mean feature of all samples belonging to that class:", + "bbox": [ + 89, + 450, + 483, + 526 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} _ {c} = \\frac {1}{| \\mathcal {I} _ {c} |} \\sum_ {\\left(\\mathbf {x} _ {i}, y _ {i}\\right) \\in \\mathcal {I} _ {c}} f _ {\\theta} \\left(\\mathbf {x} _ {i}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 532, + 482, + 570 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $f_{\\theta}$ denotes the frozen feature extractor (large pretrained model like CLIP in this work), and $\\mathcal{I}_c$ contains all available samples for class $C$ .", + "bbox": [ + 89, + 575, + 483, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For novel classes at session $S_{t}$ , the prototypes can also be derived from few-shot samples following Eq. (1). However, these prototypes are often biased due to the limited number of training examples. Such bias can result in suboptimal classification performance, presenting a significant challenge in FSCIL.", + "bbox": [ + 89, + 621, + 483, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During inference at session $S_{t}$ , the classification probability for a test sample $\\mathbf{x}$ is calculated by measuring the distance between its feature representation $f_{\\theta}(\\mathbf{x})$ and the prototype vectors of all classes encountered (i.e. $\\bigcup_{i=0}^{t} \\mathcal{C}^i$ ) so far.", + "bbox": [ + 89, + 712, + 483, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 89, + 799, + 181, + 815 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To mitigate semantic bias in novel-class prototypes derived from scarce samples, we propose a semantically feature decomposition-recomposition approach for calibrating the prototype in a fine-grained manner at keyword attribute level. An overview is illustrated in Fig. 1.", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our proposed method consists of two stages: 1) semantic feature decomposition stage, which establishes a grounding of text attribute keywords to CLIP features using base classes. It produces a set of meaningful feature segments, each corresponding to a specific semantic keyword. These feature segments are transferable across classes, ensuring their help to novel classes. 2) Feature segments recomposition stage, which adaptively calibrates the prototypes of novel classes by recomposing the semantic-aware feature segments according to the keywords associated with the novel classes. This process enables fine-grained refinement of the prototypes at keyword attribute level, mitigating bias and improving classification performance.", + "bbox": [ + 511, + 90, + 906, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Semantic feature decomposition", + "text_level": 1, + "bbox": [ + 511, + 301, + 795, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this stage, for each base class $C$ , we begin by generating visual descriptions and attribute keywords using Large Language Models (LLMs). Next, leveraging the vision-language alignment capabilities of CLIP models, a fine-grained mapping from these keywords to visual features is established, producing a set of semantically meaningful feature segments.", + "bbox": [ + 511, + 323, + 905, + 429 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Firstly, with the help of LLMs, a textual description $\\mathcal{D}_c$ for each class $C$ is generated conditioned on class names (e.g. \"toaster\"). For instance, $\\mathcal{D}_c$ might be \"A rectangular kitchen appliance with metal or plastic housing, slots for inserting bread, and controls like dials Buttons.\" Additionally, $m$ attribute keywords $\\mathcal{K}_C = \\{k_1,\\dots,k_m\\}$ are extracted from $\\mathcal{D}_c$ . Each keyword $k_i$ describes a visual attribute of the class $C$ (e.g. \"rectangular\" and \"slots\" for class \"toaster\"). These textual descriptions and keywords capture the semantic attributes of the class.", + "bbox": [ + 511, + 430, + 905, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Secondly, we design a keyword grounding function $\\mathcal{G}_{\\phi}$ to establish fine-grained mappings from these keywords to visual feature segments of base prototypes. Leveraging CLIP's frozen visual encoder $f_{\\theta}$ and textual encoder $g_{\\phi}$ , we obtain aligned visual prototype $\\mathbf{p}_c$ (same as Eq. (1)) and textual representations $\\mathbf{t}_c = g_{\\phi}(\\mathcal{D}_c)$ , where $\\mathbf{t}_c \\in \\mathbb{R}^d$ denotes the text embedding of description $\\mathcal{D}_c$ . Since CLIP's text and image embeddings are aligned, mapping a keyword $k_i$ to a segment of the text feature space inherently implies a corresponding mapping to the associated segment in the image feature space. Therefore, for each keyword $k_i \\in \\mathcal{K}_C$ , the keyword grounding function $\\mathcal{G}_{\\phi}$ generates a grounding vector $\\mathbf{g}_i$ by evaluating the relevance of the keyword to the entire textual representation:", + "bbox": [ + 511, + 583, + 906, + 794 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {g} _ {i} = \\mathcal {G} _ {\\phi} \\left(k _ {i}, \\mathcal {D} _ {c}\\right), \\in \\{0, 1 \\} ^ {d}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 808, + 906, + 825 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here $\\mathbf{g}_i$ is a sparse binary mask, where entries with a value of 1 indicate the dimensions at which the keyword $k_{i}$ contributes most significantly to the text embedding $\\mathbf{t}_c$ of description $\\mathcal{D}_c$ .", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3155", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9c6130e2000b378c177614f4f80edfe524f8aec6579e878daf9c0a57c951b6e5.jpg", + "image_caption": [ + "Feature Decomposition of Base classes", + "Segments Decomposition for calibrating Novel class", + "Figure 1. Overall Architecture of Our Feature Decomposition-Recomposition (FDR) method. The calibration process is as follows: 1) Feature Decomposition of Base classes: For all classes, LLMs are leveraged to generate category descriptions and extract semantic keywords. CLIP encodes both image features (aggregated as prototypes) and textual descriptions. A Keyword Grounding Function then decomposes textual embeddings into semantic-aware feature segments aligned with specific keywords, achieving attribute-level decomposition of visual and textual features. 2) Segments Recomposition for calibrating Novel class: For novel classes, we match each keyword-derived textual feature segment with semantically similar segments from base classes. The corresponding visual feature segments from multiple base prototypes are recomposed to calibrate the raw novel prototype. For instance, when calibrating the cannon prototype (which suffers from missing wheel features due to limited samples), our method identifies wheeled segments from unicycle and oxcart prototypes and cylindrical segments from lipstick and water tower prototypes. This fine-grained recomposition supplements missing attributes while preserving class-discriminative characteristics." + ], + "image_footnote": [], + "bbox": [ + 98, + 89, + 898, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During the text encoding process of CLIP, the embedding of the entire text description $\\mathcal{D}_c$ inherently incorporates the contributions of tokens from each word including those visual keywords $k_{i}\\in \\mathcal{K}_{C}$ . This means that the attention mechanisms and other internal structures implicitly reveal how and where each keyword contributes to the overall textual representation. Therefore, the keyword grounding function $\\mathcal{G}_{\\phi}$ is designed to leverage CLIP's internal attention mechanisms to identify these contributions.", + "bbox": [ + 89, + 500, + 482, + 635 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, in the last transformer layer of CLIP textual encoder $g_{\\phi}$ , $\\mathcal{D}_c$ is tokenized as $n$ tokens $\\{[SOS], \\text{token}_1, \\dots, \\text{token}_{n-2}, [EOS]\\}$ . Given the attention matrix $\\mathbf{A} \\in \\mathbb{R}^{n \\times n}$ and value matrix $\\mathbf{V} \\in \\mathbb{R}^{d \\times n}$ , $\\mathbf{t}_c$ is calculated as the output of the last transformer layer as follows:", + "bbox": [ + 89, + 636, + 482, + 724 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {t} _ {c} = \\mathbf {V} (:,:) \\mathbf {A} (:, e o s) + \\mathbf {h} = \\left(a _ {1} \\mathbf {v} _ {1} + \\dots + a _ {n} \\mathbf {v} _ {n}\\right) + \\mathbf {h}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 738, + 482, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{A}$ represents the inter-token attention weights, $\\mathbf{V}$ represents transformed token representations of all words, $a_{i} = A[i,p_{\\mathrm{eos}}]$ represents the attention weight from the $i$ -th token to the [EOS] token, and $\\mathbf{v}_i$ means the value vector of the $i$ -th token from CLIP's final transformer layer.", + "bbox": [ + 89, + 763, + 482, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As can be seen from the calculation of $\\mathbf{t}_c$ in Eq. (3), the attention matrix $\\mathbf{A}$ and value matrix $\\mathbf{V}$ implicitly reveal how and where each keyword contributes to the overall textual representation. Therefore, we extract these token", + "bbox": [ + 89, + 840, + 483, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "wise contribution values as a contribution indication matrix $\\mathbf{B}\\in \\mathbb{R}^{d\\times n}$ ..", + "bbox": [ + 511, + 500, + 903, + 527 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {B} = \\left[ a _ {1} \\mathbf {v} _ {1}, \\dots , a _ {n} \\mathbf {v} _ {n} \\right]. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 530, + 903, + 546 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, the columns of $\\mathbf{B}$ describe different text tokens, while the rows represent feature channels. Each element $\\mathbf{B}(j,i)$ indicates the contribution of the $i$ -th token's $j$ -th feature dimension to the final [EOS] embedding, serving as an effective indicator for keyword-related feature dimensions.", + "bbox": [ + 511, + 551, + 905, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Therefore, to quantify the influence of each token on the text embedding $\\mathbf{B}$ , we apply a softmax operation along the token dimension (horizontal) of $\\mathbf{B}$ . This results in a matrix $\\mathbf{C}$ , where each element represents the proportion of token $i$ 's contribution to $\\mathbf{t}_c$ at feature dimension $j$ , relative to all text tokens:", + "bbox": [ + 511, + 627, + 905, + 715 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C} (i,:) = \\operatorname {S o f t m a x} (\\mathbf {B} (i,:)) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 723, + 903, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, we perform top-q binarization along the feature dimension (vertical) of $\\mathbf{B}$ , converting each column $\\mathbf{C}(:, i]$ into into a sparse binary mask $\\mathbf{g}_i$ (previously introduced in Eq. (2)) as follows:", + "bbox": [ + 511, + 750, + 905, + 808 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {g} _ {i} (j) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} \\mathbf {C} (i, j) \\text {i s w i t h i n t o p - q r e s p o n s i v e o f} \\mathbf {C} (:, j) \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 816, + 919, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This process isolates the most significant feature dimensions associated with each keyword, enabling grounding of", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3156", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "textual keywords to text feature segments.", + "bbox": [ + 89, + 90, + 369, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "According to Eqs. (2) to (6), we can derive a fine-grained mapping from each keyword to its corresponding feature segments. While this mapping is computed independently for each class, ideally, the same keyword associated with different classes should yield identical mappings. However, in practice, slight discrepancies often arise due to variations in class contexts. To ensure consistency, for keywords that appear across multiple base classes, we unify all selected feature dimensions by taking the union of their respective binary masks. This strategy preserves all potentially relevant dimensions for the keyword, enhancing robustness and consistency in the grounding process.", + "bbox": [ + 89, + 106, + 483, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since CLIP's visual and textual feature spaces are aligned, given a keyword $k_{i}$ , we can obtain the visual feature segment $\\mathbf{s}_i^c$ of prototype $\\mathbf{p}_c$ , by simply performing an element-wise multiplication as follows:", + "bbox": [ + 89, + 287, + 483, + 348 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {s} _ {i} ^ {c} = \\mathbf {p} _ {c} \\odot \\mathbf {g} _ {i}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 236, + 354, + 482, + 371 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this way, we achieve semantic-aware decomposition of base class prototypes guided by textual keywords of visual attributes, where each feature segment $\\mathbf{s}_i^c$ retains only the dimensions most relevant to the corresponding semantic keyword.", + "bbox": [ + 89, + 377, + 483, + 453 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Feature segments decomposition for fine-grained calibration", + "text_level": 1, + "bbox": [ + 89, + 460, + 482, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The prototypes of novel classes are often biased due to the limited number of shots available per class. While the full-dimensional features of these prototypes can only be derived from the novel class samples, certain partial features—such as visual attribute features—frequently overlap with those of base class prototypes. For instance, features like \"wheels\" are shared between the novel class \"cannon\" and base classes \"unicycle\", \"oxcart\", and \"race car\". Leveraging this inherent property, we calibrate the prototypes of novel classes by piecewise refinement of their semantic feature segments. This is achieved by assembling similar segments from the base classes, thereby enhancing the representational accuracy and reducing bias in the novel class prototypes.", + "bbox": [ + 89, + 497, + 483, + 708 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this process, given any novel class $\\widetilde{C}$ , we first generate its initial prototype $\\mathbf{p}_{\\widetilde{C}}$ and visual keywords $\\kappa_{\\widetilde{C}}$ following the same process as base classes.", + "bbox": [ + 89, + 708, + 482, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, for each keyword $k_{i} \\in \\mathcal{K}_{\\widetilde{C}}$ , we retrieve the corresponding feature segments $\\mathbf{s}_i^{c_j}$ from the base-class prototypes that share the same keywords. These segments are then combined using a weighted summation to calibrate the feature segment of the novel class:", + "bbox": [ + 89, + 753, + 483, + 829 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {r} _ {i} = \\sum_ {j = 1} ^ {M} w _ {j} \\cdot \\mathbf {s} _ {i} ^ {c _ {j}}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 835, + 482, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, the $\\mathbf{s}_i^{c_j}$ represents the prototype segment of base class $C_j$ corresponding to the keyword $k_{i}$ . The weight $w_{j}$ reflects", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/87a5604a38b60b14516f41952a8c2ecce84ee4c5b91d57742768105e91385c2b.jpg", + "image_caption": [ + "Figure 2. Implementation of keyword grounding function $\\mathcal{G}_{\\phi}$ . The upper describes the computation of $\\mathbf{t}_c$ using CLIP, while the lower details the process for generating the grounding vector $\\mathbf{g}_c$ ." + ], + "image_footnote": [], + "bbox": [ + 519, + 92, + 901, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the similarity between the feature segments of $\\widetilde{C}$ and base class $C_j$ . This similarity is approximated by computing the textual embedding similarity between the textual segment of novel class $\\mathbf{t}_{\\widetilde{c}} \\odot \\mathbf{g}_i$ and that of every base class $\\mathbf{t}_{c_j} \\odot \\mathbf{g}_i, C_j \\in C^0$ . The top-r similarity scores are preserved after using a softmax function, while the remaining similarity scores are set to zero. This ensures that only the most semantically relevant base-class feature segments are retained, yielding the final weights $w_j$ .", + "bbox": [ + 511, + 458, + 905, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After obtaining feature segments for all keywords associated with novel class $\\widetilde{C}$ , keyword-specific adjustments are iteratively applied to novel prototype $\\mathbf{p}_{\\widetilde{c}}$ to obtain the final calibrated prototype as below:", + "bbox": [ + 511, + 598, + 905, + 659 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} _ {\\tilde {c}} \\leftarrow (1 - \\alpha) \\cdot \\mathbf {p} _ {\\tilde {c}} + \\alpha \\cdot \\mathbf {r} _ {i}, \\quad i \\in \\{1, \\dots , m \\} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 696, + 906, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ is a trade-off hyperparameter that balances the prototype directly from the few shots and calibrated one.", + "bbox": [ + 511, + 731, + 905, + 761 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Eq. (9), each calibration $\\mathbf{r}_i$ is performed individually and is a linear combination of similar feature segments from base-class prototypes. Importantly, this linear combination varies for different feature segments, as the weights and contributing segments depend on the specific keyword $k_i$ . Consequently, the overall calibration is semantic-segmentwise linear but nonlinear overall, ensuring fine-grained and more accurate prototype calibration for few-shot novel classes.", + "bbox": [ + 511, + 763, + 906, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3157", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/040f950dfd326803f0fd3876105c9c9091421288d4cb4d82d60b468592b80450.jpg", + "table_caption": [ + "Table 1. Comparison with existing methods (5-shot setting). $A_{base}$ for base session accuracy, $A_{avg}$ for AvgAcc across all sessions" + ], + "table_footnote": [], + "table_body": "
MethodBackboneTraining FreeCIFAR100miniImageNetCUB200IN1K-FSCIL
\\(A_{base}\\)\\(A_{avg}\\)\\(A_{base}\\)\\(A_{avg}\\)\\(A_{base}\\)\\(A_{avg}\\)\\(A_{base}\\)\\(A_{avg}\\)
CLOSER[2]ResNet-18/20No75.763.176.062.879.469.1--
Yourself[6]ViT-B-IN21KNo82.967.084.068.883.469.9--
LP-DiF[4]ViT-B-CLIPNo80.275.196.393.883.974.0--
PriViLego[5]ViT-B-CLIPNo84.277.295.492.782.174.272.168.9
EASE[3]ViT-B-IN21KNo85.178.096.893.983.075.173.370.4
FDR(Ours)ViT-B-CLIPNo87.779.395.389.284.477.773.670.9
BaselineViT-B-CLIPYes75.767.992.088.680.973.669.265.9
TEEN[1]ViT-B-CLIPYes75.768.492.088.780.973.769.266.1
FDR(Ours)ViT-B-CLIPYes75.768.592.089.180.974.169.266.4
", + "bbox": [ + 96, + 108, + 901, + 272 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/80f66db2c85eb94da4bf1ba10a3ba77d2c0f120e3cb384b13da4c5fe25e5294f.jpg", + "table_caption": [ + "Table 2. Comparison with existing methods on CUB200-1-shot benchmark (Accuracy %). Bold indicates best performance. $\\Delta$ shows improvement of our method over the previous best. Due to space limitations, only the first three and last three sessions are shown." + ], + "table_footnote": [], + "table_body": "
Session Method01238910
AvgAccAvgAccNAccAvgAccNAccAvgAccNAccAvgAccNAccAvgAccNAccAvgAccNAcc
Baseline [21]80.8574.1814.2568.8618.6665.0018.3952.1218.2251.6420.3049.7519.86
TEEN [1]80.8576.4035.1671.8938.2666.7030.0056.6029.7556.1432.1954.4831.56
FDR(Ours)80.8577.2654.8272.9350.0968.3839.7458.3336.4558.7638.9157.3738.40
Δ-+0.86+19.66+1.04+11.83+1.68+9.74+1.73+6.70+2.62+6.72+2.89+6.84
", + "bbox": [ + 94, + 316, + 901, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 436, + 222, + 452 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This section begins by detailing the datasets and implementation specifics. Subsequently, it presents a comparison with existing approaches to benchmark performance, followed by an ablation study to assess the effectiveness of individual components of our method.", + "bbox": [ + 89, + 460, + 482, + 536 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Datasets and Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 546, + 410, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The evaluations are conducted on four benchmarks, following standard Few-Shot Class-Incremental Learning (FSCIL) protocols: CIFAR100, miniImageNet, CUB200, and IN1K-FSCIL.", + "bbox": [ + 89, + 568, + 482, + 627 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CIFAR100, miniImageNet, and CUB200 are widely used benchmarks for FSCIL. CUB200 contains 200 bird species, while miniImageNet and CIFAR100 each comprise 100 common everyday object categories. This section details the experimental results on CUB200; results for the other two datasets are provided in the supplementary material. We adopt the same dataset split for CUB200 as used in existing works [1, 30, 32], denoted as CUB200-5-shot. Specifically, this follows an 11-session split: Session 0 uses 100 base classes (3,000 images), followed by 10 incremental sessions, each introducing 10 novel classes with 5-shot support per class. To further evaluate the effectiveness of prototype calibration under conditions of extreme data scarcity, we also evaluate a 10-way 1-shot configuration for CUB200, where only one sample per novel class is available. This configuration is denoted as CUB200-1-shot.", + "bbox": [ + 89, + 628, + 482, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "IN1K-FSCIL. Furthermore, to assess scalability in real-world scenarios with more classes, we construct a challeng-", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ing ImageNet1K[11] variant by reorganizing 1,000 classes into 800 base classes (200 images per class) and 200 novel classes divided across 10 incremental sessions (20-way 1-shot), which is denoted as IN1K-FSCIL. This setting introduces higher inter-class density and diversity compared to existing benchmarks, better simulating practical applications requiring large-scale incremental learning. The results are provided in the supplementary material.", + "bbox": [ + 511, + 436, + 903, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation: For our method, all experiments are implemented in PyTorch with CLIP-ViT/B-16 [17] as the backbone. 512-dimensional visual and textual features are extracted from CLIP's aligned embedding space, with the backbone frozen during both the base and incremental phases, making FDR a training-free method. To further bridge the domain gap between CLIP's pretraining data and the target dataset, we apply LoRA fine-tuning on CLIP during the base phase by leveraging an image-text contrastive loss, while refraining from any training in the incremental phase.", + "bbox": [ + 511, + 563, + 903, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics: For evaluation, we adopt the commonly used metric of average accuracy (AvgAcc). Specifically, at session $t$ , the average classification accuracy is computed across all classes encountered up to that session (i.e. $M + t * N$ ). While AvgAcc provides a comprehensive measure of overall performance, it is less sensitive to the accuracy of newly learned novel classes. To address this, we also introduce the average accuracy for novel classes, denoted as NAcc. NAcc is calculated similarly to AvgAcc (i.e., classification over all classes), but the proportion of correctly classified instances is computed exclu", + "bbox": [ + 511, + 734, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3158", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/76a5e151bf3b602fd6fc043a74df7db22c46206dba1052d731b3b03fdf6db414.jpg", + "table_caption": [ + "Table 3. Comparison with existing methods on CUB200-5-shot benchmark (Accuracy %). Bold indicates best performance. $\\Delta$ shows improvement of our method over the previous best. Due to space limitations, only the first three and last three sessions are shown." + ], + "table_footnote": [], + "table_body": "
MethodBackboneAccuracy in each session (%)
01238910
AvgAccAvgAccNAccAvgAccNAccAvgAccNAccAvgAccNAccAvgAccNAccAvgAccNAcc
Baseline[21]ResNet1874.1468.0912.3361.869.5957.628.5645.558.8543.3310.1841.4810.06
TOPIC[22]ResNet1868.6862.49-54.81-49.99-32.22-28.31-26.26-
CEC[30]ResNet1875.8571.94-68.50-63.50-54.83-53.52-52.28-
FACT[32]ResNet1875.9073.23-70.84-66.13-58.41-57.89-56.94-
TEEN[1]ResNet1877.2676.13-72.81-68.16-61.19-60.32-59.31-
Baseline[21]CLIP80.8579.1165.0277.4965.0273.9857.2069.5356.1169.3558.3169.2059.13
TEEN [1]CLIP80.8579.2167.2377.5466.1873.8758.2169.5656.8669.4558.0169.3758.83
FDR(Ours)CLIP80.8579.3873.7677.7070.7374.2862.5170.1359.6870.0861.8170.1562.79
Δ-+0.17+6.53+0.16+4.55+0.30+4.30+0.57+2.82+0.63+3.80+0.78+3.96
", + "bbox": [ + 93, + 121, + 903, + 296 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sively within the novel classes.", + "bbox": [ + 89, + 309, + 295, + 323 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Comparison with Existing Methods", + "text_level": 1, + "bbox": [ + 89, + 334, + 398, + 349 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We have made comparisons against recent SOTA methods [2-6]. As shown in Table 1, our FDR method surpasses all training-free approaches, consistent with the results of the main paper. For approaches that require training[2-6], we apply the CLIP backbone fine-tuned on base classes. The fine-tuning significantly improves the base session accuracy of our method, making it outperform other methods on CIFAR100, CUB200, and IN1K-FSCIL.", + "bbox": [ + 89, + 356, + 482, + 476 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2 and Table 3 present the detailed comparison results on the CUB200 dataset. All methods utilize the same frozen CLIP backbone (without fine-tuning) or ResNet18 to ensure a fair comparison. The baseline represents vanilla Prototypical Networks [21], which directly uses CLIP features without prototype calibration. The recently proposed method, TEEN, demonstrates significantly better performance than the baseline, benefiting from its comprehensive prototype calibration. Furthermore, our method achieves superior performance across all sessions in both 5-shot and 1-shot settings. Notably, in the 1-shot setting, our method achieves a $6.70\\% \\sim 19.66\\%$ improvement in NAcc across all incremental sessions when comparing with holistic calibration method [1], highlighting its effectiveness in bias correction even under extreme data scarcity. Additionally, the progressive increases in AvgAcc, with improvements ranging from $+0.86\\%$ to $+2.89\\%$ , suggest cumulative advantages from our composition strategy as the incremental sessions progress. present the comparison results on the CUB200 dataset. All methods utilize the same frozen CLIP backbone or ResNet18 to ensure a fair comparison. The baseline represents vanilla Prototypical Networks [21], which directly uses CLIP features without prototype calibration. The recently proposed method, TEEN, demonstrates significantly better performance than the baseline, benefiting from its comprehensive prototype calibration. Furthermore, our method achieves superior performance across all sessions in both 5-shot and 1-shot set", + "bbox": [ + 89, + 478, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tings. Notably, in the 1-shot setting, our method achieves a $6.70\\% \\sim 19.66\\%$ improvement in NAcc across all incremental sessions when comparing with holistic calibration method [1], highlighting its effectiveness in bias correction even under extreme data scarcity. Additionally, the progressive increases in AvgAcc, with improvements ranging from $+0.86\\%$ to $+2.89\\%$ , suggest cumulative advantages from our composition strategy as the incremental sessions progress.", + "bbox": [ + 511, + 309, + 906, + 445 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/99b7c41ac559c70868dd9c25ec3691cf02eb5dfa73e14a351b6be6b9d23290a6.jpg", + "table_caption": [ + "Table 4. Ablation study on CUB200-1-shot (Session 10)." + ], + "table_footnote": [], + "table_body": "
MethodAvgAccNAcc
No Calibration49.7519.86
Holistic Calibration only54.4831.56
Random Feature Segments54.8332.08
Naive Keyword Grounding56.6136.25
Full (Ours)57.3738.40
", + "bbox": [ + 529, + 476, + 890, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 588, + 679, + 603 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This section presents ablation experiments on the CUB200-1-shot dataset to quantify the contribution of each component and assess the impact of hyperparameters. The performance is evaluated using the final session's metrics, including the average accuracy (AvgAcc) for all classes and the novel class accuracy (NAcc) for all novel classes.", + "bbox": [ + 511, + 609, + 905, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Component. We dissect the key components of our method through several controlled experiments:", + "bbox": [ + 511, + 700, + 905, + 744 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- No Calibration: This strategy directly aggregates features of each class provided by the CLIP backbone to form prototypes. For novel classes, the prototype is calculated directly from the few-shot samples without any calibration. As a result, this approach serves as a vanilla baseline.", + "- Holistic Calibration: This method corresponds to the TEEN approach [1], which employs holistic calibration for prototypes. It is used as a comparative baseline to dissect and evaluate the effectiveness of the fine-grained calibration proposed in our method." + ], + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3159", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eb124ccf06927eaa0bdf9ae42edcb4994b0dafc92da504fd06769a9eb51aeac3.jpg", + "image_caption": [ + "(a) NAcc w.r.t. Top-q" + ], + "image_footnote": [], + "bbox": [ + 98, + 90, + 359, + 223 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d119e743b9305a427f59f1da63a47cb613759d632644ec853252b86a01cee728.jpg", + "image_caption": [ + "(b) NAcc w.r.t. Top-r", + "Figure 3. Impact of hyperparameters on average accuracy over all novel classes at last session (i.e. NAcc), where $q$ stands for the size of feature segments, $r$ for the number of base-class feature segments for calibrating, and $\\alpha$ for the calibration strength of each feature segment." + ], + "image_footnote": [], + "bbox": [ + 366, + 90, + 627, + 222 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b7e3d5426c9f0ad308bdfdc13f6ada4642169333a992ec3e3636fa183145609a.jpg", + "image_caption": [ + "(c) NAcc w.r.t. $\\alpha$" + ], + "image_footnote": [], + "bbox": [ + 635, + 90, + 895, + 222 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Random Feature Segments: This strategy represents a degraded version of our method, where the feature segments are replaced by random divisions of visual features without any semantic guidance. This approach is used to dissect the necessity of the keyword grounding function.", + "- Naive Keyword Grounding: This strategy also represents a degraded version of our method, where the mapping of each keyword to the feature segment is simplified by multiplying the [EOS] token embedding with the keyword token embedding, rather than incorporating the attention matrix $\\mathbf{A}$ and value matrix $\\mathbf{V}$ . This approach is used to dissect and demonstrate that whether the design of the keyword grounding function is trivial." + ], + "bbox": [ + 89, + 285, + 482, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The evaluation results are shown in Table 4. As can be observed, the $11.70\\%$ NAcc gain of \"Holistic Calibration only\" over \"No Calibration\" confirms that prototype calibration is beneficial. Furthermore, our method outperforms \"Holistic Calibration only\", demonstrating that fine-grained calibration of few-shot class prototypes is both effective and necessary. The inferior performance of \"Random Feature Segments\" and \"Naive Keyword Grounding\" compared to our full method indicates that the effectiveness of fine-grained calibration for few-shot class prototypes relies heavily on its elaborate design. This underscores the value of our carefully designed keyword grounding function, highlighting its critical role in achieving superior performance.", + "bbox": [ + 89, + 486, + 482, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of Hyperparameter. We analyze the impact of three key parameters in controlled experiments:", + "bbox": [ + 89, + 699, + 482, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "- Top- $q$ binarization are conducted on contribution indication matrix $\\mathbf{B}$ along the feature dimensions, where $q$ controls the size of feature segments. From Fig. 3a, we observe that feature segments obtained the best performance with $q = 50$ in NAcc. Smaller values of $q$ (e.g., 10 and 30) result in smaller feature segments, which may be insufficiently captured by the corresponding keywords, leading to suboptimal performance. On the other hand, larger values of $q$ (e.g., 200, 512) degrade performance, as they tend to approach holistic calibration and fail to yield improvements. This indicates that an intermediate granu", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "larity in feature segmentation, which well corresponds to the keyword, is essential.", + "bbox": [ + 524, + 284, + 903, + 314 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Top-r similar prototype segments from base classes are combined to calibrate the novel class, where $r$ controls the number of base-class feature segments for calibrating. In Fig. 3b, we see that increasing the number of similar prototypes (i.e. a larger $r$ ) enhances performance up to $r = 50$ , after which further increases do not yield substantial benefits. This suggests that a small set of highly relevant prototypes is sufficient for effectively calibrating novel class prototypes. Including too many segments may introduce noise or less relevant information, diminishing the calibration quality.", + "- Calibration ratio $\\alpha$ controls the calibration strength of each feature segment. As shown in Fig. 3c, the calibration ratio $\\alpha = 0.5$ yields the best results, balancing the original and calibrated features effectively. Lower values of $\\alpha$ reduce the impact of calibration, while higher values overly prioritize the reference prototypes, leading to suboptimal results." + ], + "bbox": [ + 513, + 315, + 903, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 511, + 625, + 782, + 641 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose a semantic Feature Decomposition-Recomposition method that unlocks the full potential of CLIP's visual-semantic features for few-shot class-incremental learning. Our core innovation lies in decomposing CLIP's frozen features into semantic-aware segments through guidance of textual keywords, making these segments transfer-ready through incremental learning to be selectively recomposed for novel class calibration. By establishing explicit mappings between LLM-generated attribute keywords and their corresponding feature segments, our method enables precise inheritance of visual-semantic characteristics from multiple base classes. The improvement of our method over existing methods on several settings demonstrates its unique capability to mitigate prototype bias while preserving discriminative features.", + "bbox": [ + 511, + 659, + 906, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3160", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 263, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is partially supported by the National Natural Science Foundation of China (No.62495082 and No.62461160331).", + "bbox": [ + 89, + 114, + 485, + 161 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 176, + 187, + 193 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Few-shot class-incremental learning via training-free prototype calibration. In NIPS, 2023. 1, 2, 3, 6, 7", + "[2] Closer: Towards better representation learning for few-shot class-incremental learning. In ECCV, 2024. 6, 7", + "[3] Expandable subspace ensemble for pre-trained model-based class-incremental learning. In CVPR, 2024. 6", + "[4] Learning prompt with distribution-based feature replay for few-shot class-incremental learning. arXiv preprint, 2024. 6", + "[5] Pre-trained vision and language transformers are few-shot incremental learners. In CVPR, 2024. 6", + "[6] Rethinking few-shot class-incremental learning: Learning from yourself. In ECCV, 2024. 6, 7", + "[7] Afra Feyza Akyurek, Ekin Akyurek, Derry Tanti Wijaya, and Jacob Andreas. Subspace regularizers for few-shot class incremental learning. arXiv preprint arXiv:2110.07059, 2021. 1, 2, 3", + "[8] Arslan Chaudhry, Marcus Rohrbach, Mohamed Elhoseiny, Thalaiyasingam Ajanthan, Puneet K Dokania, Philip HS Torr, and Marc'Aurelio Ranzato. On tiny episodic memories in continual learning. arXiv preprint arXiv:1902.10486, 2019. 2", + "[9] Kuilin Chen and Chi-Guhn Lee. Incremental few-shot learning via vector quantization in deep embedded space. In International conference on learning representations, 2021. 1", + "[10] Ali Cheraghian, Shafin Rahman, Pengfei Fang, Soumava Kumar Roy, Lars Petersson, and Mehrtash Harandi. Semantic-aware knowledge distillation for few-shot class-incremental learning. In CVPR, pages 2534-2543, 2021. 1, 2, 3", + "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 6", + "[12] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In CVPR, pages 831-839, 2019. 2", + "[13] David Isele and Akansel Cosgun. Selective experience replay for lifelong learning. In Proceedings of the AAAI conference on artificial intelligence, 2018. 2", + "[14] Gregory Koch, Richard Zemel, Ruslan Salakhutdinov, et al. Siamese neural networks for one-shot image recognition. In ICML deep learning workshop, pages 1-30. Lille, 2015. 2", + "[15] Anna Kukleva, Hilde Kuehne, and Bernt Schiele. Generalized and incremental few-shot learning by explicit learning and calibration without forgetting. In ICCV, 2021. 1", + "[16] Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In Psychology of learning and motivation. Elsevier, 1989. 1, 2" + ], + "bbox": [ + 93, + 203, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning. PmLR, 2021. 6", + "[18] David Rolnick, Arun Ahuja, Jonathan Schwarz, Timothy Lillicrap, and Gregory Wayne. Experience replay for continual learning. NIPS, 2019. 2", + "[19] Guangyuan Shi, Jiaxin Chen, Wenlong Zhang, Li-Ming Zhan, and Xiao-Ming Wu. Overcoming catastrophic forgetting in incremental few-shot learning by finding flat minima. NIPS, 34:6747-6761, 2021. 1", + "[20] Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay. NIPS, 30, 2017. 2", + "[21] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. NeurIPS, 30, 2017. 1, 2, 3, 6, 7", + "[22] Xiaoyu Tao, Xiaopeng Hong, Xinyuan Chang, Songlin Dong, Xing Wei, and Yihong Gong. Few-shot class-incremental learning. In CVPR, 2020. 1, 2, 3, 7", + "[23] Songsong Tian, Lusi Li, Weijun Li, Hang Ran, Xin Ning, and Prayag Tiwari. A survey on few-shot class-incremental learning. Neural Networks, 169, 2024. 1", + "[24] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al. Matching networks for one shot learning. NIPS, 29, 2016. 2", + "[25] Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: Theory, method and application. IEEE TPAMI, 2024. 2", + "[26] Yaqing Wang, Quanming Yao, James T Kwok, and Lionel M Ni. Generalizing from a few examples: A survey on few-shot learning. ACM computing surveys (csur), 53(3), 2020. 1, 2", + "[27] Chenshen Wu, Luis Herranz, Xialei Liu, Joost Van De Weijer, Bogdan Raducanu, et al. Memory replay gans: Learning to generate new categories without forgetting. NIPS, 31, 2018. 2", + "[28] Boyu Yang, Mingbao Lin, Binghao Liu, Mengying Fu, Chang Liu, Rongrong Ji, and Qixiang Ye. Learnable expansion-and-compression network for few-shot class incremental learning. arXiv preprint arXiv:2104.02281, 2021.1", + "[29] Boyu Yang, Mingbao Lin, Yunxiao Zhang, Binghao Liu, Xiaodan Liang, Rongrong Ji, and Qixiang Ye. Dynamic support network for few-shot class incremental learning. IEEE TPAMI, 45(3):2945-2951, 2022. 1", + "[30] Chi Zhang, Nan Song, Guosheng Lin, Yun Zheng, Pan Pan, and Yinghui Xu. Few-shot incremental learning with continually evolved classifiers. In CVPR, pages 12455-12464, 2021. 1, 6, 7", + "[31] Guangtao Zheng and Aidong Zhang. Few-shot class-incremental learning with meta-learned class structures. In 2021 International Conference on Data Mining Workshops (ICDMW), pages 421–430. IEEE, 2021. 1", + "[32] Da-Wei Zhou, Fu-Yun Wang, Han-Jia Ye, Liang Ma, Shiliang Pu, and De-Chuan Zhan. Forward compatible few" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3161", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "shot class-incremental learning. In CVPR, pages 9046-9056, 2022. 1, 6, 7", + "[33] Kai Zhu, Yang Cao, Wei Zhai, Jie Cheng, and Zheng-Jun Zha. Self-promoted prototype refinement for few-shot class incremental learning. In CVPR, pages 6801-6810, 2021. 1" + ], + "bbox": [ + 91, + 90, + 480, + 162 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3162", + "bbox": [ + 482, + 945, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Feature Decomposition-Recomposition in Large Vision-Language Model for Few-Shot Class-Incremental Learning/df51b4e9-12ac-40bf-8801-dbd6097351a5_model.json b/2025/Feature Decomposition-Recomposition in Large Vision-Language Model for Few-Shot Class-Incremental Learning/df51b4e9-12ac-40bf-8801-dbd6097351a5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..73828e2f596f415df278b7fe5b649c7a47ad69f1 --- /dev/null +++ b/2025/Feature Decomposition-Recomposition in Large Vision-Language Model for Few-Shot Class-Incremental Learning/df51b4e9-12ac-40bf-8801-dbd6097351a5_model.json @@ -0,0 +1,1980 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.118, + 0.13, + 0.882, + 0.177 + ], + "angle": 0, + "content": "Feature Decomposition-Recomposition in Large Vision-Language Model for Few-Shot Class-Incremental Learning" + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.203, + 0.757, + 0.221 + ], + "angle": 0, + "content": "Zongyao Xue\\(^{1,2}\\), Meina Kan\\(^{1,2}\\), Shiguang Shan\\(^{1,2,3}\\), Xin Chen\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.221, + 0.862, + 0.24 + ], + "angle": 0, + "content": "\\(^{1}\\)Institute of Computing Technology, Chinese Academy of Sciences, Beijing, 100090, China" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.239, + 0.773, + 0.257 + ], + "angle": 0, + "content": "\\(^{2}\\)University of Chinese Academy of Sciences, Beijing, 100090, China" + }, + { + "type": "text", + "bbox": [ + 0.294, + 0.257, + 0.7, + 0.275 + ], + "angle": 0, + "content": "\\(^{3}\\)Peng Cheng Laboratory, Shenzhen, 518055, China" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.342, + 0.486, + 0.812 + ], + "angle": 0, + "content": "Few-Shot Class-Incremental Learning (FSCIL) focuses on incrementally learning novel classes using only a limited number of samples from novel classes, which faces dual challenges: catastrophic forgetting of previously learned classes and over-fitting to novel classes with few available samples. Recent advances in large pre-trained vision-language models (VLMs), such as CLIP, provide rich feature representations that generalize well across diverse classes. Therefore, freezing the pre-trained backbone and aggregating class features as prototypes becomes an intuitive and effective way to mitigate catastrophic forgetting. However, this strategy fails to address the overfitting challenge, and the prototypes of novel classes exhibit semantic bias due to the few samples per class. To address these limitations, we propose a semantic Feature Decomposition-Recomposition (FDR) method based on VLMs. Firstly, we decompose the CLIP features into semantically distinct segments guided by text keywords from base classes. Then, these segments are adaptively recomposed at the attribute level given text descriptions, forming calibrated prototypes for novel classes. The recomposition process operates linearly at the attribute level but induces nonlinear adjustments across the entire prototype. This fine-grained and non-linear recomposition inherits the generalization capabilities of VLMs and the adaptive recomposition ability of base classes, leading to enhanced performance in FSCIL. Extensive experiments demonstrate our method's effectiveness, particularly in 1-shot scenarios where it achieves improvements between \\(6.70\\%\\) and \\(19.66\\%\\) for novel classes over previous prototype-based methods on CUB200. Code is available at https://github.com/herb1999/FDR." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.844, + 0.222, + 0.86 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.87, + 0.483, + 0.902 + ], + "angle": 0, + "content": "The remarkable success of deep neural networks (DNNs) is largely based on training with large-scale static datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.31, + 0.908, + 0.568 + ], + "angle": 0, + "content": "[23]. However, real-world scenarios such as autonomous driving and robotic field operations often involve continuously arriving data streams, where new categories emerge incrementally with only sparse samples available. Although humans excel at adapting to such dynamic environments, existing learning-based algorithms struggle under these constraints. To bridge this gap, Few-Shot Class-Incremental Learning (FSCIL)[22] is introduced as a more practical and challenging task, aiming to incrementally recognize novel classes from few-shot examples while preserving performance on base classes. The FSCIL task inherits dual challenges from few-shot learning and continual learning. First, models must overcome the unreliable empirical risk minimization problem [26] when generalizing with scarce novel class samples. Second, they must maintain stability against catastrophic forgetting [16] of base classes while adapting to incremental updates." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.909, + 0.719 + ], + "angle": 0, + "content": "Since its introduction in 2020 [22], FSCIL has attracted significant attention due to its potential in real-world applications, leading to a variety of early methodological developments, including meta learning[9, 31, 33], specialized model architectures[28, 29], optimization strategies [19] and replay-based methods[10, 15]. However, these approaches face a longstanding plasticity-stability dilemma, wherein models struggle to simultaneously preserve base class knowledge while effectively accommodating novel class acquisition." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Recently, leveraging the generalizable representations of pre-trained models such as CLIP, prototype-based metric learning approaches [1, 21, 30, 32] have emerged as a prominent solution to mitigate catastrophic forgetting. These methods typically freeze the pre-trained backbone networks and aggregate features from samples within each class to construct prototypes, enabling effective knowledge preservation and incremental learning. However, the scarce samples of novel classes hinder the full exploitation of features from large pre-trained models, yielding semantically biased novel prototypes that deviate from true class distributions. To address these biases, existing methods [1, 7, 10]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3153" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "leverage supplementary information from base class prototypes to calibrate novel ones. Nevertheless, these approaches directly employ entire similar base prototypes to compensate for the biased novel ones, lacking mechanisms to identify which specific components are transferable. This oversight leads to reduced inter-class distances between novel and base classes, causing feature space entanglement." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.199, + 0.487, + 0.56 + ], + "angle": 0, + "content": "To address these limitations, we propose a Feature Decomposition-Recomposition (FDR) framework that enables fine-grained prototype calibration through attribute-level knowledge transfer. Our key insight is that although the feature representations from large pre-trained models contain sufficient information for classifying diverse classes, their correct classification depends on strategically extracting generalizable features, particularly for few-shot classes. Specifically, given features from pre-trained VLMs like CLIP, we first decompose the visual prototypes of base classes into semantic-aware feature segments by establishing mappings between CLIP visual features and attribute keywords through a keyword grounding function. These semantic feature segments are transferable across classes. Next, for novel classes, given a set of keywords describing them, each corresponding feature segment of the biased prototype derived from few-shot samples is calibrated by recombining segments from similar base class prototypes (e.g., calibrating the \"cannon\" prototype by combining \"wheel\" segments from classes like bicycles and \"cylindrical\" segments from classes like lipsticks). The combinatorial possibilities for synthesizing novel prototypes grow exponentially as the number of base classes increases, effectively addressing the prototype entanglement problem." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.561, + 0.405, + 0.574 + ], + "angle": 0, + "content": "The contributions of this work are threefold:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.575, + 0.483, + 0.65 + ], + "angle": 0, + "content": "1. We propose a semantic feature decomposition and recomposition method, FDR, that divides CLIP's frozen features into reusable attribute-level feature segments, enabling precise and easy knowledge transfer from base classes to novel ones for enhanced incremental learning." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.651, + 0.483, + 0.725 + ], + "angle": 0, + "content": "2. We design a keyword grounding function to decompose CLIP's frozen features into semantic-aware segments, and introduce a segment-wise recomposition method to calibrate biased prototypes at the fine-grained attribute-level." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.726, + 0.483, + 0.786 + ], + "angle": 0, + "content": "3. Our approach is training-free, simple yet highly effective, achieving improvements of \\(6.70\\% - 19.66\\%\\) in novel class accuracy over state-of-the-art methods, showcasing its practical utility." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.575, + 0.483, + 0.786 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.8, + 0.233, + 0.815 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Incremental learning (IL), also referred to as continual or lifelong learning, distinguishes itself from conventional machine learning by addressing dynamic data distributions where models must continually integrate new knowledge from sequential tasks while preserving performance on pre" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.304 + ], + "angle": 0, + "content": "viously learned ones [25]. The core challenge lies in balancing learning plasticity for new knowledge and memory stability for old knowledge [16]. An intuitive approach to ensure the stability of old knowledge is to retain representative samples from previous tasks and replay them during training on new tasks[8, 13, 18]. However, growing memory demands and privacy concerns limit their scalability. To enhance storage efficiency, retaining features of old samples and replaying them later is also an option[12]. Another method for preserving old knowledge is generative replay[20, 27], where a generative model learns the distribution of old data and replays generated samples or features. These methods primarily focus on preserving old knowledge rather than leveraging it to enhance new task learning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.308, + 0.905, + 0.563 + ], + "angle": 0, + "content": "Few-Shot Learning (FSL) requires models to learn from a limited number of labeled samples, making it suitable for real-world scenarios where data is scarce due to safety, privacy, or ethical concerns[26]. Metric-based methods [14, 21, 24] have proven to be simple and effective solutions by learning transferable embedding spaces. Prototypical Networks[21], for instance, define a metric space where classification is performed by computing distances to prototype representations of each class. These prototypes, computed as class-mean feature vectors, naturally accommodate incremental updates by freezing the embedding backbone - a property that makes them particularly suitable for few-shot class-incremental learning [22]. However, prototype quality degrades significantly under extreme data scarcity (e.g., 1-shot settings), as limited samples fail to capture true class distributions, leading to biased prototypes that misalign with semantic realities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.57, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Few-Shot Class-Incremental Learning (FSCIL) aims to incrementally learn novel classes given only its few shots[22], which merges challenges from IL and FSL: preventing catastrophic forgetting of base classes while generalizing to novel classes with sparse samples. A common strategy to mitigate this is to use semantically similar base classes to correct biases in novel classes[1, 7, 10]. Additionally, incorporating side information from language has been explored to enhance model performance[7, 10]. The closest to our work is [1], which reduces novel prototype bias by globally calibrating them with similar base prototypes. However, such coarse-grained adjustments compress interclass distances, leading to prototype entanglement in later phases. Our method addresses these limitations through fine-grained calibration, decomposing prototypes into semantically meaningful sub-features for targeted composition. This enables precise inheritance of relevant visual attributes from multiple base classes while maintaining interclass separability. The integration of language semantics also resolves ambiguity in visual similarity measurements, particularly for classes sharing partial attributes rather than global appearances." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3154" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.09, + 0.221, + 0.108 + ], + "angle": 0, + "content": "3. Background" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.485, + 0.176 + ], + "angle": 0, + "content": "Following the existing works [1, 7, 10, 22], the Few-Shot Class-Incremental Learning (FSCIL) task is defined as a process consisting of \\((1 + T)\\) sequential learning sessions, which are divided into two consecutive phases:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.178, + 0.483, + 0.223 + ], + "angle": 0, + "content": "- Base phase (the first session \\( S_0 \\)): during \\( S_0 \\), the model is trained on \\( M \\) base classes \\( \\mathcal{C}^0 = \\{C_1, \\dots, C_M\\} \\) with abundant samples per class." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.223, + 0.484, + 0.329 + ], + "angle": 0, + "content": "- Incremental phase (sessions \\( S_{1} \\sim S_{T} \\)): following the N-way K-shot protocol, at each session \\( S_{t}|_{t=1}^{T} \\), the model learns \\( N \\) novel classes with \\( K \\) samples per class (typically \\( K \\leq 5 \\)). The novel classes are denoted as \\( \\mathcal{C}^t = \\{C_{M^t + 1}, \\dots, C_{M^t + N}\\} \\) with \\( M^t = M + (t - 1) \\times N \\). The label spaces across distinct sessions remain strictly disjoint, i.e. \\( \\mathcal{C}^i \\cap \\mathcal{C}^j = \\emptyset \\quad \\forall i, j \\in \\{0, \\dots, T\\}, i \\neq j \\)." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.178, + 0.484, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.33, + 0.484, + 0.451 + ], + "angle": 0, + "content": "The model in few-shot class-incremental learning is expected to maintain classification accuracy across all base classes while achieving precise classification for novel classes. When leveraging a large pre-trained model as the feature extractor, it can extract rich and informative features for both base and novel classes attributing to its generalization capability. Consequently, obtaining a well-defined prototype is often sufficient for classification." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.452, + 0.484, + 0.527 + ], + "angle": 0, + "content": "For base classes, prototypes can be easily derived by aggregating abundant samples, as demonstrated by existing methods like [21]. Specifically, each class \\(C\\) is represented by a prototype vector \\(\\mathbf{p}_c\\in \\mathbb{R}^d\\), computed as the mean feature of all samples belonging to that class:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.533, + 0.483, + 0.571 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} _ {c} = \\frac {1}{| \\mathcal {I} _ {c} |} \\sum_ {\\left(\\mathbf {x} _ {i}, y _ {i}\\right) \\in \\mathcal {I} _ {c}} f _ {\\theta} \\left(\\mathbf {x} _ {i}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.577, + 0.484, + 0.621 + ], + "angle": 0, + "content": "where \\( f_{\\theta} \\) denotes the frozen feature extractor (large pretrained model like CLIP in this work), and \\( \\mathcal{I}_c \\) contains all available samples for class \\( C \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.622, + 0.484, + 0.712 + ], + "angle": 0, + "content": "For novel classes at session \\( S_{t} \\), the prototypes can also be derived from few-shot samples following Eq. (1). However, these prototypes are often biased due to the limited number of training examples. Such bias can result in suboptimal classification performance, presenting a significant challenge in FSCIL." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.713, + 0.484, + 0.788 + ], + "angle": 0, + "content": "During inference at session \\( S_{t} \\), the classification probability for a test sample \\( \\mathbf{x} \\) is calculated by measuring the distance between its feature representation \\( f_{\\theta}(\\mathbf{x}) \\) and the prototype vectors of all classes encountered (i.e. \\( \\bigcup_{i=0}^{t} \\mathcal{C}^i \\)) so far." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.8, + 0.182, + 0.816 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To mitigate semantic bias in novel-class prototypes derived from scarce samples, we propose a semantically feature decomposition-recomposition approach for calibrating the prototype in a fine-grained manner at keyword attribute level. An overview is illustrated in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.289 + ], + "angle": 0, + "content": "Our proposed method consists of two stages: 1) semantic feature decomposition stage, which establishes a grounding of text attribute keywords to CLIP features using base classes. It produces a set of meaningful feature segments, each corresponding to a specific semantic keyword. These feature segments are transferable across classes, ensuring their help to novel classes. 2) Feature segments recomposition stage, which adaptively calibrates the prototypes of novel classes by recomposing the semantic-aware feature segments according to the keywords associated with the novel classes. This process enables fine-grained refinement of the prototypes at keyword attribute level, mitigating bias and improving classification performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.302, + 0.796, + 0.318 + ], + "angle": 0, + "content": "4.1. Semantic feature decomposition" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.324, + 0.906, + 0.43 + ], + "angle": 0, + "content": "In this stage, for each base class \\( C \\), we begin by generating visual descriptions and attribute keywords using Large Language Models (LLMs). Next, leveraging the vision-language alignment capabilities of CLIP models, a fine-grained mapping from these keywords to visual features is established, producing a set of semantically meaningful feature segments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.906, + 0.581 + ], + "angle": 0, + "content": "Firstly, with the help of LLMs, a textual description \\(\\mathcal{D}_c\\) for each class \\(C\\) is generated conditioned on class names (e.g. \"toaster\"). For instance, \\(\\mathcal{D}_c\\) might be \"A rectangular kitchen appliance with metal or plastic housing, slots for inserting bread, and controls like dials Buttons.\" Additionally, \\(m\\) attribute keywords \\(\\mathcal{K}_C = \\{k_1,\\dots,k_m\\}\\) are extracted from \\(\\mathcal{D}_c\\). Each keyword \\(k_i\\) describes a visual attribute of the class \\(C\\) (e.g. \"rectangular\" and \"slots\" for class \"toaster\"). These textual descriptions and keywords capture the semantic attributes of the class." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.907, + 0.795 + ], + "angle": 0, + "content": "Secondly, we design a keyword grounding function \\(\\mathcal{G}_{\\phi}\\) to establish fine-grained mappings from these keywords to visual feature segments of base prototypes. Leveraging CLIP's frozen visual encoder \\(f_{\\theta}\\) and textual encoder \\(g_{\\phi}\\), we obtain aligned visual prototype \\(\\mathbf{p}_c\\) (same as Eq. (1)) and textual representations \\(\\mathbf{t}_c = g_{\\phi}(\\mathcal{D}_c)\\), where \\(\\mathbf{t}_c \\in \\mathbb{R}^d\\) denotes the text embedding of description \\(\\mathcal{D}_c\\). Since CLIP's text and image embeddings are aligned, mapping a keyword \\(k_i\\) to a segment of the text feature space inherently implies a corresponding mapping to the associated segment in the image feature space. Therefore, for each keyword \\(k_i \\in \\mathcal{K}_C\\), the keyword grounding function \\(\\mathcal{G}_{\\phi}\\) generates a grounding vector \\(\\mathbf{g}_i\\) by evaluating the relevance of the keyword to the entire textual representation:" + }, + { + "type": "equation", + "bbox": [ + 0.615, + 0.809, + 0.907, + 0.827 + ], + "angle": 0, + "content": "\\[\n\\mathbf {g} _ {i} = \\mathcal {G} _ {\\phi} \\left(k _ {i}, \\mathcal {D} _ {c}\\right), \\in \\{0, 1 \\} ^ {d}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Here \\(\\mathbf{g}_i\\) is a sparse binary mask, where entries with a value of 1 indicate the dimensions at which the keyword \\(k_{i}\\) contributes most significantly to the text embedding \\(\\mathbf{t}_c\\) of description \\(\\mathcal{D}_c\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3155" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.09, + 0.9, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.227, + 0.322, + 0.456, + 0.335 + ], + "angle": 0, + "content": "Feature Decomposition of Base classes" + }, + { + "type": "image_caption", + "bbox": [ + 0.589, + 0.322, + 0.896, + 0.335 + ], + "angle": 0, + "content": "Segments Decomposition for calibrating Novel class" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.348, + 0.907, + 0.486 + ], + "angle": 0, + "content": "Figure 1. Overall Architecture of Our Feature Decomposition-Recomposition (FDR) method. The calibration process is as follows: 1) Feature Decomposition of Base classes: For all classes, LLMs are leveraged to generate category descriptions and extract semantic keywords. CLIP encodes both image features (aggregated as prototypes) and textual descriptions. A Keyword Grounding Function then decomposes textual embeddings into semantic-aware feature segments aligned with specific keywords, achieving attribute-level decomposition of visual and textual features. 2) Segments Recomposition for calibrating Novel class: For novel classes, we match each keyword-derived textual feature segment with semantically similar segments from base classes. The corresponding visual feature segments from multiple base prototypes are recomposed to calibrate the raw novel prototype. For instance, when calibrating the cannon prototype (which suffers from missing wheel features due to limited samples), our method identifies wheeled segments from unicycle and oxcart prototypes and cylindrical segments from lipstick and water tower prototypes. This fine-grained recomposition supplements missing attributes while preserving class-discriminative characteristics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.501, + 0.483, + 0.636 + ], + "angle": 0, + "content": "During the text encoding process of CLIP, the embedding of the entire text description \\(\\mathcal{D}_c\\) inherently incorporates the contributions of tokens from each word including those visual keywords \\(k_{i}\\in \\mathcal{K}_{C}\\). This means that the attention mechanisms and other internal structures implicitly reveal how and where each keyword contributes to the overall textual representation. Therefore, the keyword grounding function \\(\\mathcal{G}_{\\phi}\\) is designed to leverage CLIP's internal attention mechanisms to identify these contributions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.637, + 0.483, + 0.725 + ], + "angle": 0, + "content": "Specifically, in the last transformer layer of CLIP textual encoder \\( g_{\\phi} \\), \\( \\mathcal{D}_c \\) is tokenized as \\( n \\) tokens \\( \\{[SOS], \\text{token}_1, \\dots, \\text{token}_{n-2}, [EOS]\\} \\). Given the attention matrix \\( \\mathbf{A} \\in \\mathbb{R}^{n \\times n} \\) and value matrix \\( \\mathbf{V} \\in \\mathbb{R}^{d \\times n} \\), \\( \\mathbf{t}_c \\) is calculated as the output of the last transformer layer as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.739, + 0.483, + 0.755 + ], + "angle": 0, + "content": "\\[\n\\mathbf {t} _ {c} = \\mathbf {V} (:,:) \\mathbf {A} (:, e o s) + \\mathbf {h} = \\left(a _ {1} \\mathbf {v} _ {1} + \\dots + a _ {n} \\mathbf {v} _ {n}\\right) + \\mathbf {h}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.84 + ], + "angle": 0, + "content": "where \\(\\mathbf{A}\\) represents the inter-token attention weights, \\(\\mathbf{V}\\) represents transformed token representations of all words, \\(a_{i} = A[i,p_{\\mathrm{eos}}]\\) represents the attention weight from the \\(i\\)-th token to the [EOS] token, and \\(\\mathbf{v}_i\\) means the value vector of the \\(i\\)-th token from CLIP's final transformer layer." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.484, + 0.9 + ], + "angle": 0, + "content": "As can be seen from the calculation of \\(\\mathbf{t}_c\\) in Eq. (3), the attention matrix \\(\\mathbf{A}\\) and value matrix \\(\\mathbf{V}\\) implicitly reveal how and where each keyword contributes to the overall textual representation. Therefore, we extract these token" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.501, + 0.905, + 0.529 + ], + "angle": 0, + "content": "wise contribution values as a contribution indication matrix \\(\\mathbf{B}\\in \\mathbb{R}^{d\\times n}\\) .." + }, + { + "type": "equation", + "bbox": [ + 0.628, + 0.531, + 0.905, + 0.547 + ], + "angle": 0, + "content": "\\[\n\\mathbf {B} = \\left[ a _ {1} \\mathbf {v} _ {1}, \\dots , a _ {n} \\mathbf {v} _ {n} \\right]. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.552, + 0.906, + 0.627 + ], + "angle": 0, + "content": "Here, the columns of \\(\\mathbf{B}\\) describe different text tokens, while the rows represent feature channels. Each element \\(\\mathbf{B}(j,i)\\) indicates the contribution of the \\(i\\)-th token's \\(j\\)-th feature dimension to the final [EOS] embedding, serving as an effective indicator for keyword-related feature dimensions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.906, + 0.716 + ], + "angle": 0, + "content": "Therefore, to quantify the influence of each token on the text embedding \\(\\mathbf{B}\\), we apply a softmax operation along the token dimension (horizontal) of \\(\\mathbf{B}\\). This results in a matrix \\(\\mathbf{C}\\), where each element represents the proportion of token \\(i\\)'s contribution to \\(\\mathbf{t}_c\\) at feature dimension \\(j\\), relative to all text tokens:" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.724, + 0.905, + 0.74 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C} (i,:) = \\operatorname {S o f t m a x} (\\mathbf {B} (i,:)) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.906, + 0.809 + ], + "angle": 0, + "content": "Then, we perform top-q binarization along the feature dimension (vertical) of \\(\\mathbf{B}\\), converting each column \\(\\mathbf{C}(:, i]\\) into into a sparse binary mask \\(\\mathbf{g}_i\\) (previously introduced in Eq. (2)) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.818, + 0.92, + 0.87 + ], + "angle": 0, + "content": "\\[\n\\mathbf {g} _ {i} (j) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} \\mathbf {C} (i, j) \\text {i s w i t h i n t o p - q r e s p o n s i v e o f} \\mathbf {C} (:, j) \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "This process isolates the most significant feature dimensions associated with each keyword, enabling grounding of" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3156" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.37, + 0.106 + ], + "angle": 0, + "content": "textual keywords to text feature segments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.107, + 0.484, + 0.288 + ], + "angle": 0, + "content": "According to Eqs. (2) to (6), we can derive a fine-grained mapping from each keyword to its corresponding feature segments. While this mapping is computed independently for each class, ideally, the same keyword associated with different classes should yield identical mappings. However, in practice, slight discrepancies often arise due to variations in class contexts. To ensure consistency, for keywords that appear across multiple base classes, we unify all selected feature dimensions by taking the union of their respective binary masks. This strategy preserves all potentially relevant dimensions for the keyword, enhancing robustness and consistency in the grounding process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.288, + 0.484, + 0.349 + ], + "angle": 0, + "content": "Since CLIP's visual and textual feature spaces are aligned, given a keyword \\( k_{i} \\), we can obtain the visual feature segment \\( \\mathbf{s}_i^c \\) of prototype \\( \\mathbf{p}_c \\), by simply performing an element-wise multiplication as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.237, + 0.356, + 0.483, + 0.372 + ], + "angle": 0, + "content": "\\[\n\\mathbf {s} _ {i} ^ {c} = \\mathbf {p} _ {c} \\odot \\mathbf {g} _ {i}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.378, + 0.484, + 0.454 + ], + "angle": 0, + "content": "In this way, we achieve semantic-aware decomposition of base class prototypes guided by textual keywords of visual attributes, where each feature segment \\(\\mathbf{s}_i^c\\) retains only the dimensions most relevant to the corresponding semantic keyword." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.461, + 0.483, + 0.492 + ], + "angle": 0, + "content": "4.2. Feature segments decomposition for fine-grained calibration" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.498, + 0.484, + 0.709 + ], + "angle": 0, + "content": "The prototypes of novel classes are often biased due to the limited number of shots available per class. While the full-dimensional features of these prototypes can only be derived from the novel class samples, certain partial features—such as visual attribute features—frequently overlap with those of base class prototypes. For instance, features like \"wheels\" are shared between the novel class \"cannon\" and base classes \"unicycle\", \"oxcart\", and \"race car\". Leveraging this inherent property, we calibrate the prototypes of novel classes by piecewise refinement of their semantic feature segments. This is achieved by assembling similar segments from the base classes, thereby enhancing the representational accuracy and reducing bias in the novel class prototypes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.709, + 0.483, + 0.755 + ], + "angle": 0, + "content": "In this process, given any novel class \\(\\widetilde{C}\\), we first generate its initial prototype \\(\\mathbf{p}_{\\widetilde{C}}\\) and visual keywords \\(\\kappa_{\\widetilde{C}}\\) following the same process as base classes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.755, + 0.484, + 0.83 + ], + "angle": 0, + "content": "Next, for each keyword \\(k_{i} \\in \\mathcal{K}_{\\widetilde{C}}\\), we retrieve the corresponding feature segments \\(\\mathbf{s}_i^{c_j}\\) from the base-class prototypes that share the same keywords. These segments are then combined using a weighted summation to calibrate the feature segment of the novel class:" + }, + { + "type": "equation", + "bbox": [ + 0.214, + 0.836, + 0.483, + 0.865 + ], + "angle": 0, + "content": "\\[\n\\mathbf {r} _ {i} = \\sum_ {j = 1} ^ {M} w _ {j} \\cdot \\mathbf {s} _ {i} ^ {c _ {j}}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.87, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Here, the \\(\\mathbf{s}_i^{c_j}\\) represents the prototype segment of base class \\(C_j\\) corresponding to the keyword \\(k_{i}\\). The weight \\(w_{j}\\) reflects" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.093, + 0.903, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.384, + 0.907, + 0.427 + ], + "angle": 0, + "content": "Figure 2. Implementation of keyword grounding function \\(\\mathcal{G}_{\\phi}\\). The upper describes the computation of \\(\\mathbf{t}_c\\) using CLIP, while the lower details the process for generating the grounding vector \\(\\mathbf{g}_c\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.459, + 0.906, + 0.596 + ], + "angle": 0, + "content": "the similarity between the feature segments of \\(\\widetilde{C}\\) and base class \\(C_j\\). This similarity is approximated by computing the textual embedding similarity between the textual segment of novel class \\(\\mathbf{t}_{\\widetilde{c}} \\odot \\mathbf{g}_i\\) and that of every base class \\(\\mathbf{t}_{c_j} \\odot \\mathbf{g}_i, C_j \\in C^0\\). The top-r similarity scores are preserved after using a softmax function, while the remaining similarity scores are set to zero. This ensures that only the most semantically relevant base-class feature segments are retained, yielding the final weights \\(w_j\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.906, + 0.66 + ], + "angle": 0, + "content": "After obtaining feature segments for all keywords associated with novel class \\(\\widetilde{C}\\), keyword-specific adjustments are iteratively applied to novel prototype \\(\\mathbf{p}_{\\widetilde{c}}\\) to obtain the final calibrated prototype as below:" + }, + { + "type": "equation", + "bbox": [ + 0.559, + 0.697, + 0.907, + 0.713 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} _ {\\tilde {c}} \\leftarrow (1 - \\alpha) \\cdot \\mathbf {p} _ {\\tilde {c}} + \\alpha \\cdot \\mathbf {r} _ {i}, \\quad i \\in \\{1, \\dots , m \\} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.732, + 0.906, + 0.762 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is a trade-off hyperparameter that balances the prototype directly from the few shots and calibrated one." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.9 + ], + "angle": 0, + "content": "In Eq. (9), each calibration \\(\\mathbf{r}_i\\) is performed individually and is a linear combination of similar feature segments from base-class prototypes. Importantly, this linear combination varies for different feature segments, as the weights and contributing segments depend on the specific keyword \\(k_i\\). Consequently, the overall calibration is semantic-segmentwise linear but nonlinear overall, ensuring fine-grained and more accurate prototype calibration for few-shot novel classes." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3157" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.09, + 0.884, + 0.104 + ], + "angle": 0, + "content": "Table 1. Comparison with existing methods (5-shot setting). \\(A_{base}\\) for base session accuracy, \\(A_{avg}\\) for AvgAcc across all sessions" + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.109, + 0.902, + 0.273 + ], + "angle": 0, + "content": "
MethodBackboneTraining FreeCIFAR100miniImageNetCUB200IN1K-FSCIL
\\(A_{base}\\)\\(A_{avg}\\)\\(A_{base}\\)\\(A_{avg}\\)\\(A_{base}\\)\\(A_{avg}\\)\\(A_{base}\\)\\(A_{avg}\\)
CLOSER[2]ResNet-18/20No75.763.176.062.879.469.1--
Yourself[6]ViT-B-IN21KNo82.967.084.068.883.469.9--
LP-DiF[4]ViT-B-CLIPNo80.275.196.393.883.974.0--
PriViLego[5]ViT-B-CLIPNo84.277.295.492.782.174.272.168.9
EASE[3]ViT-B-IN21KNo85.178.096.893.983.075.173.370.4
FDR(Ours)ViT-B-CLIPNo87.779.395.389.284.477.773.670.9
BaselineViT-B-CLIPYes75.767.992.088.680.973.669.265.9
TEEN[1]ViT-B-CLIPYes75.768.492.088.780.973.769.266.1
FDR(Ours)ViT-B-CLIPYes75.768.592.089.180.974.169.266.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.285, + 0.905, + 0.313 + ], + "angle": 0, + "content": "Table 2. Comparison with existing methods on CUB200-1-shot benchmark (Accuracy %). Bold indicates best performance. \\(\\Delta\\) shows improvement of our method over the previous best. Due to space limitations, only the first three and last three sessions are shown." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.318, + 0.902, + 0.413 + ], + "angle": 0, + "content": "
Session Method01238910
AvgAccAvgAccNAccAvgAccNAccAvgAccNAccAvgAccNAccAvgAccNAccAvgAccNAcc
Baseline [21]80.8574.1814.2568.8618.6665.0018.3952.1218.2251.6420.3049.7519.86
TEEN [1]80.8576.4035.1671.8938.2666.7030.0056.6029.7556.1432.1954.4831.56
FDR(Ours)80.8577.2654.8272.9350.0968.3839.7458.3336.4558.7638.9157.3738.40
Δ-+0.86+19.66+1.04+11.83+1.68+9.74+1.73+6.70+2.62+6.72+2.89+6.84
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.437, + 0.223, + 0.453 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.462, + 0.483, + 0.537 + ], + "angle": 0, + "content": "This section begins by detailing the datasets and implementation specifics. Subsequently, it presents a comparison with existing approaches to benchmark performance, followed by an ablation study to assess the effectiveness of individual components of our method." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.547, + 0.411, + 0.563 + ], + "angle": 0, + "content": "5.1. Datasets and Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.483, + 0.628 + ], + "angle": 0, + "content": "The evaluations are conducted on four benchmarks, following standard Few-Shot Class-Incremental Learning (FSCIL) protocols: CIFAR100, miniImageNet, CUB200, and IN1K-FSCIL." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.483, + 0.87 + ], + "angle": 0, + "content": "CIFAR100, miniImageNet, and CUB200 are widely used benchmarks for FSCIL. CUB200 contains 200 bird species, while miniImageNet and CIFAR100 each comprise 100 common everyday object categories. This section details the experimental results on CUB200; results for the other two datasets are provided in the supplementary material. We adopt the same dataset split for CUB200 as used in existing works [1, 30, 32], denoted as CUB200-5-shot. Specifically, this follows an 11-session split: Session 0 uses 100 base classes (3,000 images), followed by 10 incremental sessions, each introducing 10 novel classes with 5-shot support per class. To further evaluate the effectiveness of prototype calibration under conditions of extreme data scarcity, we also evaluate a 10-way 1-shot configuration for CUB200, where only one sample per novel class is available. This configuration is denoted as CUB200-1-shot." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "IN1K-FSCIL. Furthermore, to assess scalability in real-world scenarios with more classes, we construct a challeng-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.438, + 0.905, + 0.558 + ], + "angle": 0, + "content": "ing ImageNet1K[11] variant by reorganizing 1,000 classes into 800 base classes (200 images per class) and 200 novel classes divided across 10 incremental sessions (20-way 1-shot), which is denoted as IN1K-FSCIL. This setting introduces higher inter-class density and diversity compared to existing benchmarks, better simulating practical applications requiring large-scale incremental learning. The results are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.564, + 0.905, + 0.73 + ], + "angle": 0, + "content": "Implementation: For our method, all experiments are implemented in PyTorch with CLIP-ViT/B-16 [17] as the backbone. 512-dimensional visual and textual features are extracted from CLIP's aligned embedding space, with the backbone frozen during both the base and incremental phases, making FDR a training-free method. To further bridge the domain gap between CLIP's pretraining data and the target dataset, we apply LoRA fine-tuning on CLIP during the base phase by leveraging an image-text contrastive loss, while refraining from any training in the incremental phase." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Evaluation Metrics: For evaluation, we adopt the commonly used metric of average accuracy (AvgAcc). Specifically, at session \\( t \\), the average classification accuracy is computed across all classes encountered up to that session (i.e. \\( M + t * N \\)). While AvgAcc provides a comprehensive measure of overall performance, it is less sensitive to the accuracy of newly learned novel classes. To address this, we also introduce the average accuracy for novel classes, denoted as NAcc. NAcc is calculated similarly to AvgAcc (i.e., classification over all classes), but the proportion of correctly classified instances is computed exclu" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3158" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 3. Comparison with existing methods on CUB200-5-shot benchmark (Accuracy %). Bold indicates best performance. \\(\\Delta\\) shows improvement of our method over the previous best. Due to space limitations, only the first three and last three sessions are shown." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.122, + 0.905, + 0.297 + ], + "angle": 0, + "content": "
MethodBackboneAccuracy in each session (%)
01238910
AvgAccAvgAccNAccAvgAccNAccAvgAccNAccAvgAccNAccAvgAccNAccAvgAccNAcc
Baseline[21]ResNet1874.1468.0912.3361.869.5957.628.5645.558.8543.3310.1841.4810.06
TOPIC[22]ResNet1868.6862.49-54.81-49.99-32.22-28.31-26.26-
CEC[30]ResNet1875.8571.94-68.50-63.50-54.83-53.52-52.28-
FACT[32]ResNet1875.9073.23-70.84-66.13-58.41-57.89-56.94-
TEEN[1]ResNet1877.2676.13-72.81-68.16-61.19-60.32-59.31-
Baseline[21]CLIP80.8579.1165.0277.4965.0273.9857.2069.5356.1169.3558.3169.2059.13
TEEN [1]CLIP80.8579.2167.2377.5466.1873.8758.2169.5656.8669.4558.0169.3758.83
FDR(Ours)CLIP80.8579.3873.7677.7070.7374.2862.5170.1359.6870.0861.8170.1562.79
Δ-+0.17+6.53+0.16+4.55+0.30+4.30+0.57+2.82+0.63+3.80+0.78+3.96
" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.31, + 0.296, + 0.324 + ], + "angle": 0, + "content": "sively within the novel classes." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.335, + 0.4, + 0.351 + ], + "angle": 0, + "content": "5.2. Comparison with Existing Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.357, + 0.483, + 0.477 + ], + "angle": 0, + "content": "We have made comparisons against recent SOTA methods [2-6]. As shown in Table 1, our FDR method surpasses all training-free approaches, consistent with the results of the main paper. For approaches that require training[2-6], we apply the CLIP backbone fine-tuned on base classes. The fine-tuning significantly improves the base session accuracy of our method, making it outperform other methods on CIFAR100, CUB200, and IN1K-FSCIL." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.479, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Table 2 and Table 3 present the detailed comparison results on the CUB200 dataset. All methods utilize the same frozen CLIP backbone (without fine-tuning) or ResNet18 to ensure a fair comparison. The baseline represents vanilla Prototypical Networks [21], which directly uses CLIP features without prototype calibration. The recently proposed method, TEEN, demonstrates significantly better performance than the baseline, benefiting from its comprehensive prototype calibration. Furthermore, our method achieves superior performance across all sessions in both 5-shot and 1-shot settings. Notably, in the 1-shot setting, our method achieves a \\(6.70\\% \\sim 19.66\\%\\) improvement in NAcc across all incremental sessions when comparing with holistic calibration method [1], highlighting its effectiveness in bias correction even under extreme data scarcity. Additionally, the progressive increases in AvgAcc, with improvements ranging from \\(+0.86\\%\\) to \\(+2.89\\%\\), suggest cumulative advantages from our composition strategy as the incremental sessions progress. present the comparison results on the CUB200 dataset. All methods utilize the same frozen CLIP backbone or ResNet18 to ensure a fair comparison. The baseline represents vanilla Prototypical Networks [21], which directly uses CLIP features without prototype calibration. The recently proposed method, TEEN, demonstrates significantly better performance than the baseline, benefiting from its comprehensive prototype calibration. Furthermore, our method achieves superior performance across all sessions in both 5-shot and 1-shot set" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.31, + 0.907, + 0.446 + ], + "angle": 0, + "content": "tings. Notably, in the 1-shot setting, our method achieves a \\(6.70\\% \\sim 19.66\\%\\) improvement in NAcc across all incremental sessions when comparing with holistic calibration method [1], highlighting its effectiveness in bias correction even under extreme data scarcity. Additionally, the progressive increases in AvgAcc, with improvements ranging from \\(+0.86\\%\\) to \\(+2.89\\%\\), suggest cumulative advantages from our composition strategy as the incremental sessions progress." + }, + { + "type": "table_caption", + "bbox": [ + 0.54, + 0.459, + 0.879, + 0.473 + ], + "angle": 0, + "content": "Table 4. Ablation study on CUB200-1-shot (Session 10)." + }, + { + "type": "table", + "bbox": [ + 0.53, + 0.477, + 0.891, + 0.577 + ], + "angle": 0, + "content": "
MethodAvgAccNAcc
No Calibration49.7519.86
Holistic Calibration only54.4831.56
Random Feature Segments54.8332.08
Naive Keyword Grounding56.6136.25
Full (Ours)57.3738.40
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.589, + 0.68, + 0.604 + ], + "angle": 0, + "content": "5.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.611, + 0.906, + 0.701 + ], + "angle": 0, + "content": "This section presents ablation experiments on the CUB200-1-shot dataset to quantify the contribution of each component and assess the impact of hyperparameters. The performance is evaluated using the final session's metrics, including the average accuracy (AvgAcc) for all classes and the novel class accuracy (NAcc) for all novel classes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.702, + 0.906, + 0.746 + ], + "angle": 0, + "content": "Effectiveness of Component. We dissect the key components of our method through several controlled experiments:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.75, + 0.907, + 0.826 + ], + "angle": 0, + "content": "- No Calibration: This strategy directly aggregates features of each class provided by the CLIP backbone to form prototypes. For novel classes, the prototype is calculated directly from the few-shot samples without any calibration. As a result, this approach serves as a vanilla baseline." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.826, + 0.907, + 0.901 + ], + "angle": 0, + "content": "- Holistic Calibration: This method corresponds to the TEEN approach [1], which employs holistic calibration for prototypes. It is used as a comparative baseline to dissect and evaluate the effectiveness of the fine-grained calibration proposed in our method." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3159" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.092, + 0.36, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.171, + 0.231, + 0.286, + 0.243 + ], + "angle": 0, + "content": "(a) NAcc w.r.t. Top-q" + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.092, + 0.628, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.441, + 0.231, + 0.553, + 0.243 + ], + "angle": 0, + "content": "(b) NAcc w.r.t. Top-r" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.092, + 0.897, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.721, + 0.231, + 0.812, + 0.242 + ], + "angle": 0, + "content": "(c) NAcc w.r.t. \\(\\alpha\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.244, + 0.907, + 0.271 + ], + "angle": 0, + "content": "Figure 3. Impact of hyperparameters on average accuracy over all novel classes at last session (i.e. NAcc), where \\( q \\) stands for the size of feature segments, \\( r \\) for the number of base-class feature segments for calibrating, and \\( \\alpha \\) for the calibration strength of each feature segment." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.286, + 0.483, + 0.36 + ], + "angle": 0, + "content": "- Random Feature Segments: This strategy represents a degraded version of our method, where the feature segments are replaced by random divisions of visual features without any semantic guidance. This approach is used to dissect the necessity of the keyword grounding function." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.362, + 0.483, + 0.481 + ], + "angle": 0, + "content": "- Naive Keyword Grounding: This strategy also represents a degraded version of our method, where the mapping of each keyword to the feature segment is simplified by multiplying the [EOS] token embedding with the keyword token embedding, rather than incorporating the attention matrix \\(\\mathbf{A}\\) and value matrix \\(\\mathbf{V}\\). This approach is used to dissect and demonstrate that whether the design of the keyword grounding function is trivial." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.286, + 0.483, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.487, + 0.483, + 0.697 + ], + "angle": 0, + "content": "The evaluation results are shown in Table 4. As can be observed, the \\(11.70\\%\\) NAcc gain of \"Holistic Calibration only\" over \"No Calibration\" confirms that prototype calibration is beneficial. Furthermore, our method outperforms \"Holistic Calibration only\", demonstrating that fine-grained calibration of few-shot class prototypes is both effective and necessary. The inferior performance of \"Random Feature Segments\" and \"Naive Keyword Grounding\" compared to our full method indicates that the effectiveness of fine-grained calibration for few-shot class prototypes relies heavily on its elaborate design. This underscores the value of our carefully designed keyword grounding function, highlighting its critical role in achieving superior performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.7, + 0.483, + 0.731 + ], + "angle": 0, + "content": "Impact of Hyperparameter. We analyze the impact of three key parameters in controlled experiments:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- Top-\\(q\\) binarization are conducted on contribution indication matrix \\(\\mathbf{B}\\) along the feature dimensions, where \\(q\\) controls the size of feature segments. From Fig. 3a, we observe that feature segments obtained the best performance with \\(q = 50\\) in NAcc. Smaller values of \\(q\\) (e.g., 10 and 30) result in smaller feature segments, which may be insufficiently captured by the corresponding keywords, leading to suboptimal performance. On the other hand, larger values of \\(q\\) (e.g., 200, 512) degrade performance, as they tend to approach holistic calibration and fail to yield improvements. This indicates that an intermediate granu" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.285, + 0.905, + 0.315 + ], + "angle": 0, + "content": "larity in feature segmentation, which well corresponds to the keyword, is essential." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.316, + 0.905, + 0.481 + ], + "angle": 0, + "content": "- Top-r similar prototype segments from base classes are combined to calibrate the novel class, where \\( r \\) controls the number of base-class feature segments for calibrating. In Fig. 3b, we see that increasing the number of similar prototypes (i.e. a larger \\( r \\)) enhances performance up to \\( r = 50 \\), after which further increases do not yield substantial benefits. This suggests that a small set of highly relevant prototypes is sufficient for effectively calibrating novel class prototypes. Including too many segments may introduce noise or less relevant information, diminishing the calibration quality." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.482, + 0.905, + 0.587 + ], + "angle": 0, + "content": "- Calibration ratio \\(\\alpha\\) controls the calibration strength of each feature segment. As shown in Fig. 3c, the calibration ratio \\(\\alpha = 0.5\\) yields the best results, balancing the original and calibrated features effectively. Lower values of \\(\\alpha\\) reduce the impact of calibration, while higher values overly prioritize the reference prototypes, leading to suboptimal results." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.316, + 0.905, + 0.587 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.626, + 0.783, + 0.642 + ], + "angle": 0, + "content": "6. Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.907, + 0.9 + ], + "angle": 0, + "content": "In this work, we propose a semantic Feature Decomposition-Recomposition method that unlocks the full potential of CLIP's visual-semantic features for few-shot class-incremental learning. Our core innovation lies in decomposing CLIP's frozen features into semantic-aware segments through guidance of textual keywords, making these segments transfer-ready through incremental learning to be selectively recomposed for novel class calibration. By establishing explicit mappings between LLM-generated attribute keywords and their corresponding feature segments, our method enables precise inheritance of visual-semantic characteristics from multiple base classes. The improvement of our method over existing methods on several settings demonstrates its unique capability to mitigate prototype bias while preserving discriminative features." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3160" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.264, + 0.108 + ], + "angle": 0, + "content": "7. Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.486, + 0.162 + ], + "angle": 0, + "content": "This work is partially supported by the National Natural Science Foundation of China (No.62495082 and No.62461160331)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.178, + 0.188, + 0.194 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.204, + 0.483, + 0.231 + ], + "angle": 0, + "content": "[1] Few-shot class-incremental learning via training-free prototype calibration. In NIPS, 2023. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.233, + 0.484, + 0.261 + ], + "angle": 0, + "content": "[2] Closer: Towards better representation learning for few-shot class-incremental learning. In ECCV, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.264, + 0.483, + 0.29 + ], + "angle": 0, + "content": "[3] Expandable subspace ensemble for pre-trained model-based class-incremental learning. In CVPR, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.294, + 0.483, + 0.321 + ], + "angle": 0, + "content": "[4] Learning prompt with distribution-based feature replay for few-shot class-incremental learning. arXiv preprint, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.324, + 0.483, + 0.35 + ], + "angle": 0, + "content": "[5] Pre-trained vision and language transformers are few-shot incremental learners. In CVPR, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.353, + 0.483, + 0.38 + ], + "angle": 0, + "content": "[6] Rethinking few-shot class-incremental learning: Learning from yourself. In ECCV, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.383, + 0.483, + 0.436 + ], + "angle": 0, + "content": "[7] Afra Feyza Akyurek, Ekin Akyurek, Derry Tanti Wijaya, and Jacob Andreas. Subspace regularizers for few-shot class incremental learning. arXiv preprint arXiv:2110.07059, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.441, + 0.483, + 0.508 + ], + "angle": 0, + "content": "[8] Arslan Chaudhry, Marcus Rohrbach, Mohamed Elhoseiny, Thalaiyasingam Ajanthan, Puneet K Dokania, Philip HS Torr, and Marc'Aurelio Ranzato. On tiny episodic memories in continual learning. arXiv preprint arXiv:1902.10486, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.512, + 0.483, + 0.553 + ], + "angle": 0, + "content": "[9] Kuilin Chen and Chi-Guhn Lee. Incremental few-shot learning via vector quantization in deep embedded space. In International conference on learning representations, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.556, + 0.483, + 0.623 + ], + "angle": 0, + "content": "[10] Ali Cheraghian, Shafin Rahman, Pengfei Fang, Soumava Kumar Roy, Lars Petersson, and Mehrtash Harandi. Semantic-aware knowledge distillation for few-shot class-incremental learning. In CVPR, pages 2534-2543, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.627, + 0.483, + 0.667 + ], + "angle": 0, + "content": "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.671, + 0.483, + 0.712 + ], + "angle": 0, + "content": "[12] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In CVPR, pages 831-839, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.715, + 0.483, + 0.756 + ], + "angle": 0, + "content": "[13] David Isele and Akansel Cosgun. Selective experience replay for lifelong learning. In Proceedings of the AAAI conference on artificial intelligence, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.759, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[14] Gregory Koch, Richard Zemel, Ruslan Salakhutdinov, et al. Siamese neural networks for one-shot image recognition. In ICML deep learning workshop, pages 1-30. Lille, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[15] Anna Kukleva, Hilde Kuehne, and Bernt Schiele. Generalized and incremental few-shot learning by explicit learning and calibration without forgetting. In ICCV, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[16] Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In Psychology of learning and motivation. Elsevier, 1989. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.204, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.174 + ], + "angle": 0, + "content": "[17] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning. PmLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[18] David Rolnick, Arun Ahuja, Jonathan Schwarz, Timothy Lillicrap, and Gregory Wayne. Experience replay for continual learning. NIPS, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.221, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[19] Guangyuan Shi, Jiaxin Chen, Wenlong Zhang, Li-Ming Zhan, and Xiao-Ming Wu. Overcoming catastrophic forgetting in incremental few-shot learning by finding flat minima. NIPS, 34:6747-6761, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.904, + 0.316 + ], + "angle": 0, + "content": "[20] Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay. NIPS, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.359 + ], + "angle": 0, + "content": "[21] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. NeurIPS, 30, 2017. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.904, + 0.403 + ], + "angle": 0, + "content": "[22] Xiaoyu Tao, Xiaopeng Hong, Xinyuan Chang, Songlin Dong, Xing Wei, and Yihong Gong. Few-shot class-incremental learning. In CVPR, 2020. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[23] Songsong Tian, Lusi Li, Weijun Li, Hang Ran, Xin Ning, and Prayag Tiwari. A survey on few-shot class-incremental learning. Neural Networks, 169, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.904, + 0.488 + ], + "angle": 0, + "content": "[24] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al. Matching networks for one shot learning. NIPS, 29, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[25] Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: Theory, method and application. IEEE TPAMI, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.534, + 0.905, + 0.575 + ], + "angle": 0, + "content": "[26] Yaqing Wang, Quanming Yao, James T Kwok, and Lionel M Ni. Generalizing from a few examples: A survey on few-shot learning. ACM computing surveys (csur), 53(3), 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.577, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[27] Chenshen Wu, Luis Herranz, Xialei Liu, Joost Van De Weijer, Bogdan Raducanu, et al. Memory replay gans: Learning to generate new categories without forgetting. NIPS, 31, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.633, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[28] Boyu Yang, Mingbao Lin, Binghao Liu, Mengying Fu, Chang Liu, Rongrong Ji, and Qixiang Ye. Learnable expansion-and-compression network for few-shot class incremental learning. arXiv preprint arXiv:2104.02281, 2021.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.704, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[29] Boyu Yang, Mingbao Lin, Yunxiao Zhang, Binghao Liu, Xiaodan Liang, Rongrong Ji, and Qixiang Ye. Dynamic support network for few-shot class incremental learning. IEEE TPAMI, 45(3):2945-2951, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[30] Chi Zhang, Nan Song, Guosheng Lin, Yun Zheng, Pan Pan, and Yinghui Xu. Few-shot incremental learning with continually evolved classifiers. In CVPR, pages 12455-12464, 2021. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[31] Guangtao Zheng and Aidong Zhang. Few-shot class-incremental learning with meta-learned class structures. In 2021 International Conference on Data Mining Workshops (ICDMW), pages 421–430. IEEE, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[32] Da-Wei Zhou, Fu-Yun Wang, Han-Jia Ye, Liang Ma, Shiliang Pu, and De-Chuan Zhan. Forward compatible few" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3161" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "shot class-incremental learning. In CVPR, pages 9046-9056, 2022. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.482, + 0.163 + ], + "angle": 0, + "content": "[33] Kai Zhu, Yang Cao, Wei Zhai, Jie Cheng, and Zheng-Jun Zha. Self-promoted prototype refinement for few-shot class incremental learning. In CVPR, pages 6801-6810, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.517, + 0.956 + ], + "angle": 0, + "content": "3162" + } + ] +] \ No newline at end of file diff --git a/2025/Feature Decomposition-Recomposition in Large Vision-Language Model for Few-Shot Class-Incremental Learning/df51b4e9-12ac-40bf-8801-dbd6097351a5_origin.pdf b/2025/Feature Decomposition-Recomposition in Large Vision-Language Model for Few-Shot Class-Incremental Learning/df51b4e9-12ac-40bf-8801-dbd6097351a5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bd53f9e13432c43fc3da9276d68fdf0375e38832 --- /dev/null +++ b/2025/Feature Decomposition-Recomposition in Large Vision-Language Model for Few-Shot Class-Incremental Learning/df51b4e9-12ac-40bf-8801-dbd6097351a5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cdc3458c5f78dc0aa9f6068bef2dbbe8d308a54a31cc1f68cb94be969a366fb3 +size 1650860 diff --git a/2025/Feature Decomposition-Recomposition in Large Vision-Language Model for Few-Shot Class-Incremental Learning/full.md b/2025/Feature Decomposition-Recomposition in Large Vision-Language Model for Few-Shot Class-Incremental Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6b94b72959b94622c42939747ac4d67d4bf109a5 --- /dev/null +++ b/2025/Feature Decomposition-Recomposition in Large Vision-Language Model for Few-Shot Class-Incremental Learning/full.md @@ -0,0 +1,288 @@ +# Feature Decomposition-Recomposition in Large Vision-Language Model for Few-Shot Class-Incremental Learning + +Zongyao Xue $^{1,2}$ , Meina Kan $^{1,2}$ , Shiguang Shan $^{1,2,3}$ , Xin Chen $^{1,2}$ + +$^{1}$ Institute of Computing Technology, Chinese Academy of Sciences, Beijing, 100090, China + +$^{2}$ University of Chinese Academy of Sciences, Beijing, 100090, China + +$^{3}$ Peng Cheng Laboratory, Shenzhen, 518055, China + +# Abstract + +Few-Shot Class-Incremental Learning (FSCIL) focuses on incrementally learning novel classes using only a limited number of samples from novel classes, which faces dual challenges: catastrophic forgetting of previously learned classes and over-fitting to novel classes with few available samples. Recent advances in large pre-trained vision-language models (VLMs), such as CLIP, provide rich feature representations that generalize well across diverse classes. Therefore, freezing the pre-trained backbone and aggregating class features as prototypes becomes an intuitive and effective way to mitigate catastrophic forgetting. However, this strategy fails to address the overfitting challenge, and the prototypes of novel classes exhibit semantic bias due to the few samples per class. To address these limitations, we propose a semantic Feature Decomposition-Recomposition (FDR) method based on VLMs. Firstly, we decompose the CLIP features into semantically distinct segments guided by text keywords from base classes. Then, these segments are adaptively recomposed at the attribute level given text descriptions, forming calibrated prototypes for novel classes. The recomposition process operates linearly at the attribute level but induces nonlinear adjustments across the entire prototype. This fine-grained and non-linear recomposition inherits the generalization capabilities of VLMs and the adaptive recomposition ability of base classes, leading to enhanced performance in FSCIL. Extensive experiments demonstrate our method's effectiveness, particularly in 1-shot scenarios where it achieves improvements between $6.70\%$ and $19.66\%$ for novel classes over previous prototype-based methods on CUB200. Code is available at https://github.com/herb1999/FDR. + +# 1. Introduction + +The remarkable success of deep neural networks (DNNs) is largely based on training with large-scale static datasets + +[23]. However, real-world scenarios such as autonomous driving and robotic field operations often involve continuously arriving data streams, where new categories emerge incrementally with only sparse samples available. Although humans excel at adapting to such dynamic environments, existing learning-based algorithms struggle under these constraints. To bridge this gap, Few-Shot Class-Incremental Learning (FSCIL)[22] is introduced as a more practical and challenging task, aiming to incrementally recognize novel classes from few-shot examples while preserving performance on base classes. The FSCIL task inherits dual challenges from few-shot learning and continual learning. First, models must overcome the unreliable empirical risk minimization problem [26] when generalizing with scarce novel class samples. Second, they must maintain stability against catastrophic forgetting [16] of base classes while adapting to incremental updates. + +Since its introduction in 2020 [22], FSCIL has attracted significant attention due to its potential in real-world applications, leading to a variety of early methodological developments, including meta learning[9, 31, 33], specialized model architectures[28, 29], optimization strategies [19] and replay-based methods[10, 15]. However, these approaches face a longstanding plasticity-stability dilemma, wherein models struggle to simultaneously preserve base class knowledge while effectively accommodating novel class acquisition. + +Recently, leveraging the generalizable representations of pre-trained models such as CLIP, prototype-based metric learning approaches [1, 21, 30, 32] have emerged as a prominent solution to mitigate catastrophic forgetting. These methods typically freeze the pre-trained backbone networks and aggregate features from samples within each class to construct prototypes, enabling effective knowledge preservation and incremental learning. However, the scarce samples of novel classes hinder the full exploitation of features from large pre-trained models, yielding semantically biased novel prototypes that deviate from true class distributions. To address these biases, existing methods [1, 7, 10] + +leverage supplementary information from base class prototypes to calibrate novel ones. Nevertheless, these approaches directly employ entire similar base prototypes to compensate for the biased novel ones, lacking mechanisms to identify which specific components are transferable. This oversight leads to reduced inter-class distances between novel and base classes, causing feature space entanglement. + +To address these limitations, we propose a Feature Decomposition-Recomposition (FDR) framework that enables fine-grained prototype calibration through attribute-level knowledge transfer. Our key insight is that although the feature representations from large pre-trained models contain sufficient information for classifying diverse classes, their correct classification depends on strategically extracting generalizable features, particularly for few-shot classes. Specifically, given features from pre-trained VLMs like CLIP, we first decompose the visual prototypes of base classes into semantic-aware feature segments by establishing mappings between CLIP visual features and attribute keywords through a keyword grounding function. These semantic feature segments are transferable across classes. Next, for novel classes, given a set of keywords describing them, each corresponding feature segment of the biased prototype derived from few-shot samples is calibrated by recombining segments from similar base class prototypes (e.g., calibrating the "cannon" prototype by combining "wheel" segments from classes like bicycles and "cylindrical" segments from classes like lipsticks). The combinatorial possibilities for synthesizing novel prototypes grow exponentially as the number of base classes increases, effectively addressing the prototype entanglement problem. + +The contributions of this work are threefold: + +1. We propose a semantic feature decomposition and recomposition method, FDR, that divides CLIP's frozen features into reusable attribute-level feature segments, enabling precise and easy knowledge transfer from base classes to novel ones for enhanced incremental learning. +2. We design a keyword grounding function to decompose CLIP's frozen features into semantic-aware segments, and introduce a segment-wise recomposition method to calibrate biased prototypes at the fine-grained attribute-level. +3. Our approach is training-free, simple yet highly effective, achieving improvements of $6.70\% - 19.66\%$ in novel class accuracy over state-of-the-art methods, showcasing its practical utility. + +# 2. Related Work + +Incremental learning (IL), also referred to as continual or lifelong learning, distinguishes itself from conventional machine learning by addressing dynamic data distributions where models must continually integrate new knowledge from sequential tasks while preserving performance on pre + +viously learned ones [25]. The core challenge lies in balancing learning plasticity for new knowledge and memory stability for old knowledge [16]. An intuitive approach to ensure the stability of old knowledge is to retain representative samples from previous tasks and replay them during training on new tasks[8, 13, 18]. However, growing memory demands and privacy concerns limit their scalability. To enhance storage efficiency, retaining features of old samples and replaying them later is also an option[12]. Another method for preserving old knowledge is generative replay[20, 27], where a generative model learns the distribution of old data and replays generated samples or features. These methods primarily focus on preserving old knowledge rather than leveraging it to enhance new task learning. + +Few-Shot Learning (FSL) requires models to learn from a limited number of labeled samples, making it suitable for real-world scenarios where data is scarce due to safety, privacy, or ethical concerns[26]. Metric-based methods [14, 21, 24] have proven to be simple and effective solutions by learning transferable embedding spaces. Prototypical Networks[21], for instance, define a metric space where classification is performed by computing distances to prototype representations of each class. These prototypes, computed as class-mean feature vectors, naturally accommodate incremental updates by freezing the embedding backbone - a property that makes them particularly suitable for few-shot class-incremental learning [22]. However, prototype quality degrades significantly under extreme data scarcity (e.g., 1-shot settings), as limited samples fail to capture true class distributions, leading to biased prototypes that misalign with semantic realities. + +Few-Shot Class-Incremental Learning (FSCIL) aims to incrementally learn novel classes given only its few shots[22], which merges challenges from IL and FSL: preventing catastrophic forgetting of base classes while generalizing to novel classes with sparse samples. A common strategy to mitigate this is to use semantically similar base classes to correct biases in novel classes[1, 7, 10]. Additionally, incorporating side information from language has been explored to enhance model performance[7, 10]. The closest to our work is [1], which reduces novel prototype bias by globally calibrating them with similar base prototypes. However, such coarse-grained adjustments compress interclass distances, leading to prototype entanglement in later phases. Our method addresses these limitations through fine-grained calibration, decomposing prototypes into semantically meaningful sub-features for targeted composition. This enables precise inheritance of relevant visual attributes from multiple base classes while maintaining interclass separability. The integration of language semantics also resolves ambiguity in visual similarity measurements, particularly for classes sharing partial attributes rather than global appearances. + +# 3. Background + +Following the existing works [1, 7, 10, 22], the Few-Shot Class-Incremental Learning (FSCIL) task is defined as a process consisting of $(1 + T)$ sequential learning sessions, which are divided into two consecutive phases: + +- Base phase (the first session $S_0$ ): during $S_0$ , the model is trained on $M$ base classes $\mathcal{C}^0 = \{C_1, \dots, C_M\}$ with abundant samples per class. +- Incremental phase (sessions $S_{1} \sim S_{T}$ ): following the N-way K-shot protocol, at each session $S_{t}|_{t=1}^{T}$ , the model learns $N$ novel classes with $K$ samples per class (typically $K \leq 5$ ). The novel classes are denoted as $\mathcal{C}^t = \{C_{M^t + 1}, \dots, C_{M^t + N}\}$ with $M^t = M + (t - 1) \times N$ . The label spaces across distinct sessions remain strictly disjoint, i.e. $\mathcal{C}^i \cap \mathcal{C}^j = \emptyset \quad \forall i, j \in \{0, \dots, T\}, i \neq j$ . + +The model in few-shot class-incremental learning is expected to maintain classification accuracy across all base classes while achieving precise classification for novel classes. When leveraging a large pre-trained model as the feature extractor, it can extract rich and informative features for both base and novel classes attributing to its generalization capability. Consequently, obtaining a well-defined prototype is often sufficient for classification. + +For base classes, prototypes can be easily derived by aggregating abundant samples, as demonstrated by existing methods like [21]. Specifically, each class $C$ is represented by a prototype vector $\mathbf{p}_c\in \mathbb{R}^d$ , computed as the mean feature of all samples belonging to that class: + +$$ +\mathbf {p} _ {c} = \frac {1}{| \mathcal {I} _ {c} |} \sum_ {\left(\mathbf {x} _ {i}, y _ {i}\right) \in \mathcal {I} _ {c}} f _ {\theta} \left(\mathbf {x} _ {i}\right), \tag {1} +$$ + +where $f_{\theta}$ denotes the frozen feature extractor (large pretrained model like CLIP in this work), and $\mathcal{I}_c$ contains all available samples for class $C$ . + +For novel classes at session $S_{t}$ , the prototypes can also be derived from few-shot samples following Eq. (1). However, these prototypes are often biased due to the limited number of training examples. Such bias can result in suboptimal classification performance, presenting a significant challenge in FSCIL. + +During inference at session $S_{t}$ , the classification probability for a test sample $\mathbf{x}$ is calculated by measuring the distance between its feature representation $f_{\theta}(\mathbf{x})$ and the prototype vectors of all classes encountered (i.e. $\bigcup_{i=0}^{t} \mathcal{C}^i$ ) so far. + +# 4. Method + +To mitigate semantic bias in novel-class prototypes derived from scarce samples, we propose a semantically feature decomposition-recomposition approach for calibrating the prototype in a fine-grained manner at keyword attribute level. An overview is illustrated in Fig. 1. + +Our proposed method consists of two stages: 1) semantic feature decomposition stage, which establishes a grounding of text attribute keywords to CLIP features using base classes. It produces a set of meaningful feature segments, each corresponding to a specific semantic keyword. These feature segments are transferable across classes, ensuring their help to novel classes. 2) Feature segments recomposition stage, which adaptively calibrates the prototypes of novel classes by recomposing the semantic-aware feature segments according to the keywords associated with the novel classes. This process enables fine-grained refinement of the prototypes at keyword attribute level, mitigating bias and improving classification performance. + +# 4.1. Semantic feature decomposition + +In this stage, for each base class $C$ , we begin by generating visual descriptions and attribute keywords using Large Language Models (LLMs). Next, leveraging the vision-language alignment capabilities of CLIP models, a fine-grained mapping from these keywords to visual features is established, producing a set of semantically meaningful feature segments. + +Firstly, with the help of LLMs, a textual description $\mathcal{D}_c$ for each class $C$ is generated conditioned on class names (e.g. "toaster"). For instance, $\mathcal{D}_c$ might be "A rectangular kitchen appliance with metal or plastic housing, slots for inserting bread, and controls like dials Buttons." Additionally, $m$ attribute keywords $\mathcal{K}_C = \{k_1,\dots,k_m\}$ are extracted from $\mathcal{D}_c$ . Each keyword $k_i$ describes a visual attribute of the class $C$ (e.g. "rectangular" and "slots" for class "toaster"). These textual descriptions and keywords capture the semantic attributes of the class. + +Secondly, we design a keyword grounding function $\mathcal{G}_{\phi}$ to establish fine-grained mappings from these keywords to visual feature segments of base prototypes. Leveraging CLIP's frozen visual encoder $f_{\theta}$ and textual encoder $g_{\phi}$ , we obtain aligned visual prototype $\mathbf{p}_c$ (same as Eq. (1)) and textual representations $\mathbf{t}_c = g_{\phi}(\mathcal{D}_c)$ , where $\mathbf{t}_c \in \mathbb{R}^d$ denotes the text embedding of description $\mathcal{D}_c$ . Since CLIP's text and image embeddings are aligned, mapping a keyword $k_i$ to a segment of the text feature space inherently implies a corresponding mapping to the associated segment in the image feature space. Therefore, for each keyword $k_i \in \mathcal{K}_C$ , the keyword grounding function $\mathcal{G}_{\phi}$ generates a grounding vector $\mathbf{g}_i$ by evaluating the relevance of the keyword to the entire textual representation: + +$$ +\mathbf {g} _ {i} = \mathcal {G} _ {\phi} \left(k _ {i}, \mathcal {D} _ {c}\right), \in \{0, 1 \} ^ {d}. \tag {2} +$$ + +Here $\mathbf{g}_i$ is a sparse binary mask, where entries with a value of 1 indicate the dimensions at which the keyword $k_{i}$ contributes most significantly to the text embedding $\mathbf{t}_c$ of description $\mathcal{D}_c$ . + +![](images/9c6130e2000b378c177614f4f80edfe524f8aec6579e878daf9c0a57c951b6e5.jpg) +Feature Decomposition of Base classes +Segments Decomposition for calibrating Novel class +Figure 1. Overall Architecture of Our Feature Decomposition-Recomposition (FDR) method. The calibration process is as follows: 1) Feature Decomposition of Base classes: For all classes, LLMs are leveraged to generate category descriptions and extract semantic keywords. CLIP encodes both image features (aggregated as prototypes) and textual descriptions. A Keyword Grounding Function then decomposes textual embeddings into semantic-aware feature segments aligned with specific keywords, achieving attribute-level decomposition of visual and textual features. 2) Segments Recomposition for calibrating Novel class: For novel classes, we match each keyword-derived textual feature segment with semantically similar segments from base classes. The corresponding visual feature segments from multiple base prototypes are recomposed to calibrate the raw novel prototype. For instance, when calibrating the cannon prototype (which suffers from missing wheel features due to limited samples), our method identifies wheeled segments from unicycle and oxcart prototypes and cylindrical segments from lipstick and water tower prototypes. This fine-grained recomposition supplements missing attributes while preserving class-discriminative characteristics. + +During the text encoding process of CLIP, the embedding of the entire text description $\mathcal{D}_c$ inherently incorporates the contributions of tokens from each word including those visual keywords $k_{i}\in \mathcal{K}_{C}$ . This means that the attention mechanisms and other internal structures implicitly reveal how and where each keyword contributes to the overall textual representation. Therefore, the keyword grounding function $\mathcal{G}_{\phi}$ is designed to leverage CLIP's internal attention mechanisms to identify these contributions. + +Specifically, in the last transformer layer of CLIP textual encoder $g_{\phi}$ , $\mathcal{D}_c$ is tokenized as $n$ tokens $\{[SOS], \text{token}_1, \dots, \text{token}_{n-2}, [EOS]\}$ . Given the attention matrix $\mathbf{A} \in \mathbb{R}^{n \times n}$ and value matrix $\mathbf{V} \in \mathbb{R}^{d \times n}$ , $\mathbf{t}_c$ is calculated as the output of the last transformer layer as follows: + +$$ +\mathbf {t} _ {c} = \mathbf {V} (:,:) \mathbf {A} (:, e o s) + \mathbf {h} = \left(a _ {1} \mathbf {v} _ {1} + \dots + a _ {n} \mathbf {v} _ {n}\right) + \mathbf {h}, \tag {3} +$$ + +where $\mathbf{A}$ represents the inter-token attention weights, $\mathbf{V}$ represents transformed token representations of all words, $a_{i} = A[i,p_{\mathrm{eos}}]$ represents the attention weight from the $i$ -th token to the [EOS] token, and $\mathbf{v}_i$ means the value vector of the $i$ -th token from CLIP's final transformer layer. + +As can be seen from the calculation of $\mathbf{t}_c$ in Eq. (3), the attention matrix $\mathbf{A}$ and value matrix $\mathbf{V}$ implicitly reveal how and where each keyword contributes to the overall textual representation. Therefore, we extract these token + +wise contribution values as a contribution indication matrix $\mathbf{B}\in \mathbb{R}^{d\times n}$ .. + +$$ +\mathbf {B} = \left[ a _ {1} \mathbf {v} _ {1}, \dots , a _ {n} \mathbf {v} _ {n} \right]. \tag {4} +$$ + +Here, the columns of $\mathbf{B}$ describe different text tokens, while the rows represent feature channels. Each element $\mathbf{B}(j,i)$ indicates the contribution of the $i$ -th token's $j$ -th feature dimension to the final [EOS] embedding, serving as an effective indicator for keyword-related feature dimensions. + +Therefore, to quantify the influence of each token on the text embedding $\mathbf{B}$ , we apply a softmax operation along the token dimension (horizontal) of $\mathbf{B}$ . This results in a matrix $\mathbf{C}$ , where each element represents the proportion of token $i$ 's contribution to $\mathbf{t}_c$ at feature dimension $j$ , relative to all text tokens: + +$$ +\mathbf {C} (i,:) = \operatorname {S o f t m a x} (\mathbf {B} (i,:)) \tag {5} +$$ + +Then, we perform top-q binarization along the feature dimension (vertical) of $\mathbf{B}$ , converting each column $\mathbf{C}(:, i]$ into into a sparse binary mask $\mathbf{g}_i$ (previously introduced in Eq. (2)) as follows: + +$$ +\mathbf {g} _ {i} (j) = \left\{ \begin{array}{l l} 1, & \text {i f} \mathbf {C} (i, j) \text {i s w i t h i n t o p - q r e s p o n s i v e o f} \mathbf {C} (:, j) \\ 0, & \text {o t h e r w i s e .} \end{array} \right. \tag {6} +$$ + +This process isolates the most significant feature dimensions associated with each keyword, enabling grounding of + +textual keywords to text feature segments. + +According to Eqs. (2) to (6), we can derive a fine-grained mapping from each keyword to its corresponding feature segments. While this mapping is computed independently for each class, ideally, the same keyword associated with different classes should yield identical mappings. However, in practice, slight discrepancies often arise due to variations in class contexts. To ensure consistency, for keywords that appear across multiple base classes, we unify all selected feature dimensions by taking the union of their respective binary masks. This strategy preserves all potentially relevant dimensions for the keyword, enhancing robustness and consistency in the grounding process. + +Since CLIP's visual and textual feature spaces are aligned, given a keyword $k_{i}$ , we can obtain the visual feature segment $\mathbf{s}_i^c$ of prototype $\mathbf{p}_c$ , by simply performing an element-wise multiplication as follows: + +$$ +\mathbf {s} _ {i} ^ {c} = \mathbf {p} _ {c} \odot \mathbf {g} _ {i}. \tag {7} +$$ + +In this way, we achieve semantic-aware decomposition of base class prototypes guided by textual keywords of visual attributes, where each feature segment $\mathbf{s}_i^c$ retains only the dimensions most relevant to the corresponding semantic keyword. + +# 4.2. Feature segments decomposition for fine-grained calibration + +The prototypes of novel classes are often biased due to the limited number of shots available per class. While the full-dimensional features of these prototypes can only be derived from the novel class samples, certain partial features—such as visual attribute features—frequently overlap with those of base class prototypes. For instance, features like "wheels" are shared between the novel class "cannon" and base classes "unicycle", "oxcart", and "race car". Leveraging this inherent property, we calibrate the prototypes of novel classes by piecewise refinement of their semantic feature segments. This is achieved by assembling similar segments from the base classes, thereby enhancing the representational accuracy and reducing bias in the novel class prototypes. + +In this process, given any novel class $\widetilde{C}$ , we first generate its initial prototype $\mathbf{p}_{\widetilde{C}}$ and visual keywords $\kappa_{\widetilde{C}}$ following the same process as base classes. + +Next, for each keyword $k_{i} \in \mathcal{K}_{\widetilde{C}}$ , we retrieve the corresponding feature segments $\mathbf{s}_i^{c_j}$ from the base-class prototypes that share the same keywords. These segments are then combined using a weighted summation to calibrate the feature segment of the novel class: + +$$ +\mathbf {r} _ {i} = \sum_ {j = 1} ^ {M} w _ {j} \cdot \mathbf {s} _ {i} ^ {c _ {j}}. \tag {8} +$$ + +Here, the $\mathbf{s}_i^{c_j}$ represents the prototype segment of base class $C_j$ corresponding to the keyword $k_{i}$ . The weight $w_{j}$ reflects + +![](images/87a5604a38b60b14516f41952a8c2ecce84ee4c5b91d57742768105e91385c2b.jpg) +Figure 2. Implementation of keyword grounding function $\mathcal{G}_{\phi}$ . The upper describes the computation of $\mathbf{t}_c$ using CLIP, while the lower details the process for generating the grounding vector $\mathbf{g}_c$ . + +the similarity between the feature segments of $\widetilde{C}$ and base class $C_j$ . This similarity is approximated by computing the textual embedding similarity between the textual segment of novel class $\mathbf{t}_{\widetilde{c}} \odot \mathbf{g}_i$ and that of every base class $\mathbf{t}_{c_j} \odot \mathbf{g}_i, C_j \in C^0$ . The top-r similarity scores are preserved after using a softmax function, while the remaining similarity scores are set to zero. This ensures that only the most semantically relevant base-class feature segments are retained, yielding the final weights $w_j$ . + +After obtaining feature segments for all keywords associated with novel class $\widetilde{C}$ , keyword-specific adjustments are iteratively applied to novel prototype $\mathbf{p}_{\widetilde{c}}$ to obtain the final calibrated prototype as below: + +$$ +\mathbf {p} _ {\tilde {c}} \leftarrow (1 - \alpha) \cdot \mathbf {p} _ {\tilde {c}} + \alpha \cdot \mathbf {r} _ {i}, \quad i \in \{1, \dots , m \} \tag {9} +$$ + +where $\alpha$ is a trade-off hyperparameter that balances the prototype directly from the few shots and calibrated one. + +In Eq. (9), each calibration $\mathbf{r}_i$ is performed individually and is a linear combination of similar feature segments from base-class prototypes. Importantly, this linear combination varies for different feature segments, as the weights and contributing segments depend on the specific keyword $k_i$ . Consequently, the overall calibration is semantic-segmentwise linear but nonlinear overall, ensuring fine-grained and more accurate prototype calibration for few-shot novel classes. + +Table 1. Comparison with existing methods (5-shot setting). $A_{base}$ for base session accuracy, $A_{avg}$ for AvgAcc across all sessions + +
MethodBackboneTraining FreeCIFAR100miniImageNetCUB200IN1K-FSCIL
\(A_{base}\)\(A_{avg}\)\(A_{base}\)\(A_{avg}\)\(A_{base}\)\(A_{avg}\)\(A_{base}\)\(A_{avg}\)
CLOSER[2]ResNet-18/20No75.763.176.062.879.469.1--
Yourself[6]ViT-B-IN21KNo82.967.084.068.883.469.9--
LP-DiF[4]ViT-B-CLIPNo80.275.196.393.883.974.0--
PriViLego[5]ViT-B-CLIPNo84.277.295.492.782.174.272.168.9
EASE[3]ViT-B-IN21KNo85.178.096.893.983.075.173.370.4
FDR(Ours)ViT-B-CLIPNo87.779.395.389.284.477.773.670.9
BaselineViT-B-CLIPYes75.767.992.088.680.973.669.265.9
TEEN[1]ViT-B-CLIPYes75.768.492.088.780.973.769.266.1
FDR(Ours)ViT-B-CLIPYes75.768.592.089.180.974.169.266.4
+ +Table 2. Comparison with existing methods on CUB200-1-shot benchmark (Accuracy %). Bold indicates best performance. $\Delta$ shows improvement of our method over the previous best. Due to space limitations, only the first three and last three sessions are shown. + +
Session Method01238910
AvgAccAvgAccNAccAvgAccNAccAvgAccNAccAvgAccNAccAvgAccNAccAvgAccNAcc
Baseline [21]80.8574.1814.2568.8618.6665.0018.3952.1218.2251.6420.3049.7519.86
TEEN [1]80.8576.4035.1671.8938.2666.7030.0056.6029.7556.1432.1954.4831.56
FDR(Ours)80.8577.2654.8272.9350.0968.3839.7458.3336.4558.7638.9157.3738.40
Δ-+0.86+19.66+1.04+11.83+1.68+9.74+1.73+6.70+2.62+6.72+2.89+6.84
+ +# 5. Experiments + +This section begins by detailing the datasets and implementation specifics. Subsequently, it presents a comparison with existing approaches to benchmark performance, followed by an ablation study to assess the effectiveness of individual components of our method. + +# 5.1. Datasets and Implementation Details + +The evaluations are conducted on four benchmarks, following standard Few-Shot Class-Incremental Learning (FSCIL) protocols: CIFAR100, miniImageNet, CUB200, and IN1K-FSCIL. + +CIFAR100, miniImageNet, and CUB200 are widely used benchmarks for FSCIL. CUB200 contains 200 bird species, while miniImageNet and CIFAR100 each comprise 100 common everyday object categories. This section details the experimental results on CUB200; results for the other two datasets are provided in the supplementary material. We adopt the same dataset split for CUB200 as used in existing works [1, 30, 32], denoted as CUB200-5-shot. Specifically, this follows an 11-session split: Session 0 uses 100 base classes (3,000 images), followed by 10 incremental sessions, each introducing 10 novel classes with 5-shot support per class. To further evaluate the effectiveness of prototype calibration under conditions of extreme data scarcity, we also evaluate a 10-way 1-shot configuration for CUB200, where only one sample per novel class is available. This configuration is denoted as CUB200-1-shot. + +IN1K-FSCIL. Furthermore, to assess scalability in real-world scenarios with more classes, we construct a challeng- + +ing ImageNet1K[11] variant by reorganizing 1,000 classes into 800 base classes (200 images per class) and 200 novel classes divided across 10 incremental sessions (20-way 1-shot), which is denoted as IN1K-FSCIL. This setting introduces higher inter-class density and diversity compared to existing benchmarks, better simulating practical applications requiring large-scale incremental learning. The results are provided in the supplementary material. + +Implementation: For our method, all experiments are implemented in PyTorch with CLIP-ViT/B-16 [17] as the backbone. 512-dimensional visual and textual features are extracted from CLIP's aligned embedding space, with the backbone frozen during both the base and incremental phases, making FDR a training-free method. To further bridge the domain gap between CLIP's pretraining data and the target dataset, we apply LoRA fine-tuning on CLIP during the base phase by leveraging an image-text contrastive loss, while refraining from any training in the incremental phase. + +Evaluation Metrics: For evaluation, we adopt the commonly used metric of average accuracy (AvgAcc). Specifically, at session $t$ , the average classification accuracy is computed across all classes encountered up to that session (i.e. $M + t * N$ ). While AvgAcc provides a comprehensive measure of overall performance, it is less sensitive to the accuracy of newly learned novel classes. To address this, we also introduce the average accuracy for novel classes, denoted as NAcc. NAcc is calculated similarly to AvgAcc (i.e., classification over all classes), but the proportion of correctly classified instances is computed exclu + +Table 3. Comparison with existing methods on CUB200-5-shot benchmark (Accuracy %). Bold indicates best performance. $\Delta$ shows improvement of our method over the previous best. Due to space limitations, only the first three and last three sessions are shown. + +
MethodBackboneAccuracy in each session (%)
01238910
AvgAccAvgAccNAccAvgAccNAccAvgAccNAccAvgAccNAccAvgAccNAccAvgAccNAcc
Baseline[21]ResNet1874.1468.0912.3361.869.5957.628.5645.558.8543.3310.1841.4810.06
TOPIC[22]ResNet1868.6862.49-54.81-49.99-32.22-28.31-26.26-
CEC[30]ResNet1875.8571.94-68.50-63.50-54.83-53.52-52.28-
FACT[32]ResNet1875.9073.23-70.84-66.13-58.41-57.89-56.94-
TEEN[1]ResNet1877.2676.13-72.81-68.16-61.19-60.32-59.31-
Baseline[21]CLIP80.8579.1165.0277.4965.0273.9857.2069.5356.1169.3558.3169.2059.13
TEEN [1]CLIP80.8579.2167.2377.5466.1873.8758.2169.5656.8669.4558.0169.3758.83
FDR(Ours)CLIP80.8579.3873.7677.7070.7374.2862.5170.1359.6870.0861.8170.1562.79
Δ-+0.17+6.53+0.16+4.55+0.30+4.30+0.57+2.82+0.63+3.80+0.78+3.96
+ +sively within the novel classes. + +# 5.2. Comparison with Existing Methods + +We have made comparisons against recent SOTA methods [2-6]. As shown in Table 1, our FDR method surpasses all training-free approaches, consistent with the results of the main paper. For approaches that require training[2-6], we apply the CLIP backbone fine-tuned on base classes. The fine-tuning significantly improves the base session accuracy of our method, making it outperform other methods on CIFAR100, CUB200, and IN1K-FSCIL. + +Table 2 and Table 3 present the detailed comparison results on the CUB200 dataset. All methods utilize the same frozen CLIP backbone (without fine-tuning) or ResNet18 to ensure a fair comparison. The baseline represents vanilla Prototypical Networks [21], which directly uses CLIP features without prototype calibration. The recently proposed method, TEEN, demonstrates significantly better performance than the baseline, benefiting from its comprehensive prototype calibration. Furthermore, our method achieves superior performance across all sessions in both 5-shot and 1-shot settings. Notably, in the 1-shot setting, our method achieves a $6.70\% \sim 19.66\%$ improvement in NAcc across all incremental sessions when comparing with holistic calibration method [1], highlighting its effectiveness in bias correction even under extreme data scarcity. Additionally, the progressive increases in AvgAcc, with improvements ranging from $+0.86\%$ to $+2.89\%$ , suggest cumulative advantages from our composition strategy as the incremental sessions progress. present the comparison results on the CUB200 dataset. All methods utilize the same frozen CLIP backbone or ResNet18 to ensure a fair comparison. The baseline represents vanilla Prototypical Networks [21], which directly uses CLIP features without prototype calibration. The recently proposed method, TEEN, demonstrates significantly better performance than the baseline, benefiting from its comprehensive prototype calibration. Furthermore, our method achieves superior performance across all sessions in both 5-shot and 1-shot set + +tings. Notably, in the 1-shot setting, our method achieves a $6.70\% \sim 19.66\%$ improvement in NAcc across all incremental sessions when comparing with holistic calibration method [1], highlighting its effectiveness in bias correction even under extreme data scarcity. Additionally, the progressive increases in AvgAcc, with improvements ranging from $+0.86\%$ to $+2.89\%$ , suggest cumulative advantages from our composition strategy as the incremental sessions progress. + +Table 4. Ablation study on CUB200-1-shot (Session 10). + +
MethodAvgAccNAcc
No Calibration49.7519.86
Holistic Calibration only54.4831.56
Random Feature Segments54.8332.08
Naive Keyword Grounding56.6136.25
Full (Ours)57.3738.40
+ +# 5.3. Ablation Studies + +This section presents ablation experiments on the CUB200-1-shot dataset to quantify the contribution of each component and assess the impact of hyperparameters. The performance is evaluated using the final session's metrics, including the average accuracy (AvgAcc) for all classes and the novel class accuracy (NAcc) for all novel classes. + +Effectiveness of Component. We dissect the key components of our method through several controlled experiments: + +- No Calibration: This strategy directly aggregates features of each class provided by the CLIP backbone to form prototypes. For novel classes, the prototype is calculated directly from the few-shot samples without any calibration. As a result, this approach serves as a vanilla baseline. +- Holistic Calibration: This method corresponds to the TEEN approach [1], which employs holistic calibration for prototypes. It is used as a comparative baseline to dissect and evaluate the effectiveness of the fine-grained calibration proposed in our method. + +![](images/eb124ccf06927eaa0bdf9ae42edcb4994b0dafc92da504fd06769a9eb51aeac3.jpg) +(a) NAcc w.r.t. Top-q + +![](images/d119e743b9305a427f59f1da63a47cb613759d632644ec853252b86a01cee728.jpg) +(b) NAcc w.r.t. Top-r +Figure 3. Impact of hyperparameters on average accuracy over all novel classes at last session (i.e. NAcc), where $q$ stands for the size of feature segments, $r$ for the number of base-class feature segments for calibrating, and $\alpha$ for the calibration strength of each feature segment. + +![](images/b7e3d5426c9f0ad308bdfdc13f6ada4642169333a992ec3e3636fa183145609a.jpg) +(c) NAcc w.r.t. $\alpha$ + +- Random Feature Segments: This strategy represents a degraded version of our method, where the feature segments are replaced by random divisions of visual features without any semantic guidance. This approach is used to dissect the necessity of the keyword grounding function. +- Naive Keyword Grounding: This strategy also represents a degraded version of our method, where the mapping of each keyword to the feature segment is simplified by multiplying the [EOS] token embedding with the keyword token embedding, rather than incorporating the attention matrix $\mathbf{A}$ and value matrix $\mathbf{V}$ . This approach is used to dissect and demonstrate that whether the design of the keyword grounding function is trivial. + +The evaluation results are shown in Table 4. As can be observed, the $11.70\%$ NAcc gain of "Holistic Calibration only" over "No Calibration" confirms that prototype calibration is beneficial. Furthermore, our method outperforms "Holistic Calibration only", demonstrating that fine-grained calibration of few-shot class prototypes is both effective and necessary. The inferior performance of "Random Feature Segments" and "Naive Keyword Grounding" compared to our full method indicates that the effectiveness of fine-grained calibration for few-shot class prototypes relies heavily on its elaborate design. This underscores the value of our carefully designed keyword grounding function, highlighting its critical role in achieving superior performance. + +Impact of Hyperparameter. We analyze the impact of three key parameters in controlled experiments: + +- Top- $q$ binarization are conducted on contribution indication matrix $\mathbf{B}$ along the feature dimensions, where $q$ controls the size of feature segments. From Fig. 3a, we observe that feature segments obtained the best performance with $q = 50$ in NAcc. Smaller values of $q$ (e.g., 10 and 30) result in smaller feature segments, which may be insufficiently captured by the corresponding keywords, leading to suboptimal performance. On the other hand, larger values of $q$ (e.g., 200, 512) degrade performance, as they tend to approach holistic calibration and fail to yield improvements. This indicates that an intermediate granu + +larity in feature segmentation, which well corresponds to the keyword, is essential. + +- Top-r similar prototype segments from base classes are combined to calibrate the novel class, where $r$ controls the number of base-class feature segments for calibrating. In Fig. 3b, we see that increasing the number of similar prototypes (i.e. a larger $r$ ) enhances performance up to $r = 50$ , after which further increases do not yield substantial benefits. This suggests that a small set of highly relevant prototypes is sufficient for effectively calibrating novel class prototypes. Including too many segments may introduce noise or less relevant information, diminishing the calibration quality. +- Calibration ratio $\alpha$ controls the calibration strength of each feature segment. As shown in Fig. 3c, the calibration ratio $\alpha = 0.5$ yields the best results, balancing the original and calibrated features effectively. Lower values of $\alpha$ reduce the impact of calibration, while higher values overly prioritize the reference prototypes, leading to suboptimal results. + +# 6. Conclusion and Future Work + +In this work, we propose a semantic Feature Decomposition-Recomposition method that unlocks the full potential of CLIP's visual-semantic features for few-shot class-incremental learning. Our core innovation lies in decomposing CLIP's frozen features into semantic-aware segments through guidance of textual keywords, making these segments transfer-ready through incremental learning to be selectively recomposed for novel class calibration. By establishing explicit mappings between LLM-generated attribute keywords and their corresponding feature segments, our method enables precise inheritance of visual-semantic characteristics from multiple base classes. The improvement of our method over existing methods on several settings demonstrates its unique capability to mitigate prototype bias while preserving discriminative features. + +# 7. Acknowledgment + +This work is partially supported by the National Natural Science Foundation of China (No.62495082 and No.62461160331). + +# References + +[1] Few-shot class-incremental learning via training-free prototype calibration. In NIPS, 2023. 1, 2, 3, 6, 7 +[2] Closer: Towards better representation learning for few-shot class-incremental learning. In ECCV, 2024. 6, 7 +[3] Expandable subspace ensemble for pre-trained model-based class-incremental learning. In CVPR, 2024. 6 +[4] Learning prompt with distribution-based feature replay for few-shot class-incremental learning. arXiv preprint, 2024. 6 +[5] Pre-trained vision and language transformers are few-shot incremental learners. In CVPR, 2024. 6 +[6] Rethinking few-shot class-incremental learning: Learning from yourself. In ECCV, 2024. 6, 7 +[7] Afra Feyza Akyurek, Ekin Akyurek, Derry Tanti Wijaya, and Jacob Andreas. Subspace regularizers for few-shot class incremental learning. arXiv preprint arXiv:2110.07059, 2021. 1, 2, 3 +[8] Arslan Chaudhry, Marcus Rohrbach, Mohamed Elhoseiny, Thalaiyasingam Ajanthan, Puneet K Dokania, Philip HS Torr, and Marc'Aurelio Ranzato. On tiny episodic memories in continual learning. arXiv preprint arXiv:1902.10486, 2019. 2 +[9] Kuilin Chen and Chi-Guhn Lee. Incremental few-shot learning via vector quantization in deep embedded space. In International conference on learning representations, 2021. 1 +[10] Ali Cheraghian, Shafin Rahman, Pengfei Fang, Soumava Kumar Roy, Lars Petersson, and Mehrtash Harandi. Semantic-aware knowledge distillation for few-shot class-incremental learning. In CVPR, pages 2534-2543, 2021. 1, 2, 3 +[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 6 +[12] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In CVPR, pages 831-839, 2019. 2 +[13] David Isele and Akansel Cosgun. Selective experience replay for lifelong learning. In Proceedings of the AAAI conference on artificial intelligence, 2018. 2 +[14] Gregory Koch, Richard Zemel, Ruslan Salakhutdinov, et al. Siamese neural networks for one-shot image recognition. In ICML deep learning workshop, pages 1-30. Lille, 2015. 2 +[15] Anna Kukleva, Hilde Kuehne, and Bernt Schiele. Generalized and incremental few-shot learning by explicit learning and calibration without forgetting. In ICCV, 2021. 1 +[16] Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In Psychology of learning and motivation. Elsevier, 1989. 1, 2 + +[17] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning. PmLR, 2021. 6 +[18] David Rolnick, Arun Ahuja, Jonathan Schwarz, Timothy Lillicrap, and Gregory Wayne. Experience replay for continual learning. NIPS, 2019. 2 +[19] Guangyuan Shi, Jiaxin Chen, Wenlong Zhang, Li-Ming Zhan, and Xiao-Ming Wu. Overcoming catastrophic forgetting in incremental few-shot learning by finding flat minima. NIPS, 34:6747-6761, 2021. 1 +[20] Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay. NIPS, 30, 2017. 2 +[21] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. NeurIPS, 30, 2017. 1, 2, 3, 6, 7 +[22] Xiaoyu Tao, Xiaopeng Hong, Xinyuan Chang, Songlin Dong, Xing Wei, and Yihong Gong. Few-shot class-incremental learning. In CVPR, 2020. 1, 2, 3, 7 +[23] Songsong Tian, Lusi Li, Weijun Li, Hang Ran, Xin Ning, and Prayag Tiwari. A survey on few-shot class-incremental learning. Neural Networks, 169, 2024. 1 +[24] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al. Matching networks for one shot learning. NIPS, 29, 2016. 2 +[25] Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: Theory, method and application. IEEE TPAMI, 2024. 2 +[26] Yaqing Wang, Quanming Yao, James T Kwok, and Lionel M Ni. Generalizing from a few examples: A survey on few-shot learning. ACM computing surveys (csur), 53(3), 2020. 1, 2 +[27] Chenshen Wu, Luis Herranz, Xialei Liu, Joost Van De Weijer, Bogdan Raducanu, et al. Memory replay gans: Learning to generate new categories without forgetting. NIPS, 31, 2018. 2 +[28] Boyu Yang, Mingbao Lin, Binghao Liu, Mengying Fu, Chang Liu, Rongrong Ji, and Qixiang Ye. Learnable expansion-and-compression network for few-shot class incremental learning. arXiv preprint arXiv:2104.02281, 2021.1 +[29] Boyu Yang, Mingbao Lin, Yunxiao Zhang, Binghao Liu, Xiaodan Liang, Rongrong Ji, and Qixiang Ye. Dynamic support network for few-shot class incremental learning. IEEE TPAMI, 45(3):2945-2951, 2022. 1 +[30] Chi Zhang, Nan Song, Guosheng Lin, Yun Zheng, Pan Pan, and Yinghui Xu. Few-shot incremental learning with continually evolved classifiers. In CVPR, pages 12455-12464, 2021. 1, 6, 7 +[31] Guangtao Zheng and Aidong Zhang. Few-shot class-incremental learning with meta-learned class structures. In 2021 International Conference on Data Mining Workshops (ICDMW), pages 421–430. IEEE, 2021. 1 +[32] Da-Wei Zhou, Fu-Yun Wang, Han-Jia Ye, Liang Ma, Shiliang Pu, and De-Chuan Zhan. Forward compatible few + +shot class-incremental learning. In CVPR, pages 9046-9056, 2022. 1, 6, 7 +[33] Kai Zhu, Yang Cao, Wei Zhai, Jie Cheng, and Zheng-Jun Zha. Self-promoted prototype refinement for few-shot class incremental learning. In CVPR, pages 6801-6810, 2021. 1 \ No newline at end of file diff --git a/2025/Feature Decomposition-Recomposition in Large Vision-Language Model for Few-Shot Class-Incremental Learning/images.zip b/2025/Feature Decomposition-Recomposition in Large Vision-Language Model for Few-Shot Class-Incremental Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3da927c723b4d4bb5957ea70f556475325982f9b --- /dev/null +++ b/2025/Feature Decomposition-Recomposition in Large Vision-Language Model for Few-Shot Class-Incremental Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9b6ca61a3775c12ddb97bc4a3ca8c9db4da0082d43765352621ce25e6af8f73f +size 425763 diff --git a/2025/Feature Decomposition-Recomposition in Large Vision-Language Model for Few-Shot Class-Incremental Learning/layout.json b/2025/Feature Decomposition-Recomposition in Large Vision-Language Model for Few-Shot Class-Incremental Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..82863449489557ec471aa038eeb2c7aa4ce991ff --- /dev/null +++ b/2025/Feature Decomposition-Recomposition in Large Vision-Language Model for Few-Shot Class-Incremental Learning/layout.json @@ -0,0 +1,8849 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 72, + 102, + 539, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 102, + 539, + 140 + ], + "spans": [ + { + "bbox": [ + 72, + 102, + 539, + 140 + ], + "type": "text", + "content": "Feature Decomposition-Recomposition in Large Vision-Language Model for Few-Shot Class-Incremental Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 160, + 463, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 160, + 463, + 175 + ], + "spans": [ + { + "bbox": [ + 146, + 160, + 463, + 175 + ], + "type": "text", + "content": "Zongyao Xue" + }, + { + "bbox": [ + 146, + 160, + 463, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 146, + 160, + 463, + 175 + ], + "type": "text", + "content": ", Meina Kan" + }, + { + "bbox": [ + 146, + 160, + 463, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 146, + 160, + 463, + 175 + ], + "type": "text", + "content": ", Shiguang Shan" + }, + { + "bbox": [ + 146, + 160, + 463, + 175 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 146, + 160, + 463, + 175 + ], + "type": "text", + "content": ", Xin Chen" + }, + { + "bbox": [ + 146, + 160, + 463, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 83, + 175, + 527, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 175, + 527, + 190 + ], + "spans": [ + { + "bbox": [ + 83, + 175, + 527, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 83, + 175, + 527, + 190 + ], + "type": "text", + "content": "Institute of Computing Technology, Chinese Academy of Sciences, Beijing, 100090, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 189, + 473, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 189, + 473, + 203 + ], + "spans": [ + { + "bbox": [ + 137, + 189, + 473, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 137, + 189, + 473, + 203 + ], + "type": "text", + "content": "University of Chinese Academy of Sciences, Beijing, 100090, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 179, + 203, + 428, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 203, + 428, + 217 + ], + "spans": [ + { + "bbox": [ + 179, + 203, + 428, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 179, + 203, + 428, + 217 + ], + "type": "text", + "content": "Peng Cheng Laboratory, Shenzhen, 518055, China" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 270, + 297, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 270, + 297, + 643 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 297, + 643 + ], + "type": "text", + "content": "Few-Shot Class-Incremental Learning (FSCIL) focuses on incrementally learning novel classes using only a limited number of samples from novel classes, which faces dual challenges: catastrophic forgetting of previously learned classes and over-fitting to novel classes with few available samples. Recent advances in large pre-trained vision-language models (VLMs), such as CLIP, provide rich feature representations that generalize well across diverse classes. Therefore, freezing the pre-trained backbone and aggregating class features as prototypes becomes an intuitive and effective way to mitigate catastrophic forgetting. However, this strategy fails to address the overfitting challenge, and the prototypes of novel classes exhibit semantic bias due to the few samples per class. To address these limitations, we propose a semantic Feature Decomposition-Recomposition (FDR) method based on VLMs. Firstly, we decompose the CLIP features into semantically distinct segments guided by text keywords from base classes. Then, these segments are adaptively recomposed at the attribute level given text descriptions, forming calibrated prototypes for novel classes. The recomposition process operates linearly at the attribute level but induces nonlinear adjustments across the entire prototype. This fine-grained and non-linear recomposition inherits the generalization capabilities of VLMs and the adaptive recomposition ability of base classes, leading to enhanced performance in FSCIL. Extensive experiments demonstrate our method's effectiveness, particularly in 1-shot scenarios where it achieves improvements between " + }, + { + "bbox": [ + 54, + 270, + 297, + 643 + ], + "type": "inline_equation", + "content": "6.70\\%" + }, + { + "bbox": [ + 54, + 270, + 297, + 643 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 270, + 297, + 643 + ], + "type": "inline_equation", + "content": "19.66\\%" + }, + { + "bbox": [ + 54, + 270, + 297, + 643 + ], + "type": "text", + "content": " for novel classes over previous prototype-based methods on CUB200. Code is available at https://github.com/herb1999/FDR." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 668, + 135, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 668, + 135, + 681 + ], + "spans": [ + { + "bbox": [ + 56, + 668, + 135, + 681 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "The remarkable success of deep neural networks (DNNs) is largely based on training with large-scale static datasets" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 245, + 555, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 245, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 555, + 449 + ], + "type": "text", + "content": "[23]. However, real-world scenarios such as autonomous driving and robotic field operations often involve continuously arriving data streams, where new categories emerge incrementally with only sparse samples available. Although humans excel at adapting to such dynamic environments, existing learning-based algorithms struggle under these constraints. To bridge this gap, Few-Shot Class-Incremental Learning (FSCIL)[22] is introduced as a more practical and challenging task, aiming to incrementally recognize novel classes from few-shot examples while preserving performance on base classes. The FSCIL task inherits dual challenges from few-shot learning and continual learning. First, models must overcome the unreliable empirical risk minimization problem [26] when generalizing with scarce novel class samples. Second, they must maintain stability against catastrophic forgetting [16] of base classes while adapting to incremental updates." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 449, + 556, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 556, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 556, + 569 + ], + "type": "text", + "content": "Since its introduction in 2020 [22], FSCIL has attracted significant attention due to its potential in real-world applications, leading to a variety of early methodological developments, including meta learning[9, 31, 33], specialized model architectures[28, 29], optimization strategies [19] and replay-based methods[10, 15]. However, these approaches face a longstanding plasticity-stability dilemma, wherein models struggle to simultaneously preserve base class knowledge while effectively accommodating novel class acquisition." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": "Recently, leveraging the generalizable representations of pre-trained models such as CLIP, prototype-based metric learning approaches [1, 21, 30, 32] have emerged as a prominent solution to mitigate catastrophic forgetting. These methods typically freeze the pre-trained backbone networks and aggregate features from samples within each class to construct prototypes, enabling effective knowledge preservation and incremental learning. However, the scarce samples of novel classes hinder the full exploitation of features from large pre-trained models, yielding semantically biased novel prototypes that deviate from true class distributions. To address these biases, existing methods [1, 7, 10]" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3153" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "leverage supplementary information from base class prototypes to calibrate novel ones. Nevertheless, these approaches directly employ entire similar base prototypes to compensate for the biased novel ones, lacking mechanisms to identify which specific components are transferable. This oversight leads to reduced inter-class distances between novel and base classes, causing feature space entanglement." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 157, + 298, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 157, + 298, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 157, + 298, + 443 + ], + "type": "text", + "content": "To address these limitations, we propose a Feature Decomposition-Recomposition (FDR) framework that enables fine-grained prototype calibration through attribute-level knowledge transfer. Our key insight is that although the feature representations from large pre-trained models contain sufficient information for classifying diverse classes, their correct classification depends on strategically extracting generalizable features, particularly for few-shot classes. Specifically, given features from pre-trained VLMs like CLIP, we first decompose the visual prototypes of base classes into semantic-aware feature segments by establishing mappings between CLIP visual features and attribute keywords through a keyword grounding function. These semantic feature segments are transferable across classes. Next, for novel classes, given a set of keywords describing them, each corresponding feature segment of the biased prototype derived from few-shot samples is calibrated by recombining segments from similar base class prototypes (e.g., calibrating the \"cannon\" prototype by combining \"wheel\" segments from classes like bicycles and \"cylindrical\" segments from classes like lipsticks). The combinatorial possibilities for synthesizing novel prototypes grow exponentially as the number of base classes increases, effectively addressing the prototype entanglement problem." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 444, + 247, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 444, + 247, + 454 + ], + "spans": [ + { + "bbox": [ + 67, + 444, + 247, + 454 + ], + "type": "text", + "content": "The contributions of this work are threefold:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 455, + 295, + 622 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 56, + 455, + 295, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 455, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 295, + 514 + ], + "type": "text", + "content": "1. We propose a semantic feature decomposition and recomposition method, FDR, that divides CLIP's frozen features into reusable attribute-level feature segments, enabling precise and easy knowledge transfer from base classes to novel ones for enhanced incremental learning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 515, + 295, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 515, + 295, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 515, + 295, + 574 + ], + "type": "text", + "content": "2. We design a keyword grounding function to decompose CLIP's frozen features into semantic-aware segments, and introduce a segment-wise recomposition method to calibrate biased prototypes at the fine-grained attribute-level." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 574, + 295, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 295, + 622 + ], + "type": "text", + "content": "3. Our approach is training-free, simple yet highly effective, achieving improvements of " + }, + { + "bbox": [ + 55, + 574, + 295, + 622 + ], + "type": "inline_equation", + "content": "6.70\\% - 19.66\\%" + }, + { + "bbox": [ + 55, + 574, + 295, + 622 + ], + "type": "text", + "content": " in novel class accuracy over state-of-the-art methods, showcasing its practical utility." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 633, + 142, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 633, + 142, + 645 + ], + "spans": [ + { + "bbox": [ + 55, + 633, + 142, + 645 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 654, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 715 + ], + "type": "text", + "content": "Incremental learning (IL), also referred to as continual or lifelong learning, distinguishes itself from conventional machine learning by addressing dynamic data distributions where models must continually integrate new knowledge from sequential tasks while preserving performance on pre" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "text", + "content": "viously learned ones [25]. The core challenge lies in balancing learning plasticity for new knowledge and memory stability for old knowledge [16]. An intuitive approach to ensure the stability of old knowledge is to retain representative samples from previous tasks and replay them during training on new tasks[8, 13, 18]. However, growing memory demands and privacy concerns limit their scalability. To enhance storage efficiency, retaining features of old samples and replaying them later is also an option[12]. Another method for preserving old knowledge is generative replay[20, 27], where a generative model learns the distribution of old data and replays generated samples or features. These methods primarily focus on preserving old knowledge rather than leveraging it to enhance new task learning." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 243, + 553, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 553, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 553, + 445 + ], + "type": "text", + "content": "Few-Shot Learning (FSL) requires models to learn from a limited number of labeled samples, making it suitable for real-world scenarios where data is scarce due to safety, privacy, or ethical concerns[26]. Metric-based methods [14, 21, 24] have proven to be simple and effective solutions by learning transferable embedding spaces. Prototypical Networks[21], for instance, define a metric space where classification is performed by computing distances to prototype representations of each class. These prototypes, computed as class-mean feature vectors, naturally accommodate incremental updates by freezing the embedding backbone - a property that makes them particularly suitable for few-shot class-incremental learning [22]. However, prototype quality degrades significantly under extreme data scarcity (e.g., 1-shot settings), as limited samples fail to capture true class distributions, leading to biased prototypes that misalign with semantic realities." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 451, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 553, + 713 + ], + "type": "text", + "content": "Few-Shot Class-Incremental Learning (FSCIL) aims to incrementally learn novel classes given only its few shots[22], which merges challenges from IL and FSL: preventing catastrophic forgetting of base classes while generalizing to novel classes with sparse samples. A common strategy to mitigate this is to use semantically similar base classes to correct biases in novel classes[1, 7, 10]. Additionally, incorporating side information from language has been explored to enhance model performance[7, 10]. The closest to our work is [1], which reduces novel prototype bias by globally calibrating them with similar base prototypes. However, such coarse-grained adjustments compress interclass distances, leading to prototype entanglement in later phases. Our method addresses these limitations through fine-grained calibration, decomposing prototypes into semantically meaningful sub-features for targeted composition. This enables precise inheritance of relevant visual attributes from multiple base classes while maintaining interclass separability. The integration of language semantics also resolves ambiguity in visual similarity measurements, particularly for classes sharing partial attributes rather than global appearances." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3154" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 135, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 135, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 135, + 85 + ], + "type": "text", + "content": "3. Background" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 139 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 139 + ], + "type": "text", + "content": "Following the existing works [1, 7, 10, 22], the Few-Shot Class-Incremental Learning (FSCIL) task is defined as a process consisting of " + }, + { + "bbox": [ + 55, + 91, + 296, + 139 + ], + "type": "inline_equation", + "content": "(1 + T)" + }, + { + "bbox": [ + 55, + 91, + 296, + 139 + ], + "type": "text", + "content": " sequential learning sessions, which are divided into two consecutive phases:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 140, + 296, + 260 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 55, + 140, + 295, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 140, + 295, + 176 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 295, + 176 + ], + "type": "text", + "content": "- Base phase (the first session " + }, + { + "bbox": [ + 55, + 140, + 295, + 176 + ], + "type": "inline_equation", + "content": "S_0" + }, + { + "bbox": [ + 55, + 140, + 295, + 176 + ], + "type": "text", + "content": "): during " + }, + { + "bbox": [ + 55, + 140, + 295, + 176 + ], + "type": "inline_equation", + "content": "S_0" + }, + { + "bbox": [ + 55, + 140, + 295, + 176 + ], + "type": "text", + "content": ", the model is trained on " + }, + { + "bbox": [ + 55, + 140, + 295, + 176 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 140, + 295, + 176 + ], + "type": "text", + "content": " base classes " + }, + { + "bbox": [ + 55, + 140, + 295, + 176 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^0 = \\{C_1, \\dots, C_M\\}" + }, + { + "bbox": [ + 55, + 140, + 295, + 176 + ], + "type": "text", + "content": " with abundant samples per class." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 176, + 296, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 176, + 296, + 260 + ], + "spans": [ + { + "bbox": [ + 55, + 176, + 296, + 260 + ], + "type": "text", + "content": "- Incremental phase (sessions " + }, + { + "bbox": [ + 55, + 176, + 296, + 260 + ], + "type": "inline_equation", + "content": "S_{1} \\sim S_{T}" + }, + { + "bbox": [ + 55, + 176, + 296, + 260 + ], + "type": "text", + "content": "): following the N-way K-shot protocol, at each session " + }, + { + "bbox": [ + 55, + 176, + 296, + 260 + ], + "type": "inline_equation", + "content": "S_{t}|_{t=1}^{T}" + }, + { + "bbox": [ + 55, + 176, + 296, + 260 + ], + "type": "text", + "content": ", the model learns " + }, + { + "bbox": [ + 55, + 176, + 296, + 260 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 176, + 296, + 260 + ], + "type": "text", + "content": " novel classes with " + }, + { + "bbox": [ + 55, + 176, + 296, + 260 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 176, + 296, + 260 + ], + "type": "text", + "content": " samples per class (typically " + }, + { + "bbox": [ + 55, + 176, + 296, + 260 + ], + "type": "inline_equation", + "content": "K \\leq 5" + }, + { + "bbox": [ + 55, + 176, + 296, + 260 + ], + "type": "text", + "content": "). The novel classes are denoted as " + }, + { + "bbox": [ + 55, + 176, + 296, + 260 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^t = \\{C_{M^t + 1}, \\dots, C_{M^t + N}\\}" + }, + { + "bbox": [ + 55, + 176, + 296, + 260 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 176, + 296, + 260 + ], + "type": "inline_equation", + "content": "M^t = M + (t - 1) \\times N" + }, + { + "bbox": [ + 55, + 176, + 296, + 260 + ], + "type": "text", + "content": ". The label spaces across distinct sessions remain strictly disjoint, i.e. " + }, + { + "bbox": [ + 55, + 176, + 296, + 260 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^i \\cap \\mathcal{C}^j = \\emptyset \\quad \\forall i, j \\in \\{0, \\dots, T\\}, i \\neq j" + }, + { + "bbox": [ + 55, + 176, + 296, + 260 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 261, + 296, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 261, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 261, + 296, + 357 + ], + "type": "text", + "content": "The model in few-shot class-incremental learning is expected to maintain classification accuracy across all base classes while achieving precise classification for novel classes. When leveraging a large pre-trained model as the feature extractor, it can extract rich and informative features for both base and novel classes attributing to its generalization capability. Consequently, obtaining a well-defined prototype is often sufficient for classification." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 357, + 296, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 296, + 417 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 296, + 417 + ], + "type": "text", + "content": "For base classes, prototypes can be easily derived by aggregating abundant samples, as demonstrated by existing methods like [21]. Specifically, each class " + }, + { + "bbox": [ + 55, + 357, + 296, + 417 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 357, + 296, + 417 + ], + "type": "text", + "content": " is represented by a prototype vector " + }, + { + "bbox": [ + 55, + 357, + 296, + 417 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_c\\in \\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 357, + 296, + 417 + ], + "type": "text", + "content": ", computed as the mean feature of all samples belonging to that class:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 117, + 422, + 295, + 452 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 422, + 295, + 452 + ], + "spans": [ + { + "bbox": [ + 117, + 422, + 295, + 452 + ], + "type": "interline_equation", + "content": "\\mathbf {p} _ {c} = \\frac {1}{| \\mathcal {I} _ {c} |} \\sum_ {\\left(\\mathbf {x} _ {i}, y _ {i}\\right) \\in \\mathcal {I} _ {c}} f _ {\\theta} \\left(\\mathbf {x} _ {i}\\right), \\tag {1}", + "image_path": "16c61a4d9ee824cb7d4c8d78035d93a77943346dadb6682595135a587e3b2734.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 456, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 296, + 491 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 456, + 296, + 491 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 55, + 456, + 296, + 491 + ], + "type": "text", + "content": " denotes the frozen feature extractor (large pretrained model like CLIP in this work), and " + }, + { + "bbox": [ + 55, + 456, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_c" + }, + { + "bbox": [ + 55, + 456, + 296, + 491 + ], + "type": "text", + "content": " contains all available samples for class " + }, + { + "bbox": [ + 55, + 456, + 296, + 491 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 456, + 296, + 491 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 492, + 296, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 492, + 296, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 492, + 296, + 563 + ], + "type": "text", + "content": "For novel classes at session " + }, + { + "bbox": [ + 55, + 492, + 296, + 563 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 55, + 492, + 296, + 563 + ], + "type": "text", + "content": ", the prototypes can also be derived from few-shot samples following Eq. (1). However, these prototypes are often biased due to the limited number of training examples. Such bias can result in suboptimal classification performance, presenting a significant challenge in FSCIL." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 564, + 296, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 296, + 624 + ], + "type": "text", + "content": "During inference at session " + }, + { + "bbox": [ + 55, + 564, + 296, + 624 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 55, + 564, + 296, + 624 + ], + "type": "text", + "content": ", the classification probability for a test sample " + }, + { + "bbox": [ + 55, + 564, + 296, + 624 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 564, + 296, + 624 + ], + "type": "text", + "content": " is calculated by measuring the distance between its feature representation " + }, + { + "bbox": [ + 55, + 564, + 296, + 624 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 564, + 296, + 624 + ], + "type": "text", + "content": " and the prototype vectors of all classes encountered (i.e. " + }, + { + "bbox": [ + 55, + 564, + 296, + 624 + ], + "type": "inline_equation", + "content": "\\bigcup_{i=0}^{t} \\mathcal{C}^i" + }, + { + "bbox": [ + 55, + 564, + 296, + 624 + ], + "type": "text", + "content": ") so far." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 633, + 111, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 633, + 111, + 646 + ], + "spans": [ + { + "bbox": [ + 55, + 633, + 111, + 646 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "To mitigate semantic bias in novel-class prototypes derived from scarce samples, we propose a semantically feature decomposition-recomposition approach for calibrating the prototype in a fine-grained manner at keyword attribute level. An overview is illustrated in Fig. 1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "content": "Our proposed method consists of two stages: 1) semantic feature decomposition stage, which establishes a grounding of text attribute keywords to CLIP features using base classes. It produces a set of meaningful feature segments, each corresponding to a specific semantic keyword. These feature segments are transferable across classes, ensuring their help to novel classes. 2) Feature segments recomposition stage, which adaptively calibrates the prototypes of novel classes by recomposing the semantic-aware feature segments according to the keywords associated with the novel classes. This process enables fine-grained refinement of the prototypes at keyword attribute level, mitigating bias and improving classification performance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 239, + 487, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 487, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 487, + 251 + ], + "type": "text", + "content": "4.1. Semantic feature decomposition" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 256, + 554, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 256, + 554, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 256, + 554, + 340 + ], + "type": "text", + "content": "In this stage, for each base class " + }, + { + "bbox": [ + 313, + 256, + 554, + 340 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 256, + 554, + 340 + ], + "type": "text", + "content": ", we begin by generating visual descriptions and attribute keywords using Large Language Models (LLMs). Next, leveraging the vision-language alignment capabilities of CLIP models, a fine-grained mapping from these keywords to visual features is established, producing a set of semantically meaningful feature segments." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 341, + 554, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 554, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 554, + 460 + ], + "type": "text", + "content": "Firstly, with the help of LLMs, a textual description " + }, + { + "bbox": [ + 313, + 341, + 554, + 460 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_c" + }, + { + "bbox": [ + 313, + 341, + 554, + 460 + ], + "type": "text", + "content": " for each class " + }, + { + "bbox": [ + 313, + 341, + 554, + 460 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 341, + 554, + 460 + ], + "type": "text", + "content": " is generated conditioned on class names (e.g. \"toaster\"). For instance, " + }, + { + "bbox": [ + 313, + 341, + 554, + 460 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_c" + }, + { + "bbox": [ + 313, + 341, + 554, + 460 + ], + "type": "text", + "content": " might be \"A rectangular kitchen appliance with metal or plastic housing, slots for inserting bread, and controls like dials Buttons.\" Additionally, " + }, + { + "bbox": [ + 313, + 341, + 554, + 460 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 341, + 554, + 460 + ], + "type": "text", + "content": " attribute keywords " + }, + { + "bbox": [ + 313, + 341, + 554, + 460 + ], + "type": "inline_equation", + "content": "\\mathcal{K}_C = \\{k_1,\\dots,k_m\\}" + }, + { + "bbox": [ + 313, + 341, + 554, + 460 + ], + "type": "text", + "content": " are extracted from " + }, + { + "bbox": [ + 313, + 341, + 554, + 460 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_c" + }, + { + "bbox": [ + 313, + 341, + 554, + 460 + ], + "type": "text", + "content": ". Each keyword " + }, + { + "bbox": [ + 313, + 341, + 554, + 460 + ], + "type": "inline_equation", + "content": "k_i" + }, + { + "bbox": [ + 313, + 341, + 554, + 460 + ], + "type": "text", + "content": " describes a visual attribute of the class " + }, + { + "bbox": [ + 313, + 341, + 554, + 460 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 341, + 554, + 460 + ], + "type": "text", + "content": " (e.g. \"rectangular\" and \"slots\" for class \"toaster\"). These textual descriptions and keywords capture the semantic attributes of the class." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 462, + 555, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 629 + ], + "type": "text", + "content": "Secondly, we design a keyword grounding function " + }, + { + "bbox": [ + 313, + 462, + 555, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\phi}" + }, + { + "bbox": [ + 313, + 462, + 555, + 629 + ], + "type": "text", + "content": " to establish fine-grained mappings from these keywords to visual feature segments of base prototypes. Leveraging CLIP's frozen visual encoder " + }, + { + "bbox": [ + 313, + 462, + 555, + 629 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 313, + 462, + 555, + 629 + ], + "type": "text", + "content": " and textual encoder " + }, + { + "bbox": [ + 313, + 462, + 555, + 629 + ], + "type": "inline_equation", + "content": "g_{\\phi}" + }, + { + "bbox": [ + 313, + 462, + 555, + 629 + ], + "type": "text", + "content": ", we obtain aligned visual prototype " + }, + { + "bbox": [ + 313, + 462, + 555, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_c" + }, + { + "bbox": [ + 313, + 462, + 555, + 629 + ], + "type": "text", + "content": " (same as Eq. (1)) and textual representations " + }, + { + "bbox": [ + 313, + 462, + 555, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_c = g_{\\phi}(\\mathcal{D}_c)" + }, + { + "bbox": [ + 313, + 462, + 555, + 629 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 462, + 555, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_c \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 462, + 555, + 629 + ], + "type": "text", + "content": " denotes the text embedding of description " + }, + { + "bbox": [ + 313, + 462, + 555, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_c" + }, + { + "bbox": [ + 313, + 462, + 555, + 629 + ], + "type": "text", + "content": ". Since CLIP's text and image embeddings are aligned, mapping a keyword " + }, + { + "bbox": [ + 313, + 462, + 555, + 629 + ], + "type": "inline_equation", + "content": "k_i" + }, + { + "bbox": [ + 313, + 462, + 555, + 629 + ], + "type": "text", + "content": " to a segment of the text feature space inherently implies a corresponding mapping to the associated segment in the image feature space. Therefore, for each keyword " + }, + { + "bbox": [ + 313, + 462, + 555, + 629 + ], + "type": "inline_equation", + "content": "k_i \\in \\mathcal{K}_C" + }, + { + "bbox": [ + 313, + 462, + 555, + 629 + ], + "type": "text", + "content": ", the keyword grounding function " + }, + { + "bbox": [ + 313, + 462, + 555, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\phi}" + }, + { + "bbox": [ + 313, + 462, + 555, + 629 + ], + "type": "text", + "content": " generates a grounding vector " + }, + { + "bbox": [ + 313, + 462, + 555, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_i" + }, + { + "bbox": [ + 313, + 462, + 555, + 629 + ], + "type": "text", + "content": " by evaluating the relevance of the keyword to the entire textual representation:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 376, + 640, + 555, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 640, + 555, + 654 + ], + "spans": [ + { + "bbox": [ + 376, + 640, + 555, + 654 + ], + "type": "interline_equation", + "content": "\\mathbf {g} _ {i} = \\mathcal {G} _ {\\phi} \\left(k _ {i}, \\mathcal {D} _ {c}\\right), \\in \\{0, 1 \\} ^ {d}. \\tag {2}", + "image_path": "4a3e6742cb035e21c818c611a28183e6fa513977958ee291c2797f10eaa42b1e.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_i" + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": " is a sparse binary mask, where entries with a value of 1 indicate the dimensions at which the keyword " + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "inline_equation", + "content": "k_{i}" + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": " contributes most significantly to the text embedding " + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_c" + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": " of description " + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_c" + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3155" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 71, + 550, + 251 + ], + "blocks": [ + { + "bbox": [ + 60, + 71, + 550, + 251 + ], + "lines": [ + { + "bbox": [ + 60, + 71, + 550, + 251 + ], + "spans": [ + { + "bbox": [ + 60, + 71, + 550, + 251 + ], + "type": "image", + "image_path": "9c6130e2000b378c177614f4f80edfe524f8aec6579e878daf9c0a57c951b6e5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 255, + 279, + 265 + ], + "lines": [ + { + "bbox": [ + 138, + 255, + 279, + 265 + ], + "spans": [ + { + "bbox": [ + 138, + 255, + 279, + 265 + ], + "type": "text", + "content": "Feature Decomposition of Base classes" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 360, + 255, + 548, + 265 + ], + "lines": [ + { + "bbox": [ + 360, + 255, + 548, + 265 + ], + "spans": [ + { + "bbox": [ + 360, + 255, + 548, + 265 + ], + "type": "text", + "content": "Segments Decomposition for calibrating Novel class" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 275, + 555, + 384 + ], + "lines": [ + { + "bbox": [ + 55, + 275, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 555, + 384 + ], + "type": "text", + "content": "Figure 1. Overall Architecture of Our Feature Decomposition-Recomposition (FDR) method. The calibration process is as follows: 1) Feature Decomposition of Base classes: For all classes, LLMs are leveraged to generate category descriptions and extract semantic keywords. CLIP encodes both image features (aggregated as prototypes) and textual descriptions. A Keyword Grounding Function then decomposes textual embeddings into semantic-aware feature segments aligned with specific keywords, achieving attribute-level decomposition of visual and textual features. 2) Segments Recomposition for calibrating Novel class: For novel classes, we match each keyword-derived textual feature segment with semantically similar segments from base classes. The corresponding visual feature segments from multiple base prototypes are recomposed to calibrate the raw novel prototype. For instance, when calibrating the cannon prototype (which suffers from missing wheel features due to limited samples), our method identifies wheeled segments from unicycle and oxcart prototypes and cylindrical segments from lipstick and water tower prototypes. This fine-grained recomposition supplements missing attributes while preserving class-discriminative characteristics." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 396, + 295, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 295, + 503 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 295, + 503 + ], + "type": "text", + "content": "During the text encoding process of CLIP, the embedding of the entire text description " + }, + { + "bbox": [ + 55, + 396, + 295, + 503 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_c" + }, + { + "bbox": [ + 55, + 396, + 295, + 503 + ], + "type": "text", + "content": " inherently incorporates the contributions of tokens from each word including those visual keywords " + }, + { + "bbox": [ + 55, + 396, + 295, + 503 + ], + "type": "inline_equation", + "content": "k_{i}\\in \\mathcal{K}_{C}" + }, + { + "bbox": [ + 55, + 396, + 295, + 503 + ], + "type": "text", + "content": ". This means that the attention mechanisms and other internal structures implicitly reveal how and where each keyword contributes to the overall textual representation. Therefore, the keyword grounding function " + }, + { + "bbox": [ + 55, + 396, + 295, + 503 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\phi}" + }, + { + "bbox": [ + 55, + 396, + 295, + 503 + ], + "type": "text", + "content": " is designed to leverage CLIP's internal attention mechanisms to identify these contributions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 504, + 295, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 504, + 295, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 504, + 295, + 574 + ], + "type": "text", + "content": "Specifically, in the last transformer layer of CLIP textual encoder " + }, + { + "bbox": [ + 55, + 504, + 295, + 574 + ], + "type": "inline_equation", + "content": "g_{\\phi}" + }, + { + "bbox": [ + 55, + 504, + 295, + 574 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 504, + 295, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_c" + }, + { + "bbox": [ + 55, + 504, + 295, + 574 + ], + "type": "text", + "content": " is tokenized as " + }, + { + "bbox": [ + 55, + 504, + 295, + 574 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 504, + 295, + 574 + ], + "type": "text", + "content": " tokens " + }, + { + "bbox": [ + 55, + 504, + 295, + 574 + ], + "type": "inline_equation", + "content": "\\{[SOS], \\text{token}_1, \\dots, \\text{token}_{n-2}, [EOS]\\}" + }, + { + "bbox": [ + 55, + 504, + 295, + 574 + ], + "type": "text", + "content": ". Given the attention matrix " + }, + { + "bbox": [ + 55, + 504, + 295, + 574 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in \\mathbb{R}^{n \\times n}" + }, + { + "bbox": [ + 55, + 504, + 295, + 574 + ], + "type": "text", + "content": " and value matrix " + }, + { + "bbox": [ + 55, + 504, + 295, + 574 + ], + "type": "inline_equation", + "content": "\\mathbf{V} \\in \\mathbb{R}^{d \\times n}" + }, + { + "bbox": [ + 55, + 504, + 295, + 574 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 504, + 295, + 574 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_c" + }, + { + "bbox": [ + 55, + 504, + 295, + 574 + ], + "type": "text", + "content": " is calculated as the output of the last transformer layer as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 585, + 295, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 585, + 295, + 597 + ], + "spans": [ + { + "bbox": [ + 61, + 585, + 295, + 597 + ], + "type": "interline_equation", + "content": "\\mathbf {t} _ {c} = \\mathbf {V} (:,:) \\mathbf {A} (:, e o s) + \\mathbf {h} = \\left(a _ {1} \\mathbf {v} _ {1} + \\dots + a _ {n} \\mathbf {v} _ {n}\\right) + \\mathbf {h}, \\tag {3}", + "image_path": "fd882f602c6645cee9252a5173111c991ef5a5837c75514506011400a219bf66.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 605, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 55, + 605, + 295, + 665 + ], + "type": "text", + "content": " represents the inter-token attention weights, " + }, + { + "bbox": [ + 55, + 605, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 55, + 605, + 295, + 665 + ], + "type": "text", + "content": " represents transformed token representations of all words, " + }, + { + "bbox": [ + 55, + 605, + 295, + 665 + ], + "type": "inline_equation", + "content": "a_{i} = A[i,p_{\\mathrm{eos}}]" + }, + { + "bbox": [ + 55, + 605, + 295, + 665 + ], + "type": "text", + "content": " represents the attention weight from the " + }, + { + "bbox": [ + 55, + 605, + 295, + 665 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 605, + 295, + 665 + ], + "type": "text", + "content": "-th token to the [EOS] token, and " + }, + { + "bbox": [ + 55, + 605, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i" + }, + { + "bbox": [ + 55, + 605, + 295, + 665 + ], + "type": "text", + "content": " means the value vector of the " + }, + { + "bbox": [ + 55, + 605, + 295, + 665 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 605, + 295, + 665 + ], + "type": "text", + "content": "-th token from CLIP's final transformer layer." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 666, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 712 + ], + "type": "text", + "content": "As can be seen from the calculation of " + }, + { + "bbox": [ + 55, + 666, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_c" + }, + { + "bbox": [ + 55, + 666, + 296, + 712 + ], + "type": "text", + "content": " in Eq. (3), the attention matrix " + }, + { + "bbox": [ + 55, + 666, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 55, + 666, + 296, + 712 + ], + "type": "text", + "content": " and value matrix " + }, + { + "bbox": [ + 55, + 666, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 55, + 666, + 296, + 712 + ], + "type": "text", + "content": " implicitly reveal how and where each keyword contributes to the overall textual representation. Therefore, we extract these token" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 396, + 553, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 553, + 418 + ], + "type": "text", + "content": "wise contribution values as a contribution indication matrix " + }, + { + "bbox": [ + 313, + 396, + 553, + 418 + ], + "type": "inline_equation", + "content": "\\mathbf{B}\\in \\mathbb{R}^{d\\times n}" + }, + { + "bbox": [ + 313, + 396, + 553, + 418 + ], + "type": "text", + "content": " .." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 384, + 420, + 553, + 433 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 420, + 553, + 433 + ], + "spans": [ + { + "bbox": [ + 384, + 420, + 553, + 433 + ], + "type": "interline_equation", + "content": "\\mathbf {B} = \\left[ a _ {1} \\mathbf {v} _ {1}, \\dots , a _ {n} \\mathbf {v} _ {n} \\right]. \\tag {4}", + "image_path": "b27bcc9b637fd59e43ee6b00d1664e2c07f48f6bfe26953020a87ffefff37357.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 437, + 554, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 496 + ], + "type": "text", + "content": "Here, the columns of " + }, + { + "bbox": [ + 313, + 437, + 554, + 496 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 313, + 437, + 554, + 496 + ], + "type": "text", + "content": " describe different text tokens, while the rows represent feature channels. Each element " + }, + { + "bbox": [ + 313, + 437, + 554, + 496 + ], + "type": "inline_equation", + "content": "\\mathbf{B}(j,i)" + }, + { + "bbox": [ + 313, + 437, + 554, + 496 + ], + "type": "text", + "content": " indicates the contribution of the " + }, + { + "bbox": [ + 313, + 437, + 554, + 496 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 437, + 554, + 496 + ], + "type": "text", + "content": "-th token's " + }, + { + "bbox": [ + 313, + 437, + 554, + 496 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 437, + 554, + 496 + ], + "type": "text", + "content": "-th feature dimension to the final [EOS] embedding, serving as an effective indicator for keyword-related feature dimensions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 497, + 554, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 554, + 567 + ], + "type": "text", + "content": "Therefore, to quantify the influence of each token on the text embedding " + }, + { + "bbox": [ + 313, + 497, + 554, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 313, + 497, + 554, + 567 + ], + "type": "text", + "content": ", we apply a softmax operation along the token dimension (horizontal) of " + }, + { + "bbox": [ + 313, + 497, + 554, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 313, + 497, + 554, + 567 + ], + "type": "text", + "content": ". This results in a matrix " + }, + { + "bbox": [ + 313, + 497, + 554, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 313, + 497, + 554, + 567 + ], + "type": "text", + "content": ", where each element represents the proportion of token " + }, + { + "bbox": [ + 313, + 497, + 554, + 567 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 497, + 554, + 567 + ], + "type": "text", + "content": "'s contribution to " + }, + { + "bbox": [ + 313, + 497, + 554, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_c" + }, + { + "bbox": [ + 313, + 497, + 554, + 567 + ], + "type": "text", + "content": " at feature dimension " + }, + { + "bbox": [ + 313, + 497, + 554, + 567 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 497, + 554, + 567 + ], + "type": "text", + "content": ", relative to all text tokens:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 375, + 573, + 553, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 573, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 375, + 573, + 553, + 586 + ], + "type": "interline_equation", + "content": "\\mathbf {C} (i,:) = \\operatorname {S o f t m a x} (\\mathbf {B} (i,:)) \\tag {5}", + "image_path": "7b7397665367633670f2ee4a87251868a8f3c704ce9b98390adbcfdaf027898a.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 594, + 554, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 554, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 554, + 640 + ], + "type": "text", + "content": "Then, we perform top-q binarization along the feature dimension (vertical) of " + }, + { + "bbox": [ + 313, + 594, + 554, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 313, + 594, + 554, + 640 + ], + "type": "text", + "content": ", converting each column " + }, + { + "bbox": [ + 313, + 594, + 554, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{C}(:, i]" + }, + { + "bbox": [ + 313, + 594, + 554, + 640 + ], + "type": "text", + "content": " into into a sparse binary mask " + }, + { + "bbox": [ + 313, + 594, + 554, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_i" + }, + { + "bbox": [ + 313, + 594, + 554, + 640 + ], + "type": "text", + "content": " (previously introduced in Eq. (2)) as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 647, + 563, + 689 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 563, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 563, + 689 + ], + "type": "interline_equation", + "content": "\\mathbf {g} _ {i} (j) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} \\mathbf {C} (i, j) \\text {i s w i t h i n t o p - q r e s p o n s i v e o f} \\mathbf {C} (:, j) \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {6}", + "image_path": "ce3fbedc19928e6f2bb5b292f438bcb747637ff4c7828d4efeb8545ece6fdebd.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "This process isolates the most significant feature dimensions associated with each keyword, enabling grounding of" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3156" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 226, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 226, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 226, + 83 + ], + "type": "text", + "content": "textual keywords to text feature segments." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 84, + 296, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 84, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 84, + 296, + 228 + ], + "type": "text", + "content": "According to Eqs. (2) to (6), we can derive a fine-grained mapping from each keyword to its corresponding feature segments. While this mapping is computed independently for each class, ideally, the same keyword associated with different classes should yield identical mappings. However, in practice, slight discrepancies often arise due to variations in class contexts. To ensure consistency, for keywords that appear across multiple base classes, we unify all selected feature dimensions by taking the union of their respective binary masks. This strategy preserves all potentially relevant dimensions for the keyword, enhancing robustness and consistency in the grounding process." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 228, + 296, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 296, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 296, + 276 + ], + "type": "text", + "content": "Since CLIP's visual and textual feature spaces are aligned, given a keyword " + }, + { + "bbox": [ + 55, + 228, + 296, + 276 + ], + "type": "inline_equation", + "content": "k_{i}" + }, + { + "bbox": [ + 55, + 228, + 296, + 276 + ], + "type": "text", + "content": ", we can obtain the visual feature segment " + }, + { + "bbox": [ + 55, + 228, + 296, + 276 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_i^c" + }, + { + "bbox": [ + 55, + 228, + 296, + 276 + ], + "type": "text", + "content": " of prototype " + }, + { + "bbox": [ + 55, + 228, + 296, + 276 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_c" + }, + { + "bbox": [ + 55, + 228, + 296, + 276 + ], + "type": "text", + "content": ", by simply performing an element-wise multiplication as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 145, + 281, + 295, + 294 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 281, + 295, + 294 + ], + "spans": [ + { + "bbox": [ + 145, + 281, + 295, + 294 + ], + "type": "interline_equation", + "content": "\\mathbf {s} _ {i} ^ {c} = \\mathbf {p} _ {c} \\odot \\mathbf {g} _ {i}. \\tag {7}", + "image_path": "3cad6e2de9197eb3302614240931f14339258789dac3caa8a0e818060c0249d4.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 299, + 296, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 299, + 296, + 359 + ], + "spans": [ + { + "bbox": [ + 55, + 299, + 296, + 359 + ], + "type": "text", + "content": "In this way, we achieve semantic-aware decomposition of base class prototypes guided by textual keywords of visual attributes, where each feature segment " + }, + { + "bbox": [ + 55, + 299, + 296, + 359 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_i^c" + }, + { + "bbox": [ + 55, + 299, + 296, + 359 + ], + "type": "text", + "content": " retains only the dimensions most relevant to the corresponding semantic keyword." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 365, + 295, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 389 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 389 + ], + "type": "text", + "content": "4.2. Feature segments decomposition for fine-grained calibration" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 394, + 296, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 394, + 296, + 561 + ], + "spans": [ + { + "bbox": [ + 55, + 394, + 296, + 561 + ], + "type": "text", + "content": "The prototypes of novel classes are often biased due to the limited number of shots available per class. While the full-dimensional features of these prototypes can only be derived from the novel class samples, certain partial features—such as visual attribute features—frequently overlap with those of base class prototypes. For instance, features like \"wheels\" are shared between the novel class \"cannon\" and base classes \"unicycle\", \"oxcart\", and \"race car\". Leveraging this inherent property, we calibrate the prototypes of novel classes by piecewise refinement of their semantic feature segments. This is achieved by assembling similar segments from the base classes, thereby enhancing the representational accuracy and reducing bias in the novel class prototypes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 561, + 295, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 561, + 295, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 561, + 295, + 597 + ], + "type": "text", + "content": "In this process, given any novel class " + }, + { + "bbox": [ + 55, + 561, + 295, + 597 + ], + "type": "inline_equation", + "content": "\\widetilde{C}" + }, + { + "bbox": [ + 55, + 561, + 295, + 597 + ], + "type": "text", + "content": ", we first generate its initial prototype " + }, + { + "bbox": [ + 55, + 561, + 295, + 597 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{\\widetilde{C}}" + }, + { + "bbox": [ + 55, + 561, + 295, + 597 + ], + "type": "text", + "content": " and visual keywords " + }, + { + "bbox": [ + 55, + 561, + 295, + 597 + ], + "type": "inline_equation", + "content": "\\kappa_{\\widetilde{C}}" + }, + { + "bbox": [ + 55, + 561, + 295, + 597 + ], + "type": "text", + "content": " following the same process as base classes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 597, + 296, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 296, + 657 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 296, + 657 + ], + "type": "text", + "content": "Next, for each keyword " + }, + { + "bbox": [ + 55, + 597, + 296, + 657 + ], + "type": "inline_equation", + "content": "k_{i} \\in \\mathcal{K}_{\\widetilde{C}}" + }, + { + "bbox": [ + 55, + 597, + 296, + 657 + ], + "type": "text", + "content": ", we retrieve the corresponding feature segments " + }, + { + "bbox": [ + 55, + 597, + 296, + 657 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_i^{c_j}" + }, + { + "bbox": [ + 55, + 597, + 296, + 657 + ], + "type": "text", + "content": " from the base-class prototypes that share the same keywords. These segments are then combined using a weighted summation to calibrate the feature segment of the novel class:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 662, + 295, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 662, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 130, + 662, + 295, + 685 + ], + "type": "interline_equation", + "content": "\\mathbf {r} _ {i} = \\sum_ {j = 1} ^ {M} w _ {j} \\cdot \\mathbf {s} _ {i} ^ {c _ {j}}. \\tag {8}", + "image_path": "2f28d6b15d26364e73593c8b510feb47e098891666751c74024ef6d021ade400.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Here, the " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_i^{c_j}" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": " represents the prototype segment of base class " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "C_j" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": " corresponding to the keyword " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "k_{i}" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": ". The weight " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "w_{j}" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": " reflects" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 318, + 73, + 552, + 293 + ], + "blocks": [ + { + "bbox": [ + 318, + 73, + 552, + 293 + ], + "lines": [ + { + "bbox": [ + 318, + 73, + 552, + 293 + ], + "spans": [ + { + "bbox": [ + 318, + 73, + 552, + 293 + ], + "type": "image", + "image_path": "87a5604a38b60b14516f41952a8c2ecce84ee4c5b91d57742768105e91385c2b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 304, + 555, + 338 + ], + "lines": [ + { + "bbox": [ + 313, + 304, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 555, + 338 + ], + "type": "text", + "content": "Figure 2. Implementation of keyword grounding function " + }, + { + "bbox": [ + 313, + 304, + 555, + 338 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\phi}" + }, + { + "bbox": [ + 313, + 304, + 555, + 338 + ], + "type": "text", + "content": ". The upper describes the computation of " + }, + { + "bbox": [ + 313, + 304, + 555, + 338 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_c" + }, + { + "bbox": [ + 313, + 304, + 555, + 338 + ], + "type": "text", + "content": " using CLIP, while the lower details the process for generating the grounding vector " + }, + { + "bbox": [ + 313, + 304, + 555, + 338 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_c" + }, + { + "bbox": [ + 313, + 304, + 555, + 338 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 363, + 554, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 554, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 554, + 472 + ], + "type": "text", + "content": "the similarity between the feature segments of " + }, + { + "bbox": [ + 313, + 363, + 554, + 472 + ], + "type": "inline_equation", + "content": "\\widetilde{C}" + }, + { + "bbox": [ + 313, + 363, + 554, + 472 + ], + "type": "text", + "content": " and base class " + }, + { + "bbox": [ + 313, + 363, + 554, + 472 + ], + "type": "inline_equation", + "content": "C_j" + }, + { + "bbox": [ + 313, + 363, + 554, + 472 + ], + "type": "text", + "content": ". This similarity is approximated by computing the textual embedding similarity between the textual segment of novel class " + }, + { + "bbox": [ + 313, + 363, + 554, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{\\widetilde{c}} \\odot \\mathbf{g}_i" + }, + { + "bbox": [ + 313, + 363, + 554, + 472 + ], + "type": "text", + "content": " and that of every base class " + }, + { + "bbox": [ + 313, + 363, + 554, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{c_j} \\odot \\mathbf{g}_i, C_j \\in C^0" + }, + { + "bbox": [ + 313, + 363, + 554, + 472 + ], + "type": "text", + "content": ". The top-r similarity scores are preserved after using a softmax function, while the remaining similarity scores are set to zero. This ensures that only the most semantically relevant base-class feature segments are retained, yielding the final weights " + }, + { + "bbox": [ + 313, + 363, + 554, + 472 + ], + "type": "inline_equation", + "content": "w_j" + }, + { + "bbox": [ + 313, + 363, + 554, + 472 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 474, + 554, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 554, + 522 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 554, + 522 + ], + "type": "text", + "content": "After obtaining feature segments for all keywords associated with novel class " + }, + { + "bbox": [ + 313, + 474, + 554, + 522 + ], + "type": "inline_equation", + "content": "\\widetilde{C}" + }, + { + "bbox": [ + 313, + 474, + 554, + 522 + ], + "type": "text", + "content": ", keyword-specific adjustments are iteratively applied to novel prototype " + }, + { + "bbox": [ + 313, + 474, + 554, + 522 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{\\widetilde{c}}" + }, + { + "bbox": [ + 313, + 474, + 554, + 522 + ], + "type": "text", + "content": " to obtain the final calibrated prototype as below:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 342, + 552, + 555, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 552, + 555, + 564 + ], + "spans": [ + { + "bbox": [ + 342, + 552, + 555, + 564 + ], + "type": "interline_equation", + "content": "\\mathbf {p} _ {\\tilde {c}} \\leftarrow (1 - \\alpha) \\cdot \\mathbf {p} _ {\\tilde {c}} + \\alpha \\cdot \\mathbf {r} _ {i}, \\quad i \\in \\{1, \\dots , m \\} \\tag {9}", + "image_path": "de9dd83ccc8b32e54d27815c6e6eb62c1544d78108a51ff1763030d0cdde493a.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 579, + 554, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 579, + 554, + 603 + ], + "spans": [ + { + "bbox": [ + 313, + 579, + 554, + 603 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 579, + 554, + 603 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 579, + 554, + 603 + ], + "type": "text", + "content": " is a trade-off hyperparameter that balances the prototype directly from the few shots and calibrated one." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 605, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 712 + ], + "type": "text", + "content": "In Eq. (9), each calibration " + }, + { + "bbox": [ + 313, + 605, + 555, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_i" + }, + { + "bbox": [ + 313, + 605, + 555, + 712 + ], + "type": "text", + "content": " is performed individually and is a linear combination of similar feature segments from base-class prototypes. Importantly, this linear combination varies for different feature segments, as the weights and contributing segments depend on the specific keyword " + }, + { + "bbox": [ + 313, + 605, + 555, + 712 + ], + "type": "inline_equation", + "content": "k_i" + }, + { + "bbox": [ + 313, + 605, + 555, + 712 + ], + "type": "text", + "content": ". Consequently, the overall calibration is semantic-segmentwise linear but nonlinear overall, ensuring fine-grained and more accurate prototype calibration for few-shot novel classes." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3157" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 86, + 552, + 216 + ], + "blocks": [ + { + "bbox": [ + 69, + 71, + 541, + 82 + ], + "lines": [ + { + "bbox": [ + 69, + 71, + 541, + 82 + ], + "spans": [ + { + "bbox": [ + 69, + 71, + 541, + 82 + ], + "type": "text", + "content": "Table 1. Comparison with existing methods (5-shot setting). " + }, + { + "bbox": [ + 69, + 71, + 541, + 82 + ], + "type": "inline_equation", + "content": "A_{base}" + }, + { + "bbox": [ + 69, + 71, + 541, + 82 + ], + "type": "text", + "content": " for base session accuracy, " + }, + { + "bbox": [ + 69, + 71, + 541, + 82 + ], + "type": "inline_equation", + "content": "A_{avg}" + }, + { + "bbox": [ + 69, + 71, + 541, + 82 + ], + "type": "text", + "content": " for AvgAcc across all sessions" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 86, + 552, + 216 + ], + "lines": [ + { + "bbox": [ + 59, + 86, + 552, + 216 + ], + "spans": [ + { + "bbox": [ + 59, + 86, + 552, + 216 + ], + "type": "table", + "html": "
MethodBackboneTraining FreeCIFAR100miniImageNetCUB200IN1K-FSCIL
\\(A_{base}\\)\\(A_{avg}\\)\\(A_{base}\\)\\(A_{avg}\\)\\(A_{base}\\)\\(A_{avg}\\)\\(A_{base}\\)\\(A_{avg}\\)
CLOSER[2]ResNet-18/20No75.763.176.062.879.469.1--
Yourself[6]ViT-B-IN21KNo82.967.084.068.883.469.9--
LP-DiF[4]ViT-B-CLIPNo80.275.196.393.883.974.0--
PriViLego[5]ViT-B-CLIPNo84.277.295.492.782.174.272.168.9
EASE[3]ViT-B-IN21KNo85.178.096.893.983.075.173.370.4
FDR(Ours)ViT-B-CLIPNo87.779.395.389.284.477.773.670.9
BaselineViT-B-CLIPYes75.767.992.088.680.973.669.265.9
TEEN[1]ViT-B-CLIPYes75.768.492.088.780.973.769.266.1
FDR(Ours)ViT-B-CLIPYes75.768.592.089.180.974.169.266.4
", + "image_path": "040f950dfd326803f0fd3876105c9c9091421288d4cb4d82d60b468592b80450.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 58, + 251, + 552, + 327 + ], + "blocks": [ + { + "bbox": [ + 55, + 225, + 553, + 247 + ], + "lines": [ + { + "bbox": [ + 55, + 225, + 553, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 225, + 553, + 247 + ], + "type": "text", + "content": "Table 2. Comparison with existing methods on CUB200-1-shot benchmark (Accuracy %). Bold indicates best performance. " + }, + { + "bbox": [ + 55, + 225, + 553, + 247 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 55, + 225, + 553, + 247 + ], + "type": "text", + "content": " shows improvement of our method over the previous best. Due to space limitations, only the first three and last three sessions are shown." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 251, + 552, + 327 + ], + "lines": [ + { + "bbox": [ + 58, + 251, + 552, + 327 + ], + "spans": [ + { + "bbox": [ + 58, + 251, + 552, + 327 + ], + "type": "table", + "html": "
Session Method01238910
AvgAccAvgAccNAccAvgAccNAccAvgAccNAccAvgAccNAccAvgAccNAccAvgAccNAcc
Baseline [21]80.8574.1814.2568.8618.6665.0018.3952.1218.2251.6420.3049.7519.86
TEEN [1]80.8576.4035.1671.8938.2666.7030.0056.6029.7556.1432.1954.4831.56
FDR(Ours)80.8577.2654.8272.9350.0968.3839.7458.3336.4558.7638.9157.3738.40
Δ-+0.86+19.66+1.04+11.83+1.68+9.74+1.73+6.70+2.62+6.72+2.89+6.84
", + "image_path": "80f66db2c85eb94da4bf1ba10a3ba77d2c0f120e3cb384b13da4c5fe25e5294f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 346, + 136, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 136, + 358 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 136, + 358 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 365, + 295, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 425 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 425 + ], + "type": "text", + "content": "This section begins by detailing the datasets and implementation specifics. Subsequently, it presents a comparison with existing approaches to benchmark performance, followed by an ablation study to assess the effectiveness of individual components of our method." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 433, + 251, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 251, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 251, + 445 + ], + "type": "text", + "content": "5.1. Datasets and Implementation Details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 450, + 295, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 497 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 497 + ], + "type": "text", + "content": "The evaluations are conducted on four benchmarks, following standard Few-Shot Class-Incremental Learning (FSCIL) protocols: CIFAR100, miniImageNet, CUB200, and IN1K-FSCIL." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 498, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 689 + ], + "type": "text", + "content": "CIFAR100, miniImageNet, and CUB200 are widely used benchmarks for FSCIL. CUB200 contains 200 bird species, while miniImageNet and CIFAR100 each comprise 100 common everyday object categories. This section details the experimental results on CUB200; results for the other two datasets are provided in the supplementary material. We adopt the same dataset split for CUB200 as used in existing works [1, 30, 32], denoted as CUB200-5-shot. Specifically, this follows an 11-session split: Session 0 uses 100 base classes (3,000 images), followed by 10 incremental sessions, each introducing 10 novel classes with 5-shot support per class. To further evaluate the effectiveness of prototype calibration under conditions of extreme data scarcity, we also evaluate a 10-way 1-shot configuration for CUB200, where only one sample per novel class is available. This configuration is denoted as CUB200-1-shot." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "IN1K-FSCIL. Furthermore, to assess scalability in real-world scenarios with more classes, we construct a challeng-" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 346, + 553, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 553, + 441 + ], + "type": "text", + "content": "ing ImageNet1K[11] variant by reorganizing 1,000 classes into 800 base classes (200 images per class) and 200 novel classes divided across 10 incremental sessions (20-way 1-shot), which is denoted as IN1K-FSCIL. This setting introduces higher inter-class density and diversity compared to existing benchmarks, better simulating practical applications requiring large-scale incremental learning. The results are provided in the supplementary material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 446, + 553, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 446, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 446, + 553, + 578 + ], + "type": "text", + "content": "Implementation: For our method, all experiments are implemented in PyTorch with CLIP-ViT/B-16 [17] as the backbone. 512-dimensional visual and textual features are extracted from CLIP's aligned embedding space, with the backbone frozen during both the base and incremental phases, making FDR a training-free method. To further bridge the domain gap between CLIP's pretraining data and the target dataset, we apply LoRA fine-tuning on CLIP during the base phase by leveraging an image-text contrastive loss, while refraining from any training in the incremental phase." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "text", + "content": "Evaluation Metrics: For evaluation, we adopt the commonly used metric of average accuracy (AvgAcc). Specifically, at session " + }, + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "text", + "content": ", the average classification accuracy is computed across all classes encountered up to that session (i.e. " + }, + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "inline_equation", + "content": "M + t * N" + }, + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "text", + "content": "). While AvgAcc provides a comprehensive measure of overall performance, it is less sensitive to the accuracy of newly learned novel classes. To address this, we also introduce the average accuracy for novel classes, denoted as NAcc. NAcc is calculated similarly to AvgAcc (i.e., classification over all classes), but the proportion of correctly classified instances is computed exclu" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3158" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 96, + 553, + 235 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 3. Comparison with existing methods on CUB200-5-shot benchmark (Accuracy %). Bold indicates best performance. " + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": " shows improvement of our method over the previous best. Due to space limitations, only the first three and last three sessions are shown." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 96, + 553, + 235 + ], + "lines": [ + { + "bbox": [ + 57, + 96, + 553, + 235 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 553, + 235 + ], + "type": "table", + "html": "
MethodBackboneAccuracy in each session (%)
01238910
AvgAccAvgAccNAccAvgAccNAccAvgAccNAccAvgAccNAccAvgAccNAccAvgAccNAcc
Baseline[21]ResNet1874.1468.0912.3361.869.5957.628.5645.558.8543.3310.1841.4810.06
TOPIC[22]ResNet1868.6862.49-54.81-49.99-32.22-28.31-26.26-
CEC[30]ResNet1875.8571.94-68.50-63.50-54.83-53.52-52.28-
FACT[32]ResNet1875.9073.23-70.84-66.13-58.41-57.89-56.94-
TEEN[1]ResNet1877.2676.13-72.81-68.16-61.19-60.32-59.31-
Baseline[21]CLIP80.8579.1165.0277.4965.0273.9857.2069.5356.1169.3558.3169.2059.13
TEEN [1]CLIP80.8579.2167.2377.5466.1873.8758.2169.5656.8669.4558.0169.3758.83
FDR(Ours)CLIP80.8579.3873.7677.7070.7374.2862.5170.1359.6870.0861.8170.1562.79
Δ-+0.17+6.53+0.16+4.55+0.30+4.30+0.57+2.82+0.63+3.80+0.78+3.96
", + "image_path": "76a5e151bf3b602fd6fc043a74df7db22c46206dba1052d731b3b03fdf6db414.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 245, + 181, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 245, + 181, + 256 + ], + "spans": [ + { + "bbox": [ + 55, + 245, + 181, + 256 + ], + "type": "text", + "content": "sively within the novel classes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 265, + 244, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 265, + 244, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 244, + 277 + ], + "type": "text", + "content": "5.2. Comparison with Existing Methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 282, + 295, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 282, + 295, + 377 + ], + "spans": [ + { + "bbox": [ + 55, + 282, + 295, + 377 + ], + "type": "text", + "content": "We have made comparisons against recent SOTA methods [2-6]. As shown in Table 1, our FDR method surpasses all training-free approaches, consistent with the results of the main paper. For approaches that require training[2-6], we apply the CLIP backbone fine-tuned on base classes. The fine-tuning significantly improves the base session accuracy of our method, making it outperform other methods on CIFAR100, CUB200, and IN1K-FSCIL." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 379, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 379, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 379, + 295, + 713 + ], + "type": "text", + "content": "Table 2 and Table 3 present the detailed comparison results on the CUB200 dataset. All methods utilize the same frozen CLIP backbone (without fine-tuning) or ResNet18 to ensure a fair comparison. The baseline represents vanilla Prototypical Networks [21], which directly uses CLIP features without prototype calibration. The recently proposed method, TEEN, demonstrates significantly better performance than the baseline, benefiting from its comprehensive prototype calibration. Furthermore, our method achieves superior performance across all sessions in both 5-shot and 1-shot settings. Notably, in the 1-shot setting, our method achieves a " + }, + { + "bbox": [ + 55, + 379, + 295, + 713 + ], + "type": "inline_equation", + "content": "6.70\\% \\sim 19.66\\%" + }, + { + "bbox": [ + 55, + 379, + 295, + 713 + ], + "type": "text", + "content": " improvement in NAcc across all incremental sessions when comparing with holistic calibration method [1], highlighting its effectiveness in bias correction even under extreme data scarcity. Additionally, the progressive increases in AvgAcc, with improvements ranging from " + }, + { + "bbox": [ + 55, + 379, + 295, + 713 + ], + "type": "inline_equation", + "content": "+0.86\\%" + }, + { + "bbox": [ + 55, + 379, + 295, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 379, + 295, + 713 + ], + "type": "inline_equation", + "content": "+2.89\\%" + }, + { + "bbox": [ + 55, + 379, + 295, + 713 + ], + "type": "text", + "content": ", suggest cumulative advantages from our composition strategy as the incremental sessions progress. present the comparison results on the CUB200 dataset. All methods utilize the same frozen CLIP backbone or ResNet18 to ensure a fair comparison. The baseline represents vanilla Prototypical Networks [21], which directly uses CLIP features without prototype calibration. The recently proposed method, TEEN, demonstrates significantly better performance than the baseline, benefiting from its comprehensive prototype calibration. Furthermore, our method achieves superior performance across all sessions in both 5-shot and 1-shot set" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 245, + 555, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 245, + 555, + 353 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 555, + 353 + ], + "type": "text", + "content": "tings. Notably, in the 1-shot setting, our method achieves a " + }, + { + "bbox": [ + 313, + 245, + 555, + 353 + ], + "type": "inline_equation", + "content": "6.70\\% \\sim 19.66\\%" + }, + { + "bbox": [ + 313, + 245, + 555, + 353 + ], + "type": "text", + "content": " improvement in NAcc across all incremental sessions when comparing with holistic calibration method [1], highlighting its effectiveness in bias correction even under extreme data scarcity. Additionally, the progressive increases in AvgAcc, with improvements ranging from " + }, + { + "bbox": [ + 313, + 245, + 555, + 353 + ], + "type": "inline_equation", + "content": "+0.86\\%" + }, + { + "bbox": [ + 313, + 245, + 555, + 353 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 245, + 555, + 353 + ], + "type": "inline_equation", + "content": "+2.89\\%" + }, + { + "bbox": [ + 313, + 245, + 555, + 353 + ], + "type": "text", + "content": ", suggest cumulative advantages from our composition strategy as the incremental sessions progress." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 324, + 377, + 545, + 456 + ], + "blocks": [ + { + "bbox": [ + 330, + 363, + 537, + 374 + ], + "lines": [ + { + "bbox": [ + 330, + 363, + 537, + 374 + ], + "spans": [ + { + "bbox": [ + 330, + 363, + 537, + 374 + ], + "type": "text", + "content": "Table 4. Ablation study on CUB200-1-shot (Session 10)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 324, + 377, + 545, + 456 + ], + "lines": [ + { + "bbox": [ + 324, + 377, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 324, + 377, + 545, + 456 + ], + "type": "table", + "html": "
MethodAvgAccNAcc
No Calibration49.7519.86
Holistic Calibration only54.4831.56
Random Feature Segments54.8332.08
Naive Keyword Grounding56.6136.25
Full (Ours)57.3738.40
", + "image_path": "99b7c41ac559c70868dd9c25ec3691cf02eb5dfa73e14a351b6be6b9d23290a6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 466, + 416, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 466, + 416, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 466, + 416, + 478 + ], + "type": "text", + "content": "5.3. Ablation Studies" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 483, + 554, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 554, + 555 + ], + "type": "text", + "content": "This section presents ablation experiments on the CUB200-1-shot dataset to quantify the contribution of each component and assess the impact of hyperparameters. The performance is evaluated using the final session's metrics, including the average accuracy (AvgAcc) for all classes and the novel class accuracy (NAcc) for all novel classes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 555, + 554, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 554, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 554, + 590 + ], + "type": "text", + "content": "Effectiveness of Component. We dissect the key components of our method through several controlled experiments:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 313, + 594, + 555, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 654 + ], + "type": "text", + "content": "- No Calibration: This strategy directly aggregates features of each class provided by the CLIP backbone to form prototypes. For novel classes, the prototype is calculated directly from the few-shot samples without any calibration. As a result, this approach serves as a vanilla baseline." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": "- Holistic Calibration: This method corresponds to the TEEN approach [1], which employs holistic calibration for prototypes. It is used as a comparative baseline to dissect and evaluate the effectiveness of the fine-grained calibration proposed in our method." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3159" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 72, + 220, + 177 + ], + "blocks": [ + { + "bbox": [ + 60, + 72, + 220, + 177 + ], + "lines": [ + { + "bbox": [ + 60, + 72, + 220, + 177 + ], + "spans": [ + { + "bbox": [ + 60, + 72, + 220, + 177 + ], + "type": "image", + "image_path": "eb124ccf06927eaa0bdf9ae42edcb4994b0dafc92da504fd06769a9eb51aeac3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 104, + 182, + 175, + 192 + ], + "lines": [ + { + "bbox": [ + 104, + 182, + 175, + 192 + ], + "spans": [ + { + "bbox": [ + 104, + 182, + 175, + 192 + ], + "type": "text", + "content": "(a) NAcc w.r.t. Top-q" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 224, + 72, + 384, + 176 + ], + "blocks": [ + { + "bbox": [ + 224, + 72, + 384, + 176 + ], + "lines": [ + { + "bbox": [ + 224, + 72, + 384, + 176 + ], + "spans": [ + { + "bbox": [ + 224, + 72, + 384, + 176 + ], + "type": "image", + "image_path": "d119e743b9305a427f59f1da63a47cb613759d632644ec853252b86a01cee728.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 269, + 182, + 338, + 192 + ], + "lines": [ + { + "bbox": [ + 269, + 182, + 338, + 192 + ], + "spans": [ + { + "bbox": [ + 269, + 182, + 338, + 192 + ], + "type": "text", + "content": "(b) NAcc w.r.t. Top-r" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 193, + 555, + 214 + ], + "lines": [ + { + "bbox": [ + 55, + 193, + 555, + 214 + ], + "spans": [ + { + "bbox": [ + 55, + 193, + 555, + 214 + ], + "type": "text", + "content": "Figure 3. Impact of hyperparameters on average accuracy over all novel classes at last session (i.e. NAcc), where " + }, + { + "bbox": [ + 55, + 193, + 555, + 214 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 193, + 555, + 214 + ], + "type": "text", + "content": " stands for the size of feature segments, " + }, + { + "bbox": [ + 55, + 193, + 555, + 214 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 193, + 555, + 214 + ], + "type": "text", + "content": " for the number of base-class feature segments for calibrating, and " + }, + { + "bbox": [ + 55, + 193, + 555, + 214 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 193, + 555, + 214 + ], + "type": "text", + "content": " for the calibration strength of each feature segment." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 389, + 72, + 548, + 176 + ], + "blocks": [ + { + "bbox": [ + 389, + 72, + 548, + 176 + ], + "lines": [ + { + "bbox": [ + 389, + 72, + 548, + 176 + ], + "spans": [ + { + "bbox": [ + 389, + 72, + 548, + 176 + ], + "type": "image", + "image_path": "b7e3d5426c9f0ad308bdfdc13f6ada4642169333a992ec3e3636fa183145609a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 441, + 182, + 496, + 191 + ], + "lines": [ + { + "bbox": [ + 441, + 182, + 496, + 191 + ], + "spans": [ + { + "bbox": [ + 441, + 182, + 496, + 191 + ], + "type": "text", + "content": "(c) NAcc w.r.t. " + }, + { + "bbox": [ + 441, + 182, + 496, + 191 + ], + "type": "inline_equation", + "content": "\\alpha" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 226, + 295, + 380 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 55, + 226, + 295, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 226, + 295, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 226, + 295, + 285 + ], + "type": "text", + "content": "- Random Feature Segments: This strategy represents a degraded version of our method, where the feature segments are replaced by random divisions of visual features without any semantic guidance. This approach is used to dissect the necessity of the keyword grounding function." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 286, + 295, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 286, + 295, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 286, + 295, + 380 + ], + "type": "text", + "content": "- Naive Keyword Grounding: This strategy also represents a degraded version of our method, where the mapping of each keyword to the feature segment is simplified by multiplying the [EOS] token embedding with the keyword token embedding, rather than incorporating the attention matrix " + }, + { + "bbox": [ + 55, + 286, + 295, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 55, + 286, + 295, + 380 + ], + "type": "text", + "content": " and value matrix " + }, + { + "bbox": [ + 55, + 286, + 295, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 55, + 286, + 295, + 380 + ], + "type": "text", + "content": ". This approach is used to dissect and demonstrate that whether the design of the keyword grounding function is trivial." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 385, + 295, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 385, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 385, + 295, + 552 + ], + "type": "text", + "content": "The evaluation results are shown in Table 4. As can be observed, the " + }, + { + "bbox": [ + 55, + 385, + 295, + 552 + ], + "type": "inline_equation", + "content": "11.70\\%" + }, + { + "bbox": [ + 55, + 385, + 295, + 552 + ], + "type": "text", + "content": " NAcc gain of \"Holistic Calibration only\" over \"No Calibration\" confirms that prototype calibration is beneficial. Furthermore, our method outperforms \"Holistic Calibration only\", demonstrating that fine-grained calibration of few-shot class prototypes is both effective and necessary. The inferior performance of \"Random Feature Segments\" and \"Naive Keyword Grounding\" compared to our full method indicates that the effectiveness of fine-grained calibration for few-shot class prototypes relies heavily on its elaborate design. This underscores the value of our carefully designed keyword grounding function, highlighting its critical role in achieving superior performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 554, + 295, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 554, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 554, + 295, + 578 + ], + "type": "text", + "content": "Impact of Hyperparameter. We analyze the impact of three key parameters in controlled experiments:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "- Top-" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": " binarization are conducted on contribution indication matrix " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": " along the feature dimensions, where " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": " controls the size of feature segments. From Fig. 3a, we observe that feature segments obtained the best performance with " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "q = 50" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": " in NAcc. Smaller values of " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": " (e.g., 10 and 30) result in smaller feature segments, which may be insufficiently captured by the corresponding keywords, leading to suboptimal performance. On the other hand, larger values of " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": " (e.g., 200, 512) degrade performance, as they tend to approach holistic calibration and fail to yield improvements. This indicates that an intermediate granu" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 321, + 225, + 553, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 225, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 321, + 225, + 553, + 249 + ], + "type": "text", + "content": "larity in feature segmentation, which well corresponds to the keyword, is essential." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 250, + 553, + 464 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 314, + 250, + 553, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 250, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 314, + 250, + 553, + 380 + ], + "type": "text", + "content": "- Top-r similar prototype segments from base classes are combined to calibrate the novel class, where " + }, + { + "bbox": [ + 314, + 250, + 553, + 380 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 314, + 250, + 553, + 380 + ], + "type": "text", + "content": " controls the number of base-class feature segments for calibrating. In Fig. 3b, we see that increasing the number of similar prototypes (i.e. a larger " + }, + { + "bbox": [ + 314, + 250, + 553, + 380 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 314, + 250, + 553, + 380 + ], + "type": "text", + "content": ") enhances performance up to " + }, + { + "bbox": [ + 314, + 250, + 553, + 380 + ], + "type": "inline_equation", + "content": "r = 50" + }, + { + "bbox": [ + 314, + 250, + 553, + 380 + ], + "type": "text", + "content": ", after which further increases do not yield substantial benefits. This suggests that a small set of highly relevant prototypes is sufficient for effectively calibrating novel class prototypes. Including too many segments may introduce noise or less relevant information, diminishing the calibration quality." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 381, + 553, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 381, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 314, + 381, + 553, + 464 + ], + "type": "text", + "content": "- Calibration ratio " + }, + { + "bbox": [ + 314, + 381, + 553, + 464 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 314, + 381, + 553, + 464 + ], + "type": "text", + "content": " controls the calibration strength of each feature segment. As shown in Fig. 3c, the calibration ratio " + }, + { + "bbox": [ + 314, + 381, + 553, + 464 + ], + "type": "inline_equation", + "content": "\\alpha = 0.5" + }, + { + "bbox": [ + 314, + 381, + 553, + 464 + ], + "type": "text", + "content": " yields the best results, balancing the original and calibrated features effectively. Lower values of " + }, + { + "bbox": [ + 314, + 381, + 553, + 464 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 314, + 381, + 553, + 464 + ], + "type": "text", + "content": " reduce the impact of calibration, while higher values overly prioritize the reference prototypes, leading to suboptimal results." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 495, + 479, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 479, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 479, + 508 + ], + "type": "text", + "content": "6. Conclusion and Future Work" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 522, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 555, + 712 + ], + "type": "text", + "content": "In this work, we propose a semantic Feature Decomposition-Recomposition method that unlocks the full potential of CLIP's visual-semantic features for few-shot class-incremental learning. Our core innovation lies in decomposing CLIP's frozen features into semantic-aware segments through guidance of textual keywords, making these segments transfer-ready through incremental learning to be selectively recomposed for novel class calibration. By establishing explicit mappings between LLM-generated attribute keywords and their corresponding feature segments, our method enables precise inheritance of visual-semantic characteristics from multiple base classes. The improvement of our method over existing methods on several settings demonstrates its unique capability to mitigate prototype bias while preserving discriminative features." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3160" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 161, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 161, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 161, + 85 + ], + "type": "text", + "content": "7. Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 128 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 128 + ], + "type": "text", + "content": "This work is partially supported by the National Natural Science Foundation of China (No.62495082 and No.62461160331)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 115, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 115, + 153 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 115, + 153 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 161, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 61, + 161, + 295, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 161, + 295, + 182 + ], + "spans": [ + { + "bbox": [ + 61, + 161, + 295, + 182 + ], + "type": "text", + "content": "[1] Few-shot class-incremental learning via training-free prototype calibration. In NIPS, 2023. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 184, + 296, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 184, + 296, + 206 + ], + "spans": [ + { + "bbox": [ + 62, + 184, + 296, + 206 + ], + "type": "text", + "content": "[2] Closer: Towards better representation learning for few-shot class-incremental learning. In ECCV, 2024. 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 209, + 295, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 209, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 62, + 209, + 295, + 229 + ], + "type": "text", + "content": "[3] Expandable subspace ensemble for pre-trained model-based class-incremental learning. In CVPR, 2024. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 232, + 295, + 254 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 232, + 295, + 254 + ], + "spans": [ + { + "bbox": [ + 62, + 232, + 295, + 254 + ], + "type": "text", + "content": "[4] Learning prompt with distribution-based feature replay for few-shot class-incremental learning. arXiv preprint, 2024. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 256, + 295, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 256, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 62, + 256, + 295, + 277 + ], + "type": "text", + "content": "[5] Pre-trained vision and language transformers are few-shot incremental learners. In CVPR, 2024. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 279, + 295, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 279, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 62, + 279, + 295, + 300 + ], + "type": "text", + "content": "[6] Rethinking few-shot class-incremental learning: Learning from yourself. In ECCV, 2024. 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 303, + 295, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 303, + 295, + 345 + ], + "spans": [ + { + "bbox": [ + 62, + 303, + 295, + 345 + ], + "type": "text", + "content": "[7] Afra Feyza Akyurek, Ekin Akyurek, Derry Tanti Wijaya, and Jacob Andreas. Subspace regularizers for few-shot class incremental learning. arXiv preprint arXiv:2110.07059, 2021. 1, 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 349, + 295, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 349, + 295, + 402 + ], + "spans": [ + { + "bbox": [ + 62, + 349, + 295, + 402 + ], + "type": "text", + "content": "[8] Arslan Chaudhry, Marcus Rohrbach, Mohamed Elhoseiny, Thalaiyasingam Ajanthan, Puneet K Dokania, Philip HS Torr, and Marc'Aurelio Ranzato. On tiny episodic memories in continual learning. arXiv preprint arXiv:1902.10486, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 405, + 295, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 405, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 62, + 405, + 295, + 437 + ], + "type": "text", + "content": "[9] Kuilin Chen and Chi-Guhn Lee. Incremental few-shot learning via vector quantization in deep embedded space. In International conference on learning representations, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 440, + 295, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 440, + 295, + 493 + ], + "spans": [ + { + "bbox": [ + 57, + 440, + 295, + 493 + ], + "type": "text", + "content": "[10] Ali Cheraghian, Shafin Rahman, Pengfei Fang, Soumava Kumar Roy, Lars Petersson, and Mehrtash Harandi. Semantic-aware knowledge distillation for few-shot class-incremental learning. In CVPR, pages 2534-2543, 2021. 1, 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 496, + 295, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 496, + 295, + 528 + ], + "spans": [ + { + "bbox": [ + 57, + 496, + 295, + 528 + ], + "type": "text", + "content": "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 531, + 295, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 531, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 57, + 531, + 295, + 563 + ], + "type": "text", + "content": "[12] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In CVPR, pages 831-839, 2019. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 566, + 295, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 566, + 295, + 598 + ], + "spans": [ + { + "bbox": [ + 57, + 566, + 295, + 598 + ], + "type": "text", + "content": "[13] David Isele and Akansel Cosgun. Selective experience replay for lifelong learning. In Proceedings of the AAAI conference on artificial intelligence, 2018. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 601, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 295, + 633 + ], + "type": "text", + "content": "[14] Gregory Koch, Richard Zemel, Ruslan Salakhutdinov, et al. Siamese neural networks for one-shot image recognition. In ICML deep learning workshop, pages 1-30. Lille, 2015. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 57, + 635, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 667 + ], + "type": "text", + "content": "[15] Anna Kukleva, Hilde Kuehne, and Bernt Schiele. Generalized and incremental few-shot learning by explicit learning and calibration without forgetting. In ICCV, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "text", + "content": "[16] Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In Psychology of learning and motivation. Elsevier, 1989. 1, 2" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 137 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 137 + ], + "type": "text", + "content": "[17] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning. PmLR, 2021. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 140, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 172 + ], + "type": "text", + "content": "[18] David Rolnick, Arun Ahuja, Jonathan Schwarz, Timothy Lillicrap, and Gregory Wayne. Experience replay for continual learning. NIPS, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 175, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 553, + 217 + ], + "type": "text", + "content": "[19] Guangyuan Shi, Jiaxin Chen, Wenlong Zhang, Li-Ming Zhan, and Xiao-Ming Wu. Overcoming catastrophic forgetting in incremental few-shot learning by finding flat minima. NIPS, 34:6747-6761, 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 219, + 553, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 250 + ], + "type": "text", + "content": "[20] Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay. NIPS, 30, 2017. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 253, + 553, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 284 + ], + "type": "text", + "content": "[21] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. NeurIPS, 30, 2017. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 287, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 319 + ], + "type": "text", + "content": "[22] Xiaoyu Tao, Xiaopeng Hong, Xinyuan Chang, Songlin Dong, Xing Wei, and Yihong Gong. Few-shot class-incremental learning. In CVPR, 2020. 1, 2, 3, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 321, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 553, + 353 + ], + "type": "text", + "content": "[23] Songsong Tian, Lusi Li, Weijun Li, Hang Ran, Xin Ning, and Prayag Tiwari. A survey on few-shot class-incremental learning. Neural Networks, 169, 2024. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 355, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 553, + 386 + ], + "type": "text", + "content": "[24] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al. Matching networks for one shot learning. NIPS, 29, 2016. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 388, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 421 + ], + "type": "text", + "content": "[25] Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: Theory, method and application. IEEE TPAMI, 2024. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 422, + 553, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 455 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 455 + ], + "type": "text", + "content": "[26] Yaqing Wang, Quanming Yao, James T Kwok, and Lionel M Ni. Generalizing from a few examples: A survey on few-shot learning. ACM computing surveys (csur), 53(3), 2020. 1, 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 456, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 553, + 498 + ], + "type": "text", + "content": "[27] Chenshen Wu, Luis Herranz, Xialei Liu, Joost Van De Weijer, Bogdan Raducanu, et al. Memory replay gans: Learning to generate new categories without forgetting. NIPS, 31, 2018. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 501, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 501, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 501, + 553, + 555 + ], + "type": "text", + "content": "[28] Boyu Yang, Mingbao Lin, Binghao Liu, Mengying Fu, Chang Liu, Rongrong Ji, and Qixiang Ye. Learnable expansion-and-compression network for few-shot class incremental learning. arXiv preprint arXiv:2104.02281, 2021.1" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 557, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 557, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 317, + 557, + 553, + 600 + ], + "type": "text", + "content": "[29] Boyu Yang, Mingbao Lin, Yunxiao Zhang, Binghao Liu, Xiaodan Liang, Rongrong Ji, and Qixiang Ye. Dynamic support network for few-shot class incremental learning. IEEE TPAMI, 45(3):2945-2951, 2022. 1" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 601, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 644 + ], + "type": "text", + "content": "[30] Chi Zhang, Nan Song, Guosheng Lin, Yun Zheng, Pan Pan, and Yinghui Xu. Few-shot incremental learning with continually evolved classifiers. In CVPR, pages 12455-12464, 2021. 1, 6, 7" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[31] Guangtao Zheng and Aidong Zhang. Few-shot class-incremental learning with meta-learned class structures. In 2021 International Conference on Data Mining Workshops (ICDMW), pages 421–430. IEEE, 2021. 1" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[32] Da-Wei Zhou, Fu-Yun Wang, Han-Jia Ye, Liang Ma, Shiliang Pu, and De-Chuan Zhan. Forward compatible few" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3161" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 129 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "shot class-incremental learning. In CVPR, pages 9046-9056, 2022. 1, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 294, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 294, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 294, + 129 + ], + "type": "text", + "content": "[33] Kai Zhu, Yang Cao, Wei Zhai, Jie Cheng, and Zheng-Jun Zha. Self-promoted prototype refinement for few-shot class incremental learning. In CVPR, pages 6801-6810, 2021. 1" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "text", + "content": "3162" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Feature Extraction and Representation of Pre-training Point Cloud Based on Diffusion Models/0c060285-4bbf-430f-b421-79f016d6dd03_content_list.json b/2025/Feature Extraction and Representation of Pre-training Point Cloud Based on Diffusion Models/0c060285-4bbf-430f-b421-79f016d6dd03_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6276f06150b538cac82c7fafd96e6a95485c7a65 --- /dev/null +++ b/2025/Feature Extraction and Representation of Pre-training Point Cloud Based on Diffusion Models/0c060285-4bbf-430f-b421-79f016d6dd03_content_list.json @@ -0,0 +1,1832 @@ +[ + { + "type": "text", + "text": "Feature Extraction and Representation of Pre-training Point Cloud Based on Diffusion Models", + "text_level": 1, + "bbox": [ + 112, + 128, + 883, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chang Qiu", + "bbox": [ + 235, + 204, + 326, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Southeast University", + "bbox": [ + 197, + 222, + 364, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "230238519@seu.edu.cn", + "bbox": [ + 191, + 242, + 370, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Feipeng Da*", + "bbox": [ + 446, + 204, + 550, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Southeast University", + "bbox": [ + 411, + 222, + 578, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "dafp@seu.edu.cn", + "bbox": [ + 428, + 242, + 563, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zilei Zhang", + "bbox": [ + 663, + 204, + 759, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Southeast University", + "bbox": [ + 627, + 222, + 795, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "230238524@seu.edu.cn", + "bbox": [ + 620, + 242, + 799, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The pretrain-finetune paradigm of pre-training a model on large amounts of image and text data and then fine-tuning the model for a specific task has led to significant progress in many 2D image and natural language processing tasks. Similarly, the use of pre-training methods in point cloud data can also enhance the working performance and generalization ability of the model. Therefore, in this paper, we propose a pre-training framework based on a diffusion model called PreDifPoint. It is able to accomplish the pretraining of the model's backbone network through a diffusion process of gradual denoising. We aggregate the potential features extracted from the backbone network, input them as conditions into the subsequent diffusion model, and direct the point-to-point mapping relationship of the noisy point clouds at neighboring time steps, so as to generate high-quality point clouds and at the same time better perform various downstream tasks of the point clouds. We also introduce a bi-directional covariate attention (DXCA-Attention) mechanism for capturing complex feature interactions, fusing local and global features, and improving the detail recovery of point clouds. In addition, we propose a density-adaptive sampling strategy, which can help the model dynamically adjust the sampling strategy between different time steps, and guide the model to pay more attention to the denser regions in the point cloud, thus improving the effectiveness of the model in point cloud recovery. Our PreDifPoint framework achieves more competitive results on various real-world datasets. Specifically, PreDifPoint achieves an overall accuracy of $87.96\\%$ , which is $0.35\\%$ higher than PointDif, on the classification task on PB-T50-395RS, a variant of ScanObjectNN dataset.", + "bbox": [ + 89, + 321, + 483, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 820, + 220, + 835 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent work has demonstrated that the pretrain-finetune paradigm, where a model is first pretrained on large", + "bbox": [ + 89, + 845, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "amounts of image and text data and then fine-tuned for specific tasks, has led to substantial advancements in many 2D image and natural language processing tasks [11, 16, 27]. Pretraining the backbone on large-scale datasets enables the model to learn rich feature representations while also avoiding the need for large amounts of data and computational resources required for training a model from scratch [8, 9, 32]. Pretrained models, after fine-tuning, can learn task-specific knowledge, exhibit better generalization capabilities, and converge to optimal performance levels more rapidly [6, 14]. Similarly, using pretraining methods for point cloud data can also enhance the model's performance and generalization ability [44, 46, 48].", + "bbox": [ + 511, + 292, + 906, + 489 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generative pretraining methods, through generative models, capture the complex structure and intricate details of point cloud data, thereby obtaining more powerful feature representations. Early works such as PointBERT [44], Point-MAE [23], and Point-M2AE [45] employed self-supervised learning by reconstructing masked local point patches, forcing the model to learn local point cloud information. However, this masking and reconstruction approach may overlook the contextual relationships between masked and unmasked regions, preventing the model from effectively understanding the global structure of the object or scene. Inspired by diffusion models [10], which can understand global structure and long-term dependencies through the process of adding and removing noise, we propose using a diffusion model to guide the pretraining process of point clouds.", + "bbox": [ + 511, + 489, + 908, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Thus, we introduce a diffusion-based pretraining framework, named PreDifPoint. This framework pretrains only the backbone network of the model, and then fine-tunes the task-specific parts. The pretraining of the backbone network is achieved through a progressive denoising diffusion process, as shown in Fig. 1. By pretraining the backbone, it can learn meaningful features during the feature extraction phase, which also significantly enhances the model's generalization ability. At the same time, this reduces the number of parameters that need to be adjusted during training, accelerating convergence and reducing overfitting.", + "bbox": [ + 511, + 734, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 886, + 236, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26559", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c66474988fa9fbe369b0b9455a8976898db1d7f067ccb1487469a4f66e888563.jpg", + "image_caption": [ + "Figure 1. Schematic illustration of our PreDiffPoint. Our PreDiffPoint can be used to train different backbones and reconstruct the original point cloud from the noisy one through a diffusion process." + ], + "image_footnote": [], + "bbox": [ + 94, + 90, + 480, + 204 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, the PreDifPoint framework consists of three core modules: the Conditional Point Generation Network (CGNet), the Conditional Denoising Network (CDNet), and the Conditional Diffusion Network (CDPNet). CGNet aggregates the latent features extracted by the backbone and feeds them as conditional inputs into the subsequent diffusion model. CDNet utilizes these conditional inputs to perform denoising and restore point cloud details. CDPNet, through the point-to-point mapping of noisy point clouds between adjacent time steps in the diffusion process [48], ensures that the generated point clouds excel in both global consistency and local detail. With this design, the model can generate high-quality point clouds based on various conditions, exhibiting strong diversity and adaptability to perform various downstream tasks.", + "bbox": [ + 88, + 303, + 482, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Additionally, we propose a density-adaptive sampling strategy. We divide the entire time step range into several subintervals, each containing multiple time steps. Within each subinterval, the sampling probability is dynamically adjusted according to the local density of the point cloud at each time step, prioritizing sampling from steps with higher density. This strategy allows for more fine-grained control over the sampling process at each time step and helps dynamically adjust the sampling strategy across time steps, improving the model's point cloud recovery performance.", + "bbox": [ + 89, + 531, + 482, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions can be summarized as follows:", + "bbox": [ + 109, + 683, + 472, + 698 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a diffusion-based point cloud pretraining framework, PreDifPoint, which leverages conditional guidance for generation, denoising, and recovery, as well as global and local feature modeling. This framework enhances the performance of point cloud generation and denoising tasks and provides high-quality feature representations for downstream tasks.", + "- We introduce a DXCA-Attention mechanism into the model to capture complex global feature interactions and generate more instructive condition vectors to enhance the detail reconstruction of the point cloud. The condition vectors obtained through CGNet will be combined with time-step information and passed to the subsequent" + ], + "bbox": [ + 89, + 704, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "network for denoising.", + "bbox": [ + 527, + 92, + 679, + 106 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a density-adaptive sampling strategy, which can help the model dynamically adjust the sampling strategy between different time steps, and guide the model to pay more attention to the denser regions in the point cloud, thus improving the effectiveness of the model in point cloud recovery.", + "- The PreDifPoint framework requires only fine-tuning for task-specific components (such as classifiers), allowing the model to perform optimally across various tasks, particularly in data-limited or computationally constrained environments. Notably, the PreDifPoint framework demonstrates competitive performance in a variety of practical downstream tasks." + ], + "bbox": [ + 511, + 106, + 903, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 319, + 663, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Point Cloud Pretraining", + "text_level": 1, + "bbox": [ + 511, + 345, + 735, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In deep learning, pretraining has become a key technique for enhancing model performance, especially when data is scarce or labels are limited. For point cloud learning tasks, pretraining is particularly important. Point cloud data typically has complex spatial structures and sparsity, and traditional training methods often rely on large amounts of labeled data, which is difficult to obtain in real-world applications. Therefore, learning effective point cloud representations through pretraining not only reduces reliance on labeled data but also improves the model's generalization and robustness in downstream tasks.", + "bbox": [ + 511, + 368, + 906, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Currently, point cloud pretraining methods can be broadly categorized into three types: self-supervised learning, generative models [29, 34, 43], and transfer learning. Each approach has its unique advantages and application scenarios.", + "bbox": [ + 511, + 536, + 905, + 611 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-supervised methods [4, 30, 31] can mask parts of the data and train the model to reconstruct the missing parts or optimize the model's representation ability through strategies like contrastive learning. Point-MAE [23] randomly masks certain points in the point cloud and trains the model to recover these missing points, thereby effectively learning the geometric features of the point cloud. Point-M2AE [45] for hierarchical self-supervised learning of 3D point clouds modifies the encoder and decoder into a pyramid architecture to incrementally model the spatial geometry and capture the fine-grained and high-level semantics of 3D shapes. PointContrast [42] employs contrastive learning to treat point cloud representation as a learning task, improving the model's ability to distinguish between similar and different point cloud segments.", + "bbox": [ + 511, + 612, + 905, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generative models [19, 20, 24, 37, 41], particularly denoising diffusion models and variational autoencoders, have made significant progress in point cloud pretraining in recent years. These models learn the generative distribution of", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26560", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "point cloud data, capturing the details of the point cloud and completing missing data. DDPM [22] demonstrates how to use pretrained denoising diffusion models for conditional point cloud generation and completion, significantly improving the quality of point cloud reconstruction. PVD [49] introduces a conditional variational diffusion model (CV-Diffusion) that can generate high-quality 3D shapes through conditional information, effectively applied to point cloud generation and completion tasks.", + "bbox": [ + 89, + 90, + 480, + 226 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Transfer learning methods [3, 12, 15] involve pretraining a model on a large-scale dataset and then transferring it to a specific task, significantly improving performance when data is limited. Meta-PointNet [17] proposes a meta-learning-based pretraining method for point cloud classification and segmentation, where knowledge is shared across multiple tasks, greatly improving performance on few-shot tasks.", + "bbox": [ + 89, + 227, + 480, + 348 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Diffusion Models", + "text_level": 1, + "bbox": [ + 89, + 361, + 259, + 376 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The core idea of diffusion models [13, 28, 36, 39, 51] is to simulate a process of gradually adding noise and then recovering the data through reverse denoising. Diffusion models have been widely used in recognition tasks such as object detection [5] and semantic discrimination [1, 2, 38]. During training, the model learns how to recover real data from noise, which helps generate high-quality samples. This paper uses conditional diffusion models, which introduce additional conditional information into the diffusion process to guide the generated model. These conditional inputs can include labels, textual descriptions, or other external features that help the model generate data that meets specific conditions. Conditional diffusion models have shown excellent performance in multimodal generation tasks, particularly in image generation, point cloud generation, and speech synthesis.", + "bbox": [ + 89, + 383, + 482, + 626 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion probabilistic models [21] proposed a point cloud probabilistic generation model inspired by nonequilibrium thermodynamics that utilizes a backward diffusion process to learn point distributions. And they achieved competitive performance in point cloud generation and automatic coding, and comparable results in unsupervised representation learning. PointDif [48] proposes a conditional diffusion model for point cloud pretraining tasks. This model can more precisely control the structure and details of the point cloud during the generation or denoising process, performing particularly well in point cloud denoising and generation tasks. ICDDPM [47] can uses 2D images as prior information to guide end-to-end 3D point cloud generation, and obtain high-quality 3D point clouds of different types of targets from real-world images.", + "bbox": [ + 89, + 628, + 482, + 853 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The PreDifPoint framework we have developed pretrains the network by recovering the original point cloud from randomly noisy point clouds, forcing the network to learn local", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and global geometric priors of point clouds. The extracted point cloud features are passed as conditional information into the diffusion process, guiding the generation of point clouds step by step. Additionally, we propose a density-adaptive sampling strategy, which allows for finer control over the sampling process at each time step and helps dynamically adjust the sampling strategy between time steps, thus improving the model's ability to recover point clouds.", + "bbox": [ + 511, + 90, + 903, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 224, + 604, + 239 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our pre-training process, the backbone of the overall framework PreDifPoint is a transformer encoder. Notably, this backbone can be replaced with any other backbone network for pre-training. The input point cloud is grouped and features are extracted, followed by masking operations and contextual enhancement via a Transformer with positional encoding. Subsequently, CGNet further processes these features to generate conditional vectors that serve as conditional inputs to CDNet. In CDNet, time-step embedding is combined with conditional vectors. Finally, the CDPNet network integrates CGNet and CDNet to generate the original point cloud by stepwise denoising the data through a diffusion model. The pipeline of our PreDifPoint is shown in Fig. 2.", + "bbox": [ + 511, + 250, + 906, + 462 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Diffusion Process", + "text_level": 1, + "bbox": [ + 511, + 470, + 683, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the diffusion model, a clean point cloud $X^0 \\epsilon R^{n \\times 3}$ is gradually corrupted into a fully noisy point cloud $X^T$ by incrementally adding noise. The diffusion process at each time step $t$ follows:", + "bbox": [ + 511, + 491, + 905, + 551 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(X ^ {t} \\mid X ^ {t - 1}\\right) = \\mathcal {N} \\left(X ^ {t}; \\sqrt {\\alpha_ {t}} X ^ {t - 1}, (1 - \\alpha_ {t}) \\mathrm {I}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 577, + 906, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha_{t} = 1 - \\beta_{t}$ denotes the noise attenuation coefficient at time step, and $\\beta_{t}$ represents the noise level. Through this process, noise is gradually added to the point cloud until it becomes nearly random at $X^T$ .", + "bbox": [ + 511, + 604, + 905, + 665 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This reverse process predicts the denoising signals at each time step by the neural network $p_{\\theta}(X^t | X^{t-1}, c)$ , which gradually recovers the original structure of the point cloud as shown in the following equation:", + "bbox": [ + 511, + 666, + 905, + 726 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(X ^ {t} \\mid X ^ {t - 1}, c\\right) = \\mathcal {N} \\left(X ^ {t - 1}; \\mu_ {\\theta} \\left(X ^ {t}, t, c\\right), \\sigma_ {t} ^ {2} \\mathrm {I}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 750, + 906, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mu_{\\theta}(X^{t},t,c)$ is the mean predicted by the network, $X^t$ representing the denoised result at step $t$ , and $\\sigma_t^2$ is the noise variance.", + "bbox": [ + 511, + 777, + 905, + 823 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Conditional Generation Network (CGNet)", + "text_level": 1, + "bbox": [ + 511, + 833, + 872, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CGNet is responsible for extracting global and local features from the point cloud, fusing these features through a DXCA-Attention mechanism to more comprehensively", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26561", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/51f7b387b6e06eb17d4312d18e311f641af80d8d8a5343d7d3a8b8c97bfd5f32.jpg", + "image_caption": [ + "Figure 2. Pipeline of PreDifPoint. We first group the input point cloud and extract features, which are then subjected to masking operations and contextual enhancement via a transformer with position encoding. Subsequently, CGNet further processes these features to generate conditional vectors, which are used as conditional inputs to CDNet. In CDNet, time-step embedding is combined with conditional vectors. Finally, the CDPNet network integrates CGNet and CDNet to generate the original point cloud by stepwise denoising the data through a diffusion model." + ], + "image_footnote": [], + "bbox": [ + 174, + 90, + 821, + 310 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "model the interactions between the features and better adapt to different data patterns and structures, as shown in Fig. 3. Finally, a conditional vector $c$ associated with the density distribution is generated, which is used to guide the subsequent denoising process.", + "bbox": [ + 88, + 420, + 480, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first project the input feature $F_{loc}$ from the point cloud chunking encoding to the high dimensional space through the multilayer perceptron (MLP) to obtain the projected representation $F_{pro}$ of the local features, and then obtain the global feature $F_{glo}$ through the global pooling operation on $F_{pro}$ as shown in the following equation:", + "bbox": [ + 88, + 496, + 482, + 587 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} F _ {p r o} = M L P (F), \\\\ F _ {g l o} = G l o b a l M a x P o o l \\left(F _ {p r o}\\right). \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 598, + 482, + 640 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, the attention weights are obtained by calculating the covariance matrix in both directions to capture the global contextual relationships and complex feature interactions. In this process, we need to first obtain the important parameters, $Q$ , $K$ , and $V$ , as shown in Eq. (4).", + "bbox": [ + 89, + 646, + 483, + 722 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} Q = \\operatorname {r e s h a p e} \\left(F _ {p r o}\\right) \\in \\mathbb {R} ^ {B \\times G \\times C}, \\\\ K = \\operatorname {r e s h a p e} \\left(F _ {g l o}\\right) \\in \\mathbb {R} ^ {B \\times G \\times C}, \\\\ V _ {1} = \\operatorname {r e s h a p e} \\left(F _ {g l o}\\right) \\in \\mathbb {R} ^ {B \\times G \\times C}, \\\\ V _ {2} = \\operatorname {r e s h a p e} \\left(F _ {p r o}\\right) \\in \\mathbb {R} ^ {B \\times G \\times C}. \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 732, + 482, + 815 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then centering along the feature dimension $C$ as shown in Eq. (5).", + "bbox": [ + 89, + 820, + 482, + 851 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} \\bar {Q} = Q - \\mathbb {E} _ {c} [ Q ], \\\\ \\bar {K} = K - \\mathbb {E} _ {c} [ K ]. \\end{array} \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 863, + 482, + 904 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then the scaling factor $\\sqrt{C}$ is introduced to compute the covariance matrix, as shown in Eq. (6).", + "bbox": [ + 511, + 419, + 906, + 450 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} A t t e n S c o r e _ {1} = S o f t m a x \\left(\\frac {\\bar {Q} \\cdot \\bar {K} ^ {T}}{\\sqrt {C}}\\right) \\in \\mathbb {R} ^ {B \\times G \\times C}, \\\\ A t t e n S c o r e _ {2} = S o f t m a x \\left(\\frac {\\bar {K} \\cdot \\bar {Q} ^ {T}}{\\sqrt {C}}\\right) \\in \\mathbb {R} ^ {B \\times G \\times C}. \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 470, + 903, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "At this point, the bi-directional covariance features can be aggregated as shown in Eq. (7).", + "bbox": [ + 511, + 575, + 906, + 606 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} O _ {1} = A t t e n S c o r e _ {1} \\cdot V _ {1} \\in \\mathbb {R} ^ {B \\times G \\times C}, \\\\ O _ {2} = A t t e n S c o r e _ {2} \\cdot V _ {2} \\in \\mathbb {R} ^ {B \\times G \\times C}, \\\\ O = \\operatorname {c o n c a t} \\left(O _ {1}, O _ {2}\\right) \\in \\mathbb {R} ^ {B \\times G \\times 2 C}. \\end{array} \\right. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 614, + 903, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we fuse the projected features, global features and the attention-weighted features to output the conditioned vector $c$ , as shown in Eq. (8).", + "bbox": [ + 511, + 680, + 906, + 727 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nc = M L P \\left(\\text {C o n c a t} \\left(F _ {\\text {p r o}}, A, F _ {\\text {g l o}}\\right)\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 739, + 903, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Conditional Denoising Network (CDNet)", + "text_level": 1, + "bbox": [ + 511, + 772, + 861, + 789 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CDNet is responsible for gradually removing the noise and restoring the point cloud structure during the diffusion process, and its core is to achieve layer-level feature optimization through residual connectivity.", + "bbox": [ + 511, + 794, + 906, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, the conditional vector $c$ and the time embedding $t_{emd}$ are fused into the point cloud features through a dynamic gating mechanism to achieve density-aware local de", + "bbox": [ + 511, + 854, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26562", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fcd4c81dc0cd386c1eb1b288656edc7c907211fb8eff1db34e11cc40aef08ceb.jpg", + "image_caption": [ + "Figure 3. DXCA-Attention Module. This attention mechanism achieves cross-level information fusion through the two-way interaction between global and local features. In the local-to-global direction, query vectors and key vectors are generated by centering local features and global features respectively, and the contribution weights of local regions to global semantics are computed through the covariance matrix, and the semantic consistency of local features is enhanced after weighted aggregation. In the global-to-local direction, the global features are used as queries and local features as keys, and the modulation relation of global context to local details is captured by covariance attention." + ], + "image_footnote": [], + "bbox": [ + 94, + 90, + 480, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "noising. The time embedding $t_{emd}$ and the fused features can be represented as:", + "bbox": [ + 89, + 407, + 483, + 438 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} t _ {e m d} = \\left[ \\beta_ {t}, \\sin \\left(\\beta_ {t}\\right), \\cos \\left(\\beta_ {t}\\right) \\right], \\\\ H _ {l} = \\sigma \\left(W _ {g} [ c \\oplus t _ {e m d} ]\\right) \\odot \\left(W _ {f} H _ {l - 1}\\right) + W _ {b} [ c \\oplus t _ {e m d} ], \\end{array} \\right. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 463, + 480, + 515 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\sigma$ is the Sigmoid function for generating gating weights in the range [0, 1], and $W_{g}, W_{f}, W_{b}$ are the gating generation weights, feature transformation weights and bias generation weights, respectively.", + "bbox": [ + 89, + 516, + 483, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, to alleviate the gradient vanishing problem, the output of each of our layers can be connected using residuals such that the output of each layer is represented as:", + "bbox": [ + 89, + 577, + 483, + 622 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nH _ {l} = H _ {l - 1} + C D N e t \\left(H _ {l - 1}, c, t _ {e m d}\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 638, + 482, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Conditional Diffusion Point Network (CDPNet)", + "text_level": 1, + "bbox": [ + 89, + 662, + 482, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The CDPNet integrates a density-weighted diffusion process, a conditional point generation network (CGNet), and a conditional denoising network (CDNet) to achieve stepwise generation from noisy to clean point clouds.", + "bbox": [ + 89, + 686, + 483, + 746 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In a nutshell, it is to first get a completely noisy point cloud $X^T$ from a clean point cloud $X^0$ in the forward diffusion process, then obtain the conditional vector $\\mathbf{c}$ by CGNet, and then update $X^T$ step by step from $t = T$ to $t = 1$ in CDNet, and finally get the reconstructed point cloud $X^0$ .", + "bbox": [ + 89, + 747, + 483, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Density-Adaptive Sampling", + "text_level": 1, + "bbox": [ + 89, + 832, + 339, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To dynamically adjust the time step scheduling in the diffusion model, we propose a density-adaptive sampling strategy. This strategy weights the sampling probability of each", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "time step based on the local density of points, prioritizing denser regions (e.g., edges) during training to enhance detail reconstruction.", + "bbox": [ + 511, + 90, + 903, + 133 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "First, the local density $\\rho_{i}$ of each point $x_{i}\\epsilon X_{0}$ is computed using $k$ -nearest neighbors:", + "bbox": [ + 511, + 136, + 903, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\rho \\left(x _ {i}\\right) = \\frac {1}{\\sum_ {j = 1} ^ {k} \\left\\| x _ {i} - x _ {j} \\right\\| + \\epsilon}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 179, + 903, + 215 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $k$ is the neighborhood size, $\\| x_{i} - x_{j}\\|$ is the Euclidean distance, and $\\epsilon$ is a small constant to avoid division by zero.", + "bbox": [ + 511, + 223, + 903, + 268 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The time step interval $[1, T]$ is uniformly divided into $h$ sub-intervals $\\{Q_1, Q_2, \\dots, Q_h\\}$ , each corresponding to different noise levels. The sampling probability for $Q_i$ is weighted by the local density within the interval:", + "bbox": [ + 511, + 268, + 905, + 330 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP \\left(Q _ {i} \\mid \\rho\\right) = \\frac {\\rho_ {i}}{\\sum_ {i = 1} ^ {T} \\rho_ {i}}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 342, + 903, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This weighting strategy ensures that more samples can be performed in the denser regions, thus improving the reconstruction accuracy in these regions.", + "bbox": [ + 511, + 381, + 903, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Within each interval $Q_{i}$ , we randomly sample a time step $t_{i}$ from that interval and compute the diffusion process at that time step. Since the sampling is based on local density weighting, the sampling probability for each interval reflects the denseness of the region. In updating the sampling, we use the following weighted loss function to calculate the loss at each time step:", + "bbox": [ + 511, + 426, + 905, + 532 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL (\\theta , \\rho) = \\frac {1}{h} \\sum_ {i = 0} ^ {h - 1} L \\left(\\theta , \\rho , t _ {i} \\sim Q _ {i}\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 545, + 903, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $L(\\theta, \\rho, t_i \\sim Q_i)$ is the loss for the sampled $t_i$ .", + "bbox": [ + 532, + 595, + 872, + 611 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.6. Density-Weighted Diffusion Loss", + "text_level": 1, + "bbox": [ + 511, + 621, + 802, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the training of diffusion models combined with density-adaptive sampling, the loss function consists mainly of the error in the prediction noise and the error in the density weighting. Our goal is to minimize the difference between the predicted noise and the real noise at each time step.", + "bbox": [ + 511, + 643, + 905, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The standard diffusion model loss is achieved by calculating the mean square error (MSE) between the predicted noise and the real noise, as shown in Eq. (14):", + "bbox": [ + 511, + 720, + 905, + 765 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {d e n o i s e} (\\theta) = \\mathbb {E} _ {x _ {0}, t, \\epsilon} [ \\| \\epsilon - \\epsilon_ {\\theta} (\\sqrt {\\alpha_ {t}} x _ {0} + \\sqrt {1 - \\alpha_ {t}} \\epsilon , t, c) \\| ^ {2} ], \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 789, + 910, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $x_0$ is the input original point cloud, $t$ is the current time step, $\\epsilon_{\\theta}$ is the noise predicted by the network, $\\epsilon$ is the noise actually added, $\\alpha_{t}$ is the noise attenuation coefficient at each time step, $c$ is the conditional information generated by the conditional aggregation network CGNet, and $\\mathbb{E}_{x_0,t,\\epsilon}$", + "bbox": [ + 511, + 825, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26563", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "is the indication of the need for averaging all the possible values of these variables to calculate the final loss.", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During the density-adaptive sampling process, we need to adjust the computation of the loss to be consistent with the sampling probability of each interval. The final training loss combining density-weighted sampling and the standard MSE loss can be expressed as:", + "bbox": [ + 89, + 122, + 483, + 198 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L (\\theta , \\rho) = \\frac {1}{h} \\sum_ {i = 0} ^ {h - 1} \\mathbb {E} _ {t _ {i} \\sim Q _ {i}} [ \\\\ \\begin{array}{l} \\mathbb {E} _ {x _ {0}, \\epsilon} \\left[ \\left\\| \\epsilon - \\epsilon_ {\\theta} \\left(\\sqrt {\\alpha_ {t _ {i}}} x _ {0} + \\sqrt {1 - \\alpha_ {t _ {i}}} \\epsilon , t _ {i}, c\\right) \\right\\| ^ {2} \\right] \\\\ ], \\end{array} \\tag {15} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 223, + 478, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $t_i$ is the time step sampled from the interval $Q_i$ , and the loss function enhances the reconstruction of these regions by weighting the samples at each time step to ensure that denser regions can have higher weights.", + "bbox": [ + 89, + 334, + 483, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 409, + 222, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Pre-training Setup", + "text_level": 1, + "bbox": [ + 89, + 434, + 269, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use ShapeNet to pre-train our model, with the core version of this dataset containing 51,300 3D models across 55 common object categories. For each 3D shape, we sample 1,024 points to serve as the input for the model. Following the approach in PointDif [48], we use 41,952 shapes as the training set. Additionally, we employ the KNN algorithm to select the $k = 32$ nearest points as point patches, resulting in 64 point cloud patches. Furthermore, we set the embedding dimension of the Transformer encoder to 384, with 6 attention heads, and the condition dimension to 768.", + "bbox": [ + 89, + 457, + 483, + 607 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During pre-training, we set the weight decay to 0.05 and the learning rate to 0.001. Our model is pre-trained for 300 epochs with a batch size of 192.", + "bbox": [ + 89, + 609, + 483, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Additionally, we visualize the point clouds generated by the PreDifPoint framework to demonstrate the effectiveness of our pre-training. As shown in Fig. 4, we use the masked point cloud features, extracted and used as conditions, to guide the diffusion model in generating the original point cloud, with a masking rate of 0.8. The visualization results indicate that the PreDifPoint framework generates high-quality point clouds and that the learned geometric priors effectively guide subsequent tasks such as recognition and detection.", + "bbox": [ + 89, + 655, + 483, + 805 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Downstream Tasks", + "text_level": 1, + "bbox": [ + 89, + 816, + 272, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The performance of a pre-trained model, obtained in a fine-tuning task, is a test of the effectiveness of the pre-trained model. Since a high-quality point cloud pre-trained model should have a high hierarchical geometric prior, we tested", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/43369277d394f2ca882e82548da3f31cfab1edb5028405624bfc12f6ca4b5184.jpg", + "image_caption": [ + "Figure 4. Visualization results on the ShapeNet validation set. We show a total of six categories of point clouds, each showing the input point cloud, the masked point cloud, and the reconstructed point cloud from left to right." + ], + "image_footnote": [], + "bbox": [ + 517, + 95, + 903, + 247 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the performance of the pre-trained model on a large number of real-world datasets.", + "bbox": [ + 511, + 349, + 905, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Object Classification.", + "text_level": 1, + "bbox": [ + 513, + 378, + 666, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first evaluated the classification ability of the PreD-ifPoint framework in the ScanObjectNNs dataset.", + "bbox": [ + 511, + 396, + 905, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The ScanObjectNN dataset has 15 classes, which is actually more challenging due to the complexity of its contexts and object locations. The ScanObjectNN dataset is divided into three subsets: OBJ-ONLY, (only objects), OBJ-BG(objects and background), and PB-T50-RS(objects, background, and artificially added perturbations). We take the Overall Accuracy on these three subsets as the evaluation metric, and the detailed experimental results are summarized in Tab. 2.", + "bbox": [ + 511, + 428, + 905, + 563 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As can be seen from the table, our PreDifPoint framework achieves excellent classification results in the ScanObjectNN dataset, regardless of the presence or absence of the complex background of artificially added perturbations. It demonstrates the strong adaptability and robustness of our model in the face of complex and changing environments. Compared with other state-of-the-art models, the PreDif-Point framework also achieves the best performance.", + "bbox": [ + 511, + 565, + 905, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Object Segmentation.", + "text_level": 1, + "bbox": [ + 513, + 686, + 668, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also validated the ability of our model for object segmentation in the ShapeNetPart dataset. ShapeNetPart contains a total of 16 large categories, which are labeled with semantic information for the semantic segmentation task of the point cloud.", + "bbox": [ + 511, + 702, + 905, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use mIoU to evaluate the object segmentation performance of the model, and from Tab. 1, we can see that overall, PreDifPoint has the best segmentation performance, which is $0.8\\%$ better than MPCT. Among other things, our model also achieves the best performance in multiple categories. This indicates that our model has been pre-trained to learn the generalized feature representation, which can extract the key features in the input data more efficiently and", + "bbox": [ + 511, + 779, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26564", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/227b5ea366b0980a7cbd9b448affaa79c312f7197b3610f86924fdb917e7166f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodmIoUairplanebagcapcarchairearphoneguitarknifelamplaptopmotorbikemugpistolrocketskateboardtable
PointNet [25]83.783.478.782.574.989.673.091.585.980.895.365.293.081.257.972.880.6
PointNet++ [26]85.182.479.087.777.390.871.891.085.983.795.371.694.181.358.776.482.6
DGCNN [33]85.284.083.486.777.890.674.791.287.582.895.766.394.981.163.574.582.6
PointCNN [18]86.184.186.586.080.890.679.792.388.485.396.177.295.284.264.280.083.0
DRNet [50]86.484.385.088.379.591.279.391.889.085.295.772.294.282.060.676.884.2
PCT [7]86.485.082.489.081.291.971.591.388.186.395.864.695.883.662.277.683.7
MPCT [40]86.784.985.789.380.992.278.692.688.486.696.573.695.783.563.780.983.6
PreDiffPoint(ours)87.585.385.289.081.591.276.592.888.287.396.276.294.784.864.976.381.5
", + "bbox": [ + 96, + 89, + 901, + 214 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/58afa463ba8034b2aa9a38dfc741fca49958b8521a7f125e49a42a11f93cee9d.jpg", + "table_caption": [ + "Table 1. Part Segmentation Results (%) on the ShapeNetPart Dataset." + ], + "table_footnote": [], + "table_body": "
MethodsPre.OBJ-ONLYOBJ-BGPB-T50-RS
PointNet [25]-79.273.368.0
PointNet++ [26]-84.382.377.9
PointCNN [18]-85.586.178.5
DGCNN [33]-86.282.878.1
Transformer [44]-80.5579.8677.24
Point-Bert [44]✓88.1287.4383.07
Point-MAE [23]✓88.2990.0285.18
TAP [35]✓89.5090.3685.67
PointDif [48]✓91.9193.2987.61
PreDifPoint(ours)✓92.2593.4287.96
", + "bbox": [ + 91, + 263, + 498, + 446 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "thus improve the segmentation accuracy. Indoor Semantic Segmentation.", + "bbox": [ + 89, + 515, + 359, + 545 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further validate the semantic segmentation capability of our model in the indoor dataset S3DIS. The S3DIS dataset contains point cloud data and detailed semantic labels for several indoor environments, which contains a total of 6 regions (Area1, Area2, Area3, Area4, Area5, Area6), which are further divided into 271 rooms or spaces. We tested our model in Area5 while training in other areas.", + "bbox": [ + 89, + 546, + 482, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We used mIoU and mAcc to evaluate the performance of the model, and from Tab. 3, we can see that our mAcc lags slightly behind compared to PointDif, probably due to the model's limited ability to extract features in certain categories, but still achieves a more competitive accuracy. In addition, PreDifPoint improves by $1.8\\%$ in mIoU, which indicates that our model is able to understand the spatial relationships in the image well and segment different regions accurately. It is clear that our model obtains more meaningful geometric prior knowledge, which enables it to better understand the contextual semantics and local geometric relationships in complex scenes and achieve better performance.", + "bbox": [ + 89, + 654, + 482, + 848 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 863, + 243, + 878 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Masking ratio.", + "bbox": [ + 89, + 885, + 199, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/442e1d02e3496407b105d03e1874600ee66ab0f10f0f4da357fd95ec8528b14d.jpg", + "table_caption": [ + "Table 2. Object classification results on ScanObjectNN. We report the Overall Accuracy $(\\%)$ ." + ], + "table_footnote": [], + "table_body": "
MethodsPre.mIoUmAcc
PointNet [25]-41.149.0
PointNet++ [26]-53.5-
PointCNN [18]-57.363.9
KPConv [40]-67.172.8
Point-Bert [44]✓68.976.1
Point-MAE [23]✓68.476.2
PointDif [48]✓70.077.1
PreDifPoint(ours)✓71.876.5
", + "bbox": [ + 514, + 263, + 906, + 416 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Semantic Segmentation Results (%) Obtained on the S3DIS Dataset, Evaluated on Area 5.", + "bbox": [ + 511, + 426, + 906, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The masking operations that we include in the pre-training will affect the performance of the model in the downstream task. Thus, we first verified the effect of different mask ratios on the downstream tasks. As in Tab. 4, we examine the classification and segmentation ability of the model on the ScanObjectNN dataset and the S3DIS dataset, respectively. It is clear that the model achieves the best classification and segmentation performance when the mask ratio is 0.8.", + "bbox": [ + 511, + 487, + 906, + 608 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This may be due to the fact that the self-supervision task is too difficult when the masking ratio is too high, which may cause the model to have difficulty in learning effective feature representations, and will affect the model's ability to extract enough information from the limited training data to cope with the prediction task. However, when the masking ratio is too low, the model cannot be forced to learn deep feature representations and the model cannot obtain better prior knowledge. Only when the masking ratio is 0.8, it may just be beneficial for the model to learn effective feature representations.", + "bbox": [ + 511, + 609, + 908, + 776 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Attention.", + "text_level": 1, + "bbox": [ + 513, + 777, + 589, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The DXCA-Attention mechanism that we use in CGNet enhances model comprehension by considering feature relationships in both directions simultaneously, enabling a more comprehensive capture of information. To validate the utility of this attention mechanism, we report results using MLP, one-way covariate attention mechanism and DXCA-Attention mechanism on the ScanObjectNN dataset, respec", + "bbox": [ + 511, + 794, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26565", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4b66558844d29c73d80e8492172a1e86ccba167f749388a1419dc4eb15fdc5fd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetsScanObjectNNS3DIS(Area5)
Mask RatioOAmIoU
091.5869.1
0.590.9668.2
0.691.2369.5
0.791.8570.3
0.893.4271.8
0.992.3470.6
", + "bbox": [ + 91, + 88, + 491, + 232 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/15ce3738c2c45f1aacdadcc381b9c0f37b1b64a7d23c9f73c49f3928edc00819.jpg", + "table_caption": [ + "Table 4. The Overall Accuracy (%) on ScanObjectNN and the mIoU (%) on S3DIS with different masking ratios." + ], + "table_footnote": [], + "table_body": "
AttentionOA
MLP92.56
Covariance Attention93.19
DXCA-Attention93.42
", + "bbox": [ + 138, + 289, + 437, + 367 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tively.", + "bbox": [ + 89, + 424, + 135, + 439 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Tab. 5, our proposed DXCA-Attention mechanism obtains the best performance. It may be because MLP is able to capture the complex relationship between features through nonlinear transformations, but it is unable to take into account the specific interaction patterns between features. The one-way covariate attention mechanism, on the other hand, only considers the covariance relationship between features in one direction, which may have led to the loss of information. Only the DXCA-Attention mechanism can model the interactions between features more comprehensively, capture more complex nonlinear relationships, enhance the robustness of the model, and facilitate feature fusion, thus enhancing the expressive power of the model.", + "bbox": [ + 88, + 441, + 482, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Density-Adaptive Sampling Strategy.", + "text_level": 1, + "bbox": [ + 89, + 654, + 352, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The density-adaptive sampling strategy that we use in the diffusion process enables the model to focus more on the denser regions in the point cloud, which makes it easier for the model to recover the structural details of the denser regions. We compare this with the uniform timestep sampling strategy [48], and as shown in Tab. 6, the OA achieved by the density-adaptive sampling strategy is 0.11 higher than the uniform time-step sampling strategy.", + "bbox": [ + 89, + 671, + 482, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "It indicates that uniform time-step sampling may lead to under-sampling in regions with higher density and oversampling in regions with lower density. Since many real-world applications of point cloud data tend to have nonuniform densities, the density-adaptive sampling strategy can utilize the data more efficiently, especially in those high-density regions that are critical to model performance.", + "bbox": [ + 89, + 795, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f45a7be82f3ca020f1d9c4fc94b47ffd69129442f2dab709f637b5855bf59bb2.jpg", + "table_caption": [ + "Table 5. The Forms of The Self-Attention Operators." + ], + "table_footnote": [], + "table_body": "
Sampling StrategynOA
Recurrent uniform sampling93.31
Density-Adaptive Sampling93.42
", + "bbox": [ + 539, + 88, + 883, + 150 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. The Forms of The Sampling Strategy.", + "bbox": [ + 570, + 160, + 846, + 175 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 202, + 633, + 217 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a pre-training framework based on a diffusion model called PreDifPoint. Specifically, we propose a conditional generation network to guide the model through the point-to-point denoising process. In this context, the two-way covariate attention mechanism used facilitates the extraction and representation of point cloud features, thus enhancing the model's comprehension. We also introduce a density-adaptive sampling strategy based on the time step to guide the model to pay more attention to the denser regions in the point cloud, thus making it easier for the model to recover the structural details of the dense regions. Extensive experiments have shown that our method has good feature learning and representation capabilities, and is able to achieve advanced results in a variety of real-world datasets, which further validates the effectiveness of the proposed method. In the future, we hope that our work will further improve the performance of point cloud analysis.", + "bbox": [ + 511, + 227, + 906, + 498 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 514, + 515, + 609, + 530 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Tomer Amit, Tal Shaharbany, Eliya Nachmani, and Lior Wolf. Segdiff: Image segmentation with diffusion probabilistic models, 2022. 3", + "[2] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models, 2022. 3", + "[3] Adam Bretherton, Joshua J. Bon, David J. Warne, Kerrie Mengersen, and Christopher Drovandi. A principled approach to bayesian transfer learning, 2025. 3", + "[4] Tolgahan Cakaloglu and Xiaowei Xu. Mrnn: A multi-resolution neural network with duplex attention for document retrieval in the context of question answering, 2019. 2", + "[5] Shoufa Chen, Peize Sun, Yibing Song, and Ping Luo. Diffusional: Diffusion model for object detection, 2023. 3", + "[6] Sachin Goyal, Ananya Kumar, Sankalp Garg, Zico Kolter, and Aditi Raghunathan. Finetune like you pretrain: Improved finetuning of zero-shot vision models, 2022. 1", + "[7] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R. Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 7(2):187-199, 2021. 7", + "[8] Jing Hao, Song Chen, Xiaodi Wang, and Shumin Han. Language-aware multiple datasets detection pretraining for detrs, 2023. 1" + ], + "bbox": [ + 521, + 540, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26566", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[9] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners, 2021. 1", + "[10] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models, 2020. 1", + "[11] Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, Tom Hennigan, Eric Noland, Katie Millican, George van den Driessche, Bogdan Damoc, Aurelia Guy, Simon Osindero, Karen Simonyan, Erich Elsen, Jack W. Rae, Oriol Vinyals, and Laurent Sifre. Training compute-optimal large language models, 2022. 1", + "[12] Shouwei Hu, Xi Li, Banyao Ruan, and Zhihao Liu. An amplitude-encoding-based classical-quantum transfer learning framework: Outperforming classical methods in image recognition, 2025. 3", + "[13] Mohammad Rafid Ul Islam, Prasad Tadepalli, and Alan Fern. Self-attention-based diffusion model for time-series imputation in partial blackout scenarios, 2025. 3", + "[14] Liqiang Jing, Yiren Li, Junhao Xu, Yongcan Yu, Pei Shen, and Xuemeng Song. Vision enhanced generative pre-trained language model for multimodal sentence summarization. Machine Intelligence Research, 20:289-298, 2023. 1", + "[15] Li Ju, Xingyi Yang, Qi Li, and Xinchao Wang. Graphbridge: Towards arbitrary transfer learning in gnns, 2025. 3", + "[16] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything, 2023. 1", + "[17] Xudong Li, Li Feng, Lei Li, and Chen Wang. Few-shot meta-learning on point cloud for semantic segmentation, 2021. 3", + "[18] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on $\\mathcal{X}$ -transformed points, 2018. 7", + "[19] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation, 2023. 2", + "[20] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3d mesh in 45 seconds without pershape optimization, 2023. 2", + "[21] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2836-2844, 2021. 3", + "[22] Zhaoyang Lyu, Zhifeng Kong, Xudong Xu, Liang Pan, and Dahua Lin. A conditional point diffusion-refinement paradigm for 3d point cloud completion, 2022. 3", + "[23] Yatian Pang, Wenxiao Wang, Francis E. H. Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning, 2022. 1, 2, 7", + "[24] Konpat Preechakul, Nattanat Chathee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion autoencoders: Toward a meaningful and decodable representation, 2022. 2" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation, 2017. 7", + "[26] Charles R. Qi, Li Yi, Hao Su, and Leonidas J. Guibas. Point-net++: Deep hierarchical feature learning on point sets in a metric space, 2017. 7", + "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 1", + "[28] Mojtaba Safari, Shansong Wang, Zach Eidex, Qiang Li, Erik H. Middlebrooks, David S. Yu, and Xiaofeng Yang. Mri super-resolution reconstruction using efficient diffusion probabilistic model with residual shifting, 2025. 3", + "[29] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior, 2023. 2", + "[30] Aaron van den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks, 2016. 2", + "[31] Aaron van den Oord, Nal Kalchbrenner, Oriol Vinyals, Lasse Espeholt, Alex Graves, and Koray Kavukcuoglu. Conditional image generation with pixelCNN decoders, 2016. 2", + "[32] Jiahui Wang, Haiyue Zhu, Haoren Guo, Abdullah Al Mamun, Cheng Xiang, and Tong Heng Lee. Few-shot point cloud semantic segmentation via contrastive self-supervision and multi-resolution attention, 2023. 1", + "[33] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E. Sarma, Michael M. Bronstein, and Justin M. Solomon. Dynamic graph cnn for learning on point clouds, 2019. 7", + "[34] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation, 2023. 2", + "[35] Ziyi Wang, Xumin Yu, Yongming Rao, Jie Zhou, and Jiwen Lu. Take-a-photo: 3d-to-2d generative pre-training of point cloud models, 2023. 7", + "[36] Zhendong Wang, Jianmin Bao, Shuyang Gu, Dong Chen, Wengang Zhou, and Houqiang Li. Design diffusion: High-quality text-to-design image generation with diffusion models, 2025. 3", + "[37] Antoine Wehenkel and Gilles Louppe. Diffusion priors in variational autoencoders, 2021. 2", + "[38] Julia Wolleb, Robin Sandkuhler, Florentin Bieder, Philippe Valmaggia, and Philippe C. Cattin. Diffusion models for implicit image segmentation ensembles, 2021. 3", + "[39] Jay Zhangjie Wu, Yuxuan Zhang, Haithem Turki, Xuanchi Ren, Jun Gao, Mike Zheng Shou, Sanja Fidler, Zan Gojcic, and Huan Ling. Difix3d+: Improving 3d reconstructions with single-step diffusion models, 2025. 3", + "[40] Yue Wu, Jiaming Liu, Maoguo Gong, Zhixiao Liu, Qiguang Miao, and Wenping Ma. Mpct: Multiscale point cloud transformer with a residual network. IEEE Transactions on Multimedia, 26:3505-3516, 2024. 7", + "[41] Weilai Xiang, Hongyu Yang, Di Huang, and Yunhong Wang. Denoising diffusion autoencoders are unified self-supervised learners, 2023. 2" + ], + "bbox": [ + 516, + 90, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26567", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Saining Xie, Jiatao Gu, Demi Guo, Charles R. Qi, Leonidas J. Guibas, and Or Litany. Pointcontrast: Unsupervised pre-training for 3d point cloud understanding, 2020. 2", + "[43] Jiale Xu, Xintao Wang, Weihao Cheng, Yan-Pei Cao, Ying Shan, Xiaohu Qie, and Shenghua Gao. Dream3d: Zero-shot text-to-3d synthesis using 3d shape prior and text-to-image diffusion models, 2023. 2", + "[44] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling, 2022. 1, 7", + "[45] Renrui Zhang, Ziyu Guo, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, Hongsheng Li, and Peng Gao. Pointm2ae: Multi-scale masked autoencoders for hierarchical point cloud pre-training, 2022. 1, 2", + "[46] Zaiwei Zhang, Rohit Girdhar, Armand Joulin, and Ishan Misra. Self-supervised pretraining of 3d features on any point-cloud, 2021. 1", + "[47] Luda Zhao, Yihua Hu, Xing Yang, Zhenglei Dou, and Qilong Wu. Icddpm: Image-conditioned denoising diffusion probabilistic model for real-world complex point cloud single view reconstruction. Expert Systems with Applications, 259:125370, 2025. 3", + "[48] Xiao Zheng, Xiaoshui Huang, Guofeng Mei, Yuenan Hou, Zhaoyang Lyu, Bo Dai, Wanli Ouyang, and Yongshun Gong. Point cloud pre-training with diffusion models, 2023. 1, 2, 3, 6, 7, 8", + "[49] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion, 2021. 3", + "[50] Mingjian Zhu, Kai Han, Enhua Wu, Qiulin Zhang, Ying Nie, Zhenzhong Lan, and Yunhe Wang. Dynamic resolution network, 2021. 7", + "[51] Aleksei Zhuravlev, Zorah Lähner, and Vladislav Golyanik. Denoising functional maps: Diffusion models for shape correspondence, 2025. 3" + ], + "bbox": [ + 89, + 92, + 482, + 574 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26568", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Feature Extraction and Representation of Pre-training Point Cloud Based on Diffusion Models/0c060285-4bbf-430f-b421-79f016d6dd03_model.json b/2025/Feature Extraction and Representation of Pre-training Point Cloud Based on Diffusion Models/0c060285-4bbf-430f-b421-79f016d6dd03_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ae39d30e4dd691e1c1f2135afa17b466c8e49a31 --- /dev/null +++ b/2025/Feature Extraction and Representation of Pre-training Point Cloud Based on Diffusion Models/0c060285-4bbf-430f-b421-79f016d6dd03_model.json @@ -0,0 +1,2398 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.13, + 0.885, + 0.175 + ], + "angle": 0, + "content": "Feature Extraction and Representation of Pre-training Point Cloud Based on Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.236, + 0.205, + 0.327, + 0.222 + ], + "angle": 0, + "content": "Chang Qiu" + }, + { + "type": "text", + "bbox": [ + 0.199, + 0.223, + 0.365, + 0.24 + ], + "angle": 0, + "content": "Southeast University" + }, + { + "type": "text", + "bbox": [ + 0.192, + 0.243, + 0.372, + 0.256 + ], + "angle": 0, + "content": "230238519@seu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.447, + 0.205, + 0.551, + 0.222 + ], + "angle": 0, + "content": "Feipeng Da*" + }, + { + "type": "text", + "bbox": [ + 0.413, + 0.223, + 0.58, + 0.24 + ], + "angle": 0, + "content": "Southeast University" + }, + { + "type": "text", + "bbox": [ + 0.429, + 0.243, + 0.564, + 0.256 + ], + "angle": 0, + "content": "dafp@seu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.665, + 0.205, + 0.761, + 0.222 + ], + "angle": 0, + "content": "Zilei Zhang" + }, + { + "type": "text", + "bbox": [ + 0.629, + 0.223, + 0.796, + 0.24 + ], + "angle": 0, + "content": "Southeast University" + }, + { + "type": "text", + "bbox": [ + 0.622, + 0.243, + 0.8, + 0.256 + ], + "angle": 0, + "content": "230238524@seu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.323, + 0.485, + 0.794 + ], + "angle": 0, + "content": "The pretrain-finetune paradigm of pre-training a model on large amounts of image and text data and then fine-tuning the model for a specific task has led to significant progress in many 2D image and natural language processing tasks. Similarly, the use of pre-training methods in point cloud data can also enhance the working performance and generalization ability of the model. Therefore, in this paper, we propose a pre-training framework based on a diffusion model called PreDifPoint. It is able to accomplish the pretraining of the model's backbone network through a diffusion process of gradual denoising. We aggregate the potential features extracted from the backbone network, input them as conditions into the subsequent diffusion model, and direct the point-to-point mapping relationship of the noisy point clouds at neighboring time steps, so as to generate high-quality point clouds and at the same time better perform various downstream tasks of the point clouds. We also introduce a bi-directional covariate attention (DXCA-Attention) mechanism for capturing complex feature interactions, fusing local and global features, and improving the detail recovery of point clouds. In addition, we propose a density-adaptive sampling strategy, which can help the model dynamically adjust the sampling strategy between different time steps, and guide the model to pay more attention to the denser regions in the point cloud, thus improving the effectiveness of the model in point cloud recovery. Our PreDifPoint framework achieves more competitive results on various real-world datasets. Specifically, PreDifPoint achieves an overall accuracy of \\(87.96\\%\\), which is \\(0.35\\%\\) higher than PointDif, on the classification task on PB-T50-395RS, a variant of ScanObjectNN dataset." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.821, + 0.222, + 0.837 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.847, + 0.483, + 0.879 + ], + "angle": 0, + "content": "Recent work has demonstrated that the pretrain-finetune paradigm, where a model is first pretrained on large" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.293, + 0.908, + 0.49 + ], + "angle": 0, + "content": "amounts of image and text data and then fine-tuned for specific tasks, has led to substantial advancements in many 2D image and natural language processing tasks [11, 16, 27]. Pretraining the backbone on large-scale datasets enables the model to learn rich feature representations while also avoiding the need for large amounts of data and computational resources required for training a model from scratch [8, 9, 32]. Pretrained models, after fine-tuning, can learn task-specific knowledge, exhibit better generalization capabilities, and converge to optimal performance levels more rapidly [6, 14]. Similarly, using pretraining methods for point cloud data can also enhance the model's performance and generalization ability [44, 46, 48]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.491, + 0.909, + 0.734 + ], + "angle": 0, + "content": "Generative pretraining methods, through generative models, capture the complex structure and intricate details of point cloud data, thereby obtaining more powerful feature representations. Early works such as PointBERT [44], Point-MAE [23], and Point-M2AE [45] employed self-supervised learning by reconstructing masked local point patches, forcing the model to learn local point cloud information. However, this masking and reconstruction approach may overlook the contextual relationships between masked and unmasked regions, preventing the model from effectively understanding the global structure of the object or scene. Inspired by diffusion models [10], which can understand global structure and long-term dependencies through the process of adding and removing noise, we propose using a diffusion model to guide the pretraining process of point clouds." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Thus, we introduce a diffusion-based pretraining framework, named PreDifPoint. This framework pretrains only the backbone network of the model, and then fine-tunes the task-specific parts. The pretraining of the backbone network is achieved through a progressive denoising diffusion process, as shown in Fig. 1. By pretraining the backbone, it can learn meaningful features during the feature extraction phase, which also significantly enhances the model's generalization ability. At the same time, this reduces the number of parameters that need to be adjusted during training, accelerating convergence and reducing overfitting." + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.237, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26559" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.092, + 0.482, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.218, + 0.485, + 0.276 + ], + "angle": 0, + "content": "Figure 1. Schematic illustration of our PreDiffPoint. Our PreDiffPoint can be used to train different backbones and reconstruct the original point cloud from the noisy one through a diffusion process." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.304, + 0.483, + 0.529 + ], + "angle": 0, + "content": "Specifically, the PreDifPoint framework consists of three core modules: the Conditional Point Generation Network (CGNet), the Conditional Denoising Network (CDNet), and the Conditional Diffusion Network (CDPNet). CGNet aggregates the latent features extracted by the backbone and feeds them as conditional inputs into the subsequent diffusion model. CDNet utilizes these conditional inputs to perform denoising and restore point cloud details. CDPNet, through the point-to-point mapping of noisy point clouds between adjacent time steps in the diffusion process [48], ensures that the generated point clouds excel in both global consistency and local detail. With this design, the model can generate high-quality point clouds based on various conditions, exhibiting strong diversity and adaptability to perform various downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.532, + 0.483, + 0.683 + ], + "angle": 0, + "content": "Additionally, we propose a density-adaptive sampling strategy. We divide the entire time step range into several subintervals, each containing multiple time steps. Within each subinterval, the sampling probability is dynamically adjusted according to the local density of the point cloud at each time step, prioritizing sampling from steps with higher density. This strategy allows for more fine-grained control over the sampling process at each time step and helps dynamically adjust the sampling strategy across time steps, improving the model's point cloud recovery performance." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.684, + 0.473, + 0.699 + ], + "angle": 0, + "content": "Our main contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.705, + 0.483, + 0.809 + ], + "angle": 0, + "content": "- We propose a diffusion-based point cloud pretraining framework, PreDifPoint, which leverages conditional guidance for generation, denoising, and recovery, as well as global and local feature modeling. This framework enhances the performance of point cloud generation and denoising tasks and provides high-quality feature representations for downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "- We introduce a DXCA-Attention mechanism into the model to capture complex global feature interactions and generate more instructive condition vectors to enhance the detail reconstruction of the point cloud. The condition vectors obtained through CGNet will be combined with time-step information and passed to the subsequent" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.705, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.528, + 0.093, + 0.68, + 0.107 + ], + "angle": 0, + "content": "network for denoising." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.107, + 0.905, + 0.197 + ], + "angle": 0, + "content": "- We propose a density-adaptive sampling strategy, which can help the model dynamically adjust the sampling strategy between different time steps, and guide the model to pay more attention to the denser regions in the point cloud, thus improving the effectiveness of the model in point cloud recovery." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.198, + 0.905, + 0.303 + ], + "angle": 0, + "content": "- The PreDifPoint framework requires only fine-tuning for task-specific components (such as classifiers), allowing the model to perform optimally across various tasks, particularly in data-limited or computationally constrained environments. Notably, the PreDifPoint framework demonstrates competitive performance in a variety of practical downstream tasks." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.107, + 0.905, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.32, + 0.664, + 0.336 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.346, + 0.736, + 0.363 + ], + "angle": 0, + "content": "2.1. Point Cloud Pretraining" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.369, + 0.907, + 0.534 + ], + "angle": 0, + "content": "In deep learning, pretraining has become a key technique for enhancing model performance, especially when data is scarce or labels are limited. For point cloud learning tasks, pretraining is particularly important. Point cloud data typically has complex spatial structures and sparsity, and traditional training methods often rely on large amounts of labeled data, which is difficult to obtain in real-world applications. Therefore, learning effective point cloud representations through pretraining not only reduces reliance on labeled data but also improves the model's generalization and robustness in downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.906, + 0.612 + ], + "angle": 0, + "content": "Currently, point cloud pretraining methods can be broadly categorized into three types: self-supervised learning, generative models [29, 34, 43], and transfer learning. Each approach has its unique advantages and application scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.906, + 0.84 + ], + "angle": 0, + "content": "Self-supervised methods [4, 30, 31] can mask parts of the data and train the model to reconstruct the missing parts or optimize the model's representation ability through strategies like contrastive learning. Point-MAE [23] randomly masks certain points in the point cloud and trains the model to recover these missing points, thereby effectively learning the geometric features of the point cloud. Point-M2AE [45] for hierarchical self-supervised learning of 3D point clouds modifies the encoder and decoder into a pyramid architecture to incrementally model the spatial geometry and capture the fine-grained and high-level semantics of 3D shapes. PointContrast [42] employs contrastive learning to treat point cloud representation as a learning task, improving the model's ability to distinguish between similar and different point cloud segments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Generative models [19, 20, 24, 37, 41], particularly denoising diffusion models and variational autoencoders, have made significant progress in point cloud pretraining in recent years. These models learn the generative distribution of" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26560" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.227 + ], + "angle": 0, + "content": "point cloud data, capturing the details of the point cloud and completing missing data. DDPM [22] demonstrates how to use pretrained denoising diffusion models for conditional point cloud generation and completion, significantly improving the quality of point cloud reconstruction. PVD [49] introduces a conditional variational diffusion model (CV-Diffusion) that can generate high-quality 3D shapes through conditional information, effectively applied to point cloud generation and completion tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.228, + 0.482, + 0.349 + ], + "angle": 0, + "content": "Transfer learning methods [3, 12, 15] involve pretraining a model on a large-scale dataset and then transferring it to a specific task, significantly improving performance when data is limited. Meta-PointNet [17] proposes a meta-learning-based pretraining method for point cloud classification and segmentation, where knowledge is shared across multiple tasks, greatly improving performance on few-shot tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.362, + 0.26, + 0.377 + ], + "angle": 0, + "content": "2.2. Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.385, + 0.483, + 0.627 + ], + "angle": 0, + "content": "The core idea of diffusion models [13, 28, 36, 39, 51] is to simulate a process of gradually adding noise and then recovering the data through reverse denoising. Diffusion models have been widely used in recognition tasks such as object detection [5] and semantic discrimination [1, 2, 38]. During training, the model learns how to recover real data from noise, which helps generate high-quality samples. This paper uses conditional diffusion models, which introduce additional conditional information into the diffusion process to guide the generated model. These conditional inputs can include labels, textual descriptions, or other external features that help the model generate data that meets specific conditions. Conditional diffusion models have shown excellent performance in multimodal generation tasks, particularly in image generation, point cloud generation, and speech synthesis." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.854 + ], + "angle": 0, + "content": "Diffusion probabilistic models [21] proposed a point cloud probabilistic generation model inspired by nonequilibrium thermodynamics that utilizes a backward diffusion process to learn point distributions. And they achieved competitive performance in point cloud generation and automatic coding, and comparable results in unsupervised representation learning. PointDif [48] proposes a conditional diffusion model for point cloud pretraining tasks. This model can more precisely control the structure and details of the point cloud during the generation or denoising process, performing particularly well in point cloud denoising and generation tasks. ICDDPM [47] can uses 2D images as prior information to guide end-to-end 3D point cloud generation, and obtain high-quality 3D point clouds of different types of targets from real-world images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The PreDifPoint framework we have developed pretrains the network by recovering the original point cloud from randomly noisy point clouds, forcing the network to learn local" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.213 + ], + "angle": 0, + "content": "and global geometric priors of point clouds. The extracted point cloud features are passed as conditional information into the diffusion process, guiding the generation of point clouds step by step. Additionally, we propose a density-adaptive sampling strategy, which allows for finer control over the sampling process at each time step and helps dynamically adjust the sampling strategy between time steps, thus improving the model's ability to recover point clouds." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.225, + 0.605, + 0.24 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.251, + 0.907, + 0.463 + ], + "angle": 0, + "content": "In our pre-training process, the backbone of the overall framework PreDifPoint is a transformer encoder. Notably, this backbone can be replaced with any other backbone network for pre-training. The input point cloud is grouped and features are extracted, followed by masking operations and contextual enhancement via a Transformer with positional encoding. Subsequently, CGNet further processes these features to generate conditional vectors that serve as conditional inputs to CDNet. In CDNet, time-step embedding is combined with conditional vectors. Finally, the CDPNet network integrates CGNet and CDNet to generate the original point cloud by stepwise denoising the data through a diffusion model. The pipeline of our PreDifPoint is shown in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.471, + 0.684, + 0.485 + ], + "angle": 0, + "content": "3.1. Diffusion Process" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.906, + 0.553 + ], + "angle": 0, + "content": "In the diffusion model, a clean point cloud \\( X^0 \\epsilon R^{n \\times 3} \\) is gradually corrupted into a fully noisy point cloud \\( X^T \\) by incrementally adding noise. The diffusion process at each time step \\( t \\) follows:" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.578, + 0.907, + 0.597 + ], + "angle": 0, + "content": "\\[\nq \\left(X ^ {t} \\mid X ^ {t - 1}\\right) = \\mathcal {N} \\left(X ^ {t}; \\sqrt {\\alpha_ {t}} X ^ {t - 1}, (1 - \\alpha_ {t}) \\mathrm {I}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.606, + 0.906, + 0.666 + ], + "angle": 0, + "content": "where \\(\\alpha_{t} = 1 - \\beta_{t}\\) denotes the noise attenuation coefficient at time step, and \\(\\beta_{t}\\) represents the noise level. Through this process, noise is gradually added to the point cloud until it becomes nearly random at \\(X^T\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.667, + 0.906, + 0.727 + ], + "angle": 0, + "content": "This reverse process predicts the denoising signals at each time step by the neural network \\( p_{\\theta}(X^t | X^{t-1}, c) \\), which gradually recovers the original structure of the point cloud as shown in the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.751, + 0.907, + 0.77 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(X ^ {t} \\mid X ^ {t - 1}, c\\right) = \\mathcal {N} \\left(X ^ {t - 1}; \\mu_ {\\theta} \\left(X ^ {t}, t, c\\right), \\sigma_ {t} ^ {2} \\mathrm {I}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.779, + 0.906, + 0.824 + ], + "angle": 0, + "content": "where \\(\\mu_{\\theta}(X^{t},t,c)\\) is the mean predicted by the network, \\(X^t\\) representing the denoised result at step \\(t\\), and \\(\\sigma_t^2\\) is the noise variance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.834, + 0.874, + 0.849 + ], + "angle": 0, + "content": "3.2. Conditional Generation Network (CGNet)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "CGNet is responsible for extracting global and local features from the point cloud, fusing these features through a DXCA-Attention mechanism to more comprehensively" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26561" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.176, + 0.092, + 0.823, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.324, + 0.907, + 0.395 + ], + "angle": 0, + "content": "Figure 2. Pipeline of PreDifPoint. We first group the input point cloud and extract features, which are then subjected to masking operations and contextual enhancement via a transformer with position encoding. Subsequently, CGNet further processes these features to generate conditional vectors, which are used as conditional inputs to CDNet. In CDNet, time-step embedding is combined with conditional vectors. Finally, the CDPNet network integrates CGNet and CDNet to generate the original point cloud by stepwise denoising the data through a diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.421, + 0.482, + 0.496 + ], + "angle": 0, + "content": "model the interactions between the features and better adapt to different data patterns and structures, as shown in Fig. 3. Finally, a conditional vector \\( c \\) associated with the density distribution is generated, which is used to guide the subsequent denoising process." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.497, + 0.483, + 0.588 + ], + "angle": 0, + "content": "We first project the input feature \\( F_{loc} \\) from the point cloud chunking encoding to the high dimensional space through the multilayer perceptron (MLP) to obtain the projected representation \\( F_{pro} \\) of the local features, and then obtain the global feature \\( F_{glo} \\) through the global pooling operation on \\( F_{pro} \\) as shown in the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.599, + 0.483, + 0.641 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} F _ {p r o} = M L P (F), \\\\ F _ {g l o} = G l o b a l M a x P o o l \\left(F _ {p r o}\\right). \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.647, + 0.484, + 0.723 + ], + "angle": 0, + "content": "Then, the attention weights are obtained by calculating the covariance matrix in both directions to capture the global contextual relationships and complex feature interactions. In this process, we need to first obtain the important parameters, \\( Q \\), \\( K \\), and \\( V \\), as shown in Eq. (4)." + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.733, + 0.483, + 0.816 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} Q = \\operatorname {r e s h a p e} \\left(F _ {p r o}\\right) \\in \\mathbb {R} ^ {B \\times G \\times C}, \\\\ K = \\operatorname {r e s h a p e} \\left(F _ {g l o}\\right) \\in \\mathbb {R} ^ {B \\times G \\times C}, \\\\ V _ {1} = \\operatorname {r e s h a p e} \\left(F _ {g l o}\\right) \\in \\mathbb {R} ^ {B \\times G \\times C}, \\\\ V _ {2} = \\operatorname {r e s h a p e} \\left(F _ {p r o}\\right) \\in \\mathbb {R} ^ {B \\times G \\times C}. \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.821, + 0.483, + 0.852 + ], + "angle": 0, + "content": "Then centering along the feature dimension \\(C\\) as shown in Eq. (5)." + }, + { + "type": "equation", + "bbox": [ + 0.219, + 0.864, + 0.483, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} \\bar {Q} = Q - \\mathbb {E} _ {c} [ Q ], \\\\ \\bar {K} = K - \\mathbb {E} _ {c} [ K ]. \\end{array} \\right. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.42, + 0.907, + 0.452 + ], + "angle": 0, + "content": "Then the scaling factor \\(\\sqrt{C}\\) is introduced to compute the covariance matrix, as shown in Eq. (6)." + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.472, + 0.905, + 0.576 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} A t t e n S c o r e _ {1} = S o f t m a x \\left(\\frac {\\bar {Q} \\cdot \\bar {K} ^ {T}}{\\sqrt {C}}\\right) \\in \\mathbb {R} ^ {B \\times G \\times C}, \\\\ A t t e n S c o r e _ {2} = S o f t m a x \\left(\\frac {\\bar {K} \\cdot \\bar {Q} ^ {T}}{\\sqrt {C}}\\right) \\in \\mathbb {R} ^ {B \\times G \\times C}. \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.577, + 0.907, + 0.607 + ], + "angle": 0, + "content": "At this point, the bi-directional covariance features can be aggregated as shown in Eq. (7)." + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.616, + 0.905, + 0.677 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} O _ {1} = A t t e n S c o r e _ {1} \\cdot V _ {1} \\in \\mathbb {R} ^ {B \\times G \\times C}, \\\\ O _ {2} = A t t e n S c o r e _ {2} \\cdot V _ {2} \\in \\mathbb {R} ^ {B \\times G \\times C}, \\\\ O = \\operatorname {c o n c a t} \\left(O _ {1}, O _ {2}\\right) \\in \\mathbb {R} ^ {B \\times G \\times 2 C}. \\end{array} \\right. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.681, + 0.907, + 0.728 + ], + "angle": 0, + "content": "Finally, we fuse the projected features, global features and the attention-weighted features to output the conditioned vector \\( c \\), as shown in Eq. (8)." + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.741, + 0.905, + 0.759 + ], + "angle": 0, + "content": "\\[\nc = M L P \\left(\\text {C o n c a t} \\left(F _ {\\text {p r o}}, A, F _ {\\text {g l o}}\\right)\\right). \\tag {8}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.773, + 0.862, + 0.79 + ], + "angle": 0, + "content": "3.3. Conditional Denoising Network (CDNet)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.856 + ], + "angle": 0, + "content": "CDNet is responsible for gradually removing the noise and restoring the point cloud structure during the diffusion process, and its core is to achieve layer-level feature optimization through residual connectivity." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.907, + 0.902 + ], + "angle": 0, + "content": "First, the conditional vector \\( c \\) and the time embedding \\( t_{emd} \\) are fused into the point cloud features through a dynamic gating mechanism to achieve density-aware local de" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26562" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.228, + 0.485, + 0.381 + ], + "angle": 0, + "content": "Figure 3. DXCA-Attention Module. This attention mechanism achieves cross-level information fusion through the two-way interaction between global and local features. In the local-to-global direction, query vectors and key vectors are generated by centering local features and global features respectively, and the contribution weights of local regions to global semantics are computed through the covariance matrix, and the semantic consistency of local features is enhanced after weighted aggregation. In the global-to-local direction, the global features are used as queries and local features as keys, and the modulation relation of global context to local details is captured by covariance attention." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.409, + 0.484, + 0.439 + ], + "angle": 0, + "content": "noising. The time embedding \\( t_{emd} \\) and the fused features can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.464, + 0.482, + 0.516 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} t _ {e m d} = \\left[ \\beta_ {t}, \\sin \\left(\\beta_ {t}\\right), \\cos \\left(\\beta_ {t}\\right) \\right], \\\\ H _ {l} = \\sigma \\left(W _ {g} [ c \\oplus t _ {e m d} ]\\right) \\odot \\left(W _ {f} H _ {l - 1}\\right) + W _ {b} [ c \\oplus t _ {e m d} ], \\end{array} \\right. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.517, + 0.484, + 0.577 + ], + "angle": 0, + "content": "where \\(\\sigma\\) is the Sigmoid function for generating gating weights in the range [0, 1], and \\(W_{g}, W_{f}, W_{b}\\) are the gating generation weights, feature transformation weights and bias generation weights, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.578, + 0.484, + 0.623 + ], + "angle": 0, + "content": "Finally, to alleviate the gradient vanishing problem, the output of each of our layers can be connected using residuals such that the output of each layer is represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.639, + 0.483, + 0.656 + ], + "angle": 0, + "content": "\\[\nH _ {l} = H _ {l - 1} + C D N e t \\left(H _ {l - 1}, c, t _ {e m d}\\right). \\tag {10}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.664, + 0.483, + 0.68 + ], + "angle": 0, + "content": "3.4. Conditional Diffusion Point Network (CDPNet)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.687, + 0.484, + 0.747 + ], + "angle": 0, + "content": "The CDPNet integrates a density-weighted diffusion process, a conditional point generation network (CGNet), and a conditional denoising network (CDNet) to achieve stepwise generation from noisy to clean point clouds." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.748, + 0.484, + 0.824 + ], + "angle": 0, + "content": "In a nutshell, it is to first get a completely noisy point cloud \\( X^T \\) from a clean point cloud \\( X^0 \\) in the forward diffusion process, then obtain the conditional vector \\( \\mathbf{c} \\) by CGNet, and then update \\( X^T \\) step by step from \\( t = T \\) to \\( t = 1 \\) in CDNet, and finally get the reconstructed point cloud \\( X^0 \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.34, + 0.851 + ], + "angle": 0, + "content": "3.5. Density-Adaptive Sampling" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To dynamically adjust the time step scheduling in the diffusion model, we propose a density-adaptive sampling strategy. This strategy weights the sampling probability of each" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.135 + ], + "angle": 0, + "content": "time step based on the local density of points, prioritizing denser regions (e.g., edges) during training to enhance detail reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.137, + 0.905, + 0.168 + ], + "angle": 0, + "content": "First, the local density \\(\\rho_{i}\\) of each point \\(x_{i}\\epsilon X_{0}\\) is computed using \\(k\\)-nearest neighbors:" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.18, + 0.905, + 0.217 + ], + "angle": 0, + "content": "\\[\n\\rho \\left(x _ {i}\\right) = \\frac {1}{\\sum_ {j = 1} ^ {k} \\left\\| x _ {i} - x _ {j} \\right\\| + \\epsilon}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.224, + 0.905, + 0.269 + ], + "angle": 0, + "content": "where \\( k \\) is the neighborhood size, \\( \\| x_{i} - x_{j}\\| \\) is the Euclidean distance, and \\( \\epsilon \\) is a small constant to avoid division by zero." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.27, + 0.906, + 0.332 + ], + "angle": 0, + "content": "The time step interval \\([1, T]\\) is uniformly divided into \\(h\\) sub-intervals \\(\\{Q_1, Q_2, \\dots, Q_h\\}\\), each corresponding to different noise levels. The sampling probability for \\(Q_i\\) is weighted by the local density within the interval:" + }, + { + "type": "equation", + "bbox": [ + 0.636, + 0.343, + 0.905, + 0.375 + ], + "angle": 0, + "content": "\\[\nP \\left(Q _ {i} \\mid \\rho\\right) = \\frac {\\rho_ {i}}{\\sum_ {i = 1} ^ {T} \\rho_ {i}}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.382, + 0.905, + 0.427 + ], + "angle": 0, + "content": "This weighting strategy ensures that more samples can be performed in the denser regions, thus improving the reconstruction accuracy in these regions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.428, + 0.906, + 0.534 + ], + "angle": 0, + "content": "Within each interval \\( Q_{i} \\), we randomly sample a time step \\( t_{i} \\) from that interval and compute the diffusion process at that time step. Since the sampling is based on local density weighting, the sampling probability for each interval reflects the denseness of the region. In updating the sampling, we use the following weighted loss function to calculate the loss at each time step:" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.546, + 0.905, + 0.587 + ], + "angle": 0, + "content": "\\[\nL (\\theta , \\rho) = \\frac {1}{h} \\sum_ {i = 0} ^ {h - 1} L \\left(\\theta , \\rho , t _ {i} \\sim Q _ {i}\\right), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.596, + 0.873, + 0.612 + ], + "angle": 0, + "content": "where \\(L(\\theta, \\rho, t_i \\sim Q_i)\\) is the loss for the sampled \\(t_i\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.622, + 0.803, + 0.639 + ], + "angle": 0, + "content": "3.6. Density-Weighted Diffusion Loss" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.906, + 0.72 + ], + "angle": 0, + "content": "In the training of diffusion models combined with density-adaptive sampling, the loss function consists mainly of the error in the prediction noise and the error in the density weighting. Our goal is to minimize the difference between the predicted noise and the real noise at each time step." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.721, + 0.906, + 0.766 + ], + "angle": 0, + "content": "The standard diffusion model loss is achieved by calculating the mean square error (MSE) between the predicted noise and the real noise, as shown in Eq. (14):" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.79, + 0.911, + 0.825 + ], + "angle": 0, + "content": "\\[\nL _ {d e n o i s e} (\\theta) = \\mathbb {E} _ {x _ {0}, t, \\epsilon} [ \\| \\epsilon - \\epsilon_ {\\theta} (\\sqrt {\\alpha_ {t}} x _ {0} + \\sqrt {1 - \\alpha_ {t}} \\epsilon , t, c) \\| ^ {2} ], \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.902 + ], + "angle": 0, + "content": "where \\(x_0\\) is the input original point cloud, \\(t\\) is the current time step, \\(\\epsilon_{\\theta}\\) is the noise predicted by the network, \\(\\epsilon\\) is the noise actually added, \\(\\alpha_{t}\\) is the noise attenuation coefficient at each time step, \\(c\\) is the conditional information generated by the conditional aggregation network CGNet, and \\(\\mathbb{E}_{x_0,t,\\epsilon}\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26563" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "is the indication of the need for averaging all the possible values of these variables to calculate the final loss." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.123, + 0.484, + 0.199 + ], + "angle": 0, + "content": "During the density-adaptive sampling process, we need to adjust the computation of the loss to be consistent with the sampling probability of each interval. The final training loss combining density-weighted sampling and the standard MSE loss can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.224, + 0.48, + 0.332 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L (\\theta , \\rho) = \\frac {1}{h} \\sum_ {i = 0} ^ {h - 1} \\mathbb {E} _ {t _ {i} \\sim Q _ {i}} [ \\\\ \\begin{array}{l} \\mathbb {E} _ {x _ {0}, \\epsilon} \\left[ \\left\\| \\epsilon - \\epsilon_ {\\theta} \\left(\\sqrt {\\alpha_ {t _ {i}}} x _ {0} + \\sqrt {1 - \\alpha_ {t _ {i}}} \\epsilon , t _ {i}, c\\right) \\right\\| ^ {2} \\right] \\\\ ], \\end{array} \\tag {15} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.335, + 0.484, + 0.395 + ], + "angle": 0, + "content": "where \\( t_i \\) is the time step sampled from the interval \\( Q_i \\), and the loss function enhances the reconstruction of these regions by weighting the samples at each time step to ensure that denser regions can have higher weights." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.41, + 0.223, + 0.427 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.435, + 0.271, + 0.452 + ], + "angle": 0, + "content": "4.1. Pre-training Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.458, + 0.484, + 0.608 + ], + "angle": 0, + "content": "We use ShapeNet to pre-train our model, with the core version of this dataset containing 51,300 3D models across 55 common object categories. For each 3D shape, we sample 1,024 points to serve as the input for the model. Following the approach in PointDif [48], we use 41,952 shapes as the training set. Additionally, we employ the KNN algorithm to select the \\( k = 32 \\) nearest points as point patches, resulting in 64 point cloud patches. Furthermore, we set the embedding dimension of the Transformer encoder to 384, with 6 attention heads, and the condition dimension to 768." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.61, + 0.484, + 0.655 + ], + "angle": 0, + "content": "During pre-training, we set the weight decay to 0.05 and the learning rate to 0.001. Our model is pre-trained for 300 epochs with a batch size of 192." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.656, + 0.484, + 0.806 + ], + "angle": 0, + "content": "Additionally, we visualize the point clouds generated by the PreDifPoint framework to demonstrate the effectiveness of our pre-training. As shown in Fig. 4, we use the masked point cloud features, extracted and used as conditions, to guide the diffusion model in generating the original point cloud, with a masking rate of 0.8. The visualization results indicate that the PreDifPoint framework generates high-quality point clouds and that the learned geometric priors effectively guide subsequent tasks such as recognition and detection." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.273, + 0.833 + ], + "angle": 0, + "content": "4.2. Downstream Tasks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The performance of a pre-trained model, obtained in a fine-tuning task, is a test of the effectiveness of the pre-trained model. Since a high-quality point cloud pre-trained model should have a high hierarchical geometric prior, we tested" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.097, + 0.905, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.264, + 0.907, + 0.321 + ], + "angle": 0, + "content": "Figure 4. Visualization results on the ShapeNet validation set. We show a total of six categories of point clouds, each showing the input point cloud, the masked point cloud, and the reconstructed point cloud from left to right." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.35, + 0.906, + 0.379 + ], + "angle": 0, + "content": "the performance of the pre-trained model on a large number of real-world datasets." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.38, + 0.668, + 0.394 + ], + "angle": 0, + "content": "Object Classification." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.397, + 0.906, + 0.427 + ], + "angle": 0, + "content": "We first evaluated the classification ability of the PreD-ifPoint framework in the ScanObjectNNs dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.429, + 0.906, + 0.564 + ], + "angle": 0, + "content": "The ScanObjectNN dataset has 15 classes, which is actually more challenging due to the complexity of its contexts and object locations. The ScanObjectNN dataset is divided into three subsets: OBJ-ONLY, (only objects), OBJ-BG(objects and background), and PB-T50-RS(objects, background, and artificially added perturbations). We take the Overall Accuracy on these three subsets as the evaluation metric, and the detailed experimental results are summarized in Tab. 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.566, + 0.906, + 0.687 + ], + "angle": 0, + "content": "As can be seen from the table, our PreDifPoint framework achieves excellent classification results in the ScanObjectNN dataset, regardless of the presence or absence of the complex background of artificially added perturbations. It demonstrates the strong adaptability and robustness of our model in the face of complex and changing environments. Compared with other state-of-the-art models, the PreDif-Point framework also achieves the best performance." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.687, + 0.669, + 0.701 + ], + "angle": 0, + "content": "Object Segmentation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.703, + 0.906, + 0.779 + ], + "angle": 0, + "content": "We also validated the ability of our model for object segmentation in the ShapeNetPart dataset. ShapeNetPart contains a total of 16 large categories, which are labeled with semantic information for the semantic segmentation task of the point cloud." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We use mIoU to evaluate the object segmentation performance of the model, and from Tab. 1, we can see that overall, PreDifPoint has the best segmentation performance, which is \\(0.8\\%\\) better than MPCT. Among other things, our model also achieves the best performance in multiple categories. This indicates that our model has been pre-trained to learn the generalized feature representation, which can extract the key features in the input data more efficiently and" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26564" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.09, + 0.903, + 0.215 + ], + "angle": 0, + "content": "
MethodmIoUairplanebagcapcarchairearphoneguitarknifelamplaptopmotorbikemugpistolrocketskateboardtable
PointNet [25]83.783.478.782.574.989.673.091.585.980.895.365.293.081.257.972.880.6
PointNet++ [26]85.182.479.087.777.390.871.891.085.983.795.371.694.181.358.776.482.6
DGCNN [33]85.284.083.486.777.890.674.791.287.582.895.766.394.981.163.574.582.6
PointCNN [18]86.184.186.586.080.890.679.792.388.485.396.177.295.284.264.280.083.0
DRNet [50]86.484.385.088.379.591.279.391.889.085.295.772.294.282.060.676.884.2
PCT [7]86.485.082.489.081.291.971.591.388.186.395.864.695.883.662.277.683.7
MPCT [40]86.784.985.789.380.992.278.692.688.486.696.573.695.783.563.780.983.6
PreDiffPoint(ours)87.585.385.289.081.591.276.592.888.287.396.276.294.784.864.976.381.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.291, + 0.226, + 0.704, + 0.24 + ], + "angle": 0, + "content": "Table 1. Part Segmentation Results (%) on the ShapeNetPart Dataset." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.264, + 0.499, + 0.447 + ], + "angle": 0, + "content": "
MethodsPre.OBJ-ONLYOBJ-BGPB-T50-RS
PointNet [25]-79.273.368.0
PointNet++ [26]-84.382.377.9
PointCNN [18]-85.586.178.5
DGCNN [33]-86.282.878.1
Transformer [44]-80.5579.8677.24
Point-Bert [44]✓88.1287.4383.07
Point-MAE [23]✓88.2990.0285.18
TAP [35]✓89.5090.3685.67
PointDif [48]✓91.9193.2987.61
PreDifPoint(ours)✓92.2593.4287.96
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.457, + 0.483, + 0.486 + ], + "angle": 0, + "content": "Table 2. Object classification results on ScanObjectNN. We report the Overall Accuracy \\((\\%)\\) ." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.516, + 0.36, + 0.546 + ], + "angle": 0, + "content": "thus improve the segmentation accuracy. Indoor Semantic Segmentation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.547, + 0.483, + 0.652 + ], + "angle": 0, + "content": "We further validate the semantic segmentation capability of our model in the indoor dataset S3DIS. The S3DIS dataset contains point cloud data and detailed semantic labels for several indoor environments, which contains a total of 6 regions (Area1, Area2, Area3, Area4, Area5, Area6), which are further divided into 271 rooms or spaces. We tested our model in Area5 while training in other areas." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.655, + 0.483, + 0.849 + ], + "angle": 0, + "content": "We used mIoU and mAcc to evaluate the performance of the model, and from Tab. 3, we can see that our mAcc lags slightly behind compared to PointDif, probably due to the model's limited ability to extract features in certain categories, but still achieves a more competitive accuracy. In addition, PreDifPoint improves by \\(1.8\\%\\) in mIoU, which indicates that our model is able to understand the spatial relationships in the image well and segment different regions accurately. It is clear that our model obtains more meaningful geometric prior knowledge, which enables it to better understand the contextual semantics and local geometric relationships in complex scenes and achieve better performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.864, + 0.245, + 0.88 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.2, + 0.901 + ], + "angle": 0, + "content": "Masking ratio." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.264, + 0.908, + 0.417 + ], + "angle": 0, + "content": "
MethodsPre.mIoUmAcc
PointNet [25]-41.149.0
PointNet++ [26]-53.5-
PointCNN [18]-57.363.9
KPConv [40]-67.172.8
Point-Bert [44]✓68.976.1
Point-MAE [23]✓68.476.2
PointDif [48]✓70.077.1
PreDifPoint(ours)✓71.876.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.427, + 0.907, + 0.456 + ], + "angle": 0, + "content": "Table 3. Semantic Segmentation Results (%) Obtained on the S3DIS Dataset, Evaluated on Area 5." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.488, + 0.907, + 0.609 + ], + "angle": 0, + "content": "The masking operations that we include in the pre-training will affect the performance of the model in the downstream task. Thus, we first verified the effect of different mask ratios on the downstream tasks. As in Tab. 4, we examine the classification and segmentation ability of the model on the ScanObjectNN dataset and the S3DIS dataset, respectively. It is clear that the model achieves the best classification and segmentation performance when the mask ratio is 0.8." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.611, + 0.909, + 0.777 + ], + "angle": 0, + "content": "This may be due to the fact that the self-supervision task is too difficult when the masking ratio is too high, which may cause the model to have difficulty in learning effective feature representations, and will affect the model's ability to extract enough information from the limited training data to cope with the prediction task. However, when the masking ratio is too low, the model cannot be forced to learn deep feature representations and the model cannot obtain better prior knowledge. Only when the masking ratio is 0.8, it may just be beneficial for the model to learn effective feature representations." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.778, + 0.59, + 0.79 + ], + "angle": 0, + "content": "Attention." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.903 + ], + "angle": 0, + "content": "The DXCA-Attention mechanism that we use in CGNet enhances model comprehension by considering feature relationships in both directions simultaneously, enabling a more comprehensive capture of information. To validate the utility of this attention mechanism, we report results using MLP, one-way covariate attention mechanism and DXCA-Attention mechanism on the ScanObjectNN dataset, respec" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26565" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.493, + 0.233 + ], + "angle": 0, + "content": "
DatasetsScanObjectNNS3DIS(Area5)
Mask RatioOAmIoU
091.5869.1
0.590.9668.2
0.691.2369.5
0.791.8570.3
0.893.4271.8
0.992.3470.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.243, + 0.485, + 0.272 + ], + "angle": 0, + "content": "Table 4. The Overall Accuracy (%) on ScanObjectNN and the mIoU (%) on S3DIS with different masking ratios." + }, + { + "type": "table", + "bbox": [ + 0.139, + 0.29, + 0.438, + 0.368 + ], + "angle": 0, + "content": "
AttentionOA
MLP92.56
Covariance Attention93.19
DXCA-Attention93.42
" + }, + { + "type": "table_caption", + "bbox": [ + 0.13, + 0.378, + 0.444, + 0.394 + ], + "angle": 0, + "content": "Table 5. The Forms of The Self-Attention Operators." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.425, + 0.136, + 0.44 + ], + "angle": 0, + "content": "tively." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.442, + 0.483, + 0.653 + ], + "angle": 0, + "content": "As shown in Tab. 5, our proposed DXCA-Attention mechanism obtains the best performance. It may be because MLP is able to capture the complex relationship between features through nonlinear transformations, but it is unable to take into account the specific interaction patterns between features. The one-way covariate attention mechanism, on the other hand, only considers the covariance relationship between features in one direction, which may have led to the loss of information. Only the DXCA-Attention mechanism can model the interactions between features more comprehensively, capture more complex nonlinear relationships, enhance the robustness of the model, and facilitate feature fusion, thus enhancing the expressive power of the model." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.655, + 0.354, + 0.669 + ], + "angle": 0, + "content": "Density-Adaptive Sampling Strategy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.672, + 0.483, + 0.793 + ], + "angle": 0, + "content": "The density-adaptive sampling strategy that we use in the diffusion process enables the model to focus more on the denser regions in the point cloud, which makes it easier for the model to recover the structural details of the denser regions. We compare this with the uniform timestep sampling strategy [48], and as shown in Tab. 6, the OA achieved by the density-adaptive sampling strategy is 0.11 higher than the uniform time-step sampling strategy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.903 + ], + "angle": 0, + "content": "It indicates that uniform time-step sampling may lead to under-sampling in regions with higher density and oversampling in regions with lower density. Since many real-world applications of point cloud data tend to have nonuniform densities, the density-adaptive sampling strategy can utilize the data more efficiently, especially in those high-density regions that are critical to model performance." + }, + { + "type": "table", + "bbox": [ + 0.54, + 0.089, + 0.885, + 0.151 + ], + "angle": 0, + "content": "
Sampling StrategynOA
Recurrent uniform sampling93.31
Density-Adaptive Sampling93.42
" + }, + { + "type": "table_caption", + "bbox": [ + 0.571, + 0.161, + 0.848, + 0.176 + ], + "angle": 0, + "content": "Table 6. The Forms of The Sampling Strategy." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.203, + 0.634, + 0.218 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.228, + 0.907, + 0.499 + ], + "angle": 0, + "content": "In this paper, we propose a pre-training framework based on a diffusion model called PreDifPoint. Specifically, we propose a conditional generation network to guide the model through the point-to-point denoising process. In this context, the two-way covariate attention mechanism used facilitates the extraction and representation of point cloud features, thus enhancing the model's comprehension. We also introduce a density-adaptive sampling strategy based on the time step to guide the model to pay more attention to the denser regions in the point cloud, thus making it easier for the model to recover the structural details of the dense regions. Extensive experiments have shown that our method has good feature learning and representation capabilities, and is able to achieve advanced results in a variety of real-world datasets, which further validates the effectiveness of the proposed method. In the future, we hope that our work will further improve the performance of point cloud analysis." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.516, + 0.611, + 0.531 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.541, + 0.907, + 0.581 + ], + "angle": 0, + "content": "[1] Tomer Amit, Tal Shaharbany, Eliya Nachmani, and Lior Wolf. Segdiff: Image segmentation with diffusion probabilistic models, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.584, + 0.906, + 0.626 + ], + "angle": 0, + "content": "[2] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.628, + 0.906, + 0.67 + ], + "angle": 0, + "content": "[3] Adam Bretherton, Joshua J. Bon, David J. Warne, Kerrie Mengersen, and Christopher Drovandi. A principled approach to bayesian transfer learning, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.671, + 0.906, + 0.725 + ], + "angle": 0, + "content": "[4] Tolgahan Cakaloglu and Xiaowei Xu. Mrnn: A multi-resolution neural network with duplex attention for document retrieval in the context of question answering, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.728, + 0.906, + 0.757 + ], + "angle": 0, + "content": "[5] Shoufa Chen, Peize Sun, Yibing Song, and Ping Luo. Diffusional: Diffusion model for object detection, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.758, + 0.906, + 0.8 + ], + "angle": 0, + "content": "[6] Sachin Goyal, Ananya Kumar, Sankalp Garg, Zico Kolter, and Aditi Raghunathan. Finetune like you pretrain: Improved finetuning of zero-shot vision models, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.801, + 0.906, + 0.856 + ], + "angle": 0, + "content": "[7] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R. Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 7(2):187-199, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.859, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[8] Jing Hao, Song Chen, Xiaodi Wang, and Shumin Han. Language-aware multiple datasets detection pretraining for detrs, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.522, + 0.541, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26566" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[9] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.163 + ], + "angle": 0, + "content": "[10] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.483, + 0.275 + ], + "angle": 0, + "content": "[11] Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, Tom Hennigan, Eric Noland, Katie Millican, George van den Driessche, Bogdan Damoc, Aurelia Guy, Simon Osindero, Karen Simonyan, Erich Elsen, Jack W. Rae, Oriol Vinyals, and Laurent Sifre. Training compute-optimal large language models, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[12] Shouwei Hu, Xi Li, Banyao Ruan, and Zhihao Liu. An amplitude-encoding-based classical-quantum transfer learning framework: Outperforming classical methods in image recognition, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[13] Mohammad Rafid Ul Islam, Prasad Tadepalli, and Alan Fern. Self-attention-based diffusion model for time-series imputation in partial blackout scenarios, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[14] Liqiang Jing, Yiren Li, Junhao Xu, Yongcan Yu, Pei Shen, and Xuemeng Song. Vision enhanced generative pre-trained language model for multimodal sentence summarization. Machine Intelligence Research, 20:289-298, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.433, + 0.482, + 0.461 + ], + "angle": 0, + "content": "[15] Li Ju, Xingyi Yang, Qi Li, and Xinchao Wang. Graphbridge: Towards arbitrary transfer learning in gnns, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[16] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.546 + ], + "angle": 0, + "content": "[17] Xudong Li, Li Feng, Lei Li, and Chen Wang. Few-shot meta-learning on point cloud for semantic segmentation, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.589 + ], + "angle": 0, + "content": "[18] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on \\(\\mathcal{X}\\)-transformed points, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.591, + 0.482, + 0.645 + ], + "angle": 0, + "content": "[19] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[20] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3d mesh in 45 seconds without pershape optimization, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[21] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2836-2844, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[22] Zhaoyang Lyu, Zhifeng Kong, Xudong Xu, Liang Pan, and Dahua Lin. A conditional point diffusion-refinement paradigm for 3d point cloud completion, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[23] Yatian Pang, Wenxiao Wang, Francis E. H. Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning, 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[24] Konpat Preechakul, Nattanat Chathee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion autoencoders: Toward a meaningful and decodable representation, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[25] Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[26] Charles R. Qi, Li Yi, Hao Su, and Leonidas J. Guibas. Point-net++: Deep hierarchical feature learning on point sets in a metric space, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[28] Mojtaba Safari, Shansong Wang, Zach Eidex, Qiang Li, Erik H. Middlebrooks, David S. Yu, and Xiaofeng Yang. Mri super-resolution reconstruction using efficient diffusion probabilistic model with residual shifting, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[29] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[30] Aaron van den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.905, + 0.418 + ], + "angle": 0, + "content": "[31] Aaron van den Oord, Nal Kalchbrenner, Oriol Vinyals, Lasse Espeholt, Alex Graves, and Koray Kavukcuoglu. Conditional image generation with pixelCNN decoders, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[32] Jiahui Wang, Haiyue Zhu, Haoren Guo, Abdullah Al Mamun, Cheng Xiang, and Tong Heng Lee. Few-shot point cloud semantic segmentation via contrastive self-supervision and multi-resolution attention, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.518 + ], + "angle": 0, + "content": "[33] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E. Sarma, Michael M. Bronstein, and Justin M. Solomon. Dynamic graph cnn for learning on point clouds, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[34] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.575, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[35] Ziyi Wang, Xumin Yu, Yongming Rao, Jie Zhou, and Jiwen Lu. Take-a-photo: 3d-to-2d generative pre-training of point cloud models, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.618, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[36] Zhendong Wang, Jianmin Bao, Shuyang Gu, Dong Chen, Wengang Zhou, and Houqiang Li. Design diffusion: High-quality text-to-design image generation with diffusion models, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[37] Antoine Wehenkel and Gilles Louppe. Diffusion priors in variational autoencoders, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.905, + 0.746 + ], + "angle": 0, + "content": "[38] Julia Wolleb, Robin Sandkuhler, Florentin Bieder, Philippe Valmaggia, and Philippe C. Cattin. Diffusion models for implicit image segmentation ensembles, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.747, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[39] Jay Zhangjie Wu, Yuxuan Zhang, Haithem Turki, Xuanchi Ren, Jun Gao, Mike Zheng Shou, Sanja Fidler, Zan Gojcic, and Huan Ling. Difix3d+: Improving 3d reconstructions with single-step diffusion models, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[40] Yue Wu, Jiaming Liu, Maoguo Gong, Zhixiao Liu, Qiguang Miao, and Wenping Ma. Mpct: Multiscale point cloud transformer with a residual network. IEEE Transactions on Multimedia, 26:3505-3516, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[41] Weilai Xiang, Hongyu Yang, Di Huang, and Yunhong Wang. Denoising diffusion autoencoders are unified self-supervised learners, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26567" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.482, + 0.135 + ], + "angle": 0, + "content": "[42] Saining Xie, Jiatao Gu, Demi Guo, Charles R. Qi, Leonidas J. Guibas, and Or Litany. Pointcontrast: Unsupervised pre-training for 3d point cloud understanding, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.136, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[43] Jiale Xu, Xintao Wang, Weihao Cheng, Yan-Pei Cao, Ying Shan, Xiaohu Qie, and Shenghua Gao. Dream3d: Zero-shot text-to-3d synthesis using 3d shape prior and text-to-image diffusion models, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.192, + 0.482, + 0.234 + ], + "angle": 0, + "content": "[44] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling, 2022. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.236, + 0.483, + 0.29 + ], + "angle": 0, + "content": "[45] Renrui Zhang, Ziyu Guo, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, Hongsheng Li, and Peng Gao. Pointm2ae: Multi-scale masked autoencoders for hierarchical point cloud pre-training, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.292, + 0.482, + 0.333 + ], + "angle": 0, + "content": "[46] Zaiwei Zhang, Rohit Girdhar, Armand Joulin, and Ishan Misra. Self-supervised pretraining of 3d features on any point-cloud, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.335, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[47] Luda Zhao, Yihua Hu, Xing Yang, Zhenglei Dou, and Qilong Wu. Icddpm: Image-conditioned denoising diffusion probabilistic model for real-world complex point cloud single view reconstruction. Expert Systems with Applications, 259:125370, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.405, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[48] Xiao Zheng, Xiaoshui Huang, Guofeng Mei, Yuenan Hou, Zhaoyang Lyu, Bo Dai, Wanli Ouyang, and Yongshun Gong. Point cloud pre-training with diffusion models, 2023. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.462, + 0.482, + 0.49 + ], + "angle": 0, + "content": "[49] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.491, + 0.482, + 0.532 + ], + "angle": 0, + "content": "[50] Mingjian Zhu, Kai Han, Enhua Wu, Qiulin Zhang, Ying Nie, Zhenzhong Lan, and Yunhe Wang. Dynamic resolution network, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.533, + 0.482, + 0.575 + ], + "angle": 0, + "content": "[51] Aleksei Zhuravlev, Zorah Lähner, and Vladislav Golyanik. Denoising functional maps: Diffusion models for shape correspondence, 2025. 3" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.483, + 0.575 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "26568" + } + ] +] \ No newline at end of file diff --git a/2025/Feature Extraction and Representation of Pre-training Point Cloud Based on Diffusion Models/0c060285-4bbf-430f-b421-79f016d6dd03_origin.pdf b/2025/Feature Extraction and Representation of Pre-training Point Cloud Based on Diffusion Models/0c060285-4bbf-430f-b421-79f016d6dd03_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..771c7c5a731f8d62e44afec3c3f5f0f25ccba5e7 --- /dev/null +++ b/2025/Feature Extraction and Representation of Pre-training Point Cloud Based on Diffusion Models/0c060285-4bbf-430f-b421-79f016d6dd03_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd1726d78e5d9b23182aa1582201a305e399ecc892e90e5f82d77e2f1d260366 +size 1497453 diff --git a/2025/Feature Extraction and Representation of Pre-training Point Cloud Based on Diffusion Models/full.md b/2025/Feature Extraction and Representation of Pre-training Point Cloud Based on Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e75c9a45b3f504eb1b077e83a729bd67dd1727dc --- /dev/null +++ b/2025/Feature Extraction and Representation of Pre-training Point Cloud Based on Diffusion Models/full.md @@ -0,0 +1,370 @@ +# Feature Extraction and Representation of Pre-training Point Cloud Based on Diffusion Models + +Chang Qiu + +Southeast University + +230238519@seu.edu.cn + +Feipeng Da* + +Southeast University + +dafp@seu.edu.cn + +Zilei Zhang + +Southeast University + +230238524@seu.edu.cn + +# Abstract + +The pretrain-finetune paradigm of pre-training a model on large amounts of image and text data and then fine-tuning the model for a specific task has led to significant progress in many 2D image and natural language processing tasks. Similarly, the use of pre-training methods in point cloud data can also enhance the working performance and generalization ability of the model. Therefore, in this paper, we propose a pre-training framework based on a diffusion model called PreDifPoint. It is able to accomplish the pretraining of the model's backbone network through a diffusion process of gradual denoising. We aggregate the potential features extracted from the backbone network, input them as conditions into the subsequent diffusion model, and direct the point-to-point mapping relationship of the noisy point clouds at neighboring time steps, so as to generate high-quality point clouds and at the same time better perform various downstream tasks of the point clouds. We also introduce a bi-directional covariate attention (DXCA-Attention) mechanism for capturing complex feature interactions, fusing local and global features, and improving the detail recovery of point clouds. In addition, we propose a density-adaptive sampling strategy, which can help the model dynamically adjust the sampling strategy between different time steps, and guide the model to pay more attention to the denser regions in the point cloud, thus improving the effectiveness of the model in point cloud recovery. Our PreDifPoint framework achieves more competitive results on various real-world datasets. Specifically, PreDifPoint achieves an overall accuracy of $87.96\%$ , which is $0.35\%$ higher than PointDif, on the classification task on PB-T50-395RS, a variant of ScanObjectNN dataset. + +# 1. Introduction + +Recent work has demonstrated that the pretrain-finetune paradigm, where a model is first pretrained on large + +amounts of image and text data and then fine-tuned for specific tasks, has led to substantial advancements in many 2D image and natural language processing tasks [11, 16, 27]. Pretraining the backbone on large-scale datasets enables the model to learn rich feature representations while also avoiding the need for large amounts of data and computational resources required for training a model from scratch [8, 9, 32]. Pretrained models, after fine-tuning, can learn task-specific knowledge, exhibit better generalization capabilities, and converge to optimal performance levels more rapidly [6, 14]. Similarly, using pretraining methods for point cloud data can also enhance the model's performance and generalization ability [44, 46, 48]. + +Generative pretraining methods, through generative models, capture the complex structure and intricate details of point cloud data, thereby obtaining more powerful feature representations. Early works such as PointBERT [44], Point-MAE [23], and Point-M2AE [45] employed self-supervised learning by reconstructing masked local point patches, forcing the model to learn local point cloud information. However, this masking and reconstruction approach may overlook the contextual relationships between masked and unmasked regions, preventing the model from effectively understanding the global structure of the object or scene. Inspired by diffusion models [10], which can understand global structure and long-term dependencies through the process of adding and removing noise, we propose using a diffusion model to guide the pretraining process of point clouds. + +Thus, we introduce a diffusion-based pretraining framework, named PreDifPoint. This framework pretrains only the backbone network of the model, and then fine-tunes the task-specific parts. The pretraining of the backbone network is achieved through a progressive denoising diffusion process, as shown in Fig. 1. By pretraining the backbone, it can learn meaningful features during the feature extraction phase, which also significantly enhances the model's generalization ability. At the same time, this reduces the number of parameters that need to be adjusted during training, accelerating convergence and reducing overfitting. + +![](images/c66474988fa9fbe369b0b9455a8976898db1d7f067ccb1487469a4f66e888563.jpg) +Figure 1. Schematic illustration of our PreDiffPoint. Our PreDiffPoint can be used to train different backbones and reconstruct the original point cloud from the noisy one through a diffusion process. + +Specifically, the PreDifPoint framework consists of three core modules: the Conditional Point Generation Network (CGNet), the Conditional Denoising Network (CDNet), and the Conditional Diffusion Network (CDPNet). CGNet aggregates the latent features extracted by the backbone and feeds them as conditional inputs into the subsequent diffusion model. CDNet utilizes these conditional inputs to perform denoising and restore point cloud details. CDPNet, through the point-to-point mapping of noisy point clouds between adjacent time steps in the diffusion process [48], ensures that the generated point clouds excel in both global consistency and local detail. With this design, the model can generate high-quality point clouds based on various conditions, exhibiting strong diversity and adaptability to perform various downstream tasks. + +Additionally, we propose a density-adaptive sampling strategy. We divide the entire time step range into several subintervals, each containing multiple time steps. Within each subinterval, the sampling probability is dynamically adjusted according to the local density of the point cloud at each time step, prioritizing sampling from steps with higher density. This strategy allows for more fine-grained control over the sampling process at each time step and helps dynamically adjust the sampling strategy across time steps, improving the model's point cloud recovery performance. + +Our main contributions can be summarized as follows: + +- We propose a diffusion-based point cloud pretraining framework, PreDifPoint, which leverages conditional guidance for generation, denoising, and recovery, as well as global and local feature modeling. This framework enhances the performance of point cloud generation and denoising tasks and provides high-quality feature representations for downstream tasks. +- We introduce a DXCA-Attention mechanism into the model to capture complex global feature interactions and generate more instructive condition vectors to enhance the detail reconstruction of the point cloud. The condition vectors obtained through CGNet will be combined with time-step information and passed to the subsequent + +network for denoising. + +- We propose a density-adaptive sampling strategy, which can help the model dynamically adjust the sampling strategy between different time steps, and guide the model to pay more attention to the denser regions in the point cloud, thus improving the effectiveness of the model in point cloud recovery. +- The PreDifPoint framework requires only fine-tuning for task-specific components (such as classifiers), allowing the model to perform optimally across various tasks, particularly in data-limited or computationally constrained environments. Notably, the PreDifPoint framework demonstrates competitive performance in a variety of practical downstream tasks. + +# 2. Related Works + +# 2.1. Point Cloud Pretraining + +In deep learning, pretraining has become a key technique for enhancing model performance, especially when data is scarce or labels are limited. For point cloud learning tasks, pretraining is particularly important. Point cloud data typically has complex spatial structures and sparsity, and traditional training methods often rely on large amounts of labeled data, which is difficult to obtain in real-world applications. Therefore, learning effective point cloud representations through pretraining not only reduces reliance on labeled data but also improves the model's generalization and robustness in downstream tasks. + +Currently, point cloud pretraining methods can be broadly categorized into three types: self-supervised learning, generative models [29, 34, 43], and transfer learning. Each approach has its unique advantages and application scenarios. + +Self-supervised methods [4, 30, 31] can mask parts of the data and train the model to reconstruct the missing parts or optimize the model's representation ability through strategies like contrastive learning. Point-MAE [23] randomly masks certain points in the point cloud and trains the model to recover these missing points, thereby effectively learning the geometric features of the point cloud. Point-M2AE [45] for hierarchical self-supervised learning of 3D point clouds modifies the encoder and decoder into a pyramid architecture to incrementally model the spatial geometry and capture the fine-grained and high-level semantics of 3D shapes. PointContrast [42] employs contrastive learning to treat point cloud representation as a learning task, improving the model's ability to distinguish between similar and different point cloud segments. + +Generative models [19, 20, 24, 37, 41], particularly denoising diffusion models and variational autoencoders, have made significant progress in point cloud pretraining in recent years. These models learn the generative distribution of + +point cloud data, capturing the details of the point cloud and completing missing data. DDPM [22] demonstrates how to use pretrained denoising diffusion models for conditional point cloud generation and completion, significantly improving the quality of point cloud reconstruction. PVD [49] introduces a conditional variational diffusion model (CV-Diffusion) that can generate high-quality 3D shapes through conditional information, effectively applied to point cloud generation and completion tasks. + +Transfer learning methods [3, 12, 15] involve pretraining a model on a large-scale dataset and then transferring it to a specific task, significantly improving performance when data is limited. Meta-PointNet [17] proposes a meta-learning-based pretraining method for point cloud classification and segmentation, where knowledge is shared across multiple tasks, greatly improving performance on few-shot tasks. + +# 2.2. Diffusion Models + +The core idea of diffusion models [13, 28, 36, 39, 51] is to simulate a process of gradually adding noise and then recovering the data through reverse denoising. Diffusion models have been widely used in recognition tasks such as object detection [5] and semantic discrimination [1, 2, 38]. During training, the model learns how to recover real data from noise, which helps generate high-quality samples. This paper uses conditional diffusion models, which introduce additional conditional information into the diffusion process to guide the generated model. These conditional inputs can include labels, textual descriptions, or other external features that help the model generate data that meets specific conditions. Conditional diffusion models have shown excellent performance in multimodal generation tasks, particularly in image generation, point cloud generation, and speech synthesis. + +Diffusion probabilistic models [21] proposed a point cloud probabilistic generation model inspired by nonequilibrium thermodynamics that utilizes a backward diffusion process to learn point distributions. And they achieved competitive performance in point cloud generation and automatic coding, and comparable results in unsupervised representation learning. PointDif [48] proposes a conditional diffusion model for point cloud pretraining tasks. This model can more precisely control the structure and details of the point cloud during the generation or denoising process, performing particularly well in point cloud denoising and generation tasks. ICDDPM [47] can uses 2D images as prior information to guide end-to-end 3D point cloud generation, and obtain high-quality 3D point clouds of different types of targets from real-world images. + +The PreDifPoint framework we have developed pretrains the network by recovering the original point cloud from randomly noisy point clouds, forcing the network to learn local + +and global geometric priors of point clouds. The extracted point cloud features are passed as conditional information into the diffusion process, guiding the generation of point clouds step by step. Additionally, we propose a density-adaptive sampling strategy, which allows for finer control over the sampling process at each time step and helps dynamically adjust the sampling strategy between time steps, thus improving the model's ability to recover point clouds. + +# 3. Method + +In our pre-training process, the backbone of the overall framework PreDifPoint is a transformer encoder. Notably, this backbone can be replaced with any other backbone network for pre-training. The input point cloud is grouped and features are extracted, followed by masking operations and contextual enhancement via a Transformer with positional encoding. Subsequently, CGNet further processes these features to generate conditional vectors that serve as conditional inputs to CDNet. In CDNet, time-step embedding is combined with conditional vectors. Finally, the CDPNet network integrates CGNet and CDNet to generate the original point cloud by stepwise denoising the data through a diffusion model. The pipeline of our PreDifPoint is shown in Fig. 2. + +# 3.1. Diffusion Process + +In the diffusion model, a clean point cloud $X^0 \epsilon R^{n \times 3}$ is gradually corrupted into a fully noisy point cloud $X^T$ by incrementally adding noise. The diffusion process at each time step $t$ follows: + +$$ +q \left(X ^ {t} \mid X ^ {t - 1}\right) = \mathcal {N} \left(X ^ {t}; \sqrt {\alpha_ {t}} X ^ {t - 1}, (1 - \alpha_ {t}) \mathrm {I}\right), \tag {1} +$$ + +where $\alpha_{t} = 1 - \beta_{t}$ denotes the noise attenuation coefficient at time step, and $\beta_{t}$ represents the noise level. Through this process, noise is gradually added to the point cloud until it becomes nearly random at $X^T$ . + +This reverse process predicts the denoising signals at each time step by the neural network $p_{\theta}(X^t | X^{t-1}, c)$ , which gradually recovers the original structure of the point cloud as shown in the following equation: + +$$ +p _ {\theta} \left(X ^ {t} \mid X ^ {t - 1}, c\right) = \mathcal {N} \left(X ^ {t - 1}; \mu_ {\theta} \left(X ^ {t}, t, c\right), \sigma_ {t} ^ {2} \mathrm {I}\right), \tag {2} +$$ + +where $\mu_{\theta}(X^{t},t,c)$ is the mean predicted by the network, $X^t$ representing the denoised result at step $t$ , and $\sigma_t^2$ is the noise variance. + +# 3.2. Conditional Generation Network (CGNet) + +CGNet is responsible for extracting global and local features from the point cloud, fusing these features through a DXCA-Attention mechanism to more comprehensively + +![](images/51f7b387b6e06eb17d4312d18e311f641af80d8d8a5343d7d3a8b8c97bfd5f32.jpg) +Figure 2. Pipeline of PreDifPoint. We first group the input point cloud and extract features, which are then subjected to masking operations and contextual enhancement via a transformer with position encoding. Subsequently, CGNet further processes these features to generate conditional vectors, which are used as conditional inputs to CDNet. In CDNet, time-step embedding is combined with conditional vectors. Finally, the CDPNet network integrates CGNet and CDNet to generate the original point cloud by stepwise denoising the data through a diffusion model. + +model the interactions between the features and better adapt to different data patterns and structures, as shown in Fig. 3. Finally, a conditional vector $c$ associated with the density distribution is generated, which is used to guide the subsequent denoising process. + +We first project the input feature $F_{loc}$ from the point cloud chunking encoding to the high dimensional space through the multilayer perceptron (MLP) to obtain the projected representation $F_{pro}$ of the local features, and then obtain the global feature $F_{glo}$ through the global pooling operation on $F_{pro}$ as shown in the following equation: + +$$ +\left\{ \begin{array}{l} F _ {p r o} = M L P (F), \\ F _ {g l o} = G l o b a l M a x P o o l \left(F _ {p r o}\right). \end{array} \right. \tag {3} +$$ + +Then, the attention weights are obtained by calculating the covariance matrix in both directions to capture the global contextual relationships and complex feature interactions. In this process, we need to first obtain the important parameters, $Q$ , $K$ , and $V$ , as shown in Eq. (4). + +$$ +\left\{ \begin{array}{l} Q = \operatorname {r e s h a p e} \left(F _ {p r o}\right) \in \mathbb {R} ^ {B \times G \times C}, \\ K = \operatorname {r e s h a p e} \left(F _ {g l o}\right) \in \mathbb {R} ^ {B \times G \times C}, \\ V _ {1} = \operatorname {r e s h a p e} \left(F _ {g l o}\right) \in \mathbb {R} ^ {B \times G \times C}, \\ V _ {2} = \operatorname {r e s h a p e} \left(F _ {p r o}\right) \in \mathbb {R} ^ {B \times G \times C}. \end{array} \right. \tag {4} +$$ + +Then centering along the feature dimension $C$ as shown in Eq. (5). + +$$ +\left\{ \begin{array}{l} \bar {Q} = Q - \mathbb {E} _ {c} [ Q ], \\ \bar {K} = K - \mathbb {E} _ {c} [ K ]. \end{array} \right. \tag {5} +$$ + +Then the scaling factor $\sqrt{C}$ is introduced to compute the covariance matrix, as shown in Eq. (6). + +$$ +\left\{ \begin{array}{l} A t t e n S c o r e _ {1} = S o f t m a x \left(\frac {\bar {Q} \cdot \bar {K} ^ {T}}{\sqrt {C}}\right) \in \mathbb {R} ^ {B \times G \times C}, \\ A t t e n S c o r e _ {2} = S o f t m a x \left(\frac {\bar {K} \cdot \bar {Q} ^ {T}}{\sqrt {C}}\right) \in \mathbb {R} ^ {B \times G \times C}. \end{array} \right. \tag {6} +$$ + +At this point, the bi-directional covariance features can be aggregated as shown in Eq. (7). + +$$ +\left\{ \begin{array}{l} O _ {1} = A t t e n S c o r e _ {1} \cdot V _ {1} \in \mathbb {R} ^ {B \times G \times C}, \\ O _ {2} = A t t e n S c o r e _ {2} \cdot V _ {2} \in \mathbb {R} ^ {B \times G \times C}, \\ O = \operatorname {c o n c a t} \left(O _ {1}, O _ {2}\right) \in \mathbb {R} ^ {B \times G \times 2 C}. \end{array} \right. \tag {7} +$$ + +Finally, we fuse the projected features, global features and the attention-weighted features to output the conditioned vector $c$ , as shown in Eq. (8). + +$$ +c = M L P \left(\text {C o n c a t} \left(F _ {\text {p r o}}, A, F _ {\text {g l o}}\right)\right). \tag {8} +$$ + +# 3.3. Conditional Denoising Network (CDNet) + +CDNet is responsible for gradually removing the noise and restoring the point cloud structure during the diffusion process, and its core is to achieve layer-level feature optimization through residual connectivity. + +First, the conditional vector $c$ and the time embedding $t_{emd}$ are fused into the point cloud features through a dynamic gating mechanism to achieve density-aware local de + +![](images/fcd4c81dc0cd386c1eb1b288656edc7c907211fb8eff1db34e11cc40aef08ceb.jpg) +Figure 3. DXCA-Attention Module. This attention mechanism achieves cross-level information fusion through the two-way interaction between global and local features. In the local-to-global direction, query vectors and key vectors are generated by centering local features and global features respectively, and the contribution weights of local regions to global semantics are computed through the covariance matrix, and the semantic consistency of local features is enhanced after weighted aggregation. In the global-to-local direction, the global features are used as queries and local features as keys, and the modulation relation of global context to local details is captured by covariance attention. + +noising. The time embedding $t_{emd}$ and the fused features can be represented as: + +$$ +\left\{ \begin{array}{l} t _ {e m d} = \left[ \beta_ {t}, \sin \left(\beta_ {t}\right), \cos \left(\beta_ {t}\right) \right], \\ H _ {l} = \sigma \left(W _ {g} [ c \oplus t _ {e m d} ]\right) \odot \left(W _ {f} H _ {l - 1}\right) + W _ {b} [ c \oplus t _ {e m d} ], \end{array} \right. \tag {9} +$$ + +where $\sigma$ is the Sigmoid function for generating gating weights in the range [0, 1], and $W_{g}, W_{f}, W_{b}$ are the gating generation weights, feature transformation weights and bias generation weights, respectively. + +Finally, to alleviate the gradient vanishing problem, the output of each of our layers can be connected using residuals such that the output of each layer is represented as: + +$$ +H _ {l} = H _ {l - 1} + C D N e t \left(H _ {l - 1}, c, t _ {e m d}\right). \tag {10} +$$ + +# 3.4. Conditional Diffusion Point Network (CDPNet) + +The CDPNet integrates a density-weighted diffusion process, a conditional point generation network (CGNet), and a conditional denoising network (CDNet) to achieve stepwise generation from noisy to clean point clouds. + +In a nutshell, it is to first get a completely noisy point cloud $X^T$ from a clean point cloud $X^0$ in the forward diffusion process, then obtain the conditional vector $\mathbf{c}$ by CGNet, and then update $X^T$ step by step from $t = T$ to $t = 1$ in CDNet, and finally get the reconstructed point cloud $X^0$ . + +# 3.5. Density-Adaptive Sampling + +To dynamically adjust the time step scheduling in the diffusion model, we propose a density-adaptive sampling strategy. This strategy weights the sampling probability of each + +time step based on the local density of points, prioritizing denser regions (e.g., edges) during training to enhance detail reconstruction. + +First, the local density $\rho_{i}$ of each point $x_{i}\epsilon X_{0}$ is computed using $k$ -nearest neighbors: + +$$ +\rho \left(x _ {i}\right) = \frac {1}{\sum_ {j = 1} ^ {k} \left\| x _ {i} - x _ {j} \right\| + \epsilon}, \tag {11} +$$ + +where $k$ is the neighborhood size, $\| x_{i} - x_{j}\|$ is the Euclidean distance, and $\epsilon$ is a small constant to avoid division by zero. + +The time step interval $[1, T]$ is uniformly divided into $h$ sub-intervals $\{Q_1, Q_2, \dots, Q_h\}$ , each corresponding to different noise levels. The sampling probability for $Q_i$ is weighted by the local density within the interval: + +$$ +P \left(Q _ {i} \mid \rho\right) = \frac {\rho_ {i}}{\sum_ {i = 1} ^ {T} \rho_ {i}}. \tag {12} +$$ + +This weighting strategy ensures that more samples can be performed in the denser regions, thus improving the reconstruction accuracy in these regions. + +Within each interval $Q_{i}$ , we randomly sample a time step $t_{i}$ from that interval and compute the diffusion process at that time step. Since the sampling is based on local density weighting, the sampling probability for each interval reflects the denseness of the region. In updating the sampling, we use the following weighted loss function to calculate the loss at each time step: + +$$ +L (\theta , \rho) = \frac {1}{h} \sum_ {i = 0} ^ {h - 1} L \left(\theta , \rho , t _ {i} \sim Q _ {i}\right), \tag {13} +$$ + +where $L(\theta, \rho, t_i \sim Q_i)$ is the loss for the sampled $t_i$ . + +# 3.6. Density-Weighted Diffusion Loss + +In the training of diffusion models combined with density-adaptive sampling, the loss function consists mainly of the error in the prediction noise and the error in the density weighting. Our goal is to minimize the difference between the predicted noise and the real noise at each time step. + +The standard diffusion model loss is achieved by calculating the mean square error (MSE) between the predicted noise and the real noise, as shown in Eq. (14): + +$$ +L _ {d e n o i s e} (\theta) = \mathbb {E} _ {x _ {0}, t, \epsilon} [ \| \epsilon - \epsilon_ {\theta} (\sqrt {\alpha_ {t}} x _ {0} + \sqrt {1 - \alpha_ {t}} \epsilon , t, c) \| ^ {2} ], \tag {14} +$$ + +where $x_0$ is the input original point cloud, $t$ is the current time step, $\epsilon_{\theta}$ is the noise predicted by the network, $\epsilon$ is the noise actually added, $\alpha_{t}$ is the noise attenuation coefficient at each time step, $c$ is the conditional information generated by the conditional aggregation network CGNet, and $\mathbb{E}_{x_0,t,\epsilon}$ + +is the indication of the need for averaging all the possible values of these variables to calculate the final loss. + +During the density-adaptive sampling process, we need to adjust the computation of the loss to be consistent with the sampling probability of each interval. The final training loss combining density-weighted sampling and the standard MSE loss can be expressed as: + +$$ +\begin{array}{l} L (\theta , \rho) = \frac {1}{h} \sum_ {i = 0} ^ {h - 1} \mathbb {E} _ {t _ {i} \sim Q _ {i}} [ \\ \begin{array}{l} \mathbb {E} _ {x _ {0}, \epsilon} \left[ \left\| \epsilon - \epsilon_ {\theta} \left(\sqrt {\alpha_ {t _ {i}}} x _ {0} + \sqrt {1 - \alpha_ {t _ {i}}} \epsilon , t _ {i}, c\right) \right\| ^ {2} \right] \\ ], \end{array} \tag {15} \\ \end{array} +$$ + +where $t_i$ is the time step sampled from the interval $Q_i$ , and the loss function enhances the reconstruction of these regions by weighting the samples at each time step to ensure that denser regions can have higher weights. + +# 4. Experiments + +# 4.1. Pre-training Setup + +We use ShapeNet to pre-train our model, with the core version of this dataset containing 51,300 3D models across 55 common object categories. For each 3D shape, we sample 1,024 points to serve as the input for the model. Following the approach in PointDif [48], we use 41,952 shapes as the training set. Additionally, we employ the KNN algorithm to select the $k = 32$ nearest points as point patches, resulting in 64 point cloud patches. Furthermore, we set the embedding dimension of the Transformer encoder to 384, with 6 attention heads, and the condition dimension to 768. + +During pre-training, we set the weight decay to 0.05 and the learning rate to 0.001. Our model is pre-trained for 300 epochs with a batch size of 192. + +Additionally, we visualize the point clouds generated by the PreDifPoint framework to demonstrate the effectiveness of our pre-training. As shown in Fig. 4, we use the masked point cloud features, extracted and used as conditions, to guide the diffusion model in generating the original point cloud, with a masking rate of 0.8. The visualization results indicate that the PreDifPoint framework generates high-quality point clouds and that the learned geometric priors effectively guide subsequent tasks such as recognition and detection. + +# 4.2. Downstream Tasks + +The performance of a pre-trained model, obtained in a fine-tuning task, is a test of the effectiveness of the pre-trained model. Since a high-quality point cloud pre-trained model should have a high hierarchical geometric prior, we tested + +![](images/43369277d394f2ca882e82548da3f31cfab1edb5028405624bfc12f6ca4b5184.jpg) +Figure 4. Visualization results on the ShapeNet validation set. We show a total of six categories of point clouds, each showing the input point cloud, the masked point cloud, and the reconstructed point cloud from left to right. + +the performance of the pre-trained model on a large number of real-world datasets. + +# Object Classification. + +We first evaluated the classification ability of the PreD-ifPoint framework in the ScanObjectNNs dataset. + +The ScanObjectNN dataset has 15 classes, which is actually more challenging due to the complexity of its contexts and object locations. The ScanObjectNN dataset is divided into three subsets: OBJ-ONLY, (only objects), OBJ-BG(objects and background), and PB-T50-RS(objects, background, and artificially added perturbations). We take the Overall Accuracy on these three subsets as the evaluation metric, and the detailed experimental results are summarized in Tab. 2. + +As can be seen from the table, our PreDifPoint framework achieves excellent classification results in the ScanObjectNN dataset, regardless of the presence or absence of the complex background of artificially added perturbations. It demonstrates the strong adaptability and robustness of our model in the face of complex and changing environments. Compared with other state-of-the-art models, the PreDif-Point framework also achieves the best performance. + +# Object Segmentation. + +We also validated the ability of our model for object segmentation in the ShapeNetPart dataset. ShapeNetPart contains a total of 16 large categories, which are labeled with semantic information for the semantic segmentation task of the point cloud. + +We use mIoU to evaluate the object segmentation performance of the model, and from Tab. 1, we can see that overall, PreDifPoint has the best segmentation performance, which is $0.8\%$ better than MPCT. Among other things, our model also achieves the best performance in multiple categories. This indicates that our model has been pre-trained to learn the generalized feature representation, which can extract the key features in the input data more efficiently and + +
MethodmIoUairplanebagcapcarchairearphoneguitarknifelamplaptopmotorbikemugpistolrocketskateboardtable
PointNet [25]83.783.478.782.574.989.673.091.585.980.895.365.293.081.257.972.880.6
PointNet++ [26]85.182.479.087.777.390.871.891.085.983.795.371.694.181.358.776.482.6
DGCNN [33]85.284.083.486.777.890.674.791.287.582.895.766.394.981.163.574.582.6
PointCNN [18]86.184.186.586.080.890.679.792.388.485.396.177.295.284.264.280.083.0
DRNet [50]86.484.385.088.379.591.279.391.889.085.295.772.294.282.060.676.884.2
PCT [7]86.485.082.489.081.291.971.591.388.186.395.864.695.883.662.277.683.7
MPCT [40]86.784.985.789.380.992.278.692.688.486.696.573.695.783.563.780.983.6
PreDiffPoint(ours)87.585.385.289.081.591.276.592.888.287.396.276.294.784.864.976.381.5
+ +Table 1. Part Segmentation Results (%) on the ShapeNetPart Dataset. + +
MethodsPre.OBJ-ONLYOBJ-BGPB-T50-RS
PointNet [25]-79.273.368.0
PointNet++ [26]-84.382.377.9
PointCNN [18]-85.586.178.5
DGCNN [33]-86.282.878.1
Transformer [44]-80.5579.8677.24
Point-Bert [44]✓88.1287.4383.07
Point-MAE [23]✓88.2990.0285.18
TAP [35]✓89.5090.3685.67
PointDif [48]✓91.9193.2987.61
PreDifPoint(ours)✓92.2593.4287.96
+ +thus improve the segmentation accuracy. Indoor Semantic Segmentation. + +We further validate the semantic segmentation capability of our model in the indoor dataset S3DIS. The S3DIS dataset contains point cloud data and detailed semantic labels for several indoor environments, which contains a total of 6 regions (Area1, Area2, Area3, Area4, Area5, Area6), which are further divided into 271 rooms or spaces. We tested our model in Area5 while training in other areas. + +We used mIoU and mAcc to evaluate the performance of the model, and from Tab. 3, we can see that our mAcc lags slightly behind compared to PointDif, probably due to the model's limited ability to extract features in certain categories, but still achieves a more competitive accuracy. In addition, PreDifPoint improves by $1.8\%$ in mIoU, which indicates that our model is able to understand the spatial relationships in the image well and segment different regions accurately. It is clear that our model obtains more meaningful geometric prior knowledge, which enables it to better understand the contextual semantics and local geometric relationships in complex scenes and achieve better performance. + +# 4.3. Ablation Study + +Masking ratio. + +Table 2. Object classification results on ScanObjectNN. We report the Overall Accuracy $(\%)$ . + +
MethodsPre.mIoUmAcc
PointNet [25]-41.149.0
PointNet++ [26]-53.5-
PointCNN [18]-57.363.9
KPConv [40]-67.172.8
Point-Bert [44]✓68.976.1
Point-MAE [23]✓68.476.2
PointDif [48]✓70.077.1
PreDifPoint(ours)✓71.876.5
+ +Table 3. Semantic Segmentation Results (%) Obtained on the S3DIS Dataset, Evaluated on Area 5. + +The masking operations that we include in the pre-training will affect the performance of the model in the downstream task. Thus, we first verified the effect of different mask ratios on the downstream tasks. As in Tab. 4, we examine the classification and segmentation ability of the model on the ScanObjectNN dataset and the S3DIS dataset, respectively. It is clear that the model achieves the best classification and segmentation performance when the mask ratio is 0.8. + +This may be due to the fact that the self-supervision task is too difficult when the masking ratio is too high, which may cause the model to have difficulty in learning effective feature representations, and will affect the model's ability to extract enough information from the limited training data to cope with the prediction task. However, when the masking ratio is too low, the model cannot be forced to learn deep feature representations and the model cannot obtain better prior knowledge. Only when the masking ratio is 0.8, it may just be beneficial for the model to learn effective feature representations. + +# Attention. + +The DXCA-Attention mechanism that we use in CGNet enhances model comprehension by considering feature relationships in both directions simultaneously, enabling a more comprehensive capture of information. To validate the utility of this attention mechanism, we report results using MLP, one-way covariate attention mechanism and DXCA-Attention mechanism on the ScanObjectNN dataset, respec + +
DatasetsScanObjectNNS3DIS(Area5)
Mask RatioOAmIoU
091.5869.1
0.590.9668.2
0.691.2369.5
0.791.8570.3
0.893.4271.8
0.992.3470.6
+ +Table 4. The Overall Accuracy (%) on ScanObjectNN and the mIoU (%) on S3DIS with different masking ratios. + +
AttentionOA
MLP92.56
Covariance Attention93.19
DXCA-Attention93.42
+ +tively. + +As shown in Tab. 5, our proposed DXCA-Attention mechanism obtains the best performance. It may be because MLP is able to capture the complex relationship between features through nonlinear transformations, but it is unable to take into account the specific interaction patterns between features. The one-way covariate attention mechanism, on the other hand, only considers the covariance relationship between features in one direction, which may have led to the loss of information. Only the DXCA-Attention mechanism can model the interactions between features more comprehensively, capture more complex nonlinear relationships, enhance the robustness of the model, and facilitate feature fusion, thus enhancing the expressive power of the model. + +# Density-Adaptive Sampling Strategy. + +The density-adaptive sampling strategy that we use in the diffusion process enables the model to focus more on the denser regions in the point cloud, which makes it easier for the model to recover the structural details of the denser regions. We compare this with the uniform timestep sampling strategy [48], and as shown in Tab. 6, the OA achieved by the density-adaptive sampling strategy is 0.11 higher than the uniform time-step sampling strategy. + +It indicates that uniform time-step sampling may lead to under-sampling in regions with higher density and oversampling in regions with lower density. Since many real-world applications of point cloud data tend to have nonuniform densities, the density-adaptive sampling strategy can utilize the data more efficiently, especially in those high-density regions that are critical to model performance. + +Table 5. The Forms of The Self-Attention Operators. + +
Sampling StrategynOA
Recurrent uniform sampling93.31
Density-Adaptive Sampling93.42
+ +Table 6. The Forms of The Sampling Strategy. + +# 5. Conclusion + +In this paper, we propose a pre-training framework based on a diffusion model called PreDifPoint. Specifically, we propose a conditional generation network to guide the model through the point-to-point denoising process. In this context, the two-way covariate attention mechanism used facilitates the extraction and representation of point cloud features, thus enhancing the model's comprehension. We also introduce a density-adaptive sampling strategy based on the time step to guide the model to pay more attention to the denser regions in the point cloud, thus making it easier for the model to recover the structural details of the dense regions. Extensive experiments have shown that our method has good feature learning and representation capabilities, and is able to achieve advanced results in a variety of real-world datasets, which further validates the effectiveness of the proposed method. In the future, we hope that our work will further improve the performance of point cloud analysis. + +# References + +[1] Tomer Amit, Tal Shaharbany, Eliya Nachmani, and Lior Wolf. Segdiff: Image segmentation with diffusion probabilistic models, 2022. 3 +[2] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models, 2022. 3 +[3] Adam Bretherton, Joshua J. Bon, David J. Warne, Kerrie Mengersen, and Christopher Drovandi. A principled approach to bayesian transfer learning, 2025. 3 +[4] Tolgahan Cakaloglu and Xiaowei Xu. Mrnn: A multi-resolution neural network with duplex attention for document retrieval in the context of question answering, 2019. 2 +[5] Shoufa Chen, Peize Sun, Yibing Song, and Ping Luo. Diffusional: Diffusion model for object detection, 2023. 3 +[6] Sachin Goyal, Ananya Kumar, Sankalp Garg, Zico Kolter, and Aditi Raghunathan. Finetune like you pretrain: Improved finetuning of zero-shot vision models, 2022. 1 +[7] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R. Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 7(2):187-199, 2021. 7 +[8] Jing Hao, Song Chen, Xiaodi Wang, and Shumin Han. Language-aware multiple datasets detection pretraining for detrs, 2023. 1 + +[9] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners, 2021. 1 +[10] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models, 2020. 1 +[11] Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, Tom Hennigan, Eric Noland, Katie Millican, George van den Driessche, Bogdan Damoc, Aurelia Guy, Simon Osindero, Karen Simonyan, Erich Elsen, Jack W. Rae, Oriol Vinyals, and Laurent Sifre. Training compute-optimal large language models, 2022. 1 +[12] Shouwei Hu, Xi Li, Banyao Ruan, and Zhihao Liu. An amplitude-encoding-based classical-quantum transfer learning framework: Outperforming classical methods in image recognition, 2025. 3 +[13] Mohammad Rafid Ul Islam, Prasad Tadepalli, and Alan Fern. Self-attention-based diffusion model for time-series imputation in partial blackout scenarios, 2025. 3 +[14] Liqiang Jing, Yiren Li, Junhao Xu, Yongcan Yu, Pei Shen, and Xuemeng Song. Vision enhanced generative pre-trained language model for multimodal sentence summarization. Machine Intelligence Research, 20:289-298, 2023. 1 +[15] Li Ju, Xingyi Yang, Qi Li, and Xinchao Wang. Graphbridge: Towards arbitrary transfer learning in gnns, 2025. 3 +[16] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything, 2023. 1 +[17] Xudong Li, Li Feng, Lei Li, and Chen Wang. Few-shot meta-learning on point cloud for semantic segmentation, 2021. 3 +[18] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on $\mathcal{X}$ -transformed points, 2018. 7 +[19] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation, 2023. 2 +[20] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3d mesh in 45 seconds without pershape optimization, 2023. 2 +[21] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2836-2844, 2021. 3 +[22] Zhaoyang Lyu, Zhifeng Kong, Xudong Xu, Liang Pan, and Dahua Lin. A conditional point diffusion-refinement paradigm for 3d point cloud completion, 2022. 3 +[23] Yatian Pang, Wenxiao Wang, Francis E. H. Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning, 2022. 1, 2, 7 +[24] Konpat Preechakul, Nattanat Chathee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion autoencoders: Toward a meaningful and decodable representation, 2022. 2 + +[25] Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation, 2017. 7 +[26] Charles R. Qi, Li Yi, Hao Su, and Leonidas J. Guibas. Point-net++: Deep hierarchical feature learning on point sets in a metric space, 2017. 7 +[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 1 +[28] Mojtaba Safari, Shansong Wang, Zach Eidex, Qiang Li, Erik H. Middlebrooks, David S. Yu, and Xiaofeng Yang. Mri super-resolution reconstruction using efficient diffusion probabilistic model with residual shifting, 2025. 3 +[29] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior, 2023. 2 +[30] Aaron van den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks, 2016. 2 +[31] Aaron van den Oord, Nal Kalchbrenner, Oriol Vinyals, Lasse Espeholt, Alex Graves, and Koray Kavukcuoglu. Conditional image generation with pixelCNN decoders, 2016. 2 +[32] Jiahui Wang, Haiyue Zhu, Haoren Guo, Abdullah Al Mamun, Cheng Xiang, and Tong Heng Lee. Few-shot point cloud semantic segmentation via contrastive self-supervision and multi-resolution attention, 2023. 1 +[33] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E. Sarma, Michael M. Bronstein, and Justin M. Solomon. Dynamic graph cnn for learning on point clouds, 2019. 7 +[34] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation, 2023. 2 +[35] Ziyi Wang, Xumin Yu, Yongming Rao, Jie Zhou, and Jiwen Lu. Take-a-photo: 3d-to-2d generative pre-training of point cloud models, 2023. 7 +[36] Zhendong Wang, Jianmin Bao, Shuyang Gu, Dong Chen, Wengang Zhou, and Houqiang Li. Design diffusion: High-quality text-to-design image generation with diffusion models, 2025. 3 +[37] Antoine Wehenkel and Gilles Louppe. Diffusion priors in variational autoencoders, 2021. 2 +[38] Julia Wolleb, Robin Sandkuhler, Florentin Bieder, Philippe Valmaggia, and Philippe C. Cattin. Diffusion models for implicit image segmentation ensembles, 2021. 3 +[39] Jay Zhangjie Wu, Yuxuan Zhang, Haithem Turki, Xuanchi Ren, Jun Gao, Mike Zheng Shou, Sanja Fidler, Zan Gojcic, and Huan Ling. Difix3d+: Improving 3d reconstructions with single-step diffusion models, 2025. 3 +[40] Yue Wu, Jiaming Liu, Maoguo Gong, Zhixiao Liu, Qiguang Miao, and Wenping Ma. Mpct: Multiscale point cloud transformer with a residual network. IEEE Transactions on Multimedia, 26:3505-3516, 2024. 7 +[41] Weilai Xiang, Hongyu Yang, Di Huang, and Yunhong Wang. Denoising diffusion autoencoders are unified self-supervised learners, 2023. 2 + +[42] Saining Xie, Jiatao Gu, Demi Guo, Charles R. Qi, Leonidas J. Guibas, and Or Litany. Pointcontrast: Unsupervised pre-training for 3d point cloud understanding, 2020. 2 +[43] Jiale Xu, Xintao Wang, Weihao Cheng, Yan-Pei Cao, Ying Shan, Xiaohu Qie, and Shenghua Gao. Dream3d: Zero-shot text-to-3d synthesis using 3d shape prior and text-to-image diffusion models, 2023. 2 +[44] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling, 2022. 1, 7 +[45] Renrui Zhang, Ziyu Guo, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, Hongsheng Li, and Peng Gao. Pointm2ae: Multi-scale masked autoencoders for hierarchical point cloud pre-training, 2022. 1, 2 +[46] Zaiwei Zhang, Rohit Girdhar, Armand Joulin, and Ishan Misra. Self-supervised pretraining of 3d features on any point-cloud, 2021. 1 +[47] Luda Zhao, Yihua Hu, Xing Yang, Zhenglei Dou, and Qilong Wu. Icddpm: Image-conditioned denoising diffusion probabilistic model for real-world complex point cloud single view reconstruction. Expert Systems with Applications, 259:125370, 2025. 3 +[48] Xiao Zheng, Xiaoshui Huang, Guofeng Mei, Yuenan Hou, Zhaoyang Lyu, Bo Dai, Wanli Ouyang, and Yongshun Gong. Point cloud pre-training with diffusion models, 2023. 1, 2, 3, 6, 7, 8 +[49] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion, 2021. 3 +[50] Mingjian Zhu, Kai Han, Enhua Wu, Qiulin Zhang, Ying Nie, Zhenzhong Lan, and Yunhe Wang. Dynamic resolution network, 2021. 7 +[51] Aleksei Zhuravlev, Zorah Lähner, and Vladislav Golyanik. Denoising functional maps: Diffusion models for shape correspondence, 2025. 3 \ No newline at end of file diff --git a/2025/Feature Extraction and Representation of Pre-training Point Cloud Based on Diffusion Models/images.zip b/2025/Feature Extraction and Representation of Pre-training Point Cloud Based on Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3418e8480d034c7ceb3c08b2026d794f24612e66 --- /dev/null +++ b/2025/Feature Extraction and Representation of Pre-training Point Cloud Based on Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2ea6293b41898e3dd7437b58934d9122edbe3974d272106408de01c66ea14978 +size 439048 diff --git a/2025/Feature Extraction and Representation of Pre-training Point Cloud Based on Diffusion Models/layout.json b/2025/Feature Extraction and Representation of Pre-training Point Cloud Based on Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..48152337210592a173c94b34ea741f0c1650bffd --- /dev/null +++ b/2025/Feature Extraction and Representation of Pre-training Point Cloud Based on Diffusion Models/layout.json @@ -0,0 +1,8541 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 69, + 102, + 541, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 102, + 541, + 138 + ], + "spans": [ + { + "bbox": [ + 69, + 102, + 541, + 138 + ], + "type": "text", + "content": "Feature Extraction and Representation of Pre-training Point Cloud Based on Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 144, + 162, + 200, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 162, + 200, + 175 + ], + "spans": [ + { + "bbox": [ + 144, + 162, + 200, + 175 + ], + "type": "text", + "content": "Chang Qiu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 121, + 176, + 223, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 176, + 223, + 190 + ], + "spans": [ + { + "bbox": [ + 121, + 176, + 223, + 190 + ], + "type": "text", + "content": "Southeast University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 117, + 192, + 227, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 192, + 227, + 202 + ], + "spans": [ + { + "bbox": [ + 117, + 192, + 227, + 202 + ], + "type": "text", + "content": "230238519@seu.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 273, + 162, + 337, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 273, + 162, + 337, + 175 + ], + "spans": [ + { + "bbox": [ + 273, + 162, + 337, + 175 + ], + "type": "text", + "content": "Feipeng Da*" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 252, + 176, + 354, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 176, + 354, + 190 + ], + "spans": [ + { + "bbox": [ + 252, + 176, + 354, + 190 + ], + "type": "text", + "content": "Southeast University" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 262, + 192, + 345, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 262, + 192, + 345, + 202 + ], + "spans": [ + { + "bbox": [ + 262, + 192, + 345, + 202 + ], + "type": "text", + "content": "dafp@seu.edu.cn" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 406, + 162, + 465, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 162, + 465, + 175 + ], + "spans": [ + { + "bbox": [ + 406, + 162, + 465, + 175 + ], + "type": "text", + "content": "Zilei Zhang" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 384, + 176, + 487, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 176, + 487, + 190 + ], + "spans": [ + { + "bbox": [ + 384, + 176, + 487, + 190 + ], + "type": "text", + "content": "Southeast University" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 380, + 192, + 489, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 192, + 489, + 202 + ], + "spans": [ + { + "bbox": [ + 380, + 192, + 489, + 202 + ], + "type": "text", + "content": "230238524@seu.edu.cn" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 255, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 296, + 628 + ], + "type": "text", + "content": "The pretrain-finetune paradigm of pre-training a model on large amounts of image and text data and then fine-tuning the model for a specific task has led to significant progress in many 2D image and natural language processing tasks. Similarly, the use of pre-training methods in point cloud data can also enhance the working performance and generalization ability of the model. Therefore, in this paper, we propose a pre-training framework based on a diffusion model called PreDifPoint. It is able to accomplish the pretraining of the model's backbone network through a diffusion process of gradual denoising. We aggregate the potential features extracted from the backbone network, input them as conditions into the subsequent diffusion model, and direct the point-to-point mapping relationship of the noisy point clouds at neighboring time steps, so as to generate high-quality point clouds and at the same time better perform various downstream tasks of the point clouds. We also introduce a bi-directional covariate attention (DXCA-Attention) mechanism for capturing complex feature interactions, fusing local and global features, and improving the detail recovery of point clouds. In addition, we propose a density-adaptive sampling strategy, which can help the model dynamically adjust the sampling strategy between different time steps, and guide the model to pay more attention to the denser regions in the point cloud, thus improving the effectiveness of the model in point cloud recovery. Our PreDifPoint framework achieves more competitive results on various real-world datasets. Specifically, PreDifPoint achieves an overall accuracy of " + }, + { + "bbox": [ + 55, + 255, + 296, + 628 + ], + "type": "inline_equation", + "content": "87.96\\%" + }, + { + "bbox": [ + 55, + 255, + 296, + 628 + ], + "type": "text", + "content": ", which is " + }, + { + "bbox": [ + 55, + 255, + 296, + 628 + ], + "type": "inline_equation", + "content": "0.35\\%" + }, + { + "bbox": [ + 55, + 255, + 296, + 628 + ], + "type": "text", + "content": " higher than PointDif, on the classification task on PB-T50-395RS, a variant of ScanObjectNN dataset." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 650, + 135, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 650, + 135, + 662 + ], + "spans": [ + { + "bbox": [ + 56, + 650, + 135, + 662 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 670, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 670, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 670, + 295, + 696 + ], + "type": "text", + "content": "Recent work has demonstrated that the pretrain-finetune paradigm, where a model is first pretrained on large" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 232, + 555, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 555, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 555, + 388 + ], + "type": "text", + "content": "amounts of image and text data and then fine-tuned for specific tasks, has led to substantial advancements in many 2D image and natural language processing tasks [11, 16, 27]. Pretraining the backbone on large-scale datasets enables the model to learn rich feature representations while also avoiding the need for large amounts of data and computational resources required for training a model from scratch [8, 9, 32]. Pretrained models, after fine-tuning, can learn task-specific knowledge, exhibit better generalization capabilities, and converge to optimal performance levels more rapidly [6, 14]. Similarly, using pretraining methods for point cloud data can also enhance the model's performance and generalization ability [44, 46, 48]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 388, + 556, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 556, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 556, + 581 + ], + "type": "text", + "content": "Generative pretraining methods, through generative models, capture the complex structure and intricate details of point cloud data, thereby obtaining more powerful feature representations. Early works such as PointBERT [44], Point-MAE [23], and Point-M2AE [45] employed self-supervised learning by reconstructing masked local point patches, forcing the model to learn local point cloud information. However, this masking and reconstruction approach may overlook the contextual relationships between masked and unmasked regions, preventing the model from effectively understanding the global structure of the object or scene. Inspired by diffusion models [10], which can understand global structure and long-term dependencies through the process of adding and removing noise, we propose using a diffusion model to guide the pretraining process of point clouds." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "type": "text", + "content": "Thus, we introduce a diffusion-based pretraining framework, named PreDifPoint. This framework pretrains only the backbone network of the model, and then fine-tunes the task-specific parts. The pretraining of the backbone network is achieved through a progressive denoising diffusion process, as shown in Fig. 1. By pretraining the backbone, it can learn meaningful features during the feature extraction phase, which also significantly enhances the model's generalization ability. At the same time, this reduces the number of parameters that need to be adjusted during training, accelerating convergence and reducing overfitting." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 145, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 145, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 145, + 713 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26559" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 72, + 294, + 162 + ], + "blocks": [ + { + "bbox": [ + 58, + 72, + 294, + 162 + ], + "lines": [ + { + "bbox": [ + 58, + 72, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 58, + 72, + 294, + 162 + ], + "type": "image", + "image_path": "c66474988fa9fbe369b0b9455a8976898db1d7f067ccb1487469a4f66e888563.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 172, + 296, + 218 + ], + "lines": [ + { + "bbox": [ + 55, + 172, + 296, + 218 + ], + "spans": [ + { + "bbox": [ + 55, + 172, + 296, + 218 + ], + "type": "text", + "content": "Figure 1. Schematic illustration of our PreDiffPoint. Our PreDiffPoint can be used to train different backbones and reconstruct the original point cloud from the noisy one through a diffusion process." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 240, + 295, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 240, + 295, + 418 + ], + "spans": [ + { + "bbox": [ + 54, + 240, + 295, + 418 + ], + "type": "text", + "content": "Specifically, the PreDifPoint framework consists of three core modules: the Conditional Point Generation Network (CGNet), the Conditional Denoising Network (CDNet), and the Conditional Diffusion Network (CDPNet). CGNet aggregates the latent features extracted by the backbone and feeds them as conditional inputs into the subsequent diffusion model. CDNet utilizes these conditional inputs to perform denoising and restore point cloud details. CDPNet, through the point-to-point mapping of noisy point clouds between adjacent time steps in the diffusion process [48], ensures that the generated point clouds excel in both global consistency and local detail. With this design, the model can generate high-quality point clouds based on various conditions, exhibiting strong diversity and adaptability to perform various downstream tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 421, + 295, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 421, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 295, + 540 + ], + "type": "text", + "content": "Additionally, we propose a density-adaptive sampling strategy. We divide the entire time step range into several subintervals, each containing multiple time steps. Within each subinterval, the sampling probability is dynamically adjusted according to the local density of the point cloud at each time step, prioritizing sampling from steps with higher density. This strategy allows for more fine-grained control over the sampling process at each time step and helps dynamically adjust the sampling strategy across time steps, improving the model's point cloud recovery performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 541, + 289, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 541, + 289, + 553 + ], + "spans": [ + { + "bbox": [ + 67, + 541, + 289, + 553 + ], + "type": "text", + "content": "Our main contributions can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 558, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 640 + ], + "type": "text", + "content": "- We propose a diffusion-based point cloud pretraining framework, PreDifPoint, which leverages conditional guidance for generation, denoising, and recovery, as well as global and local feature modeling. This framework enhances the performance of point cloud generation and denoising tasks and provides high-quality feature representations for downstream tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "- We introduce a DXCA-Attention mechanism into the model to capture complex global feature interactions and generate more instructive condition vectors to enhance the detail reconstruction of the point cloud. The condition vectors obtained through CGNet will be combined with time-step information and passed to the subsequent" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 323, + 73, + 416, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 73, + 416, + 84 + ], + "spans": [ + { + "bbox": [ + 323, + 73, + 416, + 84 + ], + "type": "text", + "content": "network for denoising." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 84, + 553, + 239 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 313, + 84, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 84, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 84, + 553, + 156 + ], + "type": "text", + "content": "- We propose a density-adaptive sampling strategy, which can help the model dynamically adjust the sampling strategy between different time steps, and guide the model to pay more attention to the denser regions in the point cloud, thus improving the effectiveness of the model in point cloud recovery." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 156, + 553, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 553, + 239 + ], + "type": "text", + "content": "- The PreDifPoint framework requires only fine-tuning for task-specific components (such as classifiers), allowing the model to perform optimally across various tasks, particularly in data-limited or computationally constrained environments. Notably, the PreDifPoint framework demonstrates competitive performance in a variety of practical downstream tasks." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 253, + 406, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 253, + 406, + 266 + ], + "spans": [ + { + "bbox": [ + 314, + 253, + 406, + 266 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 274, + 450, + 287 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 450, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 450, + 287 + ], + "type": "text", + "content": "2.1. Point Cloud Pretraining" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 292, + 555, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 292, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 292, + 555, + 422 + ], + "type": "text", + "content": "In deep learning, pretraining has become a key technique for enhancing model performance, especially when data is scarce or labels are limited. For point cloud learning tasks, pretraining is particularly important. Point cloud data typically has complex spatial structures and sparsity, and traditional training methods often rely on large amounts of labeled data, which is difficult to obtain in real-world applications. Therefore, learning effective point cloud representations through pretraining not only reduces reliance on labeled data but also improves the model's generalization and robustness in downstream tasks." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 425, + 554, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 554, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 554, + 484 + ], + "type": "text", + "content": "Currently, point cloud pretraining methods can be broadly categorized into three types: self-supervised learning, generative models [29, 34, 43], and transfer learning. Each approach has its unique advantages and application scenarios." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 485, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 554, + 665 + ], + "type": "text", + "content": "Self-supervised methods [4, 30, 31] can mask parts of the data and train the model to reconstruct the missing parts or optimize the model's representation ability through strategies like contrastive learning. Point-MAE [23] randomly masks certain points in the point cloud and trains the model to recover these missing points, thereby effectively learning the geometric features of the point cloud. Point-M2AE [45] for hierarchical self-supervised learning of 3D point clouds modifies the encoder and decoder into a pyramid architecture to incrementally model the spatial geometry and capture the fine-grained and high-level semantics of 3D shapes. PointContrast [42] employs contrastive learning to treat point cloud representation as a learning task, improving the model's ability to distinguish between similar and different point cloud segments." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Generative models [19, 20, 24, 37, 41], particularly denoising diffusion models and variational autoencoders, have made significant progress in point cloud pretraining in recent years. These models learn the generative distribution of" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26560" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": "point cloud data, capturing the details of the point cloud and completing missing data. DDPM [22] demonstrates how to use pretrained denoising diffusion models for conditional point cloud generation and completion, significantly improving the quality of point cloud reconstruction. PVD [49] introduces a conditional variational diffusion model (CV-Diffusion) that can generate high-quality 3D shapes through conditional information, effectively applied to point cloud generation and completion tasks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 180, + 294, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 180, + 294, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 294, + 276 + ], + "type": "text", + "content": "Transfer learning methods [3, 12, 15] involve pretraining a model on a large-scale dataset and then transferring it to a specific task, significantly improving performance when data is limited. Meta-PointNet [17] proposes a meta-learning-based pretraining method for point cloud classification and segmentation, where knowledge is shared across multiple tasks, greatly improving performance on few-shot tasks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 286, + 159, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 286, + 159, + 298 + ], + "spans": [ + { + "bbox": [ + 55, + 286, + 159, + 298 + ], + "type": "text", + "content": "2.2. Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 304, + 295, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 295, + 496 + ], + "type": "text", + "content": "The core idea of diffusion models [13, 28, 36, 39, 51] is to simulate a process of gradually adding noise and then recovering the data through reverse denoising. Diffusion models have been widely used in recognition tasks such as object detection [5] and semantic discrimination [1, 2, 38]. During training, the model learns how to recover real data from noise, which helps generate high-quality samples. This paper uses conditional diffusion models, which introduce additional conditional information into the diffusion process to guide the generated model. These conditional inputs can include labels, textual descriptions, or other external features that help the model generate data that meets specific conditions. Conditional diffusion models have shown excellent performance in multimodal generation tasks, particularly in image generation, point cloud generation, and speech synthesis." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 498, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 676 + ], + "type": "text", + "content": "Diffusion probabilistic models [21] proposed a point cloud probabilistic generation model inspired by nonequilibrium thermodynamics that utilizes a backward diffusion process to learn point distributions. And they achieved competitive performance in point cloud generation and automatic coding, and comparable results in unsupervised representation learning. PointDif [48] proposes a conditional diffusion model for point cloud pretraining tasks. This model can more precisely control the structure and details of the point cloud during the generation or denoising process, performing particularly well in point cloud denoising and generation tasks. ICDDPM [47] can uses 2D images as prior information to guide end-to-end 3D point cloud generation, and obtain high-quality 3D point clouds of different types of targets from real-world images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "The PreDifPoint framework we have developed pretrains the network by recovering the original point cloud from randomly noisy point clouds, forcing the network to learn local" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "content": "and global geometric priors of point clouds. The extracted point cloud features are passed as conditional information into the diffusion process, guiding the generation of point clouds step by step. Additionally, we propose a density-adaptive sampling strategy, which allows for finer control over the sampling process at each time step and helps dynamically adjust the sampling strategy between time steps, thus improving the model's ability to recover point clouds." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 178, + 370, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 178, + 370, + 190 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 370, + 190 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 198, + 555, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 555, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 555, + 366 + ], + "type": "text", + "content": "In our pre-training process, the backbone of the overall framework PreDifPoint is a transformer encoder. Notably, this backbone can be replaced with any other backbone network for pre-training. The input point cloud is grouped and features are extracted, followed by masking operations and contextual enhancement via a Transformer with positional encoding. Subsequently, CGNet further processes these features to generate conditional vectors that serve as conditional inputs to CDNet. In CDNet, time-step embedding is combined with conditional vectors. Finally, the CDPNet network integrates CGNet and CDNet to generate the original point cloud by stepwise denoising the data through a diffusion model. The pipeline of our PreDifPoint is shown in Fig. 2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 373, + 418, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 418, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 418, + 384 + ], + "type": "text", + "content": "3.1. Diffusion Process" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 389, + 554, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 554, + 437 + ], + "type": "text", + "content": "In the diffusion model, a clean point cloud " + }, + { + "bbox": [ + 313, + 389, + 554, + 437 + ], + "type": "inline_equation", + "content": "X^0 \\epsilon R^{n \\times 3}" + }, + { + "bbox": [ + 313, + 389, + 554, + 437 + ], + "type": "text", + "content": " is gradually corrupted into a fully noisy point cloud " + }, + { + "bbox": [ + 313, + 389, + 554, + 437 + ], + "type": "inline_equation", + "content": "X^T" + }, + { + "bbox": [ + 313, + 389, + 554, + 437 + ], + "type": "text", + "content": " by incrementally adding noise. The diffusion process at each time step " + }, + { + "bbox": [ + 313, + 389, + 554, + 437 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 389, + 554, + 437 + ], + "type": "text", + "content": " follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 339, + 457, + 555, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 457, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 339, + 457, + 555, + 472 + ], + "type": "interline_equation", + "content": "q \\left(X ^ {t} \\mid X ^ {t - 1}\\right) = \\mathcal {N} \\left(X ^ {t}; \\sqrt {\\alpha_ {t}} X ^ {t - 1}, (1 - \\alpha_ {t}) \\mathrm {I}\\right), \\tag {1}", + "image_path": "ab0f95ac7db6af23289b7cfe2aec5ba6983c4823181cd0e674ea3ebd3cfc5844.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 479, + 554, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 554, + 527 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 554, + 527 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 479, + 554, + 527 + ], + "type": "inline_equation", + "content": "\\alpha_{t} = 1 - \\beta_{t}" + }, + { + "bbox": [ + 313, + 479, + 554, + 527 + ], + "type": "text", + "content": " denotes the noise attenuation coefficient at time step, and " + }, + { + "bbox": [ + 313, + 479, + 554, + 527 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 313, + 479, + 554, + 527 + ], + "type": "text", + "content": " represents the noise level. Through this process, noise is gradually added to the point cloud until it becomes nearly random at " + }, + { + "bbox": [ + 313, + 479, + 554, + 527 + ], + "type": "inline_equation", + "content": "X^T" + }, + { + "bbox": [ + 313, + 479, + 554, + 527 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 528, + 554, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 554, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 554, + 575 + ], + "type": "text", + "content": "This reverse process predicts the denoising signals at each time step by the neural network " + }, + { + "bbox": [ + 313, + 528, + 554, + 575 + ], + "type": "inline_equation", + "content": "p_{\\theta}(X^t | X^{t-1}, c)" + }, + { + "bbox": [ + 313, + 528, + 554, + 575 + ], + "type": "text", + "content": ", which gradually recovers the original structure of the point cloud as shown in the following equation:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 330, + 594, + 555, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 594, + 555, + 609 + ], + "spans": [ + { + "bbox": [ + 330, + 594, + 555, + 609 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(X ^ {t} \\mid X ^ {t - 1}, c\\right) = \\mathcal {N} \\left(X ^ {t - 1}; \\mu_ {\\theta} \\left(X ^ {t}, t, c\\right), \\sigma_ {t} ^ {2} \\mathrm {I}\\right), \\tag {2}", + "image_path": "9331c8cffb32d3505883ee7fe2ee7e584e73df0cf1c33f3876eaef7b12fe1942.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 616, + 554, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 554, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 554, + 652 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 616, + 554, + 652 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}(X^{t},t,c)" + }, + { + "bbox": [ + 313, + 616, + 554, + 652 + ], + "type": "text", + "content": " is the mean predicted by the network, " + }, + { + "bbox": [ + 313, + 616, + 554, + 652 + ], + "type": "inline_equation", + "content": "X^t" + }, + { + "bbox": [ + 313, + 616, + 554, + 652 + ], + "type": "text", + "content": " representing the denoised result at step " + }, + { + "bbox": [ + 313, + 616, + 554, + 652 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 616, + 554, + 652 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 616, + 554, + 652 + ], + "type": "inline_equation", + "content": "\\sigma_t^2" + }, + { + "bbox": [ + 313, + 616, + 554, + 652 + ], + "type": "text", + "content": " is the noise variance." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 660, + 534, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 660, + 534, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 660, + 534, + 672 + ], + "type": "text", + "content": "3.2. Conditional Generation Network (CGNet)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "CGNet is responsible for extracting global and local features from the point cloud, fusing these features through a DXCA-Attention mechanism to more comprehensively" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26561" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 72, + 503, + 246 + ], + "blocks": [ + { + "bbox": [ + 107, + 72, + 503, + 246 + ], + "lines": [ + { + "bbox": [ + 107, + 72, + 503, + 246 + ], + "spans": [ + { + "bbox": [ + 107, + 72, + 503, + 246 + ], + "type": "image", + "image_path": "51f7b387b6e06eb17d4312d18e311f641af80d8d8a5343d7d3a8b8c97bfd5f32.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 256, + 555, + 312 + ], + "lines": [ + { + "bbox": [ + 54, + 256, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 54, + 256, + 555, + 312 + ], + "type": "text", + "content": "Figure 2. Pipeline of PreDifPoint. We first group the input point cloud and extract features, which are then subjected to masking operations and contextual enhancement via a transformer with position encoding. Subsequently, CGNet further processes these features to generate conditional vectors, which are used as conditional inputs to CDNet. In CDNet, time-step embedding is combined with conditional vectors. Finally, the CDPNet network integrates CGNet and CDNet to generate the original point cloud by stepwise denoising the data through a diffusion model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 333, + 294, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 333, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 54, + 333, + 294, + 392 + ], + "type": "text", + "content": "model the interactions between the features and better adapt to different data patterns and structures, as shown in Fig. 3. Finally, a conditional vector " + }, + { + "bbox": [ + 54, + 333, + 294, + 392 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 54, + 333, + 294, + 392 + ], + "type": "text", + "content": " associated with the density distribution is generated, which is used to guide the subsequent denoising process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 393, + 295, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 393, + 295, + 465 + ], + "spans": [ + { + "bbox": [ + 54, + 393, + 295, + 465 + ], + "type": "text", + "content": "We first project the input feature " + }, + { + "bbox": [ + 54, + 393, + 295, + 465 + ], + "type": "inline_equation", + "content": "F_{loc}" + }, + { + "bbox": [ + 54, + 393, + 295, + 465 + ], + "type": "text", + "content": " from the point cloud chunking encoding to the high dimensional space through the multilayer perceptron (MLP) to obtain the projected representation " + }, + { + "bbox": [ + 54, + 393, + 295, + 465 + ], + "type": "inline_equation", + "content": "F_{pro}" + }, + { + "bbox": [ + 54, + 393, + 295, + 465 + ], + "type": "text", + "content": " of the local features, and then obtain the global feature " + }, + { + "bbox": [ + 54, + 393, + 295, + 465 + ], + "type": "inline_equation", + "content": "F_{glo}" + }, + { + "bbox": [ + 54, + 393, + 295, + 465 + ], + "type": "text", + "content": " through the global pooling operation on " + }, + { + "bbox": [ + 54, + 393, + 295, + 465 + ], + "type": "inline_equation", + "content": "F_{pro}" + }, + { + "bbox": [ + 54, + 393, + 295, + 465 + ], + "type": "text", + "content": " as shown in the following equation:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 474, + 295, + 507 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 474, + 295, + 507 + ], + "spans": [ + { + "bbox": [ + 104, + 474, + 295, + 507 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} F _ {p r o} = M L P (F), \\\\ F _ {g l o} = G l o b a l M a x P o o l \\left(F _ {p r o}\\right). \\end{array} \\right. \\tag {3}", + "image_path": "91820ee8754ab760818515e64e8839084af25804a01f37747eedd7a6ab0ea1a2.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 512, + 296, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 512, + 296, + 572 + ], + "spans": [ + { + "bbox": [ + 55, + 512, + 296, + 572 + ], + "type": "text", + "content": "Then, the attention weights are obtained by calculating the covariance matrix in both directions to capture the global contextual relationships and complex feature interactions. In this process, we need to first obtain the important parameters, " + }, + { + "bbox": [ + 55, + 512, + 296, + 572 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 512, + 296, + 572 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 512, + 296, + 572 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 512, + 296, + 572 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 512, + 296, + 572 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 512, + 296, + 572 + ], + "type": "text", + "content": ", as shown in Eq. (4)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 100, + 580, + 295, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 580, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 100, + 580, + 295, + 646 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} Q = \\operatorname {r e s h a p e} \\left(F _ {p r o}\\right) \\in \\mathbb {R} ^ {B \\times G \\times C}, \\\\ K = \\operatorname {r e s h a p e} \\left(F _ {g l o}\\right) \\in \\mathbb {R} ^ {B \\times G \\times C}, \\\\ V _ {1} = \\operatorname {r e s h a p e} \\left(F _ {g l o}\\right) \\in \\mathbb {R} ^ {B \\times G \\times C}, \\\\ V _ {2} = \\operatorname {r e s h a p e} \\left(F _ {p r o}\\right) \\in \\mathbb {R} ^ {B \\times G \\times C}. \\end{array} \\right. \\tag {4}", + "image_path": "18c16a453240439736c2f1caf7d5d41dd82425395a3eac6dde25e1b668fa2e33.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 650, + 295, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 650, + 295, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 650, + 295, + 674 + ], + "type": "text", + "content": "Then centering along the feature dimension " + }, + { + "bbox": [ + 55, + 650, + 295, + 674 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 650, + 295, + 674 + ], + "type": "text", + "content": " as shown in Eq. (5)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 134, + 684, + 295, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 684, + 295, + 716 + ], + "spans": [ + { + "bbox": [ + 134, + 684, + 295, + 716 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} \\bar {Q} = Q - \\mathbb {E} _ {c} [ Q ], \\\\ \\bar {K} = K - \\mathbb {E} _ {c} [ K ]. \\end{array} \\right. \\tag {5}", + "image_path": "84d9abfb256723530e0537e738257bd48c65188959dec14ac961cb09f57a22d3.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 332, + 555, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 332, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 555, + 357 + ], + "type": "text", + "content": "Then the scaling factor " + }, + { + "bbox": [ + 313, + 332, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\sqrt{C}" + }, + { + "bbox": [ + 313, + 332, + 555, + 357 + ], + "type": "text", + "content": " is introduced to compute the covariance matrix, as shown in Eq. (6)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 323, + 373, + 553, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 373, + 553, + 456 + ], + "spans": [ + { + "bbox": [ + 323, + 373, + 553, + 456 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} A t t e n S c o r e _ {1} = S o f t m a x \\left(\\frac {\\bar {Q} \\cdot \\bar {K} ^ {T}}{\\sqrt {C}}\\right) \\in \\mathbb {R} ^ {B \\times G \\times C}, \\\\ A t t e n S c o r e _ {2} = S o f t m a x \\left(\\frac {\\bar {K} \\cdot \\bar {Q} ^ {T}}{\\sqrt {C}}\\right) \\in \\mathbb {R} ^ {B \\times G \\times C}. \\end{array} \\right. \\tag {6}", + "image_path": "8e3cad91d53785cc68ba7bcc63c7794e3a1a590c7a02a4436716ad5d24f6c442.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 456, + 555, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 555, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 555, + 480 + ], + "type": "text", + "content": "At this point, the bi-directional covariance features can be aggregated as shown in Eq. (7)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 353, + 487, + 553, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 487, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 353, + 487, + 553, + 536 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} O _ {1} = A t t e n S c o r e _ {1} \\cdot V _ {1} \\in \\mathbb {R} ^ {B \\times G \\times C}, \\\\ O _ {2} = A t t e n S c o r e _ {2} \\cdot V _ {2} \\in \\mathbb {R} ^ {B \\times G \\times C}, \\\\ O = \\operatorname {c o n c a t} \\left(O _ {1}, O _ {2}\\right) \\in \\mathbb {R} ^ {B \\times G \\times 2 C}. \\end{array} \\right. \\tag {7}", + "image_path": "692ca56a9ccc693c892ad28422f02fba01adde5dad8ef297bbed67fed525426d.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 539, + 555, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 539, + 555, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 555, + 576 + ], + "type": "text", + "content": "Finally, we fuse the projected features, global features and the attention-weighted features to output the conditioned vector " + }, + { + "bbox": [ + 313, + 539, + 555, + 576 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 539, + 555, + 576 + ], + "type": "text", + "content": ", as shown in Eq. (8)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 359, + 586, + 553, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 586, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 359, + 586, + 553, + 601 + ], + "type": "interline_equation", + "content": "c = M L P \\left(\\text {C o n c a t} \\left(F _ {\\text {p r o}}, A, F _ {\\text {g l o}}\\right)\\right). \\tag {8}", + "image_path": "e780d1407a05cfac5cde6ab5723e26497051c1aa852cc8eb56531218c300d582.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 612, + 527, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 527, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 527, + 625 + ], + "type": "text", + "content": "3.3. Conditional Denoising Network (CDNet)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 629, + 555, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 677 + ], + "type": "text", + "content": "CDNet is responsible for gradually removing the noise and restoring the point cloud structure during the diffusion process, and its core is to achieve layer-level feature optimization through residual connectivity." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": "First, the conditional vector " + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": " and the time embedding " + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "inline_equation", + "content": "t_{emd}" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": " are fused into the point cloud features through a dynamic gating mechanism to achieve density-aware local de" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26562" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 72, + 294, + 167 + ], + "blocks": [ + { + "bbox": [ + 58, + 72, + 294, + 167 + ], + "lines": [ + { + "bbox": [ + 58, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 58, + 72, + 294, + 167 + ], + "type": "image", + "image_path": "fcd4c81dc0cd386c1eb1b288656edc7c907211fb8eff1db34e11cc40aef08ceb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 180, + 296, + 301 + ], + "lines": [ + { + "bbox": [ + 55, + 180, + 296, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 296, + 301 + ], + "type": "text", + "content": "Figure 3. DXCA-Attention Module. This attention mechanism achieves cross-level information fusion through the two-way interaction between global and local features. In the local-to-global direction, query vectors and key vectors are generated by centering local features and global features respectively, and the contribution weights of local regions to global semantics are computed through the covariance matrix, and the semantic consistency of local features is enhanced after weighted aggregation. In the global-to-local direction, the global features are used as queries and local features as keys, and the modulation relation of global context to local details is captured by covariance attention." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 323, + 296, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 296, + 347 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 296, + 347 + ], + "type": "text", + "content": "noising. The time embedding " + }, + { + "bbox": [ + 55, + 323, + 296, + 347 + ], + "type": "inline_equation", + "content": "t_{emd}" + }, + { + "bbox": [ + 55, + 323, + 296, + 347 + ], + "type": "text", + "content": " and the fused features can be represented as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 60, + 367, + 294, + 408 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 367, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 60, + 367, + 294, + 408 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} t _ {e m d} = \\left[ \\beta_ {t}, \\sin \\left(\\beta_ {t}\\right), \\cos \\left(\\beta_ {t}\\right) \\right], \\\\ H _ {l} = \\sigma \\left(W _ {g} [ c \\oplus t _ {e m d} ]\\right) \\odot \\left(W _ {f} H _ {l - 1}\\right) + W _ {b} [ c \\oplus t _ {e m d} ], \\end{array} \\right. \\tag {9}", + "image_path": "41457ab43b05e98e086d5344ecef851badb6f34017e6c34a9d9304e43dd7f264.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 409, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 409, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 296, + 456 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 409, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 409, + 296, + 456 + ], + "type": "text", + "content": " is the Sigmoid function for generating gating weights in the range [0, 1], and " + }, + { + "bbox": [ + 55, + 409, + 296, + 456 + ], + "type": "inline_equation", + "content": "W_{g}, W_{f}, W_{b}" + }, + { + "bbox": [ + 55, + 409, + 296, + 456 + ], + "type": "text", + "content": " are the gating generation weights, feature transformation weights and bias generation weights, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 457, + 296, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 457, + 296, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 457, + 296, + 493 + ], + "type": "text", + "content": "Finally, to alleviate the gradient vanishing problem, the output of each of our layers can be connected using residuals such that the output of each layer is represented as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 96, + 506, + 295, + 519 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 506, + 295, + 519 + ], + "spans": [ + { + "bbox": [ + 96, + 506, + 295, + 519 + ], + "type": "interline_equation", + "content": "H _ {l} = H _ {l - 1} + C D N e t \\left(H _ {l - 1}, c, t _ {e m d}\\right). \\tag {10}", + "image_path": "ebe1ec73e9034d8b8b13152d203dba2ef4388d49d16a35342db6a0950c771310.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 525, + 295, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 295, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 295, + 538 + ], + "type": "text", + "content": "3.4. Conditional Diffusion Point Network (CDPNet)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 544, + 296, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 296, + 591 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 296, + 591 + ], + "type": "text", + "content": "The CDPNet integrates a density-weighted diffusion process, a conditional point generation network (CGNet), and a conditional denoising network (CDNet) to achieve stepwise generation from noisy to clean point clouds." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 592, + 296, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 592, + 296, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 296, + 652 + ], + "type": "text", + "content": "In a nutshell, it is to first get a completely noisy point cloud " + }, + { + "bbox": [ + 55, + 592, + 296, + 652 + ], + "type": "inline_equation", + "content": "X^T" + }, + { + "bbox": [ + 55, + 592, + 296, + 652 + ], + "type": "text", + "content": " from a clean point cloud " + }, + { + "bbox": [ + 55, + 592, + 296, + 652 + ], + "type": "inline_equation", + "content": "X^0" + }, + { + "bbox": [ + 55, + 592, + 296, + 652 + ], + "type": "text", + "content": " in the forward diffusion process, then obtain the conditional vector " + }, + { + "bbox": [ + 55, + 592, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 55, + 592, + 296, + 652 + ], + "type": "text", + "content": " by CGNet, and then update " + }, + { + "bbox": [ + 55, + 592, + 296, + 652 + ], + "type": "inline_equation", + "content": "X^T" + }, + { + "bbox": [ + 55, + 592, + 296, + 652 + ], + "type": "text", + "content": " step by step from " + }, + { + "bbox": [ + 55, + 592, + 296, + 652 + ], + "type": "inline_equation", + "content": "t = T" + }, + { + "bbox": [ + 55, + 592, + 296, + 652 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 592, + 296, + 652 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 55, + 592, + 296, + 652 + ], + "type": "text", + "content": " in CDNet, and finally get the reconstructed point cloud " + }, + { + "bbox": [ + 55, + 592, + 296, + 652 + ], + "type": "inline_equation", + "content": "X^0" + }, + { + "bbox": [ + 55, + 592, + 296, + 652 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 659, + 208, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 208, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 208, + 673 + ], + "type": "text", + "content": "3.5. Density-Adaptive Sampling" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "To dynamically adjust the time step scheduling in the diffusion model, we propose a density-adaptive sampling strategy. This strategy weights the sampling probability of each" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 553, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 106 + ], + "type": "text", + "content": "time step based on the local density of points, prioritizing denser regions (e.g., edges) during training to enhance detail reconstruction." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 108, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 553, + 133 + ], + "type": "text", + "content": "First, the local density " + }, + { + "bbox": [ + 313, + 108, + 553, + 133 + ], + "type": "inline_equation", + "content": "\\rho_{i}" + }, + { + "bbox": [ + 313, + 108, + 553, + 133 + ], + "type": "text", + "content": " of each point " + }, + { + "bbox": [ + 313, + 108, + 553, + 133 + ], + "type": "inline_equation", + "content": "x_{i}\\epsilon X_{0}" + }, + { + "bbox": [ + 313, + 108, + 553, + 133 + ], + "type": "text", + "content": " is computed using " + }, + { + "bbox": [ + 313, + 108, + 553, + 133 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 108, + 553, + 133 + ], + "type": "text", + "content": "-nearest neighbors:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 369, + 142, + 553, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 142, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 369, + 142, + 553, + 171 + ], + "type": "interline_equation", + "content": "\\rho \\left(x _ {i}\\right) = \\frac {1}{\\sum_ {j = 1} ^ {k} \\left\\| x _ {i} - x _ {j} \\right\\| + \\epsilon}, \\tag {11}", + "image_path": "402dfad6d3efcf975d580af83e4c00470c282ccf52af8660e53f138a4d9f6315.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 177, + 553, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 177, + 553, + 213 + ], + "spans": [ + { + "bbox": [ + 313, + 177, + 553, + 213 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 177, + 553, + 213 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 177, + 553, + 213 + ], + "type": "text", + "content": " is the neighborhood size, " + }, + { + "bbox": [ + 313, + 177, + 553, + 213 + ], + "type": "inline_equation", + "content": "\\| x_{i} - x_{j}\\|" + }, + { + "bbox": [ + 313, + 177, + 553, + 213 + ], + "type": "text", + "content": " is the Euclidean distance, and " + }, + { + "bbox": [ + 313, + 177, + 553, + 213 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 177, + 553, + 213 + ], + "type": "text", + "content": " is a small constant to avoid division by zero." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 213, + 554, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 554, + 262 + ], + "type": "text", + "content": "The time step interval " + }, + { + "bbox": [ + 313, + 213, + 554, + 262 + ], + "type": "inline_equation", + "content": "[1, T]" + }, + { + "bbox": [ + 313, + 213, + 554, + 262 + ], + "type": "text", + "content": " is uniformly divided into " + }, + { + "bbox": [ + 313, + 213, + 554, + 262 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 213, + 554, + 262 + ], + "type": "text", + "content": " sub-intervals " + }, + { + "bbox": [ + 313, + 213, + 554, + 262 + ], + "type": "inline_equation", + "content": "\\{Q_1, Q_2, \\dots, Q_h\\}" + }, + { + "bbox": [ + 313, + 213, + 554, + 262 + ], + "type": "text", + "content": ", each corresponding to different noise levels. The sampling probability for " + }, + { + "bbox": [ + 313, + 213, + 554, + 262 + ], + "type": "inline_equation", + "content": "Q_i" + }, + { + "bbox": [ + 313, + 213, + 554, + 262 + ], + "type": "text", + "content": " is weighted by the local density within the interval:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 389, + 271, + 553, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 271, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 389, + 271, + 553, + 297 + ], + "type": "interline_equation", + "content": "P \\left(Q _ {i} \\mid \\rho\\right) = \\frac {\\rho_ {i}}{\\sum_ {i = 1} ^ {T} \\rho_ {i}}. \\tag {12}", + "image_path": "bda6bfb108a13665657be37ab7930466f061aedf80bedf75ba1f6fcec49c632a.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 302, + 553, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 302, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 302, + 553, + 338 + ], + "type": "text", + "content": "This weighting strategy ensures that more samples can be performed in the denser regions, thus improving the reconstruction accuracy in these regions." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 338, + 554, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 554, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 554, + 422 + ], + "type": "text", + "content": "Within each interval " + }, + { + "bbox": [ + 313, + 338, + 554, + 422 + ], + "type": "inline_equation", + "content": "Q_{i}" + }, + { + "bbox": [ + 313, + 338, + 554, + 422 + ], + "type": "text", + "content": ", we randomly sample a time step " + }, + { + "bbox": [ + 313, + 338, + 554, + 422 + ], + "type": "inline_equation", + "content": "t_{i}" + }, + { + "bbox": [ + 313, + 338, + 554, + 422 + ], + "type": "text", + "content": " from that interval and compute the diffusion process at that time step. Since the sampling is based on local density weighting, the sampling probability for each interval reflects the denseness of the region. In updating the sampling, we use the following weighted loss function to calculate the loss at each time step:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 367, + 432, + 553, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 432, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 367, + 432, + 553, + 464 + ], + "type": "interline_equation", + "content": "L (\\theta , \\rho) = \\frac {1}{h} \\sum_ {i = 0} ^ {h - 1} L \\left(\\theta , \\rho , t _ {i} \\sim Q _ {i}\\right), \\tag {13}", + "image_path": "6b1a32a83cfae9dd675bc02a4664b9b27dfacafb6a5f8e7669d400bab5668a5a.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 326, + 472, + 534, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 472, + 534, + 484 + ], + "spans": [ + { + "bbox": [ + 326, + 472, + 534, + 484 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 326, + 472, + 534, + 484 + ], + "type": "inline_equation", + "content": "L(\\theta, \\rho, t_i \\sim Q_i)" + }, + { + "bbox": [ + 326, + 472, + 534, + 484 + ], + "type": "text", + "content": " is the loss for the sampled " + }, + { + "bbox": [ + 326, + 472, + 534, + 484 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 326, + 472, + 534, + 484 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 492, + 491, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 492, + 491, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 492, + 491, + 506 + ], + "type": "text", + "content": "3.6. Density-Weighted Diffusion Loss" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 510, + 554, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 554, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 554, + 570 + ], + "type": "text", + "content": "In the training of diffusion models combined with density-adaptive sampling, the loss function consists mainly of the error in the prediction noise and the error in the density weighting. Our goal is to minimize the difference between the predicted noise and the real noise at each time step." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 571, + 554, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 571, + 554, + 606 + ], + "spans": [ + { + "bbox": [ + 313, + 571, + 554, + 606 + ], + "type": "text", + "content": "The standard diffusion model loss is achieved by calculating the mean square error (MSE) between the predicted noise and the real noise, as shown in Eq. (14):" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 314, + 625, + 557, + 653 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 625, + 557, + 653 + ], + "spans": [ + { + "bbox": [ + 314, + 625, + 557, + 653 + ], + "type": "interline_equation", + "content": "L _ {d e n o i s e} (\\theta) = \\mathbb {E} _ {x _ {0}, t, \\epsilon} [ \\| \\epsilon - \\epsilon_ {\\theta} (\\sqrt {\\alpha_ {t}} x _ {0} + \\sqrt {1 - \\alpha_ {t}} \\epsilon , t, c) \\| ^ {2} ], \\tag {14}", + "image_path": "fb2f57ca6590bec6255c6f0acb76ec282bf7ff9225705f4e5cb51e1171c4a6f5.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": " is the input original point cloud, " + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": " is the current time step, " + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": " is the noise predicted by the network, " + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": " is the noise actually added, " + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": " is the noise attenuation coefficient at each time step, " + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": " is the conditional information generated by the conditional aggregation network CGNet, and " + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{x_0,t,\\epsilon}" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26563" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "is the indication of the need for averaging all the possible values of these variables to calculate the final loss." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 97, + 296, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 97, + 296, + 157 + ], + "spans": [ + { + "bbox": [ + 55, + 97, + 296, + 157 + ], + "type": "text", + "content": "During the density-adaptive sampling process, we need to adjust the computation of the loss to be consistent with the sampling probability of each interval. The final training loss combining density-weighted sampling and the standard MSE loss can be expressed as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 177, + 293, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 177, + 293, + 262 + ], + "spans": [ + { + "bbox": [ + 58, + 177, + 293, + 262 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L (\\theta , \\rho) = \\frac {1}{h} \\sum_ {i = 0} ^ {h - 1} \\mathbb {E} _ {t _ {i} \\sim Q _ {i}} [ \\\\ \\begin{array}{l} \\mathbb {E} _ {x _ {0}, \\epsilon} \\left[ \\left\\| \\epsilon - \\epsilon_ {\\theta} \\left(\\sqrt {\\alpha_ {t _ {i}}} x _ {0} + \\sqrt {1 - \\alpha_ {t _ {i}}} \\epsilon , t _ {i}, c\\right) \\right\\| ^ {2} \\right] \\\\ ], \\end{array} \\tag {15} \\\\ \\end{array}", + "image_path": "975b3d8917bf827cdb8fe10dd9cab4a976fb9b7cccfb53fd1871113c4f39602c.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 265, + 296, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 265, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 296, + 312 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 265, + 296, + 312 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 55, + 265, + 296, + 312 + ], + "type": "text", + "content": " is the time step sampled from the interval " + }, + { + "bbox": [ + 55, + 265, + 296, + 312 + ], + "type": "inline_equation", + "content": "Q_i" + }, + { + "bbox": [ + 55, + 265, + 296, + 312 + ], + "type": "text", + "content": ", and the loss function enhances the reconstruction of these regions by weighting the samples at each time step to ensure that denser regions can have higher weights." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 324, + 136, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 136, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 136, + 338 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 344, + 165, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 344, + 165, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 344, + 165, + 357 + ], + "type": "text", + "content": "4.1. Pre-training Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 362, + 296, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 362, + 296, + 481 + ], + "spans": [ + { + "bbox": [ + 55, + 362, + 296, + 481 + ], + "type": "text", + "content": "We use ShapeNet to pre-train our model, with the core version of this dataset containing 51,300 3D models across 55 common object categories. For each 3D shape, we sample 1,024 points to serve as the input for the model. Following the approach in PointDif [48], we use 41,952 shapes as the training set. Additionally, we employ the KNN algorithm to select the " + }, + { + "bbox": [ + 55, + 362, + 296, + 481 + ], + "type": "inline_equation", + "content": "k = 32" + }, + { + "bbox": [ + 55, + 362, + 296, + 481 + ], + "type": "text", + "content": " nearest points as point patches, resulting in 64 point cloud patches. Furthermore, we set the embedding dimension of the Transformer encoder to 384, with 6 attention heads, and the condition dimension to 768." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 483, + 296, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 296, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 296, + 518 + ], + "type": "text", + "content": "During pre-training, we set the weight decay to 0.05 and the learning rate to 0.001. Our model is pre-trained for 300 epochs with a batch size of 192." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 519, + 296, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 296, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 296, + 638 + ], + "type": "text", + "content": "Additionally, we visualize the point clouds generated by the PreDifPoint framework to demonstrate the effectiveness of our pre-training. As shown in Fig. 4, we use the masked point cloud features, extracted and used as conditions, to guide the diffusion model in generating the original point cloud, with a masking rate of 0.8. The visualization results indicate that the PreDifPoint framework generates high-quality point clouds and that the learned geometric priors effectively guide subsequent tasks such as recognition and detection." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 647, + 167, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 167, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 167, + 659 + ], + "type": "text", + "content": "4.2. Downstream Tasks" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "The performance of a pre-trained model, obtained in a fine-tuning task, is a test of the effectiveness of the pre-trained model. Since a high-quality point cloud pre-trained model should have a high hierarchical geometric prior, we tested" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 76, + 553, + 196 + ], + "blocks": [ + { + "bbox": [ + 317, + 76, + 553, + 196 + ], + "lines": [ + { + "bbox": [ + 317, + 76, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 317, + 76, + 553, + 196 + ], + "type": "image", + "image_path": "43369277d394f2ca882e82548da3f31cfab1edb5028405624bfc12f6ca4b5184.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 209, + 555, + 254 + ], + "lines": [ + { + "bbox": [ + 313, + 209, + 555, + 254 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 555, + 254 + ], + "type": "text", + "content": "Figure 4. Visualization results on the ShapeNet validation set. We show a total of six categories of point clouds, each showing the input point cloud, the masked point cloud, and the reconstructed point cloud from left to right." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 277, + 554, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 554, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 554, + 300 + ], + "type": "text", + "content": "the performance of the pre-trained model on a large number of real-world datasets." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 300, + 408, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 300, + 408, + 312 + ], + "spans": [ + { + "bbox": [ + 314, + 300, + 408, + 312 + ], + "type": "text", + "content": "Object Classification." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 314, + 554, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 554, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 554, + 338 + ], + "type": "text", + "content": "We first evaluated the classification ability of the PreD-ifPoint framework in the ScanObjectNNs dataset." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 339, + 554, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 554, + 446 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 554, + 446 + ], + "type": "text", + "content": "The ScanObjectNN dataset has 15 classes, which is actually more challenging due to the complexity of its contexts and object locations. The ScanObjectNN dataset is divided into three subsets: OBJ-ONLY, (only objects), OBJ-BG(objects and background), and PB-T50-RS(objects, background, and artificially added perturbations). We take the Overall Accuracy on these three subsets as the evaluation metric, and the detailed experimental results are summarized in Tab. 2." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 448, + 554, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 554, + 544 + ], + "type": "text", + "content": "As can be seen from the table, our PreDifPoint framework achieves excellent classification results in the ScanObjectNN dataset, regardless of the presence or absence of the complex background of artificially added perturbations. It demonstrates the strong adaptability and robustness of our model in the face of complex and changing environments. Compared with other state-of-the-art models, the PreDif-Point framework also achieves the best performance." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 544, + 409, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 544, + 409, + 555 + ], + "spans": [ + { + "bbox": [ + 314, + 544, + 409, + 555 + ], + "type": "text", + "content": "Object Segmentation." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 556, + 554, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 554, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 554, + 616 + ], + "type": "text", + "content": "We also validated the ability of our model for object segmentation in the ShapeNetPart dataset. ShapeNetPart contains a total of 16 large categories, which are labeled with semantic information for the semantic segmentation task of the point cloud." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": "We use mIoU to evaluate the object segmentation performance of the model, and from Tab. 1, we can see that overall, PreDifPoint has the best segmentation performance, which is " + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": " better than MPCT. Among other things, our model also achieves the best performance in multiple categories. This indicates that our model has been pre-trained to learn the generalized feature representation, which can extract the key features in the input data more efficiently and" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26564" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 71, + 552, + 170 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 552, + 170 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 552, + 170 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 552, + 170 + ], + "type": "table", + "html": "
MethodmIoUairplanebagcapcarchairearphoneguitarknifelamplaptopmotorbikemugpistolrocketskateboardtable
PointNet [25]83.783.478.782.574.989.673.091.585.980.895.365.293.081.257.972.880.6
PointNet++ [26]85.182.479.087.777.390.871.891.085.983.795.371.694.181.358.776.482.6
DGCNN [33]85.284.083.486.777.890.674.791.287.582.895.766.394.981.163.574.582.6
PointCNN [18]86.184.186.586.080.890.679.792.388.485.396.177.295.284.264.280.083.0
DRNet [50]86.484.385.088.379.591.279.391.889.085.295.772.294.282.060.676.884.2
PCT [7]86.485.082.489.081.291.971.591.388.186.395.864.695.883.662.277.683.7
MPCT [40]86.784.985.789.380.992.278.692.688.486.696.573.695.783.563.780.983.6
PreDiffPoint(ours)87.585.385.289.081.591.276.592.888.287.396.276.294.784.864.976.381.5
", + "image_path": "227b5ea366b0980a7cbd9b448affaa79c312f7197b3610f86924fdb917e7166f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 209, + 305, + 354 + ], + "blocks": [ + { + "bbox": [ + 178, + 178, + 430, + 190 + ], + "lines": [ + { + "bbox": [ + 178, + 178, + 430, + 190 + ], + "spans": [ + { + "bbox": [ + 178, + 178, + 430, + 190 + ], + "type": "text", + "content": "Table 1. Part Segmentation Results (%) on the ShapeNetPart Dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 209, + 305, + 354 + ], + "lines": [ + { + "bbox": [ + 56, + 209, + 305, + 354 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 305, + 354 + ], + "type": "table", + "html": "
MethodsPre.OBJ-ONLYOBJ-BGPB-T50-RS
PointNet [25]-79.273.368.0
PointNet++ [26]-84.382.377.9
PointCNN [18]-85.586.178.5
DGCNN [33]-86.282.878.1
Transformer [44]-80.5579.8677.24
Point-Bert [44]✓88.1287.4383.07
Point-MAE [23]✓88.2990.0285.18
TAP [35]✓89.5090.3685.67
PointDif [48]✓91.9193.2987.61
PreDifPoint(ours)✓92.2593.4287.96
", + "image_path": "58afa463ba8034b2aa9a38dfc741fca49958b8521a7f125e49a42a11f93cee9d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 408, + 220, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 408, + 220, + 432 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 220, + 432 + ], + "type": "text", + "content": "thus improve the segmentation accuracy. Indoor Semantic Segmentation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 433, + 295, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 295, + 516 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 295, + 516 + ], + "type": "text", + "content": "We further validate the semantic segmentation capability of our model in the indoor dataset S3DIS. The S3DIS dataset contains point cloud data and detailed semantic labels for several indoor environments, which contains a total of 6 regions (Area1, Area2, Area3, Area4, Area5, Area6), which are further divided into 271 rooms or spaces. We tested our model in Area5 while training in other areas." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 518, + 295, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 295, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 295, + 672 + ], + "type": "text", + "content": "We used mIoU and mAcc to evaluate the performance of the model, and from Tab. 3, we can see that our mAcc lags slightly behind compared to PointDif, probably due to the model's limited ability to extract features in certain categories, but still achieves a more competitive accuracy. In addition, PreDifPoint improves by " + }, + { + "bbox": [ + 55, + 518, + 295, + 672 + ], + "type": "inline_equation", + "content": "1.8\\%" + }, + { + "bbox": [ + 55, + 518, + 295, + 672 + ], + "type": "text", + "content": " in mIoU, which indicates that our model is able to understand the spatial relationships in the image well and segment different regions accurately. It is clear that our model obtains more meaningful geometric prior knowledge, which enables it to better understand the contextual semantics and local geometric relationships in complex scenes and achieve better performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 684, + 149, + 696 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 684, + 149, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 684, + 149, + 696 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 701, + 122, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 122, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 122, + 713 + ], + "type": "text", + "content": "Masking ratio." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 315, + 209, + 555, + 330 + ], + "blocks": [ + { + "bbox": [ + 55, + 361, + 295, + 384 + ], + "lines": [ + { + "bbox": [ + 55, + 361, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 295, + 384 + ], + "type": "text", + "content": "Table 2. Object classification results on ScanObjectNN. We report the Overall Accuracy " + }, + { + "bbox": [ + 55, + 361, + 295, + 384 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 55, + 361, + 295, + 384 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 209, + 555, + 330 + ], + "lines": [ + { + "bbox": [ + 315, + 209, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 315, + 209, + 555, + 330 + ], + "type": "table", + "html": "
MethodsPre.mIoUmAcc
PointNet [25]-41.149.0
PointNet++ [26]-53.5-
PointCNN [18]-57.363.9
KPConv [40]-67.172.8
Point-Bert [44]✓68.976.1
Point-MAE [23]✓68.476.2
PointDif [48]✓70.077.1
PreDifPoint(ours)✓71.876.5
", + "image_path": "442e1d02e3496407b105d03e1874600ee66ab0f10f0f4da357fd95ec8528b14d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 338, + 555, + 361 + ], + "lines": [ + { + "bbox": [ + 313, + 338, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 555, + 361 + ], + "type": "text", + "content": "Table 3. Semantic Segmentation Results (%) Obtained on the S3DIS Dataset, Evaluated on Area 5." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 386, + 555, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 386, + 555, + 482 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 555, + 482 + ], + "type": "text", + "content": "The masking operations that we include in the pre-training will affect the performance of the model in the downstream task. Thus, we first verified the effect of different mask ratios on the downstream tasks. As in Tab. 4, we examine the classification and segmentation ability of the model on the ScanObjectNN dataset and the S3DIS dataset, respectively. It is clear that the model achieves the best classification and segmentation performance when the mask ratio is 0.8." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 483, + 556, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 556, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 556, + 615 + ], + "type": "text", + "content": "This may be due to the fact that the self-supervision task is too difficult when the masking ratio is too high, which may cause the model to have difficulty in learning effective feature representations, and will affect the model's ability to extract enough information from the limited training data to cope with the prediction task. However, when the masking ratio is too low, the model cannot be forced to learn deep feature representations and the model cannot obtain better prior knowledge. Only when the masking ratio is 0.8, it may just be beneficial for the model to learn effective feature representations." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 616, + 361, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 616, + 361, + 625 + ], + "spans": [ + { + "bbox": [ + 314, + 616, + 361, + 625 + ], + "type": "text", + "content": "Attention." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": "The DXCA-Attention mechanism that we use in CGNet enhances model comprehension by considering feature relationships in both directions simultaneously, enabling a more comprehensive capture of information. To validate the utility of this attention mechanism, we report results using MLP, one-way covariate attention mechanism and DXCA-Attention mechanism on the ScanObjectNN dataset, respec" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26565" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 301, + 184 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 301, + 184 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 301, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 301, + 184 + ], + "type": "table", + "html": "
DatasetsScanObjectNNS3DIS(Area5)
Mask RatioOAmIoU
091.5869.1
0.590.9668.2
0.691.2369.5
0.791.8570.3
0.893.4271.8
0.992.3470.6
", + "image_path": "4b66558844d29c73d80e8492172a1e86ccba167f749388a1419dc4eb15fdc5fd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 85, + 229, + 268, + 291 + ], + "blocks": [ + { + "bbox": [ + 55, + 192, + 296, + 215 + ], + "lines": [ + { + "bbox": [ + 55, + 192, + 296, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 296, + 215 + ], + "type": "text", + "content": "Table 4. The Overall Accuracy (%) on ScanObjectNN and the mIoU (%) on S3DIS with different masking ratios." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 85, + 229, + 268, + 291 + ], + "lines": [ + { + "bbox": [ + 85, + 229, + 268, + 291 + ], + "spans": [ + { + "bbox": [ + 85, + 229, + 268, + 291 + ], + "type": "table", + "html": "
AttentionOA
MLP92.56
Covariance Attention93.19
DXCA-Attention93.42
", + "image_path": "15ce3738c2c45f1aacdadcc381b9c0f37b1b64a7d23c9f73c49f3928edc00819.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 336, + 83, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 336, + 83, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 336, + 83, + 348 + ], + "type": "text", + "content": "tively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 350, + 295, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 350, + 295, + 517 + ], + "spans": [ + { + "bbox": [ + 54, + 350, + 295, + 517 + ], + "type": "text", + "content": "As shown in Tab. 5, our proposed DXCA-Attention mechanism obtains the best performance. It may be because MLP is able to capture the complex relationship between features through nonlinear transformations, but it is unable to take into account the specific interaction patterns between features. The one-way covariate attention mechanism, on the other hand, only considers the covariance relationship between features in one direction, which may have led to the loss of information. Only the DXCA-Attention mechanism can model the interactions between features more comprehensively, capture more complex nonlinear relationships, enhance the robustness of the model, and facilitate feature fusion, thus enhancing the expressive power of the model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 518, + 216, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 216, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 216, + 529 + ], + "type": "text", + "content": "Density-Adaptive Sampling Strategy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 532, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 295, + 628 + ], + "type": "text", + "content": "The density-adaptive sampling strategy that we use in the diffusion process enables the model to focus more on the denser regions in the point cloud, which makes it easier for the model to recover the structural details of the denser regions. We compare this with the uniform timestep sampling strategy [48], and as shown in Tab. 6, the OA achieved by the density-adaptive sampling strategy is 0.11 higher than the uniform time-step sampling strategy." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "text", + "content": "It indicates that uniform time-step sampling may lead to under-sampling in regions with higher density and oversampling in regions with lower density. Since many real-world applications of point cloud data tend to have nonuniform densities, the density-adaptive sampling strategy can utilize the data more efficiently, especially in those high-density regions that are critical to model performance." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 330, + 70, + 541, + 119 + ], + "blocks": [ + { + "bbox": [ + 79, + 299, + 271, + 312 + ], + "lines": [ + { + "bbox": [ + 79, + 299, + 271, + 312 + ], + "spans": [ + { + "bbox": [ + 79, + 299, + 271, + 312 + ], + "type": "text", + "content": "Table 5. The Forms of The Self-Attention Operators." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 330, + 70, + 541, + 119 + ], + "lines": [ + { + "bbox": [ + 330, + 70, + 541, + 119 + ], + "spans": [ + { + "bbox": [ + 330, + 70, + 541, + 119 + ], + "type": "table", + "html": "
Sampling StrategynOA
Recurrent uniform sampling93.31
Density-Adaptive Sampling93.42
", + "image_path": "f45a7be82f3ca020f1d9c4fc94b47ffd69129442f2dab709f637b5855bf59bb2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 349, + 127, + 518, + 139 + ], + "lines": [ + { + "bbox": [ + 349, + 127, + 518, + 139 + ], + "spans": [ + { + "bbox": [ + 349, + 127, + 518, + 139 + ], + "type": "text", + "content": "Table 6. The Forms of The Sampling Strategy." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 314, + 160, + 388, + 172 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 160, + 388, + 172 + ], + "spans": [ + { + "bbox": [ + 314, + 160, + 388, + 172 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 180, + 555, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 555, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 555, + 395 + ], + "type": "text", + "content": "In this paper, we propose a pre-training framework based on a diffusion model called PreDifPoint. Specifically, we propose a conditional generation network to guide the model through the point-to-point denoising process. In this context, the two-way covariate attention mechanism used facilitates the extraction and representation of point cloud features, thus enhancing the model's comprehension. We also introduce a density-adaptive sampling strategy based on the time step to guide the model to pay more attention to the denser regions in the point cloud, thus making it easier for the model to recover the structural details of the dense regions. Extensive experiments have shown that our method has good feature learning and representation capabilities, and is able to achieve advanced results in a variety of real-world datasets, which further validates the effectiveness of the proposed method. In the future, we hope that our work will further improve the performance of point cloud analysis." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 408, + 373, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 408, + 373, + 420 + ], + "spans": [ + { + "bbox": [ + 315, + 408, + 373, + 420 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 319, + 428, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 319, + 428, + 555, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 428, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 319, + 428, + 555, + 460 + ], + "type": "text", + "content": "[1] Tomer Amit, Tal Shaharbany, Eliya Nachmani, and Lior Wolf. Segdiff: Image segmentation with diffusion probabilistic models, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 319, + 462, + 554, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 462, + 554, + 495 + ], + "spans": [ + { + "bbox": [ + 319, + 462, + 554, + 495 + ], + "type": "text", + "content": "[2] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 319, + 497, + 554, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 497, + 554, + 530 + ], + "spans": [ + { + "bbox": [ + 319, + 497, + 554, + 530 + ], + "type": "text", + "content": "[3] Adam Bretherton, Joshua J. Bon, David J. Warne, Kerrie Mengersen, and Christopher Drovandi. A principled approach to bayesian transfer learning, 2025. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 319, + 531, + 554, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 531, + 554, + 574 + ], + "spans": [ + { + "bbox": [ + 319, + 531, + 554, + 574 + ], + "type": "text", + "content": "[4] Tolgahan Cakaloglu and Xiaowei Xu. Mrnn: A multi-resolution neural network with duplex attention for document retrieval in the context of question answering, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 319, + 576, + 554, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 576, + 554, + 599 + ], + "spans": [ + { + "bbox": [ + 319, + 576, + 554, + 599 + ], + "type": "text", + "content": "[5] Shoufa Chen, Peize Sun, Yibing Song, and Ping Luo. Diffusional: Diffusion model for object detection, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 319, + 600, + 554, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 600, + 554, + 633 + ], + "spans": [ + { + "bbox": [ + 319, + 600, + 554, + 633 + ], + "type": "text", + "content": "[6] Sachin Goyal, Ananya Kumar, Sankalp Garg, Zico Kolter, and Aditi Raghunathan. Finetune like you pretrain: Improved finetuning of zero-shot vision models, 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 319, + 634, + 554, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 634, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 319, + 634, + 554, + 677 + ], + "type": "text", + "content": "[7] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R. Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 7(2):187-199, 2021. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 319, + 680, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 680, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 319, + 680, + 554, + 713 + ], + "type": "text", + "content": "[8] Jing Hao, Song Chen, Xiaodi Wang, and Shumin Han. Language-aware multiple datasets detection pretraining for detrs, 2023. 1" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26566" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 294, + 106 + ], + "type": "text", + "content": "[9] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners, 2021. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 129 + ], + "type": "text", + "content": "[10] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models, 2020. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 130, + 295, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 295, + 217 + ], + "type": "text", + "content": "[11] Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, Tom Hennigan, Eric Noland, Katie Millican, George van den Driessche, Bogdan Damoc, Aurelia Guy, Simon Osindero, Karen Simonyan, Erich Elsen, Jack W. Rae, Oriol Vinyals, and Laurent Sifre. Training compute-optimal large language models, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "type": "text", + "content": "[12] Shouwei Hu, Xi Li, Banyao Ruan, and Zhihao Liu. An amplitude-encoding-based classical-quantum transfer learning framework: Outperforming classical methods in image recognition, 2025. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "text", + "content": "[13] Mohammad Rafid Ul Islam, Prasad Tadepalli, and Alan Fern. Self-attention-based diffusion model for time-series imputation in partial blackout scenarios, 2025. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 342 + ], + "type": "text", + "content": "[14] Liqiang Jing, Yiren Li, Junhao Xu, Yongcan Yu, Pei Shen, and Xuemeng Song. Vision enhanced generative pre-trained language model for multimodal sentence summarization. Machine Intelligence Research, 20:289-298, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 342, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 294, + 365 + ], + "type": "text", + "content": "[15] Li Ju, Xingyi Yang, Qi Li, and Xinchao Wang. Graphbridge: Towards arbitrary transfer learning in gnns, 2025. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 365, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 409 + ], + "type": "text", + "content": "[16] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 411, + 294, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 432 + ], + "type": "text", + "content": "[17] Xudong Li, Li Feng, Lei Li, and Chen Wang. Few-shot meta-learning on point cloud for semantic segmentation, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 434, + 294, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 466 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 466 + ], + "type": "text", + "content": "[18] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on " + }, + { + "bbox": [ + 56, + 434, + 294, + 466 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 56, + 434, + 294, + 466 + ], + "type": "text", + "content": "-transformed points, 2018. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 468, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 294, + 510 + ], + "type": "text", + "content": "[19] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 512, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 555 + ], + "type": "text", + "content": "[20] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3d mesh in 45 seconds without pershape optimization, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 557, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 600 + ], + "type": "text", + "content": "[21] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2836-2844, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 601, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 294, + 635 + ], + "type": "text", + "content": "[22] Zhaoyang Lyu, Zhifeng Kong, Xudong Xu, Liang Pan, and Dahua Lin. A conditional point diffusion-refinement paradigm for 3d point cloud completion, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 636, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 669 + ], + "type": "text", + "content": "[23] Yatian Pang, Wenxiao Wang, Francis E. H. Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning, 2022. 1, 2, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[24] Konpat Preechakul, Nattanat Chathee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion autoencoders: Toward a meaningful and decodable representation, 2022. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "text", + "content": "[25] Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation, 2017. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "text", + "content": "[26] Charles R. Qi, Li Yi, Hao Su, and Leonidas J. Guibas. Point-net++: Deep hierarchical feature learning on point sets in a metric space, 2017. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 141, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 196 + ], + "type": "text", + "content": "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 197, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 240 + ], + "type": "text", + "content": "[28] Mojtaba Safari, Shansong Wang, Zach Eidex, Qiang Li, Erik H. Middlebrooks, David S. Yu, and Xiaofeng Yang. Mri super-resolution reconstruction using efficient diffusion probabilistic model with residual shifting, 2025. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 241, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 274 + ], + "type": "text", + "content": "[29] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 276, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 297 + ], + "type": "text", + "content": "[30] Aaron van den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks, 2016. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 298, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 553, + 331 + ], + "type": "text", + "content": "[31] Aaron van den Oord, Nal Kalchbrenner, Oriol Vinyals, Lasse Espeholt, Alex Graves, and Koray Kavukcuoglu. Conditional image generation with pixelCNN decoders, 2016. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 332, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 374 + ], + "type": "text", + "content": "[32] Jiahui Wang, Haiyue Zhu, Haoren Guo, Abdullah Al Mamun, Cheng Xiang, and Tong Heng Lee. Few-shot point cloud semantic segmentation via contrastive self-supervision and multi-resolution attention, 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 376, + 553, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 410 + ], + "type": "text", + "content": "[33] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E. Sarma, Michael M. Bronstein, and Justin M. Solomon. Dynamic graph cnn for learning on point clouds, 2019. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 411, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 453 + ], + "type": "text", + "content": "[34] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 455, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 455, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 553, + 487 + ], + "type": "text", + "content": "[35] Ziyi Wang, Xumin Yu, Yongming Rao, Jie Zhou, and Jiwen Lu. Take-a-photo: 3d-to-2d generative pre-training of point cloud models, 2023. 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 489, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 489, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 489, + 553, + 533 + ], + "type": "text", + "content": "[36] Zhendong Wang, Jianmin Bao, Shuyang Gu, Dong Chen, Wengang Zhou, and Houqiang Li. Design diffusion: High-quality text-to-design image generation with diffusion models, 2025. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 534, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 553, + 555 + ], + "type": "text", + "content": "[37] Antoine Wehenkel and Gilles Louppe. Diffusion priors in variational autoencoders, 2021. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 557, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 553, + 590 + ], + "type": "text", + "content": "[38] Julia Wolleb, Robin Sandkuhler, Florentin Bieder, Philippe Valmaggia, and Philippe C. Cattin. Diffusion models for implicit image segmentation ensembles, 2021. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 591, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 591, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 591, + 553, + 635 + ], + "type": "text", + "content": "[39] Jay Zhangjie Wu, Yuxuan Zhang, Haithem Turki, Xuanchi Ren, Jun Gao, Mike Zheng Shou, Sanja Fidler, Zan Gojcic, and Huan Ling. Difix3d+: Improving 3d reconstructions with single-step diffusion models, 2025. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 635, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 679 + ], + "type": "text", + "content": "[40] Yue Wu, Jiaming Liu, Maoguo Gong, Zhixiao Liu, Qiguang Miao, and Wenping Ma. Mpct: Multiscale point cloud transformer with a residual network. IEEE Transactions on Multimedia, 26:3505-3516, 2024. 7" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "type": "text", + "content": "[41] Weilai Xiang, Hongyu Yang, Di Huang, and Yunhong Wang. Denoising diffusion autoencoders are unified self-supervised learners, 2023. 2" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26567" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 295, + 455 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[42] Saining Xie, Jiatao Gu, Demi Guo, Charles R. Qi, Leonidas J. Guibas, and Or Litany. Pointcontrast: Unsupervised pre-training for 3d point cloud understanding, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 107, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 107, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 55, + 107, + 295, + 150 + ], + "type": "text", + "content": "[43] Jiale Xu, Xintao Wang, Weihao Cheng, Yan-Pei Cao, Ying Shan, Xiaohu Qie, and Shenghua Gao. Dream3d: Zero-shot text-to-3d synthesis using 3d shape prior and text-to-image diffusion models, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 152, + 294, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 152, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 57, + 152, + 294, + 185 + ], + "type": "text", + "content": "[44] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling, 2022. 1, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 186, + 295, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 186, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 57, + 186, + 295, + 229 + ], + "type": "text", + "content": "[45] Renrui Zhang, Ziyu Guo, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, Hongsheng Li, and Peng Gao. Pointm2ae: Multi-scale masked autoencoders for hierarchical point cloud pre-training, 2022. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 231, + 294, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 231, + 294, + 263 + ], + "spans": [ + { + "bbox": [ + 57, + 231, + 294, + 263 + ], + "type": "text", + "content": "[46] Zaiwei Zhang, Rohit Girdhar, Armand Joulin, and Ishan Misra. Self-supervised pretraining of 3d features on any point-cloud, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 265, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 265, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 57, + 265, + 294, + 319 + ], + "type": "text", + "content": "[47] Luda Zhao, Yihua Hu, Xing Yang, Zhenglei Dou, and Qilong Wu. Icddpm: Image-conditioned denoising diffusion probabilistic model for real-world complex point cloud single view reconstruction. Expert Systems with Applications, 259:125370, 2025. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 320, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 320, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 57, + 320, + 294, + 363 + ], + "type": "text", + "content": "[48] Xiao Zheng, Xiaoshui Huang, Guofeng Mei, Yuenan Hou, Zhaoyang Lyu, Bo Dai, Wanli Ouyang, and Yongshun Gong. Point cloud pre-training with diffusion models, 2023. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 365, + 294, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 365, + 294, + 388 + ], + "spans": [ + { + "bbox": [ + 57, + 365, + 294, + 388 + ], + "type": "text", + "content": "[49] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 388, + 294, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 388, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 57, + 388, + 294, + 421 + ], + "type": "text", + "content": "[50] Mingjian Zhu, Kai Han, Enhua Wu, Qiulin Zhang, Ying Nie, Zhenzhong Lan, and Yunhe Wang. Dynamic resolution network, 2021. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 422, + 294, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 422, + 294, + 455 + ], + "spans": [ + { + "bbox": [ + 57, + 422, + 294, + 455 + ], + "type": "text", + "content": "[51] Aleksei Zhuravlev, Zorah Lähner, and Vladislav Golyanik. Denoising functional maps: Diffusion models for shape correspondence, 2025. 3" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26568" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Feature Purification Matters_ Suppressing Outlier Propagation for Training-Free Open-Vocabulary Semantic Segmentation/4f979197-bed4-425a-a1d2-c2b5ca2160b9_content_list.json b/2025/Feature Purification Matters_ Suppressing Outlier Propagation for Training-Free Open-Vocabulary Semantic Segmentation/4f979197-bed4-425a-a1d2-c2b5ca2160b9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cb7388e935d14d47aa692387b24ee456ade04e7c --- /dev/null +++ b/2025/Feature Purification Matters_ Suppressing Outlier Propagation for Training-Free Open-Vocabulary Semantic Segmentation/4f979197-bed4-425a-a1d2-c2b5ca2160b9_content_list.json @@ -0,0 +1,1624 @@ +[ + { + "type": "text", + "text": "Feature Purification Matters: Suppressing Outlier Propagation for Training-Free Open-Vocabulary Semantic Segmentation", + "text_level": 1, + "bbox": [ + 163, + 130, + 834, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shuo Jin $^{1,2}$ Siyue Yu $^{1*}$ Bingfeng Zhang $^{3}$ Mingjie Sun $^{4}$ Yi Dong $^{2}$ Jimin Xiao $^{1}$ $^{1}$ Xi'an Jiaotong-Liverpool University $^{2}$ University of Liverpool \n $^{3}$ China University of Petroleum (East China) $^{4}$ Soochow University \n{shuo.jin, yi.dong}@liverpool.ac.uk, {siyue.yu02, jimin.xiao}@xjtlu.edu.cn \nbingfeng.zhang@upc.edu.cn, mjsun@suda.edu.cn", + "bbox": [ + 147, + 202, + 843, + 290 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Training-free open-vocabulary semantic segmentation has advanced with vision-language models like CLIP, which exhibit strong zero-shot abilities. However, CLIP's attention mechanism often wrongly emphasises specific image tokens, namely outliers, which results in irrelevant over-activation. Existing approaches struggle with these outliers that arise in intermediate layers and propagate through the model, ultimately degrading spatial perception. In this paper, we propose a Self-adaptive Feature Purifier framework (SFP) to suppress propagated outliers and enhance semantic representations for open-vocabulary semantic segmentation. Specifically, based on an in-depth analysis of attention responses between image and class tokens, we design a self-adaptive outlier mitigator to detect and mitigate outliers at each layer for propagated feature purification. In addition, we introduce a semantic-aware attention enhancer to augment attention intensity in semantically relevant regions, which strengthens the purified feature to focus on objects. Further, we introduce a hierarchical attention integrator to aggregate multi-layer attention maps to refine spatially coherent feature representations for final segmentation. Our proposed SFP enables robust outlier suppression and object-centric feature representation, leading to a more precise segmentation. Extensive experiments show that our method achieves state-of-the-art performance and surpasses existing methods by an average of $4.6\\%$ mIoU on eight segmentation benchmarks. The code is released at: https://github.com/Kimsure/SFP.", + "bbox": [ + 89, + 358, + 483, + 781 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 808, + 220, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Open-vocabulary semantic segmentation (OVSS) [4, 58] seeks to partition an image into distinct regions and assign pixel-level labels to arbitrary semantic categories. Un", + "bbox": [ + 89, + 833, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3a809d8d6a21c9db72c8c9f3de6b62c8340aaecec366c51646fb2b91e2579c1c.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 517, + 327, + 614, + 401 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3114cb73dc14bef72cdbcbae1c9dd2645ba2f83eea07987b45cf6211178fa784.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 616, + 327, + 710, + 401 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ec6b2006f99b0a5202453aaa61a917a5f9fe0ac356e732bc8d5e233cb5244a7f.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 710, + 327, + 807, + 401 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c4f8781f13f508d9d11ad39347cc9035561e4d98b07fe6491addac41b9b76c5d.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 808, + 327, + 903, + 401 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5e9d4ab5123f1a5541ceae7175fe2b8e0c0012c6dcab1917b3e6a2cae5088baa.jpg", + "image_caption": [ + "(e)", + "Figure 1. Visualization of attention and feature maps. (a) Attention map of the selected target object token (marked as $\\star$ ). (b) Attention map of the selected background token (marked as $\\star$ ). (c) Attention map of the class token. (d) Diagonal weights of the attention map of each image token. (e) Detected outliers. (f) - (h) output feature map of CLIP [38], SCLIP [45], and our SFP." + ], + "image_footnote": [], + "bbox": [ + 517, + 411, + 614, + 484 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/985463eea3e0187c90e3448f13faec3f03e9b8bd1c7111b9d6bd794cf9bf4d1a.jpg", + "image_caption": [ + "(f)" + ], + "image_footnote": [], + "bbox": [ + 616, + 411, + 710, + 484 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/772505dcee4a892fcc1253b7db85cab2a9660fcc8a2407c5ec9890c08344c1a1.jpg", + "image_caption": [ + "(g)" + ], + "image_footnote": [], + "bbox": [ + 710, + 411, + 807, + 484 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e5703a47e2d24303ae0dd1ccdc5b4fc7b0ab8e9f4fd2f06ce3c18696f858fd6d.jpg", + "image_caption": [ + "(h)" + ], + "image_footnote": [], + "bbox": [ + 808, + 411, + 903, + 484 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "like traditional semantic segmentation, OVSS must handle an open set of classes, making it more challenging. Vision-language models (VLM), such as CLIP [38], demonstrate remarkable zero-shot performance by leveraging large-scale image-text pairs [39]. Thus, existing approaches can be categorized into two pipelines: fine-tuning CLIP or training-free CLIP. Although the fine-tuning pipeline yields superior performance, it risks overfitting and compromises generalization [30, 43]. In contrast, the training-free pipeline preserves CLIP's original generalization capability without additional training, which is more convenient and efficient.", + "bbox": [ + 511, + 626, + 906, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent training-free methods [12, 40] have demonstrated that in the original CLIP's pair-wise attention maps, there exist the same highly correlated image tokens for all input tokens, such as those marked with the white box in Fig. 1 (a) - (b). In such a situation, regardless of whether it is the target image token (Fig. 1 (a)), the background image to-", + "bbox": [ + 511, + 809, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author: siyue.yu02@xjtlu.edu.cn", + "bbox": [ + 109, + 886, + 370, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20291", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ken (Fig. 1 (b)), or the class token (Fig. 1 (c)), the pair-wise token relationships within the attention map are all over-activated for these identical tokens, i.e., outliers. Consequently, when aggregated with outliers, the derived semantic feature representation is easily polluted to overemphasize these outliers and the interpretation of the target region is impeded, e.g., the feature map of 'car' in Fig 1 (f) focuses on the outliers instead of the object itself.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To mitigate outliers and ensure more accurate semantic feature representation, many training-free approaches have presented diverse modifications to the self-attention mechanism in the last layer [14, 16, 24, 45]. While these strategies effectively restrict CLIP's attention to ignore outliers, some remain unfiltered, as shown in Fig. 1 (g). Moreover, some approaches [2, 12] rely on empirical parameter tuning to detect and remove a fixed number of outliers, which is less robust to different backbones and images. Beyond these optimization approaches, several studies [12, 40] further reveal that outliers predominantly appear in the intermediate layers. In this way, once outliers emerge, they will propagate through subsequent layers, continuously affecting semantic feature representations. Yet, previous approaches fail to analyze and address outliers in the intermediate layers so that outliers are easily propagated to pollute the final semantic feature. Thus, our objective is to adaptively eliminate outliers across multiple layers to achieve more robust semantic feature representations without training.", + "bbox": [ + 91, + 214, + 483, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building on these insights, we propose a Self-adaptive Feature Purifier framework (SFP) to resolve outliers adaptively and enhance semantic representations for CLIP-based training-free OVSS. Firstly, we design a Self-adaptive Outlier Mitigator (SOM) to identify and mitigate outliers across multiple layers. Tokens are expected to focus primarily on their relevant regions, especially themselves, rather than irrelevant regions in the attention map. However, as shown in Fig. 1 (d), we observe that the self-response of image tokens in the attention map is not always prominent, especially outliers. The comparison of Fig. 1 (c) & (d) reveals that outliers exhibit a stronger response to the class token than their self-response. Hence, our SOM can identify outliers by simply computing the difference between the self-response values of the image tokens in the attention map (i.e., the diagonal weights of the attention map) and their attention values with the class token. The detected outliers are shown in Fig. 1 (e). In this way, our parameter-free SOM enables adaptive detection across various layers to mitigate the propagating influence of outliers without manual tuning.", + "bbox": [ + 91, + 505, + 483, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Nevertheless, outliers not only cause tokens to overemphasize them but also impede tokens' responses to semantic regions in the attention map. Therefore, merely removing the outliers can't directly increase the attention weights of the related regions. To enhance the tokens' attention responses to semantic regions, we propose a Semantic-aware", + "bbox": [ + 89, + 810, + 480, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Attention Enhancer (SAE). SAE incorporates the self-self attention mechanism in CLIP's last layer and applies it to the purified feature derived from SOM at the penultimate layer, enabling a better focus on semantic regions. In addition, shallow attention maps typically contain structured object information and focus on relevant semantic regions [27, 30]. To further enhance the precise spatial perception of the final feature map, we propose a Hierarchical Attention Integrator (HAI) as an assistant to the SAE, where the attention maps in previous layers are firstly processed to mitigate the influence of the outliers and then integrated to refine the final semantic feature for prediction. The refined semantic feature for the final prediction is shown in Fig. 1 (h). Our SFP can generate a more purified and complete semantic feature for segmentation.", + "bbox": [ + 511, + 90, + 903, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct comprehensive experiments to evaluate the proposed SFP. Our method outperforms existing methods and achieves state-of-the-art (SOTA) performance across eight evaluation datasets. In summary, our contributions are concluded as follows:", + "bbox": [ + 511, + 318, + 903, + 393 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present a Self-adaptive Feature Purifier framework (SFP) for training-free OVSS, which can derive a more accurate semantic feature to strengthen the final open-vocabulary semantic segmentation.", + "- We propose a simple yet effective mechanism, Self-adaptive Outlier Mitigator (SOM), which adaptively detects and mitigates outliers across CLIP's diverse layers so that the influence of the propagated influence of outliers can be suppressed.", + "- We introduce a Semantic-aware Attention Enhancer (SAE) and a Hierarchical Attention Integrator (HAI) to further refine the semantic feature relationships on relevant regions, aggregating the more relevant semantic feature for final prediction.", + "- Comprehensive experiments demonstrate that SFP derives the best performance under a fair comparison e.g., SFP achieves a gain of $5.6\\%$ mIoU on Cityscapes and an average gain of $4.6\\%$ mIoU across eight datasets." + ], + "bbox": [ + 514, + 398, + 903, + 670 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 686, + 653, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. pre-trained Vision-Language Models", + "text_level": 1, + "bbox": [ + 511, + 710, + 831, + 728 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Pre-trained vision-language models [17, 26, 38] have drawn significant attention for their ability to bridge visual and textual modalities. CLIP [38] has emerged as a highly successful model trained on a large-scale dataset of image-text pairs using contrastive learning. OpenCLIP [9] builds upon CLIP by leveraging public datasets such as LAION [39]. Despite their strong zero-shot capabilities, these models are pretrained at the image level for classification. This presents a notable limitation: the attention maps of VLMs tend to overemphasize outliers rather than semantic image tokens, leading to insufficient capture of spatial details. This draw", + "bbox": [ + 511, + 734, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20292", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "back hinders the performance in dense prediction tasks like semantic segmentation, which demands precise pixel-level understanding. Our goal is to adaptively detect and mitigate the impact of these outliers and enhance the semantic feature representation for dense prediction tasks.", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Open-vocabulary Segmentation", + "text_level": 1, + "bbox": [ + 89, + 180, + 369, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Semantic segmentation involves the pixel-wise classification of an image. Compared with the traditional segmentation methods [6, 8, 31, 49] that are trained and evaluated on the same fixed set of seen classes. OVSS methods [36, 59] aim to recognize and segment unseen objects during inference by harnessing the zero-shot ability of VLMs. Existing works can be broadly divided into two categories: training-based and training-free.", + "bbox": [ + 89, + 203, + 483, + 323 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training-based Methods Training-based methods [18, 37, 56] typically require fine-tuning on a fixed set of categories from a pixel-level annotated dataset. Some approaches follow a two-stage pipeline, where the first stage extracts the mask proposals, and the second stage assigns semantic labels to them. For instance, OVSeg [51] first trains class-agnostic mask proposals using the query-based framework Mask2Former [8] and then fine-tunes CLIP [38] to classify the cropped and masked images. Other approaches adopt a single-stage pipeline. SAN [52] introduces a side adapter to adapt CLIP [38] for both classification and segmentation. SED [48] presents a simple encoder-decoder architecture with category early rejection for fast inference. CATSeg [10] constructs a pixel-level cost map for segmentation. However, these training-based methods are prone to overfitting the training dataset, which limits the generalization ability due to fine-tuning CLIP.", + "bbox": [ + 89, + 325, + 483, + 582 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training-free Methods Unlike training-based methods, training-free methods [20, 41] aim to directly adapt VLMs for OVSS without extra training. Several approaches leverage the vision foundation models [5, 21, 42] to improve CLIP's spatial coherence. LaVG [19] uses DINO [5] for panoptic cuts, which iteratively discovers object masks until all image pixels are covered. ProxyCLIP [25] adjusts attention weights based on affinities learned from DINO's feature correspondence. However, these methods fail to exploit CLIP's inherent potential. Other efforts aim to modify CLIP's final layer to address its poor spatial consistency. For example, MaskCLIP [14] only utilizes the value feature of the final layer for prediction. CLIPtrase [40] integrates self-self attention with clustering for post-processing. SCLIP [45] and ClearCLIP [24] both modify the last layer self-attention mechanism to better represent the local information. Nevertheless, few training-free methods address propagated outliers, which act as noisy inputs across multiple layers and degrade spatial perception. In contrast, our approach enables adaptive detection and mitigation of propagated outliers among various layers, which improves", + "bbox": [ + 89, + 584, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CLIP's semantic awareness and boosts its internal potential.", + "bbox": [ + 511, + 90, + 903, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 117, + 604, + 132 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 513, + 142, + 624, + 157 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fig. 2 depicts the whole framework of SFP, including four main modules: a frozen CLIP backbone to encode the input image and text descriptions, a SOM to detect and eliminate outliers, an SAE to augment the attention intensity on related semantic regions, and an HAI that leverages intermediate-layer attention maps to assist SAE for enhanced semantic feature representation. The complete inference process is as follows:", + "bbox": [ + 511, + 164, + 906, + 284 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) First, the image and text descriptions are input to the CLIP encoder for the image feature and the text feature. Among the CLIP image encoder, our SOM is added after each transformer layer to mitigate the outliers and generate the purified feature for propagation.", + "2) Then, the derived final feature map from the last SOM, $F_{SOM}$ , is input to our SAE, which replaces the final transformer layer. SAE emphasizes the attention values of relevant semantic regions to generate the strengthened semantic feature $F_{SAE}$ .", + "3) After that, our HAI extracts the multi-layer attention maps from the image encoder to further refine $F_{SAE}$ for the final semantic feature $F_{vis}$ .", + "4) Finally, the segmentation result is obtained by calculating the cosine similarity of the final semantic feature $F_{vis}$ and text feature $F_{txt}$ ." + ], + "bbox": [ + 511, + 285, + 903, + 526 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Preliminary", + "text_level": 1, + "bbox": [ + 513, + 534, + 642, + 550 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In general, the training-free OVSS approach requires CLIP [38], which encompasses an image encoder and a text encoder to acquire multi-model features. Subsequently, these features are aligned via linear projection layers. In detail, the input image is first partitioned into a token sequence $X_{in} = [x_{cls}, x_1, \\dots, x_N] \\in R^{(N + 1) \\times d}$ , where $x_i$ denotes the image token embedding, $x_{cls}$ denotes the class token embedding, $N$ is the image token sequence length, and $d$ is the dimension of token embeddings, respectively. Then, these tokens are input to the image encoder, which contains a stack of vision transformer layers.", + "bbox": [ + 511, + 555, + 906, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Each transformer layer consists of a multi-head self-attention (MHSA) mechanism and a feed-forward network (FFN) layer to extract the image feature. Specifically, in each MHSA, the input sequence $X_{in}$ is mapped to query, key, and value embeddings, $Q, K, V \\in R^{(N + 1)\\times d}$ . Afterwards, $Q$ and $K$ are calculated to obtain the attention map $Attn \\in R^{(N + 1)\\times (N + 1)}$ , which can be formulated as:", + "bbox": [ + 511, + 720, + 905, + 827 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nA t t n = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 619, + 832, + 903, + 866 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $T$ means matrix transposition. After that, the attention map is multiplied $V$ to aggregate the global infor", + "bbox": [ + 511, + 869, + 903, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20293", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/df20bce62542ccaf74ec7943fc6e61eeec8bbe555389af3e25a8aaa3c3419022.jpg", + "image_caption": [ + "Figure 2. Framework of our SFP. It contains four main modules: a frozen CLIP backbone for feature extraction, a Self-adaptive Outlier Mitigator (SOM) at each image encoder layer for outlier detection and elimination, a Semantic-aware Attention Enhancer (SAE) at the last layer to emphasise attention values of relevant semantic regions and a Hierarchical Attention Integrator (HAI) for object-centric feature representation. The segmentation result is generated by calculating the cosine similarity between the optimized semantic feature $F_{vis}$ and the text feature $F_{txt}$ ." + ], + "image_footnote": [], + "bbox": [ + 117, + 88, + 880, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "mation, and an FFN is used to obtain the image feature $F_{img} \\in R^{(N + 1) \\times d}$ , formulated as:", + "bbox": [ + 89, + 508, + 482, + 540 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {i m g} = \\operatorname {F F N} (A t t n V). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 547, + 482, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the class embedding $x_{cls}$ is not used to predict the segmentation result, we omit it and reshape $F_{img}$ to obtain the visual feature $F_{vis} \\in R^{h \\times w \\times d}$ , where $h$ and $w$ denote the height and width of the feature map.", + "bbox": [ + 89, + 571, + 482, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Meanwhile, the textual description is derived from the standard ImageNet [13] prompts and transformed to the text feature $F_{txt} \\in R^{c \\times d}$ with CLIP's text encoder, where $c$ represents the number of classes. By calculating the cosine similarity between $F_{vis}$ and $F_{txt}$ , the segmentation map $Seg \\in R^{H \\times W}$ can be generated like:", + "bbox": [ + 89, + 633, + 482, + 724 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS e g = U (\\underset {c} {\\arg \\max } \\cos (F _ {v i s}, F _ {t x t})), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 732, + 482, + 757 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\cos (\\cdot)$ means cosine similarity calculation and $U(\\cdot)$ denotes the upsampling operation to match the original input image size. $H$ and $W$ denote the height and width of the input image, respectively.", + "bbox": [ + 89, + 763, + 483, + 825 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Self-adaptive Outlier Mitigator", + "text_level": 1, + "bbox": [ + 89, + 833, + 366, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CLIP [38] is pre-trained at the image level for classification, which tends to concentrate on the most discriminative feature that is important for classification. In this way, all input", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tokens, including both class and image tokens, will emphasize these identical tokens (outliers) in attention maps [12]. In other words, the corresponding outlier column in the attention map will be highlighted. Further, given that the classification process relies solely on the class token, the class token may exhibit a stronger response to the outlier in attention maps compared to image tokens. This prompts us to investigate whether the self-relevance of image tokens and the class token's similarity across all image tokens could serve as potential indicators to identify outliers.", + "bbox": [ + 511, + 508, + 906, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the above assumption, to robustly mitigate outliers, we introduce a Self-adaptive Outlier Mitigator (SOM) to detect and eliminate the outliers automatically. In particular, we use the attention map $Attn$ to obtain the self-attention weights of image tokens, denoted as $Attn_{i,i}$ , and their attention weights to the class tokens, denoted as $Attn_{cls,i}$ , where $i$ denotes the $i$ -th image token and $cls$ denotes the class token. By comparing the difference between $Attn_{i,i}$ and $Attn_{cls,i}$ , we obtain the set of outliers $S$ , since the self-attention weights should be higher than others. This detection process of SOM is formulated as,", + "bbox": [ + 511, + 659, + 908, + 825 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} = \\left\\{i \\mid A t t n _ {i, i} < A t t n _ {c l s, i} \\right\\}, \\quad i \\in [ 1, N ], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 832, + 906, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $S$ denotes the set of outliers.", + "bbox": [ + 511, + 854, + 750, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After recognizing outliers, we propose eliminating them at the feature level rather than modifying the attention map", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20294", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "to prevent potential model collapse. Inspired by previous work [2], it's crucial to ensure the semantic consistency of the feature map, and directly masking outliers at the feature level is not an appropriate approach. Thus, SOM can eliminate outliers through in-place merging, where the detected outliers are replaced with the average value of its 8-neighbours. That is to say, the derived $F_{img}$ in the original attention mechanism will be updated through our SOM as", + "bbox": [ + 89, + 90, + 483, + 212 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {F} _ {i m g} [ u ] = \\left\\{ \\begin{array}{l l} \\frac {1}{8} \\sum_ {v \\in \\mathcal {N} _ {8} (u)} F _ {i m g} [ v ], & u \\in \\mathcal {S} \\\\ F _ {i m g} [ u ], & \\text {o t h e r w s i e} \\end{array} , \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 222, + 483, + 280 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{N}_8(u)$ means the 8-neighbour non-outlier tokens of the $u$ -th point and $\\hat{F}_{img}$ denotes the updated feature. In this way, our SOM initially identifies the outlier and then leverages surrounding token features to replace the detected one. This process can mitigate the adverse influence exerted by the outliers and maintain the feature distribution that is crucial for CLIP's image encoder. By doing so, the purified and superior semantic feature representation is generated.", + "bbox": [ + 89, + 289, + 483, + 410 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Furthermore, our SOM exhibits remarkable versatility in that it can be applied to each individual layer of the image encoder. This allows for the detection of outliers and the generation of purified features at multiple levels. In this manner, our parameter-free SOM establishes an efficient and precise mechanism for detecting and eliminating outliers that propagate across various layers of the image encoder. This mechanism is instrumental in preserving accurate semantic information and feature representations throughout the entire image encoder, thereby enhancing its overall performance and reliability.", + "bbox": [ + 89, + 410, + 483, + 575 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b89618275e0154c0991ecd39b58a4ca196878a512fcf7a2ac266bba76f11ac4a.jpg", + "image_caption": [ + "(a)", + "Figure 3. Comparison of different feature visualization. (a) Noisy feature with outliers. (b) Purified feature with SOM. (c) Semantic feature with SAE. It's seen that the proposed SAE strengthens the semantic regions with higher object activation." + ], + "image_footnote": [], + "bbox": [ + 94, + 589, + 215, + 696 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a0e4641041202a6a68fa1d214a2ba2478afbe719137ad550d2194c14bf9c9d49.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 215, + 588, + 338, + 696 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/afa714bc38ba978a4f3c0f86f18f12e359170e7e2c0e24253b209ff14e8dc4da.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 338, + 588, + 483, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Semantic-aware Attention Enhancer", + "text_level": 1, + "bbox": [ + 89, + 803, + 406, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The comparison of Fig. 3 (a) & (b) indicates that the purified feature $F_{SOM}$ generated by the last SOM effectively ignores the irrelevant regions of outliers, marked as white boxes. However, it's also observed from Fig. 3 (b) & (c) that the object activation of $F_{SOM}$ is not salient enough,", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which indicates that eliminating outliers can't contribute to the attention response intensity.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address this issue, we propose a Semantic-aware Attention Enhancer (SAE) to strengthen the object activation in CLIP's last layer, where the self-self attention mechanism is used to derive better feature relationships,", + "bbox": [ + 511, + 121, + 903, + 181 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA t t n _ {S A E} = \\operatorname {S o f t m a x} \\left(\\lambda \\left(\\frac {Q Q ^ {T}}{\\sqrt {d}} + \\frac {K K ^ {T}}{\\sqrt {d}}\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 188, + 903, + 224 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $Attn_{SAE}$ represents the semantic-aware attention map and $\\lambda$ denotes the logit scale used to sharpen attention scores, set to 0.5. Then, the semantic feature $F_{SAE}$ is obtained through the residual connection with the purified feature $F_{SOM}$ as follows:", + "bbox": [ + 511, + 231, + 905, + 306 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {S A E} = F _ {S O M} + \\operatorname {F F N} \\left(\\operatorname {A t t n} _ {S A E} F _ {S O M}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 316, + 903, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The self-self attention mechanism has shown a strong ability to construct feature relationships [16, 24, 45]. Therefore, applying self-self attention to our purified feature enables the establishment of more precise spatial correlations. These accurate spatial correlations play a crucial role in enabling each point to aggregate the relevant semantic feature effectively. As illustrated in Fig. 3 (b) & (c), SAE significantly augments attention intensity in semantic regions compared to the feature solely processed by SOM.", + "bbox": [ + 511, + 340, + 905, + 477 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Hierarchical Attention Integrator", + "text_level": 1, + "bbox": [ + 511, + 484, + 808, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To further refine the generated semantic feature $F_{SAE}$ , we leverage attention maps from shallow layers to capture discriminative pair-wise feature relationships since shallow layers typically focus on object structures [27]. To this end, we propose a Hierarchical Attention Integrator (HAI), which integrates multi-layer attention maps to refine $F_{SAE}$ .", + "bbox": [ + 511, + 507, + 905, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Instead of using the original attention maps with outliers, HAI integrates attention maps from each transformer layer while excluding the detected outlier weights to avoid information pollution. Firstly, HAI employs the outlier set $S$ to select the non-outlier weights of each attention map, which can be formulated as,", + "bbox": [ + 511, + 598, + 905, + 686 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\overline {{A t t n}} _ {j, i} ^ {l} = \\left\\{ \\begin{array}{l l} A t t n _ {j, i}, & \\text {i f} i \\notin \\mathcal {S} ^ {l} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 695, + 903, + 736 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\overline{Attn}^l$ and $S^l$ denote the filtered attention map and the outlier set at the $l$ -th layer. $j$ and $i$ correspond to the row and column indices of the attention map. Next, HAI integrates $\\overline{Attn}^l$ from each layer, formulated as,", + "bbox": [ + 511, + 744, + 905, + 811 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA = \\sum_ {l = 1} ^ {L - 1} \\overline {{A t t n}} ^ {l}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 653, + 821, + 903, + 862 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $A$ denotes the integrated attention map, and $L$ denotes the total number of transformer layers.", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20295", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ff0a01007c44fdaab18e654939f724cdca857c6b08fd22d99f240e12011e076e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsPub. & YearExtra backboneTraining freewith backgroundV20without backgroundAvg.
V21PC60ObjectPC59StuffADECity
GroupVit [50]CVPR'22XX50.418.727.579.723.415.39.211.129.4
SegCLIP [32]ICML'23XX52.624.726.5------
TCL [7]CVPR'23XX55.030.431.683.233.922.417.124.037.2
DINOiser† [47]ECCV'24DINOX62.132.434.880.935.924.620.031.740.3
SAM-CLIP† [46]ECCV'24SAMX60.629.2----17.1--
PnP-OVSS† [33]CVPR'24BLIP✓--36.251.328.017.914.2--
LaVG† [19]ECCV'24DINO✓62.131.634.282.534.723.215.826.238.8
ProxyCLIP† [25]ECCV'24DINO✓61.335.337.580.339.126.520.238.142.3
CLIP [38]ICML'21X✓18.69.98.149.411.15.73.16.514.1
MaskCLIP [14]ECCV'22X✓38.823.620.674.926.416.49.812.627.9
CLIPSurgery [28]PR'25X✓-29.3---21.9-31.4-
CaR [44]CVPR'24X✓48.630.536.673.739.5-17.7--
GEM [3]CVPR'24X✓46.2---32.615.7---
CLIPtrase [40]ECCV'24X✓50.929.943.681.033.822.816.421.337.5
ClearCLIP* [24]ECCV'24X✓57.032.633.080.935.923.916.730.038.8
SCLIP* [45]ECCV'24X✓59.731.733.581.534.522.716.532.339.1
NACLIP* [16]WACV'25X✓58.932.233.279.735.223.317.435.539.4
SFP (Ours)X✓63.937.237.984.539.926.420.841.144.0
", + "bbox": [ + 114, + 88, + 885, + 366 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparison of our SFP against other approaches across eight segmentation benchmark datasets. The best and second-best results are marked with **bold** and **underline**, respectively. † indicates that the additional backbone settings follow the original paper. * indicates the results are re-evaluated using the official implementation.", + "bbox": [ + 89, + 378, + 906, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After integrating multi-layer attention maps, we deploy a normalization strategy to optimize $A$ instead of simply averaging it. Inspired by previous works [22, 35], we iteratively apply row-wise and column-wise normalization to $A$ for a more object-centric attention distribution, leading to a stronger focus on relevant regions. This process can be formulated as,", + "bbox": [ + 89, + 448, + 483, + 554 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nA _ {H A I} = \\operatorname {n o r m} _ {c} \\left(\\operatorname {n o r m} _ {r} (A)\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 564, + 480, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $A_{HAI}$ denotes the normalized attention map, normc and normr denote the column-wise and row-wise normalization respectively.", + "bbox": [ + 89, + 590, + 482, + 637 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Then, $A_{HAI}$ is multiplied with the semantic feature $F_{SAE}$ to obtain the final semantic feature $F_{vis}$ for segmentation, which is formulated as follows:", + "bbox": [ + 89, + 637, + 483, + 681 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF _ {v i s} = A _ {H A I} F _ {S A E}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 693, + 482, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, we compute the cosine similarity between $F_{vis}$ and the text feature $F_{txt}$ extracted from the CLIP text encoder to predict the segmentation map, as formulated in Eq. 3.", + "bbox": [ + 89, + 719, + 483, + 766 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 89, + 777, + 215, + 795 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 89, + 803, + 297, + 819 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets We conduct comprehensive experiments to verify the effectiveness of our method on diverse benchmark datasets: PASCAL VOC 2012 [15], PASCAL Context [34], and COCO [29]. These datasets are categorized into two settings: 1) with a background class, namely V20, PC59,", + "bbox": [ + 89, + 825, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and Stuff; 2) without the background class, namely V21, PC60, and Object. We also report results on the ADE20K (ADE) [57] and Cityscapes (City) [11] datasets.", + "bbox": [ + 511, + 448, + 906, + 493 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details Similar to existing training-free methods, we adopt CLIP ViT-B/16 with the original pretrained weight [38]. Following previous works [2, 45], we combine standard ImageNet prompts [13] with the class names to construct text descriptions. For a fair comparison with previous works [25, 45], we resize images with a short side of 336 pixels (or 560 pixels for the high-resolution Cityscapes dataset) and apply sliding window inference using a $224 \\times 224$ window with a $112 \\times 112$ stride. Our experiments are conducted using one NVIDIA 3090 GPU.", + "bbox": [ + 511, + 493, + 908, + 645 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 511, + 652, + 877, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our SFP is evaluated against existing SOTA methods, including both training-based and training-free approaches. The post-processing mask refinement techniques, such as PAMR [1] or DenseCRF [23], are not included for a fair comparison. The standard mean intersection over union (mIoU) is used as the metric for evaluation.", + "bbox": [ + 511, + 674, + 906, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Evaluation The main results are summarized in Tab. 1, where all evaluations are conducted using CLIP ViT-B/16 as the backbone to ensure a fair comparison. We compare SFP against multiple methods, including training-based methods, training-free methods with extra backbones, and training-free methods without extra backbones.", + "bbox": [ + 511, + 763, + 905, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Tab. 1, our SFP outperforms existing SOTA methods, regardless of whether an additional backbone is introduced or not. In particular, compared with existing", + "bbox": [ + 511, + 854, + 905, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20296", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/456adbda35fa7d9096eb56eb60121eaea6067ac9850e0c307f84672bba38fb36.jpg", + "image_caption": [ + "Figure 4. Qualitative Results Comparison. We compare our SFP with CLIP [38], ClearCLIP [24], SCLIP [45] and NACLIP [16] without post-processing. It's observed that SFP generates more precise segmentation results." + ], + "image_footnote": [], + "bbox": [ + 133, + 88, + 864, + 385 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "training-free methods without extra backbones [16, 24, 45] or cluster methods [53-55], our SFP achieves an average gain of $4.6\\%$ mIoU over previous SOTA NACLIP [16]. Specifically, On the V20 and V21 dataset [15], SFP surpasses NACLIP [16] $4.8\\%$ and $5\\%$ mIoU, respectively. These improvements indicate that our SFP can generate more purified feature maps for more precise segmentation. Beyond that, SFP outperforms SOTA training-free methods with extra backbones, e.g., ProxyCLIP [25], with a gain of $3.0\\%$ mIoU on the Cityscape [11] dataset and an average gain of $1.7\\%$ mIoU on eight datasets. Our SFP can even beat DINOiser [47], which requires both training and an extra backbone, with an average improvement of $3.7\\%$ mIoU. These results highlight our SFP boosts CLIP's internal potential and reveal that feature purification does matter in training-free OVSS, where more purified features can predict better segmentation masks.", + "bbox": [ + 88, + 450, + 485, + 708 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Evaluation In Fig. 4, we compare the visualization results of various training-free methods without extra backbones, including CLIP [38], ClearCLIP [24], SCLIP [45], NACLIP [16] and our SFP. In particular, the segmentation maps obtained by vanilla CLIP [38] exhibit significant noises, unveiling its limitations in spatial localization. Besides, in complex scenarios, like 'sofa' in the second row, SCLIP [45] and NACLIP [16] fail to capture the complete segmentation mask. In contrast, our SFP can predict more accurate and complete segmentation maps. These results indicate that our SFP has a good ability to predict comprehensive segmentation masks.", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 450, + 666, + 467 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comprehensive ablation studies are performed to analyze the effectiveness of our proposed mechanisms. Unless otherwise specified, we utilize CLIP with ViT-B/16 backbone and evaluate on V21 [15] and ADE [57] datasets for background and non-background settings, respectively.", + "bbox": [ + 511, + 473, + 906, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect on each proposed module We first conduct several experiments to validate the effectiveness of our proposed modules of our SFP: SOM, SAE, and HAI by adding each module step by step, where ClearCLIP [24] is used as the baseline. As shown in Tab. 2, our SOM achieves a gain of $3.1\\%$ and $1.5\\%$ mIoU on V21 and ADE datasets, respectively. Further, SAE enhances attention intensity in relevant regions, leading to an average gain of $1.7\\%$ mIoU. Finally, HAI helps our SFP reach $63.9\\%$ and $20.8\\%$ mIoU on V21 and ADE datasets, which set the new SOTA performance. Due to space constraints, more ablation studies about SOM and SAE are summarized in the supplementary material.", + "bbox": [ + 511, + 547, + 908, + 731 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/aac55b2e9d4d1c012b024fe53d12b1650d8fa30a3a6e0046227b762cfceb1075.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SOMSAEHAIV21ADEAvg.
057.016.736.9
1✓60.118.239.2
2✓✓62.319.540.9
3✓✓✓63.920.842.4
", + "bbox": [ + 557, + 739, + 867, + 813 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Ablation results on each proposed module to validate the impact on performance across V21 [15] and ADE [57] datasets.", + "bbox": [ + 511, + 827, + 906, + 856 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "More visualization on the ability of SOM to suppress outlier propagation We visualize purified feature maps", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20297", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a306f95aad08c2c2619221750b1d1304fdc64beae124dc2ea0e33db12deccce1.jpg", + "image_caption": [ + "Figure 5. Visualization of the feature purification across CLIP's image encoder layers. (a) Feature maps without SOM purification. (b) Feature maps purified by SOM." + ], + "image_footnote": [], + "bbox": [ + 133, + 88, + 864, + 229 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c6be23c6829af8ab2b9909e984b217f66de15edb42067004b1360f8debf45aa3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Al=l-1l=lo1581011
V2163.963.463.062.662.5
ADE20.820.520.319.919.7
Avg.42.442.041.741.341.1
", + "bbox": [ + 133, + 294, + 444, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "across CLIP's layers, from the 2nd layer to the 9th layer, which is shown in Fig. 5. It's found that the outliers gradually affect the aggregated features, especially from layer 6 to layer 9 in Fig. 5 (a). Then, with the help of our SOM, the influence of outliers is mitigated layer by layer, like in Fig. 5 (b). Thus, our SOM can suppress the propagated outlier features and gradually purify the overall semantic features.", + "bbox": [ + 89, + 443, + 483, + 549 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect on hierarchical attention integrator A series of ablation studies is conducted to investigate the optimal way to leverage attention maps from different layers. We omit HAI from our SFP to establish the baseline model for this experiment. As summarized in Tab. 3, we take $l_{0}$ to represent the starting layer. When $l_{0} = 1$ , our HAI performs best with all attention maps selected. Besides, increasing $l_{0}$ from 1 to 11 decreases the segmentation performance, which demonstrates the effectiveness of the shallow attention maps for semantic feature refinement.", + "bbox": [ + 89, + 549, + 483, + 700 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/06423330e4a9bfd769da160ac47aab784abbd036924c28201adba36206085470.jpg", + "table_caption": [ + "Table 3. Ablation results on different selection strategies in HAI, where $A_{l = l_0}^{l = L - 1}$ represents the integrated attention map from $l_0$ to the $L - 1$ layer. '1,5,8,10,11' indicates the value of $l_0$ ." + ], + "table_footnote": [], + "table_body": "
BackboneMethodsV21ADEAvg.
ViT-B/32SCLIP [45]50.614.832.7
LaVG [19]54.815.535.2
ProxyCLIP [25]57.916.737.3
NACLIP [16]51.114.933.0
SFP60.117.138.6
ViT-L/14SCLIP [45]44.410.927.7
LaVG [19]52.117.334.7
ProxyCLIP [25]60.622.641.6
NACLIP [16]52.217.334.8
SFP64.722.143.4
", + "bbox": [ + 132, + 710, + 444, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Comparison with other SOTA methods across different CLIP backbones.", + "bbox": [ + 89, + 868, + 482, + 895 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect on different CLIP backbones To validate the robustness, we further evaluate our SFP with other methods using different CLIP backbones, such as ViT-B/32 and ViT-L/14. As summarized in Tab. 4, some methods, like SCLIP [45] and LaVG [19], suffer from performance degradation when applied to the larger backbone. In contrast, SFP maintains superior performance across different backbones and achieves an improvement of $1.3\\%$ and $1.8\\%$ mIoU over ProxyCLIP [25] with ViT-B/32 and ViT-L/14 backbone, respectively. These results indicate that SFP can robustly handle outliers derived from different backbones and generate purified features for precise segmentation.", + "bbox": [ + 511, + 297, + 906, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 503, + 633, + 520 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present a Self-adaptive Feature Purifier framework (SFP) to mitigate outlier propagation for training-free open-vocabulary semantic segmentation. Specifically, we propose SOM deployed at each layer of the image encoder to identify and mitigate outliers by comparing attention weights between image and class tokens, which generates purified features for propagation. Next, to recover the semantic information affected by outliers, we introduce SAE to strengthen object activation in the purified feature. Beyond that, we design HAI to utilize multilayer attention maps for refined object-centric representations. Extensive experiments show that SFP achieves superior segmentation performance compared to existing SOTA methods on various datasets. Our SFP shows that feature purification does matter for OVSS, and cleaner features can predict more complete and accurate segmentation masks.", + "bbox": [ + 511, + 534, + 906, + 775 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported by the National Natural Science Foundation of China (No. 62471405, 62331003, 62301451,62301613), Basic Research Program of Jiangsu (BK20241814), Suzhou Basic Research Program (SYG202316) and XJTLU REF-22-01-010 and RDF-22-02-066, XJTLU AI University Research Centre, Jiangsu Province Engineering Research Centre of Data Science and Cognitive Computation at XJTLU.", + "bbox": [ + 511, + 779, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20298", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Nikita Araslanov and Stefan Roth. Single-stage semantic segmentation from image labels. In CVPR, pages 4253-4262, 2020. 6", + "[2] Sule Bai, Yong Liu, Yifei Han, Haoji Zhang, and Yansong Tang. Self-calibrated clip for training-free open-vocabulary segmentation. arXiv preprint arXiv:2411.15869, 2024. 2, 5, 6", + "[3] Walid Bousselham, Felix Petersen, Vittorio Ferrari, and Hilde Kuehne. Grounding everything: Emerging localization properties in vision-language transformers. In CVPR, pages 3828-3837, 2024. 6", + "[4] Maxime Bucher, Tuan-Hung Vu, Matthieu Cord, and Patrick Pérez. Zero-shot semantic segmentation. In NeurlPS, pages 468-479, 2019. 1", + "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. 3", + "[6] Niccolo Cavagnero, Gabriele Rosi, Claudia Cuttano, Francesca Pistilli, Marco Ciccone, Giuseppe Avera, and Fabio Cermelli. Pem: Prototype-based efficient maskformer for image segmentation. In CVPR, pages 15804-15813, 2024. 3", + "[7] Junbum Cha, Jonghwan Mun, and Byungseok Roh. Learning to generate text-grounded mask for open-world semantic segmentation from only image-text pairs. In CVPR, pages 11165-11174, 2023. 6", + "[8] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, pages 1290–1299, 2022. 3", + "[9] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, pages 2818-2829, 2023. 2", + "[10] Seokju Cho, Heeseong Shin, Sunghwan Hong, Anurag Arnab, Paul Hongsuck Seo, and Seungryong Kim. Cat-seg: Cost aggregation for open-vocabulary semantic segmentation. In CVPR, pages 4113-4123, 2024. 3", + "[11] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, pages 3213-3223, 2016. 6, 7", + "[12] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. In ICLR, pages 1-11, 2024. 1, 2, 4", + "[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, pages 248-255, 2009. 4, 6", + "[14] Zheng Ding, Jieke Wang, and Zhuowen Tu. Open-vocabulary panoptic segmentation with maskclip. arXiv preprint arXiv:2208.08984, 2022. 2, 3, 6", + "[15] Mark Everingham, SM Ali Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "pascal visual object classes challenge: A retrospective. IJCV, 111:98-136, 2015. 6, 7", + "[16] Sina Hajimiri, Ismail Ben Ayed, and Jose Dolz. Pay attention to your neighbours: Training-free open-vocabulary semantic segmentation. In WACV, pages 5061-5071, 2025. 2, 5, 6, 7, 8", + "[17] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916, 2021. 2", + "[18] Siyu Jiao, Yunchao Wei, Yaowei Wang, Yao Zhao, and Humphrey Shi. Learning mask-aware clip representations for zero-shot segmentation. In NeurIPS, pages 35631-35653, 2022. 3", + "[19] Dahiyun Kang and Minsu Cho. In defense of lazy visual grounding for open-vocabulary semantic segmentation. In ECCV, pages 143-164, 2024. 3, 6, 8", + "[20] Laurynas Karazija, Iro Laina, Andrea Vedaldi, and Christian Rupprecht. Diffusion models for zero-shot open-vocabulary segmentation. In ECCV, pages 299-317, 2024. 3", + "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, pages 4015-4026, 2023. 3", + "[22] Philip A Knight. The sinkhorn-knopp algorithm: convergence and applications. SIAM Journal on Matrix Analysis and Applications, 30(1):261-275, 2008. 6", + "[23] Philipp Kähenbuhl and Vladlen Koltun. Efficient inference in fully connected crfs with gaussian edge potentials. In NeurIPS, pages 109-117, 2011. 6", + "[24] Mengcheng Lan, Chaofeng Chen, Yiping Ke, Xinjiang Wang, Litong Feng, and Wayne Zhang. Clearclip: Decomposing clip representations for dense vision-language inference. In ECCV, pages 143-160, 2024. 2, 3, 5, 6, 7", + "[25] Mengcheng Lan, Chaofeng Chen, Yiping Ke, Xinjiang Wang, Litong Feng, and Wayne Zhang. Proxyclip: Proxy attention improves clip for open-vocabulary segmentation. In ECCV, pages 70-88, 2024. 3, 6, 7, 8", + "[26] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900, 2022. 2", + "[27] Yunheng Li, Zhongyu Li, Quansheng Zeng, Qibin Hou, and Ming-Ming Cheng. Cascade-clip: Cascaded vision-language embeddings alignment for zero-shot semantic segmentation. In ICML, pages 28243–28258, 2024. 2, 5", + "[28] Yi Li, Hualiang Wang, Yiqun Duan, Jiheng Zhang, and Xiaomeng Li. A closer look at the explainability of contrastive language-image pre-training. Pattern Recognition, 162:111409, 2025. 6", + "[29] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 6", + "[30] Yong Liu, Sule Bai, Guanbin Li, Yitong Wang, and Yansong Tang. Open-vocabulary segmentation with semantic-assisted calibration. In CVPR, pages 3491-3500, 2024. 1, 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20299", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, pages 3431-3440, 2015. 3", + "[32] Huaishao Luo, Junwei Bao, Youzheng Wu, Xiaodong He, and Tianrui Li. Segclip: Patch aggregation with learnable centers for open-vocabulary semantic segmentation. In ICML, pages 23033-23044, 2023. 6", + "[33] Jiayun Luo, Siddhesh Khandelwal, Leonid Sigal, and Boyang Li. Emergent open-vocabulary semantic segmentation from off-the-shelf vision-language models. In CVPR, pages 4029-4040, 2024. 6", + "[34] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The role of context for object detection and semantic segmentation in the wild. In CVPR, pages 891–898, 2014. 6", + "[35] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6", + "[36] Zelin Peng, Zhengqin Xu, Zhilin Zeng, Yu Huang, Yaoming Wang, and Wei Shen. Parameter-efficient fine-tuning in hyperspherical space for open-vocabulary semantic segmentation. In CVPR, pages 15009-15020, 2025. 3", + "[37] Zelin Peng, Zhengqin Xu, Zhilin Zeng, Changsong Wen, Yu Huang, Menglin Yang, Feilong Tang, and Wei Shen. Understanding fine-tuning clip for open-vocabulary semantic segmentation in hyperbolic space. In CVPR, pages 4562-4572, 2025. 3", + "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 1, 2, 3, 4, 6, 7", + "[39] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, pages 25278–25294, 2022. 1, 2", + "[40] Tong Shao, Zhuotao Tian, Hang Zhao, and Jingyong Su. Explore the potential of clip for training-free open vocabulary semantic segmentation. In ECCV, pages 139-156, 2024. 1, 2, 3, 6", + "[41] Gyungin Shin, Weidi Xie, and Samuel Albanie. Reco: Retrieve and co-segment for zero-shot transfer. In NeurIPS, pages 33754-33767, 2022. 3", + "[42] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, pages 2256-2265, 2015. 3", + "[43] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In ICCV, pages 7262-7272, 2021. 1", + "[44] Shuyang Sun, Runjia Li, Philip Torr, Xiuye Gu, and Siyang Li. Clip as rn: Segment countless visual concepts without training endeavor. In CVPR, pages 13171-13182, 2024. 6" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[45] Feng Wang, Jieru Mei, and Alan Yuille. Sclip: Rethinking self-attention for dense vision-language inference. In ECCV, pages 315–332, 2024. 1, 2, 3, 5, 6, 7, 8", + "[46] Haoxiang Wang, Pavan Kumar Anasosalu Vasu, Fartash Faghri, Raviteja Vemulapalli, Mehrdad Farajtabar, Sachin Mehta, Mohammad Rastegari, Oncel Tuzel, and Hadi Pouransari. Sam-clip: Merging vision foundation models towards semantic and spatial understanding. In CVPR, pages 3635–3647, 2024. 6", + "[47] Monika Wysoczanska, Oriane Simeoni, Michael Ramamonjisoa, Andrei Bursuc, Tomasz Trzciński, and Patrick Pérez. Clip-dinoiser: Teaching clip a few dino tricks. In ECCV, pages 320-337, 2024. 6, 7", + "[48] Bin Xie, Jiale Cao, Jin Xie, Fahad Shahbaz Khan, and Yanwei Pang. Sed: A simple encoder-decoder for open-vocabulary semantic segmentation. In CVPR, pages 3426-3436, 2024. 3", + "[49] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. In NeurIPS, pages 12077-12090, 2021. 3", + "[50] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. Groupvit: Semantic segmentation emerges from text supervision. In CVPR, pages 18134-18144, 2022. 6", + "[51] Jilan Xu, Junlin Hou, Yuejie Zhang, Rui Feng, Yi Wang, Yu Qiao, and Weidi Xie. Learning open-vocabulary semantic segmentation models from natural language supervision. In CVPR, pages 2935-2944, 2023. 3", + "[52] Mengde Xu, Zheng Zhang, Fangyun Wei, Han Hu, and Xiang Bai. Side adapter network for open-vocabulary semantic segmentation. In CVPR, pages 2945-2954, 2023. 3", + "[53] Xihong Yang, Xiaochang Hu, Sihang Zhou, Xinwang Liu, and En Zhu. Interpolation-based contrastive learning for few-label semi-supervised learning. TNNLS, 35(2):2054-2065, 2022. 7", + "[54] Xihong Yang, Jin Jiaqi, Siwei Wang, Ke Liang, Yue Liu, Yi Wen, Suyuan Liu, Sihang Zhou, Xinwang Liu, and En Zhu. Dealmvc: Dual contrastive calibration for multi-view clustering. In ACMMM, pages 337-346, 2023.", + "[55] Xihong Yang, Siwei Wang, Fangdi Wang, Jiaqi Jin, Suyuan Liu, Yue Liu, En Zhu, Xinwang Liu, and Yueming Jin. Automatically identify and rectify: Robust deep contrastive multiview clustering in noisy scenarios. In ICML. PMLR, 2025. 7", + "[56] Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang-Chieh Chen. Convolutions die hard: Open-vocabulary segmentation with single frozen convolutional clip. In NeurIPS, pages 32215-32234, 2023. 3", + "[57] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ADE20K dataset. In CVPR, pages 5122-5130, 2017. 6, 7", + "[58] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In ECCV, pages 696-712, 2022. 1", + "[59] Ziqin Zhou, Yinjie Lei, Bowen Zhang, Lingqiao Liu, and Yifan Liu. Zegclip: Towards adapting clip for zero-shot semantic segmentation. In CVPR, pages 11175-11185, 2023." + ], + "bbox": [ + 514, + 92, + 903, + 887 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20300", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Feature Purification Matters_ Suppressing Outlier Propagation for Training-Free Open-Vocabulary Semantic Segmentation/4f979197-bed4-425a-a1d2-c2b5ca2160b9_model.json b/2025/Feature Purification Matters_ Suppressing Outlier Propagation for Training-Free Open-Vocabulary Semantic Segmentation/4f979197-bed4-425a-a1d2-c2b5ca2160b9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a8f083a138fa232b0c49d74d62c175a850e65888 --- /dev/null +++ b/2025/Feature Purification Matters_ Suppressing Outlier Propagation for Training-Free Open-Vocabulary Semantic Segmentation/4f979197-bed4-425a-a1d2-c2b5ca2160b9_model.json @@ -0,0 +1,2398 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.164, + 0.131, + 0.836, + 0.177 + ], + "angle": 0, + "content": "Feature Purification Matters: Suppressing Outlier Propagation for Training-Free Open-Vocabulary Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.148, + 0.203, + 0.844, + 0.291 + ], + "angle": 0, + "content": "Shuo Jin\\(^{1,2}\\) Siyue Yu\\(^{1*}\\) Bingfeng Zhang\\(^{3}\\) Mingjie Sun\\(^{4}\\) Yi Dong\\(^{2}\\) Jimin Xiao\\(^{1}\\) \n\\(^{1}\\)Xi'an Jiaotong-Liverpool University \\(^{2}\\)University of Liverpool \n\\(^{3}\\)China University of Petroleum (East China) \\(^{4}\\)Soochow University \n{shuo.jin, yi.dong}@liverpool.ac.uk, {siyue.yu02, jimin.xiao}@xjtlu.edu.cn \nbingfeng.zhang@upc.edu.cn, mjsun@suda.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.359, + 0.485, + 0.782 + ], + "angle": 0, + "content": "Training-free open-vocabulary semantic segmentation has advanced with vision-language models like CLIP, which exhibit strong zero-shot abilities. However, CLIP's attention mechanism often wrongly emphasises specific image tokens, namely outliers, which results in irrelevant over-activation. Existing approaches struggle with these outliers that arise in intermediate layers and propagate through the model, ultimately degrading spatial perception. In this paper, we propose a Self-adaptive Feature Purifier framework (SFP) to suppress propagated outliers and enhance semantic representations for open-vocabulary semantic segmentation. Specifically, based on an in-depth analysis of attention responses between image and class tokens, we design a self-adaptive outlier mitigator to detect and mitigate outliers at each layer for propagated feature purification. In addition, we introduce a semantic-aware attention enhancer to augment attention intensity in semantically relevant regions, which strengthens the purified feature to focus on objects. Further, we introduce a hierarchical attention integrator to aggregate multi-layer attention maps to refine spatially coherent feature representations for final segmentation. Our proposed SFP enables robust outlier suppression and object-centric feature representation, leading to a more precise segmentation. Extensive experiments show that our method achieves state-of-the-art performance and surpasses existing methods by an average of \\(4.6\\%\\) mIoU on eight segmentation benchmarks. The code is released at: https://github.com/Kimsure/SFP." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.809, + 0.222, + 0.825 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.834, + 0.484, + 0.88 + ], + "angle": 0, + "content": "Open-vocabulary semantic segmentation (OVSS) [4, 58] seeks to partition an image into distinct regions and assign pixel-level labels to arbitrary semantic categories. Un" + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.328, + 0.616, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.561, + 0.402, + 0.578, + 0.412 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.328, + 0.712, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.657, + 0.402, + 0.673, + 0.412 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.328, + 0.808, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.752, + 0.402, + 0.768, + 0.413 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.809, + 0.328, + 0.904, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.847, + 0.402, + 0.864, + 0.413 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.412, + 0.616, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.561, + 0.488, + 0.578, + 0.499 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.412, + 0.712, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.657, + 0.488, + 0.673, + 0.499 + ], + "angle": 0, + "content": "(f)" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.412, + 0.808, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.751, + 0.488, + 0.768, + 0.499 + ], + "angle": 0, + "content": "(g)" + }, + { + "type": "image", + "bbox": [ + 0.809, + 0.412, + 0.904, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.847, + 0.488, + 0.864, + 0.499 + ], + "angle": 0, + "content": "(h)" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.513, + 0.907, + 0.596 + ], + "angle": 0, + "content": "Figure 1. Visualization of attention and feature maps. (a) Attention map of the selected target object token (marked as \\(\\star\\)). (b) Attention map of the selected background token (marked as \\(\\star\\)). (c) Attention map of the class token. (d) Diagonal weights of the attention map of each image token. (e) Detected outliers. (f) - (h) output feature map of CLIP [38], SCLIP [45], and our SFP." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.627, + 0.907, + 0.807 + ], + "angle": 0, + "content": "like traditional semantic segmentation, OVSS must handle an open set of classes, making it more challenging. Vision-language models (VLM), such as CLIP [38], demonstrate remarkable zero-shot performance by leveraging large-scale image-text pairs [39]. Thus, existing approaches can be categorized into two pipelines: fine-tuning CLIP or training-free CLIP. Although the fine-tuning pipeline yields superior performance, it risks overfitting and compromises generalization [30, 43]. In contrast, the training-free pipeline preserves CLIP's original generalization capability without additional training, which is more convenient and efficient." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Recent training-free methods [12, 40] have demonstrated that in the original CLIP's pair-wise attention maps, there exist the same highly correlated image tokens for all input tokens, such as those marked with the white box in Fig. 1 (a) - (b). In such a situation, regardless of whether it is the target image token (Fig. 1 (a)), the background image to-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.372, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author: siyue.yu02@xjtlu.edu.cn" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20291" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "ken (Fig. 1 (b)), or the class token (Fig. 1 (c)), the pair-wise token relationships within the attention map are all over-activated for these identical tokens, i.e., outliers. Consequently, when aggregated with outliers, the derived semantic feature representation is easily polluted to overemphasize these outliers and the interpretation of the target region is impeded, e.g., the feature map of 'car' in Fig 1 (f) focuses on the outliers instead of the object itself." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.215, + 0.485, + 0.502 + ], + "angle": 0, + "content": "To mitigate outliers and ensure more accurate semantic feature representation, many training-free approaches have presented diverse modifications to the self-attention mechanism in the last layer [14, 16, 24, 45]. While these strategies effectively restrict CLIP's attention to ignore outliers, some remain unfiltered, as shown in Fig. 1 (g). Moreover, some approaches [2, 12] rely on empirical parameter tuning to detect and remove a fixed number of outliers, which is less robust to different backbones and images. Beyond these optimization approaches, several studies [12, 40] further reveal that outliers predominantly appear in the intermediate layers. In this way, once outliers emerge, they will propagate through subsequent layers, continuously affecting semantic feature representations. Yet, previous approaches fail to analyze and address outliers in the intermediate layers so that outliers are easily propagated to pollute the final semantic feature. Thus, our objective is to adaptively eliminate outliers across multiple layers to achieve more robust semantic feature representations without training." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.506, + 0.484, + 0.807 + ], + "angle": 0, + "content": "Building on these insights, we propose a Self-adaptive Feature Purifier framework (SFP) to resolve outliers adaptively and enhance semantic representations for CLIP-based training-free OVSS. Firstly, we design a Self-adaptive Outlier Mitigator (SOM) to identify and mitigate outliers across multiple layers. Tokens are expected to focus primarily on their relevant regions, especially themselves, rather than irrelevant regions in the attention map. However, as shown in Fig. 1 (d), we observe that the self-response of image tokens in the attention map is not always prominent, especially outliers. The comparison of Fig. 1 (c) & (d) reveals that outliers exhibit a stronger response to the class token than their self-response. Hence, our SOM can identify outliers by simply computing the difference between the self-response values of the image tokens in the attention map (i.e., the diagonal weights of the attention map) and their attention values with the class token. The detected outliers are shown in Fig. 1 (e). In this way, our parameter-free SOM enables adaptive detection across various layers to mitigate the propagating influence of outliers without manual tuning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.482, + 0.902 + ], + "angle": 0, + "content": "Nevertheless, outliers not only cause tokens to overemphasize them but also impede tokens' responses to semantic regions in the attention map. Therefore, merely removing the outliers can't directly increase the attention weights of the related regions. To enhance the tokens' attention responses to semantic regions, we propose a Semantic-aware" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.318 + ], + "angle": 0, + "content": "Attention Enhancer (SAE). SAE incorporates the self-self attention mechanism in CLIP's last layer and applies it to the purified feature derived from SOM at the penultimate layer, enabling a better focus on semantic regions. In addition, shallow attention maps typically contain structured object information and focus on relevant semantic regions [27, 30]. To further enhance the precise spatial perception of the final feature map, we propose a Hierarchical Attention Integrator (HAI) as an assistant to the SAE, where the attention maps in previous layers are firstly processed to mitigate the influence of the outliers and then integrated to refine the final semantic feature for prediction. The refined semantic feature for the final prediction is shown in Fig. 1 (h). Our SFP can generate a more purified and complete semantic feature for segmentation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.319, + 0.905, + 0.394 + ], + "angle": 0, + "content": "We conduct comprehensive experiments to evaluate the proposed SFP. Our method outperforms existing methods and achieves state-of-the-art (SOTA) performance across eight evaluation datasets. In summary, our contributions are concluded as follows:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.399, + 0.904, + 0.459 + ], + "angle": 0, + "content": "- We present a Self-adaptive Feature Purifier framework (SFP) for training-free OVSS, which can derive a more accurate semantic feature to strengthen the final open-vocabulary semantic segmentation." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.46, + 0.905, + 0.535 + ], + "angle": 0, + "content": "- We propose a simple yet effective mechanism, Self-adaptive Outlier Mitigator (SOM), which adaptively detects and mitigates outliers across CLIP's diverse layers so that the influence of the propagated influence of outliers can be suppressed." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.535, + 0.905, + 0.609 + ], + "angle": 0, + "content": "- We introduce a Semantic-aware Attention Enhancer (SAE) and a Hierarchical Attention Integrator (HAI) to further refine the semantic feature relationships on relevant regions, aggregating the more relevant semantic feature for final prediction." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.61, + 0.905, + 0.671 + ], + "angle": 0, + "content": "- Comprehensive experiments demonstrate that SFP derives the best performance under a fair comparison e.g., SFP achieves a gain of \\(5.6\\%\\) mIoU on Cityscapes and an average gain of \\(4.6\\%\\) mIoU across eight datasets." + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.399, + 0.905, + 0.671 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.687, + 0.655, + 0.702 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.712, + 0.833, + 0.729 + ], + "angle": 0, + "content": "2.1. pre-trained Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Pre-trained vision-language models [17, 26, 38] have drawn significant attention for their ability to bridge visual and textual modalities. CLIP [38] has emerged as a highly successful model trained on a large-scale dataset of image-text pairs using contrastive learning. OpenCLIP [9] builds upon CLIP by leveraging public datasets such as LAION [39]. Despite their strong zero-shot capabilities, these models are pretrained at the image level for classification. This presents a notable limitation: the attention maps of VLMs tend to overemphasize outliers rather than semantic image tokens, leading to insufficient capture of spatial details. This draw" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20292" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "back hinders the performance in dense prediction tasks like semantic segmentation, which demands precise pixel-level understanding. Our goal is to adaptively detect and mitigate the impact of these outliers and enhance the semantic feature representation for dense prediction tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.181, + 0.37, + 0.198 + ], + "angle": 0, + "content": "2.2. Open-vocabulary Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.204, + 0.484, + 0.324 + ], + "angle": 0, + "content": "Semantic segmentation involves the pixel-wise classification of an image. Compared with the traditional segmentation methods [6, 8, 31, 49] that are trained and evaluated on the same fixed set of seen classes. OVSS methods [36, 59] aim to recognize and segment unseen objects during inference by harnessing the zero-shot ability of VLMs. Existing works can be broadly divided into two categories: training-based and training-free." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.326, + 0.484, + 0.583 + ], + "angle": 0, + "content": "Training-based Methods Training-based methods [18, 37, 56] typically require fine-tuning on a fixed set of categories from a pixel-level annotated dataset. Some approaches follow a two-stage pipeline, where the first stage extracts the mask proposals, and the second stage assigns semantic labels to them. For instance, OVSeg [51] first trains class-agnostic mask proposals using the query-based framework Mask2Former [8] and then fine-tunes CLIP [38] to classify the cropped and masked images. Other approaches adopt a single-stage pipeline. SAN [52] introduces a side adapter to adapt CLIP [38] for both classification and segmentation. SED [48] presents a simple encoder-decoder architecture with category early rejection for fast inference. CATSeg [10] constructs a pixel-level cost map for segmentation. However, these training-based methods are prone to overfitting the training dataset, which limits the generalization ability due to fine-tuning CLIP." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.585, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Training-free Methods Unlike training-based methods, training-free methods [20, 41] aim to directly adapt VLMs for OVSS without extra training. Several approaches leverage the vision foundation models [5, 21, 42] to improve CLIP's spatial coherence. LaVG [19] uses DINO [5] for panoptic cuts, which iteratively discovers object masks until all image pixels are covered. ProxyCLIP [25] adjusts attention weights based on affinities learned from DINO's feature correspondence. However, these methods fail to exploit CLIP's inherent potential. Other efforts aim to modify CLIP's final layer to address its poor spatial consistency. For example, MaskCLIP [14] only utilizes the value feature of the final layer for prediction. CLIPtrase [40] integrates self-self attention with clustering for post-processing. SCLIP [45] and ClearCLIP [24] both modify the last layer self-attention mechanism to better represent the local information. Nevertheless, few training-free methods address propagated outliers, which act as noisy inputs across multiple layers and degrade spatial perception. In contrast, our approach enables adaptive detection and mitigation of propagated outliers among various layers, which improves" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.107 + ], + "angle": 0, + "content": "CLIP's semantic awareness and boosts its internal potential." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.118, + 0.605, + 0.133 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.143, + 0.625, + 0.158 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.165, + 0.907, + 0.285 + ], + "angle": 0, + "content": "Fig. 2 depicts the whole framework of SFP, including four main modules: a frozen CLIP backbone to encode the input image and text descriptions, a SOM to detect and eliminate outliers, an SAE to augment the attention intensity on related semantic regions, and an HAI that leverages intermediate-layer attention maps to assist SAE for enhanced semantic feature representation. The complete inference process is as follows:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.286, + 0.905, + 0.361 + ], + "angle": 0, + "content": "1) First, the image and text descriptions are input to the CLIP encoder for the image feature and the text feature. Among the CLIP image encoder, our SOM is added after each transformer layer to mitigate the outliers and generate the purified feature for propagation." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.362, + 0.905, + 0.436 + ], + "angle": 0, + "content": "2) Then, the derived final feature map from the last SOM, \\( F_{SOM} \\), is input to our SAE, which replaces the final transformer layer. SAE emphasizes the attention values of relevant semantic regions to generate the strengthened semantic feature \\( F_{SAE} \\)." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.437, + 0.905, + 0.481 + ], + "angle": 0, + "content": "3) After that, our HAI extracts the multi-layer attention maps from the image encoder to further refine \\( F_{SAE} \\) for the final semantic feature \\( F_{vis} \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.482, + 0.905, + 0.527 + ], + "angle": 0, + "content": "4) Finally, the segmentation result is obtained by calculating the cosine similarity of the final semantic feature \\( F_{vis} \\) and text feature \\( F_{txt} \\)." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.286, + 0.905, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.535, + 0.643, + 0.551 + ], + "angle": 0, + "content": "3.2. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.556, + 0.907, + 0.721 + ], + "angle": 0, + "content": "In general, the training-free OVSS approach requires CLIP [38], which encompasses an image encoder and a text encoder to acquire multi-model features. Subsequently, these features are aligned via linear projection layers. In detail, the input image is first partitioned into a token sequence \\( X_{in} = [x_{cls}, x_1, \\dots, x_N] \\in R^{(N + 1) \\times d} \\), where \\( x_i \\) denotes the image token embedding, \\( x_{cls} \\) denotes the class token embedding, \\( N \\) is the image token sequence length, and \\( d \\) is the dimension of token embeddings, respectively. Then, these tokens are input to the image encoder, which contains a stack of vision transformer layers." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.722, + 0.906, + 0.828 + ], + "angle": 0, + "content": "Each transformer layer consists of a multi-head self-attention (MHSA) mechanism and a feed-forward network (FFN) layer to extract the image feature. Specifically, in each MHSA, the input sequence \\( X_{in} \\) is mapped to query, key, and value embeddings, \\( Q, K, V \\in R^{(N + 1)\\times d} \\). Afterwards, \\( Q \\) and \\( K \\) are calculated to obtain the attention map \\( Attn \\in R^{(N + 1)\\times (N + 1)} \\), which can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.62, + 0.833, + 0.905, + 0.867 + ], + "angle": 0, + "content": "\\[\nA t t n = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.902 + ], + "angle": 0, + "content": "where \\(T\\) means matrix transposition. After that, the attention map is multiplied \\(V\\) to aggregate the global infor" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20293" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.089, + 0.881, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.412, + 0.908, + 0.483 + ], + "angle": 0, + "content": "Figure 2. Framework of our SFP. It contains four main modules: a frozen CLIP backbone for feature extraction, a Self-adaptive Outlier Mitigator (SOM) at each image encoder layer for outlier detection and elimination, a Semantic-aware Attention Enhancer (SAE) at the last layer to emphasise attention values of relevant semantic regions and a Hierarchical Attention Integrator (HAI) for object-centric feature representation. The segmentation result is generated by calculating the cosine similarity between the optimized semantic feature \\( F_{vis} \\) and the text feature \\( F_{txt} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.509, + 0.483, + 0.541 + ], + "angle": 0, + "content": "mation, and an FFN is used to obtain the image feature \\(F_{img} \\in R^{(N + 1) \\times d}\\), formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.208, + 0.549, + 0.483, + 0.567 + ], + "angle": 0, + "content": "\\[\nF _ {i m g} = \\operatorname {F F N} (A t t n V). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.573, + 0.483, + 0.634 + ], + "angle": 0, + "content": "Since the class embedding \\( x_{cls} \\) is not used to predict the segmentation result, we omit it and reshape \\( F_{img} \\) to obtain the visual feature \\( F_{vis} \\in R^{h \\times w \\times d} \\), where \\( h \\) and \\( w \\) denote the height and width of the feature map." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.634, + 0.483, + 0.725 + ], + "angle": 0, + "content": "Meanwhile, the textual description is derived from the standard ImageNet [13] prompts and transformed to the text feature \\( F_{txt} \\in R^{c \\times d} \\) with CLIP's text encoder, where \\( c \\) represents the number of classes. By calculating the cosine similarity between \\( F_{vis} \\) and \\( F_{txt} \\), the segmentation map \\( Seg \\in R^{H \\times W} \\) can be generated like:" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.733, + 0.483, + 0.758 + ], + "angle": 0, + "content": "\\[\nS e g = U (\\underset {c} {\\arg \\max } \\cos (F _ {v i s}, F _ {t x t})), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.827 + ], + "angle": 0, + "content": "where \\(\\cos (\\cdot)\\) means cosine similarity calculation and \\(U(\\cdot)\\) denotes the upsampling operation to match the original input image size. \\(H\\) and \\(W\\) denote the height and width of the input image, respectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.834, + 0.367, + 0.851 + ], + "angle": 0, + "content": "3.3. Self-adaptive Outlier Mitigator" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "CLIP [38] is pre-trained at the image level for classification, which tends to concentrate on the most discriminative feature that is important for classification. In this way, all input" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.907, + 0.659 + ], + "angle": 0, + "content": "tokens, including both class and image tokens, will emphasize these identical tokens (outliers) in attention maps [12]. In other words, the corresponding outlier column in the attention map will be highlighted. Further, given that the classification process relies solely on the class token, the class token may exhibit a stronger response to the outlier in attention maps compared to image tokens. This prompts us to investigate whether the self-relevance of image tokens and the class token's similarity across all image tokens could serve as potential indicators to identify outliers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.909, + 0.827 + ], + "angle": 0, + "content": "Based on the above assumption, to robustly mitigate outliers, we introduce a Self-adaptive Outlier Mitigator (SOM) to detect and eliminate the outliers automatically. In particular, we use the attention map \\( Attn \\) to obtain the self-attention weights of image tokens, denoted as \\( Attn_{i,i} \\), and their attention weights to the class tokens, denoted as \\( Attn_{cls,i} \\), where \\( i \\) denotes the \\( i \\)-th image token and \\( cls \\) denotes the class token. By comparing the difference between \\( Attn_{i,i} \\) and \\( Attn_{cls,i} \\), we obtain the set of outliers \\( S \\), since the self-attention weights should be higher than others. This detection process of SOM is formulated as," + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.833, + 0.907, + 0.851 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} = \\left\\{i \\mid A t t n _ {i, i} < A t t n _ {c l s, i} \\right\\}, \\quad i \\in [ 1, N ], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.75, + 0.87 + ], + "angle": 0, + "content": "where \\(S\\) denotes the set of outliers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "After recognizing outliers, we propose eliminating them at the feature level rather than modifying the attention map" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20294" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.213 + ], + "angle": 0, + "content": "to prevent potential model collapse. Inspired by previous work [2], it's crucial to ensure the semantic consistency of the feature map, and directly masking outliers at the feature level is not an appropriate approach. Thus, SOM can eliminate outliers through in-place merging, where the detected outliers are replaced with the average value of its 8-neighbours. That is to say, the derived \\( F_{img} \\) in the original attention mechanism will be updated through our SOM as" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.223, + 0.484, + 0.281 + ], + "angle": 0, + "content": "\\[\n\\hat {F} _ {i m g} [ u ] = \\left\\{ \\begin{array}{l l} \\frac {1}{8} \\sum_ {v \\in \\mathcal {N} _ {8} (u)} F _ {i m g} [ v ], & u \\in \\mathcal {S} \\\\ F _ {i m g} [ u ], & \\text {o t h e r w s i e} \\end{array} , \\right. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.29, + 0.484, + 0.411 + ], + "angle": 0, + "content": "where \\(\\mathcal{N}_8(u)\\) means the 8-neighbour non-outlier tokens of the \\(u\\)-th point and \\(\\hat{F}_{img}\\) denotes the updated feature. In this way, our SOM initially identifies the outlier and then leverages surrounding token features to replace the detected one. This process can mitigate the adverse influence exerted by the outliers and maintain the feature distribution that is crucial for CLIP's image encoder. By doing so, the purified and superior semantic feature representation is generated." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.411, + 0.484, + 0.577 + ], + "angle": 0, + "content": "Furthermore, our SOM exhibits remarkable versatility in that it can be applied to each individual layer of the image encoder. This allows for the detection of outliers and the generation of purified features at multiple levels. In this manner, our parameter-free SOM establishes an efficient and precise mechanism for detecting and eliminating outliers that propagate across various layers of the image encoder. This mechanism is instrumental in preserving accurate semantic information and feature representations throughout the entire image encoder, thereby enhancing its overall performance and reliability." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.59, + 0.217, + 0.697 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.148, + 0.698, + 0.166, + 0.709 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.589, + 0.339, + 0.697 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.27, + 0.698, + 0.288, + 0.709 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.339, + 0.589, + 0.484, + 0.698 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.397, + 0.699, + 0.415, + 0.71 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.722, + 0.484, + 0.779 + ], + "angle": 0, + "content": "Figure 3. Comparison of different feature visualization. (a) Noisy feature with outliers. (b) Purified feature with SOM. (c) Semantic feature with SAE. It's seen that the proposed SAE strengthens the semantic regions with higher object activation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.804, + 0.408, + 0.818 + ], + "angle": 0, + "content": "3.4. Semantic-aware Attention Enhancer" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The comparison of Fig. 3 (a) & (b) indicates that the purified feature \\( F_{SOM} \\) generated by the last SOM effectively ignores the irrelevant regions of outliers, marked as white boxes. However, it's also observed from Fig. 3 (b) & (c) that the object activation of \\( F_{SOM} \\) is not salient enough," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "which indicates that eliminating outliers can't contribute to the attention response intensity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.122, + 0.905, + 0.183 + ], + "angle": 0, + "content": "To address this issue, we propose a Semantic-aware Attention Enhancer (SAE) to strengthen the object activation in CLIP's last layer, where the self-self attention mechanism is used to derive better feature relationships," + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.189, + 0.905, + 0.226 + ], + "angle": 0, + "content": "\\[\nA t t n _ {S A E} = \\operatorname {S o f t m a x} \\left(\\lambda \\left(\\frac {Q Q ^ {T}}{\\sqrt {d}} + \\frac {K K ^ {T}}{\\sqrt {d}}\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.232, + 0.906, + 0.307 + ], + "angle": 0, + "content": "where \\( Attn_{SAE} \\) represents the semantic-aware attention map and \\( \\lambda \\) denotes the logit scale used to sharpen attention scores, set to 0.5. Then, the semantic feature \\( F_{SAE} \\) is obtained through the residual connection with the purified feature \\( F_{SOM} \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.317, + 0.905, + 0.334 + ], + "angle": 0, + "content": "\\[\nF _ {S A E} = F _ {S O M} + \\operatorname {F F N} \\left(\\operatorname {A t t n} _ {S A E} F _ {S O M}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.342, + 0.906, + 0.478 + ], + "angle": 0, + "content": "The self-self attention mechanism has shown a strong ability to construct feature relationships [16, 24, 45]. Therefore, applying self-self attention to our purified feature enables the establishment of more precise spatial correlations. These accurate spatial correlations play a crucial role in enabling each point to aggregate the relevant semantic feature effectively. As illustrated in Fig. 3 (b) & (c), SAE significantly augments attention intensity in semantic regions compared to the feature solely processed by SOM." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.486, + 0.809, + 0.502 + ], + "angle": 0, + "content": "3.5. Hierarchical Attention Integrator" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.906, + 0.599 + ], + "angle": 0, + "content": "To further refine the generated semantic feature \\( F_{SAE} \\), we leverage attention maps from shallow layers to capture discriminative pair-wise feature relationships since shallow layers typically focus on object structures [27]. To this end, we propose a Hierarchical Attention Integrator (HAI), which integrates multi-layer attention maps to refine \\( F_{SAE} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.906, + 0.688 + ], + "angle": 0, + "content": "Instead of using the original attention maps with outliers, HAI integrates attention maps from each transformer layer while excluding the detected outlier weights to avoid information pollution. Firstly, HAI employs the outlier set \\( S \\) to select the non-outlier weights of each attention map, which can be formulated as," + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.696, + 0.905, + 0.737 + ], + "angle": 0, + "content": "\\[\n\\overline {{A t t n}} _ {j, i} ^ {l} = \\left\\{ \\begin{array}{l l} A t t n _ {j, i}, & \\text {i f} i \\notin \\mathcal {S} ^ {l} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.746, + 0.906, + 0.813 + ], + "angle": 0, + "content": "where \\(\\overline{Attn}^l\\) and \\(S^l\\) denote the filtered attention map and the outlier set at the \\(l\\)-th layer. \\(j\\) and \\(i\\) correspond to the row and column indices of the attention map. Next, HAI integrates \\(\\overline{Attn}^l\\) from each layer, formulated as," + }, + { + "type": "equation", + "bbox": [ + 0.654, + 0.822, + 0.905, + 0.863 + ], + "angle": 0, + "content": "\\[\nA = \\sum_ {l = 1} ^ {L - 1} \\overline {{A t t n}} ^ {l}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "where \\(A\\) denotes the integrated attention map, and \\(L\\) denotes the total number of transformer layers." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20295" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.116, + 0.089, + 0.887, + 0.367 + ], + "angle": 0, + "content": "
MethodsPub. & YearExtra backboneTraining freewith backgroundV20without backgroundAvg.
V21PC60ObjectPC59StuffADECity
GroupVit [50]CVPR'22XX50.418.727.579.723.415.39.211.129.4
SegCLIP [32]ICML'23XX52.624.726.5------
TCL [7]CVPR'23XX55.030.431.683.233.922.417.124.037.2
DINOiser† [47]ECCV'24DINOX62.132.434.880.935.924.620.031.740.3
SAM-CLIP† [46]ECCV'24SAMX60.629.2----17.1--
PnP-OVSS† [33]CVPR'24BLIP✓--36.251.328.017.914.2--
LaVG† [19]ECCV'24DINO✓62.131.634.282.534.723.215.826.238.8
ProxyCLIP† [25]ECCV'24DINO✓61.335.337.580.339.126.520.238.142.3
CLIP [38]ICML'21X✓18.69.98.149.411.15.73.16.514.1
MaskCLIP [14]ECCV'22X✓38.823.620.674.926.416.49.812.627.9
CLIPSurgery [28]PR'25X✓-29.3---21.9-31.4-
CaR [44]CVPR'24X✓48.630.536.673.739.5-17.7--
GEM [3]CVPR'24X✓46.2---32.615.7---
CLIPtrase [40]ECCV'24X✓50.929.943.681.033.822.816.421.337.5
ClearCLIP* [24]ECCV'24X✓57.032.633.080.935.923.916.730.038.8
SCLIP* [45]ECCV'24X✓59.731.733.581.534.522.716.532.339.1
NACLIP* [16]WACV'25X✓58.932.233.279.735.223.317.435.539.4
SFP (Ours)X✓63.937.237.984.539.926.420.841.144.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.38, + 0.907, + 0.424 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison of our SFP against other approaches across eight segmentation benchmark datasets. The best and second-best results are marked with **bold** and **underline**, respectively. † indicates that the additional backbone settings follow the original paper. * indicates the results are re-evaluated using the official implementation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.449, + 0.484, + 0.555 + ], + "angle": 0, + "content": "After integrating multi-layer attention maps, we deploy a normalization strategy to optimize \\( A \\) instead of simply averaging it. Inspired by previous works [22, 35], we iteratively apply row-wise and column-wise normalization to \\( A \\) for a more object-centric attention distribution, leading to a stronger focus on relevant regions. This process can be formulated as," + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.565, + 0.482, + 0.582 + ], + "angle": 0, + "content": "\\[\nA _ {H A I} = \\operatorname {n o r m} _ {c} \\left(\\operatorname {n o r m} _ {r} (A)\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.592, + 0.483, + 0.638 + ], + "angle": 0, + "content": "where \\(A_{HAI}\\) denotes the normalized attention map, normc and normr denote the column-wise and row-wise normalization respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.638, + 0.484, + 0.682 + ], + "angle": 0, + "content": "Then, \\(A_{HAI}\\) is multiplied with the semantic feature \\(F_{SAE}\\) to obtain the final semantic feature \\(F_{vis}\\) for segmentation, which is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.694, + 0.483, + 0.71 + ], + "angle": 0, + "content": "\\[\nF _ {v i s} = A _ {H A I} F _ {S A E}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.767 + ], + "angle": 0, + "content": "Finally, we compute the cosine similarity between \\( F_{vis} \\) and the text feature \\( F_{txt} \\) extracted from the CLIP text encoder to predict the segmentation map, as formulated in Eq. 3." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.779, + 0.217, + 0.796 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.804, + 0.298, + 0.82 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Datasets We conduct comprehensive experiments to verify the effectiveness of our method on diverse benchmark datasets: PASCAL VOC 2012 [15], PASCAL Context [34], and COCO [29]. These datasets are categorized into two settings: 1) with a background class, namely V20, PC59," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.449, + 0.907, + 0.494 + ], + "angle": 0, + "content": "and Stuff; 2) without the background class, namely V21, PC60, and Object. We also report results on the ADE20K (ADE) [57] and Cityscapes (City) [11] datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.494, + 0.909, + 0.646 + ], + "angle": 0, + "content": "Implementation Details Similar to existing training-free methods, we adopt CLIP ViT-B/16 with the original pretrained weight [38]. Following previous works [2, 45], we combine standard ImageNet prompts [13] with the class names to construct text descriptions. For a fair comparison with previous works [25, 45], we resize images with a short side of 336 pixels (or 560 pixels for the high-resolution Cityscapes dataset) and apply sliding window inference using a \\(224 \\times 224\\) window with a \\(112 \\times 112\\) stride. Our experiments are conducted using one NVIDIA 3090 GPU." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.653, + 0.879, + 0.669 + ], + "angle": 0, + "content": "4.2. Comparison with State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.764 + ], + "angle": 0, + "content": "Our SFP is evaluated against existing SOTA methods, including both training-based and training-free approaches. The post-processing mask refinement techniques, such as PAMR [1] or DenseCRF [23], are not included for a fair comparison. The standard mean intersection over union (mIoU) is used as the metric for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.855 + ], + "angle": 0, + "content": "Quantitative Evaluation The main results are summarized in Tab. 1, where all evaluations are conducted using CLIP ViT-B/16 as the backbone to ensure a fair comparison. We compare SFP against multiple methods, including training-based methods, training-free methods with extra backbones, and training-free methods without extra backbones." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.902 + ], + "angle": 0, + "content": "As shown in Tab. 1, our SFP outperforms existing SOTA methods, regardless of whether an additional backbone is introduced or not. In particular, compared with existing" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20296" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.135, + 0.089, + 0.865, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.397, + 0.908, + 0.425 + ], + "angle": 0, + "content": "Figure 4. Qualitative Results Comparison. We compare our SFP with CLIP [38], ClearCLIP [24], SCLIP [45] and NACLIP [16] without post-processing. It's observed that SFP generates more precise segmentation results." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.452, + 0.486, + 0.709 + ], + "angle": 0, + "content": "training-free methods without extra backbones [16, 24, 45] or cluster methods [53-55], our SFP achieves an average gain of \\(4.6\\%\\) mIoU over previous SOTA NACLIP [16]. Specifically, On the V20 and V21 dataset [15], SFP surpasses NACLIP [16] \\(4.8\\%\\) and \\(5\\%\\) mIoU, respectively. These improvements indicate that our SFP can generate more purified feature maps for more precise segmentation. Beyond that, SFP outperforms SOTA training-free methods with extra backbones, e.g., ProxyCLIP [25], with a gain of \\(3.0\\%\\) mIoU on the Cityscape [11] dataset and an average gain of \\(1.7\\%\\) mIoU on eight datasets. Our SFP can even beat DINOiser [47], which requires both training and an extra backbone, with an average improvement of \\(3.7\\%\\) mIoU. These results highlight our SFP boosts CLIP's internal potential and reveal that feature purification does matter in training-free OVSS, where more purified features can predict better segmentation masks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.485, + 0.901 + ], + "angle": 0, + "content": "Qualitative Evaluation In Fig. 4, we compare the visualization results of various training-free methods without extra backbones, including CLIP [38], ClearCLIP [24], SCLIP [45], NACLIP [16] and our SFP. In particular, the segmentation maps obtained by vanilla CLIP [38] exhibit significant noises, unveiling its limitations in spatial localization. Besides, in complex scenarios, like 'sofa' in the second row, SCLIP [45] and NACLIP [16] fail to capture the complete segmentation mask. In contrast, our SFP can predict more accurate and complete segmentation maps. These results indicate that our SFP has a good ability to predict comprehensive segmentation masks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.451, + 0.667, + 0.468 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.474, + 0.907, + 0.548 + ], + "angle": 0, + "content": "Comprehensive ablation studies are performed to analyze the effectiveness of our proposed mechanisms. Unless otherwise specified, we utilize CLIP with ViT-B/16 backbone and evaluate on V21 [15] and ADE [57] datasets for background and non-background settings, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.549, + 0.909, + 0.732 + ], + "angle": 0, + "content": "Effect on each proposed module We first conduct several experiments to validate the effectiveness of our proposed modules of our SFP: SOM, SAE, and HAI by adding each module step by step, where ClearCLIP [24] is used as the baseline. As shown in Tab. 2, our SOM achieves a gain of \\(3.1\\%\\) and \\(1.5\\%\\) mIoU on V21 and ADE datasets, respectively. Further, SAE enhances attention intensity in relevant regions, leading to an average gain of \\(1.7\\%\\) mIoU. Finally, HAI helps our SFP reach \\(63.9\\%\\) and \\(20.8\\%\\) mIoU on V21 and ADE datasets, which set the new SOTA performance. Due to space constraints, more ablation studies about SOM and SAE are summarized in the supplementary material." + }, + { + "type": "table", + "bbox": [ + 0.558, + 0.74, + 0.869, + 0.814 + ], + "angle": 0, + "content": "
SOMSAEHAIV21ADEAvg.
057.016.736.9
1✓60.118.239.2
2✓✓62.319.540.9
3✓✓✓63.920.842.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.828, + 0.908, + 0.857 + ], + "angle": 0, + "content": "Table 2. Ablation results on each proposed module to validate the impact on performance across V21 [15] and ADE [57] datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.903 + ], + "angle": 0, + "content": "More visualization on the ability of SOM to suppress outlier propagation We visualize purified feature maps" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20297" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.135, + 0.089, + 0.865, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.242, + 0.908, + 0.272 + ], + "angle": 0, + "content": "Figure 5. Visualization of the feature purification across CLIP's image encoder layers. (a) Feature maps without SOM purification. (b) Feature maps purified by SOM." + }, + { + "type": "table", + "bbox": [ + 0.135, + 0.295, + 0.446, + 0.361 + ], + "angle": 0, + "content": "
Al=l-1l=lo1581011
V2163.963.463.062.662.5
ADE20.820.520.319.919.7
Avg.42.442.041.741.341.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.375, + 0.484, + 0.418 + ], + "angle": 0, + "content": "Table 3. Ablation results on different selection strategies in HAI, where \\(A_{l = l_0}^{l = L - 1}\\) represents the integrated attention map from \\(l_0\\) to the \\(L - 1\\) layer. '1,5,8,10,11' indicates the value of \\(l_0\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.444, + 0.484, + 0.55 + ], + "angle": 0, + "content": "across CLIP's layers, from the 2nd layer to the 9th layer, which is shown in Fig. 5. It's found that the outliers gradually affect the aggregated features, especially from layer 6 to layer 9 in Fig. 5 (a). Then, with the help of our SOM, the influence of outliers is mitigated layer by layer, like in Fig. 5 (b). Thus, our SOM can suppress the propagated outlier features and gradually purify the overall semantic features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.55, + 0.484, + 0.701 + ], + "angle": 0, + "content": "Effect on hierarchical attention integrator A series of ablation studies is conducted to investigate the optimal way to leverage attention maps from different layers. We omit HAI from our SFP to establish the baseline model for this experiment. As summarized in Tab. 3, we take \\( l_{0} \\) to represent the starting layer. When \\( l_{0} = 1 \\), our HAI performs best with all attention maps selected. Besides, increasing \\( l_{0} \\) from 1 to 11 decreases the segmentation performance, which demonstrates the effectiveness of the shallow attention maps for semantic feature refinement." + }, + { + "type": "table", + "bbox": [ + 0.133, + 0.712, + 0.446, + 0.854 + ], + "angle": 0, + "content": "
BackboneMethodsV21ADEAvg.
ViT-B/32SCLIP [45]50.614.832.7
LaVG [19]54.815.535.2
ProxyCLIP [25]57.916.737.3
NACLIP [16]51.114.933.0
SFP60.117.138.6
ViT-L/14SCLIP [45]44.410.927.7
LaVG [19]52.117.334.7
ProxyCLIP [25]60.622.641.6
NACLIP [16]52.217.334.8
SFP64.722.143.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.869, + 0.483, + 0.896 + ], + "angle": 0, + "content": "Table 4. Comparison with other SOTA methods across different CLIP backbones." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.298, + 0.907, + 0.48 + ], + "angle": 0, + "content": "Effect on different CLIP backbones To validate the robustness, we further evaluate our SFP with other methods using different CLIP backbones, such as ViT-B/32 and ViT-L/14. As summarized in Tab. 4, some methods, like SCLIP [45] and LaVG [19], suffer from performance degradation when applied to the larger backbone. In contrast, SFP maintains superior performance across different backbones and achieves an improvement of \\(1.3\\%\\) and \\(1.8\\%\\) mIoU over ProxyCLIP [25] with ViT-B/32 and ViT-L/14 backbone, respectively. These results indicate that SFP can robustly handle outliers derived from different backbones and generate purified features for precise segmentation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.505, + 0.634, + 0.521 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.907, + 0.776 + ], + "angle": 0, + "content": "In this paper, we present a Self-adaptive Feature Purifier framework (SFP) to mitigate outlier propagation for training-free open-vocabulary semantic segmentation. Specifically, we propose SOM deployed at each layer of the image encoder to identify and mitigate outliers by comparing attention weights between image and class tokens, which generates purified features for propagation. Next, to recover the semantic information affected by outliers, we introduce SAE to strengthen object activation in the purified feature. Beyond that, we design HAI to utilize multilayer attention maps for refined object-centric representations. Extensive experiments show that SFP achieves superior segmentation performance compared to existing SOTA methods on various datasets. Our SFP shows that feature purification does matter for OVSS, and cleaner features can predict more complete and accurate segmentation masks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported by the National Natural Science Foundation of China (No. 62471405, 62331003, 62301451,62301613), Basic Research Program of Jiangsu (BK20241814), Suzhou Basic Research Program (SYG202316) and XJTLU REF-22-01-010 and RDF-22-02-066, XJTLU AI University Research Centre, Jiangsu Province Engineering Research Centre of Data Science and Cognitive Computation at XJTLU." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20298" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.484, + 0.156 + ], + "angle": 0, + "content": "[1] Nikita Araslanov and Stefan Roth. Single-stage semantic segmentation from image labels. In CVPR, pages 4253-4262, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.157, + 0.484, + 0.212 + ], + "angle": 0, + "content": "[2] Sule Bai, Yong Liu, Yifei Han, Haoji Zhang, and Yansong Tang. Self-calibrated clip for training-free open-vocabulary segmentation. arXiv preprint arXiv:2411.15869, 2024. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.213, + 0.483, + 0.269 + ], + "angle": 0, + "content": "[3] Walid Bousselham, Felix Petersen, Vittorio Ferrari, and Hilde Kuehne. Grounding everything: Emerging localization properties in vision-language transformers. In CVPR, pages 3828-3837, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.27, + 0.482, + 0.311 + ], + "angle": 0, + "content": "[4] Maxime Bucher, Tuan-Hung Vu, Matthieu Cord, and Patrick Pérez. Zero-shot semantic segmentation. In NeurlPS, pages 468-479, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.312, + 0.482, + 0.368 + ], + "angle": 0, + "content": "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.369, + 0.482, + 0.436 + ], + "angle": 0, + "content": "[6] Niccolo Cavagnero, Gabriele Rosi, Claudia Cuttano, Francesca Pistilli, Marco Ciccone, Giuseppe Avera, and Fabio Cermelli. Pem: Prototype-based efficient maskformer for image segmentation. In CVPR, pages 15804-15813, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.438, + 0.482, + 0.492 + ], + "angle": 0, + "content": "[7] Junbum Cha, Jonghwan Mun, and Byungseok Roh. Learning to generate text-grounded mask for open-world semantic segmentation from only image-text pairs. In CVPR, pages 11165-11174, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.494, + 0.482, + 0.549 + ], + "angle": 0, + "content": "[8] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, pages 1290–1299, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.55, + 0.482, + 0.62 + ], + "angle": 0, + "content": "[9] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, pages 2818-2829, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.621, + 0.482, + 0.676 + ], + "angle": 0, + "content": "[10] Seokju Cho, Heeseong Shin, Sunghwan Hong, Anurag Arnab, Paul Hongsuck Seo, and Seungryong Kim. Cat-seg: Cost aggregation for open-vocabulary semantic segmentation. In CVPR, pages 4113-4123, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.677, + 0.482, + 0.746 + ], + "angle": 0, + "content": "[11] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, pages 3213-3223, 2016. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[12] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. In ICLR, pages 1-11, 2024. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, pages 248-255, 2009. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.831, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[14] Zheng Ding, Jieke Wang, and Zhuowen Tu. Open-vocabulary panoptic segmentation with maskclip. arXiv preprint arXiv:2208.08984, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[15] Mark Everingham, SM Ali Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "pascal visual object classes challenge: A retrospective. IJCV, 111:98-136, 2015. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.121, + 0.905, + 0.175 + ], + "angle": 0, + "content": "[16] Sina Hajimiri, Ismail Ben Ayed, and Jose Dolz. Pay attention to your neighbours: Training-free open-vocabulary semantic segmentation. In WACV, pages 5061-5071, 2025. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.177, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[17] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.247, + 0.905, + 0.3 + ], + "angle": 0, + "content": "[18] Siyu Jiao, Yunchao Wei, Yaowei Wang, Yao Zhao, and Humphrey Shi. Learning mask-aware clip representations for zero-shot segmentation. In NeurIPS, pages 35631-35653, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.904, + 0.343 + ], + "angle": 0, + "content": "[19] Dahiyun Kang and Minsu Cho. In defense of lazy visual grounding for open-vocabulary semantic segmentation. In ECCV, pages 143-164, 2024. 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.904, + 0.385 + ], + "angle": 0, + "content": "[20] Laurynas Karazija, Iro Laina, Andrea Vedaldi, and Christian Rupprecht. Diffusion models for zero-shot open-vocabulary segmentation. In ECCV, pages 299-317, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.904, + 0.441 + ], + "angle": 0, + "content": "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, pages 4015-4026, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.904, + 0.483 + ], + "angle": 0, + "content": "[22] Philip A Knight. The sinkhorn-knopp algorithm: convergence and applications. SIAM Journal on Matrix Analysis and Applications, 30(1):261-275, 2008. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.904, + 0.525 + ], + "angle": 0, + "content": "[23] Philipp Kähenbuhl and Vladlen Koltun. Efficient inference in fully connected crfs with gaussian edge potentials. In NeurIPS, pages 109-117, 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.904, + 0.58 + ], + "angle": 0, + "content": "[24] Mengcheng Lan, Chaofeng Chen, Yiping Ke, Xinjiang Wang, Litong Feng, and Wayne Zhang. Clearclip: Decomposing clip representations for dense vision-language inference. In ECCV, pages 143-160, 2024. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.904, + 0.636 + ], + "angle": 0, + "content": "[25] Mengcheng Lan, Chaofeng Chen, Yiping Ke, Xinjiang Wang, Litong Feng, and Wayne Zhang. Proxyclip: Proxy attention improves clip for open-vocabulary segmentation. In ECCV, pages 70-88, 2024. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.904, + 0.692 + ], + "angle": 0, + "content": "[26] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.904, + 0.748 + ], + "angle": 0, + "content": "[27] Yunheng Li, Zhongyu Li, Quansheng Zeng, Qibin Hou, and Ming-Ming Cheng. Cascade-clip: Cascaded vision-language embeddings alignment for zero-shot semantic segmentation. In ICML, pages 28243–28258, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.904, + 0.802 + ], + "angle": 0, + "content": "[28] Yi Li, Hualiang Wang, Yiqun Duan, Jiheng Zhang, and Xiaomeng Li. A closer look at the explainability of contrastive language-image pre-training. Pattern Recognition, 162:111409, 2025. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.859 + ], + "angle": 0, + "content": "[29] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[30] Yong Liu, Sule Bai, Guanbin Li, Yitong Wang, and Yansong Tang. Open-vocabulary segmentation with semantic-assisted calibration. In CVPR, pages 3491-3500, 2024. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20299" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[31] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, pages 3431-3440, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[32] Huaishao Luo, Junwei Bao, Youzheng Wu, Xiaodong He, and Tianrui Li. Segclip: Patch aggregation with learnable centers for open-vocabulary semantic segmentation. In ICML, pages 23033-23044, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.246 + ], + "angle": 0, + "content": "[33] Jiayun Luo, Siddhesh Khandelwal, Leonid Sigal, and Boyang Li. Emergent open-vocabulary semantic segmentation from off-the-shelf vision-language models. In CVPR, pages 4029-4040, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.246, + 0.482, + 0.314 + ], + "angle": 0, + "content": "[34] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The role of context for object detection and semantic segmentation in the wild. In CVPR, pages 891–898, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.385 + ], + "angle": 0, + "content": "[35] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.386, + 0.482, + 0.442 + ], + "angle": 0, + "content": "[36] Zelin Peng, Zhengqin Xu, Zhilin Zeng, Yu Huang, Yaoming Wang, and Wei Shen. Parameter-efficient fine-tuning in hyperspherical space for open-vocabulary semantic segmentation. In CVPR, pages 15009-15020, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.482, + 0.509 + ], + "angle": 0, + "content": "[37] Zelin Peng, Zhengqin Xu, Zhilin Zeng, Changsong Wen, Yu Huang, Menglin Yang, Feilong Tang, and Wei Shen. Understanding fine-tuning clip for open-vocabulary semantic segmentation in hyperbolic space. In CVPR, pages 4562-4572, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.58 + ], + "angle": 0, + "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.58, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[39] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, pages 25278–25294, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.664, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[40] Tong Shao, Zhuotao Tian, Hang Zhao, and Jingyong Su. Explore the potential of clip for training-free open vocabulary semantic segmentation. In ECCV, pages 139-156, 2024. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.482, + 0.762 + ], + "angle": 0, + "content": "[41] Gyungin Shin, Weidi Xie, and Samuel Albanie. Reco: Retrieve and co-segment for zero-shot transfer. In NeurIPS, pages 33754-33767, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[42] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, pages 2256-2265, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.859 + ], + "angle": 0, + "content": "[43] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In ICCV, pages 7262-7272, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[44] Shuyang Sun, Runjia Li, Philip Torr, Xiuye Gu, and Siyang Li. Clip as rn: Segment countless visual concepts without training endeavor. In CVPR, pages 13171-13182, 2024. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[45] Feng Wang, Jieru Mei, and Alan Yuille. Sclip: Rethinking self-attention for dense vision-language inference. In ECCV, pages 315–332, 2024. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.135, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[46] Haoxiang Wang, Pavan Kumar Anasosalu Vasu, Fartash Faghri, Raviteja Vemulapalli, Mehrdad Farajtabar, Sachin Mehta, Mohammad Rastegari, Oncel Tuzel, and Hadi Pouransari. Sam-clip: Merging vision foundation models towards semantic and spatial understanding. In CVPR, pages 3635–3647, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.218, + 0.905, + 0.273 + ], + "angle": 0, + "content": "[47] Monika Wysoczanska, Oriane Simeoni, Michael Ramamonjisoa, Andrei Bursuc, Tomasz Trzciński, and Patrick Pérez. Clip-dinoiser: Teaching clip a few dino tricks. In ECCV, pages 320-337, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.273, + 0.905, + 0.327 + ], + "angle": 0, + "content": "[48] Bin Xie, Jiale Cao, Jin Xie, Fahad Shahbaz Khan, and Yanwei Pang. Sed: A simple encoder-decoder for open-vocabulary semantic segmentation. In CVPR, pages 3426-3436, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.328, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[49] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. In NeurIPS, pages 12077-12090, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.905, + 0.442 + ], + "angle": 0, + "content": "[50] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. Groupvit: Semantic segmentation emerges from text supervision. In CVPR, pages 18134-18144, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[51] Jilan Xu, Junlin Hou, Yuejie Zhang, Rui Feng, Yi Wang, Yu Qiao, and Weidi Xie. Learning open-vocabulary semantic segmentation models from natural language supervision. In CVPR, pages 2935-2944, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[52] Mengde Xu, Zheng Zhang, Fangyun Wei, Han Hu, and Xiang Bai. Side adapter network for open-vocabulary semantic segmentation. In CVPR, pages 2945-2954, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[53] Xihong Yang, Xiaochang Hu, Sihang Zhou, Xinwang Liu, and En Zhu. Interpolation-based contrastive learning for few-label semi-supervised learning. TNNLS, 35(2):2054-2065, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.594, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[54] Xihong Yang, Jin Jiaqi, Siwei Wang, Ke Liang, Yue Liu, Yi Wen, Suyuan Liu, Sihang Zhou, Xinwang Liu, and En Zhu. Dealmvc: Dual contrastive calibration for multi-view clustering. In ACMMM, pages 337-346, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.65, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[55] Xihong Yang, Siwei Wang, Fangdi Wang, Jiaqi Jin, Suyuan Liu, Yue Liu, En Zhu, Xinwang Liu, and Yueming Jin. Automatically identify and rectify: Robust deep contrastive multiview clustering in noisy scenarios. In ICML. PMLR, 2025. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.719, + 0.905, + 0.774 + ], + "angle": 0, + "content": "[56] Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang-Chieh Chen. Convolutions die hard: Open-vocabulary segmentation with single frozen convolutional clip. In NeurIPS, pages 32215-32234, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.905, + 0.817 + ], + "angle": 0, + "content": "[57] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ADE20K dataset. In CVPR, pages 5122-5130, 2017. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[58] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In ECCV, pages 696-712, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.888 + ], + "angle": 0, + "content": "[59] Ziqin Zhou, Yinjie Lei, Bowen Zhang, Lingqiao Liu, and Yifan Liu. Zegclip: Towards adapting clip for zero-shot semantic segmentation. In CVPR, pages 11175-11185, 2023." + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.888 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20300" + } + ] +] \ No newline at end of file diff --git a/2025/Feature Purification Matters_ Suppressing Outlier Propagation for Training-Free Open-Vocabulary Semantic Segmentation/4f979197-bed4-425a-a1d2-c2b5ca2160b9_origin.pdf b/2025/Feature Purification Matters_ Suppressing Outlier Propagation for Training-Free Open-Vocabulary Semantic Segmentation/4f979197-bed4-425a-a1d2-c2b5ca2160b9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4e4e220b171612ad449f77110ab8cf6083684088 --- /dev/null +++ b/2025/Feature Purification Matters_ Suppressing Outlier Propagation for Training-Free Open-Vocabulary Semantic Segmentation/4f979197-bed4-425a-a1d2-c2b5ca2160b9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f64705affa51389ae02d0dfaeffd3285f4c2b1f22ce765497f010721a5030d56 +size 2030036 diff --git a/2025/Feature Purification Matters_ Suppressing Outlier Propagation for Training-Free Open-Vocabulary Semantic Segmentation/full.md b/2025/Feature Purification Matters_ Suppressing Outlier Propagation for Training-Free Open-Vocabulary Semantic Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7da7ca4e4fc7194ab923510f997befa8fdec6932 --- /dev/null +++ b/2025/Feature Purification Matters_ Suppressing Outlier Propagation for Training-Free Open-Vocabulary Semantic Segmentation/full.md @@ -0,0 +1,340 @@ +# Feature Purification Matters: Suppressing Outlier Propagation for Training-Free Open-Vocabulary Semantic Segmentation + +Shuo Jin $^{1,2}$ Siyue Yu $^{1*}$ Bingfeng Zhang $^{3}$ Mingjie Sun $^{4}$ Yi Dong $^{2}$ Jimin Xiao $^{1}$ $^{1}$ Xi'an Jiaotong-Liverpool University $^{2}$ University of Liverpool + $^{3}$ China University of Petroleum (East China) $^{4}$ Soochow University +{shuo.jin, yi.dong}@liverpool.ac.uk, {siyue.yu02, jimin.xiao}@xjtlu.edu.cn +bingfeng.zhang@upc.edu.cn, mjsun@suda.edu.cn + +# Abstract + +Training-free open-vocabulary semantic segmentation has advanced with vision-language models like CLIP, which exhibit strong zero-shot abilities. However, CLIP's attention mechanism often wrongly emphasises specific image tokens, namely outliers, which results in irrelevant over-activation. Existing approaches struggle with these outliers that arise in intermediate layers and propagate through the model, ultimately degrading spatial perception. In this paper, we propose a Self-adaptive Feature Purifier framework (SFP) to suppress propagated outliers and enhance semantic representations for open-vocabulary semantic segmentation. Specifically, based on an in-depth analysis of attention responses between image and class tokens, we design a self-adaptive outlier mitigator to detect and mitigate outliers at each layer for propagated feature purification. In addition, we introduce a semantic-aware attention enhancer to augment attention intensity in semantically relevant regions, which strengthens the purified feature to focus on objects. Further, we introduce a hierarchical attention integrator to aggregate multi-layer attention maps to refine spatially coherent feature representations for final segmentation. Our proposed SFP enables robust outlier suppression and object-centric feature representation, leading to a more precise segmentation. Extensive experiments show that our method achieves state-of-the-art performance and surpasses existing methods by an average of $4.6\%$ mIoU on eight segmentation benchmarks. The code is released at: https://github.com/Kimsure/SFP. + +# 1. Introduction + +Open-vocabulary semantic segmentation (OVSS) [4, 58] seeks to partition an image into distinct regions and assign pixel-level labels to arbitrary semantic categories. Un + +![](images/3a809d8d6a21c9db72c8c9f3de6b62c8340aaecec366c51646fb2b91e2579c1c.jpg) +(a) + +![](images/3114cb73dc14bef72cdbcbae1c9dd2645ba2f83eea07987b45cf6211178fa784.jpg) +(b) + +![](images/ec6b2006f99b0a5202453aaa61a917a5f9fe0ac356e732bc8d5e233cb5244a7f.jpg) +(c) + +![](images/c4f8781f13f508d9d11ad39347cc9035561e4d98b07fe6491addac41b9b76c5d.jpg) +(d) + +![](images/5e9d4ab5123f1a5541ceae7175fe2b8e0c0012c6dcab1917b3e6a2cae5088baa.jpg) +(e) +Figure 1. Visualization of attention and feature maps. (a) Attention map of the selected target object token (marked as $\star$ ). (b) Attention map of the selected background token (marked as $\star$ ). (c) Attention map of the class token. (d) Diagonal weights of the attention map of each image token. (e) Detected outliers. (f) - (h) output feature map of CLIP [38], SCLIP [45], and our SFP. + +![](images/985463eea3e0187c90e3448f13faec3f03e9b8bd1c7111b9d6bd794cf9bf4d1a.jpg) +(f) + +![](images/772505dcee4a892fcc1253b7db85cab2a9660fcc8a2407c5ec9890c08344c1a1.jpg) +(g) + +![](images/e5703a47e2d24303ae0dd1ccdc5b4fc7b0ab8e9f4fd2f06ce3c18696f858fd6d.jpg) +(h) + +like traditional semantic segmentation, OVSS must handle an open set of classes, making it more challenging. Vision-language models (VLM), such as CLIP [38], demonstrate remarkable zero-shot performance by leveraging large-scale image-text pairs [39]. Thus, existing approaches can be categorized into two pipelines: fine-tuning CLIP or training-free CLIP. Although the fine-tuning pipeline yields superior performance, it risks overfitting and compromises generalization [30, 43]. In contrast, the training-free pipeline preserves CLIP's original generalization capability without additional training, which is more convenient and efficient. + +Recent training-free methods [12, 40] have demonstrated that in the original CLIP's pair-wise attention maps, there exist the same highly correlated image tokens for all input tokens, such as those marked with the white box in Fig. 1 (a) - (b). In such a situation, regardless of whether it is the target image token (Fig. 1 (a)), the background image to- + +ken (Fig. 1 (b)), or the class token (Fig. 1 (c)), the pair-wise token relationships within the attention map are all over-activated for these identical tokens, i.e., outliers. Consequently, when aggregated with outliers, the derived semantic feature representation is easily polluted to overemphasize these outliers and the interpretation of the target region is impeded, e.g., the feature map of 'car' in Fig 1 (f) focuses on the outliers instead of the object itself. + +To mitigate outliers and ensure more accurate semantic feature representation, many training-free approaches have presented diverse modifications to the self-attention mechanism in the last layer [14, 16, 24, 45]. While these strategies effectively restrict CLIP's attention to ignore outliers, some remain unfiltered, as shown in Fig. 1 (g). Moreover, some approaches [2, 12] rely on empirical parameter tuning to detect and remove a fixed number of outliers, which is less robust to different backbones and images. Beyond these optimization approaches, several studies [12, 40] further reveal that outliers predominantly appear in the intermediate layers. In this way, once outliers emerge, they will propagate through subsequent layers, continuously affecting semantic feature representations. Yet, previous approaches fail to analyze and address outliers in the intermediate layers so that outliers are easily propagated to pollute the final semantic feature. Thus, our objective is to adaptively eliminate outliers across multiple layers to achieve more robust semantic feature representations without training. + +Building on these insights, we propose a Self-adaptive Feature Purifier framework (SFP) to resolve outliers adaptively and enhance semantic representations for CLIP-based training-free OVSS. Firstly, we design a Self-adaptive Outlier Mitigator (SOM) to identify and mitigate outliers across multiple layers. Tokens are expected to focus primarily on their relevant regions, especially themselves, rather than irrelevant regions in the attention map. However, as shown in Fig. 1 (d), we observe that the self-response of image tokens in the attention map is not always prominent, especially outliers. The comparison of Fig. 1 (c) & (d) reveals that outliers exhibit a stronger response to the class token than their self-response. Hence, our SOM can identify outliers by simply computing the difference between the self-response values of the image tokens in the attention map (i.e., the diagonal weights of the attention map) and their attention values with the class token. The detected outliers are shown in Fig. 1 (e). In this way, our parameter-free SOM enables adaptive detection across various layers to mitigate the propagating influence of outliers without manual tuning. + +Nevertheless, outliers not only cause tokens to overemphasize them but also impede tokens' responses to semantic regions in the attention map. Therefore, merely removing the outliers can't directly increase the attention weights of the related regions. To enhance the tokens' attention responses to semantic regions, we propose a Semantic-aware + +Attention Enhancer (SAE). SAE incorporates the self-self attention mechanism in CLIP's last layer and applies it to the purified feature derived from SOM at the penultimate layer, enabling a better focus on semantic regions. In addition, shallow attention maps typically contain structured object information and focus on relevant semantic regions [27, 30]. To further enhance the precise spatial perception of the final feature map, we propose a Hierarchical Attention Integrator (HAI) as an assistant to the SAE, where the attention maps in previous layers are firstly processed to mitigate the influence of the outliers and then integrated to refine the final semantic feature for prediction. The refined semantic feature for the final prediction is shown in Fig. 1 (h). Our SFP can generate a more purified and complete semantic feature for segmentation. + +We conduct comprehensive experiments to evaluate the proposed SFP. Our method outperforms existing methods and achieves state-of-the-art (SOTA) performance across eight evaluation datasets. In summary, our contributions are concluded as follows: + +- We present a Self-adaptive Feature Purifier framework (SFP) for training-free OVSS, which can derive a more accurate semantic feature to strengthen the final open-vocabulary semantic segmentation. +- We propose a simple yet effective mechanism, Self-adaptive Outlier Mitigator (SOM), which adaptively detects and mitigates outliers across CLIP's diverse layers so that the influence of the propagated influence of outliers can be suppressed. +- We introduce a Semantic-aware Attention Enhancer (SAE) and a Hierarchical Attention Integrator (HAI) to further refine the semantic feature relationships on relevant regions, aggregating the more relevant semantic feature for final prediction. +- Comprehensive experiments demonstrate that SFP derives the best performance under a fair comparison e.g., SFP achieves a gain of $5.6\%$ mIoU on Cityscapes and an average gain of $4.6\%$ mIoU across eight datasets. + +# 2. Related Work + +# 2.1. pre-trained Vision-Language Models + +Pre-trained vision-language models [17, 26, 38] have drawn significant attention for their ability to bridge visual and textual modalities. CLIP [38] has emerged as a highly successful model trained on a large-scale dataset of image-text pairs using contrastive learning. OpenCLIP [9] builds upon CLIP by leveraging public datasets such as LAION [39]. Despite their strong zero-shot capabilities, these models are pretrained at the image level for classification. This presents a notable limitation: the attention maps of VLMs tend to overemphasize outliers rather than semantic image tokens, leading to insufficient capture of spatial details. This draw + +back hinders the performance in dense prediction tasks like semantic segmentation, which demands precise pixel-level understanding. Our goal is to adaptively detect and mitigate the impact of these outliers and enhance the semantic feature representation for dense prediction tasks. + +# 2.2. Open-vocabulary Segmentation + +Semantic segmentation involves the pixel-wise classification of an image. Compared with the traditional segmentation methods [6, 8, 31, 49] that are trained and evaluated on the same fixed set of seen classes. OVSS methods [36, 59] aim to recognize and segment unseen objects during inference by harnessing the zero-shot ability of VLMs. Existing works can be broadly divided into two categories: training-based and training-free. + +Training-based Methods Training-based methods [18, 37, 56] typically require fine-tuning on a fixed set of categories from a pixel-level annotated dataset. Some approaches follow a two-stage pipeline, where the first stage extracts the mask proposals, and the second stage assigns semantic labels to them. For instance, OVSeg [51] first trains class-agnostic mask proposals using the query-based framework Mask2Former [8] and then fine-tunes CLIP [38] to classify the cropped and masked images. Other approaches adopt a single-stage pipeline. SAN [52] introduces a side adapter to adapt CLIP [38] for both classification and segmentation. SED [48] presents a simple encoder-decoder architecture with category early rejection for fast inference. CATSeg [10] constructs a pixel-level cost map for segmentation. However, these training-based methods are prone to overfitting the training dataset, which limits the generalization ability due to fine-tuning CLIP. + +Training-free Methods Unlike training-based methods, training-free methods [20, 41] aim to directly adapt VLMs for OVSS without extra training. Several approaches leverage the vision foundation models [5, 21, 42] to improve CLIP's spatial coherence. LaVG [19] uses DINO [5] for panoptic cuts, which iteratively discovers object masks until all image pixels are covered. ProxyCLIP [25] adjusts attention weights based on affinities learned from DINO's feature correspondence. However, these methods fail to exploit CLIP's inherent potential. Other efforts aim to modify CLIP's final layer to address its poor spatial consistency. For example, MaskCLIP [14] only utilizes the value feature of the final layer for prediction. CLIPtrase [40] integrates self-self attention with clustering for post-processing. SCLIP [45] and ClearCLIP [24] both modify the last layer self-attention mechanism to better represent the local information. Nevertheless, few training-free methods address propagated outliers, which act as noisy inputs across multiple layers and degrade spatial perception. In contrast, our approach enables adaptive detection and mitigation of propagated outliers among various layers, which improves + +CLIP's semantic awareness and boosts its internal potential. + +# 3. Method + +# 3.1. Overview + +Fig. 2 depicts the whole framework of SFP, including four main modules: a frozen CLIP backbone to encode the input image and text descriptions, a SOM to detect and eliminate outliers, an SAE to augment the attention intensity on related semantic regions, and an HAI that leverages intermediate-layer attention maps to assist SAE for enhanced semantic feature representation. The complete inference process is as follows: + +1) First, the image and text descriptions are input to the CLIP encoder for the image feature and the text feature. Among the CLIP image encoder, our SOM is added after each transformer layer to mitigate the outliers and generate the purified feature for propagation. +2) Then, the derived final feature map from the last SOM, $F_{SOM}$ , is input to our SAE, which replaces the final transformer layer. SAE emphasizes the attention values of relevant semantic regions to generate the strengthened semantic feature $F_{SAE}$ . +3) After that, our HAI extracts the multi-layer attention maps from the image encoder to further refine $F_{SAE}$ for the final semantic feature $F_{vis}$ . +4) Finally, the segmentation result is obtained by calculating the cosine similarity of the final semantic feature $F_{vis}$ and text feature $F_{txt}$ . + +# 3.2. Preliminary + +In general, the training-free OVSS approach requires CLIP [38], which encompasses an image encoder and a text encoder to acquire multi-model features. Subsequently, these features are aligned via linear projection layers. In detail, the input image is first partitioned into a token sequence $X_{in} = [x_{cls}, x_1, \dots, x_N] \in R^{(N + 1) \times d}$ , where $x_i$ denotes the image token embedding, $x_{cls}$ denotes the class token embedding, $N$ is the image token sequence length, and $d$ is the dimension of token embeddings, respectively. Then, these tokens are input to the image encoder, which contains a stack of vision transformer layers. + +Each transformer layer consists of a multi-head self-attention (MHSA) mechanism and a feed-forward network (FFN) layer to extract the image feature. Specifically, in each MHSA, the input sequence $X_{in}$ is mapped to query, key, and value embeddings, $Q, K, V \in R^{(N + 1)\times d}$ . Afterwards, $Q$ and $K$ are calculated to obtain the attention map $Attn \in R^{(N + 1)\times (N + 1)}$ , which can be formulated as: + +$$ +A t t n = \operatorname {S o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {d}}\right), \tag {1} +$$ + +where $T$ means matrix transposition. After that, the attention map is multiplied $V$ to aggregate the global infor + +![](images/df20bce62542ccaf74ec7943fc6e61eeec8bbe555389af3e25a8aaa3c3419022.jpg) +Figure 2. Framework of our SFP. It contains four main modules: a frozen CLIP backbone for feature extraction, a Self-adaptive Outlier Mitigator (SOM) at each image encoder layer for outlier detection and elimination, a Semantic-aware Attention Enhancer (SAE) at the last layer to emphasise attention values of relevant semantic regions and a Hierarchical Attention Integrator (HAI) for object-centric feature representation. The segmentation result is generated by calculating the cosine similarity between the optimized semantic feature $F_{vis}$ and the text feature $F_{txt}$ . + +mation, and an FFN is used to obtain the image feature $F_{img} \in R^{(N + 1) \times d}$ , formulated as: + +$$ +F _ {i m g} = \operatorname {F F N} (A t t n V). \tag {2} +$$ + +Since the class embedding $x_{cls}$ is not used to predict the segmentation result, we omit it and reshape $F_{img}$ to obtain the visual feature $F_{vis} \in R^{h \times w \times d}$ , where $h$ and $w$ denote the height and width of the feature map. + +Meanwhile, the textual description is derived from the standard ImageNet [13] prompts and transformed to the text feature $F_{txt} \in R^{c \times d}$ with CLIP's text encoder, where $c$ represents the number of classes. By calculating the cosine similarity between $F_{vis}$ and $F_{txt}$ , the segmentation map $Seg \in R^{H \times W}$ can be generated like: + +$$ +S e g = U (\underset {c} {\arg \max } \cos (F _ {v i s}, F _ {t x t})), \tag {3} +$$ + +where $\cos (\cdot)$ means cosine similarity calculation and $U(\cdot)$ denotes the upsampling operation to match the original input image size. $H$ and $W$ denote the height and width of the input image, respectively. + +# 3.3. Self-adaptive Outlier Mitigator + +CLIP [38] is pre-trained at the image level for classification, which tends to concentrate on the most discriminative feature that is important for classification. In this way, all input + +tokens, including both class and image tokens, will emphasize these identical tokens (outliers) in attention maps [12]. In other words, the corresponding outlier column in the attention map will be highlighted. Further, given that the classification process relies solely on the class token, the class token may exhibit a stronger response to the outlier in attention maps compared to image tokens. This prompts us to investigate whether the self-relevance of image tokens and the class token's similarity across all image tokens could serve as potential indicators to identify outliers. + +Based on the above assumption, to robustly mitigate outliers, we introduce a Self-adaptive Outlier Mitigator (SOM) to detect and eliminate the outliers automatically. In particular, we use the attention map $Attn$ to obtain the self-attention weights of image tokens, denoted as $Attn_{i,i}$ , and their attention weights to the class tokens, denoted as $Attn_{cls,i}$ , where $i$ denotes the $i$ -th image token and $cls$ denotes the class token. By comparing the difference between $Attn_{i,i}$ and $Attn_{cls,i}$ , we obtain the set of outliers $S$ , since the self-attention weights should be higher than others. This detection process of SOM is formulated as, + +$$ +\mathcal {S} = \left\{i \mid A t t n _ {i, i} < A t t n _ {c l s, i} \right\}, \quad i \in [ 1, N ], \tag {4} +$$ + +where $S$ denotes the set of outliers. + +After recognizing outliers, we propose eliminating them at the feature level rather than modifying the attention map + +to prevent potential model collapse. Inspired by previous work [2], it's crucial to ensure the semantic consistency of the feature map, and directly masking outliers at the feature level is not an appropriate approach. Thus, SOM can eliminate outliers through in-place merging, where the detected outliers are replaced with the average value of its 8-neighbours. That is to say, the derived $F_{img}$ in the original attention mechanism will be updated through our SOM as + +$$ +\hat {F} _ {i m g} [ u ] = \left\{ \begin{array}{l l} \frac {1}{8} \sum_ {v \in \mathcal {N} _ {8} (u)} F _ {i m g} [ v ], & u \in \mathcal {S} \\ F _ {i m g} [ u ], & \text {o t h e r w s i e} \end{array} , \right. \tag {5} +$$ + +where $\mathcal{N}_8(u)$ means the 8-neighbour non-outlier tokens of the $u$ -th point and $\hat{F}_{img}$ denotes the updated feature. In this way, our SOM initially identifies the outlier and then leverages surrounding token features to replace the detected one. This process can mitigate the adverse influence exerted by the outliers and maintain the feature distribution that is crucial for CLIP's image encoder. By doing so, the purified and superior semantic feature representation is generated. + +Furthermore, our SOM exhibits remarkable versatility in that it can be applied to each individual layer of the image encoder. This allows for the detection of outliers and the generation of purified features at multiple levels. In this manner, our parameter-free SOM establishes an efficient and precise mechanism for detecting and eliminating outliers that propagate across various layers of the image encoder. This mechanism is instrumental in preserving accurate semantic information and feature representations throughout the entire image encoder, thereby enhancing its overall performance and reliability. + +![](images/b89618275e0154c0991ecd39b58a4ca196878a512fcf7a2ac266bba76f11ac4a.jpg) +(a) +Figure 3. Comparison of different feature visualization. (a) Noisy feature with outliers. (b) Purified feature with SOM. (c) Semantic feature with SAE. It's seen that the proposed SAE strengthens the semantic regions with higher object activation. + +![](images/a0e4641041202a6a68fa1d214a2ba2478afbe719137ad550d2194c14bf9c9d49.jpg) +(b) + +![](images/afa714bc38ba978a4f3c0f86f18f12e359170e7e2c0e24253b209ff14e8dc4da.jpg) +(c) + +# 3.4. Semantic-aware Attention Enhancer + +The comparison of Fig. 3 (a) & (b) indicates that the purified feature $F_{SOM}$ generated by the last SOM effectively ignores the irrelevant regions of outliers, marked as white boxes. However, it's also observed from Fig. 3 (b) & (c) that the object activation of $F_{SOM}$ is not salient enough, + +which indicates that eliminating outliers can't contribute to the attention response intensity. + +To address this issue, we propose a Semantic-aware Attention Enhancer (SAE) to strengthen the object activation in CLIP's last layer, where the self-self attention mechanism is used to derive better feature relationships, + +$$ +A t t n _ {S A E} = \operatorname {S o f t m a x} \left(\lambda \left(\frac {Q Q ^ {T}}{\sqrt {d}} + \frac {K K ^ {T}}{\sqrt {d}}\right)\right), \tag {6} +$$ + +where $Attn_{SAE}$ represents the semantic-aware attention map and $\lambda$ denotes the logit scale used to sharpen attention scores, set to 0.5. Then, the semantic feature $F_{SAE}$ is obtained through the residual connection with the purified feature $F_{SOM}$ as follows: + +$$ +F _ {S A E} = F _ {S O M} + \operatorname {F F N} \left(\operatorname {A t t n} _ {S A E} F _ {S O M}\right). \tag {7} +$$ + +The self-self attention mechanism has shown a strong ability to construct feature relationships [16, 24, 45]. Therefore, applying self-self attention to our purified feature enables the establishment of more precise spatial correlations. These accurate spatial correlations play a crucial role in enabling each point to aggregate the relevant semantic feature effectively. As illustrated in Fig. 3 (b) & (c), SAE significantly augments attention intensity in semantic regions compared to the feature solely processed by SOM. + +# 3.5. Hierarchical Attention Integrator + +To further refine the generated semantic feature $F_{SAE}$ , we leverage attention maps from shallow layers to capture discriminative pair-wise feature relationships since shallow layers typically focus on object structures [27]. To this end, we propose a Hierarchical Attention Integrator (HAI), which integrates multi-layer attention maps to refine $F_{SAE}$ . + +Instead of using the original attention maps with outliers, HAI integrates attention maps from each transformer layer while excluding the detected outlier weights to avoid information pollution. Firstly, HAI employs the outlier set $S$ to select the non-outlier weights of each attention map, which can be formulated as, + +$$ +\overline {{A t t n}} _ {j, i} ^ {l} = \left\{ \begin{array}{l l} A t t n _ {j, i}, & \text {i f} i \notin \mathcal {S} ^ {l} \\ 0 & \text {o t h e r w i s e} \end{array} , \right. \tag {8} +$$ + +where $\overline{Attn}^l$ and $S^l$ denote the filtered attention map and the outlier set at the $l$ -th layer. $j$ and $i$ correspond to the row and column indices of the attention map. Next, HAI integrates $\overline{Attn}^l$ from each layer, formulated as, + +$$ +A = \sum_ {l = 1} ^ {L - 1} \overline {{A t t n}} ^ {l}, \tag {9} +$$ + +where $A$ denotes the integrated attention map, and $L$ denotes the total number of transformer layers. + +
MethodsPub. & YearExtra backboneTraining freewith backgroundV20without backgroundAvg.
V21PC60ObjectPC59StuffADECity
GroupVit [50]CVPR'22XX50.418.727.579.723.415.39.211.129.4
SegCLIP [32]ICML'23XX52.624.726.5------
TCL [7]CVPR'23XX55.030.431.683.233.922.417.124.037.2
DINOiser† [47]ECCV'24DINOX62.132.434.880.935.924.620.031.740.3
SAM-CLIP† [46]ECCV'24SAMX60.629.2----17.1--
PnP-OVSS† [33]CVPR'24BLIP✓--36.251.328.017.914.2--
LaVG† [19]ECCV'24DINO✓62.131.634.282.534.723.215.826.238.8
ProxyCLIP† [25]ECCV'24DINO✓61.335.337.580.339.126.520.238.142.3
CLIP [38]ICML'21X✓18.69.98.149.411.15.73.16.514.1
MaskCLIP [14]ECCV'22X✓38.823.620.674.926.416.49.812.627.9
CLIPSurgery [28]PR'25X✓-29.3---21.9-31.4-
CaR [44]CVPR'24X✓48.630.536.673.739.5-17.7--
GEM [3]CVPR'24X✓46.2---32.615.7---
CLIPtrase [40]ECCV'24X✓50.929.943.681.033.822.816.421.337.5
ClearCLIP* [24]ECCV'24X✓57.032.633.080.935.923.916.730.038.8
SCLIP* [45]ECCV'24X✓59.731.733.581.534.522.716.532.339.1
NACLIP* [16]WACV'25X✓58.932.233.279.735.223.317.435.539.4
SFP (Ours)X✓63.937.237.984.539.926.420.841.144.0
+ +Table 1. Quantitative comparison of our SFP against other approaches across eight segmentation benchmark datasets. The best and second-best results are marked with **bold** and **underline**, respectively. † indicates that the additional backbone settings follow the original paper. * indicates the results are re-evaluated using the official implementation. + +After integrating multi-layer attention maps, we deploy a normalization strategy to optimize $A$ instead of simply averaging it. Inspired by previous works [22, 35], we iteratively apply row-wise and column-wise normalization to $A$ for a more object-centric attention distribution, leading to a stronger focus on relevant regions. This process can be formulated as, + +$$ +A _ {H A I} = \operatorname {n o r m} _ {c} \left(\operatorname {n o r m} _ {r} (A)\right), \tag {10} +$$ + +where $A_{HAI}$ denotes the normalized attention map, normc and normr denote the column-wise and row-wise normalization respectively. + +Then, $A_{HAI}$ is multiplied with the semantic feature $F_{SAE}$ to obtain the final semantic feature $F_{vis}$ for segmentation, which is formulated as follows: + +$$ +F _ {v i s} = A _ {H A I} F _ {S A E}. \tag {11} +$$ + +Finally, we compute the cosine similarity between $F_{vis}$ and the text feature $F_{txt}$ extracted from the CLIP text encoder to predict the segmentation map, as formulated in Eq. 3. + +# 4. Experiment + +# 4.1. Experimental Settings + +Datasets We conduct comprehensive experiments to verify the effectiveness of our method on diverse benchmark datasets: PASCAL VOC 2012 [15], PASCAL Context [34], and COCO [29]. These datasets are categorized into two settings: 1) with a background class, namely V20, PC59, + +and Stuff; 2) without the background class, namely V21, PC60, and Object. We also report results on the ADE20K (ADE) [57] and Cityscapes (City) [11] datasets. + +Implementation Details Similar to existing training-free methods, we adopt CLIP ViT-B/16 with the original pretrained weight [38]. Following previous works [2, 45], we combine standard ImageNet prompts [13] with the class names to construct text descriptions. For a fair comparison with previous works [25, 45], we resize images with a short side of 336 pixels (or 560 pixels for the high-resolution Cityscapes dataset) and apply sliding window inference using a $224 \times 224$ window with a $112 \times 112$ stride. Our experiments are conducted using one NVIDIA 3090 GPU. + +# 4.2. Comparison with State-of-the-art Methods + +Our SFP is evaluated against existing SOTA methods, including both training-based and training-free approaches. The post-processing mask refinement techniques, such as PAMR [1] or DenseCRF [23], are not included for a fair comparison. The standard mean intersection over union (mIoU) is used as the metric for evaluation. + +Quantitative Evaluation The main results are summarized in Tab. 1, where all evaluations are conducted using CLIP ViT-B/16 as the backbone to ensure a fair comparison. We compare SFP against multiple methods, including training-based methods, training-free methods with extra backbones, and training-free methods without extra backbones. + +As shown in Tab. 1, our SFP outperforms existing SOTA methods, regardless of whether an additional backbone is introduced or not. In particular, compared with existing + +![](images/456adbda35fa7d9096eb56eb60121eaea6067ac9850e0c307f84672bba38fb36.jpg) +Figure 4. Qualitative Results Comparison. We compare our SFP with CLIP [38], ClearCLIP [24], SCLIP [45] and NACLIP [16] without post-processing. It's observed that SFP generates more precise segmentation results. + +training-free methods without extra backbones [16, 24, 45] or cluster methods [53-55], our SFP achieves an average gain of $4.6\%$ mIoU over previous SOTA NACLIP [16]. Specifically, On the V20 and V21 dataset [15], SFP surpasses NACLIP [16] $4.8\%$ and $5\%$ mIoU, respectively. These improvements indicate that our SFP can generate more purified feature maps for more precise segmentation. Beyond that, SFP outperforms SOTA training-free methods with extra backbones, e.g., ProxyCLIP [25], with a gain of $3.0\%$ mIoU on the Cityscape [11] dataset and an average gain of $1.7\%$ mIoU on eight datasets. Our SFP can even beat DINOiser [47], which requires both training and an extra backbone, with an average improvement of $3.7\%$ mIoU. These results highlight our SFP boosts CLIP's internal potential and reveal that feature purification does matter in training-free OVSS, where more purified features can predict better segmentation masks. + +Qualitative Evaluation In Fig. 4, we compare the visualization results of various training-free methods without extra backbones, including CLIP [38], ClearCLIP [24], SCLIP [45], NACLIP [16] and our SFP. In particular, the segmentation maps obtained by vanilla CLIP [38] exhibit significant noises, unveiling its limitations in spatial localization. Besides, in complex scenarios, like 'sofa' in the second row, SCLIP [45] and NACLIP [16] fail to capture the complete segmentation mask. In contrast, our SFP can predict more accurate and complete segmentation maps. These results indicate that our SFP has a good ability to predict comprehensive segmentation masks. + +# 4.3. Ablation Study + +Comprehensive ablation studies are performed to analyze the effectiveness of our proposed mechanisms. Unless otherwise specified, we utilize CLIP with ViT-B/16 backbone and evaluate on V21 [15] and ADE [57] datasets for background and non-background settings, respectively. + +Effect on each proposed module We first conduct several experiments to validate the effectiveness of our proposed modules of our SFP: SOM, SAE, and HAI by adding each module step by step, where ClearCLIP [24] is used as the baseline. As shown in Tab. 2, our SOM achieves a gain of $3.1\%$ and $1.5\%$ mIoU on V21 and ADE datasets, respectively. Further, SAE enhances attention intensity in relevant regions, leading to an average gain of $1.7\%$ mIoU. Finally, HAI helps our SFP reach $63.9\%$ and $20.8\%$ mIoU on V21 and ADE datasets, which set the new SOTA performance. Due to space constraints, more ablation studies about SOM and SAE are summarized in the supplementary material. + +
SOMSAEHAIV21ADEAvg.
057.016.736.9
1✓60.118.239.2
2✓✓62.319.540.9
3✓✓✓63.920.842.4
+ +Table 2. Ablation results on each proposed module to validate the impact on performance across V21 [15] and ADE [57] datasets. + +More visualization on the ability of SOM to suppress outlier propagation We visualize purified feature maps + +![](images/a306f95aad08c2c2619221750b1d1304fdc64beae124dc2ea0e33db12deccce1.jpg) +Figure 5. Visualization of the feature purification across CLIP's image encoder layers. (a) Feature maps without SOM purification. (b) Feature maps purified by SOM. + +
Al=l-1l=lo1581011
V2163.963.463.062.662.5
ADE20.820.520.319.919.7
Avg.42.442.041.741.341.1
+ +across CLIP's layers, from the 2nd layer to the 9th layer, which is shown in Fig. 5. It's found that the outliers gradually affect the aggregated features, especially from layer 6 to layer 9 in Fig. 5 (a). Then, with the help of our SOM, the influence of outliers is mitigated layer by layer, like in Fig. 5 (b). Thus, our SOM can suppress the propagated outlier features and gradually purify the overall semantic features. + +Effect on hierarchical attention integrator A series of ablation studies is conducted to investigate the optimal way to leverage attention maps from different layers. We omit HAI from our SFP to establish the baseline model for this experiment. As summarized in Tab. 3, we take $l_{0}$ to represent the starting layer. When $l_{0} = 1$ , our HAI performs best with all attention maps selected. Besides, increasing $l_{0}$ from 1 to 11 decreases the segmentation performance, which demonstrates the effectiveness of the shallow attention maps for semantic feature refinement. + +Table 3. Ablation results on different selection strategies in HAI, where $A_{l = l_0}^{l = L - 1}$ represents the integrated attention map from $l_0$ to the $L - 1$ layer. '1,5,8,10,11' indicates the value of $l_0$ . + +
BackboneMethodsV21ADEAvg.
ViT-B/32SCLIP [45]50.614.832.7
LaVG [19]54.815.535.2
ProxyCLIP [25]57.916.737.3
NACLIP [16]51.114.933.0
SFP60.117.138.6
ViT-L/14SCLIP [45]44.410.927.7
LaVG [19]52.117.334.7
ProxyCLIP [25]60.622.641.6
NACLIP [16]52.217.334.8
SFP64.722.143.4
+ +Table 4. Comparison with other SOTA methods across different CLIP backbones. + +Effect on different CLIP backbones To validate the robustness, we further evaluate our SFP with other methods using different CLIP backbones, such as ViT-B/32 and ViT-L/14. As summarized in Tab. 4, some methods, like SCLIP [45] and LaVG [19], suffer from performance degradation when applied to the larger backbone. In contrast, SFP maintains superior performance across different backbones and achieves an improvement of $1.3\%$ and $1.8\%$ mIoU over ProxyCLIP [25] with ViT-B/32 and ViT-L/14 backbone, respectively. These results indicate that SFP can robustly handle outliers derived from different backbones and generate purified features for precise segmentation. + +# 5. Conclusion + +In this paper, we present a Self-adaptive Feature Purifier framework (SFP) to mitigate outlier propagation for training-free open-vocabulary semantic segmentation. Specifically, we propose SOM deployed at each layer of the image encoder to identify and mitigate outliers by comparing attention weights between image and class tokens, which generates purified features for propagation. Next, to recover the semantic information affected by outliers, we introduce SAE to strengthen object activation in the purified feature. Beyond that, we design HAI to utilize multilayer attention maps for refined object-centric representations. Extensive experiments show that SFP achieves superior segmentation performance compared to existing SOTA methods on various datasets. Our SFP shows that feature purification does matter for OVSS, and cleaner features can predict more complete and accurate segmentation masks. + +Acknowledgements. This work was supported by the National Natural Science Foundation of China (No. 62471405, 62331003, 62301451,62301613), Basic Research Program of Jiangsu (BK20241814), Suzhou Basic Research Program (SYG202316) and XJTLU REF-22-01-010 and RDF-22-02-066, XJTLU AI University Research Centre, Jiangsu Province Engineering Research Centre of Data Science and Cognitive Computation at XJTLU. + +# References + +[1] Nikita Araslanov and Stefan Roth. Single-stage semantic segmentation from image labels. In CVPR, pages 4253-4262, 2020. 6 +[2] Sule Bai, Yong Liu, Yifei Han, Haoji Zhang, and Yansong Tang. Self-calibrated clip for training-free open-vocabulary segmentation. arXiv preprint arXiv:2411.15869, 2024. 2, 5, 6 +[3] Walid Bousselham, Felix Petersen, Vittorio Ferrari, and Hilde Kuehne. Grounding everything: Emerging localization properties in vision-language transformers. In CVPR, pages 3828-3837, 2024. 6 +[4] Maxime Bucher, Tuan-Hung Vu, Matthieu Cord, and Patrick Pérez. Zero-shot semantic segmentation. In NeurlPS, pages 468-479, 2019. 1 +[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. 3 +[6] Niccolo Cavagnero, Gabriele Rosi, Claudia Cuttano, Francesca Pistilli, Marco Ciccone, Giuseppe Avera, and Fabio Cermelli. Pem: Prototype-based efficient maskformer for image segmentation. In CVPR, pages 15804-15813, 2024. 3 +[7] Junbum Cha, Jonghwan Mun, and Byungseok Roh. Learning to generate text-grounded mask for open-world semantic segmentation from only image-text pairs. In CVPR, pages 11165-11174, 2023. 6 +[8] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, pages 1290–1299, 2022. 3 +[9] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, pages 2818-2829, 2023. 2 +[10] Seokju Cho, Heeseong Shin, Sunghwan Hong, Anurag Arnab, Paul Hongsuck Seo, and Seungryong Kim. Cat-seg: Cost aggregation for open-vocabulary semantic segmentation. In CVPR, pages 4113-4123, 2024. 3 +[11] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, pages 3213-3223, 2016. 6, 7 +[12] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. In ICLR, pages 1-11, 2024. 1, 2, 4 +[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, pages 248-255, 2009. 4, 6 +[14] Zheng Ding, Jieke Wang, and Zhuowen Tu. Open-vocabulary panoptic segmentation with maskclip. arXiv preprint arXiv:2208.08984, 2022. 2, 3, 6 +[15] Mark Everingham, SM Ali Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The + +pascal visual object classes challenge: A retrospective. IJCV, 111:98-136, 2015. 6, 7 +[16] Sina Hajimiri, Ismail Ben Ayed, and Jose Dolz. Pay attention to your neighbours: Training-free open-vocabulary semantic segmentation. In WACV, pages 5061-5071, 2025. 2, 5, 6, 7, 8 +[17] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916, 2021. 2 +[18] Siyu Jiao, Yunchao Wei, Yaowei Wang, Yao Zhao, and Humphrey Shi. Learning mask-aware clip representations for zero-shot segmentation. In NeurIPS, pages 35631-35653, 2022. 3 +[19] Dahiyun Kang and Minsu Cho. In defense of lazy visual grounding for open-vocabulary semantic segmentation. In ECCV, pages 143-164, 2024. 3, 6, 8 +[20] Laurynas Karazija, Iro Laina, Andrea Vedaldi, and Christian Rupprecht. Diffusion models for zero-shot open-vocabulary segmentation. In ECCV, pages 299-317, 2024. 3 +[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, pages 4015-4026, 2023. 3 +[22] Philip A Knight. The sinkhorn-knopp algorithm: convergence and applications. SIAM Journal on Matrix Analysis and Applications, 30(1):261-275, 2008. 6 +[23] Philipp Kähenbuhl and Vladlen Koltun. Efficient inference in fully connected crfs with gaussian edge potentials. In NeurIPS, pages 109-117, 2011. 6 +[24] Mengcheng Lan, Chaofeng Chen, Yiping Ke, Xinjiang Wang, Litong Feng, and Wayne Zhang. Clearclip: Decomposing clip representations for dense vision-language inference. In ECCV, pages 143-160, 2024. 2, 3, 5, 6, 7 +[25] Mengcheng Lan, Chaofeng Chen, Yiping Ke, Xinjiang Wang, Litong Feng, and Wayne Zhang. Proxyclip: Proxy attention improves clip for open-vocabulary segmentation. In ECCV, pages 70-88, 2024. 3, 6, 7, 8 +[26] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900, 2022. 2 +[27] Yunheng Li, Zhongyu Li, Quansheng Zeng, Qibin Hou, and Ming-Ming Cheng. Cascade-clip: Cascaded vision-language embeddings alignment for zero-shot semantic segmentation. In ICML, pages 28243–28258, 2024. 2, 5 +[28] Yi Li, Hualiang Wang, Yiqun Duan, Jiheng Zhang, and Xiaomeng Li. A closer look at the explainability of contrastive language-image pre-training. Pattern Recognition, 162:111409, 2025. 6 +[29] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 6 +[30] Yong Liu, Sule Bai, Guanbin Li, Yitong Wang, and Yansong Tang. Open-vocabulary segmentation with semantic-assisted calibration. In CVPR, pages 3491-3500, 2024. 1, 2 + +[31] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, pages 3431-3440, 2015. 3 +[32] Huaishao Luo, Junwei Bao, Youzheng Wu, Xiaodong He, and Tianrui Li. Segclip: Patch aggregation with learnable centers for open-vocabulary semantic segmentation. In ICML, pages 23033-23044, 2023. 6 +[33] Jiayun Luo, Siddhesh Khandelwal, Leonid Sigal, and Boyang Li. Emergent open-vocabulary semantic segmentation from off-the-shelf vision-language models. In CVPR, pages 4029-4040, 2024. 6 +[34] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The role of context for object detection and semantic segmentation in the wild. In CVPR, pages 891–898, 2014. 6 +[35] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6 +[36] Zelin Peng, Zhengqin Xu, Zhilin Zeng, Yu Huang, Yaoming Wang, and Wei Shen. Parameter-efficient fine-tuning in hyperspherical space for open-vocabulary semantic segmentation. In CVPR, pages 15009-15020, 2025. 3 +[37] Zelin Peng, Zhengqin Xu, Zhilin Zeng, Changsong Wen, Yu Huang, Menglin Yang, Feilong Tang, and Wei Shen. Understanding fine-tuning clip for open-vocabulary semantic segmentation in hyperbolic space. In CVPR, pages 4562-4572, 2025. 3 +[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 1, 2, 3, 4, 6, 7 +[39] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, pages 25278–25294, 2022. 1, 2 +[40] Tong Shao, Zhuotao Tian, Hang Zhao, and Jingyong Su. Explore the potential of clip for training-free open vocabulary semantic segmentation. In ECCV, pages 139-156, 2024. 1, 2, 3, 6 +[41] Gyungin Shin, Weidi Xie, and Samuel Albanie. Reco: Retrieve and co-segment for zero-shot transfer. In NeurIPS, pages 33754-33767, 2022. 3 +[42] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, pages 2256-2265, 2015. 3 +[43] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In ICCV, pages 7262-7272, 2021. 1 +[44] Shuyang Sun, Runjia Li, Philip Torr, Xiuye Gu, and Siyang Li. Clip as rn: Segment countless visual concepts without training endeavor. In CVPR, pages 13171-13182, 2024. 6 + +[45] Feng Wang, Jieru Mei, and Alan Yuille. Sclip: Rethinking self-attention for dense vision-language inference. In ECCV, pages 315–332, 2024. 1, 2, 3, 5, 6, 7, 8 +[46] Haoxiang Wang, Pavan Kumar Anasosalu Vasu, Fartash Faghri, Raviteja Vemulapalli, Mehrdad Farajtabar, Sachin Mehta, Mohammad Rastegari, Oncel Tuzel, and Hadi Pouransari. Sam-clip: Merging vision foundation models towards semantic and spatial understanding. In CVPR, pages 3635–3647, 2024. 6 +[47] Monika Wysoczanska, Oriane Simeoni, Michael Ramamonjisoa, Andrei Bursuc, Tomasz Trzciński, and Patrick Pérez. Clip-dinoiser: Teaching clip a few dino tricks. In ECCV, pages 320-337, 2024. 6, 7 +[48] Bin Xie, Jiale Cao, Jin Xie, Fahad Shahbaz Khan, and Yanwei Pang. Sed: A simple encoder-decoder for open-vocabulary semantic segmentation. In CVPR, pages 3426-3436, 2024. 3 +[49] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. In NeurIPS, pages 12077-12090, 2021. 3 +[50] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. Groupvit: Semantic segmentation emerges from text supervision. In CVPR, pages 18134-18144, 2022. 6 +[51] Jilan Xu, Junlin Hou, Yuejie Zhang, Rui Feng, Yi Wang, Yu Qiao, and Weidi Xie. Learning open-vocabulary semantic segmentation models from natural language supervision. In CVPR, pages 2935-2944, 2023. 3 +[52] Mengde Xu, Zheng Zhang, Fangyun Wei, Han Hu, and Xiang Bai. Side adapter network for open-vocabulary semantic segmentation. In CVPR, pages 2945-2954, 2023. 3 +[53] Xihong Yang, Xiaochang Hu, Sihang Zhou, Xinwang Liu, and En Zhu. Interpolation-based contrastive learning for few-label semi-supervised learning. TNNLS, 35(2):2054-2065, 2022. 7 +[54] Xihong Yang, Jin Jiaqi, Siwei Wang, Ke Liang, Yue Liu, Yi Wen, Suyuan Liu, Sihang Zhou, Xinwang Liu, and En Zhu. Dealmvc: Dual contrastive calibration for multi-view clustering. In ACMMM, pages 337-346, 2023. +[55] Xihong Yang, Siwei Wang, Fangdi Wang, Jiaqi Jin, Suyuan Liu, Yue Liu, En Zhu, Xinwang Liu, and Yueming Jin. Automatically identify and rectify: Robust deep contrastive multiview clustering in noisy scenarios. In ICML. PMLR, 2025. 7 +[56] Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang-Chieh Chen. Convolutions die hard: Open-vocabulary segmentation with single frozen convolutional clip. In NeurIPS, pages 32215-32234, 2023. 3 +[57] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ADE20K dataset. In CVPR, pages 5122-5130, 2017. 6, 7 +[58] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In ECCV, pages 696-712, 2022. 1 +[59] Ziqin Zhou, Yinjie Lei, Bowen Zhang, Lingqiao Liu, and Yifan Liu. Zegclip: Towards adapting clip for zero-shot semantic segmentation. In CVPR, pages 11175-11185, 2023. \ No newline at end of file diff --git a/2025/Feature Purification Matters_ Suppressing Outlier Propagation for Training-Free Open-Vocabulary Semantic Segmentation/images.zip b/2025/Feature Purification Matters_ Suppressing Outlier Propagation for Training-Free Open-Vocabulary Semantic Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c86d3430c3c8d535ddd90e528842f8e4f862656a --- /dev/null +++ b/2025/Feature Purification Matters_ Suppressing Outlier Propagation for Training-Free Open-Vocabulary Semantic Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:54e730fed59f6adb5d37c8e9cf477fb688d4c07855cbd6f642680c77402f13f3 +size 693505 diff --git a/2025/Feature Purification Matters_ Suppressing Outlier Propagation for Training-Free Open-Vocabulary Semantic Segmentation/layout.json b/2025/Feature Purification Matters_ Suppressing Outlier Propagation for Training-Free Open-Vocabulary Semantic Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..aa0f5188e70261f6ad784c8ea067ea1a5e925ccc --- /dev/null +++ b/2025/Feature Purification Matters_ Suppressing Outlier Propagation for Training-Free Open-Vocabulary Semantic Segmentation/layout.json @@ -0,0 +1,9571 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 100, + 103, + 511, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 103, + 511, + 140 + ], + "spans": [ + { + "bbox": [ + 100, + 103, + 511, + 140 + ], + "type": "text", + "content": "Feature Purification Matters: Suppressing Outlier Propagation for Training-Free Open-Vocabulary Semantic Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 90, + 160, + 516, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 160, + 516, + 230 + ], + "spans": [ + { + "bbox": [ + 90, + 160, + 516, + 230 + ], + "type": "text", + "content": "Shuo Jin" + }, + { + "bbox": [ + 90, + 160, + 516, + 230 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 90, + 160, + 516, + 230 + ], + "type": "text", + "content": " Siyue Yu" + }, + { + "bbox": [ + 90, + 160, + 516, + 230 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 90, + 160, + 516, + 230 + ], + "type": "text", + "content": " Bingfeng Zhang" + }, + { + "bbox": [ + 90, + 160, + 516, + 230 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 90, + 160, + 516, + 230 + ], + "type": "text", + "content": " Mingjie Sun" + }, + { + "bbox": [ + 90, + 160, + 516, + 230 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 90, + 160, + 516, + 230 + ], + "type": "text", + "content": " Yi Dong" + }, + { + "bbox": [ + 90, + 160, + 516, + 230 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 90, + 160, + 516, + 230 + ], + "type": "text", + "content": " Jimin Xiao" + }, + { + "bbox": [ + 90, + 160, + 516, + 230 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 90, + 160, + 516, + 230 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 90, + 160, + 516, + 230 + ], + "type": "text", + "content": "Xi'an Jiaotong-Liverpool University " + }, + { + "bbox": [ + 90, + 160, + 516, + 230 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 90, + 160, + 516, + 230 + ], + "type": "text", + "content": "University of Liverpool \n" + }, + { + "bbox": [ + 90, + 160, + 516, + 230 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 90, + 160, + 516, + 230 + ], + "type": "text", + "content": "China University of Petroleum (East China) " + }, + { + "bbox": [ + 90, + 160, + 516, + 230 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 90, + 160, + 516, + 230 + ], + "type": "text", + "content": "Soochow University \n{shuo.jin, yi.dong}@liverpool.ac.uk, {siyue.yu02, jimin.xiao}@xjtlu.edu.cn \nbingfeng.zhang@upc.edu.cn, mjsun@suda.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 284, + 296, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 284, + 296, + 619 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 296, + 619 + ], + "type": "text", + "content": "Training-free open-vocabulary semantic segmentation has advanced with vision-language models like CLIP, which exhibit strong zero-shot abilities. However, CLIP's attention mechanism often wrongly emphasises specific image tokens, namely outliers, which results in irrelevant over-activation. Existing approaches struggle with these outliers that arise in intermediate layers and propagate through the model, ultimately degrading spatial perception. In this paper, we propose a Self-adaptive Feature Purifier framework (SFP) to suppress propagated outliers and enhance semantic representations for open-vocabulary semantic segmentation. Specifically, based on an in-depth analysis of attention responses between image and class tokens, we design a self-adaptive outlier mitigator to detect and mitigate outliers at each layer for propagated feature purification. In addition, we introduce a semantic-aware attention enhancer to augment attention intensity in semantically relevant regions, which strengthens the purified feature to focus on objects. Further, we introduce a hierarchical attention integrator to aggregate multi-layer attention maps to refine spatially coherent feature representations for final segmentation. Our proposed SFP enables robust outlier suppression and object-centric feature representation, leading to a more precise segmentation. Extensive experiments show that our method achieves state-of-the-art performance and surpasses existing methods by an average of " + }, + { + "bbox": [ + 55, + 284, + 296, + 619 + ], + "type": "inline_equation", + "content": "4.6\\%" + }, + { + "bbox": [ + 55, + 284, + 296, + 619 + ], + "type": "text", + "content": " mIoU on eight segmentation benchmarks. The code is released at: https://github.com/Kimsure/SFP." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 640, + 135, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 640, + 135, + 653 + ], + "spans": [ + { + "bbox": [ + 56, + 640, + 135, + 653 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 660, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 296, + 696 + ], + "type": "text", + "content": "Open-vocabulary semantic segmentation (OVSS) [4, 58] seeks to partition an image into distinct regions and assign pixel-level labels to arbitrary semantic categories. Un" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 259, + 376, + 318 + ], + "blocks": [ + { + "bbox": [ + 317, + 259, + 376, + 318 + ], + "lines": [ + { + "bbox": [ + 317, + 259, + 376, + 318 + ], + "spans": [ + { + "bbox": [ + 317, + 259, + 376, + 318 + ], + "type": "image", + "image_path": "3a809d8d6a21c9db72c8c9f3de6b62c8340aaecec366c51646fb2b91e2579c1c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 318, + 353, + 326 + ], + "lines": [ + { + "bbox": [ + 343, + 318, + 353, + 326 + ], + "spans": [ + { + "bbox": [ + 343, + 318, + 353, + 326 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 377, + 259, + 435, + 318 + ], + "blocks": [ + { + "bbox": [ + 377, + 259, + 435, + 318 + ], + "lines": [ + { + "bbox": [ + 377, + 259, + 435, + 318 + ], + "spans": [ + { + "bbox": [ + 377, + 259, + 435, + 318 + ], + "type": "image", + "image_path": "3114cb73dc14bef72cdbcbae1c9dd2645ba2f83eea07987b45cf6211178fa784.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 318, + 411, + 326 + ], + "lines": [ + { + "bbox": [ + 402, + 318, + 411, + 326 + ], + "spans": [ + { + "bbox": [ + 402, + 318, + 411, + 326 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 435, + 259, + 494, + 318 + ], + "blocks": [ + { + "bbox": [ + 435, + 259, + 494, + 318 + ], + "lines": [ + { + "bbox": [ + 435, + 259, + 494, + 318 + ], + "spans": [ + { + "bbox": [ + 435, + 259, + 494, + 318 + ], + "type": "image", + "image_path": "ec6b2006f99b0a5202453aaa61a917a5f9fe0ac356e732bc8d5e233cb5244a7f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 460, + 318, + 470, + 327 + ], + "lines": [ + { + "bbox": [ + 460, + 318, + 470, + 327 + ], + "spans": [ + { + "bbox": [ + 460, + 318, + 470, + 327 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 495, + 259, + 553, + 318 + ], + "blocks": [ + { + "bbox": [ + 495, + 259, + 553, + 318 + ], + "lines": [ + { + "bbox": [ + 495, + 259, + 553, + 318 + ], + "spans": [ + { + "bbox": [ + 495, + 259, + 553, + 318 + ], + "type": "image", + "image_path": "c4f8781f13f508d9d11ad39347cc9035561e4d98b07fe6491addac41b9b76c5d.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 518, + 318, + 528, + 327 + ], + "lines": [ + { + "bbox": [ + 518, + 318, + 528, + 327 + ], + "spans": [ + { + "bbox": [ + 518, + 318, + 528, + 327 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 317, + 326, + 376, + 384 + ], + "blocks": [ + { + "bbox": [ + 317, + 326, + 376, + 384 + ], + "lines": [ + { + "bbox": [ + 317, + 326, + 376, + 384 + ], + "spans": [ + { + "bbox": [ + 317, + 326, + 376, + 384 + ], + "type": "image", + "image_path": "5e9d4ab5123f1a5541ceae7175fe2b8e0c0012c6dcab1917b3e6a2cae5088baa.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 386, + 353, + 395 + ], + "lines": [ + { + "bbox": [ + 343, + 386, + 353, + 395 + ], + "spans": [ + { + "bbox": [ + 343, + 386, + 353, + 395 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 406, + 555, + 472 + ], + "lines": [ + { + "bbox": [ + 313, + 406, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 555, + 472 + ], + "type": "text", + "content": "Figure 1. Visualization of attention and feature maps. (a) Attention map of the selected target object token (marked as " + }, + { + "bbox": [ + 313, + 406, + 555, + 472 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 313, + 406, + 555, + 472 + ], + "type": "text", + "content": "). (b) Attention map of the selected background token (marked as " + }, + { + "bbox": [ + 313, + 406, + 555, + 472 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 313, + 406, + 555, + 472 + ], + "type": "text", + "content": "). (c) Attention map of the class token. (d) Diagonal weights of the attention map of each image token. (e) Detected outliers. (f) - (h) output feature map of CLIP [38], SCLIP [45], and our SFP." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 377, + 326, + 435, + 384 + ], + "blocks": [ + { + "bbox": [ + 377, + 326, + 435, + 384 + ], + "lines": [ + { + "bbox": [ + 377, + 326, + 435, + 384 + ], + "spans": [ + { + "bbox": [ + 377, + 326, + 435, + 384 + ], + "type": "image", + "image_path": "985463eea3e0187c90e3448f13faec3f03e9b8bd1c7111b9d6bd794cf9bf4d1a.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 386, + 411, + 395 + ], + "lines": [ + { + "bbox": [ + 402, + 386, + 411, + 395 + ], + "spans": [ + { + "bbox": [ + 402, + 386, + 411, + 395 + ], + "type": "text", + "content": "(f)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 435, + 326, + 494, + 384 + ], + "blocks": [ + { + "bbox": [ + 435, + 326, + 494, + 384 + ], + "lines": [ + { + "bbox": [ + 435, + 326, + 494, + 384 + ], + "spans": [ + { + "bbox": [ + 435, + 326, + 494, + 384 + ], + "type": "image", + "image_path": "772505dcee4a892fcc1253b7db85cab2a9660fcc8a2407c5ec9890c08344c1a1.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 459, + 386, + 470, + 395 + ], + "lines": [ + { + "bbox": [ + 459, + 386, + 470, + 395 + ], + "spans": [ + { + "bbox": [ + 459, + 386, + 470, + 395 + ], + "type": "text", + "content": "(g)" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 495, + 326, + 553, + 384 + ], + "blocks": [ + { + "bbox": [ + 495, + 326, + 553, + 384 + ], + "lines": [ + { + "bbox": [ + 495, + 326, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 495, + 326, + 553, + 384 + ], + "type": "image", + "image_path": "e5703a47e2d24303ae0dd1ccdc5b4fc7b0ab8e9f4fd2f06ce3c18696f858fd6d.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 518, + 386, + 528, + 395 + ], + "lines": [ + { + "bbox": [ + 518, + 386, + 528, + 395 + ], + "spans": [ + { + "bbox": [ + 518, + 386, + 528, + 395 + ], + "type": "text", + "content": "(h)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 496, + 555, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 555, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 555, + 639 + ], + "type": "text", + "content": "like traditional semantic segmentation, OVSS must handle an open set of classes, making it more challenging. Vision-language models (VLM), such as CLIP [38], demonstrate remarkable zero-shot performance by leveraging large-scale image-text pairs [39]. Thus, existing approaches can be categorized into two pipelines: fine-tuning CLIP or training-free CLIP. Although the fine-tuning pipeline yields superior performance, it risks overfitting and compromises generalization [30, 43]. In contrast, the training-free pipeline preserves CLIP's original generalization capability without additional training, which is more convenient and efficient." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": "Recent training-free methods [12, 40] have demonstrated that in the original CLIP's pair-wise attention maps, there exist the same highly correlated image tokens for all input tokens, such as those marked with the white box in Fig. 1 (a) - (b). In such a situation, regardless of whether it is the target image token (Fig. 1 (a)), the background image to-" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 227, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 227, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 227, + 713 + ], + "type": "text", + "content": "*Corresponding author: siyue.yu02@xjtlu.edu.cn" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20291" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "ken (Fig. 1 (b)), or the class token (Fig. 1 (c)), the pair-wise token relationships within the attention map are all over-activated for these identical tokens, i.e., outliers. Consequently, when aggregated with outliers, the derived semantic feature representation is easily polluted to overemphasize these outliers and the interpretation of the target region is impeded, e.g., the feature map of 'car' in Fig 1 (f) focuses on the outliers instead of the object itself." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 170, + 296, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 170, + 296, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 170, + 296, + 397 + ], + "type": "text", + "content": "To mitigate outliers and ensure more accurate semantic feature representation, many training-free approaches have presented diverse modifications to the self-attention mechanism in the last layer [14, 16, 24, 45]. While these strategies effectively restrict CLIP's attention to ignore outliers, some remain unfiltered, as shown in Fig. 1 (g). Moreover, some approaches [2, 12] rely on empirical parameter tuning to detect and remove a fixed number of outliers, which is less robust to different backbones and images. Beyond these optimization approaches, several studies [12, 40] further reveal that outliers predominantly appear in the intermediate layers. In this way, once outliers emerge, they will propagate through subsequent layers, continuously affecting semantic feature representations. Yet, previous approaches fail to analyze and address outliers in the intermediate layers so that outliers are easily propagated to pollute the final semantic feature. Thus, our objective is to adaptively eliminate outliers across multiple layers to achieve more robust semantic feature representations without training." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 400, + 296, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 400, + 296, + 639 + ], + "spans": [ + { + "bbox": [ + 56, + 400, + 296, + 639 + ], + "type": "text", + "content": "Building on these insights, we propose a Self-adaptive Feature Purifier framework (SFP) to resolve outliers adaptively and enhance semantic representations for CLIP-based training-free OVSS. Firstly, we design a Self-adaptive Outlier Mitigator (SOM) to identify and mitigate outliers across multiple layers. Tokens are expected to focus primarily on their relevant regions, especially themselves, rather than irrelevant regions in the attention map. However, as shown in Fig. 1 (d), we observe that the self-response of image tokens in the attention map is not always prominent, especially outliers. The comparison of Fig. 1 (c) & (d) reveals that outliers exhibit a stronger response to the class token than their self-response. Hence, our SOM can identify outliers by simply computing the difference between the self-response values of the image tokens in the attention map (i.e., the diagonal weights of the attention map) and their attention values with the class token. The detected outliers are shown in Fig. 1 (e). In this way, our parameter-free SOM enables adaptive detection across various layers to mitigate the propagating influence of outliers without manual tuning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 642, + 294, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 294, + 714 + ], + "type": "text", + "content": "Nevertheless, outliers not only cause tokens to overemphasize them but also impede tokens' responses to semantic regions in the attention map. Therefore, merely removing the outliers can't directly increase the attention weights of the related regions. To enhance the tokens' attention responses to semantic regions, we propose a Semantic-aware" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "content": "Attention Enhancer (SAE). SAE incorporates the self-self attention mechanism in CLIP's last layer and applies it to the purified feature derived from SOM at the penultimate layer, enabling a better focus on semantic regions. In addition, shallow attention maps typically contain structured object information and focus on relevant semantic regions [27, 30]. To further enhance the precise spatial perception of the final feature map, we propose a Hierarchical Attention Integrator (HAI) as an assistant to the SAE, where the attention maps in previous layers are firstly processed to mitigate the influence of the outliers and then integrated to refine the final semantic feature for prediction. The refined semantic feature for the final prediction is shown in Fig. 1 (h). Our SFP can generate a more purified and complete semantic feature for segmentation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 252, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 252, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 252, + 553, + 312 + ], + "type": "text", + "content": "We conduct comprehensive experiments to evaluate the proposed SFP. Our method outperforms existing methods and achieves state-of-the-art (SOTA) performance across eight evaluation datasets. In summary, our contributions are concluded as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 315, + 316, + 553, + 531 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 315, + 316, + 553, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 316, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 315, + 316, + 553, + 363 + ], + "type": "text", + "content": "- We present a Self-adaptive Feature Purifier framework (SFP) for training-free OVSS, which can derive a more accurate semantic feature to strengthen the final open-vocabulary semantic segmentation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 364, + 553, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 364, + 553, + 423 + ], + "spans": [ + { + "bbox": [ + 315, + 364, + 553, + 423 + ], + "type": "text", + "content": "- We propose a simple yet effective mechanism, Self-adaptive Outlier Mitigator (SOM), which adaptively detects and mitigates outliers across CLIP's diverse layers so that the influence of the propagated influence of outliers can be suppressed." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 423, + 553, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 423, + 553, + 482 + ], + "spans": [ + { + "bbox": [ + 315, + 423, + 553, + 482 + ], + "type": "text", + "content": "- We introduce a Semantic-aware Attention Enhancer (SAE) and a Hierarchical Attention Integrator (HAI) to further refine the semantic feature relationships on relevant regions, aggregating the more relevant semantic feature for final prediction." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 483, + 553, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 483, + 553, + 531 + ], + "spans": [ + { + "bbox": [ + 315, + 483, + 553, + 531 + ], + "type": "text", + "content": "- Comprehensive experiments demonstrate that SFP derives the best performance under a fair comparison e.g., SFP achieves a gain of " + }, + { + "bbox": [ + 315, + 483, + 553, + 531 + ], + "type": "inline_equation", + "content": "5.6\\%" + }, + { + "bbox": [ + 315, + 483, + 553, + 531 + ], + "type": "text", + "content": " mIoU on Cityscapes and an average gain of " + }, + { + "bbox": [ + 315, + 483, + 553, + 531 + ], + "type": "inline_equation", + "content": "4.6\\%" + }, + { + "bbox": [ + 315, + 483, + 553, + 531 + ], + "type": "text", + "content": " mIoU across eight datasets." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 544, + 400, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 544, + 400, + 555 + ], + "spans": [ + { + "bbox": [ + 314, + 544, + 400, + 555 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 563, + 509, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 509, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 509, + 577 + ], + "type": "text", + "content": "2.1. pre-trained Vision-Language Models" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "text", + "content": "Pre-trained vision-language models [17, 26, 38] have drawn significant attention for their ability to bridge visual and textual modalities. CLIP [38] has emerged as a highly successful model trained on a large-scale dataset of image-text pairs using contrastive learning. OpenCLIP [9] builds upon CLIP by leveraging public datasets such as LAION [39]. Despite their strong zero-shot capabilities, these models are pretrained at the image level for classification. This presents a notable limitation: the attention maps of VLMs tend to overemphasize outliers rather than semantic image tokens, leading to insufficient capture of spatial details. This draw" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20292" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "back hinders the performance in dense prediction tasks like semantic segmentation, which demands precise pixel-level understanding. Our goal is to adaptively detect and mitigate the impact of these outliers and enhance the semantic feature representation for dense prediction tasks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 143, + 226, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 143, + 226, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 143, + 226, + 156 + ], + "type": "text", + "content": "2.2. Open-vocabulary Segmentation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 161, + 296, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 161, + 296, + 256 + ], + "spans": [ + { + "bbox": [ + 55, + 161, + 296, + 256 + ], + "type": "text", + "content": "Semantic segmentation involves the pixel-wise classification of an image. Compared with the traditional segmentation methods [6, 8, 31, 49] that are trained and evaluated on the same fixed set of seen classes. OVSS methods [36, 59] aim to recognize and segment unseen objects during inference by harnessing the zero-shot ability of VLMs. Existing works can be broadly divided into two categories: training-based and training-free." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 258, + 296, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 296, + 461 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 296, + 461 + ], + "type": "text", + "content": "Training-based Methods Training-based methods [18, 37, 56] typically require fine-tuning on a fixed set of categories from a pixel-level annotated dataset. Some approaches follow a two-stage pipeline, where the first stage extracts the mask proposals, and the second stage assigns semantic labels to them. For instance, OVSeg [51] first trains class-agnostic mask proposals using the query-based framework Mask2Former [8] and then fine-tunes CLIP [38] to classify the cropped and masked images. Other approaches adopt a single-stage pipeline. SAN [52] introduces a side adapter to adapt CLIP [38] for both classification and segmentation. SED [48] presents a simple encoder-decoder architecture with category early rejection for fast inference. CATSeg [10] constructs a pixel-level cost map for segmentation. However, these training-based methods are prone to overfitting the training dataset, which limits the generalization ability due to fine-tuning CLIP." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 463, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 296, + 714 + ], + "type": "text", + "content": "Training-free Methods Unlike training-based methods, training-free methods [20, 41] aim to directly adapt VLMs for OVSS without extra training. Several approaches leverage the vision foundation models [5, 21, 42] to improve CLIP's spatial coherence. LaVG [19] uses DINO [5] for panoptic cuts, which iteratively discovers object masks until all image pixels are covered. ProxyCLIP [25] adjusts attention weights based on affinities learned from DINO's feature correspondence. However, these methods fail to exploit CLIP's inherent potential. Other efforts aim to modify CLIP's final layer to address its poor spatial consistency. For example, MaskCLIP [14] only utilizes the value feature of the final layer for prediction. CLIPtrase [40] integrates self-self attention with clustering for post-processing. SCLIP [45] and ClearCLIP [24] both modify the last layer self-attention mechanism to better represent the local information. Nevertheless, few training-free methods address propagated outliers, which act as noisy inputs across multiple layers and degrade spatial perception. In contrast, our approach enables adaptive detection and mitigation of propagated outliers among various layers, which improves" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 84 + ], + "type": "text", + "content": "CLIP's semantic awareness and boosts its internal potential." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 93, + 370, + 105 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 93, + 370, + 105 + ], + "spans": [ + { + "bbox": [ + 314, + 93, + 370, + 105 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 113, + 382, + 125 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 113, + 382, + 125 + ], + "spans": [ + { + "bbox": [ + 314, + 113, + 382, + 125 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 130, + 555, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 130, + 555, + 225 + ], + "spans": [ + { + "bbox": [ + 313, + 130, + 555, + 225 + ], + "type": "text", + "content": "Fig. 2 depicts the whole framework of SFP, including four main modules: a frozen CLIP backbone to encode the input image and text descriptions, a SOM to detect and eliminate outliers, an SAE to augment the attention intensity on related semantic regions, and an HAI that leverages intermediate-layer attention maps to assist SAE for enhanced semantic feature representation. The complete inference process is as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 226, + 553, + 417 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 315, + 226, + 553, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 226, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 315, + 226, + 553, + 285 + ], + "type": "text", + "content": "1) First, the image and text descriptions are input to the CLIP encoder for the image feature and the text feature. Among the CLIP image encoder, our SOM is added after each transformer layer to mitigate the outliers and generate the purified feature for propagation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 286, + 553, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 286, + 553, + 345 + ], + "spans": [ + { + "bbox": [ + 315, + 286, + 553, + 345 + ], + "type": "text", + "content": "2) Then, the derived final feature map from the last SOM, " + }, + { + "bbox": [ + 315, + 286, + 553, + 345 + ], + "type": "inline_equation", + "content": "F_{SOM}" + }, + { + "bbox": [ + 315, + 286, + 553, + 345 + ], + "type": "text", + "content": ", is input to our SAE, which replaces the final transformer layer. SAE emphasizes the attention values of relevant semantic regions to generate the strengthened semantic feature " + }, + { + "bbox": [ + 315, + 286, + 553, + 345 + ], + "type": "inline_equation", + "content": "F_{SAE}" + }, + { + "bbox": [ + 315, + 286, + 553, + 345 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 346, + 553, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 346, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 315, + 346, + 553, + 380 + ], + "type": "text", + "content": "3) After that, our HAI extracts the multi-layer attention maps from the image encoder to further refine " + }, + { + "bbox": [ + 315, + 346, + 553, + 380 + ], + "type": "inline_equation", + "content": "F_{SAE}" + }, + { + "bbox": [ + 315, + 346, + 553, + 380 + ], + "type": "text", + "content": " for the final semantic feature " + }, + { + "bbox": [ + 315, + 346, + 553, + 380 + ], + "type": "inline_equation", + "content": "F_{vis}" + }, + { + "bbox": [ + 315, + 346, + 553, + 380 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 381, + 553, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 381, + 553, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 553, + 417 + ], + "type": "text", + "content": "4) Finally, the segmentation result is obtained by calculating the cosine similarity of the final semantic feature " + }, + { + "bbox": [ + 313, + 381, + 553, + 417 + ], + "type": "inline_equation", + "content": "F_{vis}" + }, + { + "bbox": [ + 313, + 381, + 553, + 417 + ], + "type": "text", + "content": " and text feature " + }, + { + "bbox": [ + 313, + 381, + 553, + 417 + ], + "type": "inline_equation", + "content": "F_{txt}" + }, + { + "bbox": [ + 313, + 381, + 553, + 417 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 423, + 393, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 423, + 393, + 436 + ], + "spans": [ + { + "bbox": [ + 314, + 423, + 393, + 436 + ], + "type": "text", + "content": "3.2. Preliminary" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 440, + 555, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 440, + 555, + 571 + ], + "spans": [ + { + "bbox": [ + 313, + 440, + 555, + 571 + ], + "type": "text", + "content": "In general, the training-free OVSS approach requires CLIP [38], which encompasses an image encoder and a text encoder to acquire multi-model features. Subsequently, these features are aligned via linear projection layers. In detail, the input image is first partitioned into a token sequence " + }, + { + "bbox": [ + 313, + 440, + 555, + 571 + ], + "type": "inline_equation", + "content": "X_{in} = [x_{cls}, x_1, \\dots, x_N] \\in R^{(N + 1) \\times d}" + }, + { + "bbox": [ + 313, + 440, + 555, + 571 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 440, + 555, + 571 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 313, + 440, + 555, + 571 + ], + "type": "text", + "content": " denotes the image token embedding, " + }, + { + "bbox": [ + 313, + 440, + 555, + 571 + ], + "type": "inline_equation", + "content": "x_{cls}" + }, + { + "bbox": [ + 313, + 440, + 555, + 571 + ], + "type": "text", + "content": " denotes the class token embedding, " + }, + { + "bbox": [ + 313, + 440, + 555, + 571 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 440, + 555, + 571 + ], + "type": "text", + "content": " is the image token sequence length, and " + }, + { + "bbox": [ + 313, + 440, + 555, + 571 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 440, + 555, + 571 + ], + "type": "text", + "content": " is the dimension of token embeddings, respectively. Then, these tokens are input to the image encoder, which contains a stack of vision transformer layers." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 571, + 554, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 571, + 554, + 655 + ], + "spans": [ + { + "bbox": [ + 313, + 571, + 554, + 655 + ], + "type": "text", + "content": "Each transformer layer consists of a multi-head self-attention (MHSA) mechanism and a feed-forward network (FFN) layer to extract the image feature. Specifically, in each MHSA, the input sequence " + }, + { + "bbox": [ + 313, + 571, + 554, + 655 + ], + "type": "inline_equation", + "content": "X_{in}" + }, + { + "bbox": [ + 313, + 571, + 554, + 655 + ], + "type": "text", + "content": " is mapped to query, key, and value embeddings, " + }, + { + "bbox": [ + 313, + 571, + 554, + 655 + ], + "type": "inline_equation", + "content": "Q, K, V \\in R^{(N + 1)\\times d}" + }, + { + "bbox": [ + 313, + 571, + 554, + 655 + ], + "type": "text", + "content": ". Afterwards, " + }, + { + "bbox": [ + 313, + 571, + 554, + 655 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 571, + 554, + 655 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 571, + 554, + 655 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 571, + 554, + 655 + ], + "type": "text", + "content": " are calculated to obtain the attention map " + }, + { + "bbox": [ + 313, + 571, + 554, + 655 + ], + "type": "inline_equation", + "content": "Attn \\in R^{(N + 1)\\times (N + 1)}" + }, + { + "bbox": [ + 313, + 571, + 554, + 655 + ], + "type": "text", + "content": ", which can be formulated as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 379, + 659, + 553, + 686 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 379, + 659, + 553, + 686 + ], + "spans": [ + { + "bbox": [ + 379, + 659, + 553, + 686 + ], + "type": "interline_equation", + "content": "A t t n = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}}\\right), \\tag {1}", + "image_path": "5896f3135a4ac7d9c3658d17939371ffd29fd72940e188f978122e0ac28bd34c.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": " means matrix transposition. After that, the attention map is multiplied " + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": " to aggregate the global infor" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20293" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 70, + 539, + 319 + ], + "blocks": [ + { + "bbox": [ + 72, + 70, + 539, + 319 + ], + "lines": [ + { + "bbox": [ + 72, + 70, + 539, + 319 + ], + "spans": [ + { + "bbox": [ + 72, + 70, + 539, + 319 + ], + "type": "image", + "image_path": "df20bce62542ccaf74ec7943fc6e61eeec8bbe555389af3e25a8aaa3c3419022.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 326, + 555, + 382 + ], + "lines": [ + { + "bbox": [ + 54, + 326, + 555, + 382 + ], + "spans": [ + { + "bbox": [ + 54, + 326, + 555, + 382 + ], + "type": "text", + "content": "Figure 2. Framework of our SFP. It contains four main modules: a frozen CLIP backbone for feature extraction, a Self-adaptive Outlier Mitigator (SOM) at each image encoder layer for outlier detection and elimination, a Semantic-aware Attention Enhancer (SAE) at the last layer to emphasise attention values of relevant semantic regions and a Hierarchical Attention Integrator (HAI) for object-centric feature representation. The segmentation result is generated by calculating the cosine similarity between the optimized semantic feature " + }, + { + "bbox": [ + 54, + 326, + 555, + 382 + ], + "type": "inline_equation", + "content": "F_{vis}" + }, + { + "bbox": [ + 54, + 326, + 555, + 382 + ], + "type": "text", + "content": " and the text feature " + }, + { + "bbox": [ + 54, + 326, + 555, + 382 + ], + "type": "inline_equation", + "content": "F_{txt}" + }, + { + "bbox": [ + 54, + 326, + 555, + 382 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 403, + 295, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 295, + 428 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 295, + 428 + ], + "type": "text", + "content": "mation, and an FFN is used to obtain the image feature " + }, + { + "bbox": [ + 55, + 403, + 295, + 428 + ], + "type": "inline_equation", + "content": "F_{img} \\in R^{(N + 1) \\times d}" + }, + { + "bbox": [ + 55, + 403, + 295, + 428 + ], + "type": "text", + "content": ", formulated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 127, + 434, + 295, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 434, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 127, + 434, + 295, + 449 + ], + "type": "interline_equation", + "content": "F _ {i m g} = \\operatorname {F F N} (A t t n V). \\tag {2}", + "image_path": "fd788d192544138c592f229137012c689947cdd39801d5fd99a99930f25b3150.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 453, + 295, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 295, + 502 + ], + "type": "text", + "content": "Since the class embedding " + }, + { + "bbox": [ + 55, + 453, + 295, + 502 + ], + "type": "inline_equation", + "content": "x_{cls}" + }, + { + "bbox": [ + 55, + 453, + 295, + 502 + ], + "type": "text", + "content": " is not used to predict the segmentation result, we omit it and reshape " + }, + { + "bbox": [ + 55, + 453, + 295, + 502 + ], + "type": "inline_equation", + "content": "F_{img}" + }, + { + "bbox": [ + 55, + 453, + 295, + 502 + ], + "type": "text", + "content": " to obtain the visual feature " + }, + { + "bbox": [ + 55, + 453, + 295, + 502 + ], + "type": "inline_equation", + "content": "F_{vis} \\in R^{h \\times w \\times d}" + }, + { + "bbox": [ + 55, + 453, + 295, + 502 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 453, + 295, + 502 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 453, + 295, + 502 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 453, + 295, + 502 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 453, + 295, + 502 + ], + "type": "text", + "content": " denote the height and width of the feature map." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 502, + 295, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 295, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 295, + 574 + ], + "type": "text", + "content": "Meanwhile, the textual description is derived from the standard ImageNet [13] prompts and transformed to the text feature " + }, + { + "bbox": [ + 55, + 502, + 295, + 574 + ], + "type": "inline_equation", + "content": "F_{txt} \\in R^{c \\times d}" + }, + { + "bbox": [ + 55, + 502, + 295, + 574 + ], + "type": "text", + "content": " with CLIP's text encoder, where " + }, + { + "bbox": [ + 55, + 502, + 295, + 574 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 502, + 295, + 574 + ], + "type": "text", + "content": " represents the number of classes. By calculating the cosine similarity between " + }, + { + "bbox": [ + 55, + 502, + 295, + 574 + ], + "type": "inline_equation", + "content": "F_{vis}" + }, + { + "bbox": [ + 55, + 502, + 295, + 574 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 502, + 295, + 574 + ], + "type": "inline_equation", + "content": "F_{txt}" + }, + { + "bbox": [ + 55, + 502, + 295, + 574 + ], + "type": "text", + "content": ", the segmentation map " + }, + { + "bbox": [ + 55, + 502, + 295, + 574 + ], + "type": "inline_equation", + "content": "Seg \\in R^{H \\times W}" + }, + { + "bbox": [ + 55, + 502, + 295, + 574 + ], + "type": "text", + "content": " can be generated like:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 101, + 580, + 295, + 600 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 580, + 295, + 600 + ], + "spans": [ + { + "bbox": [ + 101, + 580, + 295, + 600 + ], + "type": "interline_equation", + "content": "S e g = U (\\underset {c} {\\arg \\max } \\cos (F _ {v i s}, F _ {t x t})), \\tag {3}", + "image_path": "7fa15e73491ca0cbd883668a0ce9f98bc66a3d559c80c9cffde17706109917f5.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 296, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 654 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 605, + 296, + 654 + ], + "type": "inline_equation", + "content": "\\cos (\\cdot)" + }, + { + "bbox": [ + 55, + 605, + 296, + 654 + ], + "type": "text", + "content": " means cosine similarity calculation and " + }, + { + "bbox": [ + 55, + 605, + 296, + 654 + ], + "type": "inline_equation", + "content": "U(\\cdot)" + }, + { + "bbox": [ + 55, + 605, + 296, + 654 + ], + "type": "text", + "content": " denotes the upsampling operation to match the original input image size. " + }, + { + "bbox": [ + 55, + 605, + 296, + 654 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 605, + 296, + 654 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 605, + 296, + 654 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 605, + 296, + 654 + ], + "type": "text", + "content": " denote the height and width of the input image, respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 660, + 224, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 224, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 224, + 673 + ], + "type": "text", + "content": "3.3. Self-adaptive Outlier Mitigator" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "CLIP [38] is pre-trained at the image level for classification, which tends to concentrate on the most discriminative feature that is important for classification. In this way, all input" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 403, + 555, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 555, + 521 + ], + "type": "text", + "content": "tokens, including both class and image tokens, will emphasize these identical tokens (outliers) in attention maps [12]. In other words, the corresponding outlier column in the attention map will be highlighted. Further, given that the classification process relies solely on the class token, the class token may exhibit a stronger response to the outlier in attention maps compared to image tokens. This prompts us to investigate whether the self-relevance of image tokens and the class token's similarity across all image tokens could serve as potential indicators to identify outliers." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "type": "text", + "content": "Based on the above assumption, to robustly mitigate outliers, we introduce a Self-adaptive Outlier Mitigator (SOM) to detect and eliminate the outliers automatically. In particular, we use the attention map " + }, + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "type": "inline_equation", + "content": "Attn" + }, + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "type": "text", + "content": " to obtain the self-attention weights of image tokens, denoted as " + }, + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "type": "inline_equation", + "content": "Attn_{i,i}" + }, + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "type": "text", + "content": ", and their attention weights to the class tokens, denoted as " + }, + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "type": "inline_equation", + "content": "Attn_{cls,i}" + }, + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "type": "text", + "content": "-th image token and " + }, + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "type": "inline_equation", + "content": "cls" + }, + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "type": "text", + "content": " denotes the class token. By comparing the difference between " + }, + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "type": "inline_equation", + "content": "Attn_{i,i}" + }, + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "type": "inline_equation", + "content": "Attn_{cls,i}" + }, + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "type": "text", + "content": ", we obtain the set of outliers " + }, + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 522, + 556, + 654 + ], + "type": "text", + "content": ", since the self-attention weights should be higher than others. This detection process of SOM is formulated as," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 341, + 659, + 555, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 659, + 555, + 673 + ], + "spans": [ + { + "bbox": [ + 341, + 659, + 555, + 673 + ], + "type": "interline_equation", + "content": "\\mathcal {S} = \\left\\{i \\mid A t t n _ {i, i} < A t t n _ {c l s, i} \\right\\}, \\quad i \\in [ 1, N ], \\tag {4}", + "image_path": "d0954acf7e5235e21b0fcf6218113eb8331d04b2a791c998b6f059ef3ba53bbc.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 459, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 459, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 459, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 677, + 459, + 689 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 677, + 459, + 689 + ], + "type": "text", + "content": " denotes the set of outliers." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "After recognizing outliers, we propose eliminating them at the feature level rather than modifying the attention map" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20294" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": "to prevent potential model collapse. Inspired by previous work [2], it's crucial to ensure the semantic consistency of the feature map, and directly masking outliers at the feature level is not an appropriate approach. Thus, SOM can eliminate outliers through in-place merging, where the detected outliers are replaced with the average value of its 8-neighbours. That is to say, the derived " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "F_{img}" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": " in the original attention mechanism will be updated through our SOM as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 80, + 176, + 296, + 222 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 176, + 296, + 222 + ], + "spans": [ + { + "bbox": [ + 80, + 176, + 296, + 222 + ], + "type": "interline_equation", + "content": "\\hat {F} _ {i m g} [ u ] = \\left\\{ \\begin{array}{l l} \\frac {1}{8} \\sum_ {v \\in \\mathcal {N} _ {8} (u)} F _ {i m g} [ v ], & u \\in \\mathcal {S} \\\\ F _ {i m g} [ u ], & \\text {o t h e r w s i e} \\end{array} , \\right. \\tag {5}", + "image_path": "86e2478ff00a778ff654ef646b65c24ee83bda68e55c63aa56a27df29cc719c0.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 229, + 296, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 229, + 296, + 325 + ], + "spans": [ + { + "bbox": [ + 55, + 229, + 296, + 325 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 229, + 296, + 325 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_8(u)" + }, + { + "bbox": [ + 55, + 229, + 296, + 325 + ], + "type": "text", + "content": " means the 8-neighbour non-outlier tokens of the " + }, + { + "bbox": [ + 55, + 229, + 296, + 325 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 55, + 229, + 296, + 325 + ], + "type": "text", + "content": "-th point and " + }, + { + "bbox": [ + 55, + 229, + 296, + 325 + ], + "type": "inline_equation", + "content": "\\hat{F}_{img}" + }, + { + "bbox": [ + 55, + 229, + 296, + 325 + ], + "type": "text", + "content": " denotes the updated feature. In this way, our SOM initially identifies the outlier and then leverages surrounding token features to replace the detected one. This process can mitigate the adverse influence exerted by the outliers and maintain the feature distribution that is crucial for CLIP's image encoder. By doing so, the purified and superior semantic feature representation is generated." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 325, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 325, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 325, + 296, + 456 + ], + "type": "text", + "content": "Furthermore, our SOM exhibits remarkable versatility in that it can be applied to each individual layer of the image encoder. This allows for the detection of outliers and the generation of purified features at multiple levels. In this manner, our parameter-free SOM establishes an efficient and precise mechanism for detecting and eliminating outliers that propagate across various layers of the image encoder. This mechanism is instrumental in preserving accurate semantic information and feature representations throughout the entire image encoder, thereby enhancing its overall performance and reliability." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 58, + 467, + 132, + 552 + ], + "blocks": [ + { + "bbox": [ + 58, + 467, + 132, + 552 + ], + "lines": [ + { + "bbox": [ + 58, + 467, + 132, + 552 + ], + "spans": [ + { + "bbox": [ + 58, + 467, + 132, + 552 + ], + "type": "image", + "image_path": "b89618275e0154c0991ecd39b58a4ca196878a512fcf7a2ac266bba76f11ac4a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 90, + 552, + 101, + 561 + ], + "lines": [ + { + "bbox": [ + 90, + 552, + 101, + 561 + ], + "spans": [ + { + "bbox": [ + 90, + 552, + 101, + 561 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 571, + 296, + 616 + ], + "lines": [ + { + "bbox": [ + 55, + 571, + 296, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 296, + 616 + ], + "type": "text", + "content": "Figure 3. Comparison of different feature visualization. (a) Noisy feature with outliers. (b) Purified feature with SOM. (c) Semantic feature with SAE. It's seen that the proposed SAE strengthens the semantic regions with higher object activation." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 132, + 466, + 207, + 552 + ], + "blocks": [ + { + "bbox": [ + 132, + 466, + 207, + 552 + ], + "lines": [ + { + "bbox": [ + 132, + 466, + 207, + 552 + ], + "spans": [ + { + "bbox": [ + 132, + 466, + 207, + 552 + ], + "type": "image", + "image_path": "a0e4641041202a6a68fa1d214a2ba2478afbe719137ad550d2194c14bf9c9d49.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 165, + 552, + 176, + 561 + ], + "lines": [ + { + "bbox": [ + 165, + 552, + 176, + 561 + ], + "spans": [ + { + "bbox": [ + 165, + 552, + 176, + 561 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 207, + 466, + 296, + 552 + ], + "blocks": [ + { + "bbox": [ + 207, + 466, + 296, + 552 + ], + "lines": [ + { + "bbox": [ + 207, + 466, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 207, + 466, + 296, + 552 + ], + "type": "image", + "image_path": "afa714bc38ba978a4f3c0f86f18f12e359170e7e2c0e24253b209ff14e8dc4da.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 242, + 553, + 253, + 562 + ], + "lines": [ + { + "bbox": [ + 242, + 553, + 253, + 562 + ], + "spans": [ + { + "bbox": [ + 242, + 553, + 253, + 562 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 636, + 249, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 636, + 249, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 636, + 249, + 647 + ], + "type": "text", + "content": "3.4. Semantic-aware Attention Enhancer" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "The comparison of Fig. 3 (a) & (b) indicates that the purified feature " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "F_{SOM}" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " generated by the last SOM effectively ignores the irrelevant regions of outliers, marked as white boxes. However, it's also observed from Fig. 3 (b) & (c) that the object activation of " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "F_{SOM}" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " is not salient enough," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "which indicates that eliminating outliers can't contribute to the attention response intensity." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 96, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 553, + 144 + ], + "type": "text", + "content": "To address this issue, we propose a Semantic-aware Attention Enhancer (SAE) to strengthen the object activation in CLIP's last layer, where the self-self attention mechanism is used to derive better feature relationships," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 344, + 149, + 553, + 178 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 149, + 553, + 178 + ], + "spans": [ + { + "bbox": [ + 344, + 149, + 553, + 178 + ], + "type": "interline_equation", + "content": "A t t n _ {S A E} = \\operatorname {S o f t m a x} \\left(\\lambda \\left(\\frac {Q Q ^ {T}}{\\sqrt {d}} + \\frac {K K ^ {T}}{\\sqrt {d}}\\right)\\right), \\tag {6}", + "image_path": "95a787c2c423d1e6a130cae4a8b545f76a392d2debad223102cd97541ed05ae2.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 183, + 554, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 183, + 554, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 183, + 554, + 243 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 183, + 554, + 243 + ], + "type": "inline_equation", + "content": "Attn_{SAE}" + }, + { + "bbox": [ + 313, + 183, + 554, + 243 + ], + "type": "text", + "content": " represents the semantic-aware attention map and " + }, + { + "bbox": [ + 313, + 183, + 554, + 243 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 183, + 554, + 243 + ], + "type": "text", + "content": " denotes the logit scale used to sharpen attention scores, set to 0.5. Then, the semantic feature " + }, + { + "bbox": [ + 313, + 183, + 554, + 243 + ], + "type": "inline_equation", + "content": "F_{SAE}" + }, + { + "bbox": [ + 313, + 183, + 554, + 243 + ], + "type": "text", + "content": " is obtained through the residual connection with the purified feature " + }, + { + "bbox": [ + 313, + 183, + 554, + 243 + ], + "type": "inline_equation", + "content": "F_{SOM}" + }, + { + "bbox": [ + 313, + 183, + 554, + 243 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 345, + 251, + 553, + 264 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 251, + 553, + 264 + ], + "spans": [ + { + "bbox": [ + 345, + 251, + 553, + 264 + ], + "type": "interline_equation", + "content": "F _ {S A E} = F _ {S O M} + \\operatorname {F F N} \\left(\\operatorname {A t t n} _ {S A E} F _ {S O M}\\right). \\tag {7}", + "image_path": "0cf44cec21e6ad1957f420f9b2ebfead6d0eb12b9144f437acac122493d36966.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 270, + 554, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 554, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 554, + 378 + ], + "type": "text", + "content": "The self-self attention mechanism has shown a strong ability to construct feature relationships [16, 24, 45]. Therefore, applying self-self attention to our purified feature enables the establishment of more precise spatial correlations. These accurate spatial correlations play a crucial role in enabling each point to aggregate the relevant semantic feature effectively. As illustrated in Fig. 3 (b) & (c), SAE significantly augments attention intensity in semantic regions compared to the feature solely processed by SOM." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 384, + 495, + 397 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 495, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 495, + 397 + ], + "type": "text", + "content": "3.5. Hierarchical Attention Integrator" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 402, + 554, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 554, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 554, + 474 + ], + "type": "text", + "content": "To further refine the generated semantic feature " + }, + { + "bbox": [ + 313, + 402, + 554, + 474 + ], + "type": "inline_equation", + "content": "F_{SAE}" + }, + { + "bbox": [ + 313, + 402, + 554, + 474 + ], + "type": "text", + "content": ", we leverage attention maps from shallow layers to capture discriminative pair-wise feature relationships since shallow layers typically focus on object structures [27]. To this end, we propose a Hierarchical Attention Integrator (HAI), which integrates multi-layer attention maps to refine " + }, + { + "bbox": [ + 313, + 402, + 554, + 474 + ], + "type": "inline_equation", + "content": "F_{SAE}" + }, + { + "bbox": [ + 313, + 402, + 554, + 474 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 474, + 554, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 554, + 544 + ], + "type": "text", + "content": "Instead of using the original attention maps with outliers, HAI integrates attention maps from each transformer layer while excluding the detected outlier weights to avoid information pollution. Firstly, HAI employs the outlier set " + }, + { + "bbox": [ + 313, + 474, + 554, + 544 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 474, + 554, + 544 + ], + "type": "text", + "content": " to select the non-outlier weights of each attention map, which can be formulated as," + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 361, + 551, + 553, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 551, + 553, + 583 + ], + "spans": [ + { + "bbox": [ + 361, + 551, + 553, + 583 + ], + "type": "interline_equation", + "content": "\\overline {{A t t n}} _ {j, i} ^ {l} = \\left\\{ \\begin{array}{l l} A t t n _ {j, i}, & \\text {i f} i \\notin \\mathcal {S} ^ {l} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {8}", + "image_path": "c3310528f270a136585c881d8a8915a738f8ccf0fe3cbb2618cd07d7a5f94580.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 590, + 554, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 554, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 554, + 643 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 590, + 554, + 643 + ], + "type": "inline_equation", + "content": "\\overline{Attn}^l" + }, + { + "bbox": [ + 313, + 590, + 554, + 643 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 590, + 554, + 643 + ], + "type": "inline_equation", + "content": "S^l" + }, + { + "bbox": [ + 313, + 590, + 554, + 643 + ], + "type": "text", + "content": " denote the filtered attention map and the outlier set at the " + }, + { + "bbox": [ + 313, + 590, + 554, + 643 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 590, + 554, + 643 + ], + "type": "text", + "content": "-th layer. " + }, + { + "bbox": [ + 313, + 590, + 554, + 643 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 590, + 554, + 643 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 590, + 554, + 643 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 590, + 554, + 643 + ], + "type": "text", + "content": " correspond to the row and column indices of the attention map. Next, HAI integrates " + }, + { + "bbox": [ + 313, + 590, + 554, + 643 + ], + "type": "inline_equation", + "content": "\\overline{Attn}^l" + }, + { + "bbox": [ + 313, + 590, + 554, + 643 + ], + "type": "text", + "content": " from each layer, formulated as," + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 400, + 651, + 553, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 651, + 553, + 683 + ], + "spans": [ + { + "bbox": [ + 400, + 651, + 553, + 683 + ], + "type": "interline_equation", + "content": "A = \\sum_ {l = 1} ^ {L - 1} \\overline {{A t t n}} ^ {l}, \\tag {9}", + "image_path": "3ea80aa9f797533c057b4f6d2663b5eec576726afb7b02dd1ed32b0014fdf865.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": " denotes the integrated attention map, and " + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": " denotes the total number of transformer layers." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20295" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 70, + 542, + 290 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 542, + 290 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 542, + 290 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 542, + 290 + ], + "type": "table", + "html": "
MethodsPub. & YearExtra backboneTraining freewith backgroundV20without backgroundAvg.
V21PC60ObjectPC59StuffADECity
GroupVit [50]CVPR'22XX50.418.727.579.723.415.39.211.129.4
SegCLIP [32]ICML'23XX52.624.726.5------
TCL [7]CVPR'23XX55.030.431.683.233.922.417.124.037.2
DINOiser† [47]ECCV'24DINOX62.132.434.880.935.924.620.031.740.3
SAM-CLIP† [46]ECCV'24SAMX60.629.2----17.1--
PnP-OVSS† [33]CVPR'24BLIP✓--36.251.328.017.914.2--
LaVG† [19]ECCV'24DINO✓62.131.634.282.534.723.215.826.238.8
ProxyCLIP† [25]ECCV'24DINO✓61.335.337.580.339.126.520.238.142.3
CLIP [38]ICML'21X✓18.69.98.149.411.15.73.16.514.1
MaskCLIP [14]ECCV'22X✓38.823.620.674.926.416.49.812.627.9
CLIPSurgery [28]PR'25X✓-29.3---21.9-31.4-
CaR [44]CVPR'24X✓48.630.536.673.739.5-17.7--
GEM [3]CVPR'24X✓46.2---32.615.7---
CLIPtrase [40]ECCV'24X✓50.929.943.681.033.822.816.421.337.5
ClearCLIP* [24]ECCV'24X✓57.032.633.080.935.923.916.730.038.8
SCLIP* [45]ECCV'24X✓59.731.733.581.534.522.716.532.339.1
NACLIP* [16]WACV'25X✓58.932.233.279.735.223.317.435.539.4
SFP (Ours)X✓63.937.237.984.539.926.420.841.144.0
", + "image_path": "ff0a01007c44fdaab18e654939f724cdca857c6b08fd22d99f240e12011e076e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 300, + 555, + 335 + ], + "lines": [ + { + "bbox": [ + 55, + 300, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 555, + 335 + ], + "type": "text", + "content": "Table 1. Quantitative comparison of our SFP against other approaches across eight segmentation benchmark datasets. The best and second-best results are marked with **bold** and **underline**, respectively. † indicates that the additional backbone settings follow the original paper. * indicates the results are re-evaluated using the official implementation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 355, + 296, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 355, + 296, + 439 + ], + "spans": [ + { + "bbox": [ + 55, + 355, + 296, + 439 + ], + "type": "text", + "content": "After integrating multi-layer attention maps, we deploy a normalization strategy to optimize " + }, + { + "bbox": [ + 55, + 355, + 296, + 439 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 355, + 296, + 439 + ], + "type": "text", + "content": " instead of simply averaging it. Inspired by previous works [22, 35], we iteratively apply row-wise and column-wise normalization to " + }, + { + "bbox": [ + 55, + 355, + 296, + 439 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 355, + 296, + 439 + ], + "type": "text", + "content": " for a more object-centric attention distribution, leading to a stronger focus on relevant regions. This process can be formulated as," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 447, + 294, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 447, + 294, + 460 + ], + "spans": [ + { + "bbox": [ + 113, + 447, + 294, + 460 + ], + "type": "interline_equation", + "content": "A _ {H A I} = \\operatorname {n o r m} _ {c} \\left(\\operatorname {n o r m} _ {r} (A)\\right), \\tag {10}", + "image_path": "e0ff1c5ddc9478c6ccb65adcbef0abe83e70b6008c6f754ecc66704e18e1f5af.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 468, + 295, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 295, + 505 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 295, + 505 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 468, + 295, + 505 + ], + "type": "inline_equation", + "content": "A_{HAI}" + }, + { + "bbox": [ + 55, + 468, + 295, + 505 + ], + "type": "text", + "content": " denotes the normalized attention map, normc and normr denote the column-wise and row-wise normalization respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 505, + 296, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 505, + 296, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 505, + 296, + 540 + ], + "type": "text", + "content": "Then, " + }, + { + "bbox": [ + 55, + 505, + 296, + 540 + ], + "type": "inline_equation", + "content": "A_{HAI}" + }, + { + "bbox": [ + 55, + 505, + 296, + 540 + ], + "type": "text", + "content": " is multiplied with the semantic feature " + }, + { + "bbox": [ + 55, + 505, + 296, + 540 + ], + "type": "inline_equation", + "content": "F_{SAE}" + }, + { + "bbox": [ + 55, + 505, + 296, + 540 + ], + "type": "text", + "content": " to obtain the final semantic feature " + }, + { + "bbox": [ + 55, + 505, + 296, + 540 + ], + "type": "inline_equation", + "content": "F_{vis}" + }, + { + "bbox": [ + 55, + 505, + 296, + 540 + ], + "type": "text", + "content": " for segmentation, which is formulated as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 549, + 295, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 549, + 295, + 562 + ], + "spans": [ + { + "bbox": [ + 132, + 549, + 295, + 562 + ], + "type": "interline_equation", + "content": "F _ {v i s} = A _ {H A I} F _ {S A E}. \\tag {11}", + "image_path": "49679706e5d2640f76e5481bc6c5fcdde65d2bd37cf2854c26423dc87c53fa3e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 570, + 296, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 607 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 607 + ], + "type": "text", + "content": "Finally, we compute the cosine similarity between " + }, + { + "bbox": [ + 55, + 570, + 296, + 607 + ], + "type": "inline_equation", + "content": "F_{vis}" + }, + { + "bbox": [ + 55, + 570, + 296, + 607 + ], + "type": "text", + "content": " and the text feature " + }, + { + "bbox": [ + 55, + 570, + 296, + 607 + ], + "type": "inline_equation", + "content": "F_{txt}" + }, + { + "bbox": [ + 55, + 570, + 296, + 607 + ], + "type": "text", + "content": " extracted from the CLIP text encoder to predict the segmentation map, as formulated in Eq. 3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 616, + 132, + 630 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 132, + 630 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 132, + 630 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 636, + 182, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 636, + 182, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 636, + 182, + 649 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": "Datasets We conduct comprehensive experiments to verify the effectiveness of our method on diverse benchmark datasets: PASCAL VOC 2012 [15], PASCAL Context [34], and COCO [29]. These datasets are categorized into two settings: 1) with a background class, namely V20, PC59," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 355, + 555, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 355, + 555, + 391 + ], + "spans": [ + { + "bbox": [ + 313, + 355, + 555, + 391 + ], + "type": "text", + "content": "and Stuff; 2) without the background class, namely V21, PC60, and Object. We also report results on the ADE20K (ADE) [57] and Cityscapes (City) [11] datasets." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 391, + 556, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 391, + 556, + 511 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 556, + 511 + ], + "type": "text", + "content": "Implementation Details Similar to existing training-free methods, we adopt CLIP ViT-B/16 with the original pretrained weight [38]. Following previous works [2, 45], we combine standard ImageNet prompts [13] with the class names to construct text descriptions. For a fair comparison with previous works [25, 45], we resize images with a short side of 336 pixels (or 560 pixels for the high-resolution Cityscapes dataset) and apply sliding window inference using a " + }, + { + "bbox": [ + 313, + 391, + 556, + 511 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 313, + 391, + 556, + 511 + ], + "type": "text", + "content": " window with a " + }, + { + "bbox": [ + 313, + 391, + 556, + 511 + ], + "type": "inline_equation", + "content": "112 \\times 112" + }, + { + "bbox": [ + 313, + 391, + 556, + 511 + ], + "type": "text", + "content": " stride. Our experiments are conducted using one NVIDIA 3090 GPU." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 517, + 537, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 537, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 537, + 529 + ], + "type": "text", + "content": "4.2. Comparison with State-of-the-art Methods" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 534, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 605 + ], + "type": "text", + "content": "Our SFP is evaluated against existing SOTA methods, including both training-based and training-free approaches. The post-processing mask refinement techniques, such as PAMR [1] or DenseCRF [23], are not included for a fair comparison. The standard mean intersection over union (mIoU) is used as the metric for evaluation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "type": "text", + "content": "Quantitative Evaluation The main results are summarized in Tab. 1, where all evaluations are conducted using CLIP ViT-B/16 as the backbone to ensure a fair comparison. We compare SFP against multiple methods, including training-based methods, training-free methods with extra backbones, and training-free methods without extra backbones." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": "As shown in Tab. 1, our SFP outperforms existing SOTA methods, regardless of whether an additional backbone is introduced or not. In particular, compared with existing" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "20296" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 70, + 529, + 305 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 529, + 305 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 529, + 305 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 529, + 305 + ], + "type": "image", + "image_path": "456adbda35fa7d9096eb56eb60121eaea6067ac9850e0c307f84672bba38fb36.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 314, + 555, + 336 + ], + "lines": [ + { + "bbox": [ + 55, + 314, + 555, + 336 + ], + "spans": [ + { + "bbox": [ + 55, + 314, + 555, + 336 + ], + "type": "text", + "content": "Figure 4. Qualitative Results Comparison. We compare our SFP with CLIP [38], ClearCLIP [24], SCLIP [45] and NACLIP [16] without post-processing. It's observed that SFP generates more precise segmentation results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 357, + 297, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 357, + 297, + 561 + ], + "spans": [ + { + "bbox": [ + 54, + 357, + 297, + 561 + ], + "type": "text", + "content": "training-free methods without extra backbones [16, 24, 45] or cluster methods [53-55], our SFP achieves an average gain of " + }, + { + "bbox": [ + 54, + 357, + 297, + 561 + ], + "type": "inline_equation", + "content": "4.6\\%" + }, + { + "bbox": [ + 54, + 357, + 297, + 561 + ], + "type": "text", + "content": " mIoU over previous SOTA NACLIP [16]. Specifically, On the V20 and V21 dataset [15], SFP surpasses NACLIP [16] " + }, + { + "bbox": [ + 54, + 357, + 297, + 561 + ], + "type": "inline_equation", + "content": "4.8\\%" + }, + { + "bbox": [ + 54, + 357, + 297, + 561 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 357, + 297, + 561 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 54, + 357, + 297, + 561 + ], + "type": "text", + "content": " mIoU, respectively. These improvements indicate that our SFP can generate more purified feature maps for more precise segmentation. Beyond that, SFP outperforms SOTA training-free methods with extra backbones, e.g., ProxyCLIP [25], with a gain of " + }, + { + "bbox": [ + 54, + 357, + 297, + 561 + ], + "type": "inline_equation", + "content": "3.0\\%" + }, + { + "bbox": [ + 54, + 357, + 297, + 561 + ], + "type": "text", + "content": " mIoU on the Cityscape [11] dataset and an average gain of " + }, + { + "bbox": [ + 54, + 357, + 297, + 561 + ], + "type": "inline_equation", + "content": "1.7\\%" + }, + { + "bbox": [ + 54, + 357, + 297, + 561 + ], + "type": "text", + "content": " mIoU on eight datasets. Our SFP can even beat DINOiser [47], which requires both training and an extra backbone, with an average improvement of " + }, + { + "bbox": [ + 54, + 357, + 297, + 561 + ], + "type": "inline_equation", + "content": "3.7\\%" + }, + { + "bbox": [ + 54, + 357, + 297, + 561 + ], + "type": "text", + "content": " mIoU. These results highlight our SFP boosts CLIP's internal potential and reveal that feature purification does matter in training-free OVSS, where more purified features can predict better segmentation masks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "Qualitative Evaluation In Fig. 4, we compare the visualization results of various training-free methods without extra backbones, including CLIP [38], ClearCLIP [24], SCLIP [45], NACLIP [16] and our SFP. In particular, the segmentation maps obtained by vanilla CLIP [38] exhibit significant noises, unveiling its limitations in spatial localization. Besides, in complex scenarios, like 'sofa' in the second row, SCLIP [45] and NACLIP [16] fail to capture the complete segmentation mask. In contrast, our SFP can predict more accurate and complete segmentation maps. These results indicate that our SFP has a good ability to predict comprehensive segmentation masks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 357, + 408, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 408, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 408, + 370 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 375, + 555, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 375, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 555, + 434 + ], + "type": "text", + "content": "Comprehensive ablation studies are performed to analyze the effectiveness of our proposed mechanisms. Unless otherwise specified, we utilize CLIP with ViT-B/16 backbone and evaluate on V21 [15] and ADE [57] datasets for background and non-background settings, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 434, + 556, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 556, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 556, + 579 + ], + "type": "text", + "content": "Effect on each proposed module We first conduct several experiments to validate the effectiveness of our proposed modules of our SFP: SOM, SAE, and HAI by adding each module step by step, where ClearCLIP [24] is used as the baseline. As shown in Tab. 2, our SOM achieves a gain of " + }, + { + "bbox": [ + 313, + 434, + 556, + 579 + ], + "type": "inline_equation", + "content": "3.1\\%" + }, + { + "bbox": [ + 313, + 434, + 556, + 579 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 434, + 556, + 579 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 313, + 434, + 556, + 579 + ], + "type": "text", + "content": " mIoU on V21 and ADE datasets, respectively. Further, SAE enhances attention intensity in relevant regions, leading to an average gain of " + }, + { + "bbox": [ + 313, + 434, + 556, + 579 + ], + "type": "inline_equation", + "content": "1.7\\%" + }, + { + "bbox": [ + 313, + 434, + 556, + 579 + ], + "type": "text", + "content": " mIoU. Finally, HAI helps our SFP reach " + }, + { + "bbox": [ + 313, + 434, + 556, + 579 + ], + "type": "inline_equation", + "content": "63.9\\%" + }, + { + "bbox": [ + 313, + 434, + 556, + 579 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 434, + 556, + 579 + ], + "type": "inline_equation", + "content": "20.8\\%" + }, + { + "bbox": [ + 313, + 434, + 556, + 579 + ], + "type": "text", + "content": " mIoU on V21 and ADE datasets, which set the new SOTA performance. Due to space constraints, more ablation studies about SOM and SAE are summarized in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 341, + 586, + 531, + 644 + ], + "blocks": [ + { + "bbox": [ + 341, + 586, + 531, + 644 + ], + "lines": [ + { + "bbox": [ + 341, + 586, + 531, + 644 + ], + "spans": [ + { + "bbox": [ + 341, + 586, + 531, + 644 + ], + "type": "table", + "html": "
SOMSAEHAIV21ADEAvg.
057.016.736.9
1✓60.118.239.2
2✓✓62.319.540.9
3✓✓✓63.920.842.4
", + "image_path": "aac55b2e9d4d1c012b024fe53d12b1650d8fa30a3a6e0046227b762cfceb1075.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 655, + 555, + 678 + ], + "lines": [ + { + "bbox": [ + 313, + 655, + 555, + 678 + ], + "spans": [ + { + "bbox": [ + 313, + 655, + 555, + 678 + ], + "type": "text", + "content": "Table 2. Ablation results on each proposed module to validate the impact on performance across V21 [15] and ADE [57] datasets." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "More visualization on the ability of SOM to suppress outlier propagation We visualize purified feature maps" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20297" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 70, + 529, + 182 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 529, + 182 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 529, + 182 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 529, + 182 + ], + "type": "image", + "image_path": "a306f95aad08c2c2619221750b1d1304fdc64beae124dc2ea0e33db12deccce1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 191, + 555, + 215 + ], + "lines": [ + { + "bbox": [ + 55, + 191, + 555, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 191, + 555, + 215 + ], + "type": "text", + "content": "Figure 5. Visualization of the feature purification across CLIP's image encoder layers. (a) Feature maps without SOM purification. (b) Feature maps purified by SOM." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 82, + 233, + 272, + 285 + ], + "blocks": [ + { + "bbox": [ + 82, + 233, + 272, + 285 + ], + "lines": [ + { + "bbox": [ + 82, + 233, + 272, + 285 + ], + "spans": [ + { + "bbox": [ + 82, + 233, + 272, + 285 + ], + "type": "table", + "html": "
Al=l-1l=lo1581011
V2163.963.463.062.662.5
ADE20.820.520.319.919.7
Avg.42.442.041.741.341.1
", + "image_path": "c6be23c6829af8ab2b9909e984b217f66de15edb42067004b1360f8debf45aa3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 351, + 296, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 351, + 296, + 435 + ], + "spans": [ + { + "bbox": [ + 55, + 351, + 296, + 435 + ], + "type": "text", + "content": "across CLIP's layers, from the 2nd layer to the 9th layer, which is shown in Fig. 5. It's found that the outliers gradually affect the aggregated features, especially from layer 6 to layer 9 in Fig. 5 (a). Then, with the help of our SOM, the influence of outliers is mitigated layer by layer, like in Fig. 5 (b). Thus, our SOM can suppress the propagated outlier features and gradually purify the overall semantic features." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 435, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 435, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 296, + 555 + ], + "type": "text", + "content": "Effect on hierarchical attention integrator A series of ablation studies is conducted to investigate the optimal way to leverage attention maps from different layers. We omit HAI from our SFP to establish the baseline model for this experiment. As summarized in Tab. 3, we take " + }, + { + "bbox": [ + 55, + 435, + 296, + 555 + ], + "type": "inline_equation", + "content": "l_{0}" + }, + { + "bbox": [ + 55, + 435, + 296, + 555 + ], + "type": "text", + "content": " to represent the starting layer. When " + }, + { + "bbox": [ + 55, + 435, + 296, + 555 + ], + "type": "inline_equation", + "content": "l_{0} = 1" + }, + { + "bbox": [ + 55, + 435, + 296, + 555 + ], + "type": "text", + "content": ", our HAI performs best with all attention maps selected. Besides, increasing " + }, + { + "bbox": [ + 55, + 435, + 296, + 555 + ], + "type": "inline_equation", + "content": "l_{0}" + }, + { + "bbox": [ + 55, + 435, + 296, + 555 + ], + "type": "text", + "content": " from 1 to 11 decreases the segmentation performance, which demonstrates the effectiveness of the shallow attention maps for semantic feature refinement." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 81, + 563, + 272, + 676 + ], + "blocks": [ + { + "bbox": [ + 55, + 297, + 296, + 331 + ], + "lines": [ + { + "bbox": [ + 55, + 297, + 296, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 296, + 331 + ], + "type": "text", + "content": "Table 3. Ablation results on different selection strategies in HAI, where " + }, + { + "bbox": [ + 55, + 297, + 296, + 331 + ], + "type": "inline_equation", + "content": "A_{l = l_0}^{l = L - 1}" + }, + { + "bbox": [ + 55, + 297, + 296, + 331 + ], + "type": "text", + "content": " represents the integrated attention map from " + }, + { + "bbox": [ + 55, + 297, + 296, + 331 + ], + "type": "inline_equation", + "content": "l_0" + }, + { + "bbox": [ + 55, + 297, + 296, + 331 + ], + "type": "text", + "content": " to the " + }, + { + "bbox": [ + 55, + 297, + 296, + 331 + ], + "type": "inline_equation", + "content": "L - 1" + }, + { + "bbox": [ + 55, + 297, + 296, + 331 + ], + "type": "text", + "content": " layer. '1,5,8,10,11' indicates the value of " + }, + { + "bbox": [ + 55, + 297, + 296, + 331 + ], + "type": "inline_equation", + "content": "l_0" + }, + { + "bbox": [ + 55, + 297, + 296, + 331 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 81, + 563, + 272, + 676 + ], + "lines": [ + { + "bbox": [ + 81, + 563, + 272, + 676 + ], + "spans": [ + { + "bbox": [ + 81, + 563, + 272, + 676 + ], + "type": "table", + "html": "
BackboneMethodsV21ADEAvg.
ViT-B/32SCLIP [45]50.614.832.7
LaVG [19]54.815.535.2
ProxyCLIP [25]57.916.737.3
NACLIP [16]51.114.933.0
SFP60.117.138.6
ViT-L/14SCLIP [45]44.410.927.7
LaVG [19]52.117.334.7
ProxyCLIP [25]60.622.641.6
NACLIP [16]52.217.334.8
SFP64.722.143.4
", + "image_path": "06423330e4a9bfd769da160ac47aab784abbd036924c28201adba36206085470.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 688, + 295, + 709 + ], + "lines": [ + { + "bbox": [ + 55, + 688, + 295, + 709 + ], + "spans": [ + { + "bbox": [ + 55, + 688, + 295, + 709 + ], + "type": "text", + "content": "Table 4. Comparison with other SOTA methods across different CLIP backbones." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 236, + 555, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 555, + 380 + ], + "type": "text", + "content": "Effect on different CLIP backbones To validate the robustness, we further evaluate our SFP with other methods using different CLIP backbones, such as ViT-B/32 and ViT-L/14. As summarized in Tab. 4, some methods, like SCLIP [45] and LaVG [19], suffer from performance degradation when applied to the larger backbone. In contrast, SFP maintains superior performance across different backbones and achieves an improvement of " + }, + { + "bbox": [ + 313, + 236, + 555, + 380 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 313, + 236, + 555, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 236, + 555, + 380 + ], + "type": "inline_equation", + "content": "1.8\\%" + }, + { + "bbox": [ + 313, + 236, + 555, + 380 + ], + "type": "text", + "content": " mIoU over ProxyCLIP [25] with ViT-B/32 and ViT-L/14 backbone, respectively. These results indicate that SFP can robustly handle outliers derived from different backbones and generate purified features for precise segmentation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 399, + 388, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 399, + 388, + 412 + ], + "spans": [ + { + "bbox": [ + 314, + 399, + 388, + 412 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 423, + 555, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 555, + 614 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 555, + 614 + ], + "type": "text", + "content": "In this paper, we present a Self-adaptive Feature Purifier framework (SFP) to mitigate outlier propagation for training-free open-vocabulary semantic segmentation. Specifically, we propose SOM deployed at each layer of the image encoder to identify and mitigate outliers by comparing attention weights between image and class tokens, which generates purified features for propagation. Next, to recover the semantic information affected by outliers, we introduce SAE to strengthen object activation in the purified feature. Beyond that, we design HAI to utilize multilayer attention maps for refined object-centric representations. Extensive experiments show that SFP achieves superior segmentation performance compared to existing SOTA methods on various datasets. Our SFP shows that feature purification does matter for OVSS, and cleaner features can predict more complete and accurate segmentation masks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": "Acknowledgements. This work was supported by the National Natural Science Foundation of China (No. 62471405, 62331003, 62301451,62301613), Basic Research Program of Jiangsu (BK20241814), Suzhou Basic Research Program (SYG202316) and XJTLU REF-22-01-010 and RDF-22-02-066, XJTLU AI University Research Centre, Jiangsu Province Engineering Research Centre of Data Science and Cognitive Computation at XJTLU." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20298" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "text", + "content": "[1] Nikita Araslanov and Stefan Roth. Single-stage semantic segmentation from image labels. In CVPR, pages 4253-4262, 2020. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 124, + 296, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 124, + 296, + 167 + ], + "spans": [ + { + "bbox": [ + 61, + 124, + 296, + 167 + ], + "type": "text", + "content": "[2] Sule Bai, Yong Liu, Yifei Han, Haoji Zhang, and Yansong Tang. Self-calibrated clip for training-free open-vocabulary segmentation. arXiv preprint arXiv:2411.15869, 2024. 2, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 168, + 295, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 168, + 295, + 213 + ], + "spans": [ + { + "bbox": [ + 62, + 168, + 295, + 213 + ], + "type": "text", + "content": "[3] Walid Bousselham, Felix Petersen, Vittorio Ferrari, and Hilde Kuehne. Grounding everything: Emerging localization properties in vision-language transformers. In CVPR, pages 3828-3837, 2024. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 213, + 294, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 213, + 294, + 246 + ], + "spans": [ + { + "bbox": [ + 62, + 213, + 294, + 246 + ], + "type": "text", + "content": "[4] Maxime Bucher, Tuan-Hung Vu, Matthieu Cord, and Patrick Pérez. Zero-shot semantic segmentation. In NeurlPS, pages 468-479, 2019. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 247, + 294, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 294, + 291 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 294, + 291 + ], + "type": "text", + "content": "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 292, + 294, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 292, + 294, + 345 + ], + "spans": [ + { + "bbox": [ + 62, + 292, + 294, + 345 + ], + "type": "text", + "content": "[6] Niccolo Cavagnero, Gabriele Rosi, Claudia Cuttano, Francesca Pistilli, Marco Ciccone, Giuseppe Avera, and Fabio Cermelli. Pem: Prototype-based efficient maskformer for image segmentation. In CVPR, pages 15804-15813, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 346, + 294, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 346, + 294, + 389 + ], + "spans": [ + { + "bbox": [ + 62, + 346, + 294, + 389 + ], + "type": "text", + "content": "[7] Junbum Cha, Jonghwan Mun, and Byungseok Roh. Learning to generate text-grounded mask for open-world semantic segmentation from only image-text pairs. In CVPR, pages 11165-11174, 2023. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 391, + 294, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 391, + 294, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 391, + 294, + 434 + ], + "type": "text", + "content": "[8] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, pages 1290–1299, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 435, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 435, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 435, + 294, + 491 + ], + "type": "text", + "content": "[9] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, pages 2818-2829, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 491, + 294, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 491, + 294, + 535 + ], + "spans": [ + { + "bbox": [ + 57, + 491, + 294, + 535 + ], + "type": "text", + "content": "[10] Seokju Cho, Heeseong Shin, Sunghwan Hong, Anurag Arnab, Paul Hongsuck Seo, and Seungryong Kim. Cat-seg: Cost aggregation for open-vocabulary semantic segmentation. In CVPR, pages 4113-4123, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 536, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 536, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 536, + 294, + 590 + ], + "type": "text", + "content": "[11] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, pages 3213-3223, 2016. 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 591, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 294, + 624 + ], + "type": "text", + "content": "[12] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. In ICLR, pages 1-11, 2024. 1, 2, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 624, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 294, + 657 + ], + "type": "text", + "content": "[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, pages 248-255, 2009. 4, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 658, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 294, + 690 + ], + "type": "text", + "content": "[14] Zheng Ding, Jieke Wang, and Zhuowen Tu. Open-vocabulary panoptic segmentation with maskclip. arXiv preprint arXiv:2208.08984, 2022. 2, 3, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "text", + "content": "[15] Mark Everingham, SM Ali Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "pascal visual object classes challenge: A retrospective. IJCV, 111:98-136, 2015. 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 95, + 553, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 95, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 317, + 95, + 553, + 138 + ], + "type": "text", + "content": "[16] Sina Hajimiri, Ismail Ben Ayed, and Jose Dolz. Pay attention to your neighbours: Training-free open-vocabulary semantic segmentation. In WACV, pages 5061-5071, 2025. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 140, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 194 + ], + "type": "text", + "content": "[17] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 195, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 195, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 317, + 195, + 553, + 237 + ], + "type": "text", + "content": "[18] Siyu Jiao, Yunchao Wei, Yaowei Wang, Yao Zhao, and Humphrey Shi. Learning mask-aware clip representations for zero-shot segmentation. In NeurIPS, pages 35631-35653, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 239, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 271 + ], + "type": "text", + "content": "[19] Dahiyun Kang and Minsu Cho. In defense of lazy visual grounding for open-vocabulary semantic segmentation. In ECCV, pages 143-164, 2024. 3, 6, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 272, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 304 + ], + "type": "text", + "content": "[20] Laurynas Karazija, Iro Laina, Andrea Vedaldi, and Christian Rupprecht. Diffusion models for zero-shot open-vocabulary segmentation. In ECCV, pages 299-317, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 305, + 553, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 553, + 349 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 553, + 349 + ], + "type": "text", + "content": "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, pages 4015-4026, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 350, + 553, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 382 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 382 + ], + "type": "text", + "content": "[22] Philip A Knight. The sinkhorn-knopp algorithm: convergence and applications. SIAM Journal on Matrix Analysis and Applications, 30(1):261-275, 2008. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 383, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 415 + ], + "type": "text", + "content": "[23] Philipp Kähenbuhl and Vladlen Koltun. Efficient inference in fully connected crfs with gaussian edge potentials. In NeurIPS, pages 109-117, 2011. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 416, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 553, + 459 + ], + "type": "text", + "content": "[24] Mengcheng Lan, Chaofeng Chen, Yiping Ke, Xinjiang Wang, Litong Feng, and Wayne Zhang. Clearclip: Decomposing clip representations for dense vision-language inference. In ECCV, pages 143-160, 2024. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 460, + 553, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 503 + ], + "type": "text", + "content": "[25] Mengcheng Lan, Chaofeng Chen, Yiping Ke, Xinjiang Wang, Litong Feng, and Wayne Zhang. Proxyclip: Proxy attention improves clip for open-vocabulary segmentation. In ECCV, pages 70-88, 2024. 3, 6, 7, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 504, + 553, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 553, + 548 + ], + "type": "text", + "content": "[26] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 548, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 592 + ], + "type": "text", + "content": "[27] Yunheng Li, Zhongyu Li, Quansheng Zeng, Qibin Hou, and Ming-Ming Cheng. Cascade-clip: Cascaded vision-language embeddings alignment for zero-shot semantic segmentation. In ICML, pages 28243–28258, 2024. 2, 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 593, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 553, + 635 + ], + "type": "text", + "content": "[28] Yi Li, Hualiang Wang, Yiqun Duan, Jiheng Zhang, and Xiaomeng Li. A closer look at the explainability of contrastive language-image pre-training. Pattern Recognition, 162:111409, 2025. 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 636, + 553, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 680 + ], + "type": "text", + "content": "[29] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[30] Yong Liu, Sule Bai, Guanbin Li, Yitong Wang, and Yansong Tang. Open-vocabulary segmentation with semantic-assisted calibration. In CVPR, pages 3491-3500, 2024. 1, 2" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20299" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[31] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, pages 3431-3440, 2015. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 294, + 150 + ], + "type": "text", + "content": "[32] Huaishao Luo, Junwei Bao, Youzheng Wu, Xiaodong He, and Tianrui Li. Segclip: Patch aggregation with learnable centers for open-vocabulary semantic segmentation. In ICML, pages 23033-23044, 2023. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "type": "text", + "content": "[33] Jiayun Luo, Siddhesh Khandelwal, Leonid Sigal, and Boyang Li. Emergent open-vocabulary semantic segmentation from off-the-shelf vision-language models. In CVPR, pages 4029-4040, 2024. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 194, + 294, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 294, + 248 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 294, + 248 + ], + "type": "text", + "content": "[34] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The role of context for object detection and semantic segmentation in the wild. In CVPR, pages 891–898, 2014. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 250, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 304 + ], + "type": "text", + "content": "[35] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 305, + 294, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 305, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 294, + 350 + ], + "type": "text", + "content": "[36] Zelin Peng, Zhengqin Xu, Zhilin Zeng, Yu Huang, Yaoming Wang, and Wei Shen. Parameter-efficient fine-tuning in hyperspherical space for open-vocabulary semantic segmentation. In CVPR, pages 15009-15020, 2025. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 350, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 294, + 403 + ], + "type": "text", + "content": "[37] Zelin Peng, Zhengqin Xu, Zhilin Zeng, Changsong Wen, Yu Huang, Menglin Yang, Feilong Tang, and Wei Shen. Understanding fine-tuning clip for open-vocabulary semantic segmentation in hyperbolic space. In CVPR, pages 4562-4572, 2025. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 404, + 294, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 459 + ], + "type": "text", + "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 459, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 459, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 459, + 294, + 525 + ], + "type": "text", + "content": "[39] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, pages 25278–25294, 2022. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 525, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 294, + 568 + ], + "type": "text", + "content": "[40] Tong Shao, Zhuotao Tian, Hang Zhao, and Jingyong Su. Explore the potential of clip for training-free open vocabulary semantic segmentation. In ECCV, pages 139-156, 2024. 1, 2, 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 570, + 294, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 294, + 603 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 294, + 603 + ], + "type": "text", + "content": "[41] Gyungin Shin, Weidi Xie, and Samuel Albanie. Reco: Retrieve and co-segment for zero-shot transfer. In NeurIPS, pages 33754-33767, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 603, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 646 + ], + "type": "text", + "content": "[42] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, pages 2256-2265, 2015. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 294, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 680 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 680 + ], + "type": "text", + "content": "[43] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In ICCV, pages 7262-7272, 2021. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[44] Shuyang Sun, Runjia Li, Philip Torr, Xiuye Gu, and Siyang Li. Clip as rn: Segment countless visual concepts without training endeavor. In CVPR, pages 13171-13182, 2024. 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 553, + 703 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 315, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 553, + 106 + ], + "type": "text", + "content": "[45] Feng Wang, Jieru Mei, and Alan Yuille. Sclip: Rethinking self-attention for dense vision-language inference. In ECCV, pages 315–332, 2024. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 106, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 106, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 106, + 553, + 171 + ], + "type": "text", + "content": "[46] Haoxiang Wang, Pavan Kumar Anasosalu Vasu, Fartash Faghri, Raviteja Vemulapalli, Mehrdad Farajtabar, Sachin Mehta, Mohammad Rastegari, Oncel Tuzel, and Hadi Pouransari. Sam-clip: Merging vision foundation models towards semantic and spatial understanding. In CVPR, pages 3635–3647, 2024. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 172, + 553, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 172, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 316, + 172, + 553, + 216 + ], + "type": "text", + "content": "[47] Monika Wysoczanska, Oriane Simeoni, Michael Ramamonjisoa, Andrei Bursuc, Tomasz Trzciński, and Patrick Pérez. Clip-dinoiser: Teaching clip a few dino tricks. In ECCV, pages 320-337, 2024. 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 216, + 553, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 216, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 316, + 216, + 553, + 258 + ], + "type": "text", + "content": "[48] Bin Xie, Jiale Cao, Jin Xie, Fahad Shahbaz Khan, and Yanwei Pang. Sed: A simple encoder-decoder for open-vocabulary semantic segmentation. In CVPR, pages 3426-3436, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 259, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 259, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 259, + 553, + 304 + ], + "type": "text", + "content": "[49] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. In NeurIPS, pages 12077-12090, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 305, + 553, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 553, + 350 + ], + "type": "text", + "content": "[50] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. Groupvit: Semantic segmentation emerges from text supervision. In CVPR, pages 18134-18144, 2022. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "text", + "content": "[51] Jilan Xu, Junlin Hou, Yuejie Zhang, Rui Feng, Yi Wang, Yu Qiao, and Weidi Xie. Learning open-vocabulary semantic segmentation models from natural language supervision. In CVPR, pages 2935-2944, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 393, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 553, + 426 + ], + "type": "text", + "content": "[52] Mengde Xu, Zheng Zhang, Fangyun Wei, Han Hu, and Xiang Bai. Side adapter network for open-vocabulary semantic segmentation. In CVPR, pages 2945-2954, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 426, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 469 + ], + "type": "text", + "content": "[53] Xihong Yang, Xiaochang Hu, Sihang Zhou, Xinwang Liu, and En Zhu. Interpolation-based contrastive learning for few-label semi-supervised learning. TNNLS, 35(2):2054-2065, 2022. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 470, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 470, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 470, + 553, + 514 + ], + "type": "text", + "content": "[54] Xihong Yang, Jin Jiaqi, Siwei Wang, Ke Liang, Yue Liu, Yi Wen, Suyuan Liu, Sihang Zhou, Xinwang Liu, and En Zhu. Dealmvc: Dual contrastive calibration for multi-view clustering. In ACMMM, pages 337-346, 2023." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 514, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 514, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 316, + 514, + 553, + 568 + ], + "type": "text", + "content": "[55] Xihong Yang, Siwei Wang, Fangdi Wang, Jiaqi Jin, Suyuan Liu, Yue Liu, En Zhu, Xinwang Liu, and Yueming Jin. Automatically identify and rectify: Robust deep contrastive multiview clustering in noisy scenarios. In ICML. PMLR, 2025. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 569, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 569, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 569, + 553, + 613 + ], + "type": "text", + "content": "[56] Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang-Chieh Chen. Convolutions die hard: Open-vocabulary segmentation with single frozen convolutional clip. In NeurIPS, pages 32215-32234, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 613, + 553, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 647 + ], + "type": "text", + "content": "[57] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ADE20K dataset. In CVPR, pages 5122-5130, 2017. 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 647, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 669 + ], + "type": "text", + "content": "[58] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In ECCV, pages 696-712, 2022. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 670, + 553, + 703 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 703 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 703 + ], + "type": "text", + "content": "[59] Ziqin Zhou, Yinjie Lei, Bowen Zhang, Lingqiao Liu, and Yifan Liu. Zegclip: Towards adapting clip for zero-shot semantic segmentation. In CVPR, pages 11175-11185, 2023." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20300" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/FedAGC_ Federated Continual Learning with Asymmetric Gradient Correction/69b6bc67-ec8c-4e1a-b957-09f50cc21aab_content_list.json b/2025/FedAGC_ Federated Continual Learning with Asymmetric Gradient Correction/69b6bc67-ec8c-4e1a-b957-09f50cc21aab_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f4cf010ad62f9afc26633ebe6adf608c365afa3f --- /dev/null +++ b/2025/FedAGC_ Federated Continual Learning with Asymmetric Gradient Correction/69b6bc67-ec8c-4e1a-b957-09f50cc21aab_content_list.json @@ -0,0 +1,1664 @@ +[ + { + "type": "text", + "text": "FedAGC: Federated Continual Learning with Asymmetric Gradient Correction", + "text_level": 1, + "bbox": [ + 96, + 130, + 898, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chengchao Zhang $^{1}$ , Fanhua Shang $^{1\\dagger}$ , Hongyin Liu $^{2\\dagger}$ , Liang Wan $^{1,2}$ , Wei Feng $^{1}$ , $^{1}$ College of Intelligence and Computing, Tianjin University, China $^{2}$ Medical College, Tianjin University, China", + "bbox": [ + 187, + 179, + 807, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{chengchaozhang, fhshang, hyliu2009, lwan, wfeng}@tju.edu.cn", + "bbox": [ + 228, + 236, + 759, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Federated Continual Learning (FCL) has emerged as a prominent distributed learning paradigm and aims at addressing model learning challenges in both federated and continual learning settings. Efficient personalization in FCL remains a major challenge, as it must handle not only conflicts between old and new knowledge within parallel task streams but also heterogeneous knowledge conflicts from different clients. Recent approaches attempt to mitigate these issues through gradient correction. However, they often overlook the combined impact of gradient magnitude and direction, leading to unsatisfactory gradient solutions. To address these issues, we propose a novel federated continual learning method (called FedAGC) with asymmetric gradient correction, which performs memory rehearsal using representative samples selected via a centroid-based approach from historical tasks. By formulating the problem as a multi-objective optimization problem, FedAGC derives more effective gradients while incorporating group-level personalization to facilitate useful knowledge integration and irrelevant knowledge isolation, effectively mitigating both temporal and spatial catastrophic forgetting. Extensive experiments confirm the effectiveness of FedAGC.", + "bbox": [ + 88, + 318, + 483, + 651 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 681, + 220, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Federated Learning (FL) [21, 22, 29], as a novel distributed machine learning paradigm, aims to train a global model by leveraging independent and diverse data across clients in a privacy-preserving and collaborative manner. FL has been widely applied in various domains, such as medical image analysis [23] and object detection [24]. It has been proven to effectively address the issue of data silos [33] without compromising client privacy. Most existing FL methods are limited to single-task learning, assuming that client data remains static over time. However, in real-world scenarios, data is continuously evolving, forming sequential tasks with", + "bbox": [ + 89, + 708, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5f89dc36fcd47dabdddc7f66f2e9a0e08331de42e8d0486e356c90553ec4ee86.jpg", + "image_caption": [ + "Figure 1. Spatial-Temporal Catastrophic Forgetting. (a) Direct access to new tasks leads to a significant performance decline on old tasks. (b) Non-IID data exacerbates the weight disparity." + ], + "image_footnote": [ + "Local update of client 1", + "Local update of client $N$", + "Aggregation of server", + "$\\star$ Global optimum", + "Server global model", + "Client local model", + "- Initial model", + "$\\iff$ Parameter divergence", + "Client drift", + "(b) Spatial catastrophic forgetting" + ], + "bbox": [ + 516, + 284, + 924, + 412 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "unpredictable future changes. As shown in Fig. 1a, severe temporal catastrophic forgetting (TemporalCF) [38, 45] occurs when the local model on each client learns a newly introduced task, leading to performance degradation on previously learned tasks as new knowledge is acquired.", + "bbox": [ + 511, + 612, + 906, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Improving the generalizability of FL in multi-task evolving scenarios has emerged as a critical and challenging research topic. Recently, Continual Learning (CL) [17] has made significant advances in addressing TemporalCF, enabling effective mitigation of this problem. Based on such characteristics, the concept of Federated Continual Learning (FCL) [6, 46, 47] came into being, which integrates CL methods into the FL framework. In the FCL setting, FL can be regarded as a distributed instance of conventional CL, while CL methods assist in addressing FL challenges.", + "bbox": [ + 511, + 688, + 906, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "FCL breaks the static limitation of traditional FL, where each client is confined to learning a single task sequence. However, integrating CL into the FL framework poses substantial challenges. Besides TemporalCF, FCL also", + "bbox": [ + 511, + 839, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 109, + 886, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3841", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "faces spatial catastrophic forgetting (SpatialCF). Specifically, the private data of different clients in FCL may follow drastically different distributions. This situation, where data across clients is Non-Independent and Identically Distributed (Non-IID), which is referred to as statistical heterogeneity [26]. During server-side aggregation, where gradients from multiple clients are combined to initialize the next round of local CL training, statistical heterogeneity can lead to issues such as parameter divergence and client drift [14, 37], as illustrated in Fig. 1b. Consequently, the aggregated global model exhibits reduced generalizability, further exacerbating temporal catastrophic forgetting.", + "bbox": [ + 89, + 90, + 480, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the aforementioned challenges, recent studies [42, 47, 48] have explored the problem of Personalized Federated Continual Learning (PFCL). Unlike traditional FCL that aim to obtain a universal global model, PFCL leverages efficient server-side aggregation to facilitate personalized training. Existing PFCL methods mainly focus on optimizing local models for clients, ensuring strong performance on both new and previous tasks in a parallel, independent, and heterogeneous task stream across clients. The most relevant work to ours is FedMes [42], which builds on the gradient correction technique of A-GEM [2] and enhances local models through KNN-based inference. However, existing methods typically rely solely on gradient direction to determine similarity, ignoring the impact of magnitude, leading to suboptimal gradients. In contrast, we propose a new Asymmetric Gradient Correction (AGC) technique, which adaptively balances the combined impact of direction and magnitude, enabling a smoother optimization transition.", + "bbox": [ + 91, + 275, + 482, + 546 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We formulate the PFCL problem as a Multi-Objective Optimization (MOO) problem to obtain the optimal gradient, where each objective minimizes the distance to a target gradient. Specifically, to mitigate Spatial-Temporal Catastrophic Forgetting (STCF), each client extracts a small set of representative samples after completing local training for each task and employs them for rehearsal in subsequent learning. Subsequently, we correct the gradients by considering both their magnitudes and directions, ultimately obtaining an optimal gradient that does not damage any of old or new knowledge at each update from a set of class-level gradients. Additionally, empirical evidence suggests that a single global model has limited generalization across all clients [34]. Therefore, we introduce a group-level personalization strategy, where clustering is performed based on the similarity of task gradients uploaded by participating clients. These clusters are then divided into groups, enabling joint training of multiple global models.", + "bbox": [ + 89, + 547, + 482, + 820 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions can be summarized as follows:", + "bbox": [ + 109, + 823, + 470, + 837 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To the best of our knowledge, this is the first attempt to unify both spatial and temporal catastrophic forgetting in PFCL through a multi-objective optimization framework.", + "- We propose a novel asymmetric gradient correction algo" + ], + "bbox": [ + 89, + 839, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "rithm that performs memory rehearsal using representative samples selected via a centroid-based approach from historical tasks, transforming multi-task updates into an optimal gradient search problem that minimizes inter-task discrepancies and mitigates TemporalCF.", + "bbox": [ + 526, + 90, + 903, + 167 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a new group-level personalization approach that clusters clients by gradient similarity and aggregates multiple global models, isolating irrelevant knowledge interference while fostering the positive fusion of related knowledge to mitigate SpatialCF.", + "- Our method outperforms all the baselines in terms of accuracy and forgetting rate, demonstrating its efficacy in addressing the PFCL problem." + ], + "bbox": [ + 511, + 167, + 903, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Background and Related Works", + "text_level": 1, + "bbox": [ + 511, + 301, + 807, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "FL Algorithms for Statistical Heterogeneity. The client-server architecture-based FL paradigm is currently the most widely adopted framework. This distributed learning model involves a central server for aggregating network model parameters or gradients and $N$ clients $\\{F_1, F_2, \\ldots, F_N\\}$ for participating in the training of local private data, where each client possesses a local private dataset $\\{D_1, D_2, \\ldots, D_N\\}$ . For the $i$ -th client $F_i$ , its local dataset includes $n_i$ samples, denoted as $\\{D_i = \\{(X_k^i, Y_k^i)\\}_{k=1}^{n_i}$ . Different clients may exhibit distinct features and labels, such that $X_i \\cap X_j = \\emptyset$ for $\\forall i, j \\in [1, N]$ , and $Y_k \\neq \\emptyset$ , $Y_i = \\emptyset$ for $\\forall i \\in [1, N] \\setminus \\{k\\}$ . The objective of FL is to collaboratively and privately train a global statistical model across multiple distributed clients under the supervision of a central server. The key formulation controlling FL in the classical FedAvg [29] is as follows: $\\min_{w \\in \\mathbb{R}^d} f(w) = \\frac{1}{m} \\sum_{i=1}^{m} f_i(w)$ . where $f_i(w) = \\mathbb{E}_{(x_i, y_i) \\sim \\mathcal{P}_i}[\\mathcal{L}_i(w, x_i, y_i)]$ represents the local loss function of the $i$ -th client based on its local data, and $\\mathcal{P}_i$ denotes the local data distribution of the $i$ -th client. To address the challenge of statistical heterogeneity and mitigate client drift caused by Non-IID data distributions, FedProx [19] extends FedAvg by introducing a proximal term that effectively constrains local updates, thereby enhancing the overall stability of model convergence. Another representative approach, SCAFFOLD [14], employs control variates to alleviate client drift. However, under extreme Non-IID conditions, these methods often struggle to produce a highly generalizable global model. Consequently, Personalized FL [20, 45] shifts its focus to providing specialized local models for each client, enabling personalized services. PerFedAvg [7] combines meta-learning techniques during a pre-training phase to obtain a well-initialized model. While these approaches have demonstrated strong performance in FL, they are primarily designed to address SpatialCF and may not be sufficient for more complex scenarios.", + "bbox": [ + 511, + 325, + 906, + 853 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Gradient Correction-Based CL Algorithms. In contrast to traditional FL designed for single-task scenarios, in a standard continual learning setting, there are a sequence", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3842", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/cc19b29212009502a2e6eeab4157dd9df203fc025f7a5c96d141eb41cacd00b0.jpg", + "image_caption": [ + "Figure 2. The workflow of the proposed FedAGC method. Our FedAGC consists of the following three modules: On the client side, the Centroid-based Representative Sample Selection (CBRSS) module rehearses the gradients of previous tasks to assist the Asymmetric Correction Gradient Updating (ACGU) module in generating optimal gradients. On the server side, the Group Aggregation Gradient Updating (GAGU) module clusters clients based on gradient similarity and aggregates multiple global models." + ], + "image_footnote": [], + "bbox": [ + 137, + 85, + 859, + 392 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of tasks $\\mathcal{T} = \\{\\mathcal{T}_1,\\mathcal{T}_2,\\dots ,\\mathcal{T}_T\\}$ , where $T$ is the number of tasks. For the $t$ -th task $\\mathcal{T}_t$ , its dataset is composed of $n_t$ pairs of data and labels: $\\{x_i^t,y_i^t\\}_{i = 1}^{n_t}$ . Typically, $\\mathcal{T}$ refers to a task stream where the arrival order of new tasks is unknown. As a result, the model can only access $\\mathcal{T}_t$ during the training of the $t$ -th task, and access to it is no longer available afterward. Therefore, for $\\mathcal{T}_t$ , given the set of model parameters $w^{(t)}$ learned so far, the learning objective at task $t$ is: $\\min_{w^{(t)}}\\mathcal{L}(w^{(t)};w^{(t - 1)},\\mathcal{T}^t)$ . In recent years, numerous CL algorithms have adopted gradient correction strategies to mitigate TemporalCF, typically involving two key steps: Memory Rehearsal and Optimal Gradient Computation [15, 27]. A representative example is GEM [25], which retrieves gradients from previously stored memories and integrates them with the current task's gradient to obtain an update direction suitable for all tasks. As one variant, A-GEM [2], consolidates multiple memories to enhance computational efficiency. Meanwhile, ER [3] combines data from episodic memory with the current task during training. Although ER is simple and effective, its performance depends heavily on the quality of the stored memories. Another related approach is parameter regularization [5, 8], which introduces additional regularization terms to consolidate knowledge from previous tasks and reduce interference between new and old task parameters. Among these methods, EWC [38] employs the Fisher information matrix to quantify the importance of parameters for previously learned tasks.", + "bbox": [ + 91, + 470, + 483, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While FL and CL have achieved remarkable success in", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "their respective domains, effectively mitigating the STCF problem in PFCL remains a challenging endeavor [12, 45]. FedCurv [38] incorporates EWC to handle statistical heterogeneity; however, it incurs high computational overhead and lacks a dedicated mechanism for addressing TemporalCF. FedKNOW [44] deploys a gradient integrator at each client to reduce negative knowledge transfer by effectively combining signature tasks identified from past and current tasks of other clients through the global model. Recent advances in vision foundation models have driven an alternative prompt-based approach [9, 11, 43, 48] that mitigates STCF by adapting to new tasks through training only a small subset of prompts and classifier parameters.", + "bbox": [ + 511, + 472, + 908, + 670 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. The Proposed Method", + "text_level": 1, + "bbox": [ + 511, + 676, + 723, + 695 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we elaborate on the proposed FedAGC. For the first time, FedAGC formulates the PFCL problem (see Sec. 3.1) as a MOO problem. On the client-side, after training each task, we select representative samples to construct a memory cache (see Sec. 3.2) and use them to perform asymmetric gradient correction for the subsequent new tasks (see Sec. 3.3), thus avoiding TemporalCF. On the server-side, we group the average gradient updates uploaded by the clients (see Sec. 3.4) and aggregate multiple global models to avoid SpatialCF, while also satisfying the personalized local requirements.", + "bbox": [ + 511, + 703, + 906, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall framework of FedAGC is shown in Figure 2, the algorithm is summarized in Algorithm 1.", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3843", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Personalized Federated Continual Learning", + "text_level": 1, + "bbox": [ + 89, + 90, + 464, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We extend the traditional CL and FL methods to PFCL, which enables the independent and parallel integration of multiple task sequences across clients. This approach achieving both the fusion of local knowledge from new and old tasks and the integration of heterogeneous task knowledge across clients. Unlike FCL, which aims to obtain a general model that integrates global knowledge, PFCL addresses the diverse local needs of clients in highly heterogeneous scenarios, resulting in personalized local models. Specifically, each client learns its local model by optimizing the following objective:", + "bbox": [ + 89, + 112, + 483, + 279 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {w ^ {(t)}} \\mathcal {L} \\left(w ^ {(t)}; \\mathcal {T} ^ {t}\\right) + \\frac {\\lambda}{2} \\| w ^ {(t)} - w _ {0} \\| ^ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 286, + 482, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}(\\cdot)$ represents the empirical risk loss, $\\lambda$ is the regularization coefficient, and $w_{0}$ denotes the initial model.", + "bbox": [ + 89, + 325, + 482, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The PFCL scenario we consider, involving independent parallel multi-task sequences, comprises three key steps:", + "bbox": [ + 89, + 354, + 482, + 386 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) Given $N$ clients $\\{\\mathcal{C}_1,\\mathcal{C}_2,\\dots ,\\mathcal{C}_N\\}$ and a server $\\mathcal{S}$ . There are task streams $\\{\\mathcal{T}^1,\\mathcal{T}^2,\\ldots ,\\mathcal{T}^N\\}$ , within which each client $\\mathcal{C}_c$ engages in training a local CL model on its unique, private, and accessible task sequence $\\mathcal{T}^c = \\{\\mathcal{T}_1^c,\\mathcal{T}_2^c,\\dots ,\\mathcal{T}_T^c\\}$ . $\\mathcal{T}^c$ consists of $T$ distinct tasks, where the $t$ -th task $\\mathcal{T}_t^c$ includes $K_{t}^{c}$ classes, and the total set of classes for $\\mathcal{C}_c$ is represented as $K^{c} = \\{K_{1}^{c},K_{2}^{c},\\dots ,K_{T}^{c}\\}$ .", + "(2) For the current task $\\mathcal{T}_t^c$ being learned, the participating client receives the model $\\theta_g^t$ globally broadcasted by $S$ as the initial model, where $\\theta_g^t$ incorporates knowledge from the previous $t - 1$ tasks contributed by various clients. After $e$ rounds of local training on $\\mathcal{T}_t^c$ , $\\mathcal{C}_c$ obtains a new continual learning model $\\theta_{l,c}^{t}$ which performs well on $K^c$ , and the corresponding average gradient update $G^{c}$ .", + "(3) In each communication round $r$ , the participating client $\\mathcal{C}_c$ upload its respective $G^c$ . $S$ employs a group personalization strategy to aggregate them, which is then used to update $A$ global models $\\theta_{g,A}^t$ . For the $a$ -th global model $\\theta_{g,a}^t$ , it encapsulates the generalized knowledge of the associated client group's learning tasks and serves as the initial model for the next training round." + ], + "bbox": [ + 89, + 386, + 482, + 704 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Memory Rehearsal in FCL", + "text_level": 1, + "bbox": [ + 89, + 712, + 334, + 728 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Benefiting from the efficient utilization of past experience, memory rehearsal methods [2, 10, 27] are widely used in CL as simple and effective forgetting suppression techniques. These methods involve storing partial data or knowledge from previous tasks to enable retraining during the learning of new tasks, thereby mitigating forgetting of prior tasks. Now, we extend it from CL to FCL.", + "bbox": [ + 89, + 734, + 482, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Memory Design. Each client utilizes its limited local memory to construct a small cache space, denoted as $M_{c}$ . Upon the completion of the training process for the $t$ -th task of $\\mathcal{C}_c$ , rehearsal samples $M_t^c$ are sampled at a fixed proportion", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mu$ from the training set $D_{t}^{c}$ of that task. For the initial task, $M_0^c = \\emptyset$ , and for the final task, no additional sampling operations are performed. Thus, $M_{c}$ can be represented as", + "bbox": [ + 511, + 90, + 906, + 137 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM _ {c} = \\cup_ {t = 1} ^ {T} M _ {t} ^ {c}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 653, + 143, + 906, + 162 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the size of $M_t^c$ is significantly smaller than $D_t^c$ .", + "bbox": [ + 511, + 170, + 870, + 186 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Centroid-Based Representative Sample Selection. Clearly, the quality of samples drawn from the old task dataset directly affects the quality of the memory cache. Therefore, we propose a Centroid-Based Representative Sample Selection (CBRSS) method to effectively acquire high-quality rehearsal samples. In particular, for client $\\mathcal{C}_c$ , given the distance threshold $v$ for centroids and the input data $x$ , the constraint condition for the minimum preset distance is defined as:", + "bbox": [ + 511, + 185, + 906, + 321 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {i} \\| m _ {i} ^ {c} - f _ {\\theta , c} (x) \\| > v, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 330, + 906, + 352 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $m_i^c$ is the $i$ -th centroid, and $f_{\\theta,c}(x)$ is the predictor of $\\mathcal{C}_c$ for the current task, used to map $x$ to the feature space. If the condition in Eq. (3) is satisfied, it indicates that this data point is independent and can be considered as a new centroid $m_{new,c} = f_{\\theta,c}(x)$ . Otherwise, it means that this data point belongs to the existing centroid with the smallest distance to it, and then the target centroid is updated by", + "bbox": [ + 511, + 361, + 906, + 467 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nm _ {i ^ {*}} = \\frac {N _ {i ^ {*}} ^ {k} \\times m _ {i} + f _ {\\theta , c} (x)}{N _ {i ^ {*}} ^ {k} + 1}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 611, + 474, + 905, + 510 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $i^{*} = \\underset {i}{\\arg \\min}\\| f_{\\theta ,c}(x) - m_{i}^{c}\\|$ , and $N_{i^*}^k$ represents the total number of samples in the $k$ -th category captured by the centroid closest to $x$ in the current task.", + "bbox": [ + 511, + 517, + 905, + 570 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overall, each client stores the representative data of each category for the trained task after completing the training. A similar approach is employed by FedMes [42], which utilizes a random sampling method to construct the memory cache. However, due to the small sample size, this approach inevitably results in significant sampling bias. The resulting $M_{t}^{c}$ lacks representativeness and fails to reflect the overall distribution of $D_{t}^{c}$ . This leads to a drift in the representation of old task data in the feature space, which in turn triggers the problem of catastrophic forgetting.", + "bbox": [ + 511, + 570, + 906, + 722 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Asymmetrically Correcting Gradient Updates", + "text_level": 1, + "bbox": [ + 511, + 729, + 901, + 746 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To integrate the knowledge from previous tasks with that of the current task during the learning process of the $t$ -th task on $\\mathcal{C}_c$ , we reformulate the training problem as a MOO problem, which is formally described as follows:", + "bbox": [ + 511, + 751, + 905, + 811 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta_ {c}, \\theta_ {c} ^ {1: t}} \\ell \\left(f _ {\\theta_ {c}}, f _ {\\theta_ {c} ^ {t}}, D _ {t} ^ {c}\\right) \\cup \\ell \\left(f _ {\\theta_ {c}}, f _ {\\theta_ {c} ^ {k}}, M _ {k} ^ {c}\\right) \\mid \\forall k < t. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 534, + 821, + 906, + 845 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\ell(\\cdot)$ denotes the loss function, and $\\theta_c$ and $\\theta_c^t$ represent the backbone model for shared knowledge across multiple tasks and the task-specific classifier for $\\mathcal{C}_c$ , respectively.", + "bbox": [ + 511, + 854, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3844", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the parameters of the new and old task classifiers are independent and can be updated directly using their respective gradients $\\theta_c^k\\gets \\theta_c^k -\\alpha \\nabla \\ell (f_{\\theta_c},f_{\\theta_c^k})$ with step size $\\alpha$ our focus is on updating the shared backbone model, which is updated using the gradients $g_{k}^{c} = \\nabla \\ell (f_{\\theta_{c}},f_{\\theta_{c}^{k}})$ of all $t$ tasks. Let $\\beta$ be the step size for optimization. The problem of the backbone model update can be formulated as follows:", + "bbox": [ + 89, + 90, + 483, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {c} \\leftarrow \\theta_ {c} - \\beta d _ {c} ^ {*}, d _ {c} ^ {*} = \\arg \\min _ {d _ {c}} D \\left(g _ {k} ^ {c}, d _ {c}\\right) \\mid \\forall k \\leq t, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 208, + 483, + 233 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $d_c^*$ represents the optimal gradient, which approximates the gradient of the new task while avoiding conflicts with the gradient of the old task. In Eq. (6), we define it as a min-distance multi-objective problem. Therefore, the critical issue lies in determining the method used to compute the differences between gradients, ensuring that the solution for $d_c^*$ is as close as possible to all gradients.", + "bbox": [ + 89, + 243, + 483, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Symmetric distance metrics, such as Euclidean distance and cosine distance, are widely applied in deep neural networks. These metrics share a fundamental property: the forward impact (from $x$ to $y$ ) is equal to the backward impact (from $y$ to $x$ ). However, local model updates are highly dependent on the magnitude and direction of gradients, both of which are inherently asymmetric. In brief, given an old task gradient $g_{old}$ , trained on $M_c^k$ , and a new task gradient $g_{new}$ , trained on $D_t^c$ , their mutual influences on each other's tasks are significantly different. In previous studies [2, 42, 49], the condition $\\langle g_{old}, g_{new} \\rangle \\geq 0$ has been employed to signify the absence of conflicts between two tasks. However, given that the old task has been integrated earlier, $\\| g_{old} \\|$ is typically smaller while $\\| g_{new} \\|$ tends to be larger. In cases where $\\| g_{old} \\| \\ll \\| g_{new} \\|$ , even if $\\langle g_{old}, g_{new} \\rangle \\geq 0$ holds true, the updates pertaining to the new task may have a substantial impact on the performance of the old task.", + "bbox": [ + 89, + 349, + 483, + 608 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Definition 1 (Asymmetric Gradient Distance). Given two gradients $g_{A}$ and $g_{B}$ , the asymmetric gradient distance is defined as", + "bbox": [ + 89, + 617, + 483, + 662 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nD \\left(g _ {A}, g _ {B}\\right) = \\frac {\\left\\| g _ {A} - g _ {B} \\right\\|}{\\left\\| g _ {B} \\right\\| + \\left\\| g _ {A} - g _ {B} \\right\\|}.\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 672, + 403, + 708 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this paper, an asymmetric distance calculation method is used to explore the measure of difference between the new task and the old task. Definition 1 considers both the gradient direction and magnitude. Notably, $D(g_{A}, g_{B}) = 0$ when $g_{A} = g_{B} = 0$ .", + "bbox": [ + 89, + 719, + 483, + 795 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, solving Eq. (6) directly for the parallel task flow on clients presents significant challenges. This difficulty arises from the fact that both the decision space of this problem and the model parameters are identically high-dimensional. Consequently, following [28, 35], we employ linear scalarization to determine a weight variable $w^{*}$ . Specifically, when client $C_c$ is training on the $t$ -th task, the", + "bbox": [ + 89, + 795, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 FedAGC", + "text_level": 1, + "bbox": [ + 514, + 90, + 666, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input: $N$ clients with their private parallel task sequence $\\mathcal{T}^{1:N}$ and local datasets $\\{D_t^{1:N}\\}_{t=1}^T$ , global model parameter $\\theta_g$ , communication rounds $R$ , memory space $M_{1:c} \\gets \\{\\}$", + "bbox": [ + 511, + 108, + 906, + 157 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Output: Multiple global parameters $\\theta_{g}^{i:n}$ , local parameter $\\theta_{l,1:N}$ and classification head $\\theta_{l,1:N}^{k}$", + "bbox": [ + 511, + 156, + 905, + 189 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Initialization", + "2: while task number $t \\leq T$ do", + "3: for each round $r = 1,2,\\ldots ,R$ do", + "4: Server broadcasts $\\theta_g^i$ to each group of clients;", + "5: for each client $c = 1$ to $N$ in parallel do", + "6: Minimizing local CL problems by Eq. (1);", + "7: $g^{c} = \\nabla \\ell (f_{\\theta_{c}},f_{\\theta^{t}},D_{t}^{c});$", + "8: $g_{k}^{c} = \\nabla \\ell (f_{\\theta_{c}},f_{\\theta_{k}^{k}},M_{k}^{c})$ for all $k < t$", + "9: $d_c^* \\gets \\mathsf{ACGU}(g^c, g_1^c, \\ldots, g_{t-1}^k)$ ;", + "10: $\\theta_{l,c}^{t}\\gets \\theta_{l,c}^{t} - \\alpha d_{c}^{*};$", + "11: end for", + "12: Send the $G^{c} = -\\triangle \\theta_{l,c}^{t} / \\eta$ to the central server;", + "13: $\\{S_1,S_2,\\ldots ,S_n\\} \\gets \\mathrm{GAGU}(G^1,G^2,\\ldots ,G^N);$", + "14: $\\theta_{g}^{i}\\gets \\theta_{g}^{i} - \\frac{1}{|S_{i}|}\\sum_{j\\in S_{i}}G^{j}$ for $i\\in \\{1,2,\\ldots ,n\\}$", + "15: end for", + "16: Each client samples $M_t^c$ from $D_t^c$ by Eqs. (3) and (4);", + "17: end while" + ], + "bbox": [ + 519, + 190, + 903, + 449 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "gradient update $g_{t}^{c}$ of the new task and $g_{1:t-1}^{c}$ from the old tasks and computed through", + "bbox": [ + 511, + 476, + 906, + 506 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw ^ {*} = \\arg \\min _ {w} \\left\\{\\frac {1}{t} \\sum_ {i} ^ {t} D \\left(\\sum_ {j \\neq i} ^ {t} w _ {j} g _ {j} ^ {c}, g _ {i} ^ {c}\\right) \\right\\}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 512, + 905, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\sum_{j=1}^{t} w_{j} = 1$ , $\\forall w_{j} \\geq 0$ , and $D(\\sum_{j \\neq i}^{t} w_{j} g_{j}^{c}, g_{i}^{c})$ represents the divergence between task $i$ and the combination of all other tasks. By minimizing this divergence, we achieve the suppression of the influence from the combination on task $i$ . Note that this optimization process does not modify individual gradient updates.", + "bbox": [ + 511, + 568, + 905, + 660 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To prevent training failure on new tasks, we introduce the Multiple Gradient Descent Algorithm (MGDA) [4] to minimize the gradient norm, thereby achieving a balanced solution. It is defined as", + "bbox": [ + 511, + 660, + 905, + 719 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw _ {b} ^ {*} = \\underset {w} {\\arg \\min } \\left\\| \\sum_ {i} ^ {t} w _ {i} g _ {i} ^ {c} \\right\\| ^ {2}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 724, + 905, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When the objective function value of MGDA is zero, the obtained solution satisfies the Karush-Kuhn-Tucker conditions. Otherwise, the solution provides a Pareto descent direction that does not harm any task.", + "bbox": [ + 511, + 773, + 905, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In summary, our proposed Asymmetric Correction Gradient Updating (ACGU) algorithm determines $d_c^*$ by", + "bbox": [ + 511, + 835, + 903, + 866 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nd _ {c} ^ {*} = \\sum_ {i} W _ {i} g _ {i} ^ {c}, \\quad \\text {w h e r e} W = \\gamma w ^ {*} + (1 - \\gamma) w _ {b} ^ {*}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 872, + 905, + 904 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3845", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/2d478b58aaf089f4cbbd11d634db7052a0618d900114a8414df24e262e1c0894.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
AlgorithmAttributes FCL Per GCBTask ID under Non-IID Setting (Across 8 Different Datasets)\\( A_{avg} \\)(%)
\\( A_1 \\)(%)\\( A_2 \\)(%)\\( A_3 \\)(%)\\( A_4 \\)(%)\\( A_5 \\)(%)
FedAvg× × ×54.73 ± 1.4856.06 ± 3.1663.99 ± 3.4357.90 ± 2.9678.49 ± 2.1861.83 ± 0.46
FedProx× × ×58.25 ± 1.6056.65 ± 0.5066.15 ± 2.3361.72 ± 2.2182.24 ± 1.0265.00 ± 0.62
Ditto× ✓ ×58.92 ± 1.4059.13 ± 1.6370.58 ± 3.0763.00 ± 2.7983.97 ± 0.9567.12 ± 0.85
FedCurv✓ × ✓60.18 ± 3.4354.88 ± 1.9968.05 ± 3.7859.25 ± 2.3782.19 ± 0.5864.91 ± 0.56
FedAGEM✓ × ✓69.62 ± 0.8566.91 ± 1.0975.06 ± 1.6766.73 ± 2.4781.73 ± 1.1672.01 ± 0.81
FedER✓ × ×73.34 ± 0.4568.77 ± 1.8980.18 ± 1.3970.19 ± 1.3079.43 ± 1.3274.38 ± 0.27
FedMes✓ ✓ ✓70.84 ± 0.7967.22 ± 1.0577.25 ± 0.8967.18 ± 1.5285.21 ± 0.3673.54 ± 0.55
Ours-w/o-ACGU× ✓ ×58.62 ± 0.3658.06 ± 0.1665.62 ± 0.1963.29 ± 1.8682.46 ± 1.6865.61 ± 0.80
Ours-w/o-GAGU✓ × ✓76.19 ± 0.5271.59 ± 1.1282.24 ± 1.4272.89 ± 2.0481.82 ± 2.3676.95 ± 1.38
Ours-w/o-CBRSS✓ ✓ ✓75.40 ± 0.9072.39 ± 0.5582.88 ± 0.9275.42 ± 0.5486.68 ± 0.6278.56 ± 0.57
FedAGC (Ours)✓ ✓ ✓76.21 ± 0.3672.56 ± 0.4884.90 ± 0.8577.60 ± 0.5787.28 ± 0.1879.71 ± 0.23
", + "bbox": [ + 93, + 88, + 906, + 289 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Performance (avg ± std) of local models for each task on 10 clients under Non-IID setting. \"w/o-Module\" indicates performance without the corresponding module. It also includes Per (Personalized) and GCB (Gradient Correction-Based). For the proposed FedAGC, we used $\\lambda = 0.01$ and $\\gamma = 0.5$ in Eqs. (1) and (9).", + "bbox": [ + 89, + 299, + 906, + 340 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Group-Level Personalization", + "text_level": 1, + "bbox": [ + 89, + 348, + 346, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To integrate heterogeneous knowledge from different clients, some studies [39, 45] adopt personalization strategies at the client level, where each client continuously adjusts the received global model to better fit its local tasks. However, this approach struggles to meet the needs of all clients in large-scale parallel task scenarios. In contrast, we consider a group-level personalization strategy, enabling the broadcast of global models tailored to meet the specific requirements of each client.", + "bbox": [ + 88, + 369, + 482, + 505 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this work, we propose Group Aggregation Gradient Updating (GAGU) for client-uploaded gradients. Specifically, during the $r$ -th communication round, $N$ clients upload their respective averaged gradient updates $\\{G^1,G^2,\\dots ,G^N\\}$ obtained after $e$ local training iterations, where $G^{c} = -\\triangle \\theta_{l,c}^{t} / \\alpha$ . We compute a similarity matrix $S$ to measure the similarity between pairs of averaged gradient updates. The elements of $S$ are given by", + "bbox": [ + 89, + 506, + 483, + 628 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\ns _ {i, j} = \\frac {G ^ {i} G ^ {j}}{\\| G ^ {i} \\| \\| G ^ {j} \\|.} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 637, + 480, + 671 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We construct a graph $P = \\{G, E\\}$ to represent $A$ , where $G$ denotes the set of nodes in the graph, and each node corresponds to the $G^c$ . The set of edges $E = \\{s_{i,j}\\} \\forall i \\leq j \\leq N$ represents the similarity between $G^i$ and $G^j$ . The objective of GAGU is to reorder $P$ such that similar average gradient updates are grouped together, while dissimilar ones are isolated. To achieve this, we first define the similarity between any two groups $S_A$ and $S_B$ ( $S_A \\cap S_B = \\emptyset$ , $\\forall A \\neq B$ ):", + "bbox": [ + 89, + 680, + 483, + 801 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nh \\left(S _ {A}, S _ {B}\\right) = \\sum_ {i \\in S _ {A}, j \\in S _ {B}} s _ {i, j}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 811, + 482, + 844 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "GAGU aims to maximize intra-group similarity while minimizing inter-group similarity, resulting in an optimal graph partition $a^{*} = \\{S_{1}, S_{2}, \\ldots, S_{n}\\}$ , where $n$ represents the", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "number of groups in $a^*$ . Inspired by spectrum-clustering [13] and Ncut [36], We reformulate this optimization problem into the following form:", + "bbox": [ + 511, + 349, + 906, + 393 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {a \\in A} \\sum_ {S, S ^ {\\prime} \\in a} h \\left(S, S ^ {\\prime}\\right) - h \\left(S ^ {\\prime}, S ^ {\\prime}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 401, + 906, + 434 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $A$ is all possible grouping results. Additionally, we introduce the modularity $Q$ to evaluate the quality of the current partition during the process of determining the optimal number of groups and partitioning results. This metric measures the distribution of connectivity within and between groups.", + "bbox": [ + 511, + 441, + 905, + 532 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nQ = \\frac {1}{2 W} \\sum_ {i, j} \\left[ s _ {i, j} - \\frac {d _ {i} d _ {j}}{2 W} \\right] \\delta \\left(S _ {i}, S _ {j}\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 537, + 906, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $W = \\sum_{i,j} s_{i,j}$ represents the total edge weight in the current graph, $d_i = \\sum_j s_{i,j}$ denotes the degree of node $i$ , and $\\delta(S_i, S_j)$ is the Kronecker delta function.", + "bbox": [ + 511, + 584, + 905, + 631 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 642, + 648, + 659 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 667, + 705, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Benchmarks. We validate FedAGC in large-scale multitask parallel FCL scenarios. The experimental setup follows the Task-Incremental Learning paradigm and primarily includes two different benchmarks: (1) IID Setting (Single CIFAR-100): The 100 classes of CIFAR-100 [16] are divided into 20 tasks, each containing 5 classes. These tasks are randomly assigned to clients, ensuring that each client does not train on the same task multiple times. (2) Non-IID Setting (Across 8 Different Datasets): We partition and distribute eight widely used datasets in the FCL domain, such as CIFAR-10/100 [16], MNIST [18], SVHN [30], FashionMNIST [41], Not-MNIST [1], FaceScrub [32], and TrafficSigns [31], in the same manner as (1). In both benchmarks, each client is assigned a unique task sequence.", + "bbox": [ + 511, + 688, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3846", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/817087d984ad0d80005ca64d6ef802a09904874b1459d3fd45954f26e0390054.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
AlgorithmAttributes FCL Per GCBTask ID under IID Setting (Single CIFAR-100)\\( A_{avg} \\)(%)
\\( A_1 \\)(%)\\( A_2 \\)(%)\\( A_3 \\)(%)\\( A_4 \\)(%)\\( A_5 \\)(%)
FedAvg× × ×54.32 ± 0.9460.53 ± 2.0861.89 ± 1.5665.63 ± 1.6070.77 ± 0.8462.63 ± 0.55
FedProx× × ×53.96 ± 0.9861.53 ± 2.2362.90 ± 1.9465.70 ± 1.1370.78 ± 0.9162.97 ± 0.93
Ditto× ✓ ×55.48 ± 0.9262.81 ± 1.9663.12 ± 1.5866.59 ± 1.8671.01 ± 0.9363.67 ± 0.72
FedCurv✓ × ✓54.36 ± 0.9560.77 ± 0.8663.43 ± 0.9166.46 ± 0.9771.52 ± 0.7563.31 ± 0.16
FedAGEM✓ × ✓60.95 ± 0.9067.15 ± 1.2366.39 ± 0.6967.81 ± 0.9071.28 ± 0.6666.72 ± 0.23
FedER✓ × ×63.05 ± 0.7667.83 ± 0.9166.11 ± 0.3265.39 ± 0.4468.84 ± 1.0166.44 ± 0.42
FedMes✓ ✓ ✓61.18 ± 0.3467.68 ± 0.1167.11 ± 0.4967.42 ± 0.1572.12 ± 0.9067.13 ± 0.16
Ours-w/o-ACGU× ✓ ×55.45 ± 0.2761.95 ± 0.8763.32 ± 1.7265.51 ± 0.7371.44 ± 0.6063.53 ± 0.31
Ours-w/o-GAGU✓ × ✓63.92 ± 0.5566.73 ± 0.9367.22 ± 0.6166.48 ± 0.6568.53 ± 0.2466.58 ± 0.41
Ours-w/o-CBRSS✓ ✓ ✓66.14 ± 0.3769.05 ± 0.8069.69 ± 0.4969.56 ± 2.2074.12 ± 0.0969.75 ± 0.78
FedAGC (Ours)✓ ✓ ✓66.38 ± 0.4869.48 ± 0.7870.38 ± 1.3270.58 ± 1.8573.98 ± 0.4170.16 ± 0.77
", + "bbox": [ + 94, + 88, + 906, + 289 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Performance (avg ± std) of local models for each task on 10 clients under IID setting.", + "bbox": [ + 215, + 299, + 777, + 313 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/065edf3297a1e284be73cca5f916af60ca88546822aaf1d5946914a4ef13402a.jpg", + "image_caption": [ + "(a) $F_{t}$ under Non-IID Setting", + "Figure 3. Comparison of average forgetting rates." + ], + "image_footnote": [], + "bbox": [ + 94, + 319, + 285, + 465 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4b5ee6f8a9295d9655513e46cba23c8834e4f81d409c631b3361571035ae705c.jpg", + "image_caption": [ + "(b) $F_{t}$ under IID Setting" + ], + "image_footnote": [], + "bbox": [ + 289, + 319, + 480, + 465 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. We compare FedAGC with state-of-the-art (SOTA) methods, including the fundamental FedAvg [29] and its proximal variant FedProx [19], along with the personalized FL method Ditto [20]. To verify the superiority of FedAGC, we also compare it with gradient correction-based FCL methods, including FedCurv [38], FedAGEM [2] and the latest FedMes [42]. Additionally, we consider the rehearsal-based method FedER [3].", + "bbox": [ + 89, + 502, + 483, + 622 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metrics. We employ two critical metrics to evaluate algorithm performance: Average Accuracy $(A_{avg})$ and Forgetting Rate $(F_t)$ . Let $a_{i,t}^c$ denote the test accuracy of the $i$ -th task learned by the $c$ -th client after sequential learning of $t$ subsequent tasks, where $t + i \\leq T$ . In this formulation, the initial accuracy of task $i$ (immediately after its own training) and the final accuracy after training on all tasks are defined as $a_{i,0}^c$ and $a_{i,T - i}^c$ , respectively. The forgetting rate for the $i$ -th task after learning $t$ subsequent tasks is calculated as: $f_{i,t}^{c} = \\frac{a_{i,0}^{c} - a_{i,t}^{c}}{a_{i,0}^{c}}$ . To aggregate results, we compute client-wise averages: $A_i = \\frac{1}{N}\\sum_{c=1}^N a_{i,T-i}^c$ and $f_{i,t} = \\frac{1}{N}\\sum_{c=1}^N f_{i,t}^c$ . Then, $A_{avg}$ and $F_t$ are defined as:", + "bbox": [ + 89, + 625, + 483, + 820 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nA _ {\\text {a v g}} = \\frac {1}{T} \\sum_ {i = 1} ^ {T} A _ {i}, \\quad F _ {t} = \\frac {1}{T - t} \\sum_ {i = 1} ^ {T - t} f _ {i, t}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 832, + 482, + 873 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation details. Following the experimental setup", + "bbox": [ + 89, + 885, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ed150f3019ad6825d5ac902e367243c47790fbf8c4a456df9789eb7fb1dfa2b9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 320, + 700, + 411 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d11c41af08c1f352b50d3273bdb02d13ba31314a436c9be0da172a1e46db85a3.jpg", + "image_caption": [ + "(a) Impact of $\\gamma$ on FedAGC" + ], + "image_footnote": [], + "bbox": [ + 517, + 426, + 705, + 522 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/86e4413525e67630fc5d1bf946e0fd9c18c4829f2f90681d0e32be8ce15f94fe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 320, + 901, + 411 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/aaa4303fe61e51880f01edfafa2d12c9ac4e94b2a13e4fc9863b12a8349fca2e.jpg", + "image_caption": [ + "(b) Impact of $\\lambda$ on FedAGC" + ], + "image_footnote": [], + "bbox": [ + 712, + 426, + 901, + 522 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/225cb13e94af64123351d55667487139ec065af95c06600e19b7d0bd165bdfdb.jpg", + "image_caption": [ + "(c) Non-IID Setting, $\\lambda = 0.01$", + "(e) IID Setting, $\\lambda = 0.01$", + "Figure 4. Impact of core hyperparameters on FedAGC." + ], + "image_footnote": [], + "bbox": [ + 517, + 540, + 705, + 635 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6a9898dc138632433d22d7fcc7f82a0fe26f29ba033401bff1c58d18abfc4297.jpg", + "image_caption": [ + "(d) Non-IID Setting, $\\gamma = 0.5$", + "(f) IID Setting, $\\gamma = 0.5$" + ], + "image_footnote": [], + "bbox": [ + 712, + 540, + 901, + 635 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of [47], we adopted a variant of the LeNet architecture as the backbone. For any newly introduced task, we initialized a new classifier. The Adam optimizer was used for all datasets, with an initial learning rate of $10^{-3}$ and a weight decay of $10^{-4}$ . The experimental setup consisted of 10 clients, each learning 5 sequential and non-overlapping tasks. The batch size was set to 100 per step, with $\\mu$ and $v$ set to 0.1 and 8, respectively. To ensure convergence, all experiments were conducted for 100 federated communication rounds. Each experiment was repeated five times, and the reported results were averaged accordingly.", + "bbox": [ + 511, + 672, + 906, + 840 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 511, + 848, + 651, + 862 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tabs. 1 and 2 record the average accuracy of 10 clients on each task under two data condition setups, along with the", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3847", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "final results considering all tasks. In Fig. 3, we compare the degree of Spatial-TemporalCF between FedAGC and various methods. It can be observed that FedProx and Ditto mitigate the impact of statistical heterogeneity using penalty terms, showing significant performance improvements over FedAvg in Non-IID settings. However, since they do not employ any CL strategies, their accuracy is lower compared to FCL methods such as FedAGEM, FedER, and FedMes, with higher forgetting levels. Notably, FedCurv's experimental results are suboptimal as it leverages curvature information from other clients to guide updates for the current client, which is prone to knowledge conflicts in cases of non-overlapping client data distributions. These FCL methods based on symmetric gradient correction focus solely on TemporalCF without considering the interference of knowledge across clients, leading to SpatialCF. In contrast, the proposed FedAGC outperforms other methods in both task-specific accuracy and overall performance across both settings, with a lower degree of forgetting. This demonstrates that memory buffers constructed through importance sampling, assisted by asymmetric gradient correction and group-level personalization, can significantly improve the accuracy of PFCL models.", + "bbox": [ + 89, + 90, + 483, + 439 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Performance Analysis of FedAGC", + "text_level": 1, + "bbox": [ + 89, + 452, + 387, + 468 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Studies. We conducted three different ablation studies under the same experimental setup, with the results recorded in Tabs. 1 and 2, as well as Fig. 3. We progressively removed components from FedAGC, including: (1) Omitting the ACGU module (FedAGC-w/o-ACGU), which means that only the gradients from local training are used to guide updates. Clearly, in this case, the CBRSS module also loses its significance. The performance degradation is significant, with a drop of $17.7\\%$ and $9.5\\%$ , as no strategy for mitigating TemporalCF is applied. (2) Removing the GAGU module (FedAGC-w/o-GAGU) and switching to the FedAvg average aggregation method. In the IID setting, FedAGC-w/o-GAGU's overall performance is overtaken by FedMes by $0.86\\%$ . (3) Disabling the CBRSS module (FedAGC-w/o-CBRSS) and replacing it with a random sampling method. In this case, performance decreases by $1.44\\%$ and $0.58\\%$ in the Non-IID and IID settings, respectively. However, since the core component for mitigating forgetting is retained, its accuracy on individual tasks and overall performance remains superior to other baselines. These results verify that our three proposed modules of FedAGC play a crucial role in alleviating Spatial-TemporalCF and achieving personalization, and can effectively be combined to enhance performance.", + "bbox": [ + 89, + 476, + 482, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of core hyperparameters. We evaluate FedAGC with different values of $\\gamma$ and $\\lambda$ . As shown in Fig. 4, setting $\\lambda$ too large or too small causes FedAGC to overly rely on either special or balanced solution, impairing its ability to", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8afb4c9a20a3e080b95b48912d0f19e7fb31dbbf3440484387aa96ea12bdf009.jpg", + "image_caption": [ + "(a) $A_{i}$ on the CIFAR-100 dataset, $N = 30$" + ], + "image_footnote": [], + "bbox": [ + 571, + 90, + 848, + 232 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cf5e28fab530917c16438c069921c6bfdc0f579e4ca5820b5f39f1ebaecc235c.jpg", + "image_caption": [ + "(b) $F_{t}$ on the CIFAR-100 dataset, $N = 30$", + "Figure 5. Performance on large-scale clients." + ], + "image_footnote": [], + "bbox": [ + 571, + 250, + 844, + 386 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "mitigate forgetting. Similarly, an excessively large $\\gamma$ biases the model toward the inversion of previous tasks, hindering the learning of current task, while a small $\\gamma$ has the opposite effect. Empirically, we find that the best performance of FedAGC is achieved when $\\gamma = 0.5$ and $\\lambda = 0.01$ .", + "bbox": [ + 511, + 422, + 906, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Performance on large-scale clients. To explore the scalability of FedAGC to larger client scenarios, we tripled the number of clients from the previous setting of 10 and continued comparing it with various SOTA methods. The results recorded in Fig. 5 indicate that FedAGC outperforms other baselines in accuracy on previous tasks and demonstrates greater stability in mitigating forgetting. This result highlights the superiority of FedAGC.", + "bbox": [ + 511, + 500, + 908, + 621 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 633, + 633, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To tackle the challenge of STCF in PFCL, this paper proposed FedAGC as a valid solution. After local training, FedAGC selects representative samples and stores them in a buffer for rehearsal to alleviate forgetting. To refine this process, we introduced an asymmetric gradient metric to quantify discrepancies between past and current tasks. Asymmetric gradient correction is then employed to adjust the gradient updates, jointly considering direction and magnitude for optimal adaptation. Moreover, FedAGC adopts group-level personalization by clustering clients based on the similarity of their uploaded averaged gradient updates, which enhances the fusion of valuable knowledge while isolating ineffective knowledge. It was extensively demonstrated in this paper that the effective integration of the proposed modules can enhance the model's resistance to STCF, enabling FedAGC to outperform SOTA approaches.", + "bbox": [ + 511, + 657, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3848", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledge", + "text_level": 1, + "bbox": [ + 91, + 90, + 209, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We want to thank the anonymous reviewers for their valuable suggestions and comments. This work was supported by the National Natural Science Foundation of China (No. 62276182), Tianjin University under the Emerging Direction Cultivation Project of Interdisciplinary Center (Intelligent Protection and Utilization of Digital Cultural Heritage), National Key Research and Development Program of China (No. 2023YFF0906204), Peng Cheng Lab Program (No. PCL2023A08), Tianjin Natural Science Foundation (Nos. 24JCYBJC01230, 24JCYBJC01460), and Tianjin Municipal Education Commission Research Plan (No. 2024ZX008).", + "bbox": [ + 89, + 114, + 485, + 297 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 309, + 187, + 325 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yaroslav Bulatov. Notmnist dataset. http://yaroslavvb.blogspot.it/2011/09/notmnist-dataset.html.6", + "[2] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a-gem. In Proceedings of the International Conference on Learning Representations, 2018. 2, 3, 4, 5, 7", + "[3] Arslan Chaudhry, Marcus Rohrbach, Mohamed Elhoseiny, Thalaiyasingam Ajanthan, Puneet K. Dokania, Philip H. S. Torr, and MarcÁurelio Ranzato. On tiny episodic memories in continual learning. In arXiv Preprint arXiv:1902.10486, 2019. 3, 7", + "[4] Désideri and Jean-Antoine. Multiple-gradient descent algorithm (mgda) for multiobjective optimization. Comptes Rendus Mathematique, 350(5-6):313-318, 2012. 5", + "[5] Prithviraj Dhar, Rajat Vikram Singh, Kuan-Chuan Peng, Ziyan Wu, and Rama Chellappa. Learning without memorizing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5138-5146, 2019. 3", + "[6] Jiahua Dong, Lixu Wang, Zhen Fang, Gan Sun, Shichao Xu, Xiao Wang, and Qi Zhu. Federated class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10164-10173, 2022. 1", + "[7] Alireza Fallah, Aryan Mokhtari, and Asuman Ozdaglar. Personalized federated learning with theoretical guarantees: A model-agnostic meta-learning approach. Advances in neural information processing systems, 33:3557–3568, 2020. 2", + "[8] Mehrdad Farajtabar, Navig Azizan, Alex Mott, and Ang Li. Orthogonal gradient descent for continual learning. In International Conference on Artificial Intelligence and Statistics, PMLR, pages 3762-3773, 2020. 3", + "[9] Bagwe G, Yuan X, Pan M, and et al. Fed-cprompt: Contrastive prompt for rehearsal-free federated continual learning. arXiv preprint arXiv:2307.04869, 2023. 3", + "[10] Yunhui Guo, Mingrui Liu, Tianbao Yang, and Tajana Rosing. Improved schemes for episodic memory-based lifelong learning. Advances in Neural Information Processing Systems, 33:1023-1035, 2020. 4" + ], + "bbox": [ + 93, + 335, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Piao H, Wu Y, Wu D, and et al. Federated continual learning via prompt-based dual knowledge transfer. In *Forty-First International Conference on Machine Learning (ICML)*. PMLR, 2024. 3", + "[12] Parisa Hamedi, Roozbeh Razavi-Far, and Ehsan Hallaji. Federated continual learning: Concepts, challenges, and solutions. Neurocomputing, 651:130844, 2025. 3", + "[13] Hongjie Jia, Shifei Ding, Xinzheng Xu, and Ru Nie. The latest research progress on spectral clustering. Neural Computing and Applications, 24:1477-1486, 2014. 6", + "[14] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank J.Reddi, Sebastian U.Stich, and Ananda Theertha Suresh. Scaffold: Stochastic controlled averaging for federated learning. In International conference on machine learning. PMLR, pages 5132-5143, 2020. 2", + "[15] Alexander Krawczyk and Alexander Gepperth. An analysis of best-practice strategies for replay and rehearsal in continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4196-4204, 2024. 3", + "[16] Alex Krizhevsky and Geoffrey Hinton. Learning Multiple Layers of Features from Tiny Images. 2009. 6", + "[17] Matthias De Lange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Aleš Leonardis, Gregory Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying forgetting in classification tasks. In IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 3366-3385, 2021. 1", + "[18] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 6", + "[19] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. In Proceedings of Machine Learning and Systems, pages 429-450, 2020. 2, 7", + "[20] Tian Li, Shengyuan Hu, Ahmad Beirami, and Virginia Smith. Ditto: Fair and robust federated learning through personalization. In International Conference on Machine Learning, PMLR, pages 6357-6368, 2021. 2, 7", + "[21] Junkang Liu, Yuanyuan Liu, Fanhua Shang, Hongying Liu, Jin Liu, and Wei Feng. Improving generalization in federated learning with highly heterogeneous data via momentum-based stochastic controlled weight averaging. In Forty-second International Conference on Machine Learning. 1", + "[22] Junkang Liu, Fanhua Shang, Yuanyuan Liu, Hongying Liu, Yuangang Li, and YunXiang Gong. Fedbcgd: Communication-efficient accelerated block coordinate gradient descent for federated learning. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 2955-2963, 2024. 1", + "[23] Quande Liu, Cheng Chen, Jing Qin, Qi Dou, and Pheng-Ann Heng. Feddg: Federated domain generalization on medical image segmentation via episodic learning in continuous frequency space. In CVPR, pages 1013–1023, 2021. 1", + "[24] Yang Liu, Anbu Huang, Yun Luo, He Huang, Youzhi Liu, Yuanyuan Chen, Lican Feng, Tianjian Chen, Han Yu, and" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3849", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Qiang Yang. Fedvision: An online visual object detection platform powered by federated learning. In AAAI, pages 13172-13179, 2020. 1", + "[25] David Lopez-Paz and MarcÁurelio Ranzato. Gradient episodic memory for continual learning. In Advances in neural information processing systems, 2017. 3", + "[26] Bing Luo, Wenli Xiao, Shiqiang Wang, Jianwei Huang, and Leandros Tassiulas. Tackling system and statistical heterogeneity for federated learning with adaptive client sampling. In IEEE INFOCOM 2022 - IEEE Conference on Computer Communications, pages 1739-1748, 2022. 2", + "[27] Fan Lyu, Shuai Wang, Wei Feng, Zihan Ye, Fuyuan Hu, and Song Wang. Multi-domain multi-task rehearsal for lifelong learning. Proceedings of the AAAI Conference on Artificial Intelligence, 35(10):8819-8827, 2021. 3, 4", + "[28] Fan Lyu, Qing Sun, Fanhua Shang, Liang Wan, and Wei Feng. Measuring asymmetric gradient discrepancy in parallel continual learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11411-11420, 2023. 5", + "[29] H. Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial Intelligence and Statistics, PMLR, pages 1273-1282, 2017. 1, 2, 7", + "[30] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading digits in natural images with unsupervised feature learning. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning, page 4, 2011. 6", + "[31] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading digits in natural images with unsupervised feature learning. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning, page 4, 2011. 6", + "[32] Hong-Wei Ng and Stefan Winkler. A data-driven approach to cleaning large face datasets. In Proceedings of the 2014 IEEE International Conference on Image Processing (ICIP), pages 343–347. IEEE, 2014. 6", + "[33] Li Qinbin, Diao Yiqun, Chen Quan, and He Bingsheng. Federated learning on non-iid data silos: An experimental study. In 2022 IEEE 38th International Conference on Data Engineering (ICDE), pages 965-978, 2022. 1", + "[34] Felix Sattler, Simon Wiedemann, Klaus-Robert Müller, and Wojciech Samek. Robust and communication-efficient federated learning from non-iid data. IEEE Transactions on Neural Networks and Learning Systems, 31(9):3400-3413, 2019. 2", + "[35] Ozan Sener and Vladlen Koltun. Multi-task learning as multi-objective optimization. In Advances in Neural Information Processing Systems, 2018. 5", + "[36] Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 22(8):888-905, 2000. 6", + "[37] Yong Shi, Yuanying Zhang, Yang Xiao, and Lingfeng Niu. Optimization strategies for client drift in federated learning: A review. Procedia Computer Science, 214:1168-1173, 2022. 2" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Neta Shoham, Tomer Avidor, Aviv Keren, Nadav Israe, Daniel Benditkis, Liron Mor-Yosef, and Itai Zeitak. Overcoming forgetting in federated learning on non-iid data. In arXiv Preprint arXiv:1910.07796, 2019. 1, 3, 7", + "[39] Hongda Wu and Ping Wang. Fast-convergent federated learning with adaptive weighting. IEEE Transactions on Cognitive Communications and Networking, 7(4):1078-1088, 2021. 6", + "[40] Yichen Wu, Hong Wang, Peilin Zhao, Yefeng Zheng, Ying Wei, and Long-Kai Huang. Mitigating catastrophic forgetting in online continual learning by modeling previous task interrelations via pareto optimization. In *Forty-first international conference on machine learning*, 2024.", + "[41] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: A novel image dataset for benchmarking machine learning algorithms. In arXiv Preprint arXiv:1708.07747, 2017. 6", + "[42] Jin Xie, Chenqing Zhu, and Songze Li. Fedmes: Personalized federated continual learning leveraging local memory. In arXiv Preprint arXiv:2404.12710, 2024. 2, 4, 5, 7", + "[43] He Y, Shen C, Wang X, and et al. Fppl: An efficient and non-iid robust federated continual learning framework. In 2024 IEEE International Conference on Big Data (BigData), pages 3692-3701. IEEE, 2024. 3", + "[44] Luopan Y, Han R, Zhang Q, and et al. Fedknow: Federated continual learning with signature task knowledge integration at edge. In 2023 IEEE 39th International Conference on Data Engineering (ICDE), pages 341-354. IEEE, 2023. 3", + "[45] Xin Yang, Hao Yu, Xin Gao, Hao Wang, Junbo Zhang, and Tianrui Li. Federated continual learning via knowledge fusion: A survey. In IEEE Transactions on Knowledge and Data Engineering, 2024. 1, 2, 3, 6", + "[46] Xin Yao and Lifeng Sun. Continual local training for better initialization of federated models. In 2020 IEEE International Conference on Image Processing (ICIP), pages 1736-1740, 2020. 1", + "[47] Jaehong Yoon, Wonyong Jeong, Giwoong Lee, Eunho Yang, and Sung Ju Hwang. Federated continual learning with weighted inter-client transfer. In International Conference on Machine Learning, pages 12073-12086, 2021. 1, 2, 7", + "[48] Hao Yu, Xin Yang, Xin Gao, Yan Kang, Hao Wang, Junbo Zhang, and Tianrui Li. Personalized federated continual learning via multi-granularity prompt. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 4023-4034, 2024. 2, 3", + "[49] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. In Advances in Neural Information Processing Systems, pages 5824-5836, 2020. 5" + ], + "bbox": [ + 516, + 92, + 903, + 782 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3850", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/FedAGC_ Federated Continual Learning with Asymmetric Gradient Correction/69b6bc67-ec8c-4e1a-b957-09f50cc21aab_model.json b/2025/FedAGC_ Federated Continual Learning with Asymmetric Gradient Correction/69b6bc67-ec8c-4e1a-b957-09f50cc21aab_model.json new file mode 100644 index 0000000000000000000000000000000000000000..84d97c26676a2aae6683b7f8f1d95647649316ee --- /dev/null +++ b/2025/FedAGC_ Federated Continual Learning with Asymmetric Gradient Correction/69b6bc67-ec8c-4e1a-b957-09f50cc21aab_model.json @@ -0,0 +1,2596 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.098, + 0.131, + 0.9, + 0.152 + ], + "angle": 0, + "content": "FedAGC: Federated Continual Learning with Asymmetric Gradient Correction" + }, + { + "type": "text", + "bbox": [ + 0.189, + 0.18, + 0.808, + 0.235 + ], + "angle": 0, + "content": "Chengchao Zhang\\(^{1}\\), Fanhua Shang\\(^{1\\dagger}\\), Hongyin Liu\\(^{2\\dagger}\\), Liang Wan\\(^{1,2}\\), Wei Feng\\(^{1}\\), \\(^{1}\\)College of Intelligence and Computing, Tianjin University, China \\(^{2}\\)Medical College, Tianjin University, China" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.237, + 0.76, + 0.253 + ], + "angle": 0, + "content": "{chengchaozhang, fhshang, hyliu2009, lwan, wfeng}@tju.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.319, + 0.485, + 0.652 + ], + "angle": 0, + "content": "Federated Continual Learning (FCL) has emerged as a prominent distributed learning paradigm and aims at addressing model learning challenges in both federated and continual learning settings. Efficient personalization in FCL remains a major challenge, as it must handle not only conflicts between old and new knowledge within parallel task streams but also heterogeneous knowledge conflicts from different clients. Recent approaches attempt to mitigate these issues through gradient correction. However, they often overlook the combined impact of gradient magnitude and direction, leading to unsatisfactory gradient solutions. To address these issues, we propose a novel federated continual learning method (called FedAGC) with asymmetric gradient correction, which performs memory rehearsal using representative samples selected via a centroid-based approach from historical tasks. By formulating the problem as a multi-objective optimization problem, FedAGC derives more effective gradients while incorporating group-level personalization to facilitate useful knowledge integration and irrelevant knowledge isolation, effectively mitigating both temporal and spatial catastrophic forgetting. Extensive experiments confirm the effectiveness of FedAGC." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.683, + 0.222, + 0.699 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.709, + 0.484, + 0.875 + ], + "angle": 0, + "content": "Federated Learning (FL) [21, 22, 29], as a novel distributed machine learning paradigm, aims to train a global model by leveraging independent and diverse data across clients in a privacy-preserving and collaborative manner. FL has been widely applied in various domains, such as medical image analysis [23] and object detection [24]. It has been proven to effectively address the issue of data silos [33] without compromising client privacy. Most existing FL methods are limited to single-task learning, assuming that client data remains static over time. However, in real-world scenarios, data is continuously evolving, forming sequential tasks with" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.285, + 0.925, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.518, + 0.415, + 0.66, + 0.425 + ], + "angle": 0, + "content": "Local update of client 1" + }, + { + "type": "image_footnote", + "bbox": [ + 0.518, + 0.425, + 0.66, + 0.438 + ], + "angle": 0, + "content": "Local update of client \\(N\\)" + }, + { + "type": "image_footnote", + "bbox": [ + 0.518, + 0.439, + 0.649, + 0.451 + ], + "angle": 0, + "content": "Aggregation of server" + }, + { + "type": "image_footnote", + "bbox": [ + 0.522, + 0.452, + 0.624, + 0.464 + ], + "angle": 0, + "content": "\\(\\star\\) Global optimum" + }, + { + "type": "image_footnote", + "bbox": [ + 0.523, + 0.465, + 0.64, + 0.477 + ], + "angle": 0, + "content": "Server global model" + }, + { + "type": "image_footnote", + "bbox": [ + 0.523, + 0.479, + 0.631, + 0.489 + ], + "angle": 0, + "content": "Client local model" + }, + { + "type": "image_footnote", + "bbox": [ + 0.523, + 0.49, + 0.605, + 0.502 + ], + "angle": 0, + "content": "- Initial model" + }, + { + "type": "image_footnote", + "bbox": [ + 0.518, + 0.503, + 0.647, + 0.516 + ], + "angle": 0, + "content": "\\(\\iff\\) Parameter divergence" + }, + { + "type": "image_footnote", + "bbox": [ + 0.518, + 0.517, + 0.599, + 0.528 + ], + "angle": 0, + "content": "Client drift" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.285, + 0.925, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.631, + 0.53, + 0.808, + 0.543 + ], + "angle": 0, + "content": "(b) Spatial catastrophic forgetting" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.555, + 0.907, + 0.598 + ], + "angle": 0, + "content": "Figure 1. Spatial-Temporal Catastrophic Forgetting. (a) Direct access to new tasks leads to a significant performance decline on old tasks. (b) Non-IID data exacerbates the weight disparity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.907, + 0.688 + ], + "angle": 0, + "content": "unpredictable future changes. As shown in Fig. 1a, severe temporal catastrophic forgetting (TemporalCF) [38, 45] occurs when the local model on each client learns a newly introduced task, leading to performance degradation on previously learned tasks as new knowledge is acquired." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.839 + ], + "angle": 0, + "content": "Improving the generalizability of FL in multi-task evolving scenarios has emerged as a critical and challenging research topic. Recently, Continual Learning (CL) [17] has made significant advances in addressing TemporalCF, enabling effective mitigation of this problem. Based on such characteristics, the concept of Federated Continual Learning (FCL) [6, 46, 47] came into being, which integrates CL methods into the FL framework. In the FCL setting, FL can be regarded as a distributed instance of conventional CL, while CL methods assist in addressing FL challenges." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.901 + ], + "angle": 0, + "content": "FCL breaks the static limitation of traditional FL, where each client is confined to learning a single task sequence. However, integrating CL into the FL framework poses substantial challenges. Besides TemporalCF, FCL also" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.236, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3841" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.273 + ], + "angle": 0, + "content": "faces spatial catastrophic forgetting (SpatialCF). Specifically, the private data of different clients in FCL may follow drastically different distributions. This situation, where data across clients is Non-Independent and Identically Distributed (Non-IID), which is referred to as statistical heterogeneity [26]. During server-side aggregation, where gradients from multiple clients are combined to initialize the next round of local CL training, statistical heterogeneity can lead to issues such as parameter divergence and client drift [14, 37], as illustrated in Fig. 1b. Consequently, the aggregated global model exhibits reduced generalizability, further exacerbating temporal catastrophic forgetting." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.276, + 0.483, + 0.547 + ], + "angle": 0, + "content": "To address the aforementioned challenges, recent studies [42, 47, 48] have explored the problem of Personalized Federated Continual Learning (PFCL). Unlike traditional FCL that aim to obtain a universal global model, PFCL leverages efficient server-side aggregation to facilitate personalized training. Existing PFCL methods mainly focus on optimizing local models for clients, ensuring strong performance on both new and previous tasks in a parallel, independent, and heterogeneous task stream across clients. The most relevant work to ours is FedMes [42], which builds on the gradient correction technique of A-GEM [2] and enhances local models through KNN-based inference. However, existing methods typically rely solely on gradient direction to determine similarity, ignoring the impact of magnitude, leading to suboptimal gradients. In contrast, we propose a new Asymmetric Gradient Correction (AGC) technique, which adaptively balances the combined impact of direction and magnitude, enabling a smoother optimization transition." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.549, + 0.483, + 0.821 + ], + "angle": 0, + "content": "We formulate the PFCL problem as a Multi-Objective Optimization (MOO) problem to obtain the optimal gradient, where each objective minimizes the distance to a target gradient. Specifically, to mitigate Spatial-Temporal Catastrophic Forgetting (STCF), each client extracts a small set of representative samples after completing local training for each task and employs them for rehearsal in subsequent learning. Subsequently, we correct the gradients by considering both their magnitudes and directions, ultimately obtaining an optimal gradient that does not damage any of old or new knowledge at each update from a set of class-level gradients. Additionally, empirical evidence suggests that a single global model has limited generalization across all clients [34]. Therefore, we introduce a group-level personalization strategy, where clustering is performed based on the similarity of task gradients uploaded by participating clients. These clusters are then divided into groups, enabling joint training of multiple global models." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.824, + 0.472, + 0.838 + ], + "angle": 0, + "content": "Our main contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.84, + 0.483, + 0.885 + ], + "angle": 0, + "content": "- To the best of our knowledge, this is the first attempt to unify both spatial and temporal catastrophic forgetting in PFCL through a multi-objective optimization framework." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.887, + 0.483, + 0.902 + ], + "angle": 0, + "content": "- We propose a novel asymmetric gradient correction algo" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "rithm that performs memory rehearsal using representative samples selected via a centroid-based approach from historical tasks, transforming multi-task updates into an optimal gradient search problem that minimizes inter-task discrepancies and mitigates TemporalCF." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.168, + 0.905, + 0.242 + ], + "angle": 0, + "content": "- We propose a new group-level personalization approach that clusters clients by gradient similarity and aggregates multiple global models, isolating irrelevant knowledge interference while fostering the positive fusion of related knowledge to mitigate SpatialCF." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.243, + 0.905, + 0.288 + ], + "angle": 0, + "content": "- Our method outperforms all the baselines in terms of accuracy and forgetting rate, demonstrating its efficacy in addressing the PFCL problem." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.168, + 0.905, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.302, + 0.808, + 0.318 + ], + "angle": 0, + "content": "2. Background and Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.327, + 0.907, + 0.854 + ], + "angle": 0, + "content": "FL Algorithms for Statistical Heterogeneity. The client-server architecture-based FL paradigm is currently the most widely adopted framework. This distributed learning model involves a central server for aggregating network model parameters or gradients and \\(N\\) clients \\(\\{F_1, F_2, \\ldots, F_N\\}\\) for participating in the training of local private data, where each client possesses a local private dataset \\(\\{D_1, D_2, \\ldots, D_N\\}\\). For the \\(i\\)-th client \\(F_i\\), its local dataset includes \\(n_i\\) samples, denoted as \\(\\{D_i = \\{(X_k^i, Y_k^i)\\}_{k=1}^{n_i}\\). Different clients may exhibit distinct features and labels, such that \\(X_i \\cap X_j = \\emptyset\\) for \\(\\forall i, j \\in [1, N]\\), and \\(Y_k \\neq \\emptyset\\), \\(Y_i = \\emptyset\\) for \\(\\forall i \\in [1, N] \\setminus \\{k\\}\\). The objective of FL is to collaboratively and privately train a global statistical model across multiple distributed clients under the supervision of a central server. The key formulation controlling FL in the classical FedAvg [29] is as follows: \\(\\min_{w \\in \\mathbb{R}^d} f(w) = \\frac{1}{m} \\sum_{i=1}^{m} f_i(w)\\). where \\(f_i(w) = \\mathbb{E}_{(x_i, y_i) \\sim \\mathcal{P}_i}[\\mathcal{L}_i(w, x_i, y_i)]\\) represents the local loss function of the \\(i\\)-th client based on its local data, and \\(\\mathcal{P}_i\\) denotes the local data distribution of the \\(i\\)-th client. To address the challenge of statistical heterogeneity and mitigate client drift caused by Non-IID data distributions, FedProx [19] extends FedAvg by introducing a proximal term that effectively constrains local updates, thereby enhancing the overall stability of model convergence. Another representative approach, SCAFFOLD [14], employs control variates to alleviate client drift. However, under extreme Non-IID conditions, these methods often struggle to produce a highly generalizable global model. Consequently, Personalized FL [20, 45] shifts its focus to providing specialized local models for each client, enabling personalized services. PerFedAvg [7] combines meta-learning techniques during a pre-training phase to obtain a well-initialized model. While these approaches have demonstrated strong performance in FL, they are primarily designed to address SpatialCF and may not be sufficient for more complex scenarios." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Gradient Correction-Based CL Algorithms. In contrast to traditional FL designed for single-task scenarios, in a standard continual learning setting, there are a sequence" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3842" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.138, + 0.087, + 0.861, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.402, + 0.908, + 0.462 + ], + "angle": 0, + "content": "Figure 2. The workflow of the proposed FedAGC method. Our FedAGC consists of the following three modules: On the client side, the Centroid-based Representative Sample Selection (CBRSS) module rehearses the gradients of previous tasks to assist the Asymmetric Correction Gradient Updating (ACGU) module in generating optimal gradients. On the server side, the Group Aggregation Gradient Updating (GAGU) module clusters clients based on gradient similarity and aggregates multiple global models." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.472, + 0.485, + 0.884 + ], + "angle": 0, + "content": "of tasks \\(\\mathcal{T} = \\{\\mathcal{T}_1,\\mathcal{T}_2,\\dots ,\\mathcal{T}_T\\}\\), where \\(T\\) is the number of tasks. For the \\(t\\)-th task \\(\\mathcal{T}_t\\), its dataset is composed of \\(n_t\\) pairs of data and labels: \\(\\{x_i^t,y_i^t\\}_{i = 1}^{n_t}\\). Typically, \\(\\mathcal{T}\\) refers to a task stream where the arrival order of new tasks is unknown. As a result, the model can only access \\(\\mathcal{T}_t\\) during the training of the \\(t\\)-th task, and access to it is no longer available afterward. Therefore, for \\(\\mathcal{T}_t\\), given the set of model parameters \\(w^{(t)}\\) learned so far, the learning objective at task \\(t\\) is: \\(\\min_{w^{(t)}}\\mathcal{L}(w^{(t)};w^{(t - 1)},\\mathcal{T}^t)\\). In recent years, numerous CL algorithms have adopted gradient correction strategies to mitigate TemporalCF, typically involving two key steps: Memory Rehearsal and Optimal Gradient Computation [15, 27]. A representative example is GEM [25], which retrieves gradients from previously stored memories and integrates them with the current task's gradient to obtain an update direction suitable for all tasks. As one variant, A-GEM [2], consolidates multiple memories to enhance computational efficiency. Meanwhile, ER [3] combines data from episodic memory with the current task during training. Although ER is simple and effective, its performance depends heavily on the quality of the stored memories. Another related approach is parameter regularization [5, 8], which introduces additional regularization terms to consolidate knowledge from previous tasks and reduce interference between new and old task parameters. Among these methods, EWC [38] employs the Fisher information matrix to quantify the importance of parameters for previously learned tasks." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "While FL and CL have achieved remarkable success in" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.473, + 0.909, + 0.671 + ], + "angle": 0, + "content": "their respective domains, effectively mitigating the STCF problem in PFCL remains a challenging endeavor [12, 45]. FedCurv [38] incorporates EWC to handle statistical heterogeneity; however, it incurs high computational overhead and lacks a dedicated mechanism for addressing TemporalCF. FedKNOW [44] deploys a gradient integrator at each client to reduce negative knowledge transfer by effectively combining signature tasks identified from past and current tasks of other clients through the global model. Recent advances in vision foundation models have driven an alternative prompt-based approach [9, 11, 43, 48] that mitigates STCF by adapting to new tasks through training only a small subset of prompts and classifier parameters." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.678, + 0.725, + 0.696 + ], + "angle": 0, + "content": "3. The Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.87 + ], + "angle": 0, + "content": "In this section, we elaborate on the proposed FedAGC. For the first time, FedAGC formulates the PFCL problem (see Sec. 3.1) as a MOO problem. On the client-side, after training each task, we select representative samples to construct a memory cache (see Sec. 3.2) and use them to perform asymmetric gradient correction for the subsequent new tasks (see Sec. 3.3), thus avoiding TemporalCF. On the server-side, we group the average gradient updates uploaded by the clients (see Sec. 3.4) and aggregate multiple global models to avoid SpatialCF, while also satisfying the personalized local requirements." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.903 + ], + "angle": 0, + "content": "The overall framework of FedAGC is shown in Figure 2, the algorithm is summarized in Algorithm 1." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3843" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.465, + 0.108 + ], + "angle": 0, + "content": "3.1. Personalized Federated Continual Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.485, + 0.28 + ], + "angle": 0, + "content": "We extend the traditional CL and FL methods to PFCL, which enables the independent and parallel integration of multiple task sequences across clients. This approach achieving both the fusion of local knowledge from new and old tasks and the integration of heterogeneous task knowledge across clients. Unlike FCL, which aims to obtain a general model that integrates global knowledge, PFCL addresses the diverse local needs of clients in highly heterogeneous scenarios, resulting in personalized local models. Specifically, each client learns its local model by optimizing the following objective:" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.287, + 0.483, + 0.318 + ], + "angle": 0, + "content": "\\[\n\\min _ {w ^ {(t)}} \\mathcal {L} \\left(w ^ {(t)}; \\mathcal {T} ^ {t}\\right) + \\frac {\\lambda}{2} \\| w ^ {(t)} - w _ {0} \\| ^ {2}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.326, + 0.483, + 0.356 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}(\\cdot)\\) represents the empirical risk loss, \\(\\lambda\\) is the regularization coefficient, and \\(w_{0}\\) denotes the initial model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.356, + 0.483, + 0.387 + ], + "angle": 0, + "content": "The PFCL scenario we consider, involving independent parallel multi-task sequences, comprises three key steps:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.387, + 0.483, + 0.493 + ], + "angle": 0, + "content": "(1) Given \\(N\\) clients \\(\\{\\mathcal{C}_1,\\mathcal{C}_2,\\dots ,\\mathcal{C}_N\\}\\) and a server \\(\\mathcal{S}\\). There are task streams \\(\\{\\mathcal{T}^1,\\mathcal{T}^2,\\ldots ,\\mathcal{T}^N\\}\\), within which each client \\(\\mathcal{C}_c\\) engages in training a local CL model on its unique, private, and accessible task sequence \\(\\mathcal{T}^c = \\{\\mathcal{T}_1^c,\\mathcal{T}_2^c,\\dots ,\\mathcal{T}_T^c\\}\\). \\(\\mathcal{T}^c\\) consists of \\(T\\) distinct tasks, where the \\(t\\)-th task \\(\\mathcal{T}_t^c\\) includes \\(K_{t}^{c}\\) classes, and the total set of classes for \\(\\mathcal{C}_c\\) is represented as \\(K^{c} = \\{K_{1}^{c},K_{2}^{c},\\dots ,K_{T}^{c}\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.483, + 0.598 + ], + "angle": 0, + "content": "(2) For the current task \\(\\mathcal{T}_t^c\\) being learned, the participating client receives the model \\(\\theta_g^t\\) globally broadcasted by \\(S\\) as the initial model, where \\(\\theta_g^t\\) incorporates knowledge from the previous \\(t - 1\\) tasks contributed by various clients. After \\(e\\) rounds of local training on \\(\\mathcal{T}_t^c\\), \\(\\mathcal{C}_c\\) obtains a new continual learning model \\(\\theta_{l,c}^{t}\\) which performs well on \\(K^c\\), and the corresponding average gradient update \\(G^{c}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.483, + 0.705 + ], + "angle": 0, + "content": "(3) In each communication round \\( r \\), the participating client \\( \\mathcal{C}_c \\) upload its respective \\( G^c \\). \\( S \\) employs a group personalization strategy to aggregate them, which is then used to update \\( A \\) global models \\( \\theta_{g,A}^t \\). For the \\( a \\)-th global model \\( \\theta_{g,a}^t \\), it encapsulates the generalized knowledge of the associated client group's learning tasks and serves as the initial model for the next training round." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.387, + 0.483, + 0.705 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.336, + 0.729 + ], + "angle": 0, + "content": "3.2. Memory Rehearsal in FCL" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.84 + ], + "angle": 0, + "content": "Benefiting from the efficient utilization of past experience, memory rehearsal methods [2, 10, 27] are widely used in CL as simple and effective forgetting suppression techniques. These methods involve storing partial data or knowledge from previous tasks to enable retraining during the learning of new tasks, thereby mitigating forgetting of prior tasks. Now, we extend it from CL to FCL." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Memory Design. Each client utilizes its limited local memory to construct a small cache space, denoted as \\( M_{c} \\). Upon the completion of the training process for the \\( t \\)-th task of \\( \\mathcal{C}_c \\), rehearsal samples \\( M_t^c \\) are sampled at a fixed proportion" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.138 + ], + "angle": 0, + "content": "\\(\\mu\\) from the training set \\(D_{t}^{c}\\) of that task. For the initial task, \\(M_0^c = \\emptyset\\), and for the final task, no additional sampling operations are performed. Thus, \\(M_{c}\\) can be represented as" + }, + { + "type": "equation", + "bbox": [ + 0.655, + 0.144, + 0.907, + 0.163 + ], + "angle": 0, + "content": "\\[\nM _ {c} = \\cup_ {t = 1} ^ {T} M _ {t} ^ {c}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.171, + 0.871, + 0.187 + ], + "angle": 0, + "content": "where the size of \\(M_t^c\\) is significantly smaller than \\(D_t^c\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.186, + 0.907, + 0.322 + ], + "angle": 0, + "content": "Centroid-Based Representative Sample Selection. Clearly, the quality of samples drawn from the old task dataset directly affects the quality of the memory cache. Therefore, we propose a Centroid-Based Representative Sample Selection (CBRSS) method to effectively acquire high-quality rehearsal samples. In particular, for client \\( \\mathcal{C}_c \\), given the distance threshold \\( v \\) for centroids and the input data \\( x \\), the constraint condition for the minimum preset distance is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.332, + 0.907, + 0.353 + ], + "angle": 0, + "content": "\\[\n\\min _ {i} \\| m _ {i} ^ {c} - f _ {\\theta , c} (x) \\| > v, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.362, + 0.907, + 0.468 + ], + "angle": 0, + "content": "where \\(m_i^c\\) is the \\(i\\)-th centroid, and \\(f_{\\theta,c}(x)\\) is the predictor of \\(\\mathcal{C}_c\\) for the current task, used to map \\(x\\) to the feature space. If the condition in Eq. (3) is satisfied, it indicates that this data point is independent and can be considered as a new centroid \\(m_{new,c} = f_{\\theta,c}(x)\\). Otherwise, it means that this data point belongs to the existing centroid with the smallest distance to it, and then the target centroid is updated by" + }, + { + "type": "equation", + "bbox": [ + 0.612, + 0.475, + 0.906, + 0.511 + ], + "angle": 0, + "content": "\\[\nm _ {i ^ {*}} = \\frac {N _ {i ^ {*}} ^ {k} \\times m _ {i} + f _ {\\theta , c} (x)}{N _ {i ^ {*}} ^ {k} + 1}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.518, + 0.906, + 0.571 + ], + "angle": 0, + "content": "where \\( i^{*} = \\underset {i}{\\arg \\min}\\| f_{\\theta ,c}(x) - m_{i}^{c}\\| \\), and \\( N_{i^*}^k \\) represents the total number of samples in the \\( k \\)-th category captured by the centroid closest to \\( x \\) in the current task." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.571, + 0.907, + 0.723 + ], + "angle": 0, + "content": "Overall, each client stores the representative data of each category for the trained task after completing the training. A similar approach is employed by FedMes [42], which utilizes a random sampling method to construct the memory cache. However, due to the small sample size, this approach inevitably results in significant sampling bias. The resulting \\( M_{t}^{c} \\) lacks representativeness and fails to reflect the overall distribution of \\( D_{t}^{c} \\). This leads to a drift in the representation of old task data in the feature space, which in turn triggers the problem of catastrophic forgetting." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.731, + 0.902, + 0.747 + ], + "angle": 0, + "content": "3.3. Asymmetrically Correcting Gradient Updates" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.752, + 0.906, + 0.813 + ], + "angle": 0, + "content": "To integrate the knowledge from previous tasks with that of the current task during the learning process of the \\(t\\)-th task on \\(\\mathcal{C}_c\\), we reformulate the training problem as a MOO problem, which is formally described as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.535, + 0.822, + 0.907, + 0.846 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta_ {c}, \\theta_ {c} ^ {1: t}} \\ell \\left(f _ {\\theta_ {c}}, f _ {\\theta_ {c} ^ {t}}, D _ {t} ^ {c}\\right) \\cup \\ell \\left(f _ {\\theta_ {c}}, f _ {\\theta_ {c} ^ {k}}, M _ {k} ^ {c}\\right) \\mid \\forall k < t. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Here, \\(\\ell(\\cdot)\\) denotes the loss function, and \\(\\theta_c\\) and \\(\\theta_c^t\\) represent the backbone model for shared knowledge across multiple tasks and the task-specific classifier for \\(\\mathcal{C}_c\\), respectively." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3844" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.198 + ], + "angle": 0, + "content": "Since the parameters of the new and old task classifiers are independent and can be updated directly using their respective gradients \\(\\theta_c^k\\gets \\theta_c^k -\\alpha \\nabla \\ell (f_{\\theta_c},f_{\\theta_c^k})\\) with step size \\(\\alpha\\) our focus is on updating the shared backbone model, which is updated using the gradients \\(g_{k}^{c} = \\nabla \\ell (f_{\\theta_{c}},f_{\\theta_{c}^{k}})\\) of all \\(t\\) tasks. Let \\(\\beta\\) be the step size for optimization. The problem of the backbone model update can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.209, + 0.484, + 0.234 + ], + "angle": 0, + "content": "\\[\n\\theta_ {c} \\leftarrow \\theta_ {c} - \\beta d _ {c} ^ {*}, d _ {c} ^ {*} = \\arg \\min _ {d _ {c}} D \\left(g _ {k} ^ {c}, d _ {c}\\right) \\mid \\forall k \\leq t, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.244, + 0.485, + 0.35 + ], + "angle": 0, + "content": "where \\( d_c^* \\) represents the optimal gradient, which approximates the gradient of the new task while avoiding conflicts with the gradient of the old task. In Eq. (6), we define it as a min-distance multi-objective problem. Therefore, the critical issue lies in determining the method used to compute the differences between gradients, ensuring that the solution for \\( d_c^* \\) is as close as possible to all gradients." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.351, + 0.485, + 0.609 + ], + "angle": 0, + "content": "Symmetric distance metrics, such as Euclidean distance and cosine distance, are widely applied in deep neural networks. These metrics share a fundamental property: the forward impact (from \\( x \\) to \\( y \\)) is equal to the backward impact (from \\( y \\) to \\( x \\)). However, local model updates are highly dependent on the magnitude and direction of gradients, both of which are inherently asymmetric. In brief, given an old task gradient \\( g_{old} \\), trained on \\( M_c^k \\), and a new task gradient \\( g_{new} \\), trained on \\( D_t^c \\), their mutual influences on each other's tasks are significantly different. In previous studies [2, 42, 49], the condition \\( \\langle g_{old}, g_{new} \\rangle \\geq 0 \\) has been employed to signify the absence of conflicts between two tasks. However, given that the old task has been integrated earlier, \\( \\| g_{old} \\| \\) is typically smaller while \\( \\| g_{new} \\| \\) tends to be larger. In cases where \\( \\| g_{old} \\| \\ll \\| g_{new} \\| \\), even if \\( \\langle g_{old}, g_{new} \\rangle \\geq 0 \\) holds true, the updates pertaining to the new task may have a substantial impact on the performance of the old task." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.618, + 0.484, + 0.664 + ], + "angle": 0, + "content": "Definition 1 (Asymmetric Gradient Distance). Given two gradients \\( g_{A} \\) and \\( g_{B} \\), the asymmetric gradient distance is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.674, + 0.405, + 0.709 + ], + "angle": 0, + "content": "\\[\nD \\left(g _ {A}, g _ {B}\\right) = \\frac {\\left\\| g _ {A} - g _ {B} \\right\\|}{\\left\\| g _ {B} \\right\\| + \\left\\| g _ {A} - g _ {B} \\right\\|}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.796 + ], + "angle": 0, + "content": "In this paper, an asymmetric distance calculation method is used to explore the measure of difference between the new task and the old task. Definition 1 considers both the gradient direction and magnitude. Notably, \\( D(g_{A}, g_{B}) = 0 \\) when \\( g_{A} = g_{B} = 0 \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.902 + ], + "angle": 0, + "content": "However, solving Eq. (6) directly for the parallel task flow on clients presents significant challenges. This difficulty arises from the fact that both the decision space of this problem and the model parameters are identically high-dimensional. Consequently, following [28, 35], we employ linear scalarization to determine a weight variable \\( w^{*} \\). Specifically, when client \\( C_c \\) is training on the \\( t \\)-th task, the" + }, + { + "type": "title", + "bbox": [ + 0.516, + 0.092, + 0.667, + 0.106 + ], + "angle": 0, + "content": "Algorithm 1 FedAGC" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.109, + 0.907, + 0.158 + ], + "angle": 0, + "content": "Input: \\(N\\) clients with their private parallel task sequence \\(\\mathcal{T}^{1:N}\\) and local datasets \\(\\{D_t^{1:N}\\}_{t=1}^T\\), global model parameter \\(\\theta_g\\), communication rounds \\(R\\), memory space \\(M_{1:c} \\gets \\{\\}\\)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.157, + 0.906, + 0.19 + ], + "angle": 0, + "content": "Output: Multiple global parameters \\(\\theta_{g}^{i:n}\\), local parameter \\(\\theta_{l,1:N}\\) and classification head \\(\\theta_{l,1:N}^{k}\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.191, + 0.631, + 0.204 + ], + "angle": 0, + "content": "1: Initialization" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.206, + 0.734, + 0.219 + ], + "angle": 0, + "content": "2: while task number \\( t \\leq T \\) do" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.221, + 0.787, + 0.235 + ], + "angle": 0, + "content": "3: for each round \\(r = 1,2,\\ldots ,R\\) do" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.236, + 0.875, + 0.252 + ], + "angle": 0, + "content": "4: Server broadcasts \\(\\theta_g^i\\) to each group of clients;" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.252, + 0.842, + 0.265 + ], + "angle": 0, + "content": "5: for each client \\( c = 1 \\) to \\( N \\) in parallel do" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.266, + 0.872, + 0.28 + ], + "angle": 0, + "content": "6: Minimizing local CL problems by Eq. (1);" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.281, + 0.747, + 0.297 + ], + "angle": 0, + "content": "7: \\(g^{c} = \\nabla \\ell (f_{\\theta_{c}},f_{\\theta^{t}},D_{t}^{c});\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.297, + 0.837, + 0.312 + ], + "angle": 0, + "content": "8: \\(g_{k}^{c} = \\nabla \\ell (f_{\\theta_{c}},f_{\\theta_{k}^{k}},M_{k}^{c})\\) for all \\(k < t\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.313, + 0.8, + 0.328 + ], + "angle": 0, + "content": "9: \\(d_c^* \\gets \\mathsf{ACGU}(g^c, g_1^c, \\ldots, g_{t-1}^k)\\);" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.329, + 0.716, + 0.345 + ], + "angle": 0, + "content": "10: \\(\\theta_{l,c}^{t}\\gets \\theta_{l,c}^{t} - \\alpha d_{c}^{*};\\)" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.346, + 0.63, + 0.356 + ], + "angle": 0, + "content": "11: end for" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.358, + 0.885, + 0.374 + ], + "angle": 0, + "content": "12: Send the \\(G^{c} = -\\triangle \\theta_{l,c}^{t} / \\eta\\) to the central server;" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.375, + 0.888, + 0.39 + ], + "angle": 0, + "content": "13: \\(\\{S_1,S_2,\\ldots ,S_n\\} \\gets \\mathrm{GAGU}(G^1,G^2,\\ldots ,G^N);\\)" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.391, + 0.893, + 0.408 + ], + "angle": 0, + "content": "14: \\(\\theta_{g}^{i}\\gets \\theta_{g}^{i} - \\frac{1}{|S_{i}|}\\sum_{j\\in S_{i}}G^{j}\\) for \\(i\\in \\{1,2,\\ldots ,n\\}\\)" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.409, + 0.612, + 0.42 + ], + "angle": 0, + "content": "15: end for" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.422, + 0.905, + 0.437 + ], + "angle": 0, + "content": "16: Each client samples \\( M_t^c \\) from \\( D_t^c \\) by Eqs. (3) and (4);" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.438, + 0.613, + 0.45 + ], + "angle": 0, + "content": "17: end while" + }, + { + "type": "list", + "bbox": [ + 0.52, + 0.191, + 0.905, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.907, + 0.507 + ], + "angle": 0, + "content": "gradient update \\( g_{t}^{c} \\) of the new task and \\( g_{1:t-1}^{c} \\) from the old tasks and computed through" + }, + { + "type": "equation", + "bbox": [ + 0.556, + 0.513, + 0.906, + 0.563 + ], + "angle": 0, + "content": "\\[\nw ^ {*} = \\arg \\min _ {w} \\left\\{\\frac {1}{t} \\sum_ {i} ^ {t} D \\left(\\sum_ {j \\neq i} ^ {t} w _ {j} g _ {j} ^ {c}, g _ {i} ^ {c}\\right) \\right\\}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.906, + 0.661 + ], + "angle": 0, + "content": "where \\(\\sum_{j=1}^{t} w_{j} = 1\\), \\(\\forall w_{j} \\geq 0\\), and \\(D(\\sum_{j \\neq i}^{t} w_{j} g_{j}^{c}, g_{i}^{c})\\) represents the divergence between task \\(i\\) and the combination of all other tasks. By minimizing this divergence, we achieve the suppression of the influence from the combination on task \\(i\\). Note that this optimization process does not modify individual gradient updates." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.661, + 0.906, + 0.72 + ], + "angle": 0, + "content": "To prevent training failure on new tasks, we introduce the Multiple Gradient Descent Algorithm (MGDA) [4] to minimize the gradient norm, thereby achieving a balanced solution. It is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.725, + 0.906, + 0.77 + ], + "angle": 0, + "content": "\\[\nw _ {b} ^ {*} = \\underset {w} {\\arg \\min } \\left\\| \\sum_ {i} ^ {t} w _ {i} g _ {i} ^ {c} \\right\\| ^ {2}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.775, + 0.906, + 0.835 + ], + "angle": 0, + "content": "When the objective function value of MGDA is zero, the obtained solution satisfies the Karush-Kuhn-Tucker conditions. Otherwise, the solution provides a Pareto descent direction that does not harm any task." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.836, + 0.905, + 0.867 + ], + "angle": 0, + "content": "In summary, our proposed Asymmetric Correction Gradient Updating (ACGU) algorithm determines \\(d_c^*\\) by" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.873, + 0.906, + 0.905 + ], + "angle": 0, + "content": "\\[\nd _ {c} ^ {*} = \\sum_ {i} W _ {i} g _ {i} ^ {c}, \\quad \\text {w h e r e} W = \\gamma w ^ {*} + (1 - \\gamma) w _ {b} ^ {*}. \\tag {9}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3845" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.907, + 0.29 + ], + "angle": 0, + "content": "
AlgorithmAttributes FCL Per GCBTask ID under Non-IID Setting (Across 8 Different Datasets)\\( A_{avg} \\)(%)
\\( A_1 \\)(%)\\( A_2 \\)(%)\\( A_3 \\)(%)\\( A_4 \\)(%)\\( A_5 \\)(%)
FedAvg× × ×54.73 ± 1.4856.06 ± 3.1663.99 ± 3.4357.90 ± 2.9678.49 ± 2.1861.83 ± 0.46
FedProx× × ×58.25 ± 1.6056.65 ± 0.5066.15 ± 2.3361.72 ± 2.2182.24 ± 1.0265.00 ± 0.62
Ditto× ✓ ×58.92 ± 1.4059.13 ± 1.6370.58 ± 3.0763.00 ± 2.7983.97 ± 0.9567.12 ± 0.85
FedCurv✓ × ✓60.18 ± 3.4354.88 ± 1.9968.05 ± 3.7859.25 ± 2.3782.19 ± 0.5864.91 ± 0.56
FedAGEM✓ × ✓69.62 ± 0.8566.91 ± 1.0975.06 ± 1.6766.73 ± 2.4781.73 ± 1.1672.01 ± 0.81
FedER✓ × ×73.34 ± 0.4568.77 ± 1.8980.18 ± 1.3970.19 ± 1.3079.43 ± 1.3274.38 ± 0.27
FedMes✓ ✓ ✓70.84 ± 0.7967.22 ± 1.0577.25 ± 0.8967.18 ± 1.5285.21 ± 0.3673.54 ± 0.55
Ours-w/o-ACGU× ✓ ×58.62 ± 0.3658.06 ± 0.1665.62 ± 0.1963.29 ± 1.8682.46 ± 1.6865.61 ± 0.80
Ours-w/o-GAGU✓ × ✓76.19 ± 0.5271.59 ± 1.1282.24 ± 1.4272.89 ± 2.0481.82 ± 2.3676.95 ± 1.38
Ours-w/o-CBRSS✓ ✓ ✓75.40 ± 0.9072.39 ± 0.5582.88 ± 0.9275.42 ± 0.5486.68 ± 0.6278.56 ± 0.57
FedAGC (Ours)✓ ✓ ✓76.21 ± 0.3672.56 ± 0.4884.90 ± 0.8577.60 ± 0.5787.28 ± 0.1879.71 ± 0.23
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.3, + 0.907, + 0.342 + ], + "angle": 0, + "content": "Table 1. Performance (avg ± std) of local models for each task on 10 clients under Non-IID setting. \"w/o-Module\" indicates performance without the corresponding module. It also includes Per (Personalized) and GCB (Gradient Correction-Based). For the proposed FedAGC, we used \\(\\lambda = 0.01\\) and \\(\\gamma = 0.5\\) in Eqs. (1) and (9)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.349, + 0.348, + 0.365 + ], + "angle": 0, + "content": "3.4. Group-Level Personalization" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.371, + 0.483, + 0.506 + ], + "angle": 0, + "content": "To integrate heterogeneous knowledge from different clients, some studies [39, 45] adopt personalization strategies at the client level, where each client continuously adjusts the received global model to better fit its local tasks. However, this approach struggles to meet the needs of all clients in large-scale parallel task scenarios. In contrast, we consider a group-level personalization strategy, enabling the broadcast of global models tailored to meet the specific requirements of each client." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.507, + 0.484, + 0.629 + ], + "angle": 0, + "content": "In this work, we propose Group Aggregation Gradient Updating (GAGU) for client-uploaded gradients. Specifically, during the \\(r\\)-th communication round, \\(N\\) clients upload their respective averaged gradient updates \\(\\{G^1,G^2,\\dots ,G^N\\}\\) obtained after \\(e\\) local training iterations, where \\(G^{c} = -\\triangle \\theta_{l,c}^{t} / \\alpha\\). We compute a similarity matrix \\(S\\) to measure the similarity between pairs of averaged gradient updates. The elements of \\(S\\) are given by" + }, + { + "type": "equation", + "bbox": [ + 0.221, + 0.638, + 0.482, + 0.672 + ], + "angle": 0, + "content": "\\[\ns _ {i, j} = \\frac {G ^ {i} G ^ {j}}{\\| G ^ {i} \\| \\| G ^ {j} \\|.} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.681, + 0.484, + 0.803 + ], + "angle": 0, + "content": "We construct a graph \\( P = \\{G, E\\} \\) to represent \\( A \\), where \\( G \\) denotes the set of nodes in the graph, and each node corresponds to the \\( G^c \\). The set of edges \\( E = \\{s_{i,j}\\} \\forall i \\leq j \\leq N \\) represents the similarity between \\( G^i \\) and \\( G^j \\). The objective of GAGU is to reorder \\( P \\) such that similar average gradient updates are grouped together, while dissimilar ones are isolated. To achieve this, we first define the similarity between any two groups \\( S_A \\) and \\( S_B \\) (\\( S_A \\cap S_B = \\emptyset \\), \\( \\forall A \\neq B \\)):" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.813, + 0.483, + 0.845 + ], + "angle": 0, + "content": "\\[\nh \\left(S _ {A}, S _ {B}\\right) = \\sum_ {i \\in S _ {A}, j \\in S _ {B}} s _ {i, j}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "GAGU aims to maximize intra-group similarity while minimizing inter-group similarity, resulting in an optimal graph partition \\( a^{*} = \\{S_{1}, S_{2}, \\ldots, S_{n}\\} \\), where \\( n \\) represents the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.35, + 0.907, + 0.395 + ], + "angle": 0, + "content": "number of groups in \\( a^* \\). Inspired by spectrum-clustering [13] and Ncut [36], We reformulate this optimization problem into the following form:" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.402, + 0.907, + 0.435 + ], + "angle": 0, + "content": "\\[\n\\min _ {a \\in A} \\sum_ {S, S ^ {\\prime} \\in a} h \\left(S, S ^ {\\prime}\\right) - h \\left(S ^ {\\prime}, S ^ {\\prime}\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.442, + 0.906, + 0.533 + ], + "angle": 0, + "content": "where \\( A \\) is all possible grouping results. Additionally, we introduce the modularity \\( Q \\) to evaluate the quality of the current partition during the process of determining the optimal number of groups and partitioning results. This metric measures the distribution of connectivity within and between groups." + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.539, + 0.907, + 0.577 + ], + "angle": 0, + "content": "\\[\nQ = \\frac {1}{2 W} \\sum_ {i, j} \\left[ s _ {i, j} - \\frac {d _ {i} d _ {j}}{2 W} \\right] \\delta \\left(S _ {i}, S _ {j}\\right), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.585, + 0.906, + 0.632 + ], + "angle": 0, + "content": "where \\( W = \\sum_{i,j} s_{i,j} \\) represents the total edge weight in the current graph, \\( d_i = \\sum_j s_{i,j} \\) denotes the degree of node \\( i \\), and \\( \\delta(S_i, S_j) \\) is the Kronecker delta function." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.643, + 0.649, + 0.66 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.668, + 0.707, + 0.684 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Benchmarks. We validate FedAGC in large-scale multitask parallel FCL scenarios. The experimental setup follows the Task-Incremental Learning paradigm and primarily includes two different benchmarks: (1) IID Setting (Single CIFAR-100): The 100 classes of CIFAR-100 [16] are divided into 20 tasks, each containing 5 classes. These tasks are randomly assigned to clients, ensuring that each client does not train on the same task multiple times. (2) Non-IID Setting (Across 8 Different Datasets): We partition and distribute eight widely used datasets in the FCL domain, such as CIFAR-10/100 [16], MNIST [18], SVHN [30], FashionMNIST [41], Not-MNIST [1], FaceScrub [32], and TrafficSigns [31], in the same manner as (1). In both benchmarks, each client is assigned a unique task sequence." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3846" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.907, + 0.29 + ], + "angle": 0, + "content": "
AlgorithmAttributes FCL Per GCBTask ID under IID Setting (Single CIFAR-100)\\( A_{avg} \\)(%)
\\( A_1 \\)(%)\\( A_2 \\)(%)\\( A_3 \\)(%)\\( A_4 \\)(%)\\( A_5 \\)(%)
FedAvg× × ×54.32 ± 0.9460.53 ± 2.0861.89 ± 1.5665.63 ± 1.6070.77 ± 0.8462.63 ± 0.55
FedProx× × ×53.96 ± 0.9861.53 ± 2.2362.90 ± 1.9465.70 ± 1.1370.78 ± 0.9162.97 ± 0.93
Ditto× ✓ ×55.48 ± 0.9262.81 ± 1.9663.12 ± 1.5866.59 ± 1.8671.01 ± 0.9363.67 ± 0.72
FedCurv✓ × ✓54.36 ± 0.9560.77 ± 0.8663.43 ± 0.9166.46 ± 0.9771.52 ± 0.7563.31 ± 0.16
FedAGEM✓ × ✓60.95 ± 0.9067.15 ± 1.2366.39 ± 0.6967.81 ± 0.9071.28 ± 0.6666.72 ± 0.23
FedER✓ × ×63.05 ± 0.7667.83 ± 0.9166.11 ± 0.3265.39 ± 0.4468.84 ± 1.0166.44 ± 0.42
FedMes✓ ✓ ✓61.18 ± 0.3467.68 ± 0.1167.11 ± 0.4967.42 ± 0.1572.12 ± 0.9067.13 ± 0.16
Ours-w/o-ACGU× ✓ ×55.45 ± 0.2761.95 ± 0.8763.32 ± 1.7265.51 ± 0.7371.44 ± 0.6063.53 ± 0.31
Ours-w/o-GAGU✓ × ✓63.92 ± 0.5566.73 ± 0.9367.22 ± 0.6166.48 ± 0.6568.53 ± 0.2466.58 ± 0.41
Ours-w/o-CBRSS✓ ✓ ✓66.14 ± 0.3769.05 ± 0.8069.69 ± 0.4969.56 ± 2.2074.12 ± 0.0969.75 ± 0.78
FedAGC (Ours)✓ ✓ ✓66.38 ± 0.4869.48 ± 0.7870.38 ± 1.3270.58 ± 1.8573.98 ± 0.4170.16 ± 0.77
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.3, + 0.778, + 0.314 + ], + "angle": 0, + "content": "Table 2. Performance (avg ± std) of local models for each task on 10 clients under IID setting." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.32, + 0.287, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.112, + 0.471, + 0.268, + 0.484 + ], + "angle": 0, + "content": "(a) \\(F_{t}\\) under Non-IID Setting" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.32, + 0.482, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.32, + 0.471, + 0.451, + 0.484 + ], + "angle": 0, + "content": "(b) \\(F_{t}\\) under IID Setting" + }, + { + "type": "image_caption", + "bbox": [ + 0.139, + 0.486, + 0.434, + 0.5 + ], + "angle": 0, + "content": "Figure 3. Comparison of average forgetting rates." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.503, + 0.484, + 0.623 + ], + "angle": 0, + "content": "Baselines. We compare FedAGC with state-of-the-art (SOTA) methods, including the fundamental FedAvg [29] and its proximal variant FedProx [19], along with the personalized FL method Ditto [20]. To verify the superiority of FedAGC, we also compare it with gradient correction-based FCL methods, including FedCurv [38], FedAGEM [2] and the latest FedMes [42]. Additionally, we consider the rehearsal-based method FedER [3]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.625, + 0.484, + 0.821 + ], + "angle": 0, + "content": "Metrics. We employ two critical metrics to evaluate algorithm performance: Average Accuracy \\((A_{avg})\\) and Forgetting Rate \\((F_t)\\). Let \\(a_{i,t}^c\\) denote the test accuracy of the \\(i\\)-th task learned by the \\(c\\)-th client after sequential learning of \\(t\\) subsequent tasks, where \\(t + i \\leq T\\). In this formulation, the initial accuracy of task \\(i\\) (immediately after its own training) and the final accuracy after training on all tasks are defined as \\(a_{i,0}^c\\) and \\(a_{i,T - i}^c\\), respectively. The forgetting rate for the \\(i\\)-th task after learning \\(t\\) subsequent tasks is calculated as: \\(f_{i,t}^{c} = \\frac{a_{i,0}^{c} - a_{i,t}^{c}}{a_{i,0}^{c}}\\). To aggregate results, we compute client-wise averages: \\(A_i = \\frac{1}{N}\\sum_{c=1}^N a_{i,T-i}^c\\) and \\(f_{i,t} = \\frac{1}{N}\\sum_{c=1}^N f_{i,t}^c\\). Then, \\(A_{avg}\\) and \\(F_t\\) are defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.833, + 0.483, + 0.874 + ], + "angle": 0, + "content": "\\[\nA _ {\\text {a v g}} = \\frac {1}{T} \\sum_ {i = 1} ^ {T} A _ {i}, \\quad F _ {t} = \\frac {1}{T - t} \\sum_ {i = 1} ^ {T - t} f _ {i, t}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Implementation details. Following the experimental setup" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.321, + 0.702, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.535, + 0.416, + 0.683, + 0.427 + ], + "angle": 0, + "content": "(a) Impact of \\(\\gamma\\) on FedAGC" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.428, + 0.706, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.321, + 0.903, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.738, + 0.416, + 0.887, + 0.427 + ], + "angle": 0, + "content": "(b) Impact of \\(\\lambda\\) on FedAGC" + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.428, + 0.902, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.531, + 0.528, + 0.693, + 0.54 + ], + "angle": 0, + "content": "(c) Non-IID Setting, \\(\\lambda = 0.01\\)" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.541, + 0.706, + 0.636 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.641, + 0.68, + 0.653 + ], + "angle": 0, + "content": "(e) IID Setting, \\(\\lambda = 0.01\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.738, + 0.528, + 0.881, + 0.54 + ], + "angle": 0, + "content": "(d) Non-IID Setting, \\(\\gamma = 0.5\\)" + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.541, + 0.902, + 0.636 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.751, + 0.641, + 0.866, + 0.653 + ], + "angle": 0, + "content": "(f) IID Setting, \\(\\gamma = 0.5\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.656, + 0.873, + 0.67 + ], + "angle": 0, + "content": "Figure 4. Impact of core hyperparameters on FedAGC." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.907, + 0.841 + ], + "angle": 0, + "content": "of [47], we adopted a variant of the LeNet architecture as the backbone. For any newly introduced task, we initialized a new classifier. The Adam optimizer was used for all datasets, with an initial learning rate of \\(10^{-3}\\) and a weight decay of \\(10^{-4}\\). The experimental setup consisted of 10 clients, each learning 5 sequential and non-overlapping tasks. The batch size was set to 100 per step, with \\(\\mu\\) and \\(v\\) set to 0.1 and 8, respectively. To ensure convergence, all experiments were conducted for 100 federated communication rounds. Each experiment was repeated five times, and the reported results were averaged accordingly." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.653, + 0.863 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Tabs. 1 and 2 record the average accuracy of 10 clients on each task under two data condition setups, along with the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3847" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.44 + ], + "angle": 0, + "content": "final results considering all tasks. In Fig. 3, we compare the degree of Spatial-TemporalCF between FedAGC and various methods. It can be observed that FedProx and Ditto mitigate the impact of statistical heterogeneity using penalty terms, showing significant performance improvements over FedAvg in Non-IID settings. However, since they do not employ any CL strategies, their accuracy is lower compared to FCL methods such as FedAGEM, FedER, and FedMes, with higher forgetting levels. Notably, FedCurv's experimental results are suboptimal as it leverages curvature information from other clients to guide updates for the current client, which is prone to knowledge conflicts in cases of non-overlapping client data distributions. These FCL methods based on symmetric gradient correction focus solely on TemporalCF without considering the interference of knowledge across clients, leading to SpatialCF. In contrast, the proposed FedAGC outperforms other methods in both task-specific accuracy and overall performance across both settings, with a lower degree of forgetting. This demonstrates that memory buffers constructed through importance sampling, assisted by asymmetric gradient correction and group-level personalization, can significantly improve the accuracy of PFCL models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.453, + 0.388, + 0.469 + ], + "angle": 0, + "content": "4.3. Performance Analysis of FedAGC" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.477, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Ablation Studies. We conducted three different ablation studies under the same experimental setup, with the results recorded in Tabs. 1 and 2, as well as Fig. 3. We progressively removed components from FedAGC, including: (1) Omitting the ACGU module (FedAGC-w/o-ACGU), which means that only the gradients from local training are used to guide updates. Clearly, in this case, the CBRSS module also loses its significance. The performance degradation is significant, with a drop of \\(17.7\\%\\) and \\(9.5\\%\\), as no strategy for mitigating TemporalCF is applied. (2) Removing the GAGU module (FedAGC-w/o-GAGU) and switching to the FedAvg average aggregation method. In the IID setting, FedAGC-w/o-GAGU's overall performance is overtaken by FedMes by \\(0.86\\%\\). (3) Disabling the CBRSS module (FedAGC-w/o-CBRSS) and replacing it with a random sampling method. In this case, performance decreases by \\(1.44\\%\\) and \\(0.58\\%\\) in the Non-IID and IID settings, respectively. However, since the core component for mitigating forgetting is retained, its accuracy on individual tasks and overall performance remains superior to other baselines. These results verify that our three proposed modules of FedAGC play a crucial role in alleviating Spatial-TemporalCF and achieving personalization, and can effectively be combined to enhance performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Impact of core hyperparameters. We evaluate FedAGC with different values of \\(\\gamma\\) and \\(\\lambda\\). As shown in Fig. 4, setting \\(\\lambda\\) too large or too small causes FedAGC to overly rely on either special or balanced solution, impairing its ability to" + }, + { + "type": "image", + "bbox": [ + 0.573, + 0.091, + 0.849, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.604, + 0.236, + 0.819, + 0.248 + ], + "angle": 0, + "content": "(a) \\(A_{i}\\) on the CIFAR-100 dataset, \\(N = 30\\)" + }, + { + "type": "image", + "bbox": [ + 0.573, + 0.25, + 0.846, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.604, + 0.393, + 0.818, + 0.405 + ], + "angle": 0, + "content": "(b) \\(F_{t}\\) on the CIFAR-100 dataset, \\(N = 30\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.576, + 0.407, + 0.843, + 0.422 + ], + "angle": 0, + "content": "Figure 5. Performance on large-scale clients." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.424, + 0.907, + 0.5 + ], + "angle": 0, + "content": "mitigate forgetting. Similarly, an excessively large \\(\\gamma\\) biases the model toward the inversion of previous tasks, hindering the learning of current task, while a small \\(\\gamma\\) has the opposite effect. Empirically, we find that the best performance of FedAGC is achieved when \\(\\gamma = 0.5\\) and \\(\\lambda = 0.01\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.501, + 0.909, + 0.622 + ], + "angle": 0, + "content": "Performance on large-scale clients. To explore the scalability of FedAGC to larger client scenarios, we tripled the number of clients from the previous setting of 10 and continued comparing it with various SOTA methods. The results recorded in Fig. 5 indicate that FedAGC outperforms other baselines in accuracy on previous tasks and demonstrates greater stability in mitigating forgetting. This result highlights the superiority of FedAGC." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.634, + 0.634, + 0.65 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To tackle the challenge of STCF in PFCL, this paper proposed FedAGC as a valid solution. After local training, FedAGC selects representative samples and stores them in a buffer for rehearsal to alleviate forgetting. To refine this process, we introduced an asymmetric gradient metric to quantify discrepancies between past and current tasks. Asymmetric gradient correction is then employed to adjust the gradient updates, jointly considering direction and magnitude for optimal adaptation. Moreover, FedAGC adopts group-level personalization by clustering clients based on the similarity of their uploaded averaged gradient updates, which enhances the fusion of valuable knowledge while isolating ineffective knowledge. It was extensively demonstrated in this paper that the effective integration of the proposed modules can enhance the model's resistance to STCF, enabling FedAGC to outperform SOTA approaches." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3848" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.21, + 0.108 + ], + "angle": 0, + "content": "Acknowledge" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.298 + ], + "angle": 0, + "content": "We want to thank the anonymous reviewers for their valuable suggestions and comments. This work was supported by the National Natural Science Foundation of China (No. 62276182), Tianjin University under the Emerging Direction Cultivation Project of Interdisciplinary Center (Intelligent Protection and Utilization of Digital Cultural Heritage), National Key Research and Development Program of China (No. 2023YFF0906204), Peng Cheng Lab Program (No. PCL2023A08), Tianjin Natural Science Foundation (Nos. 24JCYBJC01230, 24JCYBJC01460), and Tianjin Municipal Education Commission Research Plan (No. 2024ZX008)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.31, + 0.188, + 0.326 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.336, + 0.482, + 0.376 + ], + "angle": 0, + "content": "[1] Yaroslav Bulatov. Notmnist dataset. http://yaroslavvb.blogspot.it/2011/09/notmnist-dataset.html.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.379, + 0.482, + 0.434 + ], + "angle": 0, + "content": "[2] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a-gem. In Proceedings of the International Conference on Learning Representations, 2018. 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.436, + 0.482, + 0.504 + ], + "angle": 0, + "content": "[3] Arslan Chaudhry, Marcus Rohrbach, Mohamed Elhoseiny, Thalaiyasingam Ajanthan, Puneet K. Dokania, Philip H. S. Torr, and MarcÁurelio Ranzato. On tiny episodic memories in continual learning. In arXiv Preprint arXiv:1902.10486, 2019. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.507, + 0.483, + 0.547 + ], + "angle": 0, + "content": "[4] Désideri and Jean-Antoine. Multiple-gradient descent algorithm (mgda) for multiobjective optimization. Comptes Rendus Mathematique, 350(5-6):313-318, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.549, + 0.482, + 0.616 + ], + "angle": 0, + "content": "[5] Prithviraj Dhar, Rajat Vikram Singh, Kuan-Chuan Peng, Ziyan Wu, and Rama Chellappa. Learning without memorizing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5138-5146, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.62, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[6] Jiahua Dong, Lixu Wang, Zhen Fang, Gan Sun, Shichao Xu, Xiao Wang, and Qi Zhu. Federated class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10164-10173, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.69, + 0.482, + 0.745 + ], + "angle": 0, + "content": "[7] Alireza Fallah, Aryan Mokhtari, and Asuman Ozdaglar. Personalized federated learning with theoretical guarantees: A model-agnostic meta-learning approach. Advances in neural information processing systems, 33:3557–3568, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.747, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[8] Mehrdad Farajtabar, Navig Azizan, Alex Mott, and Ang Li. Orthogonal gradient descent for continual learning. In International Conference on Artificial Intelligence and Statistics, PMLR, pages 3762-3773, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.804, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[9] Bagwe G, Yuan X, Pan M, and et al. Fed-cprompt: Contrastive prompt for rehearsal-free federated continual learning. arXiv preprint arXiv:2307.04869, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.847, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[10] Yunhui Guo, Mingrui Liu, Tianbao Yang, and Tajana Rosing. Improved schemes for episodic memory-based lifelong learning. Advances in Neural Information Processing Systems, 33:1023-1035, 2020. 4" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.336, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[11] Piao H, Wu Y, Wu D, and et al. Federated continual learning via prompt-based dual knowledge transfer. In *Forty-First International Conference on Machine Learning (ICML)*. PMLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.191 + ], + "angle": 0, + "content": "[12] Parisa Hamedi, Roozbeh Razavi-Far, and Ehsan Hallaji. Federated continual learning: Concepts, challenges, and solutions. Neurocomputing, 651:130844, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.193, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[13] Hongjie Jia, Shifei Ding, Xinzheng Xu, and Ru Nie. The latest research progress on spectral clustering. Neural Computing and Applications, 24:1477-1486, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.236, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[14] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank J.Reddi, Sebastian U.Stich, and Ananda Theertha Suresh. Scaffold: Stochastic controlled averaging for federated learning. In International conference on machine learning. PMLR, pages 5132-5143, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.307, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[15] Alexander Krawczyk and Alexander Gepperth. An analysis of best-practice strategies for replay and rehearsal in continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4196-4204, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.378, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[16] Alex Krizhevsky and Geoffrey Hinton. Learning Multiple Layers of Features from Tiny Images. 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.407, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[17] Matthias De Lange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Aleš Leonardis, Gregory Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying forgetting in classification tasks. In IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 3366-3385, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.491, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[18] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[19] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. In Proceedings of Machine Learning and Systems, pages 429-450, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.605, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[20] Tian Li, Shengyuan Hu, Ahmad Beirami, and Virginia Smith. Ditto: Fair and robust federated learning through personalization. In International Conference on Machine Learning, PMLR, pages 6357-6368, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.662, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[21] Junkang Liu, Yuanyuan Liu, Fanhua Shang, Hongying Liu, Jin Liu, and Wei Feng. Improving generalization in federated learning with highly heterogeneous data via momentum-based stochastic controlled weight averaging. In Forty-second International Conference on Machine Learning. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.733, + 0.905, + 0.813 + ], + "angle": 0, + "content": "[22] Junkang Liu, Fanhua Shang, Yuanyuan Liu, Hongying Liu, Yuangang Li, and YunXiang Gong. Fedbcgd: Communication-efficient accelerated block coordinate gradient descent for federated learning. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 2955-2963, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[23] Quande Liu, Cheng Chen, Jing Qin, Qi Dou, and Pheng-Ann Heng. Feddg: Federated domain generalization on medical image segmentation via episodic learning in continuous frequency space. In CVPR, pages 1013–1023, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] Yang Liu, Anbu Huang, Yun Luo, He Huang, Youzhi Liu, Yuanyuan Chen, Lican Feng, Tianjian Chen, Han Yu, and" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3849" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "Qiang Yang. Fedvision: An online visual object detection platform powered by federated learning. In AAAI, pages 13172-13179, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.135, + 0.483, + 0.176 + ], + "angle": 0, + "content": "[25] David Lopez-Paz and MarcÁurelio Ranzato. Gradient episodic memory for continual learning. In Advances in neural information processing systems, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[26] Bing Luo, Wenli Xiao, Shiqiang Wang, Jianwei Huang, and Leandros Tassiulas. Tackling system and statistical heterogeneity for federated learning with adaptive client sampling. In IEEE INFOCOM 2022 - IEEE Conference on Computer Communications, pages 1739-1748, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[27] Fan Lyu, Shuai Wang, Wei Feng, Zihan Ye, Fuyuan Hu, and Song Wang. Multi-domain multi-task rehearsal for lifelong learning. Proceedings of the AAAI Conference on Artificial Intelligence, 35(10):8819-8827, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.303, + 0.482, + 0.37 + ], + "angle": 0, + "content": "[28] Fan Lyu, Qing Sun, Fanhua Shang, Liang Wan, and Wei Feng. Measuring asymmetric gradient discrepancy in parallel continual learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11411-11420, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.482, + 0.44 + ], + "angle": 0, + "content": "[29] H. Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial Intelligence and Statistics, PMLR, pages 1273-1282, 2017. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.482, + 0.51 + ], + "angle": 0, + "content": "[30] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading digits in natural images with unsupervised feature learning. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning, page 4, 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.58 + ], + "angle": 0, + "content": "[31] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading digits in natural images with unsupervised feature learning. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning, page 4, 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.636 + ], + "angle": 0, + "content": "[32] Hong-Wei Ng and Stefan Winkler. A data-driven approach to cleaning large face datasets. In Proceedings of the 2014 IEEE International Conference on Image Processing (ICIP), pages 343–347. IEEE, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.482, + 0.692 + ], + "angle": 0, + "content": "[33] Li Qinbin, Diao Yiqun, Chen Quan, and He Bingsheng. Federated learning on non-iid data silos: An experimental study. In 2022 IEEE 38th International Conference on Data Engineering (ICDE), pages 965-978, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[34] Felix Sattler, Simon Wiedemann, Klaus-Robert Müller, and Wojciech Samek. Robust and communication-efficient federated learning from non-iid data. IEEE Transactions on Neural Networks and Learning Systems, 31(9):3400-3413, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[35] Ozan Sener and Vladlen Koltun. Multi-task learning as multi-objective optimization. In Advances in Neural Information Processing Systems, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[36] Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 22(8):888-905, 2000. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[37] Yong Shi, Yuanying Zhang, Yang Xiao, and Lingfeng Niu. Optimization strategies for client drift in federated learning: A review. Procedia Computer Science, 214:1168-1173, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[38] Neta Shoham, Tomer Avidor, Aviv Keren, Nadav Israe, Daniel Benditkis, Liron Mor-Yosef, and Itai Zeitak. Overcoming forgetting in federated learning on non-iid data. In arXiv Preprint arXiv:1910.07796, 2019. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[39] Hongda Wu and Ping Wang. Fast-convergent federated learning with adaptive weighting. IEEE Transactions on Cognitive Communications and Networking, 7(4):1078-1088, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.206, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[40] Yichen Wu, Hong Wang, Peilin Zhao, Yefeng Zheng, Ying Wei, and Long-Kai Huang. Mitigating catastrophic forgetting in online continual learning by modeling previous task interrelations via pareto optimization. In *Forty-first international conference on machine learning*, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.33 + ], + "angle": 0, + "content": "[41] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: A novel image dataset for benchmarking machine learning algorithms. In arXiv Preprint arXiv:1708.07747, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.332, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[42] Jin Xie, Chenqing Zhu, and Songze Li. Fedmes: Personalized federated continual learning leveraging local memory. In arXiv Preprint arXiv:2404.12710, 2024. 2, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.375, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[43] He Y, Shen C, Wang X, and et al. Fppl: An efficient and non-iid robust federated continual learning framework. In 2024 IEEE International Conference on Big Data (BigData), pages 3692-3701. IEEE, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[44] Luopan Y, Han R, Zhang Q, and et al. Fedknow: Federated continual learning with signature task knowledge integration at edge. In 2023 IEEE 39th International Conference on Data Engineering (ICDE), pages 341-354. IEEE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.489, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[45] Xin Yang, Hao Yu, Xin Gao, Hao Wang, Junbo Zhang, and Tianrui Li. Federated continual learning via knowledge fusion: A survey. In IEEE Transactions on Knowledge and Data Engineering, 2024. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.546, + 0.905, + 0.599 + ], + "angle": 0, + "content": "[46] Xin Yao and Lifeng Sun. Continual local training for better initialization of federated models. In 2020 IEEE International Conference on Image Processing (ICIP), pages 1736-1740, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.602, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[47] Jaehong Yoon, Wonyong Jeong, Giwoong Lee, Eunho Yang, and Sung Ju Hwang. Federated continual learning with weighted inter-client transfer. In International Conference on Machine Learning, pages 12073-12086, 2021. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.659, + 0.905, + 0.728 + ], + "angle": 0, + "content": "[48] Hao Yu, Xin Yang, Xin Gao, Yan Kang, Hao Wang, Junbo Zhang, and Tianrui Li. Personalized federated continual learning via multi-granularity prompt. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 4023-4034, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.729, + 0.905, + 0.784 + ], + "angle": 0, + "content": "[49] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. In Advances in Neural Information Processing Systems, pages 5824-5836, 2020. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.784 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3850" + } + ] +] \ No newline at end of file diff --git a/2025/FedAGC_ Federated Continual Learning with Asymmetric Gradient Correction/69b6bc67-ec8c-4e1a-b957-09f50cc21aab_origin.pdf b/2025/FedAGC_ Federated Continual Learning with Asymmetric Gradient Correction/69b6bc67-ec8c-4e1a-b957-09f50cc21aab_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..76fa987feb7bd6608b3e900d5e456e39ca7ae269 --- /dev/null +++ b/2025/FedAGC_ Federated Continual Learning with Asymmetric Gradient Correction/69b6bc67-ec8c-4e1a-b957-09f50cc21aab_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e3f89df1bafdd18de7be96123288f43f627ce98956bf092a628446ee919b3c24 +size 837124 diff --git a/2025/FedAGC_ Federated Continual Learning with Asymmetric Gradient Correction/full.md b/2025/FedAGC_ Federated Continual Learning with Asymmetric Gradient Correction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..fc12dc670ab849338ca56879696c0fc4887125c1 --- /dev/null +++ b/2025/FedAGC_ Federated Continual Learning with Asymmetric Gradient Correction/full.md @@ -0,0 +1,362 @@ +# FedAGC: Federated Continual Learning with Asymmetric Gradient Correction + +Chengchao Zhang $^{1}$ , Fanhua Shang $^{1\dagger}$ , Hongyin Liu $^{2\dagger}$ , Liang Wan $^{1,2}$ , Wei Feng $^{1}$ , $^{1}$ College of Intelligence and Computing, Tianjin University, China $^{2}$ Medical College, Tianjin University, China + +{chengchaozhang, fhshang, hyliu2009, lwan, wfeng}@tju.edu.cn + +# Abstract + +Federated Continual Learning (FCL) has emerged as a prominent distributed learning paradigm and aims at addressing model learning challenges in both federated and continual learning settings. Efficient personalization in FCL remains a major challenge, as it must handle not only conflicts between old and new knowledge within parallel task streams but also heterogeneous knowledge conflicts from different clients. Recent approaches attempt to mitigate these issues through gradient correction. However, they often overlook the combined impact of gradient magnitude and direction, leading to unsatisfactory gradient solutions. To address these issues, we propose a novel federated continual learning method (called FedAGC) with asymmetric gradient correction, which performs memory rehearsal using representative samples selected via a centroid-based approach from historical tasks. By formulating the problem as a multi-objective optimization problem, FedAGC derives more effective gradients while incorporating group-level personalization to facilitate useful knowledge integration and irrelevant knowledge isolation, effectively mitigating both temporal and spatial catastrophic forgetting. Extensive experiments confirm the effectiveness of FedAGC. + +# 1. Introduction + +Federated Learning (FL) [21, 22, 29], as a novel distributed machine learning paradigm, aims to train a global model by leveraging independent and diverse data across clients in a privacy-preserving and collaborative manner. FL has been widely applied in various domains, such as medical image analysis [23] and object detection [24]. It has been proven to effectively address the issue of data silos [33] without compromising client privacy. Most existing FL methods are limited to single-task learning, assuming that client data remains static over time. However, in real-world scenarios, data is continuously evolving, forming sequential tasks with + +Figure 1. Spatial-Temporal Catastrophic Forgetting. (a) Direct access to new tasks leads to a significant performance decline on old tasks. (b) Non-IID data exacerbates the weight disparity. +![](images/5f89dc36fcd47dabdddc7f66f2e9a0e08331de42e8d0486e356c90553ec4ee86.jpg) +Local update of client 1 +Local update of client $N$ +Aggregation of server +$\star$ Global optimum +Server global model +Client local model +- Initial model +$\iff$ Parameter divergence +Client drift +(b) Spatial catastrophic forgetting + +unpredictable future changes. As shown in Fig. 1a, severe temporal catastrophic forgetting (TemporalCF) [38, 45] occurs when the local model on each client learns a newly introduced task, leading to performance degradation on previously learned tasks as new knowledge is acquired. + +Improving the generalizability of FL in multi-task evolving scenarios has emerged as a critical and challenging research topic. Recently, Continual Learning (CL) [17] has made significant advances in addressing TemporalCF, enabling effective mitigation of this problem. Based on such characteristics, the concept of Federated Continual Learning (FCL) [6, 46, 47] came into being, which integrates CL methods into the FL framework. In the FCL setting, FL can be regarded as a distributed instance of conventional CL, while CL methods assist in addressing FL challenges. + +FCL breaks the static limitation of traditional FL, where each client is confined to learning a single task sequence. However, integrating CL into the FL framework poses substantial challenges. Besides TemporalCF, FCL also + +faces spatial catastrophic forgetting (SpatialCF). Specifically, the private data of different clients in FCL may follow drastically different distributions. This situation, where data across clients is Non-Independent and Identically Distributed (Non-IID), which is referred to as statistical heterogeneity [26]. During server-side aggregation, where gradients from multiple clients are combined to initialize the next round of local CL training, statistical heterogeneity can lead to issues such as parameter divergence and client drift [14, 37], as illustrated in Fig. 1b. Consequently, the aggregated global model exhibits reduced generalizability, further exacerbating temporal catastrophic forgetting. + +To address the aforementioned challenges, recent studies [42, 47, 48] have explored the problem of Personalized Federated Continual Learning (PFCL). Unlike traditional FCL that aim to obtain a universal global model, PFCL leverages efficient server-side aggregation to facilitate personalized training. Existing PFCL methods mainly focus on optimizing local models for clients, ensuring strong performance on both new and previous tasks in a parallel, independent, and heterogeneous task stream across clients. The most relevant work to ours is FedMes [42], which builds on the gradient correction technique of A-GEM [2] and enhances local models through KNN-based inference. However, existing methods typically rely solely on gradient direction to determine similarity, ignoring the impact of magnitude, leading to suboptimal gradients. In contrast, we propose a new Asymmetric Gradient Correction (AGC) technique, which adaptively balances the combined impact of direction and magnitude, enabling a smoother optimization transition. + +We formulate the PFCL problem as a Multi-Objective Optimization (MOO) problem to obtain the optimal gradient, where each objective minimizes the distance to a target gradient. Specifically, to mitigate Spatial-Temporal Catastrophic Forgetting (STCF), each client extracts a small set of representative samples after completing local training for each task and employs them for rehearsal in subsequent learning. Subsequently, we correct the gradients by considering both their magnitudes and directions, ultimately obtaining an optimal gradient that does not damage any of old or new knowledge at each update from a set of class-level gradients. Additionally, empirical evidence suggests that a single global model has limited generalization across all clients [34]. Therefore, we introduce a group-level personalization strategy, where clustering is performed based on the similarity of task gradients uploaded by participating clients. These clusters are then divided into groups, enabling joint training of multiple global models. + +Our main contributions can be summarized as follows: + +- To the best of our knowledge, this is the first attempt to unify both spatial and temporal catastrophic forgetting in PFCL through a multi-objective optimization framework. +- We propose a novel asymmetric gradient correction algo + +rithm that performs memory rehearsal using representative samples selected via a centroid-based approach from historical tasks, transforming multi-task updates into an optimal gradient search problem that minimizes inter-task discrepancies and mitigates TemporalCF. + +- We propose a new group-level personalization approach that clusters clients by gradient similarity and aggregates multiple global models, isolating irrelevant knowledge interference while fostering the positive fusion of related knowledge to mitigate SpatialCF. +- Our method outperforms all the baselines in terms of accuracy and forgetting rate, demonstrating its efficacy in addressing the PFCL problem. + +# 2. Background and Related Works + +FL Algorithms for Statistical Heterogeneity. The client-server architecture-based FL paradigm is currently the most widely adopted framework. This distributed learning model involves a central server for aggregating network model parameters or gradients and $N$ clients $\{F_1, F_2, \ldots, F_N\}$ for participating in the training of local private data, where each client possesses a local private dataset $\{D_1, D_2, \ldots, D_N\}$ . For the $i$ -th client $F_i$ , its local dataset includes $n_i$ samples, denoted as $\{D_i = \{(X_k^i, Y_k^i)\}_{k=1}^{n_i}$ . Different clients may exhibit distinct features and labels, such that $X_i \cap X_j = \emptyset$ for $\forall i, j \in [1, N]$ , and $Y_k \neq \emptyset$ , $Y_i = \emptyset$ for $\forall i \in [1, N] \setminus \{k\}$ . The objective of FL is to collaboratively and privately train a global statistical model across multiple distributed clients under the supervision of a central server. The key formulation controlling FL in the classical FedAvg [29] is as follows: $\min_{w \in \mathbb{R}^d} f(w) = \frac{1}{m} \sum_{i=1}^{m} f_i(w)$ . where $f_i(w) = \mathbb{E}_{(x_i, y_i) \sim \mathcal{P}_i}[\mathcal{L}_i(w, x_i, y_i)]$ represents the local loss function of the $i$ -th client based on its local data, and $\mathcal{P}_i$ denotes the local data distribution of the $i$ -th client. To address the challenge of statistical heterogeneity and mitigate client drift caused by Non-IID data distributions, FedProx [19] extends FedAvg by introducing a proximal term that effectively constrains local updates, thereby enhancing the overall stability of model convergence. Another representative approach, SCAFFOLD [14], employs control variates to alleviate client drift. However, under extreme Non-IID conditions, these methods often struggle to produce a highly generalizable global model. Consequently, Personalized FL [20, 45] shifts its focus to providing specialized local models for each client, enabling personalized services. PerFedAvg [7] combines meta-learning techniques during a pre-training phase to obtain a well-initialized model. While these approaches have demonstrated strong performance in FL, they are primarily designed to address SpatialCF and may not be sufficient for more complex scenarios. + +Gradient Correction-Based CL Algorithms. In contrast to traditional FL designed for single-task scenarios, in a standard continual learning setting, there are a sequence + +![](images/cc19b29212009502a2e6eeab4157dd9df203fc025f7a5c96d141eb41cacd00b0.jpg) +Figure 2. The workflow of the proposed FedAGC method. Our FedAGC consists of the following three modules: On the client side, the Centroid-based Representative Sample Selection (CBRSS) module rehearses the gradients of previous tasks to assist the Asymmetric Correction Gradient Updating (ACGU) module in generating optimal gradients. On the server side, the Group Aggregation Gradient Updating (GAGU) module clusters clients based on gradient similarity and aggregates multiple global models. + +of tasks $\mathcal{T} = \{\mathcal{T}_1,\mathcal{T}_2,\dots ,\mathcal{T}_T\}$ , where $T$ is the number of tasks. For the $t$ -th task $\mathcal{T}_t$ , its dataset is composed of $n_t$ pairs of data and labels: $\{x_i^t,y_i^t\}_{i = 1}^{n_t}$ . Typically, $\mathcal{T}$ refers to a task stream where the arrival order of new tasks is unknown. As a result, the model can only access $\mathcal{T}_t$ during the training of the $t$ -th task, and access to it is no longer available afterward. Therefore, for $\mathcal{T}_t$ , given the set of model parameters $w^{(t)}$ learned so far, the learning objective at task $t$ is: $\min_{w^{(t)}}\mathcal{L}(w^{(t)};w^{(t - 1)},\mathcal{T}^t)$ . In recent years, numerous CL algorithms have adopted gradient correction strategies to mitigate TemporalCF, typically involving two key steps: Memory Rehearsal and Optimal Gradient Computation [15, 27]. A representative example is GEM [25], which retrieves gradients from previously stored memories and integrates them with the current task's gradient to obtain an update direction suitable for all tasks. As one variant, A-GEM [2], consolidates multiple memories to enhance computational efficiency. Meanwhile, ER [3] combines data from episodic memory with the current task during training. Although ER is simple and effective, its performance depends heavily on the quality of the stored memories. Another related approach is parameter regularization [5, 8], which introduces additional regularization terms to consolidate knowledge from previous tasks and reduce interference between new and old task parameters. Among these methods, EWC [38] employs the Fisher information matrix to quantify the importance of parameters for previously learned tasks. + +While FL and CL have achieved remarkable success in + +their respective domains, effectively mitigating the STCF problem in PFCL remains a challenging endeavor [12, 45]. FedCurv [38] incorporates EWC to handle statistical heterogeneity; however, it incurs high computational overhead and lacks a dedicated mechanism for addressing TemporalCF. FedKNOW [44] deploys a gradient integrator at each client to reduce negative knowledge transfer by effectively combining signature tasks identified from past and current tasks of other clients through the global model. Recent advances in vision foundation models have driven an alternative prompt-based approach [9, 11, 43, 48] that mitigates STCF by adapting to new tasks through training only a small subset of prompts and classifier parameters. + +# 3. The Proposed Method + +In this section, we elaborate on the proposed FedAGC. For the first time, FedAGC formulates the PFCL problem (see Sec. 3.1) as a MOO problem. On the client-side, after training each task, we select representative samples to construct a memory cache (see Sec. 3.2) and use them to perform asymmetric gradient correction for the subsequent new tasks (see Sec. 3.3), thus avoiding TemporalCF. On the server-side, we group the average gradient updates uploaded by the clients (see Sec. 3.4) and aggregate multiple global models to avoid SpatialCF, while also satisfying the personalized local requirements. + +The overall framework of FedAGC is shown in Figure 2, the algorithm is summarized in Algorithm 1. + +# 3.1. Personalized Federated Continual Learning + +We extend the traditional CL and FL methods to PFCL, which enables the independent and parallel integration of multiple task sequences across clients. This approach achieving both the fusion of local knowledge from new and old tasks and the integration of heterogeneous task knowledge across clients. Unlike FCL, which aims to obtain a general model that integrates global knowledge, PFCL addresses the diverse local needs of clients in highly heterogeneous scenarios, resulting in personalized local models. Specifically, each client learns its local model by optimizing the following objective: + +$$ +\min _ {w ^ {(t)}} \mathcal {L} \left(w ^ {(t)}; \mathcal {T} ^ {t}\right) + \frac {\lambda}{2} \| w ^ {(t)} - w _ {0} \| ^ {2}, \tag {1} +$$ + +where $\mathcal{L}(\cdot)$ represents the empirical risk loss, $\lambda$ is the regularization coefficient, and $w_{0}$ denotes the initial model. + +The PFCL scenario we consider, involving independent parallel multi-task sequences, comprises three key steps: + +(1) Given $N$ clients $\{\mathcal{C}_1,\mathcal{C}_2,\dots ,\mathcal{C}_N\}$ and a server $\mathcal{S}$ . There are task streams $\{\mathcal{T}^1,\mathcal{T}^2,\ldots ,\mathcal{T}^N\}$ , within which each client $\mathcal{C}_c$ engages in training a local CL model on its unique, private, and accessible task sequence $\mathcal{T}^c = \{\mathcal{T}_1^c,\mathcal{T}_2^c,\dots ,\mathcal{T}_T^c\}$ . $\mathcal{T}^c$ consists of $T$ distinct tasks, where the $t$ -th task $\mathcal{T}_t^c$ includes $K_{t}^{c}$ classes, and the total set of classes for $\mathcal{C}_c$ is represented as $K^{c} = \{K_{1}^{c},K_{2}^{c},\dots ,K_{T}^{c}\}$ . +(2) For the current task $\mathcal{T}_t^c$ being learned, the participating client receives the model $\theta_g^t$ globally broadcasted by $S$ as the initial model, where $\theta_g^t$ incorporates knowledge from the previous $t - 1$ tasks contributed by various clients. After $e$ rounds of local training on $\mathcal{T}_t^c$ , $\mathcal{C}_c$ obtains a new continual learning model $\theta_{l,c}^{t}$ which performs well on $K^c$ , and the corresponding average gradient update $G^{c}$ . +(3) In each communication round $r$ , the participating client $\mathcal{C}_c$ upload its respective $G^c$ . $S$ employs a group personalization strategy to aggregate them, which is then used to update $A$ global models $\theta_{g,A}^t$ . For the $a$ -th global model $\theta_{g,a}^t$ , it encapsulates the generalized knowledge of the associated client group's learning tasks and serves as the initial model for the next training round. + +# 3.2. Memory Rehearsal in FCL + +Benefiting from the efficient utilization of past experience, memory rehearsal methods [2, 10, 27] are widely used in CL as simple and effective forgetting suppression techniques. These methods involve storing partial data or knowledge from previous tasks to enable retraining during the learning of new tasks, thereby mitigating forgetting of prior tasks. Now, we extend it from CL to FCL. + +Memory Design. Each client utilizes its limited local memory to construct a small cache space, denoted as $M_{c}$ . Upon the completion of the training process for the $t$ -th task of $\mathcal{C}_c$ , rehearsal samples $M_t^c$ are sampled at a fixed proportion + +$\mu$ from the training set $D_{t}^{c}$ of that task. For the initial task, $M_0^c = \emptyset$ , and for the final task, no additional sampling operations are performed. Thus, $M_{c}$ can be represented as + +$$ +M _ {c} = \cup_ {t = 1} ^ {T} M _ {t} ^ {c}, \tag {2} +$$ + +where the size of $M_t^c$ is significantly smaller than $D_t^c$ . + +Centroid-Based Representative Sample Selection. Clearly, the quality of samples drawn from the old task dataset directly affects the quality of the memory cache. Therefore, we propose a Centroid-Based Representative Sample Selection (CBRSS) method to effectively acquire high-quality rehearsal samples. In particular, for client $\mathcal{C}_c$ , given the distance threshold $v$ for centroids and the input data $x$ , the constraint condition for the minimum preset distance is defined as: + +$$ +\min _ {i} \| m _ {i} ^ {c} - f _ {\theta , c} (x) \| > v, \tag {3} +$$ + +where $m_i^c$ is the $i$ -th centroid, and $f_{\theta,c}(x)$ is the predictor of $\mathcal{C}_c$ for the current task, used to map $x$ to the feature space. If the condition in Eq. (3) is satisfied, it indicates that this data point is independent and can be considered as a new centroid $m_{new,c} = f_{\theta,c}(x)$ . Otherwise, it means that this data point belongs to the existing centroid with the smallest distance to it, and then the target centroid is updated by + +$$ +m _ {i ^ {*}} = \frac {N _ {i ^ {*}} ^ {k} \times m _ {i} + f _ {\theta , c} (x)}{N _ {i ^ {*}} ^ {k} + 1}, \tag {4} +$$ + +where $i^{*} = \underset {i}{\arg \min}\| f_{\theta ,c}(x) - m_{i}^{c}\|$ , and $N_{i^*}^k$ represents the total number of samples in the $k$ -th category captured by the centroid closest to $x$ in the current task. + +Overall, each client stores the representative data of each category for the trained task after completing the training. A similar approach is employed by FedMes [42], which utilizes a random sampling method to construct the memory cache. However, due to the small sample size, this approach inevitably results in significant sampling bias. The resulting $M_{t}^{c}$ lacks representativeness and fails to reflect the overall distribution of $D_{t}^{c}$ . This leads to a drift in the representation of old task data in the feature space, which in turn triggers the problem of catastrophic forgetting. + +# 3.3. Asymmetrically Correcting Gradient Updates + +To integrate the knowledge from previous tasks with that of the current task during the learning process of the $t$ -th task on $\mathcal{C}_c$ , we reformulate the training problem as a MOO problem, which is formally described as follows: + +$$ +\min _ {\theta_ {c}, \theta_ {c} ^ {1: t}} \ell \left(f _ {\theta_ {c}}, f _ {\theta_ {c} ^ {t}}, D _ {t} ^ {c}\right) \cup \ell \left(f _ {\theta_ {c}}, f _ {\theta_ {c} ^ {k}}, M _ {k} ^ {c}\right) \mid \forall k < t. \tag {5} +$$ + +Here, $\ell(\cdot)$ denotes the loss function, and $\theta_c$ and $\theta_c^t$ represent the backbone model for shared knowledge across multiple tasks and the task-specific classifier for $\mathcal{C}_c$ , respectively. + +Since the parameters of the new and old task classifiers are independent and can be updated directly using their respective gradients $\theta_c^k\gets \theta_c^k -\alpha \nabla \ell (f_{\theta_c},f_{\theta_c^k})$ with step size $\alpha$ our focus is on updating the shared backbone model, which is updated using the gradients $g_{k}^{c} = \nabla \ell (f_{\theta_{c}},f_{\theta_{c}^{k}})$ of all $t$ tasks. Let $\beta$ be the step size for optimization. The problem of the backbone model update can be formulated as follows: + +$$ +\theta_ {c} \leftarrow \theta_ {c} - \beta d _ {c} ^ {*}, d _ {c} ^ {*} = \arg \min _ {d _ {c}} D \left(g _ {k} ^ {c}, d _ {c}\right) \mid \forall k \leq t, \tag {6} +$$ + +where $d_c^*$ represents the optimal gradient, which approximates the gradient of the new task while avoiding conflicts with the gradient of the old task. In Eq. (6), we define it as a min-distance multi-objective problem. Therefore, the critical issue lies in determining the method used to compute the differences between gradients, ensuring that the solution for $d_c^*$ is as close as possible to all gradients. + +Symmetric distance metrics, such as Euclidean distance and cosine distance, are widely applied in deep neural networks. These metrics share a fundamental property: the forward impact (from $x$ to $y$ ) is equal to the backward impact (from $y$ to $x$ ). However, local model updates are highly dependent on the magnitude and direction of gradients, both of which are inherently asymmetric. In brief, given an old task gradient $g_{old}$ , trained on $M_c^k$ , and a new task gradient $g_{new}$ , trained on $D_t^c$ , their mutual influences on each other's tasks are significantly different. In previous studies [2, 42, 49], the condition $\langle g_{old}, g_{new} \rangle \geq 0$ has been employed to signify the absence of conflicts between two tasks. However, given that the old task has been integrated earlier, $\| g_{old} \|$ is typically smaller while $\| g_{new} \|$ tends to be larger. In cases where $\| g_{old} \| \ll \| g_{new} \|$ , even if $\langle g_{old}, g_{new} \rangle \geq 0$ holds true, the updates pertaining to the new task may have a substantial impact on the performance of the old task. + +Definition 1 (Asymmetric Gradient Distance). Given two gradients $g_{A}$ and $g_{B}$ , the asymmetric gradient distance is defined as + +$$ +D \left(g _ {A}, g _ {B}\right) = \frac {\left\| g _ {A} - g _ {B} \right\|}{\left\| g _ {B} \right\| + \left\| g _ {A} - g _ {B} \right\|}. +$$ + +In this paper, an asymmetric distance calculation method is used to explore the measure of difference between the new task and the old task. Definition 1 considers both the gradient direction and magnitude. Notably, $D(g_{A}, g_{B}) = 0$ when $g_{A} = g_{B} = 0$ . + +However, solving Eq. (6) directly for the parallel task flow on clients presents significant challenges. This difficulty arises from the fact that both the decision space of this problem and the model parameters are identically high-dimensional. Consequently, following [28, 35], we employ linear scalarization to determine a weight variable $w^{*}$ . Specifically, when client $C_c$ is training on the $t$ -th task, the + +# Algorithm 1 FedAGC + +Input: $N$ clients with their private parallel task sequence $\mathcal{T}^{1:N}$ and local datasets $\{D_t^{1:N}\}_{t=1}^T$ , global model parameter $\theta_g$ , communication rounds $R$ , memory space $M_{1:c} \gets \{\}$ + +Output: Multiple global parameters $\theta_{g}^{i:n}$ , local parameter $\theta_{l,1:N}$ and classification head $\theta_{l,1:N}^{k}$ + +1: Initialization +2: while task number $t \leq T$ do +3: for each round $r = 1,2,\ldots ,R$ do +4: Server broadcasts $\theta_g^i$ to each group of clients; +5: for each client $c = 1$ to $N$ in parallel do +6: Minimizing local CL problems by Eq. (1); +7: $g^{c} = \nabla \ell (f_{\theta_{c}},f_{\theta^{t}},D_{t}^{c});$ +8: $g_{k}^{c} = \nabla \ell (f_{\theta_{c}},f_{\theta_{k}^{k}},M_{k}^{c})$ for all $k < t$ +9: $d_c^* \gets \mathsf{ACGU}(g^c, g_1^c, \ldots, g_{t-1}^k)$ ; +10: $\theta_{l,c}^{t}\gets \theta_{l,c}^{t} - \alpha d_{c}^{*};$ +11: end for +12: Send the $G^{c} = -\triangle \theta_{l,c}^{t} / \eta$ to the central server; +13: $\{S_1,S_2,\ldots ,S_n\} \gets \mathrm{GAGU}(G^1,G^2,\ldots ,G^N);$ +14: $\theta_{g}^{i}\gets \theta_{g}^{i} - \frac{1}{|S_{i}|}\sum_{j\in S_{i}}G^{j}$ for $i\in \{1,2,\ldots ,n\}$ +15: end for +16: Each client samples $M_t^c$ from $D_t^c$ by Eqs. (3) and (4); +17: end while + +gradient update $g_{t}^{c}$ of the new task and $g_{1:t-1}^{c}$ from the old tasks and computed through + +$$ +w ^ {*} = \arg \min _ {w} \left\{\frac {1}{t} \sum_ {i} ^ {t} D \left(\sum_ {j \neq i} ^ {t} w _ {j} g _ {j} ^ {c}, g _ {i} ^ {c}\right) \right\}, \tag {7} +$$ + +where $\sum_{j=1}^{t} w_{j} = 1$ , $\forall w_{j} \geq 0$ , and $D(\sum_{j \neq i}^{t} w_{j} g_{j}^{c}, g_{i}^{c})$ represents the divergence between task $i$ and the combination of all other tasks. By minimizing this divergence, we achieve the suppression of the influence from the combination on task $i$ . Note that this optimization process does not modify individual gradient updates. + +To prevent training failure on new tasks, we introduce the Multiple Gradient Descent Algorithm (MGDA) [4] to minimize the gradient norm, thereby achieving a balanced solution. It is defined as + +$$ +w _ {b} ^ {*} = \underset {w} {\arg \min } \left\| \sum_ {i} ^ {t} w _ {i} g _ {i} ^ {c} \right\| ^ {2}. \tag {8} +$$ + +When the objective function value of MGDA is zero, the obtained solution satisfies the Karush-Kuhn-Tucker conditions. Otherwise, the solution provides a Pareto descent direction that does not harm any task. + +In summary, our proposed Asymmetric Correction Gradient Updating (ACGU) algorithm determines $d_c^*$ by + +$$ +d _ {c} ^ {*} = \sum_ {i} W _ {i} g _ {i} ^ {c}, \quad \text {w h e r e} W = \gamma w ^ {*} + (1 - \gamma) w _ {b} ^ {*}. \tag {9} +$$ + +
AlgorithmAttributes FCL Per GCBTask ID under Non-IID Setting (Across 8 Different Datasets)\( A_{avg} \)(%)
\( A_1 \)(%)\( A_2 \)(%)\( A_3 \)(%)\( A_4 \)(%)\( A_5 \)(%)
FedAvg× × ×54.73 ± 1.4856.06 ± 3.1663.99 ± 3.4357.90 ± 2.9678.49 ± 2.1861.83 ± 0.46
FedProx× × ×58.25 ± 1.6056.65 ± 0.5066.15 ± 2.3361.72 ± 2.2182.24 ± 1.0265.00 ± 0.62
Ditto× ✓ ×58.92 ± 1.4059.13 ± 1.6370.58 ± 3.0763.00 ± 2.7983.97 ± 0.9567.12 ± 0.85
FedCurv✓ × ✓60.18 ± 3.4354.88 ± 1.9968.05 ± 3.7859.25 ± 2.3782.19 ± 0.5864.91 ± 0.56
FedAGEM✓ × ✓69.62 ± 0.8566.91 ± 1.0975.06 ± 1.6766.73 ± 2.4781.73 ± 1.1672.01 ± 0.81
FedER✓ × ×73.34 ± 0.4568.77 ± 1.8980.18 ± 1.3970.19 ± 1.3079.43 ± 1.3274.38 ± 0.27
FedMes✓ ✓ ✓70.84 ± 0.7967.22 ± 1.0577.25 ± 0.8967.18 ± 1.5285.21 ± 0.3673.54 ± 0.55
Ours-w/o-ACGU× ✓ ×58.62 ± 0.3658.06 ± 0.1665.62 ± 0.1963.29 ± 1.8682.46 ± 1.6865.61 ± 0.80
Ours-w/o-GAGU✓ × ✓76.19 ± 0.5271.59 ± 1.1282.24 ± 1.4272.89 ± 2.0481.82 ± 2.3676.95 ± 1.38
Ours-w/o-CBRSS✓ ✓ ✓75.40 ± 0.9072.39 ± 0.5582.88 ± 0.9275.42 ± 0.5486.68 ± 0.6278.56 ± 0.57
FedAGC (Ours)✓ ✓ ✓76.21 ± 0.3672.56 ± 0.4884.90 ± 0.8577.60 ± 0.5787.28 ± 0.1879.71 ± 0.23
+ +Table 1. Performance (avg ± std) of local models for each task on 10 clients under Non-IID setting. "w/o-Module" indicates performance without the corresponding module. It also includes Per (Personalized) and GCB (Gradient Correction-Based). For the proposed FedAGC, we used $\lambda = 0.01$ and $\gamma = 0.5$ in Eqs. (1) and (9). + +# 3.4. Group-Level Personalization + +To integrate heterogeneous knowledge from different clients, some studies [39, 45] adopt personalization strategies at the client level, where each client continuously adjusts the received global model to better fit its local tasks. However, this approach struggles to meet the needs of all clients in large-scale parallel task scenarios. In contrast, we consider a group-level personalization strategy, enabling the broadcast of global models tailored to meet the specific requirements of each client. + +In this work, we propose Group Aggregation Gradient Updating (GAGU) for client-uploaded gradients. Specifically, during the $r$ -th communication round, $N$ clients upload their respective averaged gradient updates $\{G^1,G^2,\dots ,G^N\}$ obtained after $e$ local training iterations, where $G^{c} = -\triangle \theta_{l,c}^{t} / \alpha$ . We compute a similarity matrix $S$ to measure the similarity between pairs of averaged gradient updates. The elements of $S$ are given by + +$$ +s _ {i, j} = \frac {G ^ {i} G ^ {j}}{\| G ^ {i} \| \| G ^ {j} \|.} \tag {10} +$$ + +We construct a graph $P = \{G, E\}$ to represent $A$ , where $G$ denotes the set of nodes in the graph, and each node corresponds to the $G^c$ . The set of edges $E = \{s_{i,j}\} \forall i \leq j \leq N$ represents the similarity between $G^i$ and $G^j$ . The objective of GAGU is to reorder $P$ such that similar average gradient updates are grouped together, while dissimilar ones are isolated. To achieve this, we first define the similarity between any two groups $S_A$ and $S_B$ ( $S_A \cap S_B = \emptyset$ , $\forall A \neq B$ ): + +$$ +h \left(S _ {A}, S _ {B}\right) = \sum_ {i \in S _ {A}, j \in S _ {B}} s _ {i, j}. \tag {11} +$$ + +GAGU aims to maximize intra-group similarity while minimizing inter-group similarity, resulting in an optimal graph partition $a^{*} = \{S_{1}, S_{2}, \ldots, S_{n}\}$ , where $n$ represents the + +number of groups in $a^*$ . Inspired by spectrum-clustering [13] and Ncut [36], We reformulate this optimization problem into the following form: + +$$ +\min _ {a \in A} \sum_ {S, S ^ {\prime} \in a} h \left(S, S ^ {\prime}\right) - h \left(S ^ {\prime}, S ^ {\prime}\right), \tag {12} +$$ + +where $A$ is all possible grouping results. Additionally, we introduce the modularity $Q$ to evaluate the quality of the current partition during the process of determining the optimal number of groups and partitioning results. This metric measures the distribution of connectivity within and between groups. + +$$ +Q = \frac {1}{2 W} \sum_ {i, j} \left[ s _ {i, j} - \frac {d _ {i} d _ {j}}{2 W} \right] \delta \left(S _ {i}, S _ {j}\right), \tag {13} +$$ + +where $W = \sum_{i,j} s_{i,j}$ represents the total edge weight in the current graph, $d_i = \sum_j s_{i,j}$ denotes the degree of node $i$ , and $\delta(S_i, S_j)$ is the Kronecker delta function. + +# 4. Experiments + +# 4.1. Experimental Setup + +Benchmarks. We validate FedAGC in large-scale multitask parallel FCL scenarios. The experimental setup follows the Task-Incremental Learning paradigm and primarily includes two different benchmarks: (1) IID Setting (Single CIFAR-100): The 100 classes of CIFAR-100 [16] are divided into 20 tasks, each containing 5 classes. These tasks are randomly assigned to clients, ensuring that each client does not train on the same task multiple times. (2) Non-IID Setting (Across 8 Different Datasets): We partition and distribute eight widely used datasets in the FCL domain, such as CIFAR-10/100 [16], MNIST [18], SVHN [30], FashionMNIST [41], Not-MNIST [1], FaceScrub [32], and TrafficSigns [31], in the same manner as (1). In both benchmarks, each client is assigned a unique task sequence. + +
AlgorithmAttributes FCL Per GCBTask ID under IID Setting (Single CIFAR-100)\( A_{avg} \)(%)
\( A_1 \)(%)\( A_2 \)(%)\( A_3 \)(%)\( A_4 \)(%)\( A_5 \)(%)
FedAvg× × ×54.32 ± 0.9460.53 ± 2.0861.89 ± 1.5665.63 ± 1.6070.77 ± 0.8462.63 ± 0.55
FedProx× × ×53.96 ± 0.9861.53 ± 2.2362.90 ± 1.9465.70 ± 1.1370.78 ± 0.9162.97 ± 0.93
Ditto× ✓ ×55.48 ± 0.9262.81 ± 1.9663.12 ± 1.5866.59 ± 1.8671.01 ± 0.9363.67 ± 0.72
FedCurv✓ × ✓54.36 ± 0.9560.77 ± 0.8663.43 ± 0.9166.46 ± 0.9771.52 ± 0.7563.31 ± 0.16
FedAGEM✓ × ✓60.95 ± 0.9067.15 ± 1.2366.39 ± 0.6967.81 ± 0.9071.28 ± 0.6666.72 ± 0.23
FedER✓ × ×63.05 ± 0.7667.83 ± 0.9166.11 ± 0.3265.39 ± 0.4468.84 ± 1.0166.44 ± 0.42
FedMes✓ ✓ ✓61.18 ± 0.3467.68 ± 0.1167.11 ± 0.4967.42 ± 0.1572.12 ± 0.9067.13 ± 0.16
Ours-w/o-ACGU× ✓ ×55.45 ± 0.2761.95 ± 0.8763.32 ± 1.7265.51 ± 0.7371.44 ± 0.6063.53 ± 0.31
Ours-w/o-GAGU✓ × ✓63.92 ± 0.5566.73 ± 0.9367.22 ± 0.6166.48 ± 0.6568.53 ± 0.2466.58 ± 0.41
Ours-w/o-CBRSS✓ ✓ ✓66.14 ± 0.3769.05 ± 0.8069.69 ± 0.4969.56 ± 2.2074.12 ± 0.0969.75 ± 0.78
FedAGC (Ours)✓ ✓ ✓66.38 ± 0.4869.48 ± 0.7870.38 ± 1.3270.58 ± 1.8573.98 ± 0.4170.16 ± 0.77
+ +Table 2. Performance (avg ± std) of local models for each task on 10 clients under IID setting. + +![](images/065edf3297a1e284be73cca5f916af60ca88546822aaf1d5946914a4ef13402a.jpg) +(a) $F_{t}$ under Non-IID Setting +Figure 3. Comparison of average forgetting rates. + +![](images/4b5ee6f8a9295d9655513e46cba23c8834e4f81d409c631b3361571035ae705c.jpg) +(b) $F_{t}$ under IID Setting + +Baselines. We compare FedAGC with state-of-the-art (SOTA) methods, including the fundamental FedAvg [29] and its proximal variant FedProx [19], along with the personalized FL method Ditto [20]. To verify the superiority of FedAGC, we also compare it with gradient correction-based FCL methods, including FedCurv [38], FedAGEM [2] and the latest FedMes [42]. Additionally, we consider the rehearsal-based method FedER [3]. + +Metrics. We employ two critical metrics to evaluate algorithm performance: Average Accuracy $(A_{avg})$ and Forgetting Rate $(F_t)$ . Let $a_{i,t}^c$ denote the test accuracy of the $i$ -th task learned by the $c$ -th client after sequential learning of $t$ subsequent tasks, where $t + i \leq T$ . In this formulation, the initial accuracy of task $i$ (immediately after its own training) and the final accuracy after training on all tasks are defined as $a_{i,0}^c$ and $a_{i,T - i}^c$ , respectively. The forgetting rate for the $i$ -th task after learning $t$ subsequent tasks is calculated as: $f_{i,t}^{c} = \frac{a_{i,0}^{c} - a_{i,t}^{c}}{a_{i,0}^{c}}$ . To aggregate results, we compute client-wise averages: $A_i = \frac{1}{N}\sum_{c=1}^N a_{i,T-i}^c$ and $f_{i,t} = \frac{1}{N}\sum_{c=1}^N f_{i,t}^c$ . Then, $A_{avg}$ and $F_t$ are defined as: + +$$ +A _ {\text {a v g}} = \frac {1}{T} \sum_ {i = 1} ^ {T} A _ {i}, \quad F _ {t} = \frac {1}{T - t} \sum_ {i = 1} ^ {T - t} f _ {i, t}. \tag {14} +$$ + +Implementation details. Following the experimental setup + +![](images/ed150f3019ad6825d5ac902e367243c47790fbf8c4a456df9789eb7fb1dfa2b9.jpg) + +![](images/d11c41af08c1f352b50d3273bdb02d13ba31314a436c9be0da172a1e46db85a3.jpg) +(a) Impact of $\gamma$ on FedAGC + +![](images/86e4413525e67630fc5d1bf946e0fd9c18c4829f2f90681d0e32be8ce15f94fe.jpg) + +![](images/aaa4303fe61e51880f01edfafa2d12c9ac4e94b2a13e4fc9863b12a8349fca2e.jpg) +(b) Impact of $\lambda$ on FedAGC + +![](images/225cb13e94af64123351d55667487139ec065af95c06600e19b7d0bd165bdfdb.jpg) +(c) Non-IID Setting, $\lambda = 0.01$ +(e) IID Setting, $\lambda = 0.01$ +Figure 4. Impact of core hyperparameters on FedAGC. + +![](images/6a9898dc138632433d22d7fcc7f82a0fe26f29ba033401bff1c58d18abfc4297.jpg) +(d) Non-IID Setting, $\gamma = 0.5$ +(f) IID Setting, $\gamma = 0.5$ + +of [47], we adopted a variant of the LeNet architecture as the backbone. For any newly introduced task, we initialized a new classifier. The Adam optimizer was used for all datasets, with an initial learning rate of $10^{-3}$ and a weight decay of $10^{-4}$ . The experimental setup consisted of 10 clients, each learning 5 sequential and non-overlapping tasks. The batch size was set to 100 per step, with $\mu$ and $v$ set to 0.1 and 8, respectively. To ensure convergence, all experiments were conducted for 100 federated communication rounds. Each experiment was repeated five times, and the reported results were averaged accordingly. + +# 4.2. Main Results + +Tabs. 1 and 2 record the average accuracy of 10 clients on each task under two data condition setups, along with the + +final results considering all tasks. In Fig. 3, we compare the degree of Spatial-TemporalCF between FedAGC and various methods. It can be observed that FedProx and Ditto mitigate the impact of statistical heterogeneity using penalty terms, showing significant performance improvements over FedAvg in Non-IID settings. However, since they do not employ any CL strategies, their accuracy is lower compared to FCL methods such as FedAGEM, FedER, and FedMes, with higher forgetting levels. Notably, FedCurv's experimental results are suboptimal as it leverages curvature information from other clients to guide updates for the current client, which is prone to knowledge conflicts in cases of non-overlapping client data distributions. These FCL methods based on symmetric gradient correction focus solely on TemporalCF without considering the interference of knowledge across clients, leading to SpatialCF. In contrast, the proposed FedAGC outperforms other methods in both task-specific accuracy and overall performance across both settings, with a lower degree of forgetting. This demonstrates that memory buffers constructed through importance sampling, assisted by asymmetric gradient correction and group-level personalization, can significantly improve the accuracy of PFCL models. + +# 4.3. Performance Analysis of FedAGC + +Ablation Studies. We conducted three different ablation studies under the same experimental setup, with the results recorded in Tabs. 1 and 2, as well as Fig. 3. We progressively removed components from FedAGC, including: (1) Omitting the ACGU module (FedAGC-w/o-ACGU), which means that only the gradients from local training are used to guide updates. Clearly, in this case, the CBRSS module also loses its significance. The performance degradation is significant, with a drop of $17.7\%$ and $9.5\%$ , as no strategy for mitigating TemporalCF is applied. (2) Removing the GAGU module (FedAGC-w/o-GAGU) and switching to the FedAvg average aggregation method. In the IID setting, FedAGC-w/o-GAGU's overall performance is overtaken by FedMes by $0.86\%$ . (3) Disabling the CBRSS module (FedAGC-w/o-CBRSS) and replacing it with a random sampling method. In this case, performance decreases by $1.44\%$ and $0.58\%$ in the Non-IID and IID settings, respectively. However, since the core component for mitigating forgetting is retained, its accuracy on individual tasks and overall performance remains superior to other baselines. These results verify that our three proposed modules of FedAGC play a crucial role in alleviating Spatial-TemporalCF and achieving personalization, and can effectively be combined to enhance performance. + +Impact of core hyperparameters. We evaluate FedAGC with different values of $\gamma$ and $\lambda$ . As shown in Fig. 4, setting $\lambda$ too large or too small causes FedAGC to overly rely on either special or balanced solution, impairing its ability to + +![](images/8afb4c9a20a3e080b95b48912d0f19e7fb31dbbf3440484387aa96ea12bdf009.jpg) +(a) $A_{i}$ on the CIFAR-100 dataset, $N = 30$ + +![](images/cf5e28fab530917c16438c069921c6bfdc0f579e4ca5820b5f39f1ebaecc235c.jpg) +(b) $F_{t}$ on the CIFAR-100 dataset, $N = 30$ +Figure 5. Performance on large-scale clients. + +mitigate forgetting. Similarly, an excessively large $\gamma$ biases the model toward the inversion of previous tasks, hindering the learning of current task, while a small $\gamma$ has the opposite effect. Empirically, we find that the best performance of FedAGC is achieved when $\gamma = 0.5$ and $\lambda = 0.01$ . + +Performance on large-scale clients. To explore the scalability of FedAGC to larger client scenarios, we tripled the number of clients from the previous setting of 10 and continued comparing it with various SOTA methods. The results recorded in Fig. 5 indicate that FedAGC outperforms other baselines in accuracy on previous tasks and demonstrates greater stability in mitigating forgetting. This result highlights the superiority of FedAGC. + +# 5. Conclusion + +To tackle the challenge of STCF in PFCL, this paper proposed FedAGC as a valid solution. After local training, FedAGC selects representative samples and stores them in a buffer for rehearsal to alleviate forgetting. To refine this process, we introduced an asymmetric gradient metric to quantify discrepancies between past and current tasks. Asymmetric gradient correction is then employed to adjust the gradient updates, jointly considering direction and magnitude for optimal adaptation. Moreover, FedAGC adopts group-level personalization by clustering clients based on the similarity of their uploaded averaged gradient updates, which enhances the fusion of valuable knowledge while isolating ineffective knowledge. It was extensively demonstrated in this paper that the effective integration of the proposed modules can enhance the model's resistance to STCF, enabling FedAGC to outperform SOTA approaches. + +# Acknowledge + +We want to thank the anonymous reviewers for their valuable suggestions and comments. This work was supported by the National Natural Science Foundation of China (No. 62276182), Tianjin University under the Emerging Direction Cultivation Project of Interdisciplinary Center (Intelligent Protection and Utilization of Digital Cultural Heritage), National Key Research and Development Program of China (No. 2023YFF0906204), Peng Cheng Lab Program (No. PCL2023A08), Tianjin Natural Science Foundation (Nos. 24JCYBJC01230, 24JCYBJC01460), and Tianjin Municipal Education Commission Research Plan (No. 2024ZX008). + +# References + +[1] Yaroslav Bulatov. Notmnist dataset. http://yaroslavvb.blogspot.it/2011/09/notmnist-dataset.html.6 +[2] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a-gem. In Proceedings of the International Conference on Learning Representations, 2018. 2, 3, 4, 5, 7 +[3] Arslan Chaudhry, Marcus Rohrbach, Mohamed Elhoseiny, Thalaiyasingam Ajanthan, Puneet K. Dokania, Philip H. S. Torr, and MarcÁurelio Ranzato. On tiny episodic memories in continual learning. In arXiv Preprint arXiv:1902.10486, 2019. 3, 7 +[4] Désideri and Jean-Antoine. Multiple-gradient descent algorithm (mgda) for multiobjective optimization. Comptes Rendus Mathematique, 350(5-6):313-318, 2012. 5 +[5] Prithviraj Dhar, Rajat Vikram Singh, Kuan-Chuan Peng, Ziyan Wu, and Rama Chellappa. Learning without memorizing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5138-5146, 2019. 3 +[6] Jiahua Dong, Lixu Wang, Zhen Fang, Gan Sun, Shichao Xu, Xiao Wang, and Qi Zhu. Federated class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10164-10173, 2022. 1 +[7] Alireza Fallah, Aryan Mokhtari, and Asuman Ozdaglar. Personalized federated learning with theoretical guarantees: A model-agnostic meta-learning approach. Advances in neural information processing systems, 33:3557–3568, 2020. 2 +[8] Mehrdad Farajtabar, Navig Azizan, Alex Mott, and Ang Li. Orthogonal gradient descent for continual learning. In International Conference on Artificial Intelligence and Statistics, PMLR, pages 3762-3773, 2020. 3 +[9] Bagwe G, Yuan X, Pan M, and et al. Fed-cprompt: Contrastive prompt for rehearsal-free federated continual learning. arXiv preprint arXiv:2307.04869, 2023. 3 +[10] Yunhui Guo, Mingrui Liu, Tianbao Yang, and Tajana Rosing. Improved schemes for episodic memory-based lifelong learning. Advances in Neural Information Processing Systems, 33:1023-1035, 2020. 4 + +[11] Piao H, Wu Y, Wu D, and et al. Federated continual learning via prompt-based dual knowledge transfer. In *Forty-First International Conference on Machine Learning (ICML)*. PMLR, 2024. 3 +[12] Parisa Hamedi, Roozbeh Razavi-Far, and Ehsan Hallaji. Federated continual learning: Concepts, challenges, and solutions. Neurocomputing, 651:130844, 2025. 3 +[13] Hongjie Jia, Shifei Ding, Xinzheng Xu, and Ru Nie. The latest research progress on spectral clustering. Neural Computing and Applications, 24:1477-1486, 2014. 6 +[14] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank J.Reddi, Sebastian U.Stich, and Ananda Theertha Suresh. Scaffold: Stochastic controlled averaging for federated learning. In International conference on machine learning. PMLR, pages 5132-5143, 2020. 2 +[15] Alexander Krawczyk and Alexander Gepperth. An analysis of best-practice strategies for replay and rehearsal in continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4196-4204, 2024. 3 +[16] Alex Krizhevsky and Geoffrey Hinton. Learning Multiple Layers of Features from Tiny Images. 2009. 6 +[17] Matthias De Lange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Aleš Leonardis, Gregory Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying forgetting in classification tasks. In IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 3366-3385, 2021. 1 +[18] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 6 +[19] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. In Proceedings of Machine Learning and Systems, pages 429-450, 2020. 2, 7 +[20] Tian Li, Shengyuan Hu, Ahmad Beirami, and Virginia Smith. Ditto: Fair and robust federated learning through personalization. In International Conference on Machine Learning, PMLR, pages 6357-6368, 2021. 2, 7 +[21] Junkang Liu, Yuanyuan Liu, Fanhua Shang, Hongying Liu, Jin Liu, and Wei Feng. Improving generalization in federated learning with highly heterogeneous data via momentum-based stochastic controlled weight averaging. In Forty-second International Conference on Machine Learning. 1 +[22] Junkang Liu, Fanhua Shang, Yuanyuan Liu, Hongying Liu, Yuangang Li, and YunXiang Gong. Fedbcgd: Communication-efficient accelerated block coordinate gradient descent for federated learning. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 2955-2963, 2024. 1 +[23] Quande Liu, Cheng Chen, Jing Qin, Qi Dou, and Pheng-Ann Heng. Feddg: Federated domain generalization on medical image segmentation via episodic learning in continuous frequency space. In CVPR, pages 1013–1023, 2021. 1 +[24] Yang Liu, Anbu Huang, Yun Luo, He Huang, Youzhi Liu, Yuanyuan Chen, Lican Feng, Tianjian Chen, Han Yu, and + +Qiang Yang. Fedvision: An online visual object detection platform powered by federated learning. In AAAI, pages 13172-13179, 2020. 1 +[25] David Lopez-Paz and MarcÁurelio Ranzato. Gradient episodic memory for continual learning. In Advances in neural information processing systems, 2017. 3 +[26] Bing Luo, Wenli Xiao, Shiqiang Wang, Jianwei Huang, and Leandros Tassiulas. Tackling system and statistical heterogeneity for federated learning with adaptive client sampling. In IEEE INFOCOM 2022 - IEEE Conference on Computer Communications, pages 1739-1748, 2022. 2 +[27] Fan Lyu, Shuai Wang, Wei Feng, Zihan Ye, Fuyuan Hu, and Song Wang. Multi-domain multi-task rehearsal for lifelong learning. Proceedings of the AAAI Conference on Artificial Intelligence, 35(10):8819-8827, 2021. 3, 4 +[28] Fan Lyu, Qing Sun, Fanhua Shang, Liang Wan, and Wei Feng. Measuring asymmetric gradient discrepancy in parallel continual learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11411-11420, 2023. 5 +[29] H. Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial Intelligence and Statistics, PMLR, pages 1273-1282, 2017. 1, 2, 7 +[30] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading digits in natural images with unsupervised feature learning. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning, page 4, 2011. 6 +[31] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading digits in natural images with unsupervised feature learning. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning, page 4, 2011. 6 +[32] Hong-Wei Ng and Stefan Winkler. A data-driven approach to cleaning large face datasets. In Proceedings of the 2014 IEEE International Conference on Image Processing (ICIP), pages 343–347. IEEE, 2014. 6 +[33] Li Qinbin, Diao Yiqun, Chen Quan, and He Bingsheng. Federated learning on non-iid data silos: An experimental study. In 2022 IEEE 38th International Conference on Data Engineering (ICDE), pages 965-978, 2022. 1 +[34] Felix Sattler, Simon Wiedemann, Klaus-Robert Müller, and Wojciech Samek. Robust and communication-efficient federated learning from non-iid data. IEEE Transactions on Neural Networks and Learning Systems, 31(9):3400-3413, 2019. 2 +[35] Ozan Sener and Vladlen Koltun. Multi-task learning as multi-objective optimization. In Advances in Neural Information Processing Systems, 2018. 5 +[36] Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 22(8):888-905, 2000. 6 +[37] Yong Shi, Yuanying Zhang, Yang Xiao, and Lingfeng Niu. Optimization strategies for client drift in federated learning: A review. Procedia Computer Science, 214:1168-1173, 2022. 2 + +[38] Neta Shoham, Tomer Avidor, Aviv Keren, Nadav Israe, Daniel Benditkis, Liron Mor-Yosef, and Itai Zeitak. Overcoming forgetting in federated learning on non-iid data. In arXiv Preprint arXiv:1910.07796, 2019. 1, 3, 7 +[39] Hongda Wu and Ping Wang. Fast-convergent federated learning with adaptive weighting. IEEE Transactions on Cognitive Communications and Networking, 7(4):1078-1088, 2021. 6 +[40] Yichen Wu, Hong Wang, Peilin Zhao, Yefeng Zheng, Ying Wei, and Long-Kai Huang. Mitigating catastrophic forgetting in online continual learning by modeling previous task interrelations via pareto optimization. In *Forty-first international conference on machine learning*, 2024. +[41] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: A novel image dataset for benchmarking machine learning algorithms. In arXiv Preprint arXiv:1708.07747, 2017. 6 +[42] Jin Xie, Chenqing Zhu, and Songze Li. Fedmes: Personalized federated continual learning leveraging local memory. In arXiv Preprint arXiv:2404.12710, 2024. 2, 4, 5, 7 +[43] He Y, Shen C, Wang X, and et al. Fppl: An efficient and non-iid robust federated continual learning framework. In 2024 IEEE International Conference on Big Data (BigData), pages 3692-3701. IEEE, 2024. 3 +[44] Luopan Y, Han R, Zhang Q, and et al. Fedknow: Federated continual learning with signature task knowledge integration at edge. In 2023 IEEE 39th International Conference on Data Engineering (ICDE), pages 341-354. IEEE, 2023. 3 +[45] Xin Yang, Hao Yu, Xin Gao, Hao Wang, Junbo Zhang, and Tianrui Li. Federated continual learning via knowledge fusion: A survey. In IEEE Transactions on Knowledge and Data Engineering, 2024. 1, 2, 3, 6 +[46] Xin Yao and Lifeng Sun. Continual local training for better initialization of federated models. In 2020 IEEE International Conference on Image Processing (ICIP), pages 1736-1740, 2020. 1 +[47] Jaehong Yoon, Wonyong Jeong, Giwoong Lee, Eunho Yang, and Sung Ju Hwang. Federated continual learning with weighted inter-client transfer. In International Conference on Machine Learning, pages 12073-12086, 2021. 1, 2, 7 +[48] Hao Yu, Xin Yang, Xin Gao, Yan Kang, Hao Wang, Junbo Zhang, and Tianrui Li. Personalized federated continual learning via multi-granularity prompt. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 4023-4034, 2024. 2, 3 +[49] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. In Advances in Neural Information Processing Systems, pages 5824-5836, 2020. 5 \ No newline at end of file diff --git a/2025/FedAGC_ Federated Continual Learning with Asymmetric Gradient Correction/images.zip b/2025/FedAGC_ Federated Continual Learning with Asymmetric Gradient Correction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c66014a0790c1052762cf9f670bc0c619e3438a6 --- /dev/null +++ b/2025/FedAGC_ Federated Continual Learning with Asymmetric Gradient Correction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:20ddd5f8625bae50ea8505bc90ac6decd1a0def261fbb8701c0989d0bc7bfa45 +size 692518 diff --git a/2025/FedAGC_ Federated Continual Learning with Asymmetric Gradient Correction/layout.json b/2025/FedAGC_ Federated Continual Learning with Asymmetric Gradient Correction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b1fe957b21073349ad8564b306b6ff6bcee04e09 --- /dev/null +++ b/2025/FedAGC_ Federated Continual Learning with Asymmetric Gradient Correction/layout.json @@ -0,0 +1,12488 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 59, + 103, + 550, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 103, + 550, + 120 + ], + "spans": [ + { + "bbox": [ + 59, + 103, + 550, + 120 + ], + "type": "text", + "content": "FedAGC: Federated Continual Learning with Asymmetric Gradient Correction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 142, + 494, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 142, + 494, + 186 + ], + "spans": [ + { + "bbox": [ + 115, + 142, + 494, + 186 + ], + "type": "text", + "content": "Chengchao Zhang" + }, + { + "bbox": [ + 115, + 142, + 494, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 115, + 142, + 494, + 186 + ], + "type": "text", + "content": ", Fanhua Shang" + }, + { + "bbox": [ + 115, + 142, + 494, + 186 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 115, + 142, + 494, + 186 + ], + "type": "text", + "content": ", Hongyin Liu" + }, + { + "bbox": [ + 115, + 142, + 494, + 186 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 115, + 142, + 494, + 186 + ], + "type": "text", + "content": ", Liang Wan" + }, + { + "bbox": [ + 115, + 142, + 494, + 186 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 115, + 142, + 494, + 186 + ], + "type": "text", + "content": ", Wei Feng" + }, + { + "bbox": [ + 115, + 142, + 494, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 115, + 142, + 494, + 186 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 115, + 142, + 494, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 115, + 142, + 494, + 186 + ], + "type": "text", + "content": "College of Intelligence and Computing, Tianjin University, China " + }, + { + "bbox": [ + 115, + 142, + 494, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 115, + 142, + 494, + 186 + ], + "type": "text", + "content": "Medical College, Tianjin University, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 140, + 187, + 465, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 187, + 465, + 200 + ], + "spans": [ + { + "bbox": [ + 140, + 187, + 465, + 200 + ], + "type": "text", + "content": "{chengchaozhang, fhshang, hyliu2009, lwan, wfeng}@tju.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 252, + 296, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 252, + 296, + 516 + ], + "spans": [ + { + "bbox": [ + 54, + 252, + 296, + 516 + ], + "type": "text", + "content": "Federated Continual Learning (FCL) has emerged as a prominent distributed learning paradigm and aims at addressing model learning challenges in both federated and continual learning settings. Efficient personalization in FCL remains a major challenge, as it must handle not only conflicts between old and new knowledge within parallel task streams but also heterogeneous knowledge conflicts from different clients. Recent approaches attempt to mitigate these issues through gradient correction. However, they often overlook the combined impact of gradient magnitude and direction, leading to unsatisfactory gradient solutions. To address these issues, we propose a novel federated continual learning method (called FedAGC) with asymmetric gradient correction, which performs memory rehearsal using representative samples selected via a centroid-based approach from historical tasks. By formulating the problem as a multi-objective optimization problem, FedAGC derives more effective gradients while incorporating group-level personalization to facilitate useful knowledge integration and irrelevant knowledge isolation, effectively mitigating both temporal and spatial catastrophic forgetting. Extensive experiments confirm the effectiveness of FedAGC." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 540, + 135, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 135, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 135, + 553 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 561, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 561, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 561, + 296, + 693 + ], + "type": "text", + "content": "Federated Learning (FL) [21, 22, 29], as a novel distributed machine learning paradigm, aims to train a global model by leveraging independent and diverse data across clients in a privacy-preserving and collaborative manner. FL has been widely applied in various domains, such as medical image analysis [23] and object detection [24]. It has been proven to effectively address the issue of data silos [33] without compromising client privacy. Most existing FL methods are limited to single-task learning, assuming that client data remains static over time. However, in real-world scenarios, data is continuously evolving, forming sequential tasks with" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 225, + 566, + 327 + ], + "blocks": [ + { + "bbox": [ + 316, + 225, + 566, + 327 + ], + "lines": [ + { + "bbox": [ + 316, + 225, + 566, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 225, + 566, + 327 + ], + "type": "image", + "image_path": "5f89dc36fcd47dabdddc7f66f2e9a0e08331de42e8d0486e356c90553ec4ee86.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 328, + 403, + 336 + ], + "lines": [ + { + "bbox": [ + 317, + 328, + 403, + 336 + ], + "spans": [ + { + "bbox": [ + 317, + 328, + 403, + 336 + ], + "type": "text", + "content": "Local update of client 1" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 317, + 336, + 403, + 346 + ], + "lines": [ + { + "bbox": [ + 317, + 336, + 403, + 346 + ], + "spans": [ + { + "bbox": [ + 317, + 336, + 403, + 346 + ], + "type": "text", + "content": "Local update of client " + }, + { + "bbox": [ + 317, + 336, + 403, + 346 + ], + "type": "inline_equation", + "content": "N" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 317, + 347, + 397, + 357 + ], + "lines": [ + { + "bbox": [ + 317, + 347, + 397, + 357 + ], + "spans": [ + { + "bbox": [ + 317, + 347, + 397, + 357 + ], + "type": "text", + "content": "Aggregation of server" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 319, + 357, + 381, + 367 + ], + "lines": [ + { + "bbox": [ + 319, + 357, + 381, + 367 + ], + "spans": [ + { + "bbox": [ + 319, + 357, + 381, + 367 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 319, + 357, + 381, + 367 + ], + "type": "text", + "content": " Global optimum" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 320, + 368, + 391, + 377 + ], + "lines": [ + { + "bbox": [ + 320, + 368, + 391, + 377 + ], + "spans": [ + { + "bbox": [ + 320, + 368, + 391, + 377 + ], + "type": "text", + "content": "Server global model" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 320, + 379, + 386, + 387 + ], + "lines": [ + { + "bbox": [ + 320, + 379, + 386, + 387 + ], + "spans": [ + { + "bbox": [ + 320, + 379, + 386, + 387 + ], + "type": "text", + "content": "Client local model" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 320, + 388, + 370, + 397 + ], + "lines": [ + { + "bbox": [ + 320, + 388, + 370, + 397 + ], + "spans": [ + { + "bbox": [ + 320, + 388, + 370, + 397 + ], + "type": "text", + "content": "- Initial model" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 317, + 398, + 395, + 408 + ], + "lines": [ + { + "bbox": [ + 317, + 398, + 395, + 408 + ], + "spans": [ + { + "bbox": [ + 317, + 398, + 395, + 408 + ], + "type": "inline_equation", + "content": "\\iff" + }, + { + "bbox": [ + 317, + 398, + 395, + 408 + ], + "type": "text", + "content": " Parameter divergence" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 317, + 409, + 366, + 418 + ], + "lines": [ + { + "bbox": [ + 317, + 409, + 366, + 418 + ], + "spans": [ + { + "bbox": [ + 317, + 409, + 366, + 418 + ], + "type": "text", + "content": "Client drift" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 386, + 419, + 494, + 430 + ], + "lines": [ + { + "bbox": [ + 386, + 419, + 494, + 430 + ], + "spans": [ + { + "bbox": [ + 386, + 419, + 494, + 430 + ], + "type": "text", + "content": "(b) Spatial catastrophic forgetting" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 313, + 439, + 555, + 473 + ], + "lines": [ + { + "bbox": [ + 313, + 439, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 439, + 555, + 473 + ], + "type": "text", + "content": "Figure 1. Spatial-Temporal Catastrophic Forgetting. (a) Direct access to new tasks leads to a significant performance decline on old tasks. (b) Non-IID data exacerbates the weight disparity." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 485, + 555, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 555, + 544 + ], + "type": "text", + "content": "unpredictable future changes. As shown in Fig. 1a, severe temporal catastrophic forgetting (TemporalCF) [38, 45] occurs when the local model on each client learns a newly introduced task, leading to performance degradation on previously learned tasks as new knowledge is acquired." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 545, + 555, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 664 + ], + "type": "text", + "content": "Improving the generalizability of FL in multi-task evolving scenarios has emerged as a critical and challenging research topic. Recently, Continual Learning (CL) [17] has made significant advances in addressing TemporalCF, enabling effective mitigation of this problem. Based on such characteristics, the concept of Federated Continual Learning (FCL) [6, 46, 47] came into being, which integrates CL methods into the FL framework. In the FCL setting, FL can be regarded as a distributed instance of conventional CL, while CL methods assist in addressing FL challenges." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "content": "FCL breaks the static limitation of traditional FL, where each client is confined to learning a single task sequence. However, integrating CL into the FL framework poses substantial challenges. Besides TemporalCF, FCL also" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 144, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3841" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "content": "faces spatial catastrophic forgetting (SpatialCF). Specifically, the private data of different clients in FCL may follow drastically different distributions. This situation, where data across clients is Non-Independent and Identically Distributed (Non-IID), which is referred to as statistical heterogeneity [26]. During server-side aggregation, where gradients from multiple clients are combined to initialize the next round of local CL training, statistical heterogeneity can lead to issues such as parameter divergence and client drift [14, 37], as illustrated in Fig. 1b. Consequently, the aggregated global model exhibits reduced generalizability, further exacerbating temporal catastrophic forgetting." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 218, + 295, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 218, + 295, + 433 + ], + "spans": [ + { + "bbox": [ + 56, + 218, + 295, + 433 + ], + "type": "text", + "content": "To address the aforementioned challenges, recent studies [42, 47, 48] have explored the problem of Personalized Federated Continual Learning (PFCL). Unlike traditional FCL that aim to obtain a universal global model, PFCL leverages efficient server-side aggregation to facilitate personalized training. Existing PFCL methods mainly focus on optimizing local models for clients, ensuring strong performance on both new and previous tasks in a parallel, independent, and heterogeneous task stream across clients. The most relevant work to ours is FedMes [42], which builds on the gradient correction technique of A-GEM [2] and enhances local models through KNN-based inference. However, existing methods typically rely solely on gradient direction to determine similarity, ignoring the impact of magnitude, leading to suboptimal gradients. In contrast, we propose a new Asymmetric Gradient Correction (AGC) technique, which adaptively balances the combined impact of direction and magnitude, enabling a smoother optimization transition." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 434, + 295, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 295, + 650 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 295, + 650 + ], + "type": "text", + "content": "We formulate the PFCL problem as a Multi-Objective Optimization (MOO) problem to obtain the optimal gradient, where each objective minimizes the distance to a target gradient. Specifically, to mitigate Spatial-Temporal Catastrophic Forgetting (STCF), each client extracts a small set of representative samples after completing local training for each task and employs them for rehearsal in subsequent learning. Subsequently, we correct the gradients by considering both their magnitudes and directions, ultimately obtaining an optimal gradient that does not damage any of old or new knowledge at each update from a set of class-level gradients. Additionally, empirical evidence suggests that a single global model has limited generalization across all clients [34]. Therefore, we introduce a group-level personalization strategy, where clustering is performed based on the similarity of task gradients uploaded by participating clients. These clusters are then divided into groups, enabling joint training of multiple global models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 652, + 288, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 652, + 288, + 663 + ], + "spans": [ + { + "bbox": [ + 67, + 652, + 288, + 663 + ], + "type": "text", + "content": "Our main contributions can be summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 665, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 700 + ], + "type": "text", + "content": "- To the best of our knowledge, this is the first attempt to unify both spatial and temporal catastrophic forgetting in PFCL through a multi-objective optimization framework." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 702, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 702, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 702, + 295, + 714 + ], + "type": "text", + "content": "- We propose a novel asymmetric gradient correction algo" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 322, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 553, + 133 + ], + "type": "text", + "content": "rithm that performs memory rehearsal using representative samples selected via a centroid-based approach from historical tasks, transforming multi-task updates into an optimal gradient search problem that minimizes inter-task discrepancies and mitigates TemporalCF." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 133, + 553, + 228 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 133, + 553, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 133, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 314, + 133, + 553, + 191 + ], + "type": "text", + "content": "- We propose a new group-level personalization approach that clusters clients by gradient similarity and aggregates multiple global models, isolating irrelevant knowledge interference while fostering the positive fusion of related knowledge to mitigate SpatialCF." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 192, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 553, + 228 + ], + "type": "text", + "content": "- Our method outperforms all the baselines in terms of accuracy and forgetting rate, demonstrating its efficacy in addressing the PFCL problem." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 239, + 494, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 494, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 494, + 251 + ], + "type": "text", + "content": "2. Background and Related Works" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 258, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 555, + 676 + ], + "type": "text", + "content": "FL Algorithms for Statistical Heterogeneity. The client-server architecture-based FL paradigm is currently the most widely adopted framework. This distributed learning model involves a central server for aggregating network model parameters or gradients and " + }, + { + "bbox": [ + 313, + 258, + 555, + 676 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 258, + 555, + 676 + ], + "type": "text", + "content": " clients " + }, + { + "bbox": [ + 313, + 258, + 555, + 676 + ], + "type": "inline_equation", + "content": "\\{F_1, F_2, \\ldots, F_N\\}" + }, + { + "bbox": [ + 313, + 258, + 555, + 676 + ], + "type": "text", + "content": " for participating in the training of local private data, where each client possesses a local private dataset " + }, + { + "bbox": [ + 313, + 258, + 555, + 676 + ], + "type": "inline_equation", + "content": "\\{D_1, D_2, \\ldots, D_N\\}" + }, + { + "bbox": [ + 313, + 258, + 555, + 676 + ], + "type": "text", + "content": ". For the " + }, + { + "bbox": [ + 313, + 258, + 555, + 676 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 258, + 555, + 676 + ], + "type": "text", + "content": "-th client " + }, + { + "bbox": [ + 313, + 258, + 555, + 676 + ], + "type": "inline_equation", + "content": "F_i" + }, + { + "bbox": [ + 313, + 258, + 555, + 676 + ], + "type": "text", + "content": ", its local dataset includes " + }, + { + "bbox": [ + 313, + 258, + 555, + 676 + ], + "type": "inline_equation", + "content": "n_i" + }, + { + "bbox": [ + 313, + 258, + 555, + 676 + ], + "type": "text", + "content": " samples, denoted as " + }, + { + "bbox": [ + 313, + 258, + 555, + 676 + ], + "type": "inline_equation", + "content": "\\{D_i = \\{(X_k^i, Y_k^i)\\}_{k=1}^{n_i}" + }, + { + "bbox": [ + 313, + 258, + 555, + 676 + ], + "type": "text", + "content": ". Different clients may exhibit distinct features and labels, such that " + }, + { + "bbox": [ + 313, + 258, + 555, + 676 + ], + "type": "inline_equation", + "content": "X_i \\cap X_j = \\emptyset" + }, + { + "bbox": [ + 313, + 258, + 555, + 676 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 258, + 555, + 676 + ], + "type": "inline_equation", + "content": "\\forall i, j \\in [1, N]" + }, + { + "bbox": [ + 313, + 258, + 555, + 676 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 258, + 555, + 676 + ], + "type": "inline_equation", + "content": "Y_k \\neq \\emptyset" + }, + { + "bbox": [ + 313, + 258, + 555, + 676 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 258, + 555, + 676 + ], + "type": "inline_equation", + "content": "Y_i = \\emptyset" + }, + { + "bbox": [ + 313, + 258, + 555, + 676 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 258, + 555, + 676 + ], + "type": "inline_equation", + "content": "\\forall i \\in [1, N] \\setminus \\{k\\}" + }, + { + "bbox": [ + 313, + 258, + 555, + 676 + ], + "type": "text", + "content": ". The objective of FL is to collaboratively and privately train a global statistical model across multiple distributed clients under the supervision of a central server. The key formulation controlling FL in the classical FedAvg [29] is as follows: " + }, + { + "bbox": [ + 313, + 258, + 555, + 676 + ], + "type": "inline_equation", + "content": "\\min_{w \\in \\mathbb{R}^d} f(w) = \\frac{1}{m} \\sum_{i=1}^{m} f_i(w)" + }, + { + "bbox": [ + 313, + 258, + 555, + 676 + ], + "type": "text", + "content": ". where " + }, + { + "bbox": [ + 313, + 258, + 555, + 676 + ], + "type": "inline_equation", + "content": "f_i(w) = \\mathbb{E}_{(x_i, y_i) \\sim \\mathcal{P}_i}[\\mathcal{L}_i(w, x_i, y_i)]" + }, + { + "bbox": [ + 313, + 258, + 555, + 676 + ], + "type": "text", + "content": " represents the local loss function of the " + }, + { + "bbox": [ + 313, + 258, + 555, + 676 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 258, + 555, + 676 + ], + "type": "text", + "content": "-th client based on its local data, and " + }, + { + "bbox": [ + 313, + 258, + 555, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 313, + 258, + 555, + 676 + ], + "type": "text", + "content": " denotes the local data distribution of the " + }, + { + "bbox": [ + 313, + 258, + 555, + 676 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 258, + 555, + 676 + ], + "type": "text", + "content": "-th client. To address the challenge of statistical heterogeneity and mitigate client drift caused by Non-IID data distributions, FedProx [19] extends FedAvg by introducing a proximal term that effectively constrains local updates, thereby enhancing the overall stability of model convergence. Another representative approach, SCAFFOLD [14], employs control variates to alleviate client drift. However, under extreme Non-IID conditions, these methods often struggle to produce a highly generalizable global model. Consequently, Personalized FL [20, 45] shifts its focus to providing specialized local models for each client, enabling personalized services. PerFedAvg [7] combines meta-learning techniques during a pre-training phase to obtain a well-initialized model. While these approaches have demonstrated strong performance in FL, they are primarily designed to address SpatialCF and may not be sufficient for more complex scenarios." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Gradient Correction-Based CL Algorithms. In contrast to traditional FL designed for single-task scenarios, in a standard continual learning setting, there are a sequence" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3842" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 84, + 68, + 526, + 311 + ], + "blocks": [ + { + "bbox": [ + 84, + 68, + 526, + 311 + ], + "lines": [ + { + "bbox": [ + 84, + 68, + 526, + 311 + ], + "spans": [ + { + "bbox": [ + 84, + 68, + 526, + 311 + ], + "type": "image", + "image_path": "cc19b29212009502a2e6eeab4157dd9df203fc025f7a5c96d141eb41cacd00b0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 318, + 555, + 365 + ], + "lines": [ + { + "bbox": [ + 54, + 318, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 54, + 318, + 555, + 365 + ], + "type": "text", + "content": "Figure 2. The workflow of the proposed FedAGC method. Our FedAGC consists of the following three modules: On the client side, the Centroid-based Representative Sample Selection (CBRSS) module rehearses the gradients of previous tasks to assist the Asymmetric Correction Gradient Updating (ACGU) module in generating optimal gradients. On the server side, the Group Aggregation Gradient Updating (GAGU) module clusters clients based on gradient similarity and aggregates multiple global models." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 373, + 296, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 373, + 296, + 700 + ], + "spans": [ + { + "bbox": [ + 56, + 373, + 296, + 700 + ], + "type": "text", + "content": "of tasks " + }, + { + "bbox": [ + 56, + 373, + 296, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{\\mathcal{T}_1,\\mathcal{T}_2,\\dots ,\\mathcal{T}_T\\}" + }, + { + "bbox": [ + 56, + 373, + 296, + 700 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 56, + 373, + 296, + 700 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 56, + 373, + 296, + 700 + ], + "type": "text", + "content": " is the number of tasks. For the " + }, + { + "bbox": [ + 56, + 373, + 296, + 700 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 56, + 373, + 296, + 700 + ], + "type": "text", + "content": "-th task " + }, + { + "bbox": [ + 56, + 373, + 296, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_t" + }, + { + "bbox": [ + 56, + 373, + 296, + 700 + ], + "type": "text", + "content": ", its dataset is composed of " + }, + { + "bbox": [ + 56, + 373, + 296, + 700 + ], + "type": "inline_equation", + "content": "n_t" + }, + { + "bbox": [ + 56, + 373, + 296, + 700 + ], + "type": "text", + "content": " pairs of data and labels: " + }, + { + "bbox": [ + 56, + 373, + 296, + 700 + ], + "type": "inline_equation", + "content": "\\{x_i^t,y_i^t\\}_{i = 1}^{n_t}" + }, + { + "bbox": [ + 56, + 373, + 296, + 700 + ], + "type": "text", + "content": ". Typically, " + }, + { + "bbox": [ + 56, + 373, + 296, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 56, + 373, + 296, + 700 + ], + "type": "text", + "content": " refers to a task stream where the arrival order of new tasks is unknown. As a result, the model can only access " + }, + { + "bbox": [ + 56, + 373, + 296, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_t" + }, + { + "bbox": [ + 56, + 373, + 296, + 700 + ], + "type": "text", + "content": " during the training of the " + }, + { + "bbox": [ + 56, + 373, + 296, + 700 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 56, + 373, + 296, + 700 + ], + "type": "text", + "content": "-th task, and access to it is no longer available afterward. Therefore, for " + }, + { + "bbox": [ + 56, + 373, + 296, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_t" + }, + { + "bbox": [ + 56, + 373, + 296, + 700 + ], + "type": "text", + "content": ", given the set of model parameters " + }, + { + "bbox": [ + 56, + 373, + 296, + 700 + ], + "type": "inline_equation", + "content": "w^{(t)}" + }, + { + "bbox": [ + 56, + 373, + 296, + 700 + ], + "type": "text", + "content": " learned so far, the learning objective at task " + }, + { + "bbox": [ + 56, + 373, + 296, + 700 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 56, + 373, + 296, + 700 + ], + "type": "text", + "content": " is: " + }, + { + "bbox": [ + 56, + 373, + 296, + 700 + ], + "type": "inline_equation", + "content": "\\min_{w^{(t)}}\\mathcal{L}(w^{(t)};w^{(t - 1)},\\mathcal{T}^t)" + }, + { + "bbox": [ + 56, + 373, + 296, + 700 + ], + "type": "text", + "content": ". In recent years, numerous CL algorithms have adopted gradient correction strategies to mitigate TemporalCF, typically involving two key steps: Memory Rehearsal and Optimal Gradient Computation [15, 27]. A representative example is GEM [25], which retrieves gradients from previously stored memories and integrates them with the current task's gradient to obtain an update direction suitable for all tasks. As one variant, A-GEM [2], consolidates multiple memories to enhance computational efficiency. Meanwhile, ER [3] combines data from episodic memory with the current task during training. Although ER is simple and effective, its performance depends heavily on the quality of the stored memories. Another related approach is parameter regularization [5, 8], which introduces additional regularization terms to consolidate knowledge from previous tasks and reduce interference between new and old task parameters. Among these methods, EWC [38] employs the Fisher information matrix to quantify the importance of parameters for previously learned tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "While FL and CL have achieved remarkable success in" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 374, + 556, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 556, + 531 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 556, + 531 + ], + "type": "text", + "content": "their respective domains, effectively mitigating the STCF problem in PFCL remains a challenging endeavor [12, 45]. FedCurv [38] incorporates EWC to handle statistical heterogeneity; however, it incurs high computational overhead and lacks a dedicated mechanism for addressing TemporalCF. FedKNOW [44] deploys a gradient integrator at each client to reduce negative knowledge transfer by effectively combining signature tasks identified from past and current tasks of other clients through the global model. Recent advances in vision foundation models have driven an alternative prompt-based approach [9, 11, 43, 48] that mitigates STCF by adapting to new tasks through training only a small subset of prompts and classifier parameters." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 536, + 443, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 443, + 551 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 443, + 551 + ], + "type": "text", + "content": "3. The Proposed Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 557, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 689 + ], + "type": "text", + "content": "In this section, we elaborate on the proposed FedAGC. For the first time, FedAGC formulates the PFCL problem (see Sec. 3.1) as a MOO problem. On the client-side, after training each task, we select representative samples to construct a memory cache (see Sec. 3.2) and use them to perform asymmetric gradient correction for the subsequent new tasks (see Sec. 3.3), thus avoiding TemporalCF. On the server-side, we group the average gradient updates uploaded by the clients (see Sec. 3.4) and aggregate multiple global models to avoid SpatialCF, while also satisfying the personalized local requirements." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "The overall framework of FedAGC is shown in Figure 2, the algorithm is summarized in Algorithm 1." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3843" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 284, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 284, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 284, + 85 + ], + "type": "text", + "content": "3.1. Personalized Federated Continual Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 221 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 221 + ], + "type": "text", + "content": "We extend the traditional CL and FL methods to PFCL, which enables the independent and parallel integration of multiple task sequences across clients. This approach achieving both the fusion of local knowledge from new and old tasks and the integration of heterogeneous task knowledge across clients. Unlike FCL, which aims to obtain a general model that integrates global knowledge, PFCL addresses the diverse local needs of clients in highly heterogeneous scenarios, resulting in personalized local models. Specifically, each client learns its local model by optimizing the following objective:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 102, + 227, + 295, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 227, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 102, + 227, + 295, + 251 + ], + "type": "interline_equation", + "content": "\\min _ {w ^ {(t)}} \\mathcal {L} \\left(w ^ {(t)}; \\mathcal {T} ^ {t}\\right) + \\frac {\\lambda}{2} \\| w ^ {(t)} - w _ {0} \\| ^ {2}, \\tag {1}", + "image_path": "44af05b51f599ceaea20fb73fc53fa6f487f6020b153fa4380d92f1a18dc3776.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 258, + 295, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 295, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 295, + 281 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 258, + 295, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\cdot)" + }, + { + "bbox": [ + 55, + 258, + 295, + 281 + ], + "type": "text", + "content": " represents the empirical risk loss, " + }, + { + "bbox": [ + 55, + 258, + 295, + 281 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 258, + 295, + 281 + ], + "type": "text", + "content": " is the regularization coefficient, and " + }, + { + "bbox": [ + 55, + 258, + 295, + 281 + ], + "type": "inline_equation", + "content": "w_{0}" + }, + { + "bbox": [ + 55, + 258, + 295, + 281 + ], + "type": "text", + "content": " denotes the initial model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 281, + 295, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 295, + 306 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 295, + 306 + ], + "type": "text", + "content": "The PFCL scenario we consider, involving independent parallel multi-task sequences, comprises three key steps:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 306, + 295, + 558 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 55, + 306, + 295, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 306, + 295, + 390 + ], + "spans": [ + { + "bbox": [ + 55, + 306, + 295, + 390 + ], + "type": "text", + "content": "(1) Given " + }, + { + "bbox": [ + 55, + 306, + 295, + 390 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 306, + 295, + 390 + ], + "type": "text", + "content": " clients " + }, + { + "bbox": [ + 55, + 306, + 295, + 390 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{C}_1,\\mathcal{C}_2,\\dots ,\\mathcal{C}_N\\}" + }, + { + "bbox": [ + 55, + 306, + 295, + 390 + ], + "type": "text", + "content": " and a server " + }, + { + "bbox": [ + 55, + 306, + 295, + 390 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 55, + 306, + 295, + 390 + ], + "type": "text", + "content": ". There are task streams " + }, + { + "bbox": [ + 55, + 306, + 295, + 390 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{T}^1,\\mathcal{T}^2,\\ldots ,\\mathcal{T}^N\\}" + }, + { + "bbox": [ + 55, + 306, + 295, + 390 + ], + "type": "text", + "content": ", within which each client " + }, + { + "bbox": [ + 55, + 306, + 295, + 390 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_c" + }, + { + "bbox": [ + 55, + 306, + 295, + 390 + ], + "type": "text", + "content": " engages in training a local CL model on its unique, private, and accessible task sequence " + }, + { + "bbox": [ + 55, + 306, + 295, + 390 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^c = \\{\\mathcal{T}_1^c,\\mathcal{T}_2^c,\\dots ,\\mathcal{T}_T^c\\}" + }, + { + "bbox": [ + 55, + 306, + 295, + 390 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 306, + 295, + 390 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^c" + }, + { + "bbox": [ + 55, + 306, + 295, + 390 + ], + "type": "text", + "content": " consists of " + }, + { + "bbox": [ + 55, + 306, + 295, + 390 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 306, + 295, + 390 + ], + "type": "text", + "content": " distinct tasks, where the " + }, + { + "bbox": [ + 55, + 306, + 295, + 390 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 306, + 295, + 390 + ], + "type": "text", + "content": "-th task " + }, + { + "bbox": [ + 55, + 306, + 295, + 390 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_t^c" + }, + { + "bbox": [ + 55, + 306, + 295, + 390 + ], + "type": "text", + "content": " includes " + }, + { + "bbox": [ + 55, + 306, + 295, + 390 + ], + "type": "inline_equation", + "content": "K_{t}^{c}" + }, + { + "bbox": [ + 55, + 306, + 295, + 390 + ], + "type": "text", + "content": " classes, and the total set of classes for " + }, + { + "bbox": [ + 55, + 306, + 295, + 390 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_c" + }, + { + "bbox": [ + 55, + 306, + 295, + 390 + ], + "type": "text", + "content": " is represented as " + }, + { + "bbox": [ + 55, + 306, + 295, + 390 + ], + "type": "inline_equation", + "content": "K^{c} = \\{K_{1}^{c},K_{2}^{c},\\dots ,K_{T}^{c}\\}" + }, + { + "bbox": [ + 55, + 306, + 295, + 390 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 389, + 295, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 295, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 295, + 473 + ], + "type": "text", + "content": "(2) For the current task " + }, + { + "bbox": [ + 55, + 389, + 295, + 473 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_t^c" + }, + { + "bbox": [ + 55, + 389, + 295, + 473 + ], + "type": "text", + "content": " being learned, the participating client receives the model " + }, + { + "bbox": [ + 55, + 389, + 295, + 473 + ], + "type": "inline_equation", + "content": "\\theta_g^t" + }, + { + "bbox": [ + 55, + 389, + 295, + 473 + ], + "type": "text", + "content": " globally broadcasted by " + }, + { + "bbox": [ + 55, + 389, + 295, + 473 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 389, + 295, + 473 + ], + "type": "text", + "content": " as the initial model, where " + }, + { + "bbox": [ + 55, + 389, + 295, + 473 + ], + "type": "inline_equation", + "content": "\\theta_g^t" + }, + { + "bbox": [ + 55, + 389, + 295, + 473 + ], + "type": "text", + "content": " incorporates knowledge from the previous " + }, + { + "bbox": [ + 55, + 389, + 295, + 473 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 55, + 389, + 295, + 473 + ], + "type": "text", + "content": " tasks contributed by various clients. After " + }, + { + "bbox": [ + 55, + 389, + 295, + 473 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 55, + 389, + 295, + 473 + ], + "type": "text", + "content": " rounds of local training on " + }, + { + "bbox": [ + 55, + 389, + 295, + 473 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_t^c" + }, + { + "bbox": [ + 55, + 389, + 295, + 473 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 389, + 295, + 473 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_c" + }, + { + "bbox": [ + 55, + 389, + 295, + 473 + ], + "type": "text", + "content": " obtains a new continual learning model " + }, + { + "bbox": [ + 55, + 389, + 295, + 473 + ], + "type": "inline_equation", + "content": "\\theta_{l,c}^{t}" + }, + { + "bbox": [ + 55, + 389, + 295, + 473 + ], + "type": "text", + "content": " which performs well on " + }, + { + "bbox": [ + 55, + 389, + 295, + 473 + ], + "type": "inline_equation", + "content": "K^c" + }, + { + "bbox": [ + 55, + 389, + 295, + 473 + ], + "type": "text", + "content": ", and the corresponding average gradient update " + }, + { + "bbox": [ + 55, + 389, + 295, + 473 + ], + "type": "inline_equation", + "content": "G^{c}" + }, + { + "bbox": [ + 55, + 389, + 295, + 473 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 473, + 295, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 295, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 295, + 558 + ], + "type": "text", + "content": "(3) In each communication round " + }, + { + "bbox": [ + 55, + 473, + 295, + 558 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 473, + 295, + 558 + ], + "type": "text", + "content": ", the participating client " + }, + { + "bbox": [ + 55, + 473, + 295, + 558 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_c" + }, + { + "bbox": [ + 55, + 473, + 295, + 558 + ], + "type": "text", + "content": " upload its respective " + }, + { + "bbox": [ + 55, + 473, + 295, + 558 + ], + "type": "inline_equation", + "content": "G^c" + }, + { + "bbox": [ + 55, + 473, + 295, + 558 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 473, + 295, + 558 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 473, + 295, + 558 + ], + "type": "text", + "content": " employs a group personalization strategy to aggregate them, which is then used to update " + }, + { + "bbox": [ + 55, + 473, + 295, + 558 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 473, + 295, + 558 + ], + "type": "text", + "content": " global models " + }, + { + "bbox": [ + 55, + 473, + 295, + 558 + ], + "type": "inline_equation", + "content": "\\theta_{g,A}^t" + }, + { + "bbox": [ + 55, + 473, + 295, + 558 + ], + "type": "text", + "content": ". For the " + }, + { + "bbox": [ + 55, + 473, + 295, + 558 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 55, + 473, + 295, + 558 + ], + "type": "text", + "content": "-th global model " + }, + { + "bbox": [ + 55, + 473, + 295, + 558 + ], + "type": "inline_equation", + "content": "\\theta_{g,a}^t" + }, + { + "bbox": [ + 55, + 473, + 295, + 558 + ], + "type": "text", + "content": ", it encapsulates the generalized knowledge of the associated client group's learning tasks and serves as the initial model for the next training round." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 564, + 205, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 205, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 205, + 577 + ], + "type": "text", + "content": "3.2. Memory Rehearsal in FCL" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 582, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 665 + ], + "type": "text", + "content": "Benefiting from the efficient utilization of past experience, memory rehearsal methods [2, 10, 27] are widely used in CL as simple and effective forgetting suppression techniques. These methods involve storing partial data or knowledge from previous tasks to enable retraining during the learning of new tasks, thereby mitigating forgetting of prior tasks. Now, we extend it from CL to FCL." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Memory Design. Each client utilizes its limited local memory to construct a small cache space, denoted as " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "M_{c}" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ". Upon the completion of the training process for the " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "-th task of " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_c" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ", rehearsal samples " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "M_t^c" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " are sampled at a fixed proportion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "content": " from the training set " + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "inline_equation", + "content": "D_{t}^{c}" + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "content": " of that task. For the initial task, " + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "inline_equation", + "content": "M_0^c = \\emptyset" + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "content": ", and for the final task, no additional sampling operations are performed. Thus, " + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "inline_equation", + "content": "M_{c}" + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "content": " can be represented as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 400, + 114, + 555, + 129 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 114, + 555, + 129 + ], + "spans": [ + { + "bbox": [ + 400, + 114, + 555, + 129 + ], + "type": "interline_equation", + "content": "M _ {c} = \\cup_ {t = 1} ^ {T} M _ {t} ^ {c}, \\tag {2}", + "image_path": "bfacbfd1da0afa092e74fbc7830e8c20386e29746ee57b1ca452333bd8fa0a56.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 135, + 533, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 135, + 533, + 148 + ], + "spans": [ + { + "bbox": [ + 313, + 135, + 533, + 148 + ], + "type": "text", + "content": "where the size of " + }, + { + "bbox": [ + 313, + 135, + 533, + 148 + ], + "type": "inline_equation", + "content": "M_t^c" + }, + { + "bbox": [ + 313, + 135, + 533, + 148 + ], + "type": "text", + "content": " is significantly smaller than " + }, + { + "bbox": [ + 313, + 135, + 533, + 148 + ], + "type": "inline_equation", + "content": "D_t^c" + }, + { + "bbox": [ + 313, + 135, + 533, + 148 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 147, + 555, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 147, + 555, + 255 + ], + "spans": [ + { + "bbox": [ + 313, + 147, + 555, + 255 + ], + "type": "text", + "content": "Centroid-Based Representative Sample Selection. Clearly, the quality of samples drawn from the old task dataset directly affects the quality of the memory cache. Therefore, we propose a Centroid-Based Representative Sample Selection (CBRSS) method to effectively acquire high-quality rehearsal samples. In particular, for client " + }, + { + "bbox": [ + 313, + 147, + 555, + 255 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_c" + }, + { + "bbox": [ + 313, + 147, + 555, + 255 + ], + "type": "text", + "content": ", given the distance threshold " + }, + { + "bbox": [ + 313, + 147, + 555, + 255 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 147, + 555, + 255 + ], + "type": "text", + "content": " for centroids and the input data " + }, + { + "bbox": [ + 313, + 147, + 555, + 255 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 147, + 555, + 255 + ], + "type": "text", + "content": ", the constraint condition for the minimum preset distance is defined as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 380, + 262, + 555, + 279 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 262, + 555, + 279 + ], + "spans": [ + { + "bbox": [ + 380, + 262, + 555, + 279 + ], + "type": "interline_equation", + "content": "\\min _ {i} \\| m _ {i} ^ {c} - f _ {\\theta , c} (x) \\| > v, \\tag {3}", + "image_path": "00201a688e69849fba63f205028815f2d4c3018f7d47f723244fff06f6756f7e.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 286, + 555, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 286, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 286, + 555, + 370 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 286, + 555, + 370 + ], + "type": "inline_equation", + "content": "m_i^c" + }, + { + "bbox": [ + 313, + 286, + 555, + 370 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 313, + 286, + 555, + 370 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 286, + 555, + 370 + ], + "type": "text", + "content": "-th centroid, and " + }, + { + "bbox": [ + 313, + 286, + 555, + 370 + ], + "type": "inline_equation", + "content": "f_{\\theta,c}(x)" + }, + { + "bbox": [ + 313, + 286, + 555, + 370 + ], + "type": "text", + "content": " is the predictor of " + }, + { + "bbox": [ + 313, + 286, + 555, + 370 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_c" + }, + { + "bbox": [ + 313, + 286, + 555, + 370 + ], + "type": "text", + "content": " for the current task, used to map " + }, + { + "bbox": [ + 313, + 286, + 555, + 370 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 286, + 555, + 370 + ], + "type": "text", + "content": " to the feature space. If the condition in Eq. (3) is satisfied, it indicates that this data point is independent and can be considered as a new centroid " + }, + { + "bbox": [ + 313, + 286, + 555, + 370 + ], + "type": "inline_equation", + "content": "m_{new,c} = f_{\\theta,c}(x)" + }, + { + "bbox": [ + 313, + 286, + 555, + 370 + ], + "type": "text", + "content": ". Otherwise, it means that this data point belongs to the existing centroid with the smallest distance to it, and then the target centroid is updated by" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 374, + 376, + 554, + 404 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 376, + 554, + 404 + ], + "spans": [ + { + "bbox": [ + 374, + 376, + 554, + 404 + ], + "type": "interline_equation", + "content": "m _ {i ^ {*}} = \\frac {N _ {i ^ {*}} ^ {k} \\times m _ {i} + f _ {\\theta , c} (x)}{N _ {i ^ {*}} ^ {k} + 1}, \\tag {4}", + "image_path": "8450d3367a8ce1bb25e4ae69933a66c1373095847ba5b39270e15971d40b210d.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 410, + 554, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 410, + 554, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 554, + 452 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 410, + 554, + 452 + ], + "type": "inline_equation", + "content": "i^{*} = \\underset {i}{\\arg \\min}\\| f_{\\theta ,c}(x) - m_{i}^{c}\\|" + }, + { + "bbox": [ + 313, + 410, + 554, + 452 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 410, + 554, + 452 + ], + "type": "inline_equation", + "content": "N_{i^*}^k" + }, + { + "bbox": [ + 313, + 410, + 554, + 452 + ], + "type": "text", + "content": " represents the total number of samples in the " + }, + { + "bbox": [ + 313, + 410, + 554, + 452 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 410, + 554, + 452 + ], + "type": "text", + "content": "-th category captured by the centroid closest to " + }, + { + "bbox": [ + 313, + 410, + 554, + 452 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 410, + 554, + 452 + ], + "type": "text", + "content": " in the current task." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 452, + 555, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 452, + 555, + 572 + ], + "spans": [ + { + "bbox": [ + 313, + 452, + 555, + 572 + ], + "type": "text", + "content": "Overall, each client stores the representative data of each category for the trained task after completing the training. A similar approach is employed by FedMes [42], which utilizes a random sampling method to construct the memory cache. However, due to the small sample size, this approach inevitably results in significant sampling bias. The resulting " + }, + { + "bbox": [ + 313, + 452, + 555, + 572 + ], + "type": "inline_equation", + "content": "M_{t}^{c}" + }, + { + "bbox": [ + 313, + 452, + 555, + 572 + ], + "type": "text", + "content": " lacks representativeness and fails to reflect the overall distribution of " + }, + { + "bbox": [ + 313, + 452, + 555, + 572 + ], + "type": "inline_equation", + "content": "D_{t}^{c}" + }, + { + "bbox": [ + 313, + 452, + 555, + 572 + ], + "type": "text", + "content": ". This leads to a drift in the representation of old task data in the feature space, which in turn triggers the problem of catastrophic forgetting." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 578, + 552, + 591 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 552, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 552, + 591 + ], + "type": "text", + "content": "3.3. Asymmetrically Correcting Gradient Updates" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 595, + 554, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 595, + 554, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 595, + 554, + 643 + ], + "type": "text", + "content": "To integrate the knowledge from previous tasks with that of the current task during the learning process of the " + }, + { + "bbox": [ + 313, + 595, + 554, + 643 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 595, + 554, + 643 + ], + "type": "text", + "content": "-th task on " + }, + { + "bbox": [ + 313, + 595, + 554, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_c" + }, + { + "bbox": [ + 313, + 595, + 554, + 643 + ], + "type": "text", + "content": ", we reformulate the training problem as a MOO problem, which is formally described as follows:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 327, + 651, + 555, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 651, + 555, + 670 + ], + "spans": [ + { + "bbox": [ + 327, + 651, + 555, + 670 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta_ {c}, \\theta_ {c} ^ {1: t}} \\ell \\left(f _ {\\theta_ {c}}, f _ {\\theta_ {c} ^ {t}}, D _ {t} ^ {c}\\right) \\cup \\ell \\left(f _ {\\theta_ {c}}, f _ {\\theta_ {c} ^ {k}}, M _ {k} ^ {c}\\right) \\mid \\forall k < t. \\tag {5}", + "image_path": "1fe589af198175e5913477f05c8b11478f7bfa19e2468876241a1ad9f251d767.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\ell(\\cdot)" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": " denotes the loss function, and " + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\theta_c" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\theta_c^t" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": " represent the backbone model for shared knowledge across multiple tasks and the task-specific classifier for " + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_c" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3844" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": "Since the parameters of the new and old task classifiers are independent and can be updated directly using their respective gradients " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\theta_c^k\\gets \\theta_c^k -\\alpha \\nabla \\ell (f_{\\theta_c},f_{\\theta_c^k})" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " with step size " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " our focus is on updating the shared backbone model, which is updated using the gradients " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "g_{k}^{c} = \\nabla \\ell (f_{\\theta_{c}},f_{\\theta_{c}^{k}})" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " of all " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " tasks. Let " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " be the step size for optimization. The problem of the backbone model update can be formulated as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 64, + 165, + 296, + 185 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 165, + 296, + 185 + ], + "spans": [ + { + "bbox": [ + 64, + 165, + 296, + 185 + ], + "type": "interline_equation", + "content": "\\theta_ {c} \\leftarrow \\theta_ {c} - \\beta d _ {c} ^ {*}, d _ {c} ^ {*} = \\arg \\min _ {d _ {c}} D \\left(g _ {k} ^ {c}, d _ {c}\\right) \\mid \\forall k \\leq t, \\tag {6}", + "image_path": "061ec46a02c747ee58ba1d7c7cda87ffd69006cbafd8dd8035abd9f3dca4635c.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 193, + 296, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 193, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 193, + 296, + 277 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 193, + 296, + 277 + ], + "type": "inline_equation", + "content": "d_c^*" + }, + { + "bbox": [ + 55, + 193, + 296, + 277 + ], + "type": "text", + "content": " represents the optimal gradient, which approximates the gradient of the new task while avoiding conflicts with the gradient of the old task. In Eq. (6), we define it as a min-distance multi-objective problem. Therefore, the critical issue lies in determining the method used to compute the differences between gradients, ensuring that the solution for " + }, + { + "bbox": [ + 55, + 193, + 296, + 277 + ], + "type": "inline_equation", + "content": "d_c^*" + }, + { + "bbox": [ + 55, + 193, + 296, + 277 + ], + "type": "text", + "content": " is as close as possible to all gradients." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 277, + 296, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 296, + 482 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 296, + 482 + ], + "type": "text", + "content": "Symmetric distance metrics, such as Euclidean distance and cosine distance, are widely applied in deep neural networks. These metrics share a fundamental property: the forward impact (from " + }, + { + "bbox": [ + 55, + 277, + 296, + 482 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 277, + 296, + 482 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 277, + 296, + 482 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 277, + 296, + 482 + ], + "type": "text", + "content": ") is equal to the backward impact (from " + }, + { + "bbox": [ + 55, + 277, + 296, + 482 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 277, + 296, + 482 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 277, + 296, + 482 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 277, + 296, + 482 + ], + "type": "text", + "content": "). However, local model updates are highly dependent on the magnitude and direction of gradients, both of which are inherently asymmetric. In brief, given an old task gradient " + }, + { + "bbox": [ + 55, + 277, + 296, + 482 + ], + "type": "inline_equation", + "content": "g_{old}" + }, + { + "bbox": [ + 55, + 277, + 296, + 482 + ], + "type": "text", + "content": ", trained on " + }, + { + "bbox": [ + 55, + 277, + 296, + 482 + ], + "type": "inline_equation", + "content": "M_c^k" + }, + { + "bbox": [ + 55, + 277, + 296, + 482 + ], + "type": "text", + "content": ", and a new task gradient " + }, + { + "bbox": [ + 55, + 277, + 296, + 482 + ], + "type": "inline_equation", + "content": "g_{new}" + }, + { + "bbox": [ + 55, + 277, + 296, + 482 + ], + "type": "text", + "content": ", trained on " + }, + { + "bbox": [ + 55, + 277, + 296, + 482 + ], + "type": "inline_equation", + "content": "D_t^c" + }, + { + "bbox": [ + 55, + 277, + 296, + 482 + ], + "type": "text", + "content": ", their mutual influences on each other's tasks are significantly different. In previous studies [2, 42, 49], the condition " + }, + { + "bbox": [ + 55, + 277, + 296, + 482 + ], + "type": "inline_equation", + "content": "\\langle g_{old}, g_{new} \\rangle \\geq 0" + }, + { + "bbox": [ + 55, + 277, + 296, + 482 + ], + "type": "text", + "content": " has been employed to signify the absence of conflicts between two tasks. However, given that the old task has been integrated earlier, " + }, + { + "bbox": [ + 55, + 277, + 296, + 482 + ], + "type": "inline_equation", + "content": "\\| g_{old} \\|" + }, + { + "bbox": [ + 55, + 277, + 296, + 482 + ], + "type": "text", + "content": " is typically smaller while " + }, + { + "bbox": [ + 55, + 277, + 296, + 482 + ], + "type": "inline_equation", + "content": "\\| g_{new} \\|" + }, + { + "bbox": [ + 55, + 277, + 296, + 482 + ], + "type": "text", + "content": " tends to be larger. In cases where " + }, + { + "bbox": [ + 55, + 277, + 296, + 482 + ], + "type": "inline_equation", + "content": "\\| g_{old} \\| \\ll \\| g_{new} \\|" + }, + { + "bbox": [ + 55, + 277, + 296, + 482 + ], + "type": "text", + "content": ", even if " + }, + { + "bbox": [ + 55, + 277, + 296, + 482 + ], + "type": "inline_equation", + "content": "\\langle g_{old}, g_{new} \\rangle \\geq 0" + }, + { + "bbox": [ + 55, + 277, + 296, + 482 + ], + "type": "text", + "content": " holds true, the updates pertaining to the new task may have a substantial impact on the performance of the old task." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 489, + 296, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 489, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 489, + 296, + 525 + ], + "type": "text", + "content": "Definition 1 (Asymmetric Gradient Distance). Given two gradients " + }, + { + "bbox": [ + 55, + 489, + 296, + 525 + ], + "type": "inline_equation", + "content": "g_{A}" + }, + { + "bbox": [ + 55, + 489, + 296, + 525 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 489, + 296, + 525 + ], + "type": "inline_equation", + "content": "g_{B}" + }, + { + "bbox": [ + 55, + 489, + 296, + 525 + ], + "type": "text", + "content": ", the asymmetric gradient distance is defined as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 533, + 247, + 561 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 533, + 247, + 561 + ], + "spans": [ + { + "bbox": [ + 104, + 533, + 247, + 561 + ], + "type": "interline_equation", + "content": "D \\left(g _ {A}, g _ {B}\\right) = \\frac {\\left\\| g _ {A} - g _ {B} \\right\\|}{\\left\\| g _ {B} \\right\\| + \\left\\| g _ {A} - g _ {B} \\right\\|}.", + "image_path": "d69a8d51e6a346d772ce17e08cf5f9875a0701c82b03a6727e59ee3dd13722d9.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 570, + 296, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 630 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 630 + ], + "type": "text", + "content": "In this paper, an asymmetric distance calculation method is used to explore the measure of difference between the new task and the old task. Definition 1 considers both the gradient direction and magnitude. Notably, " + }, + { + "bbox": [ + 55, + 570, + 296, + 630 + ], + "type": "inline_equation", + "content": "D(g_{A}, g_{B}) = 0" + }, + { + "bbox": [ + 55, + 570, + 296, + 630 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 55, + 570, + 296, + 630 + ], + "type": "inline_equation", + "content": "g_{A} = g_{B} = 0" + }, + { + "bbox": [ + 55, + 570, + 296, + 630 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": "However, solving Eq. (6) directly for the parallel task flow on clients presents significant challenges. This difficulty arises from the fact that both the decision space of this problem and the model parameters are identically high-dimensional. Consequently, following [28, 35], we employ linear scalarization to determine a weight variable " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "w^{*}" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": ". Specifically, when client " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "C_c" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": " is training on the " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": "-th task, the" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 72, + 408, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 408, + 83 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 408, + 83 + ], + "type": "text", + "content": "Algorithm 1 FedAGC" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 86, + 555, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 86, + 555, + 125 + ], + "spans": [ + { + "bbox": [ + 313, + 86, + 555, + 125 + ], + "type": "text", + "content": "Input: " + }, + { + "bbox": [ + 313, + 86, + 555, + 125 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 86, + 555, + 125 + ], + "type": "text", + "content": " clients with their private parallel task sequence " + }, + { + "bbox": [ + 313, + 86, + 555, + 125 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^{1:N}" + }, + { + "bbox": [ + 313, + 86, + 555, + 125 + ], + "type": "text", + "content": " and local datasets " + }, + { + "bbox": [ + 313, + 86, + 555, + 125 + ], + "type": "inline_equation", + "content": "\\{D_t^{1:N}\\}_{t=1}^T" + }, + { + "bbox": [ + 313, + 86, + 555, + 125 + ], + "type": "text", + "content": ", global model parameter " + }, + { + "bbox": [ + 313, + 86, + 555, + 125 + ], + "type": "inline_equation", + "content": "\\theta_g" + }, + { + "bbox": [ + 313, + 86, + 555, + 125 + ], + "type": "text", + "content": ", communication rounds " + }, + { + "bbox": [ + 313, + 86, + 555, + 125 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 86, + 555, + 125 + ], + "type": "text", + "content": ", memory space " + }, + { + "bbox": [ + 313, + 86, + 555, + 125 + ], + "type": "inline_equation", + "content": "M_{1:c} \\gets \\{\\}" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 124, + 554, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 124, + 554, + 150 + ], + "spans": [ + { + "bbox": [ + 313, + 124, + 554, + 150 + ], + "type": "text", + "content": "Output: Multiple global parameters " + }, + { + "bbox": [ + 313, + 124, + 554, + 150 + ], + "type": "inline_equation", + "content": "\\theta_{g}^{i:n}" + }, + { + "bbox": [ + 313, + 124, + 554, + 150 + ], + "type": "text", + "content": ", local parameter " + }, + { + "bbox": [ + 313, + 124, + 554, + 150 + ], + "type": "inline_equation", + "content": "\\theta_{l,1:N}" + }, + { + "bbox": [ + 313, + 124, + 554, + 150 + ], + "type": "text", + "content": " and classification head " + }, + { + "bbox": [ + 313, + 124, + 554, + 150 + ], + "type": "inline_equation", + "content": "\\theta_{l,1:N}^{k}" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 318, + 151, + 553, + 356 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 321, + 151, + 386, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 151, + 386, + 161 + ], + "spans": [ + { + "bbox": [ + 321, + 151, + 386, + 161 + ], + "type": "text", + "content": "1: Initialization" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 321, + 163, + 449, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 163, + 449, + 173 + ], + "spans": [ + { + "bbox": [ + 321, + 163, + 449, + 173 + ], + "type": "text", + "content": "2: while task number " + }, + { + "bbox": [ + 321, + 163, + 449, + 173 + ], + "type": "inline_equation", + "content": "t \\leq T" + }, + { + "bbox": [ + 321, + 163, + 449, + 173 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 321, + 175, + 481, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 175, + 481, + 186 + ], + "spans": [ + { + "bbox": [ + 321, + 175, + 481, + 186 + ], + "type": "text", + "content": "3: for each round " + }, + { + "bbox": [ + 321, + 175, + 481, + 186 + ], + "type": "inline_equation", + "content": "r = 1,2,\\ldots ,R" + }, + { + "bbox": [ + 321, + 175, + 481, + 186 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 321, + 186, + 535, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 186, + 535, + 199 + ], + "spans": [ + { + "bbox": [ + 321, + 186, + 535, + 199 + ], + "type": "text", + "content": "4: Server broadcasts " + }, + { + "bbox": [ + 321, + 186, + 535, + 199 + ], + "type": "inline_equation", + "content": "\\theta_g^i" + }, + { + "bbox": [ + 321, + 186, + 535, + 199 + ], + "type": "text", + "content": " to each group of clients;" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 199, + 515, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 199, + 515, + 209 + ], + "spans": [ + { + "bbox": [ + 321, + 199, + 515, + 209 + ], + "type": "text", + "content": "5: for each client " + }, + { + "bbox": [ + 321, + 199, + 515, + 209 + ], + "type": "inline_equation", + "content": "c = 1" + }, + { + "bbox": [ + 321, + 199, + 515, + 209 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 321, + 199, + 515, + 209 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 321, + 199, + 515, + 209 + ], + "type": "text", + "content": " in parallel do" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 210, + 533, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 210, + 533, + 221 + ], + "spans": [ + { + "bbox": [ + 321, + 210, + 533, + 221 + ], + "type": "text", + "content": "6: Minimizing local CL problems by Eq. (1);" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 222, + 457, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 222, + 457, + 235 + ], + "spans": [ + { + "bbox": [ + 321, + 222, + 457, + 235 + ], + "type": "text", + "content": "7: " + }, + { + "bbox": [ + 321, + 222, + 457, + 235 + ], + "type": "inline_equation", + "content": "g^{c} = \\nabla \\ell (f_{\\theta_{c}},f_{\\theta^{t}},D_{t}^{c});" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 235, + 512, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 235, + 512, + 247 + ], + "spans": [ + { + "bbox": [ + 321, + 235, + 512, + 247 + ], + "type": "text", + "content": "8: " + }, + { + "bbox": [ + 321, + 235, + 512, + 247 + ], + "type": "inline_equation", + "content": "g_{k}^{c} = \\nabla \\ell (f_{\\theta_{c}},f_{\\theta_{k}^{k}},M_{k}^{c})" + }, + { + "bbox": [ + 321, + 235, + 512, + 247 + ], + "type": "text", + "content": " for all " + }, + { + "bbox": [ + 321, + 235, + 512, + 247 + ], + "type": "inline_equation", + "content": "k < t" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 247, + 489, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 247, + 489, + 259 + ], + "spans": [ + { + "bbox": [ + 321, + 247, + 489, + 259 + ], + "type": "text", + "content": "9: " + }, + { + "bbox": [ + 321, + 247, + 489, + 259 + ], + "type": "inline_equation", + "content": "d_c^* \\gets \\mathsf{ACGU}(g^c, g_1^c, \\ldots, g_{t-1}^k)" + }, + { + "bbox": [ + 321, + 247, + 489, + 259 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 318, + 260, + 438, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 260, + 438, + 273 + ], + "spans": [ + { + "bbox": [ + 318, + 260, + 438, + 273 + ], + "type": "text", + "content": "10: " + }, + { + "bbox": [ + 318, + 260, + 438, + 273 + ], + "type": "inline_equation", + "content": "\\theta_{l,c}^{t}\\gets \\theta_{l,c}^{t} - \\alpha d_{c}^{*};" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 318, + 274, + 385, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 274, + 385, + 281 + ], + "spans": [ + { + "bbox": [ + 318, + 274, + 385, + 281 + ], + "type": "text", + "content": "11: end for" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 318, + 283, + 541, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 283, + 541, + 296 + ], + "spans": [ + { + "bbox": [ + 318, + 283, + 541, + 296 + ], + "type": "text", + "content": "12: Send the " + }, + { + "bbox": [ + 318, + 283, + 541, + 296 + ], + "type": "inline_equation", + "content": "G^{c} = -\\triangle \\theta_{l,c}^{t} / \\eta" + }, + { + "bbox": [ + 318, + 283, + 541, + 296 + ], + "type": "text", + "content": " to the central server;" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 318, + 297, + 543, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 297, + 543, + 308 + ], + "spans": [ + { + "bbox": [ + 318, + 297, + 543, + 308 + ], + "type": "text", + "content": "13: " + }, + { + "bbox": [ + 318, + 297, + 543, + 308 + ], + "type": "inline_equation", + "content": "\\{S_1,S_2,\\ldots ,S_n\\} \\gets \\mathrm{GAGU}(G^1,G^2,\\ldots ,G^N);" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 318, + 309, + 546, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 309, + 546, + 323 + ], + "spans": [ + { + "bbox": [ + 318, + 309, + 546, + 323 + ], + "type": "text", + "content": "14: " + }, + { + "bbox": [ + 318, + 309, + 546, + 323 + ], + "type": "inline_equation", + "content": "\\theta_{g}^{i}\\gets \\theta_{g}^{i} - \\frac{1}{|S_{i}|}\\sum_{j\\in S_{i}}G^{j}" + }, + { + "bbox": [ + 318, + 309, + 546, + 323 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 318, + 309, + 546, + 323 + ], + "type": "inline_equation", + "content": "i\\in \\{1,2,\\ldots ,n\\}" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 318, + 323, + 374, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 323, + 374, + 332 + ], + "spans": [ + { + "bbox": [ + 318, + 323, + 374, + 332 + ], + "type": "text", + "content": "15: end for" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 318, + 334, + 553, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 334, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 318, + 334, + 553, + 346 + ], + "type": "text", + "content": "16: Each client samples " + }, + { + "bbox": [ + 318, + 334, + 553, + 346 + ], + "type": "inline_equation", + "content": "M_t^c" + }, + { + "bbox": [ + 318, + 334, + 553, + 346 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 318, + 334, + 553, + 346 + ], + "type": "inline_equation", + "content": "D_t^c" + }, + { + "bbox": [ + 318, + 334, + 553, + 346 + ], + "type": "text", + "content": " by Eqs. (3) and (4);" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 318, + 346, + 375, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 346, + 375, + 356 + ], + "spans": [ + { + "bbox": [ + 318, + 346, + 375, + 356 + ], + "type": "text", + "content": "17: end while" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 377, + 555, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 555, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 555, + 401 + ], + "type": "text", + "content": "gradient update " + }, + { + "bbox": [ + 313, + 377, + 555, + 401 + ], + "type": "inline_equation", + "content": "g_{t}^{c}" + }, + { + "bbox": [ + 313, + 377, + 555, + 401 + ], + "type": "text", + "content": " of the new task and " + }, + { + "bbox": [ + 313, + 377, + 555, + 401 + ], + "type": "inline_equation", + "content": "g_{1:t-1}^{c}" + }, + { + "bbox": [ + 313, + 377, + 555, + 401 + ], + "type": "text", + "content": " from the old tasks and computed through" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 340, + 406, + 554, + 445 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 406, + 554, + 445 + ], + "spans": [ + { + "bbox": [ + 340, + 406, + 554, + 445 + ], + "type": "interline_equation", + "content": "w ^ {*} = \\arg \\min _ {w} \\left\\{\\frac {1}{t} \\sum_ {i} ^ {t} D \\left(\\sum_ {j \\neq i} ^ {t} w _ {j} g _ {j} ^ {c}, g _ {i} ^ {c}\\right) \\right\\}, \\tag {7}", + "image_path": "28e76bab8bcddcb14c64725c71ce953d23a9f1af5c75d39890ce8bc3dad33a01.jpg" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 450, + 554, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 554, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 554, + 523 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 450, + 554, + 523 + ], + "type": "inline_equation", + "content": "\\sum_{j=1}^{t} w_{j} = 1" + }, + { + "bbox": [ + 313, + 450, + 554, + 523 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 450, + 554, + 523 + ], + "type": "inline_equation", + "content": "\\forall w_{j} \\geq 0" + }, + { + "bbox": [ + 313, + 450, + 554, + 523 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 450, + 554, + 523 + ], + "type": "inline_equation", + "content": "D(\\sum_{j \\neq i}^{t} w_{j} g_{j}^{c}, g_{i}^{c})" + }, + { + "bbox": [ + 313, + 450, + 554, + 523 + ], + "type": "text", + "content": " represents the divergence between task " + }, + { + "bbox": [ + 313, + 450, + 554, + 523 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 450, + 554, + 523 + ], + "type": "text", + "content": " and the combination of all other tasks. By minimizing this divergence, we achieve the suppression of the influence from the combination on task " + }, + { + "bbox": [ + 313, + 450, + 554, + 523 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 450, + 554, + 523 + ], + "type": "text", + "content": ". Note that this optimization process does not modify individual gradient updates." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 523, + 554, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 523, + 554, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 523, + 554, + 570 + ], + "type": "text", + "content": "To prevent training failure on new tasks, we introduce the Multiple Gradient Descent Algorithm (MGDA) [4] to minimize the gradient norm, thereby achieving a balanced solution. It is defined as" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 375, + 574, + 554, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 574, + 554, + 609 + ], + "spans": [ + { + "bbox": [ + 375, + 574, + 554, + 609 + ], + "type": "interline_equation", + "content": "w _ {b} ^ {*} = \\underset {w} {\\arg \\min } \\left\\| \\sum_ {i} ^ {t} w _ {i} g _ {i} ^ {c} \\right\\| ^ {2}. \\tag {8}", + "image_path": "1d45543873e0375250796774b1fcc17a0885918793f01abd30a9e91c96be2277.jpg" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 613, + 554, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 613, + 554, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 613, + 554, + 661 + ], + "type": "text", + "content": "When the objective function value of MGDA is zero, the obtained solution satisfies the Karush-Kuhn-Tucker conditions. Otherwise, the solution provides a Pareto descent direction that does not harm any task." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 662, + 553, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 662, + 553, + 686 + ], + "spans": [ + { + "bbox": [ + 313, + 662, + 553, + 686 + ], + "type": "text", + "content": "In summary, our proposed Asymmetric Correction Gradient Updating (ACGU) algorithm determines " + }, + { + "bbox": [ + 313, + 662, + 553, + 686 + ], + "type": "inline_equation", + "content": "d_c^*" + }, + { + "bbox": [ + 313, + 662, + 553, + 686 + ], + "type": "text", + "content": " by" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 325, + 691, + 554, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 691, + 554, + 716 + ], + "spans": [ + { + "bbox": [ + 325, + 691, + 554, + 716 + ], + "type": "interline_equation", + "content": "d _ {c} ^ {*} = \\sum_ {i} W _ {i} g _ {i} ^ {c}, \\quad \\text {w h e r e} W = \\gamma w ^ {*} + (1 - \\gamma) w _ {b} ^ {*}. \\tag {9}", + "image_path": "6e5f0d2eba5819506df7255273d46333a1cfdb4fa3c6a97b5e07232af0c8952f.jpg" + } + ] + } + ], + "index": 36 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3845" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 555, + 229 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 555, + 229 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 555, + 229 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 555, + 229 + ], + "type": "table", + "html": "
AlgorithmAttributes FCL Per GCBTask ID under Non-IID Setting (Across 8 Different Datasets)\\( A_{avg} \\)(%)
\\( A_1 \\)(%)\\( A_2 \\)(%)\\( A_3 \\)(%)\\( A_4 \\)(%)\\( A_5 \\)(%)
FedAvg× × ×54.73 ± 1.4856.06 ± 3.1663.99 ± 3.4357.90 ± 2.9678.49 ± 2.1861.83 ± 0.46
FedProx× × ×58.25 ± 1.6056.65 ± 0.5066.15 ± 2.3361.72 ± 2.2182.24 ± 1.0265.00 ± 0.62
Ditto× ✓ ×58.92 ± 1.4059.13 ± 1.6370.58 ± 3.0763.00 ± 2.7983.97 ± 0.9567.12 ± 0.85
FedCurv✓ × ✓60.18 ± 3.4354.88 ± 1.9968.05 ± 3.7859.25 ± 2.3782.19 ± 0.5864.91 ± 0.56
FedAGEM✓ × ✓69.62 ± 0.8566.91 ± 1.0975.06 ± 1.6766.73 ± 2.4781.73 ± 1.1672.01 ± 0.81
FedER✓ × ×73.34 ± 0.4568.77 ± 1.8980.18 ± 1.3970.19 ± 1.3079.43 ± 1.3274.38 ± 0.27
FedMes✓ ✓ ✓70.84 ± 0.7967.22 ± 1.0577.25 ± 0.8967.18 ± 1.5285.21 ± 0.3673.54 ± 0.55
Ours-w/o-ACGU× ✓ ×58.62 ± 0.3658.06 ± 0.1665.62 ± 0.1963.29 ± 1.8682.46 ± 1.6865.61 ± 0.80
Ours-w/o-GAGU✓ × ✓76.19 ± 0.5271.59 ± 1.1282.24 ± 1.4272.89 ± 2.0481.82 ± 2.3676.95 ± 1.38
Ours-w/o-CBRSS✓ ✓ ✓75.40 ± 0.9072.39 ± 0.5582.88 ± 0.9275.42 ± 0.5486.68 ± 0.6278.56 ± 0.57
FedAGC (Ours)✓ ✓ ✓76.21 ± 0.3672.56 ± 0.4884.90 ± 0.8577.60 ± 0.5787.28 ± 0.1879.71 ± 0.23
", + "image_path": "2d478b58aaf089f4cbbd11d634db7052a0618d900114a8414df24e262e1c0894.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 237, + 555, + 270 + ], + "lines": [ + { + "bbox": [ + 55, + 237, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 237, + 555, + 270 + ], + "type": "text", + "content": "Table 1. Performance (avg ± std) of local models for each task on 10 clients under Non-IID setting. \"w/o-Module\" indicates performance without the corresponding module. It also includes Per (Personalized) and GCB (Gradient Correction-Based). For the proposed FedAGC, we used " + }, + { + "bbox": [ + 55, + 237, + 555, + 270 + ], + "type": "inline_equation", + "content": "\\lambda = 0.01" + }, + { + "bbox": [ + 55, + 237, + 555, + 270 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 237, + 555, + 270 + ], + "type": "inline_equation", + "content": "\\gamma = 0.5" + }, + { + "bbox": [ + 55, + 237, + 555, + 270 + ], + "type": "text", + "content": " in Eqs. (1) and (9)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 276, + 212, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 276, + 212, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 276, + 212, + 289 + ], + "type": "text", + "content": "3.4. Group-Level Personalization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 293, + 295, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 293, + 295, + 400 + ], + "spans": [ + { + "bbox": [ + 54, + 293, + 295, + 400 + ], + "type": "text", + "content": "To integrate heterogeneous knowledge from different clients, some studies [39, 45] adopt personalization strategies at the client level, where each client continuously adjusts the received global model to better fit its local tasks. However, this approach struggles to meet the needs of all clients in large-scale parallel task scenarios. In contrast, we consider a group-level personalization strategy, enabling the broadcast of global models tailored to meet the specific requirements of each client." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 401, + 296, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 296, + 498 + ], + "type": "text", + "content": "In this work, we propose Group Aggregation Gradient Updating (GAGU) for client-uploaded gradients. Specifically, during the " + }, + { + "bbox": [ + 55, + 401, + 296, + 498 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 401, + 296, + 498 + ], + "type": "text", + "content": "-th communication round, " + }, + { + "bbox": [ + 55, + 401, + 296, + 498 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 401, + 296, + 498 + ], + "type": "text", + "content": " clients upload their respective averaged gradient updates " + }, + { + "bbox": [ + 55, + 401, + 296, + 498 + ], + "type": "inline_equation", + "content": "\\{G^1,G^2,\\dots ,G^N\\}" + }, + { + "bbox": [ + 55, + 401, + 296, + 498 + ], + "type": "text", + "content": " obtained after " + }, + { + "bbox": [ + 55, + 401, + 296, + 498 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 55, + 401, + 296, + 498 + ], + "type": "text", + "content": " local training iterations, where " + }, + { + "bbox": [ + 55, + 401, + 296, + 498 + ], + "type": "inline_equation", + "content": "G^{c} = -\\triangle \\theta_{l,c}^{t} / \\alpha" + }, + { + "bbox": [ + 55, + 401, + 296, + 498 + ], + "type": "text", + "content": ". We compute a similarity matrix " + }, + { + "bbox": [ + 55, + 401, + 296, + 498 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 401, + 296, + 498 + ], + "type": "text", + "content": " to measure the similarity between pairs of averaged gradient updates. The elements of " + }, + { + "bbox": [ + 55, + 401, + 296, + 498 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 401, + 296, + 498 + ], + "type": "text", + "content": " are given by" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 135, + 505, + 294, + 532 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 505, + 294, + 532 + ], + "spans": [ + { + "bbox": [ + 135, + 505, + 294, + 532 + ], + "type": "interline_equation", + "content": "s _ {i, j} = \\frac {G ^ {i} G ^ {j}}{\\| G ^ {i} \\| \\| G ^ {j} \\|.} \\tag {10}", + "image_path": "0e5d7ea31041f584f7920a2db8a3e830ca5f3f9d04905c6b7509d963686198aa.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "text", + "content": "We construct a graph " + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "inline_equation", + "content": "P = \\{G, E\\}" + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "text", + "content": " to represent " + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "text", + "content": " denotes the set of nodes in the graph, and each node corresponds to the " + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "inline_equation", + "content": "G^c" + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "text", + "content": ". The set of edges " + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "inline_equation", + "content": "E = \\{s_{i,j}\\} \\forall i \\leq j \\leq N" + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "text", + "content": " represents the similarity between " + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "inline_equation", + "content": "G^i" + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "inline_equation", + "content": "G^j" + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "text", + "content": ". The objective of GAGU is to reorder " + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "text", + "content": " such that similar average gradient updates are grouped together, while dissimilar ones are isolated. To achieve this, we first define the similarity between any two groups " + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "inline_equation", + "content": "S_A" + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "inline_equation", + "content": "S_B" + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "inline_equation", + "content": "S_A \\cap S_B = \\emptyset" + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "inline_equation", + "content": "\\forall A \\neq B" + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "text", + "content": "):" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 643, + 295, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 643, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 115, + 643, + 295, + 669 + ], + "type": "interline_equation", + "content": "h \\left(S _ {A}, S _ {B}\\right) = \\sum_ {i \\in S _ {A}, j \\in S _ {B}} s _ {i, j}. \\tag {11}", + "image_path": "20247b2c3dbbdbfb6811275e8dc12bb7f8ed3a2ac573b836efd6e59a0142936e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "GAGU aims to maximize intra-group similarity while minimizing inter-group similarity, resulting in an optimal graph partition " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "a^{*} = \\{S_{1}, S_{2}, \\ldots, S_{n}\\}" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " represents the" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 277, + 555, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 555, + 312 + ], + "type": "text", + "content": "number of groups in " + }, + { + "bbox": [ + 313, + 277, + 555, + 312 + ], + "type": "inline_equation", + "content": "a^*" + }, + { + "bbox": [ + 313, + 277, + 555, + 312 + ], + "type": "text", + "content": ". Inspired by spectrum-clustering [13] and Ncut [36], We reformulate this optimization problem into the following form:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 367, + 318, + 555, + 344 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 318, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 367, + 318, + 555, + 344 + ], + "type": "interline_equation", + "content": "\\min _ {a \\in A} \\sum_ {S, S ^ {\\prime} \\in a} h \\left(S, S ^ {\\prime}\\right) - h \\left(S ^ {\\prime}, S ^ {\\prime}\\right), \\tag {12}", + "image_path": "228cd333536aa8de077912bc0c6e9e987adc6eabbf18d81b779e63c16a367d7f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 350, + 554, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 554, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 554, + 422 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 350, + 554, + 422 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 350, + 554, + 422 + ], + "type": "text", + "content": " is all possible grouping results. Additionally, we introduce the modularity " + }, + { + "bbox": [ + 313, + 350, + 554, + 422 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 350, + 554, + 422 + ], + "type": "text", + "content": " to evaluate the quality of the current partition during the process of determining the optimal number of groups and partitioning results. This metric measures the distribution of connectivity within and between groups." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 354, + 426, + 555, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 426, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 354, + 426, + 555, + 456 + ], + "type": "interline_equation", + "content": "Q = \\frac {1}{2 W} \\sum_ {i, j} \\left[ s _ {i, j} - \\frac {d _ {i} d _ {j}}{2 W} \\right] \\delta \\left(S _ {i}, S _ {j}\\right), \\tag {13}", + "image_path": "3cba87e5a81b899ae0d420f4b67588176e810d2f7739804e5d41a2cc3c182c35.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 463, + 554, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 554, + 500 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 554, + 500 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 463, + 554, + 500 + ], + "type": "inline_equation", + "content": "W = \\sum_{i,j} s_{i,j}" + }, + { + "bbox": [ + 313, + 463, + 554, + 500 + ], + "type": "text", + "content": " represents the total edge weight in the current graph, " + }, + { + "bbox": [ + 313, + 463, + 554, + 500 + ], + "type": "inline_equation", + "content": "d_i = \\sum_j s_{i,j}" + }, + { + "bbox": [ + 313, + 463, + 554, + 500 + ], + "type": "text", + "content": " denotes the degree of node " + }, + { + "bbox": [ + 313, + 463, + 554, + 500 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 463, + 554, + 500 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 463, + 554, + 500 + ], + "type": "inline_equation", + "content": "\\delta(S_i, S_j)" + }, + { + "bbox": [ + 313, + 463, + 554, + 500 + ], + "type": "text", + "content": " is the Kronecker delta function." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 509, + 397, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 397, + 522 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 397, + 522 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 529, + 432, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 432, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 432, + 541 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": "Benchmarks. We validate FedAGC in large-scale multitask parallel FCL scenarios. The experimental setup follows the Task-Incremental Learning paradigm and primarily includes two different benchmarks: (1) IID Setting (Single CIFAR-100): The 100 classes of CIFAR-100 [16] are divided into 20 tasks, each containing 5 classes. These tasks are randomly assigned to clients, ensuring that each client does not train on the same task multiple times. (2) Non-IID Setting (Across 8 Different Datasets): We partition and distribute eight widely used datasets in the FCL domain, such as CIFAR-10/100 [16], MNIST [18], SVHN [30], FashionMNIST [41], Not-MNIST [1], FaceScrub [32], and TrafficSigns [31], in the same manner as (1). In both benchmarks, each client is assigned a unique task sequence." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3846" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 555, + 229 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 555, + 229 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 555, + 229 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 555, + 229 + ], + "type": "table", + "html": "
AlgorithmAttributes FCL Per GCBTask ID under IID Setting (Single CIFAR-100)\\( A_{avg} \\)(%)
\\( A_1 \\)(%)\\( A_2 \\)(%)\\( A_3 \\)(%)\\( A_4 \\)(%)\\( A_5 \\)(%)
FedAvg× × ×54.32 ± 0.9460.53 ± 2.0861.89 ± 1.5665.63 ± 1.6070.77 ± 0.8462.63 ± 0.55
FedProx× × ×53.96 ± 0.9861.53 ± 2.2362.90 ± 1.9465.70 ± 1.1370.78 ± 0.9162.97 ± 0.93
Ditto× ✓ ×55.48 ± 0.9262.81 ± 1.9663.12 ± 1.5866.59 ± 1.8671.01 ± 0.9363.67 ± 0.72
FedCurv✓ × ✓54.36 ± 0.9560.77 ± 0.8663.43 ± 0.9166.46 ± 0.9771.52 ± 0.7563.31 ± 0.16
FedAGEM✓ × ✓60.95 ± 0.9067.15 ± 1.2366.39 ± 0.6967.81 ± 0.9071.28 ± 0.6666.72 ± 0.23
FedER✓ × ×63.05 ± 0.7667.83 ± 0.9166.11 ± 0.3265.39 ± 0.4468.84 ± 1.0166.44 ± 0.42
FedMes✓ ✓ ✓61.18 ± 0.3467.68 ± 0.1167.11 ± 0.4967.42 ± 0.1572.12 ± 0.9067.13 ± 0.16
Ours-w/o-ACGU× ✓ ×55.45 ± 0.2761.95 ± 0.8763.32 ± 1.7265.51 ± 0.7371.44 ± 0.6063.53 ± 0.31
Ours-w/o-GAGU✓ × ✓63.92 ± 0.5566.73 ± 0.9367.22 ± 0.6166.48 ± 0.6568.53 ± 0.2466.58 ± 0.41
Ours-w/o-CBRSS✓ ✓ ✓66.14 ± 0.3769.05 ± 0.8069.69 ± 0.4969.56 ± 2.2074.12 ± 0.0969.75 ± 0.78
FedAGC (Ours)✓ ✓ ✓66.38 ± 0.4869.48 ± 0.7870.38 ± 1.3270.58 ± 1.8573.98 ± 0.4170.16 ± 0.77
", + "image_path": "817087d984ad0d80005ca64d6ef802a09904874b1459d3fd45954f26e0390054.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 237, + 476, + 248 + ], + "lines": [ + { + "bbox": [ + 132, + 237, + 476, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 476, + 248 + ], + "type": "text", + "content": "Table 2. Performance (avg ± std) of local models for each task on 10 clients under IID setting." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 58, + 253, + 175, + 369 + ], + "blocks": [ + { + "bbox": [ + 58, + 253, + 175, + 369 + ], + "lines": [ + { + "bbox": [ + 58, + 253, + 175, + 369 + ], + "spans": [ + { + "bbox": [ + 58, + 253, + 175, + 369 + ], + "type": "image", + "image_path": "065edf3297a1e284be73cca5f916af60ca88546822aaf1d5946914a4ef13402a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 68, + 373, + 164, + 383 + ], + "lines": [ + { + "bbox": [ + 68, + 373, + 164, + 383 + ], + "spans": [ + { + "bbox": [ + 68, + 373, + 164, + 383 + ], + "type": "text", + "content": "(a) " + }, + { + "bbox": [ + 68, + 373, + 164, + 383 + ], + "type": "inline_equation", + "content": "F_{t}" + }, + { + "bbox": [ + 68, + 373, + 164, + 383 + ], + "type": "text", + "content": " under Non-IID Setting" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 85, + 384, + 265, + 396 + ], + "lines": [ + { + "bbox": [ + 85, + 384, + 265, + 396 + ], + "spans": [ + { + "bbox": [ + 85, + 384, + 265, + 396 + ], + "type": "text", + "content": "Figure 3. Comparison of average forgetting rates." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 177, + 253, + 294, + 369 + ], + "blocks": [ + { + "bbox": [ + 177, + 253, + 294, + 369 + ], + "lines": [ + { + "bbox": [ + 177, + 253, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 177, + 253, + 294, + 369 + ], + "type": "image", + "image_path": "4b5ee6f8a9295d9655513e46cba23c8834e4f81d409c631b3361571035ae705c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 195, + 373, + 276, + 383 + ], + "lines": [ + { + "bbox": [ + 195, + 373, + 276, + 383 + ], + "spans": [ + { + "bbox": [ + 195, + 373, + 276, + 383 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 195, + 373, + 276, + 383 + ], + "type": "inline_equation", + "content": "F_{t}" + }, + { + "bbox": [ + 195, + 373, + 276, + 383 + ], + "type": "text", + "content": " under IID Setting" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 398, + 296, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 398, + 296, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 398, + 296, + 493 + ], + "type": "text", + "content": "Baselines. We compare FedAGC with state-of-the-art (SOTA) methods, including the fundamental FedAvg [29] and its proximal variant FedProx [19], along with the personalized FL method Ditto [20]. To verify the superiority of FedAGC, we also compare it with gradient correction-based FCL methods, including FedCurv [38], FedAGEM [2] and the latest FedMes [42]. Additionally, we consider the rehearsal-based method FedER [3]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 495, + 296, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 296, + 650 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 296, + 650 + ], + "type": "text", + "content": "Metrics. We employ two critical metrics to evaluate algorithm performance: Average Accuracy " + }, + { + "bbox": [ + 55, + 495, + 296, + 650 + ], + "type": "inline_equation", + "content": "(A_{avg})" + }, + { + "bbox": [ + 55, + 495, + 296, + 650 + ], + "type": "text", + "content": " and Forgetting Rate " + }, + { + "bbox": [ + 55, + 495, + 296, + 650 + ], + "type": "inline_equation", + "content": "(F_t)" + }, + { + "bbox": [ + 55, + 495, + 296, + 650 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 55, + 495, + 296, + 650 + ], + "type": "inline_equation", + "content": "a_{i,t}^c" + }, + { + "bbox": [ + 55, + 495, + 296, + 650 + ], + "type": "text", + "content": " denote the test accuracy of the " + }, + { + "bbox": [ + 55, + 495, + 296, + 650 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 495, + 296, + 650 + ], + "type": "text", + "content": "-th task learned by the " + }, + { + "bbox": [ + 55, + 495, + 296, + 650 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 495, + 296, + 650 + ], + "type": "text", + "content": "-th client after sequential learning of " + }, + { + "bbox": [ + 55, + 495, + 296, + 650 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 495, + 296, + 650 + ], + "type": "text", + "content": " subsequent tasks, where " + }, + { + "bbox": [ + 55, + 495, + 296, + 650 + ], + "type": "inline_equation", + "content": "t + i \\leq T" + }, + { + "bbox": [ + 55, + 495, + 296, + 650 + ], + "type": "text", + "content": ". In this formulation, the initial accuracy of task " + }, + { + "bbox": [ + 55, + 495, + 296, + 650 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 495, + 296, + 650 + ], + "type": "text", + "content": " (immediately after its own training) and the final accuracy after training on all tasks are defined as " + }, + { + "bbox": [ + 55, + 495, + 296, + 650 + ], + "type": "inline_equation", + "content": "a_{i,0}^c" + }, + { + "bbox": [ + 55, + 495, + 296, + 650 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 495, + 296, + 650 + ], + "type": "inline_equation", + "content": "a_{i,T - i}^c" + }, + { + "bbox": [ + 55, + 495, + 296, + 650 + ], + "type": "text", + "content": ", respectively. The forgetting rate for the " + }, + { + "bbox": [ + 55, + 495, + 296, + 650 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 495, + 296, + 650 + ], + "type": "text", + "content": "-th task after learning " + }, + { + "bbox": [ + 55, + 495, + 296, + 650 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 495, + 296, + 650 + ], + "type": "text", + "content": " subsequent tasks is calculated as: " + }, + { + "bbox": [ + 55, + 495, + 296, + 650 + ], + "type": "inline_equation", + "content": "f_{i,t}^{c} = \\frac{a_{i,0}^{c} - a_{i,t}^{c}}{a_{i,0}^{c}}" + }, + { + "bbox": [ + 55, + 495, + 296, + 650 + ], + "type": "text", + "content": ". To aggregate results, we compute client-wise averages: " + }, + { + "bbox": [ + 55, + 495, + 296, + 650 + ], + "type": "inline_equation", + "content": "A_i = \\frac{1}{N}\\sum_{c=1}^N a_{i,T-i}^c" + }, + { + "bbox": [ + 55, + 495, + 296, + 650 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 495, + 296, + 650 + ], + "type": "inline_equation", + "content": "f_{i,t} = \\frac{1}{N}\\sum_{c=1}^N f_{i,t}^c" + }, + { + "bbox": [ + 55, + 495, + 296, + 650 + ], + "type": "text", + "content": ". Then, " + }, + { + "bbox": [ + 55, + 495, + 296, + 650 + ], + "type": "inline_equation", + "content": "A_{avg}" + }, + { + "bbox": [ + 55, + 495, + 296, + 650 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 495, + 296, + 650 + ], + "type": "inline_equation", + "content": "F_t" + }, + { + "bbox": [ + 55, + 495, + 296, + 650 + ], + "type": "text", + "content": " are defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 90, + 659, + 295, + 692 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 659, + 295, + 692 + ], + "spans": [ + { + "bbox": [ + 90, + 659, + 295, + 692 + ], + "type": "interline_equation", + "content": "A _ {\\text {a v g}} = \\frac {1}{T} \\sum_ {i = 1} ^ {T} A _ {i}, \\quad F _ {t} = \\frac {1}{T - t} \\sum_ {i = 1} ^ {T - t} f _ {i, t}. \\tag {14}", + "image_path": "5b813c45b213a2f40f7b62dd38e8e35c2313ae229f3f08f6bddbab3f76250728.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "content": "Implementation details. Following the experimental setup" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 254, + 429, + 326 + ], + "blocks": [ + { + "bbox": [ + 317, + 254, + 429, + 326 + ], + "lines": [ + { + "bbox": [ + 317, + 254, + 429, + 326 + ], + "spans": [ + { + "bbox": [ + 317, + 254, + 429, + 326 + ], + "type": "image", + "image_path": "ed150f3019ad6825d5ac902e367243c47790fbf8c4a456df9789eb7fb1dfa2b9.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 317, + 338, + 432, + 414 + ], + "blocks": [ + { + "bbox": [ + 327, + 329, + 417, + 338 + ], + "lines": [ + { + "bbox": [ + 327, + 329, + 417, + 338 + ], + "spans": [ + { + "bbox": [ + 327, + 329, + 417, + 338 + ], + "type": "text", + "content": "(a) Impact of " + }, + { + "bbox": [ + 327, + 329, + 417, + 338 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 327, + 329, + 417, + 338 + ], + "type": "text", + "content": " on FedAGC" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 338, + 432, + 414 + ], + "lines": [ + { + "bbox": [ + 317, + 338, + 432, + 414 + ], + "spans": [ + { + "bbox": [ + 317, + 338, + 432, + 414 + ], + "type": "image", + "image_path": "d11c41af08c1f352b50d3273bdb02d13ba31314a436c9be0da172a1e46db85a3.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 441, + 254, + 552, + 326 + ], + "blocks": [ + { + "bbox": [ + 441, + 254, + 552, + 326 + ], + "lines": [ + { + "bbox": [ + 441, + 254, + 552, + 326 + ], + "spans": [ + { + "bbox": [ + 441, + 254, + 552, + 326 + ], + "type": "image", + "image_path": "86e4413525e67630fc5d1bf946e0fd9c18c4829f2f90681d0e32be8ce15f94fe.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 436, + 338, + 552, + 414 + ], + "blocks": [ + { + "bbox": [ + 451, + 329, + 542, + 338 + ], + "lines": [ + { + "bbox": [ + 451, + 329, + 542, + 338 + ], + "spans": [ + { + "bbox": [ + 451, + 329, + 542, + 338 + ], + "type": "text", + "content": "(b) Impact of " + }, + { + "bbox": [ + 451, + 329, + 542, + 338 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 451, + 329, + 542, + 338 + ], + "type": "text", + "content": " on FedAGC" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 436, + 338, + 552, + 414 + ], + "lines": [ + { + "bbox": [ + 436, + 338, + 552, + 414 + ], + "spans": [ + { + "bbox": [ + 436, + 338, + 552, + 414 + ], + "type": "image", + "image_path": "aaa4303fe61e51880f01edfafa2d12c9ac4e94b2a13e4fc9863b12a8349fca2e.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 317, + 428, + 432, + 503 + ], + "blocks": [ + { + "bbox": [ + 324, + 418, + 424, + 427 + ], + "lines": [ + { + "bbox": [ + 324, + 418, + 424, + 427 + ], + "spans": [ + { + "bbox": [ + 324, + 418, + 424, + 427 + ], + "type": "text", + "content": "(c) Non-IID Setting, " + }, + { + "bbox": [ + 324, + 418, + 424, + 427 + ], + "type": "inline_equation", + "content": "\\lambda = 0.01" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 428, + 432, + 503 + ], + "lines": [ + { + "bbox": [ + 317, + 428, + 432, + 503 + ], + "spans": [ + { + "bbox": [ + 317, + 428, + 432, + 503 + ], + "type": "image", + "image_path": "225cb13e94af64123351d55667487139ec065af95c06600e19b7d0bd165bdfdb.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 507, + 416, + 517 + ], + "lines": [ + { + "bbox": [ + 333, + 507, + 416, + 517 + ], + "spans": [ + { + "bbox": [ + 333, + 507, + 416, + 517 + ], + "type": "text", + "content": "(e) IID Setting, " + }, + { + "bbox": [ + 333, + 507, + 416, + 517 + ], + "type": "inline_equation", + "content": "\\lambda = 0.01" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 333, + 519, + 534, + 530 + ], + "lines": [ + { + "bbox": [ + 333, + 519, + 534, + 530 + ], + "spans": [ + { + "bbox": [ + 333, + 519, + 534, + 530 + ], + "type": "text", + "content": "Figure 4. Impact of core hyperparameters on FedAGC." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 436, + 428, + 552, + 503 + ], + "blocks": [ + { + "bbox": [ + 451, + 418, + 539, + 427 + ], + "lines": [ + { + "bbox": [ + 451, + 418, + 539, + 427 + ], + "spans": [ + { + "bbox": [ + 451, + 418, + 539, + 427 + ], + "type": "text", + "content": "(d) Non-IID Setting, " + }, + { + "bbox": [ + 451, + 418, + 539, + 427 + ], + "type": "inline_equation", + "content": "\\gamma = 0.5" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 436, + 428, + 552, + 503 + ], + "lines": [ + { + "bbox": [ + 436, + 428, + 552, + 503 + ], + "spans": [ + { + "bbox": [ + 436, + 428, + 552, + 503 + ], + "type": "image", + "image_path": "6a9898dc138632433d22d7fcc7f82a0fe26f29ba033401bff1c58d18abfc4297.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 459, + 507, + 529, + 517 + ], + "lines": [ + { + "bbox": [ + 459, + 507, + 529, + 517 + ], + "spans": [ + { + "bbox": [ + 459, + 507, + 529, + 517 + ], + "type": "text", + "content": "(f) IID Setting, " + }, + { + "bbox": [ + 459, + 507, + 529, + 517 + ], + "type": "inline_equation", + "content": "\\gamma = 0.5" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 533, + 555, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 666 + ], + "type": "text", + "content": "of [47], we adopted a variant of the LeNet architecture as the backbone. For any newly introduced task, we initialized a new classifier. The Adam optimizer was used for all datasets, with an initial learning rate of " + }, + { + "bbox": [ + 313, + 533, + 555, + 666 + ], + "type": "inline_equation", + "content": "10^{-3}" + }, + { + "bbox": [ + 313, + 533, + 555, + 666 + ], + "type": "text", + "content": " and a weight decay of " + }, + { + "bbox": [ + 313, + 533, + 555, + 666 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 313, + 533, + 555, + 666 + ], + "type": "text", + "content": ". The experimental setup consisted of 10 clients, each learning 5 sequential and non-overlapping tasks. The batch size was set to 100 per step, with " + }, + { + "bbox": [ + 313, + 533, + 555, + 666 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 533, + 555, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 533, + 555, + 666 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 533, + 555, + 666 + ], + "type": "text", + "content": " set to 0.1 and 8, respectively. To ensure convergence, all experiments were conducted for 100 federated communication rounds. Each experiment was repeated five times, and the reported results were averaged accordingly." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 672, + 399, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 399, + 683 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 399, + 683 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Tabs. 1 and 2 record the average accuracy of 10 clients on each task under two data condition setups, along with the" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3847" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 348 + ], + "type": "text", + "content": "final results considering all tasks. In Fig. 3, we compare the degree of Spatial-TemporalCF between FedAGC and various methods. It can be observed that FedProx and Ditto mitigate the impact of statistical heterogeneity using penalty terms, showing significant performance improvements over FedAvg in Non-IID settings. However, since they do not employ any CL strategies, their accuracy is lower compared to FCL methods such as FedAGEM, FedER, and FedMes, with higher forgetting levels. Notably, FedCurv's experimental results are suboptimal as it leverages curvature information from other clients to guide updates for the current client, which is prone to knowledge conflicts in cases of non-overlapping client data distributions. These FCL methods based on symmetric gradient correction focus solely on TemporalCF without considering the interference of knowledge across clients, leading to SpatialCF. In contrast, the proposed FedAGC outperforms other methods in both task-specific accuracy and overall performance across both settings, with a lower degree of forgetting. This demonstrates that memory buffers constructed through importance sampling, assisted by asymmetric gradient correction and group-level personalization, can significantly improve the accuracy of PFCL models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 358, + 237, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 237, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 237, + 371 + ], + "type": "text", + "content": "4.3. Performance Analysis of FedAGC" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 377, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 377, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 295, + 664 + ], + "type": "text", + "content": "Ablation Studies. We conducted three different ablation studies under the same experimental setup, with the results recorded in Tabs. 1 and 2, as well as Fig. 3. We progressively removed components from FedAGC, including: (1) Omitting the ACGU module (FedAGC-w/o-ACGU), which means that only the gradients from local training are used to guide updates. Clearly, in this case, the CBRSS module also loses its significance. The performance degradation is significant, with a drop of " + }, + { + "bbox": [ + 55, + 377, + 295, + 664 + ], + "type": "inline_equation", + "content": "17.7\\%" + }, + { + "bbox": [ + 55, + 377, + 295, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 377, + 295, + 664 + ], + "type": "inline_equation", + "content": "9.5\\%" + }, + { + "bbox": [ + 55, + 377, + 295, + 664 + ], + "type": "text", + "content": ", as no strategy for mitigating TemporalCF is applied. (2) Removing the GAGU module (FedAGC-w/o-GAGU) and switching to the FedAvg average aggregation method. In the IID setting, FedAGC-w/o-GAGU's overall performance is overtaken by FedMes by " + }, + { + "bbox": [ + 55, + 377, + 295, + 664 + ], + "type": "inline_equation", + "content": "0.86\\%" + }, + { + "bbox": [ + 55, + 377, + 295, + 664 + ], + "type": "text", + "content": ". (3) Disabling the CBRSS module (FedAGC-w/o-CBRSS) and replacing it with a random sampling method. In this case, performance decreases by " + }, + { + "bbox": [ + 55, + 377, + 295, + 664 + ], + "type": "inline_equation", + "content": "1.44\\%" + }, + { + "bbox": [ + 55, + 377, + 295, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 377, + 295, + 664 + ], + "type": "inline_equation", + "content": "0.58\\%" + }, + { + "bbox": [ + 55, + 377, + 295, + 664 + ], + "type": "text", + "content": " in the Non-IID and IID settings, respectively. However, since the core component for mitigating forgetting is retained, its accuracy on individual tasks and overall performance remains superior to other baselines. These results verify that our three proposed modules of FedAGC play a crucial role in alleviating Spatial-TemporalCF and achieving personalization, and can effectively be combined to enhance performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Impact of core hyperparameters. We evaluate FedAGC with different values of " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ". As shown in Fig. 4, setting " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " too large or too small causes FedAGC to overly rely on either special or balanced solution, impairing its ability to" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 350, + 72, + 519, + 184 + ], + "blocks": [ + { + "bbox": [ + 350, + 72, + 519, + 184 + ], + "lines": [ + { + "bbox": [ + 350, + 72, + 519, + 184 + ], + "spans": [ + { + "bbox": [ + 350, + 72, + 519, + 184 + ], + "type": "image", + "image_path": "8afb4c9a20a3e080b95b48912d0f19e7fb31dbbf3440484387aa96ea12bdf009.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 186, + 501, + 196 + ], + "lines": [ + { + "bbox": [ + 369, + 186, + 501, + 196 + ], + "spans": [ + { + "bbox": [ + 369, + 186, + 501, + 196 + ], + "type": "text", + "content": "(a) " + }, + { + "bbox": [ + 369, + 186, + 501, + 196 + ], + "type": "inline_equation", + "content": "A_{i}" + }, + { + "bbox": [ + 369, + 186, + 501, + 196 + ], + "type": "text", + "content": " on the CIFAR-100 dataset, " + }, + { + "bbox": [ + 369, + 186, + 501, + 196 + ], + "type": "inline_equation", + "content": "N = 30" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 350, + 198, + 517, + 306 + ], + "blocks": [ + { + "bbox": [ + 350, + 198, + 517, + 306 + ], + "lines": [ + { + "bbox": [ + 350, + 198, + 517, + 306 + ], + "spans": [ + { + "bbox": [ + 350, + 198, + 517, + 306 + ], + "type": "image", + "image_path": "cf5e28fab530917c16438c069921c6bfdc0f579e4ca5820b5f39f1ebaecc235c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 311, + 500, + 320 + ], + "lines": [ + { + "bbox": [ + 369, + 311, + 500, + 320 + ], + "spans": [ + { + "bbox": [ + 369, + 311, + 500, + 320 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 369, + 311, + 500, + 320 + ], + "type": "inline_equation", + "content": "F_{t}" + }, + { + "bbox": [ + 369, + 311, + 500, + 320 + ], + "type": "text", + "content": " on the CIFAR-100 dataset, " + }, + { + "bbox": [ + 369, + 311, + 500, + 320 + ], + "type": "inline_equation", + "content": "N = 30" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 352, + 322, + 515, + 334 + ], + "lines": [ + { + "bbox": [ + 352, + 322, + 515, + 334 + ], + "spans": [ + { + "bbox": [ + 352, + 322, + 515, + 334 + ], + "type": "text", + "content": "Figure 5. Performance on large-scale clients." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 335, + 555, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 555, + 396 + ], + "type": "text", + "content": "mitigate forgetting. Similarly, an excessively large " + }, + { + "bbox": [ + 313, + 335, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 335, + 555, + 396 + ], + "type": "text", + "content": " biases the model toward the inversion of previous tasks, hindering the learning of current task, while a small " + }, + { + "bbox": [ + 313, + 335, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 335, + 555, + 396 + ], + "type": "text", + "content": " has the opposite effect. Empirically, we find that the best performance of FedAGC is achieved when " + }, + { + "bbox": [ + 313, + 335, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\gamma = 0.5" + }, + { + "bbox": [ + 313, + 335, + 555, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 335, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\lambda = 0.01" + }, + { + "bbox": [ + 313, + 335, + 555, + 396 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 396, + 556, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 556, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 556, + 492 + ], + "type": "text", + "content": "Performance on large-scale clients. To explore the scalability of FedAGC to larger client scenarios, we tripled the number of clients from the previous setting of 10 and continued comparing it with various SOTA methods. The results recorded in Fig. 5 indicate that FedAGC outperforms other baselines in accuracy on previous tasks and demonstrates greater stability in mitigating forgetting. This result highlights the superiority of FedAGC." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 502, + 388, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 388, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 388, + 514 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": "To tackle the challenge of STCF in PFCL, this paper proposed FedAGC as a valid solution. After local training, FedAGC selects representative samples and stores them in a buffer for rehearsal to alleviate forgetting. To refine this process, we introduced an asymmetric gradient metric to quantify discrepancies between past and current tasks. Asymmetric gradient correction is then employed to adjust the gradient updates, jointly considering direction and magnitude for optimal adaptation. Moreover, FedAGC adopts group-level personalization by clustering clients based on the similarity of their uploaded averaged gradient updates, which enhances the fusion of valuable knowledge while isolating ineffective knowledge. It was extensively demonstrated in this paper that the effective integration of the proposed modules can enhance the model's resistance to STCF, enabling FedAGC to outperform SOTA approaches." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3848" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 128, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 128, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 128, + 85 + ], + "type": "text", + "content": "Acknowledge" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 236 + ], + "type": "text", + "content": "We want to thank the anonymous reviewers for their valuable suggestions and comments. This work was supported by the National Natural Science Foundation of China (No. 62276182), Tianjin University under the Emerging Direction Cultivation Project of Interdisciplinary Center (Intelligent Protection and Utilization of Digital Cultural Heritage), National Key Research and Development Program of China (No. 2023YFF0906204), Peng Cheng Lab Program (No. PCL2023A08), Tianjin Natural Science Foundation (Nos. 24JCYBJC01230, 24JCYBJC01460), and Tianjin Municipal Education Commission Research Plan (No. 2024ZX008)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 245, + 115, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 245, + 115, + 258 + ], + "spans": [ + { + "bbox": [ + 56, + 245, + 115, + 258 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 266, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 266, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 266, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 61, + 266, + 294, + 297 + ], + "type": "text", + "content": "[1] Yaroslav Bulatov. Notmnist dataset. http://yaroslavvb.blogspot.it/2011/09/notmnist-dataset.html.6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 300, + 294, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 300, + 294, + 343 + ], + "spans": [ + { + "bbox": [ + 61, + 300, + 294, + 343 + ], + "type": "text", + "content": "[2] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a-gem. In Proceedings of the International Conference on Learning Representations, 2018. 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 345, + 294, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 345, + 294, + 399 + ], + "spans": [ + { + "bbox": [ + 62, + 345, + 294, + 399 + ], + "type": "text", + "content": "[3] Arslan Chaudhry, Marcus Rohrbach, Mohamed Elhoseiny, Thalaiyasingam Ajanthan, Puneet K. Dokania, Philip H. S. Torr, and MarcÁurelio Ranzato. On tiny episodic memories in continual learning. In arXiv Preprint arXiv:1902.10486, 2019. 3, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 401, + 295, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 401, + 295, + 433 + ], + "spans": [ + { + "bbox": [ + 62, + 401, + 295, + 433 + ], + "type": "text", + "content": "[4] Désideri and Jean-Antoine. Multiple-gradient descent algorithm (mgda) for multiobjective optimization. Comptes Rendus Mathematique, 350(5-6):313-318, 2012. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 434, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 434, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 62, + 434, + 294, + 487 + ], + "type": "text", + "content": "[5] Prithviraj Dhar, Rajat Vikram Singh, Kuan-Chuan Peng, Ziyan Wu, and Rama Chellappa. Learning without memorizing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5138-5146, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 491, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 294, + 544 + ], + "type": "text", + "content": "[6] Jiahua Dong, Lixu Wang, Zhen Fang, Gan Sun, Shichao Xu, Xiao Wang, and Qi Zhu. Federated class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10164-10173, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 546, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 546, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 62, + 546, + 294, + 590 + ], + "type": "text", + "content": "[7] Alireza Fallah, Aryan Mokhtari, and Asuman Ozdaglar. Personalized federated learning with theoretical guarantees: A model-agnostic meta-learning approach. Advances in neural information processing systems, 33:3557–3568, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 591, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 591, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 62, + 591, + 294, + 635 + ], + "type": "text", + "content": "[8] Mehrdad Farajtabar, Navig Azizan, Alex Mott, and Ang Li. Orthogonal gradient descent for continual learning. In International Conference on Artificial Intelligence and Statistics, PMLR, pages 3762-3773, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 636, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 636, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 62, + 636, + 294, + 669 + ], + "type": "text", + "content": "[9] Bagwe G, Yuan X, Pan M, and et al. Fed-cprompt: Contrastive prompt for rehearsal-free federated continual learning. arXiv preprint arXiv:2307.04869, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "text", + "content": "[10] Yunhui Guo, Mingrui Liu, Tianbao Yang, and Tajana Rosing. Improved schemes for episodic memory-based lifelong learning. Advances in Neural Information Processing Systems, 33:1023-1035, 2020. 4" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[11] Piao H, Wu Y, Wu D, and et al. Federated continual learning via prompt-based dual knowledge transfer. In *Forty-First International Conference on Machine Learning (ICML)*. PMLR, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 119, + 553, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 151 + ], + "type": "text", + "content": "[12] Parisa Hamedi, Roozbeh Razavi-Far, and Ehsan Hallaji. Federated continual learning: Concepts, challenges, and solutions. Neurocomputing, 651:130844, 2025. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 152, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 553, + 185 + ], + "type": "text", + "content": "[13] Hongjie Jia, Shifei Ding, Xinzheng Xu, and Ru Nie. The latest research progress on spectral clustering. Neural Computing and Applications, 24:1477-1486, 2014. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 186, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 241 + ], + "type": "text", + "content": "[14] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank J.Reddi, Sebastian U.Stich, and Ananda Theertha Suresh. Scaffold: Stochastic controlled averaging for federated learning. In International conference on machine learning. PMLR, pages 5132-5143, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 243, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 243, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 317, + 243, + 553, + 296 + ], + "type": "text", + "content": "[15] Alexander Krawczyk and Alexander Gepperth. An analysis of best-practice strategies for replay and rehearsal in continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4196-4204, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 299, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 299, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 299, + 553, + 319 + ], + "type": "text", + "content": "[16] Alex Krizhevsky and Geoffrey Hinton. Learning Multiple Layers of Features from Tiny Images. 2009. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 322, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 322, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 317, + 322, + 553, + 386 + ], + "type": "text", + "content": "[17] Matthias De Lange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Aleš Leonardis, Gregory Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying forgetting in classification tasks. In IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 3366-3385, 2021. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 388, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 388, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 317, + 388, + 553, + 431 + ], + "type": "text", + "content": "[18] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 434, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 477 + ], + "type": "text", + "content": "[19] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. In Proceedings of Machine Learning and Systems, pages 429-450, 2020. 2, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 479, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 479, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 317, + 479, + 553, + 521 + ], + "type": "text", + "content": "[20] Tian Li, Shengyuan Hu, Ahmad Beirami, and Virginia Smith. Ditto: Fair and robust federated learning through personalization. In International Conference on Machine Learning, PMLR, pages 6357-6368, 2021. 2, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 524, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 524, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 524, + 553, + 578 + ], + "type": "text", + "content": "[21] Junkang Liu, Yuanyuan Liu, Fanhua Shang, Hongying Liu, Jin Liu, and Wei Feng. Improving generalization in federated learning with highly heterogeneous data via momentum-based stochastic controlled weight averaging. In Forty-second International Conference on Machine Learning. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 580, + 553, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 580, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 317, + 580, + 553, + 643 + ], + "type": "text", + "content": "[22] Junkang Liu, Fanhua Shang, Yuanyuan Liu, Hongying Liu, Yuangang Li, and YunXiang Gong. Fedbcgd: Communication-efficient accelerated block coordinate gradient descent for federated learning. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 2955-2963, 2024. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "text", + "content": "[23] Quande Liu, Cheng Chen, Jing Qin, Qi Dou, and Pheng-Ann Heng. Feddg: Federated domain generalization on medical image segmentation via episodic learning in continuous frequency space. In CVPR, pages 1013–1023, 2021. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[24] Yang Liu, Anbu Huang, Yun Luo, He Huang, Youzhi Liu, Yuanyuan Chen, Lican Feng, Tianjian Chen, Han Yu, and" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3849" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 77, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 294, + 106 + ], + "type": "text", + "content": "Qiang Yang. Fedvision: An online visual object detection platform powered by federated learning. In AAAI, pages 13172-13179, 2020. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 139 + ], + "type": "text", + "content": "[25] David Lopez-Paz and MarcÁurelio Ranzato. Gradient episodic memory for continual learning. In Advances in neural information processing systems, 2017. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 194 + ], + "type": "text", + "content": "[26] Bing Luo, Wenli Xiao, Shiqiang Wang, Jianwei Huang, and Leandros Tassiulas. Tackling system and statistical heterogeneity for federated learning with adaptive client sampling. In IEEE INFOCOM 2022 - IEEE Conference on Computer Communications, pages 1739-1748, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 195, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 294, + 239 + ], + "type": "text", + "content": "[27] Fan Lyu, Shuai Wang, Wei Feng, Zihan Ye, Fuyuan Hu, and Song Wang. Multi-domain multi-task rehearsal for lifelong learning. Proceedings of the AAAI Conference on Artificial Intelligence, 35(10):8819-8827, 2021. 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 293 + ], + "type": "text", + "content": "[28] Fan Lyu, Qing Sun, Fanhua Shang, Liang Wan, and Wei Feng. Measuring asymmetric gradient discrepancy in parallel continual learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11411-11420, 2023. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 294, + 294, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 294, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 294, + 348 + ], + "type": "text", + "content": "[29] H. Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial Intelligence and Statistics, PMLR, pages 1273-1282, 2017. 1, 2, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 350, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 294, + 403 + ], + "type": "text", + "content": "[30] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading digits in natural images with unsupervised feature learning. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning, page 4, 2011. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 404, + 294, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 459 + ], + "type": "text", + "content": "[31] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading digits in natural images with unsupervised feature learning. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning, page 4, 2011. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 460, + 294, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 503 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 503 + ], + "type": "text", + "content": "[32] Hong-Wei Ng and Stefan Winkler. A data-driven approach to cleaning large face datasets. In Proceedings of the 2014 IEEE International Conference on Image Processing (ICIP), pages 343–347. IEEE, 2014. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 504, + 294, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 294, + 548 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 294, + 548 + ], + "type": "text", + "content": "[33] Li Qinbin, Diao Yiqun, Chen Quan, and He Bingsheng. Federated learning on non-iid data silos: An experimental study. In 2022 IEEE 38th International Conference on Data Engineering (ICDE), pages 965-978, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "type": "text", + "content": "[34] Felix Sattler, Simon Wiedemann, Klaus-Robert Müller, and Wojciech Samek. Robust and communication-efficient federated learning from non-iid data. IEEE Transactions on Neural Networks and Learning Systems, 31(9):3400-3413, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 603, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 635 + ], + "type": "text", + "content": "[35] Ozan Sener and Vladlen Koltun. Multi-task learning as multi-objective optimization. In Advances in Neural Information Processing Systems, 2018. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 636, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 669 + ], + "type": "text", + "content": "[36] Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 22(8):888-905, 2000. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[37] Yong Shi, Yuanying Zhang, Yang Xiao, and Lingfeng Niu. Optimization strategies for client drift in federated learning: A review. Procedia Computer Science, 214:1168-1173, 2022. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 620 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[38] Neta Shoham, Tomer Avidor, Aviv Keren, Nadav Israe, Daniel Benditkis, Liron Mor-Yosef, and Itai Zeitak. Overcoming forgetting in federated learning on non-iid data. In arXiv Preprint arXiv:1910.07796, 2019. 1, 3, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "text", + "content": "[39] Hongda Wu and Ping Wang. Fast-convergent federated learning with adaptive weighting. IEEE Transactions on Cognitive Communications and Networking, 7(4):1078-1088, 2021. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "text", + "content": "[40] Yichen Wu, Hong Wang, Peilin Zhao, Yefeng Zheng, Ying Wei, and Long-Kai Huang. Mitigating catastrophic forgetting in online continual learning by modeling previous task interrelations via pareto optimization. In *Forty-first international conference on machine learning*, 2024." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 219, + 553, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 261 + ], + "type": "text", + "content": "[41] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: A novel image dataset for benchmarking machine learning algorithms. In arXiv Preprint arXiv:1708.07747, 2017. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 262, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 553, + 296 + ], + "type": "text", + "content": "[42] Jin Xie, Chenqing Zhu, and Songze Li. Fedmes: Personalized federated continual learning leveraging local memory. In arXiv Preprint arXiv:2404.12710, 2024. 2, 4, 5, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 297, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 342 + ], + "type": "text", + "content": "[43] He Y, Shen C, Wang X, and et al. Fppl: An efficient and non-iid robust federated continual learning framework. In 2024 IEEE International Conference on Big Data (BigData), pages 3692-3701. IEEE, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 342, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 386 + ], + "type": "text", + "content": "[44] Luopan Y, Han R, Zhang Q, and et al. Fedknow: Federated continual learning with signature task knowledge integration at edge. In 2023 IEEE 39th International Conference on Data Engineering (ICDE), pages 341-354. IEEE, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 387, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 387, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 387, + 553, + 431 + ], + "type": "text", + "content": "[45] Xin Yang, Hao Yu, Xin Gao, Hao Wang, Junbo Zhang, and Tianrui Li. Federated continual learning via knowledge fusion: A survey. In IEEE Transactions on Knowledge and Data Engineering, 2024. 1, 2, 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 432, + 553, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 432, + 553, + 474 + ], + "spans": [ + { + "bbox": [ + 316, + 432, + 553, + 474 + ], + "type": "text", + "content": "[46] Xin Yao and Lifeng Sun. Continual local training for better initialization of federated models. In 2020 IEEE International Conference on Image Processing (ICIP), pages 1736-1740, 2020. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 476, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 476, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 476, + 553, + 520 + ], + "type": "text", + "content": "[47] Jaehong Yoon, Wonyong Jeong, Giwoong Lee, Eunho Yang, and Sung Ju Hwang. Federated continual learning with weighted inter-client transfer. In International Conference on Machine Learning, pages 12073-12086, 2021. 1, 2, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 521, + 553, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 521, + 553, + 576 + ], + "spans": [ + { + "bbox": [ + 316, + 521, + 553, + 576 + ], + "type": "text", + "content": "[48] Hao Yu, Xin Yang, Xin Gao, Yan Kang, Hao Wang, Junbo Zhang, and Tianrui Li. Personalized federated continual learning via multi-granularity prompt. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 4023-4034, 2024. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 577, + 553, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 577, + 553, + 620 + ], + "spans": [ + { + "bbox": [ + 316, + 577, + 553, + 620 + ], + "type": "text", + "content": "[49] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. In Advances in Neural Information Processing Systems, pages 5824-5836, 2020. 5" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3850" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/FedDifRC_ Unlocking the Potential of Text-to-Image Diffusion Models in Heterogeneous Federated Learning/d48a8fed-132a-4154-9395-1fe6c487ac0e_content_list.json b/2025/FedDifRC_ Unlocking the Potential of Text-to-Image Diffusion Models in Heterogeneous Federated Learning/d48a8fed-132a-4154-9395-1fe6c487ac0e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bb5a7978a479690b9c833cfb664389854cfc9a74 --- /dev/null +++ b/2025/FedDifRC_ Unlocking the Potential of Text-to-Image Diffusion Models in Heterogeneous Federated Learning/d48a8fed-132a-4154-9395-1fe6c487ac0e_content_list.json @@ -0,0 +1,2222 @@ +[ + { + "type": "text", + "text": "FedDifRC: Unlocking the Potential of Text-to-Image Diffusion Models in Heterogeneous Federated Learning", + "text_level": 1, + "bbox": [ + 133, + 128, + 867, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Huan Wang $^{1,\\ast}$ , Haoran Li $^{1,3,\\ast}$ , Huaming Chen $^{2}$ , Jun Yan $^{1}$ , Jiahua Shi $^{3}$ , Jun Shen $^{1,\\dagger}$ $^{1}$ School of Computing and Information Technology, University of Wollongong, Wollongong, Australia \n $^{2}$ School of Electrical and Computer Engineering, The University of Sydney, Sydney, Australia \n $^{3}$ QLD Alliance for Agriculture and Food Innovation, The University of Queensland, Brisbane, Australia \nhw226@uowmail.edu.au *Equal Contribution †Corresponding Author", + "bbox": [ + 101, + 202, + 923, + 294 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Federated learning aims at training models collaboratively across participants while protecting privacy. However, one major challenge for this paradigm is the data heterogeneity issue, where biased data preferences across multiple clients, harming the model's convergence and performance. In this paper, we first introduce powerful diffusion models into the federated learning paradigm and show that diffusion representations are effective steers during federated training. To explore the possibility of using diffusion representations in handling data heterogeneity, we propose a novel diffusion-inspired Federated paradigm with Diffusion Representation Collaboration, termed FedDifRC, leveraging meaningful guidance of diffusion models to mitigate data heterogeneity. The key idea is to construct text-driven diffusion contrasting and noise-driven diffusion regularization, aiming to provide abundant class-related semantic information and consistent convergence signals. On the one hand, we exploit the conditional feedback from the diffusion model for different text prompts to build a text-driven contrastive learning strategy. On the other hand, we introduce a noise-driven consistency regularization to align local instances with diffusion denoising representations, constraining the optimization region in the feature space. In addition, FedDifRC can be extended to a self-supervised scheme without relying on any labeled data. We also provide a theoretical analysis for FedDifRC to ensure convergence under non-convex objectives. The experiments on different scenarios validate the effectiveness of FedDifRC and the efficiency of crucial components. Code is available at https://github.com/hwang52/FedDifRC.", + "bbox": [ + 88, + 359, + 485, + 797 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 828, + 223, + 844 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Federated Learning (FL) [34] has become increasingly popular as a decentralized training paradigm, with multiple distributed clients train their local models without sharing their", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4daf08b7f68934a45627e2d5ccb5014341fe08e7eb4fa784852b91c48d022de6.jpg", + "image_caption": [ + "(a) Text-Driven Diffusion Contrastive Learning (TDCL)" + ], + "image_footnote": [], + "bbox": [ + 511, + 328, + 906, + 450 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b3b5bcf8dc4bfda73e7eed4c30ae8896bfba56f425d6305ad06f91e081f9583e.jpg", + "image_caption": [ + "(b) Noise-Driven Diffusion Consistent Restriction (NDCR)", + "Figure 1. The main idea of FedDifRC with two key modules TDCL and NDCR. (a) We leverage conditional generative feedback of the pre-trained diffusion model on various text prompts to devise a text-driven contrastive learning strategy, where the input embedding is injected as a condition. (b) We extract the diffusion denoising feature maps of intermediate layers from the pre-trained diffusion model at a specific time step $t$ during the diffusion backward process to construct a noise-driven regularization." + ], + "image_footnote": [], + "bbox": [ + 511, + 474, + 905, + 604 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "private dataset [16, 28] (i.e., training client models locally and aggregating them globally [29]). One major challenge faced in FL is the potential discrepancies in the local data distributions across distributed clients, which is known as the data heterogeneity issue [32, 35, 44]. Specifically, the private data is collected from different sources with diverse preferences and presents non-independent and identically distributed (non-iid) distribution [27, 75]. Such non-iid data", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3726", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "discrepancies can lead to inconsistency in the clients' local objective functions and optimization directions, making the model convergence slow and unstable [17, 29, 30, 39, 65].", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To mitigate the negative impact of the data heterogeneity, a mainstream of subsequent efforts delves into introducing a variety of suited signals to regulate the private model, there are mainly two perspectives: i) local model optimization at the client side and ii) global model aggregation at the server side. The former focuses on optimization strategies to make the variance of local updates limited [6, 29]. Another aims to improve the efficacy of global model aggregation [36, 69, 73]. In fact, most of these previous methods force the local models to be consistent to the global model. Although they have a certain effect by reducing gradient inconsistency, the gradually enlarged parameter deviation persists. Hence, we argue that local model overfits the local data distribution, which is magnified under the data heterogeneity and thus hinders the improvement of model's performance in FL.", + "bbox": [ + 89, + 138, + 482, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, the development of diffusion models (DMs) [4, 11, 37, 46] provides fresh opportunities. The diffusion models have demonstrated remarkable power in generating the diverse and realistic images, showing strong ability to capture intricate details and patterns from underlying data. For instance, with proper guidance, stable diffusion (SD) [46] models can effectively align a text prompt with the corresponding image context, which suggests their potential ability to understand both high-level and low-level contextual visual concepts regarding what an image contains. This semantic generative and visual comprehension capabilities of SD models motivate us to think: is it possible to exploit the inherent fine-grained visual representation capabilities of diffusion models to promote heterogeneous FL training?", + "bbox": [ + 89, + 367, + 482, + 579 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Taking into account both the effectiveness and efficiency, we revisit the underlying representations in the SD models, which are intermediate feature maps in a diffusion process (e.g., denoise or noise an image) produced by multiple calling a UNet [47]. We argue that diffusion representations are effective steers for FL training: i) extensive general knowledge inherent in diffusion models enhance local semantic diversity in FL; ii) smooth correspondences about semantic objects in diffusion models is a natural guiding signal in FL. Building on these two insights, we first propose conditional diffusion representations by imposing various text prompts in the SD model. In this way, each instance is contrastively enhanced with a set of text-driven diffusion representations, capturing rich domain variance. Then, we devise denoising diffusion representations from semantic relations between visual and linguistic concepts in the SD model. We extract the feature maps from different hierarchies of the UNet decoder in the SD model's denoising process, serving as guidance signals to avoid over-optimization in FL local training.", + "bbox": [ + 89, + 580, + 482, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a diffusion-inspired Federated framework with Diffusion Representation Collaboration, as", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "FedDifRC, to handle the data heterogeneity issue in FL, which consists of two key modules (see Fig. 1): i) First, we devise Text-driven Diffusion Contrastive Learning (TDCL), which utilizes conditional diffusion representations to construct inter-contrastive learning [1, 9, 26, 58] in a text-driven diffusion manner (the sample's embedding is injected as a condition). By maximizing the agreement between the local instance and its corresponding representations, TDCL adaptively captures class-relevant information and semantically meaningful knowledge. This objective encourages sample's embedding closer to conditional diffusion representations with same semantics and away from other negatives, thus maintaining a clear decision boundary. ii) Second, we use the denoising diffusion representations to build a flat and stable convergent target and further propose Noise-driven Diffusion Consistent Restriction (NDCR). Concretely, we extract denoising feature maps of the noisy input sample in the SD model's backward process, as a regularization term to constrain the optimization region of the local model. The local sample is forced to align the embedding with its denoising representations at the feature space to promote the uniformity for local training.", + "bbox": [ + 511, + 90, + 903, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are summarized as follows:", + "bbox": [ + 532, + 424, + 872, + 438 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We explore the possibility of using diffusion models to tackle data heterogeneity in FL and devise the conditional and denoising diffusion representations as effective steers to regulate model training in heterogeneous FL.", + "- We propose FedDifRC, a diffusion-inspired FL framework to handle data heterogeneity by leveraging the complementary advantages of two key components: i) TDCL facilitates the local model to learn more general class-relevant knowledge; ii) NDCR enforces the uniformity in the feature space for local training.", + "- We conduct extensive experiments for different FL scenarios (e.g., label shift, domain shift, and long-tailed) and different application scenarios (i.e., intra-domain and cross-domain) to demonstrate the effectiveness of our FedDifRC and the indispensability of each component." + ], + "bbox": [ + 513, + 441, + 903, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 684, + 653, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Heterogeneous Federated Learning", + "text_level": 1, + "bbox": [ + 511, + 710, + 820, + 727 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Federated learning is proposed to address privacy concerns in a distributed learning environment (e.g., Federated Averaging (FedAvg) [34]). Several studies have improved the local update process on the client side or the global aggregation process on the server side to tackle data heterogeneity in FL. The former focuses on local optimization strategies to make the diversity between the client model and the global model limited in the parameter level [6, 12, 13, 17, 26, 29, 35, 52]. The another branch aims to adopt aggregation mechanisms on the server to alleviate the negative influence of data heterogeneity [23, 43, 60, 62, 69, 73].", + "bbox": [ + 511, + 734, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3727", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, existing FL works that apply diffusion models mainly focus on building synthetic datasets [25, 51, 67, 74]. With the assistance of synthetic datasets, the FL global model can be retrained by this data augmentation scheme to mitigate data heterogeneity. Although these generated data have comparable quality and diversity to client images, the local model may overfit local domain distribution, and the inherent heterogeneity across clients may not have been addressed. In contrast, we focus on how to effectively leverage the fine-grained visual representation from diffusion models to promote a generalizable FL global model.", + "bbox": [ + 89, + 90, + 483, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Diffusion Model Representations", + "text_level": 1, + "bbox": [ + 89, + 277, + 380, + 292 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, diffusion models have shown significant advances in image generation, editing, and stylized [4, 11, 37, 46, 66], suggesting that they contain rich underlying representations to be exploited for different downstream tasks. DIFT [54] is used to extract the diffusion features without explicit supervision. Diffusion Hyperfeatures [31] integrates multi-scale and multi-timesteps diffusion features into feature descriptors. The diffusion representations have shown impressive results on the vision branch, including being testified as certifiable classifiers [3, 24, 63] and semantics correspondence learners [54, 64, 72]. In this paper, we expand the diffusion models to FL and construct the conditional and denoising diffusion representations, offering fruitful class-relevant information and consistent convergence signals.", + "bbox": [ + 89, + 303, + 483, + 515 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Contrastive Learning", + "text_level": 1, + "bbox": [ + 89, + 534, + 294, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Contrastive learning (CL) has been widely applied to self-supervised learning scenarios [1, 7, 9]. Many works are focused on learning the encoder where the embeddings of the same sample are pulled closer and those of different samples are pushed apart [2, 5, 18, 55]. They mainly construct a positive pair and some negative pairs for each data instance and use InfoNCE [38] to contrast positiveness against negativity. There are some works that incorporate contrastive learning into FL to assist local training to achieve higher model performance [8, 12, 26, 53]. In our work, Text-driven Diffusion Contrastive Learning (TDCL) is devised to promote local model by attracting samples to positive representations while pushing away other negatives.", + "bbox": [ + 89, + 560, + 483, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Diffusion Models Help Federated Learning", + "text_level": 1, + "bbox": [ + 89, + 780, + 475, + 799 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we begin with a brief overview of the general FL paradigm and the generative diffusion models (Sec. 3.1). Then, we analyze the effectiveness of diffusion representations and illustrate our core insights (Sec. 3.2). Finally, we will introduce how to extract the conditional and denoising diffusion representations in Sec. 3.3 and Sec. 3.4.", + "bbox": [ + 89, + 809, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries and Background", + "text_level": 1, + "bbox": [ + 513, + 90, + 785, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Federated Learning: We focus on the case with $K$ clients holding the heterogeneous data partition $\\{\\mathcal{D}_1,\\mathcal{D}_2,\\ldots ,\\mathcal{D}_K\\}$ with private dataset $\\mathcal{D}_k = \\{x_i,y_i\\}_{i = 1}^{n_k},y_i\\in \\{1,\\dots ,|\\mathcal{C}|\\}$ for the $k$ -th client, where $n_k$ denotes the local dataset scale. Let $n_k^j$ be the number of samples with label $j$ at client $k$ and $\\mathcal{D}_k^j = \\{(x,y)\\in \\mathcal{D}_k|y = j\\}$ denotes the set of samples with label $j$ on the $k$ -th client, so the number of samples in $j$ -th class is $n^j = \\sum_{k = 1}^{K}n_k^j$ . The objective of the general FL is to learn an optimal global model $w$ across $K$ clients:", + "bbox": [ + 511, + 112, + 906, + 250 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {w} \\mathcal {L} (w) = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\frac {n _ {k}}{N} \\mathcal {L} _ {k} (w; \\mathcal {D} _ {k}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 256, + 906, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_k$ is the $k$ -th client local loss function and $w$ is the global model parameters. $N$ is the total number of samples among all clients. In the FL training setup, we consider the local model $f$ with parameters $w = \\{u,v\\}$ . It has two modules: i) a feature extractor $h(u)$ with the parameters $u$ maps each sample $x_{i}$ to a $d$ -dim (default as 512) vector as $z_{i} = h(u;x_{i})$ ; ii) a classifier $g(v)$ with parameters $v$ maps $z_{i}$ into a $|\\mathcal{C}|$ -dim output logits as $l_{i} = g(v;z_{i})$ .", + "bbox": [ + 511, + 301, + 906, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Models: Given the image sample $x$ from an underlying distribution $p(x)$ , a forward diffusion process as a Markov chain to add random Gaussian noise $\\epsilon \\sim \\mathcal{N}(0, \\mathbf{I})$ to the original sample $x_0$ , and $x_t$ is the noise sample after adding $t$ -steps noise:", + "bbox": [ + 511, + 424, + 905, + 498 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t} = \\sqrt {1 - \\gamma_ {t}} x _ {t - 1} + \\sqrt {\\gamma_ {t}} \\epsilon_ {t}, t \\in \\{1, \\dots , T \\}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 503, + 906, + 522 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $T$ is the number of diffusion timesteps, $\\gamma_t \\in (0,1)$ is a adjustable time variance schedule [11]. By leveraging the additive property of the Gaussian distribution, Eq. (2) can be further reformulated as $x_t = \\sqrt{\\bar{\\alpha}_t} x_0 + \\sqrt{1 - \\bar{\\alpha}_t} \\epsilon_t$ , where $\\alpha_t = 1 - \\gamma_t$ and $\\bar{\\alpha}_t \\doteq \\prod_{i=1}^t \\alpha_i$ . On this basis, the image $x_0$ can be obtained from a random noise $x_t \\sim \\mathcal{N}(0,\\mathbf{I})$ by reversing the above forward diffusion process [11]:", + "bbox": [ + 511, + 527, + 905, + 633 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(x _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} \\left(x _ {t}, t\\right)\\right), t \\in \\{T, \\dots , 1 \\}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 638, + 905, + 671 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon_{\\theta}$ denotes the denoiser network with the parameters $\\theta$ to predict noise $\\epsilon$ , and the denoiser network is usually implemented as a UNet architecture [47]. Further, the denoiser $\\epsilon_{\\theta}$ is optimized using the loss $T_{\\theta}$ as follows:", + "bbox": [ + 511, + 676, + 905, + 737 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} _ {\\theta} = \\mathbb {E} _ {\\epsilon , x _ {0}, t} [ \\| \\epsilon_ {\\theta} (x _ {t}, t) - \\epsilon \\| _ {2} ^ {2} ], \\epsilon \\sim \\mathcal {N} (0, \\mathbf {I}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 742, + 903, + 760 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon$ is the noise in the forward diffusion process, and $\\epsilon_{\\theta}$ is usually implemented as a UNet [47]. Besides, the diffusion model's objective in Eq. (4) can be easily extended to a conditional diffusion paradigm by adding a condition c to the $\\epsilon_{\\theta}$ . Thus, the training objective should be modified as:", + "bbox": [ + 511, + 765, + 905, + 842 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} _ {\\theta} = \\mathbb {E} _ {\\epsilon , x _ {0}, t, \\mathbf {c}} [ | | \\epsilon_ {\\theta} (x _ {t}, t, \\mathbf {c}) - \\epsilon | | _ {2} ^ {2} ], \\epsilon \\sim \\mathcal {N} (0, \\mathbf {I}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 847, + 903, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with the denoiser $\\epsilon_{\\theta}(x_t,t,\\mathbf{c})$ in Eq. (5), we can generate data with a specific condition $\\mathbf{c}$ by solving a reverse process.", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3728", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Motivation and Analysis", + "text_level": 1, + "bbox": [ + 89, + 90, + 313, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The diffusion model is a probabilistic generative model designed for denoising by systematically reversing a progressive-noising process, and recent studies [61, 68] have verified that diffusion-based generative models can be effective representation learners. Stable diffusion models are able to weave from scratch with fruitful and innovative depictions of high fidelity and further capture the underlying properties, positions, and correspondences from visual concepts. Such a phenomenon shows that SD models can align a textual prompt with its corresponding image context, demonstrating their potential ability in understanding both high-level (e.g., semantic relations) and low-level (e.g., textures, edge, and structures) visual concepts in an image.", + "bbox": [ + 88, + 112, + 482, + 308 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Autoencoders [71, 76] can learn to reconstruct their input at the output and have been one of the dominant approaches for representation learning. We assume a linear autoencoder expressed as $W_{a}W_{b}\\cdot x + W_{c}\\cdot x$ to imitate UNet [47] used in diffusion models, where $W_{a}\\in \\mathbb{R}^{d\\times d^{*}}$ , $W_{b}\\in \\mathbb{R}^{d^{*}\\times d}$ , and $W_{c}\\in \\mathbb{R}^{d\\times d}$ , $d^{*} < d$ . For $x_{t} = \\sqrt{\\bar{\\alpha}_{t}} x_{0} + \\sqrt{1 - \\bar{\\alpha}_{t}}\\epsilon_{t}$ and we set $\\mathbb{E}[x_0] = 0$ , thus minimizing the loss $\\mathcal{T}_{\\theta}$ in Eq. (4) is equivalent to minimize the modified objective $\\mathcal{T}_{\\theta}^{*}$ :", + "bbox": [ + 89, + 309, + 483, + 431 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{r l} \\mathcal {T} _ {\\theta} ^ {*} = & \\mathbb {E} _ {\\epsilon , x _ {0}, t} [ \\| \\epsilon_ {\\theta} (\\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t}) - \\epsilon \\| _ {2} ^ {2} ], \\\\ & \\equiv \\min \\left\\{\\bar {\\alpha} _ {t} \\left(W _ {a} W _ {b} + W _ {c}\\right) ^ {\\top} \\Sigma_ {x _ {0}} \\left(W _ {a} W _ {b} + W _ {c}\\right) \\right\\}, \\end{array} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 441, + 483, + 478 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Sigma_{x_0}$ means covariance matrix of $x_0$ . Eq. (6) ensures a meaningful latent space for the diffusion models, similar to the linear autoencoder that encodes data into the principal components space. Therefore, the diffusion training process is equivalent to encourage the disentanglement of the latent representations until good-quality reconstruction is reached.", + "bbox": [ + 89, + 492, + 483, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To validate the representation ability in diffusion models, we use t-SNE [57] to visualize the latent features, shown in Fig. 2, without training on the specific MNIST [22] dataset. Concretely, we use the pre-trained SD model to extract features of the UNet decoder $\\epsilon_{\\theta}$ , $T$ means diffusion time-steps, $L$ means the layer of the UNet decoder $\\epsilon_{\\theta}$ ( $L \\in \\{1, 2, 3, 4\\}$ ). Our results in Fig. 2 indicate that, even without supervision, diffusion models are able to group input samples at properly selected time-steps $T$ and layer $L$ . However, when the $T$ or $L$ is too large (e.g., $T = 999$ or $L = 4$ ), the representations become inseparable, leading to blurred decision boundaries. Furthermore, we apply K-Means clustering (cluster size set as 5) on the latent features from the diffusion UNet decoder and visually examine whether the diffusion features contain consistent semantic information. In Fig. 3, different parts of the visual objects are clustered, showing that earlier layers (e.g., $L = 2$ ) capture coarse yet consistent semantics, while later layers (e.g., $L = 3$ ) focus on low-level textural details. Both t-SNE and K-Means observation results motivate us to leverage the fine-grained visual representation capabilities from the SD model to promote heterogeneous FL training.", + "bbox": [ + 89, + 584, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f8a6c63c46d755a35a6ffff60731a424d3f285a268df8ff1f96c7c9045251e21.jpg", + "image_caption": [ + "$\\mathrm{T} = 100$ $\\mathrm{L} = 1$" + ], + "image_footnote": [], + "bbox": [ + 516, + 88, + 612, + 135 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3a67be20e7b2fa7dea25cf54aeee10550fa233edc80ef511bd75c0a32dba453e.jpg", + "image_caption": [ + "$\\mathrm{T} = 100, \\mathrm{L} = 2$" + ], + "image_footnote": [], + "bbox": [ + 616, + 90, + 705, + 135 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/636c14c8e2b0ef1c96ad52b480b999aeddd00272a308840208646698dabcab93.jpg", + "image_caption": [ + "$\\mathrm{T} = 100, \\mathrm{L} = 3$" + ], + "image_footnote": [], + "bbox": [ + 709, + 89, + 803, + 135 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ac0778df95d294b66dd811290369d870d1189996d3f8f939f30da54a3259096d.jpg", + "image_caption": [ + "$\\mathrm{T} = 100, \\mathrm{L} = 4$" + ], + "image_footnote": [], + "bbox": [ + 807, + 89, + 903, + 135 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8f3dea7801cb18ee61ea8cc8202d2539d0362686418ad5b2c0599d676ebf616b.jpg", + "image_caption": [ + "$\\mathrm{T} = 100, \\mathrm{L} = 2$" + ], + "image_footnote": [], + "bbox": [ + 516, + 151, + 612, + 196 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0612edbe36ba27a6efb4d65a7d2454bb7cd5106497a9e1e729cb9bdffae76c33.jpg", + "image_caption": [ + "$\\mathrm{T} = 300$ $\\mathrm{L} = 2$" + ], + "image_footnote": [], + "bbox": [ + 616, + 151, + 707, + 198 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b5ac2f9cb919bc3bd8789d3d122db37ff8664819bd0f8d8728a60e258fb92019.jpg", + "image_caption": [ + "$\\mathrm{T} = 500$ $\\mathrm{L} = 2$" + ], + "image_footnote": [], + "bbox": [ + 709, + 151, + 803, + 198 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/49fec3ae084372be889e25e172bd4c71f8b0abe8bf2aca0967bc5eb06b318a0f.jpg", + "image_caption": [ + "$\\mathrm{T} = 999, \\mathrm{L} = 2$" + ], + "image_footnote": [], + "bbox": [ + 807, + 151, + 903, + 198 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7e8df0660532b3ecf50da30033198c24741b53c5ad94df89dfd4e887f077b15c.jpg", + "image_caption": [ + "Figure 2. The t-SNE feature visualization on the MNIST dataset from the diffusion UNet decoder, without training, where diffusion time-steps $T \\in \\{100, 300, 500, 999\\}$ and layer $L \\in \\{1, 2, 3, 4\\}$ ." + ], + "image_footnote": [], + "bbox": [ + 516, + 282, + 584, + 335 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4e176ea07de581eeabc07115f2217e77dbf22c2d51a899cb9152bd0b198084c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 337, + 584, + 391 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8d53eedb8fc17507590e81858cd79f3f732586882b3b6cef15cd6fbdba102d72.jpg", + "image_caption": [ + "Sample" + ], + "image_footnote": [], + "bbox": [ + 516, + 393, + 584, + 446 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3a48b6ffa541373ec0a7ac0c6a64d553b2365d202a359fdd995129792049d321.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 594, + 282, + 665, + 335 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4cd45332ef21f842c993fc9885e7458cb0052572a160c49e8a6ffd21bf40ad05.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 337, + 663, + 391 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9271bb3537daa97e75ba5164c4ed302a703d2f2cbab43ff5dc03fabdb110fa70.jpg", + "image_caption": [ + "Layer $\\mathrm{L} = 1$" + ], + "image_footnote": [], + "bbox": [ + 596, + 393, + 665, + 446 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a566f150a4f4598c07fc59202fef3f261e16d993fbb9901e67ea0dccf9a9f8f2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 282, + 743, + 335 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/be6576963a4e3bcbd2421fcf63fd7c2cb2067b4fbcba4e5a47f0053c66c42e89.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 337, + 743, + 391 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f53f98e68a2d16cc8116be8e0433fdd0b83af261cf2c0caf7d751d3d54e8ea9a.jpg", + "image_caption": [ + "Layer $\\mathrm{L} = 2$" + ], + "image_footnote": [], + "bbox": [ + 674, + 393, + 743, + 446 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d37db70802b268e552fadeee8f1e0f6f56d61ef3f6a60cbfabe5b718ff58fe6d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 282, + 823, + 335 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/25b1343cd9d6f5d90e063b0436637742948ecc1a97a9d2a54a0c563e129bde26.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 337, + 823, + 391 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/684dea9a71546e911a440b052a480f881082113ca032a2d2f07b26bc45e80915.jpg", + "image_caption": [ + "Layer $\\mathrm{L} = 3$" + ], + "image_footnote": [], + "bbox": [ + 754, + 393, + 823, + 446 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/63b94aa5355f2b8cd2d400d6e28f0935ce05e80457426ac5aef17be979d60175.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 834, + 282, + 903, + 335 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9a4072b7bff2a1f37b9c139730ac474828e105edb1c833f6ac2cc89c166c87c2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 834, + 337, + 903, + 391 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/63d28bfb0bdf8bdf0d037d5996ff91236fd594adc490751ff284b6cf047f94cb.jpg", + "image_caption": [ + "Layer L=4" + ], + "image_footnote": [], + "bbox": [ + 834, + 393, + 903, + 446 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/235e2f1b18ac52bbb92c196436c9d08492973c42d19545bf9d893b01d660e043.jpg", + "image_caption": [ + "Figure 3. The K-Means clustering feature visualization (each color represents a cluster) from the diffusion UNet decoder's layers." + ], + "image_footnote": [], + "bbox": [ + 516, + 520, + 584, + 574 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b60abde78623d328711c61c2a319a1da70e85a976831f4eec0f61ebf388a54c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 575, + 584, + 630 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a2a19a0e4d27aad4549756fd7424b2824cc3a6bd763d36427eed0c0a49061fb0.jpg", + "image_caption": [ + "Sample" + ], + "image_footnote": [], + "bbox": [ + 516, + 632, + 584, + 686 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b3fb60e29d3b8928ff02be35e06d79b5ea4103f83d37afa7e3f5030e2961c0cf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 521, + 663, + 574 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2be7cb8d134bd81d251ecc69eff81fa36b38e5132dde342f577daaecb1cba954.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 575, + 663, + 630 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d5485b794e5d93b1443ccc7eb506b6910483bbad041609356317fda952e28abf.jpg", + "image_caption": [ + "Layer $L = 2$" + ], + "image_footnote": [], + "bbox": [ + 596, + 632, + 663, + 686 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/34285cf22b2b7ad3d94c5d074149060a33a743cdb931914febe3ca71893dd38f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 521, + 743, + 574 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/de4b2a908528032c4ee7e0ab4063267994b2816e052a4be081747022546fcba7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 575, + 743, + 630 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6c47cea1cb100428c7a1881367ae23f4884ca57853d724e12cad253a6ef65b11.jpg", + "image_caption": [ + "Layer $\\mathrm{L} = 3$" + ], + "image_footnote": [], + "bbox": [ + 674, + 632, + 743, + 686 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b3fec03c3b8ea8199f936d8baa20830a15e78f6404f1430e9b4051e554f8e354.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 521, + 823, + 574 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/85bb727f570431443904a0b95e060875dbdfbb73c1a2f3b34491ed408b4da011.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 575, + 823, + 630 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/652398c7f9ecdf1a7abecd29e32d01ea9b3463422d82cf2be49348821ae6eae9.jpg", + "image_caption": [ + "Layer L=4", + "Figure 4. The PCA visualization of early (layer 2), later (layer 4), and fused features (see Eq. (8)), where the first three components of PCA-computed serve as color channels for each image." + ], + "image_footnote": [], + "bbox": [ + 754, + 632, + 823, + 686 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a60727202fe7205c4a39ee3040def9e46bf97f57cb3a5dc6e79503fab55b88b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 834, + 521, + 903, + 574 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/09cf4bc4d629bf2f1cf8eea5a44b1cb9561f5245ff380917365cc69ff3c891a0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 834, + 575, + 903, + 630 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0f1dd56c24bdc84536d1738b6e48f6e8da2e928851787d9884bfb47734b33195.jpg", + "image_caption": [ + "Fused" + ], + "image_footnote": [], + "bbox": [ + 834, + 632, + 903, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Conditional Diffusion Representations", + "text_level": 1, + "bbox": [ + 511, + 786, + 843, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The key idea of our conditional diffusion representations is based on cognitive diversity of generative diffusion models for different text prompts, via a condition-guided manner. Specifically, for the client $k$ , given an input image $x_{i}$ of the local dataset $\\mathcal{D}_k = \\{x_i,y_i\\}_{i = 1}^{n_k},y_i\\in \\{1,\\dots,|\\mathcal{C}|\\}$ means the label of $x_{i}$ , and we let $S$ represents the set of all class names,", + "bbox": [ + 511, + 810, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3729", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "so we use $S_{y_i}$ to denote the class name of the label $y_i$ . From our implementation, the image $x_i$ is fed into the encoder $h_k$ of the local model $f_k$ to obtain the embedding $\\mathbf{c}_i$ , then $\\mathbf{c}_i$ is treated as a condition and injected into the diffusion model. For a pre-trained SD model, $\\mathbf{c}_i$ is further combined with a relevant text prompt $\\mathcal{P}_{y_i} =$ 'a photo of a $S_{y_i}$ ' and sent to the UNet decoder to produce diffusion representations $\\mathcal{F}_i$ . Note that since we do not add noise, we simply set the time-steps $T = 0$ of the diffusion process. Typically, the image $x_i$ size is $224 \\times 224$ , and $\\mathcal{F}_i$ contains 4 feature maps, where the $L$ -th layer's feature map $\\mathcal{F}_{i(L)}$ has the spatial size $\\{7,14,28,28\\}$ ( $L \\in \\{1,2,3,4\\}$ ), and $\\{1280,1280,640,320\\}$ are channels. Formally, conditional diffusion representations $\\mathcal{F}_i$ are computed by the $\\epsilon_{\\theta}$ in Eq. (5), image $x_i$ , condition $\\mathbf{c}_i$ , text $\\mathcal{P}_{y_i}$ :", + "bbox": [ + 89, + 90, + 483, + 305 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {i} = \\epsilon_ {\\theta} \\left(x _ {i}, \\tau_ {i}, T = 0, \\mathbf {c} _ {i}\\right), \\tau_ {i} \\leftarrow \\varphi \\left(\\mathcal {P} _ {y _ {i}}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 318, + 482, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\varphi$ denotes the text encoder of the SD model and $\\tau_{i}$ is the text features with the prompt $\\mathcal{P}_{y_i}$ . Inspired by the results in Fig. 3, which motivate us to combine features at different layers of $\\mathcal{F}_i$ to capture both semantics and details. A simple concatenation, however, can lead to an unnecessarily high-dimensional feature $(1280 + 640 + 320 = 2240)$ . To reduce high dimension, we apply PCA for each layer $L\\in \\{2,3,4\\}$ (due to the spatial resolution too low, we do not consider the $L = 1$ ), and we aggregate them to the same resolution, then we connect them together to obtain the fused features $\\widetilde{\\mathcal{F}}_i$ :", + "bbox": [ + 89, + 342, + 483, + 494 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {\\mathcal {F}} _ {i} = \\mathrm {C o n c a t} (\\widetilde {\\mathcal {F}} _ {i (L)}), \\widetilde {\\mathcal {F}} _ {i (L)} = \\mathrm {P C A} (\\mathcal {F} _ {i (L)}), \\quad (8)\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 503, + 482, + 525 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $L\\in \\{2,3,4\\}$ we perform PCA: i) for $\\mathcal{F}_{i(2)}$ we first upsample to the same resolution and apply PCA $\\mathbb{R}^{1280\\to 256}$ ii) for $\\mathcal{F}_{i(3)}$ and $\\mathcal{F}_{i(4)}$ , we apply PCA $\\mathbb{R}^{640\\to 128}$ and PCA $\\mathbb{R}^{320\\to 128}$ . Then, we combine them to form the final fused features $\\widetilde{\\mathcal{F}}_i$ . As shown in the last column of Fig. 4, the fused features strike a balance between high-level and low-level visual concepts, focusing on both semantics and textures in the image. On the example in the second row of Fig. 4, the fused feature yields more accurate matches (e.g., face and computer borders). The comparison results in Table. 2 also confirm this viewpoint. We define $\\widetilde{\\mathcal{F}}_i$ calculated by Eq. (8) as conditional diffusion representations of the sample $x_{i}$", + "bbox": [ + 88, + 535, + 483, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Denoising Diffusion Representations", + "text_level": 1, + "bbox": [ + 89, + 727, + 406, + 743 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although we construct the conditional diffusion representations from the diffusion model's generative feedback, $\\widetilde{\\mathcal{F}}_i$ is still highly variable and depends on the condition generated by the local model, which can not ensure a consistent stable convergence signal. Based on Eq. (4), we explore the inherent denoising property in diffusion models: SD model has learned extensive general knowledge about identical semantics via large pre-training, reflected in the UNet denoising process. Therefore, we aim to extract the denoising diffusion representations contained in a pre-trained SD model, to", + "bbox": [ + 89, + 750, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "align the sample's embedding with the class semantics uniformly in the feature space. Given the image pair $\\{x_i, y_i\\}$ of $k$ -th client's dataset $\\mathcal{D}_k$ , we first perform a forward process to add $t$ time-steps random Gaussian noise on $x_i$ as $\\tilde{x}_i^t = \\sqrt{\\bar{\\alpha}_t} x_i + \\sqrt{1 - \\bar{\\alpha}_t} \\epsilon_t$ , similar to [46], where $\\bar{\\alpha}_t$ and $\\epsilon$ are the same as defined in the Eq. (3). Next, we set the text prompt of $\\tilde{x}_i^t$ as $\\mathcal{P}_{y_i}$ = 'a photo of a $S_{y_i}$ ' used in the SD model pre-training [46], to avoid destroying correspondence between visual and linguistic concepts, and the $S_{y_i}$ denotes the class name of the label $y_i$ . Formally, the denoising diffusion representations $\\tilde{\\mathcal{H}}_i$ of the sample $x_i$ are calculated by the $\\epsilon_\\theta$ in Eq. (4), noisy sample $\\tilde{x}_i^t$ , text $\\mathcal{P}_{y_i}$ :", + "bbox": [ + 511, + 90, + 906, + 273 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {H} _ {i} = \\epsilon_ {\\theta} (\\tilde {x} _ {i} ^ {t}, \\tau_ {i}, T = t), \\tau_ {i} \\gets \\varphi (\\mathcal {P} _ {y _ {i}}),\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 284, + 808, + 303 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {\\mathcal {H}} _ {i} = \\operatorname {C o n c a t} \\left(\\widetilde {\\mathcal {H}} _ {i (L)}\\right), \\quad \\widetilde {\\mathcal {H}} _ {i (L)} = \\operatorname {P C A} \\left(\\mathcal {H} _ {i (L)}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 299, + 903, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\varphi$ denotes the text encoder of the SD model and $\\tau_{i}$ is the text features of the text prompt $\\mathcal{P}_{y_i}$ . We further analyze the effects of the time-steps $t$ in Fig. 6. Then, we combine the features of layers $L\\in \\{2,3,4\\}$ , like Eq. (8), to obtain fused features $\\widetilde{\\mathcal{H}}_i$ . Intuitively, $\\widetilde{\\mathcal{H}}_i$ is considered as a virtual teacher to calibrate the embedding of $x_{i}$ at the feature-level. We define the denoising features $\\widetilde{\\mathcal{H}}_i$ calculated by Eq. (9) as denoising diffusion representations of the sample $x_{i}$ .", + "bbox": [ + 511, + 334, + 906, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 511, + 470, + 650, + 489 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our proposed FedDifRC leverage conditional and denoising diffusion representations to obtain diverse class-relevant knowledge and consistent convergence signals by two complementary components: Text-driven Diffusion Contrastive Learning in Sec. 4.1 and Noise-driven Diffusion Consistent Restriction in Sec. 4.2. Further, we discuss how to extend FedDifRC to a self-supervised scheme (Sec. 4.3). Finally, we provide a theoretical analysis of FedDifRC under nonconvex case (Sec. 4.4). Illustration of FedDifRC in Fig. 5.", + "bbox": [ + 511, + 497, + 906, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Text-Driven Diffusion Contrasting", + "text_level": 1, + "bbox": [ + 511, + 643, + 815, + 660 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inspired by success of contrastive learning [1, 9, 26], based on Sec. 3.3, conditional diffusion representations $\\widetilde{\\mathcal{F}}_i$ of sample $\\{x_i,y_i\\} \\in \\mathcal{D}_k$ can be obtained with the condition $\\mathbf{c}_i$ and the text prompt $\\mathcal{P}_{y_i} = \\text{'a photo of a } S_{y_i}$ , where $S_{y_i}$ is the class name of the label $y_{i}$ . For each sample's embedding $z_{i} = h_{k}(x_{i})\\in \\mathbb{R}^{d}$ , we enforce $z_{i}$ to be similar to respective positive pair $\\widetilde{\\mathcal{F}}_i$ and dissimilar to negative sample pairs $\\widetilde{\\mathcal{F}}_{\\mathcal{N}_i}$ where $\\widetilde{\\mathcal{F}}_{\\mathcal{N}_i}$ means the conditional diffusion representations with the same condition $\\mathbf{c}_i$ and the unmatched text prompts $\\mathcal{P}_j = \\text{'a photo of a } S_j$ ( $S_{j}\\in S$ and $S_{j}\\neq S_{y_{i}}$ ). Then, we define the similarity between $z_{i}$ and $\\widetilde{\\mathcal{F}}_i$ as follows:", + "bbox": [ + 511, + 665, + 906, + 838 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns \\left(z _ {i}, \\widetilde {\\mathcal {F}} _ {i}\\right) = \\frac {z _ {i} \\cdot \\widetilde {\\mathcal {F}} _ {i}}{\\mathcal {U} \\left(\\left\\| z _ {i} \\right\\| _ {2} \\times \\left\\| \\widetilde {\\mathcal {F}} _ {i} \\right\\| _ {2}\\right)}, \\mathcal {U} = \\frac {1}{n _ {k}} \\sum_ {m = 1} ^ {n _ {k}} \\| z _ {m} - \\widetilde {\\mathcal {F}} _ {i} \\| _ {2}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 848, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3730", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3f5f6e3e5ecb3bee521b4d2bd1782a532d785516bcdbb4e1af3e2a984eb420ca.jpg", + "image_caption": [ + "Training Process for Local Client", + "Figure 5. Illustration of FedDifRC, including two complementary modules: TDCL (Sec. 4.1) builds an inter-contrastive learning strategy by conditional diffusion representations (Sec. 3.3), NDCR (Sec. 4.2) designs a consistent regularization penalty via denoising diffusion representations (Sec. 3.4), jointly promoting FL training." + ], + "image_footnote": [], + "bbox": [ + 94, + 104, + 480, + 281 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $n_k$ is total number of samples at client $k$ , $\\mathcal{U}$ is a normalization factor for similarity, where we set $\\mathcal{U}$ to denote the average distance between $z_m|_{m=1}^{n_k}$ and $\\widetilde{\\mathcal{F}}_i$ on dataset $\\mathcal{D}_k$ . Next, we introduce TDCL to quantify and regulate the local model training, which can be formulated as follows:", + "bbox": [ + 89, + 385, + 483, + 460 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {T D C L} = \\log (1 + \\frac {\\sum_ {\\widetilde {\\mathcal {F}} _ {j} \\in \\widetilde {\\mathcal {F}} _ {\\mathcal {N} _ {i}}} \\exp \\left(s \\left(z _ {i} , \\widetilde {\\mathcal {F}} _ {j}\\right) / \\tau\\right)}{\\exp \\left(s \\left(z _ {i} , \\widetilde {\\mathcal {F}} _ {i}\\right) / \\tau\\right)}), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 465, + 482, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\tau$ denotes a temperature parameter to control representation strength [1]. By minimizing Eq. (11), for client $k$ , the local model brings each sample's embedding $z_{i}$ closer to the positive pair and away from other negative pairs.", + "bbox": [ + 89, + 513, + 483, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Noise-Driven Diffusion Regularization", + "text_level": 1, + "bbox": [ + 89, + 580, + 419, + 597 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Despite each local model aligning sample's embedding with conditional diffusion representations by optimizing $\\mathcal{L}_{TDCL}$ in Eq. (11). Since conditional diffusion representations are dynamically built from text-driven cross-modal matching and depend on generated condition $\\mathbf{c}_i$ at each round, which could not offer a stable convergence direction. Therefore, we aim to leverage denoising diffusion representations $(\\widetilde{\\mathcal{H}}_i)$ in Eq. (9)) to devise a consistent regularization penalty, constraining model's optimization region in the feature space. We design Noise-driven Diffusion Consistent Restriction (NDCR) and utilize a regularization term to pull the sample's embedding $z_{i}$ closer to the respective $\\widetilde{\\mathcal{H}}_i$ :", + "bbox": [ + 89, + 603, + 483, + 785 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {N D C R} = \\sum_ {q = 1} ^ {d} \\left(z _ {i (q)} - \\widetilde {\\mathcal {H}} _ {i (q)}\\right) ^ {2}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 178, + 792, + 482, + 833 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $q$ indexes the dimension $d$ of the embedding vector. Based on Eq. (12), we expect to align the sample's embedding with the corresponding denoising diffusion representations, enforcing semantics uniformity of the feature space.", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Besides, we construct CrossEntropy loss and use the logits output vector $l_{i} = g(z_{i}) = f(x_{i})$ with the original label $y_{i}$ to optimize the local discriminative ability for each client. Formally, given the sample pair $\\{x_{i},y_{i}\\}$ to calculate $\\mathcal{L}_{CE}$ :", + "bbox": [ + 511, + 90, + 905, + 152 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C E} = - \\mathbf {1} _ {y _ {i}} \\log (\\psi (f (x _ {i}))), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 611, + 159, + 906, + 176 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\psi$ denotes the Softmax function, $f$ is the local model with parameters $\\{u,v\\}$ maps $x_{i}$ into a $|\\mathcal{C}|$ -dim output logits $l_{i} = f(x_{i}) = g(h(x_{i}))$ . Finally, for the $k$ -th client, we carry out the following optimization objective in local training:", + "bbox": [ + 511, + 183, + 906, + 244 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {T D C L} (\\mathrm {E q .} (1 1)) + \\mathcal {L} _ {N D C R} (\\mathrm {E q .} (1 2)) + \\mathcal {L} _ {C E}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 252, + 906, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The algorithm of FedDifRC is given in Appendix A.", + "bbox": [ + 511, + 275, + 870, + 291 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Expanding to Self-Supervised Scheme", + "text_level": 1, + "bbox": [ + 511, + 297, + 841, + 314 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "FedDifRC can be extended to the self-supervised scheme without relying on any labeled data source. Specifically, for $\\mathcal{L}_{TDCL}$ in Eq. (11), we only need to modify text template slightly: i) we set the text prompt $\\mathcal{P}_{y_i} = \\text{'a photo of a similar object'}$ in Eq. (7) to generate $\\widetilde{\\mathcal{F}}_i$ as the positive pair; ii) then we randomly select from the set of all class names in TinyImageNet [21] to get $S_{\\mathcal{N}_i}$ , and set $\\mathcal{P}_{y_i} = \\text{'a photo of a } S_{\\mathcal{N}_i}$ to obtain $\\widetilde{\\mathcal{F}}_{\\mathcal{N}_i}$ as negative pairs. For the $\\mathcal{L}_{NDCR}$ in Eq. (12), we set $\\mathcal{P}_{y_i} = \\text{'a photo of a visual object'}$ in Eq. (9) to get $\\widetilde{\\mathcal{H}}_i$ as consistent signal to align the sample's embedding. Note that it is also possible to use class names of other datasets (e.g., CIFAR100 [20]) to generate $\\widetilde{\\mathcal{F}}_{\\mathcal{N}_i}$ as negative pairs. By this self-supervised way, the objective term of local training becomes $\\mathcal{L}_{TDCL} + \\mathcal{L}_{NDCR}$ . The results in Table. 5 validate the effectiveness of FedDifRC's self-supervised scheme.", + "bbox": [ + 511, + 320, + 906, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Convergence Analysis", + "text_level": 1, + "bbox": [ + 511, + 560, + 718, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We proved convergence of FedDifRC in non-convex case. We denote $\\mathcal{L}$ (Eq. (14)) as $\\mathcal{L}_r$ with the round $r$ , assumptions (Appendix B.2) are similar to existing works [6, 29, 30, 69].", + "bbox": [ + 511, + 582, + 905, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Convergence Rate of FedDifRC: Let the round $r$ from 0 to $R - 1$ , given any $\\xi > 0$ , the $\\mathcal{L}$ will converge when,", + "bbox": [ + 511, + 627, + 905, + 657 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nR > \\frac {2 \\left(\\mathcal {L} _ {0} - \\mathcal {L} ^ {*}\\right)}{\\xi E \\eta \\left(2 - L _ {1} \\eta\\right) - \\Omega_ {1} - \\Omega_ {2}}, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 664, + 903, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and we can further get the following condition for $\\eta$ ,", + "bbox": [ + 511, + 703, + 861, + 718 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\eta < \\frac {2 \\xi - 2 (| \\mathcal {C} | - 1) L _ {2} B}{L _ {1} (\\xi + \\sigma^ {2})}, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 723, + 903, + 757 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\Omega_{1} = 2(|\\mathcal{C}| - 1)L_{2}E\\eta B$ , $\\Omega_{2} = L_{1}E\\eta^{2}\\sigma^{2}$ , and $\\mathcal{L}^*$ denotes the optimal solution of $\\mathcal{L}$ , $\\{L_1,L_2,B,\\sigma^2\\}$ are constants in the assumptions from Appendix B.2, and $|\\mathcal{C}|$ as the number of classes, $E$ as local epochs. Eq. (15) and Eq. (16) ensure convergence rate of FedDifRC, after selecting the number of rounds $R$ and the learning rate $\\eta$ . So the smaller $\\xi$ is, the larger $R$ is, which means that the tighter the bound is, the more rounds $R$ is required. The detailed assumptions and proofs are formally provided in Appendix B.", + "bbox": [ + 511, + 763, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3731", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6e31980dfb4ef5d2b7c98e7db683f60e661ea86f1bc920a4fd5270958307c603.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TDCLNDCRCIFAR10 (keep ρ = 1.0)
NID10.05NID10.2NID10.5NID2AVG
78.2784.6586.1172.6080.41
✓81.3986.0388.1675.6782.81
✓80.3586.4087.5475.3382.40
✓✓83.1488.2789.3176.4584.29
TDCLNDCRCIFAR10-LT (keep NID10.2)
ρ = 10ρ = 50ρ = 100ρ = 200AVG
75.9463.2552.9749.6960.46
✓76.6264.6653.6250.3361.31
✓76.1964.7554.3950.6761.50
✓✓77.5865.3055.0151.1262.25
", + "bbox": [ + 94, + 88, + 482, + 243 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9d2bf7a622cd7528623db9e232068555b1e30f14385edbb8f02c9f4830125a85.jpg", + "table_caption": [ + "Table 1. Ablation study for two key modules: TDCL and NDCR." + ], + "table_footnote": [], + "table_body": "
ScenariosL=1L=2L=3L=4Fused△
NID10.0580.2182.6082.1981.3383.14+0.54
NID10.286.1587.2887.8185.9288.27+0.46
NID10.587.1089.0388.9387.1789.31+0.28
NID274.2675.7375.6175.0076.45+0.72
", + "bbox": [ + 94, + 284, + 482, + 354 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Comparison of features from different layers $L$ and fused features $\\left( {L \\in \\{ 2,3,4\\} }\\right)$ in Eq. (8) and Eq. (9) to obtain ${\\widetilde{\\mathcal{F}}}_{i}$ and ${\\widetilde{\\mathcal{H}}}_{i}$ , best in bold and second with underline, $\\bigtriangleup$ means accuracy gap.", + "bbox": [ + 89, + 364, + 482, + 407 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/517db9bf2184c3a176aea6e8b9629536e038b6f96eba6ac97f1de245771dfd26.jpg", + "image_caption": [ + "Figure 6. Analysis of FedDifRC with time-steps $t$ in Eq. (9) (left) and temperature $\\tau$ in Eq. (11) (right) on CIFAR10 with NID10.2." + ], + "image_footnote": [], + "bbox": [ + 91, + 422, + 290, + 542 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ebd4fbe975aeb5c865726fdcd36c6ab2d7f042b84bcff6ce3b41d9248560b9ac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 290, + 422, + 482, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 609, + 223, + 626 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 635, + 281, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Datasets: We adopt popular benchmark datasets, including CIFAR10 & CIFAR100 [20], TinyImageNet [21]; we build unbalanced versions of these datasets by [15]: CIFAR10-LT & CIFAR100-LT, and TinyImageNet-LT; we use Digits [41] and Office-Caltech [19] for the domain shift task in FL.", + "bbox": [ + 89, + 657, + 482, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Scenarios: i) label shift: we consider two non-iid settings: $\\mathrm{NID1}_{\\alpha}$ follows Dirichlet distribution [59], where $\\alpha$ denotes the heterogeneity level; NID2 is a more extreme setting consists of 6 biased clients (each has only a single category) and 1 unbiased client has all categories. ii) imbalance shift: we shape the original dataset into a long-tailed distribution with $\\rho$ follow [15], $\\rho$ means the ratio between sample sizes of the most frequent and least frequent class, as $\\rho = \\max_{j} / \\min_{j}$ . iii) domain shift: Digits [41] includes 4 domains as MNIST (M), USPS (U), SVHN (SV), SYN (SY) with 10 categories. Office-Caltech [19] also consists 4 domains as Caltech (C),", + "bbox": [ + 89, + 734, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Webcam (W), Amazon (A), DSLR (D) with 10 overlapping categories. We initialize 10 clients and assign them domains randomly: Digits (M: 1, U: 4, SV: 3, SY: 2), Office-Caltech (C: 3, W: 2, A: 2, D: 3). Each client's dataset is randomly selected from domains (Digits: $10\\%$ , Office-Caltech: $20\\%$ ).", + "bbox": [ + 511, + 90, + 903, + 166 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Configurations: The global rounds $R$ as 100 and clients $K$ as 10, all methods have little or no gain with more rounds $R$ . To facilitate fair comparison, we follow in [26, 29], and use SGD optimizer to update models with a learning rate 0.01, a momentum of 0.9, a batch size of 64, the weight decay of $10^{-5}$ , and local epochs $E = 10$ . We use MobileNetV2 [48] for TinyImageNet, ResNet-10 [14] for others. The dimension $d$ as 512, $\\tau$ (Eq. (11)) as 0.06, $t$ (Eq. (9)) as 150.", + "bbox": [ + 511, + 166, + 903, + 286 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines: We compare our against several SOTA methods: standard-based FedAvg [34], FedProx [29]; contrast-based MOON [26], FedRCL [50]; aggregate-based FedCDA [60], FedDisco [69]; unsupervised FedEMA [77], FedLID [42]; prototype-based FedProto [52], FedNH [70].", + "bbox": [ + 511, + 287, + 903, + 363 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Validation Analysis", + "text_level": 1, + "bbox": [ + 511, + 375, + 697, + 390 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We give ablation study in Table. 1: i) TDCL brings significant performance improvements over the baseline, showing that TDCL is able to promote semantic spread-out property; ii) NDCR also yields solid gains, proving the importance of aligning instance embedding via consistency regularization; iii) combining TDCL and NDCR attain better performance, supporting our motivation to unleash meaningful guidance of diffusion representations to promote FL local training.", + "bbox": [ + 511, + 396, + 903, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Then, we prove the advantage of fused features (Eq. (8)). As shown in Table. 2, results on CIFAR10 reveal that fused features capture both semantics and local details, better than features on a single layer $L$ , and if the features with too low resolution ( $L = 1$ ) or over-matching ( $L = 4$ ) would damage the model's performance. To explore the influence of different time-steps $t$ (Eq. (9)), we conduct experiments on CIFAR10 with NID10.2 in Fig. 6 left, we add different amount of noise to get $\\tilde{x}_i^t$ in Eq. (9). We find that FedDifRC is robust to the choice of $t \\in [100, 300]$ ( $t = 150$ achieves the highest value). As shown in Fig. 6 right, we present results with different $\\tau$ in Eq. (11), which reveals that FedDifRC is not sensitive to the $\\tau \\in [0.04, 0.08]$ , indicating the effectiveness of $\\mathcal{U}$ (Eq. (10)), to balance relative distances among different samples. The accuracy no longer increases after $\\tau = 0.06$ . Extra results of $t$ and $\\tau$ are provided in Appendix C.2.5.", + "bbox": [ + 511, + 518, + 905, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Evaluation Results", + "text_level": 1, + "bbox": [ + 511, + 771, + 694, + 786 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results on different settings (NID1, NID2, $\\rho$ ) in Table. 3, FedDifRC consistently outperforms these counterparts, it confirms that FedDifRC effectively leverages the inherent fine-grained visual representation capabilities of diffusion models to promote heterogeneous FL training. The results of Office-Caltech with domain shift are shown in Table. 4, indicating that FedDifRC can acquire well-generalizable", + "bbox": [ + 511, + 794, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3732", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0e4ef91db3f3354ab7d51682e7c24789cad36c92892f859a6778e609c61cdac2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
FL BaselinesCIFAR10CIFAR100TinyImageNetCIFAR10-LT (keep NID10.2)
NID10.05NID10.2NID2NID10.05NID10.2NID2NID10.05NID10.2NID2ρ = 10ρ = 50ρ = 100ρ = 200
FedAvg[AISTAT'17]78.2784.6572.6055.9760.0850.5640.4142.8435.1575.9463.2552.9749.69
FedProx[MLSys'20]78.4284.5972.8156.2760.2150.2940.2042.1635.6276.0263.4352.8649.10
MOON[CVPR'21]80.7986.1073.3556.7961.4851.8140.7943.6336.1175.7363.8053.1749.31
FedProto[AAAI'22]77.8683.9070.8454.4859.2648.6039.2941.0234.6874.3162.2350.3747.26
FedNH[AAAI'23]80.2585.8273.6556.6861.4551.3140.2543.3036.2176.1563.3652.4948.92
FedDisco[ICML'23]81.3586.6374.7257.3461.7952.0540.8343.9237.6976.2764.4053.3349.75
FedCDA[ICLR'24]81.7086.8975.0757.7662.1752.3541.1844.0938.2476.4564.3053.8550.08
FedRCL[CVPR'24]82.0287.1175.1357.9862.0552.4941.3344.2138.8676.3464.1753.4149.91
FedDifRC83.1488.2776.4559.2263.3253.8742.6345.2840.0577.5865.3055.0151.12
", + "bbox": [ + 96, + 88, + 903, + 242 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e1ada7d8e5795b2fe78c76f3d38b250fe3f61e0243cf2a9d9de9a6656f3e36b2.jpg", + "table_caption": [ + "Table 3. Comparison results on several settings (label shift and imbalance shift) and datasets, the best in bold and second with underline." + ], + "table_footnote": [], + "table_body": "
FL BaselinesOffice-Caltech
CaltechWebcamAmazonDSLRAVG
FedDisco[ICML'23]62.3945.9577.3145.2057.71
FedCDA[ICLR'24]62.2546.5376.5846.1457.87
FedRCL[CVPR'24]60.1342.7675.6043.7755.56
FedDifRC63.3250.1978.6748.1260.08
", + "bbox": [ + 96, + 291, + 480, + 388 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a06ec8bf1d833be452a56e1ab841a9f8cd5340e9662047ed51ab796ecb4631a0.jpg", + "table_caption": [ + "Table 4. Comparison results on Office-Caltech with domain shift." + ], + "table_footnote": [], + "table_body": "
FL BaselinesClientsCIFAR100
NID10.05NID10.2NID10.5NID2
FedEMA[ICLR'22]K=1026.9029.4130.6421.52
FedLID[ICCV'23]K=1028.8430.4331.2223.65
FedDifRCK=1030.6131.4532.3025.73
", + "bbox": [ + 96, + 431, + 480, + 513 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1245e05e977f82990b29d227f4c15e5ffb1869b4caa32d6a8d2010474197ef62.jpg", + "table_caption": [ + "Table 5. Comparison results on self-supervised setting (Sec. 4.3)." + ], + "table_footnote": [], + "table_body": "
FL BaselinesCIFAR10CIFAR10-LT (NID102)
NID10.05NID10.2NID2ρ = 10ρ = 50ρ = 100
FedDDA[PRCV-23]82.0586.9474.4976.3163.9553.15
FedDISC[AAAF24]82.3387.1575.4276.6064.2354.07
FedRCL (♂/†)82.30♂/80.95†87.25♂/86.16†75.37♂/73.85†76.50♂/76.12†64.26♂/62.91†53.65♂/52.20†
FedRCL (♀ + †)81.5286.4874.3576.2263.6952.93
FedDifRC83.1488.2776.4577.5865.3055.01
", + "bbox": [ + 94, + 558, + 480, + 623 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7835b4b5a3eb1b642cf2d953750fbcacff39693e827201a7a9b4ce91429ad8a7.jpg", + "table_caption": [ + "Table 6. More comparison results on CIFAR10 and CIFAR10-LT." + ], + "table_footnote": [], + "table_body": "
FL BaselinesHAM10000 [56]EuroSAT [40]
NID10.2NID10.5NID2NID10.2NID10.5NID2
FedDisco[ICML'23]50.2653.6844.6378.7682.3670.34
FedCDA[ICLR'24]51.6554.0545.1680.4583.3871.10
FedRCL[CVPR'24]52.2154.2045.4879.8083.6970.73
FedDifRC53.7955.6047.1782.4986.0573.14
", + "bbox": [ + 94, + 667, + 480, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "property in the feature space. With self-supervised scenario, in Table. 5, our FedDifRC achieves excellent performance by leveraging the visual perception capability of diffusion models, without requiring any labeled data. Please refer to Appendix C for more details and results (e.g., domain shift, convergence rates, different SD versions, etc).", + "bbox": [ + 89, + 810, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d2a3fa09853521fd4ccab6eb97c7c23353de060e0636270c8f12ae20e5825408.jpg", + "table_caption": [ + "Table 7. Comparison results on the HAM10000 and EuroSAT." + ], + "table_footnote": [], + "table_body": "
Pre-trained ModelsCIFAR10CIFAR10-LT (NID10.2)
NID10.05NID10.2NID2ρ = 10ρ = 50ρ = 100
CLIP-ViT-B/16 [45]79.7284.9173.9774.8962.4450.87
MAE-ViT-B/16 [10]80.7485.3873.2475.4563.2752.55
DINO-ViT-B/16 [33]80.5786.1474.1675.2562.7051.92
FedDifRC83.1488.2776.4577.5865.3055.01
", + "bbox": [ + 517, + 291, + 903, + 372 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. Experiments on aligning features of pre-trained models.", + "bbox": [ + 516, + 383, + 900, + 397 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further elaborate on the core idea of our FedDifRC, we then perform the following experiments:", + "bbox": [ + 511, + 424, + 903, + 453 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "i) To ensure a fair comparison, we also compare diffusion-based FL methods FedDISC [67] and FedDDA [74], we generate data for each client (same amount) by the SD model as $\\diamond$ , and align SD features (the image is fed directly into the SD model) by L2 loss as $\\dagger$ . Table. 6 shows that directly matching SD features may even hurt the performance, which further illustrates the superiority of our FedDifRC. ii) We focus on visual perception ability of diffusion models to understand an image (e.g., textures and structures) rather than pre-trained prior knowledge. We conduct experiments on medical and remote sensing datasets HAM10000 [56] and EuroSAT [40], which are not crossed in SD pre-training dataset LAION-5B [49]. Table. 7 show that the success of FedDifRC is based on the fine-grained visual representation rather than the large prior knowledge in SD model. iii) The SD model can implicitly provide both high-level and low-level visual concepts of semantic objects, while foundation models focus more on sparse visual matching. The Table. 8 show that FedDifRC can capture underlying properties and correspondences of visual objects.", + "bbox": [ + 511, + 454, + 906, + 755 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 768, + 633, + 784 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we aim to exploit the fine-grained visual representation capabilities of diffusion models to handle data heterogeneity in FL. We propose a novel diffusion-inspired FL framework FedDifRC to regulate local training via denoising and conditional diffusion representations. By leveraging complementary strength of TDCL and NDCR, FedDifRC achieves superior performance in heterogeneous FL.", + "bbox": [ + 511, + 794, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3733", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported by the scholarship from the China Scholarship Council (CSC) while the first author pursues his PhD degree at the University of Wollongong. This work was also partially supported by Australian Research Council Linkage Project LP210300009 and LP230100083.", + "bbox": [ + 89, + 114, + 483, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 205, + 187, + 220 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, pages 1597-1607, 2020. 2, 3, 5, 6", + "[2] Ching-Yao Chuang, Joshua Robinson, Yen-Chen Lin, Antonio Torralba, et al. Debiased contrastive learning. In NeurIPS, pages 8765-8775, 2020. 3", + "[3] Kevin Clark and Priyank Jain. Text-to-image diffusion models are zero shot classifiers. In NeurIPS, 2023. 3", + "[4] Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah. Diffusion models in vision: A survey. IEEE TPAMI, 45(9):10850-10869, 2023. 2, 3", + "[5] Jiequan Cui, Zhisheng Zhong, Shu Liu, Bei Yu, et al. Parametric contrastive learning. In ICCV, pages 715-724, 2021. 3", + "[6] Liang Gao, Huazhu Fu, Li Li, Yingwen Chen, et al. Feddc: Federated learning with non-iid data via local drift decoupling and correction. In CVPR, pages 10112-10121, 2022, 2, 6", + "[7] Jie Gui, Tuo Chen, Jing Zhang, Qiong Cao, et al. A survey on self-supervised learning: Algorithms, applications, and future trends. IEEE TPAMI, 46(12):9052-9071, 2024. 3", + "[8] Sungwon Han, Sungwon Park, Fangzhao Wu, Sundong Kim, et al. Fedx: Unsupervised federated learning with cross knowledge distillation. In ECCV, pages 691-707, 2022. 3", + "[9] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, et al. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9729-9738, 2020. 2, 3, 5", + "[10] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, et al. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 8", + "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, pages 6840-6851, 2020. 2, 3", + "[12] Wenke Huang, Mang Ye, and Bo Du. Learn from others and be yourself in heterogeneous federated learning. In CVPR, pages 10143-10153, 2022. 2, 3", + "[13] Wenke Huang, Mang Ye, Zekun Shi, He Li, and Bo Du. Rethinking federated learning with domain shift: A prototype view. In CVPR, pages 16312-16322, 2023. 2", + "[14] Cheng Ju, Aurélien Bibaut, and Mark van der Laan. The relative performance of ensemble methods with deep convolutional neural networks for image classification. Journal of Applied Statistics, 45:2800-2818, 2018. 7", + "[15] Cao Kaidi, Colin Wei, Adrien Gaidon, Nikos Arechiga, et al. Learning imbalanced datasets with label-distribution-aware margin loss. In NeurIPS, 2019. 7" + ], + "bbox": [ + 93, + 229, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Peter Kairouz, H Brendan McMahan, Brendan Avent, Aurélien Bellet, et al. Advances and open problems in federated learning. Foundations and Trends in Machine Learning, 14(1):1-210, 2021. 1", + "[17] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank Reddi, et al. Scaffold: Stochastic controlled averaging for federated learning. In ICML, pages 5132-5143, 2020. 2", + "[18] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, et al. Supervised contrastive learning. In NeurIPS, pages 18661-18673, 2020. 3", + "[19] Boqing Kong, Yuan Shi, Fei Sha, and Kristen Grauman. Geodesic flow kernel for unsupervised domain adaptation. In CVPR, pages 2066-2073, 2012. 7", + "[20] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. In Technical report, University of Toronto, 2009. 6, 7", + "[21] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7, 2015. 6, 7", + "[22] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 4", + "[23] Sunwoo Lee, Tuo Zhang, and A Salman Avestimehr. Layerwise adaptive model aggregation for scalable federated learning. In AAAI, pages 8491-8499, 2023. 2", + "[24] Alexander C Li, Mihir Prabhudesai, Shivam Duggal, Ellis Brown, et al. Your diffusion model is secretly a zero-shot classifier. In ICCV, pages 2206-2217, 2023. 3", + "[25] Daixun Li, Weiying Xie, Zixuan Wang, Yibing Lu, et al. Feddiff: Diffusion model driven federated learning for multimodal and multi-clients. IEEE TCSVT, 34(10):10353-10367, 2024. 3", + "[26] Qinbin Li, Bingsheng He, and Dawn Song. Model-contrastive federated learning. In CVPR, pages 10713-10722, 2021. 2, 3, 5, 7", + "[27] Qinbin Li, Yiqun Diao, Quan Chen, and Bingsheng He. Federated learning on non-iid data silos: An experimental study. In ICDE, pages 965–978, 2022. 1", + "[28] Tian Li, Anit Kumar Sahu, Ameet Talwalkar, and Virginia Smith. Federated learning: Challenges, methods, and future directions. IEEE SPM, 37(3):50-60, 2020. 1", + "[29] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, et al. Federated optimization in heterogeneous networks. MLSys, 2:429-450, 2020. 1, 2, 6, 7", + "[30] Xiang Li, Kaixuan Huang, Wenhao Yang, Shusen Wang, et al. On the convergence of fedavg on non-iid data. In ICLR, 2020. 2, 6", + "[31] Grace Luo, Lisa Dunlap, Dong Huk Park, Aleksander Holynski, et al. Diffusion hyperfeatures: Searching through time and space for semantic correspondence. In NeurIPS, 2023. 3", + "[32] Mi Luo, Fei Chen, Dapeng Hu, Yifan Zhang, et al. No fear of heterogeneity: Classifier calibration for federated learning with non-iid data. In NeurIPS, pages 5972-5984, 2021. 1", + "[33] Caron Mathilde, Hugo Touvron, Ishan Misra, Hervé Jégou, et al. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. 8" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3734", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, et al. Communication-efficient learning of deep networks from decentralized data. In AISTATS, pages 1273-1282, 2017. 1, 2, 7", + "[35] Matias Mendieta, Taojiannan Yang, Pu Wang, Minwoo Lee, et al. Local learning matters: Rethinking data heterogeneity in federated learning. In CVPR, pages 8397-8406, 2022. 1, 2", + "[36] John Nguyen, Kshitiz Malik, Hongyuan Zhan, Ashkan Yousefpour, et al. Federated learning with buffered asynchronous aggregation. In AISTATS, pages 3581-3607, 2022. 2", + "[37] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, pages 8162-8171, 2021. 2, 3", + "[38] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. In NeurIPS, 2019. 3", + "[39] Reese Pathak and Martin J Wainwright. Fedsplit: An algorithmic framework for fast federated optimization. In NeurIPS, pages 7057-7066, 2020. 2", + "[40] Helber Patrick, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE J-STARS, 12(7):2217-2226, 2019. 8", + "[41] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, et al. Moment matching for multi-source domain adaptation. In ICCV, pages 1406-1415, 2019. 7", + "[42] Athanasios Psaltis, Anestis Kastellos, Charalampos Z Patrikakis, and Petros Daras. Fedlid: Self-supervised federated learning for leveraging limited image data. In ICCV, pages 1039-1048, 2023. 7", + "[43] Pian Qi, Diletta Chiaro, Antonella Guzzo, Michele Ianni, et al. Model aggregation techniques in federated learning: A comprehensive survey. Future Generation Computer Systems, 150:272-293, 2024. 2", + "[44] Liangqiong Qu, Yuyin Zhou, Paul Pu Liang, Yingda Xia, et al. Rethinking architecture design for tackling data heterogeneity in federated learning. In CVPR, pages 10061-10071, 2022. 1", + "[45] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 8", + "[46] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, et al. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2, 3, 5", + "[47] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234-241, 2015. 2, 3, 4", + "[48] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhominov, et al. Mobilenetv2: Inverted residuals and linear bottlenecks. In CVPR, pages 4510-4520, 2018. 7", + "[49] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, pages 25278–25294, 2022. 8" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Seonguk Seo, Jinkyu Kim, Geeho Kim, and Bohyung Han. Relaxed contrastive learning for federated learning. In CVPR, pages 12279-12288, 2024. 7", + "[51] Fiona Victoria Stanley Jothiraj and Afra Mashhadi. Phoenix: A federated generative diffusion model. In ACM WWW, pages 1568-1577, 2024. 3", + "[52] Yue Tan, Guodong Long, Lu Liu, Tianyi Zhou, et al. FedProto: Federated prototype learning across heterogeneous clients. In AAAI, pages 8432-8440, 2022. 2, 7", + "[53] Yue Tan, Guodong Long, Jie Ma, Lu Liu, et al. Federated learning from pre-trained models: A contrastive learning approach. In NeurIPS, pages 19332-19344, 2022. 3", + "[54] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, et al. Emergent correspondence from image diffusion. In NeurIPS, pages 1363-1389, 2023. 3", + "[55] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, et al. What makes for good views for contrastive learning? In NeurIPS, pages 6827-6839, 2020. 3", + "[56] Philipp Tschandl, Cliff Rosendahl, and Harald Kittler. The ham10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. Nature Scientific Data, 5(1):1-9, 2018. 8", + "[57] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9 (11):2579-2605, 2008. 4", + "[58] Feng Wang and Huaping Liu. Understanding the behaviour of contrastive loss. In CVPR, pages 2495-2504, 2021. 2", + "[59] Hongyi Wang, Mikhail Yurochkin, Yuekai Sun, Dimitris Papailiopoulos, et al. Federated learning with matched averaging. In ICLR, 2020. 7", + "[60] Haozhao Wang, Haoran Xu, Yichen Li, Yuan Xu, et al. Fed-cda: Federated learning with cross-rounds divergence-aware aggregation. In ICLR, 2024. 2, 7", + "[61] Yingheng Wang, Yair Schiff, Aaron Gokaslan, Weishen Pan, et al. Infodiffusion: Representation learning using information maximizing diffusion models. In ICML, pages 36336-36354, 2023. 4", + "[62] Yuan Wang, Huazhu Fu, Renuga Kanagavelu, Qingsong Wei, et al. An aggregation-free federated learning for tackling data heterogeneity. In CVPR, pages 26233-26242, 2024. 2", + "[63] Chen Wei, Karttikeya Mangalam, Po-Yao Huang, Yanghao Li, et al. Diffusion models as masked autoencoders. In ICCV, pages 16284-16294, 2023. 3", + "[64] Weilai Xiang, Hongyu Yang, Di Huang, and Yunhong Wang. Denoising diffusion autoencoders are unified self-supervised learners. In ICCV, pages 15802-15812, 2023. 3", + "[65] Yikai Yan, Chaoyue Niu, Yucheng Ding, Zhenzhe Zheng, et al. Federated optimization under intermittent client availability. INFORMS Journal on Computing, 36(1):185-202, 2024. 2", + "[66] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, et al. Diffusion models: A comprehensive survey of methods and applications. ACM Computing Surveys, 56(4):1-39, 2023. 3", + "[67] Mingzhao Yang, Shangchao Su, Bin Li, and Xiangyang Xue. Exploring one-shot semi-supervised federated learning with pre-trained diffusion models. In AAAI, pages 16325-16333, 2024. 3, 8" + ], + "bbox": [ + 516, + 90, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3735", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[68] Xingyi Yang and Xinchao Wang. Diffusion model as representation learner. In ICCV, pages 18938-18949, 2023. 4", + "[69] Rui Ye, Mingkai Xu, Jianyu Wang, Chenxin Xu, et al. Feddisco: Federated learning with discrepancy-aware collaboration. In ICML, pages 39879-39902, 2023. 2, 6, 7", + "[70] Dai Yutong, Zeyuan Chen, Junnan Li, Shelby Heinecke, et al. Tackling data heterogeneity in federated learning with class prototypes. In AAAI, pages 7314-7322, 2023. 7", + "[71] Daokun Zhang, Jie Yin, Xingquan Zhu, and Chengqi Zhang. Network representation learning: A survey. IEEE Transactions on Big Data, 6:3-28, 2018. 4", + "[72] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, et al. A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence. In NeurIPS, 2023. 3", + "[73] Jianqing Zhang, Yang Hua, Hao Wang, Tao Song, et al. Fedala: Adaptive local aggregation for personalized federated learning. In AAAI, pages 11237-11244, 2023. 2", + "[74] Zhuang Zhao, Feng Yang, and Guirong Liang. Federated learning based on diffusion model to cope with non-iid data. In PRCV, pages 220-231, 2023. 3, 8", + "[75] Hangyu Zhu, Jinjin Xu, Shiqing Liu, and Yaochu Jin. Federated learning on non-iid data: A survey. Elsevier Neurocomputing, 465:371-390, 2021. 1", + "[76] Fuzhen Zhuang, Xiaohu Cheng, Ping Luo, Sinno Jialin Pan, et al. Supervised representation learning: Transfer learning with deep autoencoders. In IJCAI, 2015. 4", + "[77] Weiming Zhuang, Yonggang Wen, and Shuai Zhang. Divergence-aware federated self-supervised learning. In ICLR, 2022. 7" + ], + "bbox": [ + 91, + 92, + 480, + 517 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "3736", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/FedDifRC_ Unlocking the Potential of Text-to-Image Diffusion Models in Heterogeneous Federated Learning/d48a8fed-132a-4154-9395-1fe6c487ac0e_model.json b/2025/FedDifRC_ Unlocking the Potential of Text-to-Image Diffusion Models in Heterogeneous Federated Learning/d48a8fed-132a-4154-9395-1fe6c487ac0e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..42eae8ac5d843561881d5b458992940b7684194d --- /dev/null +++ b/2025/FedDifRC_ Unlocking the Potential of Text-to-Image Diffusion Models in Heterogeneous Federated Learning/d48a8fed-132a-4154-9395-1fe6c487ac0e_model.json @@ -0,0 +1,3181 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.134, + 0.13, + 0.868, + 0.177 + ], + "angle": 0, + "content": "FedDifRC: Unlocking the Potential of Text-to-Image Diffusion Models in Heterogeneous Federated Learning" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.203, + 0.924, + 0.295 + ], + "angle": 0, + "content": "Huan Wang\\(^{1,\\ast}\\), Haoran Li\\(^{1,3,\\ast}\\), Huaming Chen\\(^{2}\\), Jun Yan\\(^{1}\\), Jiahua Shi\\(^{3}\\), Jun Shen\\(^{1,\\dagger}\\) \n\\(^{1}\\)School of Computing and Information Technology, University of Wollongong, Wollongong, Australia \n\\(^{2}\\)School of Electrical and Computer Engineering, The University of Sydney, Sydney, Australia \n\\(^{3}\\)QLD Alliance for Agriculture and Food Innovation, The University of Queensland, Brisbane, Australia \nhw226@uowmail.edu.au *Equal Contribution †Corresponding Author" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.328, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.36, + 0.486, + 0.799 + ], + "angle": 0, + "content": "Federated learning aims at training models collaboratively across participants while protecting privacy. However, one major challenge for this paradigm is the data heterogeneity issue, where biased data preferences across multiple clients, harming the model's convergence and performance. In this paper, we first introduce powerful diffusion models into the federated learning paradigm and show that diffusion representations are effective steers during federated training. To explore the possibility of using diffusion representations in handling data heterogeneity, we propose a novel diffusion-inspired Federated paradigm with Diffusion Representation Collaboration, termed FedDifRC, leveraging meaningful guidance of diffusion models to mitigate data heterogeneity. The key idea is to construct text-driven diffusion contrasting and noise-driven diffusion regularization, aiming to provide abundant class-related semantic information and consistent convergence signals. On the one hand, we exploit the conditional feedback from the diffusion model for different text prompts to build a text-driven contrastive learning strategy. On the other hand, we introduce a noise-driven consistency regularization to align local instances with diffusion denoising representations, constraining the optimization region in the feature space. In addition, FedDifRC can be extended to a self-supervised scheme without relying on any labeled data. We also provide a theoretical analysis for FedDifRC to ensure convergence under non-convex objectives. The experiments on different scenarios validate the effectiveness of FedDifRC and the efficiency of crucial components. Code is available at https://github.com/hwang52/FedDifRC." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.829, + 0.224, + 0.845 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.855, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Federated Learning (FL) [34] has become increasingly popular as a decentralized training paradigm, with multiple distributed clients train their local models without sharing their" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.329, + 0.907, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.455, + 0.874, + 0.47 + ], + "angle": 0, + "content": "(a) Text-Driven Diffusion Contrastive Learning (TDCL)" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.476, + 0.906, + 0.605 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.542, + 0.609, + 0.885, + 0.624 + ], + "angle": 0, + "content": "(b) Noise-Driven Diffusion Consistent Restriction (NDCR)" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.633, + 0.909, + 0.746 + ], + "angle": 0, + "content": "Figure 1. The main idea of FedDifRC with two key modules TDCL and NDCR. (a) We leverage conditional generative feedback of the pre-trained diffusion model on various text prompts to devise a text-driven contrastive learning strategy, where the input embedding is injected as a condition. (b) We extract the diffusion denoising feature maps of intermediate layers from the pre-trained diffusion model at a specific time step \\( t \\) during the diffusion backward process to construct a noise-driven regularization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "private dataset [16, 28] (i.e., training client models locally and aggregating them globally [29]). One major challenge faced in FL is the potential discrepancies in the local data distributions across distributed clients, which is known as the data heterogeneity issue [32, 35, 44]. Specifically, the private data is collected from different sources with diverse preferences and presents non-independent and identically distributed (non-iid) distribution [27, 75]. Such non-iid data" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "3726" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "discrepancies can lead to inconsistency in the clients' local objective functions and optimization directions, making the model convergence slow and unstable [17, 29, 30, 39, 65]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.139, + 0.483, + 0.365 + ], + "angle": 0, + "content": "To mitigate the negative impact of the data heterogeneity, a mainstream of subsequent efforts delves into introducing a variety of suited signals to regulate the private model, there are mainly two perspectives: i) local model optimization at the client side and ii) global model aggregation at the server side. The former focuses on optimization strategies to make the variance of local updates limited [6, 29]. Another aims to improve the efficacy of global model aggregation [36, 69, 73]. In fact, most of these previous methods force the local models to be consistent to the global model. Although they have a certain effect by reducing gradient inconsistency, the gradually enlarged parameter deviation persists. Hence, we argue that local model overfits the local data distribution, which is magnified under the data heterogeneity and thus hinders the improvement of model's performance in FL." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.368, + 0.483, + 0.58 + ], + "angle": 0, + "content": "Recently, the development of diffusion models (DMs) [4, 11, 37, 46] provides fresh opportunities. The diffusion models have demonstrated remarkable power in generating the diverse and realistic images, showing strong ability to capture intricate details and patterns from underlying data. For instance, with proper guidance, stable diffusion (SD) [46] models can effectively align a text prompt with the corresponding image context, which suggests their potential ability to understand both high-level and low-level contextual visual concepts regarding what an image contains. This semantic generative and visual comprehension capabilities of SD models motivate us to think: is it possible to exploit the inherent fine-grained visual representation capabilities of diffusion models to promote heterogeneous FL training?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.483, + 0.868 + ], + "angle": 0, + "content": "Taking into account both the effectiveness and efficiency, we revisit the underlying representations in the SD models, which are intermediate feature maps in a diffusion process (e.g., denoise or noise an image) produced by multiple calling a UNet [47]. We argue that diffusion representations are effective steers for FL training: i) extensive general knowledge inherent in diffusion models enhance local semantic diversity in FL; ii) smooth correspondences about semantic objects in diffusion models is a natural guiding signal in FL. Building on these two insights, we first propose conditional diffusion representations by imposing various text prompts in the SD model. In this way, each instance is contrastively enhanced with a set of text-driven diffusion representations, capturing rich domain variance. Then, we devise denoising diffusion representations from semantic relations between visual and linguistic concepts in the SD model. We extract the feature maps from different hierarchies of the UNet decoder in the SD model's denoising process, serving as guidance signals to avoid over-optimization in FL local training." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose a diffusion-inspired Federated framework with Diffusion Representation Collaboration, as" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.424 + ], + "angle": 0, + "content": "FedDifRC, to handle the data heterogeneity issue in FL, which consists of two key modules (see Fig. 1): i) First, we devise Text-driven Diffusion Contrastive Learning (TDCL), which utilizes conditional diffusion representations to construct inter-contrastive learning [1, 9, 26, 58] in a text-driven diffusion manner (the sample's embedding is injected as a condition). By maximizing the agreement between the local instance and its corresponding representations, TDCL adaptively captures class-relevant information and semantically meaningful knowledge. This objective encourages sample's embedding closer to conditional diffusion representations with same semantics and away from other negatives, thus maintaining a clear decision boundary. ii) Second, we use the denoising diffusion representations to build a flat and stable convergent target and further propose Noise-driven Diffusion Consistent Restriction (NDCR). Concretely, we extract denoising feature maps of the noisy input sample in the SD model's backward process, as a regularization term to constrain the optimization region of the local model. The local sample is forced to align the embedding with its denoising representations at the feature space to promote the uniformity for local training." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.425, + 0.874, + 0.439 + ], + "angle": 0, + "content": "Our main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.442, + 0.905, + 0.501 + ], + "angle": 0, + "content": "- We explore the possibility of using diffusion models to tackle data heterogeneity in FL and devise the conditional and denoising diffusion representations as effective steers to regulate model training in heterogeneous FL." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.502, + 0.905, + 0.591 + ], + "angle": 0, + "content": "- We propose FedDifRC, a diffusion-inspired FL framework to handle data heterogeneity by leveraging the complementary advantages of two key components: i) TDCL facilitates the local model to learn more general class-relevant knowledge; ii) NDCR enforces the uniformity in the feature space for local training." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.592, + 0.905, + 0.668 + ], + "angle": 0, + "content": "- We conduct extensive experiments for different FL scenarios (e.g., label shift, domain shift, and long-tailed) and different application scenarios (i.e., intra-domain and cross-domain) to demonstrate the effectiveness of our FedDifRC and the indispensability of each component." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.442, + 0.905, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.685, + 0.655, + 0.701 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.712, + 0.821, + 0.728 + ], + "angle": 0, + "content": "2.1. Heterogeneous Federated Learning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Federated learning is proposed to address privacy concerns in a distributed learning environment (e.g., Federated Averaging (FedAvg) [34]). Several studies have improved the local update process on the client side or the global aggregation process on the server side to tackle data heterogeneity in FL. The former focuses on local optimization strategies to make the diversity between the client model and the global model limited in the parameter level [6, 12, 13, 17, 26, 29, 35, 52]. The another branch aims to adopt aggregation mechanisms on the server to alleviate the negative influence of data heterogeneity [23, 43, 60, 62, 69, 73]." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3727" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.258 + ], + "angle": 0, + "content": "Recently, existing FL works that apply diffusion models mainly focus on building synthetic datasets [25, 51, 67, 74]. With the assistance of synthetic datasets, the FL global model can be retrained by this data augmentation scheme to mitigate data heterogeneity. Although these generated data have comparable quality and diversity to client images, the local model may overfit local domain distribution, and the inherent heterogeneity across clients may not have been addressed. In contrast, we focus on how to effectively leverage the fine-grained visual representation from diffusion models to promote a generalizable FL global model." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.278, + 0.381, + 0.294 + ], + "angle": 0, + "content": "2.2. Diffusion Model Representations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.304, + 0.485, + 0.516 + ], + "angle": 0, + "content": "Recently, diffusion models have shown significant advances in image generation, editing, and stylized [4, 11, 37, 46, 66], suggesting that they contain rich underlying representations to be exploited for different downstream tasks. DIFT [54] is used to extract the diffusion features without explicit supervision. Diffusion Hyperfeatures [31] integrates multi-scale and multi-timesteps diffusion features into feature descriptors. The diffusion representations have shown impressive results on the vision branch, including being testified as certifiable classifiers [3, 24, 63] and semantics correspondence learners [54, 64, 72]. In this paper, we expand the diffusion models to FL and construct the conditional and denoising diffusion representations, offering fruitful class-relevant information and consistent convergence signals." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.535, + 0.295, + 0.552 + ], + "angle": 0, + "content": "2.3. Contrastive Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.561, + 0.484, + 0.759 + ], + "angle": 0, + "content": "Contrastive learning (CL) has been widely applied to self-supervised learning scenarios [1, 7, 9]. Many works are focused on learning the encoder where the embeddings of the same sample are pulled closer and those of different samples are pushed apart [2, 5, 18, 55]. They mainly construct a positive pair and some negative pairs for each data instance and use InfoNCE [38] to contrast positiveness against negativity. There are some works that incorporate contrastive learning into FL to assist local training to achieve higher model performance [8, 12, 26, 53]. In our work, Text-driven Diffusion Contrastive Learning (TDCL) is devised to promote local model by attracting samples to positive representations while pushing away other negatives." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.781, + 0.476, + 0.8 + ], + "angle": 0, + "content": "3. Diffusion Models Help Federated Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In this section, we begin with a brief overview of the general FL paradigm and the generative diffusion models (Sec. 3.1). Then, we analyze the effectiveness of diffusion representations and illustrate our core insights (Sec. 3.2). Finally, we will introduce how to extract the conditional and denoising diffusion representations in Sec. 3.3 and Sec. 3.4." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.787, + 0.108 + ], + "angle": 0, + "content": "3.1. Preliminaries and Background" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.113, + 0.907, + 0.251 + ], + "angle": 0, + "content": "Federated Learning: We focus on the case with \\(K\\) clients holding the heterogeneous data partition \\(\\{\\mathcal{D}_1,\\mathcal{D}_2,\\ldots ,\\mathcal{D}_K\\}\\) with private dataset \\(\\mathcal{D}_k = \\{x_i,y_i\\}_{i = 1}^{n_k},y_i\\in \\{1,\\dots ,|\\mathcal{C}|\\}\\) for the \\(k\\) -th client, where \\(n_k\\) denotes the local dataset scale. Let \\(n_k^j\\) be the number of samples with label \\(j\\) at client \\(k\\) and \\(\\mathcal{D}_k^j = \\{(x,y)\\in \\mathcal{D}_k|y = j\\}\\) denotes the set of samples with label \\(j\\) on the \\(k\\) -th client, so the number of samples in \\(j\\) -th class is \\(n^j = \\sum_{k = 1}^{K}n_k^j\\) . The objective of the general FL is to learn an optimal global model \\(w\\) across \\(K\\) clients:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.257, + 0.907, + 0.299 + ], + "angle": 0, + "content": "\\[\n\\min _ {w} \\mathcal {L} (w) = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\frac {n _ {k}}{N} \\mathcal {L} _ {k} (w; \\mathcal {D} _ {k}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.303, + 0.907, + 0.424 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_k\\) is the \\(k\\)-th client local loss function and \\(w\\) is the global model parameters. \\(N\\) is the total number of samples among all clients. In the FL training setup, we consider the local model \\(f\\) with parameters \\(w = \\{u,v\\}\\). It has two modules: i) a feature extractor \\(h(u)\\) with the parameters \\(u\\) maps each sample \\(x_{i}\\) to a \\(d\\)-dim (default as 512) vector as \\(z_{i} = h(u;x_{i})\\); ii) a classifier \\(g(v)\\) with parameters \\(v\\) maps \\(z_{i}\\) into a \\(|\\mathcal{C}|\\)-dim output logits as \\(l_{i} = g(v;z_{i})\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.425, + 0.906, + 0.499 + ], + "angle": 0, + "content": "Diffusion Models: Given the image sample \\( x \\) from an underlying distribution \\( p(x) \\), a forward diffusion process as a Markov chain to add random Gaussian noise \\( \\epsilon \\sim \\mathcal{N}(0, \\mathbf{I}) \\) to the original sample \\( x_0 \\), and \\( x_t \\) is the noise sample after adding \\( t \\)-steps noise:" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.505, + 0.907, + 0.523 + ], + "angle": 0, + "content": "\\[\nx _ {t} = \\sqrt {1 - \\gamma_ {t}} x _ {t - 1} + \\sqrt {\\gamma_ {t}} \\epsilon_ {t}, t \\in \\{1, \\dots , T \\}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.529, + 0.906, + 0.635 + ], + "angle": 0, + "content": "where \\(T\\) is the number of diffusion timesteps, \\(\\gamma_t \\in (0,1)\\) is a adjustable time variance schedule [11]. By leveraging the additive property of the Gaussian distribution, Eq. (2) can be further reformulated as \\(x_t = \\sqrt{\\bar{\\alpha}_t} x_0 + \\sqrt{1 - \\bar{\\alpha}_t} \\epsilon_t\\), where \\(\\alpha_t = 1 - \\gamma_t\\) and \\(\\bar{\\alpha}_t \\doteq \\prod_{i=1}^t \\alpha_i\\). On this basis, the image \\(x_0\\) can be obtained from a random noise \\(x_t \\sim \\mathcal{N}(0,\\mathbf{I})\\) by reversing the above forward diffusion process [11]:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.64, + 0.906, + 0.672 + ], + "angle": 0, + "content": "\\[\nx _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(x _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} \\left(x _ {t}, t\\right)\\right), t \\in \\{T, \\dots , 1 \\}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.677, + 0.906, + 0.738 + ], + "angle": 0, + "content": "where \\(\\epsilon_{\\theta}\\) denotes the denoiser network with the parameters \\(\\theta\\) to predict noise \\(\\epsilon\\), and the denoiser network is usually implemented as a UNet architecture [47]. Further, the denoiser \\(\\epsilon_{\\theta}\\) is optimized using the loss \\(T_{\\theta}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.743, + 0.905, + 0.761 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} _ {\\theta} = \\mathbb {E} _ {\\epsilon , x _ {0}, t} [ \\| \\epsilon_ {\\theta} (x _ {t}, t) - \\epsilon \\| _ {2} ^ {2} ], \\epsilon \\sim \\mathcal {N} (0, \\mathbf {I}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.906, + 0.843 + ], + "angle": 0, + "content": "where \\(\\epsilon\\) is the noise in the forward diffusion process, and \\(\\epsilon_{\\theta}\\) is usually implemented as a UNet [47]. Besides, the diffusion model's objective in Eq. (4) can be easily extended to a conditional diffusion paradigm by adding a condition c to the \\(\\epsilon_{\\theta}\\). Thus, the training objective should be modified as:" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.848, + 0.905, + 0.866 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} _ {\\theta} = \\mathbb {E} _ {\\epsilon , x _ {0}, t, \\mathbf {c}} [ | | \\epsilon_ {\\theta} (x _ {t}, t, \\mathbf {c}) - \\epsilon | | _ {2} ^ {2} ], \\epsilon \\sim \\mathcal {N} (0, \\mathbf {I}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "with the denoiser \\(\\epsilon_{\\theta}(x_t,t,\\mathbf{c})\\) in Eq. (5), we can generate data with a specific condition \\(\\mathbf{c}\\) by solving a reverse process." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3728" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.315, + 0.108 + ], + "angle": 0, + "content": "3.2. Motivation and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.113, + 0.483, + 0.309 + ], + "angle": 0, + "content": "The diffusion model is a probabilistic generative model designed for denoising by systematically reversing a progressive-noising process, and recent studies [61, 68] have verified that diffusion-based generative models can be effective representation learners. Stable diffusion models are able to weave from scratch with fruitful and innovative depictions of high fidelity and further capture the underlying properties, positions, and correspondences from visual concepts. Such a phenomenon shows that SD models can align a textual prompt with its corresponding image context, demonstrating their potential ability in understanding both high-level (e.g., semantic relations) and low-level (e.g., textures, edge, and structures) visual concepts in an image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.31, + 0.484, + 0.432 + ], + "angle": 0, + "content": "Autoencoders [71, 76] can learn to reconstruct their input at the output and have been one of the dominant approaches for representation learning. We assume a linear autoencoder expressed as \\( W_{a}W_{b}\\cdot x + W_{c}\\cdot x \\) to imitate UNet [47] used in diffusion models, where \\( W_{a}\\in \\mathbb{R}^{d\\times d^{*}} \\), \\( W_{b}\\in \\mathbb{R}^{d^{*}\\times d} \\), and \\( W_{c}\\in \\mathbb{R}^{d\\times d} \\), \\( d^{*} < d \\). For \\( x_{t} = \\sqrt{\\bar{\\alpha}_{t}} x_{0} + \\sqrt{1 - \\bar{\\alpha}_{t}}\\epsilon_{t} \\) and we set \\( \\mathbb{E}[x_0] = 0 \\), thus minimizing the loss \\( \\mathcal{T}_{\\theta} \\) in Eq. (4) is equivalent to minimize the modified objective \\( \\mathcal{T}_{\\theta}^{*} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.103, + 0.442, + 0.484, + 0.479 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{r l} \\mathcal {T} _ {\\theta} ^ {*} = & \\mathbb {E} _ {\\epsilon , x _ {0}, t} [ \\| \\epsilon_ {\\theta} (\\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t}) - \\epsilon \\| _ {2} ^ {2} ], \\\\ & \\equiv \\min \\left\\{\\bar {\\alpha} _ {t} \\left(W _ {a} W _ {b} + W _ {c}\\right) ^ {\\top} \\Sigma_ {x _ {0}} \\left(W _ {a} W _ {b} + W _ {c}\\right) \\right\\}, \\end{array} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.493, + 0.484, + 0.584 + ], + "angle": 0, + "content": "where \\(\\Sigma_{x_0}\\) means covariance matrix of \\(x_0\\). Eq. (6) ensures a meaningful latent space for the diffusion models, similar to the linear autoencoder that encodes data into the principal components space. Therefore, the diffusion training process is equivalent to encourage the disentanglement of the latent representations until good-quality reconstruction is reached." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.585, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To validate the representation ability in diffusion models, we use t-SNE [57] to visualize the latent features, shown in Fig. 2, without training on the specific MNIST [22] dataset. Concretely, we use the pre-trained SD model to extract features of the UNet decoder \\(\\epsilon_{\\theta}\\), \\(T\\) means diffusion time-steps, \\(L\\) means the layer of the UNet decoder \\(\\epsilon_{\\theta}\\) (\\(L \\in \\{1, 2, 3, 4\\}\\)). Our results in Fig. 2 indicate that, even without supervision, diffusion models are able to group input samples at properly selected time-steps \\(T\\) and layer \\(L\\). However, when the \\(T\\) or \\(L\\) is too large (e.g., \\(T = 999\\) or \\(L = 4\\)), the representations become inseparable, leading to blurred decision boundaries. Furthermore, we apply K-Means clustering (cluster size set as 5) on the latent features from the diffusion UNet decoder and visually examine whether the diffusion features contain consistent semantic information. In Fig. 3, different parts of the visual objects are clustered, showing that earlier layers (e.g., \\(L = 2\\)) capture coarse yet consistent semantics, while later layers (e.g., \\(L = 3\\)) focus on low-level textural details. Both t-SNE and K-Means observation results motivate us to leverage the fine-grained visual representation capabilities from the SD model to promote heterogeneous FL training." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.614, + 0.136 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.139, + 0.597, + 0.15 + ], + "angle": 0, + "content": "\\(\\mathrm{T} = 100\\) \\(\\mathrm{L} = 1\\)" + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.091, + 0.707, + 0.136 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.63, + 0.139, + 0.695, + 0.15 + ], + "angle": 0, + "content": "\\(\\mathrm{T} = 100, \\mathrm{L} = 2\\)" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.09, + 0.805, + 0.136 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.726, + 0.139, + 0.791, + 0.15 + ], + "angle": 0, + "content": "\\(\\mathrm{T} = 100, \\mathrm{L} = 3\\)" + }, + { + "type": "image", + "bbox": [ + 0.808, + 0.09, + 0.904, + 0.136 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.824, + 0.139, + 0.888, + 0.15 + ], + "angle": 0, + "content": "\\(\\mathrm{T} = 100, \\mathrm{L} = 4\\)" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.152, + 0.614, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.201, + 0.598, + 0.212 + ], + "angle": 0, + "content": "\\(\\mathrm{T} = 100, \\mathrm{L} = 2\\)" + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.152, + 0.708, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.63, + 0.201, + 0.695, + 0.212 + ], + "angle": 0, + "content": "\\(\\mathrm{T} = 300\\) \\(\\mathrm{L} = 2\\)" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.152, + 0.805, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.726, + 0.201, + 0.792, + 0.212 + ], + "angle": 0, + "content": "\\(\\mathrm{T} = 500\\) \\(\\mathrm{L} = 2\\)" + }, + { + "type": "image", + "bbox": [ + 0.808, + 0.152, + 0.904, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.824, + 0.201, + 0.888, + 0.212 + ], + "angle": 0, + "content": "\\(\\mathrm{T} = 999, \\mathrm{L} = 2\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.224, + 0.907, + 0.267 + ], + "angle": 0, + "content": "Figure 2. The t-SNE feature visualization on the MNIST dataset from the diffusion UNet decoder, without training, where diffusion time-steps \\( T \\in \\{100, 300, 500, 999\\} \\) and layer \\( L \\in \\{1, 2, 3, 4\\} \\)." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.283, + 0.586, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.338, + 0.586, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.395, + 0.586, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.451, + 0.574, + 0.464 + ], + "angle": 0, + "content": "Sample" + }, + { + "type": "image", + "bbox": [ + 0.596, + 0.283, + 0.666, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.338, + 0.665, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.395, + 0.666, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.599, + 0.451, + 0.662, + 0.464 + ], + "angle": 0, + "content": "Layer \\(\\mathrm{L} = 1\\)" + }, + { + "type": "image", + "bbox": [ + 0.676, + 0.283, + 0.745, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.676, + 0.338, + 0.744, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.676, + 0.394, + 0.744, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.679, + 0.451, + 0.743, + 0.464 + ], + "angle": 0, + "content": "Layer \\( \\mathrm{L} = 2 \\)" + }, + { + "type": "image", + "bbox": [ + 0.756, + 0.283, + 0.824, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.338, + 0.824, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.756, + 0.394, + 0.824, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.759, + 0.451, + 0.822, + 0.464 + ], + "angle": 0, + "content": "Layer \\( \\mathrm{L} = 3 \\)" + }, + { + "type": "image", + "bbox": [ + 0.836, + 0.283, + 0.904, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.836, + 0.338, + 0.904, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.836, + 0.394, + 0.904, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.838, + 0.451, + 0.902, + 0.464 + ], + "angle": 0, + "content": "Layer L=4" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.477, + 0.907, + 0.505 + ], + "angle": 0, + "content": "Figure 3. The K-Means clustering feature visualization (each color represents a cluster) from the diffusion UNet decoder's layers." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.521, + 0.586, + 0.575 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.577, + 0.585, + 0.631 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.633, + 0.585, + 0.687 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.69, + 0.574, + 0.703 + ], + "angle": 0, + "content": "Sample" + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.522, + 0.665, + 0.575 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.577, + 0.664, + 0.631 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.633, + 0.664, + 0.687 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.599, + 0.69, + 0.663, + 0.702 + ], + "angle": 0, + "content": "Layer \\(L = 2\\)" + }, + { + "type": "image", + "bbox": [ + 0.676, + 0.522, + 0.745, + 0.575 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.676, + 0.577, + 0.744, + 0.631 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.676, + 0.633, + 0.744, + 0.687 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.679, + 0.69, + 0.742, + 0.702 + ], + "angle": 0, + "content": "Layer \\( \\mathrm{L} = 3 \\)" + }, + { + "type": "image", + "bbox": [ + 0.756, + 0.522, + 0.824, + 0.575 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.756, + 0.577, + 0.824, + 0.631 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.756, + 0.633, + 0.824, + 0.687 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.759, + 0.69, + 0.822, + 0.702 + ], + "angle": 0, + "content": "Layer L=4" + }, + { + "type": "image", + "bbox": [ + 0.836, + 0.522, + 0.904, + 0.575 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.836, + 0.577, + 0.904, + 0.631 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.836, + 0.633, + 0.904, + 0.687 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.851, + 0.69, + 0.891, + 0.702 + ], + "angle": 0, + "content": "Fused" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.716, + 0.907, + 0.758 + ], + "angle": 0, + "content": "Figure 4. The PCA visualization of early (layer 2), later (layer 4), and fused features (see Eq. (8)), where the first three components of PCA-computed serve as color channels for each image." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.787, + 0.844, + 0.803 + ], + "angle": 0, + "content": "3.3. Conditional Diffusion Representations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.901 + ], + "angle": 0, + "content": "The key idea of our conditional diffusion representations is based on cognitive diversity of generative diffusion models for different text prompts, via a condition-guided manner. Specifically, for the client \\( k \\), given an input image \\( x_{i} \\) of the local dataset \\( \\mathcal{D}_k = \\{x_i,y_i\\}_{i = 1}^{n_k},y_i\\in \\{1,\\dots,|\\mathcal{C}|\\} \\) means the label of \\( x_{i} \\), and we let \\( S \\) represents the set of all class names," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3729" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.306 + ], + "angle": 0, + "content": "so we use \\( S_{y_i} \\) to denote the class name of the label \\( y_i \\). From our implementation, the image \\( x_i \\) is fed into the encoder \\( h_k \\) of the local model \\( f_k \\) to obtain the embedding \\( \\mathbf{c}_i \\), then \\( \\mathbf{c}_i \\) is treated as a condition and injected into the diffusion model. For a pre-trained SD model, \\( \\mathbf{c}_i \\) is further combined with a relevant text prompt \\( \\mathcal{P}_{y_i} = \\) 'a photo of a \\( S_{y_i} \\)' and sent to the UNet decoder to produce diffusion representations \\( \\mathcal{F}_i \\). Note that since we do not add noise, we simply set the time-steps \\( T = 0 \\) of the diffusion process. Typically, the image \\( x_i \\) size is \\( 224 \\times 224 \\), and \\( \\mathcal{F}_i \\) contains 4 feature maps, where the \\( L \\)-th layer's feature map \\( \\mathcal{F}_{i(L)} \\) has the spatial size \\( \\{7,14,28,28\\} \\) (\\( L \\in \\{1,2,3,4\\} \\)), and \\( \\{1280,1280,640,320\\} \\) are channels. Formally, conditional diffusion representations \\( \\mathcal{F}_i \\) are computed by the \\( \\epsilon_{\\theta} \\) in Eq. (5), image \\( x_i \\), condition \\( \\mathbf{c}_i \\), text \\( \\mathcal{P}_{y_i} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.319, + 0.483, + 0.337 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {i} = \\epsilon_ {\\theta} \\left(x _ {i}, \\tau_ {i}, T = 0, \\mathbf {c} _ {i}\\right), \\tau_ {i} \\leftarrow \\varphi \\left(\\mathcal {P} _ {y _ {i}}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.343, + 0.484, + 0.495 + ], + "angle": 0, + "content": "where \\(\\varphi\\) denotes the text encoder of the SD model and \\(\\tau_{i}\\) is the text features with the prompt \\(\\mathcal{P}_{y_i}\\). Inspired by the results in Fig. 3, which motivate us to combine features at different layers of \\(\\mathcal{F}_i\\) to capture both semantics and details. A simple concatenation, however, can lead to an unnecessarily high-dimensional feature \\((1280 + 640 + 320 = 2240)\\). To reduce high dimension, we apply PCA for each layer \\(L\\in \\{2,3,4\\}\\) (due to the spatial resolution too low, we do not consider the \\(L = 1\\)), and we aggregate them to the same resolution, then we connect them together to obtain the fused features \\(\\widetilde{\\mathcal{F}}_i\\):" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.504, + 0.483, + 0.526 + ], + "angle": 0, + "content": "\\[\n\\widetilde {\\mathcal {F}} _ {i} = \\mathrm {C o n c a t} (\\widetilde {\\mathcal {F}} _ {i (L)}), \\widetilde {\\mathcal {F}} _ {i (L)} = \\mathrm {P C A} (\\mathcal {F} _ {i (L)}), \\quad (8)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.536, + 0.485, + 0.718 + ], + "angle": 0, + "content": "where \\(L\\in \\{2,3,4\\}\\) we perform PCA: i) for \\(\\mathcal{F}_{i(2)}\\) we first upsample to the same resolution and apply PCA \\(\\mathbb{R}^{1280\\to 256}\\) ii) for \\(\\mathcal{F}_{i(3)}\\) and \\(\\mathcal{F}_{i(4)}\\) , we apply PCA \\(\\mathbb{R}^{640\\to 128}\\) and PCA \\(\\mathbb{R}^{320\\to 128}\\) . Then, we combine them to form the final fused features \\(\\widetilde{\\mathcal{F}}_i\\) . As shown in the last column of Fig. 4, the fused features strike a balance between high-level and low-level visual concepts, focusing on both semantics and textures in the image. On the example in the second row of Fig. 4, the fused feature yields more accurate matches (e.g., face and computer borders). The comparison results in Table. 2 also confirm this viewpoint. We define \\(\\widetilde{\\mathcal{F}}_i\\) calculated by Eq. (8) as conditional diffusion representations of the sample \\(x_{i}\\)" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.728, + 0.407, + 0.744 + ], + "angle": 0, + "content": "3.4. Denoising Diffusion Representations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Although we construct the conditional diffusion representations from the diffusion model's generative feedback, \\(\\widetilde{\\mathcal{F}}_i\\) is still highly variable and depends on the condition generated by the local model, which can not ensure a consistent stable convergence signal. Based on Eq. (4), we explore the inherent denoising property in diffusion models: SD model has learned extensive general knowledge about identical semantics via large pre-training, reflected in the UNet denoising process. Therefore, we aim to extract the denoising diffusion representations contained in a pre-trained SD model, to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.275 + ], + "angle": 0, + "content": "align the sample's embedding with the class semantics uniformly in the feature space. Given the image pair \\(\\{x_i, y_i\\}\\) of \\(k\\)-th client's dataset \\(\\mathcal{D}_k\\), we first perform a forward process to add \\(t\\) time-steps random Gaussian noise on \\(x_i\\) as \\(\\tilde{x}_i^t = \\sqrt{\\bar{\\alpha}_t} x_i + \\sqrt{1 - \\bar{\\alpha}_t} \\epsilon_t\\), similar to [46], where \\(\\bar{\\alpha}_t\\) and \\(\\epsilon\\) are the same as defined in the Eq. (3). Next, we set the text prompt of \\(\\tilde{x}_i^t\\) as \\(\\mathcal{P}_{y_i}\\) = 'a photo of a \\(S_{y_i}\\)' used in the SD model pre-training [46], to avoid destroying correspondence between visual and linguistic concepts, and the \\(S_{y_i}\\) denotes the class name of the label \\(y_i\\). Formally, the denoising diffusion representations \\(\\tilde{\\mathcal{H}}_i\\) of the sample \\(x_i\\) are calculated by the \\(\\epsilon_\\theta\\) in Eq. (4), noisy sample \\(\\tilde{x}_i^t\\), text \\(\\mathcal{P}_{y_i}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.285, + 0.81, + 0.304 + ], + "angle": 0, + "content": "\\[\n\\mathcal {H} _ {i} = \\epsilon_ {\\theta} (\\tilde {x} _ {i} ^ {t}, \\tau_ {i}, T = t), \\tau_ {i} \\gets \\varphi (\\mathcal {P} _ {y _ {i}}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.3, + 0.905, + 0.326 + ], + "angle": 0, + "content": "\\[\n\\widetilde {\\mathcal {H}} _ {i} = \\operatorname {C o n c a t} \\left(\\widetilde {\\mathcal {H}} _ {i (L)}\\right), \\quad \\widetilde {\\mathcal {H}} _ {i (L)} = \\operatorname {P C A} \\left(\\mathcal {H} _ {i (L)}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.335, + 0.907, + 0.459 + ], + "angle": 0, + "content": "where \\(\\varphi\\) denotes the text encoder of the SD model and \\(\\tau_{i}\\) is the text features of the text prompt \\(\\mathcal{P}_{y_i}\\). We further analyze the effects of the time-steps \\(t\\) in Fig. 6. Then, we combine the features of layers \\(L\\in \\{2,3,4\\}\\), like Eq. (8), to obtain fused features \\(\\widetilde{\\mathcal{H}}_i\\). Intuitively, \\(\\widetilde{\\mathcal{H}}_i\\) is considered as a virtual teacher to calibrate the embedding of \\(x_{i}\\) at the feature-level. We define the denoising features \\(\\widetilde{\\mathcal{H}}_i\\) calculated by Eq. (9) as denoising diffusion representations of the sample \\(x_{i}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.472, + 0.651, + 0.49 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.498, + 0.907, + 0.635 + ], + "angle": 0, + "content": "Our proposed FedDifRC leverage conditional and denoising diffusion representations to obtain diverse class-relevant knowledge and consistent convergence signals by two complementary components: Text-driven Diffusion Contrastive Learning in Sec. 4.1 and Noise-driven Diffusion Consistent Restriction in Sec. 4.2. Further, we discuss how to extend FedDifRC to a self-supervised scheme (Sec. 4.3). Finally, we provide a theoretical analysis of FedDifRC under nonconvex case (Sec. 4.4). Illustration of FedDifRC in Fig. 5." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.644, + 0.816, + 0.661 + ], + "angle": 0, + "content": "4.1. Text-Driven Diffusion Contrasting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.666, + 0.907, + 0.839 + ], + "angle": 0, + "content": "Inspired by success of contrastive learning [1, 9, 26], based on Sec. 3.3, conditional diffusion representations \\(\\widetilde{\\mathcal{F}}_i\\) of sample \\(\\{x_i,y_i\\} \\in \\mathcal{D}_k\\) can be obtained with the condition \\(\\mathbf{c}_i\\) and the text prompt \\(\\mathcal{P}_{y_i} = \\text{'a photo of a } S_{y_i}\\) , where \\(S_{y_i}\\) is the class name of the label \\(y_{i}\\) . For each sample's embedding \\(z_{i} = h_{k}(x_{i})\\in \\mathbb{R}^{d}\\) , we enforce \\(z_{i}\\) to be similar to respective positive pair \\(\\widetilde{\\mathcal{F}}_i\\) and dissimilar to negative sample pairs \\(\\widetilde{\\mathcal{F}}_{\\mathcal{N}_i}\\) where \\(\\widetilde{\\mathcal{F}}_{\\mathcal{N}_i}\\) means the conditional diffusion representations with the same condition \\(\\mathbf{c}_i\\) and the unmatched text prompts \\(\\mathcal{P}_j = \\text{'a photo of a } S_j\\) (\\(S_{j}\\in S\\) and \\(S_{j}\\neq S_{y_{i}}\\)). Then, we define the similarity between \\(z_{i}\\) and \\(\\widetilde{\\mathcal{F}}_i\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.849, + 0.907, + 0.902 + ], + "angle": 0, + "content": "\\[\ns \\left(z _ {i}, \\widetilde {\\mathcal {F}} _ {i}\\right) = \\frac {z _ {i} \\cdot \\widetilde {\\mathcal {F}} _ {i}}{\\mathcal {U} \\left(\\left\\| z _ {i} \\right\\| _ {2} \\times \\left\\| \\widetilde {\\mathcal {F}} _ {i} \\right\\| _ {2}\\right)}, \\mathcal {U} = \\frac {1}{n _ {k}} \\sum_ {m = 1} ^ {n _ {k}} \\| z _ {m} - \\widetilde {\\mathcal {F}} _ {i} \\| _ {2}, \\tag {10}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "3730" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.184, + 0.092, + 0.393, + 0.104 + ], + "angle": 0, + "content": "Training Process for Local Client" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.105, + 0.481, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.293, + 0.483, + 0.364 + ], + "angle": 0, + "content": "Figure 5. Illustration of FedDifRC, including two complementary modules: TDCL (Sec. 4.1) builds an inter-contrastive learning strategy by conditional diffusion representations (Sec. 3.3), NDCR (Sec. 4.2) designs a consistent regularization penalty via denoising diffusion representations (Sec. 3.4), jointly promoting FL training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.386, + 0.484, + 0.462 + ], + "angle": 0, + "content": "where \\(n_k\\) is total number of samples at client \\(k\\), \\(\\mathcal{U}\\) is a normalization factor for similarity, where we set \\(\\mathcal{U}\\) to denote the average distance between \\(z_m|_{m=1}^{n_k}\\) and \\(\\widetilde{\\mathcal{F}}_i\\) on dataset \\(\\mathcal{D}_k\\). Next, we introduce TDCL to quantify and regulate the local model training, which can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.467, + 0.483, + 0.51 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {T D C L} = \\log (1 + \\frac {\\sum_ {\\widetilde {\\mathcal {F}} _ {j} \\in \\widetilde {\\mathcal {F}} _ {\\mathcal {N} _ {i}}} \\exp \\left(s \\left(z _ {i} , \\widetilde {\\mathcal {F}} _ {j}\\right) / \\tau\\right)}{\\exp \\left(s \\left(z _ {i} , \\widetilde {\\mathcal {F}} _ {i}\\right) / \\tau\\right)}), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.515, + 0.484, + 0.576 + ], + "angle": 0, + "content": "where \\(\\tau\\) denotes a temperature parameter to control representation strength [1]. By minimizing Eq. (11), for client \\(k\\), the local model brings each sample's embedding \\(z_{i}\\) closer to the positive pair and away from other negative pairs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.582, + 0.421, + 0.598 + ], + "angle": 0, + "content": "4.2. Noise-Driven Diffusion Regularization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.604, + 0.484, + 0.786 + ], + "angle": 0, + "content": "Despite each local model aligning sample's embedding with conditional diffusion representations by optimizing \\(\\mathcal{L}_{TDCL}\\) in Eq. (11). Since conditional diffusion representations are dynamically built from text-driven cross-modal matching and depend on generated condition \\(\\mathbf{c}_i\\) at each round, which could not offer a stable convergence direction. Therefore, we aim to leverage denoising diffusion representations \\((\\widetilde{\\mathcal{H}}_i)\\) in Eq. (9)) to devise a consistent regularization penalty, constraining model's optimization region in the feature space. We design Noise-driven Diffusion Consistent Restriction (NDCR) and utilize a regularization term to pull the sample's embedding \\(z_{i}\\) closer to the respective \\(\\widetilde{\\mathcal{H}}_i\\):" + }, + { + "type": "equation", + "bbox": [ + 0.179, + 0.793, + 0.483, + 0.834 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {N D C R} = \\sum_ {q = 1} ^ {d} \\left(z _ {i (q)} - \\widetilde {\\mathcal {H}} _ {i (q)}\\right) ^ {2}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(q\\) indexes the dimension \\(d\\) of the embedding vector. Based on Eq. (12), we expect to align the sample's embedding with the corresponding denoising diffusion representations, enforcing semantics uniformity of the feature space." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.906, + 0.153 + ], + "angle": 0, + "content": "Besides, we construct CrossEntropy loss and use the logits output vector \\( l_{i} = g(z_{i}) = f(x_{i}) \\) with the original label \\( y_{i} \\) to optimize the local discriminative ability for each client. Formally, given the sample pair \\( \\{x_{i},y_{i}\\} \\) to calculate \\( \\mathcal{L}_{CE} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.612, + 0.16, + 0.907, + 0.178 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C E} = - \\mathbf {1} _ {y _ {i}} \\log (\\psi (f (x _ {i}))), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.184, + 0.907, + 0.245 + ], + "angle": 0, + "content": "where \\(\\psi\\) denotes the Softmax function, \\(f\\) is the local model with parameters \\(\\{u,v\\}\\) maps \\(x_{i}\\) into a \\(|\\mathcal{C}|\\)-dim output logits \\(l_{i} = f(x_{i}) = g(h(x_{i}))\\). Finally, for the \\(k\\)-th client, we carry out the following optimization objective in local training:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.253, + 0.907, + 0.27 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {T D C L} (\\mathrm {E q .} (1 1)) + \\mathcal {L} _ {N D C R} (\\mathrm {E q .} (1 2)) + \\mathcal {L} _ {C E}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.276, + 0.871, + 0.292 + ], + "angle": 0, + "content": "The algorithm of FedDifRC is given in Appendix A." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.299, + 0.842, + 0.315 + ], + "angle": 0, + "content": "4.3. Expanding to Self-Supervised Scheme" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.321, + 0.907, + 0.553 + ], + "angle": 0, + "content": "FedDifRC can be extended to the self-supervised scheme without relying on any labeled data source. Specifically, for \\(\\mathcal{L}_{TDCL}\\) in Eq. (11), we only need to modify text template slightly: i) we set the text prompt \\(\\mathcal{P}_{y_i} = \\text{'a photo of a similar object'}\\) in Eq. (7) to generate \\(\\widetilde{\\mathcal{F}}_i\\) as the positive pair; ii) then we randomly select from the set of all class names in TinyImageNet [21] to get \\(S_{\\mathcal{N}_i}\\), and set \\(\\mathcal{P}_{y_i} = \\text{'a photo of a } S_{\\mathcal{N}_i}\\) to obtain \\(\\widetilde{\\mathcal{F}}_{\\mathcal{N}_i}\\) as negative pairs. For the \\(\\mathcal{L}_{NDCR}\\) in Eq. (12), we set \\(\\mathcal{P}_{y_i} = \\text{'a photo of a visual object'}\\) in Eq. (9) to get \\(\\widetilde{\\mathcal{H}}_i\\) as consistent signal to align the sample's embedding. Note that it is also possible to use class names of other datasets (e.g., CIFAR100 [20]) to generate \\(\\widetilde{\\mathcal{F}}_{\\mathcal{N}_i}\\) as negative pairs. By this self-supervised way, the objective term of local training becomes \\(\\mathcal{L}_{TDCL} + \\mathcal{L}_{NDCR}\\). The results in Table. 5 validate the effectiveness of FedDifRC's self-supervised scheme." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.561, + 0.72, + 0.577 + ], + "angle": 0, + "content": "4.4. Convergence Analysis" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.583, + 0.906, + 0.627 + ], + "angle": 0, + "content": "We proved convergence of FedDifRC in non-convex case. We denote \\(\\mathcal{L}\\) (Eq. (14)) as \\(\\mathcal{L}_r\\) with the round \\(r\\), assumptions (Appendix B.2) are similar to existing works [6, 29, 30, 69]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.628, + 0.906, + 0.658 + ], + "angle": 0, + "content": "Convergence Rate of FedDifRC: Let the round \\( r \\) from 0 to \\( R - 1 \\), given any \\( \\xi > 0 \\), the \\( \\mathcal{L} \\) will converge when," + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.665, + 0.905, + 0.699 + ], + "angle": 0, + "content": "\\[\nR > \\frac {2 \\left(\\mathcal {L} _ {0} - \\mathcal {L} ^ {*}\\right)}{\\xi E \\eta \\left(2 - L _ {1} \\eta\\right) - \\Omega_ {1} - \\Omega_ {2}}, \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.704, + 0.862, + 0.719 + ], + "angle": 0, + "content": "and we can further get the following condition for \\(\\eta\\)," + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.724, + 0.905, + 0.758 + ], + "angle": 0, + "content": "\\[\n\\eta < \\frac {2 \\xi - 2 (| \\mathcal {C} | - 1) L _ {2} B}{L _ {1} (\\xi + \\sigma^ {2})}, \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.764, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\(\\Omega_{1} = 2(|\\mathcal{C}| - 1)L_{2}E\\eta B\\), \\(\\Omega_{2} = L_{1}E\\eta^{2}\\sigma^{2}\\), and \\(\\mathcal{L}^*\\) denotes the optimal solution of \\(\\mathcal{L}\\), \\(\\{L_1,L_2,B,\\sigma^2\\}\\) are constants in the assumptions from Appendix B.2, and \\(|\\mathcal{C}|\\) as the number of classes, \\(E\\) as local epochs. Eq. (15) and Eq. (16) ensure convergence rate of FedDifRC, after selecting the number of rounds \\(R\\) and the learning rate \\(\\eta\\). So the smaller \\(\\xi\\) is, the larger \\(R\\) is, which means that the tighter the bound is, the more rounds \\(R\\) is required. The detailed assumptions and proofs are formally provided in Appendix B." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3731" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.483, + 0.244 + ], + "angle": 0, + "content": "
TDCLNDCRCIFAR10 (keep ρ = 1.0)
NID10.05NID10.2NID10.5NID2AVG
78.2784.6586.1172.6080.41
✓81.3986.0388.1675.6782.81
✓80.3586.4087.5475.3382.40
✓✓83.1488.2789.3176.4584.29
TDCLNDCRCIFAR10-LT (keep NID10.2)
ρ = 10ρ = 50ρ = 100ρ = 200AVG
75.9463.2552.9749.6960.46
✓76.6264.6653.6250.3361.31
✓76.1964.7554.3950.6761.50
✓✓77.5865.3055.0151.1262.25
" + }, + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.254, + 0.48, + 0.267 + ], + "angle": 0, + "content": "Table 1. Ablation study for two key modules: TDCL and NDCR." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.285, + 0.483, + 0.355 + ], + "angle": 0, + "content": "
ScenariosL=1L=2L=3L=4Fused△
NID10.0580.2182.6082.1981.3383.14+0.54
NID10.286.1587.2887.8185.9288.27+0.46
NID10.587.1089.0388.9387.1789.31+0.28
NID274.2675.7375.6175.0076.45+0.72
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.365, + 0.483, + 0.409 + ], + "angle": 0, + "content": "Table 2. Comparison of features from different layers \\( L \\) and fused features \\( \\left( {L \\in \\{ 2,3,4\\} }\\right) \\) in Eq. (8) and Eq. (9) to obtain \\( {\\widetilde{\\mathcal{F}}}_{i} \\) and \\( {\\widetilde{\\mathcal{H}}}_{i} \\) , best in bold and second with underline, \\( \\bigtriangleup \\) means accuracy gap." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.424, + 0.291, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.424, + 0.483, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.553, + 0.483, + 0.583 + ], + "angle": 0, + "content": "Figure 6. Analysis of FedDifRC with time-steps \\( t \\) in Eq. (9) (left) and temperature \\( \\tau \\) in Eq. (11) (right) on CIFAR10 with NID10.2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.61, + 0.224, + 0.627 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.636, + 0.282, + 0.654 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.735 + ], + "angle": 0, + "content": "Datasets: We adopt popular benchmark datasets, including CIFAR10 & CIFAR100 [20], TinyImageNet [21]; we build unbalanced versions of these datasets by [15]: CIFAR10-LT & CIFAR100-LT, and TinyImageNet-LT; we use Digits [41] and Office-Caltech [19] for the domain shift task in FL." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Scenarios: i) label shift: we consider two non-iid settings: \\(\\mathrm{NID1}_{\\alpha}\\) follows Dirichlet distribution [59], where \\(\\alpha\\) denotes the heterogeneity level; NID2 is a more extreme setting consists of 6 biased clients (each has only a single category) and 1 unbiased client has all categories. ii) imbalance shift: we shape the original dataset into a long-tailed distribution with \\(\\rho\\) follow [15], \\(\\rho\\) means the ratio between sample sizes of the most frequent and least frequent class, as \\(\\rho = \\max_{j} / \\min_{j}\\). iii) domain shift: Digits [41] includes 4 domains as MNIST (M), USPS (U), SVHN (SV), SYN (SY) with 10 categories. Office-Caltech [19] also consists 4 domains as Caltech (C)," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.167 + ], + "angle": 0, + "content": "Webcam (W), Amazon (A), DSLR (D) with 10 overlapping categories. We initialize 10 clients and assign them domains randomly: Digits (M: 1, U: 4, SV: 3, SY: 2), Office-Caltech (C: 3, W: 2, A: 2, D: 3). Each client's dataset is randomly selected from domains (Digits: \\(10\\%\\), Office-Caltech: \\(20\\%\\))." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.167, + 0.905, + 0.287 + ], + "angle": 0, + "content": "Configurations: The global rounds \\( R \\) as 100 and clients \\( K \\) as 10, all methods have little or no gain with more rounds \\( R \\). To facilitate fair comparison, we follow in [26, 29], and use SGD optimizer to update models with a learning rate 0.01, a momentum of 0.9, a batch size of 64, the weight decay of \\( 10^{-5} \\), and local epochs \\( E = 10 \\). We use MobileNetV2 [48] for TinyImageNet, ResNet-10 [14] for others. The dimension \\( d \\) as 512, \\( \\tau \\) (Eq. (11)) as 0.06, \\( t \\) (Eq. (9)) as 150." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.905, + 0.364 + ], + "angle": 0, + "content": "Baselines: We compare our against several SOTA methods: standard-based FedAvg [34], FedProx [29]; contrast-based MOON [26], FedRCL [50]; aggregate-based FedCDA [60], FedDisco [69]; unsupervised FedEMA [77], FedLID [42]; prototype-based FedProto [52], FedNH [70]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.375, + 0.699, + 0.391 + ], + "angle": 0, + "content": "5.2. Validation Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.397, + 0.905, + 0.519 + ], + "angle": 0, + "content": "We give ablation study in Table. 1: i) TDCL brings significant performance improvements over the baseline, showing that TDCL is able to promote semantic spread-out property; ii) NDCR also yields solid gains, proving the importance of aligning instance embedding via consistency regularization; iii) combining TDCL and NDCR attain better performance, supporting our motivation to unleash meaningful guidance of diffusion representations to promote FL local training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.906, + 0.763 + ], + "angle": 0, + "content": "Then, we prove the advantage of fused features (Eq. (8)). As shown in Table. 2, results on CIFAR10 reveal that fused features capture both semantics and local details, better than features on a single layer \\( L \\), and if the features with too low resolution (\\( L = 1 \\)) or over-matching (\\( L = 4 \\)) would damage the model's performance. To explore the influence of different time-steps \\( t \\) (Eq. (9)), we conduct experiments on CIFAR10 with NID10.2 in Fig. 6 left, we add different amount of noise to get \\( \\tilde{x}_i^t \\) in Eq. (9). We find that FedDifRC is robust to the choice of \\( t \\in [100, 300] \\) (\\( t = 150 \\) achieves the highest value). As shown in Fig. 6 right, we present results with different \\( \\tau \\) in Eq. (11), which reveals that FedDifRC is not sensitive to the \\( \\tau \\in [0.04, 0.08] \\), indicating the effectiveness of \\( \\mathcal{U} \\) (Eq. (10)), to balance relative distances among different samples. The accuracy no longer increases after \\( \\tau = 0.06 \\). Extra results of \\( t \\) and \\( \\tau \\) are provided in Appendix C.2.5." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.772, + 0.695, + 0.787 + ], + "angle": 0, + "content": "5.3. Evaluation Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.905, + 0.901 + ], + "angle": 0, + "content": "The results on different settings (NID1, NID2, \\(\\rho\\)) in Table. 3, FedDifRC consistently outperforms these counterparts, it confirms that FedDifRC effectively leverages the inherent fine-grained visual representation capabilities of diffusion models to promote heterogeneous FL training. The results of Office-Caltech with domain shift are shown in Table. 4, indicating that FedDifRC can acquire well-generalizable" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3732" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.089, + 0.905, + 0.243 + ], + "angle": 0, + "content": "
FL BaselinesCIFAR10CIFAR100TinyImageNetCIFAR10-LT (keep NID10.2)
NID10.05NID10.2NID2NID10.05NID10.2NID2NID10.05NID10.2NID2ρ = 10ρ = 50ρ = 100ρ = 200
FedAvg[AISTAT'17]78.2784.6572.6055.9760.0850.5640.4142.8435.1575.9463.2552.9749.69
FedProx[MLSys'20]78.4284.5972.8156.2760.2150.2940.2042.1635.6276.0263.4352.8649.10
MOON[CVPR'21]80.7986.1073.3556.7961.4851.8140.7943.6336.1175.7363.8053.1749.31
FedProto[AAAI'22]77.8683.9070.8454.4859.2648.6039.2941.0234.6874.3162.2350.3747.26
FedNH[AAAI'23]80.2585.8273.6556.6861.4551.3140.2543.3036.2176.1563.3652.4948.92
FedDisco[ICML'23]81.3586.6374.7257.3461.7952.0540.8343.9237.6976.2764.4053.3349.75
FedCDA[ICLR'24]81.7086.8975.0757.7662.1752.3541.1844.0938.2476.4564.3053.8550.08
FedRCL[CVPR'24]82.0287.1175.1357.9862.0552.4941.3344.2138.8676.3464.1753.4149.91
FedDifRC83.1488.2776.4559.2263.3253.8742.6345.2840.0577.5865.3055.0151.12
" + }, + { + "type": "table_caption", + "bbox": [ + 0.095, + 0.253, + 0.9, + 0.268 + ], + "angle": 0, + "content": "Table 3. Comparison results on several settings (label shift and imbalance shift) and datasets, the best in bold and second with underline." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.292, + 0.482, + 0.389 + ], + "angle": 0, + "content": "
FL BaselinesOffice-Caltech
CaltechWebcamAmazonDSLRAVG
FedDisco[ICML'23]62.3945.9577.3145.2057.71
FedCDA[ICLR'24]62.2546.5376.5846.1457.87
FedRCL[CVPR'24]60.1342.7675.6043.7755.56
FedDifRC63.3250.1978.6748.1260.08
" + }, + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.4, + 0.48, + 0.414 + ], + "angle": 0, + "content": "Table 4. Comparison results on Office-Caltech with domain shift." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.432, + 0.482, + 0.515 + ], + "angle": 0, + "content": "
FL BaselinesClientsCIFAR100
NID10.05NID10.2NID10.5NID2
FedEMA[ICLR'22]K=1026.9029.4130.6421.52
FedLID[ICCV'23]K=1028.8430.4331.2223.65
FedDifRCK=1030.6131.4532.3025.73
" + }, + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.527, + 0.48, + 0.541 + ], + "angle": 0, + "content": "Table 5. Comparison results on self-supervised setting (Sec. 4.3)." + }, + { + "type": "table", + "bbox": [ + 0.096, + 0.559, + 0.481, + 0.624 + ], + "angle": 0, + "content": "
FL BaselinesCIFAR10CIFAR10-LT (NID102)
NID10.05NID10.2NID2ρ = 10ρ = 50ρ = 100
FedDDA[PRCV-23]82.0586.9474.4976.3163.9553.15
FedDISC[AAAF24]82.3387.1575.4276.6064.2354.07
FedRCL (♂/†)82.30♂/80.95†87.25♂/86.16†75.37♂/73.85†76.50♂/76.12†64.26♂/62.91†53.65♂/52.20†
FedRCL (♀ + †)81.5286.4874.3576.2263.6952.93
FedDifRC83.1488.2776.4577.5865.3055.01
" + }, + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.637, + 0.481, + 0.65 + ], + "angle": 0, + "content": "Table 6. More comparison results on CIFAR10 and CIFAR10-LT." + }, + { + "type": "table", + "bbox": [ + 0.096, + 0.668, + 0.481, + 0.754 + ], + "angle": 0, + "content": "
FL BaselinesHAM10000 [56]EuroSAT [40]
NID10.2NID10.5NID2NID10.2NID10.5NID2
FedDisco[ICML'23]50.2653.6844.6378.7682.3670.34
FedCDA[ICLR'24]51.6554.0545.1680.4583.3871.10
FedRCL[CVPR'24]52.2154.2045.4879.8083.6970.73
FedDifRC53.7955.6047.1782.4986.0573.14
" + }, + { + "type": "table_caption", + "bbox": [ + 0.102, + 0.766, + 0.472, + 0.779 + ], + "angle": 0, + "content": "Table 7. Comparison results on the HAM10000 and EuroSAT." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.901 + ], + "angle": 0, + "content": "property in the feature space. With self-supervised scenario, in Table. 5, our FedDifRC achieves excellent performance by leveraging the visual perception capability of diffusion models, without requiring any labeled data. Please refer to Appendix C for more details and results (e.g., domain shift, convergence rates, different SD versions, etc)." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.292, + 0.905, + 0.373 + ], + "angle": 0, + "content": "
Pre-trained ModelsCIFAR10CIFAR10-LT (NID10.2)
NID10.05NID10.2NID2ρ = 10ρ = 50ρ = 100
CLIP-ViT-B/16 [45]79.7284.9173.9774.8962.4450.87
MAE-ViT-B/16 [10]80.7485.3873.2475.4563.2752.55
DINO-ViT-B/16 [33]80.5786.1474.1675.2562.7051.92
FedDifRC83.1488.2776.4577.5865.3055.01
" + }, + { + "type": "table_caption", + "bbox": [ + 0.517, + 0.384, + 0.901, + 0.398 + ], + "angle": 0, + "content": "Table 8. Experiments on aligning features of pre-trained models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.425, + 0.905, + 0.454 + ], + "angle": 0, + "content": "To further elaborate on the core idea of our FedDifRC, we then perform the following experiments:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.455, + 0.907, + 0.756 + ], + "angle": 0, + "content": "i) To ensure a fair comparison, we also compare diffusion-based FL methods FedDISC [67] and FedDDA [74], we generate data for each client (same amount) by the SD model as \\(\\diamond\\), and align SD features (the image is fed directly into the SD model) by L2 loss as \\(\\dagger\\). Table. 6 shows that directly matching SD features may even hurt the performance, which further illustrates the superiority of our FedDifRC. ii) We focus on visual perception ability of diffusion models to understand an image (e.g., textures and structures) rather than pre-trained prior knowledge. We conduct experiments on medical and remote sensing datasets HAM10000 [56] and EuroSAT [40], which are not crossed in SD pre-training dataset LAION-5B [49]. Table. 7 show that the success of FedDifRC is based on the fine-grained visual representation rather than the large prior knowledge in SD model. iii) The SD model can implicitly provide both high-level and low-level visual concepts of semantic objects, while foundation models focus more on sparse visual matching. The Table. 8 show that FedDifRC can capture underlying properties and correspondences of visual objects." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.77, + 0.634, + 0.785 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.901 + ], + "angle": 0, + "content": "In this paper, we aim to exploit the fine-grained visual representation capabilities of diffusion models to handle data heterogeneity in FL. We propose a novel diffusion-inspired FL framework FedDifRC to regulate local training via denoising and conditional diffusion representations. By leveraging complementary strength of TDCL and NDCR, FedDifRC achieves superior performance in heterogeneous FL." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3733" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.192 + ], + "angle": 0, + "content": "This work is supported by the scholarship from the China Scholarship Council (CSC) while the first author pursues his PhD degree at the University of Wollongong. This work was also partially supported by Australian Research Council Linkage Project LP210300009 and LP230100083." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.206, + 0.188, + 0.221 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.23, + 0.484, + 0.286 + ], + "angle": 0, + "content": "[1] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, pages 1597-1607, 2020. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.288, + 0.484, + 0.328 + ], + "angle": 0, + "content": "[2] Ching-Yao Chuang, Joshua Robinson, Yen-Chen Lin, Antonio Torralba, et al. Debiased contrastive learning. In NeurIPS, pages 8765-8775, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.33, + 0.483, + 0.357 + ], + "angle": 0, + "content": "[3] Kevin Clark and Priyank Jain. Text-to-image diffusion models are zero shot classifiers. In NeurIPS, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.36, + 0.482, + 0.4 + ], + "angle": 0, + "content": "[4] Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah. Diffusion models in vision: A survey. IEEE TPAMI, 45(9):10850-10869, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.403, + 0.482, + 0.441 + ], + "angle": 0, + "content": "[5] Jiequan Cui, Zhisheng Zhong, Shu Liu, Bei Yu, et al. Parametric contrastive learning. In ICCV, pages 715-724, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.445, + 0.482, + 0.499 + ], + "angle": 0, + "content": "[6] Liang Gao, Huazhu Fu, Li Li, Yingwen Chen, et al. Feddc: Federated learning with non-iid data via local drift decoupling and correction. In CVPR, pages 10112-10121, 2022, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.502, + 0.482, + 0.543 + ], + "angle": 0, + "content": "[7] Jie Gui, Tuo Chen, Jing Zhang, Qiong Cao, et al. A survey on self-supervised learning: Algorithms, applications, and future trends. IEEE TPAMI, 46(12):9052-9071, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.545, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[8] Sungwon Han, Sungwon Park, Fangzhao Wu, Sundong Kim, et al. Fedx: Unsupervised federated learning with cross knowledge distillation. In ECCV, pages 691-707, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.589, + 0.482, + 0.629 + ], + "angle": 0, + "content": "[9] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, et al. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9729-9738, 2020. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.631, + 0.482, + 0.672 + ], + "angle": 0, + "content": "[10] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, et al. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.674, + 0.482, + 0.714 + ], + "angle": 0, + "content": "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, pages 6840-6851, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.717, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[12] Wenke Huang, Mang Ye, and Bo Du. Learn from others and be yourself in heterogeneous federated learning. In CVPR, pages 10143-10153, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.76, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[13] Wenke Huang, Mang Ye, Zekun Shi, He Li, and Bo Du. Rethinking federated learning with domain shift: A prototype view. In CVPR, pages 16312-16322, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[14] Cheng Ju, Aurélien Bibaut, and Mark van der Laan. The relative performance of ensemble methods with deep convolutional neural networks for image classification. Journal of Applied Statistics, 45:2800-2818, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[15] Cao Kaidi, Colin Wei, Adrien Gaidon, Nikos Arechiga, et al. Learning imbalanced datasets with label-distribution-aware margin loss. In NeurIPS, 2019. 7" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.23, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[16] Peter Kairouz, H Brendan McMahan, Brendan Avent, Aurélien Bellet, et al. Advances and open problems in federated learning. Foundations and Trends in Machine Learning, 14(1):1-210, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[17] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank Reddi, et al. Scaffold: Stochastic controlled averaging for federated learning. In ICML, pages 5132-5143, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.206, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[18] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, et al. Supervised contrastive learning. In NeurIPS, pages 18661-18673, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.904, + 0.289 + ], + "angle": 0, + "content": "[19] Boqing Kong, Yuan Shi, Fei Sha, and Kristen Grauman. Geodesic flow kernel for unsupervised domain adaptation. In CVPR, pages 2066-2073, 2012. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[20] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. In Technical report, University of Toronto, 2009. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.904, + 0.36 + ], + "angle": 0, + "content": "[21] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7, 2015. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.416 + ], + "angle": 0, + "content": "[22] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[23] Sunwoo Lee, Tuo Zhang, and A Salman Avestimehr. Layerwise adaptive model aggregation for scalable federated learning. In AAAI, pages 8491-8499, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[24] Alexander C Li, Mihir Prabhudesai, Shivam Duggal, Ellis Brown, et al. Your diffusion model is secretly a zero-shot classifier. In ICCV, pages 2206-2217, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.905, + 0.558 + ], + "angle": 0, + "content": "[25] Daixun Li, Weiying Xie, Zixuan Wang, Yibing Lu, et al. Feddiff: Diffusion model driven federated learning for multimodal and multi-clients. IEEE TCSVT, 34(10):10353-10367, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.905, + 0.601 + ], + "angle": 0, + "content": "[26] Qinbin Li, Bingsheng He, and Dawn Song. Model-contrastive federated learning. In CVPR, pages 10713-10722, 2021. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[27] Qinbin Li, Yiqun Diao, Quan Chen, and Bingsheng He. Federated learning on non-iid data silos: An experimental study. In ICDE, pages 965–978, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[28] Tian Li, Anit Kumar Sahu, Ameet Talwalkar, and Virginia Smith. Federated learning: Challenges, methods, and future directions. IEEE SPM, 37(3):50-60, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[29] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, et al. Federated optimization in heterogeneous networks. MLSys, 2:429-450, 2020. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.733, + 0.905, + 0.772 + ], + "angle": 0, + "content": "[30] Xiang Li, Kaixuan Huang, Wenhao Yang, Shusen Wang, et al. On the convergence of fedavg on non-iid data. In ICLR, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[31] Grace Luo, Lisa Dunlap, Dong Huk Park, Aleksander Holynski, et al. Diffusion hyperfeatures: Searching through time and space for semantic correspondence. In NeurIPS, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[32] Mi Luo, Fei Chen, Dapeng Hu, Yifan Zhang, et al. No fear of heterogeneity: Classifier calibration for federated learning with non-iid data. In NeurIPS, pages 5972-5984, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[33] Caron Mathilde, Hugo Touvron, Ishan Misra, Hervé Jégou, et al. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3734" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.147 + ], + "angle": 0, + "content": "[34] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, et al. Communication-efficient learning of deep networks from decentralized data. In AISTATS, pages 1273-1282, 2017. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.203 + ], + "angle": 0, + "content": "[35] Matias Mendieta, Taojiannan Yang, Pu Wang, Minwoo Lee, et al. Local learning matters: Rethinking data heterogeneity in federated learning. In CVPR, pages 8397-8406, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.483, + 0.258 + ], + "angle": 0, + "content": "[36] John Nguyen, Kshitiz Malik, Hongyuan Zhan, Ashkan Yousefpour, et al. Federated learning with buffered asynchronous aggregation. In AISTATS, pages 3581-3607, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.26, + 0.483, + 0.3 + ], + "angle": 0, + "content": "[37] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, pages 8162-8171, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.483, + 0.342 + ], + "angle": 0, + "content": "[38] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. In NeurIPS, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.483, + 0.385 + ], + "angle": 0, + "content": "[39] Reese Pathak and Martin J Wainwright. Fedsplit: An algorithmic framework for fast federated optimization. In NeurIPS, pages 7057-7066, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.386, + 0.483, + 0.44 + ], + "angle": 0, + "content": "[40] Helber Patrick, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE J-STARS, 12(7):2217-2226, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.483, + 0.482 + ], + "angle": 0, + "content": "[41] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, et al. Moment matching for multi-source domain adaptation. In ICCV, pages 1406-1415, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.483, + 0.537 + ], + "angle": 0, + "content": "[42] Athanasios Psaltis, Anestis Kastellos, Charalampos Z Patrikakis, and Petros Daras. Fedlid: Self-supervised federated learning for leveraging limited image data. In ICCV, pages 1039-1048, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.483, + 0.593 + ], + "angle": 0, + "content": "[43] Pian Qi, Diletta Chiaro, Antonella Guzzo, Michele Ianni, et al. Model aggregation techniques in federated learning: A comprehensive survey. Future Generation Computer Systems, 150:272-293, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.483, + 0.648 + ], + "angle": 0, + "content": "[44] Liangqiong Qu, Yuyin Zhou, Paul Pu Liang, Yingda Xia, et al. Rethinking architecture design for tackling data heterogeneity in federated learning. In CVPR, pages 10061-10071, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.65, + 0.483, + 0.704 + ], + "angle": 0, + "content": "[45] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.483, + 0.76 + ], + "angle": 0, + "content": "[46] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, et al. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.483, + 0.803 + ], + "angle": 0, + "content": "[47] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234-241, 2015. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.805, + 0.483, + 0.845 + ], + "angle": 0, + "content": "[48] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhominov, et al. Mobilenetv2: Inverted residuals and linear bottlenecks. In CVPR, pages 4510-4520, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.847, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[49] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, pages 25278–25294, 2022. 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[50] Seonguk Seo, Jinkyu Kim, Geeho Kim, and Bohyung Han. Relaxed contrastive learning for federated learning. In CVPR, pages 12279-12288, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[51] Fiona Victoria Stanley Jothiraj and Afra Mashhadi. Phoenix: A federated generative diffusion model. In ACM WWW, pages 1568-1577, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[52] Yue Tan, Guodong Long, Lu Liu, Tianyi Zhou, et al. FedProto: Federated prototype learning across heterogeneous clients. In AAAI, pages 8432-8440, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[53] Yue Tan, Guodong Long, Jie Ma, Lu Liu, et al. Federated learning from pre-trained models: A contrastive learning approach. In NeurIPS, pages 19332-19344, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[54] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, et al. Emergent correspondence from image diffusion. In NeurIPS, pages 1363-1389, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.344 + ], + "angle": 0, + "content": "[55] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, et al. What makes for good views for contrastive learning? In NeurIPS, pages 6827-6839, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[56] Philipp Tschandl, Cliff Rosendahl, and Harald Kittler. The ham10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. Nature Scientific Data, 5(1):1-9, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[57] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9 (11):2579-2605, 2008. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.441, + 0.905, + 0.469 + ], + "angle": 0, + "content": "[58] Feng Wang and Huaping Liu. Understanding the behaviour of contrastive loss. In CVPR, pages 2495-2504, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.511 + ], + "angle": 0, + "content": "[59] Hongyi Wang, Mikhail Yurochkin, Yuekai Sun, Dimitris Papailiopoulos, et al. Federated learning with matched averaging. In ICLR, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[60] Haozhao Wang, Haoran Xu, Yichen Li, Yuan Xu, et al. Fed-cda: Federated learning with cross-rounds divergence-aware aggregation. In ICLR, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[61] Yingheng Wang, Yair Schiff, Aaron Gokaslan, Weishen Pan, et al. Infodiffusion: Representation learning using information maximizing diffusion models. In ICML, pages 36336-36354, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.905, + 0.662 + ], + "angle": 0, + "content": "[62] Yuan Wang, Huazhu Fu, Renuga Kanagavelu, Qingsong Wei, et al. An aggregation-free federated learning for tackling data heterogeneity. In CVPR, pages 26233-26242, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[63] Chen Wei, Karttikeya Mangalam, Po-Yao Huang, Yanghao Li, et al. Diffusion models as masked autoencoders. In ICCV, pages 16284-16294, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.905, + 0.748 + ], + "angle": 0, + "content": "[64] Weilai Xiang, Hongyu Yang, Di Huang, and Yunhong Wang. Denoising diffusion autoencoders are unified self-supervised learners. In ICCV, pages 15802-15812, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[65] Yikai Yan, Chaoyue Niu, Yucheng Ding, Zhenzhe Zheng, et al. Federated optimization under intermittent client availability. INFORMS Journal on Computing, 36(1):185-202, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[66] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, et al. Diffusion models: A comprehensive survey of methods and applications. ACM Computing Surveys, 56(4):1-39, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[67] Mingzhao Yang, Shangchao Su, Bin Li, and Xiangyang Xue. Exploring one-shot semi-supervised federated learning with pre-trained diffusion models. In AAAI, pages 16325-16333, 2024. 3, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3735" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "[68] Xingyi Yang and Xinchao Wang. Diffusion model as representation learner. In ICCV, pages 18938-18949, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.482, + 0.163 + ], + "angle": 0, + "content": "[69] Rui Ye, Mingkai Xu, Jianyu Wang, Chenxin Xu, et al. Feddisco: Federated learning with discrepancy-aware collaboration. In ICML, pages 39879-39902, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.482, + 0.207 + ], + "angle": 0, + "content": "[70] Dai Yutong, Zeyuan Chen, Junnan Li, Shelby Heinecke, et al. Tackling data heterogeneity in federated learning with class prototypes. In AAAI, pages 7314-7322, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[71] Daokun Zhang, Jie Yin, Xingquan Zhu, and Chengqi Zhang. Network representation learning: A survey. IEEE Transactions on Big Data, 6:3-28, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.251, + 0.482, + 0.305 + ], + "angle": 0, + "content": "[72] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, et al. A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence. In NeurIPS, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.482, + 0.348 + ], + "angle": 0, + "content": "[73] Jianqing Zhang, Yang Hua, Hao Wang, Tao Song, et al. Fedala: Adaptive local aggregation for personalized federated learning. In AAAI, pages 11237-11244, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.35, + 0.482, + 0.391 + ], + "angle": 0, + "content": "[74] Zhuang Zhao, Feng Yang, and Guirong Liang. Federated learning based on diffusion model to cope with non-iid data. In PRCV, pages 220-231, 2023. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.393, + 0.482, + 0.434 + ], + "angle": 0, + "content": "[75] Hangyu Zhu, Jinjin Xu, Shiqing Liu, and Yaochu Jin. Federated learning on non-iid data: A survey. Elsevier Neurocomputing, 465:371-390, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.436, + 0.482, + 0.476 + ], + "angle": 0, + "content": "[76] Fuzhen Zhuang, Xiaohu Cheng, Ping Luo, Sinno Jialin Pan, et al. Supervised representation learning: Transfer learning with deep autoencoders. In IJCAI, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.478, + 0.482, + 0.518 + ], + "angle": 0, + "content": "[77] Weiming Zhuang, Yonggang Wen, and Shuai Zhang. Divergence-aware federated self-supervised learning. In ICLR, 2022. 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.518 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3736" + } + ] +] \ No newline at end of file diff --git a/2025/FedDifRC_ Unlocking the Potential of Text-to-Image Diffusion Models in Heterogeneous Federated Learning/d48a8fed-132a-4154-9395-1fe6c487ac0e_origin.pdf b/2025/FedDifRC_ Unlocking the Potential of Text-to-Image Diffusion Models in Heterogeneous Federated Learning/d48a8fed-132a-4154-9395-1fe6c487ac0e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8c1ad47050d339a12e044a533644c79fa96732a1 --- /dev/null +++ b/2025/FedDifRC_ Unlocking the Potential of Text-to-Image Diffusion Models in Heterogeneous Federated Learning/d48a8fed-132a-4154-9395-1fe6c487ac0e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1b063f3af7d4b031f7b0ba90ba2739f3d8df0ca12a64ebf7d5714b93b2c06888 +size 16348384 diff --git a/2025/FedDifRC_ Unlocking the Potential of Text-to-Image Diffusion Models in Heterogeneous Federated Learning/full.md b/2025/FedDifRC_ Unlocking the Potential of Text-to-Image Diffusion Models in Heterogeneous Federated Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..64e2cc1ac3276d41efbac4d61aee858679bad60e --- /dev/null +++ b/2025/FedDifRC_ Unlocking the Potential of Text-to-Image Diffusion Models in Heterogeneous Federated Learning/full.md @@ -0,0 +1,471 @@ +# FedDifRC: Unlocking the Potential of Text-to-Image Diffusion Models in Heterogeneous Federated Learning + +Huan Wang $^{1,\ast}$ , Haoran Li $^{1,3,\ast}$ , Huaming Chen $^{2}$ , Jun Yan $^{1}$ , Jiahua Shi $^{3}$ , Jun Shen $^{1,\dagger}$ $^{1}$ School of Computing and Information Technology, University of Wollongong, Wollongong, Australia + $^{2}$ School of Electrical and Computer Engineering, The University of Sydney, Sydney, Australia + $^{3}$ QLD Alliance for Agriculture and Food Innovation, The University of Queensland, Brisbane, Australia +hw226@uowmail.edu.au *Equal Contribution †Corresponding Author + +# Abstract + +Federated learning aims at training models collaboratively across participants while protecting privacy. However, one major challenge for this paradigm is the data heterogeneity issue, where biased data preferences across multiple clients, harming the model's convergence and performance. In this paper, we first introduce powerful diffusion models into the federated learning paradigm and show that diffusion representations are effective steers during federated training. To explore the possibility of using diffusion representations in handling data heterogeneity, we propose a novel diffusion-inspired Federated paradigm with Diffusion Representation Collaboration, termed FedDifRC, leveraging meaningful guidance of diffusion models to mitigate data heterogeneity. The key idea is to construct text-driven diffusion contrasting and noise-driven diffusion regularization, aiming to provide abundant class-related semantic information and consistent convergence signals. On the one hand, we exploit the conditional feedback from the diffusion model for different text prompts to build a text-driven contrastive learning strategy. On the other hand, we introduce a noise-driven consistency regularization to align local instances with diffusion denoising representations, constraining the optimization region in the feature space. In addition, FedDifRC can be extended to a self-supervised scheme without relying on any labeled data. We also provide a theoretical analysis for FedDifRC to ensure convergence under non-convex objectives. The experiments on different scenarios validate the effectiveness of FedDifRC and the efficiency of crucial components. Code is available at https://github.com/hwang52/FedDifRC. + +# 1. Introduction + +Federated Learning (FL) [34] has become increasingly popular as a decentralized training paradigm, with multiple distributed clients train their local models without sharing their + +![](images/4daf08b7f68934a45627e2d5ccb5014341fe08e7eb4fa784852b91c48d022de6.jpg) +(a) Text-Driven Diffusion Contrastive Learning (TDCL) + +![](images/b3b5bcf8dc4bfda73e7eed4c30ae8896bfba56f425d6305ad06f91e081f9583e.jpg) +(b) Noise-Driven Diffusion Consistent Restriction (NDCR) +Figure 1. The main idea of FedDifRC with two key modules TDCL and NDCR. (a) We leverage conditional generative feedback of the pre-trained diffusion model on various text prompts to devise a text-driven contrastive learning strategy, where the input embedding is injected as a condition. (b) We extract the diffusion denoising feature maps of intermediate layers from the pre-trained diffusion model at a specific time step $t$ during the diffusion backward process to construct a noise-driven regularization. + +private dataset [16, 28] (i.e., training client models locally and aggregating them globally [29]). One major challenge faced in FL is the potential discrepancies in the local data distributions across distributed clients, which is known as the data heterogeneity issue [32, 35, 44]. Specifically, the private data is collected from different sources with diverse preferences and presents non-independent and identically distributed (non-iid) distribution [27, 75]. Such non-iid data + +discrepancies can lead to inconsistency in the clients' local objective functions and optimization directions, making the model convergence slow and unstable [17, 29, 30, 39, 65]. + +To mitigate the negative impact of the data heterogeneity, a mainstream of subsequent efforts delves into introducing a variety of suited signals to regulate the private model, there are mainly two perspectives: i) local model optimization at the client side and ii) global model aggregation at the server side. The former focuses on optimization strategies to make the variance of local updates limited [6, 29]. Another aims to improve the efficacy of global model aggregation [36, 69, 73]. In fact, most of these previous methods force the local models to be consistent to the global model. Although they have a certain effect by reducing gradient inconsistency, the gradually enlarged parameter deviation persists. Hence, we argue that local model overfits the local data distribution, which is magnified under the data heterogeneity and thus hinders the improvement of model's performance in FL. + +Recently, the development of diffusion models (DMs) [4, 11, 37, 46] provides fresh opportunities. The diffusion models have demonstrated remarkable power in generating the diverse and realistic images, showing strong ability to capture intricate details and patterns from underlying data. For instance, with proper guidance, stable diffusion (SD) [46] models can effectively align a text prompt with the corresponding image context, which suggests their potential ability to understand both high-level and low-level contextual visual concepts regarding what an image contains. This semantic generative and visual comprehension capabilities of SD models motivate us to think: is it possible to exploit the inherent fine-grained visual representation capabilities of diffusion models to promote heterogeneous FL training? + +Taking into account both the effectiveness and efficiency, we revisit the underlying representations in the SD models, which are intermediate feature maps in a diffusion process (e.g., denoise or noise an image) produced by multiple calling a UNet [47]. We argue that diffusion representations are effective steers for FL training: i) extensive general knowledge inherent in diffusion models enhance local semantic diversity in FL; ii) smooth correspondences about semantic objects in diffusion models is a natural guiding signal in FL. Building on these two insights, we first propose conditional diffusion representations by imposing various text prompts in the SD model. In this way, each instance is contrastively enhanced with a set of text-driven diffusion representations, capturing rich domain variance. Then, we devise denoising diffusion representations from semantic relations between visual and linguistic concepts in the SD model. We extract the feature maps from different hierarchies of the UNet decoder in the SD model's denoising process, serving as guidance signals to avoid over-optimization in FL local training. + +In this paper, we propose a diffusion-inspired Federated framework with Diffusion Representation Collaboration, as + +FedDifRC, to handle the data heterogeneity issue in FL, which consists of two key modules (see Fig. 1): i) First, we devise Text-driven Diffusion Contrastive Learning (TDCL), which utilizes conditional diffusion representations to construct inter-contrastive learning [1, 9, 26, 58] in a text-driven diffusion manner (the sample's embedding is injected as a condition). By maximizing the agreement between the local instance and its corresponding representations, TDCL adaptively captures class-relevant information and semantically meaningful knowledge. This objective encourages sample's embedding closer to conditional diffusion representations with same semantics and away from other negatives, thus maintaining a clear decision boundary. ii) Second, we use the denoising diffusion representations to build a flat and stable convergent target and further propose Noise-driven Diffusion Consistent Restriction (NDCR). Concretely, we extract denoising feature maps of the noisy input sample in the SD model's backward process, as a regularization term to constrain the optimization region of the local model. The local sample is forced to align the embedding with its denoising representations at the feature space to promote the uniformity for local training. + +Our main contributions are summarized as follows: + +- We explore the possibility of using diffusion models to tackle data heterogeneity in FL and devise the conditional and denoising diffusion representations as effective steers to regulate model training in heterogeneous FL. +- We propose FedDifRC, a diffusion-inspired FL framework to handle data heterogeneity by leveraging the complementary advantages of two key components: i) TDCL facilitates the local model to learn more general class-relevant knowledge; ii) NDCR enforces the uniformity in the feature space for local training. +- We conduct extensive experiments for different FL scenarios (e.g., label shift, domain shift, and long-tailed) and different application scenarios (i.e., intra-domain and cross-domain) to demonstrate the effectiveness of our FedDifRC and the indispensability of each component. + +# 2. Related Work + +# 2.1. Heterogeneous Federated Learning + +Federated learning is proposed to address privacy concerns in a distributed learning environment (e.g., Federated Averaging (FedAvg) [34]). Several studies have improved the local update process on the client side or the global aggregation process on the server side to tackle data heterogeneity in FL. The former focuses on local optimization strategies to make the diversity between the client model and the global model limited in the parameter level [6, 12, 13, 17, 26, 29, 35, 52]. The another branch aims to adopt aggregation mechanisms on the server to alleviate the negative influence of data heterogeneity [23, 43, 60, 62, 69, 73]. + +Recently, existing FL works that apply diffusion models mainly focus on building synthetic datasets [25, 51, 67, 74]. With the assistance of synthetic datasets, the FL global model can be retrained by this data augmentation scheme to mitigate data heterogeneity. Although these generated data have comparable quality and diversity to client images, the local model may overfit local domain distribution, and the inherent heterogeneity across clients may not have been addressed. In contrast, we focus on how to effectively leverage the fine-grained visual representation from diffusion models to promote a generalizable FL global model. + +# 2.2. Diffusion Model Representations + +Recently, diffusion models have shown significant advances in image generation, editing, and stylized [4, 11, 37, 46, 66], suggesting that they contain rich underlying representations to be exploited for different downstream tasks. DIFT [54] is used to extract the diffusion features without explicit supervision. Diffusion Hyperfeatures [31] integrates multi-scale and multi-timesteps diffusion features into feature descriptors. The diffusion representations have shown impressive results on the vision branch, including being testified as certifiable classifiers [3, 24, 63] and semantics correspondence learners [54, 64, 72]. In this paper, we expand the diffusion models to FL and construct the conditional and denoising diffusion representations, offering fruitful class-relevant information and consistent convergence signals. + +# 2.3. Contrastive Learning + +Contrastive learning (CL) has been widely applied to self-supervised learning scenarios [1, 7, 9]. Many works are focused on learning the encoder where the embeddings of the same sample are pulled closer and those of different samples are pushed apart [2, 5, 18, 55]. They mainly construct a positive pair and some negative pairs for each data instance and use InfoNCE [38] to contrast positiveness against negativity. There are some works that incorporate contrastive learning into FL to assist local training to achieve higher model performance [8, 12, 26, 53]. In our work, Text-driven Diffusion Contrastive Learning (TDCL) is devised to promote local model by attracting samples to positive representations while pushing away other negatives. + +# 3. Diffusion Models Help Federated Learning + +In this section, we begin with a brief overview of the general FL paradigm and the generative diffusion models (Sec. 3.1). Then, we analyze the effectiveness of diffusion representations and illustrate our core insights (Sec. 3.2). Finally, we will introduce how to extract the conditional and denoising diffusion representations in Sec. 3.3 and Sec. 3.4. + +# 3.1. Preliminaries and Background + +Federated Learning: We focus on the case with $K$ clients holding the heterogeneous data partition $\{\mathcal{D}_1,\mathcal{D}_2,\ldots ,\mathcal{D}_K\}$ with private dataset $\mathcal{D}_k = \{x_i,y_i\}_{i = 1}^{n_k},y_i\in \{1,\dots ,|\mathcal{C}|\}$ for the $k$ -th client, where $n_k$ denotes the local dataset scale. Let $n_k^j$ be the number of samples with label $j$ at client $k$ and $\mathcal{D}_k^j = \{(x,y)\in \mathcal{D}_k|y = j\}$ denotes the set of samples with label $j$ on the $k$ -th client, so the number of samples in $j$ -th class is $n^j = \sum_{k = 1}^{K}n_k^j$ . The objective of the general FL is to learn an optimal global model $w$ across $K$ clients: + +$$ +\min _ {w} \mathcal {L} (w) = \frac {1}{K} \sum_ {k = 1} ^ {K} \frac {n _ {k}}{N} \mathcal {L} _ {k} (w; \mathcal {D} _ {k}), \tag {1} +$$ + +where $\mathcal{L}_k$ is the $k$ -th client local loss function and $w$ is the global model parameters. $N$ is the total number of samples among all clients. In the FL training setup, we consider the local model $f$ with parameters $w = \{u,v\}$ . It has two modules: i) a feature extractor $h(u)$ with the parameters $u$ maps each sample $x_{i}$ to a $d$ -dim (default as 512) vector as $z_{i} = h(u;x_{i})$ ; ii) a classifier $g(v)$ with parameters $v$ maps $z_{i}$ into a $|\mathcal{C}|$ -dim output logits as $l_{i} = g(v;z_{i})$ . + +Diffusion Models: Given the image sample $x$ from an underlying distribution $p(x)$ , a forward diffusion process as a Markov chain to add random Gaussian noise $\epsilon \sim \mathcal{N}(0, \mathbf{I})$ to the original sample $x_0$ , and $x_t$ is the noise sample after adding $t$ -steps noise: + +$$ +x _ {t} = \sqrt {1 - \gamma_ {t}} x _ {t - 1} + \sqrt {\gamma_ {t}} \epsilon_ {t}, t \in \{1, \dots , T \}, \tag {2} +$$ + +where $T$ is the number of diffusion timesteps, $\gamma_t \in (0,1)$ is a adjustable time variance schedule [11]. By leveraging the additive property of the Gaussian distribution, Eq. (2) can be further reformulated as $x_t = \sqrt{\bar{\alpha}_t} x_0 + \sqrt{1 - \bar{\alpha}_t} \epsilon_t$ , where $\alpha_t = 1 - \gamma_t$ and $\bar{\alpha}_t \doteq \prod_{i=1}^t \alpha_i$ . On this basis, the image $x_0$ can be obtained from a random noise $x_t \sim \mathcal{N}(0,\mathbf{I})$ by reversing the above forward diffusion process [11]: + +$$ +x _ {t - 1} = \frac {1}{\sqrt {\alpha_ {t}}} \left(x _ {t} - \frac {1 - \alpha_ {t}}{\sqrt {1 - \bar {\alpha} _ {t}}} \epsilon_ {\theta} \left(x _ {t}, t\right)\right), t \in \{T, \dots , 1 \}, \tag {3} +$$ + +where $\epsilon_{\theta}$ denotes the denoiser network with the parameters $\theta$ to predict noise $\epsilon$ , and the denoiser network is usually implemented as a UNet architecture [47]. Further, the denoiser $\epsilon_{\theta}$ is optimized using the loss $T_{\theta}$ as follows: + +$$ +\mathcal {T} _ {\theta} = \mathbb {E} _ {\epsilon , x _ {0}, t} [ \| \epsilon_ {\theta} (x _ {t}, t) - \epsilon \| _ {2} ^ {2} ], \epsilon \sim \mathcal {N} (0, \mathbf {I}), \tag {4} +$$ + +where $\epsilon$ is the noise in the forward diffusion process, and $\epsilon_{\theta}$ is usually implemented as a UNet [47]. Besides, the diffusion model's objective in Eq. (4) can be easily extended to a conditional diffusion paradigm by adding a condition c to the $\epsilon_{\theta}$ . Thus, the training objective should be modified as: + +$$ +\mathcal {T} _ {\theta} = \mathbb {E} _ {\epsilon , x _ {0}, t, \mathbf {c}} [ | | \epsilon_ {\theta} (x _ {t}, t, \mathbf {c}) - \epsilon | | _ {2} ^ {2} ], \epsilon \sim \mathcal {N} (0, \mathbf {I}), \tag {5} +$$ + +with the denoiser $\epsilon_{\theta}(x_t,t,\mathbf{c})$ in Eq. (5), we can generate data with a specific condition $\mathbf{c}$ by solving a reverse process. + +# 3.2. Motivation and Analysis + +The diffusion model is a probabilistic generative model designed for denoising by systematically reversing a progressive-noising process, and recent studies [61, 68] have verified that diffusion-based generative models can be effective representation learners. Stable diffusion models are able to weave from scratch with fruitful and innovative depictions of high fidelity and further capture the underlying properties, positions, and correspondences from visual concepts. Such a phenomenon shows that SD models can align a textual prompt with its corresponding image context, demonstrating their potential ability in understanding both high-level (e.g., semantic relations) and low-level (e.g., textures, edge, and structures) visual concepts in an image. + +Autoencoders [71, 76] can learn to reconstruct their input at the output and have been one of the dominant approaches for representation learning. We assume a linear autoencoder expressed as $W_{a}W_{b}\cdot x + W_{c}\cdot x$ to imitate UNet [47] used in diffusion models, where $W_{a}\in \mathbb{R}^{d\times d^{*}}$ , $W_{b}\in \mathbb{R}^{d^{*}\times d}$ , and $W_{c}\in \mathbb{R}^{d\times d}$ , $d^{*} < d$ . For $x_{t} = \sqrt{\bar{\alpha}_{t}} x_{0} + \sqrt{1 - \bar{\alpha}_{t}}\epsilon_{t}$ and we set $\mathbb{E}[x_0] = 0$ , thus minimizing the loss $\mathcal{T}_{\theta}$ in Eq. (4) is equivalent to minimize the modified objective $\mathcal{T}_{\theta}^{*}$ : + +$$ +\begin{array}{r l} \mathcal {T} _ {\theta} ^ {*} = & \mathbb {E} _ {\epsilon , x _ {0}, t} [ \| \epsilon_ {\theta} (\sqrt {\bar {\alpha} _ {t}} x _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon_ {t}) - \epsilon \| _ {2} ^ {2} ], \\ & \equiv \min \left\{\bar {\alpha} _ {t} \left(W _ {a} W _ {b} + W _ {c}\right) ^ {\top} \Sigma_ {x _ {0}} \left(W _ {a} W _ {b} + W _ {c}\right) \right\}, \end{array} \tag {6} +$$ + +where $\Sigma_{x_0}$ means covariance matrix of $x_0$ . Eq. (6) ensures a meaningful latent space for the diffusion models, similar to the linear autoencoder that encodes data into the principal components space. Therefore, the diffusion training process is equivalent to encourage the disentanglement of the latent representations until good-quality reconstruction is reached. + +To validate the representation ability in diffusion models, we use t-SNE [57] to visualize the latent features, shown in Fig. 2, without training on the specific MNIST [22] dataset. Concretely, we use the pre-trained SD model to extract features of the UNet decoder $\epsilon_{\theta}$ , $T$ means diffusion time-steps, $L$ means the layer of the UNet decoder $\epsilon_{\theta}$ ( $L \in \{1, 2, 3, 4\}$ ). Our results in Fig. 2 indicate that, even without supervision, diffusion models are able to group input samples at properly selected time-steps $T$ and layer $L$ . However, when the $T$ or $L$ is too large (e.g., $T = 999$ or $L = 4$ ), the representations become inseparable, leading to blurred decision boundaries. Furthermore, we apply K-Means clustering (cluster size set as 5) on the latent features from the diffusion UNet decoder and visually examine whether the diffusion features contain consistent semantic information. In Fig. 3, different parts of the visual objects are clustered, showing that earlier layers (e.g., $L = 2$ ) capture coarse yet consistent semantics, while later layers (e.g., $L = 3$ ) focus on low-level textural details. Both t-SNE and K-Means observation results motivate us to leverage the fine-grained visual representation capabilities from the SD model to promote heterogeneous FL training. + +![](images/f8a6c63c46d755a35a6ffff60731a424d3f285a268df8ff1f96c7c9045251e21.jpg) +$\mathrm{T} = 100$ $\mathrm{L} = 1$ + +![](images/3a67be20e7b2fa7dea25cf54aeee10550fa233edc80ef511bd75c0a32dba453e.jpg) +$\mathrm{T} = 100, \mathrm{L} = 2$ + +![](images/636c14c8e2b0ef1c96ad52b480b999aeddd00272a308840208646698dabcab93.jpg) +$\mathrm{T} = 100, \mathrm{L} = 3$ + +![](images/ac0778df95d294b66dd811290369d870d1189996d3f8f939f30da54a3259096d.jpg) +$\mathrm{T} = 100, \mathrm{L} = 4$ + +![](images/8f3dea7801cb18ee61ea8cc8202d2539d0362686418ad5b2c0599d676ebf616b.jpg) +$\mathrm{T} = 100, \mathrm{L} = 2$ + +![](images/0612edbe36ba27a6efb4d65a7d2454bb7cd5106497a9e1e729cb9bdffae76c33.jpg) +$\mathrm{T} = 300$ $\mathrm{L} = 2$ + +![](images/b5ac2f9cb919bc3bd8789d3d122db37ff8664819bd0f8d8728a60e258fb92019.jpg) +$\mathrm{T} = 500$ $\mathrm{L} = 2$ + +![](images/49fec3ae084372be889e25e172bd4c71f8b0abe8bf2aca0967bc5eb06b318a0f.jpg) +$\mathrm{T} = 999, \mathrm{L} = 2$ + +![](images/7e8df0660532b3ecf50da30033198c24741b53c5ad94df89dfd4e887f077b15c.jpg) +Figure 2. The t-SNE feature visualization on the MNIST dataset from the diffusion UNet decoder, without training, where diffusion time-steps $T \in \{100, 300, 500, 999\}$ and layer $L \in \{1, 2, 3, 4\}$ . + +![](images/4e176ea07de581eeabc07115f2217e77dbf22c2d51a899cb9152bd0b198084c8.jpg) + +![](images/8d53eedb8fc17507590e81858cd79f3f732586882b3b6cef15cd6fbdba102d72.jpg) +Sample + +![](images/3a48b6ffa541373ec0a7ac0c6a64d553b2365d202a359fdd995129792049d321.jpg) + +![](images/4cd45332ef21f842c993fc9885e7458cb0052572a160c49e8a6ffd21bf40ad05.jpg) + +![](images/9271bb3537daa97e75ba5164c4ed302a703d2f2cbab43ff5dc03fabdb110fa70.jpg) +Layer $\mathrm{L} = 1$ + +![](images/a566f150a4f4598c07fc59202fef3f261e16d993fbb9901e67ea0dccf9a9f8f2.jpg) + +![](images/be6576963a4e3bcbd2421fcf63fd7c2cb2067b4fbcba4e5a47f0053c66c42e89.jpg) + +![](images/f53f98e68a2d16cc8116be8e0433fdd0b83af261cf2c0caf7d751d3d54e8ea9a.jpg) +Layer $\mathrm{L} = 2$ + +![](images/d37db70802b268e552fadeee8f1e0f6f56d61ef3f6a60cbfabe5b718ff58fe6d.jpg) + +![](images/25b1343cd9d6f5d90e063b0436637742948ecc1a97a9d2a54a0c563e129bde26.jpg) + +![](images/684dea9a71546e911a440b052a480f881082113ca032a2d2f07b26bc45e80915.jpg) +Layer $\mathrm{L} = 3$ + +![](images/63b94aa5355f2b8cd2d400d6e28f0935ce05e80457426ac5aef17be979d60175.jpg) + +![](images/9a4072b7bff2a1f37b9c139730ac474828e105edb1c833f6ac2cc89c166c87c2.jpg) + +![](images/63d28bfb0bdf8bdf0d037d5996ff91236fd594adc490751ff284b6cf047f94cb.jpg) +Layer L=4 + +![](images/235e2f1b18ac52bbb92c196436c9d08492973c42d19545bf9d893b01d660e043.jpg) +Figure 3. The K-Means clustering feature visualization (each color represents a cluster) from the diffusion UNet decoder's layers. + +![](images/b60abde78623d328711c61c2a319a1da70e85a976831f4eec0f61ebf388a54c8.jpg) + +![](images/a2a19a0e4d27aad4549756fd7424b2824cc3a6bd763d36427eed0c0a49061fb0.jpg) +Sample + +![](images/b3fb60e29d3b8928ff02be35e06d79b5ea4103f83d37afa7e3f5030e2961c0cf.jpg) + +![](images/2be7cb8d134bd81d251ecc69eff81fa36b38e5132dde342f577daaecb1cba954.jpg) + +![](images/d5485b794e5d93b1443ccc7eb506b6910483bbad041609356317fda952e28abf.jpg) +Layer $L = 2$ + +![](images/34285cf22b2b7ad3d94c5d074149060a33a743cdb931914febe3ca71893dd38f.jpg) + +![](images/de4b2a908528032c4ee7e0ab4063267994b2816e052a4be081747022546fcba7.jpg) + +![](images/6c47cea1cb100428c7a1881367ae23f4884ca57853d724e12cad253a6ef65b11.jpg) +Layer $\mathrm{L} = 3$ + +![](images/b3fec03c3b8ea8199f936d8baa20830a15e78f6404f1430e9b4051e554f8e354.jpg) + +![](images/85bb727f570431443904a0b95e060875dbdfbb73c1a2f3b34491ed408b4da011.jpg) + +![](images/652398c7f9ecdf1a7abecd29e32d01ea9b3463422d82cf2be49348821ae6eae9.jpg) +Layer L=4 +Figure 4. The PCA visualization of early (layer 2), later (layer 4), and fused features (see Eq. (8)), where the first three components of PCA-computed serve as color channels for each image. + +![](images/a60727202fe7205c4a39ee3040def9e46bf97f57cb3a5dc6e79503fab55b88b5.jpg) + +![](images/09cf4bc4d629bf2f1cf8eea5a44b1cb9561f5245ff380917365cc69ff3c891a0.jpg) + +![](images/0f1dd56c24bdc84536d1738b6e48f6e8da2e928851787d9884bfb47734b33195.jpg) +Fused + +# 3.3. Conditional Diffusion Representations + +The key idea of our conditional diffusion representations is based on cognitive diversity of generative diffusion models for different text prompts, via a condition-guided manner. Specifically, for the client $k$ , given an input image $x_{i}$ of the local dataset $\mathcal{D}_k = \{x_i,y_i\}_{i = 1}^{n_k},y_i\in \{1,\dots,|\mathcal{C}|\}$ means the label of $x_{i}$ , and we let $S$ represents the set of all class names, + +so we use $S_{y_i}$ to denote the class name of the label $y_i$ . From our implementation, the image $x_i$ is fed into the encoder $h_k$ of the local model $f_k$ to obtain the embedding $\mathbf{c}_i$ , then $\mathbf{c}_i$ is treated as a condition and injected into the diffusion model. For a pre-trained SD model, $\mathbf{c}_i$ is further combined with a relevant text prompt $\mathcal{P}_{y_i} =$ 'a photo of a $S_{y_i}$ ' and sent to the UNet decoder to produce diffusion representations $\mathcal{F}_i$ . Note that since we do not add noise, we simply set the time-steps $T = 0$ of the diffusion process. Typically, the image $x_i$ size is $224 \times 224$ , and $\mathcal{F}_i$ contains 4 feature maps, where the $L$ -th layer's feature map $\mathcal{F}_{i(L)}$ has the spatial size $\{7,14,28,28\}$ ( $L \in \{1,2,3,4\}$ ), and $\{1280,1280,640,320\}$ are channels. Formally, conditional diffusion representations $\mathcal{F}_i$ are computed by the $\epsilon_{\theta}$ in Eq. (5), image $x_i$ , condition $\mathbf{c}_i$ , text $\mathcal{P}_{y_i}$ : + +$$ +\mathcal {F} _ {i} = \epsilon_ {\theta} \left(x _ {i}, \tau_ {i}, T = 0, \mathbf {c} _ {i}\right), \tau_ {i} \leftarrow \varphi \left(\mathcal {P} _ {y _ {i}}\right), \tag {7} +$$ + +where $\varphi$ denotes the text encoder of the SD model and $\tau_{i}$ is the text features with the prompt $\mathcal{P}_{y_i}$ . Inspired by the results in Fig. 3, which motivate us to combine features at different layers of $\mathcal{F}_i$ to capture both semantics and details. A simple concatenation, however, can lead to an unnecessarily high-dimensional feature $(1280 + 640 + 320 = 2240)$ . To reduce high dimension, we apply PCA for each layer $L\in \{2,3,4\}$ (due to the spatial resolution too low, we do not consider the $L = 1$ ), and we aggregate them to the same resolution, then we connect them together to obtain the fused features $\widetilde{\mathcal{F}}_i$ : + +$$ +\widetilde {\mathcal {F}} _ {i} = \mathrm {C o n c a t} (\widetilde {\mathcal {F}} _ {i (L)}), \widetilde {\mathcal {F}} _ {i (L)} = \mathrm {P C A} (\mathcal {F} _ {i (L)}), \quad (8) +$$ + +where $L\in \{2,3,4\}$ we perform PCA: i) for $\mathcal{F}_{i(2)}$ we first upsample to the same resolution and apply PCA $\mathbb{R}^{1280\to 256}$ ii) for $\mathcal{F}_{i(3)}$ and $\mathcal{F}_{i(4)}$ , we apply PCA $\mathbb{R}^{640\to 128}$ and PCA $\mathbb{R}^{320\to 128}$ . Then, we combine them to form the final fused features $\widetilde{\mathcal{F}}_i$ . As shown in the last column of Fig. 4, the fused features strike a balance between high-level and low-level visual concepts, focusing on both semantics and textures in the image. On the example in the second row of Fig. 4, the fused feature yields more accurate matches (e.g., face and computer borders). The comparison results in Table. 2 also confirm this viewpoint. We define $\widetilde{\mathcal{F}}_i$ calculated by Eq. (8) as conditional diffusion representations of the sample $x_{i}$ + +# 3.4. Denoising Diffusion Representations + +Although we construct the conditional diffusion representations from the diffusion model's generative feedback, $\widetilde{\mathcal{F}}_i$ is still highly variable and depends on the condition generated by the local model, which can not ensure a consistent stable convergence signal. Based on Eq. (4), we explore the inherent denoising property in diffusion models: SD model has learned extensive general knowledge about identical semantics via large pre-training, reflected in the UNet denoising process. Therefore, we aim to extract the denoising diffusion representations contained in a pre-trained SD model, to + +align the sample's embedding with the class semantics uniformly in the feature space. Given the image pair $\{x_i, y_i\}$ of $k$ -th client's dataset $\mathcal{D}_k$ , we first perform a forward process to add $t$ time-steps random Gaussian noise on $x_i$ as $\tilde{x}_i^t = \sqrt{\bar{\alpha}_t} x_i + \sqrt{1 - \bar{\alpha}_t} \epsilon_t$ , similar to [46], where $\bar{\alpha}_t$ and $\epsilon$ are the same as defined in the Eq. (3). Next, we set the text prompt of $\tilde{x}_i^t$ as $\mathcal{P}_{y_i}$ = 'a photo of a $S_{y_i}$ ' used in the SD model pre-training [46], to avoid destroying correspondence between visual and linguistic concepts, and the $S_{y_i}$ denotes the class name of the label $y_i$ . Formally, the denoising diffusion representations $\tilde{\mathcal{H}}_i$ of the sample $x_i$ are calculated by the $\epsilon_\theta$ in Eq. (4), noisy sample $\tilde{x}_i^t$ , text $\mathcal{P}_{y_i}$ : + +$$ +\mathcal {H} _ {i} = \epsilon_ {\theta} (\tilde {x} _ {i} ^ {t}, \tau_ {i}, T = t), \tau_ {i} \gets \varphi (\mathcal {P} _ {y _ {i}}), +$$ + +$$ +\widetilde {\mathcal {H}} _ {i} = \operatorname {C o n c a t} \left(\widetilde {\mathcal {H}} _ {i (L)}\right), \quad \widetilde {\mathcal {H}} _ {i (L)} = \operatorname {P C A} \left(\mathcal {H} _ {i (L)}\right), \tag {9} +$$ + +where $\varphi$ denotes the text encoder of the SD model and $\tau_{i}$ is the text features of the text prompt $\mathcal{P}_{y_i}$ . We further analyze the effects of the time-steps $t$ in Fig. 6. Then, we combine the features of layers $L\in \{2,3,4\}$ , like Eq. (8), to obtain fused features $\widetilde{\mathcal{H}}_i$ . Intuitively, $\widetilde{\mathcal{H}}_i$ is considered as a virtual teacher to calibrate the embedding of $x_{i}$ at the feature-level. We define the denoising features $\widetilde{\mathcal{H}}_i$ calculated by Eq. (9) as denoising diffusion representations of the sample $x_{i}$ . + +# 4. Methodology + +Our proposed FedDifRC leverage conditional and denoising diffusion representations to obtain diverse class-relevant knowledge and consistent convergence signals by two complementary components: Text-driven Diffusion Contrastive Learning in Sec. 4.1 and Noise-driven Diffusion Consistent Restriction in Sec. 4.2. Further, we discuss how to extend FedDifRC to a self-supervised scheme (Sec. 4.3). Finally, we provide a theoretical analysis of FedDifRC under nonconvex case (Sec. 4.4). Illustration of FedDifRC in Fig. 5. + +# 4.1. Text-Driven Diffusion Contrasting + +Inspired by success of contrastive learning [1, 9, 26], based on Sec. 3.3, conditional diffusion representations $\widetilde{\mathcal{F}}_i$ of sample $\{x_i,y_i\} \in \mathcal{D}_k$ can be obtained with the condition $\mathbf{c}_i$ and the text prompt $\mathcal{P}_{y_i} = \text{'a photo of a } S_{y_i}$ , where $S_{y_i}$ is the class name of the label $y_{i}$ . For each sample's embedding $z_{i} = h_{k}(x_{i})\in \mathbb{R}^{d}$ , we enforce $z_{i}$ to be similar to respective positive pair $\widetilde{\mathcal{F}}_i$ and dissimilar to negative sample pairs $\widetilde{\mathcal{F}}_{\mathcal{N}_i}$ where $\widetilde{\mathcal{F}}_{\mathcal{N}_i}$ means the conditional diffusion representations with the same condition $\mathbf{c}_i$ and the unmatched text prompts $\mathcal{P}_j = \text{'a photo of a } S_j$ ( $S_{j}\in S$ and $S_{j}\neq S_{y_{i}}$ ). Then, we define the similarity between $z_{i}$ and $\widetilde{\mathcal{F}}_i$ as follows: + +$$ +s \left(z _ {i}, \widetilde {\mathcal {F}} _ {i}\right) = \frac {z _ {i} \cdot \widetilde {\mathcal {F}} _ {i}}{\mathcal {U} \left(\left\| z _ {i} \right\| _ {2} \times \left\| \widetilde {\mathcal {F}} _ {i} \right\| _ {2}\right)}, \mathcal {U} = \frac {1}{n _ {k}} \sum_ {m = 1} ^ {n _ {k}} \| z _ {m} - \widetilde {\mathcal {F}} _ {i} \| _ {2}, \tag {10} +$$ + +![](images/3f5f6e3e5ecb3bee521b4d2bd1782a532d785516bcdbb4e1af3e2a984eb420ca.jpg) +Training Process for Local Client +Figure 5. Illustration of FedDifRC, including two complementary modules: TDCL (Sec. 4.1) builds an inter-contrastive learning strategy by conditional diffusion representations (Sec. 3.3), NDCR (Sec. 4.2) designs a consistent regularization penalty via denoising diffusion representations (Sec. 3.4), jointly promoting FL training. + +where $n_k$ is total number of samples at client $k$ , $\mathcal{U}$ is a normalization factor for similarity, where we set $\mathcal{U}$ to denote the average distance between $z_m|_{m=1}^{n_k}$ and $\widetilde{\mathcal{F}}_i$ on dataset $\mathcal{D}_k$ . Next, we introduce TDCL to quantify and regulate the local model training, which can be formulated as follows: + +$$ +\mathcal {L} _ {T D C L} = \log (1 + \frac {\sum_ {\widetilde {\mathcal {F}} _ {j} \in \widetilde {\mathcal {F}} _ {\mathcal {N} _ {i}}} \exp \left(s \left(z _ {i} , \widetilde {\mathcal {F}} _ {j}\right) / \tau\right)}{\exp \left(s \left(z _ {i} , \widetilde {\mathcal {F}} _ {i}\right) / \tau\right)}), \tag {11} +$$ + +where $\tau$ denotes a temperature parameter to control representation strength [1]. By minimizing Eq. (11), for client $k$ , the local model brings each sample's embedding $z_{i}$ closer to the positive pair and away from other negative pairs. + +# 4.2. Noise-Driven Diffusion Regularization + +Despite each local model aligning sample's embedding with conditional diffusion representations by optimizing $\mathcal{L}_{TDCL}$ in Eq. (11). Since conditional diffusion representations are dynamically built from text-driven cross-modal matching and depend on generated condition $\mathbf{c}_i$ at each round, which could not offer a stable convergence direction. Therefore, we aim to leverage denoising diffusion representations $(\widetilde{\mathcal{H}}_i)$ in Eq. (9)) to devise a consistent regularization penalty, constraining model's optimization region in the feature space. We design Noise-driven Diffusion Consistent Restriction (NDCR) and utilize a regularization term to pull the sample's embedding $z_{i}$ closer to the respective $\widetilde{\mathcal{H}}_i$ : + +$$ +\mathcal {L} _ {N D C R} = \sum_ {q = 1} ^ {d} \left(z _ {i (q)} - \widetilde {\mathcal {H}} _ {i (q)}\right) ^ {2}, \tag {12} +$$ + +where $q$ indexes the dimension $d$ of the embedding vector. Based on Eq. (12), we expect to align the sample's embedding with the corresponding denoising diffusion representations, enforcing semantics uniformity of the feature space. + +Besides, we construct CrossEntropy loss and use the logits output vector $l_{i} = g(z_{i}) = f(x_{i})$ with the original label $y_{i}$ to optimize the local discriminative ability for each client. Formally, given the sample pair $\{x_{i},y_{i}\}$ to calculate $\mathcal{L}_{CE}$ : + +$$ +\mathcal {L} _ {C E} = - \mathbf {1} _ {y _ {i}} \log (\psi (f (x _ {i}))), \tag {13} +$$ + +where $\psi$ denotes the Softmax function, $f$ is the local model with parameters $\{u,v\}$ maps $x_{i}$ into a $|\mathcal{C}|$ -dim output logits $l_{i} = f(x_{i}) = g(h(x_{i}))$ . Finally, for the $k$ -th client, we carry out the following optimization objective in local training: + +$$ +\mathcal {L} = \mathcal {L} _ {T D C L} (\mathrm {E q .} (1 1)) + \mathcal {L} _ {N D C R} (\mathrm {E q .} (1 2)) + \mathcal {L} _ {C E}. \tag {14} +$$ + +The algorithm of FedDifRC is given in Appendix A. + +# 4.3. Expanding to Self-Supervised Scheme + +FedDifRC can be extended to the self-supervised scheme without relying on any labeled data source. Specifically, for $\mathcal{L}_{TDCL}$ in Eq. (11), we only need to modify text template slightly: i) we set the text prompt $\mathcal{P}_{y_i} = \text{'a photo of a similar object'}$ in Eq. (7) to generate $\widetilde{\mathcal{F}}_i$ as the positive pair; ii) then we randomly select from the set of all class names in TinyImageNet [21] to get $S_{\mathcal{N}_i}$ , and set $\mathcal{P}_{y_i} = \text{'a photo of a } S_{\mathcal{N}_i}$ to obtain $\widetilde{\mathcal{F}}_{\mathcal{N}_i}$ as negative pairs. For the $\mathcal{L}_{NDCR}$ in Eq. (12), we set $\mathcal{P}_{y_i} = \text{'a photo of a visual object'}$ in Eq. (9) to get $\widetilde{\mathcal{H}}_i$ as consistent signal to align the sample's embedding. Note that it is also possible to use class names of other datasets (e.g., CIFAR100 [20]) to generate $\widetilde{\mathcal{F}}_{\mathcal{N}_i}$ as negative pairs. By this self-supervised way, the objective term of local training becomes $\mathcal{L}_{TDCL} + \mathcal{L}_{NDCR}$ . The results in Table. 5 validate the effectiveness of FedDifRC's self-supervised scheme. + +# 4.4. Convergence Analysis + +We proved convergence of FedDifRC in non-convex case. We denote $\mathcal{L}$ (Eq. (14)) as $\mathcal{L}_r$ with the round $r$ , assumptions (Appendix B.2) are similar to existing works [6, 29, 30, 69]. + +Convergence Rate of FedDifRC: Let the round $r$ from 0 to $R - 1$ , given any $\xi > 0$ , the $\mathcal{L}$ will converge when, + +$$ +R > \frac {2 \left(\mathcal {L} _ {0} - \mathcal {L} ^ {*}\right)}{\xi E \eta \left(2 - L _ {1} \eta\right) - \Omega_ {1} - \Omega_ {2}}, \tag {15} +$$ + +and we can further get the following condition for $\eta$ , + +$$ +\eta < \frac {2 \xi - 2 (| \mathcal {C} | - 1) L _ {2} B}{L _ {1} (\xi + \sigma^ {2})}, \tag {16} +$$ + +where $\Omega_{1} = 2(|\mathcal{C}| - 1)L_{2}E\eta B$ , $\Omega_{2} = L_{1}E\eta^{2}\sigma^{2}$ , and $\mathcal{L}^*$ denotes the optimal solution of $\mathcal{L}$ , $\{L_1,L_2,B,\sigma^2\}$ are constants in the assumptions from Appendix B.2, and $|\mathcal{C}|$ as the number of classes, $E$ as local epochs. Eq. (15) and Eq. (16) ensure convergence rate of FedDifRC, after selecting the number of rounds $R$ and the learning rate $\eta$ . So the smaller $\xi$ is, the larger $R$ is, which means that the tighter the bound is, the more rounds $R$ is required. The detailed assumptions and proofs are formally provided in Appendix B. + +
TDCLNDCRCIFAR10 (keep ρ = 1.0)
NID10.05NID10.2NID10.5NID2AVG
78.2784.6586.1172.6080.41
✓81.3986.0388.1675.6782.81
✓80.3586.4087.5475.3382.40
✓✓83.1488.2789.3176.4584.29
TDCLNDCRCIFAR10-LT (keep NID10.2)
ρ = 10ρ = 50ρ = 100ρ = 200AVG
75.9463.2552.9749.6960.46
✓76.6264.6653.6250.3361.31
✓76.1964.7554.3950.6761.50
✓✓77.5865.3055.0151.1262.25
+ +Table 1. Ablation study for two key modules: TDCL and NDCR. + +
ScenariosL=1L=2L=3L=4Fused△
NID10.0580.2182.6082.1981.3383.14+0.54
NID10.286.1587.2887.8185.9288.27+0.46
NID10.587.1089.0388.9387.1789.31+0.28
NID274.2675.7375.6175.0076.45+0.72
+ +Table 2. Comparison of features from different layers $L$ and fused features $\left( {L \in \{ 2,3,4\} }\right)$ in Eq. (8) and Eq. (9) to obtain ${\widetilde{\mathcal{F}}}_{i}$ and ${\widetilde{\mathcal{H}}}_{i}$ , best in bold and second with underline, $\bigtriangleup$ means accuracy gap. + +![](images/517db9bf2184c3a176aea6e8b9629536e038b6f96eba6ac97f1de245771dfd26.jpg) +Figure 6. Analysis of FedDifRC with time-steps $t$ in Eq. (9) (left) and temperature $\tau$ in Eq. (11) (right) on CIFAR10 with NID10.2. + +![](images/ebd4fbe975aeb5c865726fdcd36c6ab2d7f042b84bcff6ce3b41d9248560b9ac.jpg) + +# 5. Experiments + +# 5.1. Experimental Setup + +Datasets: We adopt popular benchmark datasets, including CIFAR10 & CIFAR100 [20], TinyImageNet [21]; we build unbalanced versions of these datasets by [15]: CIFAR10-LT & CIFAR100-LT, and TinyImageNet-LT; we use Digits [41] and Office-Caltech [19] for the domain shift task in FL. + +Scenarios: i) label shift: we consider two non-iid settings: $\mathrm{NID1}_{\alpha}$ follows Dirichlet distribution [59], where $\alpha$ denotes the heterogeneity level; NID2 is a more extreme setting consists of 6 biased clients (each has only a single category) and 1 unbiased client has all categories. ii) imbalance shift: we shape the original dataset into a long-tailed distribution with $\rho$ follow [15], $\rho$ means the ratio between sample sizes of the most frequent and least frequent class, as $\rho = \max_{j} / \min_{j}$ . iii) domain shift: Digits [41] includes 4 domains as MNIST (M), USPS (U), SVHN (SV), SYN (SY) with 10 categories. Office-Caltech [19] also consists 4 domains as Caltech (C), + +Webcam (W), Amazon (A), DSLR (D) with 10 overlapping categories. We initialize 10 clients and assign them domains randomly: Digits (M: 1, U: 4, SV: 3, SY: 2), Office-Caltech (C: 3, W: 2, A: 2, D: 3). Each client's dataset is randomly selected from domains (Digits: $10\%$ , Office-Caltech: $20\%$ ). + +Configurations: The global rounds $R$ as 100 and clients $K$ as 10, all methods have little or no gain with more rounds $R$ . To facilitate fair comparison, we follow in [26, 29], and use SGD optimizer to update models with a learning rate 0.01, a momentum of 0.9, a batch size of 64, the weight decay of $10^{-5}$ , and local epochs $E = 10$ . We use MobileNetV2 [48] for TinyImageNet, ResNet-10 [14] for others. The dimension $d$ as 512, $\tau$ (Eq. (11)) as 0.06, $t$ (Eq. (9)) as 150. + +Baselines: We compare our against several SOTA methods: standard-based FedAvg [34], FedProx [29]; contrast-based MOON [26], FedRCL [50]; aggregate-based FedCDA [60], FedDisco [69]; unsupervised FedEMA [77], FedLID [42]; prototype-based FedProto [52], FedNH [70]. + +# 5.2. Validation Analysis + +We give ablation study in Table. 1: i) TDCL brings significant performance improvements over the baseline, showing that TDCL is able to promote semantic spread-out property; ii) NDCR also yields solid gains, proving the importance of aligning instance embedding via consistency regularization; iii) combining TDCL and NDCR attain better performance, supporting our motivation to unleash meaningful guidance of diffusion representations to promote FL local training. + +Then, we prove the advantage of fused features (Eq. (8)). As shown in Table. 2, results on CIFAR10 reveal that fused features capture both semantics and local details, better than features on a single layer $L$ , and if the features with too low resolution ( $L = 1$ ) or over-matching ( $L = 4$ ) would damage the model's performance. To explore the influence of different time-steps $t$ (Eq. (9)), we conduct experiments on CIFAR10 with NID10.2 in Fig. 6 left, we add different amount of noise to get $\tilde{x}_i^t$ in Eq. (9). We find that FedDifRC is robust to the choice of $t \in [100, 300]$ ( $t = 150$ achieves the highest value). As shown in Fig. 6 right, we present results with different $\tau$ in Eq. (11), which reveals that FedDifRC is not sensitive to the $\tau \in [0.04, 0.08]$ , indicating the effectiveness of $\mathcal{U}$ (Eq. (10)), to balance relative distances among different samples. The accuracy no longer increases after $\tau = 0.06$ . Extra results of $t$ and $\tau$ are provided in Appendix C.2.5. + +# 5.3. Evaluation Results + +The results on different settings (NID1, NID2, $\rho$ ) in Table. 3, FedDifRC consistently outperforms these counterparts, it confirms that FedDifRC effectively leverages the inherent fine-grained visual representation capabilities of diffusion models to promote heterogeneous FL training. The results of Office-Caltech with domain shift are shown in Table. 4, indicating that FedDifRC can acquire well-generalizable + +
FL BaselinesCIFAR10CIFAR100TinyImageNetCIFAR10-LT (keep NID10.2)
NID10.05NID10.2NID2NID10.05NID10.2NID2NID10.05NID10.2NID2ρ = 10ρ = 50ρ = 100ρ = 200
FedAvg[AISTAT'17]78.2784.6572.6055.9760.0850.5640.4142.8435.1575.9463.2552.9749.69
FedProx[MLSys'20]78.4284.5972.8156.2760.2150.2940.2042.1635.6276.0263.4352.8649.10
MOON[CVPR'21]80.7986.1073.3556.7961.4851.8140.7943.6336.1175.7363.8053.1749.31
FedProto[AAAI'22]77.8683.9070.8454.4859.2648.6039.2941.0234.6874.3162.2350.3747.26
FedNH[AAAI'23]80.2585.8273.6556.6861.4551.3140.2543.3036.2176.1563.3652.4948.92
FedDisco[ICML'23]81.3586.6374.7257.3461.7952.0540.8343.9237.6976.2764.4053.3349.75
FedCDA[ICLR'24]81.7086.8975.0757.7662.1752.3541.1844.0938.2476.4564.3053.8550.08
FedRCL[CVPR'24]82.0287.1175.1357.9862.0552.4941.3344.2138.8676.3464.1753.4149.91
FedDifRC83.1488.2776.4559.2263.3253.8742.6345.2840.0577.5865.3055.0151.12
+ +Table 3. Comparison results on several settings (label shift and imbalance shift) and datasets, the best in bold and second with underline. + +
FL BaselinesOffice-Caltech
CaltechWebcamAmazonDSLRAVG
FedDisco[ICML'23]62.3945.9577.3145.2057.71
FedCDA[ICLR'24]62.2546.5376.5846.1457.87
FedRCL[CVPR'24]60.1342.7675.6043.7755.56
FedDifRC63.3250.1978.6748.1260.08
+ +Table 4. Comparison results on Office-Caltech with domain shift. + +
FL BaselinesClientsCIFAR100
NID10.05NID10.2NID10.5NID2
FedEMA[ICLR'22]K=1026.9029.4130.6421.52
FedLID[ICCV'23]K=1028.8430.4331.2223.65
FedDifRCK=1030.6131.4532.3025.73
+ +Table 5. Comparison results on self-supervised setting (Sec. 4.3). + +
FL BaselinesCIFAR10CIFAR10-LT (NID102)
NID10.05NID10.2NID2ρ = 10ρ = 50ρ = 100
FedDDA[PRCV-23]82.0586.9474.4976.3163.9553.15
FedDISC[AAAF24]82.3387.1575.4276.6064.2354.07
FedRCL (♂/†)82.30♂/80.95†87.25♂/86.16†75.37♂/73.85†76.50♂/76.12†64.26♂/62.91†53.65♂/52.20†
FedRCL (♀ + †)81.5286.4874.3576.2263.6952.93
FedDifRC83.1488.2776.4577.5865.3055.01
+ +Table 6. More comparison results on CIFAR10 and CIFAR10-LT. + +
FL BaselinesHAM10000 [56]EuroSAT [40]
NID10.2NID10.5NID2NID10.2NID10.5NID2
FedDisco[ICML'23]50.2653.6844.6378.7682.3670.34
FedCDA[ICLR'24]51.6554.0545.1680.4583.3871.10
FedRCL[CVPR'24]52.2154.2045.4879.8083.6970.73
FedDifRC53.7955.6047.1782.4986.0573.14
+ +property in the feature space. With self-supervised scenario, in Table. 5, our FedDifRC achieves excellent performance by leveraging the visual perception capability of diffusion models, without requiring any labeled data. Please refer to Appendix C for more details and results (e.g., domain shift, convergence rates, different SD versions, etc). + +Table 7. Comparison results on the HAM10000 and EuroSAT. + +
Pre-trained ModelsCIFAR10CIFAR10-LT (NID10.2)
NID10.05NID10.2NID2ρ = 10ρ = 50ρ = 100
CLIP-ViT-B/16 [45]79.7284.9173.9774.8962.4450.87
MAE-ViT-B/16 [10]80.7485.3873.2475.4563.2752.55
DINO-ViT-B/16 [33]80.5786.1474.1675.2562.7051.92
FedDifRC83.1488.2776.4577.5865.3055.01
+ +Table 8. Experiments on aligning features of pre-trained models. + +To further elaborate on the core idea of our FedDifRC, we then perform the following experiments: + +i) To ensure a fair comparison, we also compare diffusion-based FL methods FedDISC [67] and FedDDA [74], we generate data for each client (same amount) by the SD model as $\diamond$ , and align SD features (the image is fed directly into the SD model) by L2 loss as $\dagger$ . Table. 6 shows that directly matching SD features may even hurt the performance, which further illustrates the superiority of our FedDifRC. ii) We focus on visual perception ability of diffusion models to understand an image (e.g., textures and structures) rather than pre-trained prior knowledge. We conduct experiments on medical and remote sensing datasets HAM10000 [56] and EuroSAT [40], which are not crossed in SD pre-training dataset LAION-5B [49]. Table. 7 show that the success of FedDifRC is based on the fine-grained visual representation rather than the large prior knowledge in SD model. iii) The SD model can implicitly provide both high-level and low-level visual concepts of semantic objects, while foundation models focus more on sparse visual matching. The Table. 8 show that FedDifRC can capture underlying properties and correspondences of visual objects. + +# 6. Conclusion + +In this paper, we aim to exploit the fine-grained visual representation capabilities of diffusion models to handle data heterogeneity in FL. We propose a novel diffusion-inspired FL framework FedDifRC to regulate local training via denoising and conditional diffusion representations. By leveraging complementary strength of TDCL and NDCR, FedDifRC achieves superior performance in heterogeneous FL. + +# Acknowledgment + +This work is supported by the scholarship from the China Scholarship Council (CSC) while the first author pursues his PhD degree at the University of Wollongong. This work was also partially supported by Australian Research Council Linkage Project LP210300009 and LP230100083. + +# References + +[1] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, pages 1597-1607, 2020. 2, 3, 5, 6 +[2] Ching-Yao Chuang, Joshua Robinson, Yen-Chen Lin, Antonio Torralba, et al. Debiased contrastive learning. In NeurIPS, pages 8765-8775, 2020. 3 +[3] Kevin Clark and Priyank Jain. Text-to-image diffusion models are zero shot classifiers. In NeurIPS, 2023. 3 +[4] Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah. Diffusion models in vision: A survey. IEEE TPAMI, 45(9):10850-10869, 2023. 2, 3 +[5] Jiequan Cui, Zhisheng Zhong, Shu Liu, Bei Yu, et al. Parametric contrastive learning. In ICCV, pages 715-724, 2021. 3 +[6] Liang Gao, Huazhu Fu, Li Li, Yingwen Chen, et al. Feddc: Federated learning with non-iid data via local drift decoupling and correction. In CVPR, pages 10112-10121, 2022, 2, 6 +[7] Jie Gui, Tuo Chen, Jing Zhang, Qiong Cao, et al. A survey on self-supervised learning: Algorithms, applications, and future trends. IEEE TPAMI, 46(12):9052-9071, 2024. 3 +[8] Sungwon Han, Sungwon Park, Fangzhao Wu, Sundong Kim, et al. Fedx: Unsupervised federated learning with cross knowledge distillation. In ECCV, pages 691-707, 2022. 3 +[9] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, et al. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9729-9738, 2020. 2, 3, 5 +[10] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, et al. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 8 +[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, pages 6840-6851, 2020. 2, 3 +[12] Wenke Huang, Mang Ye, and Bo Du. Learn from others and be yourself in heterogeneous federated learning. In CVPR, pages 10143-10153, 2022. 2, 3 +[13] Wenke Huang, Mang Ye, Zekun Shi, He Li, and Bo Du. Rethinking federated learning with domain shift: A prototype view. In CVPR, pages 16312-16322, 2023. 2 +[14] Cheng Ju, Aurélien Bibaut, and Mark van der Laan. The relative performance of ensemble methods with deep convolutional neural networks for image classification. Journal of Applied Statistics, 45:2800-2818, 2018. 7 +[15] Cao Kaidi, Colin Wei, Adrien Gaidon, Nikos Arechiga, et al. Learning imbalanced datasets with label-distribution-aware margin loss. In NeurIPS, 2019. 7 + +[16] Peter Kairouz, H Brendan McMahan, Brendan Avent, Aurélien Bellet, et al. Advances and open problems in federated learning. Foundations and Trends in Machine Learning, 14(1):1-210, 2021. 1 +[17] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank Reddi, et al. Scaffold: Stochastic controlled averaging for federated learning. In ICML, pages 5132-5143, 2020. 2 +[18] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, et al. Supervised contrastive learning. In NeurIPS, pages 18661-18673, 2020. 3 +[19] Boqing Kong, Yuan Shi, Fei Sha, and Kristen Grauman. Geodesic flow kernel for unsupervised domain adaptation. In CVPR, pages 2066-2073, 2012. 7 +[20] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. In Technical report, University of Toronto, 2009. 6, 7 +[21] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7, 2015. 6, 7 +[22] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 4 +[23] Sunwoo Lee, Tuo Zhang, and A Salman Avestimehr. Layerwise adaptive model aggregation for scalable federated learning. In AAAI, pages 8491-8499, 2023. 2 +[24] Alexander C Li, Mihir Prabhudesai, Shivam Duggal, Ellis Brown, et al. Your diffusion model is secretly a zero-shot classifier. In ICCV, pages 2206-2217, 2023. 3 +[25] Daixun Li, Weiying Xie, Zixuan Wang, Yibing Lu, et al. Feddiff: Diffusion model driven federated learning for multimodal and multi-clients. IEEE TCSVT, 34(10):10353-10367, 2024. 3 +[26] Qinbin Li, Bingsheng He, and Dawn Song. Model-contrastive federated learning. In CVPR, pages 10713-10722, 2021. 2, 3, 5, 7 +[27] Qinbin Li, Yiqun Diao, Quan Chen, and Bingsheng He. Federated learning on non-iid data silos: An experimental study. In ICDE, pages 965–978, 2022. 1 +[28] Tian Li, Anit Kumar Sahu, Ameet Talwalkar, and Virginia Smith. Federated learning: Challenges, methods, and future directions. IEEE SPM, 37(3):50-60, 2020. 1 +[29] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, et al. Federated optimization in heterogeneous networks. MLSys, 2:429-450, 2020. 1, 2, 6, 7 +[30] Xiang Li, Kaixuan Huang, Wenhao Yang, Shusen Wang, et al. On the convergence of fedavg on non-iid data. In ICLR, 2020. 2, 6 +[31] Grace Luo, Lisa Dunlap, Dong Huk Park, Aleksander Holynski, et al. Diffusion hyperfeatures: Searching through time and space for semantic correspondence. In NeurIPS, 2023. 3 +[32] Mi Luo, Fei Chen, Dapeng Hu, Yifan Zhang, et al. No fear of heterogeneity: Classifier calibration for federated learning with non-iid data. In NeurIPS, pages 5972-5984, 2021. 1 +[33] Caron Mathilde, Hugo Touvron, Ishan Misra, Hervé Jégou, et al. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. 8 + +[34] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, et al. Communication-efficient learning of deep networks from decentralized data. In AISTATS, pages 1273-1282, 2017. 1, 2, 7 +[35] Matias Mendieta, Taojiannan Yang, Pu Wang, Minwoo Lee, et al. Local learning matters: Rethinking data heterogeneity in federated learning. In CVPR, pages 8397-8406, 2022. 1, 2 +[36] John Nguyen, Kshitiz Malik, Hongyuan Zhan, Ashkan Yousefpour, et al. Federated learning with buffered asynchronous aggregation. In AISTATS, pages 3581-3607, 2022. 2 +[37] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, pages 8162-8171, 2021. 2, 3 +[38] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. In NeurIPS, 2019. 3 +[39] Reese Pathak and Martin J Wainwright. Fedsplit: An algorithmic framework for fast federated optimization. In NeurIPS, pages 7057-7066, 2020. 2 +[40] Helber Patrick, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE J-STARS, 12(7):2217-2226, 2019. 8 +[41] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, et al. Moment matching for multi-source domain adaptation. In ICCV, pages 1406-1415, 2019. 7 +[42] Athanasios Psaltis, Anestis Kastellos, Charalampos Z Patrikakis, and Petros Daras. Fedlid: Self-supervised federated learning for leveraging limited image data. In ICCV, pages 1039-1048, 2023. 7 +[43] Pian Qi, Diletta Chiaro, Antonella Guzzo, Michele Ianni, et al. Model aggregation techniques in federated learning: A comprehensive survey. Future Generation Computer Systems, 150:272-293, 2024. 2 +[44] Liangqiong Qu, Yuyin Zhou, Paul Pu Liang, Yingda Xia, et al. Rethinking architecture design for tackling data heterogeneity in federated learning. In CVPR, pages 10061-10071, 2022. 1 +[45] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 8 +[46] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, et al. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2, 3, 5 +[47] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234-241, 2015. 2, 3, 4 +[48] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhominov, et al. Mobilenetv2: Inverted residuals and linear bottlenecks. In CVPR, pages 4510-4520, 2018. 7 +[49] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, pages 25278–25294, 2022. 8 + +[50] Seonguk Seo, Jinkyu Kim, Geeho Kim, and Bohyung Han. Relaxed contrastive learning for federated learning. In CVPR, pages 12279-12288, 2024. 7 +[51] Fiona Victoria Stanley Jothiraj and Afra Mashhadi. Phoenix: A federated generative diffusion model. In ACM WWW, pages 1568-1577, 2024. 3 +[52] Yue Tan, Guodong Long, Lu Liu, Tianyi Zhou, et al. FedProto: Federated prototype learning across heterogeneous clients. In AAAI, pages 8432-8440, 2022. 2, 7 +[53] Yue Tan, Guodong Long, Jie Ma, Lu Liu, et al. Federated learning from pre-trained models: A contrastive learning approach. In NeurIPS, pages 19332-19344, 2022. 3 +[54] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, et al. Emergent correspondence from image diffusion. In NeurIPS, pages 1363-1389, 2023. 3 +[55] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, et al. What makes for good views for contrastive learning? In NeurIPS, pages 6827-6839, 2020. 3 +[56] Philipp Tschandl, Cliff Rosendahl, and Harald Kittler. The ham10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. Nature Scientific Data, 5(1):1-9, 2018. 8 +[57] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9 (11):2579-2605, 2008. 4 +[58] Feng Wang and Huaping Liu. Understanding the behaviour of contrastive loss. In CVPR, pages 2495-2504, 2021. 2 +[59] Hongyi Wang, Mikhail Yurochkin, Yuekai Sun, Dimitris Papailiopoulos, et al. Federated learning with matched averaging. In ICLR, 2020. 7 +[60] Haozhao Wang, Haoran Xu, Yichen Li, Yuan Xu, et al. Fed-cda: Federated learning with cross-rounds divergence-aware aggregation. In ICLR, 2024. 2, 7 +[61] Yingheng Wang, Yair Schiff, Aaron Gokaslan, Weishen Pan, et al. Infodiffusion: Representation learning using information maximizing diffusion models. In ICML, pages 36336-36354, 2023. 4 +[62] Yuan Wang, Huazhu Fu, Renuga Kanagavelu, Qingsong Wei, et al. An aggregation-free federated learning for tackling data heterogeneity. In CVPR, pages 26233-26242, 2024. 2 +[63] Chen Wei, Karttikeya Mangalam, Po-Yao Huang, Yanghao Li, et al. Diffusion models as masked autoencoders. In ICCV, pages 16284-16294, 2023. 3 +[64] Weilai Xiang, Hongyu Yang, Di Huang, and Yunhong Wang. Denoising diffusion autoencoders are unified self-supervised learners. In ICCV, pages 15802-15812, 2023. 3 +[65] Yikai Yan, Chaoyue Niu, Yucheng Ding, Zhenzhe Zheng, et al. Federated optimization under intermittent client availability. INFORMS Journal on Computing, 36(1):185-202, 2024. 2 +[66] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, et al. Diffusion models: A comprehensive survey of methods and applications. ACM Computing Surveys, 56(4):1-39, 2023. 3 +[67] Mingzhao Yang, Shangchao Su, Bin Li, and Xiangyang Xue. Exploring one-shot semi-supervised federated learning with pre-trained diffusion models. In AAAI, pages 16325-16333, 2024. 3, 8 + +[68] Xingyi Yang and Xinchao Wang. Diffusion model as representation learner. In ICCV, pages 18938-18949, 2023. 4 +[69] Rui Ye, Mingkai Xu, Jianyu Wang, Chenxin Xu, et al. Feddisco: Federated learning with discrepancy-aware collaboration. In ICML, pages 39879-39902, 2023. 2, 6, 7 +[70] Dai Yutong, Zeyuan Chen, Junnan Li, Shelby Heinecke, et al. Tackling data heterogeneity in federated learning with class prototypes. In AAAI, pages 7314-7322, 2023. 7 +[71] Daokun Zhang, Jie Yin, Xingquan Zhu, and Chengqi Zhang. Network representation learning: A survey. IEEE Transactions on Big Data, 6:3-28, 2018. 4 +[72] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, et al. A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence. In NeurIPS, 2023. 3 +[73] Jianqing Zhang, Yang Hua, Hao Wang, Tao Song, et al. Fedala: Adaptive local aggregation for personalized federated learning. In AAAI, pages 11237-11244, 2023. 2 +[74] Zhuang Zhao, Feng Yang, and Guirong Liang. Federated learning based on diffusion model to cope with non-iid data. In PRCV, pages 220-231, 2023. 3, 8 +[75] Hangyu Zhu, Jinjin Xu, Shiqing Liu, and Yaochu Jin. Federated learning on non-iid data: A survey. Elsevier Neurocomputing, 465:371-390, 2021. 1 +[76] Fuzhen Zhuang, Xiaohu Cheng, Ping Luo, Sinno Jialin Pan, et al. Supervised representation learning: Transfer learning with deep autoencoders. In IJCAI, 2015. 4 +[77] Weiming Zhuang, Yonggang Wen, and Shuai Zhang. Divergence-aware federated self-supervised learning. In ICLR, 2022. 7 \ No newline at end of file diff --git a/2025/FedDifRC_ Unlocking the Potential of Text-to-Image Diffusion Models in Heterogeneous Federated Learning/images.zip b/2025/FedDifRC_ Unlocking the Potential of Text-to-Image Diffusion Models in Heterogeneous Federated Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5668d400503bc19b8af2afd8c540ab7079cd47b0 --- /dev/null +++ b/2025/FedDifRC_ Unlocking the Potential of Text-to-Image Diffusion Models in Heterogeneous Federated Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c322d6a76220710418f26fd6011acda5eac87d4202e0643105a0bc8987c9c92c +size 642738 diff --git a/2025/FedDifRC_ Unlocking the Potential of Text-to-Image Diffusion Models in Heterogeneous Federated Learning/layout.json b/2025/FedDifRC_ Unlocking the Potential of Text-to-Image Diffusion Models in Heterogeneous Federated Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..842a20e5c460dd314d6570ba6ddad0610ec7e9a1 --- /dev/null +++ b/2025/FedDifRC_ Unlocking the Potential of Text-to-Image Diffusion Models in Heterogeneous Federated Learning/layout.json @@ -0,0 +1,15801 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 82, + 102, + 531, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 102, + 531, + 140 + ], + "spans": [ + { + "bbox": [ + 82, + 102, + 531, + 140 + ], + "type": "text", + "content": "FedDifRC: Unlocking the Potential of Text-to-Image Diffusion Models in Heterogeneous Federated Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 160, + 565, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 160, + 565, + 233 + ], + "spans": [ + { + "bbox": [ + 62, + 160, + 565, + 233 + ], + "type": "text", + "content": "Huan Wang" + }, + { + "bbox": [ + 62, + 160, + 565, + 233 + ], + "type": "inline_equation", + "content": "^{1,\\ast}" + }, + { + "bbox": [ + 62, + 160, + 565, + 233 + ], + "type": "text", + "content": ", Haoran Li" + }, + { + "bbox": [ + 62, + 160, + 565, + 233 + ], + "type": "inline_equation", + "content": "^{1,3,\\ast}" + }, + { + "bbox": [ + 62, + 160, + 565, + 233 + ], + "type": "text", + "content": ", Huaming Chen" + }, + { + "bbox": [ + 62, + 160, + 565, + 233 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 62, + 160, + 565, + 233 + ], + "type": "text", + "content": ", Jun Yan" + }, + { + "bbox": [ + 62, + 160, + 565, + 233 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 160, + 565, + 233 + ], + "type": "text", + "content": ", Jiahua Shi" + }, + { + "bbox": [ + 62, + 160, + 565, + 233 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 62, + 160, + 565, + 233 + ], + "type": "text", + "content": ", Jun Shen" + }, + { + "bbox": [ + 62, + 160, + 565, + 233 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + }, + { + "bbox": [ + 62, + 160, + 565, + 233 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 160, + 565, + 233 + ], + "type": "text", + "content": "School of Computing and Information Technology, University of Wollongong, Wollongong, Australia \n" + }, + { + "bbox": [ + 62, + 160, + 565, + 233 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 62, + 160, + 565, + 233 + ], + "type": "text", + "content": "School of Electrical and Computer Engineering, The University of Sydney, Sydney, Australia \n" + }, + { + "bbox": [ + 62, + 160, + 565, + 233 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 62, + 160, + 565, + 233 + ], + "type": "text", + "content": "QLD Alliance for Agriculture and Food Innovation, The University of Queensland, Brisbane, Australia \nhw226@uowmail.edu.au *Equal Contribution †Corresponding Author" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 285, + 297, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 285, + 297, + 632 + ], + "spans": [ + { + "bbox": [ + 54, + 285, + 297, + 632 + ], + "type": "text", + "content": "Federated learning aims at training models collaboratively across participants while protecting privacy. However, one major challenge for this paradigm is the data heterogeneity issue, where biased data preferences across multiple clients, harming the model's convergence and performance. In this paper, we first introduce powerful diffusion models into the federated learning paradigm and show that diffusion representations are effective steers during federated training. To explore the possibility of using diffusion representations in handling data heterogeneity, we propose a novel diffusion-inspired Federated paradigm with Diffusion Representation Collaboration, termed FedDifRC, leveraging meaningful guidance of diffusion models to mitigate data heterogeneity. The key idea is to construct text-driven diffusion contrasting and noise-driven diffusion regularization, aiming to provide abundant class-related semantic information and consistent convergence signals. On the one hand, we exploit the conditional feedback from the diffusion model for different text prompts to build a text-driven contrastive learning strategy. On the other hand, we introduce a noise-driven consistency regularization to align local instances with diffusion denoising representations, constraining the optimization region in the feature space. In addition, FedDifRC can be extended to a self-supervised scheme without relying on any labeled data. We also provide a theoretical analysis for FedDifRC to ensure convergence under non-convex objectives. The experiments on different scenarios validate the effectiveness of FedDifRC and the efficiency of crucial components. Code is available at https://github.com/hwang52/FedDifRC." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 656, + 137, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 656, + 137, + 669 + ], + "spans": [ + { + "bbox": [ + 55, + 656, + 137, + 669 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Federated Learning (FL) [34] has become increasingly popular as a decentralized training paradigm, with multiple distributed clients train their local models without sharing their" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 313, + 260, + 555, + 357 + ], + "blocks": [ + { + "bbox": [ + 313, + 260, + 555, + 357 + ], + "lines": [ + { + "bbox": [ + 313, + 260, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 555, + 357 + ], + "type": "image", + "image_path": "4daf08b7f68934a45627e2d5ccb5014341fe08e7eb4fa784852b91c48d022de6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 360, + 534, + 372 + ], + "lines": [ + { + "bbox": [ + 333, + 360, + 534, + 372 + ], + "spans": [ + { + "bbox": [ + 333, + 360, + 534, + 372 + ], + "type": "text", + "content": "(a) Text-Driven Diffusion Contrastive Learning (TDCL)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 313, + 376, + 554, + 479 + ], + "blocks": [ + { + "bbox": [ + 313, + 376, + 554, + 479 + ], + "lines": [ + { + "bbox": [ + 313, + 376, + 554, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 554, + 479 + ], + "type": "image", + "image_path": "b3b5bcf8dc4bfda73e7eed4c30ae8896bfba56f425d6305ad06f91e081f9583e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 482, + 541, + 494 + ], + "lines": [ + { + "bbox": [ + 331, + 482, + 541, + 494 + ], + "spans": [ + { + "bbox": [ + 331, + 482, + 541, + 494 + ], + "type": "text", + "content": "(b) Noise-Driven Diffusion Consistent Restriction (NDCR)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 501, + 556, + 590 + ], + "lines": [ + { + "bbox": [ + 313, + 501, + 556, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 501, + 556, + 590 + ], + "type": "text", + "content": "Figure 1. The main idea of FedDifRC with two key modules TDCL and NDCR. (a) We leverage conditional generative feedback of the pre-trained diffusion model on various text prompts to devise a text-driven contrastive learning strategy, where the input embedding is injected as a condition. (b) We extract the diffusion denoising feature maps of intermediate layers from the pre-trained diffusion model at a specific time step " + }, + { + "bbox": [ + 313, + 501, + 556, + 590 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 501, + 556, + 590 + ], + "type": "text", + "content": " during the diffusion backward process to construct a noise-driven regularization." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "private dataset [16, 28] (i.e., training client models locally and aggregating them globally [29]). One major challenge faced in FL is the potential discrepancies in the local data distributions across distributed clients, which is known as the data heterogeneity issue [32, 35, 44]. Specifically, the private data is collected from different sources with diverse preferences and presents non-independent and identically distributed (non-iid) distribution [27, 75]. Such non-iid data" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3726" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "discrepancies can lead to inconsistency in the clients' local objective functions and optimization directions, making the model convergence slow and unstable [17, 29, 30, 39, 65]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 110, + 295, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 110, + 295, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 110, + 295, + 289 + ], + "type": "text", + "content": "To mitigate the negative impact of the data heterogeneity, a mainstream of subsequent efforts delves into introducing a variety of suited signals to regulate the private model, there are mainly two perspectives: i) local model optimization at the client side and ii) global model aggregation at the server side. The former focuses on optimization strategies to make the variance of local updates limited [6, 29]. Another aims to improve the efficacy of global model aggregation [36, 69, 73]. In fact, most of these previous methods force the local models to be consistent to the global model. Although they have a certain effect by reducing gradient inconsistency, the gradually enlarged parameter deviation persists. Hence, we argue that local model overfits the local data distribution, which is magnified under the data heterogeneity and thus hinders the improvement of model's performance in FL." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 291, + 295, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 291, + 295, + 459 + ], + "spans": [ + { + "bbox": [ + 55, + 291, + 295, + 459 + ], + "type": "text", + "content": "Recently, the development of diffusion models (DMs) [4, 11, 37, 46] provides fresh opportunities. The diffusion models have demonstrated remarkable power in generating the diverse and realistic images, showing strong ability to capture intricate details and patterns from underlying data. For instance, with proper guidance, stable diffusion (SD) [46] models can effectively align a text prompt with the corresponding image context, which suggests their potential ability to understand both high-level and low-level contextual visual concepts regarding what an image contains. This semantic generative and visual comprehension capabilities of SD models motivate us to think: is it possible to exploit the inherent fine-grained visual representation capabilities of diffusion models to promote heterogeneous FL training?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 460, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 687 + ], + "type": "text", + "content": "Taking into account both the effectiveness and efficiency, we revisit the underlying representations in the SD models, which are intermediate feature maps in a diffusion process (e.g., denoise or noise an image) produced by multiple calling a UNet [47]. We argue that diffusion representations are effective steers for FL training: i) extensive general knowledge inherent in diffusion models enhance local semantic diversity in FL; ii) smooth correspondences about semantic objects in diffusion models is a natural guiding signal in FL. Building on these two insights, we first propose conditional diffusion representations by imposing various text prompts in the SD model. In this way, each instance is contrastively enhanced with a set of text-driven diffusion representations, capturing rich domain variance. Then, we devise denoising diffusion representations from semantic relations between visual and linguistic concepts in the SD model. We extract the feature maps from different hierarchies of the UNet decoder in the SD model's denoising process, serving as guidance signals to avoid over-optimization in FL local training." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "In this paper, we propose a diffusion-inspired Federated framework with Diffusion Representation Collaboration, as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 335 + ], + "type": "text", + "content": "FedDifRC, to handle the data heterogeneity issue in FL, which consists of two key modules (see Fig. 1): i) First, we devise Text-driven Diffusion Contrastive Learning (TDCL), which utilizes conditional diffusion representations to construct inter-contrastive learning [1, 9, 26, 58] in a text-driven diffusion manner (the sample's embedding is injected as a condition). By maximizing the agreement between the local instance and its corresponding representations, TDCL adaptively captures class-relevant information and semantically meaningful knowledge. This objective encourages sample's embedding closer to conditional diffusion representations with same semantics and away from other negatives, thus maintaining a clear decision boundary. ii) Second, we use the denoising diffusion representations to build a flat and stable convergent target and further propose Noise-driven Diffusion Consistent Restriction (NDCR). Concretely, we extract denoising feature maps of the noisy input sample in the SD model's backward process, as a regularization term to constrain the optimization region of the local model. The local sample is forced to align the embedding with its denoising representations at the feature space to promote the uniformity for local training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 326, + 336, + 534, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 336, + 534, + 347 + ], + "spans": [ + { + "bbox": [ + 326, + 336, + 534, + 347 + ], + "type": "text", + "content": "Our main contributions are summarized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 350, + 553, + 529 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 350, + 553, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 350, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 314, + 350, + 553, + 396 + ], + "type": "text", + "content": "- We explore the possibility of using diffusion models to tackle data heterogeneity in FL and devise the conditional and denoising diffusion representations as effective steers to regulate model training in heterogeneous FL." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 397, + 553, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 397, + 553, + 468 + ], + "spans": [ + { + "bbox": [ + 314, + 397, + 553, + 468 + ], + "type": "text", + "content": "- We propose FedDifRC, a diffusion-inspired FL framework to handle data heterogeneity by leveraging the complementary advantages of two key components: i) TDCL facilitates the local model to learn more general class-relevant knowledge; ii) NDCR enforces the uniformity in the feature space for local training." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 468, + 553, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 468, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 314, + 468, + 553, + 529 + ], + "type": "text", + "content": "- We conduct extensive experiments for different FL scenarios (e.g., label shift, domain shift, and long-tailed) and different application scenarios (i.e., intra-domain and cross-domain) to demonstrate the effectiveness of our FedDifRC and the indispensability of each component." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 542, + 400, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 542, + 400, + 555 + ], + "spans": [ + { + "bbox": [ + 314, + 542, + 400, + 555 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 563, + 502, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 502, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 502, + 576 + ], + "type": "text", + "content": "2.1. Heterogeneous Federated Learning" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "text", + "content": "Federated learning is proposed to address privacy concerns in a distributed learning environment (e.g., Federated Averaging (FedAvg) [34]). Several studies have improved the local update process on the client side or the global aggregation process on the server side to tackle data heterogeneity in FL. The former focuses on local optimization strategies to make the diversity between the client model and the global model limited in the parameter level [6, 12, 13, 17, 26, 29, 35, 52]. The another branch aims to adopt aggregation mechanisms on the server to alleviate the negative influence of data heterogeneity [23, 43, 60, 62, 69, 73]." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3727" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": "Recently, existing FL works that apply diffusion models mainly focus on building synthetic datasets [25, 51, 67, 74]. With the assistance of synthetic datasets, the FL global model can be retrained by this data augmentation scheme to mitigate data heterogeneity. Although these generated data have comparable quality and diversity to client images, the local model may overfit local domain distribution, and the inherent heterogeneity across clients may not have been addressed. In contrast, we focus on how to effectively leverage the fine-grained visual representation from diffusion models to promote a generalizable FL global model." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 220, + 233, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 220, + 233, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 233, + 232 + ], + "type": "text", + "content": "2.2. Diffusion Model Representations" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 240, + 296, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 240, + 296, + 408 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 296, + 408 + ], + "type": "text", + "content": "Recently, diffusion models have shown significant advances in image generation, editing, and stylized [4, 11, 37, 46, 66], suggesting that they contain rich underlying representations to be exploited for different downstream tasks. DIFT [54] is used to extract the diffusion features without explicit supervision. Diffusion Hyperfeatures [31] integrates multi-scale and multi-timesteps diffusion features into feature descriptors. The diffusion representations have shown impressive results on the vision branch, including being testified as certifiable classifiers [3, 24, 63] and semantics correspondence learners [54, 64, 72]. In this paper, we expand the diffusion models to FL and construct the conditional and denoising diffusion representations, offering fruitful class-relevant information and consistent convergence signals." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 423, + 180, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 180, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 180, + 437 + ], + "type": "text", + "content": "2.3. Contrastive Learning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 444, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 444, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 444, + 296, + 601 + ], + "type": "text", + "content": "Contrastive learning (CL) has been widely applied to self-supervised learning scenarios [1, 7, 9]. Many works are focused on learning the encoder where the embeddings of the same sample are pulled closer and those of different samples are pushed apart [2, 5, 18, 55]. They mainly construct a positive pair and some negative pairs for each data instance and use InfoNCE [38] to contrast positiveness against negativity. There are some works that incorporate contrastive learning into FL to assist local training to achieve higher model performance [8, 12, 26, 53]. In our work, Text-driven Diffusion Contrastive Learning (TDCL) is devised to promote local model by attracting samples to positive representations while pushing away other negatives." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 618, + 291, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 291, + 633 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 291, + 633 + ], + "type": "text", + "content": "3. Diffusion Models Help Federated Learning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": "In this section, we begin with a brief overview of the general FL paradigm and the generative diffusion models (Sec. 3.1). Then, we analyze the effectiveness of diffusion representations and illustrate our core insights (Sec. 3.2). Finally, we will introduce how to extract the conditional and denoising diffusion representations in Sec. 3.3 and Sec. 3.4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 72, + 481, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 481, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 481, + 85 + ], + "type": "text", + "content": "3.1. Preliminaries and Background" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 89, + 555, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 555, + 198 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 555, + 198 + ], + "type": "text", + "content": "Federated Learning: We focus on the case with " + }, + { + "bbox": [ + 313, + 89, + 555, + 198 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 89, + 555, + 198 + ], + "type": "text", + "content": " clients holding the heterogeneous data partition " + }, + { + "bbox": [ + 313, + 89, + 555, + 198 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{D}_1,\\mathcal{D}_2,\\ldots ,\\mathcal{D}_K\\}" + }, + { + "bbox": [ + 313, + 89, + 555, + 198 + ], + "type": "text", + "content": " with private dataset " + }, + { + "bbox": [ + 313, + 89, + 555, + 198 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_k = \\{x_i,y_i\\}_{i = 1}^{n_k},y_i\\in \\{1,\\dots ,|\\mathcal{C}|\\}" + }, + { + "bbox": [ + 313, + 89, + 555, + 198 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 313, + 89, + 555, + 198 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 89, + 555, + 198 + ], + "type": "text", + "content": " -th client, where " + }, + { + "bbox": [ + 313, + 89, + 555, + 198 + ], + "type": "inline_equation", + "content": "n_k" + }, + { + "bbox": [ + 313, + 89, + 555, + 198 + ], + "type": "text", + "content": " denotes the local dataset scale. Let " + }, + { + "bbox": [ + 313, + 89, + 555, + 198 + ], + "type": "inline_equation", + "content": "n_k^j" + }, + { + "bbox": [ + 313, + 89, + 555, + 198 + ], + "type": "text", + "content": " be the number of samples with label " + }, + { + "bbox": [ + 313, + 89, + 555, + 198 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 89, + 555, + 198 + ], + "type": "text", + "content": " at client " + }, + { + "bbox": [ + 313, + 89, + 555, + 198 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 89, + 555, + 198 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 89, + 555, + 198 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_k^j = \\{(x,y)\\in \\mathcal{D}_k|y = j\\}" + }, + { + "bbox": [ + 313, + 89, + 555, + 198 + ], + "type": "text", + "content": " denotes the set of samples with label " + }, + { + "bbox": [ + 313, + 89, + 555, + 198 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 89, + 555, + 198 + ], + "type": "text", + "content": " on the " + }, + { + "bbox": [ + 313, + 89, + 555, + 198 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 89, + 555, + 198 + ], + "type": "text", + "content": " -th client, so the number of samples in " + }, + { + "bbox": [ + 313, + 89, + 555, + 198 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 89, + 555, + 198 + ], + "type": "text", + "content": " -th class is " + }, + { + "bbox": [ + 313, + 89, + 555, + 198 + ], + "type": "inline_equation", + "content": "n^j = \\sum_{k = 1}^{K}n_k^j" + }, + { + "bbox": [ + 313, + 89, + 555, + 198 + ], + "type": "text", + "content": " . The objective of the general FL is to learn an optimal global model " + }, + { + "bbox": [ + 313, + 89, + 555, + 198 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 89, + 555, + 198 + ], + "type": "text", + "content": " across " + }, + { + "bbox": [ + 313, + 89, + 555, + 198 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 89, + 555, + 198 + ], + "type": "text", + "content": " clients:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 361, + 203, + 555, + 236 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 203, + 555, + 236 + ], + "spans": [ + { + "bbox": [ + 361, + 203, + 555, + 236 + ], + "type": "interline_equation", + "content": "\\min _ {w} \\mathcal {L} (w) = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\frac {n _ {k}}{N} \\mathcal {L} _ {k} (w; \\mathcal {D} _ {k}), \\tag {1}", + "image_path": "e1e550cda32dadf86e285b2a44ea8d9e867b235e2f593feca5d7f169766118d6.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_k" + }, + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "text", + "content": "-th client local loss function and " + }, + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "text", + "content": " is the global model parameters. " + }, + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "text", + "content": " is the total number of samples among all clients. In the FL training setup, we consider the local model " + }, + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "inline_equation", + "content": "w = \\{u,v\\}" + }, + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "text", + "content": ". It has two modules: i) a feature extractor " + }, + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "inline_equation", + "content": "h(u)" + }, + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "text", + "content": " with the parameters " + }, + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "text", + "content": " maps each sample " + }, + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "text", + "content": " to a " + }, + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "text", + "content": "-dim (default as 512) vector as " + }, + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "inline_equation", + "content": "z_{i} = h(u;x_{i})" + }, + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "text", + "content": "; ii) a classifier " + }, + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "inline_equation", + "content": "g(v)" + }, + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "text", + "content": " maps " + }, + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "text", + "content": " into a " + }, + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "inline_equation", + "content": "|\\mathcal{C}|" + }, + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "text", + "content": "-dim output logits as " + }, + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "inline_equation", + "content": "l_{i} = g(v;z_{i})" + }, + { + "bbox": [ + 313, + 239, + 555, + 335 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 336, + 554, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 554, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 554, + 395 + ], + "type": "text", + "content": "Diffusion Models: Given the image sample " + }, + { + "bbox": [ + 313, + 336, + 554, + 395 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 336, + 554, + 395 + ], + "type": "text", + "content": " from an underlying distribution " + }, + { + "bbox": [ + 313, + 336, + 554, + 395 + ], + "type": "inline_equation", + "content": "p(x)" + }, + { + "bbox": [ + 313, + 336, + 554, + 395 + ], + "type": "text", + "content": ", a forward diffusion process as a Markov chain to add random Gaussian noise " + }, + { + "bbox": [ + 313, + 336, + 554, + 395 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0, \\mathbf{I})" + }, + { + "bbox": [ + 313, + 336, + 554, + 395 + ], + "type": "text", + "content": " to the original sample " + }, + { + "bbox": [ + 313, + 336, + 554, + 395 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 336, + 554, + 395 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 336, + 554, + 395 + ], + "type": "inline_equation", + "content": "x_t" + }, + { + "bbox": [ + 313, + 336, + 554, + 395 + ], + "type": "text", + "content": " is the noise sample after adding " + }, + { + "bbox": [ + 313, + 336, + 554, + 395 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 336, + 554, + 395 + ], + "type": "text", + "content": "-steps noise:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 342, + 399, + 555, + 414 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 399, + 555, + 414 + ], + "spans": [ + { + "bbox": [ + 342, + 399, + 555, + 414 + ], + "type": "interline_equation", + "content": "x _ {t} = \\sqrt {1 - \\gamma_ {t}} x _ {t - 1} + \\sqrt {\\gamma_ {t}} \\epsilon_ {t}, t \\in \\{1, \\dots , T \\}, \\tag {2}", + "image_path": "7f361ad19dbc9b4d3b5968564b5811d5912c20cc803f02990e161b36542b3d93.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 418, + 554, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 554, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 554, + 502 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 418, + 554, + 502 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 418, + 554, + 502 + ], + "type": "text", + "content": " is the number of diffusion timesteps, " + }, + { + "bbox": [ + 313, + 418, + 554, + 502 + ], + "type": "inline_equation", + "content": "\\gamma_t \\in (0,1)" + }, + { + "bbox": [ + 313, + 418, + 554, + 502 + ], + "type": "text", + "content": " is a adjustable time variance schedule [11]. By leveraging the additive property of the Gaussian distribution, Eq. (2) can be further reformulated as " + }, + { + "bbox": [ + 313, + 418, + 554, + 502 + ], + "type": "inline_equation", + "content": "x_t = \\sqrt{\\bar{\\alpha}_t} x_0 + \\sqrt{1 - \\bar{\\alpha}_t} \\epsilon_t" + }, + { + "bbox": [ + 313, + 418, + 554, + 502 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 418, + 554, + 502 + ], + "type": "inline_equation", + "content": "\\alpha_t = 1 - \\gamma_t" + }, + { + "bbox": [ + 313, + 418, + 554, + 502 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 418, + 554, + 502 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t \\doteq \\prod_{i=1}^t \\alpha_i" + }, + { + "bbox": [ + 313, + 418, + 554, + 502 + ], + "type": "text", + "content": ". On this basis, the image " + }, + { + "bbox": [ + 313, + 418, + 554, + 502 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 418, + 554, + 502 + ], + "type": "text", + "content": " can be obtained from a random noise " + }, + { + "bbox": [ + 313, + 418, + 554, + 502 + ], + "type": "inline_equation", + "content": "x_t \\sim \\mathcal{N}(0,\\mathbf{I})" + }, + { + "bbox": [ + 313, + 418, + 554, + 502 + ], + "type": "text", + "content": " by reversing the above forward diffusion process [11]:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 318, + 506, + 554, + 532 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 506, + 554, + 532 + ], + "spans": [ + { + "bbox": [ + 318, + 506, + 554, + 532 + ], + "type": "interline_equation", + "content": "x _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(x _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} \\left(x _ {t}, t\\right)\\right), t \\in \\{T, \\dots , 1 \\}, \\tag {3}", + "image_path": "d249915ecb4cefa6cb03de1bf7398e743c749777fb74d0a9465d7d52073fb677.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 536, + 554, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 554, + 584 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 554, + 584 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 536, + 554, + 584 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 536, + 554, + 584 + ], + "type": "text", + "content": " denotes the denoiser network with the parameters " + }, + { + "bbox": [ + 313, + 536, + 554, + 584 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 536, + 554, + 584 + ], + "type": "text", + "content": " to predict noise " + }, + { + "bbox": [ + 313, + 536, + 554, + 584 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 536, + 554, + 584 + ], + "type": "text", + "content": ", and the denoiser network is usually implemented as a UNet architecture [47]. Further, the denoiser " + }, + { + "bbox": [ + 313, + 536, + 554, + 584 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 536, + 554, + 584 + ], + "type": "text", + "content": " is optimized using the loss " + }, + { + "bbox": [ + 313, + 536, + 554, + 584 + ], + "type": "inline_equation", + "content": "T_{\\theta}" + }, + { + "bbox": [ + 313, + 536, + 554, + 584 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 343, + 588, + 553, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 588, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 343, + 588, + 553, + 602 + ], + "type": "interline_equation", + "content": "\\mathcal {T} _ {\\theta} = \\mathbb {E} _ {\\epsilon , x _ {0}, t} [ \\| \\epsilon_ {\\theta} (x _ {t}, t) - \\epsilon \\| _ {2} ^ {2} ], \\epsilon \\sim \\mathcal {N} (0, \\mathbf {I}), \\tag {4}", + "image_path": "f3d00077116fcf9906275bf04c26f96a7f72a4df423ac7fd21b00eb24078a641.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 606, + 554, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 554, + 667 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 554, + 667 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 606, + 554, + 667 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 606, + 554, + 667 + ], + "type": "text", + "content": " is the noise in the forward diffusion process, and " + }, + { + "bbox": [ + 313, + 606, + 554, + 667 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 606, + 554, + 667 + ], + "type": "text", + "content": " is usually implemented as a UNet [47]. Besides, the diffusion model's objective in Eq. (4) can be easily extended to a conditional diffusion paradigm by adding a condition c to the " + }, + { + "bbox": [ + 313, + 606, + 554, + 667 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 606, + 554, + 667 + ], + "type": "text", + "content": ". Thus, the training objective should be modified as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 329, + 671, + 553, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 671, + 553, + 685 + ], + "spans": [ + { + "bbox": [ + 329, + 671, + 553, + 685 + ], + "type": "interline_equation", + "content": "\\mathcal {T} _ {\\theta} = \\mathbb {E} _ {\\epsilon , x _ {0}, t, \\mathbf {c}} [ | | \\epsilon_ {\\theta} (x _ {t}, t, \\mathbf {c}) - \\epsilon | | _ {2} ^ {2} ], \\epsilon \\sim \\mathcal {N} (0, \\mathbf {I}), \\tag {5}", + "image_path": "0f39d74eaedac628471c3aa370251fc49341a8c8098bcc2475bdeca64292153c.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "with the denoiser " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t,t,\\mathbf{c})" + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": " in Eq. (5), we can generate data with a specific condition " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": " by solving a reverse process." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "3728" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 192, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 192, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 192, + 85 + ], + "type": "text", + "content": "3.2. Motivation and Analysis" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 89, + 295, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 89, + 295, + 244 + ], + "spans": [ + { + "bbox": [ + 54, + 89, + 295, + 244 + ], + "type": "text", + "content": "The diffusion model is a probabilistic generative model designed for denoising by systematically reversing a progressive-noising process, and recent studies [61, 68] have verified that diffusion-based generative models can be effective representation learners. Stable diffusion models are able to weave from scratch with fruitful and innovative depictions of high fidelity and further capture the underlying properties, positions, and correspondences from visual concepts. Such a phenomenon shows that SD models can align a textual prompt with its corresponding image context, demonstrating their potential ability in understanding both high-level (e.g., semantic relations) and low-level (e.g., textures, edge, and structures) visual concepts in an image." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 245, + 296, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 245, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 245, + 296, + 342 + ], + "type": "text", + "content": "Autoencoders [71, 76] can learn to reconstruct their input at the output and have been one of the dominant approaches for representation learning. We assume a linear autoencoder expressed as " + }, + { + "bbox": [ + 55, + 245, + 296, + 342 + ], + "type": "inline_equation", + "content": "W_{a}W_{b}\\cdot x + W_{c}\\cdot x" + }, + { + "bbox": [ + 55, + 245, + 296, + 342 + ], + "type": "text", + "content": " to imitate UNet [47] used in diffusion models, where " + }, + { + "bbox": [ + 55, + 245, + 296, + 342 + ], + "type": "inline_equation", + "content": "W_{a}\\in \\mathbb{R}^{d\\times d^{*}}" + }, + { + "bbox": [ + 55, + 245, + 296, + 342 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 245, + 296, + 342 + ], + "type": "inline_equation", + "content": "W_{b}\\in \\mathbb{R}^{d^{*}\\times d}" + }, + { + "bbox": [ + 55, + 245, + 296, + 342 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 245, + 296, + 342 + ], + "type": "inline_equation", + "content": "W_{c}\\in \\mathbb{R}^{d\\times d}" + }, + { + "bbox": [ + 55, + 245, + 296, + 342 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 245, + 296, + 342 + ], + "type": "inline_equation", + "content": "d^{*} < d" + }, + { + "bbox": [ + 55, + 245, + 296, + 342 + ], + "type": "text", + "content": ". For " + }, + { + "bbox": [ + 55, + 245, + 296, + 342 + ], + "type": "inline_equation", + "content": "x_{t} = \\sqrt{\\bar{\\alpha}_{t}} x_{0} + \\sqrt{1 - \\bar{\\alpha}_{t}}\\epsilon_{t}" + }, + { + "bbox": [ + 55, + 245, + 296, + 342 + ], + "type": "text", + "content": " and we set " + }, + { + "bbox": [ + 55, + 245, + 296, + 342 + ], + "type": "inline_equation", + "content": "\\mathbb{E}[x_0] = 0" + }, + { + "bbox": [ + 55, + 245, + 296, + 342 + ], + "type": "text", + "content": ", thus minimizing the loss " + }, + { + "bbox": [ + 55, + 245, + 296, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\theta}" + }, + { + "bbox": [ + 55, + 245, + 296, + 342 + ], + "type": "text", + "content": " in Eq. (4) is equivalent to minimize the modified objective " + }, + { + "bbox": [ + 55, + 245, + 296, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\theta}^{*}" + }, + { + "bbox": [ + 55, + 245, + 296, + 342 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 350, + 296, + 379 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 350, + 296, + 379 + ], + "spans": [ + { + "bbox": [ + 63, + 350, + 296, + 379 + ], + "type": "interline_equation", + "content": "\\begin{array}{r l} \\mathcal {T} _ {\\theta} ^ {*} = & \\mathbb {E} _ {\\epsilon , x _ {0}, t} [ \\| \\epsilon_ {\\theta} (\\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t}) - \\epsilon \\| _ {2} ^ {2} ], \\\\ & \\equiv \\min \\left\\{\\bar {\\alpha} _ {t} \\left(W _ {a} W _ {b} + W _ {c}\\right) ^ {\\top} \\Sigma_ {x _ {0}} \\left(W _ {a} W _ {b} + W _ {c}\\right) \\right\\}, \\end{array} \\tag {6}", + "image_path": "5f367bdb9eb7405b4f8256a13c433bdd5aa3ec7793a12742196d30b7f7f479c4.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 390, + 296, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 296, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 296, + 462 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 390, + 296, + 462 + ], + "type": "inline_equation", + "content": "\\Sigma_{x_0}" + }, + { + "bbox": [ + 55, + 390, + 296, + 462 + ], + "type": "text", + "content": " means covariance matrix of " + }, + { + "bbox": [ + 55, + 390, + 296, + 462 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 55, + 390, + 296, + 462 + ], + "type": "text", + "content": ". Eq. (6) ensures a meaningful latent space for the diffusion models, similar to the linear autoencoder that encodes data into the principal components space. Therefore, the diffusion training process is equivalent to encourage the disentanglement of the latent representations until good-quality reconstruction is reached." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 463, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 296, + 714 + ], + "type": "text", + "content": "To validate the representation ability in diffusion models, we use t-SNE [57] to visualize the latent features, shown in Fig. 2, without training on the specific MNIST [22] dataset. Concretely, we use the pre-trained SD model to extract features of the UNet decoder " + }, + { + "bbox": [ + 55, + 463, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 55, + 463, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 463, + 296, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 463, + 296, + 714 + ], + "type": "text", + "content": " means diffusion time-steps, " + }, + { + "bbox": [ + 55, + 463, + 296, + 714 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 463, + 296, + 714 + ], + "type": "text", + "content": " means the layer of the UNet decoder " + }, + { + "bbox": [ + 55, + 463, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 55, + 463, + 296, + 714 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 463, + 296, + 714 + ], + "type": "inline_equation", + "content": "L \\in \\{1, 2, 3, 4\\}" + }, + { + "bbox": [ + 55, + 463, + 296, + 714 + ], + "type": "text", + "content": "). Our results in Fig. 2 indicate that, even without supervision, diffusion models are able to group input samples at properly selected time-steps " + }, + { + "bbox": [ + 55, + 463, + 296, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 463, + 296, + 714 + ], + "type": "text", + "content": " and layer " + }, + { + "bbox": [ + 55, + 463, + 296, + 714 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 463, + 296, + 714 + ], + "type": "text", + "content": ". However, when the " + }, + { + "bbox": [ + 55, + 463, + 296, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 463, + 296, + 714 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 55, + 463, + 296, + 714 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 463, + 296, + 714 + ], + "type": "text", + "content": " is too large (e.g., " + }, + { + "bbox": [ + 55, + 463, + 296, + 714 + ], + "type": "inline_equation", + "content": "T = 999" + }, + { + "bbox": [ + 55, + 463, + 296, + 714 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 55, + 463, + 296, + 714 + ], + "type": "inline_equation", + "content": "L = 4" + }, + { + "bbox": [ + 55, + 463, + 296, + 714 + ], + "type": "text", + "content": "), the representations become inseparable, leading to blurred decision boundaries. Furthermore, we apply K-Means clustering (cluster size set as 5) on the latent features from the diffusion UNet decoder and visually examine whether the diffusion features contain consistent semantic information. In Fig. 3, different parts of the visual objects are clustered, showing that earlier layers (e.g., " + }, + { + "bbox": [ + 55, + 463, + 296, + 714 + ], + "type": "inline_equation", + "content": "L = 2" + }, + { + "bbox": [ + 55, + 463, + 296, + 714 + ], + "type": "text", + "content": ") capture coarse yet consistent semantics, while later layers (e.g., " + }, + { + "bbox": [ + 55, + 463, + 296, + 714 + ], + "type": "inline_equation", + "content": "L = 3" + }, + { + "bbox": [ + 55, + 463, + 296, + 714 + ], + "type": "text", + "content": ") focus on low-level textural details. Both t-SNE and K-Means observation results motivate us to leverage the fine-grained visual representation capabilities from the SD model to promote heterogeneous FL training." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 375, + 107 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 375, + 107 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 375, + 107 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 375, + 107 + ], + "type": "image", + "image_path": "f8a6c63c46d755a35a6ffff60731a424d3f285a268df8ff1f96c7c9045251e21.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 110, + 365, + 118 + ], + "lines": [ + { + "bbox": [ + 326, + 110, + 365, + 118 + ], + "spans": [ + { + "bbox": [ + 326, + 110, + 365, + 118 + ], + "type": "inline_equation", + "content": "\\mathrm{T} = 100" + }, + { + "bbox": [ + 326, + 110, + 365, + 118 + ], + "type": "inline_equation", + "content": "\\mathrm{L} = 1" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 377, + 72, + 432, + 107 + ], + "blocks": [ + { + "bbox": [ + 377, + 72, + 432, + 107 + ], + "lines": [ + { + "bbox": [ + 377, + 72, + 432, + 107 + ], + "spans": [ + { + "bbox": [ + 377, + 72, + 432, + 107 + ], + "type": "image", + "image_path": "3a67be20e7b2fa7dea25cf54aeee10550fa233edc80ef511bd75c0a32dba453e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 385, + 110, + 425, + 118 + ], + "lines": [ + { + "bbox": [ + 385, + 110, + 425, + 118 + ], + "spans": [ + { + "bbox": [ + 385, + 110, + 425, + 118 + ], + "type": "inline_equation", + "content": "\\mathrm{T} = 100, \\mathrm{L} = 2" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 434, + 71, + 492, + 107 + ], + "blocks": [ + { + "bbox": [ + 434, + 71, + 492, + 107 + ], + "lines": [ + { + "bbox": [ + 434, + 71, + 492, + 107 + ], + "spans": [ + { + "bbox": [ + 434, + 71, + 492, + 107 + ], + "type": "image", + "image_path": "636c14c8e2b0ef1c96ad52b480b999aeddd00272a308840208646698dabcab93.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 444, + 110, + 484, + 118 + ], + "lines": [ + { + "bbox": [ + 444, + 110, + 484, + 118 + ], + "spans": [ + { + "bbox": [ + 444, + 110, + 484, + 118 + ], + "type": "inline_equation", + "content": "\\mathrm{T} = 100, \\mathrm{L} = 3" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 494, + 71, + 553, + 107 + ], + "blocks": [ + { + "bbox": [ + 494, + 71, + 553, + 107 + ], + "lines": [ + { + "bbox": [ + 494, + 71, + 553, + 107 + ], + "spans": [ + { + "bbox": [ + 494, + 71, + 553, + 107 + ], + "type": "image", + "image_path": "ac0778df95d294b66dd811290369d870d1189996d3f8f939f30da54a3259096d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 504, + 110, + 543, + 118 + ], + "lines": [ + { + "bbox": [ + 504, + 110, + 543, + 118 + ], + "spans": [ + { + "bbox": [ + 504, + 110, + 543, + 118 + ], + "type": "inline_equation", + "content": "\\mathrm{T} = 100, \\mathrm{L} = 4" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 316, + 120, + 375, + 156 + ], + "blocks": [ + { + "bbox": [ + 316, + 120, + 375, + 156 + ], + "lines": [ + { + "bbox": [ + 316, + 120, + 375, + 156 + ], + "spans": [ + { + "bbox": [ + 316, + 120, + 375, + 156 + ], + "type": "image", + "image_path": "8f3dea7801cb18ee61ea8cc8202d2539d0362686418ad5b2c0599d676ebf616b.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 159, + 365, + 167 + ], + "lines": [ + { + "bbox": [ + 326, + 159, + 365, + 167 + ], + "spans": [ + { + "bbox": [ + 326, + 159, + 365, + 167 + ], + "type": "inline_equation", + "content": "\\mathrm{T} = 100, \\mathrm{L} = 2" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 377, + 120, + 433, + 157 + ], + "blocks": [ + { + "bbox": [ + 377, + 120, + 433, + 157 + ], + "lines": [ + { + "bbox": [ + 377, + 120, + 433, + 157 + ], + "spans": [ + { + "bbox": [ + 377, + 120, + 433, + 157 + ], + "type": "image", + "image_path": "0612edbe36ba27a6efb4d65a7d2454bb7cd5106497a9e1e729cb9bdffae76c33.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 385, + 159, + 425, + 167 + ], + "lines": [ + { + "bbox": [ + 385, + 159, + 425, + 167 + ], + "spans": [ + { + "bbox": [ + 385, + 159, + 425, + 167 + ], + "type": "inline_equation", + "content": "\\mathrm{T} = 300" + }, + { + "bbox": [ + 385, + 159, + 425, + 167 + ], + "type": "inline_equation", + "content": "\\mathrm{L} = 2" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 434, + 120, + 492, + 157 + ], + "blocks": [ + { + "bbox": [ + 434, + 120, + 492, + 157 + ], + "lines": [ + { + "bbox": [ + 434, + 120, + 492, + 157 + ], + "spans": [ + { + "bbox": [ + 434, + 120, + 492, + 157 + ], + "type": "image", + "image_path": "b5ac2f9cb919bc3bd8789d3d122db37ff8664819bd0f8d8728a60e258fb92019.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 444, + 159, + 484, + 167 + ], + "lines": [ + { + "bbox": [ + 444, + 159, + 484, + 167 + ], + "spans": [ + { + "bbox": [ + 444, + 159, + 484, + 167 + ], + "type": "inline_equation", + "content": "\\mathrm{T} = 500" + }, + { + "bbox": [ + 444, + 159, + 484, + 167 + ], + "type": "inline_equation", + "content": "\\mathrm{L} = 2" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 494, + 120, + 553, + 157 + ], + "blocks": [ + { + "bbox": [ + 494, + 120, + 553, + 157 + ], + "lines": [ + { + "bbox": [ + 494, + 120, + 553, + 157 + ], + "spans": [ + { + "bbox": [ + 494, + 120, + 553, + 157 + ], + "type": "image", + "image_path": "49fec3ae084372be889e25e172bd4c71f8b0abe8bf2aca0967bc5eb06b318a0f.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 504, + 159, + 543, + 167 + ], + "lines": [ + { + "bbox": [ + 504, + 159, + 543, + 167 + ], + "spans": [ + { + "bbox": [ + 504, + 159, + 543, + 167 + ], + "type": "inline_equation", + "content": "\\mathrm{T} = 999, \\mathrm{L} = 2" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 316, + 224, + 358, + 266 + ], + "blocks": [ + { + "bbox": [ + 313, + 177, + 555, + 211 + ], + "lines": [ + { + "bbox": [ + 313, + 177, + 555, + 211 + ], + "spans": [ + { + "bbox": [ + 313, + 177, + 555, + 211 + ], + "type": "text", + "content": "Figure 2. The t-SNE feature visualization on the MNIST dataset from the diffusion UNet decoder, without training, where diffusion time-steps " + }, + { + "bbox": [ + 313, + 177, + 555, + 211 + ], + "type": "inline_equation", + "content": "T \\in \\{100, 300, 500, 999\\}" + }, + { + "bbox": [ + 313, + 177, + 555, + 211 + ], + "type": "text", + "content": " and layer " + }, + { + "bbox": [ + 313, + 177, + 555, + 211 + ], + "type": "inline_equation", + "content": "L \\in \\{1, 2, 3, 4\\}" + }, + { + "bbox": [ + 313, + 177, + 555, + 211 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 224, + 358, + 266 + ], + "lines": [ + { + "bbox": [ + 316, + 224, + 358, + 266 + ], + "spans": [ + { + "bbox": [ + 316, + 224, + 358, + 266 + ], + "type": "image", + "image_path": "7e8df0660532b3ecf50da30033198c24741b53c5ad94df89dfd4e887f077b15c.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 316, + 267, + 358, + 310 + ], + "blocks": [ + { + "bbox": [ + 316, + 267, + 358, + 310 + ], + "lines": [ + { + "bbox": [ + 316, + 267, + 358, + 310 + ], + "spans": [ + { + "bbox": [ + 316, + 267, + 358, + 310 + ], + "type": "image", + "image_path": "4e176ea07de581eeabc07115f2217e77dbf22c2d51a899cb9152bd0b198084c8.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 316, + 312, + 358, + 354 + ], + "blocks": [ + { + "bbox": [ + 316, + 312, + 358, + 354 + ], + "lines": [ + { + "bbox": [ + 316, + 312, + 358, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 312, + 358, + 354 + ], + "type": "image", + "image_path": "8d53eedb8fc17507590e81858cd79f3f732586882b3b6cef15cd6fbdba102d72.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 357, + 351, + 367 + ], + "lines": [ + { + "bbox": [ + 323, + 357, + 351, + 367 + ], + "spans": [ + { + "bbox": [ + 323, + 357, + 351, + 367 + ], + "type": "text", + "content": "Sample" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 364, + 224, + 407, + 266 + ], + "blocks": [ + { + "bbox": [ + 364, + 224, + 407, + 266 + ], + "lines": [ + { + "bbox": [ + 364, + 224, + 407, + 266 + ], + "spans": [ + { + "bbox": [ + 364, + 224, + 407, + 266 + ], + "type": "image", + "image_path": "3a48b6ffa541373ec0a7ac0c6a64d553b2365d202a359fdd995129792049d321.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 365, + 267, + 406, + 310 + ], + "blocks": [ + { + "bbox": [ + 365, + 267, + 406, + 310 + ], + "lines": [ + { + "bbox": [ + 365, + 267, + 406, + 310 + ], + "spans": [ + { + "bbox": [ + 365, + 267, + 406, + 310 + ], + "type": "image", + "image_path": "4cd45332ef21f842c993fc9885e7458cb0052572a160c49e8a6ffd21bf40ad05.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 365, + 312, + 407, + 354 + ], + "blocks": [ + { + "bbox": [ + 365, + 312, + 407, + 354 + ], + "lines": [ + { + "bbox": [ + 365, + 312, + 407, + 354 + ], + "spans": [ + { + "bbox": [ + 365, + 312, + 407, + 354 + ], + "type": "image", + "image_path": "9271bb3537daa97e75ba5164c4ed302a703d2f2cbab43ff5dc03fabdb110fa70.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 366, + 357, + 405, + 367 + ], + "lines": [ + { + "bbox": [ + 366, + 357, + 405, + 367 + ], + "spans": [ + { + "bbox": [ + 366, + 357, + 405, + 367 + ], + "type": "text", + "content": "Layer " + }, + { + "bbox": [ + 366, + 357, + 405, + 367 + ], + "type": "inline_equation", + "content": "\\mathrm{L} = 1" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 413, + 224, + 455, + 266 + ], + "blocks": [ + { + "bbox": [ + 413, + 224, + 455, + 266 + ], + "lines": [ + { + "bbox": [ + 413, + 224, + 455, + 266 + ], + "spans": [ + { + "bbox": [ + 413, + 224, + 455, + 266 + ], + "type": "image", + "image_path": "a566f150a4f4598c07fc59202fef3f261e16d993fbb9901e67ea0dccf9a9f8f2.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 413, + 267, + 455, + 310 + ], + "blocks": [ + { + "bbox": [ + 413, + 267, + 455, + 310 + ], + "lines": [ + { + "bbox": [ + 413, + 267, + 455, + 310 + ], + "spans": [ + { + "bbox": [ + 413, + 267, + 455, + 310 + ], + "type": "image", + "image_path": "be6576963a4e3bcbd2421fcf63fd7c2cb2067b4fbcba4e5a47f0053c66c42e89.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 413, + 312, + 455, + 354 + ], + "blocks": [ + { + "bbox": [ + 413, + 312, + 455, + 354 + ], + "lines": [ + { + "bbox": [ + 413, + 312, + 455, + 354 + ], + "spans": [ + { + "bbox": [ + 413, + 312, + 455, + 354 + ], + "type": "image", + "image_path": "f53f98e68a2d16cc8116be8e0433fdd0b83af261cf2c0caf7d751d3d54e8ea9a.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 415, + 357, + 454, + 367 + ], + "lines": [ + { + "bbox": [ + 415, + 357, + 454, + 367 + ], + "spans": [ + { + "bbox": [ + 415, + 357, + 454, + 367 + ], + "type": "text", + "content": "Layer " + }, + { + "bbox": [ + 415, + 357, + 454, + 367 + ], + "type": "inline_equation", + "content": "\\mathrm{L} = 2" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 462, + 224, + 504, + 266 + ], + "blocks": [ + { + "bbox": [ + 462, + 224, + 504, + 266 + ], + "lines": [ + { + "bbox": [ + 462, + 224, + 504, + 266 + ], + "spans": [ + { + "bbox": [ + 462, + 224, + 504, + 266 + ], + "type": "image", + "image_path": "d37db70802b268e552fadeee8f1e0f6f56d61ef3f6a60cbfabe5b718ff58fe6d.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 462, + 267, + 504, + 310 + ], + "blocks": [ + { + "bbox": [ + 462, + 267, + 504, + 310 + ], + "lines": [ + { + "bbox": [ + 462, + 267, + 504, + 310 + ], + "spans": [ + { + "bbox": [ + 462, + 267, + 504, + 310 + ], + "type": "image", + "image_path": "25b1343cd9d6f5d90e063b0436637742948ecc1a97a9d2a54a0c563e129bde26.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 462, + 312, + 504, + 354 + ], + "blocks": [ + { + "bbox": [ + 462, + 312, + 504, + 354 + ], + "lines": [ + { + "bbox": [ + 462, + 312, + 504, + 354 + ], + "spans": [ + { + "bbox": [ + 462, + 312, + 504, + 354 + ], + "type": "image", + "image_path": "684dea9a71546e911a440b052a480f881082113ca032a2d2f07b26bc45e80915.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 464, + 357, + 503, + 367 + ], + "lines": [ + { + "bbox": [ + 464, + 357, + 503, + 367 + ], + "spans": [ + { + "bbox": [ + 464, + 357, + 503, + 367 + ], + "type": "text", + "content": "Layer " + }, + { + "bbox": [ + 464, + 357, + 503, + 367 + ], + "type": "inline_equation", + "content": "\\mathrm{L} = 3" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 511, + 224, + 553, + 266 + ], + "blocks": [ + { + "bbox": [ + 511, + 224, + 553, + 266 + ], + "lines": [ + { + "bbox": [ + 511, + 224, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 511, + 224, + 553, + 266 + ], + "type": "image", + "image_path": "63b94aa5355f2b8cd2d400d6e28f0935ce05e80457426ac5aef17be979d60175.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 511, + 267, + 553, + 310 + ], + "blocks": [ + { + "bbox": [ + 511, + 267, + 553, + 310 + ], + "lines": [ + { + "bbox": [ + 511, + 267, + 553, + 310 + ], + "spans": [ + { + "bbox": [ + 511, + 267, + 553, + 310 + ], + "type": "image", + "image_path": "9a4072b7bff2a1f37b9c139730ac474828e105edb1c833f6ac2cc89c166c87c2.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 511, + 312, + 553, + 354 + ], + "blocks": [ + { + "bbox": [ + 511, + 312, + 553, + 354 + ], + "lines": [ + { + "bbox": [ + 511, + 312, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 511, + 312, + 553, + 354 + ], + "type": "image", + "image_path": "63d28bfb0bdf8bdf0d037d5996ff91236fd594adc490751ff284b6cf047f94cb.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 512, + 357, + 552, + 367 + ], + "lines": [ + { + "bbox": [ + 512, + 357, + 552, + 367 + ], + "spans": [ + { + "bbox": [ + 512, + 357, + 552, + 367 + ], + "type": "text", + "content": "Layer L=4" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 316, + 412, + 358, + 455 + ], + "blocks": [ + { + "bbox": [ + 313, + 377, + 555, + 399 + ], + "lines": [ + { + "bbox": [ + 313, + 377, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 555, + 399 + ], + "type": "text", + "content": "Figure 3. The K-Means clustering feature visualization (each color represents a cluster) from the diffusion UNet decoder's layers." + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 412, + 358, + 455 + ], + "lines": [ + { + "bbox": [ + 316, + 412, + 358, + 455 + ], + "spans": [ + { + "bbox": [ + 316, + 412, + 358, + 455 + ], + "type": "image", + "image_path": "235e2f1b18ac52bbb92c196436c9d08492973c42d19545bf9d893b01d660e043.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 316, + 456, + 358, + 499 + ], + "blocks": [ + { + "bbox": [ + 316, + 456, + 358, + 499 + ], + "lines": [ + { + "bbox": [ + 316, + 456, + 358, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 358, + 499 + ], + "type": "image", + "image_path": "b60abde78623d328711c61c2a319a1da70e85a976831f4eec0f61ebf388a54c8.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 316, + 501, + 358, + 544 + ], + "blocks": [ + { + "bbox": [ + 316, + 501, + 358, + 544 + ], + "lines": [ + { + "bbox": [ + 316, + 501, + 358, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 358, + 544 + ], + "type": "image", + "image_path": "a2a19a0e4d27aad4549756fd7424b2824cc3a6bd763d36427eed0c0a49061fb0.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 546, + 351, + 556 + ], + "lines": [ + { + "bbox": [ + 323, + 546, + 351, + 556 + ], + "spans": [ + { + "bbox": [ + 323, + 546, + 351, + 556 + ], + "type": "text", + "content": "Sample" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 365, + 413, + 406, + 455 + ], + "blocks": [ + { + "bbox": [ + 365, + 413, + 406, + 455 + ], + "lines": [ + { + "bbox": [ + 365, + 413, + 406, + 455 + ], + "spans": [ + { + "bbox": [ + 365, + 413, + 406, + 455 + ], + "type": "image", + "image_path": "b3fb60e29d3b8928ff02be35e06d79b5ea4103f83d37afa7e3f5030e2961c0cf.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 365, + 456, + 406, + 499 + ], + "blocks": [ + { + "bbox": [ + 365, + 456, + 406, + 499 + ], + "lines": [ + { + "bbox": [ + 365, + 456, + 406, + 499 + ], + "spans": [ + { + "bbox": [ + 365, + 456, + 406, + 499 + ], + "type": "image", + "image_path": "2be7cb8d134bd81d251ecc69eff81fa36b38e5132dde342f577daaecb1cba954.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_body" + } + ], + "index": 49 + }, + { + "type": "image", + "bbox": [ + 365, + 501, + 406, + 544 + ], + "blocks": [ + { + "bbox": [ + 365, + 501, + 406, + 544 + ], + "lines": [ + { + "bbox": [ + 365, + 501, + 406, + 544 + ], + "spans": [ + { + "bbox": [ + 365, + 501, + 406, + 544 + ], + "type": "image", + "image_path": "d5485b794e5d93b1443ccc7eb506b6910483bbad041609356317fda952e28abf.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 366, + 546, + 405, + 555 + ], + "lines": [ + { + "bbox": [ + 366, + 546, + 405, + 555 + ], + "spans": [ + { + "bbox": [ + 366, + 546, + 405, + 555 + ], + "type": "text", + "content": "Layer " + }, + { + "bbox": [ + 366, + 546, + 405, + 555 + ], + "type": "inline_equation", + "content": "L = 2" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 413, + 413, + 455, + 455 + ], + "blocks": [ + { + "bbox": [ + 413, + 413, + 455, + 455 + ], + "lines": [ + { + "bbox": [ + 413, + 413, + 455, + 455 + ], + "spans": [ + { + "bbox": [ + 413, + 413, + 455, + 455 + ], + "type": "image", + "image_path": "34285cf22b2b7ad3d94c5d074149060a33a743cdb931914febe3ca71893dd38f.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 413, + 456, + 455, + 499 + ], + "blocks": [ + { + "bbox": [ + 413, + 456, + 455, + 499 + ], + "lines": [ + { + "bbox": [ + 413, + 456, + 455, + 499 + ], + "spans": [ + { + "bbox": [ + 413, + 456, + 455, + 499 + ], + "type": "image", + "image_path": "de4b2a908528032c4ee7e0ab4063267994b2816e052a4be081747022546fcba7.jpg" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_body" + } + ], + "index": 53 + }, + { + "type": "image", + "bbox": [ + 413, + 501, + 455, + 544 + ], + "blocks": [ + { + "bbox": [ + 413, + 501, + 455, + 544 + ], + "lines": [ + { + "bbox": [ + 413, + 501, + 455, + 544 + ], + "spans": [ + { + "bbox": [ + 413, + 501, + 455, + 544 + ], + "type": "image", + "image_path": "6c47cea1cb100428c7a1881367ae23f4884ca57853d724e12cad253a6ef65b11.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 415, + 546, + 454, + 555 + ], + "lines": [ + { + "bbox": [ + 415, + 546, + 454, + 555 + ], + "spans": [ + { + "bbox": [ + 415, + 546, + 454, + 555 + ], + "type": "text", + "content": "Layer " + }, + { + "bbox": [ + 415, + 546, + 454, + 555 + ], + "type": "inline_equation", + "content": "\\mathrm{L} = 3" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_caption" + } + ], + "index": 54 + }, + { + "type": "image", + "bbox": [ + 462, + 413, + 504, + 455 + ], + "blocks": [ + { + "bbox": [ + 462, + 413, + 504, + 455 + ], + "lines": [ + { + "bbox": [ + 462, + 413, + 504, + 455 + ], + "spans": [ + { + "bbox": [ + 462, + 413, + 504, + 455 + ], + "type": "image", + "image_path": "b3fec03c3b8ea8199f936d8baa20830a15e78f6404f1430e9b4051e554f8e354.jpg" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_body" + } + ], + "index": 56 + }, + { + "type": "image", + "bbox": [ + 462, + 456, + 504, + 499 + ], + "blocks": [ + { + "bbox": [ + 462, + 456, + 504, + 499 + ], + "lines": [ + { + "bbox": [ + 462, + 456, + 504, + 499 + ], + "spans": [ + { + "bbox": [ + 462, + 456, + 504, + 499 + ], + "type": "image", + "image_path": "85bb727f570431443904a0b95e060875dbdfbb73c1a2f3b34491ed408b4da011.jpg" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_body" + } + ], + "index": 57 + }, + { + "type": "image", + "bbox": [ + 462, + 501, + 504, + 544 + ], + "blocks": [ + { + "bbox": [ + 462, + 501, + 504, + 544 + ], + "lines": [ + { + "bbox": [ + 462, + 501, + 504, + 544 + ], + "spans": [ + { + "bbox": [ + 462, + 501, + 504, + 544 + ], + "type": "image", + "image_path": "652398c7f9ecdf1a7abecd29e32d01ea9b3463422d82cf2be49348821ae6eae9.jpg" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 464, + 546, + 503, + 555 + ], + "lines": [ + { + "bbox": [ + 464, + 546, + 503, + 555 + ], + "spans": [ + { + "bbox": [ + 464, + 546, + 503, + 555 + ], + "type": "text", + "content": "Layer L=4" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 567, + 555, + 600 + ], + "lines": [ + { + "bbox": [ + 313, + 567, + 555, + 600 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 555, + 600 + ], + "type": "text", + "content": "Figure 4. The PCA visualization of early (layer 2), later (layer 4), and fused features (see Eq. (8)), where the first three components of PCA-computed serve as color channels for each image." + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "image_caption" + } + ], + "index": 58 + }, + { + "type": "image", + "bbox": [ + 511, + 413, + 553, + 455 + ], + "blocks": [ + { + "bbox": [ + 511, + 413, + 553, + 455 + ], + "lines": [ + { + "bbox": [ + 511, + 413, + 553, + 455 + ], + "spans": [ + { + "bbox": [ + 511, + 413, + 553, + 455 + ], + "type": "image", + "image_path": "a60727202fe7205c4a39ee3040def9e46bf97f57cb3a5dc6e79503fab55b88b5.jpg" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_body" + } + ], + "index": 60 + }, + { + "type": "image", + "bbox": [ + 511, + 456, + 553, + 499 + ], + "blocks": [ + { + "bbox": [ + 511, + 456, + 553, + 499 + ], + "lines": [ + { + "bbox": [ + 511, + 456, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 511, + 456, + 553, + 499 + ], + "type": "image", + "image_path": "09cf4bc4d629bf2f1cf8eea5a44b1cb9561f5245ff380917365cc69ff3c891a0.jpg" + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_body" + } + ], + "index": 61 + }, + { + "type": "image", + "bbox": [ + 511, + 501, + 553, + 544 + ], + "blocks": [ + { + "bbox": [ + 511, + 501, + 553, + 544 + ], + "lines": [ + { + "bbox": [ + 511, + 501, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 511, + 501, + 553, + 544 + ], + "type": "image", + "image_path": "0f1dd56c24bdc84536d1738b6e48f6e8da2e928851787d9884bfb47734b33195.jpg" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 520, + 546, + 545, + 555 + ], + "lines": [ + { + "bbox": [ + 520, + 546, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 520, + 546, + 545, + 555 + ], + "type": "text", + "content": "Fused" + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "image_caption" + } + ], + "index": 62 + }, + { + "bbox": [ + 313, + 623, + 516, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 623, + 516, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 623, + 516, + 635 + ], + "type": "text", + "content": "3.3. Conditional Diffusion Representations" + } + ] + } + ], + "index": 65 + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": "The key idea of our conditional diffusion representations is based on cognitive diversity of generative diffusion models for different text prompts, via a condition-guided manner. Specifically, for the client " + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": ", given an input image " + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": " of the local dataset " + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_k = \\{x_i,y_i\\}_{i = 1}^{n_k},y_i\\in \\{1,\\dots,|\\mathcal{C}|\\}" + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": " means the label of " + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": ", and we let " + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": " represents the set of all class names," + } + ] + } + ], + "index": 66 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "3729" + } + ] + } + ], + "index": 67 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "text", + "content": "so we use " + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "inline_equation", + "content": "S_{y_i}" + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "text", + "content": " to denote the class name of the label " + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "inline_equation", + "content": "y_i" + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "text", + "content": ". From our implementation, the image " + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "text", + "content": " is fed into the encoder " + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "inline_equation", + "content": "h_k" + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "text", + "content": " of the local model " + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "inline_equation", + "content": "f_k" + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "text", + "content": " to obtain the embedding " + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i" + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "text", + "content": ", then " + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i" + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "text", + "content": " is treated as a condition and injected into the diffusion model. For a pre-trained SD model, " + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i" + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "text", + "content": " is further combined with a relevant text prompt " + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{y_i} =" + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "text", + "content": " 'a photo of a " + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "inline_equation", + "content": "S_{y_i}" + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "text", + "content": "' and sent to the UNet decoder to produce diffusion representations " + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_i" + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "text", + "content": ". Note that since we do not add noise, we simply set the time-steps " + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "inline_equation", + "content": "T = 0" + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "text", + "content": " of the diffusion process. Typically, the image " + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "text", + "content": " size is " + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_i" + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "text", + "content": " contains 4 feature maps, where the " + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "text", + "content": "-th layer's feature map " + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{i(L)}" + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "text", + "content": " has the spatial size " + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "inline_equation", + "content": "\\{7,14,28,28\\}" + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "inline_equation", + "content": "L \\in \\{1,2,3,4\\}" + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "text", + "content": "), and " + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "inline_equation", + "content": "\\{1280,1280,640,320\\}" + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "text", + "content": " are channels. Formally, conditional diffusion representations " + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_i" + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "text", + "content": " are computed by the " + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "text", + "content": " in Eq. (5), image " + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "text", + "content": ", condition " + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i" + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "text", + "content": ", text " + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{y_i}" + }, + { + "bbox": [ + 55, + 72, + 296, + 242 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 90, + 252, + 295, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 252, + 295, + 266 + ], + "spans": [ + { + "bbox": [ + 90, + 252, + 295, + 266 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {i} = \\epsilon_ {\\theta} \\left(x _ {i}, \\tau_ {i}, T = 0, \\mathbf {c} _ {i}\\right), \\tau_ {i} \\leftarrow \\varphi \\left(\\mathcal {P} _ {y _ {i}}\\right), \\tag {7}", + "image_path": "1449ce8de14b6791f00a14272ea189ec0864d423f73b02041201810badaff742.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 271, + 296, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 271, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 296, + 392 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 271, + 296, + 392 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 55, + 271, + 296, + 392 + ], + "type": "text", + "content": " denotes the text encoder of the SD model and " + }, + { + "bbox": [ + 55, + 271, + 296, + 392 + ], + "type": "inline_equation", + "content": "\\tau_{i}" + }, + { + "bbox": [ + 55, + 271, + 296, + 392 + ], + "type": "text", + "content": " is the text features with the prompt " + }, + { + "bbox": [ + 55, + 271, + 296, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{y_i}" + }, + { + "bbox": [ + 55, + 271, + 296, + 392 + ], + "type": "text", + "content": ". Inspired by the results in Fig. 3, which motivate us to combine features at different layers of " + }, + { + "bbox": [ + 55, + 271, + 296, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_i" + }, + { + "bbox": [ + 55, + 271, + 296, + 392 + ], + "type": "text", + "content": " to capture both semantics and details. A simple concatenation, however, can lead to an unnecessarily high-dimensional feature " + }, + { + "bbox": [ + 55, + 271, + 296, + 392 + ], + "type": "inline_equation", + "content": "(1280 + 640 + 320 = 2240)" + }, + { + "bbox": [ + 55, + 271, + 296, + 392 + ], + "type": "text", + "content": ". To reduce high dimension, we apply PCA for each layer " + }, + { + "bbox": [ + 55, + 271, + 296, + 392 + ], + "type": "inline_equation", + "content": "L\\in \\{2,3,4\\}" + }, + { + "bbox": [ + 55, + 271, + 296, + 392 + ], + "type": "text", + "content": " (due to the spatial resolution too low, we do not consider the " + }, + { + "bbox": [ + 55, + 271, + 296, + 392 + ], + "type": "inline_equation", + "content": "L = 1" + }, + { + "bbox": [ + 55, + 271, + 296, + 392 + ], + "type": "text", + "content": "), and we aggregate them to the same resolution, then we connect them together to obtain the fused features " + }, + { + "bbox": [ + 55, + 271, + 296, + 392 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{F}}_i" + }, + { + "bbox": [ + 55, + 271, + 296, + 392 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 82, + 399, + 295, + 416 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 399, + 295, + 416 + ], + "spans": [ + { + "bbox": [ + 82, + 399, + 295, + 416 + ], + "type": "interline_equation", + "content": "\\widetilde {\\mathcal {F}} _ {i} = \\mathrm {C o n c a t} (\\widetilde {\\mathcal {F}} _ {i (L)}), \\widetilde {\\mathcal {F}} _ {i (L)} = \\mathrm {P C A} (\\mathcal {F} _ {i (L)}), \\quad (8)", + "image_path": "4b268e86e535542ec84587a568f0068908d06103a49ec62564ce05d9cb5b6375.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 424, + 296, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 424, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 54, + 424, + 296, + 568 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 54, + 424, + 296, + 568 + ], + "type": "inline_equation", + "content": "L\\in \\{2,3,4\\}" + }, + { + "bbox": [ + 54, + 424, + 296, + 568 + ], + "type": "text", + "content": " we perform PCA: i) for " + }, + { + "bbox": [ + 54, + 424, + 296, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{i(2)}" + }, + { + "bbox": [ + 54, + 424, + 296, + 568 + ], + "type": "text", + "content": " we first upsample to the same resolution and apply PCA " + }, + { + "bbox": [ + 54, + 424, + 296, + 568 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{1280\\to 256}" + }, + { + "bbox": [ + 54, + 424, + 296, + 568 + ], + "type": "text", + "content": " ii) for " + }, + { + "bbox": [ + 54, + 424, + 296, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{i(3)}" + }, + { + "bbox": [ + 54, + 424, + 296, + 568 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 424, + 296, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{i(4)}" + }, + { + "bbox": [ + 54, + 424, + 296, + 568 + ], + "type": "text", + "content": " , we apply PCA " + }, + { + "bbox": [ + 54, + 424, + 296, + 568 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{640\\to 128}" + }, + { + "bbox": [ + 54, + 424, + 296, + 568 + ], + "type": "text", + "content": " and PCA " + }, + { + "bbox": [ + 54, + 424, + 296, + 568 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{320\\to 128}" + }, + { + "bbox": [ + 54, + 424, + 296, + 568 + ], + "type": "text", + "content": " . Then, we combine them to form the final fused features " + }, + { + "bbox": [ + 54, + 424, + 296, + 568 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{F}}_i" + }, + { + "bbox": [ + 54, + 424, + 296, + 568 + ], + "type": "text", + "content": " . As shown in the last column of Fig. 4, the fused features strike a balance between high-level and low-level visual concepts, focusing on both semantics and textures in the image. On the example in the second row of Fig. 4, the fused feature yields more accurate matches (e.g., face and computer borders). The comparison results in Table. 2 also confirm this viewpoint. We define " + }, + { + "bbox": [ + 54, + 424, + 296, + 568 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{F}}_i" + }, + { + "bbox": [ + 54, + 424, + 296, + 568 + ], + "type": "text", + "content": " calculated by Eq. (8) as conditional diffusion representations of the sample " + }, + { + "bbox": [ + 54, + 424, + 296, + 568 + ], + "type": "inline_equation", + "content": "x_{i}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 576, + 249, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 249, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 249, + 589 + ], + "type": "text", + "content": "3.4. Denoising Diffusion Representations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": "Although we construct the conditional diffusion representations from the diffusion model's generative feedback, " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{F}}_i" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " is still highly variable and depends on the condition generated by the local model, which can not ensure a consistent stable convergence signal. Based on Eq. (4), we explore the inherent denoising property in diffusion models: SD model has learned extensive general knowledge about identical semantics via large pre-training, reflected in the UNet denoising process. Therefore, we aim to extract the denoising diffusion representations contained in a pre-trained SD model, to" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 217 + ], + "type": "text", + "content": "align the sample's embedding with the class semantics uniformly in the feature space. Given the image pair " + }, + { + "bbox": [ + 313, + 72, + 555, + 217 + ], + "type": "inline_equation", + "content": "\\{x_i, y_i\\}" + }, + { + "bbox": [ + 313, + 72, + 555, + 217 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 72, + 555, + 217 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 72, + 555, + 217 + ], + "type": "text", + "content": "-th client's dataset " + }, + { + "bbox": [ + 313, + 72, + 555, + 217 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_k" + }, + { + "bbox": [ + 313, + 72, + 555, + 217 + ], + "type": "text", + "content": ", we first perform a forward process to add " + }, + { + "bbox": [ + 313, + 72, + 555, + 217 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 72, + 555, + 217 + ], + "type": "text", + "content": " time-steps random Gaussian noise on " + }, + { + "bbox": [ + 313, + 72, + 555, + 217 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 313, + 72, + 555, + 217 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 72, + 555, + 217 + ], + "type": "inline_equation", + "content": "\\tilde{x}_i^t = \\sqrt{\\bar{\\alpha}_t} x_i + \\sqrt{1 - \\bar{\\alpha}_t} \\epsilon_t" + }, + { + "bbox": [ + 313, + 72, + 555, + 217 + ], + "type": "text", + "content": ", similar to [46], where " + }, + { + "bbox": [ + 313, + 72, + 555, + 217 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t" + }, + { + "bbox": [ + 313, + 72, + 555, + 217 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 217 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 72, + 555, + 217 + ], + "type": "text", + "content": " are the same as defined in the Eq. (3). Next, we set the text prompt of " + }, + { + "bbox": [ + 313, + 72, + 555, + 217 + ], + "type": "inline_equation", + "content": "\\tilde{x}_i^t" + }, + { + "bbox": [ + 313, + 72, + 555, + 217 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 72, + 555, + 217 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{y_i}" + }, + { + "bbox": [ + 313, + 72, + 555, + 217 + ], + "type": "text", + "content": " = 'a photo of a " + }, + { + "bbox": [ + 313, + 72, + 555, + 217 + ], + "type": "inline_equation", + "content": "S_{y_i}" + }, + { + "bbox": [ + 313, + 72, + 555, + 217 + ], + "type": "text", + "content": "' used in the SD model pre-training [46], to avoid destroying correspondence between visual and linguistic concepts, and the " + }, + { + "bbox": [ + 313, + 72, + 555, + 217 + ], + "type": "inline_equation", + "content": "S_{y_i}" + }, + { + "bbox": [ + 313, + 72, + 555, + 217 + ], + "type": "text", + "content": " denotes the class name of the label " + }, + { + "bbox": [ + 313, + 72, + 555, + 217 + ], + "type": "inline_equation", + "content": "y_i" + }, + { + "bbox": [ + 313, + 72, + 555, + 217 + ], + "type": "text", + "content": ". Formally, the denoising diffusion representations " + }, + { + "bbox": [ + 313, + 72, + 555, + 217 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{H}}_i" + }, + { + "bbox": [ + 313, + 72, + 555, + 217 + ], + "type": "text", + "content": " of the sample " + }, + { + "bbox": [ + 313, + 72, + 555, + 217 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 313, + 72, + 555, + 217 + ], + "type": "text", + "content": " are calculated by the " + }, + { + "bbox": [ + 313, + 72, + 555, + 217 + ], + "type": "inline_equation", + "content": "\\epsilon_\\theta" + }, + { + "bbox": [ + 313, + 72, + 555, + 217 + ], + "type": "text", + "content": " in Eq. (4), noisy sample " + }, + { + "bbox": [ + 313, + 72, + 555, + 217 + ], + "type": "inline_equation", + "content": "\\tilde{x}_i^t" + }, + { + "bbox": [ + 313, + 72, + 555, + 217 + ], + "type": "text", + "content": ", text " + }, + { + "bbox": [ + 313, + 72, + 555, + 217 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{y_i}" + }, + { + "bbox": [ + 313, + 72, + 555, + 217 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 336, + 225, + 495, + 240 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 225, + 495, + 240 + ], + "spans": [ + { + "bbox": [ + 336, + 225, + 495, + 240 + ], + "type": "interline_equation", + "content": "\\mathcal {H} _ {i} = \\epsilon_ {\\theta} (\\tilde {x} _ {i} ^ {t}, \\tau_ {i}, T = t), \\tau_ {i} \\gets \\varphi (\\mathcal {P} _ {y _ {i}}),", + "image_path": "1340deae6ec1a154d4ada1ff72a5d0d72b496e0c707226cdef8e50fc81952297.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 339, + 237, + 553, + 258 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 237, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 339, + 237, + 553, + 258 + ], + "type": "interline_equation", + "content": "\\widetilde {\\mathcal {H}} _ {i} = \\operatorname {C o n c a t} \\left(\\widetilde {\\mathcal {H}} _ {i (L)}\\right), \\quad \\widetilde {\\mathcal {H}} _ {i (L)} = \\operatorname {P C A} \\left(\\mathcal {H} _ {i (L)}\\right), \\tag {9}", + "image_path": "fa66c7e3f43f33c3970d078d17cc7e6d1848d21151737a4751b40f00aee9e789.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 265, + 555, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 265, + 555, + 363 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 555, + 363 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 265, + 555, + 363 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 313, + 265, + 555, + 363 + ], + "type": "text", + "content": " denotes the text encoder of the SD model and " + }, + { + "bbox": [ + 313, + 265, + 555, + 363 + ], + "type": "inline_equation", + "content": "\\tau_{i}" + }, + { + "bbox": [ + 313, + 265, + 555, + 363 + ], + "type": "text", + "content": " is the text features of the text prompt " + }, + { + "bbox": [ + 313, + 265, + 555, + 363 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{y_i}" + }, + { + "bbox": [ + 313, + 265, + 555, + 363 + ], + "type": "text", + "content": ". We further analyze the effects of the time-steps " + }, + { + "bbox": [ + 313, + 265, + 555, + 363 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 265, + 555, + 363 + ], + "type": "text", + "content": " in Fig. 6. Then, we combine the features of layers " + }, + { + "bbox": [ + 313, + 265, + 555, + 363 + ], + "type": "inline_equation", + "content": "L\\in \\{2,3,4\\}" + }, + { + "bbox": [ + 313, + 265, + 555, + 363 + ], + "type": "text", + "content": ", like Eq. (8), to obtain fused features " + }, + { + "bbox": [ + 313, + 265, + 555, + 363 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{H}}_i" + }, + { + "bbox": [ + 313, + 265, + 555, + 363 + ], + "type": "text", + "content": ". Intuitively, " + }, + { + "bbox": [ + 313, + 265, + 555, + 363 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{H}}_i" + }, + { + "bbox": [ + 313, + 265, + 555, + 363 + ], + "type": "text", + "content": " is considered as a virtual teacher to calibrate the embedding of " + }, + { + "bbox": [ + 313, + 265, + 555, + 363 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 265, + 555, + 363 + ], + "type": "text", + "content": " at the feature-level. We define the denoising features " + }, + { + "bbox": [ + 313, + 265, + 555, + 363 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{H}}_i" + }, + { + "bbox": [ + 313, + 265, + 555, + 363 + ], + "type": "text", + "content": " calculated by Eq. (9) as denoising diffusion representations of the sample " + }, + { + "bbox": [ + 313, + 265, + 555, + 363 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 265, + 555, + 363 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 373, + 398, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 398, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 398, + 388 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 394, + 555, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 394, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 555, + 502 + ], + "type": "text", + "content": "Our proposed FedDifRC leverage conditional and denoising diffusion representations to obtain diverse class-relevant knowledge and consistent convergence signals by two complementary components: Text-driven Diffusion Contrastive Learning in Sec. 4.1 and Noise-driven Diffusion Consistent Restriction in Sec. 4.2. Further, we discuss how to extend FedDifRC to a self-supervised scheme (Sec. 4.3). Finally, we provide a theoretical analysis of FedDifRC under nonconvex case (Sec. 4.4). Illustration of FedDifRC in Fig. 5." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 510, + 499, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 499, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 499, + 523 + ], + "type": "text", + "content": "4.1. Text-Driven Diffusion Contrasting" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 527, + 555, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 527, + 555, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 527, + 555, + 664 + ], + "type": "text", + "content": "Inspired by success of contrastive learning [1, 9, 26], based on Sec. 3.3, conditional diffusion representations " + }, + { + "bbox": [ + 313, + 527, + 555, + 664 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{F}}_i" + }, + { + "bbox": [ + 313, + 527, + 555, + 664 + ], + "type": "text", + "content": " of sample " + }, + { + "bbox": [ + 313, + 527, + 555, + 664 + ], + "type": "inline_equation", + "content": "\\{x_i,y_i\\} \\in \\mathcal{D}_k" + }, + { + "bbox": [ + 313, + 527, + 555, + 664 + ], + "type": "text", + "content": " can be obtained with the condition " + }, + { + "bbox": [ + 313, + 527, + 555, + 664 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i" + }, + { + "bbox": [ + 313, + 527, + 555, + 664 + ], + "type": "text", + "content": " and the text prompt " + }, + { + "bbox": [ + 313, + 527, + 555, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{y_i} = \\text{'a photo of a } S_{y_i}" + }, + { + "bbox": [ + 313, + 527, + 555, + 664 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 313, + 527, + 555, + 664 + ], + "type": "inline_equation", + "content": "S_{y_i}" + }, + { + "bbox": [ + 313, + 527, + 555, + 664 + ], + "type": "text", + "content": " is the class name of the label " + }, + { + "bbox": [ + 313, + 527, + 555, + 664 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 313, + 527, + 555, + 664 + ], + "type": "text", + "content": " . For each sample's embedding " + }, + { + "bbox": [ + 313, + 527, + 555, + 664 + ], + "type": "inline_equation", + "content": "z_{i} = h_{k}(x_{i})\\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 313, + 527, + 555, + 664 + ], + "type": "text", + "content": " , we enforce " + }, + { + "bbox": [ + 313, + 527, + 555, + 664 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 313, + 527, + 555, + 664 + ], + "type": "text", + "content": " to be similar to respective positive pair " + }, + { + "bbox": [ + 313, + 527, + 555, + 664 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{F}}_i" + }, + { + "bbox": [ + 313, + 527, + 555, + 664 + ], + "type": "text", + "content": " and dissimilar to negative sample pairs " + }, + { + "bbox": [ + 313, + 527, + 555, + 664 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{F}}_{\\mathcal{N}_i}" + }, + { + "bbox": [ + 313, + 527, + 555, + 664 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 527, + 555, + 664 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{F}}_{\\mathcal{N}_i}" + }, + { + "bbox": [ + 313, + 527, + 555, + 664 + ], + "type": "text", + "content": " means the conditional diffusion representations with the same condition " + }, + { + "bbox": [ + 313, + 527, + 555, + 664 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i" + }, + { + "bbox": [ + 313, + 527, + 555, + 664 + ], + "type": "text", + "content": " and the unmatched text prompts " + }, + { + "bbox": [ + 313, + 527, + 555, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_j = \\text{'a photo of a } S_j" + }, + { + "bbox": [ + 313, + 527, + 555, + 664 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 527, + 555, + 664 + ], + "type": "inline_equation", + "content": "S_{j}\\in S" + }, + { + "bbox": [ + 313, + 527, + 555, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 527, + 555, + 664 + ], + "type": "inline_equation", + "content": "S_{j}\\neq S_{y_{i}}" + }, + { + "bbox": [ + 313, + 527, + 555, + 664 + ], + "type": "text", + "content": "). Then, we define the similarity between " + }, + { + "bbox": [ + 313, + 527, + 555, + 664 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 313, + 527, + 555, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 527, + 555, + 664 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{F}}_i" + }, + { + "bbox": [ + 313, + 527, + 555, + 664 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 672, + 555, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 555, + 714 + ], + "type": "interline_equation", + "content": "s \\left(z _ {i}, \\widetilde {\\mathcal {F}} _ {i}\\right) = \\frac {z _ {i} \\cdot \\widetilde {\\mathcal {F}} _ {i}}{\\mathcal {U} \\left(\\left\\| z _ {i} \\right\\| _ {2} \\times \\left\\| \\widetilde {\\mathcal {F}} _ {i} \\right\\| _ {2}\\right)}, \\mathcal {U} = \\frac {1}{n _ {k}} \\sum_ {m = 1} ^ {n _ {k}} \\| z _ {m} - \\widetilde {\\mathcal {F}} _ {i} \\| _ {2}, \\tag {10}", + "image_path": "f3401fc17808ee0bccd22f304b798ae21ce2a5a01300e1a8b9432a89c813c276.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3730" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 83, + 294, + 223 + ], + "blocks": [ + { + "bbox": [ + 112, + 72, + 240, + 82 + ], + "lines": [ + { + "bbox": [ + 112, + 72, + 240, + 82 + ], + "spans": [ + { + "bbox": [ + 112, + 72, + 240, + 82 + ], + "type": "text", + "content": "Training Process for Local Client" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 83, + 294, + 223 + ], + "lines": [ + { + "bbox": [ + 58, + 83, + 294, + 223 + ], + "spans": [ + { + "bbox": [ + 58, + 83, + 294, + 223 + ], + "type": "image", + "image_path": "3f5f6e3e5ecb3bee521b4d2bd1782a532d785516bcdbb4e1af3e2a984eb420ca.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 232, + 295, + 288 + ], + "lines": [ + { + "bbox": [ + 55, + 232, + 295, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 295, + 288 + ], + "type": "text", + "content": "Figure 5. Illustration of FedDifRC, including two complementary modules: TDCL (Sec. 4.1) builds an inter-contrastive learning strategy by conditional diffusion representations (Sec. 3.3), NDCR (Sec. 4.2) designs a consistent regularization penalty via denoising diffusion representations (Sec. 3.4), jointly promoting FL training." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 305, + 296, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 305, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 305, + 296, + 365 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 305, + 296, + 365 + ], + "type": "inline_equation", + "content": "n_k" + }, + { + "bbox": [ + 55, + 305, + 296, + 365 + ], + "type": "text", + "content": " is total number of samples at client " + }, + { + "bbox": [ + 55, + 305, + 296, + 365 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 305, + 296, + 365 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 305, + 296, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 55, + 305, + 296, + 365 + ], + "type": "text", + "content": " is a normalization factor for similarity, where we set " + }, + { + "bbox": [ + 55, + 305, + 296, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 55, + 305, + 296, + 365 + ], + "type": "text", + "content": " to denote the average distance between " + }, + { + "bbox": [ + 55, + 305, + 296, + 365 + ], + "type": "inline_equation", + "content": "z_m|_{m=1}^{n_k}" + }, + { + "bbox": [ + 55, + 305, + 296, + 365 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 305, + 296, + 365 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{F}}_i" + }, + { + "bbox": [ + 55, + 305, + 296, + 365 + ], + "type": "text", + "content": " on dataset " + }, + { + "bbox": [ + 55, + 305, + 296, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_k" + }, + { + "bbox": [ + 55, + 305, + 296, + 365 + ], + "type": "text", + "content": ". Next, we introduce TDCL to quantify and regulate the local model training, which can be formulated as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 65, + 369, + 295, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 369, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 65, + 369, + 295, + 403 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {T D C L} = \\log (1 + \\frac {\\sum_ {\\widetilde {\\mathcal {F}} _ {j} \\in \\widetilde {\\mathcal {F}} _ {\\mathcal {N} _ {i}}} \\exp \\left(s \\left(z _ {i} , \\widetilde {\\mathcal {F}} _ {j}\\right) / \\tau\\right)}{\\exp \\left(s \\left(z _ {i} , \\widetilde {\\mathcal {F}} _ {i}\\right) / \\tau\\right)}), \\tag {11}", + "image_path": "2ab9d786c0da94c898b160df75bc7182a36564d11ec76b2b6da0d17a6097639f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 407, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 296, + 456 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 407, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 407, + 296, + 456 + ], + "type": "text", + "content": " denotes a temperature parameter to control representation strength [1]. By minimizing Eq. (11), for client " + }, + { + "bbox": [ + 55, + 407, + 296, + 456 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 407, + 296, + 456 + ], + "type": "text", + "content": ", the local model brings each sample's embedding " + }, + { + "bbox": [ + 55, + 407, + 296, + 456 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 55, + 407, + 296, + 456 + ], + "type": "text", + "content": " closer to the positive pair and away from other negative pairs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 460, + 257, + 473 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 257, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 257, + 473 + ], + "type": "text", + "content": "4.2. Noise-Driven Diffusion Regularization" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 478, + 296, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 478, + 296, + 622 + ], + "spans": [ + { + "bbox": [ + 55, + 478, + 296, + 622 + ], + "type": "text", + "content": "Despite each local model aligning sample's embedding with conditional diffusion representations by optimizing " + }, + { + "bbox": [ + 55, + 478, + 296, + 622 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{TDCL}" + }, + { + "bbox": [ + 55, + 478, + 296, + 622 + ], + "type": "text", + "content": " in Eq. (11). Since conditional diffusion representations are dynamically built from text-driven cross-modal matching and depend on generated condition " + }, + { + "bbox": [ + 55, + 478, + 296, + 622 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i" + }, + { + "bbox": [ + 55, + 478, + 296, + 622 + ], + "type": "text", + "content": " at each round, which could not offer a stable convergence direction. Therefore, we aim to leverage denoising diffusion representations " + }, + { + "bbox": [ + 55, + 478, + 296, + 622 + ], + "type": "inline_equation", + "content": "(\\widetilde{\\mathcal{H}}_i)" + }, + { + "bbox": [ + 55, + 478, + 296, + 622 + ], + "type": "text", + "content": " in Eq. (9)) to devise a consistent regularization penalty, constraining model's optimization region in the feature space. We design Noise-driven Diffusion Consistent Restriction (NDCR) and utilize a regularization term to pull the sample's embedding " + }, + { + "bbox": [ + 55, + 478, + 296, + 622 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 55, + 478, + 296, + 622 + ], + "type": "text", + "content": " closer to the respective " + }, + { + "bbox": [ + 55, + 478, + 296, + 622 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{H}}_i" + }, + { + "bbox": [ + 55, + 478, + 296, + 622 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 109, + 628, + 295, + 660 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 628, + 295, + 660 + ], + "spans": [ + { + "bbox": [ + 109, + 628, + 295, + 660 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {N D C R} = \\sum_ {q = 1} ^ {d} \\left(z _ {i (q)} - \\widetilde {\\mathcal {H}} _ {i (q)}\\right) ^ {2}, \\tag {12}", + "image_path": "df3b2a7126686d4b40e8ee616759edc8da4f76f650a6dde675eceed2e9816d97.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " indexes the dimension " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " of the embedding vector. Based on Eq. (12), we expect to align the sample's embedding with the corresponding denoising diffusion representations, enforcing semantics uniformity of the feature space." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 554, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 121 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 121 + ], + "type": "text", + "content": "Besides, we construct CrossEntropy loss and use the logits output vector " + }, + { + "bbox": [ + 313, + 72, + 554, + 121 + ], + "type": "inline_equation", + "content": "l_{i} = g(z_{i}) = f(x_{i})" + }, + { + "bbox": [ + 313, + 72, + 554, + 121 + ], + "type": "text", + "content": " with the original label " + }, + { + "bbox": [ + 313, + 72, + 554, + 121 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 313, + 72, + 554, + 121 + ], + "type": "text", + "content": " to optimize the local discriminative ability for each client. Formally, given the sample pair " + }, + { + "bbox": [ + 313, + 72, + 554, + 121 + ], + "type": "inline_equation", + "content": "\\{x_{i},y_{i}\\}" + }, + { + "bbox": [ + 313, + 72, + 554, + 121 + ], + "type": "text", + "content": " to calculate " + }, + { + "bbox": [ + 313, + 72, + 554, + 121 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CE}" + }, + { + "bbox": [ + 313, + 72, + 554, + 121 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 374, + 126, + 555, + 140 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 126, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 374, + 126, + 555, + 140 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C E} = - \\mathbf {1} _ {y _ {i}} \\log (\\psi (f (x _ {i}))), \\tag {13}", + "image_path": "2cfdfefbee0a5d9e47fec2613ecb7c6e56cdeb6c51fc903fbb211d34cae53163.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 145, + 555, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 145, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 313, + 145, + 555, + 194 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 145, + 555, + 194 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 313, + 145, + 555, + 194 + ], + "type": "text", + "content": " denotes the Softmax function, " + }, + { + "bbox": [ + 313, + 145, + 555, + 194 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 145, + 555, + 194 + ], + "type": "text", + "content": " is the local model with parameters " + }, + { + "bbox": [ + 313, + 145, + 555, + 194 + ], + "type": "inline_equation", + "content": "\\{u,v\\}" + }, + { + "bbox": [ + 313, + 145, + 555, + 194 + ], + "type": "text", + "content": " maps " + }, + { + "bbox": [ + 313, + 145, + 555, + 194 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 145, + 555, + 194 + ], + "type": "text", + "content": " into a " + }, + { + "bbox": [ + 313, + 145, + 555, + 194 + ], + "type": "inline_equation", + "content": "|\\mathcal{C}|" + }, + { + "bbox": [ + 313, + 145, + 555, + 194 + ], + "type": "text", + "content": "-dim output logits " + }, + { + "bbox": [ + 313, + 145, + 555, + 194 + ], + "type": "inline_equation", + "content": "l_{i} = f(x_{i}) = g(h(x_{i}))" + }, + { + "bbox": [ + 313, + 145, + 555, + 194 + ], + "type": "text", + "content": ". Finally, for the " + }, + { + "bbox": [ + 313, + 145, + 555, + 194 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 145, + 555, + 194 + ], + "type": "text", + "content": "-th client, we carry out the following optimization objective in local training:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 318, + 200, + 555, + 213 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 200, + 555, + 213 + ], + "spans": [ + { + "bbox": [ + 318, + 200, + 555, + 213 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {T D C L} (\\mathrm {E q .} (1 1)) + \\mathcal {L} _ {N D C R} (\\mathrm {E q .} (1 2)) + \\mathcal {L} _ {C E}. \\tag {14}", + "image_path": "9e9f5e8b6db574c128d166bd34cdceaf325183dc8c978d1889f3031b2ed18644.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 218, + 533, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 218, + 533, + 231 + ], + "spans": [ + { + "bbox": [ + 313, + 218, + 533, + 231 + ], + "type": "text", + "content": "The algorithm of FedDifRC is given in Appendix A." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 236, + 515, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 515, + 249 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 515, + 249 + ], + "type": "text", + "content": "4.3. Expanding to Self-Supervised Scheme" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 254, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 254, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 254, + 555, + 437 + ], + "type": "text", + "content": "FedDifRC can be extended to the self-supervised scheme without relying on any labeled data source. Specifically, for " + }, + { + "bbox": [ + 313, + 254, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{TDCL}" + }, + { + "bbox": [ + 313, + 254, + 555, + 437 + ], + "type": "text", + "content": " in Eq. (11), we only need to modify text template slightly: i) we set the text prompt " + }, + { + "bbox": [ + 313, + 254, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{y_i} = \\text{'a photo of a similar object'}" + }, + { + "bbox": [ + 313, + 254, + 555, + 437 + ], + "type": "text", + "content": " in Eq. (7) to generate " + }, + { + "bbox": [ + 313, + 254, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{F}}_i" + }, + { + "bbox": [ + 313, + 254, + 555, + 437 + ], + "type": "text", + "content": " as the positive pair; ii) then we randomly select from the set of all class names in TinyImageNet [21] to get " + }, + { + "bbox": [ + 313, + 254, + 555, + 437 + ], + "type": "inline_equation", + "content": "S_{\\mathcal{N}_i}" + }, + { + "bbox": [ + 313, + 254, + 555, + 437 + ], + "type": "text", + "content": ", and set " + }, + { + "bbox": [ + 313, + 254, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{y_i} = \\text{'a photo of a } S_{\\mathcal{N}_i}" + }, + { + "bbox": [ + 313, + 254, + 555, + 437 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 313, + 254, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{F}}_{\\mathcal{N}_i}" + }, + { + "bbox": [ + 313, + 254, + 555, + 437 + ], + "type": "text", + "content": " as negative pairs. For the " + }, + { + "bbox": [ + 313, + 254, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{NDCR}" + }, + { + "bbox": [ + 313, + 254, + 555, + 437 + ], + "type": "text", + "content": " in Eq. (12), we set " + }, + { + "bbox": [ + 313, + 254, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{y_i} = \\text{'a photo of a visual object'}" + }, + { + "bbox": [ + 313, + 254, + 555, + 437 + ], + "type": "text", + "content": " in Eq. (9) to get " + }, + { + "bbox": [ + 313, + 254, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{H}}_i" + }, + { + "bbox": [ + 313, + 254, + 555, + 437 + ], + "type": "text", + "content": " as consistent signal to align the sample's embedding. Note that it is also possible to use class names of other datasets (e.g., CIFAR100 [20]) to generate " + }, + { + "bbox": [ + 313, + 254, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{F}}_{\\mathcal{N}_i}" + }, + { + "bbox": [ + 313, + 254, + 555, + 437 + ], + "type": "text", + "content": " as negative pairs. By this self-supervised way, the objective term of local training becomes " + }, + { + "bbox": [ + 313, + 254, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{TDCL} + \\mathcal{L}_{NDCR}" + }, + { + "bbox": [ + 313, + 254, + 555, + 437 + ], + "type": "text", + "content": ". The results in Table. 5 validate the effectiveness of FedDifRC's self-supervised scheme." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 444, + 440, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 440, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 440, + 456 + ], + "type": "text", + "content": "4.4. Convergence Analysis" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 461, + 554, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 554, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 554, + 496 + ], + "type": "text", + "content": "We proved convergence of FedDifRC in non-convex case. We denote " + }, + { + "bbox": [ + 313, + 461, + 554, + 496 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 313, + 461, + 554, + 496 + ], + "type": "text", + "content": " (Eq. (14)) as " + }, + { + "bbox": [ + 313, + 461, + 554, + 496 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_r" + }, + { + "bbox": [ + 313, + 461, + 554, + 496 + ], + "type": "text", + "content": " with the round " + }, + { + "bbox": [ + 313, + 461, + 554, + 496 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 461, + 554, + 496 + ], + "type": "text", + "content": ", assumptions (Appendix B.2) are similar to existing works [6, 29, 30, 69]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 497, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 554, + 521 + ], + "type": "text", + "content": "Convergence Rate of FedDifRC: Let the round " + }, + { + "bbox": [ + 313, + 497, + 554, + 521 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 497, + 554, + 521 + ], + "type": "text", + "content": " from 0 to " + }, + { + "bbox": [ + 313, + 497, + 554, + 521 + ], + "type": "inline_equation", + "content": "R - 1" + }, + { + "bbox": [ + 313, + 497, + 554, + 521 + ], + "type": "text", + "content": ", given any " + }, + { + "bbox": [ + 313, + 497, + 554, + 521 + ], + "type": "inline_equation", + "content": "\\xi > 0" + }, + { + "bbox": [ + 313, + 497, + 554, + 521 + ], + "type": "text", + "content": ", the " + }, + { + "bbox": [ + 313, + 497, + 554, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 313, + 497, + 554, + 521 + ], + "type": "text", + "content": " will converge when," + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 365, + 526, + 553, + 553 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 526, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 365, + 526, + 553, + 553 + ], + "type": "interline_equation", + "content": "R > \\frac {2 \\left(\\mathcal {L} _ {0} - \\mathcal {L} ^ {*}\\right)}{\\xi E \\eta \\left(2 - L _ {1} \\eta\\right) - \\Omega_ {1} - \\Omega_ {2}}, \\tag {15}", + "image_path": "69706a6f40429e4cf134252791879fbc82da83c315901724694b4fa5233316a1.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 557, + 527, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 527, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 527, + 569 + ], + "type": "text", + "content": "and we can further get the following condition for " + }, + { + "bbox": [ + 313, + 557, + 527, + 569 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 557, + 527, + 569 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 378, + 573, + 553, + 600 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 573, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 378, + 573, + 553, + 600 + ], + "type": "interline_equation", + "content": "\\eta < \\frac {2 \\xi - 2 (| \\mathcal {C} | - 1) L _ {2} B}{L _ {1} (\\xi + \\sigma^ {2})}, \\tag {16}", + "image_path": "646ec833563d81af924e511f018d6897bae2c9fcdb5061e1565d65e18672e554.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\Omega_{1} = 2(|\\mathcal{C}| - 1)L_{2}E\\eta B" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\Omega_{2} = L_{1}E\\eta^{2}\\sigma^{2}" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^*" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": " denotes the optimal solution of " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\{L_1,L_2,B,\\sigma^2\\}" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": " are constants in the assumptions from Appendix B.2, and " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "|\\mathcal{C}|" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": " as the number of classes, " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": " as local epochs. Eq. (15) and Eq. (16) ensure convergence rate of FedDifRC, after selecting the number of rounds " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": " and the learning rate " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": ". So the smaller " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": " is, the larger " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": " is, which means that the tighter the bound is, the more rounds " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": " is required. The detailed assumptions and proofs are formally provided in Appendix B." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "3731" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 295, + 193 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 295, + 193 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 295, + 193 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 295, + 193 + ], + "type": "table", + "html": "
TDCLNDCRCIFAR10 (keep ρ = 1.0)
NID10.05NID10.2NID10.5NID2AVG
78.2784.6586.1172.6080.41
✓81.3986.0388.1675.6782.81
✓80.3586.4087.5475.3382.40
✓✓83.1488.2789.3176.4584.29
TDCLNDCRCIFAR10-LT (keep NID10.2)
ρ = 10ρ = 50ρ = 100ρ = 200AVG
75.9463.2552.9749.6960.46
✓76.6264.6653.6250.3361.31
✓76.1964.7554.3950.6761.50
✓✓77.5865.3055.0151.1262.25
", + "image_path": "6e31980dfb4ef5d2b7c98e7db683f60e661ea86f1bc920a4fd5270958307c603.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 225, + 295, + 281 + ], + "blocks": [ + { + "bbox": [ + 56, + 201, + 293, + 211 + ], + "lines": [ + { + "bbox": [ + 56, + 201, + 293, + 211 + ], + "spans": [ + { + "bbox": [ + 56, + 201, + 293, + 211 + ], + "type": "text", + "content": "Table 1. Ablation study for two key modules: TDCL and NDCR." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 225, + 295, + 281 + ], + "lines": [ + { + "bbox": [ + 58, + 225, + 295, + 281 + ], + "spans": [ + { + "bbox": [ + 58, + 225, + 295, + 281 + ], + "type": "table", + "html": "
ScenariosL=1L=2L=3L=4Fused△
NID10.0580.2182.6082.1981.3383.14+0.54
NID10.286.1587.2887.8185.9288.27+0.46
NID10.587.1089.0388.9387.1789.31+0.28
NID274.2675.7375.6175.0076.45+0.72
", + "image_path": "9d2bf7a622cd7528623db9e232068555b1e30f14385edbb8f02c9f4830125a85.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 289, + 295, + 323 + ], + "lines": [ + { + "bbox": [ + 55, + 289, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 295, + 323 + ], + "type": "text", + "content": "Table 2. Comparison of features from different layers " + }, + { + "bbox": [ + 55, + 289, + 295, + 323 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 289, + 295, + 323 + ], + "type": "text", + "content": " and fused features " + }, + { + "bbox": [ + 55, + 289, + 295, + 323 + ], + "type": "inline_equation", + "content": "\\left( {L \\in \\{ 2,3,4\\} }\\right)" + }, + { + "bbox": [ + 55, + 289, + 295, + 323 + ], + "type": "text", + "content": " in Eq. (8) and Eq. (9) to obtain " + }, + { + "bbox": [ + 55, + 289, + 295, + 323 + ], + "type": "inline_equation", + "content": "{\\widetilde{\\mathcal{F}}}_{i}" + }, + { + "bbox": [ + 55, + 289, + 295, + 323 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 289, + 295, + 323 + ], + "type": "inline_equation", + "content": "{\\widetilde{\\mathcal{H}}}_{i}" + }, + { + "bbox": [ + 55, + 289, + 295, + 323 + ], + "type": "text", + "content": " , best in bold and second with underline, " + }, + { + "bbox": [ + 55, + 289, + 295, + 323 + ], + "type": "inline_equation", + "content": "\\bigtriangleup" + }, + { + "bbox": [ + 55, + 289, + 295, + 323 + ], + "type": "text", + "content": " means accuracy gap." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 56, + 335, + 178, + 430 + ], + "blocks": [ + { + "bbox": [ + 56, + 335, + 178, + 430 + ], + "lines": [ + { + "bbox": [ + 56, + 335, + 178, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 335, + 178, + 430 + ], + "type": "image", + "image_path": "517db9bf2184c3a176aea6e8b9629536e038b6f96eba6ac97f1de245771dfd26.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 437, + 295, + 461 + ], + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 461 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 461 + ], + "type": "text", + "content": "Figure 6. Analysis of FedDifRC with time-steps " + }, + { + "bbox": [ + 55, + 437, + 295, + 461 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 437, + 295, + 461 + ], + "type": "text", + "content": " in Eq. (9) (left) and temperature " + }, + { + "bbox": [ + 55, + 437, + 295, + 461 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 437, + 295, + 461 + ], + "type": "text", + "content": " in Eq. (11) (right) on CIFAR10 with NID10.2." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 178, + 335, + 295, + 430 + ], + "blocks": [ + { + "bbox": [ + 178, + 335, + 295, + 430 + ], + "lines": [ + { + "bbox": [ + 178, + 335, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 178, + 335, + 295, + 430 + ], + "type": "image", + "image_path": "ebd4fbe975aeb5c865726fdcd36c6ab2d7f042b84bcff6ce3b41d9248560b9ac.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 483, + 137, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 137, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 137, + 496 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 503, + 172, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 172, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 172, + 517 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 521, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 582 + ], + "type": "text", + "content": "Datasets: We adopt popular benchmark datasets, including CIFAR10 & CIFAR100 [20], TinyImageNet [21]; we build unbalanced versions of these datasets by [15]: CIFAR10-LT & CIFAR100-LT, and TinyImageNet-LT; we use Digits [41] and Office-Caltech [19] for the domain shift task in FL." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": "Scenarios: i) label shift: we consider two non-iid settings: " + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{NID1}_{\\alpha}" + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": " follows Dirichlet distribution [59], where " + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": " denotes the heterogeneity level; NID2 is a more extreme setting consists of 6 biased clients (each has only a single category) and 1 unbiased client has all categories. ii) imbalance shift: we shape the original dataset into a long-tailed distribution with " + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": " follow [15], " + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": " means the ratio between sample sizes of the most frequent and least frequent class, as " + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\rho = \\max_{j} / \\min_{j}" + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": ". iii) domain shift: Digits [41] includes 4 domains as MNIST (M), USPS (U), SVHN (SV), SYN (SY) with 10 categories. Office-Caltech [19] also consists 4 domains as Caltech (C)," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "type": "text", + "content": "Webcam (W), Amazon (A), DSLR (D) with 10 overlapping categories. We initialize 10 clients and assign them domains randomly: Digits (M: 1, U: 4, SV: 3, SY: 2), Office-Caltech (C: 3, W: 2, A: 2, D: 3). Each client's dataset is randomly selected from domains (Digits: " + }, + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "type": "text", + "content": ", Office-Caltech: " + }, + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 132, + 553, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 132, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 132, + 553, + 227 + ], + "type": "text", + "content": "Configurations: The global rounds " + }, + { + "bbox": [ + 313, + 132, + 553, + 227 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 132, + 553, + 227 + ], + "type": "text", + "content": " as 100 and clients " + }, + { + "bbox": [ + 313, + 132, + 553, + 227 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 132, + 553, + 227 + ], + "type": "text", + "content": " as 10, all methods have little or no gain with more rounds " + }, + { + "bbox": [ + 313, + 132, + 553, + 227 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 132, + 553, + 227 + ], + "type": "text", + "content": ". To facilitate fair comparison, we follow in [26, 29], and use SGD optimizer to update models with a learning rate 0.01, a momentum of 0.9, a batch size of 64, the weight decay of " + }, + { + "bbox": [ + 313, + 132, + 553, + 227 + ], + "type": "inline_equation", + "content": "10^{-5}" + }, + { + "bbox": [ + 313, + 132, + 553, + 227 + ], + "type": "text", + "content": ", and local epochs " + }, + { + "bbox": [ + 313, + 132, + 553, + 227 + ], + "type": "inline_equation", + "content": "E = 10" + }, + { + "bbox": [ + 313, + 132, + 553, + 227 + ], + "type": "text", + "content": ". We use MobileNetV2 [48] for TinyImageNet, ResNet-10 [14] for others. The dimension " + }, + { + "bbox": [ + 313, + 132, + 553, + 227 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 132, + 553, + 227 + ], + "type": "text", + "content": " as 512, " + }, + { + "bbox": [ + 313, + 132, + 553, + 227 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 132, + 553, + 227 + ], + "type": "text", + "content": " (Eq. (11)) as 0.06, " + }, + { + "bbox": [ + 313, + 132, + 553, + 227 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 132, + 553, + 227 + ], + "type": "text", + "content": " (Eq. (9)) as 150." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 228, + 553, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 553, + 288 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 553, + 288 + ], + "type": "text", + "content": "Baselines: We compare our against several SOTA methods: standard-based FedAvg [34], FedProx [29]; contrast-based MOON [26], FedRCL [50]; aggregate-based FedCDA [60], FedDisco [69]; unsupervised FedEMA [77], FedLID [42]; prototype-based FedProto [52], FedNH [70]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 297, + 427, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 427, + 309 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 427, + 309 + ], + "type": "text", + "content": "5.2. Validation Analysis" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 314, + 553, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 553, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 553, + 411 + ], + "type": "text", + "content": "We give ablation study in Table. 1: i) TDCL brings significant performance improvements over the baseline, showing that TDCL is able to promote semantic spread-out property; ii) NDCR also yields solid gains, proving the importance of aligning instance embedding via consistency regularization; iii) combining TDCL and NDCR attain better performance, supporting our motivation to unleash meaningful guidance of diffusion representations to promote FL local training." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 411, + 554, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 554, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 554, + 604 + ], + "type": "text", + "content": "Then, we prove the advantage of fused features (Eq. (8)). As shown in Table. 2, results on CIFAR10 reveal that fused features capture both semantics and local details, better than features on a single layer " + }, + { + "bbox": [ + 313, + 411, + 554, + 604 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 411, + 554, + 604 + ], + "type": "text", + "content": ", and if the features with too low resolution (" + }, + { + "bbox": [ + 313, + 411, + 554, + 604 + ], + "type": "inline_equation", + "content": "L = 1" + }, + { + "bbox": [ + 313, + 411, + 554, + 604 + ], + "type": "text", + "content": ") or over-matching (" + }, + { + "bbox": [ + 313, + 411, + 554, + 604 + ], + "type": "inline_equation", + "content": "L = 4" + }, + { + "bbox": [ + 313, + 411, + 554, + 604 + ], + "type": "text", + "content": ") would damage the model's performance. To explore the influence of different time-steps " + }, + { + "bbox": [ + 313, + 411, + 554, + 604 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 411, + 554, + 604 + ], + "type": "text", + "content": " (Eq. (9)), we conduct experiments on CIFAR10 with NID10.2 in Fig. 6 left, we add different amount of noise to get " + }, + { + "bbox": [ + 313, + 411, + 554, + 604 + ], + "type": "inline_equation", + "content": "\\tilde{x}_i^t" + }, + { + "bbox": [ + 313, + 411, + 554, + 604 + ], + "type": "text", + "content": " in Eq. (9). We find that FedDifRC is robust to the choice of " + }, + { + "bbox": [ + 313, + 411, + 554, + 604 + ], + "type": "inline_equation", + "content": "t \\in [100, 300]" + }, + { + "bbox": [ + 313, + 411, + 554, + 604 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 411, + 554, + 604 + ], + "type": "inline_equation", + "content": "t = 150" + }, + { + "bbox": [ + 313, + 411, + 554, + 604 + ], + "type": "text", + "content": " achieves the highest value). As shown in Fig. 6 right, we present results with different " + }, + { + "bbox": [ + 313, + 411, + 554, + 604 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 411, + 554, + 604 + ], + "type": "text", + "content": " in Eq. (11), which reveals that FedDifRC is not sensitive to the " + }, + { + "bbox": [ + 313, + 411, + 554, + 604 + ], + "type": "inline_equation", + "content": "\\tau \\in [0.04, 0.08]" + }, + { + "bbox": [ + 313, + 411, + 554, + 604 + ], + "type": "text", + "content": ", indicating the effectiveness of " + }, + { + "bbox": [ + 313, + 411, + 554, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 313, + 411, + 554, + 604 + ], + "type": "text", + "content": " (Eq. (10)), to balance relative distances among different samples. The accuracy no longer increases after " + }, + { + "bbox": [ + 313, + 411, + 554, + 604 + ], + "type": "inline_equation", + "content": "\\tau = 0.06" + }, + { + "bbox": [ + 313, + 411, + 554, + 604 + ], + "type": "text", + "content": ". Extra results of " + }, + { + "bbox": [ + 313, + 411, + 554, + 604 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 411, + 554, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 411, + 554, + 604 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 411, + 554, + 604 + ], + "type": "text", + "content": " are provided in Appendix C.2.5." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 611, + 425, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 425, + 623 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 425, + 623 + ], + "type": "text", + "content": "5.3. Evaluation Results" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "content": "The results on different settings (NID1, NID2, " + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "content": ") in Table. 3, FedDifRC consistently outperforms these counterparts, it confirms that FedDifRC effectively leverages the inherent fine-grained visual representation capabilities of diffusion models to promote heterogeneous FL training. The results of Office-Caltech with domain shift are shown in Table. 4, indicating that FedDifRC can acquire well-generalizable" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3732" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 553, + 192 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 553, + 192 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 553, + 192 + ], + "type": "table", + "html": "
FL BaselinesCIFAR10CIFAR100TinyImageNetCIFAR10-LT (keep NID10.2)
NID10.05NID10.2NID2NID10.05NID10.2NID2NID10.05NID10.2NID2ρ = 10ρ = 50ρ = 100ρ = 200
FedAvg[AISTAT'17]78.2784.6572.6055.9760.0850.5640.4142.8435.1575.9463.2552.9749.69
FedProx[MLSys'20]78.4284.5972.8156.2760.2150.2940.2042.1635.6276.0263.4352.8649.10
MOON[CVPR'21]80.7986.1073.3556.7961.4851.8140.7943.6336.1175.7363.8053.1749.31
FedProto[AAAI'22]77.8683.9070.8454.4859.2648.6039.2941.0234.6874.3162.2350.3747.26
FedNH[AAAI'23]80.2585.8273.6556.6861.4551.3140.2543.3036.2176.1563.3652.4948.92
FedDisco[ICML'23]81.3586.6374.7257.3461.7952.0540.8343.9237.6976.2764.4053.3349.75
FedCDA[ICLR'24]81.7086.8975.0757.7662.1752.3541.1844.0938.2476.4564.3053.8550.08
FedRCL[CVPR'24]82.0287.1175.1357.9862.0552.4941.3344.2138.8676.3464.1753.4149.91
FedDifRC83.1488.2776.4559.2263.3253.8742.6345.2840.0577.5865.3055.0151.12
", + "image_path": "0e4ef91db3f3354ab7d51682e7c24789cad36c92892f859a6778e609c61cdac2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 231, + 294, + 308 + ], + "blocks": [ + { + "bbox": [ + 58, + 200, + 550, + 212 + ], + "lines": [ + { + "bbox": [ + 58, + 200, + 550, + 212 + ], + "spans": [ + { + "bbox": [ + 58, + 200, + 550, + 212 + ], + "type": "text", + "content": "Table 3. Comparison results on several settings (label shift and imbalance shift) and datasets, the best in bold and second with underline." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 231, + 294, + 308 + ], + "lines": [ + { + "bbox": [ + 59, + 231, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 59, + 231, + 294, + 308 + ], + "type": "table", + "html": "
FL BaselinesOffice-Caltech
CaltechWebcamAmazonDSLRAVG
FedDisco[ICML'23]62.3945.9577.3145.2057.71
FedCDA[ICLR'24]62.2546.5376.5846.1457.87
FedRCL[CVPR'24]60.1342.7675.6043.7755.56
FedDifRC63.3250.1978.6748.1260.08
", + "image_path": "e1ada7d8e5795b2fe78c76f3d38b250fe3f61e0243cf2a9d9de9a6656f3e36b2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 59, + 342, + 294, + 407 + ], + "blocks": [ + { + "bbox": [ + 56, + 316, + 293, + 327 + ], + "lines": [ + { + "bbox": [ + 56, + 316, + 293, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 293, + 327 + ], + "type": "text", + "content": "Table 4. Comparison results on Office-Caltech with domain shift." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 342, + 294, + 407 + ], + "lines": [ + { + "bbox": [ + 59, + 342, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 59, + 342, + 294, + 407 + ], + "type": "table", + "html": "
FL BaselinesClientsCIFAR100
NID10.05NID10.2NID10.5NID2
FedEMA[ICLR'22]K=1026.9029.4130.6421.52
FedLID[ICCV'23]K=1028.8430.4331.2223.65
FedDifRCK=1030.6131.4532.3025.73
", + "image_path": "a06ec8bf1d833be452a56e1ab841a9f8cd5340e9662047ed51ab796ecb4631a0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 58, + 442, + 294, + 494 + ], + "blocks": [ + { + "bbox": [ + 56, + 417, + 293, + 428 + ], + "lines": [ + { + "bbox": [ + 56, + 417, + 293, + 428 + ], + "spans": [ + { + "bbox": [ + 56, + 417, + 293, + 428 + ], + "type": "text", + "content": "Table 5. Comparison results on self-supervised setting (Sec. 4.3)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 442, + 294, + 494 + ], + "lines": [ + { + "bbox": [ + 58, + 442, + 294, + 494 + ], + "spans": [ + { + "bbox": [ + 58, + 442, + 294, + 494 + ], + "type": "table", + "html": "
FL BaselinesCIFAR10CIFAR10-LT (NID102)
NID10.05NID10.2NID2ρ = 10ρ = 50ρ = 100
FedDDA[PRCV-23]82.0586.9474.4976.3163.9553.15
FedDISC[AAAF24]82.3387.1575.4276.6064.2354.07
FedRCL (♂/†)82.30♂/80.95†87.25♂/86.16†75.37♂/73.85†76.50♂/76.12†64.26♂/62.91†53.65♂/52.20†
FedRCL (♀ + †)81.5286.4874.3576.2263.6952.93
FedDifRC83.1488.2776.4577.5865.3055.01
", + "image_path": "1245e05e977f82990b29d227f4c15e5ffb1869b4caa32d6a8d2010474197ef62.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 58, + 529, + 294, + 597 + ], + "blocks": [ + { + "bbox": [ + 56, + 504, + 294, + 514 + ], + "lines": [ + { + "bbox": [ + 56, + 504, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 294, + 514 + ], + "type": "text", + "content": "Table 6. More comparison results on CIFAR10 and CIFAR10-LT." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 529, + 294, + 597 + ], + "lines": [ + { + "bbox": [ + 58, + 529, + 294, + 597 + ], + "spans": [ + { + "bbox": [ + 58, + 529, + 294, + 597 + ], + "type": "table", + "html": "
FL BaselinesHAM10000 [56]EuroSAT [40]
NID10.2NID10.5NID2NID10.2NID10.5NID2
FedDisco[ICML'23]50.2653.6844.6378.7682.3670.34
FedCDA[ICLR'24]51.6554.0545.1680.4583.3871.10
FedRCL[CVPR'24]52.2154.2045.4879.8083.6970.73
FedDifRC53.7955.6047.1782.4986.0573.14
", + "image_path": "7835b4b5a3eb1b642cf2d953750fbcacff39693e827201a7a9b4ce91429ad8a7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "type": "text", + "content": "property in the feature space. With self-supervised scenario, in Table. 5, our FedDifRC achieves excellent performance by leveraging the visual perception capability of diffusion models, without requiring any labeled data. Please refer to Appendix C for more details and results (e.g., domain shift, convergence rates, different SD versions, etc)." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 317, + 231, + 553, + 295 + ], + "blocks": [ + { + "bbox": [ + 62, + 606, + 288, + 616 + ], + "lines": [ + { + "bbox": [ + 62, + 606, + 288, + 616 + ], + "spans": [ + { + "bbox": [ + 62, + 606, + 288, + 616 + ], + "type": "text", + "content": "Table 7. Comparison results on the HAM10000 and EuroSAT." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 231, + 553, + 295 + ], + "lines": [ + { + "bbox": [ + 317, + 231, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 317, + 231, + 553, + 295 + ], + "type": "table", + "html": "
Pre-trained ModelsCIFAR10CIFAR10-LT (NID10.2)
NID10.05NID10.2NID2ρ = 10ρ = 50ρ = 100
CLIP-ViT-B/16 [45]79.7284.9173.9774.8962.4450.87
MAE-ViT-B/16 [10]80.7485.3873.2475.4563.2752.55
DINO-ViT-B/16 [33]80.5786.1474.1675.2562.7051.92
FedDifRC83.1488.2776.4577.5865.3055.01
", + "image_path": "d2a3fa09853521fd4ccab6eb97c7c23353de060e0636270c8f12ae20e5825408.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 304, + 551, + 315 + ], + "lines": [ + { + "bbox": [ + 316, + 304, + 551, + 315 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 551, + 315 + ], + "type": "text", + "content": "Table 8. Experiments on aligning features of pre-trained models." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 336, + 553, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 553, + 359 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 553, + 359 + ], + "type": "text", + "content": "To further elaborate on the core idea of our FedDifRC, we then perform the following experiments:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 360, + 555, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 555, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 555, + 598 + ], + "type": "text", + "content": "i) To ensure a fair comparison, we also compare diffusion-based FL methods FedDISC [67] and FedDDA [74], we generate data for each client (same amount) by the SD model as " + }, + { + "bbox": [ + 313, + 360, + 555, + 598 + ], + "type": "inline_equation", + "content": "\\diamond" + }, + { + "bbox": [ + 313, + 360, + 555, + 598 + ], + "type": "text", + "content": ", and align SD features (the image is fed directly into the SD model) by L2 loss as " + }, + { + "bbox": [ + 313, + 360, + 555, + 598 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 313, + 360, + 555, + 598 + ], + "type": "text", + "content": ". Table. 6 shows that directly matching SD features may even hurt the performance, which further illustrates the superiority of our FedDifRC. ii) We focus on visual perception ability of diffusion models to understand an image (e.g., textures and structures) rather than pre-trained prior knowledge. We conduct experiments on medical and remote sensing datasets HAM10000 [56] and EuroSAT [40], which are not crossed in SD pre-training dataset LAION-5B [49]. Table. 7 show that the success of FedDifRC is based on the fine-grained visual representation rather than the large prior knowledge in SD model. iii) The SD model can implicitly provide both high-level and low-level visual concepts of semantic objects, while foundation models focus more on sparse visual matching. The Table. 8 show that FedDifRC can capture underlying properties and correspondences of visual objects." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 609, + 388, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 609, + 388, + 621 + ], + "spans": [ + { + "bbox": [ + 314, + 609, + 388, + 621 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": "In this paper, we aim to exploit the fine-grained visual representation capabilities of diffusion models to handle data heterogeneity in FL. We propose a novel diffusion-inspired FL framework FedDifRC to regulate local training via denoising and conditional diffusion representations. By leveraging complementary strength of TDCL and NDCR, FedDifRC achieves superior performance in heterogeneous FL." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "3733" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "content": "This work is supported by the scholarship from the China Scholarship Council (CSC) while the first author pursues his PhD degree at the University of Wollongong. This work was also partially supported by Australian Research Council Linkage Project LP210300009 and LP230100083." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 115, + 175 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 115, + 175 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 115, + 175 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 182, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 61, + 182, + 296, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 182, + 296, + 226 + ], + "spans": [ + { + "bbox": [ + 61, + 182, + 296, + 226 + ], + "type": "text", + "content": "[1] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, pages 1597-1607, 2020. 2, 3, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 228, + 296, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 228, + 296, + 259 + ], + "spans": [ + { + "bbox": [ + 61, + 228, + 296, + 259 + ], + "type": "text", + "content": "[2] Ching-Yao Chuang, Joshua Robinson, Yen-Chen Lin, Antonio Torralba, et al. Debiased contrastive learning. In NeurIPS, pages 8765-8775, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 261, + 295, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 261, + 295, + 282 + ], + "spans": [ + { + "bbox": [ + 61, + 261, + 295, + 282 + ], + "type": "text", + "content": "[3] Kevin Clark and Priyank Jain. Text-to-image diffusion models are zero shot classifiers. In NeurIPS, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 285, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 285, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 61, + 285, + 294, + 316 + ], + "type": "text", + "content": "[4] Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah. Diffusion models in vision: A survey. IEEE TPAMI, 45(9):10850-10869, 2023. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 319, + 294, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 319, + 294, + 349 + ], + "spans": [ + { + "bbox": [ + 61, + 319, + 294, + 349 + ], + "type": "text", + "content": "[5] Jiequan Cui, Zhisheng Zhong, Shu Liu, Bei Yu, et al. Parametric contrastive learning. In ICCV, pages 715-724, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 352, + 294, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 352, + 294, + 395 + ], + "spans": [ + { + "bbox": [ + 61, + 352, + 294, + 395 + ], + "type": "text", + "content": "[6] Liang Gao, Huazhu Fu, Li Li, Yingwen Chen, et al. Feddc: Federated learning with non-iid data via local drift decoupling and correction. In CVPR, pages 10112-10121, 2022, 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 397, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 397, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 397, + 294, + 430 + ], + "type": "text", + "content": "[7] Jie Gui, Tuo Chen, Jing Zhang, Qiong Cao, et al. A survey on self-supervised learning: Algorithms, applications, and future trends. IEEE TPAMI, 46(12):9052-9071, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 431, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 431, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 62, + 431, + 294, + 464 + ], + "type": "text", + "content": "[8] Sungwon Han, Sungwon Park, Fangzhao Wu, Sundong Kim, et al. Fedx: Unsupervised federated learning with cross knowledge distillation. In ECCV, pages 691-707, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 466, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 466, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 62, + 466, + 294, + 498 + ], + "type": "text", + "content": "[9] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, et al. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9729-9738, 2020. 2, 3, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 499, + 294, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 499, + 294, + 532 + ], + "spans": [ + { + "bbox": [ + 57, + 499, + 294, + 532 + ], + "type": "text", + "content": "[10] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, et al. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 533, + 294, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 533, + 294, + 565 + ], + "spans": [ + { + "bbox": [ + 57, + 533, + 294, + 565 + ], + "type": "text", + "content": "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, pages 6840-6851, 2020. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 567, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 567, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 57, + 567, + 294, + 600 + ], + "type": "text", + "content": "[12] Wenke Huang, Mang Ye, and Bo Du. Learn from others and be yourself in heterogeneous federated learning. In CVPR, pages 10143-10153, 2022. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 601, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 294, + 634 + ], + "type": "text", + "content": "[13] Wenke Huang, Mang Ye, Zekun Shi, He Li, and Bo Du. Rethinking federated learning with domain shift: A prototype view. In CVPR, pages 16312-16322, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 635, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 679 + ], + "type": "text", + "content": "[14] Cheng Ju, Aurélien Bibaut, and Mark van der Laan. The relative performance of ensemble methods with deep convolutional neural networks for image classification. Journal of Applied Statistics, 45:2800-2818, 2018. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "type": "text", + "content": "[15] Cao Kaidi, Colin Wei, Adrien Gaidon, Nikos Arechiga, et al. Learning imbalanced datasets with label-distribution-aware margin loss. In NeurIPS, 2019. 7" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 37, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[16] Peter Kairouz, H Brendan McMahan, Brendan Avent, Aurélien Bellet, et al. Advances and open problems in federated learning. Foundations and Trends in Machine Learning, 14(1):1-210, 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 118, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 160 + ], + "type": "text", + "content": "[17] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank Reddi, et al. Scaffold: Stochastic controlled averaging for federated learning. In ICML, pages 5132-5143, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 163, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 194 + ], + "type": "text", + "content": "[18] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, et al. Supervised contrastive learning. In NeurIPS, pages 18661-18673, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 197, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 228 + ], + "type": "text", + "content": "[19] Boqing Kong, Yuan Shi, Fei Sha, and Kristen Grauman. Geodesic flow kernel for unsupervised domain adaptation. In CVPR, pages 2066-2073, 2012. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 230, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 262 + ], + "type": "text", + "content": "[20] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. In Technical report, University of Toronto, 2009. 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 264, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 285 + ], + "type": "text", + "content": "[21] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7, 2015. 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 287, + 553, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 329 + ], + "type": "text", + "content": "[22] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 332, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 364 + ], + "type": "text", + "content": "[23] Sunwoo Lee, Tuo Zhang, and A Salman Avestimehr. Layerwise adaptive model aggregation for scalable federated learning. In AAAI, pages 8491-8499, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 365, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 398 + ], + "type": "text", + "content": "[24] Alexander C Li, Mihir Prabhudesai, Shivam Duggal, Ellis Brown, et al. Your diffusion model is secretly a zero-shot classifier. In ICCV, pages 2206-2217, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 399, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 441 + ], + "type": "text", + "content": "[25] Daixun Li, Weiying Xie, Zixuan Wang, Yibing Lu, et al. Feddiff: Diffusion model driven federated learning for multimodal and multi-clients. IEEE TCSVT, 34(10):10353-10367, 2024. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 444, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 553, + 475 + ], + "type": "text", + "content": "[26] Qinbin Li, Bingsheng He, and Dawn Song. Model-contrastive federated learning. In CVPR, pages 10713-10722, 2021. 2, 3, 5, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 478, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 510 + ], + "type": "text", + "content": "[27] Qinbin Li, Yiqun Diao, Quan Chen, and Bingsheng He. Federated learning on non-iid data silos: An experimental study. In ICDE, pages 965–978, 2022. 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 512, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 553, + 544 + ], + "type": "text", + "content": "[28] Tian Li, Anit Kumar Sahu, Ameet Talwalkar, and Virginia Smith. Federated learning: Challenges, methods, and future directions. IEEE SPM, 37(3):50-60, 2020. 1" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 545, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 578 + ], + "type": "text", + "content": "[29] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, et al. Federated optimization in heterogeneous networks. MLSys, 2:429-450, 2020. 1, 2, 6, 7" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 580, + 553, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 580, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 316, + 580, + 553, + 611 + ], + "type": "text", + "content": "[30] Xiang Li, Kaixuan Huang, Wenhao Yang, Shusen Wang, et al. On the convergence of fedavg on non-iid data. In ICLR, 2020. 2, 6" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 613, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 645 + ], + "type": "text", + "content": "[31] Grace Luo, Lisa Dunlap, Dong Huk Park, Aleksander Holynski, et al. Diffusion hyperfeatures: Searching through time and space for semantic correspondence. In NeurIPS, 2023. 3" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 316, + 647, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 679 + ], + "type": "text", + "content": "[32] Mi Luo, Fei Chen, Dapeng Hu, Yifan Zhang, et al. No fear of heterogeneity: Classifier calibration for federated learning with non-iid data. In NeurIPS, pages 5972-5984, 2021. 1" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "type": "text", + "content": "[33] Caron Mathilde, Hugo Touvron, Ishan Misra, Hervé Jégou, et al. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. 8" + } + ] + } + ], + "index": 36 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3734" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "type": "text", + "content": "[34] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, et al. Communication-efficient learning of deep networks from decentralized data. In AISTATS, pages 1273-1282, 2017. 1, 2, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 160 + ], + "type": "text", + "content": "[35] Matias Mendieta, Taojiannan Yang, Pu Wang, Minwoo Lee, et al. Local learning matters: Rethinking data heterogeneity in federated learning. In CVPR, pages 8397-8406, 2022. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 295, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 295, + 204 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 295, + 204 + ], + "type": "text", + "content": "[36] John Nguyen, Kshitiz Malik, Hongyuan Zhan, Ashkan Yousefpour, et al. Federated learning with buffered asynchronous aggregation. In AISTATS, pages 3581-3607, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 205, + 295, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 295, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 295, + 237 + ], + "type": "text", + "content": "[37] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, pages 8162-8171, 2021. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 295, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 295, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 295, + 270 + ], + "type": "text", + "content": "[38] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. In NeurIPS, 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 272, + 295, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 295, + 304 + ], + "type": "text", + "content": "[39] Reese Pathak and Martin J Wainwright. Fedsplit: An algorithmic framework for fast federated optimization. In NeurIPS, pages 7057-7066, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 305, + 295, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 305, + 295, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 295, + 348 + ], + "type": "text", + "content": "[40] Helber Patrick, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE J-STARS, 12(7):2217-2226, 2019. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 350, + 295, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 295, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 295, + 381 + ], + "type": "text", + "content": "[41] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, et al. Moment matching for multi-source domain adaptation. In ICCV, pages 1406-1415, 2019. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 383, + 295, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 295, + 425 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 295, + 425 + ], + "type": "text", + "content": "[42] Athanasios Psaltis, Anestis Kastellos, Charalampos Z Patrikakis, and Petros Daras. Fedlid: Self-supervised federated learning for leveraging limited image data. In ICCV, pages 1039-1048, 2023. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 426, + 295, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 295, + 469 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 295, + 469 + ], + "type": "text", + "content": "[43] Pian Qi, Diletta Chiaro, Antonella Guzzo, Michele Ianni, et al. Model aggregation techniques in federated learning: A comprehensive survey. Future Generation Computer Systems, 150:272-293, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 471, + 295, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 295, + 513 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 295, + 513 + ], + "type": "text", + "content": "[44] Liangqiong Qu, Yuyin Zhou, Paul Pu Liang, Yingda Xia, et al. Rethinking architecture design for tackling data heterogeneity in federated learning. In CVPR, pages 10061-10071, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 514, + 295, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 295, + 557 + ], + "type": "text", + "content": "[45] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 559, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 295, + 601 + ], + "type": "text", + "content": "[46] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, et al. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2, 3, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 603, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 295, + 635 + ], + "type": "text", + "content": "[47] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234-241, 2015. 2, 3, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 637, + 295, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 637, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 637, + 295, + 669 + ], + "type": "text", + "content": "[48] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhominov, et al. Mobilenetv2: Inverted residuals and linear bottlenecks. In CVPR, pages 4510-4520, 2018. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "type": "text", + "content": "[49] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, pages 25278–25294, 2022. 8" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "text", + "content": "[50] Seonguk Seo, Jinkyu Kim, Geeho Kim, and Bohyung Han. Relaxed contrastive learning for federated learning. In CVPR, pages 12279-12288, 2024. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 107, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 139 + ], + "type": "text", + "content": "[51] Fiona Victoria Stanley Jothiraj and Afra Mashhadi. Phoenix: A federated generative diffusion model. In ACM WWW, pages 1568-1577, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 140, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 172 + ], + "type": "text", + "content": "[52] Yue Tan, Guodong Long, Lu Liu, Tianyi Zhou, et al. FedProto: Federated prototype learning across heterogeneous clients. In AAAI, pages 8432-8440, 2022. 2, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 173, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 553, + 205 + ], + "type": "text", + "content": "[53] Yue Tan, Guodong Long, Jie Ma, Lu Liu, et al. Federated learning from pre-trained models: A contrastive learning approach. In NeurIPS, pages 19332-19344, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 206, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 239 + ], + "type": "text", + "content": "[54] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, et al. Emergent correspondence from image diffusion. In NeurIPS, pages 1363-1389, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 239, + 553, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 272 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 272 + ], + "type": "text", + "content": "[55] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, et al. What makes for good views for contrastive learning? In NeurIPS, pages 6827-6839, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 272, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 316 + ], + "type": "text", + "content": "[56] Philipp Tschandl, Cliff Rosendahl, and Harald Kittler. The ham10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. Nature Scientific Data, 5(1):1-9, 2018. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 316, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 348 + ], + "type": "text", + "content": "[57] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9 (11):2579-2605, 2008. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 349, + 553, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 349, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 349, + 553, + 371 + ], + "type": "text", + "content": "[58] Feng Wang and Huaping Liu. Understanding the behaviour of contrastive loss. In CVPR, pages 2495-2504, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 372, + 553, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 404 + ], + "type": "text", + "content": "[59] Hongyi Wang, Mikhail Yurochkin, Yuekai Sun, Dimitris Papailiopoulos, et al. Federated learning with matched averaging. In ICLR, 2020. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 404, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 437 + ], + "type": "text", + "content": "[60] Haozhao Wang, Haoran Xu, Yichen Li, Yuan Xu, et al. Fed-cda: Federated learning with cross-rounds divergence-aware aggregation. In ICLR, 2024. 2, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 437, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 553, + 480 + ], + "type": "text", + "content": "[61] Yingheng Wang, Yair Schiff, Aaron Gokaslan, Weishen Pan, et al. Infodiffusion: Representation learning using information maximizing diffusion models. In ICML, pages 36336-36354, 2023. 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 482, + 553, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 553, + 524 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 553, + 524 + ], + "type": "text", + "content": "[62] Yuan Wang, Huazhu Fu, Renuga Kanagavelu, Qingsong Wei, et al. An aggregation-free federated learning for tackling data heterogeneity. In CVPR, pages 26233-26242, 2024. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 526, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 553, + 558 + ], + "type": "text", + "content": "[63] Chen Wei, Karttikeya Mangalam, Po-Yao Huang, Yanghao Li, et al. Diffusion models as masked autoencoders. In ICCV, pages 16284-16294, 2023. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 559, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 592 + ], + "type": "text", + "content": "[64] Weilai Xiang, Hongyu Yang, Di Huang, and Yunhong Wang. Denoising diffusion autoencoders are unified self-supervised learners. In ICCV, pages 15802-15812, 2023. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 593, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 553, + 635 + ], + "type": "text", + "content": "[65] Yikai Yan, Chaoyue Niu, Yucheng Ding, Zhenzhe Zheng, et al. Federated optimization under intermittent client availability. INFORMS Journal on Computing, 36(1):185-202, 2024. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 636, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 669 + ], + "type": "text", + "content": "[66] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, et al. Diffusion models: A comprehensive survey of methods and applications. ACM Computing Surveys, 56(4):1-39, 2023. 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[67] Mingzhao Yang, Shangchao Su, Bin Li, and Xiangyang Xue. Exploring one-shot semi-supervised federated learning with pre-trained diffusion models. In AAAI, pages 16325-16333, 2024. 3, 8" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3735" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 410 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 95 + ], + "type": "text", + "content": "[68] Xingyi Yang and Xinchao Wang. Diffusion model as representation learner. In ICCV, pages 18938-18949, 2023. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 294, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 294, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 294, + 129 + ], + "type": "text", + "content": "[69] Rui Ye, Mingkai Xu, Jianyu Wang, Chenxin Xu, et al. Feddisco: Federated learning with discrepancy-aware collaboration. In ICML, pages 39879-39902, 2023. 2, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 130, + 294, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 294, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 294, + 163 + ], + "type": "text", + "content": "[70] Dai Yutong, Zeyuan Chen, Junnan Li, Shelby Heinecke, et al. Tackling data heterogeneity in federated learning with class prototypes. In AAAI, pages 7314-7322, 2023. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 164, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 294, + 196 + ], + "type": "text", + "content": "[71] Daokun Zhang, Jie Yin, Xingquan Zhu, and Chengqi Zhang. Network representation learning: A survey. IEEE Transactions on Big Data, 6:3-28, 2018. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 198, + 294, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 241 + ], + "type": "text", + "content": "[72] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, et al. A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence. In NeurIPS, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 243, + 294, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 294, + 275 + ], + "type": "text", + "content": "[73] Jianqing Zhang, Yang Hua, Hao Wang, Tao Song, et al. Fedala: Adaptive local aggregation for personalized federated learning. In AAAI, pages 11237-11244, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 277, + 294, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 277, + 294, + 309 + ], + "spans": [ + { + "bbox": [ + 56, + 277, + 294, + 309 + ], + "type": "text", + "content": "[74] Zhuang Zhao, Feng Yang, and Guirong Liang. Federated learning based on diffusion model to cope with non-iid data. In PRCV, pages 220-231, 2023. 3, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 311, + 294, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 311, + 294, + 343 + ], + "spans": [ + { + "bbox": [ + 56, + 311, + 294, + 343 + ], + "type": "text", + "content": "[75] Hangyu Zhu, Jinjin Xu, Shiqing Liu, and Yaochu Jin. Federated learning on non-iid data: A survey. Elsevier Neurocomputing, 465:371-390, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 345, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 345, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 345, + 294, + 376 + ], + "type": "text", + "content": "[76] Fuzhen Zhuang, Xiaohu Cheng, Ping Luo, Sinno Jialin Pan, et al. Supervised representation learning: Transfer learning with deep autoencoders. In IJCAI, 2015. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 378, + 294, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 378, + 294, + 410 + ], + "spans": [ + { + "bbox": [ + 56, + 378, + 294, + 410 + ], + "type": "text", + "content": "[77] Weiming Zhuang, Yonggang Wen, and Shuai Zhang. Divergence-aware federated self-supervised learning. In ICLR, 2022. 7" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "3736" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/FedMVP_ Federated Multimodal Visual Prompt Tuning for Vision-Language Models/9a8979ca-f59a-40f4-a042-065bd341c994_content_list.json b/2025/FedMVP_ Federated Multimodal Visual Prompt Tuning for Vision-Language Models/9a8979ca-f59a-40f4-a042-065bd341c994_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..aa7adf0d91df23ef8a6adc7ec2c3a78b5aeea797 --- /dev/null +++ b/2025/FedMVP_ Federated Multimodal Visual Prompt Tuning for Vision-Language Models/9a8979ca-f59a-40f4-a042-065bd341c994_content_list.json @@ -0,0 +1,1650 @@ +[ + { + "type": "text", + "text": "FedMVP: Federated Multimodal Visual Prompt Tuning for Vision-Language Models", + "text_level": 1, + "bbox": [ + 112, + 128, + 883, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mainak Singha1 Subhankar Roy2 Sarthak Mehrotra3 Ankit Jha4 Moloud Abdar5 Biplab Banerjee3 Elisa Ricci1,6", + "bbox": [ + 181, + 202, + 815, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 University of Trento, Italy", + "bbox": [ + 140, + 241, + 308, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 LNMIIT Jaipur, India", + "bbox": [ + 156, + 258, + 297, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ University of Bergamo, Italy", + "bbox": [ + 348, + 241, + 532, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "5 The University of Queensland, Australia", + "bbox": [ + 341, + 258, + 593, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Indian Institute of Technology Bombay, India", + "bbox": [ + 575, + 241, + 856, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fondazione Bruno Kessler, Italy", + "bbox": [ + 637, + 258, + 841, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 308, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In federated learning, textual prompt tuning adapts Vision-Language Models (e.g., CLIP) by tuning lightweight input tokens (or prompts) on local client data, while keeping network weights frozen. After training, only the prompts are shared by the clients with the central server for aggregation. However, textual prompt tuning suffers from overfitting to known concepts, limiting its generalizability to unseen concepts. To address this limitation, we propose Multimodal Visual Prompt Tuning (FedMVP) that conditions the prompts on multimodal contextual information - derived from the input image and textual attribute features of a class. At the core of FedMVP is a PromptFormer module that synergistically aligns textual and visual features through a cross-attention mechanism. The dynamically generated multimodal visual prompts are then input to the frozen vision encoder of CLIP, and trained with a combination of CLIP similarity loss and a consistency loss. Extensive evaluation on 20 datasets, spanning three generalization settings, demonstrates that FedMVP not only preserves performance on in-distribution classes and domains, but also displays higher generalizability to unseen classes and domains, surpassing state-of-the-art methods by a notable margin of $+1.57\\% - 2.26\\%$ . Code is available at https://github.com/mainaksingha01/FedMVP.", + "bbox": [ + 89, + 343, + 483, + 705 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 722, + 220, + 738 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A successful learning paradigm involves training neural networks on large and centralized datasets, enabling models to learn robust and generalizable representations [22, 34]. However, practical constraints often prevent centralized training (e.g., privacy regulations) creating a critical need for decentralized training frameworks that achieve high performance without central data access. As a solution, federated learning (FL) [16] enables multiple clients to collaboratively train a global model without requiring direct data sharing. In FL, clients train models locally and trans", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cad9b44ec8e2dd22c7f3c0f33c25ad62cf44a95a38038d12e2976470c5e4af54.jpg", + "image_caption": [ + "Figure 1. Comparison of federated prompt tuning methods: (a) a textual prompt tuning method (e.g., FedTPG [21]) encodes unimodal contextual information, i.e., class names, in the learnable prompts, (b) our proposed method FedMVP injects multimodal contextual information – image and attributes – into the prompts." + ], + "image_footnote": [], + "bbox": [ + 522, + 310, + 893, + 481 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mit only model updates (e.g., gradients or parameters) to a central server, which aggregates these updates to improve the global model. The research in FL is primarily aimed at: reducing communication rounds between clients and server [31, 32], faster and lightweight training algorithms on clients [19, 23] and improving generalization while exhibiting high data heterogeneity across clients [25, 33].", + "bbox": [ + 511, + 566, + 906, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pretrained Vision-Language Models (VLMs) (e.g., CLIP [22]), have emerged as promising candidates for FL due to their strong generalization capabilities. However, the high communication overhead and computational demands of parameter-heavy VLMs pose challenges for FL, especially in bandwidth- or compute-constrained environments (e.g., robots). To make CLIP viable for the FL setting, prompt tuning (e.g., CoOp [36], VPT [11]) has been adopted as a preferred technique that appends lightweight learnable tokens (or prompts) to the input tokens of the CLIP image and/or text encoder. Each client trains the prompts on the local data, while keeping the CLIP backbone frozen, and then dispatches these prompts to the server, where they are updated through aggregation and sent back. This speeds up client training as each client needs access", + "bbox": [ + 511, + 674, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17869", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to only a few samples per class (e.g., 16/class [36]), and reduces communication overhead as only the lightweight prompts (e.g., $0.37\\%$ of network parameters) need to be communicated, instead of all the parameters.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although prompt tuning has proven to be effective in the FL setting, it suffers from reduced generalization to unseen classes and unseen domains - a typical heterogeneous setup native to non independent and identically distributed (non-i.i.d.) FL [21] - due to severe overfitting on the local training data [30]. It is attributed to the fact that prompts learn a static context from a set of seen classes, which being fixed once learned, fail to capture the generalizable elements spread across clients. To circumvent the issue of reduced generalization, recent CLIP-based FL methods condition the prompts either on the class information, as in FedTPG [21] (see Fig. 1a), or on visual information, as in FedCoCoOp [35]. This auxiliary conditioning on the input, or contextual information, makes the prompts dynamic and thus offers a better generalization to unseen distributions.", + "bbox": [ + 89, + 154, + 482, + 380 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we argue that the FL setup being highly heterogeneous, where clients have data from disjoint classes and domains [21], it is not sufficient to solely exploit either of the two modalities for contextual information. Rather, FL demands more comprehensive contextual information that encompasses all the available modalities to condition input prompts. To this end, we propose Federated Multimodal Visual Prompt Tuning (FedMVP) that conditions the prompts on multimodal contextual information, which comprise: (i) visual features of the input image itself, and (ii) textual features of attributes of a class (e.g., \"two legs\", \"beak\", etc. for the class \"hen\"). The key intuition is that (ii) promotes learning transferable prompts for unseen classes that share attributes similar to seen classes (e.g., unseen class \"seagull\" will share some attributes with the seen class \"hen\"), and (i) promotes transferability for unseen classes and domains where attributes cannot be described via text (e.g., texture or abstract concepts). In a nutshell, the dual conditioning offers richer contextual information to the prompts, which helps in generalizing to classes or domains that share similar properties.", + "bbox": [ + 89, + 383, + 482, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In detail, the proposed FedMVP framework contains a PromptFormer network (see Fig. 1b) that synergistically merges the information from visual and textual modalities with cross-attention mechanism [26] to generate the prompts. The resulting multimodal prompts are then injected at the visual space of the frozen vision encoder of CLIP, unlike the existing methods [7, 21, 30] that inject prompts through text encoder. The key differences between our FedMVP and existing CLIP-based FL methods have been summarized in Tab. 1. The PromptFormer is trained in an end-to-end manner with the CLIP loss and a consistency loss. Once trained, each participating client shares the lightweight PromptFormer network parameters with the", + "bbox": [ + 89, + 704, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Summary of federated prompt-tuning methods. $(X)$ indicates the learnable prompts are initialized randomly, $(\\checkmark)$ signifies learnable prompts are conditioned on contextual information (e.g., textual or visual inputs). Prompting column denotes the CLIP encoder wherein learnable prompts are given as input (e.g., \"Textual\" means the text encoder of CLIP takes prompts as input).", + "bbox": [ + 511, + 89, + 903, + 172 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/2a3e8f268ca86406aff990874457d254657e8df5752ccd99c0698e12f41e2df9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodContextual informationPrompting
TextualVisual
FedKgCoOp [30]XXTextual
FedVPT [11]XXVisual
FedTPG [21]✓XTextual
FedCoCoOp [35]X✓Textual
FedMaPLe [13]XXMultimodal
FedMVP (Ours)✓✓Visual
", + "bbox": [ + 544, + 183, + 870, + 284 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "server for aggregation through FedAvg [16], and thus enabling knowledge sharing among clients. During inference, the cross-attention learned by the PromptFormer enables the model to dynamically adjust the prompts, enhancing generalizability to unseen classes and domains. We evaluated FedMVP on three generalization settings encompassing 20 datasets that measure generalization to unseen classes, domains, and combinations of them. Experiments demonstrate that FedMVP consistently outperforms the state-of-the-art CLIP-based FL methods by $1.57\\% - 2.26\\%$ .", + "bbox": [ + 511, + 301, + 903, + 452 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our main contributions are: (i) We highlight the importance of leveraging multimodal contextual information in prompt tuning based FL setup in order to alleviate the issue of reduced generalizability to unseen data. (ii) We propose FedMVP that synergistically combines the visual features of the image and textual features of the attributes of classes to generate multimodal prompts. This provides richer contextual information to the model and higher transferability to unseen classes and domains.", + "bbox": [ + 511, + 454, + 906, + 589 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 603, + 653, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prompt tuning of VLMs. VLMs, such as CLIP [22] and ALIGN [10], leverage large image-text datasets to enable multimodal understanding, performing well in zero-shot classification tasks. However, fully fine-tuning VLMs for downstream tasks without sacrificing their ability to generalize to unseen concepts remains a challenge [27].", + "bbox": [ + 511, + 628, + 903, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prompt tuning approaches [11, 13, 36] have been proposed for parameter efficient and effective adaptation of CLIP by introducing a set of continuous learnable tokens (or prompts) onto the frozen CLIP backbone. These approaches can be broadly categorized into three categories depending on the manner in which CLIP is prompted: (i) textual prompting that inject prompts in the input space of the text encoder [30, 36], (ii) visual prompting that injects prompts in the input space of the vision encoder [11], and (iii) multimodal prompting that injects prompts in both the text and vision encoder [13, 28].", + "bbox": [ + 511, + 719, + 905, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In particular, MaPLe [13], which injects prompts to both", + "bbox": [ + 532, + 885, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17870", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the text and vision encoders, is not suitable for memory and compute efficient FL since it requires parts of both the encoders to be unlocked during tuning, risking overfitting. Moreover, it does not allow for image-conditioned dynamic prompting, which has shown to be effective for generalization [35]. In contrast, our proposed FedMVP enables multimodal prompting by involving only the vision encoder and supports instance-specific prompting.", + "bbox": [ + 89, + 90, + 480, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Federated prompt tuning of VLMs. Prompt tuning has emerged as an effective technique for FL of VLMs since it allows clients to collaboratively train on a heterogeneously distributed dataset and adapt faster with fewer samples. The fact that the prompts are lightweight lends well to the FL setup as it reduces communication overhead. Due to the benefits offered by prompt tuning, several works [2, 7, 21, 24] have specifically adapted prompt tuning approaches to the FL setup. For example, PromptFL [7] extended CoOp to FL via sharing prompts with the server, enhancing the generalization performance of client-server interactions in vision tasks. FedCLIP [15] learns a lightweight adapter in each client, similar to the CLIP-Adapter [5], and shares parameters with the server. FedOTP [14] introduced global and local prompts, optimizing shared parameters through optimal transport, while DiPrompt [2] studied multi-domain learning by allowing multiple domains per client.", + "bbox": [ + 89, + 214, + 482, + 472 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike existing methods, a couple of recent works [21, 28] argue that prompt tuning is prone to overfitting to the seen classes due to learning static context. The closest work to ours is FedTPG [21], where a prompt generation network is learned across multiple clients. Additionally, this network leverages textual inputs, enhancing its ability to generalize to unseen classes. Our work significantly differs from previous methods in two key ways: (i) we exploit both attribute information derived from an LLM and image features to condition the prompts, and (ii) we leverage multimodal prompts from the proposed PromptFormer for visual prompt tuning, instead of textual prompt tuning.", + "bbox": [ + 89, + 474, + 482, + 656 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Classification via textual descriptions. Zero-shot accuracy of CLIP on downstream tasks is sensitive to the quality of text prompts [22]. In detail, some works rely on simple hand-crafted prompts (e.g., \"a photo of a [CLS]\") [22] or others augment the hand-crafted prompts with richer attributes obtained from LLMs [17, 20] (e.g., a \"hen\" has two legs, white feathers, etc.). There also exists a family of methods, such as LaCLIP [4], LaBo [29] and VFC [18], that have proposed to further refine CLIP weights with enriched captions from LLMs for improved performance. Recognizing the importance of class attributes for improved generalization to unseen concepts and domains – an important desideratum in the FL set up – we, for the first time in FL, introduce LLM generated class attributes for enriching the image-conditioned multimodal prompts through the PromptFormer network.", + "bbox": [ + 89, + 659, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 513, + 89, + 612, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We introduce FL setup in Sec. 3.1, preliminaries on prompt tuning in Sec. 3.2 and our proposed solution in Sec. 3.3.", + "bbox": [ + 511, + 114, + 906, + 148 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem formulation", + "text_level": 1, + "bbox": [ + 513, + 157, + 712, + 172 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We consider a FL setup in which multiple remote clients train models on local data, and a central server aggregates models from the clients. Formally, let each remote client $i$ has access to a dataset $\\mathcal{D}_i = \\{\\mathbf{x}_{i,j},y_{i,j}\\}_{j = 1}^{n_i}$ , where $\\mathbf{x}_{i,j}$ is the $j^{\\mathrm{th}}$ image from the $i^{\\mathrm{th}}$ client, $y_{i,j}$ is the corresponding class label and $n_i$ is the number of samples in client $i$ . Let the set of text class names in client $i$ be denoted as $\\mathcal{C}_i = \\{c_{i,k}\\}_{k = 1}^{K_i}$ , where $K_{i}$ is the total number of classes. As in FedTPG [21], we follow a non-IID FL setup, where the samples in each client are drawn from a disjoint set of classes, i.e., $\\mathcal{C}_i\\cap \\mathcal{C}_l = \\emptyset$ . As an example, the training data on each mobile device usually depend on the user, which may not be representative of the general population [16].", + "bbox": [ + 511, + 181, + 903, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In particular, we investigate two kinds of non-IID FL setups, where, in addition to the in-distribution performance, we also evaluate the generalization ability of the aggregated server model on (i) unseen classes, and (ii) unseen domains or datasets. The need for generalization to unseen classes and domains makes the non-IID FL setup more challenging than the traditional FL setting [12].", + "bbox": [ + 511, + 381, + 905, + 486 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Preliminaries on Prompt Tuning", + "text_level": 1, + "bbox": [ + 513, + 500, + 799, + 515 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Textual prompt tuning (TPT), i.e., CoOp [36], has emerged as an efficient technique in CLIP-based FL that keeps the CLIP backbone frozen and locally learns lightweight prompt vectors on each client. The locally learned prompts are then communicated to the server, as in PromptFL [7]. Specifically, the hand-crafted text prompts (e.g., \"a photo of a [CLS]\") input to the text encoder $\\mathcal{E}_t$ is augmented with the learnable soft prompts $v_{1}, v_{2}, \\ldots, v_{m}$ . However, as shown in [35], TPT can often lead to overfitting and reduced generalization due to the static context learned on the seen classes. This can be exacerbated in the non-IID FL setup due to the high data heterogeneity among the clients [21].", + "bbox": [ + 511, + 522, + 905, + 703 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Alternatively, visual prompt tuning (VPT) [11] adapts pre-trained vision encoder $\\mathcal{E}_v$ by injecting small number of visual learnable parameters into ViT's input space. More formally, let an input image $\\mathbf{x}$ be divided in $b$ patches, yielding a collection of patch embeddings (augmented with positional encodings) at the input of the vision encoder as $\\mathbf{E} \\in \\mathbb{R}^{b \\times d_v}$ , where $d_v$ is the embedding dimension. A [CLS] token $\\mathbf{z}$ is concatenated with $\\mathbf{E}$ , encapsulating the initial input representation for the ViT layers of $\\mathcal{E}_v$ . VPT further augments the input by concatenating a set of randomly initialized learnable visual prompts $\\mathbf{P} = \\{p_1, p_2, \\ldots, p_m\\}$ , where $m$ is the prompt length. The first layer input of $\\mathcal{E}_v$ is then redefined as $\\mathbf{I} = [\\mathbf{z}; \\mathbf{E}; \\mathbf{P}] \\in \\mathbb{R}^{(1 + b + m) \\times d_v}$ , where", + "bbox": [ + 511, + 704, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17871", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8d12d753ddcaf747b6524df2d4c26ff3aae9fb9a19b8357fefababb4f875ae81.jpg", + "image_caption": [ + "Figure 2. Overview of Federated Multimodal Visual Prompt Tuning (FedMVP). Each client $i$ trains the parameters of the PromptFormer network $f_{\\Theta}$ and multi-head self-attention on its local data. At each round, the server aggregates the trainable parameters from the clients and sends the updated parameters back to the clients. The PromptFormer uses cross-attention mechanism to fuse the visual and textual attribute feature to generate multimodal visual prompts $\\mathbf{P}$ , which are then used for visual prompt tuning." + ], + "image_footnote": [], + "bbox": [ + 98, + 89, + 331, + 325 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7c528344f045fb89837191fb059d5e5051c445481371681448063518fdd92c12.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 90, + 602, + 324 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2ece183bd23848b010a1a9ce2ceff94d4ceb58813e90864c10c8b8702d125c59.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 90, + 843, + 325 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "“;” means concatenation. Although VPT enables more nuanced understanding of the visual space, it suffers from the same issues as TPT, i.e., limited generalization (see Sec. 4).", + "bbox": [ + 88, + 393, + 480, + 440 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Federated Multimodal Visual Prompt Tuning", + "text_level": 1, + "bbox": [ + 89, + 450, + 475, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our proposed method FedMVP deviates from the TPT and VPT paradigms and incorporates multimodal contextual information from both the visual and textual modalities to improve generalization in the FL setup. As shown in Fig. 2, at the core of our novel architecture is a prompt generator, called PromptFormer, which conditions the visual prompts $\\mathbf{P}$ on the task-related information: input image and text attributes corresponding to the class names, rather than randomly initializing them. This makes $\\mathbf{P}$ context-aware, generalizing better to unseen classes and domains. Once trained on each client, only the parameters of the lightweight PromptFormer network are communicated to the server, thus, keeping the communication overhead low. Next, we describe PromptFormer (Sec. 3.3.1), training objectives (Sec. 3.3.2) and how we conduct client-side training and server-side aggregation (Sec. 3.3.3).", + "bbox": [ + 88, + 472, + 482, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1. Visual Prompt Generation with PromptFormer", + "text_level": 1, + "bbox": [ + 89, + 723, + 465, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We introduce the PromptFormer network $f_{\\Theta}$ , parameterized by $\\Theta$ , designed to generate contextually rich visual tokens (or prompts) for each client $i$ by leveraging both attribute features and visual patch embeddings (see Fig. 2). Concretely, we use the CLIP text encoder $\\mathcal{E}_t$ to extract embeddings $\\mathbf{A}_i = \\{\\mathcal{E}_t(\\mathrm{LLM}(c_k))\\}_{k=1}^K$ of the text attributes $\\mathcal{A}_i$ corresponding to the text class names $\\mathcal{C}_i$ , generated using a large lanuage model (LLM) [1] (see Sec. A.1 of Supp. Mat.", + "bbox": [ + 89, + 742, + 482, + 863 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for Attribute Generation process). Going forward, we drop the client index $i$ for notational clarity. The PromptFormer takes as input $b$ visual patch embeddings $\\mathbf{E}$ , corresponding to an image $\\pmb{x}$ , and the text embeddings $\\mathbf{A}$ to generate a set of enriched $m$ multimodal visual prompts $\\mathbf{P} \\in \\mathbb{R}^{m \\times d_v}$ :", + "bbox": [ + 511, + 393, + 905, + 469 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} = \\left\\{p _ {j} \\right\\} _ {j = 1} ^ {m} = f _ {\\Theta} (\\mathbf {A}, \\mathbf {E}). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 478, + 903, + 496 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Consequently, the input to the first layer of the visual encoder $\\mathcal{E}_v$ is redefined as $[\\mathbf{z};\\mathbf{E};\\mathbf{P}]\\in \\mathbb{R}^{(1 + b + m)\\times d_v}$ . Intuitively, multimodal visual prompts carry richer contextual information, and helps the network in generalizing to unseen classes and domains that share commonalities.", + "bbox": [ + 511, + 507, + 905, + 582 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Architecture of PromptFormer. The PromptFormer architecture comprises: lightweight multi-head cross-attention (MHCA) modules, feed-forward network (FFN) and a linear projection layer; transforming multimodal features into comprehensive visual prompts. In detail, the visual patch embedding $\\mathbf{E}$ interacts with the linearly projected attribute features $\\mathbf{A}' = T_{\\mathrm{proj}}(\\mathbf{A})$ to construct multimodal visual prompts as:", + "bbox": [ + 511, + 595, + 905, + 715 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} \\left(\\mathbf {A} ^ {\\prime}, \\mathbf {E}\\right) = \\operatorname {F F N} \\left(\\text {C r o s s A t t e n t i o n} \\left(\\mathbf {Q} _ {\\mathbf {E}}, \\mathbf {K} _ {\\mathbf {A} ^ {\\prime}}, \\mathbf {V} _ {\\mathbf {A} ^ {\\prime}}\\right)\\right);\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 727, + 883, + 744 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\text {w i t h} \\mathbf {Q} _ {\\mathbf {E}} = \\mathbf {E} W _ {\\mathbf {Q}}, \\mathbf {K} _ {\\mathbf {A} ^ {\\prime}} = \\mathbf {A} ^ {\\prime} W _ {\\mathbf {K}}, \\mathbf {V} _ {\\mathbf {A} ^ {\\prime}} = \\mathbf {A} ^ {\\prime} W _ {\\mathbf {V}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 748, + 903, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where, $\\mathbf{E}$ are transformed into query vectors $\\mathbf{Q}_{\\mathbf{E}}$ using query matrices $W_{\\mathbf{Q}}$ ; and the projected attribute features $\\mathbf{A}'$ are transformed into key and value vectors $\\mathbf{K}_{\\mathbf{A}'}$ and $\\mathbf{V}_{\\mathbf{A}'}$ , using key and value matrices $W_{\\mathbf{K}}$ and $W_{\\mathbf{V}}$ , respectively.", + "bbox": [ + 511, + 779, + 903, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Intuitively, through the cross-attention mechanism, visual features learn to attend to the relevant parts of the corresponding attribute features. For example, the patches depicting the legs of the class \"dog\" will learn to attend to", + "bbox": [ + 511, + 840, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1For example, for the class name \"giraffe\" the LLM outputs \"Exceptionally long neck, unique coat pattern with irregular brown patches, ...\"", + "bbox": [ + 89, + 875, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17872", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the attributes four legged. When an unseen class (e.g., \"giraffe\") comprises an animal with four legs, the final prompts $\\mathbf{P}$ will contain this relevant information derived from the attributes four legged. Additional details of PromptFormer architecture are reported in Sec. A.3 of the Supp. Mat.", + "bbox": [ + 89, + 90, + 483, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.2. Training of FedMVP", + "text_level": 1, + "bbox": [ + 89, + 174, + 282, + 188 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After the visual prompts are obtained with PromptFormer as per Eq. 2, we concatenate the [CLS] token, the patch embeddings, and the visual prompts to get $\\mathbf{I} = [\\mathbf{z};\\mathbf{E};\\mathbf{P}]$ following Sec. 3.2, and feed it to the visual encoder $\\mathcal{E}_v$ .", + "bbox": [ + 89, + 193, + 482, + 253 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The parameters of PromptFormer $\\rho = \\{\\Theta\\}$ are the only trainable parameters of FedMVP, and the rest of the weights are kept frozen. To train the model, we employ the CLIP cross-entropy loss [22], formulated as:", + "bbox": [ + 89, + 253, + 483, + 315 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c e}} = - \\underset {(\\mathbf {x}, y) \\sim \\mathcal {D}} {\\mathbb {E}} y \\log p (y | \\mathbf {I}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 321, + 482, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $y$ denotes the ground truth label, and the prediction probability of CLIP between the visual features $\\mathbf{v} = \\mathcal{E}_v(\\mathbf{I})$ and text features $\\mathbf{t}_k = \\mathcal{E}_t([A\\text{ photo of [CLASS]};\\mathrm{LLM}(c_k)])$ for class $k$ is computed as:", + "bbox": [ + 89, + 357, + 483, + 417 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np (y = k | \\mathbf {I}) = \\frac {\\exp \\left(\\cos \\left(\\mathbf {v} , \\mathbf {t} _ {k}\\right)\\right) / \\tau)}{\\sum_ {k ^ {\\prime} = 1} ^ {K} \\exp \\left(\\cos \\left(\\mathbf {v} , \\mathbf {t} _ {k ^ {\\prime}}\\right)\\right) / \\tau}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 419, + 482, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\cos (\\cdot ,\\cdot)$ as a similarity function, $\\tau$ as the temperature parameter, and $[\\cdot ]$ denotes concatenation. Details on how $\\mathbf{t}_k$ is computed are provided in Sec. A.1 of the Supp. Mat.", + "bbox": [ + 89, + 465, + 483, + 511 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition to the $\\mathcal{L}_{\\mathrm{ce}}$ , we employ a regularization constraint to ensure consistency between two augmented views of an image. We define the consistency loss $\\mathcal{L}_{\\mathrm{con}}$ as:", + "bbox": [ + 89, + 512, + 483, + 556 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c o n}} = 1 - \\cos \\left(\\mathcal {E} _ {v} (\\mathbf {I}), \\mathcal {E} _ {v} \\left(\\mathbf {x} ^ {\\prime}\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 569, + 482, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{x}'$ is the augmented version of $\\mathbf{x}$ . Finally, the total loss function is given as:", + "bbox": [ + 89, + 597, + 482, + 628 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\mathrm {c e}} + \\alpha \\cdot \\mathcal {L} _ {\\mathrm {c o n}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 632, + 482, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ is a balancing coefficient for the regularization term, ensuring both alignment and consistency.", + "bbox": [ + 89, + 662, + 482, + 693 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.3. Client-side training and server-side aggregation", + "text_level": 1, + "bbox": [ + 89, + 700, + 467, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our approach, multiple remote clients participate in distinct image classification tasks, collaboratively training the PromptFormer network, parameterized by $\\rho = \\{\\Theta\\}$ . In each communication round, the client transmits $\\rho$ to the server, which is aggregated at the server. Through training on a diverse set of classification tasks distributed across clients, and aggregating the parameters at the server, the resulting model achieves generalization capability to classes and domains unseen during training. Pseudo-code of FedMVP can be found in Sec. A.2 of the Supp. Mat.", + "bbox": [ + 89, + 719, + 482, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Initially, the server randomly initializes the parameters $\\rho^0$ and in each communication round only a random subset", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of remote clients $S^r$ will be chosen by the server for sharing the updated parameters $\\rho$ . Each communication round $r$ involves the following steps:", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(i) Model initialization: Each client $i$ in $S_{r}$ receives the latest model parameters $\\rho^r$ to configure its local model.", + "(ii) Local optimization: Following Eq. (6), client $i$ optimizes its local parameters $\\rho^r$ using an SGD optimizer over $J$ iterations with learning rate $\\eta^r$ . Updated parameters:" + ], + "bbox": [ + 511, + 136, + 903, + 212 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\rho_ {i} ^ {r + 1} = \\operatorname {S G D} _ {J} \\left(\\eta^ {r}, \\rho^ {r}, \\mathcal {C} _ {i}, \\mathcal {D} _ {i}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 219, + 903, + 238 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(iii) Parameter aggregation: After local training, each client $i$ in $S_{r}$ transmits its updated parameters $\\rho_{i}^{r + 1}$ back to the server, where they are aggregated [16] as:", + "bbox": [ + 511, + 250, + 903, + 294 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\rho^ {r + 1} = \\frac {1}{\\left| S _ {r} \\right|} \\sum_ {i \\in S _ {r}} \\rho_ {i} ^ {r + 1}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 303, + 903, + 339 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The above steps are repeated for $R$ communication rounds, after which the server converges on the final model parameters $\\rho^R$ . This iterative process facilitates collaborative and distributed optimization of $\\rho$ across client data.", + "bbox": [ + 511, + 349, + 903, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Lightweight fine-tuning in clients. As the FL setup exhibits high data heterogeneity, different clients possess varying numbers of samples. Thus, treating every client as equal and repeatedly fine-tuning the parameters $\\rho$ can lead to severe overfitting in some clients, especially those with very limited training data. To prevent this, we propose to inject low-rank decomposition matrices, or LoRA [9], to the PromptFormer module, and fine-tune these LoRA matrices if training loss $\\mathcal{L}_{\\mathrm{total}}$ in a client starts below a certain loss value $\\sigma$ , which we set to 0.5. In particular, only the LoRA weights corresponding to the query, key and value matrices of the attention layers are trained, while freezing the parameters $\\rho$ . In that case, the client shares only the LoRA matrices with the server. This not only helps in boosting generalization, but also brings down the communication overhead of clients by around $\\times 267$ .", + "bbox": [ + 511, + 409, + 906, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 664, + 643, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our proposed FedMVP in three distinct settings that are aimed to gauge the (i) generalization to unseen categories (or Base-to-New Generalization) in Sec. 4.1, (ii) generalization to unseen domains (or Domain Generalization) in Sec. 4.2, and (iii) generalization to unseen datasets (or Cross-Dataset Generalization) in Sec. 4.3. Finally, we report the results of our ablation study in Sec. 4.4 that are aimed to assess impact of the design choices made in FedMVP. Due to the lack of space, we report additional experimental results in Sec. B of the Supp. Mat.", + "bbox": [ + 511, + 688, + 903, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Benchmarks and datasets. Following FedTPG [21], we have reported results on the Base-to-New Generalization setting that encompasses nine medium-sized image recognition datasets. For the Domain Generalization setting we", + "bbox": [ + 511, + 839, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17873", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "have used the ImageNet [3] and DomainBed [6] benchmarks. In addition, we have also included EuroSAT [8] for the Cross-Dataset Generalization setting, alongside all nine datasets of the Base-to-New Generalization setting. In total, we have used 20 datasets for a comprehensive experimental evaluation. Detailed dataset descriptions and statistics can be found in Sec. B.1 of the Supp. Mat.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We have compared our method with the state-of-the-art prompt tuning approaches that use the CLIP backbone [22]: KgCoOp [30], CoCoOp [35], VPT [11], and MaPLe [13], which were originally designed for offline and non-FL setup, but adapted to the federated setting, as in FedTPG [21]. Furthermore, we have compared with CLIP-based methods specifically designed to work in the FL setting: FedTPG [21], FedOTP [14], PromptFL [7], and FedCLIP [15]. We also included zero-shot CLIP (ZS-CLIP).", + "bbox": [ + 89, + 198, + 480, + 333 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. We have employed the frozen ViT-B/16 CLIP backbone as the vision encoder, and GPT-4o [1] as the LLM in our experiments. Each MHCA module of $f_{\\Theta}$ , consists of a 4-head cross-attention layers, complemented by layer normalization and each FFN comprising of a two-layer bottleneck structure (Linear-GeLU-Linear). The query prompt $\\mathbf{Q}$ has a length $m = 4$ and matches the patch embedding dimension $d_v = 768$ . The projection layer $T_{\\mathrm{proj}}$ maps the textual dimension of 512 to the patch embedding dimension via a linear transformation. In Eq. (6), we keep $\\alpha = 10$ . We have set the batch size to 128, number of shots per class to 8, and the learning rate equal to 0.003 with a decay rate of 1e-5 and momentum of 0.9.", + "bbox": [ + 89, + 335, + 482, + 531 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Base-to-New Generalization", + "text_level": 1, + "bbox": [ + 89, + 544, + 341, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experimental setup. To recap, the base-to-new generalization setting encompasses nine different datasets. The setting is characterized by splitting the classes from each dataset into two groups – base and new. The samples of the base classes are used for training, and the samples of the new classes are held out to test the generalization performance [21]. Furthermore, the samples of bases classes from all nine datasets are distributed to the clients, such that each client does not see samples from more than 20 disjoint classes (referred to as local classes). We report the model's classification accuracy on: (i) the client's local classes, (ii) all the base classes (spread across clients), and (iii) the held-out new classes that were never seen by any client. Note that (ii) and (iii) is computed using the aggregated model on the server, and (i) is computed using the model on the client (which is the same server model after convergence).", + "bbox": [ + 89, + 566, + 482, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Main results. In Table 2 we report the performance in the base-to-new generalization setting, summarized across nine datasets. The results show that the textual and multimodal prompting methods, such as PromptFL, FedKgCoOp, and FedMaPLe, excel on the seen local and base class accuracy, but have at-par or worse generalization on the new", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/737a9e648ae7bf683d085c56308d6820e164c500361b0ec03f28a72bafa2e4ee.jpg", + "table_caption": [ + "Table 2. Base-to-New Generalization setting. Accuracies $(\\%)$ on client's local (seen) classes, base (seen) classes and new (unseen) classes. Harmonic mean (HM) is computed with Base and New class accuracies. Best number in each column is in bold and second-best underlined. Improvement row reports the absolute difference between the best and second-best numbers." + ], + "table_footnote": [], + "table_body": "
MethodsLocalBaseNewHM
ZS-CLIP [22]76.7270.5175.7874.24
FedOTP [14]74.8265.2257.0464.89
FedCoCoOp [35]81.4673.7666.0073.20
FedVPT [11]76.2970.4374.8973.79
FedCLIP [15]76.8771.0475.0674.24
FedMaPLe [13]81.6374.4470.6275.29
FedKgCoOp [30]78.3872.1875.8775.39
PromptFL [7]81.7574.4771.7075.74
FedTPG [21]80.7573.6876.0276.70
FedMVP (Ours)81.8975.3777.8278.27
Improvement+0.14+0.90+1.79+1.57
", + "bbox": [ + 578, + 176, + 834, + 321 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "classes than the ZS-CLIP model. This hints at the potential overfitting caused by learning static context from the seen classes. In contrast, our proposed FedMVP that considers multimodal contextual information and dynamic conditioning demonstrates superior performance in both the seen and unseen classes, achieving the highest HM accuracy. While FedTPG also generates prompts based on task-related input, it does not outperform our FedMVP due to lack of comprehensive visual and attribute information in the prompts. In summary, FedMVP surpasses the state-of-the-art methods by a notable margin of $+1.57\\%$ in the base-to-new generalization setting. Detailed breakdown for each of the nine datasets is reported in Tab. A.4 in the Supp. Mat.", + "bbox": [ + 511, + 333, + 906, + 530 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Domain Generalization", + "text_level": 1, + "bbox": [ + 511, + 542, + 728, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experimental setup. We conduct experiments under two distinct domain generalization (DG) setups: Multi-source Single-target (MSST) and Single-source Multi-target (SSMT). In the MSST DG setting for a given benchmark, we follow a leave-one-domain-out protocol for evaluation, where we hold out one domain for measuring generalization and use the rest of the domains for training the clients. Training classes are split among clients under the constraint that each client can see images from a single domain. This setting mirrors a practical scenario where images taken with a smartphone uniquely reflect the taste of the user or socio-economic biases. In a similar manner, in the SSMT DG setting, we use one domain for training the clients in a distributed manner, while the rest of the domains are used for evaluation. More details on this experimental setup can be found in Sec. B.4 of the Supp. Mat.", + "bbox": [ + 511, + 566, + 903, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Main results. In Table 3, we report the performance of FedMVP and its competitors in the MSST and SSMT DG settings. In particular, SSMT DG is a more challenging setting than the MSST DG setting, as the clients are trained only on a single domain, and expected to generalize to multiple unseen domains. We observe that ZS-CLIP is a strong base", + "bbox": [ + 511, + 810, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17874", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/afac0969f22d01a46f161a6ba53574ae21c7b84f7298a7298bdff112b78bed58.jpg", + "table_caption": [ + "Table 3. Domain Generalization setting on the DomainBed [6] benchmark. Accuracies $(\\%)$ on the unseen target domain(s) for the Multi-source Single-target (MSST) and Single-source Multi-target (SSMT) settings. Best number in each column is in bold and second-best underlined. Improvement row reports the absolute difference between the best and second-best numbers." + ], + "table_footnote": [], + "table_body": "
MethodMulti-source Single-targetSingle-source Multi-target
PACSOfficeHomeVLCSTerra Inc.DomainNetAveragePACSOfficeHomeVLCSTerra Inc.DomainNetAverage
ZS-CLIP [22]96.1681.4983.2933.9857.1370.4196.1681.4983.2933.9857.1370.41
FedOTP [14]90.7176.4267.4113.2449.6759.4991.1774.5365.1615.1139.3757.07
FedCoCoOp [35]85.0681.4261.7323.6857.0861.7984.5650.8761.2523.0758.2855.61
FedTPG [21]90.9982.7869.7726.7956.8265.4390.7182.6067.6322.5158.3464.36
PromptFL [7]95.3781.7474.8725.0256.8766.7795.8381.7266.9724.1758.2765.39
FedKgCoOp [30]95.4281.8274.9025.0356.9266.8295.8181.7067.4924.2058.3065.50
FedMaPLe [13]94.5182.0371.7936.3058.8868.7094.2581.4872.0833.5957.6267.80
FedVPT [11]95.3681.7683.1933.6255.9869.9894.7980.9982.4432.2355.6669.22
FedCLIP [15]96.2981.7482.7036.5857.8571.0396.2981.6281.8136.4457.4270.72
FedMVP (Ours)97.2884.1585.1237.3661.1773.0297.1783.8984.6136.9260.5672.63
Improvement+0.99+1.37+1.83+0.78+2.29+1.99+0.88+1.29+1.32+0.48+2.22+1.91
", + "bbox": [ + 142, + 133, + 849, + 287 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e874b3ba358e559effbf085b9557308680bb4cb377e03fedaf8a4e6704c9a7cb.jpg", + "table_caption": [ + "Table 4. Domain Generalization setting on the ImageNet benchmark. Accuracies (\\%) on the seen source domain ImageNet (IN) and unseen target domains: ImageNet-V2 (INV2), ImageNet-Sketch (IN-S), ImageNet-A (IN-A) and ImageNet-R (IN-R). Best number in each column is in bold and second-best underlined. Improvement row reports the absolute difference between the best and second-best numbers." + ], + "table_footnote": [], + "table_body": "
MethodSourceTarget
ININV2IN-SIN-AIN-RAverage
ZS-CLIP [22]66.7560.7946.1247.7974.0057.18
FedOTP [14]51.6845.5334.7316.6463.9840.22
FedMaPLe [13]66.9660.6544.6946.2474.6256.55
FedVPT [11]66.9260.3446.4348.0374.5657.34
PromptFL [7]67.8061.5945.6148.7874.4957.62
FedCLIP [15]67.2661.3546.6648.2474.3657.65
FedKCoOp [30]67.5361.6046.6948.3774.7157.84
FedCoCoOp [35]68.5162.2946.9050.3376.4959.00
FedTPG [21]69.5162.9047.6549.9776.3559.22
FedMVP (Ours)70.8763.7250.9351.7677.2360.91
Improvement+1.36+0.82+3.28+1.43+0.74+1.69
", + "bbox": [ + 122, + 393, + 444, + 539 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "line that outperforms all textual and multimodal prompting methods, with the exception of FedCLIP and our proposed FedMVP. This suggests that suboptimal prompt tuning can lead to overfitting on the source training domains, thus hurting generalization. Our FedMVP owing to the use of contextual information learns more transferable representations that generalize better to unseen target domains, thus outperforming all competitor methods in most of the benchmarks. More detailed results in Sec. B.4 of the Supp. Mat.", + "bbox": [ + 88, + 546, + 482, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Table. 4 we report the results in the SSMT DG setting on ImageNet benchmark. We observe that FedTPG and FedCoCoOp come close to our FedMVP in terms of average performance, with FedMVP outperforming all baselines by $+1.69\\%$ . In particular, we believe FedMVP improves the performance on ImageNet-Sketch (IN-S) by a large margin of $+3.28\\%$ due to the use of attribute information, which remains unchanged between real and sketch images.", + "bbox": [ + 88, + 681, + 482, + 803 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Cross-dataset Generalization", + "text_level": 1, + "bbox": [ + 89, + 808, + 351, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Experimental setup. In the Cross-Dataset Generalization setting the classes from ImageNet are split among the clients for training, in a disjoint manner, and the resulting network is evaluated on 10 held out datasets not seen during training. In other words, it is a combination of base-to-new", + "bbox": [ + 88, + 824, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "generalization and DG settings, since it requires generalization to both unseen class and data distributions.", + "bbox": [ + 511, + 294, + 903, + 321 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Main results. In Table. 5 we report the results of the Cross-Dataset Generalization setting. This setting is considerably challenging, as the model should generalize to different datasets. We find that the results are consistent with the previous findings, where our proposed FedMVP generalizes better to unseen classes and data distributions, outperforming all the text-based and multimodal-based prompting techniques by $+2.26\\%$ on average. However, with FedMVP we notice a drop in performance with respect to the ZS-CLIP baseline for some fine-grained datasets, such as OxfordPets and StanfordCars, which is potentially due to overlapping attributes for visually similar categories.", + "bbox": [ + 511, + 325, + 906, + 507 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 517, + 666, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of proposed components in FedMVP. In Tab. 6 we ablate each component of FedMVP on the base-to-new and MSST DG settings. We observe that each component contributes positively to the final performance. First, Prompt-Former $f_{\\Theta}$ plays a major role in combining textual information with visual features, thus improving semantic alignment between the two modalities. Second, adding $\\mathcal{L}_{\\mathrm{con}}$ (as in Eq. (5)) improves the performance by a small margin. Finally, the impact of LoRA can be observed in the penultimate row, where we fine-tune all the parameters of FedMVP. It shows the dangers of overfitting in the FL setup, as some clients may have limited data.", + "bbox": [ + 511, + 535, + 905, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Participation Rate of the Clients. In Fig. 3 we vary the participation rate from $10\\%$ to $100\\%$ and show how the participation rate of the clients impacts performance on both seen and unseen classes and domains. We observe that FedMVP always exhibits a gentle, yet positive slope in both settings, indicating that continuous parameter aggregation improves generalization. In contrast, for some competitor methods, such as PromptFL, increased participation leads to saturated performance or negative slope in the plot.", + "bbox": [ + 511, + 717, + 905, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Computational analysis. In Fig. 4 we plot the accuracy vs. communication rounds for FedMVP and competitors in base-to-new and MSST settings. While FedMVP trans", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17875", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1f081649efc02334f5a1f2abb0feaa9524a68cd0b295b41a4ec940e19607bb3b.jpg", + "table_caption": [ + "Table 5. Cross-Dataset Generalization setting. Accuracies $(\\%)$ on the seen classes of the source dataset, i.e., ImageNet, and the unseen classes of the target datasets. Best number in each column is in bold and second-best underlined. Improvement row reports the absolute difference between the best and second-best numbers." + ], + "table_footnote": [], + "table_body": "
MethodSourceTarget
ImageNetCaltech101Flowers102FGVCAircraftUCF101OxfordPetsFood101DTDStanfordCarsSUN397EuroSATAverage
ZS-CLIP [22]66.7592.9071.2924.7266.7589.1586.0944.3365.2962.5947.6865.08
FedOTP [14]51.6888.6048.3610.9546.1073.4361.0141.5529.6148.3450.2649.82
FedMaPLe [13]66.9692.4968.2523.5260.3289.6783.5244.6860.1661.8545.3862.98
PromptFL [7]67.8091.8768.1321.4464.1388.7085.8542.4363.5962.7743.2663.22
FedCLIP [15]67.2693.3967.1924.0365.6488.2885.1444.4465.5063.4241.6763.87
FedKgCoOp [30]67.5393.6369.3123.0664.4688.5585.3744.7464.9963.8543.2964.13
FedVPT [11]68.5691.4269.2421.0965.7087.7985.4544.5765.0064.6847.8264.28
FedCoCoOp [35]68.5194.1166.3420.7962.7589.0485.4043.2063.9864.0255.4064.50
FedTPG [21]69.5193.7570.0423.2264.7290.6085.9146.2563.9866.7847.8665.31
FedMVP (Ours)70.8795.3772.8025.9470.5889.2787.0649.7865.8368.1950.8467.57
Improvement+1.36+1.62+1.51+1.22+3.83-1.33+0.97+3.53+0.33+1.41-4.56+2.26
", + "bbox": [ + 133, + 133, + 859, + 281 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1a6e221c5f65c28c7f4e36273ddf7540337b55e66cbd44c4c08eb08c6c473f74.jpg", + "table_caption": [ + "Table 6. Ablating FedMVP components. We report numbers for the Base-to-New and MMST DG (on DomainBed) settings." + ], + "table_footnote": [], + "table_body": "
ComponentsBase-to-NewMSST DG
ZS-CLIP74.2470.41
fΘ only75.9471.85
fΘ + Lcon76.2772.14
w/o LoRA77.4172.58
FedMVP (Full)78.2773.02
", + "bbox": [ + 173, + 316, + 400, + 401 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/46620df8a887974449ef6f18e1c867ef88ecf2192563189d5743674b1f469045.jpg", + "image_caption": [ + "(a) Base-to-New" + ], + "image_footnote": [], + "bbox": [ + 94, + 405, + 279, + 506 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/67bbe100bb32c1646d18fefb2e85d755dc3522b07aae680627c4d0d50d22db46.jpg", + "image_caption": [ + "(b) MSST DG", + "Figure 3. Sensitivity to participation rate of the clients, on (a) Base-to-New and (b) MSST DG (DomainBed) settings." + ], + "image_footnote": [], + "bbox": [ + 290, + 404, + 475, + 505 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "mits $2 \\times$ more parameters than FedTPG, it converges nearly $10 \\times$ faster, requiring fewer rounds. In particular, as training progresses, only the LoRA parameters in PromptFormer are fine-tuned in a client, thus reducing the trainable parameter count by $267 \\times$ . Overall, FedMVP has lower total computational cost with respect to other methods when normalized by the number of communication rounds.", + "bbox": [ + 89, + 574, + 483, + 679 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/58af3d4289b3246397c461095f4d89041ff90d39e97ffc30fc0f3a3f561f5a94.jpg", + "image_caption": [ + "(a) Base-to-New", + "Figure 4. Computational analysis of the methods on (a) Baseto-New and (b) MSST DG (DomainNet) setting." + ], + "image_footnote": [], + "bbox": [ + 94, + 696, + 279, + 797 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/167988ae91ba94fdfe0a58fe677573c94a829f001f23902fd6280f66e8c0a4cb.jpg", + "image_caption": [ + "(b) MSST DG" + ], + "image_footnote": [], + "bbox": [ + 290, + 698, + 475, + 797 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Size of the clients. In Fig. 5, we vary the number of training classes in each client $K = \\{5,10,20\\}$ while keeping the size of the dataset fixed. In other words, decreasing the", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "number of classes in each client results in an increase in the total number of clients participating in the federated setup. We observe that for all the baselines a higher heterogeneity (or fewer classes per client) leads to lower performance and vice versa. Nevertheless, FedMVP still exhibits a much superior performance when compared to the baselines, outperforming them while having 4x more clients.", + "bbox": [ + 511, + 287, + 906, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Additional analyses relating to number of shots, choice of LLMs, hyperparameters, and performance in nonfederated setting are reported in Sec. B.5 of the Supp. Mat.", + "bbox": [ + 511, + 393, + 905, + 439 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3d3f7ced71ce3b5d8b3d8e4c47678e50b5baae3d59b6ba9ecde997b81fe49231.jpg", + "image_caption": [ + "(a) Base-to-New", + "Figure 5. Sensitivity to number of classes per client on (a) Baseto-New and (b) MSST DG (DomainNet) setting." + ], + "image_footnote": [], + "bbox": [ + 516, + 446, + 702, + 549 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2d05eb353be34e67921fd9206d04b20888df9223d9bdd61e6f974783837ae0ef.jpg", + "image_caption": [ + "(b) MSST DG" + ], + "image_footnote": [], + "bbox": [ + 712, + 446, + 898, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 609, + 633, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we proposed FedMVP, a novel federated prompt tuning framework for CLIP, to address the challenge of reduced generalization to unseen domains and classes. FedMVP combines visual embeddings with textual attribute features to generate multimodal visual prompts, which are tuned using visual prompt tuning. Our findings revealed that the multimodal contextual information in the prompts enhances the generalizability to unseen classes and domains. FedMVP outperforms all its competitors by a notable margin of $+1.57\\% - 2.26\\%$ across benchmarks.", + "bbox": [ + 511, + 628, + 906, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported by: JPNP23019, subsidized by the New Energy and Industrial Technology Development Organization (NEDO); SERICS (PE00000014) under the NRRP MUR program funded by the EU - NGEU; EU project ELIAS (no: 101120237), ANT (no: 101169439), ELLIOT (no: 101214398) and Ministero delle Imprese e del Made in Italy (IPCEI Cloud DM 27 giugno 2022 - IPCEI-CL-0000007).", + "bbox": [ + 511, + 779, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17876", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 4, 6", + "[2] Sikai Bai, Jie Zhang, Song Guo, Shuaicheng Li, Jingcai Guo, Jun Hou, Tao Han, and Xiaocheng Lu. Diprompt: Disentangled prompt tuning for multiple latent domain generalization in federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27284-27293, 2024. 3", + "[3] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255. IEEE, 2009. 6", + "[4] Lijie Fan, Dilip Krishnan, Phillip Isola, Dina Katabi, and Yonglong Tian. Improving clip training with language rewrites. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[5] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2): 581-595, 2024. 3", + "[6] Ishaan Gulrajani and David Lopez-Paz. In search of lost domain generalization. In International Conference on Learning Representations, 2021. 6, 7", + "[7] Tao Guo, Song Guo, Junxiao Wang, Xueyang Tang, and Wenchao Xu. Promptfl: Let federated participants cooperatively learn prompts instead of models-federated learning in age of foundation model. IEEE Transactions on Mobile Computing, 2023. 2, 3, 6, 7, 8", + "[8] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 6", + "[9] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, page 3, 2022. 5", + "[10] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 2", + "[11] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 1, 2, 3, 6, 7, 8", + "[12] Peter Kairouz, H Brendan McMahan, Brendan Avent, Aurélien Bellet, Mehdi Bennis, Arjun Nitin Bhagoji, Kallista Bonawitz, Zachary Charles, Graham Cormode, Rachel Cummings, et al. Advances and open problems in federated learn" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ing. Foundations and trends® in machine learning, 14(1-2): 1-210, 2021. 3", + "[13] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19113-19122, 2023. 2, 6, 7, 8", + "[14] Hongxia Li, Wei Huang, Jingya Wang, and Ye Shi. Global and local prompts cooperation via optimal transport for federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12151-12161, 2024. 3, 6, 7, 8", + "[15] Wang Lu, HU Xixu, Jindong Wang, and Xing Xie. Fedclip: Fast generalization and personalization for clip in federated learning. In ICLR Workshop on Trustworthy and Reliable Large-Scale Machine Learning Models, 2023. 3, 6, 7, 8", + "[16] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 1, 2, 3, 5", + "[17] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. In International Conference on Learning Representations, 2023. 3", + "[18] Liliane Momeni, Mathilde Caron, Arsha Nagrani, Andrew Zisserman, and Cordelia Schmid. Verbs in action: Improving verb understanding in video-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15579-15591, 2023. 3", + "[19] Khalil Muhammad, Qinqin Wang, Diarmuid O'Reilly-Morgan, Elias Tragos, Barry Smyth, Neil Hurley, James Geraci, and Aonghus Lawlor. Fedfast: Going beyond average for faster training of federated recommender systems. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1234-1242, 2020. 1", + "[20] Sarah Pratt, Ian Covert, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15691-15701, 2023. 3", + "[21] Chen Qiu, Xingyu Li, Chaithanya Kumar Mummadi, Madan Ravi Ganesh, Zhenzhen Li, Lu Peng, and Wan-Yi Lin. Federated text-driven prompt generation for vision-language models. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 3, 5, 6, 7, 8", + "[22] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 5, 6, 7, 8", + "[23] Jinhyun So, Chaoyang He, Chien-Sheng Yang, Songze Li, Qian Yu, Ramy E Ali, Basak Guler, and Salman Avestimehr. Lightsecagg: a lightweight and versatile design for secure aggregation in federated learning. Proceedings of Machine Learning and Systems, 4:694-720, 2022. 1" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17877", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Guangyu Sun, Matias Mendieta, Aritra Dutta, Xin Li, and Chen Chen. Towards multi-modal transformers in federated learning. In European Conference on Computer Vision, pages 229-246. Springer, 2024. 3", + "[25] Zhenyu Sun, Xiaochun Niu, and Ermin Wei. Understanding generalization of federated learning via stability: Heterogeneity matters. In International Conference on Artificial Intelligence and Statistics, pages 676-684. PMLR, 2024. 1", + "[26] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, 2017. 2", + "[27] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, et al. Robust fine-tuning of zero-shot models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7959-7971, 2022. 2", + "[28] Yinghui Xing, Qirui Wu, De Cheng, Shizhou Zhang, Guoqiang Liang, Peng Wang, and Yanning Zhang. Dual modality prompt tuning for vision-language pre-trained model. IEEE Transactions on Multimedia, 26:2056-2068, 2023. 2, 3", + "[29] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language in a bottle: Language model guided concept bottlenecks for interpretable image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19187-19197, 2023. 3", + "[30] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6757-6767, 2023. 2, 6, 7, 8", + "[31] Xin Yao, Chaofeng Huang, and Lifeng Sun. Two-stream federated learning: Reduce the communication costs. In 2018 IEEE Visual Communications and Image Processing (VCIP), pages 1-4. IEEE, 2018. 1", + "[32] Xin Yao, Tianchi Huang, Chenglei Wu, Rui-Xiao Zhang, and Lifeng Sun. Federated learning with additional mechanisms on clients to reduce communication costs. arXiv preprint arXiv:1908.05891, 2019. 1", + "[33] Liping Yi, Gang Wang, Xiaoguang Liu, Zhuan Shi, and Han Yu. Fedgh: Heterogeneous federated learning with generalized global header. In Proceedings of the 31st ACM International Conference on Multimedia, pages 8686-8696, 2023. 1", + "[34] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11975-11986, 2023. 1", + "[35] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 2, 3, 6, 7, 8", + "[36] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. In-" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "ternational Journal of Computer Vision, 130(9):2337-2348, 2022. 1, 2, 3", + "bbox": [ + 545, + 92, + 903, + 119 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17878", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/FedMVP_ Federated Multimodal Visual Prompt Tuning for Vision-Language Models/9a8979ca-f59a-40f4-a042-065bd341c994_model.json b/2025/FedMVP_ Federated Multimodal Visual Prompt Tuning for Vision-Language Models/9a8979ca-f59a-40f4-a042-065bd341c994_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e7c3eef49eabc7752a58d6535e2b639bb7231339 --- /dev/null +++ b/2025/FedMVP_ Federated Multimodal Visual Prompt Tuning for Vision-Language Models/9a8979ca-f59a-40f4-a042-065bd341c994_model.json @@ -0,0 +1,2134 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.13, + 0.885, + 0.175 + ], + "angle": 0, + "content": "FedMVP: Federated Multimodal Visual Prompt Tuning for Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.182, + 0.203, + 0.816, + 0.24 + ], + "angle": 0, + "content": "Mainak Singha1 Subhankar Roy2 Sarthak Mehrotra3 Ankit Jha4 Moloud Abdar5 Biplab Banerjee3 Elisa Ricci1,6" + }, + { + "type": "text", + "bbox": [ + 0.141, + 0.242, + 0.31, + 0.257 + ], + "angle": 0, + "content": "1 University of Trento, Italy" + }, + { + "type": "text", + "bbox": [ + 0.157, + 0.259, + 0.299, + 0.274 + ], + "angle": 0, + "content": "4 LNMIIT Jaipur, India" + }, + { + "type": "text", + "bbox": [ + 0.349, + 0.242, + 0.534, + 0.257 + ], + "angle": 0, + "content": "\\(^{2}\\) University of Bergamo, Italy" + }, + { + "type": "text", + "bbox": [ + 0.342, + 0.259, + 0.594, + 0.274 + ], + "angle": 0, + "content": "5 The University of Queensland, Australia" + }, + { + "type": "text", + "bbox": [ + 0.576, + 0.242, + 0.857, + 0.257 + ], + "angle": 0, + "content": "3 Indian Institute of Technology Bombay, India" + }, + { + "type": "text", + "bbox": [ + 0.638, + 0.259, + 0.842, + 0.274 + ], + "angle": 0, + "content": "Fondazione Bruno Kessler, Italy" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.309, + 0.327, + 0.326 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.344, + 0.485, + 0.707 + ], + "angle": 0, + "content": "In federated learning, textual prompt tuning adapts Vision-Language Models (e.g., CLIP) by tuning lightweight input tokens (or prompts) on local client data, while keeping network weights frozen. After training, only the prompts are shared by the clients with the central server for aggregation. However, textual prompt tuning suffers from overfitting to known concepts, limiting its generalizability to unseen concepts. To address this limitation, we propose Multimodal Visual Prompt Tuning (FedMVP) that conditions the prompts on multimodal contextual information - derived from the input image and textual attribute features of a class. At the core of FedMVP is a PromptFormer module that synergistically aligns textual and visual features through a cross-attention mechanism. The dynamically generated multimodal visual prompts are then input to the frozen vision encoder of CLIP, and trained with a combination of CLIP similarity loss and a consistency loss. Extensive evaluation on 20 datasets, spanning three generalization settings, demonstrates that FedMVP not only preserves performance on in-distribution classes and domains, but also displays higher generalizability to unseen classes and domains, surpassing state-of-the-art methods by a notable margin of \\(+1.57\\% - 2.26\\%\\). Code is available at https://github.com/mainaksingha01/FedMVP." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.723, + 0.222, + 0.739 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.901 + ], + "angle": 0, + "content": "A successful learning paradigm involves training neural networks on large and centralized datasets, enabling models to learn robust and generalizable representations [22, 34]. However, practical constraints often prevent centralized training (e.g., privacy regulations) creating a critical need for decentralized training frameworks that achieve high performance without central data access. As a solution, federated learning (FL) [16] enables multiple clients to collaboratively train a global model without requiring direct data sharing. In FL, clients train models locally and trans" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.311, + 0.895, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.491, + 0.907, + 0.561 + ], + "angle": 0, + "content": "Figure 1. Comparison of federated prompt tuning methods: (a) a textual prompt tuning method (e.g., FedTPG [21]) encodes unimodal contextual information, i.e., class names, in the learnable prompts, (b) our proposed method FedMVP injects multimodal contextual information – image and attributes – into the prompts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.907, + 0.674 + ], + "angle": 0, + "content": "mit only model updates (e.g., gradients or parameters) to a central server, which aggregates these updates to improve the global model. The research in FL is primarily aimed at: reducing communication rounds between clients and server [31, 32], faster and lightweight training algorithms on clients [19, 23] and improving generalization while exhibiting high data heterogeneity across clients [25, 33]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Pretrained Vision-Language Models (VLMs) (e.g., CLIP [22]), have emerged as promising candidates for FL due to their strong generalization capabilities. However, the high communication overhead and computational demands of parameter-heavy VLMs pose challenges for FL, especially in bandwidth- or compute-constrained environments (e.g., robots). To make CLIP viable for the FL setting, prompt tuning (e.g., CoOp [36], VPT [11]) has been adopted as a preferred technique that appends lightweight learnable tokens (or prompts) to the input tokens of the CLIP image and/or text encoder. Each client trains the prompts on the local data, while keeping the CLIP backbone frozen, and then dispatches these prompts to the server, where they are updated through aggregation and sent back. This speeds up client training as each client needs access" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17869" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "to only a few samples per class (e.g., 16/class [36]), and reduces communication overhead as only the lightweight prompts (e.g., \\(0.37\\%\\) of network parameters) need to be communicated, instead of all the parameters." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.155, + 0.483, + 0.381 + ], + "angle": 0, + "content": "Although prompt tuning has proven to be effective in the FL setting, it suffers from reduced generalization to unseen classes and unseen domains - a typical heterogeneous setup native to non independent and identically distributed (non-i.i.d.) FL [21] - due to severe overfitting on the local training data [30]. It is attributed to the fact that prompts learn a static context from a set of seen classes, which being fixed once learned, fail to capture the generalizable elements spread across clients. To circumvent the issue of reduced generalization, recent CLIP-based FL methods condition the prompts either on the class information, as in FedTPG [21] (see Fig. 1a), or on visual information, as in FedCoCoOp [35]. This auxiliary conditioning on the input, or contextual information, makes the prompts dynamic and thus offers a better generalization to unseen distributions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.385, + 0.483, + 0.702 + ], + "angle": 0, + "content": "In this work, we argue that the FL setup being highly heterogeneous, where clients have data from disjoint classes and domains [21], it is not sufficient to solely exploit either of the two modalities for contextual information. Rather, FL demands more comprehensive contextual information that encompasses all the available modalities to condition input prompts. To this end, we propose Federated Multimodal Visual Prompt Tuning (FedMVP) that conditions the prompts on multimodal contextual information, which comprise: (i) visual features of the input image itself, and (ii) textual features of attributes of a class (e.g., \"two legs\", \"beak\", etc. for the class \"hen\"). The key intuition is that (ii) promotes learning transferable prompts for unseen classes that share attributes similar to seen classes (e.g., unseen class \"seagull\" will share some attributes with the seen class \"hen\"), and (i) promotes transferability for unseen classes and domains where attributes cannot be described via text (e.g., texture or abstract concepts). In a nutshell, the dual conditioning offers richer contextual information to the prompts, which helps in generalizing to classes or domains that share similar properties." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In detail, the proposed FedMVP framework contains a PromptFormer network (see Fig. 1b) that synergistically merges the information from visual and textual modalities with cross-attention mechanism [26] to generate the prompts. The resulting multimodal prompts are then injected at the visual space of the frozen vision encoder of CLIP, unlike the existing methods [7, 21, 30] that inject prompts through text encoder. The key differences between our FedMVP and existing CLIP-based FL methods have been summarized in Tab. 1. The PromptFormer is trained in an end-to-end manner with the CLIP loss and a consistency loss. Once trained, each participating client shares the lightweight PromptFormer network parameters with the" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.09, + 0.905, + 0.174 + ], + "angle": 0, + "content": "Table 1. Summary of federated prompt-tuning methods. \\((X)\\) indicates the learnable prompts are initialized randomly, \\((\\checkmark)\\) signifies learnable prompts are conditioned on contextual information (e.g., textual or visual inputs). Prompting column denotes the CLIP encoder wherein learnable prompts are given as input (e.g., \"Textual\" means the text encoder of CLIP takes prompts as input)." + }, + { + "type": "table", + "bbox": [ + 0.545, + 0.184, + 0.871, + 0.285 + ], + "angle": 0, + "content": "
MethodContextual informationPrompting
TextualVisual
FedKgCoOp [30]XXTextual
FedVPT [11]XXVisual
FedTPG [21]✓XTextual
FedCoCoOp [35]X✓Textual
FedMaPLe [13]XXMultimodal
FedMVP (Ours)✓✓Visual
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.303, + 0.905, + 0.453 + ], + "angle": 0, + "content": "server for aggregation through FedAvg [16], and thus enabling knowledge sharing among clients. During inference, the cross-attention learned by the PromptFormer enables the model to dynamically adjust the prompts, enhancing generalizability to unseen classes and domains. We evaluated FedMVP on three generalization settings encompassing 20 datasets that measure generalization to unseen classes, domains, and combinations of them. Experiments demonstrate that FedMVP consistently outperforms the state-of-the-art CLIP-based FL methods by \\(1.57\\% - 2.26\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.455, + 0.907, + 0.59 + ], + "angle": 0, + "content": "In summary, our main contributions are: (i) We highlight the importance of leveraging multimodal contextual information in prompt tuning based FL setup in order to alleviate the issue of reduced generalizability to unseen data. (ii) We propose FedMVP that synergistically combines the visual features of the image and textual features of the attributes of classes to generate multimodal prompts. This provides richer contextual information to the model and higher transferability to unseen classes and domains." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.604, + 0.655, + 0.62 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.905, + 0.719 + ], + "angle": 0, + "content": "Prompt tuning of VLMs. VLMs, such as CLIP [22] and ALIGN [10], leverage large image-text datasets to enable multimodal understanding, performing well in zero-shot classification tasks. However, fully fine-tuning VLMs for downstream tasks without sacrificing their ability to generalize to unseen concepts remains a challenge [27]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.885 + ], + "angle": 0, + "content": "Prompt tuning approaches [11, 13, 36] have been proposed for parameter efficient and effective adaptation of CLIP by introducing a set of continuous learnable tokens (or prompts) onto the frozen CLIP backbone. These approaches can be broadly categorized into three categories depending on the manner in which CLIP is prompted: (i) textual prompting that inject prompts in the input space of the text encoder [30, 36], (ii) visual prompting that injects prompts in the input space of the vision encoder [11], and (iii) multimodal prompting that injects prompts in both the text and vision encoder [13, 28]." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "In particular, MaPLe [13], which injects prompts to both" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17870" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.213 + ], + "angle": 0, + "content": "the text and vision encoders, is not suitable for memory and compute efficient FL since it requires parts of both the encoders to be unlocked during tuning, risking overfitting. Moreover, it does not allow for image-conditioned dynamic prompting, which has shown to be effective for generalization [35]. In contrast, our proposed FedMVP enables multimodal prompting by involving only the vision encoder and supports instance-specific prompting." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.215, + 0.483, + 0.473 + ], + "angle": 0, + "content": "Federated prompt tuning of VLMs. Prompt tuning has emerged as an effective technique for FL of VLMs since it allows clients to collaboratively train on a heterogeneously distributed dataset and adapt faster with fewer samples. The fact that the prompts are lightweight lends well to the FL setup as it reduces communication overhead. Due to the benefits offered by prompt tuning, several works [2, 7, 21, 24] have specifically adapted prompt tuning approaches to the FL setup. For example, PromptFL [7] extended CoOp to FL via sharing prompts with the server, enhancing the generalization performance of client-server interactions in vision tasks. FedCLIP [15] learns a lightweight adapter in each client, similar to the CLIP-Adapter [5], and shares parameters with the server. FedOTP [14] introduced global and local prompts, optimizing shared parameters through optimal transport, while DiPrompt [2] studied multi-domain learning by allowing multiple domains per client." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.476, + 0.483, + 0.657 + ], + "angle": 0, + "content": "Unlike existing methods, a couple of recent works [21, 28] argue that prompt tuning is prone to overfitting to the seen classes due to learning static context. The closest work to ours is FedTPG [21], where a prompt generation network is learned across multiple clients. Additionally, this network leverages textual inputs, enhancing its ability to generalize to unseen classes. Our work significantly differs from previous methods in two key ways: (i) we exploit both attribute information derived from an LLM and image features to condition the prompts, and (ii) we leverage multimodal prompts from the proposed PromptFormer for visual prompt tuning, instead of textual prompt tuning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Classification via textual descriptions. Zero-shot accuracy of CLIP on downstream tasks is sensitive to the quality of text prompts [22]. In detail, some works rely on simple hand-crafted prompts (e.g., \"a photo of a [CLS]\") [22] or others augment the hand-crafted prompts with richer attributes obtained from LLMs [17, 20] (e.g., a \"hen\" has two legs, white feathers, etc.). There also exists a family of methods, such as LaCLIP [4], LaBo [29] and VFC [18], that have proposed to further refine CLIP weights with enriched captions from LLMs for improved performance. Recognizing the importance of class attributes for improved generalization to unseen concepts and domains – an important desideratum in the FL set up – we, for the first time in FL, introduce LLM generated class attributes for enriching the image-conditioned multimodal prompts through the PromptFormer network." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.613, + 0.105 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.116, + 0.907, + 0.149 + ], + "angle": 0, + "content": "We introduce FL setup in Sec. 3.1, preliminaries on prompt tuning in Sec. 3.2 and our proposed solution in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.159, + 0.714, + 0.174 + ], + "angle": 0, + "content": "3.1. Problem formulation" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.182, + 0.905, + 0.381 + ], + "angle": 0, + "content": "We consider a FL setup in which multiple remote clients train models on local data, and a central server aggregates models from the clients. Formally, let each remote client \\(i\\) has access to a dataset \\(\\mathcal{D}_i = \\{\\mathbf{x}_{i,j},y_{i,j}\\}_{j = 1}^{n_i}\\), where \\(\\mathbf{x}_{i,j}\\) is the \\(j^{\\mathrm{th}}\\) image from the \\(i^{\\mathrm{th}}\\) client, \\(y_{i,j}\\) is the corresponding class label and \\(n_i\\) is the number of samples in client \\(i\\). Let the set of text class names in client \\(i\\) be denoted as \\(\\mathcal{C}_i = \\{c_{i,k}\\}_{k = 1}^{K_i}\\), where \\(K_{i}\\) is the total number of classes. As in FedTPG [21], we follow a non-IID FL setup, where the samples in each client are drawn from a disjoint set of classes, i.e., \\(\\mathcal{C}_i\\cap \\mathcal{C}_l = \\emptyset\\). As an example, the training data on each mobile device usually depend on the user, which may not be representative of the general population [16]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.382, + 0.906, + 0.487 + ], + "angle": 0, + "content": "In particular, we investigate two kinds of non-IID FL setups, where, in addition to the in-distribution performance, we also evaluate the generalization ability of the aggregated server model on (i) unseen classes, and (ii) unseen domains or datasets. The need for generalization to unseen classes and domains makes the non-IID FL setup more challenging than the traditional FL setting [12]." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.5, + 0.8, + 0.516 + ], + "angle": 0, + "content": "3.2. Preliminaries on Prompt Tuning" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.523, + 0.906, + 0.704 + ], + "angle": 0, + "content": "Textual prompt tuning (TPT), i.e., CoOp [36], has emerged as an efficient technique in CLIP-based FL that keeps the CLIP backbone frozen and locally learns lightweight prompt vectors on each client. The locally learned prompts are then communicated to the server, as in PromptFL [7]. Specifically, the hand-crafted text prompts (e.g., \"a photo of a [CLS]\") input to the text encoder \\(\\mathcal{E}_t\\) is augmented with the learnable soft prompts \\(v_{1}, v_{2}, \\ldots, v_{m}\\). However, as shown in [35], TPT can often lead to overfitting and reduced generalization due to the static context learned on the seen classes. This can be exacerbated in the non-IID FL setup due to the high data heterogeneity among the clients [21]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.705, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Alternatively, visual prompt tuning (VPT) [11] adapts pre-trained vision encoder \\(\\mathcal{E}_v\\) by injecting small number of visual learnable parameters into ViT's input space. More formally, let an input image \\(\\mathbf{x}\\) be divided in \\(b\\) patches, yielding a collection of patch embeddings (augmented with positional encodings) at the input of the vision encoder as \\(\\mathbf{E} \\in \\mathbb{R}^{b \\times d_v}\\), where \\(d_v\\) is the embedding dimension. A [CLS] token \\(\\mathbf{z}\\) is concatenated with \\(\\mathbf{E}\\), encapsulating the initial input representation for the ViT layers of \\(\\mathcal{E}_v\\). VPT further augments the input by concatenating a set of randomly initialized learnable visual prompts \\(\\mathbf{P} = \\{p_1, p_2, \\ldots, p_m\\}\\), where \\(m\\) is the prompt length. The first layer input of \\(\\mathcal{E}_v\\) is then redefined as \\(\\mathbf{I} = [\\mathbf{z}; \\mathbf{E}; \\mathbf{P}] \\in \\mathbb{R}^{(1 + b + m) \\times d_v}\\), where" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17871" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.09, + 0.332, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.333, + 0.091, + 0.603, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.091, + 0.844, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.33, + 0.906, + 0.388 + ], + "angle": 0, + "content": "Figure 2. Overview of Federated Multimodal Visual Prompt Tuning (FedMVP). Each client \\(i\\) trains the parameters of the PromptFormer network \\(f_{\\Theta}\\) and multi-head self-attention on its local data. At each round, the server aggregates the trainable parameters from the clients and sends the updated parameters back to the clients. The PromptFormer uses cross-attention mechanism to fuse the visual and textual attribute feature to generate multimodal visual prompts \\(\\mathbf{P}\\), which are then used for visual prompt tuning." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.395, + 0.482, + 0.441 + ], + "angle": 0, + "content": "“;” means concatenation. Although VPT enables more nuanced understanding of the visual space, it suffers from the same issues as TPT, i.e., limited generalization (see Sec. 4)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.451, + 0.476, + 0.467 + ], + "angle": 0, + "content": "3.3. Federated Multimodal Visual Prompt Tuning" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.473, + 0.483, + 0.716 + ], + "angle": 0, + "content": "Our proposed method FedMVP deviates from the TPT and VPT paradigms and incorporates multimodal contextual information from both the visual and textual modalities to improve generalization in the FL setup. As shown in Fig. 2, at the core of our novel architecture is a prompt generator, called PromptFormer, which conditions the visual prompts \\(\\mathbf{P}\\) on the task-related information: input image and text attributes corresponding to the class names, rather than randomly initializing them. This makes \\(\\mathbf{P}\\) context-aware, generalizing better to unseen classes and domains. Once trained on each client, only the parameters of the lightweight PromptFormer network are communicated to the server, thus, keeping the communication overhead low. Next, we describe PromptFormer (Sec. 3.3.1), training objectives (Sec. 3.3.2) and how we conduct client-side training and server-side aggregation (Sec. 3.3.3)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.724, + 0.466, + 0.74 + ], + "angle": 0, + "content": "3.3.1. Visual Prompt Generation with PromptFormer" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.743, + 0.483, + 0.864 + ], + "angle": 0, + "content": "We introduce the PromptFormer network \\( f_{\\Theta} \\), parameterized by \\( \\Theta \\), designed to generate contextually rich visual tokens (or prompts) for each client \\( i \\) by leveraging both attribute features and visual patch embeddings (see Fig. 2). Concretely, we use the CLIP text encoder \\( \\mathcal{E}_t \\) to extract embeddings \\( \\mathbf{A}_i = \\{\\mathcal{E}_t(\\mathrm{LLM}(c_k))\\}_{k=1}^K \\) of the text attributes \\( \\mathcal{A}_i \\) corresponding to the text class names \\( \\mathcal{C}_i \\), generated using a large lanuage model (LLM) [1] (see Sec. A.1 of Supp. Mat." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.395, + 0.906, + 0.47 + ], + "angle": 0, + "content": "for Attribute Generation process). Going forward, we drop the client index \\(i\\) for notational clarity. The PromptFormer takes as input \\(b\\) visual patch embeddings \\(\\mathbf{E}\\), corresponding to an image \\(\\pmb{x}\\), and the text embeddings \\(\\mathbf{A}\\) to generate a set of enriched \\(m\\) multimodal visual prompts \\(\\mathbf{P} \\in \\mathbb{R}^{m \\times d_v}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.479, + 0.905, + 0.497 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} = \\left\\{p _ {j} \\right\\} _ {j = 1} ^ {m} = f _ {\\Theta} (\\mathbf {A}, \\mathbf {E}). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.906, + 0.583 + ], + "angle": 0, + "content": "Consequently, the input to the first layer of the visual encoder \\(\\mathcal{E}_v\\) is redefined as \\([\\mathbf{z};\\mathbf{E};\\mathbf{P}]\\in \\mathbb{R}^{(1 + b + m)\\times d_v}\\). Intuitively, multimodal visual prompts carry richer contextual information, and helps the network in generalizing to unseen classes and domains that share commonalities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.906, + 0.716 + ], + "angle": 0, + "content": "Architecture of PromptFormer. The PromptFormer architecture comprises: lightweight multi-head cross-attention (MHCA) modules, feed-forward network (FFN) and a linear projection layer; transforming multimodal features into comprehensive visual prompts. In detail, the visual patch embedding \\(\\mathbf{E}\\) interacts with the linearly projected attribute features \\(\\mathbf{A}' = T_{\\mathrm{proj}}(\\mathbf{A})\\) to construct multimodal visual prompts as:" + }, + { + "type": "equation", + "bbox": [ + 0.533, + 0.728, + 0.885, + 0.746 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} \\left(\\mathbf {A} ^ {\\prime}, \\mathbf {E}\\right) = \\operatorname {F F N} \\left(\\text {C r o s s A t t e n t i o n} \\left(\\mathbf {Q} _ {\\mathbf {E}}, \\mathbf {K} _ {\\mathbf {A} ^ {\\prime}}, \\mathbf {V} _ {\\mathbf {A} ^ {\\prime}}\\right)\\right);\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.749, + 0.905, + 0.78 + ], + "angle": 0, + "content": "\\[\n\\text {w i t h} \\mathbf {Q} _ {\\mathbf {E}} = \\mathbf {E} W _ {\\mathbf {Q}}, \\mathbf {K} _ {\\mathbf {A} ^ {\\prime}} = \\mathbf {A} ^ {\\prime} W _ {\\mathbf {K}}, \\mathbf {V} _ {\\mathbf {A} ^ {\\prime}} = \\mathbf {A} ^ {\\prime} W _ {\\mathbf {V}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.905, + 0.841 + ], + "angle": 0, + "content": "where, \\(\\mathbf{E}\\) are transformed into query vectors \\(\\mathbf{Q}_{\\mathbf{E}}\\) using query matrices \\(W_{\\mathbf{Q}}\\); and the projected attribute features \\(\\mathbf{A}'\\) are transformed into key and value vectors \\(\\mathbf{K}_{\\mathbf{A}'}\\) and \\(\\mathbf{V}_{\\mathbf{A}'}\\), using key and value matrices \\(W_{\\mathbf{K}}\\) and \\(W_{\\mathbf{V}}\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Intuitively, through the cross-attention mechanism, visual features learn to attend to the relevant parts of the corresponding attribute features. For example, the patches depicting the legs of the class \"dog\" will learn to attend to" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.875, + 0.483, + 0.902 + ], + "angle": 0, + "content": "1For example, for the class name \"giraffe\" the LLM outputs \"Exceptionally long neck, unique coat pattern with irregular brown patches, ...\"" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17872" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.168 + ], + "angle": 0, + "content": "the attributes four legged. When an unseen class (e.g., \"giraffe\") comprises an animal with four legs, the final prompts \\(\\mathbf{P}\\) will contain this relevant information derived from the attributes four legged. Additional details of PromptFormer architecture are reported in Sec. A.3 of the Supp. Mat." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.175, + 0.283, + 0.189 + ], + "angle": 0, + "content": "3.3.2. Training of FedMVP" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.194, + 0.483, + 0.255 + ], + "angle": 0, + "content": "After the visual prompts are obtained with PromptFormer as per Eq. 2, we concatenate the [CLS] token, the patch embeddings, and the visual prompts to get \\(\\mathbf{I} = [\\mathbf{z};\\mathbf{E};\\mathbf{P}]\\) following Sec. 3.2, and feed it to the visual encoder \\(\\mathcal{E}_v\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.255, + 0.484, + 0.316 + ], + "angle": 0, + "content": "The parameters of PromptFormer \\(\\rho = \\{\\Theta\\}\\) are the only trainable parameters of FedMVP, and the rest of the weights are kept frozen. To train the model, we employ the CLIP cross-entropy loss [22], formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.323, + 0.483, + 0.348 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c e}} = - \\underset {(\\mathbf {x}, y) \\sim \\mathcal {D}} {\\mathbb {E}} y \\log p (y | \\mathbf {I}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.358, + 0.484, + 0.418 + ], + "angle": 0, + "content": "where \\(y\\) denotes the ground truth label, and the prediction probability of CLIP between the visual features \\(\\mathbf{v} = \\mathcal{E}_v(\\mathbf{I})\\) and text features \\(\\mathbf{t}_k = \\mathcal{E}_t([A\\text{ photo of [CLASS]};\\mathrm{LLM}(c_k)])\\) for class \\(k\\) is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.42, + 0.483, + 0.457 + ], + "angle": 0, + "content": "\\[\np (y = k | \\mathbf {I}) = \\frac {\\exp \\left(\\cos \\left(\\mathbf {v} , \\mathbf {t} _ {k}\\right)\\right) / \\tau)}{\\sum_ {k ^ {\\prime} = 1} ^ {K} \\exp \\left(\\cos \\left(\\mathbf {v} , \\mathbf {t} _ {k ^ {\\prime}}\\right)\\right) / \\tau}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.467, + 0.484, + 0.512 + ], + "angle": 0, + "content": "where \\(\\cos (\\cdot ,\\cdot)\\) as a similarity function, \\(\\tau\\) as the temperature parameter, and \\([\\cdot ]\\) denotes concatenation. Details on how \\(\\mathbf{t}_k\\) is computed are provided in Sec. A.1 of the Supp. Mat." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.513, + 0.484, + 0.558 + ], + "angle": 0, + "content": "In addition to the \\(\\mathcal{L}_{\\mathrm{ce}}\\), we employ a regularization constraint to ensure consistency between two augmented views of an image. We define the consistency loss \\(\\mathcal{L}_{\\mathrm{con}}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.57, + 0.483, + 0.587 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c o n}} = 1 - \\cos \\left(\\mathcal {E} _ {v} (\\mathbf {I}), \\mathcal {E} _ {v} \\left(\\mathbf {x} ^ {\\prime}\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.483, + 0.629 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}'\\) is the augmented version of \\(\\mathbf{x}\\). Finally, the total loss function is given as:" + }, + { + "type": "equation", + "bbox": [ + 0.209, + 0.633, + 0.483, + 0.649 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\mathrm {c e}} + \\alpha \\cdot \\mathcal {L} _ {\\mathrm {c o n}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.663, + 0.483, + 0.694 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is a balancing coefficient for the regularization term, ensuring both alignment and consistency." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.701, + 0.468, + 0.717 + ], + "angle": 0, + "content": "3.3.3. Client-side training and server-side aggregation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.87 + ], + "angle": 0, + "content": "In our approach, multiple remote clients participate in distinct image classification tasks, collaboratively training the PromptFormer network, parameterized by \\(\\rho = \\{\\Theta\\}\\). In each communication round, the client transmits \\(\\rho\\) to the server, which is aggregated at the server. Through training on a diverse set of classification tasks distributed across clients, and aggregating the parameters at the server, the resulting model achieves generalization capability to classes and domains unseen during training. Pseudo-code of FedMVP can be found in Sec. A.2 of the Supp. Mat." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Initially, the server randomly initializes the parameters \\(\\rho^0\\) and in each communication round only a random subset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "of remote clients \\( S^r \\) will be chosen by the server for sharing the updated parameters \\( \\rho \\). Each communication round \\( r \\) involves the following steps:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.137, + 0.905, + 0.167 + ], + "angle": 0, + "content": "(i) Model initialization: Each client \\(i\\) in \\(S_{r}\\) receives the latest model parameters \\(\\rho^r\\) to configure its local model." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.168, + 0.905, + 0.213 + ], + "angle": 0, + "content": "(ii) Local optimization: Following Eq. (6), client \\(i\\) optimizes its local parameters \\(\\rho^r\\) using an SGD optimizer over \\(J\\) iterations with learning rate \\(\\eta^r\\). Updated parameters:" + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.137, + 0.905, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.22, + 0.905, + 0.239 + ], + "angle": 0, + "content": "\\[\n\\rho_ {i} ^ {r + 1} = \\operatorname {S G D} _ {J} \\left(\\eta^ {r}, \\rho^ {r}, \\mathcal {C} _ {i}, \\mathcal {D} _ {i}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.25, + 0.905, + 0.295 + ], + "angle": 0, + "content": "(iii) Parameter aggregation: After local training, each client \\(i\\) in \\(S_{r}\\) transmits its updated parameters \\(\\rho_{i}^{r + 1}\\) back to the server, where they are aggregated [16] as:" + }, + { + "type": "equation", + "bbox": [ + 0.629, + 0.304, + 0.905, + 0.34 + ], + "angle": 0, + "content": "\\[\n\\rho^ {r + 1} = \\frac {1}{\\left| S _ {r} \\right|} \\sum_ {i \\in S _ {r}} \\rho_ {i} ^ {r + 1}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.35, + 0.905, + 0.409 + ], + "angle": 0, + "content": "The above steps are repeated for \\( R \\) communication rounds, after which the server converges on the final model parameters \\( \\rho^R \\). This iterative process facilitates collaborative and distributed optimization of \\( \\rho \\) across client data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.907, + 0.652 + ], + "angle": 0, + "content": "Lightweight fine-tuning in clients. As the FL setup exhibits high data heterogeneity, different clients possess varying numbers of samples. Thus, treating every client as equal and repeatedly fine-tuning the parameters \\(\\rho\\) can lead to severe overfitting in some clients, especially those with very limited training data. To prevent this, we propose to inject low-rank decomposition matrices, or LoRA [9], to the PromptFormer module, and fine-tune these LoRA matrices if training loss \\(\\mathcal{L}_{\\mathrm{total}}\\) in a client starts below a certain loss value \\(\\sigma\\), which we set to 0.5. In particular, only the LoRA weights corresponding to the query, key and value matrices of the attention layers are trained, while freezing the parameters \\(\\rho\\). In that case, the client shares only the LoRA matrices with the server. This not only helps in boosting generalization, but also brings down the communication overhead of clients by around \\(\\times 267\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.665, + 0.645, + 0.682 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.905, + 0.84 + ], + "angle": 0, + "content": "We evaluate our proposed FedMVP in three distinct settings that are aimed to gauge the (i) generalization to unseen categories (or Base-to-New Generalization) in Sec. 4.1, (ii) generalization to unseen domains (or Domain Generalization) in Sec. 4.2, and (iii) generalization to unseen datasets (or Cross-Dataset Generalization) in Sec. 4.3. Finally, we report the results of our ablation study in Sec. 4.4 that are aimed to assess impact of the design choices made in FedMVP. Due to the lack of space, we report additional experimental results in Sec. B of the Supp. Mat." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Benchmarks and datasets. Following FedTPG [21], we have reported results on the Base-to-New Generalization setting that encompasses nine medium-sized image recognition datasets. For the Domain Generalization setting we" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17873" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "have used the ImageNet [3] and DomainBed [6] benchmarks. In addition, we have also included EuroSAT [8] for the Cross-Dataset Generalization setting, alongside all nine datasets of the Base-to-New Generalization setting. In total, we have used 20 datasets for a comprehensive experimental evaluation. Detailed dataset descriptions and statistics can be found in Sec. B.1 of the Supp. Mat." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.199, + 0.482, + 0.334 + ], + "angle": 0, + "content": "Baselines. We have compared our method with the state-of-the-art prompt tuning approaches that use the CLIP backbone [22]: KgCoOp [30], CoCoOp [35], VPT [11], and MaPLe [13], which were originally designed for offline and non-FL setup, but adapted to the federated setting, as in FedTPG [21]. Furthermore, we have compared with CLIP-based methods specifically designed to work in the FL setting: FedTPG [21], FedOTP [14], PromptFL [7], and FedCLIP [15]. We also included zero-shot CLIP (ZS-CLIP)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.336, + 0.483, + 0.532 + ], + "angle": 0, + "content": "Implementation details. We have employed the frozen ViT-B/16 CLIP backbone as the vision encoder, and GPT-4o [1] as the LLM in our experiments. Each MHCA module of \\( f_{\\Theta} \\), consists of a 4-head cross-attention layers, complemented by layer normalization and each FFN comprising of a two-layer bottleneck structure (Linear-GeLU-Linear). The query prompt \\( \\mathbf{Q} \\) has a length \\( m = 4 \\) and matches the patch embedding dimension \\( d_v = 768 \\). The projection layer \\( T_{\\mathrm{proj}} \\) maps the textual dimension of 512 to the patch embedding dimension via a linear transformation. In Eq. (6), we keep \\( \\alpha = 10 \\). We have set the batch size to 128, number of shots per class to 8, and the learning rate equal to 0.003 with a decay rate of 1e-5 and momentum of 0.9." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.545, + 0.342, + 0.56 + ], + "angle": 0, + "content": "4.1. Base-to-New Generalization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.483, + 0.809 + ], + "angle": 0, + "content": "Experimental setup. To recap, the base-to-new generalization setting encompasses nine different datasets. The setting is characterized by splitting the classes from each dataset into two groups – base and new. The samples of the base classes are used for training, and the samples of the new classes are held out to test the generalization performance [21]. Furthermore, the samples of bases classes from all nine datasets are distributed to the clients, such that each client does not see samples from more than 20 disjoint classes (referred to as local classes). We report the model's classification accuracy on: (i) the client's local classes, (ii) all the base classes (spread across clients), and (iii) the held-out new classes that were never seen by any client. Note that (ii) and (iii) is computed using the aggregated model on the server, and (i) is computed using the model on the client (which is the same server model after convergence)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Main results. In Table 2 we report the performance in the base-to-new generalization setting, summarized across nine datasets. The results show that the textual and multimodal prompting methods, such as PromptFL, FedKgCoOp, and FedMaPLe, excel on the seen local and base class accuracy, but have at-par or worse generalization on the new" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.905, + 0.173 + ], + "angle": 0, + "content": "Table 2. Base-to-New Generalization setting. Accuracies \\((\\%)\\) on client's local (seen) classes, base (seen) classes and new (unseen) classes. Harmonic mean (HM) is computed with Base and New class accuracies. Best number in each column is in bold and second-best underlined. Improvement row reports the absolute difference between the best and second-best numbers." + }, + { + "type": "table", + "bbox": [ + 0.58, + 0.178, + 0.835, + 0.323 + ], + "angle": 0, + "content": "
MethodsLocalBaseNewHM
ZS-CLIP [22]76.7270.5175.7874.24
FedOTP [14]74.8265.2257.0464.89
FedCoCoOp [35]81.4673.7666.0073.20
FedVPT [11]76.2970.4374.8973.79
FedCLIP [15]76.8771.0475.0674.24
FedMaPLe [13]81.6374.4470.6275.29
FedKgCoOp [30]78.3872.1875.8775.39
PromptFL [7]81.7574.4771.7075.74
FedTPG [21]80.7573.6876.0276.70
FedMVP (Ours)81.8975.3777.8278.27
Improvement+0.14+0.90+1.79+1.57
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.334, + 0.907, + 0.531 + ], + "angle": 0, + "content": "classes than the ZS-CLIP model. This hints at the potential overfitting caused by learning static context from the seen classes. In contrast, our proposed FedMVP that considers multimodal contextual information and dynamic conditioning demonstrates superior performance in both the seen and unseen classes, achieving the highest HM accuracy. While FedTPG also generates prompts based on task-related input, it does not outperform our FedMVP due to lack of comprehensive visual and attribute information in the prompts. In summary, FedMVP surpasses the state-of-the-art methods by a notable margin of \\(+1.57\\%\\) in the base-to-new generalization setting. Detailed breakdown for each of the nine datasets is reported in Tab. A.4 in the Supp. Mat." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.544, + 0.729, + 0.558 + ], + "angle": 0, + "content": "4.2. Domain Generalization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.905, + 0.809 + ], + "angle": 0, + "content": "Experimental setup. We conduct experiments under two distinct domain generalization (DG) setups: Multi-source Single-target (MSST) and Single-source Multi-target (SSMT). In the MSST DG setting for a given benchmark, we follow a leave-one-domain-out protocol for evaluation, where we hold out one domain for measuring generalization and use the rest of the domains for training the clients. Training classes are split among clients under the constraint that each client can see images from a single domain. This setting mirrors a practical scenario where images taken with a smartphone uniquely reflect the taste of the user or socio-economic biases. In a similar manner, in the SSMT DG setting, we use one domain for training the clients in a distributed manner, while the rest of the domains are used for evaluation. More details on this experimental setup can be found in Sec. B.4 of the Supp. Mat." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Main results. In Table 3, we report the performance of FedMVP and its competitors in the MSST and SSMT DG settings. In particular, SSMT DG is a more challenging setting than the MSST DG setting, as the clients are trained only on a single domain, and expected to generalize to multiple unseen domains. We observe that ZS-CLIP is a strong base" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "17874" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.131 + ], + "angle": 0, + "content": "Table 3. Domain Generalization setting on the DomainBed [6] benchmark. Accuracies \\((\\%)\\) on the unseen target domain(s) for the Multi-source Single-target (MSST) and Single-source Multi-target (SSMT) settings. Best number in each column is in bold and second-best underlined. Improvement row reports the absolute difference between the best and second-best numbers." + }, + { + "type": "table", + "bbox": [ + 0.143, + 0.135, + 0.85, + 0.288 + ], + "angle": 0, + "content": "
MethodMulti-source Single-targetSingle-source Multi-target
PACSOfficeHomeVLCSTerra Inc.DomainNetAveragePACSOfficeHomeVLCSTerra Inc.DomainNetAverage
ZS-CLIP [22]96.1681.4983.2933.9857.1370.4196.1681.4983.2933.9857.1370.41
FedOTP [14]90.7176.4267.4113.2449.6759.4991.1774.5365.1615.1139.3757.07
FedCoCoOp [35]85.0681.4261.7323.6857.0861.7984.5650.8761.2523.0758.2855.61
FedTPG [21]90.9982.7869.7726.7956.8265.4390.7182.6067.6322.5158.3464.36
PromptFL [7]95.3781.7474.8725.0256.8766.7795.8381.7266.9724.1758.2765.39
FedKgCoOp [30]95.4281.8274.9025.0356.9266.8295.8181.7067.4924.2058.3065.50
FedMaPLe [13]94.5182.0371.7936.3058.8868.7094.2581.4872.0833.5957.6267.80
FedVPT [11]95.3681.7683.1933.6255.9869.9894.7980.9982.4432.2355.6669.22
FedCLIP [15]96.2981.7482.7036.5857.8571.0396.2981.6281.8136.4457.4270.72
FedMVP (Ours)97.2884.1585.1237.3661.1773.0297.1783.8984.6136.9260.5672.63
Improvement+0.99+1.37+1.83+0.78+2.29+1.99+0.88+1.29+1.32+0.48+2.22+1.91
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.293, + 0.483, + 0.389 + ], + "angle": 0, + "content": "Table 4. Domain Generalization setting on the ImageNet benchmark. Accuracies (\\%) on the seen source domain ImageNet (IN) and unseen target domains: ImageNet-V2 (INV2), ImageNet-Sketch (IN-S), ImageNet-A (IN-A) and ImageNet-R (IN-R). Best number in each column is in bold and second-best underlined. Improvement row reports the absolute difference between the best and second-best numbers." + }, + { + "type": "table", + "bbox": [ + 0.124, + 0.394, + 0.446, + 0.54 + ], + "angle": 0, + "content": "
MethodSourceTarget
ININV2IN-SIN-AIN-RAverage
ZS-CLIP [22]66.7560.7946.1247.7974.0057.18
FedOTP [14]51.6845.5334.7316.6463.9840.22
FedMaPLe [13]66.9660.6544.6946.2474.6256.55
FedVPT [11]66.9260.3446.4348.0374.5657.34
PromptFL [7]67.8061.5945.6148.7874.4957.62
FedCLIP [15]67.2661.3546.6648.2474.3657.65
FedKCoOp [30]67.5361.6046.6948.3774.7157.84
FedCoCoOp [35]68.5162.2946.9050.3376.4959.00
FedTPG [21]69.5162.9047.6549.9776.3559.22
FedMVP (Ours)70.8763.7250.9351.7677.2360.91
Improvement+1.36+0.82+3.28+1.43+0.74+1.69
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.547, + 0.483, + 0.682 + ], + "angle": 0, + "content": "line that outperforms all textual and multimodal prompting methods, with the exception of FedCLIP and our proposed FedMVP. This suggests that suboptimal prompt tuning can lead to overfitting on the source training domains, thus hurting generalization. Our FedMVP owing to the use of contextual information learns more transferable representations that generalize better to unseen target domains, thus outperforming all competitor methods in most of the benchmarks. More detailed results in Sec. B.4 of the Supp. Mat." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.683, + 0.483, + 0.804 + ], + "angle": 0, + "content": "In Table. 4 we report the results in the SSMT DG setting on ImageNet benchmark. We observe that FedTPG and FedCoCoOp come close to our FedMVP in terms of average performance, with FedMVP outperforming all baselines by \\(+1.69\\%\\). In particular, we believe FedMVP improves the performance on ImageNet-Sketch (IN-S) by a large margin of \\(+3.28\\%\\) due to the use of attribute information, which remains unchanged between real and sketch images." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.809, + 0.352, + 0.823 + ], + "angle": 0, + "content": "4.3. Cross-dataset Generalization" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.825, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Experimental setup. In the Cross-Dataset Generalization setting the classes from ImageNet are split among the clients for training, in a disjoint manner, and the resulting network is evaluated on 10 held out datasets not seen during training. In other words, it is a combination of base-to-new" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.295, + 0.905, + 0.323 + ], + "angle": 0, + "content": "generalization and DG settings, since it requires generalization to both unseen class and data distributions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.326, + 0.907, + 0.508 + ], + "angle": 0, + "content": "Main results. In Table. 5 we report the results of the Cross-Dataset Generalization setting. This setting is considerably challenging, as the model should generalize to different datasets. We find that the results are consistent with the previous findings, where our proposed FedMVP generalizes better to unseen classes and data distributions, outperforming all the text-based and multimodal-based prompting techniques by \\(+2.26\\%\\) on average. However, with FedMVP we notice a drop in performance with respect to the ZS-CLIP baseline for some fine-grained datasets, such as OxfordPets and StanfordCars, which is potentially due to overlapping attributes for visually similar categories." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.518, + 0.667, + 0.534 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.536, + 0.906, + 0.717 + ], + "angle": 0, + "content": "Impact of proposed components in FedMVP. In Tab. 6 we ablate each component of FedMVP on the base-to-new and MSST DG settings. We observe that each component contributes positively to the final performance. First, Prompt-Former \\( f_{\\Theta} \\) plays a major role in combining textual information with visual features, thus improving semantic alignment between the two modalities. Second, adding \\( \\mathcal{L}_{\\mathrm{con}} \\) (as in Eq. (5)) improves the performance by a small margin. Finally, the impact of LoRA can be observed in the penultimate row, where we fine-tune all the parameters of FedMVP. It shows the dangers of overfitting in the FL setup, as some clients may have limited data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.718, + 0.906, + 0.854 + ], + "angle": 0, + "content": "Participation Rate of the Clients. In Fig. 3 we vary the participation rate from \\(10\\%\\) to \\(100\\%\\) and show how the participation rate of the clients impacts performance on both seen and unseen classes and domains. We observe that FedMVP always exhibits a gentle, yet positive slope in both settings, indicating that continuous parameter aggregation improves generalization. In contrast, for some competitor methods, such as PromptFL, increased participation leads to saturated performance or negative slope in the plot." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Computational analysis. In Fig. 4 we plot the accuracy vs. communication rounds for FedMVP and competitors in base-to-new and MSST settings. While FedMVP trans" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "17875" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.131 + ], + "angle": 0, + "content": "Table 5. Cross-Dataset Generalization setting. Accuracies \\((\\%)\\) on the seen classes of the source dataset, i.e., ImageNet, and the unseen classes of the target datasets. Best number in each column is in bold and second-best underlined. Improvement row reports the absolute difference between the best and second-best numbers." + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.135, + 0.86, + 0.282 + ], + "angle": 0, + "content": "
MethodSourceTarget
ImageNetCaltech101Flowers102FGVCAircraftUCF101OxfordPetsFood101DTDStanfordCarsSUN397EuroSATAverage
ZS-CLIP [22]66.7592.9071.2924.7266.7589.1586.0944.3365.2962.5947.6865.08
FedOTP [14]51.6888.6048.3610.9546.1073.4361.0141.5529.6148.3450.2649.82
FedMaPLe [13]66.9692.4968.2523.5260.3289.6783.5244.6860.1661.8545.3862.98
PromptFL [7]67.8091.8768.1321.4464.1388.7085.8542.4363.5962.7743.2663.22
FedCLIP [15]67.2693.3967.1924.0365.6488.2885.1444.4465.5063.4241.6763.87
FedKgCoOp [30]67.5393.6369.3123.0664.4688.5585.3744.7464.9963.8543.2964.13
FedVPT [11]68.5691.4269.2421.0965.7087.7985.4544.5765.0064.6847.8264.28
FedCoCoOp [35]68.5194.1166.3420.7962.7589.0485.4043.2063.9864.0255.4064.50
FedTPG [21]69.5193.7570.0423.2264.7290.6085.9146.2563.9866.7847.8665.31
FedMVP (Ours)70.8795.3772.8025.9470.5889.2787.0649.7865.8368.1950.8467.57
Improvement+1.36+1.62+1.51+1.22+3.83-1.33+0.97+3.53+0.33+1.41-4.56+2.26
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.286, + 0.484, + 0.315 + ], + "angle": 0, + "content": "Table 6. Ablating FedMVP components. We report numbers for the Base-to-New and MMST DG (on DomainBed) settings." + }, + { + "type": "table", + "bbox": [ + 0.174, + 0.318, + 0.401, + 0.402 + ], + "angle": 0, + "content": "
ComponentsBase-to-NewMSST DG
ZS-CLIP74.2470.41
fΘ only75.9471.85
fΘ + Lcon76.2772.14
w/o LoRA77.4172.58
FedMVP (Full)78.2773.02
" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.406, + 0.28, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.143, + 0.511, + 0.232, + 0.522 + ], + "angle": 0, + "content": "(a) Base-to-New" + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.405, + 0.476, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.345, + 0.51, + 0.424, + 0.522 + ], + "angle": 0, + "content": "(b) MSST DG" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.528, + 0.484, + 0.556 + ], + "angle": 0, + "content": "Figure 3. Sensitivity to participation rate of the clients, on (a) Base-to-New and (b) MSST DG (DomainBed) settings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.575, + 0.484, + 0.68 + ], + "angle": 0, + "content": "mits \\(2 \\times\\) more parameters than FedTPG, it converges nearly \\(10 \\times\\) faster, requiring fewer rounds. In particular, as training progresses, only the LoRA parameters in PromptFormer are fine-tuned in a client, thus reducing the trainable parameter count by \\(267 \\times\\). Overall, FedMVP has lower total computational cost with respect to other methods when normalized by the number of communication rounds." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.697, + 0.28, + 0.798 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.143, + 0.803, + 0.232, + 0.814 + ], + "angle": 0, + "content": "(a) Base-to-New" + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.699, + 0.476, + 0.798 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.345, + 0.802, + 0.423, + 0.814 + ], + "angle": 0, + "content": "(b) MSST DG" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.82, + 0.483, + 0.848 + ], + "angle": 0, + "content": "Figure 4. Computational analysis of the methods on (a) Baseto-New and (b) MSST DG (DomainNet) setting." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Size of the clients. In Fig. 5, we vary the number of training classes in each client \\( K = \\{5,10,20\\} \\) while keeping the size of the dataset fixed. In other words, decreasing the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.907, + 0.394 + ], + "angle": 0, + "content": "number of classes in each client results in an increase in the total number of clients participating in the federated setup. We observe that for all the baselines a higher heterogeneity (or fewer classes per client) leads to lower performance and vice versa. Nevertheless, FedMVP still exhibits a much superior performance when compared to the baselines, outperforming them while having 4x more clients." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.395, + 0.906, + 0.44 + ], + "angle": 0, + "content": "Additional analyses relating to number of shots, choice of LLMs, hyperparameters, and performance in nonfederated setting are reported in Sec. B.5 of the Supp. Mat." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.447, + 0.703, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.554, + 0.655, + 0.565 + ], + "angle": 0, + "content": "(a) Base-to-New" + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.447, + 0.899, + 0.549 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.768, + 0.553, + 0.846, + 0.565 + ], + "angle": 0, + "content": "(b) MSST DG" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.571, + 0.906, + 0.599 + ], + "angle": 0, + "content": "Figure 5. Sensitivity to number of classes per client on (a) Baseto-New and (b) MSST DG (DomainNet) setting." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.61, + 0.634, + 0.625 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.779 + ], + "angle": 0, + "content": "In this work, we proposed FedMVP, a novel federated prompt tuning framework for CLIP, to address the challenge of reduced generalization to unseen domains and classes. FedMVP combines visual embeddings with textual attribute features to generate multimodal visual prompts, which are tuned using visual prompt tuning. Our findings revealed that the multimodal contextual information in the prompts enhances the generalizability to unseen classes and domains. FedMVP outperforms all its competitors by a notable margin of \\(+1.57\\% - 2.26\\%\\) across benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported by: JPNP23019, subsidized by the New Energy and Industrial Technology Development Organization (NEDO); SERICS (PE00000014) under the NRRP MUR program funded by the EU - NGEU; EU project ELIAS (no: 101120237), ANT (no: 101169439), ELLIOT (no: 101214398) and Ministero delle Imprese e del Made in Italy (IPCEI Cloud DM 27 giugno 2022 - IPCEI-CL-0000007)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17876" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.185, + 0.484, + 0.268 + ], + "angle": 0, + "content": "[2] Sikai Bai, Jie Zhang, Song Guo, Shuaicheng Li, Jingcai Guo, Jun Hou, Tao Han, and Xiaocheng Lu. Diprompt: Disentangled prompt tuning for multiple latent domain generalization in federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27284-27293, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.269, + 0.483, + 0.325 + ], + "angle": 0, + "content": "[3] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255. IEEE, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.326, + 0.483, + 0.38 + ], + "angle": 0, + "content": "[4] Lijie Fan, Dilip Krishnan, Phillip Isola, Dina Katabi, and Yonglong Tian. Improving clip training with language rewrites. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.382, + 0.482, + 0.45 + ], + "angle": 0, + "content": "[5] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2): 581-595, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.452, + 0.482, + 0.494 + ], + "angle": 0, + "content": "[6] Ishaan Gulrajani and David Lopez-Paz. In search of lost domain generalization. In International Conference on Learning Representations, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.495, + 0.482, + 0.564 + ], + "angle": 0, + "content": "[7] Tao Guo, Song Guo, Junxiao Wang, Xueyang Tang, and Wenchao Xu. Promptfl: Let federated participants cooperatively learn prompts instead of models-federated learning in age of foundation model. IEEE Transactions on Mobile Computing, 2023. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.482, + 0.634 + ], + "angle": 0, + "content": "[8] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.635, + 0.482, + 0.704 + ], + "angle": 0, + "content": "[9] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, page 3, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.705, + 0.483, + 0.787 + ], + "angle": 0, + "content": "[10] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.483, + 0.845 + ], + "angle": 0, + "content": "[11] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Peter Kairouz, H Brendan McMahan, Brendan Avent, Aurélien Bellet, Mehdi Bennis, Arjun Nitin Bhagoji, Kallista Bonawitz, Zachary Charles, Graham Cormode, Rachel Cummings, et al. Advances and open problems in federated learn" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "ing. Foundations and trends® in machine learning, 14(1-2): 1-210, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[13] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19113-19122, 2023. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.194, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[14] Hongxia Li, Wei Huang, Jingya Wang, and Ye Shi. Global and local prompts cooperation via optimal transport for federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12151-12161, 2024. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.32 + ], + "angle": 0, + "content": "[15] Wang Lu, HU Xixu, Jindong Wang, and Xing Xie. Fedclip: Fast generalization and personalization for clip in federated learning. In ICLR Workshop on Trustworthy and Reliable Large-Scale Machine Learning Models, 2023. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.39 + ], + "angle": 0, + "content": "[16] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.905, + 0.434 + ], + "angle": 0, + "content": "[17] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. In International Conference on Learning Representations, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.436, + 0.905, + 0.505 + ], + "angle": 0, + "content": "[18] Liliane Momeni, Mathilde Caron, Arsha Nagrani, Andrew Zisserman, and Cordelia Schmid. Verbs in action: Improving verb understanding in video-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15579-15591, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.507, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[19] Khalil Muhammad, Qinqin Wang, Diarmuid O'Reilly-Morgan, Elias Tragos, Barry Smyth, Neil Hurley, James Geraci, and Aonghus Lawlor. Fedfast: Going beyond average for faster training of federated recommender systems. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1234-1242, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.605, + 0.905, + 0.674 + ], + "angle": 0, + "content": "[20] Sarah Pratt, Ian Covert, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15691-15701, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.676, + 0.905, + 0.745 + ], + "angle": 0, + "content": "[21] Chen Qiu, Xingyu Li, Chaithanya Kumar Mummadi, Madan Ravi Ganesh, Zhenzhen Li, Lu Peng, and Wan-Yi Lin. Federated text-driven prompt generation for vision-language models. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.747, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[22] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[23] Jinhyun So, Chaoyang He, Chien-Sheng Yang, Songze Li, Qian Yu, Ramy E Ali, Basak Guler, and Salman Avestimehr. Lightsecagg: a lightweight and versatile design for secure aggregation in federated learning. Proceedings of Machine Learning and Systems, 4:694-720, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17877" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[24] Guangyu Sun, Matias Mendieta, Aritra Dutta, Xin Li, and Chen Chen. Towards multi-modal transformers in federated learning. In European Conference on Computer Vision, pages 229-246. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[25] Zhenyu Sun, Xiaochun Niu, and Ermin Wei. Understanding generalization of federated learning via stability: Heterogeneity matters. In International Conference on Artificial Intelligence and Statistics, pages 676-684. PMLR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.483, + 0.262 + ], + "angle": 0, + "content": "[26] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.347 + ], + "angle": 0, + "content": "[27] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, et al. Robust fine-tuning of zero-shot models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7959-7971, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[28] Yinghui Xing, Qirui Wu, De Cheng, Shizhou Zhang, Guoqiang Liang, Peng Wang, and Yanning Zhang. Dual modality prompt tuning for vision-language pre-trained model. IEEE Transactions on Multimedia, 26:2056-2068, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.483, + 0.488 + ], + "angle": 0, + "content": "[29] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language in a bottle: Language model guided concept bottlenecks for interpretable image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19187-19197, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.483, + 0.558 + ], + "angle": 0, + "content": "[30] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6757-6767, 2023. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.561, + 0.483, + 0.616 + ], + "angle": 0, + "content": "[31] Xin Yao, Chaofeng Huang, and Lifeng Sun. Two-stream federated learning: Reduce the communication costs. In 2018 IEEE Visual Communications and Image Processing (VCIP), pages 1-4. IEEE, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.483, + 0.672 + ], + "angle": 0, + "content": "[32] Xin Yao, Tianchi Huang, Chenglei Wu, Rui-Xiao Zhang, and Lifeng Sun. Federated learning with additional mechanisms on clients to reduce communication costs. arXiv preprint arXiv:1908.05891, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.483, + 0.741 + ], + "angle": 0, + "content": "[33] Liping Yi, Gang Wang, Xiaoguang Liu, Zhuan Shi, and Han Yu. Fedgh: Heterogeneous federated learning with generalized global header. In Proceedings of the 31st ACM International Conference on Multimedia, pages 8686-8696, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.483, + 0.801 + ], + "angle": 0, + "content": "[34] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11975-11986, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[35] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[36] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. In-" + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.547, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "ternational Journal of Computer Vision, 130(9):2337-2348, 2022. 1, 2, 3" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "17878" + } + ] +] \ No newline at end of file diff --git a/2025/FedMVP_ Federated Multimodal Visual Prompt Tuning for Vision-Language Models/9a8979ca-f59a-40f4-a042-065bd341c994_origin.pdf b/2025/FedMVP_ Federated Multimodal Visual Prompt Tuning for Vision-Language Models/9a8979ca-f59a-40f4-a042-065bd341c994_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..256e15a83f3de57145931a5496b3b0941061f667 --- /dev/null +++ b/2025/FedMVP_ Federated Multimodal Visual Prompt Tuning for Vision-Language Models/9a8979ca-f59a-40f4-a042-065bd341c994_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:37dd83eb5e81f0279b2557fbe3d4822d7245d13829e53c53550275a3ff43c4a7 +size 832495 diff --git a/2025/FedMVP_ Federated Multimodal Visual Prompt Tuning for Vision-Language Models/full.md b/2025/FedMVP_ Federated Multimodal Visual Prompt Tuning for Vision-Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..16e0458df9e313b853e751f9ae720eb48845ba99 --- /dev/null +++ b/2025/FedMVP_ Federated Multimodal Visual Prompt Tuning for Vision-Language Models/full.md @@ -0,0 +1,321 @@ +# FedMVP: Federated Multimodal Visual Prompt Tuning for Vision-Language Models + +Mainak Singha1 Subhankar Roy2 Sarthak Mehrotra3 Ankit Jha4 Moloud Abdar5 Biplab Banerjee3 Elisa Ricci1,6 + +1 University of Trento, Italy + +4 LNMIIT Jaipur, India + +$^{2}$ University of Bergamo, Italy + +5 The University of Queensland, Australia + +3 Indian Institute of Technology Bombay, India + +Fondazione Bruno Kessler, Italy + +# Abstract + +In federated learning, textual prompt tuning adapts Vision-Language Models (e.g., CLIP) by tuning lightweight input tokens (or prompts) on local client data, while keeping network weights frozen. After training, only the prompts are shared by the clients with the central server for aggregation. However, textual prompt tuning suffers from overfitting to known concepts, limiting its generalizability to unseen concepts. To address this limitation, we propose Multimodal Visual Prompt Tuning (FedMVP) that conditions the prompts on multimodal contextual information - derived from the input image and textual attribute features of a class. At the core of FedMVP is a PromptFormer module that synergistically aligns textual and visual features through a cross-attention mechanism. The dynamically generated multimodal visual prompts are then input to the frozen vision encoder of CLIP, and trained with a combination of CLIP similarity loss and a consistency loss. Extensive evaluation on 20 datasets, spanning three generalization settings, demonstrates that FedMVP not only preserves performance on in-distribution classes and domains, but also displays higher generalizability to unseen classes and domains, surpassing state-of-the-art methods by a notable margin of $+1.57\% - 2.26\%$ . Code is available at https://github.com/mainaksingha01/FedMVP. + +# 1. Introduction + +A successful learning paradigm involves training neural networks on large and centralized datasets, enabling models to learn robust and generalizable representations [22, 34]. However, practical constraints often prevent centralized training (e.g., privacy regulations) creating a critical need for decentralized training frameworks that achieve high performance without central data access. As a solution, federated learning (FL) [16] enables multiple clients to collaboratively train a global model without requiring direct data sharing. In FL, clients train models locally and trans + +![](images/cad9b44ec8e2dd22c7f3c0f33c25ad62cf44a95a38038d12e2976470c5e4af54.jpg) +Figure 1. Comparison of federated prompt tuning methods: (a) a textual prompt tuning method (e.g., FedTPG [21]) encodes unimodal contextual information, i.e., class names, in the learnable prompts, (b) our proposed method FedMVP injects multimodal contextual information – image and attributes – into the prompts. + +mit only model updates (e.g., gradients or parameters) to a central server, which aggregates these updates to improve the global model. The research in FL is primarily aimed at: reducing communication rounds between clients and server [31, 32], faster and lightweight training algorithms on clients [19, 23] and improving generalization while exhibiting high data heterogeneity across clients [25, 33]. + +Pretrained Vision-Language Models (VLMs) (e.g., CLIP [22]), have emerged as promising candidates for FL due to their strong generalization capabilities. However, the high communication overhead and computational demands of parameter-heavy VLMs pose challenges for FL, especially in bandwidth- or compute-constrained environments (e.g., robots). To make CLIP viable for the FL setting, prompt tuning (e.g., CoOp [36], VPT [11]) has been adopted as a preferred technique that appends lightweight learnable tokens (or prompts) to the input tokens of the CLIP image and/or text encoder. Each client trains the prompts on the local data, while keeping the CLIP backbone frozen, and then dispatches these prompts to the server, where they are updated through aggregation and sent back. This speeds up client training as each client needs access + +to only a few samples per class (e.g., 16/class [36]), and reduces communication overhead as only the lightweight prompts (e.g., $0.37\%$ of network parameters) need to be communicated, instead of all the parameters. + +Although prompt tuning has proven to be effective in the FL setting, it suffers from reduced generalization to unseen classes and unseen domains - a typical heterogeneous setup native to non independent and identically distributed (non-i.i.d.) FL [21] - due to severe overfitting on the local training data [30]. It is attributed to the fact that prompts learn a static context from a set of seen classes, which being fixed once learned, fail to capture the generalizable elements spread across clients. To circumvent the issue of reduced generalization, recent CLIP-based FL methods condition the prompts either on the class information, as in FedTPG [21] (see Fig. 1a), or on visual information, as in FedCoCoOp [35]. This auxiliary conditioning on the input, or contextual information, makes the prompts dynamic and thus offers a better generalization to unseen distributions. + +In this work, we argue that the FL setup being highly heterogeneous, where clients have data from disjoint classes and domains [21], it is not sufficient to solely exploit either of the two modalities for contextual information. Rather, FL demands more comprehensive contextual information that encompasses all the available modalities to condition input prompts. To this end, we propose Federated Multimodal Visual Prompt Tuning (FedMVP) that conditions the prompts on multimodal contextual information, which comprise: (i) visual features of the input image itself, and (ii) textual features of attributes of a class (e.g., "two legs", "beak", etc. for the class "hen"). The key intuition is that (ii) promotes learning transferable prompts for unseen classes that share attributes similar to seen classes (e.g., unseen class "seagull" will share some attributes with the seen class "hen"), and (i) promotes transferability for unseen classes and domains where attributes cannot be described via text (e.g., texture or abstract concepts). In a nutshell, the dual conditioning offers richer contextual information to the prompts, which helps in generalizing to classes or domains that share similar properties. + +In detail, the proposed FedMVP framework contains a PromptFormer network (see Fig. 1b) that synergistically merges the information from visual and textual modalities with cross-attention mechanism [26] to generate the prompts. The resulting multimodal prompts are then injected at the visual space of the frozen vision encoder of CLIP, unlike the existing methods [7, 21, 30] that inject prompts through text encoder. The key differences between our FedMVP and existing CLIP-based FL methods have been summarized in Tab. 1. The PromptFormer is trained in an end-to-end manner with the CLIP loss and a consistency loss. Once trained, each participating client shares the lightweight PromptFormer network parameters with the + +Table 1. Summary of federated prompt-tuning methods. $(X)$ indicates the learnable prompts are initialized randomly, $(\checkmark)$ signifies learnable prompts are conditioned on contextual information (e.g., textual or visual inputs). Prompting column denotes the CLIP encoder wherein learnable prompts are given as input (e.g., "Textual" means the text encoder of CLIP takes prompts as input). + +
MethodContextual informationPrompting
TextualVisual
FedKgCoOp [30]XXTextual
FedVPT [11]XXVisual
FedTPG [21]✓XTextual
FedCoCoOp [35]X✓Textual
FedMaPLe [13]XXMultimodal
FedMVP (Ours)✓✓Visual
+ +server for aggregation through FedAvg [16], and thus enabling knowledge sharing among clients. During inference, the cross-attention learned by the PromptFormer enables the model to dynamically adjust the prompts, enhancing generalizability to unseen classes and domains. We evaluated FedMVP on three generalization settings encompassing 20 datasets that measure generalization to unseen classes, domains, and combinations of them. Experiments demonstrate that FedMVP consistently outperforms the state-of-the-art CLIP-based FL methods by $1.57\% - 2.26\%$ . + +In summary, our main contributions are: (i) We highlight the importance of leveraging multimodal contextual information in prompt tuning based FL setup in order to alleviate the issue of reduced generalizability to unseen data. (ii) We propose FedMVP that synergistically combines the visual features of the image and textual features of the attributes of classes to generate multimodal prompts. This provides richer contextual information to the model and higher transferability to unseen classes and domains. + +# 2. Related Work + +Prompt tuning of VLMs. VLMs, such as CLIP [22] and ALIGN [10], leverage large image-text datasets to enable multimodal understanding, performing well in zero-shot classification tasks. However, fully fine-tuning VLMs for downstream tasks without sacrificing their ability to generalize to unseen concepts remains a challenge [27]. + +Prompt tuning approaches [11, 13, 36] have been proposed for parameter efficient and effective adaptation of CLIP by introducing a set of continuous learnable tokens (or prompts) onto the frozen CLIP backbone. These approaches can be broadly categorized into three categories depending on the manner in which CLIP is prompted: (i) textual prompting that inject prompts in the input space of the text encoder [30, 36], (ii) visual prompting that injects prompts in the input space of the vision encoder [11], and (iii) multimodal prompting that injects prompts in both the text and vision encoder [13, 28]. + +In particular, MaPLe [13], which injects prompts to both + +the text and vision encoders, is not suitable for memory and compute efficient FL since it requires parts of both the encoders to be unlocked during tuning, risking overfitting. Moreover, it does not allow for image-conditioned dynamic prompting, which has shown to be effective for generalization [35]. In contrast, our proposed FedMVP enables multimodal prompting by involving only the vision encoder and supports instance-specific prompting. + +Federated prompt tuning of VLMs. Prompt tuning has emerged as an effective technique for FL of VLMs since it allows clients to collaboratively train on a heterogeneously distributed dataset and adapt faster with fewer samples. The fact that the prompts are lightweight lends well to the FL setup as it reduces communication overhead. Due to the benefits offered by prompt tuning, several works [2, 7, 21, 24] have specifically adapted prompt tuning approaches to the FL setup. For example, PromptFL [7] extended CoOp to FL via sharing prompts with the server, enhancing the generalization performance of client-server interactions in vision tasks. FedCLIP [15] learns a lightweight adapter in each client, similar to the CLIP-Adapter [5], and shares parameters with the server. FedOTP [14] introduced global and local prompts, optimizing shared parameters through optimal transport, while DiPrompt [2] studied multi-domain learning by allowing multiple domains per client. + +Unlike existing methods, a couple of recent works [21, 28] argue that prompt tuning is prone to overfitting to the seen classes due to learning static context. The closest work to ours is FedTPG [21], where a prompt generation network is learned across multiple clients. Additionally, this network leverages textual inputs, enhancing its ability to generalize to unseen classes. Our work significantly differs from previous methods in two key ways: (i) we exploit both attribute information derived from an LLM and image features to condition the prompts, and (ii) we leverage multimodal prompts from the proposed PromptFormer for visual prompt tuning, instead of textual prompt tuning. + +Classification via textual descriptions. Zero-shot accuracy of CLIP on downstream tasks is sensitive to the quality of text prompts [22]. In detail, some works rely on simple hand-crafted prompts (e.g., "a photo of a [CLS]") [22] or others augment the hand-crafted prompts with richer attributes obtained from LLMs [17, 20] (e.g., a "hen" has two legs, white feathers, etc.). There also exists a family of methods, such as LaCLIP [4], LaBo [29] and VFC [18], that have proposed to further refine CLIP weights with enriched captions from LLMs for improved performance. Recognizing the importance of class attributes for improved generalization to unseen concepts and domains – an important desideratum in the FL set up – we, for the first time in FL, introduce LLM generated class attributes for enriching the image-conditioned multimodal prompts through the PromptFormer network. + +# 3. Methods + +We introduce FL setup in Sec. 3.1, preliminaries on prompt tuning in Sec. 3.2 and our proposed solution in Sec. 3.3. + +# 3.1. Problem formulation + +We consider a FL setup in which multiple remote clients train models on local data, and a central server aggregates models from the clients. Formally, let each remote client $i$ has access to a dataset $\mathcal{D}_i = \{\mathbf{x}_{i,j},y_{i,j}\}_{j = 1}^{n_i}$ , where $\mathbf{x}_{i,j}$ is the $j^{\mathrm{th}}$ image from the $i^{\mathrm{th}}$ client, $y_{i,j}$ is the corresponding class label and $n_i$ is the number of samples in client $i$ . Let the set of text class names in client $i$ be denoted as $\mathcal{C}_i = \{c_{i,k}\}_{k = 1}^{K_i}$ , where $K_{i}$ is the total number of classes. As in FedTPG [21], we follow a non-IID FL setup, where the samples in each client are drawn from a disjoint set of classes, i.e., $\mathcal{C}_i\cap \mathcal{C}_l = \emptyset$ . As an example, the training data on each mobile device usually depend on the user, which may not be representative of the general population [16]. + +In particular, we investigate two kinds of non-IID FL setups, where, in addition to the in-distribution performance, we also evaluate the generalization ability of the aggregated server model on (i) unseen classes, and (ii) unseen domains or datasets. The need for generalization to unseen classes and domains makes the non-IID FL setup more challenging than the traditional FL setting [12]. + +# 3.2. Preliminaries on Prompt Tuning + +Textual prompt tuning (TPT), i.e., CoOp [36], has emerged as an efficient technique in CLIP-based FL that keeps the CLIP backbone frozen and locally learns lightweight prompt vectors on each client. The locally learned prompts are then communicated to the server, as in PromptFL [7]. Specifically, the hand-crafted text prompts (e.g., "a photo of a [CLS]") input to the text encoder $\mathcal{E}_t$ is augmented with the learnable soft prompts $v_{1}, v_{2}, \ldots, v_{m}$ . However, as shown in [35], TPT can often lead to overfitting and reduced generalization due to the static context learned on the seen classes. This can be exacerbated in the non-IID FL setup due to the high data heterogeneity among the clients [21]. + +Alternatively, visual prompt tuning (VPT) [11] adapts pre-trained vision encoder $\mathcal{E}_v$ by injecting small number of visual learnable parameters into ViT's input space. More formally, let an input image $\mathbf{x}$ be divided in $b$ patches, yielding a collection of patch embeddings (augmented with positional encodings) at the input of the vision encoder as $\mathbf{E} \in \mathbb{R}^{b \times d_v}$ , where $d_v$ is the embedding dimension. A [CLS] token $\mathbf{z}$ is concatenated with $\mathbf{E}$ , encapsulating the initial input representation for the ViT layers of $\mathcal{E}_v$ . VPT further augments the input by concatenating a set of randomly initialized learnable visual prompts $\mathbf{P} = \{p_1, p_2, \ldots, p_m\}$ , where $m$ is the prompt length. The first layer input of $\mathcal{E}_v$ is then redefined as $\mathbf{I} = [\mathbf{z}; \mathbf{E}; \mathbf{P}] \in \mathbb{R}^{(1 + b + m) \times d_v}$ , where + +![](images/8d12d753ddcaf747b6524df2d4c26ff3aae9fb9a19b8357fefababb4f875ae81.jpg) +Figure 2. Overview of Federated Multimodal Visual Prompt Tuning (FedMVP). Each client $i$ trains the parameters of the PromptFormer network $f_{\Theta}$ and multi-head self-attention on its local data. At each round, the server aggregates the trainable parameters from the clients and sends the updated parameters back to the clients. The PromptFormer uses cross-attention mechanism to fuse the visual and textual attribute feature to generate multimodal visual prompts $\mathbf{P}$ , which are then used for visual prompt tuning. + +![](images/7c528344f045fb89837191fb059d5e5051c445481371681448063518fdd92c12.jpg) + +![](images/2ece183bd23848b010a1a9ce2ceff94d4ceb58813e90864c10c8b8702d125c59.jpg) + +“;” means concatenation. Although VPT enables more nuanced understanding of the visual space, it suffers from the same issues as TPT, i.e., limited generalization (see Sec. 4). + +# 3.3. Federated Multimodal Visual Prompt Tuning + +Our proposed method FedMVP deviates from the TPT and VPT paradigms and incorporates multimodal contextual information from both the visual and textual modalities to improve generalization in the FL setup. As shown in Fig. 2, at the core of our novel architecture is a prompt generator, called PromptFormer, which conditions the visual prompts $\mathbf{P}$ on the task-related information: input image and text attributes corresponding to the class names, rather than randomly initializing them. This makes $\mathbf{P}$ context-aware, generalizing better to unseen classes and domains. Once trained on each client, only the parameters of the lightweight PromptFormer network are communicated to the server, thus, keeping the communication overhead low. Next, we describe PromptFormer (Sec. 3.3.1), training objectives (Sec. 3.3.2) and how we conduct client-side training and server-side aggregation (Sec. 3.3.3). + +# 3.3.1. Visual Prompt Generation with PromptFormer + +We introduce the PromptFormer network $f_{\Theta}$ , parameterized by $\Theta$ , designed to generate contextually rich visual tokens (or prompts) for each client $i$ by leveraging both attribute features and visual patch embeddings (see Fig. 2). Concretely, we use the CLIP text encoder $\mathcal{E}_t$ to extract embeddings $\mathbf{A}_i = \{\mathcal{E}_t(\mathrm{LLM}(c_k))\}_{k=1}^K$ of the text attributes $\mathcal{A}_i$ corresponding to the text class names $\mathcal{C}_i$ , generated using a large lanuage model (LLM) [1] (see Sec. A.1 of Supp. Mat. + +for Attribute Generation process). Going forward, we drop the client index $i$ for notational clarity. The PromptFormer takes as input $b$ visual patch embeddings $\mathbf{E}$ , corresponding to an image $\pmb{x}$ , and the text embeddings $\mathbf{A}$ to generate a set of enriched $m$ multimodal visual prompts $\mathbf{P} \in \mathbb{R}^{m \times d_v}$ : + +$$ +\mathbf {P} = \left\{p _ {j} \right\} _ {j = 1} ^ {m} = f _ {\Theta} (\mathbf {A}, \mathbf {E}). \tag {1} +$$ + +Consequently, the input to the first layer of the visual encoder $\mathcal{E}_v$ is redefined as $[\mathbf{z};\mathbf{E};\mathbf{P}]\in \mathbb{R}^{(1 + b + m)\times d_v}$ . Intuitively, multimodal visual prompts carry richer contextual information, and helps the network in generalizing to unseen classes and domains that share commonalities. + +Architecture of PromptFormer. The PromptFormer architecture comprises: lightweight multi-head cross-attention (MHCA) modules, feed-forward network (FFN) and a linear projection layer; transforming multimodal features into comprehensive visual prompts. In detail, the visual patch embedding $\mathbf{E}$ interacts with the linearly projected attribute features $\mathbf{A}' = T_{\mathrm{proj}}(\mathbf{A})$ to construct multimodal visual prompts as: + +$$ +\mathbf {P} \left(\mathbf {A} ^ {\prime}, \mathbf {E}\right) = \operatorname {F F N} \left(\text {C r o s s A t t e n t i o n} \left(\mathbf {Q} _ {\mathbf {E}}, \mathbf {K} _ {\mathbf {A} ^ {\prime}}, \mathbf {V} _ {\mathbf {A} ^ {\prime}}\right)\right); +$$ + +$$ +\text {w i t h} \mathbf {Q} _ {\mathbf {E}} = \mathbf {E} W _ {\mathbf {Q}}, \mathbf {K} _ {\mathbf {A} ^ {\prime}} = \mathbf {A} ^ {\prime} W _ {\mathbf {K}}, \mathbf {V} _ {\mathbf {A} ^ {\prime}} = \mathbf {A} ^ {\prime} W _ {\mathbf {V}}, \tag {2} +$$ + +where, $\mathbf{E}$ are transformed into query vectors $\mathbf{Q}_{\mathbf{E}}$ using query matrices $W_{\mathbf{Q}}$ ; and the projected attribute features $\mathbf{A}'$ are transformed into key and value vectors $\mathbf{K}_{\mathbf{A}'}$ and $\mathbf{V}_{\mathbf{A}'}$ , using key and value matrices $W_{\mathbf{K}}$ and $W_{\mathbf{V}}$ , respectively. + +Intuitively, through the cross-attention mechanism, visual features learn to attend to the relevant parts of the corresponding attribute features. For example, the patches depicting the legs of the class "dog" will learn to attend to + +the attributes four legged. When an unseen class (e.g., "giraffe") comprises an animal with four legs, the final prompts $\mathbf{P}$ will contain this relevant information derived from the attributes four legged. Additional details of PromptFormer architecture are reported in Sec. A.3 of the Supp. Mat. + +# 3.3.2. Training of FedMVP + +After the visual prompts are obtained with PromptFormer as per Eq. 2, we concatenate the [CLS] token, the patch embeddings, and the visual prompts to get $\mathbf{I} = [\mathbf{z};\mathbf{E};\mathbf{P}]$ following Sec. 3.2, and feed it to the visual encoder $\mathcal{E}_v$ . + +The parameters of PromptFormer $\rho = \{\Theta\}$ are the only trainable parameters of FedMVP, and the rest of the weights are kept frozen. To train the model, we employ the CLIP cross-entropy loss [22], formulated as: + +$$ +\mathcal {L} _ {\mathrm {c e}} = - \underset {(\mathbf {x}, y) \sim \mathcal {D}} {\mathbb {E}} y \log p (y | \mathbf {I}), \tag {3} +$$ + +where $y$ denotes the ground truth label, and the prediction probability of CLIP between the visual features $\mathbf{v} = \mathcal{E}_v(\mathbf{I})$ and text features $\mathbf{t}_k = \mathcal{E}_t([A\text{ photo of [CLASS]};\mathrm{LLM}(c_k)])$ for class $k$ is computed as: + +$$ +p (y = k | \mathbf {I}) = \frac {\exp \left(\cos \left(\mathbf {v} , \mathbf {t} _ {k}\right)\right) / \tau)}{\sum_ {k ^ {\prime} = 1} ^ {K} \exp \left(\cos \left(\mathbf {v} , \mathbf {t} _ {k ^ {\prime}}\right)\right) / \tau}, \tag {4} +$$ + +where $\cos (\cdot ,\cdot)$ as a similarity function, $\tau$ as the temperature parameter, and $[\cdot ]$ denotes concatenation. Details on how $\mathbf{t}_k$ is computed are provided in Sec. A.1 of the Supp. Mat. + +In addition to the $\mathcal{L}_{\mathrm{ce}}$ , we employ a regularization constraint to ensure consistency between two augmented views of an image. We define the consistency loss $\mathcal{L}_{\mathrm{con}}$ as: + +$$ +\mathcal {L} _ {\mathrm {c o n}} = 1 - \cos \left(\mathcal {E} _ {v} (\mathbf {I}), \mathcal {E} _ {v} \left(\mathbf {x} ^ {\prime}\right)\right), \tag {5} +$$ + +where $\mathbf{x}'$ is the augmented version of $\mathbf{x}$ . Finally, the total loss function is given as: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\mathrm {c e}} + \alpha \cdot \mathcal {L} _ {\mathrm {c o n}}, \tag {6} +$$ + +where $\alpha$ is a balancing coefficient for the regularization term, ensuring both alignment and consistency. + +# 3.3.3. Client-side training and server-side aggregation + +In our approach, multiple remote clients participate in distinct image classification tasks, collaboratively training the PromptFormer network, parameterized by $\rho = \{\Theta\}$ . In each communication round, the client transmits $\rho$ to the server, which is aggregated at the server. Through training on a diverse set of classification tasks distributed across clients, and aggregating the parameters at the server, the resulting model achieves generalization capability to classes and domains unseen during training. Pseudo-code of FedMVP can be found in Sec. A.2 of the Supp. Mat. + +Initially, the server randomly initializes the parameters $\rho^0$ and in each communication round only a random subset + +of remote clients $S^r$ will be chosen by the server for sharing the updated parameters $\rho$ . Each communication round $r$ involves the following steps: + +(i) Model initialization: Each client $i$ in $S_{r}$ receives the latest model parameters $\rho^r$ to configure its local model. +(ii) Local optimization: Following Eq. (6), client $i$ optimizes its local parameters $\rho^r$ using an SGD optimizer over $J$ iterations with learning rate $\eta^r$ . Updated parameters: + +$$ +\rho_ {i} ^ {r + 1} = \operatorname {S G D} _ {J} \left(\eta^ {r}, \rho^ {r}, \mathcal {C} _ {i}, \mathcal {D} _ {i}\right). \tag {7} +$$ + +(iii) Parameter aggregation: After local training, each client $i$ in $S_{r}$ transmits its updated parameters $\rho_{i}^{r + 1}$ back to the server, where they are aggregated [16] as: + +$$ +\rho^ {r + 1} = \frac {1}{\left| S _ {r} \right|} \sum_ {i \in S _ {r}} \rho_ {i} ^ {r + 1}. \tag {8} +$$ + +The above steps are repeated for $R$ communication rounds, after which the server converges on the final model parameters $\rho^R$ . This iterative process facilitates collaborative and distributed optimization of $\rho$ across client data. + +Lightweight fine-tuning in clients. As the FL setup exhibits high data heterogeneity, different clients possess varying numbers of samples. Thus, treating every client as equal and repeatedly fine-tuning the parameters $\rho$ can lead to severe overfitting in some clients, especially those with very limited training data. To prevent this, we propose to inject low-rank decomposition matrices, or LoRA [9], to the PromptFormer module, and fine-tune these LoRA matrices if training loss $\mathcal{L}_{\mathrm{total}}$ in a client starts below a certain loss value $\sigma$ , which we set to 0.5. In particular, only the LoRA weights corresponding to the query, key and value matrices of the attention layers are trained, while freezing the parameters $\rho$ . In that case, the client shares only the LoRA matrices with the server. This not only helps in boosting generalization, but also brings down the communication overhead of clients by around $\times 267$ . + +# 4. Experiments + +We evaluate our proposed FedMVP in three distinct settings that are aimed to gauge the (i) generalization to unseen categories (or Base-to-New Generalization) in Sec. 4.1, (ii) generalization to unseen domains (or Domain Generalization) in Sec. 4.2, and (iii) generalization to unseen datasets (or Cross-Dataset Generalization) in Sec. 4.3. Finally, we report the results of our ablation study in Sec. 4.4 that are aimed to assess impact of the design choices made in FedMVP. Due to the lack of space, we report additional experimental results in Sec. B of the Supp. Mat. + +Benchmarks and datasets. Following FedTPG [21], we have reported results on the Base-to-New Generalization setting that encompasses nine medium-sized image recognition datasets. For the Domain Generalization setting we + +have used the ImageNet [3] and DomainBed [6] benchmarks. In addition, we have also included EuroSAT [8] for the Cross-Dataset Generalization setting, alongside all nine datasets of the Base-to-New Generalization setting. In total, we have used 20 datasets for a comprehensive experimental evaluation. Detailed dataset descriptions and statistics can be found in Sec. B.1 of the Supp. Mat. + +Baselines. We have compared our method with the state-of-the-art prompt tuning approaches that use the CLIP backbone [22]: KgCoOp [30], CoCoOp [35], VPT [11], and MaPLe [13], which were originally designed for offline and non-FL setup, but adapted to the federated setting, as in FedTPG [21]. Furthermore, we have compared with CLIP-based methods specifically designed to work in the FL setting: FedTPG [21], FedOTP [14], PromptFL [7], and FedCLIP [15]. We also included zero-shot CLIP (ZS-CLIP). + +Implementation details. We have employed the frozen ViT-B/16 CLIP backbone as the vision encoder, and GPT-4o [1] as the LLM in our experiments. Each MHCA module of $f_{\Theta}$ , consists of a 4-head cross-attention layers, complemented by layer normalization and each FFN comprising of a two-layer bottleneck structure (Linear-GeLU-Linear). The query prompt $\mathbf{Q}$ has a length $m = 4$ and matches the patch embedding dimension $d_v = 768$ . The projection layer $T_{\mathrm{proj}}$ maps the textual dimension of 512 to the patch embedding dimension via a linear transformation. In Eq. (6), we keep $\alpha = 10$ . We have set the batch size to 128, number of shots per class to 8, and the learning rate equal to 0.003 with a decay rate of 1e-5 and momentum of 0.9. + +# 4.1. Base-to-New Generalization + +Experimental setup. To recap, the base-to-new generalization setting encompasses nine different datasets. The setting is characterized by splitting the classes from each dataset into two groups – base and new. The samples of the base classes are used for training, and the samples of the new classes are held out to test the generalization performance [21]. Furthermore, the samples of bases classes from all nine datasets are distributed to the clients, such that each client does not see samples from more than 20 disjoint classes (referred to as local classes). We report the model's classification accuracy on: (i) the client's local classes, (ii) all the base classes (spread across clients), and (iii) the held-out new classes that were never seen by any client. Note that (ii) and (iii) is computed using the aggregated model on the server, and (i) is computed using the model on the client (which is the same server model after convergence). + +Main results. In Table 2 we report the performance in the base-to-new generalization setting, summarized across nine datasets. The results show that the textual and multimodal prompting methods, such as PromptFL, FedKgCoOp, and FedMaPLe, excel on the seen local and base class accuracy, but have at-par or worse generalization on the new + +Table 2. Base-to-New Generalization setting. Accuracies $(\%)$ on client's local (seen) classes, base (seen) classes and new (unseen) classes. Harmonic mean (HM) is computed with Base and New class accuracies. Best number in each column is in bold and second-best underlined. Improvement row reports the absolute difference between the best and second-best numbers. + +
MethodsLocalBaseNewHM
ZS-CLIP [22]76.7270.5175.7874.24
FedOTP [14]74.8265.2257.0464.89
FedCoCoOp [35]81.4673.7666.0073.20
FedVPT [11]76.2970.4374.8973.79
FedCLIP [15]76.8771.0475.0674.24
FedMaPLe [13]81.6374.4470.6275.29
FedKgCoOp [30]78.3872.1875.8775.39
PromptFL [7]81.7574.4771.7075.74
FedTPG [21]80.7573.6876.0276.70
FedMVP (Ours)81.8975.3777.8278.27
Improvement+0.14+0.90+1.79+1.57
+ +classes than the ZS-CLIP model. This hints at the potential overfitting caused by learning static context from the seen classes. In contrast, our proposed FedMVP that considers multimodal contextual information and dynamic conditioning demonstrates superior performance in both the seen and unseen classes, achieving the highest HM accuracy. While FedTPG also generates prompts based on task-related input, it does not outperform our FedMVP due to lack of comprehensive visual and attribute information in the prompts. In summary, FedMVP surpasses the state-of-the-art methods by a notable margin of $+1.57\%$ in the base-to-new generalization setting. Detailed breakdown for each of the nine datasets is reported in Tab. A.4 in the Supp. Mat. + +# 4.2. Domain Generalization + +Experimental setup. We conduct experiments under two distinct domain generalization (DG) setups: Multi-source Single-target (MSST) and Single-source Multi-target (SSMT). In the MSST DG setting for a given benchmark, we follow a leave-one-domain-out protocol for evaluation, where we hold out one domain for measuring generalization and use the rest of the domains for training the clients. Training classes are split among clients under the constraint that each client can see images from a single domain. This setting mirrors a practical scenario where images taken with a smartphone uniquely reflect the taste of the user or socio-economic biases. In a similar manner, in the SSMT DG setting, we use one domain for training the clients in a distributed manner, while the rest of the domains are used for evaluation. More details on this experimental setup can be found in Sec. B.4 of the Supp. Mat. + +Main results. In Table 3, we report the performance of FedMVP and its competitors in the MSST and SSMT DG settings. In particular, SSMT DG is a more challenging setting than the MSST DG setting, as the clients are trained only on a single domain, and expected to generalize to multiple unseen domains. We observe that ZS-CLIP is a strong base + +Table 3. Domain Generalization setting on the DomainBed [6] benchmark. Accuracies $(\%)$ on the unseen target domain(s) for the Multi-source Single-target (MSST) and Single-source Multi-target (SSMT) settings. Best number in each column is in bold and second-best underlined. Improvement row reports the absolute difference between the best and second-best numbers. + +
MethodMulti-source Single-targetSingle-source Multi-target
PACSOfficeHomeVLCSTerra Inc.DomainNetAveragePACSOfficeHomeVLCSTerra Inc.DomainNetAverage
ZS-CLIP [22]96.1681.4983.2933.9857.1370.4196.1681.4983.2933.9857.1370.41
FedOTP [14]90.7176.4267.4113.2449.6759.4991.1774.5365.1615.1139.3757.07
FedCoCoOp [35]85.0681.4261.7323.6857.0861.7984.5650.8761.2523.0758.2855.61
FedTPG [21]90.9982.7869.7726.7956.8265.4390.7182.6067.6322.5158.3464.36
PromptFL [7]95.3781.7474.8725.0256.8766.7795.8381.7266.9724.1758.2765.39
FedKgCoOp [30]95.4281.8274.9025.0356.9266.8295.8181.7067.4924.2058.3065.50
FedMaPLe [13]94.5182.0371.7936.3058.8868.7094.2581.4872.0833.5957.6267.80
FedVPT [11]95.3681.7683.1933.6255.9869.9894.7980.9982.4432.2355.6669.22
FedCLIP [15]96.2981.7482.7036.5857.8571.0396.2981.6281.8136.4457.4270.72
FedMVP (Ours)97.2884.1585.1237.3661.1773.0297.1783.8984.6136.9260.5672.63
Improvement+0.99+1.37+1.83+0.78+2.29+1.99+0.88+1.29+1.32+0.48+2.22+1.91
+ +Table 4. Domain Generalization setting on the ImageNet benchmark. Accuracies (\%) on the seen source domain ImageNet (IN) and unseen target domains: ImageNet-V2 (INV2), ImageNet-Sketch (IN-S), ImageNet-A (IN-A) and ImageNet-R (IN-R). Best number in each column is in bold and second-best underlined. Improvement row reports the absolute difference between the best and second-best numbers. + +
MethodSourceTarget
ININV2IN-SIN-AIN-RAverage
ZS-CLIP [22]66.7560.7946.1247.7974.0057.18
FedOTP [14]51.6845.5334.7316.6463.9840.22
FedMaPLe [13]66.9660.6544.6946.2474.6256.55
FedVPT [11]66.9260.3446.4348.0374.5657.34
PromptFL [7]67.8061.5945.6148.7874.4957.62
FedCLIP [15]67.2661.3546.6648.2474.3657.65
FedKCoOp [30]67.5361.6046.6948.3774.7157.84
FedCoCoOp [35]68.5162.2946.9050.3376.4959.00
FedTPG [21]69.5162.9047.6549.9776.3559.22
FedMVP (Ours)70.8763.7250.9351.7677.2360.91
Improvement+1.36+0.82+3.28+1.43+0.74+1.69
+ +line that outperforms all textual and multimodal prompting methods, with the exception of FedCLIP and our proposed FedMVP. This suggests that suboptimal prompt tuning can lead to overfitting on the source training domains, thus hurting generalization. Our FedMVP owing to the use of contextual information learns more transferable representations that generalize better to unseen target domains, thus outperforming all competitor methods in most of the benchmarks. More detailed results in Sec. B.4 of the Supp. Mat. + +In Table. 4 we report the results in the SSMT DG setting on ImageNet benchmark. We observe that FedTPG and FedCoCoOp come close to our FedMVP in terms of average performance, with FedMVP outperforming all baselines by $+1.69\%$ . In particular, we believe FedMVP improves the performance on ImageNet-Sketch (IN-S) by a large margin of $+3.28\%$ due to the use of attribute information, which remains unchanged between real and sketch images. + +# 4.3. Cross-dataset Generalization + +Experimental setup. In the Cross-Dataset Generalization setting the classes from ImageNet are split among the clients for training, in a disjoint manner, and the resulting network is evaluated on 10 held out datasets not seen during training. In other words, it is a combination of base-to-new + +generalization and DG settings, since it requires generalization to both unseen class and data distributions. + +Main results. In Table. 5 we report the results of the Cross-Dataset Generalization setting. This setting is considerably challenging, as the model should generalize to different datasets. We find that the results are consistent with the previous findings, where our proposed FedMVP generalizes better to unseen classes and data distributions, outperforming all the text-based and multimodal-based prompting techniques by $+2.26\%$ on average. However, with FedMVP we notice a drop in performance with respect to the ZS-CLIP baseline for some fine-grained datasets, such as OxfordPets and StanfordCars, which is potentially due to overlapping attributes for visually similar categories. + +# 4.4. Ablation Study + +Impact of proposed components in FedMVP. In Tab. 6 we ablate each component of FedMVP on the base-to-new and MSST DG settings. We observe that each component contributes positively to the final performance. First, Prompt-Former $f_{\Theta}$ plays a major role in combining textual information with visual features, thus improving semantic alignment between the two modalities. Second, adding $\mathcal{L}_{\mathrm{con}}$ (as in Eq. (5)) improves the performance by a small margin. Finally, the impact of LoRA can be observed in the penultimate row, where we fine-tune all the parameters of FedMVP. It shows the dangers of overfitting in the FL setup, as some clients may have limited data. + +Participation Rate of the Clients. In Fig. 3 we vary the participation rate from $10\%$ to $100\%$ and show how the participation rate of the clients impacts performance on both seen and unseen classes and domains. We observe that FedMVP always exhibits a gentle, yet positive slope in both settings, indicating that continuous parameter aggregation improves generalization. In contrast, for some competitor methods, such as PromptFL, increased participation leads to saturated performance or negative slope in the plot. + +Computational analysis. In Fig. 4 we plot the accuracy vs. communication rounds for FedMVP and competitors in base-to-new and MSST settings. While FedMVP trans + +Table 5. Cross-Dataset Generalization setting. Accuracies $(\%)$ on the seen classes of the source dataset, i.e., ImageNet, and the unseen classes of the target datasets. Best number in each column is in bold and second-best underlined. Improvement row reports the absolute difference between the best and second-best numbers. + +
MethodSourceTarget
ImageNetCaltech101Flowers102FGVCAircraftUCF101OxfordPetsFood101DTDStanfordCarsSUN397EuroSATAverage
ZS-CLIP [22]66.7592.9071.2924.7266.7589.1586.0944.3365.2962.5947.6865.08
FedOTP [14]51.6888.6048.3610.9546.1073.4361.0141.5529.6148.3450.2649.82
FedMaPLe [13]66.9692.4968.2523.5260.3289.6783.5244.6860.1661.8545.3862.98
PromptFL [7]67.8091.8768.1321.4464.1388.7085.8542.4363.5962.7743.2663.22
FedCLIP [15]67.2693.3967.1924.0365.6488.2885.1444.4465.5063.4241.6763.87
FedKgCoOp [30]67.5393.6369.3123.0664.4688.5585.3744.7464.9963.8543.2964.13
FedVPT [11]68.5691.4269.2421.0965.7087.7985.4544.5765.0064.6847.8264.28
FedCoCoOp [35]68.5194.1166.3420.7962.7589.0485.4043.2063.9864.0255.4064.50
FedTPG [21]69.5193.7570.0423.2264.7290.6085.9146.2563.9866.7847.8665.31
FedMVP (Ours)70.8795.3772.8025.9470.5889.2787.0649.7865.8368.1950.8467.57
Improvement+1.36+1.62+1.51+1.22+3.83-1.33+0.97+3.53+0.33+1.41-4.56+2.26
+ +Table 6. Ablating FedMVP components. We report numbers for the Base-to-New and MMST DG (on DomainBed) settings. + +
ComponentsBase-to-NewMSST DG
ZS-CLIP74.2470.41
fΘ only75.9471.85
fΘ + Lcon76.2772.14
w/o LoRA77.4172.58
FedMVP (Full)78.2773.02
+ +![](images/46620df8a887974449ef6f18e1c867ef88ecf2192563189d5743674b1f469045.jpg) +(a) Base-to-New + +![](images/67bbe100bb32c1646d18fefb2e85d755dc3522b07aae680627c4d0d50d22db46.jpg) +(b) MSST DG +Figure 3. Sensitivity to participation rate of the clients, on (a) Base-to-New and (b) MSST DG (DomainBed) settings. + +mits $2 \times$ more parameters than FedTPG, it converges nearly $10 \times$ faster, requiring fewer rounds. In particular, as training progresses, only the LoRA parameters in PromptFormer are fine-tuned in a client, thus reducing the trainable parameter count by $267 \times$ . Overall, FedMVP has lower total computational cost with respect to other methods when normalized by the number of communication rounds. + +![](images/58af3d4289b3246397c461095f4d89041ff90d39e97ffc30fc0f3a3f561f5a94.jpg) +(a) Base-to-New +Figure 4. Computational analysis of the methods on (a) Baseto-New and (b) MSST DG (DomainNet) setting. + +![](images/167988ae91ba94fdfe0a58fe677573c94a829f001f23902fd6280f66e8c0a4cb.jpg) +(b) MSST DG + +Size of the clients. In Fig. 5, we vary the number of training classes in each client $K = \{5,10,20\}$ while keeping the size of the dataset fixed. In other words, decreasing the + +number of classes in each client results in an increase in the total number of clients participating in the federated setup. We observe that for all the baselines a higher heterogeneity (or fewer classes per client) leads to lower performance and vice versa. Nevertheless, FedMVP still exhibits a much superior performance when compared to the baselines, outperforming them while having 4x more clients. + +Additional analyses relating to number of shots, choice of LLMs, hyperparameters, and performance in nonfederated setting are reported in Sec. B.5 of the Supp. Mat. + +![](images/3d3f7ced71ce3b5d8b3d8e4c47678e50b5baae3d59b6ba9ecde997b81fe49231.jpg) +(a) Base-to-New +Figure 5. Sensitivity to number of classes per client on (a) Baseto-New and (b) MSST DG (DomainNet) setting. + +![](images/2d05eb353be34e67921fd9206d04b20888df9223d9bdd61e6f974783837ae0ef.jpg) +(b) MSST DG + +# 5. Conclusion + +In this work, we proposed FedMVP, a novel federated prompt tuning framework for CLIP, to address the challenge of reduced generalization to unseen domains and classes. FedMVP combines visual embeddings with textual attribute features to generate multimodal visual prompts, which are tuned using visual prompt tuning. Our findings revealed that the multimodal contextual information in the prompts enhances the generalizability to unseen classes and domains. FedMVP outperforms all its competitors by a notable margin of $+1.57\% - 2.26\%$ across benchmarks. + +Acknowledgements. This work was supported by: JPNP23019, subsidized by the New Energy and Industrial Technology Development Organization (NEDO); SERICS (PE00000014) under the NRRP MUR program funded by the EU - NGEU; EU project ELIAS (no: 101120237), ANT (no: 101169439), ELLIOT (no: 101214398) and Ministero delle Imprese e del Made in Italy (IPCEI Cloud DM 27 giugno 2022 - IPCEI-CL-0000007). + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 4, 6 +[2] Sikai Bai, Jie Zhang, Song Guo, Shuaicheng Li, Jingcai Guo, Jun Hou, Tao Han, and Xiaocheng Lu. Diprompt: Disentangled prompt tuning for multiple latent domain generalization in federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27284-27293, 2024. 3 +[3] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255. IEEE, 2009. 6 +[4] Lijie Fan, Dilip Krishnan, Phillip Isola, Dina Katabi, and Yonglong Tian. Improving clip training with language rewrites. Advances in Neural Information Processing Systems, 36, 2024. 3 +[5] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2): 581-595, 2024. 3 +[6] Ishaan Gulrajani and David Lopez-Paz. In search of lost domain generalization. In International Conference on Learning Representations, 2021. 6, 7 +[7] Tao Guo, Song Guo, Junxiao Wang, Xueyang Tang, and Wenchao Xu. Promptfl: Let federated participants cooperatively learn prompts instead of models-federated learning in age of foundation model. IEEE Transactions on Mobile Computing, 2023. 2, 3, 6, 7, 8 +[8] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 6 +[9] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, page 3, 2022. 5 +[10] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 2 +[11] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 1, 2, 3, 6, 7, 8 +[12] Peter Kairouz, H Brendan McMahan, Brendan Avent, Aurélien Bellet, Mehdi Bennis, Arjun Nitin Bhagoji, Kallista Bonawitz, Zachary Charles, Graham Cormode, Rachel Cummings, et al. Advances and open problems in federated learn + +ing. Foundations and trends® in machine learning, 14(1-2): 1-210, 2021. 3 +[13] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19113-19122, 2023. 2, 6, 7, 8 +[14] Hongxia Li, Wei Huang, Jingya Wang, and Ye Shi. Global and local prompts cooperation via optimal transport for federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12151-12161, 2024. 3, 6, 7, 8 +[15] Wang Lu, HU Xixu, Jindong Wang, and Xing Xie. Fedclip: Fast generalization and personalization for clip in federated learning. In ICLR Workshop on Trustworthy and Reliable Large-Scale Machine Learning Models, 2023. 3, 6, 7, 8 +[16] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 1, 2, 3, 5 +[17] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. In International Conference on Learning Representations, 2023. 3 +[18] Liliane Momeni, Mathilde Caron, Arsha Nagrani, Andrew Zisserman, and Cordelia Schmid. Verbs in action: Improving verb understanding in video-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15579-15591, 2023. 3 +[19] Khalil Muhammad, Qinqin Wang, Diarmuid O'Reilly-Morgan, Elias Tragos, Barry Smyth, Neil Hurley, James Geraci, and Aonghus Lawlor. Fedfast: Going beyond average for faster training of federated recommender systems. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1234-1242, 2020. 1 +[20] Sarah Pratt, Ian Covert, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15691-15701, 2023. 3 +[21] Chen Qiu, Xingyu Li, Chaithanya Kumar Mummadi, Madan Ravi Ganesh, Zhenzhen Li, Lu Peng, and Wan-Yi Lin. Federated text-driven prompt generation for vision-language models. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 3, 5, 6, 7, 8 +[22] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 5, 6, 7, 8 +[23] Jinhyun So, Chaoyang He, Chien-Sheng Yang, Songze Li, Qian Yu, Ramy E Ali, Basak Guler, and Salman Avestimehr. Lightsecagg: a lightweight and versatile design for secure aggregation in federated learning. Proceedings of Machine Learning and Systems, 4:694-720, 2022. 1 + +[24] Guangyu Sun, Matias Mendieta, Aritra Dutta, Xin Li, and Chen Chen. Towards multi-modal transformers in federated learning. In European Conference on Computer Vision, pages 229-246. Springer, 2024. 3 +[25] Zhenyu Sun, Xiaochun Niu, and Ermin Wei. Understanding generalization of federated learning via stability: Heterogeneity matters. In International Conference on Artificial Intelligence and Statistics, pages 676-684. PMLR, 2024. 1 +[26] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, 2017. 2 +[27] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, et al. Robust fine-tuning of zero-shot models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7959-7971, 2022. 2 +[28] Yinghui Xing, Qirui Wu, De Cheng, Shizhou Zhang, Guoqiang Liang, Peng Wang, and Yanning Zhang. Dual modality prompt tuning for vision-language pre-trained model. IEEE Transactions on Multimedia, 26:2056-2068, 2023. 2, 3 +[29] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language in a bottle: Language model guided concept bottlenecks for interpretable image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19187-19197, 2023. 3 +[30] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6757-6767, 2023. 2, 6, 7, 8 +[31] Xin Yao, Chaofeng Huang, and Lifeng Sun. Two-stream federated learning: Reduce the communication costs. In 2018 IEEE Visual Communications and Image Processing (VCIP), pages 1-4. IEEE, 2018. 1 +[32] Xin Yao, Tianchi Huang, Chenglei Wu, Rui-Xiao Zhang, and Lifeng Sun. Federated learning with additional mechanisms on clients to reduce communication costs. arXiv preprint arXiv:1908.05891, 2019. 1 +[33] Liping Yi, Gang Wang, Xiaoguang Liu, Zhuan Shi, and Han Yu. Fedgh: Heterogeneous federated learning with generalized global header. In Proceedings of the 31st ACM International Conference on Multimedia, pages 8686-8696, 2023. 1 +[34] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11975-11986, 2023. 1 +[35] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 2, 3, 6, 7, 8 +[36] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. In- + +ternational Journal of Computer Vision, 130(9):2337-2348, 2022. 1, 2, 3 \ No newline at end of file diff --git a/2025/FedMVP_ Federated Multimodal Visual Prompt Tuning for Vision-Language Models/images.zip b/2025/FedMVP_ Federated Multimodal Visual Prompt Tuning for Vision-Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..da1141799065144a90d529d30167f6534aa5c524 --- /dev/null +++ b/2025/FedMVP_ Federated Multimodal Visual Prompt Tuning for Vision-Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:690003d8182371ac69c25d0f9cc4b5f5c721d59d64f1f5a66f74e67844de7b8b +size 509508 diff --git a/2025/FedMVP_ Federated Multimodal Visual Prompt Tuning for Vision-Language Models/layout.json b/2025/FedMVP_ Federated Multimodal Visual Prompt Tuning for Vision-Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..66cfa23748270773201206effdc557efbaaa77a2 --- /dev/null +++ b/2025/FedMVP_ Federated Multimodal Visual Prompt Tuning for Vision-Language Models/layout.json @@ -0,0 +1,9093 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 69, + 102, + 541, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 102, + 541, + 138 + ], + "spans": [ + { + "bbox": [ + 69, + 102, + 541, + 138 + ], + "type": "text", + "content": "FedMVP: Federated Multimodal Visual Prompt Tuning for Vision-Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 160, + 499, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 160, + 499, + 190 + ], + "spans": [ + { + "bbox": [ + 111, + 160, + 499, + 190 + ], + "type": "text", + "content": "Mainak Singha1 Subhankar Roy2 Sarthak Mehrotra3 Ankit Jha4 Moloud Abdar5 Biplab Banerjee3 Elisa Ricci1,6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 86, + 191, + 189, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 191, + 189, + 203 + ], + "spans": [ + { + "bbox": [ + 86, + 191, + 189, + 203 + ], + "type": "text", + "content": "1 University of Trento, Italy" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 96, + 205, + 182, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 205, + 182, + 217 + ], + "spans": [ + { + "bbox": [ + 96, + 205, + 182, + 217 + ], + "type": "text", + "content": "4 LNMIIT Jaipur, India" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 213, + 191, + 326, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 191, + 326, + 203 + ], + "spans": [ + { + "bbox": [ + 213, + 191, + 326, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 213, + 191, + 326, + 203 + ], + "type": "text", + "content": " University of Bergamo, Italy" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 209, + 205, + 363, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 205, + 363, + 217 + ], + "spans": [ + { + "bbox": [ + 209, + 205, + 363, + 217 + ], + "type": "text", + "content": "5 The University of Queensland, Australia" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 352, + 191, + 524, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 191, + 524, + 203 + ], + "spans": [ + { + "bbox": [ + 352, + 191, + 524, + 203 + ], + "type": "text", + "content": "3 Indian Institute of Technology Bombay, India" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 390, + 205, + 515, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 205, + 515, + 217 + ], + "spans": [ + { + "bbox": [ + 390, + 205, + 515, + 217 + ], + "type": "text", + "content": "Fondazione Bruno Kessler, Italy" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 152, + 244, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 244, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 152, + 244, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 272, + 296, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 272, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 272, + 296, + 559 + ], + "type": "text", + "content": "In federated learning, textual prompt tuning adapts Vision-Language Models (e.g., CLIP) by tuning lightweight input tokens (or prompts) on local client data, while keeping network weights frozen. After training, only the prompts are shared by the clients with the central server for aggregation. However, textual prompt tuning suffers from overfitting to known concepts, limiting its generalizability to unseen concepts. To address this limitation, we propose Multimodal Visual Prompt Tuning (FedMVP) that conditions the prompts on multimodal contextual information - derived from the input image and textual attribute features of a class. At the core of FedMVP is a PromptFormer module that synergistically aligns textual and visual features through a cross-attention mechanism. The dynamically generated multimodal visual prompts are then input to the frozen vision encoder of CLIP, and trained with a combination of CLIP similarity loss and a consistency loss. Extensive evaluation on 20 datasets, spanning three generalization settings, demonstrates that FedMVP not only preserves performance on in-distribution classes and domains, but also displays higher generalizability to unseen classes and domains, surpassing state-of-the-art methods by a notable margin of " + }, + { + "bbox": [ + 55, + 272, + 296, + 559 + ], + "type": "inline_equation", + "content": "+1.57\\% - 2.26\\%" + }, + { + "bbox": [ + 55, + 272, + 296, + 559 + ], + "type": "text", + "content": ". Code is available at https://github.com/mainaksingha01/FedMVP." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 572, + 135, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 572, + 135, + 585 + ], + "spans": [ + { + "bbox": [ + 56, + 572, + 135, + 585 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "A successful learning paradigm involves training neural networks on large and centralized datasets, enabling models to learn robust and generalizable representations [22, 34]. However, practical constraints often prevent centralized training (e.g., privacy regulations) creating a critical need for decentralized training frameworks that achieve high performance without central data access. As a solution, federated learning (FL) [16] enables multiple clients to collaboratively train a global model without requiring direct data sharing. In FL, clients train models locally and trans" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 320, + 246, + 547, + 381 + ], + "blocks": [ + { + "bbox": [ + 320, + 246, + 547, + 381 + ], + "lines": [ + { + "bbox": [ + 320, + 246, + 547, + 381 + ], + "spans": [ + { + "bbox": [ + 320, + 246, + 547, + 381 + ], + "type": "image", + "image_path": "cad9b44ec8e2dd22c7f3c0f33c25ad62cf44a95a38038d12e2976470c5e4af54.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 388, + 555, + 444 + ], + "lines": [ + { + "bbox": [ + 313, + 388, + 555, + 444 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 555, + 444 + ], + "type": "text", + "content": "Figure 1. Comparison of federated prompt tuning methods: (a) a textual prompt tuning method (e.g., FedTPG [21]) encodes unimodal contextual information, i.e., class names, in the learnable prompts, (b) our proposed method FedMVP injects multimodal contextual information – image and attributes – into the prompts." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 449, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 533 + ], + "type": "text", + "content": "mit only model updates (e.g., gradients or parameters) to a central server, which aggregates these updates to improve the global model. The research in FL is primarily aimed at: reducing communication rounds between clients and server [31, 32], faster and lightweight training algorithms on clients [19, 23] and improving generalization while exhibiting high data heterogeneity across clients [25, 33]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "content": "Pretrained Vision-Language Models (VLMs) (e.g., CLIP [22]), have emerged as promising candidates for FL due to their strong generalization capabilities. However, the high communication overhead and computational demands of parameter-heavy VLMs pose challenges for FL, especially in bandwidth- or compute-constrained environments (e.g., robots). To make CLIP viable for the FL setting, prompt tuning (e.g., CoOp [36], VPT [11]) has been adopted as a preferred technique that appends lightweight learnable tokens (or prompts) to the input tokens of the CLIP image and/or text encoder. Each client trains the prompts on the local data, while keeping the CLIP backbone frozen, and then dispatches these prompts to the server, where they are updated through aggregation and sent back. This speeds up client training as each client needs access" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17869" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "to only a few samples per class (e.g., 16/class [36]), and reduces communication overhead as only the lightweight prompts (e.g., " + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "0.37\\%" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": " of network parameters) need to be communicated, instead of all the parameters." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 122, + 295, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 122, + 295, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 122, + 295, + 301 + ], + "type": "text", + "content": "Although prompt tuning has proven to be effective in the FL setting, it suffers from reduced generalization to unseen classes and unseen domains - a typical heterogeneous setup native to non independent and identically distributed (non-i.i.d.) FL [21] - due to severe overfitting on the local training data [30]. It is attributed to the fact that prompts learn a static context from a set of seen classes, which being fixed once learned, fail to capture the generalizable elements spread across clients. To circumvent the issue of reduced generalization, recent CLIP-based FL methods condition the prompts either on the class information, as in FedTPG [21] (see Fig. 1a), or on visual information, as in FedCoCoOp [35]. This auxiliary conditioning on the input, or contextual information, makes the prompts dynamic and thus offers a better generalization to unseen distributions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 304, + 295, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 295, + 555 + ], + "type": "text", + "content": "In this work, we argue that the FL setup being highly heterogeneous, where clients have data from disjoint classes and domains [21], it is not sufficient to solely exploit either of the two modalities for contextual information. Rather, FL demands more comprehensive contextual information that encompasses all the available modalities to condition input prompts. To this end, we propose Federated Multimodal Visual Prompt Tuning (FedMVP) that conditions the prompts on multimodal contextual information, which comprise: (i) visual features of the input image itself, and (ii) textual features of attributes of a class (e.g., \"two legs\", \"beak\", etc. for the class \"hen\"). The key intuition is that (ii) promotes learning transferable prompts for unseen classes that share attributes similar to seen classes (e.g., unseen class \"seagull\" will share some attributes with the seen class \"hen\"), and (i) promotes transferability for unseen classes and domains where attributes cannot be described via text (e.g., texture or abstract concepts). In a nutshell, the dual conditioning offers richer contextual information to the prompts, which helps in generalizing to classes or domains that share similar properties." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": "In detail, the proposed FedMVP framework contains a PromptFormer network (see Fig. 1b) that synergistically merges the information from visual and textual modalities with cross-attention mechanism [26] to generate the prompts. The resulting multimodal prompts are then injected at the visual space of the frozen vision encoder of CLIP, unlike the existing methods [7, 21, 30] that inject prompts through text encoder. The key differences between our FedMVP and existing CLIP-based FL methods have been summarized in Tab. 1. The PromptFormer is trained in an end-to-end manner with the CLIP loss and a consistency loss. Once trained, each participating client shares the lightweight PromptFormer network parameters with the" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 71, + 553, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 71, + 553, + 137 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 553, + 137 + ], + "type": "text", + "content": "Table 1. Summary of federated prompt-tuning methods. " + }, + { + "bbox": [ + 313, + 71, + 553, + 137 + ], + "type": "inline_equation", + "content": "(X)" + }, + { + "bbox": [ + 313, + 71, + 553, + 137 + ], + "type": "text", + "content": " indicates the learnable prompts are initialized randomly, " + }, + { + "bbox": [ + 313, + 71, + 553, + 137 + ], + "type": "inline_equation", + "content": "(\\checkmark)" + }, + { + "bbox": [ + 313, + 71, + 553, + 137 + ], + "type": "text", + "content": " signifies learnable prompts are conditioned on contextual information (e.g., textual or visual inputs). Prompting column denotes the CLIP encoder wherein learnable prompts are given as input (e.g., \"Textual\" means the text encoder of CLIP takes prompts as input)." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 333, + 145, + 533, + 225 + ], + "blocks": [ + { + "bbox": [ + 333, + 145, + 533, + 225 + ], + "lines": [ + { + "bbox": [ + 333, + 145, + 533, + 225 + ], + "spans": [ + { + "bbox": [ + 333, + 145, + 533, + 225 + ], + "type": "table", + "html": "
MethodContextual informationPrompting
TextualVisual
FedKgCoOp [30]XXTextual
FedVPT [11]XXVisual
FedTPG [21]✓XTextual
FedCoCoOp [35]X✓Textual
FedMaPLe [13]XXMultimodal
FedMVP (Ours)✓✓Visual
", + "image_path": "2a3e8f268ca86406aff990874457d254657e8df5752ccd99c0698e12f41e2df9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 239, + 553, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 553, + 358 + ], + "type": "text", + "content": "server for aggregation through FedAvg [16], and thus enabling knowledge sharing among clients. During inference, the cross-attention learned by the PromptFormer enables the model to dynamically adjust the prompts, enhancing generalizability to unseen classes and domains. We evaluated FedMVP on three generalization settings encompassing 20 datasets that measure generalization to unseen classes, domains, and combinations of them. Experiments demonstrate that FedMVP consistently outperforms the state-of-the-art CLIP-based FL methods by " + }, + { + "bbox": [ + 313, + 239, + 553, + 358 + ], + "type": "inline_equation", + "content": "1.57\\% - 2.26\\%" + }, + { + "bbox": [ + 313, + 239, + 553, + 358 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 360, + 555, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 555, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 555, + 467 + ], + "type": "text", + "content": "In summary, our main contributions are: (i) We highlight the importance of leveraging multimodal contextual information in prompt tuning based FL setup in order to alleviate the issue of reduced generalizability to unseen data. (ii) We propose FedMVP that synergistically combines the visual features of the image and textual features of the attributes of classes to generate multimodal prompts. This provides richer contextual information to the model and higher transferability to unseen classes and domains." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 478, + 400, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 478, + 400, + 491 + ], + "spans": [ + { + "bbox": [ + 314, + 478, + 400, + 491 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 498, + 553, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 553, + 569 + ], + "type": "text", + "content": "Prompt tuning of VLMs. VLMs, such as CLIP [22] and ALIGN [10], leverage large image-text datasets to enable multimodal understanding, performing well in zero-shot classification tasks. However, fully fine-tuning VLMs for downstream tasks without sacrificing their ability to generalize to unseen concepts remains a challenge [27]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 570, + 554, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 700 + ], + "type": "text", + "content": "Prompt tuning approaches [11, 13, 36] have been proposed for parameter efficient and effective adaptation of CLIP by introducing a set of continuous learnable tokens (or prompts) onto the frozen CLIP backbone. These approaches can be broadly categorized into three categories depending on the manner in which CLIP is prompted: (i) textual prompting that inject prompts in the input space of the text encoder [30, 36], (ii) visual prompting that injects prompts in the input space of the vision encoder [11], and (iii) multimodal prompting that injects prompts in both the text and vision encoder [13, 28]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "type": "text", + "content": "In particular, MaPLe [13], which injects prompts to both" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17870" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "content": "the text and vision encoders, is not suitable for memory and compute efficient FL since it requires parts of both the encoders to be unlocked during tuning, risking overfitting. Moreover, it does not allow for image-conditioned dynamic prompting, which has shown to be effective for generalization [35]. In contrast, our proposed FedMVP enables multimodal prompting by involving only the vision encoder and supports instance-specific prompting." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 170, + 295, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 170, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 170, + 295, + 374 + ], + "type": "text", + "content": "Federated prompt tuning of VLMs. Prompt tuning has emerged as an effective technique for FL of VLMs since it allows clients to collaboratively train on a heterogeneously distributed dataset and adapt faster with fewer samples. The fact that the prompts are lightweight lends well to the FL setup as it reduces communication overhead. Due to the benefits offered by prompt tuning, several works [2, 7, 21, 24] have specifically adapted prompt tuning approaches to the FL setup. For example, PromptFL [7] extended CoOp to FL via sharing prompts with the server, enhancing the generalization performance of client-server interactions in vision tasks. FedCLIP [15] learns a lightweight adapter in each client, similar to the CLIP-Adapter [5], and shares parameters with the server. FedOTP [14] introduced global and local prompts, optimizing shared parameters through optimal transport, while DiPrompt [2] studied multi-domain learning by allowing multiple domains per client." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 376, + 295, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 295, + 520 + ], + "type": "text", + "content": "Unlike existing methods, a couple of recent works [21, 28] argue that prompt tuning is prone to overfitting to the seen classes due to learning static context. The closest work to ours is FedTPG [21], where a prompt generation network is learned across multiple clients. Additionally, this network leverages textual inputs, enhancing its ability to generalize to unseen classes. Our work significantly differs from previous methods in two key ways: (i) we exploit both attribute information derived from an LLM and image features to condition the prompts, and (ii) we leverage multimodal prompts from the proposed PromptFormer for visual prompt tuning, instead of textual prompt tuning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": "Classification via textual descriptions. Zero-shot accuracy of CLIP on downstream tasks is sensitive to the quality of text prompts [22]. In detail, some works rely on simple hand-crafted prompts (e.g., \"a photo of a [CLS]\") [22] or others augment the hand-crafted prompts with richer attributes obtained from LLMs [17, 20] (e.g., a \"hen\" has two legs, white feathers, etc.). There also exists a family of methods, such as LaCLIP [4], LaBo [29] and VFC [18], that have proposed to further refine CLIP weights with enriched captions from LLMs for improved performance. Recognizing the importance of class attributes for improved generalization to unseen concepts and domains – an important desideratum in the FL set up – we, for the first time in FL, introduce LLM generated class attributes for enriching the image-conditioned multimodal prompts through the PromptFormer network." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 71, + 375, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 375, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 375, + 83 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 91, + 555, + 118 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 118 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 118 + ], + "type": "text", + "content": "We introduce FL setup in Sec. 3.1, preliminaries on prompt tuning in Sec. 3.2 and our proposed solution in Sec. 3.3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 125, + 436, + 137 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 125, + 436, + 137 + ], + "spans": [ + { + "bbox": [ + 314, + 125, + 436, + 137 + ], + "type": "text", + "content": "3.1. Problem formulation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 144, + 553, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 553, + 301 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 553, + 301 + ], + "type": "text", + "content": "We consider a FL setup in which multiple remote clients train models on local data, and a central server aggregates models from the clients. Formally, let each remote client " + }, + { + "bbox": [ + 313, + 144, + 553, + 301 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 144, + 553, + 301 + ], + "type": "text", + "content": " has access to a dataset " + }, + { + "bbox": [ + 313, + 144, + 553, + 301 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_i = \\{\\mathbf{x}_{i,j},y_{i,j}\\}_{j = 1}^{n_i}" + }, + { + "bbox": [ + 313, + 144, + 553, + 301 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 144, + 553, + 301 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i,j}" + }, + { + "bbox": [ + 313, + 144, + 553, + 301 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 313, + 144, + 553, + 301 + ], + "type": "inline_equation", + "content": "j^{\\mathrm{th}}" + }, + { + "bbox": [ + 313, + 144, + 553, + 301 + ], + "type": "text", + "content": " image from the " + }, + { + "bbox": [ + 313, + 144, + 553, + 301 + ], + "type": "inline_equation", + "content": "i^{\\mathrm{th}}" + }, + { + "bbox": [ + 313, + 144, + 553, + 301 + ], + "type": "text", + "content": " client, " + }, + { + "bbox": [ + 313, + 144, + 553, + 301 + ], + "type": "inline_equation", + "content": "y_{i,j}" + }, + { + "bbox": [ + 313, + 144, + 553, + 301 + ], + "type": "text", + "content": " is the corresponding class label and " + }, + { + "bbox": [ + 313, + 144, + 553, + 301 + ], + "type": "inline_equation", + "content": "n_i" + }, + { + "bbox": [ + 313, + 144, + 553, + 301 + ], + "type": "text", + "content": " is the number of samples in client " + }, + { + "bbox": [ + 313, + 144, + 553, + 301 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 144, + 553, + 301 + ], + "type": "text", + "content": ". Let the set of text class names in client " + }, + { + "bbox": [ + 313, + 144, + 553, + 301 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 144, + 553, + 301 + ], + "type": "text", + "content": " be denoted as " + }, + { + "bbox": [ + 313, + 144, + 553, + 301 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i = \\{c_{i,k}\\}_{k = 1}^{K_i}" + }, + { + "bbox": [ + 313, + 144, + 553, + 301 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 144, + 553, + 301 + ], + "type": "inline_equation", + "content": "K_{i}" + }, + { + "bbox": [ + 313, + 144, + 553, + 301 + ], + "type": "text", + "content": " is the total number of classes. As in FedTPG [21], we follow a non-IID FL setup, where the samples in each client are drawn from a disjoint set of classes, i.e., " + }, + { + "bbox": [ + 313, + 144, + 553, + 301 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i\\cap \\mathcal{C}_l = \\emptyset" + }, + { + "bbox": [ + 313, + 144, + 553, + 301 + ], + "type": "text", + "content": ". As an example, the training data on each mobile device usually depend on the user, which may not be representative of the general population [16]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 302, + 554, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 302, + 554, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 302, + 554, + 385 + ], + "type": "text", + "content": "In particular, we investigate two kinds of non-IID FL setups, where, in addition to the in-distribution performance, we also evaluate the generalization ability of the aggregated server model on (i) unseen classes, and (ii) unseen domains or datasets. The need for generalization to unseen classes and domains makes the non-IID FL setup more challenging than the traditional FL setting [12]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 396, + 489, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 396, + 489, + 408 + ], + "spans": [ + { + "bbox": [ + 314, + 396, + 489, + 408 + ], + "type": "text", + "content": "3.2. Preliminaries on Prompt Tuning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 414, + 554, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 554, + 557 + ], + "type": "text", + "content": "Textual prompt tuning (TPT), i.e., CoOp [36], has emerged as an efficient technique in CLIP-based FL that keeps the CLIP backbone frozen and locally learns lightweight prompt vectors on each client. The locally learned prompts are then communicated to the server, as in PromptFL [7]. Specifically, the hand-crafted text prompts (e.g., \"a photo of a [CLS]\") input to the text encoder " + }, + { + "bbox": [ + 313, + 414, + 554, + 557 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_t" + }, + { + "bbox": [ + 313, + 414, + 554, + 557 + ], + "type": "text", + "content": " is augmented with the learnable soft prompts " + }, + { + "bbox": [ + 313, + 414, + 554, + 557 + ], + "type": "inline_equation", + "content": "v_{1}, v_{2}, \\ldots, v_{m}" + }, + { + "bbox": [ + 313, + 414, + 554, + 557 + ], + "type": "text", + "content": ". However, as shown in [35], TPT can often lead to overfitting and reduced generalization due to the static context learned on the seen classes. This can be exacerbated in the non-IID FL setup due to the high data heterogeneity among the clients [21]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "type": "text", + "content": "Alternatively, visual prompt tuning (VPT) [11] adapts pre-trained vision encoder " + }, + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_v" + }, + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "type": "text", + "content": " by injecting small number of visual learnable parameters into ViT's input space. More formally, let an input image " + }, + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "type": "text", + "content": " be divided in " + }, + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "type": "text", + "content": " patches, yielding a collection of patch embeddings (augmented with positional encodings) at the input of the vision encoder as " + }, + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{E} \\in \\mathbb{R}^{b \\times d_v}" + }, + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "type": "inline_equation", + "content": "d_v" + }, + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "type": "text", + "content": " is the embedding dimension. A [CLS] token " + }, + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "type": "text", + "content": " is concatenated with " + }, + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "type": "text", + "content": ", encapsulating the initial input representation for the ViT layers of " + }, + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_v" + }, + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "type": "text", + "content": ". VPT further augments the input by concatenating a set of randomly initialized learnable visual prompts " + }, + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{P} = \\{p_1, p_2, \\ldots, p_m\\}" + }, + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "type": "text", + "content": " is the prompt length. The first layer input of " + }, + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_v" + }, + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "type": "text", + "content": " is then redefined as " + }, + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{I} = [\\mathbf{z}; \\mathbf{E}; \\mathbf{P}] \\in \\mathbb{R}^{(1 + b + m) \\times d_v}" + }, + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "type": "text", + "content": ", where" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17871" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 71, + 203, + 258 + ], + "blocks": [ + { + "bbox": [ + 60, + 71, + 203, + 258 + ], + "lines": [ + { + "bbox": [ + 60, + 71, + 203, + 258 + ], + "spans": [ + { + "bbox": [ + 60, + 71, + 203, + 258 + ], + "type": "image", + "image_path": "8d12d753ddcaf747b6524df2d4c26ff3aae9fb9a19b8357fefababb4f875ae81.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 261, + 554, + 307 + ], + "lines": [ + { + "bbox": [ + 54, + 261, + 554, + 307 + ], + "spans": [ + { + "bbox": [ + 54, + 261, + 554, + 307 + ], + "type": "text", + "content": "Figure 2. Overview of Federated Multimodal Visual Prompt Tuning (FedMVP). Each client " + }, + { + "bbox": [ + 54, + 261, + 554, + 307 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 54, + 261, + 554, + 307 + ], + "type": "text", + "content": " trains the parameters of the PromptFormer network " + }, + { + "bbox": [ + 54, + 261, + 554, + 307 + ], + "type": "inline_equation", + "content": "f_{\\Theta}" + }, + { + "bbox": [ + 54, + 261, + 554, + 307 + ], + "type": "text", + "content": " and multi-head self-attention on its local data. At each round, the server aggregates the trainable parameters from the clients and sends the updated parameters back to the clients. The PromptFormer uses cross-attention mechanism to fuse the visual and textual attribute feature to generate multimodal visual prompts " + }, + { + "bbox": [ + 54, + 261, + 554, + 307 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 54, + 261, + 554, + 307 + ], + "type": "text", + "content": ", which are then used for visual prompt tuning." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 203, + 72, + 369, + 257 + ], + "blocks": [ + { + "bbox": [ + 203, + 72, + 369, + 257 + ], + "lines": [ + { + "bbox": [ + 203, + 72, + 369, + 257 + ], + "spans": [ + { + "bbox": [ + 203, + 72, + 369, + 257 + ], + "type": "image", + "image_path": "7c528344f045fb89837191fb059d5e5051c445481371681448063518fdd92c12.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 369, + 72, + 516, + 258 + ], + "blocks": [ + { + "bbox": [ + 369, + 72, + 516, + 258 + ], + "lines": [ + { + "bbox": [ + 369, + 72, + 516, + 258 + ], + "spans": [ + { + "bbox": [ + 369, + 72, + 516, + 258 + ], + "type": "image", + "image_path": "2ece183bd23848b010a1a9ce2ceff94d4ceb58813e90864c10c8b8702d125c59.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 312, + 294, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 312, + 294, + 349 + ], + "spans": [ + { + "bbox": [ + 54, + 312, + 294, + 349 + ], + "type": "text", + "content": "“;” means concatenation. Although VPT enables more nuanced understanding of the visual space, it suffers from the same issues as TPT, i.e., limited generalization (see Sec. 4)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 357, + 291, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 291, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 291, + 369 + ], + "type": "text", + "content": "3.3. Federated Multimodal Visual Prompt Tuning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 374, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 374, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 54, + 374, + 295, + 567 + ], + "type": "text", + "content": "Our proposed method FedMVP deviates from the TPT and VPT paradigms and incorporates multimodal contextual information from both the visual and textual modalities to improve generalization in the FL setup. As shown in Fig. 2, at the core of our novel architecture is a prompt generator, called PromptFormer, which conditions the visual prompts " + }, + { + "bbox": [ + 54, + 374, + 295, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 54, + 374, + 295, + 567 + ], + "type": "text", + "content": " on the task-related information: input image and text attributes corresponding to the class names, rather than randomly initializing them. This makes " + }, + { + "bbox": [ + 54, + 374, + 295, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 54, + 374, + 295, + 567 + ], + "type": "text", + "content": " context-aware, generalizing better to unseen classes and domains. Once trained on each client, only the parameters of the lightweight PromptFormer network are communicated to the server, thus, keeping the communication overhead low. Next, we describe PromptFormer (Sec. 3.3.1), training objectives (Sec. 3.3.2) and how we conduct client-side training and server-side aggregation (Sec. 3.3.3)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 573, + 285, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 573, + 285, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 573, + 285, + 586 + ], + "type": "text", + "content": "3.3.1. Visual Prompt Generation with PromptFormer" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 588, + 295, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 295, + 684 + ], + "type": "text", + "content": "We introduce the PromptFormer network " + }, + { + "bbox": [ + 55, + 588, + 295, + 684 + ], + "type": "inline_equation", + "content": "f_{\\Theta}" + }, + { + "bbox": [ + 55, + 588, + 295, + 684 + ], + "type": "text", + "content": ", parameterized by " + }, + { + "bbox": [ + 55, + 588, + 295, + 684 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 55, + 588, + 295, + 684 + ], + "type": "text", + "content": ", designed to generate contextually rich visual tokens (or prompts) for each client " + }, + { + "bbox": [ + 55, + 588, + 295, + 684 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 588, + 295, + 684 + ], + "type": "text", + "content": " by leveraging both attribute features and visual patch embeddings (see Fig. 2). Concretely, we use the CLIP text encoder " + }, + { + "bbox": [ + 55, + 588, + 295, + 684 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_t" + }, + { + "bbox": [ + 55, + 588, + 295, + 684 + ], + "type": "text", + "content": " to extract embeddings " + }, + { + "bbox": [ + 55, + 588, + 295, + 684 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i = \\{\\mathcal{E}_t(\\mathrm{LLM}(c_k))\\}_{k=1}^K" + }, + { + "bbox": [ + 55, + 588, + 295, + 684 + ], + "type": "text", + "content": " of the text attributes " + }, + { + "bbox": [ + 55, + 588, + 295, + 684 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_i" + }, + { + "bbox": [ + 55, + 588, + 295, + 684 + ], + "type": "text", + "content": " corresponding to the text class names " + }, + { + "bbox": [ + 55, + 588, + 295, + 684 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i" + }, + { + "bbox": [ + 55, + 588, + 295, + 684 + ], + "type": "text", + "content": ", generated using a large lanuage model (LLM) [1] (see Sec. A.1 of Supp. Mat." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 312, + 554, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 554, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 554, + 372 + ], + "type": "text", + "content": "for Attribute Generation process). Going forward, we drop the client index " + }, + { + "bbox": [ + 313, + 312, + 554, + 372 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 312, + 554, + 372 + ], + "type": "text", + "content": " for notational clarity. The PromptFormer takes as input " + }, + { + "bbox": [ + 313, + 312, + 554, + 372 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 312, + 554, + 372 + ], + "type": "text", + "content": " visual patch embeddings " + }, + { + "bbox": [ + 313, + 312, + 554, + 372 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 313, + 312, + 554, + 372 + ], + "type": "text", + "content": ", corresponding to an image " + }, + { + "bbox": [ + 313, + 312, + 554, + 372 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 313, + 312, + 554, + 372 + ], + "type": "text", + "content": ", and the text embeddings " + }, + { + "bbox": [ + 313, + 312, + 554, + 372 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 313, + 312, + 554, + 372 + ], + "type": "text", + "content": " to generate a set of enriched " + }, + { + "bbox": [ + 313, + 312, + 554, + 372 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 312, + 554, + 372 + ], + "type": "text", + "content": " multimodal visual prompts " + }, + { + "bbox": [ + 313, + 312, + 554, + 372 + ], + "type": "inline_equation", + "content": "\\mathbf{P} \\in \\mathbb{R}^{m \\times d_v}" + }, + { + "bbox": [ + 313, + 312, + 554, + 372 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 376, + 379, + 553, + 393 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 379, + 553, + 393 + ], + "spans": [ + { + "bbox": [ + 376, + 379, + 553, + 393 + ], + "type": "interline_equation", + "content": "\\mathbf {P} = \\left\\{p _ {j} \\right\\} _ {j = 1} ^ {m} = f _ {\\Theta} (\\mathbf {A}, \\mathbf {E}). \\tag {1}", + "image_path": "530653259d88842169a5afe7a113b84b0b469babaf52d985abaf518276caf73a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 402, + 554, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 554, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 554, + 461 + ], + "type": "text", + "content": "Consequently, the input to the first layer of the visual encoder " + }, + { + "bbox": [ + 313, + 402, + 554, + 461 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_v" + }, + { + "bbox": [ + 313, + 402, + 554, + 461 + ], + "type": "text", + "content": " is redefined as " + }, + { + "bbox": [ + 313, + 402, + 554, + 461 + ], + "type": "inline_equation", + "content": "[\\mathbf{z};\\mathbf{E};\\mathbf{P}]\\in \\mathbb{R}^{(1 + b + m)\\times d_v}" + }, + { + "bbox": [ + 313, + 402, + 554, + 461 + ], + "type": "text", + "content": ". Intuitively, multimodal visual prompts carry richer contextual information, and helps the network in generalizing to unseen classes and domains that share commonalities." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 472, + 554, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 554, + 567 + ], + "type": "text", + "content": "Architecture of PromptFormer. The PromptFormer architecture comprises: lightweight multi-head cross-attention (MHCA) modules, feed-forward network (FFN) and a linear projection layer; transforming multimodal features into comprehensive visual prompts. In detail, the visual patch embedding " + }, + { + "bbox": [ + 313, + 472, + 554, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 313, + 472, + 554, + 567 + ], + "type": "text", + "content": " interacts with the linearly projected attribute features " + }, + { + "bbox": [ + 313, + 472, + 554, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{A}' = T_{\\mathrm{proj}}(\\mathbf{A})" + }, + { + "bbox": [ + 313, + 472, + 554, + 567 + ], + "type": "text", + "content": " to construct multimodal visual prompts as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 326, + 576, + 541, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 576, + 541, + 590 + ], + "spans": [ + { + "bbox": [ + 326, + 576, + 541, + 590 + ], + "type": "interline_equation", + "content": "\\mathbf {P} \\left(\\mathbf {A} ^ {\\prime}, \\mathbf {E}\\right) = \\operatorname {F F N} \\left(\\text {C r o s s A t t e n t i o n} \\left(\\mathbf {Q} _ {\\mathbf {E}}, \\mathbf {K} _ {\\mathbf {A} ^ {\\prime}}, \\mathbf {V} _ {\\mathbf {A} ^ {\\prime}}\\right)\\right);", + "image_path": "b85f28ca5c57f64702196caedeaa513a26121b30e30dffddbebbc8da00adb07f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 329, + 593, + 553, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 593, + 553, + 617 + ], + "spans": [ + { + "bbox": [ + 329, + 593, + 553, + 617 + ], + "type": "interline_equation", + "content": "\\text {w i t h} \\mathbf {Q} _ {\\mathbf {E}} = \\mathbf {E} W _ {\\mathbf {Q}}, \\mathbf {K} _ {\\mathbf {A} ^ {\\prime}} = \\mathbf {A} ^ {\\prime} W _ {\\mathbf {K}}, \\mathbf {V} _ {\\mathbf {A} ^ {\\prime}} = \\mathbf {A} ^ {\\prime} W _ {\\mathbf {V}}, \\tag {2}", + "image_path": "4c76d15bfdd1ebada79e971c847954239d3ac4771aa7aa4d0763d51a15ba580f.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 617, + 553, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 553, + 666 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 313, + 617, + 553, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 313, + 617, + 553, + 666 + ], + "type": "text", + "content": " are transformed into query vectors " + }, + { + "bbox": [ + 313, + 617, + 553, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathbf{E}}" + }, + { + "bbox": [ + 313, + 617, + 553, + 666 + ], + "type": "text", + "content": " using query matrices " + }, + { + "bbox": [ + 313, + 617, + 553, + 666 + ], + "type": "inline_equation", + "content": "W_{\\mathbf{Q}}" + }, + { + "bbox": [ + 313, + 617, + 553, + 666 + ], + "type": "text", + "content": "; and the projected attribute features " + }, + { + "bbox": [ + 313, + 617, + 553, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{A}'" + }, + { + "bbox": [ + 313, + 617, + 553, + 666 + ], + "type": "text", + "content": " are transformed into key and value vectors " + }, + { + "bbox": [ + 313, + 617, + 553, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_{\\mathbf{A}'}" + }, + { + "bbox": [ + 313, + 617, + 553, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 617, + 553, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{\\mathbf{A}'}" + }, + { + "bbox": [ + 313, + 617, + 553, + 666 + ], + "type": "text", + "content": ", using key and value matrices " + }, + { + "bbox": [ + 313, + 617, + 553, + 666 + ], + "type": "inline_equation", + "content": "W_{\\mathbf{K}}" + }, + { + "bbox": [ + 313, + 617, + 553, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 617, + 553, + 666 + ], + "type": "inline_equation", + "content": "W_{\\mathbf{V}}" + }, + { + "bbox": [ + 313, + 617, + 553, + 666 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "content": "Intuitively, through the cross-attention mechanism, visual features learn to attend to the relevant parts of the corresponding attribute features. For example, the patches depicting the legs of the class \"dog\" will learn to attend to" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 714 + ], + "type": "text", + "content": "1For example, for the class name \"giraffe\" the LLM outputs \"Exceptionally long neck, unique coat pattern with irregular brown patches, ...\"" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17872" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": "the attributes four legged. When an unseen class (e.g., \"giraffe\") comprises an animal with four legs, the final prompts " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": " will contain this relevant information derived from the attributes four legged. Additional details of PromptFormer architecture are reported in Sec. A.3 of the Supp. Mat." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 138, + 173, + 149 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 138, + 173, + 149 + ], + "spans": [ + { + "bbox": [ + 55, + 138, + 173, + 149 + ], + "type": "text", + "content": "3.3.2. Training of FedMVP" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 153, + 295, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 153, + 295, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 153, + 295, + 201 + ], + "type": "text", + "content": "After the visual prompts are obtained with PromptFormer as per Eq. 2, we concatenate the [CLS] token, the patch embeddings, and the visual prompts to get " + }, + { + "bbox": [ + 55, + 153, + 295, + 201 + ], + "type": "inline_equation", + "content": "\\mathbf{I} = [\\mathbf{z};\\mathbf{E};\\mathbf{P}]" + }, + { + "bbox": [ + 55, + 153, + 295, + 201 + ], + "type": "text", + "content": " following Sec. 3.2, and feed it to the visual encoder " + }, + { + "bbox": [ + 55, + 153, + 295, + 201 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_v" + }, + { + "bbox": [ + 55, + 153, + 295, + 201 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 201, + 296, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 201, + 296, + 250 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 296, + 250 + ], + "type": "text", + "content": "The parameters of PromptFormer " + }, + { + "bbox": [ + 55, + 201, + 296, + 250 + ], + "type": "inline_equation", + "content": "\\rho = \\{\\Theta\\}" + }, + { + "bbox": [ + 55, + 201, + 296, + 250 + ], + "type": "text", + "content": " are the only trainable parameters of FedMVP, and the rest of the weights are kept frozen. To train the model, we employ the CLIP cross-entropy loss [22], formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 255, + 295, + 275 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 255, + 295, + 275 + ], + "spans": [ + { + "bbox": [ + 116, + 255, + 295, + 275 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c e}} = - \\underset {(\\mathbf {x}, y) \\sim \\mathcal {D}} {\\mathbb {E}} y \\log p (y | \\mathbf {I}), \\tag {3}", + "image_path": "4c6934901e5c775a424440ac6cbc46ec551bffcac3390b40782553c0bbc7f0cb.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 283, + 296, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 296, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 296, + 331 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 283, + 296, + 331 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 283, + 296, + 331 + ], + "type": "text", + "content": " denotes the ground truth label, and the prediction probability of CLIP between the visual features " + }, + { + "bbox": [ + 55, + 283, + 296, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{v} = \\mathcal{E}_v(\\mathbf{I})" + }, + { + "bbox": [ + 55, + 283, + 296, + 331 + ], + "type": "text", + "content": " and text features " + }, + { + "bbox": [ + 55, + 283, + 296, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_k = \\mathcal{E}_t([A\\text{ photo of [CLASS]};\\mathrm{LLM}(c_k)])" + }, + { + "bbox": [ + 55, + 283, + 296, + 331 + ], + "type": "text", + "content": " for class " + }, + { + "bbox": [ + 55, + 283, + 296, + 331 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 283, + 296, + 331 + ], + "type": "text", + "content": " is computed as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 86, + 332, + 295, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 332, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 86, + 332, + 295, + 361 + ], + "type": "interline_equation", + "content": "p (y = k | \\mathbf {I}) = \\frac {\\exp \\left(\\cos \\left(\\mathbf {v} , \\mathbf {t} _ {k}\\right)\\right) / \\tau)}{\\sum_ {k ^ {\\prime} = 1} ^ {K} \\exp \\left(\\cos \\left(\\mathbf {v} , \\mathbf {t} _ {k ^ {\\prime}}\\right)\\right) / \\tau}, \\tag {4}", + "image_path": "81c293c8c3fce62ce05630f73e3cd82f32e29de98d9cb981530a6bf670f733ae.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 369, + 296, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 296, + 405 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 296, + 405 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 369, + 296, + 405 + ], + "type": "inline_equation", + "content": "\\cos (\\cdot ,\\cdot)" + }, + { + "bbox": [ + 55, + 369, + 296, + 405 + ], + "type": "text", + "content": " as a similarity function, " + }, + { + "bbox": [ + 55, + 369, + 296, + 405 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 369, + 296, + 405 + ], + "type": "text", + "content": " as the temperature parameter, and " + }, + { + "bbox": [ + 55, + 369, + 296, + 405 + ], + "type": "inline_equation", + "content": "[\\cdot ]" + }, + { + "bbox": [ + 55, + 369, + 296, + 405 + ], + "type": "text", + "content": " denotes concatenation. Details on how " + }, + { + "bbox": [ + 55, + 369, + 296, + 405 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_k" + }, + { + "bbox": [ + 55, + 369, + 296, + 405 + ], + "type": "text", + "content": " is computed are provided in Sec. A.1 of the Supp. Mat." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 406, + 296, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 406, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 406, + 296, + 441 + ], + "type": "text", + "content": "In addition to the " + }, + { + "bbox": [ + 55, + 406, + 296, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ce}}" + }, + { + "bbox": [ + 55, + 406, + 296, + 441 + ], + "type": "text", + "content": ", we employ a regularization constraint to ensure consistency between two augmented views of an image. We define the consistency loss " + }, + { + "bbox": [ + 55, + 406, + 296, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{con}}" + }, + { + "bbox": [ + 55, + 406, + 296, + 441 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 111, + 451, + 295, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 451, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 111, + 451, + 295, + 464 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c o n}} = 1 - \\cos \\left(\\mathcal {E} _ {v} (\\mathbf {I}), \\mathcal {E} _ {v} \\left(\\mathbf {x} ^ {\\prime}\\right)\\right), \\tag {5}", + "image_path": "db48142a235a13ea53e4403e5045185506167a1a3a50dc43bb79ea96a2d0bb2a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 473, + 295, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 295, + 498 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 473, + 295, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{x}'" + }, + { + "bbox": [ + 55, + 473, + 295, + 498 + ], + "type": "text", + "content": " is the augmented version of " + }, + { + "bbox": [ + 55, + 473, + 295, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 473, + 295, + 498 + ], + "type": "text", + "content": ". Finally, the total loss function is given as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 127, + 501, + 295, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 501, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 127, + 501, + 295, + 514 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\mathrm {c e}} + \\alpha \\cdot \\mathcal {L} _ {\\mathrm {c o n}}, \\tag {6}", + "image_path": "102a90f451762a01bfab2f4eed2393c16b765ee4fce4c335366b2423f5780c83.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 525, + 295, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 295, + 549 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 295, + 549 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 525, + 295, + 549 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 525, + 295, + 549 + ], + "type": "text", + "content": " is a balancing coefficient for the regularization term, ensuring both alignment and consistency." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 555, + 286, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 286, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 286, + 567 + ], + "type": "text", + "content": "3.3.3. Client-side training and server-side aggregation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 570, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 689 + ], + "type": "text", + "content": "In our approach, multiple remote clients participate in distinct image classification tasks, collaboratively training the PromptFormer network, parameterized by " + }, + { + "bbox": [ + 55, + 570, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\rho = \\{\\Theta\\}" + }, + { + "bbox": [ + 55, + 570, + 295, + 689 + ], + "type": "text", + "content": ". In each communication round, the client transmits " + }, + { + "bbox": [ + 55, + 570, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 55, + 570, + 295, + 689 + ], + "type": "text", + "content": " to the server, which is aggregated at the server. Through training on a diverse set of classification tasks distributed across clients, and aggregating the parameters at the server, the resulting model achieves generalization capability to classes and domains unseen during training. Pseudo-code of FedMVP can be found in Sec. A.2 of the Supp. Mat." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Initially, the server randomly initializes the parameters " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\rho^0" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " and in each communication round only a random subset" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "of remote clients " + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "inline_equation", + "content": "S^r" + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": " will be chosen by the server for sharing the updated parameters " + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": ". Each communication round " + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": " involves the following steps:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 108, + 553, + 168 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 313, + 108, + 553, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 553, + 132 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 553, + 132 + ], + "type": "text", + "content": "(i) Model initialization: Each client " + }, + { + "bbox": [ + 313, + 108, + 553, + 132 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 108, + 553, + 132 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 108, + 553, + 132 + ], + "type": "inline_equation", + "content": "S_{r}" + }, + { + "bbox": [ + 313, + 108, + 553, + 132 + ], + "type": "text", + "content": " receives the latest model parameters " + }, + { + "bbox": [ + 313, + 108, + 553, + 132 + ], + "type": "inline_equation", + "content": "\\rho^r" + }, + { + "bbox": [ + 313, + 108, + 553, + 132 + ], + "type": "text", + "content": " to configure its local model." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 133, + 553, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 553, + 168 + ], + "type": "text", + "content": "(ii) Local optimization: Following Eq. (6), client " + }, + { + "bbox": [ + 313, + 133, + 553, + 168 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 133, + 553, + 168 + ], + "type": "text", + "content": " optimizes its local parameters " + }, + { + "bbox": [ + 313, + 133, + 553, + 168 + ], + "type": "inline_equation", + "content": "\\rho^r" + }, + { + "bbox": [ + 313, + 133, + 553, + 168 + ], + "type": "text", + "content": " using an SGD optimizer over " + }, + { + "bbox": [ + 313, + 133, + 553, + 168 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 313, + 133, + 553, + 168 + ], + "type": "text", + "content": " iterations with learning rate " + }, + { + "bbox": [ + 313, + 133, + 553, + 168 + ], + "type": "inline_equation", + "content": "\\eta^r" + }, + { + "bbox": [ + 313, + 133, + 553, + 168 + ], + "type": "text", + "content": ". Updated parameters:" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 373, + 174, + 553, + 189 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 174, + 553, + 189 + ], + "spans": [ + { + "bbox": [ + 373, + 174, + 553, + 189 + ], + "type": "interline_equation", + "content": "\\rho_ {i} ^ {r + 1} = \\operatorname {S G D} _ {J} \\left(\\eta^ {r}, \\rho^ {r}, \\mathcal {C} _ {i}, \\mathcal {D} _ {i}\\right). \\tag {7}", + "image_path": "093c4a863473f2686cb82cca9fd8b8592a37e04dbf54184badec3e7db9afc616.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 198, + 553, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 553, + 233 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 553, + 233 + ], + "type": "text", + "content": "(iii) Parameter aggregation: After local training, each client " + }, + { + "bbox": [ + 313, + 198, + 553, + 233 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 198, + 553, + 233 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 198, + 553, + 233 + ], + "type": "inline_equation", + "content": "S_{r}" + }, + { + "bbox": [ + 313, + 198, + 553, + 233 + ], + "type": "text", + "content": " transmits its updated parameters " + }, + { + "bbox": [ + 313, + 198, + 553, + 233 + ], + "type": "inline_equation", + "content": "\\rho_{i}^{r + 1}" + }, + { + "bbox": [ + 313, + 198, + 553, + 233 + ], + "type": "text", + "content": " back to the server, where they are aggregated [16] as:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 384, + 240, + 553, + 269 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 240, + 553, + 269 + ], + "spans": [ + { + "bbox": [ + 384, + 240, + 553, + 269 + ], + "type": "interline_equation", + "content": "\\rho^ {r + 1} = \\frac {1}{\\left| S _ {r} \\right|} \\sum_ {i \\in S _ {r}} \\rho_ {i} ^ {r + 1}. \\tag {8}", + "image_path": "c90344c920a11da4b4e6a708d238c6ae739c3de245eb73366ddd586c1f9688ce.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 277, + 553, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 553, + 323 + ], + "type": "text", + "content": "The above steps are repeated for " + }, + { + "bbox": [ + 313, + 277, + 553, + 323 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 277, + 553, + 323 + ], + "type": "text", + "content": " communication rounds, after which the server converges on the final model parameters " + }, + { + "bbox": [ + 313, + 277, + 553, + 323 + ], + "type": "inline_equation", + "content": "\\rho^R" + }, + { + "bbox": [ + 313, + 277, + 553, + 323 + ], + "type": "text", + "content": ". This iterative process facilitates collaborative and distributed optimization of " + }, + { + "bbox": [ + 313, + 277, + 553, + 323 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 313, + 277, + 553, + 323 + ], + "type": "text", + "content": " across client data." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 324, + 555, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 555, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 555, + 516 + ], + "type": "text", + "content": "Lightweight fine-tuning in clients. As the FL setup exhibits high data heterogeneity, different clients possess varying numbers of samples. Thus, treating every client as equal and repeatedly fine-tuning the parameters " + }, + { + "bbox": [ + 313, + 324, + 555, + 516 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 313, + 324, + 555, + 516 + ], + "type": "text", + "content": " can lead to severe overfitting in some clients, especially those with very limited training data. To prevent this, we propose to inject low-rank decomposition matrices, or LoRA [9], to the PromptFormer module, and fine-tune these LoRA matrices if training loss " + }, + { + "bbox": [ + 313, + 324, + 555, + 516 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{total}}" + }, + { + "bbox": [ + 313, + 324, + 555, + 516 + ], + "type": "text", + "content": " in a client starts below a certain loss value " + }, + { + "bbox": [ + 313, + 324, + 555, + 516 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 324, + 555, + 516 + ], + "type": "text", + "content": ", which we set to 0.5. In particular, only the LoRA weights corresponding to the query, key and value matrices of the attention layers are trained, while freezing the parameters " + }, + { + "bbox": [ + 313, + 324, + 555, + 516 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 313, + 324, + 555, + 516 + ], + "type": "text", + "content": ". In that case, the client shares only the LoRA matrices with the server. This not only helps in boosting generalization, but also brings down the communication overhead of clients by around " + }, + { + "bbox": [ + 313, + 324, + 555, + 516 + ], + "type": "inline_equation", + "content": "\\times 267" + }, + { + "bbox": [ + 313, + 324, + 555, + 516 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 526, + 394, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 394, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 394, + 540 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 545, + 553, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 553, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 553, + 665 + ], + "type": "text", + "content": "We evaluate our proposed FedMVP in three distinct settings that are aimed to gauge the (i) generalization to unseen categories (or Base-to-New Generalization) in Sec. 4.1, (ii) generalization to unseen domains (or Domain Generalization) in Sec. 4.2, and (iii) generalization to unseen datasets (or Cross-Dataset Generalization) in Sec. 4.3. Finally, we report the results of our ablation study in Sec. 4.4 that are aimed to assess impact of the design choices made in FedMVP. Due to the lack of space, we report additional experimental results in Sec. B of the Supp. Mat." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": "Benchmarks and datasets. Following FedTPG [21], we have reported results on the Base-to-New Generalization setting that encompasses nine medium-sized image recognition datasets. For the Domain Generalization setting we" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17873" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "have used the ImageNet [3] and DomainBed [6] benchmarks. In addition, we have also included EuroSAT [8] for the Cross-Dataset Generalization setting, alongside all nine datasets of the Base-to-New Generalization setting. In total, we have used 20 datasets for a comprehensive experimental evaluation. Detailed dataset descriptions and statistics can be found in Sec. B.1 of the Supp. Mat." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 157, + 294, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 157, + 294, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 157, + 294, + 264 + ], + "type": "text", + "content": "Baselines. We have compared our method with the state-of-the-art prompt tuning approaches that use the CLIP backbone [22]: KgCoOp [30], CoCoOp [35], VPT [11], and MaPLe [13], which were originally designed for offline and non-FL setup, but adapted to the federated setting, as in FedTPG [21]. Furthermore, we have compared with CLIP-based methods specifically designed to work in the FL setting: FedTPG [21], FedOTP [14], PromptFL [7], and FedCLIP [15]. We also included zero-shot CLIP (ZS-CLIP)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 266, + 295, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 295, + 421 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 295, + 421 + ], + "type": "text", + "content": "Implementation details. We have employed the frozen ViT-B/16 CLIP backbone as the vision encoder, and GPT-4o [1] as the LLM in our experiments. Each MHCA module of " + }, + { + "bbox": [ + 55, + 266, + 295, + 421 + ], + "type": "inline_equation", + "content": "f_{\\Theta}" + }, + { + "bbox": [ + 55, + 266, + 295, + 421 + ], + "type": "text", + "content": ", consists of a 4-head cross-attention layers, complemented by layer normalization and each FFN comprising of a two-layer bottleneck structure (Linear-GeLU-Linear). The query prompt " + }, + { + "bbox": [ + 55, + 266, + 295, + 421 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 55, + 266, + 295, + 421 + ], + "type": "text", + "content": " has a length " + }, + { + "bbox": [ + 55, + 266, + 295, + 421 + ], + "type": "inline_equation", + "content": "m = 4" + }, + { + "bbox": [ + 55, + 266, + 295, + 421 + ], + "type": "text", + "content": " and matches the patch embedding dimension " + }, + { + "bbox": [ + 55, + 266, + 295, + 421 + ], + "type": "inline_equation", + "content": "d_v = 768" + }, + { + "bbox": [ + 55, + 266, + 295, + 421 + ], + "type": "text", + "content": ". The projection layer " + }, + { + "bbox": [ + 55, + 266, + 295, + 421 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{proj}}" + }, + { + "bbox": [ + 55, + 266, + 295, + 421 + ], + "type": "text", + "content": " maps the textual dimension of 512 to the patch embedding dimension via a linear transformation. In Eq. (6), we keep " + }, + { + "bbox": [ + 55, + 266, + 295, + 421 + ], + "type": "inline_equation", + "content": "\\alpha = 10" + }, + { + "bbox": [ + 55, + 266, + 295, + 421 + ], + "type": "text", + "content": ". We have set the batch size to 128, number of shots per class to 8, and the learning rate equal to 0.003 with a decay rate of 1e-5 and momentum of 0.9." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 431, + 209, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 431, + 209, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 431, + 209, + 443 + ], + "type": "text", + "content": "4.1. Base-to-New Generalization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 449, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 640 + ], + "type": "text", + "content": "Experimental setup. To recap, the base-to-new generalization setting encompasses nine different datasets. The setting is characterized by splitting the classes from each dataset into two groups – base and new. The samples of the base classes are used for training, and the samples of the new classes are held out to test the generalization performance [21]. Furthermore, the samples of bases classes from all nine datasets are distributed to the clients, such that each client does not see samples from more than 20 disjoint classes (referred to as local classes). We report the model's classification accuracy on: (i) the client's local classes, (ii) all the base classes (spread across clients), and (iii) the held-out new classes that were never seen by any client. Note that (ii) and (iii) is computed using the aggregated model on the server, and (i) is computed using the model on the client (which is the same server model after convergence)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "Main results. In Table 2 we report the performance in the base-to-new generalization setting, summarized across nine datasets. The results show that the textual and multimodal prompting methods, such as PromptFL, FedKgCoOp, and FedMaPLe, excel on the seen local and base class accuracy, but have at-par or worse generalization on the new" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 354, + 140, + 511, + 255 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 553, + 137 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 553, + 137 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 553, + 137 + ], + "type": "text", + "content": "Table 2. Base-to-New Generalization setting. Accuracies " + }, + { + "bbox": [ + 313, + 71, + 553, + 137 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 313, + 71, + 553, + 137 + ], + "type": "text", + "content": " on client's local (seen) classes, base (seen) classes and new (unseen) classes. Harmonic mean (HM) is computed with Base and New class accuracies. Best number in each column is in bold and second-best underlined. Improvement row reports the absolute difference between the best and second-best numbers." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 354, + 140, + 511, + 255 + ], + "lines": [ + { + "bbox": [ + 354, + 140, + 511, + 255 + ], + "spans": [ + { + "bbox": [ + 354, + 140, + 511, + 255 + ], + "type": "table", + "html": "
MethodsLocalBaseNewHM
ZS-CLIP [22]76.7270.5175.7874.24
FedOTP [14]74.8265.2257.0464.89
FedCoCoOp [35]81.4673.7666.0073.20
FedVPT [11]76.2970.4374.8973.79
FedCLIP [15]76.8771.0475.0674.24
FedMaPLe [13]81.6374.4470.6275.29
FedKgCoOp [30]78.3872.1875.8775.39
PromptFL [7]81.7574.4771.7075.74
FedTPG [21]80.7573.6876.0276.70
FedMVP (Ours)81.8975.3777.8278.27
Improvement+0.14+0.90+1.79+1.57
", + "image_path": "737a9e648ae7bf683d085c56308d6820e164c500361b0ec03f28a72bafa2e4ee.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 264, + 555, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 555, + 420 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 555, + 420 + ], + "type": "text", + "content": "classes than the ZS-CLIP model. This hints at the potential overfitting caused by learning static context from the seen classes. In contrast, our proposed FedMVP that considers multimodal contextual information and dynamic conditioning demonstrates superior performance in both the seen and unseen classes, achieving the highest HM accuracy. While FedTPG also generates prompts based on task-related input, it does not outperform our FedMVP due to lack of comprehensive visual and attribute information in the prompts. In summary, FedMVP surpasses the state-of-the-art methods by a notable margin of " + }, + { + "bbox": [ + 313, + 264, + 555, + 420 + ], + "type": "inline_equation", + "content": "+1.57\\%" + }, + { + "bbox": [ + 313, + 264, + 555, + 420 + ], + "type": "text", + "content": " in the base-to-new generalization setting. Detailed breakdown for each of the nine datasets is reported in Tab. A.4 in the Supp. Mat." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 430, + 446, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 446, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 446, + 441 + ], + "type": "text", + "content": "4.2. Domain Generalization" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 449, + 553, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 553, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 553, + 640 + ], + "type": "text", + "content": "Experimental setup. We conduct experiments under two distinct domain generalization (DG) setups: Multi-source Single-target (MSST) and Single-source Multi-target (SSMT). In the MSST DG setting for a given benchmark, we follow a leave-one-domain-out protocol for evaluation, where we hold out one domain for measuring generalization and use the rest of the domains for training the clients. Training classes are split among clients under the constraint that each client can see images from a single domain. This setting mirrors a practical scenario where images taken with a smartphone uniquely reflect the taste of the user or socio-economic biases. In a similar manner, in the SSMT DG setting, we use one domain for training the clients in a distributed manner, while the rest of the domains are used for evaluation. More details on this experimental setup can be found in Sec. B.4 of the Supp. Mat." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "content": "Main results. In Table 3, we report the performance of FedMVP and its competitors in the MSST and SSMT DG settings. In particular, SSMT DG is a more challenging setting than the MSST DG setting, as the clients are trained only on a single domain, and expected to generalize to multiple unseen domains. We observe that ZS-CLIP is a strong base" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17874" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 87, + 106, + 520, + 228 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "text", + "content": "Table 3. Domain Generalization setting on the DomainBed [6] benchmark. Accuracies " + }, + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "text", + "content": " on the unseen target domain(s) for the Multi-source Single-target (MSST) and Single-source Multi-target (SSMT) settings. Best number in each column is in bold and second-best underlined. Improvement row reports the absolute difference between the best and second-best numbers." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 87, + 106, + 520, + 228 + ], + "lines": [ + { + "bbox": [ + 87, + 106, + 520, + 228 + ], + "spans": [ + { + "bbox": [ + 87, + 106, + 520, + 228 + ], + "type": "table", + "html": "
MethodMulti-source Single-targetSingle-source Multi-target
PACSOfficeHomeVLCSTerra Inc.DomainNetAveragePACSOfficeHomeVLCSTerra Inc.DomainNetAverage
ZS-CLIP [22]96.1681.4983.2933.9857.1370.4196.1681.4983.2933.9857.1370.41
FedOTP [14]90.7176.4267.4113.2449.6759.4991.1774.5365.1615.1139.3757.07
FedCoCoOp [35]85.0681.4261.7323.6857.0861.7984.5650.8761.2523.0758.2855.61
FedTPG [21]90.9982.7869.7726.7956.8265.4390.7182.6067.6322.5158.3464.36
PromptFL [7]95.3781.7474.8725.0256.8766.7795.8381.7266.9724.1758.2765.39
FedKgCoOp [30]95.4281.8274.9025.0356.9266.8295.8181.7067.4924.2058.3065.50
FedMaPLe [13]94.5182.0371.7936.3058.8868.7094.2581.4872.0833.5957.6267.80
FedVPT [11]95.3681.7683.1933.6255.9869.9894.7980.9982.4432.2355.6669.22
FedCLIP [15]96.2981.7482.7036.5857.8571.0396.2981.6281.8136.4457.4270.72
FedMVP (Ours)97.2884.1585.1237.3661.1773.0297.1783.8984.6136.9260.5672.63
Improvement+0.99+1.37+1.83+0.78+2.29+1.99+0.88+1.29+1.32+0.48+2.22+1.91
", + "image_path": "afac0969f22d01a46f161a6ba53574ae21c7b84f7298a7298bdff112b78bed58.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 75, + 312, + 272, + 427 + ], + "blocks": [ + { + "bbox": [ + 55, + 232, + 295, + 308 + ], + "lines": [ + { + "bbox": [ + 55, + 232, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 295, + 308 + ], + "type": "text", + "content": "Table 4. Domain Generalization setting on the ImageNet benchmark. Accuracies (\\%) on the seen source domain ImageNet (IN) and unseen target domains: ImageNet-V2 (INV2), ImageNet-Sketch (IN-S), ImageNet-A (IN-A) and ImageNet-R (IN-R). Best number in each column is in bold and second-best underlined. Improvement row reports the absolute difference between the best and second-best numbers." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 75, + 312, + 272, + 427 + ], + "lines": [ + { + "bbox": [ + 75, + 312, + 272, + 427 + ], + "spans": [ + { + "bbox": [ + 75, + 312, + 272, + 427 + ], + "type": "table", + "html": "
MethodSourceTarget
ININV2IN-SIN-AIN-RAverage
ZS-CLIP [22]66.7560.7946.1247.7974.0057.18
FedOTP [14]51.6845.5334.7316.6463.9840.22
FedMaPLe [13]66.9660.6544.6946.2474.6256.55
FedVPT [11]66.9260.3446.4348.0374.5657.34
PromptFL [7]67.8061.5945.6148.7874.4957.62
FedCLIP [15]67.2661.3546.6648.2474.3657.65
FedKCoOp [30]67.5361.6046.6948.3774.7157.84
FedCoCoOp [35]68.5162.2946.9050.3376.4959.00
FedTPG [21]69.5162.9047.6549.9776.3559.22
FedMVP (Ours)70.8763.7250.9351.7677.2360.91
Improvement+1.36+0.82+3.28+1.43+0.74+1.69
", + "image_path": "e874b3ba358e559effbf085b9557308680bb4cb377e03fedaf8a4e6704c9a7cb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 433, + 295, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 433, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 54, + 433, + 295, + 540 + ], + "type": "text", + "content": "line that outperforms all textual and multimodal prompting methods, with the exception of FedCLIP and our proposed FedMVP. This suggests that suboptimal prompt tuning can lead to overfitting on the source training domains, thus hurting generalization. Our FedMVP owing to the use of contextual information learns more transferable representations that generalize better to unseen target domains, thus outperforming all competitor methods in most of the benchmarks. More detailed results in Sec. B.4 of the Supp. Mat." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 540, + 295, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 540, + 295, + 636 + ], + "spans": [ + { + "bbox": [ + 54, + 540, + 295, + 636 + ], + "type": "text", + "content": "In Table. 4 we report the results in the SSMT DG setting on ImageNet benchmark. We observe that FedTPG and FedCoCoOp come close to our FedMVP in terms of average performance, with FedMVP outperforming all baselines by " + }, + { + "bbox": [ + 54, + 540, + 295, + 636 + ], + "type": "inline_equation", + "content": "+1.69\\%" + }, + { + "bbox": [ + 54, + 540, + 295, + 636 + ], + "type": "text", + "content": ". In particular, we believe FedMVP improves the performance on ImageNet-Sketch (IN-S) by a large margin of " + }, + { + "bbox": [ + 54, + 540, + 295, + 636 + ], + "type": "inline_equation", + "content": "+3.28\\%" + }, + { + "bbox": [ + 54, + 540, + 295, + 636 + ], + "type": "text", + "content": " due to the use of attribute information, which remains unchanged between real and sketch images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 640, + 215, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 640, + 215, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 640, + 215, + 651 + ], + "type": "text", + "content": "4.3. Cross-dataset Generalization" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 653, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 653, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 653, + 295, + 713 + ], + "type": "text", + "content": "Experimental setup. In the Cross-Dataset Generalization setting the classes from ImageNet are split among the clients for training, in a disjoint manner, and the resulting network is evaluated on 10 held out datasets not seen during training. In other words, it is a combination of base-to-new" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 233, + 553, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 233, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 313, + 233, + 553, + 255 + ], + "type": "text", + "content": "generalization and DG settings, since it requires generalization to both unseen class and data distributions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 258, + 555, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 555, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 555, + 402 + ], + "type": "text", + "content": "Main results. In Table. 5 we report the results of the Cross-Dataset Generalization setting. This setting is considerably challenging, as the model should generalize to different datasets. We find that the results are consistent with the previous findings, where our proposed FedMVP generalizes better to unseen classes and data distributions, outperforming all the text-based and multimodal-based prompting techniques by " + }, + { + "bbox": [ + 313, + 258, + 555, + 402 + ], + "type": "inline_equation", + "content": "+2.26\\%" + }, + { + "bbox": [ + 313, + 258, + 555, + 402 + ], + "type": "text", + "content": " on average. However, with FedMVP we notice a drop in performance with respect to the ZS-CLIP baseline for some fine-grained datasets, such as OxfordPets and StanfordCars, which is potentially due to overlapping attributes for visually similar categories." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 410, + 408, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 410, + 408, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 408, + 422 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 424, + 554, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 424, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 424, + 554, + 567 + ], + "type": "text", + "content": "Impact of proposed components in FedMVP. In Tab. 6 we ablate each component of FedMVP on the base-to-new and MSST DG settings. We observe that each component contributes positively to the final performance. First, Prompt-Former " + }, + { + "bbox": [ + 313, + 424, + 554, + 567 + ], + "type": "inline_equation", + "content": "f_{\\Theta}" + }, + { + "bbox": [ + 313, + 424, + 554, + 567 + ], + "type": "text", + "content": " plays a major role in combining textual information with visual features, thus improving semantic alignment between the two modalities. Second, adding " + }, + { + "bbox": [ + 313, + 424, + 554, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{con}}" + }, + { + "bbox": [ + 313, + 424, + 554, + 567 + ], + "type": "text", + "content": " (as in Eq. (5)) improves the performance by a small margin. Finally, the impact of LoRA can be observed in the penultimate row, where we fine-tune all the parameters of FedMVP. It shows the dangers of overfitting in the FL setup, as some clients may have limited data." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 568, + 554, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 554, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 554, + 676 + ], + "type": "text", + "content": "Participation Rate of the Clients. In Fig. 3 we vary the participation rate from " + }, + { + "bbox": [ + 313, + 568, + 554, + 676 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 568, + 554, + 676 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 568, + 554, + 676 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 313, + 568, + 554, + 676 + ], + "type": "text", + "content": " and show how the participation rate of the clients impacts performance on both seen and unseen classes and domains. We observe that FedMVP always exhibits a gentle, yet positive slope in both settings, indicating that continuous parameter aggregation improves generalization. In contrast, for some competitor methods, such as PromptFL, increased participation leads to saturated performance or negative slope in the plot." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Computational analysis. In Fig. 4 we plot the accuracy vs. communication rounds for FedMVP and competitors in base-to-new and MSST settings. While FedMVP trans" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17875" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 82, + 106, + 526, + 223 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "text", + "content": "Table 5. Cross-Dataset Generalization setting. Accuracies " + }, + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "text", + "content": " on the seen classes of the source dataset, i.e., ImageNet, and the unseen classes of the target datasets. Best number in each column is in bold and second-best underlined. Improvement row reports the absolute difference between the best and second-best numbers." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 106, + 526, + 223 + ], + "lines": [ + { + "bbox": [ + 82, + 106, + 526, + 223 + ], + "spans": [ + { + "bbox": [ + 82, + 106, + 526, + 223 + ], + "type": "table", + "html": "
MethodSourceTarget
ImageNetCaltech101Flowers102FGVCAircraftUCF101OxfordPetsFood101DTDStanfordCarsSUN397EuroSATAverage
ZS-CLIP [22]66.7592.9071.2924.7266.7589.1586.0944.3365.2962.5947.6865.08
FedOTP [14]51.6888.6048.3610.9546.1073.4361.0141.5529.6148.3450.2649.82
FedMaPLe [13]66.9692.4968.2523.5260.3289.6783.5244.6860.1661.8545.3862.98
PromptFL [7]67.8091.8768.1321.4464.1388.7085.8542.4363.5962.7743.2663.22
FedCLIP [15]67.2693.3967.1924.0365.6488.2885.1444.4465.5063.4241.6763.87
FedKgCoOp [30]67.5393.6369.3123.0664.4688.5585.3744.7464.9963.8543.2964.13
FedVPT [11]68.5691.4269.2421.0965.7087.7985.4544.5765.0064.6847.8264.28
FedCoCoOp [35]68.5194.1166.3420.7962.7589.0485.4043.2063.9864.0255.4064.50
FedTPG [21]69.5193.7570.0423.2264.7290.6085.9146.2563.9866.7847.8665.31
FedMVP (Ours)70.8795.3772.8025.9470.5889.2787.0649.7865.8368.1950.8467.57
Improvement+1.36+1.62+1.51+1.22+3.83-1.33+0.97+3.53+0.33+1.41-4.56+2.26
", + "image_path": "1f081649efc02334f5a1f2abb0feaa9524a68cd0b295b41a4ec940e19607bb3b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 106, + 251, + 245, + 318 + ], + "blocks": [ + { + "bbox": [ + 55, + 226, + 296, + 249 + ], + "lines": [ + { + "bbox": [ + 55, + 226, + 296, + 249 + ], + "spans": [ + { + "bbox": [ + 55, + 226, + 296, + 249 + ], + "type": "text", + "content": "Table 6. Ablating FedMVP components. We report numbers for the Base-to-New and MMST DG (on DomainBed) settings." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 106, + 251, + 245, + 318 + ], + "lines": [ + { + "bbox": [ + 106, + 251, + 245, + 318 + ], + "spans": [ + { + "bbox": [ + 106, + 251, + 245, + 318 + ], + "type": "table", + "html": "
ComponentsBase-to-NewMSST DG
ZS-CLIP74.2470.41
fΘ only75.9471.85
fΘ + Lcon76.2772.14
w/o LoRA77.4172.58
FedMVP (Full)78.2773.02
", + "image_path": "1a6e221c5f65c28c7f4e36273ddf7540337b55e66cbd44c4c08eb08c6c473f74.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 58, + 321, + 171, + 401 + ], + "blocks": [ + { + "bbox": [ + 58, + 321, + 171, + 401 + ], + "lines": [ + { + "bbox": [ + 58, + 321, + 171, + 401 + ], + "spans": [ + { + "bbox": [ + 58, + 321, + 171, + 401 + ], + "type": "image", + "image_path": "46620df8a887974449ef6f18e1c867ef88ecf2192563189d5743674b1f469045.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 87, + 404, + 141, + 413 + ], + "lines": [ + { + "bbox": [ + 87, + 404, + 141, + 413 + ], + "spans": [ + { + "bbox": [ + 87, + 404, + 141, + 413 + ], + "type": "text", + "content": "(a) Base-to-New" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 178, + 320, + 291, + 400 + ], + "blocks": [ + { + "bbox": [ + 178, + 320, + 291, + 400 + ], + "lines": [ + { + "bbox": [ + 178, + 320, + 291, + 400 + ], + "spans": [ + { + "bbox": [ + 178, + 320, + 291, + 400 + ], + "type": "image", + "image_path": "67bbe100bb32c1646d18fefb2e85d755dc3522b07aae680627c4d0d50d22db46.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 403, + 259, + 413 + ], + "lines": [ + { + "bbox": [ + 211, + 403, + 259, + 413 + ], + "spans": [ + { + "bbox": [ + 211, + 403, + 259, + 413 + ], + "type": "text", + "content": "(b) MSST DG" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 418, + 296, + 440 + ], + "lines": [ + { + "bbox": [ + 55, + 418, + 296, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 296, + 440 + ], + "type": "text", + "content": "Figure 3. Sensitivity to participation rate of the clients, on (a) Base-to-New and (b) MSST DG (DomainBed) settings." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 455, + 296, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 296, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 296, + 538 + ], + "type": "text", + "content": "mits " + }, + { + "bbox": [ + 55, + 455, + 296, + 538 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 55, + 455, + 296, + 538 + ], + "type": "text", + "content": " more parameters than FedTPG, it converges nearly " + }, + { + "bbox": [ + 55, + 455, + 296, + 538 + ], + "type": "inline_equation", + "content": "10 \\times" + }, + { + "bbox": [ + 55, + 455, + 296, + 538 + ], + "type": "text", + "content": " faster, requiring fewer rounds. In particular, as training progresses, only the LoRA parameters in PromptFormer are fine-tuned in a client, thus reducing the trainable parameter count by " + }, + { + "bbox": [ + 55, + 455, + 296, + 538 + ], + "type": "inline_equation", + "content": "267 \\times" + }, + { + "bbox": [ + 55, + 455, + 296, + 538 + ], + "type": "text", + "content": ". Overall, FedMVP has lower total computational cost with respect to other methods when normalized by the number of communication rounds." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 58, + 552, + 171, + 632 + ], + "blocks": [ + { + "bbox": [ + 58, + 552, + 171, + 632 + ], + "lines": [ + { + "bbox": [ + 58, + 552, + 171, + 632 + ], + "spans": [ + { + "bbox": [ + 58, + 552, + 171, + 632 + ], + "type": "image", + "image_path": "58af3d4289b3246397c461095f4d89041ff90d39e97ffc30fc0f3a3f561f5a94.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 87, + 635, + 141, + 644 + ], + "lines": [ + { + "bbox": [ + 87, + 635, + 141, + 644 + ], + "spans": [ + { + "bbox": [ + 87, + 635, + 141, + 644 + ], + "type": "text", + "content": "(a) Base-to-New" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 649, + 295, + 671 + ], + "lines": [ + { + "bbox": [ + 55, + 649, + 295, + 671 + ], + "spans": [ + { + "bbox": [ + 55, + 649, + 295, + 671 + ], + "type": "text", + "content": "Figure 4. Computational analysis of the methods on (a) Baseto-New and (b) MSST DG (DomainNet) setting." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 178, + 553, + 291, + 632 + ], + "blocks": [ + { + "bbox": [ + 178, + 553, + 291, + 632 + ], + "lines": [ + { + "bbox": [ + 178, + 553, + 291, + 632 + ], + "spans": [ + { + "bbox": [ + 178, + 553, + 291, + 632 + ], + "type": "image", + "image_path": "167988ae91ba94fdfe0a58fe677573c94a829f001f23902fd6280f66e8c0a4cb.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 635, + 258, + 644 + ], + "lines": [ + { + "bbox": [ + 211, + 635, + 258, + 644 + ], + "spans": [ + { + "bbox": [ + 211, + 635, + 258, + 644 + ], + "type": "text", + "content": "(b) MSST DG" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Size of the clients. In Fig. 5, we vary the number of training classes in each client " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "K = \\{5,10,20\\}" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " while keeping the size of the dataset fixed. In other words, decreasing the" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 228, + 555, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 312 + ], + "type": "text", + "content": "number of classes in each client results in an increase in the total number of clients participating in the federated setup. We observe that for all the baselines a higher heterogeneity (or fewer classes per client) leads to lower performance and vice versa. Nevertheless, FedMVP still exhibits a much superior performance when compared to the baselines, outperforming them while having 4x more clients." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 312, + 554, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 554, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 554, + 348 + ], + "type": "text", + "content": "Additional analyses relating to number of shots, choice of LLMs, hyperparameters, and performance in nonfederated setting are reported in Sec. B.5 of the Supp. Mat." + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 316, + 354, + 430, + 435 + ], + "blocks": [ + { + "bbox": [ + 316, + 354, + 430, + 435 + ], + "lines": [ + { + "bbox": [ + 316, + 354, + 430, + 435 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 430, + 435 + ], + "type": "image", + "image_path": "3d3f7ced71ce3b5d8b3d8e4c47678e50b5baae3d59b6ba9ecde997b81fe49231.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 438, + 400, + 447 + ], + "lines": [ + { + "bbox": [ + 345, + 438, + 400, + 447 + ], + "spans": [ + { + "bbox": [ + 345, + 438, + 400, + 447 + ], + "type": "text", + "content": "(a) Base-to-New" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 452, + 554, + 474 + ], + "lines": [ + { + "bbox": [ + 313, + 452, + 554, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 452, + 554, + 474 + ], + "type": "text", + "content": "Figure 5. Sensitivity to number of classes per client on (a) Baseto-New and (b) MSST DG (DomainNet) setting." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 436, + 354, + 550, + 434 + ], + "blocks": [ + { + "bbox": [ + 436, + 354, + 550, + 434 + ], + "lines": [ + { + "bbox": [ + 436, + 354, + 550, + 434 + ], + "spans": [ + { + "bbox": [ + 436, + 354, + 550, + 434 + ], + "type": "image", + "image_path": "2d05eb353be34e67921fd9206d04b20888df9223d9bdd61e6f974783837ae0ef.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 470, + 437, + 517, + 447 + ], + "lines": [ + { + "bbox": [ + 470, + 437, + 517, + 447 + ], + "spans": [ + { + "bbox": [ + 470, + 437, + 517, + 447 + ], + "type": "text", + "content": "(b) MSST DG" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 483, + 388, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 388, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 388, + 495 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 498, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 616 + ], + "type": "text", + "content": "In this work, we proposed FedMVP, a novel federated prompt tuning framework for CLIP, to address the challenge of reduced generalization to unseen domains and classes. FedMVP combines visual embeddings with textual attribute features to generate multimodal visual prompts, which are tuned using visual prompt tuning. Our findings revealed that the multimodal contextual information in the prompts enhances the generalizability to unseen classes and domains. FedMVP outperforms all its competitors by a notable margin of " + }, + { + "bbox": [ + 313, + 498, + 555, + 616 + ], + "type": "inline_equation", + "content": "+1.57\\% - 2.26\\%" + }, + { + "bbox": [ + 313, + 498, + 555, + 616 + ], + "type": "text", + "content": " across benchmarks." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": "Acknowledgements. This work was supported by: JPNP23019, subsidized by the New Energy and Industrial Technology Development Organization (NEDO); SERICS (PE00000014) under the NRRP MUR program funded by the EU - NGEU; EU project ELIAS (no: 101120237), ANT (no: 101169439), ELLIOT (no: 101214398) and Ministero delle Imprese e del Made in Italy (IPCEI Cloud DM 27 giugno 2022 - IPCEI-CL-0000007)." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17876" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 4, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 146, + 296, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 146, + 296, + 212 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 296, + 212 + ], + "type": "text", + "content": "[2] Sikai Bai, Jie Zhang, Song Guo, Shuaicheng Li, Jingcai Guo, Jun Hou, Tao Han, and Xiaocheng Lu. Diprompt: Disentangled prompt tuning for multiple latent domain generalization in federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27284-27293, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 213, + 295, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 213, + 295, + 257 + ], + "spans": [ + { + "bbox": [ + 62, + 213, + 295, + 257 + ], + "type": "text", + "content": "[3] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255. IEEE, 2009. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 258, + 295, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 295, + 300 + ], + "type": "text", + "content": "[4] Lijie Fan, Dilip Krishnan, Phillip Isola, Dina Katabi, and Yonglong Tian. Improving clip training with language rewrites. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 302, + 294, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 302, + 294, + 356 + ], + "spans": [ + { + "bbox": [ + 62, + 302, + 294, + 356 + ], + "type": "text", + "content": "[5] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2): 581-595, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 357, + 294, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 357, + 294, + 391 + ], + "spans": [ + { + "bbox": [ + 62, + 357, + 294, + 391 + ], + "type": "text", + "content": "[6] Ishaan Gulrajani and David Lopez-Paz. In search of lost domain generalization. In International Conference on Learning Representations, 2021. 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 392, + 294, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 294, + 446 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 294, + 446 + ], + "type": "text", + "content": "[7] Tao Guo, Song Guo, Junxiao Wang, Xueyang Tang, and Wenchao Xu. Promptfl: Let federated participants cooperatively learn prompts instead of models-federated learning in age of foundation model. IEEE Transactions on Mobile Computing, 2023. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 447, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 294, + 502 + ], + "type": "text", + "content": "[8] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 502, + 294, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 294, + 557 + ], + "type": "text", + "content": "[9] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, page 3, 2022. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 558, + 295, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 558, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 57, + 558, + 295, + 623 + ], + "type": "text", + "content": "[10] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 624, + 295, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 669 + ], + "type": "text", + "content": "[11] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[12] Peter Kairouz, H Brendan McMahan, Brendan Avent, Aurélien Bellet, Mehdi Bennis, Arjun Nitin Bhagoji, Kallista Bonawitz, Zachary Charles, Graham Cormode, Rachel Cummings, et al. Advances and open problems in federated learn" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "ing. Foundations and trends® in machine learning, 14(1-2): 1-210, 2021. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 152 + ], + "type": "text", + "content": "[13] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19113-19122, 2023. 2, 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 153, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 153, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 153, + 553, + 207 + ], + "type": "text", + "content": "[14] Hongxia Li, Wei Huang, Jingya Wang, and Ye Shi. Global and local prompts cooperation via optimal transport for federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12151-12161, 2024. 3, 6, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 209, + 553, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 253 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 253 + ], + "type": "text", + "content": "[15] Wang Lu, HU Xixu, Jindong Wang, and Xing Xie. Fedclip: Fast generalization and personalization for clip in federated learning. In ICLR Workshop on Trustworthy and Reliable Large-Scale Machine Learning Models, 2023. 3, 6, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 254, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 308 + ], + "type": "text", + "content": "[16] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 1, 2, 3, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 310, + 553, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 553, + 343 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 553, + 343 + ], + "type": "text", + "content": "[17] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. In International Conference on Learning Representations, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 345, + 553, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 345, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 316, + 345, + 553, + 399 + ], + "type": "text", + "content": "[18] Liliane Momeni, Mathilde Caron, Arsha Nagrani, Andrew Zisserman, and Cordelia Schmid. Verbs in action: Improving verb understanding in video-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15579-15591, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 401, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 401, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 401, + 553, + 477 + ], + "type": "text", + "content": "[19] Khalil Muhammad, Qinqin Wang, Diarmuid O'Reilly-Morgan, Elias Tragos, Barry Smyth, Neil Hurley, James Geraci, and Aonghus Lawlor. Fedfast: Going beyond average for faster training of federated recommender systems. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1234-1242, 2020. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 479, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 479, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 479, + 553, + 533 + ], + "type": "text", + "content": "[20] Sarah Pratt, Ian Covert, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15691-15701, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 535, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 535, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 535, + 553, + 590 + ], + "type": "text", + "content": "[21] Chen Qiu, Xingyu Li, Chaithanya Kumar Mummadi, Madan Ravi Ganesh, Zhenzhen Li, Lu Peng, and Wan-Yi Lin. Federated text-driven prompt generation for vision-language models. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 591, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 591, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 591, + 553, + 657 + ], + "type": "text", + "content": "[22] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "text", + "content": "[23] Jinhyun So, Chaoyang He, Chien-Sheng Yang, Songze Li, Qian Yu, Ramy E Ali, Basak Guler, and Salman Avestimehr. Lightsecagg: a lightweight and versatile design for secure aggregation in federated learning. Proceedings of Machine Learning and Systems, 4:694-720, 2022. 1" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17877" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[24] Guangyu Sun, Matias Mendieta, Aritra Dutta, Xin Li, and Chen Chen. Towards multi-modal transformers in federated learning. In European Conference on Computer Vision, pages 229-246. Springer, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 162 + ], + "type": "text", + "content": "[25] Zhenyu Sun, Xiaochun Niu, and Ermin Wei. Understanding generalization of federated learning via stability: Heterogeneity matters. In International Conference on Artificial Intelligence and Statistics, pages 676-684. PMLR, 2024. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 295, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 295, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 295, + 207 + ], + "type": "text", + "content": "[26] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 274 + ], + "type": "text", + "content": "[27] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, et al. Robust fine-tuning of zero-shot models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7959-7971, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "type": "text", + "content": "[28] Yinghui Xing, Qirui Wu, De Cheng, Shizhou Zhang, Guoqiang Liang, Peng Wang, and Yanning Zhang. Dual modality prompt tuning for vision-language pre-trained model. IEEE Transactions on Multimedia, 26:2056-2068, 2023. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 321, + 295, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 295, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 295, + 386 + ], + "type": "text", + "content": "[29] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language in a bottle: Language model guided concept bottlenecks for interpretable image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19187-19197, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 295, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 295, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 295, + 441 + ], + "type": "text", + "content": "[30] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6757-6767, 2023. 2, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 444, + 295, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 444, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 444, + 295, + 487 + ], + "type": "text", + "content": "[31] Xin Yao, Chaofeng Huang, and Lifeng Sun. Two-stream federated learning: Reduce the communication costs. In 2018 IEEE Visual Communications and Image Processing (VCIP), pages 1-4. IEEE, 2018. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 489, + 295, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 295, + 532 + ], + "type": "text", + "content": "[32] Xin Yao, Tianchi Huang, Chenglei Wu, Rui-Xiao Zhang, and Lifeng Sun. Federated learning with additional mechanisms on clients to reduce communication costs. arXiv preprint arXiv:1908.05891, 2019. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 534, + 295, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 295, + 586 + ], + "type": "text", + "content": "[33] Liping Yi, Gang Wang, Xiaoguang Liu, Zhuan Shi, and Han Yu. Fedgh: Heterogeneous federated learning with generalized global header. In Proceedings of the 31st ACM International Conference on Multimedia, pages 8686-8696, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 590, + 295, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 295, + 634 + ], + "type": "text", + "content": "[34] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11975-11986, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 295, + 689 + ], + "type": "text", + "content": "[35] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "text", + "content": "[36] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. In-" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 334, + 73, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 334, + 73, + 553, + 95 + ], + "type": "text", + "content": "ternational Journal of Computer Vision, 130(9):2337-2348, 2022. 1, 2, 3" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "17878" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/FedMeNF_ Privacy-Preserving Federated Meta-Learning for Neural Fields/97cb26d5-5659-4687-b57e-cc32fd64dcc4_content_list.json b/2025/FedMeNF_ Privacy-Preserving Federated Meta-Learning for Neural Fields/97cb26d5-5659-4687-b57e-cc32fd64dcc4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..13fc40f4a5a8cfaedf232b085fc38aecf8dae47d --- /dev/null +++ b/2025/FedMeNF_ Privacy-Preserving Federated Meta-Learning for Neural Fields/97cb26d5-5659-4687-b57e-cc32fd64dcc4_content_list.json @@ -0,0 +1,2323 @@ +[ + { + "type": "text", + "text": "FedMeNF: Privacy-Preserving Federated Meta-Learning for Neural Fields", + "text_level": 1, + "bbox": [ + 122, + 128, + 875, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Junhyeog Yun Minui Hong Gunhee Kim Seoul National University", + "bbox": [ + 316, + 180, + 679, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{junhyeog,alsdml123,gunhee}@snu.ac.kr", + "bbox": [ + 326, + 218, + 671, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/junhyeog/FedMeNF", + "bbox": [ + 339, + 236, + 651, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 328, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural fields provide a memory-efficient representation of data, which can effectively handle diverse modalities and large-scale data. However, learning to map neural fields often requires large amounts of training data and computations, which can be limited to resource-constrained edge devices. One approach to tackle this limitation is to leverage Federated Meta-Learning (FML), but traditional FML approaches suffer from privacy leakage. To address these issues, we introduce a novel FML approach called FedMeNF. FedMeNF utilizes a new privacy-preserving loss function that regulates privacy leakage in the local meta-optimization. This enables the local meta-learner to optimize quickly and efficiently without retaining the client's private data. Our experiments demonstrate that FedMeNF achieves fast optimization speed and robust reconstruction performance, even with few-shot or non-IID data across diverse data modalities, while preserving client data privacy.", + "bbox": [ + 88, + 316, + 485, + 575 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 603, + 220, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural Fields (NFs), also known as Implicit Neural Representations (INRs), use a deep neural network to approximate continuous signals [12, 46, 47, 57, 60]. They enable memory-efficient representation and fast processing of data [14]. Thanks to this, NFs effectively handle both large-scale data and diverse modalities, including audio, images, videos, and 3D objects [10, 11, 33, 40, 45, 48, 52, 53, 71].", + "bbox": [ + 89, + 628, + 482, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, optimizing NFs is often slow and requires large amounts of data, which can be a serious issue on resource-constrained edge devices. Consider a practical scenario: a user takes a few photos of an object using their phone and wants to get a high-fidelity 3D model of it. Training an NF from scratch with such limited data and resources would be time-consuming and likely yield poor results. One approach to address this issue is meta-learning [19, 44], which aims to train a meta-learner that can quickly adapt to new tasks with only a few samples. Using the meta-learner, the user can quickly optimize a high-performing NF for the", + "bbox": [ + 89, + 734, + 482, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "new object with just a few images.", + "bbox": [ + 511, + 287, + 743, + 303 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nevertheless, meta-learning itself requires a large and diverse set of tasks to learn an effective initialization or optimization strategy. Unfortunately, in many practical scenarios, users only have data for one or a few tasks. For example, most users typically take photos of only a single instance of privacy-sensitive objects such as their face, body, pet, or car. To enable meta-learning in such scenarios, we can leverage Federated Meta-Learning (FML) [8, 18, 29], which collaboratively trains a global meta-learner using private data from numerous distributed clients. FML builds upon Federated Learning (FL) [4, 13, 27, 36, 41], a paradigm designed to train a global model while ensuring client data remains on local devices to preserve privacy.", + "bbox": [ + 511, + 305, + 906, + 501 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yet, applying FML to NFs introduces a critical challenge that fundamentally undermines the privacy guarantee of FL. The privacy protection in traditional FL relies on the assumption that sharing trained model parameters, rather than the raw data itself, is safe. This assumption completely breaks down for NFs for two main reasons. First, in our scenario, each client has data for only a single or a few task instances (e.g., a single image or 3D object). If even one client has only a single task instance, the local meta-optimization is like standard supervised training on that specific task [26]. Consequently, the trained local meta-learner functions as an optimized NF for that private data. Second, NFs are, in essence, a compressed representation of the data and inherently encapsulate it. Thus, even if only the NF parameters are shared, a malicious server can reconstruct the private data. Due to this inherent privacy leakage issue, existing studies often limit FL applications for NFs to public data such as landmarks [59] or cityscapes [24, 55, 68], where privacy is less of a concern. For these reasons, naively applying FML methods to private data results in severe privacy issues, as the server can reconstruct the private data from the shared local meta-learner.", + "bbox": [ + 511, + 505, + 908, + 837 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address this critical challenge, we propose FedMeNF (Federated Meta-Learning for Neural Fields), a novel FML framework designed to collaboratively train a global NF meta-learner while preserving privacy as overviewed in", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 805, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2161", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0f30597defe4686a7a2bfcaad6c3c63efe9343e25e9f8260a02cd3d6c3d69952.jpg", + "image_caption": [ + "(a) FML with a single task per client." + ], + "image_footnote": [], + "bbox": [ + 94, + 87, + 321, + 358 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8888ecd7e33472cc1c28fe1cdf03739c13a8500a76df8f359b1d4f398a5d83f2.jpg", + "image_caption": [ + "(b) Comparison of FedMeNF with Local and FML in terms of fast or few-shot optimization on a new task, and privacy preservation." + ], + "image_footnote": [], + "bbox": [ + 375, + 87, + 620, + 159 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7be94016b2bbba070827b47748f3e92025c57419bf10938eaebd3854450c20bc.jpg", + "image_caption": [ + "(c) The client has photos of only one pet, car, body, or face, and aims to optimize a neural field for each instance." + ], + "image_footnote": [], + "bbox": [ + 390, + 210, + 607, + 344 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c582a7d216a6f056675d0688f116fbff6e3783c3f9bca8835688f5337e990bb6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 658, + 88, + 885, + 357 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4f24c78639b3839c85ef30ca23fd2f9c6285513cbc897f1be62349485ec3d24f.jpg", + "image_caption": [ + "(d) FedMeNF with a single task per client.", + "Figure 1. Overview of our approach. In each communication round, the server sends the global meta-learner $\\theta$ to clients, each of which uses its private data to train a local meta-learner $w$ based on the global meta-learner. The trained local meta-learners are then sent back to the server to update the global meta-learner. (a) In existing federated meta-learning (FML) approaches, the local meta-learner contains privacy-sensitive information when a client has only a single task, as shown in (c). The server can infer the client's private data from the shared local meta-learner. (d) In contrast, our FedMeNF ensures that the local meta-learner only learns how to efficiently optimize NFs while preserving privacy-sensitive details. (b, c) Local refers to using only the client's data without sharing any information with the server." + ], + "image_footnote": [], + "bbox": [ + 145, + 376, + 851, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fig. 1, which then empowers each client to rapidly optimize a high-performance NF even with a small amount of new private data. FedMeNF can be utilized in various privacy-sensitive applications. For example, suppose that a user buys a new car and takes a few pictures of it. We can then train the NF of the car with this small set of images using a global meta-learner trained with other car images in other devices. As another example, with only a handful of facial images, users can employ the global facial NF meta-learner to generate images showing a range of facial expressions. Also, they can upscale low-resolution photos to high-resolution or generate higher frame-rate videos from personal data, such as pet photos or golf swing videos.", + "bbox": [ + 88, + 523, + 482, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To quantify and control privacy leakage, we define a privacy metric, $\\mathrm{PSNR}_p$ . We theoretically analyze how $\\mathrm{PSNR}_p$ increases under existing FML approaches. By regularizing the increase of $\\mathrm{PSNR}_p$ , FedMeNF forces the local meta-learner to learn only how to optimize the NF faster with limited data without memorizing the private data. We demonstrate that FedMeNF achieves performance comparable to state-of-the-art methods while better preserving privacy across diverse data modalities, including images, videos, and neural radiance fields (NeRFs). Additionally, FedMeNF shows robust performance even when local data are very few or highly diverse among clients.", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be outlined as follows:", + "bbox": [ + 532, + 523, + 831, + 537 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. This is the first study on FL for NFs on private data.", + "2. We theoretically and empirically show how privacy leakage occurs during the FML for NFs.", + "3. We propose FedMeNF that preserves the privacy of local data with minimal impact on optimization speed and reconstruction quality.", + "4. We conduct comprehensive experiments on FedMeNF across various data modalities, private data sizes, and levels of data diversity, outperforming baseline methods." + ], + "bbox": [ + 511, + 541, + 905, + 678 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Preliminaries", + "text_level": 1, + "bbox": [ + 513, + 693, + 651, + 709 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Coordinate-based Neural Fields. Coordinate-based neural fields [3, 9, 14, 43, 51, 56, 57] represent continuous functions that map spatial coordinates to signal values such as color or density. The objective for each client $m$ is to find the optimal NF parameter $\\varphi_{i}^{m}$ that best represents its $i$ -th local task data $T_{i}^{m}$ . This is formulated as", + "bbox": [ + 511, + 719, + 906, + 811 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\varphi_ {i} ^ {m} = \\arg \\min _ {\\varphi} \\sum_ {(\\mathbf {x} _ {j}, \\mathbf {y} _ {j}) \\in T _ {i} ^ {m}} L \\left(f _ {\\varphi} (\\mathbf {x} _ {j}), \\mathbf {y} _ {j}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 821, + 906, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $f_{\\varphi}$ is the forward map parameterized by $\\varphi$ , $\\mathbf{x}_j$ represents the input coordinates (e.g., spatial positions), and $\\mathbf{y}_j$", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2162", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "denotes the target values (e.g., pixel colors). $L$ is a suitable loss function such as mean squared error (MSE).", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Federated Learning. We consider a centralized federated learning (FL) environment [2, 41], where clients do not share their data with either the server or other clients and communicate the model parameters exclusively with the centralized server. The centralized FL aims to optimize the global model that minimizes a given objective function as", + "bbox": [ + 89, + 121, + 483, + 212 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\arg \\min _ {\\theta} \\sum_ {m = 0} ^ {M - 1} \\alpha^ {m} L (\\theta , D ^ {m}), \\text {w h e r e} \\alpha^ {m} = \\frac {\\left| D ^ {m} \\right|}{\\Sigma_ {i} \\left| D ^ {i} \\right|}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 220, + 483, + 262 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\theta$ is the parameter of the global meta-learner and $M$ is the number of clients. $D^{m}$ is the local dataset of client $m$ . $\\alpha^{m}$ is the proportion of the number of data samples of client $m$ . $L(\\theta, D^{m})$ represents the loss of model $\\theta$ on a dataset $D^{m}$ . This objective ensures that the global model $\\theta$ is optimized across all clients by weighting the contribution of each client based on its local dataset size.", + "bbox": [ + 89, + 268, + 483, + 375 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Federated Meta-learning. Federated meta-learning [7, 8, 18, 29] is a variant of FL where the client trains the local model using a meta-learning approach with the objective:", + "bbox": [ + 89, + 375, + 483, + 421 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\arg \\min _ {\\theta} \\sum_ {m = 0} ^ {M - 1} \\alpha^ {m} L \\left(\\varphi^ {m}, Q ^ {m}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 441, + 482, + 483 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\text {w h e r e} \\varphi^ {m} = \\arg \\min _ {\\theta} L (\\theta , S ^ {m}). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 484, + 480, + 507 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$S^m$ and $Q^m$ are the support set and query set of the client $m$ . Each client $m$ first trains the local model $\\varphi^m$ on the support set $S^m$ (inner loop). Then, starting from the global model $\\theta$ , the local meta-learner minimizes the loss of $\\varphi^m$ on the query set $Q^m$ (outer loop). This local meta-optimization enables the local meta-learner to adapt quickly to new tasks. The trained local meta-learner is shared with the server, which aggregates the local meta-learners received from participating clients to update the global meta-learner.", + "bbox": [ + 89, + 515, + 483, + 651 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Differential Privacy. Differential privacy (DP) is one of the most popular privacy frameworks with rich theoretical guarantees [1, 5, 15, 23, 35, 63]. It guarantees that when a single entry in the database changes, the change in the probability distribution of the algorithm outputs is bounded by predetermined factors $\\epsilon$ and $\\delta$ [16]. The privacy risk can be regulated by controlling these factors. However, traditional DP methods are designed for large datasets; yet neural fields often capture just a single signal. This causes ambiguity in defining privacy leakage for individual training samples, as well as a significant trade-off between privacy preservation and reconstruction performance.", + "bbox": [ + 89, + 651, + 483, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Problem Statement", + "text_level": 1, + "bbox": [ + 89, + 844, + 277, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we assume a scenario where multiple users collaboratively train a global neural field meta-learner using", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "federated meta-learning. Each user has data about a personal object stored locally on their device. For instance, each user holds a few images of his or her own car on their device. Our goal is to optimize neural fields despite limited data and low computational resources on local devices. Thus, optimizing a specific neural field becomes a local task. The local task data are divided into a support set and a query set, each containing images of the object. During training, the support set is used to train the neural field, while the query set is utilized to optimize the local meta-learner. In the testing phase, the support set enables test-time optimization, and the query set is used to assess its performance. The detailed FML process is described in Fig. 1a and Algorithm 1, excluding the highlighted parts which will be discussed in Sec. 5.", + "bbox": [ + 511, + 90, + 906, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. First-order Approximation of Meta-gradients", + "text_level": 1, + "bbox": [ + 511, + 325, + 898, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During the local meta-optimization, the local meta-learner $w$ is updated for each outer step $i$ , where task data $T_{i} = \\{S_{i}, Q_{i}\\}$ are sampled from the client's training dataset $D_{\\mathrm{train}}$ . Then, the local NF $\\varphi$ , initialized by $w_{i}$ , are optimized with the support set $S_{i}$ (Eq. (8)), and the local meta-learner $w_{i}$ are optimized with the query set $Q_{i}$ (Eq. (9)).", + "bbox": [ + 511, + 347, + 906, + 439 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Proposition 1. Let $g_{M} = \\nabla_{w_{i}}L(\\varphi_{K},B_{K})$ and $g_{k} = \\nabla_{\\varphi_{0}}L(\\varphi_{0},B_{k})$ . Then, the first-order approximation of the meta-gradient $g_{M}$ becomes", + "bbox": [ + 511, + 444, + 906, + 491 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ng _ {M} \\approx g _ {K} - \\lambda_ {i} \\mathcal {I} _ {K}, w h e r e \\mathcal {I} _ {K} = \\sum_ {k = 0} ^ {K - 1} \\nabla_ {\\varphi_ {0}} \\left\\langle g _ {K}, g _ {k} \\right\\rangle . \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 497, + 905, + 539 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$K$ is the # of inner steps, $B_{0\\dots K - 1}$ are minibatches of $S_{i}$ , $B_{K}$ is a minibatch of $Q_{i}$ , and $\\lambda_{i}$ is an inner loop learning rate. The proof is provided in Appendix H.", + "bbox": [ + 511, + 544, + 905, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Fast Optimization via $\\mathcal{I}_K$", + "text_level": 1, + "bbox": [ + 511, + 598, + 741, + 614 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "According to Proposition 1, the meta-gradient $g_{M}$ can be expressed with $g_{K}$ and the inner product term $\\mathcal{I}_K$ , which is the sum of inner products between $g_{K}$ and $g_{0},\\dots ,g_{K - 1}$ . This term maximizes (i) the alignment of the gradients from the query set $(g_{K})$ with those from the support set $(g_0,\\dots ,g_{K - 1})$ , and (ii) the alignment of the gradients computed on different minibatches within the support set. This enables the local meta-learner (i) to be more generalized and (ii) to quickly optimize the neural field for a new task.", + "bbox": [ + 511, + 619, + 906, + 756 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Privacy Leakage in Local Meta-Learning", + "text_level": 1, + "bbox": [ + 511, + 767, + 883, + 785 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We define $\\mathrm{PSNR}_p$ , an intuitive privacy metric to quantify the privacy leakage in FML for neural fields:", + "bbox": [ + 511, + 792, + 905, + 823 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {P S N R} _ {p} = \\mathrm {P S N R} \\left(Q ^ {m}, f _ {w} \\left(\\operatorname {C o o r d} \\left(Q ^ {m}\\right)\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 830, + 906, + 849 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $Q^m$ is the query set of client $m$ , $\\operatorname{Coord}(Q^m)$ is the coordinates of $Q^m$ , and $f_w$ is the forward map parameterized by the local meta-learner $w$ . $\\mathrm{PSNR}_p$ assesses how well", + "bbox": [ + 511, + 854, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2163", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the server can reconstruct $Q^{m}$ with the shared local metalearner $w$ . It is a simple variant of PSNR that measures similarity between images reconstructed with $w$ on the server and the client's local images. A better reconstruction indicates a greater violation of the client's data privacy, meaning that a larger $\\mathrm{PSNR}_p$ indicates more privacy leakage.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We define our main privacy metric based on PSNR for two reasons. First, neural fields can handle various types of signals such as vision, speech, or sensor data, so does the PSNR. Second, PSNR is one of the most popular metrics for evaluating reconstruction attacks in FL [21, 25, 66], as well as reconstruction tasks [9, 31]. As variants, we also define $\\mathrm{SSIM}_p$ and $\\mathrm{LPIPS}_p$ as privacy metrics to assess structural and perceptual similarities [62, 67]; it can overcome the limitation of $\\mathrm{PSNR}_p$ that is based on only pixel-wise MSE. Please refer to Appendix F for more details.", + "bbox": [ + 89, + 181, + 483, + 333 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Privacy Leakage via $g_{K}$", + "text_level": 1, + "bbox": [ + 89, + 344, + 305, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since $\\mathrm{PSNR}_p$ is defined by $10\\log_{10}(R / L(w,Q^m))$ where $R$ is a constant, it becomes larger as the MSE loss $L(w,Q^{m})$ is smaller. We analyze how $L(w,Q^{m})$ changes during the client's local meta-optimization, which helps understand how $\\mathrm{PSNR}_p$ evolves.", + "bbox": [ + 89, + 364, + 483, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proposition 2. Let $\\Delta L_{i + 1} = L(w_{i + 1},B_K) - L(w_i,B_K)$ . The first-order approximation of $\\Delta L_{i + 1}$ is then", + "bbox": [ + 89, + 452, + 480, + 484 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta L _ {i + 1} \\approx - \\lambda_ {o} \\cdot g _ {K} ^ {2} \\leq 0. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 494, + 480, + 512 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proof is provided in Appendix I. Note that $Q^m$ is replaced by $B_K$ , a minibatch of the $Q^m$ . $\\Delta L_{i+1}$ denotes the change in the loss of the local meta-learner $w$ on the client's query set. According to Proposition 2, $\\Delta L_{i+1}$ is always negative, which means that $L(w, Q^m)$ decreases with each outer loop iteration. Since $\\mathrm{PSNR}_p$ is inversely proportional to $L(w, Q^m)$ , $\\mathrm{PSNR}_p$ increases every outer loop iteration. Consequently, to prevent the increase of $\\mathrm{PSNR}_p$ , we regularize $\\Delta L_{i+1}$ to be close to zero, thereby mitigating the privacy leakage in the local meta-optimization.", + "bbox": [ + 89, + 523, + 483, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. Approach", + "text_level": 1, + "bbox": [ + 89, + 690, + 197, + 708 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We reveal that federated meta-learning (FML) approaches suffer from privacy leakage as their local meta-gradients increase the privacy metric. To address this issue, we propose FedMeNF, a novel FML framework for neural fields that ensures privacy-preserving local meta-optimization. FedMeNF employs a privacy-preserving loss function $L_{pp}$ , designed to regulate the increase in the privacy metric. Please refer to Algorithm 1 and Fig. 1d for its workflow.", + "bbox": [ + 89, + 715, + 483, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.1. The Privacy-Preserving Loss", + "text_level": 1, + "bbox": [ + 89, + 848, + 348, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proposition 1-2 theoretically show how client data privacy is leaked in federated meta-learning. Now, we introduce", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 FedMeNF (see Appendix A for notation)", + "bbox": [ + 514, + 90, + 875, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Input: (1) $R$ : # of communication rounds, (2) $N$ : # of clients, (3) $M$ : # of participants, (4) $\\lambda_{i}$ : inner loop learning rate, (5) $\\lambda_{o}$ : outer loop learning rate, (6) $\\alpha^{m}$ : weight proportional to the client $m$ 's dataset size (see Eq. (2)).", + "bbox": [ + 527, + 109, + 906, + 172 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Initialize the global meta-learner $\\theta_0$", + "bbox": [ + 529, + 174, + 766, + 189 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for each communication round $r = 0,\\dots,R - 1$ do", + "bbox": [ + 529, + 189, + 869, + 203 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sample $M$ clients from 1,..., $N$ clients", + "bbox": [ + 553, + 204, + 813, + 219 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for $m = 1,\\dots,M$ do", + "bbox": [ + 555, + 220, + 694, + 233 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw _ {*} ^ {m} \\leftarrow \\text {L O C A L U P D A T E} \\left(\\theta_ {r}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 234, + 764, + 250 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "end for", + "bbox": [ + 555, + 251, + 609, + 263 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {r + 1} \\leftarrow \\Sigma_ {m} \\alpha^ {m} w _ {*} ^ {m} \\quad \\triangleright \\text {S e r v e r a g g r e g a t i o n}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 265, + 903, + 280 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "end for", + "bbox": [ + 531, + 281, + 584, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "procedure LOCALUPDATE(θ)", + "bbox": [ + 531, + 297, + 735, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Initialize the local meta-learner $w_0 \\gets \\theta$", + "bbox": [ + 555, + 314, + 818, + 328 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for each step $i = 0,\\dots,E - 1$ do $\\triangleright$ Outer loop", + "bbox": [ + 555, + 329, + 903, + 344 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sample task data $T_{i} = \\{S_{i},Q_{i}\\}$ from $D_{\\mathrm{train}}$", + "bbox": [ + 580, + 344, + 866, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\varphi_0 \\gets w_i$ Initialize the neural field for $T_{i}$", + "bbox": [ + 580, + 359, + 903, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for each step $k = 0,\\dots,K - 1$ do $\\triangleright$ Inner loop", + "bbox": [ + 580, + 375, + 903, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sample batch $B_{k}$ from the support set $S_{i}$", + "bbox": [ + 604, + 388, + 874, + 404 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\varphi_ {k + 1} \\leftarrow \\varphi_ {k} - \\lambda_ {i} \\nabla_ {\\varphi_ {k}} L \\left(\\varphi_ {k}, B _ {k}\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 419, + 903, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "end for", + "bbox": [ + 578, + 444, + 633, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sample batch $B_{K}$ from the query set $Q_{i}$", + "bbox": [ + 578, + 458, + 844, + 473 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Select the regularization coefficient $\\gamma$", + "bbox": [ + 578, + 474, + 828, + 488 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw _ {i + 1} \\leftarrow w _ {i} - \\lambda_ {o} \\nabla_ {w _ {i}} \\left(L \\left(\\varphi_ {K}, B _ {K}\\right) - \\gamma L \\left(w _ {i}, B _ {K}\\right)\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 489, + 903, + 506 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mid \\text {f o r} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 506, + 903, + 525 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "end for", + "bbox": [ + 555, + 518, + 609, + 529 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Send $w_{E}$", + "bbox": [ + 555, + 532, + 617, + 546 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "end procedure", + "bbox": [ + 531, + 547, + 635, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "a privacy-preserving loss $L_{pp}$ as follows, ensuring that it regularizes the $g_{K}$ term that increases $\\mathrm{PSNR}_p$ :", + "bbox": [ + 511, + 593, + 906, + 625 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {p p} (\\gamma , w _ {i}, \\varphi_ {K}, B _ {K}) = L \\left(\\varphi_ {K}, B _ {K}\\right) - \\gamma L \\left(w _ {i}, B _ {K}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 635, + 903, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\gamma$ is a regularization coefficient that determines the portion of the $g_{K}$ . In Proposition 3 below, we can obtain the privacy-preserving local meta-gradient $g_{pp}$ using $L_{pp}$ .", + "bbox": [ + 511, + 662, + 905, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proposition 3. Let $g_{pp} = \\nabla_{w_i}L_{pp}(\\gamma, w_i, \\varphi_K, B_K)$ . Then, the first-order approximation of $g_{pp}$ and $\\Delta L_{i+1}$ become", + "bbox": [ + 511, + 718, + 903, + 750 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ng _ {p p} \\approx (1 - \\gamma) \\cdot g _ {K} - \\lambda_ {i} \\mathcal {I} _ {K}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 619, + 762, + 903, + 777 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta L _ {i + 1} \\approx - \\lambda_ {o} (1 - \\gamma) \\left(g _ {K}\\right) ^ {2} \\leq 0. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 780, + 903, + 797 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note $\\mathcal{I}_K$ is defined in Eq. (5). At $\\gamma = 1$ , the $g_K$ term, which causes privacy leakage, is completely removed from $g_{pp}$ . This means that the local meta-learner no longer contains information about the client's local dataset and learns only fast optimization by maximizing the inner products between gradients across minibatches. Each client can choose", + "bbox": [ + 511, + 810, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2164", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\bullet$ local meta-learner", + "bbox": [ + 102, + 95, + 228, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$E$ : # of outer loop steps", + "bbox": [ + 106, + 109, + 241, + 122 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$= = \\gg$ $\\nabla_{w_i}L(\\varphi_K,B_K)$", + "bbox": [ + 102, + 125, + 227, + 138 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\longrightarrow \\nabla_{\\varphi_k}L(\\varphi_k,B_k)$", + "bbox": [ + 102, + 142, + 222, + 155 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$\\star$ optimal neural field", + "i: outer loop step", + "- $\\triangleright -\\gamma \\nabla_{w_i}L(w_i,B_K)$", + "$\\Rightarrow \\nabla_{w_i}L_{pp}(\\gamma ,w_i,\\varphi_K,B_K)$" + ], + "bbox": [ + 292, + 94, + 467, + 155 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c3fd9092c10ec35fac0117b7357c7502aee7e931e6f9f90a978305c8e1e3d0a9.jpg", + "image_caption": [ + "(a) $i = 0$" + ], + "image_footnote": [], + "bbox": [ + 99, + 208, + 223, + 275 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/37f204dc2ac283e8215ef575a521f8a84a48ac47afaf0bc4e0b876f38b8c2f1e.jpg", + "image_caption": [ + "(b) $i = 1$" + ], + "image_footnote": [], + "bbox": [ + 230, + 195, + 352, + 275 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/05e671ad1b7b490da56a8027cb584a26696148948f55ca2d834b81290292eee3.jpg", + "image_caption": [ + "(c) $i = E - 1$", + "Figure 2. Illustration of FedMeNF's local meta-optimization in a client. The local meta-learner $w$ does not get closer to the optimal neural field and maintains the initial gap (the distance between the global meta-learner $\\theta$ ( $= w_0$ ) and the optimal neural field). Nevertheless, as meta-optimization progresses, gradients are being aligned, and the local meta-learner can find $w_E$ from which the optimal neural field is reached with a few gradient steps." + ], + "image_footnote": [], + "bbox": [ + 367, + 167, + 482, + 275 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "a different $\\gamma$ that aligns with its desired privacy guarantee. We will empirically show that as $\\gamma$ increases, the privacy leakage from the client's local dataset decreases in Sec. 6.2.", + "bbox": [ + 89, + 425, + 482, + 469 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "There are two ways to set $\\gamma$ : as a fixed hyperparameter or adjusted dynamically. While we use the former approach as our default, we also propose the following adaptive method. Similar to how $\\epsilon$ defines the privacy boundary in Differential Privacy, we introduce a privacy budget $\\zeta$ , limiting the total magnitude of the loss change defined in Eq. (12):", + "bbox": [ + 89, + 470, + 483, + 561 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left| \\Delta L _ {i + 1} \\right| \\cdot R \\cdot E \\cdot M / N \\leq \\zeta , \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 574, + 480, + 592 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\gamma = \\min \\left(\\max \\left(1 - N \\zeta / \\left(R E M \\lambda_ {o} \\left(g _ {K}\\right) ^ {2}\\right), 0\\right), 1\\right). \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 594, + 480, + 612 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Privacy-Preserving Local Meta-Optimization", + "text_level": 1, + "bbox": [ + 89, + 625, + 473, + 641 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 2 illustrates FedMeNF's local meta-optimization. In existing FML approaches, each outer loop iteration brings the local meta-learner $w_{i}$ closer to the optimal neural field for the client's private data. Sharing the optimized local meta-learner $w_{E}$ with the server leads to significant privacy leakage, as the server can directly exploit $w_{E}$ to reconstruct the original private data. In contrast, FedMeNF explicitly prevents the $w_{i}$ from converging to the optimal neural field, focusing solely on gradient alignment. Thus, FedMeNF trains a local meta-learner that retains minimal private data while still enabling efficient optimization.", + "bbox": [ + 89, + 647, + 483, + 814 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 89, + 829, + 223, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Setup. We assume a scenario where $N = 50$ clients participate in the training. The training consists of 1000 communication rounds, with 5 participants randomly selected", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/01816b1cae4d92a36ee0cc792a50f019324ce6f48a8a288ccab53136c2f4380a.jpg", + "image_caption": [ + "Figure 3. Test-time optimization of the baselines and our Fed-MeNF $(\\gamma = 0.75)$ using FedAvg [41] on the Cars dataset [6]." + ], + "image_footnote": [], + "bbox": [ + 526, + 92, + 890, + 227 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/de1d98148821e44e2ee6e5166424d2fbbd83f8109cb520b9a3c8957748e1a33d.jpg", + "image_caption": [ + "(a) $\\mathrm{PSNR}_p / \\mathrm{Rounds}$", + "Figure 4. Comparison of privacy preservation between baselines and our FedMeNF with different $\\gamma$ values using FedAvg [41] on the Cars dataset [6]. Baselines like MAML, which excel in reconstruction (PSNR $\\uparrow$ ), exhibit high PSNR $_p$ value, while methods like Reptile, which have lower PSNR $_p$ value, perform poorly in reconstruction (PSNR $\\downarrow$ ). FedMeNF effectively adjusts $\\gamma$ to maintain an acceptable PSNR $_p$ value while achieving a comparable PSNR." + ], + "image_footnote": [], + "bbox": [ + 517, + 277, + 681, + 419 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/240aff57adbb309f2e7855b1b799977f02148224b13c1c58187e2665571a6fcb.jpg", + "image_caption": [ + "(b) PSNR / $\\mathrm{PSNR}_p$" + ], + "image_footnote": [], + "bbox": [ + 686, + 279, + 901, + 419 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "per round. We employ the SIREN architecture [51] for image and video modalities, which consists of 6 layers with a hidden dimension of 128. For neural radiance fields (NeRF) [43], we adopt the model used in [9, 58], which consists of 6 layers with a hidden dimension of 256. Please refer to Appendix D for more implementation details.", + "bbox": [ + 511, + 566, + 905, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We conduct experiments across the datasets of various modalities, including images, videos, and NeRFs. To evaluate performance with limited local training data per client, we select scenarios in which each client has only one or very few tasks. For images, we use the cat category from the PetFace dataset [49], assuming each client has images of a unique cat instance. For videos, we use the GolfDB dataset [42], where each client has videos of only one person's golf swings. For NeRFs, we use the Cars category from the ShapeNet dataset [6] and the FaceScape dataset [65, 70] for human faces. Please refer to Appendix E for detailed distribution information for each dataset.", + "bbox": [ + 511, + 657, + 906, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We test a lower bound baseline referred to as Local, using only local dataset without an FL approach. We construct baselines by combining an FL algorithm and a meta-learning algorithm [8, 18]. For the FL algo", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2165", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/5a69397d8e3339daabeb066e4fef44a4a6903d365a6fed572efac9875243f36c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Modality DatasetImage PetFace [49]Video GolfDB [42]3D (NeRF) Cars [6]FaceScape [65, 70]
Method\\Metric\\(PSNR_p\\)\\(PSNR\\)\\(\\Delta(↑)\\)\\(PSNR_p\\)\\(PSNR\\)\\(\\Delta(↑)\\)\\(PSNR_p\\)\\(PSNR\\)\\(\\Delta(↑)\\)\\(PSNR_p\\)\\(PSNR\\)\\(\\Delta(↑)\\)
Local-22.29--26.92--17.13--23.67-
FedAvg + MAML16.5727.3910.8221.2129.688.4719.7323.083.3521.3128.597.28
+ FOMAML18.5223.154.6320.8228.577.7519.7323.663.9321.2428.657.41
+ Reptile17.3922.525.1319.8927.227.3319.9621.982.0221.9228.246.32
+ meta-NSGD12.495.15-7.3410.964.85-6.116.8510.623.777.7211.293.57
+ Ours14.772712.2317.3128.8911.5812.1524.0511.915.1627.8812.72
FedProx + MAML16.5827.4910.9120.9329.268.3319.6722.943.2721.128.547.44
+ FOMAML18.5324.325.7921.2128.767.5519.6723.734.0621.0228.527.5
+ Reptile17.3722.55.1319.7927.217.4219.9522.112.1621.8928.216.32
+ meta-NSGD12.495.15-7.3410.964.84-6.126.8510.914.067.7211.293.57
+ Ours14.4427.0812.6415.6828.9613.2812.1423.9811.8414.0927.5413.45
Scaffold + MAML16.7127.6610.9521.329.117.8119.0724.215.1421.0928.517.42
+ FOMAML18.5523.985.4321.2928.827.5319.0724.475.421.0128.517.5
+ Reptile17.3722.475.12027.237.2319.8122.62.7921.7928.146.35
+ meta-NSGD12.495.13-7.3610.964.84-6.126.8510.633.787.7211.293.57
+ Ours14.9327.1912.2616.2129.2313.0214.1424.349.9413.9427.5313.59
FedNova + MAML16.5227.5110.9921.4729.648.1719.7223.633.9121.3728.597.22
+ FOMAML18.524.15.620.1228.698.5719.7223.273.5521.328.627.32
+ Reptile17.3822.525.1419.9327.267.3319.9622.72.7421.9828.176.19
+ meta-NSGD12.495.14-7.3510.964.84-6.126.8510.63.757.7211.293.57
+ Ours14.9427.1512.2115.7128.9813.2712.1424.1211.9815.4527.8612.41
FedExP + MAML16.5727.3910.8221.1829.588.419.8122.873.0621.3327.646.31
+ FOMAML18.5223.154.6320.8228.67.7819.8122.81321.2627.666.4
+ Reptile17.3922.645.2519.8927.237.3420.9122.031.1221.9328.276.35
+ meta-NSGD12.495.15-7.3410.964.85-6.116.8510.613.767.7211.293.57
+ Ours14.6526.8612.2115.3828.9413.5612.1724.0511.8813.6926.1212.43
FedACG + MAML16.6327.510.8721.0929.518.4219.9222.122.2828.025.74
+ FOMAML18.4824.045.5621.1228.77.5819.921.952.0522.2428.035.79
+ Reptile17.3822.635.2519.8327.237.420.1322.262.1322.0728.326.25
+ meta-NSGD12.495.15-7.3410.964.84-6.126.8510.593.747.7211.293.57
+ Ours14.7126.9712.2615.9428.9913.0510.9322.4511.5214.8327.4612.63
", + "bbox": [ + 94, + 87, + 903, + 588 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison of reconstruction (PSNR(↑)) and privacy (PSNRp(↓)) quality across different modalities: images, videos, and 3D NeRF scenes. $\\Delta$ is defined as PSNR - PSNRp where a larger value indicates a better trade-off between reconstruction quality and privacy protection. FedMeNF demonstrates a balanced performance by maintaining high PSNR for reconstruction and low PSNRp for privacy. The redhighlighted cases indicate cases where PSNRp is high enough to qualitatively confirm privacy leakage, and the gray-highlighted cases indicate cases where the reconstruction performance is lower than that of the Local.", + "bbox": [ + 89, + 597, + 906, + 667 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/30bfe43aa9058b497e70d6d564c071706d83e3cc81509f8685758ae286416718.jpg", + "image_caption": [ + "(a) GT (PSNR)" + ], + "image_footnote": [], + "bbox": [ + 122, + 691, + 178, + 746 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6a8ed5e90f4834cbac859f71963801e160942e0363786dc1f182e01b98a3bd6d.jpg", + "image_caption": [ + "(b) Local (14.97)" + ], + "image_footnote": [], + "bbox": [ + 259, + 679, + 316, + 747 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9bca89313a2bedfb1d98a8609761513392c458211f8a5da97ddb444de2a5548f.jpg", + "image_caption": [ + "(c) MAML (21.22)" + ], + "image_footnote": [], + "bbox": [ + 398, + 691, + 454, + 747 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2b89c157dc39217a7d87a3f5b6a3892500bb6c8aa7f0b7dad9e5c2ab52d13728.jpg", + "image_caption": [ + "(d) FOMAML (21.44)" + ], + "image_footnote": [], + "bbox": [ + 535, + 691, + 589, + 747 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/abe7be2f581da1b4ee89431eb1cae6b683ff13b1110504d818a7caf50f62c472.jpg", + "image_caption": [ + "(e) Reptile (21.18)" + ], + "image_footnote": [], + "bbox": [ + 674, + 691, + 728, + 747 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0d9655a6a68b363a9c951cfb5ea19449d40bb61aa53311eac4a569098053cafb.jpg", + "image_caption": [ + "(f) FedMeNF (21.92)" + ], + "image_footnote": [], + "bbox": [ + 810, + 691, + 866, + 747 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/adf67963b6502cdaed9f62900c7ae0272f474411b89c6a450b9f640fa9199e09.jpg", + "image_caption": [ + "(g) GT (PSNR)" + ], + "image_footnote": [], + "bbox": [ + 109, + 767, + 199, + 835 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/02946aba121673d9b6fcdc9e0646309758d3a19b157d654f71c77be6ad30f3c2.jpg", + "image_caption": [ + "(h) Local (32.69)" + ], + "image_footnote": [], + "bbox": [ + 246, + 768, + 336, + 835 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0963bd913cce74e77db27b13d8d4b241723865689909287a13543838fa46b515.jpg", + "image_caption": [ + "(i) MAML (33.17)", + "Figure 5. Qualitative results on the Cars [6] and FaceScape [65, 70] datasets. Test-time optimization was performed for 8192 steps on the Cars dataset and 32768 steps on the FaceScape dataset." + ], + "image_footnote": [], + "bbox": [ + 383, + 768, + 472, + 835 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/47e2f4e4b7f2471ab5e446ae8a809dffb36500ed4334a9e44b020255935d882e.jpg", + "image_caption": [ + "(j) FOMAML (33.26)" + ], + "image_footnote": [], + "bbox": [ + 519, + 768, + 609, + 835 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/798fbfe9e346d394933131b51eccc9135a07c3cd96bd75e56fec1d1c4deb6f94.jpg", + "image_caption": [ + "(k) Reptile (32.28)" + ], + "image_footnote": [], + "bbox": [ + 656, + 767, + 746, + 835 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ccae8ecfe990562d533365d85b09fbccb7e82894916324db588189e568735eaf.jpg", + "image_caption": [ + "(1) FedMeNF (33.54)" + ], + "image_footnote": [], + "bbox": [ + 792, + 767, + 883, + 835 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2166", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "rithm, we use not only FedAvg [41] as the most standard method, but also state-of-the-art methods including FedProx [38], Scaffold [30], FedNova [61], FedExP [28], and FedACG [32]. For the meta-learning algorithm, we select widely used methods such as MAML [19], FOMAML [19], and Reptile [44], alongside HF-MAML [17, 18] and A-MAML [37]. We also include meta-NSGD [69], the meta-learning algorithm with task-level differential privacy.", + "bbox": [ + 89, + 90, + 483, + 212 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.1. Fast Optimization", + "text_level": 1, + "bbox": [ + 89, + 222, + 266, + 238 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 3 illustrates the test-time optimization (TTO) performance of FedMeNF against baselines. Compared to Local, FML approaches reach higher PSNR more quickly. FedMeNF achieves $40.4\\%$ and $74.64\\%$ higher PSNR than Local with and without TTO, respectively. FedMeNF also achieves a higher PSNR in just 39 steps, making it about $210\\mathrm{x}$ faster than Local, which requires 8192 steps.", + "bbox": [ + 89, + 244, + 483, + 351 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2. Privacy Preservation", + "text_level": 1, + "bbox": [ + 89, + 361, + 287, + 376 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 4 illustrates the privacy preservation performance across varying values of the regularization coefficient $\\gamma$ . Fig. 4a presents the privacy metric $\\mathrm{PSNR}_p$ over communication rounds. FedMeNF achieves an $18.25\\%$ , $35.17\\%$ , $38.42\\%$ , and $65.28\\%$ improvement in the privacy metric $\\mathrm{PSNR}_p$ at $\\gamma = 0.25, 0.5$ , and $0.75$ , compared to the baseline. Fig. 4b shows the trade-off between privacy preservation $(\\mathrm{PSNR}_p\\downarrow)$ and reconstruction performance $(\\mathrm{PSNR}\\uparrow)$ . Meta-NSGD, which utilizes the differential privacy mechanism, achieves lower $\\mathrm{PSNR}_p$ values, indicating stronger privacy protection. However, it suffers from significant degradation in reconstruction performance. In contrast, existing meta-learning methods like MAML achieve high PSNR at the cost of greater vulnerability to privacy leakage with their higher $\\mathrm{PSNR}_p$ . FedMeNF establishes an efficient frontier that balances privacy protection and reconstruction performance, effectively reducing $\\mathrm{PSNR}_p$ with minimal impact on PSNR. Qualitative results are presented in Figs. 5 and 6.", + "bbox": [ + 89, + 383, + 483, + 654 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Tab. 1, FedMeNF achieves comparable PSNR while maintaining an acceptable $\\mathrm{PSNR}_p$ across diverse modalities, as well as across various FL algorithms. FedMeNF also performs well on other metrics, such as SSIM and LPIPS, as detailed in Appendix F.", + "bbox": [ + 89, + 656, + 483, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.3. Justification on the Privacy Metrics", + "text_level": 1, + "bbox": [ + 89, + 742, + 398, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To investigate the practical relevance of the privacy metrics in capturing privacy leakage, we simulate two standard privacy attacks on the Cars dataset [6], assuming an honest-but-curious server. The target models for these attacks are defined as the final local meta-learners from each client in the last communication round. The first is the Membership Inference Attack [50], which determines if a given data sample was in the training set of a black-box model. We assess how accurately the server can predict the membership", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/346fd516277afdbb2e487ac80d55b506d3ffebb64720e0349b7fd329891b7c90.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 532, + 87, + 638, + 169 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a562eafe4f9ad02621d3366237a54a72ea5ee7f9c60c0587c8351c8bad8a0a81.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 661, + 88, + 766, + 169 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c6d47c7949e73ff24a57289db91d133396ef1109d2f62e1270b717eeb398a1eb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 790, + 88, + 895, + 169 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f074aec58906f74817bb3411f355b8bfe534e707e725da5da6e0960ae544de33.jpg", + "image_caption": [ + "(a) GT (PSNR ${}_{p}$ )", + "(b) MAML (17.33)", + "(c) FedMeNF (5.84)", + "(d) MAML (PSNR $p = 23.43$" + ], + "image_footnote": [], + "bbox": [ + 526, + 181, + 893, + 219 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bae1a89f076ecaf7faedd61420b43916b067c49570eae3504f56e74ff01bd0e8.jpg", + "image_caption": [ + "(e) FedMeNF (PSNR $p = 8.38$" + ], + "image_footnote": [], + "bbox": [ + 527, + 233, + 893, + 270 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/81c9455ca49a80572e8c4c5a0fc9c9ba787f2c398ef117ba0020ac7ddd30aa76.jpg", + "image_caption": [ + "(f) GT (PSNR)" + ], + "image_footnote": [], + "bbox": [ + 532, + 292, + 638, + 375 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/102e6271192cebeb87c738a9b452c3138b8a026afd0c13d67837cb5892a47b15.jpg", + "image_caption": [ + "(g) MAML (35.74)" + ], + "image_footnote": [], + "bbox": [ + 661, + 292, + 766, + 373 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c0d3d363bc2f5f15bb3e6e9f67a14e0c79de98a5cc9ccd28b017dcdbb35d763d.jpg", + "image_caption": [ + "(h) FedMeNF (35.16)" + ], + "image_footnote": [], + "bbox": [ + 790, + 292, + 895, + 373 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8054425fce175f7d6e170e2b40f8e2524424d0cd7ea1d44fffe1bf7dece6470a.jpg", + "image_caption": [ + "(i) MAML (PSNR = 29.65)" + ], + "image_footnote": [], + "bbox": [ + 526, + 388, + 893, + 424 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/305aec68b39ebd23d161afae99c35d4166218c9e7228649800ad2347997a4bf4.jpg", + "image_caption": [ + "(j) FedMeNF (PSNR = 29.15)", + "Figure 6. Upper. Reconstruction results of the client's private image/video on the server: (b, d) using MAML and (c, e) using FedMeNF. Lower. Reconstruction results of a new private image/video on the client: (g, i) using MAML and (h, j) using FedMeNF. Our FedMeNF effectively enhances privacy as in (c, e) with minimal impact on the reconstruction quality as in (h, j)." + ], + "image_footnote": [], + "bbox": [ + 524, + 438, + 890, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of the client's local data using the client's local meta-learner as the black-box model. The second is the Property Inference Attack [20], which infers a property of a given model's training set. We evaluate the server's accuracy in classifying the vehicle category of the client's car using the client's local meta-learner. The setup is detailed in Appendix C. Table 2 shows that worse privacy metrics correlate with greater vulnerability to privacy attacks.", + "bbox": [ + 511, + 604, + 906, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Moreover, we examine the correlation between the privacy metrics and $\\epsilon$ of the differential privacy framework using meta-NSGD [69]. Figure 7 shows that the privacy metrics degrade as $\\epsilon$ increases, supporting their generalizability as a measure of privacy leakage.", + "bbox": [ + 511, + 727, + 906, + 803 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.4. Generalization on Non-IID Settings", + "text_level": 1, + "bbox": [ + 511, + 816, + 821, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate the generalization capability of FedMeNF on non-IID settings, we vary the data quantity per client using a Dirichlet distribution with parameter $\\alpha$ ; for the NeRF scenes (where each client has one task), we vary the num", + "bbox": [ + 511, + 839, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2167", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/41c5224fa421ca3c2da650b0e65e13f92794fef814dbaf0f7aa43acbabd34afd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Metric MethodMIA Acc↓PIA Acc↓Privacy MetricsNVS PSNR↑
PSNRp↓SSIMp↓LPIPSp↑
MAML95.8348.7819.730.8270.34323.08
FOMAML97.2248.7819.730.8270.34323.66
Reptile95.8343.9019.960.8390.31121.98
HF-MAML88.8943.9019.240.8270.34123.58
A-MAML95.8343.9019.640.8260.34323.13
Ours55.5629.2712.150.7980.53724.05
Ours (adaptive)55.5629.2712.090.7970.53823.19
", + "bbox": [ + 94, + 88, + 478, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4c390122f4c083c3c3c73865f17bf8bac6613ff22516fda7c9c7ff9e9d37a85f.jpg", + "image_caption": [ + "Figure 7. Correlation between $\\epsilon$ and privacy metrics." + ], + "image_footnote": [], + "bbox": [ + 99, + 292, + 473, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ber of input views, while for the image and video modalities (where each task is a single image or video), we vary the number of tasks. A larger $\\alpha$ indicates that the data quantity per client is more uniform. See Appendix E.3 for details.", + "bbox": [ + 89, + 443, + 482, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 1 presents the results for $\\alpha = 5.0$ , while Tab. 3 shows the results for $\\alpha = 10,5.0,1.0$ . For the baselines, a small $\\alpha$ creates two problems. First, local meta-optimization on a few views tends to result in over-fitted local models, while training on many views leads to well-trained ones. In both cases, the server can reconstruct clients' private data with high fidelity, leading to privacy leakage. Second, over-fitted models perform poorly on novel view synthesis, degrading the global meta-learner. In contrast, FedMeNF is robust to this data imbalance because it only learns how to optimize rather than simply memorizing data. This approach prevents over-fitting and thus achieves robust performance in both privacy preservation and novel view synthesis, regardless of the data quantity.", + "bbox": [ + 88, + 506, + 482, + 719 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.5. Few-shot Novel View Synthesis", + "text_level": 1, + "bbox": [ + 89, + 738, + 364, + 755 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct experiments with varying average number of input views using FedAvg [41] on the Cars [6] dataset, and the results are shown in Tab. 4. For all methods, the view synthesis performance (PSNR) improves as the average number of input views increases. FML approaches outperform the Local across all cases. Notably, FedMeNF achieves the highest PSNR with acceptable $\\mathrm{PSNR}_p$ , regardless of the average number of input views. This confirms that FedMeNF performs well even in few-shot settings.", + "bbox": [ + 89, + 763, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2244669e6d8fd4bf5d1034a5d009887e27f87486c42cc8ad275e7fb3c457e12e.jpg", + "table_caption": [ + "Table 2. Comparison of Membership Inference Attack (MIA) accuracy, Property Inference Attack (PIA) accuracy, Privacy Metrics, and Novel View Synthesis (NVS) quality." + ], + "table_footnote": [], + "table_body": "
Dirichlet α\nMethodα = 10α = 5.0α = 1.0
PSNRpPSNRPSNRpPSNRPSNRpPSNR
MAML19.6223.3319.7323.082021.24
FOMAML19.6223.119.7323.662021.25
Reptile19.9622.4719.9621.9819.9521.29
HF-MAML19.1823.4919.2423.5818.8422.2
A-MAML19.5723.2919.6423.1319.7521.44
Ours12.1323.4712.1524.0512.7324.32
Ours (adaptive)12.123.5212.0923.1913.7320.06
", + "bbox": [ + 517, + 88, + 903, + 224 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/688a9050e24e94816069d3c5e9177d572e4ee56b2ba2c4801bddd89d6d0a8422.jpg", + "table_caption": [ + "Table 3. Results of the novel view synthesis performance and privacy metric with different Dirichlet parameters $\\alpha = \\left\\lbrack {{10},{5.0},{1.0}}\\right\\rbrack$ ." + ], + "table_footnote": [], + "table_body": "
Avg. # of views Method2-view4-view8-view
\\(PSNR_p\\)\\(PSNR\\)\\(PSNR_p\\)\\(PSNR\\)\\(PSNR_p\\)\\(PSNR\\)
MAML19.7320.2319.7323.0819.5426.03
FOMAML19.7420.3919.7323.6619.5425.97
Reptile18.5920.819.9621.9820.3426.3
HF-MAML18.8321.1519.2423.5818.4125.62
A-MAML19.7320.2519.6423.1319.5725.22
Ours15.5621.2712.1524.0511.826.31
Ours (adaptive)15.8420.2112.0923.1912.0325.82
", + "bbox": [ + 517, + 273, + 903, + 410 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Results of the novel view synthesis performance and privacy metric for 2, 4, and 8 input views per task.", + "bbox": [ + 511, + 415, + 903, + 445 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Discussion", + "text_level": 1, + "bbox": [ + 513, + 470, + 627, + 486 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Since FedMeNF is designed for distributed devices, we use a relatively simple NF model. Although recent work [9, 31, 34] has achieved generalizable NFs using large models like transformers, these methods are often impractical for FL with edge devices due to their high communication and computational demands. Future work should focus on scalable, generalizable NFs aligned with FL constraints.", + "bbox": [ + 511, + 496, + 906, + 603 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 617, + 633, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We proposed FedMeNF, the first federated meta-learning framework explicitly designed for neural fields of private data. FedMeNF regulates the increase of our intuitive and widely applicable privacy metric to train the global neural fields meta-learner while preserving privacy. Through experiments across diverse data modalities, including images, videos, and NeRFs, we demonstrate that FedMeNF effectively preserves client privacy while achieving rapid optimization speed and high reconstruction performance, even when local data are highly limited or non-IID. A potential limitation of this study is that we used relatively simple neural field models and vision data only. Future work will investigate generalizable, privacy-preserving neural fields for a broader range of personal data. Measuring and preserving privacy in neural fields is a crucial and novel challenge which has received less attention. We believe this study will serve as an essential starting point for further research.", + "bbox": [ + 511, + 643, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2168", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 259, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was partially supported by the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. RS-2022-II220156, Fundamental research on continual meta-learning for quality enhancement of casual videos and their 3D metaverse transformation; No. RS-2021-II211343, Artificial Intelligence Graduate School Program (Seoul National University)), National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. 2023R1A2C2005573), and Center for Applied Research in Artificial Intelligence (CARAI) grant funded by Defense Acquisition Program Administration (DAPA) and Agency for Defense Development (ADD) (UD230017TD). Gunhee Kim is the corresponding author.", + "bbox": [ + 89, + 114, + 485, + 327 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 339, + 187, + 354 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Martin Abadi, Andy Chu, Ian Goodfellow, H Brendan McMahan, Ilya Mironov, Kunal Talwar, and Li Zhang. Deep learning with differential privacy. In ACM CCS, 2016. 3", + "[2] Sawsan AbdulRahman, Hanine Tout, Hakima Ould-Slimane, Azzam Mourad, Chamseddine Talhi, and Mohsen Guizani. A survey on federated learning: The journey from centralized to distributed on-site learning and beyond. IEEE Internet of Things Journal, 2020. 3", + "[3] Rohith Agaram, Shaurya Dewan, Rahul Sajnani, Adrien Poulenard, Madhava Krishna, and Srinath Sridhar. Canonical fields: Self-supervised learning of pose-canonicalized neural fields. In CVPR, 2023. 2", + "[4] Mohammed Aledhari, Rehma Razzak, Reza M Parizi, and Fahad Saeed. Federated learning: A survey on enabling technologies, protocols, and applications. IEEE Access, 2020. 1", + "[5] Mark Bun and Thomas Steinke. Concentrated differential privacy: Simplifications, extensions, and lower bounds. In Theory of Cryptography, 2016. 3", + "[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv:1512.03012, 2015. 5, 6, 7, 8, 3, 4, 10, 13", + "[7] Daoyuan Chen, Dawei Gao, Weirui Kuang, Yaliang Li, and Bolin Ding. pfl-bench: A comprehensive benchmark for personalized federated learning. NeurIPS, 2022. 3, 4, 6", + "[8] Fei Chen, Mi Luo, Zhenhua Dong, Zhenguo Li, and Xiuqiang He. Federated meta-learning with fast convergence and efficient communication. arXiv:1802.07876, 2018. 1, 3, 5", + "[9] Yinbo Chen and Xiaolong Wang. Transformers as metalearners for implicit neural representations. In ECCV, 2022. 2, 4, 5, 8, 1, 3", + "[10] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In CVPR, 2021. 1", + "[11] Zeyuan Chen, Yinbo Chen, Jingwen Liu, Xingqian Xu, Vedit Goel, Zhangyang Wang, Humphrey Shi, and Xiaolong" + ], + "bbox": [ + 93, + 364, + 480, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Wang. Videoirr: Learning video implicit neural representation for continuous space-time super-resolution. In CVPR, 2022. 1", + "[12] Stephen Coombes, Peter beim Graben, Roland Potthast, and James Wright. Neural fields: theory and applications. Springer, 2014. 1", + "[13] Moming Duan, Qinbin Li, Linshan Jiang, and Bingsheng He. Towards open federated learning platforms: Survey and vision from technical and legal perspectives, 2024. 1", + "[14] Emilien Dupont, Hyunjik Kim, SM Eslami, Danilo Rezende, and Dan Rosenbaum. From data to functa: Your data point is a function and you can treat it like one. arXiv:2201.12204, 2022. 1, 2, 3", + "[15] Cynthia Dwork and Guy N Rothblum. Concentrated differential privacy. arXiv:1603.01887, 2016. 3", + "[16] Cynthia Dwork, Krishnamaram Kenthapadi, Frank McSherry, Ilya Mironov, and Moni Naor. Our data, ourselves: Privacy via distributed noise generation. In EUROCRYPT, 2006. 3", + "[17] Alireza Fallah, Aryan Mokhtari, and Asuman Ozdaglar. On the convergence theory of gradient-based model-agnostic meta-learning algorithms. In AISTATS, 2020. 7, 1", + "[18] Alireza Fallah, Aryan Mokhtari, and Asuman Ozdaglar. Personalized federated learning with theoretical guarantees: A model-agnostic meta-learning approach. NeurIPS, 2020. 1, 3, 5, 7", + "[19] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In ICML, 2017. 1, 7, 6, 10", + "[20] Karan Ganju, Qi Wang, Wei Yang, Carl A Gunter, and Nikita Borisov. Property inference attacks on fully connected neural networks using permutation invariant representations. In ACM CCS, pages 619-633, 2018. 7, 3", + "[21] Wei Gao, Shangwei Guo, Tianwei Zhang, Han Qiu, Yonggang Wen, and Yang Liu. Privacy-preserving collaborative learning with automatic transformation search. In CVPR, 2021. 4", + "[22] Jeffrey Gu, Kuan-Chieh Wang, and Serena Yeung. Generalizable neural fields as partially observed neural processes. In ICCV, 2023. 1", + "[23] Jamie Hayes, Borja Balle, and Saeed Mahloujifar. Bounding training data reconstruction in dp-sgd. In NeurIPS, 2023. 3", + "[24] Lachlan Holden, Feras Dayoub, David Harvey, and Tat-Jun Chin. Federated neural radiance fields. arXiv:2305.01163, 2023. 1, 2", + "[25] Minui Hong, Junhyeog Yun, Insu Jeon, and Gunhee Kim. Fedavp: Augment local data via shared policy in federated learning. NeurIPS, 2024. 4", + "[26] Yu Huang, Yingbin Liang, and Longbo Huang. Provable generalization of overparameterized meta-learning trained with sgd. NeurIPS, 2022. 1", + "[27] Ahmed Imteaj, Urmish Thakker, Shiqiang Wang, Jian Li, and M Hadi Amini. A survey on federated learning for resource-constrained IoT devices. IEEE Internet of Things Journal, 2021. 1", + "[28] Divyansh J Hunjunwala, Shiqiang Wang, and Gauri Joshi. Fedexp: Speeding up federated averaging via extrapolation. arXiv:2301.09604, 2023. 7, 1, 11, 12, 13, 14" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2169", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Yihan Jiang, Jakub Konečný, Keith Rush, and Sreeram Kannan. Improving federated learning personalization via model agnostic meta learning. arXiv:1909.12488, 2019. 1, 3", + "[30] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank Reddi, Sebastian Stich, and Ananda Theertha Suresh. Scaffold: Stochastic controlled averaging for federated learning. In ICML, 2020. 7, 1, 11, 12, 13, 14", + "[31] Chiheon Kim, Doyup Lee, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Generalizable implicit neural representations via instance pattern composers. In CVPR, 2023. 4, 8, 1", + "[32] Geeho Kim, Jinkyu Kim, and Bohyung Han. Communication-efficient federated learning with accelerated client gradient. In CVPR, 2024. 7, 1, 11, 12, 13, 14", + "[33] Subin Kim, Sihyun Yu, Jaeho Lee, and Jinwoo Shin. Scalable neural video representations with learnable positional features. NeurIPS, 2022. 1", + "[34] Doyup Lee, Chiheon Kim, Minsu Cho, and WOOK SHIN HAN. Locality-aware generalizable implicit neural representation. NeurIPS, 2024. 8", + "[35] Jeffrey Li, Mikhail Khodak, Sebastian Caldas, and Ameet Talwalkar. Differentially private meta-learning. In ICLR, 2020. 3", + "[36] Qinbin Li, Yiqun Diao, Quan Chen, and Bingsheng He. Federated learning on non-iid data silos: An experimental study. In IEEE International Conference on Data Engineering, 2022. 1", + "[37] Shibo Li, Zheng Wang, Akil Narayan, Robert Kirby, and Shandian Zhe. Meta-learning with adjoint methods. In AISTATS, pages 7239-7251. PMLR, 2023. 7", + "[38] Xiang Li, Kaixuan Huang, Wenhao Yang, Shusen Wang, and Zhihua Zhang. On the convergence of fedavg on non-iid data. arXiv:1907.02189, 2019. 7, 1, 11, 12, 13, 14", + "[39] Zhenguo Li, Fengwei Zhou, Fei Chen, and Hang Li. Meta-sgd: Learning to learn quickly for few-shot learning. arXiv:1707.09835, 2017. 1", + "[40] Susan Liang, Chao Huang, Yapeng Tian, Anurag Kumar, and Chenliang Xu. Av-nerf: Learning neural fields for real-world audio-visual scene synthesis. NeurIPS, 2023. 1", + "[41] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In AISTATS, 2017. 1, 3, 5, 7, 8, 6, 10, 11, 12, 13, 14", + "[42] William McNally, Kanav Vats, Tyler Pinto, Chris Dulhanty, John McPhee, and Alexander Wong. Golfdb: A video database for golf swing sequencing. In CVPRW, 2019. 5, 6, 3, 4, 7, 12", + "[43] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 2021. 2, 5, 3, 4", + "[44] Alex Nichol, Joshua Achiam, and John Schulman. On first-order meta-learning algorithms. arXiv:1803.02999, 2018. 1, 7", + "[45] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learn" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ing implicit 3d representations without 3d supervision. In CVPR, 2020. 1", + "[46] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, 2019. 1", + "[47] Vishwanath Saragadam, Jasper Tan, Guha Balakrishnan, Richard G Baraniuk, and Ashok Veeraraghavan. Miner: Multiscale implicit neural representation. In ECCV, 2022. 1", + "[48] Bipasha Sen, Aditya Agarwal, Vinay P Namboodiri, and CV Jawahar. Inr-v: A continuous representation space for videobased generative tasks. arXiv:2210.16579, 2022. 1", + "[49] Risa Shinoda and Kaede Shiohara. Petface: A large-scale dataset and benchmark for animal identification. In ECCV, 2025. 5, 6, 3, 4, 7, 11", + "[50] Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. Membership inference attacks against machine learning models. In IEEE Symposium on Security and Privacy, pages 3-18. IEEE, 2017. 7, 2", + "[51] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. NeurIPS, 2020. 2, 5, 3", + "[52] Ivan Skorokhodov, Savva Ignatyev, and Mohamed Elhoseiny. Adversarial generation of continuous images. In CVPR, 2021. 1", + "[53] Kun Su, Mingfei Chen, and Eli Shlizerman. Inras: Implicit neural representation for audio scenes. NeurIPS, 2022. 1", + "[54] Terpeei Suzuki. Federated learning for large-scale scene modeling with neural radiance fields. arXiv:2309.06030, 2023.2", + "[55] Terpei Suzuki. Fed3dgs: Scalable 3d gaussian splatting with federated learning. arXiv:2403.11460, 2024. 1, 2", + "[56] Jihoon Tack, Subin Kim, Sihyun Yu, Jaeho Lee, Jinwoo Shin, and Jonathan Richard Schwarz. Learning large-scale neural fields via context pruned meta-learning. NeurIPS, 2024. 2, 3", + "[57] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. NeurIPS, 2020. 1, 2", + "[58] Matthew Tancik, Ben Mildenhall, Terrance Wang, Divi Schmidt, Pratul P Srinivasan, Jonathan T Barron, and Ren Ng. Learned initializations for optimizing coordinate-based neural representations. In CVPR, 2021. 5, 1, 3", + "[59] Zaid Tasneem, Akshit Dave, Abhishek Singh, Kushagra Tiwary, Praneeth Vepakomma, Ashok Veeraraghavan, and Ramesh Raskar. Decentnerfs: Decentralized neural radiance fields from crowdsourced images. arXiv:2403.13199, 2024. 1", + "[60] Ayush Tewari, Ohad Fried, Justus Thies, Vincent Sitzmann, Stephen Lombardi, Kalyan Sunkavalli, Ricardo MartinBrualla, Tomas Simon, Jason Saragih, Matthias Nießner, et al. State of the art on neural rendering. In Computer Graphics Forum, 2020. 1" + ], + "bbox": [ + 516, + 90, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2170", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[61] Jianyu Wang, Qinghua Liu, Hao Liang, Gauri Joshi, and H Vincent Poor. Tackling the objective inconsistency problem in heterogeneous federated optimization. NeurIPS, 2020. 7, 1, 11, 12, 13, 14", + "[62] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 2004. 4", + "[63] Kang Wei, Jun Li, Ming Ding, Chuan Ma, Howard H. Yang, Farhad Farokhi, Shi Jin, Tony Q. S. Quek, and H. Vincent Poor. Federated learning with differential privacy: Algorithms and performance analysis. IEEE Transactions on Information Forensics and Security, 2020. 3", + "[64] Yuexiang Xie, Zhen Wang, Dawei Gao, Daoyuan Chen, Liuyi Yao, Weirui Kuang, Yaliang Li, Bolin Ding, and Jingren Zhou. Federatedscope: A flexible federated learning platform for heterogeneity. arXiv:2204.05011, 2022. 4, 6", + "[65] Haotian Yang, Hao Zhu, Yanru Wang, Mingkai Huang, Qiu Shen, Ruigang Yang, and Xun Cao. Facescape: a large-scale high quality 3d face dataset and detailed riggable 3d face prediction. In CVPR, 2020. 5, 6, 3, 7, 14", + "[66] Kai Yue, Richeng Jin, Chau-Wai Wong, Dror Baron, and Huaiyu Dai. Gradient obfuscation gives a false sense of security in federated learning. In USENIX Security Symposium, 2023. 4", + "[67] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 4", + "[68] Yintian Zhang and Ziyu Shao. Federated neural radiance field for distributed intelligence. arXiv:2406.10474, 2024. 1, 2", + "[69] Xinyu Zhou and Raef Bassily. Task-level differentially private meta learning. NeurIPS, 2022. 7, 1", + "[70] Hao Zhu, Haotian Yang, Longwei Guo, Yidi Zhang, Yanru Wang, Mingkai Huang, Menghua Wu, Qiu Shen, Ruigang Yang, and Xun Cao. Facescape: 3d facial dataset and benchmark for single-view 3d face reconstruction. IEEE TPAMI, 2023. 5, 6, 3, 7, 14", + "[71] Yiyu Zhuang, Hao Zhu, Xusen Sun, and Xun Cao. Mofanerf: Morphable facial neural radiance field. In ECCV, 2022. 1" + ], + "bbox": [ + 91, + 92, + 480, + 657 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2171", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/FedMeNF_ Privacy-Preserving Federated Meta-Learning for Neural Fields/97cb26d5-5659-4687-b57e-cc32fd64dcc4_model.json b/2025/FedMeNF_ Privacy-Preserving Federated Meta-Learning for Neural Fields/97cb26d5-5659-4687-b57e-cc32fd64dcc4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f040ef13a65c903df1ebaadc4348fa9267dc5872 --- /dev/null +++ b/2025/FedMeNF_ Privacy-Preserving Federated Meta-Learning for Neural Fields/97cb26d5-5659-4687-b57e-cc32fd64dcc4_model.json @@ -0,0 +1,3390 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.807, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.123, + 0.13, + 0.877, + 0.153 + ], + "angle": 0, + "content": "FedMeNF: Privacy-Preserving Federated Meta-Learning for Neural Fields" + }, + { + "type": "text", + "bbox": [ + 0.318, + 0.181, + 0.681, + 0.218 + ], + "angle": 0, + "content": "Junhyeog Yun Minui Hong Gunhee Kim Seoul National University" + }, + { + "type": "text", + "bbox": [ + 0.327, + 0.219, + 0.672, + 0.235 + ], + "angle": 0, + "content": "{junhyeog,alsdml123,gunhee}@snu.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.34, + 0.237, + 0.653, + 0.252 + ], + "angle": 0, + "content": "https://github.com/junhyeog/FedMeNF" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.329, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.318, + 0.486, + 0.576 + ], + "angle": 0, + "content": "Neural fields provide a memory-efficient representation of data, which can effectively handle diverse modalities and large-scale data. However, learning to map neural fields often requires large amounts of training data and computations, which can be limited to resource-constrained edge devices. One approach to tackle this limitation is to leverage Federated Meta-Learning (FML), but traditional FML approaches suffer from privacy leakage. To address these issues, we introduce a novel FML approach called FedMeNF. FedMeNF utilizes a new privacy-preserving loss function that regulates privacy leakage in the local meta-optimization. This enables the local meta-learner to optimize quickly and efficiently without retaining the client's private data. Our experiments demonstrate that FedMeNF achieves fast optimization speed and robust reconstruction performance, even with few-shot or non-IID data across diverse data modalities, while preserving client data privacy." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.604, + 0.222, + 0.62 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.735 + ], + "angle": 0, + "content": "Neural Fields (NFs), also known as Implicit Neural Representations (INRs), use a deep neural network to approximate continuous signals [12, 46, 47, 57, 60]. They enable memory-efficient representation and fast processing of data [14]. Thanks to this, NFs effectively handle both large-scale data and diverse modalities, including audio, images, videos, and 3D objects [10, 11, 33, 40, 45, 48, 52, 53, 71]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.902 + ], + "angle": 0, + "content": "However, optimizing NFs is often slow and requires large amounts of data, which can be a serious issue on resource-constrained edge devices. Consider a practical scenario: a user takes a few photos of an object using their phone and wants to get a high-fidelity 3D model of it. Training an NF from scratch with such limited data and resources would be time-consuming and likely yield poor results. One approach to address this issue is meta-learning [19, 44], which aims to train a meta-learner that can quickly adapt to new tasks with only a few samples. Using the meta-learner, the user can quickly optimize a high-performing NF for the" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.289, + 0.744, + 0.304 + ], + "angle": 0, + "content": "new object with just a few images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.306, + 0.907, + 0.502 + ], + "angle": 0, + "content": "Nevertheless, meta-learning itself requires a large and diverse set of tasks to learn an effective initialization or optimization strategy. Unfortunately, in many practical scenarios, users only have data for one or a few tasks. For example, most users typically take photos of only a single instance of privacy-sensitive objects such as their face, body, pet, or car. To enable meta-learning in such scenarios, we can leverage Federated Meta-Learning (FML) [8, 18, 29], which collaboratively trains a global meta-learner using private data from numerous distributed clients. FML builds upon Federated Learning (FL) [4, 13, 27, 36, 41], a paradigm designed to train a global model while ensuring client data remains on local devices to preserve privacy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.506, + 0.909, + 0.838 + ], + "angle": 0, + "content": "Yet, applying FML to NFs introduces a critical challenge that fundamentally undermines the privacy guarantee of FL. The privacy protection in traditional FL relies on the assumption that sharing trained model parameters, rather than the raw data itself, is safe. This assumption completely breaks down for NFs for two main reasons. First, in our scenario, each client has data for only a single or a few task instances (e.g., a single image or 3D object). If even one client has only a single task instance, the local meta-optimization is like standard supervised training on that specific task [26]. Consequently, the trained local meta-learner functions as an optimized NF for that private data. Second, NFs are, in essence, a compressed representation of the data and inherently encapsulate it. Thus, even if only the NF parameters are shared, a malicious server can reconstruct the private data. Due to this inherent privacy leakage issue, existing studies often limit FL applications for NFs to public data such as landmarks [59] or cityscapes [24, 55, 68], where privacy is less of a concern. For these reasons, naively applying FML methods to private data results in severe privacy issues, as the server can reconstruct the private data from the shared local meta-learner." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To address this critical challenge, we propose FedMeNF (Federated Meta-Learning for Neural Fields), a novel FML framework designed to collaboratively train a global NF meta-learner while preserving privacy as overviewed in" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2161" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.088, + 0.322, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.361, + 0.314, + 0.373 + ], + "angle": 0, + "content": "(a) FML with a single task per client." + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.088, + 0.622, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.346, + 0.163, + 0.651, + 0.2 + ], + "angle": 0, + "content": "(b) Comparison of FedMeNF with Local and FML in terms of fast or few-shot optimization on a new task, and privacy preservation." + }, + { + "type": "image", + "bbox": [ + 0.391, + 0.211, + 0.608, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.346, + 0.348, + 0.651, + 0.373 + ], + "angle": 0, + "content": "(c) The client has photos of only one pet, car, body, or face, and aims to optimize a neural field for each instance." + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.089, + 0.887, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.67, + 0.36, + 0.893, + 0.372 + ], + "angle": 0, + "content": "(d) FedMeNF with a single task per client." + }, + { + "type": "image", + "bbox": [ + 0.147, + 0.377, + 0.852, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.419, + 0.907, + 0.504 + ], + "angle": 0, + "content": "Figure 1. Overview of our approach. In each communication round, the server sends the global meta-learner \\(\\theta\\) to clients, each of which uses its private data to train a local meta-learner \\(w\\) based on the global meta-learner. The trained local meta-learners are then sent back to the server to update the global meta-learner. (a) In existing federated meta-learning (FML) approaches, the local meta-learner contains privacy-sensitive information when a client has only a single task, as shown in (c). The server can infer the client's private data from the shared local meta-learner. (d) In contrast, our FedMeNF ensures that the local meta-learner only learns how to efficiently optimize NFs while preserving privacy-sensitive details. (b, c) Local refers to using only the client's data without sharing any information with the server." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.524, + 0.483, + 0.719 + ], + "angle": 0, + "content": "Fig. 1, which then empowers each client to rapidly optimize a high-performance NF even with a small amount of new private data. FedMeNF can be utilized in various privacy-sensitive applications. For example, suppose that a user buys a new car and takes a few pictures of it. We can then train the NF of the car with this small set of images using a global meta-learner trained with other car images in other devices. As another example, with only a handful of facial images, users can employ the global facial NF meta-learner to generate images showing a range of facial expressions. Also, they can upscale low-resolution photos to high-resolution or generate higher frame-rate videos from personal data, such as pet photos or golf swing videos." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To quantify and control privacy leakage, we define a privacy metric, \\(\\mathrm{PSNR}_p\\). We theoretically analyze how \\(\\mathrm{PSNR}_p\\) increases under existing FML approaches. By regularizing the increase of \\(\\mathrm{PSNR}_p\\), FedMeNF forces the local meta-learner to learn only how to optimize the NF faster with limited data without memorizing the private data. We demonstrate that FedMeNF achieves performance comparable to state-of-the-art methods while better preserving privacy across diverse data modalities, including images, videos, and neural radiance fields (NeRFs). Additionally, FedMeNF shows robust performance even when local data are very few or highly diverse among clients." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.524, + 0.833, + 0.538 + ], + "angle": 0, + "content": "Our contributions can be outlined as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.542, + 0.878, + 0.557 + ], + "angle": 0, + "content": "1. This is the first study on FL for NFs on private data." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.557, + 0.905, + 0.587 + ], + "angle": 0, + "content": "2. We theoretically and empirically show how privacy leakage occurs during the FML for NFs." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.587, + 0.906, + 0.633 + ], + "angle": 0, + "content": "3. We propose FedMeNF that preserves the privacy of local data with minimal impact on optimization speed and reconstruction quality." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.633, + 0.906, + 0.679 + ], + "angle": 0, + "content": "4. We conduct comprehensive experiments on FedMeNF across various data modalities, private data sizes, and levels of data diversity, outperforming baseline methods." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.542, + 0.906, + 0.679 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.694, + 0.652, + 0.71 + ], + "angle": 0, + "content": "2. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.812 + ], + "angle": 0, + "content": "Coordinate-based Neural Fields. Coordinate-based neural fields [3, 9, 14, 43, 51, 56, 57] represent continuous functions that map spatial coordinates to signal values such as color or density. The objective for each client \\( m \\) is to find the optimal NF parameter \\( \\varphi_{i}^{m} \\) that best represents its \\( i \\)-th local task data \\( T_{i}^{m} \\). This is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.823, + 0.907, + 0.86 + ], + "angle": 0, + "content": "\\[\n\\varphi_ {i} ^ {m} = \\arg \\min _ {\\varphi} \\sum_ {(\\mathbf {x} _ {j}, \\mathbf {y} _ {j}) \\in T _ {i} ^ {m}} L \\left(f _ {\\varphi} (\\mathbf {x} _ {j}), \\mathbf {y} _ {j}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.903 + ], + "angle": 0, + "content": "where \\(f_{\\varphi}\\) is the forward map parameterized by \\(\\varphi\\), \\(\\mathbf{x}_j\\) represents the input coordinates (e.g., spatial positions), and \\(\\mathbf{y}_j\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2162" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "denotes the target values (e.g., pixel colors). \\( L \\) is a suitable loss function such as mean squared error (MSE)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.484, + 0.213 + ], + "angle": 0, + "content": "Federated Learning. We consider a centralized federated learning (FL) environment [2, 41], where clients do not share their data with either the server or other clients and communicate the model parameters exclusively with the centralized server. The centralized FL aims to optimize the global model that minimizes a given objective function as" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.222, + 0.484, + 0.263 + ], + "angle": 0, + "content": "\\[\n\\arg \\min _ {\\theta} \\sum_ {m = 0} ^ {M - 1} \\alpha^ {m} L (\\theta , D ^ {m}), \\text {w h e r e} \\alpha^ {m} = \\frac {\\left| D ^ {m} \\right|}{\\Sigma_ {i} \\left| D ^ {i} \\right|}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.27, + 0.484, + 0.375 + ], + "angle": 0, + "content": "\\(\\theta\\) is the parameter of the global meta-learner and \\(M\\) is the number of clients. \\(D^{m}\\) is the local dataset of client \\(m\\). \\(\\alpha^{m}\\) is the proportion of the number of data samples of client \\(m\\). \\(L(\\theta, D^{m})\\) represents the loss of model \\(\\theta\\) on a dataset \\(D^{m}\\). This objective ensures that the global model \\(\\theta\\) is optimized across all clients by weighting the contribution of each client based on its local dataset size." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.376, + 0.484, + 0.422 + ], + "angle": 0, + "content": "Federated Meta-learning. Federated meta-learning [7, 8, 18, 29] is a variant of FL where the client trains the local model using a meta-learning approach with the objective:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.442, + 0.483, + 0.484 + ], + "angle": 0, + "content": "\\[\n\\arg \\min _ {\\theta} \\sum_ {m = 0} ^ {M - 1} \\alpha^ {m} L \\left(\\varphi^ {m}, Q ^ {m}\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.485, + 0.482, + 0.508 + ], + "angle": 0, + "content": "\\[\n\\text {w h e r e} \\varphi^ {m} = \\arg \\min _ {\\theta} L (\\theta , S ^ {m}). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.516, + 0.484, + 0.652 + ], + "angle": 0, + "content": "\\(S^m\\) and \\(Q^m\\) are the support set and query set of the client \\(m\\). Each client \\(m\\) first trains the local model \\(\\varphi^m\\) on the support set \\(S^m\\) (inner loop). Then, starting from the global model \\(\\theta\\), the local meta-learner minimizes the loss of \\(\\varphi^m\\) on the query set \\(Q^m\\) (outer loop). This local meta-optimization enables the local meta-learner to adapt quickly to new tasks. The trained local meta-learner is shared with the server, which aggregates the local meta-learners received from participating clients to update the global meta-learner." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.652, + 0.484, + 0.833 + ], + "angle": 0, + "content": "Differential Privacy. Differential privacy (DP) is one of the most popular privacy frameworks with rich theoretical guarantees [1, 5, 15, 23, 35, 63]. It guarantees that when a single entry in the database changes, the change in the probability distribution of the algorithm outputs is bounded by predetermined factors \\(\\epsilon\\) and \\(\\delta\\) [16]. The privacy risk can be regulated by controlling these factors. However, traditional DP methods are designed for large datasets; yet neural fields often capture just a single signal. This causes ambiguity in defining privacy leakage for individual training samples, as well as a significant trade-off between privacy preservation and reconstruction performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.845, + 0.279, + 0.861 + ], + "angle": 0, + "content": "3. Problem Statement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In this work, we assume a scenario where multiple users collaboratively train a global neural field meta-learner using" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.318 + ], + "angle": 0, + "content": "federated meta-learning. Each user has data about a personal object stored locally on their device. For instance, each user holds a few images of his or her own car on their device. Our goal is to optimize neural fields despite limited data and low computational resources on local devices. Thus, optimizing a specific neural field becomes a local task. The local task data are divided into a support set and a query set, each containing images of the object. During training, the support set is used to train the neural field, while the query set is utilized to optimize the local meta-learner. In the testing phase, the support set enables test-time optimization, and the query set is used to assess its performance. The detailed FML process is described in Fig. 1a and Algorithm 1, excluding the highlighted parts which will be discussed in Sec. 5." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.326, + 0.9, + 0.342 + ], + "angle": 0, + "content": "3.1. First-order Approximation of Meta-gradients" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.348, + 0.907, + 0.44 + ], + "angle": 0, + "content": "During the local meta-optimization, the local meta-learner \\( w \\) is updated for each outer step \\( i \\), where task data \\( T_{i} = \\{S_{i}, Q_{i}\\} \\) are sampled from the client's training dataset \\( D_{\\mathrm{train}} \\). Then, the local NF \\( \\varphi \\), initialized by \\( w_{i} \\), are optimized with the support set \\( S_{i} \\) (Eq. (8)), and the local meta-learner \\( w_{i} \\) are optimized with the query set \\( Q_{i} \\) (Eq. (9))." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.445, + 0.907, + 0.492 + ], + "angle": 0, + "content": "Proposition 1. Let \\( g_{M} = \\nabla_{w_{i}}L(\\varphi_{K},B_{K}) \\) and \\( g_{k} = \\nabla_{\\varphi_{0}}L(\\varphi_{0},B_{k}) \\). Then, the first-order approximation of the meta-gradient \\( g_{M} \\) becomes" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.498, + 0.906, + 0.54 + ], + "angle": 0, + "content": "\\[\ng _ {M} \\approx g _ {K} - \\lambda_ {i} \\mathcal {I} _ {K}, w h e r e \\mathcal {I} _ {K} = \\sum_ {k = 0} ^ {K - 1} \\nabla_ {\\varphi_ {0}} \\left\\langle g _ {K}, g _ {k} \\right\\rangle . \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.545, + 0.906, + 0.592 + ], + "angle": 0, + "content": "\\(K\\) is the # of inner steps, \\(B_{0\\dots K - 1}\\) are minibatches of \\(S_{i}\\), \\(B_{K}\\) is a minibatch of \\(Q_{i}\\), and \\(\\lambda_{i}\\) is an inner loop learning rate. The proof is provided in Appendix H." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.599, + 0.742, + 0.615 + ], + "angle": 0, + "content": "3.2. Fast Optimization via \\(\\mathcal{I}_K\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.621, + 0.907, + 0.757 + ], + "angle": 0, + "content": "According to Proposition 1, the meta-gradient \\( g_{M} \\) can be expressed with \\( g_{K} \\) and the inner product term \\( \\mathcal{I}_K \\), which is the sum of inner products between \\( g_{K} \\) and \\( g_{0},\\dots ,g_{K - 1} \\). This term maximizes (i) the alignment of the gradients from the query set \\( (g_{K}) \\) with those from the support set \\( (g_0,\\dots ,g_{K - 1}) \\), and (ii) the alignment of the gradients computed on different minibatches within the support set. This enables the local meta-learner (i) to be more generalized and (ii) to quickly optimize the neural field for a new task." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.768, + 0.884, + 0.786 + ], + "angle": 0, + "content": "4. Privacy Leakage in Local Meta-Learning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.794, + 0.906, + 0.824 + ], + "angle": 0, + "content": "We define \\(\\mathrm{PSNR}_p\\), an intuitive privacy metric to quantify the privacy leakage in FML for neural fields:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.832, + 0.907, + 0.85 + ], + "angle": 0, + "content": "\\[\n\\mathrm {P S N R} _ {p} = \\mathrm {P S N R} \\left(Q ^ {m}, f _ {w} \\left(\\operatorname {C o o r d} \\left(Q ^ {m}\\right)\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\( Q^m \\) is the query set of client \\( m \\), \\( \\operatorname{Coord}(Q^m) \\) is the coordinates of \\( Q^m \\), and \\( f_w \\) is the forward map parameterized by the local meta-learner \\( w \\). \\( \\mathrm{PSNR}_p \\) assesses how well" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2163" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.183 + ], + "angle": 0, + "content": "the server can reconstruct \\( Q^{m} \\) with the shared local metalearner \\( w \\). It is a simple variant of PSNR that measures similarity between images reconstructed with \\( w \\) on the server and the client's local images. A better reconstruction indicates a greater violation of the client's data privacy, meaning that a larger \\( \\mathrm{PSNR}_p \\) indicates more privacy leakage." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.183, + 0.484, + 0.334 + ], + "angle": 0, + "content": "We define our main privacy metric based on PSNR for two reasons. First, neural fields can handle various types of signals such as vision, speech, or sensor data, so does the PSNR. Second, PSNR is one of the most popular metrics for evaluating reconstruction attacks in FL [21, 25, 66], as well as reconstruction tasks [9, 31]. As variants, we also define \\(\\mathrm{SSIM}_p\\) and \\(\\mathrm{LPIPS}_p\\) as privacy metrics to assess structural and perceptual similarities [62, 67]; it can overcome the limitation of \\(\\mathrm{PSNR}_p\\) that is based on only pixel-wise MSE. Please refer to Appendix F for more details." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.345, + 0.307, + 0.361 + ], + "angle": 0, + "content": "4.1. Privacy Leakage via \\( g_{K} \\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.366, + 0.484, + 0.443 + ], + "angle": 0, + "content": "Since \\(\\mathrm{PSNR}_p\\) is defined by \\(10\\log_{10}(R / L(w,Q^m))\\) where \\(R\\) is a constant, it becomes larger as the MSE loss \\(L(w,Q^{m})\\) is smaller. We analyze how \\(L(w,Q^{m})\\) changes during the client's local meta-optimization, which helps understand how \\(\\mathrm{PSNR}_p\\) evolves." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.453, + 0.482, + 0.485 + ], + "angle": 0, + "content": "Proposition 2. Let \\(\\Delta L_{i + 1} = L(w_{i + 1},B_K) - L(w_i,B_K)\\). The first-order approximation of \\(\\Delta L_{i + 1}\\) is then" + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.495, + 0.482, + 0.513 + ], + "angle": 0, + "content": "\\[\n\\Delta L _ {i + 1} \\approx - \\lambda_ {o} \\cdot g _ {K} ^ {2} \\leq 0. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.525, + 0.484, + 0.678 + ], + "angle": 0, + "content": "The proof is provided in Appendix I. Note that \\( Q^m \\) is replaced by \\( B_K \\), a minibatch of the \\( Q^m \\). \\( \\Delta L_{i+1} \\) denotes the change in the loss of the local meta-learner \\( w \\) on the client's query set. According to Proposition 2, \\( \\Delta L_{i+1} \\) is always negative, which means that \\( L(w, Q^m) \\) decreases with each outer loop iteration. Since \\( \\mathrm{PSNR}_p \\) is inversely proportional to \\( L(w, Q^m) \\), \\( \\mathrm{PSNR}_p \\) increases every outer loop iteration. Consequently, to prevent the increase of \\( \\mathrm{PSNR}_p \\), we regularize \\( \\Delta L_{i+1} \\) to be close to zero, thereby mitigating the privacy leakage in the local meta-optimization." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.691, + 0.199, + 0.709 + ], + "angle": 0, + "content": "5. Approach" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.717, + 0.484, + 0.839 + ], + "angle": 0, + "content": "We reveal that federated meta-learning (FML) approaches suffer from privacy leakage as their local meta-gradients increase the privacy metric. To address this issue, we propose FedMeNF, a novel FML framework for neural fields that ensures privacy-preserving local meta-optimization. FedMeNF employs a privacy-preserving loss function \\( L_{pp} \\), designed to regulate the increase in the privacy metric. Please refer to Algorithm 1 and Fig. 1d for its workflow." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.349, + 0.865 + ], + "angle": 0, + "content": "5.1. The Privacy-Preserving Loss" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Proposition 1-2 theoretically show how client data privacy is leaked in federated meta-learning. Now, we introduce" + }, + { + "type": "code_caption", + "bbox": [ + 0.516, + 0.092, + 0.876, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 FedMeNF (see Appendix A for notation)" + }, + { + "type": "text", + "bbox": [ + 0.529, + 0.11, + 0.907, + 0.173 + ], + "angle": 0, + "content": "Input: (1) \\(R\\): # of communication rounds, (2) \\(N\\): # of clients, (3) \\(M\\): # of participants, (4) \\(\\lambda_{i}\\): inner loop learning rate, (5) \\(\\lambda_{o}\\): outer loop learning rate, (6) \\(\\alpha^{m}\\): weight proportional to the client \\(m\\)'s dataset size (see Eq. (2))." + }, + { + "type": "text", + "bbox": [ + 0.53, + 0.175, + 0.767, + 0.19 + ], + "angle": 0, + "content": "Initialize the global meta-learner \\(\\theta_0\\)" + }, + { + "type": "text", + "bbox": [ + 0.53, + 0.19, + 0.87, + 0.204 + ], + "angle": 0, + "content": "for each communication round \\(r = 0,\\dots,R - 1\\) do" + }, + { + "type": "text", + "bbox": [ + 0.555, + 0.205, + 0.815, + 0.22 + ], + "angle": 0, + "content": "Sample \\(M\\) clients from 1,..., \\(N\\) clients" + }, + { + "type": "text", + "bbox": [ + 0.556, + 0.221, + 0.695, + 0.234 + ], + "angle": 0, + "content": "for \\(m = 1,\\dots,M\\) do" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.235, + 0.765, + 0.251 + ], + "angle": 0, + "content": "\\[\nw _ {*} ^ {m} \\leftarrow \\text {L O C A L U P D A T E} \\left(\\theta_ {r}\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.556, + 0.252, + 0.61, + 0.264 + ], + "angle": 0, + "content": "end for" + }, + { + "type": "equation", + "bbox": [ + 0.556, + 0.266, + 0.905, + 0.281 + ], + "angle": 0, + "content": "\\[\n\\theta_ {r + 1} \\leftarrow \\Sigma_ {m} \\alpha^ {m} w _ {*} ^ {m} \\quad \\triangleright \\text {S e r v e r a g g r e g a t i o n}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.282, + 0.586, + 0.294 + ], + "angle": 0, + "content": "end for" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.299, + 0.736, + 0.313 + ], + "angle": 0, + "content": "procedure LOCALUPDATE(θ)" + }, + { + "type": "text", + "bbox": [ + 0.556, + 0.315, + 0.82, + 0.329 + ], + "angle": 0, + "content": "Initialize the local meta-learner \\( w_0 \\gets \\theta \\)" + }, + { + "type": "text", + "bbox": [ + 0.556, + 0.33, + 0.905, + 0.345 + ], + "angle": 0, + "content": "for each step \\(i = 0,\\dots,E - 1\\) do \\(\\triangleright\\) Outer loop" + }, + { + "type": "text", + "bbox": [ + 0.581, + 0.345, + 0.867, + 0.36 + ], + "angle": 0, + "content": "Sample task data \\(T_{i} = \\{S_{i},Q_{i}\\}\\) from \\(D_{\\mathrm{train}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.581, + 0.36, + 0.905, + 0.375 + ], + "angle": 0, + "content": "\\(\\varphi_0 \\gets w_i\\) Initialize the neural field for \\(T_{i}\\)" + }, + { + "type": "text", + "bbox": [ + 0.581, + 0.376, + 0.905, + 0.389 + ], + "angle": 0, + "content": "for each step \\(k = 0,\\dots,K - 1\\) do \\(\\triangleright\\) Inner loop" + }, + { + "type": "text", + "bbox": [ + 0.606, + 0.39, + 0.875, + 0.405 + ], + "angle": 0, + "content": "Sample batch \\(B_{k}\\) from the support set \\(S_{i}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.609, + 0.42, + 0.905, + 0.437 + ], + "angle": 0, + "content": "\\[\n\\varphi_ {k + 1} \\leftarrow \\varphi_ {k} - \\lambda_ {i} \\nabla_ {\\varphi_ {k}} L \\left(\\varphi_ {k}, B _ {k}\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.58, + 0.445, + 0.634, + 0.457 + ], + "angle": 0, + "content": "end for" + }, + { + "type": "text", + "bbox": [ + 0.58, + 0.459, + 0.845, + 0.474 + ], + "angle": 0, + "content": "Sample batch \\(B_{K}\\) from the query set \\(Q_{i}\\)" + }, + { + "type": "text", + "bbox": [ + 0.58, + 0.475, + 0.829, + 0.489 + ], + "angle": 0, + "content": "Select the regularization coefficient \\(\\gamma\\)" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.49, + 0.904, + 0.507 + ], + "angle": 0, + "content": "\\[\nw _ {i + 1} \\leftarrow w _ {i} - \\lambda_ {o} \\nabla_ {w _ {i}} \\left(L \\left(\\varphi_ {K}, B _ {K}\\right) - \\gamma L \\left(w _ {i}, B _ {K}\\right)\\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.507, + 0.905, + 0.526 + ], + "angle": 0, + "content": "\\[\n\\mid \\text {f o r} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.556, + 0.52, + 0.61, + 0.53 + ], + "angle": 0, + "content": "end for" + }, + { + "type": "text", + "bbox": [ + 0.556, + 0.533, + 0.619, + 0.547 + ], + "angle": 0, + "content": "Send \\(w_{E}\\)" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.549, + 0.636, + 0.563 + ], + "angle": 0, + "content": "end procedure" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.594, + 0.907, + 0.625 + ], + "angle": 0, + "content": "a privacy-preserving loss \\( L_{pp} \\) as follows, ensuring that it regularizes the \\( g_{K} \\) term that increases \\( \\mathrm{PSNR}_p \\):" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.636, + 0.905, + 0.653 + ], + "angle": 0, + "content": "\\[\nL _ {p p} (\\gamma , w _ {i}, \\varphi_ {K}, B _ {K}) = L \\left(\\varphi_ {K}, B _ {K}\\right) - \\gamma L \\left(w _ {i}, B _ {K}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.664, + 0.906, + 0.711 + ], + "angle": 0, + "content": "where \\(\\gamma\\) is a regularization coefficient that determines the portion of the \\(g_{K}\\). In Proposition 3 below, we can obtain the privacy-preserving local meta-gradient \\(g_{pp}\\) using \\(L_{pp}\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.719, + 0.904, + 0.751 + ], + "angle": 0, + "content": "Proposition 3. Let \\( g_{pp} = \\nabla_{w_i}L_{pp}(\\gamma, w_i, \\varphi_K, B_K) \\). Then, the first-order approximation of \\( g_{pp} \\) and \\( \\Delta L_{i+1} \\) become" + }, + { + "type": "equation", + "bbox": [ + 0.62, + 0.763, + 0.905, + 0.779 + ], + "angle": 0, + "content": "\\[\ng _ {p p} \\approx (1 - \\gamma) \\cdot g _ {K} - \\lambda_ {i} \\mathcal {I} _ {K}, \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.781, + 0.905, + 0.799 + ], + "angle": 0, + "content": "\\[\n\\Delta L _ {i + 1} \\approx - \\lambda_ {o} (1 - \\gamma) \\left(g _ {K}\\right) ^ {2} \\leq 0. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Note \\(\\mathcal{I}_K\\) is defined in Eq. (5). At \\(\\gamma = 1\\), the \\(g_K\\) term, which causes privacy leakage, is completely removed from \\(g_{pp}\\). This means that the local meta-learner no longer contains information about the client's local dataset and learns only fast optimization by maximizing the inner products between gradients across minibatches. Each client can choose" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2164" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.104, + 0.096, + 0.23, + 0.107 + ], + "angle": 0, + "content": "\\(\\bullet\\) local meta-learner" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.111, + 0.242, + 0.123 + ], + "angle": 0, + "content": "\\(E\\) : # of outer loop steps" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.126, + 0.228, + 0.14 + ], + "angle": 0, + "content": "\\(= = \\gg\\) \\(\\nabla_{w_i}L(\\varphi_K,B_K)\\)" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.143, + 0.223, + 0.156 + ], + "angle": 0, + "content": "\\(\\longrightarrow \\nabla_{\\varphi_k}L(\\varphi_k,B_k)\\)" + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.095, + 0.428, + 0.107 + ], + "angle": 0, + "content": "\\(\\star\\) optimal neural field" + }, + { + "type": "text", + "bbox": [ + 0.301, + 0.111, + 0.403, + 0.123 + ], + "angle": 0, + "content": "i: outer loop step" + }, + { + "type": "text", + "bbox": [ + 0.294, + 0.127, + 0.432, + 0.14 + ], + "angle": 0, + "content": "- \\(\\triangleright -\\gamma \\nabla_{w_i}L(w_i,B_K)\\)" + }, + { + "type": "text", + "bbox": [ + 0.294, + 0.142, + 0.468, + 0.156 + ], + "angle": 0, + "content": "\\(\\Rightarrow \\nabla_{w_i}L_{pp}(\\gamma ,w_i,\\varphi_K,B_K)\\)" + }, + { + "type": "list", + "bbox": [ + 0.294, + 0.095, + 0.468, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.209, + 0.224, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.132, + 0.278, + 0.183, + 0.29 + ], + "angle": 0, + "content": "(a) \\(i = 0\\)" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.196, + 0.354, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.262, + 0.278, + 0.313, + 0.29 + ], + "angle": 0, + "content": "(b) \\(i = 1\\)" + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.168, + 0.483, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.379, + 0.278, + 0.457, + 0.29 + ], + "angle": 0, + "content": "(c) \\(i = E - 1\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.302, + 0.484, + 0.4 + ], + "angle": 0, + "content": "Figure 2. Illustration of FedMeNF's local meta-optimization in a client. The local meta-learner \\( w \\) does not get closer to the optimal neural field and maintains the initial gap (the distance between the global meta-learner \\( \\theta \\) (\\( = w_0 \\)) and the optimal neural field). Nevertheless, as meta-optimization progresses, gradients are being aligned, and the local meta-learner can find \\( w_E \\) from which the optimal neural field is reached with a few gradient steps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.426, + 0.483, + 0.47 + ], + "angle": 0, + "content": "a different \\(\\gamma\\) that aligns with its desired privacy guarantee. We will empirically show that as \\(\\gamma\\) increases, the privacy leakage from the client's local dataset decreases in Sec. 6.2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.472, + 0.484, + 0.563 + ], + "angle": 0, + "content": "There are two ways to set \\(\\gamma\\): as a fixed hyperparameter or adjusted dynamically. While we use the former approach as our default, we also propose the following adaptive method. Similar to how \\(\\epsilon\\) defines the privacy boundary in Differential Privacy, we introduce a privacy budget \\(\\zeta\\), limiting the total magnitude of the loss change defined in Eq. (12):" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.575, + 0.482, + 0.593 + ], + "angle": 0, + "content": "\\[\n\\left| \\Delta L _ {i + 1} \\right| \\cdot R \\cdot E \\cdot M / N \\leq \\zeta , \\tag {13}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.595, + 0.482, + 0.613 + ], + "angle": 0, + "content": "\\[\n\\gamma = \\min \\left(\\max \\left(1 - N \\zeta / \\left(R E M \\lambda_ {o} \\left(g _ {K}\\right) ^ {2}\\right), 0\\right), 1\\right). \\tag {14}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.625, + 0.475, + 0.642 + ], + "angle": 0, + "content": "5.2. Privacy-Preserving Local Meta-Optimization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.648, + 0.484, + 0.815 + ], + "angle": 0, + "content": "Figure 2 illustrates FedMeNF's local meta-optimization. In existing FML approaches, each outer loop iteration brings the local meta-learner \\( w_{i} \\) closer to the optimal neural field for the client's private data. Sharing the optimized local meta-learner \\( w_{E} \\) with the server leads to significant privacy leakage, as the server can directly exploit \\( w_{E} \\) to reconstruct the original private data. In contrast, FedMeNF explicitly prevents the \\( w_{i} \\) from converging to the optimal neural field, focusing solely on gradient alignment. Thus, FedMeNF trains a local meta-learner that retains minimal private data while still enabling efficient optimization." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.83, + 0.224, + 0.847 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Setup. We assume a scenario where \\( N = 50 \\) clients participate in the training. The training consists of 1000 communication rounds, with 5 participants randomly selected" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.093, + 0.891, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.233, + 0.905, + 0.262 + ], + "angle": 0, + "content": "Figure 3. Test-time optimization of the baselines and our Fed-MeNF \\((\\gamma = 0.75)\\) using FedAvg [41] on the Cars dataset [6]." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.279, + 0.683, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.547, + 0.425, + 0.656, + 0.438 + ], + "angle": 0, + "content": "(a) \\(\\mathrm{PSNR}_p / \\mathrm{Rounds}\\)" + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.28, + 0.902, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.743, + 0.425, + 0.846, + 0.438 + ], + "angle": 0, + "content": "(b) PSNR / \\(\\mathrm{PSNR}_p\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.443, + 0.906, + 0.542 + ], + "angle": 0, + "content": "Figure 4. Comparison of privacy preservation between baselines and our FedMeNF with different \\(\\gamma\\) values using FedAvg [41] on the Cars dataset [6]. Baselines like MAML, which excel in reconstruction (PSNR\\(\\uparrow\\)), exhibit high PSNR\\(_p\\) value, while methods like Reptile, which have lower PSNR\\(_p\\) value, perform poorly in reconstruction (PSNR\\(\\downarrow\\)). FedMeNF effectively adjusts \\(\\gamma\\) to maintain an acceptable PSNR\\(_p\\) value while achieving a comparable PSNR." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.906, + 0.658 + ], + "angle": 0, + "content": "per round. We employ the SIREN architecture [51] for image and video modalities, which consists of 6 layers with a hidden dimension of 128. For neural radiance fields (NeRF) [43], we adopt the model used in [9, 58], which consists of 6 layers with a hidden dimension of 256. Please refer to Appendix D for more implementation details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.839 + ], + "angle": 0, + "content": "Datasets. We conduct experiments across the datasets of various modalities, including images, videos, and NeRFs. To evaluate performance with limited local training data per client, we select scenarios in which each client has only one or very few tasks. For images, we use the cat category from the PetFace dataset [49], assuming each client has images of a unique cat instance. For videos, we use the GolfDB dataset [42], where each client has videos of only one person's golf swings. For NeRFs, we use the Cars category from the ShapeNet dataset [6] and the FaceScape dataset [65, 70] for human faces. Please refer to Appendix E for detailed distribution information for each dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Baselines. We test a lower bound baseline referred to as Local, using only local dataset without an FL approach. We construct baselines by combining an FL algorithm and a meta-learning algorithm [8, 18]. For the FL algo" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2165" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.088, + 0.904, + 0.589 + ], + "angle": 0, + "content": "
Modality DatasetImage PetFace [49]Video GolfDB [42]3D (NeRF) Cars [6]FaceScape [65, 70]
Method\\Metric\\(PSNR_p\\)\\(PSNR\\)\\(\\Delta(↑)\\)\\(PSNR_p\\)\\(PSNR\\)\\(\\Delta(↑)\\)\\(PSNR_p\\)\\(PSNR\\)\\(\\Delta(↑)\\)\\(PSNR_p\\)\\(PSNR\\)\\(\\Delta(↑)\\)
Local-22.29--26.92--17.13--23.67-
FedAvg + MAML16.5727.3910.8221.2129.688.4719.7323.083.3521.3128.597.28
+ FOMAML18.5223.154.6320.8228.577.7519.7323.663.9321.2428.657.41
+ Reptile17.3922.525.1319.8927.227.3319.9621.982.0221.9228.246.32
+ meta-NSGD12.495.15-7.3410.964.85-6.116.8510.623.777.7211.293.57
+ Ours14.772712.2317.3128.8911.5812.1524.0511.915.1627.8812.72
FedProx + MAML16.5827.4910.9120.9329.268.3319.6722.943.2721.128.547.44
+ FOMAML18.5324.325.7921.2128.767.5519.6723.734.0621.0228.527.5
+ Reptile17.3722.55.1319.7927.217.4219.9522.112.1621.8928.216.32
+ meta-NSGD12.495.15-7.3410.964.84-6.126.8510.914.067.7211.293.57
+ Ours14.4427.0812.6415.6828.9613.2812.1423.9811.8414.0927.5413.45
Scaffold + MAML16.7127.6610.9521.329.117.8119.0724.215.1421.0928.517.42
+ FOMAML18.5523.985.4321.2928.827.5319.0724.475.421.0128.517.5
+ Reptile17.3722.475.12027.237.2319.8122.62.7921.7928.146.35
+ meta-NSGD12.495.13-7.3610.964.84-6.126.8510.633.787.7211.293.57
+ Ours14.9327.1912.2616.2129.2313.0214.1424.349.9413.9427.5313.59
FedNova + MAML16.5227.5110.9921.4729.648.1719.7223.633.9121.3728.597.22
+ FOMAML18.524.15.620.1228.698.5719.7223.273.5521.328.627.32
+ Reptile17.3822.525.1419.9327.267.3319.9622.72.7421.9828.176.19
+ meta-NSGD12.495.14-7.3510.964.84-6.126.8510.63.757.7211.293.57
+ Ours14.9427.1512.2115.7128.9813.2712.1424.1211.9815.4527.8612.41
FedExP + MAML16.5727.3910.8221.1829.588.419.8122.873.0621.3327.646.31
+ FOMAML18.5223.154.6320.8228.67.7819.8122.81321.2627.666.4
+ Reptile17.3922.645.2519.8927.237.3420.9122.031.1221.9328.276.35
+ meta-NSGD12.495.15-7.3410.964.85-6.116.8510.613.767.7211.293.57
+ Ours14.6526.8612.2115.3828.9413.5612.1724.0511.8813.6926.1212.43
FedACG + MAML16.6327.510.8721.0929.518.4219.9222.122.2828.025.74
+ FOMAML18.4824.045.5621.1228.77.5819.921.952.0522.2428.035.79
+ Reptile17.3822.635.2519.8327.237.420.1322.262.1322.0728.326.25
+ meta-NSGD12.495.15-7.3410.964.84-6.126.8510.593.747.7211.293.57
+ Ours14.7126.9712.2615.9428.9913.0510.9322.4511.5214.8327.4612.63
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.598, + 0.907, + 0.669 + ], + "angle": 0, + "content": "Table 1. Comparison of reconstruction (PSNR(↑)) and privacy (PSNRp(↓)) quality across different modalities: images, videos, and 3D NeRF scenes. \\(\\Delta\\) is defined as PSNR - PSNRp where a larger value indicates a better trade-off between reconstruction quality and privacy protection. FedMeNF demonstrates a balanced performance by maintaining high PSNR for reconstruction and low PSNRp for privacy. The redhighlighted cases indicate cases where PSNRp is high enough to qualitatively confirm privacy leakage, and the gray-highlighted cases indicate cases where the reconstruction performance is lower than that of the Local." + }, + { + "type": "image", + "bbox": [ + 0.123, + 0.693, + 0.179, + 0.747 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.751, + 0.192, + 0.762 + ], + "angle": 0, + "content": "(a) GT (PSNR)" + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.68, + 0.317, + 0.748 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.252, + 0.751, + 0.333, + 0.762 + ], + "angle": 0, + "content": "(b) Local (14.97)" + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.693, + 0.455, + 0.748 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.385, + 0.751, + 0.475, + 0.762 + ], + "angle": 0, + "content": "(c) MAML (21.22)" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.693, + 0.591, + 0.748 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.751, + 0.62, + 0.762 + ], + "angle": 0, + "content": "(d) FOMAML (21.44)" + }, + { + "type": "image", + "bbox": [ + 0.675, + 0.693, + 0.729, + 0.748 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.661, + 0.751, + 0.748, + 0.762 + ], + "angle": 0, + "content": "(e) Reptile (21.18)" + }, + { + "type": "image", + "bbox": [ + 0.812, + 0.693, + 0.867, + 0.748 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.793, + 0.751, + 0.892, + 0.762 + ], + "angle": 0, + "content": "(f) FedMeNF (21.92)" + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.768, + 0.2, + 0.837 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.839, + 0.191, + 0.85 + ], + "angle": 0, + "content": "(g) GT (PSNR)" + }, + { + "type": "image", + "bbox": [ + 0.247, + 0.769, + 0.337, + 0.836 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.252, + 0.838, + 0.333, + 0.85 + ], + "angle": 0, + "content": "(h) Local (32.69)" + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.769, + 0.473, + 0.836 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.385, + 0.838, + 0.473, + 0.85 + ], + "angle": 0, + "content": "(i) MAML (33.17)" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.769, + 0.61, + 0.836 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.838, + 0.616, + 0.85 + ], + "angle": 0, + "content": "(j) FOMAML (33.26)" + }, + { + "type": "image", + "bbox": [ + 0.657, + 0.768, + 0.747, + 0.836 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.659, + 0.838, + 0.746, + 0.85 + ], + "angle": 0, + "content": "(k) Reptile (32.28)" + }, + { + "type": "image", + "bbox": [ + 0.794, + 0.768, + 0.884, + 0.836 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.789, + 0.838, + 0.888, + 0.85 + ], + "angle": 0, + "content": "(1) FedMeNF (33.54)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.855, + 0.907, + 0.884 + ], + "angle": 0, + "content": "Figure 5. Qualitative results on the Cars [6] and FaceScape [65, 70] datasets. Test-time optimization was performed for 8192 steps on the Cars dataset and 32768 steps on the FaceScape dataset." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2166" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.213 + ], + "angle": 0, + "content": "rithm, we use not only FedAvg [41] as the most standard method, but also state-of-the-art methods including FedProx [38], Scaffold [30], FedNova [61], FedExP [28], and FedACG [32]. For the meta-learning algorithm, we select widely used methods such as MAML [19], FOMAML [19], and Reptile [44], alongside HF-MAML [17, 18] and A-MAML [37]. We also include meta-NSGD [69], the meta-learning algorithm with task-level differential privacy." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.223, + 0.267, + 0.239 + ], + "angle": 0, + "content": "6.1. Fast Optimization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.245, + 0.485, + 0.352 + ], + "angle": 0, + "content": "Figure 3 illustrates the test-time optimization (TTO) performance of FedMeNF against baselines. Compared to Local, FML approaches reach higher PSNR more quickly. FedMeNF achieves \\(40.4\\%\\) and \\(74.64\\%\\) higher PSNR than Local with and without TTO, respectively. FedMeNF also achieves a higher PSNR in just 39 steps, making it about \\(210\\mathrm{x}\\) faster than Local, which requires 8192 steps." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.362, + 0.288, + 0.377 + ], + "angle": 0, + "content": "6.2. Privacy Preservation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.385, + 0.484, + 0.655 + ], + "angle": 0, + "content": "Figure 4 illustrates the privacy preservation performance across varying values of the regularization coefficient \\(\\gamma\\). Fig. 4a presents the privacy metric \\(\\mathrm{PSNR}_p\\) over communication rounds. FedMeNF achieves an \\(18.25\\%\\), \\(35.17\\%\\), \\(38.42\\%\\), and \\(65.28\\%\\) improvement in the privacy metric \\(\\mathrm{PSNR}_p\\) at \\(\\gamma = 0.25, 0.5\\), and \\(0.75\\), compared to the baseline. Fig. 4b shows the trade-off between privacy preservation \\((\\mathrm{PSNR}_p\\downarrow)\\) and reconstruction performance \\((\\mathrm{PSNR}\\uparrow)\\). Meta-NSGD, which utilizes the differential privacy mechanism, achieves lower \\(\\mathrm{PSNR}_p\\) values, indicating stronger privacy protection. However, it suffers from significant degradation in reconstruction performance. In contrast, existing meta-learning methods like MAML achieve high PSNR at the cost of greater vulnerability to privacy leakage with their higher \\(\\mathrm{PSNR}_p\\). FedMeNF establishes an efficient frontier that balances privacy protection and reconstruction performance, effectively reducing \\(\\mathrm{PSNR}_p\\) with minimal impact on PSNR. Qualitative results are presented in Figs. 5 and 6." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.657, + 0.484, + 0.732 + ], + "angle": 0, + "content": "As shown in Tab. 1, FedMeNF achieves comparable PSNR while maintaining an acceptable \\(\\mathrm{PSNR}_p\\) across diverse modalities, as well as across various FL algorithms. FedMeNF also performs well on other metrics, such as SSIM and LPIPS, as detailed in Appendix F." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.4, + 0.759 + ], + "angle": 0, + "content": "6.3. Justification on the Privacy Metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To investigate the practical relevance of the privacy metrics in capturing privacy leakage, we simulate two standard privacy attacks on the Cars dataset [6], assuming an honest-but-curious server. The target models for these attacks are defined as the final local meta-learners from each client in the last communication round. The first is the Membership Inference Attack [50], which determines if a given data sample was in the training set of a black-box model. We assess how accurately the server can predict the membership" + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.088, + 0.64, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.172, + 0.625, + 0.183 + ], + "angle": 0, + "content": "(a) GT (PSNR \\( {}_{p} \\) )" + }, + { + "type": "image", + "bbox": [ + 0.662, + 0.089, + 0.767, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.67, + 0.172, + 0.759, + 0.182 + ], + "angle": 0, + "content": "(b) MAML (17.33)" + }, + { + "type": "image", + "bbox": [ + 0.791, + 0.089, + 0.896, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.796, + 0.172, + 0.891, + 0.182 + ], + "angle": 0, + "content": "(c) FedMeNF (5.84)" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.183, + 0.895, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.637, + 0.222, + 0.785, + 0.233 + ], + "angle": 0, + "content": "(d) MAML (PSNR \\(p = 23.43\\)" + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.234, + 0.895, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.637, + 0.272, + 0.785, + 0.284 + ], + "angle": 0, + "content": "(e) FedMeNF (PSNR \\(p = 8.38\\)" + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.293, + 0.639, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.376, + 0.621, + 0.387 + ], + "angle": 0, + "content": "(f) GT (PSNR)" + }, + { + "type": "image", + "bbox": [ + 0.662, + 0.293, + 0.767, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.669, + 0.376, + 0.759, + 0.387 + ], + "angle": 0, + "content": "(g) MAML (35.74)" + }, + { + "type": "image", + "bbox": [ + 0.791, + 0.293, + 0.896, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.793, + 0.376, + 0.893, + 0.387 + ], + "angle": 0, + "content": "(h) FedMeNF (35.16)" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.39, + 0.894, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.644, + 0.426, + 0.777, + 0.437 + ], + "angle": 0, + "content": "(i) MAML (PSNR = 29.65)" + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.439, + 0.892, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.634, + 0.477, + 0.783, + 0.488 + ], + "angle": 0, + "content": "(j) FedMeNF (PSNR = 29.15)" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.492, + 0.907, + 0.577 + ], + "angle": 0, + "content": "Figure 6. Upper. Reconstruction results of the client's private image/video on the server: (b, d) using MAML and (c, e) using FedMeNF. Lower. Reconstruction results of a new private image/video on the client: (g, i) using MAML and (h, j) using FedMeNF. Our FedMeNF effectively enhances privacy as in (c, e) with minimal impact on the reconstruction quality as in (h, j)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.605, + 0.907, + 0.727 + ], + "angle": 0, + "content": "of the client's local data using the client's local meta-learner as the black-box model. The second is the Property Inference Attack [20], which infers a property of a given model's training set. We evaluate the server's accuracy in classifying the vehicle category of the client's car using the client's local meta-learner. The setup is detailed in Appendix C. Table 2 shows that worse privacy metrics correlate with greater vulnerability to privacy attacks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.728, + 0.907, + 0.804 + ], + "angle": 0, + "content": "Moreover, we examine the correlation between the privacy metrics and \\(\\epsilon\\) of the differential privacy framework using meta-NSGD [69]. Figure 7 shows that the privacy metrics degrade as \\(\\epsilon\\) increases, supporting their generalizability as a measure of privacy leakage." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.817, + 0.822, + 0.833 + ], + "angle": 0, + "content": "6.4. Generalization on Non-IID Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.901 + ], + "angle": 0, + "content": "To evaluate the generalization capability of FedMeNF on non-IID settings, we vary the data quantity per client using a Dirichlet distribution with parameter \\(\\alpha\\); for the NeRF scenes (where each client has one task), we vary the num" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2167" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.48, + 0.226 + ], + "angle": 0, + "content": "
Metric MethodMIA Acc↓PIA Acc↓Privacy MetricsNVS PSNR↑
PSNRp↓SSIMp↓LPIPSp↑
MAML95.8348.7819.730.8270.34323.08
FOMAML97.2248.7819.730.8270.34323.66
Reptile95.8343.9019.960.8390.31121.98
HF-MAML88.8943.9019.240.8270.34123.58
A-MAML95.8343.9019.640.8260.34323.13
Ours55.5629.2712.150.7980.53724.05
Ours (adaptive)55.5629.2712.090.7970.53823.19
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.231, + 0.483, + 0.274 + ], + "angle": 0, + "content": "Table 2. Comparison of Membership Inference Attack (MIA) accuracy, Property Inference Attack (PIA) accuracy, Privacy Metrics, and Novel View Synthesis (NVS) quality." + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.293, + 0.474, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.13, + 0.396, + 0.444, + 0.41 + ], + "angle": 0, + "content": "Figure 7. Correlation between \\(\\epsilon\\) and privacy metrics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.444, + 0.483, + 0.505 + ], + "angle": 0, + "content": "ber of input views, while for the image and video modalities (where each task is a single image or video), we vary the number of tasks. A larger \\(\\alpha\\) indicates that the data quantity per client is more uniform. See Appendix E.3 for details." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.507, + 0.483, + 0.72 + ], + "angle": 0, + "content": "Table 1 presents the results for \\(\\alpha = 5.0\\), while Tab. 3 shows the results for \\(\\alpha = 10,5.0,1.0\\). For the baselines, a small \\(\\alpha\\) creates two problems. First, local meta-optimization on a few views tends to result in over-fitted local models, while training on many views leads to well-trained ones. In both cases, the server can reconstruct clients' private data with high fidelity, leading to privacy leakage. Second, over-fitted models perform poorly on novel view synthesis, degrading the global meta-learner. In contrast, FedMeNF is robust to this data imbalance because it only learns how to optimize rather than simply memorizing data. This approach prevents over-fitting and thus achieves robust performance in both privacy preservation and novel view synthesis, regardless of the data quantity." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.739, + 0.365, + 0.756 + ], + "angle": 0, + "content": "6.5. Few-shot Novel View Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We conduct experiments with varying average number of input views using FedAvg [41] on the Cars [6] dataset, and the results are shown in Tab. 4. For all methods, the view synthesis performance (PSNR) improves as the average number of input views increases. FML approaches outperform the Local across all cases. Notably, FedMeNF achieves the highest PSNR with acceptable \\(\\mathrm{PSNR}_p\\), regardless of the average number of input views. This confirms that FedMeNF performs well even in few-shot settings." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.226 + ], + "angle": 0, + "content": "
Dirichlet α\nMethodα = 10α = 5.0α = 1.0
PSNRpPSNRPSNRpPSNRPSNRpPSNR
MAML19.6223.3319.7323.082021.24
FOMAML19.6223.119.7323.662021.25
Reptile19.9622.4719.9621.9819.9521.29
HF-MAML19.1823.4919.2423.5818.8422.2
A-MAML19.5723.2919.6423.1319.7521.44
Ours12.1323.4712.1524.0512.7324.32
Ours (adaptive)12.123.5212.0923.1913.7320.06
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.231, + 0.905, + 0.261 + ], + "angle": 0, + "content": "Table 3. Results of the novel view synthesis performance and privacy metric with different Dirichlet parameters \\( \\alpha = \\left\\lbrack {{10},{5.0},{1.0}}\\right\\rbrack \\) ." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.274, + 0.905, + 0.411 + ], + "angle": 0, + "content": "
Avg. # of views Method2-view4-view8-view
\\(PSNR_p\\)\\(PSNR\\)\\(PSNR_p\\)\\(PSNR\\)\\(PSNR_p\\)\\(PSNR\\)
MAML19.7320.2319.7323.0819.5426.03
FOMAML19.7420.3919.7323.6619.5425.97
Reptile18.5920.819.9621.9820.3426.3
HF-MAML18.8321.1519.2423.5818.4125.62
A-MAML19.7320.2519.6423.1319.5725.22
Ours15.5621.2712.1524.0511.826.31
Ours (adaptive)15.8420.2112.0923.1912.0325.82
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.416, + 0.905, + 0.446 + ], + "angle": 0, + "content": "Table 4. Results of the novel view synthesis performance and privacy metric for 2, 4, and 8 input views per task." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.471, + 0.628, + 0.487 + ], + "angle": 0, + "content": "7. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.497, + 0.907, + 0.604 + ], + "angle": 0, + "content": "Since FedMeNF is designed for distributed devices, we use a relatively simple NF model. Although recent work [9, 31, 34] has achieved generalizable NFs using large models like transformers, these methods are often impractical for FL with edge devices due to their high communication and computational demands. Future work should focus on scalable, generalizable NFs aligned with FL constraints." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.618, + 0.634, + 0.635 + ], + "angle": 0, + "content": "8. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We proposed FedMeNF, the first federated meta-learning framework explicitly designed for neural fields of private data. FedMeNF regulates the increase of our intuitive and widely applicable privacy metric to train the global neural fields meta-learner while preserving privacy. Through experiments across diverse data modalities, including images, videos, and NeRFs, we demonstrate that FedMeNF effectively preserves client privacy while achieving rapid optimization speed and high reconstruction performance, even when local data are highly limited or non-IID. A potential limitation of this study is that we used relatively simple neural field models and vision data only. Future work will investigate generalizable, privacy-preserving neural fields for a broader range of personal data. Measuring and preserving privacy in neural fields is a crucial and novel challenge which has received less attention. We believe this study will serve as an essential starting point for further research." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2168" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.26, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.328 + ], + "angle": 0, + "content": "This work was partially supported by the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. RS-2022-II220156, Fundamental research on continual meta-learning for quality enhancement of casual videos and their 3D metaverse transformation; No. RS-2021-II211343, Artificial Intelligence Graduate School Program (Seoul National University)), National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. 2023R1A2C2005573), and Center for Applied Research in Artificial Intelligence (CARAI) grant funded by Defense Acquisition Program Administration (DAPA) and Agency for Defense Development (ADD) (UD230017TD). Gunhee Kim is the corresponding author." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.34, + 0.188, + 0.356 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.365, + 0.482, + 0.407 + ], + "angle": 0, + "content": "[1] Martin Abadi, Andy Chu, Ian Goodfellow, H Brendan McMahan, Ilya Mironov, Kunal Talwar, and Li Zhang. Deep learning with differential privacy. In ACM CCS, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.409, + 0.482, + 0.477 + ], + "angle": 0, + "content": "[2] Sawsan AbdulRahman, Hanine Tout, Hakima Ould-Slimane, Azzam Mourad, Chamseddine Talhi, and Mohsen Guizani. A survey on federated learning: The journey from centralized to distributed on-site learning and beyond. IEEE Internet of Things Journal, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.479, + 0.482, + 0.532 + ], + "angle": 0, + "content": "[3] Rohith Agaram, Shaurya Dewan, Rahul Sajnani, Adrien Poulenard, Madhava Krishna, and Srinath Sridhar. Canonical fields: Self-supervised learning of pose-canonicalized neural fields. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.535, + 0.482, + 0.576 + ], + "angle": 0, + "content": "[4] Mohammed Aledhari, Rehma Razzak, Reza M Parizi, and Fahad Saeed. Federated learning: A survey on enabling technologies, protocols, and applications. IEEE Access, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.103, + 0.578, + 0.482, + 0.618 + ], + "angle": 0, + "content": "[5] Mark Bun and Thomas Steinke. Concentrated differential privacy: Simplifications, extensions, and lower bounds. In Theory of Cryptography, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.62, + 0.482, + 0.688 + ], + "angle": 0, + "content": "[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv:1512.03012, 2015. 5, 6, 7, 8, 3, 4, 10, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.69, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[7] Daoyuan Chen, Dawei Gao, Weirui Kuang, Yaliang Li, and Bolin Ding. pfl-bench: A comprehensive benchmark for personalized federated learning. NeurIPS, 2022. 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.733, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[8] Fei Chen, Mi Luo, Zhenhua Dong, Zhenguo Li, and Xiuqiang He. Federated meta-learning with fast convergence and efficient communication. arXiv:1802.07876, 2018. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.789, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[9] Yinbo Chen and Xiaolong Wang. Transformers as metalearners for implicit neural representations. In ECCV, 2022. 2, 4, 5, 8, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.831, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[10] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In CVPR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[11] Zeyuan Chen, Yinbo Chen, Jingwen Liu, Xingqian Xu, Vedit Goel, Zhangyang Wang, Humphrey Shi, and Xiaolong" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.365, + 0.482, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "Wang. Videoirr: Learning video implicit neural representation for continuous space-time super-resolution. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[12] Stephen Coombes, Peter beim Graben, Roland Potthast, and James Wright. Neural fields: theory and applications. Springer, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.904, + 0.22 + ], + "angle": 0, + "content": "[13] Moming Duan, Qinbin Li, Linshan Jiang, and Bingsheng He. Towards open federated learning platforms: Survey and vision from technical and legal perspectives, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.222, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[14] Emilien Dupont, Hyunjik Kim, SM Eslami, Danilo Rezende, and Dan Rosenbaum. From data to functa: Your data point is a function and you can treat it like one. arXiv:2201.12204, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.904, + 0.304 + ], + "angle": 0, + "content": "[15] Cynthia Dwork and Guy N Rothblum. Concentrated differential privacy. arXiv:1603.01887, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.904, + 0.347 + ], + "angle": 0, + "content": "[16] Cynthia Dwork, Krishnamaram Kenthapadi, Frank McSherry, Ilya Mironov, and Moni Naor. Our data, ourselves: Privacy via distributed noise generation. In EUROCRYPT, 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.904, + 0.39 + ], + "angle": 0, + "content": "[17] Alireza Fallah, Aryan Mokhtari, and Asuman Ozdaglar. On the convergence theory of gradient-based model-agnostic meta-learning algorithms. In AISTATS, 2020. 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.904, + 0.445 + ], + "angle": 0, + "content": "[18] Alireza Fallah, Aryan Mokhtari, and Asuman Ozdaglar. Personalized federated learning with theoretical guarantees: A model-agnostic meta-learning approach. NeurIPS, 2020. 1, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.904, + 0.489 + ], + "angle": 0, + "content": "[19] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In ICML, 2017. 1, 7, 6, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.904, + 0.545 + ], + "angle": 0, + "content": "[20] Karan Ganju, Qi Wang, Wei Yang, Carl A Gunter, and Nikita Borisov. Property inference attacks on fully connected neural networks using permutation invariant representations. In ACM CCS, pages 619-633, 2018. 7, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.904, + 0.6 + ], + "angle": 0, + "content": "[21] Wei Gao, Shangwei Guo, Tianwei Zhang, Han Qiu, Yonggang Wen, and Yang Liu. Privacy-preserving collaborative learning with automatic transformation search. In CVPR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.904, + 0.643 + ], + "angle": 0, + "content": "[22] Jeffrey Gu, Kuan-Chieh Wang, and Serena Yeung. Generalizable neural fields as partially observed neural processes. In ICCV, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.646, + 0.904, + 0.673 + ], + "angle": 0, + "content": "[23] Jamie Hayes, Borja Balle, and Saeed Mahloujifar. Bounding training data reconstruction in dp-sgd. In NeurIPS, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.676, + 0.904, + 0.715 + ], + "angle": 0, + "content": "[24] Lachlan Holden, Feras Dayoub, David Harvey, and Tat-Jun Chin. Federated neural radiance fields. arXiv:2305.01163, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.904, + 0.759 + ], + "angle": 0, + "content": "[25] Minui Hong, Junhyeog Yun, Insu Jeon, and Gunhee Kim. Fedavp: Augment local data via shared policy in federated learning. NeurIPS, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.904, + 0.801 + ], + "angle": 0, + "content": "[26] Yu Huang, Yingbin Liang, and Longbo Huang. Provable generalization of overparameterized meta-learning trained with sgd. NeurIPS, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.904, + 0.857 + ], + "angle": 0, + "content": "[27] Ahmed Imteaj, Urmish Thakker, Shiqiang Wang, Jian Li, and M Hadi Amini. A survey on federated learning for resource-constrained IoT devices. IEEE Internet of Things Journal, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.904, + 0.9 + ], + "angle": 0, + "content": "[28] Divyansh J Hunjunwala, Shiqiang Wang, and Gauri Joshi. Fedexp: Speeding up federated averaging via extrapolation. arXiv:2301.09604, 2023. 7, 1, 11, 12, 13, 14" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2169" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[29] Yihan Jiang, Jakub Konečný, Keith Rush, and Sreeram Kannan. Improving federated learning personalization via model agnostic meta learning. arXiv:1909.12488, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[30] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank Reddi, Sebastian Stich, and Ananda Theertha Suresh. Scaffold: Stochastic controlled averaging for federated learning. In ICML, 2020. 7, 1, 11, 12, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[31] Chiheon Kim, Doyup Lee, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Generalizable implicit neural representations via instance pattern composers. In CVPR, 2023. 4, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.303 + ], + "angle": 0, + "content": "[32] Geeho Kim, Jinkyu Kim, and Bohyung Han. Communication-efficient federated learning with accelerated client gradient. In CVPR, 2024. 7, 1, 11, 12, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[33] Subin Kim, Sihyun Yu, Jaeho Lee, and Jinwoo Shin. Scalable neural video representations with learnable positional features. NeurIPS, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[34] Doyup Lee, Chiheon Kim, Minsu Cho, and WOOK SHIN HAN. Locality-aware generalizable implicit neural representation. NeurIPS, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[35] Jeffrey Li, Mikhail Khodak, Sebastian Caldas, and Ameet Talwalkar. Differentially private meta-learning. In ICLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[36] Qinbin Li, Yiqun Diao, Quan Chen, and Bingsheng He. Federated learning on non-iid data silos: An experimental study. In IEEE International Conference on Data Engineering, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.482, + 0.532 + ], + "angle": 0, + "content": "[37] Shibo Li, Zheng Wang, Akil Narayan, Robert Kirby, and Shandian Zhe. Meta-learning with adjoint methods. In AISTATS, pages 7239-7251. PMLR, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.482, + 0.574 + ], + "angle": 0, + "content": "[38] Xiang Li, Kaixuan Huang, Wenhao Yang, Shusen Wang, and Zhihua Zhang. On the convergence of fedavg on non-iid data. arXiv:1907.02189, 2019. 7, 1, 11, 12, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.482, + 0.616 + ], + "angle": 0, + "content": "[39] Zhenguo Li, Fengwei Zhou, Fei Chen, and Hang Li. Meta-sgd: Learning to learn quickly for few-shot learning. arXiv:1707.09835, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.482, + 0.659 + ], + "angle": 0, + "content": "[40] Susan Liang, Chao Huang, Yapeng Tian, Anurag Kumar, and Chenliang Xu. Av-nerf: Learning neural fields for real-world audio-visual scene synthesis. NeurIPS, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.716 + ], + "angle": 0, + "content": "[41] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In AISTATS, 2017. 1, 3, 5, 7, 8, 6, 10, 11, 12, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.772 + ], + "angle": 0, + "content": "[42] William McNally, Kanav Vats, Tyler Pinto, Chris Dulhanty, John McPhee, and Alexander Wong. Golfdb: A video database for golf swing sequencing. In CVPRW, 2019. 5, 6, 3, 4, 7, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.774, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[43] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 2021. 2, 5, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[44] Alex Nichol, Joshua Achiam, and John Schulman. On first-order meta-learning algorithms. arXiv:1803.02999, 2018. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[45] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learn" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.905, + 0.12 + ], + "angle": 0, + "content": "ing implicit 3d representations without 3d supervision. In CVPR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[46] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[47] Vishwanath Saragadam, Jasper Tan, Guha Balakrishnan, Richard G Baraniuk, and Ashok Veeraraghavan. Miner: Multiscale implicit neural representation. In ECCV, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.276 + ], + "angle": 0, + "content": "[48] Bipasha Sen, Aditya Agarwal, Vinay P Namboodiri, and CV Jawahar. Inr-v: A continuous representation space for videobased generative tasks. arXiv:2210.16579, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[49] Risa Shinoda and Kaede Shiohara. Petface: A large-scale dataset and benchmark for animal identification. In ECCV, 2025. 5, 6, 3, 4, 7, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.376 + ], + "angle": 0, + "content": "[50] Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. Membership inference attacks against machine learning models. In IEEE Symposium on Security and Privacy, pages 3-18. IEEE, 2017. 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.378, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[51] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. NeurIPS, 2020. 2, 5, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.475 + ], + "angle": 0, + "content": "[52] Ivan Skorokhodov, Savva Ignatyev, and Mohamed Elhoseiny. Adversarial generation of continuous images. In CVPR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.905, + 0.505 + ], + "angle": 0, + "content": "[53] Kun Su, Mingfei Chen, and Eli Shlizerman. Inras: Implicit neural representation for audio scenes. NeurIPS, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.507, + 0.905, + 0.546 + ], + "angle": 0, + "content": "[54] Terpeei Suzuki. Federated learning for large-scale scene modeling with neural radiance fields. arXiv:2309.06030, 2023.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.905, + 0.576 + ], + "angle": 0, + "content": "[55] Terpei Suzuki. Fed3dgs: Scalable 3d gaussian splatting with federated learning. arXiv:2403.11460, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.578, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[56] Jihoon Tack, Subin Kim, Sihyun Yu, Jaeho Lee, Jinwoo Shin, and Jonathan Richard Schwarz. Learning large-scale neural fields via context pruned meta-learning. NeurIPS, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.634, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[57] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. NeurIPS, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.905, + 0.76 + ], + "angle": 0, + "content": "[58] Matthew Tancik, Ben Mildenhall, Terrance Wang, Divi Schmidt, Pratul P Srinivasan, Jonathan T Barron, and Ren Ng. Learned initializations for optimizing coordinate-based neural representations. In CVPR, 2021. 5, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[59] Zaid Tasneem, Akshit Dave, Abhishek Singh, Kushagra Tiwary, Praneeth Vepakomma, Ashok Veeraraghavan, and Ramesh Raskar. Decentnerfs: Decentralized neural radiance fields from crowdsourced images. arXiv:2403.13199, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[60] Ayush Tewari, Ohad Fried, Justus Thies, Vincent Sitzmann, Stephen Lombardi, Kalyan Sunkavalli, Ricardo MartinBrualla, Tomas Simon, Jason Saragih, Matthias Nießner, et al. State of the art on neural rendering. In Computer Graphics Forum, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "2170" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[61] Jianyu Wang, Qinghua Liu, Hao Liang, Gauri Joshi, and H Vincent Poor. Tackling the objective inconsistency problem in heterogeneous federated optimization. NeurIPS, 2020. 7, 1, 11, 12, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[62] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 2004. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[63] Kang Wei, Jun Li, Ming Ding, Chuan Ma, Howard H. Yang, Farhad Farokhi, Shi Jin, Tony Q. S. Quek, and H. Vincent Poor. Federated learning with differential privacy: Algorithms and performance analysis. IEEE Transactions on Information Forensics and Security, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[64] Yuexiang Xie, Zhen Wang, Dawei Gao, Daoyuan Chen, Liuyi Yao, Weirui Kuang, Yaliang Li, Bolin Ding, and Jingren Zhou. Federatedscope: A flexible federated learning platform for heterogeneity. arXiv:2204.05011, 2022. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[65] Haotian Yang, Hao Zhu, Yanru Wang, Mingkai Huang, Qiu Shen, Ruigang Yang, and Xun Cao. Facescape: a large-scale high quality 3d face dataset and detailed riggable 3d face prediction. In CVPR, 2020. 5, 6, 3, 7, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.444 + ], + "angle": 0, + "content": "[66] Kai Yue, Richeng Jin, Chau-Wai Wong, Dror Baron, and Huaiyu Dai. Gradient obfuscation gives a false sense of security in federated learning. In USENIX Security Symposium, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.446, + 0.482, + 0.488 + ], + "angle": 0, + "content": "[67] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.482, + 0.529 + ], + "angle": 0, + "content": "[68] Yintian Zhang and Ziyu Shao. Federated neural radiance field for distributed intelligence. arXiv:2406.10474, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.532, + 0.482, + 0.56 + ], + "angle": 0, + "content": "[69] Xinyu Zhou and Raef Bassily. Task-level differentially private meta learning. NeurIPS, 2022. 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.629 + ], + "angle": 0, + "content": "[70] Hao Zhu, Haotian Yang, Longwei Guo, Yidi Zhang, Yanru Wang, Mingkai Huang, Menghua Wu, Qiu Shen, Ruigang Yang, and Xun Cao. Facescape: 3d facial dataset and benchmark for single-view 3d face reconstruction. IEEE TPAMI, 2023. 5, 6, 3, 7, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.482, + 0.659 + ], + "angle": 0, + "content": "[71] Yiyu Zhuang, Hao Zhu, Xusen Sun, and Xun Cao. Mofanerf: Morphable facial neural radiance field. In ECCV, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.659 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.956 + ], + "angle": 0, + "content": "2171" + } + ] +] \ No newline at end of file diff --git a/2025/FedMeNF_ Privacy-Preserving Federated Meta-Learning for Neural Fields/97cb26d5-5659-4687-b57e-cc32fd64dcc4_origin.pdf b/2025/FedMeNF_ Privacy-Preserving Federated Meta-Learning for Neural Fields/97cb26d5-5659-4687-b57e-cc32fd64dcc4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8a1392d713e169927ca9fdfe5564dd0e82a88c85 --- /dev/null +++ b/2025/FedMeNF_ Privacy-Preserving Federated Meta-Learning for Neural Fields/97cb26d5-5659-4687-b57e-cc32fd64dcc4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6207f2bd16324acc65edada98187d0516ca9263965297ab51e2dd088023d4465 +size 2390231 diff --git a/2025/FedMeNF_ Privacy-Preserving Federated Meta-Learning for Neural Fields/full.md b/2025/FedMeNF_ Privacy-Preserving Federated Meta-Learning for Neural Fields/full.md new file mode 100644 index 0000000000000000000000000000000000000000..094170d0a4a220c65eb76e28fd2131bb7ac87b48 --- /dev/null +++ b/2025/FedMeNF_ Privacy-Preserving Federated Meta-Learning for Neural Fields/full.md @@ -0,0 +1,494 @@ +# FedMeNF: Privacy-Preserving Federated Meta-Learning for Neural Fields + +Junhyeog Yun Minui Hong Gunhee Kim Seoul National University + +{junhyeog,alsdml123,gunhee}@snu.ac.kr + +https://github.com/junhyeog/FedMeNF + +# Abstract + +Neural fields provide a memory-efficient representation of data, which can effectively handle diverse modalities and large-scale data. However, learning to map neural fields often requires large amounts of training data and computations, which can be limited to resource-constrained edge devices. One approach to tackle this limitation is to leverage Federated Meta-Learning (FML), but traditional FML approaches suffer from privacy leakage. To address these issues, we introduce a novel FML approach called FedMeNF. FedMeNF utilizes a new privacy-preserving loss function that regulates privacy leakage in the local meta-optimization. This enables the local meta-learner to optimize quickly and efficiently without retaining the client's private data. Our experiments demonstrate that FedMeNF achieves fast optimization speed and robust reconstruction performance, even with few-shot or non-IID data across diverse data modalities, while preserving client data privacy. + +# 1. Introduction + +Neural Fields (NFs), also known as Implicit Neural Representations (INRs), use a deep neural network to approximate continuous signals [12, 46, 47, 57, 60]. They enable memory-efficient representation and fast processing of data [14]. Thanks to this, NFs effectively handle both large-scale data and diverse modalities, including audio, images, videos, and 3D objects [10, 11, 33, 40, 45, 48, 52, 53, 71]. + +However, optimizing NFs is often slow and requires large amounts of data, which can be a serious issue on resource-constrained edge devices. Consider a practical scenario: a user takes a few photos of an object using their phone and wants to get a high-fidelity 3D model of it. Training an NF from scratch with such limited data and resources would be time-consuming and likely yield poor results. One approach to address this issue is meta-learning [19, 44], which aims to train a meta-learner that can quickly adapt to new tasks with only a few samples. Using the meta-learner, the user can quickly optimize a high-performing NF for the + +new object with just a few images. + +Nevertheless, meta-learning itself requires a large and diverse set of tasks to learn an effective initialization or optimization strategy. Unfortunately, in many practical scenarios, users only have data for one or a few tasks. For example, most users typically take photos of only a single instance of privacy-sensitive objects such as their face, body, pet, or car. To enable meta-learning in such scenarios, we can leverage Federated Meta-Learning (FML) [8, 18, 29], which collaboratively trains a global meta-learner using private data from numerous distributed clients. FML builds upon Federated Learning (FL) [4, 13, 27, 36, 41], a paradigm designed to train a global model while ensuring client data remains on local devices to preserve privacy. + +Yet, applying FML to NFs introduces a critical challenge that fundamentally undermines the privacy guarantee of FL. The privacy protection in traditional FL relies on the assumption that sharing trained model parameters, rather than the raw data itself, is safe. This assumption completely breaks down for NFs for two main reasons. First, in our scenario, each client has data for only a single or a few task instances (e.g., a single image or 3D object). If even one client has only a single task instance, the local meta-optimization is like standard supervised training on that specific task [26]. Consequently, the trained local meta-learner functions as an optimized NF for that private data. Second, NFs are, in essence, a compressed representation of the data and inherently encapsulate it. Thus, even if only the NF parameters are shared, a malicious server can reconstruct the private data. Due to this inherent privacy leakage issue, existing studies often limit FL applications for NFs to public data such as landmarks [59] or cityscapes [24, 55, 68], where privacy is less of a concern. For these reasons, naively applying FML methods to private data results in severe privacy issues, as the server can reconstruct the private data from the shared local meta-learner. + +To address this critical challenge, we propose FedMeNF (Federated Meta-Learning for Neural Fields), a novel FML framework designed to collaboratively train a global NF meta-learner while preserving privacy as overviewed in + +![](images/0f30597defe4686a7a2bfcaad6c3c63efe9343e25e9f8260a02cd3d6c3d69952.jpg) +(a) FML with a single task per client. + +![](images/8888ecd7e33472cc1c28fe1cdf03739c13a8500a76df8f359b1d4f398a5d83f2.jpg) +(b) Comparison of FedMeNF with Local and FML in terms of fast or few-shot optimization on a new task, and privacy preservation. + +![](images/7be94016b2bbba070827b47748f3e92025c57419bf10938eaebd3854450c20bc.jpg) +(c) The client has photos of only one pet, car, body, or face, and aims to optimize a neural field for each instance. + +![](images/c582a7d216a6f056675d0688f116fbff6e3783c3f9bca8835688f5337e990bb6.jpg) + +![](images/4f24c78639b3839c85ef30ca23fd2f9c6285513cbc897f1be62349485ec3d24f.jpg) +(d) FedMeNF with a single task per client. +Figure 1. Overview of our approach. In each communication round, the server sends the global meta-learner $\theta$ to clients, each of which uses its private data to train a local meta-learner $w$ based on the global meta-learner. The trained local meta-learners are then sent back to the server to update the global meta-learner. (a) In existing federated meta-learning (FML) approaches, the local meta-learner contains privacy-sensitive information when a client has only a single task, as shown in (c). The server can infer the client's private data from the shared local meta-learner. (d) In contrast, our FedMeNF ensures that the local meta-learner only learns how to efficiently optimize NFs while preserving privacy-sensitive details. (b, c) Local refers to using only the client's data without sharing any information with the server. + +Fig. 1, which then empowers each client to rapidly optimize a high-performance NF even with a small amount of new private data. FedMeNF can be utilized in various privacy-sensitive applications. For example, suppose that a user buys a new car and takes a few pictures of it. We can then train the NF of the car with this small set of images using a global meta-learner trained with other car images in other devices. As another example, with only a handful of facial images, users can employ the global facial NF meta-learner to generate images showing a range of facial expressions. Also, they can upscale low-resolution photos to high-resolution or generate higher frame-rate videos from personal data, such as pet photos or golf swing videos. + +To quantify and control privacy leakage, we define a privacy metric, $\mathrm{PSNR}_p$ . We theoretically analyze how $\mathrm{PSNR}_p$ increases under existing FML approaches. By regularizing the increase of $\mathrm{PSNR}_p$ , FedMeNF forces the local meta-learner to learn only how to optimize the NF faster with limited data without memorizing the private data. We demonstrate that FedMeNF achieves performance comparable to state-of-the-art methods while better preserving privacy across diverse data modalities, including images, videos, and neural radiance fields (NeRFs). Additionally, FedMeNF shows robust performance even when local data are very few or highly diverse among clients. + +Our contributions can be outlined as follows: + +1. This is the first study on FL for NFs on private data. +2. We theoretically and empirically show how privacy leakage occurs during the FML for NFs. +3. We propose FedMeNF that preserves the privacy of local data with minimal impact on optimization speed and reconstruction quality. +4. We conduct comprehensive experiments on FedMeNF across various data modalities, private data sizes, and levels of data diversity, outperforming baseline methods. + +# 2. Preliminaries + +Coordinate-based Neural Fields. Coordinate-based neural fields [3, 9, 14, 43, 51, 56, 57] represent continuous functions that map spatial coordinates to signal values such as color or density. The objective for each client $m$ is to find the optimal NF parameter $\varphi_{i}^{m}$ that best represents its $i$ -th local task data $T_{i}^{m}$ . This is formulated as + +$$ +\varphi_ {i} ^ {m} = \arg \min _ {\varphi} \sum_ {(\mathbf {x} _ {j}, \mathbf {y} _ {j}) \in T _ {i} ^ {m}} L \left(f _ {\varphi} (\mathbf {x} _ {j}), \mathbf {y} _ {j}\right), \tag {1} +$$ + +where $f_{\varphi}$ is the forward map parameterized by $\varphi$ , $\mathbf{x}_j$ represents the input coordinates (e.g., spatial positions), and $\mathbf{y}_j$ + +denotes the target values (e.g., pixel colors). $L$ is a suitable loss function such as mean squared error (MSE). + +Federated Learning. We consider a centralized federated learning (FL) environment [2, 41], where clients do not share their data with either the server or other clients and communicate the model parameters exclusively with the centralized server. The centralized FL aims to optimize the global model that minimizes a given objective function as + +$$ +\arg \min _ {\theta} \sum_ {m = 0} ^ {M - 1} \alpha^ {m} L (\theta , D ^ {m}), \text {w h e r e} \alpha^ {m} = \frac {\left| D ^ {m} \right|}{\Sigma_ {i} \left| D ^ {i} \right|}. \tag {2} +$$ + +$\theta$ is the parameter of the global meta-learner and $M$ is the number of clients. $D^{m}$ is the local dataset of client $m$ . $\alpha^{m}$ is the proportion of the number of data samples of client $m$ . $L(\theta, D^{m})$ represents the loss of model $\theta$ on a dataset $D^{m}$ . This objective ensures that the global model $\theta$ is optimized across all clients by weighting the contribution of each client based on its local dataset size. + +Federated Meta-learning. Federated meta-learning [7, 8, 18, 29] is a variant of FL where the client trains the local model using a meta-learning approach with the objective: + +$$ +\arg \min _ {\theta} \sum_ {m = 0} ^ {M - 1} \alpha^ {m} L \left(\varphi^ {m}, Q ^ {m}\right), \tag {3} +$$ + +$$ +\text {w h e r e} \varphi^ {m} = \arg \min _ {\theta} L (\theta , S ^ {m}). \tag {4} +$$ + +$S^m$ and $Q^m$ are the support set and query set of the client $m$ . Each client $m$ first trains the local model $\varphi^m$ on the support set $S^m$ (inner loop). Then, starting from the global model $\theta$ , the local meta-learner minimizes the loss of $\varphi^m$ on the query set $Q^m$ (outer loop). This local meta-optimization enables the local meta-learner to adapt quickly to new tasks. The trained local meta-learner is shared with the server, which aggregates the local meta-learners received from participating clients to update the global meta-learner. + +Differential Privacy. Differential privacy (DP) is one of the most popular privacy frameworks with rich theoretical guarantees [1, 5, 15, 23, 35, 63]. It guarantees that when a single entry in the database changes, the change in the probability distribution of the algorithm outputs is bounded by predetermined factors $\epsilon$ and $\delta$ [16]. The privacy risk can be regulated by controlling these factors. However, traditional DP methods are designed for large datasets; yet neural fields often capture just a single signal. This causes ambiguity in defining privacy leakage for individual training samples, as well as a significant trade-off between privacy preservation and reconstruction performance. + +# 3. Problem Statement + +In this work, we assume a scenario where multiple users collaboratively train a global neural field meta-learner using + +federated meta-learning. Each user has data about a personal object stored locally on their device. For instance, each user holds a few images of his or her own car on their device. Our goal is to optimize neural fields despite limited data and low computational resources on local devices. Thus, optimizing a specific neural field becomes a local task. The local task data are divided into a support set and a query set, each containing images of the object. During training, the support set is used to train the neural field, while the query set is utilized to optimize the local meta-learner. In the testing phase, the support set enables test-time optimization, and the query set is used to assess its performance. The detailed FML process is described in Fig. 1a and Algorithm 1, excluding the highlighted parts which will be discussed in Sec. 5. + +# 3.1. First-order Approximation of Meta-gradients + +During the local meta-optimization, the local meta-learner $w$ is updated for each outer step $i$ , where task data $T_{i} = \{S_{i}, Q_{i}\}$ are sampled from the client's training dataset $D_{\mathrm{train}}$ . Then, the local NF $\varphi$ , initialized by $w_{i}$ , are optimized with the support set $S_{i}$ (Eq. (8)), and the local meta-learner $w_{i}$ are optimized with the query set $Q_{i}$ (Eq. (9)). + +Proposition 1. Let $g_{M} = \nabla_{w_{i}}L(\varphi_{K},B_{K})$ and $g_{k} = \nabla_{\varphi_{0}}L(\varphi_{0},B_{k})$ . Then, the first-order approximation of the meta-gradient $g_{M}$ becomes + +$$ +g _ {M} \approx g _ {K} - \lambda_ {i} \mathcal {I} _ {K}, w h e r e \mathcal {I} _ {K} = \sum_ {k = 0} ^ {K - 1} \nabla_ {\varphi_ {0}} \left\langle g _ {K}, g _ {k} \right\rangle . \tag {5} +$$ + +$K$ is the # of inner steps, $B_{0\dots K - 1}$ are minibatches of $S_{i}$ , $B_{K}$ is a minibatch of $Q_{i}$ , and $\lambda_{i}$ is an inner loop learning rate. The proof is provided in Appendix H. + +# 3.2. Fast Optimization via $\mathcal{I}_K$ + +According to Proposition 1, the meta-gradient $g_{M}$ can be expressed with $g_{K}$ and the inner product term $\mathcal{I}_K$ , which is the sum of inner products between $g_{K}$ and $g_{0},\dots ,g_{K - 1}$ . This term maximizes (i) the alignment of the gradients from the query set $(g_{K})$ with those from the support set $(g_0,\dots ,g_{K - 1})$ , and (ii) the alignment of the gradients computed on different minibatches within the support set. This enables the local meta-learner (i) to be more generalized and (ii) to quickly optimize the neural field for a new task. + +# 4. Privacy Leakage in Local Meta-Learning + +We define $\mathrm{PSNR}_p$ , an intuitive privacy metric to quantify the privacy leakage in FML for neural fields: + +$$ +\mathrm {P S N R} _ {p} = \mathrm {P S N R} \left(Q ^ {m}, f _ {w} \left(\operatorname {C o o r d} \left(Q ^ {m}\right)\right)\right), \tag {6} +$$ + +where $Q^m$ is the query set of client $m$ , $\operatorname{Coord}(Q^m)$ is the coordinates of $Q^m$ , and $f_w$ is the forward map parameterized by the local meta-learner $w$ . $\mathrm{PSNR}_p$ assesses how well + +the server can reconstruct $Q^{m}$ with the shared local metalearner $w$ . It is a simple variant of PSNR that measures similarity between images reconstructed with $w$ on the server and the client's local images. A better reconstruction indicates a greater violation of the client's data privacy, meaning that a larger $\mathrm{PSNR}_p$ indicates more privacy leakage. + +We define our main privacy metric based on PSNR for two reasons. First, neural fields can handle various types of signals such as vision, speech, or sensor data, so does the PSNR. Second, PSNR is one of the most popular metrics for evaluating reconstruction attacks in FL [21, 25, 66], as well as reconstruction tasks [9, 31]. As variants, we also define $\mathrm{SSIM}_p$ and $\mathrm{LPIPS}_p$ as privacy metrics to assess structural and perceptual similarities [62, 67]; it can overcome the limitation of $\mathrm{PSNR}_p$ that is based on only pixel-wise MSE. Please refer to Appendix F for more details. + +# 4.1. Privacy Leakage via $g_{K}$ + +Since $\mathrm{PSNR}_p$ is defined by $10\log_{10}(R / L(w,Q^m))$ where $R$ is a constant, it becomes larger as the MSE loss $L(w,Q^{m})$ is smaller. We analyze how $L(w,Q^{m})$ changes during the client's local meta-optimization, which helps understand how $\mathrm{PSNR}_p$ evolves. + +Proposition 2. Let $\Delta L_{i + 1} = L(w_{i + 1},B_K) - L(w_i,B_K)$ . The first-order approximation of $\Delta L_{i + 1}$ is then + +$$ +\Delta L _ {i + 1} \approx - \lambda_ {o} \cdot g _ {K} ^ {2} \leq 0. \tag {7} +$$ + +The proof is provided in Appendix I. Note that $Q^m$ is replaced by $B_K$ , a minibatch of the $Q^m$ . $\Delta L_{i+1}$ denotes the change in the loss of the local meta-learner $w$ on the client's query set. According to Proposition 2, $\Delta L_{i+1}$ is always negative, which means that $L(w, Q^m)$ decreases with each outer loop iteration. Since $\mathrm{PSNR}_p$ is inversely proportional to $L(w, Q^m)$ , $\mathrm{PSNR}_p$ increases every outer loop iteration. Consequently, to prevent the increase of $\mathrm{PSNR}_p$ , we regularize $\Delta L_{i+1}$ to be close to zero, thereby mitigating the privacy leakage in the local meta-optimization. + +# 5. Approach + +We reveal that federated meta-learning (FML) approaches suffer from privacy leakage as their local meta-gradients increase the privacy metric. To address this issue, we propose FedMeNF, a novel FML framework for neural fields that ensures privacy-preserving local meta-optimization. FedMeNF employs a privacy-preserving loss function $L_{pp}$ , designed to regulate the increase in the privacy metric. Please refer to Algorithm 1 and Fig. 1d for its workflow. + +# 5.1. The Privacy-Preserving Loss + +Proposition 1-2 theoretically show how client data privacy is leaked in federated meta-learning. Now, we introduce + +Algorithm 1 FedMeNF (see Appendix A for notation) + +Input: (1) $R$ : # of communication rounds, (2) $N$ : # of clients, (3) $M$ : # of participants, (4) $\lambda_{i}$ : inner loop learning rate, (5) $\lambda_{o}$ : outer loop learning rate, (6) $\alpha^{m}$ : weight proportional to the client $m$ 's dataset size (see Eq. (2)). + +Initialize the global meta-learner $\theta_0$ + +for each communication round $r = 0,\dots,R - 1$ do + +Sample $M$ clients from 1,..., $N$ clients + +for $m = 1,\dots,M$ do + +$$ +w _ {*} ^ {m} \leftarrow \text {L O C A L U P D A T E} \left(\theta_ {r}\right) +$$ + +end for + +$$ +\theta_ {r + 1} \leftarrow \Sigma_ {m} \alpha^ {m} w _ {*} ^ {m} \quad \triangleright \text {S e r v e r a g g r e g a t i o n} +$$ + +end for + +procedure LOCALUPDATE(θ) + +Initialize the local meta-learner $w_0 \gets \theta$ + +for each step $i = 0,\dots,E - 1$ do $\triangleright$ Outer loop + +Sample task data $T_{i} = \{S_{i},Q_{i}\}$ from $D_{\mathrm{train}}$ + +$\varphi_0 \gets w_i$ Initialize the neural field for $T_{i}$ + +for each step $k = 0,\dots,K - 1$ do $\triangleright$ Inner loop + +Sample batch $B_{k}$ from the support set $S_{i}$ + +$$ +\varphi_ {k + 1} \leftarrow \varphi_ {k} - \lambda_ {i} \nabla_ {\varphi_ {k}} L \left(\varphi_ {k}, B _ {k}\right) \tag {8} +$$ + +end for + +Sample batch $B_{K}$ from the query set $Q_{i}$ + +Select the regularization coefficient $\gamma$ + +$$ +w _ {i + 1} \leftarrow w _ {i} - \lambda_ {o} \nabla_ {w _ {i}} \left(L \left(\varphi_ {K}, B _ {K}\right) - \gamma L \left(w _ {i}, B _ {K}\right)\right) +$$ + +$$ +\mid \text {f o r} \tag {9} +$$ + +end for + +Send $w_{E}$ + +end procedure + +a privacy-preserving loss $L_{pp}$ as follows, ensuring that it regularizes the $g_{K}$ term that increases $\mathrm{PSNR}_p$ : + +$$ +L _ {p p} (\gamma , w _ {i}, \varphi_ {K}, B _ {K}) = L \left(\varphi_ {K}, B _ {K}\right) - \gamma L \left(w _ {i}, B _ {K}\right), \tag {10} +$$ + +where $\gamma$ is a regularization coefficient that determines the portion of the $g_{K}$ . In Proposition 3 below, we can obtain the privacy-preserving local meta-gradient $g_{pp}$ using $L_{pp}$ . + +Proposition 3. Let $g_{pp} = \nabla_{w_i}L_{pp}(\gamma, w_i, \varphi_K, B_K)$ . Then, the first-order approximation of $g_{pp}$ and $\Delta L_{i+1}$ become + +$$ +g _ {p p} \approx (1 - \gamma) \cdot g _ {K} - \lambda_ {i} \mathcal {I} _ {K}, \tag {11} +$$ + +$$ +\Delta L _ {i + 1} \approx - \lambda_ {o} (1 - \gamma) \left(g _ {K}\right) ^ {2} \leq 0. \tag {12} +$$ + +Note $\mathcal{I}_K$ is defined in Eq. (5). At $\gamma = 1$ , the $g_K$ term, which causes privacy leakage, is completely removed from $g_{pp}$ . This means that the local meta-learner no longer contains information about the client's local dataset and learns only fast optimization by maximizing the inner products between gradients across minibatches. Each client can choose + +$\bullet$ local meta-learner + +$E$ : # of outer loop steps + +$= = \gg$ $\nabla_{w_i}L(\varphi_K,B_K)$ + +$\longrightarrow \nabla_{\varphi_k}L(\varphi_k,B_k)$ + +$\star$ optimal neural field +i: outer loop step +- $\triangleright -\gamma \nabla_{w_i}L(w_i,B_K)$ +$\Rightarrow \nabla_{w_i}L_{pp}(\gamma ,w_i,\varphi_K,B_K)$ + +![](images/c3fd9092c10ec35fac0117b7357c7502aee7e931e6f9f90a978305c8e1e3d0a9.jpg) +(a) $i = 0$ + +![](images/37f204dc2ac283e8215ef575a521f8a84a48ac47afaf0bc4e0b876f38b8c2f1e.jpg) +(b) $i = 1$ + +![](images/05e671ad1b7b490da56a8027cb584a26696148948f55ca2d834b81290292eee3.jpg) +(c) $i = E - 1$ +Figure 2. Illustration of FedMeNF's local meta-optimization in a client. The local meta-learner $w$ does not get closer to the optimal neural field and maintains the initial gap (the distance between the global meta-learner $\theta$ ( $= w_0$ ) and the optimal neural field). Nevertheless, as meta-optimization progresses, gradients are being aligned, and the local meta-learner can find $w_E$ from which the optimal neural field is reached with a few gradient steps. + +a different $\gamma$ that aligns with its desired privacy guarantee. We will empirically show that as $\gamma$ increases, the privacy leakage from the client's local dataset decreases in Sec. 6.2. + +There are two ways to set $\gamma$ : as a fixed hyperparameter or adjusted dynamically. While we use the former approach as our default, we also propose the following adaptive method. Similar to how $\epsilon$ defines the privacy boundary in Differential Privacy, we introduce a privacy budget $\zeta$ , limiting the total magnitude of the loss change defined in Eq. (12): + +$$ +\left| \Delta L _ {i + 1} \right| \cdot R \cdot E \cdot M / N \leq \zeta , \tag {13} +$$ + +$$ +\gamma = \min \left(\max \left(1 - N \zeta / \left(R E M \lambda_ {o} \left(g _ {K}\right) ^ {2}\right), 0\right), 1\right). \tag {14} +$$ + +# 5.2. Privacy-Preserving Local Meta-Optimization + +Figure 2 illustrates FedMeNF's local meta-optimization. In existing FML approaches, each outer loop iteration brings the local meta-learner $w_{i}$ closer to the optimal neural field for the client's private data. Sharing the optimized local meta-learner $w_{E}$ with the server leads to significant privacy leakage, as the server can directly exploit $w_{E}$ to reconstruct the original private data. In contrast, FedMeNF explicitly prevents the $w_{i}$ from converging to the optimal neural field, focusing solely on gradient alignment. Thus, FedMeNF trains a local meta-learner that retains minimal private data while still enabling efficient optimization. + +# 6. Experiments + +Setup. We assume a scenario where $N = 50$ clients participate in the training. The training consists of 1000 communication rounds, with 5 participants randomly selected + +![](images/01816b1cae4d92a36ee0cc792a50f019324ce6f48a8a288ccab53136c2f4380a.jpg) +Figure 3. Test-time optimization of the baselines and our Fed-MeNF $(\gamma = 0.75)$ using FedAvg [41] on the Cars dataset [6]. + +![](images/de1d98148821e44e2ee6e5166424d2fbbd83f8109cb520b9a3c8957748e1a33d.jpg) +(a) $\mathrm{PSNR}_p / \mathrm{Rounds}$ +Figure 4. Comparison of privacy preservation between baselines and our FedMeNF with different $\gamma$ values using FedAvg [41] on the Cars dataset [6]. Baselines like MAML, which excel in reconstruction (PSNR $\uparrow$ ), exhibit high PSNR $_p$ value, while methods like Reptile, which have lower PSNR $_p$ value, perform poorly in reconstruction (PSNR $\downarrow$ ). FedMeNF effectively adjusts $\gamma$ to maintain an acceptable PSNR $_p$ value while achieving a comparable PSNR. + +![](images/240aff57adbb309f2e7855b1b799977f02148224b13c1c58187e2665571a6fcb.jpg) +(b) PSNR / $\mathrm{PSNR}_p$ + +per round. We employ the SIREN architecture [51] for image and video modalities, which consists of 6 layers with a hidden dimension of 128. For neural radiance fields (NeRF) [43], we adopt the model used in [9, 58], which consists of 6 layers with a hidden dimension of 256. Please refer to Appendix D for more implementation details. + +Datasets. We conduct experiments across the datasets of various modalities, including images, videos, and NeRFs. To evaluate performance with limited local training data per client, we select scenarios in which each client has only one or very few tasks. For images, we use the cat category from the PetFace dataset [49], assuming each client has images of a unique cat instance. For videos, we use the GolfDB dataset [42], where each client has videos of only one person's golf swings. For NeRFs, we use the Cars category from the ShapeNet dataset [6] and the FaceScape dataset [65, 70] for human faces. Please refer to Appendix E for detailed distribution information for each dataset. + +Baselines. We test a lower bound baseline referred to as Local, using only local dataset without an FL approach. We construct baselines by combining an FL algorithm and a meta-learning algorithm [8, 18]. For the FL algo + +
Modality DatasetImage PetFace [49]Video GolfDB [42]3D (NeRF) Cars [6]FaceScape [65, 70]
Method\Metric\(PSNR_p\)\(PSNR\)\(\Delta(↑)\)\(PSNR_p\)\(PSNR\)\(\Delta(↑)\)\(PSNR_p\)\(PSNR\)\(\Delta(↑)\)\(PSNR_p\)\(PSNR\)\(\Delta(↑)\)
Local-22.29--26.92--17.13--23.67-
FedAvg + MAML16.5727.3910.8221.2129.688.4719.7323.083.3521.3128.597.28
+ FOMAML18.5223.154.6320.8228.577.7519.7323.663.9321.2428.657.41
+ Reptile17.3922.525.1319.8927.227.3319.9621.982.0221.9228.246.32
+ meta-NSGD12.495.15-7.3410.964.85-6.116.8510.623.777.7211.293.57
+ Ours14.772712.2317.3128.8911.5812.1524.0511.915.1627.8812.72
FedProx + MAML16.5827.4910.9120.9329.268.3319.6722.943.2721.128.547.44
+ FOMAML18.5324.325.7921.2128.767.5519.6723.734.0621.0228.527.5
+ Reptile17.3722.55.1319.7927.217.4219.9522.112.1621.8928.216.32
+ meta-NSGD12.495.15-7.3410.964.84-6.126.8510.914.067.7211.293.57
+ Ours14.4427.0812.6415.6828.9613.2812.1423.9811.8414.0927.5413.45
Scaffold + MAML16.7127.6610.9521.329.117.8119.0724.215.1421.0928.517.42
+ FOMAML18.5523.985.4321.2928.827.5319.0724.475.421.0128.517.5
+ Reptile17.3722.475.12027.237.2319.8122.62.7921.7928.146.35
+ meta-NSGD12.495.13-7.3610.964.84-6.126.8510.633.787.7211.293.57
+ Ours14.9327.1912.2616.2129.2313.0214.1424.349.9413.9427.5313.59
FedNova + MAML16.5227.5110.9921.4729.648.1719.7223.633.9121.3728.597.22
+ FOMAML18.524.15.620.1228.698.5719.7223.273.5521.328.627.32
+ Reptile17.3822.525.1419.9327.267.3319.9622.72.7421.9828.176.19
+ meta-NSGD12.495.14-7.3510.964.84-6.126.8510.63.757.7211.293.57
+ Ours14.9427.1512.2115.7128.9813.2712.1424.1211.9815.4527.8612.41
FedExP + MAML16.5727.3910.8221.1829.588.419.8122.873.0621.3327.646.31
+ FOMAML18.5223.154.6320.8228.67.7819.8122.81321.2627.666.4
+ Reptile17.3922.645.2519.8927.237.3420.9122.031.1221.9328.276.35
+ meta-NSGD12.495.15-7.3410.964.85-6.116.8510.613.767.7211.293.57
+ Ours14.6526.8612.2115.3828.9413.5612.1724.0511.8813.6926.1212.43
FedACG + MAML16.6327.510.8721.0929.518.4219.9222.122.2828.025.74
+ FOMAML18.4824.045.5621.1228.77.5819.921.952.0522.2428.035.79
+ Reptile17.3822.635.2519.8327.237.420.1322.262.1322.0728.326.25
+ meta-NSGD12.495.15-7.3410.964.84-6.126.8510.593.747.7211.293.57
+ Ours14.7126.9712.2615.9428.9913.0510.9322.4511.5214.8327.4612.63
+ +Table 1. Comparison of reconstruction (PSNR(↑)) and privacy (PSNRp(↓)) quality across different modalities: images, videos, and 3D NeRF scenes. $\Delta$ is defined as PSNR - PSNRp where a larger value indicates a better trade-off between reconstruction quality and privacy protection. FedMeNF demonstrates a balanced performance by maintaining high PSNR for reconstruction and low PSNRp for privacy. The redhighlighted cases indicate cases where PSNRp is high enough to qualitatively confirm privacy leakage, and the gray-highlighted cases indicate cases where the reconstruction performance is lower than that of the Local. + +![](images/30bfe43aa9058b497e70d6d564c071706d83e3cc81509f8685758ae286416718.jpg) +(a) GT (PSNR) + +![](images/6a8ed5e90f4834cbac859f71963801e160942e0363786dc1f182e01b98a3bd6d.jpg) +(b) Local (14.97) + +![](images/9bca89313a2bedfb1d98a8609761513392c458211f8a5da97ddb444de2a5548f.jpg) +(c) MAML (21.22) + +![](images/2b89c157dc39217a7d87a3f5b6a3892500bb6c8aa7f0b7dad9e5c2ab52d13728.jpg) +(d) FOMAML (21.44) + +![](images/abe7be2f581da1b4ee89431eb1cae6b683ff13b1110504d818a7caf50f62c472.jpg) +(e) Reptile (21.18) + +![](images/0d9655a6a68b363a9c951cfb5ea19449d40bb61aa53311eac4a569098053cafb.jpg) +(f) FedMeNF (21.92) + +![](images/adf67963b6502cdaed9f62900c7ae0272f474411b89c6a450b9f640fa9199e09.jpg) +(g) GT (PSNR) + +![](images/02946aba121673d9b6fcdc9e0646309758d3a19b157d654f71c77be6ad30f3c2.jpg) +(h) Local (32.69) + +![](images/0963bd913cce74e77db27b13d8d4b241723865689909287a13543838fa46b515.jpg) +(i) MAML (33.17) +Figure 5. Qualitative results on the Cars [6] and FaceScape [65, 70] datasets. Test-time optimization was performed for 8192 steps on the Cars dataset and 32768 steps on the FaceScape dataset. + +![](images/47e2f4e4b7f2471ab5e446ae8a809dffb36500ed4334a9e44b020255935d882e.jpg) +(j) FOMAML (33.26) + +![](images/798fbfe9e346d394933131b51eccc9135a07c3cd96bd75e56fec1d1c4deb6f94.jpg) +(k) Reptile (32.28) + +![](images/ccae8ecfe990562d533365d85b09fbccb7e82894916324db588189e568735eaf.jpg) +(1) FedMeNF (33.54) + +rithm, we use not only FedAvg [41] as the most standard method, but also state-of-the-art methods including FedProx [38], Scaffold [30], FedNova [61], FedExP [28], and FedACG [32]. For the meta-learning algorithm, we select widely used methods such as MAML [19], FOMAML [19], and Reptile [44], alongside HF-MAML [17, 18] and A-MAML [37]. We also include meta-NSGD [69], the meta-learning algorithm with task-level differential privacy. + +# 6.1. Fast Optimization + +Figure 3 illustrates the test-time optimization (TTO) performance of FedMeNF against baselines. Compared to Local, FML approaches reach higher PSNR more quickly. FedMeNF achieves $40.4\%$ and $74.64\%$ higher PSNR than Local with and without TTO, respectively. FedMeNF also achieves a higher PSNR in just 39 steps, making it about $210\mathrm{x}$ faster than Local, which requires 8192 steps. + +# 6.2. Privacy Preservation + +Figure 4 illustrates the privacy preservation performance across varying values of the regularization coefficient $\gamma$ . Fig. 4a presents the privacy metric $\mathrm{PSNR}_p$ over communication rounds. FedMeNF achieves an $18.25\%$ , $35.17\%$ , $38.42\%$ , and $65.28\%$ improvement in the privacy metric $\mathrm{PSNR}_p$ at $\gamma = 0.25, 0.5$ , and $0.75$ , compared to the baseline. Fig. 4b shows the trade-off between privacy preservation $(\mathrm{PSNR}_p\downarrow)$ and reconstruction performance $(\mathrm{PSNR}\uparrow)$ . Meta-NSGD, which utilizes the differential privacy mechanism, achieves lower $\mathrm{PSNR}_p$ values, indicating stronger privacy protection. However, it suffers from significant degradation in reconstruction performance. In contrast, existing meta-learning methods like MAML achieve high PSNR at the cost of greater vulnerability to privacy leakage with their higher $\mathrm{PSNR}_p$ . FedMeNF establishes an efficient frontier that balances privacy protection and reconstruction performance, effectively reducing $\mathrm{PSNR}_p$ with minimal impact on PSNR. Qualitative results are presented in Figs. 5 and 6. + +As shown in Tab. 1, FedMeNF achieves comparable PSNR while maintaining an acceptable $\mathrm{PSNR}_p$ across diverse modalities, as well as across various FL algorithms. FedMeNF also performs well on other metrics, such as SSIM and LPIPS, as detailed in Appendix F. + +# 6.3. Justification on the Privacy Metrics + +To investigate the practical relevance of the privacy metrics in capturing privacy leakage, we simulate two standard privacy attacks on the Cars dataset [6], assuming an honest-but-curious server. The target models for these attacks are defined as the final local meta-learners from each client in the last communication round. The first is the Membership Inference Attack [50], which determines if a given data sample was in the training set of a black-box model. We assess how accurately the server can predict the membership + +![](images/346fd516277afdbb2e487ac80d55b506d3ffebb64720e0349b7fd329891b7c90.jpg) + +![](images/a562eafe4f9ad02621d3366237a54a72ea5ee7f9c60c0587c8351c8bad8a0a81.jpg) + +![](images/c6d47c7949e73ff24a57289db91d133396ef1109d2f62e1270b717eeb398a1eb.jpg) + +![](images/f074aec58906f74817bb3411f355b8bfe534e707e725da5da6e0960ae544de33.jpg) +(a) GT (PSNR ${}_{p}$ ) +(b) MAML (17.33) +(c) FedMeNF (5.84) +(d) MAML (PSNR $p = 23.43$ + +![](images/bae1a89f076ecaf7faedd61420b43916b067c49570eae3504f56e74ff01bd0e8.jpg) +(e) FedMeNF (PSNR $p = 8.38$ + +![](images/81c9455ca49a80572e8c4c5a0fc9c9ba787f2c398ef117ba0020ac7ddd30aa76.jpg) +(f) GT (PSNR) + +![](images/102e6271192cebeb87c738a9b452c3138b8a026afd0c13d67837cb5892a47b15.jpg) +(g) MAML (35.74) + +![](images/c0d3d363bc2f5f15bb3e6e9f67a14e0c79de98a5cc9ccd28b017dcdbb35d763d.jpg) +(h) FedMeNF (35.16) + +![](images/8054425fce175f7d6e170e2b40f8e2524424d0cd7ea1d44fffe1bf7dece6470a.jpg) +(i) MAML (PSNR = 29.65) + +![](images/305aec68b39ebd23d161afae99c35d4166218c9e7228649800ad2347997a4bf4.jpg) +(j) FedMeNF (PSNR = 29.15) +Figure 6. Upper. Reconstruction results of the client's private image/video on the server: (b, d) using MAML and (c, e) using FedMeNF. Lower. Reconstruction results of a new private image/video on the client: (g, i) using MAML and (h, j) using FedMeNF. Our FedMeNF effectively enhances privacy as in (c, e) with minimal impact on the reconstruction quality as in (h, j). + +of the client's local data using the client's local meta-learner as the black-box model. The second is the Property Inference Attack [20], which infers a property of a given model's training set. We evaluate the server's accuracy in classifying the vehicle category of the client's car using the client's local meta-learner. The setup is detailed in Appendix C. Table 2 shows that worse privacy metrics correlate with greater vulnerability to privacy attacks. + +Moreover, we examine the correlation between the privacy metrics and $\epsilon$ of the differential privacy framework using meta-NSGD [69]. Figure 7 shows that the privacy metrics degrade as $\epsilon$ increases, supporting their generalizability as a measure of privacy leakage. + +# 6.4. Generalization on Non-IID Settings + +To evaluate the generalization capability of FedMeNF on non-IID settings, we vary the data quantity per client using a Dirichlet distribution with parameter $\alpha$ ; for the NeRF scenes (where each client has one task), we vary the num + +
Metric MethodMIA Acc↓PIA Acc↓Privacy MetricsNVS PSNR↑
PSNRp↓SSIMp↓LPIPSp↑
MAML95.8348.7819.730.8270.34323.08
FOMAML97.2248.7819.730.8270.34323.66
Reptile95.8343.9019.960.8390.31121.98
HF-MAML88.8943.9019.240.8270.34123.58
A-MAML95.8343.9019.640.8260.34323.13
Ours55.5629.2712.150.7980.53724.05
Ours (adaptive)55.5629.2712.090.7970.53823.19
+ +![](images/4c390122f4c083c3c3c73865f17bf8bac6613ff22516fda7c9c7ff9e9d37a85f.jpg) +Figure 7. Correlation between $\epsilon$ and privacy metrics. + +ber of input views, while for the image and video modalities (where each task is a single image or video), we vary the number of tasks. A larger $\alpha$ indicates that the data quantity per client is more uniform. See Appendix E.3 for details. + +Table 1 presents the results for $\alpha = 5.0$ , while Tab. 3 shows the results for $\alpha = 10,5.0,1.0$ . For the baselines, a small $\alpha$ creates two problems. First, local meta-optimization on a few views tends to result in over-fitted local models, while training on many views leads to well-trained ones. In both cases, the server can reconstruct clients' private data with high fidelity, leading to privacy leakage. Second, over-fitted models perform poorly on novel view synthesis, degrading the global meta-learner. In contrast, FedMeNF is robust to this data imbalance because it only learns how to optimize rather than simply memorizing data. This approach prevents over-fitting and thus achieves robust performance in both privacy preservation and novel view synthesis, regardless of the data quantity. + +# 6.5. Few-shot Novel View Synthesis + +We conduct experiments with varying average number of input views using FedAvg [41] on the Cars [6] dataset, and the results are shown in Tab. 4. For all methods, the view synthesis performance (PSNR) improves as the average number of input views increases. FML approaches outperform the Local across all cases. Notably, FedMeNF achieves the highest PSNR with acceptable $\mathrm{PSNR}_p$ , regardless of the average number of input views. This confirms that FedMeNF performs well even in few-shot settings. + +Table 2. Comparison of Membership Inference Attack (MIA) accuracy, Property Inference Attack (PIA) accuracy, Privacy Metrics, and Novel View Synthesis (NVS) quality. + +
Dirichlet α +Methodα = 10α = 5.0α = 1.0
PSNRpPSNRPSNRpPSNRPSNRpPSNR
MAML19.6223.3319.7323.082021.24
FOMAML19.6223.119.7323.662021.25
Reptile19.9622.4719.9621.9819.9521.29
HF-MAML19.1823.4919.2423.5818.8422.2
A-MAML19.5723.2919.6423.1319.7521.44
Ours12.1323.4712.1524.0512.7324.32
Ours (adaptive)12.123.5212.0923.1913.7320.06
+ +Table 3. Results of the novel view synthesis performance and privacy metric with different Dirichlet parameters $\alpha = \left\lbrack {{10},{5.0},{1.0}}\right\rbrack$ . + +
Avg. # of views Method2-view4-view8-view
\(PSNR_p\)\(PSNR\)\(PSNR_p\)\(PSNR\)\(PSNR_p\)\(PSNR\)
MAML19.7320.2319.7323.0819.5426.03
FOMAML19.7420.3919.7323.6619.5425.97
Reptile18.5920.819.9621.9820.3426.3
HF-MAML18.8321.1519.2423.5818.4125.62
A-MAML19.7320.2519.6423.1319.5725.22
Ours15.5621.2712.1524.0511.826.31
Ours (adaptive)15.8420.2112.0923.1912.0325.82
+ +Table 4. Results of the novel view synthesis performance and privacy metric for 2, 4, and 8 input views per task. + +# 7. Discussion + +Since FedMeNF is designed for distributed devices, we use a relatively simple NF model. Although recent work [9, 31, 34] has achieved generalizable NFs using large models like transformers, these methods are often impractical for FL with edge devices due to their high communication and computational demands. Future work should focus on scalable, generalizable NFs aligned with FL constraints. + +# 8. Conclusion + +We proposed FedMeNF, the first federated meta-learning framework explicitly designed for neural fields of private data. FedMeNF regulates the increase of our intuitive and widely applicable privacy metric to train the global neural fields meta-learner while preserving privacy. Through experiments across diverse data modalities, including images, videos, and NeRFs, we demonstrate that FedMeNF effectively preserves client privacy while achieving rapid optimization speed and high reconstruction performance, even when local data are highly limited or non-IID. A potential limitation of this study is that we used relatively simple neural field models and vision data only. Future work will investigate generalizable, privacy-preserving neural fields for a broader range of personal data. Measuring and preserving privacy in neural fields is a crucial and novel challenge which has received less attention. We believe this study will serve as an essential starting point for further research. + +# Acknowledgements + +This work was partially supported by the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. RS-2022-II220156, Fundamental research on continual meta-learning for quality enhancement of casual videos and their 3D metaverse transformation; No. RS-2021-II211343, Artificial Intelligence Graduate School Program (Seoul National University)), National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. 2023R1A2C2005573), and Center for Applied Research in Artificial Intelligence (CARAI) grant funded by Defense Acquisition Program Administration (DAPA) and Agency for Defense Development (ADD) (UD230017TD). Gunhee Kim is the corresponding author. + +# References + +[1] Martin Abadi, Andy Chu, Ian Goodfellow, H Brendan McMahan, Ilya Mironov, Kunal Talwar, and Li Zhang. Deep learning with differential privacy. In ACM CCS, 2016. 3 +[2] Sawsan AbdulRahman, Hanine Tout, Hakima Ould-Slimane, Azzam Mourad, Chamseddine Talhi, and Mohsen Guizani. A survey on federated learning: The journey from centralized to distributed on-site learning and beyond. IEEE Internet of Things Journal, 2020. 3 +[3] Rohith Agaram, Shaurya Dewan, Rahul Sajnani, Adrien Poulenard, Madhava Krishna, and Srinath Sridhar. Canonical fields: Self-supervised learning of pose-canonicalized neural fields. In CVPR, 2023. 2 +[4] Mohammed Aledhari, Rehma Razzak, Reza M Parizi, and Fahad Saeed. Federated learning: A survey on enabling technologies, protocols, and applications. IEEE Access, 2020. 1 +[5] Mark Bun and Thomas Steinke. Concentrated differential privacy: Simplifications, extensions, and lower bounds. In Theory of Cryptography, 2016. 3 +[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv:1512.03012, 2015. 5, 6, 7, 8, 3, 4, 10, 13 +[7] Daoyuan Chen, Dawei Gao, Weirui Kuang, Yaliang Li, and Bolin Ding. pfl-bench: A comprehensive benchmark for personalized federated learning. NeurIPS, 2022. 3, 4, 6 +[8] Fei Chen, Mi Luo, Zhenhua Dong, Zhenguo Li, and Xiuqiang He. Federated meta-learning with fast convergence and efficient communication. arXiv:1802.07876, 2018. 1, 3, 5 +[9] Yinbo Chen and Xiaolong Wang. Transformers as metalearners for implicit neural representations. In ECCV, 2022. 2, 4, 5, 8, 1, 3 +[10] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In CVPR, 2021. 1 +[11] Zeyuan Chen, Yinbo Chen, Jingwen Liu, Xingqian Xu, Vedit Goel, Zhangyang Wang, Humphrey Shi, and Xiaolong + +Wang. Videoirr: Learning video implicit neural representation for continuous space-time super-resolution. In CVPR, 2022. 1 +[12] Stephen Coombes, Peter beim Graben, Roland Potthast, and James Wright. Neural fields: theory and applications. Springer, 2014. 1 +[13] Moming Duan, Qinbin Li, Linshan Jiang, and Bingsheng He. Towards open federated learning platforms: Survey and vision from technical and legal perspectives, 2024. 1 +[14] Emilien Dupont, Hyunjik Kim, SM Eslami, Danilo Rezende, and Dan Rosenbaum. From data to functa: Your data point is a function and you can treat it like one. arXiv:2201.12204, 2022. 1, 2, 3 +[15] Cynthia Dwork and Guy N Rothblum. Concentrated differential privacy. arXiv:1603.01887, 2016. 3 +[16] Cynthia Dwork, Krishnamaram Kenthapadi, Frank McSherry, Ilya Mironov, and Moni Naor. Our data, ourselves: Privacy via distributed noise generation. In EUROCRYPT, 2006. 3 +[17] Alireza Fallah, Aryan Mokhtari, and Asuman Ozdaglar. On the convergence theory of gradient-based model-agnostic meta-learning algorithms. In AISTATS, 2020. 7, 1 +[18] Alireza Fallah, Aryan Mokhtari, and Asuman Ozdaglar. Personalized federated learning with theoretical guarantees: A model-agnostic meta-learning approach. NeurIPS, 2020. 1, 3, 5, 7 +[19] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In ICML, 2017. 1, 7, 6, 10 +[20] Karan Ganju, Qi Wang, Wei Yang, Carl A Gunter, and Nikita Borisov. Property inference attacks on fully connected neural networks using permutation invariant representations. In ACM CCS, pages 619-633, 2018. 7, 3 +[21] Wei Gao, Shangwei Guo, Tianwei Zhang, Han Qiu, Yonggang Wen, and Yang Liu. Privacy-preserving collaborative learning with automatic transformation search. In CVPR, 2021. 4 +[22] Jeffrey Gu, Kuan-Chieh Wang, and Serena Yeung. Generalizable neural fields as partially observed neural processes. In ICCV, 2023. 1 +[23] Jamie Hayes, Borja Balle, and Saeed Mahloujifar. Bounding training data reconstruction in dp-sgd. In NeurIPS, 2023. 3 +[24] Lachlan Holden, Feras Dayoub, David Harvey, and Tat-Jun Chin. Federated neural radiance fields. arXiv:2305.01163, 2023. 1, 2 +[25] Minui Hong, Junhyeog Yun, Insu Jeon, and Gunhee Kim. Fedavp: Augment local data via shared policy in federated learning. NeurIPS, 2024. 4 +[26] Yu Huang, Yingbin Liang, and Longbo Huang. Provable generalization of overparameterized meta-learning trained with sgd. NeurIPS, 2022. 1 +[27] Ahmed Imteaj, Urmish Thakker, Shiqiang Wang, Jian Li, and M Hadi Amini. A survey on federated learning for resource-constrained IoT devices. IEEE Internet of Things Journal, 2021. 1 +[28] Divyansh J Hunjunwala, Shiqiang Wang, and Gauri Joshi. Fedexp: Speeding up federated averaging via extrapolation. arXiv:2301.09604, 2023. 7, 1, 11, 12, 13, 14 + +[29] Yihan Jiang, Jakub Konečný, Keith Rush, and Sreeram Kannan. Improving federated learning personalization via model agnostic meta learning. arXiv:1909.12488, 2019. 1, 3 +[30] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank Reddi, Sebastian Stich, and Ananda Theertha Suresh. Scaffold: Stochastic controlled averaging for federated learning. In ICML, 2020. 7, 1, 11, 12, 13, 14 +[31] Chiheon Kim, Doyup Lee, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Generalizable implicit neural representations via instance pattern composers. In CVPR, 2023. 4, 8, 1 +[32] Geeho Kim, Jinkyu Kim, and Bohyung Han. Communication-efficient federated learning with accelerated client gradient. In CVPR, 2024. 7, 1, 11, 12, 13, 14 +[33] Subin Kim, Sihyun Yu, Jaeho Lee, and Jinwoo Shin. Scalable neural video representations with learnable positional features. NeurIPS, 2022. 1 +[34] Doyup Lee, Chiheon Kim, Minsu Cho, and WOOK SHIN HAN. Locality-aware generalizable implicit neural representation. NeurIPS, 2024. 8 +[35] Jeffrey Li, Mikhail Khodak, Sebastian Caldas, and Ameet Talwalkar. Differentially private meta-learning. In ICLR, 2020. 3 +[36] Qinbin Li, Yiqun Diao, Quan Chen, and Bingsheng He. Federated learning on non-iid data silos: An experimental study. In IEEE International Conference on Data Engineering, 2022. 1 +[37] Shibo Li, Zheng Wang, Akil Narayan, Robert Kirby, and Shandian Zhe. Meta-learning with adjoint methods. In AISTATS, pages 7239-7251. PMLR, 2023. 7 +[38] Xiang Li, Kaixuan Huang, Wenhao Yang, Shusen Wang, and Zhihua Zhang. On the convergence of fedavg on non-iid data. arXiv:1907.02189, 2019. 7, 1, 11, 12, 13, 14 +[39] Zhenguo Li, Fengwei Zhou, Fei Chen, and Hang Li. Meta-sgd: Learning to learn quickly for few-shot learning. arXiv:1707.09835, 2017. 1 +[40] Susan Liang, Chao Huang, Yapeng Tian, Anurag Kumar, and Chenliang Xu. Av-nerf: Learning neural fields for real-world audio-visual scene synthesis. NeurIPS, 2023. 1 +[41] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In AISTATS, 2017. 1, 3, 5, 7, 8, 6, 10, 11, 12, 13, 14 +[42] William McNally, Kanav Vats, Tyler Pinto, Chris Dulhanty, John McPhee, and Alexander Wong. Golfdb: A video database for golf swing sequencing. In CVPRW, 2019. 5, 6, 3, 4, 7, 12 +[43] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 2021. 2, 5, 3, 4 +[44] Alex Nichol, Joshua Achiam, and John Schulman. On first-order meta-learning algorithms. arXiv:1803.02999, 2018. 1, 7 +[45] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learn + +ing implicit 3d representations without 3d supervision. In CVPR, 2020. 1 +[46] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, 2019. 1 +[47] Vishwanath Saragadam, Jasper Tan, Guha Balakrishnan, Richard G Baraniuk, and Ashok Veeraraghavan. Miner: Multiscale implicit neural representation. In ECCV, 2022. 1 +[48] Bipasha Sen, Aditya Agarwal, Vinay P Namboodiri, and CV Jawahar. Inr-v: A continuous representation space for videobased generative tasks. arXiv:2210.16579, 2022. 1 +[49] Risa Shinoda and Kaede Shiohara. Petface: A large-scale dataset and benchmark for animal identification. In ECCV, 2025. 5, 6, 3, 4, 7, 11 +[50] Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. Membership inference attacks against machine learning models. In IEEE Symposium on Security and Privacy, pages 3-18. IEEE, 2017. 7, 2 +[51] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. NeurIPS, 2020. 2, 5, 3 +[52] Ivan Skorokhodov, Savva Ignatyev, and Mohamed Elhoseiny. Adversarial generation of continuous images. In CVPR, 2021. 1 +[53] Kun Su, Mingfei Chen, and Eli Shlizerman. Inras: Implicit neural representation for audio scenes. NeurIPS, 2022. 1 +[54] Terpeei Suzuki. Federated learning for large-scale scene modeling with neural radiance fields. arXiv:2309.06030, 2023.2 +[55] Terpei Suzuki. Fed3dgs: Scalable 3d gaussian splatting with federated learning. arXiv:2403.11460, 2024. 1, 2 +[56] Jihoon Tack, Subin Kim, Sihyun Yu, Jaeho Lee, Jinwoo Shin, and Jonathan Richard Schwarz. Learning large-scale neural fields via context pruned meta-learning. NeurIPS, 2024. 2, 3 +[57] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. NeurIPS, 2020. 1, 2 +[58] Matthew Tancik, Ben Mildenhall, Terrance Wang, Divi Schmidt, Pratul P Srinivasan, Jonathan T Barron, and Ren Ng. Learned initializations for optimizing coordinate-based neural representations. In CVPR, 2021. 5, 1, 3 +[59] Zaid Tasneem, Akshit Dave, Abhishek Singh, Kushagra Tiwary, Praneeth Vepakomma, Ashok Veeraraghavan, and Ramesh Raskar. Decentnerfs: Decentralized neural radiance fields from crowdsourced images. arXiv:2403.13199, 2024. 1 +[60] Ayush Tewari, Ohad Fried, Justus Thies, Vincent Sitzmann, Stephen Lombardi, Kalyan Sunkavalli, Ricardo MartinBrualla, Tomas Simon, Jason Saragih, Matthias Nießner, et al. State of the art on neural rendering. In Computer Graphics Forum, 2020. 1 + +[61] Jianyu Wang, Qinghua Liu, Hao Liang, Gauri Joshi, and H Vincent Poor. Tackling the objective inconsistency problem in heterogeneous federated optimization. NeurIPS, 2020. 7, 1, 11, 12, 13, 14 +[62] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 2004. 4 +[63] Kang Wei, Jun Li, Ming Ding, Chuan Ma, Howard H. Yang, Farhad Farokhi, Shi Jin, Tony Q. S. Quek, and H. Vincent Poor. Federated learning with differential privacy: Algorithms and performance analysis. IEEE Transactions on Information Forensics and Security, 2020. 3 +[64] Yuexiang Xie, Zhen Wang, Dawei Gao, Daoyuan Chen, Liuyi Yao, Weirui Kuang, Yaliang Li, Bolin Ding, and Jingren Zhou. Federatedscope: A flexible federated learning platform for heterogeneity. arXiv:2204.05011, 2022. 4, 6 +[65] Haotian Yang, Hao Zhu, Yanru Wang, Mingkai Huang, Qiu Shen, Ruigang Yang, and Xun Cao. Facescape: a large-scale high quality 3d face dataset and detailed riggable 3d face prediction. In CVPR, 2020. 5, 6, 3, 7, 14 +[66] Kai Yue, Richeng Jin, Chau-Wai Wong, Dror Baron, and Huaiyu Dai. Gradient obfuscation gives a false sense of security in federated learning. In USENIX Security Symposium, 2023. 4 +[67] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 4 +[68] Yintian Zhang and Ziyu Shao. Federated neural radiance field for distributed intelligence. arXiv:2406.10474, 2024. 1, 2 +[69] Xinyu Zhou and Raef Bassily. Task-level differentially private meta learning. NeurIPS, 2022. 7, 1 +[70] Hao Zhu, Haotian Yang, Longwei Guo, Yidi Zhang, Yanru Wang, Mingkai Huang, Menghua Wu, Qiu Shen, Ruigang Yang, and Xun Cao. Facescape: 3d facial dataset and benchmark for single-view 3d face reconstruction. IEEE TPAMI, 2023. 5, 6, 3, 7, 14 +[71] Yiyu Zhuang, Hao Zhu, Xusen Sun, and Xun Cao. Mofanerf: Morphable facial neural radiance field. In ECCV, 2022. 1 \ No newline at end of file diff --git a/2025/FedMeNF_ Privacy-Preserving Federated Meta-Learning for Neural Fields/images.zip b/2025/FedMeNF_ Privacy-Preserving Federated Meta-Learning for Neural Fields/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4852ff82c818266b9e5fa350bb66bee6e27712bc --- /dev/null +++ b/2025/FedMeNF_ Privacy-Preserving Federated Meta-Learning for Neural Fields/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:baaf9d9aa538646c386d75605bffd68ed23b55f4503cb3b37bbc4a68c6b1de67 +size 791920 diff --git a/2025/FedMeNF_ Privacy-Preserving Federated Meta-Learning for Neural Fields/layout.json b/2025/FedMeNF_ Privacy-Preserving Federated Meta-Learning for Neural Fields/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..47f394d677b28aec1795b2b0931647a8f8ae4338 --- /dev/null +++ b/2025/FedMeNF_ Privacy-Preserving Federated Meta-Learning for Neural Fields/layout.json @@ -0,0 +1,14359 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 75, + 102, + 536, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 102, + 536, + 121 + ], + "spans": [ + { + "bbox": [ + 75, + 102, + 536, + 121 + ], + "type": "text", + "content": "FedMeNF: Privacy-Preserving Federated Meta-Learning for Neural Fields" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 194, + 143, + 416, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 143, + 416, + 172 + ], + "spans": [ + { + "bbox": [ + 194, + 143, + 416, + 172 + ], + "type": "text", + "content": "Junhyeog Yun Minui Hong Gunhee Kim Seoul National University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 200, + 173, + 411, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 173, + 411, + 186 + ], + "spans": [ + { + "bbox": [ + 200, + 173, + 411, + 186 + ], + "type": "text", + "content": "{junhyeog,alsdml123,gunhee}@snu.ac.kr" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 208, + 187, + 399, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 187, + 399, + 199 + ], + "spans": [ + { + "bbox": [ + 208, + 187, + 399, + 199 + ], + "type": "text", + "content": "https://github.com/junhyeog/FedMeNF" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 227, + 201, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 201, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 201, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 251, + 297, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 251, + 297, + 456 + ], + "spans": [ + { + "bbox": [ + 54, + 251, + 297, + 456 + ], + "type": "text", + "content": "Neural fields provide a memory-efficient representation of data, which can effectively handle diverse modalities and large-scale data. However, learning to map neural fields often requires large amounts of training data and computations, which can be limited to resource-constrained edge devices. One approach to tackle this limitation is to leverage Federated Meta-Learning (FML), but traditional FML approaches suffer from privacy leakage. To address these issues, we introduce a novel FML approach called FedMeNF. FedMeNF utilizes a new privacy-preserving loss function that regulates privacy leakage in the local meta-optimization. This enables the local meta-learner to optimize quickly and efficiently without retaining the client's private data. Our experiments demonstrate that FedMeNF achieves fast optimization speed and robust reconstruction performance, even with few-shot or non-IID data across diverse data modalities, while preserving client data privacy." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 478, + 135, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 478, + 135, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 478, + 135, + 491 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 498, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 582 + ], + "type": "text", + "content": "Neural Fields (NFs), also known as Implicit Neural Representations (INRs), use a deep neural network to approximate continuous signals [12, 46, 47, 57, 60]. They enable memory-efficient representation and fast processing of data [14]. Thanks to this, NFs effectively handle both large-scale data and diverse modalities, including audio, images, videos, and 3D objects [10, 11, 33, 40, 45, 48, 52, 53, 71]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": "However, optimizing NFs is often slow and requires large amounts of data, which can be a serious issue on resource-constrained edge devices. Consider a practical scenario: a user takes a few photos of an object using their phone and wants to get a high-fidelity 3D model of it. Training an NF from scratch with such limited data and resources would be time-consuming and likely yield poor results. One approach to address this issue is meta-learning [19, 44], which aims to train a meta-learner that can quickly adapt to new tasks with only a few samples. Using the meta-learner, the user can quickly optimize a high-performing NF for the" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 228, + 455, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 455, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 455, + 240 + ], + "type": "text", + "content": "new object with just a few images." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 242, + 555, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 242, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 242, + 555, + 397 + ], + "type": "text", + "content": "Nevertheless, meta-learning itself requires a large and diverse set of tasks to learn an effective initialization or optimization strategy. Unfortunately, in many practical scenarios, users only have data for one or a few tasks. For example, most users typically take photos of only a single instance of privacy-sensitive objects such as their face, body, pet, or car. To enable meta-learning in such scenarios, we can leverage Federated Meta-Learning (FML) [8, 18, 29], which collaboratively trains a global meta-learner using private data from numerous distributed clients. FML builds upon Federated Learning (FL) [4, 13, 27, 36, 41], a paradigm designed to train a global model while ensuring client data remains on local devices to preserve privacy." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 400, + 556, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 556, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 556, + 663 + ], + "type": "text", + "content": "Yet, applying FML to NFs introduces a critical challenge that fundamentally undermines the privacy guarantee of FL. The privacy protection in traditional FL relies on the assumption that sharing trained model parameters, rather than the raw data itself, is safe. This assumption completely breaks down for NFs for two main reasons. First, in our scenario, each client has data for only a single or a few task instances (e.g., a single image or 3D object). If even one client has only a single task instance, the local meta-optimization is like standard supervised training on that specific task [26]. Consequently, the trained local meta-learner functions as an optimized NF for that private data. Second, NFs are, in essence, a compressed representation of the data and inherently encapsulate it. Thus, even if only the NF parameters are shared, a malicious server can reconstruct the private data. Due to this inherent privacy leakage issue, existing studies often limit FL applications for NFs to public data such as landmarks [59] or cityscapes [24, 55, 68], where privacy is less of a concern. For these reasons, naively applying FML methods to private data results in severe privacy issues, as the server can reconstruct the private data from the shared local meta-learner." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "To address this critical challenge, we propose FedMeNF (Federated Meta-Learning for Neural Fields), a novel FML framework designed to collaboratively train a global NF meta-learner while preserving privacy as overviewed in" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2161" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 69, + 197, + 284 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 197, + 284 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 197, + 284 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 197, + 284 + ], + "type": "image", + "image_path": "0f30597defe4686a7a2bfcaad6c3c63efe9343e25e9f8260a02cd3d6c3d69952.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 71, + 285, + 192, + 295 + ], + "lines": [ + { + "bbox": [ + 71, + 285, + 192, + 295 + ], + "spans": [ + { + "bbox": [ + 71, + 285, + 192, + 295 + ], + "type": "text", + "content": "(a) FML with a single task per client." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 230, + 69, + 380, + 126 + ], + "blocks": [ + { + "bbox": [ + 230, + 69, + 380, + 126 + ], + "lines": [ + { + "bbox": [ + 230, + 69, + 380, + 126 + ], + "spans": [ + { + "bbox": [ + 230, + 69, + 380, + 126 + ], + "type": "image", + "image_path": "8888ecd7e33472cc1c28fe1cdf03739c13a8500a76df8f359b1d4f398a5d83f2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 129, + 398, + 158 + ], + "lines": [ + { + "bbox": [ + 211, + 129, + 398, + 158 + ], + "spans": [ + { + "bbox": [ + 211, + 129, + 398, + 158 + ], + "type": "text", + "content": "(b) Comparison of FedMeNF with Local and FML in terms of fast or few-shot optimization on a new task, and privacy preservation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 239, + 167, + 372, + 273 + ], + "blocks": [ + { + "bbox": [ + 239, + 167, + 372, + 273 + ], + "lines": [ + { + "bbox": [ + 239, + 167, + 372, + 273 + ], + "spans": [ + { + "bbox": [ + 239, + 167, + 372, + 273 + ], + "type": "image", + "image_path": "7be94016b2bbba070827b47748f3e92025c57419bf10938eaebd3854450c20bc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 275, + 398, + 295 + ], + "lines": [ + { + "bbox": [ + 211, + 275, + 398, + 295 + ], + "spans": [ + { + "bbox": [ + 211, + 275, + 398, + 295 + ], + "type": "text", + "content": "(c) The client has photos of only one pet, car, body, or face, and aims to optimize a neural field for each instance." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 403, + 70, + 542, + 283 + ], + "blocks": [ + { + "bbox": [ + 403, + 70, + 542, + 283 + ], + "lines": [ + { + "bbox": [ + 403, + 70, + 542, + 283 + ], + "spans": [ + { + "bbox": [ + 403, + 70, + 542, + 283 + ], + "type": "image", + "image_path": "c582a7d216a6f056675d0688f116fbff6e3783c3f9bca8835688f5337e990bb6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 89, + 298, + 521, + 326 + ], + "blocks": [ + { + "bbox": [ + 410, + 285, + 546, + 294 + ], + "lines": [ + { + "bbox": [ + 410, + 285, + 546, + 294 + ], + "spans": [ + { + "bbox": [ + 410, + 285, + 546, + 294 + ], + "type": "text", + "content": "(d) FedMeNF with a single task per client." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 89, + 298, + 521, + 326 + ], + "lines": [ + { + "bbox": [ + 89, + 298, + 521, + 326 + ], + "spans": [ + { + "bbox": [ + 89, + 298, + 521, + 326 + ], + "type": "image", + "image_path": "4f24c78639b3839c85ef30ca23fd2f9c6285513cbc897f1be62349485ec3d24f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 331, + 555, + 399 + ], + "lines": [ + { + "bbox": [ + 55, + 331, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 555, + 399 + ], + "type": "text", + "content": "Figure 1. Overview of our approach. In each communication round, the server sends the global meta-learner " + }, + { + "bbox": [ + 55, + 331, + 555, + 399 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 331, + 555, + 399 + ], + "type": "text", + "content": " to clients, each of which uses its private data to train a local meta-learner " + }, + { + "bbox": [ + 55, + 331, + 555, + 399 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 331, + 555, + 399 + ], + "type": "text", + "content": " based on the global meta-learner. The trained local meta-learners are then sent back to the server to update the global meta-learner. (a) In existing federated meta-learning (FML) approaches, the local meta-learner contains privacy-sensitive information when a client has only a single task, as shown in (c). The server can infer the client's private data from the shared local meta-learner. (d) In contrast, our FedMeNF ensures that the local meta-learner only learns how to efficiently optimize NFs while preserving privacy-sensitive details. (b, c) Local refers to using only the client's data without sharing any information with the server." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 415, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 415, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 54, + 415, + 295, + 569 + ], + "type": "text", + "content": "Fig. 1, which then empowers each client to rapidly optimize a high-performance NF even with a small amount of new private data. FedMeNF can be utilized in various privacy-sensitive applications. For example, suppose that a user buys a new car and takes a few pictures of it. We can then train the NF of the car with this small set of images using a global meta-learner trained with other car images in other devices. As another example, with only a handful of facial images, users can employ the global facial NF meta-learner to generate images showing a range of facial expressions. Also, they can upscale low-resolution photos to high-resolution or generate higher frame-rate videos from personal data, such as pet photos or golf swing videos." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "To quantify and control privacy leakage, we define a privacy metric, " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{PSNR}_p" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": ". We theoretically analyze how " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{PSNR}_p" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " increases under existing FML approaches. By regularizing the increase of " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{PSNR}_p" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": ", FedMeNF forces the local meta-learner to learn only how to optimize the NF faster with limited data without memorizing the private data. We demonstrate that FedMeNF achieves performance comparable to state-of-the-art methods while better preserving privacy across diverse data modalities, including images, videos, and neural radiance fields (NeRFs). Additionally, FedMeNF shows robust performance even when local data are very few or highly diverse among clients." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 326, + 415, + 509, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 415, + 509, + 426 + ], + "spans": [ + { + "bbox": [ + 326, + 415, + 509, + 426 + ], + "type": "text", + "content": "Our contributions can be outlined as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 429, + 554, + 537 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 314, + 429, + 537, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 429, + 537, + 441 + ], + "spans": [ + { + "bbox": [ + 314, + 429, + 537, + 441 + ], + "type": "text", + "content": "1. This is the first study on FL for NFs on private data." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 441, + 553, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 553, + 464 + ], + "type": "text", + "content": "2. We theoretically and empirically show how privacy leakage occurs during the FML for NFs." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 464, + 554, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 554, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 554, + 501 + ], + "type": "text", + "content": "3. We propose FedMeNF that preserves the privacy of local data with minimal impact on optimization speed and reconstruction quality." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 501, + 554, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 501, + 554, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 501, + 554, + 537 + ], + "type": "text", + "content": "4. We conduct comprehensive experiments on FedMeNF across various data modalities, private data sizes, and levels of data diversity, outperforming baseline methods." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 549, + 399, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 549, + 399, + 562 + ], + "spans": [ + { + "bbox": [ + 314, + 549, + 399, + 562 + ], + "type": "text", + "content": "2. Preliminaries" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 570, + 555, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 643 + ], + "type": "text", + "content": "Coordinate-based Neural Fields. Coordinate-based neural fields [3, 9, 14, 43, 51, 56, 57] represent continuous functions that map spatial coordinates to signal values such as color or density. The objective for each client " + }, + { + "bbox": [ + 313, + 570, + 555, + 643 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 570, + 555, + 643 + ], + "type": "text", + "content": " is to find the optimal NF parameter " + }, + { + "bbox": [ + 313, + 570, + 555, + 643 + ], + "type": "inline_equation", + "content": "\\varphi_{i}^{m}" + }, + { + "bbox": [ + 313, + 570, + 555, + 643 + ], + "type": "text", + "content": " that best represents its " + }, + { + "bbox": [ + 313, + 570, + 555, + 643 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 570, + 555, + 643 + ], + "type": "text", + "content": "-th local task data " + }, + { + "bbox": [ + 313, + 570, + 555, + 643 + ], + "type": "inline_equation", + "content": "T_{i}^{m}" + }, + { + "bbox": [ + 313, + 570, + 555, + 643 + ], + "type": "text", + "content": ". This is formulated as" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 347, + 651, + 555, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 651, + 555, + 681 + ], + "spans": [ + { + "bbox": [ + 347, + 651, + 555, + 681 + ], + "type": "interline_equation", + "content": "\\varphi_ {i} ^ {m} = \\arg \\min _ {\\varphi} \\sum_ {(\\mathbf {x} _ {j}, \\mathbf {y} _ {j}) \\in T _ {i} ^ {m}} L \\left(f _ {\\varphi} (\\mathbf {x} _ {j}), \\mathbf {y} _ {j}\\right), \\tag {1}", + "image_path": "660c67bfede8c1f52a4b4b8380162b0c03bf1b841d8031ae18baad10a3c3af31.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "inline_equation", + "content": "f_{\\varphi}" + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": " is the forward map parameterized by " + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_j" + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": " represents the input coordinates (e.g., spatial positions), and " + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_j" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2162" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "denotes the target values (e.g., pixel colors). " + }, + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": " is a suitable loss function such as mean squared error (MSE)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "type": "text", + "content": "Federated Learning. We consider a centralized federated learning (FL) environment [2, 41], where clients do not share their data with either the server or other clients and communicate the model parameters exclusively with the centralized server. The centralized FL aims to optimize the global model that minimizes a given objective function as" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 175, + 296, + 208 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 175, + 296, + 208 + ], + "spans": [ + { + "bbox": [ + 67, + 175, + 296, + 208 + ], + "type": "interline_equation", + "content": "\\arg \\min _ {\\theta} \\sum_ {m = 0} ^ {M - 1} \\alpha^ {m} L (\\theta , D ^ {m}), \\text {w h e r e} \\alpha^ {m} = \\frac {\\left| D ^ {m} \\right|}{\\Sigma_ {i} \\left| D ^ {i} \\right|}. \\tag {2}", + "image_path": "67a256345d4346e120fe81f60d5457ae26c756f826bb84e83edc7173e83be292.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 213, + 296, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 213, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 296, + 297 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 213, + 296, + 297 + ], + "type": "text", + "content": " is the parameter of the global meta-learner and " + }, + { + "bbox": [ + 55, + 213, + 296, + 297 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 213, + 296, + 297 + ], + "type": "text", + "content": " is the number of clients. " + }, + { + "bbox": [ + 55, + 213, + 296, + 297 + ], + "type": "inline_equation", + "content": "D^{m}" + }, + { + "bbox": [ + 55, + 213, + 296, + 297 + ], + "type": "text", + "content": " is the local dataset of client " + }, + { + "bbox": [ + 55, + 213, + 296, + 297 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 213, + 296, + 297 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 213, + 296, + 297 + ], + "type": "inline_equation", + "content": "\\alpha^{m}" + }, + { + "bbox": [ + 55, + 213, + 296, + 297 + ], + "type": "text", + "content": " is the proportion of the number of data samples of client " + }, + { + "bbox": [ + 55, + 213, + 296, + 297 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 213, + 296, + 297 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 213, + 296, + 297 + ], + "type": "inline_equation", + "content": "L(\\theta, D^{m})" + }, + { + "bbox": [ + 55, + 213, + 296, + 297 + ], + "type": "text", + "content": " represents the loss of model " + }, + { + "bbox": [ + 55, + 213, + 296, + 297 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 213, + 296, + 297 + ], + "type": "text", + "content": " on a dataset " + }, + { + "bbox": [ + 55, + 213, + 296, + 297 + ], + "type": "inline_equation", + "content": "D^{m}" + }, + { + "bbox": [ + 55, + 213, + 296, + 297 + ], + "type": "text", + "content": ". This objective ensures that the global model " + }, + { + "bbox": [ + 55, + 213, + 296, + 297 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 213, + 296, + 297 + ], + "type": "text", + "content": " is optimized across all clients by weighting the contribution of each client based on its local dataset size." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 297, + 296, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 296, + 334 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 296, + 334 + ], + "type": "text", + "content": "Federated Meta-learning. Federated meta-learning [7, 8, 18, 29] is a variant of FL where the client trains the local model using a meta-learning approach with the objective:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 110, + 350, + 295, + 383 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 350, + 295, + 383 + ], + "spans": [ + { + "bbox": [ + 110, + 350, + 295, + 383 + ], + "type": "interline_equation", + "content": "\\arg \\min _ {\\theta} \\sum_ {m = 0} ^ {M - 1} \\alpha^ {m} L \\left(\\varphi^ {m}, Q ^ {m}\\right), \\tag {3}", + "image_path": "f7962ea60fb5e46c11590fe6f7e986197f5b543752d17f3c2d3861fe34c5f3cc.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 110, + 384, + 294, + 402 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 384, + 294, + 402 + ], + "spans": [ + { + "bbox": [ + 110, + 384, + 294, + 402 + ], + "type": "interline_equation", + "content": "\\text {w h e r e} \\varphi^ {m} = \\arg \\min _ {\\theta} L (\\theta , S ^ {m}). \\tag {4}", + "image_path": "e0b580371746de723a5fa9bfbd50e1d8dab5388fd7b069e911cb3686e2f1f6e9.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 408, + 296, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 408, + 296, + 516 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 296, + 516 + ], + "type": "inline_equation", + "content": "S^m" + }, + { + "bbox": [ + 55, + 408, + 296, + 516 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 408, + 296, + 516 + ], + "type": "inline_equation", + "content": "Q^m" + }, + { + "bbox": [ + 55, + 408, + 296, + 516 + ], + "type": "text", + "content": " are the support set and query set of the client " + }, + { + "bbox": [ + 55, + 408, + 296, + 516 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 408, + 296, + 516 + ], + "type": "text", + "content": ". Each client " + }, + { + "bbox": [ + 55, + 408, + 296, + 516 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 408, + 296, + 516 + ], + "type": "text", + "content": " first trains the local model " + }, + { + "bbox": [ + 55, + 408, + 296, + 516 + ], + "type": "inline_equation", + "content": "\\varphi^m" + }, + { + "bbox": [ + 55, + 408, + 296, + 516 + ], + "type": "text", + "content": " on the support set " + }, + { + "bbox": [ + 55, + 408, + 296, + 516 + ], + "type": "inline_equation", + "content": "S^m" + }, + { + "bbox": [ + 55, + 408, + 296, + 516 + ], + "type": "text", + "content": " (inner loop). Then, starting from the global model " + }, + { + "bbox": [ + 55, + 408, + 296, + 516 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 408, + 296, + 516 + ], + "type": "text", + "content": ", the local meta-learner minimizes the loss of " + }, + { + "bbox": [ + 55, + 408, + 296, + 516 + ], + "type": "inline_equation", + "content": "\\varphi^m" + }, + { + "bbox": [ + 55, + 408, + 296, + 516 + ], + "type": "text", + "content": " on the query set " + }, + { + "bbox": [ + 55, + 408, + 296, + 516 + ], + "type": "inline_equation", + "content": "Q^m" + }, + { + "bbox": [ + 55, + 408, + 296, + 516 + ], + "type": "text", + "content": " (outer loop). This local meta-optimization enables the local meta-learner to adapt quickly to new tasks. The trained local meta-learner is shared with the server, which aggregates the local meta-learners received from participating clients to update the global meta-learner." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 516, + 296, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 296, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 296, + 659 + ], + "type": "text", + "content": "Differential Privacy. Differential privacy (DP) is one of the most popular privacy frameworks with rich theoretical guarantees [1, 5, 15, 23, 35, 63]. It guarantees that when a single entry in the database changes, the change in the probability distribution of the algorithm outputs is bounded by predetermined factors " + }, + { + "bbox": [ + 55, + 516, + 296, + 659 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 516, + 296, + 659 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 516, + 296, + 659 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 516, + 296, + 659 + ], + "type": "text", + "content": " [16]. The privacy risk can be regulated by controlling these factors. However, traditional DP methods are designed for large datasets; yet neural fields often capture just a single signal. This causes ambiguity in defining privacy leakage for individual training samples, as well as a significant trade-off between privacy preservation and reconstruction performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 669, + 170, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 669, + 170, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 669, + 170, + 681 + ], + "type": "text", + "content": "3. Problem Statement" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "In this work, we assume a scenario where multiple users collaboratively train a global neural field meta-learner using" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 555, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 251 + ], + "type": "text", + "content": "federated meta-learning. Each user has data about a personal object stored locally on their device. For instance, each user holds a few images of his or her own car on their device. Our goal is to optimize neural fields despite limited data and low computational resources on local devices. Thus, optimizing a specific neural field becomes a local task. The local task data are divided into a support set and a query set, each containing images of the object. During training, the support set is used to train the neural field, while the query set is utilized to optimize the local meta-learner. In the testing phase, the support set enables test-time optimization, and the query set is used to assess its performance. The detailed FML process is described in Fig. 1a and Algorithm 1, excluding the highlighted parts which will be discussed in Sec. 5." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 258, + 550, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 550, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 550, + 270 + ], + "type": "text", + "content": "3.1. First-order Approximation of Meta-gradients" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 275, + 555, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 555, + 348 + ], + "type": "text", + "content": "During the local meta-optimization, the local meta-learner " + }, + { + "bbox": [ + 313, + 275, + 555, + 348 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 275, + 555, + 348 + ], + "type": "text", + "content": " is updated for each outer step " + }, + { + "bbox": [ + 313, + 275, + 555, + 348 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 275, + 555, + 348 + ], + "type": "text", + "content": ", where task data " + }, + { + "bbox": [ + 313, + 275, + 555, + 348 + ], + "type": "inline_equation", + "content": "T_{i} = \\{S_{i}, Q_{i}\\}" + }, + { + "bbox": [ + 313, + 275, + 555, + 348 + ], + "type": "text", + "content": " are sampled from the client's training dataset " + }, + { + "bbox": [ + 313, + 275, + 555, + 348 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{train}}" + }, + { + "bbox": [ + 313, + 275, + 555, + 348 + ], + "type": "text", + "content": ". Then, the local NF " + }, + { + "bbox": [ + 313, + 275, + 555, + 348 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 313, + 275, + 555, + 348 + ], + "type": "text", + "content": ", initialized by " + }, + { + "bbox": [ + 313, + 275, + 555, + 348 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 313, + 275, + 555, + 348 + ], + "type": "text", + "content": ", are optimized with the support set " + }, + { + "bbox": [ + 313, + 275, + 555, + 348 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 313, + 275, + 555, + 348 + ], + "type": "text", + "content": " (Eq. (8)), and the local meta-learner " + }, + { + "bbox": [ + 313, + 275, + 555, + 348 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 313, + 275, + 555, + 348 + ], + "type": "text", + "content": " are optimized with the query set " + }, + { + "bbox": [ + 313, + 275, + 555, + 348 + ], + "type": "inline_equation", + "content": "Q_{i}" + }, + { + "bbox": [ + 313, + 275, + 555, + 348 + ], + "type": "text", + "content": " (Eq. (9))." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 352, + 555, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 352, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 352, + 555, + 389 + ], + "type": "text", + "content": "Proposition 1. Let " + }, + { + "bbox": [ + 313, + 352, + 555, + 389 + ], + "type": "inline_equation", + "content": "g_{M} = \\nabla_{w_{i}}L(\\varphi_{K},B_{K})" + }, + { + "bbox": [ + 313, + 352, + 555, + 389 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 352, + 555, + 389 + ], + "type": "inline_equation", + "content": "g_{k} = \\nabla_{\\varphi_{0}}L(\\varphi_{0},B_{k})" + }, + { + "bbox": [ + 313, + 352, + 555, + 389 + ], + "type": "text", + "content": ". Then, the first-order approximation of the meta-gradient " + }, + { + "bbox": [ + 313, + 352, + 555, + 389 + ], + "type": "inline_equation", + "content": "g_{M}" + }, + { + "bbox": [ + 313, + 352, + 555, + 389 + ], + "type": "text", + "content": " becomes" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 319, + 394, + 554, + 427 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 394, + 554, + 427 + ], + "spans": [ + { + "bbox": [ + 319, + 394, + 554, + 427 + ], + "type": "interline_equation", + "content": "g _ {M} \\approx g _ {K} - \\lambda_ {i} \\mathcal {I} _ {K}, w h e r e \\mathcal {I} _ {K} = \\sum_ {k = 0} ^ {K - 1} \\nabla_ {\\varphi_ {0}} \\left\\langle g _ {K}, g _ {k} \\right\\rangle . \\tag {5}", + "image_path": "180ea63d3f47f4f1fe4af91340d7a962d0fd553633427b26ea92090a301e91a6.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 431, + 554, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 554, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 554, + 468 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 431, + 554, + 468 + ], + "type": "text", + "content": " is the # of inner steps, " + }, + { + "bbox": [ + 313, + 431, + 554, + 468 + ], + "type": "inline_equation", + "content": "B_{0\\dots K - 1}" + }, + { + "bbox": [ + 313, + 431, + 554, + 468 + ], + "type": "text", + "content": " are minibatches of " + }, + { + "bbox": [ + 313, + 431, + 554, + 468 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 313, + 431, + 554, + 468 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 431, + 554, + 468 + ], + "type": "inline_equation", + "content": "B_{K}" + }, + { + "bbox": [ + 313, + 431, + 554, + 468 + ], + "type": "text", + "content": " is a minibatch of " + }, + { + "bbox": [ + 313, + 431, + 554, + 468 + ], + "type": "inline_equation", + "content": "Q_{i}" + }, + { + "bbox": [ + 313, + 431, + 554, + 468 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 431, + 554, + 468 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + }, + { + "bbox": [ + 313, + 431, + 554, + 468 + ], + "type": "text", + "content": " is an inner loop learning rate. The proof is provided in Appendix H." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 474, + 454, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 454, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 454, + 487 + ], + "type": "text", + "content": "3.2. Fast Optimization via " + }, + { + "bbox": [ + 313, + 474, + 454, + 487 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_K" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 491, + 555, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 555, + 599 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 555, + 599 + ], + "type": "text", + "content": "According to Proposition 1, the meta-gradient " + }, + { + "bbox": [ + 313, + 491, + 555, + 599 + ], + "type": "inline_equation", + "content": "g_{M}" + }, + { + "bbox": [ + 313, + 491, + 555, + 599 + ], + "type": "text", + "content": " can be expressed with " + }, + { + "bbox": [ + 313, + 491, + 555, + 599 + ], + "type": "inline_equation", + "content": "g_{K}" + }, + { + "bbox": [ + 313, + 491, + 555, + 599 + ], + "type": "text", + "content": " and the inner product term " + }, + { + "bbox": [ + 313, + 491, + 555, + 599 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_K" + }, + { + "bbox": [ + 313, + 491, + 555, + 599 + ], + "type": "text", + "content": ", which is the sum of inner products between " + }, + { + "bbox": [ + 313, + 491, + 555, + 599 + ], + "type": "inline_equation", + "content": "g_{K}" + }, + { + "bbox": [ + 313, + 491, + 555, + 599 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 491, + 555, + 599 + ], + "type": "inline_equation", + "content": "g_{0},\\dots ,g_{K - 1}" + }, + { + "bbox": [ + 313, + 491, + 555, + 599 + ], + "type": "text", + "content": ". This term maximizes (i) the alignment of the gradients from the query set " + }, + { + "bbox": [ + 313, + 491, + 555, + 599 + ], + "type": "inline_equation", + "content": "(g_{K})" + }, + { + "bbox": [ + 313, + 491, + 555, + 599 + ], + "type": "text", + "content": " with those from the support set " + }, + { + "bbox": [ + 313, + 491, + 555, + 599 + ], + "type": "inline_equation", + "content": "(g_0,\\dots ,g_{K - 1})" + }, + { + "bbox": [ + 313, + 491, + 555, + 599 + ], + "type": "text", + "content": ", and (ii) the alignment of the gradients computed on different minibatches within the support set. This enables the local meta-learner (i) to be more generalized and (ii) to quickly optimize the neural field for a new task." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 608, + 541, + 622 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 608, + 541, + 622 + ], + "spans": [ + { + "bbox": [ + 313, + 608, + 541, + 622 + ], + "type": "text", + "content": "4. Privacy Leakage in Local Meta-Learning" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 628, + 554, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 554, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 554, + 652 + ], + "type": "text", + "content": "We define " + }, + { + "bbox": [ + 313, + 628, + 554, + 652 + ], + "type": "inline_equation", + "content": "\\mathrm{PSNR}_p" + }, + { + "bbox": [ + 313, + 628, + 554, + 652 + ], + "type": "text", + "content": ", an intuitive privacy metric to quantify the privacy leakage in FML for neural fields:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 349, + 658, + 555, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 658, + 555, + 673 + ], + "spans": [ + { + "bbox": [ + 349, + 658, + 555, + 673 + ], + "type": "interline_equation", + "content": "\\mathrm {P S N R} _ {p} = \\mathrm {P S N R} \\left(Q ^ {m}, f _ {w} \\left(\\operatorname {C o o r d} \\left(Q ^ {m}\\right)\\right)\\right), \\tag {6}", + "image_path": "37add89e8ae3e592776a3359fa9d0e59983fa191501c39e42adf1d1e088816e3.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "inline_equation", + "content": "Q^m" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": " is the query set of client " + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\operatorname{Coord}(Q^m)" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": " is the coordinates of " + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "inline_equation", + "content": "Q^m" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "inline_equation", + "content": "f_w" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": " is the forward map parameterized by the local meta-learner " + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{PSNR}_p" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": " assesses how well" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2163" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "the server can reconstruct " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "Q^{m}" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " with the shared local metalearner " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": ". It is a simple variant of PSNR that measures similarity between images reconstructed with " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " on the server and the client's local images. A better reconstruction indicates a greater violation of the client's data privacy, meaning that a larger " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "\\mathrm{PSNR}_p" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " indicates more privacy leakage." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 144, + 296, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 296, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 296, + 264 + ], + "type": "text", + "content": "We define our main privacy metric based on PSNR for two reasons. First, neural fields can handle various types of signals such as vision, speech, or sensor data, so does the PSNR. Second, PSNR is one of the most popular metrics for evaluating reconstruction attacks in FL [21, 25, 66], as well as reconstruction tasks [9, 31]. As variants, we also define " + }, + { + "bbox": [ + 55, + 144, + 296, + 264 + ], + "type": "inline_equation", + "content": "\\mathrm{SSIM}_p" + }, + { + "bbox": [ + 55, + 144, + 296, + 264 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 144, + 296, + 264 + ], + "type": "inline_equation", + "content": "\\mathrm{LPIPS}_p" + }, + { + "bbox": [ + 55, + 144, + 296, + 264 + ], + "type": "text", + "content": " as privacy metrics to assess structural and perceptual similarities [62, 67]; it can overcome the limitation of " + }, + { + "bbox": [ + 55, + 144, + 296, + 264 + ], + "type": "inline_equation", + "content": "\\mathrm{PSNR}_p" + }, + { + "bbox": [ + 55, + 144, + 296, + 264 + ], + "type": "text", + "content": " that is based on only pixel-wise MSE. Please refer to Appendix F for more details." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 273, + 187, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 273, + 187, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 273, + 187, + 285 + ], + "type": "text", + "content": "4.1. Privacy Leakage via " + }, + { + "bbox": [ + 55, + 273, + 187, + 285 + ], + "type": "inline_equation", + "content": "g_{K}" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 289, + 296, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 296, + 350 + ], + "type": "text", + "content": "Since " + }, + { + "bbox": [ + 55, + 289, + 296, + 350 + ], + "type": "inline_equation", + "content": "\\mathrm{PSNR}_p" + }, + { + "bbox": [ + 55, + 289, + 296, + 350 + ], + "type": "text", + "content": " is defined by " + }, + { + "bbox": [ + 55, + 289, + 296, + 350 + ], + "type": "inline_equation", + "content": "10\\log_{10}(R / L(w,Q^m))" + }, + { + "bbox": [ + 55, + 289, + 296, + 350 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 289, + 296, + 350 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 289, + 296, + 350 + ], + "type": "text", + "content": " is a constant, it becomes larger as the MSE loss " + }, + { + "bbox": [ + 55, + 289, + 296, + 350 + ], + "type": "inline_equation", + "content": "L(w,Q^{m})" + }, + { + "bbox": [ + 55, + 289, + 296, + 350 + ], + "type": "text", + "content": " is smaller. We analyze how " + }, + { + "bbox": [ + 55, + 289, + 296, + 350 + ], + "type": "inline_equation", + "content": "L(w,Q^{m})" + }, + { + "bbox": [ + 55, + 289, + 296, + 350 + ], + "type": "text", + "content": " changes during the client's local meta-optimization, which helps understand how " + }, + { + "bbox": [ + 55, + 289, + 296, + 350 + ], + "type": "inline_equation", + "content": "\\mathrm{PSNR}_p" + }, + { + "bbox": [ + 55, + 289, + 296, + 350 + ], + "type": "text", + "content": " evolves." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 358, + 294, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 294, + 384 + ], + "type": "text", + "content": "Proposition 2. Let " + }, + { + "bbox": [ + 55, + 358, + 294, + 384 + ], + "type": "inline_equation", + "content": "\\Delta L_{i + 1} = L(w_{i + 1},B_K) - L(w_i,B_K)" + }, + { + "bbox": [ + 55, + 358, + 294, + 384 + ], + "type": "text", + "content": ". The first-order approximation of " + }, + { + "bbox": [ + 55, + 358, + 294, + 384 + ], + "type": "inline_equation", + "content": "\\Delta L_{i + 1}" + }, + { + "bbox": [ + 55, + 358, + 294, + 384 + ], + "type": "text", + "content": " is then" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 121, + 392, + 294, + 406 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 392, + 294, + 406 + ], + "spans": [ + { + "bbox": [ + 121, + 392, + 294, + 406 + ], + "type": "interline_equation", + "content": "\\Delta L _ {i + 1} \\approx - \\lambda_ {o} \\cdot g _ {K} ^ {2} \\leq 0. \\tag {7}", + "image_path": "f480340ee0e0026bed07590074e2bb4a37aabc171504a47b1b7d95f10ee7539f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "type": "text", + "content": "The proof is provided in Appendix I. Note that " + }, + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "type": "inline_equation", + "content": "Q^m" + }, + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "type": "text", + "content": " is replaced by " + }, + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "type": "inline_equation", + "content": "B_K" + }, + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "type": "text", + "content": ", a minibatch of the " + }, + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "type": "inline_equation", + "content": "Q^m" + }, + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "type": "inline_equation", + "content": "\\Delta L_{i+1}" + }, + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "type": "text", + "content": " denotes the change in the loss of the local meta-learner " + }, + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "type": "text", + "content": " on the client's query set. According to Proposition 2, " + }, + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "type": "inline_equation", + "content": "\\Delta L_{i+1}" + }, + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "type": "text", + "content": " is always negative, which means that " + }, + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "type": "inline_equation", + "content": "L(w, Q^m)" + }, + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "type": "text", + "content": " decreases with each outer loop iteration. Since " + }, + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "type": "inline_equation", + "content": "\\mathrm{PSNR}_p" + }, + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "type": "text", + "content": " is inversely proportional to " + }, + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "type": "inline_equation", + "content": "L(w, Q^m)" + }, + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "type": "inline_equation", + "content": "\\mathrm{PSNR}_p" + }, + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "type": "text", + "content": " increases every outer loop iteration. Consequently, to prevent the increase of " + }, + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "type": "inline_equation", + "content": "\\mathrm{PSNR}_p" + }, + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "type": "text", + "content": ", we regularize " + }, + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "type": "inline_equation", + "content": "\\Delta L_{i+1}" + }, + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "type": "text", + "content": " to be close to zero, thereby mitigating the privacy leakage in the local meta-optimization." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 547, + 121, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 547, + 121, + 561 + ], + "spans": [ + { + "bbox": [ + 55, + 547, + 121, + 561 + ], + "type": "text", + "content": "5. Approach" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 567, + 296, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 296, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 296, + 664 + ], + "type": "text", + "content": "We reveal that federated meta-learning (FML) approaches suffer from privacy leakage as their local meta-gradients increase the privacy metric. To address this issue, we propose FedMeNF, a novel FML framework for neural fields that ensures privacy-preserving local meta-optimization. FedMeNF employs a privacy-preserving loss function " + }, + { + "bbox": [ + 55, + 567, + 296, + 664 + ], + "type": "inline_equation", + "content": "L_{pp}" + }, + { + "bbox": [ + 55, + 567, + 296, + 664 + ], + "type": "text", + "content": ", designed to regulate the increase in the privacy metric. Please refer to Algorithm 1 and Fig. 1d for its workflow." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 672, + 213, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 213, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 213, + 685 + ], + "type": "text", + "content": "5.1. The Privacy-Preserving Loss" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Proposition 1-2 theoretically show how client data privacy is leaked in federated meta-learning. Now, we introduce" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 72, + 536, + 84 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 536, + 84 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 536, + 84 + ], + "type": "text", + "content": "Algorithm 1 FedMeNF (see Appendix A for notation)" + } + ] + } + ], + "index": 11, + "type": "text" + }, + { + "bbox": [ + 323, + 87, + 555, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 87, + 555, + 137 + ], + "spans": [ + { + "bbox": [ + 323, + 87, + 555, + 137 + ], + "type": "text", + "content": "Input: (1) " + }, + { + "bbox": [ + 323, + 87, + 555, + 137 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 323, + 87, + 555, + 137 + ], + "type": "text", + "content": ": # of communication rounds, (2) " + }, + { + "bbox": [ + 323, + 87, + 555, + 137 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 323, + 87, + 555, + 137 + ], + "type": "text", + "content": ": # of clients, (3) " + }, + { + "bbox": [ + 323, + 87, + 555, + 137 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 323, + 87, + 555, + 137 + ], + "type": "text", + "content": ": # of participants, (4) " + }, + { + "bbox": [ + 323, + 87, + 555, + 137 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + }, + { + "bbox": [ + 323, + 87, + 555, + 137 + ], + "type": "text", + "content": ": inner loop learning rate, (5) " + }, + { + "bbox": [ + 323, + 87, + 555, + 137 + ], + "type": "inline_equation", + "content": "\\lambda_{o}" + }, + { + "bbox": [ + 323, + 87, + 555, + 137 + ], + "type": "text", + "content": ": outer loop learning rate, (6) " + }, + { + "bbox": [ + 323, + 87, + 555, + 137 + ], + "type": "inline_equation", + "content": "\\alpha^{m}" + }, + { + "bbox": [ + 323, + 87, + 555, + 137 + ], + "type": "text", + "content": ": weight proportional to the client " + }, + { + "bbox": [ + 323, + 87, + 555, + 137 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 323, + 87, + 555, + 137 + ], + "type": "text", + "content": "'s dataset size (see Eq. (2))." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 324, + 138, + 469, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 138, + 469, + 150 + ], + "spans": [ + { + "bbox": [ + 324, + 138, + 469, + 150 + ], + "type": "text", + "content": "Initialize the global meta-learner " + }, + { + "bbox": [ + 324, + 138, + 469, + 150 + ], + "type": "inline_equation", + "content": "\\theta_0" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 324, + 150, + 532, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 150, + 532, + 161 + ], + "spans": [ + { + "bbox": [ + 324, + 150, + 532, + 161 + ], + "type": "text", + "content": "for each communication round " + }, + { + "bbox": [ + 324, + 150, + 532, + 161 + ], + "type": "inline_equation", + "content": "r = 0,\\dots,R - 1" + }, + { + "bbox": [ + 324, + 150, + 532, + 161 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 339, + 162, + 498, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 162, + 498, + 174 + ], + "spans": [ + { + "bbox": [ + 339, + 162, + 498, + 174 + ], + "type": "text", + "content": "Sample " + }, + { + "bbox": [ + 339, + 162, + 498, + 174 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 339, + 162, + 498, + 174 + ], + "type": "text", + "content": " clients from 1,..., " + }, + { + "bbox": [ + 339, + 162, + 498, + 174 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 339, + 162, + 498, + 174 + ], + "type": "text", + "content": " clients" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 340, + 175, + 425, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 175, + 425, + 185 + ], + "spans": [ + { + "bbox": [ + 340, + 175, + 425, + 185 + ], + "type": "text", + "content": "for " + }, + { + "bbox": [ + 340, + 175, + 425, + 185 + ], + "type": "inline_equation", + "content": "m = 1,\\dots,M" + }, + { + "bbox": [ + 340, + 175, + 425, + 185 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 354, + 186, + 468, + 198 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 186, + 468, + 198 + ], + "spans": [ + { + "bbox": [ + 354, + 186, + 468, + 198 + ], + "type": "interline_equation", + "content": "w _ {*} ^ {m} \\leftarrow \\text {L O C A L U P D A T E} \\left(\\theta_ {r}\\right)", + "image_path": "9c2ac52a2692ab2e031c293ec08118a9954246d7ad30708548350e506c1f6f17.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 340, + 199, + 373, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 199, + 373, + 209 + ], + "spans": [ + { + "bbox": [ + 340, + 199, + 373, + 209 + ], + "type": "text", + "content": "end for" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 340, + 210, + 553, + 222 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 210, + 553, + 222 + ], + "spans": [ + { + "bbox": [ + 340, + 210, + 553, + 222 + ], + "type": "interline_equation", + "content": "\\theta_ {r + 1} \\leftarrow \\Sigma_ {m} \\alpha^ {m} w _ {*} ^ {m} \\quad \\triangleright \\text {S e r v e r a g g r e g a t i o n}", + "image_path": "c2dd7f1ee1fb99e18757dda4f65215cbb39b158bc75add0ef887bebc62497ac9.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 325, + 223, + 358, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 223, + 358, + 232 + ], + "spans": [ + { + "bbox": [ + 325, + 223, + 358, + 232 + ], + "type": "text", + "content": "end for" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 325, + 236, + 450, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 236, + 450, + 247 + ], + "spans": [ + { + "bbox": [ + 325, + 236, + 450, + 247 + ], + "type": "text", + "content": "procedure LOCALUPDATE(θ)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 340, + 249, + 501, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 249, + 501, + 260 + ], + "spans": [ + { + "bbox": [ + 340, + 249, + 501, + 260 + ], + "type": "text", + "content": "Initialize the local meta-learner " + }, + { + "bbox": [ + 340, + 249, + 501, + 260 + ], + "type": "inline_equation", + "content": "w_0 \\gets \\theta" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 340, + 261, + 553, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 261, + 553, + 273 + ], + "spans": [ + { + "bbox": [ + 340, + 261, + 553, + 273 + ], + "type": "text", + "content": "for each step " + }, + { + "bbox": [ + 340, + 261, + 553, + 273 + ], + "type": "inline_equation", + "content": "i = 0,\\dots,E - 1" + }, + { + "bbox": [ + 340, + 261, + 553, + 273 + ], + "type": "text", + "content": " do " + }, + { + "bbox": [ + 340, + 261, + 553, + 273 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 340, + 261, + 553, + 273 + ], + "type": "text", + "content": " Outer loop" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 355, + 273, + 530, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 273, + 530, + 285 + ], + "spans": [ + { + "bbox": [ + 355, + 273, + 530, + 285 + ], + "type": "text", + "content": "Sample task data " + }, + { + "bbox": [ + 355, + 273, + 530, + 285 + ], + "type": "inline_equation", + "content": "T_{i} = \\{S_{i},Q_{i}\\}" + }, + { + "bbox": [ + 355, + 273, + 530, + 285 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 355, + 273, + 530, + 285 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{train}}" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 355, + 285, + 553, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 285, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 355, + 285, + 553, + 297 + ], + "type": "inline_equation", + "content": "\\varphi_0 \\gets w_i" + }, + { + "bbox": [ + 355, + 285, + 553, + 297 + ], + "type": "text", + "content": " Initialize the neural field for " + }, + { + "bbox": [ + 355, + 285, + 553, + 297 + ], + "type": "inline_equation", + "content": "T_{i}" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 355, + 297, + 553, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 297, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 355, + 297, + 553, + 308 + ], + "type": "text", + "content": "for each step " + }, + { + "bbox": [ + 355, + 297, + 553, + 308 + ], + "type": "inline_equation", + "content": "k = 0,\\dots,K - 1" + }, + { + "bbox": [ + 355, + 297, + 553, + 308 + ], + "type": "text", + "content": " do " + }, + { + "bbox": [ + 355, + 297, + 553, + 308 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 355, + 297, + 553, + 308 + ], + "type": "text", + "content": " Inner loop" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 370, + 308, + 535, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 308, + 535, + 320 + ], + "spans": [ + { + "bbox": [ + 370, + 308, + 535, + 320 + ], + "type": "text", + "content": "Sample batch " + }, + { + "bbox": [ + 370, + 308, + 535, + 320 + ], + "type": "inline_equation", + "content": "B_{k}" + }, + { + "bbox": [ + 370, + 308, + 535, + 320 + ], + "type": "text", + "content": " from the support set " + }, + { + "bbox": [ + 370, + 308, + 535, + 320 + ], + "type": "inline_equation", + "content": "S_{i}" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 372, + 332, + 553, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 332, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 372, + 332, + 553, + 346 + ], + "type": "interline_equation", + "content": "\\varphi_ {k + 1} \\leftarrow \\varphi_ {k} - \\lambda_ {i} \\nabla_ {\\varphi_ {k}} L \\left(\\varphi_ {k}, B _ {k}\\right) \\tag {8}", + "image_path": "2388c81249ce33e6be92c8f7897a4dab88cf414b7e17180c59fbd2b9c66e2cb6.jpg" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 354, + 352, + 388, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 352, + 388, + 361 + ], + "spans": [ + { + "bbox": [ + 354, + 352, + 388, + 361 + ], + "type": "text", + "content": "end for" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 354, + 363, + 517, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 363, + 517, + 375 + ], + "spans": [ + { + "bbox": [ + 354, + 363, + 517, + 375 + ], + "type": "text", + "content": "Sample batch " + }, + { + "bbox": [ + 354, + 363, + 517, + 375 + ], + "type": "inline_equation", + "content": "B_{K}" + }, + { + "bbox": [ + 354, + 363, + 517, + 375 + ], + "type": "text", + "content": " from the query set " + }, + { + "bbox": [ + 354, + 363, + 517, + 375 + ], + "type": "inline_equation", + "content": "Q_{i}" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 354, + 376, + 507, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 376, + 507, + 387 + ], + "spans": [ + { + "bbox": [ + 354, + 376, + 507, + 387 + ], + "type": "text", + "content": "Select the regularization coefficient " + }, + { + "bbox": [ + 354, + 376, + 507, + 387 + ], + "type": "inline_equation", + "content": "\\gamma" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 355, + 388, + 553, + 401 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 388, + 553, + 401 + ], + "spans": [ + { + "bbox": [ + 355, + 388, + 553, + 401 + ], + "type": "interline_equation", + "content": "w _ {i + 1} \\leftarrow w _ {i} - \\lambda_ {o} \\nabla_ {w _ {i}} \\left(L \\left(\\varphi_ {K}, B _ {K}\\right) - \\gamma L \\left(w _ {i}, B _ {K}\\right)\\right)", + "image_path": "4ae6264076eba4f29e89bc1785ed3ff93fc06e36eb3948c47232ee96058419cc.jpg" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 355, + 401, + 553, + 416 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 401, + 553, + 416 + ], + "spans": [ + { + "bbox": [ + 355, + 401, + 553, + 416 + ], + "type": "interline_equation", + "content": "\\mid \\text {f o r} \\tag {9}", + "image_path": "7b90b1ecd085bf77a828e16d485c44f4d82f9c996f714e5555f3f3d2d2f61e6c.jpg" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 340, + 411, + 373, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 411, + 373, + 419 + ], + "spans": [ + { + "bbox": [ + 340, + 411, + 373, + 419 + ], + "type": "text", + "content": "end for" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 340, + 422, + 378, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 422, + 378, + 433 + ], + "spans": [ + { + "bbox": [ + 340, + 422, + 378, + 433 + ], + "type": "text", + "content": "Send " + }, + { + "bbox": [ + 340, + 422, + 378, + 433 + ], + "type": "inline_equation", + "content": "w_{E}" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 325, + 434, + 389, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 434, + 389, + 445 + ], + "spans": [ + { + "bbox": [ + 325, + 434, + 389, + 445 + ], + "type": "text", + "content": "end procedure" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 313, + 470, + 555, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 470, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 555, + 495 + ], + "type": "text", + "content": "a privacy-preserving loss " + }, + { + "bbox": [ + 313, + 470, + 555, + 495 + ], + "type": "inline_equation", + "content": "L_{pp}" + }, + { + "bbox": [ + 313, + 470, + 555, + 495 + ], + "type": "text", + "content": " as follows, ensuring that it regularizes the " + }, + { + "bbox": [ + 313, + 470, + 555, + 495 + ], + "type": "inline_equation", + "content": "g_{K}" + }, + { + "bbox": [ + 313, + 470, + 555, + 495 + ], + "type": "text", + "content": " term that increases " + }, + { + "bbox": [ + 313, + 470, + 555, + 495 + ], + "type": "inline_equation", + "content": "\\mathrm{PSNR}_p" + }, + { + "bbox": [ + 313, + 470, + 555, + 495 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 319, + 503, + 553, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 503, + 553, + 517 + ], + "spans": [ + { + "bbox": [ + 319, + 503, + 553, + 517 + ], + "type": "interline_equation", + "content": "L _ {p p} (\\gamma , w _ {i}, \\varphi_ {K}, B _ {K}) = L \\left(\\varphi_ {K}, B _ {K}\\right) - \\gamma L \\left(w _ {i}, B _ {K}\\right), \\tag {10}", + "image_path": "aba7e3a6e393cb70d8906f2d2db249bd092b5b6e0404feb1138571efb0ec9006.jpg" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 313, + 525, + 554, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 554, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 554, + 563 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 525, + 554, + 563 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 525, + 554, + 563 + ], + "type": "text", + "content": " is a regularization coefficient that determines the portion of the " + }, + { + "bbox": [ + 313, + 525, + 554, + 563 + ], + "type": "inline_equation", + "content": "g_{K}" + }, + { + "bbox": [ + 313, + 525, + 554, + 563 + ], + "type": "text", + "content": ". In Proposition 3 below, we can obtain the privacy-preserving local meta-gradient " + }, + { + "bbox": [ + 313, + 525, + 554, + 563 + ], + "type": "inline_equation", + "content": "g_{pp}" + }, + { + "bbox": [ + 313, + 525, + 554, + 563 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 313, + 525, + 554, + 563 + ], + "type": "inline_equation", + "content": "L_{pp}" + }, + { + "bbox": [ + 313, + 525, + 554, + 563 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 313, + 569, + 553, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 553, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 553, + 594 + ], + "type": "text", + "content": "Proposition 3. Let " + }, + { + "bbox": [ + 313, + 569, + 553, + 594 + ], + "type": "inline_equation", + "content": "g_{pp} = \\nabla_{w_i}L_{pp}(\\gamma, w_i, \\varphi_K, B_K)" + }, + { + "bbox": [ + 313, + 569, + 553, + 594 + ], + "type": "text", + "content": ". Then, the first-order approximation of " + }, + { + "bbox": [ + 313, + 569, + 553, + 594 + ], + "type": "inline_equation", + "content": "g_{pp}" + }, + { + "bbox": [ + 313, + 569, + 553, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 569, + 553, + 594 + ], + "type": "inline_equation", + "content": "\\Delta L_{i+1}" + }, + { + "bbox": [ + 313, + 569, + 553, + 594 + ], + "type": "text", + "content": " become" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 379, + 604, + 553, + 616 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 379, + 604, + 553, + 616 + ], + "spans": [ + { + "bbox": [ + 379, + 604, + 553, + 616 + ], + "type": "interline_equation", + "content": "g _ {p p} \\approx (1 - \\gamma) \\cdot g _ {K} - \\lambda_ {i} \\mathcal {I} _ {K}, \\tag {11}", + "image_path": "4129fd4ad5f015a0d278cf141618a60ddfeacff510ffea2529d2cff939e866fe.jpg" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 365, + 618, + 553, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 618, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 365, + 618, + 553, + 632 + ], + "type": "interline_equation", + "content": "\\Delta L _ {i + 1} \\approx - \\lambda_ {o} (1 - \\gamma) \\left(g _ {K}\\right) ^ {2} \\leq 0. \\tag {12}", + "image_path": "ff9068c997c3ee5c2bb21920bf0f3a144c10165d3f75076c89db1510638cd9b0.jpg" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": "Note " + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_K" + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": " is defined in Eq. (5). At " + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\gamma = 1" + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": ", the " + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "inline_equation", + "content": "g_K" + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": " term, which causes privacy leakage, is completely removed from " + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "inline_equation", + "content": "g_{pp}" + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": ". This means that the local meta-learner no longer contains information about the client's local dataset and learns only fast optimization by maximizing the inner products between gradients across minibatches. Each client can choose" + } + ] + } + ], + "index": 43 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2164" + } + ] + } + ], + "index": 44 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 76, + 140, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 76, + 140, + 84 + ], + "spans": [ + { + "bbox": [ + 63, + 76, + 140, + 84 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 63, + 76, + 140, + 84 + ], + "type": "text", + "content": " local meta-learner" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 87, + 148, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 87, + 148, + 97 + ], + "spans": [ + { + "bbox": [ + 65, + 87, + 148, + 97 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 65, + 87, + 148, + 97 + ], + "type": "text", + "content": " : # of outer loop steps" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 63, + 99, + 139, + 110 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 99, + 139, + 110 + ], + "spans": [ + { + "bbox": [ + 63, + 99, + 139, + 110 + ], + "type": "inline_equation", + "content": "= = \\gg" + }, + { + "bbox": [ + 63, + 99, + 139, + 110 + ], + "type": "inline_equation", + "content": "\\nabla_{w_i}L(\\varphi_K,B_K)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 113, + 136, + 123 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 113, + 136, + 123 + ], + "spans": [ + { + "bbox": [ + 63, + 113, + 136, + 123 + ], + "type": "inline_equation", + "content": "\\longrightarrow \\nabla_{\\varphi_k}L(\\varphi_k,B_k)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 179, + 75, + 286, + 123 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 182, + 75, + 261, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 75, + 261, + 84 + ], + "spans": [ + { + "bbox": [ + 182, + 75, + 261, + 84 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 182, + 75, + 261, + 84 + ], + "type": "text", + "content": " optimal neural field" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 184, + 87, + 246, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 87, + 246, + 97 + ], + "spans": [ + { + "bbox": [ + 184, + 87, + 246, + 97 + ], + "type": "text", + "content": "i: outer loop step" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 179, + 100, + 264, + 110 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 100, + 264, + 110 + ], + "spans": [ + { + "bbox": [ + 179, + 100, + 264, + 110 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 179, + 100, + 264, + 110 + ], + "type": "inline_equation", + "content": "\\triangleright -\\gamma \\nabla_{w_i}L(w_i,B_K)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 179, + 112, + 286, + 123 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 112, + 286, + 123 + ], + "spans": [ + { + "bbox": [ + 179, + 112, + 286, + 123 + ], + "type": "inline_equation", + "content": "\\Rightarrow \\nabla_{w_i}L_{pp}(\\gamma ,w_i,\\varphi_K,B_K)" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 61, + 165, + 137, + 218 + ], + "blocks": [ + { + "bbox": [ + 61, + 165, + 137, + 218 + ], + "lines": [ + { + "bbox": [ + 61, + 165, + 137, + 218 + ], + "spans": [ + { + "bbox": [ + 61, + 165, + 137, + 218 + ], + "type": "image", + "image_path": "c3fd9092c10ec35fac0117b7357c7502aee7e931e6f9f90a978305c8e1e3d0a9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 220, + 111, + 229 + ], + "lines": [ + { + "bbox": [ + 80, + 220, + 111, + 229 + ], + "spans": [ + { + "bbox": [ + 80, + 220, + 111, + 229 + ], + "type": "text", + "content": "(a) " + }, + { + "bbox": [ + 80, + 220, + 111, + 229 + ], + "type": "inline_equation", + "content": "i = 0" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 141, + 155, + 216, + 218 + ], + "blocks": [ + { + "bbox": [ + 141, + 155, + 216, + 218 + ], + "lines": [ + { + "bbox": [ + 141, + 155, + 216, + 218 + ], + "spans": [ + { + "bbox": [ + 141, + 155, + 216, + 218 + ], + "type": "image", + "image_path": "37f204dc2ac283e8215ef575a521f8a84a48ac47afaf0bc4e0b876f38b8c2f1e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 160, + 220, + 191, + 229 + ], + "lines": [ + { + "bbox": [ + 160, + 220, + 191, + 229 + ], + "spans": [ + { + "bbox": [ + 160, + 220, + 191, + 229 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 160, + 220, + 191, + 229 + ], + "type": "inline_equation", + "content": "i = 1" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 225, + 133, + 295, + 218 + ], + "blocks": [ + { + "bbox": [ + 225, + 133, + 295, + 218 + ], + "lines": [ + { + "bbox": [ + 225, + 133, + 295, + 218 + ], + "spans": [ + { + "bbox": [ + 225, + 133, + 295, + 218 + ], + "type": "image", + "image_path": "05e671ad1b7b490da56a8027cb584a26696148948f55ca2d834b81290292eee3.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 231, + 220, + 279, + 229 + ], + "lines": [ + { + "bbox": [ + 231, + 220, + 279, + 229 + ], + "spans": [ + { + "bbox": [ + 231, + 220, + 279, + 229 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 231, + 220, + 279, + 229 + ], + "type": "inline_equation", + "content": "i = E - 1" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 239, + 296, + 316 + ], + "lines": [ + { + "bbox": [ + 55, + 239, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 296, + 316 + ], + "type": "text", + "content": "Figure 2. Illustration of FedMeNF's local meta-optimization in a client. The local meta-learner " + }, + { + "bbox": [ + 55, + 239, + 296, + 316 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 239, + 296, + 316 + ], + "type": "text", + "content": " does not get closer to the optimal neural field and maintains the initial gap (the distance between the global meta-learner " + }, + { + "bbox": [ + 55, + 239, + 296, + 316 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 239, + 296, + 316 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 239, + 296, + 316 + ], + "type": "inline_equation", + "content": "= w_0" + }, + { + "bbox": [ + 55, + 239, + 296, + 316 + ], + "type": "text", + "content": ") and the optimal neural field). Nevertheless, as meta-optimization progresses, gradients are being aligned, and the local meta-learner can find " + }, + { + "bbox": [ + 55, + 239, + 296, + 316 + ], + "type": "inline_equation", + "content": "w_E" + }, + { + "bbox": [ + 55, + 239, + 296, + 316 + ], + "type": "text", + "content": " from which the optimal neural field is reached with a few gradient steps." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 337, + 295, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 337, + 295, + 372 + ], + "spans": [ + { + "bbox": [ + 55, + 337, + 295, + 372 + ], + "type": "text", + "content": "a different " + }, + { + "bbox": [ + 55, + 337, + 295, + 372 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 337, + 295, + 372 + ], + "type": "text", + "content": " that aligns with its desired privacy guarantee. We will empirically show that as " + }, + { + "bbox": [ + 55, + 337, + 295, + 372 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 337, + 295, + 372 + ], + "type": "text", + "content": " increases, the privacy leakage from the client's local dataset decreases in Sec. 6.2." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 373, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 296, + 445 + ], + "type": "text", + "content": "There are two ways to set " + }, + { + "bbox": [ + 55, + 373, + 296, + 445 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 373, + 296, + 445 + ], + "type": "text", + "content": ": as a fixed hyperparameter or adjusted dynamically. While we use the former approach as our default, we also propose the following adaptive method. Similar to how " + }, + { + "bbox": [ + 55, + 373, + 296, + 445 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 373, + 296, + 445 + ], + "type": "text", + "content": " defines the privacy boundary in Differential Privacy, we introduce a privacy budget " + }, + { + "bbox": [ + 55, + 373, + 296, + 445 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 55, + 373, + 296, + 445 + ], + "type": "text", + "content": ", limiting the total magnitude of the loss change defined in Eq. (12):" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 455, + 294, + 469 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 455, + 294, + 469 + ], + "spans": [ + { + "bbox": [ + 116, + 455, + 294, + 469 + ], + "type": "interline_equation", + "content": "\\left| \\Delta L _ {i + 1} \\right| \\cdot R \\cdot E \\cdot M / N \\leq \\zeta , \\tag {13}", + "image_path": "5782ee9e36d04a2813fad6658f333eebe6cb9e0dec41f0afce3b64ad63a82cb5.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 67, + 471, + 294, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 471, + 294, + 485 + ], + "spans": [ + { + "bbox": [ + 67, + 471, + 294, + 485 + ], + "type": "interline_equation", + "content": "\\gamma = \\min \\left(\\max \\left(1 - N \\zeta / \\left(R E M \\lambda_ {o} \\left(g _ {K}\\right) ^ {2}\\right), 0\\right), 1\\right). \\tag {14}", + "image_path": "d63d523b66b126f471eaca4ec8f29ade68afd6744734c9c842f774a032088f5b.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 495, + 290, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 290, + 508 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 290, + 508 + ], + "type": "text", + "content": "5.2. Privacy-Preserving Local Meta-Optimization" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 513, + 296, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 513, + 296, + 645 + ], + "spans": [ + { + "bbox": [ + 55, + 513, + 296, + 645 + ], + "type": "text", + "content": "Figure 2 illustrates FedMeNF's local meta-optimization. In existing FML approaches, each outer loop iteration brings the local meta-learner " + }, + { + "bbox": [ + 55, + 513, + 296, + 645 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 55, + 513, + 296, + 645 + ], + "type": "text", + "content": " closer to the optimal neural field for the client's private data. Sharing the optimized local meta-learner " + }, + { + "bbox": [ + 55, + 513, + 296, + 645 + ], + "type": "inline_equation", + "content": "w_{E}" + }, + { + "bbox": [ + 55, + 513, + 296, + 645 + ], + "type": "text", + "content": " with the server leads to significant privacy leakage, as the server can directly exploit " + }, + { + "bbox": [ + 55, + 513, + 296, + 645 + ], + "type": "inline_equation", + "content": "w_{E}" + }, + { + "bbox": [ + 55, + 513, + 296, + 645 + ], + "type": "text", + "content": " to reconstruct the original private data. In contrast, FedMeNF explicitly prevents the " + }, + { + "bbox": [ + 55, + 513, + 296, + 645 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 55, + 513, + 296, + 645 + ], + "type": "text", + "content": " from converging to the optimal neural field, focusing solely on gradient alignment. Thus, FedMeNF trains a local meta-learner that retains minimal private data while still enabling efficient optimization." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 657, + 137, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 137, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 137, + 670 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Setup. We assume a scenario where " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "N = 50" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " clients participate in the training. The training consists of 1000 communication rounds, with 5 participants randomly selected" + } + ] + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 322, + 73, + 545, + 180 + ], + "blocks": [ + { + "bbox": [ + 322, + 73, + 545, + 180 + ], + "lines": [ + { + "bbox": [ + 322, + 73, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 322, + 73, + 545, + 180 + ], + "type": "image", + "image_path": "01816b1cae4d92a36ee0cc792a50f019324ce6f48a8a288ccab53136c2f4380a.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 184, + 553, + 207 + ], + "lines": [ + { + "bbox": [ + 313, + 184, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 313, + 184, + 553, + 207 + ], + "type": "text", + "content": "Figure 3. Test-time optimization of the baselines and our Fed-MeNF " + }, + { + "bbox": [ + 313, + 184, + 553, + 207 + ], + "type": "inline_equation", + "content": "(\\gamma = 0.75)" + }, + { + "bbox": [ + 313, + 184, + 553, + 207 + ], + "type": "text", + "content": " using FedAvg [41] on the Cars dataset [6]." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 317, + 220, + 417, + 332 + ], + "blocks": [ + { + "bbox": [ + 317, + 220, + 417, + 332 + ], + "lines": [ + { + "bbox": [ + 317, + 220, + 417, + 332 + ], + "spans": [ + { + "bbox": [ + 317, + 220, + 417, + 332 + ], + "type": "image", + "image_path": "de1d98148821e44e2ee6e5166424d2fbbd83f8109cb520b9a3c8957748e1a33d.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 336, + 401, + 346 + ], + "lines": [ + { + "bbox": [ + 334, + 336, + 401, + 346 + ], + "spans": [ + { + "bbox": [ + 334, + 336, + 401, + 346 + ], + "type": "text", + "content": "(a) " + }, + { + "bbox": [ + 334, + 336, + 401, + 346 + ], + "type": "inline_equation", + "content": "\\mathrm{PSNR}_p / \\mathrm{Rounds}" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 350, + 554, + 429 + ], + "lines": [ + { + "bbox": [ + 313, + 350, + 554, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 554, + 429 + ], + "type": "text", + "content": "Figure 4. Comparison of privacy preservation between baselines and our FedMeNF with different " + }, + { + "bbox": [ + 313, + 350, + 554, + 429 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 350, + 554, + 429 + ], + "type": "text", + "content": " values using FedAvg [41] on the Cars dataset [6]. Baselines like MAML, which excel in reconstruction (PSNR" + }, + { + "bbox": [ + 313, + 350, + 554, + 429 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 313, + 350, + 554, + 429 + ], + "type": "text", + "content": "), exhibit high PSNR" + }, + { + "bbox": [ + 313, + 350, + 554, + 429 + ], + "type": "inline_equation", + "content": "_p" + }, + { + "bbox": [ + 313, + 350, + 554, + 429 + ], + "type": "text", + "content": " value, while methods like Reptile, which have lower PSNR" + }, + { + "bbox": [ + 313, + 350, + 554, + 429 + ], + "type": "inline_equation", + "content": "_p" + }, + { + "bbox": [ + 313, + 350, + 554, + 429 + ], + "type": "text", + "content": " value, perform poorly in reconstruction (PSNR" + }, + { + "bbox": [ + 313, + 350, + 554, + 429 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 313, + 350, + 554, + 429 + ], + "type": "text", + "content": "). FedMeNF effectively adjusts " + }, + { + "bbox": [ + 313, + 350, + 554, + 429 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 350, + 554, + 429 + ], + "type": "text", + "content": " to maintain an acceptable PSNR" + }, + { + "bbox": [ + 313, + 350, + 554, + 429 + ], + "type": "inline_equation", + "content": "_p" + }, + { + "bbox": [ + 313, + 350, + 554, + 429 + ], + "type": "text", + "content": " value while achieving a comparable PSNR." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 420, + 221, + 552, + 332 + ], + "blocks": [ + { + "bbox": [ + 420, + 221, + 552, + 332 + ], + "lines": [ + { + "bbox": [ + 420, + 221, + 552, + 332 + ], + "spans": [ + { + "bbox": [ + 420, + 221, + 552, + 332 + ], + "type": "image", + "image_path": "240aff57adbb309f2e7855b1b799977f02148224b13c1c58187e2665571a6fcb.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 454, + 336, + 517, + 346 + ], + "lines": [ + { + "bbox": [ + 454, + 336, + 517, + 346 + ], + "spans": [ + { + "bbox": [ + 454, + 336, + 517, + 346 + ], + "type": "text", + "content": "(b) PSNR / " + }, + { + "bbox": [ + 454, + 336, + 517, + 346 + ], + "type": "inline_equation", + "content": "\\mathrm{PSNR}_p" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 449, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 554, + 521 + ], + "type": "text", + "content": "per round. We employ the SIREN architecture [51] for image and video modalities, which consists of 6 layers with a hidden dimension of 128. For neural radiance fields (NeRF) [43], we adopt the model used in [9, 58], which consists of 6 layers with a hidden dimension of 256. Please refer to Appendix D for more implementation details." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 521, + 555, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 664 + ], + "type": "text", + "content": "Datasets. We conduct experiments across the datasets of various modalities, including images, videos, and NeRFs. To evaluate performance with limited local training data per client, we select scenarios in which each client has only one or very few tasks. For images, we use the cat category from the PetFace dataset [49], assuming each client has images of a unique cat instance. For videos, we use the GolfDB dataset [42], where each client has videos of only one person's golf swings. For NeRFs, we use the Cars category from the ShapeNet dataset [6] and the FaceScape dataset [65, 70] for human faces. Please refer to Appendix E for detailed distribution information for each dataset." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "Baselines. We test a lower bound baseline referred to as Local, using only local dataset without an FL approach. We construct baselines by combining an FL algorithm and a meta-learning algorithm [8, 18]. For the FL algo" + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2165" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 69, + 553, + 466 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 553, + 466 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 553, + 466 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 553, + 466 + ], + "type": "table", + "html": "
Modality DatasetImage PetFace [49]Video GolfDB [42]3D (NeRF) Cars [6]FaceScape [65, 70]
Method\\Metric\\(PSNR_p\\)\\(PSNR\\)\\(\\Delta(↑)\\)\\(PSNR_p\\)\\(PSNR\\)\\(\\Delta(↑)\\)\\(PSNR_p\\)\\(PSNR\\)\\(\\Delta(↑)\\)\\(PSNR_p\\)\\(PSNR\\)\\(\\Delta(↑)\\)
Local-22.29--26.92--17.13--23.67-
FedAvg + MAML16.5727.3910.8221.2129.688.4719.7323.083.3521.3128.597.28
+ FOMAML18.5223.154.6320.8228.577.7519.7323.663.9321.2428.657.41
+ Reptile17.3922.525.1319.8927.227.3319.9621.982.0221.9228.246.32
+ meta-NSGD12.495.15-7.3410.964.85-6.116.8510.623.777.7211.293.57
+ Ours14.772712.2317.3128.8911.5812.1524.0511.915.1627.8812.72
FedProx + MAML16.5827.4910.9120.9329.268.3319.6722.943.2721.128.547.44
+ FOMAML18.5324.325.7921.2128.767.5519.6723.734.0621.0228.527.5
+ Reptile17.3722.55.1319.7927.217.4219.9522.112.1621.8928.216.32
+ meta-NSGD12.495.15-7.3410.964.84-6.126.8510.914.067.7211.293.57
+ Ours14.4427.0812.6415.6828.9613.2812.1423.9811.8414.0927.5413.45
Scaffold + MAML16.7127.6610.9521.329.117.8119.0724.215.1421.0928.517.42
+ FOMAML18.5523.985.4321.2928.827.5319.0724.475.421.0128.517.5
+ Reptile17.3722.475.12027.237.2319.8122.62.7921.7928.146.35
+ meta-NSGD12.495.13-7.3610.964.84-6.126.8510.633.787.7211.293.57
+ Ours14.9327.1912.2616.2129.2313.0214.1424.349.9413.9427.5313.59
FedNova + MAML16.5227.5110.9921.4729.648.1719.7223.633.9121.3728.597.22
+ FOMAML18.524.15.620.1228.698.5719.7223.273.5521.328.627.32
+ Reptile17.3822.525.1419.9327.267.3319.9622.72.7421.9828.176.19
+ meta-NSGD12.495.14-7.3510.964.84-6.126.8510.63.757.7211.293.57
+ Ours14.9427.1512.2115.7128.9813.2712.1424.1211.9815.4527.8612.41
FedExP + MAML16.5727.3910.8221.1829.588.419.8122.873.0621.3327.646.31
+ FOMAML18.5223.154.6320.8228.67.7819.8122.81321.2627.666.4
+ Reptile17.3922.645.2519.8927.237.3420.9122.031.1221.9328.276.35
+ meta-NSGD12.495.15-7.3410.964.85-6.116.8510.613.767.7211.293.57
+ Ours14.6526.8612.2115.3828.9413.5612.1724.0511.8813.6926.1212.43
FedACG + MAML16.6327.510.8721.0929.518.4219.9222.122.2828.025.74
+ FOMAML18.4824.045.5621.1228.77.5819.921.952.0522.2428.035.79
+ Reptile17.3822.635.2519.8327.237.420.1322.262.1322.0728.326.25
+ meta-NSGD12.495.15-7.3410.964.84-6.126.8510.593.747.7211.293.57
+ Ours14.7126.9712.2615.9428.9913.0510.9322.4511.5214.8327.4612.63
", + "image_path": "5a69397d8e3339daabeb066e4fef44a4a6903d365a6fed572efac9875243f36c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 473, + 555, + 529 + ], + "lines": [ + { + "bbox": [ + 55, + 473, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 555, + 529 + ], + "type": "text", + "content": "Table 1. Comparison of reconstruction (PSNR(↑)) and privacy (PSNRp(↓)) quality across different modalities: images, videos, and 3D NeRF scenes. " + }, + { + "bbox": [ + 55, + 473, + 555, + 529 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 55, + 473, + 555, + 529 + ], + "type": "text", + "content": " is defined as PSNR - PSNRp where a larger value indicates a better trade-off between reconstruction quality and privacy protection. FedMeNF demonstrates a balanced performance by maintaining high PSNR for reconstruction and low PSNRp for privacy. The redhighlighted cases indicate cases where PSNRp is high enough to qualitatively confirm privacy leakage, and the gray-highlighted cases indicate cases where the reconstruction performance is lower than that of the Local." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 75, + 548, + 109, + 591 + ], + "blocks": [ + { + "bbox": [ + 75, + 548, + 109, + 591 + ], + "lines": [ + { + "bbox": [ + 75, + 548, + 109, + 591 + ], + "spans": [ + { + "bbox": [ + 75, + 548, + 109, + 591 + ], + "type": "image", + "image_path": "30bfe43aa9058b497e70d6d564c071706d83e3cc81509f8685758ae286416718.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 594, + 117, + 603 + ], + "lines": [ + { + "bbox": [ + 72, + 594, + 117, + 603 + ], + "spans": [ + { + "bbox": [ + 72, + 594, + 117, + 603 + ], + "type": "text", + "content": "(a) GT (PSNR)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 159, + 538, + 194, + 592 + ], + "blocks": [ + { + "bbox": [ + 159, + 538, + 194, + 592 + ], + "lines": [ + { + "bbox": [ + 159, + 538, + 194, + 592 + ], + "spans": [ + { + "bbox": [ + 159, + 538, + 194, + 592 + ], + "type": "image", + "image_path": "6a8ed5e90f4834cbac859f71963801e160942e0363786dc1f182e01b98a3bd6d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 154, + 594, + 203, + 603 + ], + "lines": [ + { + "bbox": [ + 154, + 594, + 203, + 603 + ], + "spans": [ + { + "bbox": [ + 154, + 594, + 203, + 603 + ], + "type": "text", + "content": "(b) Local (14.97)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 244, + 548, + 278, + 592 + ], + "blocks": [ + { + "bbox": [ + 244, + 548, + 278, + 592 + ], + "lines": [ + { + "bbox": [ + 244, + 548, + 278, + 592 + ], + "spans": [ + { + "bbox": [ + 244, + 548, + 278, + 592 + ], + "type": "image", + "image_path": "9bca89313a2bedfb1d98a8609761513392c458211f8a5da97ddb444de2a5548f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 235, + 594, + 290, + 603 + ], + "lines": [ + { + "bbox": [ + 235, + 594, + 290, + 603 + ], + "spans": [ + { + "bbox": [ + 235, + 594, + 290, + 603 + ], + "type": "text", + "content": "(c) MAML (21.22)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 328, + 548, + 361, + 592 + ], + "blocks": [ + { + "bbox": [ + 328, + 548, + 361, + 592 + ], + "lines": [ + { + "bbox": [ + 328, + 548, + 361, + 592 + ], + "spans": [ + { + "bbox": [ + 328, + 548, + 361, + 592 + ], + "type": "image", + "image_path": "2b89c157dc39217a7d87a3f5b6a3892500bb6c8aa7f0b7dad9e5c2ab52d13728.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 594, + 379, + 603 + ], + "lines": [ + { + "bbox": [ + 315, + 594, + 379, + 603 + ], + "spans": [ + { + "bbox": [ + 315, + 594, + 379, + 603 + ], + "type": "text", + "content": "(d) FOMAML (21.44)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 413, + 548, + 446, + 592 + ], + "blocks": [ + { + "bbox": [ + 413, + 548, + 446, + 592 + ], + "lines": [ + { + "bbox": [ + 413, + 548, + 446, + 592 + ], + "spans": [ + { + "bbox": [ + 413, + 548, + 446, + 592 + ], + "type": "image", + "image_path": "abe7be2f581da1b4ee89431eb1cae6b683ff13b1110504d818a7caf50f62c472.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 404, + 594, + 457, + 603 + ], + "lines": [ + { + "bbox": [ + 404, + 594, + 457, + 603 + ], + "spans": [ + { + "bbox": [ + 404, + 594, + 457, + 603 + ], + "type": "text", + "content": "(e) Reptile (21.18)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 496, + 548, + 530, + 592 + ], + "blocks": [ + { + "bbox": [ + 496, + 548, + 530, + 592 + ], + "lines": [ + { + "bbox": [ + 496, + 548, + 530, + 592 + ], + "spans": [ + { + "bbox": [ + 496, + 548, + 530, + 592 + ], + "type": "image", + "image_path": "0d9655a6a68b363a9c951cfb5ea19449d40bb61aa53311eac4a569098053cafb.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 485, + 594, + 545, + 603 + ], + "lines": [ + { + "bbox": [ + 485, + 594, + 545, + 603 + ], + "spans": [ + { + "bbox": [ + 485, + 594, + 545, + 603 + ], + "type": "text", + "content": "(f) FedMeNF (21.92)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 67, + 608, + 122, + 662 + ], + "blocks": [ + { + "bbox": [ + 67, + 608, + 122, + 662 + ], + "lines": [ + { + "bbox": [ + 67, + 608, + 122, + 662 + ], + "spans": [ + { + "bbox": [ + 67, + 608, + 122, + 662 + ], + "type": "image", + "image_path": "adf67963b6502cdaed9f62900c7ae0272f474411b89c6a450b9f640fa9199e09.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 664, + 116, + 673 + ], + "lines": [ + { + "bbox": [ + 72, + 664, + 116, + 673 + ], + "spans": [ + { + "bbox": [ + 72, + 664, + 116, + 673 + ], + "type": "text", + "content": "(g) GT (PSNR)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 151, + 609, + 206, + 662 + ], + "blocks": [ + { + "bbox": [ + 151, + 609, + 206, + 662 + ], + "lines": [ + { + "bbox": [ + 151, + 609, + 206, + 662 + ], + "spans": [ + { + "bbox": [ + 151, + 609, + 206, + 662 + ], + "type": "image", + "image_path": "02946aba121673d9b6fcdc9e0646309758d3a19b157d654f71c77be6ad30f3c2.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 154, + 663, + 203, + 673 + ], + "lines": [ + { + "bbox": [ + 154, + 663, + 203, + 673 + ], + "spans": [ + { + "bbox": [ + 154, + 663, + 203, + 673 + ], + "type": "text", + "content": "(h) Local (32.69)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 235, + 609, + 289, + 662 + ], + "blocks": [ + { + "bbox": [ + 235, + 609, + 289, + 662 + ], + "lines": [ + { + "bbox": [ + 235, + 609, + 289, + 662 + ], + "spans": [ + { + "bbox": [ + 235, + 609, + 289, + 662 + ], + "type": "image", + "image_path": "0963bd913cce74e77db27b13d8d4b241723865689909287a13543838fa46b515.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 235, + 663, + 289, + 673 + ], + "lines": [ + { + "bbox": [ + 235, + 663, + 289, + 673 + ], + "spans": [ + { + "bbox": [ + 235, + 663, + 289, + 673 + ], + "type": "text", + "content": "(i) MAML (33.17)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 677, + 555, + 700 + ], + "lines": [ + { + "bbox": [ + 55, + 677, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 555, + 700 + ], + "type": "text", + "content": "Figure 5. Qualitative results on the Cars [6] and FaceScape [65, 70] datasets. Test-time optimization was performed for 8192 steps on the Cars dataset and 32768 steps on the FaceScape dataset." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 318, + 609, + 373, + 662 + ], + "blocks": [ + { + "bbox": [ + 318, + 609, + 373, + 662 + ], + "lines": [ + { + "bbox": [ + 318, + 609, + 373, + 662 + ], + "spans": [ + { + "bbox": [ + 318, + 609, + 373, + 662 + ], + "type": "image", + "image_path": "47e2f4e4b7f2471ab5e446ae8a809dffb36500ed4334a9e44b020255935d882e.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 663, + 376, + 673 + ], + "lines": [ + { + "bbox": [ + 315, + 663, + 376, + 673 + ], + "spans": [ + { + "bbox": [ + 315, + 663, + 376, + 673 + ], + "type": "text", + "content": "(j) FOMAML (33.26)" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 402, + 608, + 457, + 662 + ], + "blocks": [ + { + "bbox": [ + 402, + 608, + 457, + 662 + ], + "lines": [ + { + "bbox": [ + 402, + 608, + 457, + 662 + ], + "spans": [ + { + "bbox": [ + 402, + 608, + 457, + 662 + ], + "type": "image", + "image_path": "798fbfe9e346d394933131b51eccc9135a07c3cd96bd75e56fec1d1c4deb6f94.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 663, + 456, + 673 + ], + "lines": [ + { + "bbox": [ + 403, + 663, + 456, + 673 + ], + "spans": [ + { + "bbox": [ + 403, + 663, + 456, + 673 + ], + "type": "text", + "content": "(k) Reptile (32.28)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 485, + 608, + 541, + 662 + ], + "blocks": [ + { + "bbox": [ + 485, + 608, + 541, + 662 + ], + "lines": [ + { + "bbox": [ + 485, + 608, + 541, + 662 + ], + "spans": [ + { + "bbox": [ + 485, + 608, + 541, + 662 + ], + "type": "image", + "image_path": "ccae8ecfe990562d533365d85b09fbccb7e82894916324db588189e568735eaf.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 482, + 663, + 543, + 673 + ], + "lines": [ + { + "bbox": [ + 482, + 663, + 543, + 673 + ], + "spans": [ + { + "bbox": [ + 482, + 663, + 543, + 673 + ], + "type": "text", + "content": "(1) FedMeNF (33.54)" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2166" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": "rithm, we use not only FedAvg [41] as the most standard method, but also state-of-the-art methods including FedProx [38], Scaffold [30], FedNova [61], FedExP [28], and FedACG [32]. For the meta-learning algorithm, we select widely used methods such as MAML [19], FOMAML [19], and Reptile [44], alongside HF-MAML [17, 18] and A-MAML [37]. We also include meta-NSGD [69], the meta-learning algorithm with task-level differential privacy." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 176, + 163, + 189 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 176, + 163, + 189 + ], + "spans": [ + { + "bbox": [ + 55, + 176, + 163, + 189 + ], + "type": "text", + "content": "6.1. Fast Optimization" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 194, + 296, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 194, + 296, + 278 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 296, + 278 + ], + "type": "text", + "content": "Figure 3 illustrates the test-time optimization (TTO) performance of FedMeNF against baselines. Compared to Local, FML approaches reach higher PSNR more quickly. FedMeNF achieves " + }, + { + "bbox": [ + 55, + 194, + 296, + 278 + ], + "type": "inline_equation", + "content": "40.4\\%" + }, + { + "bbox": [ + 55, + 194, + 296, + 278 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 194, + 296, + 278 + ], + "type": "inline_equation", + "content": "74.64\\%" + }, + { + "bbox": [ + 55, + 194, + 296, + 278 + ], + "type": "text", + "content": " higher PSNR than Local with and without TTO, respectively. FedMeNF also achieves a higher PSNR in just 39 steps, making it about " + }, + { + "bbox": [ + 55, + 194, + 296, + 278 + ], + "type": "inline_equation", + "content": "210\\mathrm{x}" + }, + { + "bbox": [ + 55, + 194, + 296, + 278 + ], + "type": "text", + "content": " faster than Local, which requires 8192 steps." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 286, + 176, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 286, + 176, + 298 + ], + "spans": [ + { + "bbox": [ + 55, + 286, + 176, + 298 + ], + "type": "text", + "content": "6.2. Privacy Preservation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 304, + 296, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 296, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 296, + 518 + ], + "type": "text", + "content": "Figure 4 illustrates the privacy preservation performance across varying values of the regularization coefficient " + }, + { + "bbox": [ + 55, + 304, + 296, + 518 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 304, + 296, + 518 + ], + "type": "text", + "content": ". Fig. 4a presents the privacy metric " + }, + { + "bbox": [ + 55, + 304, + 296, + 518 + ], + "type": "inline_equation", + "content": "\\mathrm{PSNR}_p" + }, + { + "bbox": [ + 55, + 304, + 296, + 518 + ], + "type": "text", + "content": " over communication rounds. FedMeNF achieves an " + }, + { + "bbox": [ + 55, + 304, + 296, + 518 + ], + "type": "inline_equation", + "content": "18.25\\%" + }, + { + "bbox": [ + 55, + 304, + 296, + 518 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 304, + 296, + 518 + ], + "type": "inline_equation", + "content": "35.17\\%" + }, + { + "bbox": [ + 55, + 304, + 296, + 518 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 304, + 296, + 518 + ], + "type": "inline_equation", + "content": "38.42\\%" + }, + { + "bbox": [ + 55, + 304, + 296, + 518 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 304, + 296, + 518 + ], + "type": "inline_equation", + "content": "65.28\\%" + }, + { + "bbox": [ + 55, + 304, + 296, + 518 + ], + "type": "text", + "content": " improvement in the privacy metric " + }, + { + "bbox": [ + 55, + 304, + 296, + 518 + ], + "type": "inline_equation", + "content": "\\mathrm{PSNR}_p" + }, + { + "bbox": [ + 55, + 304, + 296, + 518 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 55, + 304, + 296, + 518 + ], + "type": "inline_equation", + "content": "\\gamma = 0.25, 0.5" + }, + { + "bbox": [ + 55, + 304, + 296, + 518 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 304, + 296, + 518 + ], + "type": "inline_equation", + "content": "0.75" + }, + { + "bbox": [ + 55, + 304, + 296, + 518 + ], + "type": "text", + "content": ", compared to the baseline. Fig. 4b shows the trade-off between privacy preservation " + }, + { + "bbox": [ + 55, + 304, + 296, + 518 + ], + "type": "inline_equation", + "content": "(\\mathrm{PSNR}_p\\downarrow)" + }, + { + "bbox": [ + 55, + 304, + 296, + 518 + ], + "type": "text", + "content": " and reconstruction performance " + }, + { + "bbox": [ + 55, + 304, + 296, + 518 + ], + "type": "inline_equation", + "content": "(\\mathrm{PSNR}\\uparrow)" + }, + { + "bbox": [ + 55, + 304, + 296, + 518 + ], + "type": "text", + "content": ". Meta-NSGD, which utilizes the differential privacy mechanism, achieves lower " + }, + { + "bbox": [ + 55, + 304, + 296, + 518 + ], + "type": "inline_equation", + "content": "\\mathrm{PSNR}_p" + }, + { + "bbox": [ + 55, + 304, + 296, + 518 + ], + "type": "text", + "content": " values, indicating stronger privacy protection. However, it suffers from significant degradation in reconstruction performance. In contrast, existing meta-learning methods like MAML achieve high PSNR at the cost of greater vulnerability to privacy leakage with their higher " + }, + { + "bbox": [ + 55, + 304, + 296, + 518 + ], + "type": "inline_equation", + "content": "\\mathrm{PSNR}_p" + }, + { + "bbox": [ + 55, + 304, + 296, + 518 + ], + "type": "text", + "content": ". FedMeNF establishes an efficient frontier that balances privacy protection and reconstruction performance, effectively reducing " + }, + { + "bbox": [ + 55, + 304, + 296, + 518 + ], + "type": "inline_equation", + "content": "\\mathrm{PSNR}_p" + }, + { + "bbox": [ + 55, + 304, + 296, + 518 + ], + "type": "text", + "content": " with minimal impact on PSNR. Qualitative results are presented in Figs. 5 and 6." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 520, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 296, + 579 + ], + "type": "text", + "content": "As shown in Tab. 1, FedMeNF achieves comparable PSNR while maintaining an acceptable " + }, + { + "bbox": [ + 55, + 520, + 296, + 579 + ], + "type": "inline_equation", + "content": "\\mathrm{PSNR}_p" + }, + { + "bbox": [ + 55, + 520, + 296, + 579 + ], + "type": "text", + "content": " across diverse modalities, as well as across various FL algorithms. FedMeNF also performs well on other metrics, such as SSIM and LPIPS, as detailed in Appendix F." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 588, + 244, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 244, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 244, + 601 + ], + "type": "text", + "content": "6.3. Justification on the Privacy Metrics" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "To investigate the practical relevance of the privacy metrics in capturing privacy leakage, we simulate two standard privacy attacks on the Cars dataset [6], assuming an honest-but-curious server. The target models for these attacks are defined as the final local meta-learners from each client in the last communication round. The first is the Membership Inference Attack [50], which determines if a given data sample was in the training set of a black-box model. We assess how accurately the server can predict the membership" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 326, + 69, + 391, + 134 + ], + "blocks": [ + { + "bbox": [ + 326, + 69, + 391, + 134 + ], + "lines": [ + { + "bbox": [ + 326, + 69, + 391, + 134 + ], + "spans": [ + { + "bbox": [ + 326, + 69, + 391, + 134 + ], + "type": "image", + "image_path": "346fd516277afdbb2e487ac80d55b506d3ffebb64720e0349b7fd329891b7c90.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 405, + 70, + 469, + 134 + ], + "blocks": [ + { + "bbox": [ + 405, + 70, + 469, + 134 + ], + "lines": [ + { + "bbox": [ + 405, + 70, + 469, + 134 + ], + "spans": [ + { + "bbox": [ + 405, + 70, + 469, + 134 + ], + "type": "image", + "image_path": "a562eafe4f9ad02621d3366237a54a72ea5ee7f9c60c0587c8351c8bad8a0a81.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 484, + 70, + 548, + 134 + ], + "blocks": [ + { + "bbox": [ + 484, + 70, + 548, + 134 + ], + "lines": [ + { + "bbox": [ + 484, + 70, + 548, + 134 + ], + "spans": [ + { + "bbox": [ + 484, + 70, + 548, + 134 + ], + "type": "image", + "image_path": "c6d47c7949e73ff24a57289db91d133396ef1109d2f62e1270b717eeb398a1eb.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 322, + 144, + 547, + 174 + ], + "blocks": [ + { + "bbox": [ + 333, + 136, + 382, + 144 + ], + "lines": [ + { + "bbox": [ + 333, + 136, + 382, + 144 + ], + "spans": [ + { + "bbox": [ + 333, + 136, + 382, + 144 + ], + "type": "text", + "content": "(a) GT (PSNR " + }, + { + "bbox": [ + 333, + 136, + 382, + 144 + ], + "type": "inline_equation", + "content": "{}_{p}" + }, + { + "bbox": [ + 333, + 136, + 382, + 144 + ], + "type": "text", + "content": " )" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 410, + 136, + 464, + 144 + ], + "lines": [ + { + "bbox": [ + 410, + 136, + 464, + 144 + ], + "spans": [ + { + "bbox": [ + 410, + 136, + 464, + 144 + ], + "type": "text", + "content": "(b) MAML (17.33)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 487, + 136, + 545, + 144 + ], + "lines": [ + { + "bbox": [ + 487, + 136, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 487, + 136, + 545, + 144 + ], + "type": "text", + "content": "(c) FedMeNF (5.84)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 322, + 144, + 547, + 174 + ], + "lines": [ + { + "bbox": [ + 322, + 144, + 547, + 174 + ], + "spans": [ + { + "bbox": [ + 322, + 144, + 547, + 174 + ], + "type": "image", + "image_path": "f074aec58906f74817bb3411f355b8bfe534e707e725da5da6e0960ae544de33.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 175, + 480, + 184 + ], + "lines": [ + { + "bbox": [ + 389, + 175, + 480, + 184 + ], + "spans": [ + { + "bbox": [ + 389, + 175, + 480, + 184 + ], + "type": "text", + "content": "(d) MAML (PSNR " + }, + { + "bbox": [ + 389, + 175, + 480, + 184 + ], + "type": "inline_equation", + "content": "p = 23.43" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 323, + 185, + 547, + 214 + ], + "blocks": [ + { + "bbox": [ + 323, + 185, + 547, + 214 + ], + "lines": [ + { + "bbox": [ + 323, + 185, + 547, + 214 + ], + "spans": [ + { + "bbox": [ + 323, + 185, + 547, + 214 + ], + "type": "image", + "image_path": "bae1a89f076ecaf7faedd61420b43916b067c49570eae3504f56e74ff01bd0e8.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 215, + 480, + 224 + ], + "lines": [ + { + "bbox": [ + 389, + 215, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 389, + 215, + 480, + 224 + ], + "type": "text", + "content": "(e) FedMeNF (PSNR " + }, + { + "bbox": [ + 389, + 215, + 480, + 224 + ], + "type": "inline_equation", + "content": "p = 8.38" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 326, + 232, + 391, + 297 + ], + "blocks": [ + { + "bbox": [ + 326, + 232, + 391, + 297 + ], + "lines": [ + { + "bbox": [ + 326, + 232, + 391, + 297 + ], + "spans": [ + { + "bbox": [ + 326, + 232, + 391, + 297 + ], + "type": "image", + "image_path": "81c9455ca49a80572e8c4c5a0fc9c9ba787f2c398ef117ba0020ac7ddd30aa76.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 336, + 297, + 380, + 306 + ], + "lines": [ + { + "bbox": [ + 336, + 297, + 380, + 306 + ], + "spans": [ + { + "bbox": [ + 336, + 297, + 380, + 306 + ], + "type": "text", + "content": "(f) GT (PSNR)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 405, + 232, + 469, + 296 + ], + "blocks": [ + { + "bbox": [ + 405, + 232, + 469, + 296 + ], + "lines": [ + { + "bbox": [ + 405, + 232, + 469, + 296 + ], + "spans": [ + { + "bbox": [ + 405, + 232, + 469, + 296 + ], + "type": "image", + "image_path": "102e6271192cebeb87c738a9b452c3138b8a026afd0c13d67837cb5892a47b15.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 409, + 297, + 464, + 306 + ], + "lines": [ + { + "bbox": [ + 409, + 297, + 464, + 306 + ], + "spans": [ + { + "bbox": [ + 409, + 297, + 464, + 306 + ], + "type": "text", + "content": "(g) MAML (35.74)" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 484, + 232, + 548, + 296 + ], + "blocks": [ + { + "bbox": [ + 484, + 232, + 548, + 296 + ], + "lines": [ + { + "bbox": [ + 484, + 232, + 548, + 296 + ], + "spans": [ + { + "bbox": [ + 484, + 232, + 548, + 296 + ], + "type": "image", + "image_path": "c0d3d363bc2f5f15bb3e6e9f67a14e0c79de98a5cc9ccd28b017dcdbb35d763d.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 485, + 297, + 546, + 306 + ], + "lines": [ + { + "bbox": [ + 485, + 297, + 546, + 306 + ], + "spans": [ + { + "bbox": [ + 485, + 297, + 546, + 306 + ], + "type": "text", + "content": "(h) FedMeNF (35.16)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 322, + 308, + 547, + 336 + ], + "blocks": [ + { + "bbox": [ + 322, + 308, + 547, + 336 + ], + "lines": [ + { + "bbox": [ + 322, + 308, + 547, + 336 + ], + "spans": [ + { + "bbox": [ + 322, + 308, + 547, + 336 + ], + "type": "image", + "image_path": "8054425fce175f7d6e170e2b40f8e2524424d0cd7ea1d44fffe1bf7dece6470a.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 394, + 337, + 475, + 346 + ], + "lines": [ + { + "bbox": [ + 394, + 337, + 475, + 346 + ], + "spans": [ + { + "bbox": [ + 394, + 337, + 475, + 346 + ], + "type": "text", + "content": "(i) MAML (PSNR = 29.65)" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 321, + 347, + 545, + 376 + ], + "blocks": [ + { + "bbox": [ + 321, + 347, + 545, + 376 + ], + "lines": [ + { + "bbox": [ + 321, + 347, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 321, + 347, + 545, + 376 + ], + "type": "image", + "image_path": "305aec68b39ebd23d161afae99c35d4166218c9e7228649800ad2347997a4bf4.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 377, + 479, + 386 + ], + "lines": [ + { + "bbox": [ + 388, + 377, + 479, + 386 + ], + "spans": [ + { + "bbox": [ + 388, + 377, + 479, + 386 + ], + "type": "text", + "content": "(j) FedMeNF (PSNR = 29.15)" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 389, + 555, + 456 + ], + "lines": [ + { + "bbox": [ + 313, + 389, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 555, + 456 + ], + "type": "text", + "content": "Figure 6. Upper. Reconstruction results of the client's private image/video on the server: (b, d) using MAML and (c, e) using FedMeNF. Lower. Reconstruction results of a new private image/video on the client: (g, i) using MAML and (h, j) using FedMeNF. Our FedMeNF effectively enhances privacy as in (c, e) with minimal impact on the reconstruction quality as in (h, j)." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 479, + 555, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 555, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 555, + 575 + ], + "type": "text", + "content": "of the client's local data using the client's local meta-learner as the black-box model. The second is the Property Inference Attack [20], which infers a property of a given model's training set. We evaluate the server's accuracy in classifying the vehicle category of the client's car using the client's local meta-learner. The setup is detailed in Appendix C. Table 2 shows that worse privacy metrics correlate with greater vulnerability to privacy attacks." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 576, + 555, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 555, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 555, + 636 + ], + "type": "text", + "content": "Moreover, we examine the correlation between the privacy metrics and " + }, + { + "bbox": [ + 313, + 576, + 555, + 636 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 576, + 555, + 636 + ], + "type": "text", + "content": " of the differential privacy framework using meta-NSGD [69]. Figure 7 shows that the privacy metrics degrade as " + }, + { + "bbox": [ + 313, + 576, + 555, + 636 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 576, + 555, + 636 + ], + "type": "text", + "content": " increases, supporting their generalizability as a measure of privacy leakage." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 647, + 503, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 503, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 503, + 659 + ], + "type": "text", + "content": "6.4. Generalization on Non-IID Settings" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "To evaluate the generalization capability of FedMeNF on non-IID settings, we vary the data quantity per client using a Dirichlet distribution with parameter " + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "; for the NeRF scenes (where each client has one task), we vary the num" + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2167" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 293, + 178 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 293, + 178 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 293, + 178 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 293, + 178 + ], + "type": "table", + "html": "
Metric MethodMIA Acc↓PIA Acc↓Privacy MetricsNVS PSNR↑
PSNRp↓SSIMp↓LPIPSp↑
MAML95.8348.7819.730.8270.34323.08
FOMAML97.2248.7819.730.8270.34323.66
Reptile95.8343.9019.960.8390.31121.98
HF-MAML88.8943.9019.240.8270.34123.58
A-MAML95.8343.9019.640.8260.34323.13
Ours55.5629.2712.150.7980.53724.05
Ours (adaptive)55.5629.2712.090.7970.53823.19
", + "image_path": "41c5224fa421ca3c2da650b0e65e13f92794fef814dbaf0f7aa43acbabd34afd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 61, + 232, + 290, + 312 + ], + "blocks": [ + { + "bbox": [ + 61, + 232, + 290, + 312 + ], + "lines": [ + { + "bbox": [ + 61, + 232, + 290, + 312 + ], + "spans": [ + { + "bbox": [ + 61, + 232, + 290, + 312 + ], + "type": "image", + "image_path": "4c390122f4c083c3c3c73865f17bf8bac6613ff22516fda7c9c7ff9e9d37a85f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 79, + 313, + 271, + 324 + ], + "lines": [ + { + "bbox": [ + 79, + 313, + 271, + 324 + ], + "spans": [ + { + "bbox": [ + 79, + 313, + 271, + 324 + ], + "type": "text", + "content": "Figure 7. Correlation between " + }, + { + "bbox": [ + 79, + 313, + 271, + 324 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 79, + 313, + 271, + 324 + ], + "type": "text", + "content": " and privacy metrics." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 351, + 295, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 351, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 351, + 295, + 399 + ], + "type": "text", + "content": "ber of input views, while for the image and video modalities (where each task is a single image or video), we vary the number of tasks. A larger " + }, + { + "bbox": [ + 55, + 351, + 295, + 399 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 351, + 295, + 399 + ], + "type": "text", + "content": " indicates that the data quantity per client is more uniform. See Appendix E.3 for details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 401, + 295, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 401, + 295, + 570 + ], + "spans": [ + { + "bbox": [ + 54, + 401, + 295, + 570 + ], + "type": "text", + "content": "Table 1 presents the results for " + }, + { + "bbox": [ + 54, + 401, + 295, + 570 + ], + "type": "inline_equation", + "content": "\\alpha = 5.0" + }, + { + "bbox": [ + 54, + 401, + 295, + 570 + ], + "type": "text", + "content": ", while Tab. 3 shows the results for " + }, + { + "bbox": [ + 54, + 401, + 295, + 570 + ], + "type": "inline_equation", + "content": "\\alpha = 10,5.0,1.0" + }, + { + "bbox": [ + 54, + 401, + 295, + 570 + ], + "type": "text", + "content": ". For the baselines, a small " + }, + { + "bbox": [ + 54, + 401, + 295, + 570 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 54, + 401, + 295, + 570 + ], + "type": "text", + "content": " creates two problems. First, local meta-optimization on a few views tends to result in over-fitted local models, while training on many views leads to well-trained ones. In both cases, the server can reconstruct clients' private data with high fidelity, leading to privacy leakage. Second, over-fitted models perform poorly on novel view synthesis, degrading the global meta-learner. In contrast, FedMeNF is robust to this data imbalance because it only learns how to optimize rather than simply memorizing data. This approach prevents over-fitting and thus achieves robust performance in both privacy preservation and novel view synthesis, regardless of the data quantity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 585, + 223, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 585, + 223, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 585, + 223, + 598 + ], + "type": "text", + "content": "6.5. Few-shot Novel View Synthesis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": "We conduct experiments with varying average number of input views using FedAvg [41] on the Cars [6] dataset, and the results are shown in Tab. 4. For all methods, the view synthesis performance (PSNR) improves as the average number of input views increases. FML approaches outperform the Local across all cases. Notably, FedMeNF achieves the highest PSNR with acceptable " + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{PSNR}_p" + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": ", regardless of the average number of input views. This confirms that FedMeNF performs well even in few-shot settings." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 553, + 178 + ], + "blocks": [ + { + "bbox": [ + 55, + 182, + 295, + 217 + ], + "lines": [ + { + "bbox": [ + 55, + 182, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 295, + 217 + ], + "type": "text", + "content": "Table 2. Comparison of Membership Inference Attack (MIA) accuracy, Property Inference Attack (PIA) accuracy, Privacy Metrics, and Novel View Synthesis (NVS) quality." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 70, + 553, + 178 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 178 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 178 + ], + "type": "table", + "html": "
Dirichlet α\nMethodα = 10α = 5.0α = 1.0
PSNRpPSNRPSNRpPSNRPSNRpPSNR
MAML19.6223.3319.7323.082021.24
FOMAML19.6223.119.7323.662021.25
Reptile19.9622.4719.9621.9819.9521.29
HF-MAML19.1823.4919.2423.5818.8422.2
A-MAML19.5723.2919.6423.1319.7521.44
Ours12.1323.4712.1524.0512.7324.32
Ours (adaptive)12.123.5212.0923.1913.7320.06
", + "image_path": "2244669e6d8fd4bf5d1034a5d009887e27f87486c42cc8ad275e7fb3c457e12e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 317, + 217, + 553, + 325 + ], + "blocks": [ + { + "bbox": [ + 314, + 182, + 553, + 206 + ], + "lines": [ + { + "bbox": [ + 314, + 182, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 314, + 182, + 553, + 206 + ], + "type": "text", + "content": "Table 3. Results of the novel view synthesis performance and privacy metric with different Dirichlet parameters " + }, + { + "bbox": [ + 314, + 182, + 553, + 206 + ], + "type": "inline_equation", + "content": "\\alpha = \\left\\lbrack {{10},{5.0},{1.0}}\\right\\rbrack" + }, + { + "bbox": [ + 314, + 182, + 553, + 206 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 217, + 553, + 325 + ], + "lines": [ + { + "bbox": [ + 317, + 217, + 553, + 325 + ], + "spans": [ + { + "bbox": [ + 317, + 217, + 553, + 325 + ], + "type": "table", + "html": "
Avg. # of views Method2-view4-view8-view
\\(PSNR_p\\)\\(PSNR\\)\\(PSNR_p\\)\\(PSNR\\)\\(PSNR_p\\)\\(PSNR\\)
MAML19.7320.2319.7323.0819.5426.03
FOMAML19.7420.3919.7323.6619.5425.97
Reptile18.5920.819.9621.9820.3426.3
HF-MAML18.8321.1519.2423.5818.4125.62
A-MAML19.7320.2519.6423.1319.5725.22
Ours15.5621.2712.1524.0511.826.31
Ours (adaptive)15.8420.2112.0923.1912.0325.82
", + "image_path": "688a9050e24e94816069d3c5e9177d572e4ee56b2ba2c4801bddd89d6d0a8422.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 329, + 553, + 353 + ], + "lines": [ + { + "bbox": [ + 313, + 329, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 553, + 353 + ], + "type": "text", + "content": "Table 4. Results of the novel view synthesis performance and privacy metric for 2, 4, and 8 input views per task." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 314, + 373, + 384, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 373, + 384, + 385 + ], + "spans": [ + { + "bbox": [ + 314, + 373, + 384, + 385 + ], + "type": "text", + "content": "7. Discussion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 393, + 555, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 555, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 555, + 478 + ], + "type": "text", + "content": "Since FedMeNF is designed for distributed devices, we use a relatively simple NF model. Although recent work [9, 31, 34] has achieved generalizable NFs using large models like transformers, these methods are often impractical for FL with edge devices due to their high communication and computational demands. Future work should focus on scalable, generalizable NFs aligned with FL constraints." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 489, + 388, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 489, + 388, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 489, + 388, + 502 + ], + "type": "text", + "content": "8. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": "We proposed FedMeNF, the first federated meta-learning framework explicitly designed for neural fields of private data. FedMeNF regulates the increase of our intuitive and widely applicable privacy metric to train the global neural fields meta-learner while preserving privacy. Through experiments across diverse data modalities, including images, videos, and NeRFs, we demonstrate that FedMeNF effectively preserves client privacy while achieving rapid optimization speed and high reconstruction performance, even when local data are highly limited or non-IID. A potential limitation of this study is that we used relatively simple neural field models and vision data only. Future work will investigate generalizable, privacy-preserving neural fields for a broader range of personal data. Measuring and preserving privacy in neural fields is a crucial and novel challenge which has received less attention. We believe this study will serve as an essential starting point for further research." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2168" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 159, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 159, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 159, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 259 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 259 + ], + "type": "text", + "content": "This work was partially supported by the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. RS-2022-II220156, Fundamental research on continual meta-learning for quality enhancement of casual videos and their 3D metaverse transformation; No. RS-2021-II211343, Artificial Intelligence Graduate School Program (Seoul National University)), National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. 2023R1A2C2005573), and Center for Applied Research in Artificial Intelligence (CARAI) grant funded by Defense Acquisition Program Administration (DAPA) and Agency for Defense Development (ADD) (UD230017TD). Gunhee Kim is the corresponding author." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 269, + 115, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 269, + 115, + 281 + ], + "spans": [ + { + "bbox": [ + 56, + 269, + 115, + 281 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 289, + 294, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 289, + 294, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 289, + 294, + 322 + ], + "spans": [ + { + "bbox": [ + 61, + 289, + 294, + 322 + ], + "type": "text", + "content": "[1] Martin Abadi, Andy Chu, Ian Goodfellow, H Brendan McMahan, Ilya Mironov, Kunal Talwar, and Li Zhang. Deep learning with differential privacy. In ACM CCS, 2016. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 323, + 294, + 377 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 323, + 294, + 377 + ], + "spans": [ + { + "bbox": [ + 61, + 323, + 294, + 377 + ], + "type": "text", + "content": "[2] Sawsan AbdulRahman, Hanine Tout, Hakima Ould-Slimane, Azzam Mourad, Chamseddine Talhi, and Mohsen Guizani. A survey on federated learning: The journey from centralized to distributed on-site learning and beyond. IEEE Internet of Things Journal, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 379, + 294, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 379, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 62, + 379, + 294, + 421 + ], + "type": "text", + "content": "[3] Rohith Agaram, Shaurya Dewan, Rahul Sajnani, Adrien Poulenard, Madhava Krishna, and Srinath Sridhar. Canonical fields: Self-supervised learning of pose-canonicalized neural fields. In CVPR, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 423, + 294, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 423, + 294, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 423, + 294, + 456 + ], + "type": "text", + "content": "[4] Mohammed Aledhari, Rehma Razzak, Reza M Parizi, and Fahad Saeed. Federated learning: A survey on enabling technologies, protocols, and applications. IEEE Access, 2020. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 63, + 457, + 294, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 457, + 294, + 489 + ], + "spans": [ + { + "bbox": [ + 63, + 457, + 294, + 489 + ], + "type": "text", + "content": "[5] Mark Bun and Thomas Steinke. Concentrated differential privacy: Simplifications, extensions, and lower bounds. In Theory of Cryptography, 2016. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 491, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 294, + 544 + ], + "type": "text", + "content": "[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv:1512.03012, 2015. 5, 6, 7, 8, 3, 4, 10, 13" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 546, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 546, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 546, + 294, + 578 + ], + "type": "text", + "content": "[7] Daoyuan Chen, Dawei Gao, Weirui Kuang, Yaliang Li, and Bolin Ding. pfl-bench: A comprehensive benchmark for personalized federated learning. NeurIPS, 2022. 3, 4, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 580, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 580, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 62, + 580, + 294, + 622 + ], + "type": "text", + "content": "[8] Fei Chen, Mi Luo, Zhenhua Dong, Zhenguo Li, and Xiuqiang He. Federated meta-learning with fast convergence and efficient communication. arXiv:1802.07876, 2018. 1, 3, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 624, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 624, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 62, + 624, + 294, + 656 + ], + "type": "text", + "content": "[9] Yinbo Chen and Xiaolong Wang. Transformers as metalearners for implicit neural representations. In ECCV, 2022. 2, 4, 5, 8, 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 58, + 658, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 658, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 58, + 658, + 294, + 689 + ], + "type": "text", + "content": "[10] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In CVPR, 2021. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 692, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 294, + 714 + ], + "type": "text", + "content": "[11] Zeyuan Chen, Yinbo Chen, Jingwen Liu, Xingqian Xu, Vedit Goel, Zhangyang Wang, Humphrey Shi, and Xiaolong" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "text", + "content": "Wang. Videoirr: Learning video implicit neural representation for continuous space-time super-resolution. In CVPR, 2022. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 107, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 139 + ], + "type": "text", + "content": "[12] Stephen Coombes, Peter beim Graben, Roland Potthast, and James Wright. Neural fields: theory and applications. Springer, 2014. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 141, + 553, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 174 + ], + "type": "text", + "content": "[13] Moming Duan, Qinbin Li, Linshan Jiang, and Bingsheng He. Towards open federated learning platforms: Survey and vision from technical and legal perspectives, 2024. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "type": "text", + "content": "[14] Emilien Dupont, Hyunjik Kim, SM Eslami, Danilo Rezende, and Dan Rosenbaum. From data to functa: Your data point is a function and you can treat it like one. arXiv:2201.12204, 2022. 1, 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 220, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 240 + ], + "type": "text", + "content": "[15] Cynthia Dwork and Guy N Rothblum. Concentrated differential privacy. arXiv:1603.01887, 2016. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 243, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 274 + ], + "type": "text", + "content": "[16] Cynthia Dwork, Krishnamaram Kenthapadi, Frank McSherry, Ilya Mironov, and Moni Naor. Our data, ourselves: Privacy via distributed noise generation. In EUROCRYPT, 2006. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 276, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 308 + ], + "type": "text", + "content": "[17] Alireza Fallah, Aryan Mokhtari, and Asuman Ozdaglar. On the convergence theory of gradient-based model-agnostic meta-learning algorithms. In AISTATS, 2020. 7, 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 310, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 553, + 352 + ], + "type": "text", + "content": "[18] Alireza Fallah, Aryan Mokhtari, and Asuman Ozdaglar. Personalized federated learning with theoretical guarantees: A model-agnostic meta-learning approach. NeurIPS, 2020. 1, 3, 5, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 354, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 387 + ], + "type": "text", + "content": "[19] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In ICML, 2017. 1, 7, 6, 10" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "type": "text", + "content": "[20] Karan Ganju, Qi Wang, Wei Yang, Carl A Gunter, and Nikita Borisov. Property inference attacks on fully connected neural networks using permutation invariant representations. In ACM CCS, pages 619-633, 2018. 7, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 434, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 475 + ], + "type": "text", + "content": "[21] Wei Gao, Shangwei Guo, Tianwei Zhang, Han Qiu, Yonggang Wen, and Yang Liu. Privacy-preserving collaborative learning with automatic transformation search. In CVPR, 2021. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 478, + 553, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 509 + ], + "type": "text", + "content": "[22] Jeffrey Gu, Kuan-Chieh Wang, and Serena Yeung. Generalizable neural fields as partially observed neural processes. In ICCV, 2023. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 511, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 511, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 511, + 553, + 533 + ], + "type": "text", + "content": "[23] Jamie Hayes, Borja Balle, and Saeed Mahloujifar. Bounding training data reconstruction in dp-sgd. In NeurIPS, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 535, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 535, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 316, + 535, + 553, + 566 + ], + "type": "text", + "content": "[24] Lachlan Holden, Feras Dayoub, David Harvey, and Tat-Jun Chin. Federated neural radiance fields. arXiv:2305.01163, 2023. 1, 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 568, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 601 + ], + "type": "text", + "content": "[25] Minui Hong, Junhyeog Yun, Insu Jeon, and Gunhee Kim. Fedavp: Augment local data via shared policy in federated learning. NeurIPS, 2024. 4" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 602, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 553, + 634 + ], + "type": "text", + "content": "[26] Yu Huang, Yingbin Liang, and Longbo Huang. Provable generalization of overparameterized meta-learning trained with sgd. NeurIPS, 2022. 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 636, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 678 + ], + "type": "text", + "content": "[27] Ahmed Imteaj, Urmish Thakker, Shiqiang Wang, Jian Li, and M Hadi Amini. A survey on federated learning for resource-constrained IoT devices. IEEE Internet of Things Journal, 2021. 1" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "type": "text", + "content": "[28] Divyansh J Hunjunwala, Shiqiang Wang, and Gauri Joshi. Fedexp: Speeding up federated averaging via extrapolation. arXiv:2301.09604, 2023. 7, 1, 11, 12, 13, 14" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2169" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[29] Yihan Jiang, Jakub Konečný, Keith Rush, and Sreeram Kannan. Improving federated learning personalization via model agnostic meta learning. arXiv:1909.12488, 2019. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "text", + "content": "[30] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank Reddi, Sebastian Stich, and Ananda Theertha Suresh. Scaffold: Stochastic controlled averaging for federated learning. In ICML, 2020. 7, 1, 11, 12, 13, 14" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "text", + "content": "[31] Chiheon Kim, Doyup Lee, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Generalizable implicit neural representations via instance pattern composers. In CVPR, 2023. 4, 8, 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 239 + ], + "type": "text", + "content": "[32] Geeho Kim, Jinkyu Kim, and Bohyung Han. Communication-efficient federated learning with accelerated client gradient. In CVPR, 2024. 7, 1, 11, 12, 13, 14" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 274 + ], + "type": "text", + "content": "[33] Subin Kim, Sihyun Yu, Jaeho Lee, and Jinwoo Shin. Scalable neural video representations with learnable positional features. NeurIPS, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 275, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 294, + 308 + ], + "type": "text", + "content": "[34] Doyup Lee, Chiheon Kim, Minsu Cho, and WOOK SHIN HAN. Locality-aware generalizable implicit neural representation. NeurIPS, 2024. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 309, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 342 + ], + "type": "text", + "content": "[35] Jeffrey Li, Mikhail Khodak, Sebastian Caldas, and Ameet Talwalkar. Differentially private meta-learning. In ICLR, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 343, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 294, + 387 + ], + "type": "text", + "content": "[36] Qinbin Li, Yiqun Diao, Quan Chen, and Bingsheng He. Federated learning on non-iid data silos: An experimental study. In IEEE International Conference on Data Engineering, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 388, + 294, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 421 + ], + "type": "text", + "content": "[37] Shibo Li, Zheng Wang, Akil Narayan, Robert Kirby, and Shandian Zhe. Meta-learning with adjoint methods. In AISTATS, pages 7239-7251. PMLR, 2023. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 422, + 294, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 454 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 454 + ], + "type": "text", + "content": "[38] Xiang Li, Kaixuan Huang, Wenhao Yang, Shusen Wang, and Zhihua Zhang. On the convergence of fedavg on non-iid data. arXiv:1907.02189, 2019. 7, 1, 11, 12, 13, 14" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 456, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 487 + ], + "type": "text", + "content": "[39] Zhenguo Li, Fengwei Zhou, Fei Chen, and Hang Li. Meta-sgd: Learning to learn quickly for few-shot learning. arXiv:1707.09835, 2017. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 489, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 294, + 521 + ], + "type": "text", + "content": "[40] Susan Liang, Chao Huang, Yapeng Tian, Anurag Kumar, and Chenliang Xu. Av-nerf: Learning neural fields for real-world audio-visual scene synthesis. NeurIPS, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 523, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 567 + ], + "type": "text", + "content": "[41] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In AISTATS, 2017. 1, 3, 5, 7, 8, 6, 10, 11, 12, 13, 14" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 568, + 294, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 611 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 611 + ], + "type": "text", + "content": "[42] William McNally, Kanav Vats, Tyler Pinto, Chris Dulhanty, John McPhee, and Alexander Wong. Golfdb: A video database for golf swing sequencing. In CVPRW, 2019. 5, 6, 3, 4, 7, 12" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 613, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 656 + ], + "type": "text", + "content": "[43] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 2021. 2, 5, 3, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "type": "text", + "content": "[44] Alex Nichol, Joshua Achiam, and John Schulman. On first-order meta-learning algorithms. arXiv:1803.02999, 2018. 1, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[45] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learn" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 333, + 72, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 553, + 95 + ], + "type": "text", + "content": "ing implicit 3d representations without 3d supervision. In CVPR, 2020. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "text", + "content": "[46] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, 2019. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "text", + "content": "[47] Vishwanath Saragadam, Jasper Tan, Guha Balakrishnan, Richard G Baraniuk, and Ashok Veeraraghavan. Miner: Multiscale implicit neural representation. In ECCV, 2022. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 186, + 553, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 218 + ], + "type": "text", + "content": "[48] Bipasha Sen, Aditya Agarwal, Vinay P Namboodiri, and CV Jawahar. Inr-v: A continuous representation space for videobased generative tasks. arXiv:2210.16579, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 220, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 252 + ], + "type": "text", + "content": "[49] Risa Shinoda and Kaede Shiohara. Petface: A large-scale dataset and benchmark for animal identification. In ECCV, 2025. 5, 6, 3, 4, 7, 11" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "type": "text", + "content": "[50] Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. Membership inference attacks against machine learning models. In IEEE Symposium on Security and Privacy, pages 3-18. IEEE, 2017. 7, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 299, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 299, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 299, + 553, + 342 + ], + "type": "text", + "content": "[51] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. NeurIPS, 2020. 2, 5, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 343, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 376 + ], + "type": "text", + "content": "[52] Ivan Skorokhodov, Savva Ignatyev, and Mohamed Elhoseiny. Adversarial generation of continuous images. In CVPR, 2021. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 377, + 553, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 553, + 399 + ], + "type": "text", + "content": "[53] Kun Su, Mingfei Chen, and Eli Shlizerman. Inras: Implicit neural representation for audio scenes. NeurIPS, 2022. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 401, + 553, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 401, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 316, + 401, + 553, + 432 + ], + "type": "text", + "content": "[54] Terpeei Suzuki. Federated learning for large-scale scene modeling with neural radiance fields. arXiv:2309.06030, 2023.2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 434, + 553, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 456 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 456 + ], + "type": "text", + "content": "[55] Terpei Suzuki. Fed3dgs: Scalable 3d gaussian splatting with federated learning. arXiv:2403.11460, 2024. 1, 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 457, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 457, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 457, + 553, + 499 + ], + "type": "text", + "content": "[56] Jihoon Tack, Subin Kim, Sihyun Yu, Jaeho Lee, Jinwoo Shin, and Jonathan Richard Schwarz. Learning large-scale neural fields via context pruned meta-learning. NeurIPS, 2024. 2, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 502, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 502, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 502, + 553, + 555 + ], + "type": "text", + "content": "[57] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. NeurIPS, 2020. 1, 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 557, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 553, + 601 + ], + "type": "text", + "content": "[58] Matthew Tancik, Ben Mildenhall, Terrance Wang, Divi Schmidt, Pratul P Srinivasan, Jonathan T Barron, and Ren Ng. Learned initializations for optimizing coordinate-based neural representations. In CVPR, 2021. 5, 1, 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 602, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 553, + 656 + ], + "type": "text", + "content": "[59] Zaid Tasneem, Akshit Dave, Abhishek Singh, Kushagra Tiwary, Praneeth Vepakomma, Ashok Veeraraghavan, and Ramesh Raskar. Decentnerfs: Decentralized neural radiance fields from crowdsourced images. arXiv:2403.13199, 2024. 1" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "text", + "content": "[60] Ayush Tewari, Ohad Fried, Justus Thies, Vincent Sitzmann, Stephen Lombardi, Kalyan Sunkavalli, Ricardo MartinBrualla, Tomas Simon, Jason Saragih, Matthias Nießner, et al. State of the art on neural rendering. In Computer Graphics Forum, 2020. 1" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2170" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 521 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[61] Jianyu Wang, Qinghua Liu, Hao Liang, Gauri Joshi, and H Vincent Poor. Tackling the objective inconsistency problem in heterogeneous federated optimization. NeurIPS, 2020. 7, 1, 11, 12, 13, 14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "text", + "content": "[62] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 2004. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "text", + "content": "[63] Kang Wei, Jun Li, Ming Ding, Chuan Ma, Howard H. Yang, Farhad Farokhi, Shi Jin, Tony Q. S. Quek, and H. Vincent Poor. Federated learning with differential privacy: Algorithms and performance analysis. IEEE Transactions on Information Forensics and Security, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "type": "text", + "content": "[64] Yuexiang Xie, Zhen Wang, Dawei Gao, Daoyuan Chen, Liuyi Yao, Weirui Kuang, Yaliang Li, Bolin Ding, and Jingren Zhou. Federatedscope: A flexible federated learning platform for heterogeneity. arXiv:2204.05011, 2022. 4, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 308 + ], + "type": "text", + "content": "[65] Haotian Yang, Hao Zhu, Yanru Wang, Mingkai Huang, Qiu Shen, Ruigang Yang, and Xun Cao. Facescape: a large-scale high quality 3d face dataset and detailed riggable 3d face prediction. In CVPR, 2020. 5, 6, 3, 7, 14" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 309, + 294, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 351 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 351 + ], + "type": "text", + "content": "[66] Kai Yue, Richeng Jin, Chau-Wai Wong, Dror Baron, and Huaiyu Dai. Gradient obfuscation gives a false sense of security in federated learning. In USENIX Security Symposium, 2023. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 353, + 294, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 353, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 353, + 294, + 386 + ], + "type": "text", + "content": "[67] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 294, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 418 + ], + "type": "text", + "content": "[68] Yintian Zhang and Ziyu Shao. Federated neural radiance field for distributed intelligence. arXiv:2406.10474, 2024. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 421, + 294, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 421, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 421, + 294, + 443 + ], + "type": "text", + "content": "[69] Xinyu Zhou and Raef Bassily. Task-level differentially private meta learning. NeurIPS, 2022. 7, 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 445, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 498 + ], + "type": "text", + "content": "[70] Hao Zhu, Haotian Yang, Longwei Guo, Yidi Zhang, Yanru Wang, Mingkai Huang, Menghua Wu, Qiu Shen, Ruigang Yang, and Xun Cao. Facescape: 3d facial dataset and benchmark for single-view 3d face reconstruction. IEEE TPAMI, 2023. 5, 6, 3, 7, 14" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 500, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 294, + 521 + ], + "type": "text", + "content": "[71] Yiyu Zhuang, Hao Zhu, Xusen Sun, and Xun Cao. Mofanerf: Morphable facial neural radiance field. In ECCV, 2022. 1" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2171" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/FedPall_ Prototype-based Adversarial and Collaborative Learning for Federated Learning with Feature Drift/f6c6ce84-c8dc-4de0-bad0-3f6c4ed86bd8_content_list.json b/2025/FedPall_ Prototype-based Adversarial and Collaborative Learning for Federated Learning with Feature Drift/f6c6ce84-c8dc-4de0-bad0-3f6c4ed86bd8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..af4155589c6b13b31490d7fe224b26a799e539ae --- /dev/null +++ b/2025/FedPall_ Prototype-based Adversarial and Collaborative Learning for Federated Learning with Feature Drift/f6c6ce84-c8dc-4de0-bad0-3f6c4ed86bd8_content_list.json @@ -0,0 +1,1688 @@ +[ + { + "type": "text", + "text": "FedPall: Prototype-based Adversarial and Collaborative Learning for Federated Learning with Feature Drift", + "text_level": 1, + "bbox": [ + 94, + 128, + 903, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yong Zhang $^{1,3}$ , Feng Liang $^{1*}$ , Guanghu Yuan $^{1}$ , Min Yang $^{2}$ , Chengming Li $^{1}$ and Xiping Hu $^{1,3*}$ \nArtificial Intelligence Research Institute, Shenzhen MSU-BIT University $^{1}$ \nShenzhen Institutes of Advanced Technology, Chinese Academy of Sciences $^{2}$ \nSchool of Medical Technology, Beijing Institute of Technology $^{3}$", + "bbox": [ + 130, + 202, + 870, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhangyong2023@bit.edu.cn, fliang@smbu.edu.cn, yangh@mail.ustc.edu.cn min.yang@siat.ac.cn, licm@smbu.edu.cn, huxp@smbu.edu.cn}", + "bbox": [ + 187, + 277, + 807, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 344, + 326, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Federated learning (FL) enables collaborative training of a global model in the centralized server with data from multiple parties while preserving privacy. However, data heterogeneity can significantly degrade the performance of the global model when each party uses datasets from different sources to train a local model, thereby affecting personalized local models. Among various cases of data heterogeneity, feature drift, feature space difference among parties, is prevalent in real-life data but remains largely unexplored. Feature drift can distract feature extraction learning in clients and thus lead to poor feature extraction and classification performance. To tackle the problem of feature drift in FL, we propose FedPall, an FL framework that utilizes prototype-based adversarial learning to unify feature spaces and collaborative learning to reinforce class information within the features. Moreover, FedPall leverages mixed features generated from global prototypes and local features to enhance the global classifier with classification-relevant information from a global perspective. Evaluation results on three representative feature-drifted datasets demonstrate FedPall's consistently superior performance in classification with feature-drifted data in the FL scenario.", + "bbox": [ + 88, + 375, + 483, + 708 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 752, + 220, + 767 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Today, in computer vision, researchers often utilize large amounts of data from various parties to improve the accuracy of algorithms. However, this raises concerns, such as the potential for user privacy leakage caused by sharing private data [12]. Federated learning (FL) [22] is proposed as a", + "bbox": [ + 89, + 777, + 483, + 854 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5f28ad42f31865a0a78a661ecc226f3937acd0999dda84a27f498848b5e69682.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 345, + 823, + 372 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4095f66beff724dc93a136d99e045de0ef20920ffa4e2a6024f239976a493e06.jpg", + "image_caption": [ + "(a) Feature drift" + ], + "image_footnote": [], + "bbox": [ + 526, + 377, + 661, + 454 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a568814547d2163997a8a18e9df0123816b80d05ec44cfd4b154313ea88cec32.jpg", + "image_caption": [ + "(b) Traditional classification" + ], + "image_footnote": [], + "bbox": [ + 756, + 383, + 892, + 455 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/12024965fe707cc42954bd98dd10ed6f64b8a06dfa6494cc03f3f7b37108bb0a.jpg", + "image_caption": [ + "(c) De-clientization" + ], + "image_footnote": [], + "bbox": [ + 526, + 498, + 661, + 554 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f3d3cb11ceb294bc25984d978289c61d11e553e6549ac6b8968239d7dbcccfea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 756, + 477, + 893, + 515 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ecfb34b8fa4ce36d7ac8a2e09bb1554982f0b7e59df31c5a84c58553b97c8d20.jpg", + "image_caption": [ + "(d) Prototype contrastive", + "Figure 1. We show a schematic diagram of feature drift and use different techniques to drive feature distribution to update in different directions." + ], + "image_footnote": [], + "bbox": [ + 756, + 523, + 890, + 554 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "privacy-preserving distributed learning paradigm to address these challenges. In the FL paradigm, each party maintains a local client model and collaborates with others to train a global model on the server without sharing the original data, effectively protecting user privacy.", + "bbox": [ + 509, + 657, + 906, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Particularly, the FL paradigm faces challenges from data heterogeneity [14]. Due to the limited local view of each client, data distribution discrepancies arise across clients (commonly referred to as the non-independent and identically distributed data issue, or non-independent and identically distributed(non-IID) data issue), which can increase generalization error for local models and degrade the performance of the globally aggregated model [10]. Although recent work on non-IID data in FL primarily addresses issues such as stability, client drift, and heterogeneous label distributions among clients [11, 18, 31], feature drift(i.e.,", + "bbox": [ + 509, + 734, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors.", + "bbox": [ + 107, + 862, + 241, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "The code is available at https://github.com/DistriAI/FedPall.", + "bbox": [ + 91, + 875, + 480, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3111", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "variations in feature distributions across clients) is a prevalent yet underexplored challenge in FL. As shown in Fig. 1a, feature drift refers to the phenomenon where samples of the same class exhibit different feature distributions across different clients due to variations in data collection methods, devices, and other factors. This leads to ambiguous decision boundaries, severely impacting the classification performance of federated learning models. However, traditional classification losses (e.g., cross-entropy loss (CE loss), as shown in Fig. 1b) do not account for feature drift. As a result, the collaborative effect among different clients causes the feature space of the same-class samples to influence each other, leading to only slight or even no clearer distinction between decision boundaries for various classes within the same client. As illustrated in Fig. 1c, decentralization is a common approach to addressing feature drift, which involves ignoring inter-client differences to make the feature space of same-class samples across different clients more clustered. In feature drift scenarios, some state-of-the-art algorithms do not explore the role of loss functions but instead optimize local update algorithms to achieve decentralization. For example, FedBN [19] compresses the feature space across clients by adding batch normalization layers to local models. Some methods [1, 34] align feature spaces by sharing partial data for synthetic data augmentation. Additionally, ADCOL sends raw features to the server to update a amplifier and uses the Kullback-Leibler (KL) loss function to achieve decentralization. However, these methods have drawbacks: on the one hand, they may lead to loss of class information, and on the other hand, they pose potential risks of privacy leakage.", + "bbox": [ + 93, + 90, + 480, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle the above challenge, we propose FedPall, a novel prototype-based adversarial and collaborative learning framework for FL with feature drift. FedPall applies adversarial learning between clients and the server to unify feature spaces, as well as collaborative learning across clients to enhance global decision boundaries. Specifically, FedPall uses adversarial learning to train a feature enhancer and uses KL divergence to align heterogeneous feature spaces between clients, while using prototype contrastive loss to reinforce class information (see Fig. 1d). Finally, adversially aligned features are securely mixed with the global prototypes and uploaded to the server, where a global-view classifier is trained to enhance overall performance. Our contributions are summarized as follows:", + "bbox": [ + 93, + 561, + 480, + 771 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel FL framework, FedPall, to address the feature drift problem. FedPall introduces adversarial learning between clients and the server, and collaborative learning among clients aiming to project feature representations into a unified feature space and reinforce the intrinsic class information. This approach effectively mitigates the feature drift problem in FL settings.", + "- We develop a technical strategy that hierarchically in" + ], + "bbox": [ + 89, + 780, + 480, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tegrates the global prototypes with local features to orchestrate client-server collaboration. The mixed prototype features are then used to train a global classifier, which induces the classifier to distill discriminative patterns through cross-client knowledge consolidation.", + "bbox": [ + 526, + 90, + 903, + 166 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Empirical evaluation on three typical feature-drifted benchmarks demonstrates that our proposed method achieves state-of-the-art classification accuracy.", + "bbox": [ + 513, + 167, + 903, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 224, + 653, + 239 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In federated learning (FL), each client's limited local view exacerbates the feature drift problem. Variations in data distributions can cause the same class label to have different feature representations, leading to poor generalization of local models. Existing solutions typically focus on two main approaches: feature space alignment and contrastive prototype learning.", + "bbox": [ + 511, + 251, + 903, + 356 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Some studies address feature drift through feature representation alignment. FRAug [1] employs data augmentation to generate synthetic embeddings that capture both global and client-specific information. FedSea [26] aligns feature distributions to transform raw features into an IID format, mitigating feature drift. FedCiR [20] maximizes mutual information between representations and labels, while minimizing mutual information between client-specific representations conditioned on labels. FedHEAL [2] builds a fair aggregation target on the server side, indirectly aligning feature spaces. MOON [16] encourages local models to align with the global feature distribution through constrained updates. ADCOL [17], unlike traditional aggregation frameworks, uses adversarial learning to enforce a unified representation across clients, alleviating feature drift. However, ADCOL's weak collaboration does not address global class boundaries, and its adversarial mechanism introduces potential privacy risks. Our approach enhances global class boundaries through stronger collaborative learning, while preserving privacy via adversarial learning to mitigate inter-client feature drift.", + "bbox": [ + 511, + 357, + 903, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Some studies have explored prototype-driven federated learning (FL) paradigms, where prototypes compact feature embeddings, reducing communication bandwidth and preserving privacy [30]. Tan et al.[27, 28] proposed a supervised contrastive loss function using both global and local prototypes to minimize the distance between feature representations and class prototypes, addressing data heterogeneity. However, using the average feature as a class prototype may overlook intra-class variability. To address this, MP-FedCL[25] employs clustering on the client side to generate multiple prototypes per class, capturing intra-class variation and mitigating feature drift. Building on MP-FedCL, FedPLVM [29] improves local training with a two-stage clustering process between clients and the server, and incorporates an $\\alpha$ -sparsity prototype loss to further op", + "bbox": [ + 511, + 674, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3112", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "timize performance. By leveraging prototypes' privacy-preserving nature, this method addresses privacy concerns while reinforcing class-specific information in feature representations. In the FedPall framework, we enhance client-server collaboration through global prototypes. The server provides global category information to the client, helping bring similar data closer and separate dissimilar data. Additionally, we use mixed features with global category information to strengthen the global classifier.", + "bbox": [ + 89, + 90, + 483, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 239, + 181, + 255 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Description", + "text_level": 1, + "bbox": [ + 89, + 263, + 285, + 280 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We define feature drift as follows: Given a dataset $\\mathbb{D}$ with features $x$ and labels $y$ , while the conditional distribution $P_{i}(x|y)$ differs across clients, the marginal distribution $P(y)$ remains the same. This means that the same label may have significantly different features across clients. For example, due to variations in environment, geographic location, and cultural differences, the structural features of houses can vary widely.", + "bbox": [ + 89, + 286, + 483, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With feature drift in FL, our goal is to optimize each client's personalized model loss while leveraging the potential performance gains from collaborative learning across clients [33]. In the FedPall framework, there are a total of $N$ clients, each client $n$ has a private dataset $D_{n}$ . Based on this goal, we formulate the overall optimization objective of the FedPall framework as follows:", + "bbox": [ + 89, + 407, + 483, + 512 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta_ {1}, \\theta_ {2}, \\dots , \\theta_ {N} \\in \\mathbb {R} ^ {d _ {1}}} F (\\theta) := \\frac {1}{N} \\sum_ {n = 1} ^ {N} f _ {n} \\left(\\theta_ {n}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 518, + 482, + 560 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $f_{n}$ represents the expected loss obtained from client $n$ using the global model parameters under the dataset $D_{n}$ , and $\\theta_{i}$ represents the local model parameters of client $i$ .", + "bbox": [ + 89, + 568, + 483, + 613 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. FedPall Framework", + "text_level": 1, + "bbox": [ + 89, + 622, + 279, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing approaches to addressing the feature drift problem in FL typically focus on either collaborative learning or adversarial learning in isolation. This can result in models that either fail to adequately capture class-related information in the feature representations or exhibit persistent discrepancies in feature distributions across clients. To address these limitations, we propose integrating both adversarial and collaborative learning to effectively mitigate feature drift in FL settings. In this section, we present the FedPall framework by elaborating on its adversarial and collaborative learning. The framework of the overall approach is shown in Fig. 2a. It is structured into four key procedures: generating global prototypes, training local models, training global model, and decentralizing global classifier.", + "bbox": [ + 89, + 643, + 482, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We define certain model symbols here that will be used in this section. The local model $F(\\cdot)$ consists of two components, a feature extractor $G(\\cdot)$ (e.g., Resnet50 [7] for image", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "data) and a classifier $H(\\cdot)$ . We use a multilayer perceptron (MLP) as our Amplifier, and except for the output layer, the number of nodes in other layers is consistent with that of the classifier.", + "bbox": [ + 511, + 90, + 905, + 150 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.1. Generating Global Prototypes", + "text_level": 1, + "bbox": [ + 511, + 157, + 769, + 172 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Several studies [23, 27] suggest that category-centered prototypes are a privacy-friendly form of global knowledge. We leverage collaboration between clients to aggregate and generate global prototypes, as shown in Fig. 2b. Typically, each class prototype is defined as the mean of the feature vectors belonging to that class. The distinction between local and global prototypes lies in the scope of their underlying data. This section will describe the construction of both local and global prototypes.", + "bbox": [ + 511, + 176, + 905, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The local prototype for the $k$ -th category on client $n$ is defined as:", + "bbox": [ + 511, + 311, + 905, + 340 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nc _ {n} ^ {k} = \\frac {1}{N _ {n} ^ {k}} \\sum_ {(x, y) \\in D _ {n} ^ {k}} G _ {n} (x), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 352, + 906, + 391 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $D_{n}^{k}$ and $N_{n}^{k}$ represent the data samples and the number of samples for the $k$ -th category on client $n$ , respectively.", + "bbox": [ + 511, + 396, + 905, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Gathering all the local prototypes together forms a local prototype set, which can be defined as:", + "bbox": [ + 511, + 441, + 905, + 472 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {O} _ {n} = \\left\\{c _ {n} ^ {1}, c _ {n} ^ {2}, \\dots , c _ {n} ^ {K} \\right\\} \\in \\mathbb {R} ^ {K \\times d}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 479, + 903, + 498 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $K$ represents the number of categories owned by each client, and $d$ denotes the output feature dimension.", + "bbox": [ + 511, + 505, + 905, + 534 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since we are only addressing the problem of feature drift, all clients have the same number of categories. Upon receiving the local prototype set and local label proportions $\\mathcal{N} = \\{\\{N_n^k\\}_{k=1}^K | n \\in \\mathcal{A}\\}$ sent by client set $\\mathcal{A}$ , the server integrates the local prototypes from all clients to form the global prototypes:", + "bbox": [ + 511, + 535, + 905, + 626 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} ^ {k} = \\sum_ {n} \\frac {N _ {n} ^ {k}}{\\sum_ {n} N _ {n} ^ {k}} c _ {n} ^ {k}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 633, + 903, + 670 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The global prototypes set can be represented as", + "bbox": [ + 532, + 676, + 846, + 691 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} = \\left[ \\mathcal {G} ^ {1}, \\dots , \\mathcal {G} ^ {k}, \\dots , \\mathcal {G} ^ {K} \\right]. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 699, + 903, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Next, the server sends the global prototype set $\\mathcal{G}$ to each client to guide local model training.", + "bbox": [ + 511, + 724, + 905, + 755 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.2. Training Local Models", + "text_level": 1, + "bbox": [ + 511, + 761, + 718, + 775 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The goal of training local models is to train effective feature encoders that map raw data from different clients into a unified feature space, where feature distributions are aligned and class-related information is enhanced.", + "bbox": [ + 511, + 779, + 905, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As mentioned earlier, due to feature drift, training a local classifier alone is not sufficient for accurately classifying data with feature drift. To address this, we apply adversarial learning to train a feature encoder, as illustrated", + "bbox": [ + 511, + 840, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3113", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1b0fc8c559823d3f6863fb3f196c03b5f51fa498a103b6fdced56d8693c9af2c.jpg", + "image_caption": [ + "(a) The overview of FedPall" + ], + "image_footnote": [], + "bbox": [ + 96, + 87, + 903, + 231 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/050c4c4db9b7ad9a07ca628e7e6558ca4d87a136f07607029f77cc39ad050fbe.jpg", + "image_caption": [ + "Figure 2. The FedPall framework and detailed phases" + ], + "image_footnote": [], + "bbox": [ + 98, + 250, + 906, + 420 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in Fig. 2c. Specifically, we use a global amplifier, trained on the server, which amplifies the heterogeneous information in the features from different clients. At the client's side, we apply the amplifier and use Kullback-Leibler (KL) divergence to reduce the heterogeneous information in the features, thereby creating an adversarial learning setup between the client and server. The objective is to improve the generalization ability of the feature encoders across clients while minimizing the client-specific heterogeneity in the feature representations. Let $x^{(i)}$ denote the $i$ -th dimension of vector $x$ , the KL divergence is calculated as:", + "bbox": [ + 88, + 468, + 482, + 633 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {K L} = \\sum_ {(\\mathbf {x}, \\mathbf {y}) \\in D _ {n}} D _ {K L} \\left(A _ {n} (G (\\mathbf {x})) | | [ \\frac {1}{N} ] ^ {N}\\right) \\\\ = \\sum_ {(\\mathbf {x}, y) \\in D _ {n}} \\sum_ {i} ^ {N} A _ {n} (G (\\mathbf {x})) ^ {(i)} \\log N A _ {n} (G (\\mathbf {x})) ^ {(i)}, \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 647, + 480, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $x^{(i)}$ denotes the $i$ -th dimension of vector $x$ , $A_{n}(\\cdot)$ represents the amplifier of the $n$ -th client, $D_{KL}(P||Q)$ represents the KL divergence of $P$ and $Q$ .", + "bbox": [ + 89, + 747, + 482, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After adversarial learning, although the feature representations of different clients are mapped to similar feature distributions, the class-related information may be blurred. To address this problem, we propose using contrastive learning to reinforce the class-related information within the feature encoder. To prevent information leakage, we enable collaboration between the global prototypes (server-side) and", + "bbox": [ + 89, + 795, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "local features (client-side). Specifically, we employ the InfoNCE loss to minimize the distance between local features and their corresponding global prototypes, while maximizing the distance between local features and global prototypes of other classes. This approach strengthens the class-related representation within the feature encoder. The formula for the global-prototype contrastive loss is as follows:", + "bbox": [ + 511, + 468, + 906, + 574 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {i n f o N C E}} = \\sum_ {(\\mathrm {x}, \\mathrm {y}) \\in D _ {n}} - \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(G (\\mathrm {x}) , \\mathcal {G} ^ {\\mathrm {y}}\\right) / \\tau\\right)}{\\sum_ {\\mathrm {y} _ {\\alpha} \\in A (\\mathrm {y})} \\exp \\left(\\operatorname {s i m} \\left(\\mathrm {x} , \\mathcal {G} ^ {\\left(\\mathrm {y} _ {\\alpha}\\right)}\\right) / \\tau\\right)}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 601, + 903, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $A(y) \\coloneqq \\{y_{\\alpha} \\in [1, |\\mathcal{G}|] : y_{\\alpha} \\neq y\\}$ is the set of labels distinct from $y$ , $\\tau$ is the temperature to adjust the tolerance for feature difference, and $sim(x, y)$ represents the cosine similarity of $x$ and $y$ .", + "bbox": [ + 511, + 648, + 906, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We combine adversarial learning and collaborative learning to address the feature drift problem in FL. By leveraging the two loss functions defined above, along with the local cross-entropy loss, we progressively train the local feature encoder at each client. The overall loss function for each client is as follows:", + "bbox": [ + 511, + 712, + 905, + 801 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {C E (\\mathrm {x}, \\mathrm {y}) \\sim D _ {n}} (F (\\mathrm {x}), \\mathrm {y}) + \\mu \\mathcal {L} _ {K L} + \\delta \\mathcal {L} _ {\\text {i n f o N C E}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 820, + 903, + 837 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mu$ denotes the weight of the $\\mathcal{L}_{KL}$ divergence and $\\mathcal{L}_{CE}$ denotes the cross-entropy loss.", + "bbox": [ + 511, + 852, + 903, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, the local model $F(\\cdot)$ is updated using Eq. (8).", + "bbox": [ + 531, + 885, + 888, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3114", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.3. Training Global Model", + "text_level": 1, + "bbox": [ + 89, + 90, + 297, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Due to the client's limited local view, training an accurate classifier is challenging. To address this issue, encrypted mixed features are uploaded to the server, enabling classifier training with a global perspective. Additionally, the server leverages its global view to train an amplifier, which engages in adversarial learning with the client to enhance model robustness. Fig. 2d illustrates this procedure, and the detailed process is described below.", + "bbox": [ + 89, + 109, + 483, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For each feature $z_{n}^{k}$ of class $k$ on client $n$ , we obtain a prototype mixed feature by performing a weighted fusion with the corresponding global prototype:", + "bbox": [ + 89, + 229, + 483, + 276 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr _ {n} ^ {k} = \\alpha \\times z _ {n} ^ {k} + (1 - \\alpha) \\times \\mathcal {G} ^ {k}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 286, + 483, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $z_{n}^{k}\\in Z_{n}^{k} = \\{z_{n}^{i,k}\\}_{i = 1}^{S}$ . $\\alpha \\sim U(u_f,u_r)$ are the mix parameters, with $u_{f}$ and $u_{d}$ representing two hyperparameters of a uniform distribution. $S$ represents the number of samples of the $k$ -th category in client $n$ .", + "bbox": [ + 89, + 314, + 483, + 376 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To further reduce the risk of privacy leakage, we employ a Bernoulli mask, represented as:", + "bbox": [ + 89, + 376, + 483, + 407 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} M a s k = \\left\\{X _ {1}, X _ {2}, \\dots , X _ {d} \\right\\}, \\tag {10} \\\\ X _ {i} \\sim \\operatorname {B e r n o u l l i} (\\beta), \\quad \\forall i \\in [ 1, d ]. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 417, + 480, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final output of the prototype mixing mechanism is derived by selecting the mixed feature elements based on the noise mask:", + "bbox": [ + 89, + 464, + 483, + 507 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {r} _ {n} ^ {k} = \\operatorname {M a s k} \\odot r _ {n} ^ {k}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 222, + 507, + 480, + 526 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\odot$ is an element-wise and operator.", + "bbox": [ + 89, + 532, + 370, + 546 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After generating the prototype mixed features, the client will form the set $\\mathcal{D}_{R_L}(R,Y)$ using the prototype mixed feature set $R_{n} = \\{\\tilde{r}_{n}^{1},\\dots,\\tilde{r}_{n}^{k},\\dots,\\tilde{r}_{n}^{K}\\} \\in \\mathbb{R}^{K\\times d}$ and the corresponding label set $Y$ , which will be sent to the server.", + "bbox": [ + 89, + 547, + 483, + 607 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The server updates the global amplifier $A$ using the mixed feature sets from each client along with the corresponding client IDs. Specifically, we first construct the dataset for training the amplifier, denoted as $\\mathcal{D}_{R_I}(R,I)$ where $I$ represents the client IDs. The amplifier is then updated by minimizing the empirical risk:", + "bbox": [ + 89, + 607, + 483, + 698 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {(R, I) \\sim \\mathcal {D} _ {R _ {I}}} \\ell_ {C E} (R, I). \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 710, + 480, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At the same time, the server updates the global classifier $C_g$ using the mixed prototype features and the class labels from the clients $Y$ . This is done by minimizing the empirical risk:", + "bbox": [ + 89, + 738, + 483, + 797 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {(R, Y) \\sim \\mathcal {D} _ {R _ {L}}} \\ell_ {C E} (R, Y). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 797, + 480, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.4. Decentralizing Global Classifier", + "text_level": 1, + "bbox": [ + 89, + 821, + 359, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we deploy the global classifier $C_g$ to each client to replace the original local classifiers $C_c$ , as shown in Fig. 2e. The purpose of this is to obtain a more generalized classifier that can alleviate the feature drift problem. To allow", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the global classifier to adapt to the personalized characteristics of local data, we retrain it on the client's local data, thereby enhancing the classifier's accuracy and improving its performance on individual client data distributions.", + "bbox": [ + 511, + 90, + 905, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Discussion", + "text_level": 1, + "bbox": [ + 511, + 162, + 627, + 178 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Computational Cost Compared to the standard federated learning model, our adversarial collaborative learning approach introduces an amplifier and a global classifier. However, the number of parameters for these two components is much smaller than those of the other components. In our design, both the amplifier and the classifier are designed as three-layer MLPs. Compared to the feature extractor (with a total of 23.5M parameters), the classifier (with a total of 1.32M parameters) and the amplifier (with a total of 1.33M parameters) account for approximately $5.61\\%$ and $5.59\\%$ , respectively. Moreover, the client-side amplifier remains frozen, functioning exclusively in the forward pass for loss calculation while being disabled during backpropagation cycles.", + "bbox": [ + 511, + 188, + 906, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Communication Efficiency Unlike traditional federated learning methods such as FedAvg, which transmit full local model parameters, our approach uses prototype-mixed features as the communication medium. Assuming consistent tensor representations between model parameters and prototype features, each client employs a ResNet-50-based feature extractor that generates 2048-dimensional embeddings. During upload, clients transmit an average of 12,122 prototype-mixed features, compared to approximately 24.8 million model parameters in FedAvg. For downloading, only the amplifier and global classifier parameters are transmitted, reducing communication cost to about $10.6\\%$ of that in FedAvg (see previous paragraph).", + "bbox": [ + 511, + 417, + 906, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Limitation While our framework currently specializes in image recognition tasks, its extension to NLP or time-series analysis remains unexplored. Successful cross-domain adaptation requires two key developments: (1) establishing domain-specific feature representations and prototype definitions, and (2) redesigning loss functions according to task semantics. For NLP applications, this implies reconfiguring the standard classification paradigm into autoregressive prediction frameworks. Architectural adaptations are equally crucial - particularly the incorporation of RNN-based structures with inherent temporal modeling capabilities for sequential data processing.", + "bbox": [ + 511, + 631, + 905, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 823, + 645, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 848, + 702, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets We conduct experiments on three publicly available feature drift datasets: Digits [32], Office-10 [6], and", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3115", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/79da3f786bfa25206e0967e2bc272ba425c40a6eaf69aa58c0f032b4acecbe52.jpg", + "table_caption": [ + "Table 1. The top-1 accuracy (\\%) of each algorithm on each sub-dataset of the Office-10, Digits, and PACS datasets is compared, along with the average top-1 accuracy across all sub-datasets. The mean and standard deviation (std) from three random trials (using different random seeds, with other experimental settings remaining the same) are reported. The highest accuracy for each dataset is highlighted in bold, and the second-highest accuracy is underlined." + ], + "table_footnote": [], + "table_body": "
SingleSetFedAvgFedProxPerfedAvgFedRepFedBNMOONFedProtoADCOLRUCRFedHEALours(FedPall)
Office-10amazon74.0(2.7)56.9(2.5)56.6(2.6)57.1(2.2)45.3(1.9)40.8(15.8)51.7(16.1)69.4(2.1)73.3(4.4)52.1(8.5)65.1(3.3)72.9(1.4)
caltech44.7(3.2)46.5(4.6)51.0(5.2)50.8(1.6)38.4(4.9)33.9(6.5)41.3(13.6)39.4(6.3)37.2(1.7)44.3(1.0)44.6(3.2)44.7(8.7)
dslr60.2(6.7)30.1(4.9)33.3(10.4)31.2(4.9)34.4(4.9)38.7(3.2)24.7(1.9)65.6(4.9)76.3(4.9)30.1(6.7)67.7(1.9)77.4(3.2)
webcam71.3(2.6)37.9(6.2)43.7(7.2)47.1(7.8)55.8(2.6)30.5(6.1)33.3(12.7)71.3(4.3)71.3(2.6)37.4(5.0)60.9(1.0)74.7(1.0)
avg62.5(0.4)42.9(1.2)46.1(2.6)46.6(2.9)43.5(1.3)36.0(6.5)37.8(10.9)61.4(1.7)64.5(1.8)41.0(0.6)59.6(0.5)67.5(2.7)
DigitsMNIST95.5(0.2)92.9(2.2)91.8(3.0)90.1(4.8)86.5(6.1)96.7(0.1)93.4(1.1)96.4(0.5)96.3(0.4)92.6(2.0)93.6(0.6)97.2(0.4)
SVHN71.1(0.9)77.4(0.2)76.9(0.3)75.6(0.4)67.2(1.7)79.4(0.3)79.6(0.8)72.5(0.3)75.1(2.1)77.9(0.3)68.3(2.0)78.0(0.4)
USPS86.4(0.3)89.3(0.9)89.2(1.4)88.7(0.7)90.0(3.0)90.1(0.5)81.8(0.7)87.0(0.8)86.7(1.3)88.9(2.4)87.0(0.5)87.3(1.3)
SynthDigits95.2(0.1)95.5(0.1)95.4(0.1)95.0(0.2)94.2(0.8)95.6(0.1)96.6(0.2)95.3(0.6)96.4(0.3)96.0(0.2)89.3(1.6)95.3(0.4)
MNIST-M76.6(0.4)73.8(1.5)74.0(1.5)73.2(0.8)69.1(0.9)76.3(0.4)72.2(0.9)78.3(1.2)78.3(4.4)72.7(0.4)67.8(1.9)85.9(1.4)
avg84.9(0.1)85.8(0.9)85.5(1.1)84.5(1.3)81.4(2.5)87.6(0.1)84.7(0.6)85.9(0.2)86.6(1.3)85.6(0.9)81.2(1.3)88.7(0.2)
PACSart-painting33.6(0.8)25.8(1.9)24.3(4.1)26.5(2.2)26.9(3.3)36.7(1.8)30.6(2.0)32.7(0.7)34.9(1.2)24.7(1.1)31.2(1.2)35.6(0.6)
cartoon58.5(2.5)45.4(2.3)51.4(0.6)48.3(1.2)44.4(2.1)55.6(2.0)51.5(1.8)57.3(1.5)57.2(0.8)47.5(3.3)50.8(0.4)59.7(2.3)
photo63.0(1.9)48.7(3.1)49.6(2.0)46.9(2.6)41.9(2.8)66.1(1.0)53.0(3.3)64.0(1.3)62.1(2.0)47.5(6.2)61.1(2.0)64.7(1.3)
sketch79.7(0.1)49.0(2.0)40.7(1.5)44.4(3.8)40.5(1.3)79.6(1.7)55.1(1.4)79.6(0.8)80.1(1.0)42.2(2.4)73.8(0.1)82.2(0.7)
avg58.7(1.2)42.2(1.6)41.5(1.8)41.5(1.9)38.4(1.1)59.5(1.4)47.6(0.9)58.4(0.3)58.6(0.6)40.5(2.0)54.2(0.5)60.6(0.4)
", + "bbox": [ + 98, + 155, + 903, + 354 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "PACS [15]. Specifically, (1) the Digits dataset consists of five different domain sources: MNIST [13], SVHN [24], USPS [9], SynthDigits [5], and MNIST-M [5]; (2) the Office-10 dataset includes four distinct sources: Amazon, Caltech, DSLR, and WebCam; (3) the PACS dataset consists of four sources: Art Painting, Cartoon, Photo, and Sketch. Datasets Office-10 and PACS are real-world images from natural scenes, which inherently exhibit feature drift due to the diversity of their sources. Digits is a digit recognition dataset. In line with [19, 28], we do not use the entire Digits dataset for feature transformation experiments but rather a subset of $10\\%$ of the data. For datasets Office-10 and PACS, we used all of the datasets for the experiment. Additionally, we split each dataset into training and testing sets with an 8:2 ratio.", + "bbox": [ + 88, + 378, + 480, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines We compare FedPall with ten baselines, including SingSet (where each client independently trains a model). FedAvg [22] is the most classic federated learning algorithm, while FedProx [18], PerFedAvg [4], and FedRep [3] are personalized federated learning methods. FedBN [19], ADCOL [17], MOON [16], FedProto [27] and FedHEAL [2] are personalized federated learning algorithms for cross-domain learning, all of which address the issue of non-IID features to some extent. In addition, we explored the ability of RUCR [8] to solve the feature drift problem.", + "bbox": [ + 88, + 625, + 482, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Model and Hyper-parameter Setup All algorithms adopt identical local model architectures for fair comparison. Each local model contains: (1) a ResNet-50 feature extractor (excluding classifier layer), (2) a three-layer MLP classifier with 512 hidden units, and (3) a three-layer MLP amplifier with 2048 input/512 hidden units. Output dimen", + "bbox": [ + 88, + 809, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "sions for classifier and amplifier are set according to dataset categories and data sources, respectively. We maintain client count equal to data sources (one source per client). Local training uses 5 epochs for Digits, and 10 epochs for Office-10 and PACS. Global training employs 100 epochs throughout. Optimization uses SGD (lr=0.01). Except for the digits dataset, for which the values of $\\mu$ and $\\delta$ are set to 0.7 and 0.3 respectively, the values of $\\mu$ and $\\delta$ are set to 0.1 for all other datasets. Common loss hyperparameters remain fixed across algorithms. All experiments run on NVIDIA RTX 4090 GPUs.", + "bbox": [ + 511, + 378, + 906, + 544 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Experimental Analysis", + "text_level": 1, + "bbox": [ + 511, + 559, + 723, + 574 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct the evaluation on three publicly available feature-drifted datasets (Digits, Office-10, and PACS) and compare the performance of the FedPall framework with classical and state-of-the-art baselines. As shown in Table 1, our proposed framework achieves state-of-the-art accuracy on all three datasets.", + "bbox": [ + 511, + 582, + 906, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first discuss the experimental results based on each individual dataset. On the Office-10 dataset, the overall accuracy of the FedPall framework surpasses that of the second-best method, ADCOL, by approximately 3 percentage points. On the Digits dataset, it is evident that FedPall outperforms all other models, achieving an accuracy that is approximately 1.1 percentage points higher than the second-best model, FedBN. The Digits dataset contains images that are relatively easy to classify, and the degree of feature drift is smaller compared to the Office-10 dataset. All baseline models achieve reasonably good accuracy on this dataset. Specifically, adversarial learning helps mitigate the heterogeneous information in the MNIST-M client. Similarly, our algorithm demonstrates strong performance on the PACS dataset, achieving an overall ac", + "bbox": [ + 511, + 674, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3116", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/77b6419042e50b6d2dcb0ac1b2b488f4d56d28a11746740e997f93384ae3d09a.jpg", + "image_caption": [ + "Figure 3. We evaluate the top-1 accuracy averaged over all clients using different loss function combinations on different datasets." + ], + "image_footnote": [], + "bbox": [ + 112, + 95, + 460, + 250 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "curacy that is approximately 1.1 percentage points higher than the second-highest result produced by FedBN. FedPall achieves the highest or second-highest accuracy across all sub-datasets.", + "bbox": [ + 89, + 321, + 483, + 380 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also discuss the performance of FedPall as compared to other state-of-the-art baselines across all three datasets. The average accuracy of FedPall consistently outperforms that of ADCOL in all three datasets, achieving an increase ranging from about 1.1 to 2.9 percentage points. In addition, even though FedBN can achieve accuracy comparable to our method on datasets Digits and PACS, our method outperforms it significantly by 31.5 percentage points on datasets Office-10. As mentioned earlier, the Office-10 dataset comes from real-world data, where feature drift is particularly prominent, and there is also a significant distribution difference between the training and testing sets, leading to the suboptimal performance of the FedBN method on this dataset. In contrast, the special design incorporating both adversarial and collaborative learning in FedPall enables it to adapt well to the Office-10 dataset.", + "bbox": [ + 89, + 382, + 483, + 623 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 635, + 243, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of loss combination In this section, we analyze the impact of KL loss and InfoNCE loss on the performance of the local feature encoder. We conduct ablation studies with three configurations: (1) removing both KL and InfoNCE losses, (2) removing only KL loss, and (3) removing only InfoNCE loss. As shown in Fig 3, the algorithm performs best when all losses are retained, which validates the reliability of the loss combination we designed.", + "bbox": [ + 89, + 657, + 482, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Specifically, on the Office-10 dataset, our method outperforms the model trained with only CE loss by nearly 4 percentage points. Notably, combining CE loss with InfoNCE loss yields worse results than using CE loss alone, suggesting that in the presence of severe feature drift, reinforcing category information through InfoNCE may amplify the drift. While the $\\mathrm{CE} + \\mathrm{KL}$ loss combination performs better than CE loss alone on Office-10, it un", + "bbox": [ + 89, + 780, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b85d870e4e032c96afd52993b09d62786b0c64350e2dd81af09bf89efbd5b840.jpg", + "image_caption": [ + "Figure 4. We plotted the feature distribution of different categories under different clients, corresponding to the four loss combination strategies of Fig. 3" + ], + "image_footnote": [], + "bbox": [ + 516, + 90, + 906, + 292 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "derperforms on the PACS dataset—even falling below CE loss—indicating reduced robustness. This suggests that CE + KL may compromise category information, leading to instability across heterogeneous datasets.", + "bbox": [ + 511, + 375, + 906, + 434 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To assess the impact of KL and InfoNCE losses on feature distributions, we visualize class-wise features across clients in the Office-10 dataset using t-SNE projections of randomly sampled data points (Fig 4). The CE-only model poorly mitigates feature drift, with Client 1 showing unclear decision boundaries. Adding InfoNCE improves intra-client class separation but fails to resolve inter-client drift, leading to ambiguous global boundaries. The $\\mathrm{CE + KL}$ combination reduces cross-client distances for the same class, yielding clearer global boundaries; however, it compresses intra-class spacing in Client 1, causing overlapping clusters and outliers that hinder local classification. In contrast, our unified loss balances these effects: KL aligns same-class features across clients, while CE and InfoNCE promote intra-client separation. This coordination produces compact, well-separated clusters, improving classification and validating our method's effectiveness.", + "bbox": [ + 511, + 435, + 908, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In summary, our method outperforms other loss combinations in the simple handwritten digit recognition task and achieves superior results on real-world datasets, demonstrating strong robustness and generalization.", + "bbox": [ + 511, + 694, + 906, + 753 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Hyperparameter sensitivity analysis. We conducted a hyperparameter sensitivity analysis on the Office-10 dataset by varying the loss weights $\\mu$ and $\\delta$ over $\\{0.1, 0.2, 0.5, 1.0\\}$ . To visualize their impact, we generated an interpolated heatmap of average accuracy. As shown in Fig 5, accuracy remains high and stable when $\\mu$ is within [0.1, 0.4], with $\\delta$ having minimal influence in this range. Notably, the highest accuracy of $69.12\\%$ occurs at $\\mu = \\delta = 0.2$ , fol", + "bbox": [ + 511, + 779, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3117", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/05d10ac64c159e6d145090d86683797efc2e9bcf3b718101ee4d294aa85c670c.jpg", + "image_caption": [ + "Figure 5. The top-1 accuracy $(\\%)$ of different $\\mu$ and $\\delta$ under Office-10 dataset" + ], + "image_footnote": [], + "bbox": [ + 147, + 97, + 426, + 265 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "lowed by $68.62\\%$ at $\\mu = 1.0, \\delta = 0.2$ .", + "bbox": [ + 89, + 342, + 344, + 357 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The results show that our algorithm is not sensitive to the parameter changes of the loss function within a certain range and can maintain high performance.", + "bbox": [ + 89, + 358, + 482, + 405 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3a9f74a729a28a6e768f7469e5d9d6579bc76f63d51b39e7f2153eaac939eebe.jpg", + "image_caption": [ + "(a) Digits Dataset" + ], + "image_footnote": [], + "bbox": [ + 148, + 440, + 424, + 561 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/75462e1586aa30bfad5a6b7e7aff6a76d67e27e61d3530527a077aa6a3ec6057.jpg", + "image_caption": [ + "(b) Office-10 Dataset" + ], + "image_footnote": [], + "bbox": [ + 91, + 575, + 285, + 661 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/164f4c0a75ba8067b0b170a0ae6862c554347873696d46804b64989d9450b530.jpg", + "image_caption": [ + "(c) PACS Dataset", + "Figure 6. Comparison of accuracy with and without training the global classifier on the three datasets." + ], + "image_footnote": [], + "bbox": [ + 289, + 575, + 483, + 661 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison of different classifier replacement methods We evaluate the effectiveness of the global classifier through an ablation study by removing it. As shown in Fig 6, although the simplicity of the Digits dataset results in slightly lower accuracy for the global classifier on some sub-datasets (Fig. 6a), it still outperforms the baseline without a global classifier across multiple sub-datasets. Notably, our method achieves 3.6 percentage points higher accuracy on MNIST-M. The benefits are more pronounced on datasets with substantial feature drift: on Office-10 (Fig. 6b), the global classifier surpasses the baseline on", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "nearly all sub-datasets, achieving $49.3\\%$ on Caltech—a 15.6 percentage points improvement. On PACS (Fig. 6c), it consistently outperforms the baseline across all sub-datasets, with gains of up to 3 percentage points.", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "These results confirm the necessity of FedPall's global classifier, which captures cross-client category information to enhance client-server collaboration and improve the framework's generalization against feature drift.", + "bbox": [ + 511, + 151, + 903, + 212 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Privacy Leakage Risk We evaluate the privacy risk of prototype mixture features using the Data Efficient Mutual Information Neural Estimator (DEMINE) [21]. On the Office-10 training set, the mutual information (MI) scores for: (1) standard Gaussian noise, (2) prototype mixture only, (3) Bernoulli masking only, and (4) our method (prototype mixture + Bernoulli masking) are 2.96, 3.06, 3.04, and 2.10, respectively (lower is better). The corresponding average accuracies are $66.4\\%$ , $64.5\\%$ , $63.7\\%$ , and $67.6\\%$ , compared to $65.8\\%$ in the noise-free case. Our approach not only offers stronger privacy protection but also improves accuracy, attributed to the consistent update direction between the global prototype and feature encoder. Additionally, since encryption is applied post-training with linear time complexity, the computational overhead remains negligible.", + "bbox": [ + 511, + 229, + 906, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 482, + 633, + 497 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this study, we focus on the feature drift problem in FL. The feature drift problem causes the same class samples on different clients to have distinct feature distributions, making it difficult for traditional model aggregation methods to handle such data heterogeneity. To tackle this problem, we design a prototype-based adversarial collaborative framework to unify feature spaces and enhance classification boundaries. The global classifier is retrained with mixed features to further grasp classification-relevant information from a global perspective. Our method has empirically achieved state-of-the-art performance in popular feature-drifted datasets with multiple data sources.", + "bbox": [ + 511, + 507, + 906, + 689 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 511, + 700, + 671, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by the Innovation Team Project of Guangdong Province of China (2024KCXTD017) and Guangdong-Hong Kong-Macau University Joint Laboratory (2023LSYS005).", + "bbox": [ + 511, + 724, + 903, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 514, + 806, + 609, + 821 + ], + "page_idx": 7 + }, + { + "type": "ref_text", + "text": "[1] Haokun Chen, Ahmed Frikha, Denis Krompass, Jindong Gu, and Volker Tresp. Frau: Tackling federated learning with non-iid features via representation augmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4849-4859, 2023. 2", + "bbox": [ + 521, + 830, + 906, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3118", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[2] Yuhang Chen, Wenke Huang, and Mang Ye. Fair federated learning under domain skew with local consistency and domain diversity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12077-12086, 2024. 2, 6", + "[3] Liam Collins, Hamed Hassani, Aryan Mokhtari, and Sanjay Shakkottai. Exploiting shared representations for personalized federated learning. In International conference on machine learning, pages 2089-2099. PMLR, 2021. 6", + "[4] Alireza Fallah, Aryan Mokhtari, and Asuman Ozdaglar. Personalized federated learning with theoretical guarantees: A model-agnostic meta-learning approach. Advances in neural information processing systems, 33:3557–3568, 2020. 6", + "[5] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International conference on machine learning, pages 1180-1189. PMLR, 2015. 6", + "[6] Boqing Gong, Yuan Shi, Fei Sha, and Kristen Grauman. Geodesic flow kernel for unsupervised domain adaptation. In 2012 IEEE conference on computer vision and pattern recognition, pages 2066-2073. IEEE, 2012. 5", + "[7] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3", + "[8] Wenke Huang, Yuxia Liu, Mang Ye, Jun Chen, and Bo Du. Federated learning with long-tailed data via representation unification and classifier rectification. IEEE Transactions on Information Forensics and Security, 2024. 6", + "[9] Jonathan J. Hull. A database for handwritten text recognition research. IEEE Transactions on pattern analysis and machine intelligence, 16(5):550-554, 1994. 6", + "[10] Yihan Jiang, Jakub Konečný, Keith Rush, and Sreeram Kannan. Improving federated learning personalization via model agnostic meta learning. arXiv preprint arXiv:1909.12488, 2019. 1", + "[11] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank J Reddi, Sebastian U Stich, and Ananda Theertha Suresh. Scaffold: Stochastic controlled averaging for ondevice federated learning. arXiv preprint arXiv:1910.06378, 2(6), 2019. 1", + "[12] Jakub Konečny, Brendan McMahan, and Daniel Ramage. Federated optimization: Distributed optimization beyond the datacenter. arXiv preprint arXiv:1511.03575, 2015. 1", + "[13] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 6", + "[14] Boyuan Li, Shengbo Chen, and Zihao Peng. New generation federated learning. Sensors, 22(21):8475, 2022. 1", + "[15] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. In Proceedings of the IEEE international conference on computer vision, pages 5542-5550, 2017. 6", + "[16] Qinbin Li, Bingsheng He, and Dawn Song. Model-contrastive federated learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10713–10722, 2021. 2, 6" + ], + "bbox": [ + 93, + 92, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Qinbin Li, Bingsheng He, and Dawn Song. Adversarial collaborative learning on non-iid features. In International Conference on Machine Learning, pages 19504-19526. PMLR, 2023. 2, 6", + "[18] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. Proceedings of Machine learning and systems, 2:429-450, 2020. 1, 6", + "[19] Xiaoxiao Li, Meirui JIANG, Xiaofei Zhang, Michael Kamp, and Qi Dou. Fedbn: Federated learning on non-iid features via local batch normalization. In International Conference on Learning Representations. 2, 6", + "[20] Zijian Li, Zehong Lin, Jiawei Shao, Yuyi Mao, and Jun Zhang. Fedcir: Client-invariant representation learning for federated non-iid features. IEEE Transactions on Mobile Computing, 2024. 2", + "[21] Xiao Lin, Indranil Sur, Samuel A Nastase, Ajay Divakaran, Uri Hasson, and Mohamed R Amer. Data-efficient mutual information neural estimator. arXiv preprint arXiv:1905.03319, 2019. 8", + "[22] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 1, 6", + "[23] Xutong Mu, Yulong Shen, Ke Cheng, Xueli Geng, Jiaxuan Fu, Tao Zhang, and Zhiwei Zhang. Fedproc: Prototypical contrastive federated learning on non-iid data. Future Generation Computer Systems, 143:93-104, 2023. 3", + "[24] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Baolin Wu, Andrew Y Ng, et al. Reading digits in natural images with unsupervised feature learning. In NIPS workshop on deep learning and unsupervised feature learning, page 4. Granada, 2011. 6", + "[25] Yu Qiao, Md Shirajum Munir, Apurba Adhikary, Huy Q Le, Avi Deb Raha, Chaoning Zhang, and Choong Seon Hong. Mp-fedcl: Multi-prototype federated contrastive learning for edge intelligence. IEEE Internet of Things journal, 2023. 2", + "[26] Min Tan, Yinfu Feng, Lingqiang Chu, Jingcheng Shi, Rong Xiao, Haihong Tang, and Jun Yu. Fedsea: Federated learning via selective feature alignment for non-iid multimodal data. IEEE Transactions on Multimedia, 2023. 2", + "[27] Yue Tan, Guodong Long, Lu Liu, Tianyi Zhou, Qinghua Lu, Jing Jiang, and Chengqi Zhang. Fedproto: Federated prototype learning across heterogeneous clients. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8432-8440, 2022. 2, 3, 6", + "[28] Yue Tan, Guodong Long, Jie Ma, Lu Liu, Tianyi Zhou, and Jing Jiang. Federated learning from pre-trained models: A contrastive learning approach. Advances in neural information processing systems, 35:19332-19344, 2022. 2, 6", + "[29] Lei Wang, Jieming Bian, Letian Zhang, Chen Chen, and Jie Xu. Taming cross-domain representation variance in federated prototype learning with heterogeneous data domains. arXiv preprint arXiv:2403.09048, 2024. 2", + "[30] Li Wang, Quangui Zhang, Lei Sang, Qiang Wu, and Min Xu. Federated prototype-based contrastive learning for privacy-" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3119", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "preserving cross-domain recommendation. arXiv preprint arXiv:2409.03294, 2024. 2", + "[31] Yue Zhao, Meng Li, Liangzhen Lai, Naveen Suda, Damon Civin, and Vikas Chandra. Federated learning with non-iid data. arXiv preprint arXiv:1806.00582, 2018. 1", + "[32] Kaiyang Zhou, Yongxin Yang, Timothy Hospedales, and Tao Xiang. Learning to generate novel domains for domain generalization. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16, pages 561-578. Springer, 2020. 5", + "[33] Tailin Zhou, Jun Zhang, and Danny HK Tsang. Fedfa: Federated learning with feature anchors to align features and classifiers for heterogeneous data. IEEE Transactions on Mobile Computing, 2023. 3", + "[34] Zhuangdi Zhu, Junyuan Hong, and Jiayu Zhou. Data-free knowledge distillation for heterogeneous federated learning. In International conference on machine learning, pages 12878-12889. PMLR, 2021. 2" + ], + "bbox": [ + 91, + 92, + 483, + 345 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3120", + "bbox": [ + 482, + 945, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/FedPall_ Prototype-based Adversarial and Collaborative Learning for Federated Learning with Feature Drift/f6c6ce84-c8dc-4de0-bad0-3f6c4ed86bd8_model.json b/2025/FedPall_ Prototype-based Adversarial and Collaborative Learning for Federated Learning with Feature Drift/f6c6ce84-c8dc-4de0-bad0-3f6c4ed86bd8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c7757aa5ef2bfddac302437f10653a3d3f7bbba1 --- /dev/null +++ b/2025/FedPall_ Prototype-based Adversarial and Collaborative Learning for Federated Learning with Feature Drift/f6c6ce84-c8dc-4de0-bad0-3f6c4ed86bd8_model.json @@ -0,0 +1,2145 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.095, + 0.13, + 0.905, + 0.177 + ], + "angle": 0, + "content": "FedPall: Prototype-based Adversarial and Collaborative Learning for Federated Learning with Feature Drift" + }, + { + "type": "text", + "bbox": [ + 0.132, + 0.203, + 0.871, + 0.276 + ], + "angle": 0, + "content": "Yong Zhang\\(^{1,3}\\), Feng Liang\\(^{1*}\\), Guanghu Yuan\\(^{1}\\), Min Yang\\(^{2}\\), Chengming Li\\(^{1}\\) and Xiping Hu\\(^{1,3*}\\) \nArtificial Intelligence Research Institute, Shenzhen MSU-BIT University\\(^{1}\\) \nShenzhen Institutes of Advanced Technology, Chinese Academy of Sciences\\(^{2}\\) \nSchool of Medical Technology, Beijing Institute of Technology\\(^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.189, + 0.278, + 0.808, + 0.31 + ], + "angle": 0, + "content": "{zhangyong2023@bit.edu.cn, fliang@smbu.edu.cn, yangh@mail.ustc.edu.cn min.yang@siat.ac.cn, licm@smbu.edu.cn, huxp@smbu.edu.cn}" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.345, + 0.327, + 0.361 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.376, + 0.485, + 0.709 + ], + "angle": 0, + "content": "Federated learning (FL) enables collaborative training of a global model in the centralized server with data from multiple parties while preserving privacy. However, data heterogeneity can significantly degrade the performance of the global model when each party uses datasets from different sources to train a local model, thereby affecting personalized local models. Among various cases of data heterogeneity, feature drift, feature space difference among parties, is prevalent in real-life data but remains largely unexplored. Feature drift can distract feature extraction learning in clients and thus lead to poor feature extraction and classification performance. To tackle the problem of feature drift in FL, we propose FedPall, an FL framework that utilizes prototype-based adversarial learning to unify feature spaces and collaborative learning to reinforce class information within the features. Moreover, FedPall leverages mixed features generated from global prototypes and local features to enhance the global classifier with classification-relevant information from a global perspective. Evaluation results on three representative feature-drifted datasets demonstrate FedPall's consistently superior performance in classification with feature-drifted data in the FL scenario." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.753, + 0.222, + 0.768 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.778, + 0.484, + 0.855 + ], + "angle": 0, + "content": "Today, in computer vision, researchers often utilize large amounts of data from various parties to improve the accuracy of algorithms. However, this raises concerns, such as the potential for user privacy leakage caused by sharing private data [12]. Federated learning (FL) [22] is proposed as a" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.346, + 0.825, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.378, + 0.662, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.551, + 0.461, + 0.638, + 0.472 + ], + "angle": 0, + "content": "(a) Feature drift" + }, + { + "type": "image", + "bbox": [ + 0.757, + 0.384, + 0.893, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.751, + 0.461, + 0.901, + 0.473 + ], + "angle": 0, + "content": "(b) Traditional classification" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.499, + 0.663, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.543, + 0.561, + 0.647, + 0.572 + ], + "angle": 0, + "content": "(c) De-clientization" + }, + { + "type": "image", + "bbox": [ + 0.757, + 0.478, + 0.894, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.758, + 0.525, + 0.891, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.761, + 0.561, + 0.892, + 0.573 + ], + "angle": 0, + "content": "(d) Prototype contrastive" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.585, + 0.907, + 0.627 + ], + "angle": 0, + "content": "Figure 1. We show a schematic diagram of feature drift and use different techniques to drive feature distribution to update in different directions." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.658, + 0.907, + 0.734 + ], + "angle": 0, + "content": "privacy-preserving distributed learning paradigm to address these challenges. In the FL paradigm, each party maintains a local client model and collaborates with others to train a global model on the server without sharing the original data, effectively protecting user privacy." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.735, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Particularly, the FL paradigm faces challenges from data heterogeneity [14]. Due to the limited local view of each client, data distribution discrepancies arise across clients (commonly referred to as the non-independent and identically distributed data issue, or non-independent and identically distributed(non-IID) data issue), which can increase generalization error for local models and degrade the performance of the globally aggregated model [10]. Although recent work on non-IID data in FL primarily addresses issues such as stability, client drift, and heterogeneous label distributions among clients [11, 18, 31], feature drift(i.e.," + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.863, + 0.243, + 0.876 + ], + "angle": 0, + "content": "*Corresponding authors." + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.876, + 0.482, + 0.9 + ], + "angle": 0, + "content": "The code is available at https://github.com/DistriAI/FedPall." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.863, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3111" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.094, + 0.092, + 0.482, + 0.56 + ], + "angle": 0, + "content": "variations in feature distributions across clients) is a prevalent yet underexplored challenge in FL. As shown in Fig. 1a, feature drift refers to the phenomenon where samples of the same class exhibit different feature distributions across different clients due to variations in data collection methods, devices, and other factors. This leads to ambiguous decision boundaries, severely impacting the classification performance of federated learning models. However, traditional classification losses (e.g., cross-entropy loss (CE loss), as shown in Fig. 1b) do not account for feature drift. As a result, the collaborative effect among different clients causes the feature space of the same-class samples to influence each other, leading to only slight or even no clearer distinction between decision boundaries for various classes within the same client. As illustrated in Fig. 1c, decentralization is a common approach to addressing feature drift, which involves ignoring inter-client differences to make the feature space of same-class samples across different clients more clustered. In feature drift scenarios, some state-of-the-art algorithms do not explore the role of loss functions but instead optimize local update algorithms to achieve decentralization. For example, FedBN [19] compresses the feature space across clients by adding batch normalization layers to local models. Some methods [1, 34] align feature spaces by sharing partial data for synthetic data augmentation. Additionally, ADCOL sends raw features to the server to update a amplifier and uses the Kullback-Leibler (KL) loss function to achieve decentralization. However, these methods have drawbacks: on the one hand, they may lead to loss of class information, and on the other hand, they pose potential risks of privacy leakage." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.562, + 0.482, + 0.772 + ], + "angle": 0, + "content": "To tackle the above challenge, we propose FedPall, a novel prototype-based adversarial and collaborative learning framework for FL with feature drift. FedPall applies adversarial learning between clients and the server to unify feature spaces, as well as collaborative learning across clients to enhance global decision boundaries. Specifically, FedPall uses adversarial learning to train a feature enhancer and uses KL divergence to align heterogeneous feature spaces between clients, while using prototype contrastive loss to reinforce class information (see Fig. 1d). Finally, adversially aligned features are securely mixed with the global prototypes and uploaded to the server, where a global-view classifier is trained to enhance overall performance. Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.781, + 0.481, + 0.884 + ], + "angle": 0, + "content": "- We propose a novel FL framework, FedPall, to address the feature drift problem. FedPall introduces adversarial learning between clients and the server, and collaborative learning among clients aiming to project feature representations into a unified feature space and reinforce the intrinsic class information. This approach effectively mitigates the feature drift problem in FL settings." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.481, + 0.899 + ], + "angle": 0, + "content": "- We develop a technical strategy that hierarchically in" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.781, + 0.481, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.905, + 0.167 + ], + "angle": 0, + "content": "tegrates the global prototypes with local features to orchestrate client-server collaboration. The mixed prototype features are then used to train a global classifier, which induces the classifier to distill discriminative patterns through cross-client knowledge consolidation." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.168, + 0.905, + 0.212 + ], + "angle": 0, + "content": "- Empirical evaluation on three typical feature-drifted benchmarks demonstrates that our proposed method achieves state-of-the-art classification accuracy." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.226, + 0.654, + 0.241 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.252, + 0.905, + 0.357 + ], + "angle": 0, + "content": "In federated learning (FL), each client's limited local view exacerbates the feature drift problem. Variations in data distributions can cause the same class label to have different feature representations, leading to poor generalization of local models. Existing solutions typically focus on two main approaches: feature space alignment and contrastive prototype learning." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.358, + 0.905, + 0.673 + ], + "angle": 0, + "content": "Some studies address feature drift through feature representation alignment. FRAug [1] employs data augmentation to generate synthetic embeddings that capture both global and client-specific information. FedSea [26] aligns feature distributions to transform raw features into an IID format, mitigating feature drift. FedCiR [20] maximizes mutual information between representations and labels, while minimizing mutual information between client-specific representations conditioned on labels. FedHEAL [2] builds a fair aggregation target on the server side, indirectly aligning feature spaces. MOON [16] encourages local models to align with the global feature distribution through constrained updates. ADCOL [17], unlike traditional aggregation frameworks, uses adversarial learning to enforce a unified representation across clients, alleviating feature drift. However, ADCOL's weak collaboration does not address global class boundaries, and its adversarial mechanism introduces potential privacy risks. Our approach enhances global class boundaries through stronger collaborative learning, while preserving privacy via adversarial learning to mitigate inter-client feature drift." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.675, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Some studies have explored prototype-driven federated learning (FL) paradigms, where prototypes compact feature embeddings, reducing communication bandwidth and preserving privacy [30]. Tan et al.[27, 28] proposed a supervised contrastive loss function using both global and local prototypes to minimize the distance between feature representations and class prototypes, addressing data heterogeneity. However, using the average feature as a class prototype may overlook intra-class variability. To address this, MP-FedCL[25] employs clustering on the client side to generate multiple prototypes per class, capturing intra-class variation and mitigating feature drift. Building on MP-FedCL, FedPLVM [29] improves local training with a two-stage clustering process between clients and the server, and incorporates an \\(\\alpha\\)-sparsity prototype loss to further op" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3112" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.228 + ], + "angle": 0, + "content": "timize performance. By leveraging prototypes' privacy-preserving nature, this method addresses privacy concerns while reinforcing class-specific information in feature representations. In the FedPall framework, we enhance client-server collaboration through global prototypes. The server provides global category information to the client, helping bring similar data closer and separate dissimilar data. Additionally, we use mixed features with global category information to strengthen the global classifier." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.24, + 0.183, + 0.256 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.265, + 0.287, + 0.281 + ], + "angle": 0, + "content": "3.1. Problem Description" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.287, + 0.485, + 0.408 + ], + "angle": 0, + "content": "We define feature drift as follows: Given a dataset \\(\\mathbb{D}\\) with features \\(x\\) and labels \\(y\\), while the conditional distribution \\(P_{i}(x|y)\\) differs across clients, the marginal distribution \\(P(y)\\) remains the same. This means that the same label may have significantly different features across clients. For example, due to variations in environment, geographic location, and cultural differences, the structural features of houses can vary widely." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.408, + 0.484, + 0.513 + ], + "angle": 0, + "content": "With feature drift in FL, our goal is to optimize each client's personalized model loss while leveraging the potential performance gains from collaborative learning across clients [33]. In the FedPall framework, there are a total of \\(N\\) clients, each client \\(n\\) has a private dataset \\(D_{n}\\). Based on this goal, we formulate the overall optimization objective of the FedPall framework as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.52, + 0.483, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta_ {1}, \\theta_ {2}, \\dots , \\theta_ {N} \\in \\mathbb {R} ^ {d _ {1}}} F (\\theta) := \\frac {1}{N} \\sum_ {n = 1} ^ {N} f _ {n} \\left(\\theta_ {n}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.484, + 0.614 + ], + "angle": 0, + "content": "where \\( f_{n} \\) represents the expected loss obtained from client \\( n \\) using the global model parameters under the dataset \\( D_{n} \\), and \\( \\theta_{i} \\) represents the local model parameters of client \\( i \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.623, + 0.281, + 0.637 + ], + "angle": 0, + "content": "3.2. FedPall Framework" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.483, + 0.855 + ], + "angle": 0, + "content": "Existing approaches to addressing the feature drift problem in FL typically focus on either collaborative learning or adversarial learning in isolation. This can result in models that either fail to adequately capture class-related information in the feature representations or exhibit persistent discrepancies in feature distributions across clients. To address these limitations, we propose integrating both adversarial and collaborative learning to effectively mitigate feature drift in FL settings. In this section, we present the FedPall framework by elaborating on its adversarial and collaborative learning. The framework of the overall approach is shown in Fig. 2a. It is structured into four key procedures: generating global prototypes, training local models, training global model, and decentralizing global classifier." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We define certain model symbols here that will be used in this section. The local model \\( F(\\cdot) \\) consists of two components, a feature extractor \\( G(\\cdot) \\) (e.g., Resnet50 [7] for image" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.151 + ], + "angle": 0, + "content": "data) and a classifier \\( H(\\cdot) \\). We use a multilayer perceptron (MLP) as our Amplifier, and except for the output layer, the number of nodes in other layers is consistent with that of the classifier." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.158, + 0.771, + 0.174 + ], + "angle": 0, + "content": "3.2.1. Generating Global Prototypes" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.177, + 0.906, + 0.313 + ], + "angle": 0, + "content": "Several studies [23, 27] suggest that category-centered prototypes are a privacy-friendly form of global knowledge. We leverage collaboration between clients to aggregate and generate global prototypes, as shown in Fig. 2b. Typically, each class prototype is defined as the mean of the feature vectors belonging to that class. The distinction between local and global prototypes lies in the scope of their underlying data. This section will describe the construction of both local and global prototypes." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.313, + 0.906, + 0.342 + ], + "angle": 0, + "content": "The local prototype for the \\(k\\)-th category on client \\(n\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.353, + 0.907, + 0.392 + ], + "angle": 0, + "content": "\\[\nc _ {n} ^ {k} = \\frac {1}{N _ {n} ^ {k}} \\sum_ {(x, y) \\in D _ {n} ^ {k}} G _ {n} (x), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.397, + 0.906, + 0.442 + ], + "angle": 0, + "content": "where \\(D_{n}^{k}\\) and \\(N_{n}^{k}\\) represent the data samples and the number of samples for the \\(k\\)-th category on client \\(n\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.443, + 0.906, + 0.473 + ], + "angle": 0, + "content": "Gathering all the local prototypes together forms a local prototype set, which can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.48, + 0.905, + 0.499 + ], + "angle": 0, + "content": "\\[\n\\mathcal {O} _ {n} = \\left\\{c _ {n} ^ {1}, c _ {n} ^ {2}, \\dots , c _ {n} ^ {K} \\right\\} \\in \\mathbb {R} ^ {K \\times d}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.506, + 0.906, + 0.535 + ], + "angle": 0, + "content": "where \\( K \\) represents the number of categories owned by each client, and \\( d \\) denotes the output feature dimension." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.536, + 0.906, + 0.627 + ], + "angle": 0, + "content": "Since we are only addressing the problem of feature drift, all clients have the same number of categories. Upon receiving the local prototype set and local label proportions \\(\\mathcal{N} = \\{\\{N_n^k\\}_{k=1}^K | n \\in \\mathcal{A}\\}\\) sent by client set \\(\\mathcal{A}\\), the server integrates the local prototypes from all clients to form the global prototypes:" + }, + { + "type": "equation", + "bbox": [ + 0.635, + 0.634, + 0.905, + 0.671 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} ^ {k} = \\sum_ {n} \\frac {N _ {n} ^ {k}}{\\sum_ {n} N _ {n} ^ {k}} c _ {n} ^ {k}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.678, + 0.848, + 0.693 + ], + "angle": 0, + "content": "The global prototypes set can be represented as" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.7, + 0.905, + 0.718 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} = \\left[ \\mathcal {G} ^ {1}, \\dots , \\mathcal {G} ^ {k}, \\dots , \\mathcal {G} ^ {K} \\right]. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.725, + 0.906, + 0.756 + ], + "angle": 0, + "content": "Next, the server sends the global prototype set \\(\\mathcal{G}\\) to each client to guide local model training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.762, + 0.719, + 0.776 + ], + "angle": 0, + "content": "3.2.2. Training Local Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.839 + ], + "angle": 0, + "content": "The goal of training local models is to train effective feature encoders that map raw data from different clients into a unified feature space, where feature distributions are aligned and class-related information is enhanced." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.906, + 0.901 + ], + "angle": 0, + "content": "As mentioned earlier, due to feature drift, training a local classifier alone is not sufficient for accurately classifying data with feature drift. To address this, we apply adversarial learning to train a feature encoder, as illustrated" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3113" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.088, + 0.904, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.425, + 0.234, + 0.574, + 0.246 + ], + "angle": 0, + "content": "(a) The overview of FedPall" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.251, + 0.907, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.428, + 0.658, + 0.442 + ], + "angle": 0, + "content": "Figure 2. The FedPall framework and detailed phases" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.469, + 0.483, + 0.635 + ], + "angle": 0, + "content": "in Fig. 2c. Specifically, we use a global amplifier, trained on the server, which amplifies the heterogeneous information in the features from different clients. At the client's side, we apply the amplifier and use Kullback-Leibler (KL) divergence to reduce the heterogeneous information in the features, thereby creating an adversarial learning setup between the client and server. The objective is to improve the generalization ability of the feature encoders across clients while minimizing the client-specific heterogeneity in the feature representations. Let \\( x^{(i)} \\) denote the \\( i \\)-th dimension of vector \\( x \\), the KL divergence is calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.648, + 0.482, + 0.734 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {K L} = \\sum_ {(\\mathbf {x}, \\mathbf {y}) \\in D _ {n}} D _ {K L} \\left(A _ {n} (G (\\mathbf {x})) | | [ \\frac {1}{N} ] ^ {N}\\right) \\\\ = \\sum_ {(\\mathbf {x}, y) \\in D _ {n}} \\sum_ {i} ^ {N} A _ {n} (G (\\mathbf {x})) ^ {(i)} \\log N A _ {n} (G (\\mathbf {x})) ^ {(i)}, \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.748, + 0.483, + 0.794 + ], + "angle": 0, + "content": "where \\( x^{(i)} \\) denotes the \\( i \\)-th dimension of vector \\( x \\), \\( A_{n}(\\cdot) \\) represents the amplifier of the \\( n \\)-th client, \\( D_{KL}(P||Q) \\) represents the KL divergence of \\( P \\) and \\( Q \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.902 + ], + "angle": 0, + "content": "After adversarial learning, although the feature representations of different clients are mapped to similar feature distributions, the class-related information may be blurred. To address this problem, we propose using contrastive learning to reinforce the class-related information within the feature encoder. To prevent information leakage, we enable collaboration between the global prototypes (server-side) and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.469, + 0.907, + 0.575 + ], + "angle": 0, + "content": "local features (client-side). Specifically, we employ the InfoNCE loss to minimize the distance between local features and their corresponding global prototypes, while maximizing the distance between local features and global prototypes of other classes. This approach strengthens the class-related representation within the feature encoder. The formula for the global-prototype contrastive loss is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.602, + 0.905, + 0.65 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {i n f o N C E}} = \\sum_ {(\\mathrm {x}, \\mathrm {y}) \\in D _ {n}} - \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(G (\\mathrm {x}) , \\mathcal {G} ^ {\\mathrm {y}}\\right) / \\tau\\right)}{\\sum_ {\\mathrm {y} _ {\\alpha} \\in A (\\mathrm {y})} \\exp \\left(\\operatorname {s i m} \\left(\\mathrm {x} , \\mathcal {G} ^ {\\left(\\mathrm {y} _ {\\alpha}\\right)}\\right) / \\tau\\right)}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.65, + 0.907, + 0.711 + ], + "angle": 0, + "content": "where \\( A(y) \\coloneqq \\{y_{\\alpha} \\in [1, |\\mathcal{G}|] : y_{\\alpha} \\neq y\\} \\) is the set of labels distinct from \\( y \\), \\( \\tau \\) is the temperature to adjust the tolerance for feature difference, and \\( sim(x, y) \\) represents the cosine similarity of \\( x \\) and \\( y \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.713, + 0.906, + 0.803 + ], + "angle": 0, + "content": "We combine adversarial learning and collaborative learning to address the feature drift problem in FL. By leveraging the two loss functions defined above, along with the local cross-entropy loss, we progressively train the local feature encoder at each client. The overall loss function for each client is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.821, + 0.905, + 0.838 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {C E (\\mathrm {x}, \\mathrm {y}) \\sim D _ {n}} (F (\\mathrm {x}), \\mathrm {y}) + \\mu \\mathcal {L} _ {K L} + \\delta \\mathcal {L} _ {\\text {i n f o N C E}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.853, + 0.905, + 0.884 + ], + "angle": 0, + "content": "where \\(\\mu\\) denotes the weight of the \\(\\mathcal{L}_{KL}\\) divergence and \\(\\mathcal{L}_{CE}\\) denotes the cross-entropy loss." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.889, + 0.902 + ], + "angle": 0, + "content": "Finally, the local model \\(F(\\cdot)\\) is updated using Eq. (8)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3114" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.092, + 0.298, + 0.106 + ], + "angle": 0, + "content": "3.2.3. Training Global Model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.11, + 0.484, + 0.231 + ], + "angle": 0, + "content": "Due to the client's limited local view, training an accurate classifier is challenging. To address this issue, encrypted mixed features are uploaded to the server, enabling classifier training with a global perspective. Additionally, the server leverages its global view to train an amplifier, which engages in adversarial learning with the client to enhance model robustness. Fig. 2d illustrates this procedure, and the detailed process is described below." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.231, + 0.484, + 0.277 + ], + "angle": 0, + "content": "For each feature \\( z_{n}^{k} \\) of class \\( k \\) on client \\( n \\), we obtain a prototype mixed feature by performing a weighted fusion with the corresponding global prototype:" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.287, + 0.484, + 0.307 + ], + "angle": 0, + "content": "\\[\nr _ {n} ^ {k} = \\alpha \\times z _ {n} ^ {k} + (1 - \\alpha) \\times \\mathcal {G} ^ {k}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.315, + 0.484, + 0.377 + ], + "angle": 0, + "content": "where \\( z_{n}^{k}\\in Z_{n}^{k} = \\{z_{n}^{i,k}\\}_{i = 1}^{S} \\). \\( \\alpha \\sim U(u_f,u_r) \\) are the mix parameters, with \\( u_{f} \\) and \\( u_{d} \\) representing two hyperparameters of a uniform distribution. \\( S \\) represents the number of samples of the \\( k \\)-th category in client \\( n \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.377, + 0.484, + 0.408 + ], + "angle": 0, + "content": "To further reduce the risk of privacy leakage, we employ a Bernoulli mask, represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.418, + 0.482, + 0.455 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} M a s k = \\left\\{X _ {1}, X _ {2}, \\dots , X _ {d} \\right\\}, \\tag {10} \\\\ X _ {i} \\sim \\operatorname {B e r n o u l l i} (\\beta), \\quad \\forall i \\in [ 1, d ]. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.465, + 0.484, + 0.508 + ], + "angle": 0, + "content": "The final output of the prototype mixing mechanism is derived by selecting the mixed feature elements based on the noise mask:" + }, + { + "type": "equation", + "bbox": [ + 0.223, + 0.508, + 0.482, + 0.527 + ], + "angle": 0, + "content": "\\[\n\\tilde {r} _ {n} ^ {k} = \\operatorname {M a s k} \\odot r _ {n} ^ {k}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.533, + 0.371, + 0.547 + ], + "angle": 0, + "content": "where \\(\\odot\\) is an element-wise and operator." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.548, + 0.484, + 0.608 + ], + "angle": 0, + "content": "After generating the prototype mixed features, the client will form the set \\(\\mathcal{D}_{R_L}(R,Y)\\) using the prototype mixed feature set \\(R_{n} = \\{\\tilde{r}_{n}^{1},\\dots,\\tilde{r}_{n}^{k},\\dots,\\tilde{r}_{n}^{K}\\} \\in \\mathbb{R}^{K\\times d}\\) and the corresponding label set \\(Y\\) , which will be sent to the server." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.608, + 0.484, + 0.699 + ], + "angle": 0, + "content": "The server updates the global amplifier \\(A\\) using the mixed feature sets from each client along with the corresponding client IDs. Specifically, we first construct the dataset for training the amplifier, denoted as \\(\\mathcal{D}_{R_I}(R,I)\\) where \\(I\\) represents the client IDs. The amplifier is then updated by minimizing the empirical risk:" + }, + { + "type": "equation", + "bbox": [ + 0.208, + 0.711, + 0.482, + 0.73 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {(R, I) \\sim \\mathcal {D} _ {R _ {I}}} \\ell_ {C E} (R, I). \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.739, + 0.484, + 0.798 + ], + "angle": 0, + "content": "At the same time, the server updates the global classifier \\( C_g \\) using the mixed prototype features and the class labels from the clients \\( Y \\). This is done by minimizing the empirical risk:" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.799, + 0.482, + 0.818 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {(R, Y) \\sim \\mathcal {D} _ {R _ {L}}} \\ell_ {C E} (R, Y). \\tag {13}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.822, + 0.36, + 0.837 + ], + "angle": 0, + "content": "3.2.4. Decentralizing Global Classifier" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Finally, we deploy the global classifier \\( C_g \\) to each client to replace the original local classifiers \\( C_c \\), as shown in Fig. 2e. The purpose of this is to obtain a more generalized classifier that can alleviate the feature drift problem. To allow" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.152 + ], + "angle": 0, + "content": "the global classifier to adapt to the personalized characteristics of local data, we retrain it on the client's local data, thereby enhancing the classifier's accuracy and improving its performance on individual client data distributions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.163, + 0.628, + 0.179 + ], + "angle": 0, + "content": "4. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.189, + 0.907, + 0.4 + ], + "angle": 0, + "content": "Computational Cost Compared to the standard federated learning model, our adversarial collaborative learning approach introduces an amplifier and a global classifier. However, the number of parameters for these two components is much smaller than those of the other components. In our design, both the amplifier and the classifier are designed as three-layer MLPs. Compared to the feature extractor (with a total of 23.5M parameters), the classifier (with a total of 1.32M parameters) and the amplifier (with a total of 1.33M parameters) account for approximately \\(5.61\\%\\) and \\(5.59\\%\\), respectively. Moreover, the client-side amplifier remains frozen, functioning exclusively in the forward pass for loss calculation while being disabled during backpropagation cycles." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.418, + 0.907, + 0.614 + ], + "angle": 0, + "content": "Communication Efficiency Unlike traditional federated learning methods such as FedAvg, which transmit full local model parameters, our approach uses prototype-mixed features as the communication medium. Assuming consistent tensor representations between model parameters and prototype features, each client employs a ResNet-50-based feature extractor that generates 2048-dimensional embeddings. During upload, clients transmit an average of 12,122 prototype-mixed features, compared to approximately 24.8 million model parameters in FedAvg. For downloading, only the amplifier and global classifier parameters are transmitted, reducing communication cost to about \\(10.6\\%\\) of that in FedAvg (see previous paragraph)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.632, + 0.906, + 0.813 + ], + "angle": 0, + "content": "Limitation While our framework currently specializes in image recognition tasks, its extension to NLP or time-series analysis remains unexplored. Successful cross-domain adaptation requires two key developments: (1) establishing domain-specific feature representations and prototype definitions, and (2) redesigning loss functions according to task semantics. For NLP applications, this implies reconfiguring the standard classification paradigm into autoregressive prediction frameworks. Architectural adaptations are equally crucial - particularly the incorporation of RNN-based structures with inherent temporal modeling capabilities for sequential data processing." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.824, + 0.646, + 0.841 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.704, + 0.866 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Datasets We conduct experiments on three publicly available feature drift datasets: Digits [32], Office-10 [6], and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3115" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.146 + ], + "angle": 0, + "content": "Table 1. The top-1 accuracy (\\%) of each algorithm on each sub-dataset of the Office-10, Digits, and PACS datasets is compared, along with the average top-1 accuracy across all sub-datasets. The mean and standard deviation (std) from three random trials (using different random seeds, with other experimental settings remaining the same) are reported. The highest accuracy for each dataset is highlighted in bold, and the second-highest accuracy is underlined." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.156, + 0.905, + 0.356 + ], + "angle": 0, + "content": "
SingleSetFedAvgFedProxPerfedAvgFedRepFedBNMOONFedProtoADCOLRUCRFedHEALours(FedPall)
Office-10amazon74.0(2.7)56.9(2.5)56.6(2.6)57.1(2.2)45.3(1.9)40.8(15.8)51.7(16.1)69.4(2.1)73.3(4.4)52.1(8.5)65.1(3.3)72.9(1.4)
caltech44.7(3.2)46.5(4.6)51.0(5.2)50.8(1.6)38.4(4.9)33.9(6.5)41.3(13.6)39.4(6.3)37.2(1.7)44.3(1.0)44.6(3.2)44.7(8.7)
dslr60.2(6.7)30.1(4.9)33.3(10.4)31.2(4.9)34.4(4.9)38.7(3.2)24.7(1.9)65.6(4.9)76.3(4.9)30.1(6.7)67.7(1.9)77.4(3.2)
webcam71.3(2.6)37.9(6.2)43.7(7.2)47.1(7.8)55.8(2.6)30.5(6.1)33.3(12.7)71.3(4.3)71.3(2.6)37.4(5.0)60.9(1.0)74.7(1.0)
avg62.5(0.4)42.9(1.2)46.1(2.6)46.6(2.9)43.5(1.3)36.0(6.5)37.8(10.9)61.4(1.7)64.5(1.8)41.0(0.6)59.6(0.5)67.5(2.7)
DigitsMNIST95.5(0.2)92.9(2.2)91.8(3.0)90.1(4.8)86.5(6.1)96.7(0.1)93.4(1.1)96.4(0.5)96.3(0.4)92.6(2.0)93.6(0.6)97.2(0.4)
SVHN71.1(0.9)77.4(0.2)76.9(0.3)75.6(0.4)67.2(1.7)79.4(0.3)79.6(0.8)72.5(0.3)75.1(2.1)77.9(0.3)68.3(2.0)78.0(0.4)
USPS86.4(0.3)89.3(0.9)89.2(1.4)88.7(0.7)90.0(3.0)90.1(0.5)81.8(0.7)87.0(0.8)86.7(1.3)88.9(2.4)87.0(0.5)87.3(1.3)
SynthDigits95.2(0.1)95.5(0.1)95.4(0.1)95.0(0.2)94.2(0.8)95.6(0.1)96.6(0.2)95.3(0.6)96.4(0.3)96.0(0.2)89.3(1.6)95.3(0.4)
MNIST-M76.6(0.4)73.8(1.5)74.0(1.5)73.2(0.8)69.1(0.9)76.3(0.4)72.2(0.9)78.3(1.2)78.3(4.4)72.7(0.4)67.8(1.9)85.9(1.4)
avg84.9(0.1)85.8(0.9)85.5(1.1)84.5(1.3)81.4(2.5)87.6(0.1)84.7(0.6)85.9(0.2)86.6(1.3)85.6(0.9)81.2(1.3)88.7(0.2)
PACSart-painting33.6(0.8)25.8(1.9)24.3(4.1)26.5(2.2)26.9(3.3)36.7(1.8)30.6(2.0)32.7(0.7)34.9(1.2)24.7(1.1)31.2(1.2)35.6(0.6)
cartoon58.5(2.5)45.4(2.3)51.4(0.6)48.3(1.2)44.4(2.1)55.6(2.0)51.5(1.8)57.3(1.5)57.2(0.8)47.5(3.3)50.8(0.4)59.7(2.3)
photo63.0(1.9)48.7(3.1)49.6(2.0)46.9(2.6)41.9(2.8)66.1(1.0)53.0(3.3)64.0(1.3)62.1(2.0)47.5(6.2)61.1(2.0)64.7(1.3)
sketch79.7(0.1)49.0(2.0)40.7(1.5)44.4(3.8)40.5(1.3)79.6(1.7)55.1(1.4)79.6(0.8)80.1(1.0)42.2(2.4)73.8(0.1)82.2(0.7)
avg58.7(1.2)42.2(1.6)41.5(1.8)41.5(1.9)38.4(1.1)59.5(1.4)47.6(0.9)58.4(0.3)58.6(0.6)40.5(2.0)54.2(0.5)60.6(0.4)
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.38, + 0.482, + 0.606 + ], + "angle": 0, + "content": "PACS [15]. Specifically, (1) the Digits dataset consists of five different domain sources: MNIST [13], SVHN [24], USPS [9], SynthDigits [5], and MNIST-M [5]; (2) the Office-10 dataset includes four distinct sources: Amazon, Caltech, DSLR, and WebCam; (3) the PACS dataset consists of four sources: Art Painting, Cartoon, Photo, and Sketch. Datasets Office-10 and PACS are real-world images from natural scenes, which inherently exhibit feature drift due to the diversity of their sources. Digits is a digit recognition dataset. In line with [19, 28], we do not use the entire Digits dataset for feature transformation experiments but rather a subset of \\(10\\%\\) of the data. For datasets Office-10 and PACS, we used all of the datasets for the experiment. Additionally, we split each dataset into training and testing sets with an 8:2 ratio." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.625, + 0.483, + 0.793 + ], + "angle": 0, + "content": "Baselines We compare FedPall with ten baselines, including SingSet (where each client independently trains a model). FedAvg [22] is the most classic federated learning algorithm, while FedProx [18], PerFedAvg [4], and FedRep [3] are personalized federated learning methods. FedBN [19], ADCOL [17], MOON [16], FedProto [27] and FedHEAL [2] are personalized federated learning algorithms for cross-domain learning, all of which address the issue of non-IID features to some extent. In addition, we explored the ability of RUCR [8] to solve the feature drift problem." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.81, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Model and Hyper-parameter Setup All algorithms adopt identical local model architectures for fair comparison. Each local model contains: (1) a ResNet-50 feature extractor (excluding classifier layer), (2) a three-layer MLP classifier with 512 hidden units, and (3) a three-layer MLP amplifier with 2048 input/512 hidden units. Output dimen" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.38, + 0.908, + 0.545 + ], + "angle": 0, + "content": "sions for classifier and amplifier are set according to dataset categories and data sources, respectively. We maintain client count equal to data sources (one source per client). Local training uses 5 epochs for Digits, and 10 epochs for Office-10 and PACS. Global training employs 100 epochs throughout. Optimization uses SGD (lr=0.01). Except for the digits dataset, for which the values of \\(\\mu\\) and \\(\\delta\\) are set to 0.7 and 0.3 respectively, the values of \\(\\mu\\) and \\(\\delta\\) are set to 0.1 for all other datasets. Common loss hyperparameters remain fixed across algorithms. All experiments run on NVIDIA RTX 4090 GPUs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.56, + 0.725, + 0.575 + ], + "angle": 0, + "content": "5.2. Experimental Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.907, + 0.673 + ], + "angle": 0, + "content": "We conduct the evaluation on three publicly available feature-drifted datasets (Digits, Office-10, and PACS) and compare the performance of the FedPall framework with classical and state-of-the-art baselines. As shown in Table 1, our proposed framework achieves state-of-the-art accuracy on all three datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.902 + ], + "angle": 0, + "content": "We first discuss the experimental results based on each individual dataset. On the Office-10 dataset, the overall accuracy of the FedPall framework surpasses that of the second-best method, ADCOL, by approximately 3 percentage points. On the Digits dataset, it is evident that FedPall outperforms all other models, achieving an accuracy that is approximately 1.1 percentage points higher than the second-best model, FedBN. The Digits dataset contains images that are relatively easy to classify, and the degree of feature drift is smaller compared to the Office-10 dataset. All baseline models achieve reasonably good accuracy on this dataset. Specifically, adversarial learning helps mitigate the heterogeneous information in the MNIST-M client. Similarly, our algorithm demonstrates strong performance on the PACS dataset, achieving an overall ac" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "3116" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.097, + 0.462, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.265, + 0.483, + 0.294 + ], + "angle": 0, + "content": "Figure 3. We evaluate the top-1 accuracy averaged over all clients using different loss function combinations on different datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.322, + 0.484, + 0.381 + ], + "angle": 0, + "content": "curacy that is approximately 1.1 percentage points higher than the second-highest result produced by FedBN. FedPall achieves the highest or second-highest accuracy across all sub-datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.383, + 0.484, + 0.624 + ], + "angle": 0, + "content": "We also discuss the performance of FedPall as compared to other state-of-the-art baselines across all three datasets. The average accuracy of FedPall consistently outperforms that of ADCOL in all three datasets, achieving an increase ranging from about 1.1 to 2.9 percentage points. In addition, even though FedBN can achieve accuracy comparable to our method on datasets Digits and PACS, our method outperforms it significantly by 31.5 percentage points on datasets Office-10. As mentioned earlier, the Office-10 dataset comes from real-world data, where feature drift is particularly prominent, and there is also a significant distribution difference between the training and testing sets, leading to the suboptimal performance of the FedBN method on this dataset. In contrast, the special design incorporating both adversarial and collaborative learning in FedPall enables it to adapt well to the Office-10 dataset." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.636, + 0.245, + 0.652 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.779 + ], + "angle": 0, + "content": "Effect of loss combination In this section, we analyze the impact of KL loss and InfoNCE loss on the performance of the local feature encoder. We conduct ablation studies with three configurations: (1) removing both KL and InfoNCE losses, (2) removing only KL loss, and (3) removing only InfoNCE loss. As shown in Fig 3, the algorithm performs best when all losses are retained, which validates the reliability of the loss combination we designed." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Specifically, on the Office-10 dataset, our method outperforms the model trained with only CE loss by nearly 4 percentage points. Notably, combining CE loss with InfoNCE loss yields worse results than using CE loss alone, suggesting that in the presence of severe feature drift, reinforcing category information through InfoNCE may amplify the drift. While the \\(\\mathrm{CE} + \\mathrm{KL}\\) loss combination performs better than CE loss alone on Office-10, it un" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.091, + 0.907, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.303, + 0.907, + 0.346 + ], + "angle": 0, + "content": "Figure 4. We plotted the feature distribution of different categories under different clients, corresponding to the four loss combination strategies of Fig. 3" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.375, + 0.907, + 0.435 + ], + "angle": 0, + "content": "derperforms on the PACS dataset—even falling below CE loss—indicating reduced robustness. This suggests that CE + KL may compromise category information, leading to instability across heterogeneous datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.436, + 0.909, + 0.693 + ], + "angle": 0, + "content": "To assess the impact of KL and InfoNCE losses on feature distributions, we visualize class-wise features across clients in the Office-10 dataset using t-SNE projections of randomly sampled data points (Fig 4). The CE-only model poorly mitigates feature drift, with Client 1 showing unclear decision boundaries. Adding InfoNCE improves intra-client class separation but fails to resolve inter-client drift, leading to ambiguous global boundaries. The \\(\\mathrm{CE + KL}\\) combination reduces cross-client distances for the same class, yielding clearer global boundaries; however, it compresses intra-class spacing in Client 1, causing overlapping clusters and outliers that hinder local classification. In contrast, our unified loss balances these effects: KL aligns same-class features across clients, while CE and InfoNCE promote intra-client separation. This coordination produces compact, well-separated clusters, improving classification and validating our method's effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.695, + 0.907, + 0.755 + ], + "angle": 0, + "content": "In summary, our method outperforms other loss combinations in the simple handwritten digit recognition task and achieves superior results on real-world datasets, demonstrating strong robustness and generalization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Hyperparameter sensitivity analysis. We conducted a hyperparameter sensitivity analysis on the Office-10 dataset by varying the loss weights \\(\\mu\\) and \\(\\delta\\) over \\(\\{0.1, 0.2, 0.5, 1.0\\}\\). To visualize their impact, we generated an interpolated heatmap of average accuracy. As shown in Fig 5, accuracy remains high and stable when \\(\\mu\\) is within [0.1, 0.4], with \\(\\delta\\) having minimal influence in this range. Notably, the highest accuracy of \\(69.12\\%\\) occurs at \\(\\mu = \\delta = 0.2\\), fol" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3117" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.148, + 0.098, + 0.428, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.286, + 0.483, + 0.315 + ], + "angle": 0, + "content": "Figure 5. The top-1 accuracy \\((\\%)\\) of different \\(\\mu\\) and \\(\\delta\\) under Office-10 dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.343, + 0.345, + 0.358 + ], + "angle": 0, + "content": "lowed by \\(68.62\\%\\) at \\(\\mu = 1.0, \\delta = 0.2\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.359, + 0.483, + 0.406 + ], + "angle": 0, + "content": "The results show that our algorithm is not sensitive to the parameter changes of the loss function within a certain range and can maintain high performance." + }, + { + "type": "image", + "bbox": [ + 0.149, + 0.441, + 0.426, + 0.562 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.24, + 0.564, + 0.335, + 0.575 + ], + "angle": 0, + "content": "(a) Digits Dataset" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.577, + 0.287, + 0.662 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.133, + 0.665, + 0.246, + 0.676 + ], + "angle": 0, + "content": "(b) Office-10 Dataset" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.577, + 0.484, + 0.662 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.339, + 0.665, + 0.433, + 0.676 + ], + "angle": 0, + "content": "(c) PACS Dataset" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.716 + ], + "angle": 0, + "content": "Figure 6. Comparison of accuracy with and without training the global classifier on the three datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Comparison of different classifier replacement methods We evaluate the effectiveness of the global classifier through an ablation study by removing it. As shown in Fig 6, although the simplicity of the Digits dataset results in slightly lower accuracy for the global classifier on some sub-datasets (Fig. 6a), it still outperforms the baseline without a global classifier across multiple sub-datasets. Notably, our method achieves 3.6 percentage points higher accuracy on MNIST-M. The benefits are more pronounced on datasets with substantial feature drift: on Office-10 (Fig. 6b), the global classifier surpasses the baseline on" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "nearly all sub-datasets, achieving \\(49.3\\%\\) on Caltech—a 15.6 percentage points improvement. On PACS (Fig. 6c), it consistently outperforms the baseline across all sub-datasets, with gains of up to 3 percentage points." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.152, + 0.905, + 0.213 + ], + "angle": 0, + "content": "These results confirm the necessity of FedPall's global classifier, which captures cross-client category information to enhance client-server collaboration and improve the framework's generalization against feature drift." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.23, + 0.907, + 0.472 + ], + "angle": 0, + "content": "Privacy Leakage Risk We evaluate the privacy risk of prototype mixture features using the Data Efficient Mutual Information Neural Estimator (DEMINE) [21]. On the Office-10 training set, the mutual information (MI) scores for: (1) standard Gaussian noise, (2) prototype mixture only, (3) Bernoulli masking only, and (4) our method (prototype mixture + Bernoulli masking) are 2.96, 3.06, 3.04, and 2.10, respectively (lower is better). The corresponding average accuracies are \\(66.4\\%\\), \\(64.5\\%\\), \\(63.7\\%\\), and \\(67.6\\%\\), compared to \\(65.8\\%\\) in the noise-free case. Our approach not only offers stronger privacy protection but also improves accuracy, attributed to the consistent update direction between the global prototype and feature encoder. Additionally, since encryption is applied post-training with linear time complexity, the computational overhead remains negligible." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.483, + 0.634, + 0.498 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.907, + 0.69 + ], + "angle": 0, + "content": "In this study, we focus on the feature drift problem in FL. The feature drift problem causes the same class samples on different clients to have distinct feature distributions, making it difficult for traditional model aggregation methods to handle such data heterogeneity. To tackle this problem, we design a prototype-based adversarial collaborative framework to unify feature spaces and enhance classification boundaries. The global classifier is retrained with mixed features to further grasp classification-relevant information from a global perspective. Our method has empirically achieved state-of-the-art performance in popular feature-drifted datasets with multiple data sources." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.701, + 0.673, + 0.717 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.725, + 0.905, + 0.795 + ], + "angle": 0, + "content": "This work was supported in part by the Innovation Team Project of Guangdong Province of China (2024KCXTD017) and Guangdong-Hong Kong-Macau University Joint Laboratory (2023LSYS005)." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.807, + 0.611, + 0.822 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.831, + 0.907, + 0.902 + ], + "angle": 0, + "content": "[1] Haokun Chen, Ahmed Frikha, Denis Krompass, Jindong Gu, and Volker Tresp. Frau: Tackling federated learning with non-iid features via representation augmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4849-4859, 2023. 2" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3118" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.484, + 0.161 + ], + "angle": 0, + "content": "[2] Yuhang Chen, Wenke Huang, and Mang Ye. Fair federated learning under domain skew with local consistency and domain diversity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12077-12086, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.164, + 0.484, + 0.219 + ], + "angle": 0, + "content": "[3] Liam Collins, Hamed Hassani, Aryan Mokhtari, and Sanjay Shakkottai. Exploiting shared representations for personalized federated learning. In International conference on machine learning, pages 2089-2099. PMLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.221, + 0.484, + 0.276 + ], + "angle": 0, + "content": "[4] Alireza Fallah, Aryan Mokhtari, and Asuman Ozdaglar. Personalized federated learning with theoretical guarantees: A model-agnostic meta-learning approach. Advances in neural information processing systems, 33:3557–3568, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.278, + 0.483, + 0.319 + ], + "angle": 0, + "content": "[5] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International conference on machine learning, pages 1180-1189. PMLR, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.321, + 0.483, + 0.375 + ], + "angle": 0, + "content": "[6] Boqing Gong, Yuan Shi, Fei Sha, and Kristen Grauman. Geodesic flow kernel for unsupervised domain adaptation. In 2012 IEEE conference on computer vision and pattern recognition, pages 2066-2073. IEEE, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.378, + 0.483, + 0.432 + ], + "angle": 0, + "content": "[7] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.434, + 0.483, + 0.489 + ], + "angle": 0, + "content": "[8] Wenke Huang, Yuxia Liu, Mang Ye, Jun Chen, and Bo Du. Federated learning with long-tailed data via representation unification and classifier rectification. IEEE Transactions on Information Forensics and Security, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.491, + 0.484, + 0.532 + ], + "angle": 0, + "content": "[9] Jonathan J. Hull. A database for handwritten text recognition research. IEEE Transactions on pattern analysis and machine intelligence, 16(5):550-554, 1994. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.534, + 0.483, + 0.587 + ], + "angle": 0, + "content": "[10] Yihan Jiang, Jakub Konečný, Keith Rush, and Sreeram Kannan. Improving federated learning personalization via model agnostic meta learning. arXiv preprint arXiv:1909.12488, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.59, + 0.483, + 0.658 + ], + "angle": 0, + "content": "[11] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank J Reddi, Sebastian U Stich, and Ananda Theertha Suresh. Scaffold: Stochastic controlled averaging for ondevice federated learning. arXiv preprint arXiv:1910.06378, 2(6), 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.661, + 0.483, + 0.702 + ], + "angle": 0, + "content": "[12] Jakub Konečny, Brendan McMahan, and Daniel Ramage. Federated optimization: Distributed optimization beyond the datacenter. arXiv preprint arXiv:1511.03575, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.704, + 0.483, + 0.758 + ], + "angle": 0, + "content": "[13] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.483, + 0.788 + ], + "angle": 0, + "content": "[14] Boyuan Li, Shengbo Chen, and Zihao Peng. New generation federated learning. Sensors, 22(21):8475, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[15] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. In Proceedings of the IEEE international conference on computer vision, pages 5542-5550, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.847, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[16] Qinbin Li, Bingsheng He, and Dawn Song. Model-contrastive federated learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10713–10722, 2021. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[17] Qinbin Li, Bingsheng He, and Dawn Song. Adversarial collaborative learning on non-iid features. In International Conference on Machine Learning, pages 19504-19526. PMLR, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[18] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. Proceedings of Machine learning and systems, 2:429-450, 2020. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.208, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[19] Xiaoxiao Li, Meirui JIANG, Xiaofei Zhang, Michael Kamp, and Qi Dou. Fedbn: Federated learning on non-iid features via local batch normalization. In International Conference on Learning Representations. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.264, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[20] Zijian Li, Zehong Lin, Jiawei Shao, Yuyi Mao, and Jun Zhang. Fedcir: Client-invariant representation learning for federated non-iid features. IEEE Transactions on Mobile Computing, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.321, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[21] Xiao Lin, Indranil Sur, Samuel A Nastase, Ajay Divakaran, Uri Hasson, and Mohamed R Amer. Data-efficient mutual information neural estimator. arXiv preprint arXiv:1905.03319, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.378, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[22] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.448, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[23] Xutong Mu, Yulong Shen, Ke Cheng, Xueli Geng, Jiaxuan Fu, Tao Zhang, and Zhiwei Zhang. Fedproc: Prototypical contrastive federated learning on non-iid data. Future Generation Computer Systems, 143:93-104, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.505, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[24] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Baolin Wu, Andrew Y Ng, et al. Reading digits in natural images with unsupervised feature learning. In NIPS workshop on deep learning and unsupervised feature learning, page 4. Granada, 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.576, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[25] Yu Qiao, Md Shirajum Munir, Apurba Adhikary, Huy Q Le, Avi Deb Raha, Chaoning Zhang, and Choong Seon Hong. Mp-fedcl: Multi-prototype federated contrastive learning for edge intelligence. IEEE Internet of Things journal, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.633, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[26] Min Tan, Yinfu Feng, Lingqiang Chu, Jingcheng Shi, Rong Xiao, Haihong Tang, and Jun Yu. Fedsea: Federated learning via selective feature alignment for non-iid multimodal data. IEEE Transactions on Multimedia, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.689, + 0.907, + 0.757 + ], + "angle": 0, + "content": "[27] Yue Tan, Guodong Long, Lu Liu, Tianyi Zhou, Qinghua Lu, Jing Jiang, and Chengqi Zhang. Fedproto: Federated prototype learning across heterogeneous clients. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8432-8440, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.76, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[28] Yue Tan, Guodong Long, Jie Ma, Lu Liu, Tianyi Zhou, and Jing Jiang. Federated learning from pre-trained models: A contrastive learning approach. Advances in neural information processing systems, 35:19332-19344, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.817, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[29] Lei Wang, Jieming Bian, Letian Zhang, Chen Chen, and Jie Xu. Taming cross-domain representation variance in federated prototype learning with heterogeneous data domains. arXiv preprint arXiv:2403.09048, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[30] Li Wang, Quangui Zhang, Lei Sang, Qiang Wu, and Min Xu. Federated prototype-based contrastive learning for privacy-" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3119" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.484, + 0.12 + ], + "angle": 0, + "content": "preserving cross-domain recommendation. arXiv preprint arXiv:2409.03294, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.163 + ], + "angle": 0, + "content": "[31] Yue Zhao, Meng Li, Liangzhen Lai, Naveen Suda, Damon Civin, and Vikas Chandra. Federated learning with non-iid data. arXiv preprint arXiv:1806.00582, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.484, + 0.234 + ], + "angle": 0, + "content": "[32] Kaiyang Zhou, Yongxin Yang, Timothy Hospedales, and Tao Xiang. Learning to generate novel domains for domain generalization. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16, pages 561-578. Springer, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.484, + 0.29 + ], + "angle": 0, + "content": "[33] Tailin Zhou, Jun Zhang, and Danny HK Tsang. Fedfa: Federated learning with feature anchors to align features and classifiers for heterogeneous data. IEEE Transactions on Mobile Computing, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.484, + 0.346 + ], + "angle": 0, + "content": "[34] Zhuangdi Zhu, Junyuan Hong, and Jiayu Zhou. Data-free knowledge distillation for heterogeneous federated learning. In International conference on machine learning, pages 12878-12889. PMLR, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.517, + 0.956 + ], + "angle": 0, + "content": "3120" + } + ] +] \ No newline at end of file diff --git a/2025/FedPall_ Prototype-based Adversarial and Collaborative Learning for Federated Learning with Feature Drift/f6c6ce84-c8dc-4de0-bad0-3f6c4ed86bd8_origin.pdf b/2025/FedPall_ Prototype-based Adversarial and Collaborative Learning for Federated Learning with Feature Drift/f6c6ce84-c8dc-4de0-bad0-3f6c4ed86bd8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..82e517165e78861eebb7469fad970b02aff0e99f --- /dev/null +++ b/2025/FedPall_ Prototype-based Adversarial and Collaborative Learning for Federated Learning with Feature Drift/f6c6ce84-c8dc-4de0-bad0-3f6c4ed86bd8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7d94487d8b0827ce79939cfdec40fe016b2d013486d24300fadc1394f6215ea5 +size 1957797 diff --git a/2025/FedPall_ Prototype-based Adversarial and Collaborative Learning for Federated Learning with Feature Drift/full.md b/2025/FedPall_ Prototype-based Adversarial and Collaborative Learning for Federated Learning with Feature Drift/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c0df38442fa75e2106b96f1cfc78069d091e7788 --- /dev/null +++ b/2025/FedPall_ Prototype-based Adversarial and Collaborative Learning for Federated Learning with Feature Drift/full.md @@ -0,0 +1,330 @@ +# FedPall: Prototype-based Adversarial and Collaborative Learning for Federated Learning with Feature Drift + +Yong Zhang $^{1,3}$ , Feng Liang $^{1*}$ , Guanghu Yuan $^{1}$ , Min Yang $^{2}$ , Chengming Li $^{1}$ and Xiping Hu $^{1,3*}$ +Artificial Intelligence Research Institute, Shenzhen MSU-BIT University $^{1}$ +Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences $^{2}$ +School of Medical Technology, Beijing Institute of Technology $^{3}$ + +{zhangyong2023@bit.edu.cn, fliang@smbu.edu.cn, yangh@mail.ustc.edu.cn min.yang@siat.ac.cn, licm@smbu.edu.cn, huxp@smbu.edu.cn} + +# Abstract + +Federated learning (FL) enables collaborative training of a global model in the centralized server with data from multiple parties while preserving privacy. However, data heterogeneity can significantly degrade the performance of the global model when each party uses datasets from different sources to train a local model, thereby affecting personalized local models. Among various cases of data heterogeneity, feature drift, feature space difference among parties, is prevalent in real-life data but remains largely unexplored. Feature drift can distract feature extraction learning in clients and thus lead to poor feature extraction and classification performance. To tackle the problem of feature drift in FL, we propose FedPall, an FL framework that utilizes prototype-based adversarial learning to unify feature spaces and collaborative learning to reinforce class information within the features. Moreover, FedPall leverages mixed features generated from global prototypes and local features to enhance the global classifier with classification-relevant information from a global perspective. Evaluation results on three representative feature-drifted datasets demonstrate FedPall's consistently superior performance in classification with feature-drifted data in the FL scenario. + +# 1. Introduction + +Today, in computer vision, researchers often utilize large amounts of data from various parties to improve the accuracy of algorithms. However, this raises concerns, such as the potential for user privacy leakage caused by sharing private data [12]. Federated learning (FL) [22] is proposed as a + +![](images/5f28ad42f31865a0a78a661ecc226f3937acd0999dda84a27f498848b5e69682.jpg) + +![](images/4095f66beff724dc93a136d99e045de0ef20920ffa4e2a6024f239976a493e06.jpg) +(a) Feature drift + +![](images/a568814547d2163997a8a18e9df0123816b80d05ec44cfd4b154313ea88cec32.jpg) +(b) Traditional classification + +![](images/12024965fe707cc42954bd98dd10ed6f64b8a06dfa6494cc03f3f7b37108bb0a.jpg) +(c) De-clientization + +![](images/f3d3cb11ceb294bc25984d978289c61d11e553e6549ac6b8968239d7dbcccfea.jpg) + +![](images/ecfb34b8fa4ce36d7ac8a2e09bb1554982f0b7e59df31c5a84c58553b97c8d20.jpg) +(d) Prototype contrastive +Figure 1. We show a schematic diagram of feature drift and use different techniques to drive feature distribution to update in different directions. + +privacy-preserving distributed learning paradigm to address these challenges. In the FL paradigm, each party maintains a local client model and collaborates with others to train a global model on the server without sharing the original data, effectively protecting user privacy. + +Particularly, the FL paradigm faces challenges from data heterogeneity [14]. Due to the limited local view of each client, data distribution discrepancies arise across clients (commonly referred to as the non-independent and identically distributed data issue, or non-independent and identically distributed(non-IID) data issue), which can increase generalization error for local models and degrade the performance of the globally aggregated model [10]. Although recent work on non-IID data in FL primarily addresses issues such as stability, client drift, and heterogeneous label distributions among clients [11, 18, 31], feature drift(i.e., + +variations in feature distributions across clients) is a prevalent yet underexplored challenge in FL. As shown in Fig. 1a, feature drift refers to the phenomenon where samples of the same class exhibit different feature distributions across different clients due to variations in data collection methods, devices, and other factors. This leads to ambiguous decision boundaries, severely impacting the classification performance of federated learning models. However, traditional classification losses (e.g., cross-entropy loss (CE loss), as shown in Fig. 1b) do not account for feature drift. As a result, the collaborative effect among different clients causes the feature space of the same-class samples to influence each other, leading to only slight or even no clearer distinction between decision boundaries for various classes within the same client. As illustrated in Fig. 1c, decentralization is a common approach to addressing feature drift, which involves ignoring inter-client differences to make the feature space of same-class samples across different clients more clustered. In feature drift scenarios, some state-of-the-art algorithms do not explore the role of loss functions but instead optimize local update algorithms to achieve decentralization. For example, FedBN [19] compresses the feature space across clients by adding batch normalization layers to local models. Some methods [1, 34] align feature spaces by sharing partial data for synthetic data augmentation. Additionally, ADCOL sends raw features to the server to update a amplifier and uses the Kullback-Leibler (KL) loss function to achieve decentralization. However, these methods have drawbacks: on the one hand, they may lead to loss of class information, and on the other hand, they pose potential risks of privacy leakage. + +To tackle the above challenge, we propose FedPall, a novel prototype-based adversarial and collaborative learning framework for FL with feature drift. FedPall applies adversarial learning between clients and the server to unify feature spaces, as well as collaborative learning across clients to enhance global decision boundaries. Specifically, FedPall uses adversarial learning to train a feature enhancer and uses KL divergence to align heterogeneous feature spaces between clients, while using prototype contrastive loss to reinforce class information (see Fig. 1d). Finally, adversially aligned features are securely mixed with the global prototypes and uploaded to the server, where a global-view classifier is trained to enhance overall performance. Our contributions are summarized as follows: + +- We propose a novel FL framework, FedPall, to address the feature drift problem. FedPall introduces adversarial learning between clients and the server, and collaborative learning among clients aiming to project feature representations into a unified feature space and reinforce the intrinsic class information. This approach effectively mitigates the feature drift problem in FL settings. +- We develop a technical strategy that hierarchically in + +tegrates the global prototypes with local features to orchestrate client-server collaboration. The mixed prototype features are then used to train a global classifier, which induces the classifier to distill discriminative patterns through cross-client knowledge consolidation. + +- Empirical evaluation on three typical feature-drifted benchmarks demonstrates that our proposed method achieves state-of-the-art classification accuracy. + +# 2. Related Work + +In federated learning (FL), each client's limited local view exacerbates the feature drift problem. Variations in data distributions can cause the same class label to have different feature representations, leading to poor generalization of local models. Existing solutions typically focus on two main approaches: feature space alignment and contrastive prototype learning. + +Some studies address feature drift through feature representation alignment. FRAug [1] employs data augmentation to generate synthetic embeddings that capture both global and client-specific information. FedSea [26] aligns feature distributions to transform raw features into an IID format, mitigating feature drift. FedCiR [20] maximizes mutual information between representations and labels, while minimizing mutual information between client-specific representations conditioned on labels. FedHEAL [2] builds a fair aggregation target on the server side, indirectly aligning feature spaces. MOON [16] encourages local models to align with the global feature distribution through constrained updates. ADCOL [17], unlike traditional aggregation frameworks, uses adversarial learning to enforce a unified representation across clients, alleviating feature drift. However, ADCOL's weak collaboration does not address global class boundaries, and its adversarial mechanism introduces potential privacy risks. Our approach enhances global class boundaries through stronger collaborative learning, while preserving privacy via adversarial learning to mitigate inter-client feature drift. + +Some studies have explored prototype-driven federated learning (FL) paradigms, where prototypes compact feature embeddings, reducing communication bandwidth and preserving privacy [30]. Tan et al.[27, 28] proposed a supervised contrastive loss function using both global and local prototypes to minimize the distance between feature representations and class prototypes, addressing data heterogeneity. However, using the average feature as a class prototype may overlook intra-class variability. To address this, MP-FedCL[25] employs clustering on the client side to generate multiple prototypes per class, capturing intra-class variation and mitigating feature drift. Building on MP-FedCL, FedPLVM [29] improves local training with a two-stage clustering process between clients and the server, and incorporates an $\alpha$ -sparsity prototype loss to further op + +timize performance. By leveraging prototypes' privacy-preserving nature, this method addresses privacy concerns while reinforcing class-specific information in feature representations. In the FedPall framework, we enhance client-server collaboration through global prototypes. The server provides global category information to the client, helping bring similar data closer and separate dissimilar data. Additionally, we use mixed features with global category information to strengthen the global classifier. + +# 3. Method + +# 3.1. Problem Description + +We define feature drift as follows: Given a dataset $\mathbb{D}$ with features $x$ and labels $y$ , while the conditional distribution $P_{i}(x|y)$ differs across clients, the marginal distribution $P(y)$ remains the same. This means that the same label may have significantly different features across clients. For example, due to variations in environment, geographic location, and cultural differences, the structural features of houses can vary widely. + +With feature drift in FL, our goal is to optimize each client's personalized model loss while leveraging the potential performance gains from collaborative learning across clients [33]. In the FedPall framework, there are a total of $N$ clients, each client $n$ has a private dataset $D_{n}$ . Based on this goal, we formulate the overall optimization objective of the FedPall framework as follows: + +$$ +\min _ {\theta_ {1}, \theta_ {2}, \dots , \theta_ {N} \in \mathbb {R} ^ {d _ {1}}} F (\theta) := \frac {1}{N} \sum_ {n = 1} ^ {N} f _ {n} \left(\theta_ {n}\right), \tag {1} +$$ + +where $f_{n}$ represents the expected loss obtained from client $n$ using the global model parameters under the dataset $D_{n}$ , and $\theta_{i}$ represents the local model parameters of client $i$ . + +# 3.2. FedPall Framework + +Existing approaches to addressing the feature drift problem in FL typically focus on either collaborative learning or adversarial learning in isolation. This can result in models that either fail to adequately capture class-related information in the feature representations or exhibit persistent discrepancies in feature distributions across clients. To address these limitations, we propose integrating both adversarial and collaborative learning to effectively mitigate feature drift in FL settings. In this section, we present the FedPall framework by elaborating on its adversarial and collaborative learning. The framework of the overall approach is shown in Fig. 2a. It is structured into four key procedures: generating global prototypes, training local models, training global model, and decentralizing global classifier. + +We define certain model symbols here that will be used in this section. The local model $F(\cdot)$ consists of two components, a feature extractor $G(\cdot)$ (e.g., Resnet50 [7] for image + +data) and a classifier $H(\cdot)$ . We use a multilayer perceptron (MLP) as our Amplifier, and except for the output layer, the number of nodes in other layers is consistent with that of the classifier. + +# 3.2.1. Generating Global Prototypes + +Several studies [23, 27] suggest that category-centered prototypes are a privacy-friendly form of global knowledge. We leverage collaboration between clients to aggregate and generate global prototypes, as shown in Fig. 2b. Typically, each class prototype is defined as the mean of the feature vectors belonging to that class. The distinction between local and global prototypes lies in the scope of their underlying data. This section will describe the construction of both local and global prototypes. + +The local prototype for the $k$ -th category on client $n$ is defined as: + +$$ +c _ {n} ^ {k} = \frac {1}{N _ {n} ^ {k}} \sum_ {(x, y) \in D _ {n} ^ {k}} G _ {n} (x), \tag {2} +$$ + +where $D_{n}^{k}$ and $N_{n}^{k}$ represent the data samples and the number of samples for the $k$ -th category on client $n$ , respectively. + +Gathering all the local prototypes together forms a local prototype set, which can be defined as: + +$$ +\mathcal {O} _ {n} = \left\{c _ {n} ^ {1}, c _ {n} ^ {2}, \dots , c _ {n} ^ {K} \right\} \in \mathbb {R} ^ {K \times d}, \tag {3} +$$ + +where $K$ represents the number of categories owned by each client, and $d$ denotes the output feature dimension. + +Since we are only addressing the problem of feature drift, all clients have the same number of categories. Upon receiving the local prototype set and local label proportions $\mathcal{N} = \{\{N_n^k\}_{k=1}^K | n \in \mathcal{A}\}$ sent by client set $\mathcal{A}$ , the server integrates the local prototypes from all clients to form the global prototypes: + +$$ +\mathcal {G} ^ {k} = \sum_ {n} \frac {N _ {n} ^ {k}}{\sum_ {n} N _ {n} ^ {k}} c _ {n} ^ {k}. \tag {4} +$$ + +The global prototypes set can be represented as + +$$ +\mathcal {G} = \left[ \mathcal {G} ^ {1}, \dots , \mathcal {G} ^ {k}, \dots , \mathcal {G} ^ {K} \right]. \tag {5} +$$ + +Next, the server sends the global prototype set $\mathcal{G}$ to each client to guide local model training. + +# 3.2.2. Training Local Models + +The goal of training local models is to train effective feature encoders that map raw data from different clients into a unified feature space, where feature distributions are aligned and class-related information is enhanced. + +As mentioned earlier, due to feature drift, training a local classifier alone is not sufficient for accurately classifying data with feature drift. To address this, we apply adversarial learning to train a feature encoder, as illustrated + +![](images/1b0fc8c559823d3f6863fb3f196c03b5f51fa498a103b6fdced56d8693c9af2c.jpg) +(a) The overview of FedPall + +![](images/050c4c4db9b7ad9a07ca628e7e6558ca4d87a136f07607029f77cc39ad050fbe.jpg) +Figure 2. The FedPall framework and detailed phases + +in Fig. 2c. Specifically, we use a global amplifier, trained on the server, which amplifies the heterogeneous information in the features from different clients. At the client's side, we apply the amplifier and use Kullback-Leibler (KL) divergence to reduce the heterogeneous information in the features, thereby creating an adversarial learning setup between the client and server. The objective is to improve the generalization ability of the feature encoders across clients while minimizing the client-specific heterogeneity in the feature representations. Let $x^{(i)}$ denote the $i$ -th dimension of vector $x$ , the KL divergence is calculated as: + +$$ +\begin{array}{l} \mathcal {L} _ {K L} = \sum_ {(\mathbf {x}, \mathbf {y}) \in D _ {n}} D _ {K L} \left(A _ {n} (G (\mathbf {x})) | | [ \frac {1}{N} ] ^ {N}\right) \\ = \sum_ {(\mathbf {x}, y) \in D _ {n}} \sum_ {i} ^ {N} A _ {n} (G (\mathbf {x})) ^ {(i)} \log N A _ {n} (G (\mathbf {x})) ^ {(i)}, \tag {6} \\ \end{array} +$$ + +where $x^{(i)}$ denotes the $i$ -th dimension of vector $x$ , $A_{n}(\cdot)$ represents the amplifier of the $n$ -th client, $D_{KL}(P||Q)$ represents the KL divergence of $P$ and $Q$ . + +After adversarial learning, although the feature representations of different clients are mapped to similar feature distributions, the class-related information may be blurred. To address this problem, we propose using contrastive learning to reinforce the class-related information within the feature encoder. To prevent information leakage, we enable collaboration between the global prototypes (server-side) and + +local features (client-side). Specifically, we employ the InfoNCE loss to minimize the distance between local features and their corresponding global prototypes, while maximizing the distance between local features and global prototypes of other classes. This approach strengthens the class-related representation within the feature encoder. The formula for the global-prototype contrastive loss is as follows: + +$$ +\mathcal {L} _ {\text {i n f o N C E}} = \sum_ {(\mathrm {x}, \mathrm {y}) \in D _ {n}} - \log \frac {\exp \left(\operatorname {s i m} \left(G (\mathrm {x}) , \mathcal {G} ^ {\mathrm {y}}\right) / \tau\right)}{\sum_ {\mathrm {y} _ {\alpha} \in A (\mathrm {y})} \exp \left(\operatorname {s i m} \left(\mathrm {x} , \mathcal {G} ^ {\left(\mathrm {y} _ {\alpha}\right)}\right) / \tau\right)}, \tag {7} +$$ + +where $A(y) \coloneqq \{y_{\alpha} \in [1, |\mathcal{G}|] : y_{\alpha} \neq y\}$ is the set of labels distinct from $y$ , $\tau$ is the temperature to adjust the tolerance for feature difference, and $sim(x, y)$ represents the cosine similarity of $x$ and $y$ . + +We combine adversarial learning and collaborative learning to address the feature drift problem in FL. By leveraging the two loss functions defined above, along with the local cross-entropy loss, we progressively train the local feature encoder at each client. The overall loss function for each client is as follows: + +$$ +\mathcal {L} = \mathcal {L} _ {C E (\mathrm {x}, \mathrm {y}) \sim D _ {n}} (F (\mathrm {x}), \mathrm {y}) + \mu \mathcal {L} _ {K L} + \delta \mathcal {L} _ {\text {i n f o N C E}}, \tag {8} +$$ + +where $\mu$ denotes the weight of the $\mathcal{L}_{KL}$ divergence and $\mathcal{L}_{CE}$ denotes the cross-entropy loss. + +Finally, the local model $F(\cdot)$ is updated using Eq. (8). + +# 3.2.3. Training Global Model + +Due to the client's limited local view, training an accurate classifier is challenging. To address this issue, encrypted mixed features are uploaded to the server, enabling classifier training with a global perspective. Additionally, the server leverages its global view to train an amplifier, which engages in adversarial learning with the client to enhance model robustness. Fig. 2d illustrates this procedure, and the detailed process is described below. + +For each feature $z_{n}^{k}$ of class $k$ on client $n$ , we obtain a prototype mixed feature by performing a weighted fusion with the corresponding global prototype: + +$$ +r _ {n} ^ {k} = \alpha \times z _ {n} ^ {k} + (1 - \alpha) \times \mathcal {G} ^ {k}, \tag {9} +$$ + +where $z_{n}^{k}\in Z_{n}^{k} = \{z_{n}^{i,k}\}_{i = 1}^{S}$ . $\alpha \sim U(u_f,u_r)$ are the mix parameters, with $u_{f}$ and $u_{d}$ representing two hyperparameters of a uniform distribution. $S$ represents the number of samples of the $k$ -th category in client $n$ . + +To further reduce the risk of privacy leakage, we employ a Bernoulli mask, represented as: + +$$ +\begin{array}{l} M a s k = \left\{X _ {1}, X _ {2}, \dots , X _ {d} \right\}, \tag {10} \\ X _ {i} \sim \operatorname {B e r n o u l l i} (\beta), \quad \forall i \in [ 1, d ]. \\ \end{array} +$$ + +The final output of the prototype mixing mechanism is derived by selecting the mixed feature elements based on the noise mask: + +$$ +\tilde {r} _ {n} ^ {k} = \operatorname {M a s k} \odot r _ {n} ^ {k}, \tag {11} +$$ + +where $\odot$ is an element-wise and operator. + +After generating the prototype mixed features, the client will form the set $\mathcal{D}_{R_L}(R,Y)$ using the prototype mixed feature set $R_{n} = \{\tilde{r}_{n}^{1},\dots,\tilde{r}_{n}^{k},\dots,\tilde{r}_{n}^{K}\} \in \mathbb{R}^{K\times d}$ and the corresponding label set $Y$ , which will be sent to the server. + +The server updates the global amplifier $A$ using the mixed feature sets from each client along with the corresponding client IDs. Specifically, we first construct the dataset for training the amplifier, denoted as $\mathcal{D}_{R_I}(R,I)$ where $I$ represents the client IDs. The amplifier is then updated by minimizing the empirical risk: + +$$ +\mathbb {E} _ {(R, I) \sim \mathcal {D} _ {R _ {I}}} \ell_ {C E} (R, I). \tag {12} +$$ + +At the same time, the server updates the global classifier $C_g$ using the mixed prototype features and the class labels from the clients $Y$ . This is done by minimizing the empirical risk: + +$$ +\mathbb {E} _ {(R, Y) \sim \mathcal {D} _ {R _ {L}}} \ell_ {C E} (R, Y). \tag {13} +$$ + +# 3.2.4. Decentralizing Global Classifier + +Finally, we deploy the global classifier $C_g$ to each client to replace the original local classifiers $C_c$ , as shown in Fig. 2e. The purpose of this is to obtain a more generalized classifier that can alleviate the feature drift problem. To allow + +the global classifier to adapt to the personalized characteristics of local data, we retrain it on the client's local data, thereby enhancing the classifier's accuracy and improving its performance on individual client data distributions. + +# 4. Discussion + +Computational Cost Compared to the standard federated learning model, our adversarial collaborative learning approach introduces an amplifier and a global classifier. However, the number of parameters for these two components is much smaller than those of the other components. In our design, both the amplifier and the classifier are designed as three-layer MLPs. Compared to the feature extractor (with a total of 23.5M parameters), the classifier (with a total of 1.32M parameters) and the amplifier (with a total of 1.33M parameters) account for approximately $5.61\%$ and $5.59\%$ , respectively. Moreover, the client-side amplifier remains frozen, functioning exclusively in the forward pass for loss calculation while being disabled during backpropagation cycles. + +Communication Efficiency Unlike traditional federated learning methods such as FedAvg, which transmit full local model parameters, our approach uses prototype-mixed features as the communication medium. Assuming consistent tensor representations between model parameters and prototype features, each client employs a ResNet-50-based feature extractor that generates 2048-dimensional embeddings. During upload, clients transmit an average of 12,122 prototype-mixed features, compared to approximately 24.8 million model parameters in FedAvg. For downloading, only the amplifier and global classifier parameters are transmitted, reducing communication cost to about $10.6\%$ of that in FedAvg (see previous paragraph). + +Limitation While our framework currently specializes in image recognition tasks, its extension to NLP or time-series analysis remains unexplored. Successful cross-domain adaptation requires two key developments: (1) establishing domain-specific feature representations and prototype definitions, and (2) redesigning loss functions according to task semantics. For NLP applications, this implies reconfiguring the standard classification paradigm into autoregressive prediction frameworks. Architectural adaptations are equally crucial - particularly the incorporation of RNN-based structures with inherent temporal modeling capabilities for sequential data processing. + +# 5. Experiments + +# 5.1. Experimental Setup + +Datasets We conduct experiments on three publicly available feature drift datasets: Digits [32], Office-10 [6], and + +Table 1. The top-1 accuracy (\%) of each algorithm on each sub-dataset of the Office-10, Digits, and PACS datasets is compared, along with the average top-1 accuracy across all sub-datasets. The mean and standard deviation (std) from three random trials (using different random seeds, with other experimental settings remaining the same) are reported. The highest accuracy for each dataset is highlighted in bold, and the second-highest accuracy is underlined. + +
SingleSetFedAvgFedProxPerfedAvgFedRepFedBNMOONFedProtoADCOLRUCRFedHEALours(FedPall)
Office-10amazon74.0(2.7)56.9(2.5)56.6(2.6)57.1(2.2)45.3(1.9)40.8(15.8)51.7(16.1)69.4(2.1)73.3(4.4)52.1(8.5)65.1(3.3)72.9(1.4)
caltech44.7(3.2)46.5(4.6)51.0(5.2)50.8(1.6)38.4(4.9)33.9(6.5)41.3(13.6)39.4(6.3)37.2(1.7)44.3(1.0)44.6(3.2)44.7(8.7)
dslr60.2(6.7)30.1(4.9)33.3(10.4)31.2(4.9)34.4(4.9)38.7(3.2)24.7(1.9)65.6(4.9)76.3(4.9)30.1(6.7)67.7(1.9)77.4(3.2)
webcam71.3(2.6)37.9(6.2)43.7(7.2)47.1(7.8)55.8(2.6)30.5(6.1)33.3(12.7)71.3(4.3)71.3(2.6)37.4(5.0)60.9(1.0)74.7(1.0)
avg62.5(0.4)42.9(1.2)46.1(2.6)46.6(2.9)43.5(1.3)36.0(6.5)37.8(10.9)61.4(1.7)64.5(1.8)41.0(0.6)59.6(0.5)67.5(2.7)
DigitsMNIST95.5(0.2)92.9(2.2)91.8(3.0)90.1(4.8)86.5(6.1)96.7(0.1)93.4(1.1)96.4(0.5)96.3(0.4)92.6(2.0)93.6(0.6)97.2(0.4)
SVHN71.1(0.9)77.4(0.2)76.9(0.3)75.6(0.4)67.2(1.7)79.4(0.3)79.6(0.8)72.5(0.3)75.1(2.1)77.9(0.3)68.3(2.0)78.0(0.4)
USPS86.4(0.3)89.3(0.9)89.2(1.4)88.7(0.7)90.0(3.0)90.1(0.5)81.8(0.7)87.0(0.8)86.7(1.3)88.9(2.4)87.0(0.5)87.3(1.3)
SynthDigits95.2(0.1)95.5(0.1)95.4(0.1)95.0(0.2)94.2(0.8)95.6(0.1)96.6(0.2)95.3(0.6)96.4(0.3)96.0(0.2)89.3(1.6)95.3(0.4)
MNIST-M76.6(0.4)73.8(1.5)74.0(1.5)73.2(0.8)69.1(0.9)76.3(0.4)72.2(0.9)78.3(1.2)78.3(4.4)72.7(0.4)67.8(1.9)85.9(1.4)
avg84.9(0.1)85.8(0.9)85.5(1.1)84.5(1.3)81.4(2.5)87.6(0.1)84.7(0.6)85.9(0.2)86.6(1.3)85.6(0.9)81.2(1.3)88.7(0.2)
PACSart-painting33.6(0.8)25.8(1.9)24.3(4.1)26.5(2.2)26.9(3.3)36.7(1.8)30.6(2.0)32.7(0.7)34.9(1.2)24.7(1.1)31.2(1.2)35.6(0.6)
cartoon58.5(2.5)45.4(2.3)51.4(0.6)48.3(1.2)44.4(2.1)55.6(2.0)51.5(1.8)57.3(1.5)57.2(0.8)47.5(3.3)50.8(0.4)59.7(2.3)
photo63.0(1.9)48.7(3.1)49.6(2.0)46.9(2.6)41.9(2.8)66.1(1.0)53.0(3.3)64.0(1.3)62.1(2.0)47.5(6.2)61.1(2.0)64.7(1.3)
sketch79.7(0.1)49.0(2.0)40.7(1.5)44.4(3.8)40.5(1.3)79.6(1.7)55.1(1.4)79.6(0.8)80.1(1.0)42.2(2.4)73.8(0.1)82.2(0.7)
avg58.7(1.2)42.2(1.6)41.5(1.8)41.5(1.9)38.4(1.1)59.5(1.4)47.6(0.9)58.4(0.3)58.6(0.6)40.5(2.0)54.2(0.5)60.6(0.4)
+ +PACS [15]. Specifically, (1) the Digits dataset consists of five different domain sources: MNIST [13], SVHN [24], USPS [9], SynthDigits [5], and MNIST-M [5]; (2) the Office-10 dataset includes four distinct sources: Amazon, Caltech, DSLR, and WebCam; (3) the PACS dataset consists of four sources: Art Painting, Cartoon, Photo, and Sketch. Datasets Office-10 and PACS are real-world images from natural scenes, which inherently exhibit feature drift due to the diversity of their sources. Digits is a digit recognition dataset. In line with [19, 28], we do not use the entire Digits dataset for feature transformation experiments but rather a subset of $10\%$ of the data. For datasets Office-10 and PACS, we used all of the datasets for the experiment. Additionally, we split each dataset into training and testing sets with an 8:2 ratio. + +Baselines We compare FedPall with ten baselines, including SingSet (where each client independently trains a model). FedAvg [22] is the most classic federated learning algorithm, while FedProx [18], PerFedAvg [4], and FedRep [3] are personalized federated learning methods. FedBN [19], ADCOL [17], MOON [16], FedProto [27] and FedHEAL [2] are personalized federated learning algorithms for cross-domain learning, all of which address the issue of non-IID features to some extent. In addition, we explored the ability of RUCR [8] to solve the feature drift problem. + +Model and Hyper-parameter Setup All algorithms adopt identical local model architectures for fair comparison. Each local model contains: (1) a ResNet-50 feature extractor (excluding classifier layer), (2) a three-layer MLP classifier with 512 hidden units, and (3) a three-layer MLP amplifier with 2048 input/512 hidden units. Output dimen + +sions for classifier and amplifier are set according to dataset categories and data sources, respectively. We maintain client count equal to data sources (one source per client). Local training uses 5 epochs for Digits, and 10 epochs for Office-10 and PACS. Global training employs 100 epochs throughout. Optimization uses SGD (lr=0.01). Except for the digits dataset, for which the values of $\mu$ and $\delta$ are set to 0.7 and 0.3 respectively, the values of $\mu$ and $\delta$ are set to 0.1 for all other datasets. Common loss hyperparameters remain fixed across algorithms. All experiments run on NVIDIA RTX 4090 GPUs. + +# 5.2. Experimental Analysis + +We conduct the evaluation on three publicly available feature-drifted datasets (Digits, Office-10, and PACS) and compare the performance of the FedPall framework with classical and state-of-the-art baselines. As shown in Table 1, our proposed framework achieves state-of-the-art accuracy on all three datasets. + +We first discuss the experimental results based on each individual dataset. On the Office-10 dataset, the overall accuracy of the FedPall framework surpasses that of the second-best method, ADCOL, by approximately 3 percentage points. On the Digits dataset, it is evident that FedPall outperforms all other models, achieving an accuracy that is approximately 1.1 percentage points higher than the second-best model, FedBN. The Digits dataset contains images that are relatively easy to classify, and the degree of feature drift is smaller compared to the Office-10 dataset. All baseline models achieve reasonably good accuracy on this dataset. Specifically, adversarial learning helps mitigate the heterogeneous information in the MNIST-M client. Similarly, our algorithm demonstrates strong performance on the PACS dataset, achieving an overall ac + +![](images/77b6419042e50b6d2dcb0ac1b2b488f4d56d28a11746740e997f93384ae3d09a.jpg) +Figure 3. We evaluate the top-1 accuracy averaged over all clients using different loss function combinations on different datasets. + +curacy that is approximately 1.1 percentage points higher than the second-highest result produced by FedBN. FedPall achieves the highest or second-highest accuracy across all sub-datasets. + +We also discuss the performance of FedPall as compared to other state-of-the-art baselines across all three datasets. The average accuracy of FedPall consistently outperforms that of ADCOL in all three datasets, achieving an increase ranging from about 1.1 to 2.9 percentage points. In addition, even though FedBN can achieve accuracy comparable to our method on datasets Digits and PACS, our method outperforms it significantly by 31.5 percentage points on datasets Office-10. As mentioned earlier, the Office-10 dataset comes from real-world data, where feature drift is particularly prominent, and there is also a significant distribution difference between the training and testing sets, leading to the suboptimal performance of the FedBN method on this dataset. In contrast, the special design incorporating both adversarial and collaborative learning in FedPall enables it to adapt well to the Office-10 dataset. + +# 5.3. Ablation Study + +Effect of loss combination In this section, we analyze the impact of KL loss and InfoNCE loss on the performance of the local feature encoder. We conduct ablation studies with three configurations: (1) removing both KL and InfoNCE losses, (2) removing only KL loss, and (3) removing only InfoNCE loss. As shown in Fig 3, the algorithm performs best when all losses are retained, which validates the reliability of the loss combination we designed. + +Specifically, on the Office-10 dataset, our method outperforms the model trained with only CE loss by nearly 4 percentage points. Notably, combining CE loss with InfoNCE loss yields worse results than using CE loss alone, suggesting that in the presence of severe feature drift, reinforcing category information through InfoNCE may amplify the drift. While the $\mathrm{CE} + \mathrm{KL}$ loss combination performs better than CE loss alone on Office-10, it un + +![](images/b85d870e4e032c96afd52993b09d62786b0c64350e2dd81af09bf89efbd5b840.jpg) +Figure 4. We plotted the feature distribution of different categories under different clients, corresponding to the four loss combination strategies of Fig. 3 + +derperforms on the PACS dataset—even falling below CE loss—indicating reduced robustness. This suggests that CE + KL may compromise category information, leading to instability across heterogeneous datasets. + +To assess the impact of KL and InfoNCE losses on feature distributions, we visualize class-wise features across clients in the Office-10 dataset using t-SNE projections of randomly sampled data points (Fig 4). The CE-only model poorly mitigates feature drift, with Client 1 showing unclear decision boundaries. Adding InfoNCE improves intra-client class separation but fails to resolve inter-client drift, leading to ambiguous global boundaries. The $\mathrm{CE + KL}$ combination reduces cross-client distances for the same class, yielding clearer global boundaries; however, it compresses intra-class spacing in Client 1, causing overlapping clusters and outliers that hinder local classification. In contrast, our unified loss balances these effects: KL aligns same-class features across clients, while CE and InfoNCE promote intra-client separation. This coordination produces compact, well-separated clusters, improving classification and validating our method's effectiveness. + +In summary, our method outperforms other loss combinations in the simple handwritten digit recognition task and achieves superior results on real-world datasets, demonstrating strong robustness and generalization. + +Hyperparameter sensitivity analysis. We conducted a hyperparameter sensitivity analysis on the Office-10 dataset by varying the loss weights $\mu$ and $\delta$ over $\{0.1, 0.2, 0.5, 1.0\}$ . To visualize their impact, we generated an interpolated heatmap of average accuracy. As shown in Fig 5, accuracy remains high and stable when $\mu$ is within [0.1, 0.4], with $\delta$ having minimal influence in this range. Notably, the highest accuracy of $69.12\%$ occurs at $\mu = \delta = 0.2$ , fol + +![](images/05d10ac64c159e6d145090d86683797efc2e9bcf3b718101ee4d294aa85c670c.jpg) +Figure 5. The top-1 accuracy $(\%)$ of different $\mu$ and $\delta$ under Office-10 dataset + +lowed by $68.62\%$ at $\mu = 1.0, \delta = 0.2$ . + +The results show that our algorithm is not sensitive to the parameter changes of the loss function within a certain range and can maintain high performance. + +![](images/3a9f74a729a28a6e768f7469e5d9d6579bc76f63d51b39e7f2153eaac939eebe.jpg) +(a) Digits Dataset + +![](images/75462e1586aa30bfad5a6b7e7aff6a76d67e27e61d3530527a077aa6a3ec6057.jpg) +(b) Office-10 Dataset + +![](images/164f4c0a75ba8067b0b170a0ae6862c554347873696d46804b64989d9450b530.jpg) +(c) PACS Dataset +Figure 6. Comparison of accuracy with and without training the global classifier on the three datasets. + +Comparison of different classifier replacement methods We evaluate the effectiveness of the global classifier through an ablation study by removing it. As shown in Fig 6, although the simplicity of the Digits dataset results in slightly lower accuracy for the global classifier on some sub-datasets (Fig. 6a), it still outperforms the baseline without a global classifier across multiple sub-datasets. Notably, our method achieves 3.6 percentage points higher accuracy on MNIST-M. The benefits are more pronounced on datasets with substantial feature drift: on Office-10 (Fig. 6b), the global classifier surpasses the baseline on + +nearly all sub-datasets, achieving $49.3\%$ on Caltech—a 15.6 percentage points improvement. On PACS (Fig. 6c), it consistently outperforms the baseline across all sub-datasets, with gains of up to 3 percentage points. + +These results confirm the necessity of FedPall's global classifier, which captures cross-client category information to enhance client-server collaboration and improve the framework's generalization against feature drift. + +Privacy Leakage Risk We evaluate the privacy risk of prototype mixture features using the Data Efficient Mutual Information Neural Estimator (DEMINE) [21]. On the Office-10 training set, the mutual information (MI) scores for: (1) standard Gaussian noise, (2) prototype mixture only, (3) Bernoulli masking only, and (4) our method (prototype mixture + Bernoulli masking) are 2.96, 3.06, 3.04, and 2.10, respectively (lower is better). The corresponding average accuracies are $66.4\%$ , $64.5\%$ , $63.7\%$ , and $67.6\%$ , compared to $65.8\%$ in the noise-free case. Our approach not only offers stronger privacy protection but also improves accuracy, attributed to the consistent update direction between the global prototype and feature encoder. Additionally, since encryption is applied post-training with linear time complexity, the computational overhead remains negligible. + +# 6. Conclusion + +In this study, we focus on the feature drift problem in FL. The feature drift problem causes the same class samples on different clients to have distinct feature distributions, making it difficult for traditional model aggregation methods to handle such data heterogeneity. To tackle this problem, we design a prototype-based adversarial collaborative framework to unify feature spaces and enhance classification boundaries. The global classifier is retrained with mixed features to further grasp classification-relevant information from a global perspective. Our method has empirically achieved state-of-the-art performance in popular feature-drifted datasets with multiple data sources. + +# Acknowledgement + +This work was supported in part by the Innovation Team Project of Guangdong Province of China (2024KCXTD017) and Guangdong-Hong Kong-Macau University Joint Laboratory (2023LSYS005). + +# References + +[1] Haokun Chen, Ahmed Frikha, Denis Krompass, Jindong Gu, and Volker Tresp. Frau: Tackling federated learning with non-iid features via representation augmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4849-4859, 2023. 2 + +[2] Yuhang Chen, Wenke Huang, and Mang Ye. Fair federated learning under domain skew with local consistency and domain diversity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12077-12086, 2024. 2, 6 +[3] Liam Collins, Hamed Hassani, Aryan Mokhtari, and Sanjay Shakkottai. Exploiting shared representations for personalized federated learning. In International conference on machine learning, pages 2089-2099. PMLR, 2021. 6 +[4] Alireza Fallah, Aryan Mokhtari, and Asuman Ozdaglar. Personalized federated learning with theoretical guarantees: A model-agnostic meta-learning approach. Advances in neural information processing systems, 33:3557–3568, 2020. 6 +[5] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International conference on machine learning, pages 1180-1189. PMLR, 2015. 6 +[6] Boqing Gong, Yuan Shi, Fei Sha, and Kristen Grauman. Geodesic flow kernel for unsupervised domain adaptation. In 2012 IEEE conference on computer vision and pattern recognition, pages 2066-2073. IEEE, 2012. 5 +[7] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3 +[8] Wenke Huang, Yuxia Liu, Mang Ye, Jun Chen, and Bo Du. Federated learning with long-tailed data via representation unification and classifier rectification. IEEE Transactions on Information Forensics and Security, 2024. 6 +[9] Jonathan J. Hull. A database for handwritten text recognition research. IEEE Transactions on pattern analysis and machine intelligence, 16(5):550-554, 1994. 6 +[10] Yihan Jiang, Jakub Konečný, Keith Rush, and Sreeram Kannan. Improving federated learning personalization via model agnostic meta learning. arXiv preprint arXiv:1909.12488, 2019. 1 +[11] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank J Reddi, Sebastian U Stich, and Ananda Theertha Suresh. Scaffold: Stochastic controlled averaging for ondevice federated learning. arXiv preprint arXiv:1910.06378, 2(6), 2019. 1 +[12] Jakub Konečny, Brendan McMahan, and Daniel Ramage. Federated optimization: Distributed optimization beyond the datacenter. arXiv preprint arXiv:1511.03575, 2015. 1 +[13] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 6 +[14] Boyuan Li, Shengbo Chen, and Zihao Peng. New generation federated learning. Sensors, 22(21):8475, 2022. 1 +[15] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. In Proceedings of the IEEE international conference on computer vision, pages 5542-5550, 2017. 6 +[16] Qinbin Li, Bingsheng He, and Dawn Song. Model-contrastive federated learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10713–10722, 2021. 2, 6 + +[17] Qinbin Li, Bingsheng He, and Dawn Song. Adversarial collaborative learning on non-iid features. In International Conference on Machine Learning, pages 19504-19526. PMLR, 2023. 2, 6 +[18] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. Proceedings of Machine learning and systems, 2:429-450, 2020. 1, 6 +[19] Xiaoxiao Li, Meirui JIANG, Xiaofei Zhang, Michael Kamp, and Qi Dou. Fedbn: Federated learning on non-iid features via local batch normalization. In International Conference on Learning Representations. 2, 6 +[20] Zijian Li, Zehong Lin, Jiawei Shao, Yuyi Mao, and Jun Zhang. Fedcir: Client-invariant representation learning for federated non-iid features. IEEE Transactions on Mobile Computing, 2024. 2 +[21] Xiao Lin, Indranil Sur, Samuel A Nastase, Ajay Divakaran, Uri Hasson, and Mohamed R Amer. Data-efficient mutual information neural estimator. arXiv preprint arXiv:1905.03319, 2019. 8 +[22] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 1, 6 +[23] Xutong Mu, Yulong Shen, Ke Cheng, Xueli Geng, Jiaxuan Fu, Tao Zhang, and Zhiwei Zhang. Fedproc: Prototypical contrastive federated learning on non-iid data. Future Generation Computer Systems, 143:93-104, 2023. 3 +[24] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Baolin Wu, Andrew Y Ng, et al. Reading digits in natural images with unsupervised feature learning. In NIPS workshop on deep learning and unsupervised feature learning, page 4. Granada, 2011. 6 +[25] Yu Qiao, Md Shirajum Munir, Apurba Adhikary, Huy Q Le, Avi Deb Raha, Chaoning Zhang, and Choong Seon Hong. Mp-fedcl: Multi-prototype federated contrastive learning for edge intelligence. IEEE Internet of Things journal, 2023. 2 +[26] Min Tan, Yinfu Feng, Lingqiang Chu, Jingcheng Shi, Rong Xiao, Haihong Tang, and Jun Yu. Fedsea: Federated learning via selective feature alignment for non-iid multimodal data. IEEE Transactions on Multimedia, 2023. 2 +[27] Yue Tan, Guodong Long, Lu Liu, Tianyi Zhou, Qinghua Lu, Jing Jiang, and Chengqi Zhang. Fedproto: Federated prototype learning across heterogeneous clients. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8432-8440, 2022. 2, 3, 6 +[28] Yue Tan, Guodong Long, Jie Ma, Lu Liu, Tianyi Zhou, and Jing Jiang. Federated learning from pre-trained models: A contrastive learning approach. Advances in neural information processing systems, 35:19332-19344, 2022. 2, 6 +[29] Lei Wang, Jieming Bian, Letian Zhang, Chen Chen, and Jie Xu. Taming cross-domain representation variance in federated prototype learning with heterogeneous data domains. arXiv preprint arXiv:2403.09048, 2024. 2 +[30] Li Wang, Quangui Zhang, Lei Sang, Qiang Wu, and Min Xu. Federated prototype-based contrastive learning for privacy- + +preserving cross-domain recommendation. arXiv preprint arXiv:2409.03294, 2024. 2 +[31] Yue Zhao, Meng Li, Liangzhen Lai, Naveen Suda, Damon Civin, and Vikas Chandra. Federated learning with non-iid data. arXiv preprint arXiv:1806.00582, 2018. 1 +[32] Kaiyang Zhou, Yongxin Yang, Timothy Hospedales, and Tao Xiang. Learning to generate novel domains for domain generalization. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16, pages 561-578. Springer, 2020. 5 +[33] Tailin Zhou, Jun Zhang, and Danny HK Tsang. Fedfa: Federated learning with feature anchors to align features and classifiers for heterogeneous data. IEEE Transactions on Mobile Computing, 2023. 3 +[34] Zhuangdi Zhu, Junyuan Hong, and Jiayu Zhou. Data-free knowledge distillation for heterogeneous federated learning. In International conference on machine learning, pages 12878-12889. PMLR, 2021. 2 \ No newline at end of file diff --git a/2025/FedPall_ Prototype-based Adversarial and Collaborative Learning for Federated Learning with Feature Drift/images.zip b/2025/FedPall_ Prototype-based Adversarial and Collaborative Learning for Federated Learning with Feature Drift/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c65136bc9fa679fbae5c10210d8420742491df65 --- /dev/null +++ b/2025/FedPall_ Prototype-based Adversarial and Collaborative Learning for Federated Learning with Feature Drift/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7c4ccd3042471d2cdf1174df61ab797c6a7fd52529dabaaeeb921e81b6392904 +size 542801 diff --git a/2025/FedPall_ Prototype-based Adversarial and Collaborative Learning for Federated Learning with Feature Drift/layout.json b/2025/FedPall_ Prototype-based Adversarial and Collaborative Learning for Federated Learning with Feature Drift/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bd6cb5a85507202da01edf8f29d20985532344f7 --- /dev/null +++ b/2025/FedPall_ Prototype-based Adversarial and Collaborative Learning for Federated Learning with Feature Drift/layout.json @@ -0,0 +1,8700 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 58, + 102, + 553, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 102, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 58, + 102, + 553, + 140 + ], + "type": "text", + "content": "FedPall: Prototype-based Adversarial and Collaborative Learning for Federated Learning with Feature Drift" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 160, + 533, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 160, + 533, + 218 + ], + "spans": [ + { + "bbox": [ + 80, + 160, + 533, + 218 + ], + "type": "text", + "content": "Yong Zhang" + }, + { + "bbox": [ + 80, + 160, + 533, + 218 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 80, + 160, + 533, + 218 + ], + "type": "text", + "content": ", Feng Liang" + }, + { + "bbox": [ + 80, + 160, + 533, + 218 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 80, + 160, + 533, + 218 + ], + "type": "text", + "content": ", Guanghu Yuan" + }, + { + "bbox": [ + 80, + 160, + 533, + 218 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 80, + 160, + 533, + 218 + ], + "type": "text", + "content": ", Min Yang" + }, + { + "bbox": [ + 80, + 160, + 533, + 218 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 80, + 160, + 533, + 218 + ], + "type": "text", + "content": ", Chengming Li" + }, + { + "bbox": [ + 80, + 160, + 533, + 218 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 80, + 160, + 533, + 218 + ], + "type": "text", + "content": " and Xiping Hu" + }, + { + "bbox": [ + 80, + 160, + 533, + 218 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 80, + 160, + 533, + 218 + ], + "type": "text", + "content": " \nArtificial Intelligence Research Institute, Shenzhen MSU-BIT University" + }, + { + "bbox": [ + 80, + 160, + 533, + 218 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 80, + 160, + 533, + 218 + ], + "type": "text", + "content": " \nShenzhen Institutes of Advanced Technology, Chinese Academy of Sciences" + }, + { + "bbox": [ + 80, + 160, + 533, + 218 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 80, + 160, + 533, + 218 + ], + "type": "text", + "content": " \nSchool of Medical Technology, Beijing Institute of Technology" + }, + { + "bbox": [ + 80, + 160, + 533, + 218 + ], + "type": "inline_equation", + "content": "^{3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 220, + 494, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 220, + 494, + 245 + ], + "spans": [ + { + "bbox": [ + 115, + 220, + 494, + 245 + ], + "type": "text", + "content": "{zhangyong2023@bit.edu.cn, fliang@smbu.edu.cn, yangh@mail.ustc.edu.cn min.yang@siat.ac.cn, licm@smbu.edu.cn, huxp@smbu.edu.cn}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 152, + 273, + 200, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 273, + 200, + 285 + ], + "spans": [ + { + "bbox": [ + 152, + 273, + 200, + 285 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 297, + 296, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 297, + 296, + 561 + ], + "spans": [ + { + "bbox": [ + 54, + 297, + 296, + 561 + ], + "type": "text", + "content": "Federated learning (FL) enables collaborative training of a global model in the centralized server with data from multiple parties while preserving privacy. However, data heterogeneity can significantly degrade the performance of the global model when each party uses datasets from different sources to train a local model, thereby affecting personalized local models. Among various cases of data heterogeneity, feature drift, feature space difference among parties, is prevalent in real-life data but remains largely unexplored. Feature drift can distract feature extraction learning in clients and thus lead to poor feature extraction and classification performance. To tackle the problem of feature drift in FL, we propose FedPall, an FL framework that utilizes prototype-based adversarial learning to unify feature spaces and collaborative learning to reinforce class information within the features. Moreover, FedPall leverages mixed features generated from global prototypes and local features to enhance the global classifier with classification-relevant information from a global perspective. Evaluation results on three representative feature-drifted datasets demonstrate FedPall's consistently superior performance in classification with feature-drifted data in the FL scenario." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 596, + 135, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 596, + 135, + 608 + ], + "spans": [ + { + "bbox": [ + 55, + 596, + 135, + 608 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 616, + 296, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 296, + 677 + ], + "type": "text", + "content": "Today, in computer vision, researchers often utilize large amounts of data from various parties to improve the accuracy of algorithms. However, this raises concerns, such as the potential for user privacy leakage caused by sharing private data [12]. Federated learning (FL) [22] is proposed as a" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 365, + 274, + 504, + 295 + ], + "blocks": [ + { + "bbox": [ + 365, + 274, + 504, + 295 + ], + "lines": [ + { + "bbox": [ + 365, + 274, + 504, + 295 + ], + "spans": [ + { + "bbox": [ + 365, + 274, + 504, + 295 + ], + "type": "image", + "image_path": "5f28ad42f31865a0a78a661ecc226f3937acd0999dda84a27f498848b5e69682.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 322, + 299, + 405, + 360 + ], + "blocks": [ + { + "bbox": [ + 322, + 299, + 405, + 360 + ], + "lines": [ + { + "bbox": [ + 322, + 299, + 405, + 360 + ], + "spans": [ + { + "bbox": [ + 322, + 299, + 405, + 360 + ], + "type": "image", + "image_path": "4095f66beff724dc93a136d99e045de0ef20920ffa4e2a6024f239976a493e06.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 337, + 365, + 390, + 373 + ], + "lines": [ + { + "bbox": [ + 337, + 365, + 390, + 373 + ], + "spans": [ + { + "bbox": [ + 337, + 365, + 390, + 373 + ], + "type": "text", + "content": "(a) Feature drift" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 463, + 304, + 546, + 361 + ], + "blocks": [ + { + "bbox": [ + 463, + 304, + 546, + 361 + ], + "lines": [ + { + "bbox": [ + 463, + 304, + 546, + 361 + ], + "spans": [ + { + "bbox": [ + 463, + 304, + 546, + 361 + ], + "type": "image", + "image_path": "a568814547d2163997a8a18e9df0123816b80d05ec44cfd4b154313ea88cec32.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 459, + 365, + 551, + 374 + ], + "lines": [ + { + "bbox": [ + 459, + 365, + 551, + 374 + ], + "spans": [ + { + "bbox": [ + 459, + 365, + 551, + 374 + ], + "type": "text", + "content": "(b) Traditional classification" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 322, + 395, + 405, + 439 + ], + "blocks": [ + { + "bbox": [ + 322, + 395, + 405, + 439 + ], + "lines": [ + { + "bbox": [ + 322, + 395, + 405, + 439 + ], + "spans": [ + { + "bbox": [ + 322, + 395, + 405, + 439 + ], + "type": "image", + "image_path": "12024965fe707cc42954bd98dd10ed6f64b8a06dfa6494cc03f3f7b37108bb0a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 332, + 444, + 395, + 453 + ], + "lines": [ + { + "bbox": [ + 332, + 444, + 395, + 453 + ], + "spans": [ + { + "bbox": [ + 332, + 444, + 395, + 453 + ], + "type": "text", + "content": "(c) De-clientization" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 463, + 378, + 547, + 408 + ], + "blocks": [ + { + "bbox": [ + 463, + 378, + 547, + 408 + ], + "lines": [ + { + "bbox": [ + 463, + 378, + 547, + 408 + ], + "spans": [ + { + "bbox": [ + 463, + 378, + 547, + 408 + ], + "type": "image", + "image_path": "f3d3cb11ceb294bc25984d978289c61d11e553e6549ac6b8968239d7dbcccfea.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 463, + 415, + 545, + 439 + ], + "blocks": [ + { + "bbox": [ + 463, + 415, + 545, + 439 + ], + "lines": [ + { + "bbox": [ + 463, + 415, + 545, + 439 + ], + "spans": [ + { + "bbox": [ + 463, + 415, + 545, + 439 + ], + "type": "image", + "image_path": "ecfb34b8fa4ce36d7ac8a2e09bb1554982f0b7e59df31c5a84c58553b97c8d20.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 465, + 444, + 545, + 453 + ], + "lines": [ + { + "bbox": [ + 465, + 444, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 465, + 444, + 545, + 453 + ], + "type": "text", + "content": "(d) Prototype contrastive" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 463, + 555, + 496 + ], + "lines": [ + { + "bbox": [ + 313, + 463, + 555, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 555, + 496 + ], + "type": "text", + "content": "Figure 1. We show a schematic diagram of feature drift and use different techniques to drive feature distribution to update in different directions." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 521, + 555, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 521, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 312, + 521, + 555, + 581 + ], + "type": "text", + "content": "privacy-preserving distributed learning paradigm to address these challenges. In the FL paradigm, each party maintains a local client model and collaborates with others to train a global model on the server without sharing the original data, effectively protecting user privacy." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 582, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 582, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 582, + 556, + 714 + ], + "type": "text", + "content": "Particularly, the FL paradigm faces challenges from data heterogeneity [14]. Due to the limited local view of each client, data distribution discrepancies arise across clients (commonly referred to as the non-independent and identically distributed data issue, or non-independent and identically distributed(non-IID) data issue), which can increase generalization error for local models and degrade the performance of the globally aggregated model [10]. Although recent work on non-IID data in FL primarily addresses issues such as stability, client drift, and heterogeneous label distributions among clients [11, 18, 31], feature drift(i.e.," + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 683, + 148, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 683, + 148, + 693 + ], + "spans": [ + { + "bbox": [ + 66, + 683, + 148, + 693 + ], + "type": "text", + "content": "*Corresponding authors." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 56, + 693, + 294, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 693, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 693, + 294, + 712 + ], + "type": "text", + "content": "The code is available at https://github.com/DistriAI/FedPall." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3111" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 294, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 72, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 294, + 443 + ], + "type": "text", + "content": "variations in feature distributions across clients) is a prevalent yet underexplored challenge in FL. As shown in Fig. 1a, feature drift refers to the phenomenon where samples of the same class exhibit different feature distributions across different clients due to variations in data collection methods, devices, and other factors. This leads to ambiguous decision boundaries, severely impacting the classification performance of federated learning models. However, traditional classification losses (e.g., cross-entropy loss (CE loss), as shown in Fig. 1b) do not account for feature drift. As a result, the collaborative effect among different clients causes the feature space of the same-class samples to influence each other, leading to only slight or even no clearer distinction between decision boundaries for various classes within the same client. As illustrated in Fig. 1c, decentralization is a common approach to addressing feature drift, which involves ignoring inter-client differences to make the feature space of same-class samples across different clients more clustered. In feature drift scenarios, some state-of-the-art algorithms do not explore the role of loss functions but instead optimize local update algorithms to achieve decentralization. For example, FedBN [19] compresses the feature space across clients by adding batch normalization layers to local models. Some methods [1, 34] align feature spaces by sharing partial data for synthetic data augmentation. Additionally, ADCOL sends raw features to the server to update a amplifier and uses the Kullback-Leibler (KL) loss function to achieve decentralization. However, these methods have drawbacks: on the one hand, they may lead to loss of class information, and on the other hand, they pose potential risks of privacy leakage." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 445, + 294, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 445, + 294, + 611 + ], + "spans": [ + { + "bbox": [ + 57, + 445, + 294, + 611 + ], + "type": "text", + "content": "To tackle the above challenge, we propose FedPall, a novel prototype-based adversarial and collaborative learning framework for FL with feature drift. FedPall applies adversarial learning between clients and the server to unify feature spaces, as well as collaborative learning across clients to enhance global decision boundaries. Specifically, FedPall uses adversarial learning to train a feature enhancer and uses KL divergence to align heterogeneous feature spaces between clients, while using prototype contrastive loss to reinforce class information (see Fig. 1d). Finally, adversially aligned features are securely mixed with the global prototypes and uploaded to the server, where a global-view classifier is trained to enhance overall performance. Our contributions are summarized as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 618, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 55, + 618, + 294, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 294, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 294, + 700 + ], + "type": "text", + "content": "- We propose a novel FL framework, FedPall, to address the feature drift problem. FedPall introduces adversarial learning between clients and the server, and collaborative learning among clients aiming to project feature representations into a unified feature space and reinforce the intrinsic class information. This approach effectively mitigates the feature drift problem in FL settings." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 701, + 294, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 294, + 712 + ], + "type": "text", + "content": "- We develop a technical strategy that hierarchically in" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 322, + 72, + 553, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 553, + 132 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 553, + 132 + ], + "type": "text", + "content": "tegrates the global prototypes with local features to orchestrate client-server collaboration. The mixed prototype features are then used to train a global classifier, which induces the classifier to distill discriminative patterns through cross-client knowledge consolidation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 133, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 133, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 314, + 133, + 553, + 167 + ], + "type": "text", + "content": "- Empirical evaluation on three typical feature-drifted benchmarks demonstrates that our proposed method achieves state-of-the-art classification accuracy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 178, + 400, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 178, + 400, + 190 + ], + "spans": [ + { + "bbox": [ + 314, + 178, + 400, + 190 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 199, + 553, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 199, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 313, + 199, + 553, + 282 + ], + "type": "text", + "content": "In federated learning (FL), each client's limited local view exacerbates the feature drift problem. Variations in data distributions can cause the same class label to have different feature representations, leading to poor generalization of local models. Existing solutions typically focus on two main approaches: feature space alignment and contrastive prototype learning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 283, + 553, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 283, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 283, + 553, + 533 + ], + "type": "text", + "content": "Some studies address feature drift through feature representation alignment. FRAug [1] employs data augmentation to generate synthetic embeddings that capture both global and client-specific information. FedSea [26] aligns feature distributions to transform raw features into an IID format, mitigating feature drift. FedCiR [20] maximizes mutual information between representations and labels, while minimizing mutual information between client-specific representations conditioned on labels. FedHEAL [2] builds a fair aggregation target on the server side, indirectly aligning feature spaces. MOON [16] encourages local models to align with the global feature distribution through constrained updates. ADCOL [17], unlike traditional aggregation frameworks, uses adversarial learning to enforce a unified representation across clients, alleviating feature drift. However, ADCOL's weak collaboration does not address global class boundaries, and its adversarial mechanism introduces potential privacy risks. Our approach enhances global class boundaries through stronger collaborative learning, while preserving privacy via adversarial learning to mitigate inter-client feature drift." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "type": "text", + "content": "Some studies have explored prototype-driven federated learning (FL) paradigms, where prototypes compact feature embeddings, reducing communication bandwidth and preserving privacy [30]. Tan et al.[27, 28] proposed a supervised contrastive loss function using both global and local prototypes to minimize the distance between feature representations and class prototypes, addressing data heterogeneity. However, using the average feature as a class prototype may overlook intra-class variability. To address this, MP-FedCL[25] employs clustering on the client side to generate multiple prototypes per class, capturing intra-class variation and mitigating feature drift. Building on MP-FedCL, FedPLVM [29] improves local training with a two-stage clustering process between clients and the server, and incorporates an " + }, + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "type": "text", + "content": "-sparsity prototype loss to further op" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "3112" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": "timize performance. By leveraging prototypes' privacy-preserving nature, this method addresses privacy concerns while reinforcing class-specific information in feature representations. In the FedPall framework, we enhance client-server collaboration through global prototypes. The server provides global category information to the client, helping bring similar data closer and separate dissimilar data. Additionally, we use mixed features with global category information to strengthen the global classifier." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 190, + 111, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 190, + 111, + 202 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 111, + 202 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 209, + 175, + 222 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 209, + 175, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 175, + 222 + ], + "type": "text", + "content": "3.1. Problem Description" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 227, + 296, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 227, + 296, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 227, + 296, + 323 + ], + "type": "text", + "content": "We define feature drift as follows: Given a dataset " + }, + { + "bbox": [ + 55, + 227, + 296, + 323 + ], + "type": "inline_equation", + "content": "\\mathbb{D}" + }, + { + "bbox": [ + 55, + 227, + 296, + 323 + ], + "type": "text", + "content": " with features " + }, + { + "bbox": [ + 55, + 227, + 296, + 323 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 227, + 296, + 323 + ], + "type": "text", + "content": " and labels " + }, + { + "bbox": [ + 55, + 227, + 296, + 323 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 227, + 296, + 323 + ], + "type": "text", + "content": ", while the conditional distribution " + }, + { + "bbox": [ + 55, + 227, + 296, + 323 + ], + "type": "inline_equation", + "content": "P_{i}(x|y)" + }, + { + "bbox": [ + 55, + 227, + 296, + 323 + ], + "type": "text", + "content": " differs across clients, the marginal distribution " + }, + { + "bbox": [ + 55, + 227, + 296, + 323 + ], + "type": "inline_equation", + "content": "P(y)" + }, + { + "bbox": [ + 55, + 227, + 296, + 323 + ], + "type": "text", + "content": " remains the same. This means that the same label may have significantly different features across clients. For example, due to variations in environment, geographic location, and cultural differences, the structural features of houses can vary widely." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 323, + 296, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 296, + 406 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 296, + 406 + ], + "type": "text", + "content": "With feature drift in FL, our goal is to optimize each client's personalized model loss while leveraging the potential performance gains from collaborative learning across clients [33]. In the FedPall framework, there are a total of " + }, + { + "bbox": [ + 55, + 323, + 296, + 406 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 323, + 296, + 406 + ], + "type": "text", + "content": " clients, each client " + }, + { + "bbox": [ + 55, + 323, + 296, + 406 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 323, + 296, + 406 + ], + "type": "text", + "content": " has a private dataset " + }, + { + "bbox": [ + 55, + 323, + 296, + 406 + ], + "type": "inline_equation", + "content": "D_{n}" + }, + { + "bbox": [ + 55, + 323, + 296, + 406 + ], + "type": "text", + "content": ". Based on this goal, we formulate the overall optimization objective of the FedPall framework as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 95, + 411, + 295, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 411, + 295, + 444 + ], + "spans": [ + { + "bbox": [ + 95, + 411, + 295, + 444 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta_ {1}, \\theta_ {2}, \\dots , \\theta_ {N} \\in \\mathbb {R} ^ {d _ {1}}} F (\\theta) := \\frac {1}{N} \\sum_ {n = 1} ^ {N} f _ {n} \\left(\\theta_ {n}\\right), \\tag {1}", + "image_path": "d64a5e004ed5dbdb3ebc5ceb72f7fe8b30e8a797951d80477a476d7557e537cc.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 450, + 296, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 296, + 486 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 450, + 296, + 486 + ], + "type": "inline_equation", + "content": "f_{n}" + }, + { + "bbox": [ + 55, + 450, + 296, + 486 + ], + "type": "text", + "content": " represents the expected loss obtained from client " + }, + { + "bbox": [ + 55, + 450, + 296, + 486 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 450, + 296, + 486 + ], + "type": "text", + "content": " using the global model parameters under the dataset " + }, + { + "bbox": [ + 55, + 450, + 296, + 486 + ], + "type": "inline_equation", + "content": "D_{n}" + }, + { + "bbox": [ + 55, + 450, + 296, + 486 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 450, + 296, + 486 + ], + "type": "inline_equation", + "content": "\\theta_{i}" + }, + { + "bbox": [ + 55, + 450, + 296, + 486 + ], + "type": "text", + "content": " represents the local model parameters of client " + }, + { + "bbox": [ + 55, + 450, + 296, + 486 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 450, + 296, + 486 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 493, + 171, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 493, + 171, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 493, + 171, + 504 + ], + "type": "text", + "content": "3.2. FedPall Framework" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 510, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 677 + ], + "type": "text", + "content": "Existing approaches to addressing the feature drift problem in FL typically focus on either collaborative learning or adversarial learning in isolation. This can result in models that either fail to adequately capture class-related information in the feature representations or exhibit persistent discrepancies in feature distributions across clients. To address these limitations, we propose integrating both adversarial and collaborative learning to effectively mitigate feature drift in FL settings. In this section, we present the FedPall framework by elaborating on its adversarial and collaborative learning. The framework of the overall approach is shown in Fig. 2a. It is structured into four key procedures: generating global prototypes, training local models, training global model, and decentralizing global classifier." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "We define certain model symbols here that will be used in this section. The local model " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "F(\\cdot)" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " consists of two components, a feature extractor " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "G(\\cdot)" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " (e.g., Resnet50 [7] for image" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 554, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 119 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 119 + ], + "type": "text", + "content": "data) and a classifier " + }, + { + "bbox": [ + 313, + 72, + 554, + 119 + ], + "type": "inline_equation", + "content": "H(\\cdot)" + }, + { + "bbox": [ + 313, + 72, + 554, + 119 + ], + "type": "text", + "content": ". We use a multilayer perceptron (MLP) as our Amplifier, and except for the output layer, the number of nodes in other layers is consistent with that of the classifier." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 125, + 471, + 137 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 125, + 471, + 137 + ], + "spans": [ + { + "bbox": [ + 313, + 125, + 471, + 137 + ], + "type": "text", + "content": "3.2.1. Generating Global Prototypes" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 140, + 554, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 140, + 554, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 140, + 554, + 247 + ], + "type": "text", + "content": "Several studies [23, 27] suggest that category-centered prototypes are a privacy-friendly form of global knowledge. We leverage collaboration between clients to aggregate and generate global prototypes, as shown in Fig. 2b. Typically, each class prototype is defined as the mean of the feature vectors belonging to that class. The distinction between local and global prototypes lies in the scope of their underlying data. This section will describe the construction of both local and global prototypes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 247, + 554, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 554, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 554, + 270 + ], + "type": "text", + "content": "The local prototype for the " + }, + { + "bbox": [ + 313, + 247, + 554, + 270 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 247, + 554, + 270 + ], + "type": "text", + "content": "-th category on client " + }, + { + "bbox": [ + 313, + 247, + 554, + 270 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 247, + 554, + 270 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 378, + 279, + 555, + 310 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 279, + 555, + 310 + ], + "spans": [ + { + "bbox": [ + 378, + 279, + 555, + 310 + ], + "type": "interline_equation", + "content": "c _ {n} ^ {k} = \\frac {1}{N _ {n} ^ {k}} \\sum_ {(x, y) \\in D _ {n} ^ {k}} G _ {n} (x), \\tag {2}", + "image_path": "796869882f97660b06ebc43d37ffd8098fbe330851d4e72e939548da3f3db0fd.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 314, + 554, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 554, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 554, + 350 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 314, + 554, + 350 + ], + "type": "inline_equation", + "content": "D_{n}^{k}" + }, + { + "bbox": [ + 313, + 314, + 554, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 314, + 554, + 350 + ], + "type": "inline_equation", + "content": "N_{n}^{k}" + }, + { + "bbox": [ + 313, + 314, + 554, + 350 + ], + "type": "text", + "content": " represent the data samples and the number of samples for the " + }, + { + "bbox": [ + 313, + 314, + 554, + 350 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 314, + 554, + 350 + ], + "type": "text", + "content": "-th category on client " + }, + { + "bbox": [ + 313, + 314, + 554, + 350 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 314, + 554, + 350 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 350, + 554, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 554, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 554, + 374 + ], + "type": "text", + "content": "Gathering all the local prototypes together forms a local prototype set, which can be defined as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 367, + 380, + 553, + 395 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 380, + 553, + 395 + ], + "spans": [ + { + "bbox": [ + 367, + 380, + 553, + 395 + ], + "type": "interline_equation", + "content": "\\mathcal {O} _ {n} = \\left\\{c _ {n} ^ {1}, c _ {n} ^ {2}, \\dots , c _ {n} ^ {K} \\right\\} \\in \\mathbb {R} ^ {K \\times d}, \\tag {3}", + "image_path": "613e5bb4ac5698d989c4172efe27fa2b65f9b41dad1798a128cd4b04176d693e.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 400, + 554, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 554, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 554, + 423 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 400, + 554, + 423 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 400, + 554, + 423 + ], + "type": "text", + "content": " represents the number of categories owned by each client, and " + }, + { + "bbox": [ + 313, + 400, + 554, + 423 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 400, + 554, + 423 + ], + "type": "text", + "content": " denotes the output feature dimension." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 424, + 554, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 424, + 554, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 424, + 554, + 496 + ], + "type": "text", + "content": "Since we are only addressing the problem of feature drift, all clients have the same number of categories. Upon receiving the local prototype set and local label proportions " + }, + { + "bbox": [ + 313, + 424, + 554, + 496 + ], + "type": "inline_equation", + "content": "\\mathcal{N} = \\{\\{N_n^k\\}_{k=1}^K | n \\in \\mathcal{A}\\}" + }, + { + "bbox": [ + 313, + 424, + 554, + 496 + ], + "type": "text", + "content": " sent by client set " + }, + { + "bbox": [ + 313, + 424, + 554, + 496 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 424, + 554, + 496 + ], + "type": "text", + "content": ", the server integrates the local prototypes from all clients to form the global prototypes:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 388, + 502, + 553, + 531 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 502, + 553, + 531 + ], + "spans": [ + { + "bbox": [ + 388, + 502, + 553, + 531 + ], + "type": "interline_equation", + "content": "\\mathcal {G} ^ {k} = \\sum_ {n} \\frac {N _ {n} ^ {k}}{\\sum_ {n} N _ {n} ^ {k}} c _ {n} ^ {k}. \\tag {4}", + "image_path": "802143633fb6316b6a5cffc77e3f8b6c74e2f16883575fed4c909ad6fb797378.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 326, + 536, + 518, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 536, + 518, + 548 + ], + "spans": [ + { + "bbox": [ + 326, + 536, + 518, + 548 + ], + "type": "text", + "content": "The global prototypes set can be represented as" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 377, + 554, + 553, + 568 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 554, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 377, + 554, + 553, + 568 + ], + "type": "interline_equation", + "content": "\\mathcal {G} = \\left[ \\mathcal {G} ^ {1}, \\dots , \\mathcal {G} ^ {k}, \\dots , \\mathcal {G} ^ {K} \\right]. \\tag {5}", + "image_path": "f2d502d91ada149d3fd25690af38d4469da4c691e586dd79867cab73efa710c9.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 574, + 554, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 554, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 554, + 598 + ], + "type": "text", + "content": "Next, the server sends the global prototype set " + }, + { + "bbox": [ + 313, + 574, + 554, + 598 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 574, + 554, + 598 + ], + "type": "text", + "content": " to each client to guide local model training." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 603, + 440, + 614 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 603, + 440, + 614 + ], + "spans": [ + { + "bbox": [ + 313, + 603, + 440, + 614 + ], + "type": "text", + "content": "3.2.2. Training Local Models" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 617, + 554, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 664 + ], + "type": "text", + "content": "The goal of training local models is to train effective feature encoders that map raw data from different clients into a unified feature space, where feature distributions are aligned and class-related information is enhanced." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "content": "As mentioned earlier, due to feature drift, training a local classifier alone is not sufficient for accurately classifying data with feature drift. To address this, we apply adversarial learning to train a feature encoder, as illustrated" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3113" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 69, + 553, + 183 + ], + "blocks": [ + { + "bbox": [ + 59, + 69, + 553, + 183 + ], + "lines": [ + { + "bbox": [ + 59, + 69, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 59, + 69, + 553, + 183 + ], + "type": "image", + "image_path": "1b0fc8c559823d3f6863fb3f196c03b5f51fa498a103b6fdced56d8693c9af2c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 260, + 185, + 351, + 194 + ], + "lines": [ + { + "bbox": [ + 260, + 185, + 351, + 194 + ], + "spans": [ + { + "bbox": [ + 260, + 185, + 351, + 194 + ], + "type": "text", + "content": "(a) The overview of FedPall" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 60, + 198, + 555, + 333 + ], + "blocks": [ + { + "bbox": [ + 60, + 198, + 555, + 333 + ], + "lines": [ + { + "bbox": [ + 60, + 198, + 555, + 333 + ], + "spans": [ + { + "bbox": [ + 60, + 198, + 555, + 333 + ], + "type": "image", + "image_path": "050c4c4db9b7ad9a07ca628e7e6558ca4d87a136f07607029f77cc39ad050fbe.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 338, + 402, + 350 + ], + "lines": [ + { + "bbox": [ + 206, + 338, + 402, + 350 + ], + "spans": [ + { + "bbox": [ + 206, + 338, + 402, + 350 + ], + "type": "text", + "content": "Figure 2. The FedPall framework and detailed phases" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 371, + 295, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 371, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 54, + 371, + 295, + 502 + ], + "type": "text", + "content": "in Fig. 2c. Specifically, we use a global amplifier, trained on the server, which amplifies the heterogeneous information in the features from different clients. At the client's side, we apply the amplifier and use Kullback-Leibler (KL) divergence to reduce the heterogeneous information in the features, thereby creating an adversarial learning setup between the client and server. The objective is to improve the generalization ability of the feature encoders across clients while minimizing the client-specific heterogeneity in the feature representations. Let " + }, + { + "bbox": [ + 54, + 371, + 295, + 502 + ], + "type": "inline_equation", + "content": "x^{(i)}" + }, + { + "bbox": [ + 54, + 371, + 295, + 502 + ], + "type": "text", + "content": " denote the " + }, + { + "bbox": [ + 54, + 371, + 295, + 502 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 54, + 371, + 295, + 502 + ], + "type": "text", + "content": "-th dimension of vector " + }, + { + "bbox": [ + 54, + 371, + 295, + 502 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 54, + 371, + 295, + 502 + ], + "type": "text", + "content": ", the KL divergence is calculated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 60, + 513, + 294, + 581 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 513, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 60, + 513, + 294, + 581 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {K L} = \\sum_ {(\\mathbf {x}, \\mathbf {y}) \\in D _ {n}} D _ {K L} \\left(A _ {n} (G (\\mathbf {x})) | | [ \\frac {1}{N} ] ^ {N}\\right) \\\\ = \\sum_ {(\\mathbf {x}, y) \\in D _ {n}} \\sum_ {i} ^ {N} A _ {n} (G (\\mathbf {x})) ^ {(i)} \\log N A _ {n} (G (\\mathbf {x})) ^ {(i)}, \\tag {6} \\\\ \\end{array}", + "image_path": "7e95d67b05fde0a900934558db3c3e1cf951df2cf2962c66b032dc42fb20992e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 592, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 592, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 295, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 592, + 295, + 628 + ], + "type": "inline_equation", + "content": "x^{(i)}" + }, + { + "bbox": [ + 55, + 592, + 295, + 628 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 55, + 592, + 295, + 628 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 592, + 295, + 628 + ], + "type": "text", + "content": "-th dimension of vector " + }, + { + "bbox": [ + 55, + 592, + 295, + 628 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 592, + 295, + 628 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 592, + 295, + 628 + ], + "type": "inline_equation", + "content": "A_{n}(\\cdot)" + }, + { + "bbox": [ + 55, + 592, + 295, + 628 + ], + "type": "text", + "content": " represents the amplifier of the " + }, + { + "bbox": [ + 55, + 592, + 295, + 628 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 592, + 295, + 628 + ], + "type": "text", + "content": "-th client, " + }, + { + "bbox": [ + 55, + 592, + 295, + 628 + ], + "type": "inline_equation", + "content": "D_{KL}(P||Q)" + }, + { + "bbox": [ + 55, + 592, + 295, + 628 + ], + "type": "text", + "content": " represents the KL divergence of " + }, + { + "bbox": [ + 55, + 592, + 295, + 628 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 592, + 295, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 592, + 295, + 628 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 592, + 295, + 628 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": "After adversarial learning, although the feature representations of different clients are mapped to similar feature distributions, the class-related information may be blurred. To address this problem, we propose using contrastive learning to reinforce the class-related information within the feature encoder. To prevent information leakage, we enable collaboration between the global prototypes (server-side) and" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 371, + 555, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 555, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 555, + 455 + ], + "type": "text", + "content": "local features (client-side). Specifically, we employ the InfoNCE loss to minimize the distance between local features and their corresponding global prototypes, while maximizing the distance between local features and global prototypes of other classes. This approach strengthens the class-related representation within the feature encoder. The formula for the global-prototype contrastive loss is as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 476, + 553, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 476, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 315, + 476, + 553, + 514 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {i n f o N C E}} = \\sum_ {(\\mathrm {x}, \\mathrm {y}) \\in D _ {n}} - \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(G (\\mathrm {x}) , \\mathcal {G} ^ {\\mathrm {y}}\\right) / \\tau\\right)}{\\sum_ {\\mathrm {y} _ {\\alpha} \\in A (\\mathrm {y})} \\exp \\left(\\operatorname {s i m} \\left(\\mathrm {x} , \\mathcal {G} ^ {\\left(\\mathrm {y} _ {\\alpha}\\right)}\\right) / \\tau\\right)}, \\tag {7}", + "image_path": "5ce3ee3c4825ec9910508ce683c64b8fc0b2545170061b96c9f0bcaeecf34dd5.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 514, + 555, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 555, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 555, + 563 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 514, + 555, + 563 + ], + "type": "inline_equation", + "content": "A(y) \\coloneqq \\{y_{\\alpha} \\in [1, |\\mathcal{G}|] : y_{\\alpha} \\neq y\\}" + }, + { + "bbox": [ + 313, + 514, + 555, + 563 + ], + "type": "text", + "content": " is the set of labels distinct from " + }, + { + "bbox": [ + 313, + 514, + 555, + 563 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 514, + 555, + 563 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 514, + 555, + 563 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 514, + 555, + 563 + ], + "type": "text", + "content": " is the temperature to adjust the tolerance for feature difference, and " + }, + { + "bbox": [ + 313, + 514, + 555, + 563 + ], + "type": "inline_equation", + "content": "sim(x, y)" + }, + { + "bbox": [ + 313, + 514, + 555, + 563 + ], + "type": "text", + "content": " represents the cosine similarity of " + }, + { + "bbox": [ + 313, + 514, + 555, + 563 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 514, + 555, + 563 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 514, + 555, + 563 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 514, + 555, + 563 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 564, + 554, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 564, + 554, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 564, + 554, + 635 + ], + "type": "text", + "content": "We combine adversarial learning and collaborative learning to address the feature drift problem in FL. By leveraging the two loss functions defined above, along with the local cross-entropy loss, we progressively train the local feature encoder at each client. The overall loss function for each client is as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 319, + 650, + 553, + 663 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 650, + 553, + 663 + ], + "spans": [ + { + "bbox": [ + 319, + 650, + 553, + 663 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {C E (\\mathrm {x}, \\mathrm {y}) \\sim D _ {n}} (F (\\mathrm {x}), \\mathrm {y}) + \\mu \\mathcal {L} _ {K L} + \\delta \\mathcal {L} _ {\\text {i n f o N C E}}, \\tag {8}", + "image_path": "65e0e31e966c15e58a600d9f087897e47532db3317f7e85fdd2fb5d12fcd561e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 675, + 553, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 675, + 553, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 675, + 553, + 700 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 675, + 553, + 700 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 675, + 553, + 700 + ], + "type": "text", + "content": " denotes the weight of the " + }, + { + "bbox": [ + 313, + 675, + 553, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{KL}" + }, + { + "bbox": [ + 313, + 675, + 553, + 700 + ], + "type": "text", + "content": " divergence and " + }, + { + "bbox": [ + 313, + 675, + 553, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CE}" + }, + { + "bbox": [ + 313, + 675, + 553, + 700 + ], + "type": "text", + "content": " denotes the cross-entropy loss." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 325, + 701, + 544, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 544, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 544, + 714 + ], + "type": "text", + "content": "Finally, the local model " + }, + { + "bbox": [ + 325, + 701, + 544, + 714 + ], + "type": "inline_equation", + "content": "F(\\cdot)" + }, + { + "bbox": [ + 325, + 701, + 544, + 714 + ], + "type": "text", + "content": " is updated using Eq. (8)." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3114" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 182, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 182, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 182, + 83 + ], + "type": "text", + "content": "3.2.3. Training Global Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 87, + 296, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 87, + 296, + 182 + ], + "spans": [ + { + "bbox": [ + 55, + 87, + 296, + 182 + ], + "type": "text", + "content": "Due to the client's limited local view, training an accurate classifier is challenging. To address this issue, encrypted mixed features are uploaded to the server, enabling classifier training with a global perspective. Additionally, the server leverages its global view to train an amplifier, which engages in adversarial learning with the client to enhance model robustness. Fig. 2d illustrates this procedure, and the detailed process is described below." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 182, + 296, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 182, + 296, + 219 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 296, + 219 + ], + "type": "text", + "content": "For each feature " + }, + { + "bbox": [ + 55, + 182, + 296, + 219 + ], + "type": "inline_equation", + "content": "z_{n}^{k}" + }, + { + "bbox": [ + 55, + 182, + 296, + 219 + ], + "type": "text", + "content": " of class " + }, + { + "bbox": [ + 55, + 182, + 296, + 219 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 182, + 296, + 219 + ], + "type": "text", + "content": " on client " + }, + { + "bbox": [ + 55, + 182, + 296, + 219 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 182, + 296, + 219 + ], + "type": "text", + "content": ", we obtain a prototype mixed feature by performing a weighted fusion with the corresponding global prototype:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 227, + 296, + 243 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 227, + 296, + 243 + ], + "spans": [ + { + "bbox": [ + 112, + 227, + 296, + 243 + ], + "type": "interline_equation", + "content": "r _ {n} ^ {k} = \\alpha \\times z _ {n} ^ {k} + (1 - \\alpha) \\times \\mathcal {G} ^ {k}, \\tag {9}", + "image_path": "47ccb0b64d004b40e7cd1848cc8a1bae63b1befaa96c2209bcfa83ceac9c9e5c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 249, + 296, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 249, + 296, + 298 + ], + "spans": [ + { + "bbox": [ + 55, + 249, + 296, + 298 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 249, + 296, + 298 + ], + "type": "inline_equation", + "content": "z_{n}^{k}\\in Z_{n}^{k} = \\{z_{n}^{i,k}\\}_{i = 1}^{S}" + }, + { + "bbox": [ + 55, + 249, + 296, + 298 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 249, + 296, + 298 + ], + "type": "inline_equation", + "content": "\\alpha \\sim U(u_f,u_r)" + }, + { + "bbox": [ + 55, + 249, + 296, + 298 + ], + "type": "text", + "content": " are the mix parameters, with " + }, + { + "bbox": [ + 55, + 249, + 296, + 298 + ], + "type": "inline_equation", + "content": "u_{f}" + }, + { + "bbox": [ + 55, + 249, + 296, + 298 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 249, + 296, + 298 + ], + "type": "inline_equation", + "content": "u_{d}" + }, + { + "bbox": [ + 55, + 249, + 296, + 298 + ], + "type": "text", + "content": " representing two hyperparameters of a uniform distribution. " + }, + { + "bbox": [ + 55, + 249, + 296, + 298 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 249, + 296, + 298 + ], + "type": "text", + "content": " represents the number of samples of the " + }, + { + "bbox": [ + 55, + 249, + 296, + 298 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 249, + 296, + 298 + ], + "type": "text", + "content": "-th category in client " + }, + { + "bbox": [ + 55, + 249, + 296, + 298 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 249, + 296, + 298 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 298, + 296, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 298, + 296, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 298, + 296, + 323 + ], + "type": "text", + "content": "To further reduce the risk of privacy leakage, we employ a Bernoulli mask, represented as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 97, + 331, + 294, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 331, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 97, + 331, + 294, + 360 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} M a s k = \\left\\{X _ {1}, X _ {2}, \\dots , X _ {d} \\right\\}, \\tag {10} \\\\ X _ {i} \\sim \\operatorname {B e r n o u l l i} (\\beta), \\quad \\forall i \\in [ 1, d ]. \\\\ \\end{array}", + "image_path": "2c43aaca3f20e4507b11a1d6d7c4e9d7d198639bb9f68c3c0455f80e4940139d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 368, + 296, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 368, + 296, + 402 + ], + "spans": [ + { + "bbox": [ + 55, + 368, + 296, + 402 + ], + "type": "text", + "content": "The final output of the prototype mixing mechanism is derived by selecting the mixed feature elements based on the noise mask:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 136, + 402, + 294, + 417 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 402, + 294, + 417 + ], + "spans": [ + { + "bbox": [ + 136, + 402, + 294, + 417 + ], + "type": "interline_equation", + "content": "\\tilde {r} _ {n} ^ {k} = \\operatorname {M a s k} \\odot r _ {n} ^ {k}, \\tag {11}", + "image_path": "3728ecb365973c78b9c4432f14e8a85e88aad237b321d189971c10be1d399f47.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 422, + 227, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 227, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 227, + 433 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 422, + 227, + 433 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 422, + 227, + 433 + ], + "type": "text", + "content": " is an element-wise and operator." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 434, + 296, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 296, + 481 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 296, + 481 + ], + "type": "text", + "content": "After generating the prototype mixed features, the client will form the set " + }, + { + "bbox": [ + 55, + 434, + 296, + 481 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{R_L}(R,Y)" + }, + { + "bbox": [ + 55, + 434, + 296, + 481 + ], + "type": "text", + "content": " using the prototype mixed feature set " + }, + { + "bbox": [ + 55, + 434, + 296, + 481 + ], + "type": "inline_equation", + "content": "R_{n} = \\{\\tilde{r}_{n}^{1},\\dots,\\tilde{r}_{n}^{k},\\dots,\\tilde{r}_{n}^{K}\\} \\in \\mathbb{R}^{K\\times d}" + }, + { + "bbox": [ + 55, + 434, + 296, + 481 + ], + "type": "text", + "content": " and the corresponding label set " + }, + { + "bbox": [ + 55, + 434, + 296, + 481 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 55, + 434, + 296, + 481 + ], + "type": "text", + "content": " , which will be sent to the server." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 481, + 296, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 481, + 296, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 481, + 296, + 553 + ], + "type": "text", + "content": "The server updates the global amplifier " + }, + { + "bbox": [ + 55, + 481, + 296, + 553 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 481, + 296, + 553 + ], + "type": "text", + "content": " using the mixed feature sets from each client along with the corresponding client IDs. Specifically, we first construct the dataset for training the amplifier, denoted as " + }, + { + "bbox": [ + 55, + 481, + 296, + 553 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{R_I}(R,I)" + }, + { + "bbox": [ + 55, + 481, + 296, + 553 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 481, + 296, + 553 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 481, + 296, + 553 + ], + "type": "text", + "content": " represents the client IDs. The amplifier is then updated by minimizing the empirical risk:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 127, + 563, + 294, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 563, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 127, + 563, + 294, + 578 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {(R, I) \\sim \\mathcal {D} _ {R _ {I}}} \\ell_ {C E} (R, I). \\tag {12}", + "image_path": "a630f7ecb7c3ff98b1f7b042e917a7ae294916cd6aa723b66a5b0063130566c7.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 585, + 296, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 585, + 296, + 632 + ], + "spans": [ + { + "bbox": [ + 55, + 585, + 296, + 632 + ], + "type": "text", + "content": "At the same time, the server updates the global classifier " + }, + { + "bbox": [ + 55, + 585, + 296, + 632 + ], + "type": "inline_equation", + "content": "C_g" + }, + { + "bbox": [ + 55, + 585, + 296, + 632 + ], + "type": "text", + "content": " using the mixed prototype features and the class labels from the clients " + }, + { + "bbox": [ + 55, + 585, + 296, + 632 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 55, + 585, + 296, + 632 + ], + "type": "text", + "content": ". This is done by minimizing the empirical risk:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 123, + 632, + 294, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 632, + 294, + 647 + ], + "spans": [ + { + "bbox": [ + 123, + 632, + 294, + 647 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {(R, Y) \\sim \\mathcal {D} _ {R _ {L}}} \\ell_ {C E} (R, Y). \\tag {13}", + "image_path": "affe5b98e64c0cda163295ee5606e45554dd216696f801092b208024b19e36ae.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 651, + 220, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 220, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 220, + 662 + ], + "type": "text", + "content": "3.2.4. Decentralizing Global Classifier" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "Finally, we deploy the global classifier " + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "inline_equation", + "content": "C_g" + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": " to each client to replace the original local classifiers " + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "inline_equation", + "content": "C_c" + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": ", as shown in Fig. 2e. The purpose of this is to obtain a more generalized classifier that can alleviate the feature drift problem. To allow" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": "the global classifier to adapt to the personalized characteristics of local data, we retrain it on the client's local data, thereby enhancing the classifier's accuracy and improving its performance on individual client data distributions." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 129, + 384, + 141 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 129, + 384, + 141 + ], + "spans": [ + { + "bbox": [ + 313, + 129, + 384, + 141 + ], + "type": "text", + "content": "4. Discussion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 149, + 555, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 149, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 149, + 555, + 316 + ], + "type": "text", + "content": "Computational Cost Compared to the standard federated learning model, our adversarial collaborative learning approach introduces an amplifier and a global classifier. However, the number of parameters for these two components is much smaller than those of the other components. In our design, both the amplifier and the classifier are designed as three-layer MLPs. Compared to the feature extractor (with a total of 23.5M parameters), the classifier (with a total of 1.32M parameters) and the amplifier (with a total of 1.33M parameters) account for approximately " + }, + { + "bbox": [ + 313, + 149, + 555, + 316 + ], + "type": "inline_equation", + "content": "5.61\\%" + }, + { + "bbox": [ + 313, + 149, + 555, + 316 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 149, + 555, + 316 + ], + "type": "inline_equation", + "content": "5.59\\%" + }, + { + "bbox": [ + 313, + 149, + 555, + 316 + ], + "type": "text", + "content": ", respectively. Moreover, the client-side amplifier remains frozen, functioning exclusively in the forward pass for loss calculation while being disabled during backpropagation cycles." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 331, + 555, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 555, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 555, + 486 + ], + "type": "text", + "content": "Communication Efficiency Unlike traditional federated learning methods such as FedAvg, which transmit full local model parameters, our approach uses prototype-mixed features as the communication medium. Assuming consistent tensor representations between model parameters and prototype features, each client employs a ResNet-50-based feature extractor that generates 2048-dimensional embeddings. During upload, clients transmit an average of 12,122 prototype-mixed features, compared to approximately 24.8 million model parameters in FedAvg. For downloading, only the amplifier and global classifier parameters are transmitted, reducing communication cost to about " + }, + { + "bbox": [ + 313, + 331, + 555, + 486 + ], + "type": "inline_equation", + "content": "10.6\\%" + }, + { + "bbox": [ + 313, + 331, + 555, + 486 + ], + "type": "text", + "content": " of that in FedAvg (see previous paragraph)." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 500, + 554, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 500, + 554, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 500, + 554, + 643 + ], + "type": "text", + "content": "Limitation While our framework currently specializes in image recognition tasks, its extension to NLP or time-series analysis remains unexplored. Successful cross-domain adaptation requires two key developments: (1) establishing domain-specific feature representations and prototype definitions, and (2) redesigning loss functions according to task semantics. For NLP applications, this implies reconfiguring the standard classification paradigm into autoregressive prediction frameworks. Architectural adaptations are equally crucial - particularly the incorporation of RNN-based structures with inherent temporal modeling capabilities for sequential data processing." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 652, + 395, + 666 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 652, + 395, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 652, + 395, + 666 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 672, + 430, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 430, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 430, + 685 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "Datasets We conduct experiments on three publicly available feature drift datasets: Digits [32], Office-10 [6], and" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3115" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 123, + 553, + 281 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "text", + "content": "Table 1. The top-1 accuracy (\\%) of each algorithm on each sub-dataset of the Office-10, Digits, and PACS datasets is compared, along with the average top-1 accuracy across all sub-datasets. The mean and standard deviation (std) from three random trials (using different random seeds, with other experimental settings remaining the same) are reported. The highest accuracy for each dataset is highlighted in bold, and the second-highest accuracy is underlined." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 123, + 553, + 281 + ], + "lines": [ + { + "bbox": [ + 60, + 123, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 60, + 123, + 553, + 281 + ], + "type": "table", + "html": "
SingleSetFedAvgFedProxPerfedAvgFedRepFedBNMOONFedProtoADCOLRUCRFedHEALours(FedPall)
Office-10amazon74.0(2.7)56.9(2.5)56.6(2.6)57.1(2.2)45.3(1.9)40.8(15.8)51.7(16.1)69.4(2.1)73.3(4.4)52.1(8.5)65.1(3.3)72.9(1.4)
caltech44.7(3.2)46.5(4.6)51.0(5.2)50.8(1.6)38.4(4.9)33.9(6.5)41.3(13.6)39.4(6.3)37.2(1.7)44.3(1.0)44.6(3.2)44.7(8.7)
dslr60.2(6.7)30.1(4.9)33.3(10.4)31.2(4.9)34.4(4.9)38.7(3.2)24.7(1.9)65.6(4.9)76.3(4.9)30.1(6.7)67.7(1.9)77.4(3.2)
webcam71.3(2.6)37.9(6.2)43.7(7.2)47.1(7.8)55.8(2.6)30.5(6.1)33.3(12.7)71.3(4.3)71.3(2.6)37.4(5.0)60.9(1.0)74.7(1.0)
avg62.5(0.4)42.9(1.2)46.1(2.6)46.6(2.9)43.5(1.3)36.0(6.5)37.8(10.9)61.4(1.7)64.5(1.8)41.0(0.6)59.6(0.5)67.5(2.7)
DigitsMNIST95.5(0.2)92.9(2.2)91.8(3.0)90.1(4.8)86.5(6.1)96.7(0.1)93.4(1.1)96.4(0.5)96.3(0.4)92.6(2.0)93.6(0.6)97.2(0.4)
SVHN71.1(0.9)77.4(0.2)76.9(0.3)75.6(0.4)67.2(1.7)79.4(0.3)79.6(0.8)72.5(0.3)75.1(2.1)77.9(0.3)68.3(2.0)78.0(0.4)
USPS86.4(0.3)89.3(0.9)89.2(1.4)88.7(0.7)90.0(3.0)90.1(0.5)81.8(0.7)87.0(0.8)86.7(1.3)88.9(2.4)87.0(0.5)87.3(1.3)
SynthDigits95.2(0.1)95.5(0.1)95.4(0.1)95.0(0.2)94.2(0.8)95.6(0.1)96.6(0.2)95.3(0.6)96.4(0.3)96.0(0.2)89.3(1.6)95.3(0.4)
MNIST-M76.6(0.4)73.8(1.5)74.0(1.5)73.2(0.8)69.1(0.9)76.3(0.4)72.2(0.9)78.3(1.2)78.3(4.4)72.7(0.4)67.8(1.9)85.9(1.4)
avg84.9(0.1)85.8(0.9)85.5(1.1)84.5(1.3)81.4(2.5)87.6(0.1)84.7(0.6)85.9(0.2)86.6(1.3)85.6(0.9)81.2(1.3)88.7(0.2)
PACSart-painting33.6(0.8)25.8(1.9)24.3(4.1)26.5(2.2)26.9(3.3)36.7(1.8)30.6(2.0)32.7(0.7)34.9(1.2)24.7(1.1)31.2(1.2)35.6(0.6)
cartoon58.5(2.5)45.4(2.3)51.4(0.6)48.3(1.2)44.4(2.1)55.6(2.0)51.5(1.8)57.3(1.5)57.2(0.8)47.5(3.3)50.8(0.4)59.7(2.3)
photo63.0(1.9)48.7(3.1)49.6(2.0)46.9(2.6)41.9(2.8)66.1(1.0)53.0(3.3)64.0(1.3)62.1(2.0)47.5(6.2)61.1(2.0)64.7(1.3)
sketch79.7(0.1)49.0(2.0)40.7(1.5)44.4(3.8)40.5(1.3)79.6(1.7)55.1(1.4)79.6(0.8)80.1(1.0)42.2(2.4)73.8(0.1)82.2(0.7)
avg58.7(1.2)42.2(1.6)41.5(1.8)41.5(1.9)38.4(1.1)59.5(1.4)47.6(0.9)58.4(0.3)58.6(0.6)40.5(2.0)54.2(0.5)60.6(0.4)
", + "image_path": "79da3f786bfa25206e0967e2bc272ba425c40a6eaf69aa58c0f032b4acecbe52.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 300, + 294, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 300, + 294, + 479 + ], + "spans": [ + { + "bbox": [ + 54, + 300, + 294, + 479 + ], + "type": "text", + "content": "PACS [15]. Specifically, (1) the Digits dataset consists of five different domain sources: MNIST [13], SVHN [24], USPS [9], SynthDigits [5], and MNIST-M [5]; (2) the Office-10 dataset includes four distinct sources: Amazon, Caltech, DSLR, and WebCam; (3) the PACS dataset consists of four sources: Art Painting, Cartoon, Photo, and Sketch. Datasets Office-10 and PACS are real-world images from natural scenes, which inherently exhibit feature drift due to the diversity of their sources. Digits is a digit recognition dataset. In line with [19, 28], we do not use the entire Digits dataset for feature transformation experiments but rather a subset of " + }, + { + "bbox": [ + 54, + 300, + 294, + 479 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 54, + 300, + 294, + 479 + ], + "type": "text", + "content": " of the data. For datasets Office-10 and PACS, we used all of the datasets for the experiment. Additionally, we split each dataset into training and testing sets with an 8:2 ratio." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 495, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 495, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 54, + 495, + 295, + 628 + ], + "type": "text", + "content": "Baselines We compare FedPall with ten baselines, including SingSet (where each client independently trains a model). FedAvg [22] is the most classic federated learning algorithm, while FedProx [18], PerFedAvg [4], and FedRep [3] are personalized federated learning methods. FedBN [19], ADCOL [17], MOON [16], FedProto [27] and FedHEAL [2] are personalized federated learning algorithms for cross-domain learning, all of which address the issue of non-IID features to some extent. In addition, we explored the ability of RUCR [8] to solve the feature drift problem." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 641, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 641, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 641, + 295, + 715 + ], + "type": "text", + "content": "Model and Hyper-parameter Setup All algorithms adopt identical local model architectures for fair comparison. Each local model contains: (1) a ResNet-50 feature extractor (excluding classifier layer), (2) a three-layer MLP classifier with 512 hidden units, and (3) a three-layer MLP amplifier with 2048 input/512 hidden units. Output dimen" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 300, + 555, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 555, + 431 + ], + "type": "text", + "content": "sions for classifier and amplifier are set according to dataset categories and data sources, respectively. We maintain client count equal to data sources (one source per client). Local training uses 5 epochs for Digits, and 10 epochs for Office-10 and PACS. Global training employs 100 epochs throughout. Optimization uses SGD (lr=0.01). Except for the digits dataset, for which the values of " + }, + { + "bbox": [ + 313, + 300, + 555, + 431 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 300, + 555, + 431 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 300, + 555, + 431 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 300, + 555, + 431 + ], + "type": "text", + "content": " are set to 0.7 and 0.3 respectively, the values of " + }, + { + "bbox": [ + 313, + 300, + 555, + 431 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 300, + 555, + 431 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 300, + 555, + 431 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 300, + 555, + 431 + ], + "type": "text", + "content": " are set to 0.1 for all other datasets. Common loss hyperparameters remain fixed across algorithms. All experiments run on NVIDIA RTX 4090 GPUs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 443, + 443, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 443, + 443, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 443, + 455 + ], + "type": "text", + "content": "5.2. Experimental Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 461, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 555, + 533 + ], + "type": "text", + "content": "We conduct the evaluation on three publicly available feature-drifted datasets (Digits, Office-10, and PACS) and compare the performance of the FedPall framework with classical and state-of-the-art baselines. As shown in Table 1, our proposed framework achieves state-of-the-art accuracy on all three datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "content": "We first discuss the experimental results based on each individual dataset. On the Office-10 dataset, the overall accuracy of the FedPall framework surpasses that of the second-best method, ADCOL, by approximately 3 percentage points. On the Digits dataset, it is evident that FedPall outperforms all other models, achieving an accuracy that is approximately 1.1 percentage points higher than the second-best model, FedBN. The Digits dataset contains images that are relatively easy to classify, and the degree of feature drift is smaller compared to the Office-10 dataset. All baseline models achieve reasonably good accuracy on this dataset. Specifically, adversarial learning helps mitigate the heterogeneous information in the MNIST-M client. Similarly, our algorithm demonstrates strong performance on the PACS dataset, achieving an overall ac" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3116" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 76, + 282, + 198 + ], + "blocks": [ + { + "bbox": [ + 69, + 76, + 282, + 198 + ], + "lines": [ + { + "bbox": [ + 69, + 76, + 282, + 198 + ], + "spans": [ + { + "bbox": [ + 69, + 76, + 282, + 198 + ], + "type": "image", + "image_path": "77b6419042e50b6d2dcb0ac1b2b488f4d56d28a11746740e997f93384ae3d09a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 209, + 295, + 232 + ], + "lines": [ + { + "bbox": [ + 55, + 209, + 295, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 295, + 232 + ], + "type": "text", + "content": "Figure 3. We evaluate the top-1 accuracy averaged over all clients using different loss function combinations on different datasets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 255, + 296, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 296, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 296, + 301 + ], + "type": "text", + "content": "curacy that is approximately 1.1 percentage points higher than the second-highest result produced by FedBN. FedPall achieves the highest or second-highest accuracy across all sub-datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 303, + 296, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 303, + 296, + 494 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 296, + 494 + ], + "type": "text", + "content": "We also discuss the performance of FedPall as compared to other state-of-the-art baselines across all three datasets. The average accuracy of FedPall consistently outperforms that of ADCOL in all three datasets, achieving an increase ranging from about 1.1 to 2.9 percentage points. In addition, even though FedBN can achieve accuracy comparable to our method on datasets Digits and PACS, our method outperforms it significantly by 31.5 percentage points on datasets Office-10. As mentioned earlier, the Office-10 dataset comes from real-world data, where feature drift is particularly prominent, and there is also a significant distribution difference between the training and testing sets, leading to the suboptimal performance of the FedBN method on this dataset. In contrast, the special design incorporating both adversarial and collaborative learning in FedPall enables it to adapt well to the Office-10 dataset." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 503, + 149, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 149, + 516 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 149, + 516 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 521, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 616 + ], + "type": "text", + "content": "Effect of loss combination In this section, we analyze the impact of KL loss and InfoNCE loss on the performance of the local feature encoder. We conduct ablation studies with three configurations: (1) removing both KL and InfoNCE losses, (2) removing only KL loss, and (3) removing only InfoNCE loss. As shown in Fig 3, the algorithm performs best when all losses are retained, which validates the reliability of the loss combination we designed." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "content": "Specifically, on the Office-10 dataset, our method outperforms the model trained with only CE loss by nearly 4 percentage points. Notably, combining CE loss with InfoNCE loss yields worse results than using CE loss alone, suggesting that in the presence of severe feature drift, reinforcing category information through InfoNCE may amplify the drift. While the " + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{CE} + \\mathrm{KL}" + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "content": " loss combination performs better than CE loss alone on Office-10, it un" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 72, + 555, + 232 + ], + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 232 + ], + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 232 + ], + "type": "image", + "image_path": "b85d870e4e032c96afd52993b09d62786b0c64350e2dd81af09bf89efbd5b840.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 239, + 555, + 274 + ], + "lines": [ + { + "bbox": [ + 313, + 239, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 555, + 274 + ], + "type": "text", + "content": "Figure 4. We plotted the feature distribution of different categories under different clients, corresponding to the four loss combination strategies of Fig. 3" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 297, + 555, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 555, + 344 + ], + "type": "text", + "content": "derperforms on the PACS dataset—even falling below CE loss—indicating reduced robustness. This suggests that CE + KL may compromise category information, leading to instability across heterogeneous datasets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 345, + 556, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 345, + 556, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 556, + 548 + ], + "type": "text", + "content": "To assess the impact of KL and InfoNCE losses on feature distributions, we visualize class-wise features across clients in the Office-10 dataset using t-SNE projections of randomly sampled data points (Fig 4). The CE-only model poorly mitigates feature drift, with Client 1 showing unclear decision boundaries. Adding InfoNCE improves intra-client class separation but fails to resolve inter-client drift, leading to ambiguous global boundaries. The " + }, + { + "bbox": [ + 313, + 345, + 556, + 548 + ], + "type": "inline_equation", + "content": "\\mathrm{CE + KL}" + }, + { + "bbox": [ + 313, + 345, + 556, + 548 + ], + "type": "text", + "content": " combination reduces cross-client distances for the same class, yielding clearer global boundaries; however, it compresses intra-class spacing in Client 1, causing overlapping clusters and outliers that hinder local classification. In contrast, our unified loss balances these effects: KL aligns same-class features across clients, while CE and InfoNCE promote intra-client separation. This coordination produces compact, well-separated clusters, improving classification and validating our method's effectiveness." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 550, + 555, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 555, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 555, + 597 + ], + "type": "text", + "content": "In summary, our method outperforms other loss combinations in the simple handwritten digit recognition task and achieves superior results on real-world datasets, demonstrating strong robustness and generalization." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": "Hyperparameter sensitivity analysis. We conducted a hyperparameter sensitivity analysis on the Office-10 dataset by varying the loss weights " + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\{0.1, 0.2, 0.5, 1.0\\}" + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": ". To visualize their impact, we generated an interpolated heatmap of average accuracy. As shown in Fig 5, accuracy remains high and stable when " + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": " is within [0.1, 0.4], with " + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": " having minimal influence in this range. Notably, the highest accuracy of " + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "inline_equation", + "content": "69.12\\%" + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": " occurs at " + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mu = \\delta = 0.2" + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": ", fol" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3117" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 90, + 77, + 261, + 210 + ], + "blocks": [ + { + "bbox": [ + 90, + 77, + 261, + 210 + ], + "lines": [ + { + "bbox": [ + 90, + 77, + 261, + 210 + ], + "spans": [ + { + "bbox": [ + 90, + 77, + 261, + 210 + ], + "type": "image", + "image_path": "05d10ac64c159e6d145090d86683797efc2e9bcf3b718101ee4d294aa85c670c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 226, + 295, + 249 + ], + "lines": [ + { + "bbox": [ + 55, + 226, + 295, + 249 + ], + "spans": [ + { + "bbox": [ + 55, + 226, + 295, + 249 + ], + "type": "text", + "content": "Figure 5. The top-1 accuracy " + }, + { + "bbox": [ + 55, + 226, + 295, + 249 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 55, + 226, + 295, + 249 + ], + "type": "text", + "content": " of different " + }, + { + "bbox": [ + 55, + 226, + 295, + 249 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 226, + 295, + 249 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 226, + 295, + 249 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 226, + 295, + 249 + ], + "type": "text", + "content": " under Office-10 dataset" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 271, + 211, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 271, + 211, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 211, + 283 + ], + "type": "text", + "content": "lowed by " + }, + { + "bbox": [ + 55, + 271, + 211, + 283 + ], + "type": "inline_equation", + "content": "68.62\\%" + }, + { + "bbox": [ + 55, + 271, + 211, + 283 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 55, + 271, + 211, + 283 + ], + "type": "inline_equation", + "content": "\\mu = 1.0, \\delta = 0.2" + }, + { + "bbox": [ + 55, + 271, + 211, + 283 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 284, + 295, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 284, + 295, + 321 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 295, + 321 + ], + "type": "text", + "content": "The results show that our algorithm is not sensitive to the parameter changes of the loss function within a certain range and can maintain high performance." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 91, + 349, + 260, + 445 + ], + "blocks": [ + { + "bbox": [ + 91, + 349, + 260, + 445 + ], + "lines": [ + { + "bbox": [ + 91, + 349, + 260, + 445 + ], + "spans": [ + { + "bbox": [ + 91, + 349, + 260, + 445 + ], + "type": "image", + "image_path": "3a9f74a729a28a6e768f7469e5d9d6579bc76f63d51b39e7f2153eaac939eebe.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 446, + 205, + 455 + ], + "lines": [ + { + "bbox": [ + 146, + 446, + 205, + 455 + ], + "spans": [ + { + "bbox": [ + 146, + 446, + 205, + 455 + ], + "type": "text", + "content": "(a) Digits Dataset" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 56, + 456, + 175, + 524 + ], + "blocks": [ + { + "bbox": [ + 56, + 456, + 175, + 524 + ], + "lines": [ + { + "bbox": [ + 56, + 456, + 175, + 524 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 175, + 524 + ], + "type": "image", + "image_path": "75462e1586aa30bfad5a6b7e7aff6a76d67e27e61d3530527a077aa6a3ec6057.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 81, + 526, + 150, + 535 + ], + "lines": [ + { + "bbox": [ + 81, + 526, + 150, + 535 + ], + "spans": [ + { + "bbox": [ + 81, + 526, + 150, + 535 + ], + "type": "text", + "content": "(b) Office-10 Dataset" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 177, + 456, + 296, + 524 + ], + "blocks": [ + { + "bbox": [ + 177, + 456, + 296, + 524 + ], + "lines": [ + { + "bbox": [ + 177, + 456, + 296, + 524 + ], + "spans": [ + { + "bbox": [ + 177, + 456, + 296, + 524 + ], + "type": "image", + "image_path": "164f4c0a75ba8067b0b170a0ae6862c554347873696d46804b64989d9450b530.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 207, + 526, + 264, + 535 + ], + "lines": [ + { + "bbox": [ + 207, + 526, + 264, + 535 + ], + "spans": [ + { + "bbox": [ + 207, + 526, + 264, + 535 + ], + "type": "text", + "content": "(c) PACS Dataset" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 545, + 295, + 567 + ], + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 567 + ], + "type": "text", + "content": "Figure 6. Comparison of accuracy with and without training the global classifier on the three datasets." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "Comparison of different classifier replacement methods We evaluate the effectiveness of the global classifier through an ablation study by removing it. As shown in Fig 6, although the simplicity of the Digits dataset results in slightly lower accuracy for the global classifier on some sub-datasets (Fig. 6a), it still outperforms the baseline without a global classifier across multiple sub-datasets. Notably, our method achieves 3.6 percentage points higher accuracy on MNIST-M. The benefits are more pronounced on datasets with substantial feature drift: on Office-10 (Fig. 6b), the global classifier surpasses the baseline on" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "nearly all sub-datasets, achieving " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "49.3\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": " on Caltech—a 15.6 percentage points improvement. On PACS (Fig. 6c), it consistently outperforms the baseline across all sub-datasets, with gains of up to 3 percentage points." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 120, + 553, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 120, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 120, + 553, + 168 + ], + "type": "text", + "content": "These results confirm the necessity of FedPall's global classifier, which captures cross-client category information to enhance client-server collaboration and improve the framework's generalization against feature drift." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 182, + 555, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 555, + 373 + ], + "type": "text", + "content": "Privacy Leakage Risk We evaluate the privacy risk of prototype mixture features using the Data Efficient Mutual Information Neural Estimator (DEMINE) [21]. On the Office-10 training set, the mutual information (MI) scores for: (1) standard Gaussian noise, (2) prototype mixture only, (3) Bernoulli masking only, and (4) our method (prototype mixture + Bernoulli masking) are 2.96, 3.06, 3.04, and 2.10, respectively (lower is better). The corresponding average accuracies are " + }, + { + "bbox": [ + 313, + 182, + 555, + 373 + ], + "type": "inline_equation", + "content": "66.4\\%" + }, + { + "bbox": [ + 313, + 182, + 555, + 373 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 182, + 555, + 373 + ], + "type": "inline_equation", + "content": "64.5\\%" + }, + { + "bbox": [ + 313, + 182, + 555, + 373 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 182, + 555, + 373 + ], + "type": "inline_equation", + "content": "63.7\\%" + }, + { + "bbox": [ + 313, + 182, + 555, + 373 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 182, + 555, + 373 + ], + "type": "inline_equation", + "content": "67.6\\%" + }, + { + "bbox": [ + 313, + 182, + 555, + 373 + ], + "type": "text", + "content": ", compared to " + }, + { + "bbox": [ + 313, + 182, + 555, + 373 + ], + "type": "inline_equation", + "content": "65.8\\%" + }, + { + "bbox": [ + 313, + 182, + 555, + 373 + ], + "type": "text", + "content": " in the noise-free case. Our approach not only offers stronger privacy protection but also improves accuracy, attributed to the consistent update direction between the global prototype and feature encoder. Additionally, since encryption is applied post-training with linear time complexity, the computational overhead remains negligible." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 382, + 388, + 394 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 382, + 388, + 394 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 388, + 394 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 402, + 555, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 555, + 546 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 555, + 546 + ], + "type": "text", + "content": "In this study, we focus on the feature drift problem in FL. The feature drift problem causes the same class samples on different clients to have distinct feature distributions, making it difficult for traditional model aggregation methods to handle such data heterogeneity. To tackle this problem, we design a prototype-based adversarial collaborative framework to unify feature spaces and enhance classification boundaries. The global classifier is retrained with mixed features to further grasp classification-relevant information from a global perspective. Our method has empirically achieved state-of-the-art performance in popular feature-drifted datasets with multiple data sources." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 555, + 411, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 411, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 411, + 567 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 574, + 553, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 553, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 553, + 629 + ], + "type": "text", + "content": "This work was supported in part by the Innovation Team Project of Guangdong Province of China (2024KCXTD017) and Guangdong-Hong Kong-Macau University Joint Laboratory (2023LSYS005)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 639, + 373, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 639, + 373, + 651 + ], + "spans": [ + { + "bbox": [ + 315, + 639, + 373, + 651 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 319, + 658, + 555, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 658, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 319, + 658, + 555, + 714 + ], + "type": "text", + "content": "[1] Haokun Chen, Ahmed Frikha, Denis Krompass, Jindong Gu, and Volker Tresp. Frau: Tackling federated learning with non-iid features via representation augmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4849-4859, 2023. 2" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3118" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 73, + 296, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 296, + 127 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 296, + 127 + ], + "type": "text", + "content": "[2] Yuhang Chen, Wenke Huang, and Mang Ye. Fair federated learning under domain skew with local consistency and domain diversity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12077-12086, 2024. 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 129, + 296, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 129, + 296, + 173 + ], + "spans": [ + { + "bbox": [ + 61, + 129, + 296, + 173 + ], + "type": "text", + "content": "[3] Liam Collins, Hamed Hassani, Aryan Mokhtari, and Sanjay Shakkottai. Exploiting shared representations for personalized federated learning. In International conference on machine learning, pages 2089-2099. PMLR, 2021. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 175, + 296, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 175, + 296, + 218 + ], + "spans": [ + { + "bbox": [ + 62, + 175, + 296, + 218 + ], + "type": "text", + "content": "[4] Alireza Fallah, Aryan Mokhtari, and Asuman Ozdaglar. Personalized federated learning with theoretical guarantees: A model-agnostic meta-learning approach. Advances in neural information processing systems, 33:3557–3568, 2020. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 220, + 295, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 220, + 295, + 252 + ], + "spans": [ + { + "bbox": [ + 62, + 220, + 295, + 252 + ], + "type": "text", + "content": "[5] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International conference on machine learning, pages 1180-1189. PMLR, 2015. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 254, + 295, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 254, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 62, + 254, + 295, + 297 + ], + "type": "text", + "content": "[6] Boqing Gong, Yuan Shi, Fei Sha, and Kristen Grauman. Geodesic flow kernel for unsupervised domain adaptation. In 2012 IEEE conference on computer vision and pattern recognition, pages 2066-2073. IEEE, 2012. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 299, + 295, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 299, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 62, + 299, + 295, + 342 + ], + "type": "text", + "content": "[7] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 343, + 295, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 343, + 295, + 387 + ], + "spans": [ + { + "bbox": [ + 62, + 343, + 295, + 387 + ], + "type": "text", + "content": "[8] Wenke Huang, Yuxia Liu, Mang Ye, Jun Chen, and Bo Du. Federated learning with long-tailed data via representation unification and classifier rectification. IEEE Transactions on Information Forensics and Security, 2024. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 388, + 296, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 388, + 296, + 421 + ], + "spans": [ + { + "bbox": [ + 62, + 388, + 296, + 421 + ], + "type": "text", + "content": "[9] Jonathan J. Hull. A database for handwritten text recognition research. IEEE Transactions on pattern analysis and machine intelligence, 16(5):550-554, 1994. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 422, + 295, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 422, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 57, + 422, + 295, + 464 + ], + "type": "text", + "content": "[10] Yihan Jiang, Jakub Konečný, Keith Rush, and Sreeram Kannan. Improving federated learning personalization via model agnostic meta learning. arXiv preprint arXiv:1909.12488, 2019. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 467, + 295, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 467, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 57, + 467, + 295, + 521 + ], + "type": "text", + "content": "[11] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank J Reddi, Sebastian U Stich, and Ananda Theertha Suresh. Scaffold: Stochastic controlled averaging for ondevice federated learning. arXiv preprint arXiv:1910.06378, 2(6), 2019. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 523, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 523, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 57, + 523, + 295, + 555 + ], + "type": "text", + "content": "[12] Jakub Konečny, Brendan McMahan, and Daniel Ramage. Federated optimization: Distributed optimization beyond the datacenter. arXiv preprint arXiv:1511.03575, 2015. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 557, + 295, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 557, + 295, + 600 + ], + "spans": [ + { + "bbox": [ + 57, + 557, + 295, + 600 + ], + "type": "text", + "content": "[13] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 602, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 295, + 624 + ], + "type": "text", + "content": "[14] Boyuan Li, Shengbo Chen, and Zihao Peng. New generation federated learning. Sensors, 22(21):8475, 2022. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 625, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 295, + 668 + ], + "type": "text", + "content": "[15] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. In Proceedings of the IEEE international conference on computer vision, pages 5542-5550, 2017. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[16] Qinbin Li, Bingsheng He, and Dawn Song. Model-contrastive federated learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10713–10722, 2021. 2, 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[17] Qinbin Li, Bingsheng He, and Dawn Song. Adversarial collaborative learning on non-iid features. In International Conference on Machine Learning, pages 19504-19526. PMLR, 2023. 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 119, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 162 + ], + "type": "text", + "content": "[18] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. Proceedings of Machine learning and systems, 2:429-450, 2020. 1, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 164, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 164, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 317, + 164, + 553, + 207 + ], + "type": "text", + "content": "[19] Xiaoxiao Li, Meirui JIANG, Xiaofei Zhang, Michael Kamp, and Qi Dou. Fedbn: Federated learning on non-iid features via local batch normalization. In International Conference on Learning Representations. 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 209, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 553, + 252 + ], + "type": "text", + "content": "[20] Zijian Li, Zehong Lin, Jiawei Shao, Yuyi Mao, and Jun Zhang. Fedcir: Client-invariant representation learning for federated non-iid features. IEEE Transactions on Mobile Computing, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 254, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 254, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 317, + 254, + 553, + 297 + ], + "type": "text", + "content": "[21] Xiao Lin, Indranil Sur, Samuel A Nastase, Ajay Divakaran, Uri Hasson, and Mohamed R Amer. Data-efficient mutual information neural estimator. arXiv preprint arXiv:1905.03319, 2019. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 299, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 299, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 317, + 299, + 553, + 352 + ], + "type": "text", + "content": "[22] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 1, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 354, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 354, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 317, + 354, + 553, + 398 + ], + "type": "text", + "content": "[23] Xutong Mu, Yulong Shen, Ke Cheng, Xueli Geng, Jiaxuan Fu, Tao Zhang, and Zhiwei Zhang. Fedproc: Prototypical contrastive federated learning on non-iid data. Future Generation Computer Systems, 143:93-104, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 399, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 553, + 454 + ], + "type": "text", + "content": "[24] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Baolin Wu, Andrew Y Ng, et al. Reading digits in natural images with unsupervised feature learning. In NIPS workshop on deep learning and unsupervised feature learning, page 4. Granada, 2011. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 456, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 553, + 499 + ], + "type": "text", + "content": "[25] Yu Qiao, Md Shirajum Munir, Apurba Adhikary, Huy Q Le, Avi Deb Raha, Chaoning Zhang, and Choong Seon Hong. Mp-fedcl: Multi-prototype federated contrastive learning for edge intelligence. IEEE Internet of Things journal, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 501, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 501, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 501, + 553, + 544 + ], + "type": "text", + "content": "[26] Min Tan, Yinfu Feng, Lingqiang Chu, Jingcheng Shi, Rong Xiao, Haihong Tang, and Jun Yu. Fedsea: Federated learning via selective feature alignment for non-iid multimodal data. IEEE Transactions on Multimedia, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 545, + 555, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 545, + 555, + 599 + ], + "spans": [ + { + "bbox": [ + 317, + 545, + 555, + 599 + ], + "type": "text", + "content": "[27] Yue Tan, Guodong Long, Lu Liu, Tianyi Zhou, Qinghua Lu, Jing Jiang, and Chengqi Zhang. Fedproto: Federated prototype learning across heterogeneous clients. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8432-8440, 2022. 2, 3, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 601, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 601, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 317, + 601, + 553, + 645 + ], + "type": "text", + "content": "[28] Yue Tan, Guodong Long, Jie Ma, Lu Liu, Tianyi Zhou, and Jing Jiang. Federated learning from pre-trained models: A contrastive learning approach. Advances in neural information processing systems, 35:19332-19344, 2022. 2, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "type": "text", + "content": "[29] Lei Wang, Jieming Bian, Letian Zhang, Chen Chen, and Jie Xu. Taming cross-domain representation variance in federated prototype learning with heterogeneous data domains. arXiv preprint arXiv:2403.09048, 2024. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[30] Li Wang, Quangui Zhang, Lei Sang, Qiang Wu, and Min Xu. Federated prototype-based contrastive learning for privacy-" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3119" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 274 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 76, + 73, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 296, + 95 + ], + "type": "text", + "content": "preserving cross-domain recommendation. arXiv preprint arXiv:2409.03294, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 129 + ], + "type": "text", + "content": "[31] Yue Zhao, Meng Li, Liangzhen Lai, Naveen Suda, Damon Civin, and Vikas Chandra. Federated learning with non-iid data. arXiv preprint arXiv:1806.00582, 2018. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 130, + 296, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 296, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 296, + 185 + ], + "type": "text", + "content": "[32] Kaiyang Zhou, Yongxin Yang, Timothy Hospedales, and Tao Xiang. Learning to generate novel domains for domain generalization. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16, pages 561-578. Springer, 2020. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 296, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 296, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 296, + 229 + ], + "type": "text", + "content": "[33] Tailin Zhou, Jun Zhang, and Danny HK Tsang. Fedfa: Federated learning with feature anchors to align features and classifiers for heterogeneous data. IEEE Transactions on Mobile Computing, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 231, + 296, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 296, + 274 + ], + "type": "text", + "content": "[34] Zhuangdi Zhu, Junyuan Hong, and Jiayu Zhou. Data-free knowledge distillation for heterogeneous federated learning. In International conference on machine learning, pages 12878-12889. PMLR, 2021. 2" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "text", + "content": "3120" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/FedVLA_ Federated Vision-Language-Action Learning with Dual Gating Mixture-of-Experts for Robotic Manipulation/9830aa12-506c-4606-9232-07331ac373a7_content_list.json b/2025/FedVLA_ Federated Vision-Language-Action Learning with Dual Gating Mixture-of-Experts for Robotic Manipulation/9830aa12-506c-4606-9232-07331ac373a7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..60a5a5ce31f2840d84c73d3c785f31acf5783b49 --- /dev/null +++ b/2025/FedVLA_ Federated Vision-Language-Action Learning with Dual Gating Mixture-of-Experts for Robotic Manipulation/9830aa12-506c-4606-9232-07331ac373a7_content_list.json @@ -0,0 +1,1578 @@ +[ + { + "type": "text", + "text": "FedVLA: Federated Vision-Language-Action Learning with Dual Gating Mixture-of-Experts for Robotic Manipulation", + "text_level": 1, + "bbox": [ + 133, + 128, + 867, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cui Miao $^{1}$ Tao Chang $^{1}$ Meihan Wu $^{1}$ Hongbin Xu $^{2}$ Chun Li $^{3}$ Ming Li $^{4*}$ Xiaodong Wang $^{1}$ National University of Defense Technology \n $^{2}$ Bytedance Seed Shenzhen MSU-BIT University \nLong Laboratory of Artificial Intelligence and Digital Economics", + "bbox": [ + 266, + 202, + 723, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 342 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-language-action (VLA) models have significantly advanced robotic manipulation by enabling robots to interpret language instructions for task execution. However, training these models often relies on large-scale user-specific data, raising concerns about privacy and security, which in turn limits their broader adoption. To address this, we propose FedVLA, the first federated VLA learning framework, enabling distributed model training that preserves data privacy without compromising performance. Our framework integrates task-aware representation learning, adaptive expert selection, and expert-driven federated aggregation, enabling efficient and privacy-preserving training of VLA models. Specifically, we introduce an Instruction-Oriented Scene-Parsing mechanism, which decomposes and enhances object-level features based on task instructions, improving contextual understanding. To effectively learn diverse task patterns, we design a Dual Gating Mixture-of-Experts (DGMoE) mechanism, where not only input tokens but also self-aware experts adaptively decide their activation. Finally, we propose an Expert-Driven Aggregation strategy at the federated server, where model aggregation is guided by activated experts, ensuring effective cross-client knowledge transfer. Extensive simulations and real-world robotic experiments demonstrate the effectiveness of our proposals. Notably, DGMoE significantly improves computational efficiency compared to its vanilla counterpart, while FedVLA achieves task success rates comparable to centralized training, effectively preserving data privacy.", + "bbox": [ + 89, + 357, + 483, + 782 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 808, + 220, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-language-action (VLA) models, which integrate visual perception, linguistic understanding, and robotic control, have significantly enhanced robotic manipulation by", + "bbox": [ + 89, + 833, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3cd25ba72ac5b657c4f1fe871e3689f016b13fae0b1aa74685f7b1e332cfc4dd.jpg", + "image_caption": [ + "(a) Centralized VLA learning", + "Figure 1. Comparison between centralized and federated VLA training. Centralized training collects all user data on the cloud, raising privacy concerns due to potential exposure of sensitive information. Our federated VLA framework enables decentralized training on user devices, preserving privacy while utilizing expert-driven aggregation to enhance model generalization across diverse tasks." + ], + "image_footnote": [], + "bbox": [ + 532, + 325, + 696, + 496 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2ce0c8bce08eb2f30ee766efc3ed1934089c63469e74c0ee47f8c907d7572d49.jpg", + "image_caption": [ + "(b) Federated VLA learning" + ], + "image_footnote": [], + "bbox": [ + 707, + 325, + 890, + 494 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "enabling robots to comprehend and execute complex tasks through natural language instructions [5, 14, 28]. However, training these models requires extensive indoor datasets that capture diverse, user-specific environments, raising serious concerns about data privacy. While recent advancements have improved the capability of VLA models in robotic systems, the issue of privacy leakage remains largely unexplored [19].", + "bbox": [ + 511, + 654, + 906, + 775 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Federated learning (FL) provides an effective privacy-preserving solution for training VLA models [22]. Unlike traditional centralized training, which requires aggregating all user data on a central server, FL enables distributed model training across multiple clients without transferring raw data. As illustrated in Figure 1a, centralized VLA training collects and processes user data in the cloud, posing risks of privacy leakage, such as exposure of personal pref", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 887, + 233, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6904", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "erences and sensitive health conditions. In contrast, figure 1b presents a federated VLA learning paradigm, where local models are trained on user devices, and only model updates are transmitted to the central server. This decentralized paradigm mitigates privacy risks while allowing collaborative model optimization.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing federated learning approaches are not well-suited for VLA learning, where robotic manipulation tasks exhibit substantial task heterogeneity across different clients. Typical FL methods, such as FedAvg [22], naively aggregate models across tasks, considering only shared synergies while overlooking their fundamental distinctions. Although recent works attempt to address this issue through merge-and-split training [33] or decoupled model aggregation [20], they remain limited to single-modal settings, assuming a uniform input modality across tasks. In contrast, VLA models operate in multi-modal environments, requiring the joint processing of visual observations, language instructions, and robotic actions, which significantly increases the complexity of federated training.", + "bbox": [ + 89, + 185, + 482, + 396 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Additionally, mixture of experts (MoE) has shown promise in multi-task learning by assigning specialized experts to different tasks [7, 12, 29]. However, existing methods typically select a fixed number of experts, lacking adaptability to tasks of varying complexity. This rigidity makes them inefficient for resource-constrained clients, where optimizing inference speed is critical for real-time robotic responses. These limitations highlight the need for a task-adaptive and flexible FL framework, specifically designed for multi-modal robotic learning.", + "bbox": [ + 89, + 398, + 482, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce FedVLA, a novel federated vision-language-action (VLA) learning framework with Dual Gating MoE, designed to preserve privacy while maintaining high task performance, as illustrated in Figure 1b. To address task heterogeneity and enable effective knowledge aggregation in federated settings, we propose three key components. First, we introduce an Instruction-Oriented Scene-Parsing (IOSP) module to enhance task-aware feature extraction in decentralized environments, where raw images across clients exhibit diverse backgrounds and object distributions. IOSP decomposes observation images into object-level representations guided by task instructions and leverages vision-language alignment techniques to improve contextual understanding. Second, to enhance computational efficiency in federated learning, we develop a Dual Gating MoE (DGMoE) mechanism that enables adaptive knowledge routing. Unlike conventional MoE, which assigns tokens to a fixed number of experts, DGMoE introduces self-aware experts that can accept or reject token assignments, enabling a bidirectional selection process. This dual gating mechanism improves efficiency while preserving task performance. Lastly, to effectively aggregate similar knowledge from diverse client models, we propose the", + "bbox": [ + 89, + 553, + 482, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Expert-Driven Aggregation (EDA) strategy, which dynamically assigns aggregation weights based on expert similarity across clients. By prioritizing updates from clients that activate similar experts, EDA ensures that semantically aligned knowledge is more prominently integrated at the server side, resulting in a more coherent and specialized global model.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions in this work can be summarized as follows:", + "bbox": [ + 511, + 183, + 903, + 212 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose FedVLA, the first privacy-preserving federated learning framework for VLA training, ensuring that user data remains private and never shared externally. Extensive experiments in both simulation and real-world environments demonstrate that FedVLA achieves performance comparable to centralized training while preserving data privacy.", + "- We introduce the Dual Gating Mixture-of-Experts, where self-aware experts autonomously determine their participation, enabling bidirectional token selection to optimize computational efficiency while maintaining task performance.", + "- We propose the Expert-Driven Aggregation strategy, which dynamically assigns aggregation weights based on expert similarity across clients, ensuring that semantically aligned knowledge is more effectively integrated, thereby enhancing generalization across heterogeneous tasks." + ], + "bbox": [ + 513, + 214, + 903, + 469 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 484, + 661, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Vision-Language-Action Models", + "text_level": 1, + "bbox": [ + 513, + 511, + 795, + 526 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision-Language Models (VLMs), which are pre-trained on large-scale datasets, have demonstrated remarkable capabilities in understanding and generating multimodal data [17, 18, 26]. Inspired by VLMs, Vision-Language-Action (VLA) models extend their capabilities by integrating robotic action prediction. For instance, RoboFlamingo [15], RT-2 [5], and OpenVLA [14] are all built upon VLMs such as OpenFlamingo [3], PaLI-X [6], and Prismatic [13], respectively. Recent work has begun exploring the training of generalist robots [28], followed by fine-tuning on diverse robot datasets, enabling generalization to novel instructions, unseen objects, and distribution shifts [31]. This progress relies heavily on the availability of large-scale public datasets, such as Open-X Embodiment dataset [24]. During the finetuning phase, samples from real-world scenarios are required, which are often generated by human activities and distributed across users' home environments. However, these samples frequently expose privacy-sensitive information.", + "bbox": [ + 511, + 532, + 903, + 820 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Privacy-Preserving Training", + "text_level": 1, + "bbox": [ + 513, + 832, + 767, + 848 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Traditional centralized learning methods require uploading data from multiple sources, which poses significant privacy risks. To mitigate these concerns, researchers have pro", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6905", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/69ecaee2a94e1b08f56c1b73f3ce3fdc656be67076c1ddc3fbd69875c5ca0d3b.jpg", + "image_caption": [ + "Figure 2. Illustration of the proposed FedVLA framework and its key component. (a) An overview of the FedVLA, which consists of multiple clients and a server. Each client processes vision-language-action inputs through the Stem, Trunk, and Head modules. Clients send their updated trunk parameters to the server, where Expert-Driven Aggregation selectively integrates expert-specific updates on server. The global trunk module is then distributed to clients for the next training round. (b) Details of the proposed Dual Gating MoE module. The input tokens first pass through the token-level gate $G_{t}$ , computing the gating scores for each expert with residual scores from the previous MoE layer. Only the computed scores surpass the expert-level gate $G_{e}$ , activating the corresponding expert to process the token. Finally, the expert outputs are combined to generate the final output tokens." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 908, + 324 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "posed various privacy-preserving techniques, including federated learning (FL) [22], differential privacy [9], secure multi-party computation [10], and homomorphic encryption [2]. However, Most privacy-preserving training techniques have been developed for single-modality tasks, and seldomly study the data privacy of embodied AI [32]. In robotic manipulation, the models process multimodal sensor data, including images, language instructions, and action sequences, which can contain sensitive user interaction data. Unlike static datasets, robotic systems continuously collect and process dynamic, task-specific information, making privacy protection more complex. In this context, FL as a solution by allowing decentralized model training without sharing raw data. However, FL for robotic manipulation remains challenging due to task heterogeneity and the need for efficient model aggregation across clients.", + "bbox": [ + 88, + 458, + 485, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Mixture of Experts in Large Models", + "text_level": 1, + "bbox": [ + 89, + 710, + 405, + 728 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Mixture of experts [4, 11] is an architecture where multiple experts are trained collaboratively, with each expert responsible for a specific portion of the input space. An MoE consists of two components: a routing mechanism and a set of experts. The experts can either be part of a feedforward network (FFN) or the entire FFN layer, and the routing mechanism allows each token to select experts for processing. By activating a specific number of experts, MoE can reduce resource consumption and improve inference speed. This feature makes MoE particularly effective when integrated into large models, as seen in models like LLaVA-MoE [16],", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "DeepSeekMoE [8], and GPT-4 [1]. In this work, we apply MoE to a generalist robot VLA model and introduce a self-aware variant that can adaptively decide whether to be selected, significantly improving deployment efficiency.", + "bbox": [ + 511, + 458, + 906, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. FedVLA", + "text_level": 1, + "bbox": [ + 511, + 530, + 609, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We aim to train a vision-language-action model in a federated learning (FL) framework, ensuring privacy preservation while enabling task generalization across clients. Formally, given $N$ clients $\\mathbb{C} = \\{C_1,C_2,\\dots,C_N\\}$ , $N$ local tasks $\\mathbb{T} = \\{T_1,T_2,\\dots,T_N\\}$ , and $N$ local datasets $\\mathbb{D} = \\{\\mathbb{D}_1,\\mathbb{D}_2,\\dots,\\mathbb{D}_N\\}$ , each client $C_i$ is associated with a local task $T_{i}$ and a corresponding dataset $\\mathbb{D}_i$ . The dataset $\\mathbb{D}_i$ contains visual perception image $o_i$ , language instruction $l_{i}$ of the task, proprioception $s_i$ of the robot, and action sequence $a_{i}$ executed by the robot:", + "bbox": [ + 511, + 556, + 908, + 707 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {D} _ {i} = \\left\\{\\left(o _ {i}, l _ {i}, s _ {i}, a _ {i}\\right) _ {j} | j = 1, 2, \\dots , M \\right\\}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 709, + 906, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $M$ is the number of samples in dataset $\\mathbb{D}_i$ . To train a federated VLA model $\\theta_G$ , the overall federated optimization objective can be formulated as:", + "bbox": [ + 511, + 729, + 906, + 775 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta_ {G}} \\sum_ {i = 1} ^ {N} w _ {i} \\sum_ {(o _ {i}, l _ {i}, a _ {i}, s _ {i}) \\in \\mathbb {D} _ {i}} \\mathcal {L} _ {i} (f _ {i} (\\theta_ {G}; o _ {i}, l _ {i}, s _ {i}), a _ {i}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 777, + 906, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_i$ is the Huber loss function computed over each client model's local dataset $\\mathbb{D}_i$ , $f_i$ represents the local client model's predicted action sequence, and $w_i$ denotes the weight assigned to each client during aggregation, which is further detailed in Section 3.3.", + "bbox": [ + 511, + 824, + 908, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6906", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0c5055f55c82abdd9529d01f59dbe56cd508721fe706971b900405a0997767a6.jpg", + "image_caption": [ + "Figure 3. Instruction-Oriented Scene-Parsing module, which decomposes observations into object-level representations guided by task instructions, enhancing task-aware feature extraction through vision-language alignment." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 480, + 170 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overview of our FedVLA framework is shown in Figure 2. The clients train their models locally and communicate with the server per round until the model converges. The following sections introduce our key proposals: (1) Instruction-Oriented Scene-Parsing (Section 3.1) that decomposes the observation image into object-level embeddings guided by task instructions. (2) Dual Gating Mixture-of-Experts (Section 3.2) that can adaptively select a varying number of experts for tokens based on the relevance to the specific task. (3) Expert-Driven Aggregation (Section 3.3) that dynamically aggregates the weights based on expert similarities across different clients in the server.", + "bbox": [ + 89, + 265, + 483, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Instruction-Oriented Scene-Parsing", + "text_level": 1, + "bbox": [ + 89, + 455, + 401, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To ensure the robot focuses on task-relevant information while maintaining contextual awareness to avoid collisions, we decouple objects in the scene based on the specific robotic task. To achieve this, we propose an Instruction-Oriented Scene-Parsing (IOSP) module to represent objects within the task scene.", + "bbox": [ + 89, + 477, + 483, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As illustrated in Figure 3, given an instruction and an image, we first extract Target Objects (TOs) from the instruction using named entity recognition [23]. Simultaneously, we employ YOLOv8 [27] to detect and classify Foreground Objects (FOs) and Background Objects (BOs) in the image. The scene objects are categorized into three groups based on the cosine similarity measured by a pre-trained CLIP model [25] between the instruction and object names, following these matching rules: (1) Target Objects (TOs): Objects explicitly mentioned in the instruction and directly involved in the action. (2) Surrounding Objects (SOs): Foreground objects not in TOs, providing spatial and contextual cues. (3) Background Objects (BOs): Static environmental elements ensuring scene consistency but not directly affecting the task.", + "bbox": [ + 89, + 569, + 482, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, we encode the grouped objects using the CLIP model. The text encoder $H$ maps object names from TOs, SOs, and BOs into textual embeddings $H(\\mathrm{TOs}) \\in \\mathbb{R}^{t \\times D}$ , $H(\\mathrm{BOs}) \\in \\mathbb{R}^{t \\times D}$ , and $H(\\mathrm{SOs}) \\in \\mathbb{R}^{t \\times D}$ . Meanwhile, the vision encoder $P$ processes the entire image into $P(X) \\in \\mathbb{R}^{t \\times D}$ . We assign image tokens to object groups by computing the cosine similarity between token and text embed", + "bbox": [ + 89, + 795, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "dings, selecting the top eight tokens per group. To enhance intra-group feature representation, each group is processed through an MoE module, producing refined image token embeddings. The enhanced tokens are then concatenated with the remaining tokens for subsequent processing. This IOSP module enables the model to focus on critical objects, improving task-specific manipulation.", + "bbox": [ + 511, + 90, + 905, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Dual Gating Mixture-of-Experts", + "text_level": 1, + "bbox": [ + 511, + 204, + 797, + 220 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we introduce Dual Gating Mixture-of-Expert (DGMoE) mechanism that supports adaptive knowledge routing. Different from vanilla MoE, which assigns a fixed number of experts to each token, our DGMoE can automatically determine the acceptance or rejection of the token assignments under the Self-Aware Expert.", + "bbox": [ + 511, + 226, + 903, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Figure 2, each DGMoE layer is comprised of $K$ self-aware experts $\\mathbb{E} = \\{E_1,E_2,\\dots ,E_K\\}$ . To enable the automatic token-driven expert selection, we introduce two gating mechanisms: Token-side gating mechanism $G_{t}$ and Expert-side gating mechanism $G_{e}$ . (1) For Token-side gating, $G_{t}$ serves as the gating module on the token side, comprised of a soft router module and a gating residual module. The soft router first estimates the selection scores for the hidden inputs, indicating the token preference for different experts. Then the estimated scores from previous layer are incorporated with the scores of current layer, allowing the tokens to inherit the selection prior of the most suitable experts from previous layers. (2) For Expert-side gating, $G_{e}$ can autonomously accept or reject tokens based on their acceptance threshold. The selection scores predicted by $G_{t}$ is filtered with a learnable threshold, and only the experts with the most relevant tokens are activated.", + "bbox": [ + 511, + 316, + 905, + 573 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given an input token $x$ , the token-side gate $G_{t}$ first computes the scores at the $j$ -th layer by combining the soft router output and the gating residual module:", + "bbox": [ + 511, + 574, + 905, + 619 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nG _ {t} ^ {j} (x) = \\left\\{ \\begin{array}{l l} W _ {t} ^ {j} x, & \\text {i f} j = 1, \\\\ W _ {t} ^ {j} x + W _ {t} ^ {g} G _ {t} ^ {(j - 1)} (x), & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 625, + 906, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $W_{t}^{j}\\in \\mathbb{R}^{D}$ is the trainable weight matrix for soft router, and $W_{t}^{g}\\in \\mathbb{R}^{N\\times N}$ is a trainable transformation matrix designed to aggregate the scores from the previous layer and the current layer, ensuring smooth expertise knowledge transitions across layers. Then, the token-to-expert selection scores $s_t(x)$ are obtained by normalizing $G_{t}(x)$ with a softmax function, which determines the probability distribution over the available experts.", + "bbox": [ + 511, + 672, + 905, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "On the self-aware expert side, the expert-side gate $G_{e}$ evaluates the token scores and determines the final expert activation scores $s_{e}(x)$ as follows:", + "bbox": [ + 511, + 795, + 903, + 840 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns _ {e} (x) = \\operatorname {s i g n} \\left(s _ {t} (x) - \\lambda W _ {e}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 611, + 847, + 903, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $W_{e}$ is a set of trainable self-aware parameters corresponding to the gating values of each expert, and $\\lambda$ is a", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6907", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "scaling factor that modulates the gating threshold. In our setting, we set $\\lambda = 0.5$ to balance expert selection, ensuring neither excessive expert activation nor overly strict filtering of tokens.", + "bbox": [ + 89, + 90, + 482, + 148 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, combining the dual gating mechanism of token-side and exper-side gating, the expert selection function is defined as follows:", + "bbox": [ + 89, + 151, + 483, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ng (x) = \\left\\{ \\begin{array}{l l} s _ {t} (x), & \\text {i f} s _ {e} (x) > 0, \\\\ 0, & \\text {i f} s _ {e} (x) \\leq 0, \\end{array} \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 200, + 482, + 241 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the final output token $y$ of the DGMoE layer can be computed as the weighted sum of the selected expert outputs:", + "bbox": [ + 89, + 248, + 483, + 291 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ny = \\sum_ {i = 1} ^ {K} g _ {i} (x) E _ {i} (x). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 282, + 482, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the help of self-aware expert gating mechanism, our DGMoE can dynamically activate the most relevant experts, reducing computational overhead while maintaining task-relevant feature learning.", + "bbox": [ + 89, + 325, + 483, + 386 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Expert-Driven Aggregation", + "text_level": 1, + "bbox": [ + 89, + 393, + 338, + 410 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To mitigate performance degradation caused by task heterogeneity, we propose an Expert-Driven Aggregation (EDA) strategy, which aggregates model parameters based on the similarity of expert selections across clients. And as shown in Figure 2, we selectively aggregates only the trunk module, while the stem and head modules remain personalized and are excluded from aggregation.", + "bbox": [ + 89, + 415, + 483, + 521 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a trunk module with $L$ layers, where each layer contains $K$ experts, each client $C_i$ records an expert selection matrix $\\mathbf{V}_i\\in \\mathbb{R}^{L\\times K}$ during each round through DGMoE, where each element $V_{i}^{(l,k)}$ represents the number of times expert $k$ is activated in the $l$ -th layer. We extract the expert selection vector for layer $l$ of client $C_i$ as:", + "bbox": [ + 89, + 521, + 483, + 614 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {i} ^ {(l)} = \\left(V _ {i} ^ {(l, 1)}, V _ {i} ^ {(l, 2)}, \\dots , V _ {i} ^ {(l, K)}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 619, + 482, + 642 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then we compute the pairwise similarity between clients based on expert selection vectors. The similarity between clients $C_i$ and $C_j$ at layer $l$ is defined as follows:", + "bbox": [ + 89, + 648, + 483, + 694 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns _ {i, j} ^ {(l)} = \\frac {\\mathbf {v} _ {i} ^ {(l)} \\cdot \\mathbf {v} _ {j} ^ {(l)}}{\\| \\mathbf {v} _ {i} ^ {(l)} \\| \\| \\mathbf {v} _ {j} ^ {(l)} \\|\\}.} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 700, + 482, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Using this similarity, we determine the aggregation weight $w_{l,i}$ for each client $C_i$ at layer $l$ as follows:", + "bbox": [ + 89, + 751, + 482, + 781 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw _ {l, i} = \\frac {\\sum_ {C _ {j} \\in C} s _ {i , j} ^ {(l)}}{\\sum_ {C _ {j} \\in C} \\sum_ {C _ {i} \\in C} s _ {i , j} ^ {(l)}}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 789, + 482, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This weighting strategy ensures that clients with higher expert selection similarity contribute more to each other's model updates, allowing for more effective and task-aligned aggregation across task-heterogeneous clients.", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 FedVLA: Clients", + "bbox": [ + 514, + 90, + 727, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input: Each client $C_i$ has a local task $T_i$ and dataset $D_i = \\{o_i, l_i, s_i, a_i\\}$ , along with local parameters $\\phi_i$ , $w_i$ , and $\\theta_i$ corresponding to the stem, trunk, and head modules, respectively.", + "bbox": [ + 511, + 107, + 906, + 167 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Local_Train $(C_i, \\phi_i, \\theta_i)$ :", + "2: for each local epoch $e$ from 1 to $E_{c}$ do", + "3: for each batch $(o_i, l_i, s_i, a_i)$ sampled from $D_i$ do", + "4: // Step 1: Instruction-Oriented Scene-Parsing", + "5: TOs, SOs, BOs $\\leftarrow \\{o_i, l_i\\}$", + "6: $F_{T O s}, F_{S O s}, F_{B O s} \\gets \\mathrm{CLIP}(T O s, S O s, B O s)$", + "7: $F_{O}\\gets \\mathrm{CLIP}(o_{i})$", + "8: $X_{\\mathrm{TO}} \\gets \\mathrm{topk}(\\mathrm{cosine}(F_{\\mathrm{TOs}}, F_O), F_O)$", + "9: $X_{\\mathrm{SO}} \\gets \\mathrm{topk}(\\mathrm{cosine}(F_{\\mathrm{SOs}}, F_O), F_O)$", + "10: $X_{\\mathrm{BO}} \\gets \\mathrm{topk}(\\mathrm{cosine}(F_{\\mathrm{BOs}}, F_O), F_O)$", + "11: $X_{\\mathrm{in}} = \\mathrm{concat}(X_{\\mathrm{TO}},X_{\\mathrm{SO}},X_{\\mathrm{BO}},s_i)$", + "12: // Step 2: Forward pass with DGMoE", + "13: Let $x \\in MHA(X_{\\mathrm{in}})$", + "$s_t(x) = \\mathrm{softmax}\\bigl (W_t(x)\\bigr)$ // token-gate", + "14: $s_e(x) = \\mathrm{sign}(s_t(x) - \\lambda W_e)$ // expert-gate", + "15: $g(x) = s_t(x)\\cdot \\mathbb{I}(s_e(x) > 0)$ // expert activation", + "16: $\\mathbf{v}^{(l)},y = \\sum_{k = 1}^{K}g(x)E_k(x)$ //final output", + "17: // Step 3: Model parameter updates", + "18: $\\mathcal{L}_i = \\mathrm{Loss}(f(\\phi_i, \\theta_i, w_i), a_i)$", + "19: $(\\phi_i,\\theta_i,w_i)\\gets (\\phi_i,\\theta_i,w_i) - \\eta \\nabla \\mathcal{L}_i$", + "20: end for", + "21: end for", + "22: return $\\theta, \\mathbf{v}^{(l)}$ to the server." + ], + "bbox": [ + 519, + 185, + 903, + 582 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Algorithms", + "text_level": 1, + "bbox": [ + 513, + 625, + 637, + 641 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Clients. The overall workflow of our FedVLA is shown in Algorithm 1 and Algorithm 2. At the start of each training round $t$ , each client processes task-specific data using Instruction-Oriented Scene-Parsing to extract structured features, followed by Dual Gating MoE for adaptive expert selection. The updated trunk parameters and expert selection counts are then transmitted to the server for global aggregation. The detailed client-side process is described in Algorithm 1.", + "bbox": [ + 511, + 652, + 906, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Server. On the server side, the server receives expert selection statistics and trunk updates from all participating clients and performs Expert-Driven Aggregation, which dynamically assigns aggregation weights based on expert selection similarities. The aggregated global trunk module is then redistributed to clients for the next training round. The process is elaborated in Algorithm 2.", + "bbox": [ + 511, + 795, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6908", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 2 FedVLA: Server", + "text_level": 1, + "bbox": [ + 91, + 90, + 302, + 104 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Input: $T$ is the number of training rounds. $\\theta$ represents the global trunk parameters. Each client provides $\\mathbf{v}^{(l)}$ , the expert selection counts vector for each layer $l$ .", + "bbox": [ + 89, + 107, + 483, + 154 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Server Execute:", + "2: for each round $t$ from 1 to $T$ do", + "3: // Expert-Driven Aggregation", + "4: for each trunk layer $l$ do", + "5: for each pair of clients $C_i, C_j$ do", + "6: $s_{i,j}^{(l)} = \\frac{\\mathbf{v}_i^{(l)}\\cdot\\mathbf{v}_j^{(l)}}{\\|\\mathbf{v}_i^{(l)}\\|\\|\\mathbf{v}_j^{(l)}\\|} //$ expert selection similarity", + "7: $w^{(l,i)} = \\frac{\\sum_{j\\in C}s_{i,j}^{(l)}}{\\sum_{j\\in C}\\sum_{i\\in C}s_{i,j}^{(l)}}$ // aggregation weight", + "8: end for", + "9: $\\theta^{(l)}\\gets \\sum_{i = 1}^{N}w^{(l,i)}\\cdot \\theta^{(l,i)}$ //update $\\theta$", + "10: end for", + "11: Broadcast updated $\\theta$ to all clients.", + "12: end for" + ], + "bbox": [ + 96, + 170, + 475, + 382 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 412, + 223, + 430 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We validate our FedVLA in both simulation and real-world scenarios, where each client performs a single task, resulting in a non-iid data distribution. We employ the pretrained HPT [28] as the backbone of our VLA model and train it for 1,000 communication rounds between clients and the server, with each client performing local training for 5 epochs per round. In simulation, the model is trained with a learning rate of $5 \\times 10^{-6}$ , while in real-world settings, the learning rate is set to $2 \\times 10^{-5}$ . Each client trains locally with a batch size of 256 using the Adam optimizer. The experiments are conducted on a workstation equipped with 2 GeForce GTX 4090Ti GPUs, 48 Intel Xeon CPUs, and 128GB of memory.", + "bbox": [ + 88, + 439, + 483, + 636 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation. For evaluation, the success and failure of a trial are recorded as 1 and 0. Each task takes for 15 trials and repeats for 5 times. Since this is the first work on federated VLA learning, we compare FedVLA with two methods. The centralized training (Centralized) aggregates all data on a single server for joint training, serving as an upper bound in performance. The federated averaging (FedAvg) [22] is a strong FL baseline where models are trained locally on each client and simply averaged on the server.", + "bbox": [ + 88, + 637, + 483, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Simulation", + "text_level": 1, + "bbox": [ + 89, + 786, + 210, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Environment. We utilize the MuJoCo engine [21] to simulate tasks from the Meta-World robot manipulation benchmark [30]. All tasks are performed using a Sawyer robot with a parallel gripper, positioned at the center of a dinner table. Visual data is captured by three fixed RGB cameras with a resolution of $128 \\times 128$ , providing corner, top, and", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "shoulder views. For collision detection and dynamics simulation, we employ official physics engines to ensure accurate robotic interactions within the simulation environment. Tasks. We evaluate our FedVLA on four household tasks: Door Lock, Close Printer, Sweep Into, and Open Window. For each task, we collect approximately 30-80 episodes, each consisting of 40-100 steps. The initial pose of robot and environment layouts are randomized to assess the generalization ability of FedVLA.", + "bbox": [ + 511, + 90, + 906, + 227 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5233da4f7d05f1bbbdea441724eabaf99a4c2a03ae3d1f78e99e334a27d94481.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAvg. SuccessDoor LockCloseDrawerSweep IntoOpen Window
Centralized65.086.773.353.346.7
FedAvg51.766.773.340.026.7
FedVLA (Ours)63.380.080.053.340.0
", + "bbox": [ + 521, + 239, + 898, + 314 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Simulation evaluation across four tasks.", + "bbox": [ + 563, + 325, + 854, + 338 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Performance Comparison. The evaluation results in the simulation environment are presented in Table 1. FedVLA achieves an average success rate of $63.3\\%$ , closely approaching the $65.0\\%$ of centralized training. Notably, in the Close Drawn task, FedVLA surpasses centralized training by a success rate margin of about $7.0\\%$ . Furthermore, FedVLA consistently outperforms FedAvg, which only achieves an average success rate of $51.7\\%$ . This performance gap can be attributed to FedAvg overlooking task heterogeneity in federated learning. In contrast, FedVLA explicitly accounts for task heterogeneity during aggregation by leveraging expert-driven model updates, ensuring that knowledge from diverse clients is effectively integrated to enhance generalization across different tasks.", + "bbox": [ + 511, + 354, + 906, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Real-World", + "text_level": 1, + "bbox": [ + 511, + 575, + 638, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Hardware. For real-world experiments, we employ a UR3 robotic arm with six degrees of freedom, equipped with a one-DoF parallel gripper. A RealSense D435i RGB-D camera is mounted above the robot, capturing RGB images in real time at a resolution of $1280 \\times 720$ with a frame rate of approximately $30\\mathrm{Hz}$ . To simulate real-world variations, we introduce moderate randomness in object placement and lighting conditions to better reflect practical deployment scenarios.", + "bbox": [ + 511, + 598, + 906, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tasks. We collect real-world robotic demonstrations for household-related tasks, including Clean Up,Trash Collection, OpenDrawer,and Sorting Pills, as shown in Figure 4. For data collection, we employ an expert teaching method, using inverse kinematics to compute the robot poses corresponding to key point data while recording RGB images and pose data throughout the trajectory. Each task includes approximately 50 demonstrations, with each trajectory consisting of 20 to 80 steps.", + "bbox": [ + 511, + 734, + 905, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Performance Comparison. The real-world experimental results are summarized in Table 2. FedVLA achieves an", + "bbox": [ + 511, + 869, + 905, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6909", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/575e642ab35fa047dadbb235fe3d844d2a04e1091cf58c1b5194923d6c3ee09c.jpg", + "image_caption": [ + "Figure 4. Real-World Tasks. Each column represents a different task, while each row demonstrates keyframes of task execution. From left to right, the tasks are: Clean Up, where the robot wipes the trash from plate using a cloth. Trash Collection, in which it throws crumpled paper into a trash bin. Open Printer, which requires grasping and pulling a drawer open, and Sorting Pills, where pill bottles are placed into trays of designated color." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 478, + 258 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "average task success rate of $63.3\\%$ , closely matching the $63.4\\%$ of centralized training. Notably, it attains the same success rates as the centralized model in complex tasks such as Trash Collection $(46.7\\%)$ and Sorting Pills $(73.3\\%)$ . In contrast, FedAvg exhibits a significant performance drop across all tasks, with an average success rate of $53.3\\%$ , which is $10.0\\%$ lower than our method. These results indicate that FedVLA effectively bridges the performance gap between federated and centralized training. While FedAvg merely averages model parameters across clients, it struggles to generalize effectively in VLA learning, where multimodal information—including textual instructions, scene images, and execution trajectories—is involved, requiring strong feature extraction capabilities. The proposed Dual Gating MoE and Expert-Driven Aggregation mechanisms enable the model to adaptively learn diverse patterns for robotic manipulation, enhancing its ability to handle complex tasks in federated settings.", + "bbox": [ + 88, + 393, + 482, + 667 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f5e324c3a5f23658093b504e726fecf94f99b32e2719cd838c2d20fca9ab0ca5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAvg. SuccessClean UpTrash CollectionOpenDrawerSorting Pills
Centralized63.446.746.786.773.3
FedAvg53.346.740.060.066.7
FedVLA (Ours)63.353.346.780.073.3
", + "bbox": [ + 96, + 678, + 477, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 801, + 254, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further explore the effectiveness of the IOSP, DGMOE and EDA in FedVLA, we conduct ablation experiments by individually removing each module while keeping the other components. Specifically, the experimental setups are as follows: (1) remove the IOSP module and directly input the", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "entire scene image and proprioception data into the encoder (w/o IOSP), (2) replace the DGMoE layers with standard FFN, eliminating adaptive expert selection (w/o DGMOE), and (3) perform vanilla FedAvg aggregation instead of EDA (w/o EDA). We compare the task success rate and record the validation loss during the training process.", + "bbox": [ + 511, + 90, + 906, + 183 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7af02fb37ce2849a58ad62402955c7d235e9366293c7deb5e035e5a279439d9a.jpg", + "table_caption": [ + "Table 2. Real-world evaluation across four tasks." + ], + "table_footnote": [], + "table_body": "
MethodAvg. SuccessClean UpTrash CollectionOpen DriverSorting Pills
w/o IOSP41.140.013.366.746.7
w/o DGMoE31.720.020.046.740.0
w/o EDA26.726.720.033.326.7
FedVLA (Ours)63.353.346.780.073.3
", + "bbox": [ + 519, + 198, + 898, + 280 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Ablation studies of proposed FedVLA without IOSP, DG-MoE, and EDA.", + "bbox": [ + 511, + 291, + 903, + 319 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "From Table 3, we can see that each module contributes significantly to the performance of FedVLA, and the lack of each module will lead to a decrease in the performance of the entire system. Specifically, removing IOSP leads to the highest performance drop in the Trash Collection task from $46.7\\%$ to $13.3\\%$ , because the number of objects in Trash Collection task is more than the others. The IOSP module discriminates each object from the other for taking the right action to adapt to multi-objects scenarios. DGMoE module strongly relates to the Clean Up task, leading to a success rate drop from $53.3\\%$ to $20.0\\%$ . The Clean Up task is harder in the manipulation aspect as it requires the robot interaction for multiple steps. The FedVLA depends on the DGMoE module to adaptively select the appropriate number of experts based on different types of tokens, enabling collaborative learning for interaction-rich tasks. Furthermore, the removal of EDA gives the largest reduction in average success rate from $63.3\\%$ down to $26.7\\%$ As a common aggregation cancels out the knowledge of the experts in the DGMoE module.", + "bbox": [ + 511, + 338, + 906, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 5 presents the validation loss of the ablation experiments during the first 500 training epochs. We can see that FedVLA consistently achieves the fastest convergence speed, along with the lowest and most stable validation losses across all tasks. Specifically, we can discover that (1) without IOSP causes validation losses notably increase in the multi-object scenarios (depicted in Figure 5b and Figure 5c), (2) the removal of DGMoE results in significantly increased and fluctuating losses the interaction-rich tasks, (3) removing the EDA module causes relatively higher validation losses in most tasks. These findings strongly support that IOSP helps decompose complex visual scenes into task-relevant object representations, DGMoE module routes the adaptive knowledge, EDA ensures the integrity and effectiveness of the aggregation. The integration of these modules together results in a architecture that supports FedVLA's robustness and adaptability across diverse tasks.", + "bbox": [ + 511, + 643, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6910", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1b39fcdc049f9c01fe695682f2ced6b37a9dccdb68f513eacd765d960adef0c4.jpg", + "image_caption": [ + "(a) OpenDrawer" + ], + "image_footnote": [], + "bbox": [ + 102, + 88, + 297, + 183 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8908323b8fc7a92162f28910efaf4b3eac7dd9f7da53318f93c22794a1ca093c.jpg", + "image_caption": [ + "(b) Sorting Pills" + ], + "image_footnote": [], + "bbox": [ + 300, + 88, + 496, + 185 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/92e15a63c2cc7bbedc895ee3cdbf4b1742f8105388b8c8543b4a68b0c512567d.jpg", + "image_caption": [ + "(c)Trash Collection" + ], + "image_footnote": [], + "bbox": [ + 500, + 89, + 694, + 184 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fd52ec18fc227287ca6c603c525477c6d8787087d4398eeeb4791a7052a20221.jpg", + "image_caption": [ + "(d) Clean Up" + ], + "image_footnote": [], + "bbox": [ + 697, + 89, + 893, + 184 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/18bfc2db1917015802b6753fc3818732878a18e59f16cc10db9c0d962daa0454.jpg", + "image_caption": [ + "(a) OpenDrawer" + ], + "image_footnote": [], + "bbox": [ + 94, + 270, + 287, + 349 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5cc549dc7830d07a30f94e6c4af71392b579a2bfee876f4ac8eb9220bb7790c5.jpg", + "image_caption": [ + "Figure 5. Validation loss comparison in ablation study across four tasks. The x-axis represents the training rounds, while the y-axis represents the validation loss. Each subplot includes a zoomed-in inset focusing on the validation loss fluctuations between rounds 400 and 500, illustrating the fine-grained differences in convergence trends.", + "(b) Sorting Pills" + ], + "image_footnote": [], + "bbox": [ + 299, + 268, + 491, + 349 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e4ac3ae30df5329201be0cdaf3623bc2891437791451964469ff21055030eba4.jpg", + "image_caption": [ + "(c)Trash Collection", + "Figure 6. Visualization of expert activation distribution per task for different object types. The x-axis represents the three object groups (target objects, surrounding objects, and background objects), where each group contains eight sub-bars corresponding to different experts selected. The left y-axis indicates the activation count, showing how frequently experts are selected for each object type. The right y-axis represents the density of experts per token, illustrating the average number of experts assigned per token in each object category. The dashed line represents the average expert density per token across different object groups, highlighting the dynamic expert selection mechanism of DGMoE." + ], + "image_footnote": [], + "bbox": [ + 504, + 268, + 697, + 349 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f7f92d7595ba46a56735707137bb573cfa0fb6ef1fe81f744e437838fd8f28cb.jpg", + "image_caption": [ + "(d) Clean Up" + ], + "image_footnote": [], + "bbox": [ + 709, + 268, + 901, + 349 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Analysis on DGMoE", + "text_level": 1, + "bbox": [ + 89, + 486, + 284, + 502 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The DGMoE is designed to enhance model efficiency by selectively activating experts based on task complexity and their corresponding object types. In contrast, supporting objects and background objects, which provide contextual spatial information, activate fewer experts compared to target objects. Notably, target objects tend to activate experts 4, 7, and 8 more frequently, supporting objects prefer to select experts 6, 7, and 8, whereas background objects favor experts 2, 3, and 5. This demonstrates that DGMoE not only dynamically allocates computational resources based on object relevance but also selectively routes information to specialized experts, enhancing efficiency and adaptability across different object categories.", + "bbox": [ + 88, + 515, + 483, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Furthermore, we quantitatively measure the average expert density per token across different tasks, also as shown in Figure 6. The proposed DGMoE achieves consistently lower expert activation densities compared to vanilla top- $k$ MoE, which assigns tokens to a fixed number of experts, where $k$ is typically an integer greater than one. The average number of activated experts per token is 1.219, 1.229, 1.225, and 1.227 for Open Drawing, Sorting Pills,Trash Collection, and Clean Up, respectively. Such dynamic and sparse activation directly translates into significantly reduced forward-pass computation, making DGMoE especially suitable for deployment on resource-constrained edge devices.", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 486, + 632, + 502 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduced FedVLA, a novel federated vision-language action learning framework that preserves privacy and effectively balances data privacy and task performance. To address the challenges of task heterogeneity and efficient knowledge aggregation in decentralized settings, we proposed three key components: the Instruction-Oriented Scene-Parsing module, which enhances feature extraction by decomposing raw observations into structured object-level representations guided by task instructions. The Dual Gating Mixture-of-Experts, which introduces self-aware experts capable of bidirectional token selection, optimizing computational efficiency while preserving task performance, and the Expert-Driven Aggregation strategy, which dynamically assigns aggregation weights based on expert similarity across clients. Extensive experiments in both simulated and real-world environments demonstrate that FedVLA achieves performance comparable to centralized training while ensuring privacy protection. We believe these results may contribute to the development of more generalizable and scalable federated learning frameworks for robotic manipulation.", + "bbox": [ + 511, + 518, + 906, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This research was supported by Guangdong Basic and Applied Basic Research Foundation (No. 2024A1515011774).", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6911", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3", + "[2] Martin Albrecht, Melissa Chase, Hao Chen, Jintai Ding, Shafi Goldwasser, Sergey Gorbunov, Shai Halevi, Jeffrey Hoffstein, Kim Laine, Kristin Lauter, et al. Homomorphic encryption standard. Protecting privacy through homomorphic encryption, pages 31-62, 2021. 3", + "[3] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. Openflamingo: An opensource framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. 2", + "[4] Denis Blessing, Onur Celik, Xiaogang Jia, Moritz Reuss, Maximilian Li, Rudolf Lioutikov, and Gerhard Neumann. Information maximizing curriculum: A curriculum-based approach for learning versatile skills. Advances in Neural Information Processing Systems, 36:51536-51561, 2023. 3", + "[5] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818, 2023. 1, 2", + "[6] Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa, Soravit Changpinyo, Jialin Wu, Carlos Riquelme Ruiz, Sebastian Goodman, Xiao Wang, Yi Tay, et al. Pali-x: On scaling up a multilingual vision and language model. arXiv preprint arXiv:2305.18565, 2023. 2", + "[7] Zitian Chen, Yikang Shen, Mingyu Ding, Zhenfang Chen, Hengshuang Zhao, Erik G Learned-Miller, and Chuang Gan. Mod-squad: Designing mixtures of experts as modular multi-task learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11828-11837, 2023. 2", + "[8] Damai Dai, Chengqi Deng, Chenggang Zhao, RX Xu, Huazuo Gao, Deli Chen, Jiashi Li, Wangding Zeng, Xingkai Yu, Y Wu, et al. Deepseekmoe: Towards ultimate expert specialization in mixture-of-experts language models. arXiv preprint arXiv:2401.06066, 2024. 3", + "[9] Cynthia Dwork. Differential privacy. In International colloquium on automata, languages, and programming, pages 1-12. Springer, 2006. 3", + "[10] Oded Goldreich. Secure multi-party computation. Manuscript. Preliminary version, 78(110):1-108, 1998. 3", + "[11] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. Neural computation, 3(1):79-87, 1991. 3", + "[12] Shen Jiang, Guanghui Zhu, Yue Wang, Chunfeng Yuan, and Yihua Huang. Automatic multi-task learning framework with neural architecture search in recommendations. In Proceedings of the 30th ACM SIGKDD Conference on Knowl" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "edge Discovery and Data Mining, pages 1290-1300, 2024. 2", + "[13] Siddharth Karamcheti, Suraj Nair, Ashwin Balakrishna, Percy Liang, Thomas Kollar, and Dorsa Sadigh. Prismatic vlms: Investigating the design space of visually-conditioned language models. arXiv preprint arXiv:2402.07865, 2024. 2", + "[14] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, Quan Vuong, Thomas Kollar, Benjamin Burchfiel, Russ Tedrake, Dorsa Sadigh, Sergey Levine, Percy Liang, and Chelsea Finn. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 1, 2", + "[15] Xinghang Li, Minghuan Liu, Hanbo Zhang, Cunjun Yu, Jie Xu, Hongtao Wu, Chilam Cheang, Ya Jing, Weinan Zhang, Huaping Liu, et al. Vision-language foundation models as effective robot imitators. arXiv preprint arXiv:2311.01378, 2023. 2", + "[16] Bin Lin, Zhenyu Tang, Yang Ye, Jiaxi Cui, Bin Zhu, Peng Jin, Junwu Zhang, Munan Ning, and Li Yuan. Moe-llava: Mixture of experts for large vision-language models. arXiv preprint arXiv:2401.15947, 2024.3", + "[17] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2", + "[18] Hao Liu, Wilson Yan, Matei Zaharia, and Pieter Abbeel. World model on million-length video and language with ringattention. arXiv e-prints, pages arXiv-2402, 2024. 2", + "[19] Yang Liu, Weixing Chen, Yongjie Bai, Xiaodan Liang, Guanbin Li, Wen Gao, and Liang Lin. Aligning cyber space with physical world: A comprehensive survey on embodied ai. arXiv preprint arXiv:2407.06886, 2024. 1", + "[20] Yuxiang Lu, Suizhi Huang, Yuwen Yang, Shalayiding Sirejiding, Yue Ding, and Hongtao Lu. Fedhca2: Towards hetero-client federated multi-task learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5599-5609, 2024. 2", + "[21] Ajay Mandlekar, Danfei Xu, Josiah Wong, Soroush Nasiriyani, Chen Wang, Rohun Kulkarni, Li Fei-Fei, Silvio Savarese, Yuke Zhu, and Roberto Martin-Martin. What matters in learning from offline human demonstrations for robot manipulation. arXiv preprint arXiv:2108.03298, 2021. 6", + "[22] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 1, 2, 3, 6", + "[23] David Nadeau and Satoshi Sekine. A survey of named entity recognition and classification. *Lingvisticae Investigationes*, 30(1):3-26, 2007. 4", + "[24] Abby O'Neill, Abdul Rehman, Abhiram Maddukuri, Abhishek Gupta, Abhishek Padalkar, Abraham Lee, Acorn Pooley, Agrim Gupta, Ajay Mandlekar, Ajinkya Jain, et al. Open x-embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collaboration 0. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 6892–6903. IEEE, 2024. 2" + ], + "bbox": [ + 517, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6912", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 4", + "[26] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 2", + "[27] Rejin Varghese and M Sambath. Yolov8: A novel object detection algorithm with enhanced performance and robustness. In 2024 International Conference on Advances in Data Engineering and Intelligent Computing Systems (ADICS), pages 1-6. IEEE, 2024. 4", + "[28] Lirui Wang, Xinlei Chen, Jialiang Zhao, and Kaiming He. Scaling proprioceptive-visual learning with heterogeneous pre-trained transformers. Advances in Neural Information Processing Systems, 37:124420-124450, 2025. 1, 2, 6", + "[29] Jiahui Xu, Lu Sun, and Dengji Zhao. Mone: Mixture-of-masked-experts for efficient multi-task recommendation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2527-2531, 2024. 2", + "[30] Tianhe Yu, Deirdre Quillen, Zhanpeng He, Ryan Julian, Karol Hausman, Chelsea Finn, and Sergey Levine. Metaworld: A benchmark and evaluation for multi-task and meta reinforcement learning. In Conference on robot learning, pages 1094-1100. PMLR, 2020. 6", + "[31] Michal Zawalski, William Chen, Karl Pertsch, Oier Mees, Chelsea Finn, and Sergey Levine. Robotic control via embodied chain-of-thought reasoning. arXiv preprint arXiv:2407.08693, 2024. 2", + "[32] Kaiwen Zhou and Xin Eric Wang. Fedvln: Privacy-preserving federated vision-and-language navigation. In European Conference on Computer Vision, pages 682-699. Springer, 2022. 3", + "[33] Weiming Zhuang, Yonggang Wen, Lingjuan Lyu, and Shuai Zhang. Mas: Towards resource-efficient federated multipletask learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23414-23424, 2023. 2" + ], + "bbox": [ + 91, + 92, + 480, + 694 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6913", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/FedVLA_ Federated Vision-Language-Action Learning with Dual Gating Mixture-of-Experts for Robotic Manipulation/9830aa12-506c-4606-9232-07331ac373a7_model.json b/2025/FedVLA_ Federated Vision-Language-Action Learning with Dual Gating Mixture-of-Experts for Robotic Manipulation/9830aa12-506c-4606-9232-07331ac373a7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8419cc9d0edcb4b8aa1abfca8ed45add5c792fd7 --- /dev/null +++ b/2025/FedVLA_ Federated Vision-Language-Action Learning with Dual Gating Mixture-of-Experts for Robotic Manipulation/9830aa12-506c-4606-9232-07331ac373a7_model.json @@ -0,0 +1,2398 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.134, + 0.13, + 0.868, + 0.177 + ], + "angle": 0, + "content": "FedVLA: Federated Vision-Language-Action Learning with Dual Gating Mixture-of-Experts for Robotic Manipulation" + }, + { + "type": "text", + "bbox": [ + 0.267, + 0.203, + 0.725, + 0.294 + ], + "angle": 0, + "content": "Cui Miao\\(^{1}\\) Tao Chang\\(^{1}\\) Meihan Wu\\(^{1}\\) Hongbin Xu\\(^{2}\\) Chun Li\\(^{3}\\) Ming Li\\(^{4*}\\) Xiaodong Wang\\(^{1}\\) National University of Defense Technology \n\\(^{2}\\)Bytedance Seed Shenzhen MSU-BIT University \nLong Laboratory of Artificial Intelligence and Digital Economics" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.343 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.358, + 0.485, + 0.784 + ], + "angle": 0, + "content": "Vision-language-action (VLA) models have significantly advanced robotic manipulation by enabling robots to interpret language instructions for task execution. However, training these models often relies on large-scale user-specific data, raising concerns about privacy and security, which in turn limits their broader adoption. To address this, we propose FedVLA, the first federated VLA learning framework, enabling distributed model training that preserves data privacy without compromising performance. Our framework integrates task-aware representation learning, adaptive expert selection, and expert-driven federated aggregation, enabling efficient and privacy-preserving training of VLA models. Specifically, we introduce an Instruction-Oriented Scene-Parsing mechanism, which decomposes and enhances object-level features based on task instructions, improving contextual understanding. To effectively learn diverse task patterns, we design a Dual Gating Mixture-of-Experts (DGMoE) mechanism, where not only input tokens but also self-aware experts adaptively decide their activation. Finally, we propose an Expert-Driven Aggregation strategy at the federated server, where model aggregation is guided by activated experts, ensuring effective cross-client knowledge transfer. Extensive simulations and real-world robotic experiments demonstrate the effectiveness of our proposals. Notably, DGMoE significantly improves computational efficiency compared to its vanilla counterpart, while FedVLA achieves task success rates comparable to centralized training, effectively preserving data privacy." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.809, + 0.222, + 0.825 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.834, + 0.484, + 0.88 + ], + "angle": 0, + "content": "Vision-language-action (VLA) models, which integrate visual perception, linguistic understanding, and robotic control, have significantly enhanced robotic manipulation by" + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.327, + 0.697, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.498, + 0.691, + 0.512 + ], + "angle": 0, + "content": "(a) Centralized VLA learning" + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.327, + 0.891, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.724, + 0.498, + 0.873, + 0.512 + ], + "angle": 0, + "content": "(b) Federated VLA learning" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.523, + 0.907, + 0.62 + ], + "angle": 0, + "content": "Figure 1. Comparison between centralized and federated VLA training. Centralized training collects all user data on the cloud, raising privacy concerns due to potential exposure of sensitive information. Our federated VLA framework enables decentralized training on user devices, preserving privacy while utilizing expert-driven aggregation to enhance model generalization across diverse tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.655, + 0.907, + 0.776 + ], + "angle": 0, + "content": "enabling robots to comprehend and execute complex tasks through natural language instructions [5, 14, 28]. However, training these models requires extensive indoor datasets that capture diverse, user-specific environments, raising serious concerns about data privacy. While recent advancements have improved the capability of VLA models in robotic systems, the issue of privacy leakage remains largely unexplored [19]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Federated learning (FL) provides an effective privacy-preserving solution for training VLA models [22]. Unlike traditional centralized training, which requires aggregating all user data on a central server, FL enables distributed model training across multiple clients without transferring raw data. As illustrated in Figure 1a, centralized VLA training collects and processes user data in the cloud, posing risks of privacy leakage, such as exposure of personal pref" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.235, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6904" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "erences and sensitive health conditions. In contrast, figure 1b presents a federated VLA learning paradigm, where local models are trained on user devices, and only model updates are transmitted to the central server. This decentralized paradigm mitigates privacy risks while allowing collaborative model optimization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.186, + 0.483, + 0.397 + ], + "angle": 0, + "content": "Existing federated learning approaches are not well-suited for VLA learning, where robotic manipulation tasks exhibit substantial task heterogeneity across different clients. Typical FL methods, such as FedAvg [22], naively aggregate models across tasks, considering only shared synergies while overlooking their fundamental distinctions. Although recent works attempt to address this issue through merge-and-split training [33] or decoupled model aggregation [20], they remain limited to single-modal settings, assuming a uniform input modality across tasks. In contrast, VLA models operate in multi-modal environments, requiring the joint processing of visual observations, language instructions, and robotic actions, which significantly increases the complexity of federated training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.4, + 0.483, + 0.552 + ], + "angle": 0, + "content": "Additionally, mixture of experts (MoE) has shown promise in multi-task learning by assigning specialized experts to different tasks [7, 12, 29]. However, existing methods typically select a fixed number of experts, lacking adaptability to tasks of varying complexity. This rigidity makes them inefficient for resource-constrained clients, where optimizing inference speed is critical for real-time robotic responses. These limitations highlight the need for a task-adaptive and flexible FL framework, specifically designed for multi-modal robotic learning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.483, + 0.903 + ], + "angle": 0, + "content": "In this paper, we introduce FedVLA, a novel federated vision-language-action (VLA) learning framework with Dual Gating MoE, designed to preserve privacy while maintaining high task performance, as illustrated in Figure 1b. To address task heterogeneity and enable effective knowledge aggregation in federated settings, we propose three key components. First, we introduce an Instruction-Oriented Scene-Parsing (IOSP) module to enhance task-aware feature extraction in decentralized environments, where raw images across clients exhibit diverse backgrounds and object distributions. IOSP decomposes observation images into object-level representations guided by task instructions and leverages vision-language alignment techniques to improve contextual understanding. Second, to enhance computational efficiency in federated learning, we develop a Dual Gating MoE (DGMoE) mechanism that enables adaptive knowledge routing. Unlike conventional MoE, which assigns tokens to a fixed number of experts, DGMoE introduces self-aware experts that can accept or reject token assignments, enabling a bidirectional selection process. This dual gating mechanism improves efficiency while preserving task performance. Lastly, to effectively aggregate similar knowledge from diverse client models, we propose the" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "Expert-Driven Aggregation (EDA) strategy, which dynamically assigns aggregation weights based on expert similarity across clients. By prioritizing updates from clients that activate similar experts, EDA ensures that semantically aligned knowledge is more prominently integrated at the server side, resulting in a more coherent and specialized global model." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.184, + 0.905, + 0.213 + ], + "angle": 0, + "content": "Our main contributions in this work can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.215, + 0.905, + 0.319 + ], + "angle": 0, + "content": "- We propose FedVLA, the first privacy-preserving federated learning framework for VLA training, ensuring that user data remains private and never shared externally. Extensive experiments in both simulation and real-world environments demonstrate that FedVLA achieves performance comparable to centralized training while preserving data privacy." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.32, + 0.905, + 0.393 + ], + "angle": 0, + "content": "- We introduce the Dual Gating Mixture-of-Experts, where self-aware experts autonomously determine their participation, enabling bidirectional token selection to optimize computational efficiency while maintaining task performance." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.396, + 0.905, + 0.47 + ], + "angle": 0, + "content": "- We propose the Expert-Driven Aggregation strategy, which dynamically assigns aggregation weights based on expert similarity across clients, ensuring that semantically aligned knowledge is more effectively integrated, thereby enhancing generalization across heterogeneous tasks." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.215, + 0.905, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.486, + 0.663, + 0.501 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.512, + 0.797, + 0.527 + ], + "angle": 0, + "content": "2.1. Vision-Language-Action Models" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.534, + 0.905, + 0.821 + ], + "angle": 0, + "content": "Vision-Language Models (VLMs), which are pre-trained on large-scale datasets, have demonstrated remarkable capabilities in understanding and generating multimodal data [17, 18, 26]. Inspired by VLMs, Vision-Language-Action (VLA) models extend their capabilities by integrating robotic action prediction. For instance, RoboFlamingo [15], RT-2 [5], and OpenVLA [14] are all built upon VLMs such as OpenFlamingo [3], PaLI-X [6], and Prismatic [13], respectively. Recent work has begun exploring the training of generalist robots [28], followed by fine-tuning on diverse robot datasets, enabling generalization to novel instructions, unseen objects, and distribution shifts [31]. This progress relies heavily on the availability of large-scale public datasets, such as Open-X Embodiment dataset [24]. During the finetuning phase, samples from real-world scenarios are required, which are often generated by human activities and distributed across users' home environments. However, these samples frequently expose privacy-sensitive information." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.833, + 0.768, + 0.849 + ], + "angle": 0, + "content": "2.2. Privacy-Preserving Training" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Traditional centralized learning methods require uploading data from multiple sources, which poses significant privacy risks. To mitigate these concerns, researchers have pro" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6905" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.91, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.334, + 0.908, + 0.435 + ], + "angle": 0, + "content": "Figure 2. Illustration of the proposed FedVLA framework and its key component. (a) An overview of the FedVLA, which consists of multiple clients and a server. Each client processes vision-language-action inputs through the Stem, Trunk, and Head modules. Clients send their updated trunk parameters to the server, where Expert-Driven Aggregation selectively integrates expert-specific updates on server. The global trunk module is then distributed to clients for the next training round. (b) Details of the proposed Dual Gating MoE module. The input tokens first pass through the token-level gate \\(G_{t}\\), computing the gating scores for each expert with residual scores from the previous MoE layer. Only the computed scores surpass the expert-level gate \\(G_{e}\\), activating the corresponding expert to process the token. Finally, the expert outputs are combined to generate the final output tokens." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.459, + 0.486, + 0.702 + ], + "angle": 0, + "content": "posed various privacy-preserving techniques, including federated learning (FL) [22], differential privacy [9], secure multi-party computation [10], and homomorphic encryption [2]. However, Most privacy-preserving training techniques have been developed for single-modality tasks, and seldomly study the data privacy of embodied AI [32]. In robotic manipulation, the models process multimodal sensor data, including images, language instructions, and action sequences, which can contain sensitive user interaction data. Unlike static datasets, robotic systems continuously collect and process dynamic, task-specific information, making privacy protection more complex. In this context, FL as a solution by allowing decentralized model training without sharing raw data. However, FL for robotic manipulation remains challenging due to task heterogeneity and the need for efficient model aggregation across clients." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.712, + 0.406, + 0.729 + ], + "angle": 0, + "content": "2.3. Mixture of Experts in Large Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Mixture of experts [4, 11] is an architecture where multiple experts are trained collaboratively, with each expert responsible for a specific portion of the input space. An MoE consists of two components: a routing mechanism and a set of experts. The experts can either be part of a feedforward network (FFN) or the entire FFN layer, and the routing mechanism allows each token to select experts for processing. By activating a specific number of experts, MoE can reduce resource consumption and improve inference speed. This feature makes MoE particularly effective when integrated into large models, as seen in models like LLaVA-MoE [16]," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.459, + 0.907, + 0.52 + ], + "angle": 0, + "content": "DeepSeekMoE [8], and GPT-4 [1]. In this work, we apply MoE to a generalist robot VLA model and introduce a self-aware variant that can adaptively decide whether to be selected, significantly improving deployment efficiency." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.531, + 0.611, + 0.547 + ], + "angle": 0, + "content": "3. FedVLA" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.557, + 0.909, + 0.708 + ], + "angle": 0, + "content": "We aim to train a vision-language-action model in a federated learning (FL) framework, ensuring privacy preservation while enabling task generalization across clients. Formally, given \\(N\\) clients \\(\\mathbb{C} = \\{C_1,C_2,\\dots,C_N\\}\\), \\(N\\) local tasks \\(\\mathbb{T} = \\{T_1,T_2,\\dots,T_N\\}\\), and \\(N\\) local datasets \\(\\mathbb{D} = \\{\\mathbb{D}_1,\\mathbb{D}_2,\\dots,\\mathbb{D}_N\\}\\), each client \\(C_i\\) is associated with a local task \\(T_{i}\\) and a corresponding dataset \\(\\mathbb{D}_i\\). The dataset \\(\\mathbb{D}_i\\) contains visual perception image \\(o_i\\), language instruction \\(l_{i}\\) of the task, proprioception \\(s_i\\) of the robot, and action sequence \\(a_{i}\\) executed by the robot:" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.71, + 0.907, + 0.727 + ], + "angle": 0, + "content": "\\[\n\\mathbb {D} _ {i} = \\left\\{\\left(o _ {i}, l _ {i}, s _ {i}, a _ {i}\\right) _ {j} | j = 1, 2, \\dots , M \\right\\}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.731, + 0.908, + 0.776 + ], + "angle": 0, + "content": "where \\(M\\) is the number of samples in dataset \\(\\mathbb{D}_i\\). To train a federated VLA model \\(\\theta_G\\), the overall federated optimization objective can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.778, + 0.907, + 0.822 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta_ {G}} \\sum_ {i = 1} ^ {N} w _ {i} \\sum_ {(o _ {i}, l _ {i}, a _ {i}, s _ {i}) \\in \\mathbb {D} _ {i}} \\mathcal {L} _ {i} (f _ {i} (\\theta_ {G}; o _ {i}, l _ {i}, s _ {i}), a _ {i}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.909, + 0.9 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_i\\) is the Huber loss function computed over each client model's local dataset \\(\\mathbb{D}_i\\), \\(f_i\\) represents the local client model's predicted action sequence, and \\(w_i\\) denotes the weight assigned to each client during aggregation, which is further detailed in Section 3.3." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6906" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.482, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.183, + 0.483, + 0.242 + ], + "angle": 0, + "content": "Figure 3. Instruction-Oriented Scene-Parsing module, which decomposes observations into object-level representations guided by task instructions, enhancing task-aware feature extraction through vision-language alignment." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.266, + 0.484, + 0.448 + ], + "angle": 0, + "content": "The overview of our FedVLA framework is shown in Figure 2. The clients train their models locally and communicate with the server per round until the model converges. The following sections introduce our key proposals: (1) Instruction-Oriented Scene-Parsing (Section 3.1) that decomposes the observation image into object-level embeddings guided by task instructions. (2) Dual Gating Mixture-of-Experts (Section 3.2) that can adaptively select a varying number of experts for tokens based on the relevance to the specific task. (3) Expert-Driven Aggregation (Section 3.3) that dynamically aggregates the weights based on expert similarities across different clients in the server." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.456, + 0.403, + 0.472 + ], + "angle": 0, + "content": "3.1. Instruction-Oriented Scene-Parsing" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.478, + 0.484, + 0.568 + ], + "angle": 0, + "content": "To ensure the robot focuses on task-relevant information while maintaining contextual awareness to avoid collisions, we decouple objects in the scene based on the specific robotic task. To achieve this, we propose an Instruction-Oriented Scene-Parsing (IOSP) module to represent objects within the task scene." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.57, + 0.483, + 0.795 + ], + "angle": 0, + "content": "As illustrated in Figure 3, given an instruction and an image, we first extract Target Objects (TOs) from the instruction using named entity recognition [23]. Simultaneously, we employ YOLOv8 [27] to detect and classify Foreground Objects (FOs) and Background Objects (BOs) in the image. The scene objects are categorized into three groups based on the cosine similarity measured by a pre-trained CLIP model [25] between the instruction and object names, following these matching rules: (1) Target Objects (TOs): Objects explicitly mentioned in the instruction and directly involved in the action. (2) Surrounding Objects (SOs): Foreground objects not in TOs, providing spatial and contextual cues. (3) Background Objects (BOs): Static environmental elements ensuring scene consistency but not directly affecting the task." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Then, we encode the grouped objects using the CLIP model. The text encoder \\(H\\) maps object names from TOs, SOs, and BOs into textual embeddings \\(H(\\mathrm{TOs}) \\in \\mathbb{R}^{t \\times D}\\), \\(H(\\mathrm{BOs}) \\in \\mathbb{R}^{t \\times D}\\), and \\(H(\\mathrm{SOs}) \\in \\mathbb{R}^{t \\times D}\\). Meanwhile, the vision encoder \\(P\\) processes the entire image into \\(P(X) \\in \\mathbb{R}^{t \\times D}\\). We assign image tokens to object groups by computing the cosine similarity between token and text embed" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.198 + ], + "angle": 0, + "content": "dings, selecting the top eight tokens per group. To enhance intra-group feature representation, each group is processed through an MoE module, producing refined image token embeddings. The enhanced tokens are then concatenated with the remaining tokens for subsequent processing. This IOSP module enables the model to focus on critical objects, improving task-specific manipulation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.205, + 0.799, + 0.221 + ], + "angle": 0, + "content": "3.2. Dual Gating Mixture-of-Experts" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.227, + 0.905, + 0.317 + ], + "angle": 0, + "content": "In this section, we introduce Dual Gating Mixture-of-Expert (DGMoE) mechanism that supports adaptive knowledge routing. Different from vanilla MoE, which assigns a fixed number of experts to each token, our DGMoE can automatically determine the acceptance or rejection of the token assignments under the Self-Aware Expert." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.318, + 0.906, + 0.574 + ], + "angle": 0, + "content": "As shown in Figure 2, each DGMoE layer is comprised of \\(K\\) self-aware experts \\(\\mathbb{E} = \\{E_1,E_2,\\dots ,E_K\\}\\). To enable the automatic token-driven expert selection, we introduce two gating mechanisms: Token-side gating mechanism \\(G_{t}\\) and Expert-side gating mechanism \\(G_{e}\\). (1) For Token-side gating, \\(G_{t}\\) serves as the gating module on the token side, comprised of a soft router module and a gating residual module. The soft router first estimates the selection scores for the hidden inputs, indicating the token preference for different experts. Then the estimated scores from previous layer are incorporated with the scores of current layer, allowing the tokens to inherit the selection prior of the most suitable experts from previous layers. (2) For Expert-side gating, \\(G_{e}\\) can autonomously accept or reject tokens based on their acceptance threshold. The selection scores predicted by \\(G_{t}\\) is filtered with a learnable threshold, and only the experts with the most relevant tokens are activated." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.575, + 0.906, + 0.62 + ], + "angle": 0, + "content": "Given an input token \\( x \\), the token-side gate \\( G_{t} \\) first computes the scores at the \\( j \\)-th layer by combining the soft router output and the gating residual module:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.626, + 0.907, + 0.667 + ], + "angle": 0, + "content": "\\[\nG _ {t} ^ {j} (x) = \\left\\{ \\begin{array}{l l} W _ {t} ^ {j} x, & \\text {i f} j = 1, \\\\ W _ {t} ^ {j} x + W _ {t} ^ {g} G _ {t} ^ {(j - 1)} (x), & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.906, + 0.795 + ], + "angle": 0, + "content": "where \\( W_{t}^{j}\\in \\mathbb{R}^{D} \\) is the trainable weight matrix for soft router, and \\( W_{t}^{g}\\in \\mathbb{R}^{N\\times N} \\) is a trainable transformation matrix designed to aggregate the scores from the previous layer and the current layer, ensuring smooth expertise knowledge transitions across layers. Then, the token-to-expert selection scores \\( s_t(x) \\) are obtained by normalizing \\( G_{t}(x) \\) with a softmax function, which determines the probability distribution over the available experts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.905, + 0.841 + ], + "angle": 0, + "content": "On the self-aware expert side, the expert-side gate \\( G_{e} \\) evaluates the token scores and determines the final expert activation scores \\( s_{e}(x) \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.612, + 0.848, + 0.905, + 0.866 + ], + "angle": 0, + "content": "\\[\ns _ {e} (x) = \\operatorname {s i g n} \\left(s _ {t} (x) - \\lambda W _ {e}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\( W_{e} \\) is a set of trainable self-aware parameters corresponding to the gating values of each expert, and \\( \\lambda \\) is a" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6907" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.15 + ], + "angle": 0, + "content": "scaling factor that modulates the gating threshold. In our setting, we set \\(\\lambda = 0.5\\) to balance expert selection, ensuring neither excessive expert activation nor overly strict filtering of tokens." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.152, + 0.484, + 0.197 + ], + "angle": 0, + "content": "Then, combining the dual gating mechanism of token-side and exper-side gating, the expert selection function is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.202, + 0.483, + 0.242 + ], + "angle": 0, + "content": "\\[\ng (x) = \\left\\{ \\begin{array}{l l} s _ {t} (x), & \\text {i f} s _ {e} (x) > 0, \\\\ 0, & \\text {i f} s _ {e} (x) \\leq 0, \\end{array} \\right. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.249, + 0.484, + 0.292 + ], + "angle": 0, + "content": "where the final output token \\(y\\) of the DGMoE layer can be computed as the weighted sum of the selected expert outputs:" + }, + { + "type": "equation", + "bbox": [ + 0.216, + 0.283, + 0.483, + 0.323 + ], + "angle": 0, + "content": "\\[\ny = \\sum_ {i = 1} ^ {K} g _ {i} (x) E _ {i} (x). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.326, + 0.484, + 0.387 + ], + "angle": 0, + "content": "With the help of self-aware expert gating mechanism, our DGMoE can dynamically activate the most relevant experts, reducing computational overhead while maintaining task-relevant feature learning." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.395, + 0.339, + 0.411 + ], + "angle": 0, + "content": "3.3. Expert-Driven Aggregation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.416, + 0.484, + 0.522 + ], + "angle": 0, + "content": "To mitigate performance degradation caused by task heterogeneity, we propose an Expert-Driven Aggregation (EDA) strategy, which aggregates model parameters based on the similarity of expert selections across clients. And as shown in Figure 2, we selectively aggregates only the trunk module, while the stem and head modules remain personalized and are excluded from aggregation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.522, + 0.484, + 0.615 + ], + "angle": 0, + "content": "Given a trunk module with \\(L\\) layers, where each layer contains \\(K\\) experts, each client \\(C_i\\) records an expert selection matrix \\(\\mathbf{V}_i\\in \\mathbb{R}^{L\\times K}\\) during each round through DGMoE, where each element \\(V_{i}^{(l,k)}\\) represents the number of times expert \\(k\\) is activated in the \\(l\\)-th layer. We extract the expert selection vector for layer \\(l\\) of client \\(C_i\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.621, + 0.483, + 0.643 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {i} ^ {(l)} = \\left(V _ {i} ^ {(l, 1)}, V _ {i} ^ {(l, 2)}, \\dots , V _ {i} ^ {(l, K)}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.649, + 0.484, + 0.695 + ], + "angle": 0, + "content": "Then we compute the pairwise similarity between clients based on expert selection vectors. The similarity between clients \\( C_i \\) and \\( C_j \\) at layer \\( l \\) is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.216, + 0.702, + 0.483, + 0.746 + ], + "angle": 0, + "content": "\\[\ns _ {i, j} ^ {(l)} = \\frac {\\mathbf {v} _ {i} ^ {(l)} \\cdot \\mathbf {v} _ {j} ^ {(l)}}{\\| \\mathbf {v} _ {i} ^ {(l)} \\| \\| \\mathbf {v} _ {j} ^ {(l)} \\|\\}.} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.752, + 0.483, + 0.782 + ], + "angle": 0, + "content": "Using this similarity, we determine the aggregation weight \\( w_{l,i} \\) for each client \\( C_i \\) at layer \\( l \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.79, + 0.483, + 0.835 + ], + "angle": 0, + "content": "\\[\nw _ {l, i} = \\frac {\\sum_ {C _ {j} \\in C} s _ {i , j} ^ {(l)}}{\\sum_ {C _ {j} \\in C} \\sum_ {C _ {i} \\in C} s _ {i , j} ^ {(l)}}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "This weighting strategy ensures that clients with higher expert selection similarity contribute more to each other's model updates, allowing for more effective and task-aligned aggregation across task-heterogeneous clients." + }, + { + "type": "code_caption", + "bbox": [ + 0.515, + 0.092, + 0.728, + 0.106 + ], + "angle": 0, + "content": "Algorithm 1 FedVLA: Clients" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.108, + 0.907, + 0.169 + ], + "angle": 0, + "content": "Input: Each client \\( C_i \\) has a local task \\( T_i \\) and dataset \\( D_i = \\{o_i, l_i, s_i, a_i\\} \\), along with local parameters \\( \\phi_i \\), \\( w_i \\), and \\( \\theta_i \\) corresponding to the stem, trunk, and head modules, respectively." + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.186, + 0.707, + 0.2 + ], + "angle": 0, + "content": "1: Local_Train \\((C_i, \\phi_i, \\theta_i)\\):" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.202, + 0.8, + 0.216 + ], + "angle": 0, + "content": "2: for each local epoch \\( e \\) from 1 to \\( E_{c} \\) do" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.217, + 0.884, + 0.232 + ], + "angle": 0, + "content": "3: for each batch \\((o_i, l_i, s_i, a_i)\\) sampled from \\(D_i\\) do" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.237, + 0.873, + 0.251 + ], + "angle": 0, + "content": "4: // Step 1: Instruction-Oriented Scene-Parsing" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.253, + 0.753, + 0.266 + ], + "angle": 0, + "content": "5: TOs, SOs, BOs \\(\\leftarrow \\{o_i, l_i\\}\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.268, + 0.875, + 0.282 + ], + "angle": 0, + "content": "6: \\(F_{T O s}, F_{S O s}, F_{B O s} \\gets \\mathrm{CLIP}(T O s, S O s, B O s)\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.284, + 0.69, + 0.298 + ], + "angle": 0, + "content": "7: \\(F_{O}\\gets \\mathrm{CLIP}(o_{i})\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.3, + 0.83, + 0.315 + ], + "angle": 0, + "content": "8: \\(X_{\\mathrm{TO}} \\gets \\mathrm{topk}(\\mathrm{cosine}(F_{\\mathrm{TOs}}, F_O), F_O)\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.316, + 0.828, + 0.331 + ], + "angle": 0, + "content": "9: \\(X_{\\mathrm{SO}} \\gets \\mathrm{topk}(\\mathrm{cosine}(F_{\\mathrm{SOs}}, F_O), F_O)\\)" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.333, + 0.829, + 0.347 + ], + "angle": 0, + "content": "10: \\(X_{\\mathrm{BO}} \\gets \\mathrm{topk}(\\mathrm{cosine}(F_{\\mathrm{BOs}}, F_O), F_O)\\)" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.349, + 0.825, + 0.364 + ], + "angle": 0, + "content": "11: \\(X_{\\mathrm{in}} = \\mathrm{concat}(X_{\\mathrm{TO}},X_{\\mathrm{SO}},X_{\\mathrm{BO}},s_i)\\)" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.369, + 0.823, + 0.383 + ], + "angle": 0, + "content": "12: // Step 2: Forward pass with DGMoE" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.385, + 0.715, + 0.399 + ], + "angle": 0, + "content": "13: Let \\(x \\in MHA(X_{\\mathrm{in}})\\)" + }, + { + "type": "text", + "bbox": [ + 0.576, + 0.4, + 0.869, + 0.416 + ], + "angle": 0, + "content": "\\(s_t(x) = \\mathrm{softmax}\\bigl (W_t(x)\\bigr)\\) // token-gate" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.418, + 0.874, + 0.432 + ], + "angle": 0, + "content": "14: \\(s_e(x) = \\mathrm{sign}(s_t(x) - \\lambda W_e)\\) // expert-gate" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.434, + 0.904, + 0.448 + ], + "angle": 0, + "content": "15: \\(g(x) = s_t(x)\\cdot \\mathbb{I}(s_e(x) > 0)\\) // expert activation" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.45, + 0.852, + 0.488 + ], + "angle": 0, + "content": "16: \\(\\mathbf{v}^{(l)},y = \\sum_{k = 1}^{K}g(x)E_k(x)\\) //final output" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.493, + 0.808, + 0.507 + ], + "angle": 0, + "content": "17: // Step 3: Model parameter updates" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.508, + 0.772, + 0.522 + ], + "angle": 0, + "content": "18: \\(\\mathcal{L}_i = \\mathrm{Loss}(f(\\phi_i, \\theta_i, w_i), a_i)\\)" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.524, + 0.813, + 0.539 + ], + "angle": 0, + "content": "19: \\((\\phi_i,\\theta_i,w_i)\\gets (\\phi_i,\\theta_i,w_i) - \\eta \\nabla \\mathcal{L}_i\\)" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.54, + 0.613, + 0.552 + ], + "angle": 0, + "content": "20: end for" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.555, + 0.597, + 0.567 + ], + "angle": 0, + "content": "21: end for" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.569, + 0.724, + 0.583 + ], + "angle": 0, + "content": "22: return \\(\\theta, \\mathbf{v}^{(l)}\\) to the server." + }, + { + "type": "list", + "bbox": [ + 0.52, + 0.186, + 0.904, + 0.583 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.626, + 0.638, + 0.642 + ], + "angle": 0, + "content": "3.4. Algorithms" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.653, + 0.907, + 0.789 + ], + "angle": 0, + "content": "Clients. The overall workflow of our FedVLA is shown in Algorithm 1 and Algorithm 2. At the start of each training round \\( t \\), each client processes task-specific data using Instruction-Oriented Scene-Parsing to extract structured features, followed by Dual Gating MoE for adaptive expert selection. The updated trunk parameters and expert selection counts are then transmitted to the server for global aggregation. The detailed client-side process is described in Algorithm 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Server. On the server side, the server receives expert selection statistics and trunk updates from all participating clients and performs Expert-Driven Aggregation, which dynamically assigns aggregation weights based on expert selection similarities. The aggregated global trunk module is then redistributed to clients for the next training round. The process is elaborated in Algorithm 2." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6908" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.092, + 0.303, + 0.106 + ], + "angle": 0, + "content": "Algorithm 2 FedVLA: Server" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.108, + 0.484, + 0.155 + ], + "angle": 0, + "content": "Input: \\(T\\) is the number of training rounds. \\(\\theta\\) represents the global trunk parameters. Each client provides \\(\\mathbf{v}^{(l)}\\), the expert selection counts vector for each layer \\(l\\)." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.171, + 0.237, + 0.184 + ], + "angle": 0, + "content": "1: Server Execute:" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.186, + 0.333, + 0.2 + ], + "angle": 0, + "content": "2: for each round \\(t\\) from 1 to \\(T\\) do" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.202, + 0.333, + 0.216 + ], + "angle": 0, + "content": "3: // Expert-Driven Aggregation" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.218, + 0.301, + 0.231 + ], + "angle": 0, + "content": "4: for each trunk layer \\(l\\) do" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.232, + 0.373, + 0.247 + ], + "angle": 0, + "content": "5: for each pair of clients \\( C_i, C_j \\) do" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.247, + 0.477, + 0.276 + ], + "angle": 0, + "content": "6: \\(s_{i,j}^{(l)} = \\frac{\\mathbf{v}_i^{(l)}\\cdot\\mathbf{v}_j^{(l)}}{\\|\\mathbf{v}_i^{(l)}\\|\\|\\mathbf{v}_j^{(l)}\\|} //\\) expert selection similarity" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.278, + 0.473, + 0.31 + ], + "angle": 0, + "content": "7: \\(w^{(l,i)} = \\frac{\\sum_{j\\in C}s_{i,j}^{(l)}}{\\sum_{j\\in C}\\sum_{i\\in C}s_{i,j}^{(l)}}\\) // aggregation weight" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.311, + 0.207, + 0.322 + ], + "angle": 0, + "content": "8: end for" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.323, + 0.416, + 0.34 + ], + "angle": 0, + "content": "9: \\(\\theta^{(l)}\\gets \\sum_{i = 1}^{N}w^{(l,i)}\\cdot \\theta^{(l,i)}\\) //update \\(\\theta\\)" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.341, + 0.19, + 0.353 + ], + "angle": 0, + "content": "10: end for" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.355, + 0.36, + 0.37 + ], + "angle": 0, + "content": "11: Broadcast updated \\(\\theta\\) to all clients." + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.371, + 0.174, + 0.383 + ], + "angle": 0, + "content": "12: end for" + }, + { + "type": "list", + "bbox": [ + 0.098, + 0.171, + 0.477, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.414, + 0.224, + 0.431 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.44, + 0.485, + 0.637 + ], + "angle": 0, + "content": "Implementation Details. We validate our FedVLA in both simulation and real-world scenarios, where each client performs a single task, resulting in a non-iid data distribution. We employ the pretrained HPT [28] as the backbone of our VLA model and train it for 1,000 communication rounds between clients and the server, with each client performing local training for 5 epochs per round. In simulation, the model is trained with a learning rate of \\(5 \\times 10^{-6}\\), while in real-world settings, the learning rate is set to \\(2 \\times 10^{-5}\\). Each client trains locally with a batch size of 256 using the Adam optimizer. The experiments are conducted on a workstation equipped with 2 GeForce GTX 4090Ti GPUs, 48 Intel Xeon CPUs, and 128GB of memory." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.638, + 0.484, + 0.775 + ], + "angle": 0, + "content": "Evaluation. For evaluation, the success and failure of a trial are recorded as 1 and 0. Each task takes for 15 trials and repeats for 5 times. Since this is the first work on federated VLA learning, we compare FedVLA with two methods. The centralized training (Centralized) aggregates all data on a single server for joint training, serving as an upper bound in performance. The federated averaging (FedAvg) [22] is a strong FL baseline where models are trained locally on each client and simply averaged on the server." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.787, + 0.212, + 0.802 + ], + "angle": 0, + "content": "4.1. Simulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Environment. We utilize the MuJoCo engine [21] to simulate tasks from the Meta-World robot manipulation benchmark [30]. All tasks are performed using a Sawyer robot with a parallel gripper, positioned at the center of a dinner table. Visual data is captured by three fixed RGB cameras with a resolution of \\(128 \\times 128\\), providing corner, top, and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.228 + ], + "angle": 0, + "content": "shoulder views. For collision detection and dynamics simulation, we employ official physics engines to ensure accurate robotic interactions within the simulation environment. Tasks. We evaluate our FedVLA on four household tasks: Door Lock, Close Printer, Sweep Into, and Open Window. For each task, we collect approximately 30-80 episodes, each consisting of 40-100 steps. The initial pose of robot and environment layouts are randomized to assess the generalization ability of FedVLA." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.24, + 0.9, + 0.315 + ], + "angle": 0, + "content": "
MethodAvg. SuccessDoor LockCloseDrawerSweep IntoOpen Window
Centralized65.086.773.353.346.7
FedAvg51.766.773.340.026.7
FedVLA (Ours)63.380.080.053.340.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.565, + 0.326, + 0.855, + 0.339 + ], + "angle": 0, + "content": "Table 1. Simulation evaluation across four tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.356, + 0.907, + 0.568 + ], + "angle": 0, + "content": "Performance Comparison. The evaluation results in the simulation environment are presented in Table 1. FedVLA achieves an average success rate of \\(63.3\\%\\), closely approaching the \\(65.0\\%\\) of centralized training. Notably, in the Close Drawn task, FedVLA surpasses centralized training by a success rate margin of about \\(7.0\\%\\). Furthermore, FedVLA consistently outperforms FedAvg, which only achieves an average success rate of \\(51.7\\%\\). This performance gap can be attributed to FedAvg overlooking task heterogeneity in federated learning. In contrast, FedVLA explicitly accounts for task heterogeneity during aggregation by leveraging expert-driven model updates, ensuring that knowledge from diverse clients is effectively integrated to enhance generalization across different tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.577, + 0.64, + 0.592 + ], + "angle": 0, + "content": "4.2. Real-World" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.734 + ], + "angle": 0, + "content": "Hardware. For real-world experiments, we employ a UR3 robotic arm with six degrees of freedom, equipped with a one-DoF parallel gripper. A RealSense D435i RGB-D camera is mounted above the robot, capturing RGB images in real time at a resolution of \\(1280 \\times 720\\) with a frame rate of approximately \\(30\\mathrm{Hz}\\). To simulate real-world variations, we introduce moderate randomness in object placement and lighting conditions to better reflect practical deployment scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.871 + ], + "angle": 0, + "content": "Tasks. We collect real-world robotic demonstrations for household-related tasks, including Clean Up,Trash Collection, OpenDrawer,and Sorting Pills, as shown in Figure 4. For data collection, we employ an expert teaching method, using inverse kinematics to compute the robot poses corresponding to key point data while recording RGB images and pose data throughout the trajectory. Each task includes approximately 50 demonstrations, with each trajectory consisting of 20 to 80 steps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.9 + ], + "angle": 0, + "content": "Performance Comparison. The real-world experimental results are summarized in Table 2. FedVLA achieves an" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6909" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.48, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.27, + 0.485, + 0.368 + ], + "angle": 0, + "content": "Figure 4. Real-World Tasks. Each column represents a different task, while each row demonstrates keyframes of task execution. From left to right, the tasks are: Clean Up, where the robot wipes the trash from plate using a cloth. Trash Collection, in which it throws crumpled paper into a trash bin. Open Printer, which requires grasping and pulling a drawer open, and Sorting Pills, where pill bottles are placed into trays of designated color." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.395, + 0.483, + 0.668 + ], + "angle": 0, + "content": "average task success rate of \\(63.3\\%\\), closely matching the \\(63.4\\%\\) of centralized training. Notably, it attains the same success rates as the centralized model in complex tasks such as Trash Collection \\((46.7\\%)\\) and Sorting Pills \\((73.3\\%)\\). In contrast, FedAvg exhibits a significant performance drop across all tasks, with an average success rate of \\(53.3\\%\\), which is \\(10.0\\%\\) lower than our method. These results indicate that FedVLA effectively bridges the performance gap between federated and centralized training. While FedAvg merely averages model parameters across clients, it struggles to generalize effectively in VLA learning, where multimodal information—including textual instructions, scene images, and execution trajectories—is involved, requiring strong feature extraction capabilities. The proposed Dual Gating MoE and Expert-Driven Aggregation mechanisms enable the model to adaptively learn diverse patterns for robotic manipulation, enhancing its ability to handle complex tasks in federated settings." + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.679, + 0.478, + 0.751 + ], + "angle": 0, + "content": "
MethodAvg. SuccessClean UpTrash CollectionOpenDrawerSorting Pills
Centralized63.446.746.786.773.3
FedAvg53.346.740.060.066.7
FedVLA (Ours)63.353.346.780.073.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.141, + 0.762, + 0.434, + 0.775 + ], + "angle": 0, + "content": "Table 2. Real-world evaluation across four tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.803, + 0.255, + 0.818 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.485, + 0.903 + ], + "angle": 0, + "content": "To further explore the effectiveness of the IOSP, DGMOE and EDA in FedVLA, we conduct ablation experiments by individually removing each module while keeping the other components. Specifically, the experimental setups are as follows: (1) remove the IOSP module and directly input the" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.184 + ], + "angle": 0, + "content": "entire scene image and proprioception data into the encoder (w/o IOSP), (2) replace the DGMoE layers with standard FFN, eliminating adaptive expert selection (w/o DGMOE), and (3) perform vanilla FedAvg aggregation instead of EDA (w/o EDA). We compare the task success rate and record the validation loss during the training process." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.199, + 0.9, + 0.281 + ], + "angle": 0, + "content": "
MethodAvg. SuccessClean UpTrash CollectionOpen DriverSorting Pills
w/o IOSP41.140.013.366.746.7
w/o DGMoE31.720.020.046.740.0
w/o EDA26.726.720.033.326.7
FedVLA (Ours)63.353.346.780.073.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.292, + 0.905, + 0.32 + ], + "angle": 0, + "content": "Table 3. Ablation studies of proposed FedVLA without IOSP, DG-MoE, and EDA." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.339, + 0.907, + 0.64 + ], + "angle": 0, + "content": "From Table 3, we can see that each module contributes significantly to the performance of FedVLA, and the lack of each module will lead to a decrease in the performance of the entire system. Specifically, removing IOSP leads to the highest performance drop in the Trash Collection task from \\(46.7\\%\\) to \\(13.3\\%\\), because the number of objects in Trash Collection task is more than the others. The IOSP module discriminates each object from the other for taking the right action to adapt to multi-objects scenarios. DGMoE module strongly relates to the Clean Up task, leading to a success rate drop from \\(53.3\\%\\) to \\(20.0\\%\\). The Clean Up task is harder in the manipulation aspect as it requires the robot interaction for multiple steps. The FedVLA depends on the DGMoE module to adaptively select the appropriate number of experts based on different types of tokens, enabling collaborative learning for interaction-rich tasks. Furthermore, the removal of EDA gives the largest reduction in average success rate from \\(63.3\\%\\) down to \\(26.7\\%\\) As a common aggregation cancels out the knowledge of the experts in the DGMoE module." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Figure 5 presents the validation loss of the ablation experiments during the first 500 training epochs. We can see that FedVLA consistently achieves the fastest convergence speed, along with the lowest and most stable validation losses across all tasks. Specifically, we can discover that (1) without IOSP causes validation losses notably increase in the multi-object scenarios (depicted in Figure 5b and Figure 5c), (2) the removal of DGMoE results in significantly increased and fluctuating losses the interaction-rich tasks, (3) removing the EDA module causes relatively higher validation losses in most tasks. These findings strongly support that IOSP helps decompose complex visual scenes into task-relevant object representations, DGMoE module routes the adaptive knowledge, EDA ensures the integrity and effectiveness of the aggregation. The integration of these modules together results in a architecture that supports FedVLA's robustness and adaptability across diverse tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6910" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.089, + 0.299, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.157, + 0.189, + 0.248, + 0.202 + ], + "angle": 0, + "content": "(a) OpenDrawer" + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.089, + 0.498, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.358, + 0.189, + 0.442, + 0.202 + ], + "angle": 0, + "content": "(b) Sorting Pills" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.09, + 0.696, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.546, + 0.189, + 0.652, + 0.201 + ], + "angle": 0, + "content": "(c)Trash Collection" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.09, + 0.895, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.762, + 0.189, + 0.832, + 0.202 + ], + "angle": 0, + "content": "(d) Clean Up" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.214, + 0.908, + 0.257 + ], + "angle": 0, + "content": "Figure 5. Validation loss comparison in ablation study across four tasks. The x-axis represents the training rounds, while the y-axis represents the validation loss. Each subplot includes a zoomed-in inset focusing on the validation loss fluctuations between rounds 400 and 500, illustrating the fine-grained differences in convergence trends." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.271, + 0.289, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.147, + 0.354, + 0.237, + 0.366 + ], + "angle": 0, + "content": "(a) OpenDrawer" + }, + { + "type": "image", + "bbox": [ + 0.3, + 0.27, + 0.493, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.354, + 0.354, + 0.439, + 0.366 + ], + "angle": 0, + "content": "(b) Sorting Pills" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.27, + 0.698, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.354, + 0.655, + 0.365 + ], + "angle": 0, + "content": "(c)Trash Collection" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.27, + 0.903, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.772, + 0.354, + 0.842, + 0.366 + ], + "angle": 0, + "content": "(d) Clean Up" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.378, + 0.908, + 0.462 + ], + "angle": 0, + "content": "Figure 6. Visualization of expert activation distribution per task for different object types. The x-axis represents the three object groups (target objects, surrounding objects, and background objects), where each group contains eight sub-bars corresponding to different experts selected. The left y-axis indicates the activation count, showing how frequently experts are selected for each object type. The right y-axis represents the density of experts per token, illustrating the average number of experts assigned per token in each object category. The dashed line represents the average expert density per token across different object groups, highlighting the dynamic expert selection mechanism of DGMoE." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.487, + 0.285, + 0.503 + ], + "angle": 0, + "content": "4.4. Analysis on DGMoE" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.516, + 0.484, + 0.713 + ], + "angle": 0, + "content": "The DGMoE is designed to enhance model efficiency by selectively activating experts based on task complexity and their corresponding object types. In contrast, supporting objects and background objects, which provide contextual spatial information, activate fewer experts compared to target objects. Notably, target objects tend to activate experts 4, 7, and 8 more frequently, supporting objects prefer to select experts 6, 7, and 8, whereas background objects favor experts 2, 3, and 5. This demonstrates that DGMoE not only dynamically allocates computational resources based on object relevance but also selectively routes information to specialized experts, enhancing efficiency and adaptability across different object categories." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Furthermore, we quantitatively measure the average expert density per token across different tasks, also as shown in Figure 6. The proposed DGMoE achieves consistently lower expert activation densities compared to vanilla top-\\(k\\) MoE, which assigns tokens to a fixed number of experts, where \\(k\\) is typically an integer greater than one. The average number of activated experts per token is 1.219, 1.229, 1.225, and 1.227 for Open Drawing, Sorting Pills,Trash Collection, and Clean Up, respectively. Such dynamic and sparse activation directly translates into significantly reduced forward-pass computation, making DGMoE especially suitable for deployment on resource-constrained edge devices." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.487, + 0.633, + 0.503 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.519, + 0.907, + 0.837 + ], + "angle": 0, + "content": "In this work, we introduced FedVLA, a novel federated vision-language action learning framework that preserves privacy and effectively balances data privacy and task performance. To address the challenges of task heterogeneity and efficient knowledge aggregation in decentralized settings, we proposed three key components: the Instruction-Oriented Scene-Parsing module, which enhances feature extraction by decomposing raw observations into structured object-level representations guided by task instructions. The Dual Gating Mixture-of-Experts, which introduces self-aware experts capable of bidirectional token selection, optimizing computational efficiency while preserving task performance, and the Expert-Driven Aggregation strategy, which dynamically assigns aggregation weights based on expert similarity across clients. Extensive experiments in both simulated and real-world environments demonstrate that FedVLA achieves performance comparable to centralized training while ensuring privacy protection. We believe these results may contribute to the development of more generalizable and scalable federated learning frameworks for robotic manipulation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. This research was supported by Guangdong Basic and Applied Basic Research Foundation (No. 2024A1515011774)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "6911" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.484, + 0.257 + ], + "angle": 0, + "content": "[2] Martin Albrecht, Melissa Chase, Hao Chen, Jintai Ding, Shafi Goldwasser, Sergey Gorbunov, Shai Halevi, Jeffrey Hoffstein, Kim Laine, Kristin Lauter, et al. Homomorphic encryption standard. Protecting privacy through homomorphic encryption, pages 31-62, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.259, + 0.484, + 0.34 + ], + "angle": 0, + "content": "[3] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. Openflamingo: An opensource framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.343, + 0.483, + 0.413 + ], + "angle": 0, + "content": "[4] Denis Blessing, Onur Celik, Xiaogang Jia, Moritz Reuss, Maximilian Li, Rudolf Lioutikov, and Gerhard Neumann. Information maximizing curriculum: A curriculum-based approach for learning versatile skills. Advances in Neural Information Processing Systems, 36:51536-51561, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.415, + 0.483, + 0.496 + ], + "angle": 0, + "content": "[5] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.5, + 0.483, + 0.57 + ], + "angle": 0, + "content": "[6] Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa, Soravit Changpinyo, Jialin Wu, Carlos Riquelme Ruiz, Sebastian Goodman, Xiao Wang, Yi Tay, et al. Pali-x: On scaling up a multilingual vision and language model. arXiv preprint arXiv:2305.18565, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.572, + 0.483, + 0.655 + ], + "angle": 0, + "content": "[7] Zitian Chen, Yikang Shen, Mingyu Ding, Zhenfang Chen, Hengshuang Zhao, Erik G Learned-Miller, and Chuang Gan. Mod-squad: Designing mixtures of experts as modular multi-task learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11828-11837, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.657, + 0.483, + 0.726 + ], + "angle": 0, + "content": "[8] Damai Dai, Chengqi Deng, Chenggang Zhao, RX Xu, Huazuo Gao, Deli Chen, Jiashi Li, Wangding Zeng, Xingkai Yu, Y Wu, et al. Deepseekmoe: Towards ultimate expert specialization in mixture-of-experts language models. arXiv preprint arXiv:2401.06066, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.728, + 0.483, + 0.77 + ], + "angle": 0, + "content": "[9] Cynthia Dwork. Differential privacy. In International colloquium on automata, languages, and programming, pages 1-12. Springer, 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.772, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[10] Oded Goldreich. Secure multi-party computation. Manuscript. Preliminary version, 78(110):1-108, 1998. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[11] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. Neural computation, 3(1):79-87, 1991. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Shen Jiang, Guanghui Zhu, Yue Wang, Chunfeng Yuan, and Yihua Huang. Automatic multi-task learning framework with neural architecture search in recommendations. In Proceedings of the 30th ACM SIGKDD Conference on Knowl" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "edge Discovery and Data Mining, pages 1290-1300, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[13] Siddharth Karamcheti, Suraj Nair, Ashwin Balakrishna, Percy Liang, Thomas Kollar, and Dorsa Sadigh. Prismatic vlms: Investigating the design space of visually-conditioned language models. arXiv preprint arXiv:2402.07865, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[14] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, Quan Vuong, Thomas Kollar, Benjamin Burchfiel, Russ Tedrake, Dorsa Sadigh, Sergey Levine, Percy Liang, and Chelsea Finn. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.278, + 0.905, + 0.345 + ], + "angle": 0, + "content": "[15] Xinghang Li, Minghuan Liu, Hanbo Zhang, Cunjun Yu, Jie Xu, Hongtao Wu, Chilam Cheang, Ya Jing, Weinan Zhang, Huaping Liu, et al. Vision-language foundation models as effective robot imitators. arXiv preprint arXiv:2311.01378, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.349, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[16] Bin Lin, Zhenyu Tang, Yang Ye, Jiaxi Cui, Bin Zhu, Peng Jin, Junwu Zhang, Munan Ning, and Li Yuan. Moe-llava: Mixture of experts for large vision-language models. arXiv preprint arXiv:2401.15947, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.406, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[17] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.449, + 0.905, + 0.49 + ], + "angle": 0, + "content": "[18] Hao Liu, Wilson Yan, Matei Zaharia, and Pieter Abbeel. World model on million-length video and language with ringattention. arXiv e-prints, pages arXiv-2402, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.492, + 0.905, + 0.547 + ], + "angle": 0, + "content": "[19] Yang Liu, Weixing Chen, Yongjie Bai, Xiaodan Liang, Guanbin Li, Wen Gao, and Liang Lin. Aligning cyber space with physical world: A comprehensive survey on embodied ai. arXiv preprint arXiv:2407.06886, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.549, + 0.905, + 0.618 + ], + "angle": 0, + "content": "[20] Yuxiang Lu, Suizhi Huang, Yuwen Yang, Shalayiding Sirejiding, Yue Ding, and Hongtao Lu. Fedhca2: Towards hetero-client federated multi-task learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5599-5609, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.62, + 0.905, + 0.689 + ], + "angle": 0, + "content": "[21] Ajay Mandlekar, Danfei Xu, Josiah Wong, Soroush Nasiriyani, Chen Wang, Rohun Kulkarni, Li Fei-Fei, Silvio Savarese, Yuke Zhu, and Roberto Martin-Martin. What matters in learning from offline human demonstrations for robot manipulation. arXiv preprint arXiv:2108.03298, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.691, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[22] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.761, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[23] David Nadeau and Satoshi Sekine. A survey of named entity recognition and classification. *Lingvisticae Investigationes*, 30(1):3-26, 2007. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] Abby O'Neill, Abdul Rehman, Abhiram Maddukuri, Abhishek Gupta, Abhishek Padalkar, Abraham Lee, Acorn Pooley, Agrim Gupta, Ajay Mandlekar, Ajinkya Jain, et al. Open x-embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collaboration 0. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 6892–6903. IEEE, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6912" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.174 + ], + "angle": 0, + "content": "[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[26] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.316 + ], + "angle": 0, + "content": "[27] Rejin Varghese and M Sambath. Yolov8: A novel object detection algorithm with enhanced performance and robustness. In 2024 International Conference on Advances in Data Engineering and Intelligent Computing Systems (ADICS), pages 1-6. IEEE, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.319, + 0.482, + 0.373 + ], + "angle": 0, + "content": "[28] Lirui Wang, Xinlei Chen, Jialiang Zhao, and Kaiming He. Scaling proprioceptive-visual learning with heterogeneous pre-trained transformers. Advances in Neural Information Processing Systems, 37:124420-124450, 2025. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.443 + ], + "angle": 0, + "content": "[29] Jiahui Xu, Lu Sun, and Dengji Zhao. Mone: Mixture-of-masked-experts for efficient multi-task recommendation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2527-2531, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.446, + 0.482, + 0.514 + ], + "angle": 0, + "content": "[30] Tianhe Yu, Deirdre Quillen, Zhanpeng He, Ryan Julian, Karol Hausman, Chelsea Finn, and Sergey Levine. Metaworld: A benchmark and evaluation for multi-task and meta reinforcement learning. In Conference on robot learning, pages 1094-1100. PMLR, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.517, + 0.482, + 0.569 + ], + "angle": 0, + "content": "[31] Michal Zawalski, William Chen, Karl Pertsch, Oier Mees, Chelsea Finn, and Sergey Levine. Robotic control via embodied chain-of-thought reasoning. arXiv preprint arXiv:2407.08693, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.573, + 0.482, + 0.627 + ], + "angle": 0, + "content": "[32] Kaiwen Zhou and Xin Eric Wang. Fedvln: Privacy-preserving federated vision-and-language navigation. In European Conference on Computer Vision, pages 682-699. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.63, + 0.482, + 0.695 + ], + "angle": 0, + "content": "[33] Weiming Zhuang, Yonggang Wen, Lingjuan Lyu, and Shuai Zhang. Mas: Towards resource-efficient federated multipletask learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23414-23424, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.695 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "6913" + } + ] +] \ No newline at end of file diff --git a/2025/FedVLA_ Federated Vision-Language-Action Learning with Dual Gating Mixture-of-Experts for Robotic Manipulation/9830aa12-506c-4606-9232-07331ac373a7_origin.pdf b/2025/FedVLA_ Federated Vision-Language-Action Learning with Dual Gating Mixture-of-Experts for Robotic Manipulation/9830aa12-506c-4606-9232-07331ac373a7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a3bdf0310e9b1668f6b2c258b492c9f01e92c07a --- /dev/null +++ b/2025/FedVLA_ Federated Vision-Language-Action Learning with Dual Gating Mixture-of-Experts for Robotic Manipulation/9830aa12-506c-4606-9232-07331ac373a7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a72b8d5bbd19c80849293f84daf5329572478dfffab5d59cca0f811502aea8a +size 835966 diff --git a/2025/FedVLA_ Federated Vision-Language-Action Learning with Dual Gating Mixture-of-Experts for Robotic Manipulation/full.md b/2025/FedVLA_ Federated Vision-Language-Action Learning with Dual Gating Mixture-of-Experts for Robotic Manipulation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..beda50ef6270950f9f5707692d7bda8ad1c4a667 --- /dev/null +++ b/2025/FedVLA_ Federated Vision-Language-Action Learning with Dual Gating Mixture-of-Experts for Robotic Manipulation/full.md @@ -0,0 +1,331 @@ +# FedVLA: Federated Vision-Language-Action Learning with Dual Gating Mixture-of-Experts for Robotic Manipulation + +Cui Miao $^{1}$ Tao Chang $^{1}$ Meihan Wu $^{1}$ Hongbin Xu $^{2}$ Chun Li $^{3}$ Ming Li $^{4*}$ Xiaodong Wang $^{1}$ National University of Defense Technology + $^{2}$ Bytedance Seed Shenzhen MSU-BIT University +Long Laboratory of Artificial Intelligence and Digital Economics + +# Abstract + +Vision-language-action (VLA) models have significantly advanced robotic manipulation by enabling robots to interpret language instructions for task execution. However, training these models often relies on large-scale user-specific data, raising concerns about privacy and security, which in turn limits their broader adoption. To address this, we propose FedVLA, the first federated VLA learning framework, enabling distributed model training that preserves data privacy without compromising performance. Our framework integrates task-aware representation learning, adaptive expert selection, and expert-driven federated aggregation, enabling efficient and privacy-preserving training of VLA models. Specifically, we introduce an Instruction-Oriented Scene-Parsing mechanism, which decomposes and enhances object-level features based on task instructions, improving contextual understanding. To effectively learn diverse task patterns, we design a Dual Gating Mixture-of-Experts (DGMoE) mechanism, where not only input tokens but also self-aware experts adaptively decide their activation. Finally, we propose an Expert-Driven Aggregation strategy at the federated server, where model aggregation is guided by activated experts, ensuring effective cross-client knowledge transfer. Extensive simulations and real-world robotic experiments demonstrate the effectiveness of our proposals. Notably, DGMoE significantly improves computational efficiency compared to its vanilla counterpart, while FedVLA achieves task success rates comparable to centralized training, effectively preserving data privacy. + +# 1. Introduction + +Vision-language-action (VLA) models, which integrate visual perception, linguistic understanding, and robotic control, have significantly enhanced robotic manipulation by + +![](images/3cd25ba72ac5b657c4f1fe871e3689f016b13fae0b1aa74685f7b1e332cfc4dd.jpg) +(a) Centralized VLA learning +Figure 1. Comparison between centralized and federated VLA training. Centralized training collects all user data on the cloud, raising privacy concerns due to potential exposure of sensitive information. Our federated VLA framework enables decentralized training on user devices, preserving privacy while utilizing expert-driven aggregation to enhance model generalization across diverse tasks. + +![](images/2ce0c8bce08eb2f30ee766efc3ed1934089c63469e74c0ee47f8c907d7572d49.jpg) +(b) Federated VLA learning + +enabling robots to comprehend and execute complex tasks through natural language instructions [5, 14, 28]. However, training these models requires extensive indoor datasets that capture diverse, user-specific environments, raising serious concerns about data privacy. While recent advancements have improved the capability of VLA models in robotic systems, the issue of privacy leakage remains largely unexplored [19]. + +Federated learning (FL) provides an effective privacy-preserving solution for training VLA models [22]. Unlike traditional centralized training, which requires aggregating all user data on a central server, FL enables distributed model training across multiple clients without transferring raw data. As illustrated in Figure 1a, centralized VLA training collects and processes user data in the cloud, posing risks of privacy leakage, such as exposure of personal pref + +erences and sensitive health conditions. In contrast, figure 1b presents a federated VLA learning paradigm, where local models are trained on user devices, and only model updates are transmitted to the central server. This decentralized paradigm mitigates privacy risks while allowing collaborative model optimization. + +Existing federated learning approaches are not well-suited for VLA learning, where robotic manipulation tasks exhibit substantial task heterogeneity across different clients. Typical FL methods, such as FedAvg [22], naively aggregate models across tasks, considering only shared synergies while overlooking their fundamental distinctions. Although recent works attempt to address this issue through merge-and-split training [33] or decoupled model aggregation [20], they remain limited to single-modal settings, assuming a uniform input modality across tasks. In contrast, VLA models operate in multi-modal environments, requiring the joint processing of visual observations, language instructions, and robotic actions, which significantly increases the complexity of federated training. + +Additionally, mixture of experts (MoE) has shown promise in multi-task learning by assigning specialized experts to different tasks [7, 12, 29]. However, existing methods typically select a fixed number of experts, lacking adaptability to tasks of varying complexity. This rigidity makes them inefficient for resource-constrained clients, where optimizing inference speed is critical for real-time robotic responses. These limitations highlight the need for a task-adaptive and flexible FL framework, specifically designed for multi-modal robotic learning. + +In this paper, we introduce FedVLA, a novel federated vision-language-action (VLA) learning framework with Dual Gating MoE, designed to preserve privacy while maintaining high task performance, as illustrated in Figure 1b. To address task heterogeneity and enable effective knowledge aggregation in federated settings, we propose three key components. First, we introduce an Instruction-Oriented Scene-Parsing (IOSP) module to enhance task-aware feature extraction in decentralized environments, where raw images across clients exhibit diverse backgrounds and object distributions. IOSP decomposes observation images into object-level representations guided by task instructions and leverages vision-language alignment techniques to improve contextual understanding. Second, to enhance computational efficiency in federated learning, we develop a Dual Gating MoE (DGMoE) mechanism that enables adaptive knowledge routing. Unlike conventional MoE, which assigns tokens to a fixed number of experts, DGMoE introduces self-aware experts that can accept or reject token assignments, enabling a bidirectional selection process. This dual gating mechanism improves efficiency while preserving task performance. Lastly, to effectively aggregate similar knowledge from diverse client models, we propose the + +Expert-Driven Aggregation (EDA) strategy, which dynamically assigns aggregation weights based on expert similarity across clients. By prioritizing updates from clients that activate similar experts, EDA ensures that semantically aligned knowledge is more prominently integrated at the server side, resulting in a more coherent and specialized global model. + +Our main contributions in this work can be summarized as follows: + +- We propose FedVLA, the first privacy-preserving federated learning framework for VLA training, ensuring that user data remains private and never shared externally. Extensive experiments in both simulation and real-world environments demonstrate that FedVLA achieves performance comparable to centralized training while preserving data privacy. +- We introduce the Dual Gating Mixture-of-Experts, where self-aware experts autonomously determine their participation, enabling bidirectional token selection to optimize computational efficiency while maintaining task performance. +- We propose the Expert-Driven Aggregation strategy, which dynamically assigns aggregation weights based on expert similarity across clients, ensuring that semantically aligned knowledge is more effectively integrated, thereby enhancing generalization across heterogeneous tasks. + +# 2. Related Works + +# 2.1. Vision-Language-Action Models + +Vision-Language Models (VLMs), which are pre-trained on large-scale datasets, have demonstrated remarkable capabilities in understanding and generating multimodal data [17, 18, 26]. Inspired by VLMs, Vision-Language-Action (VLA) models extend their capabilities by integrating robotic action prediction. For instance, RoboFlamingo [15], RT-2 [5], and OpenVLA [14] are all built upon VLMs such as OpenFlamingo [3], PaLI-X [6], and Prismatic [13], respectively. Recent work has begun exploring the training of generalist robots [28], followed by fine-tuning on diverse robot datasets, enabling generalization to novel instructions, unseen objects, and distribution shifts [31]. This progress relies heavily on the availability of large-scale public datasets, such as Open-X Embodiment dataset [24]. During the finetuning phase, samples from real-world scenarios are required, which are often generated by human activities and distributed across users' home environments. However, these samples frequently expose privacy-sensitive information. + +# 2.2. Privacy-Preserving Training + +Traditional centralized learning methods require uploading data from multiple sources, which poses significant privacy risks. To mitigate these concerns, researchers have pro + +![](images/69ecaee2a94e1b08f56c1b73f3ce3fdc656be67076c1ddc3fbd69875c5ca0d3b.jpg) +Figure 2. Illustration of the proposed FedVLA framework and its key component. (a) An overview of the FedVLA, which consists of multiple clients and a server. Each client processes vision-language-action inputs through the Stem, Trunk, and Head modules. Clients send their updated trunk parameters to the server, where Expert-Driven Aggregation selectively integrates expert-specific updates on server. The global trunk module is then distributed to clients for the next training round. (b) Details of the proposed Dual Gating MoE module. The input tokens first pass through the token-level gate $G_{t}$ , computing the gating scores for each expert with residual scores from the previous MoE layer. Only the computed scores surpass the expert-level gate $G_{e}$ , activating the corresponding expert to process the token. Finally, the expert outputs are combined to generate the final output tokens. + +posed various privacy-preserving techniques, including federated learning (FL) [22], differential privacy [9], secure multi-party computation [10], and homomorphic encryption [2]. However, Most privacy-preserving training techniques have been developed for single-modality tasks, and seldomly study the data privacy of embodied AI [32]. In robotic manipulation, the models process multimodal sensor data, including images, language instructions, and action sequences, which can contain sensitive user interaction data. Unlike static datasets, robotic systems continuously collect and process dynamic, task-specific information, making privacy protection more complex. In this context, FL as a solution by allowing decentralized model training without sharing raw data. However, FL for robotic manipulation remains challenging due to task heterogeneity and the need for efficient model aggregation across clients. + +# 2.3. Mixture of Experts in Large Models + +Mixture of experts [4, 11] is an architecture where multiple experts are trained collaboratively, with each expert responsible for a specific portion of the input space. An MoE consists of two components: a routing mechanism and a set of experts. The experts can either be part of a feedforward network (FFN) or the entire FFN layer, and the routing mechanism allows each token to select experts for processing. By activating a specific number of experts, MoE can reduce resource consumption and improve inference speed. This feature makes MoE particularly effective when integrated into large models, as seen in models like LLaVA-MoE [16], + +DeepSeekMoE [8], and GPT-4 [1]. In this work, we apply MoE to a generalist robot VLA model and introduce a self-aware variant that can adaptively decide whether to be selected, significantly improving deployment efficiency. + +# 3. FedVLA + +We aim to train a vision-language-action model in a federated learning (FL) framework, ensuring privacy preservation while enabling task generalization across clients. Formally, given $N$ clients $\mathbb{C} = \{C_1,C_2,\dots,C_N\}$ , $N$ local tasks $\mathbb{T} = \{T_1,T_2,\dots,T_N\}$ , and $N$ local datasets $\mathbb{D} = \{\mathbb{D}_1,\mathbb{D}_2,\dots,\mathbb{D}_N\}$ , each client $C_i$ is associated with a local task $T_{i}$ and a corresponding dataset $\mathbb{D}_i$ . The dataset $\mathbb{D}_i$ contains visual perception image $o_i$ , language instruction $l_{i}$ of the task, proprioception $s_i$ of the robot, and action sequence $a_{i}$ executed by the robot: + +$$ +\mathbb {D} _ {i} = \left\{\left(o _ {i}, l _ {i}, s _ {i}, a _ {i}\right) _ {j} | j = 1, 2, \dots , M \right\}, \tag {1} +$$ + +where $M$ is the number of samples in dataset $\mathbb{D}_i$ . To train a federated VLA model $\theta_G$ , the overall federated optimization objective can be formulated as: + +$$ +\min _ {\theta_ {G}} \sum_ {i = 1} ^ {N} w _ {i} \sum_ {(o _ {i}, l _ {i}, a _ {i}, s _ {i}) \in \mathbb {D} _ {i}} \mathcal {L} _ {i} (f _ {i} (\theta_ {G}; o _ {i}, l _ {i}, s _ {i}), a _ {i}), \tag {2} +$$ + +where $\mathcal{L}_i$ is the Huber loss function computed over each client model's local dataset $\mathbb{D}_i$ , $f_i$ represents the local client model's predicted action sequence, and $w_i$ denotes the weight assigned to each client during aggregation, which is further detailed in Section 3.3. + +![](images/0c5055f55c82abdd9529d01f59dbe56cd508721fe706971b900405a0997767a6.jpg) +Figure 3. Instruction-Oriented Scene-Parsing module, which decomposes observations into object-level representations guided by task instructions, enhancing task-aware feature extraction through vision-language alignment. + +The overview of our FedVLA framework is shown in Figure 2. The clients train their models locally and communicate with the server per round until the model converges. The following sections introduce our key proposals: (1) Instruction-Oriented Scene-Parsing (Section 3.1) that decomposes the observation image into object-level embeddings guided by task instructions. (2) Dual Gating Mixture-of-Experts (Section 3.2) that can adaptively select a varying number of experts for tokens based on the relevance to the specific task. (3) Expert-Driven Aggregation (Section 3.3) that dynamically aggregates the weights based on expert similarities across different clients in the server. + +# 3.1. Instruction-Oriented Scene-Parsing + +To ensure the robot focuses on task-relevant information while maintaining contextual awareness to avoid collisions, we decouple objects in the scene based on the specific robotic task. To achieve this, we propose an Instruction-Oriented Scene-Parsing (IOSP) module to represent objects within the task scene. + +As illustrated in Figure 3, given an instruction and an image, we first extract Target Objects (TOs) from the instruction using named entity recognition [23]. Simultaneously, we employ YOLOv8 [27] to detect and classify Foreground Objects (FOs) and Background Objects (BOs) in the image. The scene objects are categorized into three groups based on the cosine similarity measured by a pre-trained CLIP model [25] between the instruction and object names, following these matching rules: (1) Target Objects (TOs): Objects explicitly mentioned in the instruction and directly involved in the action. (2) Surrounding Objects (SOs): Foreground objects not in TOs, providing spatial and contextual cues. (3) Background Objects (BOs): Static environmental elements ensuring scene consistency but not directly affecting the task. + +Then, we encode the grouped objects using the CLIP model. The text encoder $H$ maps object names from TOs, SOs, and BOs into textual embeddings $H(\mathrm{TOs}) \in \mathbb{R}^{t \times D}$ , $H(\mathrm{BOs}) \in \mathbb{R}^{t \times D}$ , and $H(\mathrm{SOs}) \in \mathbb{R}^{t \times D}$ . Meanwhile, the vision encoder $P$ processes the entire image into $P(X) \in \mathbb{R}^{t \times D}$ . We assign image tokens to object groups by computing the cosine similarity between token and text embed + +dings, selecting the top eight tokens per group. To enhance intra-group feature representation, each group is processed through an MoE module, producing refined image token embeddings. The enhanced tokens are then concatenated with the remaining tokens for subsequent processing. This IOSP module enables the model to focus on critical objects, improving task-specific manipulation. + +# 3.2. Dual Gating Mixture-of-Experts + +In this section, we introduce Dual Gating Mixture-of-Expert (DGMoE) mechanism that supports adaptive knowledge routing. Different from vanilla MoE, which assigns a fixed number of experts to each token, our DGMoE can automatically determine the acceptance or rejection of the token assignments under the Self-Aware Expert. + +As shown in Figure 2, each DGMoE layer is comprised of $K$ self-aware experts $\mathbb{E} = \{E_1,E_2,\dots ,E_K\}$ . To enable the automatic token-driven expert selection, we introduce two gating mechanisms: Token-side gating mechanism $G_{t}$ and Expert-side gating mechanism $G_{e}$ . (1) For Token-side gating, $G_{t}$ serves as the gating module on the token side, comprised of a soft router module and a gating residual module. The soft router first estimates the selection scores for the hidden inputs, indicating the token preference for different experts. Then the estimated scores from previous layer are incorporated with the scores of current layer, allowing the tokens to inherit the selection prior of the most suitable experts from previous layers. (2) For Expert-side gating, $G_{e}$ can autonomously accept or reject tokens based on their acceptance threshold. The selection scores predicted by $G_{t}$ is filtered with a learnable threshold, and only the experts with the most relevant tokens are activated. + +Given an input token $x$ , the token-side gate $G_{t}$ first computes the scores at the $j$ -th layer by combining the soft router output and the gating residual module: + +$$ +G _ {t} ^ {j} (x) = \left\{ \begin{array}{l l} W _ {t} ^ {j} x, & \text {i f} j = 1, \\ W _ {t} ^ {j} x + W _ {t} ^ {g} G _ {t} ^ {(j - 1)} (x), & \text {o t h e r w i s e}, \end{array} \right. \tag {3} +$$ + +where $W_{t}^{j}\in \mathbb{R}^{D}$ is the trainable weight matrix for soft router, and $W_{t}^{g}\in \mathbb{R}^{N\times N}$ is a trainable transformation matrix designed to aggregate the scores from the previous layer and the current layer, ensuring smooth expertise knowledge transitions across layers. Then, the token-to-expert selection scores $s_t(x)$ are obtained by normalizing $G_{t}(x)$ with a softmax function, which determines the probability distribution over the available experts. + +On the self-aware expert side, the expert-side gate $G_{e}$ evaluates the token scores and determines the final expert activation scores $s_{e}(x)$ as follows: + +$$ +s _ {e} (x) = \operatorname {s i g n} \left(s _ {t} (x) - \lambda W _ {e}\right), \tag {4} +$$ + +where $W_{e}$ is a set of trainable self-aware parameters corresponding to the gating values of each expert, and $\lambda$ is a + +scaling factor that modulates the gating threshold. In our setting, we set $\lambda = 0.5$ to balance expert selection, ensuring neither excessive expert activation nor overly strict filtering of tokens. + +Then, combining the dual gating mechanism of token-side and exper-side gating, the expert selection function is defined as follows: + +$$ +g (x) = \left\{ \begin{array}{l l} s _ {t} (x), & \text {i f} s _ {e} (x) > 0, \\ 0, & \text {i f} s _ {e} (x) \leq 0, \end{array} \right. \tag {5} +$$ + +where the final output token $y$ of the DGMoE layer can be computed as the weighted sum of the selected expert outputs: + +$$ +y = \sum_ {i = 1} ^ {K} g _ {i} (x) E _ {i} (x). \tag {6} +$$ + +With the help of self-aware expert gating mechanism, our DGMoE can dynamically activate the most relevant experts, reducing computational overhead while maintaining task-relevant feature learning. + +# 3.3. Expert-Driven Aggregation + +To mitigate performance degradation caused by task heterogeneity, we propose an Expert-Driven Aggregation (EDA) strategy, which aggregates model parameters based on the similarity of expert selections across clients. And as shown in Figure 2, we selectively aggregates only the trunk module, while the stem and head modules remain personalized and are excluded from aggregation. + +Given a trunk module with $L$ layers, where each layer contains $K$ experts, each client $C_i$ records an expert selection matrix $\mathbf{V}_i\in \mathbb{R}^{L\times K}$ during each round through DGMoE, where each element $V_{i}^{(l,k)}$ represents the number of times expert $k$ is activated in the $l$ -th layer. We extract the expert selection vector for layer $l$ of client $C_i$ as: + +$$ +\mathbf {v} _ {i} ^ {(l)} = \left(V _ {i} ^ {(l, 1)}, V _ {i} ^ {(l, 2)}, \dots , V _ {i} ^ {(l, K)}\right). \tag {7} +$$ + +Then we compute the pairwise similarity between clients based on expert selection vectors. The similarity between clients $C_i$ and $C_j$ at layer $l$ is defined as follows: + +$$ +s _ {i, j} ^ {(l)} = \frac {\mathbf {v} _ {i} ^ {(l)} \cdot \mathbf {v} _ {j} ^ {(l)}}{\| \mathbf {v} _ {i} ^ {(l)} \| \| \mathbf {v} _ {j} ^ {(l)} \|\}.} \tag {8} +$$ + +Using this similarity, we determine the aggregation weight $w_{l,i}$ for each client $C_i$ at layer $l$ as follows: + +$$ +w _ {l, i} = \frac {\sum_ {C _ {j} \in C} s _ {i , j} ^ {(l)}}{\sum_ {C _ {j} \in C} \sum_ {C _ {i} \in C} s _ {i , j} ^ {(l)}}. \tag {9} +$$ + +This weighting strategy ensures that clients with higher expert selection similarity contribute more to each other's model updates, allowing for more effective and task-aligned aggregation across task-heterogeneous clients. + +Algorithm 1 FedVLA: Clients + +Input: Each client $C_i$ has a local task $T_i$ and dataset $D_i = \{o_i, l_i, s_i, a_i\}$ , along with local parameters $\phi_i$ , $w_i$ , and $\theta_i$ corresponding to the stem, trunk, and head modules, respectively. + +1: Local_Train $(C_i, \phi_i, \theta_i)$ : +2: for each local epoch $e$ from 1 to $E_{c}$ do +3: for each batch $(o_i, l_i, s_i, a_i)$ sampled from $D_i$ do +4: // Step 1: Instruction-Oriented Scene-Parsing +5: TOs, SOs, BOs $\leftarrow \{o_i, l_i\}$ +6: $F_{T O s}, F_{S O s}, F_{B O s} \gets \mathrm{CLIP}(T O s, S O s, B O s)$ +7: $F_{O}\gets \mathrm{CLIP}(o_{i})$ +8: $X_{\mathrm{TO}} \gets \mathrm{topk}(\mathrm{cosine}(F_{\mathrm{TOs}}, F_O), F_O)$ +9: $X_{\mathrm{SO}} \gets \mathrm{topk}(\mathrm{cosine}(F_{\mathrm{SOs}}, F_O), F_O)$ +10: $X_{\mathrm{BO}} \gets \mathrm{topk}(\mathrm{cosine}(F_{\mathrm{BOs}}, F_O), F_O)$ +11: $X_{\mathrm{in}} = \mathrm{concat}(X_{\mathrm{TO}},X_{\mathrm{SO}},X_{\mathrm{BO}},s_i)$ +12: // Step 2: Forward pass with DGMoE +13: Let $x \in MHA(X_{\mathrm{in}})$ +$s_t(x) = \mathrm{softmax}\bigl (W_t(x)\bigr)$ // token-gate +14: $s_e(x) = \mathrm{sign}(s_t(x) - \lambda W_e)$ // expert-gate +15: $g(x) = s_t(x)\cdot \mathbb{I}(s_e(x) > 0)$ // expert activation +16: $\mathbf{v}^{(l)},y = \sum_{k = 1}^{K}g(x)E_k(x)$ //final output +17: // Step 3: Model parameter updates +18: $\mathcal{L}_i = \mathrm{Loss}(f(\phi_i, \theta_i, w_i), a_i)$ +19: $(\phi_i,\theta_i,w_i)\gets (\phi_i,\theta_i,w_i) - \eta \nabla \mathcal{L}_i$ +20: end for +21: end for +22: return $\theta, \mathbf{v}^{(l)}$ to the server. + +# 3.4. Algorithms + +Clients. The overall workflow of our FedVLA is shown in Algorithm 1 and Algorithm 2. At the start of each training round $t$ , each client processes task-specific data using Instruction-Oriented Scene-Parsing to extract structured features, followed by Dual Gating MoE for adaptive expert selection. The updated trunk parameters and expert selection counts are then transmitted to the server for global aggregation. The detailed client-side process is described in Algorithm 1. + +Server. On the server side, the server receives expert selection statistics and trunk updates from all participating clients and performs Expert-Driven Aggregation, which dynamically assigns aggregation weights based on expert selection similarities. The aggregated global trunk module is then redistributed to clients for the next training round. The process is elaborated in Algorithm 2. + +# Algorithm 2 FedVLA: Server + +Input: $T$ is the number of training rounds. $\theta$ represents the global trunk parameters. Each client provides $\mathbf{v}^{(l)}$ , the expert selection counts vector for each layer $l$ . + +1: Server Execute: +2: for each round $t$ from 1 to $T$ do +3: // Expert-Driven Aggregation +4: for each trunk layer $l$ do +5: for each pair of clients $C_i, C_j$ do +6: $s_{i,j}^{(l)} = \frac{\mathbf{v}_i^{(l)}\cdot\mathbf{v}_j^{(l)}}{\|\mathbf{v}_i^{(l)}\|\|\mathbf{v}_j^{(l)}\|} //$ expert selection similarity +7: $w^{(l,i)} = \frac{\sum_{j\in C}s_{i,j}^{(l)}}{\sum_{j\in C}\sum_{i\in C}s_{i,j}^{(l)}}$ // aggregation weight +8: end for +9: $\theta^{(l)}\gets \sum_{i = 1}^{N}w^{(l,i)}\cdot \theta^{(l,i)}$ //update $\theta$ +10: end for +11: Broadcast updated $\theta$ to all clients. +12: end for + +# 4. Experiments + +Implementation Details. We validate our FedVLA in both simulation and real-world scenarios, where each client performs a single task, resulting in a non-iid data distribution. We employ the pretrained HPT [28] as the backbone of our VLA model and train it for 1,000 communication rounds between clients and the server, with each client performing local training for 5 epochs per round. In simulation, the model is trained with a learning rate of $5 \times 10^{-6}$ , while in real-world settings, the learning rate is set to $2 \times 10^{-5}$ . Each client trains locally with a batch size of 256 using the Adam optimizer. The experiments are conducted on a workstation equipped with 2 GeForce GTX 4090Ti GPUs, 48 Intel Xeon CPUs, and 128GB of memory. + +Evaluation. For evaluation, the success and failure of a trial are recorded as 1 and 0. Each task takes for 15 trials and repeats for 5 times. Since this is the first work on federated VLA learning, we compare FedVLA with two methods. The centralized training (Centralized) aggregates all data on a single server for joint training, serving as an upper bound in performance. The federated averaging (FedAvg) [22] is a strong FL baseline where models are trained locally on each client and simply averaged on the server. + +# 4.1. Simulation + +Environment. We utilize the MuJoCo engine [21] to simulate tasks from the Meta-World robot manipulation benchmark [30]. All tasks are performed using a Sawyer robot with a parallel gripper, positioned at the center of a dinner table. Visual data is captured by three fixed RGB cameras with a resolution of $128 \times 128$ , providing corner, top, and + +shoulder views. For collision detection and dynamics simulation, we employ official physics engines to ensure accurate robotic interactions within the simulation environment. Tasks. We evaluate our FedVLA on four household tasks: Door Lock, Close Printer, Sweep Into, and Open Window. For each task, we collect approximately 30-80 episodes, each consisting of 40-100 steps. The initial pose of robot and environment layouts are randomized to assess the generalization ability of FedVLA. + +
MethodAvg. SuccessDoor LockCloseDrawerSweep IntoOpen Window
Centralized65.086.773.353.346.7
FedAvg51.766.773.340.026.7
FedVLA (Ours)63.380.080.053.340.0
+ +Table 1. Simulation evaluation across four tasks. + +Performance Comparison. The evaluation results in the simulation environment are presented in Table 1. FedVLA achieves an average success rate of $63.3\%$ , closely approaching the $65.0\%$ of centralized training. Notably, in the Close Drawn task, FedVLA surpasses centralized training by a success rate margin of about $7.0\%$ . Furthermore, FedVLA consistently outperforms FedAvg, which only achieves an average success rate of $51.7\%$ . This performance gap can be attributed to FedAvg overlooking task heterogeneity in federated learning. In contrast, FedVLA explicitly accounts for task heterogeneity during aggregation by leveraging expert-driven model updates, ensuring that knowledge from diverse clients is effectively integrated to enhance generalization across different tasks. + +# 4.2. Real-World + +Hardware. For real-world experiments, we employ a UR3 robotic arm with six degrees of freedom, equipped with a one-DoF parallel gripper. A RealSense D435i RGB-D camera is mounted above the robot, capturing RGB images in real time at a resolution of $1280 \times 720$ with a frame rate of approximately $30\mathrm{Hz}$ . To simulate real-world variations, we introduce moderate randomness in object placement and lighting conditions to better reflect practical deployment scenarios. + +Tasks. We collect real-world robotic demonstrations for household-related tasks, including Clean Up,Trash Collection, OpenDrawer,and Sorting Pills, as shown in Figure 4. For data collection, we employ an expert teaching method, using inverse kinematics to compute the robot poses corresponding to key point data while recording RGB images and pose data throughout the trajectory. Each task includes approximately 50 demonstrations, with each trajectory consisting of 20 to 80 steps. + +Performance Comparison. The real-world experimental results are summarized in Table 2. FedVLA achieves an + +![](images/575e642ab35fa047dadbb235fe3d844d2a04e1091cf58c1b5194923d6c3ee09c.jpg) +Figure 4. Real-World Tasks. Each column represents a different task, while each row demonstrates keyframes of task execution. From left to right, the tasks are: Clean Up, where the robot wipes the trash from plate using a cloth. Trash Collection, in which it throws crumpled paper into a trash bin. Open Printer, which requires grasping and pulling a drawer open, and Sorting Pills, where pill bottles are placed into trays of designated color. + +average task success rate of $63.3\%$ , closely matching the $63.4\%$ of centralized training. Notably, it attains the same success rates as the centralized model in complex tasks such as Trash Collection $(46.7\%)$ and Sorting Pills $(73.3\%)$ . In contrast, FedAvg exhibits a significant performance drop across all tasks, with an average success rate of $53.3\%$ , which is $10.0\%$ lower than our method. These results indicate that FedVLA effectively bridges the performance gap between federated and centralized training. While FedAvg merely averages model parameters across clients, it struggles to generalize effectively in VLA learning, where multimodal information—including textual instructions, scene images, and execution trajectories—is involved, requiring strong feature extraction capabilities. The proposed Dual Gating MoE and Expert-Driven Aggregation mechanisms enable the model to adaptively learn diverse patterns for robotic manipulation, enhancing its ability to handle complex tasks in federated settings. + +
MethodAvg. SuccessClean UpTrash CollectionOpenDrawerSorting Pills
Centralized63.446.746.786.773.3
FedAvg53.346.740.060.066.7
FedVLA (Ours)63.353.346.780.073.3
+ +# 4.3. Ablation Studies + +To further explore the effectiveness of the IOSP, DGMOE and EDA in FedVLA, we conduct ablation experiments by individually removing each module while keeping the other components. Specifically, the experimental setups are as follows: (1) remove the IOSP module and directly input the + +entire scene image and proprioception data into the encoder (w/o IOSP), (2) replace the DGMoE layers with standard FFN, eliminating adaptive expert selection (w/o DGMOE), and (3) perform vanilla FedAvg aggregation instead of EDA (w/o EDA). We compare the task success rate and record the validation loss during the training process. + +Table 2. Real-world evaluation across four tasks. + +
MethodAvg. SuccessClean UpTrash CollectionOpen DriverSorting Pills
w/o IOSP41.140.013.366.746.7
w/o DGMoE31.720.020.046.740.0
w/o EDA26.726.720.033.326.7
FedVLA (Ours)63.353.346.780.073.3
+ +Table 3. Ablation studies of proposed FedVLA without IOSP, DG-MoE, and EDA. + +From Table 3, we can see that each module contributes significantly to the performance of FedVLA, and the lack of each module will lead to a decrease in the performance of the entire system. Specifically, removing IOSP leads to the highest performance drop in the Trash Collection task from $46.7\%$ to $13.3\%$ , because the number of objects in Trash Collection task is more than the others. The IOSP module discriminates each object from the other for taking the right action to adapt to multi-objects scenarios. DGMoE module strongly relates to the Clean Up task, leading to a success rate drop from $53.3\%$ to $20.0\%$ . The Clean Up task is harder in the manipulation aspect as it requires the robot interaction for multiple steps. The FedVLA depends on the DGMoE module to adaptively select the appropriate number of experts based on different types of tokens, enabling collaborative learning for interaction-rich tasks. Furthermore, the removal of EDA gives the largest reduction in average success rate from $63.3\%$ down to $26.7\%$ As a common aggregation cancels out the knowledge of the experts in the DGMoE module. + +Figure 5 presents the validation loss of the ablation experiments during the first 500 training epochs. We can see that FedVLA consistently achieves the fastest convergence speed, along with the lowest and most stable validation losses across all tasks. Specifically, we can discover that (1) without IOSP causes validation losses notably increase in the multi-object scenarios (depicted in Figure 5b and Figure 5c), (2) the removal of DGMoE results in significantly increased and fluctuating losses the interaction-rich tasks, (3) removing the EDA module causes relatively higher validation losses in most tasks. These findings strongly support that IOSP helps decompose complex visual scenes into task-relevant object representations, DGMoE module routes the adaptive knowledge, EDA ensures the integrity and effectiveness of the aggregation. The integration of these modules together results in a architecture that supports FedVLA's robustness and adaptability across diverse tasks. + +![](images/1b39fcdc049f9c01fe695682f2ced6b37a9dccdb68f513eacd765d960adef0c4.jpg) +(a) OpenDrawer + +![](images/8908323b8fc7a92162f28910efaf4b3eac7dd9f7da53318f93c22794a1ca093c.jpg) +(b) Sorting Pills + +![](images/92e15a63c2cc7bbedc895ee3cdbf4b1742f8105388b8c8543b4a68b0c512567d.jpg) +(c)Trash Collection + +![](images/fd52ec18fc227287ca6c603c525477c6d8787087d4398eeeb4791a7052a20221.jpg) +(d) Clean Up + +![](images/18bfc2db1917015802b6753fc3818732878a18e59f16cc10db9c0d962daa0454.jpg) +(a) OpenDrawer + +![](images/5cc549dc7830d07a30f94e6c4af71392b579a2bfee876f4ac8eb9220bb7790c5.jpg) +Figure 5. Validation loss comparison in ablation study across four tasks. The x-axis represents the training rounds, while the y-axis represents the validation loss. Each subplot includes a zoomed-in inset focusing on the validation loss fluctuations between rounds 400 and 500, illustrating the fine-grained differences in convergence trends. +(b) Sorting Pills + +![](images/e4ac3ae30df5329201be0cdaf3623bc2891437791451964469ff21055030eba4.jpg) +(c)Trash Collection +Figure 6. Visualization of expert activation distribution per task for different object types. The x-axis represents the three object groups (target objects, surrounding objects, and background objects), where each group contains eight sub-bars corresponding to different experts selected. The left y-axis indicates the activation count, showing how frequently experts are selected for each object type. The right y-axis represents the density of experts per token, illustrating the average number of experts assigned per token in each object category. The dashed line represents the average expert density per token across different object groups, highlighting the dynamic expert selection mechanism of DGMoE. + +![](images/f7f92d7595ba46a56735707137bb573cfa0fb6ef1fe81f744e437838fd8f28cb.jpg) +(d) Clean Up + +# 4.4. Analysis on DGMoE + +The DGMoE is designed to enhance model efficiency by selectively activating experts based on task complexity and their corresponding object types. In contrast, supporting objects and background objects, which provide contextual spatial information, activate fewer experts compared to target objects. Notably, target objects tend to activate experts 4, 7, and 8 more frequently, supporting objects prefer to select experts 6, 7, and 8, whereas background objects favor experts 2, 3, and 5. This demonstrates that DGMoE not only dynamically allocates computational resources based on object relevance but also selectively routes information to specialized experts, enhancing efficiency and adaptability across different object categories. + +Furthermore, we quantitatively measure the average expert density per token across different tasks, also as shown in Figure 6. The proposed DGMoE achieves consistently lower expert activation densities compared to vanilla top- $k$ MoE, which assigns tokens to a fixed number of experts, where $k$ is typically an integer greater than one. The average number of activated experts per token is 1.219, 1.229, 1.225, and 1.227 for Open Drawing, Sorting Pills,Trash Collection, and Clean Up, respectively. Such dynamic and sparse activation directly translates into significantly reduced forward-pass computation, making DGMoE especially suitable for deployment on resource-constrained edge devices. + +# 5. Conclusion + +In this work, we introduced FedVLA, a novel federated vision-language action learning framework that preserves privacy and effectively balances data privacy and task performance. To address the challenges of task heterogeneity and efficient knowledge aggregation in decentralized settings, we proposed three key components: the Instruction-Oriented Scene-Parsing module, which enhances feature extraction by decomposing raw observations into structured object-level representations guided by task instructions. The Dual Gating Mixture-of-Experts, which introduces self-aware experts capable of bidirectional token selection, optimizing computational efficiency while preserving task performance, and the Expert-Driven Aggregation strategy, which dynamically assigns aggregation weights based on expert similarity across clients. Extensive experiments in both simulated and real-world environments demonstrate that FedVLA achieves performance comparable to centralized training while ensuring privacy protection. We believe these results may contribute to the development of more generalizable and scalable federated learning frameworks for robotic manipulation. + +Acknowledgements. This research was supported by Guangdong Basic and Applied Basic Research Foundation (No. 2024A1515011774). + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3 +[2] Martin Albrecht, Melissa Chase, Hao Chen, Jintai Ding, Shafi Goldwasser, Sergey Gorbunov, Shai Halevi, Jeffrey Hoffstein, Kim Laine, Kristin Lauter, et al. Homomorphic encryption standard. Protecting privacy through homomorphic encryption, pages 31-62, 2021. 3 +[3] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. Openflamingo: An opensource framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. 2 +[4] Denis Blessing, Onur Celik, Xiaogang Jia, Moritz Reuss, Maximilian Li, Rudolf Lioutikov, and Gerhard Neumann. Information maximizing curriculum: A curriculum-based approach for learning versatile skills. Advances in Neural Information Processing Systems, 36:51536-51561, 2023. 3 +[5] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818, 2023. 1, 2 +[6] Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa, Soravit Changpinyo, Jialin Wu, Carlos Riquelme Ruiz, Sebastian Goodman, Xiao Wang, Yi Tay, et al. Pali-x: On scaling up a multilingual vision and language model. arXiv preprint arXiv:2305.18565, 2023. 2 +[7] Zitian Chen, Yikang Shen, Mingyu Ding, Zhenfang Chen, Hengshuang Zhao, Erik G Learned-Miller, and Chuang Gan. Mod-squad: Designing mixtures of experts as modular multi-task learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11828-11837, 2023. 2 +[8] Damai Dai, Chengqi Deng, Chenggang Zhao, RX Xu, Huazuo Gao, Deli Chen, Jiashi Li, Wangding Zeng, Xingkai Yu, Y Wu, et al. Deepseekmoe: Towards ultimate expert specialization in mixture-of-experts language models. arXiv preprint arXiv:2401.06066, 2024. 3 +[9] Cynthia Dwork. Differential privacy. In International colloquium on automata, languages, and programming, pages 1-12. Springer, 2006. 3 +[10] Oded Goldreich. Secure multi-party computation. Manuscript. Preliminary version, 78(110):1-108, 1998. 3 +[11] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. Neural computation, 3(1):79-87, 1991. 3 +[12] Shen Jiang, Guanghui Zhu, Yue Wang, Chunfeng Yuan, and Yihua Huang. Automatic multi-task learning framework with neural architecture search in recommendations. In Proceedings of the 30th ACM SIGKDD Conference on Knowl + +edge Discovery and Data Mining, pages 1290-1300, 2024. 2 +[13] Siddharth Karamcheti, Suraj Nair, Ashwin Balakrishna, Percy Liang, Thomas Kollar, and Dorsa Sadigh. Prismatic vlms: Investigating the design space of visually-conditioned language models. arXiv preprint arXiv:2402.07865, 2024. 2 +[14] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, Quan Vuong, Thomas Kollar, Benjamin Burchfiel, Russ Tedrake, Dorsa Sadigh, Sergey Levine, Percy Liang, and Chelsea Finn. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 1, 2 +[15] Xinghang Li, Minghuan Liu, Hanbo Zhang, Cunjun Yu, Jie Xu, Hongtao Wu, Chilam Cheang, Ya Jing, Weinan Zhang, Huaping Liu, et al. Vision-language foundation models as effective robot imitators. arXiv preprint arXiv:2311.01378, 2023. 2 +[16] Bin Lin, Zhenyu Tang, Yang Ye, Jiaxi Cui, Bin Zhu, Peng Jin, Junwu Zhang, Munan Ning, and Li Yuan. Moe-llava: Mixture of experts for large vision-language models. arXiv preprint arXiv:2401.15947, 2024.3 +[17] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2 +[18] Hao Liu, Wilson Yan, Matei Zaharia, and Pieter Abbeel. World model on million-length video and language with ringattention. arXiv e-prints, pages arXiv-2402, 2024. 2 +[19] Yang Liu, Weixing Chen, Yongjie Bai, Xiaodan Liang, Guanbin Li, Wen Gao, and Liang Lin. Aligning cyber space with physical world: A comprehensive survey on embodied ai. arXiv preprint arXiv:2407.06886, 2024. 1 +[20] Yuxiang Lu, Suizhi Huang, Yuwen Yang, Shalayiding Sirejiding, Yue Ding, and Hongtao Lu. Fedhca2: Towards hetero-client federated multi-task learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5599-5609, 2024. 2 +[21] Ajay Mandlekar, Danfei Xu, Josiah Wong, Soroush Nasiriyani, Chen Wang, Rohun Kulkarni, Li Fei-Fei, Silvio Savarese, Yuke Zhu, and Roberto Martin-Martin. What matters in learning from offline human demonstrations for robot manipulation. arXiv preprint arXiv:2108.03298, 2021. 6 +[22] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 1, 2, 3, 6 +[23] David Nadeau and Satoshi Sekine. A survey of named entity recognition and classification. *Lingvisticae Investigationes*, 30(1):3-26, 2007. 4 +[24] Abby O'Neill, Abdul Rehman, Abhiram Maddukuri, Abhishek Gupta, Abhishek Padalkar, Abraham Lee, Acorn Pooley, Agrim Gupta, Ajay Mandlekar, Ajinkya Jain, et al. Open x-embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collaboration 0. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 6892–6903. IEEE, 2024. 2 + +[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 4 +[26] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 2 +[27] Rejin Varghese and M Sambath. Yolov8: A novel object detection algorithm with enhanced performance and robustness. In 2024 International Conference on Advances in Data Engineering and Intelligent Computing Systems (ADICS), pages 1-6. IEEE, 2024. 4 +[28] Lirui Wang, Xinlei Chen, Jialiang Zhao, and Kaiming He. Scaling proprioceptive-visual learning with heterogeneous pre-trained transformers. Advances in Neural Information Processing Systems, 37:124420-124450, 2025. 1, 2, 6 +[29] Jiahui Xu, Lu Sun, and Dengji Zhao. Mone: Mixture-of-masked-experts for efficient multi-task recommendation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2527-2531, 2024. 2 +[30] Tianhe Yu, Deirdre Quillen, Zhanpeng He, Ryan Julian, Karol Hausman, Chelsea Finn, and Sergey Levine. Metaworld: A benchmark and evaluation for multi-task and meta reinforcement learning. In Conference on robot learning, pages 1094-1100. PMLR, 2020. 6 +[31] Michal Zawalski, William Chen, Karl Pertsch, Oier Mees, Chelsea Finn, and Sergey Levine. Robotic control via embodied chain-of-thought reasoning. arXiv preprint arXiv:2407.08693, 2024. 2 +[32] Kaiwen Zhou and Xin Eric Wang. Fedvln: Privacy-preserving federated vision-and-language navigation. In European Conference on Computer Vision, pages 682-699. Springer, 2022. 3 +[33] Weiming Zhuang, Yonggang Wen, Lingjuan Lyu, and Shuai Zhang. Mas: Towards resource-efficient federated multipletask learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23414-23424, 2023. 2 \ No newline at end of file diff --git a/2025/FedVLA_ Federated Vision-Language-Action Learning with Dual Gating Mixture-of-Experts for Robotic Manipulation/images.zip b/2025/FedVLA_ Federated Vision-Language-Action Learning with Dual Gating Mixture-of-Experts for Robotic Manipulation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5619dfe2e75a550b9362c6d86c1aa8ff0454d014 --- /dev/null +++ b/2025/FedVLA_ Federated Vision-Language-Action Learning with Dual Gating Mixture-of-Experts for Robotic Manipulation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d5a2ba928d55079b9c5ba55cda8ca0e82dd26d14e79f1afc99c19f6df7638671 +size 396041 diff --git a/2025/FedVLA_ Federated Vision-Language-Action Learning with Dual Gating Mixture-of-Experts for Robotic Manipulation/layout.json b/2025/FedVLA_ Federated Vision-Language-Action Learning with Dual Gating Mixture-of-Experts for Robotic Manipulation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f30fff35b145b06a1cfa6f3629dc5ddfb4f8d859 --- /dev/null +++ b/2025/FedVLA_ Federated Vision-Language-Action Learning with Dual Gating Mixture-of-Experts for Robotic Manipulation/layout.json @@ -0,0 +1,9914 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 82, + 102, + 531, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 102, + 531, + 140 + ], + "spans": [ + { + "bbox": [ + 82, + 102, + 531, + 140 + ], + "type": "text", + "content": "FedVLA: Federated Vision-Language-Action Learning with Dual Gating Mixture-of-Experts for Robotic Manipulation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 163, + 160, + 443, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 163, + 160, + 443, + 232 + ], + "spans": [ + { + "bbox": [ + 163, + 160, + 443, + 232 + ], + "type": "text", + "content": "Cui Miao" + }, + { + "bbox": [ + 163, + 160, + 443, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 163, + 160, + 443, + 232 + ], + "type": "text", + "content": " Tao Chang" + }, + { + "bbox": [ + 163, + 160, + 443, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 163, + 160, + 443, + 232 + ], + "type": "text", + "content": " Meihan Wu" + }, + { + "bbox": [ + 163, + 160, + 443, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 163, + 160, + 443, + 232 + ], + "type": "text", + "content": " Hongbin Xu" + }, + { + "bbox": [ + 163, + 160, + 443, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 163, + 160, + 443, + 232 + ], + "type": "text", + "content": " Chun Li" + }, + { + "bbox": [ + 163, + 160, + 443, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 163, + 160, + 443, + 232 + ], + "type": "text", + "content": " Ming Li" + }, + { + "bbox": [ + 163, + 160, + 443, + 232 + ], + "type": "inline_equation", + "content": "^{4*}" + }, + { + "bbox": [ + 163, + 160, + 443, + 232 + ], + "type": "text", + "content": " Xiaodong Wang" + }, + { + "bbox": [ + 163, + 160, + 443, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 163, + 160, + 443, + 232 + ], + "type": "text", + "content": " National University of Defense Technology \n" + }, + { + "bbox": [ + 163, + 160, + 443, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 163, + 160, + 443, + 232 + ], + "type": "text", + "content": "Bytedance Seed Shenzhen MSU-BIT University \nLong Laboratory of Artificial Intelligence and Digital Economics" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 283, + 296, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 296, + 620 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 296, + 620 + ], + "type": "text", + "content": "Vision-language-action (VLA) models have significantly advanced robotic manipulation by enabling robots to interpret language instructions for task execution. However, training these models often relies on large-scale user-specific data, raising concerns about privacy and security, which in turn limits their broader adoption. To address this, we propose FedVLA, the first federated VLA learning framework, enabling distributed model training that preserves data privacy without compromising performance. Our framework integrates task-aware representation learning, adaptive expert selection, and expert-driven federated aggregation, enabling efficient and privacy-preserving training of VLA models. Specifically, we introduce an Instruction-Oriented Scene-Parsing mechanism, which decomposes and enhances object-level features based on task instructions, improving contextual understanding. To effectively learn diverse task patterns, we design a Dual Gating Mixture-of-Experts (DGMoE) mechanism, where not only input tokens but also self-aware experts adaptively decide their activation. Finally, we propose an Expert-Driven Aggregation strategy at the federated server, where model aggregation is guided by activated experts, ensuring effective cross-client knowledge transfer. Extensive simulations and real-world robotic experiments demonstrate the effectiveness of our proposals. Notably, DGMoE significantly improves computational efficiency compared to its vanilla counterpart, while FedVLA achieves task success rates comparable to centralized training, effectively preserving data privacy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 640, + 135, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 640, + 135, + 653 + ], + "spans": [ + { + "bbox": [ + 56, + 640, + 135, + 653 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 660, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 296, + 696 + ], + "type": "text", + "content": "Vision-language-action (VLA) models, which integrate visual perception, linguistic understanding, and robotic control, have significantly enhanced robotic manipulation by" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 326, + 258, + 426, + 393 + ], + "blocks": [ + { + "bbox": [ + 326, + 258, + 426, + 393 + ], + "lines": [ + { + "bbox": [ + 326, + 258, + 426, + 393 + ], + "spans": [ + { + "bbox": [ + 326, + 258, + 426, + 393 + ], + "type": "image", + "image_path": "3cd25ba72ac5b657c4f1fe871e3689f016b13fae0b1aa74685f7b1e332cfc4dd.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 394, + 422, + 405 + ], + "lines": [ + { + "bbox": [ + 326, + 394, + 422, + 405 + ], + "spans": [ + { + "bbox": [ + 326, + 394, + 422, + 405 + ], + "type": "text", + "content": "(a) Centralized VLA learning" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 414, + 555, + 491 + ], + "lines": [ + { + "bbox": [ + 313, + 414, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 555, + 491 + ], + "type": "text", + "content": "Figure 1. Comparison between centralized and federated VLA training. Centralized training collects all user data on the cloud, raising privacy concerns due to potential exposure of sensitive information. Our federated VLA framework enables decentralized training on user devices, preserving privacy while utilizing expert-driven aggregation to enhance model generalization across diverse tasks." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 433, + 258, + 545, + 392 + ], + "blocks": [ + { + "bbox": [ + 433, + 258, + 545, + 392 + ], + "lines": [ + { + "bbox": [ + 433, + 258, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 433, + 258, + 545, + 392 + ], + "type": "image", + "image_path": "2ce0c8bce08eb2f30ee766efc3ed1934089c63469e74c0ee47f8c907d7572d49.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 394, + 534, + 405 + ], + "lines": [ + { + "bbox": [ + 443, + 394, + 534, + 405 + ], + "spans": [ + { + "bbox": [ + 443, + 394, + 534, + 405 + ], + "type": "text", + "content": "(b) Federated VLA learning" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 518, + 555, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 555, + 614 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 555, + 614 + ], + "type": "text", + "content": "enabling robots to comprehend and execute complex tasks through natural language instructions [5, 14, 28]. However, training these models requires extensive indoor datasets that capture diverse, user-specific environments, raising serious concerns about data privacy. While recent advancements have improved the capability of VLA models in robotic systems, the issue of privacy leakage remains largely unexplored [19]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "Federated learning (FL) provides an effective privacy-preserving solution for training VLA models [22]. Unlike traditional centralized training, which requires aggregating all user data on a central server, FL enables distributed model training across multiple clients without transferring raw data. As illustrated in Figure 1a, centralized VLA training collects and processes user data in the cloud, posing risks of privacy leakage, such as exposure of personal pref" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 143, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 143, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 143, + 713 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6904" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "erences and sensitive health conditions. In contrast, figure 1b presents a federated VLA learning paradigm, where local models are trained on user devices, and only model updates are transmitted to the central server. This decentralized paradigm mitigates privacy risks while allowing collaborative model optimization." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 147, + 295, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 147, + 295, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 147, + 295, + 314 + ], + "type": "text", + "content": "Existing federated learning approaches are not well-suited for VLA learning, where robotic manipulation tasks exhibit substantial task heterogeneity across different clients. Typical FL methods, such as FedAvg [22], naively aggregate models across tasks, considering only shared synergies while overlooking their fundamental distinctions. Although recent works attempt to address this issue through merge-and-split training [33] or decoupled model aggregation [20], they remain limited to single-modal settings, assuming a uniform input modality across tasks. In contrast, VLA models operate in multi-modal environments, requiring the joint processing of visual observations, language instructions, and robotic actions, which significantly increases the complexity of federated training." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 316, + 295, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 295, + 437 + ], + "type": "text", + "content": "Additionally, mixture of experts (MoE) has shown promise in multi-task learning by assigning specialized experts to different tasks [7, 12, 29]. However, existing methods typically select a fixed number of experts, lacking adaptability to tasks of varying complexity. This rigidity makes them inefficient for resource-constrained clients, where optimizing inference speed is critical for real-time robotic responses. These limitations highlight the need for a task-adaptive and flexible FL framework, specifically designed for multi-modal robotic learning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 438, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 295, + 715 + ], + "type": "text", + "content": "In this paper, we introduce FedVLA, a novel federated vision-language-action (VLA) learning framework with Dual Gating MoE, designed to preserve privacy while maintaining high task performance, as illustrated in Figure 1b. To address task heterogeneity and enable effective knowledge aggregation in federated settings, we propose three key components. First, we introduce an Instruction-Oriented Scene-Parsing (IOSP) module to enhance task-aware feature extraction in decentralized environments, where raw images across clients exhibit diverse backgrounds and object distributions. IOSP decomposes observation images into object-level representations guided by task instructions and leverages vision-language alignment techniques to improve contextual understanding. Second, to enhance computational efficiency in federated learning, we develop a Dual Gating MoE (DGMoE) mechanism that enables adaptive knowledge routing. Unlike conventional MoE, which assigns tokens to a fixed number of experts, DGMoE introduces self-aware experts that can accept or reject token assignments, enabling a bidirectional selection process. This dual gating mechanism improves efficiency while preserving task performance. Lastly, to effectively aggregate similar knowledge from diverse client models, we propose the" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "Expert-Driven Aggregation (EDA) strategy, which dynamically assigns aggregation weights based on expert similarity across clients. By prioritizing updates from clients that activate similar experts, EDA ensures that semantically aligned knowledge is more prominently integrated at the server side, resulting in a more coherent and specialized global model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 145, + 553, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 145, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 145, + 553, + 168 + ], + "type": "text", + "content": "Our main contributions in this work can be summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 170, + 553, + 372 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 170, + 553, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 170, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 314, + 170, + 553, + 252 + ], + "type": "text", + "content": "- We propose FedVLA, the first privacy-preserving federated learning framework for VLA training, ensuring that user data remains private and never shared externally. Extensive experiments in both simulation and real-world environments demonstrate that FedVLA achieves performance comparable to centralized training while preserving data privacy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 253, + 553, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 253, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 314, + 253, + 553, + 311 + ], + "type": "text", + "content": "- We introduce the Dual Gating Mixture-of-Experts, where self-aware experts autonomously determine their participation, enabling bidirectional token selection to optimize computational efficiency while maintaining task performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 313, + 553, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 313, + 553, + 372 + ], + "spans": [ + { + "bbox": [ + 314, + 313, + 553, + 372 + ], + "type": "text", + "content": "- We propose the Expert-Driven Aggregation strategy, which dynamically assigns aggregation weights based on expert similarity across clients, ensuring that semantically aligned knowledge is more effectively integrated, thereby enhancing generalization across heterogeneous tasks." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 384, + 405, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 384, + 405, + 396 + ], + "spans": [ + { + "bbox": [ + 314, + 384, + 405, + 396 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 405, + 487, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 405, + 487, + 417 + ], + "spans": [ + { + "bbox": [ + 314, + 405, + 487, + 417 + ], + "type": "text", + "content": "2.1. Vision-Language-Action Models" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 422, + 553, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 553, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 553, + 650 + ], + "type": "text", + "content": "Vision-Language Models (VLMs), which are pre-trained on large-scale datasets, have demonstrated remarkable capabilities in understanding and generating multimodal data [17, 18, 26]. Inspired by VLMs, Vision-Language-Action (VLA) models extend their capabilities by integrating robotic action prediction. For instance, RoboFlamingo [15], RT-2 [5], and OpenVLA [14] are all built upon VLMs such as OpenFlamingo [3], PaLI-X [6], and Prismatic [13], respectively. Recent work has begun exploring the training of generalist robots [28], followed by fine-tuning on diverse robot datasets, enabling generalization to novel instructions, unseen objects, and distribution shifts [31]. This progress relies heavily on the availability of large-scale public datasets, such as Open-X Embodiment dataset [24]. During the finetuning phase, samples from real-world scenarios are required, which are often generated by human activities and distributed across users' home environments. However, these samples frequently expose privacy-sensitive information." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 659, + 470, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 659, + 470, + 672 + ], + "spans": [ + { + "bbox": [ + 314, + 659, + 470, + 672 + ], + "type": "text", + "content": "2.2. Privacy-Preserving Training" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Traditional centralized learning methods require uploading data from multiple sources, which poses significant privacy risks. To mitigate these concerns, researchers have pro" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6905" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 556, + 257 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 556, + 257 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 556, + 257 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 556, + 257 + ], + "type": "image", + "image_path": "69ecaee2a94e1b08f56c1b73f3ce3fdc656be67076c1ddc3fbd69875c5ca0d3b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 264, + 555, + 344 + ], + "lines": [ + { + "bbox": [ + 54, + 264, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 54, + 264, + 555, + 344 + ], + "type": "text", + "content": "Figure 2. Illustration of the proposed FedVLA framework and its key component. (a) An overview of the FedVLA, which consists of multiple clients and a server. Each client processes vision-language-action inputs through the Stem, Trunk, and Head modules. Clients send their updated trunk parameters to the server, where Expert-Driven Aggregation selectively integrates expert-specific updates on server. The global trunk module is then distributed to clients for the next training round. (b) Details of the proposed Dual Gating MoE module. The input tokens first pass through the token-level gate " + }, + { + "bbox": [ + 54, + 264, + 555, + 344 + ], + "type": "inline_equation", + "content": "G_{t}" + }, + { + "bbox": [ + 54, + 264, + 555, + 344 + ], + "type": "text", + "content": ", computing the gating scores for each expert with residual scores from the previous MoE layer. Only the computed scores surpass the expert-level gate " + }, + { + "bbox": [ + 54, + 264, + 555, + 344 + ], + "type": "inline_equation", + "content": "G_{e}" + }, + { + "bbox": [ + 54, + 264, + 555, + 344 + ], + "type": "text", + "content": ", activating the corresponding expert to process the token. Finally, the expert outputs are combined to generate the final output tokens." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 363, + 297, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 363, + 297, + 555 + ], + "spans": [ + { + "bbox": [ + 54, + 363, + 297, + 555 + ], + "type": "text", + "content": "posed various privacy-preserving techniques, including federated learning (FL) [22], differential privacy [9], secure multi-party computation [10], and homomorphic encryption [2]. However, Most privacy-preserving training techniques have been developed for single-modality tasks, and seldomly study the data privacy of embodied AI [32]. In robotic manipulation, the models process multimodal sensor data, including images, language instructions, and action sequences, which can contain sensitive user interaction data. Unlike static datasets, robotic systems continuously collect and process dynamic, task-specific information, making privacy protection more complex. In this context, FL as a solution by allowing decentralized model training without sharing raw data. However, FL for robotic manipulation remains challenging due to task heterogeneity and the need for efficient model aggregation across clients." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 563, + 248, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 248, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 248, + 577 + ], + "type": "text", + "content": "2.3. Mixture of Experts in Large Models" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "Mixture of experts [4, 11] is an architecture where multiple experts are trained collaboratively, with each expert responsible for a specific portion of the input space. An MoE consists of two components: a routing mechanism and a set of experts. The experts can either be part of a feedforward network (FFN) or the entire FFN layer, and the routing mechanism allows each token to select experts for processing. By activating a specific number of experts, MoE can reduce resource consumption and improve inference speed. This feature makes MoE particularly effective when integrated into large models, as seen in models like LLaVA-MoE [16]," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 363, + 555, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 555, + 411 + ], + "type": "text", + "content": "DeepSeekMoE [8], and GPT-4 [1]. In this work, we apply MoE to a generalist robot VLA model and introduce a self-aware variant that can adaptively decide whether to be selected, significantly improving deployment efficiency." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 420, + 373, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 373, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 373, + 433 + ], + "type": "text", + "content": "3. FedVLA" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 441, + 556, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 556, + 560 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 556, + 560 + ], + "type": "text", + "content": "We aim to train a vision-language-action model in a federated learning (FL) framework, ensuring privacy preservation while enabling task generalization across clients. Formally, given " + }, + { + "bbox": [ + 313, + 441, + 556, + 560 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 441, + 556, + 560 + ], + "type": "text", + "content": " clients " + }, + { + "bbox": [ + 313, + 441, + 556, + 560 + ], + "type": "inline_equation", + "content": "\\mathbb{C} = \\{C_1,C_2,\\dots,C_N\\}" + }, + { + "bbox": [ + 313, + 441, + 556, + 560 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 441, + 556, + 560 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 441, + 556, + 560 + ], + "type": "text", + "content": " local tasks " + }, + { + "bbox": [ + 313, + 441, + 556, + 560 + ], + "type": "inline_equation", + "content": "\\mathbb{T} = \\{T_1,T_2,\\dots,T_N\\}" + }, + { + "bbox": [ + 313, + 441, + 556, + 560 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 441, + 556, + 560 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 441, + 556, + 560 + ], + "type": "text", + "content": " local datasets " + }, + { + "bbox": [ + 313, + 441, + 556, + 560 + ], + "type": "inline_equation", + "content": "\\mathbb{D} = \\{\\mathbb{D}_1,\\mathbb{D}_2,\\dots,\\mathbb{D}_N\\}" + }, + { + "bbox": [ + 313, + 441, + 556, + 560 + ], + "type": "text", + "content": ", each client " + }, + { + "bbox": [ + 313, + 441, + 556, + 560 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 313, + 441, + 556, + 560 + ], + "type": "text", + "content": " is associated with a local task " + }, + { + "bbox": [ + 313, + 441, + 556, + 560 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 313, + 441, + 556, + 560 + ], + "type": "text", + "content": " and a corresponding dataset " + }, + { + "bbox": [ + 313, + 441, + 556, + 560 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_i" + }, + { + "bbox": [ + 313, + 441, + 556, + 560 + ], + "type": "text", + "content": ". The dataset " + }, + { + "bbox": [ + 313, + 441, + 556, + 560 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_i" + }, + { + "bbox": [ + 313, + 441, + 556, + 560 + ], + "type": "text", + "content": " contains visual perception image " + }, + { + "bbox": [ + 313, + 441, + 556, + 560 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 313, + 441, + 556, + 560 + ], + "type": "text", + "content": ", language instruction " + }, + { + "bbox": [ + 313, + 441, + 556, + 560 + ], + "type": "inline_equation", + "content": "l_{i}" + }, + { + "bbox": [ + 313, + 441, + 556, + 560 + ], + "type": "text", + "content": " of the task, proprioception " + }, + { + "bbox": [ + 313, + 441, + 556, + 560 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 313, + 441, + 556, + 560 + ], + "type": "text", + "content": " of the robot, and action sequence " + }, + { + "bbox": [ + 313, + 441, + 556, + 560 + ], + "type": "inline_equation", + "content": "a_{i}" + }, + { + "bbox": [ + 313, + 441, + 556, + 560 + ], + "type": "text", + "content": " executed by the robot:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 354, + 562, + 555, + 575 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 562, + 555, + 575 + ], + "spans": [ + { + "bbox": [ + 354, + 562, + 555, + 575 + ], + "type": "interline_equation", + "content": "\\mathbb {D} _ {i} = \\left\\{\\left(o _ {i}, l _ {i}, s _ {i}, a _ {i}\\right) _ {j} | j = 1, 2, \\dots , M \\right\\}, \\tag {1}", + "image_path": "7820a249c6fdebfe90579496bf9ff3b85c65b3ecf81d55d57d5e81ce8fce183b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 578, + 555, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 555, + 614 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 555, + 614 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 578, + 555, + 614 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 578, + 555, + 614 + ], + "type": "text", + "content": " is the number of samples in dataset " + }, + { + "bbox": [ + 313, + 578, + 555, + 614 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_i" + }, + { + "bbox": [ + 313, + 578, + 555, + 614 + ], + "type": "text", + "content": ". To train a federated VLA model " + }, + { + "bbox": [ + 313, + 578, + 555, + 614 + ], + "type": "inline_equation", + "content": "\\theta_G" + }, + { + "bbox": [ + 313, + 578, + 555, + 614 + ], + "type": "text", + "content": ", the overall federated optimization objective can be formulated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 326, + 616, + 555, + 651 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 616, + 555, + 651 + ], + "spans": [ + { + "bbox": [ + 326, + 616, + 555, + 651 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta_ {G}} \\sum_ {i = 1} ^ {N} w _ {i} \\sum_ {(o _ {i}, l _ {i}, a _ {i}, s _ {i}) \\in \\mathbb {D} _ {i}} \\mathcal {L} _ {i} (f _ {i} (\\theta_ {G}; o _ {i}, l _ {i}, s _ {i}), a _ {i}), \\tag {2}", + "image_path": "62604772b22c71d19c4e22e570a9a2cee7a394f60590e5b01369c198338a6d2b.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 653, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 712 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 653, + 556, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_i" + }, + { + "bbox": [ + 313, + 653, + 556, + 712 + ], + "type": "text", + "content": " is the Huber loss function computed over each client model's local dataset " + }, + { + "bbox": [ + 313, + 653, + 556, + 712 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_i" + }, + { + "bbox": [ + 313, + 653, + 556, + 712 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 653, + 556, + 712 + ], + "type": "inline_equation", + "content": "f_i" + }, + { + "bbox": [ + 313, + 653, + 556, + 712 + ], + "type": "text", + "content": " represents the local client model's predicted action sequence, and " + }, + { + "bbox": [ + 313, + 653, + 556, + 712 + ], + "type": "inline_equation", + "content": "w_i" + }, + { + "bbox": [ + 313, + 653, + 556, + 712 + ], + "type": "text", + "content": " denotes the weight assigned to each client during aggregation, which is further detailed in Section 3.3." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6906" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 294, + 135 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 294, + 135 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 294, + 135 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 294, + 135 + ], + "type": "image", + "image_path": "0c5055f55c82abdd9529d01f59dbe56cd508721fe706971b900405a0997767a6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 144, + 295, + 191 + ], + "lines": [ + { + "bbox": [ + 55, + 144, + 295, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 295, + 191 + ], + "type": "text", + "content": "Figure 3. Instruction-Oriented Scene-Parsing module, which decomposes observations into object-level representations guided by task instructions, enhancing task-aware feature extraction through vision-language alignment." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 210, + 296, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 210, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 210, + 296, + 354 + ], + "type": "text", + "content": "The overview of our FedVLA framework is shown in Figure 2. The clients train their models locally and communicate with the server per round until the model converges. The following sections introduce our key proposals: (1) Instruction-Oriented Scene-Parsing (Section 3.1) that decomposes the observation image into object-level embeddings guided by task instructions. (2) Dual Gating Mixture-of-Experts (Section 3.2) that can adaptively select a varying number of experts for tokens based on the relevance to the specific task. (3) Expert-Driven Aggregation (Section 3.3) that dynamically aggregates the weights based on expert similarities across different clients in the server." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 361, + 246, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 246, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 246, + 373 + ], + "type": "text", + "content": "3.1. Instruction-Oriented Scene-Parsing" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 378, + 296, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 296, + 449 + ], + "type": "text", + "content": "To ensure the robot focuses on task-relevant information while maintaining contextual awareness to avoid collisions, we decouple objects in the scene based on the specific robotic task. To achieve this, we propose an Instruction-Oriented Scene-Parsing (IOSP) module to represent objects within the task scene." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 451, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 295, + 629 + ], + "type": "text", + "content": "As illustrated in Figure 3, given an instruction and an image, we first extract Target Objects (TOs) from the instruction using named entity recognition [23]. Simultaneously, we employ YOLOv8 [27] to detect and classify Foreground Objects (FOs) and Background Objects (BOs) in the image. The scene objects are categorized into three groups based on the cosine similarity measured by a pre-trained CLIP model [25] between the instruction and object names, following these matching rules: (1) Target Objects (TOs): Objects explicitly mentioned in the instruction and directly involved in the action. (2) Surrounding Objects (SOs): Foreground objects not in TOs, providing spatial and contextual cues. (3) Background Objects (BOs): Static environmental elements ensuring scene consistency but not directly affecting the task." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": "Then, we encode the grouped objects using the CLIP model. The text encoder " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": " maps object names from TOs, SOs, and BOs into textual embeddings " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "H(\\mathrm{TOs}) \\in \\mathbb{R}^{t \\times D}" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "H(\\mathrm{BOs}) \\in \\mathbb{R}^{t \\times D}" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "H(\\mathrm{SOs}) \\in \\mathbb{R}^{t \\times D}" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": ". Meanwhile, the vision encoder " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": " processes the entire image into " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "P(X) \\in \\mathbb{R}^{t \\times D}" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": ". We assign image tokens to object groups by computing the cosine similarity between token and text embed" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "content": "dings, selecting the top eight tokens per group. To enhance intra-group feature representation, each group is processed through an MoE module, producing refined image token embeddings. The enhanced tokens are then concatenated with the remaining tokens for subsequent processing. This IOSP module enables the model to focus on critical objects, improving task-specific manipulation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 162, + 488, + 175 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 162, + 488, + 175 + ], + "spans": [ + { + "bbox": [ + 313, + 162, + 488, + 175 + ], + "type": "text", + "content": "3.2. Dual Gating Mixture-of-Experts" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 179, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 179, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 179, + 553, + 251 + ], + "type": "text", + "content": "In this section, we introduce Dual Gating Mixture-of-Expert (DGMoE) mechanism that supports adaptive knowledge routing. Different from vanilla MoE, which assigns a fixed number of experts to each token, our DGMoE can automatically determine the acceptance or rejection of the token assignments under the Self-Aware Expert." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 251, + 554, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 554, + 454 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 554, + 454 + ], + "type": "text", + "content": "As shown in Figure 2, each DGMoE layer is comprised of " + }, + { + "bbox": [ + 313, + 251, + 554, + 454 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 251, + 554, + 454 + ], + "type": "text", + "content": " self-aware experts " + }, + { + "bbox": [ + 313, + 251, + 554, + 454 + ], + "type": "inline_equation", + "content": "\\mathbb{E} = \\{E_1,E_2,\\dots ,E_K\\}" + }, + { + "bbox": [ + 313, + 251, + 554, + 454 + ], + "type": "text", + "content": ". To enable the automatic token-driven expert selection, we introduce two gating mechanisms: Token-side gating mechanism " + }, + { + "bbox": [ + 313, + 251, + 554, + 454 + ], + "type": "inline_equation", + "content": "G_{t}" + }, + { + "bbox": [ + 313, + 251, + 554, + 454 + ], + "type": "text", + "content": " and Expert-side gating mechanism " + }, + { + "bbox": [ + 313, + 251, + 554, + 454 + ], + "type": "inline_equation", + "content": "G_{e}" + }, + { + "bbox": [ + 313, + 251, + 554, + 454 + ], + "type": "text", + "content": ". (1) For Token-side gating, " + }, + { + "bbox": [ + 313, + 251, + 554, + 454 + ], + "type": "inline_equation", + "content": "G_{t}" + }, + { + "bbox": [ + 313, + 251, + 554, + 454 + ], + "type": "text", + "content": " serves as the gating module on the token side, comprised of a soft router module and a gating residual module. The soft router first estimates the selection scores for the hidden inputs, indicating the token preference for different experts. Then the estimated scores from previous layer are incorporated with the scores of current layer, allowing the tokens to inherit the selection prior of the most suitable experts from previous layers. (2) For Expert-side gating, " + }, + { + "bbox": [ + 313, + 251, + 554, + 454 + ], + "type": "inline_equation", + "content": "G_{e}" + }, + { + "bbox": [ + 313, + 251, + 554, + 454 + ], + "type": "text", + "content": " can autonomously accept or reject tokens based on their acceptance threshold. The selection scores predicted by " + }, + { + "bbox": [ + 313, + 251, + 554, + 454 + ], + "type": "inline_equation", + "content": "G_{t}" + }, + { + "bbox": [ + 313, + 251, + 554, + 454 + ], + "type": "text", + "content": " is filtered with a learnable threshold, and only the experts with the most relevant tokens are activated." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 455, + 554, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 554, + 491 + ], + "type": "text", + "content": "Given an input token " + }, + { + "bbox": [ + 313, + 455, + 554, + 491 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 455, + 554, + 491 + ], + "type": "text", + "content": ", the token-side gate " + }, + { + "bbox": [ + 313, + 455, + 554, + 491 + ], + "type": "inline_equation", + "content": "G_{t}" + }, + { + "bbox": [ + 313, + 455, + 554, + 491 + ], + "type": "text", + "content": " first computes the scores at the " + }, + { + "bbox": [ + 313, + 455, + 554, + 491 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 455, + 554, + 491 + ], + "type": "text", + "content": "-th layer by combining the soft router output and the gating residual module:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 330, + 495, + 555, + 528 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 495, + 555, + 528 + ], + "spans": [ + { + "bbox": [ + 330, + 495, + 555, + 528 + ], + "type": "interline_equation", + "content": "G _ {t} ^ {j} (x) = \\left\\{ \\begin{array}{l l} W _ {t} ^ {j} x, & \\text {i f} j = 1, \\\\ W _ {t} ^ {j} x + W _ {t} ^ {g} G _ {t} ^ {(j - 1)} (x), & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {3}", + "image_path": "9f4d23eb5674c8cc6adca82502152a5d871d0e79ff29707438b467dadd514484.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 533, + 554, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 554, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 554, + 629 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 533, + 554, + 629 + ], + "type": "inline_equation", + "content": "W_{t}^{j}\\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 313, + 533, + 554, + 629 + ], + "type": "text", + "content": " is the trainable weight matrix for soft router, and " + }, + { + "bbox": [ + 313, + 533, + 554, + 629 + ], + "type": "inline_equation", + "content": "W_{t}^{g}\\in \\mathbb{R}^{N\\times N}" + }, + { + "bbox": [ + 313, + 533, + 554, + 629 + ], + "type": "text", + "content": " is a trainable transformation matrix designed to aggregate the scores from the previous layer and the current layer, ensuring smooth expertise knowledge transitions across layers. Then, the token-to-expert selection scores " + }, + { + "bbox": [ + 313, + 533, + 554, + 629 + ], + "type": "inline_equation", + "content": "s_t(x)" + }, + { + "bbox": [ + 313, + 533, + 554, + 629 + ], + "type": "text", + "content": " are obtained by normalizing " + }, + { + "bbox": [ + 313, + 533, + 554, + 629 + ], + "type": "inline_equation", + "content": "G_{t}(x)" + }, + { + "bbox": [ + 313, + 533, + 554, + 629 + ], + "type": "text", + "content": " with a softmax function, which determines the probability distribution over the available experts." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 630, + 553, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 553, + 666 + ], + "type": "text", + "content": "On the self-aware expert side, the expert-side gate " + }, + { + "bbox": [ + 313, + 630, + 553, + 666 + ], + "type": "inline_equation", + "content": "G_{e}" + }, + { + "bbox": [ + 313, + 630, + 553, + 666 + ], + "type": "text", + "content": " evaluates the token scores and determines the final expert activation scores " + }, + { + "bbox": [ + 313, + 630, + 553, + 666 + ], + "type": "inline_equation", + "content": "s_{e}(x)" + }, + { + "bbox": [ + 313, + 630, + 553, + 666 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 374, + 671, + 553, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 671, + 553, + 685 + ], + "spans": [ + { + "bbox": [ + 374, + 671, + 553, + 685 + ], + "type": "interline_equation", + "content": "s _ {e} (x) = \\operatorname {s i g n} \\left(s _ {t} (x) - \\lambda W _ {e}\\right), \\tag {4}", + "image_path": "498da8de64a0e3f11c0daf61b8540cfffbd98bbcba920b7a1a550600f106ed2d.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "W_{e}" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": " is a set of trainable self-aware parameters corresponding to the gating values of each expert, and " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": " is a" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6907" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 118 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 118 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 118 + ], + "type": "text", + "content": "scaling factor that modulates the gating threshold. In our setting, we set " + }, + { + "bbox": [ + 55, + 72, + 295, + 118 + ], + "type": "inline_equation", + "content": "\\lambda = 0.5" + }, + { + "bbox": [ + 55, + 72, + 295, + 118 + ], + "type": "text", + "content": " to balance expert selection, ensuring neither excessive expert activation nor overly strict filtering of tokens." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 120, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 120, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 120, + 296, + 156 + ], + "type": "text", + "content": "Then, combining the dual gating mechanism of token-side and exper-side gating, the expert selection function is defined as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 110, + 159, + 295, + 191 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 159, + 295, + 191 + ], + "spans": [ + { + "bbox": [ + 110, + 159, + 295, + 191 + ], + "type": "interline_equation", + "content": "g (x) = \\left\\{ \\begin{array}{l l} s _ {t} (x), & \\text {i f} s _ {e} (x) > 0, \\\\ 0, & \\text {i f} s _ {e} (x) \\leq 0, \\end{array} \\right. \\tag {5}", + "image_path": "544d6406e705c9637d5493d91c863078ae41830d3d2756a52bff2634b41058d7.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 197, + 296, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 197, + 296, + 231 + ], + "spans": [ + { + "bbox": [ + 55, + 197, + 296, + 231 + ], + "type": "text", + "content": "where the final output token " + }, + { + "bbox": [ + 55, + 197, + 296, + 231 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 197, + 296, + 231 + ], + "type": "text", + "content": " of the DGMoE layer can be computed as the weighted sum of the selected expert outputs:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 224, + 295, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 295, + 255 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 295, + 255 + ], + "type": "interline_equation", + "content": "y = \\sum_ {i = 1} ^ {K} g _ {i} (x) E _ {i} (x). \\tag {6}", + "image_path": "2ea12ebbda2dcde172e72a1ef935c367934f1eb18c64f157d8d25a31e3ea4673.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 258, + 296, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 296, + 306 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 296, + 306 + ], + "type": "text", + "content": "With the help of self-aware expert gating mechanism, our DGMoE can dynamically activate the most relevant experts, reducing computational overhead while maintaining task-relevant feature learning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 312, + 207, + 325 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 207, + 325 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 207, + 325 + ], + "type": "text", + "content": "3.3. Expert-Driven Aggregation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 329, + 296, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 329, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 329, + 296, + 413 + ], + "type": "text", + "content": "To mitigate performance degradation caused by task heterogeneity, we propose an Expert-Driven Aggregation (EDA) strategy, which aggregates model parameters based on the similarity of expert selections across clients. And as shown in Figure 2, we selectively aggregates only the trunk module, while the stem and head modules remain personalized and are excluded from aggregation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 413, + 296, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 296, + 487 + ], + "type": "text", + "content": "Given a trunk module with " + }, + { + "bbox": [ + 55, + 413, + 296, + 487 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 413, + 296, + 487 + ], + "type": "text", + "content": " layers, where each layer contains " + }, + { + "bbox": [ + 55, + 413, + 296, + 487 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 413, + 296, + 487 + ], + "type": "text", + "content": " experts, each client " + }, + { + "bbox": [ + 55, + 413, + 296, + 487 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 55, + 413, + 296, + 487 + ], + "type": "text", + "content": " records an expert selection matrix " + }, + { + "bbox": [ + 55, + 413, + 296, + 487 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_i\\in \\mathbb{R}^{L\\times K}" + }, + { + "bbox": [ + 55, + 413, + 296, + 487 + ], + "type": "text", + "content": " during each round through DGMoE, where each element " + }, + { + "bbox": [ + 55, + 413, + 296, + 487 + ], + "type": "inline_equation", + "content": "V_{i}^{(l,k)}" + }, + { + "bbox": [ + 55, + 413, + 296, + 487 + ], + "type": "text", + "content": " represents the number of times expert " + }, + { + "bbox": [ + 55, + 413, + 296, + 487 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 413, + 296, + 487 + ], + "type": "text", + "content": " is activated in the " + }, + { + "bbox": [ + 55, + 413, + 296, + 487 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 413, + 296, + 487 + ], + "type": "text", + "content": "-th layer. We extract the expert selection vector for layer " + }, + { + "bbox": [ + 55, + 413, + 296, + 487 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 413, + 296, + 487 + ], + "type": "text", + "content": " of client " + }, + { + "bbox": [ + 55, + 413, + 296, + 487 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 55, + 413, + 296, + 487 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 491, + 295, + 509 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 491, + 295, + 509 + ], + "spans": [ + { + "bbox": [ + 104, + 491, + 295, + 509 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {i} ^ {(l)} = \\left(V _ {i} ^ {(l, 1)}, V _ {i} ^ {(l, 2)}, \\dots , V _ {i} ^ {(l, K)}\\right). \\tag {7}", + "image_path": "92955606f55dcf010e0a8df2d2f7c3a9273d4da92e798c416150163b6f63072d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 514, + 296, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 296, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 296, + 550 + ], + "type": "text", + "content": "Then we compute the pairwise similarity between clients based on expert selection vectors. The similarity between clients " + }, + { + "bbox": [ + 55, + 514, + 296, + 550 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 55, + 514, + 296, + 550 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 514, + 296, + 550 + ], + "type": "inline_equation", + "content": "C_j" + }, + { + "bbox": [ + 55, + 514, + 296, + 550 + ], + "type": "text", + "content": " at layer " + }, + { + "bbox": [ + 55, + 514, + 296, + 550 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 514, + 296, + 550 + ], + "type": "text", + "content": " is defined as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 555, + 295, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 295, + 590 + ], + "type": "interline_equation", + "content": "s _ {i, j} ^ {(l)} = \\frac {\\mathbf {v} _ {i} ^ {(l)} \\cdot \\mathbf {v} _ {j} ^ {(l)}}{\\| \\mathbf {v} _ {i} ^ {(l)} \\| \\| \\mathbf {v} _ {j} ^ {(l)} \\|\\}.} \\tag {8}", + "image_path": "45872fb897346d9beb3c61f16a66ff5e05d67579d5cd081052c3e84146b2c83a.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 595, + 295, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 595, + 295, + 619 + ], + "spans": [ + { + "bbox": [ + 55, + 595, + 295, + 619 + ], + "type": "text", + "content": "Using this similarity, we determine the aggregation weight " + }, + { + "bbox": [ + 55, + 595, + 295, + 619 + ], + "type": "inline_equation", + "content": "w_{l,i}" + }, + { + "bbox": [ + 55, + 595, + 295, + 619 + ], + "type": "text", + "content": " for each client " + }, + { + "bbox": [ + 55, + 595, + 295, + 619 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 55, + 595, + 295, + 619 + ], + "type": "text", + "content": " at layer " + }, + { + "bbox": [ + 55, + 595, + 295, + 619 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 595, + 295, + 619 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 625, + 295, + 661 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 625, + 295, + 661 + ], + "spans": [ + { + "bbox": [ + 116, + 625, + 295, + 661 + ], + "type": "interline_equation", + "content": "w _ {l, i} = \\frac {\\sum_ {C _ {j} \\in C} s _ {i , j} ^ {(l)}}{\\sum_ {C _ {j} \\in C} \\sum_ {C _ {i} \\in C} s _ {i , j} ^ {(l)}}. \\tag {9}", + "image_path": "bfa55f44681511c2d56ed4631105317cce4fb3fd2c2426d2ee17b2fe8279d850.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "This weighting strategy ensures that clients with higher expert selection similarity contribute more to each other's model updates, allowing for more effective and task-aligned aggregation across task-heterogeneous clients." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 72, + 445, + 83 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 445, + 83 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 445, + 83 + ], + "type": "text", + "content": "Algorithm 1 FedVLA: Clients" + } + ] + } + ], + "index": 15, + "type": "text" + }, + { + "bbox": [ + 313, + 85, + 555, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 85, + 555, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 85, + 555, + 133 + ], + "type": "text", + "content": "Input: Each client " + }, + { + "bbox": [ + 313, + 85, + 555, + 133 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 313, + 85, + 555, + 133 + ], + "type": "text", + "content": " has a local task " + }, + { + "bbox": [ + 313, + 85, + 555, + 133 + ], + "type": "inline_equation", + "content": "T_i" + }, + { + "bbox": [ + 313, + 85, + 555, + 133 + ], + "type": "text", + "content": " and dataset " + }, + { + "bbox": [ + 313, + 85, + 555, + 133 + ], + "type": "inline_equation", + "content": "D_i = \\{o_i, l_i, s_i, a_i\\}" + }, + { + "bbox": [ + 313, + 85, + 555, + 133 + ], + "type": "text", + "content": ", along with local parameters " + }, + { + "bbox": [ + 313, + 85, + 555, + 133 + ], + "type": "inline_equation", + "content": "\\phi_i" + }, + { + "bbox": [ + 313, + 85, + 555, + 133 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 85, + 555, + 133 + ], + "type": "inline_equation", + "content": "w_i" + }, + { + "bbox": [ + 313, + 85, + 555, + 133 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 85, + 555, + 133 + ], + "type": "inline_equation", + "content": "\\theta_i" + }, + { + "bbox": [ + 313, + 85, + 555, + 133 + ], + "type": "text", + "content": " corresponding to the stem, trunk, and head modules, respectively." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 318, + 147, + 553, + 461 + ], + "type": "list", + "angle": 0, + "index": 40, + "blocks": [ + { + "bbox": [ + 321, + 147, + 432, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 147, + 432, + 158 + ], + "spans": [ + { + "bbox": [ + 321, + 147, + 432, + 158 + ], + "type": "text", + "content": "1: Local_Train " + }, + { + "bbox": [ + 321, + 147, + 432, + 158 + ], + "type": "inline_equation", + "content": "(C_i, \\phi_i, \\theta_i)" + }, + { + "bbox": [ + 321, + 147, + 432, + 158 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 159, + 489, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 159, + 489, + 171 + ], + "spans": [ + { + "bbox": [ + 321, + 159, + 489, + 171 + ], + "type": "text", + "content": "2: for each local epoch " + }, + { + "bbox": [ + 321, + 159, + 489, + 171 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 321, + 159, + 489, + 171 + ], + "type": "text", + "content": " from 1 to " + }, + { + "bbox": [ + 321, + 159, + 489, + 171 + ], + "type": "inline_equation", + "content": "E_{c}" + }, + { + "bbox": [ + 321, + 159, + 489, + 171 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 171, + 541, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 171, + 541, + 183 + ], + "spans": [ + { + "bbox": [ + 321, + 171, + 541, + 183 + ], + "type": "text", + "content": "3: for each batch " + }, + { + "bbox": [ + 321, + 171, + 541, + 183 + ], + "type": "inline_equation", + "content": "(o_i, l_i, s_i, a_i)" + }, + { + "bbox": [ + 321, + 171, + 541, + 183 + ], + "type": "text", + "content": " sampled from " + }, + { + "bbox": [ + 321, + 171, + 541, + 183 + ], + "type": "inline_equation", + "content": "D_i" + }, + { + "bbox": [ + 321, + 171, + 541, + 183 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 187, + 534, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 187, + 534, + 198 + ], + "spans": [ + { + "bbox": [ + 321, + 187, + 534, + 198 + ], + "type": "text", + "content": "4: // Step 1: Instruction-Oriented Scene-Parsing" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 200, + 460, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 200, + 460, + 210 + ], + "spans": [ + { + "bbox": [ + 321, + 200, + 460, + 210 + ], + "type": "text", + "content": "5: TOs, SOs, BOs " + }, + { + "bbox": [ + 321, + 200, + 460, + 210 + ], + "type": "inline_equation", + "content": "\\leftarrow \\{o_i, l_i\\}" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 212, + 535, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 212, + 535, + 223 + ], + "spans": [ + { + "bbox": [ + 321, + 212, + 535, + 223 + ], + "type": "text", + "content": "6: " + }, + { + "bbox": [ + 321, + 212, + 535, + 223 + ], + "type": "inline_equation", + "content": "F_{T O s}, F_{S O s}, F_{B O s} \\gets \\mathrm{CLIP}(T O s, S O s, B O s)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 224, + 422, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 224, + 422, + 236 + ], + "spans": [ + { + "bbox": [ + 321, + 224, + 422, + 236 + ], + "type": "text", + "content": "7: " + }, + { + "bbox": [ + 321, + 224, + 422, + 236 + ], + "type": "inline_equation", + "content": "F_{O}\\gets \\mathrm{CLIP}(o_{i})" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 237, + 507, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 237, + 507, + 249 + ], + "spans": [ + { + "bbox": [ + 321, + 237, + 507, + 249 + ], + "type": "text", + "content": "8: " + }, + { + "bbox": [ + 321, + 237, + 507, + 249 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{TO}} \\gets \\mathrm{topk}(\\mathrm{cosine}(F_{\\mathrm{TOs}}, F_O), F_O)" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 250, + 506, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 250, + 506, + 262 + ], + "spans": [ + { + "bbox": [ + 321, + 250, + 506, + 262 + ], + "type": "text", + "content": "9: " + }, + { + "bbox": [ + 321, + 250, + 506, + 262 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{SO}} \\gets \\mathrm{topk}(\\mathrm{cosine}(F_{\\mathrm{SOs}}, F_O), F_O)" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 318, + 263, + 507, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 263, + 507, + 274 + ], + "spans": [ + { + "bbox": [ + 318, + 263, + 507, + 274 + ], + "type": "text", + "content": "10: " + }, + { + "bbox": [ + 318, + 263, + 507, + 274 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{BO}} \\gets \\mathrm{topk}(\\mathrm{cosine}(F_{\\mathrm{BOs}}, F_O), F_O)" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 318, + 276, + 504, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 276, + 504, + 288 + ], + "spans": [ + { + "bbox": [ + 318, + 276, + 504, + 288 + ], + "type": "text", + "content": "11: " + }, + { + "bbox": [ + 318, + 276, + 504, + 288 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{in}} = \\mathrm{concat}(X_{\\mathrm{TO}},X_{\\mathrm{SO}},X_{\\mathrm{BO}},s_i)" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 318, + 292, + 503, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 292, + 503, + 303 + ], + "spans": [ + { + "bbox": [ + 318, + 292, + 503, + 303 + ], + "type": "text", + "content": "12: // Step 2: Forward pass with DGMoE" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 318, + 304, + 437, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 304, + 437, + 316 + ], + "spans": [ + { + "bbox": [ + 318, + 304, + 437, + 316 + ], + "type": "text", + "content": "13: Let " + }, + { + "bbox": [ + 318, + 304, + 437, + 316 + ], + "type": "inline_equation", + "content": "x \\in MHA(X_{\\mathrm{in}})" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 352, + 316, + 531, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 316, + 531, + 329 + ], + "spans": [ + { + "bbox": [ + 352, + 316, + 531, + 329 + ], + "type": "inline_equation", + "content": "s_t(x) = \\mathrm{softmax}\\bigl (W_t(x)\\bigr)" + }, + { + "bbox": [ + 352, + 316, + 531, + 329 + ], + "type": "text", + "content": " // token-gate" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 318, + 331, + 534, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 331, + 534, + 342 + ], + "spans": [ + { + "bbox": [ + 318, + 331, + 534, + 342 + ], + "type": "text", + "content": "14: " + }, + { + "bbox": [ + 318, + 331, + 534, + 342 + ], + "type": "inline_equation", + "content": "s_e(x) = \\mathrm{sign}(s_t(x) - \\lambda W_e)" + }, + { + "bbox": [ + 318, + 331, + 534, + 342 + ], + "type": "text", + "content": " // expert-gate" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 318, + 343, + 553, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 343, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 318, + 343, + 553, + 354 + ], + "type": "text", + "content": "15: " + }, + { + "bbox": [ + 318, + 343, + 553, + 354 + ], + "type": "inline_equation", + "content": "g(x) = s_t(x)\\cdot \\mathbb{I}(s_e(x) > 0)" + }, + { + "bbox": [ + 318, + 343, + 553, + 354 + ], + "type": "text", + "content": " // expert activation" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 318, + 356, + 521, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 356, + 521, + 386 + ], + "spans": [ + { + "bbox": [ + 318, + 356, + 521, + 386 + ], + "type": "text", + "content": "16: " + }, + { + "bbox": [ + 318, + 356, + 521, + 386 + ], + "type": "inline_equation", + "content": "\\mathbf{v}^{(l)},y = \\sum_{k = 1}^{K}g(x)E_k(x)" + }, + { + "bbox": [ + 318, + 356, + 521, + 386 + ], + "type": "text", + "content": " //final output" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 318, + 390, + 494, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 390, + 494, + 401 + ], + "spans": [ + { + "bbox": [ + 318, + 390, + 494, + 401 + ], + "type": "text", + "content": "17: // Step 3: Model parameter updates" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 318, + 402, + 472, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 402, + 472, + 413 + ], + "spans": [ + { + "bbox": [ + 318, + 402, + 472, + 413 + ], + "type": "text", + "content": "18: " + }, + { + "bbox": [ + 318, + 402, + 472, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_i = \\mathrm{Loss}(f(\\phi_i, \\theta_i, w_i), a_i)" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 318, + 415, + 497, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 415, + 497, + 426 + ], + "spans": [ + { + "bbox": [ + 318, + 415, + 497, + 426 + ], + "type": "text", + "content": "19: " + }, + { + "bbox": [ + 318, + 415, + 497, + 426 + ], + "type": "inline_equation", + "content": "(\\phi_i,\\theta_i,w_i)\\gets (\\phi_i,\\theta_i,w_i) - \\eta \\nabla \\mathcal{L}_i" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 318, + 427, + 375, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 427, + 375, + 437 + ], + "spans": [ + { + "bbox": [ + 318, + 427, + 375, + 437 + ], + "type": "text", + "content": "20: end for" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 318, + 439, + 365, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 439, + 365, + 449 + ], + "spans": [ + { + "bbox": [ + 318, + 439, + 365, + 449 + ], + "type": "text", + "content": "21: end for" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 318, + 450, + 443, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 450, + 443, + 461 + ], + "spans": [ + { + "bbox": [ + 318, + 450, + 443, + 461 + ], + "type": "text", + "content": "22: return " + }, + { + "bbox": [ + 318, + 450, + 443, + 461 + ], + "type": "inline_equation", + "content": "\\theta, \\mathbf{v}^{(l)}" + }, + { + "bbox": [ + 318, + 450, + 443, + 461 + ], + "type": "text", + "content": " to the server." + } + ] + } + ], + "index": 39 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 495, + 390, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 495, + 390, + 508 + ], + "spans": [ + { + "bbox": [ + 314, + 495, + 390, + 508 + ], + "type": "text", + "content": "3.4. Algorithms" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 313, + 517, + 555, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 555, + 624 + ], + "type": "text", + "content": "Clients. The overall workflow of our FedVLA is shown in Algorithm 1 and Algorithm 2. At the start of each training round " + }, + { + "bbox": [ + 313, + 517, + 555, + 624 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 517, + 555, + 624 + ], + "type": "text", + "content": ", each client processes task-specific data using Instruction-Oriented Scene-Parsing to extract structured features, followed by Dual Gating MoE for adaptive expert selection. The updated trunk parameters and expert selection counts are then transmitted to the server for global aggregation. The detailed client-side process is described in Algorithm 1." + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "text", + "content": "Server. On the server side, the server receives expert selection statistics and trunk updates from all participating clients and performs Expert-Driven Aggregation, which dynamically assigns aggregation weights based on expert selection similarities. The aggregated global trunk module is then redistributed to clients for the next training round. The process is elaborated in Algorithm 2." + } + ] + } + ], + "index": 43 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6908" + } + ] + } + ], + "index": 44 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 185, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 185, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 185, + 83 + ], + "type": "text", + "content": "Algorithm 2 FedVLA: Server" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 85, + 296, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 85, + 296, + 122 + ], + "spans": [ + { + "bbox": [ + 55, + 85, + 296, + 122 + ], + "type": "text", + "content": "Input: " + }, + { + "bbox": [ + 55, + 85, + 296, + 122 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 85, + 296, + 122 + ], + "type": "text", + "content": " is the number of training rounds. " + }, + { + "bbox": [ + 55, + 85, + 296, + 122 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 85, + 296, + 122 + ], + "type": "text", + "content": " represents the global trunk parameters. Each client provides " + }, + { + "bbox": [ + 55, + 85, + 296, + 122 + ], + "type": "inline_equation", + "content": "\\mathbf{v}^{(l)}" + }, + { + "bbox": [ + 55, + 85, + 296, + 122 + ], + "type": "text", + "content": ", the expert selection counts vector for each layer " + }, + { + "bbox": [ + 55, + 85, + 296, + 122 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 85, + 296, + 122 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 135, + 291, + 303 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 62, + 135, + 145, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 135, + 145, + 145 + ], + "spans": [ + { + "bbox": [ + 62, + 135, + 145, + 145 + ], + "type": "text", + "content": "1: Server Execute:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 147, + 203, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 147, + 203, + 158 + ], + "spans": [ + { + "bbox": [ + 62, + 147, + 203, + 158 + ], + "type": "text", + "content": "2: for each round " + }, + { + "bbox": [ + 62, + 147, + 203, + 158 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 62, + 147, + 203, + 158 + ], + "type": "text", + "content": " from 1 to " + }, + { + "bbox": [ + 62, + 147, + 203, + 158 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 62, + 147, + 203, + 158 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 159, + 203, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 159, + 203, + 171 + ], + "spans": [ + { + "bbox": [ + 62, + 159, + 203, + 171 + ], + "type": "text", + "content": "3: // Expert-Driven Aggregation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 172, + 184, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 172, + 184, + 182 + ], + "spans": [ + { + "bbox": [ + 62, + 172, + 184, + 182 + ], + "type": "text", + "content": "4: for each trunk layer " + }, + { + "bbox": [ + 62, + 172, + 184, + 182 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 62, + 172, + 184, + 182 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 183, + 228, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 183, + 228, + 195 + ], + "spans": [ + { + "bbox": [ + 62, + 183, + 228, + 195 + ], + "type": "text", + "content": "5: for each pair of clients " + }, + { + "bbox": [ + 62, + 183, + 228, + 195 + ], + "type": "inline_equation", + "content": "C_i, C_j" + }, + { + "bbox": [ + 62, + 183, + 228, + 195 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 195, + 291, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 195, + 291, + 218 + ], + "spans": [ + { + "bbox": [ + 62, + 195, + 291, + 218 + ], + "type": "text", + "content": "6: " + }, + { + "bbox": [ + 62, + 195, + 291, + 218 + ], + "type": "inline_equation", + "content": "s_{i,j}^{(l)} = \\frac{\\mathbf{v}_i^{(l)}\\cdot\\mathbf{v}_j^{(l)}}{\\|\\mathbf{v}_i^{(l)}\\|\\|\\mathbf{v}_j^{(l)}\\|} //" + }, + { + "bbox": [ + 62, + 195, + 291, + 218 + ], + "type": "text", + "content": " expert selection similarity" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 220, + 289, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 220, + 289, + 245 + ], + "spans": [ + { + "bbox": [ + 62, + 220, + 289, + 245 + ], + "type": "text", + "content": "7: " + }, + { + "bbox": [ + 62, + 220, + 289, + 245 + ], + "type": "inline_equation", + "content": "w^{(l,i)} = \\frac{\\sum_{j\\in C}s_{i,j}^{(l)}}{\\sum_{j\\in C}\\sum_{i\\in C}s_{i,j}^{(l)}}" + }, + { + "bbox": [ + 62, + 220, + 289, + 245 + ], + "type": "text", + "content": " // aggregation weight" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 246, + 126, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 246, + 126, + 255 + ], + "spans": [ + { + "bbox": [ + 62, + 246, + 126, + 255 + ], + "type": "text", + "content": "8: end for" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 255, + 254, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 255, + 254, + 269 + ], + "spans": [ + { + "bbox": [ + 62, + 255, + 254, + 269 + ], + "type": "text", + "content": "9: " + }, + { + "bbox": [ + 62, + 255, + 254, + 269 + ], + "type": "inline_equation", + "content": "\\theta^{(l)}\\gets \\sum_{i = 1}^{N}w^{(l,i)}\\cdot \\theta^{(l,i)}" + }, + { + "bbox": [ + 62, + 255, + 254, + 269 + ], + "type": "text", + "content": " //update " + }, + { + "bbox": [ + 62, + 255, + 254, + 269 + ], + "type": "inline_equation", + "content": "\\theta" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 61, + 270, + 116, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 270, + 116, + 279 + ], + "spans": [ + { + "bbox": [ + 61, + 270, + 116, + 279 + ], + "type": "text", + "content": "10: end for" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 61, + 281, + 220, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 281, + 220, + 293 + ], + "spans": [ + { + "bbox": [ + 61, + 281, + 220, + 293 + ], + "type": "text", + "content": "11: Broadcast updated " + }, + { + "bbox": [ + 61, + 281, + 220, + 293 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 61, + 281, + 220, + 293 + ], + "type": "text", + "content": " to all clients." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 59, + 293, + 106, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 293, + 106, + 303 + ], + "spans": [ + { + "bbox": [ + 59, + 293, + 106, + 303 + ], + "type": "text", + "content": "12: end for" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 327, + 137, + 341 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 137, + 341 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 137, + 341 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 54, + 348, + 296, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 348, + 296, + 504 + ], + "spans": [ + { + "bbox": [ + 54, + 348, + 296, + 504 + ], + "type": "text", + "content": "Implementation Details. We validate our FedVLA in both simulation and real-world scenarios, where each client performs a single task, resulting in a non-iid data distribution. We employ the pretrained HPT [28] as the backbone of our VLA model and train it for 1,000 communication rounds between clients and the server, with each client performing local training for 5 epochs per round. In simulation, the model is trained with a learning rate of " + }, + { + "bbox": [ + 54, + 348, + 296, + 504 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-6}" + }, + { + "bbox": [ + 54, + 348, + 296, + 504 + ], + "type": "text", + "content": ", while in real-world settings, the learning rate is set to " + }, + { + "bbox": [ + 54, + 348, + 296, + 504 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-5}" + }, + { + "bbox": [ + 54, + 348, + 296, + 504 + ], + "type": "text", + "content": ". Each client trains locally with a batch size of 256 using the Adam optimizer. The experiments are conducted on a workstation equipped with 2 GeForce GTX 4090Ti GPUs, 48 Intel Xeon CPUs, and 128GB of memory." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 54, + 505, + 296, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 505, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 54, + 505, + 296, + 613 + ], + "type": "text", + "content": "Evaluation. For evaluation, the success and failure of a trial are recorded as 1 and 0. Each task takes for 15 trials and repeats for 5 times. Since this is the first work on federated VLA learning, we compare FedVLA with two methods. The centralized training (Centralized) aggregates all data on a single server for joint training, serving as an upper bound in performance. The federated averaging (FedAvg) [22] is a strong FL baseline where models are trained locally on each client and simply averaged on the server." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 623, + 129, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 129, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 129, + 635 + ], + "type": "text", + "content": "4.1. Simulation" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "Environment. We utilize the MuJoCo engine [21] to simulate tasks from the Meta-World robot manipulation benchmark [30]. All tasks are performed using a Sawyer robot with a parallel gripper, positioned at the center of a dinner table. Visual data is captured by three fixed RGB cameras with a resolution of " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": ", providing corner, top, and" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": "shoulder views. For collision detection and dynamics simulation, we employ official physics engines to ensure accurate robotic interactions within the simulation environment. Tasks. We evaluate our FedVLA on four household tasks: Door Lock, Close Printer, Sweep Into, and Open Window. For each task, we collect approximately 30-80 episodes, each consisting of 40-100 steps. The initial pose of robot and environment layouts are randomized to assess the generalization ability of FedVLA." + } + ] + } + ], + "index": 20 + }, + { + "type": "table", + "bbox": [ + 319, + 190, + 550, + 249 + ], + "blocks": [ + { + "bbox": [ + 319, + 190, + 550, + 249 + ], + "lines": [ + { + "bbox": [ + 319, + 190, + 550, + 249 + ], + "spans": [ + { + "bbox": [ + 319, + 190, + 550, + 249 + ], + "type": "table", + "html": "
MethodAvg. SuccessDoor LockCloseDrawerSweep IntoOpen Window
Centralized65.086.773.353.346.7
FedAvg51.766.773.340.026.7
FedVLA (Ours)63.380.080.053.340.0
", + "image_path": "5233da4f7d05f1bbbdea441724eabaf99a4c2a03ae3d1f78e99e334a27d94481.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 345, + 258, + 523, + 268 + ], + "lines": [ + { + "bbox": [ + 345, + 258, + 523, + 268 + ], + "spans": [ + { + "bbox": [ + 345, + 258, + 523, + 268 + ], + "type": "text", + "content": "Table 1. Simulation evaluation across four tasks." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 281, + 555, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 449 + ], + "type": "text", + "content": "Performance Comparison. The evaluation results in the simulation environment are presented in Table 1. FedVLA achieves an average success rate of " + }, + { + "bbox": [ + 313, + 281, + 555, + 449 + ], + "type": "inline_equation", + "content": "63.3\\%" + }, + { + "bbox": [ + 313, + 281, + 555, + 449 + ], + "type": "text", + "content": ", closely approaching the " + }, + { + "bbox": [ + 313, + 281, + 555, + 449 + ], + "type": "inline_equation", + "content": "65.0\\%" + }, + { + "bbox": [ + 313, + 281, + 555, + 449 + ], + "type": "text", + "content": " of centralized training. Notably, in the Close Drawn task, FedVLA surpasses centralized training by a success rate margin of about " + }, + { + "bbox": [ + 313, + 281, + 555, + 449 + ], + "type": "inline_equation", + "content": "7.0\\%" + }, + { + "bbox": [ + 313, + 281, + 555, + 449 + ], + "type": "text", + "content": ". Furthermore, FedVLA consistently outperforms FedAvg, which only achieves an average success rate of " + }, + { + "bbox": [ + 313, + 281, + 555, + 449 + ], + "type": "inline_equation", + "content": "51.7\\%" + }, + { + "bbox": [ + 313, + 281, + 555, + 449 + ], + "type": "text", + "content": ". This performance gap can be attributed to FedAvg overlooking task heterogeneity in federated learning. In contrast, FedVLA explicitly accounts for task heterogeneity during aggregation by leveraging expert-driven model updates, ensuring that knowledge from diverse clients is effectively integrated to enhance generalization across different tasks." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 456, + 391, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 391, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 391, + 468 + ], + "type": "text", + "content": "4.2. Real-World" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 474, + 555, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 581 + ], + "type": "text", + "content": "Hardware. For real-world experiments, we employ a UR3 robotic arm with six degrees of freedom, equipped with a one-DoF parallel gripper. A RealSense D435i RGB-D camera is mounted above the robot, capturing RGB images in real time at a resolution of " + }, + { + "bbox": [ + 313, + 474, + 555, + 581 + ], + "type": "inline_equation", + "content": "1280 \\times 720" + }, + { + "bbox": [ + 313, + 474, + 555, + 581 + ], + "type": "text", + "content": " with a frame rate of approximately " + }, + { + "bbox": [ + 313, + 474, + 555, + 581 + ], + "type": "inline_equation", + "content": "30\\mathrm{Hz}" + }, + { + "bbox": [ + 313, + 474, + 555, + 581 + ], + "type": "text", + "content": ". To simulate real-world variations, we introduce moderate randomness in object placement and lighting conditions to better reflect practical deployment scenarios." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "type": "text", + "content": "Tasks. We collect real-world robotic demonstrations for household-related tasks, including Clean Up,Trash Collection, OpenDrawer,and Sorting Pills, as shown in Figure 4. For data collection, we employ an expert teaching method, using inverse kinematics to compute the robot poses corresponding to key point data while recording RGB images and pose data throughout the trajectory. Each task includes approximately 50 demonstrations, with each trajectory consisting of 20 to 80 steps." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 689, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 712 + ], + "type": "text", + "content": "Performance Comparison. The real-world experimental results are summarized in Table 2. FedVLA achieves an" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6909" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 293, + 205 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 293, + 205 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 293, + 205 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 293, + 205 + ], + "type": "image", + "image_path": "575e642ab35fa047dadbb235fe3d844d2a04e1091cf58c1b5194923d6c3ee09c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 213, + 296, + 291 + ], + "lines": [ + { + "bbox": [ + 55, + 213, + 296, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 296, + 291 + ], + "type": "text", + "content": "Figure 4. Real-World Tasks. Each column represents a different task, while each row demonstrates keyframes of task execution. From left to right, the tasks are: Clean Up, where the robot wipes the trash from plate using a cloth. Trash Collection, in which it throws crumpled paper into a trash bin. Open Printer, which requires grasping and pulling a drawer open, and Sorting Pills, where pill bottles are placed into trays of designated color." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 312, + 295, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 312, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 54, + 312, + 295, + 529 + ], + "type": "text", + "content": "average task success rate of " + }, + { + "bbox": [ + 54, + 312, + 295, + 529 + ], + "type": "inline_equation", + "content": "63.3\\%" + }, + { + "bbox": [ + 54, + 312, + 295, + 529 + ], + "type": "text", + "content": ", closely matching the " + }, + { + "bbox": [ + 54, + 312, + 295, + 529 + ], + "type": "inline_equation", + "content": "63.4\\%" + }, + { + "bbox": [ + 54, + 312, + 295, + 529 + ], + "type": "text", + "content": " of centralized training. Notably, it attains the same success rates as the centralized model in complex tasks such as Trash Collection " + }, + { + "bbox": [ + 54, + 312, + 295, + 529 + ], + "type": "inline_equation", + "content": "(46.7\\%)" + }, + { + "bbox": [ + 54, + 312, + 295, + 529 + ], + "type": "text", + "content": " and Sorting Pills " + }, + { + "bbox": [ + 54, + 312, + 295, + 529 + ], + "type": "inline_equation", + "content": "(73.3\\%)" + }, + { + "bbox": [ + 54, + 312, + 295, + 529 + ], + "type": "text", + "content": ". In contrast, FedAvg exhibits a significant performance drop across all tasks, with an average success rate of " + }, + { + "bbox": [ + 54, + 312, + 295, + 529 + ], + "type": "inline_equation", + "content": "53.3\\%" + }, + { + "bbox": [ + 54, + 312, + 295, + 529 + ], + "type": "text", + "content": ", which is " + }, + { + "bbox": [ + 54, + 312, + 295, + 529 + ], + "type": "inline_equation", + "content": "10.0\\%" + }, + { + "bbox": [ + 54, + 312, + 295, + 529 + ], + "type": "text", + "content": " lower than our method. These results indicate that FedVLA effectively bridges the performance gap between federated and centralized training. While FedAvg merely averages model parameters across clients, it struggles to generalize effectively in VLA learning, where multimodal information—including textual instructions, scene images, and execution trajectories—is involved, requiring strong feature extraction capabilities. The proposed Dual Gating MoE and Expert-Driven Aggregation mechanisms enable the model to adaptively learn diverse patterns for robotic manipulation, enhancing its ability to handle complex tasks in federated settings." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 59, + 537, + 292, + 594 + ], + "blocks": [ + { + "bbox": [ + 59, + 537, + 292, + 594 + ], + "lines": [ + { + "bbox": [ + 59, + 537, + 292, + 594 + ], + "spans": [ + { + "bbox": [ + 59, + 537, + 292, + 594 + ], + "type": "table", + "html": "
MethodAvg. SuccessClean UpTrash CollectionOpenDrawerSorting Pills
Centralized63.446.746.786.773.3
FedAvg53.346.740.060.066.7
FedVLA (Ours)63.353.346.780.073.3
", + "image_path": "f5e324c3a5f23658093b504e726fecf94f99b32e2719cd838c2d20fca9ab0ca5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 635, + 156, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 156, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 156, + 647 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "To further explore the effectiveness of the IOSP, DGMOE and EDA in FedVLA, we conduct ablation experiments by individually removing each module while keeping the other components. Specifically, the experimental setups are as follows: (1) remove the IOSP module and directly input the" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 145 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 145 + ], + "type": "text", + "content": "entire scene image and proprioception data into the encoder (w/o IOSP), (2) replace the DGMoE layers with standard FFN, eliminating adaptive expert selection (w/o DGMOE), and (3) perform vanilla FedAvg aggregation instead of EDA (w/o EDA). We compare the task success rate and record the validation loss during the training process." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 318, + 157, + 550, + 222 + ], + "blocks": [ + { + "bbox": [ + 86, + 603, + 265, + 613 + ], + "lines": [ + { + "bbox": [ + 86, + 603, + 265, + 613 + ], + "spans": [ + { + "bbox": [ + 86, + 603, + 265, + 613 + ], + "type": "text", + "content": "Table 2. Real-world evaluation across four tasks." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 157, + 550, + 222 + ], + "lines": [ + { + "bbox": [ + 318, + 157, + 550, + 222 + ], + "spans": [ + { + "bbox": [ + 318, + 157, + 550, + 222 + ], + "type": "table", + "html": "
MethodAvg. SuccessClean UpTrash CollectionOpen DriverSorting Pills
w/o IOSP41.140.013.366.746.7
w/o DGMoE31.720.020.046.740.0
w/o EDA26.726.720.033.326.7
FedVLA (Ours)63.353.346.780.073.3
", + "image_path": "7af02fb37ce2849a58ad62402955c7d235e9366293c7deb5e035e5a279439d9a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 231, + 553, + 253 + ], + "lines": [ + { + "bbox": [ + 313, + 231, + 553, + 253 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 553, + 253 + ], + "type": "text", + "content": "Table 3. Ablation studies of proposed FedVLA without IOSP, DG-MoE, and EDA." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 268, + 555, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 268, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 268, + 555, + 506 + ], + "type": "text", + "content": "From Table 3, we can see that each module contributes significantly to the performance of FedVLA, and the lack of each module will lead to a decrease in the performance of the entire system. Specifically, removing IOSP leads to the highest performance drop in the Trash Collection task from " + }, + { + "bbox": [ + 313, + 268, + 555, + 506 + ], + "type": "inline_equation", + "content": "46.7\\%" + }, + { + "bbox": [ + 313, + 268, + 555, + 506 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 268, + 555, + 506 + ], + "type": "inline_equation", + "content": "13.3\\%" + }, + { + "bbox": [ + 313, + 268, + 555, + 506 + ], + "type": "text", + "content": ", because the number of objects in Trash Collection task is more than the others. The IOSP module discriminates each object from the other for taking the right action to adapt to multi-objects scenarios. DGMoE module strongly relates to the Clean Up task, leading to a success rate drop from " + }, + { + "bbox": [ + 313, + 268, + 555, + 506 + ], + "type": "inline_equation", + "content": "53.3\\%" + }, + { + "bbox": [ + 313, + 268, + 555, + 506 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 268, + 555, + 506 + ], + "type": "inline_equation", + "content": "20.0\\%" + }, + { + "bbox": [ + 313, + 268, + 555, + 506 + ], + "type": "text", + "content": ". The Clean Up task is harder in the manipulation aspect as it requires the robot interaction for multiple steps. The FedVLA depends on the DGMoE module to adaptively select the appropriate number of experts based on different types of tokens, enabling collaborative learning for interaction-rich tasks. Furthermore, the removal of EDA gives the largest reduction in average success rate from " + }, + { + "bbox": [ + 313, + 268, + 555, + 506 + ], + "type": "inline_equation", + "content": "63.3\\%" + }, + { + "bbox": [ + 313, + 268, + 555, + 506 + ], + "type": "text", + "content": " down to " + }, + { + "bbox": [ + 313, + 268, + 555, + 506 + ], + "type": "inline_equation", + "content": "26.7\\%" + }, + { + "bbox": [ + 313, + 268, + 555, + 506 + ], + "type": "text", + "content": " As a common aggregation cancels out the knowledge of the experts in the DGMoE module." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": "Figure 5 presents the validation loss of the ablation experiments during the first 500 training epochs. We can see that FedVLA consistently achieves the fastest convergence speed, along with the lowest and most stable validation losses across all tasks. Specifically, we can discover that (1) without IOSP causes validation losses notably increase in the multi-object scenarios (depicted in Figure 5b and Figure 5c), (2) the removal of DGMoE results in significantly increased and fluctuating losses the interaction-rich tasks, (3) removing the EDA module causes relatively higher validation losses in most tasks. These findings strongly support that IOSP helps decompose complex visual scenes into task-relevant object representations, DGMoE module routes the adaptive knowledge, EDA ensures the integrity and effectiveness of the aggregation. The integration of these modules together results in a architecture that supports FedVLA's robustness and adaptability across diverse tasks." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6910" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 70, + 182, + 145 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 182, + 145 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 182, + 145 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 182, + 145 + ], + "type": "image", + "image_path": "1b39fcdc049f9c01fe695682f2ced6b37a9dccdb68f513eacd765d960adef0c4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 96, + 149, + 151, + 159 + ], + "lines": [ + { + "bbox": [ + 96, + 149, + 151, + 159 + ], + "spans": [ + { + "bbox": [ + 96, + 149, + 151, + 159 + ], + "type": "text", + "content": "(a) OpenDrawer" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 184, + 70, + 304, + 147 + ], + "blocks": [ + { + "bbox": [ + 184, + 70, + 304, + 147 + ], + "lines": [ + { + "bbox": [ + 184, + 70, + 304, + 147 + ], + "spans": [ + { + "bbox": [ + 184, + 70, + 304, + 147 + ], + "type": "image", + "image_path": "8908323b8fc7a92162f28910efaf4b3eac7dd9f7da53318f93c22794a1ca093c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 219, + 149, + 270, + 159 + ], + "lines": [ + { + "bbox": [ + 219, + 149, + 270, + 159 + ], + "spans": [ + { + "bbox": [ + 219, + 149, + 270, + 159 + ], + "type": "text", + "content": "(b) Sorting Pills" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 306, + 71, + 425, + 146 + ], + "blocks": [ + { + "bbox": [ + 306, + 71, + 425, + 146 + ], + "lines": [ + { + "bbox": [ + 306, + 71, + 425, + 146 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 425, + 146 + ], + "type": "image", + "image_path": "92e15a63c2cc7bbedc895ee3cdbf4b1742f8105388b8c8543b4a68b0c512567d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 149, + 399, + 159 + ], + "lines": [ + { + "bbox": [ + 334, + 149, + 399, + 159 + ], + "spans": [ + { + "bbox": [ + 334, + 149, + 399, + 159 + ], + "type": "text", + "content": "(c)Trash Collection" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 427, + 71, + 547, + 146 + ], + "blocks": [ + { + "bbox": [ + 427, + 71, + 547, + 146 + ], + "lines": [ + { + "bbox": [ + 427, + 71, + 547, + 146 + ], + "spans": [ + { + "bbox": [ + 427, + 71, + 547, + 146 + ], + "type": "image", + "image_path": "fd52ec18fc227287ca6c603c525477c6d8787087d4398eeeb4791a7052a20221.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 466, + 149, + 509, + 159 + ], + "lines": [ + { + "bbox": [ + 466, + 149, + 509, + 159 + ], + "spans": [ + { + "bbox": [ + 466, + 149, + 509, + 159 + ], + "type": "text", + "content": "(d) Clean Up" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 58, + 214, + 176, + 277 + ], + "blocks": [ + { + "bbox": [ + 58, + 214, + 176, + 277 + ], + "lines": [ + { + "bbox": [ + 58, + 214, + 176, + 277 + ], + "spans": [ + { + "bbox": [ + 58, + 214, + 176, + 277 + ], + "type": "image", + "image_path": "18bfc2db1917015802b6753fc3818732878a18e59f16cc10db9c0d962daa0454.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 89, + 280, + 145, + 289 + ], + "lines": [ + { + "bbox": [ + 89, + 280, + 145, + 289 + ], + "spans": [ + { + "bbox": [ + 89, + 280, + 145, + 289 + ], + "type": "text", + "content": "(a) OpenDrawer" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 183, + 213, + 301, + 277 + ], + "blocks": [ + { + "bbox": [ + 55, + 169, + 555, + 203 + ], + "lines": [ + { + "bbox": [ + 55, + 169, + 555, + 203 + ], + "spans": [ + { + "bbox": [ + 55, + 169, + 555, + 203 + ], + "type": "text", + "content": "Figure 5. Validation loss comparison in ablation study across four tasks. The x-axis represents the training rounds, while the y-axis represents the validation loss. Each subplot includes a zoomed-in inset focusing on the validation loss fluctuations between rounds 400 and 500, illustrating the fine-grained differences in convergence trends." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 183, + 213, + 301, + 277 + ], + "lines": [ + { + "bbox": [ + 183, + 213, + 301, + 277 + ], + "spans": [ + { + "bbox": [ + 183, + 213, + 301, + 277 + ], + "type": "image", + "image_path": "5cc549dc7830d07a30f94e6c4af71392b579a2bfee876f4ac8eb9220bb7790c5.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 216, + 280, + 268, + 289 + ], + "lines": [ + { + "bbox": [ + 216, + 280, + 268, + 289 + ], + "spans": [ + { + "bbox": [ + 216, + 280, + 268, + 289 + ], + "type": "text", + "content": "(b) Sorting Pills" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 309, + 213, + 427, + 277 + ], + "blocks": [ + { + "bbox": [ + 309, + 213, + 427, + 277 + ], + "lines": [ + { + "bbox": [ + 309, + 213, + 427, + 277 + ], + "spans": [ + { + "bbox": [ + 309, + 213, + 427, + 277 + ], + "type": "image", + "image_path": "e4ac3ae30df5329201be0cdaf3623bc2891437791451964469ff21055030eba4.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 280, + 400, + 289 + ], + "lines": [ + { + "bbox": [ + 335, + 280, + 400, + 289 + ], + "spans": [ + { + "bbox": [ + 335, + 280, + 400, + 289 + ], + "type": "text", + "content": "(c)Trash Collection" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 299, + 555, + 365 + ], + "lines": [ + { + "bbox": [ + 55, + 299, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 299, + 555, + 365 + ], + "type": "text", + "content": "Figure 6. Visualization of expert activation distribution per task for different object types. The x-axis represents the three object groups (target objects, surrounding objects, and background objects), where each group contains eight sub-bars corresponding to different experts selected. The left y-axis indicates the activation count, showing how frequently experts are selected for each object type. The right y-axis represents the density of experts per token, illustrating the average number of experts assigned per token in each object category. The dashed line represents the average expert density per token across different object groups, highlighting the dynamic expert selection mechanism of DGMoE." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 434, + 213, + 552, + 277 + ], + "blocks": [ + { + "bbox": [ + 434, + 213, + 552, + 277 + ], + "lines": [ + { + "bbox": [ + 434, + 213, + 552, + 277 + ], + "spans": [ + { + "bbox": [ + 434, + 213, + 552, + 277 + ], + "type": "image", + "image_path": "f7f92d7595ba46a56735707137bb573cfa0fb6ef1fe81f744e437838fd8f28cb.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 472, + 280, + 515, + 289 + ], + "lines": [ + { + "bbox": [ + 472, + 280, + 515, + 289 + ], + "spans": [ + { + "bbox": [ + 472, + 280, + 515, + 289 + ], + "type": "text", + "content": "(d) Clean Up" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 385, + 174, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 385, + 174, + 398 + ], + "spans": [ + { + "bbox": [ + 55, + 385, + 174, + 398 + ], + "type": "text", + "content": "4.4. Analysis on DGMoE" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 54, + 408, + 296, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 408, + 296, + 564 + ], + "spans": [ + { + "bbox": [ + 54, + 408, + 296, + 564 + ], + "type": "text", + "content": "The DGMoE is designed to enhance model efficiency by selectively activating experts based on task complexity and their corresponding object types. In contrast, supporting objects and background objects, which provide contextual spatial information, activate fewer experts compared to target objects. Notably, target objects tend to activate experts 4, 7, and 8 more frequently, supporting objects prefer to select experts 6, 7, and 8, whereas background objects favor experts 2, 3, and 5. This demonstrates that DGMoE not only dynamically allocates computational resources based on object relevance but also selectively routes information to specialized experts, enhancing efficiency and adaptability across different object categories." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "Furthermore, we quantitatively measure the average expert density per token across different tasks, also as shown in Figure 6. The proposed DGMoE achieves consistently lower expert activation densities compared to vanilla top-" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " MoE, which assigns tokens to a fixed number of experts, where " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " is typically an integer greater than one. The average number of activated experts per token is 1.219, 1.229, 1.225, and 1.227 for Open Drawing, Sorting Pills,Trash Collection, and Clean Up, respectively. Such dynamic and sparse activation directly translates into significantly reduced forward-pass computation, making DGMoE especially suitable for deployment on resource-constrained edge devices." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 314, + 385, + 387, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 385, + 387, + 398 + ], + "spans": [ + { + "bbox": [ + 314, + 385, + 387, + 398 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 411, + 555, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 555, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 555, + 662 + ], + "type": "text", + "content": "In this work, we introduced FedVLA, a novel federated vision-language action learning framework that preserves privacy and effectively balances data privacy and task performance. To address the challenges of task heterogeneity and efficient knowledge aggregation in decentralized settings, we proposed three key components: the Instruction-Oriented Scene-Parsing module, which enhances feature extraction by decomposing raw observations into structured object-level representations guided by task instructions. The Dual Gating Mixture-of-Experts, which introduces self-aware experts capable of bidirectional token selection, optimizing computational efficiency while preserving task performance, and the Expert-Driven Aggregation strategy, which dynamically assigns aggregation weights based on expert similarity across clients. Extensive experiments in both simulated and real-world environments demonstrate that FedVLA achieves performance comparable to centralized training while ensuring privacy protection. We believe these results may contribute to the development of more generalizable and scalable federated learning frameworks for robotic manipulation." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Acknowledgements. This research was supported by Guangdong Basic and Applied Basic Research Foundation (No. 2024A1515011774)." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "6911" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "type": "text", + "content": "[2] Martin Albrecht, Melissa Chase, Hao Chen, Jintai Ding, Shafi Goldwasser, Sergey Gorbunov, Shai Halevi, Jeffrey Hoffstein, Kim Laine, Kristin Lauter, et al. Homomorphic encryption standard. Protecting privacy through homomorphic encryption, pages 31-62, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 205, + 296, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 296, + 269 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 296, + 269 + ], + "type": "text", + "content": "[3] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. Openflamingo: An opensource framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 271, + 295, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 271, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 62, + 271, + 295, + 327 + ], + "type": "text", + "content": "[4] Denis Blessing, Onur Celik, Xiaogang Jia, Moritz Reuss, Maximilian Li, Rudolf Lioutikov, and Gerhard Neumann. Information maximizing curriculum: A curriculum-based approach for learning versatile skills. Advances in Neural Information Processing Systems, 36:51536-51561, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 328, + 295, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 328, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 62, + 328, + 295, + 392 + ], + "type": "text", + "content": "[5] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818, 2023. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 396, + 295, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 396, + 295, + 451 + ], + "spans": [ + { + "bbox": [ + 62, + 396, + 295, + 451 + ], + "type": "text", + "content": "[6] Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa, Soravit Changpinyo, Jialin Wu, Carlos Riquelme Ruiz, Sebastian Goodman, Xiao Wang, Yi Tay, et al. Pali-x: On scaling up a multilingual vision and language model. arXiv preprint arXiv:2305.18565, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 453, + 295, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 453, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 62, + 453, + 295, + 518 + ], + "type": "text", + "content": "[7] Zitian Chen, Yikang Shen, Mingyu Ding, Zhenfang Chen, Hengshuang Zhao, Erik G Learned-Miller, and Chuang Gan. Mod-squad: Designing mixtures of experts as modular multi-task learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11828-11837, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 520, + 295, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 520, + 295, + 574 + ], + "spans": [ + { + "bbox": [ + 62, + 520, + 295, + 574 + ], + "type": "text", + "content": "[8] Damai Dai, Chengqi Deng, Chenggang Zhao, RX Xu, Huazuo Gao, Deli Chen, Jiashi Li, Wangding Zeng, Xingkai Yu, Y Wu, et al. Deepseekmoe: Towards ultimate expert specialization in mixture-of-experts language models. arXiv preprint arXiv:2401.06066, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 576, + 295, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 576, + 295, + 609 + ], + "spans": [ + { + "bbox": [ + 62, + 576, + 295, + 609 + ], + "type": "text", + "content": "[9] Cynthia Dwork. Differential privacy. In International colloquium on automata, languages, and programming, pages 1-12. Springer, 2006. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 611, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 611, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 57, + 611, + 295, + 633 + ], + "type": "text", + "content": "[10] Oded Goldreich. Secure multi-party computation. Manuscript. Preliminary version, 78(110):1-108, 1998. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 635, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 667 + ], + "type": "text", + "content": "[11] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. Neural computation, 3(1):79-87, 1991. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[12] Shen Jiang, Guanghui Zhu, Yue Wang, Chunfeng Yuan, and Yihua Huang. Automatic multi-task learning framework with neural architecture search in recommendations. In Proceedings of the 30th ACM SIGKDD Conference on Knowl" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "edge Discovery and Data Mining, pages 1290-1300, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 140 + ], + "type": "text", + "content": "[13] Siddharth Karamcheti, Suraj Nair, Ashwin Balakrishna, Percy Liang, Thomas Kollar, and Dorsa Sadigh. Prismatic vlms: Investigating the design space of visually-conditioned language models. arXiv preprint arXiv:2402.07865, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 141, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 553, + 217 + ], + "type": "text", + "content": "[14] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, Quan Vuong, Thomas Kollar, Benjamin Burchfiel, Russ Tedrake, Dorsa Sadigh, Sergey Levine, Percy Liang, and Chelsea Finn. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 220, + 553, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 220, + 553, + 273 + ], + "spans": [ + { + "bbox": [ + 317, + 220, + 553, + 273 + ], + "type": "text", + "content": "[15] Xinghang Li, Minghuan Liu, Hanbo Zhang, Cunjun Yu, Jie Xu, Hongtao Wu, Chilam Cheang, Ya Jing, Weinan Zhang, Huaping Liu, et al. Vision-language foundation models as effective robot imitators. arXiv preprint arXiv:2311.01378, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 276, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 276, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 317, + 276, + 553, + 319 + ], + "type": "text", + "content": "[16] Bin Lin, Zhenyu Tang, Yang Ye, Jiaxi Cui, Bin Zhu, Peng Jin, Junwu Zhang, Munan Ning, and Li Yuan. Moe-llava: Mixture of experts for large vision-language models. arXiv preprint arXiv:2401.15947, 2024.3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 321, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 321, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 317, + 321, + 553, + 354 + ], + "type": "text", + "content": "[17] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 355, + 553, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 355, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 317, + 355, + 553, + 388 + ], + "type": "text", + "content": "[18] Hao Liu, Wilson Yan, Matei Zaharia, and Pieter Abbeel. World model on million-length video and language with ringattention. arXiv e-prints, pages arXiv-2402, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 389, + 553, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 389, + 553, + 433 + ], + "spans": [ + { + "bbox": [ + 317, + 389, + 553, + 433 + ], + "type": "text", + "content": "[19] Yang Liu, Weixing Chen, Yongjie Bai, Xiaodan Liang, Guanbin Li, Wen Gao, and Liang Lin. Aligning cyber space with physical world: A comprehensive survey on embodied ai. arXiv preprint arXiv:2407.06886, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 434, + 553, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 434, + 553, + 489 + ], + "spans": [ + { + "bbox": [ + 317, + 434, + 553, + 489 + ], + "type": "text", + "content": "[20] Yuxiang Lu, Suizhi Huang, Yuwen Yang, Shalayiding Sirejiding, Yue Ding, and Hongtao Lu. Fedhca2: Towards hetero-client federated multi-task learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5599-5609, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 491, + 553, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 491, + 553, + 545 + ], + "spans": [ + { + "bbox": [ + 317, + 491, + 553, + 545 + ], + "type": "text", + "content": "[21] Ajay Mandlekar, Danfei Xu, Josiah Wong, Soroush Nasiriyani, Chen Wang, Rohun Kulkarni, Li Fei-Fei, Silvio Savarese, Yuke Zhu, and Roberto Martin-Martin. What matters in learning from offline human demonstrations for robot manipulation. arXiv preprint arXiv:2108.03298, 2021. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 547, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 547, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 317, + 547, + 553, + 600 + ], + "type": "text", + "content": "[22] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 1, 2, 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 602, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 602, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 317, + 602, + 553, + 634 + ], + "type": "text", + "content": "[23] David Nadeau and Satoshi Sekine. A survey of named entity recognition and classification. *Lingvisticae Investigationes*, 30(1):3-26, 2007. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 636, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 553, + 713 + ], + "type": "text", + "content": "[24] Abby O'Neill, Abdul Rehman, Abhiram Maddukuri, Abhishek Gupta, Abhishek Padalkar, Abraham Lee, Acorn Pooley, Agrim Gupta, Ajay Mandlekar, Ajinkya Jain, et al. Open x-embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collaboration 0. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 6892–6903. IEEE, 2024. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6912" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 550 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 137 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 137 + ], + "type": "text", + "content": "[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 194 + ], + "type": "text", + "content": "[26] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 197, + 294, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 250 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 250 + ], + "type": "text", + "content": "[27] Rejin Varghese and M Sambath. Yolov8: A novel object detection algorithm with enhanced performance and robustness. In 2024 International Conference on Advances in Data Engineering and Intelligent Computing Systems (ADICS), pages 1-6. IEEE, 2024. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 252, + 294, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 252, + 294, + 295 + ], + "spans": [ + { + "bbox": [ + 56, + 252, + 294, + 295 + ], + "type": "text", + "content": "[28] Lirui Wang, Xinlei Chen, Jialiang Zhao, and Kaiming He. Scaling proprioceptive-visual learning with heterogeneous pre-trained transformers. Advances in Neural Information Processing Systems, 37:124420-124450, 2025. 1, 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 297, + 294, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 350 + ], + "type": "text", + "content": "[29] Jiahui Xu, Lu Sun, and Dengji Zhao. Mone: Mixture-of-masked-experts for efficient multi-task recommendation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2527-2531, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 353, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 353, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 353, + 294, + 407 + ], + "type": "text", + "content": "[30] Tianhe Yu, Deirdre Quillen, Zhanpeng He, Ryan Julian, Karol Hausman, Chelsea Finn, and Sergey Levine. Metaworld: A benchmark and evaluation for multi-task and meta reinforcement learning. In Conference on robot learning, pages 1094-1100. PMLR, 2020. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 409, + 294, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 409, + 294, + 450 + ], + "spans": [ + { + "bbox": [ + 56, + 409, + 294, + 450 + ], + "type": "text", + "content": "[31] Michal Zawalski, William Chen, Karl Pertsch, Oier Mees, Chelsea Finn, and Sergey Levine. Robotic control via embodied chain-of-thought reasoning. arXiv preprint arXiv:2407.08693, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 453, + 294, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 453, + 294, + 496 + ], + "spans": [ + { + "bbox": [ + 56, + 453, + 294, + 496 + ], + "type": "text", + "content": "[32] Kaiwen Zhou and Xin Eric Wang. Fedvln: Privacy-preserving federated vision-and-language navigation. In European Conference on Computer Vision, pages 682-699. Springer, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 498, + 294, + 550 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 498, + 294, + 550 + ], + "spans": [ + { + "bbox": [ + 56, + 498, + 294, + 550 + ], + "type": "text", + "content": "[33] Weiming Zhuang, Yonggang Wen, Lingjuan Lyu, and Shuai Zhang. Mas: Towards resource-efficient federated multipletask learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23414-23424, 2023. 2" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6913" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/FedWSQ_ Efficient Federated Learning with Weight Standardization and Distribution-Aware Non-Uniform Quantization/1f166dee-20e5-452e-8e86-86019ce8d0bb_content_list.json b/2025/FedWSQ_ Efficient Federated Learning with Weight Standardization and Distribution-Aware Non-Uniform Quantization/1f166dee-20e5-452e-8e86-86019ce8d0bb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f08401884bb1c3b4ca4c0c1f39378f6ed81e4d12 --- /dev/null +++ b/2025/FedWSQ_ Efficient Federated Learning with Weight Standardization and Distribution-Aware Non-Uniform Quantization/1f166dee-20e5-452e-8e86-86019ce8d0bb_content_list.json @@ -0,0 +1,1767 @@ +[ + { + "type": "text", + "text": "FedWSQ: Efficient Federated Learning with Weight Standardization and Distribution-Aware Non-Uniform Quantization", + "text_level": 1, + "bbox": [ + 129, + 128, + 869, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Seung-Wook Kim\\*1, Seongyeol Kim\\*1, Jiah Kim\\*1, Seowon Ji\\*2, and Se-Ho Lee\\*3", + "bbox": [ + 166, + 202, + 830, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Department of Intelligent Robot Engineering, Pukyong National University", + "bbox": [ + 200, + 220, + 797, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ CSE, Konkuk University $^{3}$ CS&AI, Jeonbuk National University", + "bbox": [ + 220, + 239, + 769, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Federated learning (FL) often suffers from performance degradation due to key challenges such as data heterogeneity and communication constraints. To address these limitations, we present a novel FL framework called FedWSQ, which integrates weight standardization (WS) and the proposed distribution-aware non-uniform quantization (DANUQ). WS enhances FL performance by filtering out biased components in local updates during training, thereby improving the robustness of the model against data heterogeneity and unstable client participation. In addition, DANUQ minimizes quantization errors by leveraging the statistical properties of local model updates. As a result, FedWSQ significantly reduces communication overhead while maintaining superior model accuracy. Extensive experiments on FL benchmark datasets demonstrate that FedWSQ consistently outperforms existing FL methods across various challenging FL settings, including extreme data heterogeneity and ultra-low-bit communication scenarios. The source code is available at our project page1.", + "bbox": [ + 89, + 324, + 485, + 611 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 628, + 220, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In large-scale machine learning, centralized approaches often raise privacy concerns because sensitive data from edge devices must be collected on a central server for model training. Federated learning (FL) [31] enables distributed devices to collaboratively train a shared global model without sharing raw data. A fundamental FL method, FedAvg, was proposed by McMahan et al. [31]. In this approach, each client trains its model using the local data, and then the server collects and averages these local updates to form a unified global model.", + "bbox": [ + 89, + 652, + 482, + 804 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Several studies on FL have verified its practical effectiveness [4, 28, 41, 42, 44, 49, 50]. However, real-world FL methods often face three key challenges [18, 19, 24,", + "bbox": [ + 89, + 805, + 483, + 851 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "30, 46]: 1) data heterogeneity, where clients possess non-independent and identically distributed (non-i.i.d.) data; 2) partial client participation, where only a subset of clients contributes to a global model update during each training round due to communication constraints; 3) communication bottlenecks, where limited communication bandwidth and high transmission costs hinder efficient model aggregation. These challenges worsen local gradient divergence, slow down the global model convergence, and ultimately degrade FL performance in real-world applications.", + "bbox": [ + 511, + 292, + 903, + 444 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address these limitations, we propose FedWSQ, an advanced FL framework that combines weight standardization (WS) with our proposed distribution-aware nonuniform quantization (DANUQ) to improve both learning stability and communication efficiency. WS is a plug-and-play technique that standardizes the weight vectors of convolutional or linear layers to stabilize the learning process of a neural network [35]. In FedWSQ, WS plays a crucial role in mitigating client drift by filtering out gradient components that contribute to overfitting during local training. This leads to improved generalization across heterogeneous clients and significantly enhances FL performance. Although WS accelerates global convergence, the communication cost per round remains a major bottleneck. To address this, FedWSQ integrates WS with DANUQ, a novel quantization strategy that leverages a standard normal distribution prior to minimize quantization errors, reducing communication overhead while preserving model performance. Even under ultra-low-bit quantization, the proposed FedWSQ achieves superior performance, consistently outperforming state-of-the-art (SOTA) FL methods across various datasets and FL settings.", + "bbox": [ + 511, + 445, + 905, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our main contributions are summarized as follows.", + "bbox": [ + 529, + 779, + 870, + 792 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose an effective application of WS in FL, enhancing model convergence and stability under data heterogeneity and limited client participation. By implicitly performing a gradient filtering process, WS mitigates client drift while also providing a regularization effect that stabilizes training and improves generalization.", + "- We introduce DANUQ, a novel quantization method that" + ], + "bbox": [ + 511, + 795, + 903, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*indicates equal contribution.", + "bbox": [ + 107, + 862, + 266, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding authors.", + "bbox": [ + 109, + 875, + 241, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1https://github.com/Seongyeol-kim/FedWSQ", + "bbox": [ + 109, + 887, + 424, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4616", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "employs a fixed quantization function, leveraging the statistical properties of local model parameter updates (LMPUs) to minimize quantization errors under parameter-distribution priors. By efficiently compressing LM-PUs while preserving essential information, the proposed DANUQ significantly reduces communication overhead without noticeable performance degradation.", + "bbox": [ + 102, + 90, + 482, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Extensive experiments show that FedWSQ significantly enhances performance across diverse settings. Even with an average of 2.3 bits per parameter, FedWSQ improves performance by over $5\\%$ on Tiny-ImageNet in a highly heterogeneous setting compared to SOTA FL methods.", + "bbox": [ + 88, + 196, + 480, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work and Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 284, + 385, + 299 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Federated learning (FL) FL is a decentralized approach that enables collaborative training of a global model. Using distributed devices, FL ensures efficient learning while preserving data privacy. Let $\\mathcal{F}_n(\\mathbf{w}_{\\mathrm{g}})\\coloneqq \\mathbb{E}_{\\zeta_n\\sim \\mathcal{D}_n}[\\mathcal{F}_n(\\mathbf{w}_{\\mathrm{g}};\\zeta_n)]$ be the local loss function of the client $n\\in [N]$ with a local data distribution $\\mathcal{D}_n$ , where $\\mathbf{w}_{\\mathrm{g}}$ is the global model parameter (GMP) vector. The training objective function of FL [31] is formulated as follows:", + "bbox": [ + 89, + 309, + 483, + 428 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathbf {w} _ {\\mathrm {g}}} \\left\\{\\mathcal {F} \\left(\\mathbf {w} _ {\\mathrm {g}}\\right) := \\sum_ {i \\in \\mathcal {S}} h _ {i} \\mathcal {F} _ {i} \\left(\\mathbf {w} _ {\\mathrm {g}}\\right) \\right\\}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 430, + 480, + 462 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $h_i$ is the weight assigned to the $i$ -th device, such that $h_i \\geq 0$ , $\\sum_{i} h_i = 1$ , and $h_i \\propto |\\mathcal{D}_i|$ , and $S \\subseteq [N]$ is a subset of participating clients in the current communication round. At each round $t \\in [T]$ , the server broadcasts the GMPs $\\mathbf{w}_{\\mathrm{g}}^{t-1}$ to the participating client $i$ . In FedAvg [31], the $i$ -th client performs $K$ steps of local training to obtain the local model parameter (LMP) vector, $\\mathbf{w}_i^t$ , and transmits its LMP updates (LMPUs), $\\Delta \\mathbf{w}_i^t := \\mathbf{w}_i^t - \\mathbf{w}_{\\mathrm{g}}^{t-1}$ , to the server. Finally, the server aggregates these LMPUs through weighted averaging to derive the updated GMPs.", + "bbox": [ + 89, + 463, + 482, + 613 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In real-world FL scenarios, unstable learning dynamics and slow convergence often occur due to non-i.i.d. data distributions and limited communication bandwidth. To address these issues, FedProx [27] penalizes LMPUs by adding a proximal term in the local objective to reduce the gap between the global and local loss functions. FedAvgM [13] integrates momentum into the global update process to mitigate variance in model aggregation. FedAdam [36] introduces an adaptive optimizer that improves global and local model convergence, thereby accelerating overall training progress. FedDyn [1] presents dynamic regularization for each device to diminish deviations in the LMPUs caused by data heterogeneity. FedSmoo [43] applies constraints that prevent sharp fluctuations in local updates to generalize FL performance by ensuring global consistency. Recently, FedACG [21] was proposed to improve convergence by broadcasting a global model with a lookahead gradient and aligning local models with the shared global model.", + "bbox": [ + 89, + 614, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Network quantization As neural network architectures have grown larger in scale, quantization has become essential for the memory optimization and acceleration of operations [9, 10, 16, 17, 29, 33, 55]. Let us consider a model using $B$ -bit representation. A typical quantization function $\\mathcal{Q}$ that maps a full-precision value $x$ to a quantization level (QL) $q$ can be written as", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nq = \\mathcal {Q} (x / s), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 663, + 205, + 903, + 222 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $s$ is a scaling factor, and each QL is matched with a $B$ -bit integer from $\\{0,\\dots ,2^B -1\\}$ . The primary goal of effective quantization is to determine the proper scaling factor and a set of QLs to minimize the quantization errors.", + "bbox": [ + 511, + 229, + 903, + 291 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Quantization methods can be roughly categorized into two approaches: uniform quantization (UQ) and nonuniform quantization (NUQ). UQ divides the input range into $2^{B}$ equal intervals and assigns a representative value as a QL to each interval [10, 29]. In contrast, NUQ adjusts the interval widths based on the prior data distribution [6, 8, 51]. By allocating finer intervals in dense data regions and broader intervals in sparse regions, NUQ preserves critical information more effectively than UQ. For the scaling factor $s$ , both approaches commonly employ the absmax method [16, 33] to determine the dynamic range for quantization. Furthermore, to enhance data adaptability, researchers have explored methods for dynamically learning quantization parameters (e.g., scaling factors and QLs) [9, 17, 55]. During backpropagation, these parameters are updated to reduce quantization error and better align with the underlying data distribution.", + "bbox": [ + 511, + 292, + 906, + 550 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Quantized FL In FL, communication bottlenecks arise primarily during uplink transmission, as each participating client must send its individual LMPUs to the central server. Unlike downlink transmission, where a single GMP is broadcast to multiple clients, uplink communication involves substantial data transfer. This problem is exacerbated in resource-constrained communication environments, such as mobile and IoT networks, where upload bandwidth is severely limited [5, 37]. To overcome this challenge, several studies have focused on compressing model updates. In [22], quantization and sparsification are combined to reduce the number of transmitted parameters. FedPAQ [37] employs a probabilistic rounding algorithm to prevent an excessive concentration of identical QLs. Meanwhile, FedHQ+ [5] assigns different weight factors to clients for global updates based on their individual quantization errors. Although these methods significantly improve communication efficiency, they often suffer from performance degradation under extremely low bit-width conditions. In this paper, we introduce FedWSQ, a novel quantization framework that is designed to maintain robust model performance even at ultra-low bit-width.", + "bbox": [ + 511, + 568, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4617", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4e12e521430890cc82b9f7d38086e2fc6a0e6165afd8de62ccb3dd0637e6a8c1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 181, + 90, + 815, + 242 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4b655be1c8fec3b0c32d6a8e3deed4f5a7bf375616742c6b55241bbf9874effd.jpg", + "image_caption": [ + "Figure 1. The overall framework of FedWSQ, which integrates WS and DANUQ for efficient FL. Clients transmit LMPUs quantized by the DANUQ along with scaling vectors to the server, where dequantization and aggregation are performed to update the global parameters." + ], + "image_footnote": [], + "bbox": [ + 178, + 244, + 815, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 89, + 381, + 264, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 1 illustrates the overall framework of the proposed FedWSQ, which employs WS and DANUQ for efficient FL. Each client has a standard neural network model consisting of multiple model blocks, where WS is applied to mitigate the impact of data heterogeneity between clients. After local training, the full-precision values of LMPUs are converted to a $B$ -bit representation using the global scaling vectors and carefully designed QLs. These quantized LMPUs and their corresponding scaling vectors are transmitted to the server. The server then dequantizes the received data to higher-precision values and aggregates them to update the GMPs. In addition, the global scaling vectors are refined to ensure consistent scaling across different clients for DANUQ. Using the proposed DANUQ, FedWSQ enhances communication efficiency without sacrificing model performance in challenging FL scenarios.", + "bbox": [ + 89, + 407, + 483, + 650 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Weight standardization (WS)", + "text_level": 1, + "bbox": [ + 89, + 662, + 352, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "WS [35] standardizes each weight vector of a convolution or linear layer to enhance the learning process. By ensuring consistent parameter distributions during training, WS contributes to a more stable gradient flow. Beyond this general benefit, we argue that the use of WS in FL can address the challenges posed by the distributed nature of FL and data heterogeneity. Formally, consider the $n$ -th client that has a network architecture consisting of $L$ linear or convolutional layers. Let the $l$ -th layer (without bias) map an input $\\mathbf{x}_l \\in \\mathbb{R}^{I_l}$ to an output $\\mathbf{y}_l \\in \\mathbb{R}^{O_l}$ , defined by", + "bbox": [ + 89, + 686, + 483, + 838 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {y} _ {l} = \\mathbf {W} _ {n, l} ^ {T} \\mathbf {x} _ {l} \\tag {3} \\\\ := \\left[ \\begin{array}{c c c c} \\mathbf {w} _ {n, m _ {1} ^ {l}} & \\mathbf {w} _ {n, m _ {2} ^ {l}} & \\dots & \\mathbf {w} _ {n, m _ {O _ {l}} ^ {l}} \\end{array} \\right] ^ {T} \\mathbf {x} _ {l}, \\qquad (4) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 848, + 482, + 898 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{w}_{n,m}$ , for $m\\in \\mathcal{M}_l\\coloneqq \\{m_1^l,\\ldots ,m_{O_l}^l\\}$ , is the weight vector. In WS, each $\\mathbf{w}_{n,m}$ is standardized before being used within the layer. We denote the prestandardized parameter (PSP) vector as $\\mathbf{w}_{n,m}$ , and the weight-standardized parameter (WSP) vector as $\\tilde{\\mathbf{w}}_{n,m}$ , respectively. Then, the WSP vector $\\tilde{\\mathbf{w}}_{n,m}$ is obtained by", + "bbox": [ + 511, + 382, + 906, + 474 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {w}} _ {n, m} = \\frac {\\rho}{\\sigma (\\mathbf {w} _ {n , m})} \\left(\\mathbf {I} - \\mathbf {P _ {1}}\\right) \\mathbf {w} _ {n, m}, \\qquad (5)\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 479, + 906, + 512 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{1} \\in \\mathbb{R}^{I_l}$ is a vector of all ones, $\\sigma(\\mathbf{w}_{n,m})$ is the standard deviation of the elements in $\\mathbf{w}_{n,m}$ , and $\\mathbf{P_v}$ represents the projection matrix onto the subspace spanned by the vector $\\mathbf{v}$ , which is denoted by $\\mathrm{span}\\{\\mathbf{v}\\}$ . Here, we introduce a hyper-parameter $\\rho$ to control the normalization scale. Note that we reformulate the WSP computation from [35] into Eq. (5) to better understand the gradient filtering process of WS during local training.", + "bbox": [ + 511, + 518, + 906, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Gradient filtering by WS WS plays an important role in reducing the learning diversity of local models, which is one of the most critical issues in FL. To understand its impact, we compare the gradient of a loss function $\\mathcal{L}$ with respect to the WSP vector $\\tilde{\\mathbf{w}}_{n,m}$ to that of the PSP vector $\\mathbf{w}_{n,m}$ . When WS is applied, the gradient is processed by an additional series of projections to remove specific components that may interfere with effective FL. Specifically, the gradient with respect to the PSP vector $\\mathbf{w}_{n,m}$ is obtained by", + "bbox": [ + 511, + 651, + 908, + 789 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial \\mathcal {L}}{\\partial \\mathbf {w} _ {n , m}} = \\frac {\\rho}{\\sigma (\\mathbf {w} _ {n , m})} \\left(\\mathbf {I} - \\mathbf {P} _ {\\mathbf {1}}\\right) \\left(\\mathbf {I} - \\mathbf {P} _ {\\tilde {\\mathbf {w}} _ {n, m}}\\right) \\frac {\\partial \\mathcal {L}}{\\partial \\tilde {\\mathbf {w}} _ {n , m}}. (6)\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 796, + 906, + 830 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "2Strictly speaking, convolution involves spatial structures aggregating information from local receptive fields. For simplicity, we omit spatial considerations and treat the convolution as a linear transformation. This abstraction can be easily generalized to multidimensional tensor structures.", + "bbox": [ + 511, + 838, + 906, + 887 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "3Section A of the supplementary document provides its derivation.", + "bbox": [ + 532, + 887, + 883, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4618", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c5d4cedad4a30261ec4674529f88523fee3a26460aa03e52dbe3e1a257b88a7b.jpg", + "image_caption": [ + "Figure 2. An example of gradient filtering process of WS." + ], + "image_footnote": [], + "bbox": [ + 98, + 90, + 477, + 213 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The upstream gradient $\\frac{\\partial\\mathcal{L}}{\\partial\\tilde{\\mathbf{w}}_{n,m}}$ is projected onto two subspaces, $\\mathrm{span}\\{\\tilde{\\mathbf{w}}_{n,m}\\}^{\\perp}$ and $\\mathrm{span}\\{\\mathbf{1}\\}^{\\perp}$ , where $\\mathrm{span}\\{\\mathbf{v}\\}^{\\perp}$ represents the orthogonal complement of $\\mathrm{span}\\{\\mathbf{v}\\}$ . Since $\\tilde{\\mathbf{w}}_{n,m} \\perp \\mathbf{1}$ , the upstream gradient $\\frac{\\partial\\mathcal{L}}{\\partial\\tilde{\\mathbf{w}}_{n,m}}$ is ultimately projected onto $\\mathrm{span}\\{\\tilde{\\mathbf{w}}_{n,m}, \\mathbf{1}\\}^{\\perp}$ after the two projections.", + "bbox": [ + 89, + 252, + 482, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As illustrated in Figure 2, Eq. (6) represents a gradient filtering process that applies successive projections to remove undesirable components from the gradient with respect to LMPs. In FL, data heterogeneity causes local models to overfit their local data, which leads to inconsistent updates across clients. This inconsistency results in a discrepancy between the GMPs and the optimal parameters, known as client drift. WS mitigates this issue by projecting the gradient onto $\\mathrm{span}\\{\\tilde{\\mathbf{w}}_{n,m},\\mathbf{1}\\}^{\\perp}$ , which effectively reduces the impact of client drift during training. The first projection eliminates the gradient component aligned with $\\tilde{\\mathbf{w}}_{n,m}$ , which corresponds to the LMPs that are biased due to the non-i.i.d. data distribution. The second projection removes the mean component of mini-batch gradients, which can also be biased toward the local data distribution, compounding inconsistencies across clients. By filtering out both the parameter-aligned and mean components from the local gradient, WS preserves meaningful gradient directions essential for stable global model convergence. Additionally, this projection-based filtering introduces a regularization effect by constraining the model's deviation from its current state, which can improve the model's generalization ability.", + "bbox": [ + 91, + 337, + 483, + 670 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Comparison with FedWon It is worth noting that FedWon [56], a recently proposed FL method, also employs WS to address data heterogeneity in multi-domain FL. However, a key difference lies in the parameters exchanged between clients and the server, which fundamentally alters the learning dynamics. FedWon transmits WSP, which explicitly normalizes parameters across clients to improve learning stability. However, this approach forces all clients' LMPUs to exhibit statistically similar characteristics, which potentially discards client-specific information crucial for local adaptation. In contrast, FedWSQ transmits PSP, preserving essential local information while implicitly mitigating harmful divergences through the gradient filtering process of WS, as shown in Figure 2. This design choice en", + "bbox": [ + 89, + 688, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ables FedWSQ to achieve a better balance between global stability and local information preservation. An experimental comparison between FedWon and FedWSQ will be discussed in Section 4.4.", + "bbox": [ + 511, + 90, + 905, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Quantization of LMPUs", + "text_level": 1, + "bbox": [ + 511, + 160, + 736, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Neural network quantization often involves learning quantization parameters. However, in FL, transmitting additional quantization parameters or auxiliary information for each communication round would introduce significant resource overhead. Instead, we propose a fixed quantization function that eliminates the need to transmit additional quantization information, ensuring a lightweight and efficient communication process. A fundamental assumption in NUQ is that model parameters follow a normal distribution, as demonstrated in prior studies [7, 45, 54]. Since LMPUs represent the difference between an updated LMP and the previous GMP, they may also exhibit a normal distribution.", + "bbox": [ + 511, + 181, + 906, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LMPU scaling Most quantization methods adopt the absmax strategy to manage the dynamic range of values [16, 33]. This approach scales an input tensor to the range $[-1,1]$ by dividing each element by the maximum absolute value of the tensor. Within this constrained range, conventional quantization methods determine QLs and map input values to the nearest quantization point. However, these methods are highly sensitive to outliers, which can excessively expand the dynamic range of the target tensor. Such over-scaling often results in underflow, especially in extremely low-bit representation. To solve this problem, we use the standard deviation of the input tensor as a scaling factor. Unlike the absmax value, the standard deviation is more robust to outliers. Moreover, since our DANUQ is based on a probability model of a standard normal distribution, the standard deviation provides a more reliable approximation of the target dynamic range.", + "bbox": [ + 511, + 375, + 908, + 631 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In FedWSQ, we use a shared global scaling factor to help maintain consistency in quantization. Instead of independently using scaling factors, we employ a collaborative approach in which each client contributes to a global scaling vector, ensuring that local standard deviations are taken into account. Specifically, each client $i$ computes a scaling vector, defined as $\\mathbf{s}_i = [s_{i,1},\\dots ,s_{i,L}]^T$ , where $s_{i,l}$ denotes the standard deviation of the LMPUs for the $l$ -th layer. The server then collects and aggregates $\\mathbf{s}_i$ from the participating clients $i\\in S$ , and updates the global scaling vector $\\mathbf{s_g} = [s_{g,1},\\dots ,s_{g,L}]^T$ as follows:", + "bbox": [ + 511, + 631, + 908, + 800 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {s} _ {\\mathrm {g}} \\leftarrow (1 - \\beta) \\mathbf {s} _ {\\mathrm {g}} + \\beta \\frac {1}{| \\mathcal {S} |} \\sum_ {i \\in \\mathcal {S}} \\mathbf {s} _ {i}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 808, + 906, + 844 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\beta$ is a momentum parameter that controls the update rate. This approach ensures that the global scaling factor smoothly adapts while mitigating local fluctuations. Once", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4619", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "updated, the server distributes $\\mathbf{s}_{\\mathrm{g}}$ to the participating clients. Each client then divides its LMPU updates for the $l$ -th layer by the corresponding scale $s_{\\mathrm{g},l}$ , so that the normalized values can be regarded as samples from a standard normal distribution. Based on this assumption, we apply the proposed DANUQ scheme detailed in the next section.", + "bbox": [ + 89, + 90, + 483, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "QLs considering normal distribution To determine the optimal multi-bit QLs, we formulate an optimization problem that minimizes the expected quantization error under a normal distribution. We assume that a random variable $\\Delta w$ from the normalized LMPUs follows a standard normal distribution. Since the distribution is symmetric around zero, we consider only its nonnegative half. Given a $B$ -bit representation, let $Q = \\{q_{0}, q_{1}, \\ldots, q_{R}\\}$ be the set of QLs, where $R = 2^{B - 1} - 1$ . Here, we assume that $q_{r}$ values are sorted in ascending order and set $q_{0} = 0$ as a fixed value. Then, the quantization boundaries are defined as", + "bbox": [ + 89, + 188, + 483, + 354 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nu _ {r} = \\left\\{ \\begin{array}{l l} q _ {0}, & \\text {i f} r = 0, \\\\ \\frac {q _ {r - 1} + q _ {r}}{2}, & \\text {i f} 1 \\leq r \\leq R, \\\\ + \\infty , & \\text {i f} r = R + 1. \\end{array} \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 362, + 482, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A full-precision value in the range $[u_r, u_{r+1})$ is quantized into the QL $q_r$ . The optimal QLs are determined by minimizing the expected quantization error, formulated as", + "bbox": [ + 89, + 426, + 483, + 472 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} \\left[ \\left(\\Delta w - \\Delta \\bar {w}\\right) ^ {2} \\right] = \\sum_ {r = 0} ^ {R} \\int_ {u _ {r}} ^ {u _ {r + 1}} \\left(x - q _ {r}\\right) ^ {2} p (x) d x, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 479, + 482, + 520 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Delta \\bar{w} \\in Q$ represents the quantized value, and $p(x)$ is the PDF of the standard normal distribution. From (9), we can derive the following optimization problem as", + "bbox": [ + 89, + 527, + 483, + 574 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {\\left\\{q _ {1}, \\dots , q _ {R} \\right\\}} \\left\\{\\frac {1}{2} \\left(q _ {R} ^ {2} + 1\\right) - \\sqrt {\\frac {2}{\\pi}} q _ {1} e ^ {- \\frac {u _ {1} ^ {2}}{2}} - \\frac {1}{2} q _ {1} ^ {2} \\operatorname {e r f} \\left(\\frac {u _ {1}}{\\sqrt {2}}\\right) \\right. \\\\ + \\sqrt {\\frac {2}{\\pi}} \\sum_ {r = 1} ^ {R - 1} (q _ {r} - q _ {r + 1}) e ^ {- \\frac {u _ {r + 1} ^ {2}}{2}} \\\\ + \\frac {1}{2} \\sum_ {r = 1} ^ {R - 1} \\left(q _ {r} ^ {2} - q _ {r + 1} ^ {2}\\right) \\operatorname {e r f} \\left(\\frac {u _ {r + 1}}{\\sqrt {2}}\\right) \\}, \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 580, + 480, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\operatorname{erf}(\\cdot)$ denotes the error function. A closed-form solution to Eq. (22) is intractable due to the presence of nonlinear terms, including the Gaussian integral and the error function, which lack simple analytical inverses, and the interdependent QLs due to boundary conditions. As an alternative, we employ a brute-force search algorithm to numerically determine the optimal QLs. The optimal QLs", + "bbox": [ + 89, + 713, + 483, + 820 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 FedWSQ", + "text_level": 1, + "bbox": [ + 514, + 90, + 668, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input: Initial global parameters $\\mathbf{W}_{\\mathrm{g}}^{0}$ , initial scales $\\mathbf{s}_{\\mathrm{g}}^{0}$ , global communication round $T$ , local training iteration $K$ , number of total clients $N$ , number of model layers $L$ , learning rate $\\eta$ , hyperparameter $\\beta$", + "bbox": [ + 514, + 108, + 906, + 146 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Server-side:", + "text_level": 1, + "bbox": [ + 514, + 146, + 584, + 156 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for $t = 1,\\dots ,T$ do", + "bbox": [ + 514, + 157, + 625, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Randomly sample a subset of clients $S_{t} \\subseteq [N]$", + "bbox": [ + 540, + 167, + 789, + 180 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Set the bit-precision $B_{i}$ for the client $i \\in S_{t}$", + "bbox": [ + 540, + 181, + 774, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Send $(\\mathbf{W}_{\\mathrm{g}}^{t - 1},\\mathbf{s}_{\\mathrm{g}}^{t - 1})$ to client $i\\in S_{t}$", + "bbox": [ + 540, + 191, + 738, + 204 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for each client $i\\in S_t$ , in parallel do", + "bbox": [ + 540, + 205, + 745, + 215 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\Delta \\overline {{\\mathbf {W}}} _ {i} ^ {t}, \\mathbf {s} _ {i} ^ {t}\\right) \\leftarrow \\text {C l i e n t L o c a l T r a i n i n g} \\left(\\mathbf {W} _ {\\mathrm {g}} ^ {t - 1}, \\mathbf {s} _ {\\mathrm {g}} ^ {t - 1}, i, B _ {i}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 215, + 879, + 232 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "end", + "bbox": [ + 540, + 232, + 563, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for $i\\in S_t$ do", + "bbox": [ + 540, + 243, + 614, + 253 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Obtain dequantized parameters $\\Delta_i^t$ using $(\\Delta \\overline{\\mathbf{W}}_i^t, \\mathbf{s}_i^t)$", + "bbox": [ + 547, + 255, + 846, + 270 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "end", + "bbox": [ + 540, + 270, + 563, + 279 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Aggregate local updates: $\\Delta^t\\gets \\sum_{i\\in S_t}\\Delta_i^t$", + "bbox": [ + 540, + 280, + 774, + 294 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Update global parameters: $\\mathbf{W}_{\\mathrm{g}}^{t}\\gets \\mathbf{W}_{\\mathrm{g}}^{t - 1} + \\Delta^{t}$", + "bbox": [ + 540, + 295, + 799, + 309 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Update global scales: $\\mathbf{s}_{\\mathrm{g}}^{t}\\gets (1 - \\beta)\\mathbf{s}_{\\mathrm{g}}^{t - 1} + \\beta \\frac{1}{|\\mathcal{S}_{t}|}\\sum_{i\\in \\mathcal{S}_{t}}\\mathbf{s}_{i}^{t}$", + "bbox": [ + 540, + 310, + 867, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "end", + "bbox": [ + 514, + 325, + 537, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ClientLocalTraining $(\\mathbf{W}_{\\mathrm{g}},\\mathbf{s}_{\\mathrm{g}},i,I,B)$", + "bbox": [ + 514, + 349, + 712, + 362 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Initialize local parameters: $\\mathbf{W}_j^0\\gets \\mathbf{W}_{\\mathrm{g}}$", + "bbox": [ + 514, + 363, + 728, + 373 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Set DANUQ function $\\mathcal{Q}(\\cdot)$ based on the bit-precision $B$", + "bbox": [ + 514, + 373, + 812, + 385 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for $k = 1,\\ldots ,K$ do", + "bbox": [ + 514, + 386, + 632, + 397 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Apply WS to $\\mathbf{W}_i^{k - 1}$ using (5)", + "bbox": [ + 540, + 397, + 704, + 411 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compute unbiased gradient of local loss $\\nabla f_{i}(\\mathbf{W}_{i}^{k - 1})$", + "bbox": [ + 540, + 412, + 831, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Update local parameters: $\\mathbf{W}_i^k\\gets \\mathbf{W}_i^{k - 1} - \\eta \\nabla f_i(\\mathbf{W}_i^{k - 1})$", + "bbox": [ + 540, + 426, + 859, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "end", + "bbox": [ + 514, + 441, + 537, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compute local update: $\\Delta \\mathbf{W}_i\\gets \\mathbf{W}_i^K -\\mathbf{W}_g$", + "bbox": [ + 514, + 453, + 761, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for each layer $l = 1,\\dots ,L$ do", + "bbox": [ + 514, + 467, + 689, + 477 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compute quantized local update: $\\Delta \\overline{\\mathbf{W}}_{i,l}\\gets \\mathcal{Q}(\\Delta \\mathbf{W}_{i,l} / s_{\\mathrm{g},l})$", + "bbox": [ + 540, + 478, + 870, + 492 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Calculate local scale factors: $s_{i,l} \\gets \\mathrm{std}(\\Delta \\mathbf{W}_{i,l})$", + "bbox": [ + 540, + 492, + 803, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "end", + "bbox": [ + 514, + 505, + 537, + 513 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\overline {{\\mathbf {W}}} _ {i} := (\\Delta \\mathbf {W} _ {i, 1}, \\dots , \\Delta \\mathbf {W} _ {i, L})\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 516, + 702, + 530 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {s} _ {i} := \\left[ s _ {i, 1}, \\dots , s _ {i, L} \\right]\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 530, + 635, + 542 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {r e t u r n} \\left(\\Delta \\overline {{\\mathbf {W}}} _ {i}, \\mathbf {s} _ {i}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 542, + 617, + 555 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for different bit-widths are obtained as follows: 1) 1-bit $^6$ : [-0.798, 0.798]; 2) 2-bit: [-1.224, 0, 0.765, 1.724]; 3) 4-bit: [-2.654, -1.974, -1.508, -1.149, -0.834, -0.544, -0.269, 0, 0.230, 0.465, 0.708, 0.966, 1.248, 1.568, 1.968, 2.649]. These QLs minimize quantization error under the normal distribution, which provides an effective trade-off between precision and computational efficiency.", + "bbox": [ + 511, + 583, + 906, + 689 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Mixed channel-bit allocation To further enhance communication efficiency in FL, we propose an adaptive mixed-precision strategy in which each client dynamically selects its bit representation based on its communication channel conditions. This approach effectively balances model update quality and transmission cost, reducing overall bandwidth utilization. To validate this strategy, we conduct two simulation setups: 1) Fixed-bit allocation (FBA), where each client keeps a constant bit-width selected from $\\{1,2,4\\}$ throughout training, and 2) dynamic-bit allocation (DBA), where each client is assigned a randomly", + "bbox": [ + 511, + 700, + 906, + 867 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "4The full derivation is provided in the supplementary document.", + "bbox": [ + 107, + 825, + 447, + 839 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "5We discretize the search space of possible QLs and perform an exhaustive evaluation of the cost function to identify the configuration that minimizes the quantization error. To improve computational efficiency, we restrict the search space to a reasonable range based on empirical observations and employ parallel processing to accelerate evaluation.", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "For the 1-bit case, the QLs consist of only two values. Thus, the constraint, $q_{0} = 0$ , is omitted to allow optimal placement of both QLs.", + "bbox": [ + 511, + 875, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4620", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/9e39e66dfee90a50f8aba2cc88e432892b55b588823fc560e9718aa3d3ad19b0.jpg", + "table_caption": [ + "Table 1. FL performance on the three benchmark datasets with a $5\\%$ participation rate over 100 clients for $\\alpha \\in \\{0.1, 0.3, 0.6\\}$ ." + ], + "table_footnote": [], + "table_body": "
Method#bitsCIFAR-10CIFAR-100Tiny-ImageNet
α = 0.10.30.6i.i.d.0.10.30.6i.i.d.0.10.30.6i.i.d.
FedAvg [31]3270.6582.5285.4589.2543.4547.3949.0448.6630.3734.6536.5437.39
FedProx [27]3270.8983.9186.6088.8143.7848.2648.4948.5429.9935.4437.1437.43
FedAvgM [13]3278.3085.2787.9490.0047.7052.0453.0053.4032.0037.6339.6641.02
FedADAM [36]3272.5881.7385.9788.4546.8353.4355.9657.9334.7739.9743.2746.26
FedDyn [1]3283.6388.2889.2091.0853.8356.4056.9056.9337.3240.9242.7842.65
FedMLB [21]3271.1282.4086.2389.2549.7254.1655.1756.4533.4539.0640.5541.46
FedLC [52]3274.0383.8086.2088.3543.8648.5048.6147.8131.7334.5536.7238.06
FedNTD [26]3273.2983.3486.2689.3645.0848.7350.3950.9034.3636.4439.0640.60
FedSmoo [43]3278.6180.0785.0386.4842.0244.9745.3249.4323.0927.1429.7034.96
FedDecorr [39]3275.4483.0184.4988.4044.0549.3550.4449.4031.1234.1335.5936.36
FedWon [56]3270.8978.3682.4486.5645.5551.6853.6055.2830.5236.2038.5940.73
FedRCL [38]3283.4588.1989.7491.5558.2662.2163.9565.1137.8644.9446.7547.96
FedACG [20]3283.6289.1290.3891.7558.1462.8061.8562.4639.7545.4648.2250.49
FedPAQ [37]466.8178.5281.3685.4638.0642.4243.4644.3528.5431.9333.5634.98
157.8271.2772.3679.4630.0734.1638.2437.8524.7429.4830.7433.81
FedHQ+ [5]466.8978.4281.4285.5438.2242.6643.1044.8628.5932.1834.2435.69
158.1470.6873.1679.0431.1134.1137.5039.1724.7929.1831.1433.50
FedWSQFedWS (32)85.9489.8191.0092.4764.1468.1969.3170.1447.0552.8454.2355.26
486.8489.8491.1192.2164.3467.3168.4168.5346.5150.4951.9652.13
184.7988.6289.4991.1562.0565.1466.1966.1145.1149.9351.1151.77
FBA (2.33)86.1689.6090.4491.7163.0766.1667.0867.1746.6250.6951.4351.68
DBA (2.33)85.2089.2390.2691.7863.2866.3267.2667.2246.5050.2451.5551.80
", + "bbox": [ + 119, + 107, + 875, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "selected bit-width at every communication round. With bit-width selection following a uniform distribution over $\\{1,2,4\\}$ , the expected bit-width per client is approximately 2.3 bits. The entire FedWSQ framework, including the mixed-precision strategy, is summarized in Algorithm 1.", + "bbox": [ + 89, + 411, + 480, + 488 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 505, + 223, + 522 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we evaluate the proposed FedWSQ on standard FL benchmarks, compare its performance against various FL methods on both i.i.d. and non-i.i.d. data conditions, and offer a further analysis of FedWSQ. The convergence plots obtained from the experiments are provided in Section D of the supplementary document.", + "bbox": [ + 89, + 531, + 483, + 623 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental setup", + "text_level": 1, + "bbox": [ + 89, + 635, + 277, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For experiments, we use three standard benchmark datasets: CIFAR-10 [23], CIFAR-100 [23], and Tiny-ImageNet [25]. CIFAR-10 and CIFAR-100 contain 60,000 images divided into 10 and 100 classes, respectively. Tiny-ImageNet is composed of 200 classes, providing a more complex image classification task. For i.i.d. settings, training samples are randomly assigned to clients without replacement. For the non-i.i.d. setting, we sample label ratios from a Dirichlet distribution with a concentration parameter $\\alpha \\in \\{0.1, 0.3, 0.6\\}$ , following Hsu et al. [13]. Lower values of $\\alpha$ (e.g., 0.1) indicate higher data heterogeneity, whereas larger values (e.g., 0.6) and the i.i.d. case correspond to more homogeneous distributions. The participation rate is set to $5\\%$ among 100 clients. Model accuracy is evaluated on the test set of each dataset at the 1,000th communication round. To ensure stable evaluations, we report accuracy", + "bbox": [ + 89, + 657, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "using an exponential moving average with a smoothing parameter of 0.9. To validate the effectiveness of FedWSQ, we integrate it with FedAvg [31]. Note that FedWSQ can be incorporated into any existing FL algorithm. For clarity, we refer to the proposed method without DANUQ as FedWS. For quantized FL settings, we consider three bit-allocation strategies: constant allocation, FBA, and DBA. In the constant allocation settings, a fixed bit-width is assigned to all clients. We compare FedWSQ against various FL algorithms, including FedAvg [31], FedProx [27], FedAvgM [13], FedADAM [36], FedDyn [1], FedMLB [21], FedLC [52], FedNTD [26], FedSmoo [43], FedDecorr [39], FedWon [56], FedRCL [38], FedACG [20], FedPAQ [37], and FedHQ+ [5].", + "bbox": [ + 511, + 412, + 906, + 623 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation details", + "text_level": 1, + "bbox": [ + 511, + 633, + 725, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We follow most of the implementation setups and evaluation protocols from [1, 20, 38, 48]. The ResNet-18 architecture [11] is adopted as the backbone network. Consistent with standard practice in FL [12], all BN [15] layers in ResNet-18 are replaced with GN [47] layers. Following the recommendation of [35], WS is applied before each GN layer to achieve optimal performance. The hyper-parameter of WS is set to $\\rho = 0.001$ , and the momentum for updating the scaling vector is set to $\\beta = 0.1$ . Additional information on the hyper-parameter selection is provided in Section C of the supplementary document.", + "bbox": [ + 511, + 656, + 906, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparison with existing FL methods", + "text_level": 1, + "bbox": [ + 511, + 832, + 844, + 849 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 presents the performance comparison of the proposed FedWSQ against various existing FL methods on the three benchmark datasets. The bit-width settings for each", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4621", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/757f1f071868a411527ade7a472deca08325914e7ff84e22940205653c3c80e8.jpg", + "table_caption": [ + "Table 2. Comparison of UQ and DANUQ methods with or without WS on the three benchmark datasets with a $5\\%$ participation rate over 100 clients for $\\alpha \\in \\{0.1, 0.3, 0.6\\}$ ." + ], + "table_footnote": [], + "table_body": "
DatasetWS#bitsα = 0.1α = 0.3α = 0.6i.i.d.
UQDANUQUQDANUQUQDANUQUQDANUQ
CIFAR-10X157.8272.4871.2780.8872.3683.5279.4686.92
262.3965.1974.3578.6176.7982.5982.2586.11
466.8172.8778.5284.6281.3686.2685.4688.66
✓175.4884.7980.1588.6282.5389.4985.3791.15
278.3685.9283.0288.9884.1090.3586.9591.50
482.7886.8486.3789.8487.5891.1189.8792.21
CIFAR-100X130.0740.5134.1646.6238.2448.6737.8548.55
233.5240.5837.6645.0740.4746.3140.9046.06
438.0643.5942.4248.3343.4649.2944.3548.76
✓145.5962.0550.1065.1452.1366.1954.5866.11
249.6863.3454.2265.9656.0167.1958.7267.82
456.8564.3460.2167.3161.9868.4162.9868.53
Tiny-ImageNetX124.7428.9529.4833.1430.7436.7533.8138.95
227.0230.3331.6034.1233.5836.634.8735.41
428.5431.1431.9335.4233.5636.6634.9838.13
✓132.7345.1137.6349.9339.6051.1141.9851.77
236.1645.9640.9050.1742.5151.1244.8951.12
440.4646.5144.5850.4945.5551.9646.9352.13
", + "bbox": [ + 212, + 121, + 784, + 368 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e85414b285a32afc06968838588186767798b1323243a3051a62e4ac3b3a770e.jpg", + "table_caption": [ + "Table 3. Comparison of NUQ methods using WS on CIFAR-10 and CIFAR-100 datasets for $\\alpha \\in \\{0.1, 0.3, 0.6\\}$ ." + ], + "table_footnote": [], + "table_body": "
#bitsMethodCIFAR-10CIFAR-100
α = 0.10.30.60.10.30.6
1NF/FP24.030.628.87.08.38.1
DANUQ84.888.689.562.165.166.2
2NF57.873.880.041.249.153.0
FP45.057.558.625.332.034.2
DANUQ85.989.090.463.366.067.2
4NF86.689.791.064.866.768.4
FP87.190.091.064.367.068.4
DANUQ86.889.891.164.367.368.4
", + "bbox": [ + 117, + 410, + 455, + 539 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "FL method are indicated in the table. Compared to FL methods that use quantization, such as FedPAQ and FedHQ+, FedWSQ demonstrates superior performance for 4-bit and 1-bit settings. Also, FedWSQ outperforms FL methods with full-precision representation, including SOTA methods such as FedACG and FedRCL. Notably, even with a 1-bit representation, FedWSQ demonstrates competitive performance on CIFAR-10 and achieves superior results on CIFAR-100 and Tiny/ImageNet. For example, in a highly heterogeneous setting $(\\alpha = 0.1)$ , FedWSQ (1-bit) achieves $62.05\\%$ accuracy on CIFAR-100, surpassing FedRCL and FedACG by $3.79\\%$ and $3.91\\%$ , respectively. Similarly, on Tiny/ImageNet, FedWSQ (1-bit) achieves $45.11\\%$ accuracy, outperforming FedRCL and FedACG by $7.25\\%$ and $5.36\\%$ , respectively. These results highlight the effectiveness of FedWSQ, which demonstrates its robustness even in extremely low bit-width scenarios while maintaining high model accuracy.", + "bbox": [ + 88, + 551, + 482, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Additional analyses", + "text_level": 1, + "bbox": [ + 89, + 832, + 277, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of bit representation in FedWSQ FedWSQ balances model performance and communication efficiency by adopting different bit-width settings. Table 1 presents the", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1ed1f36c553f79a20bc740e1b8baee9687320d82ed9fa44955e3e9d54d496882.jpg", + "image_caption": [ + "Figure 3. Loss landscape of FL methods trained on CIFAR-100, where the numbers indicate the Hessian top eigenvalue." + ], + "image_footnote": [], + "bbox": [ + 519, + 380, + 901, + 468 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "results for 32-bit (FedWS), 4-bit, and 1-bit representations, as well as adaptive bit-allocation strategies (FBA and DBA). Despite using lower bit-widths, FedWSQ remains robust in non-i.i.d. settings, particularly under high heterogeneity ( $\\alpha = 0.1$ ). Since local models tend to overfit under high data heterogeneity, quantizing model updates helps to regularize overfitting, achieving more robust FL performance [2, 3]. The accuracy gaps among different bit-widths are relatively small, and thus the FBA and DBA strategies enable efficient communication with marginal performance loss. As shown in Table 1, both FBA and DBA strategies achieve performance close to the 4-bit setting while reducing communication costs by half. In summary, FedWSQ confirms its ability to maintain accuracy while significantly reducing communication bandwidth usage for real-world FL scenarios.", + "bbox": [ + 511, + 512, + 906, + 753 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of WS and DANUQ Table 2 shows a comparative analysis of our proposed DANUQ and UQ under varying bit-widths and data heterogeneity. In this experiment, UQ is implemented using the FedPAQ approach [37]. As listed in the table, our DANUQ consistently outperforms the UQ method. Under extremely low-bit quantization (1-bit and 2-bit settings), the UQ suffers from significant accuracy degradation. While WS alone improves performance, combining WS with the proposed DANUQ yields the high-", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4622", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8fadb22a3ed349650b85592d01a5eb031318e5ec26d22a8d260921f7799f682d.jpg", + "table_caption": [ + "Table 4. FL performance of different backbones for $\\alpha = {0.3}$ ." + ], + "table_footnote": [], + "table_body": "
BackboneVanillaFedWSFedWSQ (DBA)
ShuffleNet [53]36.3753.3251.12
VGGNet-9 [40]46.1460.0160.48
SqueezeNet [14]39.4555.8756.10
ResNet-18 [11]47.3968.1966.32
MobileViT [32]35.4240.8041.27
", + "bbox": [ + 125, + 106, + 441, + 183 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "est performance gains. The UQ with WS shows improvements compared with its non-WS counterparts, but the synergy between WS and DANUQ provides the most robust FL performance. Furthermore, our method maintains high accuracy even with low-bit representations. This suggests that applying FBA or DBA can be highly effective, as our approach maintains strong performance across varying bit-widths and makes FedWSQ particularly practical for real-world FL deployments.", + "bbox": [ + 89, + 191, + 480, + 328 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison with other NUQ methods We evaluate the effectiveness of the proposed DANUQ by comparing it with existing NUQ approaches, FP [45] and NF [7].7 As shown in Table 3, at the 4-bit representation, all NUQ methods perform competitively, indicating that NUQ is generally robust at relatively higher bit-widths. However, under 1-bit and 2-bit representations, FP and NF suffer from severe performance degradation, particularly in highly heterogeneous data ( $\\alpha = 0.1$ ). In contrast, our DANUQ outperforms both NF and FP across all conditions, which confirms its robustness against extreme quantization constraints. These results highlight the superiority of our DANUQ method among NUQ methods for communication-efficient FL.", + "bbox": [ + 89, + 333, + 482, + 530 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Loss Landscape Analysis Figure 3 visualizes the loss landscapes of various FL methods compared to FedWSQ (with DBA). FedWSQ achieves the smoothest and most stable loss landscape, with the lowest Hessian top eigenvalue of 135.8, indicating improved generalization ability. FedSmoo exhibits a flatter landscape but converges to a much higher global minimum, reducing sharpness at the cost of performance. In contrast, FedACG reaches a global comparable to FedWSQ but has a significantly sharper landscape, with a Hessian top eigenvalue of 671.2, which indicates it is more sensitive to perturbations. These results show that FedWSQ effectively enhances the generalization, leading to more stable training and improved FL performance.", + "bbox": [ + 89, + 536, + 482, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact on different backbone architectures Table 4 presents the performance of different backbone architectures on CIFAR-100 to evaluate the impact of FedWS and FedWSQ. For all architectures, FedWS significantly improves accuracy over the vanilla models, and FedWSQ maintains a comparable performance to FedWS with improved communication efficiency. In addition, We eval", + "bbox": [ + 89, + 758, + 482, + 864 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c5543337dc8d48510ecc4d8a128bd1c3e35ed2623b8ce2ad1339587b1ee10beb.jpg", + "table_caption": [ + "Table 5. Ablation study for $\\rho$ using FedWS on CIFAR-10. For the non-i.i.d. setting, $\\alpha = {0.3}$ is used." + ], + "table_footnote": [], + "table_body": "
ρ1 × 10-41 × 10-31 × 10-21 × 10-1
non-i.i.d.87.1589.7189.6289.46
i.i.d.90.6492.4891.9792.11
", + "bbox": [ + 540, + 121, + 872, + 165 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "uate MobileViT [32], a vision transformer-based architecture, which also benefits from FedWS and FedWSQ. These results confirm their applicability beyond CNNs.", + "bbox": [ + 511, + 174, + 903, + 220 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison with FedWon We provide the comparison of FedWSQ with FedWon, which was discussed in Section 3.1. As shown in Table 1, FedWSQ substantially outperforms FedWon across all datasets and data heterogeneity settings. As illustrated in Figure 3, FedWon exhibits a much higher global minimum and a very sharp loss landscape with a Hessian top eigenvalue of 203.4, whereas FedWSQ maintains a smoother surface with an eigenvalue of 135.8. These results demonstrate FedWSQ's superior robustness and generalization ability.", + "bbox": [ + 511, + 229, + 906, + 380 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Hyper-parameters We examine the hyper-parameter $\\rho$ for WS, which handles the normalization scales. To evaluate its impact, we vary the values of $\\rho$ and assess the performance of FedWS under both non-i.i.d. and i.i.d. data distributions. As shown in Table 5, the performance of FedWS is insensitive to $\\rho$ variations. This is because during inference, normalization layers like GN cancel the effect of constant scaling, and the variance of the parameter vectors within the same convolutional layer is not significant. We select $\\rho = 1 \\times 10^{-3}$ , which yields the best results.", + "bbox": [ + 511, + 388, + 908, + 541 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 556, + 638, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presents FedWSQ, a novel FL framework using WS and the proposed DANUQ to address key FL challenges, including data heterogeneity, partial client participation, and communication bottleneck. By leveraging WS, FedWSQ enhances training stability via an implicit gradient filtering mechanism. Additionally, DANUQ efficiently compresses LMPUs while preserving essential local information, significantly reducing communication overhead even at ultra-low bit-widths. Extensive experiments on the benchmark datasets confirm that FedWSQ consistently outperforms SOTA FL methods under various conditions. Furthermore, the mixed-precision strategies FBA and DBA effectively balance communication cost and model accuracy. These results validate the robustness of the FedWSQ framework in extreme quantization scenarios, making it highly practical for real-world FL.", + "bbox": [ + 511, + 582, + 906, + 825 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments This work was supported by the National Research Foundation (NRF), South Korea, under project BK21 FOUR (Smart Robot Convergence and Application Education Research Center).", + "bbox": [ + 511, + 839, + 905, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "${}^{7}$ QLs of different methods are visualized in Section C of the supplementary document.", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4623", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Durmus Alp Emre Acar, Yue Zhao, Ramon Matas, Matthew Mattina, Paul Whatmough, and Venkatesh Saligrama. Federated learning based on dynamic regularization. In ICLR, 2021. 2, 6, 3", + "[2] MohammadHossein AskariHemmat, Reyhane Askari Hemmat, Alex Hoffman, Ivan Lazarevich, Ehsan Saboori, Olivier Mastropietro, Sudhakar Sah, Yvon Savaria, and Jean-Pierre David. QReg: On regularization effects of quantization. arXiv preprint arXiv:2206.12372, 2022. 7", + "[3] MohammadHossein AskariHemmat, Ahmadreza Jeddi, Reyhane Askari Hemmat, Ivan Lazarevich, Alexander Hoffman, Sudhakar Sah, Ehsan Saboori, Yvon Savaria, and Jean-Pierre David. QGen: On the ability to generalize in quantization aware training. arXiv preprint arXiv:2404.11769, 2024. 7", + "[4] Debraj Basu, Deepesh Data, Can Karakus, and Suhas N Diggavi. Qsparse-Local-SGD: Distributed SGD with quantization, sparsification, and local computations. IEEE Journal on Selected Areas in Information Theory, 1(1):217-226, 2020. 1", + "[5] Shengbo Chen, Cong Shen, Lanxue Zhang, and Yuanmin Tang. Dynamic aggregation for heterogeneous quantization in federated learning. IEEE Transactions on Wireless Communications, 20(10):6804-6819, 2021. 2, 6", + "[6] Jungwook Choi, Zhuo Wang, Swagath Venkataramani, Pierce I-Jen Chuang, Vijayalakshmi Srinivasan, and Kailash Gopalakrishnan. Pact: Parameterized clipping activation for quantized neural networks. In ICLR, 2018. 2", + "[7] T. Dettmers, M. Lewis, S. Shleifer, and L. Zettlemoyer. 8-bit optimizers via block-wise quantization. In ICLR, 2022. 4, 8", + "[8] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: Efficient finetuning of quantized llms. In NeurIPS, 2023. 2", + "[9] Steven K. Esser, Jeffrey L. McKinstry, Deepika Bablani, Rathinakumar Appuswamy, and Dharmendra S. Modha. Learned Step Size Quantization. In ICLR, 2020. 2", + "[10] Robert M. Gray and David L. Neuhoff. Quantization. IEEE transactions on information theory, 44(6):2325-2383, 1998. 2", + "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6, 8, 3", + "[12] Kevin Hsieh, Amar Phanishayee, Onur Mutlu, and Phillip Gibbons. The non-IID data quagmire of decentralized machine learning. In ICML, 2020. 6, 3", + "[13] Tzu-Ming Harry Hsu, Hang Qi, and Matthew Brown. Measuring the effects of non-identical data distribution for federated visual classification. arXiv preprint arXiv:1909.06335, 2019. 2, 6, 3", + "[14] Forrest N Iandola, Song Han, Matthew W Moskewicz, Khalid Ashraf, William J Dally, and Kurt Keutzer. SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and $< 0.5$ MB model size. arXiv preprint arXiv:1602.07360, 2016.8", + "[15] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015. 6" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Benoit Jacob, Skirmantas Kligys, Bo Chen, Menglong Zhu, Matthew Tang, Andrew Howard, Hartwig Adam, and Dmitry Kalenichenko. Quantization and training of neural networks for efficient integer-arithmetic-only inference. In CVPR, pages 2704-2713, 2018. 2, 4", + "[17] Sangil Jung, Changyong Son, Seohyung Lee, Jinwoo Son, Youngjun Kwak, Jae-Joon Han, Sung Ju Hwang, and Changkyu Choi. Learning to quantize deep networks by optimizing quantization intervals with task loss. In CVPR, pages 4350-4359, 2019. 2", + "[18] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank J Reddi, Sebastian U Stich, and Ananda Theertha Suresh. SCAFFOLD: Stochastic controlled averaging for ondevice federated learning. In ICML, 2020. 1", + "[19] Ahmed Khaled, Konstantin Mishchenko, and Peter Richtárik. First analysis of local GD on heterogeneous data. arXiv preprint arXiv:1909.04715, 2019. 1", + "[20] Geeho Kim, Jinkyu Kim, and Bohyung Han. Communication-efficient federated learning with accelerated client gradient. In CVPR, 2024. 6, 3", + "[21] Jinkyu Kim, Geeho Kim, and Bohyung Han. Multi-level branched regularization for federated learning. In ICML, 2022. 2, 6, 3", + "[22] Jakub Konečny, H. Brendan McMahan, Felix X. Yu, Ananda Theertha Suresh, Dave Bacon, and Peter Richtárik. Federated learning: Strategies for improving communication efficiency. In arXiv preprint, 2017. 2", + "[23] Alex Krizhevsky. Learning multiple layers of features from tiny images. Master's thesis, Department of Computer Science, University of Toronto, 2009. 6", + "[24] Khiem Le, Nhan Luong-Ha, Manh Nguyen-Duc, Danh Le-Phuoc, Cuong Do, and Kok-Seng Wong. Exploring the practicality of federated learning: A survey towards the communication perspective. arXiv preprint arXiv:2405.20431, 2024. 1", + "[25] Ya Le and Xuan Yang. Tiny ImageNet visual recognition challenge. CS 231N, 7(7):3, 2015. 6", + "[26] Gihun Lee, Minchan Jeong, Yongjin Shin, Sangmin Bae, and Se-Young Yun. Preservation of the global knowledge by not-true distillation in federated learning. In NeurIPS, 2022. 6, 3", + "[27] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. In MLSys, 2020. 2, 6, 3", + "[28] Xiang Li, Kaixuan Huang, Wenhao Yang, Shusen Wang, and Zhihua Zhang. On the convergence of FedAvg on non-IID data. In ICLR, 2020. 1", + "[29] Darryl Lin, Sachin Talathi, and Sreekanth Annapureddy. Fixed point quantization of deep convolutional networks. In International conference on machine learning, pages 2849-2858. PMLR, 2016. 2", + "[30] Wang Luping, Wang Wei, and Li Bo. CMFL: Mitigating communication overhead for federated learning. In 2019 IEEE 39th international conference on distributed computing systems (ICDCS), pages 954-964. IEEE, 2019. 1", + "[31] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4624", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "efficient learning of deep networks from decentralized data. In AISTATS, 2017. 1, 2, 6", + "[32] Sachin Mehta and Mohammad Rastegari. Mobilevit: lightweight, general-purpose, and mobile-friendly vision transformer. arXiv preprint arXiv:2110.02178, 2021. 8", + "[33] Markus Nagel, Marios Fournarakis, Rana Ali Amjad, Yelysei Bondarenko, Mart van Baalen, and Tijmen Blankevoort. A white paper on neural network quantization. arXiv preprint arXiv:2106.08295, 2021. 2, 4", + "[34] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. PyTorch: An imperative style, high-performance deep learning library. In NeurIPS, 2019. 3", + "[35] Siyuan Qiao, Huiyu Wang, Chenxi Liu, Wei Shen, and Alan Yuille. Micro-batch training with batch-channel normalization and weight standardization. arXiv preprint arXiv:1903.10520, 2019. 1, 3, 6", + "[36] Sashank J Reddi, Zachary Charles, Manzil Zaheer, Zachary Garrett, Keith Rush, Jakub Konečny, Sanjiv Kumar, and Hugh Brendan McMahan. Adaptive federated optimization. In ICLR, 2021. 2, 6, 3", + "[37] Amirhossein Reisizadeh, Aryan Mokhtari, Hamed Hassani, Ali Jabbabaie, and Ramtin Pedarsani. FedPAQ: A communication-efficient federated learning method with periodic averaging and quantization. In AISTATS, 2020. 2, 6, 7", + "[38] Seonguk Seo, Jinkyu Kim, Geeho Kim, and Bohyung Han. Relaxed contrastive learning for federated learning. In CVPR, 2024. 6, 3", + "[39] Yujun Shi, Jian Liang, Wenqing Zhang, Vincent YF Tan, and Song Bai. Towards understanding and mitigating dimensional collapse in heterogeneous federated learning. In ICLR, 2023. 6, 3", + "[40] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2014. 8", + "[41] Sebastian U Stich. Local SGD converges fast and communicates little. In ICLR, 2019. 1", + "[42] Sebastian U Stich and Sai Praneeth Karimireddy. The errorfeedback framework: Better rates for SGD with delayed gradients and compressed communication. arXiv preprint arXiv:1909.05350, 2019. 1", + "[43] Yan Sun, Li Shen, Shixiang Chen, Liang Ding, and Dacheng Tao. Dynamic regularized sharpness aware minimization in federated learning: Approaching global consistency and smooth landscape. In International Conference on Machine Learning, pages 32991-33013. PMLR, 2023. 2, 6", + "[44] Jianyu Wang and Gauri Joshi. Cooperative SGD: A unified framework for the design and analysis of local-update SGD algorithms. Journal of Machine Learning Research, 22(213): 1-50, 2021. 1", + "[45] Ruizhe Wang, Yeyun Gong, Xiao Liu, Guoshuai Zhao, Ziyue Yang, Baining Guo, Zhengjun Zha, and Peng Cheng. Optimizing large language model training using fp4 quantization. In arXiv preprint arXiv:2501.17116, 2025. 4, 8" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Jie Wen, Zhixia Zhang, Yang Lan, Zhihua Cui, Jianghui Cai, and Wensheng Zhang. A survey on federated learning: challenges and applications. International Journal of Machine Learning and Cybernetics, 14(2):513-535, 2023. 1", + "[47] Yuxin Wu and Kaiming He. Group normalization. In ECCV, 2018. 6", + "[48] Jing Xu, Sen Wang, Liwei Wang, and Andrew Chi-Chih Yao. FedCM: Federated learning with client-level momentum. arXiv preprint arXiv:2106.10874, 2021. 6, 3", + "[49] Haibo Yang, Minghong Fang, and Jia Liu. Achieving linear speedup with partial worker participation in non-IID federated learning. In ICLR, 2021. 1", + "[50] Hao Yu, Sen Yang, and Shenghuo Zhu. Parallel restarted SGD with faster convergence and less communication: Demystifying why model averaging works for deep learning. In AAAI, 2019. 1", + "[51] Dongqing Zhang, Jiaolong Yang, Dongqiangzi Ye, and Gang Hua. LQ-Nets: Learned Quantization for Highly Accurate and Compact Deep Neural Networks. In ECCV, 2018. 2", + "[52] Jie Zhang, Zhiqi Li, Bo Li, Jianghe Xu, Shuang Wu, Shouhong Ding, and Chao Wu. Federated learning with label distribution skew via logits calibration. In ICML, 2022. 6, 3", + "[53] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. ShuffleNet: An extremely efficient convolutional neural network for mobile devices. In CVPR, 2018. 8", + "[54] Sijie Zhao, Tao Yue, and Xuemei Hu. Distribution-aware adaptive multi-bit quantization. In CVPR, 2021. 4", + "[55] Chenzhuo Zhu, Song Han, Huizi Mao, and William J. Dally. Trained ternary quantization. In ICLR, 2017. 2", + "[56] Weiming Zhuang and Lingjuan Lyu. FedWon: Triumphing multi-domain federated learning without normalization. In ICLR, 2023. 4, 6" + ], + "bbox": [ + 516, + 90, + 903, + 559 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4625", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/FedWSQ_ Efficient Federated Learning with Weight Standardization and Distribution-Aware Non-Uniform Quantization/1f166dee-20e5-452e-8e86-86019ce8d0bb_model.json b/2025/FedWSQ_ Efficient Federated Learning with Weight Standardization and Distribution-Aware Non-Uniform Quantization/1f166dee-20e5-452e-8e86-86019ce8d0bb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b734de86fad3033aaad803dc9b397a39e95ab9cf --- /dev/null +++ b/2025/FedWSQ_ Efficient Federated Learning with Weight Standardization and Distribution-Aware Non-Uniform Quantization/1f166dee-20e5-452e-8e86-86019ce8d0bb_model.json @@ -0,0 +1,2420 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.13, + 0.13, + 0.87, + 0.177 + ], + "angle": 0, + "content": "FedWSQ: Efficient Federated Learning with Weight Standardization and Distribution-Aware Non-Uniform Quantization" + }, + { + "type": "text", + "bbox": [ + 0.167, + 0.203, + 0.831, + 0.221 + ], + "angle": 0, + "content": "Seung-Wook Kim\\*1, Seongyeol Kim\\*1, Jiah Kim\\*1, Seowon Ji\\*2, and Se-Ho Lee\\*3" + }, + { + "type": "text", + "bbox": [ + 0.202, + 0.222, + 0.798, + 0.239 + ], + "angle": 0, + "content": "\\(^{1}\\)Department of Intelligent Robot Engineering, Pukyong National University" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.24, + 0.771, + 0.257 + ], + "angle": 0, + "content": "\\(^{2}\\)CSE, Konkuk University \\(^{3}\\)CS&AI, Jeonbuk National University" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.325, + 0.486, + 0.612 + ], + "angle": 0, + "content": "Federated learning (FL) often suffers from performance degradation due to key challenges such as data heterogeneity and communication constraints. To address these limitations, we present a novel FL framework called FedWSQ, which integrates weight standardization (WS) and the proposed distribution-aware non-uniform quantization (DANUQ). WS enhances FL performance by filtering out biased components in local updates during training, thereby improving the robustness of the model against data heterogeneity and unstable client participation. In addition, DANUQ minimizes quantization errors by leveraging the statistical properties of local model updates. As a result, FedWSQ significantly reduces communication overhead while maintaining superior model accuracy. Extensive experiments on FL benchmark datasets demonstrate that FedWSQ consistently outperforms existing FL methods across various challenging FL settings, including extreme data heterogeneity and ultra-low-bit communication scenarios. The source code is available at our project page1." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.629, + 0.222, + 0.645 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.654, + 0.483, + 0.805 + ], + "angle": 0, + "content": "In large-scale machine learning, centralized approaches often raise privacy concerns because sensitive data from edge devices must be collected on a central server for model training. Federated learning (FL) [31] enables distributed devices to collaboratively train a shared global model without sharing raw data. A fundamental FL method, FedAvg, was proposed by McMahan et al. [31]. In this approach, each client trains its model using the local data, and then the server collects and averages these local updates to form a unified global model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.806, + 0.484, + 0.852 + ], + "angle": 0, + "content": "Several studies on FL have verified its practical effectiveness [4, 28, 41, 42, 44, 49, 50]. However, real-world FL methods often face three key challenges [18, 19, 24," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.293, + 0.905, + 0.445 + ], + "angle": 0, + "content": "30, 46]: 1) data heterogeneity, where clients possess non-independent and identically distributed (non-i.i.d.) data; 2) partial client participation, where only a subset of clients contributes to a global model update during each training round due to communication constraints; 3) communication bottlenecks, where limited communication bandwidth and high transmission costs hinder efficient model aggregation. These challenges worsen local gradient divergence, slow down the global model convergence, and ultimately degrade FL performance in real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.446, + 0.906, + 0.779 + ], + "angle": 0, + "content": "To address these limitations, we propose FedWSQ, an advanced FL framework that combines weight standardization (WS) with our proposed distribution-aware nonuniform quantization (DANUQ) to improve both learning stability and communication efficiency. WS is a plug-and-play technique that standardizes the weight vectors of convolutional or linear layers to stabilize the learning process of a neural network [35]. In FedWSQ, WS plays a crucial role in mitigating client drift by filtering out gradient components that contribute to overfitting during local training. This leads to improved generalization across heterogeneous clients and significantly enhances FL performance. Although WS accelerates global convergence, the communication cost per round remains a major bottleneck. To address this, FedWSQ integrates WS with DANUQ, a novel quantization strategy that leverages a standard normal distribution prior to minimize quantization errors, reducing communication overhead while preserving model performance. Even under ultra-low-bit quantization, the proposed FedWSQ achieves superior performance, consistently outperforming state-of-the-art (SOTA) FL methods across various datasets and FL settings." + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.78, + 0.871, + 0.793 + ], + "angle": 0, + "content": "Our main contributions are summarized as follows." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.905, + 0.886 + ], + "angle": 0, + "content": "- We propose an effective application of WS in FL, enhancing model convergence and stability under data heterogeneity and limited client participation. By implicitly performing a gradient filtering process, WS mitigates client drift while also providing a regularization effect that stabilizes training and improves generalization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "- We introduce DANUQ, a novel quantization method that" + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.796, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.863, + 0.267, + 0.876 + ], + "angle": 0, + "content": "*indicates equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.876, + 0.242, + 0.888 + ], + "angle": 0, + "content": "† Corresponding authors." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.425, + 0.9 + ], + "angle": 0, + "content": "1https://github.com/Seongyeol-kim/FedWSQ" + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.863, + 0.425, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "4616" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.103, + 0.092, + 0.483, + 0.198 + ], + "angle": 0, + "content": "employs a fixed quantization function, leveraging the statistical properties of local model parameter updates (LMPUs) to minimize quantization errors under parameter-distribution priors. By efficiently compressing LM-PUs while preserving essential information, the proposed DANUQ significantly reduces communication overhead without noticeable performance degradation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.198, + 0.482, + 0.273 + ], + "angle": 0, + "content": "- Extensive experiments show that FedWSQ significantly enhances performance across diverse settings. Even with an average of 2.3 bits per parameter, FedWSQ improves performance by over \\(5\\%\\) on Tiny-ImageNet in a highly heterogeneous setting compared to SOTA FL methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.285, + 0.387, + 0.3 + ], + "angle": 0, + "content": "2. Related Work and Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.31, + 0.484, + 0.429 + ], + "angle": 0, + "content": "Federated learning (FL) FL is a decentralized approach that enables collaborative training of a global model. Using distributed devices, FL ensures efficient learning while preserving data privacy. Let \\(\\mathcal{F}_n(\\mathbf{w}_{\\mathrm{g}})\\coloneqq \\mathbb{E}_{\\zeta_n\\sim \\mathcal{D}_n}[\\mathcal{F}_n(\\mathbf{w}_{\\mathrm{g}};\\zeta_n)]\\) be the local loss function of the client \\(n\\in [N]\\) with a local data distribution \\(\\mathcal{D}_n\\), where \\(\\mathbf{w}_{\\mathrm{g}}\\) is the global model parameter (GMP) vector. The training objective function of FL [31] is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.431, + 0.482, + 0.463 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathbf {w} _ {\\mathrm {g}}} \\left\\{\\mathcal {F} \\left(\\mathbf {w} _ {\\mathrm {g}}\\right) := \\sum_ {i \\in \\mathcal {S}} h _ {i} \\mathcal {F} _ {i} \\left(\\mathbf {w} _ {\\mathrm {g}}\\right) \\right\\}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.464, + 0.483, + 0.614 + ], + "angle": 0, + "content": "where \\( h_i \\) is the weight assigned to the \\( i \\)-th device, such that \\( h_i \\geq 0 \\), \\( \\sum_{i} h_i = 1 \\), and \\( h_i \\propto |\\mathcal{D}_i| \\), and \\( S \\subseteq [N] \\) is a subset of participating clients in the current communication round. At each round \\( t \\in [T] \\), the server broadcasts the GMPs \\( \\mathbf{w}_{\\mathrm{g}}^{t-1} \\) to the participating client \\( i \\). In FedAvg [31], the \\( i \\)-th client performs \\( K \\) steps of local training to obtain the local model parameter (LMP) vector, \\( \\mathbf{w}_i^t \\), and transmits its LMP updates (LMPUs), \\( \\Delta \\mathbf{w}_i^t := \\mathbf{w}_i^t - \\mathbf{w}_{\\mathrm{g}}^{t-1} \\), to the server. Finally, the server aggregates these LMPUs through weighted averaging to derive the updated GMPs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.615, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In real-world FL scenarios, unstable learning dynamics and slow convergence often occur due to non-i.i.d. data distributions and limited communication bandwidth. To address these issues, FedProx [27] penalizes LMPUs by adding a proximal term in the local objective to reduce the gap between the global and local loss functions. FedAvgM [13] integrates momentum into the global update process to mitigate variance in model aggregation. FedAdam [36] introduces an adaptive optimizer that improves global and local model convergence, thereby accelerating overall training progress. FedDyn [1] presents dynamic regularization for each device to diminish deviations in the LMPUs caused by data heterogeneity. FedSmoo [43] applies constraints that prevent sharp fluctuations in local updates to generalize FL performance by ensuring global consistency. Recently, FedACG [21] was proposed to improve convergence by broadcasting a global model with a lookahead gradient and aligning local models with the shared global model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "Network quantization As neural network architectures have grown larger in scale, quantization has become essential for the memory optimization and acceleration of operations [9, 10, 16, 17, 29, 33, 55]. Let us consider a model using \\( B \\)-bit representation. A typical quantization function \\( \\mathcal{Q} \\) that maps a full-precision value \\( x \\) to a quantization level (QL) \\( q \\) can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.664, + 0.207, + 0.905, + 0.223 + ], + "angle": 0, + "content": "\\[\nq = \\mathcal {Q} (x / s), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.231, + 0.905, + 0.292 + ], + "angle": 0, + "content": "where \\(s\\) is a scaling factor, and each QL is matched with a \\(B\\)-bit integer from \\(\\{0,\\dots ,2^B -1\\}\\). The primary goal of effective quantization is to determine the proper scaling factor and a set of QLs to minimize the quantization errors." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.293, + 0.907, + 0.551 + ], + "angle": 0, + "content": "Quantization methods can be roughly categorized into two approaches: uniform quantization (UQ) and nonuniform quantization (NUQ). UQ divides the input range into \\(2^{B}\\) equal intervals and assigns a representative value as a QL to each interval [10, 29]. In contrast, NUQ adjusts the interval widths based on the prior data distribution [6, 8, 51]. By allocating finer intervals in dense data regions and broader intervals in sparse regions, NUQ preserves critical information more effectively than UQ. For the scaling factor \\(s\\), both approaches commonly employ the absmax method [16, 33] to determine the dynamic range for quantization. Furthermore, to enhance data adaptability, researchers have explored methods for dynamically learning quantization parameters (e.g., scaling factors and QLs) [9, 17, 55]. During backpropagation, these parameters are updated to reduce quantization error and better align with the underlying data distribution." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Quantized FL In FL, communication bottlenecks arise primarily during uplink transmission, as each participating client must send its individual LMPUs to the central server. Unlike downlink transmission, where a single GMP is broadcast to multiple clients, uplink communication involves substantial data transfer. This problem is exacerbated in resource-constrained communication environments, such as mobile and IoT networks, where upload bandwidth is severely limited [5, 37]. To overcome this challenge, several studies have focused on compressing model updates. In [22], quantization and sparsification are combined to reduce the number of transmitted parameters. FedPAQ [37] employs a probabilistic rounding algorithm to prevent an excessive concentration of identical QLs. Meanwhile, FedHQ+ [5] assigns different weight factors to clients for global updates based on their individual quantization errors. Although these methods significantly improve communication efficiency, they often suffer from performance degradation under extremely low bit-width conditions. In this paper, we introduce FedWSQ, a novel quantization framework that is designed to maintain robust model performance even at ultra-low bit-width." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4617" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.182, + 0.091, + 0.816, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.179, + 0.245, + 0.816, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.339, + 0.908, + 0.369 + ], + "angle": 0, + "content": "Figure 1. The overall framework of FedWSQ, which integrates WS and DANUQ for efficient FL. Clients transmit LMPUs quantized by the DANUQ along with scaling vectors to the server, where dequantization and aggregation are performed to update the global parameters." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.382, + 0.265, + 0.399 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.408, + 0.484, + 0.651 + ], + "angle": 0, + "content": "Figure 1 illustrates the overall framework of the proposed FedWSQ, which employs WS and DANUQ for efficient FL. Each client has a standard neural network model consisting of multiple model blocks, where WS is applied to mitigate the impact of data heterogeneity between clients. After local training, the full-precision values of LMPUs are converted to a \\(B\\)-bit representation using the global scaling vectors and carefully designed QLs. These quantized LMPUs and their corresponding scaling vectors are transmitted to the server. The server then dequantizes the received data to higher-precision values and aggregates them to update the GMPs. In addition, the global scaling vectors are refined to ensure consistent scaling across different clients for DANUQ. Using the proposed DANUQ, FedWSQ enhances communication efficiency without sacrificing model performance in challenging FL scenarios." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.664, + 0.353, + 0.681 + ], + "angle": 0, + "content": "3.1. Weight standardization (WS)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.687, + 0.484, + 0.839 + ], + "angle": 0, + "content": "WS [35] standardizes each weight vector of a convolution or linear layer to enhance the learning process. By ensuring consistent parameter distributions during training, WS contributes to a more stable gradient flow. Beyond this general benefit, we argue that the use of WS in FL can address the challenges posed by the distributed nature of FL and data heterogeneity. Formally, consider the \\(n\\)-th client that has a network architecture consisting of \\(L\\) linear or convolutional layers. Let the \\(l\\)-th layer (without bias) map an input \\(\\mathbf{x}_l \\in \\mathbb{R}^{I_l}\\) to an output \\(\\mathbf{y}_l \\in \\mathbb{R}^{O_l}\\), defined by" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.849, + 0.483, + 0.899 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {y} _ {l} = \\mathbf {W} _ {n, l} ^ {T} \\mathbf {x} _ {l} \\tag {3} \\\\ := \\left[ \\begin{array}{c c c c} \\mathbf {w} _ {n, m _ {1} ^ {l}} & \\mathbf {w} _ {n, m _ {2} ^ {l}} & \\dots & \\mathbf {w} _ {n, m _ {O _ {l}} ^ {l}} \\end{array} \\right] ^ {T} \\mathbf {x} _ {l}, \\qquad (4) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.383, + 0.907, + 0.476 + ], + "angle": 0, + "content": "where \\(\\mathbf{w}_{n,m}\\), for \\(m\\in \\mathcal{M}_l\\coloneqq \\{m_1^l,\\ldots ,m_{O_l}^l\\}\\), is the weight vector. In WS, each \\(\\mathbf{w}_{n,m}\\) is standardized before being used within the layer. We denote the prestandardized parameter (PSP) vector as \\(\\mathbf{w}_{n,m}\\), and the weight-standardized parameter (WSP) vector as \\(\\tilde{\\mathbf{w}}_{n,m}\\), respectively. Then, the WSP vector \\(\\tilde{\\mathbf{w}}_{n,m}\\) is obtained by" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.481, + 0.907, + 0.513 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {w}} _ {n, m} = \\frac {\\rho}{\\sigma (\\mathbf {w} _ {n , m})} \\left(\\mathbf {I} - \\mathbf {P _ {1}}\\right) \\mathbf {w} _ {n, m}, \\qquad (5)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.907, + 0.643 + ], + "angle": 0, + "content": "where \\(\\mathbf{1} \\in \\mathbb{R}^{I_l}\\) is a vector of all ones, \\(\\sigma(\\mathbf{w}_{n,m})\\) is the standard deviation of the elements in \\(\\mathbf{w}_{n,m}\\), and \\(\\mathbf{P_v}\\) represents the projection matrix onto the subspace spanned by the vector \\(\\mathbf{v}\\), which is denoted by \\(\\mathrm{span}\\{\\mathbf{v}\\}\\). Here, we introduce a hyper-parameter \\(\\rho\\) to control the normalization scale. Note that we reformulate the WSP computation from [35] into Eq. (5) to better understand the gradient filtering process of WS during local training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.652, + 0.909, + 0.79 + ], + "angle": 0, + "content": "Gradient filtering by WS WS plays an important role in reducing the learning diversity of local models, which is one of the most critical issues in FL. To understand its impact, we compare the gradient of a loss function \\(\\mathcal{L}\\) with respect to the WSP vector \\(\\tilde{\\mathbf{w}}_{n,m}\\) to that of the PSP vector \\(\\mathbf{w}_{n,m}\\). When WS is applied, the gradient is processed by an additional series of projections to remove specific components that may interfere with effective FL. Specifically, the gradient with respect to the PSP vector \\(\\mathbf{w}_{n,m}\\) is obtained by" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.797, + 0.907, + 0.832 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial \\mathcal {L}}{\\partial \\mathbf {w} _ {n , m}} = \\frac {\\rho}{\\sigma (\\mathbf {w} _ {n , m})} \\left(\\mathbf {I} - \\mathbf {P} _ {\\mathbf {1}}\\right) \\left(\\mathbf {I} - \\mathbf {P} _ {\\tilde {\\mathbf {w}} _ {n, m}}\\right) \\frac {\\partial \\mathcal {L}}{\\partial \\tilde {\\mathbf {w}} _ {n , m}}. (6)\n\\]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.512, + 0.839, + 0.908, + 0.888 + ], + "angle": 0, + "content": "2Strictly speaking, convolution involves spatial structures aggregating information from local receptive fields. For simplicity, we omit spatial considerations and treat the convolution as a linear transformation. This abstraction can be easily generalized to multidimensional tensor structures." + }, + { + "type": "page_footnote", + "bbox": [ + 0.533, + 0.888, + 0.885, + 0.901 + ], + "angle": 0, + "content": "3Section A of the supplementary document provides its derivation." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.839, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4618" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.091, + 0.478, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.22, + 0.459, + 0.235 + ], + "angle": 0, + "content": "Figure 2. An example of gradient filtering process of WS." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.253, + 0.483, + 0.337 + ], + "angle": 0, + "content": "The upstream gradient \\(\\frac{\\partial\\mathcal{L}}{\\partial\\tilde{\\mathbf{w}}_{n,m}}\\) is projected onto two subspaces, \\(\\mathrm{span}\\{\\tilde{\\mathbf{w}}_{n,m}\\}^{\\perp}\\) and \\(\\mathrm{span}\\{\\mathbf{1}\\}^{\\perp}\\), where \\(\\mathrm{span}\\{\\mathbf{v}\\}^{\\perp}\\) represents the orthogonal complement of \\(\\mathrm{span}\\{\\mathbf{v}\\}\\). Since \\(\\tilde{\\mathbf{w}}_{n,m} \\perp \\mathbf{1}\\), the upstream gradient \\(\\frac{\\partial\\mathcal{L}}{\\partial\\tilde{\\mathbf{w}}_{n,m}}\\) is ultimately projected onto \\(\\mathrm{span}\\{\\tilde{\\mathbf{w}}_{n,m}, \\mathbf{1}\\}^{\\perp}\\) after the two projections." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.338, + 0.484, + 0.671 + ], + "angle": 0, + "content": "As illustrated in Figure 2, Eq. (6) represents a gradient filtering process that applies successive projections to remove undesirable components from the gradient with respect to LMPs. In FL, data heterogeneity causes local models to overfit their local data, which leads to inconsistent updates across clients. This inconsistency results in a discrepancy between the GMPs and the optimal parameters, known as client drift. WS mitigates this issue by projecting the gradient onto \\(\\mathrm{span}\\{\\tilde{\\mathbf{w}}_{n,m},\\mathbf{1}\\}^{\\perp}\\), which effectively reduces the impact of client drift during training. The first projection eliminates the gradient component aligned with \\(\\tilde{\\mathbf{w}}_{n,m}\\), which corresponds to the LMPs that are biased due to the non-i.i.d. data distribution. The second projection removes the mean component of mini-batch gradients, which can also be biased toward the local data distribution, compounding inconsistencies across clients. By filtering out both the parameter-aligned and mean components from the local gradient, WS preserves meaningful gradient directions essential for stable global model convergence. Additionally, this projection-based filtering introduces a regularization effect by constraining the model's deviation from its current state, which can improve the model's generalization ability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Comparison with FedWon It is worth noting that FedWon [56], a recently proposed FL method, also employs WS to address data heterogeneity in multi-domain FL. However, a key difference lies in the parameters exchanged between clients and the server, which fundamentally alters the learning dynamics. FedWon transmits WSP, which explicitly normalizes parameters across clients to improve learning stability. However, this approach forces all clients' LMPUs to exhibit statistically similar characteristics, which potentially discards client-specific information crucial for local adaptation. In contrast, FedWSQ transmits PSP, preserving essential local information while implicitly mitigating harmful divergences through the gradient filtering process of WS, as shown in Figure 2. This design choice en" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.152 + ], + "angle": 0, + "content": "ables FedWSQ to achieve a better balance between global stability and local information preservation. An experimental comparison between FedWon and FedWSQ will be discussed in Section 4.4." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.161, + 0.737, + 0.178 + ], + "angle": 0, + "content": "3.2. Quantization of LMPUs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.183, + 0.907, + 0.365 + ], + "angle": 0, + "content": "Neural network quantization often involves learning quantization parameters. However, in FL, transmitting additional quantization parameters or auxiliary information for each communication round would introduce significant resource overhead. Instead, we propose a fixed quantization function that eliminates the need to transmit additional quantization information, ensuring a lightweight and efficient communication process. A fundamental assumption in NUQ is that model parameters follow a normal distribution, as demonstrated in prior studies [7, 45, 54]. Since LMPUs represent the difference between an updated LMP and the previous GMP, they may also exhibit a normal distribution." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.375, + 0.909, + 0.632 + ], + "angle": 0, + "content": "LMPU scaling Most quantization methods adopt the absmax strategy to manage the dynamic range of values [16, 33]. This approach scales an input tensor to the range \\([-1,1]\\) by dividing each element by the maximum absolute value of the tensor. Within this constrained range, conventional quantization methods determine QLs and map input values to the nearest quantization point. However, these methods are highly sensitive to outliers, which can excessively expand the dynamic range of the target tensor. Such over-scaling often results in underflow, especially in extremely low-bit representation. To solve this problem, we use the standard deviation of the input tensor as a scaling factor. Unlike the absmax value, the standard deviation is more robust to outliers. Moreover, since our DANUQ is based on a probability model of a standard normal distribution, the standard deviation provides a more reliable approximation of the target dynamic range." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.632, + 0.909, + 0.801 + ], + "angle": 0, + "content": "In FedWSQ, we use a shared global scaling factor to help maintain consistency in quantization. Instead of independently using scaling factors, we employ a collaborative approach in which each client contributes to a global scaling vector, ensuring that local standard deviations are taken into account. Specifically, each client \\(i\\) computes a scaling vector, defined as \\(\\mathbf{s}_i = [s_{i,1},\\dots ,s_{i,L}]^T\\), where \\(s_{i,l}\\) denotes the standard deviation of the LMPUs for the \\(l\\)-th layer. The server then collects and aggregates \\(\\mathbf{s}_i\\) from the participating clients \\(i\\in S\\), and updates the global scaling vector \\(\\mathbf{s_g} = [s_{g,1},\\dots ,s_{g,L}]^T\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.603, + 0.809, + 0.907, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mathbf {s} _ {\\mathrm {g}} \\leftarrow (1 - \\beta) \\mathbf {s} _ {\\mathrm {g}} + \\beta \\frac {1}{| \\mathcal {S} |} \\sum_ {i \\in \\mathcal {S}} \\mathbf {s} _ {i}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "where \\(\\beta\\) is a momentum parameter that controls the update rate. This approach ensures that the global scaling factor smoothly adapts while mitigating local fluctuations. Once" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4619" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.182 + ], + "angle": 0, + "content": "updated, the server distributes \\( \\mathbf{s}_{\\mathrm{g}} \\) to the participating clients. Each client then divides its LMPU updates for the \\( l \\)-th layer by the corresponding scale \\( s_{\\mathrm{g},l} \\), so that the normalized values can be regarded as samples from a standard normal distribution. Based on this assumption, we apply the proposed DANUQ scheme detailed in the next section." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.189, + 0.484, + 0.355 + ], + "angle": 0, + "content": "QLs considering normal distribution To determine the optimal multi-bit QLs, we formulate an optimization problem that minimizes the expected quantization error under a normal distribution. We assume that a random variable \\(\\Delta w\\) from the normalized LMPUs follows a standard normal distribution. Since the distribution is symmetric around zero, we consider only its nonnegative half. Given a \\(B\\)-bit representation, let \\(Q = \\{q_{0}, q_{1}, \\ldots, q_{R}\\}\\) be the set of QLs, where \\(R = 2^{B - 1} - 1\\). Here, we assume that \\(q_{r}\\) values are sorted in ascending order and set \\(q_{0} = 0\\) as a fixed value. Then, the quantization boundaries are defined as" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.363, + 0.483, + 0.419 + ], + "angle": 0, + "content": "\\[\nu _ {r} = \\left\\{ \\begin{array}{l l} q _ {0}, & \\text {i f} r = 0, \\\\ \\frac {q _ {r - 1} + q _ {r}}{2}, & \\text {i f} 1 \\leq r \\leq R, \\\\ + \\infty , & \\text {i f} r = R + 1. \\end{array} \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.427, + 0.484, + 0.473 + ], + "angle": 0, + "content": "A full-precision value in the range \\([u_r, u_{r+1})\\) is quantized into the QL \\(q_r\\). The optimal QLs are determined by minimizing the expected quantization error, formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.48, + 0.483, + 0.521 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} \\left[ \\left(\\Delta w - \\Delta \\bar {w}\\right) ^ {2} \\right] = \\sum_ {r = 0} ^ {R} \\int_ {u _ {r}} ^ {u _ {r + 1}} \\left(x - q _ {r}\\right) ^ {2} p (x) d x, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.529, + 0.484, + 0.575 + ], + "angle": 0, + "content": "where \\(\\Delta \\bar{w} \\in Q\\) represents the quantized value, and \\(p(x)\\) is the PDF of the standard normal distribution. From (9), we can derive the following optimization problem as" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.582, + 0.482, + 0.705 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {\\left\\{q _ {1}, \\dots , q _ {R} \\right\\}} \\left\\{\\frac {1}{2} \\left(q _ {R} ^ {2} + 1\\right) - \\sqrt {\\frac {2}{\\pi}} q _ {1} e ^ {- \\frac {u _ {1} ^ {2}}{2}} - \\frac {1}{2} q _ {1} ^ {2} \\operatorname {e r f} \\left(\\frac {u _ {1}}{\\sqrt {2}}\\right) \\right. \\\\ + \\sqrt {\\frac {2}{\\pi}} \\sum_ {r = 1} ^ {R - 1} (q _ {r} - q _ {r + 1}) e ^ {- \\frac {u _ {r + 1} ^ {2}}{2}} \\\\ + \\frac {1}{2} \\sum_ {r = 1} ^ {R - 1} \\left(q _ {r} ^ {2} - q _ {r + 1} ^ {2}\\right) \\operatorname {e r f} \\left(\\frac {u _ {r + 1}}{\\sqrt {2}}\\right) \\}, \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.714, + 0.484, + 0.821 + ], + "angle": 0, + "content": "where \\(\\operatorname{erf}(\\cdot)\\) denotes the error function. A closed-form solution to Eq. (22) is intractable due to the presence of nonlinear terms, including the Gaussian integral and the error function, which lack simple analytical inverses, and the interdependent QLs due to boundary conditions. As an alternative, we employ a brute-force search algorithm to numerically determine the optimal QLs. The optimal QLs" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.827, + 0.448, + 0.84 + ], + "angle": 0, + "content": "4The full derivation is provided in the supplementary document." + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "5We discretize the search space of possible QLs and perform an exhaustive evaluation of the cost function to identify the configuration that minimizes the quantization error. To improve computational efficiency, we restrict the search space to a reasonable range based on empirical observations and employ parallel processing to accelerate evaluation." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.827, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.516, + 0.092, + 0.669, + 0.106 + ], + "angle": 0, + "content": "Algorithm 1 FedWSQ" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.109, + 0.907, + 0.147 + ], + "angle": 0, + "content": "Input: Initial global parameters \\(\\mathbf{W}_{\\mathrm{g}}^{0}\\), initial scales \\(\\mathbf{s}_{\\mathrm{g}}^{0}\\), global communication round \\(T\\), local training iteration \\(K\\), number of total clients \\(N\\), number of model layers \\(L\\), learning rate \\(\\eta\\), hyperparameter \\(\\beta\\)" + }, + { + "type": "title", + "bbox": [ + 0.516, + 0.147, + 0.586, + 0.157 + ], + "angle": 0, + "content": "Server-side:" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.158, + 0.627, + 0.168 + ], + "angle": 0, + "content": "for \\(t = 1,\\dots ,T\\) do" + }, + { + "type": "text", + "bbox": [ + 0.541, + 0.169, + 0.79, + 0.181 + ], + "angle": 0, + "content": "Randomly sample a subset of clients \\(S_{t} \\subseteq [N]\\)" + }, + { + "type": "text", + "bbox": [ + 0.541, + 0.182, + 0.775, + 0.193 + ], + "angle": 0, + "content": "Set the bit-precision \\( B_{i} \\) for the client \\( i \\in S_{t} \\)" + }, + { + "type": "text", + "bbox": [ + 0.541, + 0.193, + 0.739, + 0.205 + ], + "angle": 0, + "content": "Send \\((\\mathbf{W}_{\\mathrm{g}}^{t - 1},\\mathbf{s}_{\\mathrm{g}}^{t - 1})\\) to client \\(i\\in S_{t}\\)" + }, + { + "type": "text", + "bbox": [ + 0.541, + 0.206, + 0.746, + 0.217 + ], + "angle": 0, + "content": "for each client \\(i\\in S_t\\) , in parallel do" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.217, + 0.88, + 0.233 + ], + "angle": 0, + "content": "\\[\n\\left(\\Delta \\overline {{\\mathbf {W}}} _ {i} ^ {t}, \\mathbf {s} _ {i} ^ {t}\\right) \\leftarrow \\text {C l i e n t L o c a l T r a i n i n g} \\left(\\mathbf {W} _ {\\mathrm {g}} ^ {t - 1}, \\mathbf {s} _ {\\mathrm {g}} ^ {t - 1}, i, B _ {i}\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.541, + 0.233, + 0.564, + 0.241 + ], + "angle": 0, + "content": "end" + }, + { + "type": "text", + "bbox": [ + 0.541, + 0.244, + 0.616, + 0.254 + ], + "angle": 0, + "content": "for \\(i\\in S_t\\) do" + }, + { + "type": "text", + "bbox": [ + 0.548, + 0.256, + 0.848, + 0.271 + ], + "angle": 0, + "content": "Obtain dequantized parameters \\(\\Delta_i^t\\) using \\((\\Delta \\overline{\\mathbf{W}}_i^t, \\mathbf{s}_i^t)\\)" + }, + { + "type": "text", + "bbox": [ + 0.541, + 0.271, + 0.564, + 0.28 + ], + "angle": 0, + "content": "end" + }, + { + "type": "text", + "bbox": [ + 0.541, + 0.281, + 0.775, + 0.295 + ], + "angle": 0, + "content": "Aggregate local updates: \\(\\Delta^t\\gets \\sum_{i\\in S_t}\\Delta_i^t\\)" + }, + { + "type": "text", + "bbox": [ + 0.541, + 0.296, + 0.8, + 0.31 + ], + "angle": 0, + "content": "Update global parameters: \\(\\mathbf{W}_{\\mathrm{g}}^{t}\\gets \\mathbf{W}_{\\mathrm{g}}^{t - 1} + \\Delta^{t}\\)" + }, + { + "type": "text", + "bbox": [ + 0.541, + 0.311, + 0.869, + 0.326 + ], + "angle": 0, + "content": "Update global scales: \\(\\mathbf{s}_{\\mathrm{g}}^{t}\\gets (1 - \\beta)\\mathbf{s}_{\\mathrm{g}}^{t - 1} + \\beta \\frac{1}{|\\mathcal{S}_{t}|}\\sum_{i\\in \\mathcal{S}_{t}}\\mathbf{s}_{i}^{t}\\)" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.327, + 0.539, + 0.337 + ], + "angle": 0, + "content": "end" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.35, + 0.714, + 0.363 + ], + "angle": 0, + "content": "ClientLocalTraining \\((\\mathbf{W}_{\\mathrm{g}},\\mathbf{s}_{\\mathrm{g}},i,I,B)\\)" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.364, + 0.729, + 0.374 + ], + "angle": 0, + "content": "Initialize local parameters: \\(\\mathbf{W}_j^0\\gets \\mathbf{W}_{\\mathrm{g}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.374, + 0.813, + 0.386 + ], + "angle": 0, + "content": "Set DANUQ function \\(\\mathcal{Q}(\\cdot)\\) based on the bit-precision \\(B\\)" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.387, + 0.633, + 0.398 + ], + "angle": 0, + "content": "for \\(k = 1,\\ldots ,K\\) do" + }, + { + "type": "text", + "bbox": [ + 0.541, + 0.398, + 0.705, + 0.412 + ], + "angle": 0, + "content": "Apply WS to \\(\\mathbf{W}_i^{k - 1}\\) using (5)" + }, + { + "type": "text", + "bbox": [ + 0.541, + 0.413, + 0.832, + 0.427 + ], + "angle": 0, + "content": "Compute unbiased gradient of local loss \\(\\nabla f_{i}(\\mathbf{W}_{i}^{k - 1})\\)" + }, + { + "type": "text", + "bbox": [ + 0.541, + 0.427, + 0.86, + 0.442 + ], + "angle": 0, + "content": "Update local parameters: \\(\\mathbf{W}_i^k\\gets \\mathbf{W}_i^{k - 1} - \\eta \\nabla f_i(\\mathbf{W}_i^{k - 1})\\)" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.442, + 0.539, + 0.452 + ], + "angle": 0, + "content": "end" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.454, + 0.762, + 0.467 + ], + "angle": 0, + "content": "Compute local update: \\(\\Delta \\mathbf{W}_i\\gets \\mathbf{W}_i^K -\\mathbf{W}_g\\)" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.468, + 0.691, + 0.478 + ], + "angle": 0, + "content": "for each layer \\(l = 1,\\dots ,L\\) do" + }, + { + "type": "text", + "bbox": [ + 0.541, + 0.479, + 0.871, + 0.493 + ], + "angle": 0, + "content": "Compute quantized local update: \\(\\Delta \\overline{\\mathbf{W}}_{i,l}\\gets \\mathcal{Q}(\\Delta \\mathbf{W}_{i,l} / s_{\\mathrm{g},l})\\)" + }, + { + "type": "text", + "bbox": [ + 0.541, + 0.493, + 0.804, + 0.505 + ], + "angle": 0, + "content": "Calculate local scale factors: \\( s_{i,l} \\gets \\mathrm{std}(\\Delta \\mathbf{W}_{i,l}) \\)" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.506, + 0.539, + 0.515 + ], + "angle": 0, + "content": "end" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.517, + 0.704, + 0.531 + ], + "angle": 0, + "content": "\\[\n\\Delta \\overline {{\\mathbf {W}}} _ {i} := (\\Delta \\mathbf {W} _ {i, 1}, \\dots , \\Delta \\mathbf {W} _ {i, L})\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.531, + 0.636, + 0.544 + ], + "angle": 0, + "content": "\\[\n\\mathbf {s} _ {i} := \\left[ s _ {i, 1}, \\dots , s _ {i, L} \\right]\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.544, + 0.619, + 0.556 + ], + "angle": 0, + "content": "\\[\n\\operatorname {r e t u r n} \\left(\\Delta \\overline {{\\mathbf {W}}} _ {i}, \\mathbf {s} _ {i}\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.907, + 0.69 + ], + "angle": 0, + "content": "for different bit-widths are obtained as follows: 1) 1-bit\\(^6\\): [-0.798, 0.798]; 2) 2-bit: [-1.224, 0, 0.765, 1.724]; 3) 4-bit: [-2.654, -1.974, -1.508, -1.149, -0.834, -0.544, -0.269, 0, 0.230, 0.465, 0.708, 0.966, 1.248, 1.568, 1.968, 2.649]. These QLs minimize quantization error under the normal distribution, which provides an effective trade-off between precision and computational efficiency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.701, + 0.907, + 0.868 + ], + "angle": 0, + "content": "Mixed channel-bit allocation To further enhance communication efficiency in FL, we propose an adaptive mixed-precision strategy in which each client dynamically selects its bit representation based on its communication channel conditions. This approach effectively balances model update quality and transmission cost, reducing overall bandwidth utilization. To validate this strategy, we conduct two simulation setups: 1) Fixed-bit allocation (FBA), where each client keeps a constant bit-width selected from \\(\\{1,2,4\\}\\) throughout training, and 2) dynamic-bit allocation (DBA), where each client is assigned a randomly" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.906, + 0.901 + ], + "angle": 0, + "content": "For the 1-bit case, the QLs consist of only two values. Thus, the constraint, \\( q_{0} = 0 \\), is omitted to allow optimal placement of both QLs." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4620" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.125, + 0.089, + 0.872, + 0.105 + ], + "angle": 0, + "content": "Table 1. FL performance on the three benchmark datasets with a \\(5\\%\\) participation rate over 100 clients for \\(\\alpha \\in \\{0.1, 0.3, 0.6\\}\\)." + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.108, + 0.877, + 0.4 + ], + "angle": 0, + "content": "
Method#bitsCIFAR-10CIFAR-100Tiny-ImageNet
α = 0.10.30.6i.i.d.0.10.30.6i.i.d.0.10.30.6i.i.d.
FedAvg [31]3270.6582.5285.4589.2543.4547.3949.0448.6630.3734.6536.5437.39
FedProx [27]3270.8983.9186.6088.8143.7848.2648.4948.5429.9935.4437.1437.43
FedAvgM [13]3278.3085.2787.9490.0047.7052.0453.0053.4032.0037.6339.6641.02
FedADAM [36]3272.5881.7385.9788.4546.8353.4355.9657.9334.7739.9743.2746.26
FedDyn [1]3283.6388.2889.2091.0853.8356.4056.9056.9337.3240.9242.7842.65
FedMLB [21]3271.1282.4086.2389.2549.7254.1655.1756.4533.4539.0640.5541.46
FedLC [52]3274.0383.8086.2088.3543.8648.5048.6147.8131.7334.5536.7238.06
FedNTD [26]3273.2983.3486.2689.3645.0848.7350.3950.9034.3636.4439.0640.60
FedSmoo [43]3278.6180.0785.0386.4842.0244.9745.3249.4323.0927.1429.7034.96
FedDecorr [39]3275.4483.0184.4988.4044.0549.3550.4449.4031.1234.1335.5936.36
FedWon [56]3270.8978.3682.4486.5645.5551.6853.6055.2830.5236.2038.5940.73
FedRCL [38]3283.4588.1989.7491.5558.2662.2163.9565.1137.8644.9446.7547.96
FedACG [20]3283.6289.1290.3891.7558.1462.8061.8562.4639.7545.4648.2250.49
FedPAQ [37]466.8178.5281.3685.4638.0642.4243.4644.3528.5431.9333.5634.98
157.8271.2772.3679.4630.0734.1638.2437.8524.7429.4830.7433.81
FedHQ+ [5]466.8978.4281.4285.5438.2242.6643.1044.8628.5932.1834.2435.69
158.1470.6873.1679.0431.1134.1137.5039.1724.7929.1831.1433.50
FedWSQFedWS (32)85.9489.8191.0092.4764.1468.1969.3170.1447.0552.8454.2355.26
486.8489.8491.1192.2164.3467.3168.4168.5346.5150.4951.9652.13
184.7988.6289.4991.1562.0565.1466.1966.1145.1149.9351.1151.77
FBA (2.33)86.1689.6090.4491.7163.0766.1667.0867.1746.6250.6951.4351.68
DBA (2.33)85.2089.2390.2691.7863.2866.3267.2667.2246.5050.2451.5551.80
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.412, + 0.482, + 0.489 + ], + "angle": 0, + "content": "selected bit-width at every communication round. With bit-width selection following a uniform distribution over \\(\\{1,2,4\\}\\), the expected bit-width per client is approximately 2.3 bits. The entire FedWSQ framework, including the mixed-precision strategy, is summarized in Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.506, + 0.225, + 0.523 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.532, + 0.484, + 0.624 + ], + "angle": 0, + "content": "In this section, we evaluate the proposed FedWSQ on standard FL benchmarks, compare its performance against various FL methods on both i.i.d. and non-i.i.d. data conditions, and offer a further analysis of FedWSQ. The convergence plots obtained from the experiments are provided in Section D of the supplementary document." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.636, + 0.279, + 0.654 + ], + "angle": 0, + "content": "4.1. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.901 + ], + "angle": 0, + "content": "For experiments, we use three standard benchmark datasets: CIFAR-10 [23], CIFAR-100 [23], and Tiny-ImageNet [25]. CIFAR-10 and CIFAR-100 contain 60,000 images divided into 10 and 100 classes, respectively. Tiny-ImageNet is composed of 200 classes, providing a more complex image classification task. For i.i.d. settings, training samples are randomly assigned to clients without replacement. For the non-i.i.d. setting, we sample label ratios from a Dirichlet distribution with a concentration parameter \\(\\alpha \\in \\{0.1, 0.3, 0.6\\}\\), following Hsu et al. [13]. Lower values of \\(\\alpha\\) (e.g., 0.1) indicate higher data heterogeneity, whereas larger values (e.g., 0.6) and the i.i.d. case correspond to more homogeneous distributions. The participation rate is set to \\(5\\%\\) among 100 clients. Model accuracy is evaluated on the test set of each dataset at the 1,000th communication round. To ensure stable evaluations, we report accuracy" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.413, + 0.907, + 0.624 + ], + "angle": 0, + "content": "using an exponential moving average with a smoothing parameter of 0.9. To validate the effectiveness of FedWSQ, we integrate it with FedAvg [31]. Note that FedWSQ can be incorporated into any existing FL algorithm. For clarity, we refer to the proposed method without DANUQ as FedWS. For quantized FL settings, we consider three bit-allocation strategies: constant allocation, FBA, and DBA. In the constant allocation settings, a fixed bit-width is assigned to all clients. We compare FedWSQ against various FL algorithms, including FedAvg [31], FedProx [27], FedAvgM [13], FedADAM [36], FedDyn [1], FedMLB [21], FedLC [52], FedNTD [26], FedSmoo [43], FedDecorr [39], FedWon [56], FedRCL [38], FedACG [20], FedPAQ [37], and FedHQ+ [5]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.634, + 0.727, + 0.65 + ], + "angle": 0, + "content": "4.2. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.657, + 0.907, + 0.823 + ], + "angle": 0, + "content": "We follow most of the implementation setups and evaluation protocols from [1, 20, 38, 48]. The ResNet-18 architecture [11] is adopted as the backbone network. Consistent with standard practice in FL [12], all BN [15] layers in ResNet-18 are replaced with GN [47] layers. Following the recommendation of [35], WS is applied before each GN layer to achieve optimal performance. The hyper-parameter of WS is set to \\(\\rho = 0.001\\), and the momentum for updating the scaling vector is set to \\(\\beta = 0.1\\). Additional information on the hyper-parameter selection is provided in Section C of the supplementary document." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.845, + 0.85 + ], + "angle": 0, + "content": "4.3. Comparison with existing FL methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Table 1 presents the performance comparison of the proposed FedWSQ against various existing FL methods on the three benchmark datasets. The bit-width settings for each" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.957 + ], + "angle": 0, + "content": "4621" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 2. Comparison of UQ and DANUQ methods with or without WS on the three benchmark datasets with a \\(5\\%\\) participation rate over 100 clients for \\(\\alpha \\in \\{0.1, 0.3, 0.6\\}\\)." + }, + { + "type": "table", + "bbox": [ + 0.214, + 0.122, + 0.785, + 0.369 + ], + "angle": 0, + "content": "
DatasetWS#bitsα = 0.1α = 0.3α = 0.6i.i.d.
UQDANUQUQDANUQUQDANUQUQDANUQ
CIFAR-10X157.8272.4871.2780.8872.3683.5279.4686.92
262.3965.1974.3578.6176.7982.5982.2586.11
466.8172.8778.5284.6281.3686.2685.4688.66
✓175.4884.7980.1588.6282.5389.4985.3791.15
278.3685.9283.0288.9884.1090.3586.9591.50
482.7886.8486.3789.8487.5891.1189.8792.21
CIFAR-100X130.0740.5134.1646.6238.2448.6737.8548.55
233.5240.5837.6645.0740.4746.3140.9046.06
438.0643.5942.4248.3343.4649.2944.3548.76
✓145.5962.0550.1065.1452.1366.1954.5866.11
249.6863.3454.2265.9656.0167.1958.7267.82
456.8564.3460.2167.3161.9868.4162.9868.53
Tiny-ImageNetX124.7428.9529.4833.1430.7436.7533.8138.95
227.0230.3331.6034.1233.5836.634.8735.41
428.5431.1431.9335.4233.5636.6634.9838.13
✓132.7345.1137.6349.9339.6051.1141.9851.77
236.1645.9640.9050.1742.5151.1244.8951.12
440.4646.5144.5850.4945.5551.9646.9352.13
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.379, + 0.483, + 0.408 + ], + "angle": 0, + "content": "Table 3. Comparison of NUQ methods using WS on CIFAR-10 and CIFAR-100 datasets for \\(\\alpha \\in \\{0.1, 0.3, 0.6\\}\\)." + }, + { + "type": "table", + "bbox": [ + 0.118, + 0.411, + 0.457, + 0.54 + ], + "angle": 0, + "content": "
#bitsMethodCIFAR-10CIFAR-100
α = 0.10.30.60.10.30.6
1NF/FP24.030.628.87.08.38.1
DANUQ84.888.689.562.165.166.2
2NF57.873.880.041.249.153.0
FP45.057.558.625.332.034.2
DANUQ85.989.090.463.366.067.2
4NF86.689.791.064.866.768.4
FP87.190.091.064.367.068.4
DANUQ86.889.891.164.367.368.4
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.552, + 0.483, + 0.825 + ], + "angle": 0, + "content": "FL method are indicated in the table. Compared to FL methods that use quantization, such as FedPAQ and FedHQ+, FedWSQ demonstrates superior performance for 4-bit and 1-bit settings. Also, FedWSQ outperforms FL methods with full-precision representation, including SOTA methods such as FedACG and FedRCL. Notably, even with a 1-bit representation, FedWSQ demonstrates competitive performance on CIFAR-10 and achieves superior results on CIFAR-100 and Tiny/ImageNet. For example, in a highly heterogeneous setting \\((\\alpha = 0.1)\\), FedWSQ (1-bit) achieves \\(62.05\\%\\) accuracy on CIFAR-100, surpassing FedRCL and FedACG by \\(3.79\\%\\) and \\(3.91\\%\\), respectively. Similarly, on Tiny/ImageNet, FedWSQ (1-bit) achieves \\(45.11\\%\\) accuracy, outperforming FedRCL and FedACG by \\(7.25\\%\\) and \\(5.36\\%\\), respectively. These results highlight the effectiveness of FedWSQ, which demonstrates its robustness even in extremely low bit-width scenarios while maintaining high model accuracy." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.279, + 0.851 + ], + "angle": 0, + "content": "4.4. Additional analyses" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Impact of bit representation in FedWSQ FedWSQ balances model performance and communication efficiency by adopting different bit-width settings. Table 1 presents the" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.381, + 0.903, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.474, + 0.907, + 0.504 + ], + "angle": 0, + "content": "Figure 3. Loss landscape of FL methods trained on CIFAR-100, where the numbers indicate the Hessian top eigenvalue." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.513, + 0.907, + 0.755 + ], + "angle": 0, + "content": "results for 32-bit (FedWS), 4-bit, and 1-bit representations, as well as adaptive bit-allocation strategies (FBA and DBA). Despite using lower bit-widths, FedWSQ remains robust in non-i.i.d. settings, particularly under high heterogeneity (\\(\\alpha = 0.1\\)). Since local models tend to overfit under high data heterogeneity, quantizing model updates helps to regularize overfitting, achieving more robust FL performance [2, 3]. The accuracy gaps among different bit-widths are relatively small, and thus the FBA and DBA strategies enable efficient communication with marginal performance loss. As shown in Table 1, both FBA and DBA strategies achieve performance close to the 4-bit setting while reducing communication costs by half. In summary, FedWSQ confirms its ability to maintain accuracy while significantly reducing communication bandwidth usage for real-world FL scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Effect of WS and DANUQ Table 2 shows a comparative analysis of our proposed DANUQ and UQ under varying bit-widths and data heterogeneity. In this experiment, UQ is implemented using the FedPAQ approach [37]. As listed in the table, our DANUQ consistently outperforms the UQ method. Under extremely low-bit quantization (1-bit and 2-bit settings), the UQ suffers from significant accuracy degradation. While WS alone improves performance, combining WS with the proposed DANUQ yields the high-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4622" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.105, + 0.089, + 0.468, + 0.103 + ], + "angle": 0, + "content": "Table 4. FL performance of different backbones for \\( \\alpha = {0.3} \\) ." + }, + { + "type": "table", + "bbox": [ + 0.127, + 0.107, + 0.442, + 0.184 + ], + "angle": 0, + "content": "
BackboneVanillaFedWSFedWSQ (DBA)
ShuffleNet [53]36.3753.3251.12
VGGNet-9 [40]46.1460.0160.48
SqueezeNet [14]39.4555.8756.10
ResNet-18 [11]47.3968.1966.32
MobileViT [32]35.4240.8041.27
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.192, + 0.482, + 0.329 + ], + "angle": 0, + "content": "est performance gains. The UQ with WS shows improvements compared with its non-WS counterparts, but the synergy between WS and DANUQ provides the most robust FL performance. Furthermore, our method maintains high accuracy even with low-bit representations. This suggests that applying FBA or DBA can be highly effective, as our approach maintains strong performance across varying bit-widths and makes FedWSQ particularly practical for real-world FL deployments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.334, + 0.483, + 0.531 + ], + "angle": 0, + "content": "Comparison with other NUQ methods We evaluate the effectiveness of the proposed DANUQ by comparing it with existing NUQ approaches, FP [45] and NF [7].7 As shown in Table 3, at the 4-bit representation, all NUQ methods perform competitively, indicating that NUQ is generally robust at relatively higher bit-widths. However, under 1-bit and 2-bit representations, FP and NF suffer from severe performance degradation, particularly in highly heterogeneous data (\\(\\alpha = 0.1\\)). In contrast, our DANUQ outperforms both NF and FP across all conditions, which confirms its robustness against extreme quantization constraints. These results highlight the superiority of our DANUQ method among NUQ methods for communication-efficient FL." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.537, + 0.483, + 0.749 + ], + "angle": 0, + "content": "Loss Landscape Analysis Figure 3 visualizes the loss landscapes of various FL methods compared to FedWSQ (with DBA). FedWSQ achieves the smoothest and most stable loss landscape, with the lowest Hessian top eigenvalue of 135.8, indicating improved generalization ability. FedSmoo exhibits a flatter landscape but converges to a much higher global minimum, reducing sharpness at the cost of performance. In contrast, FedACG reaches a global comparable to FedWSQ but has a significantly sharper landscape, with a Hessian top eigenvalue of 671.2, which indicates it is more sensitive to perturbations. These results show that FedWSQ effectively enhances the generalization, leading to more stable training and improved FL performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.759, + 0.483, + 0.865 + ], + "angle": 0, + "content": "Impact on different backbone architectures Table 4 presents the performance of different backbone architectures on CIFAR-100 to evaluate the impact of FedWS and FedWSQ. For all architectures, FedWS significantly improves accuracy over the vanilla models, and FedWSQ maintains a comparable performance to FedWS with improved communication efficiency. In addition, We eval" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.905, + 0.118 + ], + "angle": 0, + "content": "Table 5. Ablation study for \\( \\rho \\) using FedWS on CIFAR-10. For the non-i.i.d. setting, \\( \\alpha = {0.3} \\) is used." + }, + { + "type": "table", + "bbox": [ + 0.541, + 0.122, + 0.874, + 0.166 + ], + "angle": 0, + "content": "
ρ1 × 10-41 × 10-31 × 10-21 × 10-1
non-i.i.d.87.1589.7189.6289.46
i.i.d.90.6492.4891.9792.11
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.175, + 0.905, + 0.221 + ], + "angle": 0, + "content": "uate MobileViT [32], a vision transformer-based architecture, which also benefits from FedWS and FedWSQ. These results confirm their applicability beyond CNNs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.23, + 0.907, + 0.381 + ], + "angle": 0, + "content": "Comparison with FedWon We provide the comparison of FedWSQ with FedWon, which was discussed in Section 3.1. As shown in Table 1, FedWSQ substantially outperforms FedWon across all datasets and data heterogeneity settings. As illustrated in Figure 3, FedWon exhibits a much higher global minimum and a very sharp loss landscape with a Hessian top eigenvalue of 203.4, whereas FedWSQ maintains a smoother surface with an eigenvalue of 135.8. These results demonstrate FedWSQ's superior robustness and generalization ability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.39, + 0.909, + 0.542 + ], + "angle": 0, + "content": "Hyper-parameters We examine the hyper-parameter \\(\\rho\\) for WS, which handles the normalization scales. To evaluate its impact, we vary the values of \\(\\rho\\) and assess the performance of FedWS under both non-i.i.d. and i.i.d. data distributions. As shown in Table 5, the performance of FedWS is insensitive to \\(\\rho\\) variations. This is because during inference, normalization layers like GN cancel the effect of constant scaling, and the variance of the parameter vectors within the same convolutional layer is not significant. We select \\(\\rho = 1 \\times 10^{-3}\\), which yields the best results." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.557, + 0.64, + 0.573 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.907, + 0.826 + ], + "angle": 0, + "content": "This paper presents FedWSQ, a novel FL framework using WS and the proposed DANUQ to address key FL challenges, including data heterogeneity, partial client participation, and communication bottleneck. By leveraging WS, FedWSQ enhances training stability via an implicit gradient filtering mechanism. Additionally, DANUQ efficiently compresses LMPUs while preserving essential local information, significantly reducing communication overhead even at ultra-low bit-widths. Extensive experiments on the benchmark datasets confirm that FedWSQ consistently outperforms SOTA FL methods under various conditions. Furthermore, the mixed-precision strategies FBA and DBA effectively balance communication cost and model accuracy. These results validate the robustness of the FedWSQ framework in extreme quantization scenarios, making it highly practical for real-world FL." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Acknowledgments This work was supported by the National Research Foundation (NRF), South Korea, under project BK21 FOUR (Smart Robot Convergence and Application Education Research Center)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "\\( {}^{7} \\) QLs of different methods are visualized in Section C of the supplementary document." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4623" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.17 + ], + "angle": 0, + "content": "[1] Durmus Alp Emre Acar, Yue Zhao, Ramon Matas, Matthew Mattina, Paul Whatmough, and Venkatesh Saligrama. Federated learning based on dynamic regularization. In ICLR, 2021. 2, 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.241 + ], + "angle": 0, + "content": "[2] MohammadHossein AskariHemmat, Reyhane Askari Hemmat, Alex Hoffman, Ivan Lazarevich, Ehsan Saboori, Olivier Mastropietro, Sudhakar Sah, Yvon Savaria, and Jean-Pierre David. QReg: On regularization effects of quantization. arXiv preprint arXiv:2206.12372, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.242, + 0.483, + 0.311 + ], + "angle": 0, + "content": "[3] MohammadHossein AskariHemmat, Ahmadreza Jeddi, Reyhane Askari Hemmat, Ivan Lazarevich, Alexander Hoffman, Sudhakar Sah, Ehsan Saboori, Yvon Savaria, and Jean-Pierre David. QGen: On the ability to generalize in quantization aware training. arXiv preprint arXiv:2404.11769, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.312, + 0.483, + 0.379 + ], + "angle": 0, + "content": "[4] Debraj Basu, Deepesh Data, Can Karakus, and Suhas N Diggavi. Qsparse-Local-SGD: Distributed SGD with quantization, sparsification, and local computations. IEEE Journal on Selected Areas in Information Theory, 1(1):217-226, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.381, + 0.483, + 0.436 + ], + "angle": 0, + "content": "[5] Shengbo Chen, Cong Shen, Lanxue Zhang, and Yuanmin Tang. Dynamic aggregation for heterogeneous quantization in federated learning. IEEE Transactions on Wireless Communications, 20(10):6804-6819, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.438, + 0.483, + 0.492 + ], + "angle": 0, + "content": "[6] Jungwook Choi, Zhuo Wang, Swagath Venkataramani, Pierce I-Jen Chuang, Vijayalakshmi Srinivasan, and Kailash Gopalakrishnan. Pact: Parameterized clipping activation for quantized neural networks. In ICLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.494, + 0.483, + 0.522 + ], + "angle": 0, + "content": "[7] T. Dettmers, M. Lewis, S. Shleifer, and L. Zettlemoyer. 8-bit optimizers via block-wise quantization. In ICLR, 2022. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.523, + 0.483, + 0.563 + ], + "angle": 0, + "content": "[8] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: Efficient finetuning of quantized llms. In NeurIPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.483, + 0.606 + ], + "angle": 0, + "content": "[9] Steven K. Esser, Jeffrey L. McKinstry, Deepika Bablani, Rathinakumar Appuswamy, and Dharmendra S. Modha. Learned Step Size Quantization. In ICLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.607, + 0.483, + 0.647 + ], + "angle": 0, + "content": "[10] Robert M. Gray and David L. Neuhoff. Quantization. IEEE transactions on information theory, 44(6):2325-2383, 1998. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.649, + 0.483, + 0.69 + ], + "angle": 0, + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6, 8, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.692, + 0.483, + 0.733 + ], + "angle": 0, + "content": "[12] Kevin Hsieh, Amar Phanishayee, Onur Mutlu, and Phillip Gibbons. The non-IID data quagmire of decentralized machine learning. In ICML, 2020. 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.734, + 0.483, + 0.788 + ], + "angle": 0, + "content": "[13] Tzu-Ming Harry Hsu, Hang Qi, and Matthew Brown. Measuring the effects of non-identical data distribution for federated visual classification. arXiv preprint arXiv:1909.06335, 2019. 2, 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[14] Forrest N Iandola, Song Han, Matthew W Moskewicz, Khalid Ashraf, William J Dally, and Kurt Keutzer. SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and \\(< 0.5\\) MB model size. arXiv preprint arXiv:1602.07360, 2016.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[15] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015. 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[16] Benoit Jacob, Skirmantas Kligys, Bo Chen, Menglong Zhu, Matthew Tang, Andrew Howard, Hartwig Adam, and Dmitry Kalenichenko. Quantization and training of neural networks for efficient integer-arithmetic-only inference. In CVPR, pages 2704-2713, 2018. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[17] Sangil Jung, Changyong Son, Seohyung Lee, Jinwoo Son, Youngjun Kwak, Jae-Joon Han, Sung Ju Hwang, and Changkyu Choi. Learning to quantize deep networks by optimizing quantization intervals with task loss. In CVPR, pages 4350-4359, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.234, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[18] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank J Reddi, Sebastian U Stich, and Ananda Theertha Suresh. SCAFFOLD: Stochastic controlled averaging for ondevice federated learning. In ICML, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.291, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[19] Ahmed Khaled, Konstantin Mishchenko, and Peter Richtárik. First analysis of local GD on heterogeneous data. arXiv preprint arXiv:1909.04715, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[20] Geeho Kim, Jinkyu Kim, and Bohyung Han. Communication-efficient federated learning with accelerated client gradient. In CVPR, 2024. 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[21] Jinkyu Kim, Geeho Kim, and Bohyung Han. Multi-level branched regularization for federated learning. In ICML, 2022. 2, 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.419, + 0.905, + 0.474 + ], + "angle": 0, + "content": "[22] Jakub Konečny, H. Brendan McMahan, Felix X. Yu, Ananda Theertha Suresh, Dave Bacon, and Peter Richtárik. Federated learning: Strategies for improving communication efficiency. In arXiv preprint, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[23] Alex Krizhevsky. Learning multiple layers of features from tiny images. Master's thesis, Department of Computer Science, University of Toronto, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.585 + ], + "angle": 0, + "content": "[24] Khiem Le, Nhan Luong-Ha, Manh Nguyen-Duc, Danh Le-Phuoc, Cuong Do, and Kok-Seng Wong. Exploring the practicality of federated learning: A survey towards the communication perspective. arXiv preprint arXiv:2405.20431, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.589, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[25] Ya Le and Xuan Yang. Tiny ImageNet visual recognition challenge. CS 231N, 7(7):3, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.618, + 0.905, + 0.671 + ], + "angle": 0, + "content": "[26] Gihun Lee, Minchan Jeong, Yongjin Shin, Sangmin Bae, and Se-Young Yun. Preservation of the global knowledge by not-true distillation in federated learning. In NeurIPS, 2022. 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.674, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[27] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. In MLSys, 2020. 2, 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.718, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[28] Xiang Li, Kaixuan Huang, Wenhao Yang, Shusen Wang, and Zhihua Zhang. On the convergence of FedAvg on non-IID data. In ICLR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.76, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[29] Darryl Lin, Sachin Talathi, and Sreekanth Annapureddy. Fixed point quantization of deep convolutional networks. In International conference on machine learning, pages 2849-2858. PMLR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[30] Wang Luping, Wang Wei, and Li Bo. CMFL: Mitigating communication overhead for federated learning. In 2019 IEEE 39th international conference on distributed computing systems (ICDCS), pages 954-964. IEEE, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[31] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4624" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "efficient learning of deep networks from decentralized data. In AISTATS, 2017. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.123, + 0.483, + 0.164 + ], + "angle": 0, + "content": "[32] Sachin Mehta and Mohammad Rastegari. Mobilevit: lightweight, general-purpose, and mobile-friendly vision transformer. arXiv preprint arXiv:2110.02178, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.167, + 0.482, + 0.222 + ], + "angle": 0, + "content": "[33] Markus Nagel, Marios Fournarakis, Rana Ali Amjad, Yelysei Bondarenko, Mart van Baalen, and Tijmen Blankevoort. A white paper on neural network quantization. arXiv preprint arXiv:2106.08295, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.224, + 0.482, + 0.293 + ], + "angle": 0, + "content": "[34] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. PyTorch: An imperative style, high-performance deep learning library. In NeurIPS, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.296, + 0.482, + 0.351 + ], + "angle": 0, + "content": "[35] Siyuan Qiao, Huiyu Wang, Chenxi Liu, Wei Shen, and Alan Yuille. Micro-batch training with batch-channel normalization and weight standardization. arXiv preprint arXiv:1903.10520, 2019. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.354, + 0.482, + 0.408 + ], + "angle": 0, + "content": "[36] Sashank J Reddi, Zachary Charles, Manzil Zaheer, Zachary Garrett, Keith Rush, Jakub Konečny, Sanjiv Kumar, and Hugh Brendan McMahan. Adaptive federated optimization. In ICLR, 2021. 2, 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.412, + 0.482, + 0.478 + ], + "angle": 0, + "content": "[37] Amirhossein Reisizadeh, Aryan Mokhtari, Hamed Hassani, Ali Jabbabaie, and Ramtin Pedarsani. FedPAQ: A communication-efficient federated learning method with periodic averaging and quantization. In AISTATS, 2020. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.482, + 0.482, + 0.523 + ], + "angle": 0, + "content": "[38] Seonguk Seo, Jinkyu Kim, Geeho Kim, and Bohyung Han. Relaxed contrastive learning for federated learning. In CVPR, 2024. 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.527, + 0.482, + 0.58 + ], + "angle": 0, + "content": "[39] Yujun Shi, Jian Liang, Wenqing Zhang, Vincent YF Tan, and Song Bai. Towards understanding and mitigating dimensional collapse in heterogeneous federated learning. In ICLR, 2023. 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.585, + 0.482, + 0.625 + ], + "angle": 0, + "content": "[40] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2014. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.629, + 0.482, + 0.655 + ], + "angle": 0, + "content": "[41] Sebastian U Stich. Local SGD converges fast and communicates little. In ICLR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.659, + 0.482, + 0.713 + ], + "angle": 0, + "content": "[42] Sebastian U Stich and Sai Praneeth Karimireddy. The errorfeedback framework: Better rates for SGD with delayed gradients and compressed communication. arXiv preprint arXiv:1909.05350, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.716, + 0.482, + 0.785 + ], + "angle": 0, + "content": "[43] Yan Sun, Li Shen, Shixiang Chen, Liang Ding, and Dacheng Tao. Dynamic regularized sharpness aware minimization in federated learning: Approaching global consistency and smooth landscape. In International Conference on Machine Learning, pages 32991-33013. PMLR, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.482, + 0.842 + ], + "angle": 0, + "content": "[44] Jianyu Wang and Gauri Joshi. Cooperative SGD: A unified framework for the design and analysis of local-update SGD algorithms. Journal of Machine Learning Research, 22(213): 1-50, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[45] Ruizhe Wang, Yeyun Gong, Xiao Liu, Guoshuai Zhao, Ziyue Yang, Baining Guo, Zhengjun Zha, and Peng Cheng. Optimizing large language model training using fp4 quantization. In arXiv preprint arXiv:2501.17116, 2025. 4, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[46] Jie Wen, Zhixia Zhang, Yang Lan, Zhihua Cui, Jianghui Cai, and Wensheng Zhang. A survey on federated learning: challenges and applications. International Journal of Machine Learning and Cybernetics, 14(2):513-535, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.175 + ], + "angle": 0, + "content": "[47] Yuxin Wu and Kaiming He. Group normalization. In ECCV, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[48] Jing Xu, Sen Wang, Liwei Wang, and Andrew Chi-Chih Yao. FedCM: Federated learning with client-level momentum. arXiv preprint arXiv:2106.10874, 2021. 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[49] Haibo Yang, Minghong Fang, and Jia Liu. Achieving linear speedup with partial worker participation in non-IID federated learning. In ICLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[50] Hao Yu, Sen Yang, and Shenghuo Zhu. Parallel restarted SGD with faster convergence and less communication: Demystifying why model averaging works for deep learning. In AAAI, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[51] Dongqing Zhang, Jiaolong Yang, Dongqiangzi Ye, and Gang Hua. LQ-Nets: Learned Quantization for Highly Accurate and Compact Deep Neural Networks. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[52] Jie Zhang, Zhiqi Li, Bo Li, Jianghe Xu, Shuang Wu, Shouhong Ding, and Chao Wu. Federated learning with label distribution skew via logits calibration. In ICML, 2022. 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[53] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. ShuffleNet: An extremely efficient convolutional neural network for mobile devices. In CVPR, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.905, + 0.49 + ], + "angle": 0, + "content": "[54] Sijie Zhao, Tao Yue, and Xuemei Hu. Distribution-aware adaptive multi-bit quantization. In CVPR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.492, + 0.905, + 0.519 + ], + "angle": 0, + "content": "[55] Chenzhuo Zhu, Song Han, Huizi Mao, and William J. Dally. Trained ternary quantization. In ICLR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.521, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[56] Weiming Zhuang and Lingjuan Lyu. FedWon: Triumphing multi-domain federated learning without normalization. In ICLR, 2023. 4, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4625" + } + ] +] \ No newline at end of file diff --git a/2025/FedWSQ_ Efficient Federated Learning with Weight Standardization and Distribution-Aware Non-Uniform Quantization/1f166dee-20e5-452e-8e86-86019ce8d0bb_origin.pdf b/2025/FedWSQ_ Efficient Federated Learning with Weight Standardization and Distribution-Aware Non-Uniform Quantization/1f166dee-20e5-452e-8e86-86019ce8d0bb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d0bfb6e31214afd3018b045a46a2a5d473f32ed7 --- /dev/null +++ b/2025/FedWSQ_ Efficient Federated Learning with Weight Standardization and Distribution-Aware Non-Uniform Quantization/1f166dee-20e5-452e-8e86-86019ce8d0bb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c805d6170e6343b4fe2fb75c4d63730f9c9a218d06d403602e498515b63a8b3a +size 1203406 diff --git a/2025/FedWSQ_ Efficient Federated Learning with Weight Standardization and Distribution-Aware Non-Uniform Quantization/full.md b/2025/FedWSQ_ Efficient Federated Learning with Weight Standardization and Distribution-Aware Non-Uniform Quantization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0725c734caa461909775473cca6ec1c90c65ca03 --- /dev/null +++ b/2025/FedWSQ_ Efficient Federated Learning with Weight Standardization and Distribution-Aware Non-Uniform Quantization/full.md @@ -0,0 +1,346 @@ +# FedWSQ: Efficient Federated Learning with Weight Standardization and Distribution-Aware Non-Uniform Quantization + +Seung-Wook Kim\*1, Seongyeol Kim\*1, Jiah Kim\*1, Seowon Ji\*2, and Se-Ho Lee\*3 + +$^{1}$ Department of Intelligent Robot Engineering, Pukyong National University + +$^{2}$ CSE, Konkuk University $^{3}$ CS&AI, Jeonbuk National University + +# Abstract + +Federated learning (FL) often suffers from performance degradation due to key challenges such as data heterogeneity and communication constraints. To address these limitations, we present a novel FL framework called FedWSQ, which integrates weight standardization (WS) and the proposed distribution-aware non-uniform quantization (DANUQ). WS enhances FL performance by filtering out biased components in local updates during training, thereby improving the robustness of the model against data heterogeneity and unstable client participation. In addition, DANUQ minimizes quantization errors by leveraging the statistical properties of local model updates. As a result, FedWSQ significantly reduces communication overhead while maintaining superior model accuracy. Extensive experiments on FL benchmark datasets demonstrate that FedWSQ consistently outperforms existing FL methods across various challenging FL settings, including extreme data heterogeneity and ultra-low-bit communication scenarios. The source code is available at our project page1. + +# 1. Introduction + +In large-scale machine learning, centralized approaches often raise privacy concerns because sensitive data from edge devices must be collected on a central server for model training. Federated learning (FL) [31] enables distributed devices to collaboratively train a shared global model without sharing raw data. A fundamental FL method, FedAvg, was proposed by McMahan et al. [31]. In this approach, each client trains its model using the local data, and then the server collects and averages these local updates to form a unified global model. + +Several studies on FL have verified its practical effectiveness [4, 28, 41, 42, 44, 49, 50]. However, real-world FL methods often face three key challenges [18, 19, 24, + +30, 46]: 1) data heterogeneity, where clients possess non-independent and identically distributed (non-i.i.d.) data; 2) partial client participation, where only a subset of clients contributes to a global model update during each training round due to communication constraints; 3) communication bottlenecks, where limited communication bandwidth and high transmission costs hinder efficient model aggregation. These challenges worsen local gradient divergence, slow down the global model convergence, and ultimately degrade FL performance in real-world applications. + +To address these limitations, we propose FedWSQ, an advanced FL framework that combines weight standardization (WS) with our proposed distribution-aware nonuniform quantization (DANUQ) to improve both learning stability and communication efficiency. WS is a plug-and-play technique that standardizes the weight vectors of convolutional or linear layers to stabilize the learning process of a neural network [35]. In FedWSQ, WS plays a crucial role in mitigating client drift by filtering out gradient components that contribute to overfitting during local training. This leads to improved generalization across heterogeneous clients and significantly enhances FL performance. Although WS accelerates global convergence, the communication cost per round remains a major bottleneck. To address this, FedWSQ integrates WS with DANUQ, a novel quantization strategy that leverages a standard normal distribution prior to minimize quantization errors, reducing communication overhead while preserving model performance. Even under ultra-low-bit quantization, the proposed FedWSQ achieves superior performance, consistently outperforming state-of-the-art (SOTA) FL methods across various datasets and FL settings. + +Our main contributions are summarized as follows. + +- We propose an effective application of WS in FL, enhancing model convergence and stability under data heterogeneity and limited client participation. By implicitly performing a gradient filtering process, WS mitigates client drift while also providing a regularization effect that stabilizes training and improves generalization. +- We introduce DANUQ, a novel quantization method that + +employs a fixed quantization function, leveraging the statistical properties of local model parameter updates (LMPUs) to minimize quantization errors under parameter-distribution priors. By efficiently compressing LM-PUs while preserving essential information, the proposed DANUQ significantly reduces communication overhead without noticeable performance degradation. + +- Extensive experiments show that FedWSQ significantly enhances performance across diverse settings. Even with an average of 2.3 bits per parameter, FedWSQ improves performance by over $5\%$ on Tiny-ImageNet in a highly heterogeneous setting compared to SOTA FL methods. + +# 2. Related Work and Preliminaries + +Federated learning (FL) FL is a decentralized approach that enables collaborative training of a global model. Using distributed devices, FL ensures efficient learning while preserving data privacy. Let $\mathcal{F}_n(\mathbf{w}_{\mathrm{g}})\coloneqq \mathbb{E}_{\zeta_n\sim \mathcal{D}_n}[\mathcal{F}_n(\mathbf{w}_{\mathrm{g}};\zeta_n)]$ be the local loss function of the client $n\in [N]$ with a local data distribution $\mathcal{D}_n$ , where $\mathbf{w}_{\mathrm{g}}$ is the global model parameter (GMP) vector. The training objective function of FL [31] is formulated as follows: + +$$ +\min _ {\mathbf {w} _ {\mathrm {g}}} \left\{\mathcal {F} \left(\mathbf {w} _ {\mathrm {g}}\right) := \sum_ {i \in \mathcal {S}} h _ {i} \mathcal {F} _ {i} \left(\mathbf {w} _ {\mathrm {g}}\right) \right\}, \tag {1} +$$ + +where $h_i$ is the weight assigned to the $i$ -th device, such that $h_i \geq 0$ , $\sum_{i} h_i = 1$ , and $h_i \propto |\mathcal{D}_i|$ , and $S \subseteq [N]$ is a subset of participating clients in the current communication round. At each round $t \in [T]$ , the server broadcasts the GMPs $\mathbf{w}_{\mathrm{g}}^{t-1}$ to the participating client $i$ . In FedAvg [31], the $i$ -th client performs $K$ steps of local training to obtain the local model parameter (LMP) vector, $\mathbf{w}_i^t$ , and transmits its LMP updates (LMPUs), $\Delta \mathbf{w}_i^t := \mathbf{w}_i^t - \mathbf{w}_{\mathrm{g}}^{t-1}$ , to the server. Finally, the server aggregates these LMPUs through weighted averaging to derive the updated GMPs. + +In real-world FL scenarios, unstable learning dynamics and slow convergence often occur due to non-i.i.d. data distributions and limited communication bandwidth. To address these issues, FedProx [27] penalizes LMPUs by adding a proximal term in the local objective to reduce the gap between the global and local loss functions. FedAvgM [13] integrates momentum into the global update process to mitigate variance in model aggregation. FedAdam [36] introduces an adaptive optimizer that improves global and local model convergence, thereby accelerating overall training progress. FedDyn [1] presents dynamic regularization for each device to diminish deviations in the LMPUs caused by data heterogeneity. FedSmoo [43] applies constraints that prevent sharp fluctuations in local updates to generalize FL performance by ensuring global consistency. Recently, FedACG [21] was proposed to improve convergence by broadcasting a global model with a lookahead gradient and aligning local models with the shared global model. + +Network quantization As neural network architectures have grown larger in scale, quantization has become essential for the memory optimization and acceleration of operations [9, 10, 16, 17, 29, 33, 55]. Let us consider a model using $B$ -bit representation. A typical quantization function $\mathcal{Q}$ that maps a full-precision value $x$ to a quantization level (QL) $q$ can be written as + +$$ +q = \mathcal {Q} (x / s), \tag {2} +$$ + +where $s$ is a scaling factor, and each QL is matched with a $B$ -bit integer from $\{0,\dots ,2^B -1\}$ . The primary goal of effective quantization is to determine the proper scaling factor and a set of QLs to minimize the quantization errors. + +Quantization methods can be roughly categorized into two approaches: uniform quantization (UQ) and nonuniform quantization (NUQ). UQ divides the input range into $2^{B}$ equal intervals and assigns a representative value as a QL to each interval [10, 29]. In contrast, NUQ adjusts the interval widths based on the prior data distribution [6, 8, 51]. By allocating finer intervals in dense data regions and broader intervals in sparse regions, NUQ preserves critical information more effectively than UQ. For the scaling factor $s$ , both approaches commonly employ the absmax method [16, 33] to determine the dynamic range for quantization. Furthermore, to enhance data adaptability, researchers have explored methods for dynamically learning quantization parameters (e.g., scaling factors and QLs) [9, 17, 55]. During backpropagation, these parameters are updated to reduce quantization error and better align with the underlying data distribution. + +Quantized FL In FL, communication bottlenecks arise primarily during uplink transmission, as each participating client must send its individual LMPUs to the central server. Unlike downlink transmission, where a single GMP is broadcast to multiple clients, uplink communication involves substantial data transfer. This problem is exacerbated in resource-constrained communication environments, such as mobile and IoT networks, where upload bandwidth is severely limited [5, 37]. To overcome this challenge, several studies have focused on compressing model updates. In [22], quantization and sparsification are combined to reduce the number of transmitted parameters. FedPAQ [37] employs a probabilistic rounding algorithm to prevent an excessive concentration of identical QLs. Meanwhile, FedHQ+ [5] assigns different weight factors to clients for global updates based on their individual quantization errors. Although these methods significantly improve communication efficiency, they often suffer from performance degradation under extremely low bit-width conditions. In this paper, we introduce FedWSQ, a novel quantization framework that is designed to maintain robust model performance even at ultra-low bit-width. + +![](images/4e12e521430890cc82b9f7d38086e2fc6a0e6165afd8de62ccb3dd0637e6a8c1.jpg) + +![](images/4b655be1c8fec3b0c32d6a8e3deed4f5a7bf375616742c6b55241bbf9874effd.jpg) +Figure 1. The overall framework of FedWSQ, which integrates WS and DANUQ for efficient FL. Clients transmit LMPUs quantized by the DANUQ along with scaling vectors to the server, where dequantization and aggregation are performed to update the global parameters. + +# 3. Proposed Method + +Figure 1 illustrates the overall framework of the proposed FedWSQ, which employs WS and DANUQ for efficient FL. Each client has a standard neural network model consisting of multiple model blocks, where WS is applied to mitigate the impact of data heterogeneity between clients. After local training, the full-precision values of LMPUs are converted to a $B$ -bit representation using the global scaling vectors and carefully designed QLs. These quantized LMPUs and their corresponding scaling vectors are transmitted to the server. The server then dequantizes the received data to higher-precision values and aggregates them to update the GMPs. In addition, the global scaling vectors are refined to ensure consistent scaling across different clients for DANUQ. Using the proposed DANUQ, FedWSQ enhances communication efficiency without sacrificing model performance in challenging FL scenarios. + +# 3.1. Weight standardization (WS) + +WS [35] standardizes each weight vector of a convolution or linear layer to enhance the learning process. By ensuring consistent parameter distributions during training, WS contributes to a more stable gradient flow. Beyond this general benefit, we argue that the use of WS in FL can address the challenges posed by the distributed nature of FL and data heterogeneity. Formally, consider the $n$ -th client that has a network architecture consisting of $L$ linear or convolutional layers. Let the $l$ -th layer (without bias) map an input $\mathbf{x}_l \in \mathbb{R}^{I_l}$ to an output $\mathbf{y}_l \in \mathbb{R}^{O_l}$ , defined by + +$$ +\begin{array}{l} \mathbf {y} _ {l} = \mathbf {W} _ {n, l} ^ {T} \mathbf {x} _ {l} \tag {3} \\ := \left[ \begin{array}{c c c c} \mathbf {w} _ {n, m _ {1} ^ {l}} & \mathbf {w} _ {n, m _ {2} ^ {l}} & \dots & \mathbf {w} _ {n, m _ {O _ {l}} ^ {l}} \end{array} \right] ^ {T} \mathbf {x} _ {l}, \qquad (4) \\ \end{array} +$$ + +where $\mathbf{w}_{n,m}$ , for $m\in \mathcal{M}_l\coloneqq \{m_1^l,\ldots ,m_{O_l}^l\}$ , is the weight vector. In WS, each $\mathbf{w}_{n,m}$ is standardized before being used within the layer. We denote the prestandardized parameter (PSP) vector as $\mathbf{w}_{n,m}$ , and the weight-standardized parameter (WSP) vector as $\tilde{\mathbf{w}}_{n,m}$ , respectively. Then, the WSP vector $\tilde{\mathbf{w}}_{n,m}$ is obtained by + +$$ +\tilde {\mathbf {w}} _ {n, m} = \frac {\rho}{\sigma (\mathbf {w} _ {n , m})} \left(\mathbf {I} - \mathbf {P _ {1}}\right) \mathbf {w} _ {n, m}, \qquad (5) +$$ + +where $\mathbf{1} \in \mathbb{R}^{I_l}$ is a vector of all ones, $\sigma(\mathbf{w}_{n,m})$ is the standard deviation of the elements in $\mathbf{w}_{n,m}$ , and $\mathbf{P_v}$ represents the projection matrix onto the subspace spanned by the vector $\mathbf{v}$ , which is denoted by $\mathrm{span}\{\mathbf{v}\}$ . Here, we introduce a hyper-parameter $\rho$ to control the normalization scale. Note that we reformulate the WSP computation from [35] into Eq. (5) to better understand the gradient filtering process of WS during local training. + +Gradient filtering by WS WS plays an important role in reducing the learning diversity of local models, which is one of the most critical issues in FL. To understand its impact, we compare the gradient of a loss function $\mathcal{L}$ with respect to the WSP vector $\tilde{\mathbf{w}}_{n,m}$ to that of the PSP vector $\mathbf{w}_{n,m}$ . When WS is applied, the gradient is processed by an additional series of projections to remove specific components that may interfere with effective FL. Specifically, the gradient with respect to the PSP vector $\mathbf{w}_{n,m}$ is obtained by + +$$ +\frac {\partial \mathcal {L}}{\partial \mathbf {w} _ {n , m}} = \frac {\rho}{\sigma (\mathbf {w} _ {n , m})} \left(\mathbf {I} - \mathbf {P} _ {\mathbf {1}}\right) \left(\mathbf {I} - \mathbf {P} _ {\tilde {\mathbf {w}} _ {n, m}}\right) \frac {\partial \mathcal {L}}{\partial \tilde {\mathbf {w}} _ {n , m}}. (6) +$$ + +![](images/c5d4cedad4a30261ec4674529f88523fee3a26460aa03e52dbe3e1a257b88a7b.jpg) +Figure 2. An example of gradient filtering process of WS. + +The upstream gradient $\frac{\partial\mathcal{L}}{\partial\tilde{\mathbf{w}}_{n,m}}$ is projected onto two subspaces, $\mathrm{span}\{\tilde{\mathbf{w}}_{n,m}\}^{\perp}$ and $\mathrm{span}\{\mathbf{1}\}^{\perp}$ , where $\mathrm{span}\{\mathbf{v}\}^{\perp}$ represents the orthogonal complement of $\mathrm{span}\{\mathbf{v}\}$ . Since $\tilde{\mathbf{w}}_{n,m} \perp \mathbf{1}$ , the upstream gradient $\frac{\partial\mathcal{L}}{\partial\tilde{\mathbf{w}}_{n,m}}$ is ultimately projected onto $\mathrm{span}\{\tilde{\mathbf{w}}_{n,m}, \mathbf{1}\}^{\perp}$ after the two projections. + +As illustrated in Figure 2, Eq. (6) represents a gradient filtering process that applies successive projections to remove undesirable components from the gradient with respect to LMPs. In FL, data heterogeneity causes local models to overfit their local data, which leads to inconsistent updates across clients. This inconsistency results in a discrepancy between the GMPs and the optimal parameters, known as client drift. WS mitigates this issue by projecting the gradient onto $\mathrm{span}\{\tilde{\mathbf{w}}_{n,m},\mathbf{1}\}^{\perp}$ , which effectively reduces the impact of client drift during training. The first projection eliminates the gradient component aligned with $\tilde{\mathbf{w}}_{n,m}$ , which corresponds to the LMPs that are biased due to the non-i.i.d. data distribution. The second projection removes the mean component of mini-batch gradients, which can also be biased toward the local data distribution, compounding inconsistencies across clients. By filtering out both the parameter-aligned and mean components from the local gradient, WS preserves meaningful gradient directions essential for stable global model convergence. Additionally, this projection-based filtering introduces a regularization effect by constraining the model's deviation from its current state, which can improve the model's generalization ability. + +Comparison with FedWon It is worth noting that FedWon [56], a recently proposed FL method, also employs WS to address data heterogeneity in multi-domain FL. However, a key difference lies in the parameters exchanged between clients and the server, which fundamentally alters the learning dynamics. FedWon transmits WSP, which explicitly normalizes parameters across clients to improve learning stability. However, this approach forces all clients' LMPUs to exhibit statistically similar characteristics, which potentially discards client-specific information crucial for local adaptation. In contrast, FedWSQ transmits PSP, preserving essential local information while implicitly mitigating harmful divergences through the gradient filtering process of WS, as shown in Figure 2. This design choice en + +ables FedWSQ to achieve a better balance between global stability and local information preservation. An experimental comparison between FedWon and FedWSQ will be discussed in Section 4.4. + +# 3.2. Quantization of LMPUs + +Neural network quantization often involves learning quantization parameters. However, in FL, transmitting additional quantization parameters or auxiliary information for each communication round would introduce significant resource overhead. Instead, we propose a fixed quantization function that eliminates the need to transmit additional quantization information, ensuring a lightweight and efficient communication process. A fundamental assumption in NUQ is that model parameters follow a normal distribution, as demonstrated in prior studies [7, 45, 54]. Since LMPUs represent the difference between an updated LMP and the previous GMP, they may also exhibit a normal distribution. + +LMPU scaling Most quantization methods adopt the absmax strategy to manage the dynamic range of values [16, 33]. This approach scales an input tensor to the range $[-1,1]$ by dividing each element by the maximum absolute value of the tensor. Within this constrained range, conventional quantization methods determine QLs and map input values to the nearest quantization point. However, these methods are highly sensitive to outliers, which can excessively expand the dynamic range of the target tensor. Such over-scaling often results in underflow, especially in extremely low-bit representation. To solve this problem, we use the standard deviation of the input tensor as a scaling factor. Unlike the absmax value, the standard deviation is more robust to outliers. Moreover, since our DANUQ is based on a probability model of a standard normal distribution, the standard deviation provides a more reliable approximation of the target dynamic range. + +In FedWSQ, we use a shared global scaling factor to help maintain consistency in quantization. Instead of independently using scaling factors, we employ a collaborative approach in which each client contributes to a global scaling vector, ensuring that local standard deviations are taken into account. Specifically, each client $i$ computes a scaling vector, defined as $\mathbf{s}_i = [s_{i,1},\dots ,s_{i,L}]^T$ , where $s_{i,l}$ denotes the standard deviation of the LMPUs for the $l$ -th layer. The server then collects and aggregates $\mathbf{s}_i$ from the participating clients $i\in S$ , and updates the global scaling vector $\mathbf{s_g} = [s_{g,1},\dots ,s_{g,L}]^T$ as follows: + +$$ +\mathbf {s} _ {\mathrm {g}} \leftarrow (1 - \beta) \mathbf {s} _ {\mathrm {g}} + \beta \frac {1}{| \mathcal {S} |} \sum_ {i \in \mathcal {S}} \mathbf {s} _ {i}, \tag {7} +$$ + +where $\beta$ is a momentum parameter that controls the update rate. This approach ensures that the global scaling factor smoothly adapts while mitigating local fluctuations. Once + +updated, the server distributes $\mathbf{s}_{\mathrm{g}}$ to the participating clients. Each client then divides its LMPU updates for the $l$ -th layer by the corresponding scale $s_{\mathrm{g},l}$ , so that the normalized values can be regarded as samples from a standard normal distribution. Based on this assumption, we apply the proposed DANUQ scheme detailed in the next section. + +QLs considering normal distribution To determine the optimal multi-bit QLs, we formulate an optimization problem that minimizes the expected quantization error under a normal distribution. We assume that a random variable $\Delta w$ from the normalized LMPUs follows a standard normal distribution. Since the distribution is symmetric around zero, we consider only its nonnegative half. Given a $B$ -bit representation, let $Q = \{q_{0}, q_{1}, \ldots, q_{R}\}$ be the set of QLs, where $R = 2^{B - 1} - 1$ . Here, we assume that $q_{r}$ values are sorted in ascending order and set $q_{0} = 0$ as a fixed value. Then, the quantization boundaries are defined as + +$$ +u _ {r} = \left\{ \begin{array}{l l} q _ {0}, & \text {i f} r = 0, \\ \frac {q _ {r - 1} + q _ {r}}{2}, & \text {i f} 1 \leq r \leq R, \\ + \infty , & \text {i f} r = R + 1. \end{array} \right. \tag {8} +$$ + +A full-precision value in the range $[u_r, u_{r+1})$ is quantized into the QL $q_r$ . The optimal QLs are determined by minimizing the expected quantization error, formulated as + +$$ +\mathbb {E} \left[ \left(\Delta w - \Delta \bar {w}\right) ^ {2} \right] = \sum_ {r = 0} ^ {R} \int_ {u _ {r}} ^ {u _ {r + 1}} \left(x - q _ {r}\right) ^ {2} p (x) d x, \tag {9} +$$ + +where $\Delta \bar{w} \in Q$ represents the quantized value, and $p(x)$ is the PDF of the standard normal distribution. From (9), we can derive the following optimization problem as + +$$ +\begin{array}{l} \min _ {\left\{q _ {1}, \dots , q _ {R} \right\}} \left\{\frac {1}{2} \left(q _ {R} ^ {2} + 1\right) - \sqrt {\frac {2}{\pi}} q _ {1} e ^ {- \frac {u _ {1} ^ {2}}{2}} - \frac {1}{2} q _ {1} ^ {2} \operatorname {e r f} \left(\frac {u _ {1}}{\sqrt {2}}\right) \right. \\ + \sqrt {\frac {2}{\pi}} \sum_ {r = 1} ^ {R - 1} (q _ {r} - q _ {r + 1}) e ^ {- \frac {u _ {r + 1} ^ {2}}{2}} \\ + \frac {1}{2} \sum_ {r = 1} ^ {R - 1} \left(q _ {r} ^ {2} - q _ {r + 1} ^ {2}\right) \operatorname {e r f} \left(\frac {u _ {r + 1}}{\sqrt {2}}\right) \}, \tag {10} \\ \end{array} +$$ + +where $\operatorname{erf}(\cdot)$ denotes the error function. A closed-form solution to Eq. (22) is intractable due to the presence of nonlinear terms, including the Gaussian integral and the error function, which lack simple analytical inverses, and the interdependent QLs due to boundary conditions. As an alternative, we employ a brute-force search algorithm to numerically determine the optimal QLs. The optimal QLs + +# Algorithm 1 FedWSQ + +Input: Initial global parameters $\mathbf{W}_{\mathrm{g}}^{0}$ , initial scales $\mathbf{s}_{\mathrm{g}}^{0}$ , global communication round $T$ , local training iteration $K$ , number of total clients $N$ , number of model layers $L$ , learning rate $\eta$ , hyperparameter $\beta$ + +# Server-side: + +for $t = 1,\dots ,T$ do + +Randomly sample a subset of clients $S_{t} \subseteq [N]$ + +Set the bit-precision $B_{i}$ for the client $i \in S_{t}$ + +Send $(\mathbf{W}_{\mathrm{g}}^{t - 1},\mathbf{s}_{\mathrm{g}}^{t - 1})$ to client $i\in S_{t}$ + +for each client $i\in S_t$ , in parallel do + +$$ +\left(\Delta \overline {{\mathbf {W}}} _ {i} ^ {t}, \mathbf {s} _ {i} ^ {t}\right) \leftarrow \text {C l i e n t L o c a l T r a i n i n g} \left(\mathbf {W} _ {\mathrm {g}} ^ {t - 1}, \mathbf {s} _ {\mathrm {g}} ^ {t - 1}, i, B _ {i}\right) +$$ + +end + +for $i\in S_t$ do + +Obtain dequantized parameters $\Delta_i^t$ using $(\Delta \overline{\mathbf{W}}_i^t, \mathbf{s}_i^t)$ + +end + +Aggregate local updates: $\Delta^t\gets \sum_{i\in S_t}\Delta_i^t$ + +Update global parameters: $\mathbf{W}_{\mathrm{g}}^{t}\gets \mathbf{W}_{\mathrm{g}}^{t - 1} + \Delta^{t}$ + +Update global scales: $\mathbf{s}_{\mathrm{g}}^{t}\gets (1 - \beta)\mathbf{s}_{\mathrm{g}}^{t - 1} + \beta \frac{1}{|\mathcal{S}_{t}|}\sum_{i\in \mathcal{S}_{t}}\mathbf{s}_{i}^{t}$ + +end + +ClientLocalTraining $(\mathbf{W}_{\mathrm{g}},\mathbf{s}_{\mathrm{g}},i,I,B)$ + +Initialize local parameters: $\mathbf{W}_j^0\gets \mathbf{W}_{\mathrm{g}}$ + +Set DANUQ function $\mathcal{Q}(\cdot)$ based on the bit-precision $B$ + +for $k = 1,\ldots ,K$ do + +Apply WS to $\mathbf{W}_i^{k - 1}$ using (5) + +Compute unbiased gradient of local loss $\nabla f_{i}(\mathbf{W}_{i}^{k - 1})$ + +Update local parameters: $\mathbf{W}_i^k\gets \mathbf{W}_i^{k - 1} - \eta \nabla f_i(\mathbf{W}_i^{k - 1})$ + +end + +Compute local update: $\Delta \mathbf{W}_i\gets \mathbf{W}_i^K -\mathbf{W}_g$ + +for each layer $l = 1,\dots ,L$ do + +Compute quantized local update: $\Delta \overline{\mathbf{W}}_{i,l}\gets \mathcal{Q}(\Delta \mathbf{W}_{i,l} / s_{\mathrm{g},l})$ + +Calculate local scale factors: $s_{i,l} \gets \mathrm{std}(\Delta \mathbf{W}_{i,l})$ + +end + +$$ +\Delta \overline {{\mathbf {W}}} _ {i} := (\Delta \mathbf {W} _ {i, 1}, \dots , \Delta \mathbf {W} _ {i, L}) +$$ + +$$ +\mathbf {s} _ {i} := \left[ s _ {i, 1}, \dots , s _ {i, L} \right] +$$ + +$$ +\operatorname {r e t u r n} \left(\Delta \overline {{\mathbf {W}}} _ {i}, \mathbf {s} _ {i}\right) +$$ + +for different bit-widths are obtained as follows: 1) 1-bit $^6$ : [-0.798, 0.798]; 2) 2-bit: [-1.224, 0, 0.765, 1.724]; 3) 4-bit: [-2.654, -1.974, -1.508, -1.149, -0.834, -0.544, -0.269, 0, 0.230, 0.465, 0.708, 0.966, 1.248, 1.568, 1.968, 2.649]. These QLs minimize quantization error under the normal distribution, which provides an effective trade-off between precision and computational efficiency. + +Mixed channel-bit allocation To further enhance communication efficiency in FL, we propose an adaptive mixed-precision strategy in which each client dynamically selects its bit representation based on its communication channel conditions. This approach effectively balances model update quality and transmission cost, reducing overall bandwidth utilization. To validate this strategy, we conduct two simulation setups: 1) Fixed-bit allocation (FBA), where each client keeps a constant bit-width selected from $\{1,2,4\}$ throughout training, and 2) dynamic-bit allocation (DBA), where each client is assigned a randomly + +Table 1. FL performance on the three benchmark datasets with a $5\%$ participation rate over 100 clients for $\alpha \in \{0.1, 0.3, 0.6\}$ . + +
Method#bitsCIFAR-10CIFAR-100Tiny-ImageNet
α = 0.10.30.6i.i.d.0.10.30.6i.i.d.0.10.30.6i.i.d.
FedAvg [31]3270.6582.5285.4589.2543.4547.3949.0448.6630.3734.6536.5437.39
FedProx [27]3270.8983.9186.6088.8143.7848.2648.4948.5429.9935.4437.1437.43
FedAvgM [13]3278.3085.2787.9490.0047.7052.0453.0053.4032.0037.6339.6641.02
FedADAM [36]3272.5881.7385.9788.4546.8353.4355.9657.9334.7739.9743.2746.26
FedDyn [1]3283.6388.2889.2091.0853.8356.4056.9056.9337.3240.9242.7842.65
FedMLB [21]3271.1282.4086.2389.2549.7254.1655.1756.4533.4539.0640.5541.46
FedLC [52]3274.0383.8086.2088.3543.8648.5048.6147.8131.7334.5536.7238.06
FedNTD [26]3273.2983.3486.2689.3645.0848.7350.3950.9034.3636.4439.0640.60
FedSmoo [43]3278.6180.0785.0386.4842.0244.9745.3249.4323.0927.1429.7034.96
FedDecorr [39]3275.4483.0184.4988.4044.0549.3550.4449.4031.1234.1335.5936.36
FedWon [56]3270.8978.3682.4486.5645.5551.6853.6055.2830.5236.2038.5940.73
FedRCL [38]3283.4588.1989.7491.5558.2662.2163.9565.1137.8644.9446.7547.96
FedACG [20]3283.6289.1290.3891.7558.1462.8061.8562.4639.7545.4648.2250.49
FedPAQ [37]466.8178.5281.3685.4638.0642.4243.4644.3528.5431.9333.5634.98
157.8271.2772.3679.4630.0734.1638.2437.8524.7429.4830.7433.81
FedHQ+ [5]466.8978.4281.4285.5438.2242.6643.1044.8628.5932.1834.2435.69
158.1470.6873.1679.0431.1134.1137.5039.1724.7929.1831.1433.50
FedWSQFedWS (32)85.9489.8191.0092.4764.1468.1969.3170.1447.0552.8454.2355.26
486.8489.8491.1192.2164.3467.3168.4168.5346.5150.4951.9652.13
184.7988.6289.4991.1562.0565.1466.1966.1145.1149.9351.1151.77
FBA (2.33)86.1689.6090.4491.7163.0766.1667.0867.1746.6250.6951.4351.68
DBA (2.33)85.2089.2390.2691.7863.2866.3267.2667.2246.5050.2451.5551.80
+ +selected bit-width at every communication round. With bit-width selection following a uniform distribution over $\{1,2,4\}$ , the expected bit-width per client is approximately 2.3 bits. The entire FedWSQ framework, including the mixed-precision strategy, is summarized in Algorithm 1. + +# 4. Experiments + +In this section, we evaluate the proposed FedWSQ on standard FL benchmarks, compare its performance against various FL methods on both i.i.d. and non-i.i.d. data conditions, and offer a further analysis of FedWSQ. The convergence plots obtained from the experiments are provided in Section D of the supplementary document. + +# 4.1. Experimental setup + +For experiments, we use three standard benchmark datasets: CIFAR-10 [23], CIFAR-100 [23], and Tiny-ImageNet [25]. CIFAR-10 and CIFAR-100 contain 60,000 images divided into 10 and 100 classes, respectively. Tiny-ImageNet is composed of 200 classes, providing a more complex image classification task. For i.i.d. settings, training samples are randomly assigned to clients without replacement. For the non-i.i.d. setting, we sample label ratios from a Dirichlet distribution with a concentration parameter $\alpha \in \{0.1, 0.3, 0.6\}$ , following Hsu et al. [13]. Lower values of $\alpha$ (e.g., 0.1) indicate higher data heterogeneity, whereas larger values (e.g., 0.6) and the i.i.d. case correspond to more homogeneous distributions. The participation rate is set to $5\%$ among 100 clients. Model accuracy is evaluated on the test set of each dataset at the 1,000th communication round. To ensure stable evaluations, we report accuracy + +using an exponential moving average with a smoothing parameter of 0.9. To validate the effectiveness of FedWSQ, we integrate it with FedAvg [31]. Note that FedWSQ can be incorporated into any existing FL algorithm. For clarity, we refer to the proposed method without DANUQ as FedWS. For quantized FL settings, we consider three bit-allocation strategies: constant allocation, FBA, and DBA. In the constant allocation settings, a fixed bit-width is assigned to all clients. We compare FedWSQ against various FL algorithms, including FedAvg [31], FedProx [27], FedAvgM [13], FedADAM [36], FedDyn [1], FedMLB [21], FedLC [52], FedNTD [26], FedSmoo [43], FedDecorr [39], FedWon [56], FedRCL [38], FedACG [20], FedPAQ [37], and FedHQ+ [5]. + +# 4.2. Implementation details + +We follow most of the implementation setups and evaluation protocols from [1, 20, 38, 48]. The ResNet-18 architecture [11] is adopted as the backbone network. Consistent with standard practice in FL [12], all BN [15] layers in ResNet-18 are replaced with GN [47] layers. Following the recommendation of [35], WS is applied before each GN layer to achieve optimal performance. The hyper-parameter of WS is set to $\rho = 0.001$ , and the momentum for updating the scaling vector is set to $\beta = 0.1$ . Additional information on the hyper-parameter selection is provided in Section C of the supplementary document. + +# 4.3. Comparison with existing FL methods + +Table 1 presents the performance comparison of the proposed FedWSQ against various existing FL methods on the three benchmark datasets. The bit-width settings for each + +Table 2. Comparison of UQ and DANUQ methods with or without WS on the three benchmark datasets with a $5\%$ participation rate over 100 clients for $\alpha \in \{0.1, 0.3, 0.6\}$ . + +
DatasetWS#bitsα = 0.1α = 0.3α = 0.6i.i.d.
UQDANUQUQDANUQUQDANUQUQDANUQ
CIFAR-10X157.8272.4871.2780.8872.3683.5279.4686.92
262.3965.1974.3578.6176.7982.5982.2586.11
466.8172.8778.5284.6281.3686.2685.4688.66
✓175.4884.7980.1588.6282.5389.4985.3791.15
278.3685.9283.0288.9884.1090.3586.9591.50
482.7886.8486.3789.8487.5891.1189.8792.21
CIFAR-100X130.0740.5134.1646.6238.2448.6737.8548.55
233.5240.5837.6645.0740.4746.3140.9046.06
438.0643.5942.4248.3343.4649.2944.3548.76
✓145.5962.0550.1065.1452.1366.1954.5866.11
249.6863.3454.2265.9656.0167.1958.7267.82
456.8564.3460.2167.3161.9868.4162.9868.53
Tiny-ImageNetX124.7428.9529.4833.1430.7436.7533.8138.95
227.0230.3331.6034.1233.5836.634.8735.41
428.5431.1431.9335.4233.5636.6634.9838.13
✓132.7345.1137.6349.9339.6051.1141.9851.77
236.1645.9640.9050.1742.5151.1244.8951.12
440.4646.5144.5850.4945.5551.9646.9352.13
+ +Table 3. Comparison of NUQ methods using WS on CIFAR-10 and CIFAR-100 datasets for $\alpha \in \{0.1, 0.3, 0.6\}$ . + +
#bitsMethodCIFAR-10CIFAR-100
α = 0.10.30.60.10.30.6
1NF/FP24.030.628.87.08.38.1
DANUQ84.888.689.562.165.166.2
2NF57.873.880.041.249.153.0
FP45.057.558.625.332.034.2
DANUQ85.989.090.463.366.067.2
4NF86.689.791.064.866.768.4
FP87.190.091.064.367.068.4
DANUQ86.889.891.164.367.368.4
+ +FL method are indicated in the table. Compared to FL methods that use quantization, such as FedPAQ and FedHQ+, FedWSQ demonstrates superior performance for 4-bit and 1-bit settings. Also, FedWSQ outperforms FL methods with full-precision representation, including SOTA methods such as FedACG and FedRCL. Notably, even with a 1-bit representation, FedWSQ demonstrates competitive performance on CIFAR-10 and achieves superior results on CIFAR-100 and Tiny/ImageNet. For example, in a highly heterogeneous setting $(\alpha = 0.1)$ , FedWSQ (1-bit) achieves $62.05\%$ accuracy on CIFAR-100, surpassing FedRCL and FedACG by $3.79\%$ and $3.91\%$ , respectively. Similarly, on Tiny/ImageNet, FedWSQ (1-bit) achieves $45.11\%$ accuracy, outperforming FedRCL and FedACG by $7.25\%$ and $5.36\%$ , respectively. These results highlight the effectiveness of FedWSQ, which demonstrates its robustness even in extremely low bit-width scenarios while maintaining high model accuracy. + +# 4.4. Additional analyses + +Impact of bit representation in FedWSQ FedWSQ balances model performance and communication efficiency by adopting different bit-width settings. Table 1 presents the + +![](images/1ed1f36c553f79a20bc740e1b8baee9687320d82ed9fa44955e3e9d54d496882.jpg) +Figure 3. Loss landscape of FL methods trained on CIFAR-100, where the numbers indicate the Hessian top eigenvalue. + +results for 32-bit (FedWS), 4-bit, and 1-bit representations, as well as adaptive bit-allocation strategies (FBA and DBA). Despite using lower bit-widths, FedWSQ remains robust in non-i.i.d. settings, particularly under high heterogeneity ( $\alpha = 0.1$ ). Since local models tend to overfit under high data heterogeneity, quantizing model updates helps to regularize overfitting, achieving more robust FL performance [2, 3]. The accuracy gaps among different bit-widths are relatively small, and thus the FBA and DBA strategies enable efficient communication with marginal performance loss. As shown in Table 1, both FBA and DBA strategies achieve performance close to the 4-bit setting while reducing communication costs by half. In summary, FedWSQ confirms its ability to maintain accuracy while significantly reducing communication bandwidth usage for real-world FL scenarios. + +Effect of WS and DANUQ Table 2 shows a comparative analysis of our proposed DANUQ and UQ under varying bit-widths and data heterogeneity. In this experiment, UQ is implemented using the FedPAQ approach [37]. As listed in the table, our DANUQ consistently outperforms the UQ method. Under extremely low-bit quantization (1-bit and 2-bit settings), the UQ suffers from significant accuracy degradation. While WS alone improves performance, combining WS with the proposed DANUQ yields the high- + +Table 4. FL performance of different backbones for $\alpha = {0.3}$ . + +
BackboneVanillaFedWSFedWSQ (DBA)
ShuffleNet [53]36.3753.3251.12
VGGNet-9 [40]46.1460.0160.48
SqueezeNet [14]39.4555.8756.10
ResNet-18 [11]47.3968.1966.32
MobileViT [32]35.4240.8041.27
+ +est performance gains. The UQ with WS shows improvements compared with its non-WS counterparts, but the synergy between WS and DANUQ provides the most robust FL performance. Furthermore, our method maintains high accuracy even with low-bit representations. This suggests that applying FBA or DBA can be highly effective, as our approach maintains strong performance across varying bit-widths and makes FedWSQ particularly practical for real-world FL deployments. + +Comparison with other NUQ methods We evaluate the effectiveness of the proposed DANUQ by comparing it with existing NUQ approaches, FP [45] and NF [7].7 As shown in Table 3, at the 4-bit representation, all NUQ methods perform competitively, indicating that NUQ is generally robust at relatively higher bit-widths. However, under 1-bit and 2-bit representations, FP and NF suffer from severe performance degradation, particularly in highly heterogeneous data ( $\alpha = 0.1$ ). In contrast, our DANUQ outperforms both NF and FP across all conditions, which confirms its robustness against extreme quantization constraints. These results highlight the superiority of our DANUQ method among NUQ methods for communication-efficient FL. + +Loss Landscape Analysis Figure 3 visualizes the loss landscapes of various FL methods compared to FedWSQ (with DBA). FedWSQ achieves the smoothest and most stable loss landscape, with the lowest Hessian top eigenvalue of 135.8, indicating improved generalization ability. FedSmoo exhibits a flatter landscape but converges to a much higher global minimum, reducing sharpness at the cost of performance. In contrast, FedACG reaches a global comparable to FedWSQ but has a significantly sharper landscape, with a Hessian top eigenvalue of 671.2, which indicates it is more sensitive to perturbations. These results show that FedWSQ effectively enhances the generalization, leading to more stable training and improved FL performance. + +Impact on different backbone architectures Table 4 presents the performance of different backbone architectures on CIFAR-100 to evaluate the impact of FedWS and FedWSQ. For all architectures, FedWS significantly improves accuracy over the vanilla models, and FedWSQ maintains a comparable performance to FedWS with improved communication efficiency. In addition, We eval + +Table 5. Ablation study for $\rho$ using FedWS on CIFAR-10. For the non-i.i.d. setting, $\alpha = {0.3}$ is used. + +
ρ1 × 10-41 × 10-31 × 10-21 × 10-1
non-i.i.d.87.1589.7189.6289.46
i.i.d.90.6492.4891.9792.11
+ +uate MobileViT [32], a vision transformer-based architecture, which also benefits from FedWS and FedWSQ. These results confirm their applicability beyond CNNs. + +Comparison with FedWon We provide the comparison of FedWSQ with FedWon, which was discussed in Section 3.1. As shown in Table 1, FedWSQ substantially outperforms FedWon across all datasets and data heterogeneity settings. As illustrated in Figure 3, FedWon exhibits a much higher global minimum and a very sharp loss landscape with a Hessian top eigenvalue of 203.4, whereas FedWSQ maintains a smoother surface with an eigenvalue of 135.8. These results demonstrate FedWSQ's superior robustness and generalization ability. + +Hyper-parameters We examine the hyper-parameter $\rho$ for WS, which handles the normalization scales. To evaluate its impact, we vary the values of $\rho$ and assess the performance of FedWS under both non-i.i.d. and i.i.d. data distributions. As shown in Table 5, the performance of FedWS is insensitive to $\rho$ variations. This is because during inference, normalization layers like GN cancel the effect of constant scaling, and the variance of the parameter vectors within the same convolutional layer is not significant. We select $\rho = 1 \times 10^{-3}$ , which yields the best results. + +# 5. Conclusions + +This paper presents FedWSQ, a novel FL framework using WS and the proposed DANUQ to address key FL challenges, including data heterogeneity, partial client participation, and communication bottleneck. By leveraging WS, FedWSQ enhances training stability via an implicit gradient filtering mechanism. Additionally, DANUQ efficiently compresses LMPUs while preserving essential local information, significantly reducing communication overhead even at ultra-low bit-widths. Extensive experiments on the benchmark datasets confirm that FedWSQ consistently outperforms SOTA FL methods under various conditions. Furthermore, the mixed-precision strategies FBA and DBA effectively balance communication cost and model accuracy. These results validate the robustness of the FedWSQ framework in extreme quantization scenarios, making it highly practical for real-world FL. + +Acknowledgments This work was supported by the National Research Foundation (NRF), South Korea, under project BK21 FOUR (Smart Robot Convergence and Application Education Research Center). + +# References + +[1] Durmus Alp Emre Acar, Yue Zhao, Ramon Matas, Matthew Mattina, Paul Whatmough, and Venkatesh Saligrama. Federated learning based on dynamic regularization. In ICLR, 2021. 2, 6, 3 +[2] MohammadHossein AskariHemmat, Reyhane Askari Hemmat, Alex Hoffman, Ivan Lazarevich, Ehsan Saboori, Olivier Mastropietro, Sudhakar Sah, Yvon Savaria, and Jean-Pierre David. QReg: On regularization effects of quantization. arXiv preprint arXiv:2206.12372, 2022. 7 +[3] MohammadHossein AskariHemmat, Ahmadreza Jeddi, Reyhane Askari Hemmat, Ivan Lazarevich, Alexander Hoffman, Sudhakar Sah, Ehsan Saboori, Yvon Savaria, and Jean-Pierre David. QGen: On the ability to generalize in quantization aware training. arXiv preprint arXiv:2404.11769, 2024. 7 +[4] Debraj Basu, Deepesh Data, Can Karakus, and Suhas N Diggavi. Qsparse-Local-SGD: Distributed SGD with quantization, sparsification, and local computations. IEEE Journal on Selected Areas in Information Theory, 1(1):217-226, 2020. 1 +[5] Shengbo Chen, Cong Shen, Lanxue Zhang, and Yuanmin Tang. Dynamic aggregation for heterogeneous quantization in federated learning. IEEE Transactions on Wireless Communications, 20(10):6804-6819, 2021. 2, 6 +[6] Jungwook Choi, Zhuo Wang, Swagath Venkataramani, Pierce I-Jen Chuang, Vijayalakshmi Srinivasan, and Kailash Gopalakrishnan. Pact: Parameterized clipping activation for quantized neural networks. In ICLR, 2018. 2 +[7] T. Dettmers, M. Lewis, S. Shleifer, and L. Zettlemoyer. 8-bit optimizers via block-wise quantization. In ICLR, 2022. 4, 8 +[8] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: Efficient finetuning of quantized llms. In NeurIPS, 2023. 2 +[9] Steven K. Esser, Jeffrey L. McKinstry, Deepika Bablani, Rathinakumar Appuswamy, and Dharmendra S. Modha. Learned Step Size Quantization. In ICLR, 2020. 2 +[10] Robert M. Gray and David L. Neuhoff. Quantization. IEEE transactions on information theory, 44(6):2325-2383, 1998. 2 +[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6, 8, 3 +[12] Kevin Hsieh, Amar Phanishayee, Onur Mutlu, and Phillip Gibbons. The non-IID data quagmire of decentralized machine learning. In ICML, 2020. 6, 3 +[13] Tzu-Ming Harry Hsu, Hang Qi, and Matthew Brown. Measuring the effects of non-identical data distribution for federated visual classification. arXiv preprint arXiv:1909.06335, 2019. 2, 6, 3 +[14] Forrest N Iandola, Song Han, Matthew W Moskewicz, Khalid Ashraf, William J Dally, and Kurt Keutzer. SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and $< 0.5$ MB model size. arXiv preprint arXiv:1602.07360, 2016.8 +[15] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015. 6 + +[16] Benoit Jacob, Skirmantas Kligys, Bo Chen, Menglong Zhu, Matthew Tang, Andrew Howard, Hartwig Adam, and Dmitry Kalenichenko. Quantization and training of neural networks for efficient integer-arithmetic-only inference. In CVPR, pages 2704-2713, 2018. 2, 4 +[17] Sangil Jung, Changyong Son, Seohyung Lee, Jinwoo Son, Youngjun Kwak, Jae-Joon Han, Sung Ju Hwang, and Changkyu Choi. Learning to quantize deep networks by optimizing quantization intervals with task loss. In CVPR, pages 4350-4359, 2019. 2 +[18] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank J Reddi, Sebastian U Stich, and Ananda Theertha Suresh. SCAFFOLD: Stochastic controlled averaging for ondevice federated learning. In ICML, 2020. 1 +[19] Ahmed Khaled, Konstantin Mishchenko, and Peter Richtárik. First analysis of local GD on heterogeneous data. arXiv preprint arXiv:1909.04715, 2019. 1 +[20] Geeho Kim, Jinkyu Kim, and Bohyung Han. Communication-efficient federated learning with accelerated client gradient. In CVPR, 2024. 6, 3 +[21] Jinkyu Kim, Geeho Kim, and Bohyung Han. Multi-level branched regularization for federated learning. In ICML, 2022. 2, 6, 3 +[22] Jakub Konečny, H. Brendan McMahan, Felix X. Yu, Ananda Theertha Suresh, Dave Bacon, and Peter Richtárik. Federated learning: Strategies for improving communication efficiency. In arXiv preprint, 2017. 2 +[23] Alex Krizhevsky. Learning multiple layers of features from tiny images. Master's thesis, Department of Computer Science, University of Toronto, 2009. 6 +[24] Khiem Le, Nhan Luong-Ha, Manh Nguyen-Duc, Danh Le-Phuoc, Cuong Do, and Kok-Seng Wong. Exploring the practicality of federated learning: A survey towards the communication perspective. arXiv preprint arXiv:2405.20431, 2024. 1 +[25] Ya Le and Xuan Yang. Tiny ImageNet visual recognition challenge. CS 231N, 7(7):3, 2015. 6 +[26] Gihun Lee, Minchan Jeong, Yongjin Shin, Sangmin Bae, and Se-Young Yun. Preservation of the global knowledge by not-true distillation in federated learning. In NeurIPS, 2022. 6, 3 +[27] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. In MLSys, 2020. 2, 6, 3 +[28] Xiang Li, Kaixuan Huang, Wenhao Yang, Shusen Wang, and Zhihua Zhang. On the convergence of FedAvg on non-IID data. In ICLR, 2020. 1 +[29] Darryl Lin, Sachin Talathi, and Sreekanth Annapureddy. Fixed point quantization of deep convolutional networks. In International conference on machine learning, pages 2849-2858. PMLR, 2016. 2 +[30] Wang Luping, Wang Wei, and Li Bo. CMFL: Mitigating communication overhead for federated learning. In 2019 IEEE 39th international conference on distributed computing systems (ICDCS), pages 954-964. IEEE, 2019. 1 +[31] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication- + +efficient learning of deep networks from decentralized data. In AISTATS, 2017. 1, 2, 6 +[32] Sachin Mehta and Mohammad Rastegari. Mobilevit: lightweight, general-purpose, and mobile-friendly vision transformer. arXiv preprint arXiv:2110.02178, 2021. 8 +[33] Markus Nagel, Marios Fournarakis, Rana Ali Amjad, Yelysei Bondarenko, Mart van Baalen, and Tijmen Blankevoort. A white paper on neural network quantization. arXiv preprint arXiv:2106.08295, 2021. 2, 4 +[34] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. PyTorch: An imperative style, high-performance deep learning library. In NeurIPS, 2019. 3 +[35] Siyuan Qiao, Huiyu Wang, Chenxi Liu, Wei Shen, and Alan Yuille. Micro-batch training with batch-channel normalization and weight standardization. arXiv preprint arXiv:1903.10520, 2019. 1, 3, 6 +[36] Sashank J Reddi, Zachary Charles, Manzil Zaheer, Zachary Garrett, Keith Rush, Jakub Konečny, Sanjiv Kumar, and Hugh Brendan McMahan. Adaptive federated optimization. In ICLR, 2021. 2, 6, 3 +[37] Amirhossein Reisizadeh, Aryan Mokhtari, Hamed Hassani, Ali Jabbabaie, and Ramtin Pedarsani. FedPAQ: A communication-efficient federated learning method with periodic averaging and quantization. In AISTATS, 2020. 2, 6, 7 +[38] Seonguk Seo, Jinkyu Kim, Geeho Kim, and Bohyung Han. Relaxed contrastive learning for federated learning. In CVPR, 2024. 6, 3 +[39] Yujun Shi, Jian Liang, Wenqing Zhang, Vincent YF Tan, and Song Bai. Towards understanding and mitigating dimensional collapse in heterogeneous federated learning. In ICLR, 2023. 6, 3 +[40] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2014. 8 +[41] Sebastian U Stich. Local SGD converges fast and communicates little. In ICLR, 2019. 1 +[42] Sebastian U Stich and Sai Praneeth Karimireddy. The errorfeedback framework: Better rates for SGD with delayed gradients and compressed communication. arXiv preprint arXiv:1909.05350, 2019. 1 +[43] Yan Sun, Li Shen, Shixiang Chen, Liang Ding, and Dacheng Tao. Dynamic regularized sharpness aware minimization in federated learning: Approaching global consistency and smooth landscape. In International Conference on Machine Learning, pages 32991-33013. PMLR, 2023. 2, 6 +[44] Jianyu Wang and Gauri Joshi. Cooperative SGD: A unified framework for the design and analysis of local-update SGD algorithms. Journal of Machine Learning Research, 22(213): 1-50, 2021. 1 +[45] Ruizhe Wang, Yeyun Gong, Xiao Liu, Guoshuai Zhao, Ziyue Yang, Baining Guo, Zhengjun Zha, and Peng Cheng. Optimizing large language model training using fp4 quantization. In arXiv preprint arXiv:2501.17116, 2025. 4, 8 + +[46] Jie Wen, Zhixia Zhang, Yang Lan, Zhihua Cui, Jianghui Cai, and Wensheng Zhang. A survey on federated learning: challenges and applications. International Journal of Machine Learning and Cybernetics, 14(2):513-535, 2023. 1 +[47] Yuxin Wu and Kaiming He. Group normalization. In ECCV, 2018. 6 +[48] Jing Xu, Sen Wang, Liwei Wang, and Andrew Chi-Chih Yao. FedCM: Federated learning with client-level momentum. arXiv preprint arXiv:2106.10874, 2021. 6, 3 +[49] Haibo Yang, Minghong Fang, and Jia Liu. Achieving linear speedup with partial worker participation in non-IID federated learning. In ICLR, 2021. 1 +[50] Hao Yu, Sen Yang, and Shenghuo Zhu. Parallel restarted SGD with faster convergence and less communication: Demystifying why model averaging works for deep learning. In AAAI, 2019. 1 +[51] Dongqing Zhang, Jiaolong Yang, Dongqiangzi Ye, and Gang Hua. LQ-Nets: Learned Quantization for Highly Accurate and Compact Deep Neural Networks. In ECCV, 2018. 2 +[52] Jie Zhang, Zhiqi Li, Bo Li, Jianghe Xu, Shuang Wu, Shouhong Ding, and Chao Wu. Federated learning with label distribution skew via logits calibration. In ICML, 2022. 6, 3 +[53] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. ShuffleNet: An extremely efficient convolutional neural network for mobile devices. In CVPR, 2018. 8 +[54] Sijie Zhao, Tao Yue, and Xuemei Hu. Distribution-aware adaptive multi-bit quantization. In CVPR, 2021. 4 +[55] Chenzhuo Zhu, Song Han, Huizi Mao, and William J. Dally. Trained ternary quantization. In ICLR, 2017. 2 +[56] Weiming Zhuang and Lingjuan Lyu. FedWon: Triumphing multi-domain federated learning without normalization. In ICLR, 2023. 4, 6 \ No newline at end of file diff --git a/2025/FedWSQ_ Efficient Federated Learning with Weight Standardization and Distribution-Aware Non-Uniform Quantization/images.zip b/2025/FedWSQ_ Efficient Federated Learning with Weight Standardization and Distribution-Aware Non-Uniform Quantization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1482ea6c375241573e32101b177992eef5a651d0 --- /dev/null +++ b/2025/FedWSQ_ Efficient Federated Learning with Weight Standardization and Distribution-Aware Non-Uniform Quantization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:507ce7f12558e90ce9603be520053be275a92cc0eee40b5e316942c137b6f08d +size 501738 diff --git a/2025/FedWSQ_ Efficient Federated Learning with Weight Standardization and Distribution-Aware Non-Uniform Quantization/layout.json b/2025/FedWSQ_ Efficient Federated Learning with Weight Standardization and Distribution-Aware Non-Uniform Quantization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1744190af568e5389f4513247ba0d4d3aec71b02 --- /dev/null +++ b/2025/FedWSQ_ Efficient Federated Learning with Weight Standardization and Distribution-Aware Non-Uniform Quantization/layout.json @@ -0,0 +1,10072 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 79, + 102, + 532, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 102, + 532, + 140 + ], + "spans": [ + { + "bbox": [ + 79, + 102, + 532, + 140 + ], + "type": "text", + "content": "FedWSQ: Efficient Federated Learning with Weight Standardization and Distribution-Aware Non-Uniform Quantization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 102, + 160, + 508, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 160, + 508, + 175 + ], + "spans": [ + { + "bbox": [ + 102, + 160, + 508, + 175 + ], + "type": "text", + "content": "Seung-Wook Kim\\*1, Seongyeol Kim\\*1, Jiah Kim\\*1, Seowon Ji\\*2, and Se-Ho Lee\\*3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 123, + 175, + 488, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 175, + 488, + 189 + ], + "spans": [ + { + "bbox": [ + 123, + 175, + 488, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 123, + 175, + 488, + 189 + ], + "type": "text", + "content": "Department of Intelligent Robot Engineering, Pukyong National University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 135, + 190, + 471, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 190, + 471, + 203 + ], + "spans": [ + { + "bbox": [ + 135, + 190, + 471, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 135, + 190, + 471, + 203 + ], + "type": "text", + "content": "CSE, Konkuk University " + }, + { + "bbox": [ + 135, + 190, + 471, + 203 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 135, + 190, + 471, + 203 + ], + "type": "text", + "content": "CS&AI, Jeonbuk National University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 257, + 297, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 257, + 297, + 484 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 297, + 484 + ], + "type": "text", + "content": "Federated learning (FL) often suffers from performance degradation due to key challenges such as data heterogeneity and communication constraints. To address these limitations, we present a novel FL framework called FedWSQ, which integrates weight standardization (WS) and the proposed distribution-aware non-uniform quantization (DANUQ). WS enhances FL performance by filtering out biased components in local updates during training, thereby improving the robustness of the model against data heterogeneity and unstable client participation. In addition, DANUQ minimizes quantization errors by leveraging the statistical properties of local model updates. As a result, FedWSQ significantly reduces communication overhead while maintaining superior model accuracy. Extensive experiments on FL benchmark datasets demonstrate that FedWSQ consistently outperforms existing FL methods across various challenging FL settings, including extreme data heterogeneity and ultra-low-bit communication scenarios. The source code is available at our project page1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 498, + 135, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 135, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 135, + 510 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 517, + 295, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 295, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 295, + 637 + ], + "type": "text", + "content": "In large-scale machine learning, centralized approaches often raise privacy concerns because sensitive data from edge devices must be collected on a central server for model training. Federated learning (FL) [31] enables distributed devices to collaboratively train a shared global model without sharing raw data. A fundamental FL method, FedAvg, was proposed by McMahan et al. [31]. In this approach, each client trains its model using the local data, and then the server collects and averages these local updates to form a unified global model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 638, + 296, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 638, + 296, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 638, + 296, + 674 + ], + "type": "text", + "content": "Several studies on FL have verified its practical effectiveness [4, 28, 41, 42, 44, 49, 50]. However, real-world FL methods often face three key challenges [18, 19, 24," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 232, + 553, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 553, + 352 + ], + "type": "text", + "content": "30, 46]: 1) data heterogeneity, where clients possess non-independent and identically distributed (non-i.i.d.) data; 2) partial client participation, where only a subset of clients contributes to a global model update during each training round due to communication constraints; 3) communication bottlenecks, where limited communication bandwidth and high transmission costs hinder efficient model aggregation. These challenges worsen local gradient divergence, slow down the global model convergence, and ultimately degrade FL performance in real-world applications." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 353, + 554, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 353, + 554, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 353, + 554, + 616 + ], + "type": "text", + "content": "To address these limitations, we propose FedWSQ, an advanced FL framework that combines weight standardization (WS) with our proposed distribution-aware nonuniform quantization (DANUQ) to improve both learning stability and communication efficiency. WS is a plug-and-play technique that standardizes the weight vectors of convolutional or linear layers to stabilize the learning process of a neural network [35]. In FedWSQ, WS plays a crucial role in mitigating client drift by filtering out gradient components that contribute to overfitting during local training. This leads to improved generalization across heterogeneous clients and significantly enhances FL performance. Although WS accelerates global convergence, the communication cost per round remains a major bottleneck. To address this, FedWSQ integrates WS with DANUQ, a novel quantization strategy that leverages a standard normal distribution prior to minimize quantization errors, reducing communication overhead while preserving model performance. Even under ultra-low-bit quantization, the proposed FedWSQ achieves superior performance, consistently outperforming state-of-the-art (SOTA) FL methods across various datasets and FL settings." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 324, + 617, + 533, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 617, + 533, + 628 + ], + "spans": [ + { + "bbox": [ + 324, + 617, + 533, + 628 + ], + "type": "text", + "content": "Our main contributions are summarized as follows." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 313, + 630, + 553, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 553, + 701 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 553, + 701 + ], + "type": "text", + "content": "- We propose an effective application of WS in FL, enhancing model convergence and stability under data heterogeneity and limited client participation. By implicitly performing a gradient filtering process, WS mitigates client drift while also providing a regularization effect that stabilizes training and improves generalization." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "type": "text", + "content": "- We introduce DANUQ, a novel quantization method that" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 683, + 163, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 683, + 163, + 693 + ], + "spans": [ + { + "bbox": [ + 66, + 683, + 163, + 693 + ], + "type": "text", + "content": "*indicates equal contribution." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 67, + 693, + 148, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 693, + 148, + 703 + ], + "spans": [ + { + "bbox": [ + 67, + 693, + 148, + 703 + ], + "type": "text", + "content": "† Corresponding authors." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 67, + 703, + 260, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 260, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 260, + 712 + ], + "type": "text", + "content": "1https://github.com/Seongyeol-kim/FedWSQ" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "4616" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 72, + 295, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 72, + 295, + 156 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 295, + 156 + ], + "type": "text", + "content": "employs a fixed quantization function, leveraging the statistical properties of local model parameter updates (LMPUs) to minimize quantization errors under parameter-distribution priors. By efficiently compressing LM-PUs while preserving essential information, the proposed DANUQ significantly reduces communication overhead without noticeable performance degradation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 156, + 294, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 156, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 54, + 156, + 294, + 216 + ], + "type": "text", + "content": "- Extensive experiments show that FedWSQ significantly enhances performance across diverse settings. Even with an average of 2.3 bits per parameter, FedWSQ improves performance by over " + }, + { + "bbox": [ + 54, + 156, + 294, + 216 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 54, + 156, + 294, + 216 + ], + "type": "text", + "content": " on Tiny-ImageNet in a highly heterogeneous setting compared to SOTA FL methods." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 225, + 236, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 225, + 236, + 237 + ], + "spans": [ + { + "bbox": [ + 55, + 225, + 236, + 237 + ], + "type": "text", + "content": "2. Related Work and Preliminaries" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 245, + 296, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 245, + 296, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 245, + 296, + 339 + ], + "type": "text", + "content": "Federated learning (FL) FL is a decentralized approach that enables collaborative training of a global model. Using distributed devices, FL ensures efficient learning while preserving data privacy. Let " + }, + { + "bbox": [ + 55, + 245, + 296, + 339 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_n(\\mathbf{w}_{\\mathrm{g}})\\coloneqq \\mathbb{E}_{\\zeta_n\\sim \\mathcal{D}_n}[\\mathcal{F}_n(\\mathbf{w}_{\\mathrm{g}};\\zeta_n)]" + }, + { + "bbox": [ + 55, + 245, + 296, + 339 + ], + "type": "text", + "content": " be the local loss function of the client " + }, + { + "bbox": [ + 55, + 245, + 296, + 339 + ], + "type": "inline_equation", + "content": "n\\in [N]" + }, + { + "bbox": [ + 55, + 245, + 296, + 339 + ], + "type": "text", + "content": " with a local data distribution " + }, + { + "bbox": [ + 55, + 245, + 296, + 339 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_n" + }, + { + "bbox": [ + 55, + 245, + 296, + 339 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 245, + 296, + 339 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{\\mathrm{g}}" + }, + { + "bbox": [ + 55, + 245, + 296, + 339 + ], + "type": "text", + "content": " is the global model parameter (GMP) vector. The training objective function of FL [31] is formulated as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 341, + 294, + 366 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 341, + 294, + 366 + ], + "spans": [ + { + "bbox": [ + 106, + 341, + 294, + 366 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathbf {w} _ {\\mathrm {g}}} \\left\\{\\mathcal {F} \\left(\\mathbf {w} _ {\\mathrm {g}}\\right) := \\sum_ {i \\in \\mathcal {S}} h _ {i} \\mathcal {F} _ {i} \\left(\\mathbf {w} _ {\\mathrm {g}}\\right) \\right\\}, \\tag {1}", + "image_path": "0effc06b5bc4e852e20a041db2998dc26c8e93dfcb3fb6cf5fb3fca6fd82b54a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 367, + 295, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 367, + 295, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 295, + 486 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 367, + 295, + 486 + ], + "type": "inline_equation", + "content": "h_i" + }, + { + "bbox": [ + 55, + 367, + 295, + 486 + ], + "type": "text", + "content": " is the weight assigned to the " + }, + { + "bbox": [ + 55, + 367, + 295, + 486 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 367, + 295, + 486 + ], + "type": "text", + "content": "-th device, such that " + }, + { + "bbox": [ + 55, + 367, + 295, + 486 + ], + "type": "inline_equation", + "content": "h_i \\geq 0" + }, + { + "bbox": [ + 55, + 367, + 295, + 486 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 367, + 295, + 486 + ], + "type": "inline_equation", + "content": "\\sum_{i} h_i = 1" + }, + { + "bbox": [ + 55, + 367, + 295, + 486 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 367, + 295, + 486 + ], + "type": "inline_equation", + "content": "h_i \\propto |\\mathcal{D}_i|" + }, + { + "bbox": [ + 55, + 367, + 295, + 486 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 367, + 295, + 486 + ], + "type": "inline_equation", + "content": "S \\subseteq [N]" + }, + { + "bbox": [ + 55, + 367, + 295, + 486 + ], + "type": "text", + "content": " is a subset of participating clients in the current communication round. At each round " + }, + { + "bbox": [ + 55, + 367, + 295, + 486 + ], + "type": "inline_equation", + "content": "t \\in [T]" + }, + { + "bbox": [ + 55, + 367, + 295, + 486 + ], + "type": "text", + "content": ", the server broadcasts the GMPs " + }, + { + "bbox": [ + 55, + 367, + 295, + 486 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{\\mathrm{g}}^{t-1}" + }, + { + "bbox": [ + 55, + 367, + 295, + 486 + ], + "type": "text", + "content": " to the participating client " + }, + { + "bbox": [ + 55, + 367, + 295, + 486 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 367, + 295, + 486 + ], + "type": "text", + "content": ". In FedAvg [31], the " + }, + { + "bbox": [ + 55, + 367, + 295, + 486 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 367, + 295, + 486 + ], + "type": "text", + "content": "-th client performs " + }, + { + "bbox": [ + 55, + 367, + 295, + 486 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 367, + 295, + 486 + ], + "type": "text", + "content": " steps of local training to obtain the local model parameter (LMP) vector, " + }, + { + "bbox": [ + 55, + 367, + 295, + 486 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_i^t" + }, + { + "bbox": [ + 55, + 367, + 295, + 486 + ], + "type": "text", + "content": ", and transmits its LMP updates (LMPUs), " + }, + { + "bbox": [ + 55, + 367, + 295, + 486 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{w}_i^t := \\mathbf{w}_i^t - \\mathbf{w}_{\\mathrm{g}}^{t-1}" + }, + { + "bbox": [ + 55, + 367, + 295, + 486 + ], + "type": "text", + "content": ", to the server. Finally, the server aggregates these LMPUs through weighted averaging to derive the updated GMPs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 487, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 295, + 713 + ], + "type": "text", + "content": "In real-world FL scenarios, unstable learning dynamics and slow convergence often occur due to non-i.i.d. data distributions and limited communication bandwidth. To address these issues, FedProx [27] penalizes LMPUs by adding a proximal term in the local objective to reduce the gap between the global and local loss functions. FedAvgM [13] integrates momentum into the global update process to mitigate variance in model aggregation. FedAdam [36] introduces an adaptive optimizer that improves global and local model convergence, thereby accelerating overall training progress. FedDyn [1] presents dynamic regularization for each device to diminish deviations in the LMPUs caused by data heterogeneity. FedSmoo [43] applies constraints that prevent sharp fluctuations in local updates to generalize FL performance by ensuring global consistency. Recently, FedACG [21] was proposed to improve convergence by broadcasting a global model with a lookahead gradient and aligning local models with the shared global model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "Network quantization As neural network architectures have grown larger in scale, quantization has become essential for the memory optimization and acceleration of operations [9, 10, 16, 17, 29, 33, 55]. Let us consider a model using " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "-bit representation. A typical quantization function " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " that maps a full-precision value " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " to a quantization level (QL) " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " can be written as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 406, + 163, + 553, + 176 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 163, + 553, + 176 + ], + "spans": [ + { + "bbox": [ + 406, + 163, + 553, + 176 + ], + "type": "interline_equation", + "content": "q = \\mathcal {Q} (x / s), \\tag {2}", + "image_path": "cb63b1c0f0fa73df33fbfef0dba0079431e85cbc7d67877e55021d9a486d8329.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 182, + 553, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 553, + 231 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 553, + 231 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 182, + 553, + 231 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 182, + 553, + 231 + ], + "type": "text", + "content": " is a scaling factor, and each QL is matched with a " + }, + { + "bbox": [ + 313, + 182, + 553, + 231 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 182, + 553, + 231 + ], + "type": "text", + "content": "-bit integer from " + }, + { + "bbox": [ + 313, + 182, + 553, + 231 + ], + "type": "inline_equation", + "content": "\\{0,\\dots ,2^B -1\\}" + }, + { + "bbox": [ + 313, + 182, + 553, + 231 + ], + "type": "text", + "content": ". The primary goal of effective quantization is to determine the proper scaling factor and a set of QLs to minimize the quantization errors." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 232, + 555, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 555, + 436 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 555, + 436 + ], + "type": "text", + "content": "Quantization methods can be roughly categorized into two approaches: uniform quantization (UQ) and nonuniform quantization (NUQ). UQ divides the input range into " + }, + { + "bbox": [ + 313, + 232, + 555, + 436 + ], + "type": "inline_equation", + "content": "2^{B}" + }, + { + "bbox": [ + 313, + 232, + 555, + 436 + ], + "type": "text", + "content": " equal intervals and assigns a representative value as a QL to each interval [10, 29]. In contrast, NUQ adjusts the interval widths based on the prior data distribution [6, 8, 51]. By allocating finer intervals in dense data regions and broader intervals in sparse regions, NUQ preserves critical information more effectively than UQ. For the scaling factor " + }, + { + "bbox": [ + 313, + 232, + 555, + 436 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 232, + 555, + 436 + ], + "type": "text", + "content": ", both approaches commonly employ the absmax method [16, 33] to determine the dynamic range for quantization. Furthermore, to enhance data adaptability, researchers have explored methods for dynamically learning quantization parameters (e.g., scaling factors and QLs) [9, 17, 55]. During backpropagation, these parameters are updated to reduce quantization error and better align with the underlying data distribution." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "content": "Quantized FL In FL, communication bottlenecks arise primarily during uplink transmission, as each participating client must send its individual LMPUs to the central server. Unlike downlink transmission, where a single GMP is broadcast to multiple clients, uplink communication involves substantial data transfer. This problem is exacerbated in resource-constrained communication environments, such as mobile and IoT networks, where upload bandwidth is severely limited [5, 37]. To overcome this challenge, several studies have focused on compressing model updates. In [22], quantization and sparsification are combined to reduce the number of transmitted parameters. FedPAQ [37] employs a probabilistic rounding algorithm to prevent an excessive concentration of identical QLs. Meanwhile, FedHQ+ [5] assigns different weight factors to clients for global updates based on their individual quantization errors. Although these methods significantly improve communication efficiency, they often suffer from performance degradation under extremely low bit-width conditions. In this paper, we introduce FedWSQ, a novel quantization framework that is designed to maintain robust model performance even at ultra-low bit-width." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4617" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 111, + 72, + 499, + 192 + ], + "blocks": [ + { + "bbox": [ + 111, + 72, + 499, + 192 + ], + "lines": [ + { + "bbox": [ + 111, + 72, + 499, + 192 + ], + "spans": [ + { + "bbox": [ + 111, + 72, + 499, + 192 + ], + "type": "image", + "image_path": "4e12e521430890cc82b9f7d38086e2fc6a0e6165afd8de62ccb3dd0637e6a8c1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 109, + 194, + 499, + 264 + ], + "blocks": [ + { + "bbox": [ + 109, + 194, + 499, + 264 + ], + "lines": [ + { + "bbox": [ + 109, + 194, + 499, + 264 + ], + "spans": [ + { + "bbox": [ + 109, + 194, + 499, + 264 + ], + "type": "image", + "image_path": "4b655be1c8fec3b0c32d6a8e3deed4f5a7bf375616742c6b55241bbf9874effd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 268, + 555, + 292 + ], + "lines": [ + { + "bbox": [ + 54, + 268, + 555, + 292 + ], + "spans": [ + { + "bbox": [ + 54, + 268, + 555, + 292 + ], + "type": "text", + "content": "Figure 1. The overall framework of FedWSQ, which integrates WS and DANUQ for efficient FL. Clients transmit LMPUs quantized by the DANUQ along with scaling vectors to the server, where dequantization and aggregation are performed to update the global parameters." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 302, + 162, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 302, + 162, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 302, + 162, + 316 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 323, + 296, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 296, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 296, + 515 + ], + "type": "text", + "content": "Figure 1 illustrates the overall framework of the proposed FedWSQ, which employs WS and DANUQ for efficient FL. Each client has a standard neural network model consisting of multiple model blocks, where WS is applied to mitigate the impact of data heterogeneity between clients. After local training, the full-precision values of LMPUs are converted to a " + }, + { + "bbox": [ + 55, + 323, + 296, + 515 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 323, + 296, + 515 + ], + "type": "text", + "content": "-bit representation using the global scaling vectors and carefully designed QLs. These quantized LMPUs and their corresponding scaling vectors are transmitted to the server. The server then dequantizes the received data to higher-precision values and aggregates them to update the GMPs. In addition, the global scaling vectors are refined to ensure consistent scaling across different clients for DANUQ. Using the proposed DANUQ, FedWSQ enhances communication efficiency without sacrificing model performance in challenging FL scenarios." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 525, + 216, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 216, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 216, + 539 + ], + "type": "text", + "content": "3.1. Weight standardization (WS)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 544, + 296, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 296, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 296, + 664 + ], + "type": "text", + "content": "WS [35] standardizes each weight vector of a convolution or linear layer to enhance the learning process. By ensuring consistent parameter distributions during training, WS contributes to a more stable gradient flow. Beyond this general benefit, we argue that the use of WS in FL can address the challenges posed by the distributed nature of FL and data heterogeneity. Formally, consider the " + }, + { + "bbox": [ + 55, + 544, + 296, + 664 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 544, + 296, + 664 + ], + "type": "text", + "content": "-th client that has a network architecture consisting of " + }, + { + "bbox": [ + 55, + 544, + 296, + 664 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 544, + 296, + 664 + ], + "type": "text", + "content": " linear or convolutional layers. Let the " + }, + { + "bbox": [ + 55, + 544, + 296, + 664 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 544, + 296, + 664 + ], + "type": "text", + "content": "-th layer (without bias) map an input " + }, + { + "bbox": [ + 55, + 544, + 296, + 664 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_l \\in \\mathbb{R}^{I_l}" + }, + { + "bbox": [ + 55, + 544, + 296, + 664 + ], + "type": "text", + "content": " to an output " + }, + { + "bbox": [ + 55, + 544, + 296, + 664 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_l \\in \\mathbb{R}^{O_l}" + }, + { + "bbox": [ + 55, + 544, + 296, + 664 + ], + "type": "text", + "content": ", defined by" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 83, + 672, + 295, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 672, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 83, + 672, + 295, + 712 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {y} _ {l} = \\mathbf {W} _ {n, l} ^ {T} \\mathbf {x} _ {l} \\tag {3} \\\\ := \\left[ \\begin{array}{c c c c} \\mathbf {w} _ {n, m _ {1} ^ {l}} & \\mathbf {w} _ {n, m _ {2} ^ {l}} & \\dots & \\mathbf {w} _ {n, m _ {O _ {l}} ^ {l}} \\end{array} \\right] ^ {T} \\mathbf {x} _ {l}, \\qquad (4) \\\\ \\end{array}", + "image_path": "d8d22ec875cd06d8c3496c3572561142c30232653a9affb01590e854153ca910.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 303, + 555, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 303, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 555, + 376 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 303, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{n,m}" + }, + { + "bbox": [ + 313, + 303, + 555, + 376 + ], + "type": "text", + "content": ", for " + }, + { + "bbox": [ + 313, + 303, + 555, + 376 + ], + "type": "inline_equation", + "content": "m\\in \\mathcal{M}_l\\coloneqq \\{m_1^l,\\ldots ,m_{O_l}^l\\}" + }, + { + "bbox": [ + 313, + 303, + 555, + 376 + ], + "type": "text", + "content": ", is the weight vector. In WS, each " + }, + { + "bbox": [ + 313, + 303, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{n,m}" + }, + { + "bbox": [ + 313, + 303, + 555, + 376 + ], + "type": "text", + "content": " is standardized before being used within the layer. We denote the prestandardized parameter (PSP) vector as " + }, + { + "bbox": [ + 313, + 303, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{n,m}" + }, + { + "bbox": [ + 313, + 303, + 555, + 376 + ], + "type": "text", + "content": ", and the weight-standardized parameter (WSP) vector as " + }, + { + "bbox": [ + 313, + 303, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{w}}_{n,m}" + }, + { + "bbox": [ + 313, + 303, + 555, + 376 + ], + "type": "text", + "content": ", respectively. Then, the WSP vector " + }, + { + "bbox": [ + 313, + 303, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{w}}_{n,m}" + }, + { + "bbox": [ + 313, + 303, + 555, + 376 + ], + "type": "text", + "content": " is obtained by" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 361, + 380, + 555, + 406 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 380, + 555, + 406 + ], + "spans": [ + { + "bbox": [ + 361, + 380, + 555, + 406 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {w}} _ {n, m} = \\frac {\\rho}{\\sigma (\\mathbf {w} _ {n , m})} \\left(\\mathbf {I} - \\mathbf {P _ {1}}\\right) \\mathbf {w} _ {n, m}, \\qquad (5)", + "image_path": "a2808681e7e835ae540b81f80aa7723abec97bd833092fae0b63bb7533d464dd.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 411, + 555, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 555, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 555, + 509 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 411, + 555, + 509 + ], + "type": "inline_equation", + "content": "\\mathbf{1} \\in \\mathbb{R}^{I_l}" + }, + { + "bbox": [ + 313, + 411, + 555, + 509 + ], + "type": "text", + "content": " is a vector of all ones, " + }, + { + "bbox": [ + 313, + 411, + 555, + 509 + ], + "type": "inline_equation", + "content": "\\sigma(\\mathbf{w}_{n,m})" + }, + { + "bbox": [ + 313, + 411, + 555, + 509 + ], + "type": "text", + "content": " is the standard deviation of the elements in " + }, + { + "bbox": [ + 313, + 411, + 555, + 509 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{n,m}" + }, + { + "bbox": [ + 313, + 411, + 555, + 509 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 411, + 555, + 509 + ], + "type": "inline_equation", + "content": "\\mathbf{P_v}" + }, + { + "bbox": [ + 313, + 411, + 555, + 509 + ], + "type": "text", + "content": " represents the projection matrix onto the subspace spanned by the vector " + }, + { + "bbox": [ + 313, + 411, + 555, + 509 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 313, + 411, + 555, + 509 + ], + "type": "text", + "content": ", which is denoted by " + }, + { + "bbox": [ + 313, + 411, + 555, + 509 + ], + "type": "inline_equation", + "content": "\\mathrm{span}\\{\\mathbf{v}\\}" + }, + { + "bbox": [ + 313, + 411, + 555, + 509 + ], + "type": "text", + "content": ". Here, we introduce a hyper-parameter " + }, + { + "bbox": [ + 313, + 411, + 555, + 509 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 313, + 411, + 555, + 509 + ], + "type": "text", + "content": " to control the normalization scale. Note that we reformulate the WSP computation from [35] into Eq. (5) to better understand the gradient filtering process of WS during local training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 516, + 556, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 516, + 556, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 516, + 556, + 625 + ], + "type": "text", + "content": "Gradient filtering by WS WS plays an important role in reducing the learning diversity of local models, which is one of the most critical issues in FL. To understand its impact, we compare the gradient of a loss function " + }, + { + "bbox": [ + 313, + 516, + 556, + 625 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 313, + 516, + 556, + 625 + ], + "type": "text", + "content": " with respect to the WSP vector " + }, + { + "bbox": [ + 313, + 516, + 556, + 625 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{w}}_{n,m}" + }, + { + "bbox": [ + 313, + 516, + 556, + 625 + ], + "type": "text", + "content": " to that of the PSP vector " + }, + { + "bbox": [ + 313, + 516, + 556, + 625 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{n,m}" + }, + { + "bbox": [ + 313, + 516, + 556, + 625 + ], + "type": "text", + "content": ". When WS is applied, the gradient is processed by an additional series of projections to remove specific components that may interfere with effective FL. Specifically, the gradient with respect to the PSP vector " + }, + { + "bbox": [ + 313, + 516, + 556, + 625 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{n,m}" + }, + { + "bbox": [ + 313, + 516, + 556, + 625 + ], + "type": "text", + "content": " is obtained by" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 319, + 631, + 555, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 631, + 555, + 658 + ], + "spans": [ + { + "bbox": [ + 319, + 631, + 555, + 658 + ], + "type": "interline_equation", + "content": "\\frac {\\partial \\mathcal {L}}{\\partial \\mathbf {w} _ {n , m}} = \\frac {\\rho}{\\sigma (\\mathbf {w} _ {n , m})} \\left(\\mathbf {I} - \\mathbf {P} _ {\\mathbf {1}}\\right) \\left(\\mathbf {I} - \\mathbf {P} _ {\\tilde {\\mathbf {w}} _ {n, m}}\\right) \\frac {\\partial \\mathcal {L}}{\\partial \\tilde {\\mathbf {w}} _ {n , m}}. (6)", + "image_path": "817f943b991db5611871cbd167e1134f3c32b8757d18d9cb4183bbede8beead8.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 664, + 555, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 664, + 555, + 703 + ], + "spans": [ + { + "bbox": [ + 313, + 664, + 555, + 703 + ], + "type": "text", + "content": "2Strictly speaking, convolution involves spatial structures aggregating information from local receptive fields. For simplicity, we omit spatial considerations and treat the convolution as a linear transformation. This abstraction can be easily generalized to multidimensional tensor structures." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 326, + 703, + 541, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 703, + 541, + 713 + ], + "spans": [ + { + "bbox": [ + 326, + 703, + 541, + 713 + ], + "type": "text", + "content": "3Section A of the supplementary document provides its derivation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4618" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 72, + 292, + 169 + ], + "blocks": [ + { + "bbox": [ + 60, + 72, + 292, + 169 + ], + "lines": [ + { + "bbox": [ + 60, + 72, + 292, + 169 + ], + "spans": [ + { + "bbox": [ + 60, + 72, + 292, + 169 + ], + "type": "image", + "image_path": "c5d4cedad4a30261ec4674529f88523fee3a26460aa03e52dbe3e1a257b88a7b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 174, + 280, + 186 + ], + "lines": [ + { + "bbox": [ + 69, + 174, + 280, + 186 + ], + "spans": [ + { + "bbox": [ + 69, + 174, + 280, + 186 + ], + "type": "text", + "content": "Figure 2. An example of gradient filtering process of WS." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 200, + 295, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 200, + 295, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 200, + 295, + 266 + ], + "type": "text", + "content": "The upstream gradient " + }, + { + "bbox": [ + 55, + 200, + 295, + 266 + ], + "type": "inline_equation", + "content": "\\frac{\\partial\\mathcal{L}}{\\partial\\tilde{\\mathbf{w}}_{n,m}}" + }, + { + "bbox": [ + 55, + 200, + 295, + 266 + ], + "type": "text", + "content": " is projected onto two subspaces, " + }, + { + "bbox": [ + 55, + 200, + 295, + 266 + ], + "type": "inline_equation", + "content": "\\mathrm{span}\\{\\tilde{\\mathbf{w}}_{n,m}\\}^{\\perp}" + }, + { + "bbox": [ + 55, + 200, + 295, + 266 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 200, + 295, + 266 + ], + "type": "inline_equation", + "content": "\\mathrm{span}\\{\\mathbf{1}\\}^{\\perp}" + }, + { + "bbox": [ + 55, + 200, + 295, + 266 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 200, + 295, + 266 + ], + "type": "inline_equation", + "content": "\\mathrm{span}\\{\\mathbf{v}\\}^{\\perp}" + }, + { + "bbox": [ + 55, + 200, + 295, + 266 + ], + "type": "text", + "content": " represents the orthogonal complement of " + }, + { + "bbox": [ + 55, + 200, + 295, + 266 + ], + "type": "inline_equation", + "content": "\\mathrm{span}\\{\\mathbf{v}\\}" + }, + { + "bbox": [ + 55, + 200, + 295, + 266 + ], + "type": "text", + "content": ". Since " + }, + { + "bbox": [ + 55, + 200, + 295, + 266 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{w}}_{n,m} \\perp \\mathbf{1}" + }, + { + "bbox": [ + 55, + 200, + 295, + 266 + ], + "type": "text", + "content": ", the upstream gradient " + }, + { + "bbox": [ + 55, + 200, + 295, + 266 + ], + "type": "inline_equation", + "content": "\\frac{\\partial\\mathcal{L}}{\\partial\\tilde{\\mathbf{w}}_{n,m}}" + }, + { + "bbox": [ + 55, + 200, + 295, + 266 + ], + "type": "text", + "content": " is ultimately projected onto " + }, + { + "bbox": [ + 55, + 200, + 295, + 266 + ], + "type": "inline_equation", + "content": "\\mathrm{span}\\{\\tilde{\\mathbf{w}}_{n,m}, \\mathbf{1}\\}^{\\perp}" + }, + { + "bbox": [ + 55, + 200, + 295, + 266 + ], + "type": "text", + "content": " after the two projections." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 267, + 296, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 267, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 56, + 267, + 296, + 531 + ], + "type": "text", + "content": "As illustrated in Figure 2, Eq. (6) represents a gradient filtering process that applies successive projections to remove undesirable components from the gradient with respect to LMPs. In FL, data heterogeneity causes local models to overfit their local data, which leads to inconsistent updates across clients. This inconsistency results in a discrepancy between the GMPs and the optimal parameters, known as client drift. WS mitigates this issue by projecting the gradient onto " + }, + { + "bbox": [ + 56, + 267, + 296, + 531 + ], + "type": "inline_equation", + "content": "\\mathrm{span}\\{\\tilde{\\mathbf{w}}_{n,m},\\mathbf{1}\\}^{\\perp}" + }, + { + "bbox": [ + 56, + 267, + 296, + 531 + ], + "type": "text", + "content": ", which effectively reduces the impact of client drift during training. The first projection eliminates the gradient component aligned with " + }, + { + "bbox": [ + 56, + 267, + 296, + 531 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{w}}_{n,m}" + }, + { + "bbox": [ + 56, + 267, + 296, + 531 + ], + "type": "text", + "content": ", which corresponds to the LMPs that are biased due to the non-i.i.d. data distribution. The second projection removes the mean component of mini-batch gradients, which can also be biased toward the local data distribution, compounding inconsistencies across clients. By filtering out both the parameter-aligned and mean components from the local gradient, WS preserves meaningful gradient directions essential for stable global model convergence. Additionally, this projection-based filtering introduces a regularization effect by constraining the model's deviation from its current state, which can improve the model's generalization ability." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": "Comparison with FedWon It is worth noting that FedWon [56], a recently proposed FL method, also employs WS to address data heterogeneity in multi-domain FL. However, a key difference lies in the parameters exchanged between clients and the server, which fundamentally alters the learning dynamics. FedWon transmits WSP, which explicitly normalizes parameters across clients to improve learning stability. However, this approach forces all clients' LMPUs to exhibit statistically similar characteristics, which potentially discards client-specific information crucial for local adaptation. In contrast, FedWSQ transmits PSP, preserving essential local information while implicitly mitigating harmful divergences through the gradient filtering process of WS, as shown in Figure 2. This design choice en" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": "ables FedWSQ to achieve a better balance between global stability and local information preservation. An experimental comparison between FedWon and FedWSQ will be discussed in Section 4.4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 127, + 451, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 127, + 451, + 140 + ], + "spans": [ + { + "bbox": [ + 313, + 127, + 451, + 140 + ], + "type": "text", + "content": "3.2. Quantization of LMPUs" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 144, + 555, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 555, + 289 + ], + "type": "text", + "content": "Neural network quantization often involves learning quantization parameters. However, in FL, transmitting additional quantization parameters or auxiliary information for each communication round would introduce significant resource overhead. Instead, we propose a fixed quantization function that eliminates the need to transmit additional quantization information, ensuring a lightweight and efficient communication process. A fundamental assumption in NUQ is that model parameters follow a normal distribution, as demonstrated in prior studies [7, 45, 54]. Since LMPUs represent the difference between an updated LMP and the previous GMP, they may also exhibit a normal distribution." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 297, + 556, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 556, + 500 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 556, + 500 + ], + "type": "text", + "content": "LMPU scaling Most quantization methods adopt the absmax strategy to manage the dynamic range of values [16, 33]. This approach scales an input tensor to the range " + }, + { + "bbox": [ + 313, + 297, + 556, + 500 + ], + "type": "inline_equation", + "content": "[-1,1]" + }, + { + "bbox": [ + 313, + 297, + 556, + 500 + ], + "type": "text", + "content": " by dividing each element by the maximum absolute value of the tensor. Within this constrained range, conventional quantization methods determine QLs and map input values to the nearest quantization point. However, these methods are highly sensitive to outliers, which can excessively expand the dynamic range of the target tensor. Such over-scaling often results in underflow, especially in extremely low-bit representation. To solve this problem, we use the standard deviation of the input tensor as a scaling factor. Unlike the absmax value, the standard deviation is more robust to outliers. Moreover, since our DANUQ is based on a probability model of a standard normal distribution, the standard deviation provides a more reliable approximation of the target dynamic range." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 500, + 556, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 500, + 556, + 634 + ], + "spans": [ + { + "bbox": [ + 313, + 500, + 556, + 634 + ], + "type": "text", + "content": "In FedWSQ, we use a shared global scaling factor to help maintain consistency in quantization. Instead of independently using scaling factors, we employ a collaborative approach in which each client contributes to a global scaling vector, ensuring that local standard deviations are taken into account. Specifically, each client " + }, + { + "bbox": [ + 313, + 500, + 556, + 634 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 500, + 556, + 634 + ], + "type": "text", + "content": " computes a scaling vector, defined as " + }, + { + "bbox": [ + 313, + 500, + 556, + 634 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_i = [s_{i,1},\\dots ,s_{i,L}]^T" + }, + { + "bbox": [ + 313, + 500, + 556, + 634 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 500, + 556, + 634 + ], + "type": "inline_equation", + "content": "s_{i,l}" + }, + { + "bbox": [ + 313, + 500, + 556, + 634 + ], + "type": "text", + "content": " denotes the standard deviation of the LMPUs for the " + }, + { + "bbox": [ + 313, + 500, + 556, + 634 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 500, + 556, + 634 + ], + "type": "text", + "content": "-th layer. The server then collects and aggregates " + }, + { + "bbox": [ + 313, + 500, + 556, + 634 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_i" + }, + { + "bbox": [ + 313, + 500, + 556, + 634 + ], + "type": "text", + "content": " from the participating clients " + }, + { + "bbox": [ + 313, + 500, + 556, + 634 + ], + "type": "inline_equation", + "content": "i\\in S" + }, + { + "bbox": [ + 313, + 500, + 556, + 634 + ], + "type": "text", + "content": ", and updates the global scaling vector " + }, + { + "bbox": [ + 313, + 500, + 556, + 634 + ], + "type": "inline_equation", + "content": "\\mathbf{s_g} = [s_{g,1},\\dots ,s_{g,L}]^T" + }, + { + "bbox": [ + 313, + 500, + 556, + 634 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 369, + 640, + 555, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 640, + 555, + 669 + ], + "spans": [ + { + "bbox": [ + 369, + 640, + 555, + 669 + ], + "type": "interline_equation", + "content": "\\mathbf {s} _ {\\mathrm {g}} \\leftarrow (1 - \\beta) \\mathbf {s} _ {\\mathrm {g}} + \\beta \\frac {1}{| \\mathcal {S} |} \\sum_ {i \\in \\mathcal {S}} \\mathbf {s} _ {i}, \\tag {7}", + "image_path": "b2484da57b5a21e586b4c1942ed4245b1a402883d80eeccd3119716a4271c95c.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": " is a momentum parameter that controls the update rate. This approach ensures that the global scaling factor smoothly adapts while mitigating local fluctuations. Once" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4619" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": "updated, the server distributes " + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{\\mathrm{g}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": " to the participating clients. Each client then divides its LMPU updates for the " + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": "-th layer by the corresponding scale " + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "inline_equation", + "content": "s_{\\mathrm{g},l}" + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": ", so that the normalized values can be regarded as samples from a standard normal distribution. Based on this assumption, we apply the proposed DANUQ scheme detailed in the next section." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 149, + 296, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 149, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 149, + 296, + 281 + ], + "type": "text", + "content": "QLs considering normal distribution To determine the optimal multi-bit QLs, we formulate an optimization problem that minimizes the expected quantization error under a normal distribution. We assume that a random variable " + }, + { + "bbox": [ + 55, + 149, + 296, + 281 + ], + "type": "inline_equation", + "content": "\\Delta w" + }, + { + "bbox": [ + 55, + 149, + 296, + 281 + ], + "type": "text", + "content": " from the normalized LMPUs follows a standard normal distribution. Since the distribution is symmetric around zero, we consider only its nonnegative half. Given a " + }, + { + "bbox": [ + 55, + 149, + 296, + 281 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 149, + 296, + 281 + ], + "type": "text", + "content": "-bit representation, let " + }, + { + "bbox": [ + 55, + 149, + 296, + 281 + ], + "type": "inline_equation", + "content": "Q = \\{q_{0}, q_{1}, \\ldots, q_{R}\\}" + }, + { + "bbox": [ + 55, + 149, + 296, + 281 + ], + "type": "text", + "content": " be the set of QLs, where " + }, + { + "bbox": [ + 55, + 149, + 296, + 281 + ], + "type": "inline_equation", + "content": "R = 2^{B - 1} - 1" + }, + { + "bbox": [ + 55, + 149, + 296, + 281 + ], + "type": "text", + "content": ". Here, we assume that " + }, + { + "bbox": [ + 55, + 149, + 296, + 281 + ], + "type": "inline_equation", + "content": "q_{r}" + }, + { + "bbox": [ + 55, + 149, + 296, + 281 + ], + "type": "text", + "content": " values are sorted in ascending order and set " + }, + { + "bbox": [ + 55, + 149, + 296, + 281 + ], + "type": "inline_equation", + "content": "q_{0} = 0" + }, + { + "bbox": [ + 55, + 149, + 296, + 281 + ], + "type": "text", + "content": " as a fixed value. Then, the quantization boundaries are defined as" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 287, + 295, + 331 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 287, + 295, + 331 + ], + "spans": [ + { + "bbox": [ + 106, + 287, + 295, + 331 + ], + "type": "interline_equation", + "content": "u _ {r} = \\left\\{ \\begin{array}{l l} q _ {0}, & \\text {i f} r = 0, \\\\ \\frac {q _ {r - 1} + q _ {r}}{2}, & \\text {i f} 1 \\leq r \\leq R, \\\\ + \\infty , & \\text {i f} r = R + 1. \\end{array} \\right. \\tag {8}", + "image_path": "72c085215e84d3b92a6fc726662146a01cc662854c2c3e7b2c583bad692f202c.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 338, + 296, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 296, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 296, + 374 + ], + "type": "text", + "content": "A full-precision value in the range " + }, + { + "bbox": [ + 55, + 338, + 296, + 374 + ], + "type": "inline_equation", + "content": "[u_r, u_{r+1})" + }, + { + "bbox": [ + 55, + 338, + 296, + 374 + ], + "type": "text", + "content": " is quantized into the QL " + }, + { + "bbox": [ + 55, + 338, + 296, + 374 + ], + "type": "inline_equation", + "content": "q_r" + }, + { + "bbox": [ + 55, + 338, + 296, + 374 + ], + "type": "text", + "content": ". The optimal QLs are determined by minimizing the expected quantization error, formulated as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 380, + 295, + 412 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 380, + 295, + 412 + ], + "spans": [ + { + "bbox": [ + 69, + 380, + 295, + 412 + ], + "type": "interline_equation", + "content": "\\mathbb {E} \\left[ \\left(\\Delta w - \\Delta \\bar {w}\\right) ^ {2} \\right] = \\sum_ {r = 0} ^ {R} \\int_ {u _ {r}} ^ {u _ {r + 1}} \\left(x - q _ {r}\\right) ^ {2} p (x) d x, \\tag {9}", + "image_path": "e6f38ae9ec3af3a1fbf19b862e81d9168b4a734be4ed6a2a1d3cc50e052fcf1c.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 418, + 296, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 296, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 296, + 455 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 418, + 296, + 455 + ], + "type": "inline_equation", + "content": "\\Delta \\bar{w} \\in Q" + }, + { + "bbox": [ + 55, + 418, + 296, + 455 + ], + "type": "text", + "content": " represents the quantized value, and " + }, + { + "bbox": [ + 55, + 418, + 296, + 455 + ], + "type": "inline_equation", + "content": "p(x)" + }, + { + "bbox": [ + 55, + 418, + 296, + 455 + ], + "type": "text", + "content": " is the PDF of the standard normal distribution. From (9), we can derive the following optimization problem as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 460, + 294, + 558 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 460, + 294, + 558 + ], + "spans": [ + { + "bbox": [ + 66, + 460, + 294, + 558 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {\\left\\{q _ {1}, \\dots , q _ {R} \\right\\}} \\left\\{\\frac {1}{2} \\left(q _ {R} ^ {2} + 1\\right) - \\sqrt {\\frac {2}{\\pi}} q _ {1} e ^ {- \\frac {u _ {1} ^ {2}}{2}} - \\frac {1}{2} q _ {1} ^ {2} \\operatorname {e r f} \\left(\\frac {u _ {1}}{\\sqrt {2}}\\right) \\right. \\\\ + \\sqrt {\\frac {2}{\\pi}} \\sum_ {r = 1} ^ {R - 1} (q _ {r} - q _ {r + 1}) e ^ {- \\frac {u _ {r + 1} ^ {2}}{2}} \\\\ + \\frac {1}{2} \\sum_ {r = 1} ^ {R - 1} \\left(q _ {r} ^ {2} - q _ {r + 1} ^ {2}\\right) \\operatorname {e r f} \\left(\\frac {u _ {r + 1}}{\\sqrt {2}}\\right) \\}, \\tag {10} \\\\ \\end{array}", + "image_path": "c5fec929828ab765f441d68cb7d0e39b9f4a41beb833e23b7c8a2bf8396143d5.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 565, + 296, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 565, + 296, + 650 + ], + "spans": [ + { + "bbox": [ + 55, + 565, + 296, + 650 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 565, + 296, + 650 + ], + "type": "inline_equation", + "content": "\\operatorname{erf}(\\cdot)" + }, + { + "bbox": [ + 55, + 565, + 296, + 650 + ], + "type": "text", + "content": " denotes the error function. A closed-form solution to Eq. (22) is intractable due to the presence of nonlinear terms, including the Gaussian integral and the error function, which lack simple analytical inverses, and the interdependent QLs due to boundary conditions. As an alternative, we employ a brute-force search algorithm to numerically determine the optimal QLs. The optimal QLs" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 72, + 409, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 409, + 83 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 409, + 83 + ], + "type": "text", + "content": "Algorithm 1 FedWSQ" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 86, + 555, + 116 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 86, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 315, + 86, + 555, + 116 + ], + "type": "text", + "content": "Input: Initial global parameters " + }, + { + "bbox": [ + 315, + 86, + 555, + 116 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathrm{g}}^{0}" + }, + { + "bbox": [ + 315, + 86, + 555, + 116 + ], + "type": "text", + "content": ", initial scales " + }, + { + "bbox": [ + 315, + 86, + 555, + 116 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{\\mathrm{g}}^{0}" + }, + { + "bbox": [ + 315, + 86, + 555, + 116 + ], + "type": "text", + "content": ", global communication round " + }, + { + "bbox": [ + 315, + 86, + 555, + 116 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 315, + 86, + 555, + 116 + ], + "type": "text", + "content": ", local training iteration " + }, + { + "bbox": [ + 315, + 86, + 555, + 116 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 315, + 86, + 555, + 116 + ], + "type": "text", + "content": ", number of total clients " + }, + { + "bbox": [ + 315, + 86, + 555, + 116 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 315, + 86, + 555, + 116 + ], + "type": "text", + "content": ", number of model layers " + }, + { + "bbox": [ + 315, + 86, + 555, + 116 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 315, + 86, + 555, + 116 + ], + "type": "text", + "content": ", learning rate " + }, + { + "bbox": [ + 315, + 86, + 555, + 116 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 315, + 86, + 555, + 116 + ], + "type": "text", + "content": ", hyperparameter " + }, + { + "bbox": [ + 315, + 86, + 555, + 116 + ], + "type": "inline_equation", + "content": "\\beta" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 116, + 358, + 124 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 116, + 358, + 124 + ], + "spans": [ + { + "bbox": [ + 315, + 116, + 358, + 124 + ], + "type": "text", + "content": "Server-side:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 125, + 383, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 125, + 383, + 133 + ], + "spans": [ + { + "bbox": [ + 315, + 125, + 383, + 133 + ], + "type": "text", + "content": "for " + }, + { + "bbox": [ + 315, + 125, + 383, + 133 + ], + "type": "inline_equation", + "content": "t = 1,\\dots ,T" + }, + { + "bbox": [ + 315, + 125, + 383, + 133 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 331, + 133, + 483, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 133, + 483, + 143 + ], + "spans": [ + { + "bbox": [ + 331, + 133, + 483, + 143 + ], + "type": "text", + "content": "Randomly sample a subset of clients " + }, + { + "bbox": [ + 331, + 133, + 483, + 143 + ], + "type": "inline_equation", + "content": "S_{t} \\subseteq [N]" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 331, + 144, + 474, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 144, + 474, + 152 + ], + "spans": [ + { + "bbox": [ + 331, + 144, + 474, + 152 + ], + "type": "text", + "content": "Set the bit-precision " + }, + { + "bbox": [ + 331, + 144, + 474, + 152 + ], + "type": "inline_equation", + "content": "B_{i}" + }, + { + "bbox": [ + 331, + 144, + 474, + 152 + ], + "type": "text", + "content": " for the client " + }, + { + "bbox": [ + 331, + 144, + 474, + 152 + ], + "type": "inline_equation", + "content": "i \\in S_{t}" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 331, + 152, + 452, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 152, + 452, + 162 + ], + "spans": [ + { + "bbox": [ + 331, + 152, + 452, + 162 + ], + "type": "text", + "content": "Send " + }, + { + "bbox": [ + 331, + 152, + 452, + 162 + ], + "type": "inline_equation", + "content": "(\\mathbf{W}_{\\mathrm{g}}^{t - 1},\\mathbf{s}_{\\mathrm{g}}^{t - 1})" + }, + { + "bbox": [ + 331, + 152, + 452, + 162 + ], + "type": "text", + "content": " to client " + }, + { + "bbox": [ + 331, + 152, + 452, + 162 + ], + "type": "inline_equation", + "content": "i\\in S_{t}" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 331, + 163, + 456, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 163, + 456, + 171 + ], + "spans": [ + { + "bbox": [ + 331, + 163, + 456, + 171 + ], + "type": "text", + "content": "for each client " + }, + { + "bbox": [ + 331, + 163, + 456, + 171 + ], + "type": "inline_equation", + "content": "i\\in S_t" + }, + { + "bbox": [ + 331, + 163, + 456, + 171 + ], + "type": "text", + "content": " , in parallel do" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 334, + 171, + 538, + 184 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 171, + 538, + 184 + ], + "spans": [ + { + "bbox": [ + 334, + 171, + 538, + 184 + ], + "type": "interline_equation", + "content": "\\left(\\Delta \\overline {{\\mathbf {W}}} _ {i} ^ {t}, \\mathbf {s} _ {i} ^ {t}\\right) \\leftarrow \\text {C l i e n t L o c a l T r a i n i n g} \\left(\\mathbf {W} _ {\\mathrm {g}} ^ {t - 1}, \\mathbf {s} _ {\\mathrm {g}} ^ {t - 1}, i, B _ {i}\\right)", + "image_path": "8929252eb7b56d8ab9656208a5e75183acea833f7775705120804ee57c3cf79c.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 331, + 184, + 345, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 184, + 345, + 190 + ], + "spans": [ + { + "bbox": [ + 331, + 184, + 345, + 190 + ], + "type": "text", + "content": "end" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 331, + 193, + 376, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 193, + 376, + 201 + ], + "spans": [ + { + "bbox": [ + 331, + 193, + 376, + 201 + ], + "type": "text", + "content": "for " + }, + { + "bbox": [ + 331, + 193, + 376, + 201 + ], + "type": "inline_equation", + "content": "i\\in S_t" + }, + { + "bbox": [ + 331, + 193, + 376, + 201 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 335, + 202, + 518, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 202, + 518, + 214 + ], + "spans": [ + { + "bbox": [ + 335, + 202, + 518, + 214 + ], + "type": "text", + "content": "Obtain dequantized parameters " + }, + { + "bbox": [ + 335, + 202, + 518, + 214 + ], + "type": "inline_equation", + "content": "\\Delta_i^t" + }, + { + "bbox": [ + 335, + 202, + 518, + 214 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 335, + 202, + 518, + 214 + ], + "type": "inline_equation", + "content": "(\\Delta \\overline{\\mathbf{W}}_i^t, \\mathbf{s}_i^t)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 331, + 214, + 345, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 214, + 345, + 221 + ], + "spans": [ + { + "bbox": [ + 331, + 214, + 345, + 221 + ], + "type": "text", + "content": "end" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 331, + 222, + 474, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 222, + 474, + 233 + ], + "spans": [ + { + "bbox": [ + 331, + 222, + 474, + 233 + ], + "type": "text", + "content": "Aggregate local updates: " + }, + { + "bbox": [ + 331, + 222, + 474, + 233 + ], + "type": "inline_equation", + "content": "\\Delta^t\\gets \\sum_{i\\in S_t}\\Delta_i^t" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 331, + 234, + 489, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 234, + 489, + 245 + ], + "spans": [ + { + "bbox": [ + 331, + 234, + 489, + 245 + ], + "type": "text", + "content": "Update global parameters: " + }, + { + "bbox": [ + 331, + 234, + 489, + 245 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathrm{g}}^{t}\\gets \\mathbf{W}_{\\mathrm{g}}^{t - 1} + \\Delta^{t}" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 331, + 246, + 531, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 246, + 531, + 258 + ], + "spans": [ + { + "bbox": [ + 331, + 246, + 531, + 258 + ], + "type": "text", + "content": "Update global scales: " + }, + { + "bbox": [ + 331, + 246, + 531, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{\\mathrm{g}}^{t}\\gets (1 - \\beta)\\mathbf{s}_{\\mathrm{g}}^{t - 1} + \\beta \\frac{1}{|\\mathcal{S}_{t}|}\\sum_{i\\in \\mathcal{S}_{t}}\\mathbf{s}_{i}^{t}" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 258, + 329, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 258, + 329, + 266 + ], + "spans": [ + { + "bbox": [ + 315, + 258, + 329, + 266 + ], + "type": "text", + "content": "end" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 315, + 277, + 436, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 277, + 436, + 287 + ], + "spans": [ + { + "bbox": [ + 315, + 277, + 436, + 287 + ], + "type": "text", + "content": "ClientLocalTraining " + }, + { + "bbox": [ + 315, + 277, + 436, + 287 + ], + "type": "inline_equation", + "content": "(\\mathbf{W}_{\\mathrm{g}},\\mathbf{s}_{\\mathrm{g}},i,I,B)" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 315, + 288, + 446, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 288, + 446, + 296 + ], + "spans": [ + { + "bbox": [ + 315, + 288, + 446, + 296 + ], + "type": "text", + "content": "Initialize local parameters: " + }, + { + "bbox": [ + 315, + 288, + 446, + 296 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_j^0\\gets \\mathbf{W}_{\\mathrm{g}}" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 315, + 296, + 497, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 296, + 497, + 305 + ], + "spans": [ + { + "bbox": [ + 315, + 296, + 497, + 305 + ], + "type": "text", + "content": "Set DANUQ function " + }, + { + "bbox": [ + 315, + 296, + 497, + 305 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}(\\cdot)" + }, + { + "bbox": [ + 315, + 296, + 497, + 305 + ], + "type": "text", + "content": " based on the bit-precision " + }, + { + "bbox": [ + 315, + 296, + 497, + 305 + ], + "type": "inline_equation", + "content": "B" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 315, + 306, + 387, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 306, + 387, + 315 + ], + "spans": [ + { + "bbox": [ + 315, + 306, + 387, + 315 + ], + "type": "text", + "content": "for " + }, + { + "bbox": [ + 315, + 306, + 387, + 315 + ], + "type": "inline_equation", + "content": "k = 1,\\ldots ,K" + }, + { + "bbox": [ + 315, + 306, + 387, + 315 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 331, + 315, + 431, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 315, + 431, + 326 + ], + "spans": [ + { + "bbox": [ + 331, + 315, + 431, + 326 + ], + "type": "text", + "content": "Apply WS to " + }, + { + "bbox": [ + 331, + 315, + 431, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_i^{k - 1}" + }, + { + "bbox": [ + 331, + 315, + 431, + 326 + ], + "type": "text", + "content": " using (5)" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 331, + 327, + 509, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 327, + 509, + 338 + ], + "spans": [ + { + "bbox": [ + 331, + 327, + 509, + 338 + ], + "type": "text", + "content": "Compute unbiased gradient of local loss " + }, + { + "bbox": [ + 331, + 327, + 509, + 338 + ], + "type": "inline_equation", + "content": "\\nabla f_{i}(\\mathbf{W}_{i}^{k - 1})" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 331, + 338, + 526, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 338, + 526, + 350 + ], + "spans": [ + { + "bbox": [ + 331, + 338, + 526, + 350 + ], + "type": "text", + "content": "Update local parameters: " + }, + { + "bbox": [ + 331, + 338, + 526, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_i^k\\gets \\mathbf{W}_i^{k - 1} - \\eta \\nabla f_i(\\mathbf{W}_i^{k - 1})" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 315, + 350, + 329, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 350, + 329, + 357 + ], + "spans": [ + { + "bbox": [ + 315, + 350, + 329, + 357 + ], + "type": "text", + "content": "end" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 315, + 359, + 466, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 359, + 466, + 369 + ], + "spans": [ + { + "bbox": [ + 315, + 359, + 466, + 369 + ], + "type": "text", + "content": "Compute local update: " + }, + { + "bbox": [ + 315, + 359, + 466, + 369 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{W}_i\\gets \\mathbf{W}_i^K -\\mathbf{W}_g" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 315, + 370, + 422, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 370, + 422, + 378 + ], + "spans": [ + { + "bbox": [ + 315, + 370, + 422, + 378 + ], + "type": "text", + "content": "for each layer " + }, + { + "bbox": [ + 315, + 370, + 422, + 378 + ], + "type": "inline_equation", + "content": "l = 1,\\dots ,L" + }, + { + "bbox": [ + 315, + 370, + 422, + 378 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 331, + 379, + 533, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 379, + 533, + 390 + ], + "spans": [ + { + "bbox": [ + 331, + 379, + 533, + 390 + ], + "type": "text", + "content": "Compute quantized local update: " + }, + { + "bbox": [ + 331, + 379, + 533, + 390 + ], + "type": "inline_equation", + "content": "\\Delta \\overline{\\mathbf{W}}_{i,l}\\gets \\mathcal{Q}(\\Delta \\mathbf{W}_{i,l} / s_{\\mathrm{g},l})" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 331, + 390, + 492, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 390, + 492, + 399 + ], + "spans": [ + { + "bbox": [ + 331, + 390, + 492, + 399 + ], + "type": "text", + "content": "Calculate local scale factors: " + }, + { + "bbox": [ + 331, + 390, + 492, + 399 + ], + "type": "inline_equation", + "content": "s_{i,l} \\gets \\mathrm{std}(\\Delta \\mathbf{W}_{i,l})" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 315, + 400, + 329, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 400, + 329, + 407 + ], + "spans": [ + { + "bbox": [ + 315, + 400, + 329, + 407 + ], + "type": "text", + "content": "end" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 315, + 409, + 430, + 420 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 409, + 430, + 420 + ], + "spans": [ + { + "bbox": [ + 315, + 409, + 430, + 420 + ], + "type": "interline_equation", + "content": "\\Delta \\overline {{\\mathbf {W}}} _ {i} := (\\Delta \\mathbf {W} _ {i, 1}, \\dots , \\Delta \\mathbf {W} _ {i, L})", + "image_path": "bdafc7163f9106fe46e0140250f4a1a58f599f3dc5b5f1e78be73b96d97aa569.jpg" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 315, + 420, + 389, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 420, + 389, + 430 + ], + "spans": [ + { + "bbox": [ + 315, + 420, + 389, + 430 + ], + "type": "interline_equation", + "content": "\\mathbf {s} _ {i} := \\left[ s _ {i, 1}, \\dots , s _ {i, L} \\right]", + "image_path": "f8e86530a08a8275d0fd2818779f503969f5b9009efb7ff66e75697f5fca8b8d.jpg" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 315, + 430, + 378, + 440 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 430, + 378, + 440 + ], + "spans": [ + { + "bbox": [ + 315, + 430, + 378, + 440 + ], + "type": "interline_equation", + "content": "\\operatorname {r e t u r n} \\left(\\Delta \\overline {{\\mathbf {W}}} _ {i}, \\mathbf {s} _ {i}\\right)", + "image_path": "0cd2380df4c0efe89a482a1ebca697e523438bb047ae25826e5f385573c1fa76.jpg" + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 313, + 462, + 555, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 546 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 546 + ], + "type": "text", + "content": "for different bit-widths are obtained as follows: 1) 1-bit" + }, + { + "bbox": [ + 313, + 462, + 555, + 546 + ], + "type": "inline_equation", + "content": "^6" + }, + { + "bbox": [ + 313, + 462, + 555, + 546 + ], + "type": "text", + "content": ": [-0.798, 0.798]; 2) 2-bit: [-1.224, 0, 0.765, 1.724]; 3) 4-bit: [-2.654, -1.974, -1.508, -1.149, -0.834, -0.544, -0.269, 0, 0.230, 0.465, 0.708, 0.966, 1.248, 1.568, 1.968, 2.649]. These QLs minimize quantization error under the normal distribution, which provides an effective trade-off between precision and computational efficiency." + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 313, + 555, + 555, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 555, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 555, + 687 + ], + "type": "text", + "content": "Mixed channel-bit allocation To further enhance communication efficiency in FL, we propose an adaptive mixed-precision strategy in which each client dynamically selects its bit representation based on its communication channel conditions. This approach effectively balances model update quality and transmission cost, reducing overall bandwidth utilization. To validate this strategy, we conduct two simulation setups: 1) Fixed-bit allocation (FBA), where each client keeps a constant bit-width selected from " + }, + { + "bbox": [ + 313, + 555, + 555, + 687 + ], + "type": "inline_equation", + "content": "\\{1,2,4\\}" + }, + { + "bbox": [ + 313, + 555, + 555, + 687 + ], + "type": "text", + "content": " throughout training, and 2) dynamic-bit allocation (DBA), where each client is assigned a randomly" + } + ] + } + ], + "index": 45 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 654, + 274, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 654, + 274, + 665 + ], + "spans": [ + { + "bbox": [ + 66, + 654, + 274, + 665 + ], + "type": "text", + "content": "4The full derivation is provided in the supplementary document." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "5We discretize the search space of possible QLs and perform an exhaustive evaluation of the cost function to identify the configuration that minimizes the quantization error. To improve computational efficiency, we restrict the search space to a reasonable range based on empirical observations and employ parallel processing to accelerate evaluation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 693, + 554, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 554, + 713 + ], + "type": "text", + "content": "For the 1-bit case, the QLs consist of only two values. Thus, the constraint, " + }, + { + "bbox": [ + 313, + 693, + 554, + 713 + ], + "type": "inline_equation", + "content": "q_{0} = 0" + }, + { + "bbox": [ + 313, + 693, + 554, + 713 + ], + "type": "text", + "content": ", is omitted to allow optimal placement of both QLs." + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4620" + } + ] + } + ], + "index": 47 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 85, + 536, + 316 + ], + "blocks": [ + { + "bbox": [ + 76, + 70, + 533, + 83 + ], + "lines": [ + { + "bbox": [ + 76, + 70, + 533, + 83 + ], + "spans": [ + { + "bbox": [ + 76, + 70, + 533, + 83 + ], + "type": "text", + "content": "Table 1. FL performance on the three benchmark datasets with a " + }, + { + "bbox": [ + 76, + 70, + 533, + 83 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 76, + 70, + 533, + 83 + ], + "type": "text", + "content": " participation rate over 100 clients for " + }, + { + "bbox": [ + 76, + 70, + 533, + 83 + ], + "type": "inline_equation", + "content": "\\alpha \\in \\{0.1, 0.3, 0.6\\}" + }, + { + "bbox": [ + 76, + 70, + 533, + 83 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 73, + 85, + 536, + 316 + ], + "lines": [ + { + "bbox": [ + 73, + 85, + 536, + 316 + ], + "spans": [ + { + "bbox": [ + 73, + 85, + 536, + 316 + ], + "type": "table", + "html": "
Method#bitsCIFAR-10CIFAR-100Tiny-ImageNet
α = 0.10.30.6i.i.d.0.10.30.6i.i.d.0.10.30.6i.i.d.
FedAvg [31]3270.6582.5285.4589.2543.4547.3949.0448.6630.3734.6536.5437.39
FedProx [27]3270.8983.9186.6088.8143.7848.2648.4948.5429.9935.4437.1437.43
FedAvgM [13]3278.3085.2787.9490.0047.7052.0453.0053.4032.0037.6339.6641.02
FedADAM [36]3272.5881.7385.9788.4546.8353.4355.9657.9334.7739.9743.2746.26
FedDyn [1]3283.6388.2889.2091.0853.8356.4056.9056.9337.3240.9242.7842.65
FedMLB [21]3271.1282.4086.2389.2549.7254.1655.1756.4533.4539.0640.5541.46
FedLC [52]3274.0383.8086.2088.3543.8648.5048.6147.8131.7334.5536.7238.06
FedNTD [26]3273.2983.3486.2689.3645.0848.7350.3950.9034.3636.4439.0640.60
FedSmoo [43]3278.6180.0785.0386.4842.0244.9745.3249.4323.0927.1429.7034.96
FedDecorr [39]3275.4483.0184.4988.4044.0549.3550.4449.4031.1234.1335.5936.36
FedWon [56]3270.8978.3682.4486.5645.5551.6853.6055.2830.5236.2038.5940.73
FedRCL [38]3283.4588.1989.7491.5558.2662.2163.9565.1137.8644.9446.7547.96
FedACG [20]3283.6289.1290.3891.7558.1462.8061.8562.4639.7545.4648.2250.49
FedPAQ [37]466.8178.5281.3685.4638.0642.4243.4644.3528.5431.9333.5634.98
157.8271.2772.3679.4630.0734.1638.2437.8524.7429.4830.7433.81
FedHQ+ [5]466.8978.4281.4285.5438.2242.6643.1044.8628.5932.1834.2435.69
158.1470.6873.1679.0431.1134.1137.5039.1724.7929.1831.1433.50
FedWSQFedWS (32)85.9489.8191.0092.4764.1468.1969.3170.1447.0552.8454.2355.26
486.8489.8491.1192.2164.3467.3168.4168.5346.5150.4951.9652.13
184.7988.6289.4991.1562.0565.1466.1966.1145.1149.9351.1151.77
FBA (2.33)86.1689.6090.4491.7163.0766.1667.0867.1746.6250.6951.4351.68
DBA (2.33)85.2089.2390.2691.7863.2866.3267.2667.2246.5050.2451.5551.80
", + "image_path": "9e39e66dfee90a50f8aba2cc88e432892b55b588823fc560e9718aa3d3ad19b0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 326, + 294, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 326, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 294, + 387 + ], + "type": "text", + "content": "selected bit-width at every communication round. With bit-width selection following a uniform distribution over " + }, + { + "bbox": [ + 55, + 326, + 294, + 387 + ], + "type": "inline_equation", + "content": "\\{1,2,4\\}" + }, + { + "bbox": [ + 55, + 326, + 294, + 387 + ], + "type": "text", + "content": ", the expected bit-width per client is approximately 2.3 bits. The entire FedWSQ framework, including the mixed-precision strategy, is summarized in Algorithm 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 400, + 137, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 400, + 137, + 414 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 137, + 414 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 421, + 296, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 421, + 296, + 494 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 296, + 494 + ], + "type": "text", + "content": "In this section, we evaluate the proposed FedWSQ on standard FL benchmarks, compare its performance against various FL methods on both i.i.d. and non-i.i.d. data conditions, and offer a further analysis of FedWSQ. The convergence plots obtained from the experiments are provided in Section D of the supplementary document." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 503, + 170, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 170, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 170, + 517 + ], + "type": "text", + "content": "4.1. Experimental setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "content": "For experiments, we use three standard benchmark datasets: CIFAR-10 [23], CIFAR-100 [23], and Tiny-ImageNet [25]. CIFAR-10 and CIFAR-100 contain 60,000 images divided into 10 and 100 classes, respectively. Tiny-ImageNet is composed of 200 classes, providing a more complex image classification task. For i.i.d. settings, training samples are randomly assigned to clients without replacement. For the non-i.i.d. setting, we sample label ratios from a Dirichlet distribution with a concentration parameter " + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\alpha \\in \\{0.1, 0.3, 0.6\\}" + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "content": ", following Hsu et al. [13]. Lower values of " + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "content": " (e.g., 0.1) indicate higher data heterogeneity, whereas larger values (e.g., 0.6) and the i.i.d. case correspond to more homogeneous distributions. The participation rate is set to " + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "content": " among 100 clients. Model accuracy is evaluated on the test set of each dataset at the 1,000th communication round. To ensure stable evaluations, we report accuracy" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 327, + 555, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 555, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 555, + 494 + ], + "type": "text", + "content": "using an exponential moving average with a smoothing parameter of 0.9. To validate the effectiveness of FedWSQ, we integrate it with FedAvg [31]. Note that FedWSQ can be incorporated into any existing FL algorithm. For clarity, we refer to the proposed method without DANUQ as FedWS. For quantized FL settings, we consider three bit-allocation strategies: constant allocation, FBA, and DBA. In the constant allocation settings, a fixed bit-width is assigned to all clients. We compare FedWSQ against various FL algorithms, including FedAvg [31], FedProx [27], FedAvgM [13], FedADAM [36], FedDyn [1], FedMLB [21], FedLC [52], FedNTD [26], FedSmoo [43], FedDecorr [39], FedWon [56], FedRCL [38], FedACG [20], FedPAQ [37], and FedHQ+ [5]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 502, + 444, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 444, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 444, + 514 + ], + "type": "text", + "content": "4.2. Implementation details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 520, + 555, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 555, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 555, + 651 + ], + "type": "text", + "content": "We follow most of the implementation setups and evaluation protocols from [1, 20, 38, 48]. The ResNet-18 architecture [11] is adopted as the backbone network. Consistent with standard practice in FL [12], all BN [15] layers in ResNet-18 are replaced with GN [47] layers. Following the recommendation of [35], WS is applied before each GN layer to achieve optimal performance. The hyper-parameter of WS is set to " + }, + { + "bbox": [ + 313, + 520, + 555, + 651 + ], + "type": "inline_equation", + "content": "\\rho = 0.001" + }, + { + "bbox": [ + 313, + 520, + 555, + 651 + ], + "type": "text", + "content": ", and the momentum for updating the scaling vector is set to " + }, + { + "bbox": [ + 313, + 520, + 555, + 651 + ], + "type": "inline_equation", + "content": "\\beta = 0.1" + }, + { + "bbox": [ + 313, + 520, + 555, + 651 + ], + "type": "text", + "content": ". Additional information on the hyper-parameter selection is provided in Section C of the supplementary document." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 659, + 517, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 517, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 517, + 673 + ], + "type": "text", + "content": "4.3. Comparison with existing FL methods" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "Table 1 presents the performance comparison of the proposed FedWSQ against various existing FL methods on the three benchmark datasets. The bit-width settings for each" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4621" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 130, + 96, + 480, + 292 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 2. Comparison of UQ and DANUQ methods with or without WS on the three benchmark datasets with a " + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": " participation rate over 100 clients for " + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "inline_equation", + "content": "\\alpha \\in \\{0.1, 0.3, 0.6\\}" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 130, + 96, + 480, + 292 + ], + "lines": [ + { + "bbox": [ + 130, + 96, + 480, + 292 + ], + "spans": [ + { + "bbox": [ + 130, + 96, + 480, + 292 + ], + "type": "table", + "html": "
DatasetWS#bitsα = 0.1α = 0.3α = 0.6i.i.d.
UQDANUQUQDANUQUQDANUQUQDANUQ
CIFAR-10X157.8272.4871.2780.8872.3683.5279.4686.92
262.3965.1974.3578.6176.7982.5982.2586.11
466.8172.8778.5284.6281.3686.2685.4688.66
✓175.4884.7980.1588.6282.5389.4985.3791.15
278.3685.9283.0288.9884.1090.3586.9591.50
482.7886.8486.3789.8487.5891.1189.8792.21
CIFAR-100X130.0740.5134.1646.6238.2448.6737.8548.55
233.5240.5837.6645.0740.4746.3140.9046.06
438.0643.5942.4248.3343.4649.2944.3548.76
✓145.5962.0550.1065.1452.1366.1954.5866.11
249.6863.3454.2265.9656.0167.1958.7267.82
456.8564.3460.2167.3161.9868.4162.9868.53
Tiny-ImageNetX124.7428.9529.4833.1430.7436.7533.8138.95
227.0230.3331.6034.1233.5836.634.8735.41
428.5431.1431.9335.4233.5636.6634.9838.13
✓132.7345.1137.6349.9339.6051.1141.9851.77
236.1645.9640.9050.1742.5151.1244.8951.12
440.4646.5144.5850.4945.5551.9646.9352.13
", + "image_path": "757f1f071868a411527ade7a472deca08325914e7ff84e22940205653c3c80e8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 72, + 325, + 279, + 427 + ], + "blocks": [ + { + "bbox": [ + 55, + 300, + 295, + 323 + ], + "lines": [ + { + "bbox": [ + 55, + 300, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 295, + 323 + ], + "type": "text", + "content": "Table 3. Comparison of NUQ methods using WS on CIFAR-10 and CIFAR-100 datasets for " + }, + { + "bbox": [ + 55, + 300, + 295, + 323 + ], + "type": "inline_equation", + "content": "\\alpha \\in \\{0.1, 0.3, 0.6\\}" + }, + { + "bbox": [ + 55, + 300, + 295, + 323 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 72, + 325, + 279, + 427 + ], + "lines": [ + { + "bbox": [ + 72, + 325, + 279, + 427 + ], + "spans": [ + { + "bbox": [ + 72, + 325, + 279, + 427 + ], + "type": "table", + "html": "
#bitsMethodCIFAR-10CIFAR-100
α = 0.10.30.60.10.30.6
1NF/FP24.030.628.87.08.38.1
DANUQ84.888.689.562.165.166.2
2NF57.873.880.041.249.153.0
FP45.057.558.625.332.034.2
DANUQ85.989.090.463.366.067.2
4NF86.689.791.064.866.768.4
FP87.190.091.064.367.068.4
DANUQ86.889.891.164.367.368.4
", + "image_path": "e85414b285a32afc06968838588186767798b1323243a3051a62e4ac3b3a770e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 437, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 437, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 54, + 437, + 295, + 653 + ], + "type": "text", + "content": "FL method are indicated in the table. Compared to FL methods that use quantization, such as FedPAQ and FedHQ+, FedWSQ demonstrates superior performance for 4-bit and 1-bit settings. Also, FedWSQ outperforms FL methods with full-precision representation, including SOTA methods such as FedACG and FedRCL. Notably, even with a 1-bit representation, FedWSQ demonstrates competitive performance on CIFAR-10 and achieves superior results on CIFAR-100 and Tiny/ImageNet. For example, in a highly heterogeneous setting " + }, + { + "bbox": [ + 54, + 437, + 295, + 653 + ], + "type": "inline_equation", + "content": "(\\alpha = 0.1)" + }, + { + "bbox": [ + 54, + 437, + 295, + 653 + ], + "type": "text", + "content": ", FedWSQ (1-bit) achieves " + }, + { + "bbox": [ + 54, + 437, + 295, + 653 + ], + "type": "inline_equation", + "content": "62.05\\%" + }, + { + "bbox": [ + 54, + 437, + 295, + 653 + ], + "type": "text", + "content": " accuracy on CIFAR-100, surpassing FedRCL and FedACG by " + }, + { + "bbox": [ + 54, + 437, + 295, + 653 + ], + "type": "inline_equation", + "content": "3.79\\%" + }, + { + "bbox": [ + 54, + 437, + 295, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 437, + 295, + 653 + ], + "type": "inline_equation", + "content": "3.91\\%" + }, + { + "bbox": [ + 54, + 437, + 295, + 653 + ], + "type": "text", + "content": ", respectively. Similarly, on Tiny/ImageNet, FedWSQ (1-bit) achieves " + }, + { + "bbox": [ + 54, + 437, + 295, + 653 + ], + "type": "inline_equation", + "content": "45.11\\%" + }, + { + "bbox": [ + 54, + 437, + 295, + 653 + ], + "type": "text", + "content": " accuracy, outperforming FedRCL and FedACG by " + }, + { + "bbox": [ + 54, + 437, + 295, + 653 + ], + "type": "inline_equation", + "content": "7.25\\%" + }, + { + "bbox": [ + 54, + 437, + 295, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 437, + 295, + 653 + ], + "type": "inline_equation", + "content": "5.36\\%" + }, + { + "bbox": [ + 54, + 437, + 295, + 653 + ], + "type": "text", + "content": ", respectively. These results highlight the effectiveness of FedWSQ, which demonstrates its robustness even in extremely low bit-width scenarios while maintaining high model accuracy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 659, + 170, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 170, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 170, + 673 + ], + "type": "text", + "content": "4.4. Additional analyses" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Impact of bit representation in FedWSQ FedWSQ balances model performance and communication efficiency by adopting different bit-width settings. Table 1 presents the" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 318, + 301, + 552, + 371 + ], + "blocks": [ + { + "bbox": [ + 318, + 301, + 552, + 371 + ], + "lines": [ + { + "bbox": [ + 318, + 301, + 552, + 371 + ], + "spans": [ + { + "bbox": [ + 318, + 301, + 552, + 371 + ], + "type": "image", + "image_path": "1ed1f36c553f79a20bc740e1b8baee9687320d82ed9fa44955e3e9d54d496882.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 375, + 555, + 399 + ], + "lines": [ + { + "bbox": [ + 313, + 375, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 555, + 399 + ], + "type": "text", + "content": "Figure 3. Loss landscape of FL methods trained on CIFAR-100, where the numbers indicate the Hessian top eigenvalue." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 406, + 555, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 555, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 555, + 597 + ], + "type": "text", + "content": "results for 32-bit (FedWS), 4-bit, and 1-bit representations, as well as adaptive bit-allocation strategies (FBA and DBA). Despite using lower bit-widths, FedWSQ remains robust in non-i.i.d. settings, particularly under high heterogeneity (" + }, + { + "bbox": [ + 313, + 406, + 555, + 597 + ], + "type": "inline_equation", + "content": "\\alpha = 0.1" + }, + { + "bbox": [ + 313, + 406, + 555, + 597 + ], + "type": "text", + "content": "). Since local models tend to overfit under high data heterogeneity, quantizing model updates helps to regularize overfitting, achieving more robust FL performance [2, 3]. The accuracy gaps among different bit-widths are relatively small, and thus the FBA and DBA strategies enable efficient communication with marginal performance loss. As shown in Table 1, both FBA and DBA strategies achieve performance close to the 4-bit setting while reducing communication costs by half. In summary, FedWSQ confirms its ability to maintain accuracy while significantly reducing communication bandwidth usage for real-world FL scenarios." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Effect of WS and DANUQ Table 2 shows a comparative analysis of our proposed DANUQ and UQ under varying bit-widths and data heterogeneity. In this experiment, UQ is implemented using the FedPAQ approach [37]. As listed in the table, our DANUQ consistently outperforms the UQ method. Under extremely low-bit quantization (1-bit and 2-bit settings), the UQ suffers from significant accuracy degradation. While WS alone improves performance, combining WS with the proposed DANUQ yields the high-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4622" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 77, + 84, + 270, + 145 + ], + "blocks": [ + { + "bbox": [ + 64, + 70, + 286, + 81 + ], + "lines": [ + { + "bbox": [ + 64, + 70, + 286, + 81 + ], + "spans": [ + { + "bbox": [ + 64, + 70, + 286, + 81 + ], + "type": "text", + "content": "Table 4. FL performance of different backbones for " + }, + { + "bbox": [ + 64, + 70, + 286, + 81 + ], + "type": "inline_equation", + "content": "\\alpha = {0.3}" + }, + { + "bbox": [ + 64, + 70, + 286, + 81 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 77, + 84, + 270, + 145 + ], + "lines": [ + { + "bbox": [ + 77, + 84, + 270, + 145 + ], + "spans": [ + { + "bbox": [ + 77, + 84, + 270, + 145 + ], + "type": "table", + "html": "
BackboneVanillaFedWSFedWSQ (DBA)
ShuffleNet [53]36.3753.3251.12
VGGNet-9 [40]46.1460.0160.48
SqueezeNet [14]39.4555.8756.10
ResNet-18 [11]47.3968.1966.32
MobileViT [32]35.4240.8041.27
", + "image_path": "8fadb22a3ed349650b85592d01a5eb031318e5ec26d22a8d260921f7799f682d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 152, + 294, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 152, + 294, + 260 + ], + "spans": [ + { + "bbox": [ + 55, + 152, + 294, + 260 + ], + "type": "text", + "content": "est performance gains. The UQ with WS shows improvements compared with its non-WS counterparts, but the synergy between WS and DANUQ provides the most robust FL performance. Furthermore, our method maintains high accuracy even with low-bit representations. This suggests that applying FBA or DBA can be highly effective, as our approach maintains strong performance across varying bit-widths and makes FedWSQ particularly practical for real-world FL deployments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 264, + 295, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 264, + 295, + 420 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 295, + 420 + ], + "type": "text", + "content": "Comparison with other NUQ methods We evaluate the effectiveness of the proposed DANUQ by comparing it with existing NUQ approaches, FP [45] and NF [7].7 As shown in Table 3, at the 4-bit representation, all NUQ methods perform competitively, indicating that NUQ is generally robust at relatively higher bit-widths. However, under 1-bit and 2-bit representations, FP and NF suffer from severe performance degradation, particularly in highly heterogeneous data (" + }, + { + "bbox": [ + 55, + 264, + 295, + 420 + ], + "type": "inline_equation", + "content": "\\alpha = 0.1" + }, + { + "bbox": [ + 55, + 264, + 295, + 420 + ], + "type": "text", + "content": "). In contrast, our DANUQ outperforms both NF and FP across all conditions, which confirms its robustness against extreme quantization constraints. These results highlight the superiority of our DANUQ method among NUQ methods for communication-efficient FL." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 425, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 295, + 593 + ], + "type": "text", + "content": "Loss Landscape Analysis Figure 3 visualizes the loss landscapes of various FL methods compared to FedWSQ (with DBA). FedWSQ achieves the smoothest and most stable loss landscape, with the lowest Hessian top eigenvalue of 135.8, indicating improved generalization ability. FedSmoo exhibits a flatter landscape but converges to a much higher global minimum, reducing sharpness at the cost of performance. In contrast, FedACG reaches a global comparable to FedWSQ but has a significantly sharper landscape, with a Hessian top eigenvalue of 671.2, which indicates it is more sensitive to perturbations. These results show that FedWSQ effectively enhances the generalization, leading to more stable training and improved FL performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 601, + 295, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 601, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 295, + 685 + ], + "type": "text", + "content": "Impact on different backbone architectures Table 4 presents the performance of different backbone architectures on CIFAR-100 to evaluate the impact of FedWS and FedWSQ. For all architectures, FedWS significantly improves accuracy over the vanilla models, and FedWSQ maintains a comparable performance to FedWS with improved communication efficiency. In addition, We eval" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 331, + 96, + 534, + 131 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 553, + 93 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 553, + 93 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 553, + 93 + ], + "type": "text", + "content": "Table 5. Ablation study for " + }, + { + "bbox": [ + 313, + 70, + 553, + 93 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 313, + 70, + 553, + 93 + ], + "type": "text", + "content": " using FedWS on CIFAR-10. For the non-i.i.d. setting, " + }, + { + "bbox": [ + 313, + 70, + 553, + 93 + ], + "type": "inline_equation", + "content": "\\alpha = {0.3}" + }, + { + "bbox": [ + 313, + 70, + 553, + 93 + ], + "type": "text", + "content": " is used." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 331, + 96, + 534, + 131 + ], + "lines": [ + { + "bbox": [ + 331, + 96, + 534, + 131 + ], + "spans": [ + { + "bbox": [ + 331, + 96, + 534, + 131 + ], + "type": "table", + "html": "
ρ1 × 10-41 × 10-31 × 10-21 × 10-1
non-i.i.d.87.1589.7189.6289.46
i.i.d.90.6492.4891.9792.11
", + "image_path": "c5543337dc8d48510ecc4d8a128bd1c3e35ed2623b8ce2ad1339587b1ee10beb.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 138, + 553, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 138, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 313, + 138, + 553, + 175 + ], + "type": "text", + "content": "uate MobileViT [32], a vision transformer-based architecture, which also benefits from FedWS and FedWSQ. These results confirm their applicability beyond CNNs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 182, + 555, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 555, + 301 + ], + "type": "text", + "content": "Comparison with FedWon We provide the comparison of FedWSQ with FedWon, which was discussed in Section 3.1. As shown in Table 1, FedWSQ substantially outperforms FedWon across all datasets and data heterogeneity settings. As illustrated in Figure 3, FedWon exhibits a much higher global minimum and a very sharp loss landscape with a Hessian top eigenvalue of 203.4, whereas FedWSQ maintains a smoother surface with an eigenvalue of 135.8. These results demonstrate FedWSQ's superior robustness and generalization ability." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 308, + 556, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 556, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 556, + 429 + ], + "type": "text", + "content": "Hyper-parameters We examine the hyper-parameter " + }, + { + "bbox": [ + 313, + 308, + 556, + 429 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 313, + 308, + 556, + 429 + ], + "type": "text", + "content": " for WS, which handles the normalization scales. To evaluate its impact, we vary the values of " + }, + { + "bbox": [ + 313, + 308, + 556, + 429 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 313, + 308, + 556, + 429 + ], + "type": "text", + "content": " and assess the performance of FedWS under both non-i.i.d. and i.i.d. data distributions. As shown in Table 5, the performance of FedWS is insensitive to " + }, + { + "bbox": [ + 313, + 308, + 556, + 429 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 313, + 308, + 556, + 429 + ], + "type": "text", + "content": " variations. This is because during inference, normalization layers like GN cancel the effect of constant scaling, and the variance of the parameter vectors within the same convolutional layer is not significant. We select " + }, + { + "bbox": [ + 313, + 308, + 556, + 429 + ], + "type": "inline_equation", + "content": "\\rho = 1 \\times 10^{-3}" + }, + { + "bbox": [ + 313, + 308, + 556, + 429 + ], + "type": "text", + "content": ", which yields the best results." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 441, + 391, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 391, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 391, + 453 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 461, + 555, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 555, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 555, + 654 + ], + "type": "text", + "content": "This paper presents FedWSQ, a novel FL framework using WS and the proposed DANUQ to address key FL challenges, including data heterogeneity, partial client participation, and communication bottleneck. By leveraging WS, FedWSQ enhances training stability via an implicit gradient filtering mechanism. Additionally, DANUQ efficiently compresses LMPUs while preserving essential local information, significantly reducing communication overhead even at ultra-low bit-widths. Extensive experiments on the benchmark datasets confirm that FedWSQ consistently outperforms SOTA FL methods under various conditions. Furthermore, the mixed-precision strategies FBA and DBA effectively balance communication cost and model accuracy. These results validate the robustness of the FedWSQ framework in extreme quantization scenarios, making it highly practical for real-world FL." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": "Acknowledgments This work was supported by the National Research Foundation (NRF), South Korea, under project BK21 FOUR (Smart Robot Convergence and Application Education Research Center)." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "inline_equation", + "content": "{}^{7}" + }, + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": " QLs of different methods are visualized in Section C of the supplementary document." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4623" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "text", + "content": "[1] Durmus Alp Emre Acar, Yue Zhao, Ramon Matas, Matthew Mattina, Paul Whatmough, and Venkatesh Saligrama. Federated learning based on dynamic regularization. In ICLR, 2021. 2, 6, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 190 + ], + "type": "text", + "content": "[2] MohammadHossein AskariHemmat, Reyhane Askari Hemmat, Alex Hoffman, Ivan Lazarevich, Ehsan Saboori, Olivier Mastropietro, Sudhakar Sah, Yvon Savaria, and Jean-Pierre David. QReg: On regularization effects of quantization. arXiv preprint arXiv:2206.12372, 2022. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 191, + 295, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 191, + 295, + 246 + ], + "spans": [ + { + "bbox": [ + 62, + 191, + 295, + 246 + ], + "type": "text", + "content": "[3] MohammadHossein AskariHemmat, Ahmadreza Jeddi, Reyhane Askari Hemmat, Ivan Lazarevich, Alexander Hoffman, Sudhakar Sah, Ehsan Saboori, Yvon Savaria, and Jean-Pierre David. QGen: On the ability to generalize in quantization aware training. arXiv preprint arXiv:2404.11769, 2024. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 247, + 295, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 295, + 300 + ], + "type": "text", + "content": "[4] Debraj Basu, Deepesh Data, Can Karakus, and Suhas N Diggavi. Qsparse-Local-SGD: Distributed SGD with quantization, sparsification, and local computations. IEEE Journal on Selected Areas in Information Theory, 1(1):217-226, 2020. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 301, + 295, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 301, + 295, + 345 + ], + "spans": [ + { + "bbox": [ + 62, + 301, + 295, + 345 + ], + "type": "text", + "content": "[5] Shengbo Chen, Cong Shen, Lanxue Zhang, and Yuanmin Tang. Dynamic aggregation for heterogeneous quantization in federated learning. IEEE Transactions on Wireless Communications, 20(10):6804-6819, 2021. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 346, + 295, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 346, + 295, + 389 + ], + "spans": [ + { + "bbox": [ + 62, + 346, + 295, + 389 + ], + "type": "text", + "content": "[6] Jungwook Choi, Zhuo Wang, Swagath Venkataramani, Pierce I-Jen Chuang, Vijayalakshmi Srinivasan, and Kailash Gopalakrishnan. Pact: Parameterized clipping activation for quantized neural networks. In ICLR, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 391, + 295, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 391, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 391, + 295, + 413 + ], + "type": "text", + "content": "[7] T. Dettmers, M. Lewis, S. Shleifer, and L. Zettlemoyer. 8-bit optimizers via block-wise quantization. In ICLR, 2022. 4, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 414, + 295, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 414, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 62, + 414, + 295, + 445 + ], + "type": "text", + "content": "[8] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: Efficient finetuning of quantized llms. In NeurIPS, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 447, + 295, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 295, + 479 + ], + "type": "text", + "content": "[9] Steven K. Esser, Jeffrey L. McKinstry, Deepika Bablani, Rathinakumar Appuswamy, and Dharmendra S. Modha. Learned Step Size Quantization. In ICLR, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 480, + 295, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 480, + 295, + 512 + ], + "spans": [ + { + "bbox": [ + 57, + 480, + 295, + 512 + ], + "type": "text", + "content": "[10] Robert M. Gray and David L. Neuhoff. Quantization. IEEE transactions on information theory, 44(6):2325-2383, 1998. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 514, + 295, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 514, + 295, + 546 + ], + "spans": [ + { + "bbox": [ + 57, + 514, + 295, + 546 + ], + "type": "text", + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6, 8, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 548, + 295, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 548, + 295, + 580 + ], + "spans": [ + { + "bbox": [ + 57, + 548, + 295, + 580 + ], + "type": "text", + "content": "[12] Kevin Hsieh, Amar Phanishayee, Onur Mutlu, and Phillip Gibbons. The non-IID data quagmire of decentralized machine learning. In ICML, 2020. 6, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 581, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 581, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 581, + 295, + 624 + ], + "type": "text", + "content": "[13] Tzu-Ming Harry Hsu, Hang Qi, and Matthew Brown. Measuring the effects of non-identical data distribution for federated visual classification. arXiv preprint arXiv:1909.06335, 2019. 2, 6, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 625, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 295, + 678 + ], + "type": "text", + "content": "[14] Forrest N Iandola, Song Han, Matthew W Moskewicz, Khalid Ashraf, William J Dally, and Kurt Keutzer. SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and " + }, + { + "bbox": [ + 57, + 625, + 295, + 678 + ], + "type": "inline_equation", + "content": "< 0.5" + }, + { + "bbox": [ + 57, + 625, + 295, + 678 + ], + "type": "text", + "content": " MB model size. arXiv preprint arXiv:1602.07360, 2016.8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 680, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 295, + 712 + ], + "type": "text", + "content": "[15] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015. 6" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[16] Benoit Jacob, Skirmantas Kligys, Bo Chen, Menglong Zhu, Matthew Tang, Andrew Howard, Hartwig Adam, and Dmitry Kalenichenko. Quantization and training of neural networks for efficient integer-arithmetic-only inference. In CVPR, pages 2704-2713, 2018. 2, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 129, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 183 + ], + "type": "text", + "content": "[17] Sangil Jung, Changyong Son, Seohyung Lee, Jinwoo Son, Youngjun Kwak, Jae-Joon Han, Sung Ju Hwang, and Changkyu Choi. Learning to quantize deep networks by optimizing quantization intervals with task loss. In CVPR, pages 4350-4359, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 185, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 185, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 185, + 553, + 228 + ], + "type": "text", + "content": "[18] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank J Reddi, Sebastian U Stich, and Ananda Theertha Suresh. SCAFFOLD: Stochastic controlled averaging for ondevice federated learning. In ICML, 2020. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 230, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 230, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 230, + 553, + 262 + ], + "type": "text", + "content": "[19] Ahmed Khaled, Konstantin Mishchenko, and Peter Richtárik. First analysis of local GD on heterogeneous data. arXiv preprint arXiv:1909.04715, 2019. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 264, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 296 + ], + "type": "text", + "content": "[20] Geeho Kim, Jinkyu Kim, and Bohyung Han. Communication-efficient federated learning with accelerated client gradient. In CVPR, 2024. 6, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 297, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 330 + ], + "type": "text", + "content": "[21] Jinkyu Kim, Geeho Kim, and Bohyung Han. Multi-level branched regularization for federated learning. In ICML, 2022. 2, 6, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 331, + 553, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 375 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 375 + ], + "type": "text", + "content": "[22] Jakub Konečny, H. Brendan McMahan, Felix X. Yu, Ananda Theertha Suresh, Dave Bacon, and Peter Richtárik. Federated learning: Strategies for improving communication efficiency. In arXiv preprint, 2017. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 376, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 409 + ], + "type": "text", + "content": "[23] Alex Krizhevsky. Learning multiple layers of features from tiny images. Master's thesis, Department of Computer Science, University of Toronto, 2009. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 411, + 553, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 463 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 463 + ], + "type": "text", + "content": "[24] Khiem Le, Nhan Luong-Ha, Manh Nguyen-Duc, Danh Le-Phuoc, Cuong Do, and Kok-Seng Wong. Exploring the practicality of federated learning: A survey towards the communication perspective. arXiv preprint arXiv:2405.20431, 2024. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 466, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 466, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 466, + 553, + 487 + ], + "type": "text", + "content": "[25] Ya Le and Xuan Yang. Tiny ImageNet visual recognition challenge. CS 231N, 7(7):3, 2015. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 489, + 553, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 489, + 553, + 531 + ], + "spans": [ + { + "bbox": [ + 317, + 489, + 553, + 531 + ], + "type": "text", + "content": "[26] Gihun Lee, Minchan Jeong, Yongjin Shin, Sangmin Bae, and Se-Young Yun. Preservation of the global knowledge by not-true distillation in federated learning. In NeurIPS, 2022. 6, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 533, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 533, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 533, + 553, + 567 + ], + "type": "text", + "content": "[27] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. In MLSys, 2020. 2, 6, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 568, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 568, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 317, + 568, + 553, + 600 + ], + "type": "text", + "content": "[28] Xiang Li, Kaixuan Huang, Wenhao Yang, Shusen Wang, and Zhihua Zhang. On the convergence of FedAvg on non-IID data. In ICLR, 2020. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 601, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 601, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 317, + 601, + 553, + 644 + ], + "type": "text", + "content": "[29] Darryl Lin, Sachin Talathi, and Sreekanth Annapureddy. Fixed point quantization of deep convolutional networks. In International conference on machine learning, pages 2849-2858. PMLR, 2016. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "text", + "content": "[30] Wang Luping, Wang Wei, and Li Bo. CMFL: Mitigating communication overhead for federated learning. In 2019 IEEE 39th international conference on distributed computing systems (ICDCS), pages 954-964. IEEE, 2019. 1" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[31] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4624" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "efficient learning of deep networks from decentralized data. In AISTATS, 2017. 1, 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 97, + 295, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 97, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 97, + 295, + 129 + ], + "type": "text", + "content": "[32] Sachin Mehta and Mohammad Rastegari. Mobilevit: lightweight, general-purpose, and mobile-friendly vision transformer. arXiv preprint arXiv:2110.02178, 2021. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 132, + 294, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 132, + 294, + 175 + ], + "spans": [ + { + "bbox": [ + 56, + 132, + 294, + 175 + ], + "type": "text", + "content": "[33] Markus Nagel, Marios Fournarakis, Rana Ali Amjad, Yelysei Bondarenko, Mart van Baalen, and Tijmen Blankevoort. A white paper on neural network quantization. arXiv preprint arXiv:2106.08295, 2021. 2, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 177, + 294, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 177, + 294, + 232 + ], + "spans": [ + { + "bbox": [ + 56, + 177, + 294, + 232 + ], + "type": "text", + "content": "[34] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. PyTorch: An imperative style, high-performance deep learning library. In NeurIPS, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 234, + 294, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 234, + 294, + 277 + ], + "spans": [ + { + "bbox": [ + 56, + 234, + 294, + 277 + ], + "type": "text", + "content": "[35] Siyuan Qiao, Huiyu Wang, Chenxi Liu, Wei Shen, and Alan Yuille. Micro-batch training with batch-channel normalization and weight standardization. arXiv preprint arXiv:1903.10520, 2019. 1, 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 280, + 294, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 280, + 294, + 323 + ], + "spans": [ + { + "bbox": [ + 56, + 280, + 294, + 323 + ], + "type": "text", + "content": "[36] Sashank J Reddi, Zachary Charles, Manzil Zaheer, Zachary Garrett, Keith Rush, Jakub Konečny, Sanjiv Kumar, and Hugh Brendan McMahan. Adaptive federated optimization. In ICLR, 2021. 2, 6, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 326, + 294, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 326, + 294, + 378 + ], + "spans": [ + { + "bbox": [ + 56, + 326, + 294, + 378 + ], + "type": "text", + "content": "[37] Amirhossein Reisizadeh, Aryan Mokhtari, Hamed Hassani, Ali Jabbabaie, and Ramtin Pedarsani. FedPAQ: A communication-efficient federated learning method with periodic averaging and quantization. In AISTATS, 2020. 2, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 381, + 294, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 381, + 294, + 414 + ], + "spans": [ + { + "bbox": [ + 56, + 381, + 294, + 414 + ], + "type": "text", + "content": "[38] Seonguk Seo, Jinkyu Kim, Geeho Kim, and Bohyung Han. Relaxed contrastive learning for federated learning. In CVPR, 2024. 6, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 417, + 294, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 417, + 294, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 417, + 294, + 459 + ], + "type": "text", + "content": "[39] Yujun Shi, Jian Liang, Wenqing Zhang, Vincent YF Tan, and Song Bai. Towards understanding and mitigating dimensional collapse in heterogeneous federated learning. In ICLR, 2023. 6, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 463, + 294, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 463, + 294, + 495 + ], + "spans": [ + { + "bbox": [ + 56, + 463, + 294, + 495 + ], + "type": "text", + "content": "[40] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2014. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 498, + 294, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 498, + 294, + 518 + ], + "spans": [ + { + "bbox": [ + 56, + 498, + 294, + 518 + ], + "type": "text", + "content": "[41] Sebastian U Stich. Local SGD converges fast and communicates little. In ICLR, 2019. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 521, + 294, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 521, + 294, + 564 + ], + "spans": [ + { + "bbox": [ + 56, + 521, + 294, + 564 + ], + "type": "text", + "content": "[42] Sebastian U Stich and Sai Praneeth Karimireddy. The errorfeedback framework: Better rates for SGD with delayed gradients and compressed communication. arXiv preprint arXiv:1909.05350, 2019. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 567, + 294, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 294, + 621 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 294, + 621 + ], + "type": "text", + "content": "[43] Yan Sun, Li Shen, Shixiang Chen, Liang Ding, and Dacheng Tao. Dynamic regularized sharpness aware minimization in federated learning: Approaching global consistency and smooth landscape. In International Conference on Machine Learning, pages 32991-33013. PMLR, 2023. 2, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 624, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 666 + ], + "type": "text", + "content": "[44] Jianyu Wang and Gauri Joshi. Cooperative SGD: A unified framework for the design and analysis of local-update SGD algorithms. Journal of Machine Learning Research, 22(213): 1-50, 2021. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[45] Ruizhe Wang, Yeyun Gong, Xiao Liu, Guoshuai Zhao, Ziyue Yang, Baining Guo, Zhengjun Zha, and Peng Cheng. Optimizing large language model training using fp4 quantization. In arXiv preprint arXiv:2501.17116, 2025. 4, 8" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 443 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "text", + "content": "[46] Jie Wen, Zhixia Zhang, Yang Lan, Zhihua Cui, Jianghui Cai, and Wensheng Zhang. A survey on federated learning: challenges and applications. International Journal of Machine Learning and Cybernetics, 14(2):513-535, 2023. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 118, + 553, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 138 + ], + "type": "text", + "content": "[47] Yuxin Wu and Kaiming He. Group normalization. In ECCV, 2018. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 141, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 173 + ], + "type": "text", + "content": "[48] Jing Xu, Sen Wang, Liwei Wang, and Andrew Chi-Chih Yao. FedCM: Federated learning with client-level momentum. arXiv preprint arXiv:2106.10874, 2021. 6, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "type": "text", + "content": "[49] Haibo Yang, Minghong Fang, and Jia Liu. Achieving linear speedup with partial worker participation in non-IID federated learning. In ICLR, 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 209, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 251 + ], + "type": "text", + "content": "[50] Hao Yu, Sen Yang, and Shenghuo Zhu. Parallel restarted SGD with faster convergence and less communication: Demystifying why model averaging works for deep learning. In AAAI, 2019. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 254, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 285 + ], + "type": "text", + "content": "[51] Dongqing Zhang, Jiaolong Yang, Dongqiangzi Ye, and Gang Hua. LQ-Nets: Learned Quantization for Highly Accurate and Compact Deep Neural Networks. In ECCV, 2018. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 287, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 330 + ], + "type": "text", + "content": "[52] Jie Zhang, Zhiqi Li, Bo Li, Jianghe Xu, Shuang Wu, Shouhong Ding, and Chao Wu. Federated learning with label distribution skew via logits calibration. In ICML, 2022. 6, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 332, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 364 + ], + "type": "text", + "content": "[53] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. ShuffleNet: An extremely efficient convolutional neural network for mobile devices. In CVPR, 2018. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 365, + 553, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 388 + ], + "type": "text", + "content": "[54] Sijie Zhao, Tao Yue, and Xuemei Hu. Distribution-aware adaptive multi-bit quantization. In CVPR, 2021. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 389, + 553, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 389, + 553, + 411 + ], + "spans": [ + { + "bbox": [ + 316, + 389, + 553, + 411 + ], + "type": "text", + "content": "[55] Chenzhuo Zhu, Song Han, Huizi Mao, and William J. Dally. Trained ternary quantization. In ICLR, 2017. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 412, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 412, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 412, + 553, + 443 + ], + "type": "text", + "content": "[56] Weiming Zhuang and Lingjuan Lyu. FedWon: Triumphing multi-domain federated learning without normalization. In ICLR, 2023. 4, 6" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4625" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/FedXDS_ Leveraging Model Attribution Methods to counteract Data Heterogeneity in Federated Learning/3d4823b4-ec38-431a-8ed1-cb351135854c_content_list.json b/2025/FedXDS_ Leveraging Model Attribution Methods to counteract Data Heterogeneity in Federated Learning/3d4823b4-ec38-431a-8ed1-cb351135854c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f1e4137b0e11a9f4ba93a7b5da2ce3eddcd592be --- /dev/null +++ b/2025/FedXDS_ Leveraging Model Attribution Methods to counteract Data Heterogeneity in Federated Learning/3d4823b4-ec38-431a-8ed1-cb351135854c_content_list.json @@ -0,0 +1,1853 @@ +[ + { + "type": "text", + "text": "FedXDS: Leveraging Model Attribution Methods to counteract Data Heterogeneity in Federated Learning", + "text_level": 1, + "bbox": [ + 153, + 128, + 843, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Maximilian Andreas Hoefler $^{1}$ * Karsten Mueller $^{1}$ Wojciech Samek $^{1,2,3}$ $^{1}$ Fraunhofer Heinrich Hertz Institute, Berlin, Germany \n $^{2}$ Technical University of Berlin, Berlin, Germany \n $^{3}$ Berlin Institute for the Foundations of Learning and Data (BIFOLD)", + "bbox": [ + 204, + 202, + 790, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Explainable AI (XAI) methods have demonstrated significant success in recent years at identifying relevant features in input data that drive deep learning model decisions, enhancing interpretability for users. However, the potential of XAI beyond providing model transparency has remained largely unexplored in adjacent machine learning domains. In this paper, we show for the first time how XAI can be utilized in the context of federated learning. Specifically, while federated learning enables collaborative model training without raw data sharing, it suffers from performance degradation when client data distributions exhibit statistical heterogeneity. We introduce FedXDS (Federated Learning via XAI-guided Data Sharing), the first approach to utilize feature attribution techniques to identify precisely which data elements should be selectively shared between clients to mitigate heterogeneity. By employing propagation-based attribution, our method identifies task-relevant features through a single backward pass, enabling selective data sharing that aligns client contributions. To protect sensitive information, we incorporate metric privacy techniques that provide formal privacy guarantees while preserving utility. Experimental results demonstrate that our approach consistently achieves higher accuracy and faster convergence compared to existing methods across varying client numbers and heterogeneity settings. We provide theoretical privacy guarantees and empirically demonstrate robustness against both membership inference and feature inversion attacks. Code is available at https://github.com/MaxH1996/FedXDS.", + "bbox": [ + 91, + 340, + 480, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 806, + 220, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Federated learning (FL) [19] has received significant attention in recent years, proving to be an effective method for collaborative distributed machine learning while keeping", + "bbox": [ + 89, + 832, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "local datasets private. However, the practical deployment of federated learning systems faces fundamental challenges regarding data heterogeneity and privacy [11, 35].", + "bbox": [ + 511, + 310, + 903, + 354 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Data heterogeneity, characterized by non-IID data across clients, leads to contradicting model updates and significantly degraded performance [12, 31]. Current approaches address this challenge by aligning client distributions using proximal optimization terms [16] and mitigating client drift [12, 14], while others focus on flattening the loss landscape to facilitate better model aggregation [20, 21]. In addition, recent works have shown that sharing even a small portion of client data globally can significantly improve local model generalization [34]. Such shared data introduces centralization, giving clients access to a more homogeneous distribution, which reduces update divergence and mitigates heterogeneity [34].", + "bbox": [ + 509, + 356, + 906, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, sharing raw data is not feasible due to privacy constraints. Introducing differential privacy [9] by adding noise to the shared data offers a potential solution, however this can cause considerable performance loss. Alternatively, instead of sharing raw data, abstract feature representations or partial data could be released using generator-type approaches [30, 33, 36]. However, the obtained features are not as potent as raw data in mitigating statistical heterogeneity, and can also be prone to privacy leaks and additionally introduce significant computational overhead due to generators. Hence, the challenge arises of how to retain the performance improvement from data sharing while simultaneously guaranteeing privacy.", + "bbox": [ + 509, + 553, + 905, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address this challenge we propose FedXDS (Federated Learning via XAI-guided Data Sharing) which entails a novel approach wherein we leverage methods from the field of eXplainable AI (XAI) [8] enabling performance enhancing data sharing while preserving privacy.", + "bbox": [ + 511, + 750, + 903, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Specifically, we leverage propagation-based attribution methods, which have been shown to reliably identify input features that consistently contribute to model predictions [2, 8]. Moreover, attribution maps yield a pixel-wise relevance score which highlights regions in the input space", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Correspondence to maximilian.andreas.hoefler@hhi.fraunhofer.de", + "bbox": [ + 107, + 886, + 465, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4572", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "which focus on semantically meaningful features, while suppressing spurious correlations and background noise. In our setup local client data is filtered through the attribution mechanism, i.e., we only share the input features which the attribution map deems most relevant for model predictions. This retains the information needed for model generalization while discarding irrelevant information.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We show that our method, when combined with differential-metric privacy techniques [6, 9], offers stronger privacy guarantees compared to applying similar privacy mechanisms directly to raw input features. Specifically, by discarding task-irrelevant information we effectively reduce the dimensionality of the raw features. This allows us to selectively protect only the task-relevant regions rather than uniformly protecting all input pixels, and thus obtain better utility.", + "bbox": [ + 89, + 196, + 480, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition, propagation-based attributions can be obtained via a single backward pass over the local dataset, which only needs to be performed once in the entire FL process.", + "bbox": [ + 89, + 333, + 480, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions can be summarized as:", + "bbox": [ + 89, + 393, + 398, + 407 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A novel federated learning algorithm leveraging propagation-based attributions to address the challenge of retaining performance from data sharing with privacy guarantees in heterogeneous environments.", + "- A privacy-preserving mechanism that utilizes attribution-guided dimensionality reduction to achieve metric differential privacy. We perform theoretical and empirical privacy evaluations on our feature sharing approach through membership inference attacks and feature inversion, demonstrating that we can attain strong privacy guarantees.", + "- Extensive experiments on image datasets benchmarking our method against current state of the art, demonstrating superior performance in terms of accuracy and convergence." + ], + "bbox": [ + 89, + 410, + 480, + 637 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 651, + 227, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Federated Learning", + "text_level": 1, + "bbox": [ + 89, + 676, + 279, + 693 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Federated Learning [19] is a distributed machine learning paradigm where multiple clients collaborate to train a model under the orchestration of a central server, without sharing their raw data. Let $K$ be the number of clients, each with a local dataset $\\mathcal{D}_k = \\{(x_i^k,y_i^k)\\}_{i = 1}^{n_k}$ , where $x_{i}^{k}\\in \\mathcal{X}$ is an input sample and $y_{i}^{k}\\in \\mathcal{V}$ is its corresponding label. The goal is to learn a global model $f_{\\theta}:\\mathcal{X}\\to \\mathcal{Y}$ parameterized by $\\theta$ , which minimizes the empirical risk:", + "bbox": [ + 89, + 698, + 480, + 819 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\frac {1}{N} \\sum_ {k = 1} ^ {K} \\sum_ {i = 1} ^ {n _ {k}} \\mathcal {L} \\left(f _ {\\theta} \\left(x _ {i} ^ {k}\\right), y _ {i} ^ {k}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 833, + 480, + 873 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $N = \\sum_{k=1}^{K} n_k$ is the total number of samples", + "bbox": [ + 109, + 883, + 480, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "across all clients, and $\\mathcal{L}$ is a suitable loss function.", + "bbox": [ + 511, + 90, + 844, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The FedAvg algorithm [19] iteratively aggregates model updates from selected clients $S_{t}$ , where each client $k$ performs local updates $\\theta_{k}^{t} = \\theta^{t - 1} - \\eta \\nabla \\mathcal{L}_{k}(\\theta^{t - 1})$ before the server averages these updates as $\\theta^t = \\frac{1}{|S_t|}\\sum_{k\\in S_t}\\theta_k^t$ .", + "bbox": [ + 511, + 106, + 903, + 170 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Differential Privacy in Metric Spaces", + "text_level": 1, + "bbox": [ + 511, + 176, + 833, + 191 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Standard differential privacy (DP) [9] provides guarantees against membership inference by bounding the output distribution change when a single dataset element is modified. For continuous data such as image embeddings, classical adjacency notions are restrictive, making metric differential privacy [6] more appropriate:", + "bbox": [ + 511, + 196, + 903, + 289 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Definition 1 (Metric Privacy). [6] Let $(X, d_X)$ be a metric space and $Z$ a set of possible outputs. A randomized mechanism $A\\colon X \\to \\Delta(Z)$ is $(\\varepsilon, \\delta)$ -metric private if for all $x, x' \\in X$ and every measurable set $U \\subseteq Z$ :", + "bbox": [ + 511, + 294, + 905, + 354 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\Pr [ A (x) \\in U ] \\leq \\exp \\left(\\varepsilon d _ {X} \\left(x, x ^ {\\prime}\\right)\\right) \\Pr [ A \\left(x ^ {\\prime}\\right) \\in U ] + \\delta .\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 362, + 903, + 387 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This definition reduces to standard DP when $d_X(x, x') = 1$ for all adjacent datasets, but naturally captures similarity in continuous spaces. In our work, we define $(X, d_X)$ where $X = \\mathbb{R}^d$ represents image data and $d_X(x, x') = ||x - x'||_2$ is the $\\ell_2$ distance between inputs.", + "bbox": [ + 511, + 392, + 903, + 468 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To guarantee privacy, we calibrate noise based on the sensitivity of a query function $f\\colon X\\to \\mathbb{R}^m$ , which measures output changes relative to input changes. In our work we choose the following sensitivity measure as used in [4, 10, 26]:", + "bbox": [ + 511, + 468, + 903, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Definition 2 (Sensitivity). For a function $f \\colon X \\to \\mathbb{R}^m$ , the sensitivity is:", + "bbox": [ + 511, + 547, + 903, + 579 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta_ {f} = \\max _ {x, x ^ {\\prime} \\in X} \\frac {\\| f (x) - f \\left(x ^ {\\prime}\\right) \\|}{\\left| \\left| x - x ^ {\\prime} \\right| \\right|}.\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 584, + 812, + 618 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This definition captures the largest possible output difference relative to input difference, directly controlling how much noise is needed for privacy. This notion of sensitivity offers several advantages, which we discuss in the supplementary material. Importantly, when $\\Delta_f$ is small, less noise is required to achieve the same privacy guarantee, preserving more signal and improving utility—particularly valuable for high-dimensional data like images.", + "bbox": [ + 511, + 623, + 903, + 743 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For privacy preservation, we employ the Gaussian mechanism, which adds noise to the output of our query function proportional to the sensitivity $\\Delta_f$ :", + "bbox": [ + 511, + 744, + 903, + 790 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Theorem 1 (Gaussian Mechanism). [6, 9] For a function $f: \\mathcal{D} \\to \\mathbb{R}^d$ with sensitivity $\\Delta_f$ , adding Gaussian noise with scale $\\sigma$ satisfying:", + "bbox": [ + 511, + 795, + 903, + 840 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma \\geq \\frac {\\Delta_ {f} \\sqrt {2 \\log (1 . 2 5 / \\delta)}}{\\varepsilon} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 845, + 903, + 878 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ensures $f(x) + \\mathcal{N}(0, \\sigma^2 I)$ is $(\\varepsilon, \\delta)$ -metric private.", + "bbox": [ + 511, + 883, + 846, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4573", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This mechanism forms the foundation of our privacy-preserving feature sharing approach, allowing us to bound and quantify the privacy guarantees of our federated learning framework.", + "bbox": [ + 89, + 90, + 482, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Neural Network Attribution Methods", + "text_level": 1, + "bbox": [ + 89, + 160, + 413, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We consider several gradient-based attribution techniques in our work. The simplest being Gradient $\\times$ Input [25], which computes attributions through element-wise multiplication of input and gradient:", + "bbox": [ + 89, + 181, + 483, + 244 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} _ {\\operatorname {g r a d}} \\left(f _ {\\theta}, \\mathbf {x}\\right) = \\mathbf {x} \\odot \\frac {\\partial f _ {\\theta}}{\\partial \\mathbf {x}} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 252, + 482, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Integrated Gradients (IG) [28] addresses gradient saturation by accumulating gradients along a path from baseline $\\mathbf{x}'$ to input:", + "bbox": [ + 89, + 290, + 483, + 335 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} _ {\\mathrm {I G}} \\left(f _ {\\theta}, \\mathbf {x}\\right) = \\left(\\mathbf {x} - \\mathbf {x} ^ {\\prime}\\right) \\odot \\int_ {0} ^ {1} \\frac {\\partial f _ {\\theta} \\left(\\mathbf {x} ^ {\\prime} + \\alpha \\left(\\mathbf {x} - \\mathbf {x} ^ {\\prime}\\right)\\right)}{\\partial \\mathbf {x}} d \\alpha \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 345, + 482, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SmoothGrad [27] reduces attribution noise by averaging gradients over perturbed inputs:", + "bbox": [ + 89, + 388, + 483, + 420 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} _ {\\text {s m o o t h}} (f _ {\\theta}, \\mathbf {x}) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\frac {\\partial f _ {\\theta}}{\\partial (\\mathbf {x} + \\epsilon)} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 429, + 482, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Lastly we use Layer-wise Relevance Propagation (LRP) with the $\\epsilon$ -rule according to [3]. This recursively propagates relevance scores $R^{(l)}$ from layer $l$ to $l - 1$ using:", + "bbox": [ + 89, + 479, + 483, + 525 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nR _ {j} = \\sum \\frac {a _ {j} w _ {j k}}{\\epsilon + \\sum_ {0 , j} a _ {j} w _ {j k}} R _ {k} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 534, + 482, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $w_{jk}$ are layer weights, $a_{j}$ are layer activations, and $\\epsilon$ is a small stabilizing term.", + "bbox": [ + 89, + 575, + 483, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Related Work", + "text_level": 1, + "bbox": [ + 89, + 618, + 233, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Federated Learning was introduced with FedAvg [19], enabling collaborative model training without sharing raw data, but suffers under statistical heterogeneity. Regularization-based methods mitigate client drift by constraining local updates, such as FedProx [16], SCAFFOLD [12], and FedDyn [1], with extensions like FedBN [17], MOON [15], FedNova [29], FedSAM [21], and FedDISCO [32] targeting specific aspects like feature shifts or loss geometry. A more direct solution is data or knowledge sharing: FedDF and FedAux [18, 24] distill client knowledge using a public dataset; FedGen [36] synthesizes class-conditional features for distribution alignment; FedFTG [33] generates pseudo-data for feature-level transfer; and FedFed [30] applies VAE-based feature distillation. While effective, these methods often introduce significant computational and privacy costs due to data synthesis and sharing. In contrast, FedXDS shares compact", + "bbox": [ + 89, + 643, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c6b902572e5b77296570c376eb8544f36a1845179476bbbb780ddf654482818a.jpg", + "image_caption": [ + "Figure 1. Illustration of the DP attribution-guided feature extraction. The process consists of four main steps: (1) Computing attribution scores using a warmup model to identify important features, (2) Creating and applying a binary mask based on the attributions, (3) Adding Gaussian noise for metric privacy." + ], + "image_footnote": [], + "bbox": [ + 532, + 87, + 890, + 305 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Algorithm 1 Attribution-Guided Private Representation Extraction.", + "bbox": [ + 514, + 402, + 905, + 430 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Require: $D_{k}$ : client dataset, $f_{\\theta}$ : model, $s$ : sparsity level, $(\\varepsilon ,\\delta)$ : privacy params, $S$ : sensitivity, $\\theta_{\\mathrm{warmup}}$", + "bbox": [ + 514, + 434, + 905, + 464 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Ensure: Private dataset $D_k^p$", + "bbox": [ + 514, + 464, + 702, + 479 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: for $\\mathbf{x}_i^k\\in D_k$ do", + "2: $\\mathbf{h} \\gets \\mathcal{A}(f_{\\theta_{\\mathrm{warmup}}}, \\mathbf{x}_i^k)$ ; $h^{(s)} \\gets \\mathrm{s}$ -th largest in $\\mathbf{h}$", + "3: $\\mathbf{m} \\gets [\\mathbf{h}[i] \\geq h^{(s)}]; \\mathbf{x}_m \\gets \\mathbf{x}_i^k \\odot \\mathbf{m}$", + "4: $\\tilde{\\mathbf{x}}_i^k\\gets \\mathbf{x}_m + \\mathcal{N}(0,\\sigma^2)$ $\\triangleright \\sigma$ from $(\\varepsilon ,\\delta)$", + "5: Add $(\\tilde{\\mathbf{x}}_i^k,y_i^k)$ to $D_k^p$", + "6: end for", + "7: return $D_k^p$" + ], + "bbox": [ + 524, + 479, + 903, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "attribution-based feature subsets derived via XAI, avoiding generators while enhancing privacy, efficiency, and performance in heterogeneous settings. Wo provide an extensive related works in the supplementary material.", + "bbox": [ + 511, + 618, + 905, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. FedXDS Approach", + "text_level": 1, + "bbox": [ + 511, + 693, + 696, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Consider a federated learning system with $N$ clients, where each client $k$ holds local data $\\mathcal{D}_k$ . In our setting all clients share the same neural network architecture $f_{\\theta}$ . Our goal is to learn a global model $\\Theta_G$ which generalizes well on all client datasets.", + "bbox": [ + 511, + 718, + 905, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Attribution-Guided Feature Selection", + "text_level": 1, + "bbox": [ + 511, + 803, + 836, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The first stage of our method identifies discriminative features using a pre-trained warmup model, with parameters $\\theta_{\\mathrm{warmup}}$ . This is achieved by training with FedAvg for $R_{\\mathrm{warmup}}$ rounds, before sharing representations. Then, for each input sample $\\mathbf{x} \\in \\mathbb{R}^{H \\times W \\times 3}$ , where $H$ and $W$ are the", + "bbox": [ + 511, + 824, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4574", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 2 Privacy-Preserving Federated Training" + ], + "code_body": "Require: $\\{D_k\\}_{k=1}^K$ : client datasets, $\\lambda$ : knowledge weight, $T$ : rounds, $E$ : local epochs, $\\eta$ : learning rate, $R_{\\text{warmup}}$ \n1: Initialize $\\theta_{\\text{warmup}}$ using $R_{\\text{warmup}}$ rounds of FedAvg \n2: Obtain $D_k^p$ using Algorithm 1 for all $k \\in [K]$ in parallel \n3: Server: $\\mathcal{D}_g \\gets \\bigcup_{k=1}^K D_k^p$ \n4: for round $t = 1$ to $T$ do \n5: Select client subset $S_t \\subseteq [K]$ \n6: for $k \\in S_t$ in parallel do \n7: $\\theta_k^t \\gets \\theta^{t-1}$ Download global model \n8: for epoch $e = 1$ to $E$ do \n9: for batch $B$ from $D_k$ , $B_g$ from $\\mathcal{D}_g$ do \n10: $\\mathcal{L}_k^t \\gets \\frac{1}{|B_k|} \\sum_{(\\mathbf{x}, y) \\in B_k} \\ell(f_{\\theta_k^t}(\\mathbf{x}), y) + \\lambda \\frac{1}{|B_g|} \\sum_{(\\mathbf{x}_g, y_g) \\in B_g} \\ell(f_{\\theta_k^t}(\\mathbf{x}_g), y_g)$ \n11: $\\theta_k^t \\gets \\theta_k^t - \\eta \\nabla \\mathcal{L}$ \n12: end for \n13: end for \n14: end for \n15: Server: $\\theta^t \\gets \\frac{1}{|S_t|} \\sum_{k \\in S_t} \\theta_k^t$ >FedAvg \n16: end for", + "bbox": [ + 91, + 109, + 486, + 410 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "dimensions of the image with three color channels, in the client dataset $\\mathcal{D}_k$ , we compute relevances using an attribution method $\\mathcal{A}$ from Section 2.3:", + "bbox": [ + 89, + 435, + 483, + 481 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} = \\mathcal {A} \\left(f _ {\\theta}, \\mathbf {x}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 238, + 491, + 480, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{h} \\in \\mathbb{R}^{H \\times W}$ is the pixel-wise importance score of an input image $\\mathbf{x}$ . These attribution scores quantify the importance of each input feature to the model's prediction. To focus on the most relevant features, we create a binary mask $\\mathbf{m}$ , as shown in Equation 8, by retaining features whose attribution scores exceed a threshold determined by the desired sparsity level $s$ :", + "bbox": [ + 89, + 517, + 483, + 625 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {m} [ i ] = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\mathbf {h} [ i ] \\geq h ^ {(s)} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 633, + 482, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $h^{(s)}$ represents the $s$ -th largest value in the attribution scores. This mask identifies the subset of features that are most crucial for the model's decision-making process.", + "bbox": [ + 89, + 681, + 482, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We define our feature selection function $f_{A} \\colon \\mathbb{R}^{H \\times W \\times 3} \\to \\mathbb{R}^{H \\times W \\times 3}$ as the application of the attribution mask $\\mathbf{m}$ to each image $\\mathbf{x} \\in \\mathcal{D}_k$ through:", + "bbox": [ + 89, + 728, + 483, + 773 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {A} (\\mathbf {x}) = \\mathbf {x} \\odot \\mathbf {m} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 784, + 482, + 800 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This operation emphasizes the most informative parts of the input while suppressing less relevant details. However, directly sharing $f_{A}(\\mathbf{x})$ would still pose privacy risks. Therefore, we design a privacy mechanism $\\mathcal{M}$ that adds calibrated noise to these masked features (Section refsubsec:privacy):", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} (\\mathbf {x}) = f _ {A} (\\mathbf {x}) + \\mathcal {N} (0, \\sigma^ {2} \\mathbf {I}) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 104, + 903, + 122 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\sigma$ is determined based on privacy requirements. The resulting privacy-protected features, along with their corresponding labels, are then aggregated into a new dataset $\\mathcal{D}_k^p = \\{(\\mathcal{M}(\\mathbf{x}),y)|(x,y)\\in \\mathcal{D}_k\\}$ and shared with the server. The full data generated process is outlined in Algorithm 1 and visualized in Figure 1.", + "bbox": [ + 511, + 125, + 906, + 215 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Local Training using Shared Data", + "text_level": 1, + "bbox": [ + 511, + 223, + 808, + 239 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After the privacy-protected features have been obtained, the server aggregates private representations from all clients into a global dataset $\\mathcal{D}_g = \\bigcup_{k=1}^N \\mathcal{D}_k^p$ , and sends this dataset to each client. The clients then optimize a composite objective that balances local task performance with knowledge from the global private dataset:", + "bbox": [ + 511, + 244, + 906, + 335 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {\\theta} \\left[ \\mathbb {E} _ {(\\mathbf {x}, y) \\sim \\mathcal {D} _ {k}} [ \\ell (f _ {\\theta} (\\mathbf {x}), y) ] \\right. \\tag {11} \\\\ + \\lambda \\mathbb {E} _ {(\\mathbf {x}, y) \\sim \\mathcal {D} _ {g}} [ \\ell (f _ {\\theta} (\\mathbf {x}), y) ] ] \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 340, + 903, + 385 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The hyperparameter $\\lambda$ controls the trade-off between local specialization and global knowledge integration. This formulation allows clients to benefit from collective learning while maintaining their own task-specific performance and ensuring privacy. The procedure is outlined in Algorithm 2.", + "bbox": [ + 511, + 388, + 905, + 464 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Privacy Discussion", + "text_level": 1, + "bbox": [ + 511, + 470, + 694, + 486 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We build on the insight that sharing features derived from raw client data can mitigate data heterogeneity in federated learning [34]. Nevertheless, directly sharing these features poses significant privacy risks. A straightforward approach would be to apply differential privacy (DP) to the raw data itself; however, the high dimensionality of image data forces the addition of very large noise levels, thereby significantly degrading utility.", + "bbox": [ + 511, + 492, + 905, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this challenge, we propose a two-step approach. First, we use an attribution-guided masking strategy (Section 4.1) to identify and retain only task-relevant pixels. Second, we add Gaussian noise (Theorem 1) to the resulting masked features to achieve the desired $\\varepsilon$ -metric privacy (Definition 1, Definition 2).", + "bbox": [ + 511, + 614, + 905, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This design offers two main advantages. First, by sparsifying input features that do not contribute to classification predictions, we reduce the overall sensitivity of the data. According to Theorem 1, lower sensitivity requires less noise to meet the same privacy budget $\\varepsilon$ , thus improving utility. Second, in contrast to random or magnitude-based sparsification that may unintentionally discard valuable information, our attribution-guided approach specifically preserves the pixels most critical for the learning task while eliminating irrelevant ones.", + "bbox": [ + 511, + 704, + 905, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The primary goal of our privacy mechanism is to protect against membership inference attacks (MIA) and feature inversion, which we empirically validate in section 6.", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4575", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the following section, we formally demonstrate how our strategy guarantees metric privacy, while a comprehensive discussion of privacy assumptions appears in the appendix.", + "bbox": [ + 89, + 90, + 480, + 137 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3.1. Sensitivity Analysis of Attribution-Based Masking", + "text_level": 1, + "bbox": [ + 89, + 143, + 482, + 159 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A critical component of ensuring privacy is understanding the sensitivity of our feature selection function $f(\\mathbf{x})$ defined in Equation 9. From Theorem 1, the noise required to attain a certain privacy level $\\varepsilon$ depends on the sensitivity $\\Delta_f$ . Our goal is to minimize this sensitivity before applying the privacy mechanism $\\mathcal{M}$ .", + "bbox": [ + 89, + 162, + 482, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For our feature selection function $f_{A}(\\mathbf{x}) = \\mathbf{x}\\odot \\mathbf{m}$ , we can derive a strict bound on sensitivity. Since each coordinate of the mask satisfies $\\mathbf{m}[i]\\in \\{0,1\\}$ , for any two inputs $\\mathbf{x},\\mathbf{x}'\\in \\mathbb{R}^{H\\times W\\times 3}$ , we have, in the worst case:", + "bbox": [ + 89, + 253, + 482, + 313 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left| \\mathbf {m} [ i ] \\mathbf {x} [ i ] - \\mathbf {m} [ i ] \\mathbf {x} ^ {\\prime} [ i ] \\right| \\leq \\left| \\mathbf {x} [ i ] - \\mathbf {x} ^ {\\prime} [ i ] \\right|\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 323, + 413, + 343 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Thus, in aggregate:", + "bbox": [ + 109, + 352, + 241, + 367 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\| f _ {A} (\\mathbf {x}) - f _ {A} \\left(\\mathbf {x} ^ {\\prime}\\right) \\right\\| _ {2} \\leq \\left\\| \\mathbf {x} - \\mathbf {x} ^ {\\prime} \\right\\| _ {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 378, + 398, + 395 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This shows that our masking operation is non-expansive with respect to the $\\ell_2$ norm. Rearranging the equation:", + "bbox": [ + 89, + 406, + 482, + 436 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta_ {f} = \\frac {\\left\\| f _ {A} (\\mathbf {x}) - f _ {A} \\left(\\mathbf {x} ^ {\\prime}\\right) \\right\\| _ {2}}{\\left\\| \\mathbf {x} - \\mathbf {x} ^ {\\prime} \\right\\| _ {2}} \\leq 1 \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 445, + 482, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Delta_f$ is the sensitivity we defined in the preliminaries in Definition 2. Without our attribution-based masking, the sensitivity would be unbounded for arbitrary transformations. Moreover, in the worst case $f_A$ is the identity, i.e. $f_A(x) = x$ , meaning that the attribution method deems all pixels important. This implies that our method would reduce to the naive approach of adding noise to the raw features. In practice, this rarely occurs, since this would already require a highly sparsified image. Nevertheless, our approach safeguards against this worst-case scenario. More discussion can be found in the supplementary material.", + "bbox": [ + 89, + 488, + 482, + 654 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3.2. Privacy Mechanism and Guarantees", + "text_level": 1, + "bbox": [ + 89, + 662, + 390, + 676 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on the sensitivity bound established above, we design our privacy mechanism $\\mathcal{M}$ (Equation 10) by adding Gaussian noise calibrated to this reduced sensitivity:", + "bbox": [ + 89, + 680, + 482, + 726 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} (\\mathbf {x}) = f _ {A} (\\mathbf {x}) + \\mathcal {N} (0, \\sigma^ {2} \\mathbf {I})\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 736, + 385, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\sigma$ is chosen according to Theorem 1 for an $(\\varepsilon, \\delta)$ -level of privacy given sensitivity $\\Delta_f \\leq 1$ . This ensures that for all input images in the client dataset, $\\mathbf{x}, \\mathbf{x}' \\in D_k$ , our method satisfies $(\\varepsilon, \\delta)$ -metric privacy.", + "bbox": [ + 89, + 763, + 482, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our theoretical privacy guarantees are empirically validated through membership inference and feature inversion attacks Section 6, where attribution-masked features consistently show stronger protection than unmasked features under equivalent privacy budgets.", + "bbox": [ + 89, + 825, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experimental", + "text_level": 1, + "bbox": [ + 513, + 89, + 653, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 513, + 114, + 702, + 131 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate FedXDS on standard image classification benchmarks (CIFAR-10, CIFAR-100, Tiny-ImageNet) and real-world federated datasets from the LEAF framework [5] (CelebA, FEMNIST) that naturally exhibit heterogeneous distributions. For the standard benchmarks, we simulate statistical heterogeneity using Dirichlet-based partitioning with concentration parameters $\\alpha \\in \\{0.05, 0.1\\}$ , where smaller values indicate greater heterogeneity. We conduct experiments with both 10 and 100 clients, setting participation rates to 0.5 and 0.1, respectively, and train for $R = 200$ communication rounds. To compute relevance-based attributions, we evaluate four attribution methods in our FL framework: Layerwise Relevance Propagation (FedXLRP) [3], Integrated Gradients (FedXIG) [28], SmoothGrad (FedXSG) [27], and Input $\\times$ Gradient (FedXGrad). Unless otherwise state we use $\\varepsilon = 20$ sparsity level $s = 70\\%$ , and $\\lambda = 0.5$ .", + "bbox": [ + 511, + 137, + 906, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Baselines", + "text_level": 1, + "bbox": [ + 513, + 405, + 620, + 420 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The experimental results in Table 1 show that FedXLRP consistently outperforms all baselines across varying client numbers and data distributions. This advantage amplifies in more challenging scenarios (K=100, $\\alpha = 0.05$ ). FedFed ranks second, while FedAvg and FedAux degrade significantly with increased client numbers. On CIFAR-100, performance gaps narrow, though FedXLRP maintains its edge, particularly with higher client counts and heterogeneity. Other attribution methods (FedXIG, FedXGrad, FedXSG) perform well but remain below FedXLRP, highlighting attribution method choice importance. We discuss this point in Subsection 7.1. In addition, communication efficiency results Table 2 reinforce FedXLRP's advantages. To reach $70\\%$ accuracy on CIFAR-10 with 10 clients, FedXLRP requires only 14 rounds versus 49 for FedAvg and 28 for FedFTG. With 100 clients, FedXLRP needs 23 rounds to reach $60\\%$ accuracy compared to FedAvg's 79. CIFAR-100 shows similar patterns, with FedXLRP achieving target accuracies in substantially fewer rounds across client configurations. Among baselines, only FedFed shows comparable efficiency.", + "bbox": [ + 511, + 428, + 906, + 746 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.3. Experiments on Real-World Datasets", + "text_level": 1, + "bbox": [ + 513, + 756, + 834, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We also evaluate FedXDS on real-world datasets using implementations from the LEAF [5] and [7] frameworks. Specifically, we consider CelebA and FEMNIST, which exhibit a distinct form of non-IIIDness known as distribution shift, introducing an additional challenge for federated learning. For CelebA, we follow the implementation of [36], while for FEMNIST, we use the setup from [7], conducting experiments with 10 and 100 clients, respectively.", + "bbox": [ + 511, + 779, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4576", + "bbox": [ + 480, + 945, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ce507b87acfb0e3a6f3350328feb0690d47b6f2cc4368d46f71aaedd597f4135.jpg", + "table_caption": [ + "Table 1. Performance comparison across different federated learning methods, including standard deviations from 5 runs. Results show the top-1 test accuracy (%) on CIFAR-10, CIFAR-100, and Tiny-ImageNet under varying numbers of clients (K) and Dirichlet concentration parameters $(\\alpha)$ . Bold values indicate the best performance in each column." + ], + "table_footnote": [], + "table_body": "
DatasetCIFAR-10CIFAR-100Tiny-ImageNet
K=10K=100K=10K=100K=10K=100
α = 0.05α = 0.1α = 0.05α = 0.1α = 0.05α = 0.1α = 0.05α = 0.1α = 0.05α = 0.1α = 0.05α = 0.1
FedAvg64.16 ± 0.7172.90 ± 0.5555.53 ± 0.8260.94 ± 0.6843.82 ± 0.4544.57 ± 0.5132.99 ± 0.7335.59 ± 0.6230.30 ± 0.5934.41 ± 0.4826.20 ± 0.8128.79 ± 0.75
FedProx71.56 ± 0.6574.31 ± 0.4951.89 ± 0.9165.01 ± 0.5344.33 ± 0.3848.67 ± 0.4737.79 ± 0.5538.43 ± 0.4933.27 ± 0.4234.54 ± 0.5124.15 ± 0.8830.23 ± 0.67
FedDyn70.62 ± 0.5878.25 ± 0.3463.09 ± 0.6169.41 ± 0.4548.53 ± 0.4149.42 ± 0.3937.14 ± 0.6841.60 ± 0.5433.03 ± 0.4835.69 ± 0.4129.52 ± 0.6332.37 ± 0.59
SCAFFOLD71.35 ± 0.6175.62 ± 0.4258.51 ± 0.7963.82 ± 0.6246.61 ± 0.3347.27 ± 0.4435.28 ± 0.2438.87 ± 0.5831.53 ± 0.5335.76 ± 0.3928.35 ± 0.5930.49 ± 0.61
FedSAM70.15 ± 0.8877.53 ± 0.5162.45 ± 0.7268.78 ± 0.5947.95 ± 0.3548.81 ± 0.4136.88 ± 0.5140.92 ± 0.4732.21 ± 0.5535.03 ± 0.4428.98 ± 0.6631.84 ± 0.52
FedDISCO69.73 ± 0.4476.98 ± 0.6361.82 ± 0.6868.11 ± 0.6547.31 ± 0.5248.15 ± 0.5836.17 ± 0.6740.25 ± 0.6131.78 ± 0.6134.68 ± 0.5228.51 ± 0.7131.33 ± 0.64
FedFed79.12 ± 0.4882.58 ± 0.3175.35 ± 0.5778.87 ± 0.4352.27 ± 0.4956.45 ± 0.3845.41 ± 0.7649.28 ± 0.5234.99 ± 0.4536.49 ± 0.3733.28 ± 0.5834.89 ± 0.47
FedFTG75.21 ± 0.5378.44 ± 0.3970.84 ± 0.3174.85 ± 0.4144.91 ± 0.6254.15 ± 0.4540.43 ± 0.3946.47 ± 0.4832.40 ± 0.5733.79 ± 0.4930.53 ± 0.5132.26 ± 0.55
FedGen68.34 ± 0.6874.65 ± 0.4758.90 ± 0.7563.20 ± 0.6644.29 ± 0.5550.12 ± 0.5136.70 ± 0.6239.82 ± 0.5734.17 ± 0.4436.33 ± 0.4229.45 ± 0.6831.60 ± 0.61
FedAux59.72 ± 0.8573.70 ± 0.5153.13 ± 0.8865.11 ± 0.5944.05 ± 0.4844.85 ± 0.5336.54 ± 0.6538.39 ± 0.6127.83 ± 0.7234.31 ± 0.4924.75 ± 0.8230.38 ± 0.69
FedDF60.56 ± 0.7772.40 ± 0.5859.20 ± 0.7263.47 ± 0.6430.13 ± 0.8145.47 ± 0.5034.46 ± 0.6936.77 ± 0.6429.07 ± 0.6534.75 ± 0.4628.42 ± 0.6230.51 ± 0.63
FedXLRP81.72 ± 0.3583.46 ± 0.2877.02 ± 0.6380.27 ± 0.3952.07 ± 0.4558.09 ± 0.3346.25 ± 0.8252.63 ± 0.4136.85 ± 0.3938.64 ± 0.3234.64 ± 0.5233.18 ± 0.51
FedXIG72.22 ± 0.5975.89 ± 0.4464.02 ± 0.6968.61 ± 0.5549.21 ± 0.4855.46 ± 0.4142.34 ± 0.5847.89 ± 0.5033.59 ± 0.5135.30 ± 0.4529.77 ± 0.6431.93 ± 0.58
FedXGrad71.38 ± 0.6273.99 ± 0.5264.52 ± 0.6566.49 ± 0.6150.13 ± 0.4354.21 ± 0.4642.11 ± 0.6046.98 ± 0.5333.18 ± 0.5434.40 ± 0.4830.00 ± 0.6130.93 ± 0.60
FedXSG71.89 ± 0.6074.63 ± 0.4963.75 ± 0.7167.80 ± 0.5851.57 ± 0.3955.22 ± 0.4240.91 ± 0.6347.61 ± 0.5133.43 ± 0.5234.74 ± 0.4729.64 ± 0.6531.53 ± 0.59
", + "bbox": [ + 91, + 141, + 906, + 357 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/411f05dea331c732f6ee5b49abfdb9637306957c1d35e573654ce456992e8621.jpg", + "table_caption": [ + "Table 2. Communication efficiency comparison across different federated learning methods. Results show the number of communication rounds (mean ± std over 5 runs) needed to achieve target accuracy thresholds (70% and 60% for CIFAR-10; 40% and 30% for CIFAR-100; 35% and 30% for Tiny-ImageNet) with different numbers of clients (K=10 and K=100) and heterogeneity parameter $\\alpha = 0.1$ . Lower values indicate better communication efficiency. Best results are in bold." + ], + "table_footnote": [], + "table_body": "
MethodCIFAR-10CIFAR-100Tiny-ImageNet
acc=70% K=10acc=60% K=100acc=40% K=10acc=30% K=100acc=35% K=10acc=30% K=100
FedAvg49 ± 579 ± 724 ± 434 ± 560 ± 690 ± 8
FedProx32 ± 355 ± 428 ± 432 ± 445 ± 465 ± 5
FedDyn29 ± 351 ± 424 ± 332 ± 348 ± 560 ± 4
SCAFFOLD33 ± 452 ± 525 ± 338 ± 446 ± 469 ± 6
FedSAM34 ± 458 ± 527 ± 435 ± 450 ± 564 ± 5
FedDISCO36 ± 361 ± 628 ± 437 ± 552 ± 567 ± 6
FedFed15 ± 222 ± 413 ± 412 ± 312 ± 215 ± 2
FedFTG28 ± 357 ± 422 ± 328 ± 330 ± 340 ± 4
FedGen51 ± 566 ± 621 ± 439 ± 555 ± 570 ± 6
FedAux38 ± 445 ± 419 ± 336 ± 442 ± 450 ± 5
FedDF84 ± 790 ± 828 ± 445 ± 5100 ± 8110 ± 9
FedXLRP14 ± 223 ± 311 ± 313 ± 210 ± 412 ± 2
FedXIG35 ± 439 ± 416 ± 221 ± 338 ± 445 ± 4
FedXGrad41 ± 446 ± 513 ± 226 ± 344 ± 450 ± 5
FedXSG41 ± 551 ± 421 ± 333 ± 442 ± 448 ± 5
", + "bbox": [ + 246, + 436, + 750, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Additional implementation details can be found in the supplementary material. The results in Table 3 show a clear trend where FedXDS consistently achieves the highest accuracy across both datasets, demonstrating its effectiveness in handling distribution shift. Notably, methods incorporating additional data generation or augmentation, such as FedFTG and FedGen, also perform well, but FedXDS surpasses them, suggesting that relevance-guided data sharing provides a stronger mechanism for improving generaliza", + "bbox": [ + 89, + 750, + 480, + 887 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tion in heterogeneous federated settings.", + "bbox": [ + 511, + 750, + 781, + 766 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The results in Table 4 demonstrate that integrating FedXDS with FL methods consistently improves the performance. In all cases, the addition of FedXDS yields substantial accuracy gains, particularly in the highly heterogeneous setting with $\\mathrm{K} = 100$ clients. For instance, FedAvg, which traditionally suffers from performance degradation in non-IID scenarios, improves from $60.94\\%$ to $80.29\\%$ when combined with FedXDS. Similar improvements are observed for", + "bbox": [ + 511, + 780, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4577", + "bbox": [ + 480, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1e38c3edf27b809d1770507e727061b3eb22e57c571b7ee3ca7f2cb2817844ee.jpg", + "table_caption": [ + "Table 3. Accuracy comparison (mean ± std % over 5 runs) on CelebA and FEMNIST. Best results are in bold. We use the LRP variant of FedXDS." + ], + "table_footnote": [], + "table_body": "
MethodCelebAFEMNIST
FedAvg87.23 ± 0.7184.71 ± 0.68
FedProx87.79 ± 0.6585.66 ± 0.61
SCAFFOLD86.36 ± 0.5684.24 ± 0.59
FedDyn88.13 ± 0.5286.45 ± 0.34
FedSAM90.03 ± 0.1587.56 ± 0.44
FedDISCO89.78 ± 0.6187.29 ± 0.62
FedFTG90.31 ± 0.4387.88 ± 0.67
FedGen89.65 ± 0.5986.37 ± 0.55
FedDF88.56 ± 0.6385.28 ± 0.66
FedAux89.27 ± 0.6085.94 ± 0.62
FedFed90.76 ± 0.8388.28 ± 0.71
FedXDS91.55 ± 0.4889.03 ± 0.35
", + "bbox": [ + 138, + 142, + 433, + 345 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bb123525ad79e91959dda314deb244513f045b98c23137bb88a65de51ef7ef95.jpg", + "table_caption": [ + "Table 4. Performance comparison (mean ± std % over 5 runs) on CIFAR-10 for different base FL methods and their FedXDS (using LRP) variants under two heterogeneity settings ( $K = 10$ and $K = 100$ ). The FedXDS variants consistently improve performance." + ], + "table_footnote": [], + "table_body": "
MethodK=10K=100
FedAvg72.90 ± 0.5560.94 ± 0.68
FedAvg + FedXDS83.46 ± 0.3180.29 ± 0.42
FedProx74.31 ± 0.4965.01 ± 0.53
FedProx + FedXDS84.05 ± 0.3580.73 ± 0.45
FedDyn70.62 ± 0.5869.41 ± 0.45
FedDyn + FedXDS84.20 ± 0.3379.56 ± 0.48
SCAFFOLD71.35 ± 0.6163.82 ± 0.62
SCAFFOLD + FedXDS83.89 ± 0.3879.72 ± 0.44
", + "bbox": [ + 101, + 426, + 468, + 588 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "FedProx, FedDyn, and SCAFFOLD, with FedXDS enhancing their robustness to data heterogeneity while maintaining their inherent advantages.", + "bbox": [ + 89, + 612, + 482, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The magnitude of improvement suggests that FedXDS effectively mitigates the adverse effects of statistical heterogeneity by facilitating better knowledge transfer across clients. Notably, FedDyn, which incorporates regularization-based local adaptation, exhibits the smallest absolute improvement with FedXDS, indicating that its existing adaptation mechanisms already partially address heterogeneity. In contrast, methods like FedAvg, without correction for client drift, benefit more from FedXDS.", + "bbox": [ + 88, + 657, + 482, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Empirical Privacy Analysis", + "text_level": 1, + "bbox": [ + 89, + 806, + 339, + 825 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.1. Membership Inference Attack", + "text_level": 1, + "bbox": [ + 89, + 832, + 359, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct membership inference attacks following [30] to evaluate our methodology's privacy guarantees. We investigate whether training on masked and noised features reveals", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/da409e42317998b20580432056ba38051b6619f098a3beeaeb66a1d62344a9c3.jpg", + "table_caption": [ + "Table 5. Test accuracy (mean ± std % over 5 runs) for different sparsification levels $s$ across various attribution methods. Best results are in bold. Results for CIFAR10 at $K = 10, \\alpha = 0.1$ ." + ], + "table_footnote": [], + "table_body": "
sGradientSmoothGradInt. Grad.LRP
6075.99 ± 0.5275.63 ± 0.5577.10 ± 0.4884.74 ± 0.31
6574.19 ± 0.5673.15 ± 0.6176.31 ± 0.5184.01 ± 0.35
7073.99 ± 0.5974.63 ± 0.5475.89 ± 0.4983.46 ± 0.28
7570.54 ± 0.6570.67 ± 0.6371.68 ± 0.5883.01 ± 0.33
8068.31 ± 0.7170.45 ± 0.6670.98 ± 0.6282.67 ± 0.36
8568.80 ± 0.6868.13 ± 0.7369.73 ± 0.6781.88 ± 0.41
", + "bbox": [ + 514, + 142, + 906, + 247 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "membership in the underlying dataset by using the globally DP data to train a shadow model, then testing if clients can infer membership by querying this model and training a random forest classifier. Figure 2a shows recall values for different privacy parameters $\\varepsilon$ . Results demonstrate that non-masked features, despite DP protection, offer inferior privacy protection compared to masked data. This can be explained by the fact that sparsity significantly mitigates attacks through norm reduction from attribution masks (Equation 8).", + "bbox": [ + 511, + 273, + 906, + 424 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2. Feature Inversion Attack", + "text_level": 1, + "bbox": [ + 511, + 435, + 741, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "While metric private features prevent direct inference of individual data features, a sophisticated adversary might train a denoising autoencoder to learn mappings between protected representations and original images. We conduct feature inversion attacks and report the SSIM score between original and reconstructed images. Figure 2b reveals inversion attack success increases with lower sparsity and higher privacy budgets. Unmasked features consistently exhibit the highest SSIM, confirming that raw features under identical DP constraints provide inferior protection compared to our approach. Additional validations in supplementary material, including visualization of inversion results for non-masked features and FedFed [30] features, confirming both cases' vulnerability to reconstruction, whereas our method provides superior protection.", + "bbox": [ + 511, + 458, + 908, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "7. Ablation Studies and Discussion", + "text_level": 1, + "bbox": [ + 511, + 700, + 807, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "7.1. XAI Method Comparison Across Sparsification", + "text_level": 1, + "bbox": [ + 511, + 726, + 906, + 743 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5 demonstrates that FedXDS maintains stable performance across different sparsification levels, with all attribution methods achieving above $68\\%$ accuracy, even at high sparsity thresholds. Notably, LRP consistently outperforms other methods, preserving accuracies between $81.88\\%$ and $84.74\\%$ , while others degrade more sharply under extreme sparsification. LRP's robustness stems from its layer-wise relevance propagation mechanism, which is fundamentally different from gradient-based approaches. Instead of measuring local sensitivity, LRP redistributes the", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4578", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d654f65426f1b09337378c9b8bb673358720c17eb2d5a29177b9ccba3e8b785e.jpg", + "image_caption": [ + "(a) Recall score of MI attack at different $\\varepsilon$ for various sparsity levels $s$ ." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 285, + 178 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/334fbfd003e818507f5a5d76ad2c6de6ee2b76487b76bccdd60b2ee70ed5646a.jpg", + "image_caption": [ + "(b) SSIM score of denoising attack at different $\\varepsilon$ and sparsity levels $s$ ." + ], + "image_footnote": [], + "bbox": [ + 297, + 90, + 491, + 179 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0507348d4aaec163abc51ebeba93ede0d6190e7357363e40077b79034915f19b.jpg", + "image_caption": [ + "(c) Effect of knowledge weight $\\lambda$ on model accuracy across datasets." + ], + "image_footnote": [], + "bbox": [ + 503, + 90, + 697, + 178 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7c0e6832821a718b23e96a981dcd2488e770e4ce2890d8fc52319f8cc3cb09ee.jpg", + "image_caption": [ + "(d) Impact of warmup rounds on model accuracy across datasets.", + "Figure 2. Evaluation of privacy-utility trade-offs and hyperparameter effects in our framework." + ], + "image_footnote": [], + "bbox": [ + 707, + 90, + 901, + 179 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "model's output prediction backward through the network, layer by layer, conserving relevance at each step. This process aggregates relevance into contiguous, semantically coherent regions that represent structured object parts. In stark contrast, methods like SmoothGrad or Integrated Gradients identify 'hotspot' pixels that are locally influential but often disconnected from each other. Consequently, under high sparsification, LRP retains meaningful visual patterns, whereas other methods are left with a sparse collection of isolated points that lack sufficient context for accurate classification. We conclude that structural coherence in saliency maps is an essential criterion when selecting the attribution method. This finding is corroborated by current literature [2, 13, 23], which confirms LRP's superior explanation quality and reinforces that our work is consistent with established findings on the importance of explanation quality for downstream tasks. More discussion can be found in the supplementary material.", + "bbox": [ + 89, + 260, + 483, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7.2. Privacy vs. Utility", + "text_level": 1, + "bbox": [ + 89, + 542, + 264, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We investigate how the privacy budget $\\varepsilon$ impacts model performance across attribution methods. Table 6 shows that increasing $\\varepsilon$ improves accuracy for all methods, demonstrating the expected privacy-utility trade-off: stronger privacy (lower $\\varepsilon$ ) leads to reduced performance due to greater noise. Among all methods, LRP consistently achieves the highest accuracy across privacy levels outperforming the next best (Integrated Gradients) by $7 - 8\\%$ and maintaining robustness even under strict privacy constraints. SmoothGrad shows modest improvements over basic gradients, though less pronounced than Integrated Gradients. These results underscore LRP's ability to extract stable, task-relevant features leading to better generalization with less privacy leakage, aligning with prior work on explanation fidelity [3, 22].", + "bbox": [ + 89, + 565, + 483, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7.3. Ablation Study on the Knowledge Weight", + "text_level": 1, + "bbox": [ + 89, + 787, + 444, + 803 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluate how the knowledge weight $\\lambda$ , as defined in Equation 11, affects performance across CIFAR-10, CIFAR-100, and Tiny-ImageNet. Figure 2c shows that intermediate values ( $\\lambda = 0.5$ ) consistently outperform both extremes, with CIFAR-10 accuracy peaking at $83.46\\%$ compared to $72.90\\%$ ( $\\lambda = 0.0$ , equivalent to FedAvg) and", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/aba453af7d5b6dc1e2d8de56fc42f279a3308965abcb304db4c6b33acffcd584.jpg", + "table_caption": [ + "Table 6. Ablation study showing test accuracy) for different privacy budgets $(\\varepsilon)$ across XAI-based federated learning methods. Results for CIFAR10 at $K = 10, \\alpha = 0.1$" + ], + "table_footnote": [], + "table_body": "
εGradientSmoothGradInt. Grad.LRP
570.55 ± 0.6271.23 ± 0.5972.45 ± 0.5280.06 ± 0.33
771.45 ± 0.5872.15 ± 0.5573.35 ± 0.4780.89 ± 0.31
1072.25 ± 0.5172.95 ± 0.4874.15 ± 0.4181.65 ± 0.29
1573.12 ± 0.4573.85 ± 0.4375.02 ± 0.3882.11 ± 0.25
2073.99 ± 0.4274.63 ± 0.3975.89 ± 0.3583.46 ± 0.28
", + "bbox": [ + 514, + 311, + 905, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "61.18% ( $\\lambda = 1.0$ , synthetic data equal weight). Our ablation thus shows that balancing the contribution of synthetic data is critical for performance.", + "bbox": [ + 511, + 430, + 905, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7.4. Warmup and Efficiency", + "text_level": 1, + "bbox": [ + 511, + 487, + 732, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 2d shows accuracy slightly improves with more warmup rounds, plateauing at 10, suggesting this is the optimal trade-off between convergence and communication cost. Our approach is also computationally efficient. Attribution masks are computed in a single backward pass during warmup, incurring minimal overhead. In contrast, methods like FedGen and FedFTG retrain generators every round, and FedFed requires costly VAE training—making FedXDS more suitable for resource-constrained settings. We provide more discussion in the supplementary material.", + "bbox": [ + 511, + 510, + 906, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 678, + 633, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduced FedXDS, leveraging attribution methods to address statistical heterogeneity, privacy, and computational efficiency in federated learning. By using attribution maps to extract and share privacy-preserved task-relevant features, our approach enables effective knowledge transfer while ensuring strong privacy guarantees. Experiments show that FedXDS consistently outperforms existing methods in accuracy and communication efficiency. Moreover, the sparsity induced by relevance attribution enhances privacy, mitigating feature inversion and membership inference risks. These findings highlight attribution methods as a promising tool for improving federated learning while preserving privacy and efficiency.", + "bbox": [ + 511, + 703, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4579", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "9. Acknowledgements", + "text_level": 1, + "bbox": [ + 89, + 90, + 277, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the European Union's Horizon Europe research and innovation programme (EU Horizon Europe) as grant ACHILLES (101189689).", + "bbox": [ + 89, + 113, + 485, + 170 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 181, + 187, + 196 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Durmus Alp Emre Acar, Yue Zhao, Ramon Matas, Matthew Mattina, Paul Whatmough, and Venkatesh Saligrama. Federated learning based on dynamic regularization. In International Conference on Learning Representations, 2021. 3", + "[2] Leila Arras, Ahmed Osman, and Wojciech Samek. Clevr-xai: A benchmark dataset for the ground truth evaluation of neural network explanations. Information Fusion, 81:14-40, 2022. 1, 8", + "[3] Sebastian Bach, Alexander Binder, Grégoire Montavon, Frederick Klauschen, Klaus-Robert Müller, and Wojciech Samek. On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. PLOS ONE, 10(7):1-46, 2015. 3, 5, 8", + "[4] Jeremiah Blocki, Avrim Blum, Anupam Datta, and Or Sheffet. Differentially private data analysis of social networks via restricted sensitivity. In Proceedings of the 4th conference on Innovations in Theoretical Computer Science, pages 87-96, 2013. 2", + "[5] Sebastian Caldas, Sai Meher Karthik Duddu, Peter Wu, Tian Li, Jakub Konečný, H Brendan McMahan, Virginia Smith, and Ameet Talwalkar. Leaf: A benchmark for federated settings. arXiv preprint arXiv:1812.01097, 2018. 5", + "[6] Konstantinos Chatzikokolakis, Miguel E Andres, Nicolás Emilio Bordenabe, and Catuscia Palamidessi. Broadening the scope of differential privacy using metrics. In Privacy Enhancing Technologies: 13th International Symposium, PETS 2013, Bloomington, IN, USA, July 10-12, 2013. Proceedings 13, pages 82-102. Springer, 2013. 2", + "[7] Daoyuan Chen, Dawei Gao, Weirui Kuang, Yaliang Li, and Bolin Ding. pfl-bench: A comprehensive benchmark for personalized federated learning. Advances in Neural Information Processing Systems, 35:9344-9360, 2022. 5", + "[8] Rudresh Dwivedi, Devam Dave, Het Naik, Smiti Singhal, Rana Omer, Pankesh Patel, Bin Qian, Zhenyu Wen, Tejal Shah, Graham Morgan, et al. Explainable ai (xai): Core ideas, techniques, and solutions. ACM Computing Surveys, 55(9):1-33, 2023. 1", + "[9] Cynthia Dwork. Differential privacy. In International colloquium on automata, languages, and programming, pages 1-12. Springer, 2006. 1, 2", + "[10] Oluwaseyi Feyisetan and Shiva Kasiviswanathan. Private release of text embedding vectors. In Proceedings of the First Workshop on Trustworthy Natural Language Processing, pages 15-27, 2021. 2", + "[11] Peter Kairouz, H Brendan McMahan, Brendan Avent, Aurélien Bellet, Mehdi Bennis, Arjun Nitin Bhagoji, Kallista Bonawitz, Zachary Charles, Graham Cormode, Rachel Cummings, et al. Advances and open problems in federated learning. Foundations and trends® in machine learning, 14(1-2): 1-210, 2021. 1" + ], + "bbox": [ + 93, + 205, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank Reddi, Sebastian Stich, and Ananda Theertha Suresh. SCAFFOLD: Stochastic controlled averaging for federated learning. In Proceedings of the 37th International Conference on Machine Learning, pages 5132-5143. PMLR, 2020. 1, 3", + "[13] Sebastian Lapuschkin, Stephan Wäldchen, Alexander Binder, Grégoire Montavon, Wojciech Samek, and Klaus-Robert Müller. Unmasking clever hans predictors and assessing what machines really learn. Nature communications, 10(1):1096, 2019. 8", + "[14] Bo Li, Mikkel N Schmidt, Tommy S Alstrom, and Sebastian U Stich. On the effectiveness of partial variance reduction in federated learning with heterogeneous data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3964-3973, 2023. 1", + "[15] Qinbin Li, Bingsheng He, and Dawn Song. Model-contrastive federated learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10713–10722, 2021. 3", + "[16] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. Proceedings of Machine learning and systems, 2:429-450, 2020. 1, 3", + "[17] Xiaoxiao Li, Meirui Jiang, Xiaofei Zhang, Michael Kamp, and Qi Dou. Fed{bn}: Federated learning on non- $\\{\\mathrm{id}\\}$ features via local batch normalization. In International Conference on Learning Representations, 2021. 3", + "[18] Tao Lin, Lingjing Kong, Sebastian U Stich, and Martin Jaggi. Ensemble distillation for robust model fusion in federated learning. Advances in neural information processing systems, 33:2351-2363, 2020. 3", + "[19] H. Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Agüera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Proceedings of the 20th International Conference on Artificial Intelligence and Statistics (AISTATS). PMLR, 2017. 1, 2, 3", + "[20] Matias Mendieta, Taojiannan Yang, Pu Wang, Minwoo Lee, Zhengming Ding, and Chen Chen. Local learning matters: Rethinking data heterogeneity in federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8397-8406, 2022. 1", + "[21] Zhe Qu, Xingyu Li, Rui Duan, Yao Liu, Bo Tang, and Zhuo Lu. Generalized federated learning via sharpness aware minimization. In International conference on machine learning, pages 18250-18280. PMLR, 2022. 1, 3", + "[22] Wojciech Samek, Alexander Binder, Grégoire Montavon, Sebastian Lapischkin, and Klaus-Robert Müller. Evaluating the visualization of what a deep neural network has learned. IEEE Transactions on Neural Networks and Learning Systems, 28(11):2660-2673, 2017. 8", + "[23] Wojciech Samek, Grégoire Montavon, Sebastian Lapischkin, Christopher J. Anders, and Klaus-Robert Müller. Explaining deep neural networks and beyond: A review of methods and applications. Proceedings of the IEEE, 109(3): 247-278, 2021. 8" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4580", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Felix Sattler, Tim Korjakow, Roman Rischke, and Wojciech Samek. Fedaux: Leveraging unlabeled auxiliary data in federated learning. IEEE Transactions on Neural Networks and Learning Systems, 34(9):5531-5543, 2021. 3", + "[25] Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside convolutional networks: Visualising image classification models and saliency maps. In International Conference on Learning Representations (ICLR), 2014. 3", + "[26] Abhishek Singh, Praneeth Vepakomma, Vivek Sharma, and Ramesh Raskar. Posthoc privacy guarantees for collaborative inference with modified propose-test-release. Advances in Neural Information Processing Systems, 36:26438-26451, 2023. 2", + "[27] Daniel Smilkov, Nikhil Thorat, Been Kim, Fernanda Viégas, and Martin Wattenberg. Smoothgrad: removing noise by adding noise. arXiv preprint arXiv:1706.03825, 2017. 3, 5", + "[28] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In International Conference on Machine Learning, pages 3319-3328. PMLR, 2017. 3, 5", + "[29] Jianyu Wang, Qinghua Liu, Hao Liang, Gauri Joshi, and H. Vincent Poor. Tackling the objective inconsistency problem in heterogeneous federated optimization. In Proceedings of the 34th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2020. Curran Associates Inc. 3", + "[30] Zhiqin Yang, Yonggang Zhang, Yu Zheng, Xinmei Tian, Hao Peng, Tongliang Liu, and Bo Han. Fedfed: Feature distillation against data heterogeneity in federated learning. Advances in Neural Information Processing Systems, 36, 2024. 1, 3, 7", + "[31] Mang Ye, Xiuwen Fang, Bo Du, Pong C Yuen, and Dacheng Tao. Heterogeneous federated learning: State-of-the-art and research challenges. ACM Computing Surveys, 56(3):1-44, 2023. 1", + "[32] Rui Ye, Mingkai Xu, Jianyu Wang, Chenxin Xu, Siheng Chen, and Yanfeng Wang. Feddisco: Federated learning with discrepancy-aware collaboration. In International Conference on Machine Learning, pages 39879-39902. PMLR, 2023. 3", + "[33] Lin Zhang, Li Shen, Liang Ding, Dacheng Tao, and Ling-Yu Duan. Fine-tuning global model via data-free knowledge distillation for non-iid federated learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10174-10183, 2022. 1, 3", + "[34] Yue Zhao, Meng Li, Liangzhen Lai, Naveen Suda, Damon Civin, and Vikas Chandra. Federated learning with non-iid data. arXiv preprint arXiv:1806.00582, 2018. 1, 4", + "[35] Yue Zhao, Meng Li, Liangzhen Lai, Naveen Suda, Damon Civin, and Vikas Chandra. Federated learning with non-iid data. arXiv preprint arXiv:1806.00582, 2018. 1", + "[36] Zhuangdi Zhu, Junyuan Hong, and Jiayu Zhou. Data-free knowledge distillation for heterogeneous federated learning. In Proceedings of the 38th International Conference on Machine Learning, pages 12878-12889. PMLR, 2021. 1, 3, 5" + ], + "bbox": [ + 91, + 92, + 480, + 854 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4581", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/FedXDS_ Leveraging Model Attribution Methods to counteract Data Heterogeneity in Federated Learning/3d4823b4-ec38-431a-8ed1-cb351135854c_model.json b/2025/FedXDS_ Leveraging Model Attribution Methods to counteract Data Heterogeneity in Federated Learning/3d4823b4-ec38-431a-8ed1-cb351135854c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..75530392f7be657aca286770d9421f25e1b7bb32 --- /dev/null +++ b/2025/FedXDS_ Leveraging Model Attribution Methods to counteract Data Heterogeneity in Federated Learning/3d4823b4-ec38-431a-8ed1-cb351135854c_model.json @@ -0,0 +1,2365 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.155, + 0.13, + 0.844, + 0.177 + ], + "angle": 0, + "content": "FedXDS: Leveraging Model Attribution Methods to counteract Data Heterogeneity in Federated Learning" + }, + { + "type": "text", + "bbox": [ + 0.205, + 0.203, + 0.792, + 0.275 + ], + "angle": 0, + "content": "Maximilian Andreas Hoefler\\(^{1}\\) * Karsten Mueller\\(^{1}\\) Wojciech Samek\\(^{1,2,3}\\) \n\\(^{1}\\)Fraunhofer Heinrich Hertz Institute, Berlin, Germany \n\\(^{2}\\)Technical University of Berlin, Berlin, Germany \n\\(^{3}\\)Berlin Institute for the Foundations of Learning and Data (BIFOLD)" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.342, + 0.482, + 0.778 + ], + "angle": 0, + "content": "Explainable AI (XAI) methods have demonstrated significant success in recent years at identifying relevant features in input data that drive deep learning model decisions, enhancing interpretability for users. However, the potential of XAI beyond providing model transparency has remained largely unexplored in adjacent machine learning domains. In this paper, we show for the first time how XAI can be utilized in the context of federated learning. Specifically, while federated learning enables collaborative model training without raw data sharing, it suffers from performance degradation when client data distributions exhibit statistical heterogeneity. We introduce FedXDS (Federated Learning via XAI-guided Data Sharing), the first approach to utilize feature attribution techniques to identify precisely which data elements should be selectively shared between clients to mitigate heterogeneity. By employing propagation-based attribution, our method identifies task-relevant features through a single backward pass, enabling selective data sharing that aligns client contributions. To protect sensitive information, we incorporate metric privacy techniques that provide formal privacy guarantees while preserving utility. Experimental results demonstrate that our approach consistently achieves higher accuracy and faster convergence compared to existing methods across varying client numbers and heterogeneity settings. We provide theoretical privacy guarantees and empirically demonstrate robustness against both membership inference and feature inversion attacks. Code is available at https://github.com/MaxH1996/FedXDS." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.808, + 0.222, + 0.823 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.833, + 0.483, + 0.879 + ], + "angle": 0, + "content": "Federated learning (FL) [19] has received significant attention in recent years, proving to be an effective method for collaborative distributed machine learning while keeping" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.311, + 0.905, + 0.356 + ], + "angle": 0, + "content": "local datasets private. However, the practical deployment of federated learning systems faces fundamental challenges regarding data heterogeneity and privacy [11, 35]." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.357, + 0.907, + 0.553 + ], + "angle": 0, + "content": "Data heterogeneity, characterized by non-IID data across clients, leads to contradicting model updates and significantly degraded performance [12, 31]. Current approaches address this challenge by aligning client distributions using proximal optimization terms [16] and mitigating client drift [12, 14], while others focus on flattening the loss landscape to facilitate better model aggregation [20, 21]. In addition, recent works have shown that sharing even a small portion of client data globally can significantly improve local model generalization [34]. Such shared data introduces centralization, giving clients access to a more homogeneous distribution, which reduces update divergence and mitigates heterogeneity [34]." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.554, + 0.906, + 0.749 + ], + "angle": 0, + "content": "However, sharing raw data is not feasible due to privacy constraints. Introducing differential privacy [9] by adding noise to the shared data offers a potential solution, however this can cause considerable performance loss. Alternatively, instead of sharing raw data, abstract feature representations or partial data could be released using generator-type approaches [30, 33, 36]. However, the obtained features are not as potent as raw data in mitigating statistical heterogeneity, and can also be prone to privacy leaks and additionally introduce significant computational overhead due to generators. Hence, the challenge arises of how to retain the performance improvement from data sharing while simultaneously guaranteeing privacy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.905, + 0.825 + ], + "angle": 0, + "content": "To address this challenge we propose FedXDS (Federated Learning via XAI-guided Data Sharing) which entails a novel approach wherein we leverage methods from the field of eXplainable AI (XAI) [8] enabling performance enhancing data sharing while preserving privacy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Specifically, we leverage propagation-based attribution methods, which have been shown to reliably identify input features that consistently contribute to model predictions [2, 8]. Moreover, attribution maps yield a pixel-wise relevance score which highlights regions in the input space" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.466, + 0.9 + ], + "angle": 0, + "content": "*Correspondence to maximilian.andreas.hoefler@hhi.fraunhofer.de" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4572" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "which focus on semantically meaningful features, while suppressing spurious correlations and background noise. In our setup local client data is filtered through the attribution mechanism, i.e., we only share the input features which the attribution map deems most relevant for model predictions. This retains the information needed for model generalization while discarding irrelevant information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.198, + 0.482, + 0.333 + ], + "angle": 0, + "content": "We show that our method, when combined with differential-metric privacy techniques [6, 9], offers stronger privacy guarantees compared to applying similar privacy mechanisms directly to raw input features. Specifically, by discarding task-irrelevant information we effectively reduce the dimensionality of the raw features. This allows us to selectively protect only the task-relevant regions rather than uniformly protecting all input pixels, and thus obtain better utility." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.334, + 0.482, + 0.395 + ], + "angle": 0, + "content": "In addition, propagation-based attributions can be obtained via a single backward pass over the local dataset, which only needs to be performed once in the entire FL process." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.395, + 0.4, + 0.409 + ], + "angle": 0, + "content": "Our main contributions can be summarized as:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.411, + 0.482, + 0.47 + ], + "angle": 0, + "content": "- A novel federated learning algorithm leveraging propagation-based attributions to address the challenge of retaining performance from data sharing with privacy guarantees in heterogeneous environments." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.471, + 0.482, + 0.576 + ], + "angle": 0, + "content": "- A privacy-preserving mechanism that utilizes attribution-guided dimensionality reduction to achieve metric differential privacy. We perform theoretical and empirical privacy evaluations on our feature sharing approach through membership inference attacks and feature inversion, demonstrating that we can attain strong privacy guarantees." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.577, + 0.482, + 0.638 + ], + "angle": 0, + "content": "- Extensive experiments on image datasets benchmarking our method against current state of the art, demonstrating superior performance in terms of accuracy and convergence." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.411, + 0.482, + 0.638 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.652, + 0.228, + 0.668 + ], + "angle": 0, + "content": "2. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.677, + 0.28, + 0.694 + ], + "angle": 0, + "content": "2.1. Federated Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.699, + 0.482, + 0.82 + ], + "angle": 0, + "content": "Federated Learning [19] is a distributed machine learning paradigm where multiple clients collaborate to train a model under the orchestration of a central server, without sharing their raw data. Let \\(K\\) be the number of clients, each with a local dataset \\(\\mathcal{D}_k = \\{(x_i^k,y_i^k)\\}_{i = 1}^{n_k}\\), where \\(x_{i}^{k}\\in \\mathcal{X}\\) is an input sample and \\(y_{i}^{k}\\in \\mathcal{V}\\) is its corresponding label. The goal is to learn a global model \\(f_{\\theta}:\\mathcal{X}\\to \\mathcal{Y}\\) parameterized by \\(\\theta\\), which minimizes the empirical risk:" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.834, + 0.482, + 0.874 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\frac {1}{N} \\sum_ {k = 1} ^ {K} \\sum_ {i = 1} ^ {n _ {k}} \\mathcal {L} \\left(f _ {\\theta} \\left(x _ {i} ^ {k}\\right), y _ {i} ^ {k}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.884, + 0.482, + 0.903 + ], + "angle": 0, + "content": "where \\(N = \\sum_{k=1}^{K} n_k\\) is the total number of samples" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.845, + 0.106 + ], + "angle": 0, + "content": "across all clients, and \\(\\mathcal{L}\\) is a suitable loss function." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.107, + 0.905, + 0.171 + ], + "angle": 0, + "content": "The FedAvg algorithm [19] iteratively aggregates model updates from selected clients \\( S_{t} \\), where each client \\( k \\) performs local updates \\( \\theta_{k}^{t} = \\theta^{t - 1} - \\eta \\nabla \\mathcal{L}_{k}(\\theta^{t - 1}) \\) before the server averages these updates as \\( \\theta^t = \\frac{1}{|S_t|}\\sum_{k\\in S_t}\\theta_k^t \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.177, + 0.834, + 0.193 + ], + "angle": 0, + "content": "2.2. Differential Privacy in Metric Spaces" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.198, + 0.905, + 0.29 + ], + "angle": 0, + "content": "Standard differential privacy (DP) [9] provides guarantees against membership inference by bounding the output distribution change when a single dataset element is modified. For continuous data such as image embeddings, classical adjacency notions are restrictive, making metric differential privacy [6] more appropriate:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.295, + 0.906, + 0.356 + ], + "angle": 0, + "content": "Definition 1 (Metric Privacy). [6] Let \\((X, d_X)\\) be a metric space and \\(Z\\) a set of possible outputs. A randomized mechanism \\(A\\colon X \\to \\Delta(Z)\\) is \\((\\varepsilon, \\delta)\\)-metric private if for all \\(x, x' \\in X\\) and every measurable set \\(U \\subseteq Z\\):" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.363, + 0.905, + 0.388 + ], + "angle": 0, + "content": "\\[\n\\Pr [ A (x) \\in U ] \\leq \\exp \\left(\\varepsilon d _ {X} \\left(x, x ^ {\\prime}\\right)\\right) \\Pr [ A \\left(x ^ {\\prime}\\right) \\in U ] + \\delta .\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.393, + 0.905, + 0.469 + ], + "angle": 0, + "content": "This definition reduces to standard DP when \\( d_X(x, x') = 1 \\) for all adjacent datasets, but naturally captures similarity in continuous spaces. In our work, we define \\( (X, d_X) \\) where \\( X = \\mathbb{R}^d \\) represents image data and \\( d_X(x, x') = ||x - x'||_2 \\) is the \\( \\ell_2 \\) distance between inputs." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.469, + 0.905, + 0.543 + ], + "angle": 0, + "content": "To guarantee privacy, we calibrate noise based on the sensitivity of a query function \\( f\\colon X\\to \\mathbb{R}^m \\), which measures output changes relative to input changes. In our work we choose the following sensitivity measure as used in [4, 10, 26]:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.549, + 0.905, + 0.58 + ], + "angle": 0, + "content": "Definition 2 (Sensitivity). For a function \\( f \\colon X \\to \\mathbb{R}^m \\), the sensitivity is:" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.585, + 0.813, + 0.619 + ], + "angle": 0, + "content": "\\[\n\\Delta_ {f} = \\max _ {x, x ^ {\\prime} \\in X} \\frac {\\| f (x) - f \\left(x ^ {\\prime}\\right) \\|}{\\left| \\left| x - x ^ {\\prime} \\right| \\right|}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.624, + 0.905, + 0.744 + ], + "angle": 0, + "content": "This definition captures the largest possible output difference relative to input difference, directly controlling how much noise is needed for privacy. This notion of sensitivity offers several advantages, which we discuss in the supplementary material. Importantly, when \\(\\Delta_f\\) is small, less noise is required to achieve the same privacy guarantee, preserving more signal and improving utility—particularly valuable for high-dimensional data like images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.745, + 0.905, + 0.791 + ], + "angle": 0, + "content": "For privacy preservation, we employ the Gaussian mechanism, which adds noise to the output of our query function proportional to the sensitivity \\(\\Delta_f\\):" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.796, + 0.905, + 0.841 + ], + "angle": 0, + "content": "Theorem 1 (Gaussian Mechanism). [6, 9] For a function \\( f: \\mathcal{D} \\to \\mathbb{R}^d \\) with sensitivity \\( \\Delta_f \\), adding Gaussian noise with scale \\( \\sigma \\) satisfying:" + }, + { + "type": "equation", + "bbox": [ + 0.624, + 0.847, + 0.905, + 0.88 + ], + "angle": 0, + "content": "\\[\n\\sigma \\geq \\frac {\\Delta_ {f} \\sqrt {2 \\log (1 . 2 5 / \\delta)}}{\\varepsilon} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.885, + 0.848, + 0.902 + ], + "angle": 0, + "content": "ensures \\( f(x) + \\mathcal{N}(0, \\sigma^2 I) \\) is \\( (\\varepsilon, \\delta) \\)-metric private." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4573" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.152 + ], + "angle": 0, + "content": "This mechanism forms the foundation of our privacy-preserving feature sharing approach, allowing us to bound and quantify the privacy guarantees of our federated learning framework." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.161, + 0.415, + 0.177 + ], + "angle": 0, + "content": "2.3. Neural Network Attribution Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.183, + 0.484, + 0.245 + ], + "angle": 0, + "content": "We consider several gradient-based attribution techniques in our work. The simplest being Gradient \\(\\times\\) Input [25], which computes attributions through element-wise multiplication of input and gradient:" + }, + { + "type": "equation", + "bbox": [ + 0.204, + 0.253, + 0.483, + 0.283 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} _ {\\operatorname {g r a d}} \\left(f _ {\\theta}, \\mathbf {x}\\right) = \\mathbf {x} \\odot \\frac {\\partial f _ {\\theta}}{\\partial \\mathbf {x}} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.291, + 0.484, + 0.336 + ], + "angle": 0, + "content": "Integrated Gradients (IG) [28] addresses gradient saturation by accumulating gradients along a path from baseline \\(\\mathbf{x}'\\) to input:" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.346, + 0.483, + 0.381 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} _ {\\mathrm {I G}} \\left(f _ {\\theta}, \\mathbf {x}\\right) = \\left(\\mathbf {x} - \\mathbf {x} ^ {\\prime}\\right) \\odot \\int_ {0} ^ {1} \\frac {\\partial f _ {\\theta} \\left(\\mathbf {x} ^ {\\prime} + \\alpha \\left(\\mathbf {x} - \\mathbf {x} ^ {\\prime}\\right)\\right)}{\\partial \\mathbf {x}} d \\alpha \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.39, + 0.484, + 0.421 + ], + "angle": 0, + "content": "SmoothGrad [27] reduces attribution noise by averaging gradients over perturbed inputs:" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.43, + 0.483, + 0.471 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} _ {\\text {s m o o t h}} (f _ {\\theta}, \\mathbf {x}) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\frac {\\partial f _ {\\theta}}{\\partial (\\mathbf {x} + \\epsilon)} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.48, + 0.484, + 0.526 + ], + "angle": 0, + "content": "Lastly we use Layer-wise Relevance Propagation (LRP) with the \\(\\epsilon\\)-rule according to [3]. This recursively propagates relevance scores \\(R^{(l)}\\) from layer \\(l\\) to \\(l - 1\\) using:" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.535, + 0.483, + 0.567 + ], + "angle": 0, + "content": "\\[\nR _ {j} = \\sum \\frac {a _ {j} w _ {j k}}{\\epsilon + \\sum_ {0 , j} a _ {j} w _ {j k}} R _ {k} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.576, + 0.484, + 0.607 + ], + "angle": 0, + "content": "where \\(w_{jk}\\) are layer weights, \\(a_{j}\\) are layer activations, and \\(\\epsilon\\) is a small stabilizing term." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.619, + 0.234, + 0.635 + ], + "angle": 0, + "content": "3. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Federated Learning was introduced with FedAvg [19], enabling collaborative model training without sharing raw data, but suffers under statistical heterogeneity. Regularization-based methods mitigate client drift by constraining local updates, such as FedProx [16], SCAFFOLD [12], and FedDyn [1], with extensions like FedBN [17], MOON [15], FedNova [29], FedSAM [21], and FedDISCO [32] targeting specific aspects like feature shifts or loss geometry. A more direct solution is data or knowledge sharing: FedDF and FedAux [18, 24] distill client knowledge using a public dataset; FedGen [36] synthesizes class-conditional features for distribution alignment; FedFTG [33] generates pseudo-data for feature-level transfer; and FedFed [30] applies VAE-based feature distillation. While effective, these methods often introduce significant computational and privacy costs due to data synthesis and sharing. In contrast, FedXDS shares compact" + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.088, + 0.891, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.317, + 0.907, + 0.388 + ], + "angle": 0, + "content": "Figure 1. Illustration of the DP attribution-guided feature extraction. The process consists of four main steps: (1) Computing attribution scores using a warmup model to identify important features, (2) Creating and applying a binary mask based on the attributions, (3) Adding Gaussian noise for metric privacy." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.403, + 0.906, + 0.431 + ], + "angle": 0, + "content": "Algorithm 1 Attribution-Guided Private Representation Extraction." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.435, + 0.906, + 0.465 + ], + "angle": 0, + "content": "Require: \\(D_{k}\\) : client dataset, \\(f_{\\theta}\\) : model, \\(s\\) : sparsity level, \\((\\varepsilon ,\\delta)\\) : privacy params, \\(S\\) : sensitivity, \\(\\theta_{\\mathrm{warmup}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.465, + 0.704, + 0.48 + ], + "angle": 0, + "content": "Ensure: Private dataset \\(D_k^p\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.48, + 0.652, + 0.495 + ], + "angle": 0, + "content": "1: for \\(\\mathbf{x}_i^k\\in D_k\\) do" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.495, + 0.863, + 0.511 + ], + "angle": 0, + "content": "2: \\(\\mathbf{h} \\gets \\mathcal{A}(f_{\\theta_{\\mathrm{warmup}}}, \\mathbf{x}_i^k)\\); \\(h^{(s)} \\gets \\mathrm{s}\\)-th largest in \\(\\mathbf{h}\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.512, + 0.804, + 0.528 + ], + "angle": 0, + "content": "3: \\(\\mathbf{m} \\gets [\\mathbf{h}[i] \\geq h^{(s)}]; \\mathbf{x}_m \\gets \\mathbf{x}_i^k \\odot \\mathbf{m}\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.528, + 0.904, + 0.543 + ], + "angle": 0, + "content": "4: \\(\\tilde{\\mathbf{x}}_i^k\\gets \\mathbf{x}_m + \\mathcal{N}(0,\\sigma^2)\\) \\(\\triangleright \\sigma\\) from \\((\\varepsilon ,\\delta)\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.543, + 0.7, + 0.559 + ], + "angle": 0, + "content": "5: Add \\((\\tilde{\\mathbf{x}}_i^k,y_i^k)\\) to \\(D_k^p\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.559, + 0.596, + 0.571 + ], + "angle": 0, + "content": "6: end for" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.573, + 0.617, + 0.59 + ], + "angle": 0, + "content": "7: return \\(D_k^p\\)" + }, + { + "type": "list", + "bbox": [ + 0.526, + 0.48, + 0.904, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.619, + 0.906, + 0.68 + ], + "angle": 0, + "content": "attribution-based feature subsets derived via XAI, avoiding generators while enhancing privacy, efficiency, and performance in heterogeneous settings. Wo provide an extensive related works in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.694, + 0.697, + 0.711 + ], + "angle": 0, + "content": "4. FedXDS Approach" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.719, + 0.906, + 0.793 + ], + "angle": 0, + "content": "Consider a federated learning system with \\(N\\) clients, where each client \\(k\\) holds local data \\(\\mathcal{D}_k\\). In our setting all clients share the same neural network architecture \\(f_{\\theta}\\). Our goal is to learn a global model \\(\\Theta_G\\) which generalizes well on all client datasets." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.804, + 0.838, + 0.818 + ], + "angle": 0, + "content": "4.1. Attribution-Guided Feature Selection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.901 + ], + "angle": 0, + "content": "The first stage of our method identifies discriminative features using a pre-trained warmup model, with parameters \\(\\theta_{\\mathrm{warmup}}\\). This is achieved by training with FedAvg for \\(R_{\\mathrm{warmup}}\\) rounds, before sharing representations. Then, for each input sample \\(\\mathbf{x} \\in \\mathbb{R}^{H \\times W \\times 3}\\), where \\(H\\) and \\(W\\) are the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4574" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.093, + 0.092, + 0.437, + 0.107 + ], + "angle": 0, + "content": "Algorithm 2 Privacy-Preserving Federated Training" + }, + { + "type": "algorithm", + "bbox": [ + 0.093, + 0.111, + 0.487, + 0.411 + ], + "angle": 0, + "content": "Require: \\(\\{D_k\\}_{k=1}^K\\) : client datasets, \\(\\lambda\\): knowledge weight, \\(T\\): rounds, \\(E\\): local epochs, \\(\\eta\\): learning rate, \\(R_{\\text{warmup}}\\) \n1: Initialize \\(\\theta_{\\text{warmup}}\\) using \\(R_{\\text{warmup}}\\) rounds of FedAvg \n2: Obtain \\(D_k^p\\) using Algorithm 1 for all \\(k \\in [K]\\) in parallel \n3: Server: \\(\\mathcal{D}_g \\gets \\bigcup_{k=1}^K D_k^p\\) \n4: for round \\(t = 1\\) to \\(T\\) do \n5: Select client subset \\(S_t \\subseteq [K]\\) \n6: for \\(k \\in S_t\\) in parallel do \n7: \\(\\theta_k^t \\gets \\theta^{t-1}\\) Download global model \n8: for epoch \\(e = 1\\) to \\(E\\) do \n9: for batch \\(B\\) from \\(D_k\\), \\(B_g\\) from \\(\\mathcal{D}_g\\) do \n10: \\(\\mathcal{L}_k^t \\gets \\frac{1}{|B_k|} \\sum_{(\\mathbf{x}, y) \\in B_k} \\ell(f_{\\theta_k^t}(\\mathbf{x}), y) + \\lambda \\frac{1}{|B_g|} \\sum_{(\\mathbf{x}_g, y_g) \\in B_g} \\ell(f_{\\theta_k^t}(\\mathbf{x}_g), y_g)\\) \n11: \\(\\theta_k^t \\gets \\theta_k^t - \\eta \\nabla \\mathcal{L}\\) \n12: end for \n13: end for \n14: end for \n15: Server: \\(\\theta^t \\gets \\frac{1}{|S_t|} \\sum_{k \\in S_t} \\theta_k^t\\) >FedAvg \n16: end for" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.436, + 0.484, + 0.482 + ], + "angle": 0, + "content": "dimensions of the image with three color channels, in the client dataset \\(\\mathcal{D}_k\\), we compute relevances using an attribution method \\(\\mathcal{A}\\) from Section 2.3:" + }, + { + "type": "equation", + "bbox": [ + 0.239, + 0.492, + 0.482, + 0.51 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} = \\mathcal {A} \\left(f _ {\\theta}, \\mathbf {x}\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.518, + 0.484, + 0.625 + ], + "angle": 0, + "content": "where \\(\\mathbf{h} \\in \\mathbb{R}^{H \\times W}\\) is the pixel-wise importance score of an input image \\(\\mathbf{x}\\). These attribution scores quantify the importance of each input feature to the model's prediction. To focus on the most relevant features, we create a binary mask \\(\\mathbf{m}\\), as shown in Equation 8, by retaining features whose attribution scores exceed a threshold determined by the desired sparsity level \\(s\\):" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.635, + 0.483, + 0.675 + ], + "angle": 0, + "content": "\\[\n\\mathbf {m} [ i ] = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\mathbf {h} [ i ] \\geq h ^ {(s)} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.683, + 0.483, + 0.728 + ], + "angle": 0, + "content": "where \\( h^{(s)} \\) represents the \\( s \\)-th largest value in the attribution scores. This mask identifies the subset of features that are most crucial for the model's decision-making process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.729, + 0.484, + 0.774 + ], + "angle": 0, + "content": "We define our feature selection function \\( f_{A} \\colon \\mathbb{R}^{H \\times W \\times 3} \\to \\mathbb{R}^{H \\times W \\times 3} \\) as the application of the attribution mask \\( \\mathbf{m} \\) to each image \\( \\mathbf{x} \\in \\mathcal{D}_k \\) through:" + }, + { + "type": "equation", + "bbox": [ + 0.23, + 0.785, + 0.483, + 0.801 + ], + "angle": 0, + "content": "\\[\nf _ {A} (\\mathbf {x}) = \\mathbf {x} \\odot \\mathbf {m} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "This operation emphasizes the most informative parts of the input while suppressing less relevant details. However, directly sharing \\( f_{A}(\\mathbf{x}) \\) would still pose privacy risks. Therefore, we design a privacy mechanism \\( \\mathcal{M} \\) that adds calibrated noise to these masked features (Section refsubsec:privacy):" + }, + { + "type": "equation", + "bbox": [ + 0.611, + 0.105, + 0.905, + 0.123 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} (\\mathbf {x}) = f _ {A} (\\mathbf {x}) + \\mathcal {N} (0, \\sigma^ {2} \\mathbf {I}) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.126, + 0.907, + 0.217 + ], + "angle": 0, + "content": "where \\(\\sigma\\) is determined based on privacy requirements. The resulting privacy-protected features, along with their corresponding labels, are then aggregated into a new dataset \\(\\mathcal{D}_k^p = \\{(\\mathcal{M}(\\mathbf{x}),y)|(x,y)\\in \\mathcal{D}_k\\}\\) and shared with the server. The full data generated process is outlined in Algorithm 1 and visualized in Figure 1." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.224, + 0.81, + 0.24 + ], + "angle": 0, + "content": "4.2. Local Training using Shared Data" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.245, + 0.907, + 0.336 + ], + "angle": 0, + "content": "After the privacy-protected features have been obtained, the server aggregates private representations from all clients into a global dataset \\(\\mathcal{D}_g = \\bigcup_{k=1}^N \\mathcal{D}_k^p\\), and sends this dataset to each client. The clients then optimize a composite objective that balances local task performance with knowledge from the global private dataset:" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.341, + 0.905, + 0.386 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {\\theta} \\left[ \\mathbb {E} _ {(\\mathbf {x}, y) \\sim \\mathcal {D} _ {k}} [ \\ell (f _ {\\theta} (\\mathbf {x}), y) ] \\right. \\tag {11} \\\\ + \\lambda \\mathbb {E} _ {(\\mathbf {x}, y) \\sim \\mathcal {D} _ {g}} [ \\ell (f _ {\\theta} (\\mathbf {x}), y) ] ] \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.389, + 0.906, + 0.465 + ], + "angle": 0, + "content": "The hyperparameter \\(\\lambda\\) controls the trade-off between local specialization and global knowledge integration. This formulation allows clients to benefit from collective learning while maintaining their own task-specific performance and ensuring privacy. The procedure is outlined in Algorithm 2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.472, + 0.695, + 0.487 + ], + "angle": 0, + "content": "4.3. Privacy Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.906, + 0.614 + ], + "angle": 0, + "content": "We build on the insight that sharing features derived from raw client data can mitigate data heterogeneity in federated learning [34]. Nevertheless, directly sharing these features poses significant privacy risks. A straightforward approach would be to apply differential privacy (DP) to the raw data itself; however, the high dimensionality of image data forces the addition of very large noise levels, thereby significantly degrading utility." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.615, + 0.906, + 0.704 + ], + "angle": 0, + "content": "To address this challenge, we propose a two-step approach. First, we use an attribution-guided masking strategy (Section 4.1) to identify and retain only task-relevant pixels. Second, we add Gaussian noise (Theorem 1) to the resulting masked features to achieve the desired \\(\\varepsilon\\)-metric privacy (Definition 1, Definition 2)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.906, + 0.855 + ], + "angle": 0, + "content": "This design offers two main advantages. First, by sparsifying input features that do not contribute to classification predictions, we reduce the overall sensitivity of the data. According to Theorem 1, lower sensitivity requires less noise to meet the same privacy budget \\(\\varepsilon\\), thus improving utility. Second, in contrast to random or magnitude-based sparsification that may unintentionally discard valuable information, our attribution-guided approach specifically preserves the pixels most critical for the learning task while eliminating irrelevant ones." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "The primary goal of our privacy mechanism is to protect against membership inference attacks (MIA) and feature inversion, which we empirically validate in section 6." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4575" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.138 + ], + "angle": 0, + "content": "In the following section, we formally demonstrate how our strategy guarantees metric privacy, while a comprehensive discussion of privacy assumptions appears in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.144, + 0.483, + 0.16 + ], + "angle": 0, + "content": "4.3.1. Sensitivity Analysis of Attribution-Based Masking" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.163, + 0.483, + 0.253 + ], + "angle": 0, + "content": "A critical component of ensuring privacy is understanding the sensitivity of our feature selection function \\( f(\\mathbf{x}) \\) defined in Equation 9. From Theorem 1, the noise required to attain a certain privacy level \\( \\varepsilon \\) depends on the sensitivity \\( \\Delta_f \\). Our goal is to minimize this sensitivity before applying the privacy mechanism \\( \\mathcal{M} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.254, + 0.483, + 0.314 + ], + "angle": 0, + "content": "For our feature selection function \\(f_{A}(\\mathbf{x}) = \\mathbf{x}\\odot \\mathbf{m}\\), we can derive a strict bound on sensitivity. Since each coordinate of the mask satisfies \\(\\mathbf{m}[i]\\in \\{0,1\\}\\), for any two inputs \\(\\mathbf{x},\\mathbf{x}'\\in \\mathbb{R}^{H\\times W\\times 3}\\), we have, in the worst case:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.324, + 0.414, + 0.344 + ], + "angle": 0, + "content": "\\[\n\\left| \\mathbf {m} [ i ] \\mathbf {x} [ i ] - \\mathbf {m} [ i ] \\mathbf {x} ^ {\\prime} [ i ] \\right| \\leq \\left| \\mathbf {x} [ i ] - \\mathbf {x} ^ {\\prime} [ i ] \\right|\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.353, + 0.242, + 0.368 + ], + "angle": 0, + "content": "Thus, in aggregate:" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.379, + 0.4, + 0.396 + ], + "angle": 0, + "content": "\\[\n\\left\\| f _ {A} (\\mathbf {x}) - f _ {A} \\left(\\mathbf {x} ^ {\\prime}\\right) \\right\\| _ {2} \\leq \\left\\| \\mathbf {x} - \\mathbf {x} ^ {\\prime} \\right\\| _ {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.407, + 0.483, + 0.438 + ], + "angle": 0, + "content": "This shows that our masking operation is non-expansive with respect to the \\(\\ell_2\\) norm. Rearranging the equation:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.446, + 0.483, + 0.48 + ], + "angle": 0, + "content": "\\[\n\\Delta_ {f} = \\frac {\\left\\| f _ {A} (\\mathbf {x}) - f _ {A} \\left(\\mathbf {x} ^ {\\prime}\\right) \\right\\| _ {2}}{\\left\\| \\mathbf {x} - \\mathbf {x} ^ {\\prime} \\right\\| _ {2}} \\leq 1 \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.489, + 0.483, + 0.655 + ], + "angle": 0, + "content": "where \\(\\Delta_f\\) is the sensitivity we defined in the preliminaries in Definition 2. Without our attribution-based masking, the sensitivity would be unbounded for arbitrary transformations. Moreover, in the worst case \\(f_A\\) is the identity, i.e. \\(f_A(x) = x\\), meaning that the attribution method deems all pixels important. This implies that our method would reduce to the naive approach of adding noise to the raw features. In practice, this rarely occurs, since this would already require a highly sparsified image. Nevertheless, our approach safeguards against this worst-case scenario. More discussion can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.663, + 0.391, + 0.677 + ], + "angle": 0, + "content": "4.3.2. Privacy Mechanism and Guarantees" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.681, + 0.483, + 0.727 + ], + "angle": 0, + "content": "Based on the sensitivity bound established above, we design our privacy mechanism \\(\\mathcal{M}\\) (Equation 10) by adding Gaussian noise calibrated to this reduced sensitivity:" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.737, + 0.386, + 0.755 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} (\\mathbf {x}) = f _ {A} (\\mathbf {x}) + \\mathcal {N} (0, \\sigma^ {2} \\mathbf {I})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.826 + ], + "angle": 0, + "content": "where \\(\\sigma\\) is chosen according to Theorem 1 for an \\((\\varepsilon, \\delta)\\)-level of privacy given sensitivity \\(\\Delta_f \\leq 1\\). This ensures that for all input images in the client dataset, \\(\\mathbf{x}, \\mathbf{x}' \\in D_k\\), our method satisfies \\((\\varepsilon, \\delta)\\)-metric privacy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Our theoretical privacy guarantees are empirically validated through membership inference and feature inversion attacks Section 6, where attribution-masked features consistently show stronger protection than unmasked features under equivalent privacy budgets." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.654, + 0.107 + ], + "angle": 0, + "content": "5. Experimental" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.116, + 0.704, + 0.132 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.138, + 0.907, + 0.395 + ], + "angle": 0, + "content": "We evaluate FedXDS on standard image classification benchmarks (CIFAR-10, CIFAR-100, Tiny-ImageNet) and real-world federated datasets from the LEAF framework [5] (CelebA, FEMNIST) that naturally exhibit heterogeneous distributions. For the standard benchmarks, we simulate statistical heterogeneity using Dirichlet-based partitioning with concentration parameters \\(\\alpha \\in \\{0.05, 0.1\\}\\), where smaller values indicate greater heterogeneity. We conduct experiments with both 10 and 100 clients, setting participation rates to 0.5 and 0.1, respectively, and train for \\(R = 200\\) communication rounds. To compute relevance-based attributions, we evaluate four attribution methods in our FL framework: Layerwise Relevance Propagation (FedXLRP) [3], Integrated Gradients (FedXIG) [28], SmoothGrad (FedXSG) [27], and Input \\(\\times\\) Gradient (FedXGrad). Unless otherwise state we use \\(\\varepsilon = 20\\) sparsity level \\(s = 70\\%\\), and \\(\\lambda = 0.5\\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.406, + 0.622, + 0.421 + ], + "angle": 0, + "content": "5.2. Baselines" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.429, + 0.907, + 0.747 + ], + "angle": 0, + "content": "The experimental results in Table 1 show that FedXLRP consistently outperforms all baselines across varying client numbers and data distributions. This advantage amplifies in more challenging scenarios (K=100, \\(\\alpha = 0.05\\)). FedFed ranks second, while FedAvg and FedAux degrade significantly with increased client numbers. On CIFAR-100, performance gaps narrow, though FedXLRP maintains its edge, particularly with higher client counts and heterogeneity. Other attribution methods (FedXIG, FedXGrad, FedXSG) perform well but remain below FedXLRP, highlighting attribution method choice importance. We discuss this point in Subsection 7.1. In addition, communication efficiency results Table 2 reinforce FedXLRP's advantages. To reach \\(70\\%\\) accuracy on CIFAR-10 with 10 clients, FedXLRP requires only 14 rounds versus 49 for FedAvg and 28 for FedFTG. With 100 clients, FedXLRP needs 23 rounds to reach \\(60\\%\\) accuracy compared to FedAvg's 79. CIFAR-100 shows similar patterns, with FedXLRP achieving target accuracies in substantially fewer rounds across client configurations. Among baselines, only FedFed shows comparable efficiency." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.757, + 0.835, + 0.773 + ], + "angle": 0, + "content": "5.3. Experiments on Real-World Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.902 + ], + "angle": 0, + "content": "We also evaluate FedXDS on real-world datasets using implementations from the LEAF [5] and [7] frameworks. Specifically, we consider CelebA and FEMNIST, which exhibit a distinct form of non-IIIDness known as distribution shift, introducing an additional challenge for federated learning. For CelebA, we follow the implementation of [36], while for FEMNIST, we use the setup from [7], conducting experiments with 10 and 100 clients, respectively." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4576" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.133 + ], + "angle": 0, + "content": "Table 1. Performance comparison across different federated learning methods, including standard deviations from 5 runs. Results show the top-1 test accuracy (%) on CIFAR-10, CIFAR-100, and Tiny-ImageNet under varying numbers of clients (K) and Dirichlet concentration parameters \\((\\alpha)\\). Bold values indicate the best performance in each column." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.142, + 0.907, + 0.358 + ], + "angle": 0, + "content": "
DatasetCIFAR-10CIFAR-100Tiny-ImageNet
K=10K=100K=10K=100K=10K=100
α = 0.05α = 0.1α = 0.05α = 0.1α = 0.05α = 0.1α = 0.05α = 0.1α = 0.05α = 0.1α = 0.05α = 0.1
FedAvg64.16 ± 0.7172.90 ± 0.5555.53 ± 0.8260.94 ± 0.6843.82 ± 0.4544.57 ± 0.5132.99 ± 0.7335.59 ± 0.6230.30 ± 0.5934.41 ± 0.4826.20 ± 0.8128.79 ± 0.75
FedProx71.56 ± 0.6574.31 ± 0.4951.89 ± 0.9165.01 ± 0.5344.33 ± 0.3848.67 ± 0.4737.79 ± 0.5538.43 ± 0.4933.27 ± 0.4234.54 ± 0.5124.15 ± 0.8830.23 ± 0.67
FedDyn70.62 ± 0.5878.25 ± 0.3463.09 ± 0.6169.41 ± 0.4548.53 ± 0.4149.42 ± 0.3937.14 ± 0.6841.60 ± 0.5433.03 ± 0.4835.69 ± 0.4129.52 ± 0.6332.37 ± 0.59
SCAFFOLD71.35 ± 0.6175.62 ± 0.4258.51 ± 0.7963.82 ± 0.6246.61 ± 0.3347.27 ± 0.4435.28 ± 0.2438.87 ± 0.5831.53 ± 0.5335.76 ± 0.3928.35 ± 0.5930.49 ± 0.61
FedSAM70.15 ± 0.8877.53 ± 0.5162.45 ± 0.7268.78 ± 0.5947.95 ± 0.3548.81 ± 0.4136.88 ± 0.5140.92 ± 0.4732.21 ± 0.5535.03 ± 0.4428.98 ± 0.6631.84 ± 0.52
FedDISCO69.73 ± 0.4476.98 ± 0.6361.82 ± 0.6868.11 ± 0.6547.31 ± 0.5248.15 ± 0.5836.17 ± 0.6740.25 ± 0.6131.78 ± 0.6134.68 ± 0.5228.51 ± 0.7131.33 ± 0.64
FedFed79.12 ± 0.4882.58 ± 0.3175.35 ± 0.5778.87 ± 0.4352.27 ± 0.4956.45 ± 0.3845.41 ± 0.7649.28 ± 0.5234.99 ± 0.4536.49 ± 0.3733.28 ± 0.5834.89 ± 0.47
FedFTG75.21 ± 0.5378.44 ± 0.3970.84 ± 0.3174.85 ± 0.4144.91 ± 0.6254.15 ± 0.4540.43 ± 0.3946.47 ± 0.4832.40 ± 0.5733.79 ± 0.4930.53 ± 0.5132.26 ± 0.55
FedGen68.34 ± 0.6874.65 ± 0.4758.90 ± 0.7563.20 ± 0.6644.29 ± 0.5550.12 ± 0.5136.70 ± 0.6239.82 ± 0.5734.17 ± 0.4436.33 ± 0.4229.45 ± 0.6831.60 ± 0.61
FedAux59.72 ± 0.8573.70 ± 0.5153.13 ± 0.8865.11 ± 0.5944.05 ± 0.4844.85 ± 0.5336.54 ± 0.6538.39 ± 0.6127.83 ± 0.7234.31 ± 0.4924.75 ± 0.8230.38 ± 0.69
FedDF60.56 ± 0.7772.40 ± 0.5859.20 ± 0.7263.47 ± 0.6430.13 ± 0.8145.47 ± 0.5034.46 ± 0.6936.77 ± 0.6429.07 ± 0.6534.75 ± 0.4628.42 ± 0.6230.51 ± 0.63
FedXLRP81.72 ± 0.3583.46 ± 0.2877.02 ± 0.6380.27 ± 0.3952.07 ± 0.4558.09 ± 0.3346.25 ± 0.8252.63 ± 0.4136.85 ± 0.3938.64 ± 0.3234.64 ± 0.5233.18 ± 0.51
FedXIG72.22 ± 0.5975.89 ± 0.4464.02 ± 0.6968.61 ± 0.5549.21 ± 0.4855.46 ± 0.4142.34 ± 0.5847.89 ± 0.5033.59 ± 0.5135.30 ± 0.4529.77 ± 0.6431.93 ± 0.58
FedXGrad71.38 ± 0.6273.99 ± 0.5264.52 ± 0.6566.49 ± 0.6150.13 ± 0.4354.21 ± 0.4642.11 ± 0.6046.98 ± 0.5333.18 ± 0.5434.40 ± 0.4830.00 ± 0.6130.93 ± 0.60
FedXSG71.89 ± 0.6074.63 ± 0.4963.75 ± 0.7167.80 ± 0.5851.57 ± 0.3955.22 ± 0.4240.91 ± 0.6347.61 ± 0.5133.43 ± 0.5234.74 ± 0.4729.64 ± 0.6531.53 ± 0.59
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.371, + 0.907, + 0.429 + ], + "angle": 0, + "content": "Table 2. Communication efficiency comparison across different federated learning methods. Results show the number of communication rounds (mean ± std over 5 runs) needed to achieve target accuracy thresholds (70% and 60% for CIFAR-10; 40% and 30% for CIFAR-100; 35% and 30% for Tiny-ImageNet) with different numbers of clients (K=10 and K=100) and heterogeneity parameter \\(\\alpha = 0.1\\). Lower values indicate better communication efficiency. Best results are in bold." + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.438, + 0.75, + 0.726 + ], + "angle": 0, + "content": "
MethodCIFAR-10CIFAR-100Tiny-ImageNet
acc=70% K=10acc=60% K=100acc=40% K=10acc=30% K=100acc=35% K=10acc=30% K=100
FedAvg49 ± 579 ± 724 ± 434 ± 560 ± 690 ± 8
FedProx32 ± 355 ± 428 ± 432 ± 445 ± 465 ± 5
FedDyn29 ± 351 ± 424 ± 332 ± 348 ± 560 ± 4
SCAFFOLD33 ± 452 ± 525 ± 338 ± 446 ± 469 ± 6
FedSAM34 ± 458 ± 527 ± 435 ± 450 ± 564 ± 5
FedDISCO36 ± 361 ± 628 ± 437 ± 552 ± 567 ± 6
FedFed15 ± 222 ± 413 ± 412 ± 312 ± 215 ± 2
FedFTG28 ± 357 ± 422 ± 328 ± 330 ± 340 ± 4
FedGen51 ± 566 ± 621 ± 439 ± 555 ± 570 ± 6
FedAux38 ± 445 ± 419 ± 336 ± 442 ± 450 ± 5
FedDF84 ± 790 ± 828 ± 445 ± 5100 ± 8110 ± 9
FedXLRP14 ± 223 ± 311 ± 313 ± 210 ± 412 ± 2
FedXIG35 ± 439 ± 416 ± 221 ± 338 ± 445 ± 4
FedXGrad41 ± 446 ± 513 ± 226 ± 344 ± 450 ± 5
FedXSG41 ± 551 ± 421 ± 333 ± 442 ± 448 ± 5
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.751, + 0.482, + 0.888 + ], + "angle": 0, + "content": "Additional implementation details can be found in the supplementary material. The results in Table 3 show a clear trend where FedXDS consistently achieves the highest accuracy across both datasets, demonstrating its effectiveness in handling distribution shift. Notably, methods incorporating additional data generation or augmentation, such as FedFTG and FedGen, also perform well, but FedXDS surpasses them, suggesting that relevance-guided data sharing provides a stronger mechanism for improving generaliza" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.751, + 0.782, + 0.767 + ], + "angle": 0, + "content": "tion in heterogeneous federated settings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.907, + 0.903 + ], + "angle": 0, + "content": "The results in Table 4 demonstrate that integrating FedXDS with FL methods consistently improves the performance. In all cases, the addition of FedXDS yields substantial accuracy gains, particularly in the highly heterogeneous setting with \\( \\mathrm{K} = 100 \\) clients. For instance, FedAvg, which traditionally suffers from performance degradation in non-IID scenarios, improves from \\( 60.94\\% \\) to \\( 80.29\\% \\) when combined with FedXDS. Similar improvements are observed for" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4577" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.131 + ], + "angle": 0, + "content": "Table 3. Accuracy comparison (mean ± std % over 5 runs) on CelebA and FEMNIST. Best results are in bold. We use the LRP variant of FedXDS." + }, + { + "type": "table", + "bbox": [ + 0.14, + 0.143, + 0.434, + 0.346 + ], + "angle": 0, + "content": "
MethodCelebAFEMNIST
FedAvg87.23 ± 0.7184.71 ± 0.68
FedProx87.79 ± 0.6585.66 ± 0.61
SCAFFOLD86.36 ± 0.5684.24 ± 0.59
FedDyn88.13 ± 0.5286.45 ± 0.34
FedSAM90.03 ± 0.1587.56 ± 0.44
FedDISCO89.78 ± 0.6187.29 ± 0.62
FedFTG90.31 ± 0.4387.88 ± 0.67
FedGen89.65 ± 0.5986.37 ± 0.55
FedDF88.56 ± 0.6385.28 ± 0.66
FedAux89.27 ± 0.6085.94 ± 0.62
FedFed90.76 ± 0.8388.28 ± 0.71
FedXDS91.55 ± 0.4889.03 ± 0.35
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.359, + 0.485, + 0.416 + ], + "angle": 0, + "content": "Table 4. Performance comparison (mean ± std % over 5 runs) on CIFAR-10 for different base FL methods and their FedXDS (using LRP) variants under two heterogeneity settings (\\( K = 10 \\) and \\( K = 100 \\)). The FedXDS variants consistently improve performance." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.427, + 0.47, + 0.589 + ], + "angle": 0, + "content": "
MethodK=10K=100
FedAvg72.90 ± 0.5560.94 ± 0.68
FedAvg + FedXDS83.46 ± 0.3180.29 ± 0.42
FedProx74.31 ± 0.4965.01 ± 0.53
FedProx + FedXDS84.05 ± 0.3580.73 ± 0.45
FedDyn70.62 ± 0.5869.41 ± 0.45
FedDyn + FedXDS84.20 ± 0.3379.56 ± 0.48
SCAFFOLD71.35 ± 0.6163.82 ± 0.62
SCAFFOLD + FedXDS83.89 ± 0.3879.72 ± 0.44
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.483, + 0.658 + ], + "angle": 0, + "content": "FedProx, FedDyn, and SCAFFOLD, with FedXDS enhancing their robustness to data heterogeneity while maintaining their inherent advantages." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.658, + 0.483, + 0.794 + ], + "angle": 0, + "content": "The magnitude of improvement suggests that FedXDS effectively mitigates the adverse effects of statistical heterogeneity by facilitating better knowledge transfer across clients. Notably, FedDyn, which incorporates regularization-based local adaptation, exhibits the smallest absolute improvement with FedXDS, indicating that its existing adaptation mechanisms already partially address heterogeneity. In contrast, methods like FedAvg, without correction for client drift, benefit more from FedXDS." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.808, + 0.341, + 0.827 + ], + "angle": 0, + "content": "6. Empirical Privacy Analysis" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.36, + 0.851 + ], + "angle": 0, + "content": "6.1. Membership Inference Attack" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.485, + 0.903 + ], + "angle": 0, + "content": "We conduct membership inference attacks following [30] to evaluate our methodology's privacy guarantees. We investigate whether training on masked and noised features reveals" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.907, + 0.132 + ], + "angle": 0, + "content": "Table 5. Test accuracy (mean ± std % over 5 runs) for different sparsification levels \\( s \\) across various attribution methods. Best results are in bold. Results for CIFAR10 at \\( K = 10, \\alpha = 0.1 \\)." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.143, + 0.907, + 0.248 + ], + "angle": 0, + "content": "
sGradientSmoothGradInt. Grad.LRP
6075.99 ± 0.5275.63 ± 0.5577.10 ± 0.4884.74 ± 0.31
6574.19 ± 0.5673.15 ± 0.6176.31 ± 0.5184.01 ± 0.35
7073.99 ± 0.5974.63 ± 0.5475.89 ± 0.4983.46 ± 0.28
7570.54 ± 0.6570.67 ± 0.6371.68 ± 0.5883.01 ± 0.33
8068.31 ± 0.7170.45 ± 0.6670.98 ± 0.6282.67 ± 0.36
8568.80 ± 0.6868.13 ± 0.7369.73 ± 0.6781.88 ± 0.41
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.274, + 0.907, + 0.425 + ], + "angle": 0, + "content": "membership in the underlying dataset by using the globally DP data to train a shadow model, then testing if clients can infer membership by querying this model and training a random forest classifier. Figure 2a shows recall values for different privacy parameters \\(\\varepsilon\\). Results demonstrate that non-masked features, despite DP protection, offer inferior privacy protection compared to masked data. This can be explained by the fact that sparsity significantly mitigates attacks through norm reduction from attribution masks (Equation 8)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.436, + 0.743, + 0.451 + ], + "angle": 0, + "content": "6.2. Feature Inversion Attack" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.459, + 0.909, + 0.687 + ], + "angle": 0, + "content": "While metric private features prevent direct inference of individual data features, a sophisticated adversary might train a denoising autoencoder to learn mappings between protected representations and original images. We conduct feature inversion attacks and report the SSIM score between original and reconstructed images. Figure 2b reveals inversion attack success increases with lower sparsity and higher privacy budgets. Unmasked features consistently exhibit the highest SSIM, confirming that raw features under identical DP constraints provide inferior protection compared to our approach. Additional validations in supplementary material, including visualization of inversion results for non-masked features and FedFed [30] features, confirming both cases' vulnerability to reconstruction, whereas our method provides superior protection." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.702, + 0.808, + 0.718 + ], + "angle": 0, + "content": "7. Ablation Studies and Discussion" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.727, + 0.907, + 0.744 + ], + "angle": 0, + "content": "7.1. XAI Method Comparison Across Sparsification" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Table 5 demonstrates that FedXDS maintains stable performance across different sparsification levels, with all attribution methods achieving above \\(68\\%\\) accuracy, even at high sparsity thresholds. Notably, LRP consistently outperforms other methods, preserving accuracies between \\(81.88\\%\\) and \\(84.74\\%\\), while others degrade more sharply under extreme sparsification. LRP's robustness stems from its layer-wise relevance propagation mechanism, which is fundamentally different from gradient-based approaches. Instead of measuring local sensitivity, LRP redistributes the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4578" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.287, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.093, + 0.183, + 0.29, + 0.208 + ], + "angle": 0, + "content": "(a) Recall score of MI attack at different \\(\\varepsilon\\) for various sparsity levels \\(s\\)." + }, + { + "type": "image", + "bbox": [ + 0.299, + 0.091, + 0.492, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.298, + 0.183, + 0.496, + 0.208 + ], + "angle": 0, + "content": "(b) SSIM score of denoising attack at different \\(\\varepsilon\\) and sparsity levels \\(s\\)." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.092, + 0.698, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.183, + 0.7, + 0.208 + ], + "angle": 0, + "content": "(c) Effect of knowledge weight \\(\\lambda\\) on model accuracy across datasets." + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.092, + 0.903, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.708, + 0.183, + 0.905, + 0.208 + ], + "angle": 0, + "content": "(d) Impact of warmup rounds on model accuracy across datasets." + }, + { + "type": "image_caption", + "bbox": [ + 0.217, + 0.22, + 0.778, + 0.235 + ], + "angle": 0, + "content": "Figure 2. Evaluation of privacy-utility trade-offs and hyperparameter effects in our framework." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.261, + 0.485, + 0.534 + ], + "angle": 0, + "content": "model's output prediction backward through the network, layer by layer, conserving relevance at each step. This process aggregates relevance into contiguous, semantically coherent regions that represent structured object parts. In stark contrast, methods like SmoothGrad or Integrated Gradients identify 'hotspot' pixels that are locally influential but often disconnected from each other. Consequently, under high sparsification, LRP retains meaningful visual patterns, whereas other methods are left with a sparse collection of isolated points that lack sufficient context for accurate classification. We conclude that structural coherence in saliency maps is an essential criterion when selecting the attribution method. This finding is corroborated by current literature [2, 13, 23], which confirms LRP's superior explanation quality and reinforces that our work is consistent with established findings on the importance of explanation quality for downstream tasks. More discussion can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.543, + 0.266, + 0.56 + ], + "angle": 0, + "content": "7.2. Privacy vs. Utility" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.566, + 0.484, + 0.779 + ], + "angle": 0, + "content": "We investigate how the privacy budget \\(\\varepsilon\\) impacts model performance across attribution methods. Table 6 shows that increasing \\(\\varepsilon\\) improves accuracy for all methods, demonstrating the expected privacy-utility trade-off: stronger privacy (lower \\(\\varepsilon\\)) leads to reduced performance due to greater noise. Among all methods, LRP consistently achieves the highest accuracy across privacy levels outperforming the next best (Integrated Gradients) by \\(7 - 8\\%\\) and maintaining robustness even under strict privacy constraints. SmoothGrad shows modest improvements over basic gradients, though less pronounced than Integrated Gradients. These results underscore LRP's ability to extract stable, task-relevant features leading to better generalization with less privacy leakage, aligning with prior work on explanation fidelity [3, 22]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.788, + 0.445, + 0.804 + ], + "angle": 0, + "content": "7.3. Ablation Study on the Knowledge Weight" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We evaluate how the knowledge weight \\(\\lambda\\), as defined in Equation 11, affects performance across CIFAR-10, CIFAR-100, and Tiny-ImageNet. Figure 2c shows that intermediate values (\\(\\lambda = 0.5\\)) consistently outperform both extremes, with CIFAR-10 accuracy peaking at \\(83.46\\%\\) compared to \\(72.90\\%\\) (\\(\\lambda = 0.0\\), equivalent to FedAvg) and" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.259, + 0.905, + 0.3 + ], + "angle": 0, + "content": "Table 6. Ablation study showing test accuracy) for different privacy budgets \\((\\varepsilon)\\) across XAI-based federated learning methods. Results for CIFAR10 at \\(K = 10, \\alpha = 0.1\\)" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.312, + 0.906, + 0.404 + ], + "angle": 0, + "content": "
εGradientSmoothGradInt. Grad.LRP
570.55 ± 0.6271.23 ± 0.5972.45 ± 0.5280.06 ± 0.33
771.45 ± 0.5872.15 ± 0.5573.35 ± 0.4780.89 ± 0.31
1072.25 ± 0.5172.95 ± 0.4874.15 ± 0.4181.65 ± 0.29
1573.12 ± 0.4573.85 ± 0.4375.02 ± 0.3882.11 ± 0.25
2073.99 ± 0.4274.63 ± 0.3975.89 ± 0.3583.46 ± 0.28
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.906, + 0.476 + ], + "angle": 0, + "content": "61.18% (\\(\\lambda = 1.0\\), synthetic data equal weight). Our ablation thus shows that balancing the contribution of synthetic data is critical for performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.488, + 0.733, + 0.505 + ], + "angle": 0, + "content": "7.4. Warmup and Efficiency" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.511, + 0.907, + 0.663 + ], + "angle": 0, + "content": "Figure 2d shows accuracy slightly improves with more warmup rounds, plateauing at 10, suggesting this is the optimal trade-off between convergence and communication cost. Our approach is also computationally efficient. Attribution masks are computed in a single backward pass during warmup, incurring minimal overhead. In contrast, methods like FedGen and FedFTG retrain generators every round, and FedFed requires costly VAE training—making FedXDS more suitable for resource-constrained settings. We provide more discussion in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.679, + 0.634, + 0.695 + ], + "angle": 0, + "content": "8. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this work, we introduced FedXDS, leveraging attribution methods to address statistical heterogeneity, privacy, and computational efficiency in federated learning. By using attribution maps to extract and share privacy-preserved task-relevant features, our approach enables effective knowledge transfer while ensuring strong privacy guarantees. Experiments show that FedXDS consistently outperforms existing methods in accuracy and communication efficiency. Moreover, the sparsity induced by relevance attribution enhances privacy, mitigating feature inversion and membership inference risks. These findings highlight attribution methods as a promising tool for improving federated learning while preserving privacy and efficiency." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4579" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.279, + 0.108 + ], + "angle": 0, + "content": "9. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.486, + 0.171 + ], + "angle": 0, + "content": "This work was supported by the European Union's Horizon Europe research and innovation programme (EU Horizon Europe) as grant ACHILLES (101189689)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.182, + 0.188, + 0.198 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.207, + 0.484, + 0.262 + ], + "angle": 0, + "content": "[1] Durmus Alp Emre Acar, Yue Zhao, Ramon Matas, Matthew Mattina, Paul Whatmough, and Venkatesh Saligrama. Federated learning based on dynamic regularization. In International Conference on Learning Representations, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.264, + 0.484, + 0.317 + ], + "angle": 0, + "content": "[2] Leila Arras, Ahmed Osman, and Wojciech Samek. Clevr-xai: A benchmark dataset for the ground truth evaluation of neural network explanations. Information Fusion, 81:14-40, 2022. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.319, + 0.484, + 0.387 + ], + "angle": 0, + "content": "[3] Sebastian Bach, Alexander Binder, Grégoire Montavon, Frederick Klauschen, Klaus-Robert Müller, and Wojciech Samek. On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. PLOS ONE, 10(7):1-46, 2015. 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.389, + 0.484, + 0.456 + ], + "angle": 0, + "content": "[4] Jeremiah Blocki, Avrim Blum, Anupam Datta, and Or Sheffet. Differentially private data analysis of social networks via restricted sensitivity. In Proceedings of the 4th conference on Innovations in Theoretical Computer Science, pages 87-96, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.458, + 0.484, + 0.512 + ], + "angle": 0, + "content": "[5] Sebastian Caldas, Sai Meher Karthik Duddu, Peter Wu, Tian Li, Jakub Konečný, H Brendan McMahan, Virginia Smith, and Ameet Talwalkar. Leaf: A benchmark for federated settings. arXiv preprint arXiv:1812.01097, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.513, + 0.484, + 0.596 + ], + "angle": 0, + "content": "[6] Konstantinos Chatzikokolakis, Miguel E Andres, Nicolás Emilio Bordenabe, and Catuscia Palamidessi. Broadening the scope of differential privacy using metrics. In Privacy Enhancing Technologies: 13th International Symposium, PETS 2013, Bloomington, IN, USA, July 10-12, 2013. Proceedings 13, pages 82-102. Springer, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.597, + 0.484, + 0.651 + ], + "angle": 0, + "content": "[7] Daoyuan Chen, Dawei Gao, Weirui Kuang, Yaliang Li, and Bolin Ding. pfl-bench: A comprehensive benchmark for personalized federated learning. Advances in Neural Information Processing Systems, 35:9344-9360, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.652, + 0.484, + 0.719 + ], + "angle": 0, + "content": "[8] Rudresh Dwivedi, Devam Dave, Het Naik, Smiti Singhal, Rana Omer, Pankesh Patel, Bin Qian, Zhenyu Wen, Tejal Shah, Graham Morgan, et al. Explainable ai (xai): Core ideas, techniques, and solutions. ACM Computing Surveys, 55(9):1-33, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.721, + 0.484, + 0.762 + ], + "angle": 0, + "content": "[9] Cynthia Dwork. Differential privacy. In International colloquium on automata, languages, and programming, pages 1-12. Springer, 2006. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.763, + 0.484, + 0.817 + ], + "angle": 0, + "content": "[10] Oluwaseyi Feyisetan and Shiva Kasiviswanathan. Private release of text embedding vectors. In Proceedings of the First Workshop on Trustworthy Natural Language Processing, pages 15-27, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[11] Peter Kairouz, H Brendan McMahan, Brendan Avent, Aurélien Bellet, Mehdi Bennis, Arjun Nitin Bhagoji, Kallista Bonawitz, Zachary Charles, Graham Cormode, Rachel Cummings, et al. Advances and open problems in federated learning. Foundations and trends® in machine learning, 14(1-2): 1-210, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.207, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.175 + ], + "angle": 0, + "content": "[12] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank Reddi, Sebastian Stich, and Ananda Theertha Suresh. SCAFFOLD: Stochastic controlled averaging for federated learning. In Proceedings of the 37th International Conference on Machine Learning, pages 5132-5143. PMLR, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.906, + 0.245 + ], + "angle": 0, + "content": "[13] Sebastian Lapuschkin, Stephan Wäldchen, Alexander Binder, Grégoire Montavon, Wojciech Samek, and Klaus-Robert Müller. Unmasking clever hans predictors and assessing what machines really learn. Nature communications, 10(1):1096, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.906, + 0.317 + ], + "angle": 0, + "content": "[14] Bo Li, Mikkel N Schmidt, Tommy S Alstrom, and Sebastian U Stich. On the effectiveness of partial variance reduction in federated learning with heterogeneous data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3964-3973, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.906, + 0.375 + ], + "angle": 0, + "content": "[15] Qinbin Li, Bingsheng He, and Dawn Song. Model-contrastive federated learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10713–10722, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.906, + 0.432 + ], + "angle": 0, + "content": "[16] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. Proceedings of Machine learning and systems, 2:429-450, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.906, + 0.489 + ], + "angle": 0, + "content": "[17] Xiaoxiao Li, Meirui Jiang, Xiaofei Zhang, Michael Kamp, and Qi Dou. Fed{bn}: Federated learning on non- \\(\\{\\mathrm{id}\\}\\) features via local batch normalization. In International Conference on Learning Representations, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.906, + 0.545 + ], + "angle": 0, + "content": "[18] Tao Lin, Lingjing Kong, Sebastian U Stich, and Martin Jaggi. Ensemble distillation for robust model fusion in federated learning. Advances in neural information processing systems, 33:2351-2363, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.906, + 0.63 + ], + "angle": 0, + "content": "[19] H. Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Agüera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Proceedings of the 20th International Conference on Artificial Intelligence and Statistics (AISTATS). PMLR, 2017. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.906, + 0.702 + ], + "angle": 0, + "content": "[20] Matias Mendieta, Taojiannan Yang, Pu Wang, Minwoo Lee, Zhengming Ding, and Chen Chen. Local learning matters: Rethinking data heterogeneity in federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8397-8406, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.906, + 0.759 + ], + "angle": 0, + "content": "[21] Zhe Qu, Xingyu Li, Rui Duan, Yao Liu, Bo Tang, and Zhuo Lu. Generalized federated learning via sharpness aware minimization. In International conference on machine learning, pages 18250-18280. PMLR, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.906, + 0.829 + ], + "angle": 0, + "content": "[22] Wojciech Samek, Alexander Binder, Grégoire Montavon, Sebastian Lapischkin, and Klaus-Robert Müller. Evaluating the visualization of what a deep neural network has learned. IEEE Transactions on Neural Networks and Learning Systems, 28(11):2660-2673, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[23] Wojciech Samek, Grégoire Montavon, Sebastian Lapischkin, Christopher J. Anders, and Klaus-Robert Müller. Explaining deep neural networks and beyond: A review of methods and applications. Proceedings of the IEEE, 109(3): 247-278, 2021. 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "4580" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[24] Felix Sattler, Tim Korjakow, Roman Rischke, and Wojciech Samek. Fedaux: Leveraging unlabeled auxiliary data in federated learning. IEEE Transactions on Neural Networks and Learning Systems, 34(9):5531-5543, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[25] Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside convolutional networks: Visualising image classification models and saliency maps. In International Conference on Learning Representations (ICLR), 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.274 + ], + "angle": 0, + "content": "[26] Abhishek Singh, Praneeth Vepakomma, Vivek Sharma, and Ramesh Raskar. Posthoc privacy guarantees for collaborative inference with modified propose-test-release. Advances in Neural Information Processing Systems, 36:26438-26451, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[27] Daniel Smilkov, Nikhil Thorat, Been Kim, Fernanda Viégas, and Martin Wattenberg. Smoothgrad: removing noise by adding noise. arXiv preprint arXiv:1706.03825, 2017. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[28] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In International Conference on Machine Learning, pages 3319-3328. PMLR, 2017. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.444 + ], + "angle": 0, + "content": "[29] Jianyu Wang, Qinghua Liu, Hao Liang, Gauri Joshi, and H. Vincent Poor. Tackling the objective inconsistency problem in heterogeneous federated optimization. In Proceedings of the 34th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2020. Curran Associates Inc. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.446, + 0.482, + 0.514 + ], + "angle": 0, + "content": "[30] Zhiqin Yang, Yonggang Zhang, Yu Zheng, Xinmei Tian, Hao Peng, Tongliang Liu, and Bo Han. Fedfed: Feature distillation against data heterogeneity in federated learning. Advances in Neural Information Processing Systems, 36, 2024. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.517, + 0.482, + 0.571 + ], + "angle": 0, + "content": "[31] Mang Ye, Xiuwen Fang, Bo Du, Pong C Yuen, and Dacheng Tao. Heterogeneous federated learning: State-of-the-art and research challenges. ACM Computing Surveys, 56(3):1-44, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.574, + 0.482, + 0.641 + ], + "angle": 0, + "content": "[32] Rui Ye, Mingkai Xu, Jianyu Wang, Chenxin Xu, Siheng Chen, and Yanfeng Wang. Feddisco: Federated learning with discrepancy-aware collaboration. In International Conference on Machine Learning, pages 39879-39902. PMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.644, + 0.482, + 0.713 + ], + "angle": 0, + "content": "[33] Lin Zhang, Li Shen, Liang Ding, Dacheng Tao, and Ling-Yu Duan. Fine-tuning global model via data-free knowledge distillation for non-iid federated learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10174-10183, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.715, + 0.482, + 0.755 + ], + "angle": 0, + "content": "[34] Yue Zhao, Meng Li, Liangzhen Lai, Naveen Suda, Damon Civin, and Vikas Chandra. Federated learning with non-iid data. arXiv preprint arXiv:1806.00582, 2018. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.758, + 0.482, + 0.798 + ], + "angle": 0, + "content": "[35] Yue Zhao, Meng Li, Liangzhen Lai, Naveen Suda, Damon Civin, and Vikas Chandra. Federated learning with non-iid data. arXiv preprint arXiv:1806.00582, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.8, + 0.482, + 0.855 + ], + "angle": 0, + "content": "[36] Zhuangdi Zhu, Junyuan Hong, and Jiayu Zhou. Data-free knowledge distillation for heterogeneous federated learning. In Proceedings of the 38th International Conference on Machine Learning, pages 12878-12889. PMLR, 2021. 1, 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.855 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4581" + } + ] +] \ No newline at end of file diff --git a/2025/FedXDS_ Leveraging Model Attribution Methods to counteract Data Heterogeneity in Federated Learning/3d4823b4-ec38-431a-8ed1-cb351135854c_origin.pdf b/2025/FedXDS_ Leveraging Model Attribution Methods to counteract Data Heterogeneity in Federated Learning/3d4823b4-ec38-431a-8ed1-cb351135854c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6b3afe75a82aa5fd153078c48db484d69e433ba6 --- /dev/null +++ b/2025/FedXDS_ Leveraging Model Attribution Methods to counteract Data Heterogeneity in Federated Learning/3d4823b4-ec38-431a-8ed1-cb351135854c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d779d38e392d75b2c979acb257ef821bc9334130b60269390de2696078308510 +size 2255747 diff --git a/2025/FedXDS_ Leveraging Model Attribution Methods to counteract Data Heterogeneity in Federated Learning/full.md b/2025/FedXDS_ Leveraging Model Attribution Methods to counteract Data Heterogeneity in Federated Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..34e3fdff18a29282687c15e3762dc160ed3f2bcf --- /dev/null +++ b/2025/FedXDS_ Leveraging Model Attribution Methods to counteract Data Heterogeneity in Federated Learning/full.md @@ -0,0 +1,393 @@ +# FedXDS: Leveraging Model Attribution Methods to counteract Data Heterogeneity in Federated Learning + +Maximilian Andreas Hoefler $^{1}$ * Karsten Mueller $^{1}$ Wojciech Samek $^{1,2,3}$ $^{1}$ Fraunhofer Heinrich Hertz Institute, Berlin, Germany + $^{2}$ Technical University of Berlin, Berlin, Germany + $^{3}$ Berlin Institute for the Foundations of Learning and Data (BIFOLD) + +# Abstract + +Explainable AI (XAI) methods have demonstrated significant success in recent years at identifying relevant features in input data that drive deep learning model decisions, enhancing interpretability for users. However, the potential of XAI beyond providing model transparency has remained largely unexplored in adjacent machine learning domains. In this paper, we show for the first time how XAI can be utilized in the context of federated learning. Specifically, while federated learning enables collaborative model training without raw data sharing, it suffers from performance degradation when client data distributions exhibit statistical heterogeneity. We introduce FedXDS (Federated Learning via XAI-guided Data Sharing), the first approach to utilize feature attribution techniques to identify precisely which data elements should be selectively shared between clients to mitigate heterogeneity. By employing propagation-based attribution, our method identifies task-relevant features through a single backward pass, enabling selective data sharing that aligns client contributions. To protect sensitive information, we incorporate metric privacy techniques that provide formal privacy guarantees while preserving utility. Experimental results demonstrate that our approach consistently achieves higher accuracy and faster convergence compared to existing methods across varying client numbers and heterogeneity settings. We provide theoretical privacy guarantees and empirically demonstrate robustness against both membership inference and feature inversion attacks. Code is available at https://github.com/MaxH1996/FedXDS. + +# 1. Introduction + +Federated learning (FL) [19] has received significant attention in recent years, proving to be an effective method for collaborative distributed machine learning while keeping + +local datasets private. However, the practical deployment of federated learning systems faces fundamental challenges regarding data heterogeneity and privacy [11, 35]. + +Data heterogeneity, characterized by non-IID data across clients, leads to contradicting model updates and significantly degraded performance [12, 31]. Current approaches address this challenge by aligning client distributions using proximal optimization terms [16] and mitigating client drift [12, 14], while others focus on flattening the loss landscape to facilitate better model aggregation [20, 21]. In addition, recent works have shown that sharing even a small portion of client data globally can significantly improve local model generalization [34]. Such shared data introduces centralization, giving clients access to a more homogeneous distribution, which reduces update divergence and mitigates heterogeneity [34]. + +However, sharing raw data is not feasible due to privacy constraints. Introducing differential privacy [9] by adding noise to the shared data offers a potential solution, however this can cause considerable performance loss. Alternatively, instead of sharing raw data, abstract feature representations or partial data could be released using generator-type approaches [30, 33, 36]. However, the obtained features are not as potent as raw data in mitigating statistical heterogeneity, and can also be prone to privacy leaks and additionally introduce significant computational overhead due to generators. Hence, the challenge arises of how to retain the performance improvement from data sharing while simultaneously guaranteeing privacy. + +To address this challenge we propose FedXDS (Federated Learning via XAI-guided Data Sharing) which entails a novel approach wherein we leverage methods from the field of eXplainable AI (XAI) [8] enabling performance enhancing data sharing while preserving privacy. + +Specifically, we leverage propagation-based attribution methods, which have been shown to reliably identify input features that consistently contribute to model predictions [2, 8]. Moreover, attribution maps yield a pixel-wise relevance score which highlights regions in the input space + +which focus on semantically meaningful features, while suppressing spurious correlations and background noise. In our setup local client data is filtered through the attribution mechanism, i.e., we only share the input features which the attribution map deems most relevant for model predictions. This retains the information needed for model generalization while discarding irrelevant information. + +We show that our method, when combined with differential-metric privacy techniques [6, 9], offers stronger privacy guarantees compared to applying similar privacy mechanisms directly to raw input features. Specifically, by discarding task-irrelevant information we effectively reduce the dimensionality of the raw features. This allows us to selectively protect only the task-relevant regions rather than uniformly protecting all input pixels, and thus obtain better utility. + +In addition, propagation-based attributions can be obtained via a single backward pass over the local dataset, which only needs to be performed once in the entire FL process. + +Our main contributions can be summarized as: + +- A novel federated learning algorithm leveraging propagation-based attributions to address the challenge of retaining performance from data sharing with privacy guarantees in heterogeneous environments. +- A privacy-preserving mechanism that utilizes attribution-guided dimensionality reduction to achieve metric differential privacy. We perform theoretical and empirical privacy evaluations on our feature sharing approach through membership inference attacks and feature inversion, demonstrating that we can attain strong privacy guarantees. +- Extensive experiments on image datasets benchmarking our method against current state of the art, demonstrating superior performance in terms of accuracy and convergence. + +# 2. Preliminaries + +# 2.1. Federated Learning + +Federated Learning [19] is a distributed machine learning paradigm where multiple clients collaborate to train a model under the orchestration of a central server, without sharing their raw data. Let $K$ be the number of clients, each with a local dataset $\mathcal{D}_k = \{(x_i^k,y_i^k)\}_{i = 1}^{n_k}$ , where $x_{i}^{k}\in \mathcal{X}$ is an input sample and $y_{i}^{k}\in \mathcal{V}$ is its corresponding label. The goal is to learn a global model $f_{\theta}:\mathcal{X}\to \mathcal{Y}$ parameterized by $\theta$ , which minimizes the empirical risk: + +$$ +\min _ {\theta} \frac {1}{N} \sum_ {k = 1} ^ {K} \sum_ {i = 1} ^ {n _ {k}} \mathcal {L} \left(f _ {\theta} \left(x _ {i} ^ {k}\right), y _ {i} ^ {k}\right) \tag {1} +$$ + +where $N = \sum_{k=1}^{K} n_k$ is the total number of samples + +across all clients, and $\mathcal{L}$ is a suitable loss function. + +The FedAvg algorithm [19] iteratively aggregates model updates from selected clients $S_{t}$ , where each client $k$ performs local updates $\theta_{k}^{t} = \theta^{t - 1} - \eta \nabla \mathcal{L}_{k}(\theta^{t - 1})$ before the server averages these updates as $\theta^t = \frac{1}{|S_t|}\sum_{k\in S_t}\theta_k^t$ . + +# 2.2. Differential Privacy in Metric Spaces + +Standard differential privacy (DP) [9] provides guarantees against membership inference by bounding the output distribution change when a single dataset element is modified. For continuous data such as image embeddings, classical adjacency notions are restrictive, making metric differential privacy [6] more appropriate: + +Definition 1 (Metric Privacy). [6] Let $(X, d_X)$ be a metric space and $Z$ a set of possible outputs. A randomized mechanism $A\colon X \to \Delta(Z)$ is $(\varepsilon, \delta)$ -metric private if for all $x, x' \in X$ and every measurable set $U \subseteq Z$ : + +$$ +\Pr [ A (x) \in U ] \leq \exp \left(\varepsilon d _ {X} \left(x, x ^ {\prime}\right)\right) \Pr [ A \left(x ^ {\prime}\right) \in U ] + \delta . +$$ + +This definition reduces to standard DP when $d_X(x, x') = 1$ for all adjacent datasets, but naturally captures similarity in continuous spaces. In our work, we define $(X, d_X)$ where $X = \mathbb{R}^d$ represents image data and $d_X(x, x') = ||x - x'||_2$ is the $\ell_2$ distance between inputs. + +To guarantee privacy, we calibrate noise based on the sensitivity of a query function $f\colon X\to \mathbb{R}^m$ , which measures output changes relative to input changes. In our work we choose the following sensitivity measure as used in [4, 10, 26]: + +Definition 2 (Sensitivity). For a function $f \colon X \to \mathbb{R}^m$ , the sensitivity is: + +$$ +\Delta_ {f} = \max _ {x, x ^ {\prime} \in X} \frac {\| f (x) - f \left(x ^ {\prime}\right) \|}{\left| \left| x - x ^ {\prime} \right| \right|}. +$$ + +This definition captures the largest possible output difference relative to input difference, directly controlling how much noise is needed for privacy. This notion of sensitivity offers several advantages, which we discuss in the supplementary material. Importantly, when $\Delta_f$ is small, less noise is required to achieve the same privacy guarantee, preserving more signal and improving utility—particularly valuable for high-dimensional data like images. + +For privacy preservation, we employ the Gaussian mechanism, which adds noise to the output of our query function proportional to the sensitivity $\Delta_f$ : + +Theorem 1 (Gaussian Mechanism). [6, 9] For a function $f: \mathcal{D} \to \mathbb{R}^d$ with sensitivity $\Delta_f$ , adding Gaussian noise with scale $\sigma$ satisfying: + +$$ +\sigma \geq \frac {\Delta_ {f} \sqrt {2 \log (1 . 2 5 / \delta)}}{\varepsilon} \tag {2} +$$ + +ensures $f(x) + \mathcal{N}(0, \sigma^2 I)$ is $(\varepsilon, \delta)$ -metric private. + +This mechanism forms the foundation of our privacy-preserving feature sharing approach, allowing us to bound and quantify the privacy guarantees of our federated learning framework. + +# 2.3. Neural Network Attribution Methods + +We consider several gradient-based attribution techniques in our work. The simplest being Gradient $\times$ Input [25], which computes attributions through element-wise multiplication of input and gradient: + +$$ +\mathcal {A} _ {\operatorname {g r a d}} \left(f _ {\theta}, \mathbf {x}\right) = \mathbf {x} \odot \frac {\partial f _ {\theta}}{\partial \mathbf {x}} \tag {3} +$$ + +Integrated Gradients (IG) [28] addresses gradient saturation by accumulating gradients along a path from baseline $\mathbf{x}'$ to input: + +$$ +\mathcal {A} _ {\mathrm {I G}} \left(f _ {\theta}, \mathbf {x}\right) = \left(\mathbf {x} - \mathbf {x} ^ {\prime}\right) \odot \int_ {0} ^ {1} \frac {\partial f _ {\theta} \left(\mathbf {x} ^ {\prime} + \alpha \left(\mathbf {x} - \mathbf {x} ^ {\prime}\right)\right)}{\partial \mathbf {x}} d \alpha \tag {4} +$$ + +SmoothGrad [27] reduces attribution noise by averaging gradients over perturbed inputs: + +$$ +\mathcal {A} _ {\text {s m o o t h}} (f _ {\theta}, \mathbf {x}) = \frac {1}{N} \sum_ {i = 1} ^ {N} \frac {\partial f _ {\theta}}{\partial (\mathbf {x} + \epsilon)} \tag {5} +$$ + +Lastly we use Layer-wise Relevance Propagation (LRP) with the $\epsilon$ -rule according to [3]. This recursively propagates relevance scores $R^{(l)}$ from layer $l$ to $l - 1$ using: + +$$ +R _ {j} = \sum \frac {a _ {j} w _ {j k}}{\epsilon + \sum_ {0 , j} a _ {j} w _ {j k}} R _ {k} \tag {6} +$$ + +where $w_{jk}$ are layer weights, $a_{j}$ are layer activations, and $\epsilon$ is a small stabilizing term. + +# 3. Related Work + +Federated Learning was introduced with FedAvg [19], enabling collaborative model training without sharing raw data, but suffers under statistical heterogeneity. Regularization-based methods mitigate client drift by constraining local updates, such as FedProx [16], SCAFFOLD [12], and FedDyn [1], with extensions like FedBN [17], MOON [15], FedNova [29], FedSAM [21], and FedDISCO [32] targeting specific aspects like feature shifts or loss geometry. A more direct solution is data or knowledge sharing: FedDF and FedAux [18, 24] distill client knowledge using a public dataset; FedGen [36] synthesizes class-conditional features for distribution alignment; FedFTG [33] generates pseudo-data for feature-level transfer; and FedFed [30] applies VAE-based feature distillation. While effective, these methods often introduce significant computational and privacy costs due to data synthesis and sharing. In contrast, FedXDS shares compact + +![](images/c6b902572e5b77296570c376eb8544f36a1845179476bbbb780ddf654482818a.jpg) +Figure 1. Illustration of the DP attribution-guided feature extraction. The process consists of four main steps: (1) Computing attribution scores using a warmup model to identify important features, (2) Creating and applying a binary mask based on the attributions, (3) Adding Gaussian noise for metric privacy. + +Algorithm 1 Attribution-Guided Private Representation Extraction. + +Require: $D_{k}$ : client dataset, $f_{\theta}$ : model, $s$ : sparsity level, $(\varepsilon ,\delta)$ : privacy params, $S$ : sensitivity, $\theta_{\mathrm{warmup}}$ + +Ensure: Private dataset $D_k^p$ + +1: for $\mathbf{x}_i^k\in D_k$ do +2: $\mathbf{h} \gets \mathcal{A}(f_{\theta_{\mathrm{warmup}}}, \mathbf{x}_i^k)$ ; $h^{(s)} \gets \mathrm{s}$ -th largest in $\mathbf{h}$ +3: $\mathbf{m} \gets [\mathbf{h}[i] \geq h^{(s)}]; \mathbf{x}_m \gets \mathbf{x}_i^k \odot \mathbf{m}$ +4: $\tilde{\mathbf{x}}_i^k\gets \mathbf{x}_m + \mathcal{N}(0,\sigma^2)$ $\triangleright \sigma$ from $(\varepsilon ,\delta)$ +5: Add $(\tilde{\mathbf{x}}_i^k,y_i^k)$ to $D_k^p$ +6: end for +7: return $D_k^p$ + +attribution-based feature subsets derived via XAI, avoiding generators while enhancing privacy, efficiency, and performance in heterogeneous settings. Wo provide an extensive related works in the supplementary material. + +# 4. FedXDS Approach + +Consider a federated learning system with $N$ clients, where each client $k$ holds local data $\mathcal{D}_k$ . In our setting all clients share the same neural network architecture $f_{\theta}$ . Our goal is to learn a global model $\Theta_G$ which generalizes well on all client datasets. + +# 4.1. Attribution-Guided Feature Selection + +The first stage of our method identifies discriminative features using a pre-trained warmup model, with parameters $\theta_{\mathrm{warmup}}$ . This is achieved by training with FedAvg for $R_{\mathrm{warmup}}$ rounds, before sharing representations. Then, for each input sample $\mathbf{x} \in \mathbb{R}^{H \times W \times 3}$ , where $H$ and $W$ are the + +Algorithm 2 Privacy-Preserving Federated Training +Require: $\{D_k\}_{k=1}^K$ : client datasets, $\lambda$ : knowledge weight, $T$ : rounds, $E$ : local epochs, $\eta$ : learning rate, $R_{\text{warmup}}$ +1: Initialize $\theta_{\text{warmup}}$ using $R_{\text{warmup}}$ rounds of FedAvg +2: Obtain $D_k^p$ using Algorithm 1 for all $k \in [K]$ in parallel +3: Server: $\mathcal{D}_g \gets \bigcup_{k=1}^K D_k^p$ +4: for round $t = 1$ to $T$ do +5: Select client subset $S_t \subseteq [K]$ +6: for $k \in S_t$ in parallel do +7: $\theta_k^t \gets \theta^{t-1}$ Download global model +8: for epoch $e = 1$ to $E$ do +9: for batch $B$ from $D_k$ , $B_g$ from $\mathcal{D}_g$ do +10: $\mathcal{L}_k^t \gets \frac{1}{|B_k|} \sum_{(\mathbf{x}, y) \in B_k} \ell(f_{\theta_k^t}(\mathbf{x}), y) + \lambda \frac{1}{|B_g|} \sum_{(\mathbf{x}_g, y_g) \in B_g} \ell(f_{\theta_k^t}(\mathbf{x}_g), y_g)$ +11: $\theta_k^t \gets \theta_k^t - \eta \nabla \mathcal{L}$ +12: end for +13: end for +14: end for +15: Server: $\theta^t \gets \frac{1}{|S_t|} \sum_{k \in S_t} \theta_k^t$ >FedAvg +16: end for + +dimensions of the image with three color channels, in the client dataset $\mathcal{D}_k$ , we compute relevances using an attribution method $\mathcal{A}$ from Section 2.3: + +$$ +\mathbf {h} = \mathcal {A} \left(f _ {\theta}, \mathbf {x}\right) \tag {7} +$$ + +where $\mathbf{h} \in \mathbb{R}^{H \times W}$ is the pixel-wise importance score of an input image $\mathbf{x}$ . These attribution scores quantify the importance of each input feature to the model's prediction. To focus on the most relevant features, we create a binary mask $\mathbf{m}$ , as shown in Equation 8, by retaining features whose attribution scores exceed a threshold determined by the desired sparsity level $s$ : + +$$ +\mathbf {m} [ i ] = \left\{ \begin{array}{l l} 1 & \text {i f} \mathbf {h} [ i ] \geq h ^ {(s)} \\ 0 & \text {o t h e r w i s e} \end{array} \right. \tag {8} +$$ + +where $h^{(s)}$ represents the $s$ -th largest value in the attribution scores. This mask identifies the subset of features that are most crucial for the model's decision-making process. + +We define our feature selection function $f_{A} \colon \mathbb{R}^{H \times W \times 3} \to \mathbb{R}^{H \times W \times 3}$ as the application of the attribution mask $\mathbf{m}$ to each image $\mathbf{x} \in \mathcal{D}_k$ through: + +$$ +f _ {A} (\mathbf {x}) = \mathbf {x} \odot \mathbf {m} \tag {9} +$$ + +This operation emphasizes the most informative parts of the input while suppressing less relevant details. However, directly sharing $f_{A}(\mathbf{x})$ would still pose privacy risks. Therefore, we design a privacy mechanism $\mathcal{M}$ that adds calibrated noise to these masked features (Section refsubsec:privacy): + +$$ +\mathcal {M} (\mathbf {x}) = f _ {A} (\mathbf {x}) + \mathcal {N} (0, \sigma^ {2} \mathbf {I}) \tag {10} +$$ + +where $\sigma$ is determined based on privacy requirements. The resulting privacy-protected features, along with their corresponding labels, are then aggregated into a new dataset $\mathcal{D}_k^p = \{(\mathcal{M}(\mathbf{x}),y)|(x,y)\in \mathcal{D}_k\}$ and shared with the server. The full data generated process is outlined in Algorithm 1 and visualized in Figure 1. + +# 4.2. Local Training using Shared Data + +After the privacy-protected features have been obtained, the server aggregates private representations from all clients into a global dataset $\mathcal{D}_g = \bigcup_{k=1}^N \mathcal{D}_k^p$ , and sends this dataset to each client. The clients then optimize a composite objective that balances local task performance with knowledge from the global private dataset: + +$$ +\begin{array}{l} \min _ {\theta} \left[ \mathbb {E} _ {(\mathbf {x}, y) \sim \mathcal {D} _ {k}} [ \ell (f _ {\theta} (\mathbf {x}), y) ] \right. \tag {11} \\ + \lambda \mathbb {E} _ {(\mathbf {x}, y) \sim \mathcal {D} _ {g}} [ \ell (f _ {\theta} (\mathbf {x}), y) ] ] \\ \end{array} +$$ + +The hyperparameter $\lambda$ controls the trade-off between local specialization and global knowledge integration. This formulation allows clients to benefit from collective learning while maintaining their own task-specific performance and ensuring privacy. The procedure is outlined in Algorithm 2. + +# 4.3. Privacy Discussion + +We build on the insight that sharing features derived from raw client data can mitigate data heterogeneity in federated learning [34]. Nevertheless, directly sharing these features poses significant privacy risks. A straightforward approach would be to apply differential privacy (DP) to the raw data itself; however, the high dimensionality of image data forces the addition of very large noise levels, thereby significantly degrading utility. + +To address this challenge, we propose a two-step approach. First, we use an attribution-guided masking strategy (Section 4.1) to identify and retain only task-relevant pixels. Second, we add Gaussian noise (Theorem 1) to the resulting masked features to achieve the desired $\varepsilon$ -metric privacy (Definition 1, Definition 2). + +This design offers two main advantages. First, by sparsifying input features that do not contribute to classification predictions, we reduce the overall sensitivity of the data. According to Theorem 1, lower sensitivity requires less noise to meet the same privacy budget $\varepsilon$ , thus improving utility. Second, in contrast to random or magnitude-based sparsification that may unintentionally discard valuable information, our attribution-guided approach specifically preserves the pixels most critical for the learning task while eliminating irrelevant ones. + +The primary goal of our privacy mechanism is to protect against membership inference attacks (MIA) and feature inversion, which we empirically validate in section 6. + +In the following section, we formally demonstrate how our strategy guarantees metric privacy, while a comprehensive discussion of privacy assumptions appears in the appendix. + +# 4.3.1. Sensitivity Analysis of Attribution-Based Masking + +A critical component of ensuring privacy is understanding the sensitivity of our feature selection function $f(\mathbf{x})$ defined in Equation 9. From Theorem 1, the noise required to attain a certain privacy level $\varepsilon$ depends on the sensitivity $\Delta_f$ . Our goal is to minimize this sensitivity before applying the privacy mechanism $\mathcal{M}$ . + +For our feature selection function $f_{A}(\mathbf{x}) = \mathbf{x}\odot \mathbf{m}$ , we can derive a strict bound on sensitivity. Since each coordinate of the mask satisfies $\mathbf{m}[i]\in \{0,1\}$ , for any two inputs $\mathbf{x},\mathbf{x}'\in \mathbb{R}^{H\times W\times 3}$ , we have, in the worst case: + +$$ +\left| \mathbf {m} [ i ] \mathbf {x} [ i ] - \mathbf {m} [ i ] \mathbf {x} ^ {\prime} [ i ] \right| \leq \left| \mathbf {x} [ i ] - \mathbf {x} ^ {\prime} [ i ] \right| +$$ + +Thus, in aggregate: + +$$ +\left\| f _ {A} (\mathbf {x}) - f _ {A} \left(\mathbf {x} ^ {\prime}\right) \right\| _ {2} \leq \left\| \mathbf {x} - \mathbf {x} ^ {\prime} \right\| _ {2} +$$ + +This shows that our masking operation is non-expansive with respect to the $\ell_2$ norm. Rearranging the equation: + +$$ +\Delta_ {f} = \frac {\left\| f _ {A} (\mathbf {x}) - f _ {A} \left(\mathbf {x} ^ {\prime}\right) \right\| _ {2}}{\left\| \mathbf {x} - \mathbf {x} ^ {\prime} \right\| _ {2}} \leq 1 \tag {12} +$$ + +where $\Delta_f$ is the sensitivity we defined in the preliminaries in Definition 2. Without our attribution-based masking, the sensitivity would be unbounded for arbitrary transformations. Moreover, in the worst case $f_A$ is the identity, i.e. $f_A(x) = x$ , meaning that the attribution method deems all pixels important. This implies that our method would reduce to the naive approach of adding noise to the raw features. In practice, this rarely occurs, since this would already require a highly sparsified image. Nevertheless, our approach safeguards against this worst-case scenario. More discussion can be found in the supplementary material. + +# 4.3.2. Privacy Mechanism and Guarantees + +Based on the sensitivity bound established above, we design our privacy mechanism $\mathcal{M}$ (Equation 10) by adding Gaussian noise calibrated to this reduced sensitivity: + +$$ +\mathcal {M} (\mathbf {x}) = f _ {A} (\mathbf {x}) + \mathcal {N} (0, \sigma^ {2} \mathbf {I}) +$$ + +where $\sigma$ is chosen according to Theorem 1 for an $(\varepsilon, \delta)$ -level of privacy given sensitivity $\Delta_f \leq 1$ . This ensures that for all input images in the client dataset, $\mathbf{x}, \mathbf{x}' \in D_k$ , our method satisfies $(\varepsilon, \delta)$ -metric privacy. + +Our theoretical privacy guarantees are empirically validated through membership inference and feature inversion attacks Section 6, where attribution-masked features consistently show stronger protection than unmasked features under equivalent privacy budgets. + +# 5. Experimental + +# 5.1. Experimental Setup + +We evaluate FedXDS on standard image classification benchmarks (CIFAR-10, CIFAR-100, Tiny-ImageNet) and real-world federated datasets from the LEAF framework [5] (CelebA, FEMNIST) that naturally exhibit heterogeneous distributions. For the standard benchmarks, we simulate statistical heterogeneity using Dirichlet-based partitioning with concentration parameters $\alpha \in \{0.05, 0.1\}$ , where smaller values indicate greater heterogeneity. We conduct experiments with both 10 and 100 clients, setting participation rates to 0.5 and 0.1, respectively, and train for $R = 200$ communication rounds. To compute relevance-based attributions, we evaluate four attribution methods in our FL framework: Layerwise Relevance Propagation (FedXLRP) [3], Integrated Gradients (FedXIG) [28], SmoothGrad (FedXSG) [27], and Input $\times$ Gradient (FedXGrad). Unless otherwise state we use $\varepsilon = 20$ sparsity level $s = 70\%$ , and $\lambda = 0.5$ . + +# 5.2. Baselines + +The experimental results in Table 1 show that FedXLRP consistently outperforms all baselines across varying client numbers and data distributions. This advantage amplifies in more challenging scenarios (K=100, $\alpha = 0.05$ ). FedFed ranks second, while FedAvg and FedAux degrade significantly with increased client numbers. On CIFAR-100, performance gaps narrow, though FedXLRP maintains its edge, particularly with higher client counts and heterogeneity. Other attribution methods (FedXIG, FedXGrad, FedXSG) perform well but remain below FedXLRP, highlighting attribution method choice importance. We discuss this point in Subsection 7.1. In addition, communication efficiency results Table 2 reinforce FedXLRP's advantages. To reach $70\%$ accuracy on CIFAR-10 with 10 clients, FedXLRP requires only 14 rounds versus 49 for FedAvg and 28 for FedFTG. With 100 clients, FedXLRP needs 23 rounds to reach $60\%$ accuracy compared to FedAvg's 79. CIFAR-100 shows similar patterns, with FedXLRP achieving target accuracies in substantially fewer rounds across client configurations. Among baselines, only FedFed shows comparable efficiency. + +# 5.3. Experiments on Real-World Datasets + +We also evaluate FedXDS on real-world datasets using implementations from the LEAF [5] and [7] frameworks. Specifically, we consider CelebA and FEMNIST, which exhibit a distinct form of non-IIIDness known as distribution shift, introducing an additional challenge for federated learning. For CelebA, we follow the implementation of [36], while for FEMNIST, we use the setup from [7], conducting experiments with 10 and 100 clients, respectively. + +Table 1. Performance comparison across different federated learning methods, including standard deviations from 5 runs. Results show the top-1 test accuracy (%) on CIFAR-10, CIFAR-100, and Tiny-ImageNet under varying numbers of clients (K) and Dirichlet concentration parameters $(\alpha)$ . Bold values indicate the best performance in each column. + +
DatasetCIFAR-10CIFAR-100Tiny-ImageNet
K=10K=100K=10K=100K=10K=100
α = 0.05α = 0.1α = 0.05α = 0.1α = 0.05α = 0.1α = 0.05α = 0.1α = 0.05α = 0.1α = 0.05α = 0.1
FedAvg64.16 ± 0.7172.90 ± 0.5555.53 ± 0.8260.94 ± 0.6843.82 ± 0.4544.57 ± 0.5132.99 ± 0.7335.59 ± 0.6230.30 ± 0.5934.41 ± 0.4826.20 ± 0.8128.79 ± 0.75
FedProx71.56 ± 0.6574.31 ± 0.4951.89 ± 0.9165.01 ± 0.5344.33 ± 0.3848.67 ± 0.4737.79 ± 0.5538.43 ± 0.4933.27 ± 0.4234.54 ± 0.5124.15 ± 0.8830.23 ± 0.67
FedDyn70.62 ± 0.5878.25 ± 0.3463.09 ± 0.6169.41 ± 0.4548.53 ± 0.4149.42 ± 0.3937.14 ± 0.6841.60 ± 0.5433.03 ± 0.4835.69 ± 0.4129.52 ± 0.6332.37 ± 0.59
SCAFFOLD71.35 ± 0.6175.62 ± 0.4258.51 ± 0.7963.82 ± 0.6246.61 ± 0.3347.27 ± 0.4435.28 ± 0.2438.87 ± 0.5831.53 ± 0.5335.76 ± 0.3928.35 ± 0.5930.49 ± 0.61
FedSAM70.15 ± 0.8877.53 ± 0.5162.45 ± 0.7268.78 ± 0.5947.95 ± 0.3548.81 ± 0.4136.88 ± 0.5140.92 ± 0.4732.21 ± 0.5535.03 ± 0.4428.98 ± 0.6631.84 ± 0.52
FedDISCO69.73 ± 0.4476.98 ± 0.6361.82 ± 0.6868.11 ± 0.6547.31 ± 0.5248.15 ± 0.5836.17 ± 0.6740.25 ± 0.6131.78 ± 0.6134.68 ± 0.5228.51 ± 0.7131.33 ± 0.64
FedFed79.12 ± 0.4882.58 ± 0.3175.35 ± 0.5778.87 ± 0.4352.27 ± 0.4956.45 ± 0.3845.41 ± 0.7649.28 ± 0.5234.99 ± 0.4536.49 ± 0.3733.28 ± 0.5834.89 ± 0.47
FedFTG75.21 ± 0.5378.44 ± 0.3970.84 ± 0.3174.85 ± 0.4144.91 ± 0.6254.15 ± 0.4540.43 ± 0.3946.47 ± 0.4832.40 ± 0.5733.79 ± 0.4930.53 ± 0.5132.26 ± 0.55
FedGen68.34 ± 0.6874.65 ± 0.4758.90 ± 0.7563.20 ± 0.6644.29 ± 0.5550.12 ± 0.5136.70 ± 0.6239.82 ± 0.5734.17 ± 0.4436.33 ± 0.4229.45 ± 0.6831.60 ± 0.61
FedAux59.72 ± 0.8573.70 ± 0.5153.13 ± 0.8865.11 ± 0.5944.05 ± 0.4844.85 ± 0.5336.54 ± 0.6538.39 ± 0.6127.83 ± 0.7234.31 ± 0.4924.75 ± 0.8230.38 ± 0.69
FedDF60.56 ± 0.7772.40 ± 0.5859.20 ± 0.7263.47 ± 0.6430.13 ± 0.8145.47 ± 0.5034.46 ± 0.6936.77 ± 0.6429.07 ± 0.6534.75 ± 0.4628.42 ± 0.6230.51 ± 0.63
FedXLRP81.72 ± 0.3583.46 ± 0.2877.02 ± 0.6380.27 ± 0.3952.07 ± 0.4558.09 ± 0.3346.25 ± 0.8252.63 ± 0.4136.85 ± 0.3938.64 ± 0.3234.64 ± 0.5233.18 ± 0.51
FedXIG72.22 ± 0.5975.89 ± 0.4464.02 ± 0.6968.61 ± 0.5549.21 ± 0.4855.46 ± 0.4142.34 ± 0.5847.89 ± 0.5033.59 ± 0.5135.30 ± 0.4529.77 ± 0.6431.93 ± 0.58
FedXGrad71.38 ± 0.6273.99 ± 0.5264.52 ± 0.6566.49 ± 0.6150.13 ± 0.4354.21 ± 0.4642.11 ± 0.6046.98 ± 0.5333.18 ± 0.5434.40 ± 0.4830.00 ± 0.6130.93 ± 0.60
FedXSG71.89 ± 0.6074.63 ± 0.4963.75 ± 0.7167.80 ± 0.5851.57 ± 0.3955.22 ± 0.4240.91 ± 0.6347.61 ± 0.5133.43 ± 0.5234.74 ± 0.4729.64 ± 0.6531.53 ± 0.59
+ +Table 2. Communication efficiency comparison across different federated learning methods. Results show the number of communication rounds (mean ± std over 5 runs) needed to achieve target accuracy thresholds (70% and 60% for CIFAR-10; 40% and 30% for CIFAR-100; 35% and 30% for Tiny-ImageNet) with different numbers of clients (K=10 and K=100) and heterogeneity parameter $\alpha = 0.1$ . Lower values indicate better communication efficiency. Best results are in bold. + +
MethodCIFAR-10CIFAR-100Tiny-ImageNet
acc=70% K=10acc=60% K=100acc=40% K=10acc=30% K=100acc=35% K=10acc=30% K=100
FedAvg49 ± 579 ± 724 ± 434 ± 560 ± 690 ± 8
FedProx32 ± 355 ± 428 ± 432 ± 445 ± 465 ± 5
FedDyn29 ± 351 ± 424 ± 332 ± 348 ± 560 ± 4
SCAFFOLD33 ± 452 ± 525 ± 338 ± 446 ± 469 ± 6
FedSAM34 ± 458 ± 527 ± 435 ± 450 ± 564 ± 5
FedDISCO36 ± 361 ± 628 ± 437 ± 552 ± 567 ± 6
FedFed15 ± 222 ± 413 ± 412 ± 312 ± 215 ± 2
FedFTG28 ± 357 ± 422 ± 328 ± 330 ± 340 ± 4
FedGen51 ± 566 ± 621 ± 439 ± 555 ± 570 ± 6
FedAux38 ± 445 ± 419 ± 336 ± 442 ± 450 ± 5
FedDF84 ± 790 ± 828 ± 445 ± 5100 ± 8110 ± 9
FedXLRP14 ± 223 ± 311 ± 313 ± 210 ± 412 ± 2
FedXIG35 ± 439 ± 416 ± 221 ± 338 ± 445 ± 4
FedXGrad41 ± 446 ± 513 ± 226 ± 344 ± 450 ± 5
FedXSG41 ± 551 ± 421 ± 333 ± 442 ± 448 ± 5
+ +Additional implementation details can be found in the supplementary material. The results in Table 3 show a clear trend where FedXDS consistently achieves the highest accuracy across both datasets, demonstrating its effectiveness in handling distribution shift. Notably, methods incorporating additional data generation or augmentation, such as FedFTG and FedGen, also perform well, but FedXDS surpasses them, suggesting that relevance-guided data sharing provides a stronger mechanism for improving generaliza + +tion in heterogeneous federated settings. + +The results in Table 4 demonstrate that integrating FedXDS with FL methods consistently improves the performance. In all cases, the addition of FedXDS yields substantial accuracy gains, particularly in the highly heterogeneous setting with $\mathrm{K} = 100$ clients. For instance, FedAvg, which traditionally suffers from performance degradation in non-IID scenarios, improves from $60.94\%$ to $80.29\%$ when combined with FedXDS. Similar improvements are observed for + +Table 3. Accuracy comparison (mean ± std % over 5 runs) on CelebA and FEMNIST. Best results are in bold. We use the LRP variant of FedXDS. + +
MethodCelebAFEMNIST
FedAvg87.23 ± 0.7184.71 ± 0.68
FedProx87.79 ± 0.6585.66 ± 0.61
SCAFFOLD86.36 ± 0.5684.24 ± 0.59
FedDyn88.13 ± 0.5286.45 ± 0.34
FedSAM90.03 ± 0.1587.56 ± 0.44
FedDISCO89.78 ± 0.6187.29 ± 0.62
FedFTG90.31 ± 0.4387.88 ± 0.67
FedGen89.65 ± 0.5986.37 ± 0.55
FedDF88.56 ± 0.6385.28 ± 0.66
FedAux89.27 ± 0.6085.94 ± 0.62
FedFed90.76 ± 0.8388.28 ± 0.71
FedXDS91.55 ± 0.4889.03 ± 0.35
+ +Table 4. Performance comparison (mean ± std % over 5 runs) on CIFAR-10 for different base FL methods and their FedXDS (using LRP) variants under two heterogeneity settings ( $K = 10$ and $K = 100$ ). The FedXDS variants consistently improve performance. + +
MethodK=10K=100
FedAvg72.90 ± 0.5560.94 ± 0.68
FedAvg + FedXDS83.46 ± 0.3180.29 ± 0.42
FedProx74.31 ± 0.4965.01 ± 0.53
FedProx + FedXDS84.05 ± 0.3580.73 ± 0.45
FedDyn70.62 ± 0.5869.41 ± 0.45
FedDyn + FedXDS84.20 ± 0.3379.56 ± 0.48
SCAFFOLD71.35 ± 0.6163.82 ± 0.62
SCAFFOLD + FedXDS83.89 ± 0.3879.72 ± 0.44
+ +FedProx, FedDyn, and SCAFFOLD, with FedXDS enhancing their robustness to data heterogeneity while maintaining their inherent advantages. + +The magnitude of improvement suggests that FedXDS effectively mitigates the adverse effects of statistical heterogeneity by facilitating better knowledge transfer across clients. Notably, FedDyn, which incorporates regularization-based local adaptation, exhibits the smallest absolute improvement with FedXDS, indicating that its existing adaptation mechanisms already partially address heterogeneity. In contrast, methods like FedAvg, without correction for client drift, benefit more from FedXDS. + +# 6. Empirical Privacy Analysis + +# 6.1. Membership Inference Attack + +We conduct membership inference attacks following [30] to evaluate our methodology's privacy guarantees. We investigate whether training on masked and noised features reveals + +Table 5. Test accuracy (mean ± std % over 5 runs) for different sparsification levels $s$ across various attribution methods. Best results are in bold. Results for CIFAR10 at $K = 10, \alpha = 0.1$ . + +
sGradientSmoothGradInt. Grad.LRP
6075.99 ± 0.5275.63 ± 0.5577.10 ± 0.4884.74 ± 0.31
6574.19 ± 0.5673.15 ± 0.6176.31 ± 0.5184.01 ± 0.35
7073.99 ± 0.5974.63 ± 0.5475.89 ± 0.4983.46 ± 0.28
7570.54 ± 0.6570.67 ± 0.6371.68 ± 0.5883.01 ± 0.33
8068.31 ± 0.7170.45 ± 0.6670.98 ± 0.6282.67 ± 0.36
8568.80 ± 0.6868.13 ± 0.7369.73 ± 0.6781.88 ± 0.41
+ +membership in the underlying dataset by using the globally DP data to train a shadow model, then testing if clients can infer membership by querying this model and training a random forest classifier. Figure 2a shows recall values for different privacy parameters $\varepsilon$ . Results demonstrate that non-masked features, despite DP protection, offer inferior privacy protection compared to masked data. This can be explained by the fact that sparsity significantly mitigates attacks through norm reduction from attribution masks (Equation 8). + +# 6.2. Feature Inversion Attack + +While metric private features prevent direct inference of individual data features, a sophisticated adversary might train a denoising autoencoder to learn mappings between protected representations and original images. We conduct feature inversion attacks and report the SSIM score between original and reconstructed images. Figure 2b reveals inversion attack success increases with lower sparsity and higher privacy budgets. Unmasked features consistently exhibit the highest SSIM, confirming that raw features under identical DP constraints provide inferior protection compared to our approach. Additional validations in supplementary material, including visualization of inversion results for non-masked features and FedFed [30] features, confirming both cases' vulnerability to reconstruction, whereas our method provides superior protection. + +# 7. Ablation Studies and Discussion + +# 7.1. XAI Method Comparison Across Sparsification + +Table 5 demonstrates that FedXDS maintains stable performance across different sparsification levels, with all attribution methods achieving above $68\%$ accuracy, even at high sparsity thresholds. Notably, LRP consistently outperforms other methods, preserving accuracies between $81.88\%$ and $84.74\%$ , while others degrade more sharply under extreme sparsification. LRP's robustness stems from its layer-wise relevance propagation mechanism, which is fundamentally different from gradient-based approaches. Instead of measuring local sensitivity, LRP redistributes the + +![](images/d654f65426f1b09337378c9b8bb673358720c17eb2d5a29177b9ccba3e8b785e.jpg) +(a) Recall score of MI attack at different $\varepsilon$ for various sparsity levels $s$ . + +![](images/334fbfd003e818507f5a5d76ad2c6de6ee2b76487b76bccdd60b2ee70ed5646a.jpg) +(b) SSIM score of denoising attack at different $\varepsilon$ and sparsity levels $s$ . + +![](images/0507348d4aaec163abc51ebeba93ede0d6190e7357363e40077b79034915f19b.jpg) +(c) Effect of knowledge weight $\lambda$ on model accuracy across datasets. + +![](images/7c0e6832821a718b23e96a981dcd2488e770e4ce2890d8fc52319f8cc3cb09ee.jpg) +(d) Impact of warmup rounds on model accuracy across datasets. +Figure 2. Evaluation of privacy-utility trade-offs and hyperparameter effects in our framework. + +model's output prediction backward through the network, layer by layer, conserving relevance at each step. This process aggregates relevance into contiguous, semantically coherent regions that represent structured object parts. In stark contrast, methods like SmoothGrad or Integrated Gradients identify 'hotspot' pixels that are locally influential but often disconnected from each other. Consequently, under high sparsification, LRP retains meaningful visual patterns, whereas other methods are left with a sparse collection of isolated points that lack sufficient context for accurate classification. We conclude that structural coherence in saliency maps is an essential criterion when selecting the attribution method. This finding is corroborated by current literature [2, 13, 23], which confirms LRP's superior explanation quality and reinforces that our work is consistent with established findings on the importance of explanation quality for downstream tasks. More discussion can be found in the supplementary material. + +# 7.2. Privacy vs. Utility + +We investigate how the privacy budget $\varepsilon$ impacts model performance across attribution methods. Table 6 shows that increasing $\varepsilon$ improves accuracy for all methods, demonstrating the expected privacy-utility trade-off: stronger privacy (lower $\varepsilon$ ) leads to reduced performance due to greater noise. Among all methods, LRP consistently achieves the highest accuracy across privacy levels outperforming the next best (Integrated Gradients) by $7 - 8\%$ and maintaining robustness even under strict privacy constraints. SmoothGrad shows modest improvements over basic gradients, though less pronounced than Integrated Gradients. These results underscore LRP's ability to extract stable, task-relevant features leading to better generalization with less privacy leakage, aligning with prior work on explanation fidelity [3, 22]. + +# 7.3. Ablation Study on the Knowledge Weight + +We evaluate how the knowledge weight $\lambda$ , as defined in Equation 11, affects performance across CIFAR-10, CIFAR-100, and Tiny-ImageNet. Figure 2c shows that intermediate values ( $\lambda = 0.5$ ) consistently outperform both extremes, with CIFAR-10 accuracy peaking at $83.46\%$ compared to $72.90\%$ ( $\lambda = 0.0$ , equivalent to FedAvg) and + +Table 6. Ablation study showing test accuracy) for different privacy budgets $(\varepsilon)$ across XAI-based federated learning methods. Results for CIFAR10 at $K = 10, \alpha = 0.1$ + +
εGradientSmoothGradInt. Grad.LRP
570.55 ± 0.6271.23 ± 0.5972.45 ± 0.5280.06 ± 0.33
771.45 ± 0.5872.15 ± 0.5573.35 ± 0.4780.89 ± 0.31
1072.25 ± 0.5172.95 ± 0.4874.15 ± 0.4181.65 ± 0.29
1573.12 ± 0.4573.85 ± 0.4375.02 ± 0.3882.11 ± 0.25
2073.99 ± 0.4274.63 ± 0.3975.89 ± 0.3583.46 ± 0.28
+ +61.18% ( $\lambda = 1.0$ , synthetic data equal weight). Our ablation thus shows that balancing the contribution of synthetic data is critical for performance. + +# 7.4. Warmup and Efficiency + +Figure 2d shows accuracy slightly improves with more warmup rounds, plateauing at 10, suggesting this is the optimal trade-off between convergence and communication cost. Our approach is also computationally efficient. Attribution masks are computed in a single backward pass during warmup, incurring minimal overhead. In contrast, methods like FedGen and FedFTG retrain generators every round, and FedFed requires costly VAE training—making FedXDS more suitable for resource-constrained settings. We provide more discussion in the supplementary material. + +# 8. Conclusion + +In this work, we introduced FedXDS, leveraging attribution methods to address statistical heterogeneity, privacy, and computational efficiency in federated learning. By using attribution maps to extract and share privacy-preserved task-relevant features, our approach enables effective knowledge transfer while ensuring strong privacy guarantees. Experiments show that FedXDS consistently outperforms existing methods in accuracy and communication efficiency. Moreover, the sparsity induced by relevance attribution enhances privacy, mitigating feature inversion and membership inference risks. These findings highlight attribution methods as a promising tool for improving federated learning while preserving privacy and efficiency. + +# 9. Acknowledgements + +This work was supported by the European Union's Horizon Europe research and innovation programme (EU Horizon Europe) as grant ACHILLES (101189689). + +# References + +[1] Durmus Alp Emre Acar, Yue Zhao, Ramon Matas, Matthew Mattina, Paul Whatmough, and Venkatesh Saligrama. Federated learning based on dynamic regularization. In International Conference on Learning Representations, 2021. 3 +[2] Leila Arras, Ahmed Osman, and Wojciech Samek. Clevr-xai: A benchmark dataset for the ground truth evaluation of neural network explanations. Information Fusion, 81:14-40, 2022. 1, 8 +[3] Sebastian Bach, Alexander Binder, Grégoire Montavon, Frederick Klauschen, Klaus-Robert Müller, and Wojciech Samek. On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. PLOS ONE, 10(7):1-46, 2015. 3, 5, 8 +[4] Jeremiah Blocki, Avrim Blum, Anupam Datta, and Or Sheffet. Differentially private data analysis of social networks via restricted sensitivity. In Proceedings of the 4th conference on Innovations in Theoretical Computer Science, pages 87-96, 2013. 2 +[5] Sebastian Caldas, Sai Meher Karthik Duddu, Peter Wu, Tian Li, Jakub Konečný, H Brendan McMahan, Virginia Smith, and Ameet Talwalkar. Leaf: A benchmark for federated settings. arXiv preprint arXiv:1812.01097, 2018. 5 +[6] Konstantinos Chatzikokolakis, Miguel E Andres, Nicolás Emilio Bordenabe, and Catuscia Palamidessi. Broadening the scope of differential privacy using metrics. In Privacy Enhancing Technologies: 13th International Symposium, PETS 2013, Bloomington, IN, USA, July 10-12, 2013. Proceedings 13, pages 82-102. Springer, 2013. 2 +[7] Daoyuan Chen, Dawei Gao, Weirui Kuang, Yaliang Li, and Bolin Ding. pfl-bench: A comprehensive benchmark for personalized federated learning. Advances in Neural Information Processing Systems, 35:9344-9360, 2022. 5 +[8] Rudresh Dwivedi, Devam Dave, Het Naik, Smiti Singhal, Rana Omer, Pankesh Patel, Bin Qian, Zhenyu Wen, Tejal Shah, Graham Morgan, et al. Explainable ai (xai): Core ideas, techniques, and solutions. ACM Computing Surveys, 55(9):1-33, 2023. 1 +[9] Cynthia Dwork. Differential privacy. In International colloquium on automata, languages, and programming, pages 1-12. Springer, 2006. 1, 2 +[10] Oluwaseyi Feyisetan and Shiva Kasiviswanathan. Private release of text embedding vectors. In Proceedings of the First Workshop on Trustworthy Natural Language Processing, pages 15-27, 2021. 2 +[11] Peter Kairouz, H Brendan McMahan, Brendan Avent, Aurélien Bellet, Mehdi Bennis, Arjun Nitin Bhagoji, Kallista Bonawitz, Zachary Charles, Graham Cormode, Rachel Cummings, et al. Advances and open problems in federated learning. Foundations and trends® in machine learning, 14(1-2): 1-210, 2021. 1 + +[12] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank Reddi, Sebastian Stich, and Ananda Theertha Suresh. SCAFFOLD: Stochastic controlled averaging for federated learning. In Proceedings of the 37th International Conference on Machine Learning, pages 5132-5143. PMLR, 2020. 1, 3 +[13] Sebastian Lapuschkin, Stephan Wäldchen, Alexander Binder, Grégoire Montavon, Wojciech Samek, and Klaus-Robert Müller. Unmasking clever hans predictors and assessing what machines really learn. Nature communications, 10(1):1096, 2019. 8 +[14] Bo Li, Mikkel N Schmidt, Tommy S Alstrom, and Sebastian U Stich. On the effectiveness of partial variance reduction in federated learning with heterogeneous data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3964-3973, 2023. 1 +[15] Qinbin Li, Bingsheng He, and Dawn Song. Model-contrastive federated learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10713–10722, 2021. 3 +[16] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. Proceedings of Machine learning and systems, 2:429-450, 2020. 1, 3 +[17] Xiaoxiao Li, Meirui Jiang, Xiaofei Zhang, Michael Kamp, and Qi Dou. Fed{bn}: Federated learning on non- $\{\mathrm{id}\}$ features via local batch normalization. In International Conference on Learning Representations, 2021. 3 +[18] Tao Lin, Lingjing Kong, Sebastian U Stich, and Martin Jaggi. Ensemble distillation for robust model fusion in federated learning. Advances in neural information processing systems, 33:2351-2363, 2020. 3 +[19] H. Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Agüera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Proceedings of the 20th International Conference on Artificial Intelligence and Statistics (AISTATS). PMLR, 2017. 1, 2, 3 +[20] Matias Mendieta, Taojiannan Yang, Pu Wang, Minwoo Lee, Zhengming Ding, and Chen Chen. Local learning matters: Rethinking data heterogeneity in federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8397-8406, 2022. 1 +[21] Zhe Qu, Xingyu Li, Rui Duan, Yao Liu, Bo Tang, and Zhuo Lu. Generalized federated learning via sharpness aware minimization. In International conference on machine learning, pages 18250-18280. PMLR, 2022. 1, 3 +[22] Wojciech Samek, Alexander Binder, Grégoire Montavon, Sebastian Lapischkin, and Klaus-Robert Müller. Evaluating the visualization of what a deep neural network has learned. IEEE Transactions on Neural Networks and Learning Systems, 28(11):2660-2673, 2017. 8 +[23] Wojciech Samek, Grégoire Montavon, Sebastian Lapischkin, Christopher J. Anders, and Klaus-Robert Müller. Explaining deep neural networks and beyond: A review of methods and applications. Proceedings of the IEEE, 109(3): 247-278, 2021. 8 + +[24] Felix Sattler, Tim Korjakow, Roman Rischke, and Wojciech Samek. Fedaux: Leveraging unlabeled auxiliary data in federated learning. IEEE Transactions on Neural Networks and Learning Systems, 34(9):5531-5543, 2021. 3 +[25] Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside convolutional networks: Visualising image classification models and saliency maps. In International Conference on Learning Representations (ICLR), 2014. 3 +[26] Abhishek Singh, Praneeth Vepakomma, Vivek Sharma, and Ramesh Raskar. Posthoc privacy guarantees for collaborative inference with modified propose-test-release. Advances in Neural Information Processing Systems, 36:26438-26451, 2023. 2 +[27] Daniel Smilkov, Nikhil Thorat, Been Kim, Fernanda Viégas, and Martin Wattenberg. Smoothgrad: removing noise by adding noise. arXiv preprint arXiv:1706.03825, 2017. 3, 5 +[28] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In International Conference on Machine Learning, pages 3319-3328. PMLR, 2017. 3, 5 +[29] Jianyu Wang, Qinghua Liu, Hao Liang, Gauri Joshi, and H. Vincent Poor. Tackling the objective inconsistency problem in heterogeneous federated optimization. In Proceedings of the 34th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2020. Curran Associates Inc. 3 +[30] Zhiqin Yang, Yonggang Zhang, Yu Zheng, Xinmei Tian, Hao Peng, Tongliang Liu, and Bo Han. Fedfed: Feature distillation against data heterogeneity in federated learning. Advances in Neural Information Processing Systems, 36, 2024. 1, 3, 7 +[31] Mang Ye, Xiuwen Fang, Bo Du, Pong C Yuen, and Dacheng Tao. Heterogeneous federated learning: State-of-the-art and research challenges. ACM Computing Surveys, 56(3):1-44, 2023. 1 +[32] Rui Ye, Mingkai Xu, Jianyu Wang, Chenxin Xu, Siheng Chen, and Yanfeng Wang. Feddisco: Federated learning with discrepancy-aware collaboration. In International Conference on Machine Learning, pages 39879-39902. PMLR, 2023. 3 +[33] Lin Zhang, Li Shen, Liang Ding, Dacheng Tao, and Ling-Yu Duan. Fine-tuning global model via data-free knowledge distillation for non-iid federated learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10174-10183, 2022. 1, 3 +[34] Yue Zhao, Meng Li, Liangzhen Lai, Naveen Suda, Damon Civin, and Vikas Chandra. Federated learning with non-iid data. arXiv preprint arXiv:1806.00582, 2018. 1, 4 +[35] Yue Zhao, Meng Li, Liangzhen Lai, Naveen Suda, Damon Civin, and Vikas Chandra. Federated learning with non-iid data. arXiv preprint arXiv:1806.00582, 2018. 1 +[36] Zhuangdi Zhu, Junyuan Hong, and Jiayu Zhou. Data-free knowledge distillation for heterogeneous federated learning. In Proceedings of the 38th International Conference on Machine Learning, pages 12878-12889. PMLR, 2021. 1, 3, 5 \ No newline at end of file diff --git a/2025/FedXDS_ Leveraging Model Attribution Methods to counteract Data Heterogeneity in Federated Learning/images.zip b/2025/FedXDS_ Leveraging Model Attribution Methods to counteract Data Heterogeneity in Federated Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..199debf5f2f1897bcc0714d9675432afe63fc0fb --- /dev/null +++ b/2025/FedXDS_ Leveraging Model Attribution Methods to counteract Data Heterogeneity in Federated Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:427befccfeb2b45df30fbee275e57e66089b44749a5328bb730c9630e9034568 +size 540115 diff --git a/2025/FedXDS_ Leveraging Model Attribution Methods to counteract Data Heterogeneity in Federated Learning/layout.json b/2025/FedXDS_ Leveraging Model Attribution Methods to counteract Data Heterogeneity in Federated Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ed6fa8d9d5e8d98a52fe3180962fe2b7c5e337cc --- /dev/null +++ b/2025/FedXDS_ Leveraging Model Attribution Methods to counteract Data Heterogeneity in Federated Learning/layout.json @@ -0,0 +1,10638 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 94, + 102, + 516, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 102, + 516, + 140 + ], + "spans": [ + { + "bbox": [ + 94, + 102, + 516, + 140 + ], + "type": "text", + "content": "FedXDS: Leveraging Model Attribution Methods to counteract Data Heterogeneity in Federated Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 125, + 160, + 484, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 160, + 484, + 217 + ], + "spans": [ + { + "bbox": [ + 125, + 160, + 484, + 217 + ], + "type": "text", + "content": "Maximilian Andreas Hoefler" + }, + { + "bbox": [ + 125, + 160, + 484, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 125, + 160, + 484, + 217 + ], + "type": "text", + "content": " * Karsten Mueller" + }, + { + "bbox": [ + 125, + 160, + 484, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 125, + 160, + 484, + 217 + ], + "type": "text", + "content": " Wojciech Samek" + }, + { + "bbox": [ + 125, + 160, + 484, + 217 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 125, + 160, + 484, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 125, + 160, + 484, + 217 + ], + "type": "text", + "content": "Fraunhofer Heinrich Hertz Institute, Berlin, Germany \n" + }, + { + "bbox": [ + 125, + 160, + 484, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 125, + 160, + 484, + 217 + ], + "type": "text", + "content": "Technical University of Berlin, Berlin, Germany \n" + }, + { + "bbox": [ + 125, + 160, + 484, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 125, + 160, + 484, + 217 + ], + "type": "text", + "content": "Berlin Institute for the Foundations of Learning and Data (BIFOLD)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 270, + 294, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 270, + 294, + 616 + ], + "spans": [ + { + "bbox": [ + 56, + 270, + 294, + 616 + ], + "type": "text", + "content": "Explainable AI (XAI) methods have demonstrated significant success in recent years at identifying relevant features in input data that drive deep learning model decisions, enhancing interpretability for users. However, the potential of XAI beyond providing model transparency has remained largely unexplored in adjacent machine learning domains. In this paper, we show for the first time how XAI can be utilized in the context of federated learning. Specifically, while federated learning enables collaborative model training without raw data sharing, it suffers from performance degradation when client data distributions exhibit statistical heterogeneity. We introduce FedXDS (Federated Learning via XAI-guided Data Sharing), the first approach to utilize feature attribution techniques to identify precisely which data elements should be selectively shared between clients to mitigate heterogeneity. By employing propagation-based attribution, our method identifies task-relevant features through a single backward pass, enabling selective data sharing that aligns client contributions. To protect sensitive information, we incorporate metric privacy techniques that provide formal privacy guarantees while preserving utility. Experimental results demonstrate that our approach consistently achieves higher accuracy and faster convergence compared to existing methods across varying client numbers and heterogeneity settings. We provide theoretical privacy guarantees and empirically demonstrate robustness against both membership inference and feature inversion attacks. Code is available at https://github.com/MaxH1996/FedXDS." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 639, + 135, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 639, + 135, + 651 + ], + "spans": [ + { + "bbox": [ + 56, + 639, + 135, + 651 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 659, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 295, + 696 + ], + "type": "text", + "content": "Federated learning (FL) [19] has received significant attention in recent years, proving to be an effective method for collaborative distributed machine learning while keeping" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 246, + 553, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 246, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 553, + 281 + ], + "type": "text", + "content": "local datasets private. However, the practical deployment of federated learning systems faces fundamental challenges regarding data heterogeneity and privacy [11, 35]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 282, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 282, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 312, + 282, + 555, + 437 + ], + "type": "text", + "content": "Data heterogeneity, characterized by non-IID data across clients, leads to contradicting model updates and significantly degraded performance [12, 31]. Current approaches address this challenge by aligning client distributions using proximal optimization terms [16] and mitigating client drift [12, 14], while others focus on flattening the loss landscape to facilitate better model aggregation [20, 21]. In addition, recent works have shown that sharing even a small portion of client data globally can significantly improve local model generalization [34]. Such shared data introduces centralization, giving clients access to a more homogeneous distribution, which reduces update divergence and mitigates heterogeneity [34]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 438, + 554, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 438, + 554, + 593 + ], + "spans": [ + { + "bbox": [ + 312, + 438, + 554, + 593 + ], + "type": "text", + "content": "However, sharing raw data is not feasible due to privacy constraints. Introducing differential privacy [9] by adding noise to the shared data offers a potential solution, however this can cause considerable performance loss. Alternatively, instead of sharing raw data, abstract feature representations or partial data could be released using generator-type approaches [30, 33, 36]. However, the obtained features are not as potent as raw data in mitigating statistical heterogeneity, and can also be prone to privacy leaks and additionally introduce significant computational overhead due to generators. Hence, the challenge arises of how to retain the performance improvement from data sharing while simultaneously guaranteeing privacy." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 594, + 553, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 553, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 553, + 653 + ], + "type": "text", + "content": "To address this challenge we propose FedXDS (Federated Learning via XAI-guided Data Sharing) which entails a novel approach wherein we leverage methods from the field of eXplainable AI (XAI) [8] enabling performance enhancing data sharing while preserving privacy." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "Specifically, we leverage propagation-based attribution methods, which have been shown to reliably identify input features that consistently contribute to model predictions [2, 8]. Moreover, attribution maps yield a pixel-wise relevance score which highlights regions in the input space" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 285, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 285, + 712 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 285, + 712 + ], + "type": "text", + "content": "*Correspondence to maximilian.andreas.hoefler@hhi.fraunhofer.de" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4572" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "which focus on semantically meaningful features, while suppressing spurious correlations and background noise. In our setup local client data is filtered through the attribution mechanism, i.e., we only share the input features which the attribution map deems most relevant for model predictions. This retains the information needed for model generalization while discarding irrelevant information." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 156, + 294, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 156, + 294, + 263 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 294, + 263 + ], + "type": "text", + "content": "We show that our method, when combined with differential-metric privacy techniques [6, 9], offers stronger privacy guarantees compared to applying similar privacy mechanisms directly to raw input features. Specifically, by discarding task-irrelevant information we effectively reduce the dimensionality of the raw features. This allows us to selectively protect only the task-relevant regions rather than uniformly protecting all input pixels, and thus obtain better utility." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 264, + 294, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 264, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 294, + 312 + ], + "type": "text", + "content": "In addition, propagation-based attributions can be obtained via a single backward pass over the local dataset, which only needs to be performed once in the entire FL process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 312, + 244, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 244, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 244, + 323 + ], + "type": "text", + "content": "Our main contributions can be summarized as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 325, + 294, + 505 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 325, + 294, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 325, + 294, + 372 + ], + "spans": [ + { + "bbox": [ + 55, + 325, + 294, + 372 + ], + "type": "text", + "content": "- A novel federated learning algorithm leveraging propagation-based attributions to address the challenge of retaining performance from data sharing with privacy guarantees in heterogeneous environments." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 373, + 294, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 294, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 294, + 456 + ], + "type": "text", + "content": "- A privacy-preserving mechanism that utilizes attribution-guided dimensionality reduction to achieve metric differential privacy. We perform theoretical and empirical privacy evaluations on our feature sharing approach through membership inference attacks and feature inversion, demonstrating that we can attain strong privacy guarantees." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 456, + 294, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 294, + 505 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 294, + 505 + ], + "type": "text", + "content": "- Extensive experiments on image datasets benchmarking our method against current state of the art, demonstrating superior performance in terms of accuracy and convergence." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 516, + 139, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 139, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 139, + 529 + ], + "type": "text", + "content": "2. Preliminaries" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 536, + 171, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 536, + 171, + 549 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 171, + 549 + ], + "type": "text", + "content": "2.1. Federated Learning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 553, + 294, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 553, + 294, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 553, + 294, + 649 + ], + "type": "text", + "content": "Federated Learning [19] is a distributed machine learning paradigm where multiple clients collaborate to train a model under the orchestration of a central server, without sharing their raw data. Let " + }, + { + "bbox": [ + 55, + 553, + 294, + 649 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 553, + 294, + 649 + ], + "type": "text", + "content": " be the number of clients, each with a local dataset " + }, + { + "bbox": [ + 55, + 553, + 294, + 649 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_k = \\{(x_i^k,y_i^k)\\}_{i = 1}^{n_k}" + }, + { + "bbox": [ + 55, + 553, + 294, + 649 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 553, + 294, + 649 + ], + "type": "inline_equation", + "content": "x_{i}^{k}\\in \\mathcal{X}" + }, + { + "bbox": [ + 55, + 553, + 294, + 649 + ], + "type": "text", + "content": " is an input sample and " + }, + { + "bbox": [ + 55, + 553, + 294, + 649 + ], + "type": "inline_equation", + "content": "y_{i}^{k}\\in \\mathcal{V}" + }, + { + "bbox": [ + 55, + 553, + 294, + 649 + ], + "type": "text", + "content": " is its corresponding label. The goal is to learn a global model " + }, + { + "bbox": [ + 55, + 553, + 294, + 649 + ], + "type": "inline_equation", + "content": "f_{\\theta}:\\mathcal{X}\\to \\mathcal{Y}" + }, + { + "bbox": [ + 55, + 553, + 294, + 649 + ], + "type": "text", + "content": " parameterized by " + }, + { + "bbox": [ + 55, + 553, + 294, + 649 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 553, + 294, + 649 + ], + "type": "text", + "content": ", which minimizes the empirical risk:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 113, + 660, + 294, + 692 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 660, + 294, + 692 + ], + "spans": [ + { + "bbox": [ + 113, + 660, + 294, + 692 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\frac {1}{N} \\sum_ {k = 1} ^ {K} \\sum_ {i = 1} ^ {n _ {k}} \\mathcal {L} \\left(f _ {\\theta} \\left(x _ {i} ^ {k}\\right), y _ {i} ^ {k}\\right) \\tag {1}", + "image_path": "5f3e2fdb26fdf7e0b5e2a1a4b46e09d235ec54d61b65104412f6e05c55e98730.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 67, + 700, + 294, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 700, + 294, + 715 + ], + "spans": [ + { + "bbox": [ + 67, + 700, + 294, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 67, + 700, + 294, + 715 + ], + "type": "inline_equation", + "content": "N = \\sum_{k=1}^{K} n_k" + }, + { + "bbox": [ + 67, + 700, + 294, + 715 + ], + "type": "text", + "content": " is the total number of samples" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 517, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 517, + 83 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 517, + 83 + ], + "type": "text", + "content": "across all clients, and " + }, + { + "bbox": [ + 313, + 72, + 517, + 83 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 313, + 72, + 517, + 83 + ], + "type": "text", + "content": " is a suitable loss function." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 84, + 553, + 135 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 84, + 553, + 135 + ], + "spans": [ + { + "bbox": [ + 313, + 84, + 553, + 135 + ], + "type": "text", + "content": "The FedAvg algorithm [19] iteratively aggregates model updates from selected clients " + }, + { + "bbox": [ + 313, + 84, + 553, + 135 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 313, + 84, + 553, + 135 + ], + "type": "text", + "content": ", where each client " + }, + { + "bbox": [ + 313, + 84, + 553, + 135 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 84, + 553, + 135 + ], + "type": "text", + "content": " performs local updates " + }, + { + "bbox": [ + 313, + 84, + 553, + 135 + ], + "type": "inline_equation", + "content": "\\theta_{k}^{t} = \\theta^{t - 1} - \\eta \\nabla \\mathcal{L}_{k}(\\theta^{t - 1})" + }, + { + "bbox": [ + 313, + 84, + 553, + 135 + ], + "type": "text", + "content": " before the server averages these updates as " + }, + { + "bbox": [ + 313, + 84, + 553, + 135 + ], + "type": "inline_equation", + "content": "\\theta^t = \\frac{1}{|S_t|}\\sum_{k\\in S_t}\\theta_k^t" + }, + { + "bbox": [ + 313, + 84, + 553, + 135 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 140, + 510, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 140, + 510, + 152 + ], + "spans": [ + { + "bbox": [ + 313, + 140, + 510, + 152 + ], + "type": "text", + "content": "2.2. Differential Privacy in Metric Spaces" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 156, + 553, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 553, + 229 + ], + "type": "text", + "content": "Standard differential privacy (DP) [9] provides guarantees against membership inference by bounding the output distribution change when a single dataset element is modified. For continuous data such as image embeddings, classical adjacency notions are restrictive, making metric differential privacy [6] more appropriate:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 233, + 554, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 233, + 554, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 233, + 554, + 281 + ], + "type": "text", + "content": "Definition 1 (Metric Privacy). [6] Let " + }, + { + "bbox": [ + 313, + 233, + 554, + 281 + ], + "type": "inline_equation", + "content": "(X, d_X)" + }, + { + "bbox": [ + 313, + 233, + 554, + 281 + ], + "type": "text", + "content": " be a metric space and " + }, + { + "bbox": [ + 313, + 233, + 554, + 281 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 313, + 233, + 554, + 281 + ], + "type": "text", + "content": " a set of possible outputs. A randomized mechanism " + }, + { + "bbox": [ + 313, + 233, + 554, + 281 + ], + "type": "inline_equation", + "content": "A\\colon X \\to \\Delta(Z)" + }, + { + "bbox": [ + 313, + 233, + 554, + 281 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 313, + 233, + 554, + 281 + ], + "type": "inline_equation", + "content": "(\\varepsilon, \\delta)" + }, + { + "bbox": [ + 313, + 233, + 554, + 281 + ], + "type": "text", + "content": "-metric private if for all " + }, + { + "bbox": [ + 313, + 233, + 554, + 281 + ], + "type": "inline_equation", + "content": "x, x' \\in X" + }, + { + "bbox": [ + 313, + 233, + 554, + 281 + ], + "type": "text", + "content": " and every measurable set " + }, + { + "bbox": [ + 313, + 233, + 554, + 281 + ], + "type": "inline_equation", + "content": "U \\subseteq Z" + }, + { + "bbox": [ + 313, + 233, + 554, + 281 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 287, + 553, + 307 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 287, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 314, + 287, + 553, + 307 + ], + "type": "interline_equation", + "content": "\\Pr [ A (x) \\in U ] \\leq \\exp \\left(\\varepsilon d _ {X} \\left(x, x ^ {\\prime}\\right)\\right) \\Pr [ A \\left(x ^ {\\prime}\\right) \\in U ] + \\delta .", + "image_path": "e81611a68943a55a28df61814e058afff741f7f42dc5b90fe2c3d1665ef41cd2.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 311, + 553, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 553, + 371 + ], + "type": "text", + "content": "This definition reduces to standard DP when " + }, + { + "bbox": [ + 313, + 311, + 553, + 371 + ], + "type": "inline_equation", + "content": "d_X(x, x') = 1" + }, + { + "bbox": [ + 313, + 311, + 553, + 371 + ], + "type": "text", + "content": " for all adjacent datasets, but naturally captures similarity in continuous spaces. In our work, we define " + }, + { + "bbox": [ + 313, + 311, + 553, + 371 + ], + "type": "inline_equation", + "content": "(X, d_X)" + }, + { + "bbox": [ + 313, + 311, + 553, + 371 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 311, + 553, + 371 + ], + "type": "inline_equation", + "content": "X = \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 311, + 553, + 371 + ], + "type": "text", + "content": " represents image data and " + }, + { + "bbox": [ + 313, + 311, + 553, + 371 + ], + "type": "inline_equation", + "content": "d_X(x, x') = ||x - x'||_2" + }, + { + "bbox": [ + 313, + 311, + 553, + 371 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 313, + 311, + 553, + 371 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 313, + 311, + 553, + 371 + ], + "type": "text", + "content": " distance between inputs." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 371, + 553, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 553, + 430 + ], + "type": "text", + "content": "To guarantee privacy, we calibrate noise based on the sensitivity of a query function " + }, + { + "bbox": [ + 313, + 371, + 553, + 430 + ], + "type": "inline_equation", + "content": "f\\colon X\\to \\mathbb{R}^m" + }, + { + "bbox": [ + 313, + 371, + 553, + 430 + ], + "type": "text", + "content": ", which measures output changes relative to input changes. In our work we choose the following sensitivity measure as used in [4, 10, 26]:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 434, + 553, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 553, + 459 + ], + "type": "text", + "content": "Definition 2 (Sensitivity). For a function " + }, + { + "bbox": [ + 313, + 434, + 553, + 459 + ], + "type": "inline_equation", + "content": "f \\colon X \\to \\mathbb{R}^m" + }, + { + "bbox": [ + 313, + 434, + 553, + 459 + ], + "type": "text", + "content": ", the sensitivity is:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 370, + 463, + 497, + 490 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 463, + 497, + 490 + ], + "spans": [ + { + "bbox": [ + 370, + 463, + 497, + 490 + ], + "type": "interline_equation", + "content": "\\Delta_ {f} = \\max _ {x, x ^ {\\prime} \\in X} \\frac {\\| f (x) - f \\left(x ^ {\\prime}\\right) \\|}{\\left| \\left| x - x ^ {\\prime} \\right| \\right|}.", + "image_path": "97a8d9d9b966bf2c205a0ccd50df9faa3db834407c8e9416ce0b9a521c34311c.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 494, + 553, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 494, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 494, + 553, + 589 + ], + "type": "text", + "content": "This definition captures the largest possible output difference relative to input difference, directly controlling how much noise is needed for privacy. This notion of sensitivity offers several advantages, which we discuss in the supplementary material. Importantly, when " + }, + { + "bbox": [ + 313, + 494, + 553, + 589 + ], + "type": "inline_equation", + "content": "\\Delta_f" + }, + { + "bbox": [ + 313, + 494, + 553, + 589 + ], + "type": "text", + "content": " is small, less noise is required to achieve the same privacy guarantee, preserving more signal and improving utility—particularly valuable for high-dimensional data like images." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 590, + 553, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 553, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 553, + 626 + ], + "type": "text", + "content": "For privacy preservation, we employ the Gaussian mechanism, which adds noise to the output of our query function proportional to the sensitivity " + }, + { + "bbox": [ + 313, + 590, + 553, + 626 + ], + "type": "inline_equation", + "content": "\\Delta_f" + }, + { + "bbox": [ + 313, + 590, + 553, + 626 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 630, + 553, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 553, + 666 + ], + "type": "text", + "content": "Theorem 1 (Gaussian Mechanism). [6, 9] For a function " + }, + { + "bbox": [ + 313, + 630, + 553, + 666 + ], + "type": "inline_equation", + "content": "f: \\mathcal{D} \\to \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 630, + 553, + 666 + ], + "type": "text", + "content": " with sensitivity " + }, + { + "bbox": [ + 313, + 630, + 553, + 666 + ], + "type": "inline_equation", + "content": "\\Delta_f" + }, + { + "bbox": [ + 313, + 630, + 553, + 666 + ], + "type": "text", + "content": ", adding Gaussian noise with scale " + }, + { + "bbox": [ + 313, + 630, + 553, + 666 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 630, + 553, + 666 + ], + "type": "text", + "content": " satisfying:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 381, + 670, + 553, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 670, + 553, + 696 + ], + "spans": [ + { + "bbox": [ + 381, + 670, + 553, + 696 + ], + "type": "interline_equation", + "content": "\\sigma \\geq \\frac {\\Delta_ {f} \\sqrt {2 \\log (1 . 2 5 / \\delta)}}{\\varepsilon} \\tag {2}", + "image_path": "b5931528d1aa7ed96ea837348e2abc0b48d39336dcb04745e26ead728c40d19b.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 700, + 518, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 700, + 518, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 700, + 518, + 714 + ], + "type": "text", + "content": "ensures " + }, + { + "bbox": [ + 313, + 700, + 518, + 714 + ], + "type": "inline_equation", + "content": "f(x) + \\mathcal{N}(0, \\sigma^2 I)" + }, + { + "bbox": [ + 313, + 700, + 518, + 714 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 313, + 700, + 518, + 714 + ], + "type": "inline_equation", + "content": "(\\varepsilon, \\delta)" + }, + { + "bbox": [ + 313, + 700, + 518, + 714 + ], + "type": "text", + "content": "-metric private." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4573" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "content": "This mechanism forms the foundation of our privacy-preserving feature sharing approach, allowing us to bound and quantify the privacy guarantees of our federated learning framework." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 127, + 253, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 127, + 253, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 127, + 253, + 140 + ], + "type": "text", + "content": "2.3. Neural Network Attribution Methods" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 144, + 296, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 296, + 194 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 296, + 194 + ], + "type": "text", + "content": "We consider several gradient-based attribution techniques in our work. The simplest being Gradient " + }, + { + "bbox": [ + 55, + 144, + 296, + 194 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 55, + 144, + 296, + 194 + ], + "type": "text", + "content": " Input [25], which computes attributions through element-wise multiplication of input and gradient:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 124, + 200, + 295, + 224 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 200, + 295, + 224 + ], + "spans": [ + { + "bbox": [ + 124, + 200, + 295, + 224 + ], + "type": "interline_equation", + "content": "\\mathcal {A} _ {\\operatorname {g r a d}} \\left(f _ {\\theta}, \\mathbf {x}\\right) = \\mathbf {x} \\odot \\frac {\\partial f _ {\\theta}}{\\partial \\mathbf {x}} \\tag {3}", + "image_path": "cde51f73a43ddde437fd04d50139eb54d7dcca387e3dd5a77384b672bfdf8e6b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 230, + 296, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 230, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 230, + 296, + 266 + ], + "type": "text", + "content": "Integrated Gradients (IG) [28] addresses gradient saturation by accumulating gradients along a path from baseline " + }, + { + "bbox": [ + 55, + 230, + 296, + 266 + ], + "type": "inline_equation", + "content": "\\mathbf{x}'" + }, + { + "bbox": [ + 55, + 230, + 296, + 266 + ], + "type": "text", + "content": " to input:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 60, + 274, + 295, + 301 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 274, + 295, + 301 + ], + "spans": [ + { + "bbox": [ + 60, + 274, + 295, + 301 + ], + "type": "interline_equation", + "content": "\\mathcal {A} _ {\\mathrm {I G}} \\left(f _ {\\theta}, \\mathbf {x}\\right) = \\left(\\mathbf {x} - \\mathbf {x} ^ {\\prime}\\right) \\odot \\int_ {0} ^ {1} \\frac {\\partial f _ {\\theta} \\left(\\mathbf {x} ^ {\\prime} + \\alpha \\left(\\mathbf {x} - \\mathbf {x} ^ {\\prime}\\right)\\right)}{\\partial \\mathbf {x}} d \\alpha \\tag {4}", + "image_path": "7b63ce175980e0477a5e3ed4b6ea30ca03a7a7960cac7b14f16314eacf3d7ded.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 308, + 296, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 308, + 296, + 333 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 296, + 333 + ], + "type": "text", + "content": "SmoothGrad [27] reduces attribution noise by averaging gradients over perturbed inputs:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 340, + 295, + 373 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 340, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 104, + 340, + 295, + 373 + ], + "type": "interline_equation", + "content": "\\mathcal {A} _ {\\text {s m o o t h}} (f _ {\\theta}, \\mathbf {x}) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\frac {\\partial f _ {\\theta}}{\\partial (\\mathbf {x} + \\epsilon)} \\tag {5}", + "image_path": "0323d8b25a6827fef92a8fcf48363ad40b8b2003d9a07071945e449a46e96b80.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 380, + 296, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 296, + 416 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 296, + 416 + ], + "type": "text", + "content": "Lastly we use Layer-wise Relevance Propagation (LRP) with the " + }, + { + "bbox": [ + 55, + 380, + 296, + 416 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 380, + 296, + 416 + ], + "type": "text", + "content": "-rule according to [3]. This recursively propagates relevance scores " + }, + { + "bbox": [ + 55, + 380, + 296, + 416 + ], + "type": "inline_equation", + "content": "R^{(l)}" + }, + { + "bbox": [ + 55, + 380, + 296, + 416 + ], + "type": "text", + "content": " from layer " + }, + { + "bbox": [ + 55, + 380, + 296, + 416 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 380, + 296, + 416 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 380, + 296, + 416 + ], + "type": "inline_equation", + "content": "l - 1" + }, + { + "bbox": [ + 55, + 380, + 296, + 416 + ], + "type": "text", + "content": " using:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 113, + 423, + 295, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 423, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 113, + 423, + 295, + 449 + ], + "type": "interline_equation", + "content": "R _ {j} = \\sum \\frac {a _ {j} w _ {j k}}{\\epsilon + \\sum_ {0 , j} a _ {j} w _ {j k}} R _ {k} \\tag {6}", + "image_path": "f7d56b937ee887fb79a159d92f398688a4dc34d907c644039e470a65a7a07df4.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 456, + 296, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 296, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 296, + 480 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 456, + 296, + 480 + ], + "type": "inline_equation", + "content": "w_{jk}" + }, + { + "bbox": [ + 55, + 456, + 296, + 480 + ], + "type": "text", + "content": " are layer weights, " + }, + { + "bbox": [ + 55, + 456, + 296, + 480 + ], + "type": "inline_equation", + "content": "a_{j}" + }, + { + "bbox": [ + 55, + 456, + 296, + 480 + ], + "type": "text", + "content": " are layer activations, and " + }, + { + "bbox": [ + 55, + 456, + 296, + 480 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 456, + 296, + 480 + ], + "type": "text", + "content": " is a small stabilizing term." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 490, + 143, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 490, + 143, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 490, + 143, + 502 + ], + "type": "text", + "content": "3. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "content": "Federated Learning was introduced with FedAvg [19], enabling collaborative model training without sharing raw data, but suffers under statistical heterogeneity. Regularization-based methods mitigate client drift by constraining local updates, such as FedProx [16], SCAFFOLD [12], and FedDyn [1], with extensions like FedBN [17], MOON [15], FedNova [29], FedSAM [21], and FedDISCO [32] targeting specific aspects like feature shifts or loss geometry. A more direct solution is data or knowledge sharing: FedDF and FedAux [18, 24] distill client knowledge using a public dataset; FedGen [36] synthesizes class-conditional features for distribution alignment; FedFTG [33] generates pseudo-data for feature-level transfer; and FedFed [30] applies VAE-based feature distillation. While effective, these methods often introduce significant computational and privacy costs due to data synthesis and sharing. In contrast, FedXDS shares compact" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 326, + 69, + 545, + 242 + ], + "blocks": [ + { + "bbox": [ + 326, + 69, + 545, + 242 + ], + "lines": [ + { + "bbox": [ + 326, + 69, + 545, + 242 + ], + "spans": [ + { + "bbox": [ + 326, + 69, + 545, + 242 + ], + "type": "image", + "image_path": "c6b902572e5b77296570c376eb8544f36a1845179476bbbb780ddf654482818a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 251, + 555, + 307 + ], + "lines": [ + { + "bbox": [ + 313, + 251, + 555, + 307 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 555, + 307 + ], + "type": "text", + "content": "Figure 1. Illustration of the DP attribution-guided feature extraction. The process consists of four main steps: (1) Computing attribution scores using a warmup model to identify important features, (2) Creating and applying a binary mask based on the attributions, (3) Adding Gaussian noise for metric privacy." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 319, + 554, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 319, + 554, + 341 + ], + "spans": [ + { + "bbox": [ + 315, + 319, + 554, + 341 + ], + "type": "text", + "content": "Algorithm 1 Attribution-Guided Private Representation Extraction." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 344, + 554, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 344, + 554, + 368 + ], + "spans": [ + { + "bbox": [ + 315, + 344, + 554, + 368 + ], + "type": "text", + "content": "Require: " + }, + { + "bbox": [ + 315, + 344, + 554, + 368 + ], + "type": "inline_equation", + "content": "D_{k}" + }, + { + "bbox": [ + 315, + 344, + 554, + 368 + ], + "type": "text", + "content": " : client dataset, " + }, + { + "bbox": [ + 315, + 344, + 554, + 368 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 315, + 344, + 554, + 368 + ], + "type": "text", + "content": " : model, " + }, + { + "bbox": [ + 315, + 344, + 554, + 368 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 315, + 344, + 554, + 368 + ], + "type": "text", + "content": " : sparsity level, " + }, + { + "bbox": [ + 315, + 344, + 554, + 368 + ], + "type": "inline_equation", + "content": "(\\varepsilon ,\\delta)" + }, + { + "bbox": [ + 315, + 344, + 554, + 368 + ], + "type": "text", + "content": " : privacy params, " + }, + { + "bbox": [ + 315, + 344, + 554, + 368 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 315, + 344, + 554, + 368 + ], + "type": "text", + "content": " : sensitivity, " + }, + { + "bbox": [ + 315, + 344, + 554, + 368 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{warmup}}" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 368, + 430, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 368, + 430, + 380 + ], + "spans": [ + { + "bbox": [ + 315, + 368, + 430, + 380 + ], + "type": "text", + "content": "Ensure: Private dataset " + }, + { + "bbox": [ + 315, + 368, + 430, + 380 + ], + "type": "inline_equation", + "content": "D_k^p" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 380, + 553, + 467 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 321, + 380, + 399, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 380, + 399, + 392 + ], + "spans": [ + { + "bbox": [ + 321, + 380, + 399, + 392 + ], + "type": "text", + "content": "1: for " + }, + { + "bbox": [ + 321, + 380, + 399, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^k\\in D_k" + }, + { + "bbox": [ + 321, + 380, + 399, + 392 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 392, + 528, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 392, + 528, + 404 + ], + "spans": [ + { + "bbox": [ + 321, + 392, + 528, + 404 + ], + "type": "text", + "content": "2: " + }, + { + "bbox": [ + 321, + 392, + 528, + 404 + ], + "type": "inline_equation", + "content": "\\mathbf{h} \\gets \\mathcal{A}(f_{\\theta_{\\mathrm{warmup}}}, \\mathbf{x}_i^k)" + }, + { + "bbox": [ + 321, + 392, + 528, + 404 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 321, + 392, + 528, + 404 + ], + "type": "inline_equation", + "content": "h^{(s)} \\gets \\mathrm{s}" + }, + { + "bbox": [ + 321, + 392, + 528, + 404 + ], + "type": "text", + "content": "-th largest in " + }, + { + "bbox": [ + 321, + 392, + 528, + 404 + ], + "type": "inline_equation", + "content": "\\mathbf{h}" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 405, + 492, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 405, + 492, + 418 + ], + "spans": [ + { + "bbox": [ + 321, + 405, + 492, + 418 + ], + "type": "text", + "content": "3: " + }, + { + "bbox": [ + 321, + 405, + 492, + 418 + ], + "type": "inline_equation", + "content": "\\mathbf{m} \\gets [\\mathbf{h}[i] \\geq h^{(s)}]; \\mathbf{x}_m \\gets \\mathbf{x}_i^k \\odot \\mathbf{m}" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 418, + 553, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 418, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 321, + 418, + 553, + 430 + ], + "type": "text", + "content": "4: " + }, + { + "bbox": [ + 321, + 418, + 553, + 430 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_i^k\\gets \\mathbf{x}_m + \\mathcal{N}(0,\\sigma^2)" + }, + { + "bbox": [ + 321, + 418, + 553, + 430 + ], + "type": "inline_equation", + "content": "\\triangleright \\sigma" + }, + { + "bbox": [ + 321, + 418, + 553, + 430 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 321, + 418, + 553, + 430 + ], + "type": "inline_equation", + "content": "(\\varepsilon ,\\delta)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 430, + 428, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 430, + 428, + 442 + ], + "spans": [ + { + "bbox": [ + 321, + 430, + 428, + 442 + ], + "type": "text", + "content": "5: Add " + }, + { + "bbox": [ + 321, + 430, + 428, + 442 + ], + "type": "inline_equation", + "content": "(\\tilde{\\mathbf{x}}_i^k,y_i^k)" + }, + { + "bbox": [ + 321, + 430, + 428, + 442 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 321, + 430, + 428, + 442 + ], + "type": "inline_equation", + "content": "D_k^p" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 442, + 364, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 442, + 364, + 452 + ], + "spans": [ + { + "bbox": [ + 321, + 442, + 364, + 452 + ], + "type": "text", + "content": "6: end for" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 453, + 377, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 453, + 377, + 467 + ], + "spans": [ + { + "bbox": [ + 321, + 453, + 377, + 467 + ], + "type": "text", + "content": "7: return " + }, + { + "bbox": [ + 321, + 453, + 377, + 467 + ], + "type": "inline_equation", + "content": "D_k^p" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 490, + 554, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 490, + 554, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 490, + 554, + 538 + ], + "type": "text", + "content": "attribution-based feature subsets derived via XAI, avoiding generators while enhancing privacy, efficiency, and performance in heterogeneous settings. Wo provide an extensive related works in the supplementary material." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 549, + 426, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 549, + 426, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 549, + 426, + 563 + ], + "type": "text", + "content": "4. FedXDS Approach" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 569, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 554, + 628 + ], + "type": "text", + "content": "Consider a federated learning system with " + }, + { + "bbox": [ + 313, + 569, + 554, + 628 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 569, + 554, + 628 + ], + "type": "text", + "content": " clients, where each client " + }, + { + "bbox": [ + 313, + 569, + 554, + 628 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 569, + 554, + 628 + ], + "type": "text", + "content": " holds local data " + }, + { + "bbox": [ + 313, + 569, + 554, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_k" + }, + { + "bbox": [ + 313, + 569, + 554, + 628 + ], + "type": "text", + "content": ". In our setting all clients share the same neural network architecture " + }, + { + "bbox": [ + 313, + 569, + 554, + 628 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 313, + 569, + 554, + 628 + ], + "type": "text", + "content": ". Our goal is to learn a global model " + }, + { + "bbox": [ + 313, + 569, + 554, + 628 + ], + "type": "inline_equation", + "content": "\\Theta_G" + }, + { + "bbox": [ + 313, + 569, + 554, + 628 + ], + "type": "text", + "content": " which generalizes well on all client datasets." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 636, + 512, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 636, + 512, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 636, + 512, + 647 + ], + "type": "text", + "content": "4.1. Attribution-Guided Feature Selection" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": "The first stage of our method identifies discriminative features using a pre-trained warmup model, with parameters " + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{warmup}}" + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": ". This is achieved by training with FedAvg for " + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "inline_equation", + "content": "R_{\\mathrm{warmup}}" + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": " rounds, before sharing representations. Then, for each input sample " + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": " are the" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4574" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 56, + 87, + 298, + 325 + ], + "blocks": [ + { + "bbox": [ + 56, + 72, + 267, + 84 + ], + "lines": [ + { + "bbox": [ + 56, + 72, + 267, + 84 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 267, + 84 + ], + "type": "text", + "content": "Algorithm 2 Privacy-Preserving Federated Training" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "lines": [ + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "spans": [ + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "text", + "content": "Require: " + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "inline_equation", + "content": "\\{D_k\\}_{k=1}^K" + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "text", + "content": " : client datasets, " + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "text", + "content": ": knowledge weight, " + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "text", + "content": ": rounds, " + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "text", + "content": ": local epochs, " + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "text", + "content": ": learning rate, " + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "inline_equation", + "content": "R_{\\text{warmup}}" + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "text", + "content": " \n1: Initialize " + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "inline_equation", + "content": "\\theta_{\\text{warmup}}" + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "inline_equation", + "content": "R_{\\text{warmup}}" + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "text", + "content": " rounds of FedAvg \n2: Obtain " + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "inline_equation", + "content": "D_k^p" + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "text", + "content": " using Algorithm 1 for all " + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "inline_equation", + "content": "k \\in [K]" + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "text", + "content": " in parallel \n3: Server: " + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_g \\gets \\bigcup_{k=1}^K D_k^p" + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "text", + "content": " \n4: for round " + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "text", + "content": " do \n5: Select client subset " + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "inline_equation", + "content": "S_t \\subseteq [K]" + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "text", + "content": " \n6: for " + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "inline_equation", + "content": "k \\in S_t" + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "text", + "content": " in parallel do \n7: " + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "inline_equation", + "content": "\\theta_k^t \\gets \\theta^{t-1}" + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "text", + "content": " Download global model \n8: for epoch " + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "inline_equation", + "content": "e = 1" + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "text", + "content": " do \n9: for batch " + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "inline_equation", + "content": "D_k" + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "inline_equation", + "content": "B_g" + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_g" + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "text", + "content": " do \n10: " + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_k^t \\gets \\frac{1}{|B_k|} \\sum_{(\\mathbf{x}, y) \\in B_k} \\ell(f_{\\theta_k^t}(\\mathbf{x}), y) + \\lambda \\frac{1}{|B_g|} \\sum_{(\\mathbf{x}_g, y_g) \\in B_g} \\ell(f_{\\theta_k^t}(\\mathbf{x}_g), y_g)" + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "text", + "content": " \n11: " + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "inline_equation", + "content": "\\theta_k^t \\gets \\theta_k^t - \\eta \\nabla \\mathcal{L}" + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "text", + "content": " \n12: end for \n13: end for \n14: end for \n15: Server: " + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "inline_equation", + "content": "\\theta^t \\gets \\frac{1}{|S_t|} \\sum_{k \\in S_t} \\theta_k^t" + }, + { + "bbox": [ + 56, + 87, + 298, + 325 + ], + "type": "text", + "content": " >FedAvg \n16: end for" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 55, + 345, + 296, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 345, + 296, + 381 + ], + "spans": [ + { + "bbox": [ + 55, + 345, + 296, + 381 + ], + "type": "text", + "content": "dimensions of the image with three color channels, in the client dataset " + }, + { + "bbox": [ + 55, + 345, + 296, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_k" + }, + { + "bbox": [ + 55, + 345, + 296, + 381 + ], + "type": "text", + "content": ", we compute relevances using an attribution method " + }, + { + "bbox": [ + 55, + 345, + 296, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 55, + 345, + 296, + 381 + ], + "type": "text", + "content": " from Section 2.3:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 389, + 294, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 389, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 146, + 389, + 294, + 403 + ], + "type": "interline_equation", + "content": "\\mathbf {h} = \\mathcal {A} \\left(f _ {\\theta}, \\mathbf {x}\\right) \\tag {7}", + "image_path": "bcd1d17a155c6dfa6f4403e61d97ae443fd39dcbe923cc6ecadedc53017626a4.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 410, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 296, + 495 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 410, + 296, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{h} \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 55, + 410, + 296, + 495 + ], + "type": "text", + "content": " is the pixel-wise importance score of an input image " + }, + { + "bbox": [ + 55, + 410, + 296, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 410, + 296, + 495 + ], + "type": "text", + "content": ". These attribution scores quantify the importance of each input feature to the model's prediction. To focus on the most relevant features, we create a binary mask " + }, + { + "bbox": [ + 55, + 410, + 296, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 55, + 410, + 296, + 495 + ], + "type": "text", + "content": ", as shown in Equation 8, by retaining features whose attribution scores exceed a threshold determined by the desired sparsity level " + }, + { + "bbox": [ + 55, + 410, + 296, + 495 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 410, + 296, + 495 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 118, + 502, + 295, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 502, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 118, + 502, + 295, + 534 + ], + "type": "interline_equation", + "content": "\\mathbf {m} [ i ] = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\mathbf {h} [ i ] \\geq h ^ {(s)} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {8}", + "image_path": "e046530391b1b2f14632f96964f95dee82e1db2600bc4d9040bbc755c8009b42.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 540, + 295, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 295, + 576 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 295, + 576 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 540, + 295, + 576 + ], + "type": "inline_equation", + "content": "h^{(s)}" + }, + { + "bbox": [ + 55, + 540, + 295, + 576 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 55, + 540, + 295, + 576 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 540, + 295, + 576 + ], + "type": "text", + "content": "-th largest value in the attribution scores. This mask identifies the subset of features that are most crucial for the model's decision-making process." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 577, + 296, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 577, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 577, + 296, + 613 + ], + "type": "text", + "content": "We define our feature selection function " + }, + { + "bbox": [ + 55, + 577, + 296, + 613 + ], + "type": "inline_equation", + "content": "f_{A} \\colon \\mathbb{R}^{H \\times W \\times 3} \\to \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 577, + 296, + 613 + ], + "type": "text", + "content": " as the application of the attribution mask " + }, + { + "bbox": [ + 55, + 577, + 296, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 55, + 577, + 296, + 613 + ], + "type": "text", + "content": " to each image " + }, + { + "bbox": [ + 55, + 577, + 296, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathcal{D}_k" + }, + { + "bbox": [ + 55, + 577, + 296, + 613 + ], + "type": "text", + "content": " through:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 140, + 621, + 295, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 621, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 140, + 621, + 295, + 634 + ], + "type": "interline_equation", + "content": "f _ {A} (\\mathbf {x}) = \\mathbf {x} \\odot \\mathbf {m} \\tag {9}", + "image_path": "4926a79dc7d1d9461ae6c6ab22fa7de6e4892d82269a09acf51b74c7fd17ec9c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "This operation emphasizes the most informative parts of the input while suppressing less relevant details. However, directly sharing " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "f_{A}(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " would still pose privacy risks. Therefore, we design a privacy mechanism " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " that adds calibrated noise to these masked features (Section refsubsec:privacy):" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 373, + 83, + 553, + 97 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 83, + 553, + 97 + ], + "spans": [ + { + "bbox": [ + 373, + 83, + 553, + 97 + ], + "type": "interline_equation", + "content": "\\mathcal {M} (\\mathbf {x}) = f _ {A} (\\mathbf {x}) + \\mathcal {N} (0, \\sigma^ {2} \\mathbf {I}) \\tag {10}", + "image_path": "7a1515fbe9be19b4c0bbbec1fcb67501e55d39865f97a7760e272ef04cc10cfe.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 99, + 555, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 99, + 555, + 171 + ], + "spans": [ + { + "bbox": [ + 313, + 99, + 555, + 171 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 99, + 555, + 171 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 99, + 555, + 171 + ], + "type": "text", + "content": " is determined based on privacy requirements. The resulting privacy-protected features, along with their corresponding labels, are then aggregated into a new dataset " + }, + { + "bbox": [ + 313, + 99, + 555, + 171 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_k^p = \\{(\\mathcal{M}(\\mathbf{x}),y)|(x,y)\\in \\mathcal{D}_k\\}" + }, + { + "bbox": [ + 313, + 99, + 555, + 171 + ], + "type": "text", + "content": " and shared with the server. The full data generated process is outlined in Algorithm 1 and visualized in Figure 1." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 177, + 495, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 177, + 495, + 190 + ], + "spans": [ + { + "bbox": [ + 313, + 177, + 495, + 190 + ], + "type": "text", + "content": "4.2. Local Training using Shared Data" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 194, + 555, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 555, + 266 + ], + "type": "text", + "content": "After the privacy-protected features have been obtained, the server aggregates private representations from all clients into a global dataset " + }, + { + "bbox": [ + 313, + 194, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_g = \\bigcup_{k=1}^N \\mathcal{D}_k^p" + }, + { + "bbox": [ + 313, + 194, + 555, + 266 + ], + "type": "text", + "content": ", and sends this dataset to each client. The clients then optimize a composite objective that balances local task performance with knowledge from the global private dataset:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 367, + 270, + 553, + 305 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 270, + 553, + 305 + ], + "spans": [ + { + "bbox": [ + 367, + 270, + 553, + 305 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {\\theta} \\left[ \\mathbb {E} _ {(\\mathbf {x}, y) \\sim \\mathcal {D} _ {k}} [ \\ell (f _ {\\theta} (\\mathbf {x}), y) ] \\right. \\tag {11} \\\\ + \\lambda \\mathbb {E} _ {(\\mathbf {x}, y) \\sim \\mathcal {D} _ {g}} [ \\ell (f _ {\\theta} (\\mathbf {x}), y) ] ] \\\\ \\end{array}", + "image_path": "78c85b2960cc856beee6361289381a1874f34d2248c2699161b2fba89a54fcee.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 308, + 554, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 554, + 368 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 554, + 368 + ], + "type": "text", + "content": "The hyperparameter " + }, + { + "bbox": [ + 313, + 308, + 554, + 368 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 308, + 554, + 368 + ], + "type": "text", + "content": " controls the trade-off between local specialization and global knowledge integration. This formulation allows clients to benefit from collective learning while maintaining their own task-specific performance and ensuring privacy. The procedure is outlined in Algorithm 2." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 373, + 425, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 425, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 425, + 385 + ], + "type": "text", + "content": "4.3. Privacy Discussion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 390, + 554, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 554, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 554, + 486 + ], + "type": "text", + "content": "We build on the insight that sharing features derived from raw client data can mitigate data heterogeneity in federated learning [34]. Nevertheless, directly sharing these features poses significant privacy risks. A straightforward approach would be to apply differential privacy (DP) to the raw data itself; however, the high dimensionality of image data forces the addition of very large noise levels, thereby significantly degrading utility." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 487, + 554, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 554, + 557 + ], + "type": "text", + "content": "To address this challenge, we propose a two-step approach. First, we use an attribution-guided masking strategy (Section 4.1) to identify and retain only task-relevant pixels. Second, we add Gaussian noise (Theorem 1) to the resulting masked features to achieve the desired " + }, + { + "bbox": [ + 313, + 487, + 554, + 557 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 313, + 487, + 554, + 557 + ], + "type": "text", + "content": "-metric privacy (Definition 1, Definition 2)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 558, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 554, + 677 + ], + "type": "text", + "content": "This design offers two main advantages. First, by sparsifying input features that do not contribute to classification predictions, we reduce the overall sensitivity of the data. According to Theorem 1, lower sensitivity requires less noise to meet the same privacy budget " + }, + { + "bbox": [ + 313, + 558, + 554, + 677 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 313, + 558, + 554, + 677 + ], + "type": "text", + "content": ", thus improving utility. Second, in contrast to random or magnitude-based sparsification that may unintentionally discard valuable information, our attribution-guided approach specifically preserves the pixels most critical for the learning task while eliminating irrelevant ones." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "The primary goal of our privacy mechanism is to protect against membership inference attacks (MIA) and feature inversion, which we empirically validate in section 6." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4575" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 109 + ], + "type": "text", + "content": "In the following section, we formally demonstrate how our strategy guarantees metric privacy, while a comprehensive discussion of privacy assumptions appears in the appendix." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 114, + 295, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 114, + 295, + 126 + ], + "spans": [ + { + "bbox": [ + 55, + 114, + 295, + 126 + ], + "type": "text", + "content": "4.3.1. Sensitivity Analysis of Attribution-Based Masking" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 129, + 295, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 129, + 295, + 200 + ], + "spans": [ + { + "bbox": [ + 55, + 129, + 295, + 200 + ], + "type": "text", + "content": "A critical component of ensuring privacy is understanding the sensitivity of our feature selection function " + }, + { + "bbox": [ + 55, + 129, + 295, + 200 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 129, + 295, + 200 + ], + "type": "text", + "content": " defined in Equation 9. From Theorem 1, the noise required to attain a certain privacy level " + }, + { + "bbox": [ + 55, + 129, + 295, + 200 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 55, + 129, + 295, + 200 + ], + "type": "text", + "content": " depends on the sensitivity " + }, + { + "bbox": [ + 55, + 129, + 295, + 200 + ], + "type": "inline_equation", + "content": "\\Delta_f" + }, + { + "bbox": [ + 55, + 129, + 295, + 200 + ], + "type": "text", + "content": ". Our goal is to minimize this sensitivity before applying the privacy mechanism " + }, + { + "bbox": [ + 55, + 129, + 295, + 200 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 129, + 295, + 200 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 201, + 295, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 201, + 295, + 248 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 295, + 248 + ], + "type": "text", + "content": "For our feature selection function " + }, + { + "bbox": [ + 55, + 201, + 295, + 248 + ], + "type": "inline_equation", + "content": "f_{A}(\\mathbf{x}) = \\mathbf{x}\\odot \\mathbf{m}" + }, + { + "bbox": [ + 55, + 201, + 295, + 248 + ], + "type": "text", + "content": ", we can derive a strict bound on sensitivity. Since each coordinate of the mask satisfies " + }, + { + "bbox": [ + 55, + 201, + 295, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{m}[i]\\in \\{0,1\\}" + }, + { + "bbox": [ + 55, + 201, + 295, + 248 + ], + "type": "text", + "content": ", for any two inputs " + }, + { + "bbox": [ + 55, + 201, + 295, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{x},\\mathbf{x}'\\in \\mathbb{R}^{H\\times W\\times 3}" + }, + { + "bbox": [ + 55, + 201, + 295, + 248 + ], + "type": "text", + "content": ", we have, in the worst case:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 97, + 256, + 253, + 272 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 256, + 253, + 272 + ], + "spans": [ + { + "bbox": [ + 97, + 256, + 253, + 272 + ], + "type": "interline_equation", + "content": "\\left| \\mathbf {m} [ i ] \\mathbf {x} [ i ] - \\mathbf {m} [ i ] \\mathbf {x} ^ {\\prime} [ i ] \\right| \\leq \\left| \\mathbf {x} [ i ] - \\mathbf {x} ^ {\\prime} [ i ] \\right|", + "image_path": "63bf567db58ae119ee40afeb78a7934092f1c191054452f6713b04d030d792dd.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 279, + 148, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 279, + 148, + 291 + ], + "spans": [ + { + "bbox": [ + 67, + 279, + 148, + 291 + ], + "type": "text", + "content": "Thus, in aggregate:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 300, + 244, + 313 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 300, + 244, + 313 + ], + "spans": [ + { + "bbox": [ + 107, + 300, + 244, + 313 + ], + "type": "interline_equation", + "content": "\\left\\| f _ {A} (\\mathbf {x}) - f _ {A} \\left(\\mathbf {x} ^ {\\prime}\\right) \\right\\| _ {2} \\leq \\left\\| \\mathbf {x} - \\mathbf {x} ^ {\\prime} \\right\\| _ {2}", + "image_path": "8ad5bfde8d845a75f5664701522dc90f5f6211bf0180602fc47bcef037a92eac.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 322, + 295, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 322, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 322, + 295, + 346 + ], + "type": "text", + "content": "This shows that our masking operation is non-expansive with respect to the " + }, + { + "bbox": [ + 55, + 322, + 295, + 346 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 55, + 322, + 295, + 346 + ], + "type": "text", + "content": " norm. Rearranging the equation:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 110, + 353, + 295, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 353, + 295, + 380 + ], + "spans": [ + { + "bbox": [ + 110, + 353, + 295, + 380 + ], + "type": "interline_equation", + "content": "\\Delta_ {f} = \\frac {\\left\\| f _ {A} (\\mathbf {x}) - f _ {A} \\left(\\mathbf {x} ^ {\\prime}\\right) \\right\\| _ {2}}{\\left\\| \\mathbf {x} - \\mathbf {x} ^ {\\prime} \\right\\| _ {2}} \\leq 1 \\tag {12}", + "image_path": "e9dc6a0e2e17fbe8890177b600ed65dec6fdfe09161ecdd5401861abe7a2eb4e.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 387, + 295, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 387, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 295, + 518 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 387, + 295, + 518 + ], + "type": "inline_equation", + "content": "\\Delta_f" + }, + { + "bbox": [ + 55, + 387, + 295, + 518 + ], + "type": "text", + "content": " is the sensitivity we defined in the preliminaries in Definition 2. Without our attribution-based masking, the sensitivity would be unbounded for arbitrary transformations. Moreover, in the worst case " + }, + { + "bbox": [ + 55, + 387, + 295, + 518 + ], + "type": "inline_equation", + "content": "f_A" + }, + { + "bbox": [ + 55, + 387, + 295, + 518 + ], + "type": "text", + "content": " is the identity, i.e. " + }, + { + "bbox": [ + 55, + 387, + 295, + 518 + ], + "type": "inline_equation", + "content": "f_A(x) = x" + }, + { + "bbox": [ + 55, + 387, + 295, + 518 + ], + "type": "text", + "content": ", meaning that the attribution method deems all pixels important. This implies that our method would reduce to the naive approach of adding noise to the raw features. In practice, this rarely occurs, since this would already require a highly sparsified image. Nevertheless, our approach safeguards against this worst-case scenario. More discussion can be found in the supplementary material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 525, + 239, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 239, + 536 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 239, + 536 + ], + "type": "text", + "content": "4.3.2. Privacy Mechanism and Guarantees" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 539, + 295, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 295, + 575 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 295, + 575 + ], + "type": "text", + "content": "Based on the sensitivity bound established above, we design our privacy mechanism " + }, + { + "bbox": [ + 55, + 539, + 295, + 575 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 539, + 295, + 575 + ], + "type": "text", + "content": " (Equation 10) by adding Gaussian noise calibrated to this reduced sensitivity:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 583, + 236, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 583, + 236, + 597 + ], + "spans": [ + { + "bbox": [ + 115, + 583, + 236, + 597 + ], + "type": "interline_equation", + "content": "\\mathcal {M} (\\mathbf {x}) = f _ {A} (\\mathbf {x}) + \\mathcal {N} (0, \\sigma^ {2} \\mathbf {I})", + "image_path": "02d966c24c12b03c88a9e2a0bba6a3633b6c9cc38986a5be8c7c1e059089f11f.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 605, + 295, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 654 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 605, + 295, + 654 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 605, + 295, + 654 + ], + "type": "text", + "content": " is chosen according to Theorem 1 for an " + }, + { + "bbox": [ + 55, + 605, + 295, + 654 + ], + "type": "inline_equation", + "content": "(\\varepsilon, \\delta)" + }, + { + "bbox": [ + 55, + 605, + 295, + 654 + ], + "type": "text", + "content": "-level of privacy given sensitivity " + }, + { + "bbox": [ + 55, + 605, + 295, + 654 + ], + "type": "inline_equation", + "content": "\\Delta_f \\leq 1" + }, + { + "bbox": [ + 55, + 605, + 295, + 654 + ], + "type": "text", + "content": ". This ensures that for all input images in the client dataset, " + }, + { + "bbox": [ + 55, + 605, + 295, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{x}, \\mathbf{x}' \\in D_k" + }, + { + "bbox": [ + 55, + 605, + 295, + 654 + ], + "type": "text", + "content": ", our method satisfies " + }, + { + "bbox": [ + 55, + 605, + 295, + 654 + ], + "type": "inline_equation", + "content": "(\\varepsilon, \\delta)" + }, + { + "bbox": [ + 55, + 605, + 295, + 654 + ], + "type": "text", + "content": "-metric privacy." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "text", + "content": "Our theoretical privacy guarantees are empirically validated through membership inference and feature inversion attacks Section 6, where attribution-masked features consistently show stronger protection than unmasked features under equivalent privacy budgets." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 71, + 400, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 400, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 400, + 84 + ], + "type": "text", + "content": "5. Experimental" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 91, + 430, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 91, + 430, + 104 + ], + "spans": [ + { + "bbox": [ + 314, + 91, + 430, + 104 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 109, + 555, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 555, + 312 + ], + "type": "text", + "content": "We evaluate FedXDS on standard image classification benchmarks (CIFAR-10, CIFAR-100, Tiny-ImageNet) and real-world federated datasets from the LEAF framework [5] (CelebA, FEMNIST) that naturally exhibit heterogeneous distributions. For the standard benchmarks, we simulate statistical heterogeneity using Dirichlet-based partitioning with concentration parameters " + }, + { + "bbox": [ + 313, + 109, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\alpha \\in \\{0.05, 0.1\\}" + }, + { + "bbox": [ + 313, + 109, + 555, + 312 + ], + "type": "text", + "content": ", where smaller values indicate greater heterogeneity. We conduct experiments with both 10 and 100 clients, setting participation rates to 0.5 and 0.1, respectively, and train for " + }, + { + "bbox": [ + 313, + 109, + 555, + 312 + ], + "type": "inline_equation", + "content": "R = 200" + }, + { + "bbox": [ + 313, + 109, + 555, + 312 + ], + "type": "text", + "content": " communication rounds. To compute relevance-based attributions, we evaluate four attribution methods in our FL framework: Layerwise Relevance Propagation (FedXLRP) [3], Integrated Gradients (FedXIG) [28], SmoothGrad (FedXSG) [27], and Input " + }, + { + "bbox": [ + 313, + 109, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 313, + 109, + 555, + 312 + ], + "type": "text", + "content": " Gradient (FedXGrad). Unless otherwise state we use " + }, + { + "bbox": [ + 313, + 109, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\varepsilon = 20" + }, + { + "bbox": [ + 313, + 109, + 555, + 312 + ], + "type": "text", + "content": " sparsity level " + }, + { + "bbox": [ + 313, + 109, + 555, + 312 + ], + "type": "inline_equation", + "content": "s = 70\\%" + }, + { + "bbox": [ + 313, + 109, + 555, + 312 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 109, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\lambda = 0.5" + }, + { + "bbox": [ + 313, + 109, + 555, + 312 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 321, + 380, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 321, + 380, + 333 + ], + "spans": [ + { + "bbox": [ + 314, + 321, + 380, + 333 + ], + "type": "text", + "content": "5.2. Baselines" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 339, + 555, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 555, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 555, + 591 + ], + "type": "text", + "content": "The experimental results in Table 1 show that FedXLRP consistently outperforms all baselines across varying client numbers and data distributions. This advantage amplifies in more challenging scenarios (K=100, " + }, + { + "bbox": [ + 313, + 339, + 555, + 591 + ], + "type": "inline_equation", + "content": "\\alpha = 0.05" + }, + { + "bbox": [ + 313, + 339, + 555, + 591 + ], + "type": "text", + "content": "). FedFed ranks second, while FedAvg and FedAux degrade significantly with increased client numbers. On CIFAR-100, performance gaps narrow, though FedXLRP maintains its edge, particularly with higher client counts and heterogeneity. Other attribution methods (FedXIG, FedXGrad, FedXSG) perform well but remain below FedXLRP, highlighting attribution method choice importance. We discuss this point in Subsection 7.1. In addition, communication efficiency results Table 2 reinforce FedXLRP's advantages. To reach " + }, + { + "bbox": [ + 313, + 339, + 555, + 591 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 313, + 339, + 555, + 591 + ], + "type": "text", + "content": " accuracy on CIFAR-10 with 10 clients, FedXLRP requires only 14 rounds versus 49 for FedAvg and 28 for FedFTG. With 100 clients, FedXLRP needs 23 rounds to reach " + }, + { + "bbox": [ + 313, + 339, + 555, + 591 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 313, + 339, + 555, + 591 + ], + "type": "text", + "content": " accuracy compared to FedAvg's 79. CIFAR-100 shows similar patterns, with FedXLRP achieving target accuracies in substantially fewer rounds across client configurations. Among baselines, only FedFed shows comparable efficiency." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 599, + 511, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 599, + 511, + 612 + ], + "spans": [ + { + "bbox": [ + 314, + 599, + 511, + 612 + ], + "type": "text", + "content": "5.3. Experiments on Real-World Datasets" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "content": "We also evaluate FedXDS on real-world datasets using implementations from the LEAF [5] and [7] frameworks. Specifically, we consider CelebA and FEMNIST, which exhibit a distinct form of non-IIIDness known as distribution shift, introducing an additional challenge for federated learning. For CelebA, we follow the implementation of [36], while for FEMNIST, we use the setup from [7], conducting experiments with 10 and 100 clients, respectively." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "type": "text", + "content": "4576" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 112, + 555, + 283 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": "Table 1. Performance comparison across different federated learning methods, including standard deviations from 5 runs. Results show the top-1 test accuracy (%) on CIFAR-10, CIFAR-100, and Tiny-ImageNet under varying numbers of clients (K) and Dirichlet concentration parameters " + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "inline_equation", + "content": "(\\alpha)" + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": ". Bold values indicate the best performance in each column." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 112, + 555, + 283 + ], + "lines": [ + { + "bbox": [ + 56, + 112, + 555, + 283 + ], + "spans": [ + { + "bbox": [ + 56, + 112, + 555, + 283 + ], + "type": "table", + "html": "
DatasetCIFAR-10CIFAR-100Tiny-ImageNet
K=10K=100K=10K=100K=10K=100
α = 0.05α = 0.1α = 0.05α = 0.1α = 0.05α = 0.1α = 0.05α = 0.1α = 0.05α = 0.1α = 0.05α = 0.1
FedAvg64.16 ± 0.7172.90 ± 0.5555.53 ± 0.8260.94 ± 0.6843.82 ± 0.4544.57 ± 0.5132.99 ± 0.7335.59 ± 0.6230.30 ± 0.5934.41 ± 0.4826.20 ± 0.8128.79 ± 0.75
FedProx71.56 ± 0.6574.31 ± 0.4951.89 ± 0.9165.01 ± 0.5344.33 ± 0.3848.67 ± 0.4737.79 ± 0.5538.43 ± 0.4933.27 ± 0.4234.54 ± 0.5124.15 ± 0.8830.23 ± 0.67
FedDyn70.62 ± 0.5878.25 ± 0.3463.09 ± 0.6169.41 ± 0.4548.53 ± 0.4149.42 ± 0.3937.14 ± 0.6841.60 ± 0.5433.03 ± 0.4835.69 ± 0.4129.52 ± 0.6332.37 ± 0.59
SCAFFOLD71.35 ± 0.6175.62 ± 0.4258.51 ± 0.7963.82 ± 0.6246.61 ± 0.3347.27 ± 0.4435.28 ± 0.2438.87 ± 0.5831.53 ± 0.5335.76 ± 0.3928.35 ± 0.5930.49 ± 0.61
FedSAM70.15 ± 0.8877.53 ± 0.5162.45 ± 0.7268.78 ± 0.5947.95 ± 0.3548.81 ± 0.4136.88 ± 0.5140.92 ± 0.4732.21 ± 0.5535.03 ± 0.4428.98 ± 0.6631.84 ± 0.52
FedDISCO69.73 ± 0.4476.98 ± 0.6361.82 ± 0.6868.11 ± 0.6547.31 ± 0.5248.15 ± 0.5836.17 ± 0.6740.25 ± 0.6131.78 ± 0.6134.68 ± 0.5228.51 ± 0.7131.33 ± 0.64
FedFed79.12 ± 0.4882.58 ± 0.3175.35 ± 0.5778.87 ± 0.4352.27 ± 0.4956.45 ± 0.3845.41 ± 0.7649.28 ± 0.5234.99 ± 0.4536.49 ± 0.3733.28 ± 0.5834.89 ± 0.47
FedFTG75.21 ± 0.5378.44 ± 0.3970.84 ± 0.3174.85 ± 0.4144.91 ± 0.6254.15 ± 0.4540.43 ± 0.3946.47 ± 0.4832.40 ± 0.5733.79 ± 0.4930.53 ± 0.5132.26 ± 0.55
FedGen68.34 ± 0.6874.65 ± 0.4758.90 ± 0.7563.20 ± 0.6644.29 ± 0.5550.12 ± 0.5136.70 ± 0.6239.82 ± 0.5734.17 ± 0.4436.33 ± 0.4229.45 ± 0.6831.60 ± 0.61
FedAux59.72 ± 0.8573.70 ± 0.5153.13 ± 0.8865.11 ± 0.5944.05 ± 0.4844.85 ± 0.5336.54 ± 0.6538.39 ± 0.6127.83 ± 0.7234.31 ± 0.4924.75 ± 0.8230.38 ± 0.69
FedDF60.56 ± 0.7772.40 ± 0.5859.20 ± 0.7263.47 ± 0.6430.13 ± 0.8145.47 ± 0.5034.46 ± 0.6936.77 ± 0.6429.07 ± 0.6534.75 ± 0.4628.42 ± 0.6230.51 ± 0.63
FedXLRP81.72 ± 0.3583.46 ± 0.2877.02 ± 0.6380.27 ± 0.3952.07 ± 0.4558.09 ± 0.3346.25 ± 0.8252.63 ± 0.4136.85 ± 0.3938.64 ± 0.3234.64 ± 0.5233.18 ± 0.51
FedXIG72.22 ± 0.5975.89 ± 0.4464.02 ± 0.6968.61 ± 0.5549.21 ± 0.4855.46 ± 0.4142.34 ± 0.5847.89 ± 0.5033.59 ± 0.5135.30 ± 0.4529.77 ± 0.6431.93 ± 0.58
FedXGrad71.38 ± 0.6273.99 ± 0.5264.52 ± 0.6566.49 ± 0.6150.13 ± 0.4354.21 ± 0.4642.11 ± 0.6046.98 ± 0.5333.18 ± 0.5434.40 ± 0.4830.00 ± 0.6130.93 ± 0.60
FedXSG71.89 ± 0.6074.63 ± 0.4963.75 ± 0.7167.80 ± 0.5851.57 ± 0.3955.22 ± 0.4240.91 ± 0.6347.61 ± 0.5133.43 ± 0.5234.74 ± 0.4729.64 ± 0.6531.53 ± 0.59
", + "image_path": "ce507b87acfb0e3a6f3350328feb0690d47b6f2cc4368d46f71aaedd597f4135.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 151, + 346, + 459, + 574 + ], + "blocks": [ + { + "bbox": [ + 55, + 293, + 555, + 339 + ], + "lines": [ + { + "bbox": [ + 55, + 293, + 555, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 555, + 339 + ], + "type": "text", + "content": "Table 2. Communication efficiency comparison across different federated learning methods. Results show the number of communication rounds (mean ± std over 5 runs) needed to achieve target accuracy thresholds (70% and 60% for CIFAR-10; 40% and 30% for CIFAR-100; 35% and 30% for Tiny-ImageNet) with different numbers of clients (K=10 and K=100) and heterogeneity parameter " + }, + { + "bbox": [ + 55, + 293, + 555, + 339 + ], + "type": "inline_equation", + "content": "\\alpha = 0.1" + }, + { + "bbox": [ + 55, + 293, + 555, + 339 + ], + "type": "text", + "content": ". Lower values indicate better communication efficiency. Best results are in bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 346, + 459, + 574 + ], + "lines": [ + { + "bbox": [ + 151, + 346, + 459, + 574 + ], + "spans": [ + { + "bbox": [ + 151, + 346, + 459, + 574 + ], + "type": "table", + "html": "
MethodCIFAR-10CIFAR-100Tiny-ImageNet
acc=70% K=10acc=60% K=100acc=40% K=10acc=30% K=100acc=35% K=10acc=30% K=100
FedAvg49 ± 579 ± 724 ± 434 ± 560 ± 690 ± 8
FedProx32 ± 355 ± 428 ± 432 ± 445 ± 465 ± 5
FedDyn29 ± 351 ± 424 ± 332 ± 348 ± 560 ± 4
SCAFFOLD33 ± 452 ± 525 ± 338 ± 446 ± 469 ± 6
FedSAM34 ± 458 ± 527 ± 435 ± 450 ± 564 ± 5
FedDISCO36 ± 361 ± 628 ± 437 ± 552 ± 567 ± 6
FedFed15 ± 222 ± 413 ± 412 ± 312 ± 215 ± 2
FedFTG28 ± 357 ± 422 ± 328 ± 330 ± 340 ± 4
FedGen51 ± 566 ± 621 ± 439 ± 555 ± 570 ± 6
FedAux38 ± 445 ± 419 ± 336 ± 442 ± 450 ± 5
FedDF84 ± 790 ± 828 ± 445 ± 5100 ± 8110 ± 9
FedXLRP14 ± 223 ± 311 ± 313 ± 210 ± 412 ± 2
FedXIG35 ± 439 ± 416 ± 221 ± 338 ± 445 ± 4
FedXGrad41 ± 446 ± 513 ± 226 ± 344 ± 450 ± 5
FedXSG41 ± 551 ± 421 ± 333 ± 442 ± 448 ± 5
", + "image_path": "411f05dea331c732f6ee5b49abfdb9637306957c1d35e573654ce456992e8621.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 594, + 294, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 294, + 703 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 294, + 703 + ], + "type": "text", + "content": "Additional implementation details can be found in the supplementary material. The results in Table 3 show a clear trend where FedXDS consistently achieves the highest accuracy across both datasets, demonstrating its effectiveness in handling distribution shift. Notably, methods incorporating additional data generation or augmentation, such as FedFTG and FedGen, also perform well, but FedXDS surpasses them, suggesting that relevance-guided data sharing provides a stronger mechanism for improving generaliza" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 594, + 478, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 478, + 607 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 478, + 607 + ], + "type": "text", + "content": "tion in heterogeneous federated settings." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 618, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 555, + 715 + ], + "type": "text", + "content": "The results in Table 4 demonstrate that integrating FedXDS with FL methods consistently improves the performance. In all cases, the addition of FedXDS yields substantial accuracy gains, particularly in the highly heterogeneous setting with " + }, + { + "bbox": [ + 313, + 618, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{K} = 100" + }, + { + "bbox": [ + 313, + 618, + 555, + 715 + ], + "type": "text", + "content": " clients. For instance, FedAvg, which traditionally suffers from performance degradation in non-IID scenarios, improves from " + }, + { + "bbox": [ + 313, + 618, + 555, + 715 + ], + "type": "inline_equation", + "content": "60.94\\%" + }, + { + "bbox": [ + 313, + 618, + 555, + 715 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 618, + 555, + 715 + ], + "type": "inline_equation", + "content": "80.29\\%" + }, + { + "bbox": [ + 313, + 618, + 555, + 715 + ], + "type": "text", + "content": " when combined with FedXDS. Similar improvements are observed for" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "type": "text", + "content": "4577" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 85, + 113, + 265, + 274 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 103 + ], + "type": "text", + "content": "Table 3. Accuracy comparison (mean ± std % over 5 runs) on CelebA and FEMNIST. Best results are in bold. We use the LRP variant of FedXDS." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 85, + 113, + 265, + 274 + ], + "lines": [ + { + "bbox": [ + 85, + 113, + 265, + 274 + ], + "spans": [ + { + "bbox": [ + 85, + 113, + 265, + 274 + ], + "type": "table", + "html": "
MethodCelebAFEMNIST
FedAvg87.23 ± 0.7184.71 ± 0.68
FedProx87.79 ± 0.6585.66 ± 0.61
SCAFFOLD86.36 ± 0.5684.24 ± 0.59
FedDyn88.13 ± 0.5286.45 ± 0.34
FedSAM90.03 ± 0.1587.56 ± 0.44
FedDISCO89.78 ± 0.6187.29 ± 0.62
FedFTG90.31 ± 0.4387.88 ± 0.67
FedGen89.65 ± 0.5986.37 ± 0.55
FedDF88.56 ± 0.6385.28 ± 0.66
FedAux89.27 ± 0.6085.94 ± 0.62
FedFed90.76 ± 0.8388.28 ± 0.71
FedXDS91.55 ± 0.4889.03 ± 0.35
", + "image_path": "1e38c3edf27b809d1770507e727061b3eb22e57c571b7ee3ca7f2cb2817844ee.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 62, + 338, + 287, + 466 + ], + "blocks": [ + { + "bbox": [ + 55, + 284, + 296, + 329 + ], + "lines": [ + { + "bbox": [ + 55, + 284, + 296, + 329 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 296, + 329 + ], + "type": "text", + "content": "Table 4. Performance comparison (mean ± std % over 5 runs) on CIFAR-10 for different base FL methods and their FedXDS (using LRP) variants under two heterogeneity settings (" + }, + { + "bbox": [ + 55, + 284, + 296, + 329 + ], + "type": "inline_equation", + "content": "K = 10" + }, + { + "bbox": [ + 55, + 284, + 296, + 329 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 284, + 296, + 329 + ], + "type": "inline_equation", + "content": "K = 100" + }, + { + "bbox": [ + 55, + 284, + 296, + 329 + ], + "type": "text", + "content": "). The FedXDS variants consistently improve performance." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 338, + 287, + 466 + ], + "lines": [ + { + "bbox": [ + 62, + 338, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 62, + 338, + 287, + 466 + ], + "type": "table", + "html": "
MethodK=10K=100
FedAvg72.90 ± 0.5560.94 ± 0.68
FedAvg + FedXDS83.46 ± 0.3180.29 ± 0.42
FedProx74.31 ± 0.4965.01 ± 0.53
FedProx + FedXDS84.05 ± 0.3580.73 ± 0.45
FedDyn70.62 ± 0.5869.41 ± 0.45
FedDyn + FedXDS84.20 ± 0.3379.56 ± 0.48
SCAFFOLD71.35 ± 0.6163.82 ± 0.62
SCAFFOLD + FedXDS83.89 ± 0.3879.72 ± 0.44
", + "image_path": "bb123525ad79e91959dda314deb244513f045b98c23137bb88a65de51ef7ef95.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 485, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 295, + 521 + ], + "type": "text", + "content": "FedProx, FedDyn, and SCAFFOLD, with FedXDS enhancing their robustness to data heterogeneity while maintaining their inherent advantages." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 521, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 521, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 54, + 521, + 295, + 628 + ], + "type": "text", + "content": "The magnitude of improvement suggests that FedXDS effectively mitigates the adverse effects of statistical heterogeneity by facilitating better knowledge transfer across clients. Notably, FedDyn, which incorporates regularization-based local adaptation, exhibits the smallest absolute improvement with FedXDS, indicating that its existing adaptation mechanisms already partially address heterogeneity. In contrast, methods like FedAvg, without correction for client drift, benefit more from FedXDS." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 639, + 208, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 639, + 208, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 639, + 208, + 654 + ], + "type": "text", + "content": "6. Empirical Privacy Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 659, + 220, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 220, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 220, + 673 + ], + "type": "text", + "content": "6.1. Membership Inference Attack" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "We conduct membership inference attacks following [30] to evaluate our methodology's privacy guarantees. We investigate whether training on masked and noised features reveals" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 315, + 113, + 555, + 196 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 555, + 104 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 104 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 104 + ], + "type": "text", + "content": "Table 5. Test accuracy (mean ± std % over 5 runs) for different sparsification levels " + }, + { + "bbox": [ + 313, + 70, + 555, + 104 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 70, + 555, + 104 + ], + "type": "text", + "content": " across various attribution methods. Best results are in bold. Results for CIFAR10 at " + }, + { + "bbox": [ + 313, + 70, + 555, + 104 + ], + "type": "inline_equation", + "content": "K = 10, \\alpha = 0.1" + }, + { + "bbox": [ + 313, + 70, + 555, + 104 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 113, + 555, + 196 + ], + "lines": [ + { + "bbox": [ + 315, + 113, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 315, + 113, + 555, + 196 + ], + "type": "table", + "html": "
sGradientSmoothGradInt. Grad.LRP
6075.99 ± 0.5275.63 ± 0.5577.10 ± 0.4884.74 ± 0.31
6574.19 ± 0.5673.15 ± 0.6176.31 ± 0.5184.01 ± 0.35
7073.99 ± 0.5974.63 ± 0.5475.89 ± 0.4983.46 ± 0.28
7570.54 ± 0.6570.67 ± 0.6371.68 ± 0.5883.01 ± 0.33
8068.31 ± 0.7170.45 ± 0.6670.98 ± 0.6282.67 ± 0.36
8568.80 ± 0.6868.13 ± 0.7369.73 ± 0.6781.88 ± 0.41
", + "image_path": "da409e42317998b20580432056ba38051b6619f098a3beeaeb66a1d62344a9c3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 217, + 555, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 555, + 336 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 555, + 336 + ], + "type": "text", + "content": "membership in the underlying dataset by using the globally DP data to train a shadow model, then testing if clients can infer membership by querying this model and training a random forest classifier. Figure 2a shows recall values for different privacy parameters " + }, + { + "bbox": [ + 313, + 217, + 555, + 336 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 313, + 217, + 555, + 336 + ], + "type": "text", + "content": ". Results demonstrate that non-masked features, despite DP protection, offer inferior privacy protection compared to masked data. This can be explained by the fact that sparsity significantly mitigates attacks through norm reduction from attribution masks (Equation 8)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 345, + 454, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 345, + 454, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 454, + 357 + ], + "type": "text", + "content": "6.2. Feature Inversion Attack" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 363, + 556, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 556, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 556, + 544 + ], + "type": "text", + "content": "While metric private features prevent direct inference of individual data features, a sophisticated adversary might train a denoising autoencoder to learn mappings between protected representations and original images. We conduct feature inversion attacks and report the SSIM score between original and reconstructed images. Figure 2b reveals inversion attack success increases with lower sparsity and higher privacy budgets. Unmasked features consistently exhibit the highest SSIM, confirming that raw features under identical DP constraints provide inferior protection compared to our approach. Additional validations in supplementary material, including visualization of inversion results for non-masked features and FedFed [30] features, confirming both cases' vulnerability to reconstruction, whereas our method provides superior protection." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 555, + 494, + 568 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 494, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 494, + 568 + ], + "type": "text", + "content": "7. Ablation Studies and Discussion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 575, + 555, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 575, + 555, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 575, + 555, + 589 + ], + "type": "text", + "content": "7.1. XAI Method Comparison Across Sparsification" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "Table 5 demonstrates that FedXDS maintains stable performance across different sparsification levels, with all attribution methods achieving above " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "68\\%" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": " accuracy, even at high sparsity thresholds. Notably, LRP consistently outperforms other methods, preserving accuracies between " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "81.88\\%" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "84.74\\%" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": ", while others degrade more sharply under extreme sparsification. LRP's robustness stems from its layer-wise relevance propagation mechanism, which is fundamentally different from gradient-based approaches. Instead of measuring local sensitivity, LRP redistributes the" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4578" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 175, + 141 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 175, + 141 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 175, + 141 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 175, + 141 + ], + "type": "image", + "image_path": "d654f65426f1b09337378c9b8bb673358720c17eb2d5a29177b9ccba3e8b785e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 144, + 177, + 164 + ], + "lines": [ + { + "bbox": [ + 56, + 144, + 177, + 164 + ], + "spans": [ + { + "bbox": [ + 56, + 144, + 177, + 164 + ], + "type": "text", + "content": "(a) Recall score of MI attack at different " + }, + { + "bbox": [ + 56, + 144, + 177, + 164 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 56, + 144, + 177, + 164 + ], + "type": "text", + "content": " for various sparsity levels " + }, + { + "bbox": [ + 56, + 144, + 177, + 164 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 56, + 144, + 177, + 164 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 182, + 72, + 301, + 142 + ], + "blocks": [ + { + "bbox": [ + 182, + 72, + 301, + 142 + ], + "lines": [ + { + "bbox": [ + 182, + 72, + 301, + 142 + ], + "spans": [ + { + "bbox": [ + 182, + 72, + 301, + 142 + ], + "type": "image", + "image_path": "334fbfd003e818507f5a5d76ad2c6de6ee2b76487b76bccdd60b2ee70ed5646a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 182, + 144, + 303, + 164 + ], + "lines": [ + { + "bbox": [ + 182, + 144, + 303, + 164 + ], + "spans": [ + { + "bbox": [ + 182, + 144, + 303, + 164 + ], + "type": "text", + "content": "(b) SSIM score of denoising attack at different " + }, + { + "bbox": [ + 182, + 144, + 303, + 164 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 182, + 144, + 303, + 164 + ], + "type": "text", + "content": " and sparsity levels " + }, + { + "bbox": [ + 182, + 144, + 303, + 164 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 182, + 144, + 303, + 164 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 308, + 72, + 427, + 141 + ], + "blocks": [ + { + "bbox": [ + 308, + 72, + 427, + 141 + ], + "lines": [ + { + "bbox": [ + 308, + 72, + 427, + 141 + ], + "spans": [ + { + "bbox": [ + 308, + 72, + 427, + 141 + ], + "type": "image", + "image_path": "0507348d4aaec163abc51ebeba93ede0d6190e7357363e40077b79034915f19b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 307, + 144, + 428, + 164 + ], + "lines": [ + { + "bbox": [ + 307, + 144, + 428, + 164 + ], + "spans": [ + { + "bbox": [ + 307, + 144, + 428, + 164 + ], + "type": "text", + "content": "(c) Effect of knowledge weight " + }, + { + "bbox": [ + 307, + 144, + 428, + 164 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 307, + 144, + 428, + 164 + ], + "type": "text", + "content": " on model accuracy across datasets." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 433, + 72, + 552, + 142 + ], + "blocks": [ + { + "bbox": [ + 433, + 72, + 552, + 142 + ], + "lines": [ + { + "bbox": [ + 433, + 72, + 552, + 142 + ], + "spans": [ + { + "bbox": [ + 433, + 72, + 552, + 142 + ], + "type": "image", + "image_path": "7c0e6832821a718b23e96a981dcd2488e770e4ce2890d8fc52319f8cc3cb09ee.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 433, + 144, + 553, + 164 + ], + "lines": [ + { + "bbox": [ + 433, + 144, + 553, + 164 + ], + "spans": [ + { + "bbox": [ + 433, + 144, + 553, + 164 + ], + "type": "text", + "content": "(d) Impact of warmup rounds on model accuracy across datasets." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 174, + 476, + 186 + ], + "lines": [ + { + "bbox": [ + 132, + 174, + 476, + 186 + ], + "spans": [ + { + "bbox": [ + 132, + 174, + 476, + 186 + ], + "type": "text", + "content": "Figure 2. Evaluation of privacy-utility trade-offs and hyperparameter effects in our framework." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 206, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 206, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 206, + 296, + 422 + ], + "type": "text", + "content": "model's output prediction backward through the network, layer by layer, conserving relevance at each step. This process aggregates relevance into contiguous, semantically coherent regions that represent structured object parts. In stark contrast, methods like SmoothGrad or Integrated Gradients identify 'hotspot' pixels that are locally influential but often disconnected from each other. Consequently, under high sparsification, LRP retains meaningful visual patterns, whereas other methods are left with a sparse collection of isolated points that lack sufficient context for accurate classification. We conclude that structural coherence in saliency maps is an essential criterion when selecting the attribution method. This finding is corroborated by current literature [2, 13, 23], which confirms LRP's superior explanation quality and reinforces that our work is consistent with established findings on the importance of explanation quality for downstream tasks. More discussion can be found in the supplementary material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 430, + 162, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 162, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 162, + 443 + ], + "type": "text", + "content": "7.2. Privacy vs. Utility" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 448, + 296, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 448, + 296, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 448, + 296, + 616 + ], + "type": "text", + "content": "We investigate how the privacy budget " + }, + { + "bbox": [ + 55, + 448, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 55, + 448, + 296, + 616 + ], + "type": "text", + "content": " impacts model performance across attribution methods. Table 6 shows that increasing " + }, + { + "bbox": [ + 55, + 448, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 55, + 448, + 296, + 616 + ], + "type": "text", + "content": " improves accuracy for all methods, demonstrating the expected privacy-utility trade-off: stronger privacy (lower " + }, + { + "bbox": [ + 55, + 448, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 55, + 448, + 296, + 616 + ], + "type": "text", + "content": ") leads to reduced performance due to greater noise. Among all methods, LRP consistently achieves the highest accuracy across privacy levels outperforming the next best (Integrated Gradients) by " + }, + { + "bbox": [ + 55, + 448, + 296, + 616 + ], + "type": "inline_equation", + "content": "7 - 8\\%" + }, + { + "bbox": [ + 55, + 448, + 296, + 616 + ], + "type": "text", + "content": " and maintaining robustness even under strict privacy constraints. SmoothGrad shows modest improvements over basic gradients, though less pronounced than Integrated Gradients. These results underscore LRP's ability to extract stable, task-relevant features leading to better generalization with less privacy leakage, aligning with prior work on explanation fidelity [3, 22]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 624, + 272, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 272, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 272, + 636 + ], + "type": "text", + "content": "7.3. Ablation Study on the Knowledge Weight" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "We evaluate how the knowledge weight " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": ", as defined in Equation 11, affects performance across CIFAR-10, CIFAR-100, and Tiny-ImageNet. Figure 2c shows that intermediate values (" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda = 0.5" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": ") consistently outperform both extremes, with CIFAR-10 accuracy peaking at " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "83.46\\%" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " compared to " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "72.90\\%" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda = 0.0" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": ", equivalent to FedAvg) and" + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 315, + 247, + 554, + 319 + ], + "blocks": [ + { + "bbox": [ + 313, + 205, + 553, + 237 + ], + "lines": [ + { + "bbox": [ + 313, + 205, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 553, + 237 + ], + "type": "text", + "content": "Table 6. Ablation study showing test accuracy) for different privacy budgets " + }, + { + "bbox": [ + 313, + 205, + 553, + 237 + ], + "type": "inline_equation", + "content": "(\\varepsilon)" + }, + { + "bbox": [ + 313, + 205, + 553, + 237 + ], + "type": "text", + "content": " across XAI-based federated learning methods. Results for CIFAR10 at " + }, + { + "bbox": [ + 313, + 205, + 553, + 237 + ], + "type": "inline_equation", + "content": "K = 10, \\alpha = 0.1" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 247, + 554, + 319 + ], + "lines": [ + { + "bbox": [ + 315, + 247, + 554, + 319 + ], + "spans": [ + { + "bbox": [ + 315, + 247, + 554, + 319 + ], + "type": "table", + "html": "
εGradientSmoothGradInt. Grad.LRP
570.55 ± 0.6271.23 ± 0.5972.45 ± 0.5280.06 ± 0.33
771.45 ± 0.5872.15 ± 0.5573.35 ± 0.4780.89 ± 0.31
1072.25 ± 0.5172.95 ± 0.4874.15 ± 0.4181.65 ± 0.29
1573.12 ± 0.4573.85 ± 0.4375.02 ± 0.3882.11 ± 0.25
2073.99 ± 0.4274.63 ± 0.3975.89 ± 0.3583.46 ± 0.28
", + "image_path": "aba453af7d5b6dc1e2d8de56fc42f279a3308965abcb304db4c6b33acffcd584.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 341, + 554, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 554, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 554, + 376 + ], + "type": "text", + "content": "61.18% (" + }, + { + "bbox": [ + 313, + 341, + 554, + 376 + ], + "type": "inline_equation", + "content": "\\lambda = 1.0" + }, + { + "bbox": [ + 313, + 341, + 554, + 376 + ], + "type": "text", + "content": ", synthetic data equal weight). Our ablation thus shows that balancing the contribution of synthetic data is critical for performance." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 386, + 448, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 386, + 448, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 448, + 399 + ], + "type": "text", + "content": "7.4. Warmup and Efficiency" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 404, + 555, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 404, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 404, + 555, + 525 + ], + "type": "text", + "content": "Figure 2d shows accuracy slightly improves with more warmup rounds, plateauing at 10, suggesting this is the optimal trade-off between convergence and communication cost. Our approach is also computationally efficient. Attribution masks are computed in a single backward pass during warmup, incurring minimal overhead. In contrast, methods like FedGen and FedFTG retrain generators every round, and FedFed requires costly VAE training—making FedXDS more suitable for resource-constrained settings. We provide more discussion in the supplementary material." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 537, + 388, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 537, + 388, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 537, + 388, + 550 + ], + "type": "text", + "content": "8. Conclusion" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "content": "In this work, we introduced FedXDS, leveraging attribution methods to address statistical heterogeneity, privacy, and computational efficiency in federated learning. By using attribution maps to extract and share privacy-preserved task-relevant features, our approach enables effective knowledge transfer while ensuring strong privacy guarantees. Experiments show that FedXDS consistently outperforms existing methods in accuracy and communication efficiency. Moreover, the sparsity induced by relevance attribution enhances privacy, mitigating feature inversion and membership inference risks. These findings highlight attribution methods as a promising tool for improving federated learning while preserving privacy and efficiency." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4579" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 170, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 170, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 170, + 85 + ], + "type": "text", + "content": "9. Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 297, + 135 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 135 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 135 + ], + "type": "text", + "content": "This work was supported by the European Union's Horizon Europe research and innovation programme (EU Horizon Europe) as grant ACHILLES (101189689)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 144, + 115, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 144, + 115, + 156 + ], + "spans": [ + { + "bbox": [ + 56, + 144, + 115, + 156 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 163, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 163, + 296, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 163, + 296, + 207 + ], + "spans": [ + { + "bbox": [ + 61, + 163, + 296, + 207 + ], + "type": "text", + "content": "[1] Durmus Alp Emre Acar, Yue Zhao, Ramon Matas, Matthew Mattina, Paul Whatmough, and Venkatesh Saligrama. Federated learning based on dynamic regularization. In International Conference on Learning Representations, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 209, + 296, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 209, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 61, + 209, + 296, + 251 + ], + "type": "text", + "content": "[2] Leila Arras, Ahmed Osman, and Wojciech Samek. Clevr-xai: A benchmark dataset for the ground truth evaluation of neural network explanations. Information Fusion, 81:14-40, 2022. 1, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 252, + 296, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 252, + 296, + 306 + ], + "spans": [ + { + "bbox": [ + 61, + 252, + 296, + 306 + ], + "type": "text", + "content": "[3] Sebastian Bach, Alexander Binder, Grégoire Montavon, Frederick Klauschen, Klaus-Robert Müller, and Wojciech Samek. On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. PLOS ONE, 10(7):1-46, 2015. 3, 5, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 308, + 296, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 308, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 62, + 308, + 296, + 361 + ], + "type": "text", + "content": "[4] Jeremiah Blocki, Avrim Blum, Anupam Datta, and Or Sheffet. Differentially private data analysis of social networks via restricted sensitivity. In Proceedings of the 4th conference on Innovations in Theoretical Computer Science, pages 87-96, 2013. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 362, + 296, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 362, + 296, + 405 + ], + "spans": [ + { + "bbox": [ + 62, + 362, + 296, + 405 + ], + "type": "text", + "content": "[5] Sebastian Caldas, Sai Meher Karthik Duddu, Peter Wu, Tian Li, Jakub Konečný, H Brendan McMahan, Virginia Smith, and Ameet Talwalkar. Leaf: A benchmark for federated settings. arXiv preprint arXiv:1812.01097, 2018. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 406, + 296, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 406, + 296, + 472 + ], + "spans": [ + { + "bbox": [ + 62, + 406, + 296, + 472 + ], + "type": "text", + "content": "[6] Konstantinos Chatzikokolakis, Miguel E Andres, Nicolás Emilio Bordenabe, and Catuscia Palamidessi. Broadening the scope of differential privacy using metrics. In Privacy Enhancing Technologies: 13th International Symposium, PETS 2013, Bloomington, IN, USA, July 10-12, 2013. Proceedings 13, pages 82-102. Springer, 2013. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 472, + 296, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 472, + 296, + 515 + ], + "spans": [ + { + "bbox": [ + 62, + 472, + 296, + 515 + ], + "type": "text", + "content": "[7] Daoyuan Chen, Dawei Gao, Weirui Kuang, Yaliang Li, and Bolin Ding. pfl-bench: A comprehensive benchmark for personalized federated learning. Advances in Neural Information Processing Systems, 35:9344-9360, 2022. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 516, + 296, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 516, + 296, + 569 + ], + "spans": [ + { + "bbox": [ + 62, + 516, + 296, + 569 + ], + "type": "text", + "content": "[8] Rudresh Dwivedi, Devam Dave, Het Naik, Smiti Singhal, Rana Omer, Pankesh Patel, Bin Qian, Zhenyu Wen, Tejal Shah, Graham Morgan, et al. Explainable ai (xai): Core ideas, techniques, and solutions. ACM Computing Surveys, 55(9):1-33, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 571, + 296, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 571, + 296, + 603 + ], + "spans": [ + { + "bbox": [ + 62, + 571, + 296, + 603 + ], + "type": "text", + "content": "[9] Cynthia Dwork. Differential privacy. In International colloquium on automata, languages, and programming, pages 1-12. Springer, 2006. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 604, + 296, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 604, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 57, + 604, + 296, + 647 + ], + "type": "text", + "content": "[10] Oluwaseyi Feyisetan and Shiva Kasiviswanathan. Private release of text embedding vectors. In Proceedings of the First Workshop on Trustworthy Natural Language Processing, pages 15-27, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 296, + 712 + ], + "type": "text", + "content": "[11] Peter Kairouz, H Brendan McMahan, Brendan Avent, Aurélien Bellet, Mehdi Bennis, Arjun Nitin Bhagoji, Kallista Bonawitz, Zachary Charles, Graham Cormode, Rachel Cummings, et al. Advances and open problems in federated learning. Foundations and trends® in machine learning, 14(1-2): 1-210, 2021. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 138 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 138 + ], + "type": "text", + "content": "[12] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank Reddi, Sebastian Stich, and Ananda Theertha Suresh. SCAFFOLD: Stochastic controlled averaging for federated learning. In Proceedings of the 37th International Conference on Machine Learning, pages 5132-5143. PMLR, 2020. 1, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 140, + 554, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 554, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 554, + 194 + ], + "type": "text", + "content": "[13] Sebastian Lapuschkin, Stephan Wäldchen, Alexander Binder, Grégoire Montavon, Wojciech Samek, and Klaus-Robert Müller. Unmasking clever hans predictors and assessing what machines really learn. Nature communications, 10(1):1096, 2019. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 197, + 554, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 554, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 554, + 251 + ], + "type": "text", + "content": "[14] Bo Li, Mikkel N Schmidt, Tommy S Alstrom, and Sebastian U Stich. On the effectiveness of partial variance reduction in federated learning with heterogeneous data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3964-3973, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 253, + 554, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 554, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 554, + 297 + ], + "type": "text", + "content": "[15] Qinbin Li, Bingsheng He, and Dawn Song. Model-contrastive federated learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10713–10722, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 298, + 554, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 554, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 554, + 342 + ], + "type": "text", + "content": "[16] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. Proceedings of Machine learning and systems, 2:429-450, 2020. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 343, + 554, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 554, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 554, + 387 + ], + "type": "text", + "content": "[17] Xiaoxiao Li, Meirui Jiang, Xiaofei Zhang, Michael Kamp, and Qi Dou. Fed{bn}: Federated learning on non- " + }, + { + "bbox": [ + 316, + 343, + 554, + 387 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{id}\\}" + }, + { + "bbox": [ + 316, + 343, + 554, + 387 + ], + "type": "text", + "content": " features via local batch normalization. In International Conference on Learning Representations, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 388, + 554, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 554, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 554, + 431 + ], + "type": "text", + "content": "[18] Tao Lin, Lingjing Kong, Sebastian U Stich, and Martin Jaggi. Ensemble distillation for robust model fusion in federated learning. Advances in neural information processing systems, 33:2351-2363, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 434, + 554, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 554, + 498 + ], + "type": "text", + "content": "[19] H. Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Agüera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Proceedings of the 20th International Conference on Artificial Intelligence and Statistics (AISTATS). PMLR, 2017. 1, 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 501, + 554, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 554, + 555 + ], + "type": "text", + "content": "[20] Matias Mendieta, Taojiannan Yang, Pu Wang, Minwoo Lee, Zhengming Ding, and Chen Chen. Local learning matters: Rethinking data heterogeneity in federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8397-8406, 2022. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 557, + 554, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 554, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 554, + 601 + ], + "type": "text", + "content": "[21] Zhe Qu, Xingyu Li, Rui Duan, Yao Liu, Bo Tang, and Zhuo Lu. Generalized federated learning via sharpness aware minimization. In International conference on machine learning, pages 18250-18280. PMLR, 2022. 1, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 602, + 554, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 554, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 554, + 656 + ], + "type": "text", + "content": "[22] Wojciech Samek, Alexander Binder, Grégoire Montavon, Sebastian Lapischkin, and Klaus-Robert Müller. Evaluating the visualization of what a deep neural network has learned. IEEE Transactions on Neural Networks and Learning Systems, 28(11):2660-2673, 2017. 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 658, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 554, + 712 + ], + "type": "text", + "content": "[23] Wojciech Samek, Grégoire Montavon, Sebastian Lapischkin, Christopher J. Anders, and Klaus-Robert Müller. Explaining deep neural networks and beyond: A review of methods and applications. Proceedings of the IEEE, 109(3): 247-278, 2021. 8" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "4580" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 677 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[24] Felix Sattler, Tim Korjakow, Roman Rischke, and Wojciech Samek. Fedaux: Leveraging unlabeled auxiliary data in federated learning. IEEE Transactions on Neural Networks and Learning Systems, 34(9):5531-5543, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "text", + "content": "[25] Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside convolutional networks: Visualising image classification models and saliency maps. In International Conference on Learning Representations (ICLR), 2014. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "text", + "content": "[26] Abhishek Singh, Praneeth Vepakomma, Vivek Sharma, and Ramesh Raskar. Posthoc privacy guarantees for collaborative inference with modified propose-test-release. Advances in Neural Information Processing Systems, 36:26438-26451, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 251 + ], + "type": "text", + "content": "[27] Daniel Smilkov, Nikhil Thorat, Been Kim, Fernanda Viégas, and Martin Wattenberg. Smoothgrad: removing noise by adding noise. arXiv preprint arXiv:1706.03825, 2017. 3, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 285 + ], + "type": "text", + "content": "[28] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In International Conference on Machine Learning, pages 3319-3328. PMLR, 2017. 3, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 286, + 294, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 351 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 351 + ], + "type": "text", + "content": "[29] Jianyu Wang, Qinghua Liu, Hao Liang, Gauri Joshi, and H. Vincent Poor. Tackling the objective inconsistency problem in heterogeneous federated optimization. In Proceedings of the 34th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2020. Curran Associates Inc. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 353, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 353, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 353, + 294, + 407 + ], + "type": "text", + "content": "[30] Zhiqin Yang, Yonggang Zhang, Yu Zheng, Xinmei Tian, Hao Peng, Tongliang Liu, and Bo Han. Fedfed: Feature distillation against data heterogeneity in federated learning. Advances in Neural Information Processing Systems, 36, 2024. 1, 3, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 409, + 294, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 409, + 294, + 452 + ], + "spans": [ + { + "bbox": [ + 56, + 409, + 294, + 452 + ], + "type": "text", + "content": "[31] Mang Ye, Xiuwen Fang, Bo Du, Pong C Yuen, and Dacheng Tao. Heterogeneous federated learning: State-of-the-art and research challenges. ACM Computing Surveys, 56(3):1-44, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 454, + 294, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 454, + 294, + 507 + ], + "spans": [ + { + "bbox": [ + 56, + 454, + 294, + 507 + ], + "type": "text", + "content": "[32] Rui Ye, Mingkai Xu, Jianyu Wang, Chenxin Xu, Siheng Chen, and Yanfeng Wang. Feddisco: Federated learning with discrepancy-aware collaboration. In International Conference on Machine Learning, pages 39879-39902. PMLR, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 510, + 294, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 510, + 294, + 564 + ], + "spans": [ + { + "bbox": [ + 56, + 510, + 294, + 564 + ], + "type": "text", + "content": "[33] Lin Zhang, Li Shen, Liang Ding, Dacheng Tao, and Ling-Yu Duan. Fine-tuning global model via data-free knowledge distillation for non-iid federated learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10174-10183, 2022. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 566, + 294, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 566, + 294, + 597 + ], + "spans": [ + { + "bbox": [ + 56, + 566, + 294, + 597 + ], + "type": "text", + "content": "[34] Yue Zhao, Meng Li, Liangzhen Lai, Naveen Suda, Damon Civin, and Vikas Chandra. Federated learning with non-iid data. arXiv preprint arXiv:1806.00582, 2018. 1, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 600, + 294, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 600, + 294, + 632 + ], + "spans": [ + { + "bbox": [ + 56, + 600, + 294, + 632 + ], + "type": "text", + "content": "[35] Yue Zhao, Meng Li, Liangzhen Lai, Naveen Suda, Damon Civin, and Vikas Chandra. Federated learning with non-iid data. arXiv preprint arXiv:1806.00582, 2018. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 633, + 294, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 633, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 633, + 294, + 677 + ], + "type": "text", + "content": "[36] Zhuangdi Zhu, Junyuan Hong, and Jiayu Zhou. Data-free knowledge distillation for heterogeneous federated learning. In Proceedings of the 38th International Conference on Machine Learning, pages 12878-12889. PMLR, 2021. 1, 3, 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "4581" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Federated Continual Instruction Tuning/c95b9507-af86-4883-a6e8-057630ebe555_content_list.json b/2025/Federated Continual Instruction Tuning/c95b9507-af86-4883-a6e8-057630ebe555_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e2ca9d001a0ab80c89559cb5d18b27784b26a2b2 --- /dev/null +++ b/2025/Federated Continual Instruction Tuning/c95b9507-af86-4883-a6e8-057630ebe555_content_list.json @@ -0,0 +1,1676 @@ +[ + { + "type": "text", + "text": "Federated Continual Instruction Tuning", + "text_level": 1, + "bbox": [ + 295, + 130, + 702, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haiyang Guo $^{1,2}$ , Fanhu Zeng $^{2,3}$ , Fei Zhu $^{4}$ , Wenzhuo Liu $^{2,3}$ , Da-Han Wang $^{5}$ , Jian Xu $^{2,3}$ , Xu-Yao Zhang $^{1,2,3*}$ , Cheng-Lin Liu $^{1,2,3}$", + "bbox": [ + 205, + 178, + 795, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Advanced Interdisciplinary Sciences, UCAS $^{2}$ MAIS, CASIA", + "bbox": [ + 217, + 215, + 785, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ School of Artificial Intelligence, UCAS $^{4}$ Centre for Artificial Intelligence and Robotics, HKISI-CAS", + "bbox": [ + 101, + 233, + 901, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{5}$ FKLPRIU, School of Computer and Information Engineering, Xiamen University of Technology", + "bbox": [ + 116, + 251, + 885, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{guohaiyang2023, zengfanhu2022, jian.xu} @ia.ac.cn, zhfei2018@gmail.com, {xyz, liucl} @nlpr.ia.ac.cn", + "bbox": [ + 153, + 270, + 844, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 321, + 326, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A vast amount of instruction tuning data is crucial for the impressive performance of Large Multimodal Models (LMMs), but the associated computational costs and data collection demands during supervised fine-tuning make it impractical for most researchers. Federated learning (FL) has the potential to leverage all distributed data and training resources to reduce the overhead of joint training. However, most existing methods assume a fixed number of tasks, while in real-world scenarios, clients continuously encounter new knowledge and often struggle to retain old tasks due to memory constraints. In this work, we introduce the Federated Continual Instruction Tuning (FCIT) benchmark to model this real-world challenge. Our benchmark includes two realistic scenarios, encompassing four different settings and twelve carefully curated instruction tuning datasets. To address the challenges posed by FCIT, we propose a dynamic knowledge organization to effectively integrate updates from different tasks during training and subspace selective activation to allocate task-specific output during inference. Extensive experimental results demonstrate that our proposed method significantly enhances model performance across varying levels of data heterogeneity and catastrophic forgetting. Code and dataset are released at https://github.com/Ghy0501/FCIT.", + "bbox": [ + 89, + 353, + 485, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 746, + 220, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large Multimodal Models (LMMs) [2, 32, 34], which integrate Large Language Model [4, 20, 46, 49] with a visual encoder and multimodal projector to bridge visual and textual modalities, have exhibited impressive visual understanding and complex reasoning abilities. A crucial factor in this success is the supervised fine-tuning of LMMs using huge and diverse visual instruction-following data [32]", + "bbox": [ + 89, + 771, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to align with human preferences. However, collecting such vast amounts of training data and computational resources for joint fine-tuning is impractical for most researchers. Federated Learning (FL) [28, 31, 38, 43], as a decentralized paradigm, offers a viable alternative by leveraging distributed data and computational resources for local training while integrating local weights to produce a unified model. This paradigm accommodates constraints on storage and computation while ensuring privacy protection.", + "bbox": [ + 511, + 323, + 905, + 459 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generally, most existing FL frameworks [7, 55, 58] are modeled in static, closed-world scenarios [3, 59], where a fixed set of tasks is predefined and unchanged. However, real-world applications are dynamic [59, 63], requiring models to continuously acquire new knowledge while retaining previously learned tasks. Taking a realistic health emergency as an example, different hospitals act as local clients when a major disease outbreak occurs. Large hospitals can leverage FL to collaboratively train on their case data, building a comprehensive virus knowledge base, while smaller clinics update their own cases using this global knowledge. Over time, both large and small hospitals need to keep updating the global knowledge base through continual learning methods to cope with the event. Additionally, the emergence of new strains necessitates the simultaneous integration of knowledge and response strategies to minimize potential losses. In this scenario, traditional FL methods struggle with newly arrived knowledge, while Continual Learning (CL) [14, 15, 33, 56, 60] methods alone do not facilitate knowledge sharing between clients. Only an organic integration of both can address this real-world problem. Recently, numerous Federated Continual Learning methods have emerged to address this challenge in traditional image classification tasks [11, 12, 57]. However, their methodological and task-setting limitations make them insufficient for current LMM applications. Therefore, a more comprehensive benchmark is needed to better simulate the practical application of LMMs in real-world scenarios.", + "bbox": [ + 511, + 460, + 906, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we first establish a Federated Continual In", + "bbox": [ + 531, + 885, + 903, + 898 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 805, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author.", + "bbox": [ + 109, + 887, + 238, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1325", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1172980ffe523979c81b514f577a783b638694ff4f32a162ac2b3c43a6229dc7.jpg", + "image_caption": [ + "Figure 1. Overview of FCIT benchmark. FCIT encompasses 2 real-world scenarios, 4 FCIT settings, and 12 curated datasets, providing a comprehensive simulation of LMMs instruction-following training in real-world applications." + ], + "image_footnote": [], + "bbox": [ + 102, + 92, + 890, + 406 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "struction Tuning (FCIT) benchmark to fill this gap. Specifically, we simulate two client-level realistic scenarios: (i) Homogeneous FCIT refers to an FL system learning a series of instruction tuning data sequentially, with different clients learning the same task at each stage. (ii) Heterogeneous FCIT, on the other hand, requires the FL system to collaborate on different tasks simultaneously, with different clients potentially learning different tasks in the same stage. Building on this, we define two settings of datasets for each scenario: Capability-related and Task-related. The former evaluates the model's ability to integrate multiple dimensions of instruction following in a short period, while the latter assesses the model's performance during long phases of continual learning. For dataset selection, we curate twelve instruction tuning datasets unseen during LMM fine-tuning or with low zero-shot performance, preventing information leakage [13, 23]. In addition, we introduce varying degrees of data heterogeneity for each setting to challenge the model's performance in the non-IID situation [27]. An illustration of our FCIT benchmark is provided in Figure 1. To the best of our knowledge, this is the first work to introduce a comprehensive benchmark for federated learning of LMM in a continual learning setting.", + "bbox": [ + 91, + 455, + 483, + 805 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To effectively address the challenges posed by FCIT, we propose a novel Dynamic knowledge organization and Subspace seleCtive activatiOn (DISCO) framework. Specifically, we identify FCIT challenges into two types: conflicts between different tasks within the same stage and conflicts between old and new tasks across different stages.", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The former can cause catastrophic forgetting by altering the parameter space of previous tasks when learning new ones without access to past data, while the latter requires the model to integrate and organize knowledge from different tasks to derive unified representations. Therefore, we first propose Dynamic Knowledge Organization (DKO), which leverages a dynamic cache at the global server to store task-specific parameters. Using a unique identity token matching mechanism, it systematically organizes knowledge for different tasks into corresponding subspaces within the cache, effectively mitigating two types of conflicts. To better utilize the organized task subspaces in the dynamic cache, we introduce Subspace Selective Activation (SSA), which selectively activates subspaces relevant to the test input while filtering out irrelevant outputs, leading to significant performance improvements. Consequently, these designs enable our framework to efficiently tackle the data heterogeneity and catastrophic forgetting in FCIT. In summary, our major contributions are:", + "bbox": [ + 511, + 455, + 906, + 743 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present the first Federated Continual Instruction Tuning (FCIT) benchmark designed for real-world scenarios, providing a comprehensive evaluation of LMMs to continuously learn new knowledge using distributed data and training resources in real applications.", + "- We propose a novel DISCO framework that integrates dynamic knowledge organization and subspace selective activation to efficiently address data heterogeneity and catastrophic forgetting in FCIT settings.", + "- Extensive experiments demonstrate that our method sig" + ], + "bbox": [ + 513, + 750, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1326", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "nificantly enhances model performance under data heterogeneity while minimizing the catastrophic forgetting, and achieves state-of-the-art performance.", + "bbox": [ + 102, + 90, + 482, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 150, + 233, + 166 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Large Multimodal Models. With the grand unification of Large Language Models (LLMs) [4, 20, 49] for various NLP tasks, Large Multimodal Models (LMMs) [2, 9, 32, 34, 61] emerge by extending LLMs to combine with visual encoder and multimodal projectors, demonstrating exceptional visual understanding and complex reasoning abilities. To better align with human preferences, these LMMs typically undergo further fine-tuning on extensive instruction-following data, ensuring they meet the demands of real-world applications [32]. However, in real-world scenarios, improving the performance of LMMs on new downstream tasks becomes a significant challenge without access to sufficient training data and computational resources. In this paper, we introduce the first Federated Continual Instruction Tuning (FCIT) benchmark to bridge this gap from the perspective of distributed training and continual learning.", + "bbox": [ + 89, + 174, + 482, + 416 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Federated Continual Learning. In classical vision tasks, Federated Continual Learning (FCL) [54] aims to adapt the global model to new data while maintaining the knowledge of the old task. From the perspective of LMMs, research in this setting faces two main challenges: (1) Most of the methods are designed for traditional vision tasks (i.e., image classification). For instance, MFCL [1] employs a generative model to synthesize images of previously learned classes, thereby mitigating forgetting. PILoRA [12] introduces a prototype re-weight module to address the classifier bias caused by data heterogeneity and obtain unified knowledge through LoRA [18] fusion. Despite progress in classification tasks, complex designs remain challenging to adapt for LMMs research. (2) FCL follows a single data composition by splitting a dataset (e.g. ImageNet [10]) into different tasks based on classes, whereas LMMs face greater challenges in continual learning due to their diverse tasks with varying styles [5, 6, 56].", + "bbox": [ + 89, + 417, + 482, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This work pioneers a Federated Continual Instruction Tuning setup for LMMs and establishes diverse scenarios to simulate real-world applications comprehensively. Notably, AFCL [45] is most relevant to our work, as it enables clients to continuously learn multiple tasks in different orders and asynchronous time slots. However, it is specifically designed for image classification tasks, whereas our study focuses on the more widely used LMM.", + "bbox": [ + 89, + 688, + 482, + 810 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Problem Formulation", + "text_level": 1, + "bbox": [ + 89, + 823, + 297, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 848, + 230, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Instruction tuning enhances LMM's ability to understand and execute human instructions by performing supervised", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "fine-tuning of pre-trained models on extensive datasets comprising instructions and responses. Formally, the instruction data $\\mathcal{D} = \\{(\\mathbf{x}_v^j,\\mathbf{x}_{ins}^j,\\mathbf{x}_{res}^j)^N\\}$ consists of image input $\\mathbf{x}_v$ , instruction $\\mathbf{x}_{ins}$ and response $\\mathbf{x}_{res}$ , where $N$ represents the total number of samples. For clarity, given a simple image-instruction pair with a response of length $L$ , the objective of an LMM is to predict the next token autoregressively based on all preceding tokens:", + "bbox": [ + 511, + 90, + 903, + 212 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np \\left(\\mathbf {x} _ {r e s} \\mid \\mathbf {x} _ {v}, \\mathbf {x} _ {i n s}\\right) = \\prod_ {i = 1} ^ {L} p _ {\\theta} \\left(x _ {i} \\mid \\mathbf {x} _ {v}, \\mathbf {x} _ {i n s}, \\mathbf {x} _ {r e s, < i}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 218, + 903, + 260 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\theta$ denotes the trainable parameters during fine-tuning, $\\mathbf{x}_{\\mathrm{res}, < i}$ denotes all response tokens preceding the current prediction token $x_{i}$ . Then, the loss function of fine-tuning LMMs can be expressed as:", + "bbox": [ + 511, + 265, + 906, + 325 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\theta} = - \\frac {1}{N} \\sum_ {j = 1} ^ {N} \\sum_ {i = 1} ^ {L _ {j}} \\log p _ {\\theta} \\left(x _ {i} ^ {j} \\mid \\mathbf {x} _ {v} ^ {j}, \\mathbf {x} _ {i n s} ^ {j}, \\mathbf {x} _ {r e s, < i} ^ {j}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 333, + 903, + 377 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Federated learning framework typically comprises a global server and several local clients, all employing the same LMM with a shared homogeneous model architecture. In each communication round1, local clients train their models on own data and upload the updated weights to the global server for aggregation, enabling collaborative optimization of the global model while preserving data privacy.", + "bbox": [ + 511, + 382, + 903, + 487 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Considering the excessive communication overhead of transferring the entire LMM between clients and the global server, we adopt LoRA [18] for efficient fine-tuning, balancing training overhead and implementation cost [55, 58]. Specifically, for a weight matrix $\\mathbf{W}_0\\in \\mathbb{R}^{d\\times k}$ , LoRA decomposes parameter updates $\\Delta \\mathbf{W}$ during fine-tuning into two low-rank subspaces:", + "bbox": [ + 511, + 488, + 905, + 592 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {W} = \\mathbf {W} _ {0} + \\Delta \\mathbf {W} = \\mathbf {W} _ {0} + \\mathbf {B A}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 601, + 903, + 616 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{B} \\in \\mathbb{R}^{d \\times r}$ , $\\mathbf{A} \\in \\mathbb{R}^{r \\times k}$ and $r \\ll \\min\\{d, k\\}$ .", + "bbox": [ + 511, + 623, + 844, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Continual learning aims to minimize the loss on the current task while retaining knowledge from previous tasks. Formally, given a sequence of datasets $\\mathcal{D}_1,\\mathcal{D}_2,\\dots ,\\mathcal{D}_T$ , the optimization objective at task $t$ is:", + "bbox": [ + 511, + 638, + 903, + 700 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\mathcal {L} (\\theta) = \\mathbb {E} _ {(x, y) \\sim \\mathcal {D} _ {t}} [ \\mathcal {L} (f _ {\\theta_ {t}} (x), y) ] + \\sum_ {i = 1} ^ {t - 1} \\epsilon_ {i}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 707, + 903, + 752 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\operatorname {s. t.} \\mathbb {E} _ {(x, y) \\sim \\mathcal {D} _ {i}} \\left[ \\mathcal {L} \\left(f _ {\\theta_ {t}} (x), y\\right) - \\mathcal {L} \\left(f _ {\\theta_ {t - 1}} (x), y\\right) \\right] \\leq \\epsilon_ {i}, \\\\ \\epsilon_ {i} \\geq 0; \\forall i \\in [ 1, \\dots , t - 1 ], \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 751, + 869, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon_{i}$ is a slack variable that allows a small increase in the loss from the old datasets, providing tolerance for minor forgetting while focusing on learning the current task. In Eq. 4, $x$ and $y$ can be viewed as the multimodal inputs $(\\mathbf{x}_v,\\mathbf{x}_{ins})$ and the response $\\mathbf{x}_{res}$ , respectively.", + "bbox": [ + 511, + 791, + 903, + 868 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "For clarification, we standardize the communication round in Section 4 to 1.", + "bbox": [ + 511, + 875, + 903, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1327", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Federated Continual Instruction Tuning", + "text_level": 1, + "bbox": [ + 89, + 90, + 434, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Figure 1, we integrate federated and continual learning for LMMs within a unified framework and propose two client-level realistic scenarios: (1) Homogeneous FCIT (Hom-FCIT). In this scenario, clients sequentially learn a series of tasks, allowing the global server to continuously update its knowledge. Each client learns the same task at a given stage and can only access data from that stage. (2) Heterogeneous FCIT (Het-FCIT). In real-world applications, different clients may learn different tasks within the same stage, enabling the global server to respond more rapidly to diverse instructions. This requires the model not only to coordinate the knowledge of different tasks learned by clients in the current stage but also to mitigate forgetting during the learning process. For each scenario, we define two dataset-level settings: capability-related and task-related, detailed as follows.", + "bbox": [ + 89, + 117, + 483, + 359 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Capability-related. Following the dataset construction and division in LLaVA-OneVision [25], we classify the 12 datasets into 4 capabilities: General, Math, Chart, and Other. The General capability includes A-OKVQA [44], ImageNet-R [16], Grounding [37], and IconQA [35]; Math comprises CLEVR-Math [8], super-CLEVR [30], and TabMWP [36]; Chart involves ArxivQA [26], FigureQA [22], and DVQA [21]; and Other encompasses OCR-VQA [39] and Flickr30k [40]. We treat these four capabilities as different stages of continual learning, with each capability comprising a mixture of datasets.", + "- Task-related. To evaluate the performance of different methods in a long-phase continual learning situation, we selected 8 datasets: ImageNet-R, ArxivQA, IconQA, CLEVR-Math, OCR-VQA, Flickr30k, FigureQA, and super-CLEVR as distinct stages of continual learning." + ], + "bbox": [ + 89, + 372, + 483, + 614 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In total, we propose 4 FCIT settings to evaluate different methods. Compared to existing works [5, 6, 55, 56, 58], we are the first to explore the organic integration of FL and CL in the context of LMM. We provide further illustrations on the settings and datasets in Appendix A.", + "bbox": [ + 89, + 628, + 483, + 704 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. The Proposed Framework: DISCO", + "text_level": 1, + "bbox": [ + 89, + 733, + 406, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overview of the Method. As shown in Figure 2 and 3, our method primarily consists of: (1) Dynamic Knowledge Organization (DKO), which dynamically integrates knowledge learned by different clients across stages during training, significantly reducing inter-task conflicts (Section 4.1); and (2) Subspace Selective Activation (SSA), which selectively activates subspace outputs based on input features during inference, effectively filtering out irrelevant information (Section 4.2). We name this framework DISCO.", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2069e575ef5e18f160271e66931eeb8285660d720ea8a7abe549298323d281bc.jpg", + "image_caption": [ + "Figure 2. Illustration of the proposed DKO. Dynamic caches store the knowledge of each subspace while matching and updating through identity tokens." + ], + "image_footnote": [], + "bbox": [ + 526, + 92, + 885, + 348 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Dynamic Knowledge Organization", + "text_level": 1, + "bbox": [ + 511, + 401, + 815, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The core challenge of FCIT is enabling the global server to effectively harmonize knowledge learned by clients, addressing both the forgetting of previous knowledge when acquiring new tasks and the conflicts that arise from integrating knowledge from different tasks within the same stage. To this end, we propose maintaining a dynamic cache at the global server to organize task-specific knowledge uploaded by clients at different stages, thereby preventing both forgetting and conflicts between knowledge. Specifically, each task is assigned a dedicated parameter space to store and update its corresponding knowledge:", + "bbox": [ + 511, + 422, + 906, + 589 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\mathbf {W} = \\mathbf {B} \\mathbf {A} \\Leftrightarrow \\underbrace {\\left\\{\\mathbf {B} _ {1} \\mathbf {A} _ {1} , \\cdots , \\mathbf {B} _ {T} \\mathbf {A} _ {T} \\right\\}} _ {\\text {t a s k - s p e c i f i c s u b s p a c e}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 597, + 906, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $T$ is the number of tasks learned. At this point, it is crucial to aggregate the parameters uploaded by clients into their respective subspaces without privacy leakage.", + "bbox": [ + 511, + 640, + 905, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by the widespread use of prototypes in FL fields [19, 47, 48], we propose to distinguish knowledge across tasks by using the feature mean of each client's training data as an Identity Token. Specifically, considering the uniqueness of textual inputs in the visual instruction tuning task, we introduce a text encoder $f_{ins}$ for each client to extract the feature of the input instruction $\\mathbf{x}_{ins,j}^{t}$ during training, and take the mean value $\\mu_k^t$ as the local identity token of the $k$ -th client on the task $t$ at the end of training:", + "bbox": [ + 511, + 686, + 905, + 821 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {k} ^ {t} = \\frac {1}{n _ {k} ^ {t}} \\sum_ {j = 1} ^ {n _ {k} ^ {t}} f _ {i n s} \\left(\\mathbf {x} _ {i n s, j} ^ {t}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 619, + 830, + 903, + 875 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $n_k^t$ is the number of training samples of client $k$ at", + "bbox": [ + 511, + 885, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1328", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/321710d579552cd17ea1e7d5f77fe3666442909e13c48150ba9e12d9c387c1ca.jpg", + "image_caption": [ + "Figure 3. Illustration of the proposed SSA. Each subspace organized by DKO dynamically adjusts its output via the intrinsic activation matrix during inference, effectively filtering irrelevant information and enhancing model performance." + ], + "image_footnote": [], + "bbox": [ + 106, + 89, + 893, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "task $t$ , and we use CLIP's text encoder [41] in our experiments. Notably, the text encoder $f_{ins}$ remains frozen during training, which prevents additional training overhead.", + "bbox": [ + 89, + 279, + 482, + 324 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Using the identity tokens, we first apply cosine similarity along with a threshold to match the uploaded local identity token $\\mu_j^t$ with the global identity token $\\tilde{\\mathbf{z}}_i$ :", + "bbox": [ + 89, + 324, + 482, + 371 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\mu_ {j} ^ {t} \\cdot \\tilde {\\mathbf {z}} _ {i}}{\\| \\mu_ {j} ^ {t} \\| \\| \\tilde {\\mathbf {z}} _ {i} \\|} \\geq \\tau , \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 233, + 382, + 482, + 421 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tau$ denotes a pre-defined threshold. For paired local identity tokens, we update the corresponding global identity token. For mismatched local identity tokens, we pair them two by two using Eq. (7) and initialize new global identity tokens. This process can be formalized as:", + "bbox": [ + 89, + 430, + 483, + 506 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {z}} _ {i} ^ {*} = \\left\\{ \\begin{array}{l l} \\frac {n _ {\\bar {\\mathbf {z}} _ {i}} \\cdot \\tilde {\\mathbf {z}} _ {i} + \\sum_ {j = 1} ^ {m} n _ {j} ^ {t} \\cdot \\mu_ {j} ^ {t}}{n _ {\\bar {\\mathbf {z}} _ {i}} + \\sum_ {j = 1} ^ {m} n _ {j} ^ {t}}, & \\text {i f} \\tilde {\\mathbf {z}} _ {i} \\text {e x i s t s}, \\\\ \\frac {\\sum_ {j = 1} ^ {m} n _ {j} ^ {t} \\cdot \\mu_ {j} ^ {t}}{\\sum_ {j = 1} ^ {m} n _ {j} ^ {t}}, & \\text {i f} \\tilde {\\mathbf {z}} _ {i} \\text {d o e s n ’ t e x i s t}, \\end{array} \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 517, + 482, + 570 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $m$ , $n_{\\overline{\\mathbf{z}}_i}$ and $n_j^t$ denote the number of matched local identity tokens, the number of samples used to form the previous $i$ -th global identity token, and the number of samples from the local client $j$ at task $t$ , respectively. Then, we leverage this matching process to guide the update or initialization of task-specific subspace $\\{\\theta_i = \\mathbf{B}_i\\mathbf{A}_i\\}$ in Eq. (5):", + "bbox": [ + 89, + 580, + 483, + 674 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {i} ^ {*} = \\mathcal {F} \\left(\\theta_ {1} ^ {t}, \\dots , \\theta_ {m} ^ {t}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 684, + 482, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{F}$ denotes the FL algorithm (e.g. FedAvg [38]) used to aggregate local weights $^2$ . As a result, we effectively prevent inter-task conflicts and integrate knowledge from different clients using the identity token matching mechanism.", + "bbox": [ + 89, + 713, + 482, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After the global server completes the aggregation, it distributes the dynamic cache to each selected client. The client then matches the identity token (Using Eq. (7)) of each subspace with its own training data, deciding whether to update the corresponding subspace or reinitialize a new one. The entire DKO process is illustrated in Figure 2.", + "bbox": [ + 89, + 773, + 483, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Subspace Selective Activation", + "text_level": 1, + "bbox": [ + 511, + 277, + 777, + 294 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Section 4.1 effectively mitigates inter-task conflicts by disentangling and organizing complex knowledge into distinct subspaces. The key challenge then lies in how to leverage these task-specific subspaces during inference.", + "bbox": [ + 511, + 300, + 905, + 359 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address this, an intuitive method is to concatenate the subspaces of the dynamic cache in low-rank dimensions [50, 52], enabling the integration of knowledge across all task spaces. However, this may introduce redundant information unrelated to the current task during inference, potentially compromising the model's output. For instance, when the desired answer is a simple word (e.g., what is the object in the picture), knowledge from other subspaces, such as those used for generating long-form descriptions, can introduce unnecessary information, leading to responses that do not align with the given instruction.", + "bbox": [ + 511, + 361, + 905, + 526 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Drawing inspiration from LoRA's intrinsic space [53], we propose subspace selective activation (SSA) without additional training to filter out irrelevant subspace outputs, ensuring alignment between responses and instructions. In particular, A vanilla LoRA can be decomposed as the product of two low-rank subspaces (i.e., $\\mathbf{A}$ , $\\mathbf{B}$ ) and an intrinsic mixing matrix $\\mathcal{W}$ :", + "bbox": [ + 511, + 527, + 905, + 632 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\mathbf {W} = \\mathbf {B} \\mathcal {W} \\mathbf {A}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 658, + 641, + 903, + 656 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{W} \\in \\mathbb{R}^{r \\times r}$ is typically an identity matrix and can thus be omitted. In this paper, we redefine $\\mathcal{W}$ as the activation matrix that dynamically responds to the test input. Specifically, we treat $\\mathcal{W}$ as the product of an identity matrix and an activation factor (i.e., $\\mathcal{W} = \\alpha \\cdot \\mathbf{I}_{r \\times r}$ ), where $\\alpha = 1$ denotes full activation, and $\\alpha = 0$ means that the output is fully masked. Therefore, for each subspace $\\{\\mathbf{B}_1\\mathbf{A}_1, \\dots, \\mathbf{B}_T\\mathbf{A}_T\\}$ in dynamic cache, we can control the activation or inhibition of the corresponding output by adjusting its activation factor $\\{\\alpha_1, \\dots, \\alpha_T\\}$ .", + "bbox": [ + 511, + 666, + 905, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To provide better flexibility in assigning activation factors, we use global identity tokens and test input features extracted by the text encoder for similarity matching:", + "bbox": [ + 511, + 816, + 905, + 863 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns _ {i} = \\frac {\\tilde {\\mathbf {z}} _ {\\mathbf {i}} \\cdot f _ {i n s} \\left(\\mathbf {x} _ {i n s} ^ {t e s t}\\right)}{\\| \\tilde {\\mathbf {z}} _ {\\mathbf {i}} \\| \\cdot \\| f _ {i n s} \\left(\\mathbf {x} _ {i n s} ^ {t e s t}\\right) \\|}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 619, + 869, + 903, + 905 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "2In Section 5.4, we implement more FL algorithms to demonstrate the compatibility of our method.", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1329", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/95ed1ae71b6a1d5015452e1f6164b6db905751ff7f82d79319861fcedb6627bd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Dataset settingCapability-related (4 task)Task-related (8 task)
Partitionβ = 0.5β = 1.0β = 5.0β = 0.5β = 1.0β = 5.0
MethodsLastAvgLastAvgLastAvgLastAvgLastAvgLastAvg
Zero-shot30.57-30.57-30.57-29.08-29.08-29.08-
Individual61.75-59.62-60.27-64.21-63.85-64.07-
Centralized MTL63.83-63.83-63.83-66.60-66.60-66.60-
Finetune48.6860.0549.6759.0150.4058.0846.2468.9347.2068.7948.0069.97
EWC49.1260.1349.4659.2849.8958.7647.5169.0447.9269.2248.1570.27
LwF48.8760.3250.0259.8850.1559.0647.8969.5747.6269.1448.2670.30
L2P48.2259.7949.5659.3449.6259.2547.3069.3148.0869.6548.4270.16
O-LoRA51.6560.1949.1357.9350.2958.1852.8771.5449.8770.2647.7670.84
M-LoRA49.0460.0850.3960.7650.5658.0650.6871.9448.5371.5848.3871.21
MoELoRA49.6961.0050.9060.1850.4359.1149.2370.9649.0270.6548.8271.08
DISCO53.7362.0055.4762.0755.0660.5357.6974.0356.2273.0355.5872.64
", + "bbox": [ + 114, + 88, + 885, + 325 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Last and Avg performance of different methods on Hom-FCIT setting. The best performance is shown in bold.", + "bbox": [ + 142, + 334, + 852, + 349 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b2202c5f0b067b022c91b6c9c651798aee492216ef68e6307e0554fb32338fa2.jpg", + "image_caption": [ + "Figure 4. Performance curves of different methods in Hom-FCIT across seen tasks under varying data heterogeneity. We plot the average performance across all seen tasks at each stage." + ], + "image_footnote": [], + "bbox": [ + 117, + 354, + 334, + 502 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a78c15ea7723b5e842794c4d1f000cf743c20b5df6ebd16a2d30b2ea1abd4895.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 338, + 354, + 553, + 502 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/549f69b58d50b25e8cd8ea9bede745cac61ec68a5e506a528ae30003b5478c11.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 354, + 875, + 502 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $s_i$ denotes the similarity between the $i$ -th global identity token and the feature of the test instruction $\\mathbf{x}_{ins}^{test}$ . The activation factor $\\alpha_i$ is then computed by applying softmax normalization over all similarities:", + "bbox": [ + 89, + 550, + 482, + 609 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {i} = \\frac {\\exp \\left(s _ {i} / \\varepsilon\\right)}{\\sum_ {j = 1} ^ {T} \\exp \\left(s _ {j} / \\varepsilon\\right)}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 613, + 482, + 650 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\varepsilon$ is the temperature coefficient.", + "bbox": [ + 89, + 655, + 344, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Overall, our proposed SSA can be formulated as follows:", + "bbox": [ + 107, + 669, + 482, + 685 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\Delta \\mathbf {W} = \\bar {\\mathbf {B}} \\bar {\\mathcal {W}} \\bar {\\mathbf {A}} \\\\ = \\bar {\\mathbf {B}} \\left[ \\begin{array}{c c c c} \\alpha_ {1} \\cdot \\mathbf {I} _ {r \\times r} & \\mathbf {0} _ {r \\times r} & \\dots & \\mathbf {0} _ {r \\times r} \\\\ \\mathbf {0} _ {r \\times r} & \\alpha_ {2} \\cdot \\mathbf {I} _ {r \\times r} & \\dots & \\mathbf {0} _ {r \\times r} \\\\ \\vdots & \\vdots & \\ddots & \\vdots \\\\ \\mathbf {0} _ {r \\times r} & \\mathbf {0} _ {r \\times r} & \\dots & \\alpha_ {T} \\cdot \\mathbf {I} _ {r \\times r} \\end{array} \\right] \\bar {\\mathbf {A}}, \\tag {13} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 690, + 482, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\bar{\\mathbf{B}}\\in \\mathbb{R}^{d\\times (r\\cdot T)}$ represents the concatenation of all $\\{\\mathbf{B}_1,\\dots ,\\mathbf{B}_T\\}$ in the dynamic cache along the low-rank dimension $r$ , and $\\bar{\\mathbf{A}}\\in \\mathbb{R}^{(r\\cdot T)\\times k}$ follows the same structure. The activation factors $\\{\\alpha_{1},\\dots ,\\alpha_{T}\\}$ , computed via Eq. (11) and Eq. (12), control each subspace's output by amplifying matched subspaces while suppressing irrelevant ones. The schematic of SSA is provided in Figure 3.", + "bbox": [ + 89, + 794, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 549, + 645, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 574, + 702, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. The dataset composition is detailed in Section 3.2. For the experimental setup, we implement two types of dataset-level settings: Capability-related and Task-related, based on the two client-level realistic scenarios, Hom-FCIT and Het-FCIT, respectively. At each stage, the own data of local clients is partitioned according to the Dirichlet distribution [27], with three partitions $\\beta$ for each setting to model different levels of data heterogeneity. More details are provided in Appendix A.", + "bbox": [ + 511, + 595, + 906, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines & Evaluation Metrics. We compare our method with continual learning methods such as LwF [29], EWC [24], L2P [51], O-LoRA [50], MoELoRA [6], and also with the LoRA merging method from the federated continual learning approach PILoRA [12], referred to as M-LoRA. We follow a rehearsal-free continual learning setting [62], where only the data for the current task is available. All comparison methods are carefully calibrated to ensure the fairness of evaluations.", + "bbox": [ + 511, + 733, + 906, + 868 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the evaluation metrics, we report the standard metrics to measure the model performance: Last refers to the", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1330", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/10a43a7c910c7ca056c65bbb2e99fcfa58e35b19c1fcd35e6eeb26aea19591de.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Dataset settingCapability-related (4 task)Task-related (8 task)
Partitionβ = 0.5β = 1.0β = 5.0β = 0.5β = 1.0β = 5.0
MethodsLastAvgLastAvgLastAvgLastAvgLastAvgLastAvg
Zero-shot30.57-30.57-30.57-29.08-29.08-29.08-
Individual61.75-59.62-60.27-64.21-63.85-64.07-
Centralized MTL63.83-63.83-63.83-66.60-66.60-66.60-
Finetune55.6555.8256.3456.8556.7457.1558.0453.0557.9654.2258.1754.87
EWC55.2154.5855.1355.7655.6956.0258.7453.7657.9654.1857.4454.30
LwF55.9256.1856.4256.8056.8157.2658.8253.7758.0154.3458.2254.78
L2P56.1056.6356.7657.0256.9557.3658.8453.8058.7354.6758.3954.66
O-LoRA58.2458.5858.3258.6558.6058.9459.6154.5559.7455.2059.5154.97
M-LoRA57.7658.0257.6557.8957.8058.1759.7654.1158.8254.0359.3554.89
MoELoRA57.6857.9557.7758.0058.1558.4459.0254.2559.1454.6958.8654.50
DISCO59.4060.0159.9459.9159.7160.1662.0859.6463.2561.9962.7860.87
", + "bbox": [ + 114, + 88, + 885, + 325 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Last and Avg performance of different methods on Het-FCIT setting. The best performance is shown in bold.", + "bbox": [ + 143, + 334, + 846, + 349 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/04b024c7588837943e90cbecd61b2dbb677a73bdbb08184309f889f7404c0219.jpg", + "image_caption": [ + "Figure 5. Performance curve of different methods in the Het-FCIT under varying degrees of data heterogeneity. We plot the average performance across all tasks at each stage. The black dashed line marks the stage where the model has encountered all tasks." + ], + "image_footnote": [], + "bbox": [ + 117, + 356, + 334, + 501 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2c2bb57e7731e238b1bf640d4bae8e55cb25d5709ceb01b1ee351eea830f7a09.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 338, + 356, + 553, + 501 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1772929cc40a3675df940e8e9ae9519aa576e5169fb6d15fcd7dee50dd9816c1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 356, + 875, + 502 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "average result across all learned tasks after the completion of learning the final task. $Avg$ is based on Last, which tracks the performance of the learned tasks at each stage of the continual learning process and reports the average results.", + "bbox": [ + 88, + 550, + 482, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation Details. We choose LLaVA-1.5-7b [32] as the base LMM due to its structural simplicity and adopt its LoRA fine-tuning strategy, training only the LoRAs and multimodal projector. The LoRA rank $r$ at each stage is set to 8, with a learning rate of 2e-4, and it is embedded solely in the FFN layers of each block. The learning rate for the projector is set to 2e-5. For identity token extraction, we use the frozen CLIP text encoder [41]. The threshold $\\tau$ and temperature coefficient $\\varepsilon$ are set to 0.9 and 0.05, respectively.", + "bbox": [ + 88, + 611, + 482, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For the setting of FCIT, we set the epoch to 1 and communication rounds to 10 in each stage, and the Dirichlet distribution coefficients $\\beta$ are set to $\\{0.5, 1.0, 5.0\\}$ . In each round, the global server randomly selects 5 clients from a pool of 50 to participate in the training, and we use FedAvg [38] as the base FL aggregation algorithm.", + "bbox": [ + 88, + 747, + 482, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Main Results", + "text_level": 1, + "bbox": [ + 89, + 848, + 228, + 862 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results are shown in Table 1, Table 2, Figure 4, and Figure 5. Under Hom-FCIT setting, our proposed DISCO", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "achieves the best performance in both capability-related and task-related dataset settings, surpassing the best baseline by an average of $4.84\\%$ and $1.43\\%$ in the Last and Avg metrics, respectively. As illustrated in Figure 4, DISCO effectively mitigates the forgetting of previous tasks while learning new ones, outperforming other methods in the challenging long-stage continual learning setting.", + "bbox": [ + 511, + 550, + 906, + 655 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Het-FCIT setting, we plot the upward curve of the average performance across all tasks in Figure 5. As shown in the figure. Before completing all tasks (Left of the black dotted line), our method maintains the fastest rate of improvement, significantly outperforming other methods. In the subsequent phases (Right of the black dashed line), it continues to consolidate previously learned knowledge and further enhance performance, demonstrating strong adaptability in dynamic real-world scenarios. On the Last and Avg metrics, our method outperforms the best comparison method by an average of $2.17\\%$ and $3.62\\%$ , respectively.", + "bbox": [ + 511, + 657, + 908, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Additionally, our method demonstrates superior robustness to varying degrees of data heterogeneity, adapting well to diverse distributions and maintaining strong performance even as heterogeneity increases, highlighting its reliability in dynamic real-world scenarios.", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1331", + "bbox": [ + 483, + 945, + 511, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/de43b4405e7f8063244bd1baf64accc0840ecae63e87a2dd430335ea936f8cf6.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3. Ablation studies on (a) identity token extraction methods; (b) calculation of activation factors in SSA; (c) location of LoRA embedding. All experiments were conducted in the task-related setting with $\\beta = 1.0$ ." + ], + "table_body": "
MethodHom-FCITHet-FCIT
LastΔAvgΔLastΔAvgΔ
(a)Text56.220.073.030.063.250.061.990.0
Image55.63-0.5972.13-0.9063.00-0.2561.80-0.19
Text & Image55.96-0.2672.68-0.3563.02-0.2361.78-0.21
(b)Softmax56.220.073.030.063.250.061.990.0
Concatenate51.74-4.4869.20-3.8360.36-2.8959.88-2.11
Cosine sim52.83-3.3970.07-2.9660.92-2.3360.13-1.86
Argmax55.74-0.4872.07-0.9662.88-0.3761.42-0.57
(c)FFN56.220.073.030.063.250.061.990.0
Attn56.03-0.1972.88-0.1563.01-0.2461.83-0.16
FFN & Attn56.30+0.0872.96-0.0763.18-0.0762.04+0.05
", + "bbox": [ + 94, + 88, + 482, + 224 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 301, + 243, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The detailed ablation studies are provided in Table 3. In this paper, we use CLIP's text encoder to extract textual features as identity tokens. This section tests two alternatives: using only visual features from CLIP's visual encoder, and combining visual and textual features. From Table 3a, we observe that both alternatives degrade performance. This is potential because, in visual instruction tuning datasets, task similarities at the image level are higher than at the textual level (e.g. CLEVR-Math and super-CLEVR), making visual features less effective than textual ones.", + "bbox": [ + 89, + 323, + 483, + 473 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3b shows the ablation studies on how to calculate the activation factors in SSA. Direct concatenation, where the activation factors of all subspaces are set to 1.0, leads to significant performance degradation due to unfiltered outputs. Using cosine similarity (Eq.(11)) to filter the outputs also fails to solve the issue. While selecting the largest similarity (Argmax) can filter out irrelevant information, it risks fully activating a subspace that, though similar, is not directly related to the current task. This can negatively impact the output, as there are inherent similarities between the textual information of different tasks. In contrast, our method normalizes the cosine similarity to better focus on the relevant task output, achieving optimal performance.", + "bbox": [ + 89, + 474, + 483, + 671 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also test the location of LoRA embedding. As can be seen in Table 3c, embedding in the attention layer alone proved less effective than in the FFN layer. Moreover, embedding LoRA in every linear layer did not offer significant improvement and resulted in higher parameter transmission. Therefore, we chose to embed LoRA only in the FFN layer as a balanced solution. We provide more ablation and visualization results in Appendix B.", + "bbox": [ + 89, + 671, + 483, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Further Analysis", + "text_level": 1, + "bbox": [ + 89, + 801, + 259, + 819 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Compatible with other FL algorithms. In this paper, we use the classical FedAvg as the FL algorithm for global server aggregation of local weights. Additionally, we implement other federated learning algorithms, including FedAvgM [17], FedAdam [42], FedAdagrad [42], and", + "bbox": [ + 89, + 825, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b568094367a773e86eebfc4344d7c3bdd79ca2a0d356a4e57a7a29a79ff8bf7e.jpg", + "table_caption": [], + "table_footnote": [ + "Table 4. Results of different FL algorithms. All experiments were conducted in the task-related setting with $\\beta = 1.0$ ." + ], + "table_body": "
MethodHom-FCITHet-FCIT
LastAvgLastAvg
FedAvg56.2273.0363.2561.99
FedAvgM56.0772.5962.4761.10
FedAdam55.7672.9162.7861.35
FedAdagrad56.0272.8062.5561.41
FedYogi56.1173.1662.8861.67
", + "bbox": [ + 547, + 88, + 877, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fa06c48127ce5152049f999ae4e2d0149a8d8058956cfc623619cb4ac5987f1f.jpg", + "image_caption": [ + "Figure 6. Visualization of activation factors during inference." + ], + "image_footnote": [], + "bbox": [ + 575, + 272, + 861, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "FedYogi [42], to extend our framework. As shown in Table 4, FedAvg achieves the best average performance.", + "bbox": [ + 511, + 513, + 905, + 544 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization of activation factors. In Figure 6, we plot the responses of activations to the test inputs of different tasks. As observed, only the activator corresponding to the current task is responsive, effectively activating the relevant subspace, while the others remain largely inhibited. This confirms the effectiveness of our proposed SSA.", + "bbox": [ + 511, + 545, + 905, + 636 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 648, + 633, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have explored for the first time the integration of federated learning and continual learning for the instruction tuning of LMMs, addressing the real-world challenge of dynamically acquiring new knowledge through distributed training resources and data. Our proposed FCIT benchmark encompasses 2 real-world scenarios, 4 distinct settings, and 12 curated datasets, providing a comprehensive evaluation of different methods. Additionally, we introduce the DISCO framework, which leverages dynamic knowledge organization (DKO) to decompose inter-task conflicts and subspace selective activation (SSA) to assign task-relevant outputs while suppressing irrelevant information. Extensive experiments demonstrate that our approach significantly improves the model's ability to learn new knowledge and handle data heterogeneity in real-world scenarios.", + "bbox": [ + 511, + 674, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1332", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the National Science and Technology Major Project (2022ZD0116500), National Natural Science Foundation of China (62222609, 62320106010), CAS Project for Young Scientists in Basic Research (YSBR-083), Strategic Priority Research Program of Chinese Academy of Sciences under Grant (XDA0480200), Major Science and Technology Plan Project on the Future Industry Fields of Xiamen City (3502Z20241027), Unveiling and Leading Projects of Xiamen (3502Z20241011) and the InnoHK program.", + "bbox": [ + 89, + 114, + 485, + 267 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 277, + 187, + 292 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sara Babakniya, Zalan Fabian, Chaoyang He, Mahdi Soltanolkotabi, and Salman Avestimehr. A data-free approach to mitigate catastrophic forgetting in federated class incremental learning for vision tasks. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.1, 3", + "[3] Abhijit Bendale and Terrance Boult. Towards open world recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1893-1902, 2015. 1", + "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1, 3", + "[5] Meng Cao, Yuyang Liu, Yingfei Liu, Tiancai Wang, Jiahua Dong, Henghui Ding, Xiangyu Zhang, Ian Reid, and Xiaodan Liang. Continual llava: Continual instruction tuning in large vision-language models. arXiv preprint arXiv:2411.02564, 2024. 3, 4", + "[6] Cheng Chen, Junchen Zhu, Xu Luo, Hengtao Shen, Lianli Gao, and Jingkuan Song. Coin: A benchmark of continual instruction tuning for multimodel large language model. arXiv preprint arXiv:2403.08350, 2024. 3, 4, 6", + "[7] Haokun Chen, Yao Zhang, Denis Krompass, Jindong Gu, and Volker Tresp. Feddat: An approach for foundation model finetuning in multi-modal heterogeneous federated learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11285-11293, 2024. 1", + "[8] Adam Dahlgren Lindstrom and Savitha Sam Abraham. Clevr-math: A dataset for compositional language, visual and mathematical reasoning. In 16th International Workshop on Neural-Symbolic Learning and Reasoning, NeSy 2022, Windsor, UK, september 28-30, 2022. CEUR-WS, 2022. 4", + "[9] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500, 2023. 3" + ], + "bbox": [ + 99, + 301, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 3", + "[11] Jiahua Dong, Lixu Wang, Zhen Fang, Gan Sun, Shichao Xu, Xiao Wang, and Qi Zhu. Federated class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10164-10173, 2022. 1", + "[12] Haiyang Guo, Fei Zhu, Wenzhuo Liu, Xu-Yao Zhang, and Cheng-Lin Liu. Pilora: Prototype guided incremental lora for federated class-incremental learning. In European Conference on Computer Vision, pages 141-159. Springer, 2024. 1, 3, 6", + "[13] Haiyang Guo, Fei Zhu, Fanhu Zeng, Bing Liu, and XuYao Zhang. Desire: Dynamic knowledge consolidation for rehearsal-free continual learning. arXiv preprint arXiv:2411.19154, 2024. 2", + "[14] Haiyang Guo, Fanhu Zeng, Ziwei Xiang, Fei Zhu, Da-Han Wang, Xu-Yao Zhang, and Cheng-Lin Liu. Hidellava: Hierarchical decoupling for continual instruction tuning of multimodal large language model. arXiv preprint arXiv:2503.12941, 2025. 1", + "[15] Haiyang Guo, Fanhu Zeng, Fei Zhu, Jiayi Wang, Xukai Wang, Jingang Zhou, Hongbo Zhao, Wenzhuo Liu, Shijie Ma, Xu-Yao Zhang, et al. A comprehensive survey on continual learning in generative models. arXiv preprint arXiv:2506.13045, 2025. 1", + "[16] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, Dawn Song, Jacob Steinhardt, and Justin Gilmer. The many faces of robustness: A critical analysis of out-of-distribution generalization. ICCV, 2021. 4", + "[17] Tzu-Ming Harry Hsu, Hang Qi, and Matthew Brown. Measuring the effects of non-identical data distribution for federated visual classification. arXiv preprint arXiv:1909.06335, 2019.8", + "[18] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3", + "[19] Wenke Huang, Mang Ye, Zekun Shi, He Li, and Bo Du. Rethinking federated learning with domain shift: A prototype view. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16312-16322. IEEE, 2023. 4", + "[20] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 1, 3", + "[21] Kushal Kafle, Brian Price, Scott Cohen, and Christopher Kanan. Dvqa: Understanding data visualizations via question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5648-5656, 2018. 4", + "[22] Samira Ebrahimi Kahou, Vincent Michalski, Adam Atkinson, Ákos Kádár, Adam Trischler, and Yoshua Bengio. Fig-" + ], + "bbox": [ + 516, + 92, + 906, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1333", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ureqa: An annotated figure dataset for visual reasoning. arXiv preprint arXiv:1710.07300, 2017. 4", + "[23] Gyuhak Kim, Changnan Xiao, Tatsuya Konishi, and Bing Liu. Learnability and algorithm for continual learning. In International Conference on Machine Learning, pages 16877-16896. PMLR, 2023. 2", + "[24] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 6", + "[25] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 4", + "[26] Lei Li, Yuqi Wang, Runxin Xu, Peiyi Wang, Xiachong Feng, Lingpeng Kong, and Qi Liu. Multimodal arxiv: A dataset for improving scientific comprehension of large vision-language models. arXiv preprint arXiv:2403.00231, 2024. 4", + "[27] Qinbin Li, Yiqun Diao, Quan Chen, and Bingsheng He. Federated learning on non-iid data silos: An experimental study. In 2022 IEEE 38th international conference on data engineering (ICDE), pages 965-978. IEEE, 2022. 2, 6", + "[28] Tian Li, Anit Kumar Sahu, Ameet Talwalkar, and Virginia Smith. Federated learning: Challenges, methods, and future directions. IEEE signal processing magazine, 37(3):50-60, 2020. 1", + "[29] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 6", + "[30] Zhuowan Li, Xingrui Wang, Elias Stengel-Eskin, Adam Kortylewski, Wufei Ma, Benjamin Van Durme, and Alan L Yuille. Super-clevr: A virtual benchmark to diagnose domain robustness in visual reasoning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14963-14973, 2023. 4", + "[31] Yi-Ming Lin, Yuan Gao, Mao-Guo Gong, Si-Jia Zhang, Yuan-Qiao Zhang, and Zhi-Yuan Li. Federated learning on multimodal data: A comprehensive survey. Machine Intelligence Research, 20(4):539-553, 2023. 1", + "[32] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 3, 7", + "[33] Wenzhuo Liu, Fei Zhu, Haiyang Guo, Longhui Wei, and Cheng-Lin Liu. Llava-c: Continual improved visual instruction tuning. arXiv preprint arXiv:2506.08666, 2025.1", + "[34] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, et al. Deepseek-vl: towards real-world vision-language understanding. arXiv preprint arXiv:2403.05525, 2024. 1, 3", + "[35] Pan Lu, Liang Qiu, Jiaqi Chen, Tony Xia, Yizhou Zhao, Wei Zhang, Zhou Yu, Xiaodan Liang, and Song-Chun Zhu. Iconqa: A new benchmark for abstract diagram understanding and visual language reasoning. In The 35th Conference on Neural Information Processing Systems (NeurIPS) Track on Datasets and Benchmarks, 2021. 4" + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Pan Lu, Liang Qiu, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, Tanmay Rajpurohit, Peter Clark, and Ashwin Kalyan. Dynamic prompt learning via policy gradient for semi-structured mathematical reasoning. arXiv preprint arXiv:2209.14610, 2022. 4", + "[37] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 4", + "[38] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 1, 5, 7", + "[39] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering by reading text in images. In 2019 international conference on document analysis and recognition (ICDAR), pages 947-952. IEEE, 2019. 4", + "[40] Bryan A. Plummer, Liwei Wang, Christopher M. Cervantes, Juan C. Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. IJCV, 123 (1):74–93, 2017. 4", + "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 5, 7", + "[42] Sashank Reddi, Zachary Charles, Manzil Zaheer, Zachary Garrett, Keith Rush, Jakub Konečny, Sanjiv Kumar, and H Brendan McMahan. Adaptive federated optimization. arXiv preprint arXiv:2003.00295, 2020. 8", + "[43] Lorenzo Sani, Alex Iacob, Zeyu Cao, Bill Marino, Yan Gao, Tomas Paulik, Wanru Zhao, William F Shen, Preslav Aleksandrov, Xinchi Qiu, et al. The future of large language model pre-training is federated. arXiv preprint arXiv:2405.10853, 2024. 1", + "[44] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge. In European conference on computer vision, pages 146–162. Springer, 2022. 4", + "[45] Donald Shenaj, Marco Toldo, Alberto Rigon, and Pietro Zanuttigh. Asynchronous federated continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5055-5063, 2023. 3", + "[46] Tianxiang Sun, Xiaotian Zhang, Zhengfu He, Peng Li, Qinyuan Cheng, Xiangyang Liu, Hang Yan, Yunfan Shao, Qiong Tang, Shiduo Zhang, et al. Moss: An open conversational large language model. Machine Intelligence Research, 21(5):888-905, 2024. 1", + "[47] Yue Tan, Guodong Long, Lu Liu, Tianyi Zhou, Qinghua Lu, Jing Jiang, and Chengqi Zhang. Fedproto: Federated prototype learning across heterogeneous clients. In Proceedings of" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1334", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the AAAI conference on artificial intelligence, pages 8432-8440, 2022. 4", + "[48] Yue Tan, Guodong Long, Jie Ma, Lu Liu, Tianyi Zhou, and Jing Jiang. Federated learning from pre-trained models: A contrastive learning approach. Advances in neural information processing systems, 35:19332-19344, 2022. 4", + "[49] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1, 3", + "[50] Xiao Wang, Tianze Chen, Qiming Ge, Han Xia, Rong Bao, Rui Zheng, Qi Zhang, Tao Gui, and Xuanjing Huang. Orthogonal subspace learning for language model continual learning. arXiv preprint arXiv:2310.14152, 2023. 5, 6", + "[51] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 139-149, 2022. 6", + "[52] Ziyao Wang, Zheyu Shen, Yexiao He, Guoheng Sun, Hongyi Wang, Lingjuan Lyu, and Ang Li. Flora: Federated fin-tuning large language models with heterogeneous low-rank adaptations. arXiv preprint arXiv:2409.05976, 2024. 5", + "[53] Taiqiang Wu, Jiahao Wang, Zhe Zhao, and Ngai Wong. Mixture-of-subspaces in low-rank adaptation. arXiv preprint arXiv:2406.11909, 2024. 5", + "[54] Xin Yang, Hao Yu, Xin Gao, Hao Wang, Junbo Zhang, and Tianrui Li. Federated continual learning via knowledge fusion: A survey. IEEE Transactions on Knowledge and Data Engineering, 2024. 3", + "[55] Rui Ye, Wenhao Wang, Jingyi Chai, Dihan Li, Zexi Li, Yinda Xu, Yaxin Du, Yanfeng Wang, and Siheng Chen. Openfedllm: Training large language models on decentralized private data via federated learning. In Proceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining, pages 6137-6147, 2024. 1, 3, 4", + "[56] Fanhu Zeng, Fei Zhu, Haiyang Guo, Xu-Yao Zhang, and Cheng-Lin Liu. Modalprompt: Dual-modality guided prompt for continual learning of large multimodal models. arXiv preprint arXiv:2410.05849, 2024. 1, 3, 4", + "[57] Jie Zhang, Chen Chen, Weiming Zhuang, and Lingjuan Lyu. Target: Federated class-continual learning via exemplar-free distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4782–4793, 2023. 1", + "[58] Jianyi Zhang, Saeed Vahidian, Martin Kuo, Chunyuan Li, Ruiyi Zhang, Tong Yu, Guoyin Wang, and Yiran Chen. Towards building the federatedgpt: Federated instruction tuning. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6915-6919. IEEE, 2024. 1, 3, 4", + "[59] Xu-Yao Zhang, Cheng-Lin Liu, and Ching Y Suen. Towards robust pattern recognition: A review. Proceedings of the IEEE, 108(6):894-922, 2020. 1", + "[60] Hongbo Zhao, Fei Zhu, Rundong Wang, Gaofeng Meng, and Zhaoxiang Zhang. Mllm-cl: Continual learning" + ], + "bbox": [ + 91, + 90, + 483, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for multimodal large language models. arXiv preprint arXiv:2506.05453, 2025.1", + "[61] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3", + "[62] Fei Zhu, Xu-Yao Zhang, Chuang Wang, Fei Yin, and ChengLin Liu. Prototype augmentation and self-supervision for incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5871-5880, 2021. 6", + "[63] Fei Zhu, Shijie Ma, Zhen Cheng, Xu-Yao Zhang, Zhaoxiang Zhang, and Cheng-Lin Liu. Open-world machine learning: A review and new outlooks. arXiv preprint arXiv:2403.01759, 2024. 1" + ], + "bbox": [ + 516, + 92, + 903, + 303 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1335", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Federated Continual Instruction Tuning/c95b9507-af86-4883-a6e8-057630ebe555_model.json b/2025/Federated Continual Instruction Tuning/c95b9507-af86-4883-a6e8-057630ebe555_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d3169770518b2e0b1b71af47e0adf3fe48cf49c0 --- /dev/null +++ b/2025/Federated Continual Instruction Tuning/c95b9507-af86-4883-a6e8-057630ebe555_model.json @@ -0,0 +1,2400 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.807, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.296, + 0.131, + 0.703, + 0.154 + ], + "angle": 0, + "content": "Federated Continual Instruction Tuning" + }, + { + "type": "text", + "bbox": [ + 0.206, + 0.179, + 0.797, + 0.216 + ], + "angle": 0, + "content": "Haiyang Guo\\(^{1,2}\\), Fanhu Zeng\\(^{2,3}\\), Fei Zhu\\(^{4}\\), Wenzhuo Liu\\(^{2,3}\\), Da-Han Wang\\(^{5}\\), Jian Xu\\(^{2,3}\\), Xu-Yao Zhang\\(^{1,2,3*}\\), Cheng-Lin Liu\\(^{1,2,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.217, + 0.787, + 0.234 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Advanced Interdisciplinary Sciences, UCAS \\(^{2}\\)MAIS, CASIA" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.234, + 0.903, + 0.251 + ], + "angle": 0, + "content": "\\(^{3}\\)School of Artificial Intelligence, UCAS \\(^{4}\\)Centre for Artificial Intelligence and Robotics, HKISI-CAS" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.252, + 0.886, + 0.269 + ], + "angle": 0, + "content": "\\(^{5}\\)FKLPRIU, School of Computer and Information Engineering, Xiamen University of Technology" + }, + { + "type": "text", + "bbox": [ + 0.155, + 0.271, + 0.846, + 0.288 + ], + "angle": 0, + "content": "{guohaiyang2023, zengfanhu2022, jian.xu} @ia.ac.cn, zhfei2018@gmail.com, {xyz, liucl} @nlpr.ia.ac.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.322, + 0.327, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.354, + 0.486, + 0.716 + ], + "angle": 0, + "content": "A vast amount of instruction tuning data is crucial for the impressive performance of Large Multimodal Models (LMMs), but the associated computational costs and data collection demands during supervised fine-tuning make it impractical for most researchers. Federated learning (FL) has the potential to leverage all distributed data and training resources to reduce the overhead of joint training. However, most existing methods assume a fixed number of tasks, while in real-world scenarios, clients continuously encounter new knowledge and often struggle to retain old tasks due to memory constraints. In this work, we introduce the Federated Continual Instruction Tuning (FCIT) benchmark to model this real-world challenge. Our benchmark includes two realistic scenarios, encompassing four different settings and twelve carefully curated instruction tuning datasets. To address the challenges posed by FCIT, we propose a dynamic knowledge organization to effectively integrate updates from different tasks during training and subspace selective activation to allocate task-specific output during inference. Extensive experimental results demonstrate that our proposed method significantly enhances model performance across varying levels of data heterogeneity and catastrophic forgetting. Code and dataset are released at https://github.com/Ghy0501/FCIT." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.747, + 0.222, + 0.763 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.772, + 0.484, + 0.88 + ], + "angle": 0, + "content": "Large Multimodal Models (LMMs) [2, 32, 34], which integrate Large Language Model [4, 20, 46, 49] with a visual encoder and multimodal projector to bridge visual and textual modalities, have exhibited impressive visual understanding and complex reasoning abilities. A crucial factor in this success is the supervised fine-tuning of LMMs using huge and diverse visual instruction-following data [32]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.324, + 0.906, + 0.46 + ], + "angle": 0, + "content": "to align with human preferences. However, collecting such vast amounts of training data and computational resources for joint fine-tuning is impractical for most researchers. Federated Learning (FL) [28, 31, 38, 43], as a decentralized paradigm, offers a viable alternative by leveraging distributed data and computational resources for local training while integrating local weights to produce a unified model. This paradigm accommodates constraints on storage and computation while ensuring privacy protection." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.907, + 0.884 + ], + "angle": 0, + "content": "Generally, most existing FL frameworks [7, 55, 58] are modeled in static, closed-world scenarios [3, 59], where a fixed set of tasks is predefined and unchanged. However, real-world applications are dynamic [59, 63], requiring models to continuously acquire new knowledge while retaining previously learned tasks. Taking a realistic health emergency as an example, different hospitals act as local clients when a major disease outbreak occurs. Large hospitals can leverage FL to collaboratively train on their case data, building a comprehensive virus knowledge base, while smaller clinics update their own cases using this global knowledge. Over time, both large and small hospitals need to keep updating the global knowledge base through continual learning methods to cope with the event. Additionally, the emergence of new strains necessitates the simultaneous integration of knowledge and response strategies to minimize potential losses. In this scenario, traditional FL methods struggle with newly arrived knowledge, while Continual Learning (CL) [14, 15, 33, 56, 60] methods alone do not facilitate knowledge sharing between clients. Only an organic integration of both can address this real-world problem. Recently, numerous Federated Continual Learning methods have emerged to address this challenge in traditional image classification tasks [11, 12, 57]. However, their methodological and task-setting limitations make them insufficient for current LMM applications. Therefore, a more comprehensive benchmark is needed to better simulate the practical application of LMMs in real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.905, + 0.9 + ], + "angle": 0, + "content": "In this work, we first establish a Federated Continual In" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.24, + 0.9 + ], + "angle": 0, + "content": "*Corresponding Author." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1325" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.093, + 0.891, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.415, + 0.908, + 0.445 + ], + "angle": 0, + "content": "Figure 1. Overview of FCIT benchmark. FCIT encompasses 2 real-world scenarios, 4 FCIT settings, and 12 curated datasets, providing a comprehensive simulation of LMMs instruction-following training in real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.457, + 0.485, + 0.806 + ], + "angle": 0, + "content": "struction Tuning (FCIT) benchmark to fill this gap. Specifically, we simulate two client-level realistic scenarios: (i) Homogeneous FCIT refers to an FL system learning a series of instruction tuning data sequentially, with different clients learning the same task at each stage. (ii) Heterogeneous FCIT, on the other hand, requires the FL system to collaborate on different tasks simultaneously, with different clients potentially learning different tasks in the same stage. Building on this, we define two settings of datasets for each scenario: Capability-related and Task-related. The former evaluates the model's ability to integrate multiple dimensions of instruction following in a short period, while the latter assesses the model's performance during long phases of continual learning. For dataset selection, we curate twelve instruction tuning datasets unseen during LMM fine-tuning or with low zero-shot performance, preventing information leakage [13, 23]. In addition, we introduce varying degrees of data heterogeneity for each setting to challenge the model's performance in the non-IID situation [27]. An illustration of our FCIT benchmark is provided in Figure 1. To the best of our knowledge, this is the first work to introduce a comprehensive benchmark for federated learning of LMM in a continual learning setting." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To effectively address the challenges posed by FCIT, we propose a novel Dynamic knowledge organization and Subspace seleCtive activatiOn (DISCO) framework. Specifically, we identify FCIT challenges into two types: conflicts between different tasks within the same stage and conflicts between old and new tasks across different stages." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.457, + 0.907, + 0.744 + ], + "angle": 0, + "content": "The former can cause catastrophic forgetting by altering the parameter space of previous tasks when learning new ones without access to past data, while the latter requires the model to integrate and organize knowledge from different tasks to derive unified representations. Therefore, we first propose Dynamic Knowledge Organization (DKO), which leverages a dynamic cache at the global server to store task-specific parameters. Using a unique identity token matching mechanism, it systematically organizes knowledge for different tasks into corresponding subspaces within the cache, effectively mitigating two types of conflicts. To better utilize the organized task subspaces in the dynamic cache, we introduce Subspace Selective Activation (SSA), which selectively activates subspaces relevant to the test input while filtering out irrelevant outputs, leading to significant performance improvements. Consequently, these designs enable our framework to efficiently tackle the data heterogeneity and catastrophic forgetting in FCIT. In summary, our major contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.75, + 0.907, + 0.826 + ], + "angle": 0, + "content": "- We present the first Federated Continual Instruction Tuning (FCIT) benchmark designed for real-world scenarios, providing a comprehensive evaluation of LMMs to continuously learn new knowledge using distributed data and training resources in real applications." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.826, + 0.909, + 0.886 + ], + "angle": 0, + "content": "- We propose a novel DISCO framework that integrates dynamic knowledge organization and subspace selective activation to efficiently address data heterogeneity and catastrophic forgetting in FCIT settings." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate that our method sig" + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.75, + 0.909, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1326" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.104, + 0.092, + 0.483, + 0.138 + ], + "angle": 0, + "content": "nificantly enhances model performance under data heterogeneity while minimizing the catastrophic forgetting, and achieves state-of-the-art performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.151, + 0.234, + 0.167 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.175, + 0.483, + 0.417 + ], + "angle": 0, + "content": "Large Multimodal Models. With the grand unification of Large Language Models (LLMs) [4, 20, 49] for various NLP tasks, Large Multimodal Models (LMMs) [2, 9, 32, 34, 61] emerge by extending LLMs to combine with visual encoder and multimodal projectors, demonstrating exceptional visual understanding and complex reasoning abilities. To better align with human preferences, these LMMs typically undergo further fine-tuning on extensive instruction-following data, ensuring they meet the demands of real-world applications [32]. However, in real-world scenarios, improving the performance of LMMs on new downstream tasks becomes a significant challenge without access to sufficient training data and computational resources. In this paper, we introduce the first Federated Continual Instruction Tuning (FCIT) benchmark to bridge this gap from the perspective of distributed training and continual learning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.418, + 0.483, + 0.689 + ], + "angle": 0, + "content": "Federated Continual Learning. In classical vision tasks, Federated Continual Learning (FCL) [54] aims to adapt the global model to new data while maintaining the knowledge of the old task. From the perspective of LMMs, research in this setting faces two main challenges: (1) Most of the methods are designed for traditional vision tasks (i.e., image classification). For instance, MFCL [1] employs a generative model to synthesize images of previously learned classes, thereby mitigating forgetting. PILoRA [12] introduces a prototype re-weight module to address the classifier bias caused by data heterogeneity and obtain unified knowledge through LoRA [18] fusion. Despite progress in classification tasks, complex designs remain challenging to adapt for LMMs research. (2) FCL follows a single data composition by splitting a dataset (e.g. ImageNet [10]) into different tasks based on classes, whereas LMMs face greater challenges in continual learning due to their diverse tasks with varying styles [5, 6, 56]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.811 + ], + "angle": 0, + "content": "This work pioneers a Federated Continual Instruction Tuning setup for LMMs and establishes diverse scenarios to simulate real-world applications comprehensively. Notably, AFCL [45] is most relevant to our work, as it enables clients to continuously learn multiple tasks in different orders and asynchronous time slots. However, it is specifically designed for image classification tasks, whereas our study focuses on the more widely used LMM." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.824, + 0.298, + 0.839 + ], + "angle": 0, + "content": "3. Problem Formulation" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.232, + 0.863 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Instruction tuning enhances LMM's ability to understand and execute human instructions by performing supervised" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.213 + ], + "angle": 0, + "content": "fine-tuning of pre-trained models on extensive datasets comprising instructions and responses. Formally, the instruction data \\(\\mathcal{D} = \\{(\\mathbf{x}_v^j,\\mathbf{x}_{ins}^j,\\mathbf{x}_{res}^j)^N\\}\\) consists of image input \\(\\mathbf{x}_v\\), instruction \\(\\mathbf{x}_{ins}\\) and response \\(\\mathbf{x}_{res}\\), where \\(N\\) represents the total number of samples. For clarity, given a simple image-instruction pair with a response of length \\(L\\), the objective of an LMM is to predict the next token autoregressively based on all preceding tokens:" + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.219, + 0.905, + 0.261 + ], + "angle": 0, + "content": "\\[\np \\left(\\mathbf {x} _ {r e s} \\mid \\mathbf {x} _ {v}, \\mathbf {x} _ {i n s}\\right) = \\prod_ {i = 1} ^ {L} p _ {\\theta} \\left(x _ {i} \\mid \\mathbf {x} _ {v}, \\mathbf {x} _ {i n s}, \\mathbf {x} _ {r e s, < i}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.266, + 0.907, + 0.327 + ], + "angle": 0, + "content": "where \\(\\theta\\) denotes the trainable parameters during fine-tuning, \\(\\mathbf{x}_{\\mathrm{res}, < i}\\) denotes all response tokens preceding the current prediction token \\(x_{i}\\). Then, the loss function of fine-tuning LMMs can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.334, + 0.905, + 0.378 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\theta} = - \\frac {1}{N} \\sum_ {j = 1} ^ {N} \\sum_ {i = 1} ^ {L _ {j}} \\log p _ {\\theta} \\left(x _ {i} ^ {j} \\mid \\mathbf {x} _ {v} ^ {j}, \\mathbf {x} _ {i n s} ^ {j}, \\mathbf {x} _ {r e s, < i} ^ {j}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.383, + 0.905, + 0.488 + ], + "angle": 0, + "content": "Federated learning framework typically comprises a global server and several local clients, all employing the same LMM with a shared homogeneous model architecture. In each communication round1, local clients train their models on own data and upload the updated weights to the global server for aggregation, enabling collaborative optimization of the global model while preserving data privacy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.489, + 0.906, + 0.593 + ], + "angle": 0, + "content": "Considering the excessive communication overhead of transferring the entire LMM between clients and the global server, we adopt LoRA [18] for efficient fine-tuning, balancing training overhead and implementation cost [55, 58]. Specifically, for a weight matrix \\(\\mathbf{W}_0\\in \\mathbb{R}^{d\\times k}\\), LoRA decomposes parameter updates \\(\\Delta \\mathbf{W}\\) during fine-tuning into two low-rank subspaces:" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.602, + 0.905, + 0.617 + ], + "angle": 0, + "content": "\\[\n\\mathbf {W} = \\mathbf {W} _ {0} + \\Delta \\mathbf {W} = \\mathbf {W} _ {0} + \\mathbf {B A}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.624, + 0.845, + 0.64 + ], + "angle": 0, + "content": "where \\(\\mathbf{B} \\in \\mathbb{R}^{d \\times r}\\), \\(\\mathbf{A} \\in \\mathbb{R}^{r \\times k}\\) and \\(r \\ll \\min\\{d, k\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.905, + 0.701 + ], + "angle": 0, + "content": "Continual learning aims to minimize the loss on the current task while retaining knowledge from previous tasks. Formally, given a sequence of datasets \\(\\mathcal{D}_1,\\mathcal{D}_2,\\dots ,\\mathcal{D}_T\\) , the optimization objective at task \\(t\\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.708, + 0.905, + 0.753 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\mathcal {L} (\\theta) = \\mathbb {E} _ {(x, y) \\sim \\mathcal {D} _ {t}} [ \\mathcal {L} (f _ {\\theta_ {t}} (x), y) ] + \\sum_ {i = 1} ^ {t - 1} \\epsilon_ {i}, \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.752, + 0.87, + 0.787 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\operatorname {s. t.} \\mathbb {E} _ {(x, y) \\sim \\mathcal {D} _ {i}} \\left[ \\mathcal {L} \\left(f _ {\\theta_ {t}} (x), y\\right) - \\mathcal {L} \\left(f _ {\\theta_ {t - 1}} (x), y\\right) \\right] \\leq \\epsilon_ {i}, \\\\ \\epsilon_ {i} \\geq 0; \\forall i \\in [ 1, \\dots , t - 1 ], \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.792, + 0.905, + 0.869 + ], + "angle": 0, + "content": "where \\(\\epsilon_{i}\\) is a slack variable that allows a small increase in the loss from the old datasets, providing tolerance for minor forgetting while focusing on learning the current task. In Eq. 4, \\(x\\) and \\(y\\) can be viewed as the multimodal inputs \\((\\mathbf{x}_v,\\mathbf{x}_{ins})\\) and the response \\(\\mathbf{x}_{res}\\), respectively." + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.905, + 0.9 + ], + "angle": 0, + "content": "For clarification, we standardize the communication round in Section 4 to 1." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1327" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.436, + 0.108 + ], + "angle": 0, + "content": "3.2. Federated Continual Instruction Tuning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.118, + 0.485, + 0.36 + ], + "angle": 0, + "content": "As shown in Figure 1, we integrate federated and continual learning for LMMs within a unified framework and propose two client-level realistic scenarios: (1) Homogeneous FCIT (Hom-FCIT). In this scenario, clients sequentially learn a series of tasks, allowing the global server to continuously update its knowledge. Each client learns the same task at a given stage and can only access data from that stage. (2) Heterogeneous FCIT (Het-FCIT). In real-world applications, different clients may learn different tasks within the same stage, enabling the global server to respond more rapidly to diverse instructions. This requires the model not only to coordinate the knowledge of different tasks learned by clients in the current stage but also to mitigate forgetting during the learning process. For each scenario, we define two dataset-level settings: capability-related and task-related, detailed as follows." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.373, + 0.484, + 0.539 + ], + "angle": 0, + "content": "- Capability-related. Following the dataset construction and division in LLaVA-OneVision [25], we classify the 12 datasets into 4 capabilities: General, Math, Chart, and Other. The General capability includes A-OKVQA [44], ImageNet-R [16], Grounding [37], and IconQA [35]; Math comprises CLEVR-Math [8], super-CLEVR [30], and TabMWP [36]; Chart involves ArxivQA [26], FigureQA [22], and DVQA [21]; and Other encompasses OCR-VQA [39] and Flickr30k [40]. We treat these four capabilities as different stages of continual learning, with each capability comprising a mixture of datasets." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.54, + 0.484, + 0.616 + ], + "angle": 0, + "content": "- Task-related. To evaluate the performance of different methods in a long-phase continual learning situation, we selected 8 datasets: ImageNet-R, ArxivQA, IconQA, CLEVR-Math, OCR-VQA, Flickr30k, FigureQA, and super-CLEVR as distinct stages of continual learning." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.373, + 0.484, + 0.616 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.484, + 0.705 + ], + "angle": 0, + "content": "In total, we propose 4 FCIT settings to evaluate different methods. Compared to existing works [5, 6, 55, 56, 58], we are the first to explore the organic integration of FL and CL in the context of LMM. We provide further illustrations on the settings and datasets in Appendix A." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.734, + 0.408, + 0.751 + ], + "angle": 0, + "content": "4. The Proposed Framework: DISCO" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Overview of the Method. As shown in Figure 2 and 3, our method primarily consists of: (1) Dynamic Knowledge Organization (DKO), which dynamically integrates knowledge learned by different clients across stages during training, significantly reducing inter-task conflicts (Section 4.1); and (2) Subspace Selective Activation (SSA), which selectively activates subspace outputs based on input features during inference, effectively filtering out irrelevant information (Section 4.2). We name this framework DISCO." + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.093, + 0.887, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.355, + 0.907, + 0.397 + ], + "angle": 0, + "content": "Figure 2. Illustration of the proposed DKO. Dynamic caches store the knowledge of each subspace while matching and updating through identity tokens." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.402, + 0.816, + 0.418 + ], + "angle": 0, + "content": "4.1. Dynamic Knowledge Organization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.423, + 0.907, + 0.59 + ], + "angle": 0, + "content": "The core challenge of FCIT is enabling the global server to effectively harmonize knowledge learned by clients, addressing both the forgetting of previous knowledge when acquiring new tasks and the conflicts that arise from integrating knowledge from different tasks within the same stage. To this end, we propose maintaining a dynamic cache at the global server to organize task-specific knowledge uploaded by clients at different stages, thereby preventing both forgetting and conflicts between knowledge. Specifically, each task is assigned a dedicated parameter space to store and update its corresponding knowledge:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.598, + 0.907, + 0.634 + ], + "angle": 0, + "content": "\\[\n\\Delta \\mathbf {W} = \\mathbf {B} \\mathbf {A} \\Leftrightarrow \\underbrace {\\left\\{\\mathbf {B} _ {1} \\mathbf {A} _ {1} , \\cdots , \\mathbf {B} _ {T} \\mathbf {A} _ {T} \\right\\}} _ {\\text {t a s k - s p e c i f i c s u b s p a c e}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.641, + 0.906, + 0.687 + ], + "angle": 0, + "content": "where \\( T \\) is the number of tasks learned. At this point, it is crucial to aggregate the parameters uploaded by clients into their respective subspaces without privacy leakage." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.687, + 0.906, + 0.822 + ], + "angle": 0, + "content": "Inspired by the widespread use of prototypes in FL fields [19, 47, 48], we propose to distinguish knowledge across tasks by using the feature mean of each client's training data as an Identity Token. Specifically, considering the uniqueness of textual inputs in the visual instruction tuning task, we introduce a text encoder \\( f_{ins} \\) for each client to extract the feature of the input instruction \\( \\mathbf{x}_{ins,j}^{t} \\) during training, and take the mean value \\( \\mu_k^t \\) as the local identity token of the \\( k \\)-th client on the task \\( t \\) at the end of training:" + }, + { + "type": "equation", + "bbox": [ + 0.62, + 0.832, + 0.905, + 0.876 + ], + "angle": 0, + "content": "\\[\n\\mu_ {k} ^ {t} = \\frac {1}{n _ {k} ^ {t}} \\sum_ {j = 1} ^ {n _ {k} ^ {t}} f _ {i n s} \\left(\\mathbf {x} _ {i n s, j} ^ {t}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.907, + 0.903 + ], + "angle": 0, + "content": "where \\( n_k^t \\) is the number of training samples of client \\( k \\) at" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1328" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.09, + 0.895, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.243, + 0.908, + 0.274 + ], + "angle": 0, + "content": "Figure 3. Illustration of the proposed SSA. Each subspace organized by DKO dynamically adjusts its output via the intrinsic activation matrix during inference, effectively filtering irrelevant information and enhancing model performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.28, + 0.483, + 0.325 + ], + "angle": 0, + "content": "task \\(t\\), and we use CLIP's text encoder [41] in our experiments. Notably, the text encoder \\(f_{ins}\\) remains frozen during training, which prevents additional training overhead." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.325, + 0.483, + 0.372 + ], + "angle": 0, + "content": "Using the identity tokens, we first apply cosine similarity along with a threshold to match the uploaded local identity token \\(\\mu_j^t\\) with the global identity token \\(\\tilde{\\mathbf{z}}_i\\):" + }, + { + "type": "equation", + "bbox": [ + 0.235, + 0.383, + 0.483, + 0.422 + ], + "angle": 0, + "content": "\\[\n\\frac {\\mu_ {j} ^ {t} \\cdot \\tilde {\\mathbf {z}} _ {i}}{\\| \\mu_ {j} ^ {t} \\| \\| \\tilde {\\mathbf {z}} _ {i} \\|} \\geq \\tau , \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.431, + 0.484, + 0.507 + ], + "angle": 0, + "content": "where \\(\\tau\\) denotes a pre-defined threshold. For paired local identity tokens, we update the corresponding global identity token. For mismatched local identity tokens, we pair them two by two using Eq. (7) and initialize new global identity tokens. This process can be formalized as:" + }, + { + "type": "equation", + "bbox": [ + 0.128, + 0.518, + 0.483, + 0.571 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {z}} _ {i} ^ {*} = \\left\\{ \\begin{array}{l l} \\frac {n _ {\\bar {\\mathbf {z}} _ {i}} \\cdot \\tilde {\\mathbf {z}} _ {i} + \\sum_ {j = 1} ^ {m} n _ {j} ^ {t} \\cdot \\mu_ {j} ^ {t}}{n _ {\\bar {\\mathbf {z}} _ {i}} + \\sum_ {j = 1} ^ {m} n _ {j} ^ {t}}, & \\text {i f} \\tilde {\\mathbf {z}} _ {i} \\text {e x i s t s}, \\\\ \\frac {\\sum_ {j = 1} ^ {m} n _ {j} ^ {t} \\cdot \\mu_ {j} ^ {t}}{\\sum_ {j = 1} ^ {m} n _ {j} ^ {t}}, & \\text {i f} \\tilde {\\mathbf {z}} _ {i} \\text {d o e s n ’ t e x i s t}, \\end{array} \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.484, + 0.675 + ], + "angle": 0, + "content": "where \\(m\\), \\(n_{\\overline{\\mathbf{z}}_i}\\) and \\(n_j^t\\) denote the number of matched local identity tokens, the number of samples used to form the previous \\(i\\)-th global identity token, and the number of samples from the local client \\(j\\) at task \\(t\\), respectively. Then, we leverage this matching process to guide the update or initialization of task-specific subspace \\(\\{\\theta_i = \\mathbf{B}_i\\mathbf{A}_i\\}\\) in Eq. (5):" + }, + { + "type": "equation", + "bbox": [ + 0.213, + 0.685, + 0.483, + 0.703 + ], + "angle": 0, + "content": "\\[\n\\theta_ {i} ^ {*} = \\mathcal {F} \\left(\\theta_ {1} ^ {t}, \\dots , \\theta_ {m} ^ {t}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.714, + 0.483, + 0.774 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}\\) denotes the FL algorithm (e.g. FedAvg [38]) used to aggregate local weights \\(^2\\). As a result, we effectively prevent inter-task conflicts and integrate knowledge from different clients using the identity token matching mechanism." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.775, + 0.484, + 0.866 + ], + "angle": 0, + "content": "After the global server completes the aggregation, it distributes the dynamic cache to each selected client. The client then matches the identity token (Using Eq. (7)) of each subspace with its own training data, deciding whether to update the corresponding subspace or reinitialize a new one. The entire DKO process is illustrated in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.279, + 0.779, + 0.295 + ], + "angle": 0, + "content": "4.2. Subspace Selective Activation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.301, + 0.906, + 0.361 + ], + "angle": 0, + "content": "Section 4.1 effectively mitigates inter-task conflicts by disentangling and organizing complex knowledge into distinct subspaces. The key challenge then lies in how to leverage these task-specific subspaces during inference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.362, + 0.906, + 0.527 + ], + "angle": 0, + "content": "To address this, an intuitive method is to concatenate the subspaces of the dynamic cache in low-rank dimensions [50, 52], enabling the integration of knowledge across all task spaces. However, this may introduce redundant information unrelated to the current task during inference, potentially compromising the model's output. For instance, when the desired answer is a simple word (e.g., what is the object in the picture), knowledge from other subspaces, such as those used for generating long-form descriptions, can introduce unnecessary information, leading to responses that do not align with the given instruction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.528, + 0.906, + 0.633 + ], + "angle": 0, + "content": "Drawing inspiration from LoRA's intrinsic space [53], we propose subspace selective activation (SSA) without additional training to filter out irrelevant subspace outputs, ensuring alignment between responses and instructions. In particular, A vanilla LoRA can be decomposed as the product of two low-rank subspaces (i.e., \\(\\mathbf{A}\\), \\(\\mathbf{B}\\)) and an intrinsic mixing matrix \\(\\mathcal{W}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.66, + 0.642, + 0.905, + 0.657 + ], + "angle": 0, + "content": "\\[\n\\Delta \\mathbf {W} = \\mathbf {B} \\mathcal {W} \\mathbf {A}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.667, + 0.906, + 0.818 + ], + "angle": 0, + "content": "where \\(\\mathcal{W} \\in \\mathbb{R}^{r \\times r}\\) is typically an identity matrix and can thus be omitted. In this paper, we redefine \\(\\mathcal{W}\\) as the activation matrix that dynamically responds to the test input. Specifically, we treat \\(\\mathcal{W}\\) as the product of an identity matrix and an activation factor (i.e., \\(\\mathcal{W} = \\alpha \\cdot \\mathbf{I}_{r \\times r}\\)), where \\(\\alpha = 1\\) denotes full activation, and \\(\\alpha = 0\\) means that the output is fully masked. Therefore, for each subspace \\(\\{\\mathbf{B}_1\\mathbf{A}_1, \\dots, \\mathbf{B}_T\\mathbf{A}_T\\}\\) in dynamic cache, we can control the activation or inhibition of the corresponding output by adjusting its activation factor \\(\\{\\alpha_1, \\dots, \\alpha_T\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.818, + 0.906, + 0.864 + ], + "angle": 0, + "content": "To provide better flexibility in assigning activation factors, we use global identity tokens and test input features extracted by the text encoder for similarity matching:" + }, + { + "type": "equation", + "bbox": [ + 0.62, + 0.871, + 0.905, + 0.906 + ], + "angle": 0, + "content": "\\[\ns _ {i} = \\frac {\\tilde {\\mathbf {z}} _ {\\mathbf {i}} \\cdot f _ {i n s} \\left(\\mathbf {x} _ {i n s} ^ {t e s t}\\right)}{\\| \\tilde {\\mathbf {z}} _ {\\mathbf {i}} \\| \\cdot \\| f _ {i n s} \\left(\\mathbf {x} _ {i n s} ^ {t e s t}\\right) \\|}, \\tag {11}\n\\]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "2In Section 5.4, we implement more FL algorithms to demonstrate the compatibility of our method." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1329" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.116, + 0.089, + 0.887, + 0.326 + ], + "angle": 0, + "content": "
Dataset settingCapability-related (4 task)Task-related (8 task)
Partitionβ = 0.5β = 1.0β = 5.0β = 0.5β = 1.0β = 5.0
MethodsLastAvgLastAvgLastAvgLastAvgLastAvgLastAvg
Zero-shot30.57-30.57-30.57-29.08-29.08-29.08-
Individual61.75-59.62-60.27-64.21-63.85-64.07-
Centralized MTL63.83-63.83-63.83-66.60-66.60-66.60-
Finetune48.6860.0549.6759.0150.4058.0846.2468.9347.2068.7948.0069.97
EWC49.1260.1349.4659.2849.8958.7647.5169.0447.9269.2248.1570.27
LwF48.8760.3250.0259.8850.1559.0647.8969.5747.6269.1448.2670.30
L2P48.2259.7949.5659.3449.6259.2547.3069.3148.0869.6548.4270.16
O-LoRA51.6560.1949.1357.9350.2958.1852.8771.5449.8770.2647.7670.84
M-LoRA49.0460.0850.3960.7650.5658.0650.6871.9448.5371.5848.3871.21
MoELoRA49.6961.0050.9060.1850.4359.1149.2370.9649.0270.6548.8271.08
DISCO53.7362.0055.4762.0755.0660.5357.6974.0356.2273.0355.5872.64
" + }, + { + "type": "table_caption", + "bbox": [ + 0.143, + 0.335, + 0.854, + 0.35 + ], + "angle": 0, + "content": "Table 1. Last and Avg performance of different methods on Hom-FCIT setting. The best performance is shown in bold." + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.356, + 0.335, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.339, + 0.356, + 0.554, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.356, + 0.876, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.508, + 0.907, + 0.537 + ], + "angle": 0, + "content": "Figure 4. Performance curves of different methods in Hom-FCIT across seen tasks under varying data heterogeneity. We plot the average performance across all seen tasks at each stage." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.483, + 0.61 + ], + "angle": 0, + "content": "where \\( s_i \\) denotes the similarity between the \\( i \\)-th global identity token and the feature of the test instruction \\( \\mathbf{x}_{ins}^{test} \\). The activation factor \\( \\alpha_i \\) is then computed by applying softmax normalization over all similarities:" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.614, + 0.483, + 0.651 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {i} = \\frac {\\exp \\left(s _ {i} / \\varepsilon\\right)}{\\sum_ {j = 1} ^ {T} \\exp \\left(s _ {j} / \\varepsilon\\right)}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.656, + 0.345, + 0.669 + ], + "angle": 0, + "content": "where \\(\\varepsilon\\) is the temperature coefficient." + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.67, + 0.483, + 0.686 + ], + "angle": 0, + "content": "Overall, our proposed SSA can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.691, + 0.483, + 0.794 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\Delta \\mathbf {W} = \\bar {\\mathbf {B}} \\bar {\\mathcal {W}} \\bar {\\mathbf {A}} \\\\ = \\bar {\\mathbf {B}} \\left[ \\begin{array}{c c c c} \\alpha_ {1} \\cdot \\mathbf {I} _ {r \\times r} & \\mathbf {0} _ {r \\times r} & \\dots & \\mathbf {0} _ {r \\times r} \\\\ \\mathbf {0} _ {r \\times r} & \\alpha_ {2} \\cdot \\mathbf {I} _ {r \\times r} & \\dots & \\mathbf {0} _ {r \\times r} \\\\ \\vdots & \\vdots & \\ddots & \\vdots \\\\ \\mathbf {0} _ {r \\times r} & \\mathbf {0} _ {r \\times r} & \\dots & \\alpha_ {T} \\cdot \\mathbf {I} _ {r \\times r} \\end{array} \\right] \\bar {\\mathbf {A}}, \\tag {13} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.902 + ], + "angle": 0, + "content": "where \\(\\bar{\\mathbf{B}}\\in \\mathbb{R}^{d\\times (r\\cdot T)}\\) represents the concatenation of all \\(\\{\\mathbf{B}_1,\\dots ,\\mathbf{B}_T\\}\\) in the dynamic cache along the low-rank dimension \\(r\\), and \\(\\bar{\\mathbf{A}}\\in \\mathbb{R}^{(r\\cdot T)\\times k}\\) follows the same structure. The activation factors \\(\\{\\alpha_{1},\\dots ,\\alpha_{T}\\}\\), computed via Eq. (11) and Eq. (12), control each subspace's output by amplifying matched subspaces while suppressing irrelevant ones. The schematic of SSA is provided in Figure 3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.55, + 0.646, + 0.567 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.575, + 0.704, + 0.591 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.597, + 0.907, + 0.734 + ], + "angle": 0, + "content": "Datasets. The dataset composition is detailed in Section 3.2. For the experimental setup, we implement two types of dataset-level settings: Capability-related and Task-related, based on the two client-level realistic scenarios, Hom-FCIT and Het-FCIT, respectively. At each stage, the own data of local clients is partitioned according to the Dirichlet distribution [27], with three partitions \\(\\beta\\) for each setting to model different levels of data heterogeneity. More details are provided in Appendix A." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.734, + 0.907, + 0.869 + ], + "angle": 0, + "content": "Baselines & Evaluation Metrics. We compare our method with continual learning methods such as LwF [29], EWC [24], L2P [51], O-LoRA [50], MoELoRA [6], and also with the LoRA merging method from the federated continual learning approach PILoRA [12], referred to as M-LoRA. We follow a rehearsal-free continual learning setting [62], where only the data for the current task is available. All comparison methods are carefully calibrated to ensure the fairness of evaluations." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "For the evaluation metrics, we report the standard metrics to measure the model performance: Last refers to the" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1330" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.116, + 0.089, + 0.887, + 0.326 + ], + "angle": 0, + "content": "
Dataset settingCapability-related (4 task)Task-related (8 task)
Partitionβ = 0.5β = 1.0β = 5.0β = 0.5β = 1.0β = 5.0
MethodsLastAvgLastAvgLastAvgLastAvgLastAvgLastAvg
Zero-shot30.57-30.57-30.57-29.08-29.08-29.08-
Individual61.75-59.62-60.27-64.21-63.85-64.07-
Centralized MTL63.83-63.83-63.83-66.60-66.60-66.60-
Finetune55.6555.8256.3456.8556.7457.1558.0453.0557.9654.2258.1754.87
EWC55.2154.5855.1355.7655.6956.0258.7453.7657.9654.1857.4454.30
LwF55.9256.1856.4256.8056.8157.2658.8253.7758.0154.3458.2254.78
L2P56.1056.6356.7657.0256.9557.3658.8453.8058.7354.6758.3954.66
O-LoRA58.2458.5858.3258.6558.6058.9459.6154.5559.7455.2059.5154.97
M-LoRA57.7658.0257.6557.8957.8058.1759.7654.1158.8254.0359.3554.89
MoELoRA57.6857.9557.7758.0058.1558.4459.0254.2559.1454.6958.8654.50
DISCO59.4060.0159.9459.9159.7160.1662.0859.6463.2561.9962.7860.87
" + }, + { + "type": "table_caption", + "bbox": [ + 0.145, + 0.335, + 0.848, + 0.35 + ], + "angle": 0, + "content": "Table 2. Last and Avg performance of different methods on Het-FCIT setting. The best performance is shown in bold." + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.357, + 0.335, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.339, + 0.357, + 0.554, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.357, + 0.876, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.508, + 0.907, + 0.537 + ], + "angle": 0, + "content": "Figure 5. Performance curve of different methods in the Het-FCIT under varying degrees of data heterogeneity. We plot the average performance across all tasks at each stage. The black dashed line marks the stage where the model has encountered all tasks." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.551, + 0.483, + 0.611 + ], + "angle": 0, + "content": "average result across all learned tasks after the completion of learning the final task. \\(Avg\\) is based on Last, which tracks the performance of the learned tasks at each stage of the continual learning process and reports the average results." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.612, + 0.483, + 0.746 + ], + "angle": 0, + "content": "Implementation Details. We choose LLaVA-1.5-7b [32] as the base LMM due to its structural simplicity and adopt its LoRA fine-tuning strategy, training only the LoRAs and multimodal projector. The LoRA rank \\( r \\) at each stage is set to 8, with a learning rate of 2e-4, and it is embedded solely in the FFN layers of each block. The learning rate for the projector is set to 2e-5. For identity token extraction, we use the frozen CLIP text encoder [41]. The threshold \\( \\tau \\) and temperature coefficient \\( \\varepsilon \\) are set to 0.9 and 0.05, respectively." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.748, + 0.483, + 0.839 + ], + "angle": 0, + "content": "For the setting of FCIT, we set the epoch to 1 and communication rounds to 10 in each stage, and the Dirichlet distribution coefficients \\(\\beta\\) are set to \\(\\{0.5, 1.0, 5.0\\}\\). In each round, the global server randomly selects 5 clients from a pool of 50 to participate in the training, and we use FedAvg [38] as the base FL aggregation algorithm." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.229, + 0.863 + ], + "angle": 0, + "content": "5.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Results are shown in Table 1, Table 2, Figure 4, and Figure 5. Under Hom-FCIT setting, our proposed DISCO" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.551, + 0.907, + 0.656 + ], + "angle": 0, + "content": "achieves the best performance in both capability-related and task-related dataset settings, surpassing the best baseline by an average of \\(4.84\\%\\) and \\(1.43\\%\\) in the Last and Avg metrics, respectively. As illustrated in Figure 4, DISCO effectively mitigates the forgetting of previous tasks while learning new ones, outperforming other methods in the challenging long-stage continual learning setting." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.909, + 0.824 + ], + "angle": 0, + "content": "In Het-FCIT setting, we plot the upward curve of the average performance across all tasks in Figure 5. As shown in the figure. Before completing all tasks (Left of the black dotted line), our method maintains the fastest rate of improvement, significantly outperforming other methods. In the subsequent phases (Right of the black dashed line), it continues to consolidate previously learned knowledge and further enhance performance, demonstrating strong adaptability in dynamic real-world scenarios. On the Last and Avg metrics, our method outperforms the best comparison method by an average of \\(2.17\\%\\) and \\(3.62\\%\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Additionally, our method demonstrates superior robustness to varying degrees of data heterogeneity, adapting well to diverse distributions and maintaining strong performance even as heterogeneity increases, highlighting its reliability in dynamic real-world scenarios." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.513, + 0.957 + ], + "angle": 0, + "content": "1331" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.483, + 0.226 + ], + "angle": 0, + "content": "
MethodHom-FCITHet-FCIT
LastΔAvgΔLastΔAvgΔ
(a)Text56.220.073.030.063.250.061.990.0
Image55.63-0.5972.13-0.9063.00-0.2561.80-0.19
Text & Image55.96-0.2672.68-0.3563.02-0.2361.78-0.21
(b)Softmax56.220.073.030.063.250.061.990.0
Concatenate51.74-4.4869.20-3.8360.36-2.8959.88-2.11
Cosine sim52.83-3.3970.07-2.9660.92-2.3360.13-1.86
Argmax55.74-0.4872.07-0.9662.88-0.3761.42-0.57
(c)FFN56.220.073.030.063.250.061.990.0
Attn56.03-0.1972.88-0.1563.01-0.2461.83-0.16
FFN & Attn56.30+0.0872.96-0.0763.18-0.0762.04+0.05
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.238, + 0.483, + 0.294 + ], + "angle": 0, + "content": "Table 3. Ablation studies on (a) identity token extraction methods; (b) calculation of activation factors in SSA; (c) location of LoRA embedding. All experiments were conducted in the task-related setting with \\(\\beta = 1.0\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.302, + 0.245, + 0.318 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.324, + 0.484, + 0.474 + ], + "angle": 0, + "content": "The detailed ablation studies are provided in Table 3. In this paper, we use CLIP's text encoder to extract textual features as identity tokens. This section tests two alternatives: using only visual features from CLIP's visual encoder, and combining visual and textual features. From Table 3a, we observe that both alternatives degrade performance. This is potential because, in visual instruction tuning datasets, task similarities at the image level are higher than at the textual level (e.g. CLEVR-Math and super-CLEVR), making visual features less effective than textual ones." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.476, + 0.484, + 0.672 + ], + "angle": 0, + "content": "Table 3b shows the ablation studies on how to calculate the activation factors in SSA. Direct concatenation, where the activation factors of all subspaces are set to 1.0, leads to significant performance degradation due to unfiltered outputs. Using cosine similarity (Eq.(11)) to filter the outputs also fails to solve the issue. While selecting the largest similarity (Argmax) can filter out irrelevant information, it risks fully activating a subspace that, though similar, is not directly related to the current task. This can negatively impact the output, as there are inherent similarities between the textual information of different tasks. In contrast, our method normalizes the cosine similarity to better focus on the relevant task output, achieving optimal performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.672, + 0.484, + 0.793 + ], + "angle": 0, + "content": "We also test the location of LoRA embedding. As can be seen in Table 3c, embedding in the attention layer alone proved less effective than in the FFN layer. Moreover, embedding LoRA in every linear layer did not offer significant improvement and resulted in higher parameter transmission. Therefore, we chose to embed LoRA only in the FFN layer as a balanced solution. We provide more ablation and visualization results in Appendix B." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.803, + 0.26, + 0.82 + ], + "angle": 0, + "content": "5.4. Further Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Compatible with other FL algorithms. In this paper, we use the classical FedAvg as the FL algorithm for global server aggregation of local weights. Additionally, we implement other federated learning algorithms, including FedAvgM [17], FedAdam [42], FedAdagrad [42], and" + }, + { + "type": "table", + "bbox": [ + 0.548, + 0.089, + 0.878, + 0.212 + ], + "angle": 0, + "content": "
MethodHom-FCITHet-FCIT
LastAvgLastAvg
FedAvg56.2273.0363.2561.99
FedAvgM56.0772.5962.4761.10
FedAdam55.7672.9162.7861.35
FedAdagrad56.0272.8062.5561.41
FedYogi56.1173.1662.8861.67
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.513, + 0.221, + 0.906, + 0.25 + ], + "angle": 0, + "content": "Table 4. Results of different FL algorithms. All experiments were conducted in the task-related setting with \\( \\beta = 1.0 \\)." + }, + { + "type": "image", + "bbox": [ + 0.576, + 0.273, + 0.862, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.487, + 0.892, + 0.502 + ], + "angle": 0, + "content": "Figure 6. Visualization of activation factors during inference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.515, + 0.906, + 0.545 + ], + "angle": 0, + "content": "FedYogi [42], to extend our framework. As shown in Table 4, FedAvg achieves the best average performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.546, + 0.906, + 0.637 + ], + "angle": 0, + "content": "Visualization of activation factors. In Figure 6, we plot the responses of activations to the test inputs of different tasks. As observed, only the activator corresponding to the current task is responsive, effectively activating the relevant subspace, while the others remain largely inhibited. This confirms the effectiveness of our proposed SSA." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.649, + 0.634, + 0.665 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We have explored for the first time the integration of federated learning and continual learning for the instruction tuning of LMMs, addressing the real-world challenge of dynamically acquiring new knowledge through distributed training resources and data. Our proposed FCIT benchmark encompasses 2 real-world scenarios, 4 distinct settings, and 12 curated datasets, providing a comprehensive evaluation of different methods. Additionally, we introduce the DISCO framework, which leverages dynamic knowledge organization (DKO) to decompose inter-task conflicts and subspace selective activation (SSA) to assign task-relevant outputs while suppressing irrelevant information. Extensive experiments demonstrate that our approach significantly improves the model's ability to learn new knowledge and handle data heterogeneity in real-world scenarios." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1332" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.268 + ], + "angle": 0, + "content": "This work was supported by the National Science and Technology Major Project (2022ZD0116500), National Natural Science Foundation of China (62222609, 62320106010), CAS Project for Young Scientists in Basic Research (YSBR-083), Strategic Priority Research Program of Chinese Academy of Sciences under Grant (XDA0480200), Major Science and Technology Plan Project on the Future Industry Fields of Xiamen City (3502Z20241027), Unveiling and Leading Projects of Xiamen (3502Z20241011) and the InnoHK program." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.278, + 0.188, + 0.294 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.303, + 0.484, + 0.373 + ], + "angle": 0, + "content": "[1] Sara Babakniya, Zalan Fabian, Chaoyang He, Mahdi Soltanolkotabi, and Salman Avestimehr. A data-free approach to mitigate catastrophic forgetting in federated class incremental learning for vision tasks. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.373, + 0.484, + 0.442 + ], + "angle": 0, + "content": "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.443, + 0.484, + 0.497 + ], + "angle": 0, + "content": "[3] Abhijit Bendale and Terrance Boult. Towards open world recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1893-1902, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.498, + 0.484, + 0.568 + ], + "angle": 0, + "content": "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.568, + 0.484, + 0.637 + ], + "angle": 0, + "content": "[5] Meng Cao, Yuyang Liu, Yingfei Liu, Tiancai Wang, Jiahua Dong, Henghui Ding, Xiangyu Zhang, Ian Reid, and Xiaodan Liang. Continual llava: Continual instruction tuning in large vision-language models. arXiv preprint arXiv:2411.02564, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.638, + 0.484, + 0.692 + ], + "angle": 0, + "content": "[6] Cheng Chen, Junchen Zhu, Xu Luo, Hengtao Shen, Lianli Gao, and Jingkuan Song. Coin: A benchmark of continual instruction tuning for multimodel large language model. arXiv preprint arXiv:2403.08350, 2024. 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.693, + 0.484, + 0.762 + ], + "angle": 0, + "content": "[7] Haokun Chen, Yao Zhang, Denis Krompass, Jindong Gu, and Volker Tresp. Feddat: An approach for foundation model finetuning in multi-modal heterogeneous federated learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11285-11293, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.763, + 0.484, + 0.831 + ], + "angle": 0, + "content": "[8] Adam Dahlgren Lindstrom and Savitha Sam Abraham. Clevr-math: A dataset for compositional language, visual and mathematical reasoning. In 16th International Workshop on Neural-Symbolic Learning and Reasoning, NeSy 2022, Windsor, UK, september 28-30, 2022. CEUR-WS, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.832, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[9] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.303, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.148 + ], + "angle": 0, + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.906, + 0.217 + ], + "angle": 0, + "content": "[11] Jiahua Dong, Lixu Wang, Zhen Fang, Gan Sun, Shichao Xu, Xiao Wang, and Qi Zhu. Federated class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10164-10173, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.218, + 0.906, + 0.288 + ], + "angle": 0, + "content": "[12] Haiyang Guo, Fei Zhu, Wenzhuo Liu, Xu-Yao Zhang, and Cheng-Lin Liu. Pilora: Prototype guided incremental lora for federated class-incremental learning. In European Conference on Computer Vision, pages 141-159. Springer, 2024. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.906, + 0.343 + ], + "angle": 0, + "content": "[13] Haiyang Guo, Fei Zhu, Fanhu Zeng, Bing Liu, and XuYao Zhang. Desire: Dynamic knowledge consolidation for rehearsal-free continual learning. arXiv preprint arXiv:2411.19154, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.906, + 0.412 + ], + "angle": 0, + "content": "[14] Haiyang Guo, Fanhu Zeng, Ziwei Xiang, Fei Zhu, Da-Han Wang, Xu-Yao Zhang, and Cheng-Lin Liu. Hidellava: Hierarchical decoupling for continual instruction tuning of multimodal large language model. arXiv preprint arXiv:2503.12941, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.413, + 0.906, + 0.482 + ], + "angle": 0, + "content": "[15] Haiyang Guo, Fanhu Zeng, Fei Zhu, Jiayi Wang, Xukai Wang, Jingang Zhou, Hongbo Zhao, Wenzhuo Liu, Shijie Ma, Xu-Yao Zhang, et al. A comprehensive survey on continual learning in generative models. arXiv preprint arXiv:2506.13045, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.484, + 0.906, + 0.553 + ], + "angle": 0, + "content": "[16] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, Dawn Song, Jacob Steinhardt, and Justin Gilmer. The many faces of robustness: A critical analysis of out-of-distribution generalization. ICCV, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.906, + 0.607 + ], + "angle": 0, + "content": "[17] Tzu-Ming Harry Hsu, Hang Qi, and Matthew Brown. Measuring the effects of non-identical data distribution for federated visual classification. arXiv preprint arXiv:1909.06335, 2019.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.608, + 0.906, + 0.664 + ], + "angle": 0, + "content": "[18] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.906, + 0.733 + ], + "angle": 0, + "content": "[19] Wenke Huang, Mang Ye, Zekun Shi, He Li, and Bo Du. Rethinking federated learning with domain shift: A prototype view. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16312-16322. IEEE, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.734, + 0.906, + 0.803 + ], + "angle": 0, + "content": "[20] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.906, + 0.871 + ], + "angle": 0, + "content": "[21] Kushal Kafle, Brian Price, Scott Cohen, and Christopher Kanan. Dvqa: Understanding data visualizations via question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5648-5656, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.872, + 0.906, + 0.903 + ], + "angle": 0, + "content": "[22] Samira Ebrahimi Kahou, Vincent Michalski, Adam Atkinson, Ákos Kádár, Adam Trischler, and Yoshua Bengio. Fig-" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1333" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "ureqa: An annotated figure dataset for visual reasoning. arXiv preprint arXiv:1710.07300, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.175 + ], + "angle": 0, + "content": "[23] Gyuhak Kim, Changnan Xiao, Tatsuya Konishi, and Bing Liu. Learnability and algorithm for continual learning. In International Conference on Machine Learning, pages 16877-16896. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[24] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[25] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.317, + 0.482, + 0.371 + ], + "angle": 0, + "content": "[26] Lei Li, Yuqi Wang, Runxin Xu, Peiyi Wang, Xiachong Feng, Lingpeng Kong, and Qi Liu. Multimodal arxiv: A dataset for improving scientific comprehension of large vision-language models. arXiv preprint arXiv:2403.00231, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.482, + 0.427 + ], + "angle": 0, + "content": "[27] Qinbin Li, Yiqun Diao, Quan Chen, and Bingsheng He. Federated learning on non-iid data silos: An experimental study. In 2022 IEEE 38th international conference on data engineering (ICDE), pages 965-978. IEEE, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.482, + 0.481 + ], + "angle": 0, + "content": "[28] Tian Li, Anit Kumar Sahu, Ameet Talwalkar, and Virginia Smith. Federated learning: Challenges, methods, and future directions. IEEE signal processing magazine, 37(3):50-60, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.524 + ], + "angle": 0, + "content": "[29] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.526, + 0.483, + 0.607 + ], + "angle": 0, + "content": "[30] Zhuowan Li, Xingrui Wang, Elias Stengel-Eskin, Adam Kortylewski, Wufei Ma, Benjamin Van Durme, and Alan L Yuille. Super-clevr: A virtual benchmark to diagnose domain robustness in visual reasoning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14963-14973, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.609, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[31] Yi-Ming Lin, Yuan Gao, Mao-Guo Gong, Si-Jia Zhang, Yuan-Qiao Zhang, and Zhi-Yuan Li. Federated learning on multimodal data: A comprehensive survey. Machine Intelligence Research, 20(4):539-553, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.482, + 0.705 + ], + "angle": 0, + "content": "[32] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[33] Wenzhuo Liu, Fei Zhu, Haiyang Guo, Longhui Wei, and Cheng-Lin Liu. Llava-c: Continual improved visual instruction tuning. arXiv preprint arXiv:2506.08666, 2025.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.749, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[34] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, et al. Deepseek-vl: towards real-world vision-language understanding. arXiv preprint arXiv:2403.05525, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[35] Pan Lu, Liang Qiu, Jiaqi Chen, Tony Xia, Yizhou Zhao, Wei Zhang, Zhou Yu, Xiaodan Liang, and Song-Chun Zhu. Iconqa: A new benchmark for abstract diagram understanding and visual language reasoning. In The 35th Conference on Neural Information Processing Systems (NeurIPS) Track on Datasets and Benchmarks, 2021. 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.16 + ], + "angle": 0, + "content": "[36] Pan Lu, Liang Qiu, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, Tanmay Rajpurohit, Peter Clark, and Ashwin Kalyan. Dynamic prompt learning via policy gradient for semi-structured mathematical reasoning. arXiv preprint arXiv:2209.14610, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[37] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[38] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 1, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[39] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering by reading text in images. In 2019 international conference on document analysis and recognition (ICDAR), pages 947-952. IEEE, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[40] Bryan A. Plummer, Liwei Wang, Christopher M. Cervantes, Juan C. Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. IJCV, 123 (1):74–93, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.905, + 0.529 + ], + "angle": 0, + "content": "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[42] Sashank Reddi, Zachary Charles, Manzil Zaheer, Zachary Garrett, Keith Rush, Jakub Konečny, Sanjiv Kumar, and H Brendan McMahan. Adaptive federated optimization. arXiv preprint arXiv:2003.00295, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.59, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[43] Lorenzo Sani, Alex Iacob, Zeyu Cao, Bill Marino, Yan Gao, Tomas Paulik, Wanru Zhao, William F Shen, Preslav Aleksandrov, Xinchi Qiu, et al. The future of large language model pre-training is federated. arXiv preprint arXiv:2405.10853, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.729 + ], + "angle": 0, + "content": "[44] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge. In European conference on computer vision, pages 146–162. Springer, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[45] Donald Shenaj, Marco Toldo, Alberto Rigon, and Pietro Zanuttigh. Asynchronous federated continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5055-5063, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[46] Tianxiang Sun, Xiaotian Zhang, Zhengfu He, Peng Li, Qinyuan Cheng, Xiangyang Liu, Hang Yan, Yunfan Shao, Qiong Tang, Shiduo Zhang, et al. Moss: An open conversational large language model. Machine Intelligence Research, 21(5):888-905, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[47] Yue Tan, Guodong Long, Lu Liu, Tianyi Zhou, Qinghua Lu, Jing Jiang, and Chengqi Zhang. Fedproto: Federated prototype learning across heterogeneous clients. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1334" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "the AAAI conference on artificial intelligence, pages 8432-8440, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.178 + ], + "angle": 0, + "content": "[48] Yue Tan, Guodong Long, Jie Ma, Lu Liu, Tianyi Zhou, and Jing Jiang. Federated learning from pre-trained models: A contrastive learning approach. Advances in neural information processing systems, 35:19332-19344, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[49] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.305 + ], + "angle": 0, + "content": "[50] Xiao Wang, Tianze Chen, Qiming Ge, Han Xia, Rong Bao, Rui Zheng, Qi Zhang, Tao Gui, and Xuanjing Huang. Orthogonal subspace learning for language model continual learning. arXiv preprint arXiv:2310.14152, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[51] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 139-149, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.446 + ], + "angle": 0, + "content": "[52] Ziyao Wang, Zheyu Shen, Yexiao He, Guoheng Sun, Hongyi Wang, Lingjuan Lyu, and Ang Li. Flora: Federated fin-tuning large language models with heterogeneous low-rank adaptations. arXiv preprint arXiv:2409.05976, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[53] Taiqiang Wu, Jiahao Wang, Zhe Zhao, and Ngai Wong. Mixture-of-subspaces in low-rank adaptation. arXiv preprint arXiv:2406.11909, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.546 + ], + "angle": 0, + "content": "[54] Xin Yang, Hao Yu, Xin Gao, Hao Wang, Junbo Zhang, and Tianrui Li. Federated continual learning via knowledge fusion: A survey. IEEE Transactions on Knowledge and Data Engineering, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.631 + ], + "angle": 0, + "content": "[55] Rui Ye, Wenhao Wang, Jingyi Chai, Dihan Li, Zexi Li, Yinda Xu, Yaxin Du, Yanfeng Wang, and Siheng Chen. Openfedllm: Training large language models on decentralized private data via federated learning. In Proceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining, pages 6137-6147, 2024. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[56] Fanhu Zeng, Fei Zhu, Haiyang Guo, Xu-Yao Zhang, and Cheng-Lin Liu. Modalprompt: Dual-modality guided prompt for continual learning of large multimodal models. arXiv preprint arXiv:2410.05849, 2024. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[57] Jie Zhang, Chen Chen, Weiming Zhuang, and Lingjuan Lyu. Target: Federated class-continual learning via exemplar-free distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4782–4793, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[58] Jianyi Zhang, Saeed Vahidian, Martin Kuo, Chunyuan Li, Ruiyi Zhang, Tong Yu, Guoyin Wang, and Yiran Chen. Towards building the federatedgpt: Federated instruction tuning. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6915-6919. IEEE, 2024. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[59] Xu-Yao Zhang, Cheng-Lin Liu, and Ching Y Suen. Towards robust pattern recognition: A review. Proceedings of the IEEE, 108(6):894-922, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[60] Hongbo Zhao, Fei Zhu, Rundong Wang, Gaofeng Meng, and Zhaoxiang Zhang. Mllm-cl: Continual learning" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "for multimodal large language models. arXiv preprint arXiv:2506.05453, 2025.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[61] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[62] Fei Zhu, Xu-Yao Zhang, Chuang Wang, Fei Yin, and ChengLin Liu. Prototype augmentation and self-supervision for incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5871-5880, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[63] Fei Zhu, Shijie Ma, Zhen Cheng, Xu-Yao Zhang, Zhaoxiang Zhang, and Cheng-Lin Liu. Open-world machine learning: A review and new outlooks. arXiv preprint arXiv:2403.01759, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1335" + } + ] +] \ No newline at end of file diff --git a/2025/Federated Continual Instruction Tuning/c95b9507-af86-4883-a6e8-057630ebe555_origin.pdf b/2025/Federated Continual Instruction Tuning/c95b9507-af86-4883-a6e8-057630ebe555_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..665495a040aa8abd631ee898235882b1bdaaefa0 --- /dev/null +++ b/2025/Federated Continual Instruction Tuning/c95b9507-af86-4883-a6e8-057630ebe555_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7541ee38e1990a9ec2e9325e6f72da68a11180672fe42d524a159c7ec39f6bb1 +size 997124 diff --git a/2025/Federated Continual Instruction Tuning/full.md b/2025/Federated Continual Instruction Tuning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..951670137fbee82ae0fcf8f828f216cae359bbfa --- /dev/null +++ b/2025/Federated Continual Instruction Tuning/full.md @@ -0,0 +1,342 @@ +# Federated Continual Instruction Tuning + +Haiyang Guo $^{1,2}$ , Fanhu Zeng $^{2,3}$ , Fei Zhu $^{4}$ , Wenzhuo Liu $^{2,3}$ , Da-Han Wang $^{5}$ , Jian Xu $^{2,3}$ , Xu-Yao Zhang $^{1,2,3*}$ , Cheng-Lin Liu $^{1,2,3}$ + +$^{1}$ School of Advanced Interdisciplinary Sciences, UCAS $^{2}$ MAIS, CASIA + +$^{3}$ School of Artificial Intelligence, UCAS $^{4}$ Centre for Artificial Intelligence and Robotics, HKISI-CAS + +$^{5}$ FKLPRIU, School of Computer and Information Engineering, Xiamen University of Technology + +{guohaiyang2023, zengfanhu2022, jian.xu} @ia.ac.cn, zhfei2018@gmail.com, {xyz, liucl} @nlpr.ia.ac.cn + +# Abstract + +A vast amount of instruction tuning data is crucial for the impressive performance of Large Multimodal Models (LMMs), but the associated computational costs and data collection demands during supervised fine-tuning make it impractical for most researchers. Federated learning (FL) has the potential to leverage all distributed data and training resources to reduce the overhead of joint training. However, most existing methods assume a fixed number of tasks, while in real-world scenarios, clients continuously encounter new knowledge and often struggle to retain old tasks due to memory constraints. In this work, we introduce the Federated Continual Instruction Tuning (FCIT) benchmark to model this real-world challenge. Our benchmark includes two realistic scenarios, encompassing four different settings and twelve carefully curated instruction tuning datasets. To address the challenges posed by FCIT, we propose a dynamic knowledge organization to effectively integrate updates from different tasks during training and subspace selective activation to allocate task-specific output during inference. Extensive experimental results demonstrate that our proposed method significantly enhances model performance across varying levels of data heterogeneity and catastrophic forgetting. Code and dataset are released at https://github.com/Ghy0501/FCIT. + +# 1. Introduction + +Large Multimodal Models (LMMs) [2, 32, 34], which integrate Large Language Model [4, 20, 46, 49] with a visual encoder and multimodal projector to bridge visual and textual modalities, have exhibited impressive visual understanding and complex reasoning abilities. A crucial factor in this success is the supervised fine-tuning of LMMs using huge and diverse visual instruction-following data [32] + +to align with human preferences. However, collecting such vast amounts of training data and computational resources for joint fine-tuning is impractical for most researchers. Federated Learning (FL) [28, 31, 38, 43], as a decentralized paradigm, offers a viable alternative by leveraging distributed data and computational resources for local training while integrating local weights to produce a unified model. This paradigm accommodates constraints on storage and computation while ensuring privacy protection. + +Generally, most existing FL frameworks [7, 55, 58] are modeled in static, closed-world scenarios [3, 59], where a fixed set of tasks is predefined and unchanged. However, real-world applications are dynamic [59, 63], requiring models to continuously acquire new knowledge while retaining previously learned tasks. Taking a realistic health emergency as an example, different hospitals act as local clients when a major disease outbreak occurs. Large hospitals can leverage FL to collaboratively train on their case data, building a comprehensive virus knowledge base, while smaller clinics update their own cases using this global knowledge. Over time, both large and small hospitals need to keep updating the global knowledge base through continual learning methods to cope with the event. Additionally, the emergence of new strains necessitates the simultaneous integration of knowledge and response strategies to minimize potential losses. In this scenario, traditional FL methods struggle with newly arrived knowledge, while Continual Learning (CL) [14, 15, 33, 56, 60] methods alone do not facilitate knowledge sharing between clients. Only an organic integration of both can address this real-world problem. Recently, numerous Federated Continual Learning methods have emerged to address this challenge in traditional image classification tasks [11, 12, 57]. However, their methodological and task-setting limitations make them insufficient for current LMM applications. Therefore, a more comprehensive benchmark is needed to better simulate the practical application of LMMs in real-world scenarios. + +In this work, we first establish a Federated Continual In + +![](images/1172980ffe523979c81b514f577a783b638694ff4f32a162ac2b3c43a6229dc7.jpg) +Figure 1. Overview of FCIT benchmark. FCIT encompasses 2 real-world scenarios, 4 FCIT settings, and 12 curated datasets, providing a comprehensive simulation of LMMs instruction-following training in real-world applications. + +struction Tuning (FCIT) benchmark to fill this gap. Specifically, we simulate two client-level realistic scenarios: (i) Homogeneous FCIT refers to an FL system learning a series of instruction tuning data sequentially, with different clients learning the same task at each stage. (ii) Heterogeneous FCIT, on the other hand, requires the FL system to collaborate on different tasks simultaneously, with different clients potentially learning different tasks in the same stage. Building on this, we define two settings of datasets for each scenario: Capability-related and Task-related. The former evaluates the model's ability to integrate multiple dimensions of instruction following in a short period, while the latter assesses the model's performance during long phases of continual learning. For dataset selection, we curate twelve instruction tuning datasets unseen during LMM fine-tuning or with low zero-shot performance, preventing information leakage [13, 23]. In addition, we introduce varying degrees of data heterogeneity for each setting to challenge the model's performance in the non-IID situation [27]. An illustration of our FCIT benchmark is provided in Figure 1. To the best of our knowledge, this is the first work to introduce a comprehensive benchmark for federated learning of LMM in a continual learning setting. + +To effectively address the challenges posed by FCIT, we propose a novel Dynamic knowledge organization and Subspace seleCtive activatiOn (DISCO) framework. Specifically, we identify FCIT challenges into two types: conflicts between different tasks within the same stage and conflicts between old and new tasks across different stages. + +The former can cause catastrophic forgetting by altering the parameter space of previous tasks when learning new ones without access to past data, while the latter requires the model to integrate and organize knowledge from different tasks to derive unified representations. Therefore, we first propose Dynamic Knowledge Organization (DKO), which leverages a dynamic cache at the global server to store task-specific parameters. Using a unique identity token matching mechanism, it systematically organizes knowledge for different tasks into corresponding subspaces within the cache, effectively mitigating two types of conflicts. To better utilize the organized task subspaces in the dynamic cache, we introduce Subspace Selective Activation (SSA), which selectively activates subspaces relevant to the test input while filtering out irrelevant outputs, leading to significant performance improvements. Consequently, these designs enable our framework to efficiently tackle the data heterogeneity and catastrophic forgetting in FCIT. In summary, our major contributions are: + +- We present the first Federated Continual Instruction Tuning (FCIT) benchmark designed for real-world scenarios, providing a comprehensive evaluation of LMMs to continuously learn new knowledge using distributed data and training resources in real applications. +- We propose a novel DISCO framework that integrates dynamic knowledge organization and subspace selective activation to efficiently address data heterogeneity and catastrophic forgetting in FCIT settings. +- Extensive experiments demonstrate that our method sig + +nificantly enhances model performance under data heterogeneity while minimizing the catastrophic forgetting, and achieves state-of-the-art performance. + +# 2. Related Work + +Large Multimodal Models. With the grand unification of Large Language Models (LLMs) [4, 20, 49] for various NLP tasks, Large Multimodal Models (LMMs) [2, 9, 32, 34, 61] emerge by extending LLMs to combine with visual encoder and multimodal projectors, demonstrating exceptional visual understanding and complex reasoning abilities. To better align with human preferences, these LMMs typically undergo further fine-tuning on extensive instruction-following data, ensuring they meet the demands of real-world applications [32]. However, in real-world scenarios, improving the performance of LMMs on new downstream tasks becomes a significant challenge without access to sufficient training data and computational resources. In this paper, we introduce the first Federated Continual Instruction Tuning (FCIT) benchmark to bridge this gap from the perspective of distributed training and continual learning. + +Federated Continual Learning. In classical vision tasks, Federated Continual Learning (FCL) [54] aims to adapt the global model to new data while maintaining the knowledge of the old task. From the perspective of LMMs, research in this setting faces two main challenges: (1) Most of the methods are designed for traditional vision tasks (i.e., image classification). For instance, MFCL [1] employs a generative model to synthesize images of previously learned classes, thereby mitigating forgetting. PILoRA [12] introduces a prototype re-weight module to address the classifier bias caused by data heterogeneity and obtain unified knowledge through LoRA [18] fusion. Despite progress in classification tasks, complex designs remain challenging to adapt for LMMs research. (2) FCL follows a single data composition by splitting a dataset (e.g. ImageNet [10]) into different tasks based on classes, whereas LMMs face greater challenges in continual learning due to their diverse tasks with varying styles [5, 6, 56]. + +This work pioneers a Federated Continual Instruction Tuning setup for LMMs and establishes diverse scenarios to simulate real-world applications comprehensively. Notably, AFCL [45] is most relevant to our work, as it enables clients to continuously learn multiple tasks in different orders and asynchronous time slots. However, it is specifically designed for image classification tasks, whereas our study focuses on the more widely used LMM. + +# 3. Problem Formulation + +# 3.1. Preliminaries + +Instruction tuning enhances LMM's ability to understand and execute human instructions by performing supervised + +fine-tuning of pre-trained models on extensive datasets comprising instructions and responses. Formally, the instruction data $\mathcal{D} = \{(\mathbf{x}_v^j,\mathbf{x}_{ins}^j,\mathbf{x}_{res}^j)^N\}$ consists of image input $\mathbf{x}_v$ , instruction $\mathbf{x}_{ins}$ and response $\mathbf{x}_{res}$ , where $N$ represents the total number of samples. For clarity, given a simple image-instruction pair with a response of length $L$ , the objective of an LMM is to predict the next token autoregressively based on all preceding tokens: + +$$ +p \left(\mathbf {x} _ {r e s} \mid \mathbf {x} _ {v}, \mathbf {x} _ {i n s}\right) = \prod_ {i = 1} ^ {L} p _ {\theta} \left(x _ {i} \mid \mathbf {x} _ {v}, \mathbf {x} _ {i n s}, \mathbf {x} _ {r e s, < i}\right), \tag {1} +$$ + +where $\theta$ denotes the trainable parameters during fine-tuning, $\mathbf{x}_{\mathrm{res}, < i}$ denotes all response tokens preceding the current prediction token $x_{i}$ . Then, the loss function of fine-tuning LMMs can be expressed as: + +$$ +\mathcal {L} _ {\theta} = - \frac {1}{N} \sum_ {j = 1} ^ {N} \sum_ {i = 1} ^ {L _ {j}} \log p _ {\theta} \left(x _ {i} ^ {j} \mid \mathbf {x} _ {v} ^ {j}, \mathbf {x} _ {i n s} ^ {j}, \mathbf {x} _ {r e s, < i} ^ {j}\right). \tag {2} +$$ + +Federated learning framework typically comprises a global server and several local clients, all employing the same LMM with a shared homogeneous model architecture. In each communication round1, local clients train their models on own data and upload the updated weights to the global server for aggregation, enabling collaborative optimization of the global model while preserving data privacy. + +Considering the excessive communication overhead of transferring the entire LMM between clients and the global server, we adopt LoRA [18] for efficient fine-tuning, balancing training overhead and implementation cost [55, 58]. Specifically, for a weight matrix $\mathbf{W}_0\in \mathbb{R}^{d\times k}$ , LoRA decomposes parameter updates $\Delta \mathbf{W}$ during fine-tuning into two low-rank subspaces: + +$$ +\mathbf {W} = \mathbf {W} _ {0} + \Delta \mathbf {W} = \mathbf {W} _ {0} + \mathbf {B A}, \tag {3} +$$ + +where $\mathbf{B} \in \mathbb{R}^{d \times r}$ , $\mathbf{A} \in \mathbb{R}^{r \times k}$ and $r \ll \min\{d, k\}$ . + +Continual learning aims to minimize the loss on the current task while retaining knowledge from previous tasks. Formally, given a sequence of datasets $\mathcal{D}_1,\mathcal{D}_2,\dots ,\mathcal{D}_T$ , the optimization objective at task $t$ is: + +$$ +\min _ {\theta} \mathcal {L} (\theta) = \mathbb {E} _ {(x, y) \sim \mathcal {D} _ {t}} [ \mathcal {L} (f _ {\theta_ {t}} (x), y) ] + \sum_ {i = 1} ^ {t - 1} \epsilon_ {i}, \tag {4} +$$ + +$$ +\begin{array}{l} \operatorname {s. t.} \mathbb {E} _ {(x, y) \sim \mathcal {D} _ {i}} \left[ \mathcal {L} \left(f _ {\theta_ {t}} (x), y\right) - \mathcal {L} \left(f _ {\theta_ {t - 1}} (x), y\right) \right] \leq \epsilon_ {i}, \\ \epsilon_ {i} \geq 0; \forall i \in [ 1, \dots , t - 1 ], \\ \end{array} +$$ + +where $\epsilon_{i}$ is a slack variable that allows a small increase in the loss from the old datasets, providing tolerance for minor forgetting while focusing on learning the current task. In Eq. 4, $x$ and $y$ can be viewed as the multimodal inputs $(\mathbf{x}_v,\mathbf{x}_{ins})$ and the response $\mathbf{x}_{res}$ , respectively. + +# 3.2. Federated Continual Instruction Tuning + +As shown in Figure 1, we integrate federated and continual learning for LMMs within a unified framework and propose two client-level realistic scenarios: (1) Homogeneous FCIT (Hom-FCIT). In this scenario, clients sequentially learn a series of tasks, allowing the global server to continuously update its knowledge. Each client learns the same task at a given stage and can only access data from that stage. (2) Heterogeneous FCIT (Het-FCIT). In real-world applications, different clients may learn different tasks within the same stage, enabling the global server to respond more rapidly to diverse instructions. This requires the model not only to coordinate the knowledge of different tasks learned by clients in the current stage but also to mitigate forgetting during the learning process. For each scenario, we define two dataset-level settings: capability-related and task-related, detailed as follows. + +- Capability-related. Following the dataset construction and division in LLaVA-OneVision [25], we classify the 12 datasets into 4 capabilities: General, Math, Chart, and Other. The General capability includes A-OKVQA [44], ImageNet-R [16], Grounding [37], and IconQA [35]; Math comprises CLEVR-Math [8], super-CLEVR [30], and TabMWP [36]; Chart involves ArxivQA [26], FigureQA [22], and DVQA [21]; and Other encompasses OCR-VQA [39] and Flickr30k [40]. We treat these four capabilities as different stages of continual learning, with each capability comprising a mixture of datasets. +- Task-related. To evaluate the performance of different methods in a long-phase continual learning situation, we selected 8 datasets: ImageNet-R, ArxivQA, IconQA, CLEVR-Math, OCR-VQA, Flickr30k, FigureQA, and super-CLEVR as distinct stages of continual learning. + +In total, we propose 4 FCIT settings to evaluate different methods. Compared to existing works [5, 6, 55, 56, 58], we are the first to explore the organic integration of FL and CL in the context of LMM. We provide further illustrations on the settings and datasets in Appendix A. + +# 4. The Proposed Framework: DISCO + +Overview of the Method. As shown in Figure 2 and 3, our method primarily consists of: (1) Dynamic Knowledge Organization (DKO), which dynamically integrates knowledge learned by different clients across stages during training, significantly reducing inter-task conflicts (Section 4.1); and (2) Subspace Selective Activation (SSA), which selectively activates subspace outputs based on input features during inference, effectively filtering out irrelevant information (Section 4.2). We name this framework DISCO. + +![](images/2069e575ef5e18f160271e66931eeb8285660d720ea8a7abe549298323d281bc.jpg) +Figure 2. Illustration of the proposed DKO. Dynamic caches store the knowledge of each subspace while matching and updating through identity tokens. + +# 4.1. Dynamic Knowledge Organization + +The core challenge of FCIT is enabling the global server to effectively harmonize knowledge learned by clients, addressing both the forgetting of previous knowledge when acquiring new tasks and the conflicts that arise from integrating knowledge from different tasks within the same stage. To this end, we propose maintaining a dynamic cache at the global server to organize task-specific knowledge uploaded by clients at different stages, thereby preventing both forgetting and conflicts between knowledge. Specifically, each task is assigned a dedicated parameter space to store and update its corresponding knowledge: + +$$ +\Delta \mathbf {W} = \mathbf {B} \mathbf {A} \Leftrightarrow \underbrace {\left\{\mathbf {B} _ {1} \mathbf {A} _ {1} , \cdots , \mathbf {B} _ {T} \mathbf {A} _ {T} \right\}} _ {\text {t a s k - s p e c i f i c s u b s p a c e}}, \tag {5} +$$ + +where $T$ is the number of tasks learned. At this point, it is crucial to aggregate the parameters uploaded by clients into their respective subspaces without privacy leakage. + +Inspired by the widespread use of prototypes in FL fields [19, 47, 48], we propose to distinguish knowledge across tasks by using the feature mean of each client's training data as an Identity Token. Specifically, considering the uniqueness of textual inputs in the visual instruction tuning task, we introduce a text encoder $f_{ins}$ for each client to extract the feature of the input instruction $\mathbf{x}_{ins,j}^{t}$ during training, and take the mean value $\mu_k^t$ as the local identity token of the $k$ -th client on the task $t$ at the end of training: + +$$ +\mu_ {k} ^ {t} = \frac {1}{n _ {k} ^ {t}} \sum_ {j = 1} ^ {n _ {k} ^ {t}} f _ {i n s} \left(\mathbf {x} _ {i n s, j} ^ {t}\right), \tag {6} +$$ + +where $n_k^t$ is the number of training samples of client $k$ at + +![](images/321710d579552cd17ea1e7d5f77fe3666442909e13c48150ba9e12d9c387c1ca.jpg) +Figure 3. Illustration of the proposed SSA. Each subspace organized by DKO dynamically adjusts its output via the intrinsic activation matrix during inference, effectively filtering irrelevant information and enhancing model performance. + +task $t$ , and we use CLIP's text encoder [41] in our experiments. Notably, the text encoder $f_{ins}$ remains frozen during training, which prevents additional training overhead. + +Using the identity tokens, we first apply cosine similarity along with a threshold to match the uploaded local identity token $\mu_j^t$ with the global identity token $\tilde{\mathbf{z}}_i$ : + +$$ +\frac {\mu_ {j} ^ {t} \cdot \tilde {\mathbf {z}} _ {i}}{\| \mu_ {j} ^ {t} \| \| \tilde {\mathbf {z}} _ {i} \|} \geq \tau , \tag {7} +$$ + +where $\tau$ denotes a pre-defined threshold. For paired local identity tokens, we update the corresponding global identity token. For mismatched local identity tokens, we pair them two by two using Eq. (7) and initialize new global identity tokens. This process can be formalized as: + +$$ +\tilde {\mathbf {z}} _ {i} ^ {*} = \left\{ \begin{array}{l l} \frac {n _ {\bar {\mathbf {z}} _ {i}} \cdot \tilde {\mathbf {z}} _ {i} + \sum_ {j = 1} ^ {m} n _ {j} ^ {t} \cdot \mu_ {j} ^ {t}}{n _ {\bar {\mathbf {z}} _ {i}} + \sum_ {j = 1} ^ {m} n _ {j} ^ {t}}, & \text {i f} \tilde {\mathbf {z}} _ {i} \text {e x i s t s}, \\ \frac {\sum_ {j = 1} ^ {m} n _ {j} ^ {t} \cdot \mu_ {j} ^ {t}}{\sum_ {j = 1} ^ {m} n _ {j} ^ {t}}, & \text {i f} \tilde {\mathbf {z}} _ {i} \text {d o e s n ’ t e x i s t}, \end{array} \right. \tag {8} +$$ + +where $m$ , $n_{\overline{\mathbf{z}}_i}$ and $n_j^t$ denote the number of matched local identity tokens, the number of samples used to form the previous $i$ -th global identity token, and the number of samples from the local client $j$ at task $t$ , respectively. Then, we leverage this matching process to guide the update or initialization of task-specific subspace $\{\theta_i = \mathbf{B}_i\mathbf{A}_i\}$ in Eq. (5): + +$$ +\theta_ {i} ^ {*} = \mathcal {F} \left(\theta_ {1} ^ {t}, \dots , \theta_ {m} ^ {t}\right), \tag {9} +$$ + +where $\mathcal{F}$ denotes the FL algorithm (e.g. FedAvg [38]) used to aggregate local weights $^2$ . As a result, we effectively prevent inter-task conflicts and integrate knowledge from different clients using the identity token matching mechanism. + +After the global server completes the aggregation, it distributes the dynamic cache to each selected client. The client then matches the identity token (Using Eq. (7)) of each subspace with its own training data, deciding whether to update the corresponding subspace or reinitialize a new one. The entire DKO process is illustrated in Figure 2. + +# 4.2. Subspace Selective Activation + +Section 4.1 effectively mitigates inter-task conflicts by disentangling and organizing complex knowledge into distinct subspaces. The key challenge then lies in how to leverage these task-specific subspaces during inference. + +To address this, an intuitive method is to concatenate the subspaces of the dynamic cache in low-rank dimensions [50, 52], enabling the integration of knowledge across all task spaces. However, this may introduce redundant information unrelated to the current task during inference, potentially compromising the model's output. For instance, when the desired answer is a simple word (e.g., what is the object in the picture), knowledge from other subspaces, such as those used for generating long-form descriptions, can introduce unnecessary information, leading to responses that do not align with the given instruction. + +Drawing inspiration from LoRA's intrinsic space [53], we propose subspace selective activation (SSA) without additional training to filter out irrelevant subspace outputs, ensuring alignment between responses and instructions. In particular, A vanilla LoRA can be decomposed as the product of two low-rank subspaces (i.e., $\mathbf{A}$ , $\mathbf{B}$ ) and an intrinsic mixing matrix $\mathcal{W}$ : + +$$ +\Delta \mathbf {W} = \mathbf {B} \mathcal {W} \mathbf {A}, \tag {10} +$$ + +where $\mathcal{W} \in \mathbb{R}^{r \times r}$ is typically an identity matrix and can thus be omitted. In this paper, we redefine $\mathcal{W}$ as the activation matrix that dynamically responds to the test input. Specifically, we treat $\mathcal{W}$ as the product of an identity matrix and an activation factor (i.e., $\mathcal{W} = \alpha \cdot \mathbf{I}_{r \times r}$ ), where $\alpha = 1$ denotes full activation, and $\alpha = 0$ means that the output is fully masked. Therefore, for each subspace $\{\mathbf{B}_1\mathbf{A}_1, \dots, \mathbf{B}_T\mathbf{A}_T\}$ in dynamic cache, we can control the activation or inhibition of the corresponding output by adjusting its activation factor $\{\alpha_1, \dots, \alpha_T\}$ . + +To provide better flexibility in assigning activation factors, we use global identity tokens and test input features extracted by the text encoder for similarity matching: + +$$ +s _ {i} = \frac {\tilde {\mathbf {z}} _ {\mathbf {i}} \cdot f _ {i n s} \left(\mathbf {x} _ {i n s} ^ {t e s t}\right)}{\| \tilde {\mathbf {z}} _ {\mathbf {i}} \| \cdot \| f _ {i n s} \left(\mathbf {x} _ {i n s} ^ {t e s t}\right) \|}, \tag {11} +$$ + +
Dataset settingCapability-related (4 task)Task-related (8 task)
Partitionβ = 0.5β = 1.0β = 5.0β = 0.5β = 1.0β = 5.0
MethodsLastAvgLastAvgLastAvgLastAvgLastAvgLastAvg
Zero-shot30.57-30.57-30.57-29.08-29.08-29.08-
Individual61.75-59.62-60.27-64.21-63.85-64.07-
Centralized MTL63.83-63.83-63.83-66.60-66.60-66.60-
Finetune48.6860.0549.6759.0150.4058.0846.2468.9347.2068.7948.0069.97
EWC49.1260.1349.4659.2849.8958.7647.5169.0447.9269.2248.1570.27
LwF48.8760.3250.0259.8850.1559.0647.8969.5747.6269.1448.2670.30
L2P48.2259.7949.5659.3449.6259.2547.3069.3148.0869.6548.4270.16
O-LoRA51.6560.1949.1357.9350.2958.1852.8771.5449.8770.2647.7670.84
M-LoRA49.0460.0850.3960.7650.5658.0650.6871.9448.5371.5848.3871.21
MoELoRA49.6961.0050.9060.1850.4359.1149.2370.9649.0270.6548.8271.08
DISCO53.7362.0055.4762.0755.0660.5357.6974.0356.2273.0355.5872.64
+ +Table 1. Last and Avg performance of different methods on Hom-FCIT setting. The best performance is shown in bold. + +![](images/b2202c5f0b067b022c91b6c9c651798aee492216ef68e6307e0554fb32338fa2.jpg) +Figure 4. Performance curves of different methods in Hom-FCIT across seen tasks under varying data heterogeneity. We plot the average performance across all seen tasks at each stage. + +![](images/a78c15ea7723b5e842794c4d1f000cf743c20b5df6ebd16a2d30b2ea1abd4895.jpg) + +![](images/549f69b58d50b25e8cd8ea9bede745cac61ec68a5e506a528ae30003b5478c11.jpg) + +where $s_i$ denotes the similarity between the $i$ -th global identity token and the feature of the test instruction $\mathbf{x}_{ins}^{test}$ . The activation factor $\alpha_i$ is then computed by applying softmax normalization over all similarities: + +$$ +\alpha_ {i} = \frac {\exp \left(s _ {i} / \varepsilon\right)}{\sum_ {j = 1} ^ {T} \exp \left(s _ {j} / \varepsilon\right)}, \tag {12} +$$ + +where $\varepsilon$ is the temperature coefficient. + +Overall, our proposed SSA can be formulated as follows: + +$$ +\begin{array}{l} \Delta \mathbf {W} = \bar {\mathbf {B}} \bar {\mathcal {W}} \bar {\mathbf {A}} \\ = \bar {\mathbf {B}} \left[ \begin{array}{c c c c} \alpha_ {1} \cdot \mathbf {I} _ {r \times r} & \mathbf {0} _ {r \times r} & \dots & \mathbf {0} _ {r \times r} \\ \mathbf {0} _ {r \times r} & \alpha_ {2} \cdot \mathbf {I} _ {r \times r} & \dots & \mathbf {0} _ {r \times r} \\ \vdots & \vdots & \ddots & \vdots \\ \mathbf {0} _ {r \times r} & \mathbf {0} _ {r \times r} & \dots & \alpha_ {T} \cdot \mathbf {I} _ {r \times r} \end{array} \right] \bar {\mathbf {A}}, \tag {13} \\ \end{array} +$$ + +where $\bar{\mathbf{B}}\in \mathbb{R}^{d\times (r\cdot T)}$ represents the concatenation of all $\{\mathbf{B}_1,\dots ,\mathbf{B}_T\}$ in the dynamic cache along the low-rank dimension $r$ , and $\bar{\mathbf{A}}\in \mathbb{R}^{(r\cdot T)\times k}$ follows the same structure. The activation factors $\{\alpha_{1},\dots ,\alpha_{T}\}$ , computed via Eq. (11) and Eq. (12), control each subspace's output by amplifying matched subspaces while suppressing irrelevant ones. The schematic of SSA is provided in Figure 3. + +# 5. Experiments + +# 5.1. Experimental Setup + +Datasets. The dataset composition is detailed in Section 3.2. For the experimental setup, we implement two types of dataset-level settings: Capability-related and Task-related, based on the two client-level realistic scenarios, Hom-FCIT and Het-FCIT, respectively. At each stage, the own data of local clients is partitioned according to the Dirichlet distribution [27], with three partitions $\beta$ for each setting to model different levels of data heterogeneity. More details are provided in Appendix A. + +Baselines & Evaluation Metrics. We compare our method with continual learning methods such as LwF [29], EWC [24], L2P [51], O-LoRA [50], MoELoRA [6], and also with the LoRA merging method from the federated continual learning approach PILoRA [12], referred to as M-LoRA. We follow a rehearsal-free continual learning setting [62], where only the data for the current task is available. All comparison methods are carefully calibrated to ensure the fairness of evaluations. + +For the evaluation metrics, we report the standard metrics to measure the model performance: Last refers to the + +
Dataset settingCapability-related (4 task)Task-related (8 task)
Partitionβ = 0.5β = 1.0β = 5.0β = 0.5β = 1.0β = 5.0
MethodsLastAvgLastAvgLastAvgLastAvgLastAvgLastAvg
Zero-shot30.57-30.57-30.57-29.08-29.08-29.08-
Individual61.75-59.62-60.27-64.21-63.85-64.07-
Centralized MTL63.83-63.83-63.83-66.60-66.60-66.60-
Finetune55.6555.8256.3456.8556.7457.1558.0453.0557.9654.2258.1754.87
EWC55.2154.5855.1355.7655.6956.0258.7453.7657.9654.1857.4454.30
LwF55.9256.1856.4256.8056.8157.2658.8253.7758.0154.3458.2254.78
L2P56.1056.6356.7657.0256.9557.3658.8453.8058.7354.6758.3954.66
O-LoRA58.2458.5858.3258.6558.6058.9459.6154.5559.7455.2059.5154.97
M-LoRA57.7658.0257.6557.8957.8058.1759.7654.1158.8254.0359.3554.89
MoELoRA57.6857.9557.7758.0058.1558.4459.0254.2559.1454.6958.8654.50
DISCO59.4060.0159.9459.9159.7160.1662.0859.6463.2561.9962.7860.87
+ +Table 2. Last and Avg performance of different methods on Het-FCIT setting. The best performance is shown in bold. + +![](images/04b024c7588837943e90cbecd61b2dbb677a73bdbb08184309f889f7404c0219.jpg) +Figure 5. Performance curve of different methods in the Het-FCIT under varying degrees of data heterogeneity. We plot the average performance across all tasks at each stage. The black dashed line marks the stage where the model has encountered all tasks. + +![](images/2c2bb57e7731e238b1bf640d4bae8e55cb25d5709ceb01b1ee351eea830f7a09.jpg) + +![](images/1772929cc40a3675df940e8e9ae9519aa576e5169fb6d15fcd7dee50dd9816c1.jpg) + +average result across all learned tasks after the completion of learning the final task. $Avg$ is based on Last, which tracks the performance of the learned tasks at each stage of the continual learning process and reports the average results. + +Implementation Details. We choose LLaVA-1.5-7b [32] as the base LMM due to its structural simplicity and adopt its LoRA fine-tuning strategy, training only the LoRAs and multimodal projector. The LoRA rank $r$ at each stage is set to 8, with a learning rate of 2e-4, and it is embedded solely in the FFN layers of each block. The learning rate for the projector is set to 2e-5. For identity token extraction, we use the frozen CLIP text encoder [41]. The threshold $\tau$ and temperature coefficient $\varepsilon$ are set to 0.9 and 0.05, respectively. + +For the setting of FCIT, we set the epoch to 1 and communication rounds to 10 in each stage, and the Dirichlet distribution coefficients $\beta$ are set to $\{0.5, 1.0, 5.0\}$ . In each round, the global server randomly selects 5 clients from a pool of 50 to participate in the training, and we use FedAvg [38] as the base FL aggregation algorithm. + +# 5.2. Main Results + +Results are shown in Table 1, Table 2, Figure 4, and Figure 5. Under Hom-FCIT setting, our proposed DISCO + +achieves the best performance in both capability-related and task-related dataset settings, surpassing the best baseline by an average of $4.84\%$ and $1.43\%$ in the Last and Avg metrics, respectively. As illustrated in Figure 4, DISCO effectively mitigates the forgetting of previous tasks while learning new ones, outperforming other methods in the challenging long-stage continual learning setting. + +In Het-FCIT setting, we plot the upward curve of the average performance across all tasks in Figure 5. As shown in the figure. Before completing all tasks (Left of the black dotted line), our method maintains the fastest rate of improvement, significantly outperforming other methods. In the subsequent phases (Right of the black dashed line), it continues to consolidate previously learned knowledge and further enhance performance, demonstrating strong adaptability in dynamic real-world scenarios. On the Last and Avg metrics, our method outperforms the best comparison method by an average of $2.17\%$ and $3.62\%$ , respectively. + +Additionally, our method demonstrates superior robustness to varying degrees of data heterogeneity, adapting well to diverse distributions and maintaining strong performance even as heterogeneity increases, highlighting its reliability in dynamic real-world scenarios. + +
MethodHom-FCITHet-FCIT
LastΔAvgΔLastΔAvgΔ
(a)Text56.220.073.030.063.250.061.990.0
Image55.63-0.5972.13-0.9063.00-0.2561.80-0.19
Text & Image55.96-0.2672.68-0.3563.02-0.2361.78-0.21
(b)Softmax56.220.073.030.063.250.061.990.0
Concatenate51.74-4.4869.20-3.8360.36-2.8959.88-2.11
Cosine sim52.83-3.3970.07-2.9660.92-2.3360.13-1.86
Argmax55.74-0.4872.07-0.9662.88-0.3761.42-0.57
(c)FFN56.220.073.030.063.250.061.990.0
Attn56.03-0.1972.88-0.1563.01-0.2461.83-0.16
FFN & Attn56.30+0.0872.96-0.0763.18-0.0762.04+0.05
+ +Table 3. Ablation studies on (a) identity token extraction methods; (b) calculation of activation factors in SSA; (c) location of LoRA embedding. All experiments were conducted in the task-related setting with $\beta = 1.0$ . + +# 5.3. Ablation Study + +The detailed ablation studies are provided in Table 3. In this paper, we use CLIP's text encoder to extract textual features as identity tokens. This section tests two alternatives: using only visual features from CLIP's visual encoder, and combining visual and textual features. From Table 3a, we observe that both alternatives degrade performance. This is potential because, in visual instruction tuning datasets, task similarities at the image level are higher than at the textual level (e.g. CLEVR-Math and super-CLEVR), making visual features less effective than textual ones. + +Table 3b shows the ablation studies on how to calculate the activation factors in SSA. Direct concatenation, where the activation factors of all subspaces are set to 1.0, leads to significant performance degradation due to unfiltered outputs. Using cosine similarity (Eq.(11)) to filter the outputs also fails to solve the issue. While selecting the largest similarity (Argmax) can filter out irrelevant information, it risks fully activating a subspace that, though similar, is not directly related to the current task. This can negatively impact the output, as there are inherent similarities between the textual information of different tasks. In contrast, our method normalizes the cosine similarity to better focus on the relevant task output, achieving optimal performance. + +We also test the location of LoRA embedding. As can be seen in Table 3c, embedding in the attention layer alone proved less effective than in the FFN layer. Moreover, embedding LoRA in every linear layer did not offer significant improvement and resulted in higher parameter transmission. Therefore, we chose to embed LoRA only in the FFN layer as a balanced solution. We provide more ablation and visualization results in Appendix B. + +# 5.4. Further Analysis + +Compatible with other FL algorithms. In this paper, we use the classical FedAvg as the FL algorithm for global server aggregation of local weights. Additionally, we implement other federated learning algorithms, including FedAvgM [17], FedAdam [42], FedAdagrad [42], and + +
MethodHom-FCITHet-FCIT
LastAvgLastAvg
FedAvg56.2273.0363.2561.99
FedAvgM56.0772.5962.4761.10
FedAdam55.7672.9162.7861.35
FedAdagrad56.0272.8062.5561.41
FedYogi56.1173.1662.8861.67
+ +Table 4. Results of different FL algorithms. All experiments were conducted in the task-related setting with $\beta = 1.0$ . + +![](images/fa06c48127ce5152049f999ae4e2d0149a8d8058956cfc623619cb4ac5987f1f.jpg) +Figure 6. Visualization of activation factors during inference. + +FedYogi [42], to extend our framework. As shown in Table 4, FedAvg achieves the best average performance. + +Visualization of activation factors. In Figure 6, we plot the responses of activations to the test inputs of different tasks. As observed, only the activator corresponding to the current task is responsive, effectively activating the relevant subspace, while the others remain largely inhibited. This confirms the effectiveness of our proposed SSA. + +# 6. Conclusion + +We have explored for the first time the integration of federated learning and continual learning for the instruction tuning of LMMs, addressing the real-world challenge of dynamically acquiring new knowledge through distributed training resources and data. Our proposed FCIT benchmark encompasses 2 real-world scenarios, 4 distinct settings, and 12 curated datasets, providing a comprehensive evaluation of different methods. Additionally, we introduce the DISCO framework, which leverages dynamic knowledge organization (DKO) to decompose inter-task conflicts and subspace selective activation (SSA) to assign task-relevant outputs while suppressing irrelevant information. Extensive experiments demonstrate that our approach significantly improves the model's ability to learn new knowledge and handle data heterogeneity in real-world scenarios. + +# Acknowledgments + +This work was supported by the National Science and Technology Major Project (2022ZD0116500), National Natural Science Foundation of China (62222609, 62320106010), CAS Project for Young Scientists in Basic Research (YSBR-083), Strategic Priority Research Program of Chinese Academy of Sciences under Grant (XDA0480200), Major Science and Technology Plan Project on the Future Industry Fields of Xiamen City (3502Z20241027), Unveiling and Leading Projects of Xiamen (3502Z20241011) and the InnoHK program. + +# References + +[1] Sara Babakniya, Zalan Fabian, Chaoyang He, Mahdi Soltanolkotabi, and Salman Avestimehr. A data-free approach to mitigate catastrophic forgetting in federated class incremental learning for vision tasks. Advances in Neural Information Processing Systems, 36, 2024. 3 +[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.1, 3 +[3] Abhijit Bendale and Terrance Boult. Towards open world recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1893-1902, 2015. 1 +[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1, 3 +[5] Meng Cao, Yuyang Liu, Yingfei Liu, Tiancai Wang, Jiahua Dong, Henghui Ding, Xiangyu Zhang, Ian Reid, and Xiaodan Liang. Continual llava: Continual instruction tuning in large vision-language models. arXiv preprint arXiv:2411.02564, 2024. 3, 4 +[6] Cheng Chen, Junchen Zhu, Xu Luo, Hengtao Shen, Lianli Gao, and Jingkuan Song. Coin: A benchmark of continual instruction tuning for multimodel large language model. arXiv preprint arXiv:2403.08350, 2024. 3, 4, 6 +[7] Haokun Chen, Yao Zhang, Denis Krompass, Jindong Gu, and Volker Tresp. Feddat: An approach for foundation model finetuning in multi-modal heterogeneous federated learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11285-11293, 2024. 1 +[8] Adam Dahlgren Lindstrom and Savitha Sam Abraham. Clevr-math: A dataset for compositional language, visual and mathematical reasoning. In 16th International Workshop on Neural-Symbolic Learning and Reasoning, NeSy 2022, Windsor, UK, september 28-30, 2022. CEUR-WS, 2022. 4 +[9] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500, 2023. 3 + +[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 3 +[11] Jiahua Dong, Lixu Wang, Zhen Fang, Gan Sun, Shichao Xu, Xiao Wang, and Qi Zhu. Federated class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10164-10173, 2022. 1 +[12] Haiyang Guo, Fei Zhu, Wenzhuo Liu, Xu-Yao Zhang, and Cheng-Lin Liu. Pilora: Prototype guided incremental lora for federated class-incremental learning. In European Conference on Computer Vision, pages 141-159. Springer, 2024. 1, 3, 6 +[13] Haiyang Guo, Fei Zhu, Fanhu Zeng, Bing Liu, and XuYao Zhang. Desire: Dynamic knowledge consolidation for rehearsal-free continual learning. arXiv preprint arXiv:2411.19154, 2024. 2 +[14] Haiyang Guo, Fanhu Zeng, Ziwei Xiang, Fei Zhu, Da-Han Wang, Xu-Yao Zhang, and Cheng-Lin Liu. Hidellava: Hierarchical decoupling for continual instruction tuning of multimodal large language model. arXiv preprint arXiv:2503.12941, 2025. 1 +[15] Haiyang Guo, Fanhu Zeng, Fei Zhu, Jiayi Wang, Xukai Wang, Jingang Zhou, Hongbo Zhao, Wenzhuo Liu, Shijie Ma, Xu-Yao Zhang, et al. A comprehensive survey on continual learning in generative models. arXiv preprint arXiv:2506.13045, 2025. 1 +[16] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, Dawn Song, Jacob Steinhardt, and Justin Gilmer. The many faces of robustness: A critical analysis of out-of-distribution generalization. ICCV, 2021. 4 +[17] Tzu-Ming Harry Hsu, Hang Qi, and Matthew Brown. Measuring the effects of non-identical data distribution for federated visual classification. arXiv preprint arXiv:1909.06335, 2019.8 +[18] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3 +[19] Wenke Huang, Mang Ye, Zekun Shi, He Li, and Bo Du. Rethinking federated learning with domain shift: A prototype view. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16312-16322. IEEE, 2023. 4 +[20] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 1, 3 +[21] Kushal Kafle, Brian Price, Scott Cohen, and Christopher Kanan. Dvqa: Understanding data visualizations via question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5648-5656, 2018. 4 +[22] Samira Ebrahimi Kahou, Vincent Michalski, Adam Atkinson, Ákos Kádár, Adam Trischler, and Yoshua Bengio. Fig- + +ureqa: An annotated figure dataset for visual reasoning. arXiv preprint arXiv:1710.07300, 2017. 4 +[23] Gyuhak Kim, Changnan Xiao, Tatsuya Konishi, and Bing Liu. Learnability and algorithm for continual learning. In International Conference on Machine Learning, pages 16877-16896. PMLR, 2023. 2 +[24] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 6 +[25] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 4 +[26] Lei Li, Yuqi Wang, Runxin Xu, Peiyi Wang, Xiachong Feng, Lingpeng Kong, and Qi Liu. Multimodal arxiv: A dataset for improving scientific comprehension of large vision-language models. arXiv preprint arXiv:2403.00231, 2024. 4 +[27] Qinbin Li, Yiqun Diao, Quan Chen, and Bingsheng He. Federated learning on non-iid data silos: An experimental study. In 2022 IEEE 38th international conference on data engineering (ICDE), pages 965-978. IEEE, 2022. 2, 6 +[28] Tian Li, Anit Kumar Sahu, Ameet Talwalkar, and Virginia Smith. Federated learning: Challenges, methods, and future directions. IEEE signal processing magazine, 37(3):50-60, 2020. 1 +[29] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 6 +[30] Zhuowan Li, Xingrui Wang, Elias Stengel-Eskin, Adam Kortylewski, Wufei Ma, Benjamin Van Durme, and Alan L Yuille. Super-clevr: A virtual benchmark to diagnose domain robustness in visual reasoning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14963-14973, 2023. 4 +[31] Yi-Ming Lin, Yuan Gao, Mao-Guo Gong, Si-Jia Zhang, Yuan-Qiao Zhang, and Zhi-Yuan Li. Federated learning on multimodal data: A comprehensive survey. Machine Intelligence Research, 20(4):539-553, 2023. 1 +[32] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 3, 7 +[33] Wenzhuo Liu, Fei Zhu, Haiyang Guo, Longhui Wei, and Cheng-Lin Liu. Llava-c: Continual improved visual instruction tuning. arXiv preprint arXiv:2506.08666, 2025.1 +[34] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, et al. Deepseek-vl: towards real-world vision-language understanding. arXiv preprint arXiv:2403.05525, 2024. 1, 3 +[35] Pan Lu, Liang Qiu, Jiaqi Chen, Tony Xia, Yizhou Zhao, Wei Zhang, Zhou Yu, Xiaodan Liang, and Song-Chun Zhu. Iconqa: A new benchmark for abstract diagram understanding and visual language reasoning. In The 35th Conference on Neural Information Processing Systems (NeurIPS) Track on Datasets and Benchmarks, 2021. 4 + +[36] Pan Lu, Liang Qiu, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, Tanmay Rajpurohit, Peter Clark, and Ashwin Kalyan. Dynamic prompt learning via policy gradient for semi-structured mathematical reasoning. arXiv preprint arXiv:2209.14610, 2022. 4 +[37] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 4 +[38] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 1, 5, 7 +[39] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering by reading text in images. In 2019 international conference on document analysis and recognition (ICDAR), pages 947-952. IEEE, 2019. 4 +[40] Bryan A. Plummer, Liwei Wang, Christopher M. Cervantes, Juan C. Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. IJCV, 123 (1):74–93, 2017. 4 +[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 5, 7 +[42] Sashank Reddi, Zachary Charles, Manzil Zaheer, Zachary Garrett, Keith Rush, Jakub Konečny, Sanjiv Kumar, and H Brendan McMahan. Adaptive federated optimization. arXiv preprint arXiv:2003.00295, 2020. 8 +[43] Lorenzo Sani, Alex Iacob, Zeyu Cao, Bill Marino, Yan Gao, Tomas Paulik, Wanru Zhao, William F Shen, Preslav Aleksandrov, Xinchi Qiu, et al. The future of large language model pre-training is federated. arXiv preprint arXiv:2405.10853, 2024. 1 +[44] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge. In European conference on computer vision, pages 146–162. Springer, 2022. 4 +[45] Donald Shenaj, Marco Toldo, Alberto Rigon, and Pietro Zanuttigh. Asynchronous federated continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5055-5063, 2023. 3 +[46] Tianxiang Sun, Xiaotian Zhang, Zhengfu He, Peng Li, Qinyuan Cheng, Xiangyang Liu, Hang Yan, Yunfan Shao, Qiong Tang, Shiduo Zhang, et al. Moss: An open conversational large language model. Machine Intelligence Research, 21(5):888-905, 2024. 1 +[47] Yue Tan, Guodong Long, Lu Liu, Tianyi Zhou, Qinghua Lu, Jing Jiang, and Chengqi Zhang. Fedproto: Federated prototype learning across heterogeneous clients. In Proceedings of + +the AAAI conference on artificial intelligence, pages 8432-8440, 2022. 4 +[48] Yue Tan, Guodong Long, Jie Ma, Lu Liu, Tianyi Zhou, and Jing Jiang. Federated learning from pre-trained models: A contrastive learning approach. Advances in neural information processing systems, 35:19332-19344, 2022. 4 +[49] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1, 3 +[50] Xiao Wang, Tianze Chen, Qiming Ge, Han Xia, Rong Bao, Rui Zheng, Qi Zhang, Tao Gui, and Xuanjing Huang. Orthogonal subspace learning for language model continual learning. arXiv preprint arXiv:2310.14152, 2023. 5, 6 +[51] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 139-149, 2022. 6 +[52] Ziyao Wang, Zheyu Shen, Yexiao He, Guoheng Sun, Hongyi Wang, Lingjuan Lyu, and Ang Li. Flora: Federated fin-tuning large language models with heterogeneous low-rank adaptations. arXiv preprint arXiv:2409.05976, 2024. 5 +[53] Taiqiang Wu, Jiahao Wang, Zhe Zhao, and Ngai Wong. Mixture-of-subspaces in low-rank adaptation. arXiv preprint arXiv:2406.11909, 2024. 5 +[54] Xin Yang, Hao Yu, Xin Gao, Hao Wang, Junbo Zhang, and Tianrui Li. Federated continual learning via knowledge fusion: A survey. IEEE Transactions on Knowledge and Data Engineering, 2024. 3 +[55] Rui Ye, Wenhao Wang, Jingyi Chai, Dihan Li, Zexi Li, Yinda Xu, Yaxin Du, Yanfeng Wang, and Siheng Chen. Openfedllm: Training large language models on decentralized private data via federated learning. In Proceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining, pages 6137-6147, 2024. 1, 3, 4 +[56] Fanhu Zeng, Fei Zhu, Haiyang Guo, Xu-Yao Zhang, and Cheng-Lin Liu. Modalprompt: Dual-modality guided prompt for continual learning of large multimodal models. arXiv preprint arXiv:2410.05849, 2024. 1, 3, 4 +[57] Jie Zhang, Chen Chen, Weiming Zhuang, and Lingjuan Lyu. Target: Federated class-continual learning via exemplar-free distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4782–4793, 2023. 1 +[58] Jianyi Zhang, Saeed Vahidian, Martin Kuo, Chunyuan Li, Ruiyi Zhang, Tong Yu, Guoyin Wang, and Yiran Chen. Towards building the federatedgpt: Federated instruction tuning. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6915-6919. IEEE, 2024. 1, 3, 4 +[59] Xu-Yao Zhang, Cheng-Lin Liu, and Ching Y Suen. Towards robust pattern recognition: A review. Proceedings of the IEEE, 108(6):894-922, 2020. 1 +[60] Hongbo Zhao, Fei Zhu, Rundong Wang, Gaofeng Meng, and Zhaoxiang Zhang. Mllm-cl: Continual learning + +for multimodal large language models. arXiv preprint arXiv:2506.05453, 2025.1 +[61] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3 +[62] Fei Zhu, Xu-Yao Zhang, Chuang Wang, Fei Yin, and ChengLin Liu. Prototype augmentation and self-supervision for incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5871-5880, 2021. 6 +[63] Fei Zhu, Shijie Ma, Zhen Cheng, Xu-Yao Zhang, Zhaoxiang Zhang, and Cheng-Lin Liu. Open-world machine learning: A review and new outlooks. arXiv preprint arXiv:2403.01759, 2024. 1 \ No newline at end of file diff --git a/2025/Federated Continual Instruction Tuning/images.zip b/2025/Federated Continual Instruction Tuning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b59ac8d4ba2d69c9137dcf006dde14fbc5cde058 --- /dev/null +++ b/2025/Federated Continual Instruction Tuning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0c68bb6e43eec5de9da0ad7e10d6a6ce4c9cd2ab34977443deeca395e8a5e9d5 +size 831238 diff --git a/2025/Federated Continual Instruction Tuning/layout.json b/2025/Federated Continual Instruction Tuning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b5e512de457cb0e735c843fa4da992bff5fba2d3 --- /dev/null +++ b/2025/Federated Continual Instruction Tuning/layout.json @@ -0,0 +1,9049 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 181, + 103, + 430, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 103, + 430, + 121 + ], + "spans": [ + { + "bbox": [ + 181, + 103, + 430, + 121 + ], + "type": "text", + "content": "Federated Continual Instruction Tuning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 126, + 141, + 487, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 141, + 487, + 171 + ], + "spans": [ + { + "bbox": [ + 126, + 141, + 487, + 171 + ], + "type": "text", + "content": "Haiyang Guo" + }, + { + "bbox": [ + 126, + 141, + 487, + 171 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 126, + 141, + 487, + 171 + ], + "type": "text", + "content": ", Fanhu Zeng" + }, + { + "bbox": [ + 126, + 141, + 487, + 171 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 126, + 141, + 487, + 171 + ], + "type": "text", + "content": ", Fei Zhu" + }, + { + "bbox": [ + 126, + 141, + 487, + 171 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 126, + 141, + 487, + 171 + ], + "type": "text", + "content": ", Wenzhuo Liu" + }, + { + "bbox": [ + 126, + 141, + 487, + 171 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 126, + 141, + 487, + 171 + ], + "type": "text", + "content": ", Da-Han Wang" + }, + { + "bbox": [ + 126, + 141, + 487, + 171 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 126, + 141, + 487, + 171 + ], + "type": "text", + "content": ", Jian Xu" + }, + { + "bbox": [ + 126, + 141, + 487, + 171 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 126, + 141, + 487, + 171 + ], + "type": "text", + "content": ", Xu-Yao Zhang" + }, + { + "bbox": [ + 126, + 141, + 487, + 171 + ], + "type": "inline_equation", + "content": "^{1,2,3*}" + }, + { + "bbox": [ + 126, + 141, + 487, + 171 + ], + "type": "text", + "content": ", Cheng-Lin Liu" + }, + { + "bbox": [ + 126, + 141, + 487, + 171 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 171, + 481, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 481, + 185 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 481, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 133, + 171, + 481, + 185 + ], + "type": "text", + "content": "School of Advanced Interdisciplinary Sciences, UCAS " + }, + { + "bbox": [ + 133, + 171, + 481, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 171, + 481, + 185 + ], + "type": "text", + "content": "MAIS, CASIA" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 185, + 552, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 185, + 552, + 198 + ], + "spans": [ + { + "bbox": [ + 62, + 185, + 552, + 198 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 62, + 185, + 552, + 198 + ], + "type": "text", + "content": "School of Artificial Intelligence, UCAS " + }, + { + "bbox": [ + 62, + 185, + 552, + 198 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 62, + 185, + 552, + 198 + ], + "type": "text", + "content": "Centre for Artificial Intelligence and Robotics, HKISI-CAS" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 71, + 199, + 542, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 199, + 542, + 213 + ], + "spans": [ + { + "bbox": [ + 71, + 199, + 542, + 213 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 71, + 199, + 542, + 213 + ], + "type": "text", + "content": "FKLPRIU, School of Computer and Information Engineering, Xiamen University of Technology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 94, + 214, + 517, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 214, + 517, + 228 + ], + "spans": [ + { + "bbox": [ + 94, + 214, + 517, + 228 + ], + "type": "text", + "content": "{guohaiyang2023, zengfanhu2022, jian.xu} @ia.ac.cn, zhfei2018@gmail.com, {xyz, liucl} @nlpr.ia.ac.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "spans": [ + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 280, + 297, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 280, + 297, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 280, + 297, + 567 + ], + "type": "text", + "content": "A vast amount of instruction tuning data is crucial for the impressive performance of Large Multimodal Models (LMMs), but the associated computational costs and data collection demands during supervised fine-tuning make it impractical for most researchers. Federated learning (FL) has the potential to leverage all distributed data and training resources to reduce the overhead of joint training. However, most existing methods assume a fixed number of tasks, while in real-world scenarios, clients continuously encounter new knowledge and often struggle to retain old tasks due to memory constraints. In this work, we introduce the Federated Continual Instruction Tuning (FCIT) benchmark to model this real-world challenge. Our benchmark includes two realistic scenarios, encompassing four different settings and twelve carefully curated instruction tuning datasets. To address the challenges posed by FCIT, we propose a dynamic knowledge organization to effectively integrate updates from different tasks during training and subspace selective activation to allocate task-specific output during inference. Extensive experimental results demonstrate that our proposed method significantly enhances model performance across varying levels of data heterogeneity and catastrophic forgetting. Code and dataset are released at https://github.com/Ghy0501/FCIT." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 591, + 135, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 591, + 135, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 591, + 135, + 604 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 611, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 296, + 696 + ], + "type": "text", + "content": "Large Multimodal Models (LMMs) [2, 32, 34], which integrate Large Language Model [4, 20, 46, 49] with a visual encoder and multimodal projector to bridge visual and textual modalities, have exhibited impressive visual understanding and complex reasoning abilities. A crucial factor in this success is the supervised fine-tuning of LMMs using huge and diverse visual instruction-following data [32]" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 256, + 554, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 256, + 554, + 364 + ], + "spans": [ + { + "bbox": [ + 313, + 256, + 554, + 364 + ], + "type": "text", + "content": "to align with human preferences. However, collecting such vast amounts of training data and computational resources for joint fine-tuning is impractical for most researchers. Federated Learning (FL) [28, 31, 38, 43], as a decentralized paradigm, offers a viable alternative by leveraging distributed data and computational resources for local training while integrating local weights to produce a unified model. This paradigm accommodates constraints on storage and computation while ensuring privacy protection." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 365, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 700 + ], + "type": "text", + "content": "Generally, most existing FL frameworks [7, 55, 58] are modeled in static, closed-world scenarios [3, 59], where a fixed set of tasks is predefined and unchanged. However, real-world applications are dynamic [59, 63], requiring models to continuously acquire new knowledge while retaining previously learned tasks. Taking a realistic health emergency as an example, different hospitals act as local clients when a major disease outbreak occurs. Large hospitals can leverage FL to collaboratively train on their case data, building a comprehensive virus knowledge base, while smaller clinics update their own cases using this global knowledge. Over time, both large and small hospitals need to keep updating the global knowledge base through continual learning methods to cope with the event. Additionally, the emergence of new strains necessitates the simultaneous integration of knowledge and response strategies to minimize potential losses. In this scenario, traditional FL methods struggle with newly arrived knowledge, while Continual Learning (CL) [14, 15, 33, 56, 60] methods alone do not facilitate knowledge sharing between clients. Only an organic integration of both can address this real-world problem. Recently, numerous Federated Continual Learning methods have emerged to address this challenge in traditional image classification tasks [11, 12, 57]. However, their methodological and task-setting limitations make them insufficient for current LMM applications. Therefore, a more comprehensive benchmark is needed to better simulate the practical application of LMMs in real-world scenarios." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 325, + 701, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 553, + 712 + ], + "type": "text", + "content": "In this work, we first establish a Federated Continual In" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "type": "text", + "content": "*Corresponding Author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1325" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 73, + 545, + 322 + ], + "blocks": [ + { + "bbox": [ + 63, + 73, + 545, + 322 + ], + "lines": [ + { + "bbox": [ + 63, + 73, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 63, + 73, + 545, + 322 + ], + "type": "image", + "image_path": "1172980ffe523979c81b514f577a783b638694ff4f32a162ac2b3c43a6229dc7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 328, + 555, + 352 + ], + "lines": [ + { + "bbox": [ + 55, + 328, + 555, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 555, + 352 + ], + "type": "text", + "content": "Figure 1. Overview of FCIT benchmark. FCIT encompasses 2 real-world scenarios, 4 FCIT settings, and 12 curated datasets, providing a comprehensive simulation of LMMs instruction-following training in real-world applications." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 361, + 296, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 296, + 638 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 296, + 638 + ], + "type": "text", + "content": "struction Tuning (FCIT) benchmark to fill this gap. Specifically, we simulate two client-level realistic scenarios: (i) Homogeneous FCIT refers to an FL system learning a series of instruction tuning data sequentially, with different clients learning the same task at each stage. (ii) Heterogeneous FCIT, on the other hand, requires the FL system to collaborate on different tasks simultaneously, with different clients potentially learning different tasks in the same stage. Building on this, we define two settings of datasets for each scenario: Capability-related and Task-related. The former evaluates the model's ability to integrate multiple dimensions of instruction following in a short period, while the latter assesses the model's performance during long phases of continual learning. For dataset selection, we curate twelve instruction tuning datasets unseen during LMM fine-tuning or with low zero-shot performance, preventing information leakage [13, 23]. In addition, we introduce varying degrees of data heterogeneity for each setting to challenge the model's performance in the non-IID situation [27]. An illustration of our FCIT benchmark is provided in Figure 1. To the best of our knowledge, this is the first work to introduce a comprehensive benchmark for federated learning of LMM in a continual learning setting." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "To effectively address the challenges posed by FCIT, we propose a novel Dynamic knowledge organization and Subspace seleCtive activatiOn (DISCO) framework. Specifically, we identify FCIT challenges into two types: conflicts between different tasks within the same stage and conflicts between old and new tasks across different stages." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 361, + 555, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 555, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 555, + 589 + ], + "type": "text", + "content": "The former can cause catastrophic forgetting by altering the parameter space of previous tasks when learning new ones without access to past data, while the latter requires the model to integrate and organize knowledge from different tasks to derive unified representations. Therefore, we first propose Dynamic Knowledge Organization (DKO), which leverages a dynamic cache at the global server to store task-specific parameters. Using a unique identity token matching mechanism, it systematically organizes knowledge for different tasks into corresponding subspaces within the cache, effectively mitigating two types of conflicts. To better utilize the organized task subspaces in the dynamic cache, we introduce Subspace Selective Activation (SSA), which selectively activates subspaces relevant to the test input while filtering out irrelevant outputs, leading to significant performance improvements. Consequently, these designs enable our framework to efficiently tackle the data heterogeneity and catastrophic forgetting in FCIT. In summary, our major contributions are:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 594, + 556, + 714 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 594, + 555, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 594, + 555, + 654 + ], + "spans": [ + { + "bbox": [ + 314, + 594, + 555, + 654 + ], + "type": "text", + "content": "- We present the first Federated Continual Instruction Tuning (FCIT) benchmark designed for real-world scenarios, providing a comprehensive evaluation of LMMs to continuously learn new knowledge using distributed data and training resources in real applications." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 654, + 556, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 654, + 556, + 701 + ], + "spans": [ + { + "bbox": [ + 314, + 654, + 556, + 701 + ], + "type": "text", + "content": "- We propose a novel DISCO framework that integrates dynamic knowledge organization and subspace selective activation to efficiently address data heterogeneity and catastrophic forgetting in FCIT settings." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 314, + 701, + 553, + 714 + ], + "type": "text", + "content": "- Extensive experiments demonstrate that our method sig" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1326" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 72, + 295, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 72, + 295, + 109 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 295, + 109 + ], + "type": "text", + "content": "nificantly enhances model performance under data heterogeneity while minimizing the catastrophic forgetting, and achieves state-of-the-art performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 119, + 143, + 132 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 119, + 143, + 132 + ], + "spans": [ + { + "bbox": [ + 55, + 119, + 143, + 132 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 138, + 295, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 138, + 295, + 330 + ], + "spans": [ + { + "bbox": [ + 55, + 138, + 295, + 330 + ], + "type": "text", + "content": "Large Multimodal Models. With the grand unification of Large Language Models (LLMs) [4, 20, 49] for various NLP tasks, Large Multimodal Models (LMMs) [2, 9, 32, 34, 61] emerge by extending LLMs to combine with visual encoder and multimodal projectors, demonstrating exceptional visual understanding and complex reasoning abilities. To better align with human preferences, these LMMs typically undergo further fine-tuning on extensive instruction-following data, ensuring they meet the demands of real-world applications [32]. However, in real-world scenarios, improving the performance of LMMs on new downstream tasks becomes a significant challenge without access to sufficient training data and computational resources. In this paper, we introduce the first Federated Continual Instruction Tuning (FCIT) benchmark to bridge this gap from the perspective of distributed training and continual learning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 331, + 295, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 295, + 545 + ], + "type": "text", + "content": "Federated Continual Learning. In classical vision tasks, Federated Continual Learning (FCL) [54] aims to adapt the global model to new data while maintaining the knowledge of the old task. From the perspective of LMMs, research in this setting faces two main challenges: (1) Most of the methods are designed for traditional vision tasks (i.e., image classification). For instance, MFCL [1] employs a generative model to synthesize images of previously learned classes, thereby mitigating forgetting. PILoRA [12] introduces a prototype re-weight module to address the classifier bias caused by data heterogeneity and obtain unified knowledge through LoRA [18] fusion. Despite progress in classification tasks, complex designs remain challenging to adapt for LMMs research. (2) FCL follows a single data composition by splitting a dataset (e.g. ImageNet [10]) into different tasks based on classes, whereas LMMs face greater challenges in continual learning due to their diverse tasks with varying styles [5, 6, 56]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 545, + 295, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 642 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 642 + ], + "type": "text", + "content": "This work pioneers a Federated Continual Instruction Tuning setup for LMMs and establishes diverse scenarios to simulate real-world applications comprehensively. Notably, AFCL [45] is most relevant to our work, as it enables clients to continuously learn multiple tasks in different orders and asynchronous time slots. However, it is specifically designed for image classification tasks, whereas our study focuses on the more widely used LMM." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 652, + 182, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 652, + 182, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 652, + 182, + 664 + ], + "type": "text", + "content": "3. Problem Formulation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 672, + 141, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 141, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 141, + 683 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "Instruction tuning enhances LMM's ability to understand and execute human instructions by performing supervised" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "content": "fine-tuning of pre-trained models on extensive datasets comprising instructions and responses. Formally, the instruction data " + }, + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(\\mathbf{x}_v^j,\\mathbf{x}_{ins}^j,\\mathbf{x}_{res}^j)^N\\}" + }, + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "content": " consists of image input " + }, + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_v" + }, + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "content": ", instruction " + }, + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{ins}" + }, + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "content": " and response " + }, + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{res}" + }, + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "content": " represents the total number of samples. For clarity, given a simple image-instruction pair with a response of length " + }, + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "content": ", the objective of an LMM is to predict the next token autoregressively based on all preceding tokens:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 338, + 173, + 553, + 206 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 173, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 338, + 173, + 553, + 206 + ], + "type": "interline_equation", + "content": "p \\left(\\mathbf {x} _ {r e s} \\mid \\mathbf {x} _ {v}, \\mathbf {x} _ {i n s}\\right) = \\prod_ {i = 1} ^ {L} p _ {\\theta} \\left(x _ {i} \\mid \\mathbf {x} _ {v}, \\mathbf {x} _ {i n s}, \\mathbf {x} _ {r e s, < i}\\right), \\tag {1}", + "image_path": "3785c6f41592d12612566338cdbc6f5d2384c478e81826ee987255c7ce9b0339.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 210, + 555, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 210, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 210, + 555, + 258 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 210, + 555, + 258 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 210, + 555, + 258 + ], + "type": "text", + "content": " denotes the trainable parameters during fine-tuning, " + }, + { + "bbox": [ + 313, + 210, + 555, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{res}, < i}" + }, + { + "bbox": [ + 313, + 210, + 555, + 258 + ], + "type": "text", + "content": " denotes all response tokens preceding the current prediction token " + }, + { + "bbox": [ + 313, + 210, + 555, + 258 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 210, + 555, + 258 + ], + "type": "text", + "content": ". Then, the loss function of fine-tuning LMMs can be expressed as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 338, + 264, + 553, + 299 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 264, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 338, + 264, + 553, + 299 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\theta} = - \\frac {1}{N} \\sum_ {j = 1} ^ {N} \\sum_ {i = 1} ^ {L _ {j}} \\log p _ {\\theta} \\left(x _ {i} ^ {j} \\mid \\mathbf {x} _ {v} ^ {j}, \\mathbf {x} _ {i n s} ^ {j}, \\mathbf {x} _ {r e s, < i} ^ {j}\\right). \\tag {2}", + "image_path": "66d24c98abff44da06eefaf3efe59ac528aa687675a7459b8f183b683ceb6b34.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 303, + 553, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 303, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 553, + 386 + ], + "type": "text", + "content": "Federated learning framework typically comprises a global server and several local clients, all employing the same LMM with a shared homogeneous model architecture. In each communication round1, local clients train their models on own data and upload the updated weights to the global server for aggregation, enabling collaborative optimization of the global model while preserving data privacy." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 387, + 554, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 554, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 554, + 469 + ], + "type": "text", + "content": "Considering the excessive communication overhead of transferring the entire LMM between clients and the global server, we adopt LoRA [18] for efficient fine-tuning, balancing training overhead and implementation cost [55, 58]. Specifically, for a weight matrix " + }, + { + "bbox": [ + 313, + 387, + 554, + 469 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_0\\in \\mathbb{R}^{d\\times k}" + }, + { + "bbox": [ + 313, + 387, + 554, + 469 + ], + "type": "text", + "content": ", LoRA decomposes parameter updates " + }, + { + "bbox": [ + 313, + 387, + 554, + 469 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{W}" + }, + { + "bbox": [ + 313, + 387, + 554, + 469 + ], + "type": "text", + "content": " during fine-tuning into two low-rank subspaces:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 370, + 476, + 553, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 476, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 370, + 476, + 553, + 488 + ], + "type": "interline_equation", + "content": "\\mathbf {W} = \\mathbf {W} _ {0} + \\Delta \\mathbf {W} = \\mathbf {W} _ {0} + \\mathbf {B A}, \\tag {3}", + "image_path": "956bc2a6486c29dbe8b50012ba79fd0933b361c28206ca2dd2231190c1bd9291.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 494, + 517, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 494, + 517, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 494, + 517, + 506 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 494, + 517, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{B} \\in \\mathbb{R}^{d \\times r}" + }, + { + "bbox": [ + 313, + 494, + 517, + 506 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 494, + 517, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in \\mathbb{R}^{r \\times k}" + }, + { + "bbox": [ + 313, + 494, + 517, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 494, + 517, + 506 + ], + "type": "inline_equation", + "content": "r \\ll \\min\\{d, k\\}" + }, + { + "bbox": [ + 313, + 494, + 517, + 506 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 506, + 553, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 553, + 555 + ], + "type": "text", + "content": "Continual learning aims to minimize the loss on the current task while retaining knowledge from previous tasks. Formally, given a sequence of datasets " + }, + { + "bbox": [ + 313, + 506, + 553, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_1,\\mathcal{D}_2,\\dots ,\\mathcal{D}_T" + }, + { + "bbox": [ + 313, + 506, + 553, + 555 + ], + "type": "text", + "content": " , the optimization objective at task " + }, + { + "bbox": [ + 313, + 506, + 553, + 555 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 506, + 553, + 555 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 343, + 560, + 553, + 596 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 560, + 553, + 596 + ], + "spans": [ + { + "bbox": [ + 343, + 560, + 553, + 596 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\mathcal {L} (\\theta) = \\mathbb {E} _ {(x, y) \\sim \\mathcal {D} _ {t}} [ \\mathcal {L} (f _ {\\theta_ {t}} (x), y) ] + \\sum_ {i = 1} ^ {t - 1} \\epsilon_ {i}, \\tag {4}", + "image_path": "0a0999a371e644edf0ea25631d272d114c30fbaa9eca13d885411ab3ee43fea8.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 325, + 595, + 532, + 623 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 595, + 532, + 623 + ], + "spans": [ + { + "bbox": [ + 325, + 595, + 532, + 623 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\operatorname {s. t.} \\mathbb {E} _ {(x, y) \\sim \\mathcal {D} _ {i}} \\left[ \\mathcal {L} \\left(f _ {\\theta_ {t}} (x), y\\right) - \\mathcal {L} \\left(f _ {\\theta_ {t - 1}} (x), y\\right) \\right] \\leq \\epsilon_ {i}, \\\\ \\epsilon_ {i} \\geq 0; \\forall i \\in [ 1, \\dots , t - 1 ], \\\\ \\end{array}", + "image_path": "b68e603f393944f18efa4585eddf8bdf3eabad2b3652ac87e11f51c86c57250e.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 627, + 553, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 627, + 553, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 627, + 553, + 688 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 627, + 553, + 688 + ], + "type": "inline_equation", + "content": "\\epsilon_{i}" + }, + { + "bbox": [ + 313, + 627, + 553, + 688 + ], + "type": "text", + "content": " is a slack variable that allows a small increase in the loss from the old datasets, providing tolerance for minor forgetting while focusing on learning the current task. In Eq. 4, " + }, + { + "bbox": [ + 313, + 627, + 553, + 688 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 627, + 553, + 688 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 627, + 553, + 688 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 627, + 553, + 688 + ], + "type": "text", + "content": " can be viewed as the multimodal inputs " + }, + { + "bbox": [ + 313, + 627, + 553, + 688 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}_v,\\mathbf{x}_{ins})" + }, + { + "bbox": [ + 313, + 627, + 553, + 688 + ], + "type": "text", + "content": " and the response " + }, + { + "bbox": [ + 313, + 627, + 553, + 688 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{res}" + }, + { + "bbox": [ + 313, + 627, + 553, + 688 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 693, + 553, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 553, + 712 + ], + "type": "text", + "content": "For clarification, we standardize the communication round in Section 4 to 1." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1327" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 266, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 266, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 266, + 85 + ], + "type": "text", + "content": "3.2. Federated Continual Instruction Tuning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 93, + 296, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 93, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 93, + 296, + 285 + ], + "type": "text", + "content": "As shown in Figure 1, we integrate federated and continual learning for LMMs within a unified framework and propose two client-level realistic scenarios: (1) Homogeneous FCIT (Hom-FCIT). In this scenario, clients sequentially learn a series of tasks, allowing the global server to continuously update its knowledge. Each client learns the same task at a given stage and can only access data from that stage. (2) Heterogeneous FCIT (Het-FCIT). In real-world applications, different clients may learn different tasks within the same stage, enabling the global server to respond more rapidly to diverse instructions. This requires the model not only to coordinate the knowledge of different tasks learned by clients in the current stage but also to mitigate forgetting during the learning process. For each scenario, we define two dataset-level settings: capability-related and task-related, detailed as follows." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 295, + 296, + 487 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 55, + 295, + 296, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 295, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 296, + 426 + ], + "type": "text", + "content": "- Capability-related. Following the dataset construction and division in LLaVA-OneVision [25], we classify the 12 datasets into 4 capabilities: General, Math, Chart, and Other. The General capability includes A-OKVQA [44], ImageNet-R [16], Grounding [37], and IconQA [35]; Math comprises CLEVR-Math [8], super-CLEVR [30], and TabMWP [36]; Chart involves ArxivQA [26], FigureQA [22], and DVQA [21]; and Other encompasses OCR-VQA [39] and Flickr30k [40]. We treat these four capabilities as different stages of continual learning, with each capability comprising a mixture of datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 427, + 296, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 427, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 55, + 427, + 296, + 487 + ], + "type": "text", + "content": "- Task-related. To evaluate the performance of different methods in a long-phase continual learning situation, we selected 8 datasets: ImageNet-R, ArxivQA, IconQA, CLEVR-Math, OCR-VQA, Flickr30k, FigureQA, and super-CLEVR as distinct stages of continual learning." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 498, + 296, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 558 + ], + "type": "text", + "content": "In total, we propose 4 FCIT settings to evaluate different methods. Compared to existing works [5, 6, 55, 56, 58], we are the first to explore the organic integration of FL and CL in the context of LMM. We provide further illustrations on the settings and datasets in Appendix A." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 581, + 249, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 249, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 249, + 594 + ], + "type": "text", + "content": "4. The Proposed Framework: DISCO" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "Overview of the Method. As shown in Figure 2 and 3, our method primarily consists of: (1) Dynamic Knowledge Organization (DKO), which dynamically integrates knowledge learned by different clients across stages during training, significantly reducing inter-task conflicts (Section 4.1); and (2) Subspace Selective Activation (SSA), which selectively activates subspace outputs based on input features during inference, effectively filtering out irrelevant information (Section 4.2). We name this framework DISCO." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 322, + 73, + 542, + 276 + ], + "blocks": [ + { + "bbox": [ + 322, + 73, + 542, + 276 + ], + "lines": [ + { + "bbox": [ + 322, + 73, + 542, + 276 + ], + "spans": [ + { + "bbox": [ + 322, + 73, + 542, + 276 + ], + "type": "image", + "image_path": "2069e575ef5e18f160271e66931eeb8285660d720ea8a7abe549298323d281bc.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 281, + 555, + 314 + ], + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 314 + ], + "type": "text", + "content": "Figure 2. Illustration of the proposed DKO. Dynamic caches store the knowledge of each subspace while matching and updating through identity tokens." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 318, + 499, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 499, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 499, + 331 + ], + "type": "text", + "content": "4.1. Dynamic Knowledge Organization" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 335, + 555, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 555, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 555, + 467 + ], + "type": "text", + "content": "The core challenge of FCIT is enabling the global server to effectively harmonize knowledge learned by clients, addressing both the forgetting of previous knowledge when acquiring new tasks and the conflicts that arise from integrating knowledge from different tasks within the same stage. To this end, we propose maintaining a dynamic cache at the global server to organize task-specific knowledge uploaded by clients at different stages, thereby preventing both forgetting and conflicts between knowledge. Specifically, each task is assigned a dedicated parameter space to store and update its corresponding knowledge:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 359, + 473, + 555, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 473, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 359, + 473, + 555, + 502 + ], + "type": "interline_equation", + "content": "\\Delta \\mathbf {W} = \\mathbf {B} \\mathbf {A} \\Leftrightarrow \\underbrace {\\left\\{\\mathbf {B} _ {1} \\mathbf {A} _ {1} , \\cdots , \\mathbf {B} _ {T} \\mathbf {A} _ {T} \\right\\}} _ {\\text {t a s k - s p e c i f i c s u b s p a c e}}, \\tag {5}", + "image_path": "34f92882a82c5fe12b6b359d9c18af3c78af7b54d0fa8c0f9581216862458546.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 507, + 554, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 554, + 544 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 507, + 554, + 544 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 507, + 554, + 544 + ], + "type": "text", + "content": " is the number of tasks learned. At this point, it is crucial to aggregate the parameters uploaded by clients into their respective subspaces without privacy leakage." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 544, + 554, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 554, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 554, + 651 + ], + "type": "text", + "content": "Inspired by the widespread use of prototypes in FL fields [19, 47, 48], we propose to distinguish knowledge across tasks by using the feature mean of each client's training data as an Identity Token. Specifically, considering the uniqueness of textual inputs in the visual instruction tuning task, we introduce a text encoder " + }, + { + "bbox": [ + 313, + 544, + 554, + 651 + ], + "type": "inline_equation", + "content": "f_{ins}" + }, + { + "bbox": [ + 313, + 544, + 554, + 651 + ], + "type": "text", + "content": " for each client to extract the feature of the input instruction " + }, + { + "bbox": [ + 313, + 544, + 554, + 651 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{ins,j}^{t}" + }, + { + "bbox": [ + 313, + 544, + 554, + 651 + ], + "type": "text", + "content": " during training, and take the mean value " + }, + { + "bbox": [ + 313, + 544, + 554, + 651 + ], + "type": "inline_equation", + "content": "\\mu_k^t" + }, + { + "bbox": [ + 313, + 544, + 554, + 651 + ], + "type": "text", + "content": " as the local identity token of the " + }, + { + "bbox": [ + 313, + 544, + 554, + 651 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 544, + 554, + 651 + ], + "type": "text", + "content": "-th client on the task " + }, + { + "bbox": [ + 313, + 544, + 554, + 651 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 544, + 554, + 651 + ], + "type": "text", + "content": " at the end of training:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 379, + 658, + 553, + 693 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 379, + 658, + 553, + 693 + ], + "spans": [ + { + "bbox": [ + 379, + 658, + 553, + 693 + ], + "type": "interline_equation", + "content": "\\mu_ {k} ^ {t} = \\frac {1}{n _ {k} ^ {t}} \\sum_ {j = 1} ^ {n _ {k} ^ {t}} f _ {i n s} \\left(\\mathbf {x} _ {i n s, j} ^ {t}\\right), \\tag {6}", + "image_path": "9d782fc0e8ed778c0804d0fafe528b2e5b9d5052cfe409fdc9637fbdffa4f2f1.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 701, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 555, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 701, + 555, + 715 + ], + "type": "inline_equation", + "content": "n_k^t" + }, + { + "bbox": [ + 313, + 701, + 555, + 715 + ], + "type": "text", + "content": " is the number of training samples of client " + }, + { + "bbox": [ + 313, + 701, + 555, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 701, + 555, + 715 + ], + "type": "text", + "content": " at" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1328" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 71, + 547, + 190 + ], + "blocks": [ + { + "bbox": [ + 65, + 71, + 547, + 190 + ], + "lines": [ + { + "bbox": [ + 65, + 71, + 547, + 190 + ], + "spans": [ + { + "bbox": [ + 65, + 71, + 547, + 190 + ], + "type": "image", + "image_path": "321710d579552cd17ea1e7d5f77fe3666442909e13c48150ba9e12d9c387c1ca.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 192, + 555, + 217 + ], + "lines": [ + { + "bbox": [ + 55, + 192, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 555, + 217 + ], + "type": "text", + "content": "Figure 3. Illustration of the proposed SSA. Each subspace organized by DKO dynamically adjusts its output via the intrinsic activation matrix during inference, effectively filtering irrelevant information and enhancing model performance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 221, + 295, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 221, + 295, + 257 + ], + "spans": [ + { + "bbox": [ + 55, + 221, + 295, + 257 + ], + "type": "text", + "content": "task " + }, + { + "bbox": [ + 55, + 221, + 295, + 257 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 221, + 295, + 257 + ], + "type": "text", + "content": ", and we use CLIP's text encoder [41] in our experiments. Notably, the text encoder " + }, + { + "bbox": [ + 55, + 221, + 295, + 257 + ], + "type": "inline_equation", + "content": "f_{ins}" + }, + { + "bbox": [ + 55, + 221, + 295, + 257 + ], + "type": "text", + "content": " remains frozen during training, which prevents additional training overhead." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 257, + 295, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 257, + 295, + 294 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 295, + 294 + ], + "type": "text", + "content": "Using the identity tokens, we first apply cosine similarity along with a threshold to match the uploaded local identity token " + }, + { + "bbox": [ + 55, + 257, + 295, + 294 + ], + "type": "inline_equation", + "content": "\\mu_j^t" + }, + { + "bbox": [ + 55, + 257, + 295, + 294 + ], + "type": "text", + "content": " with the global identity token " + }, + { + "bbox": [ + 55, + 257, + 295, + 294 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{z}}_i" + }, + { + "bbox": [ + 55, + 257, + 295, + 294 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 303, + 295, + 334 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 303, + 295, + 334 + ], + "spans": [ + { + "bbox": [ + 143, + 303, + 295, + 334 + ], + "type": "interline_equation", + "content": "\\frac {\\mu_ {j} ^ {t} \\cdot \\tilde {\\mathbf {z}} _ {i}}{\\| \\mu_ {j} ^ {t} \\| \\| \\tilde {\\mathbf {z}} _ {i} \\|} \\geq \\tau , \\tag {7}", + "image_path": "55035861bbd9a3bdeeeb84f3e29dc0b479b5fffa6b1243f2c36bc8a7fcad69c6.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 341, + 296, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 341, + 296, + 401 + ], + "spans": [ + { + "bbox": [ + 55, + 341, + 296, + 401 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 341, + 296, + 401 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 341, + 296, + 401 + ], + "type": "text", + "content": " denotes a pre-defined threshold. For paired local identity tokens, we update the corresponding global identity token. For mismatched local identity tokens, we pair them two by two using Eq. (7) and initialize new global identity tokens. This process can be formalized as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 78, + 410, + 295, + 452 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 410, + 295, + 452 + ], + "spans": [ + { + "bbox": [ + 78, + 410, + 295, + 452 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {z}} _ {i} ^ {*} = \\left\\{ \\begin{array}{l l} \\frac {n _ {\\bar {\\mathbf {z}} _ {i}} \\cdot \\tilde {\\mathbf {z}} _ {i} + \\sum_ {j = 1} ^ {m} n _ {j} ^ {t} \\cdot \\mu_ {j} ^ {t}}{n _ {\\bar {\\mathbf {z}} _ {i}} + \\sum_ {j = 1} ^ {m} n _ {j} ^ {t}}, & \\text {i f} \\tilde {\\mathbf {z}} _ {i} \\text {e x i s t s}, \\\\ \\frac {\\sum_ {j = 1} ^ {m} n _ {j} ^ {t} \\cdot \\mu_ {j} ^ {t}}{\\sum_ {j = 1} ^ {m} n _ {j} ^ {t}}, & \\text {i f} \\tilde {\\mathbf {z}} _ {i} \\text {d o e s n ’ t e x i s t}, \\end{array} \\right. \\tag {8}", + "image_path": "04987089276a7daf79b81973b8f82b70227962d3918b3f45bb911e3285026a01.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 460, + 296, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 296, + 534 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 460, + 296, + 534 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 460, + 296, + 534 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 460, + 296, + 534 + ], + "type": "inline_equation", + "content": "n_{\\overline{\\mathbf{z}}_i}" + }, + { + "bbox": [ + 55, + 460, + 296, + 534 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 460, + 296, + 534 + ], + "type": "inline_equation", + "content": "n_j^t" + }, + { + "bbox": [ + 55, + 460, + 296, + 534 + ], + "type": "text", + "content": " denote the number of matched local identity tokens, the number of samples used to form the previous " + }, + { + "bbox": [ + 55, + 460, + 296, + 534 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 460, + 296, + 534 + ], + "type": "text", + "content": "-th global identity token, and the number of samples from the local client " + }, + { + "bbox": [ + 55, + 460, + 296, + 534 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 460, + 296, + 534 + ], + "type": "text", + "content": " at task " + }, + { + "bbox": [ + 55, + 460, + 296, + 534 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 460, + 296, + 534 + ], + "type": "text", + "content": ", respectively. Then, we leverage this matching process to guide the update or initialization of task-specific subspace " + }, + { + "bbox": [ + 55, + 460, + 296, + 534 + ], + "type": "inline_equation", + "content": "\\{\\theta_i = \\mathbf{B}_i\\mathbf{A}_i\\}" + }, + { + "bbox": [ + 55, + 460, + 296, + 534 + ], + "type": "text", + "content": " in Eq. (5):" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 542, + 295, + 556 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 542, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 542, + 295, + 556 + ], + "type": "interline_equation", + "content": "\\theta_ {i} ^ {*} = \\mathcal {F} \\left(\\theta_ {1} ^ {t}, \\dots , \\theta_ {m} ^ {t}\\right), \\tag {9}", + "image_path": "e786e735276ed20c8a0149f9303989b593213928cd374640afea99d15f3fef9d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 565, + 295, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 565, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 565, + 295, + 613 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 565, + 295, + 613 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 565, + 295, + 613 + ], + "type": "text", + "content": " denotes the FL algorithm (e.g. FedAvg [38]) used to aggregate local weights " + }, + { + "bbox": [ + 55, + 565, + 295, + 613 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 55, + 565, + 295, + 613 + ], + "type": "text", + "content": ". As a result, we effectively prevent inter-task conflicts and integrate knowledge from different clients using the identity token matching mechanism." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 613, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 613, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 613, + 296, + 685 + ], + "type": "text", + "content": "After the global server completes the aggregation, it distributes the dynamic cache to each selected client. The client then matches the identity token (Using Eq. (7)) of each subspace with its own training data, deciding whether to update the corresponding subspace or reinitialize a new one. The entire DKO process is illustrated in Figure 2." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 220, + 476, + 233 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 476, + 233 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 476, + 233 + ], + "type": "text", + "content": "4.2. Subspace Selective Activation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 238, + 554, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 238, + 554, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 238, + 554, + 285 + ], + "type": "text", + "content": "Section 4.1 effectively mitigates inter-task conflicts by disentangling and organizing complex knowledge into distinct subspaces. The key challenge then lies in how to leverage these task-specific subspaces during inference." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 286, + 554, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 286, + 554, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 286, + 554, + 417 + ], + "type": "text", + "content": "To address this, an intuitive method is to concatenate the subspaces of the dynamic cache in low-rank dimensions [50, 52], enabling the integration of knowledge across all task spaces. However, this may introduce redundant information unrelated to the current task during inference, potentially compromising the model's output. For instance, when the desired answer is a simple word (e.g., what is the object in the picture), knowledge from other subspaces, such as those used for generating long-form descriptions, can introduce unnecessary information, leading to responses that do not align with the given instruction." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 418, + 554, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 554, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 554, + 501 + ], + "type": "text", + "content": "Drawing inspiration from LoRA's intrinsic space [53], we propose subspace selective activation (SSA) without additional training to filter out irrelevant subspace outputs, ensuring alignment between responses and instructions. In particular, A vanilla LoRA can be decomposed as the product of two low-rank subspaces (i.e., " + }, + { + "bbox": [ + 313, + 418, + 554, + 501 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 313, + 418, + 554, + 501 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 418, + 554, + 501 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 313, + 418, + 554, + 501 + ], + "type": "text", + "content": ") and an intrinsic mixing matrix " + }, + { + "bbox": [ + 313, + 418, + 554, + 501 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 313, + 418, + 554, + 501 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 403, + 508, + 553, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 508, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 403, + 508, + 553, + 520 + ], + "type": "interline_equation", + "content": "\\Delta \\mathbf {W} = \\mathbf {B} \\mathcal {W} \\mathbf {A}, \\tag {10}", + "image_path": "3275d88be16100499062b7155773d86b26138c5c9803b06037fb5ff2507c5c19.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 528, + 554, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 554, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 554, + 647 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 528, + 554, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{W} \\in \\mathbb{R}^{r \\times r}" + }, + { + "bbox": [ + 313, + 528, + 554, + 647 + ], + "type": "text", + "content": " is typically an identity matrix and can thus be omitted. In this paper, we redefine " + }, + { + "bbox": [ + 313, + 528, + 554, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 313, + 528, + 554, + 647 + ], + "type": "text", + "content": " as the activation matrix that dynamically responds to the test input. Specifically, we treat " + }, + { + "bbox": [ + 313, + 528, + 554, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 313, + 528, + 554, + 647 + ], + "type": "text", + "content": " as the product of an identity matrix and an activation factor (i.e., " + }, + { + "bbox": [ + 313, + 528, + 554, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{W} = \\alpha \\cdot \\mathbf{I}_{r \\times r}" + }, + { + "bbox": [ + 313, + 528, + 554, + 647 + ], + "type": "text", + "content": "), where " + }, + { + "bbox": [ + 313, + 528, + 554, + 647 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 313, + 528, + 554, + 647 + ], + "type": "text", + "content": " denotes full activation, and " + }, + { + "bbox": [ + 313, + 528, + 554, + 647 + ], + "type": "inline_equation", + "content": "\\alpha = 0" + }, + { + "bbox": [ + 313, + 528, + 554, + 647 + ], + "type": "text", + "content": " means that the output is fully masked. Therefore, for each subspace " + }, + { + "bbox": [ + 313, + 528, + 554, + 647 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{B}_1\\mathbf{A}_1, \\dots, \\mathbf{B}_T\\mathbf{A}_T\\}" + }, + { + "bbox": [ + 313, + 528, + 554, + 647 + ], + "type": "text", + "content": " in dynamic cache, we can control the activation or inhibition of the corresponding output by adjusting its activation factor " + }, + { + "bbox": [ + 313, + 528, + 554, + 647 + ], + "type": "inline_equation", + "content": "\\{\\alpha_1, \\dots, \\alpha_T\\}" + }, + { + "bbox": [ + 313, + 528, + 554, + 647 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 647, + 554, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 554, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 554, + 684 + ], + "type": "text", + "content": "To provide better flexibility in assigning activation factors, we use global identity tokens and test input features extracted by the text encoder for similarity matching:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 379, + 689, + 553, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 379, + 689, + 553, + 717 + ], + "spans": [ + { + "bbox": [ + 379, + 689, + 553, + 717 + ], + "type": "interline_equation", + "content": "s _ {i} = \\frac {\\tilde {\\mathbf {z}} _ {\\mathbf {i}} \\cdot f _ {i n s} \\left(\\mathbf {x} _ {i n s} ^ {t e s t}\\right)}{\\| \\tilde {\\mathbf {z}} _ {\\mathbf {i}} \\| \\cdot \\| f _ {i n s} \\left(\\mathbf {x} _ {i n s} ^ {t e s t}\\right) \\|}, \\tag {11}", + "image_path": "de79bbea1cc00da31017895230b6f484f61a22913f6cf8884f1604d17f287613.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "2In Section 5.4, we implement more FL algorithms to demonstrate the compatibility of our method." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1329" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 70, + 542, + 258 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 542, + 258 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 542, + 258 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 542, + 258 + ], + "type": "table", + "html": "
Dataset settingCapability-related (4 task)Task-related (8 task)
Partitionβ = 0.5β = 1.0β = 5.0β = 0.5β = 1.0β = 5.0
MethodsLastAvgLastAvgLastAvgLastAvgLastAvgLastAvg
Zero-shot30.57-30.57-30.57-29.08-29.08-29.08-
Individual61.75-59.62-60.27-64.21-63.85-64.07-
Centralized MTL63.83-63.83-63.83-66.60-66.60-66.60-
Finetune48.6860.0549.6759.0150.4058.0846.2468.9347.2068.7948.0069.97
EWC49.1260.1349.4659.2849.8958.7647.5169.0447.9269.2248.1570.27
LwF48.8760.3250.0259.8850.1559.0647.8969.5747.6269.1448.2670.30
L2P48.2259.7949.5659.3449.6259.2547.3069.3148.0869.6548.4270.16
O-LoRA51.6560.1949.1357.9350.2958.1852.8771.5449.8770.2647.7670.84
M-LoRA49.0460.0850.3960.7650.5658.0650.6871.9448.5371.5848.3871.21
MoELoRA49.6961.0050.9060.1850.4359.1149.2370.9649.0270.6548.8271.08
DISCO53.7362.0055.4762.0755.0660.5357.6974.0356.2273.0355.5872.64
", + "image_path": "95ed1ae71b6a1d5015452e1f6164b6db905751ff7f82d79319861fcedb6627bd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 87, + 265, + 522, + 277 + ], + "lines": [ + { + "bbox": [ + 87, + 265, + 522, + 277 + ], + "spans": [ + { + "bbox": [ + 87, + 265, + 522, + 277 + ], + "type": "text", + "content": "Table 1. Last and Avg performance of different methods on Hom-FCIT setting. The best performance is shown in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 72, + 281, + 205, + 398 + ], + "blocks": [ + { + "bbox": [ + 72, + 281, + 205, + 398 + ], + "lines": [ + { + "bbox": [ + 72, + 281, + 205, + 398 + ], + "spans": [ + { + "bbox": [ + 72, + 281, + 205, + 398 + ], + "type": "image", + "image_path": "b2202c5f0b067b022c91b6c9c651798aee492216ef68e6307e0554fb32338fa2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 402, + 555, + 425 + ], + "lines": [ + { + "bbox": [ + 55, + 402, + 555, + 425 + ], + "spans": [ + { + "bbox": [ + 55, + 402, + 555, + 425 + ], + "type": "text", + "content": "Figure 4. Performance curves of different methods in Hom-FCIT across seen tasks under varying data heterogeneity. We plot the average performance across all seen tasks at each stage." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 207, + 281, + 339, + 398 + ], + "blocks": [ + { + "bbox": [ + 207, + 281, + 339, + 398 + ], + "lines": [ + { + "bbox": [ + 207, + 281, + 339, + 398 + ], + "spans": [ + { + "bbox": [ + 207, + 281, + 339, + 398 + ], + "type": "image", + "image_path": "a78c15ea7723b5e842794c4d1f000cf743c20b5df6ebd16a2d30b2ea1abd4895.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 342, + 281, + 536, + 398 + ], + "blocks": [ + { + "bbox": [ + 342, + 281, + 536, + 398 + ], + "lines": [ + { + "bbox": [ + 342, + 281, + 536, + 398 + ], + "spans": [ + { + "bbox": [ + 342, + 281, + 536, + 398 + ], + "type": "image", + "image_path": "549f69b58d50b25e8cd8ea9bede745cac61ec68a5e506a528ae30003b5478c11.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 436, + 295, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 295, + 483 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 436, + 295, + 483 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 55, + 436, + 295, + 483 + ], + "type": "text", + "content": " denotes the similarity between the " + }, + { + "bbox": [ + 55, + 436, + 295, + 483 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 436, + 295, + 483 + ], + "type": "text", + "content": "-th global identity token and the feature of the test instruction " + }, + { + "bbox": [ + 55, + 436, + 295, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{ins}^{test}" + }, + { + "bbox": [ + 55, + 436, + 295, + 483 + ], + "type": "text", + "content": ". The activation factor " + }, + { + "bbox": [ + 55, + 436, + 295, + 483 + ], + "type": "inline_equation", + "content": "\\alpha_i" + }, + { + "bbox": [ + 55, + 436, + 295, + 483 + ], + "type": "text", + "content": " is then computed by applying softmax normalization over all similarities:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 125, + 486, + 295, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 486, + 295, + 515 + ], + "spans": [ + { + "bbox": [ + 125, + 486, + 295, + 515 + ], + "type": "interline_equation", + "content": "\\alpha_ {i} = \\frac {\\exp \\left(s _ {i} / \\varepsilon\\right)}{\\sum_ {j = 1} ^ {T} \\exp \\left(s _ {j} / \\varepsilon\\right)}, \\tag {12}", + "image_path": "426789007b9431393a6b49064f8d94d00705a709dc93a08775b1cdda07b2abc9.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 519, + 211, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 211, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 211, + 529 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 519, + 211, + 529 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 55, + 519, + 211, + 529 + ], + "type": "text", + "content": " is the temperature coefficient." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 530, + 295, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 530, + 295, + 543 + ], + "spans": [ + { + "bbox": [ + 66, + 530, + 295, + 543 + ], + "type": "text", + "content": "Overall, our proposed SSA can be formulated as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 64, + 547, + 295, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 547, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 64, + 547, + 295, + 628 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\Delta \\mathbf {W} = \\bar {\\mathbf {B}} \\bar {\\mathcal {W}} \\bar {\\mathbf {A}} \\\\ = \\bar {\\mathbf {B}} \\left[ \\begin{array}{c c c c} \\alpha_ {1} \\cdot \\mathbf {I} _ {r \\times r} & \\mathbf {0} _ {r \\times r} & \\dots & \\mathbf {0} _ {r \\times r} \\\\ \\mathbf {0} _ {r \\times r} & \\alpha_ {2} \\cdot \\mathbf {I} _ {r \\times r} & \\dots & \\mathbf {0} _ {r \\times r} \\\\ \\vdots & \\vdots & \\ddots & \\vdots \\\\ \\mathbf {0} _ {r \\times r} & \\mathbf {0} _ {r \\times r} & \\dots & \\alpha_ {T} \\cdot \\mathbf {I} _ {r \\times r} \\end{array} \\right] \\bar {\\mathbf {A}}, \\tag {13} \\\\ \\end{array}", + "image_path": "20851ddec672d4b2d0a7ca35f179f1a9d6f8b91b4fbe9c9aeef22f66266c7a2f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{B}}\\in \\mathbb{R}^{d\\times (r\\cdot T)}" + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "content": " represents the concatenation of all " + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{B}_1,\\dots ,\\mathbf{B}_T\\}" + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "content": " in the dynamic cache along the low-rank dimension " + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{A}}\\in \\mathbb{R}^{(r\\cdot T)\\times k}" + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "content": " follows the same structure. The activation factors " + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\{\\alpha_{1},\\dots ,\\alpha_{T}\\}" + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "content": ", computed via Eq. (11) and Eq. (12), control each subspace's output by amplifying matched subspaces while suppressing irrelevant ones. The schematic of SSA is provided in Figure 3." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 435, + 395, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 395, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 395, + 449 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 455, + 430, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 430, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 430, + 468 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 472, + 555, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 555, + 581 + ], + "type": "text", + "content": "Datasets. The dataset composition is detailed in Section 3.2. For the experimental setup, we implement two types of dataset-level settings: Capability-related and Task-related, based on the two client-level realistic scenarios, Hom-FCIT and Het-FCIT, respectively. At each stage, the own data of local clients is partitioned according to the Dirichlet distribution [27], with three partitions " + }, + { + "bbox": [ + 313, + 472, + 555, + 581 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 472, + 555, + 581 + ], + "type": "text", + "content": " for each setting to model different levels of data heterogeneity. More details are provided in Appendix A." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 581, + 555, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 555, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 555, + 688 + ], + "type": "text", + "content": "Baselines & Evaluation Metrics. We compare our method with continual learning methods such as LwF [29], EWC [24], L2P [51], O-LoRA [50], MoELoRA [6], and also with the LoRA merging method from the federated continual learning approach PILoRA [12], referred to as M-LoRA. We follow a rehearsal-free continual learning setting [62], where only the data for the current task is available. All comparison methods are carefully calibrated to ensure the fairness of evaluations." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "For the evaluation metrics, we report the standard metrics to measure the model performance: Last refers to the" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1330" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 70, + 542, + 258 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 542, + 258 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 542, + 258 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 542, + 258 + ], + "type": "table", + "html": "
Dataset settingCapability-related (4 task)Task-related (8 task)
Partitionβ = 0.5β = 1.0β = 5.0β = 0.5β = 1.0β = 5.0
MethodsLastAvgLastAvgLastAvgLastAvgLastAvgLastAvg
Zero-shot30.57-30.57-30.57-29.08-29.08-29.08-
Individual61.75-59.62-60.27-64.21-63.85-64.07-
Centralized MTL63.83-63.83-63.83-66.60-66.60-66.60-
Finetune55.6555.8256.3456.8556.7457.1558.0453.0557.9654.2258.1754.87
EWC55.2154.5855.1355.7655.6956.0258.7453.7657.9654.1857.4454.30
LwF55.9256.1856.4256.8056.8157.2658.8253.7758.0154.3458.2254.78
L2P56.1056.6356.7657.0256.9557.3658.8453.8058.7354.6758.3954.66
O-LoRA58.2458.5858.3258.6558.6058.9459.6154.5559.7455.2059.5154.97
M-LoRA57.7658.0257.6557.8957.8058.1759.7654.1158.8254.0359.3554.89
MoELoRA57.6857.9557.7758.0058.1558.4459.0254.2559.1454.6958.8654.50
DISCO59.4060.0159.9459.9159.7160.1662.0859.6463.2561.9962.7860.87
", + "image_path": "10a43a7c910c7ca056c65bbb2e99fcfa58e35b19c1fcd35e6eeb26aea19591de.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 88, + 265, + 518, + 277 + ], + "lines": [ + { + "bbox": [ + 88, + 265, + 518, + 277 + ], + "spans": [ + { + "bbox": [ + 88, + 265, + 518, + 277 + ], + "type": "text", + "content": "Table 2. Last and Avg performance of different methods on Het-FCIT setting. The best performance is shown in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 72, + 282, + 205, + 397 + ], + "blocks": [ + { + "bbox": [ + 72, + 282, + 205, + 397 + ], + "lines": [ + { + "bbox": [ + 72, + 282, + 205, + 397 + ], + "spans": [ + { + "bbox": [ + 72, + 282, + 205, + 397 + ], + "type": "image", + "image_path": "04b024c7588837943e90cbecd61b2dbb677a73bdbb08184309f889f7404c0219.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 402, + 555, + 425 + ], + "lines": [ + { + "bbox": [ + 55, + 402, + 555, + 425 + ], + "spans": [ + { + "bbox": [ + 55, + 402, + 555, + 425 + ], + "type": "text", + "content": "Figure 5. Performance curve of different methods in the Het-FCIT under varying degrees of data heterogeneity. We plot the average performance across all tasks at each stage. The black dashed line marks the stage where the model has encountered all tasks." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 207, + 282, + 339, + 397 + ], + "blocks": [ + { + "bbox": [ + 207, + 282, + 339, + 397 + ], + "lines": [ + { + "bbox": [ + 207, + 282, + 339, + 397 + ], + "spans": [ + { + "bbox": [ + 207, + 282, + 339, + 397 + ], + "type": "image", + "image_path": "2c2bb57e7731e238b1bf640d4bae8e55cb25d5709ceb01b1ee351eea830f7a09.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 342, + 282, + 536, + 398 + ], + "blocks": [ + { + "bbox": [ + 342, + 282, + 536, + 398 + ], + "lines": [ + { + "bbox": [ + 342, + 282, + 536, + 398 + ], + "spans": [ + { + "bbox": [ + 342, + 282, + 536, + 398 + ], + "type": "image", + "image_path": "1772929cc40a3675df940e8e9ae9519aa576e5169fb6d15fcd7dee50dd9816c1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 436, + 295, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 436, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 54, + 436, + 295, + 483 + ], + "type": "text", + "content": "average result across all learned tasks after the completion of learning the final task. " + }, + { + "bbox": [ + 54, + 436, + 295, + 483 + ], + "type": "inline_equation", + "content": "Avg" + }, + { + "bbox": [ + 54, + 436, + 295, + 483 + ], + "type": "text", + "content": " is based on Last, which tracks the performance of the learned tasks at each stage of the continual learning process and reports the average results." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 484, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 484, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 54, + 484, + 295, + 590 + ], + "type": "text", + "content": "Implementation Details. We choose LLaVA-1.5-7b [32] as the base LMM due to its structural simplicity and adopt its LoRA fine-tuning strategy, training only the LoRAs and multimodal projector. The LoRA rank " + }, + { + "bbox": [ + 54, + 484, + 295, + 590 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 54, + 484, + 295, + 590 + ], + "type": "text", + "content": " at each stage is set to 8, with a learning rate of 2e-4, and it is embedded solely in the FFN layers of each block. The learning rate for the projector is set to 2e-5. For identity token extraction, we use the frozen CLIP text encoder [41]. The threshold " + }, + { + "bbox": [ + 54, + 484, + 295, + 590 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 54, + 484, + 295, + 590 + ], + "type": "text", + "content": " and temperature coefficient " + }, + { + "bbox": [ + 54, + 484, + 295, + 590 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 54, + 484, + 295, + 590 + ], + "type": "text", + "content": " are set to 0.9 and 0.05, respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 592, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 592, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 54, + 592, + 295, + 664 + ], + "type": "text", + "content": "For the setting of FCIT, we set the epoch to 1 and communication rounds to 10 in each stage, and the Dirichlet distribution coefficients " + }, + { + "bbox": [ + 54, + 592, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 54, + 592, + 295, + 664 + ], + "type": "text", + "content": " are set to " + }, + { + "bbox": [ + 54, + 592, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\{0.5, 1.0, 5.0\\}" + }, + { + "bbox": [ + 54, + 592, + 295, + 664 + ], + "type": "text", + "content": ". In each round, the global server randomly selects 5 clients from a pool of 50 to participate in the training, and we use FedAvg [38] as the base FL aggregation algorithm." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 672, + 140, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 140, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 140, + 683 + ], + "type": "text", + "content": "5.2. Main Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Results are shown in Table 1, Table 2, Figure 4, and Figure 5. Under Hom-FCIT setting, our proposed DISCO" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 436, + 555, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 555, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 555, + 519 + ], + "type": "text", + "content": "achieves the best performance in both capability-related and task-related dataset settings, surpassing the best baseline by an average of " + }, + { + "bbox": [ + 313, + 436, + 555, + 519 + ], + "type": "inline_equation", + "content": "4.84\\%" + }, + { + "bbox": [ + 313, + 436, + 555, + 519 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 436, + 555, + 519 + ], + "type": "inline_equation", + "content": "1.43\\%" + }, + { + "bbox": [ + 313, + 436, + 555, + 519 + ], + "type": "text", + "content": " in the Last and Avg metrics, respectively. As illustrated in Figure 4, DISCO effectively mitigates the forgetting of previous tasks while learning new ones, outperforming other methods in the challenging long-stage continual learning setting." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 521, + 556, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 652 + ], + "type": "text", + "content": "In Het-FCIT setting, we plot the upward curve of the average performance across all tasks in Figure 5. As shown in the figure. Before completing all tasks (Left of the black dotted line), our method maintains the fastest rate of improvement, significantly outperforming other methods. In the subsequent phases (Right of the black dashed line), it continues to consolidate previously learned knowledge and further enhance performance, demonstrating strong adaptability in dynamic real-world scenarios. On the Last and Avg metrics, our method outperforms the best comparison method by an average of " + }, + { + "bbox": [ + 313, + 521, + 556, + 652 + ], + "type": "inline_equation", + "content": "2.17\\%" + }, + { + "bbox": [ + 313, + 521, + 556, + 652 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 521, + 556, + 652 + ], + "type": "inline_equation", + "content": "3.62\\%" + }, + { + "bbox": [ + 313, + 521, + 556, + 652 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "Additionally, our method demonstrates superior robustness to varying degrees of data heterogeneity, adapting well to diverse distributions and maintaining strong performance even as heterogeneity increases, highlighting its reliability in dynamic real-world scenarios." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 313, + 757 + ], + "type": "text", + "content": "1331" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 295, + 178 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 295, + 178 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 295, + 178 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 295, + 178 + ], + "type": "table", + "html": "
MethodHom-FCITHet-FCIT
LastΔAvgΔLastΔAvgΔ
(a)Text56.220.073.030.063.250.061.990.0
Image55.63-0.5972.13-0.9063.00-0.2561.80-0.19
Text & Image55.96-0.2672.68-0.3563.02-0.2361.78-0.21
(b)Softmax56.220.073.030.063.250.061.990.0
Concatenate51.74-4.4869.20-3.8360.36-2.8959.88-2.11
Cosine sim52.83-3.3970.07-2.9660.92-2.3360.13-1.86
Argmax55.74-0.4872.07-0.9662.88-0.3761.42-0.57
(c)FFN56.220.073.030.063.250.061.990.0
Attn56.03-0.1972.88-0.1563.01-0.2461.83-0.16
FFN & Attn56.30+0.0872.96-0.0763.18-0.0762.04+0.05
", + "image_path": "de43b4405e7f8063244bd1baf64accc0840ecae63e87a2dd430335ea936f8cf6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 188, + 295, + 232 + ], + "lines": [ + { + "bbox": [ + 55, + 188, + 295, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 188, + 295, + 232 + ], + "type": "text", + "content": "Table 3. Ablation studies on (a) identity token extraction methods; (b) calculation of activation factors in SSA; (c) location of LoRA embedding. All experiments were conducted in the task-related setting with " + }, + { + "bbox": [ + 55, + 188, + 295, + 232 + ], + "type": "inline_equation", + "content": "\\beta = 1.0" + }, + { + "bbox": [ + 55, + 188, + 295, + 232 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 239, + 149, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 149, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 149, + 251 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 256, + 296, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 296, + 375 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 296, + 375 + ], + "type": "text", + "content": "The detailed ablation studies are provided in Table 3. In this paper, we use CLIP's text encoder to extract textual features as identity tokens. This section tests two alternatives: using only visual features from CLIP's visual encoder, and combining visual and textual features. From Table 3a, we observe that both alternatives degrade performance. This is potential because, in visual instruction tuning datasets, task similarities at the image level are higher than at the textual level (e.g. CLEVR-Math and super-CLEVR), making visual features less effective than textual ones." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 376, + 296, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 296, + 532 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 296, + 532 + ], + "type": "text", + "content": "Table 3b shows the ablation studies on how to calculate the activation factors in SSA. Direct concatenation, where the activation factors of all subspaces are set to 1.0, leads to significant performance degradation due to unfiltered outputs. Using cosine similarity (Eq.(11)) to filter the outputs also fails to solve the issue. While selecting the largest similarity (Argmax) can filter out irrelevant information, it risks fully activating a subspace that, though similar, is not directly related to the current task. This can negatively impact the output, as there are inherent similarities between the textual information of different tasks. In contrast, our method normalizes the cosine similarity to better focus on the relevant task output, achieving optimal performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 532, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 296, + 628 + ], + "type": "text", + "content": "We also test the location of LoRA embedding. As can be seen in Table 3c, embedding in the attention layer alone proved less effective than in the FFN layer. Moreover, embedding LoRA in every linear layer did not offer significant improvement and resulted in higher parameter transmission. Therefore, we chose to embed LoRA only in the FFN layer as a balanced solution. We provide more ablation and visualization results in Appendix B." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 635, + 159, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 159, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 159, + 649 + ], + "type": "text", + "content": "5.4. Further Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": "Compatible with other FL algorithms. In this paper, we use the classical FedAvg as the FL algorithm for global server aggregation of local weights. Additionally, we implement other federated learning algorithms, including FedAvgM [17], FedAdam [42], FedAdagrad [42], and" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 335, + 70, + 537, + 167 + ], + "blocks": [ + { + "bbox": [ + 335, + 70, + 537, + 167 + ], + "lines": [ + { + "bbox": [ + 335, + 70, + 537, + 167 + ], + "spans": [ + { + "bbox": [ + 335, + 70, + 537, + 167 + ], + "type": "table", + "html": "
MethodHom-FCITHet-FCIT
LastAvgLastAvg
FedAvg56.2273.0363.2561.99
FedAvgM56.0772.5962.4761.10
FedAdam55.7672.9162.7861.35
FedAdagrad56.0272.8062.5561.41
FedYogi56.1173.1662.8861.67
", + "image_path": "b568094367a773e86eebfc4344d7c3bdd79ca2a0d356a4e57a7a29a79ff8bf7e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 313, + 175, + 554, + 198 + ], + "lines": [ + { + "bbox": [ + 313, + 175, + 554, + 198 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 554, + 198 + ], + "type": "text", + "content": "Table 4. Results of different FL algorithms. All experiments were conducted in the task-related setting with " + }, + { + "bbox": [ + 313, + 175, + 554, + 198 + ], + "type": "inline_equation", + "content": "\\beta = 1.0" + }, + { + "bbox": [ + 313, + 175, + 554, + 198 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 352, + 216, + 527, + 380 + ], + "blocks": [ + { + "bbox": [ + 352, + 216, + 527, + 380 + ], + "lines": [ + { + "bbox": [ + 352, + 216, + 527, + 380 + ], + "spans": [ + { + "bbox": [ + 352, + 216, + 527, + 380 + ], + "type": "image", + "image_path": "fa06c48127ce5152049f999ae4e2d0149a8d8058956cfc623619cb4ac5987f1f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 385, + 545, + 397 + ], + "lines": [ + { + "bbox": [ + 322, + 385, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 322, + 385, + 545, + 397 + ], + "type": "text", + "content": "Figure 6. Visualization of activation factors during inference." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 407, + 554, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 554, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 554, + 431 + ], + "type": "text", + "content": "FedYogi [42], to extend our framework. As shown in Table 4, FedAvg achieves the best average performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 432, + 554, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 554, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 554, + 504 + ], + "type": "text", + "content": "Visualization of activation factors. In Figure 6, we plot the responses of activations to the test inputs of different tasks. As observed, only the activator corresponding to the current task is responsive, effectively activating the relevant subspace, while the others remain largely inhibited. This confirms the effectiveness of our proposed SSA." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 514, + 388, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 388, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 388, + 526 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "We have explored for the first time the integration of federated learning and continual learning for the instruction tuning of LMMs, addressing the real-world challenge of dynamically acquiring new knowledge through distributed training resources and data. Our proposed FCIT benchmark encompasses 2 real-world scenarios, 4 distinct settings, and 12 curated datasets, providing a comprehensive evaluation of different methods. Additionally, we introduce the DISCO framework, which leverages dynamic knowledge organization (DKO) to decompose inter-task conflicts and subspace selective activation (SSA) to assign task-relevant outputs while suppressing irrelevant information. Extensive experiments demonstrate that our approach significantly improves the model's ability to learn new knowledge and handle data heterogeneity in real-world scenarios." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1332" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 212 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 212 + ], + "type": "text", + "content": "This work was supported by the National Science and Technology Major Project (2022ZD0116500), National Natural Science Foundation of China (62222609, 62320106010), CAS Project for Young Scientists in Basic Research (YSBR-083), Strategic Priority Research Program of Chinese Academy of Sciences under Grant (XDA0480200), Major Science and Technology Plan Project on the Future Industry Fields of Xiamen City (3502Z20241027), Unveiling and Leading Projects of Xiamen (3502Z20241011) and the InnoHK program." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 220, + 115, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 115, + 232 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 115, + 232 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 239, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 239, + 296, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 239, + 296, + 295 + ], + "spans": [ + { + "bbox": [ + 61, + 239, + 296, + 295 + ], + "type": "text", + "content": "[1] Sara Babakniya, Zalan Fabian, Chaoyang He, Mahdi Soltanolkotabi, and Salman Avestimehr. A data-free approach to mitigate catastrophic forgetting in federated class incremental learning for vision tasks. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 295, + 296, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 295, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 61, + 295, + 296, + 350 + ], + "type": "text", + "content": "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 350, + 296, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 350, + 296, + 393 + ], + "spans": [ + { + "bbox": [ + 61, + 350, + 296, + 393 + ], + "type": "text", + "content": "[3] Abhijit Bendale and Terrance Boult. Towards open world recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1893-1902, 2015. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 394, + 296, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 394, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 62, + 394, + 296, + 449 + ], + "type": "text", + "content": "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 449, + 296, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 449, + 296, + 504 + ], + "spans": [ + { + "bbox": [ + 62, + 449, + 296, + 504 + ], + "type": "text", + "content": "[5] Meng Cao, Yuyang Liu, Yingfei Liu, Tiancai Wang, Jiahua Dong, Henghui Ding, Xiangyu Zhang, Ian Reid, and Xiaodan Liang. Continual llava: Continual instruction tuning in large vision-language models. arXiv preprint arXiv:2411.02564, 2024. 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 505, + 296, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 505, + 296, + 548 + ], + "spans": [ + { + "bbox": [ + 62, + 505, + 296, + 548 + ], + "type": "text", + "content": "[6] Cheng Chen, Junchen Zhu, Xu Luo, Hengtao Shen, Lianli Gao, and Jingkuan Song. Coin: A benchmark of continual instruction tuning for multimodel large language model. arXiv preprint arXiv:2403.08350, 2024. 3, 4, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 548, + 296, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 548, + 296, + 603 + ], + "spans": [ + { + "bbox": [ + 62, + 548, + 296, + 603 + ], + "type": "text", + "content": "[7] Haokun Chen, Yao Zhang, Denis Krompass, Jindong Gu, and Volker Tresp. Feddat: An approach for foundation model finetuning in multi-modal heterogeneous federated learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11285-11293, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 604, + 296, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 604, + 296, + 658 + ], + "spans": [ + { + "bbox": [ + 62, + 604, + 296, + 658 + ], + "type": "text", + "content": "[8] Adam Dahlgren Lindstrom and Savitha Sam Abraham. Clevr-math: A dataset for compositional language, visual and mathematical reasoning. In 16th International Workshop on Neural-Symbolic Learning and Reasoning, NeSy 2022, Windsor, UK, september 28-30, 2022. CEUR-WS, 2022. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 658, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 658, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 658, + 296, + 713 + ], + "type": "text", + "content": "[9] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500, 2023. 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 715 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 117 + ], + "type": "text", + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 118, + 554, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 171 + ], + "type": "text", + "content": "[11] Jiahua Dong, Lixu Wang, Zhen Fang, Gan Sun, Shichao Xu, Xiao Wang, and Qi Zhu. Federated class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10164-10173, 2022. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 172, + 554, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 172, + 554, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 172, + 554, + 228 + ], + "type": "text", + "content": "[12] Haiyang Guo, Fei Zhu, Wenzhuo Liu, Xu-Yao Zhang, and Cheng-Lin Liu. Pilora: Prototype guided incremental lora for federated class-incremental learning. In European Conference on Computer Vision, pages 141-159. Springer, 2024. 1, 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 228, + 554, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 554, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 554, + 271 + ], + "type": "text", + "content": "[13] Haiyang Guo, Fei Zhu, Fanhu Zeng, Bing Liu, and XuYao Zhang. Desire: Dynamic knowledge consolidation for rehearsal-free continual learning. arXiv preprint arXiv:2411.19154, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 272, + 554, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 554, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 554, + 326 + ], + "type": "text", + "content": "[14] Haiyang Guo, Fanhu Zeng, Ziwei Xiang, Fei Zhu, Da-Han Wang, Xu-Yao Zhang, and Cheng-Lin Liu. Hidellava: Hierarchical decoupling for continual instruction tuning of multimodal large language model. arXiv preprint arXiv:2503.12941, 2025. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 327, + 554, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 554, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 554, + 381 + ], + "type": "text", + "content": "[15] Haiyang Guo, Fanhu Zeng, Fei Zhu, Jiayi Wang, Xukai Wang, Jingang Zhou, Hongbo Zhao, Wenzhuo Liu, Shijie Ma, Xu-Yao Zhang, et al. A comprehensive survey on continual learning in generative models. arXiv preprint arXiv:2506.13045, 2025. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 383, + 554, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 383, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 317, + 383, + 554, + 437 + ], + "type": "text", + "content": "[16] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, Dawn Song, Jacob Steinhardt, and Justin Gilmer. The many faces of robustness: A critical analysis of out-of-distribution generalization. ICCV, 2021. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 437, + 554, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 554, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 554, + 480 + ], + "type": "text", + "content": "[17] Tzu-Ming Harry Hsu, Hang Qi, and Matthew Brown. Measuring the effects of non-identical data distribution for federated visual classification. arXiv preprint arXiv:1909.06335, 2019.8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 481, + 554, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 481, + 554, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 481, + 554, + 525 + ], + "type": "text", + "content": "[18] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 526, + 554, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 554, + 580 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 554, + 580 + ], + "type": "text", + "content": "[19] Wenke Huang, Mang Ye, Zekun Shi, He Li, and Bo Du. Rethinking federated learning with domain shift: A prototype view. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16312-16322. IEEE, 2023. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 581, + 554, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 581, + 554, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 581, + 554, + 635 + ], + "type": "text", + "content": "[20] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 1, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 636, + 554, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 554, + 689 + ], + "type": "text", + "content": "[21] Kushal Kafle, Brian Price, Scott Cohen, and Christopher Kanan. Dvqa: Understanding data visualizations via question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5648-5656, 2018. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 690, + 554, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 690, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 316, + 690, + 554, + 715 + ], + "type": "text", + "content": "[22] Samira Ebrahimi Kahou, Vincent Michalski, Adam Atkinson, Ákos Kádár, Adam Trischler, and Yoshua Bengio. Fig-" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1333" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "text", + "content": "ureqa: An annotated figure dataset for visual reasoning. arXiv preprint arXiv:1710.07300, 2017. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 138 + ], + "type": "text", + "content": "[23] Gyuhak Kim, Changnan Xiao, Tatsuya Konishi, and Bing Liu. Learnability and algorithm for continual learning. In International Conference on Machine Learning, pages 16877-16896. PMLR, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 205 + ], + "type": "text", + "content": "[24] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 294, + 249 + ], + "type": "text", + "content": "[25] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 251, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 294, + 293 + ], + "type": "text", + "content": "[26] Lei Li, Yuqi Wang, Runxin Xu, Peiyi Wang, Xiachong Feng, Lingpeng Kong, and Qi Liu. Multimodal arxiv: A dataset for improving scientific comprehension of large vision-language models. arXiv preprint arXiv:2403.00231, 2024. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 294, + 294, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 294, + 338 + ], + "type": "text", + "content": "[27] Qinbin Li, Yiqun Diao, Quan Chen, and Bingsheng He. Federated learning on non-iid data silos: An experimental study. In 2022 IEEE 38th international conference on data engineering (ICDE), pages 965-978. IEEE, 2022. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 338, + 294, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 380 + ], + "type": "text", + "content": "[28] Tian Li, Anit Kumar Sahu, Ameet Talwalkar, and Virginia Smith. Federated learning: Challenges, methods, and future directions. IEEE signal processing magazine, 37(3):50-60, 2020. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 383, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 415 + ], + "type": "text", + "content": "[29] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 416, + 295, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 416, + 295, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 416, + 295, + 480 + ], + "type": "text", + "content": "[30] Zhuowan Li, Xingrui Wang, Elias Stengel-Eskin, Adam Kortylewski, Wufei Ma, Benjamin Van Durme, and Alan L Yuille. Super-clevr: A virtual benchmark to diagnose domain robustness in visual reasoning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14963-14973, 2023. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 482, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 482, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 294, + 525 + ], + "type": "text", + "content": "[31] Yi-Ming Lin, Yuan Gao, Mao-Guo Gong, Si-Jia Zhang, Yuan-Qiao Zhang, and Zhi-Yuan Li. Federated learning on multimodal data: A comprehensive survey. Machine Intelligence Research, 20(4):539-553, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 526, + 294, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 294, + 558 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 294, + 558 + ], + "type": "text", + "content": "[32] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 3, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 559, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 591 + ], + "type": "text", + "content": "[33] Wenzhuo Liu, Fei Zhu, Haiyang Guo, Longhui Wei, and Cheng-Lin Liu. Llava-c: Continual improved visual instruction tuning. arXiv preprint arXiv:2506.08666, 2025.1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 593, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 593, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 593, + 294, + 646 + ], + "type": "text", + "content": "[34] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, et al. Deepseek-vl: towards real-world vision-language understanding. arXiv preprint arXiv:2403.05525, 2024. 1, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "text", + "content": "[35] Pan Lu, Liang Qiu, Jiaqi Chen, Tony Xia, Yizhou Zhao, Wei Zhang, Zhou Yu, Xiaodan Liang, and Song-Chun Zhu. Iconqa: A new benchmark for abstract diagram understanding and visual language reasoning. In The 35th Conference on Neural Information Processing Systems (NeurIPS) Track on Datasets and Benchmarks, 2021. 4" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 126 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 126 + ], + "type": "text", + "content": "[36] Pan Lu, Liang Qiu, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, Tanmay Rajpurohit, Peter Clark, and Ashwin Kalyan. Dynamic prompt learning via policy gradient for semi-structured mathematical reasoning. arXiv preprint arXiv:2209.14610, 2022. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 130, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 184 + ], + "type": "text", + "content": "[37] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 186, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 239 + ], + "type": "text", + "content": "[38] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 1, 5, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 243, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 296 + ], + "type": "text", + "content": "[39] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering by reading text in images. In 2019 international conference on document analysis and recognition (ICDAR), pages 947-952. IEEE, 2019. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 298, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 553, + 352 + ], + "type": "text", + "content": "[40] Bryan A. Plummer, Liwei Wang, Christopher M. Cervantes, Juan C. Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. IJCV, 123 (1):74–93, 2017. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 354, + 553, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 418 + ], + "type": "text", + "content": "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 5, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 422, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 464 + ], + "type": "text", + "content": "[42] Sashank Reddi, Zachary Charles, Manzil Zaheer, Zachary Garrett, Keith Rush, Jakub Konečny, Sanjiv Kumar, and H Brendan McMahan. Adaptive federated optimization. arXiv preprint arXiv:2003.00295, 2020. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 467, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 467, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 467, + 553, + 520 + ], + "type": "text", + "content": "[43] Lorenzo Sani, Alex Iacob, Zeyu Cao, Bill Marino, Yan Gao, Tomas Paulik, Wanru Zhao, William F Shen, Preslav Aleksandrov, Xinchi Qiu, et al. The future of large language model pre-training is federated. arXiv preprint arXiv:2405.10853, 2024. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 523, + 553, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 577 + ], + "type": "text", + "content": "[44] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge. In European conference on computer vision, pages 146–162. Springer, 2022. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 579, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 553, + 623 + ], + "type": "text", + "content": "[45] Donald Shenaj, Marco Toldo, Alberto Rigon, and Pietro Zanuttigh. Asynchronous federated continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5055-5063, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "text", + "content": "[46] Tianxiang Sun, Xiaotian Zhang, Zhengfu He, Peng Li, Qinyuan Cheng, Xiangyang Liu, Hang Yan, Yunfan Shao, Qiong Tang, Shiduo Zhang, et al. Moss: An open conversational large language model. Machine Intelligence Research, 21(5):888-905, 2024. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[47] Yue Tan, Guodong Long, Lu Liu, Tianyi Zhou, Qinghua Lu, Jing Jiang, and Chengqi Zhang. Fedproto: Federated prototype learning across heterogeneous clients. In Proceedings of" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1334" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "text", + "content": "the AAAI conference on artificial intelligence, pages 8432-8440, 2022. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "text", + "content": "[48] Yue Tan, Guodong Long, Jie Ma, Lu Liu, Tianyi Zhou, and Jing Jiang. Federated learning from pre-trained models: A contrastive learning approach. Advances in neural information processing systems, 35:19332-19344, 2022. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 196 + ], + "type": "text", + "content": "[49] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 241 + ], + "type": "text", + "content": "[50] Xiao Wang, Tianze Chen, Qiming Ge, Han Xia, Rong Bao, Rui Zheng, Qi Zhang, Tao Gui, and Xuanjing Huang. Orthogonal subspace learning for language model continual learning. arXiv preprint arXiv:2310.14152, 2023. 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 243, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 294, + 308 + ], + "type": "text", + "content": "[51] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 139-149, 2022. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 309, + 294, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 353 + ], + "type": "text", + "content": "[52] Ziyao Wang, Zheyu Shen, Yexiao He, Guoheng Sun, Hongyi Wang, Lingjuan Lyu, and Ang Li. Flora: Federated fin-tuning large language models with heterogeneous low-rank adaptations. arXiv preprint arXiv:2409.05976, 2024. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 387 + ], + "type": "text", + "content": "[53] Taiqiang Wu, Jiahao Wang, Zhe Zhao, and Ngai Wong. Mixture-of-subspaces in low-rank adaptation. arXiv preprint arXiv:2406.11909, 2024. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 294, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 432 + ], + "type": "text", + "content": "[54] Xin Yang, Hao Yu, Xin Gao, Hao Wang, Junbo Zhang, and Tianrui Li. Federated continual learning via knowledge fusion: A survey. IEEE Transactions on Knowledge and Data Engineering, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 434, + 294, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 499 + ], + "type": "text", + "content": "[55] Rui Ye, Wenhao Wang, Jingyi Chai, Dihan Li, Zexi Li, Yinda Xu, Yaxin Du, Yanfeng Wang, and Siheng Chen. Openfedllm: Training large language models on decentralized private data via federated learning. In Proceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining, pages 6137-6147, 2024. 1, 3, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 501, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 544 + ], + "type": "text", + "content": "[56] Fanhu Zeng, Fei Zhu, Haiyang Guo, Xu-Yao Zhang, and Cheng-Lin Liu. Modalprompt: Dual-modality guided prompt for continual learning of large multimodal models. arXiv preprint arXiv:2410.05849, 2024. 1, 3, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 545, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 589 + ], + "type": "text", + "content": "[57] Jie Zhang, Chen Chen, Weiming Zhuang, and Lingjuan Lyu. Target: Federated class-continual learning via exemplar-free distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4782–4793, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 590, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 656 + ], + "type": "text", + "content": "[58] Jianyi Zhang, Saeed Vahidian, Martin Kuo, Chunyuan Li, Ruiyi Zhang, Tong Yu, Guoyin Wang, and Yiran Chen. Towards building the federatedgpt: Federated instruction tuning. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6915-6919. IEEE, 2024. 1, 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "type": "text", + "content": "[59] Xu-Yao Zhang, Cheng-Lin Liu, and Ching Y Suen. Towards robust pattern recognition: A review. Proceedings of the IEEE, 108(6):894-922, 2020. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "text", + "content": "[60] Hongbo Zhao, Fei Zhu, Rundong Wang, Gaofeng Meng, and Zhaoxiang Zhang. Mllm-cl: Continual learning" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 240 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "for multimodal large language models. arXiv preprint arXiv:2506.05453, 2025.1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[61] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "type": "text", + "content": "[62] Fei Zhu, Xu-Yao Zhang, Chuang Wang, Fei Yin, and ChengLin Liu. Prototype augmentation and self-supervision for incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5871-5880, 2021. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 197, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 240 + ], + "type": "text", + "content": "[63] Fei Zhu, Shijie Ma, Zhen Cheng, Xu-Yao Zhang, Zhaoxiang Zhang, and Cheng-Lin Liu. Open-world machine learning: A review and new outlooks. arXiv preprint arXiv:2403.01759, 2024. 1" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1335" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Federated Continuous Category Discovery and Learning/7332a50f-74a4-4833-9616-ad03014bd631_content_list.json b/2025/Federated Continuous Category Discovery and Learning/7332a50f-74a4-4833-9616-ad03014bd631_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..416a6bb3ebf450785dbda2f6c8f1a075e6b0fb8f --- /dev/null +++ b/2025/Federated Continuous Category Discovery and Learning/7332a50f-74a4-4833-9616-ad03014bd631_content_list.json @@ -0,0 +1,1584 @@ +[ + { + "type": "text", + "text": "Federated Continuous Category Discovery and Learning", + "text_level": 1, + "bbox": [ + 212, + 130, + 785, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lixu Wang $^{1*}$ , Chenxi Liu $^{2*}$ , Junfeng Guo $^{2}$ , Qingqing Ye $^{3}$ , Heng Huang $^{2}$ , Haibo Hu $^{3}$ , Wei Dong $^{1\\dagger}$ , Nanyang Technological University. $^{2}$ University of Maryland. $^{3}$ The Hong Kong Polytechnic University.", + "bbox": [ + 101, + 179, + 916, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lixu.wang@ntu.edu.sg, {cxliu539, gjf2023}@umd.edu, qqing.ye@polyu.edu.hk, heng@umd.edu", + "bbox": [ + 138, + 219, + 879, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "haibo.hu@polyu.edu.hk, wei_dong@ntu.edu.sg", + "bbox": [ + 326, + 237, + 692, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Federated Learning (FL) studies often assume a static data distribution, whereas real-world scenarios involve dynamic changes. To address this gap, we study Federated Continuous Category Discovery and Learning $(FC^2 DL)$ , an essential yet underexplored problem that enables FL models to evolve continuously by discovering and learning novel data categories. The key challenge in $FC^2 DL$ lies in merging and aligning categories discovered and learned by different clients, all while maintaining privacy. To tackle this, we propose the Global Prototype Alignment (GPA) framework. GPA first estimates the number of categories and constructs global prototypes by locating high-density regions in the representation space through bi-level clustering. To mitigate pseudo-label noise, GPA then employs a semantic-weighted loss to capture correlations between global prototypes and the novel data. This semantic weighting strategy is also used for contrastive loss, facilitating unsupervised novel-category learning. Besides, GPA incorporates a mixup-based mechanism for both data and models, effectively mitigating interference between known and novel categories while alleviating forgetting. Extensive experiments across multiple datasets demonstrate GPA's superiority over state-of-the-art baseline approaches. Notably, GPA achieves absolute gains of $5.7\\%$ to $13.1\\%$ in novel category accuracy while preserving known category performance. Furthermore, GPA is highly adaptable, equipping various mainstream FL algorithms with category discovery and learning capabilities, underscoring its potential for real-world deployment.", + "bbox": [ + 86, + 319, + 485, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 787, + 220, + 804 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Federated Learning [37] (FL) enables distributed devices to collaboratively train models while preserving data privacy, making it a compelling solution for privacy-sensitive applications. However, most existing FL studies assume a static", + "bbox": [ + 89, + 813, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and well-characterized data distribution [4, 13, 20, 30]. In reality, data is inherently dynamic, continuously evolving as new information emerges [41, 54, 58]. This evolving nature presents significant challenges, particularly in enabling FL systems to adapt and remain effective in the presence of novel patterns and functionalities. To address this, FL must be equipped with mechanisms to identify and incorporate emerging functionalities, ensuring adaptability to evolving data distributions. This capability is particularly critical in domains such as healthcare, finance, transportation, and the Internet of Things. For example, FL-powered diagnostic models in healthcare must continuously adapt to the emergence of new diseases and rare conditions to maintain their effectiveness [43, 55].", + "bbox": [ + 511, + 287, + 906, + 500 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To enable FL systems to adapt to continuously emerging novel functionalities, we study Federated Continuous Category Discovery and Learning $(\\mathsf{FC}^2\\mathsf{DL})$ . We consider an FL setting where multiple participants initially perform supervised training on a predefined category space, with non-independent and non-identically distributed (non-IID) data [37]. Once trained, the FL model is expected to generalize well to known categories. At this stage, the FL administrator typically stops the training to conserve resources and deploys the latest global model for inference [37]. After deployment, users provide unlabeled data and obtain model predictions. In real-world scenarios, participants frequently encounter data from previously unseen novel categories, which also follow a non-IID distribution. In this case, these novel data samples are often misclassified as existing known categories. $\\mathsf{FC}^2\\mathsf{DL}$ aims to enable FL systems to discover and learn novel categories continuously, ensuring adaptability to evolving data distributions. At the same time, $\\mathsf{FC}^2\\mathsf{DL}$ seeks to maintain strong performance on previously learned categories, striking a balance between new knowledge learning and retention of prior knowledge.", + "bbox": [ + 511, + 503, + 908, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A straightforward approach to tackling $\\mathbf{FC}^2\\mathbf{DL}$ is to integrate FL with Novel Category Discovery and Learning (NCDL) [32, 44, 62]. However, this approach presents several fundamental challenges. First, effective clustering of novel data requires prior knowledge of the number of", + "bbox": [ + 511, + 824, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contributions. †Corresponding author.", + "bbox": [ + 107, + 886, + 351, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2429", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "novel categories. Conventional NCDL assumes this information is available, as model trainers typically have access to the training data and the specific application domain [32, 44, 46]. However, in FL, this assumption does not hold: the FL administrator lacks direct data access, and participants may not have the expertise to estimate task details [10, 37]. Second, in a non-IID setting, even if participants conduct NCDL locally, the discovered novel categories vary across clients, making it difficult to merge their label spaces into a unified one. Finally, most existing NCDL methods often rely on labeled data from known categories [7, 16, 49], which is challenging to obtain in practical FL systems. Many participants have limited storage, and newly joined clients may not have any labeled data [10].", + "bbox": [ + 89, + 90, + 480, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome these challenges, we propose Global Prototype Alignment (GPA), a novel framework for $\\mathbf{FC}^2\\mathbf{DL}$ , as illustrated in Figure 1. First, GPA needs participants to construct potential cluster centroids—synthetic virtual points that do not correspond to real data and are transmitted to the server only once. The server then applies Potential Prototype Merge (PPM) to localize high-density regions among all received centroids, estimating the number of novel categories and initializing the corresponding global prototypes. To facilitate effective learning, GPA introduces Semantic-Weighted Loss (SWL), which guides local training by enforcing alignment with global prototypes. This weighting strategy further enables Contrastive Weighted Loss (CWL), allowing GPA to learn from unlabeled novel-category data effectively. Besides, GPA incorporates a mixup mechanism at both the data and model levels to mitigate interference between known and novel categories while alleviating forgetting. Extensive experiments demonstrate that GPA achieves substantial improvements in novel-category accuracy for $\\mathbf{FC}^2\\mathbf{DL}$ , outperforming state-of-the-art methods with absolute gains of $5.7\\% - 13.1\\%$ in a single-stage setting and $4.1\\% - 13.0\\%$ across five stages, while maintaining near-zero forgetting on known categories. Furthermore, GPA is adaptable and can be seamlessly integrated into various mainstream FL algorithms, equipping them with strong NCDL capabilities. In summary, our contributions include:", + "bbox": [ + 91, + 306, + 482, + 699 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We address an important and practical problem called Federated Continuous Category Discovery and Learning $(\\mathbf{FC}^2\\mathbf{DL})$ , in which the biggest challenge is to merge and align novel categories of different participants that conform to a non-IID unsupervised setting.", + "- We design Global Prototype Alignment (GPA) to tackle FC $^2$ DL. GPA mainly consists of a novel category number estimation approach Potential Prototype Merge that can find high-density regions in unlabeled data, a Semantic-Weighted Loss and a Contrastive Weighted Loss that align the unlabeled novel data with global prototypes in local training, and a mixup mechanism for both the data and model that handle category interference and forgetting." + ], + "bbox": [ + 89, + 704, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We validate GPA on multiple datasets and various $\\mathbf{FC}^2\\mathbf{DL}$ scenarios. GPA achieves significant performance improvements over state-of-the-art baseline approaches when solving $\\mathbf{FC}^2\\mathbf{DL}$ . Moreover, except for the standard FedAvg [37] algorithm, GPA is highly effective in equipping many other FL algorithms with the capability of category discovery and learning.", + "bbox": [ + 513, + 90, + 903, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 208, + 653, + 223 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Federated Learning (FL) research [13, 20, 30] mainly focuses on static scenarios, assuming minimal changes in the FL system over time. However, this assumption is overly idealistic and fails to capture many real-world scenarios [54]. Although some studies [22, 38, 59] employ adversarial training to improve out-of-distribution generalization in FL models, and Federated Open-World and OpenSet Semi-Supervised Learning [33, 40, 61] aim to enable learning of unseen categories. However, these methods are restricted to supervised static FL settings, while existing unsupervised FL approaches [35, 60] focus solely on static scenarios, primarily training representation models, with downstream tasks requiring additional adaptation. Recent research has explored dynamic FL [10, 41, 58], where data distributions evolve over time; however, these studies remain confined to the supervised setting. To the best of our knowledge, only one study [40] addresses unseen category discovery in FL, but it depends on labeled data and lacks support for continuous learning. In summary, current FL research cannot discover and learn novel categories in unsupervised settings while mitigating catastrophic forgetting of previously learned knowledge.", + "bbox": [ + 511, + 233, + 903, + 565 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Novel Category Discovery and Learning (NCDL) refers to the process by which ML models identify and learn previously unseen categories [16]. Existing NCDL approaches can be broadly categorized into label-available and label-unavailable settings. Label-available NCDL, also referred to as Open-World Semi-Supervised Learning (SSL) [2, 15], typically integrates labeled known-category data with unlabeled novel-category data, leveraging SSL [32, 56], clustering [50], or weakly-supervised learning [1, 6] to facilitate novel category discovery. In contrast, the less-explored label-unavailable NCDL addresses practical scenarios where labeled known-category data is unavailable due to privacy constraints or cost limitations [42]. These methods typically follow a two-stage process: an initial supervised learning phase with labeled known-category data, followed by pseudo-labeling or self-supervised learning [44, 62, 63] on unlabeled data containing both known and novel categories. Recent research focuses on mitigating catastrophic forgetting during the continuous execution of the second stage [3, 36, 39, 64]. $\\mathsf{FC}^{2}\\mathsf{DL}$ falls into the label-unavailable NCDL category; however, data heterogeneity in FL significantly diminishes the effectiveness of existing", + "bbox": [ + 511, + 568, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2430", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a0499a4261107fff32fe66416be83b1a8e52d44e4178c58f84213b2cfa922f92.jpg", + "image_caption": [ + "Figure 1. Overview of GPA for enabling FL models to discover and learn novel categories continuously. When the known category learning finishes and the novel category learning starts, FL participants apply Kmeans on their data to find local potential novel prototypes. These local prototypes are sent to the server only once and GPA applies PPM to merge these prototypes into global novel prototypes. Then each selected client leverages SWL, CWL, and Data Mixup to conduct local training and update the model via Model Mixup." + ], + "image_footnote": [], + "bbox": [ + 98, + 90, + 883, + 424 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "NCDL approaches.", + "bbox": [ + 89, + 508, + 220, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 540, + 228, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Novel Category Discovery and Learning", + "text_level": 1, + "bbox": [ + 89, + 566, + 434, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In NCDL, there is a labeled dataset $\\mathcal{D}^{\\mathrm{L}} = \\{(x_i^{\\mathrm{L}},y_i^{\\mathrm{L}})\\}_{i = 1}^{N^{\\mathrm{L}}}$ and all labels belong to a one-hot label space of dimension $C^\\mathrm{L}$ , i.e., $\\pmb {y}_i^{\\mathrm{L}}\\in \\mathcal{V}^{\\mathrm{L}}$ and the input marginal distribution is $\\mathcal{X}^{\\mathrm{L}}$ . Besides, there is a unlabeled dataset $\\mathcal{D}^{\\mathrm{U}} = \\{\\pmb{x}_i^{\\mathrm{U}}\\}_{i = 1}^{N^{\\mathrm{U}}}$ which contains $C^\\mathrm{U}$ unique categories in $\\mathcal{V}^{\\mathrm{U}}$ , and its input distribution is $\\mathcal{X}^{\\mathrm{U}}$ . It is assumed that the labels in $\\mathcal{V}^{\\mathrm{L}}$ and $\\mathcal{V}^{\\mathrm{U}}$ are disjoint, i.e., $\\mathcal{V}^{\\mathrm{L}}\\cap \\mathcal{V}^{\\mathrm{U}} = \\emptyset$ , which are usually called known categories and novel categories, respectively. The general objective of NCDL is to partition the novel-category data into different clusters with the help of $\\mathcal{D}^{\\mathrm{L}}$ . More specifically, we leverage a neural network $m = f_{\\theta}\\circ g_{\\omega}$ which consists of a feature extractor $f$ parameterized on $\\theta$ and a classifier $g$ parameterized on $\\omega$ , to achieve NCDL:", + "bbox": [ + 89, + 588, + 483, + 790 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {\\theta , \\omega} \\mathbb {E} _ {\\mathcal {X} ^ {\\mathrm {U}}, \\mathcal {Y} ^ {\\mathrm {U}}} \\left\\{\\mathbb {I} \\left[ g _ {\\omega} \\left(f _ {\\theta} \\left(\\boldsymbol {x} _ {i}\\right)\\right) = g _ {\\omega} \\left(f _ {\\theta} \\left(\\boldsymbol {x} _ {j}\\right)\\right) \\right] \\cdot \\mathbb {I} \\left[ \\boldsymbol {y} _ {i} = \\boldsymbol {y} _ {j} \\right] \\right\\}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 803, + 483, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbb{I}[\\cdot ]$ is an indicator function that if the inner condition is true, $\\mathbb{I}[\\mathrm{True}] = 1$ , otherwise $\\mathbb{I}[\\mathrm{False}] = 0$ . $\\pmb{y}_i,\\pmb{y}_j$ are the ground truth labels of novel samples $\\pmb{x}_i,\\pmb{x}_j$ , though unavailable during NCDL.", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Federated Learning", + "text_level": 1, + "bbox": [ + 513, + 508, + 702, + 525 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "At the beginning of FL, we assume that a global model $m$ is trained across $K$ participants, and each participant holds its local dataset $\\mathcal{D}^k = \\{(x_i^k,y_i^k)\\}_{i = 1}^{N^k}$ . The data distributions of participants are non-IID from each other [19, 37]. In this case, the label spaces of different participants $\\mathcal{V}^k$ are not the same, but they all belong to a unified space $\\mathcal{V}^{\\mathrm{L}}$ , i.e., $\\mathcal{V}^k\\subseteq \\mathcal{V}^{\\mathrm{L}}$ , which can be regarded as known categories in NCDL. For every global round, a random set of participants is selected as clients, namely $\\mathcal{K}$ , and they need to conduct local training and then upload the model updates to the server. We assume that the local training is conducted with the CrossEntropy loss and the server applies FedAvg [37] to aggregate model updates:", + "bbox": [ + 511, + 531, + 906, + 729 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C E}} = - \\log g _ {\\omega} \\left(f _ {\\theta} \\left(\\boldsymbol {x} _ {i}\\right)\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 742, + 906, + 760 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Problem Formulation of $\\mathsf{FC^2DL}$", + "text_level": 1, + "bbox": [ + 513, + 771, + 787, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "After training in a sufficient number of rounds, which we call Known Category Learning $\\mathcal{T}^{\\mathrm{L}}$ , the FL model $m$ is hopefully able to perform well on $\\mathcal{V}^{\\mathrm{L}}$ . Then, the FL administrator may stop the training to save resources and deploy the latest model $m^{\\mathrm{L}}$ to serve all participants in the FL framework. During the usage of $m^{\\mathrm{L}}$ , participants feed it with unlabeled testing data $\\pmb{x}^{\\mathrm{test}}$ and obtain corresponding", + "bbox": [ + 511, + 794, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2431", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "inference results $g_{\\omega}(f_{\\theta}(\\boldsymbol{x}^{\\mathrm{test}}))$ . Intuitively, if $\\boldsymbol{x}^{\\mathrm{test}}$ belongs to the known categories $(\\mathcal{V}^{\\mathrm{L}})$ , the inference results should be mostly correct and accurate, i.e., $\\mathbb{I}[g_{\\omega}(f_{\\theta}(\\boldsymbol{x}^{\\mathrm{test}})) = \\boldsymbol{y}^{\\mathrm{test}}] = 1$ . However, if $\\boldsymbol{x}^{\\mathrm{test}}$ comes from unseen novel categories, the current model $m^{\\mathrm{L}}$ will classify it into a certain known category, in other words, $100\\%$ misclassification $(\\mathbb{I}[g_{\\omega}(f_{\\theta}(\\boldsymbol{x}^{\\mathrm{test}})) = \\boldsymbol{y}^{\\mathrm{test}}] = 0)$ .", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Now let us more formally define the problem of $\\mathbf{FC}^2\\mathbf{DL}$ . Suppose that the FL training periodically steps into a stage called Novel Category Learning $\\mathcal{T}^{\\mathrm{U}}$ , and we assume that there are $K^{\\mathrm{U}}$ participants available in total in $\\mathcal{T}^{\\mathrm{U}}$ . Note that $K^{\\mathrm{U}}$ is uncertain and may be changing dynamically, especially since old participants are free to withdraw and new participants are welcome to join in at any time [10]. At this moment, we denote the local novel-category dataset of participant $k$ as $\\mathcal{D}^{\\mathrm{U},k} = \\{\\pmb{x}_i^{\\mathrm{U},k}\\}_{i=1}^{N^{\\mathrm{U},k}}$ , screened from its received unlabeled testing dataset $\\mathcal{D}^{\\mathrm{test}}$ . Consistent with the non-IID setting, although $\\mathcal{D}^{\\mathrm{U},k}$ has no labels, its groundtruth label space $\\mathcal{Y}^{\\mathrm{U},k}$ is distinct across different participants but also belongs to a unified novel label space $\\mathcal{Y}^{\\mathrm{U}}$ . Our objective is to conduct effective NCDL that can cluster novel-category data into different clusters and preserve good performance on known categories as well:", + "bbox": [ + 89, + 198, + 482, + 441 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\max _ {\\theta , \\omega} \\left\\{\\sum_ {\\boldsymbol {x} \\in \\mathcal {D} ^ {\\mathrm {U}}} \\mathbb {I} \\left[ g _ {\\omega} \\left(f _ {\\theta} \\left(\\boldsymbol {x} _ {i}\\right)\\right) = g _ {\\omega} \\left(f _ {\\theta} \\left(\\boldsymbol {x} _ {j}\\right)\\right) \\right] \\cdot \\mathbb {I} \\left[ \\boldsymbol {y} _ {i} = \\boldsymbol {y} _ {j} \\right] \\right. \\tag {3} \\\\ \\left. \\right.\\left. + \\mathbb {E} _ {\\boldsymbol {x} \\sim \\mathcal {X} ^ {\\mathrm {L}}; \\boldsymbol {y} \\sim \\mathcal {Y} ^ {\\mathrm {L}}} \\mathbb {I} \\left[ g _ {\\omega} \\left(f _ {\\theta} (\\boldsymbol {x})\\right) = \\boldsymbol {y} \\right]\\right\\}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 449, + 482, + 522 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Our Proposed GPA Framework", + "text_level": 1, + "bbox": [ + 89, + 531, + 375, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Known Category Learning", + "text_level": 1, + "bbox": [ + 89, + 555, + 333, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To achieve better NCDL, model $m^{\\mathrm{L}}$ should not only have a good performance on known categories but also extract high-quality representations for novel categories. To this end, we modify Prototype Contrastive Learning (PCL) [28, 34] to enhance the training in $\\mathcal{T}^{\\mathrm{L}}$ . More specifically, our modified PCL constructs positive pairs as a sample representation $z_{i} = f_{\\theta}(x_{i})$ and its corresponding category prototype $p_{y_i}^{\\mathrm{L}}$ . The negative pairs consist of two parts: all pairs between $z_{i}$ and all prototypes $\\{p_c^{\\mathrm{L}}\\}_{c = 1}^{C^{\\mathrm{L}}}$ , and sample-wise pairs between $z_{i}$ and samples belonging to other labels:", + "bbox": [ + 89, + 577, + 482, + 733 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {P C L}} = - \\log \\frac {\\exp \\left(\\boldsymbol {z} _ {i} ^ {\\mathrm {L}} \\cdot \\boldsymbol {p} _ {\\boldsymbol {y} _ {i} ^ {\\mathrm {L}}} ^ {\\mathrm {L}} / \\tau\\right)}{\\sum_ {c = 1} ^ {C ^ {\\mathrm {L}}} \\exp \\left(\\boldsymbol {z} _ {i} ^ {\\mathrm {L}} \\cdot \\boldsymbol {p} _ {c} ^ {\\mathrm {L}} / \\tau\\right) + \\sum_ {j = 1} ^ {N _ {i} ^ {-}} \\exp \\left(\\boldsymbol {z} _ {i} ^ {\\mathrm {L}} \\cdot \\boldsymbol {z} _ {j} ^ {\\mathrm {L}} / \\tau\\right)} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 739, + 486, + 800 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tau = 0.07$ (consistent with [28]) is a temperature parameter and $N_{i}^{-}$ is the set size of samples with different labels from $\\pmb{x}_i$ in a data batch. Then we apply a combined loss $\\mathcal{L}_{\\mathcal{T}^{\\mathrm{L}}} = \\frac{1}{N^{\\mathrm{L},k}}\\sum_{\\mathcal{D}^{\\mathrm{L},k}}[\\mathcal{L}_{\\mathrm{CE}} + \\eta \\mathcal{L}_{\\mathrm{PCL}}]$ to train $m$ on known category data ( $\\eta = 0.1$ ; please refer to the Appendix B.4 for sensitivity analysis).", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Preparations before Novel Category Learning", + "text_level": 1, + "bbox": [ + 511, + 90, + 898, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2.1. Novel-Category Data Filtering.", + "text_level": 1, + "bbox": [ + 511, + 112, + 774, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When stepping into Novel Category Learning $\\mathcal{T}^{\\mathrm{U}}$ , it is possible for every participant to continually receive data of both novel and known categories. Thus we first need to separate known and novel categories from $\\mathcal{D}^{\\mathrm{test},k}$ . Specifically, for each participant $k$ , we assume that there is a data memory $\\mathcal{M}^k$ employed to store $\\mathcal{D}^{\\mathrm{U},k}$ that is filtered out during model $m^{\\mathrm{L}}$ usage*. We adopt a popular strategy [62] based on the representation distance between data samples and known-category prototypes. The traditional prototype [44, 62] is constructed by averaging category-wise representations, but this is infeasible in FL as the data is distributed among participants. We solve this problem by viewing the first $C^{\\mathrm{L}}$ dimensions of neuron weights $\\omega$ of classifier $g$ as the prototypes $\\mathcal{P}^{\\mathrm{L}} = \\{\\pmb{p}_c^{\\mathrm{L}}\\}_{c=1}^{C^{\\mathrm{L}}}$ , if $g$ is composed of only one single linear layer without bias [57]. The detailed filtering mechanism is as follows:", + "bbox": [ + 511, + 130, + 903, + 375 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} ^ {\\mathrm {U}, k} = \\left\\{\\boldsymbol {x} \\in \\mathcal {D} ^ {\\text {t e s t}, k} \\Bigg | \\max \\left\\{\\frac {f _ {\\theta} (\\boldsymbol {x}) \\cdot \\boldsymbol {p} _ {c} ^ {\\mathrm {L}}}{\\| f _ {\\theta} (\\boldsymbol {x}) \\| \\| \\boldsymbol {p} _ {c} ^ {\\mathrm {L}} \\|} \\right\\} _ {c = 1} ^ {C ^ {\\mathrm {L}}} < r \\right\\}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 516, + 378, + 903, + 433 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $r = 0.5$ is a threshold. This implies that if a sample $x$ has the largest similarity with known prototypes lower than $r$ , it should belong to novel categories. Note that such filtering cannot screen the known-category data completely. To prevent potential interference between known and novel categories, we design a data-level mixup strategy. As for the screened known-category data, we test GPA by applying Softmax-based pseudo labeling on them during $\\mathcal{T}^{\\mathrm{U}}$ (detailed results can be found in the Appendix A.3).", + "bbox": [ + 511, + 434, + 906, + 569 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2.2. Challenges of Estimating the Number of Novel Categories in $\\mathsf{FC}^2\\mathsf{DL}$", + "text_level": 1, + "bbox": [ + 511, + 575, + 905, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The goal of NCDL is to accurately classify novel data into different clusters, which requires prior knowledge of the cluster number. Most existing NCDL studies [32, 44, 56] directly assume this is known, but this assumption is unreasonable in $\\mathbf{FC}^2\\mathbf{DL}$ due to the following special challenges:", + "bbox": [ + 511, + 609, + 903, + 684 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- FL system administrator cannot access any training data and cannot predict the dynamic changes over time in data or task distribution [10, 37].", + "- There are various overlapping degrees between ordered clustering distributions $\\left[c_{1}^{\\mathrm{U}, k}, \\cdots, c_{i}^{\\mathrm{U}, k}, \\cdots\\right]$ over $\\mathcal{D}^{\\mathrm{U}, k}$ among different participants when applying unsupervised clustering algorithms locally, i.e.," + ], + "bbox": [ + 513, + 686, + 903, + 790 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {P} \\left(\\boldsymbol {c} _ {i} ^ {\\mathrm {U}, k} \\neq \\boldsymbol {c} _ {i} ^ {\\mathrm {U}, k ^ {\\prime}}\\right) > 0, \\mathcal {P} \\left(\\boldsymbol {c} _ {i} ^ {\\mathrm {U}, k} = \\boldsymbol {c} _ {j} ^ {\\mathrm {U}, k ^ {\\prime}}\\right) > 0, (i \\neq j). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 797, + 903, + 832 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "*Note that old participants can directly apply the data memory that was previously used for storing the known-category data to store the screened novel data, after removing the known-category data due to memory storage limitation or moving them to somewhere else for privacy or intellectual property protection [10, 42, 53].", + "bbox": [ + 511, + 839, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2432", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Existing novel-category number estimation methods [7, 16, 50] require labeled data, either from known categories or auxiliary novel categories. This requirement violates cost-cutting and privacy regulations [14] for old participants, while new participants do not have any labeled data [10]. Besides, these methods rely on sufficient observation of the novel data distribution, i.e., they need to access a large amount of novel-category data, which is also unreasonable in practice.", + "bbox": [ + 89, + 90, + 483, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.3. Potential Prototype Merge.", + "text_level": 1, + "bbox": [ + 89, + 234, + 323, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To achieve an accurate global novel-category number estimation, all participants need to identify potential local prototypes within their local data first. Unlike studies [1, 32] that employ Sinkhorn-Knopp-based clustering [12] to cluster unlabeled data into clusters with an equal size, we adopt standard Kmeans [17] to construct local prototypes. The reason is that local data is often class-imbalanced [52]. Specifically, each participant $k$ feeds its local data $\\mathcal{D}^{\\mathrm{U},k}$ into the model $m^{\\mathrm{L}}$ to extract representations $\\boldsymbol {z}_i = f_\\theta (\\boldsymbol{x}_i^{\\mathrm{U},k})$ and then employs Kmeans clustering to group these representations into $C^\\mathrm{L}$ clusters. Subsequently, the centroids of these clusters are utilized as local prototypes:", + "bbox": [ + 89, + 253, + 483, + 435 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {Z} ^ {\\mathrm {U}, k} = \\left\\{\\frac {\\sum_ {\\boldsymbol {z} \\in \\mathcal {Z} _ {c} ^ {\\mathrm {U} , k}} \\boldsymbol {z}}{| \\mathcal {Z} _ {c} ^ {\\mathrm {U} , k} |} \\right\\} _ {c = 1} ^ {C ^ {\\mathrm {L}}}, \\mathcal {Z} _ {c} ^ {\\mathrm {U}, k} = \\left\\{\\boldsymbol {z} \\mid \\operatorname {K M} (\\boldsymbol {z}) = \\boldsymbol {c} _ {c} ^ {\\mathrm {U}, k} \\right\\}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 93, + 446, + 483, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{KM}(z)$ outputs the cluster label of $z$ when employing Kmeans clustering. Notably, these Kmeans centroids are synthetic points that achieve the minimum aggregated intra-cluster distance. In other words, they do not correspond to any real data. Moreover, these local prototypes are sent to the server only once. In these ways, privacy is protected. More discussions about why our design can protect privacy are provided in the Appendix B.6.", + "bbox": [ + 89, + 505, + 483, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The server shuffles all received local prototypes and constructs a prototype pool $\\mathcal{Z}^{\\mathrm{U}} = \\bigcup_{k=1}^{K^{\\mathrm{U}}} \\mathcal{Z}^{\\mathrm{U}, k}$ . When building PPM, we totally depart from conventional design principles in existing novel-category number estimation methods [7, 16, 50]. Instead, we leverage an intuition that data representations belonging to the same novel category appear in close proximity, forming high-density regions. Identifying these regions globally is achievable since we do not conduct any normalization when building local prototypes. Specifically, we modify DBSCAN [45] with a rising radius $\\epsilon$ and a fixed minimum cluster data size $n_{\\mathrm{size}}$ to find the high-density regions. $\\epsilon$ controls how large the nearby area should be considered for a certain data point, while $n_{\\mathrm{size}}$ defines the requirement that if a data point is a core point, it should be surrounded by $n_{\\mathrm{size}} - 1$ points. To cover as", + "bbox": [ + 89, + 626, + 483, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "many high-density regions as possible, we try to use a loose setting with $n_{\\mathrm{size}} = 2$ , which also considers the global imbalance case where minority categories correspond to very few prototypes globally [52]. As for $\\epsilon$ , we first calculate the pair-wise Euclidean distance range within the prototype pool $\\mathcal{Z}^{\\mathrm{U}}$ and set a rising $\\epsilon$ as:", + "bbox": [ + 511, + 90, + 905, + 181 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\epsilon_ {e} = \\min D + e \\cdot (\\max D - \\min D) / E \\right\\} _ {e = 0} ^ {E}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 191, + 903, + 218 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $D = \\{d(\\pmb {z}_i^{\\mathrm{U}},\\pmb {z}_j^{\\mathrm{U}})\\}_{\\pmb {z}_i^{\\mathrm{U}},\\pmb {z}_j^{\\mathrm{U}}\\in \\mathcal{Z}^{\\mathrm{U}}}$", + "bbox": [ + 553, + 210, + 790, + 233 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $d(\\cdot)$ computes the Euclidean distance, and $E = 50$ is the total rising steps. For each $\\epsilon_{e}$ , we conduct DBSCAN clustering on $\\mathcal{Z}^{\\mathrm{U}}$ and record the unique resulting cluster number $\\tilde{c}_e^{\\mathrm{U}}$ . After conducting the final step, we choose the maximum cluster number as the estimated global novel-category number $\\widetilde{C}^{\\mathrm{U}} = \\max \\left\\{\\tilde{c}_{e}^{\\mathrm{U}}\\right\\}_{e=0}^{E}$ . We believe that PPM can work effectively with much fewer requirements than existing methods, thus tackling the aforementioned challenges of estimating the novel-category number in FC²DL.", + "bbox": [ + 511, + 242, + 906, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then we apply Kmeans again on $\\mathcal{Z}^{\\mathrm{U}}$ with the cluster number $\\widetilde{C}^{\\mathrm{U}}$ and use the cluster centroids $\\hat{z}_c^{\\mathrm{U}}$ to initialize the global novel prototypes as $\\mathcal{P}^{\\mathrm{U}} = \\{\\hat{z}_c^{\\mathrm{U}}\\}_{c=1}^{\\widetilde{C}^{\\mathrm{U}}}$ .", + "bbox": [ + 511, + 378, + 905, + 428 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Representation Learning for Novel Categories", + "text_level": 1, + "bbox": [ + 511, + 435, + 901, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3.1. Semantic-Weighted Loss", + "text_level": 1, + "bbox": [ + 511, + 457, + 732, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By viewing our problem here as weakly-supervised learning, we propose a novel Semantic-Weighted Loss (SWL) to align local data with the global prototypes $\\mathcal{P}^{\\mathrm{U}}$ in representation space. Specifically, instead of assigning data samples with a single label, SWL considers the possible semantic relationships between data samples with all global prototypes, which are measured by the Softmax probability. Moreover, as $\\mathcal{P}^{\\mathrm{U}}$ is obtained by Euclidean distance-based Kmeans, we minimize the Euclidean distance between samples and prototypes. In this case, SWL is shaped as:", + "bbox": [ + 511, + 476, + 905, + 628 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {S W L}} = \\sum_ {c = 1} ^ {\\tilde {C} ^ {\\mathrm {U}}} \\frac {\\exp \\left(f _ {\\theta} \\left(\\boldsymbol {x} _ {i}\\right) \\cdot \\hat {\\boldsymbol {z}} _ {c} ^ {\\mathrm {U}} / \\tau\\right)}{\\sum_ {j = 1} ^ {\\tilde {C} ^ {\\mathrm {U}}} \\exp \\left(f _ {\\theta} \\left(\\boldsymbol {x} _ {i}\\right) \\cdot \\hat {\\boldsymbol {z}} _ {j} ^ {\\mathrm {U}} / \\tau\\right)} d \\left(f _ {\\theta} \\left(\\boldsymbol {x} _ {i}\\right), \\hat {\\boldsymbol {z}} _ {c} ^ {\\mathrm {U}}\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 637, + 903, + 699 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By weighting the Euclidean distance with the semantic similarity, SWL can guide the model to learn more distinguishable semantics and be more resistant to pseudo-label noise.", + "bbox": [ + 511, + 708, + 905, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3.2. Contrastive Weighted Loss", + "text_level": 1, + "bbox": [ + 511, + 760, + 746, + 775 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SWL's effectiveness relies on the correlations between known and novel categories captured by the model $m^{\\mathrm{L}}$ , but $m^{\\mathrm{L}}$ is trained on labeled known-category data and thus naturally cannot extract high-quality representations for unlabeled novel-category data. To this end, we employ instance discrimination [5] to facilitate better feature extraction from unlabeled novel-category data. However, standard instance discrimination requires lengthy training to converge, which", + "bbox": [ + 511, + 779, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "$^{\\dagger}$ We assume that $C^{\\mathrm{U}} < C^{\\mathrm{L}}$ as previous studies [44, 62]. Clustering local novel data into $C^{\\mathrm{L}}$ clusters can cover as complete novel label space as possible.", + "bbox": [ + 89, + 861, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2433", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "lags far behind the pace of SWL. We found that enabling the training to focus on easy samples in the early stage that is more like a curriculum learning can accelerate the convergence process. Therefore, similar to SWL, we also use the semantic relationship to weight the InfoNCE loss,", + "bbox": [ + 89, + 90, + 483, + 167 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C W L}} = - \\Delta (\\boldsymbol {x} _ {i}, \\boldsymbol {x} _ {i}) \\log \\Delta (\\boldsymbol {x} _ {i}, \\boldsymbol {x} _ {i}),\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 176, + 395, + 191 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\Delta (\\pmb {x}_i,\\pmb {x}_i) = \\frac{\\exp\\left(f_\\theta(\\mathrm{A}(\\pmb{x}_i))\\cdot f_\\theta(\\mathrm{A}(\\pmb{x}_i)) / \\tau\\right)}{\\sum_{q = 1}^{N_B^k}\\exp\\left(f_\\theta(\\mathrm{A}(\\pmb{x}_i))\\cdot f_\\theta(\\mathrm{A}(\\pmb{x}_q)) / \\tau\\right)}$ (10)", + "bbox": [ + 91, + 195, + 483, + 246 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $A(\\cdot)$ denotes using stochastic augmentation and $N_B^k$ is the size of a mini-batch.", + "bbox": [ + 89, + 247, + 483, + 276 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.3. Data Mixup", + "text_level": 1, + "bbox": [ + 89, + 285, + 220, + 299 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To preserve the knowledge learned from labeled known categories and optimize for novel categories simultaneously, each client $k$ applies a mixup mechanism on the data and model. The principle behind the data-level mixup is to handle the interference between known categories and novel categories in the novel data memory $\\mathcal{M}^k$ . For the specific design, each client $k$ constructs a relatively high-confidence dataset $\\mathcal{D}^{\\mathrm{U},k}$ from $\\mathcal{D}^{\\mathrm{U},k}$ in terms of Softmax probability $\\widetilde{\\boldsymbol{y}}$ :", + "bbox": [ + 89, + 303, + 483, + 425 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} ^ {\\mathrm {U} ^ {\\prime}, k} = \\left\\{\\boldsymbol {x} \\mid \\max \\left\\{\\frac {\\exp \\left(f _ {\\theta} \\left(\\boldsymbol {x} _ {i}\\right) \\cdot \\boldsymbol {p} _ {c} / \\tau\\right)}{C ^ {\\mathrm {L}} + \\widetilde {C} ^ {\\mathrm {U}}} \\right\\} _ {c = 1} ^ {C ^ {\\mathrm {L}} + \\widetilde {C} ^ {\\mathrm {U}}} > r ^ {\\prime} \\right\\}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 433, + 486, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$r'$ here is a threshold for selecting data to build $\\mathcal{D}^{\\mathrm{U}',k}$ , and we set it as 0.95 (please refer to the Appendix B.4 for the sensitivity analysis). Then for every instance $\\pmb{x}_i'$ in $\\mathcal{D}^{\\mathrm{U}',k}$ , client $k$ will randomly select another instance $\\pmb{x}_i$ from the residual set $\\mathcal{D}^{\\mathrm{U},k} \\setminus \\mathcal{D}^{\\mathrm{U}',k}$ and conduct the data-level mixup $\\pmb{x}_i^* = \\lambda \\pmb{x}_i' + (1 - \\lambda)\\pmb{x}_i$ , as well as the training as follows:", + "bbox": [ + 89, + 510, + 483, + 602 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {M I X}} = \\lambda \\mathcal {L} _ {\\mathrm {C E}} \\left(f _ {\\theta} \\left(\\boldsymbol {x} _ {i} ^ {*}\\right), \\widetilde {\\boldsymbol {y}} _ {i} ^ {\\prime}\\right) + (1 - \\lambda) \\mathcal {L} _ {\\mathrm {C E}} \\left(f _ {\\theta} \\left(\\boldsymbol {x} _ {i} ^ {*}\\right), \\widetilde {\\boldsymbol {y}} _ {i}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 611, + 480, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda \\sim \\mathrm{Beta}(\\lambda; 0.75, 0.75)$ that is suggested in semi-supervised learning works [9]. With the data mixup, the optimization objective of the novel category learning is", + "bbox": [ + 89, + 642, + 483, + 688 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathcal {T} ^ {\\mathrm {U}}} = \\frac {1}{N ^ {\\mathrm {U} , k}} \\sum_ {\\mathcal {D} ^ {\\mathrm {U}, k}} \\left(\\mathcal {L} _ {\\mathrm {S W L}} + \\mathcal {L} _ {\\mathrm {C W L}}\\right) + \\frac {1}{\\left| \\mathcal {D} ^ {\\mathrm {U} ^ {\\prime} , k} \\right|} \\sum_ {\\mathcal {D} ^ {\\mathrm {U} ^ {\\prime}, k}} \\left(\\mathcal {L} _ {\\mathrm {M I X}} + \\mathcal {L} _ {\\mathrm {C E}}\\right). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 696, + 496, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.4. Model Mixup", + "text_level": 1, + "bbox": [ + 89, + 753, + 232, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition to data-level mixup, we also conduct a model-level mixup to compensate for the forgetting of known categories. Specifically, each client $k$ updates $f_{\\theta}^{k}$ for each local training round:", + "bbox": [ + 89, + 773, + 483, + 834 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\theta^ {k} = \\beta \\theta^ {\\mathrm {L}} + (1 - \\beta) \\theta^ {k}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 840, + 483, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\beta$ controls the merging speed, which is set as 0.99 (please see the Appendix B.4 for sensitivity analysis). Note", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "that such model mixup differs from the standard Exponential Moving Average as the first term $\\theta^{\\mathrm{L}}$ here remains unchanged. As for the update of classifier $g_{\\omega}$ , the neuron weights corresponding to known categories are frozen during GPA, while the neuron weights corresponding to novel categories are updated freely through back-propagation on the local side and FedAvg on the global side.", + "bbox": [ + 511, + 90, + 905, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experimental Results", + "text_level": 1, + "bbox": [ + 511, + 210, + 718, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The main experimental setup, results, and analysis are introduced below. Please refer to Supplementary Materials for more details and all implementation source codes.", + "bbox": [ + 511, + 234, + 905, + 281 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. CIFAR-100 [25] consists of 100 classes, and each class contains 600 images. Tiny-ImageNet [27] contains 200 classes with 500 images per class. ImageNet-S is a subset of ImageNet [8], including 100 classes and each class has 600 images. Three fine-grained datasets are also used: CUB200 [51] (200 classes), StanfordCars [24] (196 classes), and Herbarium 19 [47] (683 classes). These three datasets are usually class-imbalanced.", + "bbox": [ + 511, + 284, + 905, + 405 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. The number of participants is 10, and 5 are randomly selected as the clients to conduct 10-epoch local training in every global round. The number of global rounds for both $\\mathcal{T}^{\\mathrm{L}}$ and $\\mathcal{T}^{\\mathrm{U}}$ is 30. To simulate the non-IID setting, the data distributions of all participants follow a Dirichlet distribution [19] with the parameter $\\alpha = 0.1$ (experiments with other values can be found in the Appendix). We use ResNet-18 [18] as the feature extractor for CIFAR-100 and Tiny-IIImageNet, and ResNet-34 for ImageNet-S, CUB-200, StanfordCars, and Herbarium-19. For all datasets, we employ a bias-free linear classifier. Notably, these ResNets are trained from scratch, as our goal is to evaluate the direct impact of the proposed algorithms by preventing the influence of factors such as pretraining. Certainly, we conduct experiments on Vision Transformers [11] that have been pre-trained on large-scale datasets, and the results are provided in the Appendix B.8. The batch size is 128 and the SGD learning rate is 0.02. All experiments are conducted repeatedly with seeds 2023, 2024, and 2025.", + "bbox": [ + 511, + 409, + 905, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. For known categories, we use ground-truth labels to calculate the classification accuracy. As for novel categories, we follow [62] to adopt Hungarian clustering accuracy [26] as the metric. We also use the overall accuracy [44] calculated on known and novel categories to measure the model performance further. The accuracy forgetting on known categories after $\\mathcal{T}^{\\mathrm{U}}$ of the experiments below can be found in the Appendix B.7.", + "bbox": [ + 511, + 700, + 905, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison baselines. For a fair comparison, we try our best to customize recent NCDL methods (AutoNovel [16], GM [62], iNCD [44], Happy [36], IIC [32], OpenCon [46]), novel-category estimation approaches (MACC [50], EMaCS [7]), a preliminary federated novel", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2434", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/024508124a7d109603533f4f783d9571cde04f5b01bd2778534b670311515b78.jpg", + "table_caption": [ + "Table 1. Performance comparison between GPA and other baseline methods in ${\\mathbf{{FC}}}^{2}\\mathbf{{DL}}$ ,when there is only one novel category learning stage. Left precision is the mean accuracy, and the right is the variance of three runs. We bold and blue the best performance, and bold the second-best; same for all following tables." + ], + "table_footnote": [], + "table_body": "
MethodCIFAR-100Tiny-ImageNetImageNet-SCUB200
knownnovelallknownnovelallknownnovelallknownnovelall
AutoNovel67.3±1.033.4±2.460.6±0.552.3±0.425.0±0.749.6±0.455.1±021.0±0.748.3±040.0±018.1±2.337.5±0.5
GM71.4±029.9±0.963.1±057.3±019.4±1.253.6±055.8±019.9±0.948.6±038.5±2.219.7±5.537.3±2.0
iNCD69.8±0.235.2±15.562.9±1.256.4±025.3±2.153.3±055.5±0.224.6±9.149.3±0.839.2±1.020.0±2.037.2±1.2
Happy70.2±1.139.2±4.464.0±0.856.4±0.327.8±2.253.5±0.255.8±029.0±1.650.4±0.340.2±0.225.8±1.138.8±0.4
IIC60.8±2.333.1±1.855.3±2.442.8±0.723.3±0.640.9±0.542.3±1.519.1±8.237.7±2.438.7±1.118.5±0.936.6±0.4
OpenCon61.5±1.718.4±1.452.9±1.547.6±2.015.6±1.744.4±2.047.4±3.918.0±10.141.6±1.029.9±3.411.2±6.728.3±2.5
Orchestra66.6±0.533.1±0.860.0±0.253.4±0.324.7±1.150.1±0.152.8±0.321.5±0.348.4±040.0±018.5±1.037.5±0.1
SemiFL44.2±0.229.1±0.541.2±0.233.5±0.125.7±1.132.7±0.638.8±0.227.5±0.536.5±0.230.0±0.718.7±1.329.1±0.6
FedoSSL62.2±0.218.2±0.753.4±0.453.0±0.18.7±1.048.6±0.455.0±017.5±0.247.5±0.138.8±014.5±0.736.9±0.3
AGCL65.6±0.235.5±0.559.6±0.243.2±0.524.5±1.441.3±0.746.9±0.321.0±0.641.7±0.336.0±020.5±1.134.8±0.7
GPA71.9±052.3±1.068.0±0.457.6±039.1±1.955.8±0.755.8±037.4±1.052.1±0.440.3±0.231.5±0.739.4±0.3
", + "bbox": [ + 135, + 122, + 856, + 301 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6cb11a414ba33bea2cdb73ad46f923d75ac600816698c9a881ff0a71a60ad854.jpg", + "table_caption": [ + "Table 2. Performance comparison between GPA and other baseline methods in ${\\mathbf{{FC}}}^{2}\\mathbf{{DL}}$ ,when there are five novel category learning stage. Left precision is the mean accuracy, and the right is the variance of three runs." + ], + "table_footnote": [], + "table_body": "
DatasetMethod1st Stage2nd Stage3rd Stage4th Stage5th Stage
knownnovelallknownnovelallknownnovelallknownnovelallknownnovelall
CIFAR-100GM72.9±1.340.2±2.469.1±1.571.5±0.837.8±1.464.4±0.970.7±1.031.4±2.759.5±1.369.5±0.627.9±1.955.0±1.168.8±2.526.0±2.051.7±1.7
iNCD71.5±0.544.4±1.768.3±0.668.9±1.241.6±1.563.2±0.866.0±1.334.5±1.457.0±1.165.7±0.833.0±1.754.3±0.965.0±1.031.2±1.951.5±1.3
Happy72.0±0.545.6±1.168.9±0.768.8±0.944.7±1.563.7±0.965.8±0.741.0±2.158.7±1.064.5±1.338.7±0.855.5±1.063.7±1.035.5±2.052.4±1.2
GPA74.0±057.8±1.472.1±0.673.5±0.656.5±1.369.9±0.773.3±0.553.0±1.067.5±0.473.0±0.951.7±1.665.6±1.172.5±1.048.2±2.362.8±1.5
T-ImageNetGM58.3±0.237.0±1.157.3±0.356.6±0.423.4±3.453.6±0.855.8±1.720.2±2.351.2±1.854.0±0.818.8±2.048.1±1.252.5±1.718.0±2.445.6±1.9
iNCD58.0±0.538.2±0.757.1±0.553.1±1.031.0±1.451.1±1.150.8±0.925.0±1.947.4±1.147.7±2.623.2±1.443.6±2.945.3±1.122.1±1.840.7±1.3
Happy58.3±0.236.5±1.557.3±0.655.7±0.930.9±2.153.4±1.353.4±1.228.0±1.450.1±1.252.0±1.126.3±1.447.7±1.251.0±0.724.9±1.845.8±1.1
GPA58.5±0.248.2±1.658.0±0.658.0±0.543.5±2.256.7±0.957.3±0.739.5±1.255.0±0.957.0±0.638.0±2.153.8±1.056.8±1.136.9±2.552.8±1.5
ImageNet-SGM57.0±037.2±0.454.7±0.156.5±0.226.7±1.950.2±0.854.8±0.519.5±1.444.7±0.653.0±0.518.2±1.040.9±0.652.9±1.517.3±1.838.7±1.2
iNCD55.5±0.236.2±0.753.2±0.354.1±0.430.2±1.349.1±0.552.0±0.324.5±0.844.1±0.449.9±1.422.3±1.940.3±1.548.0±1.020.2±1.536.9±1.2
Happy56.5±0.437.9±1.454.3±0.556.0±0.833.7±2.251.3±1.153.5±1.629.2±2.446.6±1.952.0±1.027.5±1.643.5±1.351.4±1.826.0±2.741.2±2.2
GPA57.2±046.0±1.155.9±0.457.0±0.542.2±0.753.9±0.556.4±0.937.0±2.050.9±1.455.7±1.235.1±1.548.5±1.255.5±1.332.9±3.046.5±1.9
CUB200GM42.5±0.328.9±0.941.9±0.442.0±0.523.7±1.040.3±0.641.0±0.719.5±1.438.2±0.940.2±0.518.1±1.236.5±0.738.5±0.617.7±1.634.3±0.9
iNCD41.0±0.534.0±0.940.7±0.539.8±0.624.6±1.838.4±1.037.0±1.020.0±2.034.8±1.236.5±1.118.9±1.733.6±1.333.8±0.917.9±1.430.6±1.0
Happy42.2±0.733.9±1.141.8±0.841.1±0.529.5±1.640.0±0.739.8±0.525.5±1.237.9±0.739.0±1.023.3±1.436.4±1.137.8±0.823.0±2.034.8±1.2
GPA42.8±0.238.0±0.942.6±0.342.5±0.735.3±1.441.8±0.942.2±0.731.0±1.940.7±1.041.8±0.429.4±1.239.7±0.641.5±0.828.8±1.739.0±1.0
", + "bbox": [ + 89, + 345, + 906, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "category discovery work (AGCL [40]), federated self/semi-supervised learning methods (Orchestra [35], SemiFL [9], FedoSSL [61]), and popular FL algorithms (FedProx [31], SCAFFOLD [23], Moon [29]) reasonably in the setting of $\\mathbf{FC}^2\\mathbf{DL}$ , and compare them with our GPA.", + "bbox": [ + 89, + 597, + 482, + 674 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Effectiveness of GPA on $\\mathsf{FC^2DL}$", + "text_level": 1, + "bbox": [ + 89, + 688, + 356, + 704 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison on one novel category learning stage. We first assume that there is only one $\\mathcal{T}^{\\mathrm{U}}$ in $\\mathsf{FC^{2}DL}$ , which is consistent with conventional NCDL studies. For CIFAR100, Tiny-ImageNet, ImageNet-S, CUB200, and StanfordCars, we set the novel category number as 20 and 83 for Herbarium 19, while the rest are regarded as known categories. The experiment results are shown in Tables 1 and 5. The results demonstrate that our GPA framework always achieves the best performance on all metrics, including accuracy on known, novel, and all categories. The performance improvement on novel-category accuracy is particularly promising, ranging from $5.7\\%$ on CUB200 to", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$13.1\\%$ on CIFAR-100 over the best baseline method, which demonstrates the effectiveness of GPA.", + "bbox": [ + 511, + 597, + 906, + 627 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison on multiple novel category learning stages. As mentioned in related works [10], FL participants often collect unseen novel-category data continually. For such cases, we split each dataset into a labeled subset and 5 unlabeled subsets (each contains 8 novel categories) with disjoint categories, and conduct five $\\mathcal{T}^{\\mathrm{U}}$ in sequence. For a fair comparison, we only compare GPA with GM [62], iNCD [44], and Happy [36] as only they have dedicated designs for multiple novel category learning stages. All experiments are conducted using the category number estimated by our PPM and the detailed results are shown in Table 2. We can observe that GPA achieves the best performance on all datasets, showing $4.1\\%$ to $13.0\\%$ improvement on novel categories compared with the best-performing approach across 1st and 5th novel category learning stages.", + "bbox": [ + 511, + 636, + 906, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison on estimating novel-category number. Furthermore, to validate the effectiveness of PPM, we com", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2435", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2d523bdd3c2681d91e2fc8b3b1be47e57281e892595ec92ba64b34ed156883d2.jpg", + "table_caption": [ + "Table 3. Performance comparison between PPM and other novel category number estimation approaches in various novel category number settings. The experiments here show that PPM can consistently provide the most accurate novel category number estimation." + ], + "table_footnote": [], + "table_body": "
DatasetCIFAR-100Tiny-ImageNetImageNet-SCUB200StanfordCarsHerbarium19Average\nError (%)
Class #5101520253010203040505101520102030401020304020406080
MACC4566141568122228466144613186122024918243048.9
EMaCS18173023121127452526582120341521353429194542362517384679.5
PPM4913192529819314054581217921324411202839194265868.1
", + "bbox": [ + 137, + 119, + 859, + 185 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c39bf1041eda92e63752d7b86853d95ee88280decbe7c25064bccf7bb3be6f95.jpg", + "image_caption": [ + "Figure 2. Performance evaluation of combining GPA with other FL algorithms. We assume that Kmeans can equip FedProx, SCAFFOLD, and Moon with a certain level of NCDL capability, while GPA can substantially exceed it. The FC $^2$ DL setting is the same as Table 1." + ], + "image_footnote": [], + "bbox": [ + 94, + 202, + 898, + 325 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/69c758672d66d35b5bc8a0c3d1209224a0a4f115cac0b6a00162e203434ee3a5.jpg", + "table_caption": [ + "Table 4. Ablation study of major components in GPA. The $FC^2DL$ setting is the same as Table 1." + ], + "table_footnote": [], + "table_body": "
MethodCIFAR-100Tiny-ImageNetImageNet-S
knownnovelallknownnovelallknownnovelall
GPA-w/o LpCL70.251.166.454.337.052.653.835.950.2
GPA-w/o LSWL71.837.364.957.631.155.055.830.250.7
GPA-w/o LCWL71.942.966.157.633.055.155.832.751.2
GPA-w/o LMix71.845.266.557.634.355.355.734.051.4
GPA-w/ MACC71.539.865.257.627.954.655.729.650.5
GPA-w/o M.Mix69.848.965.655.238.253.553.236.349.8
GPA71.952.368.057.639.155.855.837.452.1
", + "bbox": [ + 89, + 398, + 480, + 522 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "pare it with state-of-the-art approaches (MACC [50] and EMaCS [7]) when estimating different novel category numbers in $\\mathsf{FC^2DL}$ . The detailed category number setups and experiment results are shown in Table 3. We can observe that our PPM provides the most accurate category number estimation for all datasets, with an average error rate of $8.1\\%$ , while the other approaches have approximately $50\\%$ and even higher average error rate.", + "bbox": [ + 88, + 553, + 482, + 675 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Experiments with more FL algorithms. To further evaluate the applicability of GPA, we also carry out experiments combining GPA with FedProx [31], SCAFFOLD [23], and Moon [29]. Specifically, we assume that regular FL can possess a certain level of NCDL capability through unsupervised clustering (e.g., Kmeans) on novel-category data representations. The experiment results are shown in Figure 2. We can see that using GPA with these FL algorithms provides much better performance on novel categories.", + "bbox": [ + 88, + 681, + 482, + 818 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "More experiments in the Appendix. 1) experiments in the centralized setting—Section B.1; 2) experiments with various degrees of heterogeneity in $\\mathbf{FC}^2\\mathbf{DL}$ —Section B.2; 3) experiments with different data category partitions—Section B.3; 4) sensitivity analysis of hyper-parameters—", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Section B.4; 5) experiments with a larger number of participants—Section B.5 6) experiments on pre-trained vision transformers—Section B.8; 7) experiments of launching data reconstruction attacks for verifying GPA's characteristic of privacy-preserving—Section B.6.", + "bbox": [ + 511, + 381, + 906, + 457 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 470, + 666, + 486 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We investigate the importance of each component in GPA via an ablation study built-in $\\mathbf{FC}^2\\mathbf{DL}$ with one novel category learning stage. As shown in Table 4, 'GPA-w/o $\\mathcal{L}_{\\mathrm{PCL}}$ 'GPA-w/o $\\mathcal{L}_{\\mathrm{SWL}}$ ,GPA-w/o $\\mathcal{L}_{\\mathrm{CWL}}$ ,GPA-w/o $\\mathcal{L}_{\\mathrm{MIX}}$ GPA-w/MACC, 'GPA-w/o M.Mix' denote the performance of GPA without using PCL in $\\mathcal{T}^{\\mathrm{L}}$ , without using SWL (replacing it with a Binary CrossEntropy Loss [16]), without using CWL, without using data mixup, with MACC instead of PPM to estimate category number, and without applying model mixup (replacing it with distillation [44]), respectively. We can observe that these versions of GPA degrade evidently in performance, which shows the importance of all these modules in our approach.", + "bbox": [ + 511, + 493, + 908, + 690 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 707, + 633, + 723 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we study an important and practical problem called Federated Continuous Category Discovery and Learning (FC $^2$ DL), and propose Global Prototype Alignment (GPA) to help the FL model discover and learn novel categories. GPA can accurately estimate the novel category number, construct corresponding global prototypes, and guide the local training of different participants to tackle data heterogeneity. Extensive experiments demonstrate that GPA significantly outperforms state-of-the-art novel category learning methods and other potential solutions in various scenarios for the FC $^2$ DL problem.", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2436", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This research is supported by the National Research Foundation, Singapore, and the Cyber Security Agency of Singapore under the National Cybersecurity R&D Programme and the CyberSG R&D Programme Office (Award CRPO-GC3-NTU-001) and NTU-NAP startup grant (024584-00001). Any opinions, findings, conclusions, or recommendations expressed in these materials are those of the author(s) and do not reflect the views of the National Research Foundation, Singapore, the Cyber Security Agency of Singapore, or the CyberSG R&D Programme Office.", + "bbox": [ + 89, + 114, + 483, + 267 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 280, + 187, + 296 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Wenbin An, Feng Tian, Qinghua Zheng, Wei Ding, QianYing Wang, and Ping Chen. Generalized category discovery with decoupled prototypical network. arXiv preprint arXiv:2211.15115, 2022. 2, 5", + "[2] Kaidi Cao, Maria Brbic, and Jure Leskovec. Open-world semi-supervised learning. In International Conference on Learning Representations (ICLR), 2022. 2", + "[3] Fernando Julio Cendra, Bingchen Zhao, and Kai Han. Promptcd: Learning gaussian mixture prompt pool for continual category discovery. In European Conference on Computer Vision, pages 188-205. Springer, 2024. 2", + "[4] Mingzhe Chen, Nir Shlezinger, H Vincent Poor, Yonina C Eldar, and Shuguang Cui. Communication-efficient federated learning. Proceedings of the National Academy of Sciences, 118(17):e2024789118, 2021. 1", + "[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PmLR, 2020. 5", + "[6] Haoang Chi, Feng Liu, Wenjing Yang, Long Lan, Tongliang Liu, Bo Han, Gang Niu, Mingyuan Zhou, and Masashi Sugiyama. Meta discovery: Learning to discover novel classes given very limited data. In International Conference on Learning Representations. 2", + "[7] Florent Chiaroni, Jose Dolz, Ziko Imtiaz Masud, Amar Mitiche, and Ismail Ben Ayed. Parametric information maximization for generalized category discovery. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1729-1739, 2023. 2, 5, 6, 8, 13", + "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6", + "[9] Enmao Diao, Jie Ding, and Vahid Tarokh. Semifl: Semi-supervised federated learning for unlabeled clients with alternate training. Advances in Neural Information Processing Systems, 35:17871-17884, 2022. 6, 7, 13", + "[10] Jiahua Dong, Lixu Wang, Zhen Fang, Gan Sun, Shichao Xu, Xiao Wang, and Qi Zhu. Federated class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10164-10173, 2022. 2, 4, 5, 7, 12" + ], + "bbox": [ + 93, + 305, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2020. 6", + "[12] Aude Geneva, Gabriel Dulac-Arnold, and Jean-Philippe Vert. Differentiable deep clustering with cluster size constraints. arXiv preprint arXiv:1910.09036, 2019. 5", + "[13] Avishek Ghosh, Jichan Chung, Dong Yin, and Kannan Ramchandran. An efficient framework for clustered federated learning. Advances in Neural Information Processing Systems, 33:19586-19597, 2020. 1, 2", + "[14] Junfeng Guo, Yiming Li, Lixu Wang, Shu-Tao Xia, Heng Huang, Cong Liu, and Bo Li. Domain watermark: Effective and harmless dataset copyright protection is closed at hand. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 5", + "[15] Lan-Zhe Guo, Yi-Ge Zhang, Zhi-Fan Wu, Jie-Jing Shao, and Yu-Feng Li. Robust semi-supervised learning when not all classes have labels. Advances in Neural Information Processing Systems, 35:3305-3317, 2022. 2", + "[16] Kai Han, Sylvestre-Alvise Rebuffi, Sebastien Ehrhardt, Andrea Vedaldi, and Andrew Zisserman. Autonovel: Automatically discovering and learning novel visual categories. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6767-6781, 2021. 2, 5, 6, 8, 12, 14", + "[17] John A Hartigan and Manchek A Wong. Algorithm as 136: A k-means clustering algorithm. Journal of the royal statistical society. series c (applied statistics), 28(1):100-108, 1979. 5", + "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6", + "[19] Tzu-Ming Harry Hsu, Hang Qi, and Matthew Brown. Measuring the effects of non-identical data distribution for federated visual classification. arXiv preprint arXiv:1909.06335, 2019. 3, 6, 12", + "[20] Wenke Huang, Mang Ye, and Bo Du. Learn from others and be yourself in heterogeneous federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10143-10153, 2022. 1, 2", + "[21] Wenke Huang, Mang Ye, Zekun Shi, He Li, and Bo Du. Rethinking federated learning with domain shift: A prototype view. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16312-16322. IEEE, 2023. 16", + "[22] Liangze Jiang and Tao Lin. Test-time robust personalization for federated learning. arXiv preprint arXiv:2205.10920, 2022. 2", + "[23] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank Reddi, Sebastian Stich, and Ananda Theertha Suresh. Scaffold: Stochastic controlled averaging for federated learning. In International conference on machine learning, pages 5132-5143. PMLR, 2020. 7, 8, 12, 13", + "[24] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In" + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2437", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the IEEE international conference on computer vision workshops, pages 554-561, 2013. 6", + "[25] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6", + "[26] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 6", + "[27] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3, 2015. 6", + "[28] Junnan Li, Pan Zhou, Caiming Xiong, and Steven Hoi. Prototypical contrastive learning of unsupervised representations. In International Conference on Learning Representations, 2020. 4", + "[29] Qinbin Li, Bingsheng He, and Dawn Song. Model-contrastive federated learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10713–10722, 2021. 7, 8, 12, 13", + "[30] Qinbin Li, Zeyi Wen, Zhaomin Wu, Sixu Hu, Naibo Wang, Yuan Li, Xu Liu, and Bingsheng He. A survey on federated learning systems: Vision, hype and reality for data privacy and protection. IEEE Transactions on Knowledge and Data Engineering, 2021. 1, 2", + "[31] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. Proceedings of Machine learning and systems, 2:429-450, 2020. 7, 8, 12, 13", + "[32] Wenbin Li, Zhichen Fan, Jing Huo, and Yang Gao. Modeling inter-class and intra-class constraints in novel class discovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 4, 5, 6, 13", + "[33] Xiaoxiao Liang, Yiqun Lin, Huazhu Fu, Lei Zhu, and Xiaomeng Li. Rscfed: Random sampling consensus federated semi-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10154-10163, 2022. 2", + "[34] Chenxi Liu, Lixu Wang, Lingjuan Lyu, Chen Sun, Xiao Wang, and Qi Zhu. Deja vu: Continual model generalization for unseen domains. In The Eleventh International Conference on Learning Representations, 2022. 4", + "[35] Ekdeep Lubana, Chi Ian Tang, Fahim Kawsar, Robert Dick, and Akhil Mathur. Orchestra: Unsupervised federated learning via globally consistent clustering. In International Conference on Machine Learning, pages 14461-14484. PMLR, 2022. 2, 7, 13", + "[36] Shijie Ma, Fei Zhu, Zhun Zhong, Wenzhuo Liu, Xu-Yao Zhang, and Cheng-Lin Liu. Happy: A debiased learning framework for continual generalized category discovery. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2, 6, 7, 13, 17", + "[37] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 1, 2, 3, 4, 12", + "[38] A Tuan Nguyen, Philip Torr, and Ser Nam Lim. Fedsr: A simple and effective domain generalization method for fed-" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "erated learning. Advances in Neural Information Processing Systems, 35:38831-38843, 2022. 2", + "[39] Keon-Hee Park, Hakyung Lee, Kyungwoo Song, and Gyeong-Moon Park. Online continuous generalized category discovery. In European Conference on Computer Vision, pages 53–69. Springer, 2024. 2", + "[40] Nan Pu, Zhun Zhong, Xinyuan Ji, and Nicu Sebe. Federated generalized category discovery. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2024. 2, 7, 13", + "[41] Daiqing Qi, Handong Zhao, and Sheng Li. Better generative replay for continual federated learning. In The Eleventh International Conference on Learning Representations. 1, 2", + "[42] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 2001-2010, 2017. 2, 4", + "[43] Nicola Rieke, Jonny Hancox, Wenqi Li, Fausto Miletari, Holger R Roth, Shadi Albarqouni, Spyridon Bakas, Mathieu N Galtier, Bennett A Landman, Klaus Maier-Hein, et al. The future of digital health with federated learning. NPJ digital medicine, 3(1):119, 2020. 1", + "[44] Subhankar Roy, Mingxuan Liu, Zhun Zhong, Nicu Sebe, and Elisa Ricci. Class-incremental novel class discovery. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIII, pages 317-333. Springer, 2022. 1, 2, 4, 5, 6, 7, 8, 13, 14, 15", + "[45] Erich Schubert and et. al. Dbscan revisited, revisited: why and how you should (still) use dbscan. ACM Transactions on Database Systems, 2017. 5", + "[46] Yiyou Sun and Yixuan Li. Opencon: Open-world contrastive learning. In Transactions on Machine Learning Research, 2022. 2, 6, 13", + "[47] Kiat Chuan Tan, Yulong Liu, Barbara Ambrose, Melissa Tulig, and Serge Belongie. The herbarium challenge 2019 dataset. arXiv preprint arXiv:1906.05372, 2019. 6", + "[48] Yue Tan, Guodong Long, Lu Liu, Tianyi Zhou, Qinghua Lu, Jing Jiang, and Chengqi Zhang. Fedproto: Federated prototype learning across heterogeneous clients. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8432-8440, 2022. 16", + "[49] Colin Troisemaine, Vincent Lemaire, Stéphane Gosselin, Alexandre Reiffers-Masson, Joachim Flocon-Cholet, and Sandrine Vaton. Novel class discovery: an introduction and key concepts. arXiv preprint arXiv:2302.12028, 2023. 2", + "[50] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Generalized category discovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7492-7501, 2022. 2, 5, 6, 8, 13", + "[51] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. Caltech-ucsd birds-200-2011. Technical Report CNS-TR-2011-001, California Institute of Technology, 2011. 6", + "[52] Lixu Wang, Shichao Xu, Xiao Wang, and Qi Zhu. Addressing class imbalance in federated learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10165–10173, 2021. 5" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2438", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Lixu Wang, Shichao Xu, Ruiqi Xu, Xiao Wang, and Qi Zhu. Non-transferable learning: A new approach for model ownership verification and applicability authorization. In International Conference on Learning Representations, 2021. 4", + "[54] Lixu Wang, Yang Zhao, Jiahua Dong, Ating Yin, Qinbin Li, Xiao Wang, Dusit Niyato, and Qi Zhu. Federated learning with new knowledge: Fundamentals, advances, and futures. arXiv preprint arXiv:2402.02268, 2024. 1, 2", + "[55] Jie Xu, Benjamin S Glicksberg, Chang Su, Peter Walker, Jiang Bian, and Fei Wang. Federated learning for healthcare informatics. Journal of Healthcare Informatics Research, 5: 1-19, 2021. 1", + "[56] Muli Yang, Yuehua Zhu, Jiaping Yu, Among Wu, and Cheng Deng. Divide and conquer: Compositional experts for generalized novel class discovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14268-14277, 2022. 2, 4", + "[57] Xufeng Yao, Yang Bai, Xinyun Zhang, Yuechen Zhang, Qi Sun, Ran Chen, Ruiyu Li, and Bei Yu. Pcl: Proxy-based contrastive learning for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7097-7107, 2022. 4", + "[58] Jaehong Yoon, Wonyong Jeong, Giwoong Lee, Eunho Yang, and Sung Ju Hwang. Federated continual learning with weighted inter-client transfer. In International Conference on Machine Learning, pages 12073-12086. PMLR, 2021. 1, 2", + "[59] Shuyang Yu, Junyuan Hong, Haotao Wang, Zhangyang Wang, and Jiayu Zhou. Turning the curse of heterogeneity in federated learning into a blessing for out-of-distribution detection. In The Eleventh International Conference on Learning Representations, 2023. 2", + "[60] Fengda Zhang, Kun Kuang, Zhaoyang You, Tao Shen, Jun Xiao, Yin Zhang, Chao Wu, Yueting Zhuang, and Xiaolin Li. Federated unsupervised representation learning. arXiv preprint arXiv:2010.08982, 2020. 2", + "[61] Jie Zhang, Xiaosong Ma, Song Guo, and Wenchao Xu. Towards unbiased training in federated open-world semi-supervised learning. ICML, 2023. 2, 7, 13", + "[62] Xinwei Zhang, Jianwen Jiang, Yutong Feng, Zhi-Fan Wu, Xibin Zhao, Hai Wan, Mingqian Tang, Rong Jin, and Yue Gao. Grow and merge: A unified framework for continuous categories discovery. In Advances in Neural Information Processing Systems, 2022. 1, 2, 4, 5, 6, 7, 13, 15", + "[63] Bingchen Zhao and Oisin Mac Aodha. Incremental generalized category discovery. arXiv preprint arXiv:2304.14310, 2023. 2", + "[64] Haiyang Zheng, Nan Pu, Wenjing Li, Nicu Sebe, and Zhun Zhong. Prototypical hash encoding for on-the-fly fine-grained category discovery. Advances in Neural Information Processing Systems, 37:101428-101455, 2025. 2", + "[65] Ligeng Zhu, Zhijian Liu, and Song Han. Deep leakage from gradients. Advances in neural information processing systems, 32, 2019. 17" + ], + "bbox": [ + 91, + 92, + 480, + 853 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2439", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Federated Continuous Category Discovery and Learning/7332a50f-74a4-4833-9616-ad03014bd631_model.json b/2025/Federated Continuous Category Discovery and Learning/7332a50f-74a4-4833-9616-ad03014bd631_model.json new file mode 100644 index 0000000000000000000000000000000000000000..99a9071892751a9bfe88cf97f9e5dc5af5f971ef --- /dev/null +++ b/2025/Federated Continuous Category Discovery and Learning/7332a50f-74a4-4833-9616-ad03014bd631_model.json @@ -0,0 +1,2301 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.213, + 0.131, + 0.787, + 0.154 + ], + "angle": 0, + "content": "Federated Continuous Category Discovery and Learning" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.18, + 0.918, + 0.217 + ], + "angle": 0, + "content": "Lixu Wang\\(^{1*}\\), Chenxi Liu\\(^{2*}\\), Junfeng Guo\\(^{2}\\), Qingqing Ye\\(^{3}\\), Heng Huang\\(^{2}\\), Haibo Hu\\(^{3}\\), Wei Dong\\(^{1\\dagger}\\), Nanyang Technological University. \\(^{2}\\)University of Maryland. \\(^{3}\\)The Hong Kong Polytechnic University." + }, + { + "type": "text", + "bbox": [ + 0.139, + 0.22, + 0.88, + 0.234 + ], + "angle": 0, + "content": "lixu.wang@ntu.edu.sg, {cxliu539, gjf2023}@umd.edu, qqing.ye@polyu.edu.hk, heng@umd.edu" + }, + { + "type": "text", + "bbox": [ + 0.328, + 0.238, + 0.694, + 0.251 + ], + "angle": 0, + "content": "haibo.hu@polyu.edu.hk, wei_dong@ntu.edu.sg" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.32, + 0.486, + 0.76 + ], + "angle": 0, + "content": "Federated Learning (FL) studies often assume a static data distribution, whereas real-world scenarios involve dynamic changes. To address this gap, we study Federated Continuous Category Discovery and Learning \\((FC^2 DL)\\), an essential yet underexplored problem that enables FL models to evolve continuously by discovering and learning novel data categories. The key challenge in \\(FC^2 DL\\) lies in merging and aligning categories discovered and learned by different clients, all while maintaining privacy. To tackle this, we propose the Global Prototype Alignment (GPA) framework. GPA first estimates the number of categories and constructs global prototypes by locating high-density regions in the representation space through bi-level clustering. To mitigate pseudo-label noise, GPA then employs a semantic-weighted loss to capture correlations between global prototypes and the novel data. This semantic weighting strategy is also used for contrastive loss, facilitating unsupervised novel-category learning. Besides, GPA incorporates a mixup-based mechanism for both data and models, effectively mitigating interference between known and novel categories while alleviating forgetting. Extensive experiments across multiple datasets demonstrate GPA's superiority over state-of-the-art baseline approaches. Notably, GPA achieves absolute gains of \\(5.7\\%\\) to \\(13.1\\%\\) in novel category accuracy while preserving known category performance. Furthermore, GPA is highly adaptable, equipping various mainstream FL algorithms with category discovery and learning capabilities, underscoring its potential for real-world deployment." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.789, + 0.222, + 0.805 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.814, + 0.485, + 0.875 + ], + "angle": 0, + "content": "Federated Learning [37] (FL) enables distributed devices to collaboratively train models while preserving data privacy, making it a compelling solution for privacy-sensitive applications. However, most existing FL studies assume a static" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.908, + 0.501 + ], + "angle": 0, + "content": "and well-characterized data distribution [4, 13, 20, 30]. In reality, data is inherently dynamic, continuously evolving as new information emerges [41, 54, 58]. This evolving nature presents significant challenges, particularly in enabling FL systems to adapt and remain effective in the presence of novel patterns and functionalities. To address this, FL must be equipped with mechanisms to identify and incorporate emerging functionalities, ensuring adaptability to evolving data distributions. This capability is particularly critical in domains such as healthcare, finance, transportation, and the Internet of Things. For example, FL-powered diagnostic models in healthcare must continuously adapt to the emergence of new diseases and rare conditions to maintain their effectiveness [43, 55]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.504, + 0.909, + 0.822 + ], + "angle": 0, + "content": "To enable FL systems to adapt to continuously emerging novel functionalities, we study Federated Continuous Category Discovery and Learning \\((\\mathsf{FC}^2\\mathsf{DL})\\). We consider an FL setting where multiple participants initially perform supervised training on a predefined category space, with non-independent and non-identically distributed (non-IID) data [37]. Once trained, the FL model is expected to generalize well to known categories. At this stage, the FL administrator typically stops the training to conserve resources and deploys the latest global model for inference [37]. After deployment, users provide unlabeled data and obtain model predictions. In real-world scenarios, participants frequently encounter data from previously unseen novel categories, which also follow a non-IID distribution. In this case, these novel data samples are often misclassified as existing known categories. \\(\\mathsf{FC}^2\\mathsf{DL}\\) aims to enable FL systems to discover and learn novel categories continuously, ensuring adaptability to evolving data distributions. At the same time, \\(\\mathsf{FC}^2\\mathsf{DL}\\) seeks to maintain strong performance on previously learned categories, striking a balance between new knowledge learning and retention of prior knowledge." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.909, + 0.903 + ], + "angle": 0, + "content": "A straightforward approach to tackling \\(\\mathbf{FC}^2\\mathbf{DL}\\) is to integrate FL with Novel Category Discovery and Learning (NCDL) [32, 44, 62]. However, this approach presents several fundamental challenges. First, effective clustering of novel data requires prior knowledge of the number of" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.352, + 0.901 + ], + "angle": 0, + "content": "*Equal contributions. †Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2429" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.304 + ], + "angle": 0, + "content": "novel categories. Conventional NCDL assumes this information is available, as model trainers typically have access to the training data and the specific application domain [32, 44, 46]. However, in FL, this assumption does not hold: the FL administrator lacks direct data access, and participants may not have the expertise to estimate task details [10, 37]. Second, in a non-IID setting, even if participants conduct NCDL locally, the discovered novel categories vary across clients, making it difficult to merge their label spaces into a unified one. Finally, most existing NCDL methods often rely on labeled data from known categories [7, 16, 49], which is challenging to obtain in practical FL systems. Many participants have limited storage, and newly joined clients may not have any labeled data [10]." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.308, + 0.483, + 0.7 + ], + "angle": 0, + "content": "To overcome these challenges, we propose Global Prototype Alignment (GPA), a novel framework for \\(\\mathbf{FC}^2\\mathbf{DL}\\), as illustrated in Figure 1. First, GPA needs participants to construct potential cluster centroids—synthetic virtual points that do not correspond to real data and are transmitted to the server only once. The server then applies Potential Prototype Merge (PPM) to localize high-density regions among all received centroids, estimating the number of novel categories and initializing the corresponding global prototypes. To facilitate effective learning, GPA introduces Semantic-Weighted Loss (SWL), which guides local training by enforcing alignment with global prototypes. This weighting strategy further enables Contrastive Weighted Loss (CWL), allowing GPA to learn from unlabeled novel-category data effectively. Besides, GPA incorporates a mixup mechanism at both the data and model levels to mitigate interference between known and novel categories while alleviating forgetting. Extensive experiments demonstrate that GPA achieves substantial improvements in novel-category accuracy for \\(\\mathbf{FC}^2\\mathbf{DL}\\), outperforming state-of-the-art methods with absolute gains of \\(5.7\\% - 13.1\\%\\) in a single-stage setting and \\(4.1\\% - 13.0\\%\\) across five stages, while maintaining near-zero forgetting on known categories. Furthermore, GPA is adaptable and can be seamlessly integrated into various mainstream FL algorithms, equipping them with strong NCDL capabilities. In summary, our contributions include:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.705, + 0.482, + 0.78 + ], + "angle": 0, + "content": "- We address an important and practical problem called Federated Continuous Category Discovery and Learning \\((\\mathbf{FC}^2\\mathbf{DL})\\), in which the biggest challenge is to merge and align novel categories of different participants that conform to a non-IID unsupervised setting." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.781, + 0.482, + 0.901 + ], + "angle": 0, + "content": "- We design Global Prototype Alignment (GPA) to tackle FC\\(^2\\)DL. GPA mainly consists of a novel category number estimation approach Potential Prototype Merge that can find high-density regions in unlabeled data, a Semantic-Weighted Loss and a Contrastive Weighted Loss that align the unlabeled novel data with global prototypes in local training, and a mixup mechanism for both the data and model that handle category interference and forgetting." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.705, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "- We validate GPA on multiple datasets and various \\(\\mathbf{FC}^2\\mathbf{DL}\\) scenarios. GPA achieves significant performance improvements over state-of-the-art baseline approaches when solving \\(\\mathbf{FC}^2\\mathbf{DL}\\). Moreover, except for the standard FedAvg [37] algorithm, GPA is highly effective in equipping many other FL algorithms with the capability of category discovery and learning." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.209, + 0.655, + 0.224 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.234, + 0.905, + 0.566 + ], + "angle": 0, + "content": "Federated Learning (FL) research [13, 20, 30] mainly focuses on static scenarios, assuming minimal changes in the FL system over time. However, this assumption is overly idealistic and fails to capture many real-world scenarios [54]. Although some studies [22, 38, 59] employ adversarial training to improve out-of-distribution generalization in FL models, and Federated Open-World and OpenSet Semi-Supervised Learning [33, 40, 61] aim to enable learning of unseen categories. However, these methods are restricted to supervised static FL settings, while existing unsupervised FL approaches [35, 60] focus solely on static scenarios, primarily training representation models, with downstream tasks requiring additional adaptation. Recent research has explored dynamic FL [10, 41, 58], where data distributions evolve over time; however, these studies remain confined to the supervised setting. To the best of our knowledge, only one study [40] addresses unseen category discovery in FL, but it depends on labeled data and lacks support for continuous learning. In summary, current FL research cannot discover and learn novel categories in unsupervised settings while mitigating catastrophic forgetting of previously learned knowledge." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.569, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Novel Category Discovery and Learning (NCDL) refers to the process by which ML models identify and learn previously unseen categories [16]. Existing NCDL approaches can be broadly categorized into label-available and label-unavailable settings. Label-available NCDL, also referred to as Open-World Semi-Supervised Learning (SSL) [2, 15], typically integrates labeled known-category data with unlabeled novel-category data, leveraging SSL [32, 56], clustering [50], or weakly-supervised learning [1, 6] to facilitate novel category discovery. In contrast, the less-explored label-unavailable NCDL addresses practical scenarios where labeled known-category data is unavailable due to privacy constraints or cost limitations [42]. These methods typically follow a two-stage process: an initial supervised learning phase with labeled known-category data, followed by pseudo-labeling or self-supervised learning [44, 62, 63] on unlabeled data containing both known and novel categories. Recent research focuses on mitigating catastrophic forgetting during the continuous execution of the second stage [3, 36, 39, 64]. \\(\\mathsf{FC}^{2}\\mathsf{DL}\\) falls into the label-unavailable NCDL category; however, data heterogeneity in FL significantly diminishes the effectiveness of existing" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2430" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.092, + 0.885, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.427, + 0.908, + 0.486 + ], + "angle": 0, + "content": "Figure 1. Overview of GPA for enabling FL models to discover and learn novel categories continuously. When the known category learning finishes and the novel category learning starts, FL participants apply Kmeans on their data to find local potential novel prototypes. These local prototypes are sent to the server only once and GPA applies PPM to merge these prototypes into global novel prototypes. Then each selected client leverages SWL, CWL, and Data Mixup to conduct local training and update the model via Model Mixup." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.51, + 0.221, + 0.525 + ], + "angle": 0, + "content": "NCDL approaches." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.541, + 0.229, + 0.557 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.567, + 0.435, + 0.584 + ], + "angle": 0, + "content": "3.1. Novel Category Discovery and Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.589, + 0.484, + 0.791 + ], + "angle": 0, + "content": "In NCDL, there is a labeled dataset \\(\\mathcal{D}^{\\mathrm{L}} = \\{(x_i^{\\mathrm{L}},y_i^{\\mathrm{L}})\\}_{i = 1}^{N^{\\mathrm{L}}}\\) and all labels belong to a one-hot label space of dimension \\(C^\\mathrm{L}\\), i.e., \\(\\pmb {y}_i^{\\mathrm{L}}\\in \\mathcal{V}^{\\mathrm{L}}\\) and the input marginal distribution is \\(\\mathcal{X}^{\\mathrm{L}}\\). Besides, there is a unlabeled dataset \\(\\mathcal{D}^{\\mathrm{U}} = \\{\\pmb{x}_i^{\\mathrm{U}}\\}_{i = 1}^{N^{\\mathrm{U}}}\\) which contains \\(C^\\mathrm{U}\\) unique categories in \\(\\mathcal{V}^{\\mathrm{U}}\\), and its input distribution is \\(\\mathcal{X}^{\\mathrm{U}}\\). It is assumed that the labels in \\(\\mathcal{V}^{\\mathrm{L}}\\) and \\(\\mathcal{V}^{\\mathrm{U}}\\) are disjoint, i.e., \\(\\mathcal{V}^{\\mathrm{L}}\\cap \\mathcal{V}^{\\mathrm{U}} = \\emptyset\\), which are usually called known categories and novel categories, respectively. The general objective of NCDL is to partition the novel-category data into different clusters with the help of \\(\\mathcal{D}^{\\mathrm{L}}\\). More specifically, we leverage a neural network \\(m = f_{\\theta}\\circ g_{\\omega}\\) which consists of a feature extractor \\(f\\) parameterized on \\(\\theta\\) and a classifier \\(g\\) parameterized on \\(\\omega\\), to achieve NCDL:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.804, + 0.484, + 0.84 + ], + "angle": 0, + "content": "\\[\n\\max _ {\\theta , \\omega} \\mathbb {E} _ {\\mathcal {X} ^ {\\mathrm {U}}, \\mathcal {Y} ^ {\\mathrm {U}}} \\left\\{\\mathbb {I} \\left[ g _ {\\omega} \\left(f _ {\\theta} \\left(\\boldsymbol {x} _ {i}\\right)\\right) = g _ {\\omega} \\left(f _ {\\theta} \\left(\\boldsymbol {x} _ {j}\\right)\\right) \\right] \\cdot \\mathbb {I} \\left[ \\boldsymbol {y} _ {i} = \\boldsymbol {y} _ {j} \\right] \\right\\}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathbb{I}[\\cdot ]\\) is an indicator function that if the inner condition is true, \\(\\mathbb{I}[\\mathrm{True}] = 1\\), otherwise \\(\\mathbb{I}[\\mathrm{False}] = 0\\). \\(\\pmb{y}_i,\\pmb{y}_j\\) are the ground truth labels of novel samples \\(\\pmb{x}_i,\\pmb{x}_j\\), though unavailable during NCDL." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.51, + 0.704, + 0.526 + ], + "angle": 0, + "content": "3.2. Federated Learning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.532, + 0.907, + 0.73 + ], + "angle": 0, + "content": "At the beginning of FL, we assume that a global model \\(m\\) is trained across \\(K\\) participants, and each participant holds its local dataset \\(\\mathcal{D}^k = \\{(x_i^k,y_i^k)\\}_{i = 1}^{N^k}\\). The data distributions of participants are non-IID from each other [19, 37]. In this case, the label spaces of different participants \\(\\mathcal{V}^k\\) are not the same, but they all belong to a unified space \\(\\mathcal{V}^{\\mathrm{L}}\\), i.e., \\(\\mathcal{V}^k\\subseteq \\mathcal{V}^{\\mathrm{L}}\\), which can be regarded as known categories in NCDL. For every global round, a random set of participants is selected as clients, namely \\(\\mathcal{K}\\), and they need to conduct local training and then upload the model updates to the server. We assume that the local training is conducted with the CrossEntropy loss and the server applies FedAvg [37] to aggregate model updates:" + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.743, + 0.907, + 0.761 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C E}} = - \\log g _ {\\omega} \\left(f _ {\\theta} \\left(\\boldsymbol {x} _ {i}\\right)\\right). \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.772, + 0.788, + 0.787 + ], + "angle": 0, + "content": "3.3. Problem Formulation of \\(\\mathsf{FC^2DL}\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.903 + ], + "angle": 0, + "content": "After training in a sufficient number of rounds, which we call Known Category Learning \\(\\mathcal{T}^{\\mathrm{L}}\\), the FL model \\(m\\) is hopefully able to perform well on \\(\\mathcal{V}^{\\mathrm{L}}\\). Then, the FL administrator may stop the training to save resources and deploy the latest model \\(m^{\\mathrm{L}}\\) to serve all participants in the FL framework. During the usage of \\(m^{\\mathrm{L}}\\), participants feed it with unlabeled testing data \\(\\pmb{x}^{\\mathrm{test}}\\) and obtain corresponding" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2431" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.482, + 0.198 + ], + "angle": 0, + "content": "inference results \\( g_{\\omega}(f_{\\theta}(\\boldsymbol{x}^{\\mathrm{test}})) \\). Intuitively, if \\( \\boldsymbol{x}^{\\mathrm{test}} \\) belongs to the known categories \\( (\\mathcal{V}^{\\mathrm{L}}) \\), the inference results should be mostly correct and accurate, i.e., \\( \\mathbb{I}[g_{\\omega}(f_{\\theta}(\\boldsymbol{x}^{\\mathrm{test}})) = \\boldsymbol{y}^{\\mathrm{test}}] = 1 \\). However, if \\( \\boldsymbol{x}^{\\mathrm{test}} \\) comes from unseen novel categories, the current model \\( m^{\\mathrm{L}} \\) will classify it into a certain known category, in other words, \\( 100\\% \\) misclassification \\( (\\mathbb{I}[g_{\\omega}(f_{\\theta}(\\boldsymbol{x}^{\\mathrm{test}})) = \\boldsymbol{y}^{\\mathrm{test}}] = 0) \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.199, + 0.483, + 0.442 + ], + "angle": 0, + "content": "Now let us more formally define the problem of \\(\\mathbf{FC}^2\\mathbf{DL}\\). Suppose that the FL training periodically steps into a stage called Novel Category Learning \\(\\mathcal{T}^{\\mathrm{U}}\\), and we assume that there are \\(K^{\\mathrm{U}}\\) participants available in total in \\(\\mathcal{T}^{\\mathrm{U}}\\). Note that \\(K^{\\mathrm{U}}\\) is uncertain and may be changing dynamically, especially since old participants are free to withdraw and new participants are welcome to join in at any time [10]. At this moment, we denote the local novel-category dataset of participant \\(k\\) as \\(\\mathcal{D}^{\\mathrm{U},k} = \\{\\pmb{x}_i^{\\mathrm{U},k}\\}_{i=1}^{N^{\\mathrm{U},k}}\\), screened from its received unlabeled testing dataset \\(\\mathcal{D}^{\\mathrm{test}}\\). Consistent with the non-IID setting, although \\(\\mathcal{D}^{\\mathrm{U},k}\\) has no labels, its groundtruth label space \\(\\mathcal{Y}^{\\mathrm{U},k}\\) is distinct across different participants but also belongs to a unified novel label space \\(\\mathcal{Y}^{\\mathrm{U}}\\). Our objective is to conduct effective NCDL that can cluster novel-category data into different clusters and preserve good performance on known categories as well:" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.45, + 0.483, + 0.523 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\max _ {\\theta , \\omega} \\left\\{\\sum_ {\\boldsymbol {x} \\in \\mathcal {D} ^ {\\mathrm {U}}} \\mathbb {I} \\left[ g _ {\\omega} \\left(f _ {\\theta} \\left(\\boldsymbol {x} _ {i}\\right)\\right) = g _ {\\omega} \\left(f _ {\\theta} \\left(\\boldsymbol {x} _ {j}\\right)\\right) \\right] \\cdot \\mathbb {I} \\left[ \\boldsymbol {y} _ {i} = \\boldsymbol {y} _ {j} \\right] \\right. \\tag {3} \\\\ \\left. \\right.\\left. + \\mathbb {E} _ {\\boldsymbol {x} \\sim \\mathcal {X} ^ {\\mathrm {L}}; \\boldsymbol {y} \\sim \\mathcal {Y} ^ {\\mathrm {L}}} \\mathbb {I} \\left[ g _ {\\omega} \\left(f _ {\\theta} (\\boldsymbol {x})\\right) = \\boldsymbol {y} \\right]\\right\\}. \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.532, + 0.377, + 0.548 + ], + "angle": 0, + "content": "4. Our Proposed GPA Framework" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.556, + 0.334, + 0.572 + ], + "angle": 0, + "content": "4.1. Known Category Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.578, + 0.483, + 0.734 + ], + "angle": 0, + "content": "To achieve better NCDL, model \\( m^{\\mathrm{L}} \\) should not only have a good performance on known categories but also extract high-quality representations for novel categories. To this end, we modify Prototype Contrastive Learning (PCL) [28, 34] to enhance the training in \\( \\mathcal{T}^{\\mathrm{L}} \\). More specifically, our modified PCL constructs positive pairs as a sample representation \\( z_{i} = f_{\\theta}(x_{i}) \\) and its corresponding category prototype \\( p_{y_i}^{\\mathrm{L}} \\). The negative pairs consist of two parts: all pairs between \\( z_{i} \\) and all prototypes \\( \\{p_c^{\\mathrm{L}}\\}_{c = 1}^{C^{\\mathrm{L}}} \\), and sample-wise pairs between \\( z_{i} \\) and samples belonging to other labels:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.741, + 0.488, + 0.801 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {P C L}} = - \\log \\frac {\\exp \\left(\\boldsymbol {z} _ {i} ^ {\\mathrm {L}} \\cdot \\boldsymbol {p} _ {\\boldsymbol {y} _ {i} ^ {\\mathrm {L}}} ^ {\\mathrm {L}} / \\tau\\right)}{\\sum_ {c = 1} ^ {C ^ {\\mathrm {L}}} \\exp \\left(\\boldsymbol {z} _ {i} ^ {\\mathrm {L}} \\cdot \\boldsymbol {p} _ {c} ^ {\\mathrm {L}} / \\tau\\right) + \\sum_ {j = 1} ^ {N _ {i} ^ {-}} \\exp \\left(\\boldsymbol {z} _ {i} ^ {\\mathrm {L}} \\cdot \\boldsymbol {z} _ {j} ^ {\\mathrm {L}} / \\tau\\right)} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\tau = 0.07\\) (consistent with [28]) is a temperature parameter and \\(N_{i}^{-}\\) is the set size of samples with different labels from \\(\\pmb{x}_i\\) in a data batch. Then we apply a combined loss \\(\\mathcal{L}_{\\mathcal{T}^{\\mathrm{L}}} = \\frac{1}{N^{\\mathrm{L},k}}\\sum_{\\mathcal{D}^{\\mathrm{L},k}}[\\mathcal{L}_{\\mathrm{CE}} + \\eta \\mathcal{L}_{\\mathrm{PCL}}]\\) to train \\(m\\) on known category data (\\(\\eta = 0.1\\); please refer to the Appendix B.4 for sensitivity analysis)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.091, + 0.9, + 0.108 + ], + "angle": 0, + "content": "4.2. Preparations before Novel Category Learning" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.113, + 0.776, + 0.128 + ], + "angle": 0, + "content": "4.2.1. Novel-Category Data Filtering." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.131, + 0.905, + 0.376 + ], + "angle": 0, + "content": "When stepping into Novel Category Learning \\(\\mathcal{T}^{\\mathrm{U}}\\), it is possible for every participant to continually receive data of both novel and known categories. Thus we first need to separate known and novel categories from \\(\\mathcal{D}^{\\mathrm{test},k}\\). Specifically, for each participant \\(k\\), we assume that there is a data memory \\(\\mathcal{M}^k\\) employed to store \\(\\mathcal{D}^{\\mathrm{U},k}\\) that is filtered out during model \\(m^{\\mathrm{L}}\\) usage*. We adopt a popular strategy [62] based on the representation distance between data samples and known-category prototypes. The traditional prototype [44, 62] is constructed by averaging category-wise representations, but this is infeasible in FL as the data is distributed among participants. We solve this problem by viewing the first \\(C^{\\mathrm{L}}\\) dimensions of neuron weights \\(\\omega\\) of classifier \\(g\\) as the prototypes \\(\\mathcal{P}^{\\mathrm{L}} = \\{\\pmb{p}_c^{\\mathrm{L}}\\}_{c=1}^{C^{\\mathrm{L}}}\\), if \\(g\\) is composed of only one single linear layer without bias [57]. The detailed filtering mechanism is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.517, + 0.379, + 0.905, + 0.434 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} ^ {\\mathrm {U}, k} = \\left\\{\\boldsymbol {x} \\in \\mathcal {D} ^ {\\text {t e s t}, k} \\Bigg | \\max \\left\\{\\frac {f _ {\\theta} (\\boldsymbol {x}) \\cdot \\boldsymbol {p} _ {c} ^ {\\mathrm {L}}}{\\| f _ {\\theta} (\\boldsymbol {x}) \\| \\| \\boldsymbol {p} _ {c} ^ {\\mathrm {L}} \\|} \\right\\} _ {c = 1} ^ {C ^ {\\mathrm {L}}} < r \\right\\}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.435, + 0.907, + 0.57 + ], + "angle": 0, + "content": "where \\( r = 0.5 \\) is a threshold. This implies that if a sample \\( x \\) has the largest similarity with known prototypes lower than \\( r \\), it should belong to novel categories. Note that such filtering cannot screen the known-category data completely. To prevent potential interference between known and novel categories, we design a data-level mixup strategy. As for the screened known-category data, we test GPA by applying Softmax-based pseudo labeling on them during \\( \\mathcal{T}^{\\mathrm{U}} \\) (detailed results can be found in the Appendix A.3)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.576, + 0.906, + 0.606 + ], + "angle": 0, + "content": "4.2.2. Challenges of Estimating the Number of Novel Categories in \\(\\mathsf{FC}^2\\mathsf{DL}\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.905, + 0.685 + ], + "angle": 0, + "content": "The goal of NCDL is to accurately classify novel data into different clusters, which requires prior knowledge of the cluster number. Most existing NCDL studies [32, 44, 56] directly assume this is known, but this assumption is unreasonable in \\(\\mathbf{FC}^2\\mathbf{DL}\\) due to the following special challenges:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.687, + 0.905, + 0.731 + ], + "angle": 0, + "content": "- FL system administrator cannot access any training data and cannot predict the dynamic changes over time in data or task distribution [10, 37]." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.732, + 0.905, + 0.791 + ], + "angle": 0, + "content": "- There are various overlapping degrees between ordered clustering distributions \\(\\left[c_{1}^{\\mathrm{U}, k}, \\cdots, c_{i}^{\\mathrm{U}, k}, \\cdots\\right]\\) over \\(\\mathcal{D}^{\\mathrm{U}, k}\\) among different participants when applying unsupervised clustering algorithms locally, i.e.," + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.687, + 0.905, + 0.791 + ], + "angle": 0, + "content": null + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.798, + 0.905, + 0.833 + ], + "angle": 0, + "content": "\\[\n\\mathcal {P} \\left(\\boldsymbol {c} _ {i} ^ {\\mathrm {U}, k} \\neq \\boldsymbol {c} _ {i} ^ {\\mathrm {U}, k ^ {\\prime}}\\right) > 0, \\mathcal {P} \\left(\\boldsymbol {c} _ {i} ^ {\\mathrm {U}, k} = \\boldsymbol {c} _ {j} ^ {\\mathrm {U}, k ^ {\\prime}}\\right) > 0, (i \\neq j). \\tag {6}\n\\]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.901 + ], + "angle": 0, + "content": "*Note that old participants can directly apply the data memory that was previously used for storing the known-category data to store the screened novel data, after removing the known-category data due to memory storage limitation or moving them to somewhere else for privacy or intellectual property protection [10, 42, 53]." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2432" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.484, + 0.228 + ], + "angle": 0, + "content": "- Existing novel-category number estimation methods [7, 16, 50] require labeled data, either from known categories or auxiliary novel categories. This requirement violates cost-cutting and privacy regulations [14] for old participants, while new participants do not have any labeled data [10]. Besides, these methods rely on sufficient observation of the novel data distribution, i.e., they need to access a large amount of novel-category data, which is also unreasonable in practice." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.235, + 0.325, + 0.251 + ], + "angle": 0, + "content": "4.2.3. Potential Prototype Merge." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.254, + 0.484, + 0.436 + ], + "angle": 0, + "content": "To achieve an accurate global novel-category number estimation, all participants need to identify potential local prototypes within their local data first. Unlike studies [1, 32] that employ Sinkhorn-Knopp-based clustering [12] to cluster unlabeled data into clusters with an equal size, we adopt standard Kmeans [17] to construct local prototypes. The reason is that local data is often class-imbalanced [52]. Specifically, each participant \\(k\\) feeds its local data \\(\\mathcal{D}^{\\mathrm{U},k}\\) into the model \\(m^{\\mathrm{L}}\\) to extract representations \\(\\boldsymbol {z}_i = f_\\theta (\\boldsymbol{x}_i^{\\mathrm{U},k})\\) and then employs Kmeans clustering to group these representations into \\(C^\\mathrm{L}\\) clusters. Subsequently, the centroids of these clusters are utilized as local prototypes:" + }, + { + "type": "equation", + "bbox": [ + 0.094, + 0.447, + 0.484, + 0.505 + ], + "angle": 0, + "content": "\\[\n\\mathcal {Z} ^ {\\mathrm {U}, k} = \\left\\{\\frac {\\sum_ {\\boldsymbol {z} \\in \\mathcal {Z} _ {c} ^ {\\mathrm {U} , k}} \\boldsymbol {z}}{| \\mathcal {Z} _ {c} ^ {\\mathrm {U} , k} |} \\right\\} _ {c = 1} ^ {C ^ {\\mathrm {L}}}, \\mathcal {Z} _ {c} ^ {\\mathrm {U}, k} = \\left\\{\\boldsymbol {z} \\mid \\operatorname {K M} (\\boldsymbol {z}) = \\boldsymbol {c} _ {c} ^ {\\mathrm {U}, k} \\right\\}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.506, + 0.484, + 0.627 + ], + "angle": 0, + "content": "where \\(\\mathrm{KM}(z)\\) outputs the cluster label of \\(z\\) when employing Kmeans clustering. Notably, these Kmeans centroids are synthetic points that achieve the minimum aggregated intra-cluster distance. In other words, they do not correspond to any real data. Moreover, these local prototypes are sent to the server only once. In these ways, privacy is protected. More discussions about why our design can protect privacy are provided in the Appendix B.6." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.627, + 0.484, + 0.855 + ], + "angle": 0, + "content": "The server shuffles all received local prototypes and constructs a prototype pool \\(\\mathcal{Z}^{\\mathrm{U}} = \\bigcup_{k=1}^{K^{\\mathrm{U}}} \\mathcal{Z}^{\\mathrm{U}, k}\\). When building PPM, we totally depart from conventional design principles in existing novel-category number estimation methods [7, 16, 50]. Instead, we leverage an intuition that data representations belonging to the same novel category appear in close proximity, forming high-density regions. Identifying these regions globally is achievable since we do not conduct any normalization when building local prototypes. Specifically, we modify DBSCAN [45] with a rising radius \\(\\epsilon\\) and a fixed minimum cluster data size \\(n_{\\mathrm{size}}\\) to find the high-density regions. \\(\\epsilon\\) controls how large the nearby area should be considered for a certain data point, while \\(n_{\\mathrm{size}}\\) defines the requirement that if a data point is a core point, it should be surrounded by \\(n_{\\mathrm{size}} - 1\\) points. To cover as" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.183 + ], + "angle": 0, + "content": "many high-density regions as possible, we try to use a loose setting with \\( n_{\\mathrm{size}} = 2 \\), which also considers the global imbalance case where minority categories correspond to very few prototypes globally [52]. As for \\( \\epsilon \\), we first calculate the pair-wise Euclidean distance range within the prototype pool \\( \\mathcal{Z}^{\\mathrm{U}} \\) and set a rising \\( \\epsilon \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.192, + 0.905, + 0.219 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\epsilon_ {e} = \\min D + e \\cdot (\\max D - \\min D) / E \\right\\} _ {e = 0} ^ {E}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.555, + 0.212, + 0.792, + 0.234 + ], + "angle": 0, + "content": "where \\(D = \\{d(\\pmb {z}_i^{\\mathrm{U}},\\pmb {z}_j^{\\mathrm{U}})\\}_{\\pmb {z}_i^{\\mathrm{U}},\\pmb {z}_j^{\\mathrm{U}}\\in \\mathcal{Z}^{\\mathrm{U}}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.243, + 0.907, + 0.379 + ], + "angle": 0, + "content": "where \\(d(\\cdot)\\) computes the Euclidean distance, and \\(E = 50\\) is the total rising steps. For each \\(\\epsilon_{e}\\), we conduct DBSCAN clustering on \\(\\mathcal{Z}^{\\mathrm{U}}\\) and record the unique resulting cluster number \\(\\tilde{c}_e^{\\mathrm{U}}\\). After conducting the final step, we choose the maximum cluster number as the estimated global novel-category number \\(\\widetilde{C}^{\\mathrm{U}} = \\max \\left\\{\\tilde{c}_{e}^{\\mathrm{U}}\\right\\}_{e=0}^{E}\\). We believe that PPM can work effectively with much fewer requirements than existing methods, thus tackling the aforementioned challenges of estimating the novel-category number in FC²DL." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.379, + 0.906, + 0.429 + ], + "angle": 0, + "content": "Then we apply Kmeans again on \\(\\mathcal{Z}^{\\mathrm{U}}\\) with the cluster number \\(\\widetilde{C}^{\\mathrm{U}}\\) and use the cluster centroids \\(\\hat{z}_c^{\\mathrm{U}}\\) to initialize the global novel prototypes as \\(\\mathcal{P}^{\\mathrm{U}} = \\{\\hat{z}_c^{\\mathrm{U}}\\}_{c=1}^{\\widetilde{C}^{\\mathrm{U}}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.436, + 0.902, + 0.452 + ], + "angle": 0, + "content": "4.3. Representation Learning for Novel Categories" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.458, + 0.733, + 0.473 + ], + "angle": 0, + "content": "4.3.1. Semantic-Weighted Loss" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.906, + 0.629 + ], + "angle": 0, + "content": "By viewing our problem here as weakly-supervised learning, we propose a novel Semantic-Weighted Loss (SWL) to align local data with the global prototypes \\(\\mathcal{P}^{\\mathrm{U}}\\) in representation space. Specifically, instead of assigning data samples with a single label, SWL considers the possible semantic relationships between data samples with all global prototypes, which are measured by the Softmax probability. Moreover, as \\(\\mathcal{P}^{\\mathrm{U}}\\) is obtained by Euclidean distance-based Kmeans, we minimize the Euclidean distance between samples and prototypes. In this case, SWL is shaped as:" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.638, + 0.905, + 0.7 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {S W L}} = \\sum_ {c = 1} ^ {\\tilde {C} ^ {\\mathrm {U}}} \\frac {\\exp \\left(f _ {\\theta} \\left(\\boldsymbol {x} _ {i}\\right) \\cdot \\hat {\\boldsymbol {z}} _ {c} ^ {\\mathrm {U}} / \\tau\\right)}{\\sum_ {j = 1} ^ {\\tilde {C} ^ {\\mathrm {U}}} \\exp \\left(f _ {\\theta} \\left(\\boldsymbol {x} _ {i}\\right) \\cdot \\hat {\\boldsymbol {z}} _ {j} ^ {\\mathrm {U}} / \\tau\\right)} d \\left(f _ {\\theta} \\left(\\boldsymbol {x} _ {i}\\right), \\hat {\\boldsymbol {z}} _ {c} ^ {\\mathrm {U}}\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.709, + 0.906, + 0.755 + ], + "angle": 0, + "content": "By weighting the Euclidean distance with the semantic similarity, SWL can guide the model to learn more distinguishable semantics and be more resistant to pseudo-label noise." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.761, + 0.748, + 0.776 + ], + "angle": 0, + "content": "4.3.2. Contrastive Weighted Loss" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.902 + ], + "angle": 0, + "content": "SWL's effectiveness relies on the correlations between known and novel categories captured by the model \\( m^{\\mathrm{L}} \\), but \\( m^{\\mathrm{L}} \\) is trained on labeled known-category data and thus naturally cannot extract high-quality representations for unlabeled novel-category data. To this end, we employ instance discrimination [5] to facilitate better feature extraction from unlabeled novel-category data. However, standard instance discrimination requires lengthy training to converge, which" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.862, + 0.483, + 0.901 + ], + "angle": 0, + "content": "\\(^{\\dagger}\\)We assume that \\(C^{\\mathrm{U}} < C^{\\mathrm{L}}\\) as previous studies [44, 62]. Clustering local novel data into \\(C^{\\mathrm{L}}\\) clusters can cover as complete novel label space as possible." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2433" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.168 + ], + "angle": 0, + "content": "lags far behind the pace of SWL. We found that enabling the training to focus on easy samples in the early stage that is more like a curriculum learning can accelerate the convergence process. Therefore, similar to SWL, we also use the semantic relationship to weight the InfoNCE loss," + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.177, + 0.397, + 0.193 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C W L}} = - \\Delta (\\boldsymbol {x} _ {i}, \\boldsymbol {x} _ {i}) \\log \\Delta (\\boldsymbol {x} _ {i}, \\boldsymbol {x} _ {i}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.196, + 0.484, + 0.247 + ], + "angle": 0, + "content": "where \\(\\Delta (\\pmb {x}_i,\\pmb {x}_i) = \\frac{\\exp\\left(f_\\theta(\\mathrm{A}(\\pmb{x}_i))\\cdot f_\\theta(\\mathrm{A}(\\pmb{x}_i)) / \\tau\\right)}{\\sum_{q = 1}^{N_B^k}\\exp\\left(f_\\theta(\\mathrm{A}(\\pmb{x}_i))\\cdot f_\\theta(\\mathrm{A}(\\pmb{x}_q)) / \\tau\\right)}\\) (10)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.248, + 0.484, + 0.277 + ], + "angle": 0, + "content": "where \\(A(\\cdot)\\) denotes using stochastic augmentation and \\(N_B^k\\) is the size of a mini-batch." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.286, + 0.222, + 0.3 + ], + "angle": 0, + "content": "4.3.3. Data Mixup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.304, + 0.484, + 0.426 + ], + "angle": 0, + "content": "To preserve the knowledge learned from labeled known categories and optimize for novel categories simultaneously, each client \\( k \\) applies a mixup mechanism on the data and model. The principle behind the data-level mixup is to handle the interference between known categories and novel categories in the novel data memory \\( \\mathcal{M}^k \\). For the specific design, each client \\( k \\) constructs a relatively high-confidence dataset \\( \\mathcal{D}^{\\mathrm{U},k} \\) from \\( \\mathcal{D}^{\\mathrm{U},k} \\) in terms of Softmax probability \\( \\widetilde{\\boldsymbol{y}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.434, + 0.488, + 0.51 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} ^ {\\mathrm {U} ^ {\\prime}, k} = \\left\\{\\boldsymbol {x} \\mid \\max \\left\\{\\frac {\\exp \\left(f _ {\\theta} \\left(\\boldsymbol {x} _ {i}\\right) \\cdot \\boldsymbol {p} _ {c} / \\tau\\right)}{C ^ {\\mathrm {L}} + \\widetilde {C} ^ {\\mathrm {U}}} \\right\\} _ {c = 1} ^ {C ^ {\\mathrm {L}} + \\widetilde {C} ^ {\\mathrm {U}}} > r ^ {\\prime} \\right\\}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.511, + 0.484, + 0.603 + ], + "angle": 0, + "content": "\\(r'\\) here is a threshold for selecting data to build \\(\\mathcal{D}^{\\mathrm{U}',k}\\), and we set it as 0.95 (please refer to the Appendix B.4 for the sensitivity analysis). Then for every instance \\(\\pmb{x}_i'\\) in \\(\\mathcal{D}^{\\mathrm{U}',k}\\), client \\(k\\) will randomly select another instance \\(\\pmb{x}_i\\) from the residual set \\(\\mathcal{D}^{\\mathrm{U},k} \\setminus \\mathcal{D}^{\\mathrm{U}',k}\\) and conduct the data-level mixup \\(\\pmb{x}_i^* = \\lambda \\pmb{x}_i' + (1 - \\lambda)\\pmb{x}_i\\), as well as the training as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.612, + 0.482, + 0.643 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {M I X}} = \\lambda \\mathcal {L} _ {\\mathrm {C E}} \\left(f _ {\\theta} \\left(\\boldsymbol {x} _ {i} ^ {*}\\right), \\widetilde {\\boldsymbol {y}} _ {i} ^ {\\prime}\\right) + (1 - \\lambda) \\mathcal {L} _ {\\mathrm {C E}} \\left(f _ {\\theta} \\left(\\boldsymbol {x} _ {i} ^ {*}\\right), \\widetilde {\\boldsymbol {y}} _ {i}\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.484, + 0.689 + ], + "angle": 0, + "content": "where \\(\\lambda \\sim \\mathrm{Beta}(\\lambda; 0.75, 0.75)\\) that is suggested in semi-supervised learning works [9]. With the data mixup, the optimization objective of the novel category learning is" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.697, + 0.498, + 0.749 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathcal {T} ^ {\\mathrm {U}}} = \\frac {1}{N ^ {\\mathrm {U} , k}} \\sum_ {\\mathcal {D} ^ {\\mathrm {U}, k}} \\left(\\mathcal {L} _ {\\mathrm {S W L}} + \\mathcal {L} _ {\\mathrm {C W L}}\\right) + \\frac {1}{\\left| \\mathcal {D} ^ {\\mathrm {U} ^ {\\prime} , k} \\right|} \\sum_ {\\mathcal {D} ^ {\\mathrm {U} ^ {\\prime}, k}} \\left(\\mathcal {L} _ {\\mathrm {M I X}} + \\mathcal {L} _ {\\mathrm {C E}}\\right). \\tag {13}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.755, + 0.233, + 0.77 + ], + "angle": 0, + "content": "4.3.4. Model Mixup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.774, + 0.484, + 0.835 + ], + "angle": 0, + "content": "In addition to data-level mixup, we also conduct a model-level mixup to compensate for the forgetting of known categories. Specifically, each client \\( k \\) updates \\( f_{\\theta}^{k} \\) for each local training round:" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.842, + 0.484, + 0.862 + ], + "angle": 0, + "content": "\\[\n\\theta^ {k} = \\beta \\theta^ {\\mathrm {L}} + (1 - \\beta) \\theta^ {k}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\beta\\) controls the merging speed, which is set as 0.99 (please see the Appendix B.4 for sensitivity analysis). Note" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.198 + ], + "angle": 0, + "content": "that such model mixup differs from the standard Exponential Moving Average as the first term \\(\\theta^{\\mathrm{L}}\\) here remains unchanged. As for the update of classifier \\(g_{\\omega}\\), the neuron weights corresponding to known categories are frozen during GPA, while the neuron weights corresponding to novel categories are updated freely through back-propagation on the local side and FedAvg on the global side." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.211, + 0.72, + 0.228 + ], + "angle": 0, + "content": "5. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.236, + 0.906, + 0.282 + ], + "angle": 0, + "content": "The main experimental setup, results, and analysis are introduced below. Please refer to Supplementary Materials for more details and all implementation source codes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.285, + 0.906, + 0.406 + ], + "angle": 0, + "content": "Datasets. CIFAR-100 [25] consists of 100 classes, and each class contains 600 images. Tiny-ImageNet [27] contains 200 classes with 500 images per class. ImageNet-S is a subset of ImageNet [8], including 100 classes and each class has 600 images. Three fine-grained datasets are also used: CUB200 [51] (200 classes), StanfordCars [24] (196 classes), and Herbarium 19 [47] (683 classes). These three datasets are usually class-imbalanced." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.906, + 0.697 + ], + "angle": 0, + "content": "Implementation details. The number of participants is 10, and 5 are randomly selected as the clients to conduct 10-epoch local training in every global round. The number of global rounds for both \\(\\mathcal{T}^{\\mathrm{L}}\\) and \\(\\mathcal{T}^{\\mathrm{U}}\\) is 30. To simulate the non-IID setting, the data distributions of all participants follow a Dirichlet distribution [19] with the parameter \\(\\alpha = 0.1\\) (experiments with other values can be found in the Appendix). We use ResNet-18 [18] as the feature extractor for CIFAR-100 and Tiny-IIImageNet, and ResNet-34 for ImageNet-S, CUB-200, StanfordCars, and Herbarium-19. For all datasets, we employ a bias-free linear classifier. Notably, these ResNets are trained from scratch, as our goal is to evaluate the direct impact of the proposed algorithms by preventing the influence of factors such as pretraining. Certainly, we conduct experiments on Vision Transformers [11] that have been pre-trained on large-scale datasets, and the results are provided in the Appendix B.8. The batch size is 128 and the SGD learning rate is 0.02. All experiments are conducted repeatedly with seeds 2023, 2024, and 2025." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.701, + 0.906, + 0.822 + ], + "angle": 0, + "content": "Evaluation metrics. For known categories, we use ground-truth labels to calculate the classification accuracy. As for novel categories, we follow [62] to adopt Hungarian clustering accuracy [26] as the metric. We also use the overall accuracy [44] calculated on known and novel categories to measure the model performance further. The accuracy forgetting on known categories after \\(\\mathcal{T}^{\\mathrm{U}}\\) of the experiments below can be found in the Appendix B.7." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Comparison baselines. For a fair comparison, we try our best to customize recent NCDL methods (AutoNovel [16], GM [62], iNCD [44], Happy [36], IIC [32], OpenCon [46]), novel-category estimation approaches (MACC [50], EMaCS [7]), a preliminary federated novel" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2434" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.08, + 0.907, + 0.121 + ], + "angle": 0, + "content": "Table 1. Performance comparison between GPA and other baseline methods in \\( {\\mathbf{{FC}}}^{2}\\mathbf{{DL}} \\) ,when there is only one novel category learning stage. Left precision is the mean accuracy, and the right is the variance of three runs. We bold and blue the best performance, and bold the second-best; same for all following tables." + }, + { + "type": "table", + "bbox": [ + 0.137, + 0.123, + 0.857, + 0.302 + ], + "angle": 0, + "content": "
MethodCIFAR-100Tiny-ImageNetImageNet-SCUB200
knownnovelallknownnovelallknownnovelallknownnovelall
AutoNovel67.3±1.033.4±2.460.6±0.552.3±0.425.0±0.749.6±0.455.1±021.0±0.748.3±040.0±018.1±2.337.5±0.5
GM71.4±029.9±0.963.1±057.3±019.4±1.253.6±055.8±019.9±0.948.6±038.5±2.219.7±5.537.3±2.0
iNCD69.8±0.235.2±15.562.9±1.256.4±025.3±2.153.3±055.5±0.224.6±9.149.3±0.839.2±1.020.0±2.037.2±1.2
Happy70.2±1.139.2±4.464.0±0.856.4±0.327.8±2.253.5±0.255.8±029.0±1.650.4±0.340.2±0.225.8±1.138.8±0.4
IIC60.8±2.333.1±1.855.3±2.442.8±0.723.3±0.640.9±0.542.3±1.519.1±8.237.7±2.438.7±1.118.5±0.936.6±0.4
OpenCon61.5±1.718.4±1.452.9±1.547.6±2.015.6±1.744.4±2.047.4±3.918.0±10.141.6±1.029.9±3.411.2±6.728.3±2.5
Orchestra66.6±0.533.1±0.860.0±0.253.4±0.324.7±1.150.1±0.152.8±0.321.5±0.348.4±040.0±018.5±1.037.5±0.1
SemiFL44.2±0.229.1±0.541.2±0.233.5±0.125.7±1.132.7±0.638.8±0.227.5±0.536.5±0.230.0±0.718.7±1.329.1±0.6
FedoSSL62.2±0.218.2±0.753.4±0.453.0±0.18.7±1.048.6±0.455.0±017.5±0.247.5±0.138.8±014.5±0.736.9±0.3
AGCL65.6±0.235.5±0.559.6±0.243.2±0.524.5±1.441.3±0.746.9±0.321.0±0.641.7±0.336.0±020.5±1.134.8±0.7
GPA71.9±052.3±1.068.0±0.457.6±039.1±1.955.8±0.755.8±037.4±1.052.1±0.440.3±0.231.5±0.739.4±0.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.316, + 0.904, + 0.343 + ], + "angle": 0, + "content": "Table 2. Performance comparison between GPA and other baseline methods in \\( {\\mathbf{{FC}}}^{2}\\mathbf{{DL}} \\) ,when there are five novel category learning stage. Left precision is the mean accuracy, and the right is the variance of three runs." + }, + { + "type": "table", + "bbox": [ + 0.09, + 0.346, + 0.907, + 0.572 + ], + "angle": 0, + "content": "
DatasetMethod1st Stage2nd Stage3rd Stage4th Stage5th Stage
knownnovelallknownnovelallknownnovelallknownnovelallknownnovelall
CIFAR-100GM72.9±1.340.2±2.469.1±1.571.5±0.837.8±1.464.4±0.970.7±1.031.4±2.759.5±1.369.5±0.627.9±1.955.0±1.168.8±2.526.0±2.051.7±1.7
iNCD71.5±0.544.4±1.768.3±0.668.9±1.241.6±1.563.2±0.866.0±1.334.5±1.457.0±1.165.7±0.833.0±1.754.3±0.965.0±1.031.2±1.951.5±1.3
Happy72.0±0.545.6±1.168.9±0.768.8±0.944.7±1.563.7±0.965.8±0.741.0±2.158.7±1.064.5±1.338.7±0.855.5±1.063.7±1.035.5±2.052.4±1.2
GPA74.0±057.8±1.472.1±0.673.5±0.656.5±1.369.9±0.773.3±0.553.0±1.067.5±0.473.0±0.951.7±1.665.6±1.172.5±1.048.2±2.362.8±1.5
T-ImageNetGM58.3±0.237.0±1.157.3±0.356.6±0.423.4±3.453.6±0.855.8±1.720.2±2.351.2±1.854.0±0.818.8±2.048.1±1.252.5±1.718.0±2.445.6±1.9
iNCD58.0±0.538.2±0.757.1±0.553.1±1.031.0±1.451.1±1.150.8±0.925.0±1.947.4±1.147.7±2.623.2±1.443.6±2.945.3±1.122.1±1.840.7±1.3
Happy58.3±0.236.5±1.557.3±0.655.7±0.930.9±2.153.4±1.353.4±1.228.0±1.450.1±1.252.0±1.126.3±1.447.7±1.251.0±0.724.9±1.845.8±1.1
GPA58.5±0.248.2±1.658.0±0.658.0±0.543.5±2.256.7±0.957.3±0.739.5±1.255.0±0.957.0±0.638.0±2.153.8±1.056.8±1.136.9±2.552.8±1.5
ImageNet-SGM57.0±037.2±0.454.7±0.156.5±0.226.7±1.950.2±0.854.8±0.519.5±1.444.7±0.653.0±0.518.2±1.040.9±0.652.9±1.517.3±1.838.7±1.2
iNCD55.5±0.236.2±0.753.2±0.354.1±0.430.2±1.349.1±0.552.0±0.324.5±0.844.1±0.449.9±1.422.3±1.940.3±1.548.0±1.020.2±1.536.9±1.2
Happy56.5±0.437.9±1.454.3±0.556.0±0.833.7±2.251.3±1.153.5±1.629.2±2.446.6±1.952.0±1.027.5±1.643.5±1.351.4±1.826.0±2.741.2±2.2
GPA57.2±046.0±1.155.9±0.457.0±0.542.2±0.753.9±0.556.4±0.937.0±2.050.9±1.455.7±1.235.1±1.548.5±1.255.5±1.332.9±3.046.5±1.9
CUB200GM42.5±0.328.9±0.941.9±0.442.0±0.523.7±1.040.3±0.641.0±0.719.5±1.438.2±0.940.2±0.518.1±1.236.5±0.738.5±0.617.7±1.634.3±0.9
iNCD41.0±0.534.0±0.940.7±0.539.8±0.624.6±1.838.4±1.037.0±1.020.0±2.034.8±1.236.5±1.118.9±1.733.6±1.333.8±0.917.9±1.430.6±1.0
Happy42.2±0.733.9±1.141.8±0.841.1±0.529.5±1.640.0±0.739.8±0.525.5±1.237.9±0.739.0±1.023.3±1.436.4±1.137.8±0.823.0±2.034.8±1.2
GPA42.8±0.238.0±0.942.6±0.342.5±0.735.3±1.441.8±0.942.2±0.731.0±1.940.7±1.041.8±0.429.4±1.239.7±0.641.5±0.828.8±1.739.0±1.0
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.483, + 0.675 + ], + "angle": 0, + "content": "category discovery work (AGCL [40]), federated self/semi-supervised learning methods (Orchestra [35], SemiFL [9], FedoSSL [61]), and popular FL algorithms (FedProx [31], SCAFFOLD [23], Moon [29]) reasonably in the setting of \\(\\mathbf{FC}^2\\mathbf{DL}\\), and compare them with our GPA." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.689, + 0.357, + 0.705 + ], + "angle": 0, + "content": "5.1. Effectiveness of GPA on \\(\\mathsf{FC^2DL}\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Comparison on one novel category learning stage. We first assume that there is only one \\(\\mathcal{T}^{\\mathrm{U}}\\) in \\(\\mathsf{FC^{2}DL}\\), which is consistent with conventional NCDL studies. For CIFAR100, Tiny-ImageNet, ImageNet-S, CUB200, and StanfordCars, we set the novel category number as 20 and 83 for Herbarium 19, while the rest are regarded as known categories. The experiment results are shown in Tables 1 and 5. The results demonstrate that our GPA framework always achieves the best performance on all metrics, including accuracy on known, novel, and all categories. The performance improvement on novel-category accuracy is particularly promising, ranging from \\(5.7\\%\\) on CUB200 to" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.598, + 0.907, + 0.628 + ], + "angle": 0, + "content": "\\(13.1\\%\\) on CIFAR-100 over the best baseline method, which demonstrates the effectiveness of GPA." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.637, + 0.907, + 0.865 + ], + "angle": 0, + "content": "Comparison on multiple novel category learning stages. As mentioned in related works [10], FL participants often collect unseen novel-category data continually. For such cases, we split each dataset into a labeled subset and 5 unlabeled subsets (each contains 8 novel categories) with disjoint categories, and conduct five \\(\\mathcal{T}^{\\mathrm{U}}\\) in sequence. For a fair comparison, we only compare GPA with GM [62], iNCD [44], and Happy [36] as only they have dedicated designs for multiple novel category learning stages. All experiments are conducted using the category number estimated by our PPM and the detailed results are shown in Table 2. We can observe that GPA achieves the best performance on all datasets, showing \\(4.1\\%\\) to \\(13.0\\%\\) improvement on novel categories compared with the best-performing approach across 1st and 5th novel category learning stages." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Comparison on estimating novel-category number. Furthermore, to validate the effectiveness of PPM, we com" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2435" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.09, + 0.908, + 0.118 + ], + "angle": 0, + "content": "Table 3. Performance comparison between PPM and other novel category number estimation approaches in various novel category number settings. The experiments here show that PPM can consistently provide the most accurate novel category number estimation." + }, + { + "type": "table", + "bbox": [ + 0.138, + 0.12, + 0.86, + 0.186 + ], + "angle": 0, + "content": "
DatasetCIFAR-100Tiny-ImageNetImageNet-SCUB200StanfordCarsHerbarium19Average\nError (%)
Class #5101520253010203040505101520102030401020304020406080
MACC4566141568122228466144613186122024918243048.9
EMaCS18173023121127452526582120341521353429194542362517384679.5
PPM4913192529819314054581217921324411202839194265868.1
" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.203, + 0.9, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.328, + 0.908, + 0.357 + ], + "angle": 0, + "content": "Figure 2. Performance evaluation of combining GPA with other FL algorithms. We assume that Kmeans can equip FedProx, SCAFFOLD, and Moon with a certain level of NCDL capability, while GPA can substantially exceed it. The FC\\(^2\\)DL setting is the same as Table 1." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.37, + 0.483, + 0.397 + ], + "angle": 0, + "content": "Table 4. Ablation study of major components in GPA. The \\(FC^2DL\\) setting is the same as Table 1." + }, + { + "type": "table", + "bbox": [ + 0.09, + 0.4, + 0.481, + 0.523 + ], + "angle": 0, + "content": "
MethodCIFAR-100Tiny-ImageNetImageNet-S
knownnovelallknownnovelallknownnovelall
GPA-w/o LpCL70.251.166.454.337.052.653.835.950.2
GPA-w/o LSWL71.837.364.957.631.155.055.830.250.7
GPA-w/o LCWL71.942.966.157.633.055.155.832.751.2
GPA-w/o LMix71.845.266.557.634.355.355.734.051.4
GPA-w/ MACC71.539.865.257.627.954.655.729.650.5
GPA-w/o M.Mix69.848.965.655.238.253.553.236.349.8
GPA71.952.368.057.639.155.855.837.452.1
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.554, + 0.483, + 0.676 + ], + "angle": 0, + "content": "pare it with state-of-the-art approaches (MACC [50] and EMaCS [7]) when estimating different novel category numbers in \\(\\mathsf{FC^2DL}\\). The detailed category number setups and experiment results are shown in Table 3. We can observe that our PPM provides the most accurate category number estimation for all datasets, with an average error rate of \\(8.1\\%\\), while the other approaches have approximately \\(50\\%\\) and even higher average error rate." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.682, + 0.483, + 0.819 + ], + "angle": 0, + "content": "Experiments with more FL algorithms. To further evaluate the applicability of GPA, we also carry out experiments combining GPA with FedProx [31], SCAFFOLD [23], and Moon [29]. Specifically, we assume that regular FL can possess a certain level of NCDL capability through unsupervised clustering (e.g., Kmeans) on novel-category data representations. The experiment results are shown in Figure 2. We can see that using GPA with these FL algorithms provides much better performance on novel categories." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "More experiments in the Appendix. 1) experiments in the centralized setting—Section B.1; 2) experiments with various degrees of heterogeneity in \\(\\mathbf{FC}^2\\mathbf{DL}\\)—Section B.2; 3) experiments with different data category partitions—Section B.3; 4) sensitivity analysis of hyper-parameters—" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.382, + 0.907, + 0.458 + ], + "angle": 0, + "content": "Section B.4; 5) experiments with a larger number of participants—Section B.5 6) experiments on pre-trained vision transformers—Section B.8; 7) experiments of launching data reconstruction attacks for verifying GPA's characteristic of privacy-preserving—Section B.6." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.471, + 0.667, + 0.487 + ], + "angle": 0, + "content": "5.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.494, + 0.909, + 0.691 + ], + "angle": 0, + "content": "We investigate the importance of each component in GPA via an ablation study built-in \\(\\mathbf{FC}^2\\mathbf{DL}\\) with one novel category learning stage. As shown in Table 4, 'GPA-w/o \\(\\mathcal{L}_{\\mathrm{PCL}}\\) 'GPA-w/o \\(\\mathcal{L}_{\\mathrm{SWL}}\\) ,GPA-w/o \\(\\mathcal{L}_{\\mathrm{CWL}}\\) ,GPA-w/o \\(\\mathcal{L}_{\\mathrm{MIX}}\\) GPA-w/MACC, 'GPA-w/o M.Mix' denote the performance of GPA without using PCL in \\(\\mathcal{T}^{\\mathrm{L}}\\) , without using SWL (replacing it with a Binary CrossEntropy Loss [16]), without using CWL, without using data mixup, with MACC instead of PPM to estimate category number, and without applying model mixup (replacing it with distillation [44]), respectively. We can observe that these versions of GPA degrade evidently in performance, which shows the importance of all these modules in our approach." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.708, + 0.634, + 0.724 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this paper, we study an important and practical problem called Federated Continuous Category Discovery and Learning (FC\\(^2\\)DL), and propose Global Prototype Alignment (GPA) to help the FL model discover and learn novel categories. GPA can accurately estimate the novel category number, construct corresponding global prototypes, and guide the local training of different participants to tackle data heterogeneity. Extensive experiments demonstrate that GPA significantly outperforms state-of-the-art novel category learning methods and other potential solutions in various scenarios for the FC\\(^2\\)DL problem." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2436" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.268 + ], + "angle": 0, + "content": "This research is supported by the National Research Foundation, Singapore, and the Cyber Security Agency of Singapore under the National Cybersecurity R&D Programme and the CyberSG R&D Programme Office (Award CRPO-GC3-NTU-001) and NTU-NAP startup grant (024584-00001). Any opinions, findings, conclusions, or recommendations expressed in these materials are those of the author(s) and do not reflect the views of the National Research Foundation, Singapore, the Cyber Security Agency of Singapore, or the CyberSG R&D Programme Office." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.281, + 0.188, + 0.297 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.306, + 0.484, + 0.361 + ], + "angle": 0, + "content": "[1] Wenbin An, Feng Tian, Qinghua Zheng, Wei Ding, QianYing Wang, and Ping Chen. Generalized category discovery with decoupled prototypical network. arXiv preprint arXiv:2211.15115, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.362, + 0.484, + 0.404 + ], + "angle": 0, + "content": "[2] Kaidi Cao, Maria Brbic, and Jure Leskovec. Open-world semi-supervised learning. In International Conference on Learning Representations (ICLR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.406, + 0.484, + 0.462 + ], + "angle": 0, + "content": "[3] Fernando Julio Cendra, Bingchen Zhao, and Kai Han. Promptcd: Learning gaussian mixture prompt pool for continual category discovery. In European Conference on Computer Vision, pages 188-205. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.463, + 0.484, + 0.517 + ], + "angle": 0, + "content": "[4] Mingzhe Chen, Nir Shlezinger, H Vincent Poor, Yonina C Eldar, and Shuguang Cui. Communication-efficient federated learning. Proceedings of the National Academy of Sciences, 118(17):e2024789118, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.519, + 0.484, + 0.575 + ], + "angle": 0, + "content": "[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PmLR, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.577, + 0.484, + 0.646 + ], + "angle": 0, + "content": "[6] Haoang Chi, Feng Liu, Wenjing Yang, Long Lan, Tongliang Liu, Bo Han, Gang Niu, Mingyuan Zhou, and Masashi Sugiyama. Meta discovery: Learning to discover novel classes given very limited data. In International Conference on Learning Representations. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.647, + 0.484, + 0.716 + ], + "angle": 0, + "content": "[7] Florent Chiaroni, Jose Dolz, Ziko Imtiaz Masud, Amar Mitiche, and Ismail Ben Ayed. Parametric information maximization for generalized category discovery. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1729-1739, 2023. 2, 5, 6, 8, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.718, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.775, + 0.484, + 0.83 + ], + "angle": 0, + "content": "[9] Enmao Diao, Jie Ding, and Vahid Tarokh. Semifl: Semi-supervised federated learning for unlabeled clients with alternate training. Advances in Neural Information Processing Systems, 35:17871-17884, 2022. 6, 7, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[10] Jiahua Dong, Lixu Wang, Zhen Fang, Gan Sun, Shichao Xu, Xiao Wang, and Qi Zhu. Federated class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10164-10173, 2022. 2, 4, 5, 7, 12" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.306, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[12] Aude Geneva, Gabriel Dulac-Arnold, and Jean-Philippe Vert. Differentiable deep clustering with cluster size constraints. arXiv preprint arXiv:1910.09036, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[13] Avishek Ghosh, Jichan Chung, Dong Yin, and Kannan Ramchandran. An efficient framework for clustered federated learning. Advances in Neural Information Processing Systems, 33:19586-19597, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.263, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[14] Junfeng Guo, Yiming Li, Lixu Wang, Shu-Tao Xia, Heng Huang, Cong Liu, and Bo Li. Domain watermark: Effective and harmless dataset copyright protection is closed at hand. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[15] Lan-Zhe Guo, Yi-Ge Zhang, Zhi-Fan Wu, Jie-Jing Shao, and Yu-Feng Li. Robust semi-supervised learning when not all classes have labels. Advances in Neural Information Processing Systems, 35:3305-3317, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[16] Kai Han, Sylvestre-Alvise Rebuffi, Sebastien Ehrhardt, Andrea Vedaldi, and Andrew Zisserman. Autonovel: Automatically discovering and learning novel visual categories. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6767-6781, 2021. 2, 5, 6, 8, 12, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[17] John A Hartigan and Manchek A Wong. Algorithm as 136: A k-means clustering algorithm. Journal of the royal statistical society. series c (applied statistics), 28(1):100-108, 1979. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.518, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.575, + 0.905, + 0.629 + ], + "angle": 0, + "content": "[19] Tzu-Ming Harry Hsu, Hang Qi, and Matthew Brown. Measuring the effects of non-identical data distribution for federated visual classification. arXiv preprint arXiv:1909.06335, 2019. 3, 6, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[20] Wenke Huang, Mang Ye, and Bo Du. Learn from others and be yourself in heterogeneous federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10143-10153, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[21] Wenke Huang, Mang Ye, Zekun Shi, He Li, and Bo Du. Rethinking federated learning with domain shift: A prototype view. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16312-16322. IEEE, 2023. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.759, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[22] Liangze Jiang and Tao Lin. Test-time robust personalization for federated learning. arXiv preprint arXiv:2205.10920, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[23] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank Reddi, Sebastian Stich, and Ananda Theertha Suresh. Scaffold: Stochastic controlled averaging for federated learning. In International conference on machine learning, pages 5132-5143. PMLR, 2020. 7, 8, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[24] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2437" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Proceedings of the IEEE international conference on computer vision workshops, pages 554-561, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.483, + 0.15 + ], + "angle": 0, + "content": "[25] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.151, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[26] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.194, + 0.482, + 0.221 + ], + "angle": 0, + "content": "[27] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.223, + 0.483, + 0.277 + ], + "angle": 0, + "content": "[28] Junnan Li, Pan Zhou, Caiming Xiong, and Steven Hoi. Prototypical contrastive learning of unsupervised representations. In International Conference on Learning Representations, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.28, + 0.482, + 0.334 + ], + "angle": 0, + "content": "[29] Qinbin Li, Bingsheng He, and Dawn Song. Model-contrastive federated learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10713–10722, 2021. 7, 8, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.336, + 0.482, + 0.405 + ], + "angle": 0, + "content": "[30] Qinbin Li, Zeyi Wen, Zhaomin Wu, Sixu Hu, Naibo Wang, Yuan Li, Xu Liu, and Bingsheng He. A survey on federated learning systems: Vision, hype and reality for data privacy and protection. IEEE Transactions on Knowledge and Data Engineering, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.407, + 0.482, + 0.462 + ], + "angle": 0, + "content": "[31] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. Proceedings of Machine learning and systems, 2:429-450, 2020. 7, 8, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.464, + 0.482, + 0.532 + ], + "angle": 0, + "content": "[32] Wenbin Li, Zhichen Fan, Jing Huo, and Yang Gao. Modeling inter-class and intra-class constraints in novel class discovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 4, 5, 6, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.535, + 0.482, + 0.603 + ], + "angle": 0, + "content": "[33] Xiaoxiao Liang, Yiqun Lin, Huazhu Fu, Lei Zhu, and Xiaomeng Li. Rscfed: Random sampling consensus federated semi-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10154-10163, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.605, + 0.482, + 0.66 + ], + "angle": 0, + "content": "[34] Chenxi Liu, Lixu Wang, Lingjuan Lyu, Chen Sun, Xiao Wang, and Qi Zhu. Deja vu: Continual model generalization for unseen domains. In The Eleventh International Conference on Learning Representations, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.73 + ], + "angle": 0, + "content": "[35] Ekdeep Lubana, Chi Ian Tang, Fahim Kawsar, Robert Dick, and Akhil Mathur. Orchestra: Unsupervised federated learning via globally consistent clustering. In International Conference on Machine Learning, pages 14461-14484. PMLR, 2022. 2, 7, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[36] Shijie Ma, Fei Zhu, Zhun Zhong, Wenzhuo Liu, Xu-Yao Zhang, and Cheng-Lin Liu. Happy: A debiased learning framework for continual generalized category discovery. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2, 6, 7, 13, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[37] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 1, 2, 3, 4, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[38] A Tuan Nguyen, Philip Torr, and Ser Nam Lim. Fedsr: A simple and effective domain generalization method for fed-" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "erated learning. Advances in Neural Information Processing Systems, 35:38831-38843, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[39] Keon-Hee Park, Hakyung Lee, Kyungwoo Song, and Gyeong-Moon Park. Online continuous generalized category discovery. In European Conference on Computer Vision, pages 53–69. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.231 + ], + "angle": 0, + "content": "[40] Nan Pu, Zhun Zhong, Xinyuan Ji, and Nicu Sebe. Federated generalized category discovery. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2024. 2, 7, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[41] Daiqing Qi, Handong Zhao, and Sheng Li. Better generative replay for continual federated learning. In The Eleventh International Conference on Learning Representations. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.275, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[42] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 2001-2010, 2017. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.413 + ], + "angle": 0, + "content": "[43] Nicola Rieke, Jonny Hancox, Wenqi Li, Fausto Miletari, Holger R Roth, Shadi Albarqouni, Spyridon Bakas, Mathieu N Galtier, Bennett A Landman, Klaus Maier-Hein, et al. The future of digital health with federated learning. NPJ digital medicine, 3(1):119, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.495 + ], + "angle": 0, + "content": "[44] Subhankar Roy, Mingxuan Liu, Zhun Zhong, Nicu Sebe, and Elisa Ricci. Class-incremental novel class discovery. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIII, pages 317-333. Springer, 2022. 1, 2, 4, 5, 6, 7, 8, 13, 14, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[45] Erich Schubert and et. al. Dbscan revisited, revisited: why and how you should (still) use dbscan. ACM Transactions on Database Systems, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.579 + ], + "angle": 0, + "content": "[46] Yiyou Sun and Yixuan Li. Opencon: Open-world contrastive learning. In Transactions on Machine Learning Research, 2022. 2, 6, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.905, + 0.622 + ], + "angle": 0, + "content": "[47] Kiat Chuan Tan, Yulong Liu, Barbara Ambrose, Melissa Tulig, and Serge Belongie. The herbarium challenge 2019 dataset. arXiv preprint arXiv:1906.05372, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.907, + 0.691 + ], + "angle": 0, + "content": "[48] Yue Tan, Guodong Long, Lu Liu, Tianyi Zhou, Qinghua Lu, Jing Jiang, and Chengqi Zhang. Fedproto: Federated prototype learning across heterogeneous clients. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8432-8440, 2022. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.905, + 0.748 + ], + "angle": 0, + "content": "[49] Colin Troisemaine, Vincent Lemaire, Stéphane Gosselin, Alexandre Reiffers-Masson, Joachim Flocon-Cholet, and Sandrine Vaton. Novel class discovery: an introduction and key concepts. arXiv preprint arXiv:2302.12028, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.907, + 0.803 + ], + "angle": 0, + "content": "[50] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Generalized category discovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7492-7501, 2022. 2, 5, 6, 8, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[51] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. Caltech-ucsd birds-200-2011. Technical Report CNS-TR-2011-001, California Institute of Technology, 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.907, + 0.9 + ], + "angle": 0, + "content": "[52] Lixu Wang, Shichao Xu, Xiao Wang, and Qi Zhu. Addressing class imbalance in federated learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10165–10173, 2021. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2438" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[53] Lixu Wang, Shichao Xu, Ruiqi Xu, Xiao Wang, and Qi Zhu. Non-transferable learning: A new approach for model ownership verification and applicability authorization. In International Conference on Learning Representations, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[54] Lixu Wang, Yang Zhao, Jiahua Dong, Ating Yin, Qinbin Li, Xiao Wang, Dusit Niyato, and Qi Zhu. Federated learning with new knowledge: Fundamentals, advances, and futures. arXiv preprint arXiv:2402.02268, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[55] Jie Xu, Benjamin S Glicksberg, Chang Su, Peter Walker, Jiang Bian, and Fei Wang. Federated learning for healthcare informatics. Journal of Healthcare Informatics Research, 5: 1-19, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.331 + ], + "angle": 0, + "content": "[56] Muli Yang, Yuehua Zhu, Jiaping Yu, Among Wu, and Cheng Deng. Divide and conquer: Compositional experts for generalized novel class discovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14268-14277, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.333, + 0.482, + 0.402 + ], + "angle": 0, + "content": "[57] Xufeng Yao, Yang Bai, Xinyun Zhang, Yuechen Zhang, Qi Sun, Ran Chen, Ruiyu Li, and Bei Yu. Pcl: Proxy-based contrastive learning for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7097-7107, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.404, + 0.482, + 0.47 + ], + "angle": 0, + "content": "[58] Jaehong Yoon, Wonyong Jeong, Giwoong Lee, Eunho Yang, and Sung Ju Hwang. Federated continual learning with weighted inter-client transfer. In International Conference on Machine Learning, pages 12073-12086. PMLR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.474, + 0.482, + 0.543 + ], + "angle": 0, + "content": "[59] Shuyang Yu, Junyuan Hong, Haotao Wang, Zhangyang Wang, and Jiayu Zhou. Turning the curse of heterogeneity in federated learning into a blessing for out-of-distribution detection. In The Eleventh International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.545, + 0.482, + 0.599 + ], + "angle": 0, + "content": "[60] Fengda Zhang, Kun Kuang, Zhaoyang You, Tao Shen, Jun Xiao, Yin Zhang, Chao Wu, Yueting Zhuang, and Xiaolin Li. Federated unsupervised representation learning. arXiv preprint arXiv:2010.08982, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.602, + 0.482, + 0.642 + ], + "angle": 0, + "content": "[61] Jie Zhang, Xiaosong Ma, Song Guo, and Wenchao Xu. Towards unbiased training in federated open-world semi-supervised learning. ICML, 2023. 2, 7, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.644, + 0.482, + 0.713 + ], + "angle": 0, + "content": "[62] Xinwei Zhang, Jianwen Jiang, Yutong Feng, Zhi-Fan Wu, Xibin Zhao, Hai Wan, Mingqian Tang, Rong Jin, and Yue Gao. Grow and merge: A unified framework for continuous categories discovery. In Advances in Neural Information Processing Systems, 2022. 1, 2, 4, 5, 6, 7, 13, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.715, + 0.482, + 0.754 + ], + "angle": 0, + "content": "[63] Bingchen Zhao and Oisin Mac Aodha. Incremental generalized category discovery. arXiv preprint arXiv:2304.14310, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.757, + 0.482, + 0.812 + ], + "angle": 0, + "content": "[64] Haiyang Zheng, Nan Pu, Wenjing Li, Nicu Sebe, and Zhun Zhong. Prototypical hash encoding for on-the-fly fine-grained category discovery. Advances in Neural Information Processing Systems, 37:101428-101455, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.814, + 0.482, + 0.854 + ], + "angle": 0, + "content": "[65] Ligeng Zhu, Zhijian Liu, and Song Han. Deep leakage from gradients. Advances in neural information processing systems, 32, 2019. 17" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.854 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2439" + } + ] +] \ No newline at end of file diff --git a/2025/Federated Continuous Category Discovery and Learning/7332a50f-74a4-4833-9616-ad03014bd631_origin.pdf b/2025/Federated Continuous Category Discovery and Learning/7332a50f-74a4-4833-9616-ad03014bd631_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8c4a327844a08fac642f67d5cb5d0f748c8e2f77 --- /dev/null +++ b/2025/Federated Continuous Category Discovery and Learning/7332a50f-74a4-4833-9616-ad03014bd631_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ff50ff63951eef0a3c9bbe4941206068caab643cd8ee7b18a843930b04060ca2 +size 1438138 diff --git a/2025/Federated Continuous Category Discovery and Learning/full.md b/2025/Federated Continuous Category Discovery and Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..56f60258b73caaf00cdf3efa1a531f3518a408fb --- /dev/null +++ b/2025/Federated Continuous Category Discovery and Learning/full.md @@ -0,0 +1,330 @@ +# Federated Continuous Category Discovery and Learning + +Lixu Wang $^{1*}$ , Chenxi Liu $^{2*}$ , Junfeng Guo $^{2}$ , Qingqing Ye $^{3}$ , Heng Huang $^{2}$ , Haibo Hu $^{3}$ , Wei Dong $^{1\dagger}$ , Nanyang Technological University. $^{2}$ University of Maryland. $^{3}$ The Hong Kong Polytechnic University. + +lixu.wang@ntu.edu.sg, {cxliu539, gjf2023}@umd.edu, qqing.ye@polyu.edu.hk, heng@umd.edu + +haibo.hu@polyu.edu.hk, wei_dong@ntu.edu.sg + +# Abstract + +Federated Learning (FL) studies often assume a static data distribution, whereas real-world scenarios involve dynamic changes. To address this gap, we study Federated Continuous Category Discovery and Learning $(FC^2 DL)$ , an essential yet underexplored problem that enables FL models to evolve continuously by discovering and learning novel data categories. The key challenge in $FC^2 DL$ lies in merging and aligning categories discovered and learned by different clients, all while maintaining privacy. To tackle this, we propose the Global Prototype Alignment (GPA) framework. GPA first estimates the number of categories and constructs global prototypes by locating high-density regions in the representation space through bi-level clustering. To mitigate pseudo-label noise, GPA then employs a semantic-weighted loss to capture correlations between global prototypes and the novel data. This semantic weighting strategy is also used for contrastive loss, facilitating unsupervised novel-category learning. Besides, GPA incorporates a mixup-based mechanism for both data and models, effectively mitigating interference between known and novel categories while alleviating forgetting. Extensive experiments across multiple datasets demonstrate GPA's superiority over state-of-the-art baseline approaches. Notably, GPA achieves absolute gains of $5.7\%$ to $13.1\%$ in novel category accuracy while preserving known category performance. Furthermore, GPA is highly adaptable, equipping various mainstream FL algorithms with category discovery and learning capabilities, underscoring its potential for real-world deployment. + +# 1. Introduction + +Federated Learning [37] (FL) enables distributed devices to collaboratively train models while preserving data privacy, making it a compelling solution for privacy-sensitive applications. However, most existing FL studies assume a static + +and well-characterized data distribution [4, 13, 20, 30]. In reality, data is inherently dynamic, continuously evolving as new information emerges [41, 54, 58]. This evolving nature presents significant challenges, particularly in enabling FL systems to adapt and remain effective in the presence of novel patterns and functionalities. To address this, FL must be equipped with mechanisms to identify and incorporate emerging functionalities, ensuring adaptability to evolving data distributions. This capability is particularly critical in domains such as healthcare, finance, transportation, and the Internet of Things. For example, FL-powered diagnostic models in healthcare must continuously adapt to the emergence of new diseases and rare conditions to maintain their effectiveness [43, 55]. + +To enable FL systems to adapt to continuously emerging novel functionalities, we study Federated Continuous Category Discovery and Learning $(\mathsf{FC}^2\mathsf{DL})$ . We consider an FL setting where multiple participants initially perform supervised training on a predefined category space, with non-independent and non-identically distributed (non-IID) data [37]. Once trained, the FL model is expected to generalize well to known categories. At this stage, the FL administrator typically stops the training to conserve resources and deploys the latest global model for inference [37]. After deployment, users provide unlabeled data and obtain model predictions. In real-world scenarios, participants frequently encounter data from previously unseen novel categories, which also follow a non-IID distribution. In this case, these novel data samples are often misclassified as existing known categories. $\mathsf{FC}^2\mathsf{DL}$ aims to enable FL systems to discover and learn novel categories continuously, ensuring adaptability to evolving data distributions. At the same time, $\mathsf{FC}^2\mathsf{DL}$ seeks to maintain strong performance on previously learned categories, striking a balance between new knowledge learning and retention of prior knowledge. + +A straightforward approach to tackling $\mathbf{FC}^2\mathbf{DL}$ is to integrate FL with Novel Category Discovery and Learning (NCDL) [32, 44, 62]. However, this approach presents several fundamental challenges. First, effective clustering of novel data requires prior knowledge of the number of + +novel categories. Conventional NCDL assumes this information is available, as model trainers typically have access to the training data and the specific application domain [32, 44, 46]. However, in FL, this assumption does not hold: the FL administrator lacks direct data access, and participants may not have the expertise to estimate task details [10, 37]. Second, in a non-IID setting, even if participants conduct NCDL locally, the discovered novel categories vary across clients, making it difficult to merge their label spaces into a unified one. Finally, most existing NCDL methods often rely on labeled data from known categories [7, 16, 49], which is challenging to obtain in practical FL systems. Many participants have limited storage, and newly joined clients may not have any labeled data [10]. + +To overcome these challenges, we propose Global Prototype Alignment (GPA), a novel framework for $\mathbf{FC}^2\mathbf{DL}$ , as illustrated in Figure 1. First, GPA needs participants to construct potential cluster centroids—synthetic virtual points that do not correspond to real data and are transmitted to the server only once. The server then applies Potential Prototype Merge (PPM) to localize high-density regions among all received centroids, estimating the number of novel categories and initializing the corresponding global prototypes. To facilitate effective learning, GPA introduces Semantic-Weighted Loss (SWL), which guides local training by enforcing alignment with global prototypes. This weighting strategy further enables Contrastive Weighted Loss (CWL), allowing GPA to learn from unlabeled novel-category data effectively. Besides, GPA incorporates a mixup mechanism at both the data and model levels to mitigate interference between known and novel categories while alleviating forgetting. Extensive experiments demonstrate that GPA achieves substantial improvements in novel-category accuracy for $\mathbf{FC}^2\mathbf{DL}$ , outperforming state-of-the-art methods with absolute gains of $5.7\% - 13.1\%$ in a single-stage setting and $4.1\% - 13.0\%$ across five stages, while maintaining near-zero forgetting on known categories. Furthermore, GPA is adaptable and can be seamlessly integrated into various mainstream FL algorithms, equipping them with strong NCDL capabilities. In summary, our contributions include: + +- We address an important and practical problem called Federated Continuous Category Discovery and Learning $(\mathbf{FC}^2\mathbf{DL})$ , in which the biggest challenge is to merge and align novel categories of different participants that conform to a non-IID unsupervised setting. +- We design Global Prototype Alignment (GPA) to tackle FC $^2$ DL. GPA mainly consists of a novel category number estimation approach Potential Prototype Merge that can find high-density regions in unlabeled data, a Semantic-Weighted Loss and a Contrastive Weighted Loss that align the unlabeled novel data with global prototypes in local training, and a mixup mechanism for both the data and model that handle category interference and forgetting. + +- We validate GPA on multiple datasets and various $\mathbf{FC}^2\mathbf{DL}$ scenarios. GPA achieves significant performance improvements over state-of-the-art baseline approaches when solving $\mathbf{FC}^2\mathbf{DL}$ . Moreover, except for the standard FedAvg [37] algorithm, GPA is highly effective in equipping many other FL algorithms with the capability of category discovery and learning. + +# 2. Related Work + +Federated Learning (FL) research [13, 20, 30] mainly focuses on static scenarios, assuming minimal changes in the FL system over time. However, this assumption is overly idealistic and fails to capture many real-world scenarios [54]. Although some studies [22, 38, 59] employ adversarial training to improve out-of-distribution generalization in FL models, and Federated Open-World and OpenSet Semi-Supervised Learning [33, 40, 61] aim to enable learning of unseen categories. However, these methods are restricted to supervised static FL settings, while existing unsupervised FL approaches [35, 60] focus solely on static scenarios, primarily training representation models, with downstream tasks requiring additional adaptation. Recent research has explored dynamic FL [10, 41, 58], where data distributions evolve over time; however, these studies remain confined to the supervised setting. To the best of our knowledge, only one study [40] addresses unseen category discovery in FL, but it depends on labeled data and lacks support for continuous learning. In summary, current FL research cannot discover and learn novel categories in unsupervised settings while mitigating catastrophic forgetting of previously learned knowledge. + +Novel Category Discovery and Learning (NCDL) refers to the process by which ML models identify and learn previously unseen categories [16]. Existing NCDL approaches can be broadly categorized into label-available and label-unavailable settings. Label-available NCDL, also referred to as Open-World Semi-Supervised Learning (SSL) [2, 15], typically integrates labeled known-category data with unlabeled novel-category data, leveraging SSL [32, 56], clustering [50], or weakly-supervised learning [1, 6] to facilitate novel category discovery. In contrast, the less-explored label-unavailable NCDL addresses practical scenarios where labeled known-category data is unavailable due to privacy constraints or cost limitations [42]. These methods typically follow a two-stage process: an initial supervised learning phase with labeled known-category data, followed by pseudo-labeling or self-supervised learning [44, 62, 63] on unlabeled data containing both known and novel categories. Recent research focuses on mitigating catastrophic forgetting during the continuous execution of the second stage [3, 36, 39, 64]. $\mathsf{FC}^{2}\mathsf{DL}$ falls into the label-unavailable NCDL category; however, data heterogeneity in FL significantly diminishes the effectiveness of existing + +![](images/a0499a4261107fff32fe66416be83b1a8e52d44e4178c58f84213b2cfa922f92.jpg) +Figure 1. Overview of GPA for enabling FL models to discover and learn novel categories continuously. When the known category learning finishes and the novel category learning starts, FL participants apply Kmeans on their data to find local potential novel prototypes. These local prototypes are sent to the server only once and GPA applies PPM to merge these prototypes into global novel prototypes. Then each selected client leverages SWL, CWL, and Data Mixup to conduct local training and update the model via Model Mixup. + +NCDL approaches. + +# 3. Preliminaries + +# 3.1. Novel Category Discovery and Learning + +In NCDL, there is a labeled dataset $\mathcal{D}^{\mathrm{L}} = \{(x_i^{\mathrm{L}},y_i^{\mathrm{L}})\}_{i = 1}^{N^{\mathrm{L}}}$ and all labels belong to a one-hot label space of dimension $C^\mathrm{L}$ , i.e., $\pmb {y}_i^{\mathrm{L}}\in \mathcal{V}^{\mathrm{L}}$ and the input marginal distribution is $\mathcal{X}^{\mathrm{L}}$ . Besides, there is a unlabeled dataset $\mathcal{D}^{\mathrm{U}} = \{\pmb{x}_i^{\mathrm{U}}\}_{i = 1}^{N^{\mathrm{U}}}$ which contains $C^\mathrm{U}$ unique categories in $\mathcal{V}^{\mathrm{U}}$ , and its input distribution is $\mathcal{X}^{\mathrm{U}}$ . It is assumed that the labels in $\mathcal{V}^{\mathrm{L}}$ and $\mathcal{V}^{\mathrm{U}}$ are disjoint, i.e., $\mathcal{V}^{\mathrm{L}}\cap \mathcal{V}^{\mathrm{U}} = \emptyset$ , which are usually called known categories and novel categories, respectively. The general objective of NCDL is to partition the novel-category data into different clusters with the help of $\mathcal{D}^{\mathrm{L}}$ . More specifically, we leverage a neural network $m = f_{\theta}\circ g_{\omega}$ which consists of a feature extractor $f$ parameterized on $\theta$ and a classifier $g$ parameterized on $\omega$ , to achieve NCDL: + +$$ +\max _ {\theta , \omega} \mathbb {E} _ {\mathcal {X} ^ {\mathrm {U}}, \mathcal {Y} ^ {\mathrm {U}}} \left\{\mathbb {I} \left[ g _ {\omega} \left(f _ {\theta} \left(\boldsymbol {x} _ {i}\right)\right) = g _ {\omega} \left(f _ {\theta} \left(\boldsymbol {x} _ {j}\right)\right) \right] \cdot \mathbb {I} \left[ \boldsymbol {y} _ {i} = \boldsymbol {y} _ {j} \right] \right\}, \tag {1} +$$ + +where $\mathbb{I}[\cdot ]$ is an indicator function that if the inner condition is true, $\mathbb{I}[\mathrm{True}] = 1$ , otherwise $\mathbb{I}[\mathrm{False}] = 0$ . $\pmb{y}_i,\pmb{y}_j$ are the ground truth labels of novel samples $\pmb{x}_i,\pmb{x}_j$ , though unavailable during NCDL. + +# 3.2. Federated Learning + +At the beginning of FL, we assume that a global model $m$ is trained across $K$ participants, and each participant holds its local dataset $\mathcal{D}^k = \{(x_i^k,y_i^k)\}_{i = 1}^{N^k}$ . The data distributions of participants are non-IID from each other [19, 37]. In this case, the label spaces of different participants $\mathcal{V}^k$ are not the same, but they all belong to a unified space $\mathcal{V}^{\mathrm{L}}$ , i.e., $\mathcal{V}^k\subseteq \mathcal{V}^{\mathrm{L}}$ , which can be regarded as known categories in NCDL. For every global round, a random set of participants is selected as clients, namely $\mathcal{K}$ , and they need to conduct local training and then upload the model updates to the server. We assume that the local training is conducted with the CrossEntropy loss and the server applies FedAvg [37] to aggregate model updates: + +$$ +\mathcal {L} _ {\mathrm {C E}} = - \log g _ {\omega} \left(f _ {\theta} \left(\boldsymbol {x} _ {i}\right)\right). \tag {2} +$$ + +# 3.3. Problem Formulation of $\mathsf{FC^2DL}$ + +After training in a sufficient number of rounds, which we call Known Category Learning $\mathcal{T}^{\mathrm{L}}$ , the FL model $m$ is hopefully able to perform well on $\mathcal{V}^{\mathrm{L}}$ . Then, the FL administrator may stop the training to save resources and deploy the latest model $m^{\mathrm{L}}$ to serve all participants in the FL framework. During the usage of $m^{\mathrm{L}}$ , participants feed it with unlabeled testing data $\pmb{x}^{\mathrm{test}}$ and obtain corresponding + +inference results $g_{\omega}(f_{\theta}(\boldsymbol{x}^{\mathrm{test}}))$ . Intuitively, if $\boldsymbol{x}^{\mathrm{test}}$ belongs to the known categories $(\mathcal{V}^{\mathrm{L}})$ , the inference results should be mostly correct and accurate, i.e., $\mathbb{I}[g_{\omega}(f_{\theta}(\boldsymbol{x}^{\mathrm{test}})) = \boldsymbol{y}^{\mathrm{test}}] = 1$ . However, if $\boldsymbol{x}^{\mathrm{test}}$ comes from unseen novel categories, the current model $m^{\mathrm{L}}$ will classify it into a certain known category, in other words, $100\%$ misclassification $(\mathbb{I}[g_{\omega}(f_{\theta}(\boldsymbol{x}^{\mathrm{test}})) = \boldsymbol{y}^{\mathrm{test}}] = 0)$ . + +Now let us more formally define the problem of $\mathbf{FC}^2\mathbf{DL}$ . Suppose that the FL training periodically steps into a stage called Novel Category Learning $\mathcal{T}^{\mathrm{U}}$ , and we assume that there are $K^{\mathrm{U}}$ participants available in total in $\mathcal{T}^{\mathrm{U}}$ . Note that $K^{\mathrm{U}}$ is uncertain and may be changing dynamically, especially since old participants are free to withdraw and new participants are welcome to join in at any time [10]. At this moment, we denote the local novel-category dataset of participant $k$ as $\mathcal{D}^{\mathrm{U},k} = \{\pmb{x}_i^{\mathrm{U},k}\}_{i=1}^{N^{\mathrm{U},k}}$ , screened from its received unlabeled testing dataset $\mathcal{D}^{\mathrm{test}}$ . Consistent with the non-IID setting, although $\mathcal{D}^{\mathrm{U},k}$ has no labels, its groundtruth label space $\mathcal{Y}^{\mathrm{U},k}$ is distinct across different participants but also belongs to a unified novel label space $\mathcal{Y}^{\mathrm{U}}$ . Our objective is to conduct effective NCDL that can cluster novel-category data into different clusters and preserve good performance on known categories as well: + +$$ +\begin{array}{l} \max _ {\theta , \omega} \left\{\sum_ {\boldsymbol {x} \in \mathcal {D} ^ {\mathrm {U}}} \mathbb {I} \left[ g _ {\omega} \left(f _ {\theta} \left(\boldsymbol {x} _ {i}\right)\right) = g _ {\omega} \left(f _ {\theta} \left(\boldsymbol {x} _ {j}\right)\right) \right] \cdot \mathbb {I} \left[ \boldsymbol {y} _ {i} = \boldsymbol {y} _ {j} \right] \right. \tag {3} \\ \left. \right.\left. + \mathbb {E} _ {\boldsymbol {x} \sim \mathcal {X} ^ {\mathrm {L}}; \boldsymbol {y} \sim \mathcal {Y} ^ {\mathrm {L}}} \mathbb {I} \left[ g _ {\omega} \left(f _ {\theta} (\boldsymbol {x})\right) = \boldsymbol {y} \right]\right\}. \\ \end{array} +$$ + +# 4. Our Proposed GPA Framework + +# 4.1. Known Category Learning + +To achieve better NCDL, model $m^{\mathrm{L}}$ should not only have a good performance on known categories but also extract high-quality representations for novel categories. To this end, we modify Prototype Contrastive Learning (PCL) [28, 34] to enhance the training in $\mathcal{T}^{\mathrm{L}}$ . More specifically, our modified PCL constructs positive pairs as a sample representation $z_{i} = f_{\theta}(x_{i})$ and its corresponding category prototype $p_{y_i}^{\mathrm{L}}$ . The negative pairs consist of two parts: all pairs between $z_{i}$ and all prototypes $\{p_c^{\mathrm{L}}\}_{c = 1}^{C^{\mathrm{L}}}$ , and sample-wise pairs between $z_{i}$ and samples belonging to other labels: + +$$ +\mathcal {L} _ {\mathrm {P C L}} = - \log \frac {\exp \left(\boldsymbol {z} _ {i} ^ {\mathrm {L}} \cdot \boldsymbol {p} _ {\boldsymbol {y} _ {i} ^ {\mathrm {L}}} ^ {\mathrm {L}} / \tau\right)}{\sum_ {c = 1} ^ {C ^ {\mathrm {L}}} \exp \left(\boldsymbol {z} _ {i} ^ {\mathrm {L}} \cdot \boldsymbol {p} _ {c} ^ {\mathrm {L}} / \tau\right) + \sum_ {j = 1} ^ {N _ {i} ^ {-}} \exp \left(\boldsymbol {z} _ {i} ^ {\mathrm {L}} \cdot \boldsymbol {z} _ {j} ^ {\mathrm {L}} / \tau\right)} \tag {4} +$$ + +where $\tau = 0.07$ (consistent with [28]) is a temperature parameter and $N_{i}^{-}$ is the set size of samples with different labels from $\pmb{x}_i$ in a data batch. Then we apply a combined loss $\mathcal{L}_{\mathcal{T}^{\mathrm{L}}} = \frac{1}{N^{\mathrm{L},k}}\sum_{\mathcal{D}^{\mathrm{L},k}}[\mathcal{L}_{\mathrm{CE}} + \eta \mathcal{L}_{\mathrm{PCL}}]$ to train $m$ on known category data ( $\eta = 0.1$ ; please refer to the Appendix B.4 for sensitivity analysis). + +# 4.2. Preparations before Novel Category Learning + +# 4.2.1. Novel-Category Data Filtering. + +When stepping into Novel Category Learning $\mathcal{T}^{\mathrm{U}}$ , it is possible for every participant to continually receive data of both novel and known categories. Thus we first need to separate known and novel categories from $\mathcal{D}^{\mathrm{test},k}$ . Specifically, for each participant $k$ , we assume that there is a data memory $\mathcal{M}^k$ employed to store $\mathcal{D}^{\mathrm{U},k}$ that is filtered out during model $m^{\mathrm{L}}$ usage*. We adopt a popular strategy [62] based on the representation distance between data samples and known-category prototypes. The traditional prototype [44, 62] is constructed by averaging category-wise representations, but this is infeasible in FL as the data is distributed among participants. We solve this problem by viewing the first $C^{\mathrm{L}}$ dimensions of neuron weights $\omega$ of classifier $g$ as the prototypes $\mathcal{P}^{\mathrm{L}} = \{\pmb{p}_c^{\mathrm{L}}\}_{c=1}^{C^{\mathrm{L}}}$ , if $g$ is composed of only one single linear layer without bias [57]. The detailed filtering mechanism is as follows: + +$$ +\mathcal {D} ^ {\mathrm {U}, k} = \left\{\boldsymbol {x} \in \mathcal {D} ^ {\text {t e s t}, k} \Bigg | \max \left\{\frac {f _ {\theta} (\boldsymbol {x}) \cdot \boldsymbol {p} _ {c} ^ {\mathrm {L}}}{\| f _ {\theta} (\boldsymbol {x}) \| \| \boldsymbol {p} _ {c} ^ {\mathrm {L}} \|} \right\} _ {c = 1} ^ {C ^ {\mathrm {L}}} < r \right\}, \tag {5} +$$ + +where $r = 0.5$ is a threshold. This implies that if a sample $x$ has the largest similarity with known prototypes lower than $r$ , it should belong to novel categories. Note that such filtering cannot screen the known-category data completely. To prevent potential interference between known and novel categories, we design a data-level mixup strategy. As for the screened known-category data, we test GPA by applying Softmax-based pseudo labeling on them during $\mathcal{T}^{\mathrm{U}}$ (detailed results can be found in the Appendix A.3). + +# 4.2.2. Challenges of Estimating the Number of Novel Categories in $\mathsf{FC}^2\mathsf{DL}$ + +The goal of NCDL is to accurately classify novel data into different clusters, which requires prior knowledge of the cluster number. Most existing NCDL studies [32, 44, 56] directly assume this is known, but this assumption is unreasonable in $\mathbf{FC}^2\mathbf{DL}$ due to the following special challenges: + +- FL system administrator cannot access any training data and cannot predict the dynamic changes over time in data or task distribution [10, 37]. +- There are various overlapping degrees between ordered clustering distributions $\left[c_{1}^{\mathrm{U}, k}, \cdots, c_{i}^{\mathrm{U}, k}, \cdots\right]$ over $\mathcal{D}^{\mathrm{U}, k}$ among different participants when applying unsupervised clustering algorithms locally, i.e., + +$$ +\mathcal {P} \left(\boldsymbol {c} _ {i} ^ {\mathrm {U}, k} \neq \boldsymbol {c} _ {i} ^ {\mathrm {U}, k ^ {\prime}}\right) > 0, \mathcal {P} \left(\boldsymbol {c} _ {i} ^ {\mathrm {U}, k} = \boldsymbol {c} _ {j} ^ {\mathrm {U}, k ^ {\prime}}\right) > 0, (i \neq j). \tag {6} +$$ + +- Existing novel-category number estimation methods [7, 16, 50] require labeled data, either from known categories or auxiliary novel categories. This requirement violates cost-cutting and privacy regulations [14] for old participants, while new participants do not have any labeled data [10]. Besides, these methods rely on sufficient observation of the novel data distribution, i.e., they need to access a large amount of novel-category data, which is also unreasonable in practice. + +# 4.2.3. Potential Prototype Merge. + +To achieve an accurate global novel-category number estimation, all participants need to identify potential local prototypes within their local data first. Unlike studies [1, 32] that employ Sinkhorn-Knopp-based clustering [12] to cluster unlabeled data into clusters with an equal size, we adopt standard Kmeans [17] to construct local prototypes. The reason is that local data is often class-imbalanced [52]. Specifically, each participant $k$ feeds its local data $\mathcal{D}^{\mathrm{U},k}$ into the model $m^{\mathrm{L}}$ to extract representations $\boldsymbol {z}_i = f_\theta (\boldsymbol{x}_i^{\mathrm{U},k})$ and then employs Kmeans clustering to group these representations into $C^\mathrm{L}$ clusters. Subsequently, the centroids of these clusters are utilized as local prototypes: + +$$ +\mathcal {Z} ^ {\mathrm {U}, k} = \left\{\frac {\sum_ {\boldsymbol {z} \in \mathcal {Z} _ {c} ^ {\mathrm {U} , k}} \boldsymbol {z}}{| \mathcal {Z} _ {c} ^ {\mathrm {U} , k} |} \right\} _ {c = 1} ^ {C ^ {\mathrm {L}}}, \mathcal {Z} _ {c} ^ {\mathrm {U}, k} = \left\{\boldsymbol {z} \mid \operatorname {K M} (\boldsymbol {z}) = \boldsymbol {c} _ {c} ^ {\mathrm {U}, k} \right\}. \tag {7} +$$ + +where $\mathrm{KM}(z)$ outputs the cluster label of $z$ when employing Kmeans clustering. Notably, these Kmeans centroids are synthetic points that achieve the minimum aggregated intra-cluster distance. In other words, they do not correspond to any real data. Moreover, these local prototypes are sent to the server only once. In these ways, privacy is protected. More discussions about why our design can protect privacy are provided in the Appendix B.6. + +The server shuffles all received local prototypes and constructs a prototype pool $\mathcal{Z}^{\mathrm{U}} = \bigcup_{k=1}^{K^{\mathrm{U}}} \mathcal{Z}^{\mathrm{U}, k}$ . When building PPM, we totally depart from conventional design principles in existing novel-category number estimation methods [7, 16, 50]. Instead, we leverage an intuition that data representations belonging to the same novel category appear in close proximity, forming high-density regions. Identifying these regions globally is achievable since we do not conduct any normalization when building local prototypes. Specifically, we modify DBSCAN [45] with a rising radius $\epsilon$ and a fixed minimum cluster data size $n_{\mathrm{size}}$ to find the high-density regions. $\epsilon$ controls how large the nearby area should be considered for a certain data point, while $n_{\mathrm{size}}$ defines the requirement that if a data point is a core point, it should be surrounded by $n_{\mathrm{size}} - 1$ points. To cover as + +many high-density regions as possible, we try to use a loose setting with $n_{\mathrm{size}} = 2$ , which also considers the global imbalance case where minority categories correspond to very few prototypes globally [52]. As for $\epsilon$ , we first calculate the pair-wise Euclidean distance range within the prototype pool $\mathcal{Z}^{\mathrm{U}}$ and set a rising $\epsilon$ as: + +$$ +\left\{\epsilon_ {e} = \min D + e \cdot (\max D - \min D) / E \right\} _ {e = 0} ^ {E}, \tag {8} +$$ + +where $D = \{d(\pmb {z}_i^{\mathrm{U}},\pmb {z}_j^{\mathrm{U}})\}_{\pmb {z}_i^{\mathrm{U}},\pmb {z}_j^{\mathrm{U}}\in \mathcal{Z}^{\mathrm{U}}}$ + +where $d(\cdot)$ computes the Euclidean distance, and $E = 50$ is the total rising steps. For each $\epsilon_{e}$ , we conduct DBSCAN clustering on $\mathcal{Z}^{\mathrm{U}}$ and record the unique resulting cluster number $\tilde{c}_e^{\mathrm{U}}$ . After conducting the final step, we choose the maximum cluster number as the estimated global novel-category number $\widetilde{C}^{\mathrm{U}} = \max \left\{\tilde{c}_{e}^{\mathrm{U}}\right\}_{e=0}^{E}$ . We believe that PPM can work effectively with much fewer requirements than existing methods, thus tackling the aforementioned challenges of estimating the novel-category number in FC²DL. + +Then we apply Kmeans again on $\mathcal{Z}^{\mathrm{U}}$ with the cluster number $\widetilde{C}^{\mathrm{U}}$ and use the cluster centroids $\hat{z}_c^{\mathrm{U}}$ to initialize the global novel prototypes as $\mathcal{P}^{\mathrm{U}} = \{\hat{z}_c^{\mathrm{U}}\}_{c=1}^{\widetilde{C}^{\mathrm{U}}}$ . + +# 4.3. Representation Learning for Novel Categories + +# 4.3.1. Semantic-Weighted Loss + +By viewing our problem here as weakly-supervised learning, we propose a novel Semantic-Weighted Loss (SWL) to align local data with the global prototypes $\mathcal{P}^{\mathrm{U}}$ in representation space. Specifically, instead of assigning data samples with a single label, SWL considers the possible semantic relationships between data samples with all global prototypes, which are measured by the Softmax probability. Moreover, as $\mathcal{P}^{\mathrm{U}}$ is obtained by Euclidean distance-based Kmeans, we minimize the Euclidean distance between samples and prototypes. In this case, SWL is shaped as: + +$$ +\mathcal {L} _ {\mathrm {S W L}} = \sum_ {c = 1} ^ {\tilde {C} ^ {\mathrm {U}}} \frac {\exp \left(f _ {\theta} \left(\boldsymbol {x} _ {i}\right) \cdot \hat {\boldsymbol {z}} _ {c} ^ {\mathrm {U}} / \tau\right)}{\sum_ {j = 1} ^ {\tilde {C} ^ {\mathrm {U}}} \exp \left(f _ {\theta} \left(\boldsymbol {x} _ {i}\right) \cdot \hat {\boldsymbol {z}} _ {j} ^ {\mathrm {U}} / \tau\right)} d \left(f _ {\theta} \left(\boldsymbol {x} _ {i}\right), \hat {\boldsymbol {z}} _ {c} ^ {\mathrm {U}}\right). \tag {9} +$$ + +By weighting the Euclidean distance with the semantic similarity, SWL can guide the model to learn more distinguishable semantics and be more resistant to pseudo-label noise. + +# 4.3.2. Contrastive Weighted Loss + +SWL's effectiveness relies on the correlations between known and novel categories captured by the model $m^{\mathrm{L}}$ , but $m^{\mathrm{L}}$ is trained on labeled known-category data and thus naturally cannot extract high-quality representations for unlabeled novel-category data. To this end, we employ instance discrimination [5] to facilitate better feature extraction from unlabeled novel-category data. However, standard instance discrimination requires lengthy training to converge, which + +lags far behind the pace of SWL. We found that enabling the training to focus on easy samples in the early stage that is more like a curriculum learning can accelerate the convergence process. Therefore, similar to SWL, we also use the semantic relationship to weight the InfoNCE loss, + +$$ +\mathcal {L} _ {\mathrm {C W L}} = - \Delta (\boldsymbol {x} _ {i}, \boldsymbol {x} _ {i}) \log \Delta (\boldsymbol {x} _ {i}, \boldsymbol {x} _ {i}), +$$ + +where $\Delta (\pmb {x}_i,\pmb {x}_i) = \frac{\exp\left(f_\theta(\mathrm{A}(\pmb{x}_i))\cdot f_\theta(\mathrm{A}(\pmb{x}_i)) / \tau\right)}{\sum_{q = 1}^{N_B^k}\exp\left(f_\theta(\mathrm{A}(\pmb{x}_i))\cdot f_\theta(\mathrm{A}(\pmb{x}_q)) / \tau\right)}$ (10) + +where $A(\cdot)$ denotes using stochastic augmentation and $N_B^k$ is the size of a mini-batch. + +# 4.3.3. Data Mixup + +To preserve the knowledge learned from labeled known categories and optimize for novel categories simultaneously, each client $k$ applies a mixup mechanism on the data and model. The principle behind the data-level mixup is to handle the interference between known categories and novel categories in the novel data memory $\mathcal{M}^k$ . For the specific design, each client $k$ constructs a relatively high-confidence dataset $\mathcal{D}^{\mathrm{U},k}$ from $\mathcal{D}^{\mathrm{U},k}$ in terms of Softmax probability $\widetilde{\boldsymbol{y}}$ : + +$$ +\mathcal {D} ^ {\mathrm {U} ^ {\prime}, k} = \left\{\boldsymbol {x} \mid \max \left\{\frac {\exp \left(f _ {\theta} \left(\boldsymbol {x} _ {i}\right) \cdot \boldsymbol {p} _ {c} / \tau\right)}{C ^ {\mathrm {L}} + \widetilde {C} ^ {\mathrm {U}}} \right\} _ {c = 1} ^ {C ^ {\mathrm {L}} + \widetilde {C} ^ {\mathrm {U}}} > r ^ {\prime} \right\}. \tag {11} +$$ + +$r'$ here is a threshold for selecting data to build $\mathcal{D}^{\mathrm{U}',k}$ , and we set it as 0.95 (please refer to the Appendix B.4 for the sensitivity analysis). Then for every instance $\pmb{x}_i'$ in $\mathcal{D}^{\mathrm{U}',k}$ , client $k$ will randomly select another instance $\pmb{x}_i$ from the residual set $\mathcal{D}^{\mathrm{U},k} \setminus \mathcal{D}^{\mathrm{U}',k}$ and conduct the data-level mixup $\pmb{x}_i^* = \lambda \pmb{x}_i' + (1 - \lambda)\pmb{x}_i$ , as well as the training as follows: + +$$ +\mathcal {L} _ {\mathrm {M I X}} = \lambda \mathcal {L} _ {\mathrm {C E}} \left(f _ {\theta} \left(\boldsymbol {x} _ {i} ^ {*}\right), \widetilde {\boldsymbol {y}} _ {i} ^ {\prime}\right) + (1 - \lambda) \mathcal {L} _ {\mathrm {C E}} \left(f _ {\theta} \left(\boldsymbol {x} _ {i} ^ {*}\right), \widetilde {\boldsymbol {y}} _ {i}\right), \tag {12} +$$ + +where $\lambda \sim \mathrm{Beta}(\lambda; 0.75, 0.75)$ that is suggested in semi-supervised learning works [9]. With the data mixup, the optimization objective of the novel category learning is + +$$ +\mathcal {L} _ {\mathcal {T} ^ {\mathrm {U}}} = \frac {1}{N ^ {\mathrm {U} , k}} \sum_ {\mathcal {D} ^ {\mathrm {U}, k}} \left(\mathcal {L} _ {\mathrm {S W L}} + \mathcal {L} _ {\mathrm {C W L}}\right) + \frac {1}{\left| \mathcal {D} ^ {\mathrm {U} ^ {\prime} , k} \right|} \sum_ {\mathcal {D} ^ {\mathrm {U} ^ {\prime}, k}} \left(\mathcal {L} _ {\mathrm {M I X}} + \mathcal {L} _ {\mathrm {C E}}\right). \tag {13} +$$ + +# 4.3.4. Model Mixup + +In addition to data-level mixup, we also conduct a model-level mixup to compensate for the forgetting of known categories. Specifically, each client $k$ updates $f_{\theta}^{k}$ for each local training round: + +$$ +\theta^ {k} = \beta \theta^ {\mathrm {L}} + (1 - \beta) \theta^ {k}, \tag {14} +$$ + +where $\beta$ controls the merging speed, which is set as 0.99 (please see the Appendix B.4 for sensitivity analysis). Note + +that such model mixup differs from the standard Exponential Moving Average as the first term $\theta^{\mathrm{L}}$ here remains unchanged. As for the update of classifier $g_{\omega}$ , the neuron weights corresponding to known categories are frozen during GPA, while the neuron weights corresponding to novel categories are updated freely through back-propagation on the local side and FedAvg on the global side. + +# 5. Experimental Results + +The main experimental setup, results, and analysis are introduced below. Please refer to Supplementary Materials for more details and all implementation source codes. + +Datasets. CIFAR-100 [25] consists of 100 classes, and each class contains 600 images. Tiny-ImageNet [27] contains 200 classes with 500 images per class. ImageNet-S is a subset of ImageNet [8], including 100 classes and each class has 600 images. Three fine-grained datasets are also used: CUB200 [51] (200 classes), StanfordCars [24] (196 classes), and Herbarium 19 [47] (683 classes). These three datasets are usually class-imbalanced. + +Implementation details. The number of participants is 10, and 5 are randomly selected as the clients to conduct 10-epoch local training in every global round. The number of global rounds for both $\mathcal{T}^{\mathrm{L}}$ and $\mathcal{T}^{\mathrm{U}}$ is 30. To simulate the non-IID setting, the data distributions of all participants follow a Dirichlet distribution [19] with the parameter $\alpha = 0.1$ (experiments with other values can be found in the Appendix). We use ResNet-18 [18] as the feature extractor for CIFAR-100 and Tiny-IIImageNet, and ResNet-34 for ImageNet-S, CUB-200, StanfordCars, and Herbarium-19. For all datasets, we employ a bias-free linear classifier. Notably, these ResNets are trained from scratch, as our goal is to evaluate the direct impact of the proposed algorithms by preventing the influence of factors such as pretraining. Certainly, we conduct experiments on Vision Transformers [11] that have been pre-trained on large-scale datasets, and the results are provided in the Appendix B.8. The batch size is 128 and the SGD learning rate is 0.02. All experiments are conducted repeatedly with seeds 2023, 2024, and 2025. + +Evaluation metrics. For known categories, we use ground-truth labels to calculate the classification accuracy. As for novel categories, we follow [62] to adopt Hungarian clustering accuracy [26] as the metric. We also use the overall accuracy [44] calculated on known and novel categories to measure the model performance further. The accuracy forgetting on known categories after $\mathcal{T}^{\mathrm{U}}$ of the experiments below can be found in the Appendix B.7. + +Comparison baselines. For a fair comparison, we try our best to customize recent NCDL methods (AutoNovel [16], GM [62], iNCD [44], Happy [36], IIC [32], OpenCon [46]), novel-category estimation approaches (MACC [50], EMaCS [7]), a preliminary federated novel + +Table 1. Performance comparison between GPA and other baseline methods in ${\mathbf{{FC}}}^{2}\mathbf{{DL}}$ ,when there is only one novel category learning stage. Left precision is the mean accuracy, and the right is the variance of three runs. We bold and blue the best performance, and bold the second-best; same for all following tables. + +
MethodCIFAR-100Tiny-ImageNetImageNet-SCUB200
knownnovelallknownnovelallknownnovelallknownnovelall
AutoNovel67.3±1.033.4±2.460.6±0.552.3±0.425.0±0.749.6±0.455.1±021.0±0.748.3±040.0±018.1±2.337.5±0.5
GM71.4±029.9±0.963.1±057.3±019.4±1.253.6±055.8±019.9±0.948.6±038.5±2.219.7±5.537.3±2.0
iNCD69.8±0.235.2±15.562.9±1.256.4±025.3±2.153.3±055.5±0.224.6±9.149.3±0.839.2±1.020.0±2.037.2±1.2
Happy70.2±1.139.2±4.464.0±0.856.4±0.327.8±2.253.5±0.255.8±029.0±1.650.4±0.340.2±0.225.8±1.138.8±0.4
IIC60.8±2.333.1±1.855.3±2.442.8±0.723.3±0.640.9±0.542.3±1.519.1±8.237.7±2.438.7±1.118.5±0.936.6±0.4
OpenCon61.5±1.718.4±1.452.9±1.547.6±2.015.6±1.744.4±2.047.4±3.918.0±10.141.6±1.029.9±3.411.2±6.728.3±2.5
Orchestra66.6±0.533.1±0.860.0±0.253.4±0.324.7±1.150.1±0.152.8±0.321.5±0.348.4±040.0±018.5±1.037.5±0.1
SemiFL44.2±0.229.1±0.541.2±0.233.5±0.125.7±1.132.7±0.638.8±0.227.5±0.536.5±0.230.0±0.718.7±1.329.1±0.6
FedoSSL62.2±0.218.2±0.753.4±0.453.0±0.18.7±1.048.6±0.455.0±017.5±0.247.5±0.138.8±014.5±0.736.9±0.3
AGCL65.6±0.235.5±0.559.6±0.243.2±0.524.5±1.441.3±0.746.9±0.321.0±0.641.7±0.336.0±020.5±1.134.8±0.7
GPA71.9±052.3±1.068.0±0.457.6±039.1±1.955.8±0.755.8±037.4±1.052.1±0.440.3±0.231.5±0.739.4±0.3
+ +Table 2. Performance comparison between GPA and other baseline methods in ${\mathbf{{FC}}}^{2}\mathbf{{DL}}$ ,when there are five novel category learning stage. Left precision is the mean accuracy, and the right is the variance of three runs. + +
DatasetMethod1st Stage2nd Stage3rd Stage4th Stage5th Stage
knownnovelallknownnovelallknownnovelallknownnovelallknownnovelall
CIFAR-100GM72.9±1.340.2±2.469.1±1.571.5±0.837.8±1.464.4±0.970.7±1.031.4±2.759.5±1.369.5±0.627.9±1.955.0±1.168.8±2.526.0±2.051.7±1.7
iNCD71.5±0.544.4±1.768.3±0.668.9±1.241.6±1.563.2±0.866.0±1.334.5±1.457.0±1.165.7±0.833.0±1.754.3±0.965.0±1.031.2±1.951.5±1.3
Happy72.0±0.545.6±1.168.9±0.768.8±0.944.7±1.563.7±0.965.8±0.741.0±2.158.7±1.064.5±1.338.7±0.855.5±1.063.7±1.035.5±2.052.4±1.2
GPA74.0±057.8±1.472.1±0.673.5±0.656.5±1.369.9±0.773.3±0.553.0±1.067.5±0.473.0±0.951.7±1.665.6±1.172.5±1.048.2±2.362.8±1.5
T-ImageNetGM58.3±0.237.0±1.157.3±0.356.6±0.423.4±3.453.6±0.855.8±1.720.2±2.351.2±1.854.0±0.818.8±2.048.1±1.252.5±1.718.0±2.445.6±1.9
iNCD58.0±0.538.2±0.757.1±0.553.1±1.031.0±1.451.1±1.150.8±0.925.0±1.947.4±1.147.7±2.623.2±1.443.6±2.945.3±1.122.1±1.840.7±1.3
Happy58.3±0.236.5±1.557.3±0.655.7±0.930.9±2.153.4±1.353.4±1.228.0±1.450.1±1.252.0±1.126.3±1.447.7±1.251.0±0.724.9±1.845.8±1.1
GPA58.5±0.248.2±1.658.0±0.658.0±0.543.5±2.256.7±0.957.3±0.739.5±1.255.0±0.957.0±0.638.0±2.153.8±1.056.8±1.136.9±2.552.8±1.5
ImageNet-SGM57.0±037.2±0.454.7±0.156.5±0.226.7±1.950.2±0.854.8±0.519.5±1.444.7±0.653.0±0.518.2±1.040.9±0.652.9±1.517.3±1.838.7±1.2
iNCD55.5±0.236.2±0.753.2±0.354.1±0.430.2±1.349.1±0.552.0±0.324.5±0.844.1±0.449.9±1.422.3±1.940.3±1.548.0±1.020.2±1.536.9±1.2
Happy56.5±0.437.9±1.454.3±0.556.0±0.833.7±2.251.3±1.153.5±1.629.2±2.446.6±1.952.0±1.027.5±1.643.5±1.351.4±1.826.0±2.741.2±2.2
GPA57.2±046.0±1.155.9±0.457.0±0.542.2±0.753.9±0.556.4±0.937.0±2.050.9±1.455.7±1.235.1±1.548.5±1.255.5±1.332.9±3.046.5±1.9
CUB200GM42.5±0.328.9±0.941.9±0.442.0±0.523.7±1.040.3±0.641.0±0.719.5±1.438.2±0.940.2±0.518.1±1.236.5±0.738.5±0.617.7±1.634.3±0.9
iNCD41.0±0.534.0±0.940.7±0.539.8±0.624.6±1.838.4±1.037.0±1.020.0±2.034.8±1.236.5±1.118.9±1.733.6±1.333.8±0.917.9±1.430.6±1.0
Happy42.2±0.733.9±1.141.8±0.841.1±0.529.5±1.640.0±0.739.8±0.525.5±1.237.9±0.739.0±1.023.3±1.436.4±1.137.8±0.823.0±2.034.8±1.2
GPA42.8±0.238.0±0.942.6±0.342.5±0.735.3±1.441.8±0.942.2±0.731.0±1.940.7±1.041.8±0.429.4±1.239.7±0.641.5±0.828.8±1.739.0±1.0
+ +category discovery work (AGCL [40]), federated self/semi-supervised learning methods (Orchestra [35], SemiFL [9], FedoSSL [61]), and popular FL algorithms (FedProx [31], SCAFFOLD [23], Moon [29]) reasonably in the setting of $\mathbf{FC}^2\mathbf{DL}$ , and compare them with our GPA. + +# 5.1. Effectiveness of GPA on $\mathsf{FC^2DL}$ + +Comparison on one novel category learning stage. We first assume that there is only one $\mathcal{T}^{\mathrm{U}}$ in $\mathsf{FC^{2}DL}$ , which is consistent with conventional NCDL studies. For CIFAR100, Tiny-ImageNet, ImageNet-S, CUB200, and StanfordCars, we set the novel category number as 20 and 83 for Herbarium 19, while the rest are regarded as known categories. The experiment results are shown in Tables 1 and 5. The results demonstrate that our GPA framework always achieves the best performance on all metrics, including accuracy on known, novel, and all categories. The performance improvement on novel-category accuracy is particularly promising, ranging from $5.7\%$ on CUB200 to + +$13.1\%$ on CIFAR-100 over the best baseline method, which demonstrates the effectiveness of GPA. + +Comparison on multiple novel category learning stages. As mentioned in related works [10], FL participants often collect unseen novel-category data continually. For such cases, we split each dataset into a labeled subset and 5 unlabeled subsets (each contains 8 novel categories) with disjoint categories, and conduct five $\mathcal{T}^{\mathrm{U}}$ in sequence. For a fair comparison, we only compare GPA with GM [62], iNCD [44], and Happy [36] as only they have dedicated designs for multiple novel category learning stages. All experiments are conducted using the category number estimated by our PPM and the detailed results are shown in Table 2. We can observe that GPA achieves the best performance on all datasets, showing $4.1\%$ to $13.0\%$ improvement on novel categories compared with the best-performing approach across 1st and 5th novel category learning stages. + +Comparison on estimating novel-category number. Furthermore, to validate the effectiveness of PPM, we com + +Table 3. Performance comparison between PPM and other novel category number estimation approaches in various novel category number settings. The experiments here show that PPM can consistently provide the most accurate novel category number estimation. + +
DatasetCIFAR-100Tiny-ImageNetImageNet-SCUB200StanfordCarsHerbarium19Average +Error (%)
Class #5101520253010203040505101520102030401020304020406080
MACC4566141568122228466144613186122024918243048.9
EMaCS18173023121127452526582120341521353429194542362517384679.5
PPM4913192529819314054581217921324411202839194265868.1
+ +![](images/c39bf1041eda92e63752d7b86853d95ee88280decbe7c25064bccf7bb3be6f95.jpg) +Figure 2. Performance evaluation of combining GPA with other FL algorithms. We assume that Kmeans can equip FedProx, SCAFFOLD, and Moon with a certain level of NCDL capability, while GPA can substantially exceed it. The FC $^2$ DL setting is the same as Table 1. + +Table 4. Ablation study of major components in GPA. The $FC^2DL$ setting is the same as Table 1. + +
MethodCIFAR-100Tiny-ImageNetImageNet-S
knownnovelallknownnovelallknownnovelall
GPA-w/o LpCL70.251.166.454.337.052.653.835.950.2
GPA-w/o LSWL71.837.364.957.631.155.055.830.250.7
GPA-w/o LCWL71.942.966.157.633.055.155.832.751.2
GPA-w/o LMix71.845.266.557.634.355.355.734.051.4
GPA-w/ MACC71.539.865.257.627.954.655.729.650.5
GPA-w/o M.Mix69.848.965.655.238.253.553.236.349.8
GPA71.952.368.057.639.155.855.837.452.1
+ +pare it with state-of-the-art approaches (MACC [50] and EMaCS [7]) when estimating different novel category numbers in $\mathsf{FC^2DL}$ . The detailed category number setups and experiment results are shown in Table 3. We can observe that our PPM provides the most accurate category number estimation for all datasets, with an average error rate of $8.1\%$ , while the other approaches have approximately $50\%$ and even higher average error rate. + +Experiments with more FL algorithms. To further evaluate the applicability of GPA, we also carry out experiments combining GPA with FedProx [31], SCAFFOLD [23], and Moon [29]. Specifically, we assume that regular FL can possess a certain level of NCDL capability through unsupervised clustering (e.g., Kmeans) on novel-category data representations. The experiment results are shown in Figure 2. We can see that using GPA with these FL algorithms provides much better performance on novel categories. + +More experiments in the Appendix. 1) experiments in the centralized setting—Section B.1; 2) experiments with various degrees of heterogeneity in $\mathbf{FC}^2\mathbf{DL}$ —Section B.2; 3) experiments with different data category partitions—Section B.3; 4) sensitivity analysis of hyper-parameters— + +Section B.4; 5) experiments with a larger number of participants—Section B.5 6) experiments on pre-trained vision transformers—Section B.8; 7) experiments of launching data reconstruction attacks for verifying GPA's characteristic of privacy-preserving—Section B.6. + +# 5.2. Ablation Study + +We investigate the importance of each component in GPA via an ablation study built-in $\mathbf{FC}^2\mathbf{DL}$ with one novel category learning stage. As shown in Table 4, 'GPA-w/o $\mathcal{L}_{\mathrm{PCL}}$ 'GPA-w/o $\mathcal{L}_{\mathrm{SWL}}$ ,GPA-w/o $\mathcal{L}_{\mathrm{CWL}}$ ,GPA-w/o $\mathcal{L}_{\mathrm{MIX}}$ GPA-w/MACC, 'GPA-w/o M.Mix' denote the performance of GPA without using PCL in $\mathcal{T}^{\mathrm{L}}$ , without using SWL (replacing it with a Binary CrossEntropy Loss [16]), without using CWL, without using data mixup, with MACC instead of PPM to estimate category number, and without applying model mixup (replacing it with distillation [44]), respectively. We can observe that these versions of GPA degrade evidently in performance, which shows the importance of all these modules in our approach. + +# 6. Conclusion + +In this paper, we study an important and practical problem called Federated Continuous Category Discovery and Learning (FC $^2$ DL), and propose Global Prototype Alignment (GPA) to help the FL model discover and learn novel categories. GPA can accurately estimate the novel category number, construct corresponding global prototypes, and guide the local training of different participants to tackle data heterogeneity. Extensive experiments demonstrate that GPA significantly outperforms state-of-the-art novel category learning methods and other potential solutions in various scenarios for the FC $^2$ DL problem. + +# Acknowledgement + +This research is supported by the National Research Foundation, Singapore, and the Cyber Security Agency of Singapore under the National Cybersecurity R&D Programme and the CyberSG R&D Programme Office (Award CRPO-GC3-NTU-001) and NTU-NAP startup grant (024584-00001). Any opinions, findings, conclusions, or recommendations expressed in these materials are those of the author(s) and do not reflect the views of the National Research Foundation, Singapore, the Cyber Security Agency of Singapore, or the CyberSG R&D Programme Office. + +# References + +[1] Wenbin An, Feng Tian, Qinghua Zheng, Wei Ding, QianYing Wang, and Ping Chen. Generalized category discovery with decoupled prototypical network. arXiv preprint arXiv:2211.15115, 2022. 2, 5 +[2] Kaidi Cao, Maria Brbic, and Jure Leskovec. Open-world semi-supervised learning. In International Conference on Learning Representations (ICLR), 2022. 2 +[3] Fernando Julio Cendra, Bingchen Zhao, and Kai Han. Promptcd: Learning gaussian mixture prompt pool for continual category discovery. In European Conference on Computer Vision, pages 188-205. Springer, 2024. 2 +[4] Mingzhe Chen, Nir Shlezinger, H Vincent Poor, Yonina C Eldar, and Shuguang Cui. Communication-efficient federated learning. Proceedings of the National Academy of Sciences, 118(17):e2024789118, 2021. 1 +[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PmLR, 2020. 5 +[6] Haoang Chi, Feng Liu, Wenjing Yang, Long Lan, Tongliang Liu, Bo Han, Gang Niu, Mingyuan Zhou, and Masashi Sugiyama. Meta discovery: Learning to discover novel classes given very limited data. In International Conference on Learning Representations. 2 +[7] Florent Chiaroni, Jose Dolz, Ziko Imtiaz Masud, Amar Mitiche, and Ismail Ben Ayed. Parametric information maximization for generalized category discovery. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1729-1739, 2023. 2, 5, 6, 8, 13 +[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6 +[9] Enmao Diao, Jie Ding, and Vahid Tarokh. Semifl: Semi-supervised federated learning for unlabeled clients with alternate training. Advances in Neural Information Processing Systems, 35:17871-17884, 2022. 6, 7, 13 +[10] Jiahua Dong, Lixu Wang, Zhen Fang, Gan Sun, Shichao Xu, Xiao Wang, and Qi Zhu. Federated class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10164-10173, 2022. 2, 4, 5, 7, 12 + +[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2020. 6 +[12] Aude Geneva, Gabriel Dulac-Arnold, and Jean-Philippe Vert. Differentiable deep clustering with cluster size constraints. arXiv preprint arXiv:1910.09036, 2019. 5 +[13] Avishek Ghosh, Jichan Chung, Dong Yin, and Kannan Ramchandran. An efficient framework for clustered federated learning. Advances in Neural Information Processing Systems, 33:19586-19597, 2020. 1, 2 +[14] Junfeng Guo, Yiming Li, Lixu Wang, Shu-Tao Xia, Heng Huang, Cong Liu, and Bo Li. Domain watermark: Effective and harmless dataset copyright protection is closed at hand. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 5 +[15] Lan-Zhe Guo, Yi-Ge Zhang, Zhi-Fan Wu, Jie-Jing Shao, and Yu-Feng Li. Robust semi-supervised learning when not all classes have labels. Advances in Neural Information Processing Systems, 35:3305-3317, 2022. 2 +[16] Kai Han, Sylvestre-Alvise Rebuffi, Sebastien Ehrhardt, Andrea Vedaldi, and Andrew Zisserman. Autonovel: Automatically discovering and learning novel visual categories. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6767-6781, 2021. 2, 5, 6, 8, 12, 14 +[17] John A Hartigan and Manchek A Wong. Algorithm as 136: A k-means clustering algorithm. Journal of the royal statistical society. series c (applied statistics), 28(1):100-108, 1979. 5 +[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6 +[19] Tzu-Ming Harry Hsu, Hang Qi, and Matthew Brown. Measuring the effects of non-identical data distribution for federated visual classification. arXiv preprint arXiv:1909.06335, 2019. 3, 6, 12 +[20] Wenke Huang, Mang Ye, and Bo Du. Learn from others and be yourself in heterogeneous federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10143-10153, 2022. 1, 2 +[21] Wenke Huang, Mang Ye, Zekun Shi, He Li, and Bo Du. Rethinking federated learning with domain shift: A prototype view. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16312-16322. IEEE, 2023. 16 +[22] Liangze Jiang and Tao Lin. Test-time robust personalization for federated learning. arXiv preprint arXiv:2205.10920, 2022. 2 +[23] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank Reddi, Sebastian Stich, and Ananda Theertha Suresh. Scaffold: Stochastic controlled averaging for federated learning. In International conference on machine learning, pages 5132-5143. PMLR, 2020. 7, 8, 12, 13 +[24] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In + +Proceedings of the IEEE international conference on computer vision workshops, pages 554-561, 2013. 6 +[25] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6 +[26] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 6 +[27] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3, 2015. 6 +[28] Junnan Li, Pan Zhou, Caiming Xiong, and Steven Hoi. Prototypical contrastive learning of unsupervised representations. In International Conference on Learning Representations, 2020. 4 +[29] Qinbin Li, Bingsheng He, and Dawn Song. Model-contrastive federated learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10713–10722, 2021. 7, 8, 12, 13 +[30] Qinbin Li, Zeyi Wen, Zhaomin Wu, Sixu Hu, Naibo Wang, Yuan Li, Xu Liu, and Bingsheng He. A survey on federated learning systems: Vision, hype and reality for data privacy and protection. IEEE Transactions on Knowledge and Data Engineering, 2021. 1, 2 +[31] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. Proceedings of Machine learning and systems, 2:429-450, 2020. 7, 8, 12, 13 +[32] Wenbin Li, Zhichen Fan, Jing Huo, and Yang Gao. Modeling inter-class and intra-class constraints in novel class discovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 4, 5, 6, 13 +[33] Xiaoxiao Liang, Yiqun Lin, Huazhu Fu, Lei Zhu, and Xiaomeng Li. Rscfed: Random sampling consensus federated semi-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10154-10163, 2022. 2 +[34] Chenxi Liu, Lixu Wang, Lingjuan Lyu, Chen Sun, Xiao Wang, and Qi Zhu. Deja vu: Continual model generalization for unseen domains. In The Eleventh International Conference on Learning Representations, 2022. 4 +[35] Ekdeep Lubana, Chi Ian Tang, Fahim Kawsar, Robert Dick, and Akhil Mathur. Orchestra: Unsupervised federated learning via globally consistent clustering. In International Conference on Machine Learning, pages 14461-14484. PMLR, 2022. 2, 7, 13 +[36] Shijie Ma, Fei Zhu, Zhun Zhong, Wenzhuo Liu, Xu-Yao Zhang, and Cheng-Lin Liu. Happy: A debiased learning framework for continual generalized category discovery. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2, 6, 7, 13, 17 +[37] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 1, 2, 3, 4, 12 +[38] A Tuan Nguyen, Philip Torr, and Ser Nam Lim. Fedsr: A simple and effective domain generalization method for fed- + +erated learning. Advances in Neural Information Processing Systems, 35:38831-38843, 2022. 2 +[39] Keon-Hee Park, Hakyung Lee, Kyungwoo Song, and Gyeong-Moon Park. Online continuous generalized category discovery. In European Conference on Computer Vision, pages 53–69. Springer, 2024. 2 +[40] Nan Pu, Zhun Zhong, Xinyuan Ji, and Nicu Sebe. Federated generalized category discovery. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2024. 2, 7, 13 +[41] Daiqing Qi, Handong Zhao, and Sheng Li. Better generative replay for continual federated learning. In The Eleventh International Conference on Learning Representations. 1, 2 +[42] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 2001-2010, 2017. 2, 4 +[43] Nicola Rieke, Jonny Hancox, Wenqi Li, Fausto Miletari, Holger R Roth, Shadi Albarqouni, Spyridon Bakas, Mathieu N Galtier, Bennett A Landman, Klaus Maier-Hein, et al. The future of digital health with federated learning. NPJ digital medicine, 3(1):119, 2020. 1 +[44] Subhankar Roy, Mingxuan Liu, Zhun Zhong, Nicu Sebe, and Elisa Ricci. Class-incremental novel class discovery. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIII, pages 317-333. Springer, 2022. 1, 2, 4, 5, 6, 7, 8, 13, 14, 15 +[45] Erich Schubert and et. al. Dbscan revisited, revisited: why and how you should (still) use dbscan. ACM Transactions on Database Systems, 2017. 5 +[46] Yiyou Sun and Yixuan Li. Opencon: Open-world contrastive learning. In Transactions on Machine Learning Research, 2022. 2, 6, 13 +[47] Kiat Chuan Tan, Yulong Liu, Barbara Ambrose, Melissa Tulig, and Serge Belongie. The herbarium challenge 2019 dataset. arXiv preprint arXiv:1906.05372, 2019. 6 +[48] Yue Tan, Guodong Long, Lu Liu, Tianyi Zhou, Qinghua Lu, Jing Jiang, and Chengqi Zhang. Fedproto: Federated prototype learning across heterogeneous clients. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8432-8440, 2022. 16 +[49] Colin Troisemaine, Vincent Lemaire, Stéphane Gosselin, Alexandre Reiffers-Masson, Joachim Flocon-Cholet, and Sandrine Vaton. Novel class discovery: an introduction and key concepts. arXiv preprint arXiv:2302.12028, 2023. 2 +[50] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Generalized category discovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7492-7501, 2022. 2, 5, 6, 8, 13 +[51] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. Caltech-ucsd birds-200-2011. Technical Report CNS-TR-2011-001, California Institute of Technology, 2011. 6 +[52] Lixu Wang, Shichao Xu, Xiao Wang, and Qi Zhu. Addressing class imbalance in federated learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10165–10173, 2021. 5 + +[53] Lixu Wang, Shichao Xu, Ruiqi Xu, Xiao Wang, and Qi Zhu. Non-transferable learning: A new approach for model ownership verification and applicability authorization. In International Conference on Learning Representations, 2021. 4 +[54] Lixu Wang, Yang Zhao, Jiahua Dong, Ating Yin, Qinbin Li, Xiao Wang, Dusit Niyato, and Qi Zhu. Federated learning with new knowledge: Fundamentals, advances, and futures. arXiv preprint arXiv:2402.02268, 2024. 1, 2 +[55] Jie Xu, Benjamin S Glicksberg, Chang Su, Peter Walker, Jiang Bian, and Fei Wang. Federated learning for healthcare informatics. Journal of Healthcare Informatics Research, 5: 1-19, 2021. 1 +[56] Muli Yang, Yuehua Zhu, Jiaping Yu, Among Wu, and Cheng Deng. Divide and conquer: Compositional experts for generalized novel class discovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14268-14277, 2022. 2, 4 +[57] Xufeng Yao, Yang Bai, Xinyun Zhang, Yuechen Zhang, Qi Sun, Ran Chen, Ruiyu Li, and Bei Yu. Pcl: Proxy-based contrastive learning for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7097-7107, 2022. 4 +[58] Jaehong Yoon, Wonyong Jeong, Giwoong Lee, Eunho Yang, and Sung Ju Hwang. Federated continual learning with weighted inter-client transfer. In International Conference on Machine Learning, pages 12073-12086. PMLR, 2021. 1, 2 +[59] Shuyang Yu, Junyuan Hong, Haotao Wang, Zhangyang Wang, and Jiayu Zhou. Turning the curse of heterogeneity in federated learning into a blessing for out-of-distribution detection. In The Eleventh International Conference on Learning Representations, 2023. 2 +[60] Fengda Zhang, Kun Kuang, Zhaoyang You, Tao Shen, Jun Xiao, Yin Zhang, Chao Wu, Yueting Zhuang, and Xiaolin Li. Federated unsupervised representation learning. arXiv preprint arXiv:2010.08982, 2020. 2 +[61] Jie Zhang, Xiaosong Ma, Song Guo, and Wenchao Xu. Towards unbiased training in federated open-world semi-supervised learning. ICML, 2023. 2, 7, 13 +[62] Xinwei Zhang, Jianwen Jiang, Yutong Feng, Zhi-Fan Wu, Xibin Zhao, Hai Wan, Mingqian Tang, Rong Jin, and Yue Gao. Grow and merge: A unified framework for continuous categories discovery. In Advances in Neural Information Processing Systems, 2022. 1, 2, 4, 5, 6, 7, 13, 15 +[63] Bingchen Zhao and Oisin Mac Aodha. Incremental generalized category discovery. arXiv preprint arXiv:2304.14310, 2023. 2 +[64] Haiyang Zheng, Nan Pu, Wenjing Li, Nicu Sebe, and Zhun Zhong. Prototypical hash encoding for on-the-fly fine-grained category discovery. Advances in Neural Information Processing Systems, 37:101428-101455, 2025. 2 +[65] Ligeng Zhu, Zhijian Liu, and Song Han. Deep leakage from gradients. Advances in neural information processing systems, 32, 2019. 17 \ No newline at end of file diff --git a/2025/Federated Continuous Category Discovery and Learning/images.zip b/2025/Federated Continuous Category Discovery and Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..44b5ec27abe9841ea8ea4481e8bf6a7a90fb5bf7 --- /dev/null +++ b/2025/Federated Continuous Category Discovery and Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f6c656136b4afe73555f784213da369ac8bcc7480aefa8044221485254533638 +size 772592 diff --git a/2025/Federated Continuous Category Discovery and Learning/layout.json b/2025/Federated Continuous Category Discovery and Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a3559501a054c7ff234ddfcf7281fc64d8fe9361 --- /dev/null +++ b/2025/Federated Continuous Category Discovery and Learning/layout.json @@ -0,0 +1,10845 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 130, + 103, + 481, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 103, + 481, + 121 + ], + "spans": [ + { + "bbox": [ + 130, + 103, + 481, + 121 + ], + "type": "text", + "content": "Federated Continuous Category Discovery and Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 142, + 561, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 142, + 561, + 171 + ], + "spans": [ + { + "bbox": [ + 62, + 142, + 561, + 171 + ], + "type": "text", + "content": "Lixu Wang" + }, + { + "bbox": [ + 62, + 142, + 561, + 171 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 62, + 142, + 561, + 171 + ], + "type": "text", + "content": ", Chenxi Liu" + }, + { + "bbox": [ + 62, + 142, + 561, + 171 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 62, + 142, + 561, + 171 + ], + "type": "text", + "content": ", Junfeng Guo" + }, + { + "bbox": [ + 62, + 142, + 561, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 62, + 142, + 561, + 171 + ], + "type": "text", + "content": ", Qingqing Ye" + }, + { + "bbox": [ + 62, + 142, + 561, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 62, + 142, + 561, + 171 + ], + "type": "text", + "content": ", Heng Huang" + }, + { + "bbox": [ + 62, + 142, + 561, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 62, + 142, + 561, + 171 + ], + "type": "text", + "content": ", Haibo Hu" + }, + { + "bbox": [ + 62, + 142, + 561, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 62, + 142, + 561, + 171 + ], + "type": "text", + "content": ", Wei Dong" + }, + { + "bbox": [ + 62, + 142, + 561, + 171 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 62, + 142, + 561, + 171 + ], + "type": "text", + "content": ", Nanyang Technological University. " + }, + { + "bbox": [ + 62, + 142, + 561, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 62, + 142, + 561, + 171 + ], + "type": "text", + "content": "University of Maryland. " + }, + { + "bbox": [ + 62, + 142, + 561, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 62, + 142, + 561, + 171 + ], + "type": "text", + "content": "The Hong Kong Polytechnic University." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 85, + 174, + 538, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 174, + 538, + 185 + ], + "spans": [ + { + "bbox": [ + 85, + 174, + 538, + 185 + ], + "type": "text", + "content": "lixu.wang@ntu.edu.sg, {cxliu539, gjf2023}@umd.edu, qqing.ye@polyu.edu.hk, heng@umd.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 200, + 188, + 424, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 188, + 424, + 198 + ], + "spans": [ + { + "bbox": [ + 200, + 188, + 424, + 198 + ], + "type": "text", + "content": "haibo.hu@polyu.edu.hk, wei_dong@ntu.edu.sg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 253, + 297, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 253, + 297, + 601 + ], + "spans": [ + { + "bbox": [ + 53, + 253, + 297, + 601 + ], + "type": "text", + "content": "Federated Learning (FL) studies often assume a static data distribution, whereas real-world scenarios involve dynamic changes. To address this gap, we study Federated Continuous Category Discovery and Learning " + }, + { + "bbox": [ + 53, + 253, + 297, + 601 + ], + "type": "inline_equation", + "content": "(FC^2 DL)" + }, + { + "bbox": [ + 53, + 253, + 297, + 601 + ], + "type": "text", + "content": ", an essential yet underexplored problem that enables FL models to evolve continuously by discovering and learning novel data categories. The key challenge in " + }, + { + "bbox": [ + 53, + 253, + 297, + 601 + ], + "type": "inline_equation", + "content": "FC^2 DL" + }, + { + "bbox": [ + 53, + 253, + 297, + 601 + ], + "type": "text", + "content": " lies in merging and aligning categories discovered and learned by different clients, all while maintaining privacy. To tackle this, we propose the Global Prototype Alignment (GPA) framework. GPA first estimates the number of categories and constructs global prototypes by locating high-density regions in the representation space through bi-level clustering. To mitigate pseudo-label noise, GPA then employs a semantic-weighted loss to capture correlations between global prototypes and the novel data. This semantic weighting strategy is also used for contrastive loss, facilitating unsupervised novel-category learning. Besides, GPA incorporates a mixup-based mechanism for both data and models, effectively mitigating interference between known and novel categories while alleviating forgetting. Extensive experiments across multiple datasets demonstrate GPA's superiority over state-of-the-art baseline approaches. Notably, GPA achieves absolute gains of " + }, + { + "bbox": [ + 53, + 253, + 297, + 601 + ], + "type": "inline_equation", + "content": "5.7\\%" + }, + { + "bbox": [ + 53, + 253, + 297, + 601 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 53, + 253, + 297, + 601 + ], + "type": "inline_equation", + "content": "13.1\\%" + }, + { + "bbox": [ + 53, + 253, + 297, + 601 + ], + "type": "text", + "content": " in novel category accuracy while preserving known category performance. Furthermore, GPA is highly adaptable, equipping various mainstream FL algorithms with category discovery and learning capabilities, underscoring its potential for real-world deployment." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 624, + 135, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 135, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 135, + 637 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 644, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 644, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 644, + 296, + 693 + ], + "type": "text", + "content": "Federated Learning [37] (FL) enables distributed devices to collaboratively train models while preserving data privacy, making it a compelling solution for privacy-sensitive applications. However, most existing FL studies assume a static" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 228, + 555, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 396 + ], + "type": "text", + "content": "and well-characterized data distribution [4, 13, 20, 30]. In reality, data is inherently dynamic, continuously evolving as new information emerges [41, 54, 58]. This evolving nature presents significant challenges, particularly in enabling FL systems to adapt and remain effective in the presence of novel patterns and functionalities. To address this, FL must be equipped with mechanisms to identify and incorporate emerging functionalities, ensuring adaptability to evolving data distributions. This capability is particularly critical in domains such as healthcare, finance, transportation, and the Internet of Things. For example, FL-powered diagnostic models in healthcare must continuously adapt to the emergence of new diseases and rare conditions to maintain their effectiveness [43, 55]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 399, + 556, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 556, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 556, + 651 + ], + "type": "text", + "content": "To enable FL systems to adapt to continuously emerging novel functionalities, we study Federated Continuous Category Discovery and Learning " + }, + { + "bbox": [ + 313, + 399, + 556, + 651 + ], + "type": "inline_equation", + "content": "(\\mathsf{FC}^2\\mathsf{DL})" + }, + { + "bbox": [ + 313, + 399, + 556, + 651 + ], + "type": "text", + "content": ". We consider an FL setting where multiple participants initially perform supervised training on a predefined category space, with non-independent and non-identically distributed (non-IID) data [37]. Once trained, the FL model is expected to generalize well to known categories. At this stage, the FL administrator typically stops the training to conserve resources and deploys the latest global model for inference [37]. After deployment, users provide unlabeled data and obtain model predictions. In real-world scenarios, participants frequently encounter data from previously unseen novel categories, which also follow a non-IID distribution. In this case, these novel data samples are often misclassified as existing known categories. " + }, + { + "bbox": [ + 313, + 399, + 556, + 651 + ], + "type": "inline_equation", + "content": "\\mathsf{FC}^2\\mathsf{DL}" + }, + { + "bbox": [ + 313, + 399, + 556, + 651 + ], + "type": "text", + "content": " aims to enable FL systems to discover and learn novel categories continuously, ensuring adaptability to evolving data distributions. At the same time, " + }, + { + "bbox": [ + 313, + 399, + 556, + 651 + ], + "type": "inline_equation", + "content": "\\mathsf{FC}^2\\mathsf{DL}" + }, + { + "bbox": [ + 313, + 399, + 556, + 651 + ], + "type": "text", + "content": " seeks to maintain strong performance on previously learned categories, striking a balance between new knowledge learning and retention of prior knowledge." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": "A straightforward approach to tackling " + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{FC}^2\\mathbf{DL}" + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": " is to integrate FL with Novel Category Discovery and Learning (NCDL) [32, 44, 62]. However, this approach presents several fundamental challenges. First, effective clustering of novel data requires prior knowledge of the number of" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 215, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 215, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 215, + 713 + ], + "type": "text", + "content": "*Equal contributions. †Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2429" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "type": "text", + "content": "novel categories. Conventional NCDL assumes this information is available, as model trainers typically have access to the training data and the specific application domain [32, 44, 46]. However, in FL, this assumption does not hold: the FL administrator lacks direct data access, and participants may not have the expertise to estimate task details [10, 37]. Second, in a non-IID setting, even if participants conduct NCDL locally, the discovered novel categories vary across clients, making it difficult to merge their label spaces into a unified one. Finally, most existing NCDL methods often rely on labeled data from known categories [7, 16, 49], which is challenging to obtain in practical FL systems. Many participants have limited storage, and newly joined clients may not have any labeled data [10]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 243, + 295, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 295, + 554 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 295, + 554 + ], + "type": "text", + "content": "To overcome these challenges, we propose Global Prototype Alignment (GPA), a novel framework for " + }, + { + "bbox": [ + 56, + 243, + 295, + 554 + ], + "type": "inline_equation", + "content": "\\mathbf{FC}^2\\mathbf{DL}" + }, + { + "bbox": [ + 56, + 243, + 295, + 554 + ], + "type": "text", + "content": ", as illustrated in Figure 1. First, GPA needs participants to construct potential cluster centroids—synthetic virtual points that do not correspond to real data and are transmitted to the server only once. The server then applies Potential Prototype Merge (PPM) to localize high-density regions among all received centroids, estimating the number of novel categories and initializing the corresponding global prototypes. To facilitate effective learning, GPA introduces Semantic-Weighted Loss (SWL), which guides local training by enforcing alignment with global prototypes. This weighting strategy further enables Contrastive Weighted Loss (CWL), allowing GPA to learn from unlabeled novel-category data effectively. Besides, GPA incorporates a mixup mechanism at both the data and model levels to mitigate interference between known and novel categories while alleviating forgetting. Extensive experiments demonstrate that GPA achieves substantial improvements in novel-category accuracy for " + }, + { + "bbox": [ + 56, + 243, + 295, + 554 + ], + "type": "inline_equation", + "content": "\\mathbf{FC}^2\\mathbf{DL}" + }, + { + "bbox": [ + 56, + 243, + 295, + 554 + ], + "type": "text", + "content": ", outperforming state-of-the-art methods with absolute gains of " + }, + { + "bbox": [ + 56, + 243, + 295, + 554 + ], + "type": "inline_equation", + "content": "5.7\\% - 13.1\\%" + }, + { + "bbox": [ + 56, + 243, + 295, + 554 + ], + "type": "text", + "content": " in a single-stage setting and " + }, + { + "bbox": [ + 56, + 243, + 295, + 554 + ], + "type": "inline_equation", + "content": "4.1\\% - 13.0\\%" + }, + { + "bbox": [ + 56, + 243, + 295, + 554 + ], + "type": "text", + "content": " across five stages, while maintaining near-zero forgetting on known categories. Furthermore, GPA is adaptable and can be seamlessly integrated into various mainstream FL algorithms, equipping them with strong NCDL capabilities. In summary, our contributions include:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 558, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 55, + 558, + 294, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 294, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 294, + 617 + ], + "type": "text", + "content": "- We address an important and practical problem called Federated Continuous Category Discovery and Learning " + }, + { + "bbox": [ + 55, + 558, + 294, + 617 + ], + "type": "inline_equation", + "content": "(\\mathbf{FC}^2\\mathbf{DL})" + }, + { + "bbox": [ + 55, + 558, + 294, + 617 + ], + "type": "text", + "content": ", in which the biggest challenge is to merge and align novel categories of different participants that conform to a non-IID unsupervised setting." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 618, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 294, + 713 + ], + "type": "text", + "content": "- We design Global Prototype Alignment (GPA) to tackle FC" + }, + { + "bbox": [ + 55, + 618, + 294, + 713 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 55, + 618, + 294, + 713 + ], + "type": "text", + "content": "DL. GPA mainly consists of a novel category number estimation approach Potential Prototype Merge that can find high-density regions in unlabeled data, a Semantic-Weighted Loss and a Contrastive Weighted Loss that align the unlabeled novel data with global prototypes in local training, and a mixup mechanism for both the data and model that handle category interference and forgetting." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 553, + 156 + ], + "type": "text", + "content": "- We validate GPA on multiple datasets and various " + }, + { + "bbox": [ + 314, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{FC}^2\\mathbf{DL}" + }, + { + "bbox": [ + 314, + 72, + 553, + 156 + ], + "type": "text", + "content": " scenarios. GPA achieves significant performance improvements over state-of-the-art baseline approaches when solving " + }, + { + "bbox": [ + 314, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{FC}^2\\mathbf{DL}" + }, + { + "bbox": [ + 314, + 72, + 553, + 156 + ], + "type": "text", + "content": ". Moreover, except for the standard FedAvg [37] algorithm, GPA is highly effective in equipping many other FL algorithms with the capability of category discovery and learning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 165, + 400, + 177 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 165, + 400, + 177 + ], + "spans": [ + { + "bbox": [ + 314, + 165, + 400, + 177 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 185, + 553, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 185, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 313, + 185, + 553, + 448 + ], + "type": "text", + "content": "Federated Learning (FL) research [13, 20, 30] mainly focuses on static scenarios, assuming minimal changes in the FL system over time. However, this assumption is overly idealistic and fails to capture many real-world scenarios [54]. Although some studies [22, 38, 59] employ adversarial training to improve out-of-distribution generalization in FL models, and Federated Open-World and OpenSet Semi-Supervised Learning [33, 40, 61] aim to enable learning of unseen categories. However, these methods are restricted to supervised static FL settings, while existing unsupervised FL approaches [35, 60] focus solely on static scenarios, primarily training representation models, with downstream tasks requiring additional adaptation. Recent research has explored dynamic FL [10, 41, 58], where data distributions evolve over time; however, these studies remain confined to the supervised setting. To the best of our knowledge, only one study [40] addresses unseen category discovery in FL, but it depends on labeled data and lacks support for continuous learning. In summary, current FL research cannot discover and learn novel categories in unsupervised settings while mitigating catastrophic forgetting of previously learned knowledge." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 450, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 553, + 713 + ], + "type": "text", + "content": "Novel Category Discovery and Learning (NCDL) refers to the process by which ML models identify and learn previously unseen categories [16]. Existing NCDL approaches can be broadly categorized into label-available and label-unavailable settings. Label-available NCDL, also referred to as Open-World Semi-Supervised Learning (SSL) [2, 15], typically integrates labeled known-category data with unlabeled novel-category data, leveraging SSL [32, 56], clustering [50], or weakly-supervised learning [1, 6] to facilitate novel category discovery. In contrast, the less-explored label-unavailable NCDL addresses practical scenarios where labeled known-category data is unavailable due to privacy constraints or cost limitations [42]. These methods typically follow a two-stage process: an initial supervised learning phase with labeled known-category data, followed by pseudo-labeling or self-supervised learning [44, 62, 63] on unlabeled data containing both known and novel categories. Recent research focuses on mitigating catastrophic forgetting during the continuous execution of the second stage [3, 36, 39, 64]. " + }, + { + "bbox": [ + 313, + 450, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathsf{FC}^{2}\\mathsf{DL}" + }, + { + "bbox": [ + 313, + 450, + 553, + 713 + ], + "type": "text", + "content": " falls into the label-unavailable NCDL category; however, data heterogeneity in FL significantly diminishes the effectiveness of existing" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2430" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 72, + 541, + 336 + ], + "blocks": [ + { + "bbox": [ + 60, + 72, + 541, + 336 + ], + "lines": [ + { + "bbox": [ + 60, + 72, + 541, + 336 + ], + "spans": [ + { + "bbox": [ + 60, + 72, + 541, + 336 + ], + "type": "image", + "image_path": "a0499a4261107fff32fe66416be83b1a8e52d44e4178c58f84213b2cfa922f92.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 338, + 555, + 384 + ], + "lines": [ + { + "bbox": [ + 55, + 338, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 555, + 384 + ], + "type": "text", + "content": "Figure 1. Overview of GPA for enabling FL models to discover and learn novel categories continuously. When the known category learning finishes and the novel category learning starts, FL participants apply Kmeans on their data to find local potential novel prototypes. These local prototypes are sent to the server only once and GPA applies PPM to merge these prototypes into global novel prototypes. Then each selected client leverages SWL, CWL, and Data Mixup to conduct local training and update the model via Model Mixup." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 403, + 135, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 135, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 135, + 415 + ], + "type": "text", + "content": "NCDL approaches." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 428, + 140, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 428, + 140, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 428, + 140, + 441 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 449, + 266, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 266, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 266, + 462 + ], + "type": "text", + "content": "3.1. Novel Category Discovery and Learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 466, + 296, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 466, + 296, + 626 + ], + "spans": [ + { + "bbox": [ + 55, + 466, + 296, + 626 + ], + "type": "text", + "content": "In NCDL, there is a labeled dataset " + }, + { + "bbox": [ + 55, + 466, + 296, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{L}} = \\{(x_i^{\\mathrm{L}},y_i^{\\mathrm{L}})\\}_{i = 1}^{N^{\\mathrm{L}}}" + }, + { + "bbox": [ + 55, + 466, + 296, + 626 + ], + "type": "text", + "content": " and all labels belong to a one-hot label space of dimension " + }, + { + "bbox": [ + 55, + 466, + 296, + 626 + ], + "type": "inline_equation", + "content": "C^\\mathrm{L}" + }, + { + "bbox": [ + 55, + 466, + 296, + 626 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 55, + 466, + 296, + 626 + ], + "type": "inline_equation", + "content": "\\pmb {y}_i^{\\mathrm{L}}\\in \\mathcal{V}^{\\mathrm{L}}" + }, + { + "bbox": [ + 55, + 466, + 296, + 626 + ], + "type": "text", + "content": " and the input marginal distribution is " + }, + { + "bbox": [ + 55, + 466, + 296, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{X}^{\\mathrm{L}}" + }, + { + "bbox": [ + 55, + 466, + 296, + 626 + ], + "type": "text", + "content": ". Besides, there is a unlabeled dataset " + }, + { + "bbox": [ + 55, + 466, + 296, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{U}} = \\{\\pmb{x}_i^{\\mathrm{U}}\\}_{i = 1}^{N^{\\mathrm{U}}}" + }, + { + "bbox": [ + 55, + 466, + 296, + 626 + ], + "type": "text", + "content": " which contains " + }, + { + "bbox": [ + 55, + 466, + 296, + 626 + ], + "type": "inline_equation", + "content": "C^\\mathrm{U}" + }, + { + "bbox": [ + 55, + 466, + 296, + 626 + ], + "type": "text", + "content": " unique categories in " + }, + { + "bbox": [ + 55, + 466, + 296, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^{\\mathrm{U}}" + }, + { + "bbox": [ + 55, + 466, + 296, + 626 + ], + "type": "text", + "content": ", and its input distribution is " + }, + { + "bbox": [ + 55, + 466, + 296, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{X}^{\\mathrm{U}}" + }, + { + "bbox": [ + 55, + 466, + 296, + 626 + ], + "type": "text", + "content": ". It is assumed that the labels in " + }, + { + "bbox": [ + 55, + 466, + 296, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^{\\mathrm{L}}" + }, + { + "bbox": [ + 55, + 466, + 296, + 626 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 466, + 296, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^{\\mathrm{U}}" + }, + { + "bbox": [ + 55, + 466, + 296, + 626 + ], + "type": "text", + "content": " are disjoint, i.e., " + }, + { + "bbox": [ + 55, + 466, + 296, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^{\\mathrm{L}}\\cap \\mathcal{V}^{\\mathrm{U}} = \\emptyset" + }, + { + "bbox": [ + 55, + 466, + 296, + 626 + ], + "type": "text", + "content": ", which are usually called known categories and novel categories, respectively. The general objective of NCDL is to partition the novel-category data into different clusters with the help of " + }, + { + "bbox": [ + 55, + 466, + 296, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{L}}" + }, + { + "bbox": [ + 55, + 466, + 296, + 626 + ], + "type": "text", + "content": ". More specifically, we leverage a neural network " + }, + { + "bbox": [ + 55, + 466, + 296, + 626 + ], + "type": "inline_equation", + "content": "m = f_{\\theta}\\circ g_{\\omega}" + }, + { + "bbox": [ + 55, + 466, + 296, + 626 + ], + "type": "text", + "content": " which consists of a feature extractor " + }, + { + "bbox": [ + 55, + 466, + 296, + 626 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 466, + 296, + 626 + ], + "type": "text", + "content": " parameterized on " + }, + { + "bbox": [ + 55, + 466, + 296, + 626 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 466, + 296, + 626 + ], + "type": "text", + "content": " and a classifier " + }, + { + "bbox": [ + 55, + 466, + 296, + 626 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 466, + 296, + 626 + ], + "type": "text", + "content": " parameterized on " + }, + { + "bbox": [ + 55, + 466, + 296, + 626 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 55, + 466, + 296, + 626 + ], + "type": "text", + "content": ", to achieve NCDL:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 636, + 296, + 665 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 636, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 636, + 296, + 665 + ], + "type": "interline_equation", + "content": "\\max _ {\\theta , \\omega} \\mathbb {E} _ {\\mathcal {X} ^ {\\mathrm {U}}, \\mathcal {Y} ^ {\\mathrm {U}}} \\left\\{\\mathbb {I} \\left[ g _ {\\omega} \\left(f _ {\\theta} \\left(\\boldsymbol {x} _ {i}\\right)\\right) = g _ {\\omega} \\left(f _ {\\theta} \\left(\\boldsymbol {x} _ {j}\\right)\\right) \\right] \\cdot \\mathbb {I} \\left[ \\boldsymbol {y} _ {i} = \\boldsymbol {y} _ {j} \\right] \\right\\}, \\tag {1}", + "image_path": "5ba1f193c3c141f9a164832ef61615c0e6420cf1ff15df8cb3e1058520cf23c2.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{I}[\\cdot ]" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " is an indicator function that if the inner condition is true, " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{I}[\\mathrm{True}] = 1" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ", otherwise " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{I}[\\mathrm{False}] = 0" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{y}_i,\\pmb{y}_j" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " are the ground truth labels of novel samples " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i,\\pmb{x}_j" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ", though unavailable during NCDL." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 403, + 430, + 416 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 403, + 430, + 416 + ], + "spans": [ + { + "bbox": [ + 314, + 403, + 430, + 416 + ], + "type": "text", + "content": "3.2. Federated Learning" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 421, + 555, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 555, + 578 + ], + "type": "text", + "content": "At the beginning of FL, we assume that a global model " + }, + { + "bbox": [ + 313, + 421, + 555, + 578 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 421, + 555, + 578 + ], + "type": "text", + "content": " is trained across " + }, + { + "bbox": [ + 313, + 421, + 555, + 578 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 421, + 555, + 578 + ], + "type": "text", + "content": " participants, and each participant holds its local dataset " + }, + { + "bbox": [ + 313, + 421, + 555, + 578 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^k = \\{(x_i^k,y_i^k)\\}_{i = 1}^{N^k}" + }, + { + "bbox": [ + 313, + 421, + 555, + 578 + ], + "type": "text", + "content": ". The data distributions of participants are non-IID from each other [19, 37]. In this case, the label spaces of different participants " + }, + { + "bbox": [ + 313, + 421, + 555, + 578 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^k" + }, + { + "bbox": [ + 313, + 421, + 555, + 578 + ], + "type": "text", + "content": " are not the same, but they all belong to a unified space " + }, + { + "bbox": [ + 313, + 421, + 555, + 578 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^{\\mathrm{L}}" + }, + { + "bbox": [ + 313, + 421, + 555, + 578 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 313, + 421, + 555, + 578 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^k\\subseteq \\mathcal{V}^{\\mathrm{L}}" + }, + { + "bbox": [ + 313, + 421, + 555, + 578 + ], + "type": "text", + "content": ", which can be regarded as known categories in NCDL. For every global round, a random set of participants is selected as clients, namely " + }, + { + "bbox": [ + 313, + 421, + 555, + 578 + ], + "type": "inline_equation", + "content": "\\mathcal{K}" + }, + { + "bbox": [ + 313, + 421, + 555, + 578 + ], + "type": "text", + "content": ", and they need to conduct local training and then upload the model updates to the server. We assume that the local training is conducted with the CrossEntropy loss and the server applies FedAvg [37] to aggregate model updates:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 381, + 588, + 555, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 588, + 555, + 602 + ], + "spans": [ + { + "bbox": [ + 381, + 588, + 555, + 602 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C E}} = - \\log g _ {\\omega} \\left(f _ {\\theta} \\left(\\boldsymbol {x} _ {i}\\right)\\right). \\tag {2}", + "image_path": "0bba2167fcb3d1ff635126735238534e09f01e07f0028c5594d3924e03f739a9.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 611, + 482, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 611, + 482, + 623 + ], + "spans": [ + { + "bbox": [ + 314, + 611, + 482, + 623 + ], + "type": "text", + "content": "3.3. Problem Formulation of " + }, + { + "bbox": [ + 314, + 611, + 482, + 623 + ], + "type": "inline_equation", + "content": "\\mathsf{FC^2DL}" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": "After training in a sufficient number of rounds, which we call Known Category Learning " + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^{\\mathrm{L}}" + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": ", the FL model " + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": " is hopefully able to perform well on " + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^{\\mathrm{L}}" + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": ". Then, the FL administrator may stop the training to save resources and deploy the latest model " + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "inline_equation", + "content": "m^{\\mathrm{L}}" + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": " to serve all participants in the FL framework. During the usage of " + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "inline_equation", + "content": "m^{\\mathrm{L}}" + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": ", participants feed it with unlabeled testing data " + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{x}^{\\mathrm{test}}" + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": " and obtain corresponding" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2431" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "inference results " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "g_{\\omega}(f_{\\theta}(\\boldsymbol{x}^{\\mathrm{test}}))" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": ". Intuitively, if " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}^{\\mathrm{test}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " belongs to the known categories " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "(\\mathcal{V}^{\\mathrm{L}})" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": ", the inference results should be mostly correct and accurate, i.e., " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "\\mathbb{I}[g_{\\omega}(f_{\\theta}(\\boldsymbol{x}^{\\mathrm{test}})) = \\boldsymbol{y}^{\\mathrm{test}}] = 1" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": ". However, if " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}^{\\mathrm{test}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " comes from unseen novel categories, the current model " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "m^{\\mathrm{L}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " will classify it into a certain known category, in other words, " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " misclassification " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "(\\mathbb{I}[g_{\\omega}(f_{\\theta}(\\boldsymbol{x}^{\\mathrm{test}})) = \\boldsymbol{y}^{\\mathrm{test}}] = 0)" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 157, + 295, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 157, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 157, + 295, + 350 + ], + "type": "text", + "content": "Now let us more formally define the problem of " + }, + { + "bbox": [ + 55, + 157, + 295, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{FC}^2\\mathbf{DL}" + }, + { + "bbox": [ + 55, + 157, + 295, + 350 + ], + "type": "text", + "content": ". Suppose that the FL training periodically steps into a stage called Novel Category Learning " + }, + { + "bbox": [ + 55, + 157, + 295, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^{\\mathrm{U}}" + }, + { + "bbox": [ + 55, + 157, + 295, + 350 + ], + "type": "text", + "content": ", and we assume that there are " + }, + { + "bbox": [ + 55, + 157, + 295, + 350 + ], + "type": "inline_equation", + "content": "K^{\\mathrm{U}}" + }, + { + "bbox": [ + 55, + 157, + 295, + 350 + ], + "type": "text", + "content": " participants available in total in " + }, + { + "bbox": [ + 55, + 157, + 295, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^{\\mathrm{U}}" + }, + { + "bbox": [ + 55, + 157, + 295, + 350 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 55, + 157, + 295, + 350 + ], + "type": "inline_equation", + "content": "K^{\\mathrm{U}}" + }, + { + "bbox": [ + 55, + 157, + 295, + 350 + ], + "type": "text", + "content": " is uncertain and may be changing dynamically, especially since old participants are free to withdraw and new participants are welcome to join in at any time [10]. At this moment, we denote the local novel-category dataset of participant " + }, + { + "bbox": [ + 55, + 157, + 295, + 350 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 157, + 295, + 350 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 157, + 295, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{U},k} = \\{\\pmb{x}_i^{\\mathrm{U},k}\\}_{i=1}^{N^{\\mathrm{U},k}}" + }, + { + "bbox": [ + 55, + 157, + 295, + 350 + ], + "type": "text", + "content": ", screened from its received unlabeled testing dataset " + }, + { + "bbox": [ + 55, + 157, + 295, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{test}}" + }, + { + "bbox": [ + 55, + 157, + 295, + 350 + ], + "type": "text", + "content": ". Consistent with the non-IID setting, although " + }, + { + "bbox": [ + 55, + 157, + 295, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{U},k}" + }, + { + "bbox": [ + 55, + 157, + 295, + 350 + ], + "type": "text", + "content": " has no labels, its groundtruth label space " + }, + { + "bbox": [ + 55, + 157, + 295, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}^{\\mathrm{U},k}" + }, + { + "bbox": [ + 55, + 157, + 295, + 350 + ], + "type": "text", + "content": " is distinct across different participants but also belongs to a unified novel label space " + }, + { + "bbox": [ + 55, + 157, + 295, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}^{\\mathrm{U}}" + }, + { + "bbox": [ + 55, + 157, + 295, + 350 + ], + "type": "text", + "content": ". Our objective is to conduct effective NCDL that can cluster novel-category data into different clusters and preserve good performance on known categories as well:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 356, + 295, + 414 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 356, + 295, + 414 + ], + "spans": [ + { + "bbox": [ + 66, + 356, + 295, + 414 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\max _ {\\theta , \\omega} \\left\\{\\sum_ {\\boldsymbol {x} \\in \\mathcal {D} ^ {\\mathrm {U}}} \\mathbb {I} \\left[ g _ {\\omega} \\left(f _ {\\theta} \\left(\\boldsymbol {x} _ {i}\\right)\\right) = g _ {\\omega} \\left(f _ {\\theta} \\left(\\boldsymbol {x} _ {j}\\right)\\right) \\right] \\cdot \\mathbb {I} \\left[ \\boldsymbol {y} _ {i} = \\boldsymbol {y} _ {j} \\right] \\right. \\tag {3} \\\\ \\left. \\right.\\left. + \\mathbb {E} _ {\\boldsymbol {x} \\sim \\mathcal {X} ^ {\\mathrm {L}}; \\boldsymbol {y} \\sim \\mathcal {Y} ^ {\\mathrm {L}}} \\mathbb {I} \\left[ g _ {\\omega} \\left(f _ {\\theta} (\\boldsymbol {x})\\right) = \\boldsymbol {y} \\right]\\right\\}. \\\\ \\end{array}", + "image_path": "c845bd329ed19afbe95eb452250a45d2446df3ef88bf3d35371a24bd7f3b33f9.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 421, + 230, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 421, + 230, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 230, + 434 + ], + "type": "text", + "content": "4. Our Proposed GPA Framework" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 440, + 204, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 440, + 204, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 440, + 204, + 453 + ], + "type": "text", + "content": "4.1. Known Category Learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 457, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 457, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 457, + 295, + 581 + ], + "type": "text", + "content": "To achieve better NCDL, model " + }, + { + "bbox": [ + 55, + 457, + 295, + 581 + ], + "type": "inline_equation", + "content": "m^{\\mathrm{L}}" + }, + { + "bbox": [ + 55, + 457, + 295, + 581 + ], + "type": "text", + "content": " should not only have a good performance on known categories but also extract high-quality representations for novel categories. To this end, we modify Prototype Contrastive Learning (PCL) [28, 34] to enhance the training in " + }, + { + "bbox": [ + 55, + 457, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^{\\mathrm{L}}" + }, + { + "bbox": [ + 55, + 457, + 295, + 581 + ], + "type": "text", + "content": ". More specifically, our modified PCL constructs positive pairs as a sample representation " + }, + { + "bbox": [ + 55, + 457, + 295, + 581 + ], + "type": "inline_equation", + "content": "z_{i} = f_{\\theta}(x_{i})" + }, + { + "bbox": [ + 55, + 457, + 295, + 581 + ], + "type": "text", + "content": " and its corresponding category prototype " + }, + { + "bbox": [ + 55, + 457, + 295, + 581 + ], + "type": "inline_equation", + "content": "p_{y_i}^{\\mathrm{L}}" + }, + { + "bbox": [ + 55, + 457, + 295, + 581 + ], + "type": "text", + "content": ". The negative pairs consist of two parts: all pairs between " + }, + { + "bbox": [ + 55, + 457, + 295, + 581 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 55, + 457, + 295, + 581 + ], + "type": "text", + "content": " and all prototypes " + }, + { + "bbox": [ + 55, + 457, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\{p_c^{\\mathrm{L}}\\}_{c = 1}^{C^{\\mathrm{L}}}" + }, + { + "bbox": [ + 55, + 457, + 295, + 581 + ], + "type": "text", + "content": ", and sample-wise pairs between " + }, + { + "bbox": [ + 55, + 457, + 295, + 581 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 55, + 457, + 295, + 581 + ], + "type": "text", + "content": " and samples belonging to other labels:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 586, + 298, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 298, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 298, + 634 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {P C L}} = - \\log \\frac {\\exp \\left(\\boldsymbol {z} _ {i} ^ {\\mathrm {L}} \\cdot \\boldsymbol {p} _ {\\boldsymbol {y} _ {i} ^ {\\mathrm {L}}} ^ {\\mathrm {L}} / \\tau\\right)}{\\sum_ {c = 1} ^ {C ^ {\\mathrm {L}}} \\exp \\left(\\boldsymbol {z} _ {i} ^ {\\mathrm {L}} \\cdot \\boldsymbol {p} _ {c} ^ {\\mathrm {L}} / \\tau\\right) + \\sum_ {j = 1} ^ {N _ {i} ^ {-}} \\exp \\left(\\boldsymbol {z} _ {i} ^ {\\mathrm {L}} \\cdot \\boldsymbol {z} _ {j} ^ {\\mathrm {L}} / \\tau\\right)} \\tag {4}", + "image_path": "57c7906977b39b87eb1f19afd6b80ea2c97c79c60128a06168223c404fc0103b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\tau = 0.07" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " (consistent with [28]) is a temperature parameter and " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "N_{i}^{-}" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " is the set size of samples with different labels from " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " in a data batch. Then we apply a combined loss " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{T}^{\\mathrm{L}}} = \\frac{1}{N^{\\mathrm{L},k}}\\sum_{\\mathcal{D}^{\\mathrm{L},k}}[\\mathcal{L}_{\\mathrm{CE}} + \\eta \\mathcal{L}_{\\mathrm{PCL}}]" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " to train " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " on known category data (" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\eta = 0.1" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "; please refer to the Appendix B.4 for sensitivity analysis)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 550, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 550, + 85 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 550, + 85 + ], + "type": "text", + "content": "4.2. Preparations before Novel Category Learning" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 89, + 474, + 101 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 474, + 101 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 474, + 101 + ], + "type": "text", + "content": "4.2.1. Novel-Category Data Filtering." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 103, + 553, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 103, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 103, + 553, + 297 + ], + "type": "text", + "content": "When stepping into Novel Category Learning " + }, + { + "bbox": [ + 313, + 103, + 553, + 297 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^{\\mathrm{U}}" + }, + { + "bbox": [ + 313, + 103, + 553, + 297 + ], + "type": "text", + "content": ", it is possible for every participant to continually receive data of both novel and known categories. Thus we first need to separate known and novel categories from " + }, + { + "bbox": [ + 313, + 103, + 553, + 297 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{test},k}" + }, + { + "bbox": [ + 313, + 103, + 553, + 297 + ], + "type": "text", + "content": ". Specifically, for each participant " + }, + { + "bbox": [ + 313, + 103, + 553, + 297 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 103, + 553, + 297 + ], + "type": "text", + "content": ", we assume that there is a data memory " + }, + { + "bbox": [ + 313, + 103, + 553, + 297 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^k" + }, + { + "bbox": [ + 313, + 103, + 553, + 297 + ], + "type": "text", + "content": " employed to store " + }, + { + "bbox": [ + 313, + 103, + 553, + 297 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{U},k}" + }, + { + "bbox": [ + 313, + 103, + 553, + 297 + ], + "type": "text", + "content": " that is filtered out during model " + }, + { + "bbox": [ + 313, + 103, + 553, + 297 + ], + "type": "inline_equation", + "content": "m^{\\mathrm{L}}" + }, + { + "bbox": [ + 313, + 103, + 553, + 297 + ], + "type": "text", + "content": " usage*. We adopt a popular strategy [62] based on the representation distance between data samples and known-category prototypes. The traditional prototype [44, 62] is constructed by averaging category-wise representations, but this is infeasible in FL as the data is distributed among participants. We solve this problem by viewing the first " + }, + { + "bbox": [ + 313, + 103, + 553, + 297 + ], + "type": "inline_equation", + "content": "C^{\\mathrm{L}}" + }, + { + "bbox": [ + 313, + 103, + 553, + 297 + ], + "type": "text", + "content": " dimensions of neuron weights " + }, + { + "bbox": [ + 313, + 103, + 553, + 297 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 313, + 103, + 553, + 297 + ], + "type": "text", + "content": " of classifier " + }, + { + "bbox": [ + 313, + 103, + 553, + 297 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 103, + 553, + 297 + ], + "type": "text", + "content": " as the prototypes " + }, + { + "bbox": [ + 313, + 103, + 553, + 297 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{\\mathrm{L}} = \\{\\pmb{p}_c^{\\mathrm{L}}\\}_{c=1}^{C^{\\mathrm{L}}}" + }, + { + "bbox": [ + 313, + 103, + 553, + 297 + ], + "type": "text", + "content": ", if " + }, + { + "bbox": [ + 313, + 103, + 553, + 297 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 103, + 553, + 297 + ], + "type": "text", + "content": " is composed of only one single linear layer without bias [57]. The detailed filtering mechanism is as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 316, + 300, + 553, + 343 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 300, + 553, + 343 + ], + "spans": [ + { + "bbox": [ + 316, + 300, + 553, + 343 + ], + "type": "interline_equation", + "content": "\\mathcal {D} ^ {\\mathrm {U}, k} = \\left\\{\\boldsymbol {x} \\in \\mathcal {D} ^ {\\text {t e s t}, k} \\Bigg | \\max \\left\\{\\frac {f _ {\\theta} (\\boldsymbol {x}) \\cdot \\boldsymbol {p} _ {c} ^ {\\mathrm {L}}}{\\| f _ {\\theta} (\\boldsymbol {x}) \\| \\| \\boldsymbol {p} _ {c} ^ {\\mathrm {L}} \\|} \\right\\} _ {c = 1} ^ {C ^ {\\mathrm {L}}} < r \\right\\}, \\tag {5}", + "image_path": "585c15d80bb131dc0d0a9f221f4c432152cd9b5905083fe9d4e4a70f70194f4f.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 344, + 555, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 344, + 555, + 451 + ], + "spans": [ + { + "bbox": [ + 313, + 344, + 555, + 451 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 344, + 555, + 451 + ], + "type": "inline_equation", + "content": "r = 0.5" + }, + { + "bbox": [ + 313, + 344, + 555, + 451 + ], + "type": "text", + "content": " is a threshold. This implies that if a sample " + }, + { + "bbox": [ + 313, + 344, + 555, + 451 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 344, + 555, + 451 + ], + "type": "text", + "content": " has the largest similarity with known prototypes lower than " + }, + { + "bbox": [ + 313, + 344, + 555, + 451 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 344, + 555, + 451 + ], + "type": "text", + "content": ", it should belong to novel categories. Note that such filtering cannot screen the known-category data completely. To prevent potential interference between known and novel categories, we design a data-level mixup strategy. As for the screened known-category data, we test GPA by applying Softmax-based pseudo labeling on them during " + }, + { + "bbox": [ + 313, + 344, + 555, + 451 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^{\\mathrm{U}}" + }, + { + "bbox": [ + 313, + 344, + 555, + 451 + ], + "type": "text", + "content": " (detailed results can be found in the Appendix A.3)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 456, + 554, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 554, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 554, + 479 + ], + "type": "text", + "content": "4.2.2. Challenges of Estimating the Number of Novel Categories in " + }, + { + "bbox": [ + 313, + 456, + 554, + 479 + ], + "type": "inline_equation", + "content": "\\mathsf{FC}^2\\mathsf{DL}" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 483, + 553, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 553, + 542 + ], + "type": "text", + "content": "The goal of NCDL is to accurately classify novel data into different clusters, which requires prior knowledge of the cluster number. Most existing NCDL studies [32, 44, 56] directly assume this is known, but this assumption is unreasonable in " + }, + { + "bbox": [ + 313, + 483, + 553, + 542 + ], + "type": "inline_equation", + "content": "\\mathbf{FC}^2\\mathbf{DL}" + }, + { + "bbox": [ + 313, + 483, + 553, + 542 + ], + "type": "text", + "content": " due to the following special challenges:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 544, + 553, + 626 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 314, + 544, + 553, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 544, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 314, + 544, + 553, + 578 + ], + "type": "text", + "content": "- FL system administrator cannot access any training data and cannot predict the dynamic changes over time in data or task distribution [10, 37]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 579, + 553, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 579, + 553, + 626 + ], + "spans": [ + { + "bbox": [ + 314, + 579, + 553, + 626 + ], + "type": "text", + "content": "- There are various overlapping degrees between ordered clustering distributions " + }, + { + "bbox": [ + 314, + 579, + 553, + 626 + ], + "type": "inline_equation", + "content": "\\left[c_{1}^{\\mathrm{U}, k}, \\cdots, c_{i}^{\\mathrm{U}, k}, \\cdots\\right]" + }, + { + "bbox": [ + 314, + 579, + 553, + 626 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 314, + 579, + 553, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{U}, k}" + }, + { + "bbox": [ + 314, + 579, + 553, + 626 + ], + "type": "text", + "content": " among different participants when applying unsupervised clustering algorithms locally, i.e.," + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 329, + 632, + 553, + 659 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 632, + 553, + 659 + ], + "spans": [ + { + "bbox": [ + 329, + 632, + 553, + 659 + ], + "type": "interline_equation", + "content": "\\mathcal {P} \\left(\\boldsymbol {c} _ {i} ^ {\\mathrm {U}, k} \\neq \\boldsymbol {c} _ {i} ^ {\\mathrm {U}, k ^ {\\prime}}\\right) > 0, \\mathcal {P} \\left(\\boldsymbol {c} _ {i} ^ {\\mathrm {U}, k} = \\boldsymbol {c} _ {j} ^ {\\mathrm {U}, k ^ {\\prime}}\\right) > 0, (i \\neq j). \\tag {6}", + "image_path": "1a177211c85bee23359436e7a7e5f04fa8c2b87db3684ff61e34a79a6596df56.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": "*Note that old participants can directly apply the data memory that was previously used for storing the known-category data to store the screened novel data, after removing the known-category data due to memory storage limitation or moving them to somewhere else for privacy or intellectual property protection [10, 42, 53]." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2432" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": "- Existing novel-category number estimation methods [7, 16, 50] require labeled data, either from known categories or auxiliary novel categories. This requirement violates cost-cutting and privacy regulations [14] for old participants, while new participants do not have any labeled data [10]. Besides, these methods rely on sufficient observation of the novel data distribution, i.e., they need to access a large amount of novel-category data, which is also unreasonable in practice." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 186, + 198, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 186, + 198, + 198 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 198, + 198 + ], + "type": "text", + "content": "4.2.3. Potential Prototype Merge." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 201, + 296, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 201, + 296, + 345 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 296, + 345 + ], + "type": "text", + "content": "To achieve an accurate global novel-category number estimation, all participants need to identify potential local prototypes within their local data first. Unlike studies [1, 32] that employ Sinkhorn-Knopp-based clustering [12] to cluster unlabeled data into clusters with an equal size, we adopt standard Kmeans [17] to construct local prototypes. The reason is that local data is often class-imbalanced [52]. Specifically, each participant " + }, + { + "bbox": [ + 55, + 201, + 296, + 345 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 201, + 296, + 345 + ], + "type": "text", + "content": " feeds its local data " + }, + { + "bbox": [ + 55, + 201, + 296, + 345 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{U},k}" + }, + { + "bbox": [ + 55, + 201, + 296, + 345 + ], + "type": "text", + "content": " into the model " + }, + { + "bbox": [ + 55, + 201, + 296, + 345 + ], + "type": "inline_equation", + "content": "m^{\\mathrm{L}}" + }, + { + "bbox": [ + 55, + 201, + 296, + 345 + ], + "type": "text", + "content": " to extract representations " + }, + { + "bbox": [ + 55, + 201, + 296, + 345 + ], + "type": "inline_equation", + "content": "\\boldsymbol {z}_i = f_\\theta (\\boldsymbol{x}_i^{\\mathrm{U},k})" + }, + { + "bbox": [ + 55, + 201, + 296, + 345 + ], + "type": "text", + "content": " and then employs Kmeans clustering to group these representations into " + }, + { + "bbox": [ + 55, + 201, + 296, + 345 + ], + "type": "inline_equation", + "content": "C^\\mathrm{L}" + }, + { + "bbox": [ + 55, + 201, + 296, + 345 + ], + "type": "text", + "content": " clusters. Subsequently, the centroids of these clusters are utilized as local prototypes:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 354, + 296, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 354, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 57, + 354, + 296, + 399 + ], + "type": "interline_equation", + "content": "\\mathcal {Z} ^ {\\mathrm {U}, k} = \\left\\{\\frac {\\sum_ {\\boldsymbol {z} \\in \\mathcal {Z} _ {c} ^ {\\mathrm {U} , k}} \\boldsymbol {z}}{| \\mathcal {Z} _ {c} ^ {\\mathrm {U} , k} |} \\right\\} _ {c = 1} ^ {C ^ {\\mathrm {L}}}, \\mathcal {Z} _ {c} ^ {\\mathrm {U}, k} = \\left\\{\\boldsymbol {z} \\mid \\operatorname {K M} (\\boldsymbol {z}) = \\boldsymbol {c} _ {c} ^ {\\mathrm {U}, k} \\right\\}. \\tag {7}", + "image_path": "a6ad7857b0854e4955a6cc0ebec058ea23fd2edb4daf1f161fd98d6a8c0b757b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\mathrm{KM}(z)" + }, + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "type": "text", + "content": " outputs the cluster label of " + }, + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "type": "text", + "content": " when employing Kmeans clustering. Notably, these Kmeans centroids are synthetic points that achieve the minimum aggregated intra-cluster distance. In other words, they do not correspond to any real data. Moreover, these local prototypes are sent to the server only once. In these ways, privacy is protected. More discussions about why our design can protect privacy are provided in the Appendix B.6." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 496, + 296, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 296, + 677 + ], + "type": "text", + "content": "The server shuffles all received local prototypes and constructs a prototype pool " + }, + { + "bbox": [ + 55, + 496, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}^{\\mathrm{U}} = \\bigcup_{k=1}^{K^{\\mathrm{U}}} \\mathcal{Z}^{\\mathrm{U}, k}" + }, + { + "bbox": [ + 55, + 496, + 296, + 677 + ], + "type": "text", + "content": ". When building PPM, we totally depart from conventional design principles in existing novel-category number estimation methods [7, 16, 50]. Instead, we leverage an intuition that data representations belonging to the same novel category appear in close proximity, forming high-density regions. Identifying these regions globally is achievable since we do not conduct any normalization when building local prototypes. Specifically, we modify DBSCAN [45] with a rising radius " + }, + { + "bbox": [ + 55, + 496, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 496, + 296, + 677 + ], + "type": "text", + "content": " and a fixed minimum cluster data size " + }, + { + "bbox": [ + 55, + 496, + 296, + 677 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{size}}" + }, + { + "bbox": [ + 55, + 496, + 296, + 677 + ], + "type": "text", + "content": " to find the high-density regions. " + }, + { + "bbox": [ + 55, + 496, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 496, + 296, + 677 + ], + "type": "text", + "content": " controls how large the nearby area should be considered for a certain data point, while " + }, + { + "bbox": [ + 55, + 496, + 296, + 677 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{size}}" + }, + { + "bbox": [ + 55, + 496, + 296, + 677 + ], + "type": "text", + "content": " defines the requirement that if a data point is a core point, it should be surrounded by " + }, + { + "bbox": [ + 55, + 496, + 296, + 677 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{size}} - 1" + }, + { + "bbox": [ + 55, + 496, + 296, + 677 + ], + "type": "text", + "content": " points. To cover as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": "many high-density regions as possible, we try to use a loose setting with " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{size}} = 2" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": ", which also considers the global imbalance case where minority categories correspond to very few prototypes globally [52]. As for " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": ", we first calculate the pair-wise Euclidean distance range within the prototype pool " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}^{\\mathrm{U}}" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " and set a rising " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 338, + 152, + 553, + 173 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 152, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 338, + 152, + 553, + 173 + ], + "type": "interline_equation", + "content": "\\left\\{\\epsilon_ {e} = \\min D + e \\cdot (\\max D - \\min D) / E \\right\\} _ {e = 0} ^ {E}, \\tag {8}", + "image_path": "63ef60f9d416e3bd7b1ff188558db2a08f4af4483176488308a40ad47b295ccf.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 339, + 167, + 484, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 167, + 484, + 185 + ], + "spans": [ + { + "bbox": [ + 339, + 167, + 484, + 185 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 339, + 167, + 484, + 185 + ], + "type": "inline_equation", + "content": "D = \\{d(\\pmb {z}_i^{\\mathrm{U}},\\pmb {z}_j^{\\mathrm{U}})\\}_{\\pmb {z}_i^{\\mathrm{U}},\\pmb {z}_j^{\\mathrm{U}}\\in \\mathcal{Z}^{\\mathrm{U}}}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 192, + 555, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 555, + 300 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 192, + 555, + 300 + ], + "type": "inline_equation", + "content": "d(\\cdot)" + }, + { + "bbox": [ + 313, + 192, + 555, + 300 + ], + "type": "text", + "content": " computes the Euclidean distance, and " + }, + { + "bbox": [ + 313, + 192, + 555, + 300 + ], + "type": "inline_equation", + "content": "E = 50" + }, + { + "bbox": [ + 313, + 192, + 555, + 300 + ], + "type": "text", + "content": " is the total rising steps. For each " + }, + { + "bbox": [ + 313, + 192, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\epsilon_{e}" + }, + { + "bbox": [ + 313, + 192, + 555, + 300 + ], + "type": "text", + "content": ", we conduct DBSCAN clustering on " + }, + { + "bbox": [ + 313, + 192, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}^{\\mathrm{U}}" + }, + { + "bbox": [ + 313, + 192, + 555, + 300 + ], + "type": "text", + "content": " and record the unique resulting cluster number " + }, + { + "bbox": [ + 313, + 192, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\tilde{c}_e^{\\mathrm{U}}" + }, + { + "bbox": [ + 313, + 192, + 555, + 300 + ], + "type": "text", + "content": ". After conducting the final step, we choose the maximum cluster number as the estimated global novel-category number " + }, + { + "bbox": [ + 313, + 192, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\widetilde{C}^{\\mathrm{U}} = \\max \\left\\{\\tilde{c}_{e}^{\\mathrm{U}}\\right\\}_{e=0}^{E}" + }, + { + "bbox": [ + 313, + 192, + 555, + 300 + ], + "type": "text", + "content": ". We believe that PPM can work effectively with much fewer requirements than existing methods, thus tackling the aforementioned challenges of estimating the novel-category number in FC²DL." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 300, + 554, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 554, + 339 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 554, + 339 + ], + "type": "text", + "content": "Then we apply Kmeans again on " + }, + { + "bbox": [ + 313, + 300, + 554, + 339 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}^{\\mathrm{U}}" + }, + { + "bbox": [ + 313, + 300, + 554, + 339 + ], + "type": "text", + "content": " with the cluster number " + }, + { + "bbox": [ + 313, + 300, + 554, + 339 + ], + "type": "inline_equation", + "content": "\\widetilde{C}^{\\mathrm{U}}" + }, + { + "bbox": [ + 313, + 300, + 554, + 339 + ], + "type": "text", + "content": " and use the cluster centroids " + }, + { + "bbox": [ + 313, + 300, + 554, + 339 + ], + "type": "inline_equation", + "content": "\\hat{z}_c^{\\mathrm{U}}" + }, + { + "bbox": [ + 313, + 300, + 554, + 339 + ], + "type": "text", + "content": " to initialize the global novel prototypes as " + }, + { + "bbox": [ + 313, + 300, + 554, + 339 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{\\mathrm{U}} = \\{\\hat{z}_c^{\\mathrm{U}}\\}_{c=1}^{\\widetilde{C}^{\\mathrm{U}}}" + }, + { + "bbox": [ + 313, + 300, + 554, + 339 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 345, + 552, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 345, + 552, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 552, + 357 + ], + "type": "text", + "content": "4.3. Representation Learning for Novel Categories" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 362, + 448, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 362, + 448, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 362, + 448, + 374 + ], + "type": "text", + "content": "4.3.1. Semantic-Weighted Loss" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 377, + 554, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 554, + 498 + ], + "type": "text", + "content": "By viewing our problem here as weakly-supervised learning, we propose a novel Semantic-Weighted Loss (SWL) to align local data with the global prototypes " + }, + { + "bbox": [ + 313, + 377, + 554, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{\\mathrm{U}}" + }, + { + "bbox": [ + 313, + 377, + 554, + 498 + ], + "type": "text", + "content": " in representation space. Specifically, instead of assigning data samples with a single label, SWL considers the possible semantic relationships between data samples with all global prototypes, which are measured by the Softmax probability. Moreover, as " + }, + { + "bbox": [ + 313, + 377, + 554, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{\\mathrm{U}}" + }, + { + "bbox": [ + 313, + 377, + 554, + 498 + ], + "type": "text", + "content": " is obtained by Euclidean distance-based Kmeans, we minimize the Euclidean distance between samples and prototypes. In this case, SWL is shaped as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 320, + 505, + 553, + 554 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 505, + 553, + 554 + ], + "spans": [ + { + "bbox": [ + 320, + 505, + 553, + 554 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {S W L}} = \\sum_ {c = 1} ^ {\\tilde {C} ^ {\\mathrm {U}}} \\frac {\\exp \\left(f _ {\\theta} \\left(\\boldsymbol {x} _ {i}\\right) \\cdot \\hat {\\boldsymbol {z}} _ {c} ^ {\\mathrm {U}} / \\tau\\right)}{\\sum_ {j = 1} ^ {\\tilde {C} ^ {\\mathrm {U}}} \\exp \\left(f _ {\\theta} \\left(\\boldsymbol {x} _ {i}\\right) \\cdot \\hat {\\boldsymbol {z}} _ {j} ^ {\\mathrm {U}} / \\tau\\right)} d \\left(f _ {\\theta} \\left(\\boldsymbol {x} _ {i}\\right), \\hat {\\boldsymbol {z}} _ {c} ^ {\\mathrm {U}}\\right). \\tag {9}", + "image_path": "7c17545fba5013bf22e3699a72e5d4600560bda3b48494ec1ee974cdaac56119.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 561, + 554, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 561, + 554, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 561, + 554, + 597 + ], + "type": "text", + "content": "By weighting the Euclidean distance with the semantic similarity, SWL can guide the model to learn more distinguishable semantics and be more resistant to pseudo-label noise." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 602, + 457, + 614 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 602, + 457, + 614 + ], + "spans": [ + { + "bbox": [ + 313, + 602, + 457, + 614 + ], + "type": "text", + "content": "4.3.2. Contrastive Weighted Loss" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "content": "SWL's effectiveness relies on the correlations between known and novel categories captured by the model " + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "inline_equation", + "content": "m^{\\mathrm{L}}" + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "content": ", but " + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "inline_equation", + "content": "m^{\\mathrm{L}}" + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "content": " is trained on labeled known-category data and thus naturally cannot extract high-quality representations for unlabeled novel-category data. To this end, we employ instance discrimination [5] to facilitate better feature extraction from unlabeled novel-category data. However, standard instance discrimination requires lengthy training to converge, which" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 682, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 682, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 682, + 295, + 713 + ], + "type": "inline_equation", + "content": "^{\\dagger}" + }, + { + "bbox": [ + 55, + 682, + 295, + 713 + ], + "type": "text", + "content": "We assume that " + }, + { + "bbox": [ + 55, + 682, + 295, + 713 + ], + "type": "inline_equation", + "content": "C^{\\mathrm{U}} < C^{\\mathrm{L}}" + }, + { + "bbox": [ + 55, + 682, + 295, + 713 + ], + "type": "text", + "content": " as previous studies [44, 62]. Clustering local novel data into " + }, + { + "bbox": [ + 55, + 682, + 295, + 713 + ], + "type": "inline_equation", + "content": "C^{\\mathrm{L}}" + }, + { + "bbox": [ + 55, + 682, + 295, + 713 + ], + "type": "text", + "content": " clusters can cover as complete novel label space as possible." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2433" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": "lags far behind the pace of SWL. We found that enabling the training to focus on easy samples in the early stage that is more like a curriculum learning can accelerate the convergence process. Therefore, similar to SWL, we also use the semantic relationship to weight the InfoNCE loss," + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 96, + 140, + 242, + 152 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 140, + 242, + 152 + ], + "spans": [ + { + "bbox": [ + 96, + 140, + 242, + 152 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C W L}} = - \\Delta (\\boldsymbol {x} _ {i}, \\boldsymbol {x} _ {i}) \\log \\Delta (\\boldsymbol {x} _ {i}, \\boldsymbol {x} _ {i}),", + "image_path": "dd6bb51d3c637aa94cdb5570e9e6b1b5ad393086ef26551e591eafcd385d3a80.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 155, + 296, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 155, + 296, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 155, + 296, + 195 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 56, + 155, + 296, + 195 + ], + "type": "inline_equation", + "content": "\\Delta (\\pmb {x}_i,\\pmb {x}_i) = \\frac{\\exp\\left(f_\\theta(\\mathrm{A}(\\pmb{x}_i))\\cdot f_\\theta(\\mathrm{A}(\\pmb{x}_i)) / \\tau\\right)}{\\sum_{q = 1}^{N_B^k}\\exp\\left(f_\\theta(\\mathrm{A}(\\pmb{x}_i))\\cdot f_\\theta(\\mathrm{A}(\\pmb{x}_q)) / \\tau\\right)}" + }, + { + "bbox": [ + 56, + 155, + 296, + 195 + ], + "type": "text", + "content": " (10)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 196, + 296, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 196, + 296, + 219 + ], + "spans": [ + { + "bbox": [ + 55, + 196, + 296, + 219 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 196, + 296, + 219 + ], + "type": "inline_equation", + "content": "A(\\cdot)" + }, + { + "bbox": [ + 55, + 196, + 296, + 219 + ], + "type": "text", + "content": " denotes using stochastic augmentation and " + }, + { + "bbox": [ + 55, + 196, + 296, + 219 + ], + "type": "inline_equation", + "content": "N_B^k" + }, + { + "bbox": [ + 55, + 196, + 296, + 219 + ], + "type": "text", + "content": " is the size of a mini-batch." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 226, + 135, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 226, + 135, + 237 + ], + "spans": [ + { + "bbox": [ + 55, + 226, + 135, + 237 + ], + "type": "text", + "content": "4.3.3. Data Mixup" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 240, + 296, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 240, + 296, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 296, + 337 + ], + "type": "text", + "content": "To preserve the knowledge learned from labeled known categories and optimize for novel categories simultaneously, each client " + }, + { + "bbox": [ + 55, + 240, + 296, + 337 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 240, + 296, + 337 + ], + "type": "text", + "content": " applies a mixup mechanism on the data and model. The principle behind the data-level mixup is to handle the interference between known categories and novel categories in the novel data memory " + }, + { + "bbox": [ + 55, + 240, + 296, + 337 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^k" + }, + { + "bbox": [ + 55, + 240, + 296, + 337 + ], + "type": "text", + "content": ". For the specific design, each client " + }, + { + "bbox": [ + 55, + 240, + 296, + 337 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 240, + 296, + 337 + ], + "type": "text", + "content": " constructs a relatively high-confidence dataset " + }, + { + "bbox": [ + 55, + 240, + 296, + 337 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{U},k}" + }, + { + "bbox": [ + 55, + 240, + 296, + 337 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 240, + 296, + 337 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{U},k}" + }, + { + "bbox": [ + 55, + 240, + 296, + 337 + ], + "type": "text", + "content": " in terms of Softmax probability " + }, + { + "bbox": [ + 55, + 240, + 296, + 337 + ], + "type": "inline_equation", + "content": "\\widetilde{\\boldsymbol{y}}" + }, + { + "bbox": [ + 55, + 240, + 296, + 337 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 343, + 298, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 343, + 298, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 343, + 298, + 403 + ], + "type": "interline_equation", + "content": "\\mathcal {D} ^ {\\mathrm {U} ^ {\\prime}, k} = \\left\\{\\boldsymbol {x} \\mid \\max \\left\\{\\frac {\\exp \\left(f _ {\\theta} \\left(\\boldsymbol {x} _ {i}\\right) \\cdot \\boldsymbol {p} _ {c} / \\tau\\right)}{C ^ {\\mathrm {L}} + \\widetilde {C} ^ {\\mathrm {U}}} \\right\\} _ {c = 1} ^ {C ^ {\\mathrm {L}} + \\widetilde {C} ^ {\\mathrm {U}}} > r ^ {\\prime} \\right\\}. \\tag {11}", + "image_path": "06e112d84e63786fecdff1944149591f3ef0f48d798c707a95c0327823f09cb4.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 404, + 296, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 404, + 296, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 404, + 296, + 477 + ], + "type": "inline_equation", + "content": "r'" + }, + { + "bbox": [ + 55, + 404, + 296, + 477 + ], + "type": "text", + "content": " here is a threshold for selecting data to build " + }, + { + "bbox": [ + 55, + 404, + 296, + 477 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{U}',k}" + }, + { + "bbox": [ + 55, + 404, + 296, + 477 + ], + "type": "text", + "content": ", and we set it as 0.95 (please refer to the Appendix B.4 for the sensitivity analysis). Then for every instance " + }, + { + "bbox": [ + 55, + 404, + 296, + 477 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i'" + }, + { + "bbox": [ + 55, + 404, + 296, + 477 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 404, + 296, + 477 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{U}',k}" + }, + { + "bbox": [ + 55, + 404, + 296, + 477 + ], + "type": "text", + "content": ", client " + }, + { + "bbox": [ + 55, + 404, + 296, + 477 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 404, + 296, + 477 + ], + "type": "text", + "content": " will randomly select another instance " + }, + { + "bbox": [ + 55, + 404, + 296, + 477 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i" + }, + { + "bbox": [ + 55, + 404, + 296, + 477 + ], + "type": "text", + "content": " from the residual set " + }, + { + "bbox": [ + 55, + 404, + 296, + 477 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{U},k} \\setminus \\mathcal{D}^{\\mathrm{U}',k}" + }, + { + "bbox": [ + 55, + 404, + 296, + 477 + ], + "type": "text", + "content": " and conduct the data-level mixup " + }, + { + "bbox": [ + 55, + 404, + 296, + 477 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i^* = \\lambda \\pmb{x}_i' + (1 - \\lambda)\\pmb{x}_i" + }, + { + "bbox": [ + 55, + 404, + 296, + 477 + ], + "type": "text", + "content": ", as well as the training as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 484, + 294, + 509 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 484, + 294, + 509 + ], + "spans": [ + { + "bbox": [ + 61, + 484, + 294, + 509 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {M I X}} = \\lambda \\mathcal {L} _ {\\mathrm {C E}} \\left(f _ {\\theta} \\left(\\boldsymbol {x} _ {i} ^ {*}\\right), \\widetilde {\\boldsymbol {y}} _ {i} ^ {\\prime}\\right) + (1 - \\lambda) \\mathcal {L} _ {\\mathrm {C E}} \\left(f _ {\\theta} \\left(\\boldsymbol {x} _ {i} ^ {*}\\right), \\widetilde {\\boldsymbol {y}} _ {i}\\right), \\tag {12}", + "image_path": "652c5064c073a9ee50ad91bf3265e358d51b14d165214a9719bcf086cfa81264.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 509, + 296, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 296, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 296, + 545 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 509, + 296, + 545 + ], + "type": "inline_equation", + "content": "\\lambda \\sim \\mathrm{Beta}(\\lambda; 0.75, 0.75)" + }, + { + "bbox": [ + 55, + 509, + 296, + 545 + ], + "type": "text", + "content": " that is suggested in semi-supervised learning works [9]. With the data mixup, the optimization objective of the novel category learning is" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 552, + 304, + 593 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 304, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 304, + 593 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathcal {T} ^ {\\mathrm {U}}} = \\frac {1}{N ^ {\\mathrm {U} , k}} \\sum_ {\\mathcal {D} ^ {\\mathrm {U}, k}} \\left(\\mathcal {L} _ {\\mathrm {S W L}} + \\mathcal {L} _ {\\mathrm {C W L}}\\right) + \\frac {1}{\\left| \\mathcal {D} ^ {\\mathrm {U} ^ {\\prime} , k} \\right|} \\sum_ {\\mathcal {D} ^ {\\mathrm {U} ^ {\\prime}, k}} \\left(\\mathcal {L} _ {\\mathrm {M I X}} + \\mathcal {L} _ {\\mathrm {C E}}\\right). \\tag {13}", + "image_path": "328b8800c3ff2ef6c5bb27de3c09cf4c3b58ac519be4731ea77791163ace4fa2.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 597, + 142, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 142, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 142, + 609 + ], + "type": "text", + "content": "4.3.4. Model Mixup" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 613, + 296, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 613, + 296, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 613, + 296, + 661 + ], + "type": "text", + "content": "In addition to data-level mixup, we also conduct a model-level mixup to compensate for the forgetting of known categories. Specifically, each client " + }, + { + "bbox": [ + 55, + 613, + 296, + 661 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 613, + 296, + 661 + ], + "type": "text", + "content": " updates " + }, + { + "bbox": [ + 55, + 613, + 296, + 661 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{k}" + }, + { + "bbox": [ + 55, + 613, + 296, + 661 + ], + "type": "text", + "content": " for each local training round:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 125, + 666, + 296, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 666, + 296, + 682 + ], + "spans": [ + { + "bbox": [ + 125, + 666, + 296, + 682 + ], + "type": "interline_equation", + "content": "\\theta^ {k} = \\beta \\theta^ {\\mathrm {L}} + (1 - \\beta) \\theta^ {k}, \\tag {14}", + "image_path": "98bf872092f51e83a67f94b8fbf8395991c9bd25f15b21500ad85f7bdcfdf3e7.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " controls the merging speed, which is set as 0.99 (please see the Appendix B.4 for sensitivity analysis). Note" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "content": "that such model mixup differs from the standard Exponential Moving Average as the first term " + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "inline_equation", + "content": "\\theta^{\\mathrm{L}}" + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "content": " here remains unchanged. As for the update of classifier " + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "inline_equation", + "content": "g_{\\omega}" + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "content": ", the neuron weights corresponding to known categories are frozen during GPA, while the neuron weights corresponding to novel categories are updated freely through back-propagation on the local side and FedAvg on the global side." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 167, + 440, + 180 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 167, + 440, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 167, + 440, + 180 + ], + "type": "text", + "content": "5. Experimental Results" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 186, + 554, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 186, + 554, + 223 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 554, + 223 + ], + "type": "text", + "content": "The main experimental setup, results, and analysis are introduced below. Please refer to Supplementary Materials for more details and all implementation source codes." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 225, + 554, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 225, + 554, + 321 + ], + "spans": [ + { + "bbox": [ + 313, + 225, + 554, + 321 + ], + "type": "text", + "content": "Datasets. CIFAR-100 [25] consists of 100 classes, and each class contains 600 images. Tiny-ImageNet [27] contains 200 classes with 500 images per class. ImageNet-S is a subset of ImageNet [8], including 100 classes and each class has 600 images. Three fine-grained datasets are also used: CUB200 [51] (200 classes), StanfordCars [24] (196 classes), and Herbarium 19 [47] (683 classes). These three datasets are usually class-imbalanced." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 324, + 554, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 554, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 554, + 552 + ], + "type": "text", + "content": "Implementation details. The number of participants is 10, and 5 are randomly selected as the clients to conduct 10-epoch local training in every global round. The number of global rounds for both " + }, + { + "bbox": [ + 313, + 324, + 554, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^{\\mathrm{L}}" + }, + { + "bbox": [ + 313, + 324, + 554, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 324, + 554, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^{\\mathrm{U}}" + }, + { + "bbox": [ + 313, + 324, + 554, + 552 + ], + "type": "text", + "content": " is 30. To simulate the non-IID setting, the data distributions of all participants follow a Dirichlet distribution [19] with the parameter " + }, + { + "bbox": [ + 313, + 324, + 554, + 552 + ], + "type": "inline_equation", + "content": "\\alpha = 0.1" + }, + { + "bbox": [ + 313, + 324, + 554, + 552 + ], + "type": "text", + "content": " (experiments with other values can be found in the Appendix). We use ResNet-18 [18] as the feature extractor for CIFAR-100 and Tiny-IIImageNet, and ResNet-34 for ImageNet-S, CUB-200, StanfordCars, and Herbarium-19. For all datasets, we employ a bias-free linear classifier. Notably, these ResNets are trained from scratch, as our goal is to evaluate the direct impact of the proposed algorithms by preventing the influence of factors such as pretraining. Certainly, we conduct experiments on Vision Transformers [11] that have been pre-trained on large-scale datasets, and the results are provided in the Appendix B.8. The batch size is 128 and the SGD learning rate is 0.02. All experiments are conducted repeatedly with seeds 2023, 2024, and 2025." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 555, + 554, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 554, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 554, + 651 + ], + "type": "text", + "content": "Evaluation metrics. For known categories, we use ground-truth labels to calculate the classification accuracy. As for novel categories, we follow [62] to adopt Hungarian clustering accuracy [26] as the metric. We also use the overall accuracy [44] calculated on known and novel categories to measure the model performance further. The accuracy forgetting on known categories after " + }, + { + "bbox": [ + 313, + 555, + 554, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^{\\mathrm{U}}" + }, + { + "bbox": [ + 313, + 555, + 554, + 651 + ], + "type": "text", + "content": " of the experiments below can be found in the Appendix B.7." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "Comparison baselines. For a fair comparison, we try our best to customize recent NCDL methods (AutoNovel [16], GM [62], iNCD [44], Happy [36], IIC [32], OpenCon [46]), novel-category estimation approaches (MACC [50], EMaCS [7]), a preliminary federated novel" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2434" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 83, + 97, + 524, + 239 + ], + "blocks": [ + { + "bbox": [ + 55, + 63, + 555, + 95 + ], + "lines": [ + { + "bbox": [ + 55, + 63, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 63, + 555, + 95 + ], + "type": "text", + "content": "Table 1. Performance comparison between GPA and other baseline methods in " + }, + { + "bbox": [ + 55, + 63, + 555, + 95 + ], + "type": "inline_equation", + "content": "{\\mathbf{{FC}}}^{2}\\mathbf{{DL}}" + }, + { + "bbox": [ + 55, + 63, + 555, + 95 + ], + "type": "text", + "content": " ,when there is only one novel category learning stage. Left precision is the mean accuracy, and the right is the variance of three runs. We bold and blue the best performance, and bold the second-best; same for all following tables." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 83, + 97, + 524, + 239 + ], + "lines": [ + { + "bbox": [ + 83, + 97, + 524, + 239 + ], + "spans": [ + { + "bbox": [ + 83, + 97, + 524, + 239 + ], + "type": "table", + "html": "
MethodCIFAR-100Tiny-ImageNetImageNet-SCUB200
knownnovelallknownnovelallknownnovelallknownnovelall
AutoNovel67.3±1.033.4±2.460.6±0.552.3±0.425.0±0.749.6±0.455.1±021.0±0.748.3±040.0±018.1±2.337.5±0.5
GM71.4±029.9±0.963.1±057.3±019.4±1.253.6±055.8±019.9±0.948.6±038.5±2.219.7±5.537.3±2.0
iNCD69.8±0.235.2±15.562.9±1.256.4±025.3±2.153.3±055.5±0.224.6±9.149.3±0.839.2±1.020.0±2.037.2±1.2
Happy70.2±1.139.2±4.464.0±0.856.4±0.327.8±2.253.5±0.255.8±029.0±1.650.4±0.340.2±0.225.8±1.138.8±0.4
IIC60.8±2.333.1±1.855.3±2.442.8±0.723.3±0.640.9±0.542.3±1.519.1±8.237.7±2.438.7±1.118.5±0.936.6±0.4
OpenCon61.5±1.718.4±1.452.9±1.547.6±2.015.6±1.744.4±2.047.4±3.918.0±10.141.6±1.029.9±3.411.2±6.728.3±2.5
Orchestra66.6±0.533.1±0.860.0±0.253.4±0.324.7±1.150.1±0.152.8±0.321.5±0.348.4±040.0±018.5±1.037.5±0.1
SemiFL44.2±0.229.1±0.541.2±0.233.5±0.125.7±1.132.7±0.638.8±0.227.5±0.536.5±0.230.0±0.718.7±1.329.1±0.6
FedoSSL62.2±0.218.2±0.753.4±0.453.0±0.18.7±1.048.6±0.455.0±017.5±0.247.5±0.138.8±014.5±0.736.9±0.3
AGCL65.6±0.235.5±0.559.6±0.243.2±0.524.5±1.441.3±0.746.9±0.321.0±0.641.7±0.336.0±020.5±1.134.8±0.7
GPA71.9±052.3±1.068.0±0.457.6±039.1±1.955.8±0.755.8±037.4±1.052.1±0.440.3±0.231.5±0.739.4±0.3
", + "image_path": "024508124a7d109603533f4f783d9571cde04f5b01bd2778534b670311515b78.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 55, + 274, + 555, + 453 + ], + "blocks": [ + { + "bbox": [ + 55, + 250, + 553, + 271 + ], + "lines": [ + { + "bbox": [ + 55, + 250, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 55, + 250, + 553, + 271 + ], + "type": "text", + "content": "Table 2. Performance comparison between GPA and other baseline methods in " + }, + { + "bbox": [ + 55, + 250, + 553, + 271 + ], + "type": "inline_equation", + "content": "{\\mathbf{{FC}}}^{2}\\mathbf{{DL}}" + }, + { + "bbox": [ + 55, + 250, + 553, + 271 + ], + "type": "text", + "content": " ,when there are five novel category learning stage. Left precision is the mean accuracy, and the right is the variance of three runs." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 274, + 555, + 453 + ], + "lines": [ + { + "bbox": [ + 55, + 274, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 555, + 453 + ], + "type": "table", + "html": "
DatasetMethod1st Stage2nd Stage3rd Stage4th Stage5th Stage
knownnovelallknownnovelallknownnovelallknownnovelallknownnovelall
CIFAR-100GM72.9±1.340.2±2.469.1±1.571.5±0.837.8±1.464.4±0.970.7±1.031.4±2.759.5±1.369.5±0.627.9±1.955.0±1.168.8±2.526.0±2.051.7±1.7
iNCD71.5±0.544.4±1.768.3±0.668.9±1.241.6±1.563.2±0.866.0±1.334.5±1.457.0±1.165.7±0.833.0±1.754.3±0.965.0±1.031.2±1.951.5±1.3
Happy72.0±0.545.6±1.168.9±0.768.8±0.944.7±1.563.7±0.965.8±0.741.0±2.158.7±1.064.5±1.338.7±0.855.5±1.063.7±1.035.5±2.052.4±1.2
GPA74.0±057.8±1.472.1±0.673.5±0.656.5±1.369.9±0.773.3±0.553.0±1.067.5±0.473.0±0.951.7±1.665.6±1.172.5±1.048.2±2.362.8±1.5
T-ImageNetGM58.3±0.237.0±1.157.3±0.356.6±0.423.4±3.453.6±0.855.8±1.720.2±2.351.2±1.854.0±0.818.8±2.048.1±1.252.5±1.718.0±2.445.6±1.9
iNCD58.0±0.538.2±0.757.1±0.553.1±1.031.0±1.451.1±1.150.8±0.925.0±1.947.4±1.147.7±2.623.2±1.443.6±2.945.3±1.122.1±1.840.7±1.3
Happy58.3±0.236.5±1.557.3±0.655.7±0.930.9±2.153.4±1.353.4±1.228.0±1.450.1±1.252.0±1.126.3±1.447.7±1.251.0±0.724.9±1.845.8±1.1
GPA58.5±0.248.2±1.658.0±0.658.0±0.543.5±2.256.7±0.957.3±0.739.5±1.255.0±0.957.0±0.638.0±2.153.8±1.056.8±1.136.9±2.552.8±1.5
ImageNet-SGM57.0±037.2±0.454.7±0.156.5±0.226.7±1.950.2±0.854.8±0.519.5±1.444.7±0.653.0±0.518.2±1.040.9±0.652.9±1.517.3±1.838.7±1.2
iNCD55.5±0.236.2±0.753.2±0.354.1±0.430.2±1.349.1±0.552.0±0.324.5±0.844.1±0.449.9±1.422.3±1.940.3±1.548.0±1.020.2±1.536.9±1.2
Happy56.5±0.437.9±1.454.3±0.556.0±0.833.7±2.251.3±1.153.5±1.629.2±2.446.6±1.952.0±1.027.5±1.643.5±1.351.4±1.826.0±2.741.2±2.2
GPA57.2±046.0±1.155.9±0.457.0±0.542.2±0.753.9±0.556.4±0.937.0±2.050.9±1.455.7±1.235.1±1.548.5±1.255.5±1.332.9±3.046.5±1.9
CUB200GM42.5±0.328.9±0.941.9±0.442.0±0.523.7±1.040.3±0.641.0±0.719.5±1.438.2±0.940.2±0.518.1±1.236.5±0.738.5±0.617.7±1.634.3±0.9
iNCD41.0±0.534.0±0.940.7±0.539.8±0.624.6±1.838.4±1.037.0±1.020.0±2.034.8±1.236.5±1.118.9±1.733.6±1.333.8±0.917.9±1.430.6±1.0
Happy42.2±0.733.9±1.141.8±0.841.1±0.529.5±1.640.0±0.739.8±0.525.5±1.237.9±0.739.0±1.023.3±1.436.4±1.137.8±0.823.0±2.034.8±1.2
GPA42.8±0.238.0±0.942.6±0.342.5±0.735.3±1.441.8±0.942.2±0.731.0±1.940.7±1.041.8±0.429.4±1.239.7±0.641.5±0.828.8±1.739.0±1.0
", + "image_path": "6cb11a414ba33bea2cdb73ad46f923d75ac600816698c9a881ff0a71a60ad854.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 473, + 295, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 295, + 534 + ], + "type": "text", + "content": "category discovery work (AGCL [40]), federated self/semi-supervised learning methods (Orchestra [35], SemiFL [9], FedoSSL [61]), and popular FL algorithms (FedProx [31], SCAFFOLD [23], Moon [29]) reasonably in the setting of " + }, + { + "bbox": [ + 55, + 473, + 295, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{FC}^2\\mathbf{DL}" + }, + { + "bbox": [ + 55, + 473, + 295, + 534 + ], + "type": "text", + "content": ", and compare them with our GPA." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 545, + 218, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 218, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 218, + 558 + ], + "type": "text", + "content": "5.1. Effectiveness of GPA on " + }, + { + "bbox": [ + 55, + 545, + 218, + 558 + ], + "type": "inline_equation", + "content": "\\mathsf{FC^2DL}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "Comparison on one novel category learning stage. We first assume that there is only one " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^{\\mathrm{U}}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathsf{FC^{2}DL}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": ", which is consistent with conventional NCDL studies. For CIFAR100, Tiny-ImageNet, ImageNet-S, CUB200, and StanfordCars, we set the novel category number as 20 and 83 for Herbarium 19, while the rest are regarded as known categories. The experiment results are shown in Tables 1 and 5. The results demonstrate that our GPA framework always achieves the best performance on all metrics, including accuracy on known, novel, and all categories. The performance improvement on novel-category accuracy is particularly promising, ranging from " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "5.7\\%" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " on CUB200 to" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 473, + 555, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 555, + 497 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 555, + 497 + ], + "type": "inline_equation", + "content": "13.1\\%" + }, + { + "bbox": [ + 313, + 473, + 555, + 497 + ], + "type": "text", + "content": " on CIFAR-100 over the best baseline method, which demonstrates the effectiveness of GPA." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 504, + 555, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 555, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 555, + 685 + ], + "type": "text", + "content": "Comparison on multiple novel category learning stages. As mentioned in related works [10], FL participants often collect unseen novel-category data continually. For such cases, we split each dataset into a labeled subset and 5 unlabeled subsets (each contains 8 novel categories) with disjoint categories, and conduct five " + }, + { + "bbox": [ + 313, + 504, + 555, + 685 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^{\\mathrm{U}}" + }, + { + "bbox": [ + 313, + 504, + 555, + 685 + ], + "type": "text", + "content": " in sequence. For a fair comparison, we only compare GPA with GM [62], iNCD [44], and Happy [36] as only they have dedicated designs for multiple novel category learning stages. All experiments are conducted using the category number estimated by our PPM and the detailed results are shown in Table 2. We can observe that GPA achieves the best performance on all datasets, showing " + }, + { + "bbox": [ + 313, + 504, + 555, + 685 + ], + "type": "inline_equation", + "content": "4.1\\%" + }, + { + "bbox": [ + 313, + 504, + 555, + 685 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 504, + 555, + 685 + ], + "type": "inline_equation", + "content": "13.0\\%" + }, + { + "bbox": [ + 313, + 504, + 555, + 685 + ], + "type": "text", + "content": " improvement on novel categories compared with the best-performing approach across 1st and 5th novel category learning stages." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Comparison on estimating novel-category number. Furthermore, to validate the effectiveness of PPM, we com" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2435" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 84, + 95, + 526, + 147 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 555, + 93 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 555, + 93 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 555, + 93 + ], + "type": "text", + "content": "Table 3. Performance comparison between PPM and other novel category number estimation approaches in various novel category number settings. The experiments here show that PPM can consistently provide the most accurate novel category number estimation." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 84, + 95, + 526, + 147 + ], + "lines": [ + { + "bbox": [ + 84, + 95, + 526, + 147 + ], + "spans": [ + { + "bbox": [ + 84, + 95, + 526, + 147 + ], + "type": "table", + "html": "
DatasetCIFAR-100Tiny-ImageNetImageNet-SCUB200StanfordCarsHerbarium19Average\nError (%)
Class #5101520253010203040505101520102030401020304020406080
MACC4566141568122228466144613186122024918243048.9
EMaCS18173023121127452526582120341521353429194542362517384679.5
PPM4913192529819314054581217921324411202839194265868.1
", + "image_path": "2d523bdd3c2681d91e2fc8b3b1be47e57281e892595ec92ba64b34ed156883d2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 160, + 550, + 258 + ], + "blocks": [ + { + "bbox": [ + 58, + 160, + 550, + 258 + ], + "lines": [ + { + "bbox": [ + 58, + 160, + 550, + 258 + ], + "spans": [ + { + "bbox": [ + 58, + 160, + 550, + 258 + ], + "type": "image", + "image_path": "c39bf1041eda92e63752d7b86853d95ee88280decbe7c25064bccf7bb3be6f95.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 259, + 555, + 282 + ], + "lines": [ + { + "bbox": [ + 55, + 259, + 555, + 282 + ], + "spans": [ + { + "bbox": [ + 55, + 259, + 555, + 282 + ], + "type": "text", + "content": "Figure 2. Performance evaluation of combining GPA with other FL algorithms. We assume that Kmeans can equip FedProx, SCAFFOLD, and Moon with a certain level of NCDL capability, while GPA can substantially exceed it. The FC" + }, + { + "bbox": [ + 55, + 259, + 555, + 282 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 55, + 259, + 555, + 282 + ], + "type": "text", + "content": "DL setting is the same as Table 1." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 55, + 316, + 294, + 414 + ], + "blocks": [ + { + "bbox": [ + 55, + 293, + 295, + 314 + ], + "lines": [ + { + "bbox": [ + 55, + 293, + 295, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 295, + 314 + ], + "type": "text", + "content": "Table 4. Ablation study of major components in GPA. The " + }, + { + "bbox": [ + 55, + 293, + 295, + 314 + ], + "type": "inline_equation", + "content": "FC^2DL" + }, + { + "bbox": [ + 55, + 293, + 295, + 314 + ], + "type": "text", + "content": " setting is the same as Table 1." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 316, + 294, + 414 + ], + "lines": [ + { + "bbox": [ + 55, + 316, + 294, + 414 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 294, + 414 + ], + "type": "table", + "html": "
MethodCIFAR-100Tiny-ImageNetImageNet-S
knownnovelallknownnovelallknownnovelall
GPA-w/o LpCL70.251.166.454.337.052.653.835.950.2
GPA-w/o LSWL71.837.364.957.631.155.055.830.250.7
GPA-w/o LCWL71.942.966.157.633.055.155.832.751.2
GPA-w/o LMix71.845.266.557.634.355.355.734.051.4
GPA-w/ MACC71.539.865.257.627.954.655.729.650.5
GPA-w/o M.Mix69.848.965.655.238.253.553.236.349.8
GPA71.952.368.057.639.155.855.837.452.1
", + "image_path": "69c758672d66d35b5bc8a0c3d1209224a0a4f115cac0b6a00162e203434ee3a5.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 438, + 295, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 438, + 295, + 535 + ], + "spans": [ + { + "bbox": [ + 54, + 438, + 295, + 535 + ], + "type": "text", + "content": "pare it with state-of-the-art approaches (MACC [50] and EMaCS [7]) when estimating different novel category numbers in " + }, + { + "bbox": [ + 54, + 438, + 295, + 535 + ], + "type": "inline_equation", + "content": "\\mathsf{FC^2DL}" + }, + { + "bbox": [ + 54, + 438, + 295, + 535 + ], + "type": "text", + "content": ". The detailed category number setups and experiment results are shown in Table 3. We can observe that our PPM provides the most accurate category number estimation for all datasets, with an average error rate of " + }, + { + "bbox": [ + 54, + 438, + 295, + 535 + ], + "type": "inline_equation", + "content": "8.1\\%" + }, + { + "bbox": [ + 54, + 438, + 295, + 535 + ], + "type": "text", + "content": ", while the other approaches have approximately " + }, + { + "bbox": [ + 54, + 438, + 295, + 535 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 54, + 438, + 295, + 535 + ], + "type": "text", + "content": " and even higher average error rate." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 540, + 295, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 540, + 295, + 648 + ], + "spans": [ + { + "bbox": [ + 54, + 540, + 295, + 648 + ], + "type": "text", + "content": "Experiments with more FL algorithms. To further evaluate the applicability of GPA, we also carry out experiments combining GPA with FedProx [31], SCAFFOLD [23], and Moon [29]. Specifically, we assume that regular FL can possess a certain level of NCDL capability through unsupervised clustering (e.g., Kmeans) on novel-category data representations. The experiment results are shown in Figure 2. We can see that using GPA with these FL algorithms provides much better performance on novel categories." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "More experiments in the Appendix. 1) experiments in the centralized setting—Section B.1; 2) experiments with various degrees of heterogeneity in " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{FC}^2\\mathbf{DL}" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "—Section B.2; 3) experiments with different data category partitions—Section B.3; 4) sensitivity analysis of hyper-parameters—" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 302, + 555, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 302, + 555, + 362 + ], + "spans": [ + { + "bbox": [ + 313, + 302, + 555, + 362 + ], + "type": "text", + "content": "Section B.4; 5) experiments with a larger number of participants—Section B.5 6) experiments on pre-trained vision transformers—Section B.8; 7) experiments of launching data reconstruction attacks for verifying GPA's characteristic of privacy-preserving—Section B.6." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 373, + 408, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 408, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 408, + 385 + ], + "type": "text", + "content": "5.2. Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 391, + 556, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 391, + 556, + 547 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 556, + 547 + ], + "type": "text", + "content": "We investigate the importance of each component in GPA via an ablation study built-in " + }, + { + "bbox": [ + 313, + 391, + 556, + 547 + ], + "type": "inline_equation", + "content": "\\mathbf{FC}^2\\mathbf{DL}" + }, + { + "bbox": [ + 313, + 391, + 556, + 547 + ], + "type": "text", + "content": " with one novel category learning stage. As shown in Table 4, 'GPA-w/o " + }, + { + "bbox": [ + 313, + 391, + 556, + 547 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{PCL}}" + }, + { + "bbox": [ + 313, + 391, + 556, + 547 + ], + "type": "text", + "content": " 'GPA-w/o " + }, + { + "bbox": [ + 313, + 391, + 556, + 547 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{SWL}}" + }, + { + "bbox": [ + 313, + 391, + 556, + 547 + ], + "type": "text", + "content": " ,GPA-w/o " + }, + { + "bbox": [ + 313, + 391, + 556, + 547 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CWL}}" + }, + { + "bbox": [ + 313, + 391, + 556, + 547 + ], + "type": "text", + "content": " ,GPA-w/o " + }, + { + "bbox": [ + 313, + 391, + 556, + 547 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{MIX}}" + }, + { + "bbox": [ + 313, + 391, + 556, + 547 + ], + "type": "text", + "content": " GPA-w/MACC, 'GPA-w/o M.Mix' denote the performance of GPA without using PCL in " + }, + { + "bbox": [ + 313, + 391, + 556, + 547 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^{\\mathrm{L}}" + }, + { + "bbox": [ + 313, + 391, + 556, + 547 + ], + "type": "text", + "content": " , without using SWL (replacing it with a Binary CrossEntropy Loss [16]), without using CWL, without using data mixup, with MACC instead of PPM to estimate category number, and without applying model mixup (replacing it with distillation [44]), respectively. We can observe that these versions of GPA degrade evidently in performance, which shows the importance of all these modules in our approach." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 560, + 388, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 560, + 388, + 573 + ], + "spans": [ + { + "bbox": [ + 313, + 560, + 388, + 573 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "In this paper, we study an important and practical problem called Federated Continuous Category Discovery and Learning (FC" + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "DL), and propose Global Prototype Alignment (GPA) to help the FL model discover and learn novel categories. GPA can accurately estimate the novel category number, construct corresponding global prototypes, and guide the local training of different participants to tackle data heterogeneity. Extensive experiments demonstrate that GPA significantly outperforms state-of-the-art novel category learning methods and other potential solutions in various scenarios for the FC" + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "DL problem." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2436" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 212 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 212 + ], + "type": "text", + "content": "This research is supported by the National Research Foundation, Singapore, and the Cyber Security Agency of Singapore under the National Cybersecurity R&D Programme and the CyberSG R&D Programme Office (Award CRPO-GC3-NTU-001) and NTU-NAP startup grant (024584-00001). Any opinions, findings, conclusions, or recommendations expressed in these materials are those of the author(s) and do not reflect the views of the National Research Foundation, Singapore, the Cyber Security Agency of Singapore, or the CyberSG R&D Programme Office." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 222, + 115, + 235 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 222, + 115, + 235 + ], + "spans": [ + { + "bbox": [ + 56, + 222, + 115, + 235 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 242, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 242, + 296, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 242, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 61, + 242, + 296, + 285 + ], + "type": "text", + "content": "[1] Wenbin An, Feng Tian, Qinghua Zheng, Wei Ding, QianYing Wang, and Ping Chen. Generalized category discovery with decoupled prototypical network. arXiv preprint arXiv:2211.15115, 2022. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 286, + 296, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 286, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 61, + 286, + 296, + 319 + ], + "type": "text", + "content": "[2] Kaidi Cao, Maria Brbic, and Jure Leskovec. Open-world semi-supervised learning. In International Conference on Learning Representations (ICLR), 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 321, + 296, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 321, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 61, + 321, + 296, + 365 + ], + "type": "text", + "content": "[3] Fernando Julio Cendra, Bingchen Zhao, and Kai Han. Promptcd: Learning gaussian mixture prompt pool for continual category discovery. In European Conference on Computer Vision, pages 188-205. Springer, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 366, + 296, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 366, + 296, + 409 + ], + "spans": [ + { + "bbox": [ + 62, + 366, + 296, + 409 + ], + "type": "text", + "content": "[4] Mingzhe Chen, Nir Shlezinger, H Vincent Poor, Yonina C Eldar, and Shuguang Cui. Communication-efficient federated learning. Proceedings of the National Academy of Sciences, 118(17):e2024789118, 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 411, + 296, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 411, + 296, + 455 + ], + "spans": [ + { + "bbox": [ + 62, + 411, + 296, + 455 + ], + "type": "text", + "content": "[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PmLR, 2020. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 456, + 296, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 456, + 296, + 511 + ], + "spans": [ + { + "bbox": [ + 62, + 456, + 296, + 511 + ], + "type": "text", + "content": "[6] Haoang Chi, Feng Liu, Wenjing Yang, Long Lan, Tongliang Liu, Bo Han, Gang Niu, Mingyuan Zhou, and Masashi Sugiyama. Meta discovery: Learning to discover novel classes given very limited data. In International Conference on Learning Representations. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 512, + 296, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 512, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 62, + 512, + 296, + 567 + ], + "type": "text", + "content": "[7] Florent Chiaroni, Jose Dolz, Ziko Imtiaz Masud, Amar Mitiche, and Ismail Ben Ayed. Parametric information maximization for generalized category discovery. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1729-1739, 2023. 2, 5, 6, 8, 13" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 568, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 568, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 568, + 296, + 612 + ], + "type": "text", + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 613, + 296, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 613, + 296, + 657 + ], + "spans": [ + { + "bbox": [ + 62, + 613, + 296, + 657 + ], + "type": "text", + "content": "[9] Enmao Diao, Jie Ding, and Vahid Tarokh. Semifl: Semi-supervised federated learning for unlabeled clients with alternate training. Advances in Neural Information Processing Systems, 35:17871-17884, 2022. 6, 7, 13" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 658, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 712 + ], + "type": "text", + "content": "[10] Jiahua Dong, Lixu Wang, Zhen Fang, Gan Sun, Shichao Xu, Xiao Wang, and Qi Zhu. Federated class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10164-10173, 2022. 2, 4, 5, 7, 12" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2020. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 162 + ], + "type": "text", + "content": "[12] Aude Geneva, Gabriel Dulac-Arnold, and Jean-Philippe Vert. Differentiable deep clustering with cluster size constraints. arXiv preprint arXiv:1910.09036, 2019. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 163, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 206 + ], + "type": "text", + "content": "[13] Avishek Ghosh, Jichan Chung, Dong Yin, and Kannan Ramchandran. An efficient framework for clustered federated learning. Advances in Neural Information Processing Systems, 33:19586-19597, 2020. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 208, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 208, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 208, + 553, + 262 + ], + "type": "text", + "content": "[14] Junfeng Guo, Yiming Li, Lixu Wang, Shu-Tao Xia, Heng Huang, Cong Liu, and Bo Li. Domain watermark: Effective and harmless dataset copyright protection is closed at hand. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "text", + "content": "[15] Lan-Zhe Guo, Yi-Ge Zhang, Zhi-Fan Wu, Jie-Jing Shao, and Yu-Feng Li. Robust semi-supervised learning when not all classes have labels. Advances in Neural Information Processing Systems, 35:3305-3317, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 309, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 363 + ], + "type": "text", + "content": "[16] Kai Han, Sylvestre-Alvise Rebuffi, Sebastien Ehrhardt, Andrea Vedaldi, and Andrew Zisserman. Autonovel: Automatically discovering and learning novel visual categories. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6767-6781, 2021. 2, 5, 6, 8, 12, 14" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 365, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 408 + ], + "type": "text", + "content": "[17] John A Hartigan and Manchek A Wong. Algorithm as 136: A k-means clustering algorithm. Journal of the royal statistical society. series c (applied statistics), 28(1):100-108, 1979. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 410, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 410, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 410, + 553, + 453 + ], + "type": "text", + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 455, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 455, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 553, + 498 + ], + "type": "text", + "content": "[19] Tzu-Ming Harry Hsu, Hang Qi, and Matthew Brown. Measuring the effects of non-identical data distribution for federated visual classification. arXiv preprint arXiv:1909.06335, 2019. 3, 6, 12" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 500, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 553, + 544 + ], + "type": "text", + "content": "[20] Wenke Huang, Mang Ye, and Bo Du. Learn from others and be yourself in heterogeneous federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10143-10153, 2022. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 545, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 599 + ], + "type": "text", + "content": "[21] Wenke Huang, Mang Ye, Zekun Shi, He Li, and Bo Du. Rethinking federated learning with domain shift: A prototype view. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16312-16322. IEEE, 2023. 16" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 601, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 633 + ], + "type": "text", + "content": "[22] Liangze Jiang and Tao Lin. Test-time robust personalization for federated learning. arXiv preprint arXiv:2205.10920, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "text", + "content": "[23] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank Reddi, Sebastian Stich, and Ananda Theertha Suresh. Scaffold: Stochastic controlled averaging for federated learning. In International conference on machine learning, pages 5132-5143. PMLR, 2020. 7, 8, 12, 13" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "text", + "content": "[24] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2437" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "text", + "content": "Proceedings of the IEEE international conference on computer vision workshops, pages 554-561, 2013. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 295, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 295, + 118 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 295, + 118 + ], + "type": "text", + "content": "[25] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 119, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 119, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 119, + 295, + 150 + ], + "type": "text", + "content": "[26] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 153, + 294, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 153, + 294, + 175 + ], + "spans": [ + { + "bbox": [ + 57, + 153, + 294, + 175 + ], + "type": "text", + "content": "[27] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3, 2015. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 176, + 295, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 176, + 295, + 219 + ], + "spans": [ + { + "bbox": [ + 56, + 176, + 295, + 219 + ], + "type": "text", + "content": "[28] Junnan Li, Pan Zhou, Caiming Xiong, and Steven Hoi. Prototypical contrastive learning of unsupervised representations. In International Conference on Learning Representations, 2020. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 221, + 294, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 221, + 294, + 264 + ], + "spans": [ + { + "bbox": [ + 56, + 221, + 294, + 264 + ], + "type": "text", + "content": "[29] Qinbin Li, Bingsheng He, and Dawn Song. Model-contrastive federated learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10713–10722, 2021. 7, 8, 12, 13" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 266, + 294, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 266, + 294, + 320 + ], + "spans": [ + { + "bbox": [ + 56, + 266, + 294, + 320 + ], + "type": "text", + "content": "[30] Qinbin Li, Zeyi Wen, Zhaomin Wu, Sixu Hu, Naibo Wang, Yuan Li, Xu Liu, and Bingsheng He. A survey on federated learning systems: Vision, hype and reality for data privacy and protection. IEEE Transactions on Knowledge and Data Engineering, 2021. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 322, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 322, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 322, + 294, + 365 + ], + "type": "text", + "content": "[31] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. Proceedings of Machine learning and systems, 2:429-450, 2020. 7, 8, 12, 13" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 367, + 294, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 367, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 367, + 294, + 421 + ], + "type": "text", + "content": "[32] Wenbin Li, Zhichen Fan, Jing Huo, and Yang Gao. Modeling inter-class and intra-class constraints in novel class discovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 4, 5, 6, 13" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 423, + 294, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 423, + 294, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 423, + 294, + 477 + ], + "type": "text", + "content": "[33] Xiaoxiao Liang, Yiqun Lin, Huazhu Fu, Lei Zhu, and Xiaomeng Li. Rscfed: Random sampling consensus federated semi-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10154-10163, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 479, + 294, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 294, + 522 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 294, + 522 + ], + "type": "text", + "content": "[34] Chenxi Liu, Lixu Wang, Lingjuan Lyu, Chen Sun, Xiao Wang, and Qi Zhu. Deja vu: Continual model generalization for unseen domains. In The Eleventh International Conference on Learning Representations, 2022. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "type": "text", + "content": "[35] Ekdeep Lubana, Chi Ian Tang, Fahim Kawsar, Robert Dick, and Akhil Mathur. Orchestra: Unsupervised federated learning via globally consistent clustering. In International Conference on Machine Learning, pages 14461-14484. PMLR, 2022. 2, 7, 13" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 579, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 634 + ], + "type": "text", + "content": "[36] Shijie Ma, Fei Zhu, Zhun Zhong, Wenzhuo Liu, Xu-Yao Zhang, and Cheng-Lin Liu. Happy: A debiased learning framework for continual generalized category discovery. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2, 6, 7, 13, 17" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "text", + "content": "[37] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 1, 2, 3, 4, 12" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[38] A Tuan Nguyen, Philip Torr, and Ser Nam Lim. Fedsr: A simple and effective domain generalization method for fed-" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "erated learning. Advances in Neural Information Processing Systems, 35:38831-38843, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 95, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 553, + 139 + ], + "type": "text", + "content": "[39] Keon-Hee Park, Hakyung Lee, Kyungwoo Song, and Gyeong-Moon Park. Online continuous generalized category discovery. In European Conference on Computer Vision, pages 53–69. Springer, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 140, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 182 + ], + "type": "text", + "content": "[40] Nan Pu, Zhun Zhong, Xinyuan Ji, and Nicu Sebe. Federated generalized category discovery. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2024. 2, 7, 13" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 184, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 217 + ], + "type": "text", + "content": "[41] Daiqing Qi, Handong Zhao, and Sheng Li. Better generative replay for continual federated learning. In The Eleventh International Conference on Learning Representations. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 217, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 271 + ], + "type": "text", + "content": "[42] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 2001-2010, 2017. 2, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 272, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 327 + ], + "type": "text", + "content": "[43] Nicola Rieke, Jonny Hancox, Wenqi Li, Fausto Miletari, Holger R Roth, Shadi Albarqouni, Spyridon Bakas, Mathieu N Galtier, Bennett A Landman, Klaus Maier-Hein, et al. The future of digital health with federated learning. NPJ digital medicine, 3(1):119, 2020. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 327, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 392 + ], + "type": "text", + "content": "[44] Subhankar Roy, Mingxuan Liu, Zhun Zhong, Nicu Sebe, and Elisa Ricci. Class-incremental novel class discovery. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIII, pages 317-333. Springer, 2022. 1, 2, 4, 5, 6, 7, 8, 13, 14, 15" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 393, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 553, + 425 + ], + "type": "text", + "content": "[45] Erich Schubert and et. al. Dbscan revisited, revisited: why and how you should (still) use dbscan. ACM Transactions on Database Systems, 2017. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 426, + 553, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 458 + ], + "type": "text", + "content": "[46] Yiyou Sun and Yixuan Li. Opencon: Open-world contrastive learning. In Transactions on Machine Learning Research, 2022. 2, 6, 13" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 460, + 553, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 492 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 492 + ], + "type": "text", + "content": "[47] Kiat Chuan Tan, Yulong Liu, Barbara Ambrose, Melissa Tulig, and Serge Belongie. The herbarium challenge 2019 dataset. arXiv preprint arXiv:1906.05372, 2019. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 493, + 555, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 555, + 547 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 555, + 547 + ], + "type": "text", + "content": "[48] Yue Tan, Guodong Long, Lu Liu, Tianyi Zhou, Qinghua Lu, Jing Jiang, and Chengqi Zhang. Fedproto: Federated prototype learning across heterogeneous clients. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8432-8440, 2022. 16" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 548, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 592 + ], + "type": "text", + "content": "[49] Colin Troisemaine, Vincent Lemaire, Stéphane Gosselin, Alexandre Reiffers-Masson, Joachim Flocon-Cholet, and Sandrine Vaton. Novel class discovery: an introduction and key concepts. arXiv preprint arXiv:2302.12028, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 593, + 555, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 555, + 635 + ], + "type": "text", + "content": "[50] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Generalized category discovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7492-7501, 2022. 2, 5, 6, 8, 13" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 636, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 669 + ], + "type": "text", + "content": "[51] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. Caltech-ucsd birds-200-2011. Technical Report CNS-TR-2011-001, California Institute of Technology, 2011. 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "type": "text", + "content": "[52] Lixu Wang, Shichao Xu, Xiao Wang, and Qi Zhu. Addressing class imbalance in federated learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10165–10173, 2021. 5" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2438" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 676 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[53] Lixu Wang, Shichao Xu, Ruiqi Xu, Xiao Wang, and Qi Zhu. Non-transferable learning: A new approach for model ownership verification and applicability authorization. In International Conference on Learning Representations, 2021. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "text", + "content": "[54] Lixu Wang, Yang Zhao, Jiahua Dong, Ating Yin, Qinbin Li, Xiao Wang, Dusit Niyato, and Qi Zhu. Federated learning with new knowledge: Fundamentals, advances, and futures. arXiv preprint arXiv:2402.02268, 2024. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 205 + ], + "type": "text", + "content": "[55] Jie Xu, Benjamin S Glicksberg, Chang Su, Peter Walker, Jiang Bian, and Fei Wang. Federated learning for healthcare informatics. Journal of Healthcare Informatics Research, 5: 1-19, 2021. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "text", + "content": "[56] Muli Yang, Yuehua Zhu, Jiaping Yu, Among Wu, and Cheng Deng. Divide and conquer: Compositional experts for generalized novel class discovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14268-14277, 2022. 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 263, + 294, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 294, + 318 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 294, + 318 + ], + "type": "text", + "content": "[57] Xufeng Yao, Yang Bai, Xinyun Zhang, Yuechen Zhang, Qi Sun, Ran Chen, Ruiyu Li, and Bei Yu. Pcl: Proxy-based contrastive learning for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7097-7107, 2022. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 319, + 294, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 294, + 372 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 294, + 372 + ], + "type": "text", + "content": "[58] Jaehong Yoon, Wonyong Jeong, Giwoong Lee, Eunho Yang, and Sung Ju Hwang. Federated continual learning with weighted inter-client transfer. In International Conference on Machine Learning, pages 12073-12086. PMLR, 2021. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 375, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 375, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 375, + 294, + 430 + ], + "type": "text", + "content": "[59] Shuyang Yu, Junyuan Hong, Haotao Wang, Zhangyang Wang, and Jiayu Zhou. Turning the curse of heterogeneity in federated learning into a blessing for out-of-distribution detection. In The Eleventh International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 431, + 294, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 431, + 294, + 474 + ], + "spans": [ + { + "bbox": [ + 56, + 431, + 294, + 474 + ], + "type": "text", + "content": "[60] Fengda Zhang, Kun Kuang, Zhaoyang You, Tao Shen, Jun Xiao, Yin Zhang, Chao Wu, Yueting Zhuang, and Xiaolin Li. Federated unsupervised representation learning. arXiv preprint arXiv:2010.08982, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 476, + 294, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 476, + 294, + 508 + ], + "spans": [ + { + "bbox": [ + 56, + 476, + 294, + 508 + ], + "type": "text", + "content": "[61] Jie Zhang, Xiaosong Ma, Song Guo, and Wenchao Xu. Towards unbiased training in federated open-world semi-supervised learning. ICML, 2023. 2, 7, 13" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 510, + 294, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 510, + 294, + 564 + ], + "spans": [ + { + "bbox": [ + 56, + 510, + 294, + 564 + ], + "type": "text", + "content": "[62] Xinwei Zhang, Jianwen Jiang, Yutong Feng, Zhi-Fan Wu, Xibin Zhao, Hai Wan, Mingqian Tang, Rong Jin, and Yue Gao. Grow and merge: A unified framework for continuous categories discovery. In Advances in Neural Information Processing Systems, 2022. 1, 2, 4, 5, 6, 7, 13, 15" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 566, + 294, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 566, + 294, + 597 + ], + "spans": [ + { + "bbox": [ + 56, + 566, + 294, + 597 + ], + "type": "text", + "content": "[63] Bingchen Zhao and Oisin Mac Aodha. Incremental generalized category discovery. arXiv preprint arXiv:2304.14310, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 599, + 294, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 599, + 294, + 643 + ], + "spans": [ + { + "bbox": [ + 56, + 599, + 294, + 643 + ], + "type": "text", + "content": "[64] Haiyang Zheng, Nan Pu, Wenjing Li, Nicu Sebe, and Zhun Zhong. Prototypical hash encoding for on-the-fly fine-grained category discovery. Advances in Neural Information Processing Systems, 37:101428-101455, 2025. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 644, + 294, + 676 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 644, + 294, + 676 + ], + "spans": [ + { + "bbox": [ + 56, + 644, + 294, + 676 + ], + "type": "text", + "content": "[65] Ligeng Zhu, Zhijian Liu, and Song Han. Deep leakage from gradients. Advances in neural information processing systems, 32, 2019. 17" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2439" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Federated Domain Generalization with Domain-specific Soft Prompts Generation/b70289dd-c48a-4e53-8732-8d816d04b0c4_content_list.json b/2025/Federated Domain Generalization with Domain-specific Soft Prompts Generation/b70289dd-c48a-4e53-8732-8d816d04b0c4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9cb411af69d109ce6ec973547dd3a2ffe896e383 --- /dev/null +++ b/2025/Federated Domain Generalization with Domain-specific Soft Prompts Generation/b70289dd-c48a-4e53-8732-8d816d04b0c4_content_list.json @@ -0,0 +1,1361 @@ +[ + { + "type": "text", + "text": "Federated Domain Generalization with Domain-specific Soft Prompts Generation", + "text_level": 1, + "bbox": [ + 122, + 128, + 875, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jianhan Wu * Xiaoyang Qu * Zhangcheng Huang Jianzong Wang † Ping An Technology (Shenzhen) Co., Ltd.", + "bbox": [ + 196, + 202, + 807, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "wujianhan@mail.ustc.edu.cn, quxiaoy@gmail.com", + "bbox": [ + 295, + 242, + 699, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "william.huang@connect.polyu.hk, jzwang@188.com", + "bbox": [ + 292, + 258, + 700, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 328, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Prompt learning has become an efficient paradigm for adapting CLIP to downstream tasks. Compared with traditional fine-tuning, prompt learning optimizes a few parameters yet yields highly competitive results, especially appealing in federated learning for computational efficiency. engendering domain shift among clients and posing a formidable challenge for downstream-task adaptation. Existing federated domain generalization (FDG) methods based on prompt learning typically learn soft prompts from training samples, replacing manually designed prompts to enhance the generalization ability of federated models. However, these learned prompts exhibit limited diversity and tend to ignore information from unknown domains. We propose a novel and effective method from a generative perspective for handling FDG tasks, namely federated domain generalization with domain-specific soft prompts generation (FedDSPG). Specifically, during training, we introduce domain-specific soft prompts (DSPs) for each domain and integrate content and domain knowledge into the generative model among clients. In the inference phase, the generator is utilized to obtain DSPs for unseen target domains, thus guiding downstream tasks in unknown domains. Comprehensive evaluations across several public datasets confirm that our method outperforms existing strong baselines in FDG, achieving state-of-the-art results.", + "bbox": [ + 91, + 339, + 483, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 746, + 228, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pre-trained vision-language models (VLMs) [13], such as Contrastive Language Image Pretraining (CLIP) [12, 18, 27], have emerged as a powerful technique to", + "bbox": [ + 89, + 771, + 482, + 819 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "adapt to downstream tasks. By introducing learnable prompts [1, 9, 37] into the text or vision encoders of CLIP, this approach allows the model to dynamically adjust its representations with minimal fine-tuning. Unlike traditional fine-tuning methods [35] that require extensive parameter updates, prompt learning optimizes only a small fraction of parameters, significantly reducing computational costs while maintaining high performance. This approach has been extensively adopted in federated learning (FL) [19, 21] paradigms, particularly in scenarios characterized by substantial computational and communication overheads between server and client devices. While federated learning based on prompt learning demonstrates substantial privacy and computational benefits, a key challenge persists: data on each node is often collected in a non-IID manner in real-world scenarios, leading to domain shift [16, 30] between nodes. For example, one device may take photos mostly indoors, while another mostly outdoors [24]. The presence of the domain shift issue renders the direct application of learning with hand-crafted prompts (such as \"a photo of a [class]\") in federated learning incapable of achieving satisfactory outcomes [33].", + "bbox": [ + 511, + 309, + 906, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address the domain shift problem in FL, federated domain generalization (FDG) [6, 16, 17, 29] further considers that the test data (target domain) is unavailable during decentralized training from distributed source domains. Some methods have optimized soft prompts [14, 25] across different clients by treating them as learnable vectors and refining them via backpropagation. For example, Promptfl [9] shares prompt vectors instead of model parameters to enhance communication efficiency and privacy in FL. pFedPG [34] introduced a domain bank mechanism to embed textual domain information into soft prompts. Recently, ADAPT [33] and FedTPG [26] used lightweight neural networks (e.g., adapter [11]) to optimize soft prompts conditioned on images through residual or concatenation.", + "bbox": [ + 511, + 674, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution", + "bbox": [ + 109, + 825, + 227, + 838 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding author is Jianzong Wang. This work was supported by the National Key Research and Development Program of China (Youth Scientist Project) under Grant No. 2024YFB4504300 and the Shenzhen-Hong Kong Joint Funding Project (Category A) under Grant No. SGDX20240115103359001.", + "bbox": [ + 89, + 838, + 482, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2366", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tion mechanisms [3], which improves generalization in a scalable manner. Despite recent advances, these methods either rely heavily on the diversity of client-specific data or use a simple adapter layer to estimate directly the relationship between soft prompts and images, leading to the learned prompts failing to fully capture the complex and diverse cross-domain feature [2].", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle the aforementioned challenges, we propose a novel and effective federated domain generalization method with domain-specific soft prompt generation (FedDSPG). Our approach reconceptualizes the prompt learning mechanism from a generative perspective, leveraging a generative model to produce domain-specific soft prompts (DSPs). Through an adversarial game manner between the generator and the discriminator, the generated DSPs exhibit greater diversity and generalization in both form and semantics. This enables our prompt generator model to acquire more generalized and robust features, facilitating adaptation to complex downstream tasks. The architectural details of our proposed approach are comprehensively depicted in Figure 1, comprising a training phase with two steps and an inference phase. Specifically, in the first step of the training phase, we avoid using hand-crafted prompts or soft prompts that lack sufficient attribute focus and specific content design. We elaborately design DSPs, composed of domain and class name vectors that map domain and content information, representing the optimal prompts for images from each client. In the second step, images and their corresponding DSPs are fed into a simple yet effective conditional generative model [4], which is trained to further optimize the DSPs by learning domain and content knowledge of images. Given the significant variance in tunable prompt embedding parameters across data-heterogeneous clients, the traditional FedAvg [19] aggregation method yields suboptimal results (see Table 5a for comparative experiments). To mitigate this, we propose using a momentum-based method to aggregate tunable parameters in the first training step. During the inference stage, the generative model's generator component will generate DSPs for target domain images, which subsequently facilitate predictions in downstream tasks.", + "bbox": [ + 91, + 214, + 483, + 773 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our key contributions are summarized below:", + "bbox": [ + 109, + 776, + 447, + 792 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We are the first to optimize prompt learning from a generative perspective in FL and propose a novel soft prompt generation paradigm for FDG tasks.", + "- Unlike the current CLIP-based methods that rely on hand-crafted or naive soft prompts, we devise domain-specific soft prompt structures that significantly enhance generalization to unseen domains." + ], + "bbox": [ + 91, + 794, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, we introduce a momentum-based aggregation mechanism to consistently refine performance.", + "bbox": [ + 529, + 90, + 903, + 133 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Comprehensive empirical evaluations on standard benchmark datasets corroborate that our proposed method achieves state-of-the-art performance. Additionally, comprehensive ablation research indicates the role played by each strategy in the increased efficacy.", + "bbox": [ + 514, + 136, + 906, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 243, + 660, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Prompt Learning for Vision Language Models", + "text_level": 1, + "bbox": [ + 511, + 268, + 901, + 285 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prompt learning, also known as prompt tuning, provides an efficient alternative to conventional fine-tuning approaches for pre-trained models. By focusing on tuning only learnable prompt tokens attached to the input of models like CLIP. CoOp [40] first introduced soft prompts into vision-language models, demonstrating that appropriate prompts can enhance performance in downstream tasks. CoCoOp [39] addresses the overfitting issue in CoOp through conditioned prompt learning, specifically residual prompts. DPL [36] achieved effective domain generalization by training a lightweight prompt generator (a linear network). FUZZLE [28] integrates fuzzy C-means clustering into prompt learning to optimize cross-domain alignment between source and target distributions. In these studies, soft prompts are typically generated by simple tunable models (e.g., adapter) that predominantly rely on the diversity of client-specific data, which yields unsatisfactory results for the unseen domain. The research most pertinent to ours is SPG [2], which employs conditional adversarial generative networks to generate prompts but lacks domain-specific design, resulting in suboptimal performance in federated learning scenarios. We not only meticulously designed DSPs but also tailored the federated aggregation algorithm to better execute FDG.", + "bbox": [ + 511, + 292, + 906, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Federated Domain Generalization", + "text_level": 1, + "bbox": [ + 511, + 696, + 810, + 710 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "FDG aims to address domain shifts between source and target domains in federated settings where data is distributed across clients. Traditional FDG methods generally fall into two types. The first is domain alignment approaches that focus on learning domain-invariant representations by reducing source domain discrepancies. For instance, in KD3A [7], domain differences are reduced by explicitly matching the mean embeddings of data from different domains. These methods' disadvantage lies in stringent assumptions about data distribution that are hard to satisfy in practice. The second type involves data augmentation", + "bbox": [ + 511, + 719, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2367", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e243d6ad935e909c52ba3bd94fef06d2a1fbb81fda1950caee0bbaf6fbc6aba4.jpg", + "image_caption": [ + "Figure 1. The overall training structure of FedDSPG consists of two main stages: domain-specific prompts tuning and domain prompt generation. The first stage produces DSPs with compositional soft prompting technique [22], which encode the content and style information with large-scale pretrained VLMs. In the second stage, a GAN-based training method is employed to learn the latent data distribution, thereby generating soft prompts that adhere to the intrinsic data patterns. For the inference stage, the generative model's generator component will generate DSPs for target domain images, producing enhanced predictions for downstream tasks. Additionally, the momentum-based aggregation is conducted to enhance model generalization for FDG." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 903, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "methods that aim to expand the training data distribution to enhance out-of-distribution generalization. Additional strategies include meta-learning [32], adversarial learning [38], and more. Recently, PromptFL [9], ADAPT [33], and FedTPG [26] introduced prompt learning to extract semantic information, aligning domain information on implicit textual features to improve the model's generalization capacity. To further advance this field, we conduct an in-depth exploration of this efficient paradigm. Departing from the previous prompt learning designs that rely on handcrafted prompts or simple neural network, we present a novel and effective FDG structure that leverages generative modeling techniques to learn domain-specific soft prompts, enhancing the generalization of prompts to unseen target domains.", + "bbox": [ + 88, + 492, + 485, + 736 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 763, + 236, + 782 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first introduce the preliminary knowledge of problem setup of FDG. Following that, the proposed method FedDSPG will be introduced in detail, which includes two training steps of domain-specific soft prompts tuning and prompts generative model tuning, as well as federated parameter aggregation and federated inference.", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 492, + 653, + 507 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the standard FDG problem, there is a source domain dataset $\\mathcal{D}_S = (\\mathbf{x}_i^s,\\mathbf{y}_i^s)_{i = 1}^n$ with labeled data, where $(\\mathbf{x}_i^s,\\mathbf{y}_i^s)$ is the data distribution for client $i^{\\prime}s$ input space and corresponding labels, with n being the number of source domains. The target domain dataset can be defined as $\\mathcal{D}_T = (\\mathbf{x}_i^t,\\mathbf{y}_i^t)_{i = 1}^m$ , where the number of target domains $m$ is typically configured as one per client for experiment evaluation. FDG seeks to develop models that can achieve robust performance on target domains while being trained exclusively on distributed source domain data. The primary challenge lies in addressing the domain shift between source and target domains, which is exacerbated by the non-IID nature of federated data across clients.", + "bbox": [ + 511, + 518, + 906, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Taking image classification as an example, let $X$ and $Y$ represent the input and label spaces, respectively. Previous domain generalization methods concentrated on training a domain-invariant model for the mapping $F: X \\to Y$ from images to labels. Prompt learning methods embed learnable prompt vectors $V$ into the input to leverage the generalization ability of VLMs, reformulating the mapping as $F: (X, V) \\to Y$ . Unlike existing prompt learning methods that focus primarily on fine-tuning the prompt vectors $V$ , we propose a novel prompt learning paradigm that utilizes a", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2368", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/29025138b121b72cdea1c30a6126a06f6fa9f01b7a82895709a109cfd55ebc60.jpg", + "image_caption": [ + "Figure 2. The structure design of the second training step is based on CcGAN. In the generator, input noise $z$ is concatenated with fake image embeddings to form the joint latent representation, which is used to generate DSPs. The discriminator distinguishes the authenticity of the generated DSPs from the real image." + ], + "image_footnote": [], + "bbox": [ + 102, + 90, + 478, + 265 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "generative model $G$ to generate soft prompts, reformulating the mapping as $F:(X,G(X))\\to Y$ . The objective is to develop a generative framework that simultaneously captures both domain-specific and domain-invariant representations, enabling the dynamic generation of generalized prompts for unseen target domains.", + "bbox": [ + 89, + 377, + 483, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Two Training Steps", + "text_level": 1, + "bbox": [ + 89, + 491, + 277, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "FedDSPG is implemented in two steps: domain-specific soft prompts tuning and prompts generative model tuning. Domain-specific prompts tuning aims to embed domain and class information into prompts, and the goal of prompts generative model tuning is to incorporate such information into the generative model. Notably, domain distribution is learned by the generative model, enabling the generation of instance-specific prompts for each image, guaranteeing prompt diversity and integrating domain-specific information. Below are the detailed designs of FedDSPG.", + "bbox": [ + 89, + 513, + 483, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Domain-specific Soft Prompts Tuning. We adopt CLIP [27] as our backbone, leveraging its contrastive learning framework trained on extensive image-text pairs to establish a joint embedding space through dedicated image $(f)$ and text $(g)$ encoders. By describing the input text as structured prompt $t_i$ : \"a photo of a [CLASS]\", the similarity of image representations in the same category $i$ can be maximized to align semantic representations:", + "bbox": [ + 89, + 680, + 483, + 816 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP (\\hat {y} = i | x) = \\frac {\\exp (S i m (g (t _ {i}) , f (x)) / \\tau)}{\\sum_ {k = 1} ^ {K} \\exp (S i m (g (t _ {k}) , f (x)) / \\tau)} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 825, + 482, + 863 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Where $Sim(\\cdot)$ is the cosine similarity and $\\tau$ denotes temperature parameter. Since the soft prompts", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of previous work (e.g., CoOp [40] and SPG [2] $t_i = [v]_1[v]_2 \\ldots [v]_{M_1}[\\mathrm{CLASS}]_i$ ) cannot deal well with the distribution shift between domains, we design domain-specific soft prompts to serve as the optimal prompt representations for images from each client. Specifically, each image from a client source domain is associated with unique DSPs (seen in Figure 1). The tunable vectors of DSPs contain two parts, a domain-invariant context and a domain-specific context, denoted as:", + "bbox": [ + 511, + 90, + 903, + 242 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nt _ {i} ^ {d} = [ v ] _ {1} [ v ] _ {2} \\dots [ v ] _ {M _ {1}} [ u ] _ {1} ^ {d} [ u ] _ {2} ^ {d} \\dots [ u ] _ {M _ {2}} ^ {d} [ \\mathrm {C L S} ] _ {i} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 247, + 903, + 267 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Where $[u]_{M_2}^d$ denotes domain-specific tokens, which have the same dimension as word embeddings. $M_{1}$ and $M_2$ are the number of context tokens and domain tokens incorporated in the prompt structure, respectively. These DSPs are optimized through cross-entropy loss minimization during training on domain-specific data:", + "bbox": [ + 511, + 270, + 906, + 377 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {c _ {j}} ^ {d _ {i} *} = \\arg \\min _ {\\mathbf {v}} \\mathbb {E} _ {\\mathbf {x} _ {j} ^ {d _ {i}}, y _ {j} ^ {d _ {i}}} \\left[ - \\log p \\left(y _ {j} ^ {d _ {i}} \\mid \\mathbf {x} _ {j} ^ {d _ {i}}, \\mathbf {t} _ {j} ^ {d _ {i}}\\right) \\right] \\quad (3)\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 383, + 906, + 411 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Where $(\\mathbf{x}_j^{d_i},y_j^{d_i})$ represents the image-label pairs from client $j^{\\prime}s$ training data in $i$ -th domain. DSPs integrate domain and class information in a more coherent and structured manner than prior approaches. These DSPs constitute optimized prompt representations for each client domain, encoding comprehensive domain characteristics, which encapsulate rich domain and class information for training a generative model in the second training step.", + "bbox": [ + 511, + 415, + 905, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prompts Generative Model Tuning. To learn the domain and class knowledge in DSPs, considering the communication cost requirements (e.g., methods based on diffusion models [10] require multi-step denoising, resulting in high communication costs. The corresponding experiments can be found in the supplementary materials.) and input data continuity of federated learning, we employ an efficient generative model—Continuous Conditional Generative Adversarial Network (CcGAN) [4] to validate the effectiveness of our approach. Similar to conditional GAN [20], we manipulate the generation process by constraining the additional information (i.e., DSPs generated in the first training step) of the CcGAN. The goal is to learn the generator's distribution $p_{g}$ over DSPs $\\mathbf{v}_{c_j}^{d_i}$ , effectively transferring generalization capabilities from prompts to the generator itself. The architecture consists of two core components: a generator $G$ that synthesizes outputs conditioned on auxiliary information and a discriminator $D$ that assesses the authenticity of real data and fake outputs. These components facilitate an adversarial process that enhances the model's generative performance.", + "bbox": [ + 511, + 553, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2369", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, the detailed training process of the prompts generative model is shown in Figure 2. The generator concatenates noise variables $\\mathbf{z}$ with fake-batch embeddings $f(\\mathbf{x})$ to form joint latent codes $[\\mathbf{z}, f(\\mathbf{x})]$ , where $\\mathbf{x}$ is randomly sampled from client data. Then the generator subsequently produces the corresponding DSPs by transforming the joint latent representation $[\\mathbf{z}, f(\\mathbf{x})]$ . In the discriminator, we input both the genuine image embeddings $f(\\mathbf{x})$ from real batches and the generated DSPs with their associated image embeddings $f(\\mathbf{x})$ to distinguish real from fake. Therefore, the two-player min-max game objective function $V(G, D)$ for CcGAN can be formally expressed as:", + "bbox": [ + 89, + 90, + 483, + 289 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {G} \\max _ {D} V (G, D) = \\mathbb {E} _ {\\mathbf {v} \\sim p _ {\\mathrm {v}} (\\mathbf {v})} [ \\log D (\\mathbf {v} \\mid f (\\mathbf {x})) ] \\\\ + \\mathbb {E} _ {\\mathbf {z} \\sim p _ {z} (\\mathbf {z})} [ \\log (1 - D (G (\\mathbf {z} \\mid f (\\mathbf {x}))) ]) \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 308, + 486, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The process of prompts generative model tuning is to learn coarse-to-fine DSPs through an adversarial manner. This design not only promotes diversity in prompt generation but also significantly improves generalization across tasks and domains, facilitating robust performance in both seen and unseen conditions.", + "bbox": [ + 89, + 364, + 483, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Federated Parameters Aggregation", + "text_level": 1, + "bbox": [ + 89, + 479, + 397, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We employ distinct aggregation strategies for the learnable parameters of prompt tuning and the generative model. Assuming a centralized parameter server for uniform aggregation—though our method also supports decentralized communication—clients upload their respective soft text embedding parameters and generative model parameters to the server during each communication round. The server averages the parameters from all n clients and redistributes them. While this parameter aggregation paradigm appears viable for our method, it introduces a minor issue for soft text embeddings. Specifically, given that FDG addresses domain shifts across clients, the domain components of DSPs vary significantly between clients. We observe that direct averaging causes abrupt changes in soft text embedding parameters, negatively impacting model performance. To mitigate this, we introduce momentum updates for soft prompt parameters, utilizing exponential moving averages to maintain parameter continuity and prevent destabilizing abrupt changes during the optimization process. The aggregation formula is as follows:", + "bbox": [ + 89, + 500, + 483, + 834 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n[ w ] ^ {k} = \\alpha [ w ] ^ {k - 1} + (1 - \\alpha) [ w ] ^ {k - 2} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 842, + 482, + 862 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Where $[w]^k$ , $[w]^{k-1}$ and $[w]^{k-2}$ denote the tunable soft prompt embedding parameters at the $k, k-1$ and", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$k - 2$ step, and $\\alpha \\in [0,1]$ serves as a hyperparameter governing the smoothing intensity of DSPs. Comprehensive ablation studies examining the impact of momentum updates are presented in table 5a.", + "bbox": [ + 511, + 90, + 906, + 152 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Inference with Generator", + "text_level": 1, + "bbox": [ + 511, + 159, + 746, + 174 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The generator trained via CcGAN is utilized to generate DSPs for inference on each target image. The predicted probability that a target domain image $x$ belongs to the $i$ -th class is formally expressed as:", + "bbox": [ + 511, + 181, + 906, + 242 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np (y = i \\mid \\mathbf {x}) = \\frac {\\exp (\\langle w _ {i} , f (\\mathbf {x}) \\rangle / \\tau)}{\\sum_ {j = 1} ^ {K} \\exp (\\langle w _ {j} , f (\\mathbf {x}) \\rangle / \\tau)}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 252, + 906, + 290 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns. t. \\quad w _ {j} = g ([ G (\\mathbf {z} \\mid f (\\mathbf {x})), c _ {j} ]) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 297, + 903, + 315 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Where $\\tau$ represents the temperature parameter that modulates the sharpness of the output distribution, $g$ and $f$ represent the text and the image encoder respectively, and $K$ denotes the total number of classes. The term $G(\\mathbf{z} \\mid f(\\mathbf{x}))$ signifies the generated DSPs by the generator. We use $c_{j}$ to represent the $j$ -th tokenized class label, with $[\\cdot, \\cdot]$ denoting the standard concatenation operation. In the generative paradigm, our approach achieves enhanced adaptability and dynamic prompts generation, which are essential for accommodating diverse input distributions and downstream task requirements.", + "bbox": [ + 511, + 319, + 906, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 512, + 718, + 531 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We conducted extensive experiments for image classification on domain generalization benchmark datasets: (1) Office-Home, a large-scale visual cross-domain dataset with 65 classes across four visually distinct domains, i.e., Art (Ar), Clipart (Cl), Product (Pr), and Real World (Re) [31]. (2) DomainNet, a largescale domain generalization benchmark dataset with 345 classes distributed among six domains, i.e., Clipart (Clp), Infograph (Inf), Painting (Pnt), Quickdraw (Qdr), Real (Rel), and Sketch (Skt) [23]. (3) PACS, featuring approximately 10,000 images from seven categories across four domains, including Photo (P), Sketch (S), Cartoon (C), and Art (A) [15]. For rigorous evaluation, we adopt the leave-one-domain-out protocol, designating each domain sequentially as the target while utilizing the remaining domains as source data.", + "bbox": [ + 511, + 537, + 906, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. In FDG, all participating nodes maintain identical architectural configurations while ensuring complete domain data isolation between nodes. The data processing and division are followed [8] on the image generalization task. For our FedDSPG method, we employ a frozen CLIP backbone (ViT-B/16 [5]) pre-trained on large-scale vision-language data. The text and visual prompt tokens", + "bbox": [ + 511, + 780, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2370", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/9e2905675cd4e0f88c941c30bd3c3939f0d21a7f1d088cf3075b3b8ca27de114.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetOffice-HomeDomainNetPACS
Target DomainRePrClArAvgClpInfPntQdrRelSktAvgPACSAvg
CoOp[40]82.982.859.372.174.361.622.252.612.167.748.644.194.191.595.684.991.5
CoCoOp[39]85.081.162.568.674.372.423.360.816.472.760.551.097.594.697.987.594.4
FedVPT-D[14]85.085.261.176.076.863.343.074.854.887.267.165.099.296.598.187.295.3
FedCLIP[18]87.787.563.678.079.265.146.575.956.190.267.967.099.896.397.985.694.9
PromptFL[9]89.688.271.679.882.372.850.384.260.492.167.771.399.997.199.090.696.7
ADAPT[33]90.390.568.282.682.977.563.170.541.685.772.168.499.998.099.191.797.2
FedTPG[26]89.089.771.184.783.674.052.185.162.194.068.672.799.897.899.290.896.9
FedDSPG90.591.073.885.885.374.763.585.563.094.272.675.699.998.199.392.297.4
", + "bbox": [ + 94, + 88, + 906, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Accuracies (%) for the image classification task on Office-Home, DomainNet, and PACS datasets using CLIP with ViT-B/16 image encoder as frozen backbone. The domains displayed here are the target domains, and the others are the source domains; the best result is indicated in bold.", + "bbox": [ + 89, + 300, + 906, + 344 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "share a unified dimensionality of 512. For the learning strategies, in the first training step, the DSPs are trained by using the Adam optimizer with the learning rate of $1 \\times 10^{-5}$ , batch size of 32, and the context lengths $M_{1}$ and $M_{2}$ of the prompt number are 4. In the second training step, the CcGAN is trained by using the AdamW optimizer with weight decay $2 \\times 10^{-5}$ and $\\beta_{1} = 0.9$ , $\\beta = 0.999$ , starting with a learning rate of $1 \\times 10^{-4}$ for the OfficeHome and DomainNet datasets. The 100 epochs are trained for both our method and the baseline models for accelerating convergence, then the FedAvg and momentum-based aggregation algorithms are executed after every epoch. 8 Tesla-V100 with 16GB memory are used to experiment.", + "bbox": [ + 88, + 362, + 480, + 574 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline models. We compared three distinct SOTA prompt learning approaches: (1) methods focusing on designing trainable models with small-scale parameters on the client side, such as CoOp [40], CoCoOp [39], FedVPT-D[14], FedCLIP [18], and FedTPG [26]; (2) methods that design prompt generators on the server side to reduce communication costs, such as PromptFL [9]; and (3) methods that emphasize both semantic and domain-specific prompts, such as ADAPT [33]. To show fair competition, the results of subsequent experiments were either the result of their papers or reproduced from source code provided by their authors.", + "bbox": [ + 89, + 575, + 482, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments Analysis", + "text_level": 1, + "bbox": [ + 89, + 789, + 310, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Federated Domain Generalization Performance", + "text_level": 1, + "bbox": [ + 89, + 816, + 482, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "FedDSPG achieved effective FDG in scenarios where target domain distributions diverge significantly from source domains. Quantitative results across three", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "benchmark datasets are presented in Table 1 following the \"leave-one-out\" protocol. In general, our method outperforms all baseline methods to varying degrees. We gleaned two primary observations from the results. Firstly, our approach outperforms all baselines, achieving SOTA performance, which demonstrates the effectiveness of our proposed approach in addressing federated domain generalization challenges. This establishes FedDSPG as a highly referencable benchmark for future research on FDG. Secondly, existing methods, including methods focusing on designing small-scale parameters models (FedCLIP, FedTPG) and methods generating prompts on the server side (PromptFL), exhibit notable performance gaps. We suppose that the possible reason stems from two primary factors: (1) inadequate incorporation of domain-specific information in their prompt design strategies, and (2) insufficient adaptability of generated prompts to unseen target domain images. These potential risks are mitigated in our approach, as it designs domain-specific soft prompts structure and trains a specific generative model, enhancing the generalization and diversity of prompts to unseen domains.", + "bbox": [ + 511, + 362, + 906, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Cross-Dataset Transfer", + "text_level": 1, + "bbox": [ + 511, + 724, + 728, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Having demonstrated FedDSPG's capability for domain generalization within a single dataset, we further conducted experiments on cross-dataset domain generalization. This represents a more challenging task, as the fundamental characteristics can exhibit significant variations across different datasets. Specifically, we adopt a rigorous experimental protocol where DomainNet serves as the source domain, while OfficeHome and PACS function as target domains. We compared FedDSPG with FedTPG (best-performing", + "bbox": [ + 511, + 748, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2371", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/76026bb88c01c1995efe960033b888a0e278636babc23995bd4c688d1a2f9bef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 102, + 95, + 361, + 257 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1fb3f94cf79013e94ac01dc58f3d70cbdda093f19619477efd5fd83feb6a8c3e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 95, + 625, + 258 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/beab1ee4bb344961cf6c66e3965c1da6048c8c3a004dadfb3824bb0f631fd6eb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 95, + 893, + 257 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1e84a11f912ef61a471d1bffb39df562fa88d147f46ed92623c73a1b30d9999c.jpg", + "table_caption": [ + "Figure 3. The accuracy of FedDSPG with different data distribution." + ], + "table_footnote": [], + "table_body": "
SourceTarget
DomainNetRePrClArAvgPACSAvg
FedTPG72.781.378.566.176.475.691.293.891.686.190.7
Ours75.686.082.570.480.479.894.595.393.588.693.0
△+3.9+4.7+4.0+4.3+4.0+4.2+3.3+1.5+1.9+2.5+2.3
", + "bbox": [ + 94, + 289, + 906, + 381 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. The results for cross-dataset transfer experiment. The model is trained on the DomainNet dataset and inference on the Office-Home and PACS dataset.", + "bbox": [ + 89, + 388, + 906, + 417 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "baseline), with the primary results detailed in Table 2. Overall, both methods experienced a decline in performance when confronted with the gap between datasets. However, FedDSPG consistently outperformed FedTPG on the target dataset by a substantial margin. This further underscores our method's remarkable capability to enhance model generalization across diverse domains.", + "bbox": [ + 88, + 433, + 480, + 555 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Data Distribution Analysis", + "text_level": 1, + "bbox": [ + 89, + 573, + 331, + 589 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Considering client data distributions are inherently non-IID in federated learning scenarios, we conducted a comprehensive empirical analysis examining how varying data distributions affect FedDSPG's performance. Specifically, we varied the overlap ratio $r$ of data domains across clients, setting it to $0\\%$ , $10\\%$ , $20\\%$ , and $50\\%$ . Specifically, an overlap ratio of $r\\%$ implies that $r\\%$ of the domains are shared among multiple clients, while the remaining $1 - r\\%$ of the domains are unique to individual clients. Figure 3 shows F1/accuracy across different domain distribution settings. The model exhibits greater stability on the Office-Home dataset compared to the other evaluation datasets, and a slight performance improvement is observed only when the overlap ratio reaches $50\\%$ . We attribute this enhanced stability to the higher number of domains, which likely endows the model with greater robustness against distribution shifts. This finding underscores the importance of domain diversity in enhancing the generalization and stability of", + "bbox": [ + 91, + 598, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "federated learning models.", + "bbox": [ + 513, + 434, + 710, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 465, + 666, + 482 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation on the different prompts structure. We classify prompts into 4 types (fixed, learnable, domain-specific, generative), and three representative SOTA models from baseline methods for comparison: PromptFL [9], and ADAPT [33], FedTPG [26]. PromptFL represents methods that design a generator on the server side to produce prompts. ADAPT represents methods emphasizing domain-specific information. FedTPG represents methods focusing on designing single trainable networks to optimize prompts.", + "bbox": [ + 509, + 489, + 906, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To validate the effectiveness of our approach and the improvements, several ablation models with different prompt structures were trained. These models include: Our model with hand-crafted domain prompts (Ours-HDP), which replaces DSPs with structured prompts (i.e., \"a photo of a [class]\"). Our model with conditional soft prompts (Ours-CSP), which replaces DSPs with conditional soft prompts (i.e., $t_i = [v]_1[v]_2 \\ldots [v]_{M_1}$ ). Our model without a generative model (Ours-WGM), which omits the second training step. The results are presented in Table 3. Ours-HDP performs the worst, proving learnable prompt designing are crucial for performance. Although both ADAPT and Ours-WGM have domain-specific prompt, the improvement in results is not significant. However, Ours-CSP outperforms Ours-WGM and other baseline methods, underscoring the crucial", + "bbox": [ + 509, + 643, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2372", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/eadc8b6cd3e598cd40dd58f544c9ca823a69ef96d17d6c07057378398664a4d8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPrompt TypeRePrClArAvg
FixedLearnableDomain-specificGenerative
PromptFL [9]✓✓89.688.271.679.882.3
ADAPT [33]✓✓90.390.568.282.682.9
FedTPG [26]✓89.089.771.184.783.6
Ours-HDP✓✓86.181.470.273.277.7
Ours-CSP✓✓90.990.873.184.184.6
Ours-WGM✓✓88.187.871.682.482.5
Ours✓✓✓90.591.073.885.885.3
", + "bbox": [ + 96, + 88, + 898, + 229 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c6e80a16a26676002c9a45e7d5f18c338f812bcaed5d565d1d407f4d7149e50f.jpg", + "table_caption": [ + "Table 3. Comparisons with different design of prompt on the Office-Home dataset for FDG performance." + ], + "table_footnote": [], + "table_body": "
α0.00.10.20.30.40.50.60.70.80.91.0
Accuracy (%)84.1584.8985.3085.2884.5684.3284.6784.0183.9583.5083.39
", + "bbox": [ + 107, + 257, + 888, + 304 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/553b9ca2b4b91b05215c9e8e874651b01dc72d4fbe67d0d832d73174065ba9bb.jpg", + "table_caption": [ + "Table 4. Accuracy of FedDSPG with different coefficient $\\alpha$ ." + ], + "table_footnote": [], + "table_body": "
Modeacc.
w/momentum85.3
w/o momentum84.5
", + "bbox": [ + 99, + 352, + 295, + 441 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/574542219672456c0b11731c119b1b6d5b0d2c2ef4835960a9c2bf63932d9e1f.jpg", + "table_caption": [ + "(a) Prompt update." + ], + "table_footnote": [], + "table_body": "
epochs/roundacc.
0.585.3
185.3
284.6
", + "bbox": [ + 307, + 352, + 478, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(b) Comm. frequency", + "bbox": [ + 326, + 444, + 454, + 457 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablation studies on different momentum update and communication frequency.", + "bbox": [ + 89, + 468, + 482, + 497 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "role of the generative model in enhancing generalizability. Ultimately, our method achieves superior performance compared to all baselines, demonstrating the effective integration of DSPs and the generative model.", + "bbox": [ + 89, + 506, + 482, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Momentum update and communication frequency. As elaborated in Section 3.3, we employ a momentum-based approach to mitigate the performance degradation caused by large gradient variations. We conduct comparative experiments on parameters with and without momentum updates. As presented in Table 5a, the results reveal a 0.8 percentage point decrease in accuracy without momentum. The experimental results for the momentum hyperparameter $\\alpha$ ranging from 0 to 1 are shown in Table 4, with $\\alpha = 0.2$ yielding the best performance. Additionally, we also test communication frequency, varying it to 0.5, 1, and 2 training epochs per communication round. Results (Table 5b) showed no performance gain from more frequent aggregation compared to the default 1 epoch per round, while lower frequencies led to accuracy declines. This suggests that FedDSPG can achieve good performance by maintaining the default update frequency.", + "bbox": [ + 89, + 566, + 482, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Number of clients and shots. We conducted ablation studies on the number of clients and shots, the results are displayed in Figure 4. Regarding the number", + "bbox": [ + 89, + 854, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/aa9ff88485566362e5cdd3b1f15debdec9984a37a1d4a7b574ff377f923e1d58.jpg", + "image_caption": [ + "Figure 4. Ablation studies on different number of clients and shots." + ], + "image_footnote": [], + "bbox": [ + 521, + 358, + 705, + 498 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f9614fc4e1b8a38f5b4fd168c68945884d4abbb8f4d8c561c02cb32b5f7c79ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 358, + 897, + 498 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "of clients, although an inevitable performance drop occurs for all methods as the number of clients increases, FedDSPG exhibits only a marginal decline (approximately $1\\%$ ). For the number of shots, FedDSPG outperforms baseline methods when the number of shots exceeds one.", + "bbox": [ + 511, + 555, + 905, + 646 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 660, + 640, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work introduces FedDSPG, a novel federated domain generalization approach that leverages domain-specific soft prompts (DSPs) to enhance model generalization in non-IID federated learning scenarios. By integrating domain and content knowledge into a generative model, FedDSPG generates diverse prompts for unseen target domains, improving performance in unknown domains. Through extensive evaluations, we demonstrate our method's superior performance against existing baselines, while also revealing the untapped potential of generative techniques in FDG scenarios. These findings provide both immediate practical value and long-term research directions for advancing FL in real-world applications.", + "bbox": [ + 509, + 685, + 906, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2373", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 192, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sikai Bai, Jiewei Zhang, Shuaicheng Li, Song Guo, Jingcai Guo, Jun Hou, Tao Han, and Xiaocheng Lu. Diprompt: Disentangled prompt tuning for multiple latent domain generalization in federated learning. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1", + "[2] Shuanghao Bai, Yuedi Zhang, Wanqi Zhou, Zhirong Luan, and Badong Chen. Soft prompt generation for domain generalization. In European Conference on Computer Vision, pages 434-450. Springer, 2025. 2, 4", + "[3] Tianyu Cui, Hongxia Li, Jingya Wang, and Ye Shi. Harmonizing generalization and personalization in federated prompt learning. International Conference on Machine Learning, 2024. 2", + "[4] Xin Ding, Yongwei Wang, Zuheng Xu, William J Welch, and Z Jane Wang. Ccgan: Continuous conditional generative adversarial networks for image generation. In International Conference on Learning Representations, 2021. 2, 4", + "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations, ICLR, 2021. 5", + "[6] Yuqi Fang, Pew-Thian Yap, Weili Lin, Hongtu Zhu, and Mingxia Liu. Source-free unsupervised domain adaptation: A survey. Neural Networks, 2024. 1", + "[7] Haozhe Feng, Zhaoyang You, Minghao Chen, Tianye Zhang, Minfeng Zhu, et al. Kd3a: Unsupervised multisource decentralized domain adaptation via knowledge distillation. In ICML, pages 3274-3283, 2021. 2", + "[8] Chunjiang Ge, Rui Huang, Mixue Xie, Zihang Lai, Shiji Song, Shuang Li, and Gao Huang. Domain adaptation via prompt learning. IEEE Transactions on Neural Networks and Learning Systems, 2023. 5", + "[9] Tao Guo, Song Guo, Junxiao Wang, Xueyang Tang, and Wenchao Xu. Promptfl: Let federated participants cooperatively learn prompts instead of models-federated learning in age of foundation model. IEEE Transactions on Mobile Computing, 2023. 1, 3, 6, 7, 8", + "[10] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 4", + "[11] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. 9th International Conference on Learning Representations, ICLR, 2021. 1", + "[12] Xuefeng Hu, Ke Zhang, Lu Xia, Albert Chen, Jiajia Luo, Yuyin Sun, Ken Wang, Nan Qiao, Xiao Zeng, Min Sun, et al. Reclip: Refine contrastive language image pretraining with source free domain adaptation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2994-3003, 2024. 1", + "[13] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen" + ], + "bbox": [ + 93, + 114, + 483, + 902 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 1", + "[14] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 1, 6", + "[15] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. In Proceedings of the IEEE International Conference on Computer Vision, pages 5542-5550, 2017. 5", + "[16] Xinhui Liu, Wei Xi, Wen Li, Dong Xu, Gairui Bai, and Jizhong Zhao. Co-mda: Federated multi-source domain adaptation on black-box models. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 1", + "[17] Xinhui Liu, Zhenghao Chen, Luping Zhou, Dong Xu, Wei Xi, Gairui Bai, Yihan Zhao, and Jizhong Zhao. Ufda: Universal federated domain adaptation with practical assumptions. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 14026-14034, 2024. 1", + "[18] Wang Lu, Xixu Hu, Jindong Wang, and Xing Xie. Fedclip: Fast generalization and personalization for clip in federated learning. In ICLR Workshop on Distributed and Private Machine Learning, 2023. 1, 6", + "[19] Brendan McMahan, Eider Moore, et al. Communication-efficient learning of deep networks from decentralized data. In Artificial Intelligence and Statistics, pages 1273-1282. PMLR, 2017. 1, 2", + "[20] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. Advances in Neural Information Processing Systems, 2014. 4", + "[21] Mehryar Mohri, Gary Sivek, and Ananda Theertha Suresh. Agnostic federated learning. In International Conference on Machine Learning, pages 4615-4625. PMLR, 2019. 1", + "[22] Nihal V Nayak, Peilin Yu, and Stephen H Bach. Learning to compose soft prompts for compositional zero-shot learning. International Conference on Learning Representations, 2022. 3", + "[23] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1406-1415, 2019. 5", + "[24] Xingchao Peng, Zijun Huang, Yizhe Zhu, and Kate Saenko. Federated adversarial domain adaptation. International Conference on Learning Representations, ICLR, 2019. 1", + "[25] Chen Qiu, Xingyu Li, Chaithanya Kumar Mummadi, Madan Ravi Ganesh, Zhenzhen Li, Lu Peng, and Wan-Yi Lin. Text-driven prompt generation for vision-language models in federated learning. International Conference on Learning Representations, 2024. 1", + "[26] Chen Qiu, Xingyu Li, Chaithanya Kumar Mummadi, Madan Ravi Ganesh, Zhenzhen Li, Lu Peng, and Wan-Yi Lin. Federated text-driven prompt generation for" + ], + "bbox": [ + 516, + 92, + 903, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2374", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "vision-language models. In The Twelfth International Conference on Learning Representations, 2024. 1, 3, 6, 7, 8", + "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 4", + "[28] Kuo Shi, Jie Lu, Zhen Fang, and Guangquan Zhang. Unsupervised domain adaptation enhanced by fuzzy prompt learning. IEEE Transactions on Fuzzy Systems, 32:4038-4048, 2024. 2", + "[29] Lei Song, Chunguang Ma, Guoyin Zhang, and Yun Zhang. Privacy-preserving unsupervised domain adaptation in federated setting. IEEE Access, 8:143233-143240, 2020. 1", + "[30] Yue Tan, Chen Chen, Weiming Zhuang, Xin Dong, Lingjuan Lyu, and Guodong Long. Is heterogeneity notorious? taming heterogeneity to handle test-time shift in federated learning. Advances in Neural Information Processing Systems, 36, 2024. 1", + "[31] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5018-5027, 2017. 5", + "[32] Anna Vettoruzzo, Mohamed-Rafik Bouguelia, Joaquin Vanschoren, Thorsteinn Rognvaldsson, and KC Santosh. Advances and challenges in meta-learning: A technical review. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 4763-4779, 2024. 3", + "[33] Guoyizhe Wei, Feng Wang, Anshul Shah, and Rama Chellappa. Learning to prompt your domain for vision-language models. In arXiv, 2024. 1, 3, 6, 7, 8", + "[34] Fu-En Yang, Chien-Yi Wang, and Yu-Chiang Frank Wang. Efficient model personalization in federated learning via client-specific prompt generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19159–19168, 2023. 1", + "[35] Elad Ben Zaken, Shauli Ravfogel, and Yoav Goldberg. Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. Annual Meeting of the Association for Computational Linguistics, 2021. 1", + "[36] Xin Zhang, Shixiang Shane Gu, Yutaka Matsuo, and Yusuke Iwasawa. Domain prompt learning for efficiently adapting clip to unseen domains. Transactions of the Japanese Society for Artificial Intelligence, 38(6):B-MC2_1, 2023. 2", + "[37] Haodong Zhao, Wei Du, Fangqi Li, Peixuan Li, and Gongshen Liu. Fedprompt: Communication-efficient and privacy-preserving prompt tuning in federated learning. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 1", + "[38] Ke Zhao, Junchen Hu, Haidong Shao, and Jiabei Hu. Federated multi-source domain adversarial adaptation" + ], + "bbox": [ + 91, + 90, + 482, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "framework for machinery fault diagnosis with data privacy. Reliability Engineering & System Safety, 236:109246, 2023. 3", + "[39] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 2, 6", + "[40] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9): 2337-2348, 2022. 2, 4, 6" + ], + "bbox": [ + 516, + 90, + 903, + 260 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2375", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Federated Domain Generalization with Domain-specific Soft Prompts Generation/b70289dd-c48a-4e53-8732-8d816d04b0c4_model.json b/2025/Federated Domain Generalization with Domain-specific Soft Prompts Generation/b70289dd-c48a-4e53-8732-8d816d04b0c4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bcfec044618ada4437d4f920ea7932f3f7e7db54 --- /dev/null +++ b/2025/Federated Domain Generalization with Domain-specific Soft Prompts Generation/b70289dd-c48a-4e53-8732-8d816d04b0c4_model.json @@ -0,0 +1,1826 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.124, + 0.13, + 0.877, + 0.175 + ], + "angle": 0, + "content": "Federated Domain Generalization with Domain-specific Soft Prompts Generation" + }, + { + "type": "text", + "bbox": [ + 0.197, + 0.203, + 0.808, + 0.24 + ], + "angle": 0, + "content": "Jianhan Wu * Xiaoyang Qu * Zhangcheng Huang Jianzong Wang † Ping An Technology (Shenzhen) Co., Ltd." + }, + { + "type": "text", + "bbox": [ + 0.297, + 0.243, + 0.7, + 0.257 + ], + "angle": 0, + "content": "wujianhan@mail.ustc.edu.cn, quxiaoy@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.293, + 0.26, + 0.702, + 0.275 + ], + "angle": 0, + "content": "william.huang@connect.polyu.hk, jzwang@188.com" + }, + { + "type": "title", + "bbox": [ + 0.247, + 0.309, + 0.33, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.34, + 0.485, + 0.718 + ], + "angle": 0, + "content": "Prompt learning has become an efficient paradigm for adapting CLIP to downstream tasks. Compared with traditional fine-tuning, prompt learning optimizes a few parameters yet yields highly competitive results, especially appealing in federated learning for computational efficiency. engendering domain shift among clients and posing a formidable challenge for downstream-task adaptation. Existing federated domain generalization (FDG) methods based on prompt learning typically learn soft prompts from training samples, replacing manually designed prompts to enhance the generalization ability of federated models. However, these learned prompts exhibit limited diversity and tend to ignore information from unknown domains. We propose a novel and effective method from a generative perspective for handling FDG tasks, namely federated domain generalization with domain-specific soft prompts generation (FedDSPG). Specifically, during training, we introduce domain-specific soft prompts (DSPs) for each domain and integrate content and domain knowledge into the generative model among clients. In the inference phase, the generator is utilized to obtain DSPs for unseen target domains, thus guiding downstream tasks in unknown domains. Comprehensive evaluations across several public datasets confirm that our method outperforms existing strong baselines in FDG, achieving state-of-the-art results." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.747, + 0.23, + 0.763 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.772, + 0.483, + 0.82 + ], + "angle": 0, + "content": "Pre-trained vision-language models (VLMs) [13], such as Contrastive Language Image Pretraining (CLIP) [12, 18, 27], have emerged as a powerful technique to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.31, + 0.908, + 0.674 + ], + "angle": 0, + "content": "adapt to downstream tasks. By introducing learnable prompts [1, 9, 37] into the text or vision encoders of CLIP, this approach allows the model to dynamically adjust its representations with minimal fine-tuning. Unlike traditional fine-tuning methods [35] that require extensive parameter updates, prompt learning optimizes only a small fraction of parameters, significantly reducing computational costs while maintaining high performance. This approach has been extensively adopted in federated learning (FL) [19, 21] paradigms, particularly in scenarios characterized by substantial computational and communication overheads between server and client devices. While federated learning based on prompt learning demonstrates substantial privacy and computational benefits, a key challenge persists: data on each node is often collected in a non-IID manner in real-world scenarios, leading to domain shift [16, 30] between nodes. For example, one device may take photos mostly indoors, while another mostly outdoors [24]. The presence of the domain shift issue renders the direct application of learning with hand-crafted prompts (such as \"a photo of a [class]\") in federated learning incapable of achieving satisfactory outcomes [33]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To address the domain shift problem in FL, federated domain generalization (FDG) [6, 16, 17, 29] further considers that the test data (target domain) is unavailable during decentralized training from distributed source domains. Some methods have optimized soft prompts [14, 25] across different clients by treating them as learnable vectors and refining them via backpropagation. For example, Promptfl [9] shares prompt vectors instead of model parameters to enhance communication efficiency and privacy in FL. pFedPG [34] introduced a domain bank mechanism to embed textual domain information into soft prompts. Recently, ADAPT [33] and FedTPG [26] used lightweight neural networks (e.g., adapter [11]) to optimize soft prompts conditioned on images through residual or concatenation." + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.827, + 0.228, + 0.839 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.839, + 0.483, + 0.9 + ], + "angle": 0, + "content": "†Corresponding author is Jianzong Wang. This work was supported by the National Key Research and Development Program of China (Youth Scientist Project) under Grant No. 2024YFB4504300 and the Shenzhen-Hong Kong Joint Funding Project (Category A) under Grant No. SGDX20240115103359001." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "2366" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "tion mechanisms [3], which improves generalization in a scalable manner. Despite recent advances, these methods either rely heavily on the diversity of client-specific data or use a simple adapter layer to estimate directly the relationship between soft prompts and images, leading to the learned prompts failing to fully capture the complex and diverse cross-domain feature [2]." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.215, + 0.485, + 0.775 + ], + "angle": 0, + "content": "To tackle the aforementioned challenges, we propose a novel and effective federated domain generalization method with domain-specific soft prompt generation (FedDSPG). Our approach reconceptualizes the prompt learning mechanism from a generative perspective, leveraging a generative model to produce domain-specific soft prompts (DSPs). Through an adversarial game manner between the generator and the discriminator, the generated DSPs exhibit greater diversity and generalization in both form and semantics. This enables our prompt generator model to acquire more generalized and robust features, facilitating adaptation to complex downstream tasks. The architectural details of our proposed approach are comprehensively depicted in Figure 1, comprising a training phase with two steps and an inference phase. Specifically, in the first step of the training phase, we avoid using hand-crafted prompts or soft prompts that lack sufficient attribute focus and specific content design. We elaborately design DSPs, composed of domain and class name vectors that map domain and content information, representing the optimal prompts for images from each client. In the second step, images and their corresponding DSPs are fed into a simple yet effective conditional generative model [4], which is trained to further optimize the DSPs by learning domain and content knowledge of images. Given the significant variance in tunable prompt embedding parameters across data-heterogeneous clients, the traditional FedAvg [19] aggregation method yields suboptimal results (see Table 5a for comparative experiments). To mitigate this, we propose using a momentum-based method to aggregate tunable parameters in the first training step. During the inference stage, the generative model's generator component will generate DSPs for target domain images, which subsequently facilitate predictions in downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.777, + 0.449, + 0.793 + ], + "angle": 0, + "content": "Our key contributions are summarized below:" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.795, + 0.483, + 0.841 + ], + "angle": 0, + "content": "- We are the first to optimize prompt learning from a generative perspective in FL and propose a novel soft prompt generation paradigm for FDG tasks." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "- Unlike the current CLIP-based methods that rely on hand-crafted or naive soft prompts, we devise domain-specific soft prompt structures that significantly enhance generalization to unseen domains." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.795, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.092, + 0.905, + 0.135 + ], + "angle": 0, + "content": "Furthermore, we introduce a momentum-based aggregation mechanism to consistently refine performance." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.137, + 0.907, + 0.228 + ], + "angle": 0, + "content": "- Comprehensive empirical evaluations on standard benchmark datasets corroborate that our proposed method achieves state-of-the-art performance. Additionally, comprehensive ablation research indicates the role played by each strategy in the increased efficacy." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.244, + 0.661, + 0.26 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.269, + 0.903, + 0.286 + ], + "angle": 0, + "content": "2.1. Prompt Learning for Vision Language Models" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.293, + 0.907, + 0.685 + ], + "angle": 0, + "content": "Prompt learning, also known as prompt tuning, provides an efficient alternative to conventional fine-tuning approaches for pre-trained models. By focusing on tuning only learnable prompt tokens attached to the input of models like CLIP. CoOp [40] first introduced soft prompts into vision-language models, demonstrating that appropriate prompts can enhance performance in downstream tasks. CoCoOp [39] addresses the overfitting issue in CoOp through conditioned prompt learning, specifically residual prompts. DPL [36] achieved effective domain generalization by training a lightweight prompt generator (a linear network). FUZZLE [28] integrates fuzzy C-means clustering into prompt learning to optimize cross-domain alignment between source and target distributions. In these studies, soft prompts are typically generated by simple tunable models (e.g., adapter) that predominantly rely on the diversity of client-specific data, which yields unsatisfactory results for the unseen domain. The research most pertinent to ours is SPG [2], which employs conditional adversarial generative networks to generate prompts but lacks domain-specific design, resulting in suboptimal performance in federated learning scenarios. We not only meticulously designed DSPs but also tailored the federated aggregation algorithm to better execute FDG." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.697, + 0.812, + 0.712 + ], + "angle": 0, + "content": "2.2. Federated Domain Generalization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.903 + ], + "angle": 0, + "content": "FDG aims to address domain shifts between source and target domains in federated settings where data is distributed across clients. Traditional FDG methods generally fall into two types. The first is domain alignment approaches that focus on learning domain-invariant representations by reducing source domain discrepancies. For instance, in KD3A [7], domain differences are reduced by explicitly matching the mean embeddings of data from different domains. These methods' disadvantage lies in stringent assumptions about data distribution that are hard to satisfy in practice. The second type involves data augmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2367" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.905, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.39, + 0.908, + 0.477 + ], + "angle": 0, + "content": "Figure 1. The overall training structure of FedDSPG consists of two main stages: domain-specific prompts tuning and domain prompt generation. The first stage produces DSPs with compositional soft prompting technique [22], which encode the content and style information with large-scale pretrained VLMs. In the second stage, a GAN-based training method is employed to learn the latent data distribution, thereby generating soft prompts that adhere to the intrinsic data patterns. For the inference stage, the generative model's generator component will generate DSPs for target domain images, producing enhanced predictions for downstream tasks. Additionally, the momentum-based aggregation is conducted to enhance model generalization for FDG." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.493, + 0.486, + 0.737 + ], + "angle": 0, + "content": "methods that aim to expand the training data distribution to enhance out-of-distribution generalization. Additional strategies include meta-learning [32], adversarial learning [38], and more. Recently, PromptFL [9], ADAPT [33], and FedTPG [26] introduced prompt learning to extract semantic information, aligning domain information on implicit textual features to improve the model's generalization capacity. To further advance this field, we conduct an in-depth exploration of this efficient paradigm. Departing from the previous prompt learning designs that rely on handcrafted prompts or simple neural network, we present a novel and effective FDG structure that leverages generative modeling techniques to learn domain-specific soft prompts, enhancing the generalization of prompts to unseen target domains." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.765, + 0.238, + 0.784 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.485, + 0.903 + ], + "angle": 0, + "content": "In this section, we first introduce the preliminary knowledge of problem setup of FDG. Following that, the proposed method FedDSPG will be introduced in detail, which includes two training steps of domain-specific soft prompts tuning and prompts generative model tuning, as well as federated parameter aggregation and federated inference." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.493, + 0.655, + 0.508 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.519, + 0.907, + 0.732 + ], + "angle": 0, + "content": "In the standard FDG problem, there is a source domain dataset \\(\\mathcal{D}_S = (\\mathbf{x}_i^s,\\mathbf{y}_i^s)_{i = 1}^n\\) with labeled data, where \\((\\mathbf{x}_i^s,\\mathbf{y}_i^s)\\) is the data distribution for client \\(i^{\\prime}s\\) input space and corresponding labels, with n being the number of source domains. The target domain dataset can be defined as \\(\\mathcal{D}_T = (\\mathbf{x}_i^t,\\mathbf{y}_i^t)_{i = 1}^m\\), where the number of target domains \\(m\\) is typically configured as one per client for experiment evaluation. FDG seeks to develop models that can achieve robust performance on target domains while being trained exclusively on distributed source domain data. The primary challenge lies in addressing the domain shift between source and target domains, which is exacerbated by the non-IID nature of federated data across clients." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Taking image classification as an example, let \\( X \\) and \\( Y \\) represent the input and label spaces, respectively. Previous domain generalization methods concentrated on training a domain-invariant model for the mapping \\( F: X \\to Y \\) from images to labels. Prompt learning methods embed learnable prompt vectors \\( V \\) into the input to leverage the generalization ability of VLMs, reformulating the mapping as \\( F: (X, V) \\to Y \\). Unlike existing prompt learning methods that focus primarily on fine-tuning the prompt vectors \\( V \\), we propose a novel prompt learning paradigm that utilizes a" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2368" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.092, + 0.48, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.277, + 0.485, + 0.364 + ], + "angle": 0, + "content": "Figure 2. The structure design of the second training step is based on CcGAN. In the generator, input noise \\( z \\) is concatenated with fake image embeddings to form the joint latent representation, which is used to generate DSPs. The discriminator distinguishes the authenticity of the generated DSPs from the real image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.378, + 0.484, + 0.485 + ], + "angle": 0, + "content": "generative model \\(G\\) to generate soft prompts, reformulating the mapping as \\(F:(X,G(X))\\to Y\\). The objective is to develop a generative framework that simultaneously captures both domain-specific and domain-invariant representations, enabling the dynamic generation of generalized prompts for unseen target domains." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.492, + 0.279, + 0.51 + ], + "angle": 0, + "content": "3.2. Two Training Steps" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.515, + 0.484, + 0.681 + ], + "angle": 0, + "content": "FedDSPG is implemented in two steps: domain-specific soft prompts tuning and prompts generative model tuning. Domain-specific prompts tuning aims to embed domain and class information into prompts, and the goal of prompts generative model tuning is to incorporate such information into the generative model. Notably, domain distribution is learned by the generative model, enabling the generation of instance-specific prompts for each image, guaranteeing prompt diversity and integrating domain-specific information. Below are the detailed designs of FedDSPG." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.681, + 0.484, + 0.817 + ], + "angle": 0, + "content": "Domain-specific Soft Prompts Tuning. We adopt CLIP [27] as our backbone, leveraging its contrastive learning framework trained on extensive image-text pairs to establish a joint embedding space through dedicated image \\((f)\\) and text \\((g)\\) encoders. By describing the input text as structured prompt \\(t_i\\): \"a photo of a [CLASS]\", the similarity of image representations in the same category \\(i\\) can be maximized to align semantic representations:" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.826, + 0.483, + 0.864 + ], + "angle": 0, + "content": "\\[\nP (\\hat {y} = i | x) = \\frac {\\exp (S i m (g (t _ {i}) , f (x)) / \\tau)}{\\sum_ {k = 1} ^ {K} \\exp (S i m (g (t _ {k}) , f (x)) / \\tau)} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.87, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Where \\( Sim(\\cdot) \\) is the cosine similarity and \\( \\tau \\) denotes temperature parameter. Since the soft prompts" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.091, + 0.905, + 0.243 + ], + "angle": 0, + "content": "of previous work (e.g., CoOp [40] and SPG [2] \\(t_i = [v]_1[v]_2 \\ldots [v]_{M_1}[\\mathrm{CLASS}]_i\\)) cannot deal well with the distribution shift between domains, we design domain-specific soft prompts to serve as the optimal prompt representations for images from each client. Specifically, each image from a client source domain is associated with unique DSPs (seen in Figure 1). The tunable vectors of DSPs contain two parts, a domain-invariant context and a domain-specific context, denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.248, + 0.905, + 0.268 + ], + "angle": 0, + "content": "\\[\nt _ {i} ^ {d} = [ v ] _ {1} [ v ] _ {2} \\dots [ v ] _ {M _ {1}} [ u ] _ {1} ^ {d} [ u ] _ {2} ^ {d} \\dots [ u ] _ {M _ {2}} ^ {d} [ \\mathrm {C L S} ] _ {i} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.271, + 0.907, + 0.378 + ], + "angle": 0, + "content": "Where \\([u]_{M_2}^d\\) denotes domain-specific tokens, which have the same dimension as word embeddings. \\(M_{1}\\) and \\(M_2\\) are the number of context tokens and domain tokens incorporated in the prompt structure, respectively. These DSPs are optimized through cross-entropy loss minimization during training on domain-specific data:" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.384, + 0.907, + 0.412 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {c _ {j}} ^ {d _ {i} *} = \\arg \\min _ {\\mathbf {v}} \\mathbb {E} _ {\\mathbf {x} _ {j} ^ {d _ {i}}, y _ {j} ^ {d _ {i}}} \\left[ - \\log p \\left(y _ {j} ^ {d _ {i}} \\mid \\mathbf {x} _ {j} ^ {d _ {i}}, \\mathbf {t} _ {j} ^ {d _ {i}}\\right) \\right] \\quad (3)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.416, + 0.906, + 0.553 + ], + "angle": 0, + "content": "Where \\((\\mathbf{x}_j^{d_i},y_j^{d_i})\\) represents the image-label pairs from client \\(j^{\\prime}s\\) training data in \\(i\\)-th domain. DSPs integrate domain and class information in a more coherent and structured manner than prior approaches. These DSPs constitute optimized prompt representations for each client domain, encoding comprehensive domain characteristics, which encapsulate rich domain and class information for training a generative model in the second training step." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Prompts Generative Model Tuning. To learn the domain and class knowledge in DSPs, considering the communication cost requirements (e.g., methods based on diffusion models [10] require multi-step denoising, resulting in high communication costs. The corresponding experiments can be found in the supplementary materials.) and input data continuity of federated learning, we employ an efficient generative model—Continuous Conditional Generative Adversarial Network (CcGAN) [4] to validate the effectiveness of our approach. Similar to conditional GAN [20], we manipulate the generation process by constraining the additional information (i.e., DSPs generated in the first training step) of the CcGAN. The goal is to learn the generator's distribution \\( p_{g} \\) over DSPs \\( \\mathbf{v}_{c_j}^{d_i} \\), effectively transferring generalization capabilities from prompts to the generator itself. The architecture consists of two core components: a generator \\( G \\) that synthesizes outputs conditioned on auxiliary information and a discriminator \\( D \\) that assesses the authenticity of real data and fake outputs. These components facilitate an adversarial process that enhances the model's generative performance." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2369" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.29 + ], + "angle": 0, + "content": "Specifically, the detailed training process of the prompts generative model is shown in Figure 2. The generator concatenates noise variables \\(\\mathbf{z}\\) with fake-batch embeddings \\(f(\\mathbf{x})\\) to form joint latent codes \\([\\mathbf{z}, f(\\mathbf{x})]\\), where \\(\\mathbf{x}\\) is randomly sampled from client data. Then the generator subsequently produces the corresponding DSPs by transforming the joint latent representation \\([\\mathbf{z}, f(\\mathbf{x})]\\). In the discriminator, we input both the genuine image embeddings \\(f(\\mathbf{x})\\) from real batches and the generated DSPs with their associated image embeddings \\(f(\\mathbf{x})\\) to distinguish real from fake. Therefore, the two-player min-max game objective function \\(V(G, D)\\) for CcGAN can be formally expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.309, + 0.487, + 0.365 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {G} \\max _ {D} V (G, D) = \\mathbb {E} _ {\\mathbf {v} \\sim p _ {\\mathrm {v}} (\\mathbf {v})} [ \\log D (\\mathbf {v} \\mid f (\\mathbf {x})) ] \\\\ + \\mathbb {E} _ {\\mathbf {z} \\sim p _ {z} (\\mathbf {z})} [ \\log (1 - D (G (\\mathbf {z} \\mid f (\\mathbf {x}))) ]) \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.366, + 0.484, + 0.472 + ], + "angle": 0, + "content": "The process of prompts generative model tuning is to learn coarse-to-fine DSPs through an adversarial manner. This design not only promotes diversity in prompt generation but also significantly improves generalization across tasks and domains, facilitating robust performance in both seen and unseen conditions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.48, + 0.398, + 0.496 + ], + "angle": 0, + "content": "3.3. Federated Parameters Aggregation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.501, + 0.484, + 0.835 + ], + "angle": 0, + "content": "We employ distinct aggregation strategies for the learnable parameters of prompt tuning and the generative model. Assuming a centralized parameter server for uniform aggregation—though our method also supports decentralized communication—clients upload their respective soft text embedding parameters and generative model parameters to the server during each communication round. The server averages the parameters from all n clients and redistributes them. While this parameter aggregation paradigm appears viable for our method, it introduces a minor issue for soft text embeddings. Specifically, given that FDG addresses domain shifts across clients, the domain components of DSPs vary significantly between clients. We observe that direct averaging causes abrupt changes in soft text embedding parameters, negatively impacting model performance. To mitigate this, we introduce momentum updates for soft prompt parameters, utilizing exponential moving averages to maintain parameter continuity and prevent destabilizing abrupt changes during the optimization process. The aggregation formula is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.843, + 0.483, + 0.863 + ], + "angle": 0, + "content": "\\[\n[ w ] ^ {k} = \\alpha [ w ] ^ {k - 1} + (1 - \\alpha) [ w ] ^ {k - 2} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.87, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Where \\([w]^k\\), \\([w]^{k-1}\\) and \\([w]^{k-2}\\) denote the tunable soft prompt embedding parameters at the \\(k, k-1\\) and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.153 + ], + "angle": 0, + "content": "\\(k - 2\\) step, and \\(\\alpha \\in [0,1]\\) serves as a hyperparameter governing the smoothing intensity of DSPs. Comprehensive ablation studies examining the impact of momentum updates are presented in table 5a." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.16, + 0.747, + 0.175 + ], + "angle": 0, + "content": "3.4. Inference with Generator" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.182, + 0.907, + 0.243 + ], + "angle": 0, + "content": "The generator trained via CcGAN is utilized to generate DSPs for inference on each target image. The predicted probability that a target domain image \\( x \\) belongs to the \\( i \\)-th class is formally expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.253, + 0.907, + 0.291 + ], + "angle": 0, + "content": "\\[\np (y = i \\mid \\mathbf {x}) = \\frac {\\exp (\\langle w _ {i} , f (\\mathbf {x}) \\rangle / \\tau)}{\\sum_ {j = 1} ^ {K} \\exp (\\langle w _ {j} , f (\\mathbf {x}) \\rangle / \\tau)}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.299, + 0.905, + 0.316 + ], + "angle": 0, + "content": "\\[\ns. t. \\quad w _ {j} = g ([ G (\\mathbf {z} \\mid f (\\mathbf {x})), c _ {j} ]) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.32, + 0.907, + 0.502 + ], + "angle": 0, + "content": "Where \\(\\tau\\) represents the temperature parameter that modulates the sharpness of the output distribution, \\(g\\) and \\(f\\) represent the text and the image encoder respectively, and \\(K\\) denotes the total number of classes. The term \\(G(\\mathbf{z} \\mid f(\\mathbf{x}))\\) signifies the generated DSPs by the generator. We use \\(c_{j}\\) to represent the \\(j\\)-th tokenized class label, with \\([\\cdot, \\cdot]\\) denoting the standard concatenation operation. In the generative paradigm, our approach achieves enhanced adaptability and dynamic prompts generation, which are essential for accommodating diverse input distributions and downstream task requirements." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.513, + 0.719, + 0.532 + ], + "angle": 0, + "content": "4. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.907, + 0.78 + ], + "angle": 0, + "content": "Datasets. We conducted extensive experiments for image classification on domain generalization benchmark datasets: (1) Office-Home, a large-scale visual cross-domain dataset with 65 classes across four visually distinct domains, i.e., Art (Ar), Clipart (Cl), Product (Pr), and Real World (Re) [31]. (2) DomainNet, a largescale domain generalization benchmark dataset with 345 classes distributed among six domains, i.e., Clipart (Clp), Infograph (Inf), Painting (Pnt), Quickdraw (Qdr), Real (Rel), and Sketch (Skt) [23]. (3) PACS, featuring approximately 10,000 images from seven categories across four domains, including Photo (P), Sketch (S), Cartoon (C), and Art (A) [15]. For rigorous evaluation, we adopt the leave-one-domain-out protocol, designating each domain sequentially as the target while utilizing the remaining domains as source data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Implementation details. In FDG, all participating nodes maintain identical architectural configurations while ensuring complete domain data isolation between nodes. The data processing and division are followed [8] on the image generalization task. For our FedDSPG method, we employ a frozen CLIP backbone (ViT-B/16 [5]) pre-trained on large-scale vision-language data. The text and visual prompt tokens" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2370" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.907, + 0.293 + ], + "angle": 0, + "content": "
DatasetOffice-HomeDomainNetPACS
Target DomainRePrClArAvgClpInfPntQdrRelSktAvgPACSAvg
CoOp[40]82.982.859.372.174.361.622.252.612.167.748.644.194.191.595.684.991.5
CoCoOp[39]85.081.162.568.674.372.423.360.816.472.760.551.097.594.697.987.594.4
FedVPT-D[14]85.085.261.176.076.863.343.074.854.887.267.165.099.296.598.187.295.3
FedCLIP[18]87.787.563.678.079.265.146.575.956.190.267.967.099.896.397.985.694.9
PromptFL[9]89.688.271.679.882.372.850.384.260.492.167.771.399.997.199.090.696.7
ADAPT[33]90.390.568.282.682.977.563.170.541.685.772.168.499.998.099.191.797.2
FedTPG[26]89.089.771.184.783.674.052.185.162.194.068.672.799.897.899.290.896.9
FedDSPG90.591.073.885.885.374.763.585.563.094.272.675.699.998.199.392.297.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.301, + 0.907, + 0.345 + ], + "angle": 0, + "content": "Table 1. Accuracies (%) for the image classification task on Office-Home, DomainNet, and PACS datasets using CLIP with ViT-B/16 image encoder as frozen backbone. The domains displayed here are the target domains, and the others are the source domains; the best result is indicated in bold." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.363, + 0.482, + 0.575 + ], + "angle": 0, + "content": "share a unified dimensionality of 512. For the learning strategies, in the first training step, the DSPs are trained by using the Adam optimizer with the learning rate of \\(1 \\times 10^{-5}\\), batch size of 32, and the context lengths \\(M_{1}\\) and \\(M_{2}\\) of the prompt number are 4. In the second training step, the CcGAN is trained by using the AdamW optimizer with weight decay \\(2 \\times 10^{-5}\\) and \\(\\beta_{1} = 0.9\\), \\(\\beta = 0.999\\), starting with a learning rate of \\(1 \\times 10^{-4}\\) for the OfficeHome and DomainNet datasets. The 100 epochs are trained for both our method and the baseline models for accelerating convergence, then the FedAvg and momentum-based aggregation algorithms are executed after every epoch. 8 Tesla-V100 with 16GB memory are used to experiment." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.577, + 0.483, + 0.773 + ], + "angle": 0, + "content": "Baseline models. We compared three distinct SOTA prompt learning approaches: (1) methods focusing on designing trainable models with small-scale parameters on the client side, such as CoOp [40], CoCoOp [39], FedVPT-D[14], FedCLIP [18], and FedTPG [26]; (2) methods that design prompt generators on the server side to reduce communication costs, such as PromptFL [9]; and (3) methods that emphasize both semantic and domain-specific prompts, such as ADAPT [33]. To show fair competition, the results of subsequent experiments were either the result of their papers or reproduced from source code provided by their authors." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.79, + 0.312, + 0.809 + ], + "angle": 0, + "content": "5. Experiments Analysis" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.817, + 0.483, + 0.848 + ], + "angle": 0, + "content": "5.1. Federated Domain Generalization Performance" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.855, + 0.484, + 0.903 + ], + "angle": 0, + "content": "FedDSPG achieved effective FDG in scenarios where target domain distributions diverge significantly from source domains. Quantitative results across three" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.363, + 0.908, + 0.712 + ], + "angle": 0, + "content": "benchmark datasets are presented in Table 1 following the \"leave-one-out\" protocol. In general, our method outperforms all baseline methods to varying degrees. We gleaned two primary observations from the results. Firstly, our approach outperforms all baselines, achieving SOTA performance, which demonstrates the effectiveness of our proposed approach in addressing federated domain generalization challenges. This establishes FedDSPG as a highly referencable benchmark for future research on FDG. Secondly, existing methods, including methods focusing on designing small-scale parameters models (FedCLIP, FedTPG) and methods generating prompts on the server side (PromptFL), exhibit notable performance gaps. We suppose that the possible reason stems from two primary factors: (1) inadequate incorporation of domain-specific information in their prompt design strategies, and (2) insufficient adaptability of generated prompts to unseen target domain images. These potential risks are mitigated in our approach, as it designs domain-specific soft prompts structure and trains a specific generative model, enhancing the generalization and diversity of prompts to unseen domains." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.726, + 0.73, + 0.74 + ], + "angle": 0, + "content": "5.2. Cross-Dataset Transfer" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.749, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Having demonstrated FedDSPG's capability for domain generalization within a single dataset, we further conducted experiments on cross-dataset domain generalization. This represents a more challenging task, as the fundamental characteristics can exhibit significant variations across different datasets. Specifically, we adopt a rigorous experimental protocol where DomainNet serves as the source domain, while OfficeHome and PACS function as target domains. We compared FedDSPG with FedTPG (best-performing" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.957 + ], + "angle": 0, + "content": "2371" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.096, + 0.362, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.096, + 0.627, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.097, + 0.895, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.276, + 0.276, + 0.72, + 0.29 + ], + "angle": 0, + "content": "Figure 3. The accuracy of FedDSPG with different data distribution." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.29, + 0.907, + 0.382 + ], + "angle": 0, + "content": "
SourceTarget
DomainNetRePrClArAvgPACSAvg
FedTPG72.781.378.566.176.475.691.293.891.686.190.7
Ours75.686.082.570.480.479.894.595.393.588.693.0
△+3.9+4.7+4.0+4.3+4.0+4.2+3.3+1.5+1.9+2.5+2.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.39, + 0.907, + 0.418 + ], + "angle": 0, + "content": "Table 2. The results for cross-dataset transfer experiment. The model is trained on the DomainNet dataset and inference on the Office-Home and PACS dataset." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.434, + 0.482, + 0.556 + ], + "angle": 0, + "content": "baseline), with the primary results detailed in Table 2. Overall, both methods experienced a decline in performance when confronted with the gap between datasets. However, FedDSPG consistently outperformed FedTPG on the target dataset by a substantial margin. This further underscores our method's remarkable capability to enhance model generalization across diverse domains." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.574, + 0.332, + 0.59 + ], + "angle": 0, + "content": "5.3. Data Distribution Analysis" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.599, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Considering client data distributions are inherently non-IID in federated learning scenarios, we conducted a comprehensive empirical analysis examining how varying data distributions affect FedDSPG's performance. Specifically, we varied the overlap ratio \\( r \\) of data domains across clients, setting it to \\( 0\\% \\), \\( 10\\% \\), \\( 20\\% \\), and \\( 50\\% \\). Specifically, an overlap ratio of \\( r\\% \\) implies that \\( r\\% \\) of the domains are shared among multiple clients, while the remaining \\( 1 - r\\% \\) of the domains are unique to individual clients. Figure 3 shows F1/accuracy across different domain distribution settings. The model exhibits greater stability on the Office-Home dataset compared to the other evaluation datasets, and a slight performance improvement is observed only when the overlap ratio reaches \\( 50\\% \\). We attribute this enhanced stability to the higher number of domains, which likely endows the model with greater robustness against distribution shifts. This finding underscores the importance of domain diversity in enhancing the generalization and stability of" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.435, + 0.712, + 0.451 + ], + "angle": 0, + "content": "federated learning models." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.467, + 0.667, + 0.483 + ], + "angle": 0, + "content": "5.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.491, + 0.907, + 0.643 + ], + "angle": 0, + "content": "Ablation on the different prompts structure. We classify prompts into 4 types (fixed, learnable, domain-specific, generative), and three representative SOTA models from baseline methods for comparison: PromptFL [9], and ADAPT [33], FedTPG [26]. PromptFL represents methods that design a generator on the server side to produce prompts. ADAPT represents methods emphasizing domain-specific information. FedTPG represents methods focusing on designing single trainable networks to optimize prompts." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.644, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To validate the effectiveness of our approach and the improvements, several ablation models with different prompt structures were trained. These models include: Our model with hand-crafted domain prompts (Ours-HDP), which replaces DSPs with structured prompts (i.e., \"a photo of a [class]\"). Our model with conditional soft prompts (Ours-CSP), which replaces DSPs with conditional soft prompts (i.e., \\( t_i = [v]_1[v]_2 \\ldots [v]_{M_1} \\)). Our model without a generative model (Ours-WGM), which omits the second training step. The results are presented in Table 3. Ours-HDP performs the worst, proving learnable prompt designing are crucial for performance. Although both ADAPT and Ours-WGM have domain-specific prompt, the improvement in results is not significant. However, Ours-CSP outperforms Ours-WGM and other baseline methods, underscoring the crucial" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2372" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.089, + 0.9, + 0.23 + ], + "angle": 0, + "content": "
MethodPrompt TypeRePrClArAvg
FixedLearnableDomain-specificGenerative
PromptFL [9]✓✓89.688.271.679.882.3
ADAPT [33]✓✓90.390.568.282.682.9
FedTPG [26]✓89.089.771.184.783.6
Ours-HDP✓✓86.181.470.273.277.7
Ours-CSP✓✓90.990.873.184.184.6
Ours-WGM✓✓88.187.871.682.482.5
Ours✓✓✓90.591.073.885.885.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.157, + 0.24, + 0.838, + 0.255 + ], + "angle": 0, + "content": "Table 3. Comparisons with different design of prompt on the Office-Home dataset for FDG performance." + }, + { + "type": "table", + "bbox": [ + 0.108, + 0.258, + 0.89, + 0.305 + ], + "angle": 0, + "content": "
α0.00.10.20.30.40.50.60.70.80.91.0
Accuracy (%)84.1584.8985.3085.2884.5684.3284.6784.0183.9583.5083.39
" + }, + { + "type": "table_caption", + "bbox": [ + 0.307, + 0.315, + 0.688, + 0.329 + ], + "angle": 0, + "content": "Table 4. Accuracy of FedDSPG with different coefficient \\( \\alpha \\) ." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.353, + 0.296, + 0.442 + ], + "angle": 0, + "content": "
Modeacc.
w/momentum85.3
w/o momentum84.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.14, + 0.445, + 0.252, + 0.458 + ], + "angle": 0, + "content": "(a) Prompt update." + }, + { + "type": "table", + "bbox": [ + 0.308, + 0.353, + 0.48, + 0.443 + ], + "angle": 0, + "content": "
epochs/roundacc.
0.585.3
185.3
284.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.328, + 0.445, + 0.455, + 0.458 + ], + "angle": 0, + "content": "(b) Comm. frequency" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.469, + 0.483, + 0.498 + ], + "angle": 0, + "content": "Table 5. Ablation studies on different momentum update and communication frequency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.507, + 0.483, + 0.568 + ], + "angle": 0, + "content": "role of the generative model in enhancing generalizability. Ultimately, our method achieves superior performance compared to all baselines, demonstrating the effective integration of DSPs and the generative model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.483, + 0.854 + ], + "angle": 0, + "content": "Momentum update and communication frequency. As elaborated in Section 3.3, we employ a momentum-based approach to mitigate the performance degradation caused by large gradient variations. We conduct comparative experiments on parameters with and without momentum updates. As presented in Table 5a, the results reveal a 0.8 percentage point decrease in accuracy without momentum. The experimental results for the momentum hyperparameter \\(\\alpha\\) ranging from 0 to 1 are shown in Table 4, with \\(\\alpha = 0.2\\) yielding the best performance. Additionally, we also test communication frequency, varying it to 0.5, 1, and 2 training epochs per communication round. Results (Table 5b) showed no performance gain from more frequent aggregation compared to the default 1 epoch per round, while lower frequencies led to accuracy declines. This suggests that FedDSPG can achieve good performance by maintaining the default update frequency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Number of clients and shots. We conducted ablation studies on the number of clients and shots, the results are displayed in Figure 4. Regarding the number" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.359, + 0.707, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.709, + 0.359, + 0.898, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.516, + 0.906, + 0.543 + ], + "angle": 0, + "content": "Figure 4. Ablation studies on different number of clients and shots." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.556, + 0.906, + 0.647 + ], + "angle": 0, + "content": "of clients, although an inevitable performance drop occurs for all methods as the number of clients increases, FedDSPG exhibits only a marginal decline (approximately \\(1\\%\\)). For the number of shots, FedDSPG outperforms baseline methods when the number of shots exceeds one." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.661, + 0.641, + 0.677 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.686, + 0.907, + 0.899 + ], + "angle": 0, + "content": "This work introduces FedDSPG, a novel federated domain generalization approach that leverages domain-specific soft prompts (DSPs) to enhance model generalization in non-IID federated learning scenarios. By integrating domain and content knowledge into a generative model, FedDSPG generates diverse prompts for unseen target domains, improving performance in unknown domains. Through extensive evaluations, we demonstrate our method's superior performance against existing baselines, while also revealing the untapped potential of generative techniques in FDG scenarios. These findings provide both immediate practical value and long-term research directions for advancing FL in real-world applications." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2373" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.194, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.199 + ], + "angle": 0, + "content": "[1] Sikai Bai, Jiewei Zhang, Shuaicheng Li, Song Guo, Jingcai Guo, Jun Hou, Tao Han, and Xiaocheng Lu. Diprompt: Disentangled prompt tuning for multiple latent domain generalization in federated learning. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.2, + 0.484, + 0.257 + ], + "angle": 0, + "content": "[2] Shuanghao Bai, Yuedi Zhang, Wanqi Zhou, Zhirong Luan, and Badong Chen. Soft prompt generation for domain generalization. In European Conference on Computer Vision, pages 434-450. Springer, 2025. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.257, + 0.483, + 0.311 + ], + "angle": 0, + "content": "[3] Tianyu Cui, Hongxia Li, Jingya Wang, and Ye Shi. Harmonizing generalization and personalization in federated prompt learning. International Conference on Machine Learning, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.312, + 0.483, + 0.382 + ], + "angle": 0, + "content": "[4] Xin Ding, Yongwei Wang, Zuheng Xu, William J Welch, and Z Jane Wang. Ccgan: Continuous conditional generative adversarial networks for image generation. In International Conference on Learning Representations, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.383, + 0.483, + 0.451 + ], + "angle": 0, + "content": "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations, ICLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.452, + 0.483, + 0.494 + ], + "angle": 0, + "content": "[6] Yuqi Fang, Pew-Thian Yap, Weili Lin, Hongtu Zhu, and Mingxia Liu. Source-free unsupervised domain adaptation: A survey. Neural Networks, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.495, + 0.483, + 0.55 + ], + "angle": 0, + "content": "[7] Haozhe Feng, Zhaoyang You, Minghao Chen, Tianye Zhang, Minfeng Zhu, et al. Kd3a: Unsupervised multisource decentralized domain adaptation via knowledge distillation. In ICML, pages 3274-3283, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.483, + 0.607 + ], + "angle": 0, + "content": "[8] Chunjiang Ge, Rui Huang, Mixue Xie, Zihang Lai, Shiji Song, Shuang Li, and Gao Huang. Domain adaptation via prompt learning. IEEE Transactions on Neural Networks and Learning Systems, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.608, + 0.483, + 0.677 + ], + "angle": 0, + "content": "[9] Tao Guo, Song Guo, Junxiao Wang, Xueyang Tang, and Wenchao Xu. Promptfl: Let federated participants cooperatively learn prompts instead of models-federated learning in age of foundation model. IEEE Transactions on Mobile Computing, 2023. 1, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.678, + 0.483, + 0.718 + ], + "angle": 0, + "content": "[10] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.719, + 0.483, + 0.789 + ], + "angle": 0, + "content": "[11] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. 9th International Conference on Learning Representations, ICLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.483, + 0.873 + ], + "angle": 0, + "content": "[12] Xuefeng Hu, Ke Zhang, Lu Xia, Albert Chen, Jiajia Luo, Yuyin Sun, Ken Wang, Nan Qiao, Xiao Zeng, Min Sun, et al. Reclip: Refine contrastive language image pretraining with source free domain adaptation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2994-3003, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.903 + ], + "angle": 0, + "content": "[13] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.149 + ], + "angle": 0, + "content": "Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[14] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[15] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. In Proceedings of the IEEE International Conference on Computer Vision, pages 5542-5550, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[16] Xinhui Liu, Wei Xi, Wen Li, Dong Xu, Gairui Bai, and Jizhong Zhao. Co-mda: Federated multi-source domain adaptation on black-box models. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.4 + ], + "angle": 0, + "content": "[17] Xinhui Liu, Zhenghao Chen, Luping Zhou, Dong Xu, Wei Xi, Gairui Bai, Yihan Zhao, and Jizhong Zhao. Ufda: Universal federated domain adaptation with practical assumptions. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 14026-14034, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.905, + 0.456 + ], + "angle": 0, + "content": "[18] Wang Lu, Xixu Hu, Jindong Wang, and Xing Xie. Fedclip: Fast generalization and personalization for clip in federated learning. In ICLR Workshop on Distributed and Private Machine Learning, 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.905, + 0.512 + ], + "angle": 0, + "content": "[19] Brendan McMahan, Eider Moore, et al. Communication-efficient learning of deep networks from decentralized data. In Artificial Intelligence and Statistics, pages 1273-1282. PMLR, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.512, + 0.905, + 0.553 + ], + "angle": 0, + "content": "[20] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. Advances in Neural Information Processing Systems, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.905, + 0.608 + ], + "angle": 0, + "content": "[21] Mehryar Mohri, Gary Sivek, and Ananda Theertha Suresh. Agnostic federated learning. In International Conference on Machine Learning, pages 4615-4625. PMLR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.608, + 0.905, + 0.664 + ], + "angle": 0, + "content": "[22] Nihal V Nayak, Peilin Yu, and Stephen H Bach. Learning to compose soft prompts for compositional zero-shot learning. International Conference on Learning Representations, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.664, + 0.905, + 0.735 + ], + "angle": 0, + "content": "[23] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1406-1415, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[24] Xingchao Peng, Zijun Huang, Yizhe Zhu, and Kate Saenko. Federated adversarial domain adaptation. International Conference on Learning Representations, ICLR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.86 + ], + "angle": 0, + "content": "[25] Chen Qiu, Xingyu Li, Chaithanya Kumar Mummadi, Madan Ravi Ganesh, Zhenzhen Li, Lu Peng, and Wan-Yi Lin. Text-driven prompt generation for vision-language models in federated learning. International Conference on Learning Representations, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.903 + ], + "angle": 0, + "content": "[26] Chen Qiu, Xingyu Li, Chaithanya Kumar Mummadi, Madan Ravi Ganesh, Zhenzhen Li, Lu Peng, and Wan-Yi Lin. Federated text-driven prompt generation for" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2374" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "vision-language models. In The Twelfth International Conference on Learning Representations, 2024. 1, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[28] Kuo Shi, Jie Lu, Zhen Fang, and Guangquan Zhang. Unsupervised domain adaptation enhanced by fuzzy prompt learning. IEEE Transactions on Fuzzy Systems, 32:4038-4048, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[29] Lei Song, Chunguang Ma, Guoyin Zhang, and Yun Zhang. Privacy-preserving unsupervised domain adaptation in federated setting. IEEE Access, 8:143233-143240, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.39 + ], + "angle": 0, + "content": "[30] Yue Tan, Chen Chen, Weiming Zhuang, Xin Dong, Lingjuan Lyu, and Guodong Long. Is heterogeneity notorious? taming heterogeneity to handle test-time shift in federated learning. Advances in Neural Information Processing Systems, 36, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.482, + 0.461 + ], + "angle": 0, + "content": "[31] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5018-5027, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.463, + 0.482, + 0.532 + ], + "angle": 0, + "content": "[32] Anna Vettoruzzo, Mohamed-Rafik Bouguelia, Joaquin Vanschoren, Thorsteinn Rognvaldsson, and KC Santosh. Advances and challenges in meta-learning: A technical review. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 4763-4779, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.534, + 0.482, + 0.575 + ], + "angle": 0, + "content": "[33] Guoyizhe Wei, Feng Wang, Anshul Shah, and Rama Chellappa. Learning to prompt your domain for vision-language models. In arXiv, 2024. 1, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.577, + 0.482, + 0.646 + ], + "angle": 0, + "content": "[34] Fu-En Yang, Chien-Yi Wang, and Yu-Chiang Frank Wang. Efficient model personalization in federated learning via client-specific prompt generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19159–19168, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.648, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[35] Elad Ben Zaken, Shauli Ravfogel, and Yoav Goldberg. Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. Annual Meeting of the Association for Computational Linguistics, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[36] Xin Zhang, Shixiang Shane Gu, Yutaka Matsuo, and Yusuke Iwasawa. Domain prompt learning for efficiently adapting clip to unseen domains. Transactions of the Japanese Society for Artificial Intelligence, 38(6):B-MC2_1, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[37] Haodong Zhao, Wei Du, Fangqi Li, Peixuan Li, and Gongshen Liu. Fedprompt: Communication-efficient and privacy-preserving prompt tuning in federated learning. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[38] Ke Zhao, Junchen Hu, Haidong Shao, and Jiabei Hu. Federated multi-source domain adversarial adaptation" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.092, + 0.905, + 0.133 + ], + "angle": 0, + "content": "framework for machinery fault diagnosis with data privacy. Reliability Engineering & System Safety, 236:109246, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[39] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.206, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[40] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9): 2337-2348, 2022. 2, 4, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "2375" + } + ] +] \ No newline at end of file diff --git a/2025/Federated Domain Generalization with Domain-specific Soft Prompts Generation/b70289dd-c48a-4e53-8732-8d816d04b0c4_origin.pdf b/2025/Federated Domain Generalization with Domain-specific Soft Prompts Generation/b70289dd-c48a-4e53-8732-8d816d04b0c4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..79e21ba6068e409d5831a87415955d40d65c3214 --- /dev/null +++ b/2025/Federated Domain Generalization with Domain-specific Soft Prompts Generation/b70289dd-c48a-4e53-8732-8d816d04b0c4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f12f4240df4ace5cc866446bb48d6a76467a8eec67f8dbe058c91564178cf7db +size 611848 diff --git a/2025/Federated Domain Generalization with Domain-specific Soft Prompts Generation/full.md b/2025/Federated Domain Generalization with Domain-specific Soft Prompts Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2c4df8a6ebd4f2eccf77ac231e19cce0ad7c1af5 --- /dev/null +++ b/2025/Federated Domain Generalization with Domain-specific Soft Prompts Generation/full.md @@ -0,0 +1,260 @@ +# Federated Domain Generalization with Domain-specific Soft Prompts Generation + +Jianhan Wu * Xiaoyang Qu * Zhangcheng Huang Jianzong Wang † Ping An Technology (Shenzhen) Co., Ltd. + +wujianhan@mail.ustc.edu.cn, quxiaoy@gmail.com + +william.huang@connect.polyu.hk, jzwang@188.com + +# Abstract + +Prompt learning has become an efficient paradigm for adapting CLIP to downstream tasks. Compared with traditional fine-tuning, prompt learning optimizes a few parameters yet yields highly competitive results, especially appealing in federated learning for computational efficiency. engendering domain shift among clients and posing a formidable challenge for downstream-task adaptation. Existing federated domain generalization (FDG) methods based on prompt learning typically learn soft prompts from training samples, replacing manually designed prompts to enhance the generalization ability of federated models. However, these learned prompts exhibit limited diversity and tend to ignore information from unknown domains. We propose a novel and effective method from a generative perspective for handling FDG tasks, namely federated domain generalization with domain-specific soft prompts generation (FedDSPG). Specifically, during training, we introduce domain-specific soft prompts (DSPs) for each domain and integrate content and domain knowledge into the generative model among clients. In the inference phase, the generator is utilized to obtain DSPs for unseen target domains, thus guiding downstream tasks in unknown domains. Comprehensive evaluations across several public datasets confirm that our method outperforms existing strong baselines in FDG, achieving state-of-the-art results. + +# 1. Introduction + +Pre-trained vision-language models (VLMs) [13], such as Contrastive Language Image Pretraining (CLIP) [12, 18, 27], have emerged as a powerful technique to + +adapt to downstream tasks. By introducing learnable prompts [1, 9, 37] into the text or vision encoders of CLIP, this approach allows the model to dynamically adjust its representations with minimal fine-tuning. Unlike traditional fine-tuning methods [35] that require extensive parameter updates, prompt learning optimizes only a small fraction of parameters, significantly reducing computational costs while maintaining high performance. This approach has been extensively adopted in federated learning (FL) [19, 21] paradigms, particularly in scenarios characterized by substantial computational and communication overheads between server and client devices. While federated learning based on prompt learning demonstrates substantial privacy and computational benefits, a key challenge persists: data on each node is often collected in a non-IID manner in real-world scenarios, leading to domain shift [16, 30] between nodes. For example, one device may take photos mostly indoors, while another mostly outdoors [24]. The presence of the domain shift issue renders the direct application of learning with hand-crafted prompts (such as "a photo of a [class]") in federated learning incapable of achieving satisfactory outcomes [33]. + +To address the domain shift problem in FL, federated domain generalization (FDG) [6, 16, 17, 29] further considers that the test data (target domain) is unavailable during decentralized training from distributed source domains. Some methods have optimized soft prompts [14, 25] across different clients by treating them as learnable vectors and refining them via backpropagation. For example, Promptfl [9] shares prompt vectors instead of model parameters to enhance communication efficiency and privacy in FL. pFedPG [34] introduced a domain bank mechanism to embed textual domain information into soft prompts. Recently, ADAPT [33] and FedTPG [26] used lightweight neural networks (e.g., adapter [11]) to optimize soft prompts conditioned on images through residual or concatenation. + +tion mechanisms [3], which improves generalization in a scalable manner. Despite recent advances, these methods either rely heavily on the diversity of client-specific data or use a simple adapter layer to estimate directly the relationship between soft prompts and images, leading to the learned prompts failing to fully capture the complex and diverse cross-domain feature [2]. + +To tackle the aforementioned challenges, we propose a novel and effective federated domain generalization method with domain-specific soft prompt generation (FedDSPG). Our approach reconceptualizes the prompt learning mechanism from a generative perspective, leveraging a generative model to produce domain-specific soft prompts (DSPs). Through an adversarial game manner between the generator and the discriminator, the generated DSPs exhibit greater diversity and generalization in both form and semantics. This enables our prompt generator model to acquire more generalized and robust features, facilitating adaptation to complex downstream tasks. The architectural details of our proposed approach are comprehensively depicted in Figure 1, comprising a training phase with two steps and an inference phase. Specifically, in the first step of the training phase, we avoid using hand-crafted prompts or soft prompts that lack sufficient attribute focus and specific content design. We elaborately design DSPs, composed of domain and class name vectors that map domain and content information, representing the optimal prompts for images from each client. In the second step, images and their corresponding DSPs are fed into a simple yet effective conditional generative model [4], which is trained to further optimize the DSPs by learning domain and content knowledge of images. Given the significant variance in tunable prompt embedding parameters across data-heterogeneous clients, the traditional FedAvg [19] aggregation method yields suboptimal results (see Table 5a for comparative experiments). To mitigate this, we propose using a momentum-based method to aggregate tunable parameters in the first training step. During the inference stage, the generative model's generator component will generate DSPs for target domain images, which subsequently facilitate predictions in downstream tasks. + +Our key contributions are summarized below: + +- We are the first to optimize prompt learning from a generative perspective in FL and propose a novel soft prompt generation paradigm for FDG tasks. +- Unlike the current CLIP-based methods that rely on hand-crafted or naive soft prompts, we devise domain-specific soft prompt structures that significantly enhance generalization to unseen domains. + +Furthermore, we introduce a momentum-based aggregation mechanism to consistently refine performance. + +- Comprehensive empirical evaluations on standard benchmark datasets corroborate that our proposed method achieves state-of-the-art performance. Additionally, comprehensive ablation research indicates the role played by each strategy in the increased efficacy. + +# 2. Related Work + +# 2.1. Prompt Learning for Vision Language Models + +Prompt learning, also known as prompt tuning, provides an efficient alternative to conventional fine-tuning approaches for pre-trained models. By focusing on tuning only learnable prompt tokens attached to the input of models like CLIP. CoOp [40] first introduced soft prompts into vision-language models, demonstrating that appropriate prompts can enhance performance in downstream tasks. CoCoOp [39] addresses the overfitting issue in CoOp through conditioned prompt learning, specifically residual prompts. DPL [36] achieved effective domain generalization by training a lightweight prompt generator (a linear network). FUZZLE [28] integrates fuzzy C-means clustering into prompt learning to optimize cross-domain alignment between source and target distributions. In these studies, soft prompts are typically generated by simple tunable models (e.g., adapter) that predominantly rely on the diversity of client-specific data, which yields unsatisfactory results for the unseen domain. The research most pertinent to ours is SPG [2], which employs conditional adversarial generative networks to generate prompts but lacks domain-specific design, resulting in suboptimal performance in federated learning scenarios. We not only meticulously designed DSPs but also tailored the federated aggregation algorithm to better execute FDG. + +# 2.2. Federated Domain Generalization + +FDG aims to address domain shifts between source and target domains in federated settings where data is distributed across clients. Traditional FDG methods generally fall into two types. The first is domain alignment approaches that focus on learning domain-invariant representations by reducing source domain discrepancies. For instance, in KD3A [7], domain differences are reduced by explicitly matching the mean embeddings of data from different domains. These methods' disadvantage lies in stringent assumptions about data distribution that are hard to satisfy in practice. The second type involves data augmentation + +![](images/e243d6ad935e909c52ba3bd94fef06d2a1fbb81fda1950caee0bbaf6fbc6aba4.jpg) +Figure 1. The overall training structure of FedDSPG consists of two main stages: domain-specific prompts tuning and domain prompt generation. The first stage produces DSPs with compositional soft prompting technique [22], which encode the content and style information with large-scale pretrained VLMs. In the second stage, a GAN-based training method is employed to learn the latent data distribution, thereby generating soft prompts that adhere to the intrinsic data patterns. For the inference stage, the generative model's generator component will generate DSPs for target domain images, producing enhanced predictions for downstream tasks. Additionally, the momentum-based aggregation is conducted to enhance model generalization for FDG. + +methods that aim to expand the training data distribution to enhance out-of-distribution generalization. Additional strategies include meta-learning [32], adversarial learning [38], and more. Recently, PromptFL [9], ADAPT [33], and FedTPG [26] introduced prompt learning to extract semantic information, aligning domain information on implicit textual features to improve the model's generalization capacity. To further advance this field, we conduct an in-depth exploration of this efficient paradigm. Departing from the previous prompt learning designs that rely on handcrafted prompts or simple neural network, we present a novel and effective FDG structure that leverages generative modeling techniques to learn domain-specific soft prompts, enhancing the generalization of prompts to unseen target domains. + +# 3. Methodology + +In this section, we first introduce the preliminary knowledge of problem setup of FDG. Following that, the proposed method FedDSPG will be introduced in detail, which includes two training steps of domain-specific soft prompts tuning and prompts generative model tuning, as well as federated parameter aggregation and federated inference. + +# 3.1. Preliminaries + +In the standard FDG problem, there is a source domain dataset $\mathcal{D}_S = (\mathbf{x}_i^s,\mathbf{y}_i^s)_{i = 1}^n$ with labeled data, where $(\mathbf{x}_i^s,\mathbf{y}_i^s)$ is the data distribution for client $i^{\prime}s$ input space and corresponding labels, with n being the number of source domains. The target domain dataset can be defined as $\mathcal{D}_T = (\mathbf{x}_i^t,\mathbf{y}_i^t)_{i = 1}^m$ , where the number of target domains $m$ is typically configured as one per client for experiment evaluation. FDG seeks to develop models that can achieve robust performance on target domains while being trained exclusively on distributed source domain data. The primary challenge lies in addressing the domain shift between source and target domains, which is exacerbated by the non-IID nature of federated data across clients. + +Taking image classification as an example, let $X$ and $Y$ represent the input and label spaces, respectively. Previous domain generalization methods concentrated on training a domain-invariant model for the mapping $F: X \to Y$ from images to labels. Prompt learning methods embed learnable prompt vectors $V$ into the input to leverage the generalization ability of VLMs, reformulating the mapping as $F: (X, V) \to Y$ . Unlike existing prompt learning methods that focus primarily on fine-tuning the prompt vectors $V$ , we propose a novel prompt learning paradigm that utilizes a + +![](images/29025138b121b72cdea1c30a6126a06f6fa9f01b7a82895709a109cfd55ebc60.jpg) +Figure 2. The structure design of the second training step is based on CcGAN. In the generator, input noise $z$ is concatenated with fake image embeddings to form the joint latent representation, which is used to generate DSPs. The discriminator distinguishes the authenticity of the generated DSPs from the real image. + +generative model $G$ to generate soft prompts, reformulating the mapping as $F:(X,G(X))\to Y$ . The objective is to develop a generative framework that simultaneously captures both domain-specific and domain-invariant representations, enabling the dynamic generation of generalized prompts for unseen target domains. + +# 3.2. Two Training Steps + +FedDSPG is implemented in two steps: domain-specific soft prompts tuning and prompts generative model tuning. Domain-specific prompts tuning aims to embed domain and class information into prompts, and the goal of prompts generative model tuning is to incorporate such information into the generative model. Notably, domain distribution is learned by the generative model, enabling the generation of instance-specific prompts for each image, guaranteeing prompt diversity and integrating domain-specific information. Below are the detailed designs of FedDSPG. + +Domain-specific Soft Prompts Tuning. We adopt CLIP [27] as our backbone, leveraging its contrastive learning framework trained on extensive image-text pairs to establish a joint embedding space through dedicated image $(f)$ and text $(g)$ encoders. By describing the input text as structured prompt $t_i$ : "a photo of a [CLASS]", the similarity of image representations in the same category $i$ can be maximized to align semantic representations: + +$$ +P (\hat {y} = i | x) = \frac {\exp (S i m (g (t _ {i}) , f (x)) / \tau)}{\sum_ {k = 1} ^ {K} \exp (S i m (g (t _ {k}) , f (x)) / \tau)} \tag {1} +$$ + +Where $Sim(\cdot)$ is the cosine similarity and $\tau$ denotes temperature parameter. Since the soft prompts + +of previous work (e.g., CoOp [40] and SPG [2] $t_i = [v]_1[v]_2 \ldots [v]_{M_1}[\mathrm{CLASS}]_i$ ) cannot deal well with the distribution shift between domains, we design domain-specific soft prompts to serve as the optimal prompt representations for images from each client. Specifically, each image from a client source domain is associated with unique DSPs (seen in Figure 1). The tunable vectors of DSPs contain two parts, a domain-invariant context and a domain-specific context, denoted as: + +$$ +t _ {i} ^ {d} = [ v ] _ {1} [ v ] _ {2} \dots [ v ] _ {M _ {1}} [ u ] _ {1} ^ {d} [ u ] _ {2} ^ {d} \dots [ u ] _ {M _ {2}} ^ {d} [ \mathrm {C L S} ] _ {i} \tag {2} +$$ + +Where $[u]_{M_2}^d$ denotes domain-specific tokens, which have the same dimension as word embeddings. $M_{1}$ and $M_2$ are the number of context tokens and domain tokens incorporated in the prompt structure, respectively. These DSPs are optimized through cross-entropy loss minimization during training on domain-specific data: + +$$ +\mathbf {v} _ {c _ {j}} ^ {d _ {i} *} = \arg \min _ {\mathbf {v}} \mathbb {E} _ {\mathbf {x} _ {j} ^ {d _ {i}}, y _ {j} ^ {d _ {i}}} \left[ - \log p \left(y _ {j} ^ {d _ {i}} \mid \mathbf {x} _ {j} ^ {d _ {i}}, \mathbf {t} _ {j} ^ {d _ {i}}\right) \right] \quad (3) +$$ + +Where $(\mathbf{x}_j^{d_i},y_j^{d_i})$ represents the image-label pairs from client $j^{\prime}s$ training data in $i$ -th domain. DSPs integrate domain and class information in a more coherent and structured manner than prior approaches. These DSPs constitute optimized prompt representations for each client domain, encoding comprehensive domain characteristics, which encapsulate rich domain and class information for training a generative model in the second training step. + +Prompts Generative Model Tuning. To learn the domain and class knowledge in DSPs, considering the communication cost requirements (e.g., methods based on diffusion models [10] require multi-step denoising, resulting in high communication costs. The corresponding experiments can be found in the supplementary materials.) and input data continuity of federated learning, we employ an efficient generative model—Continuous Conditional Generative Adversarial Network (CcGAN) [4] to validate the effectiveness of our approach. Similar to conditional GAN [20], we manipulate the generation process by constraining the additional information (i.e., DSPs generated in the first training step) of the CcGAN. The goal is to learn the generator's distribution $p_{g}$ over DSPs $\mathbf{v}_{c_j}^{d_i}$ , effectively transferring generalization capabilities from prompts to the generator itself. The architecture consists of two core components: a generator $G$ that synthesizes outputs conditioned on auxiliary information and a discriminator $D$ that assesses the authenticity of real data and fake outputs. These components facilitate an adversarial process that enhances the model's generative performance. + +Specifically, the detailed training process of the prompts generative model is shown in Figure 2. The generator concatenates noise variables $\mathbf{z}$ with fake-batch embeddings $f(\mathbf{x})$ to form joint latent codes $[\mathbf{z}, f(\mathbf{x})]$ , where $\mathbf{x}$ is randomly sampled from client data. Then the generator subsequently produces the corresponding DSPs by transforming the joint latent representation $[\mathbf{z}, f(\mathbf{x})]$ . In the discriminator, we input both the genuine image embeddings $f(\mathbf{x})$ from real batches and the generated DSPs with their associated image embeddings $f(\mathbf{x})$ to distinguish real from fake. Therefore, the two-player min-max game objective function $V(G, D)$ for CcGAN can be formally expressed as: + +$$ +\begin{array}{l} \min _ {G} \max _ {D} V (G, D) = \mathbb {E} _ {\mathbf {v} \sim p _ {\mathrm {v}} (\mathbf {v})} [ \log D (\mathbf {v} \mid f (\mathbf {x})) ] \\ + \mathbb {E} _ {\mathbf {z} \sim p _ {z} (\mathbf {z})} [ \log (1 - D (G (\mathbf {z} \mid f (\mathbf {x}))) ]) \tag {4} \\ \end{array} +$$ + +The process of prompts generative model tuning is to learn coarse-to-fine DSPs through an adversarial manner. This design not only promotes diversity in prompt generation but also significantly improves generalization across tasks and domains, facilitating robust performance in both seen and unseen conditions. + +# 3.3. Federated Parameters Aggregation + +We employ distinct aggregation strategies for the learnable parameters of prompt tuning and the generative model. Assuming a centralized parameter server for uniform aggregation—though our method also supports decentralized communication—clients upload their respective soft text embedding parameters and generative model parameters to the server during each communication round. The server averages the parameters from all n clients and redistributes them. While this parameter aggregation paradigm appears viable for our method, it introduces a minor issue for soft text embeddings. Specifically, given that FDG addresses domain shifts across clients, the domain components of DSPs vary significantly between clients. We observe that direct averaging causes abrupt changes in soft text embedding parameters, negatively impacting model performance. To mitigate this, we introduce momentum updates for soft prompt parameters, utilizing exponential moving averages to maintain parameter continuity and prevent destabilizing abrupt changes during the optimization process. The aggregation formula is as follows: + +$$ +[ w ] ^ {k} = \alpha [ w ] ^ {k - 1} + (1 - \alpha) [ w ] ^ {k - 2} \tag {5} +$$ + +Where $[w]^k$ , $[w]^{k-1}$ and $[w]^{k-2}$ denote the tunable soft prompt embedding parameters at the $k, k-1$ and + +$k - 2$ step, and $\alpha \in [0,1]$ serves as a hyperparameter governing the smoothing intensity of DSPs. Comprehensive ablation studies examining the impact of momentum updates are presented in table 5a. + +# 3.4. Inference with Generator + +The generator trained via CcGAN is utilized to generate DSPs for inference on each target image. The predicted probability that a target domain image $x$ belongs to the $i$ -th class is formally expressed as: + +$$ +p (y = i \mid \mathbf {x}) = \frac {\exp (\langle w _ {i} , f (\mathbf {x}) \rangle / \tau)}{\sum_ {j = 1} ^ {K} \exp (\langle w _ {j} , f (\mathbf {x}) \rangle / \tau)}, \tag {6} +$$ + +$$ +s. t. \quad w _ {j} = g ([ G (\mathbf {z} \mid f (\mathbf {x})), c _ {j} ]) \tag {7} +$$ + +Where $\tau$ represents the temperature parameter that modulates the sharpness of the output distribution, $g$ and $f$ represent the text and the image encoder respectively, and $K$ denotes the total number of classes. The term $G(\mathbf{z} \mid f(\mathbf{x}))$ signifies the generated DSPs by the generator. We use $c_{j}$ to represent the $j$ -th tokenized class label, with $[\cdot, \cdot]$ denoting the standard concatenation operation. In the generative paradigm, our approach achieves enhanced adaptability and dynamic prompts generation, which are essential for accommodating diverse input distributions and downstream task requirements. + +# 4. Experimental Setup + +Datasets. We conducted extensive experiments for image classification on domain generalization benchmark datasets: (1) Office-Home, a large-scale visual cross-domain dataset with 65 classes across four visually distinct domains, i.e., Art (Ar), Clipart (Cl), Product (Pr), and Real World (Re) [31]. (2) DomainNet, a largescale domain generalization benchmark dataset with 345 classes distributed among six domains, i.e., Clipart (Clp), Infograph (Inf), Painting (Pnt), Quickdraw (Qdr), Real (Rel), and Sketch (Skt) [23]. (3) PACS, featuring approximately 10,000 images from seven categories across four domains, including Photo (P), Sketch (S), Cartoon (C), and Art (A) [15]. For rigorous evaluation, we adopt the leave-one-domain-out protocol, designating each domain sequentially as the target while utilizing the remaining domains as source data. + +Implementation details. In FDG, all participating nodes maintain identical architectural configurations while ensuring complete domain data isolation between nodes. The data processing and division are followed [8] on the image generalization task. For our FedDSPG method, we employ a frozen CLIP backbone (ViT-B/16 [5]) pre-trained on large-scale vision-language data. The text and visual prompt tokens + +
DatasetOffice-HomeDomainNetPACS
Target DomainRePrClArAvgClpInfPntQdrRelSktAvgPACSAvg
CoOp[40]82.982.859.372.174.361.622.252.612.167.748.644.194.191.595.684.991.5
CoCoOp[39]85.081.162.568.674.372.423.360.816.472.760.551.097.594.697.987.594.4
FedVPT-D[14]85.085.261.176.076.863.343.074.854.887.267.165.099.296.598.187.295.3
FedCLIP[18]87.787.563.678.079.265.146.575.956.190.267.967.099.896.397.985.694.9
PromptFL[9]89.688.271.679.882.372.850.384.260.492.167.771.399.997.199.090.696.7
ADAPT[33]90.390.568.282.682.977.563.170.541.685.772.168.499.998.099.191.797.2
FedTPG[26]89.089.771.184.783.674.052.185.162.194.068.672.799.897.899.290.896.9
FedDSPG90.591.073.885.885.374.763.585.563.094.272.675.699.998.199.392.297.4
+ +Table 1. Accuracies (%) for the image classification task on Office-Home, DomainNet, and PACS datasets using CLIP with ViT-B/16 image encoder as frozen backbone. The domains displayed here are the target domains, and the others are the source domains; the best result is indicated in bold. + +share a unified dimensionality of 512. For the learning strategies, in the first training step, the DSPs are trained by using the Adam optimizer with the learning rate of $1 \times 10^{-5}$ , batch size of 32, and the context lengths $M_{1}$ and $M_{2}$ of the prompt number are 4. In the second training step, the CcGAN is trained by using the AdamW optimizer with weight decay $2 \times 10^{-5}$ and $\beta_{1} = 0.9$ , $\beta = 0.999$ , starting with a learning rate of $1 \times 10^{-4}$ for the OfficeHome and DomainNet datasets. The 100 epochs are trained for both our method and the baseline models for accelerating convergence, then the FedAvg and momentum-based aggregation algorithms are executed after every epoch. 8 Tesla-V100 with 16GB memory are used to experiment. + +Baseline models. We compared three distinct SOTA prompt learning approaches: (1) methods focusing on designing trainable models with small-scale parameters on the client side, such as CoOp [40], CoCoOp [39], FedVPT-D[14], FedCLIP [18], and FedTPG [26]; (2) methods that design prompt generators on the server side to reduce communication costs, such as PromptFL [9]; and (3) methods that emphasize both semantic and domain-specific prompts, such as ADAPT [33]. To show fair competition, the results of subsequent experiments were either the result of their papers or reproduced from source code provided by their authors. + +# 5. Experiments Analysis + +# 5.1. Federated Domain Generalization Performance + +FedDSPG achieved effective FDG in scenarios where target domain distributions diverge significantly from source domains. Quantitative results across three + +benchmark datasets are presented in Table 1 following the "leave-one-out" protocol. In general, our method outperforms all baseline methods to varying degrees. We gleaned two primary observations from the results. Firstly, our approach outperforms all baselines, achieving SOTA performance, which demonstrates the effectiveness of our proposed approach in addressing federated domain generalization challenges. This establishes FedDSPG as a highly referencable benchmark for future research on FDG. Secondly, existing methods, including methods focusing on designing small-scale parameters models (FedCLIP, FedTPG) and methods generating prompts on the server side (PromptFL), exhibit notable performance gaps. We suppose that the possible reason stems from two primary factors: (1) inadequate incorporation of domain-specific information in their prompt design strategies, and (2) insufficient adaptability of generated prompts to unseen target domain images. These potential risks are mitigated in our approach, as it designs domain-specific soft prompts structure and trains a specific generative model, enhancing the generalization and diversity of prompts to unseen domains. + +# 5.2. Cross-Dataset Transfer + +Having demonstrated FedDSPG's capability for domain generalization within a single dataset, we further conducted experiments on cross-dataset domain generalization. This represents a more challenging task, as the fundamental characteristics can exhibit significant variations across different datasets. Specifically, we adopt a rigorous experimental protocol where DomainNet serves as the source domain, while OfficeHome and PACS function as target domains. We compared FedDSPG with FedTPG (best-performing + +![](images/76026bb88c01c1995efe960033b888a0e278636babc23995bd4c688d1a2f9bef.jpg) + +![](images/1fb3f94cf79013e94ac01dc58f3d70cbdda093f19619477efd5fd83feb6a8c3e.jpg) + +![](images/beab1ee4bb344961cf6c66e3965c1da6048c8c3a004dadfb3824bb0f631fd6eb.jpg) + +Figure 3. The accuracy of FedDSPG with different data distribution. + +
SourceTarget
DomainNetRePrClArAvgPACSAvg
FedTPG72.781.378.566.176.475.691.293.891.686.190.7
Ours75.686.082.570.480.479.894.595.393.588.693.0
△+3.9+4.7+4.0+4.3+4.0+4.2+3.3+1.5+1.9+2.5+2.3
+ +Table 2. The results for cross-dataset transfer experiment. The model is trained on the DomainNet dataset and inference on the Office-Home and PACS dataset. + +baseline), with the primary results detailed in Table 2. Overall, both methods experienced a decline in performance when confronted with the gap between datasets. However, FedDSPG consistently outperformed FedTPG on the target dataset by a substantial margin. This further underscores our method's remarkable capability to enhance model generalization across diverse domains. + +# 5.3. Data Distribution Analysis + +Considering client data distributions are inherently non-IID in federated learning scenarios, we conducted a comprehensive empirical analysis examining how varying data distributions affect FedDSPG's performance. Specifically, we varied the overlap ratio $r$ of data domains across clients, setting it to $0\%$ , $10\%$ , $20\%$ , and $50\%$ . Specifically, an overlap ratio of $r\%$ implies that $r\%$ of the domains are shared among multiple clients, while the remaining $1 - r\%$ of the domains are unique to individual clients. Figure 3 shows F1/accuracy across different domain distribution settings. The model exhibits greater stability on the Office-Home dataset compared to the other evaluation datasets, and a slight performance improvement is observed only when the overlap ratio reaches $50\%$ . We attribute this enhanced stability to the higher number of domains, which likely endows the model with greater robustness against distribution shifts. This finding underscores the importance of domain diversity in enhancing the generalization and stability of + +federated learning models. + +# 5.4. Ablation Study + +Ablation on the different prompts structure. We classify prompts into 4 types (fixed, learnable, domain-specific, generative), and three representative SOTA models from baseline methods for comparison: PromptFL [9], and ADAPT [33], FedTPG [26]. PromptFL represents methods that design a generator on the server side to produce prompts. ADAPT represents methods emphasizing domain-specific information. FedTPG represents methods focusing on designing single trainable networks to optimize prompts. + +To validate the effectiveness of our approach and the improvements, several ablation models with different prompt structures were trained. These models include: Our model with hand-crafted domain prompts (Ours-HDP), which replaces DSPs with structured prompts (i.e., "a photo of a [class]"). Our model with conditional soft prompts (Ours-CSP), which replaces DSPs with conditional soft prompts (i.e., $t_i = [v]_1[v]_2 \ldots [v]_{M_1}$ ). Our model without a generative model (Ours-WGM), which omits the second training step. The results are presented in Table 3. Ours-HDP performs the worst, proving learnable prompt designing are crucial for performance. Although both ADAPT and Ours-WGM have domain-specific prompt, the improvement in results is not significant. However, Ours-CSP outperforms Ours-WGM and other baseline methods, underscoring the crucial + +
MethodPrompt TypeRePrClArAvg
FixedLearnableDomain-specificGenerative
PromptFL [9]✓✓89.688.271.679.882.3
ADAPT [33]✓✓90.390.568.282.682.9
FedTPG [26]✓89.089.771.184.783.6
Ours-HDP✓✓86.181.470.273.277.7
Ours-CSP✓✓90.990.873.184.184.6
Ours-WGM✓✓88.187.871.682.482.5
Ours✓✓✓90.591.073.885.885.3
+ +Table 3. Comparisons with different design of prompt on the Office-Home dataset for FDG performance. + +
α0.00.10.20.30.40.50.60.70.80.91.0
Accuracy (%)84.1584.8985.3085.2884.5684.3284.6784.0183.9583.5083.39
+ +Table 4. Accuracy of FedDSPG with different coefficient $\alpha$ . + +
Modeacc.
w/momentum85.3
w/o momentum84.5
+ +(a) Prompt update. + +
epochs/roundacc.
0.585.3
185.3
284.6
+ +(b) Comm. frequency + +Table 5. Ablation studies on different momentum update and communication frequency. + +role of the generative model in enhancing generalizability. Ultimately, our method achieves superior performance compared to all baselines, demonstrating the effective integration of DSPs and the generative model. + +Momentum update and communication frequency. As elaborated in Section 3.3, we employ a momentum-based approach to mitigate the performance degradation caused by large gradient variations. We conduct comparative experiments on parameters with and without momentum updates. As presented in Table 5a, the results reveal a 0.8 percentage point decrease in accuracy without momentum. The experimental results for the momentum hyperparameter $\alpha$ ranging from 0 to 1 are shown in Table 4, with $\alpha = 0.2$ yielding the best performance. Additionally, we also test communication frequency, varying it to 0.5, 1, and 2 training epochs per communication round. Results (Table 5b) showed no performance gain from more frequent aggregation compared to the default 1 epoch per round, while lower frequencies led to accuracy declines. This suggests that FedDSPG can achieve good performance by maintaining the default update frequency. + +Number of clients and shots. We conducted ablation studies on the number of clients and shots, the results are displayed in Figure 4. Regarding the number + +![](images/aa9ff88485566362e5cdd3b1f15debdec9984a37a1d4a7b574ff377f923e1d58.jpg) +Figure 4. Ablation studies on different number of clients and shots. + +![](images/f9614fc4e1b8a38f5b4fd168c68945884d4abbb8f4d8c561c02cb32b5f7c79ed.jpg) + +of clients, although an inevitable performance drop occurs for all methods as the number of clients increases, FedDSPG exhibits only a marginal decline (approximately $1\%$ ). For the number of shots, FedDSPG outperforms baseline methods when the number of shots exceeds one. + +# 6. Conclusion + +This work introduces FedDSPG, a novel federated domain generalization approach that leverages domain-specific soft prompts (DSPs) to enhance model generalization in non-IID federated learning scenarios. By integrating domain and content knowledge into a generative model, FedDSPG generates diverse prompts for unseen target domains, improving performance in unknown domains. Through extensive evaluations, we demonstrate our method's superior performance against existing baselines, while also revealing the untapped potential of generative techniques in FDG scenarios. These findings provide both immediate practical value and long-term research directions for advancing FL in real-world applications. + +# References + +[1] Sikai Bai, Jiewei Zhang, Shuaicheng Li, Song Guo, Jingcai Guo, Jun Hou, Tao Han, and Xiaocheng Lu. Diprompt: Disentangled prompt tuning for multiple latent domain generalization in federated learning. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1 +[2] Shuanghao Bai, Yuedi Zhang, Wanqi Zhou, Zhirong Luan, and Badong Chen. Soft prompt generation for domain generalization. In European Conference on Computer Vision, pages 434-450. Springer, 2025. 2, 4 +[3] Tianyu Cui, Hongxia Li, Jingya Wang, and Ye Shi. Harmonizing generalization and personalization in federated prompt learning. International Conference on Machine Learning, 2024. 2 +[4] Xin Ding, Yongwei Wang, Zuheng Xu, William J Welch, and Z Jane Wang. Ccgan: Continuous conditional generative adversarial networks for image generation. In International Conference on Learning Representations, 2021. 2, 4 +[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations, ICLR, 2021. 5 +[6] Yuqi Fang, Pew-Thian Yap, Weili Lin, Hongtu Zhu, and Mingxia Liu. Source-free unsupervised domain adaptation: A survey. Neural Networks, 2024. 1 +[7] Haozhe Feng, Zhaoyang You, Minghao Chen, Tianye Zhang, Minfeng Zhu, et al. Kd3a: Unsupervised multisource decentralized domain adaptation via knowledge distillation. In ICML, pages 3274-3283, 2021. 2 +[8] Chunjiang Ge, Rui Huang, Mixue Xie, Zihang Lai, Shiji Song, Shuang Li, and Gao Huang. Domain adaptation via prompt learning. IEEE Transactions on Neural Networks and Learning Systems, 2023. 5 +[9] Tao Guo, Song Guo, Junxiao Wang, Xueyang Tang, and Wenchao Xu. Promptfl: Let federated participants cooperatively learn prompts instead of models-federated learning in age of foundation model. IEEE Transactions on Mobile Computing, 2023. 1, 3, 6, 7, 8 +[10] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 4 +[11] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. 9th International Conference on Learning Representations, ICLR, 2021. 1 +[12] Xuefeng Hu, Ke Zhang, Lu Xia, Albert Chen, Jiajia Luo, Yuyin Sun, Ken Wang, Nan Qiao, Xiao Zeng, Min Sun, et al. Reclip: Refine contrastive language image pretraining with source free domain adaptation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2994-3003, 2024. 1 +[13] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen + +Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 1 +[14] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 1, 6 +[15] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. In Proceedings of the IEEE International Conference on Computer Vision, pages 5542-5550, 2017. 5 +[16] Xinhui Liu, Wei Xi, Wen Li, Dong Xu, Gairui Bai, and Jizhong Zhao. Co-mda: Federated multi-source domain adaptation on black-box models. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 1 +[17] Xinhui Liu, Zhenghao Chen, Luping Zhou, Dong Xu, Wei Xi, Gairui Bai, Yihan Zhao, and Jizhong Zhao. Ufda: Universal federated domain adaptation with practical assumptions. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 14026-14034, 2024. 1 +[18] Wang Lu, Xixu Hu, Jindong Wang, and Xing Xie. Fedclip: Fast generalization and personalization for clip in federated learning. In ICLR Workshop on Distributed and Private Machine Learning, 2023. 1, 6 +[19] Brendan McMahan, Eider Moore, et al. Communication-efficient learning of deep networks from decentralized data. In Artificial Intelligence and Statistics, pages 1273-1282. PMLR, 2017. 1, 2 +[20] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. Advances in Neural Information Processing Systems, 2014. 4 +[21] Mehryar Mohri, Gary Sivek, and Ananda Theertha Suresh. Agnostic federated learning. In International Conference on Machine Learning, pages 4615-4625. PMLR, 2019. 1 +[22] Nihal V Nayak, Peilin Yu, and Stephen H Bach. Learning to compose soft prompts for compositional zero-shot learning. International Conference on Learning Representations, 2022. 3 +[23] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1406-1415, 2019. 5 +[24] Xingchao Peng, Zijun Huang, Yizhe Zhu, and Kate Saenko. Federated adversarial domain adaptation. International Conference on Learning Representations, ICLR, 2019. 1 +[25] Chen Qiu, Xingyu Li, Chaithanya Kumar Mummadi, Madan Ravi Ganesh, Zhenzhen Li, Lu Peng, and Wan-Yi Lin. Text-driven prompt generation for vision-language models in federated learning. International Conference on Learning Representations, 2024. 1 +[26] Chen Qiu, Xingyu Li, Chaithanya Kumar Mummadi, Madan Ravi Ganesh, Zhenzhen Li, Lu Peng, and Wan-Yi Lin. Federated text-driven prompt generation for + +vision-language models. In The Twelfth International Conference on Learning Representations, 2024. 1, 3, 6, 7, 8 +[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 4 +[28] Kuo Shi, Jie Lu, Zhen Fang, and Guangquan Zhang. Unsupervised domain adaptation enhanced by fuzzy prompt learning. IEEE Transactions on Fuzzy Systems, 32:4038-4048, 2024. 2 +[29] Lei Song, Chunguang Ma, Guoyin Zhang, and Yun Zhang. Privacy-preserving unsupervised domain adaptation in federated setting. IEEE Access, 8:143233-143240, 2020. 1 +[30] Yue Tan, Chen Chen, Weiming Zhuang, Xin Dong, Lingjuan Lyu, and Guodong Long. Is heterogeneity notorious? taming heterogeneity to handle test-time shift in federated learning. Advances in Neural Information Processing Systems, 36, 2024. 1 +[31] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5018-5027, 2017. 5 +[32] Anna Vettoruzzo, Mohamed-Rafik Bouguelia, Joaquin Vanschoren, Thorsteinn Rognvaldsson, and KC Santosh. Advances and challenges in meta-learning: A technical review. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 4763-4779, 2024. 3 +[33] Guoyizhe Wei, Feng Wang, Anshul Shah, and Rama Chellappa. Learning to prompt your domain for vision-language models. In arXiv, 2024. 1, 3, 6, 7, 8 +[34] Fu-En Yang, Chien-Yi Wang, and Yu-Chiang Frank Wang. Efficient model personalization in federated learning via client-specific prompt generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19159–19168, 2023. 1 +[35] Elad Ben Zaken, Shauli Ravfogel, and Yoav Goldberg. Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. Annual Meeting of the Association for Computational Linguistics, 2021. 1 +[36] Xin Zhang, Shixiang Shane Gu, Yutaka Matsuo, and Yusuke Iwasawa. Domain prompt learning for efficiently adapting clip to unseen domains. Transactions of the Japanese Society for Artificial Intelligence, 38(6):B-MC2_1, 2023. 2 +[37] Haodong Zhao, Wei Du, Fangqi Li, Peixuan Li, and Gongshen Liu. Fedprompt: Communication-efficient and privacy-preserving prompt tuning in federated learning. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 1 +[38] Ke Zhao, Junchen Hu, Haidong Shao, and Jiabei Hu. Federated multi-source domain adversarial adaptation + +framework for machinery fault diagnosis with data privacy. Reliability Engineering & System Safety, 236:109246, 2023. 3 +[39] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 2, 6 +[40] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9): 2337-2348, 2022. 2, 4, 6 \ No newline at end of file diff --git a/2025/Federated Domain Generalization with Domain-specific Soft Prompts Generation/images.zip b/2025/Federated Domain Generalization with Domain-specific Soft Prompts Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..75ff54e36a29d6043824bb950b8e0d34685a2e39 --- /dev/null +++ b/2025/Federated Domain Generalization with Domain-specific Soft Prompts Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d495b48eaecb5119e0accfde220f5fc8c7a1ca98eb5a66ee96c5ae77c8b45a72 +size 441479 diff --git a/2025/Federated Domain Generalization with Domain-specific Soft Prompts Generation/layout.json b/2025/Federated Domain Generalization with Domain-specific Soft Prompts Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..60a85848c87a644d5b4cbc5a8dc06057ce83bc68 --- /dev/null +++ b/2025/Federated Domain Generalization with Domain-specific Soft Prompts Generation/layout.json @@ -0,0 +1,7112 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 75, + 102, + 536, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 102, + 536, + 138 + ], + "spans": [ + { + "bbox": [ + 75, + 102, + 536, + 138 + ], + "type": "text", + "content": "Federated Domain Generalization with Domain-specific Soft Prompts Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 120, + 160, + 494, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 160, + 494, + 190 + ], + "spans": [ + { + "bbox": [ + 120, + 160, + 494, + 190 + ], + "type": "text", + "content": "Jianhan Wu * Xiaoyang Qu * Zhangcheng Huang Jianzong Wang † Ping An Technology (Shenzhen) Co., Ltd." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 181, + 192, + 428, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 192, + 428, + 203 + ], + "spans": [ + { + "bbox": [ + 181, + 192, + 428, + 203 + ], + "type": "text", + "content": "wujianhan@mail.ustc.edu.cn, quxiaoy@gmail.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 179, + 205, + 429, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 205, + 429, + 217 + ], + "spans": [ + { + "bbox": [ + 179, + 205, + 429, + 217 + ], + "type": "text", + "content": "william.huang@connect.polyu.hk, jzwang@188.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 244, + 201, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 201, + 258 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 201, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 269, + 296, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 269, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 269, + 296, + 568 + ], + "type": "text", + "content": "Prompt learning has become an efficient paradigm for adapting CLIP to downstream tasks. Compared with traditional fine-tuning, prompt learning optimizes a few parameters yet yields highly competitive results, especially appealing in federated learning for computational efficiency. engendering domain shift among clients and posing a formidable challenge for downstream-task adaptation. Existing federated domain generalization (FDG) methods based on prompt learning typically learn soft prompts from training samples, replacing manually designed prompts to enhance the generalization ability of federated models. However, these learned prompts exhibit limited diversity and tend to ignore information from unknown domains. We propose a novel and effective method from a generative perspective for handling FDG tasks, namely federated domain generalization with domain-specific soft prompts generation (FedDSPG). Specifically, during training, we introduce domain-specific soft prompts (DSPs) for each domain and integrate content and domain knowledge into the generative model among clients. In the inference phase, the generator is utilized to obtain DSPs for unseen target domains, thus guiding downstream tasks in unknown domains. Comprehensive evaluations across several public datasets confirm that our method outperforms existing strong baselines in FDG, achieving state-of-the-art results." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 591, + 140, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 591, + 140, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 591, + 140, + 604 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 611, + 295, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 295, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 295, + 649 + ], + "type": "text", + "content": "Pre-trained vision-language models (VLMs) [13], such as Contrastive Language Image Pretraining (CLIP) [12, 18, 27], have emerged as a powerful technique to" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 245, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 245, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 555, + 533 + ], + "type": "text", + "content": "adapt to downstream tasks. By introducing learnable prompts [1, 9, 37] into the text or vision encoders of CLIP, this approach allows the model to dynamically adjust its representations with minimal fine-tuning. Unlike traditional fine-tuning methods [35] that require extensive parameter updates, prompt learning optimizes only a small fraction of parameters, significantly reducing computational costs while maintaining high performance. This approach has been extensively adopted in federated learning (FL) [19, 21] paradigms, particularly in scenarios characterized by substantial computational and communication overheads between server and client devices. While federated learning based on prompt learning demonstrates substantial privacy and computational benefits, a key challenge persists: data on each node is often collected in a non-IID manner in real-world scenarios, leading to domain shift [16, 30] between nodes. For example, one device may take photos mostly indoors, while another mostly outdoors [24]. The presence of the domain shift issue renders the direct application of learning with hand-crafted prompts (such as \"a photo of a [class]\") in federated learning incapable of achieving satisfactory outcomes [33]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": "To address the domain shift problem in FL, federated domain generalization (FDG) [6, 16, 17, 29] further considers that the test data (target domain) is unavailable during decentralized training from distributed source domains. Some methods have optimized soft prompts [14, 25] across different clients by treating them as learnable vectors and refining them via backpropagation. For example, Promptfl [9] shares prompt vectors instead of model parameters to enhance communication efficiency and privacy in FL. pFedPG [34] introduced a domain bank mechanism to embed textual domain information into soft prompts. Recently, ADAPT [33] and FedTPG [26] used lightweight neural networks (e.g., adapter [11]) to optimize soft prompts conditioned on images through residual or concatenation." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 654, + 139, + 664 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 654, + 139, + 664 + ], + "spans": [ + { + "bbox": [ + 67, + 654, + 139, + 664 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 664, + 295, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 664, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 664, + 295, + 712 + ], + "type": "text", + "content": "†Corresponding author is Jianzong Wang. This work was supported by the National Key Research and Development Program of China (Youth Scientist Project) under Grant No. 2024YFB4504300 and the Shenzhen-Hong Kong Joint Funding Project (Category A) under Grant No. SGDX20240115103359001." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2366" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "tion mechanisms [3], which improves generalization in a scalable manner. Despite recent advances, these methods either rely heavily on the diversity of client-specific data or use a simple adapter layer to estimate directly the relationship between soft prompts and images, leading to the learned prompts failing to fully capture the complex and diverse cross-domain feature [2]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 170, + 296, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 170, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 170, + 296, + 613 + ], + "type": "text", + "content": "To tackle the aforementioned challenges, we propose a novel and effective federated domain generalization method with domain-specific soft prompt generation (FedDSPG). Our approach reconceptualizes the prompt learning mechanism from a generative perspective, leveraging a generative model to produce domain-specific soft prompts (DSPs). Through an adversarial game manner between the generator and the discriminator, the generated DSPs exhibit greater diversity and generalization in both form and semantics. This enables our prompt generator model to acquire more generalized and robust features, facilitating adaptation to complex downstream tasks. The architectural details of our proposed approach are comprehensively depicted in Figure 1, comprising a training phase with two steps and an inference phase. Specifically, in the first step of the training phase, we avoid using hand-crafted prompts or soft prompts that lack sufficient attribute focus and specific content design. We elaborately design DSPs, composed of domain and class name vectors that map domain and content information, representing the optimal prompts for images from each client. In the second step, images and their corresponding DSPs are fed into a simple yet effective conditional generative model [4], which is trained to further optimize the DSPs by learning domain and content knowledge of images. Given the significant variance in tunable prompt embedding parameters across data-heterogeneous clients, the traditional FedAvg [19] aggregation method yields suboptimal results (see Table 5a for comparative experiments). To mitigate this, we propose using a momentum-based method to aggregate tunable parameters in the first training step. During the inference stage, the generative model's generator component will generate DSPs for target domain images, which subsequently facilitate predictions in downstream tasks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 615, + 274, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 615, + 274, + 628 + ], + "spans": [ + { + "bbox": [ + 67, + 615, + 274, + 628 + ], + "type": "text", + "content": "Our key contributions are summarized below:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 629, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 56, + 629, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 629, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 56, + 629, + 295, + 666 + ], + "type": "text", + "content": "- We are the first to optimize prompt learning from a generative perspective in FL and propose a novel soft prompt generation paradigm for FDG tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 665, + 296, + 714 + ], + "type": "text", + "content": "- Unlike the current CLIP-based methods that rely on hand-crafted or naive soft prompts, we devise domain-specific soft prompt structures that significantly enhance generalization to unseen domains." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 324, + 72, + 553, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 324, + 72, + 553, + 106 + ], + "type": "text", + "content": "Furthermore, we introduce a momentum-based aggregation mechanism to consistently refine performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 108, + 555, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 108, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 315, + 108, + 555, + 180 + ], + "type": "text", + "content": "- Comprehensive empirical evaluations on standard benchmark datasets corroborate that our proposed method achieves state-of-the-art performance. Additionally, comprehensive ablation research indicates the role played by each strategy in the increased efficacy." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 193, + 404, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 193, + 404, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 193, + 404, + 205 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 213, + 552, + 226 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 552, + 226 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 552, + 226 + ], + "type": "text", + "content": "2.1. Prompt Learning for Vision Language Models" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 232, + 555, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 555, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 555, + 542 + ], + "type": "text", + "content": "Prompt learning, also known as prompt tuning, provides an efficient alternative to conventional fine-tuning approaches for pre-trained models. By focusing on tuning only learnable prompt tokens attached to the input of models like CLIP. CoOp [40] first introduced soft prompts into vision-language models, demonstrating that appropriate prompts can enhance performance in downstream tasks. CoCoOp [39] addresses the overfitting issue in CoOp through conditioned prompt learning, specifically residual prompts. DPL [36] achieved effective domain generalization by training a lightweight prompt generator (a linear network). FUZZLE [28] integrates fuzzy C-means clustering into prompt learning to optimize cross-domain alignment between source and target distributions. In these studies, soft prompts are typically generated by simple tunable models (e.g., adapter) that predominantly rely on the diversity of client-specific data, which yields unsatisfactory results for the unseen domain. The research most pertinent to ours is SPG [2], which employs conditional adversarial generative networks to generate prompts but lacks domain-specific design, resulting in suboptimal performance in federated learning scenarios. We not only meticulously designed DSPs but also tailored the federated aggregation algorithm to better execute FDG." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 552, + 496, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 496, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 496, + 563 + ], + "type": "text", + "content": "2.2. Federated Domain Generalization" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "text", + "content": "FDG aims to address domain shifts between source and target domains in federated settings where data is distributed across clients. Traditional FDG methods generally fall into two types. The first is domain alignment approaches that focus on learning domain-invariant representations by reducing source domain discrepancies. For instance, in KD3A [7], domain differences are reduced by explicitly matching the mean embeddings of data from different domains. These methods' disadvantage lies in stringent assumptions about data distribution that are hard to satisfy in practice. The second type involves data augmentation" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2367" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 553, + 300 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 553, + 300 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 553, + 300 + ], + "type": "image", + "image_path": "e243d6ad935e909c52ba3bd94fef06d2a1fbb81fda1950caee0bbaf6fbc6aba4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 308, + 555, + 377 + ], + "lines": [ + { + "bbox": [ + 55, + 308, + 555, + 377 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 555, + 377 + ], + "type": "text", + "content": "Figure 1. The overall training structure of FedDSPG consists of two main stages: domain-specific prompts tuning and domain prompt generation. The first stage produces DSPs with compositional soft prompting technique [22], which encode the content and style information with large-scale pretrained VLMs. In the second stage, a GAN-based training method is employed to learn the latent data distribution, thereby generating soft prompts that adhere to the intrinsic data patterns. For the inference stage, the generative model's generator component will generate DSPs for target domain images, producing enhanced predictions for downstream tasks. Additionally, the momentum-based aggregation is conducted to enhance model generalization for FDG." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 390, + 297, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 390, + 297, + 583 + ], + "spans": [ + { + "bbox": [ + 54, + 390, + 297, + 583 + ], + "type": "text", + "content": "methods that aim to expand the training data distribution to enhance out-of-distribution generalization. Additional strategies include meta-learning [32], adversarial learning [38], and more. Recently, PromptFL [9], ADAPT [33], and FedTPG [26] introduced prompt learning to extract semantic information, aligning domain information on implicit textual features to improve the model's generalization capacity. To further advance this field, we conduct an in-depth exploration of this efficient paradigm. Departing from the previous prompt learning designs that rely on handcrafted prompts or simple neural network, we present a novel and effective FDG structure that leverages generative modeling techniques to learn domain-specific soft prompts, enhancing the generalization of prompts to unseen target domains." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 605, + 145, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 145, + 620 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 145, + 620 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "In this section, we first introduce the preliminary knowledge of problem setup of FDG. Following that, the proposed method FedDSPG will be introduced in detail, which includes two training steps of domain-specific soft prompts tuning and prompts generative model tuning, as well as federated parameter aggregation and federated inference." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 390, + 400, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 400, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 400, + 402 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 411, + 555, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 555, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 555, + 579 + ], + "type": "text", + "content": "In the standard FDG problem, there is a source domain dataset " + }, + { + "bbox": [ + 313, + 411, + 555, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_S = (\\mathbf{x}_i^s,\\mathbf{y}_i^s)_{i = 1}^n" + }, + { + "bbox": [ + 313, + 411, + 555, + 579 + ], + "type": "text", + "content": " with labeled data, where " + }, + { + "bbox": [ + 313, + 411, + 555, + 579 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}_i^s,\\mathbf{y}_i^s)" + }, + { + "bbox": [ + 313, + 411, + 555, + 579 + ], + "type": "text", + "content": " is the data distribution for client " + }, + { + "bbox": [ + 313, + 411, + 555, + 579 + ], + "type": "inline_equation", + "content": "i^{\\prime}s" + }, + { + "bbox": [ + 313, + 411, + 555, + 579 + ], + "type": "text", + "content": " input space and corresponding labels, with n being the number of source domains. The target domain dataset can be defined as " + }, + { + "bbox": [ + 313, + 411, + 555, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_T = (\\mathbf{x}_i^t,\\mathbf{y}_i^t)_{i = 1}^m" + }, + { + "bbox": [ + 313, + 411, + 555, + 579 + ], + "type": "text", + "content": ", where the number of target domains " + }, + { + "bbox": [ + 313, + 411, + 555, + 579 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 411, + 555, + 579 + ], + "type": "text", + "content": " is typically configured as one per client for experiment evaluation. FDG seeks to develop models that can achieve robust performance on target domains while being trained exclusively on distributed source domain data. The primary challenge lies in addressing the domain shift between source and target domains, which is exacerbated by the non-IID nature of federated data across clients." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "Taking image classification as an example, let " + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": " represent the input and label spaces, respectively. Previous domain generalization methods concentrated on training a domain-invariant model for the mapping " + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "inline_equation", + "content": "F: X \\to Y" + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": " from images to labels. Prompt learning methods embed learnable prompt vectors " + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": " into the input to leverage the generalization ability of VLMs, reformulating the mapping as " + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "inline_equation", + "content": "F: (X, V) \\to Y" + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": ". Unlike existing prompt learning methods that focus primarily on fine-tuning the prompt vectors " + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": ", we propose a novel prompt learning paradigm that utilizes a" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2368" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 72, + 293, + 210 + ], + "blocks": [ + { + "bbox": [ + 63, + 72, + 293, + 210 + ], + "lines": [ + { + "bbox": [ + 63, + 72, + 293, + 210 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 293, + 210 + ], + "type": "image", + "image_path": "29025138b121b72cdea1c30a6126a06f6fa9f01b7a82895709a109cfd55ebc60.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 219, + 296, + 288 + ], + "lines": [ + { + "bbox": [ + 55, + 219, + 296, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 219, + 296, + 288 + ], + "type": "text", + "content": "Figure 2. The structure design of the second training step is based on CcGAN. In the generator, input noise " + }, + { + "bbox": [ + 55, + 219, + 296, + 288 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 219, + 296, + 288 + ], + "type": "text", + "content": " is concatenated with fake image embeddings to form the joint latent representation, which is used to generate DSPs. The discriminator distinguishes the authenticity of the generated DSPs from the real image." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 299, + 296, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 299, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 299, + 296, + 384 + ], + "type": "text", + "content": "generative model " + }, + { + "bbox": [ + 55, + 299, + 296, + 384 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 299, + 296, + 384 + ], + "type": "text", + "content": " to generate soft prompts, reformulating the mapping as " + }, + { + "bbox": [ + 55, + 299, + 296, + 384 + ], + "type": "inline_equation", + "content": "F:(X,G(X))\\to Y" + }, + { + "bbox": [ + 55, + 299, + 296, + 384 + ], + "type": "text", + "content": ". The objective is to develop a generative framework that simultaneously captures both domain-specific and domain-invariant representations, enabling the dynamic generation of generalized prompts for unseen target domains." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 389, + 170, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 170, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 170, + 403 + ], + "type": "text", + "content": "3.2. Two Training Steps" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 407, + 296, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 296, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 296, + 539 + ], + "type": "text", + "content": "FedDSPG is implemented in two steps: domain-specific soft prompts tuning and prompts generative model tuning. Domain-specific prompts tuning aims to embed domain and class information into prompts, and the goal of prompts generative model tuning is to incorporate such information into the generative model. Notably, domain distribution is learned by the generative model, enabling the generation of instance-specific prompts for each image, guaranteeing prompt diversity and integrating domain-specific information. Below are the detailed designs of FedDSPG." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 539, + 296, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 296, + 647 + ], + "type": "text", + "content": "Domain-specific Soft Prompts Tuning. We adopt CLIP [27] as our backbone, leveraging its contrastive learning framework trained on extensive image-text pairs to establish a joint embedding space through dedicated image " + }, + { + "bbox": [ + 55, + 539, + 296, + 647 + ], + "type": "inline_equation", + "content": "(f)" + }, + { + "bbox": [ + 55, + 539, + 296, + 647 + ], + "type": "text", + "content": " and text " + }, + { + "bbox": [ + 55, + 539, + 296, + 647 + ], + "type": "inline_equation", + "content": "(g)" + }, + { + "bbox": [ + 55, + 539, + 296, + 647 + ], + "type": "text", + "content": " encoders. By describing the input text as structured prompt " + }, + { + "bbox": [ + 55, + 539, + 296, + 647 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 55, + 539, + 296, + 647 + ], + "type": "text", + "content": ": \"a photo of a [CLASS]\", the similarity of image representations in the same category " + }, + { + "bbox": [ + 55, + 539, + 296, + 647 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 539, + 296, + 647 + ], + "type": "text", + "content": " can be maximized to align semantic representations:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 70, + 654, + 295, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 654, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 70, + 654, + 295, + 684 + ], + "type": "interline_equation", + "content": "P (\\hat {y} = i | x) = \\frac {\\exp (S i m (g (t _ {i}) , f (x)) / \\tau)}{\\sum_ {k = 1} ^ {K} \\exp (S i m (g (t _ {k}) , f (x)) / \\tau)} \\tag {1}", + "image_path": "260ac636e21467f4a17ca2bed38bd986055f81857b9f73cbde3fb5f202921564.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "Sim(\\cdot)" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": " is the cosine similarity and " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": " denotes temperature parameter. Since the soft prompts" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": "of previous work (e.g., CoOp [40] and SPG [2] " + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "inline_equation", + "content": "t_i = [v]_1[v]_2 \\ldots [v]_{M_1}[\\mathrm{CLASS}]_i" + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": ") cannot deal well with the distribution shift between domains, we design domain-specific soft prompts to serve as the optimal prompt representations for images from each client. Specifically, each image from a client source domain is associated with unique DSPs (seen in Figure 1). The tunable vectors of DSPs contain two parts, a domain-invariant context and a domain-specific context, denoted as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 339, + 196, + 553, + 212 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 196, + 553, + 212 + ], + "spans": [ + { + "bbox": [ + 339, + 196, + 553, + 212 + ], + "type": "interline_equation", + "content": "t _ {i} ^ {d} = [ v ] _ {1} [ v ] _ {2} \\dots [ v ] _ {M _ {1}} [ u ] _ {1} ^ {d} [ u ] _ {2} ^ {d} \\dots [ u ] _ {M _ {2}} ^ {d} [ \\mathrm {C L S} ] _ {i} \\tag {2}", + "image_path": "72797f00784c9a7bea9b469262354bc5c5db12d704812d4081e52f767b9b781a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 214, + 555, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 214, + 555, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 214, + 555, + 299 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 313, + 214, + 555, + 299 + ], + "type": "inline_equation", + "content": "[u]_{M_2}^d" + }, + { + "bbox": [ + 313, + 214, + 555, + 299 + ], + "type": "text", + "content": " denotes domain-specific tokens, which have the same dimension as word embeddings. " + }, + { + "bbox": [ + 313, + 214, + 555, + 299 + ], + "type": "inline_equation", + "content": "M_{1}" + }, + { + "bbox": [ + 313, + 214, + 555, + 299 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 214, + 555, + 299 + ], + "type": "inline_equation", + "content": "M_2" + }, + { + "bbox": [ + 313, + 214, + 555, + 299 + ], + "type": "text", + "content": " are the number of context tokens and domain tokens incorporated in the prompt structure, respectively. These DSPs are optimized through cross-entropy loss minimization during training on domain-specific data:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 324, + 304, + 555, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 304, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 324, + 304, + 555, + 326 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {c _ {j}} ^ {d _ {i} *} = \\arg \\min _ {\\mathbf {v}} \\mathbb {E} _ {\\mathbf {x} _ {j} ^ {d _ {i}}, y _ {j} ^ {d _ {i}}} \\left[ - \\log p \\left(y _ {j} ^ {d _ {i}} \\mid \\mathbf {x} _ {j} ^ {d _ {i}}, \\mathbf {t} _ {j} ^ {d _ {i}}\\right) \\right] \\quad (3)", + "image_path": "9513ad5994ae7adb5f7e6d049cee835c630dfe1e123e784567aec6a216287a33.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 329, + 554, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 554, + 437 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 313, + 329, + 554, + 437 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}_j^{d_i},y_j^{d_i})" + }, + { + "bbox": [ + 313, + 329, + 554, + 437 + ], + "type": "text", + "content": " represents the image-label pairs from client " + }, + { + "bbox": [ + 313, + 329, + 554, + 437 + ], + "type": "inline_equation", + "content": "j^{\\prime}s" + }, + { + "bbox": [ + 313, + 329, + 554, + 437 + ], + "type": "text", + "content": " training data in " + }, + { + "bbox": [ + 313, + 329, + 554, + 437 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 329, + 554, + 437 + ], + "type": "text", + "content": "-th domain. DSPs integrate domain and class information in a more coherent and structured manner than prior approaches. These DSPs constitute optimized prompt representations for each client domain, encoding comprehensive domain characteristics, which encapsulate rich domain and class information for training a generative model in the second training step." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "type": "text", + "content": "Prompts Generative Model Tuning. To learn the domain and class knowledge in DSPs, considering the communication cost requirements (e.g., methods based on diffusion models [10] require multi-step denoising, resulting in high communication costs. The corresponding experiments can be found in the supplementary materials.) and input data continuity of federated learning, we employ an efficient generative model—Continuous Conditional Generative Adversarial Network (CcGAN) [4] to validate the effectiveness of our approach. Similar to conditional GAN [20], we manipulate the generation process by constraining the additional information (i.e., DSPs generated in the first training step) of the CcGAN. The goal is to learn the generator's distribution " + }, + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "type": "inline_equation", + "content": "p_{g}" + }, + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "type": "text", + "content": " over DSPs " + }, + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{c_j}^{d_i}" + }, + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "type": "text", + "content": ", effectively transferring generalization capabilities from prompts to the generator itself. The architecture consists of two core components: a generator " + }, + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "type": "text", + "content": " that synthesizes outputs conditioned on auxiliary information and a discriminator " + }, + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "type": "text", + "content": " that assesses the authenticity of real data and fake outputs. These components facilitate an adversarial process that enhances the model's generative performance." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2369" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 229 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 229 + ], + "type": "text", + "content": "Specifically, the detailed training process of the prompts generative model is shown in Figure 2. The generator concatenates noise variables " + }, + { + "bbox": [ + 55, + 72, + 296, + 229 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 55, + 72, + 296, + 229 + ], + "type": "text", + "content": " with fake-batch embeddings " + }, + { + "bbox": [ + 55, + 72, + 296, + 229 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 72, + 296, + 229 + ], + "type": "text", + "content": " to form joint latent codes " + }, + { + "bbox": [ + 55, + 72, + 296, + 229 + ], + "type": "inline_equation", + "content": "[\\mathbf{z}, f(\\mathbf{x})]" + }, + { + "bbox": [ + 55, + 72, + 296, + 229 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 72, + 296, + 229 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 72, + 296, + 229 + ], + "type": "text", + "content": " is randomly sampled from client data. Then the generator subsequently produces the corresponding DSPs by transforming the joint latent representation " + }, + { + "bbox": [ + 55, + 72, + 296, + 229 + ], + "type": "inline_equation", + "content": "[\\mathbf{z}, f(\\mathbf{x})]" + }, + { + "bbox": [ + 55, + 72, + 296, + 229 + ], + "type": "text", + "content": ". In the discriminator, we input both the genuine image embeddings " + }, + { + "bbox": [ + 55, + 72, + 296, + 229 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 72, + 296, + 229 + ], + "type": "text", + "content": " from real batches and the generated DSPs with their associated image embeddings " + }, + { + "bbox": [ + 55, + 72, + 296, + 229 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 72, + 296, + 229 + ], + "type": "text", + "content": " to distinguish real from fake. Therefore, the two-player min-max game objective function " + }, + { + "bbox": [ + 55, + 72, + 296, + 229 + ], + "type": "inline_equation", + "content": "V(G, D)" + }, + { + "bbox": [ + 55, + 72, + 296, + 229 + ], + "type": "text", + "content": " for CcGAN can be formally expressed as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 244, + 298, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 244, + 298, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 244, + 298, + 289 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {G} \\max _ {D} V (G, D) = \\mathbb {E} _ {\\mathbf {v} \\sim p _ {\\mathrm {v}} (\\mathbf {v})} [ \\log D (\\mathbf {v} \\mid f (\\mathbf {x})) ] \\\\ + \\mathbb {E} _ {\\mathbf {z} \\sim p _ {z} (\\mathbf {z})} [ \\log (1 - D (G (\\mathbf {z} \\mid f (\\mathbf {x}))) ]) \\tag {4} \\\\ \\end{array}", + "image_path": "6c000ef6995410eb917e51e7f1c0039c681ef84f95a920492e8f485f26766ae3.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 289, + 296, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 296, + 373 + ], + "type": "text", + "content": "The process of prompts generative model tuning is to learn coarse-to-fine DSPs through an adversarial manner. This design not only promotes diversity in prompt generation but also significantly improves generalization across tasks and domains, facilitating robust performance in both seen and unseen conditions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 380, + 243, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 243, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 243, + 392 + ], + "type": "text", + "content": "3.3. Federated Parameters Aggregation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 396, + 296, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 296, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 296, + 661 + ], + "type": "text", + "content": "We employ distinct aggregation strategies for the learnable parameters of prompt tuning and the generative model. Assuming a centralized parameter server for uniform aggregation—though our method also supports decentralized communication—clients upload their respective soft text embedding parameters and generative model parameters to the server during each communication round. The server averages the parameters from all n clients and redistributes them. While this parameter aggregation paradigm appears viable for our method, it introduces a minor issue for soft text embeddings. Specifically, given that FDG addresses domain shifts across clients, the domain components of DSPs vary significantly between clients. We observe that direct averaging causes abrupt changes in soft text embedding parameters, negatively impacting model performance. To mitigate this, we introduce momentum updates for soft prompt parameters, utilizing exponential moving averages to maintain parameter continuity and prevent destabilizing abrupt changes during the optimization process. The aggregation formula is as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 667, + 295, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 667, + 295, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 295, + 683 + ], + "type": "interline_equation", + "content": "[ w ] ^ {k} = \\alpha [ w ] ^ {k - 1} + (1 - \\alpha) [ w ] ^ {k - 2} \\tag {5}", + "image_path": "9f0354b66cf42371065a0dd6252bf4d4b734c4f5d77a532654ba358aa06b61c4.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "[w]^k" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "[w]^{k-1}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "[w]^{k-2}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " denote the tunable soft prompt embedding parameters at the " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "k, k-1" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 121 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 121 + ], + "type": "inline_equation", + "content": "k - 2" + }, + { + "bbox": [ + 313, + 72, + 555, + 121 + ], + "type": "text", + "content": " step, and " + }, + { + "bbox": [ + 313, + 72, + 555, + 121 + ], + "type": "inline_equation", + "content": "\\alpha \\in [0,1]" + }, + { + "bbox": [ + 313, + 72, + 555, + 121 + ], + "type": "text", + "content": " serves as a hyperparameter governing the smoothing intensity of DSPs. Comprehensive ablation studies examining the impact of momentum updates are presented in table 5a." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 126, + 457, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 126, + 457, + 138 + ], + "spans": [ + { + "bbox": [ + 313, + 126, + 457, + 138 + ], + "type": "text", + "content": "3.4. Inference with Generator" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 144, + 555, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 555, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 555, + 192 + ], + "type": "text", + "content": "The generator trained via CcGAN is utilized to generate DSPs for inference on each target image. The predicted probability that a target domain image " + }, + { + "bbox": [ + 313, + 144, + 555, + 192 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 144, + 555, + 192 + ], + "type": "text", + "content": " belongs to the " + }, + { + "bbox": [ + 313, + 144, + 555, + 192 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 144, + 555, + 192 + ], + "type": "text", + "content": "-th class is formally expressed as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 346, + 200, + 555, + 230 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 200, + 555, + 230 + ], + "spans": [ + { + "bbox": [ + 346, + 200, + 555, + 230 + ], + "type": "interline_equation", + "content": "p (y = i \\mid \\mathbf {x}) = \\frac {\\exp (\\langle w _ {i} , f (\\mathbf {x}) \\rangle / \\tau)}{\\sum_ {j = 1} ^ {K} \\exp (\\langle w _ {j} , f (\\mathbf {x}) \\rangle / \\tau)}, \\tag {6}", + "image_path": "f237af93d84f283b20e6572f2522265bdc0e6c95045b1cc85b9d1243c7a1eb4a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 364, + 236, + 553, + 250 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 236, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 364, + 236, + 553, + 250 + ], + "type": "interline_equation", + "content": "s. t. \\quad w _ {j} = g ([ G (\\mathbf {z} \\mid f (\\mathbf {x})), c _ {j} ]) \\tag {7}", + "image_path": "29d2b85a60d070d846816a8819aa87fa88ddebc98b2c3e6f08a4d32bbbbfb1e9.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 253, + 555, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 555, + 397 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 313, + 253, + 555, + 397 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 253, + 555, + 397 + ], + "type": "text", + "content": " represents the temperature parameter that modulates the sharpness of the output distribution, " + }, + { + "bbox": [ + 313, + 253, + 555, + 397 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 253, + 555, + 397 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 253, + 555, + 397 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 253, + 555, + 397 + ], + "type": "text", + "content": " represent the text and the image encoder respectively, and " + }, + { + "bbox": [ + 313, + 253, + 555, + 397 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 253, + 555, + 397 + ], + "type": "text", + "content": " denotes the total number of classes. The term " + }, + { + "bbox": [ + 313, + 253, + 555, + 397 + ], + "type": "inline_equation", + "content": "G(\\mathbf{z} \\mid f(\\mathbf{x}))" + }, + { + "bbox": [ + 313, + 253, + 555, + 397 + ], + "type": "text", + "content": " signifies the generated DSPs by the generator. We use " + }, + { + "bbox": [ + 313, + 253, + 555, + 397 + ], + "type": "inline_equation", + "content": "c_{j}" + }, + { + "bbox": [ + 313, + 253, + 555, + 397 + ], + "type": "text", + "content": " to represent the " + }, + { + "bbox": [ + 313, + 253, + 555, + 397 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 253, + 555, + 397 + ], + "type": "text", + "content": "-th tokenized class label, with " + }, + { + "bbox": [ + 313, + 253, + 555, + 397 + ], + "type": "inline_equation", + "content": "[\\cdot, \\cdot]" + }, + { + "bbox": [ + 313, + 253, + 555, + 397 + ], + "type": "text", + "content": " denoting the standard concatenation operation. In the generative paradigm, our approach achieves enhanced adaptability and dynamic prompts generation, which are essential for accommodating diverse input distributions and downstream task requirements." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 406, + 440, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 440, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 440, + 421 + ], + "type": "text", + "content": "4. Experimental Setup" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 426, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 617 + ], + "type": "text", + "content": "Datasets. We conducted extensive experiments for image classification on domain generalization benchmark datasets: (1) Office-Home, a large-scale visual cross-domain dataset with 65 classes across four visually distinct domains, i.e., Art (Ar), Clipart (Cl), Product (Pr), and Real World (Re) [31]. (2) DomainNet, a largescale domain generalization benchmark dataset with 345 classes distributed among six domains, i.e., Clipart (Clp), Infograph (Inf), Painting (Pnt), Quickdraw (Qdr), Real (Rel), and Sketch (Skt) [23]. (3) PACS, featuring approximately 10,000 images from seven categories across four domains, including Photo (P), Sketch (S), Cartoon (C), and Art (A) [15]. For rigorous evaluation, we adopt the leave-one-domain-out protocol, designating each domain sequentially as the target while utilizing the remaining domains as source data." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "content": "Implementation details. In FDG, all participating nodes maintain identical architectural configurations while ensuring complete domain data isolation between nodes. The data processing and division are followed [8] on the image generalization task. For our FedDSPG method, we employ a frozen CLIP backbone (ViT-B/16 [5]) pre-trained on large-scale vision-language data. The text and visual prompt tokens" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2370" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 555, + 232 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 555, + 232 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 555, + 232 + ], + "type": "table", + "html": "
DatasetOffice-HomeDomainNetPACS
Target DomainRePrClArAvgClpInfPntQdrRelSktAvgPACSAvg
CoOp[40]82.982.859.372.174.361.622.252.612.167.748.644.194.191.595.684.991.5
CoCoOp[39]85.081.162.568.674.372.423.360.816.472.760.551.097.594.697.987.594.4
FedVPT-D[14]85.085.261.176.076.863.343.074.854.887.267.165.099.296.598.187.295.3
FedCLIP[18]87.787.563.678.079.265.146.575.956.190.267.967.099.896.397.985.694.9
PromptFL[9]89.688.271.679.882.372.850.384.260.492.167.771.399.997.199.090.696.7
ADAPT[33]90.390.568.282.682.977.563.170.541.685.772.168.499.998.099.191.797.2
FedTPG[26]89.089.771.184.783.674.052.185.162.194.068.672.799.897.899.290.896.9
FedDSPG90.591.073.885.885.374.763.585.563.094.272.675.699.998.199.392.297.4
", + "image_path": "9e2905675cd4e0f88c941c30bd3c3939f0d21a7f1d088cf3075b3b8ca27de114.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 238, + 555, + 273 + ], + "lines": [ + { + "bbox": [ + 55, + 238, + 555, + 273 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 555, + 273 + ], + "type": "text", + "content": "Table 1. Accuracies (%) for the image classification task on Office-Home, DomainNet, and PACS datasets using CLIP with ViT-B/16 image encoder as frozen backbone. The domains displayed here are the target domains, and the others are the source domains; the best result is indicated in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 287, + 294, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 287, + 294, + 455 + ], + "spans": [ + { + "bbox": [ + 54, + 287, + 294, + 455 + ], + "type": "text", + "content": "share a unified dimensionality of 512. For the learning strategies, in the first training step, the DSPs are trained by using the Adam optimizer with the learning rate of " + }, + { + "bbox": [ + 54, + 287, + 294, + 455 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-5}" + }, + { + "bbox": [ + 54, + 287, + 294, + 455 + ], + "type": "text", + "content": ", batch size of 32, and the context lengths " + }, + { + "bbox": [ + 54, + 287, + 294, + 455 + ], + "type": "inline_equation", + "content": "M_{1}" + }, + { + "bbox": [ + 54, + 287, + 294, + 455 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 287, + 294, + 455 + ], + "type": "inline_equation", + "content": "M_{2}" + }, + { + "bbox": [ + 54, + 287, + 294, + 455 + ], + "type": "text", + "content": " of the prompt number are 4. In the second training step, the CcGAN is trained by using the AdamW optimizer with weight decay " + }, + { + "bbox": [ + 54, + 287, + 294, + 455 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-5}" + }, + { + "bbox": [ + 54, + 287, + 294, + 455 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 287, + 294, + 455 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 54, + 287, + 294, + 455 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 287, + 294, + 455 + ], + "type": "inline_equation", + "content": "\\beta = 0.999" + }, + { + "bbox": [ + 54, + 287, + 294, + 455 + ], + "type": "text", + "content": ", starting with a learning rate of " + }, + { + "bbox": [ + 54, + 287, + 294, + 455 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 54, + 287, + 294, + 455 + ], + "type": "text", + "content": " for the OfficeHome and DomainNet datasets. The 100 epochs are trained for both our method and the baseline models for accelerating convergence, then the FedAvg and momentum-based aggregation algorithms are executed after every epoch. 8 Tesla-V100 with 16GB memory are used to experiment." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 456, + 295, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 295, + 612 + ], + "type": "text", + "content": "Baseline models. We compared three distinct SOTA prompt learning approaches: (1) methods focusing on designing trainable models with small-scale parameters on the client side, such as CoOp [40], CoCoOp [39], FedVPT-D[14], FedCLIP [18], and FedTPG [26]; (2) methods that design prompt generators on the server side to reduce communication costs, such as PromptFL [9]; and (3) methods that emphasize both semantic and domain-specific prompts, such as ADAPT [33]. To show fair competition, the results of subsequent experiments were either the result of their papers or reproduced from source code provided by their authors." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 625, + 190, + 640 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 625, + 190, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 625, + 190, + 640 + ], + "type": "text", + "content": "5. Experiments Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 647, + 295, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 295, + 671 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 295, + 671 + ], + "type": "text", + "content": "5.1. Federated Domain Generalization Performance" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "FedDSPG achieved effective FDG in scenarios where target domain distributions diverge significantly from source domains. Quantitative results across three" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 287, + 555, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 287, + 555, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 287, + 555, + 563 + ], + "type": "text", + "content": "benchmark datasets are presented in Table 1 following the \"leave-one-out\" protocol. In general, our method outperforms all baseline methods to varying degrees. We gleaned two primary observations from the results. Firstly, our approach outperforms all baselines, achieving SOTA performance, which demonstrates the effectiveness of our proposed approach in addressing federated domain generalization challenges. This establishes FedDSPG as a highly referencable benchmark for future research on FDG. Secondly, existing methods, including methods focusing on designing small-scale parameters models (FedCLIP, FedTPG) and methods generating prompts on the server side (PromptFL), exhibit notable performance gaps. We suppose that the possible reason stems from two primary factors: (1) inadequate incorporation of domain-specific information in their prompt design strategies, and (2) insufficient adaptability of generated prompts to unseen target domain images. These potential risks are mitigated in our approach, as it designs domain-specific soft prompts structure and trains a specific generative model, enhancing the generalization and diversity of prompts to unseen domains." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 574, + 446, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 446, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 446, + 586 + ], + "type": "text", + "content": "5.2. Cross-Dataset Transfer" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 593, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 593, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 593, + 556, + 715 + ], + "type": "text", + "content": "Having demonstrated FedDSPG's capability for domain generalization within a single dataset, we further conducted experiments on cross-dataset domain generalization. This represents a more challenging task, as the fundamental characteristics can exhibit significant variations across different datasets. Specifically, we adopt a rigorous experimental protocol where DomainNet serves as the source domain, while OfficeHome and PACS function as target domains. We compared FedDSPG with FedTPG (best-performing" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2371" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 76, + 221, + 204 + ], + "blocks": [ + { + "bbox": [ + 63, + 76, + 221, + 204 + ], + "lines": [ + { + "bbox": [ + 63, + 76, + 221, + 204 + ], + "spans": [ + { + "bbox": [ + 63, + 76, + 221, + 204 + ], + "type": "image", + "image_path": "76026bb88c01c1995efe960033b888a0e278636babc23995bd4c688d1a2f9bef.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 225, + 76, + 383, + 205 + ], + "blocks": [ + { + "bbox": [ + 225, + 76, + 383, + 205 + ], + "lines": [ + { + "bbox": [ + 225, + 76, + 383, + 205 + ], + "spans": [ + { + "bbox": [ + 225, + 76, + 383, + 205 + ], + "type": "image", + "image_path": "1fb3f94cf79013e94ac01dc58f3d70cbdda093f19619477efd5fd83feb6a8c3e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 388, + 76, + 547, + 204 + ], + "blocks": [ + { + "bbox": [ + 388, + 76, + 547, + 204 + ], + "lines": [ + { + "bbox": [ + 388, + 76, + 547, + 204 + ], + "spans": [ + { + "bbox": [ + 388, + 76, + 547, + 204 + ], + "type": "image", + "image_path": "beab1ee4bb344961cf6c66e3965c1da6048c8c3a004dadfb3824bb0f631fd6eb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 229, + 555, + 302 + ], + "blocks": [ + { + "bbox": [ + 168, + 218, + 440, + 229 + ], + "lines": [ + { + "bbox": [ + 168, + 218, + 440, + 229 + ], + "spans": [ + { + "bbox": [ + 168, + 218, + 440, + 229 + ], + "type": "text", + "content": "Figure 3. The accuracy of FedDSPG with different data distribution." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 229, + 555, + 302 + ], + "lines": [ + { + "bbox": [ + 58, + 229, + 555, + 302 + ], + "spans": [ + { + "bbox": [ + 58, + 229, + 555, + 302 + ], + "type": "table", + "html": "
SourceTarget
DomainNetRePrClArAvgPACSAvg
FedTPG72.781.378.566.176.475.691.293.891.686.190.7
Ours75.686.082.570.480.479.894.595.393.588.693.0
△+3.9+4.7+4.0+4.3+4.0+4.2+3.3+1.5+1.9+2.5+2.3
", + "image_path": "1e84a11f912ef61a471d1bffb39df562fa88d147f46ed92623c73a1b30d9999c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 308, + 555, + 331 + ], + "lines": [ + { + "bbox": [ + 55, + 308, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 555, + 331 + ], + "type": "text", + "content": "Table 2. The results for cross-dataset transfer experiment. The model is trained on the DomainNet dataset and inference on the Office-Home and PACS dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 343, + 294, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 343, + 294, + 440 + ], + "spans": [ + { + "bbox": [ + 54, + 343, + 294, + 440 + ], + "type": "text", + "content": "baseline), with the primary results detailed in Table 2. Overall, both methods experienced a decline in performance when confronted with the gap between datasets. However, FedDSPG consistently outperformed FedTPG on the target dataset by a substantial margin. This further underscores our method's remarkable capability to enhance model generalization across diverse domains." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 454, + 203, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 454, + 203, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 203, + 467 + ], + "type": "text", + "content": "5.3. Data Distribution Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "text", + "content": "Considering client data distributions are inherently non-IID in federated learning scenarios, we conducted a comprehensive empirical analysis examining how varying data distributions affect FedDSPG's performance. Specifically, we varied the overlap ratio " + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "text", + "content": " of data domains across clients, setting it to " + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "inline_equation", + "content": "0\\%" + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "text", + "content": ". Specifically, an overlap ratio of " + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "inline_equation", + "content": "r\\%" + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "text", + "content": " implies that " + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "inline_equation", + "content": "r\\%" + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "text", + "content": " of the domains are shared among multiple clients, while the remaining " + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "inline_equation", + "content": "1 - r\\%" + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "text", + "content": " of the domains are unique to individual clients. Figure 3 shows F1/accuracy across different domain distribution settings. The model exhibits greater stability on the Office-Home dataset compared to the other evaluation datasets, and a slight performance improvement is observed only when the overlap ratio reaches " + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "text", + "content": ". We attribute this enhanced stability to the higher number of domains, which likely endows the model with greater robustness against distribution shifts. This finding underscores the importance of domain diversity in enhancing the generalization and stability of" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 344, + 435, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 344, + 435, + 357 + ], + "spans": [ + { + "bbox": [ + 314, + 344, + 435, + 357 + ], + "type": "text", + "content": "federated learning models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 369, + 408, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 408, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 408, + 382 + ], + "type": "text", + "content": "5.4. Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 388, + 555, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 388, + 555, + 509 + ], + "spans": [ + { + "bbox": [ + 312, + 388, + 555, + 509 + ], + "type": "text", + "content": "Ablation on the different prompts structure. We classify prompts into 4 types (fixed, learnable, domain-specific, generative), and three representative SOTA models from baseline methods for comparison: PromptFL [9], and ADAPT [33], FedTPG [26]. PromptFL represents methods that design a generator on the server side to produce prompts. ADAPT represents methods emphasizing domain-specific information. FedTPG represents methods focusing on designing single trainable networks to optimize prompts." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 510, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 510, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 510, + 556, + 715 + ], + "type": "text", + "content": "To validate the effectiveness of our approach and the improvements, several ablation models with different prompt structures were trained. These models include: Our model with hand-crafted domain prompts (Ours-HDP), which replaces DSPs with structured prompts (i.e., \"a photo of a [class]\"). Our model with conditional soft prompts (Ours-CSP), which replaces DSPs with conditional soft prompts (i.e., " + }, + { + "bbox": [ + 312, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "t_i = [v]_1[v]_2 \\ldots [v]_{M_1}" + }, + { + "bbox": [ + 312, + 510, + 556, + 715 + ], + "type": "text", + "content": "). Our model without a generative model (Ours-WGM), which omits the second training step. The results are presented in Table 3. Ours-HDP performs the worst, proving learnable prompt designing are crucial for performance. Although both ADAPT and Ours-WGM have domain-specific prompt, the improvement in results is not significant. However, Ours-CSP outperforms Ours-WGM and other baseline methods, underscoring the crucial" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2372" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 550, + 182 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 550, + 182 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 550, + 182 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 550, + 182 + ], + "type": "table", + "html": "
MethodPrompt TypeRePrClArAvg
FixedLearnableDomain-specificGenerative
PromptFL [9]✓✓89.688.271.679.882.3
ADAPT [33]✓✓90.390.568.282.682.9
FedTPG [26]✓89.089.771.184.783.6
Ours-HDP✓✓86.181.470.273.277.7
Ours-CSP✓✓90.990.873.184.184.6
Ours-WGM✓✓88.187.871.682.482.5
Ours✓✓✓90.591.073.885.885.3
", + "image_path": "eadc8b6cd3e598cd40dd58f544c9ca823a69ef96d17d6c07057378398664a4d8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 66, + 204, + 544, + 241 + ], + "blocks": [ + { + "bbox": [ + 96, + 190, + 512, + 201 + ], + "lines": [ + { + "bbox": [ + 96, + 190, + 512, + 201 + ], + "spans": [ + { + "bbox": [ + 96, + 190, + 512, + 201 + ], + "type": "text", + "content": "Table 3. Comparisons with different design of prompt on the Office-Home dataset for FDG performance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 204, + 544, + 241 + ], + "lines": [ + { + "bbox": [ + 66, + 204, + 544, + 241 + ], + "spans": [ + { + "bbox": [ + 66, + 204, + 544, + 241 + ], + "type": "table", + "html": "
α0.00.10.20.30.40.50.60.70.80.91.0
Accuracy (%)84.1584.8985.3085.2884.5684.3284.6784.0183.9583.5083.39
", + "image_path": "c6e80a16a26676002c9a45e7d5f18c338f812bcaed5d565d1d407f4d7149e50f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 61, + 279, + 181, + 350 + ], + "blocks": [ + { + "bbox": [ + 187, + 249, + 421, + 260 + ], + "lines": [ + { + "bbox": [ + 187, + 249, + 421, + 260 + ], + "spans": [ + { + "bbox": [ + 187, + 249, + 421, + 260 + ], + "type": "text", + "content": "Table 4. Accuracy of FedDSPG with different coefficient " + }, + { + "bbox": [ + 187, + 249, + 421, + 260 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 187, + 249, + 421, + 260 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 279, + 181, + 350 + ], + "lines": [ + { + "bbox": [ + 61, + 279, + 181, + 350 + ], + "spans": [ + { + "bbox": [ + 61, + 279, + 181, + 350 + ], + "type": "table", + "html": "
Modeacc.
w/momentum85.3
w/o momentum84.5
", + "image_path": "553b9ca2b4b91b05215c9e8e874651b01dc72d4fbe67d0d832d73174065ba9bb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 188, + 279, + 293, + 350 + ], + "blocks": [ + { + "bbox": [ + 85, + 352, + 154, + 362 + ], + "lines": [ + { + "bbox": [ + 85, + 352, + 154, + 362 + ], + "spans": [ + { + "bbox": [ + 85, + 352, + 154, + 362 + ], + "type": "text", + "content": "(a) Prompt update." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 188, + 279, + 293, + 350 + ], + "lines": [ + { + "bbox": [ + 188, + 279, + 293, + 350 + ], + "spans": [ + { + "bbox": [ + 188, + 279, + 293, + 350 + ], + "type": "table", + "html": "
epochs/roundacc.
0.585.3
185.3
284.6
", + "image_path": "574542219672456c0b11731c119b1b6d5b0d2c2ef4835960a9c2bf63932d9e1f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 200, + 352, + 278, + 362 + ], + "lines": [ + { + "bbox": [ + 200, + 352, + 278, + 362 + ], + "spans": [ + { + "bbox": [ + 200, + 352, + 278, + 362 + ], + "type": "text", + "content": "(b) Comm. frequency" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 371, + 295, + 394 + ], + "lines": [ + { + "bbox": [ + 55, + 371, + 295, + 394 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 295, + 394 + ], + "type": "text", + "content": "Table 5. Ablation studies on different momentum update and communication frequency." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 401, + 295, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 295, + 449 + ], + "type": "text", + "content": "role of the generative model in enhancing generalizability. Ultimately, our method achieves superior performance compared to all baselines, demonstrating the effective integration of DSPs and the generative model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 449, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 676 + ], + "type": "text", + "content": "Momentum update and communication frequency. As elaborated in Section 3.3, we employ a momentum-based approach to mitigate the performance degradation caused by large gradient variations. We conduct comparative experiments on parameters with and without momentum updates. As presented in Table 5a, the results reveal a 0.8 percentage point decrease in accuracy without momentum. The experimental results for the momentum hyperparameter " + }, + { + "bbox": [ + 55, + 449, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 449, + 295, + 676 + ], + "type": "text", + "content": " ranging from 0 to 1 are shown in Table 4, with " + }, + { + "bbox": [ + 55, + 449, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\alpha = 0.2" + }, + { + "bbox": [ + 55, + 449, + 295, + 676 + ], + "type": "text", + "content": " yielding the best performance. Additionally, we also test communication frequency, varying it to 0.5, 1, and 2 training epochs per communication round. Results (Table 5b) showed no performance gain from more frequent aggregation compared to the default 1 epoch per round, while lower frequencies led to accuracy declines. This suggests that FedDSPG can achieve good performance by maintaining the default update frequency." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "content": "Number of clients and shots. We conducted ablation studies on the number of clients and shots, the results are displayed in Figure 4. Regarding the number" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 319, + 284, + 432, + 395 + ], + "blocks": [ + { + "bbox": [ + 319, + 284, + 432, + 395 + ], + "lines": [ + { + "bbox": [ + 319, + 284, + 432, + 395 + ], + "spans": [ + { + "bbox": [ + 319, + 284, + 432, + 395 + ], + "type": "image", + "image_path": "aa9ff88485566362e5cdd3b1f15debdec9984a37a1d4a7b574ff377f923e1d58.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 408, + 554, + 430 + ], + "lines": [ + { + "bbox": [ + 313, + 408, + 554, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 554, + 430 + ], + "type": "text", + "content": "Figure 4. Ablation studies on different number of clients and shots." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 433, + 284, + 549, + 395 + ], + "blocks": [ + { + "bbox": [ + 433, + 284, + 549, + 395 + ], + "lines": [ + { + "bbox": [ + 433, + 284, + 549, + 395 + ], + "spans": [ + { + "bbox": [ + 433, + 284, + 549, + 395 + ], + "type": "image", + "image_path": "f9614fc4e1b8a38f5b4fd168c68945884d4abbb8f4d8c561c02cb32b5f7c79ed.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 440, + 554, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 440, + 554, + 512 + ], + "spans": [ + { + "bbox": [ + 313, + 440, + 554, + 512 + ], + "type": "text", + "content": "of clients, although an inevitable performance drop occurs for all methods as the number of clients increases, FedDSPG exhibits only a marginal decline (approximately " + }, + { + "bbox": [ + 313, + 440, + 554, + 512 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 440, + 554, + 512 + ], + "type": "text", + "content": "). For the number of shots, FedDSPG outperforms baseline methods when the number of shots exceeds one." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 523, + 392, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 523, + 392, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 523, + 392, + 536 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 543, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 543, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 312, + 543, + 555, + 712 + ], + "type": "text", + "content": "This work introduces FedDSPG, a novel federated domain generalization approach that leverages domain-specific soft prompts (DSPs) to enhance model generalization in non-IID federated learning scenarios. By integrating domain and content knowledge into a generative model, FedDSPG generates diverse prompts for unseen target domains, improving performance in unknown domains. Through extensive evaluations, we demonstrate our method's superior performance against existing baselines, while also revealing the untapped potential of generative techniques in FDG scenarios. These findings provide both immediate practical value and long-term research directions for advancing FL in real-world applications." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2373" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 118, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 118, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 118, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 715 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 157 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 157 + ], + "type": "text", + "content": "[1] Sikai Bai, Jiewei Zhang, Shuaicheng Li, Song Guo, Jingcai Guo, Jun Hou, Tao Han, and Xiaocheng Lu. Diprompt: Disentangled prompt tuning for multiple latent domain generalization in federated learning. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 158, + 296, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 158, + 296, + 203 + ], + "spans": [ + { + "bbox": [ + 61, + 158, + 296, + 203 + ], + "type": "text", + "content": "[2] Shuanghao Bai, Yuedi Zhang, Wanqi Zhou, Zhirong Luan, and Badong Chen. Soft prompt generation for domain generalization. In European Conference on Computer Vision, pages 434-450. Springer, 2025. 2, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 203, + 295, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 203, + 295, + 246 + ], + "spans": [ + { + "bbox": [ + 62, + 203, + 295, + 246 + ], + "type": "text", + "content": "[3] Tianyu Cui, Hongxia Li, Jingya Wang, and Ye Shi. Harmonizing generalization and personalization in federated prompt learning. International Conference on Machine Learning, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 247, + 295, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 295, + 302 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 295, + 302 + ], + "type": "text", + "content": "[4] Xin Ding, Yongwei Wang, Zuheng Xu, William J Welch, and Z Jane Wang. Ccgan: Continuous conditional generative adversarial networks for image generation. In International Conference on Learning Representations, 2021. 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 303, + 295, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 303, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 62, + 303, + 295, + 357 + ], + "type": "text", + "content": "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations, ICLR, 2021. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 357, + 295, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 357, + 295, + 391 + ], + "spans": [ + { + "bbox": [ + 62, + 357, + 295, + 391 + ], + "type": "text", + "content": "[6] Yuqi Fang, Pew-Thian Yap, Weili Lin, Hongtu Zhu, and Mingxia Liu. Source-free unsupervised domain adaptation: A survey. Neural Networks, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 392, + 295, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 295, + 435 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 295, + 435 + ], + "type": "text", + "content": "[7] Haozhe Feng, Zhaoyang You, Minghao Chen, Tianye Zhang, Minfeng Zhu, et al. Kd3a: Unsupervised multisource decentralized domain adaptation via knowledge distillation. In ICML, pages 3274-3283, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 436, + 295, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 295, + 480 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 295, + 480 + ], + "type": "text", + "content": "[8] Chunjiang Ge, Rui Huang, Mixue Xie, Zihang Lai, Shiji Song, Shuang Li, and Gao Huang. Domain adaptation via prompt learning. IEEE Transactions on Neural Networks and Learning Systems, 2023. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 481, + 295, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 481, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 62, + 481, + 295, + 536 + ], + "type": "text", + "content": "[9] Tao Guo, Song Guo, Junxiao Wang, Xueyang Tang, and Wenchao Xu. Promptfl: Let federated participants cooperatively learn prompts instead of models-federated learning in age of foundation model. IEEE Transactions on Mobile Computing, 2023. 1, 3, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 536, + 295, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 536, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 57, + 536, + 295, + 568 + ], + "type": "text", + "content": "[10] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 569, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 569, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 569, + 295, + 624 + ], + "type": "text", + "content": "[11] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. 9th International Conference on Learning Representations, ICLR, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 625, + 295, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 295, + 691 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 295, + 691 + ], + "type": "text", + "content": "[12] Xuefeng Hu, Ke Zhang, Lu Xia, Albert Chen, Jiajia Luo, Yuyin Sun, Ken Wang, Nan Qiao, Xiao Zeng, Min Sun, et al. Reclip: Refine contrastive language image pretraining with source free domain adaptation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2994-3003, 2024. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 295, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 715 + ], + "type": "text", + "content": "[13] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 715 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 118 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 118 + ], + "type": "text", + "content": "Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "type": "text", + "content": "[14] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 1, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 162, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 553, + 205 + ], + "type": "text", + "content": "[15] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. In Proceedings of the IEEE International Conference on Computer Vision, pages 5542-5550, 2017. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 206, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 259 + ], + "type": "text", + "content": "[16] Xinhui Liu, Wei Xi, Wen Li, Dong Xu, Gairui Bai, and Jizhong Zhao. Co-mda: Federated multi-source domain adaptation on black-box models. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 261, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 316 + ], + "type": "text", + "content": "[17] Xinhui Liu, Zhenghao Chen, Luping Zhou, Dong Xu, Wei Xi, Gairui Bai, Yihan Zhao, and Jizhong Zhao. Ufda: Universal federated domain adaptation with practical assumptions. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 14026-14034, 2024. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 316, + 553, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 361 + ], + "type": "text", + "content": "[18] Wang Lu, Xixu Hu, Jindong Wang, and Xing Xie. Fedclip: Fast generalization and personalization for clip in federated learning. In ICLR Workshop on Distributed and Private Machine Learning, 2023. 1, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 361, + 553, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 405 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 405 + ], + "type": "text", + "content": "[19] Brendan McMahan, Eider Moore, et al. Communication-efficient learning of deep networks from decentralized data. In Artificial Intelligence and Statistics, pages 1273-1282. PMLR, 2017. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 405, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 405, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 405, + 553, + 437 + ], + "type": "text", + "content": "[20] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. Advances in Neural Information Processing Systems, 2014. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 437, + 553, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 553, + 481 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 553, + 481 + ], + "type": "text", + "content": "[21] Mehryar Mohri, Gary Sivek, and Ananda Theertha Suresh. Agnostic federated learning. In International Conference on Machine Learning, pages 4615-4625. PMLR, 2019. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 481, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 481, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 481, + 553, + 525 + ], + "type": "text", + "content": "[22] Nihal V Nayak, Peilin Yu, and Stephen H Bach. Learning to compose soft prompts for compositional zero-shot learning. International Conference on Learning Representations, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 525, + 553, + 582 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 525, + 553, + 582 + ], + "spans": [ + { + "bbox": [ + 317, + 525, + 553, + 582 + ], + "type": "text", + "content": "[23] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1406-1415, 2019. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "type": "text", + "content": "[24] Xingchao Peng, Zijun Huang, Yizhe Zhu, and Kate Saenko. Federated adversarial domain adaptation. International Conference on Learning Representations, ICLR, 2019. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 625, + 553, + 681 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 681 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 681 + ], + "type": "text", + "content": "[25] Chen Qiu, Xingyu Li, Chaithanya Kumar Mummadi, Madan Ravi Ganesh, Zhenzhen Li, Lu Peng, and Wan-Yi Lin. Text-driven prompt generation for vision-language models in federated learning. International Conference on Learning Representations, 2024. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 681, + 553, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 715 + ], + "type": "text", + "content": "[26] Chen Qiu, Xingyu Li, Chaithanya Kumar Mummadi, Madan Ravi Ganesh, Zhenzhen Li, Lu Peng, and Wan-Yi Lin. Federated text-driven prompt generation for" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2374" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "vision-language models. In The Twelfth International Conference on Learning Representations, 2024. 1, 3, 6, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "type": "text", + "content": "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "text", + "content": "[28] Kuo Shi, Jie Lu, Zhen Fang, and Guangquan Zhang. Unsupervised domain adaptation enhanced by fuzzy prompt learning. IEEE Transactions on Fuzzy Systems, 32:4038-4048, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "text", + "content": "[29] Lei Song, Chunguang Ma, Guoyin Zhang, and Yun Zhang. Privacy-preserving unsupervised domain adaptation in federated setting. IEEE Access, 8:143233-143240, 2020. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 308 + ], + "type": "text", + "content": "[30] Yue Tan, Chen Chen, Weiming Zhuang, Xin Dong, Lingjuan Lyu, and Guodong Long. Is heterogeneity notorious? taming heterogeneity to handle test-time shift in federated learning. Advances in Neural Information Processing Systems, 36, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 310, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 294, + 365 + ], + "type": "text", + "content": "[31] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5018-5027, 2017. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 366, + 294, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 366, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 366, + 294, + 421 + ], + "type": "text", + "content": "[32] Anna Vettoruzzo, Mohamed-Rafik Bouguelia, Joaquin Vanschoren, Thorsteinn Rognvaldsson, and KC Santosh. Advances and challenges in meta-learning: A technical review. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 4763-4779, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 294, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 455 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 455 + ], + "type": "text", + "content": "[33] Guoyizhe Wei, Feng Wang, Anshul Shah, and Rama Chellappa. Learning to prompt your domain for vision-language models. In arXiv, 2024. 1, 3, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 456, + 294, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 511 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 511 + ], + "type": "text", + "content": "[34] Fu-En Yang, Chien-Yi Wang, and Yu-Chiang Frank Wang. Efficient model personalization in federated learning via client-specific prompt generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19159–19168, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 513, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 513, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 56, + 513, + 294, + 566 + ], + "type": "text", + "content": "[35] Elad Ben Zaken, Shauli Ravfogel, and Yoav Goldberg. Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. Annual Meeting of the Association for Computational Linguistics, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 568, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 622 + ], + "type": "text", + "content": "[36] Xin Zhang, Shixiang Shane Gu, Yutaka Matsuo, and Yusuke Iwasawa. Domain prompt learning for efficiently adapting clip to unseen domains. Transactions of the Japanese Society for Artificial Intelligence, 38(6):B-MC2_1, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 689 + ], + "type": "text", + "content": "[37] Haodong Zhao, Wei Du, Fangqi Li, Peixuan Li, and Gongshen Liu. Fedprompt: Communication-efficient and privacy-preserving prompt tuning in federated learning. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "text", + "content": "[38] Ke Zhao, Junchen Hu, Haidong Shao, and Jiabei Hu. Federated multi-source domain adversarial adaptation" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 206 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 335, + 72, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 72, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 335, + 72, + 553, + 105 + ], + "type": "text", + "content": "framework for machinery fault diagnosis with data privacy. Reliability Engineering & System Safety, 236:109246, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "type": "text", + "content": "[39] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 2, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "text", + "content": "[40] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9): 2337-2348, 2022. 2, 4, 6" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2375" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Federated Prompt-Tuning with Heterogeneous and Incomplete Multimodal Client Data/c4864325-3bcd-4609-87ea-0a89ff648c25_content_list.json b/2025/Federated Prompt-Tuning with Heterogeneous and Incomplete Multimodal Client Data/c4864325-3bcd-4609-87ea-0a89ff648c25_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0f5c202527878f79db2f22b13c46ee25d33b0d90 --- /dev/null +++ b/2025/Federated Prompt-Tuning with Heterogeneous and Incomplete Multimodal Client Data/c4864325-3bcd-4609-87ea-0a89ff648c25_content_list.json @@ -0,0 +1,1542 @@ +[ + { + "type": "text", + "text": "Federated Prompt-Tuning with Heterogeneous and Incomplete Multimodal Client Data", + "text_level": 1, + "bbox": [ + 181, + 130, + 818, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Thu Hang Phung1 Duong M. Nguyen1 Thanh Trung Huynh2 Quoc Viet Hung Nguyen3 \nTrong Nghia Hoang4* Phi Le Nguyen1* \n1 Institute of AI Innovation and Societal Impact, Hanoi University of Science and Technology \n2 EPFL 3 Griffith University 4 Washington State University", + "bbox": [ + 114, + 202, + 875, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper introduces a generalized federated prompt-tuning framework for practical scenarios where local datasets are multi-modal and exhibit different distributional patterns of missing features at the input level. The proposed framework bridges the gap between federated learning and multi-modal prompt-tuning which have traditionally focused on either uni-modal or centralized data. A key challenge in this setting arises from the lack of semantic alignment between prompt instructions that encode similar distributional patterns of missing data across different clients. To address this, our framework introduces specialized client-tuning and server-aggregation designs that simultaneously optimize, align, and aggregate prompt-tuning instructions across clients and data modalities. This allows prompt instructions to complement one another and be combined effectively. Extensive evaluations on diverse multimodal benchmark datasets demonstrate that our work consistently outperforms state-of-the-art (SOTA) baselines.", + "bbox": [ + 89, + 340, + 485, + 613 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 626, + 220, + 642 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The recent emergence of exceedingly large models [3] with immense scale and versatility due to being pre-trained on large swaths of generic data has resulted in a paradigm shift in machine learning (ML). Most task-specific models are now fine-tuned versions of those large models instead of being created from de novo learning processes. For example, most state-of-the-art (SOTA) models solving specific NLP tasks are fine-tuned versions of existing large pre-trained or foundation models such as BERT [12], GPT [5], and T5 [40]. Likewise, SOTA models for various computer vision tasks have also been derived from foundation models such as the vision transformer [13] pre-trained on extensive datasets like ImageNet [11]. The standard fine-tuning approach, however, assumes centralized data which is not always possible in practical scenarios where fine-tuning data", + "bbox": [ + 89, + 651, + 485, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "are scattered across multiple private siloed systems.", + "bbox": [ + 511, + 310, + 854, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To overcome this siloed data challenge, federated fin-tuning has recently been considered to aggregate the fin-tuned updates across private devices running their own (local) fine-tuning processes [9, 32, 47]. This approach can be further combined with an existing parameter-efficient tuning method called prompt tuning [25]. This method focuses on creating lightweight signals such as additional tokens that are prepended to the input of the pre-trained models [41, 44, 59, 60], providing them valuable contextual information to perform specific tasks more effectively. Local sets of prompts can then be aggregated using the conventional parameter aggregation procedure in FL [16, 29, 47, 50, 58].", + "bbox": [ + 511, + 325, + 906, + 521 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, such federated prompt-tuning techniques are restricted to learning with uni-modal data. In more practical scenarios, advances in sensor and device manufacturing increasingly enable the capture of multiple data modalities (e.g., text, images), leading to richer datasets for fine-tuning [6]. Leveraging such diverse multi-modal information can significantly enhance model performance, as demonstrated by the success of multi-modal foundation models like CLIP [39] and ViLT [21]. This has further motivated research on multi-modal federated learning [8, 33, 37, 52] and fine-tuning pre-trained multi-modal LLMs, i.e., multi-modal federated fine-tuning [56].", + "bbox": [ + 511, + 523, + 908, + 703 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite these advances, multi-modal federated learning research often adopts specific local architectures which cannot be learned via fine-tuning existing pre-trained multimodal foundation models [8, 33, 37, 52]. Existing multimodal federated fine-tuning research on the other hand assumes that all local datasets have the same set of modalities [56]. As such, these approaches are not suitable for situations where different clients might collect data using different types of sensors. Furthermore, the multi-modal sensor data across clients might even have different missing patterns of data entries. This leads to local datasets with both intra- and inter-heterogeneities. Intra-heterogeneity occurs when individual samples within a single dataset", + "bbox": [ + 511, + 704, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Authors", + "bbox": [ + 109, + 886, + 243, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3936", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "have missing entries for different subsets of features. Interheterogeneity arises when different clients have different distributions over their data-missing patterns.", + "bbox": [ + 89, + 90, + 482, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "There also exists recent work in prompt-tuning that learns a specific prompt for each subset of missing modalities [24]. Input samples having the same set of missing modalities are then augmented with the same (learnable) prompt during training, specializing it to complement for the corresponding missing modalities. This approach is, however, restricted to centralized data setting (i.e., no interheterogeneities) and does not generalize well to federated settings where different sets of local prompts might be biased towards different distributional patterns of missing data. A naive aggregation of these prompts might therefore collapse them into less informative prompts which are ignorant of the heterogeneities in missing-data patterns across clients, resulting in poor performance (see Fig. 1).", + "bbox": [ + 88, + 138, + 482, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To bridge this gap between multi-modal prompt-tuning and federated learning, we propose a multimodal federated prompt-tuning approach that optimizes and aggregates two distinct sets of (local) prompts for each client: inter-client and intra-client prompts. Inter-client prompts breakdown and encode prototypical distribution patterns of missing input data, which can be aligned and aggregated across clients to represent shared distributional information. In contrast, intra-client prompts encode information tailored to support specific local subsets of missing modalities, independent of the distributional patterns of missing data at each client. This is achieved via the following contributions:", + "bbox": [ + 89, + 353, + 482, + 534 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We develop a local input-adaptive fine-tuning algorithm which identifies and updates a subset of inter- and intra-client prompts that are most relevant to each individual sample. This is formulated as prompt-retrieval task which focuses on learning mapping functions that push an input towards its relevant prompts on an embedding space while pulling away from irrelevant prompts on an embedding space. This helps decompose and distill local fine-tuning knowledge into different parts of the prompt set, preventing learned instructions for different fine-tuning patterns from being overloaded on the same prompts (Section 3.2).", + "2. We develop a server-side model for aggregating and aligning inter-client prompts across clients. Both generation and alignment are framed as part of a clustering task, where inter-client prompts encoding similar input-level distribution patterns of missing data are identified and merged into more comprehensive and versatile prompt instructions. The empirical effectiveness of these combined prompts in reducing local training losses is used as a learning feedback to optimize this clustering model (Section 3.3).", + "3. We evaluate and demonstrate the effectiveness of our proposed approach through extensive experiments on two benchmark datasets, MM-IMDB [1] and UPMC Food" + ], + "bbox": [ + 89, + 540, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2d42a96d9e28d4269fffcac2392f709e1fffcd35722a50fe27a27aff2ab6142a.jpg", + "image_caption": [ + "Figure 1. Our approach with prompt alignment outperforms FedAvg prompt-tuning (w/o alignment) across settings with different data missing rates." + ], + "image_footnote": [], + "bbox": [ + 535, + 90, + 883, + 267 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "101 [46], comparing its performance against a variety of existing multimodal federated learning methods as well as repurposed prompt-tuning baselines for federated settings. The reported results show that our method consistently and significantly outperforms the baselines across various modality-missing scenarios, achieving new SOTA performance for multi-modal federated learning (Section 4).", + "bbox": [ + 511, + 316, + 906, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A bird-view of our proposed framework, Federated Prompt-Tuning with Heterogeneous and Incomplete Multimodal Client Data (FED-PRIME), is also provided in Figure 2.", + "bbox": [ + 511, + 439, + 906, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 496, + 663, + 512 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-Modal Federated Learning. Federated learning (FL) [17, 31, 53, 54] is a collaborative learning paradigm that enables multiple devices to train a shared model while keeping local data private. This design addresses concerns over privacy, security, and inefficiency in transferring large datasets [45, 57]. While FL has been widely studied [10, 14, 15, 18, 26, 27, 30, 34–36, 42, 43], most existing methods focus on uni-modal data (e.g., image or text). However, with the rapid advancement of mobile phones and IoT devices [6], local data increasingly comes from multiple modalities. Integrating such heterogeneous signals enables richer, more robust representations [19, 48], motivating growing interest in multi-modal FL. Yet many current approaches rely on custom architectures that cannot exploit pre-trained multi-modal models. More recent studies explore federated fine-tuning of pre-trained multi-modal models, but they typically assume each client holds the same set of modalities [56], limiting applicability in realistic heterogeneous environments.", + "bbox": [ + 511, + 522, + 906, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Missing Data in Multi-Modal FL. In real-world settings, local devices collect data from different types of sensors, leading to datasets with varying feature modalities (inter-heterogeneity). Moreover, sensor malfunctions or data corruption can cause missing features, with devices exhibiting distinct missing patterns (intra-heterogeneity).", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3937", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing solutions either treat each modality independently [55, 61] or are restricted to centralized learning [7, 38, 51], both overlooking inter-modal dependencies. Recent FL methods address inter-heterogeneity [8, 37] but not intra-heterogeneity, while approaches like FedInMM [52] and FedMAC [33] tackle intra-client missing data but ignore cross-client alignment, limiting effectiveness. A related line of work, including FedMSplit [8] and CACMRN [49], handles modality-incomplete data under a personalized FL framework, where each client retains its model. This paradigm differs from standard FL, which aims to improve a single shared model, making these methods not directly comparable.", + "bbox": [ + 89, + 90, + 483, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Addressing Missing Data with Prompt-Tuning. Prompt-tuning [41, 44, 59, 60] provides an effective way to adapt pre-trained models by learning prompts that add contextual information to input embeddings. Recent work has extended this idea to missing-modality scenarios by learning a prompt for each subset of missing features [20, 24], so that inputs with similar missing patterns are augmented with suitable prompts. However, these methods are developed for centralized training and do not address the additional heterogeneity introduced in FL, where prompts must be aligned and aggregated across clients. Designing effective alignment and aggregation strategies for such prompts is therefore a key challenge, as demonstrated in our framework.", + "bbox": [ + 89, + 289, + 483, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Multi-Modal Federated Prompt-Tuning", + "text_level": 1, + "bbox": [ + 89, + 510, + 447, + 529 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multimodal federated learning faces significant challenges due to missing data modalities, leading to both inter-client and intra-client heterogeneities that complicate knowledge aggregation and degrade overall model performance - see Fig. 1. To address this challenge (Section 3.1), we present a federated prompt-tuning framework based on the pretrained Vision and Language Transformer (ViLT) model, which comprises a prompt-based client design and a server-side prompt-aggregation algorithm. The client design features a fine-tuning algorithm that learns and decomposes the fine-tuned knowledge into separate sets of inter-client and intra-client prompts specialized for fine-tuning at different input-level patterns of missing data and input-agnostic patterns of missing modalities (Section 3.2). The server-side algorithm then learns alignment between inter-client prompts specializing for similar input-level patterns of missing data and aggregates them to create more versatile prompts. This is also combined with a standard FEDAVG-based aggregation of intra-prompts specializing for input-agnostic patterns of missing modalities (Section 3.3) - see Fig. 2.", + "bbox": [ + 89, + 537, + 483, + 840 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 89, + 845, + 294, + 861 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In multimodal federated learning (MMFL), we assume there are $n$ clients and $r$ data modalities (e.g., text, image).", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Each client $t \\in [n]^{1}$ has access to a private dataset $D_{t} \\triangleq \\{x(M_{t,s}), z_{t,s}\\}_{s=1}^{m}$ , where $x(M_{t,s})$ denote the observed features for the set of modalities, $M_{t,s} \\subseteq [r]$ recorded for the $s$ -th data point of the $t$ -th client, and $z_{t,s}$ denotes the corresponding target label. The goal is to learn a multimodal model $w_{*}$ using all the private datasets $\\{D_{t}\\}_{t=1}^{n}$ without centralizing them. That is,", + "bbox": [ + 511, + 90, + 906, + 196 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {w} _ {*} = \\underset {\\boldsymbol {w}} {\\arg \\min } \\left\\{L (\\boldsymbol {w}) \\triangleq \\frac {1}{n} \\sum_ {t = 1} ^ {n} L _ {t} (\\boldsymbol {w}) \\right\\} \\text {w i t h}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 202, + 877, + 243 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {t} (\\boldsymbol {w}) \\triangleq \\frac {1}{n} \\sum_ {s = 1} ^ {m} \\ell \\left(F \\left(\\boldsymbol {x} \\left(M _ {t, s}\\right); \\boldsymbol {w}\\right), z _ {t, s}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 244, + 906, + 284 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $F(\\pmb{x}(M_{t,s}); \\pmb{w})$ denotes the multimodal prediction model with parameter $\\pmb{w}$ and $\\ell(F(\\pmb{x}(M_{t,s}); \\pmb{w}), z_{t,s})$ denotes the individual loss of $\\pmb{w}$ on the $s$ -th data point. Previous work [8, 33, 37, 52] in MMFL has largely focused on designing $F(\\pmb{x}(M_{t,s}); \\pmb{w})$ without leveraging existing knowledge in pre-trained foundation model. Instead, we let $F(\\pmb{x}(M_{t,s}); \\pmb{w})$ be a prompt-tuned version of a pre-trained transformer-based model,", + "bbox": [ + 511, + 291, + 906, + 412 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nF (\\boldsymbol {x} (M); \\boldsymbol {w}) \\triangleq F _ {c} \\left(F _ {p} \\left(F _ {e} (\\boldsymbol {x} (M)) \\circ \\boldsymbol {w} _ {p}\\right); \\boldsymbol {w} _ {c}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 417, + 905, + 444 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $F_{c}(:,\\boldsymbol{w}_{c})$ is the prediction head parameterized with $\\boldsymbol{w}_{c}$ , $F_{e}(\\boldsymbol{x}(M)) = F_{e}(\\boldsymbol{x}_{1}) \\circ \\dots \\circ F_{e}(\\boldsymbol{x}_{|M|})$ where $F_{e}(\\boldsymbol{x}_{i})$ denote the tokenized input sequence of each unimodal feature $\\boldsymbol{x}_{i}$ in $\\boldsymbol{x}(M) \\triangleq (\\boldsymbol{x}_{1}, \\boldsymbol{x}_{2}, \\dots, \\boldsymbol{x}_{|M|})$ , $F_{p}()$ denotes the frozen multimodal pre-trained model – e.g., ViLT [21] – and $\\boldsymbol{w}_{p}$ denotes the set of (learnable) prompts. Here, $\\circ$ denotes the set concatenation operator and $\\boldsymbol{w} = (\\boldsymbol{w}_{c}, \\boldsymbol{w}_{p})$ denotes the full set of tuning parameters. Due to the decentralized and private nature of the local datasets $\\{D_{t}\\}_{t=1}^{n}$ , solving Eq. (1) directly is not possible. Instead, following the federated learning [31] paradigm, its optimal solution $\\boldsymbol{w}_{*}$ is approximated via a number of iterations over two interleaving client- and server-side steps:", + "bbox": [ + 511, + 448, + 906, + 645 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {w} _ {t} ^ {(h + 1)} = \\operatorname {u p d a t e} \\left(L _ {t}, \\boldsymbol {w} _ {*} ^ {(h)}\\right) \\forall t \\in [ n ], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 651, + 906, + 676 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {w} _ {*} ^ {(h + 1)} = \\text {a g g r e g a t e} \\left(\\boldsymbol {w} _ {1} ^ {(h + 1)}, \\dots , \\boldsymbol {w} _ {n} ^ {(h + 1)}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 678, + 905, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\boldsymbol{w}_t^{(h)}$ denote the estimation of the local tuning parameter $\\boldsymbol{w}_t$ of client $t$ after $h$ iterations. For example, the update step can be a vanilla gradient descent (GD) optimization of the prompt $\\omega$ and the aggregate step denotes the average of the locally optimized prompts. Such vanilla approach however does not account for both the inter- and intra-heterogeneities among local prompts (see Section 1) which ultimately risks aggregating incompatible prompts and degrade the overall performance (see Section 4). To avoid this, we will introduce novel designs for client- and server-prompts as detailed in the followings.", + "bbox": [ + 511, + 710, + 906, + 878 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "$^{1}[n]\\triangleq \\{1,2,\\dots ,n\\}$", + "bbox": [ + 531, + 885, + 653, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3938", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f0ddb7f20be97d3b27ffd81002ae84c428259f17301d7fbd3677cd308e86b259.jpg", + "image_caption": [ + "Figure 2. Overview of the proposed multi-modal federated prompt-tuning framework - see Alg. 3. Each client maintains two (learnable) sets of intra- and inter-client prompts. At the beginning of each iteration, each client performs local training via Eq. (7) and Eq. (8). Local sets of intra- and inter-client prompts are then sent to the server for aggregation - see Eq. (10)-Eq. (14)." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 911, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Client Design", + "text_level": 1, + "bbox": [ + 89, + 430, + 230, + 448 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To enable effective prompt-aggregation across heterogeneous clients, we need to separate the fine-tuning treatment for input-level patterns of missing data and input-agnostic patterns of missing modalities. This can be achieved via further parameterizing $\\boldsymbol{w}_p = (\\boldsymbol{w}_p^{inter}, \\boldsymbol{w}_p^{intra})$ with two distinct subsets of $\\tau$ inter- and intra-client prompts:", + "bbox": [ + 89, + 453, + 483, + 544 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {w} _ {p} ^ {\\text {i n t e r}} \\triangleq \\left\\{\\boldsymbol {p} _ {1} ^ {\\text {i n t e r}}, \\boldsymbol {p} _ {2} ^ {\\text {i n t e r}}, \\dots , \\boldsymbol {p} _ {\\tau} ^ {\\text {i n t e r}} \\right\\}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 551, + 482, + 578 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {w} _ {p} ^ {\\text {i n t r a}} \\triangleq \\left\\{\\boldsymbol {p} _ {1} ^ {\\text {i n t r a}}, \\boldsymbol {p} _ {2} ^ {\\text {i n t r a}}, \\dots , \\boldsymbol {p} _ {\\tau} ^ {\\text {i n t r a}} \\right\\}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 579, + 482, + 606 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which are learned separately. The learned inter-client prompts $\\pmb{w}_{p}^{inter}$ across clients will be sent to the server which clusters them into different groups based on their specialization for similar input-level patterns of missing data - Section 3.3. Local sets of intra-client prompts $\\pmb{w}_{p}^{intra}$ can be averaged via FEDAVG [31] since their specialization is for input-agnostic patterns of missing modalities. Intuitively, the aggregation mechanisms will determine how different aspects of fine-tuning knowledge are encoded across the prompt sets. For example, if knowledge regarding different input-level patterns of missing data is (incorrectly) incorporated into $\\{\\pmb{p}_i^{intra}\\}_{i=1}^\\tau$ , they will be averaged out and collapsed into less informative prompts due to the non-clustering aggregation mechanism of $\\{\\pmb{p}_i^{intra}\\}_{i=1}^\\tau$ .", + "bbox": [ + 89, + 613, + 482, + 825 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As this potential loss of information will decrease the overall performance, the (learnable) prompt parameters will be updated to prevent this from happening. Otherwise, if (common) knowledge patterns regarding input-agnostic patterns of missing modalities are (incorrectly) incorporated", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "into $\\left\\{\\pmb{p}_i^{inter}\\right\\}_{i=1}^{\\tau}$ , their modeling bandwidth for inter-client heterogeneities will be reduced, which will also hamper the performance. This will, in turn, generate a prompt update signal to avoid this as well. Furthermore, within each set of inter- and intra-client prompts, there might also be different patterns that need to be separated and internalized into different subset of prompts.", + "bbox": [ + 511, + 430, + 906, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To enable this, we develop an input-adaptive learning mechanism for both sets of prompts which identifies and updates the most relevant subset of prompts for each (multimodal) input sample. This is formulated as an information-retrieval task, learning a pair of key and query functions, $k(\\pmb{p})$ and $q(\\pmb{x}(M))$ , which transform the prompt $\\pmb{p}$ and input $\\pmb{x}$ with the set $M$ observed modalities, respectively, into a shared metric space. We can then define a geometric input-prompt distance, e.g., $d(\\pmb{x}(M), \\pmb{p}) = \\cos(q(\\pmb{x}(M)), k(\\pmb{p}))$ which can be used to measure their contextual relevance. To avoid loading diverse fine-tuning knowledge into a single prompt, we regularize the (local) loss function to penalize it when a less relevant prompt is selected as fine-tuning instruction for an input $\\pmb{x}(M)$ :", + "bbox": [ + 511, + 537, + 908, + 750 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {w} _ {t} \\triangleq \\underset {\\boldsymbol {w}} {\\arg \\min } L _ {t} ^ {\\prime} (\\boldsymbol {w}) \\quad \\text {w h e r e} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 753, + 906, + 779 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {t} ^ {\\prime} (\\boldsymbol {w}) \\triangleq \\sum_ {s = 1} ^ {m} \\left\\{\\ell \\left(F \\left(\\boldsymbol {x} \\left(M _ {t, s}\\right); \\boldsymbol {w} ^ {\\prime}\\right), z _ {t, s}\\right) \\right\\} \\\\ + \\sum_ {s = 1} ^ {m} \\left\\{r \\left(\\boldsymbol {x} \\left(M _ {t, s}\\right), \\boldsymbol {w} _ {p} ^ {\\prime}\\right) \\right\\}, \\tag {8} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 781, + 903, + 866 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\ell (\\pmb {w}^{\\prime})$ is defined in Eq. (1) and $\\pmb{w}^{\\prime} = (\\pmb{w}_{c},\\pmb{w}_{p}^{\\prime})$ denote the customized set of tuning instructions for each", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3939", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d354035796080b754dc6c8bf5c32f2276441d652b5d4db4b27aa301009bf8fe3.jpg", + "image_caption": [ + "Figure 3. Workflow of the prompt-alignment algorithm. At each iteration, each client samples a subset of summarizing prompts using its query and key functions. The client then performs a local update, resulting in heterogeneous prompt sets, which are subsequently sent to the server to be clustered and summarized into new summarizing prompt sets for the next iteration." + ], + "image_footnote": [], + "bbox": [ + 102, + 90, + 472, + 265 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "input $\\pmb{x}(M_{t,s})$ where $\\pmb{w}_p' = (\\pmb{w}_p^{inter'}, \\pmb{w}_p^{intra'})$ such that $\\pmb{w}_p^{inter'} \\subseteq \\pmb{w}_p^{inter}$ and $\\pmb{w}_p^{intra'} \\subseteq \\pmb{w}_p^{intra}$ are the sets of $\\kappa$ most relevant prompts to $\\pmb{x}(M_{t,s})$ in $\\pmb{w}_p^{inter}$ and $\\pmb{w}_p^{intra}$ . The regularizer $r(\\pmb{x}(M), \\pmb{w}_p')$ is then defined as", + "bbox": [ + 89, + 383, + 483, + 455 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr \\left(\\boldsymbol {x} (M), \\boldsymbol {w} _ {p} ^ {\\prime}\\right) \\triangleq \\sum_ {\\boldsymbol {p} \\in \\boldsymbol {w} _ {p} ^ {\\prime}} \\left\\{d (\\boldsymbol {x} (M), \\boldsymbol {p}) \\right\\}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 464, + 483, + 510 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Intuitively, the regularizer penalizes the total distance between the (multimodal) input $\\pmb{x}(M)$ and the $\\kappa$ closest prompts according to the aforementioned distance metric $d(\\pmb{x}(M), \\pmb{p})$ . Minimizing both the original local loss and this regularizer allows the local model to simultaneously learn to (1) internalize fine-tuning patterns of inter- and intra-heterogeneities to minimize the original loss $L_{t}(\\pmb{w})$ ; and (2) ensure that selected prompts for a given input will be close to it to reduce the penalty $r(\\pmb{x}(M), \\pmb{w}_{p}^{\\prime})$ . A pseudocode of this client design is detailed in Suppl. A.", + "bbox": [ + 89, + 520, + 483, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Server Aggregation", + "text_level": 1, + "bbox": [ + 89, + 680, + 277, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We will now establish a principled algorithm to align and aggregate inter-client prompts specializing in similar input-level patterns of missing data2. This is not straightforward since inter-client prompts at the same position in the prompt sets across different clients might not specialize for similar patterns of missing data due to client heterogeneities: certain missing-data patterns are not observed at certain clients.", + "bbox": [ + 89, + 702, + 482, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To find the correct alignment, we formulate it as a clustering optimization task. Suppose each client $t$ communicates a set of inter-client prompts $\\boldsymbol{w}_t^{inter} = (\\boldsymbol{p}_t^i)_{i=1}^\\tau$ . Each", + "bbox": [ + 89, + 808, + 482, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "potential clustering of these prompts thus defines a candidate alignment, i.e., prompts in the same cluster are considered aligned and can be aggregated. A (learnable) cluster center can then be associated with the aggregated prompt - see Fig. 3. Both the cluster assignment and centers can be found via solving the following optimization task with discrete alignment variables $\\alpha_{t}^{pq}\\in \\{0,1\\}$ and continuous center variables $\\theta_{q}$ for each cluster $q\\in [n\\times \\tau ]$ ..", + "bbox": [ + 511, + 90, + 906, + 213 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {\\boldsymbol {\\alpha}, \\boldsymbol {\\theta}, \\gamma} G (\\boldsymbol {\\alpha}, \\boldsymbol {\\theta}, \\gamma) \\triangleq \\sum_ {t = 1} ^ {n} \\sum_ {p = 1} ^ {\\tau} \\sum_ {q = 1} ^ {n \\times \\tau} \\alpha_ {t} ^ {p q} \\cdot c \\left(\\boldsymbol {p} _ {t} ^ {p}, \\boldsymbol {\\theta} _ {q},; \\gamma\\right) (10) \\\\ \\mathrm {s . t .} \\alpha_ {t} ^ {p, 1} + \\dots + \\alpha_ {t} ^ {p, n \\times \\tau} = 1, (11) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 224, + 906, + 289 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha \\triangleq (\\alpha_{t}^{p,q})$ and $\\alpha_{t}^{p,q} \\in \\{0,1\\} \\forall (t,p,q)$ indicates whether the $p$ -th inter-client prompt of client $t$ is matched to the $q$ -th cluster. The corresponding cluster center is then optimized to have minimum total distance to all cluster members according to a (learnable) cost function $c(\\pmb{p}_{t}^{p},\\pmb{\\theta}_{q})$ . The optimized cluster center can then be used as the aggregated or summarizing prompt. Minimizing Eq. (10) thus allows us to align and summarize/aggregate the inter-client prompts across clients simultaneously. Note that Eq. (11) enforces a constraint that intuitively indicates two prompts from the same clients should not be matched to the same cluster and get aggregated. This is because local prompts from the same client have already been learned to encode different finetuning aspects. Aggregating them will therefore collapse them into less informative ones.", + "bbox": [ + 511, + 297, + 906, + 526 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition, some summarizing/aggregated prompts might be more generic than others and can be used to provide fine-tuning instructions to a wider range of (multimodal) inputs. The update of such prompts should be prioritized during training and to achieve that, we need a mechanism to capture this prompt popularity. To achieve this, we augment the optimization task in Eq. (10) above with a regularizer that penalizes alignment to less popular prompts,", + "bbox": [ + 511, + 527, + 905, + 650 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} R (\\boldsymbol {\\alpha}, \\zeta) \\triangleq \\sum_ {t = 1} ^ {n} \\sum_ {p = 1} ^ {\\tau} \\sum_ {q = 1} ^ {n \\times \\tau} \\left(\\alpha_ {t} ^ {p, q} \\cdot \\log U \\left(\\boldsymbol {\\theta} _ {q}; \\zeta\\right)\\right) \\tag {12} \\\\ + \\sum_ {t = 1} ^ {n} \\sum_ {p = 1} ^ {\\tau} \\sum_ {q = 1} ^ {n \\times \\tau} \\Big (1 - \\alpha_ {t} ^ {p, q} \\Big) \\cdot \\log \\Big (1 - U (\\boldsymbol {\\theta} _ {q}; \\zeta) \\Big), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 660, + 905, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $U(\\pmb{\\theta}_q;\\zeta)$ is the popularity function of $\\pmb{\\theta}_q$ , which is parameterized by learnable parameter $\\zeta$ . For example, $U(\\pmb{\\theta}_q;\\zeta) = \\sigma(g(\\pmb{\\theta}_q;\\zeta)) \\in (0,1)^3$ where $g(\\pmb{\\theta}_q;\\zeta)$ is a deep neural net parameterized with $\\zeta$ . Putting Eq. (13) and Eq. (10) together, we arrive at our final loss function,", + "bbox": [ + 511, + 758, + 906, + 835 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\boldsymbol {\\alpha}, \\boldsymbol {\\theta}, \\gamma , \\zeta} {\\text {m i n i m i z e}} \\left\\{G (\\boldsymbol {\\alpha}, \\boldsymbol {\\theta}, \\gamma) + R (\\boldsymbol {\\alpha}, \\zeta) \\right\\}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 847, + 906, + 875 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "$^3\\sigma(.)$ denotes the sigmoid function.", + "bbox": [ + 529, + 886, + 723, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "2This section only describes the aggregation algorithm for inter-client prompts since intra-client prompts are specialized for input-agnostic patterns of missing modalities and can be aggregated via FEDAVG [31].", + "bbox": [ + 89, + 862, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3940", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which can be optimized via alternating between (1) optimizing $(\\theta, \\zeta, \\gamma)$ while fixing $\\alpha$ ; and (2) optimizing $\\alpha$ while fixing $(\\theta, \\zeta, \\gamma)$ . The first optimization sub-task is straightforward while the latter is less trivial due to the constraint as well as the discrete nature of the optimizing variables. To sidestep the technical challenge in solving for $\\alpha$ while fixing $(\\theta, \\zeta, \\gamma)$ , we approximate it by optimizing for $\\alpha_{t} = \\{\\alpha_{t}^{p,q}\\}_{p,q}$ while also fixing $\\alpha_{-t} = \\{\\alpha_{-t}^{p,q}\\}_{p,q}$ . This results in a minimum weighted linear sum task,", + "bbox": [ + 89, + 90, + 483, + 227 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {\\boldsymbol {\\alpha} _ {t}} L \\left(\\boldsymbol {\\alpha} _ {t}\\right) \\triangleq \\sum_ {p = 1} ^ {\\tau} \\sum_ {q = 1} ^ {n \\times \\tau} \\left(\\alpha_ {t} ^ {p, q} \\cdot v (p, q)\\right) \\text {w h e r e} \\tag {14} \\\\ v (p, q) \\triangleq \\left(\\log \\left\\{\\frac {U (\\boldsymbol {\\theta} _ {q} ; \\zeta)}{1 - U (\\boldsymbol {\\theta} _ {q} ; \\zeta)} \\right\\} + c (\\boldsymbol {p} _ {t} ^ {p}, \\boldsymbol {\\theta} _ {q}; \\gamma)\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 237, + 495, + 333 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "which can be solved effectively in polynomial time using the Hungarian algorithm [23]. An overview and pseudocode of the server-side aggregation algorithm are presented in Fig. 3 and Suppl. A. A thorough evaluation of this framework is detailed next in Section 4.", + "bbox": [ + 89, + 340, + 483, + 416 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Empirical Evaluation", + "text_level": 1, + "bbox": [ + 89, + 429, + 295, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This section provides a detailed empirical evaluation of our proposed framework, FED-PRIME. Our evaluation is conducted on two datasets UPMC Food-101 [46] and MM-IMDB [1] and several baselines adapting existing works in federated fine-tuning (for unimodal data) and multi-modal prompt-tuning (for centralized data) to our novel setting on decentralized multi-modal data with missing modalities and features. Detailed description of the datasets and baselines are provided in Section 4.1. Our main results and ablation studies are reported in Section 4.2.", + "bbox": [ + 89, + 454, + 483, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experiment Settings", + "text_level": 1, + "bbox": [ + 89, + 614, + 284, + 630 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A. Datasets. Following the experiment protocol in [24], we evaluate FED-PRIME on two multimodal datasets:", + "bbox": [ + 89, + 636, + 482, + 666 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. UPMC Food-101 [46] is an image-text classification dataset consisting of noisy image-text pairs collected from Google Image Search. Its classes are aligned with those in the larger, publicly available ETHZ Food-101 dataset [4]. For the federated setting, we select only eight most frequent classes from the original UPMC Food-101.", + "2. MM-IMDB [1] is a movie genre classification dataset with both image and text modalities. Since each movie can belong to multiple genres, the task is originally a multi-label classification problem, where genres are predicted using image, text, or both. For the federated setting, we use only single-label instances and select eight most frequent classes only from the original MM-IMDB." + ], + "bbox": [ + 89, + 670, + 482, + 866 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Each dataset is first partitioned into train and test sets with the 80:20 ratio. The (heterogeneous) missing patterns on the train/test data are simulated using the following protocols.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "B. Training Data Simulation. Following [24], we simulate missing data phenomenon within a dataset using a missing rate $\\eta \\in (0,1)$ as detailed in 3 simulation cases below:", + "bbox": [ + 511, + 140, + 903, + 186 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Missing Image. We drop the image features in $\\eta$ percent of the data and keep the remaining $(1 - \\eta)$ percent complete.", + "2. Missing Text. We drop the text features in $\\eta$ percent of the data and keep the remaining $(1 - \\eta)$ percent complete.", + "3. Missing Both. We drop the image component in $\\eta /2$ percent of the data, and the text component in the other $\\eta /2$ percent, leaving the remaining $(1 - \\eta)$ percent complete." + ], + "bbox": [ + 511, + 190, + 903, + 306 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The missing-simulated train set is then partitioned into $n$ subsets for $n$ clients via random sampling. The resulting client datasets will therefore have heterogeneous distributions over missing-data patterns.", + "bbox": [ + 511, + 309, + 903, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "C. Test Data Simulation. There are 5 test scenarios, which include: (1) Full Modal (no missing data); (2) Image Only (drop all text features); (3) Text Only (drop all image features); (4) Miss Both (run Missing Both on test data); and (5) $\\sim$ Train (same simulation as the training data).", + "bbox": [ + 511, + 373, + 903, + 449 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "D. Metrics. Following prior protocol in [24], we use F1-macro and classification accuracy to evaluate the following baseline methods on MM-IMDB and UPMC Food-101.", + "bbox": [ + 511, + 453, + 903, + 498 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "E. Baseline Methods. Following [24], we adopt the pretrained multimodal transformer ViLT [21] as the backbone of our client design. Each client model is a fine-tuned version of this backbone which is generated based on its own local dataset. As no prior work has addressed modality missing in multimodal fine-tuning within federated learning settings, our comparison baselines mostly feature adapted versions of existing work in federated learning and multimodal prompt-tuning. This includes adaptations of the latest missing prompt-tuning SOTA method in [24] into the context of the classical FedAvg method as well as that of a more recent FL approach, FedMSplit [8]. These are referred to as FEDAVG-P and FEDMSPLIT-P, respectively. In addition, to underscore the importance of separating fine-tuning knowledge across intra- and inter-client prompt sets, we also include two simplified variants of FED-PRIME utilizing either inter-client prompt pooling or intra-client prompt pooling, along with their respective aggregations. We refer to these variants as FED-INTER and FED-INTRA, respectively (see more details in Supplementary B).", + "bbox": [ + 511, + 503, + 905, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Experimental Results", + "text_level": 1, + "bbox": [ + 511, + 816, + 715, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This section provides extensive performance comparison between FED-PRIME and other baseline methods across various settings of missing data in the train and test sets, as previously stated in Section 4.1. In addition, we also provide", + "bbox": [ + 511, + 839, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "4Our implementation is available at https://github.com/hangpt01/FedPrime.", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3941", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5fb19d4f2e0c8b1bdc77b26f5e9c82a5f181f1092664d5ef57d621f968613a3f.jpg", + "table_caption": [ + "Table 1. Performance achieved by the baselines on the Food-101 (left) and MM-IMDB (right) datasets under various missing-data settings in the train and test sets. This includes all combinations of the 3 train scenarios and 5 test scenarios (i.e., 15 experiments) per dataset, as previously stated in Section 4.1. The best and second-best results are highlighted in bold red, and blue, respectively. Note that in the Miss Both training scenarios, Test ( $\\sim$ Train) and Test (Miss Both) represent the same test scenario according to the definition of $\\sim$ Train. The reported result for Test (Miss Both) in this case is the same as that for Test ( $\\sim$ Train), as indicated by a dash (-)." + ], + "table_footnote": [ + "(*) Improv. shows the relative performance improvement between our proposal and the second-best. (in percentage)." + ], + "table_body": "
DatasetsUPMC Food-101MM-IMDB
TrainMethodTest(∼ Train)Test(Miss Both)Test(Full Modal)Test(Text only)Test(Image only)Test(∼ Train)Test(Miss Both)Test(Full Modal)Test(Text only)Test(Image only)
Miss TextFEDAVG-P15.71 ± 2.2714.90 ± 1.5721.56 ± 7.8116.91 ± 0.6915.36 ± 0.4322.42 ± 2.2721.89 ± 1.3430.78 ± 1.6518.40 ± 0.0614.53 ± 4.56
FEDMSPLIT-P15.62 ± 1.5117.50 ± 1.9725.27 ± 8.2218.78 ± 0.6417.50 ± 1.9721.02 ± 1.8919.97 ± 0.7424.39 ± 6.0814.38 ± 2.7618.09 ± 6.13
FED-INTER54.82 ± 19.0148.87 ± 24.6459.17 ± 27.0635.13 ± 26.7856.59 ± 15.1218.25 ± 3.5016.95 ± 3.5718.67 ± 7.6315.03 ± 4.6618.01 ± 1.95
FED-INTRA61.71 ± 17.2248.09 ± 19.1262.06 ± 26.9822.51 ± 5.9262.64 ± 11.8313.38 ± 1.7312.77 ± 0.8512.55 ± 1.6711.31 ± 0.3814.33 ± 1.80
FED-PRIME78.88 ± 0.9080.38 ± 0.6592.12 ± 0.4073.01 ± 4.2576.83 ± 1.2231.92 ± 0.2031.48 ± 0.3037.67 ± 0.0430.60 ± 1.4130.69 ± 1.41
Improv. (%)27.82 ↑64.48 ↑48.44 ↑107.83 ↑22.65 ↑42.37 ↑43.81 ↑22.35 ↑66.30 ↑69.65 ↑
Miss ImageFEDAVG-P17.35 ± 4.7715.12 ± 1.4816.84 ± 2.3718.12 ± 6.4914.81 ± 0.2427.69 ± 5.9722.55 ± 3.0631.94 ± 0.9823.76 ± 11.7212.29 ± 0.47
FEDMSPLIT-P74.16 ± 10.5648.88 ± 10.2645.64 ± 32.4388.65 ± 2.1714.81 ± 0.9019.11 ± 11.3316.61 ± 7.2218.19 ± 12.5518.12 ± 12.3012.81 ± 1.25
FED-INTER77.96 ± 11.6264.62 ± 10.2282.08 ± 7.7577.69 ± 12.3537.56 ± 6.4918.79 ± 5.2317.93 ± 3.6020.56 ± 3.5617.67 ± 6.7915.47 ± 2.51
FED-INTRA22.84 ± 3.5220.13 ± 1.7223.48 ± 1.8624.46 ± 2.9916.66 ± 1.3215.75 ± 4.3414.06 ± 3.1615.68 ± 4.7714.53 ± 3.6511.71 ± 0.42
FED-PRIME90.55 ± 0.2279.12 ± 0.4992.89 ± 0.2190.18 ± 0.2954.14 ± 2.5036.08 ± 0.3531.35 ± 0.6138.49 ± 0.5636.91 ± 0.5918.15 ± 0.66
Improv. (%)16.15 ↑22.44 ↑13.17 ↑1.73 ↑44.14 ↑30.30 ↑39.02 ↑20.51 ↑55.35 ↑17.32 ↑
Miss BothFEDAVG-P14.57 ± 1.50-17.17 ± 4.3716.40 ± 4.0513.24 ± 0.3226.45 ± 2.63-33.03 ± 2.5624.12 ± 11.3020.21 ± 1.98
FEDMSPLIT-P49.15 ± 24.76-64.78 ± 36.6264.62 ± 36.5121.49 ± 7.1924.25 ± 5.02-26.05 ± 11.1726.02 ± 9.6419.79 ± 6.20
FED-INTER56.32 ± 21.77-69.57 ± 19.4145.15 ± 34.0959.30 ± 10.8426.53 ± 0.90-31.97 ± 2.2229.69 ± 2.2121.63 ± 0.77
FED-INTRA49.28 ± 32.87-56.70 ± 37.9043.24 ± 34.1949.85 ± 25.4411.90 ± 0.37-12.47 ± 0.4511.46 ± 0.3312.83 ± 0.92
FED-PRIME84.44 ± 2.65-93.64 ± 0.5887.95 ± 0.9172.41 ± 3.8832.01 ± 2.51-38.68 ± 0.6531.00 ± 2.9726.01 ± 0.12
Improv. (%)49.93 ↑-34.60 ↑36.10 ↑22.11 ↑20.66 ↑-17.11 ↑4.41 ↑20.25 ↑
", + "bbox": [ + 91, + 169, + 906, + 405 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ablation studies on (1) performance robustness against various missing rate $\\eta$ ; (2) the ablated impact of having inter- and intra-client prompt sets; and (3) training convergence.", + "bbox": [ + 89, + 446, + 482, + 492 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.1. Main Results", + "text_level": 1, + "bbox": [ + 89, + 503, + 230, + 517 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare the performance of FED-PRIME against the other baselines on an extensive set of experiments across all datasets. This includes all combinations of the 3 missing simulation scenarios on train data and 5 test scenarios (with both missing and no missing data) as previously detailed in parts B and C of Section 4.1. All results are reported in Table 1 which show that FED-PRIME consistently outperforms all baselines in all train/test scenarios with various missing-data patterns. The performance on the UPMC Food-101 and MM-IMDB datasets is measured in terms of classification accuracy and F1-macro, respectively. The F1-macro performance on MM-IMDB is relatively low across all settings due to the class imbalance nature of the dataset. In each experiment, the relative performance improvement of FED-PRIME over the second best baseline is reported, which ranges between $1.73\\%$ and $107.83\\%$ for the UPMC Food-101 dataset and between $4.41\\%$ to $69.65\\%$ for the MM-IMDB dataset. Notably, FED-PRIME significantly outperforms both of its simplified variants FED-INTRA and FED-INTER, which either handle intra- or interheterogeneities, in all settings. This highlights the remarkable synergies and essential of handling both types of heterogeneities in multi-modal federated learning with missing data. In addition, it can also be observed that in more than half of the experiment settings, the second best base", + "bbox": [ + 89, + 523, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "line is either FED-INTRA or FED-INTER which significantly outperform both FEDAVG-P and FEDMSPLIT-P in those cases. These observations complement our findings of the isolated impact of these components in our ablation studies in Section 4.2.4.", + "bbox": [ + 511, + 446, + 906, + 521 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.2. Impact of Missing Modalities.", + "text_level": 1, + "bbox": [ + 511, + 532, + 767, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Across different testing scenarios, we observe that the highest test performance is always achieved when there is no missing data in the test set as expected. Another interesting observation is that in the Miss Text training scenario, where approximately $70\\%$ of the text features are lost on the training data, the proposed model still performs surprisingly well on the Text Only test scenarios. This suggests that the multi-modal prompt alignment is effective in recovering text information, highlighting the proposed design's adaptability. However, in the Miss Image training scenario, the test performance on the Image Only case suffers instead a significant performance drop. Upon investigation, we believe this is likely due to the VILT backbone's lightweight image processing module, which was optimized to work in tandem with text for competitive performance, hinting at a potential text-centered nature of VILT during pre-training. This is in fact well-aligned with existing observations that co-training with text often leads to improved performance.", + "bbox": [ + 511, + 551, + 906, + 825 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.3. Robustness Against Various Missing Rates.", + "text_level": 1, + "bbox": [ + 511, + 835, + 857, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To demonstrate the robustness of FED-PRIME against different missing rates $\\eta$ , we run multiple comparisons of FED-PRIME, FEDAVG-P, and its centralized variant [24]", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3942", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a8c4a2b4d053f7b04dca0657b87f5b136b5b6deaf01c181c0cc83ea1db0121ba.jpg", + "image_caption": [ + "Figure 4. Performance comparison on UMPC Food-101 under Miss Both scenarios with various missing rates." + ], + "image_footnote": [], + "bbox": [ + 89, + 85, + 326, + 227 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/96ad7d5f6bc0f66a2028428fce690da69fc4747b2de093344daa46f39ffb7128.jpg", + "image_caption": [ + "Figure 5. Plots of the train/test loss convergence on Food-101 under Miss-Text.FED-PRIME converges fastest." + ], + "image_footnote": [], + "bbox": [ + 336, + 87, + 580, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/814c2782251785cb72023581f7eb545e4c8ad9679dd6dc8d3d8c8a4956ee4f2f.jpg", + "image_caption": [ + "Figure 6. Convergence of Euclidean distance between aggregated prompt centroids on the UPMC Food-101 under Miss Image scenario." + ], + "image_footnote": [], + "bbox": [ + 629, + 87, + 870, + 229 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "across a wide range of $\\eta \\in \\{0.00, 0.25, 0.50, 0.75, 1.00\\}$ under the Miss Both training scenarios. The reported results in Fig. 1 and Fig. 4 show that our model significantly outperforms FEDAVG-P and its centralized counterpart Centralized-P. As the missing rate increases, our approach consistently maintains a competitive accuracy above $80\\%$ , closely approaching the upper-bound performance of FEDAVG-P in the centralized setting, which benefits from a centralized dataset with better-aligned modalities. These results demonstrate the robustness of our model across varying missing rates and highlight the effectiveness of our federated prompt-tuning approach, achieving performance comparable to a centralized prompt-tuning model despite having access only to decentralized local datasets with heterogeneous missing patterns.", + "bbox": [ + 88, + 297, + 485, + 525 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 535, + 243, + 550 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As explained previously in Section 4.1, FED-INTER and FED-INTRA are in variants of FED-PRIME when we drop the intra- and inter-client prompt sets from the client design, respectively. Table 1 shows that FED-PRIME consistently outperforms both FED-INTER and FED-INTRA in all test conditions with substantial gap. The lack of intra-client prompts causes FED-INTER to underperform FED-PRIME suggests that having a dedicated prompt set to capture the common and highly generalizable input-agnostic aspect of fine-tuning is essential to improve performance. Likewise, FED-INTRA, which omits the inter-client prompt set, also shows a significant performance drop. This highlights the importance of capturing and correctly aggregating input-level patterns of missing data through prompt alignment.", + "bbox": [ + 89, + 554, + 483, + 781 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2.5. Convergence Analysis", + "text_level": 1, + "bbox": [ + 89, + 790, + 290, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Model Convergence. Fig. 5 shows that FED-PRIME has superior performance, achieving faster convergence and significantly lower final loss values in both training and testing compared to FED-INTER and FED-INTRA. This indicates better learning efficiency and robustness. Both FED-INTER and FED-INTRA converge more slowly, with", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "consistently higher loss values, especially in the early rounds. Additionally, our method is more stable across server-aggregation rounds, with minimal fluctuations. In contrast, FED-INTER and FED-INTRA exhibit significantly more variability, particularly toward the end. This highlights the effectiveness of FED-PRIME, with faster convergence and improved stability across rounds.", + "bbox": [ + 511, + 297, + 906, + 405 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Prompt Convergence. Fig. 6 shows that FED-PRIME reduces the distance between the aggregated prompt centroids much faster than FEDAVG-P, which highlights the faster convergence rate of FED-PRIME over FEDAVG-P, indicating faster convergence and better prompt alignment. In contrast, FEDAVG-P converges more slowly and unstably, likely due to incorrect default prompt alignment. This again underscores the necessity of having a learnable prompt clustering mechanism to prevent incorrect prompt alignment.", + "bbox": [ + 511, + 419, + 908, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 607, + 633, + 622 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fine-tuning large pre-trained models for specific tasks has become an increasingly effective solution paradigm in machine learning. However, existing fine-tuning frameworks struggle in practical, multimodal settings where data is distributed across private silos with heterogeneous patterns of missing modalities and features. To address these challenges, this paper introduces a novel federated prompt-tuning approach that optimizes, aligns, and aggregates prompt instructions, compensating for gaps in data modalities and distributional patterns across a number of diverse and private local datasets. Rigorous evaluation on multimodal benchmark datasets demonstrates the effectiveness and robustness of our approach against various patterns of missing data in local datasets. We view this work as an initial step toward developing resilient, adaptable federated models that can fully leverage the diversity of multimodal data in real-world environments.", + "bbox": [ + 511, + 643, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3943", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This research is funded by Hanoi University of Science and Technology (HUST) under grant number T2024-TD-002. Additionally, this work utilized GPU compute resource at SDSC and ACES through allocation CIS230391 from the Advanced Cyberinfrastructure Coordination Ecosystem: Services and Support (ACCESS) program [2], which is supported by U.S. National Science Foundation grants #2138259, #2138286, #2138307, #2137603, and #2138296.", + "bbox": [ + 89, + 113, + 483, + 239 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 265, + 187, + 280 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] John Arevalo, Thamar Solorio, Manuel Montes y Gómez, and Fabio A. González. Gated multimodal units for information fusion. ArXiv, abs/1702.01992, 2017. 2, 6, 3", + "[2] Timothy J. Boerner, Stephen Deems, Thomas R. Furlani, Shelley L. Knuth, and John Towns. Access: Advancing innovation: Nsf's advanced cyberinfrastructure coordination ecosystem: Services & support. In Practice and Experience in Advanced Research Computing 2023: Computing for the Common Good, page 173-176, New York, NY, USA, 2023. Association for Computing Machinery. 9", + "[3] Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S. Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, Erik Brynjolfsson, S. Buch, Dallas Card, Rodrigo Castellon, Niladri S. Chatterji, Annie S. Chen, Kathleen A. Creel, Jared Davis, Dora Demszky, Chris Donahue, Moussa Doumbouya, Esin Durmus, Stefano Ermon, John Etchemendy, Kawin Ethayarajh, Li Fei-Fei, Chelsea Finn, Trevor Gale, Lauren E. Gillespie, Karan Goel, Noah D. Goodman, Shelby Grossman, Neel Guha, Tatsunori Hashimoto, Peter Henderson, John Hewitt, Daniel E. Ho, Jenny Hong, Kyle Hsu, Jing Huang, Thomas F. Icard, Saahil Jain, Dan Jurafsky, Pratyusha Kalluri, Siddharth Karamcheti, Geoff Keeling, Fereshte Khani, O. Khattab, Pang Wei Koh, Mark S. Krass, Ranjay Krishna, Rohith Kuditipudi, Ananya Kumar, Faisal Ladhak, Mina Lee, Tony Lee, Jure Leskovec, Isabelle Levent, Xiang Lisa Li, Xuechen Li, Tengyu Ma, Ali Malik, Christopher D. Manning, Suvir P. Mirchandani, Eric Mitchell, Zanele Munyikwa, Suraj Nair, Avanika Narayan, Deepak Narayanan, Benjamin Newman, Allen Nie, Juan Carlos Niebles, Hamed Nilforoshan, J. F. Nyarko, Giray Ogut, Laurel Orr, Isabel Papadimitriou, Joon Sung Park, Chris Piech, Eva Portelance, Christopher Potts, Aditi Raghunathan, Robert Reich, Hongyu Ren, Frieda Rong, Yusuf H. Roohani, Camilo Ruiz, Jack Ryan, Christopher R'e, Dorsa Sadigh, Shiori Sagawa, Keshav Santhanam, Andy Shih, Krishna Parasuram Srinivasan, Alex Tamkin, Rohan Taori, Armin W. Thomas, Florian Tramér, Rose E. Wang, William Wang, Bohan Wu, Jiajun Wu, Yuhuai Wu, Sang Michael Xie, Michihiro Yasunaga, Jiaxuan You, Matei A. Zaharia, Michael Zhang, Tianyi Zhang, Xikun Zhang, Yuhui Zhang, Lucia Zheng, Kaitlyn Zhou, and Percy Liang. On the opportunities and risks of foundation models. ArXiv, 2021. 1", + "[4] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool." + ], + "bbox": [ + 99, + 290, + 486, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Food-101 - mining discriminative components with random forests. In European Conference on Computer Vision, 2014. 6", + "[5] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. CoRR, abs/2005.14165, 2020. 1", + "[6] Alberto Brunete, Ernesto Gambao, Miguel Hernando, and Raquel Cedazo. Smart assistive architecture for the integration of IoT devices, robotic systems, and multimodal interfaces in healthcare environments. Sensors, 21(6):2212, 2021. 1, 2", + "[7] Jiayi Chen and Aidong Zhang. Hgmf: heterogeneous graph-based fusion for multimodal data with incompleteness. In Proceedings of the 26th ACM International Conference on Knowledge Discovery & Data Mining, pages 1295-1305, 2020. 3", + "[8] Jiayi Chen and Aidong Zhang. FedMSplit: Correlation-adaptive federated multi-task learning across multimodal split networks. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 87-96, 2022. 1, 3, 6", + "[9] Jinyu Chen, Wenchao Xu, Song Guo, Junxiao Wang, Jie Zhang, and Haozhao Wang. Fedtune: A deep dive into efficient federated fine-tuning with pre-trained transformers, 2022. 1", + "[10] Yae Jee Cho, Andre Manoel, Gauri Joshi, Robert Sim, and Dimitrios Dimitriadis. Heterogeneous ensemble knowledge transfer for training large models in Federated Learning. Computing Research Repository arXiv Preprints, 2204.12703, 2022. 2", + "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 1", + "[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota, 2019. Association for Computational Linguistics. 1", + "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 1", + "[14] Alireza Fallah, Aryan Mokhtari, and Asuman Ozdaglar. Personalized federated learning: A meta-learning ap" + ], + "bbox": [ + 514, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3944", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "proach. Computing Research Repository arXiv Preprints, 2002.07948, 2020.2", + "[15] Ziwei Fan, Hao Ding, Anoop Deoras, and Trong Nghia Hoang. Personalized federated domain adaptation for item-to-item recommendation. In Proceedings of the Thirty-Ninth Conference on Uncertainty in Artificial Intelligence, pages 560-570. PMLR, 2023. 2", + "[16] Tao Guo, Song Guo, Junxiao Wang, Xueyang Tang, and Wenchao Xu. PromptFL: Let Federated Participants Cooperatively Learn Prompts Instead of Models-Federated Learning in Age of Foundation Model. IEEE Transactions on Mobile Computing, 2023. 1", + "[17] Minh Hoang, Nghia Hoang, Bryan Kian Hsiang Low, and Carleton Kingsford. Collective model fusion for multiple black-box experts. In Proceedings of the 36th International Conference on Machine Learning, pages 2742-2750. PMLR, 2019. 2", + "[18] Trong Nghia Hoang, Quang Minh Hoang, Kian Hsiang Low, and Jonathan P. How. Collective online learning of Gaussian processes in massive multi-agent systems. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7850-7857, 2019. 2", + "[19] Parminder Kaur, Husanbir Singh Pannu, and Avleen Kaur Malhi. Comparative analysis on cross-modal information retrieval: A review. Computer Science Review, 39:100336, 2021. 2", + "[20] Muhammad Uzair khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3", + "[21] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 5583-5594. PMLR, 2021. 1, 3, 6", + "[22] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A. Shamma, Michael S. Bernstein, and Li Fei-Fei. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International Journal of Computer Vision, 123:32 - 73, 2016. 1", + "[23] Harold W Kuhn. The Hungarian Method for the Assignment Problem. Naval Research Logistics Quarterly, 2(1-2):83-97, 1955. 6", + "[24] Yi-Lun Lee, Yi-Hsuan Tsai, Wei-Chen Chiu, and Chen-Yu Lee. Multimodal prompting with missing modalities for visual recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3, 6, 7, 1", + "[25] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3045-3059, Online and Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 1", + "[26] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. Proceedings of Machine Learning and Systems, 2:429-450, 2020. 2, 4" + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Xiang Li, Kaixuan Huang, Wenhao Yang, Shusen Wang, and Zhihua Zhang. On the convergence of FedAvg on non-IID data. Computing Research Repository arXiv Preprints, 1907.02189t, 2019. 2", + "[28] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, 2014. 1", + "[29] Wang Lu, Xixu Hu, Jindong Wang, and Xing Xie. FedCLIP: Fast Generalization and Personalization for CLIP in Federated Learning, 2023. 1", + "[30] Tengfei Ma, Trong Nghia Hoang, and Jie Chen. Federated learning of models pre-trained on different features with consensus graphs. In Uncertainty in Artificial Intelligence, pages 1336-1346, 2023. 2", + "[31] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial Intelligence and Statistics, pages 1273-1282, 2017. 2, 3, 4, 5", + "[32] John Nguyen, Jianyu Wang, Kshitiz Malik, Maziar Sanjabi, and Michael Rabbat. Where to Begin? On the Impact of Pre-Training and Initialization in Federated Learning, 2023. 1", + "[33] Manh Duong Nguyen, Trung Thanh Nguyen, Huy-Hieu Pham, Trong Nghia Hoang, Phi Le Nguyen, and Thanh Trung Huynh. Fedmac: Tackling partial-modality missing in federated learning with cross-modal aggregation and contrastive regularization. 2024. 1, 3", + "[34] Nang Hung Nguyen, Phi Le Nguyen, Thuy Dung Nguyen, Trung Thanh Nguyen, Duc Long Nguyen, Thanh Hung Nguyen, Huy Hieu Pham, and Thao Nguyen Truong. FedDRL: Deep reinforcement learning-based adaptive aggregation for non-IID data in federated learning. In Proceedings of the 51st International Conference on Parallel Processing, pages 1-11, 2022. 2", + "[35] Nang Hung Nguyen, Duc Long Nguyen, Trong Bang Nguyen, Thanh-Hung Nguyen, Huy Hieu Pham, Truong Thao Nguyen, and Phi Le Nguyen. CADIS: Handling cluster-skewed Non-IID data in federated learning with clustered aggregation and knowledge distilled regularization. In Proceedings of the 23rd IEEE/ACM International Symposium on Cluster, Cloud and Internet Computing, pages 249-261, 2023.", + "[36] Quang Ha Pham, Nang Hung Nguyen, Thanh Hung Nguyen, Huy Hieu Pham, Phi Le Nguyen, and Truong Thao Nguyen. SEM: A simple yet efficient model-agnostic local training mechanism to tackle data sparsity and scarcity in federated learning. In Proceedings of the 11st International Symposium on Computing and Networking, pages 120–126, 2023. 2", + "[37] Thu Hang Phung, Binh P Nguyen, Thanh Hung Nguyen, Quoc Viet Hung Nguyen, Phi Le Nguyen, and Thanh Trung Huynh. A contrastive learning and graph-based approach for missing modalities in multimodal federated learning. In Proceedings of the 2024 International Joint Conference on Neural Networks, pages 1-8, 2024. 1, 3" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3945", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Petra Poklukar, Miguel Vasco, Hang Yin, Francisco S Melo, Ana Paiva, and Danica Kragic. Geometric multimodal contrastive representation learning. In Proceedings of the 2022 International Conference on Machine Learning, pages 17782-17800, 2022. 3", + "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. CoRR, abs/2103.00020, 2021. 1", + "[40] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21(140):1-67, 2020. 1", + "[41] Kuniaki Saito, Kihyuk Sohn, Xiang Zhang, Chun-Liang Li, Chen-Yu Lee, Kate Saenko, and Tomas Pfister. Prefix conditioning unifies language and label supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2861-2870, 2023. 1, 3", + "[42] Rachael Hwee Ling Sim, Yehong Zhang, Trong Nghia Hoang, Xinyi Xu, Bryan Kian Hsiang Low, and Patrick Jaillet. Incentives in private collaborative machine learning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2", + "[43] Canh T Dinh, Nguyen Tran, and Josh Nguyen. Personalized federated learning with Moreau envelopes. Advances in Neural Information Processing Systems, 33:21394-21405, 2020. 2", + "[44] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, S. M. Ali Eslami, Oriol Vinyals, and Felix Hill. Multimodal few-shot learning with frozen language models. In Advances in Neural Information Processing Systems, pages 200-212. Curran Associates, Inc., 2021. 1, 3", + "[45] Feng Wang, M Cenk Gursoy, and Senem Velipasalar. Communication-efficient and privacy-preserving feature-based federated transfer learning. In Proceedings of the 2022 IEEE Global Communications Conference, pages 3875-3880, 2022. 2", + "[46] Xin Wang, Devinder Kumar, Nicolas Thome, Matthieu Cord, and Frédéric Precioso. Recipe recognition with large multimodal food dataset. In 2015 IEEE International Conference on Multimedia & Expo Workshops (ICMEW), pages 1-6, 2015. 2, 6, 3", + "[47] Pei-Yau Weng, Minh Hoang, Lam M. Nguyen, My T. Thai, Tsui-Wei Weng, and Trong Nghia Hoang. Probabilistic federated prompt-tuning with non-IID and imbalanced data. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 1", + "[48] Baochen Xiong, Xiaoshan Yang, Fan Qi, and Changsheng Xu. A unified framework for multi-modal Federated Learning. Neurocomputing, 480:110-118, 2022. 2", + "[49] Baochen Xiong, Xiaoshan Yang, Yaguang Song, Yaowei Wang, and Changsheng Xu. Client-adaptive cross-model reconstruction network for modality-incomplete multimodal federated learning. In Proceedings of the 31st ACM Interna-" + ], + "bbox": [ + 91, + 90, + 485, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tional Conference on Multimedia, pages 1241-1249, 2023. 3", + "[50] Fu-En Yang, Chien-Yi Wang, and Yu-Chiang Frank Wang. Efficient Model Personalization in Federated Learning via Client-Specific Prompt Generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 19159-19168, 2023. 1", + "[51] Guan Yu, Quefeng Li, Dinggang Shen, and Yufeng Liu. Optimal sparse linear prediction for block-missing multimodality data without imputation. Journal of the American Statistical Association, 115(531):1406-1419, 2020. 3", + "[52] Songcan Yu, Junbo Wang, Walid Hussein, and Patrick C.K. Hung. Robust multimodal federated learning for incomplete modalities. Computer Communications, 214:234-243, 2024. 1, 3", + "[53] Mikhail Yurochkin, Mayank Agarwal, Soumya Ghosh, Kristjan Greenewald, Trong Nghia Hoang, and Yasaman Khazaeni. Bayesian nonparametric federated learning of neural networks. In International Conference on Machine Learning, 2019. 2", + "[54] Mikhail Yurochkin, Mayank Argawal, Soumya Ghosh, Kristjan Greenewald, and Trong Nghia Hoang. Statistical model aggregation via parameter matching. In Advances in Neural Information Processing Systems, pages 10954-10964, 2019. 2", + "[55] Changqing Zhang, Yajie Cui, Zongbo Han, Joey Tianyi Zhou, Huazhu Fu, and Qinghua Hu. Deep partial multi-view learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(5):2402-2415, 2020. 3", + "[56] Jianyi Zhang, Hao Frank Yang, Ang Li, Xin Guo, Pu Wang, Haiming Wang, Yiran Chen, and Hai Li. Mllm-fl: Multimodal large language model assisted federated learning on heterogeneous and long-tailed data. arXiv preprint arXiv:2409.06067, 2024. 1, 2", + "[57] Wei Zhang and Xiang Li. Data privacy preserving federated transfer learning in machinery fault diagnostics using prior distributions. Structural Health Monitoring, 21(4):1329-1344, 2022. 2", + "[58] Haodong Zhao, Wei Du, Fangqi Li, Peixuan Li, and Gongshen Liu. FedPrompt: Communication-Efficient and Privacy Preserving Prompt Tuning in Federated Learning. In Proc. ICASSP, 2023. 1", + "[59] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 3", + "[60] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision (IJCV), 2022. 1, 3", + "[61] Tongxue Zhou, Pierre Vera, Stephane Canu, and Su Ruan. Missing data imputation via conditional generator and correlation learning for multimodal brain tumor segmentation. Pattern Recognition Letters, 158:125-132, 2022. 3" + ], + "bbox": [ + 516, + 92, + 903, + 840 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "3946", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Federated Prompt-Tuning with Heterogeneous and Incomplete Multimodal Client Data/c4864325-3bcd-4609-87ea-0a89ff648c25_model.json b/2025/Federated Prompt-Tuning with Heterogeneous and Incomplete Multimodal Client Data/c4864325-3bcd-4609-87ea-0a89ff648c25_model.json new file mode 100644 index 0000000000000000000000000000000000000000..937f45fe5cbb8673a75b2236ff93dd9131d2e356 --- /dev/null +++ b/2025/Federated Prompt-Tuning with Heterogeneous and Incomplete Multimodal Client Data/c4864325-3bcd-4609-87ea-0a89ff648c25_model.json @@ -0,0 +1,2301 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.183, + 0.131, + 0.819, + 0.175 + ], + "angle": 0, + "content": "Federated Prompt-Tuning with Heterogeneous and Incomplete Multimodal Client Data" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.203, + 0.877, + 0.277 + ], + "angle": 0, + "content": "Thu Hang Phung1 Duong M. Nguyen1 Thanh Trung Huynh2 Quoc Viet Hung Nguyen3 \nTrong Nghia Hoang4* Phi Le Nguyen1* \n1 Institute of AI Innovation and Societal Impact, Hanoi University of Science and Technology \n2 EPFL 3 Griffith University 4 Washington State University" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.341, + 0.486, + 0.614 + ], + "angle": 0, + "content": "This paper introduces a generalized federated prompt-tuning framework for practical scenarios where local datasets are multi-modal and exhibit different distributional patterns of missing features at the input level. The proposed framework bridges the gap between federated learning and multi-modal prompt-tuning which have traditionally focused on either uni-modal or centralized data. A key challenge in this setting arises from the lack of semantic alignment between prompt instructions that encode similar distributional patterns of missing data across different clients. To address this, our framework introduces specialized client-tuning and server-aggregation designs that simultaneously optimize, align, and aggregate prompt-tuning instructions across clients and data modalities. This allows prompt instructions to complement one another and be combined effectively. Extensive evaluations on diverse multimodal benchmark datasets demonstrate that our work consistently outperforms state-of-the-art (SOTA) baselines." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.627, + 0.222, + 0.643 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.652, + 0.486, + 0.879 + ], + "angle": 0, + "content": "The recent emergence of exceedingly large models [3] with immense scale and versatility due to being pre-trained on large swaths of generic data has resulted in a paradigm shift in machine learning (ML). Most task-specific models are now fine-tuned versions of those large models instead of being created from de novo learning processes. For example, most state-of-the-art (SOTA) models solving specific NLP tasks are fine-tuned versions of existing large pre-trained or foundation models such as BERT [12], GPT [5], and T5 [40]. Likewise, SOTA models for various computer vision tasks have also been derived from foundation models such as the vision transformer [13] pre-trained on extensive datasets like ImageNet [11]. The standard fine-tuning approach, however, assumes centralized data which is not always possible in practical scenarios where fine-tuning data" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.311, + 0.855, + 0.326 + ], + "angle": 0, + "content": "are scattered across multiple private siloed systems." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.327, + 0.907, + 0.522 + ], + "angle": 0, + "content": "To overcome this siloed data challenge, federated fin-tuning has recently been considered to aggregate the fin-tuned updates across private devices running their own (local) fine-tuning processes [9, 32, 47]. This approach can be further combined with an existing parameter-efficient tuning method called prompt tuning [25]. This method focuses on creating lightweight signals such as additional tokens that are prepended to the input of the pre-trained models [41, 44, 59, 60], providing them valuable contextual information to perform specific tasks more effectively. Local sets of prompts can then be aggregated using the conventional parameter aggregation procedure in FL [16, 29, 47, 50, 58]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.909, + 0.704 + ], + "angle": 0, + "content": "However, such federated prompt-tuning techniques are restricted to learning with uni-modal data. In more practical scenarios, advances in sensor and device manufacturing increasingly enable the capture of multiple data modalities (e.g., text, images), leading to richer datasets for fine-tuning [6]. Leveraging such diverse multi-modal information can significantly enhance model performance, as demonstrated by the success of multi-modal foundation models like CLIP [39] and ViLT [21]. This has further motivated research on multi-modal federated learning [8, 33, 37, 52] and fine-tuning pre-trained multi-modal LLMs, i.e., multi-modal federated fine-tuning [56]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Despite these advances, multi-modal federated learning research often adopts specific local architectures which cannot be learned via fine-tuning existing pre-trained multimodal foundation models [8, 33, 37, 52]. Existing multimodal federated fine-tuning research on the other hand assumes that all local datasets have the same set of modalities [56]. As such, these approaches are not suitable for situations where different clients might collect data using different types of sensors. Furthermore, the multi-modal sensor data across clients might even have different missing patterns of data entries. This leads to local datasets with both intra- and inter-heterogeneities. Intra-heterogeneity occurs when individual samples within a single dataset" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.245, + 0.901 + ], + "angle": 0, + "content": "*Corresponding Authors" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "3936" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.137 + ], + "angle": 0, + "content": "have missing entries for different subsets of features. Interheterogeneity arises when different clients have different distributions over their data-missing patterns." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.139, + 0.483, + 0.351 + ], + "angle": 0, + "content": "There also exists recent work in prompt-tuning that learns a specific prompt for each subset of missing modalities [24]. Input samples having the same set of missing modalities are then augmented with the same (learnable) prompt during training, specializing it to complement for the corresponding missing modalities. This approach is, however, restricted to centralized data setting (i.e., no interheterogeneities) and does not generalize well to federated settings where different sets of local prompts might be biased towards different distributional patterns of missing data. A naive aggregation of these prompts might therefore collapse them into less informative prompts which are ignorant of the heterogeneities in missing-data patterns across clients, resulting in poor performance (see Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.354, + 0.483, + 0.535 + ], + "angle": 0, + "content": "To bridge this gap between multi-modal prompt-tuning and federated learning, we propose a multimodal federated prompt-tuning approach that optimizes and aggregates two distinct sets of (local) prompts for each client: inter-client and intra-client prompts. Inter-client prompts breakdown and encode prototypical distribution patterns of missing input data, which can be aligned and aggregated across clients to represent shared distributional information. In contrast, intra-client prompts encode information tailored to support specific local subsets of missing modalities, independent of the distributional patterns of missing data at each client. This is achieved via the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.541, + 0.483, + 0.708 + ], + "angle": 0, + "content": "1. We develop a local input-adaptive fine-tuning algorithm which identifies and updates a subset of inter- and intra-client prompts that are most relevant to each individual sample. This is formulated as prompt-retrieval task which focuses on learning mapping functions that push an input towards its relevant prompts on an embedding space while pulling away from irrelevant prompts on an embedding space. This helps decompose and distill local fine-tuning knowledge into different parts of the prompt set, preventing learned instructions for different fine-tuning patterns from being overloaded on the same prompts (Section 3.2)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.713, + 0.483, + 0.849 + ], + "angle": 0, + "content": "2. We develop a server-side model for aggregating and aligning inter-client prompts across clients. Both generation and alignment are framed as part of a clustering task, where inter-client prompts encoding similar input-level distribution patterns of missing data are identified and merged into more comprehensive and versatile prompt instructions. The empirical effectiveness of these combined prompts in reducing local training losses is used as a learning feedback to optimize this clustering model (Section 3.3)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "3. We evaluate and demonstrate the effectiveness of our proposed approach through extensive experiments on two benchmark datasets, MM-IMDB [1] and UPMC Food" + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.541, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.091, + 0.885, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.27, + 0.907, + 0.312 + ], + "angle": 0, + "content": "Figure 1. Our approach with prompt alignment outperforms FedAvg prompt-tuning (w/o alignment) across settings with different data missing rates." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.318, + 0.907, + 0.438 + ], + "angle": 0, + "content": "101 [46], comparing its performance against a variety of existing multimodal federated learning methods as well as repurposed prompt-tuning baselines for federated settings. The reported results show that our method consistently and significantly outperforms the baselines across various modality-missing scenarios, achieving new SOTA performance for multi-modal federated learning (Section 4)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.44, + 0.907, + 0.485 + ], + "angle": 0, + "content": "A bird-view of our proposed framework, Federated Prompt-Tuning with Heterogeneous and Incomplete Multimodal Client Data (FED-PRIME), is also provided in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.497, + 0.665, + 0.513 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.523, + 0.907, + 0.809 + ], + "angle": 0, + "content": "Multi-Modal Federated Learning. Federated learning (FL) [17, 31, 53, 54] is a collaborative learning paradigm that enables multiple devices to train a shared model while keeping local data private. This design addresses concerns over privacy, security, and inefficiency in transferring large datasets [45, 57]. While FL has been widely studied [10, 14, 15, 18, 26, 27, 30, 34–36, 42, 43], most existing methods focus on uni-modal data (e.g., image or text). However, with the rapid advancement of mobile phones and IoT devices [6], local data increasingly comes from multiple modalities. Integrating such heterogeneous signals enables richer, more robust representations [19, 48], motivating growing interest in multi-modal FL. Yet many current approaches rely on custom architectures that cannot exploit pre-trained multi-modal models. More recent studies explore federated fine-tuning of pre-trained multi-modal models, but they typically assume each client holds the same set of modalities [56], limiting applicability in realistic heterogeneous environments." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Missing Data in Multi-Modal FL. In real-world settings, local devices collect data from different types of sensors, leading to datasets with varying feature modalities (inter-heterogeneity). Moreover, sensor malfunctions or data corruption can cause missing features, with devices exhibiting distinct missing patterns (intra-heterogeneity)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3937" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.288 + ], + "angle": 0, + "content": "Existing solutions either treat each modality independently [55, 61] or are restricted to centralized learning [7, 38, 51], both overlooking inter-modal dependencies. Recent FL methods address inter-heterogeneity [8, 37] but not intra-heterogeneity, while approaches like FedInMM [52] and FedMAC [33] tackle intra-client missing data but ignore cross-client alignment, limiting effectiveness. A related line of work, including FedMSplit [8] and CACMRN [49], handles modality-incomplete data under a personalized FL framework, where each client retains its model. This paradigm differs from standard FL, which aims to improve a single shared model, making these methods not directly comparable." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.29, + 0.485, + 0.501 + ], + "angle": 0, + "content": "Addressing Missing Data with Prompt-Tuning. Prompt-tuning [41, 44, 59, 60] provides an effective way to adapt pre-trained models by learning prompts that add contextual information to input embeddings. Recent work has extended this idea to missing-modality scenarios by learning a prompt for each subset of missing features [20, 24], so that inputs with similar missing patterns are augmented with suitable prompts. However, these methods are developed for centralized training and do not address the additional heterogeneity introduced in FL, where prompts must be aligned and aggregated across clients. Designing effective alignment and aggregation strategies for such prompts is therefore a key challenge, as demonstrated in our framework." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.511, + 0.449, + 0.53 + ], + "angle": 0, + "content": "3. Multi-Modal Federated Prompt-Tuning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.484, + 0.842 + ], + "angle": 0, + "content": "Multimodal federated learning faces significant challenges due to missing data modalities, leading to both inter-client and intra-client heterogeneities that complicate knowledge aggregation and degrade overall model performance - see Fig. 1. To address this challenge (Section 3.1), we present a federated prompt-tuning framework based on the pretrained Vision and Language Transformer (ViLT) model, which comprises a prompt-based client design and a server-side prompt-aggregation algorithm. The client design features a fine-tuning algorithm that learns and decomposes the fine-tuned knowledge into separate sets of inter-client and intra-client prompts specialized for fine-tuning at different input-level patterns of missing data and input-agnostic patterns of missing modalities (Section 3.2). The server-side algorithm then learns alignment between inter-client prompts specializing for similar input-level patterns of missing data and aggregates them to create more versatile prompts. This is also combined with a standard FEDAVG-based aggregation of intra-prompts specializing for input-agnostic patterns of missing modalities (Section 3.3) - see Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.847, + 0.295, + 0.862 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "In multimodal federated learning (MMFL), we assume there are \\( n \\) clients and \\( r \\) data modalities (e.g., text, image)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.091, + 0.907, + 0.198 + ], + "angle": 0, + "content": "Each client \\( t \\in [n]^{1} \\) has access to a private dataset \\( D_{t} \\triangleq \\{x(M_{t,s}), z_{t,s}\\}_{s=1}^{m} \\), where \\( x(M_{t,s}) \\) denote the observed features for the set of modalities, \\( M_{t,s} \\subseteq [r] \\) recorded for the \\( s \\)-th data point of the \\( t \\)-th client, and \\( z_{t,s} \\) denotes the corresponding target label. The goal is to learn a multimodal model \\( w_{*} \\) using all the private datasets \\( \\{D_{t}\\}_{t=1}^{n} \\) without centralizing them. That is," + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.203, + 0.878, + 0.244 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {w} _ {*} = \\underset {\\boldsymbol {w}} {\\arg \\min } \\left\\{L (\\boldsymbol {w}) \\triangleq \\frac {1}{n} \\sum_ {t = 1} ^ {n} L _ {t} (\\boldsymbol {w}) \\right\\} \\text {w i t h}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.539, + 0.246, + 0.907, + 0.285 + ], + "angle": 0, + "content": "\\[\nL _ {t} (\\boldsymbol {w}) \\triangleq \\frac {1}{n} \\sum_ {s = 1} ^ {m} \\ell \\left(F \\left(\\boldsymbol {x} \\left(M _ {t, s}\\right); \\boldsymbol {w}\\right), z _ {t, s}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.292, + 0.907, + 0.414 + ], + "angle": 0, + "content": "where \\( F(\\pmb{x}(M_{t,s}); \\pmb{w}) \\) denotes the multimodal prediction model with parameter \\( \\pmb{w} \\) and \\( \\ell(F(\\pmb{x}(M_{t,s}); \\pmb{w}), z_{t,s}) \\) denotes the individual loss of \\( \\pmb{w} \\) on the \\( s \\)-th data point. Previous work [8, 33, 37, 52] in MMFL has largely focused on designing \\( F(\\pmb{x}(M_{t,s}); \\pmb{w}) \\) without leveraging existing knowledge in pre-trained foundation model. Instead, we let \\( F(\\pmb{x}(M_{t,s}); \\pmb{w}) \\) be a prompt-tuned version of a pre-trained transformer-based model," + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.418, + 0.906, + 0.445 + ], + "angle": 0, + "content": "\\[\nF (\\boldsymbol {x} (M); \\boldsymbol {w}) \\triangleq F _ {c} \\left(F _ {p} \\left(F _ {e} (\\boldsymbol {x} (M)) \\circ \\boldsymbol {w} _ {p}\\right); \\boldsymbol {w} _ {c}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.449, + 0.907, + 0.646 + ], + "angle": 0, + "content": "where \\( F_{c}(:,\\boldsymbol{w}_{c}) \\) is the prediction head parameterized with \\( \\boldsymbol{w}_{c} \\), \\( F_{e}(\\boldsymbol{x}(M)) = F_{e}(\\boldsymbol{x}_{1}) \\circ \\dots \\circ F_{e}(\\boldsymbol{x}_{|M|}) \\) where \\( F_{e}(\\boldsymbol{x}_{i}) \\) denote the tokenized input sequence of each unimodal feature \\( \\boldsymbol{x}_{i} \\) in \\( \\boldsymbol{x}(M) \\triangleq (\\boldsymbol{x}_{1}, \\boldsymbol{x}_{2}, \\dots, \\boldsymbol{x}_{|M|}) \\), \\( F_{p}() \\) denotes the frozen multimodal pre-trained model – e.g., ViLT [21] – and \\( \\boldsymbol{w}_{p} \\) denotes the set of (learnable) prompts. Here, \\( \\circ \\) denotes the set concatenation operator and \\( \\boldsymbol{w} = (\\boldsymbol{w}_{c}, \\boldsymbol{w}_{p}) \\) denotes the full set of tuning parameters. Due to the decentralized and private nature of the local datasets \\( \\{D_{t}\\}_{t=1}^{n} \\), solving Eq. (1) directly is not possible. Instead, following the federated learning [31] paradigm, its optimal solution \\( \\boldsymbol{w}_{*} \\) is approximated via a number of iterations over two interleaving client- and server-side steps:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.652, + 0.907, + 0.678 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {w} _ {t} ^ {(h + 1)} = \\operatorname {u p d a t e} \\left(L _ {t}, \\boldsymbol {w} _ {*} ^ {(h)}\\right) \\forall t \\in [ n ], \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.679, + 0.906, + 0.706 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {w} _ {*} ^ {(h + 1)} = \\text {a g g r e g a t e} \\left(\\boldsymbol {w} _ {1} ^ {(h + 1)}, \\dots , \\boldsymbol {w} _ {n} ^ {(h + 1)}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.712, + 0.907, + 0.88 + ], + "angle": 0, + "content": "where \\(\\boldsymbol{w}_t^{(h)}\\) denote the estimation of the local tuning parameter \\(\\boldsymbol{w}_t\\) of client \\(t\\) after \\(h\\) iterations. For example, the update step can be a vanilla gradient descent (GD) optimization of the prompt \\(\\omega\\) and the aggregate step denotes the average of the locally optimized prompts. Such vanilla approach however does not account for both the inter- and intra-heterogeneities among local prompts (see Section 1) which ultimately risks aggregating incompatible prompts and degrade the overall performance (see Section 4). To avoid this, we will introduce novel designs for client- and server-prompts as detailed in the followings." + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.886, + 0.655, + 0.902 + ], + "angle": 0, + "content": "\\(^{1}[n]\\triangleq \\{1,2,\\dots ,n\\}\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3938" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.087, + 0.912, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.378, + 0.908, + 0.422 + ], + "angle": 0, + "content": "Figure 2. Overview of the proposed multi-modal federated prompt-tuning framework - see Alg. 3. Each client maintains two (learnable) sets of intra- and inter-client prompts. At the beginning of each iteration, each client performs local training via Eq. (7) and Eq. (8). Local sets of intra- and inter-client prompts are then sent to the server for aggregation - see Eq. (10)-Eq. (14)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.431, + 0.232, + 0.449 + ], + "angle": 0, + "content": "3.2. Client Design" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.454, + 0.484, + 0.545 + ], + "angle": 0, + "content": "To enable effective prompt-aggregation across heterogeneous clients, we need to separate the fine-tuning treatment for input-level patterns of missing data and input-agnostic patterns of missing modalities. This can be achieved via further parameterizing \\(\\boldsymbol{w}_p = (\\boldsymbol{w}_p^{inter}, \\boldsymbol{w}_p^{intra})\\) with two distinct subsets of \\(\\tau\\) inter- and intra-client prompts:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.553, + 0.483, + 0.579 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {w} _ {p} ^ {\\text {i n t e r}} \\triangleq \\left\\{\\boldsymbol {p} _ {1} ^ {\\text {i n t e r}}, \\boldsymbol {p} _ {2} ^ {\\text {i n t e r}}, \\dots , \\boldsymbol {p} _ {\\tau} ^ {\\text {i n t e r}} \\right\\}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.58, + 0.483, + 0.607 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {w} _ {p} ^ {\\text {i n t r a}} \\triangleq \\left\\{\\boldsymbol {p} _ {1} ^ {\\text {i n t r a}}, \\boldsymbol {p} _ {2} ^ {\\text {i n t r a}}, \\dots , \\boldsymbol {p} _ {\\tau} ^ {\\text {i n t r a}} \\right\\}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.827 + ], + "angle": 0, + "content": "which are learned separately. The learned inter-client prompts \\(\\pmb{w}_{p}^{inter}\\) across clients will be sent to the server which clusters them into different groups based on their specialization for similar input-level patterns of missing data - Section 3.3. Local sets of intra-client prompts \\(\\pmb{w}_{p}^{intra}\\) can be averaged via FEDAVG [31] since their specialization is for input-agnostic patterns of missing modalities. Intuitively, the aggregation mechanisms will determine how different aspects of fine-tuning knowledge are encoded across the prompt sets. For example, if knowledge regarding different input-level patterns of missing data is (incorrectly) incorporated into \\(\\{\\pmb{p}_i^{intra}\\}_{i=1}^\\tau\\), they will be averaged out and collapsed into less informative prompts due to the non-clustering aggregation mechanism of \\(\\{\\pmb{p}_i^{intra}\\}_{i=1}^\\tau\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.827, + 0.484, + 0.902 + ], + "angle": 0, + "content": "As this potential loss of information will decrease the overall performance, the (learnable) prompt parameters will be updated to prevent this from happening. Otherwise, if (common) knowledge patterns regarding input-agnostic patterns of missing modalities are (incorrectly) incorporated" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.907, + 0.538 + ], + "angle": 0, + "content": "into \\(\\left\\{\\pmb{p}_i^{inter}\\right\\}_{i=1}^{\\tau}\\), their modeling bandwidth for inter-client heterogeneities will be reduced, which will also hamper the performance. This will, in turn, generate a prompt update signal to avoid this as well. Furthermore, within each set of inter- and intra-client prompts, there might also be different patterns that need to be separated and internalized into different subset of prompts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.909, + 0.75 + ], + "angle": 0, + "content": "To enable this, we develop an input-adaptive learning mechanism for both sets of prompts which identifies and updates the most relevant subset of prompts for each (multimodal) input sample. This is formulated as an information-retrieval task, learning a pair of key and query functions, \\( k(\\pmb{p}) \\) and \\( q(\\pmb{x}(M)) \\), which transform the prompt \\( \\pmb{p} \\) and input \\( \\pmb{x} \\) with the set \\( M \\) observed modalities, respectively, into a shared metric space. We can then define a geometric input-prompt distance, e.g., \\( d(\\pmb{x}(M), \\pmb{p}) = \\cos(q(\\pmb{x}(M)), k(\\pmb{p})) \\) which can be used to measure their contextual relevance. To avoid loading diverse fine-tuning knowledge into a single prompt, we regularize the (local) loss function to penalize it when a less relevant prompt is selected as fine-tuning instruction for an input \\( \\pmb{x}(M) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.754, + 0.907, + 0.78 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {w} _ {t} \\triangleq \\underset {\\boldsymbol {w}} {\\arg \\min } L _ {t} ^ {\\prime} (\\boldsymbol {w}) \\quad \\text {w h e r e} \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.782, + 0.905, + 0.867 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {t} ^ {\\prime} (\\boldsymbol {w}) \\triangleq \\sum_ {s = 1} ^ {m} \\left\\{\\ell \\left(F \\left(\\boldsymbol {x} \\left(M _ {t, s}\\right); \\boldsymbol {w} ^ {\\prime}\\right), z _ {t, s}\\right) \\right\\} \\\\ + \\sum_ {s = 1} ^ {m} \\left\\{r \\left(\\boldsymbol {x} \\left(M _ {t, s}\\right), \\boldsymbol {w} _ {p} ^ {\\prime}\\right) \\right\\}, \\tag {8} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.87, + 0.908, + 0.901 + ], + "angle": 0, + "content": "where \\(\\ell (\\pmb {w}^{\\prime})\\) is defined in Eq. (1) and \\(\\pmb{w}^{\\prime} = (\\pmb{w}_{c},\\pmb{w}_{p}^{\\prime})\\) denote the customized set of tuning instructions for each" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3939" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.091, + 0.473, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.276, + 0.485, + 0.361 + ], + "angle": 0, + "content": "Figure 3. Workflow of the prompt-alignment algorithm. At each iteration, each client samples a subset of summarizing prompts using its query and key functions. The client then performs a local update, resulting in heterogeneous prompt sets, which are subsequently sent to the server to be clustered and summarized into new summarizing prompt sets for the next iteration." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.385, + 0.484, + 0.456 + ], + "angle": 0, + "content": "input \\(\\pmb{x}(M_{t,s})\\) where \\(\\pmb{w}_p' = (\\pmb{w}_p^{inter'}, \\pmb{w}_p^{intra'})\\) such that \\(\\pmb{w}_p^{inter'} \\subseteq \\pmb{w}_p^{inter}\\) and \\(\\pmb{w}_p^{intra'} \\subseteq \\pmb{w}_p^{intra}\\) are the sets of \\(\\kappa\\) most relevant prompts to \\(\\pmb{x}(M_{t,s})\\) in \\(\\pmb{w}_p^{inter}\\) and \\(\\pmb{w}_p^{intra}\\). The regularizer \\(r(\\pmb{x}(M), \\pmb{w}_p')\\) is then defined as" + }, + { + "type": "equation", + "bbox": [ + 0.129, + 0.465, + 0.484, + 0.511 + ], + "angle": 0, + "content": "\\[\nr \\left(\\boldsymbol {x} (M), \\boldsymbol {w} _ {p} ^ {\\prime}\\right) \\triangleq \\sum_ {\\boldsymbol {p} \\in \\boldsymbol {w} _ {p} ^ {\\prime}} \\left\\{d (\\boldsymbol {x} (M), \\boldsymbol {p}) \\right\\}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.521, + 0.484, + 0.673 + ], + "angle": 0, + "content": "Intuitively, the regularizer penalizes the total distance between the (multimodal) input \\( \\pmb{x}(M) \\) and the \\( \\kappa \\) closest prompts according to the aforementioned distance metric \\( d(\\pmb{x}(M), \\pmb{p}) \\). Minimizing both the original local loss and this regularizer allows the local model to simultaneously learn to (1) internalize fine-tuning patterns of inter- and intra-heterogeneities to minimize the original loss \\( L_{t}(\\pmb{w}) \\); and (2) ensure that selected prompts for a given input will be close to it to reduce the penalty \\( r(\\pmb{x}(M), \\pmb{w}_{p}^{\\prime}) \\). A pseudocode of this client design is detailed in Suppl. A." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.681, + 0.279, + 0.697 + ], + "angle": 0, + "content": "3.3. Server Aggregation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.703, + 0.483, + 0.809 + ], + "angle": 0, + "content": "We will now establish a principled algorithm to align and aggregate inter-client prompts specializing in similar input-level patterns of missing data2. This is not straightforward since inter-client prompts at the same position in the prompt sets across different clients might not specialize for similar patterns of missing data due to client heterogeneities: certain missing-data patterns are not observed at certain clients." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.809, + 0.483, + 0.856 + ], + "angle": 0, + "content": "To find the correct alignment, we formulate it as a clustering optimization task. Suppose each client \\(t\\) communicates a set of inter-client prompts \\(\\boldsymbol{w}_t^{inter} = (\\boldsymbol{p}_t^i)_{i=1}^\\tau\\). Each" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.214 + ], + "angle": 0, + "content": "potential clustering of these prompts thus defines a candidate alignment, i.e., prompts in the same cluster are considered aligned and can be aggregated. A (learnable) cluster center can then be associated with the aggregated prompt - see Fig. 3. Both the cluster assignment and centers can be found via solving the following optimization task with discrete alignment variables \\(\\alpha_{t}^{pq}\\in \\{0,1\\}\\) and continuous center variables \\(\\theta_{q}\\) for each cluster \\(q\\in [n\\times \\tau ]\\) .." + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.226, + 0.907, + 0.29 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {\\boldsymbol {\\alpha}, \\boldsymbol {\\theta}, \\gamma} G (\\boldsymbol {\\alpha}, \\boldsymbol {\\theta}, \\gamma) \\triangleq \\sum_ {t = 1} ^ {n} \\sum_ {p = 1} ^ {\\tau} \\sum_ {q = 1} ^ {n \\times \\tau} \\alpha_ {t} ^ {p q} \\cdot c \\left(\\boldsymbol {p} _ {t} ^ {p}, \\boldsymbol {\\theta} _ {q},; \\gamma\\right) (10) \\\\ \\mathrm {s . t .} \\alpha_ {t} ^ {p, 1} + \\dots + \\alpha_ {t} ^ {p, n \\times \\tau} = 1, (11) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.299, + 0.907, + 0.527 + ], + "angle": 0, + "content": "where \\(\\alpha \\triangleq (\\alpha_{t}^{p,q})\\) and \\(\\alpha_{t}^{p,q} \\in \\{0,1\\} \\forall (t,p,q)\\) indicates whether the \\(p\\)-th inter-client prompt of client \\(t\\) is matched to the \\(q\\)-th cluster. The corresponding cluster center is then optimized to have minimum total distance to all cluster members according to a (learnable) cost function \\(c(\\pmb{p}_{t}^{p},\\pmb{\\theta}_{q})\\). The optimized cluster center can then be used as the aggregated or summarizing prompt. Minimizing Eq. (10) thus allows us to align and summarize/aggregate the inter-client prompts across clients simultaneously. Note that Eq. (11) enforces a constraint that intuitively indicates two prompts from the same clients should not be matched to the same cluster and get aggregated. This is because local prompts from the same client have already been learned to encode different finetuning aspects. Aggregating them will therefore collapse them into less informative ones." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.528, + 0.906, + 0.651 + ], + "angle": 0, + "content": "In addition, some summarizing/aggregated prompts might be more generic than others and can be used to provide fine-tuning instructions to a wider range of (multimodal) inputs. The update of such prompts should be prioritized during training and to achieve that, we need a mechanism to capture this prompt popularity. To achieve this, we augment the optimization task in Eq. (10) above with a regularizer that penalizes alignment to less popular prompts," + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.661, + 0.906, + 0.749 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} R (\\boldsymbol {\\alpha}, \\zeta) \\triangleq \\sum_ {t = 1} ^ {n} \\sum_ {p = 1} ^ {\\tau} \\sum_ {q = 1} ^ {n \\times \\tau} \\left(\\alpha_ {t} ^ {p, q} \\cdot \\log U \\left(\\boldsymbol {\\theta} _ {q}; \\zeta\\right)\\right) \\tag {12} \\\\ + \\sum_ {t = 1} ^ {n} \\sum_ {p = 1} ^ {\\tau} \\sum_ {q = 1} ^ {n \\times \\tau} \\Big (1 - \\alpha_ {t} ^ {p, q} \\Big) \\cdot \\log \\Big (1 - U (\\boldsymbol {\\theta} _ {q}; \\zeta) \\Big), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.76, + 0.907, + 0.837 + ], + "angle": 0, + "content": "where \\(U(\\pmb{\\theta}_q;\\zeta)\\) is the popularity function of \\(\\pmb{\\theta}_q\\), which is parameterized by learnable parameter \\(\\zeta\\). For example, \\(U(\\pmb{\\theta}_q;\\zeta) = \\sigma(g(\\pmb{\\theta}_q;\\zeta)) \\in (0,1)^3\\) where \\(g(\\pmb{\\theta}_q;\\zeta)\\) is a deep neural net parameterized with \\(\\zeta\\). Putting Eq. (13) and Eq. (10) together, we arrive at our final loss function," + }, + { + "type": "equation", + "bbox": [ + 0.556, + 0.848, + 0.907, + 0.876 + ], + "angle": 0, + "content": "\\[\n\\underset {\\boldsymbol {\\alpha}, \\boldsymbol {\\theta}, \\gamma , \\zeta} {\\text {m i n i m i z e}} \\left\\{G (\\boldsymbol {\\alpha}, \\boldsymbol {\\theta}, \\gamma) + R (\\boldsymbol {\\alpha}, \\zeta) \\right\\}, \\tag {13}\n\\]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.887, + 0.724, + 0.901 + ], + "angle": 0, + "content": "\\(^3\\sigma(.)\\) denotes the sigmoid function." + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.863, + 0.483, + 0.901 + ], + "angle": 0, + "content": "2This section only describes the aggregation algorithm for inter-client prompts since intra-client prompts are specialized for input-agnostic patterns of missing modalities and can be aggregated via FEDAVG [31]." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.863, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3940" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.228 + ], + "angle": 0, + "content": "which can be optimized via alternating between (1) optimizing \\((\\theta, \\zeta, \\gamma)\\) while fixing \\(\\alpha\\); and (2) optimizing \\(\\alpha\\) while fixing \\((\\theta, \\zeta, \\gamma)\\). The first optimization sub-task is straightforward while the latter is less trivial due to the constraint as well as the discrete nature of the optimizing variables. To sidestep the technical challenge in solving for \\(\\alpha\\) while fixing \\((\\theta, \\zeta, \\gamma)\\), we approximate it by optimizing for \\(\\alpha_{t} = \\{\\alpha_{t}^{p,q}\\}_{p,q}\\) while also fixing \\(\\alpha_{-t} = \\{\\alpha_{-t}^{p,q}\\}_{p,q}\\). This results in a minimum weighted linear sum task," + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.238, + 0.496, + 0.334 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {\\boldsymbol {\\alpha} _ {t}} L \\left(\\boldsymbol {\\alpha} _ {t}\\right) \\triangleq \\sum_ {p = 1} ^ {\\tau} \\sum_ {q = 1} ^ {n \\times \\tau} \\left(\\alpha_ {t} ^ {p, q} \\cdot v (p, q)\\right) \\text {w h e r e} \\tag {14} \\\\ v (p, q) \\triangleq \\left(\\log \\left\\{\\frac {U (\\boldsymbol {\\theta} _ {q} ; \\zeta)}{1 - U (\\boldsymbol {\\theta} _ {q} ; \\zeta)} \\right\\} + c (\\boldsymbol {p} _ {t} ^ {p}, \\boldsymbol {\\theta} _ {q}; \\gamma)\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.342, + 0.484, + 0.417 + ], + "angle": 0, + "content": "which can be solved effectively in polynomial time using the Hungarian algorithm [23]. An overview and pseudocode of the server-side aggregation algorithm are presented in Fig. 3 and Suppl. A. A thorough evaluation of this framework is detailed next in Section 4." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.43, + 0.296, + 0.447 + ], + "angle": 0, + "content": "4. Empirical Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.455, + 0.484, + 0.606 + ], + "angle": 0, + "content": "This section provides a detailed empirical evaluation of our proposed framework, FED-PRIME. Our evaluation is conducted on two datasets UPMC Food-101 [46] and MM-IMDB [1] and several baselines adapting existing works in federated fine-tuning (for unimodal data) and multi-modal prompt-tuning (for centralized data) to our novel setting on decentralized multi-modal data with missing modalities and features. Detailed description of the datasets and baselines are provided in Section 4.1. Our main results and ablation studies are reported in Section 4.2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.615, + 0.285, + 0.631 + ], + "angle": 0, + "content": "4.1. Experiment Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.637, + 0.483, + 0.667 + ], + "angle": 0, + "content": "A. Datasets. Following the experiment protocol in [24], we evaluate FED-PRIME on two multimodal datasets:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.671, + 0.483, + 0.761 + ], + "angle": 0, + "content": "1. UPMC Food-101 [46] is an image-text classification dataset consisting of noisy image-text pairs collected from Google Image Search. Its classes are aligned with those in the larger, publicly available ETHZ Food-101 dataset [4]. For the federated setting, we select only eight most frequent classes from the original UPMC Food-101." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.761, + 0.483, + 0.867 + ], + "angle": 0, + "content": "2. MM-IMDB [1] is a movie genre classification dataset with both image and text modalities. Since each movie can belong to multiple genres, the task is originally a multi-label classification problem, where genres are predicted using image, text, or both. For the federated setting, we use only single-label instances and select eight most frequent classes only from the original MM-IMDB." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.671, + 0.483, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "4Our implementation is available at https://github.com/hangpt01/FedPrime." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "Each dataset is first partitioned into train and test sets with the 80:20 ratio. The (heterogeneous) missing patterns on the train/test data are simulated using the following protocols." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.141, + 0.905, + 0.187 + ], + "angle": 0, + "content": "B. Training Data Simulation. Following [24], we simulate missing data phenomenon within a dataset using a missing rate \\(\\eta \\in (0,1)\\) as detailed in 3 simulation cases below:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.191, + 0.905, + 0.223 + ], + "angle": 0, + "content": "1. Missing Image. We drop the image features in \\(\\eta\\) percent of the data and keep the remaining \\((1 - \\eta)\\) percent complete." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.226, + 0.905, + 0.257 + ], + "angle": 0, + "content": "2. Missing Text. We drop the text features in \\(\\eta\\) percent of the data and keep the remaining \\((1 - \\eta)\\) percent complete." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.26, + 0.905, + 0.307 + ], + "angle": 0, + "content": "3. Missing Both. We drop the image component in \\(\\eta /2\\) percent of the data, and the text component in the other \\(\\eta /2\\) percent, leaving the remaining \\((1 - \\eta)\\) percent complete." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.191, + 0.905, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.31, + 0.905, + 0.37 + ], + "angle": 0, + "content": "The missing-simulated train set is then partitioned into \\( n \\) subsets for \\( n \\) clients via random sampling. The resulting client datasets will therefore have heterogeneous distributions over missing-data patterns." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.374, + 0.905, + 0.45 + ], + "angle": 0, + "content": "C. Test Data Simulation. There are 5 test scenarios, which include: (1) Full Modal (no missing data); (2) Image Only (drop all text features); (3) Text Only (drop all image features); (4) Miss Both (run Missing Both on test data); and (5) \\(\\sim\\) Train (same simulation as the training data)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.454, + 0.905, + 0.499 + ], + "angle": 0, + "content": "D. Metrics. Following prior protocol in [24], we use F1-macro and classification accuracy to evaluate the following baseline methods on MM-IMDB and UPMC Food-101." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.504, + 0.906, + 0.807 + ], + "angle": 0, + "content": "E. Baseline Methods. Following [24], we adopt the pretrained multimodal transformer ViLT [21] as the backbone of our client design. Each client model is a fine-tuned version of this backbone which is generated based on its own local dataset. As no prior work has addressed modality missing in multimodal fine-tuning within federated learning settings, our comparison baselines mostly feature adapted versions of existing work in federated learning and multimodal prompt-tuning. This includes adaptations of the latest missing prompt-tuning SOTA method in [24] into the context of the classical FedAvg method as well as that of a more recent FL approach, FedMSplit [8]. These are referred to as FEDAVG-P and FEDMSPLIT-P, respectively. In addition, to underscore the importance of separating fine-tuning knowledge across intra- and inter-client prompt sets, we also include two simplified variants of FED-PRIME utilizing either inter-client prompt pooling or intra-client prompt pooling, along with their respective aggregations. We refer to these variants as FED-INTER and FED-INTRA, respectively (see more details in Supplementary B)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.818, + 0.716, + 0.833 + ], + "angle": 0, + "content": "4.2. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.901 + ], + "angle": 0, + "content": "This section provides extensive performance comparison between FED-PRIME and other baseline methods across various settings of missing data in the train and test sets, as previously stated in Section 4.1. In addition, we also provide" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3941" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.16 + ], + "angle": 0, + "content": "Table 1. Performance achieved by the baselines on the Food-101 (left) and MM-IMDB (right) datasets under various missing-data settings in the train and test sets. This includes all combinations of the 3 train scenarios and 5 test scenarios (i.e., 15 experiments) per dataset, as previously stated in Section 4.1. The best and second-best results are highlighted in bold red, and blue, respectively. Note that in the Miss Both training scenarios, Test (\\(\\sim\\)Train) and Test (Miss Both) represent the same test scenario according to the definition of \\(\\sim\\)Train. The reported result for Test (Miss Both) in this case is the same as that for Test (\\(\\sim\\)Train), as indicated by a dash (-)." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.17, + 0.907, + 0.406 + ], + "angle": 0, + "content": "
DatasetsUPMC Food-101MM-IMDB
TrainMethodTest(∼ Train)Test(Miss Both)Test(Full Modal)Test(Text only)Test(Image only)Test(∼ Train)Test(Miss Both)Test(Full Modal)Test(Text only)Test(Image only)
Miss TextFEDAVG-P15.71 ± 2.2714.90 ± 1.5721.56 ± 7.8116.91 ± 0.6915.36 ± 0.4322.42 ± 2.2721.89 ± 1.3430.78 ± 1.6518.40 ± 0.0614.53 ± 4.56
FEDMSPLIT-P15.62 ± 1.5117.50 ± 1.9725.27 ± 8.2218.78 ± 0.6417.50 ± 1.9721.02 ± 1.8919.97 ± 0.7424.39 ± 6.0814.38 ± 2.7618.09 ± 6.13
FED-INTER54.82 ± 19.0148.87 ± 24.6459.17 ± 27.0635.13 ± 26.7856.59 ± 15.1218.25 ± 3.5016.95 ± 3.5718.67 ± 7.6315.03 ± 4.6618.01 ± 1.95
FED-INTRA61.71 ± 17.2248.09 ± 19.1262.06 ± 26.9822.51 ± 5.9262.64 ± 11.8313.38 ± 1.7312.77 ± 0.8512.55 ± 1.6711.31 ± 0.3814.33 ± 1.80
FED-PRIME78.88 ± 0.9080.38 ± 0.6592.12 ± 0.4073.01 ± 4.2576.83 ± 1.2231.92 ± 0.2031.48 ± 0.3037.67 ± 0.0430.60 ± 1.4130.69 ± 1.41
Improv. (%)27.82 ↑64.48 ↑48.44 ↑107.83 ↑22.65 ↑42.37 ↑43.81 ↑22.35 ↑66.30 ↑69.65 ↑
Miss ImageFEDAVG-P17.35 ± 4.7715.12 ± 1.4816.84 ± 2.3718.12 ± 6.4914.81 ± 0.2427.69 ± 5.9722.55 ± 3.0631.94 ± 0.9823.76 ± 11.7212.29 ± 0.47
FEDMSPLIT-P74.16 ± 10.5648.88 ± 10.2645.64 ± 32.4388.65 ± 2.1714.81 ± 0.9019.11 ± 11.3316.61 ± 7.2218.19 ± 12.5518.12 ± 12.3012.81 ± 1.25
FED-INTER77.96 ± 11.6264.62 ± 10.2282.08 ± 7.7577.69 ± 12.3537.56 ± 6.4918.79 ± 5.2317.93 ± 3.6020.56 ± 3.5617.67 ± 6.7915.47 ± 2.51
FED-INTRA22.84 ± 3.5220.13 ± 1.7223.48 ± 1.8624.46 ± 2.9916.66 ± 1.3215.75 ± 4.3414.06 ± 3.1615.68 ± 4.7714.53 ± 3.6511.71 ± 0.42
FED-PRIME90.55 ± 0.2279.12 ± 0.4992.89 ± 0.2190.18 ± 0.2954.14 ± 2.5036.08 ± 0.3531.35 ± 0.6138.49 ± 0.5636.91 ± 0.5918.15 ± 0.66
Improv. (%)16.15 ↑22.44 ↑13.17 ↑1.73 ↑44.14 ↑30.30 ↑39.02 ↑20.51 ↑55.35 ↑17.32 ↑
Miss BothFEDAVG-P14.57 ± 1.50-17.17 ± 4.3716.40 ± 4.0513.24 ± 0.3226.45 ± 2.63-33.03 ± 2.5624.12 ± 11.3020.21 ± 1.98
FEDMSPLIT-P49.15 ± 24.76-64.78 ± 36.6264.62 ± 36.5121.49 ± 7.1924.25 ± 5.02-26.05 ± 11.1726.02 ± 9.6419.79 ± 6.20
FED-INTER56.32 ± 21.77-69.57 ± 19.4145.15 ± 34.0959.30 ± 10.8426.53 ± 0.90-31.97 ± 2.2229.69 ± 2.2121.63 ± 0.77
FED-INTRA49.28 ± 32.87-56.70 ± 37.9043.24 ± 34.1949.85 ± 25.4411.90 ± 0.37-12.47 ± 0.4511.46 ± 0.3312.83 ± 0.92
FED-PRIME84.44 ± 2.65-93.64 ± 0.5887.95 ± 0.9172.41 ± 3.8832.01 ± 2.51-38.68 ± 0.6531.00 ± 2.9726.01 ± 0.12
Improv. (%)49.93 ↑-34.60 ↑36.10 ↑22.11 ↑20.66 ↑-17.11 ↑4.41 ↑20.25 ↑
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.115, + 0.408, + 0.807, + 0.422 + ], + "angle": 0, + "content": "(*) Improv. shows the relative performance improvement between our proposal and the second-best. (in percentage)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.447, + 0.483, + 0.493 + ], + "angle": 0, + "content": "ablation studies on (1) performance robustness against various missing rate \\(\\eta\\); (2) the ablated impact of having inter- and intra-client prompt sets; and (3) training convergence." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.504, + 0.231, + 0.518 + ], + "angle": 0, + "content": "4.2.1. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.524, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We compare the performance of FED-PRIME against the other baselines on an extensive set of experiments across all datasets. This includes all combinations of the 3 missing simulation scenarios on train data and 5 test scenarios (with both missing and no missing data) as previously detailed in parts B and C of Section 4.1. All results are reported in Table 1 which show that FED-PRIME consistently outperforms all baselines in all train/test scenarios with various missing-data patterns. The performance on the UPMC Food-101 and MM-IMDB datasets is measured in terms of classification accuracy and F1-macro, respectively. The F1-macro performance on MM-IMDB is relatively low across all settings due to the class imbalance nature of the dataset. In each experiment, the relative performance improvement of FED-PRIME over the second best baseline is reported, which ranges between \\(1.73\\%\\) and \\(107.83\\%\\) for the UPMC Food-101 dataset and between \\(4.41\\%\\) to \\(69.65\\%\\) for the MM-IMDB dataset. Notably, FED-PRIME significantly outperforms both of its simplified variants FED-INTRA and FED-INTER, which either handle intra- or interheterogeneities, in all settings. This highlights the remarkable synergies and essential of handling both types of heterogeneities in multi-modal federated learning with missing data. In addition, it can also be observed that in more than half of the experiment settings, the second best base" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.447, + 0.907, + 0.522 + ], + "angle": 0, + "content": "line is either FED-INTRA or FED-INTER which significantly outperform both FEDAVG-P and FEDMSPLIT-P in those cases. These observations complement our findings of the isolated impact of these components in our ablation studies in Section 4.2.4." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.533, + 0.768, + 0.548 + ], + "angle": 0, + "content": "4.2.2. Impact of Missing Modalities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.907, + 0.827 + ], + "angle": 0, + "content": "Across different testing scenarios, we observe that the highest test performance is always achieved when there is no missing data in the test set as expected. Another interesting observation is that in the Miss Text training scenario, where approximately \\(70\\%\\) of the text features are lost on the training data, the proposed model still performs surprisingly well on the Text Only test scenarios. This suggests that the multi-modal prompt alignment is effective in recovering text information, highlighting the proposed design's adaptability. However, in the Miss Image training scenario, the test performance on the Image Only case suffers instead a significant performance drop. Upon investigation, we believe this is likely due to the VILT backbone's lightweight image processing module, which was optimized to work in tandem with text for competitive performance, hinting at a potential text-centered nature of VILT during pre-training. This is in fact well-aligned with existing observations that co-training with text often leads to improved performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.836, + 0.859, + 0.851 + ], + "angle": 0, + "content": "4.2.3. Robustness Against Various Missing Rates." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "To demonstrate the robustness of FED-PRIME against different missing rates \\(\\eta\\), we run multiple comparisons of FED-PRIME, FEDAVG-P, and its centralized variant [24]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3942" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.087, + 0.327, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.238, + 0.325, + 0.281 + ], + "angle": 0, + "content": "Figure 4. Performance comparison on UMPC Food-101 under Miss Both scenarios with various missing rates." + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.088, + 0.581, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.333, + 0.238, + 0.584, + 0.281 + ], + "angle": 0, + "content": "Figure 5. Plots of the train/test loss convergence on Food-101 under Miss-Text.FED-PRIME converges fastest." + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.088, + 0.871, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.592, + 0.238, + 0.908, + 0.281 + ], + "angle": 0, + "content": "Figure 6. Convergence of Euclidean distance between aggregated prompt centroids on the UPMC Food-101 under Miss Image scenario." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.299, + 0.486, + 0.526 + ], + "angle": 0, + "content": "across a wide range of \\(\\eta \\in \\{0.00, 0.25, 0.50, 0.75, 1.00\\}\\) under the Miss Both training scenarios. The reported results in Fig. 1 and Fig. 4 show that our model significantly outperforms FEDAVG-P and its centralized counterpart Centralized-P. As the missing rate increases, our approach consistently maintains a competitive accuracy above \\(80\\%\\), closely approaching the upper-bound performance of FEDAVG-P in the centralized setting, which benefits from a centralized dataset with better-aligned modalities. These results demonstrate the robustness of our model across varying missing rates and highlight the effectiveness of our federated prompt-tuning approach, achieving performance comparable to a centralized prompt-tuning model despite having access only to decentralized local datasets with heterogeneous missing patterns." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.536, + 0.245, + 0.551 + ], + "angle": 0, + "content": "4.2.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.555, + 0.485, + 0.782 + ], + "angle": 0, + "content": "As explained previously in Section 4.1, FED-INTER and FED-INTRA are in variants of FED-PRIME when we drop the intra- and inter-client prompt sets from the client design, respectively. Table 1 shows that FED-PRIME consistently outperforms both FED-INTER and FED-INTRA in all test conditions with substantial gap. The lack of intra-client prompts causes FED-INTER to underperform FED-PRIME suggests that having a dedicated prompt set to capture the common and highly generalizable input-agnostic aspect of fine-tuning is essential to improve performance. Likewise, FED-INTRA, which omits the inter-client prompt set, also shows a significant performance drop. This highlights the importance of capturing and correctly aggregating input-level patterns of missing data through prompt alignment." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.791, + 0.291, + 0.807 + ], + "angle": 0, + "content": "4.2.5. Convergence Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Model Convergence. Fig. 5 shows that FED-PRIME has superior performance, achieving faster convergence and significantly lower final loss values in both training and testing compared to FED-INTER and FED-INTRA. This indicates better learning efficiency and robustness. Both FED-INTER and FED-INTRA converge more slowly, with" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.299, + 0.907, + 0.406 + ], + "angle": 0, + "content": "consistently higher loss values, especially in the early rounds. Additionally, our method is more stable across server-aggregation rounds, with minimal fluctuations. In contrast, FED-INTER and FED-INTRA exhibit significantly more variability, particularly toward the end. This highlights the effectiveness of FED-PRIME, with faster convergence and improved stability across rounds." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.42, + 0.909, + 0.558 + ], + "angle": 0, + "content": "Prompt Convergence. Fig. 6 shows that FED-PRIME reduces the distance between the aggregated prompt centroids much faster than FEDAVG-P, which highlights the faster convergence rate of FED-PRIME over FEDAVG-P, indicating faster convergence and better prompt alignment. In contrast, FEDAVG-P converges more slowly and unstably, likely due to incorrect default prompt alignment. This again underscores the necessity of having a learnable prompt clustering mechanism to prevent incorrect prompt alignment." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.608, + 0.634, + 0.623 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Fine-tuning large pre-trained models for specific tasks has become an increasingly effective solution paradigm in machine learning. However, existing fine-tuning frameworks struggle in practical, multimodal settings where data is distributed across private silos with heterogeneous patterns of missing modalities and features. To address these challenges, this paper introduces a novel federated prompt-tuning approach that optimizes, aligns, and aggregates prompt instructions, compensating for gaps in data modalities and distributional patterns across a number of diverse and private local datasets. Rigorous evaluation on multimodal benchmark datasets demonstrates the effectiveness and robustness of our approach against various patterns of missing data in local datasets. We view this work as an initial step toward developing resilient, adaptable federated models that can fully leverage the diversity of multimodal data in real-world environments." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3943" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.484, + 0.241 + ], + "angle": 0, + "content": "This research is funded by Hanoi University of Science and Technology (HUST) under grant number T2024-TD-002. Additionally, this work utilized GPU compute resource at SDSC and ACES through allocation CIS230391 from the Advanced Cyberinfrastructure Coordination Ecosystem: Services and Support (ACCESS) program [2], which is supported by U.S. National Science Foundation grants #2138259, #2138286, #2138307, #2137603, and #2138296." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.266, + 0.188, + 0.281 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.291, + 0.484, + 0.331 + ], + "angle": 0, + "content": "[1] John Arevalo, Thamar Solorio, Manuel Montes y Gómez, and Fabio A. González. Gated multimodal units for information fusion. ArXiv, abs/1702.01992, 2017. 2, 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.333, + 0.484, + 0.429 + ], + "angle": 0, + "content": "[2] Timothy J. Boerner, Stephen Deems, Thomas R. Furlani, Shelley L. Knuth, and John Towns. Access: Advancing innovation: Nsf's advanced cyberinfrastructure coordination ecosystem: Services & support. In Practice and Experience in Advanced Research Computing 2023: Computing for the Common Good, page 173-176, New York, NY, USA, 2023. Association for Computing Machinery. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.431, + 0.487, + 0.886 + ], + "angle": 0, + "content": "[3] Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S. Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, Erik Brynjolfsson, S. Buch, Dallas Card, Rodrigo Castellon, Niladri S. Chatterji, Annie S. Chen, Kathleen A. Creel, Jared Davis, Dora Demszky, Chris Donahue, Moussa Doumbouya, Esin Durmus, Stefano Ermon, John Etchemendy, Kawin Ethayarajh, Li Fei-Fei, Chelsea Finn, Trevor Gale, Lauren E. Gillespie, Karan Goel, Noah D. Goodman, Shelby Grossman, Neel Guha, Tatsunori Hashimoto, Peter Henderson, John Hewitt, Daniel E. Ho, Jenny Hong, Kyle Hsu, Jing Huang, Thomas F. Icard, Saahil Jain, Dan Jurafsky, Pratyusha Kalluri, Siddharth Karamcheti, Geoff Keeling, Fereshte Khani, O. Khattab, Pang Wei Koh, Mark S. Krass, Ranjay Krishna, Rohith Kuditipudi, Ananya Kumar, Faisal Ladhak, Mina Lee, Tony Lee, Jure Leskovec, Isabelle Levent, Xiang Lisa Li, Xuechen Li, Tengyu Ma, Ali Malik, Christopher D. Manning, Suvir P. Mirchandani, Eric Mitchell, Zanele Munyikwa, Suraj Nair, Avanika Narayan, Deepak Narayanan, Benjamin Newman, Allen Nie, Juan Carlos Niebles, Hamed Nilforoshan, J. F. Nyarko, Giray Ogut, Laurel Orr, Isabel Papadimitriou, Joon Sung Park, Chris Piech, Eva Portelance, Christopher Potts, Aditi Raghunathan, Robert Reich, Hongyu Ren, Frieda Rong, Yusuf H. Roohani, Camilo Ruiz, Jack Ryan, Christopher R'e, Dorsa Sadigh, Shiori Sagawa, Keshav Santhanam, Andy Shih, Krishna Parasuram Srinivasan, Alex Tamkin, Rohan Taori, Armin W. Thomas, Florian Tramér, Rose E. Wang, William Wang, Bohan Wu, Jiajun Wu, Yuhuai Wu, Sang Michael Xie, Michihiro Yasunaga, Jiaxuan You, Matei A. Zaharia, Michael Zhang, Tianyi Zhang, Xikun Zhang, Yuhui Zhang, Lucia Zheng, Kaitlyn Zhou, and Percy Liang. On the opportunities and risks of foundation models. ArXiv, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.887, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[4] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool." + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.291, + 0.487, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.133 + ], + "angle": 0, + "content": "Food-101 - mining discriminative components with random forests. In European Conference on Computer Vision, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.136, + 0.906, + 0.274 + ], + "angle": 0, + "content": "[5] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. CoRR, abs/2005.14165, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.277, + 0.906, + 0.344 + ], + "angle": 0, + "content": "[6] Alberto Brunete, Ernesto Gambao, Miguel Hernando, and Raquel Cedazo. Smart assistive architecture for the integration of IoT devices, robotic systems, and multimodal interfaces in healthcare environments. Sensors, 21(6):2212, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.347, + 0.906, + 0.415 + ], + "angle": 0, + "content": "[7] Jiayi Chen and Aidong Zhang. Hgmf: heterogeneous graph-based fusion for multimodal data with incompleteness. In Proceedings of the 26th ACM International Conference on Knowledge Discovery & Data Mining, pages 1295-1305, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.419, + 0.906, + 0.488 + ], + "angle": 0, + "content": "[8] Jiayi Chen and Aidong Zhang. FedMSplit: Correlation-adaptive federated multi-task learning across multimodal split networks. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 87-96, 2022. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.49, + 0.906, + 0.544 + ], + "angle": 0, + "content": "[9] Jinyu Chen, Wenchao Xu, Song Guo, Junxiao Wang, Jie Zhang, and Haozhao Wang. Fedtune: A deep dive into efficient federated fine-tuning with pre-trained transformers, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.547, + 0.906, + 0.615 + ], + "angle": 0, + "content": "[10] Yae Jee Cho, Andre Manoel, Gauri Joshi, Robert Sim, and Dimitrios Dimitriadis. Heterogeneous ensemble knowledge transfer for training large models in Federated Learning. Computing Research Repository arXiv Preprints, 2204.12703, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.618, + 0.906, + 0.673 + ], + "angle": 0, + "content": "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.676, + 0.906, + 0.786 + ], + "angle": 0, + "content": "[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota, 2019. Association for Computational Linguistics. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.906, + 0.87 + ], + "angle": 0, + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.873, + 0.906, + 0.902 + ], + "angle": 0, + "content": "[14] Alireza Fallah, Aryan Mokhtari, and Asuman Ozdaglar. Personalized federated learning: A meta-learning ap" + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.093, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3944" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.481, + 0.119 + ], + "angle": 0, + "content": "proach. Computing Research Repository arXiv Preprints, 2002.07948, 2020.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[15] Ziwei Fan, Hao Ding, Anoop Deoras, and Trong Nghia Hoang. Personalized federated domain adaptation for item-to-item recommendation. In Proceedings of the Thirty-Ninth Conference on Uncertainty in Artificial Intelligence, pages 560-570. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.191, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[16] Tao Guo, Song Guo, Junxiao Wang, Xueyang Tang, and Wenchao Xu. PromptFL: Let Federated Participants Cooperatively Learn Prompts Instead of Models-Federated Learning in Age of Foundation Model. IEEE Transactions on Mobile Computing, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.482, + 0.327 + ], + "angle": 0, + "content": "[17] Minh Hoang, Nghia Hoang, Bryan Kian Hsiang Low, and Carleton Kingsford. Collective model fusion for multiple black-box experts. In Proceedings of the 36th International Conference on Machine Learning, pages 2742-2750. PMLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.483, + 0.398 + ], + "angle": 0, + "content": "[18] Trong Nghia Hoang, Quang Minh Hoang, Kian Hsiang Low, and Jonathan P. How. Collective online learning of Gaussian processes in massive multi-agent systems. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7850-7857, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[19] Parminder Kaur, Husanbir Singh Pannu, and Avleen Kaur Malhi. Comparative analysis on cross-modal information retrieval: A review. Computer Science Review, 39:100336, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.456, + 0.482, + 0.51 + ], + "angle": 0, + "content": "[20] Muhammad Uzair khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.579 + ], + "angle": 0, + "content": "[21] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 5583-5594. PMLR, 2021. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[22] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A. Shamma, Michael S. Bernstein, and Li Fei-Fei. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International Journal of Computer Vision, 123:32 - 73, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.482, + 0.704 + ], + "angle": 0, + "content": "[23] Harold W Kuhn. The Hungarian Method for the Assignment Problem. Naval Research Logistics Quarterly, 2(1-2):83-97, 1955. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[24] Yi-Lun Lee, Yi-Hsuan Tsai, Wei-Chen Chiu, and Chen-Yu Lee. Multimodal prompting with missing modalities for visual recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3, 6, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.763, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[25] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3045-3059, Online and Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[26] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. Proceedings of Machine Learning and Systems, 2:429-450, 2020. 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[27] Xiang Li, Kaixuan Huang, Wenhao Yang, Shusen Wang, and Zhihua Zhang. On the convergence of FedAvg on non-IID data. Computing Research Repository arXiv Preprints, 1907.02189t, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.905, + 0.216 + ], + "angle": 0, + "content": "[28] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[29] Wang Lu, Xixu Hu, Jindong Wang, and Xing Xie. FedCLIP: Fast Generalization and Personalization for CLIP in Federated Learning, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.314 + ], + "angle": 0, + "content": "[30] Tengfei Ma, Trong Nghia Hoang, and Jie Chen. Federated learning of models pre-trained on different features with consensus graphs. In Uncertainty in Artificial Intelligence, pages 1336-1346, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.905, + 0.384 + ], + "angle": 0, + "content": "[31] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial Intelligence and Statistics, pages 1273-1282, 2017. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.905, + 0.439 + ], + "angle": 0, + "content": "[32] John Nguyen, Jianyu Wang, Kshitiz Malik, Maziar Sanjabi, and Michael Rabbat. Where to Begin? On the Impact of Pre-Training and Initialization in Federated Learning, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.51 + ], + "angle": 0, + "content": "[33] Manh Duong Nguyen, Trung Thanh Nguyen, Huy-Hieu Pham, Trong Nghia Hoang, Phi Le Nguyen, and Thanh Trung Huynh. Fedmac: Tackling partial-modality missing in federated learning with cross-modal aggregation and contrastive regularization. 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.512, + 0.905, + 0.608 + ], + "angle": 0, + "content": "[34] Nang Hung Nguyen, Phi Le Nguyen, Thuy Dung Nguyen, Trung Thanh Nguyen, Duc Long Nguyen, Thanh Hung Nguyen, Huy Hieu Pham, and Thao Nguyen Truong. FedDRL: Deep reinforcement learning-based adaptive aggregation for non-IID data in federated learning. In Proceedings of the 51st International Conference on Parallel Processing, pages 1-11, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.61, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[35] Nang Hung Nguyen, Duc Long Nguyen, Trong Bang Nguyen, Thanh-Hung Nguyen, Huy Hieu Pham, Truong Thao Nguyen, and Phi Le Nguyen. CADIS: Handling cluster-skewed Non-IID data in federated learning with clustered aggregation and knowledge distilled regularization. In Proceedings of the 23rd IEEE/ACM International Symposium on Cluster, Cloud and Internet Computing, pages 249-261, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.721, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[36] Quang Ha Pham, Nang Hung Nguyen, Thanh Hung Nguyen, Huy Hieu Pham, Phi Le Nguyen, and Truong Thao Nguyen. SEM: A simple yet efficient model-agnostic local training mechanism to tackle data sparsity and scarcity in federated learning. In Proceedings of the 11st International Symposium on Computing and Networking, pages 120–126, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[37] Thu Hang Phung, Binh P Nguyen, Thanh Hung Nguyen, Quoc Viet Hung Nguyen, Phi Le Nguyen, and Thanh Trung Huynh. A contrastive learning and graph-based approach for missing modalities in multimodal federated learning. In Proceedings of the 2024 International Joint Conference on Neural Networks, pages 1-8, 2024. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3945" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[38] Petra Poklukar, Miguel Vasco, Hang Yin, Francisco S Melo, Ana Paiva, and Danica Kragic. Geometric multimodal contrastive representation learning. In Proceedings of the 2022 International Conference on Machine Learning, pages 17782-17800, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.485, + 0.247 + ], + "angle": 0, + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. CoRR, abs/2103.00020, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.484, + 0.319 + ], + "angle": 0, + "content": "[40] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21(140):1-67, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.484, + 0.389 + ], + "angle": 0, + "content": "[41] Kuniaki Saito, Kihyuk Sohn, Xiang Zhang, Chun-Liang Li, Chen-Yu Lee, Kate Saenko, and Tomas Pfister. Prefix conditioning unifies language and label supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2861-2870, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.484, + 0.46 + ], + "angle": 0, + "content": "[42] Rachael Hwee Ling Sim, Yehong Zhang, Trong Nghia Hoang, Xinyi Xu, Bryan Kian Hsiang Low, and Patrick Jaillet. Incentives in private collaborative machine learning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.484, + 0.516 + ], + "angle": 0, + "content": "[43] Canh T Dinh, Nguyen Tran, and Josh Nguyen. Personalized federated learning with Moreau envelopes. Advances in Neural Information Processing Systems, 33:21394-21405, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.484, + 0.587 + ], + "angle": 0, + "content": "[44] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, S. M. Ali Eslami, Oriol Vinyals, and Felix Hill. Multimodal few-shot learning with frozen language models. In Advances in Neural Information Processing Systems, pages 200-212. Curran Associates, Inc., 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.484, + 0.658 + ], + "angle": 0, + "content": "[45] Feng Wang, M Cenk Gursoy, and Senem Velipasalar. Communication-efficient and privacy-preserving feature-based federated transfer learning. In Proceedings of the 2022 IEEE Global Communications Conference, pages 3875-3880, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.484, + 0.73 + ], + "angle": 0, + "content": "[46] Xin Wang, Devinder Kumar, Nicolas Thome, Matthieu Cord, and Frédéric Precioso. Recipe recognition with large multimodal food dataset. In 2015 IEEE International Conference on Multimedia & Expo Workshops (ICMEW), pages 1-6, 2015. 2, 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.484, + 0.801 + ], + "angle": 0, + "content": "[47] Pei-Yau Weng, Minh Hoang, Lam M. Nguyen, My T. Thai, Tsui-Wei Weng, and Trong Nghia Hoang. Probabilistic federated prompt-tuning with non-IID and imbalanced data. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.484, + 0.845 + ], + "angle": 0, + "content": "[48] Baochen Xiong, Xiaoshan Yang, Fan Qi, and Changsheng Xu. A unified framework for multi-modal Federated Learning. Neurocomputing, 480:110-118, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.847, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[49] Baochen Xiong, Xiaoshan Yang, Yaguang Song, Yaowei Wang, and Changsheng Xu. Client-adaptive cross-model reconstruction network for modality-incomplete multimodal federated learning. In Proceedings of the 31st ACM Interna-" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "tional Conference on Multimedia, pages 1241-1249, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[50] Fu-En Yang, Chien-Yi Wang, and Yu-Chiang Frank Wang. Efficient Model Personalization in Federated Learning via Client-Specific Prompt Generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 19159-19168, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[51] Guan Yu, Quefeng Li, Dinggang Shen, and Yufeng Liu. Optimal sparse linear prediction for block-missing multimodality data without imputation. Journal of the American Statistical Association, 115(531):1406-1419, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[52] Songcan Yu, Junbo Wang, Walid Hussein, and Patrick C.K. Hung. Robust multimodal federated learning for incomplete modalities. Computer Communications, 214:234-243, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.306, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[53] Mikhail Yurochkin, Mayank Agarwal, Soumya Ghosh, Kristjan Greenewald, Trong Nghia Hoang, and Yasaman Khazaeni. Bayesian nonparametric federated learning of neural networks. In International Conference on Machine Learning, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[54] Mikhail Yurochkin, Mayank Argawal, Soumya Ghosh, Kristjan Greenewald, and Trong Nghia Hoang. Statistical model aggregation via parameter matching. In Advances in Neural Information Processing Systems, pages 10954-10964, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[55] Changqing Zhang, Yajie Cui, Zongbo Han, Joey Tianyi Zhou, Huazhu Fu, and Qinghua Hu. Deep partial multi-view learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(5):2402-2415, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[56] Jianyi Zhang, Hao Frank Yang, Ang Li, Xin Guo, Pu Wang, Haiming Wang, Yiran Chen, and Hai Li. Mllm-fl: Multimodal large language model assisted federated learning on heterogeneous and long-tailed data. arXiv preprint arXiv:2409.06067, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.574, + 0.905, + 0.628 + ], + "angle": 0, + "content": "[57] Wei Zhang and Xiang Li. Data privacy preserving federated transfer learning in machinery fault diagnostics using prior distributions. Structural Health Monitoring, 21(4):1329-1344, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.63, + 0.905, + 0.685 + ], + "angle": 0, + "content": "[58] Haodong Zhao, Wei Du, Fangqi Li, Peixuan Li, and Gongshen Liu. FedPrompt: Communication-Efficient and Privacy Preserving Prompt Tuning in Federated Learning. In Proc. ICASSP, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.687, + 0.905, + 0.742 + ], + "angle": 0, + "content": "[59] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.744, + 0.905, + 0.785 + ], + "angle": 0, + "content": "[60] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision (IJCV), 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.787, + 0.905, + 0.842 + ], + "angle": 0, + "content": "[61] Tongxue Zhou, Pierre Vera, Stephane Canu, and Su Ruan. Missing data imputation via conditional generator and correlation learning for multimodal brain tumor segmentation. Pattern Recognition Letters, 158:125-132, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.842 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "3946" + } + ] +] \ No newline at end of file diff --git a/2025/Federated Prompt-Tuning with Heterogeneous and Incomplete Multimodal Client Data/c4864325-3bcd-4609-87ea-0a89ff648c25_origin.pdf b/2025/Federated Prompt-Tuning with Heterogeneous and Incomplete Multimodal Client Data/c4864325-3bcd-4609-87ea-0a89ff648c25_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..95cda5d72639952270fa1622c6220667a310dd7b --- /dev/null +++ b/2025/Federated Prompt-Tuning with Heterogeneous and Incomplete Multimodal Client Data/c4864325-3bcd-4609-87ea-0a89ff648c25_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9916345f534c6cee00f9345a02e8ae8d3ad1d795a1412bca08d2819c68c902dd +size 1115241 diff --git a/2025/Federated Prompt-Tuning with Heterogeneous and Incomplete Multimodal Client Data/full.md b/2025/Federated Prompt-Tuning with Heterogeneous and Incomplete Multimodal Client Data/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7b7cd2f8b2e0c010f5b13757da559671397ebd03 --- /dev/null +++ b/2025/Federated Prompt-Tuning with Heterogeneous and Incomplete Multimodal Client Data/full.md @@ -0,0 +1,319 @@ +# Federated Prompt-Tuning with Heterogeneous and Incomplete Multimodal Client Data + +Thu Hang Phung1 Duong M. Nguyen1 Thanh Trung Huynh2 Quoc Viet Hung Nguyen3 +Trong Nghia Hoang4* Phi Le Nguyen1* +1 Institute of AI Innovation and Societal Impact, Hanoi University of Science and Technology +2 EPFL 3 Griffith University 4 Washington State University + +# Abstract + +This paper introduces a generalized federated prompt-tuning framework for practical scenarios where local datasets are multi-modal and exhibit different distributional patterns of missing features at the input level. The proposed framework bridges the gap between federated learning and multi-modal prompt-tuning which have traditionally focused on either uni-modal or centralized data. A key challenge in this setting arises from the lack of semantic alignment between prompt instructions that encode similar distributional patterns of missing data across different clients. To address this, our framework introduces specialized client-tuning and server-aggregation designs that simultaneously optimize, align, and aggregate prompt-tuning instructions across clients and data modalities. This allows prompt instructions to complement one another and be combined effectively. Extensive evaluations on diverse multimodal benchmark datasets demonstrate that our work consistently outperforms state-of-the-art (SOTA) baselines. + +# 1. Introduction + +The recent emergence of exceedingly large models [3] with immense scale and versatility due to being pre-trained on large swaths of generic data has resulted in a paradigm shift in machine learning (ML). Most task-specific models are now fine-tuned versions of those large models instead of being created from de novo learning processes. For example, most state-of-the-art (SOTA) models solving specific NLP tasks are fine-tuned versions of existing large pre-trained or foundation models such as BERT [12], GPT [5], and T5 [40]. Likewise, SOTA models for various computer vision tasks have also been derived from foundation models such as the vision transformer [13] pre-trained on extensive datasets like ImageNet [11]. The standard fine-tuning approach, however, assumes centralized data which is not always possible in practical scenarios where fine-tuning data + +are scattered across multiple private siloed systems. + +To overcome this siloed data challenge, federated fin-tuning has recently been considered to aggregate the fin-tuned updates across private devices running their own (local) fine-tuning processes [9, 32, 47]. This approach can be further combined with an existing parameter-efficient tuning method called prompt tuning [25]. This method focuses on creating lightweight signals such as additional tokens that are prepended to the input of the pre-trained models [41, 44, 59, 60], providing them valuable contextual information to perform specific tasks more effectively. Local sets of prompts can then be aggregated using the conventional parameter aggregation procedure in FL [16, 29, 47, 50, 58]. + +However, such federated prompt-tuning techniques are restricted to learning with uni-modal data. In more practical scenarios, advances in sensor and device manufacturing increasingly enable the capture of multiple data modalities (e.g., text, images), leading to richer datasets for fine-tuning [6]. Leveraging such diverse multi-modal information can significantly enhance model performance, as demonstrated by the success of multi-modal foundation models like CLIP [39] and ViLT [21]. This has further motivated research on multi-modal federated learning [8, 33, 37, 52] and fine-tuning pre-trained multi-modal LLMs, i.e., multi-modal federated fine-tuning [56]. + +Despite these advances, multi-modal federated learning research often adopts specific local architectures which cannot be learned via fine-tuning existing pre-trained multimodal foundation models [8, 33, 37, 52]. Existing multimodal federated fine-tuning research on the other hand assumes that all local datasets have the same set of modalities [56]. As such, these approaches are not suitable for situations where different clients might collect data using different types of sensors. Furthermore, the multi-modal sensor data across clients might even have different missing patterns of data entries. This leads to local datasets with both intra- and inter-heterogeneities. Intra-heterogeneity occurs when individual samples within a single dataset + +have missing entries for different subsets of features. Interheterogeneity arises when different clients have different distributions over their data-missing patterns. + +There also exists recent work in prompt-tuning that learns a specific prompt for each subset of missing modalities [24]. Input samples having the same set of missing modalities are then augmented with the same (learnable) prompt during training, specializing it to complement for the corresponding missing modalities. This approach is, however, restricted to centralized data setting (i.e., no interheterogeneities) and does not generalize well to federated settings where different sets of local prompts might be biased towards different distributional patterns of missing data. A naive aggregation of these prompts might therefore collapse them into less informative prompts which are ignorant of the heterogeneities in missing-data patterns across clients, resulting in poor performance (see Fig. 1). + +To bridge this gap between multi-modal prompt-tuning and federated learning, we propose a multimodal federated prompt-tuning approach that optimizes and aggregates two distinct sets of (local) prompts for each client: inter-client and intra-client prompts. Inter-client prompts breakdown and encode prototypical distribution patterns of missing input data, which can be aligned and aggregated across clients to represent shared distributional information. In contrast, intra-client prompts encode information tailored to support specific local subsets of missing modalities, independent of the distributional patterns of missing data at each client. This is achieved via the following contributions: + +1. We develop a local input-adaptive fine-tuning algorithm which identifies and updates a subset of inter- and intra-client prompts that are most relevant to each individual sample. This is formulated as prompt-retrieval task which focuses on learning mapping functions that push an input towards its relevant prompts on an embedding space while pulling away from irrelevant prompts on an embedding space. This helps decompose and distill local fine-tuning knowledge into different parts of the prompt set, preventing learned instructions for different fine-tuning patterns from being overloaded on the same prompts (Section 3.2). +2. We develop a server-side model for aggregating and aligning inter-client prompts across clients. Both generation and alignment are framed as part of a clustering task, where inter-client prompts encoding similar input-level distribution patterns of missing data are identified and merged into more comprehensive and versatile prompt instructions. The empirical effectiveness of these combined prompts in reducing local training losses is used as a learning feedback to optimize this clustering model (Section 3.3). +3. We evaluate and demonstrate the effectiveness of our proposed approach through extensive experiments on two benchmark datasets, MM-IMDB [1] and UPMC Food + +![](images/2d42a96d9e28d4269fffcac2392f709e1fffcd35722a50fe27a27aff2ab6142a.jpg) +Figure 1. Our approach with prompt alignment outperforms FedAvg prompt-tuning (w/o alignment) across settings with different data missing rates. + +101 [46], comparing its performance against a variety of existing multimodal federated learning methods as well as repurposed prompt-tuning baselines for federated settings. The reported results show that our method consistently and significantly outperforms the baselines across various modality-missing scenarios, achieving new SOTA performance for multi-modal federated learning (Section 4). + +A bird-view of our proposed framework, Federated Prompt-Tuning with Heterogeneous and Incomplete Multimodal Client Data (FED-PRIME), is also provided in Figure 2. + +# 2. Related Works + +Multi-Modal Federated Learning. Federated learning (FL) [17, 31, 53, 54] is a collaborative learning paradigm that enables multiple devices to train a shared model while keeping local data private. This design addresses concerns over privacy, security, and inefficiency in transferring large datasets [45, 57]. While FL has been widely studied [10, 14, 15, 18, 26, 27, 30, 34–36, 42, 43], most existing methods focus on uni-modal data (e.g., image or text). However, with the rapid advancement of mobile phones and IoT devices [6], local data increasingly comes from multiple modalities. Integrating such heterogeneous signals enables richer, more robust representations [19, 48], motivating growing interest in multi-modal FL. Yet many current approaches rely on custom architectures that cannot exploit pre-trained multi-modal models. More recent studies explore federated fine-tuning of pre-trained multi-modal models, but they typically assume each client holds the same set of modalities [56], limiting applicability in realistic heterogeneous environments. + +Missing Data in Multi-Modal FL. In real-world settings, local devices collect data from different types of sensors, leading to datasets with varying feature modalities (inter-heterogeneity). Moreover, sensor malfunctions or data corruption can cause missing features, with devices exhibiting distinct missing patterns (intra-heterogeneity). + +Existing solutions either treat each modality independently [55, 61] or are restricted to centralized learning [7, 38, 51], both overlooking inter-modal dependencies. Recent FL methods address inter-heterogeneity [8, 37] but not intra-heterogeneity, while approaches like FedInMM [52] and FedMAC [33] tackle intra-client missing data but ignore cross-client alignment, limiting effectiveness. A related line of work, including FedMSplit [8] and CACMRN [49], handles modality-incomplete data under a personalized FL framework, where each client retains its model. This paradigm differs from standard FL, which aims to improve a single shared model, making these methods not directly comparable. + +Addressing Missing Data with Prompt-Tuning. Prompt-tuning [41, 44, 59, 60] provides an effective way to adapt pre-trained models by learning prompts that add contextual information to input embeddings. Recent work has extended this idea to missing-modality scenarios by learning a prompt for each subset of missing features [20, 24], so that inputs with similar missing patterns are augmented with suitable prompts. However, these methods are developed for centralized training and do not address the additional heterogeneity introduced in FL, where prompts must be aligned and aggregated across clients. Designing effective alignment and aggregation strategies for such prompts is therefore a key challenge, as demonstrated in our framework. + +# 3. Multi-Modal Federated Prompt-Tuning + +Multimodal federated learning faces significant challenges due to missing data modalities, leading to both inter-client and intra-client heterogeneities that complicate knowledge aggregation and degrade overall model performance - see Fig. 1. To address this challenge (Section 3.1), we present a federated prompt-tuning framework based on the pretrained Vision and Language Transformer (ViLT) model, which comprises a prompt-based client design and a server-side prompt-aggregation algorithm. The client design features a fine-tuning algorithm that learns and decomposes the fine-tuned knowledge into separate sets of inter-client and intra-client prompts specialized for fine-tuning at different input-level patterns of missing data and input-agnostic patterns of missing modalities (Section 3.2). The server-side algorithm then learns alignment between inter-client prompts specializing for similar input-level patterns of missing data and aggregates them to create more versatile prompts. This is also combined with a standard FEDAVG-based aggregation of intra-prompts specializing for input-agnostic patterns of missing modalities (Section 3.3) - see Fig. 2. + +# 3.1. Problem Formulation + +In multimodal federated learning (MMFL), we assume there are $n$ clients and $r$ data modalities (e.g., text, image). + +Each client $t \in [n]^{1}$ has access to a private dataset $D_{t} \triangleq \{x(M_{t,s}), z_{t,s}\}_{s=1}^{m}$ , where $x(M_{t,s})$ denote the observed features for the set of modalities, $M_{t,s} \subseteq [r]$ recorded for the $s$ -th data point of the $t$ -th client, and $z_{t,s}$ denotes the corresponding target label. The goal is to learn a multimodal model $w_{*}$ using all the private datasets $\{D_{t}\}_{t=1}^{n}$ without centralizing them. That is, + +$$ +\boldsymbol {w} _ {*} = \underset {\boldsymbol {w}} {\arg \min } \left\{L (\boldsymbol {w}) \triangleq \frac {1}{n} \sum_ {t = 1} ^ {n} L _ {t} (\boldsymbol {w}) \right\} \text {w i t h} +$$ + +$$ +L _ {t} (\boldsymbol {w}) \triangleq \frac {1}{n} \sum_ {s = 1} ^ {m} \ell \left(F \left(\boldsymbol {x} \left(M _ {t, s}\right); \boldsymbol {w}\right), z _ {t, s}\right) \tag {1} +$$ + +where $F(\pmb{x}(M_{t,s}); \pmb{w})$ denotes the multimodal prediction model with parameter $\pmb{w}$ and $\ell(F(\pmb{x}(M_{t,s}); \pmb{w}), z_{t,s})$ denotes the individual loss of $\pmb{w}$ on the $s$ -th data point. Previous work [8, 33, 37, 52] in MMFL has largely focused on designing $F(\pmb{x}(M_{t,s}); \pmb{w})$ without leveraging existing knowledge in pre-trained foundation model. Instead, we let $F(\pmb{x}(M_{t,s}); \pmb{w})$ be a prompt-tuned version of a pre-trained transformer-based model, + +$$ +F (\boldsymbol {x} (M); \boldsymbol {w}) \triangleq F _ {c} \left(F _ {p} \left(F _ {e} (\boldsymbol {x} (M)) \circ \boldsymbol {w} _ {p}\right); \boldsymbol {w} _ {c}\right) \tag {2} +$$ + +where $F_{c}(:,\boldsymbol{w}_{c})$ is the prediction head parameterized with $\boldsymbol{w}_{c}$ , $F_{e}(\boldsymbol{x}(M)) = F_{e}(\boldsymbol{x}_{1}) \circ \dots \circ F_{e}(\boldsymbol{x}_{|M|})$ where $F_{e}(\boldsymbol{x}_{i})$ denote the tokenized input sequence of each unimodal feature $\boldsymbol{x}_{i}$ in $\boldsymbol{x}(M) \triangleq (\boldsymbol{x}_{1}, \boldsymbol{x}_{2}, \dots, \boldsymbol{x}_{|M|})$ , $F_{p}()$ denotes the frozen multimodal pre-trained model – e.g., ViLT [21] – and $\boldsymbol{w}_{p}$ denotes the set of (learnable) prompts. Here, $\circ$ denotes the set concatenation operator and $\boldsymbol{w} = (\boldsymbol{w}_{c}, \boldsymbol{w}_{p})$ denotes the full set of tuning parameters. Due to the decentralized and private nature of the local datasets $\{D_{t}\}_{t=1}^{n}$ , solving Eq. (1) directly is not possible. Instead, following the federated learning [31] paradigm, its optimal solution $\boldsymbol{w}_{*}$ is approximated via a number of iterations over two interleaving client- and server-side steps: + +$$ +\boldsymbol {w} _ {t} ^ {(h + 1)} = \operatorname {u p d a t e} \left(L _ {t}, \boldsymbol {w} _ {*} ^ {(h)}\right) \forall t \in [ n ], \tag {3} +$$ + +$$ +\boldsymbol {w} _ {*} ^ {(h + 1)} = \text {a g g r e g a t e} \left(\boldsymbol {w} _ {1} ^ {(h + 1)}, \dots , \boldsymbol {w} _ {n} ^ {(h + 1)}\right), \tag {4} +$$ + +where $\boldsymbol{w}_t^{(h)}$ denote the estimation of the local tuning parameter $\boldsymbol{w}_t$ of client $t$ after $h$ iterations. For example, the update step can be a vanilla gradient descent (GD) optimization of the prompt $\omega$ and the aggregate step denotes the average of the locally optimized prompts. Such vanilla approach however does not account for both the inter- and intra-heterogeneities among local prompts (see Section 1) which ultimately risks aggregating incompatible prompts and degrade the overall performance (see Section 4). To avoid this, we will introduce novel designs for client- and server-prompts as detailed in the followings. + +![](images/f0ddb7f20be97d3b27ffd81002ae84c428259f17301d7fbd3677cd308e86b259.jpg) +Figure 2. Overview of the proposed multi-modal federated prompt-tuning framework - see Alg. 3. Each client maintains two (learnable) sets of intra- and inter-client prompts. At the beginning of each iteration, each client performs local training via Eq. (7) and Eq. (8). Local sets of intra- and inter-client prompts are then sent to the server for aggregation - see Eq. (10)-Eq. (14). + +# 3.2. Client Design + +To enable effective prompt-aggregation across heterogeneous clients, we need to separate the fine-tuning treatment for input-level patterns of missing data and input-agnostic patterns of missing modalities. This can be achieved via further parameterizing $\boldsymbol{w}_p = (\boldsymbol{w}_p^{inter}, \boldsymbol{w}_p^{intra})$ with two distinct subsets of $\tau$ inter- and intra-client prompts: + +$$ +\boldsymbol {w} _ {p} ^ {\text {i n t e r}} \triangleq \left\{\boldsymbol {p} _ {1} ^ {\text {i n t e r}}, \boldsymbol {p} _ {2} ^ {\text {i n t e r}}, \dots , \boldsymbol {p} _ {\tau} ^ {\text {i n t e r}} \right\}, \tag {5} +$$ + +$$ +\boldsymbol {w} _ {p} ^ {\text {i n t r a}} \triangleq \left\{\boldsymbol {p} _ {1} ^ {\text {i n t r a}}, \boldsymbol {p} _ {2} ^ {\text {i n t r a}}, \dots , \boldsymbol {p} _ {\tau} ^ {\text {i n t r a}} \right\}, \tag {6} +$$ + +which are learned separately. The learned inter-client prompts $\pmb{w}_{p}^{inter}$ across clients will be sent to the server which clusters them into different groups based on their specialization for similar input-level patterns of missing data - Section 3.3. Local sets of intra-client prompts $\pmb{w}_{p}^{intra}$ can be averaged via FEDAVG [31] since their specialization is for input-agnostic patterns of missing modalities. Intuitively, the aggregation mechanisms will determine how different aspects of fine-tuning knowledge are encoded across the prompt sets. For example, if knowledge regarding different input-level patterns of missing data is (incorrectly) incorporated into $\{\pmb{p}_i^{intra}\}_{i=1}^\tau$ , they will be averaged out and collapsed into less informative prompts due to the non-clustering aggregation mechanism of $\{\pmb{p}_i^{intra}\}_{i=1}^\tau$ . + +As this potential loss of information will decrease the overall performance, the (learnable) prompt parameters will be updated to prevent this from happening. Otherwise, if (common) knowledge patterns regarding input-agnostic patterns of missing modalities are (incorrectly) incorporated + +into $\left\{\pmb{p}_i^{inter}\right\}_{i=1}^{\tau}$ , their modeling bandwidth for inter-client heterogeneities will be reduced, which will also hamper the performance. This will, in turn, generate a prompt update signal to avoid this as well. Furthermore, within each set of inter- and intra-client prompts, there might also be different patterns that need to be separated and internalized into different subset of prompts. + +To enable this, we develop an input-adaptive learning mechanism for both sets of prompts which identifies and updates the most relevant subset of prompts for each (multimodal) input sample. This is formulated as an information-retrieval task, learning a pair of key and query functions, $k(\pmb{p})$ and $q(\pmb{x}(M))$ , which transform the prompt $\pmb{p}$ and input $\pmb{x}$ with the set $M$ observed modalities, respectively, into a shared metric space. We can then define a geometric input-prompt distance, e.g., $d(\pmb{x}(M), \pmb{p}) = \cos(q(\pmb{x}(M)), k(\pmb{p}))$ which can be used to measure their contextual relevance. To avoid loading diverse fine-tuning knowledge into a single prompt, we regularize the (local) loss function to penalize it when a less relevant prompt is selected as fine-tuning instruction for an input $\pmb{x}(M)$ : + +$$ +\boldsymbol {w} _ {t} \triangleq \underset {\boldsymbol {w}} {\arg \min } L _ {t} ^ {\prime} (\boldsymbol {w}) \quad \text {w h e r e} \tag {7} +$$ + +$$ +\begin{array}{l} L _ {t} ^ {\prime} (\boldsymbol {w}) \triangleq \sum_ {s = 1} ^ {m} \left\{\ell \left(F \left(\boldsymbol {x} \left(M _ {t, s}\right); \boldsymbol {w} ^ {\prime}\right), z _ {t, s}\right) \right\} \\ + \sum_ {s = 1} ^ {m} \left\{r \left(\boldsymbol {x} \left(M _ {t, s}\right), \boldsymbol {w} _ {p} ^ {\prime}\right) \right\}, \tag {8} \\ \end{array} +$$ + +where $\ell (\pmb {w}^{\prime})$ is defined in Eq. (1) and $\pmb{w}^{\prime} = (\pmb{w}_{c},\pmb{w}_{p}^{\prime})$ denote the customized set of tuning instructions for each + +![](images/d354035796080b754dc6c8bf5c32f2276441d652b5d4db4b27aa301009bf8fe3.jpg) +Figure 3. Workflow of the prompt-alignment algorithm. At each iteration, each client samples a subset of summarizing prompts using its query and key functions. The client then performs a local update, resulting in heterogeneous prompt sets, which are subsequently sent to the server to be clustered and summarized into new summarizing prompt sets for the next iteration. + +input $\pmb{x}(M_{t,s})$ where $\pmb{w}_p' = (\pmb{w}_p^{inter'}, \pmb{w}_p^{intra'})$ such that $\pmb{w}_p^{inter'} \subseteq \pmb{w}_p^{inter}$ and $\pmb{w}_p^{intra'} \subseteq \pmb{w}_p^{intra}$ are the sets of $\kappa$ most relevant prompts to $\pmb{x}(M_{t,s})$ in $\pmb{w}_p^{inter}$ and $\pmb{w}_p^{intra}$ . The regularizer $r(\pmb{x}(M), \pmb{w}_p')$ is then defined as + +$$ +r \left(\boldsymbol {x} (M), \boldsymbol {w} _ {p} ^ {\prime}\right) \triangleq \sum_ {\boldsymbol {p} \in \boldsymbol {w} _ {p} ^ {\prime}} \left\{d (\boldsymbol {x} (M), \boldsymbol {p}) \right\}. \tag {9} +$$ + +Intuitively, the regularizer penalizes the total distance between the (multimodal) input $\pmb{x}(M)$ and the $\kappa$ closest prompts according to the aforementioned distance metric $d(\pmb{x}(M), \pmb{p})$ . Minimizing both the original local loss and this regularizer allows the local model to simultaneously learn to (1) internalize fine-tuning patterns of inter- and intra-heterogeneities to minimize the original loss $L_{t}(\pmb{w})$ ; and (2) ensure that selected prompts for a given input will be close to it to reduce the penalty $r(\pmb{x}(M), \pmb{w}_{p}^{\prime})$ . A pseudocode of this client design is detailed in Suppl. A. + +# 3.3. Server Aggregation + +We will now establish a principled algorithm to align and aggregate inter-client prompts specializing in similar input-level patterns of missing data2. This is not straightforward since inter-client prompts at the same position in the prompt sets across different clients might not specialize for similar patterns of missing data due to client heterogeneities: certain missing-data patterns are not observed at certain clients. + +To find the correct alignment, we formulate it as a clustering optimization task. Suppose each client $t$ communicates a set of inter-client prompts $\boldsymbol{w}_t^{inter} = (\boldsymbol{p}_t^i)_{i=1}^\tau$ . Each + +potential clustering of these prompts thus defines a candidate alignment, i.e., prompts in the same cluster are considered aligned and can be aggregated. A (learnable) cluster center can then be associated with the aggregated prompt - see Fig. 3. Both the cluster assignment and centers can be found via solving the following optimization task with discrete alignment variables $\alpha_{t}^{pq}\in \{0,1\}$ and continuous center variables $\theta_{q}$ for each cluster $q\in [n\times \tau ]$ .. + +$$ +\begin{array}{l} \min _ {\boldsymbol {\alpha}, \boldsymbol {\theta}, \gamma} G (\boldsymbol {\alpha}, \boldsymbol {\theta}, \gamma) \triangleq \sum_ {t = 1} ^ {n} \sum_ {p = 1} ^ {\tau} \sum_ {q = 1} ^ {n \times \tau} \alpha_ {t} ^ {p q} \cdot c \left(\boldsymbol {p} _ {t} ^ {p}, \boldsymbol {\theta} _ {q},; \gamma\right) (10) \\ \mathrm {s . t .} \alpha_ {t} ^ {p, 1} + \dots + \alpha_ {t} ^ {p, n \times \tau} = 1, (11) \\ \end{array} +$$ + +where $\alpha \triangleq (\alpha_{t}^{p,q})$ and $\alpha_{t}^{p,q} \in \{0,1\} \forall (t,p,q)$ indicates whether the $p$ -th inter-client prompt of client $t$ is matched to the $q$ -th cluster. The corresponding cluster center is then optimized to have minimum total distance to all cluster members according to a (learnable) cost function $c(\pmb{p}_{t}^{p},\pmb{\theta}_{q})$ . The optimized cluster center can then be used as the aggregated or summarizing prompt. Minimizing Eq. (10) thus allows us to align and summarize/aggregate the inter-client prompts across clients simultaneously. Note that Eq. (11) enforces a constraint that intuitively indicates two prompts from the same clients should not be matched to the same cluster and get aggregated. This is because local prompts from the same client have already been learned to encode different finetuning aspects. Aggregating them will therefore collapse them into less informative ones. + +In addition, some summarizing/aggregated prompts might be more generic than others and can be used to provide fine-tuning instructions to a wider range of (multimodal) inputs. The update of such prompts should be prioritized during training and to achieve that, we need a mechanism to capture this prompt popularity. To achieve this, we augment the optimization task in Eq. (10) above with a regularizer that penalizes alignment to less popular prompts, + +$$ +\begin{array}{l} R (\boldsymbol {\alpha}, \zeta) \triangleq \sum_ {t = 1} ^ {n} \sum_ {p = 1} ^ {\tau} \sum_ {q = 1} ^ {n \times \tau} \left(\alpha_ {t} ^ {p, q} \cdot \log U \left(\boldsymbol {\theta} _ {q}; \zeta\right)\right) \tag {12} \\ + \sum_ {t = 1} ^ {n} \sum_ {p = 1} ^ {\tau} \sum_ {q = 1} ^ {n \times \tau} \Big (1 - \alpha_ {t} ^ {p, q} \Big) \cdot \log \Big (1 - U (\boldsymbol {\theta} _ {q}; \zeta) \Big), \\ \end{array} +$$ + +where $U(\pmb{\theta}_q;\zeta)$ is the popularity function of $\pmb{\theta}_q$ , which is parameterized by learnable parameter $\zeta$ . For example, $U(\pmb{\theta}_q;\zeta) = \sigma(g(\pmb{\theta}_q;\zeta)) \in (0,1)^3$ where $g(\pmb{\theta}_q;\zeta)$ is a deep neural net parameterized with $\zeta$ . Putting Eq. (13) and Eq. (10) together, we arrive at our final loss function, + +$$ +\underset {\boldsymbol {\alpha}, \boldsymbol {\theta}, \gamma , \zeta} {\text {m i n i m i z e}} \left\{G (\boldsymbol {\alpha}, \boldsymbol {\theta}, \gamma) + R (\boldsymbol {\alpha}, \zeta) \right\}, \tag {13} +$$ + +which can be optimized via alternating between (1) optimizing $(\theta, \zeta, \gamma)$ while fixing $\alpha$ ; and (2) optimizing $\alpha$ while fixing $(\theta, \zeta, \gamma)$ . The first optimization sub-task is straightforward while the latter is less trivial due to the constraint as well as the discrete nature of the optimizing variables. To sidestep the technical challenge in solving for $\alpha$ while fixing $(\theta, \zeta, \gamma)$ , we approximate it by optimizing for $\alpha_{t} = \{\alpha_{t}^{p,q}\}_{p,q}$ while also fixing $\alpha_{-t} = \{\alpha_{-t}^{p,q}\}_{p,q}$ . This results in a minimum weighted linear sum task, + +$$ +\begin{array}{l} \min _ {\boldsymbol {\alpha} _ {t}} L \left(\boldsymbol {\alpha} _ {t}\right) \triangleq \sum_ {p = 1} ^ {\tau} \sum_ {q = 1} ^ {n \times \tau} \left(\alpha_ {t} ^ {p, q} \cdot v (p, q)\right) \text {w h e r e} \tag {14} \\ v (p, q) \triangleq \left(\log \left\{\frac {U (\boldsymbol {\theta} _ {q} ; \zeta)}{1 - U (\boldsymbol {\theta} _ {q} ; \zeta)} \right\} + c (\boldsymbol {p} _ {t} ^ {p}, \boldsymbol {\theta} _ {q}; \gamma)\right), \\ \end{array} +$$ + +which can be solved effectively in polynomial time using the Hungarian algorithm [23]. An overview and pseudocode of the server-side aggregation algorithm are presented in Fig. 3 and Suppl. A. A thorough evaluation of this framework is detailed next in Section 4. + +# 4. Empirical Evaluation + +This section provides a detailed empirical evaluation of our proposed framework, FED-PRIME. Our evaluation is conducted on two datasets UPMC Food-101 [46] and MM-IMDB [1] and several baselines adapting existing works in federated fine-tuning (for unimodal data) and multi-modal prompt-tuning (for centralized data) to our novel setting on decentralized multi-modal data with missing modalities and features. Detailed description of the datasets and baselines are provided in Section 4.1. Our main results and ablation studies are reported in Section 4.2. + +# 4.1. Experiment Settings + +A. Datasets. Following the experiment protocol in [24], we evaluate FED-PRIME on two multimodal datasets: + +1. UPMC Food-101 [46] is an image-text classification dataset consisting of noisy image-text pairs collected from Google Image Search. Its classes are aligned with those in the larger, publicly available ETHZ Food-101 dataset [4]. For the federated setting, we select only eight most frequent classes from the original UPMC Food-101. +2. MM-IMDB [1] is a movie genre classification dataset with both image and text modalities. Since each movie can belong to multiple genres, the task is originally a multi-label classification problem, where genres are predicted using image, text, or both. For the federated setting, we use only single-label instances and select eight most frequent classes only from the original MM-IMDB. + +Each dataset is first partitioned into train and test sets with the 80:20 ratio. The (heterogeneous) missing patterns on the train/test data are simulated using the following protocols. + +B. Training Data Simulation. Following [24], we simulate missing data phenomenon within a dataset using a missing rate $\eta \in (0,1)$ as detailed in 3 simulation cases below: + +1. Missing Image. We drop the image features in $\eta$ percent of the data and keep the remaining $(1 - \eta)$ percent complete. +2. Missing Text. We drop the text features in $\eta$ percent of the data and keep the remaining $(1 - \eta)$ percent complete. +3. Missing Both. We drop the image component in $\eta /2$ percent of the data, and the text component in the other $\eta /2$ percent, leaving the remaining $(1 - \eta)$ percent complete. + +The missing-simulated train set is then partitioned into $n$ subsets for $n$ clients via random sampling. The resulting client datasets will therefore have heterogeneous distributions over missing-data patterns. + +C. Test Data Simulation. There are 5 test scenarios, which include: (1) Full Modal (no missing data); (2) Image Only (drop all text features); (3) Text Only (drop all image features); (4) Miss Both (run Missing Both on test data); and (5) $\sim$ Train (same simulation as the training data). + +D. Metrics. Following prior protocol in [24], we use F1-macro and classification accuracy to evaluate the following baseline methods on MM-IMDB and UPMC Food-101. + +E. Baseline Methods. Following [24], we adopt the pretrained multimodal transformer ViLT [21] as the backbone of our client design. Each client model is a fine-tuned version of this backbone which is generated based on its own local dataset. As no prior work has addressed modality missing in multimodal fine-tuning within federated learning settings, our comparison baselines mostly feature adapted versions of existing work in federated learning and multimodal prompt-tuning. This includes adaptations of the latest missing prompt-tuning SOTA method in [24] into the context of the classical FedAvg method as well as that of a more recent FL approach, FedMSplit [8]. These are referred to as FEDAVG-P and FEDMSPLIT-P, respectively. In addition, to underscore the importance of separating fine-tuning knowledge across intra- and inter-client prompt sets, we also include two simplified variants of FED-PRIME utilizing either inter-client prompt pooling or intra-client prompt pooling, along with their respective aggregations. We refer to these variants as FED-INTER and FED-INTRA, respectively (see more details in Supplementary B). + +# 4.2. Experimental Results + +This section provides extensive performance comparison between FED-PRIME and other baseline methods across various settings of missing data in the train and test sets, as previously stated in Section 4.1. In addition, we also provide + +Table 1. Performance achieved by the baselines on the Food-101 (left) and MM-IMDB (right) datasets under various missing-data settings in the train and test sets. This includes all combinations of the 3 train scenarios and 5 test scenarios (i.e., 15 experiments) per dataset, as previously stated in Section 4.1. The best and second-best results are highlighted in bold red, and blue, respectively. Note that in the Miss Both training scenarios, Test ( $\sim$ Train) and Test (Miss Both) represent the same test scenario according to the definition of $\sim$ Train. The reported result for Test (Miss Both) in this case is the same as that for Test ( $\sim$ Train), as indicated by a dash (-). + +
DatasetsUPMC Food-101MM-IMDB
TrainMethodTest(∼ Train)Test(Miss Both)Test(Full Modal)Test(Text only)Test(Image only)Test(∼ Train)Test(Miss Both)Test(Full Modal)Test(Text only)Test(Image only)
Miss TextFEDAVG-P15.71 ± 2.2714.90 ± 1.5721.56 ± 7.8116.91 ± 0.6915.36 ± 0.4322.42 ± 2.2721.89 ± 1.3430.78 ± 1.6518.40 ± 0.0614.53 ± 4.56
FEDMSPLIT-P15.62 ± 1.5117.50 ± 1.9725.27 ± 8.2218.78 ± 0.6417.50 ± 1.9721.02 ± 1.8919.97 ± 0.7424.39 ± 6.0814.38 ± 2.7618.09 ± 6.13
FED-INTER54.82 ± 19.0148.87 ± 24.6459.17 ± 27.0635.13 ± 26.7856.59 ± 15.1218.25 ± 3.5016.95 ± 3.5718.67 ± 7.6315.03 ± 4.6618.01 ± 1.95
FED-INTRA61.71 ± 17.2248.09 ± 19.1262.06 ± 26.9822.51 ± 5.9262.64 ± 11.8313.38 ± 1.7312.77 ± 0.8512.55 ± 1.6711.31 ± 0.3814.33 ± 1.80
FED-PRIME78.88 ± 0.9080.38 ± 0.6592.12 ± 0.4073.01 ± 4.2576.83 ± 1.2231.92 ± 0.2031.48 ± 0.3037.67 ± 0.0430.60 ± 1.4130.69 ± 1.41
Improv. (%)27.82 ↑64.48 ↑48.44 ↑107.83 ↑22.65 ↑42.37 ↑43.81 ↑22.35 ↑66.30 ↑69.65 ↑
Miss ImageFEDAVG-P17.35 ± 4.7715.12 ± 1.4816.84 ± 2.3718.12 ± 6.4914.81 ± 0.2427.69 ± 5.9722.55 ± 3.0631.94 ± 0.9823.76 ± 11.7212.29 ± 0.47
FEDMSPLIT-P74.16 ± 10.5648.88 ± 10.2645.64 ± 32.4388.65 ± 2.1714.81 ± 0.9019.11 ± 11.3316.61 ± 7.2218.19 ± 12.5518.12 ± 12.3012.81 ± 1.25
FED-INTER77.96 ± 11.6264.62 ± 10.2282.08 ± 7.7577.69 ± 12.3537.56 ± 6.4918.79 ± 5.2317.93 ± 3.6020.56 ± 3.5617.67 ± 6.7915.47 ± 2.51
FED-INTRA22.84 ± 3.5220.13 ± 1.7223.48 ± 1.8624.46 ± 2.9916.66 ± 1.3215.75 ± 4.3414.06 ± 3.1615.68 ± 4.7714.53 ± 3.6511.71 ± 0.42
FED-PRIME90.55 ± 0.2279.12 ± 0.4992.89 ± 0.2190.18 ± 0.2954.14 ± 2.5036.08 ± 0.3531.35 ± 0.6138.49 ± 0.5636.91 ± 0.5918.15 ± 0.66
Improv. (%)16.15 ↑22.44 ↑13.17 ↑1.73 ↑44.14 ↑30.30 ↑39.02 ↑20.51 ↑55.35 ↑17.32 ↑
Miss BothFEDAVG-P14.57 ± 1.50-17.17 ± 4.3716.40 ± 4.0513.24 ± 0.3226.45 ± 2.63-33.03 ± 2.5624.12 ± 11.3020.21 ± 1.98
FEDMSPLIT-P49.15 ± 24.76-64.78 ± 36.6264.62 ± 36.5121.49 ± 7.1924.25 ± 5.02-26.05 ± 11.1726.02 ± 9.6419.79 ± 6.20
FED-INTER56.32 ± 21.77-69.57 ± 19.4145.15 ± 34.0959.30 ± 10.8426.53 ± 0.90-31.97 ± 2.2229.69 ± 2.2121.63 ± 0.77
FED-INTRA49.28 ± 32.87-56.70 ± 37.9043.24 ± 34.1949.85 ± 25.4411.90 ± 0.37-12.47 ± 0.4511.46 ± 0.3312.83 ± 0.92
FED-PRIME84.44 ± 2.65-93.64 ± 0.5887.95 ± 0.9172.41 ± 3.8832.01 ± 2.51-38.68 ± 0.6531.00 ± 2.9726.01 ± 0.12
Improv. (%)49.93 ↑-34.60 ↑36.10 ↑22.11 ↑20.66 ↑-17.11 ↑4.41 ↑20.25 ↑
+ +(*) Improv. shows the relative performance improvement between our proposal and the second-best. (in percentage). + +ablation studies on (1) performance robustness against various missing rate $\eta$ ; (2) the ablated impact of having inter- and intra-client prompt sets; and (3) training convergence. + +# 4.2.1. Main Results + +We compare the performance of FED-PRIME against the other baselines on an extensive set of experiments across all datasets. This includes all combinations of the 3 missing simulation scenarios on train data and 5 test scenarios (with both missing and no missing data) as previously detailed in parts B and C of Section 4.1. All results are reported in Table 1 which show that FED-PRIME consistently outperforms all baselines in all train/test scenarios with various missing-data patterns. The performance on the UPMC Food-101 and MM-IMDB datasets is measured in terms of classification accuracy and F1-macro, respectively. The F1-macro performance on MM-IMDB is relatively low across all settings due to the class imbalance nature of the dataset. In each experiment, the relative performance improvement of FED-PRIME over the second best baseline is reported, which ranges between $1.73\%$ and $107.83\%$ for the UPMC Food-101 dataset and between $4.41\%$ to $69.65\%$ for the MM-IMDB dataset. Notably, FED-PRIME significantly outperforms both of its simplified variants FED-INTRA and FED-INTER, which either handle intra- or interheterogeneities, in all settings. This highlights the remarkable synergies and essential of handling both types of heterogeneities in multi-modal federated learning with missing data. In addition, it can also be observed that in more than half of the experiment settings, the second best base + +line is either FED-INTRA or FED-INTER which significantly outperform both FEDAVG-P and FEDMSPLIT-P in those cases. These observations complement our findings of the isolated impact of these components in our ablation studies in Section 4.2.4. + +# 4.2.2. Impact of Missing Modalities. + +Across different testing scenarios, we observe that the highest test performance is always achieved when there is no missing data in the test set as expected. Another interesting observation is that in the Miss Text training scenario, where approximately $70\%$ of the text features are lost on the training data, the proposed model still performs surprisingly well on the Text Only test scenarios. This suggests that the multi-modal prompt alignment is effective in recovering text information, highlighting the proposed design's adaptability. However, in the Miss Image training scenario, the test performance on the Image Only case suffers instead a significant performance drop. Upon investigation, we believe this is likely due to the VILT backbone's lightweight image processing module, which was optimized to work in tandem with text for competitive performance, hinting at a potential text-centered nature of VILT during pre-training. This is in fact well-aligned with existing observations that co-training with text often leads to improved performance. + +# 4.2.3. Robustness Against Various Missing Rates. + +To demonstrate the robustness of FED-PRIME against different missing rates $\eta$ , we run multiple comparisons of FED-PRIME, FEDAVG-P, and its centralized variant [24] + +![](images/a8c4a2b4d053f7b04dca0657b87f5b136b5b6deaf01c181c0cc83ea1db0121ba.jpg) +Figure 4. Performance comparison on UMPC Food-101 under Miss Both scenarios with various missing rates. + +![](images/96ad7d5f6bc0f66a2028428fce690da69fc4747b2de093344daa46f39ffb7128.jpg) +Figure 5. Plots of the train/test loss convergence on Food-101 under Miss-Text.FED-PRIME converges fastest. + +![](images/814c2782251785cb72023581f7eb545e4c8ad9679dd6dc8d3d8c8a4956ee4f2f.jpg) +Figure 6. Convergence of Euclidean distance between aggregated prompt centroids on the UPMC Food-101 under Miss Image scenario. + +across a wide range of $\eta \in \{0.00, 0.25, 0.50, 0.75, 1.00\}$ under the Miss Both training scenarios. The reported results in Fig. 1 and Fig. 4 show that our model significantly outperforms FEDAVG-P and its centralized counterpart Centralized-P. As the missing rate increases, our approach consistently maintains a competitive accuracy above $80\%$ , closely approaching the upper-bound performance of FEDAVG-P in the centralized setting, which benefits from a centralized dataset with better-aligned modalities. These results demonstrate the robustness of our model across varying missing rates and highlight the effectiveness of our federated prompt-tuning approach, achieving performance comparable to a centralized prompt-tuning model despite having access only to decentralized local datasets with heterogeneous missing patterns. + +# 4.2.4. Ablation Study + +As explained previously in Section 4.1, FED-INTER and FED-INTRA are in variants of FED-PRIME when we drop the intra- and inter-client prompt sets from the client design, respectively. Table 1 shows that FED-PRIME consistently outperforms both FED-INTER and FED-INTRA in all test conditions with substantial gap. The lack of intra-client prompts causes FED-INTER to underperform FED-PRIME suggests that having a dedicated prompt set to capture the common and highly generalizable input-agnostic aspect of fine-tuning is essential to improve performance. Likewise, FED-INTRA, which omits the inter-client prompt set, also shows a significant performance drop. This highlights the importance of capturing and correctly aggregating input-level patterns of missing data through prompt alignment. + +# 4.2.5. Convergence Analysis + +Model Convergence. Fig. 5 shows that FED-PRIME has superior performance, achieving faster convergence and significantly lower final loss values in both training and testing compared to FED-INTER and FED-INTRA. This indicates better learning efficiency and robustness. Both FED-INTER and FED-INTRA converge more slowly, with + +consistently higher loss values, especially in the early rounds. Additionally, our method is more stable across server-aggregation rounds, with minimal fluctuations. In contrast, FED-INTER and FED-INTRA exhibit significantly more variability, particularly toward the end. This highlights the effectiveness of FED-PRIME, with faster convergence and improved stability across rounds. + +Prompt Convergence. Fig. 6 shows that FED-PRIME reduces the distance between the aggregated prompt centroids much faster than FEDAVG-P, which highlights the faster convergence rate of FED-PRIME over FEDAVG-P, indicating faster convergence and better prompt alignment. In contrast, FEDAVG-P converges more slowly and unstably, likely due to incorrect default prompt alignment. This again underscores the necessity of having a learnable prompt clustering mechanism to prevent incorrect prompt alignment. + +# 5. Conclusion + +Fine-tuning large pre-trained models for specific tasks has become an increasingly effective solution paradigm in machine learning. However, existing fine-tuning frameworks struggle in practical, multimodal settings where data is distributed across private silos with heterogeneous patterns of missing modalities and features. To address these challenges, this paper introduces a novel federated prompt-tuning approach that optimizes, aligns, and aggregates prompt instructions, compensating for gaps in data modalities and distributional patterns across a number of diverse and private local datasets. Rigorous evaluation on multimodal benchmark datasets demonstrates the effectiveness and robustness of our approach against various patterns of missing data in local datasets. We view this work as an initial step toward developing resilient, adaptable federated models that can fully leverage the diversity of multimodal data in real-world environments. + +# Acknowledgement + +This research is funded by Hanoi University of Science and Technology (HUST) under grant number T2024-TD-002. Additionally, this work utilized GPU compute resource at SDSC and ACES through allocation CIS230391 from the Advanced Cyberinfrastructure Coordination Ecosystem: Services and Support (ACCESS) program [2], which is supported by U.S. National Science Foundation grants #2138259, #2138286, #2138307, #2137603, and #2138296. + +# References + +[1] John Arevalo, Thamar Solorio, Manuel Montes y Gómez, and Fabio A. González. Gated multimodal units for information fusion. ArXiv, abs/1702.01992, 2017. 2, 6, 3 +[2] Timothy J. Boerner, Stephen Deems, Thomas R. Furlani, Shelley L. Knuth, and John Towns. Access: Advancing innovation: Nsf's advanced cyberinfrastructure coordination ecosystem: Services & support. In Practice and Experience in Advanced Research Computing 2023: Computing for the Common Good, page 173-176, New York, NY, USA, 2023. Association for Computing Machinery. 9 +[3] Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S. Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, Erik Brynjolfsson, S. Buch, Dallas Card, Rodrigo Castellon, Niladri S. Chatterji, Annie S. Chen, Kathleen A. Creel, Jared Davis, Dora Demszky, Chris Donahue, Moussa Doumbouya, Esin Durmus, Stefano Ermon, John Etchemendy, Kawin Ethayarajh, Li Fei-Fei, Chelsea Finn, Trevor Gale, Lauren E. Gillespie, Karan Goel, Noah D. Goodman, Shelby Grossman, Neel Guha, Tatsunori Hashimoto, Peter Henderson, John Hewitt, Daniel E. Ho, Jenny Hong, Kyle Hsu, Jing Huang, Thomas F. Icard, Saahil Jain, Dan Jurafsky, Pratyusha Kalluri, Siddharth Karamcheti, Geoff Keeling, Fereshte Khani, O. Khattab, Pang Wei Koh, Mark S. Krass, Ranjay Krishna, Rohith Kuditipudi, Ananya Kumar, Faisal Ladhak, Mina Lee, Tony Lee, Jure Leskovec, Isabelle Levent, Xiang Lisa Li, Xuechen Li, Tengyu Ma, Ali Malik, Christopher D. Manning, Suvir P. Mirchandani, Eric Mitchell, Zanele Munyikwa, Suraj Nair, Avanika Narayan, Deepak Narayanan, Benjamin Newman, Allen Nie, Juan Carlos Niebles, Hamed Nilforoshan, J. F. Nyarko, Giray Ogut, Laurel Orr, Isabel Papadimitriou, Joon Sung Park, Chris Piech, Eva Portelance, Christopher Potts, Aditi Raghunathan, Robert Reich, Hongyu Ren, Frieda Rong, Yusuf H. Roohani, Camilo Ruiz, Jack Ryan, Christopher R'e, Dorsa Sadigh, Shiori Sagawa, Keshav Santhanam, Andy Shih, Krishna Parasuram Srinivasan, Alex Tamkin, Rohan Taori, Armin W. Thomas, Florian Tramér, Rose E. Wang, William Wang, Bohan Wu, Jiajun Wu, Yuhuai Wu, Sang Michael Xie, Michihiro Yasunaga, Jiaxuan You, Matei A. Zaharia, Michael Zhang, Tianyi Zhang, Xikun Zhang, Yuhui Zhang, Lucia Zheng, Kaitlyn Zhou, and Percy Liang. On the opportunities and risks of foundation models. ArXiv, 2021. 1 +[4] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. + +Food-101 - mining discriminative components with random forests. In European Conference on Computer Vision, 2014. 6 +[5] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. CoRR, abs/2005.14165, 2020. 1 +[6] Alberto Brunete, Ernesto Gambao, Miguel Hernando, and Raquel Cedazo. Smart assistive architecture for the integration of IoT devices, robotic systems, and multimodal interfaces in healthcare environments. Sensors, 21(6):2212, 2021. 1, 2 +[7] Jiayi Chen and Aidong Zhang. Hgmf: heterogeneous graph-based fusion for multimodal data with incompleteness. In Proceedings of the 26th ACM International Conference on Knowledge Discovery & Data Mining, pages 1295-1305, 2020. 3 +[8] Jiayi Chen and Aidong Zhang. FedMSplit: Correlation-adaptive federated multi-task learning across multimodal split networks. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 87-96, 2022. 1, 3, 6 +[9] Jinyu Chen, Wenchao Xu, Song Guo, Junxiao Wang, Jie Zhang, and Haozhao Wang. Fedtune: A deep dive into efficient federated fine-tuning with pre-trained transformers, 2022. 1 +[10] Yae Jee Cho, Andre Manoel, Gauri Joshi, Robert Sim, and Dimitrios Dimitriadis. Heterogeneous ensemble knowledge transfer for training large models in Federated Learning. Computing Research Repository arXiv Preprints, 2204.12703, 2022. 2 +[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 1 +[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota, 2019. Association for Computational Linguistics. 1 +[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 1 +[14] Alireza Fallah, Aryan Mokhtari, and Asuman Ozdaglar. Personalized federated learning: A meta-learning ap + +proach. Computing Research Repository arXiv Preprints, 2002.07948, 2020.2 +[15] Ziwei Fan, Hao Ding, Anoop Deoras, and Trong Nghia Hoang. Personalized federated domain adaptation for item-to-item recommendation. In Proceedings of the Thirty-Ninth Conference on Uncertainty in Artificial Intelligence, pages 560-570. PMLR, 2023. 2 +[16] Tao Guo, Song Guo, Junxiao Wang, Xueyang Tang, and Wenchao Xu. PromptFL: Let Federated Participants Cooperatively Learn Prompts Instead of Models-Federated Learning in Age of Foundation Model. IEEE Transactions on Mobile Computing, 2023. 1 +[17] Minh Hoang, Nghia Hoang, Bryan Kian Hsiang Low, and Carleton Kingsford. Collective model fusion for multiple black-box experts. In Proceedings of the 36th International Conference on Machine Learning, pages 2742-2750. PMLR, 2019. 2 +[18] Trong Nghia Hoang, Quang Minh Hoang, Kian Hsiang Low, and Jonathan P. How. Collective online learning of Gaussian processes in massive multi-agent systems. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7850-7857, 2019. 2 +[19] Parminder Kaur, Husanbir Singh Pannu, and Avleen Kaur Malhi. Comparative analysis on cross-modal information retrieval: A review. Computer Science Review, 39:100336, 2021. 2 +[20] Muhammad Uzair khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3 +[21] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 5583-5594. PMLR, 2021. 1, 3, 6 +[22] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A. Shamma, Michael S. Bernstein, and Li Fei-Fei. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International Journal of Computer Vision, 123:32 - 73, 2016. 1 +[23] Harold W Kuhn. The Hungarian Method for the Assignment Problem. Naval Research Logistics Quarterly, 2(1-2):83-97, 1955. 6 +[24] Yi-Lun Lee, Yi-Hsuan Tsai, Wei-Chen Chiu, and Chen-Yu Lee. Multimodal prompting with missing modalities for visual recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3, 6, 7, 1 +[25] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3045-3059, Online and Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 1 +[26] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. Proceedings of Machine Learning and Systems, 2:429-450, 2020. 2, 4 + +[27] Xiang Li, Kaixuan Huang, Wenhao Yang, Shusen Wang, and Zhihua Zhang. On the convergence of FedAvg on non-IID data. Computing Research Repository arXiv Preprints, 1907.02189t, 2019. 2 +[28] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, 2014. 1 +[29] Wang Lu, Xixu Hu, Jindong Wang, and Xing Xie. FedCLIP: Fast Generalization and Personalization for CLIP in Federated Learning, 2023. 1 +[30] Tengfei Ma, Trong Nghia Hoang, and Jie Chen. Federated learning of models pre-trained on different features with consensus graphs. In Uncertainty in Artificial Intelligence, pages 1336-1346, 2023. 2 +[31] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial Intelligence and Statistics, pages 1273-1282, 2017. 2, 3, 4, 5 +[32] John Nguyen, Jianyu Wang, Kshitiz Malik, Maziar Sanjabi, and Michael Rabbat. Where to Begin? On the Impact of Pre-Training and Initialization in Federated Learning, 2023. 1 +[33] Manh Duong Nguyen, Trung Thanh Nguyen, Huy-Hieu Pham, Trong Nghia Hoang, Phi Le Nguyen, and Thanh Trung Huynh. Fedmac: Tackling partial-modality missing in federated learning with cross-modal aggregation and contrastive regularization. 2024. 1, 3 +[34] Nang Hung Nguyen, Phi Le Nguyen, Thuy Dung Nguyen, Trung Thanh Nguyen, Duc Long Nguyen, Thanh Hung Nguyen, Huy Hieu Pham, and Thao Nguyen Truong. FedDRL: Deep reinforcement learning-based adaptive aggregation for non-IID data in federated learning. In Proceedings of the 51st International Conference on Parallel Processing, pages 1-11, 2022. 2 +[35] Nang Hung Nguyen, Duc Long Nguyen, Trong Bang Nguyen, Thanh-Hung Nguyen, Huy Hieu Pham, Truong Thao Nguyen, and Phi Le Nguyen. CADIS: Handling cluster-skewed Non-IID data in federated learning with clustered aggregation and knowledge distilled regularization. In Proceedings of the 23rd IEEE/ACM International Symposium on Cluster, Cloud and Internet Computing, pages 249-261, 2023. +[36] Quang Ha Pham, Nang Hung Nguyen, Thanh Hung Nguyen, Huy Hieu Pham, Phi Le Nguyen, and Truong Thao Nguyen. SEM: A simple yet efficient model-agnostic local training mechanism to tackle data sparsity and scarcity in federated learning. In Proceedings of the 11st International Symposium on Computing and Networking, pages 120–126, 2023. 2 +[37] Thu Hang Phung, Binh P Nguyen, Thanh Hung Nguyen, Quoc Viet Hung Nguyen, Phi Le Nguyen, and Thanh Trung Huynh. A contrastive learning and graph-based approach for missing modalities in multimodal federated learning. In Proceedings of the 2024 International Joint Conference on Neural Networks, pages 1-8, 2024. 1, 3 + +[38] Petra Poklukar, Miguel Vasco, Hang Yin, Francisco S Melo, Ana Paiva, and Danica Kragic. Geometric multimodal contrastive representation learning. In Proceedings of the 2022 International Conference on Machine Learning, pages 17782-17800, 2022. 3 +[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. CoRR, abs/2103.00020, 2021. 1 +[40] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21(140):1-67, 2020. 1 +[41] Kuniaki Saito, Kihyuk Sohn, Xiang Zhang, Chun-Liang Li, Chen-Yu Lee, Kate Saenko, and Tomas Pfister. Prefix conditioning unifies language and label supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2861-2870, 2023. 1, 3 +[42] Rachael Hwee Ling Sim, Yehong Zhang, Trong Nghia Hoang, Xinyi Xu, Bryan Kian Hsiang Low, and Patrick Jaillet. Incentives in private collaborative machine learning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2 +[43] Canh T Dinh, Nguyen Tran, and Josh Nguyen. Personalized federated learning with Moreau envelopes. Advances in Neural Information Processing Systems, 33:21394-21405, 2020. 2 +[44] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, S. M. Ali Eslami, Oriol Vinyals, and Felix Hill. Multimodal few-shot learning with frozen language models. In Advances in Neural Information Processing Systems, pages 200-212. Curran Associates, Inc., 2021. 1, 3 +[45] Feng Wang, M Cenk Gursoy, and Senem Velipasalar. Communication-efficient and privacy-preserving feature-based federated transfer learning. In Proceedings of the 2022 IEEE Global Communications Conference, pages 3875-3880, 2022. 2 +[46] Xin Wang, Devinder Kumar, Nicolas Thome, Matthieu Cord, and Frédéric Precioso. Recipe recognition with large multimodal food dataset. In 2015 IEEE International Conference on Multimedia & Expo Workshops (ICMEW), pages 1-6, 2015. 2, 6, 3 +[47] Pei-Yau Weng, Minh Hoang, Lam M. Nguyen, My T. Thai, Tsui-Wei Weng, and Trong Nghia Hoang. Probabilistic federated prompt-tuning with non-IID and imbalanced data. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 1 +[48] Baochen Xiong, Xiaoshan Yang, Fan Qi, and Changsheng Xu. A unified framework for multi-modal Federated Learning. Neurocomputing, 480:110-118, 2022. 2 +[49] Baochen Xiong, Xiaoshan Yang, Yaguang Song, Yaowei Wang, and Changsheng Xu. Client-adaptive cross-model reconstruction network for modality-incomplete multimodal federated learning. In Proceedings of the 31st ACM Interna- + +tional Conference on Multimedia, pages 1241-1249, 2023. 3 +[50] Fu-En Yang, Chien-Yi Wang, and Yu-Chiang Frank Wang. Efficient Model Personalization in Federated Learning via Client-Specific Prompt Generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 19159-19168, 2023. 1 +[51] Guan Yu, Quefeng Li, Dinggang Shen, and Yufeng Liu. Optimal sparse linear prediction for block-missing multimodality data without imputation. Journal of the American Statistical Association, 115(531):1406-1419, 2020. 3 +[52] Songcan Yu, Junbo Wang, Walid Hussein, and Patrick C.K. Hung. Robust multimodal federated learning for incomplete modalities. Computer Communications, 214:234-243, 2024. 1, 3 +[53] Mikhail Yurochkin, Mayank Agarwal, Soumya Ghosh, Kristjan Greenewald, Trong Nghia Hoang, and Yasaman Khazaeni. Bayesian nonparametric federated learning of neural networks. In International Conference on Machine Learning, 2019. 2 +[54] Mikhail Yurochkin, Mayank Argawal, Soumya Ghosh, Kristjan Greenewald, and Trong Nghia Hoang. Statistical model aggregation via parameter matching. In Advances in Neural Information Processing Systems, pages 10954-10964, 2019. 2 +[55] Changqing Zhang, Yajie Cui, Zongbo Han, Joey Tianyi Zhou, Huazhu Fu, and Qinghua Hu. Deep partial multi-view learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(5):2402-2415, 2020. 3 +[56] Jianyi Zhang, Hao Frank Yang, Ang Li, Xin Guo, Pu Wang, Haiming Wang, Yiran Chen, and Hai Li. Mllm-fl: Multimodal large language model assisted federated learning on heterogeneous and long-tailed data. arXiv preprint arXiv:2409.06067, 2024. 1, 2 +[57] Wei Zhang and Xiang Li. Data privacy preserving federated transfer learning in machinery fault diagnostics using prior distributions. Structural Health Monitoring, 21(4):1329-1344, 2022. 2 +[58] Haodong Zhao, Wei Du, Fangqi Li, Peixuan Li, and Gongshen Liu. FedPrompt: Communication-Efficient and Privacy Preserving Prompt Tuning in Federated Learning. In Proc. ICASSP, 2023. 1 +[59] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 3 +[60] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision (IJCV), 2022. 1, 3 +[61] Tongxue Zhou, Pierre Vera, Stephane Canu, and Su Ruan. Missing data imputation via conditional generator and correlation learning for multimodal brain tumor segmentation. Pattern Recognition Letters, 158:125-132, 2022. 3 \ No newline at end of file diff --git a/2025/Federated Prompt-Tuning with Heterogeneous and Incomplete Multimodal Client Data/images.zip b/2025/Federated Prompt-Tuning with Heterogeneous and Incomplete Multimodal Client Data/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b96a5685e18fc5a7934fe226b24fa78ad2399e13 --- /dev/null +++ b/2025/Federated Prompt-Tuning with Heterogeneous and Incomplete Multimodal Client Data/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c206aacc5256a018e8581ebfda3665eff0824ebc87b1a1da7abc4e9777b8daaa +size 542158 diff --git a/2025/Federated Prompt-Tuning with Heterogeneous and Incomplete Multimodal Client Data/layout.json b/2025/Federated Prompt-Tuning with Heterogeneous and Incomplete Multimodal Client Data/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4a67e9919c488a7b10abcb92667f173e5173ef83 --- /dev/null +++ b/2025/Federated Prompt-Tuning with Heterogeneous and Incomplete Multimodal Client Data/layout.json @@ -0,0 +1,9123 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 111, + 103, + 501, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 103, + 501, + 138 + ], + "spans": [ + { + "bbox": [ + 111, + 103, + 501, + 138 + ], + "type": "text", + "content": "Federated Prompt-Tuning with Heterogeneous and Incomplete Multimodal Client Data" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 160, + 536, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 160, + 536, + 219 + ], + "spans": [ + { + "bbox": [ + 70, + 160, + 536, + 219 + ], + "type": "text", + "content": "Thu Hang Phung1 Duong M. Nguyen1 Thanh Trung Huynh2 Quoc Viet Hung Nguyen3 \nTrong Nghia Hoang4* Phi Le Nguyen1* \n1 Institute of AI Innovation and Societal Impact, Hanoi University of Science and Technology \n2 EPFL 3 Griffith University 4 Washington State University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 270, + 297, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 297, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 297, + 486 + ], + "type": "text", + "content": "This paper introduces a generalized federated prompt-tuning framework for practical scenarios where local datasets are multi-modal and exhibit different distributional patterns of missing features at the input level. The proposed framework bridges the gap between federated learning and multi-modal prompt-tuning which have traditionally focused on either uni-modal or centralized data. A key challenge in this setting arises from the lack of semantic alignment between prompt instructions that encode similar distributional patterns of missing data across different clients. To address this, our framework introduces specialized client-tuning and server-aggregation designs that simultaneously optimize, align, and aggregate prompt-tuning instructions across clients and data modalities. This allows prompt instructions to complement one another and be combined effectively. Extensive evaluations on diverse multimodal benchmark datasets demonstrate that our work consistently outperforms state-of-the-art (SOTA) baselines." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 496, + 135, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 135, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 135, + 509 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 516, + 297, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 297, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 297, + 696 + ], + "type": "text", + "content": "The recent emergence of exceedingly large models [3] with immense scale and versatility due to being pre-trained on large swaths of generic data has resulted in a paradigm shift in machine learning (ML). Most task-specific models are now fine-tuned versions of those large models instead of being created from de novo learning processes. For example, most state-of-the-art (SOTA) models solving specific NLP tasks are fine-tuned versions of existing large pre-trained or foundation models such as BERT [12], GPT [5], and T5 [40]. Likewise, SOTA models for various computer vision tasks have also been derived from foundation models such as the vision transformer [13] pre-trained on extensive datasets like ImageNet [11]. The standard fine-tuning approach, however, assumes centralized data which is not always possible in practical scenarios where fine-tuning data" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 246, + 523, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 246, + 523, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 523, + 258 + ], + "type": "text", + "content": "are scattered across multiple private siloed systems." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 258, + 555, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 555, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 555, + 413 + ], + "type": "text", + "content": "To overcome this siloed data challenge, federated fin-tuning has recently been considered to aggregate the fin-tuned updates across private devices running their own (local) fine-tuning processes [9, 32, 47]. This approach can be further combined with an existing parameter-efficient tuning method called prompt tuning [25]. This method focuses on creating lightweight signals such as additional tokens that are prepended to the input of the pre-trained models [41, 44, 59, 60], providing them valuable contextual information to perform specific tasks more effectively. Local sets of prompts can then be aggregated using the conventional parameter aggregation procedure in FL [16, 29, 47, 50, 58]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 415, + 556, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 556, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 556, + 557 + ], + "type": "text", + "content": "However, such federated prompt-tuning techniques are restricted to learning with uni-modal data. In more practical scenarios, advances in sensor and device manufacturing increasingly enable the capture of multiple data modalities (e.g., text, images), leading to richer datasets for fine-tuning [6]. Leveraging such diverse multi-modal information can significantly enhance model performance, as demonstrated by the success of multi-modal foundation models like CLIP [39] and ViLT [21]. This has further motivated research on multi-modal federated learning [8, 33, 37, 52] and fine-tuning pre-trained multi-modal LLMs, i.e., multi-modal federated fine-tuning [56]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 558, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 557, + 715 + ], + "type": "text", + "content": "Despite these advances, multi-modal federated learning research often adopts specific local architectures which cannot be learned via fine-tuning existing pre-trained multimodal foundation models [8, 33, 37, 52]. Existing multimodal federated fine-tuning research on the other hand assumes that all local datasets have the same set of modalities [56]. As such, these approaches are not suitable for situations where different clients might collect data using different types of sensors. Furthermore, the multi-modal sensor data across clients might even have different missing patterns of data entries. This leads to local datasets with both intra- and inter-heterogeneities. Intra-heterogeneity occurs when individual samples within a single dataset" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 149, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 149, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 149, + 713 + ], + "type": "text", + "content": "*Corresponding Authors" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3936" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": "have missing entries for different subsets of features. Interheterogeneity arises when different clients have different distributions over their data-missing patterns." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 110, + 295, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 110, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 54, + 110, + 295, + 277 + ], + "type": "text", + "content": "There also exists recent work in prompt-tuning that learns a specific prompt for each subset of missing modalities [24]. Input samples having the same set of missing modalities are then augmented with the same (learnable) prompt during training, specializing it to complement for the corresponding missing modalities. This approach is, however, restricted to centralized data setting (i.e., no interheterogeneities) and does not generalize well to federated settings where different sets of local prompts might be biased towards different distributional patterns of missing data. A naive aggregation of these prompts might therefore collapse them into less informative prompts which are ignorant of the heterogeneities in missing-data patterns across clients, resulting in poor performance (see Fig. 1)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 280, + 295, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 280, + 295, + 423 + ], + "spans": [ + { + "bbox": [ + 55, + 280, + 295, + 423 + ], + "type": "text", + "content": "To bridge this gap between multi-modal prompt-tuning and federated learning, we propose a multimodal federated prompt-tuning approach that optimizes and aggregates two distinct sets of (local) prompts for each client: inter-client and intra-client prompts. Inter-client prompts breakdown and encode prototypical distribution patterns of missing input data, which can be aligned and aggregated across clients to represent shared distributional information. In contrast, intra-client prompts encode information tailored to support specific local subsets of missing modalities, independent of the distributional patterns of missing data at each client. This is achieved via the following contributions:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 428, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 428, + 295, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 428, + 295, + 560 + ], + "spans": [ + { + "bbox": [ + 55, + 428, + 295, + 560 + ], + "type": "text", + "content": "1. We develop a local input-adaptive fine-tuning algorithm which identifies and updates a subset of inter- and intra-client prompts that are most relevant to each individual sample. This is formulated as prompt-retrieval task which focuses on learning mapping functions that push an input towards its relevant prompts on an embedding space while pulling away from irrelevant prompts on an embedding space. This helps decompose and distill local fine-tuning knowledge into different parts of the prompt set, preventing learned instructions for different fine-tuning patterns from being overloaded on the same prompts (Section 3.2)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 564, + 295, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 295, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 295, + 672 + ], + "type": "text", + "content": "2. We develop a server-side model for aggregating and aligning inter-client prompts across clients. Both generation and alignment are framed as part of a clustering task, where inter-client prompts encoding similar input-level distribution patterns of missing data are identified and merged into more comprehensive and versatile prompt instructions. The empirical effectiveness of these combined prompts in reducing local training losses is used as a learning feedback to optimize this clustering model (Section 3.3)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "3. We evaluate and demonstrate the effectiveness of our proposed approach through extensive experiments on two benchmark datasets, MM-IMDB [1] and UPMC Food" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 328, + 72, + 541, + 212 + ], + "blocks": [ + { + "bbox": [ + 328, + 72, + 541, + 212 + ], + "lines": [ + { + "bbox": [ + 328, + 72, + 541, + 212 + ], + "spans": [ + { + "bbox": [ + 328, + 72, + 541, + 212 + ], + "type": "image", + "image_path": "2d42a96d9e28d4269fffcac2392f709e1fffcd35722a50fe27a27aff2ab6142a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 213, + 555, + 247 + ], + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 247 + ], + "type": "text", + "content": "Figure 1. Our approach with prompt alignment outperforms FedAvg prompt-tuning (w/o alignment) across settings with different data missing rates." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 251, + 555, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 555, + 346 + ], + "type": "text", + "content": "101 [46], comparing its performance against a variety of existing multimodal federated learning methods as well as repurposed prompt-tuning baselines for federated settings. The reported results show that our method consistently and significantly outperforms the baselines across various modality-missing scenarios, achieving new SOTA performance for multi-modal federated learning (Section 4)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 348, + 555, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 555, + 384 + ], + "type": "text", + "content": "A bird-view of our proposed framework, Federated Prompt-Tuning with Heterogeneous and Incomplete Multimodal Client Data (FED-PRIME), is also provided in Figure 2." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 393, + 406, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 393, + 406, + 406 + ], + "spans": [ + { + "bbox": [ + 314, + 393, + 406, + 406 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 414, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 555, + 640 + ], + "type": "text", + "content": "Multi-Modal Federated Learning. Federated learning (FL) [17, 31, 53, 54] is a collaborative learning paradigm that enables multiple devices to train a shared model while keeping local data private. This design addresses concerns over privacy, security, and inefficiency in transferring large datasets [45, 57]. While FL has been widely studied [10, 14, 15, 18, 26, 27, 30, 34–36, 42, 43], most existing methods focus on uni-modal data (e.g., image or text). However, with the rapid advancement of mobile phones and IoT devices [6], local data increasingly comes from multiple modalities. Integrating such heterogeneous signals enables richer, more robust representations [19, 48], motivating growing interest in multi-modal FL. Yet many current approaches rely on custom architectures that cannot exploit pre-trained multi-modal models. More recent studies explore federated fine-tuning of pre-trained multi-modal models, but they typically assume each client holds the same set of modalities [56], limiting applicability in realistic heterogeneous environments." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Missing Data in Multi-Modal FL. In real-world settings, local devices collect data from different types of sensors, leading to datasets with varying feature modalities (inter-heterogeneity). Moreover, sensor malfunctions or data corruption can cause missing features, with devices exhibiting distinct missing patterns (intra-heterogeneity)." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3937" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": "Existing solutions either treat each modality independently [55, 61] or are restricted to centralized learning [7, 38, 51], both overlooking inter-modal dependencies. Recent FL methods address inter-heterogeneity [8, 37] but not intra-heterogeneity, while approaches like FedInMM [52] and FedMAC [33] tackle intra-client missing data but ignore cross-client alignment, limiting effectiveness. A related line of work, including FedMSplit [8] and CACMRN [49], handles modality-incomplete data under a personalized FL framework, where each client retains its model. This paradigm differs from standard FL, which aims to improve a single shared model, making these methods not directly comparable." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 229, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 229, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 229, + 296, + 396 + ], + "type": "text", + "content": "Addressing Missing Data with Prompt-Tuning. Prompt-tuning [41, 44, 59, 60] provides an effective way to adapt pre-trained models by learning prompts that add contextual information to input embeddings. Recent work has extended this idea to missing-modality scenarios by learning a prompt for each subset of missing features [20, 24], so that inputs with similar missing patterns are augmented with suitable prompts. However, these methods are developed for centralized training and do not address the additional heterogeneity introduced in FL, where prompts must be aligned and aggregated across clients. Designing effective alignment and aggregation strategies for such prompts is therefore a key challenge, as demonstrated in our framework." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 404, + 274, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 404, + 274, + 419 + ], + "spans": [ + { + "bbox": [ + 55, + 404, + 274, + 419 + ], + "type": "text", + "content": "3. Multi-Modal Federated Prompt-Tuning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 426, + 296, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 666 + ], + "type": "text", + "content": "Multimodal federated learning faces significant challenges due to missing data modalities, leading to both inter-client and intra-client heterogeneities that complicate knowledge aggregation and degrade overall model performance - see Fig. 1. To address this challenge (Section 3.1), we present a federated prompt-tuning framework based on the pretrained Vision and Language Transformer (ViLT) model, which comprises a prompt-based client design and a server-side prompt-aggregation algorithm. The client design features a fine-tuning algorithm that learns and decomposes the fine-tuned knowledge into separate sets of inter-client and intra-client prompts specialized for fine-tuning at different input-level patterns of missing data and input-agnostic patterns of missing modalities (Section 3.2). The server-side algorithm then learns alignment between inter-client prompts specializing for similar input-level patterns of missing data and aggregates them to create more versatile prompts. This is also combined with a standard FEDAVG-based aggregation of intra-prompts specializing for input-agnostic patterns of missing modalities (Section 3.3) - see Fig. 2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 670, + 180, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 670, + 180, + 682 + ], + "spans": [ + { + "bbox": [ + 55, + 670, + 180, + 682 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "In multimodal federated learning (MMFL), we assume there are " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": " clients and " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": " data modalities (e.g., text, image)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": "Each client " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "t \\in [n]^{1}" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": " has access to a private dataset " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "D_{t} \\triangleq \\{x(M_{t,s}), z_{t,s}\\}_{s=1}^{m}" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "x(M_{t,s})" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": " denote the observed features for the set of modalities, " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "M_{t,s} \\subseteq [r]" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": " recorded for the " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": "-th data point of the " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": "-th client, and " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "z_{t,s}" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": " denotes the corresponding target label. The goal is to learn a multimodal model " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "w_{*}" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": " using all the private datasets " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "\\{D_{t}\\}_{t=1}^{n}" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": " without centralizing them. That is," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 343, + 160, + 537, + 193 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 160, + 537, + 193 + ], + "spans": [ + { + "bbox": [ + 343, + 160, + 537, + 193 + ], + "type": "interline_equation", + "content": "\\boldsymbol {w} _ {*} = \\underset {\\boldsymbol {w}} {\\arg \\min } \\left\\{L (\\boldsymbol {w}) \\triangleq \\frac {1}{n} \\sum_ {t = 1} ^ {n} L _ {t} (\\boldsymbol {w}) \\right\\} \\text {w i t h}", + "image_path": "a6a8f123e98f019d1ab92062ae26ccbd703ec1553ec56aa7349056f6cb4f2f97.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 329, + 194, + 555, + 225 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 194, + 555, + 225 + ], + "spans": [ + { + "bbox": [ + 329, + 194, + 555, + 225 + ], + "type": "interline_equation", + "content": "L _ {t} (\\boldsymbol {w}) \\triangleq \\frac {1}{n} \\sum_ {s = 1} ^ {m} \\ell \\left(F \\left(\\boldsymbol {x} \\left(M _ {t, s}\\right); \\boldsymbol {w}\\right), z _ {t, s}\\right) \\tag {1}", + "image_path": "e4a1788f729c022b06a8a89da15c65b85c7b244714e542d7cd97a57fced3413c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 231, + 555, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 231, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 555, + 327 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 231, + 555, + 327 + ], + "type": "inline_equation", + "content": "F(\\pmb{x}(M_{t,s}); \\pmb{w})" + }, + { + "bbox": [ + 313, + 231, + 555, + 327 + ], + "type": "text", + "content": " denotes the multimodal prediction model with parameter " + }, + { + "bbox": [ + 313, + 231, + 555, + 327 + ], + "type": "inline_equation", + "content": "\\pmb{w}" + }, + { + "bbox": [ + 313, + 231, + 555, + 327 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 231, + 555, + 327 + ], + "type": "inline_equation", + "content": "\\ell(F(\\pmb{x}(M_{t,s}); \\pmb{w}), z_{t,s})" + }, + { + "bbox": [ + 313, + 231, + 555, + 327 + ], + "type": "text", + "content": " denotes the individual loss of " + }, + { + "bbox": [ + 313, + 231, + 555, + 327 + ], + "type": "inline_equation", + "content": "\\pmb{w}" + }, + { + "bbox": [ + 313, + 231, + 555, + 327 + ], + "type": "text", + "content": " on the " + }, + { + "bbox": [ + 313, + 231, + 555, + 327 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 231, + 555, + 327 + ], + "type": "text", + "content": "-th data point. Previous work [8, 33, 37, 52] in MMFL has largely focused on designing " + }, + { + "bbox": [ + 313, + 231, + 555, + 327 + ], + "type": "inline_equation", + "content": "F(\\pmb{x}(M_{t,s}); \\pmb{w})" + }, + { + "bbox": [ + 313, + 231, + 555, + 327 + ], + "type": "text", + "content": " without leveraging existing knowledge in pre-trained foundation model. Instead, we let " + }, + { + "bbox": [ + 313, + 231, + 555, + 327 + ], + "type": "inline_equation", + "content": "F(\\pmb{x}(M_{t,s}); \\pmb{w})" + }, + { + "bbox": [ + 313, + 231, + 555, + 327 + ], + "type": "text", + "content": " be a prompt-tuned version of a pre-trained transformer-based model," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 331, + 554, + 352 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 331, + 554, + 352 + ], + "spans": [ + { + "bbox": [ + 315, + 331, + 554, + 352 + ], + "type": "interline_equation", + "content": "F (\\boldsymbol {x} (M); \\boldsymbol {w}) \\triangleq F _ {c} \\left(F _ {p} \\left(F _ {e} (\\boldsymbol {x} (M)) \\circ \\boldsymbol {w} _ {p}\\right); \\boldsymbol {w} _ {c}\\right) \\tag {2}", + "image_path": "d522cf74d27a495eb548517f88a64c456611a7e28f9ebcf95bc2eccaabbc7f08.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 355, + 555, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 355, + 555, + 511 + ], + "spans": [ + { + "bbox": [ + 313, + 355, + 555, + 511 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 355, + 555, + 511 + ], + "type": "inline_equation", + "content": "F_{c}(:,\\boldsymbol{w}_{c})" + }, + { + "bbox": [ + 313, + 355, + 555, + 511 + ], + "type": "text", + "content": " is the prediction head parameterized with " + }, + { + "bbox": [ + 313, + 355, + 555, + 511 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}_{c}" + }, + { + "bbox": [ + 313, + 355, + 555, + 511 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 355, + 555, + 511 + ], + "type": "inline_equation", + "content": "F_{e}(\\boldsymbol{x}(M)) = F_{e}(\\boldsymbol{x}_{1}) \\circ \\dots \\circ F_{e}(\\boldsymbol{x}_{|M|})" + }, + { + "bbox": [ + 313, + 355, + 555, + 511 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 355, + 555, + 511 + ], + "type": "inline_equation", + "content": "F_{e}(\\boldsymbol{x}_{i})" + }, + { + "bbox": [ + 313, + 355, + 555, + 511 + ], + "type": "text", + "content": " denote the tokenized input sequence of each unimodal feature " + }, + { + "bbox": [ + 313, + 355, + 555, + 511 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{i}" + }, + { + "bbox": [ + 313, + 355, + 555, + 511 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 355, + 555, + 511 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}(M) \\triangleq (\\boldsymbol{x}_{1}, \\boldsymbol{x}_{2}, \\dots, \\boldsymbol{x}_{|M|})" + }, + { + "bbox": [ + 313, + 355, + 555, + 511 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 355, + 555, + 511 + ], + "type": "inline_equation", + "content": "F_{p}()" + }, + { + "bbox": [ + 313, + 355, + 555, + 511 + ], + "type": "text", + "content": " denotes the frozen multimodal pre-trained model – e.g., ViLT [21] – and " + }, + { + "bbox": [ + 313, + 355, + 555, + 511 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}_{p}" + }, + { + "bbox": [ + 313, + 355, + 555, + 511 + ], + "type": "text", + "content": " denotes the set of (learnable) prompts. Here, " + }, + { + "bbox": [ + 313, + 355, + 555, + 511 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 313, + 355, + 555, + 511 + ], + "type": "text", + "content": " denotes the set concatenation operator and " + }, + { + "bbox": [ + 313, + 355, + 555, + 511 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w} = (\\boldsymbol{w}_{c}, \\boldsymbol{w}_{p})" + }, + { + "bbox": [ + 313, + 355, + 555, + 511 + ], + "type": "text", + "content": " denotes the full set of tuning parameters. Due to the decentralized and private nature of the local datasets " + }, + { + "bbox": [ + 313, + 355, + 555, + 511 + ], + "type": "inline_equation", + "content": "\\{D_{t}\\}_{t=1}^{n}" + }, + { + "bbox": [ + 313, + 355, + 555, + 511 + ], + "type": "text", + "content": ", solving Eq. (1) directly is not possible. Instead, following the federated learning [31] paradigm, its optimal solution " + }, + { + "bbox": [ + 313, + 355, + 555, + 511 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}_{*}" + }, + { + "bbox": [ + 313, + 355, + 555, + 511 + ], + "type": "text", + "content": " is approximated via a number of iterations over two interleaving client- and server-side steps:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 516, + 555, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 516, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 314, + 516, + 555, + 536 + ], + "type": "interline_equation", + "content": "\\boldsymbol {w} _ {t} ^ {(h + 1)} = \\operatorname {u p d a t e} \\left(L _ {t}, \\boldsymbol {w} _ {*} ^ {(h)}\\right) \\forall t \\in [ n ], \\tag {3}", + "image_path": "3e763da05c49ce21e0573576094fa52ec87c556ddf6fadc111e78079c970919e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 537, + 554, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 537, + 554, + 559 + ], + "spans": [ + { + "bbox": [ + 315, + 537, + 554, + 559 + ], + "type": "interline_equation", + "content": "\\boldsymbol {w} _ {*} ^ {(h + 1)} = \\text {a g g r e g a t e} \\left(\\boldsymbol {w} _ {1} ^ {(h + 1)}, \\dots , \\boldsymbol {w} _ {n} ^ {(h + 1)}\\right), \\tag {4}", + "image_path": "574bda243ddd27ba6cd101af17c517a7e5d15ce19c77474281cfe4e1b1939d03.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 563, + 555, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 555, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 555, + 696 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 563, + 555, + 696 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}_t^{(h)}" + }, + { + "bbox": [ + 313, + 563, + 555, + 696 + ], + "type": "text", + "content": " denote the estimation of the local tuning parameter " + }, + { + "bbox": [ + 313, + 563, + 555, + 696 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}_t" + }, + { + "bbox": [ + 313, + 563, + 555, + 696 + ], + "type": "text", + "content": " of client " + }, + { + "bbox": [ + 313, + 563, + 555, + 696 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 563, + 555, + 696 + ], + "type": "text", + "content": " after " + }, + { + "bbox": [ + 313, + 563, + 555, + 696 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 563, + 555, + 696 + ], + "type": "text", + "content": " iterations. For example, the update step can be a vanilla gradient descent (GD) optimization of the prompt " + }, + { + "bbox": [ + 313, + 563, + 555, + 696 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 313, + 563, + 555, + 696 + ], + "type": "text", + "content": " and the aggregate step denotes the average of the locally optimized prompts. Such vanilla approach however does not account for both the inter- and intra-heterogeneities among local prompts (see Section 1) which ultimately risks aggregating incompatible prompts and degrade the overall performance (see Section 4). To avoid this, we will introduce novel designs for client- and server-prompts as detailed in the followings." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 701, + 400, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 400, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 400, + 714 + ], + "type": "inline_equation", + "content": "^{1}[n]\\triangleq \\{1,2,\\dots ,n\\}" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3938" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 558, + 285 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 558, + 285 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 558, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 558, + 285 + ], + "type": "image", + "image_path": "f0ddb7f20be97d3b27ffd81002ae84c428259f17301d7fbd3677cd308e86b259.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 299, + 555, + 334 + ], + "lines": [ + { + "bbox": [ + 54, + 299, + 555, + 334 + ], + "spans": [ + { + "bbox": [ + 54, + 299, + 555, + 334 + ], + "type": "text", + "content": "Figure 2. Overview of the proposed multi-modal federated prompt-tuning framework - see Alg. 3. Each client maintains two (learnable) sets of intra- and inter-client prompts. At the beginning of each iteration, each client performs local training via Eq. (7) and Eq. (8). Local sets of intra- and inter-client prompts are then sent to the server for aggregation - see Eq. (10)-Eq. (14)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 341, + 141, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 341, + 141, + 355 + ], + "spans": [ + { + "bbox": [ + 55, + 341, + 141, + 355 + ], + "type": "text", + "content": "3.2. Client Design" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 359, + 296, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 359, + 296, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 359, + 296, + 431 + ], + "type": "text", + "content": "To enable effective prompt-aggregation across heterogeneous clients, we need to separate the fine-tuning treatment for input-level patterns of missing data and input-agnostic patterns of missing modalities. This can be achieved via further parameterizing " + }, + { + "bbox": [ + 55, + 359, + 296, + 431 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}_p = (\\boldsymbol{w}_p^{inter}, \\boldsymbol{w}_p^{intra})" + }, + { + "bbox": [ + 55, + 359, + 296, + 431 + ], + "type": "text", + "content": " with two distinct subsets of " + }, + { + "bbox": [ + 55, + 359, + 296, + 431 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 359, + 296, + 431 + ], + "type": "text", + "content": " inter- and intra-client prompts:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 437, + 295, + 458 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 295, + 458 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 295, + 458 + ], + "type": "interline_equation", + "content": "\\boldsymbol {w} _ {p} ^ {\\text {i n t e r}} \\triangleq \\left\\{\\boldsymbol {p} _ {1} ^ {\\text {i n t e r}}, \\boldsymbol {p} _ {2} ^ {\\text {i n t e r}}, \\dots , \\boldsymbol {p} _ {\\tau} ^ {\\text {i n t e r}} \\right\\}, \\tag {5}", + "image_path": "c0368c9e1c4fc6c5b2cd874e44a542383dabe297060ce59b523f229c17ae8fde.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 459, + 295, + 480 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 459, + 295, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 459, + 295, + 480 + ], + "type": "interline_equation", + "content": "\\boldsymbol {w} _ {p} ^ {\\text {i n t r a}} \\triangleq \\left\\{\\boldsymbol {p} _ {1} ^ {\\text {i n t r a}}, \\boldsymbol {p} _ {2} ^ {\\text {i n t r a}}, \\dots , \\boldsymbol {p} _ {\\tau} ^ {\\text {i n t r a}} \\right\\}, \\tag {6}", + "image_path": "ce16dea17833328088604ad986e086587931d8572d03346bb4d0f4af9a28bd8e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 486, + 295, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 654 + ], + "type": "text", + "content": "which are learned separately. The learned inter-client prompts " + }, + { + "bbox": [ + 55, + 486, + 295, + 654 + ], + "type": "inline_equation", + "content": "\\pmb{w}_{p}^{inter}" + }, + { + "bbox": [ + 55, + 486, + 295, + 654 + ], + "type": "text", + "content": " across clients will be sent to the server which clusters them into different groups based on their specialization for similar input-level patterns of missing data - Section 3.3. Local sets of intra-client prompts " + }, + { + "bbox": [ + 55, + 486, + 295, + 654 + ], + "type": "inline_equation", + "content": "\\pmb{w}_{p}^{intra}" + }, + { + "bbox": [ + 55, + 486, + 295, + 654 + ], + "type": "text", + "content": " can be averaged via FEDAVG [31] since their specialization is for input-agnostic patterns of missing modalities. Intuitively, the aggregation mechanisms will determine how different aspects of fine-tuning knowledge are encoded across the prompt sets. For example, if knowledge regarding different input-level patterns of missing data is (incorrectly) incorporated into " + }, + { + "bbox": [ + 55, + 486, + 295, + 654 + ], + "type": "inline_equation", + "content": "\\{\\pmb{p}_i^{intra}\\}_{i=1}^\\tau" + }, + { + "bbox": [ + 55, + 486, + 295, + 654 + ], + "type": "text", + "content": ", they will be averaged out and collapsed into less informative prompts due to the non-clustering aggregation mechanism of " + }, + { + "bbox": [ + 55, + 486, + 295, + 654 + ], + "type": "inline_equation", + "content": "\\{\\pmb{p}_i^{intra}\\}_{i=1}^\\tau" + }, + { + "bbox": [ + 55, + 486, + 295, + 654 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "As this potential loss of information will decrease the overall performance, the (learnable) prompt parameters will be updated to prevent this from happening. Otherwise, if (common) knowledge patterns regarding input-agnostic patterns of missing modalities are (incorrectly) incorporated" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 341, + 555, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 555, + 426 + ], + "type": "text", + "content": "into " + }, + { + "bbox": [ + 313, + 341, + 555, + 426 + ], + "type": "inline_equation", + "content": "\\left\\{\\pmb{p}_i^{inter}\\right\\}_{i=1}^{\\tau}" + }, + { + "bbox": [ + 313, + 341, + 555, + 426 + ], + "type": "text", + "content": ", their modeling bandwidth for inter-client heterogeneities will be reduced, which will also hamper the performance. This will, in turn, generate a prompt update signal to avoid this as well. Furthermore, within each set of inter- and intra-client prompts, there might also be different patterns that need to be separated and internalized into different subset of prompts." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 426, + 556, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 556, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 556, + 594 + ], + "type": "text", + "content": "To enable this, we develop an input-adaptive learning mechanism for both sets of prompts which identifies and updates the most relevant subset of prompts for each (multimodal) input sample. This is formulated as an information-retrieval task, learning a pair of key and query functions, " + }, + { + "bbox": [ + 313, + 426, + 556, + 594 + ], + "type": "inline_equation", + "content": "k(\\pmb{p})" + }, + { + "bbox": [ + 313, + 426, + 556, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 426, + 556, + 594 + ], + "type": "inline_equation", + "content": "q(\\pmb{x}(M))" + }, + { + "bbox": [ + 313, + 426, + 556, + 594 + ], + "type": "text", + "content": ", which transform the prompt " + }, + { + "bbox": [ + 313, + 426, + 556, + 594 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 313, + 426, + 556, + 594 + ], + "type": "text", + "content": " and input " + }, + { + "bbox": [ + 313, + 426, + 556, + 594 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 313, + 426, + 556, + 594 + ], + "type": "text", + "content": " with the set " + }, + { + "bbox": [ + 313, + 426, + 556, + 594 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 426, + 556, + 594 + ], + "type": "text", + "content": " observed modalities, respectively, into a shared metric space. We can then define a geometric input-prompt distance, e.g., " + }, + { + "bbox": [ + 313, + 426, + 556, + 594 + ], + "type": "inline_equation", + "content": "d(\\pmb{x}(M), \\pmb{p}) = \\cos(q(\\pmb{x}(M)), k(\\pmb{p}))" + }, + { + "bbox": [ + 313, + 426, + 556, + 594 + ], + "type": "text", + "content": " which can be used to measure their contextual relevance. To avoid loading diverse fine-tuning knowledge into a single prompt, we regularize the (local) loss function to penalize it when a less relevant prompt is selected as fine-tuning instruction for an input " + }, + { + "bbox": [ + 313, + 426, + 556, + 594 + ], + "type": "inline_equation", + "content": "\\pmb{x}(M)" + }, + { + "bbox": [ + 313, + 426, + 556, + 594 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 353, + 597, + 555, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 597, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 353, + 597, + 555, + 617 + ], + "type": "interline_equation", + "content": "\\boldsymbol {w} _ {t} \\triangleq \\underset {\\boldsymbol {w}} {\\arg \\min } L _ {t} ^ {\\prime} (\\boldsymbol {w}) \\quad \\text {w h e r e} \\tag {7}", + "image_path": "aa1d6731c858cfb64f258976a96df66e2fabac680800e13ef203531c78a773a6.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 334, + 619, + 553, + 686 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 619, + 553, + 686 + ], + "spans": [ + { + "bbox": [ + 334, + 619, + 553, + 686 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {t} ^ {\\prime} (\\boldsymbol {w}) \\triangleq \\sum_ {s = 1} ^ {m} \\left\\{\\ell \\left(F \\left(\\boldsymbol {x} \\left(M _ {t, s}\\right); \\boldsymbol {w} ^ {\\prime}\\right), z _ {t, s}\\right) \\right\\} \\\\ + \\sum_ {s = 1} ^ {m} \\left\\{r \\left(\\boldsymbol {x} \\left(M _ {t, s}\\right), \\boldsymbol {w} _ {p} ^ {\\prime}\\right) \\right\\}, \\tag {8} \\\\ \\end{array}", + "image_path": "1895545fdb7c2964469546470b747213a7584e4c1968585a745ff2b4f97b1225.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\ell (\\pmb {w}^{\\prime})" + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": " is defined in Eq. (1) and " + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{w}^{\\prime} = (\\pmb{w}_{c},\\pmb{w}_{p}^{\\prime})" + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": " denote the customized set of tuning instructions for each" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3939" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 72, + 289, + 210 + ], + "blocks": [ + { + "bbox": [ + 63, + 72, + 289, + 210 + ], + "lines": [ + { + "bbox": [ + 63, + 72, + 289, + 210 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 289, + 210 + ], + "type": "image", + "image_path": "d354035796080b754dc6c8bf5c32f2276441d652b5d4db4b27aa301009bf8fe3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 218, + 296, + 285 + ], + "lines": [ + { + "bbox": [ + 55, + 218, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 218, + 296, + 285 + ], + "type": "text", + "content": "Figure 3. Workflow of the prompt-alignment algorithm. At each iteration, each client samples a subset of summarizing prompts using its query and key functions. The client then performs a local update, resulting in heterogeneous prompt sets, which are subsequently sent to the server to be clustered and summarized into new summarizing prompt sets for the next iteration." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 304, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 296, + 361 + ], + "type": "text", + "content": "input " + }, + { + "bbox": [ + 55, + 304, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\pmb{x}(M_{t,s})" + }, + { + "bbox": [ + 55, + 304, + 296, + 361 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 304, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\pmb{w}_p' = (\\pmb{w}_p^{inter'}, \\pmb{w}_p^{intra'})" + }, + { + "bbox": [ + 55, + 304, + 296, + 361 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 55, + 304, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\pmb{w}_p^{inter'} \\subseteq \\pmb{w}_p^{inter}" + }, + { + "bbox": [ + 55, + 304, + 296, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 304, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\pmb{w}_p^{intra'} \\subseteq \\pmb{w}_p^{intra}" + }, + { + "bbox": [ + 55, + 304, + 296, + 361 + ], + "type": "text", + "content": " are the sets of " + }, + { + "bbox": [ + 55, + 304, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 55, + 304, + 296, + 361 + ], + "type": "text", + "content": " most relevant prompts to " + }, + { + "bbox": [ + 55, + 304, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\pmb{x}(M_{t,s})" + }, + { + "bbox": [ + 55, + 304, + 296, + 361 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 304, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\pmb{w}_p^{inter}" + }, + { + "bbox": [ + 55, + 304, + 296, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 304, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\pmb{w}_p^{intra}" + }, + { + "bbox": [ + 55, + 304, + 296, + 361 + ], + "type": "text", + "content": ". The regularizer " + }, + { + "bbox": [ + 55, + 304, + 296, + 361 + ], + "type": "inline_equation", + "content": "r(\\pmb{x}(M), \\pmb{w}_p')" + }, + { + "bbox": [ + 55, + 304, + 296, + 361 + ], + "type": "text", + "content": " is then defined as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 78, + 368, + 296, + 404 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 368, + 296, + 404 + ], + "spans": [ + { + "bbox": [ + 78, + 368, + 296, + 404 + ], + "type": "interline_equation", + "content": "r \\left(\\boldsymbol {x} (M), \\boldsymbol {w} _ {p} ^ {\\prime}\\right) \\triangleq \\sum_ {\\boldsymbol {p} \\in \\boldsymbol {w} _ {p} ^ {\\prime}} \\left\\{d (\\boldsymbol {x} (M), \\boldsymbol {p}) \\right\\}. \\tag {9}", + "image_path": "9c82ec2077cefcd6209d7879eeca220d1457d92c4ea513994d5fd56347bb79ae.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 412, + 296, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 412, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 296, + 533 + ], + "type": "text", + "content": "Intuitively, the regularizer penalizes the total distance between the (multimodal) input " + }, + { + "bbox": [ + 55, + 412, + 296, + 533 + ], + "type": "inline_equation", + "content": "\\pmb{x}(M)" + }, + { + "bbox": [ + 55, + 412, + 296, + 533 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 55, + 412, + 296, + 533 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 55, + 412, + 296, + 533 + ], + "type": "text", + "content": " closest prompts according to the aforementioned distance metric " + }, + { + "bbox": [ + 55, + 412, + 296, + 533 + ], + "type": "inline_equation", + "content": "d(\\pmb{x}(M), \\pmb{p})" + }, + { + "bbox": [ + 55, + 412, + 296, + 533 + ], + "type": "text", + "content": ". Minimizing both the original local loss and this regularizer allows the local model to simultaneously learn to (1) internalize fine-tuning patterns of inter- and intra-heterogeneities to minimize the original loss " + }, + { + "bbox": [ + 55, + 412, + 296, + 533 + ], + "type": "inline_equation", + "content": "L_{t}(\\pmb{w})" + }, + { + "bbox": [ + 55, + 412, + 296, + 533 + ], + "type": "text", + "content": "; and (2) ensure that selected prompts for a given input will be close to it to reduce the penalty " + }, + { + "bbox": [ + 55, + 412, + 296, + 533 + ], + "type": "inline_equation", + "content": "r(\\pmb{x}(M), \\pmb{w}_{p}^{\\prime})" + }, + { + "bbox": [ + 55, + 412, + 296, + 533 + ], + "type": "text", + "content": ". A pseudocode of this client design is detailed in Suppl. A." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 539, + 170, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 170, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 170, + 552 + ], + "type": "text", + "content": "3.3. Server Aggregation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 556, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 295, + 640 + ], + "type": "text", + "content": "We will now establish a principled algorithm to align and aggregate inter-client prompts specializing in similar input-level patterns of missing data2. This is not straightforward since inter-client prompts at the same position in the prompt sets across different clients might not specialize for similar patterns of missing data due to client heterogeneities: certain missing-data patterns are not observed at certain clients." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 640, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 640, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 640, + 295, + 677 + ], + "type": "text", + "content": "To find the correct alignment, we formulate it as a clustering optimization task. Suppose each client " + }, + { + "bbox": [ + 55, + 640, + 295, + 677 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 640, + 295, + 677 + ], + "type": "text", + "content": " communicates a set of inter-client prompts " + }, + { + "bbox": [ + 55, + 640, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}_t^{inter} = (\\boldsymbol{p}_t^i)_{i=1}^\\tau" + }, + { + "bbox": [ + 55, + 640, + 295, + 677 + ], + "type": "text", + "content": ". Each" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 169 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 169 + ], + "type": "text", + "content": "potential clustering of these prompts thus defines a candidate alignment, i.e., prompts in the same cluster are considered aligned and can be aggregated. A (learnable) cluster center can then be associated with the aggregated prompt - see Fig. 3. Both the cluster assignment and centers can be found via solving the following optimization task with discrete alignment variables " + }, + { + "bbox": [ + 313, + 72, + 555, + 169 + ], + "type": "inline_equation", + "content": "\\alpha_{t}^{pq}\\in \\{0,1\\}" + }, + { + "bbox": [ + 313, + 72, + 555, + 169 + ], + "type": "text", + "content": " and continuous center variables " + }, + { + "bbox": [ + 313, + 72, + 555, + 169 + ], + "type": "inline_equation", + "content": "\\theta_{q}" + }, + { + "bbox": [ + 313, + 72, + 555, + 169 + ], + "type": "text", + "content": " for each cluster " + }, + { + "bbox": [ + 313, + 72, + 555, + 169 + ], + "type": "inline_equation", + "content": "q\\in [n\\times \\tau ]" + }, + { + "bbox": [ + 313, + 72, + 555, + 169 + ], + "type": "text", + "content": " .." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 178, + 555, + 229 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 178, + 555, + 229 + ], + "spans": [ + { + "bbox": [ + 314, + 178, + 555, + 229 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {\\boldsymbol {\\alpha}, \\boldsymbol {\\theta}, \\gamma} G (\\boldsymbol {\\alpha}, \\boldsymbol {\\theta}, \\gamma) \\triangleq \\sum_ {t = 1} ^ {n} \\sum_ {p = 1} ^ {\\tau} \\sum_ {q = 1} ^ {n \\times \\tau} \\alpha_ {t} ^ {p q} \\cdot c \\left(\\boldsymbol {p} _ {t} ^ {p}, \\boldsymbol {\\theta} _ {q},; \\gamma\\right) (10) \\\\ \\mathrm {s . t .} \\alpha_ {t} ^ {p, 1} + \\dots + \\alpha_ {t} ^ {p, n \\times \\tau} = 1, (11) \\\\ \\end{array}", + "image_path": "3b5bd3c3adddd293e1dff4d341aa5ff00ab3232bfacbb6a1c7d628b692e2ac0f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 236, + 555, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 555, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 555, + 417 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 236, + 555, + 417 + ], + "type": "inline_equation", + "content": "\\alpha \\triangleq (\\alpha_{t}^{p,q})" + }, + { + "bbox": [ + 313, + 236, + 555, + 417 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 236, + 555, + 417 + ], + "type": "inline_equation", + "content": "\\alpha_{t}^{p,q} \\in \\{0,1\\} \\forall (t,p,q)" + }, + { + "bbox": [ + 313, + 236, + 555, + 417 + ], + "type": "text", + "content": " indicates whether the " + }, + { + "bbox": [ + 313, + 236, + 555, + 417 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 236, + 555, + 417 + ], + "type": "text", + "content": "-th inter-client prompt of client " + }, + { + "bbox": [ + 313, + 236, + 555, + 417 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 236, + 555, + 417 + ], + "type": "text", + "content": " is matched to the " + }, + { + "bbox": [ + 313, + 236, + 555, + 417 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 313, + 236, + 555, + 417 + ], + "type": "text", + "content": "-th cluster. The corresponding cluster center is then optimized to have minimum total distance to all cluster members according to a (learnable) cost function " + }, + { + "bbox": [ + 313, + 236, + 555, + 417 + ], + "type": "inline_equation", + "content": "c(\\pmb{p}_{t}^{p},\\pmb{\\theta}_{q})" + }, + { + "bbox": [ + 313, + 236, + 555, + 417 + ], + "type": "text", + "content": ". The optimized cluster center can then be used as the aggregated or summarizing prompt. Minimizing Eq. (10) thus allows us to align and summarize/aggregate the inter-client prompts across clients simultaneously. Note that Eq. (11) enforces a constraint that intuitively indicates two prompts from the same clients should not be matched to the same cluster and get aggregated. This is because local prompts from the same client have already been learned to encode different finetuning aspects. Aggregating them will therefore collapse them into less informative ones." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 418, + 554, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 554, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 554, + 515 + ], + "type": "text", + "content": "In addition, some summarizing/aggregated prompts might be more generic than others and can be used to provide fine-tuning instructions to a wider range of (multimodal) inputs. The update of such prompts should be prioritized during training and to achieve that, we need a mechanism to capture this prompt popularity. To achieve this, we augment the optimization task in Eq. (10) above with a regularizer that penalizes alignment to less popular prompts," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 523, + 554, + 593 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 523, + 554, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 523, + 554, + 593 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} R (\\boldsymbol {\\alpha}, \\zeta) \\triangleq \\sum_ {t = 1} ^ {n} \\sum_ {p = 1} ^ {\\tau} \\sum_ {q = 1} ^ {n \\times \\tau} \\left(\\alpha_ {t} ^ {p, q} \\cdot \\log U \\left(\\boldsymbol {\\theta} _ {q}; \\zeta\\right)\\right) \\tag {12} \\\\ + \\sum_ {t = 1} ^ {n} \\sum_ {p = 1} ^ {\\tau} \\sum_ {q = 1} ^ {n \\times \\tau} \\Big (1 - \\alpha_ {t} ^ {p, q} \\Big) \\cdot \\log \\Big (1 - U (\\boldsymbol {\\theta} _ {q}; \\zeta) \\Big), \\\\ \\end{array}", + "image_path": "6446565b11048761035a8c7f62194843d26325ea2fb4f07a0418f909fc1a050b.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 601, + 555, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 601, + 555, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 601, + 555, + 662 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 601, + 555, + 662 + ], + "type": "inline_equation", + "content": "U(\\pmb{\\theta}_q;\\zeta)" + }, + { + "bbox": [ + 313, + 601, + 555, + 662 + ], + "type": "text", + "content": " is the popularity function of " + }, + { + "bbox": [ + 313, + 601, + 555, + 662 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}_q" + }, + { + "bbox": [ + 313, + 601, + 555, + 662 + ], + "type": "text", + "content": ", which is parameterized by learnable parameter " + }, + { + "bbox": [ + 313, + 601, + 555, + 662 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 313, + 601, + 555, + 662 + ], + "type": "text", + "content": ". For example, " + }, + { + "bbox": [ + 313, + 601, + 555, + 662 + ], + "type": "inline_equation", + "content": "U(\\pmb{\\theta}_q;\\zeta) = \\sigma(g(\\pmb{\\theta}_q;\\zeta)) \\in (0,1)^3" + }, + { + "bbox": [ + 313, + 601, + 555, + 662 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 601, + 555, + 662 + ], + "type": "inline_equation", + "content": "g(\\pmb{\\theta}_q;\\zeta)" + }, + { + "bbox": [ + 313, + 601, + 555, + 662 + ], + "type": "text", + "content": " is a deep neural net parameterized with " + }, + { + "bbox": [ + 313, + 601, + 555, + 662 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 313, + 601, + 555, + 662 + ], + "type": "text", + "content": ". Putting Eq. (13) and Eq. (10) together, we arrive at our final loss function," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 340, + 671, + 555, + 693 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 671, + 555, + 693 + ], + "spans": [ + { + "bbox": [ + 340, + 671, + 555, + 693 + ], + "type": "interline_equation", + "content": "\\underset {\\boldsymbol {\\alpha}, \\boldsymbol {\\theta}, \\gamma , \\zeta} {\\text {m i n i m i z e}} \\left\\{G (\\boldsymbol {\\alpha}, \\boldsymbol {\\theta}, \\gamma) + R (\\boldsymbol {\\alpha}, \\zeta) \\right\\}, \\tag {13}", + "image_path": "ac286f73f11736da08d2c412873d5a1125f4a5d932b213a9344a5164577c1667.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 324, + 702, + 443, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 702, + 443, + 713 + ], + "spans": [ + { + "bbox": [ + 324, + 702, + 443, + 713 + ], + "type": "inline_equation", + "content": "^3\\sigma(.)" + }, + { + "bbox": [ + 324, + 702, + 443, + 713 + ], + "type": "text", + "content": " denotes the sigmoid function." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "text", + "content": "2This section only describes the aggregation algorithm for inter-client prompts since intra-client prompts are specialized for input-agnostic patterns of missing modalities and can be aggregated via FEDAVG [31]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3940" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": "which can be optimized via alternating between (1) optimizing " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "(\\theta, \\zeta, \\gamma)" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " while fixing " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": "; and (2) optimizing " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " while fixing " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "(\\theta, \\zeta, \\gamma)" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": ". The first optimization sub-task is straightforward while the latter is less trivial due to the constraint as well as the discrete nature of the optimizing variables. To sidestep the technical challenge in solving for " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " while fixing " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "(\\theta, \\zeta, \\gamma)" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": ", we approximate it by optimizing for " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\alpha_{t} = \\{\\alpha_{t}^{p,q}\\}_{p,q}" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " while also fixing " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\alpha_{-t} = \\{\\alpha_{-t}^{p,q}\\}_{p,q}" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": ". This results in a minimum weighted linear sum task," + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 188, + 303, + 264 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 188, + 303, + 264 + ], + "spans": [ + { + "bbox": [ + 56, + 188, + 303, + 264 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {\\boldsymbol {\\alpha} _ {t}} L \\left(\\boldsymbol {\\alpha} _ {t}\\right) \\triangleq \\sum_ {p = 1} ^ {\\tau} \\sum_ {q = 1} ^ {n \\times \\tau} \\left(\\alpha_ {t} ^ {p, q} \\cdot v (p, q)\\right) \\text {w h e r e} \\tag {14} \\\\ v (p, q) \\triangleq \\left(\\log \\left\\{\\frac {U (\\boldsymbol {\\theta} _ {q} ; \\zeta)}{1 - U (\\boldsymbol {\\theta} _ {q} ; \\zeta)} \\right\\} + c (\\boldsymbol {p} _ {t} ^ {p}, \\boldsymbol {\\theta} _ {q}; \\gamma)\\right), \\\\ \\end{array}", + "image_path": "1b9a73870de4f1ff29c7f40f6673775be3e8402f410948355fe58110be08fc7f.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 270, + 296, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 296, + 330 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 296, + 330 + ], + "type": "text", + "content": "which can be solved effectively in polynomial time using the Hungarian algorithm [23]. An overview and pseudocode of the server-side aggregation algorithm are presented in Fig. 3 and Suppl. A. A thorough evaluation of this framework is detailed next in Section 4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 340, + 181, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 181, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 181, + 354 + ], + "type": "text", + "content": "4. Empirical Evaluation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 360, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 296, + 479 + ], + "type": "text", + "content": "This section provides a detailed empirical evaluation of our proposed framework, FED-PRIME. Our evaluation is conducted on two datasets UPMC Food-101 [46] and MM-IMDB [1] and several baselines adapting existing works in federated fine-tuning (for unimodal data) and multi-modal prompt-tuning (for centralized data) to our novel setting on decentralized multi-modal data with missing modalities and features. Detailed description of the datasets and baselines are provided in Section 4.1. Our main results and ablation studies are reported in Section 4.2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 487, + 174, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 174, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 174, + 499 + ], + "type": "text", + "content": "4.1. Experiment Settings" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 504, + 295, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 504, + 295, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 504, + 295, + 528 + ], + "type": "text", + "content": "A. Datasets. Following the experiment protocol in [24], we evaluate FED-PRIME on two multimodal datasets:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 531, + 295, + 686 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 55, + 531, + 295, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 531, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 531, + 295, + 602 + ], + "type": "text", + "content": "1. UPMC Food-101 [46] is an image-text classification dataset consisting of noisy image-text pairs collected from Google Image Search. Its classes are aligned with those in the larger, publicly available ETHZ Food-101 dataset [4]. For the federated setting, we select only eight most frequent classes from the original UPMC Food-101." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 602, + 295, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 602, + 295, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 602, + 295, + 686 + ], + "type": "text", + "content": "2. MM-IMDB [1] is a movie genre classification dataset with both image and text modalities. Since each movie can belong to multiple genres, the task is originally a multi-label classification problem, where genres are predicted using image, text, or both. For the federated setting, we use only single-label instances and select eight most frequent classes only from the original MM-IMDB." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "Each dataset is first partitioned into train and test sets with the 80:20 ratio. The (heterogeneous) missing patterns on the train/test data are simulated using the following protocols." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 111, + 553, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 111, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 313, + 111, + 553, + 148 + ], + "type": "text", + "content": "B. Training Data Simulation. Following [24], we simulate missing data phenomenon within a dataset using a missing rate " + }, + { + "bbox": [ + 313, + 111, + 553, + 148 + ], + "type": "inline_equation", + "content": "\\eta \\in (0,1)" + }, + { + "bbox": [ + 313, + 111, + 553, + 148 + ], + "type": "text", + "content": " as detailed in 3 simulation cases below:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 151, + 553, + 243 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 313, + 151, + 553, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 151, + 553, + 176 + ], + "spans": [ + { + "bbox": [ + 313, + 151, + 553, + 176 + ], + "type": "text", + "content": "1. Missing Image. We drop the image features in " + }, + { + "bbox": [ + 313, + 151, + 553, + 176 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 151, + 553, + 176 + ], + "type": "text", + "content": " percent of the data and keep the remaining " + }, + { + "bbox": [ + 313, + 151, + 553, + 176 + ], + "type": "inline_equation", + "content": "(1 - \\eta)" + }, + { + "bbox": [ + 313, + 151, + 553, + 176 + ], + "type": "text", + "content": " percent complete." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 178, + 553, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 178, + 553, + 203 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 553, + 203 + ], + "type": "text", + "content": "2. Missing Text. We drop the text features in " + }, + { + "bbox": [ + 313, + 178, + 553, + 203 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 178, + 553, + 203 + ], + "type": "text", + "content": " percent of the data and keep the remaining " + }, + { + "bbox": [ + 313, + 178, + 553, + 203 + ], + "type": "inline_equation", + "content": "(1 - \\eta)" + }, + { + "bbox": [ + 313, + 178, + 553, + 203 + ], + "type": "text", + "content": " percent complete." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 205, + 553, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 553, + 243 + ], + "type": "text", + "content": "3. Missing Both. We drop the image component in " + }, + { + "bbox": [ + 313, + 205, + 553, + 243 + ], + "type": "inline_equation", + "content": "\\eta /2" + }, + { + "bbox": [ + 313, + 205, + 553, + 243 + ], + "type": "text", + "content": " percent of the data, and the text component in the other " + }, + { + "bbox": [ + 313, + 205, + 553, + 243 + ], + "type": "inline_equation", + "content": "\\eta /2" + }, + { + "bbox": [ + 313, + 205, + 553, + 243 + ], + "type": "text", + "content": " percent, leaving the remaining " + }, + { + "bbox": [ + 313, + 205, + 553, + 243 + ], + "type": "inline_equation", + "content": "(1 - \\eta)" + }, + { + "bbox": [ + 313, + 205, + 553, + 243 + ], + "type": "text", + "content": " percent complete." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 245, + 553, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 245, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 553, + 293 + ], + "type": "text", + "content": "The missing-simulated train set is then partitioned into " + }, + { + "bbox": [ + 313, + 245, + 553, + 293 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 245, + 553, + 293 + ], + "type": "text", + "content": " subsets for " + }, + { + "bbox": [ + 313, + 245, + 553, + 293 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 245, + 553, + 293 + ], + "type": "text", + "content": " clients via random sampling. The resulting client datasets will therefore have heterogeneous distributions over missing-data patterns." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 296, + 553, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 296, + 553, + 356 + ], + "spans": [ + { + "bbox": [ + 313, + 296, + 553, + 356 + ], + "type": "text", + "content": "C. Test Data Simulation. There are 5 test scenarios, which include: (1) Full Modal (no missing data); (2) Image Only (drop all text features); (3) Text Only (drop all image features); (4) Miss Both (run Missing Both on test data); and (5) " + }, + { + "bbox": [ + 313, + 296, + 553, + 356 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 313, + 296, + 553, + 356 + ], + "type": "text", + "content": " Train (same simulation as the training data)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 359, + 553, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 359, + 553, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 359, + 553, + 395 + ], + "type": "text", + "content": "D. Metrics. Following prior protocol in [24], we use F1-macro and classification accuracy to evaluate the following baseline methods on MM-IMDB and UPMC Food-101." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 399, + 554, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 554, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 554, + 639 + ], + "type": "text", + "content": "E. Baseline Methods. Following [24], we adopt the pretrained multimodal transformer ViLT [21] as the backbone of our client design. Each client model is a fine-tuned version of this backbone which is generated based on its own local dataset. As no prior work has addressed modality missing in multimodal fine-tuning within federated learning settings, our comparison baselines mostly feature adapted versions of existing work in federated learning and multimodal prompt-tuning. This includes adaptations of the latest missing prompt-tuning SOTA method in [24] into the context of the classical FedAvg method as well as that of a more recent FL approach, FedMSplit [8]. These are referred to as FEDAVG-P and FEDMSPLIT-P, respectively. In addition, to underscore the importance of separating fine-tuning knowledge across intra- and inter-client prompt sets, we also include two simplified variants of FED-PRIME utilizing either inter-client prompt pooling or intra-client prompt pooling, along with their respective aggregations. We refer to these variants as FED-INTER and FED-INTRA, respectively (see more details in Supplementary B)." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 647, + 438, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 438, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 438, + 659 + ], + "type": "text", + "content": "4.2. Experimental Results" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": "This section provides extensive performance comparison between FED-PRIME and other baseline methods across various settings of missing data in the train and test sets, as previously stated in Section 4.1. In addition, we also provide" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "4Our implementation is available at https://github.com/hangpt01/FedPrime." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3941" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 134, + 555, + 321 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "type": "text", + "content": "Table 1. Performance achieved by the baselines on the Food-101 (left) and MM-IMDB (right) datasets under various missing-data settings in the train and test sets. This includes all combinations of the 3 train scenarios and 5 test scenarios (i.e., 15 experiments) per dataset, as previously stated in Section 4.1. The best and second-best results are highlighted in bold red, and blue, respectively. Note that in the Miss Both training scenarios, Test (" + }, + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "type": "text", + "content": "Train) and Test (Miss Both) represent the same test scenario according to the definition of " + }, + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "type": "text", + "content": "Train. The reported result for Test (Miss Both) in this case is the same as that for Test (" + }, + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "type": "text", + "content": "Train), as indicated by a dash (-)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 134, + 555, + 321 + ], + "lines": [ + { + "bbox": [ + 56, + 134, + 555, + 321 + ], + "spans": [ + { + "bbox": [ + 56, + 134, + 555, + 321 + ], + "type": "table", + "html": "
DatasetsUPMC Food-101MM-IMDB
TrainMethodTest(∼ Train)Test(Miss Both)Test(Full Modal)Test(Text only)Test(Image only)Test(∼ Train)Test(Miss Both)Test(Full Modal)Test(Text only)Test(Image only)
Miss TextFEDAVG-P15.71 ± 2.2714.90 ± 1.5721.56 ± 7.8116.91 ± 0.6915.36 ± 0.4322.42 ± 2.2721.89 ± 1.3430.78 ± 1.6518.40 ± 0.0614.53 ± 4.56
FEDMSPLIT-P15.62 ± 1.5117.50 ± 1.9725.27 ± 8.2218.78 ± 0.6417.50 ± 1.9721.02 ± 1.8919.97 ± 0.7424.39 ± 6.0814.38 ± 2.7618.09 ± 6.13
FED-INTER54.82 ± 19.0148.87 ± 24.6459.17 ± 27.0635.13 ± 26.7856.59 ± 15.1218.25 ± 3.5016.95 ± 3.5718.67 ± 7.6315.03 ± 4.6618.01 ± 1.95
FED-INTRA61.71 ± 17.2248.09 ± 19.1262.06 ± 26.9822.51 ± 5.9262.64 ± 11.8313.38 ± 1.7312.77 ± 0.8512.55 ± 1.6711.31 ± 0.3814.33 ± 1.80
FED-PRIME78.88 ± 0.9080.38 ± 0.6592.12 ± 0.4073.01 ± 4.2576.83 ± 1.2231.92 ± 0.2031.48 ± 0.3037.67 ± 0.0430.60 ± 1.4130.69 ± 1.41
Improv. (%)27.82 ↑64.48 ↑48.44 ↑107.83 ↑22.65 ↑42.37 ↑43.81 ↑22.35 ↑66.30 ↑69.65 ↑
Miss ImageFEDAVG-P17.35 ± 4.7715.12 ± 1.4816.84 ± 2.3718.12 ± 6.4914.81 ± 0.2427.69 ± 5.9722.55 ± 3.0631.94 ± 0.9823.76 ± 11.7212.29 ± 0.47
FEDMSPLIT-P74.16 ± 10.5648.88 ± 10.2645.64 ± 32.4388.65 ± 2.1714.81 ± 0.9019.11 ± 11.3316.61 ± 7.2218.19 ± 12.5518.12 ± 12.3012.81 ± 1.25
FED-INTER77.96 ± 11.6264.62 ± 10.2282.08 ± 7.7577.69 ± 12.3537.56 ± 6.4918.79 ± 5.2317.93 ± 3.6020.56 ± 3.5617.67 ± 6.7915.47 ± 2.51
FED-INTRA22.84 ± 3.5220.13 ± 1.7223.48 ± 1.8624.46 ± 2.9916.66 ± 1.3215.75 ± 4.3414.06 ± 3.1615.68 ± 4.7714.53 ± 3.6511.71 ± 0.42
FED-PRIME90.55 ± 0.2279.12 ± 0.4992.89 ± 0.2190.18 ± 0.2954.14 ± 2.5036.08 ± 0.3531.35 ± 0.6138.49 ± 0.5636.91 ± 0.5918.15 ± 0.66
Improv. (%)16.15 ↑22.44 ↑13.17 ↑1.73 ↑44.14 ↑30.30 ↑39.02 ↑20.51 ↑55.35 ↑17.32 ↑
Miss BothFEDAVG-P14.57 ± 1.50-17.17 ± 4.3716.40 ± 4.0513.24 ± 0.3226.45 ± 2.63-33.03 ± 2.5624.12 ± 11.3020.21 ± 1.98
FEDMSPLIT-P49.15 ± 24.76-64.78 ± 36.6264.62 ± 36.5121.49 ± 7.1924.25 ± 5.02-26.05 ± 11.1726.02 ± 9.6419.79 ± 6.20
FED-INTER56.32 ± 21.77-69.57 ± 19.4145.15 ± 34.0959.30 ± 10.8426.53 ± 0.90-31.97 ± 2.2229.69 ± 2.2121.63 ± 0.77
FED-INTRA49.28 ± 32.87-56.70 ± 37.9043.24 ± 34.1949.85 ± 25.4411.90 ± 0.37-12.47 ± 0.4511.46 ± 0.3312.83 ± 0.92
FED-PRIME84.44 ± 2.65-93.64 ± 0.5887.95 ± 0.9172.41 ± 3.8832.01 ± 2.51-38.68 ± 0.6531.00 ± 2.9726.01 ± 0.12
Improv. (%)49.93 ↑-34.60 ↑36.10 ↑22.11 ↑20.66 ↑-17.11 ↑4.41 ↑20.25 ↑
", + "image_path": "5fb19d4f2e0c8b1bdc77b26f5e9c82a5f181f1092664d5ef57d621f968613a3f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 70, + 323, + 493, + 334 + ], + "lines": [ + { + "bbox": [ + 70, + 323, + 493, + 334 + ], + "spans": [ + { + "bbox": [ + 70, + 323, + 493, + 334 + ], + "type": "text", + "content": "(*) Improv. shows the relative performance improvement between our proposal and the second-best. (in percentage)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 354, + 295, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 295, + 390 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 295, + 390 + ], + "type": "text", + "content": "ablation studies on (1) performance robustness against various missing rate " + }, + { + "bbox": [ + 55, + 354, + 295, + 390 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 354, + 295, + 390 + ], + "type": "text", + "content": "; (2) the ablated impact of having inter- and intra-client prompt sets; and (3) training convergence." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 399, + 141, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 141, + 410 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 141, + 410 + ], + "type": "text", + "content": "4.2.1. Main Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 415, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 296, + 713 + ], + "type": "text", + "content": "We compare the performance of FED-PRIME against the other baselines on an extensive set of experiments across all datasets. This includes all combinations of the 3 missing simulation scenarios on train data and 5 test scenarios (with both missing and no missing data) as previously detailed in parts B and C of Section 4.1. All results are reported in Table 1 which show that FED-PRIME consistently outperforms all baselines in all train/test scenarios with various missing-data patterns. The performance on the UPMC Food-101 and MM-IMDB datasets is measured in terms of classification accuracy and F1-macro, respectively. The F1-macro performance on MM-IMDB is relatively low across all settings due to the class imbalance nature of the dataset. In each experiment, the relative performance improvement of FED-PRIME over the second best baseline is reported, which ranges between " + }, + { + "bbox": [ + 55, + 415, + 296, + 713 + ], + "type": "inline_equation", + "content": "1.73\\%" + }, + { + "bbox": [ + 55, + 415, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 415, + 296, + 713 + ], + "type": "inline_equation", + "content": "107.83\\%" + }, + { + "bbox": [ + 55, + 415, + 296, + 713 + ], + "type": "text", + "content": " for the UPMC Food-101 dataset and between " + }, + { + "bbox": [ + 55, + 415, + 296, + 713 + ], + "type": "inline_equation", + "content": "4.41\\%" + }, + { + "bbox": [ + 55, + 415, + 296, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 415, + 296, + 713 + ], + "type": "inline_equation", + "content": "69.65\\%" + }, + { + "bbox": [ + 55, + 415, + 296, + 713 + ], + "type": "text", + "content": " for the MM-IMDB dataset. Notably, FED-PRIME significantly outperforms both of its simplified variants FED-INTRA and FED-INTER, which either handle intra- or interheterogeneities, in all settings. This highlights the remarkable synergies and essential of handling both types of heterogeneities in multi-modal federated learning with missing data. In addition, it can also be observed that in more than half of the experiment settings, the second best base" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 354, + 555, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 555, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 555, + 413 + ], + "type": "text", + "content": "line is either FED-INTRA or FED-INTER which significantly outperform both FEDAVG-P and FEDMSPLIT-P in those cases. These observations complement our findings of the isolated impact of these components in our ablation studies in Section 4.2.4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 422, + 470, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 470, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 470, + 434 + ], + "type": "text", + "content": "4.2.2. Impact of Missing Modalities." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 437, + 555, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 654 + ], + "type": "text", + "content": "Across different testing scenarios, we observe that the highest test performance is always achieved when there is no missing data in the test set as expected. Another interesting observation is that in the Miss Text training scenario, where approximately " + }, + { + "bbox": [ + 313, + 437, + 555, + 654 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 313, + 437, + 555, + 654 + ], + "type": "text", + "content": " of the text features are lost on the training data, the proposed model still performs surprisingly well on the Text Only test scenarios. This suggests that the multi-modal prompt alignment is effective in recovering text information, highlighting the proposed design's adaptability. However, in the Miss Image training scenario, the test performance on the Image Only case suffers instead a significant performance drop. Upon investigation, we believe this is likely due to the VILT backbone's lightweight image processing module, which was optimized to work in tandem with text for competitive performance, hinting at a potential text-centered nature of VILT during pre-training. This is in fact well-aligned with existing observations that co-training with text often leads to improved performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 662, + 525, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 662, + 525, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 662, + 525, + 673 + ], + "type": "text", + "content": "4.2.3. Robustness Against Various Missing Rates." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "To demonstrate the robustness of FED-PRIME against different missing rates " + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": ", we run multiple comparisons of FED-PRIME, FEDAVG-P, and its centralized variant [24]" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3942" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 68, + 200, + 180 + ], + "blocks": [ + { + "bbox": [ + 55, + 68, + 200, + 180 + ], + "lines": [ + { + "bbox": [ + 55, + 68, + 200, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 68, + 200, + 180 + ], + "type": "image", + "image_path": "a8c4a2b4d053f7b04dca0657b87f5b136b5b6deaf01c181c0cc83ea1db0121ba.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 188, + 198, + 222 + ], + "lines": [ + { + "bbox": [ + 55, + 188, + 198, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 188, + 198, + 222 + ], + "type": "text", + "content": "Figure 4. Performance comparison on UMPC Food-101 under Miss Both scenarios with various missing rates." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 206, + 69, + 355, + 182 + ], + "blocks": [ + { + "bbox": [ + 206, + 69, + 355, + 182 + ], + "lines": [ + { + "bbox": [ + 206, + 69, + 355, + 182 + ], + "spans": [ + { + "bbox": [ + 206, + 69, + 355, + 182 + ], + "type": "image", + "image_path": "96ad7d5f6bc0f66a2028428fce690da69fc4747b2de093344daa46f39ffb7128.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 203, + 188, + 357, + 222 + ], + "lines": [ + { + "bbox": [ + 203, + 188, + 357, + 222 + ], + "spans": [ + { + "bbox": [ + 203, + 188, + 357, + 222 + ], + "type": "text", + "content": "Figure 5. Plots of the train/test loss convergence on Food-101 under Miss-Text.FED-PRIME converges fastest." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 385, + 69, + 533, + 182 + ], + "blocks": [ + { + "bbox": [ + 385, + 69, + 533, + 182 + ], + "lines": [ + { + "bbox": [ + 385, + 69, + 533, + 182 + ], + "spans": [ + { + "bbox": [ + 385, + 69, + 533, + 182 + ], + "type": "image", + "image_path": "814c2782251785cb72023581f7eb545e4c8ad9679dd6dc8d3d8c8a4956ee4f2f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 362, + 188, + 555, + 222 + ], + "lines": [ + { + "bbox": [ + 362, + 188, + 555, + 222 + ], + "spans": [ + { + "bbox": [ + 362, + 188, + 555, + 222 + ], + "type": "text", + "content": "Figure 6. Convergence of Euclidean distance between aggregated prompt centroids on the UPMC Food-101 under Miss Image scenario." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 236, + 297, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 236, + 297, + 416 + ], + "spans": [ + { + "bbox": [ + 54, + 236, + 297, + 416 + ], + "type": "text", + "content": "across a wide range of " + }, + { + "bbox": [ + 54, + 236, + 297, + 416 + ], + "type": "inline_equation", + "content": "\\eta \\in \\{0.00, 0.25, 0.50, 0.75, 1.00\\}" + }, + { + "bbox": [ + 54, + 236, + 297, + 416 + ], + "type": "text", + "content": " under the Miss Both training scenarios. The reported results in Fig. 1 and Fig. 4 show that our model significantly outperforms FEDAVG-P and its centralized counterpart Centralized-P. As the missing rate increases, our approach consistently maintains a competitive accuracy above " + }, + { + "bbox": [ + 54, + 236, + 297, + 416 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 54, + 236, + 297, + 416 + ], + "type": "text", + "content": ", closely approaching the upper-bound performance of FEDAVG-P in the centralized setting, which benefits from a centralized dataset with better-aligned modalities. These results demonstrate the robustness of our model across varying missing rates and highlight the effectiveness of our federated prompt-tuning approach, achieving performance comparable to a centralized prompt-tuning model despite having access only to decentralized local datasets with heterogeneous missing patterns." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 424, + 149, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 424, + 149, + 436 + ], + "spans": [ + { + "bbox": [ + 55, + 424, + 149, + 436 + ], + "type": "text", + "content": "4.2.4. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 439, + 296, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 439, + 296, + 619 + ], + "spans": [ + { + "bbox": [ + 55, + 439, + 296, + 619 + ], + "type": "text", + "content": "As explained previously in Section 4.1, FED-INTER and FED-INTRA are in variants of FED-PRIME when we drop the intra- and inter-client prompt sets from the client design, respectively. Table 1 shows that FED-PRIME consistently outperforms both FED-INTER and FED-INTRA in all test conditions with substantial gap. The lack of intra-client prompts causes FED-INTER to underperform FED-PRIME suggests that having a dedicated prompt set to capture the common and highly generalizable input-agnostic aspect of fine-tuning is essential to improve performance. Likewise, FED-INTRA, which omits the inter-client prompt set, also shows a significant performance drop. This highlights the importance of capturing and correctly aggregating input-level patterns of missing data through prompt alignment." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 626, + 178, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 626, + 178, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 626, + 178, + 639 + ], + "type": "text", + "content": "4.2.5. Convergence Analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "Model Convergence. Fig. 5 shows that FED-PRIME has superior performance, achieving faster convergence and significantly lower final loss values in both training and testing compared to FED-INTER and FED-INTRA. This indicates better learning efficiency and robustness. Both FED-INTER and FED-INTRA converge more slowly, with" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 236, + 555, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 555, + 321 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 555, + 321 + ], + "type": "text", + "content": "consistently higher loss values, especially in the early rounds. Additionally, our method is more stable across server-aggregation rounds, with minimal fluctuations. In contrast, FED-INTER and FED-INTRA exhibit significantly more variability, particularly toward the end. This highlights the effectiveness of FED-PRIME, with faster convergence and improved stability across rounds." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 332, + 556, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 332, + 556, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 556, + 441 + ], + "type": "text", + "content": "Prompt Convergence. Fig. 6 shows that FED-PRIME reduces the distance between the aggregated prompt centroids much faster than FEDAVG-P, which highlights the faster convergence rate of FED-PRIME over FEDAVG-P, indicating faster convergence and better prompt alignment. In contrast, FEDAVG-P converges more slowly and unstably, likely due to incorrect default prompt alignment. This again underscores the necessity of having a learnable prompt clustering mechanism to prevent incorrect prompt alignment." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 481, + 388, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 481, + 388, + 493 + ], + "spans": [ + { + "bbox": [ + 314, + 481, + 388, + 493 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "content": "Fine-tuning large pre-trained models for specific tasks has become an increasingly effective solution paradigm in machine learning. However, existing fine-tuning frameworks struggle in practical, multimodal settings where data is distributed across private silos with heterogeneous patterns of missing modalities and features. To address these challenges, this paper introduces a novel federated prompt-tuning approach that optimizes, aligns, and aggregates prompt instructions, compensating for gaps in data modalities and distributional patterns across a number of diverse and private local datasets. Rigorous evaluation on multimodal benchmark datasets demonstrates the effectiveness and robustness of our approach against various patterns of missing data in local datasets. We view this work as an initial step toward developing resilient, adaptable federated models that can fully leverage the diversity of multimodal data in real-world environments." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3943" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 190 + ], + "type": "text", + "content": "This research is funded by Hanoi University of Science and Technology (HUST) under grant number T2024-TD-002. Additionally, this work utilized GPU compute resource at SDSC and ACES through allocation CIS230391 from the Advanced Cyberinfrastructure Coordination Ecosystem: Services and Support (ACCESS) program [2], which is supported by U.S. National Science Foundation grants #2138259, #2138286, #2138307, #2137603, and #2138296." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 210, + 115, + 222 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 210, + 115, + 222 + ], + "spans": [ + { + "bbox": [ + 56, + 210, + 115, + 222 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 230, + 298, + 714 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 61, + 230, + 296, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 230, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 61, + 230, + 296, + 262 + ], + "type": "text", + "content": "[1] John Arevalo, Thamar Solorio, Manuel Montes y Gómez, and Fabio A. González. Gated multimodal units for information fusion. ArXiv, abs/1702.01992, 2017. 2, 6, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 263, + 296, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 263, + 296, + 339 + ], + "spans": [ + { + "bbox": [ + 61, + 263, + 296, + 339 + ], + "type": "text", + "content": "[2] Timothy J. Boerner, Stephen Deems, Thomas R. Furlani, Shelley L. Knuth, and John Towns. Access: Advancing innovation: Nsf's advanced cyberinfrastructure coordination ecosystem: Services & support. In Practice and Experience in Advanced Research Computing 2023: Computing for the Common Good, page 173-176, New York, NY, USA, 2023. Association for Computing Machinery. 9" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 341, + 298, + 701 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 341, + 298, + 701 + ], + "spans": [ + { + "bbox": [ + 61, + 341, + 298, + 701 + ], + "type": "text", + "content": "[3] Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S. Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, Erik Brynjolfsson, S. Buch, Dallas Card, Rodrigo Castellon, Niladri S. Chatterji, Annie S. Chen, Kathleen A. Creel, Jared Davis, Dora Demszky, Chris Donahue, Moussa Doumbouya, Esin Durmus, Stefano Ermon, John Etchemendy, Kawin Ethayarajh, Li Fei-Fei, Chelsea Finn, Trevor Gale, Lauren E. Gillespie, Karan Goel, Noah D. Goodman, Shelby Grossman, Neel Guha, Tatsunori Hashimoto, Peter Henderson, John Hewitt, Daniel E. Ho, Jenny Hong, Kyle Hsu, Jing Huang, Thomas F. Icard, Saahil Jain, Dan Jurafsky, Pratyusha Kalluri, Siddharth Karamcheti, Geoff Keeling, Fereshte Khani, O. Khattab, Pang Wei Koh, Mark S. Krass, Ranjay Krishna, Rohith Kuditipudi, Ananya Kumar, Faisal Ladhak, Mina Lee, Tony Lee, Jure Leskovec, Isabelle Levent, Xiang Lisa Li, Xuechen Li, Tengyu Ma, Ali Malik, Christopher D. Manning, Suvir P. Mirchandani, Eric Mitchell, Zanele Munyikwa, Suraj Nair, Avanika Narayan, Deepak Narayanan, Benjamin Newman, Allen Nie, Juan Carlos Niebles, Hamed Nilforoshan, J. F. Nyarko, Giray Ogut, Laurel Orr, Isabel Papadimitriou, Joon Sung Park, Chris Piech, Eva Portelance, Christopher Potts, Aditi Raghunathan, Robert Reich, Hongyu Ren, Frieda Rong, Yusuf H. Roohani, Camilo Ruiz, Jack Ryan, Christopher R'e, Dorsa Sadigh, Shiori Sagawa, Keshav Santhanam, Andy Shih, Krishna Parasuram Srinivasan, Alex Tamkin, Rohan Taori, Armin W. Thomas, Florian Tramér, Rose E. Wang, William Wang, Bohan Wu, Jiajun Wu, Yuhuai Wu, Sang Michael Xie, Michihiro Yasunaga, Jiaxuan You, Matei A. Zaharia, Michael Zhang, Tianyi Zhang, Xikun Zhang, Yuhui Zhang, Lucia Zheng, Kaitlyn Zhou, and Percy Liang. On the opportunities and risks of foundation models. ArXiv, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 702, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 702, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 61, + 702, + 296, + 714 + ], + "type": "text", + "content": "[4] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 105 + ], + "type": "text", + "content": "Food-101 - mining discriminative components with random forests. In European Conference on Computer Vision, 2014. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 320, + 107, + 554, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 107, + 554, + 217 + ], + "spans": [ + { + "bbox": [ + 320, + 107, + 554, + 217 + ], + "type": "text", + "content": "[5] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. CoRR, abs/2005.14165, 2020. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 320, + 219, + 554, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 219, + 554, + 272 + ], + "spans": [ + { + "bbox": [ + 320, + 219, + 554, + 272 + ], + "type": "text", + "content": "[6] Alberto Brunete, Ernesto Gambao, Miguel Hernando, and Raquel Cedazo. Smart assistive architecture for the integration of IoT devices, robotic systems, and multimodal interfaces in healthcare environments. Sensors, 21(6):2212, 2021. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 320, + 274, + 554, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 274, + 554, + 328 + ], + "spans": [ + { + "bbox": [ + 320, + 274, + 554, + 328 + ], + "type": "text", + "content": "[7] Jiayi Chen and Aidong Zhang. Hgmf: heterogeneous graph-based fusion for multimodal data with incompleteness. In Proceedings of the 26th ACM International Conference on Knowledge Discovery & Data Mining, pages 1295-1305, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 320, + 331, + 554, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 331, + 554, + 386 + ], + "spans": [ + { + "bbox": [ + 320, + 331, + 554, + 386 + ], + "type": "text", + "content": "[8] Jiayi Chen and Aidong Zhang. FedMSplit: Correlation-adaptive federated multi-task learning across multimodal split networks. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 87-96, 2022. 1, 3, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 320, + 388, + 554, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 388, + 554, + 430 + ], + "spans": [ + { + "bbox": [ + 320, + 388, + 554, + 430 + ], + "type": "text", + "content": "[9] Jinyu Chen, Wenchao Xu, Song Guo, Junxiao Wang, Jie Zhang, and Haozhao Wang. Fedtune: A deep dive into efficient federated fine-tuning with pre-trained transformers, 2022. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 433, + 554, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 433, + 554, + 487 + ], + "spans": [ + { + "bbox": [ + 315, + 433, + 554, + 487 + ], + "type": "text", + "content": "[10] Yae Jee Cho, Andre Manoel, Gauri Joshi, Robert Sim, and Dimitrios Dimitriadis. Heterogeneous ensemble knowledge transfer for training large models in Federated Learning. Computing Research Repository arXiv Preprints, 2204.12703, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 489, + 554, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 489, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 315, + 489, + 554, + 533 + ], + "type": "text", + "content": "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 535, + 554, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 535, + 554, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 535, + 554, + 622 + ], + "type": "text", + "content": "[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota, 2019. Association for Computational Linguistics. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 624, + 554, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 554, + 689 + ], + "type": "text", + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 691, + 554, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 691, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 315, + 691, + 554, + 714 + ], + "type": "text", + "content": "[14] Alireza Fallah, Aryan Mokhtari, and Asuman Ozdaglar. Personalized federated learning: A meta-learning ap" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3944" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 94 + ], + "type": "text", + "content": "proach. Computing Research Repository arXiv Preprints, 2002.07948, 2020.2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "type": "text", + "content": "[15] Ziwei Fan, Hao Ding, Anoop Deoras, and Trong Nghia Hoang. Personalized federated domain adaptation for item-to-item recommendation. In Proceedings of the Thirty-Ninth Conference on Uncertainty in Artificial Intelligence, pages 560-570. PMLR, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 151, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 151, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 151, + 294, + 205 + ], + "type": "text", + "content": "[16] Tao Guo, Song Guo, Junxiao Wang, Xueyang Tang, and Wenchao Xu. PromptFL: Let Federated Participants Cooperatively Learn Prompts Instead of Models-Federated Learning in Age of Foundation Model. IEEE Transactions on Mobile Computing, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 294, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 294, + 258 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 294, + 258 + ], + "type": "text", + "content": "[17] Minh Hoang, Nghia Hoang, Bryan Kian Hsiang Low, and Carleton Kingsford. Collective model fusion for multiple black-box experts. In Proceedings of the 36th International Conference on Machine Learning, pages 2742-2750. PMLR, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 295, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 295, + 315 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 295, + 315 + ], + "type": "text", + "content": "[18] Trong Nghia Hoang, Quang Minh Hoang, Kian Hsiang Low, and Jonathan P. How. Collective online learning of Gaussian processes in massive multi-agent systems. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7850-7857, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 316, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 358 + ], + "type": "text", + "content": "[19] Parminder Kaur, Husanbir Singh Pannu, and Avleen Kaur Malhi. Comparative analysis on cross-modal information retrieval: A review. Computer Science Review, 39:100336, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 361, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 294, + 403 + ], + "type": "text", + "content": "[20] Muhammad Uzair khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 404, + 294, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 458 + ], + "type": "text", + "content": "[21] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 5583-5594. PMLR, 2021. 1, 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 460, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 525 + ], + "type": "text", + "content": "[22] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A. Shamma, Michael S. Bernstein, and Li Fei-Fei. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International Journal of Computer Vision, 123:32 - 73, 2016. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 526, + 294, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 294, + 557 + ], + "type": "text", + "content": "[23] Harold W Kuhn. The Hungarian Method for the Assignment Problem. Naval Research Logistics Quarterly, 2(1-2):83-97, 1955. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 559, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 602 + ], + "type": "text", + "content": "[24] Yi-Lun Lee, Yi-Hsuan Tsai, Wei-Chen Chiu, and Chen-Yu Lee. Multimodal prompting with missing modalities for visual recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3, 6, 7, 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 604, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 604, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 604, + 294, + 668 + ], + "type": "text", + "content": "[25] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3045-3059, Online and Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[26] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. Proceedings of Machine Learning and Systems, 2:429-450, 2020. 2, 4" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[27] Xiang Li, Kaixuan Huang, Wenhao Yang, Shusen Wang, and Zhihua Zhang. On the convergence of FedAvg on non-IID data. Computing Research Repository arXiv Preprints, 1907.02189t, 2019. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 118, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 171 + ], + "type": "text", + "content": "[28] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, 2014. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 173, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 553, + 205 + ], + "type": "text", + "content": "[29] Wang Lu, Xixu Hu, Jindong Wang, and Xing Xie. FedCLIP: Fast Generalization and Personalization for CLIP in Federated Learning, 2023. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 206, + 553, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 248 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 248 + ], + "type": "text", + "content": "[30] Tengfei Ma, Trong Nghia Hoang, and Jie Chen. Federated learning of models pre-trained on different features with consensus graphs. In Uncertainty in Artificial Intelligence, pages 1336-1346, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 251, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 304 + ], + "type": "text", + "content": "[31] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial Intelligence and Statistics, pages 1273-1282, 2017. 2, 3, 4, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 305, + 553, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 553, + 347 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 553, + 347 + ], + "type": "text", + "content": "[32] John Nguyen, Jianyu Wang, Kshitiz Malik, Maziar Sanjabi, and Michael Rabbat. Where to Begin? On the Impact of Pre-Training and Initialization in Federated Learning, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 350, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 403 + ], + "type": "text", + "content": "[33] Manh Duong Nguyen, Trung Thanh Nguyen, Huy-Hieu Pham, Trong Nghia Hoang, Phi Le Nguyen, and Thanh Trung Huynh. Fedmac: Tackling partial-modality missing in federated learning with cross-modal aggregation and contrastive regularization. 2024. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 405, + 553, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 405, + 553, + 481 + ], + "spans": [ + { + "bbox": [ + 316, + 405, + 553, + 481 + ], + "type": "text", + "content": "[34] Nang Hung Nguyen, Phi Le Nguyen, Thuy Dung Nguyen, Trung Thanh Nguyen, Duc Long Nguyen, Thanh Hung Nguyen, Huy Hieu Pham, and Thao Nguyen Truong. FedDRL: Deep reinforcement learning-based adaptive aggregation for non-IID data in federated learning. In Proceedings of the 51st International Conference on Parallel Processing, pages 1-11, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 483, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 483, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 316, + 483, + 553, + 569 + ], + "type": "text", + "content": "[35] Nang Hung Nguyen, Duc Long Nguyen, Trong Bang Nguyen, Thanh-Hung Nguyen, Huy Hieu Pham, Truong Thao Nguyen, and Phi Le Nguyen. CADIS: Handling cluster-skewed Non-IID data in federated learning with clustered aggregation and knowledge distilled regularization. In Proceedings of the 23rd IEEE/ACM International Symposium on Cluster, Cloud and Internet Computing, pages 249-261, 2023." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 571, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 571, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 571, + 553, + 645 + ], + "type": "text", + "content": "[36] Quang Ha Pham, Nang Hung Nguyen, Thanh Hung Nguyen, Huy Hieu Pham, Phi Le Nguyen, and Truong Thao Nguyen. SEM: A simple yet efficient model-agnostic local training mechanism to tackle data sparsity and scarcity in federated learning. In Proceedings of the 11st International Symposium on Computing and Networking, pages 120–126, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "text", + "content": "[37] Thu Hang Phung, Binh P Nguyen, Thanh Hung Nguyen, Quoc Viet Hung Nguyen, Phi Le Nguyen, and Thanh Trung Huynh. A contrastive learning and graph-based approach for missing modalities in multimodal federated learning. In Proceedings of the 2024 International Joint Conference on Neural Networks, pages 1-8, 2024. 1, 3" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3945" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "text", + "content": "[38] Petra Poklukar, Miguel Vasco, Hang Yin, Francisco S Melo, Ana Paiva, and Danica Kragic. Geometric multimodal contrastive representation learning. In Proceedings of the 2022 International Conference on Machine Learning, pages 17782-17800, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 296, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 296, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 296, + 195 + ], + "type": "text", + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. CoRR, abs/2103.00020, 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 197, + 296, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 296, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 296, + 252 + ], + "type": "text", + "content": "[40] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21(140):1-67, 2020. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 253, + 296, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 296, + 308 + ], + "type": "text", + "content": "[41] Kuniaki Saito, Kihyuk Sohn, Xiang Zhang, Chun-Liang Li, Chen-Yu Lee, Kate Saenko, and Tomas Pfister. Prefix conditioning unifies language and label supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2861-2870, 2023. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 309, + 296, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 296, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 296, + 364 + ], + "type": "text", + "content": "[42] Rachael Hwee Ling Sim, Yehong Zhang, Trong Nghia Hoang, Xinyi Xu, Bryan Kian Hsiang Low, and Patrick Jaillet. Incentives in private collaborative machine learning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 365, + 296, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 296, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 296, + 408 + ], + "type": "text", + "content": "[43] Canh T Dinh, Nguyen Tran, and Josh Nguyen. Personalized federated learning with Moreau envelopes. Advances in Neural Information Processing Systems, 33:21394-21405, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 411, + 296, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 296, + 464 + ], + "type": "text", + "content": "[44] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, S. M. Ali Eslami, Oriol Vinyals, and Felix Hill. Multimodal few-shot learning with frozen language models. In Advances in Neural Information Processing Systems, pages 200-212. Curran Associates, Inc., 2021. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 467, + 296, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 296, + 521 + ], + "type": "text", + "content": "[45] Feng Wang, M Cenk Gursoy, and Senem Velipasalar. Communication-efficient and privacy-preserving feature-based federated transfer learning. In Proceedings of the 2022 IEEE Global Communications Conference, pages 3875-3880, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 523, + 296, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 296, + 578 + ], + "type": "text", + "content": "[46] Xin Wang, Devinder Kumar, Nicolas Thome, Matthieu Cord, and Frédéric Precioso. Recipe recognition with large multimodal food dataset. In 2015 IEEE International Conference on Multimedia & Expo Workshops (ICMEW), pages 1-6, 2015. 2, 6, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 579, + 296, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 296, + 634 + ], + "type": "text", + "content": "[47] Pei-Yau Weng, Minh Hoang, Lam M. Nguyen, My T. Thai, Tsui-Wei Weng, and Trong Nghia Hoang. Probabilistic federated prompt-tuning with non-IID and imbalanced data. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 635, + 296, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 296, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 296, + 669 + ], + "type": "text", + "content": "[48] Baochen Xiong, Xiaoshan Yang, Fan Qi, and Changsheng Xu. A unified framework for multi-modal Federated Learning. Neurocomputing, 480:110-118, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 670, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 296, + 714 + ], + "type": "text", + "content": "[49] Baochen Xiong, Xiaoshan Yang, Yaguang Song, Yaowei Wang, and Changsheng Xu. Client-adaptive cross-model reconstruction network for modality-incomplete multimodal federated learning. In Proceedings of the 31st ACM Interna-" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 666 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "text", + "content": "tional Conference on Multimedia, pages 1241-1249, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "text", + "content": "[50] Fu-En Yang, Chien-Yi Wang, and Yu-Chiang Frank Wang. Efficient Model Personalization in Federated Learning via Client-Specific Prompt Generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 19159-19168, 2023. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "text", + "content": "[51] Guan Yu, Quefeng Li, Dinggang Shen, and Yufeng Liu. Optimal sparse linear prediction for block-missing multimodality data without imputation. Journal of the American Statistical Association, 115(531):1406-1419, 2020. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "type": "text", + "content": "[52] Songcan Yu, Junbo Wang, Walid Hussein, and Patrick C.K. Hung. Robust multimodal federated learning for incomplete modalities. Computer Communications, 214:234-243, 2024. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 242, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 242, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 242, + 553, + 297 + ], + "type": "text", + "content": "[53] Mikhail Yurochkin, Mayank Agarwal, Soumya Ghosh, Kristjan Greenewald, Trong Nghia Hoang, and Yasaman Khazaeni. Bayesian nonparametric federated learning of neural networks. In International Conference on Machine Learning, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 297, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 352 + ], + "type": "text", + "content": "[54] Mikhail Yurochkin, Mayank Argawal, Soumya Ghosh, Kristjan Greenewald, and Trong Nghia Hoang. Statistical model aggregation via parameter matching. In Advances in Neural Information Processing Systems, pages 10954-10964, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 354, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 397 + ], + "type": "text", + "content": "[55] Changqing Zhang, Yajie Cui, Zongbo Han, Joey Tianyi Zhou, Huazhu Fu, and Qinghua Hu. Deep partial multi-view learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(5):2402-2415, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 399, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 453 + ], + "type": "text", + "content": "[56] Jianyi Zhang, Hao Frank Yang, Ang Li, Xin Guo, Pu Wang, Haiming Wang, Yiran Chen, and Hai Li. Mllm-fl: Multimodal large language model assisted federated learning on heterogeneous and long-tailed data. arXiv preprint arXiv:2409.06067, 2024. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 454, + 553, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 454, + 553, + 497 + ], + "spans": [ + { + "bbox": [ + 316, + 454, + 553, + 497 + ], + "type": "text", + "content": "[57] Wei Zhang and Xiang Li. Data privacy preserving federated transfer learning in machinery fault diagnostics using prior distributions. Structural Health Monitoring, 21(4):1329-1344, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 498, + 553, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 553, + 542 + ], + "type": "text", + "content": "[58] Haodong Zhao, Wei Du, Fangqi Li, Peixuan Li, and Gongshen Liu. FedPrompt: Communication-Efficient and Privacy Preserving Prompt Tuning in Federated Learning. In Proc. ICASSP, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 544, + 553, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 553, + 587 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 553, + 587 + ], + "type": "text", + "content": "[59] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 589, + 553, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 589, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 316, + 589, + 553, + 621 + ], + "type": "text", + "content": "[60] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision (IJCV), 2022. 1, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 623, + 553, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 623, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 316, + 623, + 553, + 666 + ], + "type": "text", + "content": "[61] Tongxue Zhou, Pierre Vera, Stephane Canu, and Su Ruan. Missing data imputation via conditional generator and correlation learning for multimodal brain tumor segmentation. Pattern Recognition Letters, 158:125-132, 2022. 3" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3946" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Federated Representation Angle Learning/f94f188a-83f1-436a-a6d8-36878318e078_content_list.json b/2025/Federated Representation Angle Learning/f94f188a-83f1-436a-a6d8-36878318e078_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ac39660e7e79604da518fd0f3696176136def162 --- /dev/null +++ b/2025/Federated Representation Angle Learning/f94f188a-83f1-436a-a6d8-36878318e078_content_list.json @@ -0,0 +1,2162 @@ +[ + { + "type": "text", + "text": "Federated Representation Angle Learning*", + "text_level": 1, + "bbox": [ + 285, + 130, + 718, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Liping Yi $^{1}$ , Han Yu $^{2}$ , Gang Wang $^{1,\\ast}$ , Xiaoguang Liu $^{1,\\ast}$ , Xiaoxiao Li $^{3,4}$ $^{1}$ College of Computer Science, TMCC, SysNet, DISSec, GTIISC, Nankai University, Tianjin, China \n $^{2}$ College of Computing and Data Science, Nanyang Technological University, Singapore \n $^{3}$ Department of Electrical and Computer Engineering, University of British Columbia, Vancouver, Canada \n $^{4}$ Vector Institute, Vancouver, Canada \n{yiliping, wgzwp, liuxg}@nbjl.nankai.edu.cn, han.yu@ntu.edu.sg, xiaoxiao.li@ece.ubc.ca", + "bbox": [ + 104, + 179, + 944, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 321, + 326, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Model-heterogeneous federated learning (MHFL) is a challenging FL paradigm designed to allow FL clients to train structurally heterogeneous models under the coordination of an FL server. Existing MHFL methods face significant limitations when it comes to transferring global knowledge to clients as a result of sharing only partial homogeneous model parameters or calculating distance loss, leading to inferior model generalization. To bridge this gap, we propose a novel model-heterogeneous Federated learning method with Representation Angle Learning (FedRAL). It consists of three innovative designs: (1) We first introduce representation angle learning into MHFL. Specifically, we embed a homogeneous square matrix into the local heterogeneous model of each client, which learns the angle information of local representations. These homogeneous representation angle square matrices are aggregated on the server to fuse representation angle knowledge shared by clients for enhancing the generalization of local representations. (2) As different clients might have heterogeneous system resources, we propose an adaptive diagonal sparsification strategy to reduce the numbers of the parameters of representation angle square matrices uploaded to the server, to improve FL communication efficiency. (3) To enable effective fusion of sparsified homogeneous local representation angle square matrices, we design an element-wise weighted aggregation approach. Experiments on 4 benchmark datasets under 2 types of non-IID divisions over 6 state-of-the-art baselines demonstrate that FedRAL achieves the best performance. It improves test accuracy, communication efficiency and computational efficiency by up to $5.03\\%$ , $12.43\\times$ and $6.49\\times$ , respectively.", + "bbox": [ + 89, + 361, + 485, + 830 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 321, + 643, + 337 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Federated learning (FL) [13, 14, 24, 36, 45, 55, 56] is a privacy-preserved distributed machine learning paradigm. Typically, an FL server is often involved in broadcast a global model to FL clients. The clients then further train the global model on local data to obtain local models. The server aggregates local model updates received from the clients to produce a new global model. In this way, client data are not exposed. Under this setting, the models trained by the clients must follow the same structure.", + "bbox": [ + 511, + 349, + 906, + 484 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In practice, clients participating in FL often have non-independent and identically distributed (non-IID) data, a.k.a., data heterogeneity [27, 42, 50, 58, 75, 78, 79], the local models trained on such data are biased, so the global model obtained by aggregating them may not perform well on all client data. Besides, FL clients also have heterogeneous system configurations in terms of computing power and communication bandwidth, a.k.a., system heterogeneity [39, 57, 59, 60, 67]. Thus, requesting all clients to train the same model might not be viable for some weaker clients, while leaving stronger clients under-utilized. More critically, some companies or institutions joining FL as clients might face intellectual property issues regarding their proprietary local models. Thus, they might be reluctant to share their private heterogeneous models, a.k.a., model heterogeneity. These challenges have inspired the field of model-heterogeneous federated learning (MHFL) with preserved data and model privacy, which advocates avoiding directly sharing private local heterogeneous models [61-65].", + "bbox": [ + 511, + 489, + 908, + 776 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing MHFL approaches mainly include: a) Modeldecoupling: each client model is decoupled into heterogeneous and homogeneous parts, with only homogeneous parts being shared with the server [11, 30]. b) Knowledge distillation: the server aggregates local representation or logits from different clients to generate global representation or logits, which is used to calculate distillation loss for local training [28, 31]. c) Mutual training: each client is", + "bbox": [ + 511, + 779, + 910, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 886, + 236, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1314", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "assigned a shared homogeneous small model and trains it with the local heterogeneous model by mutual loss between them [49, 53]. Although these methods support model heterogeneity, their capability for transferring global knowledge from the server to clients is limited due to sharing partial model parameters or using distillation loss, thereby constraining the generalization of local models. Whereas, model representations often involve more sufficient semantic knowledge than partial model parameters or distillation loss, which are explored in this work for knowledge sharing.", + "bbox": [ + 89, + 90, + 480, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A recent work [44] empirically verified that parameter fine-tuning approaches (e.g., LoRA [18]) cannot sufficiently retain semantic information of the pre-trained model. Whereas, fine-tuning the angles (i.e., gradient directions) of the pre-trained model can achieve this, thereby enhancing the generalization of the pre-trained model. The work further proposed the Orthogonal Fine-Tuning (OFT) method to fine-tune the parameter angles of the pre-trained models via an orthogonal matrix (Figure 1). In short, parameter angles involve sufficient semantic information that can enhance model generalization. Therefore, we can enhance FL model generalization through sharing parameter angles instead of model parameters. However, the FL server can not aggregate parameter angles from structure-heterogeneous client models. Considering representations also with sufficient semantic knowledge, we attempt to share homogeneous representation angles among clients to achieve generalization-enhanced model-heterogeneous FL.", + "bbox": [ + 89, + 246, + 482, + 518 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we propose a novel model-heterogeneous Federated learning method with Representation Angel Learning (FedRAL) to enhance the transfer of generalizable global knowledge from the FL server to FL clients. It consists of three innovative designs: (1) Since representations are extracted high-level features from data, with sufficient semantic information, FedRAL embeds a homogeneous square matrix into the local heterogeneous model to learn the angle information of local extracted representations. The server aggregates local homogeneous representation angle square matrices from different clients to generate a global representation angle square matrix, thereby fusing cross-client representation angle knowledge. Then, the global representation angle square matrix with sufficient generalized representation angle information is used to enhance the generalization of local representations during local model training. Different from the strong assumption of the orthogonality of the matrix in OFT, the homogeneous representation angle square matrix is initialized randomly. (2) Considering the diverse system resources (e.g., communication bandwidth) of different clients, we design an adaptive diagonal sparsification strategy for the homogeneous representation angle square matrix that allows clients to upload diagonal blocks with sizes matching with communication capacity. (3) To effectively aggregate sparsified local", + "bbox": [ + 89, + 523, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/304dd7488f93e82b9d90d9fb9eac10b293fd18933f5a8073bd4c51e1ce16d80f.jpg", + "image_caption": [ + "Figure 1. Orthogonal Fine-Tuning (OFT)." + ], + "image_footnote": [], + "bbox": [ + 516, + 87, + 906, + 161 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "homogeneous representation angle square matrices on the server, element-wise weighted aggregation is designed.", + "bbox": [ + 511, + 191, + 903, + 222 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experiments on 4 benchmark datasets under 2 modes of non-IID divisions against 6 advanced MHFL methods demonstrate the superiority of FedRAL. Compared with the state-of-the-art baseline, FedRAL improves test accuracy by up to $5.03\\%$ , while enhancing communication and computational efficiency by up to $12.43\\times$ and $6.49\\times$ .", + "bbox": [ + 511, + 222, + 903, + 313 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 328, + 653, + 343 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on different approaches used for achieving MHFL, existing methods can be divided into 4 categories.", + "bbox": [ + 511, + 354, + 903, + 383 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Heterogeneous Subnet. Some works [3-5, 12, 17, 33, 66, 77] assume that each client trains local heterogeneous subnets of the global model, and the server aggregates them according to parameter positions to re-construct a new global model. They allow each client to train personalized heterogeneous subnets to tackle data and system heterogeneity. However, uploading subnets exposes model parameters, thereby breaching model IP protection requirements.", + "bbox": [ + 511, + 385, + 905, + 505 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Model Decoupling. These methods [7, 11, 22, 30, 32, 38, 41, 62] share partial local model parameters instead of the entire local model for global aggregation. Specifically, they decouple each client's local heterogeneous model into heterogeneous and homogeneous parts, and only share homogeneous parts for cross-client knowledge exchange. Nevertheless, sharing partial model parameters still violates model IP protection needs. Furthermore, the unshared remaining parameters might overfit due to local training.", + "bbox": [ + 511, + 506, + 905, + 642 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Knowledge Distillation. These methods [1, 2, 6, 8, 9, 15, 16, 19-21, 23, 28, 29, 31, 34, 37, 40, 46, 47, 51, 52, 68, 69, 71-73, 76] aggregate the local seen-class representations or logits extracted on local private data or a public dataset to generate the global representation or logits by class, which is used to calculate distillation loss with the local representations or logits. Although they protect local model structures, the distillation loss only transfers limited knowledge from the server to clients. Besides, sharing class information might be prohibited in some data privacy-sensitive applications. The public dataset used must follow similar data distributions with local private data, which is difficult to access due to data privacy.", + "bbox": [ + 511, + 643, + 905, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Mutual Learning. These methods [25, 35, 43, 49, 53] add a global homogeneous small model shared by all clients, which is alternately trained with the local heterogeneous model by the mutual loss between the outputs of", + "bbox": [ + 511, + 839, + 905, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1315", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the two models for each sample. Although they hide local heterogeneous model structures, the mutual loss only transfers limited knowledge between the two models. The unstable mutual loss in early training rounds might result in both models not converging.", + "bbox": [ + 89, + 90, + 482, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing MHFL methods generally transfer limited generalized knowledge from the server to clients, leading to model performance bottlenecks. FedRAL effectively tackles data, system and model heterogeneity simultaneously, allowing clients to share a global homogeneous representation angle square matrix for learning generalized representation angle information. It bridges this important gap by enhancing the generalization of local representations via utilizing the global homogeneous representation angle square matrix to fine-tune local representation angles, while preserving personalized semantic information.", + "bbox": [ + 89, + 167, + 483, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 349, + 230, + 366 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Orthogonal Fine-Tuning", + "text_level": 1, + "bbox": [ + 89, + 375, + 316, + 391 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fine-tuning the parameters of a pre-trained model has been demonstrated to compromise its semantic information, while fine-tuning its parameter angles (i.e., gradient directions) can sufficiently retain its original semantic information [44]. This enables the preservation of the generalized ability when adapting to downstream tasks. Hypersphere Energy (HE) measures the uniformity of neuron distribution on the unit hypersphere. A lower HE value is beneficial to keeping pre-trained model generalization ability. To minimize HE variance between the pre-trained model $W^0 \\in \\mathbb{R}^{u \\times v}$ and the angle-fine-tuned model $W \\in \\mathbb{R}^{u \\times v}$ , orthogonal finetuning (OFT) [44] fine-tunes the parameter angles of the pre-trained model through an orthogonal matrix $O \\in \\mathbb{R}^{u \\times u}$ , $W = O \\times W^0$ , as depicted in Figure 1. To ensure orthogonality, OFT utilizes Cayley parameterization to construct matrix $O = (I + Q)(I - Q)^{-1}$ , where $Q \\in \\mathbb{R}^{u \\times u}$ is a skew-symmetric matrix satisfying $Q = -Q^T$ . Due to inverse operations, updating the orthogonal matrix $O$ during fine-tuning introduces high computational overhead, especially with a large dimension $u$ .", + "bbox": [ + 89, + 397, + 483, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike OFT, we explore fine-tuning the angles of representations with sufficient semantic information. We relax the orthogonality assumption and avoid expensive inverse operations. Instead, we randomly initialize the homogeneous representation angle square matrix, which is embedded into the local heterogeneous model. They are trained in an end-to-end manner to improve computation efficiency.", + "bbox": [ + 89, + 699, + 483, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Problem Formulation", + "text_level": 1, + "bbox": [ + 89, + 816, + 294, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In FedRAL, a server coordinates $K$ clients with heterogeneous local models for collaborative training. Client $k$ 's local model $f_{k}(\\omega_{k})$ consists of two parts: 1) a heterogeneous feature extractor $f_{k}^{ex}(\\omega_{k}^{ex})$ , and 2) a homogeneous predic", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0d11c5368c380c506206ab3935ca190e4afb5fa809eeb4bad025bc4c81090165.jpg", + "image_caption": [ + "Figure 2. The workflow of FedRAL." + ], + "image_footnote": [], + "bbox": [ + 517, + 89, + 906, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tion head $f^{hd}(\\omega_k^{hd})$ (all clients conduct the same prediction task), $f_{k}(\\omega_{k}) = f_{k}^{ex}(\\omega_{k}^{ex})\\circ f^{hd}(\\omega_{k}^{hd})$ , as depicted in Figure 2. Each client trains a local homogeneous square matrix $\\mathcal{A}_k$ to learn the angle information of local representations extracted by the heterogeneous feature extractor. The server aggregates local homogeneous representation angle square matrices from different clients to generate a global representation angle square matrix $\\mathcal{A}$ for cross-client representation angle knowledge fusion. Therefore, the training objective of FedRAL is to minimize the loss sum of all heterogeneous client models with the help of the global homogeneous representation angle square matrix $\\mathcal{A}$ :", + "bbox": [ + 511, + 382, + 906, + 565 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathcal {A}, \\left\\{\\omega_ {0}, \\dots , \\omega_ {K - 1} \\right\\}} \\sum_ {k = 0} ^ {K - 1} \\ell \\left(\\mathcal {F} _ {k} \\left(D _ {k}; \\left(\\omega_ {k} \\circ \\mathcal {A}\\right)\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 577, + 905, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\ell$ is loss. $\\mathcal{F}_k(\\omega_k\\circ \\mathcal{A})$ denotes the combination of the local heterogeneous model $\\omega_{k}$ and the global representation angle square matrix $\\mathcal{A}$ , $D_{k}$ is the non-IID data of client $k$ .", + "bbox": [ + 511, + 631, + 905, + 676 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. The Proposed FedRAL Approach", + "text_level": 1, + "bbox": [ + 511, + 691, + 818, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "FedRAL consists of three innovative designs: (1) homogeneous representation angle square matrix learning, (2) adaptive diagonal sparsification, and (3) element-wise weighted homogeneous square matrix aggregation. The workflow of FedRAL in a single communication round includes the following steps.", + "bbox": [ + 511, + 718, + 905, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview. As shown in Figure 2, in the $t$ -th communication round, FedRAL executes 4 steps:", + "bbox": [ + 511, + 809, + 903, + 840 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "① The server broadcasts the latest global representation angle square matrix $\\mathcal{A}^{t-1}$ to $K$ participating clients.", + "$②$ A client $k$ embeds the global representation angle square matrix $\\mathcal{A}^{t - 1}$ into its local model $f_{k}(\\omega_{k}^{t - 1})$ to enhance" + ], + "bbox": [ + 506, + 840, + 903, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1316", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the generalization of local representations, while learning the personalized angle information of local representations via end-to-end training.", + "bbox": [ + 102, + 90, + 482, + 135 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "③ The local homogeneous representation angle square matrix $\\mathcal{A}_k^t$ after local training is sparsified into $\\tilde{\\mathcal{A}}_k^t$ via the proposed adaptive diagonal sparsification strategy before being uploaded to the server.", + "(4) The server aggregates sparsified local homogeneous representation angle square matrices $\\{\\tilde{\\mathcal{A}}_0^t,\\dots ,\\tilde{\\mathcal{A}}_{K - 1}^t\\}$ via the proposed element-wise weighted aggregation to generate a new global representation angle square matrix $\\mathcal{A}^t$" + ], + "bbox": [ + 83, + 136, + 482, + 263 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inference. These steps are iteratively executed until heterogeneous client models $\\{f_0(\\omega_0),\\dots ,f_{K - 1}(\\omega_{K - 1})\\}$ converge, which are used for inference. Algorithm 1 describes the detailed workflow of FedRAL.", + "bbox": [ + 89, + 263, + 482, + 323 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Representation Angle Learning", + "text_level": 1, + "bbox": [ + 89, + 333, + 369, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During the $t$ -th training round, a client $k$ uses its local heterogeneous feature extractor $f_{k}^{ex}\\left(\\omega_{k}^{ex,t - 1}\\right)$ to extract the personalized representation $\\mathcal{R} \\in \\mathbb{R}^r$ of the input sample $\\pmb{x}$ :", + "bbox": [ + 89, + 354, + 482, + 401 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {R} = f _ {k} ^ {e x} \\left(\\boldsymbol {x}; \\omega_ {k} ^ {e x, t - 1}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 411, + 480, + 430 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, the extracted personalized representation $\\mathfrak{R} \\in \\mathbb{R}^r$ is multiplied by the received global representation angle square matrix $\\mathcal{A}^{t-1} \\in \\mathbb{R}^{r \\times r}$ to produce the representation $\\tilde{\\mathcal{R}} \\in \\mathbb{R}^r$ with enhanced generalization:", + "bbox": [ + 89, + 440, + 483, + 501 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathcal {R}} = \\mathcal {R} \\times \\mathcal {A} ^ {t - 1}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 512, + 480, + 529 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, $\\mathcal{R} \\in \\mathbb{R}^r$ and $\\hat{\\mathcal{R}} \\in \\mathbb{R}^r$ are summed and fed into the prediction head to make a prediction:", + "bbox": [ + 89, + 541, + 482, + 571 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} = f ^ {h d} \\left(\\mathcal {R} + \\tilde {\\mathcal {R}}; \\omega_ {k} ^ {h d, t - 1}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 580, + 480, + 601 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The loss $\\ell$ (e.g., Cross-Entropy loss [74]) between the prediction $\\hat{y}$ and the ground-truth label $y$ is used to update the heterogeneous client model $f_{k}(\\omega_{k}^{t - 1})$ and the homogeneous representation angle square matrix $\\mathcal{A}^{t - 1}$ simultaneously via end-to-end training:", + "bbox": [ + 89, + 612, + 482, + 689 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {k} \\left(\\omega_ {k} ^ {t}\\right) \\leftarrow f _ {k} \\left(\\omega_ {k} ^ {t - 1}\\right) - \\eta_ {\\omega} \\nabla \\ell (\\hat {y}, y), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 699, + 482, + 724 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mathcal {A}} _ {k} ^ {t} \\leftarrow \\boldsymbol {\\mathcal {A}} ^ {t - 1} - \\eta_ {\\boldsymbol {\\mathcal {A}}} \\nabla \\ell (\\hat {y}, y),\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 720, + 383, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\eta_{\\omega}$ and $\\eta_{\\mathcal{A}}$ are the learning rates of the local heterogeneous model and the homogeneous representation angle square matrix. We set $\\eta_{\\omega} = \\eta_{\\mathcal{A}}$ by default as they are updated at the same time.", + "bbox": [ + 89, + 750, + 482, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The shared global homogeneous representation angle square matrix serves three functions during local training: a) it transfers globally generalizable representation angle knowledge from the server to the client by producing the generalization-enhanced representation; b) it learns the personalized representation angle knowledge from the", + "bbox": [ + 89, + 810, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2e5e6187e943bfa0893d88b814a31a8597dee93330407dfd89acffe1353a16f0.jpg", + "image_caption": [ + "Figure 3. An illustrative example of diagonal sparsification." + ], + "image_footnote": [], + "bbox": [ + 563, + 87, + 856, + 309 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "extracted personalized representations, thereby transferring local personalized representation angle knowledge to the server; and c) the summed representation includes both generalized and personalized knowledge, thereby improving model expressiveness while addressing non-IID problem.", + "bbox": [ + 511, + 349, + 903, + 426 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Adaptive Diagonal Sparsification", + "text_level": 1, + "bbox": [ + 511, + 436, + 803, + 454 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To improve communication efficiency while maintaining model performance, we design an adaptive diagonal sparsification strategy to sparsify the local representation angle square matrices before uploading them to the FL server.", + "bbox": [ + 511, + 459, + 903, + 520 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Each element in $\\mathcal{R} \\in \\mathbb{R}^r$ corresponds to the element on the diagonal of $\\mathcal{A} \\in \\mathbb{R}^{r \\times r}$ . If we only retain the diagonal elements of $\\mathcal{A}$ and set other elements to be 0, the angle of each element in the representation is directly adjusted by the corresponding element in the square matrix diagonal, as depicted in Figure 3(a). When all elements of the square matrix are non-zero, non-diagonal elements might disturb representation angle adjustment due to multiplication with non-corresponding representation elements. As Figure 3(b) shows, the enhanced representation processed by the entire representation angle square matrix is significantly different from the one processed by the sparse square matrix with only diagonal elements.", + "bbox": [ + 511, + 521, + 905, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since adjacent elements in the representation are semantically correlated, retaining partial elements wrapped with the diagonal in the representation angle square matrix can enhance the angle correlation of adjacent representation elements. Therefore, the proposed sparsification strategy retains the elements from the diagonal blocks of the square matrix and masks other elements with 0, as shown in Figure 3(c). For a $\\mathbb{R}^{r\\times r}$ square matrix, $m$ diagonal blocks are retained. Each diagonal block is $\\mathbb{R}\\frac{r}{m}\\times \\frac{r}{m}$ .", + "bbox": [ + 511, + 718, + 905, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Clients with different communication bandwidths can select different numbers of diagonal blocks $m_{k}$ for the local homogeneous representation angle matrix $\\mathcal{A}_k^t$ to produce", + "bbox": [ + 511, + 854, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1317", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c84bfffbab70fda61ed42c034b425b1470b148704e38315638cb09b615ad01b7.jpg", + "image_caption": [ + "(a) Element-wise Averaging:" + ], + "image_footnote": [], + "bbox": [ + 94, + 99, + 168, + 165 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/59df6510412a5ae89ad362ae640d332f0542b789028dabc68d185101466451d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 171, + 99, + 272, + 165 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/71511d1e88e9f2a642040eca27f286bd6ccd20ceb5622a5e8203b99292402089.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 276, + 99, + 377, + 166 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/17aa1faac55a7d2153dec4449e33905b7a8164effba59e275e53316e8f4576e0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 380, + 99, + 480, + 165 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5fecec498424cf6df704e8692a1a4b82025efb7bc10537ae41711e76599c0175.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 179, + 166, + 234 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ca9ad1436886434ff7568a2ea7c22a3cdee033773fbe5a2a7fedb58beba2b13a.jpg", + "image_caption": [ + "(b) Element-wise Weighted Averaging:", + "Figure 4. Element-wise averaging (a) without or (b) with weights." + ], + "image_footnote": [], + "bbox": [ + 168, + 178, + 272, + 244 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cc35316f230016d817659cce7fd92ef307926761b45267fc3b76b89c11b0a96f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 178, + 305, + 210 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/54e56c77265e9e9106e3a06525b8bd7cbd4b79e0d59fb7fa5b8c03389841a33d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 307, + 178, + 375, + 244 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ef010b041291657e0d7599d6aa9dc3a5ec8c46a4c67e8df9db057f4cfc30080a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 375, + 178, + 406, + 210 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d24e31fd7f90d36683d55659ac068aa66e3b2ada183b9e452c261e39d00e64b6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 178, + 480, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the sparsified representation angle square matrix $\\tilde{\\mathcal{A}}_k^t$ :", + "bbox": [ + 89, + 280, + 442, + 301 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathcal {A}} _ {k} ^ {t} = \\text {D i a g o n a l - S p a r s i f i c a t i o n} \\left(\\mathcal {A} _ {k} ^ {t}; m _ {k}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 309, + 482, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Element-wise Weighted Aggregation", + "text_level": 1, + "bbox": [ + 89, + 338, + 405, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After receiving the sparsified local homogeneous representation square matrices from $K$ clients in the $t$ -th communication round, the server aggregates them according to the element positions with weighted averaging:", + "bbox": [ + 89, + 359, + 483, + 421 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mathcal {A}} ^ {t} = \\sum_ {k = 0} ^ {K - 1} \\frac {n _ {k}}{n} \\tilde {\\boldsymbol {\\mathcal {A}}} _ {k} ^ {t}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 222, + 429, + 482, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $n_k$ is the data volume of the client $k$ , and $n$ is the total data volume of $K$ clients.", + "bbox": [ + 89, + 478, + 482, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This element-wise weighted aggregation rule provides two benefits. a) All clients upload the diagonal elements of local homogeneous representation angle square matrices. Using $\\frac{n_k}{n}$ to weigh the diagonal elements can sufficiently reflect the overall data distribution across all clients, improving the generalization of the newly generated global square matrix. b) As the example given in Figure 4(a), for non-diagonal elements, if we directly average overlapped them by element, when one non-diagonal element is only provided by one client, then it is the value of this non-diagonal element. In this case, these non-overlapped non-diagonal elements may negatively affect the clients who do not provide for these elements. Instead, as shown in Figure 4(b), weighted averaging scales down such elements and then reduce negative effects while improving the generalization of the newly generated global square matrix.", + "bbox": [ + 89, + 508, + 483, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experimental Evaluation", + "text_level": 1, + "bbox": [ + 89, + 762, + 325, + 780 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To evaluate the performance of FedRAL ${}^{1}$ ,we compare it with 6 state-of-the-art existing approaches on 4 widely adopted benchmark datasets under 2 types of data heterogeneity settings. We implement FedRAL and baselines using pytorch and run them on 4 NVIDIA GeForce RTX 3090 GPUs (24GB memory).", + "bbox": [ + 89, + 787, + 483, + 878 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1: FedRAL", + "text_level": 1, + "bbox": [ + 524, + 94, + 681, + 108 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input: $K$ , total number of participating clients; $T$ , total number of communication rounds; $\\eta_{\\omega}$ , learning rate of heterogeneous client models; $\\eta_{A}$ , learning rate of the homogeneous representation angle square matrix.", + "bbox": [ + 531, + 112, + 879, + 186 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Randomly initialize the global representation angle square matrix $\\mathcal{A}^0$ and heterogeneous client models $[f_0(\\omega_0^0),\\dots ,f_k(\\omega_k^0),\\dots ,f_{K - 1}(\\omega_{K - 1}^0)]$", + "bbox": [ + 534, + 186, + 879, + 233 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for $t = 1$ to $T$ do", + "bbox": [ + 535, + 234, + 651, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "/* FL Server: */", + "bbox": [ + 560, + 247, + 674, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Broadcast $\\mathcal{A}^{t - 1}$ to participating clients;", + "bbox": [ + 560, + 260, + 823, + 276 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "/* Each FL Client $k$ : */", + "bbox": [ + 560, + 287, + 722, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "// Representation Angle Matrix Learning", + "bbox": [ + 560, + 304, + 828, + 318 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for each $(x_{i},y_{i})\\in D_{k}$ do", + "bbox": [ + 560, + 319, + 733, + 333 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {R} _ {i} = f _ {k} ^ {e x} \\left(\\boldsymbol {x} _ {i}; \\omega_ {k} ^ {e x, t - 1}\\right); \\\\ \\tilde {\\mathcal {R}} _ {i} = \\mathcal {R} _ {i} \\times \\boldsymbol {A} ^ {t - 1}; \\\\ \\hat {y} _ {i} = f ^ {h d} \\left(\\mathcal {R} _ {i} + \\tilde {\\mathcal {R}} _ {i}; \\omega_ {k} ^ {h d, t - 1}\\right); \\\\ f _ {k} \\left(\\omega_ {k} ^ {t}\\right) \\leftarrow f _ {k} \\left(\\omega_ {k} ^ {t - 1}\\right) - \\eta_ {\\omega} \\nabla \\ell \\left(\\dot {y} _ {i}, y _ {i}\\right); \\\\ \\mathcal {A} _ {k} ^ {t} \\leftarrow \\mathcal {A} ^ {t - 1} - \\eta_ {\\mathcal {A}} \\nabla \\ell (\\hat {y} _ {i}, y _ {i}); \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 334, + 830, + 414 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "end", + "text_level": 1, + "bbox": [ + 560, + 415, + 588, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "// Adaptive Diagonal Sparsification", + "bbox": [ + 560, + 429, + 792, + 443 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Obtain $\\tilde{\\mathcal{A}}_k^t$ using Eq. (6);", + "bbox": [ + 560, + 445, + 727, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Upload $\\tilde{\\mathcal{A}}_k^t$ to the FL server;", + "bbox": [ + 560, + 464, + 750, + 481 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "/* FL Server: */", + "bbox": [ + 560, + 492, + 673, + 506 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "// Element-wise Weighted Aggregation", + "bbox": [ + 560, + 508, + 816, + 522 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Generate $\\mathcal{A}^t$ by aggregation using Eq. (7);", + "bbox": [ + 560, + 523, + 841, + 539 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "1", + "text_level": 1, + "bbox": [ + 560, + 539, + 568, + 550 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Return local heterogeneous client models", + "bbox": [ + 535, + 553, + 812, + 566 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\{f _ {0} \\left(\\omega_ {0} ^ {T - 1}\\right), \\dots , f _ {k} \\left(\\omega_ {k} ^ {T - 1}\\right), \\dots , f _ {K - 1} \\left(\\omega_ {K - 1} ^ {T - 1}\\right) \\}.\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 566, + 859, + 584 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 513, + 618, + 689, + 635 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first introduce datasets, base models, comparison baselines, evaluation metrics and hyperparameter settings.", + "bbox": [ + 511, + 642, + 903, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We choose 4 image classification benchmark datasets commonly used in FL investigations.", + "bbox": [ + 511, + 672, + 903, + 702 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Fashion-MNIST² [54] contains 60,000 training and 10,000 testing grayscale $28 \\times 28$ 10-class images.", + "- CIFAR-10 [26] contains 50,000 training and 10,000 testing colourful $32 \\times 32$ 10-class images.", + "- CIFAR-100 $^4$ [26] contains 50,000 training and 10,000 testing colourful $32 \\times 32$ 100-class images.", + "- Tiny-ImageNet [10] contains 100,000 training and 10,000 testing colourful $64 \\times 64$ 200-class images." + ], + "bbox": [ + 513, + 704, + 903, + 825 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "", + "bbox": [ + 107, + 886, + 385, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "2https://github.com/zalandoresearch/fashion-mnist", + "bbox": [ + 513, + 837, + 903, + 861 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "3https://www.cs.toronto.edu/%7Ekriz/cifar.html", + "bbox": [ + 532, + 862, + 890, + 875 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "4https://www.cs.toronto.edu/%7Ekriz/cifar.html", + "bbox": [ + 532, + 875, + 890, + 886 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "$^{5}$ https://tiny-imagenet.herokuapp.com/", + "bbox": [ + 532, + 887, + 823, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1318", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/0df7b092c84881b263cb7ccff163f0c429f248c4be4944ad14547cce2f7fb834.jpg", + "table_caption": [ + "Table 1. Heterogeneous CNNs for Fashion-MNIST." + ], + "table_footnote": [ + "Note: kernel size: $5 \\times 5$ , with 20 filters for convolution layers." + ], + "table_body": "
LayerCNN-1CNN-2CNN-3CNN-4CNN-5
Conv15×5, 205×5, 205×5, 205×5, 205×5, 20
Maxpool12×22×22×22×22×2
Conv25×5, 205×5, 205×5, 205×5, 205×5, 20
Maxpool22×22×22×22×22×2
FC130020015010050
FC25050505050
FC31010101010
model size0.47 MB0.33 MB0.26 MB0.19 MB0.12 MB
", + "bbox": [ + 91, + 114, + 480, + 227 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f6a3f14249439cac1a22e2ef03bbd4200cc4007bdd3d6414e1db78fd60a8f2b2.jpg", + "table_caption": [ + "Table 2. Heterogeneous CNNs for CIFAR-10 and CIFAR-100." + ], + "table_footnote": [ + "Note: kernel size: $5 \\times 5$ , with 16 or 32 filters for convolution layers." + ], + "table_body": "
LayerCNN-1CNN-2CNN-3CNN-4CNN-5
Conv15×5, 165×5, 165×5, 165×5, 165×5, 16
Maxpool12×22×22×22×22×2
Conv25×5, 325×5, 165×5, 325×5, 325×5, 32
Maxpool22×22×22×22×22×2
FC1200020001000800500
FC2500500500500500
FC310/10010/10010/10010/10010/100
model size10.01 MB6.93 MB5.04 MB4.05 MB2.56 MB
", + "bbox": [ + 91, + 277, + 480, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Then, we partition them into non-IID modes in 2 ways:", + "bbox": [ + 109, + 425, + 475, + 440 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Pathological. Following [48], we allocate $\\{2, 2, 10, 20\\}$ classes from $\\{10, 10, 100, 200\\}$ classes to each FL client.", + "- Practical. Following [43], we allocate $\\{10, 10, 100, 200\\}$ classes to each FL client and use a Dirichlet distribution function with a hyperparameter $\\alpha = 0.4$ to generate diverse ratios of one class for different clients." + ], + "bbox": [ + 89, + 440, + 480, + 530 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Base Models. Following [62], we construct 5 heterogeneous CNN models $\\{\\mathrm{CNN - 1,}\\dots,\\mathrm{CNN - 5}\\}$ with different channels of convolutional layers or dimensions of linear layers (Tables 1 and 2) and allocate them to 100 clients. The client $k$ is assigned CNN- $(k\\% 5)$ . Note that the representation $\\mathcal{R}$ dimension $r$ (the dimension of the penultimate linear layer) is $\\{50,500,500\\}$ for Fashion-MNIST, CIFAR-10 and CIFAR-100. Thus, the dimensions of the homogeneous representation angle square matrix $\\mathcal{A}$ are $\\{50\\times 50,500\\times 500,500\\times 500\\}$ . For Tiny-ImageNet, we evenly allocate ResNet- $\\{18,34,50,101,152\\}$ to 100 clients.", + "bbox": [ + 89, + 531, + 482, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison Baselines. We compare FedRAL with 6 baselines, which are the latest representative MHFL works that support completely heterogeneous local models.", + "bbox": [ + 89, + 696, + 482, + 743 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Standalone: each client solely trains its local heterogeneous model, without FL.", + "- Model Decoupling: LG-FedAvg [30].", + "- Mutual Learning: FedKD [53], FedAPEN [43].", + "- Knowledge Distillation without Public Dataset: FedProto [51], FedTGP [72]." + ], + "bbox": [ + 89, + 746, + 482, + 837 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "Evaluation Metrics. We evaluate the performances of FedRAL and baselines with the following three metrics:", + "- Model Accuracy. We record each client's model accuracy (\\%) and calculate their average in each communica" + ], + "bbox": [ + 89, + 839, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tion round. We report the highest average accuracy.", + "bbox": [ + 526, + 90, + 867, + 106 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Communication Cost. We track the average number of transmitted parameters between clients and the server in one communication round, and we record the total rounds consumed to reach the specified target accuracy. The product of the two is the total communication cost.", + "- Computational Overhead. We track the average computational FLOPs of clients in one communication round, and we record the total communication rounds consumed to reach the specified target accuracy. The product of the two is the total computational overhead." + ], + "bbox": [ + 511, + 106, + 903, + 256 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Hyperparameters. We use grid-search to choose the optimal hyperparameters for all algorithms. For general FL hyperparameters, we set the local training epoch $E$ as $\\{1,10\\}$ and batch size $B$ as $\\{32,64,128,256,512\\}$ , and we choose the SGD gradient optimizer with a learning rate of 0.01. To ensure the sufficient convergence of algorithms, we set the total communication rounds $T$ as 500. For the unique hyperparameter - the number $m$ of diagonal blocks for sparsifying the homogeneous representation angle square matrix in FedRAL, we set it to be the factor of the representation dimension $r$ , e.g., $m = \\{1,2,\\dots,50\\}$ for Fashion-MNIST, and $m = \\{1,2,\\dots,500\\}$ for CIFAR.", + "bbox": [ + 511, + 258, + 903, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Comparison Results Analysis", + "text_level": 1, + "bbox": [ + 511, + 452, + 774, + 468 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Average Accuracy. Table 3 shows that FedRAL outperforms all baselines in terms of average test accuracy, demonstrating its effectiveness in enhancing local model generalization through the shared homogeneous representation angle square matrix. As a recent MHFL method, FedTGP achieves the second-best performance. We notice that FedAPEN fails to converge in some cases. This can be attributed to the fact that it uses the same coefficients to combine the outputs of the heterogeneous and homogeneous models for all data samples, failing to balance personalization and generalization under varying data features.", + "bbox": [ + 511, + 474, + 903, + 641 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 5 presents how the average test accuracies of FedRAL and the state-of-the-art FedTGP baseline vary as communication rounds. We see that FedRAL converges to a higher accuracy with a faster speed than FedTGP, especially under more non-IID practical FL settings. This demonstrates the effectiveness of FedRAL in alleviating data heterogeneity while supporting the collaboration of heterogeneous client models.", + "bbox": [ + 511, + 643, + 905, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Communication Cost. Table 4 presents the communication costs of the state-of-the-art FedTGP baseline and FedRAL. We can observe that FedRAL consumes a lower single-round communication cost than FedTGP (column-3), since the sparse representation angle square matrix communicated in FedRAL has fewer parameters than the label-wise average representations transmitted in FedTGP. FedRAL also requires fewer communication rounds to reach the specified target accuracy (column-5), since the de", + "bbox": [ + 511, + 763, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1319", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/215cb1c221319255a623ead291d2ac51c5e6f3f117ce9e3ab10584c39512bb90.jpg", + "table_caption": [ + "Table 3. Comparison of average test accuracy (%) on 4 datasets under pathological and practical data heterogeneity settings." + ], + "table_footnote": [ + "Note: “-” denotes failure to converge. Bold indicates the highest accuracy, and underline indicates the second highest accuracy." + ], + "table_body": "
MethodPathologicalPractical
Fashion-MNISTCIFAR-10CIFAR-100Tiny-ImageNetFashion-MNISTCIFAR-10CIFAR-100Tiny-ImageNet
Standalone99.03±0.5391.97±0.3753.04±0.2433.15±0.0273.39±0.8940.72±0.839.56±0.347.35±0.21
LG-FedAvg [30]98.87±0.4091.27±0.2545.83±0.6735.79±0.1971.97±0.2540.44±0.579.80±0.397.51±0.34
FedKD [53]58.37±0.0573.21±0.5637.21±0.9834.53±0.1645.57±0.4928.14±0.298.29±0.286.29±0.23
FedAPEN [43]44.20±0.58-------
FedProto [51]99.02±0.5092.49±0.2653.67±0.7936.24±0.2774.06±0.6340.20±0.119.27±0.717.63±0.29
FedTGP [72]99.09±0.6592.50±0.3153.20±0.7836.92±0.1073.64±0.6040.95±0.099.37±0.137.65±0.19
FedRAL99.54±0.1693.60±0.1958.70±0.0538.75±0.1378.58±0.1242.97±0.1011.70±0.029.47±0.13
", + "bbox": [ + 91, + 114, + 903, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/11085d94379bc0b80b3e79100512dfe7eaef036e3ab42fb8180a3fefeb8d6800.jpg", + "image_caption": [ + "Communication Round" + ], + "image_footnote": [], + "bbox": [ + 91, + 262, + 272, + 373 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e8f965e28d3aebdf420070565b353f1f4c7b1ded62452b2bfcf186ed88a63763.jpg", + "image_caption": [ + "Communication Round", + "Figure 5. Average test accuracy $(\\%)$ vs. number of rounds." + ], + "image_footnote": [], + "bbox": [ + 93, + 386, + 271, + 496 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9608db99494f565ae83e4cee0c50a3f661f0931e1cd8d52f833fc3d11a4530be.jpg", + "image_caption": [ + "Pathological, CIFAR-100", + "Communication Round" + ], + "image_footnote": [], + "bbox": [ + 287, + 273, + 464, + 373 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4d3466a93a94b8842746b7c9c8d6c881807eea9a38efb853a3a5e644a0753635.jpg", + "image_caption": [ + "Communication Round" + ], + "image_footnote": [], + "bbox": [ + 292, + 386, + 464, + 496 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "signed representation angle learning facilitates model convergence. Overall, FedRAL consumes lower communication costs for reaching the target accuracy than FedTGP (column-6), demonstrating its efficient communication.", + "bbox": [ + 89, + 564, + 480, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Computational Cost. Table 4 also shows the computational costs of the state-of-the-art FedTGP baseline and FedRAL. We can see that FedRAL consumes a lower single-round computational cost than FedTGP (column-4), since additionally training a homogeneous representation angle square matrix in FedRAL consumes lower FLOPs than additionally calculating label-wise average representations in FedTGP, except local model training in both methods. Since FedRAL requires fewer communication rounds to reach the specified target accuracy (column-5), its total computational costs are lower than FedTGP when reaching the specified target accuracy (column-7), substantiating its efficient computation.", + "bbox": [ + 88, + 627, + 482, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Privacy Evaluation. We conduct experiments to evaluate the privacy of FedRAL and baselines. Followed by Zhang et al. [70], considering a semi-honest FL scenario where the server is untrustworthy and may inversely infer the original data of comprised clients from transmit-", + "bbox": [ + 89, + 824, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3ed37b9d9b117ec306673ab52863f411f1f6ca4a72d2b060c3e6e7dd04bf91d8.jpg", + "table_caption": [ + "Table 4. Communication and computational costs." + ], + "table_footnote": [ + "Note: \"Comm/C/R\": a single client's communication cost (KB) in a communication round. \"Comp/C/R\": a single client's computational cost (MB) in a communication round. \"Rounds\": the communication rounds required to reach the specified $\\{90, 50, 40, 9\\}\\%$ target accuracy (the near maximum accuracy that both the state-of-the-art FedTGP baseline and FedRAL can achieve). \"Comm/C\": a single client's communication cost (MB) required to reach the target accuracy. \"Comm/C\": a single client's computational cost (GB) required to reach the target accuracy." + ], + "table_body": "
FL SettingMethodComm/C/R (KB)Comp/C/R (MB)RoundsComm/C (MB)Comp/C (GB)
Pathological CIFAR-10FedTGP3.9154.523021.1516.08
FedRAL1.5628.232010.315.54
Pathological CIFAR-100FedTGP19.5354.523546.7518.85
FedRAL9.7728.231441.373.97
Practical CIFAR-10FedTGP19.5354.873486.6418.65
FedRAL7.8128.581030.792.87
Practical CIFAR-100FedTGP195.3154.8736369.2419.45
FedRAL39.0628.581465.574.07
", + "bbox": [ + 514, + 285, + 903, + 388 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1a9cbe095d75a5204f3203c00a493637355353f3d78fdb06c385243654674579.jpg", + "table_caption": [ + "Table 5. PSNR under DLG attacks on CIFAR-100." + ], + "table_footnote": [], + "table_body": "
MethodLG-FedAvgFedKDFedAPENFedProtoFedTGPFedRAL
PSNR (dB)7.836.956.936.766.696.28
", + "bbox": [ + 514, + 523, + 903, + 549 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f7fba3cc95541f910c52ef592f7899e61c28bfd71a91b4543468111f2145c5f3.jpg", + "table_caption": [ + "Table 6. Average Accuracy $(\\%)$ under high model heterogeneity. \" -\" denotes failure to converge." + ], + "table_footnote": [], + "table_body": "
MethodLG-FedAvgFedKDFedAPENFedProtoFedTGPFedRAL
Accuracy (%)45.6345.95-48.5648.6950.79
", + "bbox": [ + 514, + 603, + 903, + 627 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ted information by launching Deep Leakage from Gradients (DLG) attack. We use the Peak Signal-to-Noise Ratio (PSNR), a common privacy measurement metric, to evaluate the privacy-protecting level. A lower PSNR denotes better privacy protection. The results in Table 5 validate that FedRAL can protect data privacy effectively.", + "bbox": [ + 511, + 652, + 905, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "High Model Heterogeneity. To evaluate how FedRAL and baselines perform under high model heterogeneity, we conduct experiments CIFAR-100 with 100 clients and heterogeneous models including ResNet-{4, 6, 8, 10, 18, 34, 50, 101, 152}, CNN, and ViT. The results in Table 6 validate FedRAL's robustness to high model heterogeneity.", + "bbox": [ + 511, + 744, + 905, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Robustness Analysis", + "text_level": 1, + "bbox": [ + 511, + 847, + 705, + 863 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This section analyzes the robustness of FedRAL to different non-IID degrees of two types.", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1320", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d12d1e619a820b815069035e927909bd39e9706d01da73fc3d9eb4e6fa38964f.jpg", + "image_caption": [ + "Figure 6. Robustness to pathological non-IIDness." + ], + "image_footnote": [], + "bbox": [ + 91, + 89, + 272, + 213 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e4b75de0008a8b0a5c0d32b7519f1a3ffbac953952a0c155c446f408b28c2445.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 89, + 467, + 214 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fe088149e870e34a087efc4e4cac6ecc7faba08c13b9369c4041a4189fbbc257.jpg", + "image_caption": [ + "Figure 7. Robustness to practical non-IIDness." + ], + "image_footnote": [], + "bbox": [ + 91, + 252, + 272, + 377 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/27800c7e49c2f1d56639e1c37049fff8ed88eacca6a25895456c0d4994c568ee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 252, + 467, + 377 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Robustness to pathological non-IIDness. We change the number of classes allocated to each client to obtain different non-IID degrees. For CIFAR-10, we allocate $\\{2,4,6,8,10\\}$ classes from 10 classes to each client. For CIFAR-100, we allocate $\\{10,30,50,70,90,100\\}$ classes out of 100 classes to each client. Figure 6 presents that FedRAL always performs higher accuracies than the state-of-the-art FedTGP baseline. As the number of classes assigned to one client increases, i.e., the non-IID degree decreases, model accuracy tends to drop since more IID data benefits generalization while compromising personalization of local models.", + "bbox": [ + 89, + 429, + 483, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Robustness to practical non-IIDness. We vary the hyperparameter $\\alpha$ of the Dirichlet distribution function to construct diverse non-IID degrees. For both CIFAR-10 and CIFAR-100, we set $\\alpha = \\{0.1, 0.2, 0.3, 0.4, 0.5\\}$ to evaluate the average model accuracies of FedRAL and the state-of-the-art FedTGP baseline. Figure 7 shows that FedRAL performs significantly higher average model accuracy than FedTGP, again demonstrating the robustness of FedRAL to non-IIDness. Similarly, as the value of $\\alpha$ increases, i.e., the non-IID degree decreases, model accuracy tends to degrade also due to the above reason.", + "bbox": [ + 89, + 611, + 483, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Ablation Analysis", + "text_level": 1, + "bbox": [ + 89, + 787, + 264, + 803 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "FedRAL involves three contributions: 1) introducing representation angle matrix learning to share cross-client representation angle knowledge for enhancing model generalization, 2) devising adaptive diagonal sparsification to improve communication efficiency while keeping important representation angle knowledge, 3) designing ordinate-wise", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b7e40be629238630afd77a67614cbb7cf48a9c49b8b64d5d9d1b2792e3c420f0.jpg", + "table_caption": [ + "Table 7. Average accuracy $(\\%)$ in ablation experiments." + ], + "table_footnote": [], + "table_body": "
CaseDiagonal SparsificationWeighted AggregationPathologicalPractical
CIFAR-10CIFAR-100CIFAR-10CIFAR-100
A✓✓93.6058.7042.9711.70
B✗✓91.6456.5841.3711.58
C✓✗93.4753.9440.529.24
", + "bbox": [ + 514, + 114, + 905, + 165 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "weighted aggregation to fuse cross-client representation angle square matrices. Contribution-1) is essential to implement FL collaboration across clients with heterogeneous models, so we conduct experiments to evaluate the necessity of the other two contributions with three cases:", + "bbox": [ + 511, + 198, + 903, + 273 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Case-A, FedRAL with both adaptive diagonal sparsification and ordinate-wise weighted aggregation, i.e., the designed method.", + "- Case-B, FedRAL without adaptive diagonal sparsification but with ordinate-wise weighted aggregation, i.e., clients upload complete representation angle square matrices to the server.", + "- Case-C, FedRAL with adaptive diagonal sparsification but with naive ordinate-wise averaging aggregation, i.e., the aggregation rule (a) displayed in Figure 4, ignoring client data distribution distinctions." + ], + "bbox": [ + 511, + 277, + 903, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7 shows that removing anyone degrades accuracy. Especially, replacing ordinate-wise weighted aggregation with naive averaging aggregation presents obvious accuracy degradations under most settings. These results reflect the effectiveness and necessity of the proposed two designs.", + "bbox": [ + 511, + 446, + 905, + 522 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 546, + 633, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work proposed FedRAL with three innovative designs: representation angle learning, adaptive diagonal sparsification, and element-wise weighted aggregation. The three designs enable enabling it to be a privacy-preserved, well-performed, communication and computation-efficient model-heterogeneous FL approach. Extensive experiments demonstrate its advantages in accuracy, communication and computation efficiency.", + "bbox": [ + 511, + 575, + 905, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 511, + 720, + 665, + 738 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Xiaoguang Liu is supported by the National Science Foundation of China under Grant 62272252 & 62272253, and the Fundamental Research Funds for the Central Universities. Han Yu is supported by the Ministry of Education, Singapore, under its Academic Research Fund Tier 1 (RG101/24); the RIE2025 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) (Award I2301E0026), administered by A\\*STAR, as well as supported by Alibaba Group and NTU Singapore through Alibaba-NTU Global e-Sustainability CorpLab (ANGEL).", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1321", + "bbox": [ + 483, + 945, + 513, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jin-Hyun Ahn et al. Wireless federated distillation for distributed edge learning with heterogeneous data. In Proc. PIMRC, pages 1-6, Istanbul, Turkey, 2019. IEEE. 2", + "[2] Jin-Hyun Ahn et al. Cooperative learning VIA federated distillation OVER fading channels. In Proc. ICASSP, pages 8856-8860, Barcelona, Spain, 2020. IEEE. 2", + "[3] Samiul Alam et al. Fedrolex: Model-heterogeneous federated learning with rolling sub-model extraction. In Proc. NeurIPS, virtual, 2022.. 2", + "[4] Sara Babakniya et al. Revisiting sparsity hunting in federated learning: Why does sparsity consensus matter? Transactions on Machine Learning Research, 1(1):1, 2023.", + "[5] Yun-Hin Chan, Rui Zhou, Running Zhao, Zhihan JIANG, and Edith C. H. Ngai. Internal cross-layer gradients for extending homogeneity to heterogeneity in federated learning. In Proc. ICLR, page 1, Vienna, Austria, 2024. OpenReview.net. 2", + "[6] Hongyan Chang et al. Cronus: Robust and heterogeneous collaborative learning with black-box knowledge transfer. In Proc. NeurIPS Workshop, virtual, 2021.. 2", + "[7] Jiangui Chen et al. Fedmatch: Federated learning over heterogeneous question answering data. In Proc. CIKM, pages 181-190, virtual, 2021. ACM. 2", + "[8] Sijie Cheng et al. Fedgems: Federated learning of larger server models via selective knowledge fusion. CoRR, abs/2110.11027, 2021. 2", + "[9] Yae Jee Cho et al. Heterogeneous ensemble knowledge transfer for training large models in federated learning. In Proc. IJCAI, pages 2881-2887, virtual, 2022. ijcai.org. 2", + "[10] Patryk Chrabaszcz, Ilya Loshchilov, and Frank Hutter. A downsampled variant of imagenet as an alternative to the CIFAR datasets. CoRR, abs/1707.08819, 2017. 5", + "[11] Liam Collins et al. Exploiting shared representations for personalized federated learning. In Proc. ICML, pages 2089-2099, virtual, 2021. PMLR. 1, 2", + "[12] Enmao Diao. Heterofl: Computation and communication efficient federated learning for heterogeneous clients. In Proc. ICLR, page 1, Virtual Event, Austria, 2021. OpenReview.net. 2", + "[13] Tao Fan, Hanlin Gu, et al. Ten challenging problems in federated foundation models. IEEE Transactions on Knowledge and Data Engineering, 2025. 1", + "[14] Randy Goebel, Han Yu, Boi Faltings, Lixin Fan, and Zehui Xiong. Trustworthy Federated Learning. Springer, Cham, 2023. 1", + "[15] Xuan Gong et al. Federated learning via input-output collaborative distillation. In Proc. AAAI, pages 22058-22066, Vancouver, Canada, 2024. AAAI Press. 2", + "[16] Chaoyang He et al. Group knowledge transfer: Federated learning of large cnns at the edge. In Proc. NeurIPS, virtual, 2020. . 2", + "[17] S. Horváth. FjORD: Fair and accurate federated learning under heterogeneous targets with ordered dropout. In Proc. NIPS, pages 12876-12889, Virtual, 2021. OpenReview.net. 2" + ], + "bbox": [ + 93, + 114, + 482, + 897 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Edward J. Hu et al. Lora: Low-rank adaptation of large language models. In ICLR, page 1, Virtual, 2022. OpenReview.net. 2", + "[19] Wenke Huang et al. Learn from others and be yourself in heterogeneous federated learning. In Proc. CVPR, pages 10133-10143, virtual, 2022. IEEE. 2", + "[20] Wenke Huang et al. Few-shot model agnostic federated learning. In Proc. MM, pages 7309-7316, Lisboa, Portugal, 2022. ACM.", + "[21] Sohei Itahara et al. Distillation-based semi-supervised federated learning for communication-efficient collaborative training with non-iid private data. IEEE Trans. Mob. Comput., 22(1):191-205, 2023. 2", + "[22] Jaehee Jang et al. Fedclassavg: Local representation learning for personalized federated learning on heterogeneous neural networks. In Proc. ICPP, pages 76:1-76:10, virtual, 2022. ACM, 2", + "[23] Eunjoeong Jeong et al. Communication-efficient on-device machine learning: Federated distillation and augmentation under non-iid private data. In Proc. NeurIPS Workshop, virtual, 2018. . 2", + "[24] Peter Kairouz et al. Advances and open problems in federated learning. Foundations and Trends in Machine Learning, 14(1-2):1-210, 2021. 1", + "[25] Shivam Kalra et al. Decentralized federated learning through proxy model sharing. Nature communications, 14(1):2899, 2023. 2", + "[26] Alex Krizhevsky et al. Learning multiple layers of features from tiny images. Toronto, ON, Canada, 2009. 5", + "[27] Meng Lei, Qi Zhuang, Wu Lei, Du Xiaoyu, Li Zhaochuan, Cui Lizhen, and Meng Xiangxu. Improving global generalization and local personalization for federated learning. IEEE Transactions on Neural Networks and Learning Systems, 36, 2024. 1", + "[28] Daliang Li and Junpu Wang. Fedmd: Heterogenous federated learning via model distillation. In Proc. NeurIPS Workshop, virtual, 2019. . 1, 2", + "[29] Qinbin Li et al. Practical one-shot federated learning for cross-silo setting. In Proc. IJCAI, pages 1484-1490, virtual, 2021. ijcai.org. 2", + "[30] Paul Pu Liang et al. Think locally, act globally: Federated learning with local and global representations. arXiv preprint arXiv:2001.01523, 1(1), 2020. 1, 2, 6, 7", + "[31] Tao Lin et al. Ensemble distillation for robust model fusion in federated learning. In Proc. NeurIPS, virtual, 2020. . 1, 2", + "[32] Chang Liu et al. Completely heterogeneous federated learning. CoRR, abs/2210.15865, 2022. 2", + "[33] Qianyu Long, Christos Anagnostopoulos, Shameem Puthiya Parambath, and Daning Bi. Feddip: Federated learning with extreme dynamic pruning and incremental regularization. In Proc.ICDM, pages 1187-1192, Shanghai, China, 2023. IEEE. 2", + "[34] Disha Makhija et al. Architecture agnostic federated learning for neural networks. In Proc. ICML, pages 14860-14870, virtual, 2022. PMLR. 2", + "[35] Koji Matsuda et al. Fedme: Federated learning via model exchange. In Proc. SDM, pages 459-467, Alexandria, VA, USA, 2022. SIAM. 2" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1322", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Brendan McMahan et al. Communication-efficient learning of deep networks from decentralized data. In Proc. AISTATS, pages 1273-1282, Fort Lauderdale, FL, USA, 2017. PMLR. 1", + "[37] Duy Phuong Nguyen et al. Enhancing heterogeneous federated learning with knowledge extraction and multi-model fusion. In Proc. SC Workshop, pages 36-43, Denver, CO, USA, 2023. ACM. 2", + "[38] Jaehoon Oh et al. Fedbabu: Toward enhanced representation for federated image classification. In Proc. ICLR, virtual, 2022. OpenReview.net. 2", + "[39] Zhengxiang Pan, Han Yu, Chunyan Miao, and Cyril Leung. Efficient collaborative crowdsourcing. In The 30th AAAI Conference on Artificial Intelligence (AAAI-16), pages 4248–4249, 2016. 1", + "[40] Sejun Park et al. Towards understanding ensemble distillation in federated learning. In Proc. ICML, pages 27132-27187, Honolulu, Hawaii, USA, 2023. PMLR. 2", + "[41] Krishna Pillutla et al. Federated learning with partial model personalization. In Proc. ICML, pages 17716-17758, virtual, 2022. PMLR. 2", + "[42] Zhuang Qi, Sijin Zhou, Lei Meng, Han Yu, and Xiangxu Meng. Federated deconfounding and debiasing learning for out-of-distribution generalization. In The 34th International Joint Conference on Artificial Intelligence, (IJ-CAI'25), pages 1–9, 2025. 1", + "[43] Zhen Qin et al. Fedapen: Personalized cross-silo federated learning with adaptability to statistical heterogeneity. In Proc. KDD, pages 1954-1964, Long Beach, CA, USA, 2023. ACM. 2, 6, 7", + "[44] Zeju Qiu et al. Controlling text-to-image diffusion by orthogonal finetuning. In Proc. NeurIPS, New Orleans, LA, USA, 2023. 2, 3", + "[45] Chao Ren, Han Yu, et al. Advances and open challenges in federated foundation models. IEEE Communications Surveys and Tutorials, 2025. 1", + "[46] Felix Sattler et al. Fedaux: Leveraging unlabeled auxiliary data in federated learning. IEEE Trans. Neural Networks Learn. Syst., 1(1):1-13, 2021. 2", + "[47] Felix Sattler et al. CFD: communication-efficient federated distillation via soft-label quantization and delta coding. IEEE Trans. Netw. Sci. Eng., 9(4):2025-2038, 2022. 2", + "[48] Aviv Shamsian et al. Personalized federated learning using hypernetworks. In Proc. ICML, pages 9489-9502, virtual, 2021. PMLR. 6", + "[49] Tao Shen et al. Federated mutual learning. CoRR, abs/2006.16765, 2020. 2", + "[50] Alysa Ziying Tan et al. Towards personalized federated learning. IEEE Trans. Neural Networks Learn. Syst., 1(1): 1-17, 2022. 1", + "[51] Yue Tan et al. Fedproto: Federated prototype learning across heterogeneous clients. In Proc. AAAI, pages 8432-8440, virtual, 2022. AAAI Press. 2, 6, 7", + "[52] Jiaqi Wang et al. Towards personalized federated learning via heterogeneous model reassembly. In Proc. NeurIPS, page 13, New Orleans, Louisiana, USA, 2023. OpenReview.net. 2" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Chuhan Wu et al. Communication-efficient federated learning via knowledge distillation. Nature Communications, 13 (1):2032, 2022. 2, 6, 7", + "[54] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms, 2017. 5", + "[55] Qiang Yang, Lixin Fan, and Han Yu. Federated Learning: Privacy and Incentive. Springer, Cham, 2020. 1", + "[56] Qiang Yang, Yang Liu, Yong Cheng, Yan Kang, Tianjian Chen, and Han Yu. Federated Learning. Springer, Cham, 2020. 1", + "[57] Liping Yi, Gang Wang, and Xiaoguang Liu. QSFL: A two-level uplink communication optimization framework for federated learning. In Proc. ICML, pages 25501-25513. PMLR, 2022. 1", + "[58] Liping Yi, Xiaorong Shi, Nan Wang, Gang Wang, Xiaoguang Liu, Zhuan Shi, and Han Yu. pfedkt: Personalized federated learning with dual knowledge transfer. Knowledge-Based Systems, 292:111633, 2024. 1", + "[59] Liping Yi, Xiaorong Shi, Nan Wang, Jinsong Zhang, Gang Wang, and Xiaoguang Liu. Fedpe: Adaptive model pruning-expanding for federated learning on mobile devices. IEEE Transactions on Mobile Computing, pages 1-18, 2024. 1", + "[60] Liping Yi, Gang Wang, Xiaofei Wang, and Xiaoguang Liu. Qsfl: Two-level communication-efficient federated learning on mobile edge devices. IEEE Transactions on Services Computing, pages 1-16, 2024. 1", + "[61] Liping Yi, Han Yu, Zhuan Shi, Gang Wang, Xiaoguang Liu, Lizhen Cui, and Xiaoxiao Li. FedSSA: Semantic Similarity-based Aggregation for Efficient Model-Heterogeneous Personalized Federated Learning. In IJCAI, 2024. 1", + "[62] Liping Yi et al. FedGH: Heterogeneous federated learning with generalized global header. In Proc. ACM MM, 2023. 2, 6", + "[63] Liping Yi et al. Federated model heterogeneous matryoshka representation learning. In Proc. NeurIPS, Vancouver, Canada, 2024.", + "[64] Liping Yi et al. pFedAFM: Adaptive Feature Mixture for Data-Level Personalization in Heterogeneous Federated Learning on Mobile Edge Devices. In Proc. ICDE, pages 1981-1994, Hong Kong SAR, China, 2025. IEEE Computer Society.", + "[65] Liping Yi et al. pfedes: Generalized proxy feature extractor sharing for model heterogeneous personalized federated learning. In Proc. AAAI, pages 22146-22154, Philadelphia, PA, USA, 2025. AAAI Press. 1", + "[66] Fuxun Yu et al. Fed2: Feature-aligned federated learning. In Proc. KDD, pages 2066-2074, virtual, 2021. ACM. 2", + "[67] Han Yu, Siyuan Liu, Alex C Kot, Chunyan Miao, and Cyril Leung. Dynamic witness selection for trustworthy distributed cooperative sensing in cognitive radio networks. In Proceedings of the 13th IEEE International Conference on Communication Technology (ICCT'11), pages 1-6, 2011. 1", + "[68] Sixing Yu et al. Resource-aware federated learning using knowledge extraction and multi-model fusion. CoRR, abs/2208.07978, 2022. 2" + ], + "bbox": [ + 516, + 90, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1323", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[69] Jie Zhang et al. Parameterized knowledge transfer for personalized federated learning. In Proc. NeurIPS, pages 10092-10104, virtual, 2021. OpenReview.net. 2", + "[70] Jianqing Zhang et al. Fedcp: Separating feature information for personalized federated learning via conditional policy. In Proc. KDD, page 1, Long Beach, CA, USA, 2023. ACM. 7", + "[71] Jie Zhang et al. Towards data-independent knowledge transfer in model-heterogeneous federated learning. IEEE Trans. Computers, 72(10):2888-2901, 2023. 2", + "[72] Jianqing Zhang et al. Fedtgp: Trainable global prototypes with adaptive-margin-enhanced contrastive learning for data and model heterogeneity in federated learning. In Proc. AAAI, pages 16768-16776, Vancouver, Canada, 2024. 1. 6, 7", + "[73] Lan Zhang et al. Fedzkt: Zero-shot knowledge transfer towards resource-constrained federated learning with heterogeneous on-device models. In Proc. ICDCS, pages 928-938, virtual, 2022. IEEE. 2", + "[74] Zhilu Zhang and Mert R. Sabuncu. Generalized cross entropy loss for training deep neural networks with noisy labels. In Proc. NeurIPS, pages 8792-8802, Montreal, Canada, 2018. Curran Associates Inc. 4", + "[75] Hangyu Zhu et al. Federated learning on non-iid data: A survey. Neurocomputing, 465:371-390, 2021. 1", + "[76] Zhuangdi Zhu et al. Data-free knowledge distillation for heterogeneous federated learning. In Proc. ICML, pages 12878-12889, virtual, 2021. PMLR. 2", + "[77] Zhuangdi Zhu et al. Resilient and communication efficient learning for heterogeneous federated systems. In Proc. ICML, pages 27504-27526, virtual, 2022. PMLR. 2", + "[78] Qi Zhuang, Meng Lei, Chen Zitan, Hu Han, Lin Hui, and Meng Xiangxu. Cross-silo prototypical calibration for federated learning with non-iid data. In Proceedings of the 31st ACM International Conference on Multimedia, pages 3099-3107, 2023. 1", + "[79] Qi Zhuang, Meng Lei, et al. Cross-silo feature space alignment for federated learning on clients with imbalanced data. In The 39th Annual AAAI Conference on Artificial Intelligence (AAAI-25), pages 19986-19994, 2025. 1" + ], + "bbox": [ + 91, + 92, + 480, + 643 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1324", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Federated Representation Angle Learning/f94f188a-83f1-436a-a6d8-36878318e078_model.json b/2025/Federated Representation Angle Learning/f94f188a-83f1-436a-a6d8-36878318e078_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1f15f22c5a280937a235df4f2eaf32d4ef2182d0 --- /dev/null +++ b/2025/Federated Representation Angle Learning/f94f188a-83f1-436a-a6d8-36878318e078_model.json @@ -0,0 +1,3302 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.286, + 0.131, + 0.72, + 0.153 + ], + "angle": 0, + "content": "Federated Representation Angle Learning*" + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.18, + 0.946, + 0.289 + ], + "angle": 0, + "content": "Liping Yi\\(^{1}\\), Han Yu\\(^{2}\\), Gang Wang\\(^{1,\\ast}\\), Xiaoguang Liu\\(^{1,\\ast}\\), Xiaoxiao Li\\(^{3,4}\\) \n\\(^{1}\\)College of Computer Science, TMCC, SysNet, DISSec, GTIISC, Nankai University, Tianjin, China \n\\(^{2}\\)College of Computing and Data Science, Nanyang Technological University, Singapore \n\\(^{3}\\)Department of Electrical and Computer Engineering, University of British Columbia, Vancouver, Canada \n\\(^{4}\\)Vector Institute, Vancouver, Canada \n{yiliping, wgzwp, liuxg}@nbjl.nankai.edu.cn, han.yu@ntu.edu.sg, xiaoxiao.li@ece.ubc.ca" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.322, + 0.327, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.362, + 0.486, + 0.832 + ], + "angle": 0, + "content": "Model-heterogeneous federated learning (MHFL) is a challenging FL paradigm designed to allow FL clients to train structurally heterogeneous models under the coordination of an FL server. Existing MHFL methods face significant limitations when it comes to transferring global knowledge to clients as a result of sharing only partial homogeneous model parameters or calculating distance loss, leading to inferior model generalization. To bridge this gap, we propose a novel model-heterogeneous Federated learning method with Representation Angle Learning (FedRAL). It consists of three innovative designs: (1) We first introduce representation angle learning into MHFL. Specifically, we embed a homogeneous square matrix into the local heterogeneous model of each client, which learns the angle information of local representations. These homogeneous representation angle square matrices are aggregated on the server to fuse representation angle knowledge shared by clients for enhancing the generalization of local representations. (2) As different clients might have heterogeneous system resources, we propose an adaptive diagonal sparsification strategy to reduce the numbers of the parameters of representation angle square matrices uploaded to the server, to improve FL communication efficiency. (3) To enable effective fusion of sparsified homogeneous local representation angle square matrices, we design an element-wise weighted aggregation approach. Experiments on 4 benchmark datasets under 2 types of non-IID divisions over 6 state-of-the-art baselines demonstrate that FedRAL achieves the best performance. It improves test accuracy, communication efficiency and computational efficiency by up to \\(5.03\\%\\), \\(12.43\\times\\) and \\(6.49\\times\\), respectively." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.322, + 0.645, + 0.338 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.351, + 0.907, + 0.486 + ], + "angle": 0, + "content": "Federated learning (FL) [13, 14, 24, 36, 45, 55, 56] is a privacy-preserved distributed machine learning paradigm. Typically, an FL server is often involved in broadcast a global model to FL clients. The clients then further train the global model on local data to obtain local models. The server aggregates local model updates received from the clients to produce a new global model. In this way, client data are not exposed. Under this setting, the models trained by the clients must follow the same structure." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.49, + 0.909, + 0.777 + ], + "angle": 0, + "content": "In practice, clients participating in FL often have non-independent and identically distributed (non-IID) data, a.k.a., data heterogeneity [27, 42, 50, 58, 75, 78, 79], the local models trained on such data are biased, so the global model obtained by aggregating them may not perform well on all client data. Besides, FL clients also have heterogeneous system configurations in terms of computing power and communication bandwidth, a.k.a., system heterogeneity [39, 57, 59, 60, 67]. Thus, requesting all clients to train the same model might not be viable for some weaker clients, while leaving stronger clients under-utilized. More critically, some companies or institutions joining FL as clients might face intellectual property issues regarding their proprietary local models. Thus, they might be reluctant to share their private heterogeneous models, a.k.a., model heterogeneity. These challenges have inspired the field of model-heterogeneous federated learning (MHFL) with preserved data and model privacy, which advocates avoiding directly sharing private local heterogeneous models [61-65]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.911, + 0.902 + ], + "angle": 0, + "content": "Existing MHFL approaches mainly include: a) Modeldecoupling: each client model is decoupled into heterogeneous and homogeneous parts, with only homogeneous parts being shared with the server [11, 30]. b) Knowledge distillation: the server aggregates local representation or logits from different clients to generate global representation or logits, which is used to calculate distillation loss for local training [28, 31]. c) Mutual training: each client is" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.237, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1314" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.243 + ], + "angle": 0, + "content": "assigned a shared homogeneous small model and trains it with the local heterogeneous model by mutual loss between them [49, 53]. Although these methods support model heterogeneity, their capability for transferring global knowledge from the server to clients is limited due to sharing partial model parameters or using distillation loss, thereby constraining the generalization of local models. Whereas, model representations often involve more sufficient semantic knowledge than partial model parameters or distillation loss, which are explored in this work for knowledge sharing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.247, + 0.483, + 0.52 + ], + "angle": 0, + "content": "A recent work [44] empirically verified that parameter fine-tuning approaches (e.g., LoRA [18]) cannot sufficiently retain semantic information of the pre-trained model. Whereas, fine-tuning the angles (i.e., gradient directions) of the pre-trained model can achieve this, thereby enhancing the generalization of the pre-trained model. The work further proposed the Orthogonal Fine-Tuning (OFT) method to fine-tune the parameter angles of the pre-trained models via an orthogonal matrix (Figure 1). In short, parameter angles involve sufficient semantic information that can enhance model generalization. Therefore, we can enhance FL model generalization through sharing parameter angles instead of model parameters. However, the FL server can not aggregate parameter angles from structure-heterogeneous client models. Considering representations also with sufficient semantic knowledge, we attempt to share homogeneous representation angles among clients to achieve generalization-enhanced model-heterogeneous FL." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.524, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To this end, we propose a novel model-heterogeneous Federated learning method with Representation Angel Learning (FedRAL) to enhance the transfer of generalizable global knowledge from the FL server to FL clients. It consists of three innovative designs: (1) Since representations are extracted high-level features from data, with sufficient semantic information, FedRAL embeds a homogeneous square matrix into the local heterogeneous model to learn the angle information of local extracted representations. The server aggregates local homogeneous representation angle square matrices from different clients to generate a global representation angle square matrix, thereby fusing cross-client representation angle knowledge. Then, the global representation angle square matrix with sufficient generalized representation angle information is used to enhance the generalization of local representations during local model training. Different from the strong assumption of the orthogonality of the matrix in OFT, the homogeneous representation angle square matrix is initialized randomly. (2) Considering the diverse system resources (e.g., communication bandwidth) of different clients, we design an adaptive diagonal sparsification strategy for the homogeneous representation angle square matrix that allows clients to upload diagonal blocks with sizes matching with communication capacity. (3) To effectively aggregate sparsified local" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.088, + 0.907, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.584, + 0.168, + 0.835, + 0.182 + ], + "angle": 0, + "content": "Figure 1. Orthogonal Fine-Tuning (OFT)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.192, + 0.905, + 0.223 + ], + "angle": 0, + "content": "homogeneous representation angle square matrices on the server, element-wise weighted aggregation is designed." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.223, + 0.905, + 0.314 + ], + "angle": 0, + "content": "Experiments on 4 benchmark datasets under 2 modes of non-IID divisions against 6 advanced MHFL methods demonstrate the superiority of FedRAL. Compared with the state-of-the-art baseline, FedRAL improves test accuracy by up to \\(5.03\\%\\), while enhancing communication and computational efficiency by up to \\(12.43\\times\\) and \\(6.49\\times\\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.329, + 0.655, + 0.344 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.355, + 0.905, + 0.385 + ], + "angle": 0, + "content": "Based on different approaches used for achieving MHFL, existing methods can be divided into 4 categories." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.386, + 0.906, + 0.506 + ], + "angle": 0, + "content": "Heterogeneous Subnet. Some works [3-5, 12, 17, 33, 66, 77] assume that each client trains local heterogeneous subnets of the global model, and the server aggregates them according to parameter positions to re-construct a new global model. They allow each client to train personalized heterogeneous subnets to tackle data and system heterogeneity. However, uploading subnets exposes model parameters, thereby breaching model IP protection requirements." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.906, + 0.643 + ], + "angle": 0, + "content": "Model Decoupling. These methods [7, 11, 22, 30, 32, 38, 41, 62] share partial local model parameters instead of the entire local model for global aggregation. Specifically, they decouple each client's local heterogeneous model into heterogeneous and homogeneous parts, and only share homogeneous parts for cross-client knowledge exchange. Nevertheless, sharing partial model parameters still violates model IP protection needs. Furthermore, the unshared remaining parameters might overfit due to local training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.906, + 0.84 + ], + "angle": 0, + "content": "Knowledge Distillation. These methods [1, 2, 6, 8, 9, 15, 16, 19-21, 23, 28, 29, 31, 34, 37, 40, 46, 47, 51, 52, 68, 69, 71-73, 76] aggregate the local seen-class representations or logits extracted on local private data or a public dataset to generate the global representation or logits by class, which is used to calculate distillation loss with the local representations or logits. Although they protect local model structures, the distillation loss only transfers limited knowledge from the server to clients. Besides, sharing class information might be prohibited in some data privacy-sensitive applications. The public dataset used must follow similar data distributions with local private data, which is difficult to access due to data privacy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Mutual Learning. These methods [25, 35, 43, 49, 53] add a global homogeneous small model shared by all clients, which is alternately trained with the local heterogeneous model by the mutual loss between the outputs of" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1315" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.168 + ], + "angle": 0, + "content": "the two models for each sample. Although they hide local heterogeneous model structures, the mutual loss only transfers limited knowledge between the two models. The unstable mutual loss in early training rounds might result in both models not converging." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.168, + 0.484, + 0.334 + ], + "angle": 0, + "content": "Existing MHFL methods generally transfer limited generalized knowledge from the server to clients, leading to model performance bottlenecks. FedRAL effectively tackles data, system and model heterogeneity simultaneously, allowing clients to share a global homogeneous representation angle square matrix for learning generalized representation angle information. It bridges this important gap by enhancing the generalization of local representations via utilizing the global homogeneous representation angle square matrix to fine-tune local representation angles, while preserving personalized semantic information." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.35, + 0.232, + 0.367 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.375, + 0.317, + 0.392 + ], + "angle": 0, + "content": "3.1. Orthogonal Fine-Tuning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.398, + 0.484, + 0.699 + ], + "angle": 0, + "content": "Fine-tuning the parameters of a pre-trained model has been demonstrated to compromise its semantic information, while fine-tuning its parameter angles (i.e., gradient directions) can sufficiently retain its original semantic information [44]. This enables the preservation of the generalized ability when adapting to downstream tasks. Hypersphere Energy (HE) measures the uniformity of neuron distribution on the unit hypersphere. A lower HE value is beneficial to keeping pre-trained model generalization ability. To minimize HE variance between the pre-trained model \\( W^0 \\in \\mathbb{R}^{u \\times v} \\) and the angle-fine-tuned model \\( W \\in \\mathbb{R}^{u \\times v} \\), orthogonal finetuning (OFT) [44] fine-tunes the parameter angles of the pre-trained model through an orthogonal matrix \\( O \\in \\mathbb{R}^{u \\times u} \\), \\( W = O \\times W^0 \\), as depicted in Figure 1. To ensure orthogonality, OFT utilizes Cayley parameterization to construct matrix \\( O = (I + Q)(I - Q)^{-1} \\), where \\( Q \\in \\mathbb{R}^{u \\times u} \\) is a skew-symmetric matrix satisfying \\( Q = -Q^T \\). Due to inverse operations, updating the orthogonal matrix \\( O \\) during fine-tuning introduces high computational overhead, especially with a large dimension \\( u \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.7, + 0.484, + 0.807 + ], + "angle": 0, + "content": "Unlike OFT, we explore fine-tuning the angles of representations with sufficient semantic information. We relax the orthogonality assumption and avoid expensive inverse operations. Instead, we randomly initialize the homogeneous representation angle square matrix, which is embedded into the local heterogeneous model. They are trained in an end-to-end manner to improve computation efficiency." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.295, + 0.833 + ], + "angle": 0, + "content": "3.2. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.903 + ], + "angle": 0, + "content": "In FedRAL, a server coordinates \\( K \\) clients with heterogeneous local models for collaborative training. Client \\( k \\)'s local model \\( f_{k}(\\omega_{k}) \\) consists of two parts: 1) a heterogeneous feature extractor \\( f_{k}^{ex}(\\omega_{k}^{ex}) \\), and 2) a homogeneous predic" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.09, + 0.907, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.6, + 0.342, + 0.82, + 0.356 + ], + "angle": 0, + "content": "Figure 2. The workflow of FedRAL." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.383, + 0.907, + 0.566 + ], + "angle": 0, + "content": "tion head \\( f^{hd}(\\omega_k^{hd}) \\) (all clients conduct the same prediction task), \\( f_{k}(\\omega_{k}) = f_{k}^{ex}(\\omega_{k}^{ex})\\circ f^{hd}(\\omega_{k}^{hd}) \\), as depicted in Figure 2. Each client trains a local homogeneous square matrix \\( \\mathcal{A}_k \\) to learn the angle information of local representations extracted by the heterogeneous feature extractor. The server aggregates local homogeneous representation angle square matrices from different clients to generate a global representation angle square matrix \\( \\mathcal{A} \\) for cross-client representation angle knowledge fusion. Therefore, the training objective of FedRAL is to minimize the loss sum of all heterogeneous client models with the help of the global homogeneous representation angle square matrix \\( \\mathcal{A} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.578, + 0.906, + 0.62 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathcal {A}, \\left\\{\\omega_ {0}, \\dots , \\omega_ {K - 1} \\right\\}} \\sum_ {k = 0} ^ {K - 1} \\ell \\left(\\mathcal {F} _ {k} \\left(D _ {k}; \\left(\\omega_ {k} \\circ \\mathcal {A}\\right)\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.632, + 0.906, + 0.678 + ], + "angle": 0, + "content": "where \\(\\ell\\) is loss. \\(\\mathcal{F}_k(\\omega_k\\circ \\mathcal{A})\\) denotes the combination of the local heterogeneous model \\(\\omega_{k}\\) and the global representation angle square matrix \\(\\mathcal{A}\\), \\(D_{k}\\) is the non-IID data of client \\(k\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.693, + 0.819, + 0.71 + ], + "angle": 0, + "content": "4. The Proposed FedRAL Approach" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.719, + 0.906, + 0.809 + ], + "angle": 0, + "content": "FedRAL consists of three innovative designs: (1) homogeneous representation angle square matrix learning, (2) adaptive diagonal sparsification, and (3) element-wise weighted homogeneous square matrix aggregation. The workflow of FedRAL in a single communication round includes the following steps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.905, + 0.841 + ], + "angle": 0, + "content": "Overview. As shown in Figure 2, in the \\(t\\)-th communication round, FedRAL executes 4 steps:" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.841, + 0.905, + 0.871 + ], + "angle": 0, + "content": "① The server broadcasts the latest global representation angle square matrix \\(\\mathcal{A}^{t-1}\\) to \\(K\\) participating clients." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.871, + 0.905, + 0.902 + ], + "angle": 0, + "content": "\\(②\\) A client \\(k\\) embeds the global representation angle square matrix \\(\\mathcal{A}^{t - 1}\\) into its local model \\(f_{k}(\\omega_{k}^{t - 1})\\) to enhance" + }, + { + "type": "list", + "bbox": [ + 0.508, + 0.841, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1316" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.103, + 0.092, + 0.483, + 0.136 + ], + "angle": 0, + "content": "the generalization of local representations, while learning the personalized angle information of local representations via end-to-end training." + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.137, + 0.483, + 0.2 + ], + "angle": 0, + "content": "③ The local homogeneous representation angle square matrix \\(\\mathcal{A}_k^t\\) after local training is sparsified into \\(\\tilde{\\mathcal{A}}_k^t\\) via the proposed adaptive diagonal sparsification strategy before being uploaded to the server." + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.201, + 0.483, + 0.264 + ], + "angle": 0, + "content": "(4) The server aggregates sparsified local homogeneous representation angle square matrices \\(\\{\\tilde{\\mathcal{A}}_0^t,\\dots ,\\tilde{\\mathcal{A}}_{K - 1}^t\\}\\) via the proposed element-wise weighted aggregation to generate a new global representation angle square matrix \\(\\mathcal{A}^t\\)" + }, + { + "type": "list", + "bbox": [ + 0.084, + 0.137, + 0.483, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.265, + 0.483, + 0.324 + ], + "angle": 0, + "content": "Inference. These steps are iteratively executed until heterogeneous client models \\(\\{f_0(\\omega_0),\\dots ,f_{K - 1}(\\omega_{K - 1})\\}\\) converge, which are used for inference. Algorithm 1 describes the detailed workflow of FedRAL." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.334, + 0.37, + 0.35 + ], + "angle": 0, + "content": "4.1. Representation Angle Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.356, + 0.483, + 0.402 + ], + "angle": 0, + "content": "During the \\(t\\)-th training round, a client \\(k\\) uses its local heterogeneous feature extractor \\(f_{k}^{ex}\\left(\\omega_{k}^{ex,t - 1}\\right)\\) to extract the personalized representation \\(\\mathcal{R} \\in \\mathbb{R}^r\\) of the input sample \\(\\pmb{x}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.412, + 0.482, + 0.431 + ], + "angle": 0, + "content": "\\[\n\\mathcal {R} = f _ {k} ^ {e x} \\left(\\boldsymbol {x}; \\omega_ {k} ^ {e x, t - 1}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.441, + 0.484, + 0.502 + ], + "angle": 0, + "content": "Then, the extracted personalized representation \\(\\mathfrak{R} \\in \\mathbb{R}^r\\) is multiplied by the received global representation angle square matrix \\(\\mathcal{A}^{t-1} \\in \\mathbb{R}^{r \\times r}\\) to produce the representation \\(\\tilde{\\mathcal{R}} \\in \\mathbb{R}^r\\) with enhanced generalization:" + }, + { + "type": "equation", + "bbox": [ + 0.229, + 0.513, + 0.482, + 0.53 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathcal {R}} = \\mathcal {R} \\times \\mathcal {A} ^ {t - 1}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.542, + 0.483, + 0.573 + ], + "angle": 0, + "content": "Then, \\(\\mathcal{R} \\in \\mathbb{R}^r\\) and \\(\\hat{\\mathcal{R}} \\in \\mathbb{R}^r\\) are summed and fed into the prediction head to make a prediction:" + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.582, + 0.482, + 0.602 + ], + "angle": 0, + "content": "\\[\n\\hat {y} = f ^ {h d} \\left(\\mathcal {R} + \\tilde {\\mathcal {R}}; \\omega_ {k} ^ {h d, t - 1}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.483, + 0.69 + ], + "angle": 0, + "content": "The loss \\(\\ell\\) (e.g., Cross-Entropy loss [74]) between the prediction \\(\\hat{y}\\) and the ground-truth label \\(y\\) is used to update the heterogeneous client model \\(f_{k}(\\omega_{k}^{t - 1})\\) and the homogeneous representation angle square matrix \\(\\mathcal{A}^{t - 1}\\) simultaneously via end-to-end training:" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.7, + 0.483, + 0.726 + ], + "angle": 0, + "content": "\\[\nf _ {k} \\left(\\omega_ {k} ^ {t}\\right) \\leftarrow f _ {k} \\left(\\omega_ {k} ^ {t - 1}\\right) - \\eta_ {\\omega} \\nabla \\ell (\\hat {y}, y), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.722, + 0.384, + 0.74 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mathcal {A}} _ {k} ^ {t} \\leftarrow \\boldsymbol {\\mathcal {A}} ^ {t - 1} - \\eta_ {\\boldsymbol {\\mathcal {A}}} \\nabla \\ell (\\hat {y}, y),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.809 + ], + "angle": 0, + "content": "where \\(\\eta_{\\omega}\\) and \\(\\eta_{\\mathcal{A}}\\) are the learning rates of the local heterogeneous model and the homogeneous representation angle square matrix. We set \\(\\eta_{\\omega} = \\eta_{\\mathcal{A}}\\) by default as they are updated at the same time." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The shared global homogeneous representation angle square matrix serves three functions during local training: a) it transfers globally generalizable representation angle knowledge from the server to the client by producing the generalization-enhanced representation; b) it learns the personalized representation angle knowledge from the" + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.088, + 0.857, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.531, + 0.32, + 0.887, + 0.335 + ], + "angle": 0, + "content": "Figure 3. An illustrative example of diagonal sparsification." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.35, + 0.905, + 0.427 + ], + "angle": 0, + "content": "extracted personalized representations, thereby transferring local personalized representation angle knowledge to the server; and c) the summed representation includes both generalized and personalized knowledge, thereby improving model expressiveness while addressing non-IID problem." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.438, + 0.804, + 0.455 + ], + "angle": 0, + "content": "4.2. Adaptive Diagonal Sparsification" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.46, + 0.905, + 0.521 + ], + "angle": 0, + "content": "To improve communication efficiency while maintaining model performance, we design an adaptive diagonal sparsification strategy to sparsify the local representation angle square matrices before uploading them to the FL server." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.906, + 0.718 + ], + "angle": 0, + "content": "Each element in \\(\\mathcal{R} \\in \\mathbb{R}^r\\) corresponds to the element on the diagonal of \\(\\mathcal{A} \\in \\mathbb{R}^{r \\times r}\\). If we only retain the diagonal elements of \\(\\mathcal{A}\\) and set other elements to be 0, the angle of each element in the representation is directly adjusted by the corresponding element in the square matrix diagonal, as depicted in Figure 3(a). When all elements of the square matrix are non-zero, non-diagonal elements might disturb representation angle adjustment due to multiplication with non-corresponding representation elements. As Figure 3(b) shows, the enhanced representation processed by the entire representation angle square matrix is significantly different from the one processed by the sparse square matrix with only diagonal elements." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.719, + 0.906, + 0.854 + ], + "angle": 0, + "content": "Since adjacent elements in the representation are semantically correlated, retaining partial elements wrapped with the diagonal in the representation angle square matrix can enhance the angle correlation of adjacent representation elements. Therefore, the proposed sparsification strategy retains the elements from the diagonal blocks of the square matrix and masks other elements with 0, as shown in Figure 3(c). For a \\(\\mathbb{R}^{r\\times r}\\) square matrix, \\(m\\) diagonal blocks are retained. Each diagonal block is \\(\\mathbb{R}\\frac{r}{m}\\times \\frac{r}{m}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Clients with different communication bandwidths can select different numbers of diagonal blocks \\(m_{k}\\) for the local homogeneous representation angle matrix \\(\\mathcal{A}_k^t\\) to produce" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1317" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.089, + 0.22, + 0.099 + ], + "angle": 0, + "content": "(a) Element-wise Averaging:" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.1, + 0.169, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.172, + 0.1, + 0.274, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.1, + 0.379, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.1, + 0.482, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.17, + 0.266, + 0.179 + ], + "angle": 0, + "content": "(b) Element-wise Weighted Averaging:" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.18, + 0.168, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.169, + 0.179, + 0.274, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.179, + 0.307, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.308, + 0.179, + 0.376, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.179, + 0.408, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.179, + 0.482, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.258, + 0.482, + 0.273 + ], + "angle": 0, + "content": "Figure 4. Element-wise averaging (a) without or (b) with weights." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.281, + 0.443, + 0.302 + ], + "angle": 0, + "content": "the sparsified representation angle square matrix \\(\\tilde{\\mathcal{A}}_k^t\\):" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.31, + 0.483, + 0.331 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathcal {A}} _ {k} ^ {t} = \\text {D i a g o n a l - S p a r s i f i c a t i o n} \\left(\\mathcal {A} _ {k} ^ {t}; m _ {k}\\right). \\tag {6}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.339, + 0.406, + 0.355 + ], + "angle": 0, + "content": "4.3. Element-wise Weighted Aggregation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.361, + 0.484, + 0.422 + ], + "angle": 0, + "content": "After receiving the sparsified local homogeneous representation square matrices from \\( K \\) clients in the \\( t \\)-th communication round, the server aggregates them according to the element positions with weighted averaging:" + }, + { + "type": "equation", + "bbox": [ + 0.223, + 0.43, + 0.483, + 0.471 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mathcal {A}} ^ {t} = \\sum_ {k = 0} ^ {K - 1} \\frac {n _ {k}}{n} \\tilde {\\boldsymbol {\\mathcal {A}}} _ {k} ^ {t}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.479, + 0.483, + 0.508 + ], + "angle": 0, + "content": "where \\( n_k \\) is the data volume of the client \\( k \\), and \\( n \\) is the total data volume of \\( K \\) clients." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.51, + 0.484, + 0.751 + ], + "angle": 0, + "content": "This element-wise weighted aggregation rule provides two benefits. a) All clients upload the diagonal elements of local homogeneous representation angle square matrices. Using \\(\\frac{n_k}{n}\\) to weigh the diagonal elements can sufficiently reflect the overall data distribution across all clients, improving the generalization of the newly generated global square matrix. b) As the example given in Figure 4(a), for non-diagonal elements, if we directly average overlapped them by element, when one non-diagonal element is only provided by one client, then it is the value of this non-diagonal element. In this case, these non-overlapped non-diagonal elements may negatively affect the clients who do not provide for these elements. Instead, as shown in Figure 4(b), weighted averaging scales down such elements and then reduce negative effects while improving the generalization of the newly generated global square matrix." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.763, + 0.326, + 0.781 + ], + "angle": 0, + "content": "5. Experimental Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.788, + 0.484, + 0.88 + ], + "angle": 0, + "content": "To evaluate the performance of FedRAL \\( {}^{1} \\) ,we compare it with 6 state-of-the-art existing approaches on 4 widely adopted benchmark datasets under 2 types of data heterogeneity settings. We implement FedRAL and baselines using pytorch and run them on 4 NVIDIA GeForce RTX 3090 GPUs (24GB memory)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.386, + 0.9 + ], + "angle": 0, + "content": "" + }, + { + "type": "title", + "bbox": [ + 0.525, + 0.095, + 0.682, + 0.109 + ], + "angle": 0, + "content": "Algorithm 1: FedRAL" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.113, + 0.88, + 0.187 + ], + "angle": 0, + "content": "Input: \\(K\\), total number of participating clients; \\(T\\), total number of communication rounds; \\(\\eta_{\\omega}\\), learning rate of heterogeneous client models; \\(\\eta_{A}\\), learning rate of the homogeneous representation angle square matrix." + }, + { + "type": "text", + "bbox": [ + 0.535, + 0.188, + 0.88, + 0.234 + ], + "angle": 0, + "content": "Randomly initialize the global representation angle square matrix \\(\\mathcal{A}^0\\) and heterogeneous client models \\([f_0(\\omega_0^0),\\dots ,f_k(\\omega_k^0),\\dots ,f_{K - 1}(\\omega_{K - 1}^0)]\\)" + }, + { + "type": "text", + "bbox": [ + 0.537, + 0.235, + 0.653, + 0.246 + ], + "angle": 0, + "content": "for \\(t = 1\\) to \\(T\\) do" + }, + { + "type": "text", + "bbox": [ + 0.561, + 0.248, + 0.675, + 0.26 + ], + "angle": 0, + "content": "/* FL Server: */" + }, + { + "type": "text", + "bbox": [ + 0.561, + 0.261, + 0.824, + 0.277 + ], + "angle": 0, + "content": "Broadcast \\(\\mathcal{A}^{t - 1}\\) to participating clients;" + }, + { + "type": "text", + "bbox": [ + 0.561, + 0.289, + 0.723, + 0.302 + ], + "angle": 0, + "content": "/* Each FL Client \\(k\\) : */" + }, + { + "type": "text", + "bbox": [ + 0.561, + 0.305, + 0.829, + 0.319 + ], + "angle": 0, + "content": "// Representation Angle Matrix Learning" + }, + { + "type": "text", + "bbox": [ + 0.561, + 0.32, + 0.735, + 0.334 + ], + "angle": 0, + "content": "for each \\((x_{i},y_{i})\\in D_{k}\\) do" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.335, + 0.831, + 0.415 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {R} _ {i} = f _ {k} ^ {e x} \\left(\\boldsymbol {x} _ {i}; \\omega_ {k} ^ {e x, t - 1}\\right); \\\\ \\tilde {\\mathcal {R}} _ {i} = \\mathcal {R} _ {i} \\times \\boldsymbol {A} ^ {t - 1}; \\\\ \\hat {y} _ {i} = f ^ {h d} \\left(\\mathcal {R} _ {i} + \\tilde {\\mathcal {R}} _ {i}; \\omega_ {k} ^ {h d, t - 1}\\right); \\\\ f _ {k} \\left(\\omega_ {k} ^ {t}\\right) \\leftarrow f _ {k} \\left(\\omega_ {k} ^ {t - 1}\\right) - \\eta_ {\\omega} \\nabla \\ell \\left(\\dot {y} _ {i}, y _ {i}\\right); \\\\ \\mathcal {A} _ {k} ^ {t} \\leftarrow \\mathcal {A} ^ {t - 1} - \\eta_ {\\mathcal {A}} \\nabla \\ell (\\hat {y} _ {i}, y _ {i}); \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.561, + 0.416, + 0.589, + 0.427 + ], + "angle": 0, + "content": "end" + }, + { + "type": "text", + "bbox": [ + 0.561, + 0.43, + 0.793, + 0.444 + ], + "angle": 0, + "content": "// Adaptive Diagonal Sparsification" + }, + { + "type": "text", + "bbox": [ + 0.561, + 0.446, + 0.728, + 0.463 + ], + "angle": 0, + "content": "Obtain \\(\\tilde{\\mathcal{A}}_k^t\\) using Eq. (6);" + }, + { + "type": "text", + "bbox": [ + 0.561, + 0.465, + 0.75, + 0.482 + ], + "angle": 0, + "content": "Upload \\(\\tilde{\\mathcal{A}}_k^t\\) to the FL server;" + }, + { + "type": "text", + "bbox": [ + 0.561, + 0.493, + 0.674, + 0.507 + ], + "angle": 0, + "content": "/* FL Server: */" + }, + { + "type": "text", + "bbox": [ + 0.561, + 0.509, + 0.818, + 0.523 + ], + "angle": 0, + "content": "// Element-wise Weighted Aggregation" + }, + { + "type": "text", + "bbox": [ + 0.561, + 0.524, + 0.842, + 0.54 + ], + "angle": 0, + "content": "Generate \\(\\mathcal{A}^t\\) by aggregation using Eq. (7);" + }, + { + "type": "title", + "bbox": [ + 0.561, + 0.54, + 0.569, + 0.551 + ], + "angle": 0, + "content": "1" + }, + { + "type": "text", + "bbox": [ + 0.537, + 0.554, + 0.813, + 0.567 + ], + "angle": 0, + "content": "Return local heterogeneous client models" + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.567, + 0.86, + 0.585 + ], + "angle": 0, + "content": "\\[\n\\{f _ {0} \\left(\\omega_ {0} ^ {T - 1}\\right), \\dots , f _ {k} \\left(\\omega_ {k} ^ {T - 1}\\right), \\dots , f _ {K - 1} \\left(\\omega_ {K - 1} ^ {T - 1}\\right) \\}.\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.619, + 0.69, + 0.636 + ], + "angle": 0, + "content": "5.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.905, + 0.673 + ], + "angle": 0, + "content": "We first introduce datasets, base models, comparison baselines, evaluation metrics and hyperparameter settings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.905, + 0.703 + ], + "angle": 0, + "content": "Datasets. We choose 4 image classification benchmark datasets commonly used in FL investigations." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.705, + 0.905, + 0.735 + ], + "angle": 0, + "content": "- Fashion-MNIST² [54] contains 60,000 training and 10,000 testing grayscale \\(28 \\times 28\\) 10-class images." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.736, + 0.905, + 0.765 + ], + "angle": 0, + "content": "- CIFAR-10 [26] contains 50,000 training and 10,000 testing colourful \\(32 \\times 32\\) 10-class images." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.766, + 0.905, + 0.795 + ], + "angle": 0, + "content": "- CIFAR-100\\(^4\\) [26] contains 50,000 training and 10,000 testing colourful \\(32 \\times 32\\) 100-class images." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.796, + 0.905, + 0.826 + ], + "angle": 0, + "content": "- Tiny-ImageNet [10] contains 100,000 training and 10,000 testing colourful \\(64 \\times 64\\) 200-class images." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.705, + 0.905, + 0.826 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.514, + 0.838, + 0.904, + 0.862 + ], + "angle": 0, + "content": "2https://github.com/zalandoresearch/fashion-mnist" + }, + { + "type": "page_footnote", + "bbox": [ + 0.534, + 0.863, + 0.892, + 0.875 + ], + "angle": 0, + "content": "3https://www.cs.toronto.edu/%7Ekriz/cifar.html" + }, + { + "type": "page_footnote", + "bbox": [ + 0.534, + 0.876, + 0.892, + 0.887 + ], + "angle": 0, + "content": "4https://www.cs.toronto.edu/%7Ekriz/cifar.html" + }, + { + "type": "page_footnote", + "bbox": [ + 0.534, + 0.888, + 0.824, + 0.9 + ], + "angle": 0, + "content": "\\(^{5}\\)https://tiny-imagenet.herokuapp.com/" + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.838, + 0.904, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1318" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.133, + 0.09, + 0.442, + 0.103 + ], + "angle": 0, + "content": "Table 1. Heterogeneous CNNs for Fashion-MNIST." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.115, + 0.482, + 0.228 + ], + "angle": 0, + "content": "
LayerCNN-1CNN-2CNN-3CNN-4CNN-5
Conv15×5, 205×5, 205×5, 205×5, 205×5, 20
Maxpool12×22×22×22×22×2
Conv25×5, 205×5, 205×5, 205×5, 205×5, 20
Maxpool22×22×22×22×22×2
FC130020015010050
FC25050505050
FC31010101010
model size0.47 MB0.33 MB0.26 MB0.19 MB0.12 MB
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.093, + 0.228, + 0.424, + 0.238 + ], + "angle": 0, + "content": "Note: kernel size: \\(5 \\times 5\\), with 20 filters for convolution layers." + }, + { + "type": "table_caption", + "bbox": [ + 0.101, + 0.253, + 0.472, + 0.266 + ], + "angle": 0, + "content": "Table 2. Heterogeneous CNNs for CIFAR-10 and CIFAR-100." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.389 + ], + "angle": 0, + "content": "
LayerCNN-1CNN-2CNN-3CNN-4CNN-5
Conv15×5, 165×5, 165×5, 165×5, 165×5, 16
Maxpool12×22×22×22×22×2
Conv25×5, 325×5, 165×5, 325×5, 325×5, 32
Maxpool22×22×22×22×22×2
FC1200020001000800500
FC2500500500500500
FC310/10010/10010/10010/10010/100
model size10.01 MB6.93 MB5.04 MB4.05 MB2.56 MB
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.093, + 0.389, + 0.455, + 0.4 + ], + "angle": 0, + "content": "Note: kernel size: \\(5 \\times 5\\), with 16 or 32 filters for convolution layers." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.426, + 0.476, + 0.441 + ], + "angle": 0, + "content": "Then, we partition them into non-IID modes in 2 ways:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.441, + 0.48, + 0.471 + ], + "angle": 0, + "content": "- Pathological. Following [48], we allocate \\(\\{2, 2, 10, 20\\}\\) classes from \\(\\{10, 10, 100, 200\\}\\) classes to each FL client." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.471, + 0.481, + 0.531 + ], + "angle": 0, + "content": "- Practical. Following [43], we allocate \\(\\{10, 10, 100, 200\\}\\) classes to each FL client and use a Dirichlet distribution function with a hyperparameter \\(\\alpha = 0.4\\) to generate diverse ratios of one class for different clients." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.441, + 0.481, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.532, + 0.483, + 0.698 + ], + "angle": 0, + "content": "Base Models. Following [62], we construct 5 heterogeneous CNN models \\(\\{\\mathrm{CNN - 1,}\\dots,\\mathrm{CNN - 5}\\}\\) with different channels of convolutional layers or dimensions of linear layers (Tables 1 and 2) and allocate them to 100 clients. The client \\(k\\) is assigned CNN- \\((k\\% 5)\\). Note that the representation \\(\\mathcal{R}\\) dimension \\(r\\) (the dimension of the penultimate linear layer) is \\(\\{50,500,500\\}\\) for Fashion-MNIST, CIFAR-10 and CIFAR-100. Thus, the dimensions of the homogeneous representation angle square matrix \\(\\mathcal{A}\\) are \\(\\{50\\times 50,500\\times 500,500\\times 500\\}\\). For Tiny-ImageNet, we evenly allocate ResNet-\\(\\{18,34,50,101,152\\}\\) to 100 clients." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.698, + 0.483, + 0.744 + ], + "angle": 0, + "content": "Comparison Baselines. We compare FedRAL with 6 baselines, which are the latest representative MHFL works that support completely heterogeneous local models." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.747, + 0.483, + 0.775 + ], + "angle": 0, + "content": "- Standalone: each client solely trains its local heterogeneous model, without FL." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.777, + 0.37, + 0.792 + ], + "angle": 0, + "content": "- Model Decoupling: LG-FedAvg [30]." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.793, + 0.432, + 0.807 + ], + "angle": 0, + "content": "- Mutual Learning: FedKD [53], FedAPEN [43]." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.808, + 0.482, + 0.838 + ], + "angle": 0, + "content": "- Knowledge Distillation without Public Dataset: FedProto [51], FedTGP [72]." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.747, + 0.483, + 0.838 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.84, + 0.483, + 0.87 + ], + "angle": 0, + "content": "Evaluation Metrics. We evaluate the performances of FedRAL and baselines with the following three metrics:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- Model Accuracy. We record each client's model accuracy (\\%) and calculate their average in each communica" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.84, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.868, + 0.107 + ], + "angle": 0, + "content": "tion round. We report the highest average accuracy." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.107, + 0.905, + 0.182 + ], + "angle": 0, + "content": "- Communication Cost. We track the average number of transmitted parameters between clients and the server in one communication round, and we record the total rounds consumed to reach the specified target accuracy. The product of the two is the total communication cost." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.182, + 0.905, + 0.257 + ], + "angle": 0, + "content": "- Computational Overhead. We track the average computational FLOPs of clients in one communication round, and we record the total communication rounds consumed to reach the specified target accuracy. The product of the two is the total computational overhead." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.107, + 0.905, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.259, + 0.905, + 0.44 + ], + "angle": 0, + "content": "Hyperparameters. We use grid-search to choose the optimal hyperparameters for all algorithms. For general FL hyperparameters, we set the local training epoch \\(E\\) as \\(\\{1,10\\}\\) and batch size \\(B\\) as \\(\\{32,64,128,256,512\\}\\), and we choose the SGD gradient optimizer with a learning rate of 0.01. To ensure the sufficient convergence of algorithms, we set the total communication rounds \\(T\\) as 500. For the unique hyperparameter - the number \\(m\\) of diagonal blocks for sparsifying the homogeneous representation angle square matrix in FedRAL, we set it to be the factor of the representation dimension \\(r\\), e.g., \\(m = \\{1,2,\\dots,50\\}\\) for Fashion-MNIST, and \\(m = \\{1,2,\\dots,500\\}\\) for CIFAR." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.453, + 0.776, + 0.469 + ], + "angle": 0, + "content": "5.2. Comparison Results Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.476, + 0.905, + 0.642 + ], + "angle": 0, + "content": "Average Accuracy. Table 3 shows that FedRAL outperforms all baselines in terms of average test accuracy, demonstrating its effectiveness in enhancing local model generalization through the shared homogeneous representation angle square matrix. As a recent MHFL method, FedTGP achieves the second-best performance. We notice that FedAPEN fails to converge in some cases. This can be attributed to the fact that it uses the same coefficients to combine the outputs of the heterogeneous and homogeneous models for all data samples, failing to balance personalization and generalization under varying data features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.906, + 0.764 + ], + "angle": 0, + "content": "Figure 5 presents how the average test accuracies of FedRAL and the state-of-the-art FedTGP baseline vary as communication rounds. We see that FedRAL converges to a higher accuracy with a faster speed than FedTGP, especially under more non-IID practical FL settings. This demonstrates the effectiveness of FedRAL in alleviating data heterogeneity while supporting the collaboration of heterogeneous client models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Communication Cost. Table 4 presents the communication costs of the state-of-the-art FedTGP baseline and FedRAL. We can observe that FedRAL consumes a lower single-round communication cost than FedTGP (column-3), since the sparse representation angle square matrix communicated in FedRAL has fewer parameters than the label-wise average representations transmitted in FedTGP. FedRAL also requires fewer communication rounds to reach the specified target accuracy (column-5), since the de" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1319" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.131, + 0.09, + 0.864, + 0.104 + ], + "angle": 0, + "content": "Table 3. Comparison of average test accuracy (%) on 4 datasets under pathological and practical data heterogeneity settings." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.115, + 0.905, + 0.226 + ], + "angle": 0, + "content": "
MethodPathologicalPractical
Fashion-MNISTCIFAR-10CIFAR-100Tiny-ImageNetFashion-MNISTCIFAR-10CIFAR-100Tiny-ImageNet
Standalone99.03±0.5391.97±0.3753.04±0.2433.15±0.0273.39±0.8940.72±0.839.56±0.347.35±0.21
LG-FedAvg [30]98.87±0.4091.27±0.2545.83±0.6735.79±0.1971.97±0.2540.44±0.579.80±0.397.51±0.34
FedKD [53]58.37±0.0573.21±0.5637.21±0.9834.53±0.1645.57±0.4928.14±0.298.29±0.286.29±0.23
FedAPEN [43]44.20±0.58-------
FedProto [51]99.02±0.5092.49±0.2653.67±0.7936.24±0.2774.06±0.6340.20±0.119.27±0.717.63±0.29
FedTGP [72]99.09±0.6592.50±0.3153.20±0.7836.92±0.1073.64±0.6040.95±0.099.37±0.137.65±0.19
FedRAL99.54±0.1693.60±0.1958.70±0.0538.75±0.1378.58±0.1242.97±0.1011.70±0.029.47±0.13
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.093, + 0.227, + 0.757, + 0.238 + ], + "angle": 0, + "content": "Note: “-” denotes failure to converge. Bold indicates the highest accuracy, and underline indicates the second highest accuracy." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.263, + 0.273, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.142, + 0.376, + 0.256, + 0.385 + ], + "angle": 0, + "content": "Communication Round" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.387, + 0.272, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.142, + 0.499, + 0.256, + 0.508 + ], + "angle": 0, + "content": "Communication Round" + }, + { + "type": "image_caption", + "bbox": [ + 0.321, + 0.263, + 0.463, + 0.275 + ], + "angle": 0, + "content": "Pathological, CIFAR-100" + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.275, + 0.465, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.334, + 0.376, + 0.449, + 0.385 + ], + "angle": 0, + "content": "Communication Round" + }, + { + "type": "image", + "bbox": [ + 0.294, + 0.387, + 0.465, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.336, + 0.499, + 0.449, + 0.508 + ], + "angle": 0, + "content": "Communication Round" + }, + { + "type": "image_caption", + "bbox": [ + 0.112, + 0.52, + 0.461, + 0.533 + ], + "angle": 0, + "content": "Figure 5. Average test accuracy \\((\\%)\\) vs. number of rounds." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.565, + 0.481, + 0.625 + ], + "angle": 0, + "content": "signed representation angle learning facilitates model convergence. Overall, FedRAL consumes lower communication costs for reaching the target accuracy than FedTGP (column-6), demonstrating its efficient communication." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.628, + 0.483, + 0.822 + ], + "angle": 0, + "content": "Computational Cost. Table 4 also shows the computational costs of the state-of-the-art FedTGP baseline and FedRAL. We can see that FedRAL consumes a lower single-round computational cost than FedTGP (column-4), since additionally training a homogeneous representation angle square matrix in FedRAL consumes lower FLOPs than additionally calculating label-wise average representations in FedTGP, except local model training in both methods. Since FedRAL requires fewer communication rounds to reach the specified target accuracy (column-5), its total computational costs are lower than FedTGP when reaching the specified target accuracy (column-7), substantiating its efficient computation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Privacy Evaluation. We conduct experiments to evaluate the privacy of FedRAL and baselines. Followed by Zhang et al. [70], considering a semi-honest FL scenario where the server is untrustworthy and may inversely infer the original data of comprised clients from transmit-" + }, + { + "type": "table_caption", + "bbox": [ + 0.559, + 0.262, + 0.859, + 0.275 + ], + "angle": 0, + "content": "Table 4. Communication and computational costs." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.286, + 0.905, + 0.39 + ], + "angle": 0, + "content": "
FL SettingMethodComm/C/R (KB)Comp/C/R (MB)RoundsComm/C (MB)Comp/C (GB)
Pathological CIFAR-10FedTGP3.9154.523021.1516.08
FedRAL1.5628.232010.315.54
Pathological CIFAR-100FedTGP19.5354.523546.7518.85
FedRAL9.7728.231441.373.97
Practical CIFAR-10FedTGP19.5354.873486.6418.65
FedRAL7.8128.581030.792.87
Practical CIFAR-100FedTGP195.3154.8736369.2419.45
FedRAL39.0628.581465.574.07
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.513, + 0.39, + 0.902, + 0.486 + ], + "angle": 0, + "content": "Note: \"Comm/C/R\": a single client's communication cost (KB) in a communication round. \"Comp/C/R\": a single client's computational cost (MB) in a communication round. \"Rounds\": the communication rounds required to reach the specified \\(\\{90, 50, 40, 9\\}\\%\\) target accuracy (the near maximum accuracy that both the state-of-the-art FedTGP baseline and FedRAL can achieve). \"Comm/C\": a single client's communication cost (MB) required to reach the target accuracy. \"Comm/C\": a single client's computational cost (GB) required to reach the target accuracy." + }, + { + "type": "table_caption", + "bbox": [ + 0.558, + 0.5, + 0.86, + 0.513 + ], + "angle": 0, + "content": "Table 5. PSNR under DLG attacks on CIFAR-100." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.525, + 0.905, + 0.55 + ], + "angle": 0, + "content": "
MethodLG-FedAvgFedKDFedAPENFedProtoFedTGPFedRAL
PSNR (dB)7.836.956.936.766.696.28
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.565, + 0.905, + 0.592 + ], + "angle": 0, + "content": "Table 6. Average Accuracy \\((\\%)\\) under high model heterogeneity. \" -\" denotes failure to converge." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.604, + 0.905, + 0.628 + ], + "angle": 0, + "content": "
MethodLG-FedAvgFedKDFedAPENFedProtoFedTGPFedRAL
Accuracy (%)45.6345.95-48.5648.6950.79
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.654, + 0.906, + 0.745 + ], + "angle": 0, + "content": "ted information by launching Deep Leakage from Gradients (DLG) attack. We use the Peak Signal-to-Noise Ratio (PSNR), a common privacy measurement metric, to evaluate the privacy-protecting level. A lower PSNR denotes better privacy protection. The results in Table 5 validate that FedRAL can protect data privacy effectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.746, + 0.906, + 0.837 + ], + "angle": 0, + "content": "High Model Heterogeneity. To evaluate how FedRAL and baselines perform under high model heterogeneity, we conduct experiments CIFAR-100 with 100 clients and heterogeneous models including ResNet-{4, 6, 8, 10, 18, 34, 50, 101, 152}, CNN, and ViT. The results in Table 6 validate FedRAL's robustness to high model heterogeneity." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.707, + 0.864 + ], + "angle": 0, + "content": "5.3. Robustness Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "This section analyzes the robustness of FedRAL to different non-IID degrees of two types." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1320" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.09, + 0.273, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.09, + 0.468, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.137, + 0.224, + 0.438, + 0.239 + ], + "angle": 0, + "content": "Figure 6. Robustness to pathological non-IIDness." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.253, + 0.273, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.253, + 0.468, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.148, + 0.389, + 0.427, + 0.404 + ], + "angle": 0, + "content": "Figure 7. Robustness to practical non-IIDness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.43, + 0.484, + 0.61 + ], + "angle": 0, + "content": "Robustness to pathological non-IIDness. We change the number of classes allocated to each client to obtain different non-IID degrees. For CIFAR-10, we allocate \\(\\{2,4,6,8,10\\}\\) classes from 10 classes to each client. For CIFAR-100, we allocate \\(\\{10,30,50,70,90,100\\}\\) classes out of 100 classes to each client. Figure 6 presents that FedRAL always performs higher accuracies than the state-of-the-art FedTGP baseline. As the number of classes assigned to one client increases, i.e., the non-IID degree decreases, model accuracy tends to drop since more IID data benefits generalization while compromising personalization of local models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.612, + 0.484, + 0.779 + ], + "angle": 0, + "content": "Robustness to practical non-IIDness. We vary the hyperparameter \\(\\alpha\\) of the Dirichlet distribution function to construct diverse non-IID degrees. For both CIFAR-10 and CIFAR-100, we set \\(\\alpha = \\{0.1, 0.2, 0.3, 0.4, 0.5\\}\\) to evaluate the average model accuracies of FedRAL and the state-of-the-art FedTGP baseline. Figure 7 shows that FedRAL performs significantly higher average model accuracy than FedTGP, again demonstrating the robustness of FedRAL to non-IIDness. Similarly, as the value of \\(\\alpha\\) increases, i.e., the non-IID degree decreases, model accuracy tends to degrade also due to the above reason." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.788, + 0.265, + 0.804 + ], + "angle": 0, + "content": "5.4. Ablation Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "FedRAL involves three contributions: 1) introducing representation angle matrix learning to share cross-client representation angle knowledge for enhancing model generalization, 2) devising adaptive diagonal sparsification to improve communication efficiency while keeping important representation angle knowledge, 3) designing ordinate-wise" + }, + { + "type": "table_caption", + "bbox": [ + 0.545, + 0.09, + 0.875, + 0.104 + ], + "angle": 0, + "content": "Table 7. Average accuracy \\((\\%)\\) in ablation experiments." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.115, + 0.906, + 0.166 + ], + "angle": 0, + "content": "
CaseDiagonal SparsificationWeighted AggregationPathologicalPractical
CIFAR-10CIFAR-100CIFAR-10CIFAR-100
A✓✓93.6058.7042.9711.70
B✗✓91.6456.5841.3711.58
C✓✗93.4753.9440.529.24
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.199, + 0.905, + 0.274 + ], + "angle": 0, + "content": "weighted aggregation to fuse cross-client representation angle square matrices. Contribution-1) is essential to implement FL collaboration across clients with heterogeneous models, so we conduct experiments to evaluate the necessity of the other two contributions with three cases:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.278, + 0.905, + 0.322 + ], + "angle": 0, + "content": "- Case-A, FedRAL with both adaptive diagonal sparsification and ordinate-wise weighted aggregation, i.e., the designed method." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.323, + 0.905, + 0.382 + ], + "angle": 0, + "content": "- Case-B, FedRAL without adaptive diagonal sparsification but with ordinate-wise weighted aggregation, i.e., clients upload complete representation angle square matrices to the server." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.384, + 0.905, + 0.442 + ], + "angle": 0, + "content": "- Case-C, FedRAL with adaptive diagonal sparsification but with naive ordinate-wise averaging aggregation, i.e., the aggregation rule (a) displayed in Figure 4, ignoring client data distribution distinctions." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.278, + 0.905, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.447, + 0.906, + 0.523 + ], + "angle": 0, + "content": "Table 7 shows that removing anyone degrades accuracy. Especially, replacing ordinate-wise weighted aggregation with naive averaging aggregation presents obvious accuracy degradations under most settings. These results reflect the effectiveness and necessity of the proposed two designs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.547, + 0.634, + 0.563 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.576, + 0.906, + 0.698 + ], + "angle": 0, + "content": "This work proposed FedRAL with three innovative designs: representation angle learning, adaptive diagonal sparsification, and element-wise weighted aggregation. The three designs enable enabling it to be a privacy-preserved, well-performed, communication and computation-efficient model-heterogeneous FL approach. Extensive experiments demonstrate its advantages in accuracy, communication and computation efficiency." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.721, + 0.666, + 0.739 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Xiaoguang Liu is supported by the National Science Foundation of China under Grant 62272252 & 62272253, and the Fundamental Research Funds for the Central Universities. Han Yu is supported by the Ministry of Education, Singapore, under its Academic Research Fund Tier 1 (RG101/24); the RIE2025 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) (Award I2301E0026), administered by A\\*STAR, as well as supported by Alibaba Group and NTU Singapore through Alibaba-NTU Global e-Sustainability CorpLab (ANGEL)." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1321" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.482, + 0.156 + ], + "angle": 0, + "content": "[1] Jin-Hyun Ahn et al. Wireless federated distillation for distributed edge learning with heterogeneous data. In Proc. PIMRC, pages 1-6, Istanbul, Turkey, 2019. IEEE. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.483, + 0.2 + ], + "angle": 0, + "content": "[2] Jin-Hyun Ahn et al. Cooperative learning VIA federated distillation OVER fading channels. In Proc. ICASSP, pages 8856-8860, Barcelona, Spain, 2020. IEEE. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.202, + 0.482, + 0.242 + ], + "angle": 0, + "content": "[3] Samiul Alam et al. Fedrolex: Model-heterogeneous federated learning with rolling sub-model extraction. In Proc. NeurIPS, virtual, 2022.. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.245, + 0.483, + 0.285 + ], + "angle": 0, + "content": "[4] Sara Babakniya et al. Revisiting sparsity hunting in federated learning: Why does sparsity consensus matter? Transactions on Machine Learning Research, 1(1):1, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.288, + 0.482, + 0.355 + ], + "angle": 0, + "content": "[5] Yun-Hin Chan, Rui Zhou, Running Zhao, Zhihan JIANG, and Edith C. H. Ngai. Internal cross-layer gradients for extending homogeneity to heterogeneity in federated learning. In Proc. ICLR, page 1, Vienna, Austria, 2024. OpenReview.net. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.359, + 0.482, + 0.4 + ], + "angle": 0, + "content": "[6] Hongyan Chang et al. Cronus: Robust and heterogeneous collaborative learning with black-box knowledge transfer. In Proc. NeurIPS Workshop, virtual, 2021.. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.402, + 0.482, + 0.442 + ], + "angle": 0, + "content": "[7] Jiangui Chen et al. Fedmatch: Federated learning over heterogeneous question answering data. In Proc. CIKM, pages 181-190, virtual, 2021. ACM. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.445, + 0.482, + 0.485 + ], + "angle": 0, + "content": "[8] Sijie Cheng et al. Fedgems: Federated learning of larger server models via selective knowledge fusion. CoRR, abs/2110.11027, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.488, + 0.482, + 0.529 + ], + "angle": 0, + "content": "[9] Yae Jee Cho et al. Heterogeneous ensemble knowledge transfer for training large models in federated learning. In Proc. IJCAI, pages 2881-2887, virtual, 2022. ijcai.org. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.531, + 0.482, + 0.571 + ], + "angle": 0, + "content": "[10] Patryk Chrabaszcz, Ilya Loshchilov, and Frank Hutter. A downsampled variant of imagenet as an alternative to the CIFAR datasets. CoRR, abs/1707.08819, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.574, + 0.482, + 0.614 + ], + "angle": 0, + "content": "[11] Liam Collins et al. Exploiting shared representations for personalized federated learning. In Proc. ICML, pages 2089-2099, virtual, 2021. PMLR. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.617, + 0.482, + 0.669 + ], + "angle": 0, + "content": "[12] Enmao Diao. Heterofl: Computation and communication efficient federated learning for heterogeneous clients. In Proc. ICLR, page 1, Virtual Event, Austria, 2021. OpenReview.net. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.674, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[13] Tao Fan, Hanlin Gu, et al. Ten challenging problems in federated foundation models. IEEE Transactions on Knowledge and Data Engineering, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.717, + 0.482, + 0.757 + ], + "angle": 0, + "content": "[14] Randy Goebel, Han Yu, Boi Faltings, Lixin Fan, and Zehui Xiong. Trustworthy Federated Learning. Springer, Cham, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.76, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[15] Xuan Gong et al. Federated learning via input-output collaborative distillation. In Proc. AAAI, pages 22058-22066, Vancouver, Canada, 2024. AAAI Press. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.482, + 0.842 + ], + "angle": 0, + "content": "[16] Chaoyang He et al. Group knowledge transfer: Federated learning of large cnns at the edge. In Proc. NeurIPS, virtual, 2020. . 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.898 + ], + "angle": 0, + "content": "[17] S. Horváth. FjORD: Fair and accurate federated learning under heterogeneous targets with ordered dropout. In Proc. NIPS, pages 12876-12889, Virtual, 2021. OpenReview.net. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.483, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "[18] Edward J. Hu et al. Lora: Low-rank adaptation of large language models. In ICLR, page 1, Virtual, 2022. OpenReview.net. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.175 + ], + "angle": 0, + "content": "[19] Wenke Huang et al. Learn from others and be yourself in heterogeneous federated learning. In Proc. CVPR, pages 10133-10143, virtual, 2022. IEEE. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[20] Wenke Huang et al. Few-shot model agnostic federated learning. In Proc. MM, pages 7309-7316, Lisboa, Portugal, 2022. ACM." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.905, + 0.273 + ], + "angle": 0, + "content": "[21] Sohei Itahara et al. Distillation-based semi-supervised federated learning for communication-efficient collaborative training with non-iid private data. IEEE Trans. Mob. Comput., 22(1):191-205, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.275, + 0.905, + 0.327 + ], + "angle": 0, + "content": "[22] Jaehee Jang et al. Fedclassavg: Local representation learning for personalized federated learning on heterogeneous neural networks. In Proc. ICPP, pages 76:1-76:10, virtual, 2022. ACM, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.383 + ], + "angle": 0, + "content": "[23] Eunjoeong Jeong et al. Communication-efficient on-device machine learning: Federated distillation and augmentation under non-iid private data. In Proc. NeurIPS Workshop, virtual, 2018. . 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.905, + 0.425 + ], + "angle": 0, + "content": "[24] Peter Kairouz et al. Advances and open problems in federated learning. Foundations and Trends in Machine Learning, 14(1-2):1-210, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.905, + 0.466 + ], + "angle": 0, + "content": "[25] Shivam Kalra et al. Decentralized federated learning through proxy model sharing. Nature communications, 14(1):2899, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.904, + 0.496 + ], + "angle": 0, + "content": "[26] Alex Krizhevsky et al. Learning multiple layers of features from tiny images. Toronto, ON, Canada, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.498, + 0.905, + 0.565 + ], + "angle": 0, + "content": "[27] Meng Lei, Qi Zhuang, Wu Lei, Du Xiaoyu, Li Zhaochuan, Cui Lizhen, and Meng Xiangxu. Improving global generalization and local personalization for federated learning. IEEE Transactions on Neural Networks and Learning Systems, 36, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.568, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[28] Daliang Li and Junpu Wang. Fedmd: Heterogenous federated learning via model distillation. In Proc. NeurIPS Workshop, virtual, 2019. . 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.609, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[29] Qinbin Li et al. Practical one-shot federated learning for cross-silo setting. In Proc. IJCAI, pages 1484-1490, virtual, 2021. ijcai.org. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.652, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[30] Paul Pu Liang et al. Think locally, act globally: Federated learning with local and global representations. arXiv preprint arXiv:2001.01523, 1(1), 2020. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.693, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[31] Tao Lin et al. Ensemble distillation for robust model fusion in federated learning. In Proc. NeurIPS, virtual, 2020. . 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.721, + 0.905, + 0.747 + ], + "angle": 0, + "content": "[32] Chang Liu et al. Completely heterogeneous federated learning. CoRR, abs/2210.15865, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.749, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[33] Qianyu Long, Christos Anagnostopoulos, Shameem Puthiya Parambath, and Daning Bi. Feddip: Federated learning with extreme dynamic pruning and incremental regularization. In Proc.ICDM, pages 1187-1192, Shanghai, China, 2023. IEEE. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.818, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[34] Disha Makhija et al. Architecture agnostic federated learning for neural networks. In Proc. ICML, pages 14860-14870, virtual, 2022. PMLR. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.86, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[35] Koji Matsuda et al. Fedme: Federated learning via model exchange. In Proc. SDM, pages 459-467, Alexandria, VA, USA, 2022. SIAM. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1322" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[36] Brendan McMahan et al. Communication-efficient learning of deep networks from decentralized data. In Proc. AISTATS, pages 1273-1282, Fort Lauderdale, FL, USA, 2017. PMLR. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[37] Duy Phuong Nguyen et al. Enhancing heterogeneous federated learning with knowledge extraction and multi-model fusion. In Proc. SC Workshop, pages 36-43, Denver, CO, USA, 2023. ACM. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.206, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[38] Jaehoon Oh et al. Fedbabu: Toward enhanced representation for federated image classification. In Proc. ICLR, virtual, 2022. OpenReview.net. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.303 + ], + "angle": 0, + "content": "[39] Zhengxiang Pan, Han Yu, Chunyan Miao, and Cyril Leung. Efficient collaborative crowdsourcing. In The 30th AAAI Conference on Artificial Intelligence (AAAI-16), pages 4248–4249, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.482, + 0.347 + ], + "angle": 0, + "content": "[40] Sejun Park et al. Towards understanding ensemble distillation in federated learning. In Proc. ICML, pages 27132-27187, Honolulu, Hawaii, USA, 2023. PMLR. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[41] Krishna Pillutla et al. Federated learning with partial model personalization. In Proc. ICML, pages 17716-17758, virtual, 2022. PMLR. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.461 + ], + "angle": 0, + "content": "[42] Zhuang Qi, Sijin Zhou, Lei Meng, Han Yu, and Xiangxu Meng. Federated deconfounding and debiasing learning for out-of-distribution generalization. In The 34th International Joint Conference on Artificial Intelligence, (IJ-CAI'25), pages 1–9, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.461, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[43] Zhen Qin et al. Fedapen: Personalized cross-silo federated learning with adaptability to statistical heterogeneity. In Proc. KDD, pages 1954-1964, Long Beach, CA, USA, 2023. ACM. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.518, + 0.482, + 0.559 + ], + "angle": 0, + "content": "[44] Zeju Qiu et al. Controlling text-to-image diffusion by orthogonal finetuning. In Proc. NeurIPS, New Orleans, LA, USA, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.561, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[45] Chao Ren, Han Yu, et al. Advances and open challenges in federated foundation models. IEEE Communications Surveys and Tutorials, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.603, + 0.482, + 0.644 + ], + "angle": 0, + "content": "[46] Felix Sattler et al. Fedaux: Leveraging unlabeled auxiliary data in federated learning. IEEE Trans. Neural Networks Learn. Syst., 1(1):1-13, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.646, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[47] Felix Sattler et al. CFD: communication-efficient federated distillation via soft-label quantization and delta coding. IEEE Trans. Netw. Sci. Eng., 9(4):2025-2038, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.73 + ], + "angle": 0, + "content": "[48] Aviv Shamsian et al. Personalized federated learning using hypernetworks. In Proc. ICML, pages 9489-9502, virtual, 2021. PMLR. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[49] Tao Shen et al. Federated mutual learning. CoRR, abs/2006.16765, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[50] Alysa Ziying Tan et al. Towards personalized federated learning. IEEE Trans. Neural Networks Learn. Syst., 1(1): 1-17, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[51] Yue Tan et al. Fedproto: Federated prototype learning across heterogeneous clients. In Proc. AAAI, pages 8432-8440, virtual, 2022. AAAI Press. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[52] Jiaqi Wang et al. Towards personalized federated learning via heterogeneous model reassembly. In Proc. NeurIPS, page 13, New Orleans, Louisiana, USA, 2023. OpenReview.net. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[53] Chuhan Wu et al. Communication-efficient federated learning via knowledge distillation. Nature Communications, 13 (1):2032, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[54] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.181, + 0.905, + 0.209 + ], + "angle": 0, + "content": "[55] Qiang Yang, Lixin Fan, and Han Yu. Federated Learning: Privacy and Incentive. Springer, Cham, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.211, + 0.905, + 0.251 + ], + "angle": 0, + "content": "[56] Qiang Yang, Yang Liu, Yong Cheng, Yan Kang, Tianjian Chen, and Han Yu. Federated Learning. Springer, Cham, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.254, + 0.905, + 0.307 + ], + "angle": 0, + "content": "[57] Liping Yi, Gang Wang, and Xiaoguang Liu. QSFL: A two-level uplink communication optimization framework for federated learning. In Proc. ICML, pages 25501-25513. PMLR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.311, + 0.905, + 0.367 + ], + "angle": 0, + "content": "[58] Liping Yi, Xiaorong Shi, Nan Wang, Gang Wang, Xiaoguang Liu, Zhuan Shi, and Han Yu. pfedkt: Personalized federated learning with dual knowledge transfer. Knowledge-Based Systems, 292:111633, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.369, + 0.905, + 0.425 + ], + "angle": 0, + "content": "[59] Liping Yi, Xiaorong Shi, Nan Wang, Jinsong Zhang, Gang Wang, and Xiaoguang Liu. Fedpe: Adaptive model pruning-expanding for federated learning on mobile devices. IEEE Transactions on Mobile Computing, pages 1-18, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.427, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[60] Liping Yi, Gang Wang, Xiaofei Wang, and Xiaoguang Liu. Qsfl: Two-level communication-efficient federated learning on mobile edge devices. IEEE Transactions on Services Computing, pages 1-16, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.485, + 0.905, + 0.54 + ], + "angle": 0, + "content": "[61] Liping Yi, Han Yu, Zhuan Shi, Gang Wang, Xiaoguang Liu, Lizhen Cui, and Xiaoxiao Li. FedSSA: Semantic Similarity-based Aggregation for Efficient Model-Heterogeneous Personalized Federated Learning. In IJCAI, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.542, + 0.905, + 0.583 + ], + "angle": 0, + "content": "[62] Liping Yi et al. FedGH: Heterogeneous federated learning with generalized global header. In Proc. ACM MM, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.585, + 0.905, + 0.626 + ], + "angle": 0, + "content": "[63] Liping Yi et al. Federated model heterogeneous matryoshka representation learning. In Proc. NeurIPS, Vancouver, Canada, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.629, + 0.905, + 0.699 + ], + "angle": 0, + "content": "[64] Liping Yi et al. pFedAFM: Adaptive Feature Mixture for Data-Level Personalization in Heterogeneous Federated Learning on Mobile Edge Devices. In Proc. ICDE, pages 1981-1994, Hong Kong SAR, China, 2025. IEEE Computer Society." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.701, + 0.905, + 0.756 + ], + "angle": 0, + "content": "[65] Liping Yi et al. pfedes: Generalized proxy feature extractor sharing for model heterogeneous personalized federated learning. In Proc. AAAI, pages 22146-22154, Philadelphia, PA, USA, 2025. AAAI Press. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.758, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[66] Fuxun Yu et al. Fed2: Feature-aligned federated learning. In Proc. KDD, pages 2066-2074, virtual, 2021. ACM. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[67] Han Yu, Siyuan Liu, Alex C Kot, Chunyan Miao, and Cyril Leung. Dynamic witness selection for trustworthy distributed cooperative sensing in cognitive radio networks. In Proceedings of the 13th IEEE International Conference on Communication Technology (ICCT'11), pages 1-6, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[68] Sixing Yu et al. Resource-aware federated learning using knowledge extraction and multi-model fusion. CoRR, abs/2208.07978, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1323" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[69] Jie Zhang et al. Parameterized knowledge transfer for personalized federated learning. In Proc. NeurIPS, pages 10092-10104, virtual, 2021. OpenReview.net. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.177 + ], + "angle": 0, + "content": "[70] Jianqing Zhang et al. Fedcp: Separating feature information for personalized federated learning via conditional policy. In Proc. KDD, page 1, Long Beach, CA, USA, 2023. ACM. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[71] Jie Zhang et al. Towards data-independent knowledge transfer in model-heterogeneous federated learning. IEEE Trans. Computers, 72(10):2888-2901, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.222, + 0.482, + 0.288 + ], + "angle": 0, + "content": "[72] Jianqing Zhang et al. Fedtgp: Trainable global prototypes with adaptive-margin-enhanced contrastive learning for data and model heterogeneity in federated learning. In Proc. AAAI, pages 16768-16776, Vancouver, Canada, 2024. 1. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[73] Lan Zhang et al. Fedzkt: Zero-shot knowledge transfer towards resource-constrained federated learning with heterogeneous on-device models. In Proc. ICDCS, pages 928-938, virtual, 2022. IEEE. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[74] Zhilu Zhang and Mert R. Sabuncu. Generalized cross entropy loss for training deep neural networks with noisy labels. In Proc. NeurIPS, pages 8792-8802, Montreal, Canada, 2018. Curran Associates Inc. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[75] Hangyu Zhu et al. Federated learning on non-iid data: A survey. Neurocomputing, 465:371-390, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.482, + 0.474 + ], + "angle": 0, + "content": "[76] Zhuangdi Zhu et al. Data-free knowledge distillation for heterogeneous federated learning. In Proc. ICML, pages 12878-12889, virtual, 2021. PMLR. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.482, + 0.518 + ], + "angle": 0, + "content": "[77] Zhuangdi Zhu et al. Resilient and communication efficient learning for heterogeneous federated systems. In Proc. ICML, pages 27504-27526, virtual, 2022. PMLR. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.52, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[78] Qi Zhuang, Meng Lei, Chen Zitan, Hu Han, Lin Hui, and Meng Xiangxu. Cross-silo prototypical calibration for federated learning with non-iid data. In Proceedings of the 31st ACM International Conference on Multimedia, pages 3099-3107, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.645 + ], + "angle": 0, + "content": "[79] Qi Zhuang, Meng Lei, et al. Cross-silo feature space alignment for federated learning on clients with imbalanced data. In The 39th Annual AAAI Conference on Artificial Intelligence (AAAI-25), pages 19986-19994, 2025. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.645 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1324" + } + ] +] \ No newline at end of file diff --git a/2025/Federated Representation Angle Learning/f94f188a-83f1-436a-a6d8-36878318e078_origin.pdf b/2025/Federated Representation Angle Learning/f94f188a-83f1-436a-a6d8-36878318e078_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0f3967f4037d077cce493aaa6546b355d3fcab6f --- /dev/null +++ b/2025/Federated Representation Angle Learning/f94f188a-83f1-436a-a6d8-36878318e078_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:87cfba14ed0a916e7506a0f28a5cba4101a424d3669b448d2b996d1d1c216528 +size 799770 diff --git a/2025/Federated Representation Angle Learning/full.md b/2025/Federated Representation Angle Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f625abc9f410ffb8b685a5fa3a3569087aa1e889 --- /dev/null +++ b/2025/Federated Representation Angle Learning/full.md @@ -0,0 +1,452 @@ +# Federated Representation Angle Learning* + +Liping Yi $^{1}$ , Han Yu $^{2}$ , Gang Wang $^{1,\ast}$ , Xiaoguang Liu $^{1,\ast}$ , Xiaoxiao Li $^{3,4}$ $^{1}$ College of Computer Science, TMCC, SysNet, DISSec, GTIISC, Nankai University, Tianjin, China + $^{2}$ College of Computing and Data Science, Nanyang Technological University, Singapore + $^{3}$ Department of Electrical and Computer Engineering, University of British Columbia, Vancouver, Canada + $^{4}$ Vector Institute, Vancouver, Canada +{yiliping, wgzwp, liuxg}@nbjl.nankai.edu.cn, han.yu@ntu.edu.sg, xiaoxiao.li@ece.ubc.ca + +# Abstract + +Model-heterogeneous federated learning (MHFL) is a challenging FL paradigm designed to allow FL clients to train structurally heterogeneous models under the coordination of an FL server. Existing MHFL methods face significant limitations when it comes to transferring global knowledge to clients as a result of sharing only partial homogeneous model parameters or calculating distance loss, leading to inferior model generalization. To bridge this gap, we propose a novel model-heterogeneous Federated learning method with Representation Angle Learning (FedRAL). It consists of three innovative designs: (1) We first introduce representation angle learning into MHFL. Specifically, we embed a homogeneous square matrix into the local heterogeneous model of each client, which learns the angle information of local representations. These homogeneous representation angle square matrices are aggregated on the server to fuse representation angle knowledge shared by clients for enhancing the generalization of local representations. (2) As different clients might have heterogeneous system resources, we propose an adaptive diagonal sparsification strategy to reduce the numbers of the parameters of representation angle square matrices uploaded to the server, to improve FL communication efficiency. (3) To enable effective fusion of sparsified homogeneous local representation angle square matrices, we design an element-wise weighted aggregation approach. Experiments on 4 benchmark datasets under 2 types of non-IID divisions over 6 state-of-the-art baselines demonstrate that FedRAL achieves the best performance. It improves test accuracy, communication efficiency and computational efficiency by up to $5.03\%$ , $12.43\times$ and $6.49\times$ , respectively. + +# 1. Introduction + +Federated learning (FL) [13, 14, 24, 36, 45, 55, 56] is a privacy-preserved distributed machine learning paradigm. Typically, an FL server is often involved in broadcast a global model to FL clients. The clients then further train the global model on local data to obtain local models. The server aggregates local model updates received from the clients to produce a new global model. In this way, client data are not exposed. Under this setting, the models trained by the clients must follow the same structure. + +In practice, clients participating in FL often have non-independent and identically distributed (non-IID) data, a.k.a., data heterogeneity [27, 42, 50, 58, 75, 78, 79], the local models trained on such data are biased, so the global model obtained by aggregating them may not perform well on all client data. Besides, FL clients also have heterogeneous system configurations in terms of computing power and communication bandwidth, a.k.a., system heterogeneity [39, 57, 59, 60, 67]. Thus, requesting all clients to train the same model might not be viable for some weaker clients, while leaving stronger clients under-utilized. More critically, some companies or institutions joining FL as clients might face intellectual property issues regarding their proprietary local models. Thus, they might be reluctant to share their private heterogeneous models, a.k.a., model heterogeneity. These challenges have inspired the field of model-heterogeneous federated learning (MHFL) with preserved data and model privacy, which advocates avoiding directly sharing private local heterogeneous models [61-65]. + +Existing MHFL approaches mainly include: a) Modeldecoupling: each client model is decoupled into heterogeneous and homogeneous parts, with only homogeneous parts being shared with the server [11, 30]. b) Knowledge distillation: the server aggregates local representation or logits from different clients to generate global representation or logits, which is used to calculate distillation loss for local training [28, 31]. c) Mutual training: each client is + +assigned a shared homogeneous small model and trains it with the local heterogeneous model by mutual loss between them [49, 53]. Although these methods support model heterogeneity, their capability for transferring global knowledge from the server to clients is limited due to sharing partial model parameters or using distillation loss, thereby constraining the generalization of local models. Whereas, model representations often involve more sufficient semantic knowledge than partial model parameters or distillation loss, which are explored in this work for knowledge sharing. + +A recent work [44] empirically verified that parameter fine-tuning approaches (e.g., LoRA [18]) cannot sufficiently retain semantic information of the pre-trained model. Whereas, fine-tuning the angles (i.e., gradient directions) of the pre-trained model can achieve this, thereby enhancing the generalization of the pre-trained model. The work further proposed the Orthogonal Fine-Tuning (OFT) method to fine-tune the parameter angles of the pre-trained models via an orthogonal matrix (Figure 1). In short, parameter angles involve sufficient semantic information that can enhance model generalization. Therefore, we can enhance FL model generalization through sharing parameter angles instead of model parameters. However, the FL server can not aggregate parameter angles from structure-heterogeneous client models. Considering representations also with sufficient semantic knowledge, we attempt to share homogeneous representation angles among clients to achieve generalization-enhanced model-heterogeneous FL. + +To this end, we propose a novel model-heterogeneous Federated learning method with Representation Angel Learning (FedRAL) to enhance the transfer of generalizable global knowledge from the FL server to FL clients. It consists of three innovative designs: (1) Since representations are extracted high-level features from data, with sufficient semantic information, FedRAL embeds a homogeneous square matrix into the local heterogeneous model to learn the angle information of local extracted representations. The server aggregates local homogeneous representation angle square matrices from different clients to generate a global representation angle square matrix, thereby fusing cross-client representation angle knowledge. Then, the global representation angle square matrix with sufficient generalized representation angle information is used to enhance the generalization of local representations during local model training. Different from the strong assumption of the orthogonality of the matrix in OFT, the homogeneous representation angle square matrix is initialized randomly. (2) Considering the diverse system resources (e.g., communication bandwidth) of different clients, we design an adaptive diagonal sparsification strategy for the homogeneous representation angle square matrix that allows clients to upload diagonal blocks with sizes matching with communication capacity. (3) To effectively aggregate sparsified local + +![](images/304dd7488f93e82b9d90d9fb9eac10b293fd18933f5a8073bd4c51e1ce16d80f.jpg) +Figure 1. Orthogonal Fine-Tuning (OFT). + +homogeneous representation angle square matrices on the server, element-wise weighted aggregation is designed. + +Experiments on 4 benchmark datasets under 2 modes of non-IID divisions against 6 advanced MHFL methods demonstrate the superiority of FedRAL. Compared with the state-of-the-art baseline, FedRAL improves test accuracy by up to $5.03\%$ , while enhancing communication and computational efficiency by up to $12.43\times$ and $6.49\times$ . + +# 2. Related Work + +Based on different approaches used for achieving MHFL, existing methods can be divided into 4 categories. + +Heterogeneous Subnet. Some works [3-5, 12, 17, 33, 66, 77] assume that each client trains local heterogeneous subnets of the global model, and the server aggregates them according to parameter positions to re-construct a new global model. They allow each client to train personalized heterogeneous subnets to tackle data and system heterogeneity. However, uploading subnets exposes model parameters, thereby breaching model IP protection requirements. + +Model Decoupling. These methods [7, 11, 22, 30, 32, 38, 41, 62] share partial local model parameters instead of the entire local model for global aggregation. Specifically, they decouple each client's local heterogeneous model into heterogeneous and homogeneous parts, and only share homogeneous parts for cross-client knowledge exchange. Nevertheless, sharing partial model parameters still violates model IP protection needs. Furthermore, the unshared remaining parameters might overfit due to local training. + +Knowledge Distillation. These methods [1, 2, 6, 8, 9, 15, 16, 19-21, 23, 28, 29, 31, 34, 37, 40, 46, 47, 51, 52, 68, 69, 71-73, 76] aggregate the local seen-class representations or logits extracted on local private data or a public dataset to generate the global representation or logits by class, which is used to calculate distillation loss with the local representations or logits. Although they protect local model structures, the distillation loss only transfers limited knowledge from the server to clients. Besides, sharing class information might be prohibited in some data privacy-sensitive applications. The public dataset used must follow similar data distributions with local private data, which is difficult to access due to data privacy. + +Mutual Learning. These methods [25, 35, 43, 49, 53] add a global homogeneous small model shared by all clients, which is alternately trained with the local heterogeneous model by the mutual loss between the outputs of + +the two models for each sample. Although they hide local heterogeneous model structures, the mutual loss only transfers limited knowledge between the two models. The unstable mutual loss in early training rounds might result in both models not converging. + +Existing MHFL methods generally transfer limited generalized knowledge from the server to clients, leading to model performance bottlenecks. FedRAL effectively tackles data, system and model heterogeneity simultaneously, allowing clients to share a global homogeneous representation angle square matrix for learning generalized representation angle information. It bridges this important gap by enhancing the generalization of local representations via utilizing the global homogeneous representation angle square matrix to fine-tune local representation angles, while preserving personalized semantic information. + +# 3. Preliminaries + +# 3.1. Orthogonal Fine-Tuning + +Fine-tuning the parameters of a pre-trained model has been demonstrated to compromise its semantic information, while fine-tuning its parameter angles (i.e., gradient directions) can sufficiently retain its original semantic information [44]. This enables the preservation of the generalized ability when adapting to downstream tasks. Hypersphere Energy (HE) measures the uniformity of neuron distribution on the unit hypersphere. A lower HE value is beneficial to keeping pre-trained model generalization ability. To minimize HE variance between the pre-trained model $W^0 \in \mathbb{R}^{u \times v}$ and the angle-fine-tuned model $W \in \mathbb{R}^{u \times v}$ , orthogonal finetuning (OFT) [44] fine-tunes the parameter angles of the pre-trained model through an orthogonal matrix $O \in \mathbb{R}^{u \times u}$ , $W = O \times W^0$ , as depicted in Figure 1. To ensure orthogonality, OFT utilizes Cayley parameterization to construct matrix $O = (I + Q)(I - Q)^{-1}$ , where $Q \in \mathbb{R}^{u \times u}$ is a skew-symmetric matrix satisfying $Q = -Q^T$ . Due to inverse operations, updating the orthogonal matrix $O$ during fine-tuning introduces high computational overhead, especially with a large dimension $u$ . + +Unlike OFT, we explore fine-tuning the angles of representations with sufficient semantic information. We relax the orthogonality assumption and avoid expensive inverse operations. Instead, we randomly initialize the homogeneous representation angle square matrix, which is embedded into the local heterogeneous model. They are trained in an end-to-end manner to improve computation efficiency. + +# 3.2. Problem Formulation + +In FedRAL, a server coordinates $K$ clients with heterogeneous local models for collaborative training. Client $k$ 's local model $f_{k}(\omega_{k})$ consists of two parts: 1) a heterogeneous feature extractor $f_{k}^{ex}(\omega_{k}^{ex})$ , and 2) a homogeneous predic + +![](images/0d11c5368c380c506206ab3935ca190e4afb5fa809eeb4bad025bc4c81090165.jpg) +Figure 2. The workflow of FedRAL. + +tion head $f^{hd}(\omega_k^{hd})$ (all clients conduct the same prediction task), $f_{k}(\omega_{k}) = f_{k}^{ex}(\omega_{k}^{ex})\circ f^{hd}(\omega_{k}^{hd})$ , as depicted in Figure 2. Each client trains a local homogeneous square matrix $\mathcal{A}_k$ to learn the angle information of local representations extracted by the heterogeneous feature extractor. The server aggregates local homogeneous representation angle square matrices from different clients to generate a global representation angle square matrix $\mathcal{A}$ for cross-client representation angle knowledge fusion. Therefore, the training objective of FedRAL is to minimize the loss sum of all heterogeneous client models with the help of the global homogeneous representation angle square matrix $\mathcal{A}$ : + +$$ +\min _ {\mathcal {A}, \left\{\omega_ {0}, \dots , \omega_ {K - 1} \right\}} \sum_ {k = 0} ^ {K - 1} \ell \left(\mathcal {F} _ {k} \left(D _ {k}; \left(\omega_ {k} \circ \mathcal {A}\right)\right)\right), \tag {1} +$$ + +where $\ell$ is loss. $\mathcal{F}_k(\omega_k\circ \mathcal{A})$ denotes the combination of the local heterogeneous model $\omega_{k}$ and the global representation angle square matrix $\mathcal{A}$ , $D_{k}$ is the non-IID data of client $k$ . + +# 4. The Proposed FedRAL Approach + +FedRAL consists of three innovative designs: (1) homogeneous representation angle square matrix learning, (2) adaptive diagonal sparsification, and (3) element-wise weighted homogeneous square matrix aggregation. The workflow of FedRAL in a single communication round includes the following steps. + +Overview. As shown in Figure 2, in the $t$ -th communication round, FedRAL executes 4 steps: + +① The server broadcasts the latest global representation angle square matrix $\mathcal{A}^{t-1}$ to $K$ participating clients. +$②$ A client $k$ embeds the global representation angle square matrix $\mathcal{A}^{t - 1}$ into its local model $f_{k}(\omega_{k}^{t - 1})$ to enhance + +the generalization of local representations, while learning the personalized angle information of local representations via end-to-end training. + +③ The local homogeneous representation angle square matrix $\mathcal{A}_k^t$ after local training is sparsified into $\tilde{\mathcal{A}}_k^t$ via the proposed adaptive diagonal sparsification strategy before being uploaded to the server. +(4) The server aggregates sparsified local homogeneous representation angle square matrices $\{\tilde{\mathcal{A}}_0^t,\dots ,\tilde{\mathcal{A}}_{K - 1}^t\}$ via the proposed element-wise weighted aggregation to generate a new global representation angle square matrix $\mathcal{A}^t$ + +Inference. These steps are iteratively executed until heterogeneous client models $\{f_0(\omega_0),\dots ,f_{K - 1}(\omega_{K - 1})\}$ converge, which are used for inference. Algorithm 1 describes the detailed workflow of FedRAL. + +# 4.1. Representation Angle Learning + +During the $t$ -th training round, a client $k$ uses its local heterogeneous feature extractor $f_{k}^{ex}\left(\omega_{k}^{ex,t - 1}\right)$ to extract the personalized representation $\mathcal{R} \in \mathbb{R}^r$ of the input sample $\pmb{x}$ : + +$$ +\mathcal {R} = f _ {k} ^ {e x} \left(\boldsymbol {x}; \omega_ {k} ^ {e x, t - 1}\right). \tag {2} +$$ + +Then, the extracted personalized representation $\mathfrak{R} \in \mathbb{R}^r$ is multiplied by the received global representation angle square matrix $\mathcal{A}^{t-1} \in \mathbb{R}^{r \times r}$ to produce the representation $\tilde{\mathcal{R}} \in \mathbb{R}^r$ with enhanced generalization: + +$$ +\tilde {\mathcal {R}} = \mathcal {R} \times \mathcal {A} ^ {t - 1}. \tag {3} +$$ + +Then, $\mathcal{R} \in \mathbb{R}^r$ and $\hat{\mathcal{R}} \in \mathbb{R}^r$ are summed and fed into the prediction head to make a prediction: + +$$ +\hat {y} = f ^ {h d} \left(\mathcal {R} + \tilde {\mathcal {R}}; \omega_ {k} ^ {h d, t - 1}\right). \tag {4} +$$ + +The loss $\ell$ (e.g., Cross-Entropy loss [74]) between the prediction $\hat{y}$ and the ground-truth label $y$ is used to update the heterogeneous client model $f_{k}(\omega_{k}^{t - 1})$ and the homogeneous representation angle square matrix $\mathcal{A}^{t - 1}$ simultaneously via end-to-end training: + +$$ +f _ {k} \left(\omega_ {k} ^ {t}\right) \leftarrow f _ {k} \left(\omega_ {k} ^ {t - 1}\right) - \eta_ {\omega} \nabla \ell (\hat {y}, y), \tag {5} +$$ + +$$ +\boldsymbol {\mathcal {A}} _ {k} ^ {t} \leftarrow \boldsymbol {\mathcal {A}} ^ {t - 1} - \eta_ {\boldsymbol {\mathcal {A}}} \nabla \ell (\hat {y}, y), +$$ + +where $\eta_{\omega}$ and $\eta_{\mathcal{A}}$ are the learning rates of the local heterogeneous model and the homogeneous representation angle square matrix. We set $\eta_{\omega} = \eta_{\mathcal{A}}$ by default as they are updated at the same time. + +The shared global homogeneous representation angle square matrix serves three functions during local training: a) it transfers globally generalizable representation angle knowledge from the server to the client by producing the generalization-enhanced representation; b) it learns the personalized representation angle knowledge from the + +![](images/2e5e6187e943bfa0893d88b814a31a8597dee93330407dfd89acffe1353a16f0.jpg) +Figure 3. An illustrative example of diagonal sparsification. + +extracted personalized representations, thereby transferring local personalized representation angle knowledge to the server; and c) the summed representation includes both generalized and personalized knowledge, thereby improving model expressiveness while addressing non-IID problem. + +# 4.2. Adaptive Diagonal Sparsification + +To improve communication efficiency while maintaining model performance, we design an adaptive diagonal sparsification strategy to sparsify the local representation angle square matrices before uploading them to the FL server. + +Each element in $\mathcal{R} \in \mathbb{R}^r$ corresponds to the element on the diagonal of $\mathcal{A} \in \mathbb{R}^{r \times r}$ . If we only retain the diagonal elements of $\mathcal{A}$ and set other elements to be 0, the angle of each element in the representation is directly adjusted by the corresponding element in the square matrix diagonal, as depicted in Figure 3(a). When all elements of the square matrix are non-zero, non-diagonal elements might disturb representation angle adjustment due to multiplication with non-corresponding representation elements. As Figure 3(b) shows, the enhanced representation processed by the entire representation angle square matrix is significantly different from the one processed by the sparse square matrix with only diagonal elements. + +Since adjacent elements in the representation are semantically correlated, retaining partial elements wrapped with the diagonal in the representation angle square matrix can enhance the angle correlation of adjacent representation elements. Therefore, the proposed sparsification strategy retains the elements from the diagonal blocks of the square matrix and masks other elements with 0, as shown in Figure 3(c). For a $\mathbb{R}^{r\times r}$ square matrix, $m$ diagonal blocks are retained. Each diagonal block is $\mathbb{R}\frac{r}{m}\times \frac{r}{m}$ . + +Clients with different communication bandwidths can select different numbers of diagonal blocks $m_{k}$ for the local homogeneous representation angle matrix $\mathcal{A}_k^t$ to produce + +![](images/c84bfffbab70fda61ed42c034b425b1470b148704e38315638cb09b615ad01b7.jpg) +(a) Element-wise Averaging: + +![](images/59df6510412a5ae89ad362ae640d332f0542b789028dabc68d185101466451d2.jpg) + +![](images/71511d1e88e9f2a642040eca27f286bd6ccd20ceb5622a5e8203b99292402089.jpg) + +![](images/17aa1faac55a7d2153dec4449e33905b7a8164effba59e275e53316e8f4576e0.jpg) + +![](images/5fecec498424cf6df704e8692a1a4b82025efb7bc10537ae41711e76599c0175.jpg) + +![](images/ca9ad1436886434ff7568a2ea7c22a3cdee033773fbe5a2a7fedb58beba2b13a.jpg) +(b) Element-wise Weighted Averaging: +Figure 4. Element-wise averaging (a) without or (b) with weights. + +![](images/cc35316f230016d817659cce7fd92ef307926761b45267fc3b76b89c11b0a96f.jpg) + +![](images/54e56c77265e9e9106e3a06525b8bd7cbd4b79e0d59fb7fa5b8c03389841a33d.jpg) + +![](images/ef010b041291657e0d7599d6aa9dc3a5ec8c46a4c67e8df9db057f4cfc30080a.jpg) + +![](images/d24e31fd7f90d36683d55659ac068aa66e3b2ada183b9e452c261e39d00e64b6.jpg) + +the sparsified representation angle square matrix $\tilde{\mathcal{A}}_k^t$ : + +$$ +\tilde {\mathcal {A}} _ {k} ^ {t} = \text {D i a g o n a l - S p a r s i f i c a t i o n} \left(\mathcal {A} _ {k} ^ {t}; m _ {k}\right). \tag {6} +$$ + +# 4.3. Element-wise Weighted Aggregation + +After receiving the sparsified local homogeneous representation square matrices from $K$ clients in the $t$ -th communication round, the server aggregates them according to the element positions with weighted averaging: + +$$ +\boldsymbol {\mathcal {A}} ^ {t} = \sum_ {k = 0} ^ {K - 1} \frac {n _ {k}}{n} \tilde {\boldsymbol {\mathcal {A}}} _ {k} ^ {t}, \tag {7} +$$ + +where $n_k$ is the data volume of the client $k$ , and $n$ is the total data volume of $K$ clients. + +This element-wise weighted aggregation rule provides two benefits. a) All clients upload the diagonal elements of local homogeneous representation angle square matrices. Using $\frac{n_k}{n}$ to weigh the diagonal elements can sufficiently reflect the overall data distribution across all clients, improving the generalization of the newly generated global square matrix. b) As the example given in Figure 4(a), for non-diagonal elements, if we directly average overlapped them by element, when one non-diagonal element is only provided by one client, then it is the value of this non-diagonal element. In this case, these non-overlapped non-diagonal elements may negatively affect the clients who do not provide for these elements. Instead, as shown in Figure 4(b), weighted averaging scales down such elements and then reduce negative effects while improving the generalization of the newly generated global square matrix. + +# 5. Experimental Evaluation + +To evaluate the performance of FedRAL ${}^{1}$ ,we compare it with 6 state-of-the-art existing approaches on 4 widely adopted benchmark datasets under 2 types of data heterogeneity settings. We implement FedRAL and baselines using pytorch and run them on 4 NVIDIA GeForce RTX 3090 GPUs (24GB memory). + +# Algorithm 1: FedRAL + +Input: $K$ , total number of participating clients; $T$ , total number of communication rounds; $\eta_{\omega}$ , learning rate of heterogeneous client models; $\eta_{A}$ , learning rate of the homogeneous representation angle square matrix. + +Randomly initialize the global representation angle square matrix $\mathcal{A}^0$ and heterogeneous client models $[f_0(\omega_0^0),\dots ,f_k(\omega_k^0),\dots ,f_{K - 1}(\omega_{K - 1}^0)]$ + +for $t = 1$ to $T$ do + +/* FL Server: */ + +Broadcast $\mathcal{A}^{t - 1}$ to participating clients; + +/* Each FL Client $k$ : */ + +// Representation Angle Matrix Learning + +for each $(x_{i},y_{i})\in D_{k}$ do + +$$ +\begin{array}{l} \mathcal {R} _ {i} = f _ {k} ^ {e x} \left(\boldsymbol {x} _ {i}; \omega_ {k} ^ {e x, t - 1}\right); \\ \tilde {\mathcal {R}} _ {i} = \mathcal {R} _ {i} \times \boldsymbol {A} ^ {t - 1}; \\ \hat {y} _ {i} = f ^ {h d} \left(\mathcal {R} _ {i} + \tilde {\mathcal {R}} _ {i}; \omega_ {k} ^ {h d, t - 1}\right); \\ f _ {k} \left(\omega_ {k} ^ {t}\right) \leftarrow f _ {k} \left(\omega_ {k} ^ {t - 1}\right) - \eta_ {\omega} \nabla \ell \left(\dot {y} _ {i}, y _ {i}\right); \\ \mathcal {A} _ {k} ^ {t} \leftarrow \mathcal {A} ^ {t - 1} - \eta_ {\mathcal {A}} \nabla \ell (\hat {y} _ {i}, y _ {i}); \\ \end{array} +$$ + +# end + +// Adaptive Diagonal Sparsification + +Obtain $\tilde{\mathcal{A}}_k^t$ using Eq. (6); + +Upload $\tilde{\mathcal{A}}_k^t$ to the FL server; + +/* FL Server: */ + +// Element-wise Weighted Aggregation + +Generate $\mathcal{A}^t$ by aggregation using Eq. (7); + +# 1 + +Return local heterogeneous client models + +$$ +\{f _ {0} \left(\omega_ {0} ^ {T - 1}\right), \dots , f _ {k} \left(\omega_ {k} ^ {T - 1}\right), \dots , f _ {K - 1} \left(\omega_ {K - 1} ^ {T - 1}\right) \}. +$$ + +# 5.1. Experiment Setup + +We first introduce datasets, base models, comparison baselines, evaluation metrics and hyperparameter settings. + +Datasets. We choose 4 image classification benchmark datasets commonly used in FL investigations. + +- Fashion-MNIST² [54] contains 60,000 training and 10,000 testing grayscale $28 \times 28$ 10-class images. +- CIFAR-10 [26] contains 50,000 training and 10,000 testing colourful $32 \times 32$ 10-class images. +- CIFAR-100 $^4$ [26] contains 50,000 training and 10,000 testing colourful $32 \times 32$ 100-class images. +- Tiny-ImageNet [10] contains 100,000 training and 10,000 testing colourful $64 \times 64$ 200-class images. + +Table 1. Heterogeneous CNNs for Fashion-MNIST. + +
LayerCNN-1CNN-2CNN-3CNN-4CNN-5
Conv15×5, 205×5, 205×5, 205×5, 205×5, 20
Maxpool12×22×22×22×22×2
Conv25×5, 205×5, 205×5, 205×5, 205×5, 20
Maxpool22×22×22×22×22×2
FC130020015010050
FC25050505050
FC31010101010
model size0.47 MB0.33 MB0.26 MB0.19 MB0.12 MB
+ +Note: kernel size: $5 \times 5$ , with 20 filters for convolution layers. + +Table 2. Heterogeneous CNNs for CIFAR-10 and CIFAR-100. + +
LayerCNN-1CNN-2CNN-3CNN-4CNN-5
Conv15×5, 165×5, 165×5, 165×5, 165×5, 16
Maxpool12×22×22×22×22×2
Conv25×5, 325×5, 165×5, 325×5, 325×5, 32
Maxpool22×22×22×22×22×2
FC1200020001000800500
FC2500500500500500
FC310/10010/10010/10010/10010/100
model size10.01 MB6.93 MB5.04 MB4.05 MB2.56 MB
+ +Note: kernel size: $5 \times 5$ , with 16 or 32 filters for convolution layers. + +Then, we partition them into non-IID modes in 2 ways: + +- Pathological. Following [48], we allocate $\{2, 2, 10, 20\}$ classes from $\{10, 10, 100, 200\}$ classes to each FL client. +- Practical. Following [43], we allocate $\{10, 10, 100, 200\}$ classes to each FL client and use a Dirichlet distribution function with a hyperparameter $\alpha = 0.4$ to generate diverse ratios of one class for different clients. + +Base Models. Following [62], we construct 5 heterogeneous CNN models $\{\mathrm{CNN - 1,}\dots,\mathrm{CNN - 5}\}$ with different channels of convolutional layers or dimensions of linear layers (Tables 1 and 2) and allocate them to 100 clients. The client $k$ is assigned CNN- $(k\% 5)$ . Note that the representation $\mathcal{R}$ dimension $r$ (the dimension of the penultimate linear layer) is $\{50,500,500\}$ for Fashion-MNIST, CIFAR-10 and CIFAR-100. Thus, the dimensions of the homogeneous representation angle square matrix $\mathcal{A}$ are $\{50\times 50,500\times 500,500\times 500\}$ . For Tiny-ImageNet, we evenly allocate ResNet- $\{18,34,50,101,152\}$ to 100 clients. + +Comparison Baselines. We compare FedRAL with 6 baselines, which are the latest representative MHFL works that support completely heterogeneous local models. + +- Standalone: each client solely trains its local heterogeneous model, without FL. +- Model Decoupling: LG-FedAvg [30]. +- Mutual Learning: FedKD [53], FedAPEN [43]. +- Knowledge Distillation without Public Dataset: FedProto [51], FedTGP [72]. + +Evaluation Metrics. We evaluate the performances of FedRAL and baselines with the following three metrics: +- Model Accuracy. We record each client's model accuracy (\%) and calculate their average in each communica + +tion round. We report the highest average accuracy. + +- Communication Cost. We track the average number of transmitted parameters between clients and the server in one communication round, and we record the total rounds consumed to reach the specified target accuracy. The product of the two is the total communication cost. +- Computational Overhead. We track the average computational FLOPs of clients in one communication round, and we record the total communication rounds consumed to reach the specified target accuracy. The product of the two is the total computational overhead. + +Hyperparameters. We use grid-search to choose the optimal hyperparameters for all algorithms. For general FL hyperparameters, we set the local training epoch $E$ as $\{1,10\}$ and batch size $B$ as $\{32,64,128,256,512\}$ , and we choose the SGD gradient optimizer with a learning rate of 0.01. To ensure the sufficient convergence of algorithms, we set the total communication rounds $T$ as 500. For the unique hyperparameter - the number $m$ of diagonal blocks for sparsifying the homogeneous representation angle square matrix in FedRAL, we set it to be the factor of the representation dimension $r$ , e.g., $m = \{1,2,\dots,50\}$ for Fashion-MNIST, and $m = \{1,2,\dots,500\}$ for CIFAR. + +# 5.2. Comparison Results Analysis + +Average Accuracy. Table 3 shows that FedRAL outperforms all baselines in terms of average test accuracy, demonstrating its effectiveness in enhancing local model generalization through the shared homogeneous representation angle square matrix. As a recent MHFL method, FedTGP achieves the second-best performance. We notice that FedAPEN fails to converge in some cases. This can be attributed to the fact that it uses the same coefficients to combine the outputs of the heterogeneous and homogeneous models for all data samples, failing to balance personalization and generalization under varying data features. + +Figure 5 presents how the average test accuracies of FedRAL and the state-of-the-art FedTGP baseline vary as communication rounds. We see that FedRAL converges to a higher accuracy with a faster speed than FedTGP, especially under more non-IID practical FL settings. This demonstrates the effectiveness of FedRAL in alleviating data heterogeneity while supporting the collaboration of heterogeneous client models. + +Communication Cost. Table 4 presents the communication costs of the state-of-the-art FedTGP baseline and FedRAL. We can observe that FedRAL consumes a lower single-round communication cost than FedTGP (column-3), since the sparse representation angle square matrix communicated in FedRAL has fewer parameters than the label-wise average representations transmitted in FedTGP. FedRAL also requires fewer communication rounds to reach the specified target accuracy (column-5), since the de + +Table 3. Comparison of average test accuracy (%) on 4 datasets under pathological and practical data heterogeneity settings. + +
MethodPathologicalPractical
Fashion-MNISTCIFAR-10CIFAR-100Tiny-ImageNetFashion-MNISTCIFAR-10CIFAR-100Tiny-ImageNet
Standalone99.03±0.5391.97±0.3753.04±0.2433.15±0.0273.39±0.8940.72±0.839.56±0.347.35±0.21
LG-FedAvg [30]98.87±0.4091.27±0.2545.83±0.6735.79±0.1971.97±0.2540.44±0.579.80±0.397.51±0.34
FedKD [53]58.37±0.0573.21±0.5637.21±0.9834.53±0.1645.57±0.4928.14±0.298.29±0.286.29±0.23
FedAPEN [43]44.20±0.58-------
FedProto [51]99.02±0.5092.49±0.2653.67±0.7936.24±0.2774.06±0.6340.20±0.119.27±0.717.63±0.29
FedTGP [72]99.09±0.6592.50±0.3153.20±0.7836.92±0.1073.64±0.6040.95±0.099.37±0.137.65±0.19
FedRAL99.54±0.1693.60±0.1958.70±0.0538.75±0.1378.58±0.1242.97±0.1011.70±0.029.47±0.13
+ +Note: “-” denotes failure to converge. Bold indicates the highest accuracy, and underline indicates the second highest accuracy. + +![](images/11085d94379bc0b80b3e79100512dfe7eaef036e3ab42fb8180a3fefeb8d6800.jpg) +Communication Round + +![](images/e8f965e28d3aebdf420070565b353f1f4c7b1ded62452b2bfcf186ed88a63763.jpg) +Communication Round +Figure 5. Average test accuracy $(\%)$ vs. number of rounds. + +![](images/9608db99494f565ae83e4cee0c50a3f661f0931e1cd8d52f833fc3d11a4530be.jpg) +Pathological, CIFAR-100 +Communication Round + +![](images/4d3466a93a94b8842746b7c9c8d6c881807eea9a38efb853a3a5e644a0753635.jpg) +Communication Round + +signed representation angle learning facilitates model convergence. Overall, FedRAL consumes lower communication costs for reaching the target accuracy than FedTGP (column-6), demonstrating its efficient communication. + +Computational Cost. Table 4 also shows the computational costs of the state-of-the-art FedTGP baseline and FedRAL. We can see that FedRAL consumes a lower single-round computational cost than FedTGP (column-4), since additionally training a homogeneous representation angle square matrix in FedRAL consumes lower FLOPs than additionally calculating label-wise average representations in FedTGP, except local model training in both methods. Since FedRAL requires fewer communication rounds to reach the specified target accuracy (column-5), its total computational costs are lower than FedTGP when reaching the specified target accuracy (column-7), substantiating its efficient computation. + +Privacy Evaluation. We conduct experiments to evaluate the privacy of FedRAL and baselines. Followed by Zhang et al. [70], considering a semi-honest FL scenario where the server is untrustworthy and may inversely infer the original data of comprised clients from transmit- + +Table 4. Communication and computational costs. + +
FL SettingMethodComm/C/R (KB)Comp/C/R (MB)RoundsComm/C (MB)Comp/C (GB)
Pathological CIFAR-10FedTGP3.9154.523021.1516.08
FedRAL1.5628.232010.315.54
Pathological CIFAR-100FedTGP19.5354.523546.7518.85
FedRAL9.7728.231441.373.97
Practical CIFAR-10FedTGP19.5354.873486.6418.65
FedRAL7.8128.581030.792.87
Practical CIFAR-100FedTGP195.3154.8736369.2419.45
FedRAL39.0628.581465.574.07
+ +Note: "Comm/C/R": a single client's communication cost (KB) in a communication round. "Comp/C/R": a single client's computational cost (MB) in a communication round. "Rounds": the communication rounds required to reach the specified $\{90, 50, 40, 9\}\%$ target accuracy (the near maximum accuracy that both the state-of-the-art FedTGP baseline and FedRAL can achieve). "Comm/C": a single client's communication cost (MB) required to reach the target accuracy. "Comm/C": a single client's computational cost (GB) required to reach the target accuracy. + +Table 5. PSNR under DLG attacks on CIFAR-100. + +
MethodLG-FedAvgFedKDFedAPENFedProtoFedTGPFedRAL
PSNR (dB)7.836.956.936.766.696.28
+ +Table 6. Average Accuracy $(\%)$ under high model heterogeneity. " -" denotes failure to converge. + +
MethodLG-FedAvgFedKDFedAPENFedProtoFedTGPFedRAL
Accuracy (%)45.6345.95-48.5648.6950.79
+ +ted information by launching Deep Leakage from Gradients (DLG) attack. We use the Peak Signal-to-Noise Ratio (PSNR), a common privacy measurement metric, to evaluate the privacy-protecting level. A lower PSNR denotes better privacy protection. The results in Table 5 validate that FedRAL can protect data privacy effectively. + +High Model Heterogeneity. To evaluate how FedRAL and baselines perform under high model heterogeneity, we conduct experiments CIFAR-100 with 100 clients and heterogeneous models including ResNet-{4, 6, 8, 10, 18, 34, 50, 101, 152}, CNN, and ViT. The results in Table 6 validate FedRAL's robustness to high model heterogeneity. + +# 5.3. Robustness Analysis + +This section analyzes the robustness of FedRAL to different non-IID degrees of two types. + +![](images/d12d1e619a820b815069035e927909bd39e9706d01da73fc3d9eb4e6fa38964f.jpg) +Figure 6. Robustness to pathological non-IIDness. + +![](images/e4b75de0008a8b0a5c0d32b7519f1a3ffbac953952a0c155c446f408b28c2445.jpg) + +![](images/fe088149e870e34a087efc4e4cac6ecc7faba08c13b9369c4041a4189fbbc257.jpg) +Figure 7. Robustness to practical non-IIDness. + +![](images/27800c7e49c2f1d56639e1c37049fff8ed88eacca6a25895456c0d4994c568ee.jpg) + +Robustness to pathological non-IIDness. We change the number of classes allocated to each client to obtain different non-IID degrees. For CIFAR-10, we allocate $\{2,4,6,8,10\}$ classes from 10 classes to each client. For CIFAR-100, we allocate $\{10,30,50,70,90,100\}$ classes out of 100 classes to each client. Figure 6 presents that FedRAL always performs higher accuracies than the state-of-the-art FedTGP baseline. As the number of classes assigned to one client increases, i.e., the non-IID degree decreases, model accuracy tends to drop since more IID data benefits generalization while compromising personalization of local models. + +Robustness to practical non-IIDness. We vary the hyperparameter $\alpha$ of the Dirichlet distribution function to construct diverse non-IID degrees. For both CIFAR-10 and CIFAR-100, we set $\alpha = \{0.1, 0.2, 0.3, 0.4, 0.5\}$ to evaluate the average model accuracies of FedRAL and the state-of-the-art FedTGP baseline. Figure 7 shows that FedRAL performs significantly higher average model accuracy than FedTGP, again demonstrating the robustness of FedRAL to non-IIDness. Similarly, as the value of $\alpha$ increases, i.e., the non-IID degree decreases, model accuracy tends to degrade also due to the above reason. + +# 5.4. Ablation Analysis + +FedRAL involves three contributions: 1) introducing representation angle matrix learning to share cross-client representation angle knowledge for enhancing model generalization, 2) devising adaptive diagonal sparsification to improve communication efficiency while keeping important representation angle knowledge, 3) designing ordinate-wise + +Table 7. Average accuracy $(\%)$ in ablation experiments. + +
CaseDiagonal SparsificationWeighted AggregationPathologicalPractical
CIFAR-10CIFAR-100CIFAR-10CIFAR-100
A✓✓93.6058.7042.9711.70
B✗✓91.6456.5841.3711.58
C✓✗93.4753.9440.529.24
+ +weighted aggregation to fuse cross-client representation angle square matrices. Contribution-1) is essential to implement FL collaboration across clients with heterogeneous models, so we conduct experiments to evaluate the necessity of the other two contributions with three cases: + +- Case-A, FedRAL with both adaptive diagonal sparsification and ordinate-wise weighted aggregation, i.e., the designed method. +- Case-B, FedRAL without adaptive diagonal sparsification but with ordinate-wise weighted aggregation, i.e., clients upload complete representation angle square matrices to the server. +- Case-C, FedRAL with adaptive diagonal sparsification but with naive ordinate-wise averaging aggregation, i.e., the aggregation rule (a) displayed in Figure 4, ignoring client data distribution distinctions. + +Table 7 shows that removing anyone degrades accuracy. Especially, replacing ordinate-wise weighted aggregation with naive averaging aggregation presents obvious accuracy degradations under most settings. These results reflect the effectiveness and necessity of the proposed two designs. + +# 6. Conclusion + +This work proposed FedRAL with three innovative designs: representation angle learning, adaptive diagonal sparsification, and element-wise weighted aggregation. The three designs enable enabling it to be a privacy-preserved, well-performed, communication and computation-efficient model-heterogeneous FL approach. Extensive experiments demonstrate its advantages in accuracy, communication and computation efficiency. + +# Acknowledgment + +Xiaoguang Liu is supported by the National Science Foundation of China under Grant 62272252 & 62272253, and the Fundamental Research Funds for the Central Universities. Han Yu is supported by the Ministry of Education, Singapore, under its Academic Research Fund Tier 1 (RG101/24); the RIE2025 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) (Award I2301E0026), administered by A\*STAR, as well as supported by Alibaba Group and NTU Singapore through Alibaba-NTU Global e-Sustainability CorpLab (ANGEL). + +# References + +[1] Jin-Hyun Ahn et al. Wireless federated distillation for distributed edge learning with heterogeneous data. In Proc. PIMRC, pages 1-6, Istanbul, Turkey, 2019. IEEE. 2 +[2] Jin-Hyun Ahn et al. Cooperative learning VIA federated distillation OVER fading channels. In Proc. ICASSP, pages 8856-8860, Barcelona, Spain, 2020. IEEE. 2 +[3] Samiul Alam et al. Fedrolex: Model-heterogeneous federated learning with rolling sub-model extraction. In Proc. NeurIPS, virtual, 2022.. 2 +[4] Sara Babakniya et al. Revisiting sparsity hunting in federated learning: Why does sparsity consensus matter? Transactions on Machine Learning Research, 1(1):1, 2023. +[5] Yun-Hin Chan, Rui Zhou, Running Zhao, Zhihan JIANG, and Edith C. H. Ngai. Internal cross-layer gradients for extending homogeneity to heterogeneity in federated learning. In Proc. ICLR, page 1, Vienna, Austria, 2024. OpenReview.net. 2 +[6] Hongyan Chang et al. Cronus: Robust and heterogeneous collaborative learning with black-box knowledge transfer. In Proc. NeurIPS Workshop, virtual, 2021.. 2 +[7] Jiangui Chen et al. Fedmatch: Federated learning over heterogeneous question answering data. In Proc. CIKM, pages 181-190, virtual, 2021. ACM. 2 +[8] Sijie Cheng et al. Fedgems: Federated learning of larger server models via selective knowledge fusion. CoRR, abs/2110.11027, 2021. 2 +[9] Yae Jee Cho et al. Heterogeneous ensemble knowledge transfer for training large models in federated learning. In Proc. IJCAI, pages 2881-2887, virtual, 2022. ijcai.org. 2 +[10] Patryk Chrabaszcz, Ilya Loshchilov, and Frank Hutter. A downsampled variant of imagenet as an alternative to the CIFAR datasets. CoRR, abs/1707.08819, 2017. 5 +[11] Liam Collins et al. Exploiting shared representations for personalized federated learning. In Proc. ICML, pages 2089-2099, virtual, 2021. PMLR. 1, 2 +[12] Enmao Diao. Heterofl: Computation and communication efficient federated learning for heterogeneous clients. In Proc. ICLR, page 1, Virtual Event, Austria, 2021. OpenReview.net. 2 +[13] Tao Fan, Hanlin Gu, et al. Ten challenging problems in federated foundation models. IEEE Transactions on Knowledge and Data Engineering, 2025. 1 +[14] Randy Goebel, Han Yu, Boi Faltings, Lixin Fan, and Zehui Xiong. Trustworthy Federated Learning. Springer, Cham, 2023. 1 +[15] Xuan Gong et al. Federated learning via input-output collaborative distillation. In Proc. AAAI, pages 22058-22066, Vancouver, Canada, 2024. AAAI Press. 2 +[16] Chaoyang He et al. Group knowledge transfer: Federated learning of large cnns at the edge. In Proc. NeurIPS, virtual, 2020. . 2 +[17] S. Horváth. FjORD: Fair and accurate federated learning under heterogeneous targets with ordered dropout. In Proc. NIPS, pages 12876-12889, Virtual, 2021. OpenReview.net. 2 + +[18] Edward J. Hu et al. Lora: Low-rank adaptation of large language models. In ICLR, page 1, Virtual, 2022. OpenReview.net. 2 +[19] Wenke Huang et al. Learn from others and be yourself in heterogeneous federated learning. In Proc. CVPR, pages 10133-10143, virtual, 2022. IEEE. 2 +[20] Wenke Huang et al. Few-shot model agnostic federated learning. In Proc. MM, pages 7309-7316, Lisboa, Portugal, 2022. ACM. +[21] Sohei Itahara et al. Distillation-based semi-supervised federated learning for communication-efficient collaborative training with non-iid private data. IEEE Trans. Mob. Comput., 22(1):191-205, 2023. 2 +[22] Jaehee Jang et al. Fedclassavg: Local representation learning for personalized federated learning on heterogeneous neural networks. In Proc. ICPP, pages 76:1-76:10, virtual, 2022. ACM, 2 +[23] Eunjoeong Jeong et al. Communication-efficient on-device machine learning: Federated distillation and augmentation under non-iid private data. In Proc. NeurIPS Workshop, virtual, 2018. . 2 +[24] Peter Kairouz et al. Advances and open problems in federated learning. Foundations and Trends in Machine Learning, 14(1-2):1-210, 2021. 1 +[25] Shivam Kalra et al. Decentralized federated learning through proxy model sharing. Nature communications, 14(1):2899, 2023. 2 +[26] Alex Krizhevsky et al. Learning multiple layers of features from tiny images. Toronto, ON, Canada, 2009. 5 +[27] Meng Lei, Qi Zhuang, Wu Lei, Du Xiaoyu, Li Zhaochuan, Cui Lizhen, and Meng Xiangxu. Improving global generalization and local personalization for federated learning. IEEE Transactions on Neural Networks and Learning Systems, 36, 2024. 1 +[28] Daliang Li and Junpu Wang. Fedmd: Heterogenous federated learning via model distillation. In Proc. NeurIPS Workshop, virtual, 2019. . 1, 2 +[29] Qinbin Li et al. Practical one-shot federated learning for cross-silo setting. In Proc. IJCAI, pages 1484-1490, virtual, 2021. ijcai.org. 2 +[30] Paul Pu Liang et al. Think locally, act globally: Federated learning with local and global representations. arXiv preprint arXiv:2001.01523, 1(1), 2020. 1, 2, 6, 7 +[31] Tao Lin et al. Ensemble distillation for robust model fusion in federated learning. In Proc. NeurIPS, virtual, 2020. . 1, 2 +[32] Chang Liu et al. Completely heterogeneous federated learning. CoRR, abs/2210.15865, 2022. 2 +[33] Qianyu Long, Christos Anagnostopoulos, Shameem Puthiya Parambath, and Daning Bi. Feddip: Federated learning with extreme dynamic pruning and incremental regularization. In Proc.ICDM, pages 1187-1192, Shanghai, China, 2023. IEEE. 2 +[34] Disha Makhija et al. Architecture agnostic federated learning for neural networks. In Proc. ICML, pages 14860-14870, virtual, 2022. PMLR. 2 +[35] Koji Matsuda et al. Fedme: Federated learning via model exchange. In Proc. SDM, pages 459-467, Alexandria, VA, USA, 2022. SIAM. 2 + +[36] Brendan McMahan et al. Communication-efficient learning of deep networks from decentralized data. In Proc. AISTATS, pages 1273-1282, Fort Lauderdale, FL, USA, 2017. PMLR. 1 +[37] Duy Phuong Nguyen et al. Enhancing heterogeneous federated learning with knowledge extraction and multi-model fusion. In Proc. SC Workshop, pages 36-43, Denver, CO, USA, 2023. ACM. 2 +[38] Jaehoon Oh et al. Fedbabu: Toward enhanced representation for federated image classification. In Proc. ICLR, virtual, 2022. OpenReview.net. 2 +[39] Zhengxiang Pan, Han Yu, Chunyan Miao, and Cyril Leung. Efficient collaborative crowdsourcing. In The 30th AAAI Conference on Artificial Intelligence (AAAI-16), pages 4248–4249, 2016. 1 +[40] Sejun Park et al. Towards understanding ensemble distillation in federated learning. In Proc. ICML, pages 27132-27187, Honolulu, Hawaii, USA, 2023. PMLR. 2 +[41] Krishna Pillutla et al. Federated learning with partial model personalization. In Proc. ICML, pages 17716-17758, virtual, 2022. PMLR. 2 +[42] Zhuang Qi, Sijin Zhou, Lei Meng, Han Yu, and Xiangxu Meng. Federated deconfounding and debiasing learning for out-of-distribution generalization. In The 34th International Joint Conference on Artificial Intelligence, (IJ-CAI'25), pages 1–9, 2025. 1 +[43] Zhen Qin et al. Fedapen: Personalized cross-silo federated learning with adaptability to statistical heterogeneity. In Proc. KDD, pages 1954-1964, Long Beach, CA, USA, 2023. ACM. 2, 6, 7 +[44] Zeju Qiu et al. Controlling text-to-image diffusion by orthogonal finetuning. In Proc. NeurIPS, New Orleans, LA, USA, 2023. 2, 3 +[45] Chao Ren, Han Yu, et al. Advances and open challenges in federated foundation models. IEEE Communications Surveys and Tutorials, 2025. 1 +[46] Felix Sattler et al. Fedaux: Leveraging unlabeled auxiliary data in federated learning. IEEE Trans. Neural Networks Learn. Syst., 1(1):1-13, 2021. 2 +[47] Felix Sattler et al. CFD: communication-efficient federated distillation via soft-label quantization and delta coding. IEEE Trans. Netw. Sci. Eng., 9(4):2025-2038, 2022. 2 +[48] Aviv Shamsian et al. Personalized federated learning using hypernetworks. In Proc. ICML, pages 9489-9502, virtual, 2021. PMLR. 6 +[49] Tao Shen et al. Federated mutual learning. CoRR, abs/2006.16765, 2020. 2 +[50] Alysa Ziying Tan et al. Towards personalized federated learning. IEEE Trans. Neural Networks Learn. Syst., 1(1): 1-17, 2022. 1 +[51] Yue Tan et al. Fedproto: Federated prototype learning across heterogeneous clients. In Proc. AAAI, pages 8432-8440, virtual, 2022. AAAI Press. 2, 6, 7 +[52] Jiaqi Wang et al. Towards personalized federated learning via heterogeneous model reassembly. In Proc. NeurIPS, page 13, New Orleans, Louisiana, USA, 2023. OpenReview.net. 2 + +[53] Chuhan Wu et al. Communication-efficient federated learning via knowledge distillation. Nature Communications, 13 (1):2032, 2022. 2, 6, 7 +[54] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms, 2017. 5 +[55] Qiang Yang, Lixin Fan, and Han Yu. Federated Learning: Privacy and Incentive. Springer, Cham, 2020. 1 +[56] Qiang Yang, Yang Liu, Yong Cheng, Yan Kang, Tianjian Chen, and Han Yu. Federated Learning. Springer, Cham, 2020. 1 +[57] Liping Yi, Gang Wang, and Xiaoguang Liu. QSFL: A two-level uplink communication optimization framework for federated learning. In Proc. ICML, pages 25501-25513. PMLR, 2022. 1 +[58] Liping Yi, Xiaorong Shi, Nan Wang, Gang Wang, Xiaoguang Liu, Zhuan Shi, and Han Yu. pfedkt: Personalized federated learning with dual knowledge transfer. Knowledge-Based Systems, 292:111633, 2024. 1 +[59] Liping Yi, Xiaorong Shi, Nan Wang, Jinsong Zhang, Gang Wang, and Xiaoguang Liu. Fedpe: Adaptive model pruning-expanding for federated learning on mobile devices. IEEE Transactions on Mobile Computing, pages 1-18, 2024. 1 +[60] Liping Yi, Gang Wang, Xiaofei Wang, and Xiaoguang Liu. Qsfl: Two-level communication-efficient federated learning on mobile edge devices. IEEE Transactions on Services Computing, pages 1-16, 2024. 1 +[61] Liping Yi, Han Yu, Zhuan Shi, Gang Wang, Xiaoguang Liu, Lizhen Cui, and Xiaoxiao Li. FedSSA: Semantic Similarity-based Aggregation for Efficient Model-Heterogeneous Personalized Federated Learning. In IJCAI, 2024. 1 +[62] Liping Yi et al. FedGH: Heterogeneous federated learning with generalized global header. In Proc. ACM MM, 2023. 2, 6 +[63] Liping Yi et al. Federated model heterogeneous matryoshka representation learning. In Proc. NeurIPS, Vancouver, Canada, 2024. +[64] Liping Yi et al. pFedAFM: Adaptive Feature Mixture for Data-Level Personalization in Heterogeneous Federated Learning on Mobile Edge Devices. In Proc. ICDE, pages 1981-1994, Hong Kong SAR, China, 2025. IEEE Computer Society. +[65] Liping Yi et al. pfedes: Generalized proxy feature extractor sharing for model heterogeneous personalized federated learning. In Proc. AAAI, pages 22146-22154, Philadelphia, PA, USA, 2025. AAAI Press. 1 +[66] Fuxun Yu et al. Fed2: Feature-aligned federated learning. In Proc. KDD, pages 2066-2074, virtual, 2021. ACM. 2 +[67] Han Yu, Siyuan Liu, Alex C Kot, Chunyan Miao, and Cyril Leung. Dynamic witness selection for trustworthy distributed cooperative sensing in cognitive radio networks. In Proceedings of the 13th IEEE International Conference on Communication Technology (ICCT'11), pages 1-6, 2011. 1 +[68] Sixing Yu et al. Resource-aware federated learning using knowledge extraction and multi-model fusion. CoRR, abs/2208.07978, 2022. 2 + +[69] Jie Zhang et al. Parameterized knowledge transfer for personalized federated learning. In Proc. NeurIPS, pages 10092-10104, virtual, 2021. OpenReview.net. 2 +[70] Jianqing Zhang et al. Fedcp: Separating feature information for personalized federated learning via conditional policy. In Proc. KDD, page 1, Long Beach, CA, USA, 2023. ACM. 7 +[71] Jie Zhang et al. Towards data-independent knowledge transfer in model-heterogeneous federated learning. IEEE Trans. Computers, 72(10):2888-2901, 2023. 2 +[72] Jianqing Zhang et al. Fedtgp: Trainable global prototypes with adaptive-margin-enhanced contrastive learning for data and model heterogeneity in federated learning. In Proc. AAAI, pages 16768-16776, Vancouver, Canada, 2024. 1. 6, 7 +[73] Lan Zhang et al. Fedzkt: Zero-shot knowledge transfer towards resource-constrained federated learning with heterogeneous on-device models. In Proc. ICDCS, pages 928-938, virtual, 2022. IEEE. 2 +[74] Zhilu Zhang and Mert R. Sabuncu. Generalized cross entropy loss for training deep neural networks with noisy labels. In Proc. NeurIPS, pages 8792-8802, Montreal, Canada, 2018. Curran Associates Inc. 4 +[75] Hangyu Zhu et al. Federated learning on non-iid data: A survey. Neurocomputing, 465:371-390, 2021. 1 +[76] Zhuangdi Zhu et al. Data-free knowledge distillation for heterogeneous federated learning. In Proc. ICML, pages 12878-12889, virtual, 2021. PMLR. 2 +[77] Zhuangdi Zhu et al. Resilient and communication efficient learning for heterogeneous federated systems. In Proc. ICML, pages 27504-27526, virtual, 2022. PMLR. 2 +[78] Qi Zhuang, Meng Lei, Chen Zitan, Hu Han, Lin Hui, and Meng Xiangxu. Cross-silo prototypical calibration for federated learning with non-iid data. In Proceedings of the 31st ACM International Conference on Multimedia, pages 3099-3107, 2023. 1 +[79] Qi Zhuang, Meng Lei, et al. Cross-silo feature space alignment for federated learning on clients with imbalanced data. In The 39th Annual AAAI Conference on Artificial Intelligence (AAAI-25), pages 19986-19994, 2025. 1 \ No newline at end of file diff --git a/2025/Federated Representation Angle Learning/images.zip b/2025/Federated Representation Angle Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1195bbbefd541386d07591e40fb9a6835a5b1664 --- /dev/null +++ b/2025/Federated Representation Angle Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3e720011ff9a2cd486bf1560b365a151da38d019d68cfd68cc9d3b1e880451d3 +size 536860 diff --git a/2025/Federated Representation Angle Learning/layout.json b/2025/Federated Representation Angle Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0123595ecf1aa5311ac7d16e9ea0d07c9a361a0c --- /dev/null +++ b/2025/Federated Representation Angle Learning/layout.json @@ -0,0 +1,12764 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 175, + 103, + 440, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 103, + 440, + 121 + ], + "spans": [ + { + "bbox": [ + 175, + 103, + 440, + 121 + ], + "type": "text", + "content": "Federated Representation Angle Learning*" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 142, + 578, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 142, + 578, + 228 + ], + "spans": [ + { + "bbox": [ + 64, + 142, + 578, + 228 + ], + "type": "text", + "content": "Liping Yi" + }, + { + "bbox": [ + 64, + 142, + 578, + 228 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 142, + 578, + 228 + ], + "type": "text", + "content": ", Han Yu" + }, + { + "bbox": [ + 64, + 142, + 578, + 228 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 64, + 142, + 578, + 228 + ], + "type": "text", + "content": ", Gang Wang" + }, + { + "bbox": [ + 64, + 142, + 578, + 228 + ], + "type": "inline_equation", + "content": "^{1,\\ast}" + }, + { + "bbox": [ + 64, + 142, + 578, + 228 + ], + "type": "text", + "content": ", Xiaoguang Liu" + }, + { + "bbox": [ + 64, + 142, + 578, + 228 + ], + "type": "inline_equation", + "content": "^{1,\\ast}" + }, + { + "bbox": [ + 64, + 142, + 578, + 228 + ], + "type": "text", + "content": ", Xiaoxiao Li" + }, + { + "bbox": [ + 64, + 142, + 578, + 228 + ], + "type": "inline_equation", + "content": "^{3,4}" + }, + { + "bbox": [ + 64, + 142, + 578, + 228 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 142, + 578, + 228 + ], + "type": "text", + "content": "College of Computer Science, TMCC, SysNet, DISSec, GTIISC, Nankai University, Tianjin, China \n" + }, + { + "bbox": [ + 64, + 142, + 578, + 228 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 64, + 142, + 578, + 228 + ], + "type": "text", + "content": "College of Computing and Data Science, Nanyang Technological University, Singapore \n" + }, + { + "bbox": [ + 64, + 142, + 578, + 228 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 64, + 142, + 578, + 228 + ], + "type": "text", + "content": "Department of Electrical and Computer Engineering, University of British Columbia, Vancouver, Canada \n" + }, + { + "bbox": [ + 64, + 142, + 578, + 228 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 64, + 142, + 578, + 228 + ], + "type": "text", + "content": "Vector Institute, Vancouver, Canada \n{yiliping, wgzwp, liuxg}@nbjl.nankai.edu.cn, han.yu@ntu.edu.sg, xiaoxiao.li@ece.ubc.ca" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "spans": [ + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 286, + 297, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 286, + 297, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 286, + 297, + 658 + ], + "type": "text", + "content": "Model-heterogeneous federated learning (MHFL) is a challenging FL paradigm designed to allow FL clients to train structurally heterogeneous models under the coordination of an FL server. Existing MHFL methods face significant limitations when it comes to transferring global knowledge to clients as a result of sharing only partial homogeneous model parameters or calculating distance loss, leading to inferior model generalization. To bridge this gap, we propose a novel model-heterogeneous Federated learning method with Representation Angle Learning (FedRAL). It consists of three innovative designs: (1) We first introduce representation angle learning into MHFL. Specifically, we embed a homogeneous square matrix into the local heterogeneous model of each client, which learns the angle information of local representations. These homogeneous representation angle square matrices are aggregated on the server to fuse representation angle knowledge shared by clients for enhancing the generalization of local representations. (2) As different clients might have heterogeneous system resources, we propose an adaptive diagonal sparsification strategy to reduce the numbers of the parameters of representation angle square matrices uploaded to the server, to improve FL communication efficiency. (3) To enable effective fusion of sparsified homogeneous local representation angle square matrices, we design an element-wise weighted aggregation approach. Experiments on 4 benchmark datasets under 2 types of non-IID divisions over 6 state-of-the-art baselines demonstrate that FedRAL achieves the best performance. It improves test accuracy, communication efficiency and computational efficiency by up to " + }, + { + "bbox": [ + 55, + 286, + 297, + 658 + ], + "type": "inline_equation", + "content": "5.03\\%" + }, + { + "bbox": [ + 55, + 286, + 297, + 658 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 286, + 297, + 658 + ], + "type": "inline_equation", + "content": "12.43\\times" + }, + { + "bbox": [ + 55, + 286, + 297, + 658 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 286, + 297, + 658 + ], + "type": "inline_equation", + "content": "6.49\\times" + }, + { + "bbox": [ + 55, + 286, + 297, + 658 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 315, + 255, + 394, + 267 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 255, + 394, + 267 + ], + "spans": [ + { + "bbox": [ + 315, + 255, + 394, + 267 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 277, + 555, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 555, + 384 + ], + "type": "text", + "content": "Federated learning (FL) [13, 14, 24, 36, 45, 55, 56] is a privacy-preserved distributed machine learning paradigm. Typically, an FL server is often involved in broadcast a global model to FL clients. The clients then further train the global model on local data to obtain local models. The server aggregates local model updates received from the clients to produce a new global model. In this way, client data are not exposed. Under this setting, the models trained by the clients must follow the same structure." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 388, + 556, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 556, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 556, + 615 + ], + "type": "text", + "content": "In practice, clients participating in FL often have non-independent and identically distributed (non-IID) data, a.k.a., data heterogeneity [27, 42, 50, 58, 75, 78, 79], the local models trained on such data are biased, so the global model obtained by aggregating them may not perform well on all client data. Besides, FL clients also have heterogeneous system configurations in terms of computing power and communication bandwidth, a.k.a., system heterogeneity [39, 57, 59, 60, 67]. Thus, requesting all clients to train the same model might not be viable for some weaker clients, while leaving stronger clients under-utilized. More critically, some companies or institutions joining FL as clients might face intellectual property issues regarding their proprietary local models. Thus, they might be reluctant to share their private heterogeneous models, a.k.a., model heterogeneity. These challenges have inspired the field of model-heterogeneous federated learning (MHFL) with preserved data and model privacy, which advocates avoiding directly sharing private local heterogeneous models [61-65]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 617, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 557, + 714 + ], + "type": "text", + "content": "Existing MHFL approaches mainly include: a) Modeldecoupling: each client model is decoupled into heterogeneous and homogeneous parts, with only homogeneous parts being shared with the server [11, 30]. b) Knowledge distillation: the server aggregates local representation or logits from different clients to generate global representation or logits, which is used to calculate distillation loss for local training [28, 31]. c) Mutual training: each client is" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 145, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 145, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 145, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1314" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "content": "assigned a shared homogeneous small model and trains it with the local heterogeneous model by mutual loss between them [49, 53]. Although these methods support model heterogeneity, their capability for transferring global knowledge from the server to clients is limited due to sharing partial model parameters or using distillation loss, thereby constraining the generalization of local models. Whereas, model representations often involve more sufficient semantic knowledge than partial model parameters or distillation loss, which are explored in this work for knowledge sharing." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 195, + 295, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 195, + 295, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 195, + 295, + 411 + ], + "type": "text", + "content": "A recent work [44] empirically verified that parameter fine-tuning approaches (e.g., LoRA [18]) cannot sufficiently retain semantic information of the pre-trained model. Whereas, fine-tuning the angles (i.e., gradient directions) of the pre-trained model can achieve this, thereby enhancing the generalization of the pre-trained model. The work further proposed the Orthogonal Fine-Tuning (OFT) method to fine-tune the parameter angles of the pre-trained models via an orthogonal matrix (Figure 1). In short, parameter angles involve sufficient semantic information that can enhance model generalization. Therefore, we can enhance FL model generalization through sharing parameter angles instead of model parameters. However, the FL server can not aggregate parameter angles from structure-heterogeneous client models. Considering representations also with sufficient semantic knowledge, we attempt to share homogeneous representation angles among clients to achieve generalization-enhanced model-heterogeneous FL." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 415, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 295, + 714 + ], + "type": "text", + "content": "To this end, we propose a novel model-heterogeneous Federated learning method with Representation Angel Learning (FedRAL) to enhance the transfer of generalizable global knowledge from the FL server to FL clients. It consists of three innovative designs: (1) Since representations are extracted high-level features from data, with sufficient semantic information, FedRAL embeds a homogeneous square matrix into the local heterogeneous model to learn the angle information of local extracted representations. The server aggregates local homogeneous representation angle square matrices from different clients to generate a global representation angle square matrix, thereby fusing cross-client representation angle knowledge. Then, the global representation angle square matrix with sufficient generalized representation angle information is used to enhance the generalization of local representations during local model training. Different from the strong assumption of the orthogonality of the matrix in OFT, the homogeneous representation angle square matrix is initialized randomly. (2) Considering the diverse system resources (e.g., communication bandwidth) of different clients, we design an adaptive diagonal sparsification strategy for the homogeneous representation angle square matrix that allows clients to upload diagonal blocks with sizes matching with communication capacity. (3) To effectively aggregate sparsified local" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 316, + 69, + 555, + 128 + ], + "blocks": [ + { + "bbox": [ + 316, + 69, + 555, + 128 + ], + "lines": [ + { + "bbox": [ + 316, + 69, + 555, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 69, + 555, + 128 + ], + "type": "image", + "image_path": "304dd7488f93e82b9d90d9fb9eac10b293fd18933f5a8073bd4c51e1ce16d80f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 357, + 133, + 511, + 144 + ], + "lines": [ + { + "bbox": [ + 357, + 133, + 511, + 144 + ], + "spans": [ + { + "bbox": [ + 357, + 133, + 511, + 144 + ], + "type": "text", + "content": "Figure 1. Orthogonal Fine-Tuning (OFT)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 152, + 553, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 152, + 553, + 176 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 553, + 176 + ], + "type": "text", + "content": "homogeneous representation angle square matrices on the server, element-wise weighted aggregation is designed." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 176, + 553, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 176, + 553, + 248 + ], + "spans": [ + { + "bbox": [ + 313, + 176, + 553, + 248 + ], + "type": "text", + "content": "Experiments on 4 benchmark datasets under 2 modes of non-IID divisions against 6 advanced MHFL methods demonstrate the superiority of FedRAL. Compared with the state-of-the-art baseline, FedRAL improves test accuracy by up to " + }, + { + "bbox": [ + 313, + 176, + 553, + 248 + ], + "type": "inline_equation", + "content": "5.03\\%" + }, + { + "bbox": [ + 313, + 176, + 553, + 248 + ], + "type": "text", + "content": ", while enhancing communication and computational efficiency by up to " + }, + { + "bbox": [ + 313, + 176, + 553, + 248 + ], + "type": "inline_equation", + "content": "12.43\\times" + }, + { + "bbox": [ + 313, + 176, + 553, + 248 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 176, + 553, + 248 + ], + "type": "inline_equation", + "content": "6.49\\times" + }, + { + "bbox": [ + 313, + 176, + 553, + 248 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 260, + 400, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 260, + 400, + 272 + ], + "spans": [ + { + "bbox": [ + 314, + 260, + 400, + 272 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 281, + 553, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 553, + 304 + ], + "type": "text", + "content": "Based on different approaches used for achieving MHFL, existing methods can be divided into 4 categories." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 305, + 554, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 305, + 554, + 400 + ], + "spans": [ + { + "bbox": [ + 313, + 305, + 554, + 400 + ], + "type": "text", + "content": "Heterogeneous Subnet. Some works [3-5, 12, 17, 33, 66, 77] assume that each client trains local heterogeneous subnets of the global model, and the server aggregates them according to parameter positions to re-construct a new global model. They allow each client to train personalized heterogeneous subnets to tackle data and system heterogeneity. However, uploading subnets exposes model parameters, thereby breaching model IP protection requirements." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 401, + 554, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 554, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 554, + 509 + ], + "type": "text", + "content": "Model Decoupling. These methods [7, 11, 22, 30, 32, 38, 41, 62] share partial local model parameters instead of the entire local model for global aggregation. Specifically, they decouple each client's local heterogeneous model into heterogeneous and homogeneous parts, and only share homogeneous parts for cross-client knowledge exchange. Nevertheless, sharing partial model parameters still violates model IP protection needs. Furthermore, the unshared remaining parameters might overfit due to local training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 510, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 554, + 665 + ], + "type": "text", + "content": "Knowledge Distillation. These methods [1, 2, 6, 8, 9, 15, 16, 19-21, 23, 28, 29, 31, 34, 37, 40, 46, 47, 51, 52, 68, 69, 71-73, 76] aggregate the local seen-class representations or logits extracted on local private data or a public dataset to generate the global representation or logits by class, which is used to calculate distillation loss with the local representations or logits. Although they protect local model structures, the distillation loss only transfers limited knowledge from the server to clients. Besides, sharing class information might be prohibited in some data privacy-sensitive applications. The public dataset used must follow similar data distributions with local private data, which is difficult to access due to data privacy." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": "Mutual Learning. These methods [25, 35, 43, 49, 53] add a global homogeneous small model shared by all clients, which is alternately trained with the local heterogeneous model by the mutual loss between the outputs of" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1315" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "text", + "content": "the two models for each sample. Although they hide local heterogeneous model structures, the mutual loss only transfers limited knowledge between the two models. The unstable mutual loss in early training rounds might result in both models not converging." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 296, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 296, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 296, + 264 + ], + "type": "text", + "content": "Existing MHFL methods generally transfer limited generalized knowledge from the server to clients, leading to model performance bottlenecks. FedRAL effectively tackles data, system and model heterogeneity simultaneously, allowing clients to share a global homogeneous representation angle square matrix for learning generalized representation angle information. It bridges this important gap by enhancing the generalization of local representations via utilizing the global homogeneous representation angle square matrix to fine-tune local representation angles, while preserving personalized semantic information." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 277, + 141, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 141, + 290 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 141, + 290 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 297, + 194, + 310 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 194, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 194, + 310 + ], + "type": "text", + "content": "3.1. Orthogonal Fine-Tuning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 315, + 296, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 315, + 296, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 296, + 553 + ], + "type": "text", + "content": "Fine-tuning the parameters of a pre-trained model has been demonstrated to compromise its semantic information, while fine-tuning its parameter angles (i.e., gradient directions) can sufficiently retain its original semantic information [44]. This enables the preservation of the generalized ability when adapting to downstream tasks. Hypersphere Energy (HE) measures the uniformity of neuron distribution on the unit hypersphere. A lower HE value is beneficial to keeping pre-trained model generalization ability. To minimize HE variance between the pre-trained model " + }, + { + "bbox": [ + 55, + 315, + 296, + 553 + ], + "type": "inline_equation", + "content": "W^0 \\in \\mathbb{R}^{u \\times v}" + }, + { + "bbox": [ + 55, + 315, + 296, + 553 + ], + "type": "text", + "content": " and the angle-fine-tuned model " + }, + { + "bbox": [ + 55, + 315, + 296, + 553 + ], + "type": "inline_equation", + "content": "W \\in \\mathbb{R}^{u \\times v}" + }, + { + "bbox": [ + 55, + 315, + 296, + 553 + ], + "type": "text", + "content": ", orthogonal finetuning (OFT) [44] fine-tunes the parameter angles of the pre-trained model through an orthogonal matrix " + }, + { + "bbox": [ + 55, + 315, + 296, + 553 + ], + "type": "inline_equation", + "content": "O \\in \\mathbb{R}^{u \\times u}" + }, + { + "bbox": [ + 55, + 315, + 296, + 553 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 315, + 296, + 553 + ], + "type": "inline_equation", + "content": "W = O \\times W^0" + }, + { + "bbox": [ + 55, + 315, + 296, + 553 + ], + "type": "text", + "content": ", as depicted in Figure 1. To ensure orthogonality, OFT utilizes Cayley parameterization to construct matrix " + }, + { + "bbox": [ + 55, + 315, + 296, + 553 + ], + "type": "inline_equation", + "content": "O = (I + Q)(I - Q)^{-1}" + }, + { + "bbox": [ + 55, + 315, + 296, + 553 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 315, + 296, + 553 + ], + "type": "inline_equation", + "content": "Q \\in \\mathbb{R}^{u \\times u}" + }, + { + "bbox": [ + 55, + 315, + 296, + 553 + ], + "type": "text", + "content": " is a skew-symmetric matrix satisfying " + }, + { + "bbox": [ + 55, + 315, + 296, + 553 + ], + "type": "inline_equation", + "content": "Q = -Q^T" + }, + { + "bbox": [ + 55, + 315, + 296, + 553 + ], + "type": "text", + "content": ". Due to inverse operations, updating the orthogonal matrix " + }, + { + "bbox": [ + 55, + 315, + 296, + 553 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 55, + 315, + 296, + 553 + ], + "type": "text", + "content": " during fine-tuning introduces high computational overhead, especially with a large dimension " + }, + { + "bbox": [ + 55, + 315, + 296, + 553 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 55, + 315, + 296, + 553 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 554, + 296, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 554, + 296, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 554, + 296, + 639 + ], + "type": "text", + "content": "Unlike OFT, we explore fine-tuning the angles of representations with sufficient semantic information. We relax the orthogonality assumption and avoid expensive inverse operations. Instead, we randomly initialize the homogeneous representation angle square matrix, which is embedded into the local heterogeneous model. They are trained in an end-to-end manner to improve computation efficiency." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 647, + 180, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 180, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 180, + 659 + ], + "type": "text", + "content": "3.2. Problem Formulation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "In FedRAL, a server coordinates " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": " clients with heterogeneous local models for collaborative training. Client " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "'s local model " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "f_{k}(\\omega_{k})" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": " consists of two parts: 1) a heterogeneous feature extractor " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "f_{k}^{ex}(\\omega_{k}^{ex})" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": ", and 2) a homogeneous predic" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 71, + 555, + 262 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 555, + 262 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 555, + 262 + ], + "type": "image", + "image_path": "0d11c5368c380c506206ab3935ca190e4afb5fa809eeb4bad025bc4c81090165.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 367, + 270, + 501, + 281 + ], + "lines": [ + { + "bbox": [ + 367, + 270, + 501, + 281 + ], + "spans": [ + { + "bbox": [ + 367, + 270, + 501, + 281 + ], + "type": "text", + "content": "Figure 2. The workflow of FedRAL." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 303, + 555, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 303, + 555, + 448 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 555, + 448 + ], + "type": "text", + "content": "tion head " + }, + { + "bbox": [ + 313, + 303, + 555, + 448 + ], + "type": "inline_equation", + "content": "f^{hd}(\\omega_k^{hd})" + }, + { + "bbox": [ + 313, + 303, + 555, + 448 + ], + "type": "text", + "content": " (all clients conduct the same prediction task), " + }, + { + "bbox": [ + 313, + 303, + 555, + 448 + ], + "type": "inline_equation", + "content": "f_{k}(\\omega_{k}) = f_{k}^{ex}(\\omega_{k}^{ex})\\circ f^{hd}(\\omega_{k}^{hd})" + }, + { + "bbox": [ + 313, + 303, + 555, + 448 + ], + "type": "text", + "content": ", as depicted in Figure 2. Each client trains a local homogeneous square matrix " + }, + { + "bbox": [ + 313, + 303, + 555, + 448 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_k" + }, + { + "bbox": [ + 313, + 303, + 555, + 448 + ], + "type": "text", + "content": " to learn the angle information of local representations extracted by the heterogeneous feature extractor. The server aggregates local homogeneous representation angle square matrices from different clients to generate a global representation angle square matrix " + }, + { + "bbox": [ + 313, + 303, + 555, + 448 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 303, + 555, + 448 + ], + "type": "text", + "content": " for cross-client representation angle knowledge fusion. Therefore, the training objective of FedRAL is to minimize the loss sum of all heterogeneous client models with the help of the global homogeneous representation angle square matrix " + }, + { + "bbox": [ + 313, + 303, + 555, + 448 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 303, + 555, + 448 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 345, + 457, + 554, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 457, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 345, + 457, + 554, + 491 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathcal {A}, \\left\\{\\omega_ {0}, \\dots , \\omega_ {K - 1} \\right\\}} \\sum_ {k = 0} ^ {K - 1} \\ell \\left(\\mathcal {F} _ {k} \\left(D _ {k}; \\left(\\omega_ {k} \\circ \\mathcal {A}\\right)\\right)\\right), \\tag {1}", + "image_path": "480bcd664b6ff2418142ed072c4c287d1c2a9cb04f7ada1dfa349dd6291f3896.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 500, + 554, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 500, + 554, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 500, + 554, + 536 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 500, + 554, + 536 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 313, + 500, + 554, + 536 + ], + "type": "text", + "content": " is loss. " + }, + { + "bbox": [ + 313, + 500, + 554, + 536 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_k(\\omega_k\\circ \\mathcal{A})" + }, + { + "bbox": [ + 313, + 500, + 554, + 536 + ], + "type": "text", + "content": " denotes the combination of the local heterogeneous model " + }, + { + "bbox": [ + 313, + 500, + 554, + 536 + ], + "type": "inline_equation", + "content": "\\omega_{k}" + }, + { + "bbox": [ + 313, + 500, + 554, + 536 + ], + "type": "text", + "content": " and the global representation angle square matrix " + }, + { + "bbox": [ + 313, + 500, + 554, + 536 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 500, + 554, + 536 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 500, + 554, + 536 + ], + "type": "inline_equation", + "content": "D_{k}" + }, + { + "bbox": [ + 313, + 500, + 554, + 536 + ], + "type": "text", + "content": " is the non-IID data of client " + }, + { + "bbox": [ + 313, + 500, + 554, + 536 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 500, + 554, + 536 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 548, + 501, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 501, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 501, + 562 + ], + "type": "text", + "content": "4. The Proposed FedRAL Approach" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 569, + 554, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 554, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 554, + 640 + ], + "type": "text", + "content": "FedRAL consists of three innovative designs: (1) homogeneous representation angle square matrix learning, (2) adaptive diagonal sparsification, and (3) element-wise weighted homogeneous square matrix aggregation. The workflow of FedRAL in a single communication round includes the following steps." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 641, + 553, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 553, + 666 + ], + "type": "text", + "content": "Overview. As shown in Figure 2, in the " + }, + { + "bbox": [ + 313, + 641, + 553, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 641, + 553, + 666 + ], + "type": "text", + "content": "-th communication round, FedRAL executes 4 steps:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 310, + 666, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 310, + 666, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 666, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 310, + 666, + 553, + 689 + ], + "type": "text", + "content": "① The server broadcasts the latest global representation angle square matrix " + }, + { + "bbox": [ + 310, + 666, + 553, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^{t-1}" + }, + { + "bbox": [ + 310, + 666, + 553, + 689 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 310, + 666, + 553, + 689 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 310, + 666, + 553, + 689 + ], + "type": "text", + "content": " participating clients." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 310, + 689, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 310, + 689, + 553, + 714 + ], + "type": "inline_equation", + "content": "②" + }, + { + "bbox": [ + 310, + 689, + 553, + 714 + ], + "type": "text", + "content": " A client " + }, + { + "bbox": [ + 310, + 689, + 553, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 310, + 689, + 553, + 714 + ], + "type": "text", + "content": " embeds the global representation angle square matrix " + }, + { + "bbox": [ + 310, + 689, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^{t - 1}" + }, + { + "bbox": [ + 310, + 689, + 553, + 714 + ], + "type": "text", + "content": " into its local model " + }, + { + "bbox": [ + 310, + 689, + 553, + 714 + ], + "type": "inline_equation", + "content": "f_{k}(\\omega_{k}^{t - 1})" + }, + { + "bbox": [ + 310, + 689, + 553, + 714 + ], + "type": "text", + "content": " to enhance" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1316" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 72, + 295, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 72, + 295, + 107 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 295, + 107 + ], + "type": "text", + "content": "the generalization of local representations, while learning the personalized angle information of local representations via end-to-end training." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 51, + 108, + 295, + 209 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 51, + 108, + 295, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 108, + 295, + 158 + ], + "spans": [ + { + "bbox": [ + 51, + 108, + 295, + 158 + ], + "type": "text", + "content": "③ The local homogeneous representation angle square matrix " + }, + { + "bbox": [ + 51, + 108, + 295, + 158 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_k^t" + }, + { + "bbox": [ + 51, + 108, + 295, + 158 + ], + "type": "text", + "content": " after local training is sparsified into " + }, + { + "bbox": [ + 51, + 108, + 295, + 158 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{A}}_k^t" + }, + { + "bbox": [ + 51, + 108, + 295, + 158 + ], + "type": "text", + "content": " via the proposed adaptive diagonal sparsification strategy before being uploaded to the server." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 51, + 159, + 295, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 159, + 295, + 209 + ], + "spans": [ + { + "bbox": [ + 51, + 159, + 295, + 209 + ], + "type": "text", + "content": "(4) The server aggregates sparsified local homogeneous representation angle square matrices " + }, + { + "bbox": [ + 51, + 159, + 295, + 209 + ], + "type": "inline_equation", + "content": "\\{\\tilde{\\mathcal{A}}_0^t,\\dots ,\\tilde{\\mathcal{A}}_{K - 1}^t\\}" + }, + { + "bbox": [ + 51, + 159, + 295, + 209 + ], + "type": "text", + "content": " via the proposed element-wise weighted aggregation to generate a new global representation angle square matrix " + }, + { + "bbox": [ + 51, + 159, + 295, + 209 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^t" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 209, + 295, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 209, + 295, + 256 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 295, + 256 + ], + "type": "text", + "content": "Inference. These steps are iteratively executed until heterogeneous client models " + }, + { + "bbox": [ + 55, + 209, + 295, + 256 + ], + "type": "inline_equation", + "content": "\\{f_0(\\omega_0),\\dots ,f_{K - 1}(\\omega_{K - 1})\\}" + }, + { + "bbox": [ + 55, + 209, + 295, + 256 + ], + "type": "text", + "content": " converge, which are used for inference. Algorithm 1 describes the detailed workflow of FedRAL." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 264, + 226, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 264, + 226, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 226, + 277 + ], + "type": "text", + "content": "4.1. Representation Angle Learning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 281, + 295, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 295, + 318 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 295, + 318 + ], + "type": "text", + "content": "During the " + }, + { + "bbox": [ + 55, + 281, + 295, + 318 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 281, + 295, + 318 + ], + "type": "text", + "content": "-th training round, a client " + }, + { + "bbox": [ + 55, + 281, + 295, + 318 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 281, + 295, + 318 + ], + "type": "text", + "content": " uses its local heterogeneous feature extractor " + }, + { + "bbox": [ + 55, + 281, + 295, + 318 + ], + "type": "inline_equation", + "content": "f_{k}^{ex}\\left(\\omega_{k}^{ex,t - 1}\\right)" + }, + { + "bbox": [ + 55, + 281, + 295, + 318 + ], + "type": "text", + "content": " to extract the personalized representation " + }, + { + "bbox": [ + 55, + 281, + 295, + 318 + ], + "type": "inline_equation", + "content": "\\mathcal{R} \\in \\mathbb{R}^r" + }, + { + "bbox": [ + 55, + 281, + 295, + 318 + ], + "type": "text", + "content": " of the input sample " + }, + { + "bbox": [ + 55, + 281, + 295, + 318 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 55, + 281, + 295, + 318 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 129, + 326, + 294, + 341 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 326, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 129, + 326, + 294, + 341 + ], + "type": "interline_equation", + "content": "\\mathcal {R} = f _ {k} ^ {e x} \\left(\\boldsymbol {x}; \\omega_ {k} ^ {e x, t - 1}\\right). \\tag {2}", + "image_path": "ebfe8e724448f4a724f5d181b5c13d10941ae47c1932090d7a8696a97248d6c3.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 349, + 296, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 349, + 296, + 397 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 296, + 397 + ], + "type": "text", + "content": "Then, the extracted personalized representation " + }, + { + "bbox": [ + 55, + 349, + 296, + 397 + ], + "type": "inline_equation", + "content": "\\mathfrak{R} \\in \\mathbb{R}^r" + }, + { + "bbox": [ + 55, + 349, + 296, + 397 + ], + "type": "text", + "content": " is multiplied by the received global representation angle square matrix " + }, + { + "bbox": [ + 55, + 349, + 296, + 397 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^{t-1} \\in \\mathbb{R}^{r \\times r}" + }, + { + "bbox": [ + 55, + 349, + 296, + 397 + ], + "type": "text", + "content": " to produce the representation " + }, + { + "bbox": [ + 55, + 349, + 296, + 397 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{R}} \\in \\mathbb{R}^r" + }, + { + "bbox": [ + 55, + 349, + 296, + 397 + ], + "type": "text", + "content": " with enhanced generalization:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 140, + 406, + 294, + 419 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 406, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 140, + 406, + 294, + 419 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathcal {R}} = \\mathcal {R} \\times \\mathcal {A} ^ {t - 1}. \\tag {3}", + "image_path": "df09b8b4627bced82ae846345177651c53d32ab847c434e4f011ff7a0d6a6555.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 429, + 295, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 429, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 429, + 295, + 453 + ], + "type": "text", + "content": "Then, " + }, + { + "bbox": [ + 55, + 429, + 295, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{R} \\in \\mathbb{R}^r" + }, + { + "bbox": [ + 55, + 429, + 295, + 453 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 429, + 295, + 453 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{R}} \\in \\mathbb{R}^r" + }, + { + "bbox": [ + 55, + 429, + 295, + 453 + ], + "type": "text", + "content": " are summed and fed into the prediction head to make a prediction:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 119, + 460, + 294, + 476 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 460, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 119, + 460, + 294, + 476 + ], + "type": "interline_equation", + "content": "\\hat {y} = f ^ {h d} \\left(\\mathcal {R} + \\tilde {\\mathcal {R}}; \\omega_ {k} ^ {h d, t - 1}\\right). \\tag {4}", + "image_path": "5f09b40541427e5f518c65b6c3371cfca4782bb4ac7e711e81284073bb6f8782.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 485, + 295, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 295, + 546 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 295, + 546 + ], + "type": "text", + "content": "The loss " + }, + { + "bbox": [ + 55, + 485, + 295, + 546 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 55, + 485, + 295, + 546 + ], + "type": "text", + "content": " (e.g., Cross-Entropy loss [74]) between the prediction " + }, + { + "bbox": [ + 55, + 485, + 295, + 546 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 55, + 485, + 295, + 546 + ], + "type": "text", + "content": " and the ground-truth label " + }, + { + "bbox": [ + 55, + 485, + 295, + 546 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 485, + 295, + 546 + ], + "type": "text", + "content": " is used to update the heterogeneous client model " + }, + { + "bbox": [ + 55, + 485, + 295, + 546 + ], + "type": "inline_equation", + "content": "f_{k}(\\omega_{k}^{t - 1})" + }, + { + "bbox": [ + 55, + 485, + 295, + 546 + ], + "type": "text", + "content": " and the homogeneous representation angle square matrix " + }, + { + "bbox": [ + 55, + 485, + 295, + 546 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^{t - 1}" + }, + { + "bbox": [ + 55, + 485, + 295, + 546 + ], + "type": "text", + "content": " simultaneously via end-to-end training:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 102, + 554, + 295, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 554, + 295, + 574 + ], + "spans": [ + { + "bbox": [ + 102, + 554, + 295, + 574 + ], + "type": "interline_equation", + "content": "f _ {k} \\left(\\omega_ {k} ^ {t}\\right) \\leftarrow f _ {k} \\left(\\omega_ {k} ^ {t - 1}\\right) - \\eta_ {\\omega} \\nabla \\ell (\\hat {y}, y), \\tag {5}", + "image_path": "b9e037fcfa430bd3ce36fe35e9f64e140fca580c98ffcf44301f1af243147a9c.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 118, + 571, + 235, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 571, + 235, + 586 + ], + "spans": [ + { + "bbox": [ + 118, + 571, + 235, + 586 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mathcal {A}} _ {k} ^ {t} \\leftarrow \\boldsymbol {\\mathcal {A}} ^ {t - 1} - \\eta_ {\\boldsymbol {\\mathcal {A}}} \\nabla \\ell (\\hat {y}, y),", + "image_path": "ecd77e712ced2820121336f0d242d7f8366be1eb1b481bb4f3a71364af391a41.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 594, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 640 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 594, + 295, + 640 + ], + "type": "inline_equation", + "content": "\\eta_{\\omega}" + }, + { + "bbox": [ + 55, + 594, + 295, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 594, + 295, + 640 + ], + "type": "inline_equation", + "content": "\\eta_{\\mathcal{A}}" + }, + { + "bbox": [ + 55, + 594, + 295, + 640 + ], + "type": "text", + "content": " are the learning rates of the local heterogeneous model and the homogeneous representation angle square matrix. We set " + }, + { + "bbox": [ + 55, + 594, + 295, + 640 + ], + "type": "inline_equation", + "content": "\\eta_{\\omega} = \\eta_{\\mathcal{A}}" + }, + { + "bbox": [ + 55, + 594, + 295, + 640 + ], + "type": "text", + "content": " by default as they are updated at the same time." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": "The shared global homogeneous representation angle square matrix serves three functions during local training: a) it transfers globally generalizable representation angle knowledge from the server to the client by producing the generalization-enhanced representation; b) it learns the personalized representation angle knowledge from the" + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 345, + 69, + 524, + 245 + ], + "blocks": [ + { + "bbox": [ + 345, + 69, + 524, + 245 + ], + "lines": [ + { + "bbox": [ + 345, + 69, + 524, + 245 + ], + "spans": [ + { + "bbox": [ + 345, + 69, + 524, + 245 + ], + "type": "image", + "image_path": "2e5e6187e943bfa0893d88b814a31a8597dee93330407dfd89acffe1353a16f0.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 253, + 542, + 265 + ], + "lines": [ + { + "bbox": [ + 324, + 253, + 542, + 265 + ], + "spans": [ + { + "bbox": [ + 324, + 253, + 542, + 265 + ], + "type": "text", + "content": "Figure 3. An illustrative example of diagonal sparsification." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 277, + 553, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 553, + 338 + ], + "type": "text", + "content": "extracted personalized representations, thereby transferring local personalized representation angle knowledge to the server; and c) the summed representation includes both generalized and personalized knowledge, thereby improving model expressiveness while addressing non-IID problem." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 346, + 492, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 492, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 492, + 360 + ], + "type": "text", + "content": "4.2. Adaptive Diagonal Sparsification" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 364, + 553, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 364, + 553, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 364, + 553, + 412 + ], + "type": "text", + "content": "To improve communication efficiency while maintaining model performance, we design an adaptive diagonal sparsification strategy to sparsify the local representation angle square matrices before uploading them to the FL server." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 413, + 554, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 554, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 554, + 568 + ], + "type": "text", + "content": "Each element in " + }, + { + "bbox": [ + 313, + 413, + 554, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{R} \\in \\mathbb{R}^r" + }, + { + "bbox": [ + 313, + 413, + 554, + 568 + ], + "type": "text", + "content": " corresponds to the element on the diagonal of " + }, + { + "bbox": [ + 313, + 413, + 554, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{A} \\in \\mathbb{R}^{r \\times r}" + }, + { + "bbox": [ + 313, + 413, + 554, + 568 + ], + "type": "text", + "content": ". If we only retain the diagonal elements of " + }, + { + "bbox": [ + 313, + 413, + 554, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 413, + 554, + 568 + ], + "type": "text", + "content": " and set other elements to be 0, the angle of each element in the representation is directly adjusted by the corresponding element in the square matrix diagonal, as depicted in Figure 3(a). When all elements of the square matrix are non-zero, non-diagonal elements might disturb representation angle adjustment due to multiplication with non-corresponding representation elements. As Figure 3(b) shows, the enhanced representation processed by the entire representation angle square matrix is significantly different from the one processed by the sparse square matrix with only diagonal elements." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 569, + 554, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 554, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 554, + 676 + ], + "type": "text", + "content": "Since adjacent elements in the representation are semantically correlated, retaining partial elements wrapped with the diagonal in the representation angle square matrix can enhance the angle correlation of adjacent representation elements. Therefore, the proposed sparsification strategy retains the elements from the diagonal blocks of the square matrix and masks other elements with 0, as shown in Figure 3(c). For a " + }, + { + "bbox": [ + 313, + 569, + 554, + 676 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{r\\times r}" + }, + { + "bbox": [ + 313, + 569, + 554, + 676 + ], + "type": "text", + "content": " square matrix, " + }, + { + "bbox": [ + 313, + 569, + 554, + 676 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 569, + 554, + 676 + ], + "type": "text", + "content": " diagonal blocks are retained. Each diagonal block is " + }, + { + "bbox": [ + 313, + 569, + 554, + 676 + ], + "type": "inline_equation", + "content": "\\mathbb{R}\\frac{r}{m}\\times \\frac{r}{m}" + }, + { + "bbox": [ + 313, + 569, + 554, + 676 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": "Clients with different communication bandwidths can select different numbers of diagonal blocks " + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "inline_equation", + "content": "m_{k}" + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": " for the local homogeneous representation angle matrix " + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_k^t" + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": " to produce" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1317" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 79, + 103, + 131 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 134, + 78 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 134, + 78 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 134, + 78 + ], + "type": "text", + "content": "(a) Element-wise Averaging:" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 79, + 103, + 131 + ], + "lines": [ + { + "bbox": [ + 58, + 79, + 103, + 131 + ], + "spans": [ + { + "bbox": [ + 58, + 79, + 103, + 131 + ], + "type": "image", + "image_path": "c84bfffbab70fda61ed42c034b425b1470b148704e38315638cb09b615ad01b7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 105, + 79, + 167, + 131 + ], + "blocks": [ + { + "bbox": [ + 105, + 79, + 167, + 131 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 167, + 131 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 167, + 131 + ], + "type": "image", + "image_path": "59df6510412a5ae89ad362ae640d332f0542b789028dabc68d185101466451d2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 169, + 79, + 231, + 132 + ], + "blocks": [ + { + "bbox": [ + 169, + 79, + 231, + 132 + ], + "lines": [ + { + "bbox": [ + 169, + 79, + 231, + 132 + ], + "spans": [ + { + "bbox": [ + 169, + 79, + 231, + 132 + ], + "type": "image", + "image_path": "71511d1e88e9f2a642040eca27f286bd6ccd20ceb5622a5e8203b99292402089.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 233, + 79, + 294, + 131 + ], + "blocks": [ + { + "bbox": [ + 233, + 79, + 294, + 131 + ], + "lines": [ + { + "bbox": [ + 233, + 79, + 294, + 131 + ], + "spans": [ + { + "bbox": [ + 233, + 79, + 294, + 131 + ], + "type": "image", + "image_path": "17aa1faac55a7d2153dec4449e33905b7a8164effba59e275e53316e8f4576e0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 58, + 142, + 102, + 186 + ], + "blocks": [ + { + "bbox": [ + 58, + 142, + 102, + 186 + ], + "lines": [ + { + "bbox": [ + 58, + 142, + 102, + 186 + ], + "spans": [ + { + "bbox": [ + 58, + 142, + 102, + 186 + ], + "type": "image", + "image_path": "5fecec498424cf6df704e8692a1a4b82025efb7bc10537ae41711e76599c0175.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 103, + 141, + 167, + 194 + ], + "blocks": [ + { + "bbox": [ + 59, + 134, + 162, + 141 + ], + "lines": [ + { + "bbox": [ + 59, + 134, + 162, + 141 + ], + "spans": [ + { + "bbox": [ + 59, + 134, + 162, + 141 + ], + "type": "text", + "content": "(b) Element-wise Weighted Averaging:" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 103, + 141, + 167, + 194 + ], + "lines": [ + { + "bbox": [ + 103, + 141, + 167, + 194 + ], + "spans": [ + { + "bbox": [ + 103, + 141, + 167, + 194 + ], + "type": "image", + "image_path": "ca9ad1436886434ff7568a2ea7c22a3cdee033773fbe5a2a7fedb58beba2b13a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 204, + 294, + 216 + ], + "lines": [ + { + "bbox": [ + 55, + 204, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 294, + 216 + ], + "type": "text", + "content": "Figure 4. Element-wise averaging (a) without or (b) with weights." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 168, + 141, + 187, + 167 + ], + "blocks": [ + { + "bbox": [ + 168, + 141, + 187, + 167 + ], + "lines": [ + { + "bbox": [ + 168, + 141, + 187, + 167 + ], + "spans": [ + { + "bbox": [ + 168, + 141, + 187, + 167 + ], + "type": "image", + "image_path": "cc35316f230016d817659cce7fd92ef307926761b45267fc3b76b89c11b0a96f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 188, + 141, + 230, + 194 + ], + "blocks": [ + { + "bbox": [ + 188, + 141, + 230, + 194 + ], + "lines": [ + { + "bbox": [ + 188, + 141, + 230, + 194 + ], + "spans": [ + { + "bbox": [ + 188, + 141, + 230, + 194 + ], + "type": "image", + "image_path": "54e56c77265e9e9106e3a06525b8bd7cbd4b79e0d59fb7fa5b8c03389841a33d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 230, + 141, + 249, + 167 + ], + "blocks": [ + { + "bbox": [ + 230, + 141, + 249, + 167 + ], + "lines": [ + { + "bbox": [ + 230, + 141, + 249, + 167 + ], + "spans": [ + { + "bbox": [ + 230, + 141, + 249, + 167 + ], + "type": "image", + "image_path": "ef010b041291657e0d7599d6aa9dc3a5ec8c46a4c67e8df9db057f4cfc30080a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 250, + 141, + 294, + 194 + ], + "blocks": [ + { + "bbox": [ + 250, + 141, + 294, + 194 + ], + "lines": [ + { + "bbox": [ + 250, + 141, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 250, + 141, + 294, + 194 + ], + "type": "image", + "image_path": "d24e31fd7f90d36683d55659ac068aa66e3b2ada183b9e452c261e39d00e64b6.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 222, + 271, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 222, + 271, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 222, + 271, + 239 + ], + "type": "text", + "content": "the sparsified representation angle square matrix " + }, + { + "bbox": [ + 55, + 222, + 271, + 239 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{A}}_k^t" + }, + { + "bbox": [ + 55, + 222, + 271, + 239 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 91, + 245, + 295, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 245, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 91, + 245, + 295, + 262 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathcal {A}} _ {k} ^ {t} = \\text {D i a g o n a l - S p a r s i f i c a t i o n} \\left(\\mathcal {A} _ {k} ^ {t}; m _ {k}\\right). \\tag {6}", + "image_path": "7f089bee62611afb3e5228ed9542da2cd19fd6fb318da78127ab2299efd169b4.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 268, + 248, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 268, + 248, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 268, + 248, + 281 + ], + "type": "text", + "content": "4.3. Element-wise Weighted Aggregation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 285, + 296, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 296, + 334 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 296, + 334 + ], + "type": "text", + "content": "After receiving the sparsified local homogeneous representation square matrices from " + }, + { + "bbox": [ + 55, + 285, + 296, + 334 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 285, + 296, + 334 + ], + "type": "text", + "content": " clients in the " + }, + { + "bbox": [ + 55, + 285, + 296, + 334 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 285, + 296, + 334 + ], + "type": "text", + "content": "-th communication round, the server aggregates them according to the element positions with weighted averaging:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 136, + 340, + 295, + 373 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 340, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 136, + 340, + 295, + 373 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mathcal {A}} ^ {t} = \\sum_ {k = 0} ^ {K - 1} \\frac {n _ {k}}{n} \\tilde {\\boldsymbol {\\mathcal {A}}} _ {k} ^ {t}, \\tag {7}", + "image_path": "6898ce01deab73d50826eec7eb0c6b9a5708364432f963b453fce20753295c1c.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 379, + 295, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 379, + 295, + 402 + ], + "spans": [ + { + "bbox": [ + 55, + 379, + 295, + 402 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 379, + 295, + 402 + ], + "type": "inline_equation", + "content": "n_k" + }, + { + "bbox": [ + 55, + 379, + 295, + 402 + ], + "type": "text", + "content": " is the data volume of the client " + }, + { + "bbox": [ + 55, + 379, + 295, + 402 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 379, + 295, + 402 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 379, + 295, + 402 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 379, + 295, + 402 + ], + "type": "text", + "content": " is the total data volume of " + }, + { + "bbox": [ + 55, + 379, + 295, + 402 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 379, + 295, + 402 + ], + "type": "text", + "content": " clients." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 403, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 296, + 594 + ], + "type": "text", + "content": "This element-wise weighted aggregation rule provides two benefits. a) All clients upload the diagonal elements of local homogeneous representation angle square matrices. Using " + }, + { + "bbox": [ + 55, + 403, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\frac{n_k}{n}" + }, + { + "bbox": [ + 55, + 403, + 296, + 594 + ], + "type": "text", + "content": " to weigh the diagonal elements can sufficiently reflect the overall data distribution across all clients, improving the generalization of the newly generated global square matrix. b) As the example given in Figure 4(a), for non-diagonal elements, if we directly average overlapped them by element, when one non-diagonal element is only provided by one client, then it is the value of this non-diagonal element. In this case, these non-overlapped non-diagonal elements may negatively affect the clients who do not provide for these elements. Instead, as shown in Figure 4(b), weighted averaging scales down such elements and then reduce negative effects while improving the generalization of the newly generated global square matrix." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 604, + 199, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 199, + 618 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 199, + 618 + ], + "type": "text", + "content": "5. Experimental Evaluation" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 624, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 296, + 696 + ], + "type": "text", + "content": "To evaluate the performance of FedRAL " + }, + { + "bbox": [ + 55, + 624, + 296, + 696 + ], + "type": "inline_equation", + "content": "{}^{1}" + }, + { + "bbox": [ + 55, + 624, + 296, + 696 + ], + "type": "text", + "content": " ,we compare it with 6 state-of-the-art existing approaches on 4 widely adopted benchmark datasets under 2 types of data heterogeneity settings. We implement FedRAL and baselines using pytorch and run them on 4 NVIDIA GeForce RTX 3090 GPUs (24GB memory)." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 75, + 417, + 86 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 75, + 417, + 86 + ], + "spans": [ + { + "bbox": [ + 321, + 75, + 417, + 86 + ], + "type": "text", + "content": "Algorithm 1: FedRAL" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 325, + 89, + 538, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 89, + 538, + 148 + ], + "spans": [ + { + "bbox": [ + 325, + 89, + 538, + 148 + ], + "type": "text", + "content": "Input: " + }, + { + "bbox": [ + 325, + 89, + 538, + 148 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 325, + 89, + 538, + 148 + ], + "type": "text", + "content": ", total number of participating clients; " + }, + { + "bbox": [ + 325, + 89, + 538, + 148 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 325, + 89, + 538, + 148 + ], + "type": "text", + "content": ", total number of communication rounds; " + }, + { + "bbox": [ + 325, + 89, + 538, + 148 + ], + "type": "inline_equation", + "content": "\\eta_{\\omega}" + }, + { + "bbox": [ + 325, + 89, + 538, + 148 + ], + "type": "text", + "content": ", learning rate of heterogeneous client models; " + }, + { + "bbox": [ + 325, + 89, + 538, + 148 + ], + "type": "inline_equation", + "content": "\\eta_{A}" + }, + { + "bbox": [ + 325, + 89, + 538, + 148 + ], + "type": "text", + "content": ", learning rate of the homogeneous representation angle square matrix." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 327, + 148, + 538, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 148, + 538, + 185 + ], + "spans": [ + { + "bbox": [ + 327, + 148, + 538, + 185 + ], + "type": "text", + "content": "Randomly initialize the global representation angle square matrix " + }, + { + "bbox": [ + 327, + 148, + 538, + 185 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^0" + }, + { + "bbox": [ + 327, + 148, + 538, + 185 + ], + "type": "text", + "content": " and heterogeneous client models " + }, + { + "bbox": [ + 327, + 148, + 538, + 185 + ], + "type": "inline_equation", + "content": "[f_0(\\omega_0^0),\\dots ,f_k(\\omega_k^0),\\dots ,f_{K - 1}(\\omega_{K - 1}^0)]" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 328, + 186, + 399, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 186, + 399, + 194 + ], + "spans": [ + { + "bbox": [ + 328, + 186, + 399, + 194 + ], + "type": "text", + "content": "for " + }, + { + "bbox": [ + 328, + 186, + 399, + 194 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 328, + 186, + 399, + 194 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 328, + 186, + 399, + 194 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 328, + 186, + 399, + 194 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 343, + 196, + 413, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 196, + 413, + 205 + ], + "spans": [ + { + "bbox": [ + 343, + 196, + 413, + 205 + ], + "type": "text", + "content": "/* FL Server: */" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 343, + 206, + 504, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 206, + 504, + 219 + ], + "spans": [ + { + "bbox": [ + 343, + 206, + 504, + 219 + ], + "type": "text", + "content": "Broadcast " + }, + { + "bbox": [ + 343, + 206, + 504, + 219 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^{t - 1}" + }, + { + "bbox": [ + 343, + 206, + 504, + 219 + ], + "type": "text", + "content": " to participating clients;" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 343, + 228, + 442, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 228, + 442, + 239 + ], + "spans": [ + { + "bbox": [ + 343, + 228, + 442, + 239 + ], + "type": "text", + "content": "/* Each FL Client " + }, + { + "bbox": [ + 343, + 228, + 442, + 239 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 343, + 228, + 442, + 239 + ], + "type": "text", + "content": " : */" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 343, + 241, + 507, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 241, + 507, + 252 + ], + "spans": [ + { + "bbox": [ + 343, + 241, + 507, + 252 + ], + "type": "text", + "content": "// Representation Angle Matrix Learning" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 343, + 253, + 449, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 253, + 449, + 264 + ], + "spans": [ + { + "bbox": [ + 343, + 253, + 449, + 264 + ], + "type": "text", + "content": "for each " + }, + { + "bbox": [ + 343, + 253, + 449, + 264 + ], + "type": "inline_equation", + "content": "(x_{i},y_{i})\\in D_{k}" + }, + { + "bbox": [ + 343, + 253, + 449, + 264 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 358, + 265, + 508, + 328 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 265, + 508, + 328 + ], + "spans": [ + { + "bbox": [ + 358, + 265, + 508, + 328 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {R} _ {i} = f _ {k} ^ {e x} \\left(\\boldsymbol {x} _ {i}; \\omega_ {k} ^ {e x, t - 1}\\right); \\\\ \\tilde {\\mathcal {R}} _ {i} = \\mathcal {R} _ {i} \\times \\boldsymbol {A} ^ {t - 1}; \\\\ \\hat {y} _ {i} = f ^ {h d} \\left(\\mathcal {R} _ {i} + \\tilde {\\mathcal {R}} _ {i}; \\omega_ {k} ^ {h d, t - 1}\\right); \\\\ f _ {k} \\left(\\omega_ {k} ^ {t}\\right) \\leftarrow f _ {k} \\left(\\omega_ {k} ^ {t - 1}\\right) - \\eta_ {\\omega} \\nabla \\ell \\left(\\dot {y} _ {i}, y _ {i}\\right); \\\\ \\mathcal {A} _ {k} ^ {t} \\leftarrow \\mathcal {A} ^ {t - 1} - \\eta_ {\\mathcal {A}} \\nabla \\ell (\\hat {y} _ {i}, y _ {i}); \\\\ \\end{array}", + "image_path": "00ad4f0322353ead02d3af2eb413d51f81fc111a39752cf4cfc88629a1e6f7e2.jpg" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 343, + 329, + 360, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 329, + 360, + 338 + ], + "spans": [ + { + "bbox": [ + 343, + 329, + 360, + 338 + ], + "type": "text", + "content": "end" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 343, + 340, + 485, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 340, + 485, + 351 + ], + "spans": [ + { + "bbox": [ + 343, + 340, + 485, + 351 + ], + "type": "text", + "content": "// Adaptive Diagonal Sparsification" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 343, + 353, + 445, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 353, + 445, + 366 + ], + "spans": [ + { + "bbox": [ + 343, + 353, + 445, + 366 + ], + "type": "text", + "content": "Obtain " + }, + { + "bbox": [ + 343, + 353, + 445, + 366 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{A}}_k^t" + }, + { + "bbox": [ + 343, + 353, + 445, + 366 + ], + "type": "text", + "content": " using Eq. (6);" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 343, + 368, + 459, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 368, + 459, + 381 + ], + "spans": [ + { + "bbox": [ + 343, + 368, + 459, + 381 + ], + "type": "text", + "content": "Upload " + }, + { + "bbox": [ + 343, + 368, + 459, + 381 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{A}}_k^t" + }, + { + "bbox": [ + 343, + 368, + 459, + 381 + ], + "type": "text", + "content": " to the FL server;" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 343, + 390, + 412, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 390, + 412, + 401 + ], + "spans": [ + { + "bbox": [ + 343, + 390, + 412, + 401 + ], + "type": "text", + "content": "/* FL Server: */" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 343, + 403, + 500, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 403, + 500, + 414 + ], + "spans": [ + { + "bbox": [ + 343, + 403, + 500, + 414 + ], + "type": "text", + "content": "// Element-wise Weighted Aggregation" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 343, + 415, + 515, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 415, + 515, + 427 + ], + "spans": [ + { + "bbox": [ + 343, + 415, + 515, + 427 + ], + "type": "text", + "content": "Generate " + }, + { + "bbox": [ + 343, + 415, + 515, + 427 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^t" + }, + { + "bbox": [ + 343, + 415, + 515, + 427 + ], + "type": "text", + "content": " by aggregation using Eq. (7);" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 343, + 427, + 348, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 427, + 348, + 436 + ], + "spans": [ + { + "bbox": [ + 343, + 427, + 348, + 436 + ], + "type": "text", + "content": "1" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 328, + 438, + 497, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 438, + 497, + 449 + ], + "spans": [ + { + "bbox": [ + 328, + 438, + 497, + 449 + ], + "type": "text", + "content": "Return local heterogeneous client models" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 333, + 449, + 526, + 463 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 449, + 526, + 463 + ], + "spans": [ + { + "bbox": [ + 333, + 449, + 526, + 463 + ], + "type": "interline_equation", + "content": "\\{f _ {0} \\left(\\omega_ {0} ^ {T - 1}\\right), \\dots , f _ {k} \\left(\\omega_ {k} ^ {T - 1}\\right), \\dots , f _ {K - 1} \\left(\\omega_ {K - 1} ^ {T - 1}\\right) \\}.", + "image_path": "cb763c7a854e639c7ea5590c17a296779c2c7cc79bb0fc1efc093e95b690db23.jpg" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 314, + 490, + 422, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 490, + 422, + 503 + ], + "spans": [ + { + "bbox": [ + 314, + 490, + 422, + 503 + ], + "type": "text", + "content": "5.1. Experiment Setup" + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 313, + 509, + 553, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 553, + 533 + ], + "type": "text", + "content": "We first introduce datasets, base models, comparison baselines, evaluation metrics and hyperparameter settings." + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 313, + 533, + 553, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 553, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 553, + 556 + ], + "type": "text", + "content": "Datasets. We choose 4 image classification benchmark datasets commonly used in FL investigations." + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 314, + 558, + 553, + 654 + ], + "type": "list", + "angle": 0, + "index": 50, + "blocks": [ + { + "bbox": [ + 314, + 558, + 553, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 558, + 553, + 582 + ], + "spans": [ + { + "bbox": [ + 314, + 558, + 553, + 582 + ], + "type": "text", + "content": "- Fashion-MNIST² [54] contains 60,000 training and 10,000 testing grayscale " + }, + { + "bbox": [ + 314, + 558, + 553, + 582 + ], + "type": "inline_equation", + "content": "28 \\times 28" + }, + { + "bbox": [ + 314, + 558, + 553, + 582 + ], + "type": "text", + "content": " 10-class images." + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 315, + 582, + 553, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 582, + 553, + 605 + ], + "spans": [ + { + "bbox": [ + 315, + 582, + 553, + 605 + ], + "type": "text", + "content": "- CIFAR-10 [26] contains 50,000 training and 10,000 testing colourful " + }, + { + "bbox": [ + 315, + 582, + 553, + 605 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 315, + 582, + 553, + 605 + ], + "type": "text", + "content": " 10-class images." + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 315, + 606, + 553, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 606, + 553, + 629 + ], + "spans": [ + { + "bbox": [ + 315, + 606, + 553, + 629 + ], + "type": "text", + "content": "- CIFAR-100" + }, + { + "bbox": [ + 315, + 606, + 553, + 629 + ], + "type": "inline_equation", + "content": "^4" + }, + { + "bbox": [ + 315, + 606, + 553, + 629 + ], + "type": "text", + "content": " [26] contains 50,000 training and 10,000 testing colourful " + }, + { + "bbox": [ + 315, + 606, + 553, + 629 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 315, + 606, + 553, + 629 + ], + "type": "text", + "content": " 100-class images." + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 315, + 630, + 553, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 630, + 553, + 654 + ], + "spans": [ + { + "bbox": [ + 315, + 630, + 553, + 654 + ], + "type": "text", + "content": "- Tiny-ImageNet [10] contains 100,000 training and 10,000 testing colourful " + }, + { + "bbox": [ + 315, + 630, + 553, + 654 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 315, + 630, + 553, + 654 + ], + "type": "text", + "content": " 200-class images." + } + ] + } + ], + "index": 49 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 236, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 236, + 712 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 236, + 712 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 314, + 663, + 553, + 682 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 663, + 553, + 682 + ], + "spans": [ + { + "bbox": [ + 314, + 663, + 553, + 682 + ], + "type": "text", + "content": "2https://github.com/zalandoresearch/fashion-mnist" + } + ] + } + ], + "index": 51 + }, + { + "bbox": [ + 326, + 683, + 545, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 683, + 545, + 693 + ], + "spans": [ + { + "bbox": [ + 326, + 683, + 545, + 693 + ], + "type": "text", + "content": "3https://www.cs.toronto.edu/%7Ekriz/cifar.html" + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 326, + 693, + 545, + 702 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 693, + 545, + 702 + ], + "spans": [ + { + "bbox": [ + 326, + 693, + 545, + 702 + ], + "type": "text", + "content": "4https://www.cs.toronto.edu/%7Ekriz/cifar.html" + } + ] + } + ], + "index": 53 + }, + { + "bbox": [ + 326, + 703, + 504, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 703, + 504, + 712 + ], + "spans": [ + { + "bbox": [ + 326, + 703, + 504, + 712 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 326, + 703, + 504, + 712 + ], + "type": "text", + "content": "https://tiny-imagenet.herokuapp.com/" + } + ] + } + ], + "index": 54 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1318" + } + ] + } + ], + "index": 56 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 91, + 294, + 180 + ], + "blocks": [ + { + "bbox": [ + 81, + 71, + 270, + 81 + ], + "lines": [ + { + "bbox": [ + 81, + 71, + 270, + 81 + ], + "spans": [ + { + "bbox": [ + 81, + 71, + 270, + 81 + ], + "type": "text", + "content": "Table 1. Heterogeneous CNNs for Fashion-MNIST." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 91, + 294, + 180 + ], + "lines": [ + { + "bbox": [ + 56, + 91, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 56, + 91, + 294, + 180 + ], + "type": "table", + "html": "
LayerCNN-1CNN-2CNN-3CNN-4CNN-5
Conv15×5, 205×5, 205×5, 205×5, 205×5, 20
Maxpool12×22×22×22×22×2
Conv25×5, 205×5, 205×5, 205×5, 205×5, 20
Maxpool22×22×22×22×22×2
FC130020015010050
FC25050505050
FC31010101010
model size0.47 MB0.33 MB0.26 MB0.19 MB0.12 MB
", + "image_path": "0df7b092c84881b263cb7ccff163f0c429f248c4be4944ad14547cce2f7fb834.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 56, + 180, + 259, + 188 + ], + "lines": [ + { + "bbox": [ + 56, + 180, + 259, + 188 + ], + "spans": [ + { + "bbox": [ + 56, + 180, + 259, + 188 + ], + "type": "text", + "content": "Note: kernel size: " + }, + { + "bbox": [ + 56, + 180, + 259, + 188 + ], + "type": "inline_equation", + "content": "5 \\times 5" + }, + { + "bbox": [ + 56, + 180, + 259, + 188 + ], + "type": "text", + "content": ", with 20 filters for convolution layers." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 220, + 294, + 308 + ], + "blocks": [ + { + "bbox": [ + 61, + 200, + 288, + 210 + ], + "lines": [ + { + "bbox": [ + 61, + 200, + 288, + 210 + ], + "spans": [ + { + "bbox": [ + 61, + 200, + 288, + 210 + ], + "type": "text", + "content": "Table 2. Heterogeneous CNNs for CIFAR-10 and CIFAR-100." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 220, + 294, + 308 + ], + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 308 + ], + "type": "table", + "html": "
LayerCNN-1CNN-2CNN-3CNN-4CNN-5
Conv15×5, 165×5, 165×5, 165×5, 165×5, 16
Maxpool12×22×22×22×22×2
Conv25×5, 325×5, 165×5, 325×5, 325×5, 32
Maxpool22×22×22×22×22×2
FC1200020001000800500
FC2500500500500500
FC310/10010/10010/10010/10010/100
model size10.01 MB6.93 MB5.04 MB4.05 MB2.56 MB
", + "image_path": "f6a3f14249439cac1a22e2ef03bbd4200cc4007bdd3d6414e1db78fd60a8f2b2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 56, + 308, + 278, + 316 + ], + "lines": [ + { + "bbox": [ + 56, + 308, + 278, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 278, + 316 + ], + "type": "text", + "content": "Note: kernel size: " + }, + { + "bbox": [ + 56, + 308, + 278, + 316 + ], + "type": "inline_equation", + "content": "5 \\times 5" + }, + { + "bbox": [ + 56, + 308, + 278, + 316 + ], + "type": "text", + "content": ", with 16 or 32 filters for convolution layers." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 337, + 291, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 337, + 291, + 349 + ], + "spans": [ + { + "bbox": [ + 67, + 337, + 291, + 349 + ], + "type": "text", + "content": "Then, we partition them into non-IID modes in 2 ways:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 349, + 294, + 420 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 55, + 349, + 293, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 349, + 293, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 293, + 373 + ], + "type": "text", + "content": "- Pathological. Following [48], we allocate " + }, + { + "bbox": [ + 55, + 349, + 293, + 373 + ], + "type": "inline_equation", + "content": "\\{2, 2, 10, 20\\}" + }, + { + "bbox": [ + 55, + 349, + 293, + 373 + ], + "type": "text", + "content": " classes from " + }, + { + "bbox": [ + 55, + 349, + 293, + 373 + ], + "type": "inline_equation", + "content": "\\{10, 10, 100, 200\\}" + }, + { + "bbox": [ + 55, + 349, + 293, + 373 + ], + "type": "text", + "content": " classes to each FL client." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 373, + 294, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 373, + 294, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 373, + 294, + 420 + ], + "type": "text", + "content": "- Practical. Following [43], we allocate " + }, + { + "bbox": [ + 56, + 373, + 294, + 420 + ], + "type": "inline_equation", + "content": "\\{10, 10, 100, 200\\}" + }, + { + "bbox": [ + 56, + 373, + 294, + 420 + ], + "type": "text", + "content": " classes to each FL client and use a Dirichlet distribution function with a hyperparameter " + }, + { + "bbox": [ + 56, + 373, + 294, + 420 + ], + "type": "inline_equation", + "content": "\\alpha = 0.4" + }, + { + "bbox": [ + 56, + 373, + 294, + 420 + ], + "type": "text", + "content": " to generate diverse ratios of one class for different clients." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 421, + 295, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 421, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 295, + 552 + ], + "type": "text", + "content": "Base Models. Following [62], we construct 5 heterogeneous CNN models " + }, + { + "bbox": [ + 55, + 421, + 295, + 552 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{CNN - 1,}\\dots,\\mathrm{CNN - 5}\\}" + }, + { + "bbox": [ + 55, + 421, + 295, + 552 + ], + "type": "text", + "content": " with different channels of convolutional layers or dimensions of linear layers (Tables 1 and 2) and allocate them to 100 clients. The client " + }, + { + "bbox": [ + 55, + 421, + 295, + 552 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 421, + 295, + 552 + ], + "type": "text", + "content": " is assigned CNN- " + }, + { + "bbox": [ + 55, + 421, + 295, + 552 + ], + "type": "inline_equation", + "content": "(k\\% 5)" + }, + { + "bbox": [ + 55, + 421, + 295, + 552 + ], + "type": "text", + "content": ". Note that the representation " + }, + { + "bbox": [ + 55, + 421, + 295, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 55, + 421, + 295, + 552 + ], + "type": "text", + "content": " dimension " + }, + { + "bbox": [ + 55, + 421, + 295, + 552 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 421, + 295, + 552 + ], + "type": "text", + "content": " (the dimension of the penultimate linear layer) is " + }, + { + "bbox": [ + 55, + 421, + 295, + 552 + ], + "type": "inline_equation", + "content": "\\{50,500,500\\}" + }, + { + "bbox": [ + 55, + 421, + 295, + 552 + ], + "type": "text", + "content": " for Fashion-MNIST, CIFAR-10 and CIFAR-100. Thus, the dimensions of the homogeneous representation angle square matrix " + }, + { + "bbox": [ + 55, + 421, + 295, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 55, + 421, + 295, + 552 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 55, + 421, + 295, + 552 + ], + "type": "inline_equation", + "content": "\\{50\\times 50,500\\times 500,500\\times 500\\}" + }, + { + "bbox": [ + 55, + 421, + 295, + 552 + ], + "type": "text", + "content": ". For Tiny-ImageNet, we evenly allocate ResNet-" + }, + { + "bbox": [ + 55, + 421, + 295, + 552 + ], + "type": "inline_equation", + "content": "\\{18,34,50,101,152\\}" + }, + { + "bbox": [ + 55, + 421, + 295, + 552 + ], + "type": "text", + "content": " to 100 clients." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 552, + 295, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 295, + 589 + ], + "type": "text", + "content": "Comparison Baselines. We compare FedRAL with 6 baselines, which are the latest representative MHFL works that support completely heterogeneous local models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 591, + 295, + 663 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 55, + 591, + 295, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 591, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 591, + 295, + 613 + ], + "type": "text", + "content": "- Standalone: each client solely trains its local heterogeneous model, without FL." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 615, + 226, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 615, + 226, + 627 + ], + "spans": [ + { + "bbox": [ + 56, + 615, + 226, + 627 + ], + "type": "text", + "content": "- Model Decoupling: LG-FedAvg [30]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 628, + 264, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 628, + 264, + 639 + ], + "spans": [ + { + "bbox": [ + 56, + 628, + 264, + 639 + ], + "type": "text", + "content": "- Mutual Learning: FedKD [53], FedAPEN [43]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 639, + 294, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 639, + 294, + 663 + ], + "spans": [ + { + "bbox": [ + 56, + 639, + 294, + 663 + ], + "type": "text", + "content": "- Knowledge Distillation without Public Dataset: FedProto [51], FedTGP [72]." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "type": "text", + "content": "Evaluation Metrics. We evaluate the performances of FedRAL and baselines with the following three metrics:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "- Model Accuracy. We record each client's model accuracy (\\%) and calculate their average in each communica" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 322, + 72, + 531, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 531, + 84 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 531, + 84 + ], + "type": "text", + "content": "tion round. We report the highest average accuracy." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 84, + 553, + 203 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 314, + 84, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 84, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 314, + 84, + 553, + 144 + ], + "type": "text", + "content": "- Communication Cost. We track the average number of transmitted parameters between clients and the server in one communication round, and we record the total rounds consumed to reach the specified target accuracy. The product of the two is the total communication cost." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 144, + 553, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 553, + 203 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 553, + 203 + ], + "type": "text", + "content": "- Computational Overhead. We track the average computational FLOPs of clients in one communication round, and we record the total communication rounds consumed to reach the specified target accuracy. The product of the two is the total computational overhead." + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 205, + 553, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 553, + 348 + ], + "type": "text", + "content": "Hyperparameters. We use grid-search to choose the optimal hyperparameters for all algorithms. For general FL hyperparameters, we set the local training epoch " + }, + { + "bbox": [ + 313, + 205, + 553, + 348 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 313, + 205, + 553, + 348 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 205, + 553, + 348 + ], + "type": "inline_equation", + "content": "\\{1,10\\}" + }, + { + "bbox": [ + 313, + 205, + 553, + 348 + ], + "type": "text", + "content": " and batch size " + }, + { + "bbox": [ + 313, + 205, + 553, + 348 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 205, + 553, + 348 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 205, + 553, + 348 + ], + "type": "inline_equation", + "content": "\\{32,64,128,256,512\\}" + }, + { + "bbox": [ + 313, + 205, + 553, + 348 + ], + "type": "text", + "content": ", and we choose the SGD gradient optimizer with a learning rate of 0.01. To ensure the sufficient convergence of algorithms, we set the total communication rounds " + }, + { + "bbox": [ + 313, + 205, + 553, + 348 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 205, + 553, + 348 + ], + "type": "text", + "content": " as 500. For the unique hyperparameter - the number " + }, + { + "bbox": [ + 313, + 205, + 553, + 348 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 205, + 553, + 348 + ], + "type": "text", + "content": " of diagonal blocks for sparsifying the homogeneous representation angle square matrix in FedRAL, we set it to be the factor of the representation dimension " + }, + { + "bbox": [ + 313, + 205, + 553, + 348 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 205, + 553, + 348 + ], + "type": "text", + "content": ", e.g., " + }, + { + "bbox": [ + 313, + 205, + 553, + 348 + ], + "type": "inline_equation", + "content": "m = \\{1,2,\\dots,50\\}" + }, + { + "bbox": [ + 313, + 205, + 553, + 348 + ], + "type": "text", + "content": " for Fashion-MNIST, and " + }, + { + "bbox": [ + 313, + 205, + 553, + 348 + ], + "type": "inline_equation", + "content": "m = \\{1,2,\\dots,500\\}" + }, + { + "bbox": [ + 313, + 205, + 553, + 348 + ], + "type": "text", + "content": " for CIFAR." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 358, + 474, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 474, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 474, + 371 + ], + "type": "text", + "content": "5.2. Comparison Results Analysis" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 376, + 553, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 553, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 553, + 508 + ], + "type": "text", + "content": "Average Accuracy. Table 3 shows that FedRAL outperforms all baselines in terms of average test accuracy, demonstrating its effectiveness in enhancing local model generalization through the shared homogeneous representation angle square matrix. As a recent MHFL method, FedTGP achieves the second-best performance. We notice that FedAPEN fails to converge in some cases. This can be attributed to the fact that it uses the same coefficients to combine the outputs of the heterogeneous and homogeneous models for all data samples, failing to balance personalization and generalization under varying data features." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 510, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 554, + 605 + ], + "type": "text", + "content": "Figure 5 presents how the average test accuracies of FedRAL and the state-of-the-art FedTGP baseline vary as communication rounds. We see that FedRAL converges to a higher accuracy with a faster speed than FedTGP, especially under more non-IID practical FL settings. This demonstrates the effectiveness of FedRAL in alleviating data heterogeneity while supporting the collaboration of heterogeneous client models." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": "Communication Cost. Table 4 presents the communication costs of the state-of-the-art FedTGP baseline and FedRAL. We can observe that FedRAL consumes a lower single-round communication cost than FedTGP (column-3), since the sparse representation angle square matrix communicated in FedRAL has fewer parameters than the label-wise average representations transmitted in FedTGP. FedRAL also requires fewer communication rounds to reach the specified target accuracy (column-5), since the de" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1319" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 91, + 553, + 178 + ], + "blocks": [ + { + "bbox": [ + 80, + 71, + 528, + 82 + ], + "lines": [ + { + "bbox": [ + 80, + 71, + 528, + 82 + ], + "spans": [ + { + "bbox": [ + 80, + 71, + 528, + 82 + ], + "type": "text", + "content": "Table 3. Comparison of average test accuracy (%) on 4 datasets under pathological and practical data heterogeneity settings." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 91, + 553, + 178 + ], + "lines": [ + { + "bbox": [ + 56, + 91, + 553, + 178 + ], + "spans": [ + { + "bbox": [ + 56, + 91, + 553, + 178 + ], + "type": "table", + "html": "
MethodPathologicalPractical
Fashion-MNISTCIFAR-10CIFAR-100Tiny-ImageNetFashion-MNISTCIFAR-10CIFAR-100Tiny-ImageNet
Standalone99.03±0.5391.97±0.3753.04±0.2433.15±0.0273.39±0.8940.72±0.839.56±0.347.35±0.21
LG-FedAvg [30]98.87±0.4091.27±0.2545.83±0.6735.79±0.1971.97±0.2540.44±0.579.80±0.397.51±0.34
FedKD [53]58.37±0.0573.21±0.5637.21±0.9834.53±0.1645.57±0.4928.14±0.298.29±0.286.29±0.23
FedAPEN [43]44.20±0.58-------
FedProto [51]99.02±0.5092.49±0.2653.67±0.7936.24±0.2774.06±0.6340.20±0.119.27±0.717.63±0.29
FedTGP [72]99.09±0.6592.50±0.3153.20±0.7836.92±0.1073.64±0.6040.95±0.099.37±0.137.65±0.19
FedRAL99.54±0.1693.60±0.1958.70±0.0538.75±0.1378.58±0.1242.97±0.1011.70±0.029.47±0.13
", + "image_path": "215cb1c221319255a623ead291d2ac51c5e6f3f117ce9e3ab10584c39512bb90.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 56, + 179, + 463, + 188 + ], + "lines": [ + { + "bbox": [ + 56, + 179, + 463, + 188 + ], + "spans": [ + { + "bbox": [ + 56, + 179, + 463, + 188 + ], + "type": "text", + "content": "Note: “-” denotes failure to converge. Bold indicates the highest accuracy, and underline indicates the second highest accuracy." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 56, + 208, + 167, + 296 + ], + "blocks": [ + { + "bbox": [ + 56, + 208, + 167, + 296 + ], + "lines": [ + { + "bbox": [ + 56, + 208, + 167, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 167, + 296 + ], + "type": "image", + "image_path": "11085d94379bc0b80b3e79100512dfe7eaef036e3ab42fb8180a3fefeb8d6800.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 297, + 156, + 304 + ], + "lines": [ + { + "bbox": [ + 86, + 297, + 156, + 304 + ], + "spans": [ + { + "bbox": [ + 86, + 297, + 156, + 304 + ], + "type": "text", + "content": "Communication Round" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 57, + 306, + 166, + 393 + ], + "blocks": [ + { + "bbox": [ + 57, + 306, + 166, + 393 + ], + "lines": [ + { + "bbox": [ + 57, + 306, + 166, + 393 + ], + "spans": [ + { + "bbox": [ + 57, + 306, + 166, + 393 + ], + "type": "image", + "image_path": "e8f965e28d3aebdf420070565b353f1f4c7b1ded62452b2bfcf186ed88a63763.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 395, + 156, + 402 + ], + "lines": [ + { + "bbox": [ + 86, + 395, + 156, + 402 + ], + "spans": [ + { + "bbox": [ + 86, + 395, + 156, + 402 + ], + "type": "text", + "content": "Communication Round" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 68, + 411, + 282, + 422 + ], + "lines": [ + { + "bbox": [ + 68, + 411, + 282, + 422 + ], + "spans": [ + { + "bbox": [ + 68, + 411, + 282, + 422 + ], + "type": "text", + "content": "Figure 5. Average test accuracy " + }, + { + "bbox": [ + 68, + 411, + 282, + 422 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 68, + 411, + 282, + 422 + ], + "type": "text", + "content": " vs. number of rounds." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 176, + 217, + 284, + 296 + ], + "blocks": [ + { + "bbox": [ + 196, + 208, + 283, + 217 + ], + "lines": [ + { + "bbox": [ + 196, + 208, + 283, + 217 + ], + "spans": [ + { + "bbox": [ + 196, + 208, + 283, + 217 + ], + "type": "text", + "content": "Pathological, CIFAR-100" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 176, + 217, + 284, + 296 + ], + "lines": [ + { + "bbox": [ + 176, + 217, + 284, + 296 + ], + "spans": [ + { + "bbox": [ + 176, + 217, + 284, + 296 + ], + "type": "image", + "image_path": "9608db99494f565ae83e4cee0c50a3f661f0931e1cd8d52f833fc3d11a4530be.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 204, + 297, + 274, + 304 + ], + "lines": [ + { + "bbox": [ + 204, + 297, + 274, + 304 + ], + "spans": [ + { + "bbox": [ + 204, + 297, + 274, + 304 + ], + "type": "text", + "content": "Communication Round" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 179, + 306, + 284, + 393 + ], + "blocks": [ + { + "bbox": [ + 179, + 306, + 284, + 393 + ], + "lines": [ + { + "bbox": [ + 179, + 306, + 284, + 393 + ], + "spans": [ + { + "bbox": [ + 179, + 306, + 284, + 393 + ], + "type": "image", + "image_path": "4d3466a93a94b8842746b7c9c8d6c881807eea9a38efb853a3a5e644a0753635.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 205, + 395, + 274, + 402 + ], + "lines": [ + { + "bbox": [ + 205, + 395, + 274, + 402 + ], + "spans": [ + { + "bbox": [ + 205, + 395, + 274, + 402 + ], + "type": "text", + "content": "Communication Round" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 447, + 294, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 294, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 294, + 495 + ], + "type": "text", + "content": "signed representation angle learning facilitates model convergence. Overall, FedRAL consumes lower communication costs for reaching the target accuracy than FedTGP (column-6), demonstrating its efficient communication." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 54, + 497, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 497, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 54, + 497, + 295, + 651 + ], + "type": "text", + "content": "Computational Cost. Table 4 also shows the computational costs of the state-of-the-art FedTGP baseline and FedRAL. We can see that FedRAL consumes a lower single-round computational cost than FedTGP (column-4), since additionally training a homogeneous representation angle square matrix in FedRAL consumes lower FLOPs than additionally calculating label-wise average representations in FedTGP, except local model training in both methods. Since FedRAL requires fewer communication rounds to reach the specified target accuracy (column-5), its total computational costs are lower than FedTGP when reaching the specified target accuracy (column-7), substantiating its efficient computation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": "Privacy Evaluation. We conduct experiments to evaluate the privacy of FedRAL and baselines. Followed by Zhang et al. [70], considering a semi-honest FL scenario where the server is untrustworthy and may inversely infer the original data of comprised clients from transmit-" + } + ] + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 315, + 226, + 553, + 308 + ], + "blocks": [ + { + "bbox": [ + 342, + 207, + 525, + 217 + ], + "lines": [ + { + "bbox": [ + 342, + 207, + 525, + 217 + ], + "spans": [ + { + "bbox": [ + 342, + 207, + 525, + 217 + ], + "type": "text", + "content": "Table 4. Communication and computational costs." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 226, + 553, + 308 + ], + "lines": [ + { + "bbox": [ + 315, + 226, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 315, + 226, + 553, + 308 + ], + "type": "table", + "html": "
FL SettingMethodComm/C/R (KB)Comp/C/R (MB)RoundsComm/C (MB)Comp/C (GB)
Pathological CIFAR-10FedTGP3.9154.523021.1516.08
FedRAL1.5628.232010.315.54
Pathological CIFAR-100FedTGP19.5354.523546.7518.85
FedRAL9.7728.231441.373.97
Practical CIFAR-10FedTGP19.5354.873486.6418.65
FedRAL7.8128.581030.792.87
Practical CIFAR-100FedTGP195.3154.8736369.2419.45
FedRAL39.0628.581465.574.07
", + "image_path": "3ed37b9d9b117ec306673ab52863f411f1f6ca4a72d2b060c3e6e7dd04bf91d8.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 313, + 308, + 552, + 384 + ], + "lines": [ + { + "bbox": [ + 313, + 308, + 552, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 552, + 384 + ], + "type": "text", + "content": "Note: \"Comm/C/R\": a single client's communication cost (KB) in a communication round. \"Comp/C/R\": a single client's computational cost (MB) in a communication round. \"Rounds\": the communication rounds required to reach the specified " + }, + { + "bbox": [ + 313, + 308, + 552, + 384 + ], + "type": "inline_equation", + "content": "\\{90, 50, 40, 9\\}\\%" + }, + { + "bbox": [ + 313, + 308, + 552, + 384 + ], + "type": "text", + "content": " target accuracy (the near maximum accuracy that both the state-of-the-art FedTGP baseline and FedRAL can achieve). \"Comm/C\": a single client's communication cost (MB) required to reach the target accuracy. \"Comm/C\": a single client's computational cost (GB) required to reach the target accuracy." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 315, + 415, + 553, + 435 + ], + "blocks": [ + { + "bbox": [ + 341, + 396, + 526, + 406 + ], + "lines": [ + { + "bbox": [ + 341, + 396, + 526, + 406 + ], + "spans": [ + { + "bbox": [ + 341, + 396, + 526, + 406 + ], + "type": "text", + "content": "Table 5. PSNR under DLG attacks on CIFAR-100." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 415, + 553, + 435 + ], + "lines": [ + { + "bbox": [ + 315, + 415, + 553, + 435 + ], + "spans": [ + { + "bbox": [ + 315, + 415, + 553, + 435 + ], + "type": "table", + "html": "
MethodLG-FedAvgFedKDFedAPENFedProtoFedTGPFedRAL
PSNR (dB)7.836.956.936.766.696.28
", + "image_path": "1a9cbe095d75a5204f3203c00a493637355353f3d78fdb06c385243654674579.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_body" + } + ], + "index": 20 + }, + { + "type": "table", + "bbox": [ + 315, + 478, + 553, + 497 + ], + "blocks": [ + { + "bbox": [ + 313, + 447, + 553, + 468 + ], + "lines": [ + { + "bbox": [ + 313, + 447, + 553, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 553, + 468 + ], + "type": "text", + "content": "Table 6. Average Accuracy " + }, + { + "bbox": [ + 313, + 447, + 553, + 468 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 313, + 447, + 553, + 468 + ], + "type": "text", + "content": " under high model heterogeneity. \" -\" denotes failure to converge." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 478, + 553, + 497 + ], + "lines": [ + { + "bbox": [ + 315, + 478, + 553, + 497 + ], + "spans": [ + { + "bbox": [ + 315, + 478, + 553, + 497 + ], + "type": "table", + "html": "
MethodLG-FedAvgFedKDFedAPENFedProtoFedTGPFedRAL
Accuracy (%)45.6345.95-48.5648.6950.79
", + "image_path": "f7fba3cc95541f910c52ef592f7899e61c28bfd71a91b4543468111f2145c5f3.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "table_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 517, + 554, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 554, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 554, + 590 + ], + "type": "text", + "content": "ted information by launching Deep Leakage from Gradients (DLG) attack. We use the Peak Signal-to-Noise Ratio (PSNR), a common privacy measurement metric, to evaluate the privacy-protecting level. A lower PSNR denotes better privacy protection. The results in Table 5 validate that FedRAL can protect data privacy effectively." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 590, + 554, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 554, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 554, + 662 + ], + "type": "text", + "content": "High Model Heterogeneity. To evaluate how FedRAL and baselines perform under high model heterogeneity, we conduct experiments CIFAR-100 with 100 clients and heterogeneous models including ResNet-{4, 6, 8, 10, 18, 34, 50, 101, 152}, CNN, and ViT. The results in Table 6 validate FedRAL's robustness to high model heterogeneity." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 671, + 432, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 432, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 432, + 684 + ], + "type": "text", + "content": "5.3. Robustness Analysis" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "This section analyzes the robustness of FedRAL to different non-IID degrees of two types." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1320" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 71, + 167, + 169 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 167, + 169 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 167, + 169 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 167, + 169 + ], + "type": "image", + "image_path": "d12d1e619a820b815069035e927909bd39e9706d01da73fc3d9eb4e6fa38964f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 177, + 268, + 189 + ], + "lines": [ + { + "bbox": [ + 83, + 177, + 268, + 189 + ], + "spans": [ + { + "bbox": [ + 83, + 177, + 268, + 189 + ], + "type": "text", + "content": "Figure 6. Robustness to pathological non-IIDness." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 71, + 286, + 170 + ], + "blocks": [ + { + "bbox": [ + 173, + 71, + 286, + 170 + ], + "lines": [ + { + "bbox": [ + 173, + 71, + 286, + 170 + ], + "spans": [ + { + "bbox": [ + 173, + 71, + 286, + 170 + ], + "type": "image", + "image_path": "e4b75de0008a8b0a5c0d32b7519f1a3ffbac953952a0c155c446f408b28c2445.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 56, + 200, + 167, + 299 + ], + "blocks": [ + { + "bbox": [ + 56, + 200, + 167, + 299 + ], + "lines": [ + { + "bbox": [ + 56, + 200, + 167, + 299 + ], + "spans": [ + { + "bbox": [ + 56, + 200, + 167, + 299 + ], + "type": "image", + "image_path": "fe088149e870e34a087efc4e4cac6ecc7faba08c13b9369c4041a4189fbbc257.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 90, + 308, + 261, + 319 + ], + "lines": [ + { + "bbox": [ + 90, + 308, + 261, + 319 + ], + "spans": [ + { + "bbox": [ + 90, + 308, + 261, + 319 + ], + "type": "text", + "content": "Figure 7. Robustness to practical non-IIDness." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 173, + 200, + 286, + 299 + ], + "blocks": [ + { + "bbox": [ + 173, + 200, + 286, + 299 + ], + "lines": [ + { + "bbox": [ + 173, + 200, + 286, + 299 + ], + "spans": [ + { + "bbox": [ + 173, + 200, + 286, + 299 + ], + "type": "image", + "image_path": "27800c7e49c2f1d56639e1c37049fff8ed88eacca6a25895456c0d4994c568ee.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 340, + 296, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 296, + 483 + ], + "type": "text", + "content": "Robustness to pathological non-IIDness. We change the number of classes allocated to each client to obtain different non-IID degrees. For CIFAR-10, we allocate " + }, + { + "bbox": [ + 55, + 340, + 296, + 483 + ], + "type": "inline_equation", + "content": "\\{2,4,6,8,10\\}" + }, + { + "bbox": [ + 55, + 340, + 296, + 483 + ], + "type": "text", + "content": " classes from 10 classes to each client. For CIFAR-100, we allocate " + }, + { + "bbox": [ + 55, + 340, + 296, + 483 + ], + "type": "inline_equation", + "content": "\\{10,30,50,70,90,100\\}" + }, + { + "bbox": [ + 55, + 340, + 296, + 483 + ], + "type": "text", + "content": " classes out of 100 classes to each client. Figure 6 presents that FedRAL always performs higher accuracies than the state-of-the-art FedTGP baseline. As the number of classes assigned to one client increases, i.e., the non-IID degree decreases, model accuracy tends to drop since more IID data benefits generalization while compromising personalization of local models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 484, + 296, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 296, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 296, + 616 + ], + "type": "text", + "content": "Robustness to practical non-IIDness. We vary the hyperparameter " + }, + { + "bbox": [ + 55, + 484, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 484, + 296, + 616 + ], + "type": "text", + "content": " of the Dirichlet distribution function to construct diverse non-IID degrees. For both CIFAR-10 and CIFAR-100, we set " + }, + { + "bbox": [ + 55, + 484, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\alpha = \\{0.1, 0.2, 0.3, 0.4, 0.5\\}" + }, + { + "bbox": [ + 55, + 484, + 296, + 616 + ], + "type": "text", + "content": " to evaluate the average model accuracies of FedRAL and the state-of-the-art FedTGP baseline. Figure 7 shows that FedRAL performs significantly higher average model accuracy than FedTGP, again demonstrating the robustness of FedRAL to non-IIDness. Similarly, as the value of " + }, + { + "bbox": [ + 55, + 484, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 484, + 296, + 616 + ], + "type": "text", + "content": " increases, i.e., the non-IID degree decreases, model accuracy tends to degrade also due to the above reason." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 624, + 162, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 162, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 162, + 636 + ], + "type": "text", + "content": "5.4. Ablation Analysis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "FedRAL involves three contributions: 1) introducing representation angle matrix learning to share cross-client representation angle knowledge for enhancing model generalization, 2) devising adaptive diagonal sparsification to improve communication efficiency while keeping important representation angle knowledge, 3) designing ordinate-wise" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 315, + 91, + 554, + 131 + ], + "blocks": [ + { + "bbox": [ + 333, + 71, + 535, + 82 + ], + "lines": [ + { + "bbox": [ + 333, + 71, + 535, + 82 + ], + "spans": [ + { + "bbox": [ + 333, + 71, + 535, + 82 + ], + "type": "text", + "content": "Table 7. Average accuracy " + }, + { + "bbox": [ + 333, + 71, + 535, + 82 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 333, + 71, + 535, + 82 + ], + "type": "text", + "content": " in ablation experiments." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 91, + 554, + 131 + ], + "lines": [ + { + "bbox": [ + 315, + 91, + 554, + 131 + ], + "spans": [ + { + "bbox": [ + 315, + 91, + 554, + 131 + ], + "type": "table", + "html": "
CaseDiagonal SparsificationWeighted AggregationPathologicalPractical
CIFAR-10CIFAR-100CIFAR-10CIFAR-100
A✓✓93.6058.7042.9711.70
B✗✓91.6456.5841.3711.58
C✓✗93.4753.9440.529.24
", + "image_path": "b7e40be629238630afd77a67614cbb7cf48a9c49b8b64d5d9d1b2792e3c420f0.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 157, + 553, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 553, + 217 + ], + "type": "text", + "content": "weighted aggregation to fuse cross-client representation angle square matrices. Contribution-1) is essential to implement FL collaboration across clients with heterogeneous models, so we conduct experiments to evaluate the necessity of the other two contributions with three cases:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 220, + 553, + 350 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 314, + 220, + 553, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 220, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 314, + 220, + 553, + 255 + ], + "type": "text", + "content": "- Case-A, FedRAL with both adaptive diagonal sparsification and ordinate-wise weighted aggregation, i.e., the designed method." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 255, + 553, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 255, + 553, + 302 + ], + "spans": [ + { + "bbox": [ + 314, + 255, + 553, + 302 + ], + "type": "text", + "content": "- Case-B, FedRAL without adaptive diagonal sparsification but with ordinate-wise weighted aggregation, i.e., clients upload complete representation angle square matrices to the server." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 304, + 553, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 553, + 350 + ], + "type": "text", + "content": "- Case-C, FedRAL with adaptive diagonal sparsification but with naive ordinate-wise averaging aggregation, i.e., the aggregation rule (a) displayed in Figure 4, ignoring client data distribution distinctions." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 354, + 554, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 554, + 414 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 554, + 414 + ], + "type": "text", + "content": "Table 7 shows that removing anyone degrades accuracy. Especially, replacing ordinate-wise weighted aggregation with naive averaging aggregation presents obvious accuracy degradations under most settings. These results reflect the effectiveness and necessity of the proposed two designs." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 433, + 388, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 388, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 388, + 445 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 456, + 554, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 554, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 554, + 552 + ], + "type": "text", + "content": "This work proposed FedRAL with three innovative designs: representation angle learning, adaptive diagonal sparsification, and element-wise weighted aggregation. The three designs enable enabling it to be a privacy-preserved, well-performed, communication and computation-efficient model-heterogeneous FL approach. Extensive experiments demonstrate its advantages in accuracy, communication and computation efficiency." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 571, + 407, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 571, + 407, + 585 + ], + "spans": [ + { + "bbox": [ + 313, + 571, + 407, + 585 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "Xiaoguang Liu is supported by the National Science Foundation of China under Grant 62272252 & 62272253, and the Fundamental Research Funds for the Central Universities. Han Yu is supported by the Ministry of Education, Singapore, under its Academic Research Fund Tier 1 (RG101/24); the RIE2025 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) (Award I2301E0026), administered by A\\*STAR, as well as supported by Alibaba Group and NTU Singapore through Alibaba-NTU Global e-Sustainability CorpLab (ANGEL)." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 314, + 757 + ], + "type": "text", + "content": "1321" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 711 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "type": "text", + "content": "[1] Jin-Hyun Ahn et al. Wireless federated distillation for distributed edge learning with heterogeneous data. In Proc. PIMRC, pages 1-6, Istanbul, Turkey, 2019. IEEE. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 295, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 295, + 158 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 295, + 158 + ], + "type": "text", + "content": "[2] Jin-Hyun Ahn et al. Cooperative learning VIA federated distillation OVER fading channels. In Proc. ICASSP, pages 8856-8860, Barcelona, Spain, 2020. IEEE. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 159, + 294, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 159, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 62, + 159, + 294, + 191 + ], + "type": "text", + "content": "[3] Samiul Alam et al. Fedrolex: Model-heterogeneous federated learning with rolling sub-model extraction. In Proc. NeurIPS, virtual, 2022.. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 194, + 295, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 295, + 225 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 295, + 225 + ], + "type": "text", + "content": "[4] Sara Babakniya et al. Revisiting sparsity hunting in federated learning: Why does sparsity consensus matter? Transactions on Machine Learning Research, 1(1):1, 2023." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 228, + 294, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 228, + 294, + 281 + ], + "spans": [ + { + "bbox": [ + 62, + 228, + 294, + 281 + ], + "type": "text", + "content": "[5] Yun-Hin Chan, Rui Zhou, Running Zhao, Zhihan JIANG, and Edith C. H. Ngai. Internal cross-layer gradients for extending homogeneity to heterogeneity in federated learning. In Proc. ICLR, page 1, Vienna, Austria, 2024. OpenReview.net. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 284, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 284, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 62, + 284, + 294, + 316 + ], + "type": "text", + "content": "[6] Hongyan Chang et al. Cronus: Robust and heterogeneous collaborative learning with black-box knowledge transfer. In Proc. NeurIPS Workshop, virtual, 2021.. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 318, + 294, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 318, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 318, + 294, + 350 + ], + "type": "text", + "content": "[7] Jiangui Chen et al. Fedmatch: Federated learning over heterogeneous question answering data. In Proc. CIKM, pages 181-190, virtual, 2021. ACM. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 352, + 294, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 352, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 352, + 294, + 384 + ], + "type": "text", + "content": "[8] Sijie Cheng et al. Fedgems: Federated learning of larger server models via selective knowledge fusion. CoRR, abs/2110.11027, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 386, + 294, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 386, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 62, + 386, + 294, + 418 + ], + "type": "text", + "content": "[9] Yae Jee Cho et al. Heterogeneous ensemble knowledge transfer for training large models in federated learning. In Proc. IJCAI, pages 2881-2887, virtual, 2022. ijcai.org. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 420, + 294, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 420, + 294, + 452 + ], + "spans": [ + { + "bbox": [ + 57, + 420, + 294, + 452 + ], + "type": "text", + "content": "[10] Patryk Chrabaszcz, Ilya Loshchilov, and Frank Hutter. A downsampled variant of imagenet as an alternative to the CIFAR datasets. CoRR, abs/1707.08819, 2017. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 454, + 294, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 454, + 294, + 486 + ], + "spans": [ + { + "bbox": [ + 57, + 454, + 294, + 486 + ], + "type": "text", + "content": "[11] Liam Collins et al. Exploiting shared representations for personalized federated learning. In Proc. ICML, pages 2089-2099, virtual, 2021. PMLR. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 488, + 294, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 488, + 294, + 529 + ], + "spans": [ + { + "bbox": [ + 57, + 488, + 294, + 529 + ], + "type": "text", + "content": "[12] Enmao Diao. Heterofl: Computation and communication efficient federated learning for heterogeneous clients. In Proc. ICLR, page 1, Virtual Event, Austria, 2021. OpenReview.net. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 533, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 533, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 57, + 533, + 294, + 566 + ], + "type": "text", + "content": "[13] Tao Fan, Hanlin Gu, et al. Ten challenging problems in federated foundation models. IEEE Transactions on Knowledge and Data Engineering, 2025. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 567, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 567, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 57, + 567, + 294, + 599 + ], + "type": "text", + "content": "[14] Randy Goebel, Han Yu, Boi Faltings, Lixin Fan, and Zehui Xiong. Trustworthy Federated Learning. Springer, Cham, 2023. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 601, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 294, + 633 + ], + "type": "text", + "content": "[15] Xuan Gong et al. Federated learning via input-output collaborative distillation. In Proc. AAAI, pages 22058-22066, Vancouver, Canada, 2024. AAAI Press. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 635, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 294, + 666 + ], + "type": "text", + "content": "[16] Chaoyang He et al. Group knowledge transfer: Federated learning of large cnns at the edge. In Proc. NeurIPS, virtual, 2020. . 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 57, + 670, + 294, + 711 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 711 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 711 + ], + "type": "text", + "content": "[17] S. Horváth. FjORD: Fair and accurate federated learning under heterogeneous targets with ordered dropout. In Proc. NIPS, pages 12876-12889, Virtual, 2021. OpenReview.net. 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 37, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "type": "text", + "content": "[18] Edward J. Hu et al. Lora: Low-rank adaptation of large language models. In ICLR, page 1, Virtual, 2022. OpenReview.net. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 107, + 553, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 138 + ], + "type": "text", + "content": "[19] Wenke Huang et al. Learn from others and be yourself in heterogeneous federated learning. In Proc. CVPR, pages 10133-10143, virtual, 2022. IEEE. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 140, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 171 + ], + "type": "text", + "content": "[20] Wenke Huang et al. Few-shot model agnostic federated learning. In Proc. MM, pages 7309-7316, Lisboa, Portugal, 2022. ACM." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 173, + 553, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 553, + 216 + ], + "type": "text", + "content": "[21] Sohei Itahara et al. Distillation-based semi-supervised federated learning for communication-efficient collaborative training with non-iid private data. IEEE Trans. Mob. Comput., 22(1):191-205, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 217, + 553, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 217, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 317, + 217, + 553, + 258 + ], + "type": "text", + "content": "[22] Jaehee Jang et al. Fedclassavg: Local representation learning for personalized federated learning on heterogeneous neural networks. In Proc. ICPP, pages 76:1-76:10, virtual, 2022. ACM, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 261, + 553, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 303 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 303 + ], + "type": "text", + "content": "[23] Eunjoeong Jeong et al. Communication-efficient on-device machine learning: Federated distillation and augmentation under non-iid private data. In Proc. NeurIPS Workshop, virtual, 2018. . 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 305, + 553, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 553, + 336 + ], + "type": "text", + "content": "[24] Peter Kairouz et al. Advances and open problems in federated learning. Foundations and Trends in Machine Learning, 14(1-2):1-210, 2021. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 338, + 553, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 369 + ], + "type": "text", + "content": "[25] Shivam Kalra et al. Decentralized federated learning through proxy model sharing. Nature communications, 14(1):2899, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 371, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 553, + 392 + ], + "type": "text", + "content": "[26] Alex Krizhevsky et al. Learning multiple layers of features from tiny images. Toronto, ON, Canada, 2009. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 394, + 553, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 394, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 317, + 394, + 553, + 447 + ], + "type": "text", + "content": "[27] Meng Lei, Qi Zhuang, Wu Lei, Du Xiaoyu, Li Zhaochuan, Cui Lizhen, and Meng Xiangxu. Improving global generalization and local personalization for federated learning. IEEE Transactions on Neural Networks and Learning Systems, 36, 2024. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 449, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 553, + 480 + ], + "type": "text", + "content": "[28] Daliang Li and Junpu Wang. Fedmd: Heterogenous federated learning via model distillation. In Proc. NeurIPS Workshop, virtual, 2019. . 1, 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 482, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 482, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 482, + 553, + 514 + ], + "type": "text", + "content": "[29] Qinbin Li et al. Practical one-shot federated learning for cross-silo setting. In Proc. IJCAI, pages 1484-1490, virtual, 2021. ijcai.org. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 516, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 516, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 317, + 516, + 553, + 547 + ], + "type": "text", + "content": "[30] Paul Pu Liang et al. Think locally, act globally: Federated learning with local and global representations. arXiv preprint arXiv:2001.01523, 1(1), 2020. 1, 2, 6, 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 548, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 553, + 569 + ], + "type": "text", + "content": "[31] Tao Lin et al. Ensemble distillation for robust model fusion in federated learning. In Proc. NeurIPS, virtual, 2020. . 1, 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 317, + 571, + 553, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 571, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 317, + 571, + 553, + 591 + ], + "type": "text", + "content": "[32] Chang Liu et al. Completely heterogeneous federated learning. CoRR, abs/2210.15865, 2022. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 317, + 593, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 593, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 317, + 593, + 553, + 645 + ], + "type": "text", + "content": "[33] Qianyu Long, Christos Anagnostopoulos, Shameem Puthiya Parambath, and Daning Bi. Feddip: Federated learning with extreme dynamic pruning and incremental regularization. In Proc.ICDM, pages 1187-1192, Shanghai, China, 2023. IEEE. 2" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 317, + 647, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 679 + ], + "type": "text", + "content": "[34] Disha Makhija et al. Architecture agnostic federated learning for neural networks. In Proc. ICML, pages 14860-14870, virtual, 2022. PMLR. 2" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 317, + 681, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 681, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 681, + 553, + 712 + ], + "type": "text", + "content": "[35] Koji Matsuda et al. Fedme: Federated learning via model exchange. In Proc. SDM, pages 459-467, Alexandria, VA, USA, 2022. SIAM. 2" + } + ] + } + ], + "index": 36 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1322" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[36] Brendan McMahan et al. Communication-efficient learning of deep networks from decentralized data. In Proc. AISTATS, pages 1273-1282, Fort Lauderdale, FL, USA, 2017. PMLR. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "text", + "content": "[37] Duy Phuong Nguyen et al. Enhancing heterogeneous federated learning with knowledge extraction and multi-model fusion. In Proc. SC Workshop, pages 36-43, Denver, CO, USA, 2023. ACM. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 196 + ], + "type": "text", + "content": "[38] Jaehoon Oh et al. Fedbabu: Toward enhanced representation for federated image classification. In Proc. ICLR, virtual, 2022. OpenReview.net. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 239 + ], + "type": "text", + "content": "[39] Zhengxiang Pan, Han Yu, Chunyan Miao, and Cyril Leung. Efficient collaborative crowdsourcing. In The 30th AAAI Conference on Artificial Intelligence (AAAI-16), pages 4248–4249, 2016. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 274 + ], + "type": "text", + "content": "[40] Sejun Park et al. Towards understanding ensemble distillation in federated learning. In Proc. ICML, pages 27132-27187, Honolulu, Hawaii, USA, 2023. PMLR. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 275, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 294, + 308 + ], + "type": "text", + "content": "[41] Krishna Pillutla et al. Federated learning with partial model personalization. In Proc. ICML, pages 17716-17758, virtual, 2022. PMLR. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 309, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 365 + ], + "type": "text", + "content": "[42] Zhuang Qi, Sijin Zhou, Lei Meng, Han Yu, and Xiangxu Meng. Federated deconfounding and debiasing learning for out-of-distribution generalization. In The 34th International Joint Conference on Artificial Intelligence, (IJ-CAI'25), pages 1–9, 2025. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 365, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 408 + ], + "type": "text", + "content": "[43] Zhen Qin et al. Fedapen: Personalized cross-silo federated learning with adaptability to statistical heterogeneity. In Proc. KDD, pages 1954-1964, Long Beach, CA, USA, 2023. ACM. 2, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 410, + 294, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 410, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 410, + 294, + 442 + ], + "type": "text", + "content": "[44] Zeju Qiu et al. Controlling text-to-image diffusion by orthogonal finetuning. In Proc. NeurIPS, New Orleans, LA, USA, 2023. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 444, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 444, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 444, + 294, + 476 + ], + "type": "text", + "content": "[45] Chao Ren, Han Yu, et al. Advances and open challenges in federated foundation models. IEEE Communications Surveys and Tutorials, 2025. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 477, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 477, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 477, + 294, + 510 + ], + "type": "text", + "content": "[46] Felix Sattler et al. Fedaux: Leveraging unlabeled auxiliary data in federated learning. IEEE Trans. Neural Networks Learn. Syst., 1(1):1-13, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 511, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 511, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 511, + 294, + 544 + ], + "type": "text", + "content": "[47] Felix Sattler et al. CFD: communication-efficient federated distillation via soft-label quantization and delta coding. IEEE Trans. Netw. Sci. Eng., 9(4):2025-2038, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 545, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 578 + ], + "type": "text", + "content": "[48] Aviv Shamsian et al. Personalized federated learning using hypernetworks. In Proc. ICML, pages 9489-9502, virtual, 2021. PMLR. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 579, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 600 + ], + "type": "text", + "content": "[49] Tao Shen et al. Federated mutual learning. CoRR, abs/2006.16765, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 602, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 634 + ], + "type": "text", + "content": "[50] Alysa Ziying Tan et al. Towards personalized federated learning. IEEE Trans. Neural Networks Learn. Syst., 1(1): 1-17, 2022. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 635, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 668 + ], + "type": "text", + "content": "[51] Yue Tan et al. Fedproto: Federated prototype learning across heterogeneous clients. In Proc. AAAI, pages 8432-8440, virtual, 2022. AAAI Press. 2, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[52] Jiaqi Wang et al. Towards personalized federated learning via heterogeneous model reassembly. In Proc. NeurIPS, page 13, New Orleans, Louisiana, USA, 2023. OpenReview.net. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "text", + "content": "[53] Chuhan Wu et al. Communication-efficient federated learning via knowledge distillation. Nature Communications, 13 (1):2032, 2022. 2, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 108, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 553, + 140 + ], + "type": "text", + "content": "[54] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms, 2017. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 143, + 553, + 165 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 143, + 553, + 165 + ], + "spans": [ + { + "bbox": [ + 316, + 143, + 553, + 165 + ], + "type": "text", + "content": "[55] Qiang Yang, Lixin Fan, and Han Yu. Federated Learning: Privacy and Incentive. Springer, Cham, 2020. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 167, + 553, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 167, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 316, + 167, + 553, + 198 + ], + "type": "text", + "content": "[56] Qiang Yang, Yang Liu, Yong Cheng, Yan Kang, Tianjian Chen, and Han Yu. Federated Learning. Springer, Cham, 2020. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 201, + 553, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 201, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 316, + 201, + 553, + 243 + ], + "type": "text", + "content": "[57] Liping Yi, Gang Wang, and Xiaoguang Liu. QSFL: A two-level uplink communication optimization framework for federated learning. In Proc. ICML, pages 25501-25513. PMLR, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 246, + 553, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 246, + 553, + 290 + ], + "spans": [ + { + "bbox": [ + 316, + 246, + 553, + 290 + ], + "type": "text", + "content": "[58] Liping Yi, Xiaorong Shi, Nan Wang, Gang Wang, Xiaoguang Liu, Zhuan Shi, and Han Yu. pfedkt: Personalized federated learning with dual knowledge transfer. Knowledge-Based Systems, 292:111633, 2024. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 292, + 553, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 292, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 316, + 292, + 553, + 336 + ], + "type": "text", + "content": "[59] Liping Yi, Xiaorong Shi, Nan Wang, Jinsong Zhang, Gang Wang, and Xiaoguang Liu. Fedpe: Adaptive model pruning-expanding for federated learning on mobile devices. IEEE Transactions on Mobile Computing, pages 1-18, 2024. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "type": "text", + "content": "[60] Liping Yi, Gang Wang, Xiaofei Wang, and Xiaoguang Liu. Qsfl: Two-level communication-efficient federated learning on mobile edge devices. IEEE Transactions on Services Computing, pages 1-16, 2024. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 384, + 553, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 384, + 553, + 427 + ], + "spans": [ + { + "bbox": [ + 316, + 384, + 553, + 427 + ], + "type": "text", + "content": "[61] Liping Yi, Han Yu, Zhuan Shi, Gang Wang, Xiaoguang Liu, Lizhen Cui, and Xiaoxiao Li. FedSSA: Semantic Similarity-based Aggregation for Efficient Model-Heterogeneous Personalized Federated Learning. In IJCAI, 2024. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 429, + 553, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 429, + 553, + 461 + ], + "spans": [ + { + "bbox": [ + 316, + 429, + 553, + 461 + ], + "type": "text", + "content": "[62] Liping Yi et al. FedGH: Heterogeneous federated learning with generalized global header. In Proc. ACM MM, 2023. 2, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 463, + 553, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 463, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 316, + 463, + 553, + 495 + ], + "type": "text", + "content": "[63] Liping Yi et al. Federated model heterogeneous matryoshka representation learning. In Proc. NeurIPS, Vancouver, Canada, 2024." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 498, + 553, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 553, + 553 + ], + "type": "text", + "content": "[64] Liping Yi et al. pFedAFM: Adaptive Feature Mixture for Data-Level Personalization in Heterogeneous Federated Learning on Mobile Edge Devices. In Proc. ICDE, pages 1981-1994, Hong Kong SAR, China, 2025. IEEE Computer Society." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 555, + 553, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 553, + 598 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 553, + 598 + ], + "type": "text", + "content": "[65] Liping Yi et al. pfedes: Generalized proxy feature extractor sharing for model heterogeneous personalized federated learning. In Proc. AAAI, pages 22146-22154, Philadelphia, PA, USA, 2025. AAAI Press. 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 600, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 600, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 600, + 553, + 623 + ], + "type": "text", + "content": "[66] Fuxun Yu et al. Fed2: Feature-aligned federated learning. In Proc. KDD, pages 2066-2074, virtual, 2021. ACM. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "text", + "content": "[67] Han Yu, Siyuan Liu, Alex C Kot, Chunyan Miao, and Cyril Leung. Dynamic witness selection for trustworthy distributed cooperative sensing in cognitive radio networks. In Proceedings of the 13th IEEE International Conference on Communication Technology (ICCT'11), pages 1-6, 2011. 1" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 680, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 712 + ], + "type": "text", + "content": "[68] Sixing Yu et al. Resource-aware federated learning using knowledge extraction and multi-model fusion. CoRR, abs/2208.07978, 2022. 2" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1323" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 510 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[69] Jie Zhang et al. Parameterized knowledge transfer for personalized federated learning. In Proc. NeurIPS, pages 10092-10104, virtual, 2021. OpenReview.net. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "type": "text", + "content": "[70] Jianqing Zhang et al. Fedcp: Separating feature information for personalized federated learning via conditional policy. In Proc. KDD, page 1, Long Beach, CA, USA, 2023. ACM. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 173 + ], + "type": "text", + "content": "[71] Jie Zhang et al. Towards data-independent knowledge transfer in model-heterogeneous federated learning. IEEE Trans. Computers, 72(10):2888-2901, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "type": "text", + "content": "[72] Jianqing Zhang et al. Fedtgp: Trainable global prototypes with adaptive-margin-enhanced contrastive learning for data and model heterogeneity in federated learning. In Proc. AAAI, pages 16768-16776, Vancouver, Canada, 2024. 1. 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "type": "text", + "content": "[73] Lan Zhang et al. Fedzkt: Zero-shot knowledge transfer towards resource-constrained federated learning with heterogeneous on-device models. In Proc. ICDCS, pages 928-938, virtual, 2022. IEEE. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "type": "text", + "content": "[74] Zhilu Zhang and Mert R. Sabuncu. Generalized cross entropy loss for training deep neural networks with noisy labels. In Proc. NeurIPS, pages 8792-8802, Montreal, Canada, 2018. Curran Associates Inc. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 321, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 294, + 342 + ], + "type": "text", + "content": "[75] Hangyu Zhu et al. Federated learning on non-iid data: A survey. Neurocomputing, 465:371-390, 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 343, + 294, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 294, + 375 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 294, + 375 + ], + "type": "text", + "content": "[76] Zhuangdi Zhu et al. Data-free knowledge distillation for heterogeneous federated learning. In Proc. ICML, pages 12878-12889, virtual, 2021. PMLR. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 377, + 294, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 294, + 410 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 294, + 410 + ], + "type": "text", + "content": "[77] Zhuangdi Zhu et al. Resilient and communication efficient learning for heterogeneous federated systems. In Proc. ICML, pages 27504-27526, virtual, 2022. PMLR. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "type": "text", + "content": "[78] Qi Zhuang, Meng Lei, Chen Zitan, Hu Han, Lin Hui, and Meng Xiangxu. Cross-silo prototypical calibration for federated learning with non-iid data. In Proceedings of the 31st ACM International Conference on Multimedia, pages 3099-3107, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "type": "text", + "content": "[79] Qi Zhuang, Meng Lei, et al. Cross-silo feature space alignment for federated learning on clients with imbalanced data. In The 39th Annual AAAI Conference on Artificial Intelligence (AAAI-25), pages 19986-19994, 2025. 1" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1324" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Feed-Forward SceneDINO for Unsupervised Semantic Scene Completion/0cb8dcd6-d20d-4d18-a810-629c106f3cdd_content_list.json b/2025/Feed-Forward SceneDINO for Unsupervised Semantic Scene Completion/0cb8dcd6-d20d-4d18-a810-629c106f3cdd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0bf0fd9d8a8014ca878a5d31c67980836c15b336 --- /dev/null +++ b/2025/Feed-Forward SceneDINO for Unsupervised Semantic Scene Completion/0cb8dcd6-d20d-4d18-a810-629c106f3cdd_content_list.json @@ -0,0 +1,1892 @@ +[ + { + "type": "text", + "text": "Feed-Forward SceneDINO for Unsupervised Semantic Scene Completion", + "text_level": 1, + "bbox": [ + 130, + 130, + 866, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Aleksandar Jevtic*1", + "bbox": [ + 217, + 179, + 377, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Oliver Hahn", + "bbox": [ + 161, + 198, + 272, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Christian Rupprecht", + "bbox": [ + 307, + 198, + 478, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ TU Munich", + "bbox": [ + 135, + 218, + 215, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ TU Darmstadt", + "bbox": [ + 236, + 219, + 331, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3University of Oxford", + "bbox": [ + 352, + 219, + 486, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Christoph Reich* 1,2,4,5", + "bbox": [ + 413, + 179, + 591, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Eich\\*1,2,4,5 F6 Stefan Roth 2,5,6", + "bbox": [ + 514, + 179, + 642, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Felix Wimbauer1,4", + "bbox": [ + 627, + 180, + 779, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4MCML 5ELIZA", + "bbox": [ + 511, + 218, + 633, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Daniel Cremers $^{1,4,5}$", + "bbox": [ + 678, + 198, + 834, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{6}$ hessian.AI *equal contribution", + "bbox": [ + 653, + 219, + 861, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://visinf.github.io/scenedino", + "bbox": [ + 364, + 236, + 630, + 250 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/262c286345295238ebd432bdf2ecccf1e0df6b5da244a4e6a3c87aefed1f9aa2.jpg", + "image_caption": [ + "Single Input Image" + ], + "image_footnote": [], + "bbox": [ + 96, + 273, + 300, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "SceneDINO", + "bbox": [ + 305, + 295, + 387, + 306 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bcd377efa08022c7bda5c5b9514b60b4f3f97d03895267ac41ed2035f74e99a0.jpg", + "image_caption": [ + "3D Feature Field", + "Figure 1. SceneDINO overview. Given a single input image (left), SceneDINO predicts both 3D scene geometry and 3D features in the form of a feature field (middle) in a feed-forward manner, capturing the structure and semantics of the scene. Unsupervised distillation and clustering of SceneDINO's feature space leads to unsupervised semantic scene completion predictions (right)." + ], + "image_footnote": [], + "bbox": [ + 410, + 263, + 599, + 339 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Distill & Cluster", + "bbox": [ + 601, + 295, + 700, + 306 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bf9fd12e6841fe47ba904bc33b9a2eb12f5ac84965bf06620b42ef234550e8e4.jpg", + "image_caption": [ + "SSC Prediction" + ], + "image_footnote": [], + "bbox": [ + 710, + 263, + 901, + 339 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 428, + 326, + 444 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Semantic scene completion (SSC) aims to infer both the 3D geometry and semantics of a scene from single images. In contrast to prior work on SSC that heavily relies on expensive ground-truth annotations, we approach SSC in an unsupervised setting. Our novel method, SceneDINO, adapts techniques from self-supervised representation learning and 2D unsupervised scene understanding to SSC. Our training exclusively utilizes multi-view consistency self-supervision without any form of semantic or geometric ground truth. Given a single input image, SceneDINO infers the 3D geometry and expressive 3D DINO features in a feed-forward manner. Through a novel 3D feature distillation approach, we obtain unsupervised 3D semantics. In both 3D and 2D unsupervised scene understanding, SceneDINO reaches state-of-the-art segmentation accuracy. Linear probing our 3D features matches the segmentation accuracy of a current supervised SSC approach. Additionally, we showcase the domain generalization and multi-view consistency of SceneDINO, taking the first steps towards a strong foundation for single image 3D scene understanding.", + "bbox": [ + 88, + 460, + 485, + 765 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 797, + 220, + 813 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Understanding the geometry and semantics of 3D scenes from image observations is a fundamental computer vision task with broad applications in robotics [26], autonomous driving [47, 66], medical image analysis [18, 113], and civil engineering [70]. The Semantic Scene Completion (SSC)", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "task unifies 3D geometry and semantic prediction from limited image observations [64, 89, 96]. Recent progress in SSC has been primarily driven by utilizing supervised learning [38, 88, 96]. However, acquiring large-scale 3D annotations is highly labor-intensive [66]. While significant resources have been invested in collecting human annotations for 2D tasks [53, 85], annotating similar amounts of data in 3D remains unapproached. This motivates approaching SSC without the need for manually annotated data.", + "bbox": [ + 511, + 429, + 906, + 565 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing SSC approaches rely on ground-truth semantic annotations and frequently utilize additional supervision from LiDAR scans [38, 46, 74, 96]. In contrast, we are the first to approach SSC in a fully unsupervised setting, i.e. without task supervision or other supervised components. In particular, we aim to approach SSC from a single image without relying on any human annotations, only learning from unlabeled multi-view images using self-supervision. This setting is extremely challenging for two reasons: first, the human-defined nature of semantic taxonomies is ambiguous, and second, a single image only provides a partial observation of the scene with many invisible areas. We take inspiration from recent advances in self-supervised learning (SSL) of 2D representations and 3D reconstruction. 2D SSL representations, such as from DINO [11], have been shown effective for 2D unsupervised scene understanding [32, 33, 104]. 3D reconstruction approaches successfully leveraged SSL from multi-view data to infer dense 3D geometry from a single image [34, 108].", + "bbox": [ + 511, + 566, + 908, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we present SceneDINO, to the best of our knowledge, the first approach for unsupervised semantic scene completion. Trained using 2D SSL features", + "bbox": [ + 511, + 854, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6784", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "from DINO [11] and multi-view self-supervision [108], SceneDINO predicts both 3D geometry and 3D features from a single image during inference in a feed-forward manner. Our general 3D feature representations enable us to approach unsupervised 3D scene understanding. Harnessing our expressive 3D features, we propose a novel 3D feature distillation approach for obtaining unsupervised semantic predictions in 3D. While we focus on the task of unsupervised SSC, SceneDINO's features are general, offering a foundation for different 3D scene-understanding tasks by building on our 3D feature field.", + "bbox": [ + 89, + 90, + 480, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, we make the following contributions: (i) We introduce SceneDINO, the first approach predicting dense 3D geometry and expressive 3D features in a feedforward manner from a single image. (ii) We effectively distill SceneDINO's feature field representation in 3D, obtaining unsupervised semantic predictions. (iii) We demonstrate the first fully unsupervised SSC results. We build a simple, yet competitive unsupervised SSC baseline, lifting unsupervised 2D semantic predictions. Our SceneDINO approach outperforms this SSC baseline in unsupervised SSC as well as established 2D approaches in 2D semantic segmentation. (iv) Finally, we also showcase the domain generalization ability and multi-view consistency of SceneDINO.", + "bbox": [ + 88, + 258, + 480, + 454 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 470, + 232, + 486 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Single-image scene reconstruction. Estimating 3D geometry from image observations is a fundamental task in computer vision and has been studied for decades [37]. Traditional approaches, such as structure from motion [90], as well as recent neural radiance fields (NeRFs) [75], perform scene reconstruction using multiple images, as reviewed by multiple surveys [35, 109, 120]. Recently, estimating dense 3D geometry from a single image have been approached [8, 34, 81, 86, 97, 103, 108, 114]. Unlike monocular depth estimation [76], these approaches predict the depth for visible and occluded regions, reconstructing a complete scene. Behind the Scenes (BTS) [108] introduced an approach for unsupervised single-image scene reconstruction using multi-view self-supervision, which infers dense 3D geometry in a feed-forward manner. Our approach extends BTS by additionally lifting self-supervised features into 3D for unsupervised 3D scene understanding.", + "bbox": [ + 89, + 502, + 482, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Semantic scene completion (SSC), also known as 3D semantic occupancy prediction, aims to jointly estimate the 3D geometry and semantics of a scene [63, 64, 96, 118]. Initial approaches used 3D semantic and geometric annotations and addressed indoor scenes [6, 13, 58-60, 68, 117], outdoor scenes with LiDAR [16, 62, 87, 88, 110], or both domains [8, 74]. Using birds-eye views has been proven effective for SSC [45, 65, 100]. To overcome the need for 3D annotations, approaches for using 2D annotations", + "bbox": [ + 89, + 763, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "have been proposed [38, 46, 82]. While SelfOcc [46] and RenderOcc [82] use multiple inference views, S4C [38] performs single-image SSC. In particular, S4C [38] employs a supervised 2D model and lifts 2D multi-view semantic predictions into 3D. In contrast to using 2D annotations, GaussTR [49] uses 2D foundation models for SSC and multiple views during inference. However, GaussTR relies on heavily supervised foundation models, including SAM [53] and Metric3Dv2 [43], and uses weak supervision from image/text pairs. To the best of our knowledge, there is no method for approaching SSC without the need for any ground-truth annotations. Our work presents the first unsupervised SSC approach, utilizing lifted SSL features and a single RGB input image for inference.", + "bbox": [ + 511, + 90, + 903, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-supervised representation learning (SSL) aims to extract general features from data without annotations, facilitating various downstream tasks such as segmentation [24]. Recent SSL methods, often based on Vision Transformers (ViTs) [23], leverage clustering [2, 9, 10, 48, 61], masked modeling [20, 29, 40, 77, 107], contrastive learning [3, 12, 14, 39, 41, 42], or negative-free [4, 5, 11, 28, 80] pretext tasks [22, 79] for large-scale training. State-of-the-art models, e.g., DINO [11], produce semantically rich, dense features, driving recent advances in 2D unsupervised scene understanding [33, 104]. We here aim to bring expressive features from DINO [11, 80] to 3D for SSC.", + "bbox": [ + 511, + 305, + 903, + 487 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2D-to-3D feature lifting. The expressiveness of 2D visual representations has motivated lifting 2D features into 3D [94, 111]. Existing approaches utilize multi-view 2D features for 3D feature lifting [30, 44, 50, 54, 73, 83, 93, 94, 98, 101, 102, 106, 111, 116]. Lifting 2D features is effective in various tasks, including few-shot semantic occupancy prediction [111], and refining 2D representations [116]. However, existing feature-lifting approaches fit to a single scene [50, 54, 93, 94, 101, 102, 111, 116], require RGB-D inputs [30, 44, 73, 98, 106], or work on 3D point cloud inputs [83]. The only feed-forward approaches that use RGB inputs and lift 2D features, which we are aware of, are GaussTR [49] and MVSplat360 [15]. However, both approaches utilize multiple input images during inference, and MVSplat360 [15] only predicts low-dimensional feature representations, which are not suitable for unsupervised scene understanding. In contrast, we propose the first feedforward approach for inferring lifted high-dimensional and rich 3D features using a single input image.", + "bbox": [ + 511, + 489, + 903, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2D unsupervised semantic segmentation partitions images automatically into semantically meaningful regions without any form of human annotations. Early deep learning-based methods [17, 36, 48] approach the problem via representation learning. Leveraging SSL features from DINO as an inductive prior, STEGO [33] distills the feature representation into a lower-dimensional space for unsupervised probing. Building up on STEGO, subsequent methods", + "bbox": [ + 511, + 779, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6785", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7a32d6794c76d05a04ce2206d89ab5f3c7736bf7547632ad0614e5dff0fc093d.jpg", + "image_caption": [ + "Figure 2. SceneDINO architecture, rendering, and training. (a) Inference: Given a single input image $\\mathbf{I}_0$ during inference, a 2D encoder-decoder $\\xi$ produces the embedding $\\mathbf{E}$ from which the local embedding $\\mathbf{e}_{\\mathbf{u}}$ is interpolated. The MLP encoder $\\phi$ takes in $\\mathbf{e}_{\\mathbf{u}}$ and 3D position $\\mathbf{x}_i$ , and predicts both the density $\\sigma_{\\mathbf{x}_i}$ and the 3D feature $f_{\\mathbf{x}_i}$ . Using a lightweight unsupervised segmentation head $h$ , we can obtain semantic predictions $p_{\\mathbf{x}_i}$ using $f_{\\mathbf{x}_i}$ . (b) Rendering: Our feature field allows for volume rendering by shooting rays through it, yielding depth $\\hat{d}$ and $\\hat{f}$ in 2D. Color $c_i$ is sampled from an another view ( $e.g., \\mathbf{I}_1$ ) using $\\mathbf{u}_s$ and rendered to obtain the reconstructed color $\\hat{c}$ . (c) Multi-view training: We render 2D views (features & images) from our feature field and reconstruct the training views." + ], + "image_footnote": [], + "bbox": [ + 106, + 90, + 346, + 255 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/30dd532230d021ee8257fae00817fed36b4bbce0d48b19a7f7209dd2246ec836.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 102, + 609, + 255 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6072acde5b5bbd7b0ebfd347875ac119c96457bbc4e47eaa0ac16f0ad1b0fef0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 650, + 88, + 883, + 255 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "[31, 51, 92, 95] propose enhancements to the distillation. Our approach follows the idea of STEGO [33], extending it to 3D and integrating feature distillation using our 3D feature field to build the first unsupervised SSC approach.", + "bbox": [ + 88, + 363, + 480, + 425 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Unsupervised Semantic Scene Completion", + "text_level": 1, + "bbox": [ + 89, + 438, + 465, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We approach semantic scene completion (SSC) without any form of manual supervision. To this end, we first describe SceneDINO, predicting 3D geometry and expressive 3D features from a single image in a feed-forward manner (Sec. 3.1), and SceneDINO's multi-view training (Sec. 3.2). Next, we present our 3D feature distillation approach to obtain unsupervised 3D semantic predictions (Sec. 3.3). An overview of our full pipeline, including inference, rendering, and multi-view self-supervision, is provided in Fig. 2.", + "bbox": [ + 89, + 464, + 482, + 601 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Notation. Let $\\mathbf{I}_0\\in [0,1]^{3\\times \\mathrm{H}\\times \\mathrm{W}}$ be a single RGB input image (for both training & inference) with corresponding pose $T_0\\in \\mathbb{R}^{4\\times 4}$ and projection matrix $K_{0}\\in \\mathbb{R}^{3\\times 4}$ . For training, let $(\\mathbf{I}_v,T_v,K_v)$ with $v\\in \\{1,2,\\dots ,n\\}$ , be $n$ additional views for multi-view self-supervision. Assuming a pinhole camera model, any 3D point $\\mathbf{x}\\in \\mathbb{R}^3$ in world coordinates can be projected onto the image plane of view $v$ and the input view $v = 0$ with the perspective projection $\\pi_v(\\mathbf{x})$ .", + "bbox": [ + 89, + 602, + 482, + 728 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. SceneDINO", + "text_level": 1, + "bbox": [ + 89, + 737, + 218, + 752 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a single input image $\\mathbf{I}_0$ , SceneDINO represents the dense geometric structure and features of a scene as a continuous mapping from world coordinates $\\mathbf{x} \\in \\mathbb{R}^3$ to a volumetric density $\\sigma_{\\mathbf{x}} \\in \\mathbb{R}_{+}$ and a feature $f_{\\mathbf{x}} \\in \\mathbb{R}^{\\mathrm{D}}$ . This continuous output representation is often called a feature field. While SceneDINO could represent any feature space, we aim for expressive SSL features from DINO [11, 80].", + "bbox": [ + 88, + 758, + 482, + 866 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Architecture & feature field inference. Our SceneDINO architecture comprises two main parts: a 2D encoder-", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "decoder $\\xi$ and an MLP decoder (cf. Fig. 2a), following BTS [108]. $\\xi$ takes in $\\mathbf{I}_0$ and produces a per-pixel embedding $\\mathbf{E} \\in \\mathbb{R}^{\\mathrm{D}_{\\mathbf{E}} \\times \\mathrm{H} \\times \\mathrm{W}}$ with $\\mathrm{D}_{\\mathbf{E}}$ dimensions. Intuitively, every spatial element of $\\mathbf{E}$ represents a camera ray through a pixel, capturing both local geometry and features.", + "bbox": [ + 511, + 363, + 906, + 438 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To infer the feature at a 3D position $\\mathbf{x}$ , we employ a two-layer MLP decoder $\\phi$ (cf. Fig. 2a). Given a position $\\mathbf{x}$ within the camera frustum, we project $\\mathbf{x}$ into the camera plane, obtaining the pixel location $\\mathbf{u} = \\pi_0(\\mathbf{x})$ . We query $\\mathbf{E}$ at the position $\\mathbf{u}$ using bilinear interpolation, obtaining the local embedding $\\mathbf{e_u}$ . Based on the embedding $\\mathbf{e_u}$ , the pixel position $\\mathbf{u}$ , and the distance $d_{\\mathbf{x}} \\in \\mathbb{R}_{+}$ of $\\mathbf{x}$ to the camera, we obtain the density $\\sigma_{\\mathbf{x}}$ and feature prediction $f_{\\mathbf{x}}$ as", + "bbox": [ + 511, + 439, + 906, + 560 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\sigma_ {\\mathbf {x}}, f _ {\\mathbf {x}}\\right) = \\phi \\left(\\mathbf {e} _ {\\mathbf {u}}, \\gamma (\\mathbf {u}, d _ {\\mathbf {x}})\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 565, + 903, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\gamma$ denotes a positional encoding [75].", + "bbox": [ + 511, + 587, + 805, + 603 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Feature, depth & color volume rendering. SceneDINO predicts a continuous feature field from a single image. This representation can be used to render features and depth in 2D from an arbitrary viewpoint (cf. Fig. 2b), following the discretization strategy of Max et al. [72]. Given a viewpoint $(T,K)$ , we sample L points $\\mathbf{x}_i$ along the ray through pixel $\\mathbf{u}_r$ , with distance $\\delta_i$ between $\\mathbf{x}_i$ and $\\mathbf{x}_{i+1}$ . Based on the volumetric densities $\\sigma_{\\mathbf{x}_i}$ (cf. Eq. 1), we can compute the probabilities $\\alpha_i$ of the ray ending between $\\mathbf{x}_i$ and $\\mathbf{x}_{i+1}$ , and accumulate these into $V_i$ , the probability of $\\mathbf{x}_i$ being visible:", + "bbox": [ + 511, + 604, + 906, + 757 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nV _ {i} = \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\quad \\text {w i t h} \\alpha_ {i} = 1 - \\exp (- \\sigma_ {\\mathbf {x} _ {i}} \\delta_ {i}). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 762, + 903, + 805 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Using $V_{i}$ and $\\alpha_{i}$ , we render depth $\\tilde{d}_{\\mathbf{u}_r}$ and feature $\\tilde{f}_{\\mathbf{u}_r}$ from the estimated features $f_{\\mathbf{x}_i}$ from Eq. (1) and distances $d_{\\mathbf{x}_i}$ to $\\mathbf{x}_i$ onto the image plane at position $\\mathbf{u}_r$ as", + "bbox": [ + 511, + 810, + 906, + 858 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {f} _ {\\mathbf {u} _ {r}} = \\sum_ {i = 1} ^ {\\mathrm {L}} V _ {i} \\alpha_ {i} f _ {\\mathbf {x} _ {i}} \\quad \\tilde {d} _ {\\mathbf {u} _ {r}} = \\sum_ {i = 1} ^ {\\mathrm {L}} V _ {i} \\alpha_ {i} d _ {\\mathbf {x} _ {i}}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 862, + 906, + 904 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6786", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The differentiability of this rendering process enables us to self-supervise SceneDINO using multi-view images and their 2D feature representations (e.g., from DINO [11]). SceneDINO predicts 3D geometry and features, but does not predict color as we focus on semantic downstream tasks. To obtain color for image reconstruction during training, we follow the color sampling approach of BTS [108].", + "bbox": [ + 89, + 90, + 483, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. 3D feature field training", + "text_level": 1, + "bbox": [ + 89, + 205, + 313, + 222 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We train SceneDINO using multi-view self-supervision (cf. Fig. 2c), aiming to obtain an expressive and view-consistent feature field without the need for any form of manual annotations. For self-supervision, we sample $n + 1$ views $\\mathbf{I}_v$ with camera parameters $T_v, K_v$ from the data and obtain dense 2D target features $\\mathbf{F}_v$ from a self-supervised ViT (e.g., DINO [11]). Note that the 2D features entail a resolution of $\\mathbf{F}_v \\in \\mathbb{R}^{\\mathrm{D} \\times \\frac{\\mathrm{H}}{p} \\times \\frac{\\mathrm{W}}{p}}$ , due to the ViT patch size $p$ . The set of training views and features $\\mathbb{V} = \\{( \\mathbf{I}_v, T_v, K_v, \\mathbf{F}_v) | v = 0, \\dots, n \\}$ is randomly partitioned into two subsets $\\mathbb{V}_{\\text{source}}$ and $\\mathbb{V}_{\\text{target}}$ . Training reconstructs the views $\\mathbb{V}_{\\text{target}}$ using the views of $\\mathbb{V}_{\\text{source}}$ . In practice, we use a randomly sampled set of image patches that align with the ViT patches instead of the full image. In the following, we still refer to images and the full image resolution for the sake of brevity.", + "bbox": [ + 89, + 227, + 483, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Image reconstruction. We aim to learn the geometry of our feature field via multi-view photometric consistency. In particular, for every image $\\mathbf{I}_t\\in \\mathbb{V}_{\\mathrm{target}}$ we derive a reconstructed image $\\hat{\\mathbf{I}}_{t,s}$ from every view $s$ in $\\mathbb{V}_{\\mathrm{source}}$ using differentiable rendering (cf. Eq. 3) and color sampling [108]. Equipped with both the reconstructed image $\\hat{\\mathbf{I}}_{t,s}$ and the target image $\\mathbf{I}_t$ , we compute our photometric loss per view as", + "bbox": [ + 89, + 459, + 483, + 568 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {p}} = \\min _ {\\mathbf {I} _ {s} \\in \\mathbb {V} _ {\\text {s o u r c e}}} \\left(\\lambda_ {1} \\mathcal {L} _ {1} \\left(\\mathbf {I} _ {t}, \\hat {\\mathbf {I}} _ {t, s}\\right) + \\lambda_ {\\text {S S I M}} \\mathcal {L} _ {\\text {S S I M}} \\left(\\mathbf {I} _ {t}, \\hat {\\mathbf {I}} _ {t, s}\\right)\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 577, + 483, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We only consider the minimum loss across the views in $\\mathbb{V}_{\\mathrm{source}}$ , in practice across patches. The scalars $\\lambda_{1}$ and $\\lambda_{\\mathrm{SSIM}}$ weight the absolute error $\\mathcal{L}_1$ and the SSIM loss $\\mathcal{L}_{\\mathrm{SSIM}}$ [105].", + "bbox": [ + 89, + 613, + 482, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To regularize our 3D geometry, we impose smoothness using an edge-aware smoothness loss [27]. We estimate 2D depth maps $\\tilde{\\mathbf{d}}_t$ for views in $\\mathbb{V}_{\\mathrm{target}}$ , using Eq. 3. From $\\tilde{\\mathbf{d}}_t$ , we obtain the inverse mean-normalized depths $\\tilde{\\mathbf{d}}_t^*$ and compute the edge-aware smoothness loss $\\mathcal{L}_{\\mathrm{s}}$ per view as", + "bbox": [ + 89, + 657, + 483, + 736 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {s}} = \\left| \\nabla_ {x} \\tilde {\\mathbf {d}} _ {t} ^ {*} \\right| e ^ {- \\left| \\nabla_ {x} \\mathbf {I} _ {t} \\right|} + \\left| \\nabla_ {y} \\tilde {\\mathbf {d}} _ {t} ^ {*} \\right| e ^ {- \\left| \\nabla_ {y} \\mathbf {I} _ {t} \\right|}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 743, + 482, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\nabla_{x}$ and $\\nabla_{y}$ denote the first spatial derivatives.", + "bbox": [ + 89, + 772, + 441, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Feature reconstruction. We learn a multi-view consistent and expressive 3D feature field using the 2D features $\\mathbf{F}_t$ from $\\mathbb{V}_{\\mathrm{target}}$ . As we aim to learn a high-resolution (continuous) feature field, we render 2D features using Eq. 3 at the full image resolution $\\hat{\\mathbf{F}}_t \\in \\mathbb{R}^{D \\times H \\times W}$ . To compensate for", + "bbox": [ + 89, + 790, + 483, + 867 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/14640c21cf9305b7701c5a14a164adcf218d304d70e5071325f4210f6cc5d0cf.jpg", + "image_caption": [ + "Figure 3. 3D feature distillation. Given an input image, SceneDINO predicts a 3D feature field. 3D features $f_{\\mathbf{X}}$ are sampled from the feature field. For $f_{\\mathbf{X}}$ , we obtain $f_{\\mathbf{Y}_{k\\mathrm{NN}}}$ and $f_{\\mathbf{Y}_{kr\\mathrm{and}}}$ from the feature buffer. The segmentation head $h$ distills the features into a low-dimensional space and is trained using $\\mathcal{L}_{\\mathrm{dist}}$ ." + ], + "image_footnote": [], + "bbox": [ + 516, + 87, + 905, + 223 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the reduced spatial dimension of $\\mathbf{F}_t$ , we employ the down-sampler $\\psi$ proposed by Fu et al. [25] to our rendered features $\\hat{\\mathbf{F}}_t$ . While current 2D SSL features capture semantics, they lack multi-view consistency, i.a., due to positional encodings used in ViTs [112], leading to different features for identical visual content at two distinct positions in an image. As we aim for multi-view consistency, we compensate for this by learning a constant decomposition $\\overline{\\mathbf{F}} \\in \\mathbb{R}^{\\mathrm{D} \\times \\frac{\\mathrm{H}}{p} \\times \\frac{\\mathrm{W}}{p}}$ of features induced by positional encodings. Our feature loss is defined per view as", + "bbox": [ + 511, + 323, + 906, + 474 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {f}} = 1 - \\cos - \\sin \\left(\\mathbf {F} _ {t}, \\psi \\left(\\hat {\\mathbf {F}} _ {t}\\right) + \\overline {{\\mathbf {F}}}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 486, + 903, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where cos-sim is the cosine similarity between two features.", + "bbox": [ + 511, + 513, + 903, + 529 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As image edges correlate with semantic edges and to further impose consistency, we regularize the rendered features $\\hat{\\mathbf{F}}_t$ using an edge-aware smoothness loss per view", + "bbox": [ + 511, + 529, + 903, + 575 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {f s}} = \\left| \\nabla_ {x} \\hat {\\mathbf {F}} _ {t} \\right| e ^ {- \\left| \\nabla_ {x} \\mathbf {I} _ {t} \\right|} + \\left| \\nabla_ {y} \\hat {\\mathbf {F}} _ {t} \\right| e ^ {- \\left| \\nabla_ {y} \\mathbf {I} _ {t} \\right|}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 583, + 903, + 603 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our final loss for training SceneDINO is a weighted sum of the photometric loss, the feature loss, and both smoothness losses, $\\mathcal{L}_{\\mathrm{SceneDINO}} = \\lambda_{\\mathrm{p}}\\mathcal{L}_{\\mathrm{p}} + \\lambda_{\\mathrm{s}}\\mathcal{L}_{\\mathrm{s}} + \\lambda_{\\mathrm{f}}\\mathcal{L}_{\\mathrm{f}} + \\lambda_{\\mathrm{fs}}\\mathcal{L}_{\\mathrm{fs}}$ , averaged over all pixels, features, and views.", + "bbox": [ + 511, + 612, + 905, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. 3D feature distillation for unsupervised SSC", + "text_level": 1, + "bbox": [ + 511, + 681, + 885, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the expressive feature field representation, we aim to obtain unsupervised semantic predictions for SSC. While naive $k$ -means [69, 71] can yield meaningful pseudo semantics, distilling features into a lower-dimensional space has been shown to be more effective in 2D semantic segmentation [33, 55]. To this end, we present a novel 3D feature distillation approach (cf. Fig. 3). We train a pointwise segmentation head $h$ , mapping $f_{\\mathbf{x}} \\in \\mathbb{R}^{\\mathrm{D}}$ to a lower-dimensional distilled representation $z_{\\mathbf{x}} \\in \\mathbb{R}^{\\mathrm{K}}$ , with $\\mathrm{K} \\ll \\mathrm{D}$ . The resulting distilled space is clustered to obtain pseudosemantic predictions $p_{\\mathbf{x}} \\in [0,1]^{\\mathrm{C}}$ , with C pseudo classes.", + "bbox": [ + 511, + 704, + 905, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Existing work in 2D unsupervised semantic segmentation has shown that SSL feature correspondence captures", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1Note, camera poses can be obtained using unsupervised visual SLAM [7], strictly adhering to the fully unsupervised setting.", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6787", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "semantic class co-occurrence [33]. This correspondence between two batches of $\\mathbf{N}$ sample points $\\mathbf{X} = [\\mathbf{x}_1, \\dots, \\mathbf{x}_N]$ and $\\mathbf{Y} = [\\mathbf{y}_1, \\dots, \\mathbf{y}_N]$ can be expressed by pairwise feature similarity $S_{i,j} = \\cos -\\sin (f_{\\mathbf{x}_i}, f_{\\mathbf{y}_j}) \\in [-1, 1]$ . Similarly, we can express the correspondence in the distilled feature space by $S_{i,j}^h = \\cos -\\sin (h(f_{\\mathbf{x}_i}), h(f_{\\mathbf{y}_j})) \\in [-1, 1]$ . We describe the sampling of the $\\mathbf{x}_i$ and $\\mathbf{y}_j$ below.", + "bbox": [ + 89, + 90, + 483, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Feature distillation. We aim to distill features such that similar features align while dissimilar features are separated. To this end, we use the contrastive correlation loss $\\mathcal{L}_{\\mathrm{corr}}$ , introduced by STEGO [33] and defined as", + "bbox": [ + 89, + 200, + 483, + 261 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o r r}} \\left(f _ {\\mathbf {X}}, f _ {\\mathbf {Y}}, b\\right) = - \\sum_ {i, j} \\left(S _ {i, j} - b\\right) \\max \\left(S _ {i, j} ^ {h}, 0\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 271, + 483, + 304 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f_{\\mathbf{X}}$ , $f_{\\mathbf{Y}}$ are the features of the two sample batches. This loss pushes $S_{i,j}^{h}$ towards 1 in case $S_{i,j}$ exceeds the threshold $b$ . Otherwise, $\\mathcal{L}_{\\mathrm{corr}}$ pushes the $S_{i,j}^{h}$ below 0.", + "bbox": [ + 89, + 310, + 482, + 358 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The correlation loss $\\mathcal{L}_{\\mathrm{corr}}$ requires informative pairs of sampled features, balancing attractive and repulsive signals. Following STEGO [33], we consider three different relations: (1) feature pairs from the same image $(f_{\\mathbf{X}}, f_{\\mathbf{X}})$ , (2) feature pairs from an image and its $k$ -nearest neighbors in feature space $(f_{\\mathbf{X}}, f_{\\mathbf{Y}_{k\\mathrm{NN}}})$ , and (3) feature pairs from an image and a randomly sampled other image $(f_{\\mathbf{X}}, f_{\\mathbf{Y}_{\\mathrm{rand}}})$ . Note that each pair is obtained from SceneDINO's 3D feature field, see below. Equipped with the three feature sample pairs, we compute the full distillation loss as", + "bbox": [ + 89, + 358, + 483, + 508 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {d i s t}} = \\lambda_ {\\mathrm {s e l f}} \\mathcal {L} _ {\\mathrm {c o r r}} \\left(f _ {\\mathbf {X}}, f _ {\\mathbf {X}}, b _ {\\mathrm {s e l f}}\\right) \\\\ + \\lambda_ {k \\mathrm {N N}} \\mathcal {L} _ {\\text {c o r r}} \\left(f _ {\\mathbf {X}}, f _ {\\mathbf {Y} _ {k \\mathrm {N N}}}, b _ {k \\mathrm {N N}}\\right) \\tag {9} \\\\ + \\lambda_ {\\text {r a n d}} \\mathcal {L} _ {\\text {c o r r}} \\left(f _ {\\mathbf {X}}, f _ {\\mathbf {Y} _ {\\text {r a n d}}}, b _ {\\text {r a n d}}\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 518, + 480, + 569 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{\\mathrm{self}}$ , $\\lambda_{k\\mathrm{NN}}$ , and $\\lambda_{\\mathrm{rand}}$ denote the scalar loss weights. $b_{\\mathrm{self}}$ , $b_{k\\mathrm{NN}}$ , and $b_{\\mathrm{rand}}$ are the contrastive thresholds.", + "bbox": [ + 89, + 580, + 482, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Feature sampling in 3D. While obtaining feature pairs using 2D rendered features is straightforward [33], we aim to take advantage of our learned 3D geometry of the scene. To this end, we introduce a novel 3D feature sampling approach for the distillation loss $\\mathcal{L}_{\\mathrm{dist}}$ from Eq. (9). Our goal is to sample features both similar and dissimilar in terms of the encoded semantic concept, which should capture rich semantics as well as different semantic concepts.", + "bbox": [ + 89, + 613, + 483, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "First, we obtain all G visible 3D surface points $\\mathbf{V} \\in \\mathbb{R}^{3 \\times \\mathrm{G}}$ and their depth $d_{\\mathbf{V}} \\in \\mathbb{R}_{+}^{\\mathrm{G}}$ from the camera. To sample points that cover different semantic concepts, we use depth as a cue and sample different depth ranges. In particular, we sort the surface points $\\mathbf{V}$ based on $d_{\\mathbf{V}}$ . The sorted surface points $\\hat{\\mathbf{V}}$ are partitioned into M equally-sized chunks; we uniformly sample a single 3D point from each chunk, resulting in M center points $\\mathbf{X} \\in \\mathbb{R}^{3 \\times \\mathrm{M}}$ .", + "bbox": [ + 89, + 734, + 483, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Equipped with the center points $\\mathbf{X}$ , we aim to extract rich semantic features from the feature field. While we could just obtain the features for $\\mathbf{X}$ , we query positions in the", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7dd58bbc0b5b72c0032e950414ba5056d6f42570149f97ded9eebfe5bfeac687.jpg", + "image_caption": [ + "Figure 4. 3D feature sampling. We first sample a center point $\\mathbf{X}_i$ from all visible surface points. Further points are sampled within the radius $r$ around the center point $\\mathbf{X}_i$ . Sampled points with sufficient density are accepted; otherwise rejected. The accepted points are used to obtain the feature batch $f_{\\mathbf{X}}$ ." + ], + "image_footnote": [], + "bbox": [ + 584, + 85, + 834, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "neighborhood of $\\mathbf{X}$ to increase semantic richness and better capture the 3D structure of the scene for distillation. In particular, for each center point, we randomly sample a point within a radius of $r = 0.5\\mathrm{m}$ . To account for samples falling into unoccupied regions in our feature field, we only keep samples with a sufficient density $\\sigma > 0.5$ . We repeat this sampling process until we obtain $N$ valid samples per center point. Using these samples, we query our feature field, resulting in a feature batch $f_{\\mathbf{X}} \\in \\mathbb{R}^{D \\times N}$ for each of the G center points in each scene (cf. Fig. 4).", + "bbox": [ + 511, + 306, + 905, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To obtain $f_{\\mathbf{Y}_{k\\mathrm{NN}}}$ and $f_{\\mathbf{Y}_{\\mathrm{rand}}}$ , we utilize a feature buffer that efficiently stores the sampled features of multiple scenes. Given a new input image, we obtain $G$ feature batches $f_{\\mathbf{X}}$ as just described. For each $f_{\\mathbf{X}}$ , we randomly sample another feature batch from the buffer to obtain $f_{\\mathbf{Y}_{\\mathrm{rand}}}$ . To obtain $f_{\\mathbf{Y}_{k\\mathrm{NN}}}$ , we search in the feature buffer for the $k$ -nearest neighbors of $f_{\\mathbf{X}}$ , using the average feature of each batch. From these $k$ -nearest neighbors, we randomly pick a feature batch to obtain $f_{\\mathbf{Y}_{k\\mathrm{NN}}}$ and compute the distillation loss $\\mathcal{L}_{\\mathrm{dist}}$ . After repeating this process for each of the current $G$ feature batches, we add the current feature batches to the feature buffer and remove the oldest batches.", + "bbox": [ + 511, + 458, + 906, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unsupervised probing. To obtain semantics, we probe the distilled K-dim. feature space using $k$ -means [69, 71]. In particular, we update cluster centers $\\theta \\in \\mathbb{R}^{K\\times C}$ using cosine similarity-based mini-batch $k$ -means [91] during distillation. We compute $p_{\\mathbf{x}} = \\mathrm{softmax}(\\cos -\\mathrm{sim}(h(f_{\\mathbf{x}}),\\theta))$ to infer C pseudo semantic class predictions.", + "bbox": [ + 511, + 643, + 905, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 750, + 645, + 766 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate SceneDINO on SSC and compare it to a simple unsupervised SSC baseline (Sec. 4.1). We also report results for 2D unsupervised segmentation, including domain generalization results (Sec. 4.2). Finally, we explore multiview feature consistency (Sec. 4.3) and present an analysis of SceneDINO and our 3D distillation (Sec. 4.4).", + "bbox": [ + 511, + 775, + 905, + 866 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We train using KITTI-360 [66], composed of clips from a moving vehicle equipped with cameras. For", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6788", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "consistency, we follow S4C [38] by sampling eight views and using the dataset's camera poses. We further provide results with estimated poses. We also show experiments for training on RealEstate10k [119], composed of monocular videos. Here, we follow the setup of BTS [108], obtaining three views. If not noted differently, we report results obtained with training on KITTI-360. For SSC and 2D semantic segmentation validation, we use the SSCBench-KITTI-360 test split [64]. Cityscapes [19] and BDD100K [115] val are used for domain generalization results. To enable evaluation in 3D and 2D, we use the 19-class taxonomy of Cityscapes and perform 2D evaluation on Cityscapes, BDD100K, and KITTI-360 on 19 classes. For SSCBench, we combine classes to adhere to the 15 SSCBench classes.", + "bbox": [ + 89, + 90, + 480, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3D evaluation. Given our unsupervised setup, we predict pseudo-semantic classes that must be aligned with the ground truth for evaluation. We follow standard practice in 2D unsupervised semantic segmentation [17, 31, 33, 51, 92, 95] by applying Hungarian matching [57] to align our pseudo semantics. For validating the aligned semantics, we follow the standardized SSCBench [64] protocol and report both semantic performance using the mean Intersection-over-Union (mIoU) and geometric performance using IoU, precision, and recall. We report all metrics on SSCBench ranges $12.8\\mathrm{m}$ , $25.6\\mathrm{m}$ , and $51.2\\mathrm{m}$ .", + "bbox": [ + 89, + 308, + 480, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "2D evaluation. Following the established evaluation protocol in 2D unsupervised semantic segmentation [17, 31, 33, 51, 92, 95], we use the all-pixel accuracy (Acc) and mean Intersection-over-Union (mIoU) metrics. Likewise, in line with prior work, 2D segmentation predictions of all models are refined using a dense Conditional Random Field [56] before computing Acc and mIoU.", + "bbox": [ + 89, + 479, + 480, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Multi-view feature consistency evaluation. We aim to evaluate the multi-view consistency of our feature field. As we are not aware of any general feed-forward 3D feature field approach, we compare against 2D SSL models. To measure multi-view consistency in 2D, we use two video frames and estimate optical flow and occlusions with RAFT [99]. We backward warp 2D features of the second frame to the first. On the aligned features, we compute the feature similarity using absolute error $(\\mathrm{L}_1)$ , the Euclidean distance $(\\mathrm{L}_2)$ , and the cosine similarity, ignoring occlusions.", + "bbox": [ + 89, + 592, + 480, + 743 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We are not aware of any existing unsupervised SSC approach. To allow for comparisons, we construct a competitive baseline for unsupervised SSC. In particular, we train the S4C approach with unsupervised semantics of the established STEGO [33] approach. For 2D semantic segmentation, we use U2Seg [78] and STEGO [33] as established unsupervised baselines. Note that U2Seg is trained on ImageNet [21] and COCO [67] using STEGO pseudo-labels. We use STEGO [33] with DINO [11] (ViT-B/8), DINOv2 [80] (ViT-B/14), and FiT3D [116] (ViT", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. SSCBench-KITTI-360 results. Semantic results using mIoU and per class IoU, and geometric results using IoU, Precision, and Recall (all in %, ↑) on SSCBench-KITTI-360 test using three depth ranges. We compare our baseline S4C + STEGO to our SceneDINO. We report S4C as a 2D supervised baseline.", + "bbox": [ + 511, + 88, + 903, + 159 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/94b74e3667e9105f79b902d364b7faf83bc77eedc8cc9397174fff38e729b998.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodS4C + STEGOSceneDINO (Ours)S4C
SupervisionUnsupervised2D supervision
Range12.8 m25.6 m51.2 m12.8 m25.6 m51.2 m12.8 m25.6 m51.2 m
Semantic validation
mIoU10.539.266.6010.7610.018.0016.9413.9410.19
car18.5714.099.2221.2415.9411.2122.5818.6411.49
bicycle0.010.010.010.000.000.000.000.000.00
motorcycle0.000.000.000.000.000.000.000.000.00
truck0.110.040.020.000.000.007.514.372.12
other-v.0.010.050.020.000.000.000.000.010.06
person0.010.010.010.000.000.000.000.000.00
road61.9752.4738.1551.1049.1239.8269.3861.4648.23
sidewalk18.7420.9518.2120.2622.3118.9745.0337.1228.45
building14.7524.4417.8112.3318.2714.3226.3428.4821.36
fence1.410.200.111.910.900.589.706.373.64
vegetation15.8316.5811.3031.2225.5719.8535.7828.0421.43
terrain26.499.954.1723.2618.0215.2235.0322.8815.08
pole0.080.040.040.050.050.051.230.940.65
traffic-sign0.000.000.000.000.000.001.570.830.36
other_obj.0.050.040.020.000.000.000.000.000.00
Geometric validation
IoU49.3241.0836.3949.5442.2737.6054.6445.5739.35
Precision54.0446.2341.9153.2746.1041.5959.7550.3443.59
Recall84.9578.6973.4387.6183.5979.6786.4782.7980.16
", + "bbox": [ + 514, + 164, + 903, + 415 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "B/14) features. FiT3D offers multi-view refined DINOv2 features [116]. Note that FiT3D reports results, concatenating the refined features with DINOv2 features. We report results using both plain features only and the concatenation. We also use rendered 2D segmentations of our S4C + STEGO baseline for 2D validation. For multi-view feature consistency, we utilize DINO [11], DINOv2, and FiT3D [116] features as baselines.", + "bbox": [ + 511, + 436, + 906, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. Our encoder-decoder uses a DINO-B/8 [11] backbone and a dense prediction decoder [84]. The MLP decoder $\\phi$ entails two layers with 128 hidden features. As rendering features is expensive, $\\phi$ predicts 64 features. We employ another MLP to upproject again to the full dimensionality $D = 768$ . If not stated differently, our target features are obtained from DINO-B/8 [11]. We train using a batch size of 4 and extract 32 patches of size $8 \\times 8$ from each image to compute $\\mathcal{L}_{\\mathrm{SceneDINO}}$ . Volume rendering samples each ray at $L = 32$ uniformly spaced points in inverse depth within $[3\\mathrm{m}, 80\\mathrm{m}]$ . We train for $100\\mathrm{k}$ steps using Adam [52] with a base learning rate of $10^{-4}$ . Training takes ca. 2 days on a single V100 GPU. We distill using a batch size of 4, 5 center points, a feature batch of size 576, and cluster with $K = 19$ . For kNN sampling, we use $k = 4$ . The feature buffer holds 256 feature batches. Refer to the supplement for more details.", + "bbox": [ + 511, + 563, + 908, + 820 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.3D semantic scene completion", + "text_level": 1, + "bbox": [ + 511, + 832, + 779, + 849 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We assess the unsupervised SSC and geometric accuracy of SceneDINO with our 3D feature distillation approach on SSCBench-KITTI-360. In particular, Tab. 1 com", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6789", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/810d3f72e739835c1aaaf5f4c812a7b34c5c4dec26fa2b38671b11d7b415a1a0.jpg", + "image_caption": [ + "Figure 5. Qualitative SSC comparison on KITTI-360. We show the input image, SceneDINO's feature field using the first three principal components and SSC prediction, the prediction of our baseline S4C + STEGO, and the ground truth. We only visualize surface voxels. Qualitative results show the expressiveness of our feature field and SceneDINO's capabilities to accurately reconstruct and label a scene." + ], + "image_footnote": [], + "bbox": [ + 93, + 85, + 908, + 224 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d57ed7dc7c538e1040e8a8fbe59df8d47fa6163efd1bda8c928cae10ed0a2317.jpg", + "table_caption": [ + "Table 2. 2D unsupervised semantic segmentation results on KITTI-360. Comparing SceneDINO to existing 2D methods and our S4C + STEGO 3D baseline, using Accuracy and mean IoU (in $\\%$ , $\\uparrow$ ) on the SSCBench-KITTI-360 test split. $\\dagger$ denotes the use of plain FiT3D features. $\\ddagger$ denotes training on ImageNet and COCO." + ], + "table_footnote": [], + "table_body": "
MethodFeaturesAccmIoU
U2Seg‡ [78]-72.8923.43
STEGO [33]DINO [11]73.3223.57
STEGO [33]DINOv2 [80]64.5424.82
STEGO [33]FiT3D [116]54.1922.29
STEGO [33]FiT3D† [116]57.2518.95
S4C [38] + STEGO [33]DINO [11]65.1621.67
SceneDINO (Ours)DINO [11]77.7425.81
", + "bbox": [ + 91, + 369, + 483, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "pares SceneDINO against our unsupervised SSC baseline S4C [38] + STEGO [33]. SceneDINO achieves a (semantic) mIoU of $8.0\\%$ for the range of $51.2\\mathrm{m}$ , significantly improving over our unsupervised baseline $(6.6\\%)$ . This demonstrates that SceneDINO effectively lifts DINO features into 3D. In terms of geometric accuracy, SceneDINO moderately improves over S4C + STEGO. Despite being fully unsupervised, SceneDINO comes within $2.2\\%$ points mIoU of the 2D-supervised S4C.", + "bbox": [ + 86, + 498, + 482, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 5 provides qualitative samples on SSCBenchKITTI-360. SceneDINO's unsupervised SSC predictions are less noisy and capture finely resolved semantics compared to S4C + STEGO. Compared to the ground truth, we observe that SceneDINO captures both the geometry and general semantics of the scene. We visualize SceneDINO's feature field (before distillation) using the first three principal components. In PCA space, we observe that our feature field captures semantically meaningful regions.", + "bbox": [ + 89, + 636, + 483, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. 2D semantic segmentation", + "text_level": 1, + "bbox": [ + 89, + 786, + 328, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2 compares the semantic predictions of SceneDINO to recent 2D approaches and our 3D baseline. We obtain 2D semantic segmentations from SceneDINO and our S4C + STEGO baseline using semantic rendering [38]. SceneDINO with our 3D distillation approach outperforms STEGO with DINO features, an established 2D unsuper", + "bbox": [ + 89, + 809, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/33bbb0735d5968570bcf451a0aaad96ecfbd3730f7b9b1a210ddda94039f514e.jpg", + "table_caption": [ + "Table 3. 2D unsupervised semantic segmentation domain generalization results. Comparing SceneDINO to existing 2D unsupervised semantic segmentation methods and S4C + STEGO 3D baseline, using Accuracy and mean IoU (in %, $\\dagger$ ). We train on KITTI-360 images and report domain generalization results on Cityscapes and BDD-100K val. $\\dagger$ denotes plain FiT3D features." + ], + "table_footnote": [], + "table_body": "
MethodFeaturesCityscapesBDD-100K
AccmIoUAccmIoU
U2Seg [78]-75.5718.6269.0017.99
STEGO [33]DINO [11]71.2119.4275.0221.41
STEGO [33]DINOv2 [80]68.4119.7365.7221.77
STEGO [33]FiT3D [116]66.9421.0165.9620.99
STEGO [33]FiT3D† [116]64.7617.1760.8319.09
S4C [38] + STEGO [33]DINO [11]54.8014.0444.9811.62
SceneDINO (Ours)DINO [11]73.1722.8172.2822.09
", + "bbox": [ + 513, + 383, + 906, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "vised semantic segmentation approach. In particular, the mIoU of SceneDINO is $2.24\\%$ points higher than for STEGO (w/ DINO). Utilizing 3D refined features from FiT3D deteriorates the baseline relative to DINO, indicating that the FiT3D refinement reduces feature expressiveness. Notably, our unsupervised 3D baseline S4C + STEGO transfers significantly worse to 2D than SceneDINO.", + "bbox": [ + 509, + 522, + 906, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also validate SceneDINO, trained on KITTI-360, on Cityscapes and BDD10K, demonstrating domain generalization. The results are reported in Tab. 3. SceneDINO outperforms all baselines in mIoU on both datasets while only falling short in Acc. S4C + STEGO falls short in generalization. We suspect this poor generalization is caused by the fact that S4C does not rely on general SSL features in the final model, while our feature field generalizes.", + "bbox": [ + 511, + 628, + 908, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Multi-view feature consistency", + "text_level": 1, + "bbox": [ + 511, + 757, + 784, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We analyze the multi-view consistency of our feature field against existing 2D SSL features in Tab. 4. We report the results of SceneDINO trained on KITTI-360 and RealEstate10K. SceneDINO trained using DINO features exhibits significant improvements in multi-view feature consistency over standard DINO features. We also train SceneDINO using target features from DINOv2 [80]. Compared to standard DINOv2 and FiT3D fea", + "bbox": [ + 511, + 780, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6790", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/52df232ea434c6fe491ce07cffa0c3df5c1f26eb54a4e09fdd31719287a4d2af.jpg", + "table_caption": [ + "Table 4. Multi-view consistency results. Comparing multi-view consistency of SceneDINO to existing 2D SSL features, using $\\mathrm{L}_1$ distance $(\\downarrow)$ , $\\mathrm{L}_2$ distance $(\\downarrow)$ , and cosine similarity $(\\uparrow)$ on KITTI-360 and RealEstate10K. We compare DINO (top) and DINOv2-based (bottom) features. $\\dagger$ denotes plain FiT3D features." + ], + "table_footnote": [], + "table_body": "
MethodKITTI-360RealEstate10K
L1L2cos-simL1L2cos-sim
DINO [11]16.060.740.7014.410.660.75
SceneDINO (w/ DINO)6.450.330.935.870.280.95
DINOv2 [80]15.830.730.7014.200.660.75
FiT3D [116]22.860.810.8219.880.720.85
FiT3D† [116]7.020.330.935.670.270.95
SceneDINO (w/ DINOv2)5.240.240.964.870.220.97
", + "bbox": [ + 91, + 164, + 483, + 277 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5dbc1df1ba260bf86ff28cc7e40fae07fcb325d4ec28a632270ddb706eff2d7e.jpg", + "table_caption": [ + "Table 5. SceneDINO analysis. We analyze the role of decomposing positional encodings, the choice of downsampling features during training, the effectiveness of the feature smoothness loss, the effect of estimated camera poses, and the choice of target features. We report the mean IoU (in %, ↑) using a range of $51.2\\mathrm{m}$ on SSCBench-KITTI-360 test. $\\Delta$ mIoU reports the absolute difference in % points to our standard model with DINO target features." + ], + "table_footnote": [], + "table_body": "
Δ mIoUmIoUConfiguration
-1.186.82No downsampler (bilinear up. + aug.)
-1.176.83No feature smoothness loss (λfs = 0)
-0.747.26No pos. enc. decomposition
-0.127.88w/ estimated ORB-SLAM3 poses
—8.00Full framework (SceneDINO)
+1.089.08DINOv2 target features (vs. DINO)
", + "bbox": [ + 91, + 390, + 483, + 481 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tures, SceneDINO's feature field yields significantly better multi-view consistency. Notably, compared against plain 3D refined features of FiT3D, SceneDINO shows a better multi-view consistency on both datasets and all metrics while also offering more expressiveness (cf. Tab. 2).", + "bbox": [ + 89, + 501, + 483, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Analyzing SceneDINO", + "text_level": 1, + "bbox": [ + 89, + 589, + 302, + 606 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To understand what core components contribute to obtaining an expressive feature field of SceneDINO, we omit or replace individual components and report the results in Tab. 5. Replacing the downsampling approach with bilinear upsampling and multi-crop augmentations, similar to [1], to obtain high-resolution target features decreases the SSC mIoU by $1.18\\%$ . Omitting the feature smoothness loss leads to a similar mIoU drop. Abolishing the constant decomposition of positional encodings leads to an mIoU drop of $0.74\\%$ . Training using unsupervised camera poses estimated by ORB-SLAM3 [7] results in an insignificant mIoU drop of only $0.12\\%$ over using KITTI-360 poses. Going from DINO to DINOv2 target features leads to an increased mIoU of $1.08\\%$ , demonstrating that SceneDINO can benefit from more expressive 2D target features.", + "bbox": [ + 88, + 612, + 482, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Tab. 6, we analyze our 3D distillation. Performing no distillation at all, just feature clustering, decreases mIoU by $1.61\\%$ . Omitting the kNN-correlation loss leads to an mIoU drop of $1.35\\%$ . Distilling only with center points,", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8c37d089a1e6c2802d19da28710db6f771709af7da02bb0e65c1deb53bc779fe.jpg", + "table_caption": [ + "Table 6. Feature distillation analysis. We analyze the effectiveness of distilling SceneDINO's features, the $k$ -NN-correlation loss, our neighborhood sampling, and our 3D sampling approach over standard 5-crop sampling. We report the mean IoU (in %, ↑) using a range of $51.2 \\mathrm{~m}$ on SSCBench-KITTI-360 test." + ], + "table_footnote": [], + "table_body": "
Δ mIoUmIoUConfiguration
-1.616.39No distillation
-1.356.65No kNN-correlation loss (λkNN = 0)
-0.977.03No neighborhood sampling (cf. Fig. 4)
-0.477.535-crop sampling [33] (instead 3D sampling)
—8.00Full framework (SceneDINO)
", + "bbox": [ + 514, + 164, + 906, + 239 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/801bc9b9593448c715c95e2c6b9091501fd7937c510b8dd4f6e6aacbe8741bdb.jpg", + "table_caption": [ + "Table 7. Probing analysis. We analyze linear and unsupervised probing of our distilled SceneDINO features on SSCBenchKITTI-360 test using mean IoU (in %, ↑). For reference, we also report S4C (2D supervised). Linear probing uses 2D annotations." + ], + "table_footnote": [], + "table_body": "
Probing approachTarget featuresmIoU
12.8 m25.6 m51.2 m
UnsupervisedDINO [11]10.7610.018.00
DINOv2 [80]13.7611.789.08
LinearDINO [11]13.6312.079.34
DINOv2 [80]15.8513.7010.57
S4C (full training)n/a16.9413.9410.19
", + "bbox": [ + 514, + 310, + 905, + 414 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "i.e., not performing neighborhood sampling (cf. Fig. 4), reduces mIoU by $0.97\\%$ . Using 5-crop feature sampling [33], instead of our proposed 3D sampling, leads to a reduced mIoU of $0.47\\%$ . This demonstrates the effectiveness of performing distillation in 3D using our novel approach.", + "bbox": [ + 511, + 435, + 905, + 511 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While focusing on unsupervised SSC, we can also linearly probe our distilled feature field (cf. Tab. 7). In particular, we train SceneDINO using different target features (DINO [11] and DINOv2 [11]), perform distillation, and probe the resulting distilled features. Using linear probing, i.e., training a single linear layer using 2D semantic labels, leads to a consistent mIoU increase over unsupervised probing. SceneDINO trained using DINOv2 target features even closes the gap to S4C, trained using 2D ground-truth semantic labels. We even surpass 2D supervised S4C slightly on the full range $(51.2\\mathrm{m})$ , suggesting the effectiveness of SceneDINO also for weakly-supervised tasks.", + "bbox": [ + 511, + 511, + 908, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 708, + 633, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented SceneDINO, to our knowledge, the first approach for unsupervised semantic scene completion. Trained using multi-view images and 2D DINO features without human supervision, SceneDINO is able to predict an expressive 3D feature field using a single input image during inference. Our novel 3D distillation approach yields state-of-the-art results in unsupervised SSC. While we focus on unsupervised SSC, our multi-view feature consistency, linear probing, and domain generalization results highlight the potential of SceneDINO as a strong foundation for various 3D scene-understanding tasks.", + "bbox": [ + 511, + 734, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6791", + "bbox": [ + 482, + 944, + 513, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This project was partially supported by the European Research Council (ERC) Advanced Grant SIMULACRON, DFG project CR 250/26-1 \"4D-YouTube\", and GNI Project \"AICC\". This project was also partially supported by the ERC under the European Union's Horizon 2020 research and innovation programme (grant agreement No. 866008). Additionally, this work has been co-funded by the LOEWE initiative (Hesse, Germany) within the emergenCITY center [LOEWE/1/12/519/03/05.001(0016)/72] and by the Deutsche Forschungsgemeinschaft (German Research Foundation, DFG) under Germany's Excellence Strategy (EXC 3066/1 \"The Adaptive Mind\", Project No. 533717223). Christoph Reich is supported by the Konrad Zuse School of Excellence in Learning and Intelligent Systems (ELIZA) through the DAAD programme Konrad Zuse Schools of Excellence in Artificial Intelligence, sponsored by the Federal Ministry of Education and Research. Christian Rupprecht is supported by an Amazon Research Award. Finally, we acknowledge the support of the European Laboratory for Learning and Intelligent Systems (ELLIS) and thank Mateo de Mayo as well as Igor Cvišić for help with estimating camera poses.", + "bbox": [ + 89, + 90, + 485, + 395 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 410, + 187, + 426 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Nikita Araslanov and Stefan Roth. Self-supervised augmentation consistency for adapting semantic segmentation. In CVPR, pages 15384-15394, 2021. 8", + "[2] Yuki Markus Asano, Christian Rupprecht, and Andrea Vedaldi. Self-labelling via simultaneous clustering and representation learning. In ICLR, 2020. 2", + "[3] Philip Bachman, R. Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. In NeurIPS*2019, pages 15509-15519. 2", + "[4] Adrien Bardes, Jean Ponce, and Yann LeCun. VI-CRegL: Self-supervised learning of local visual features. In NeurIPS*2022, pages 8799-8810. 2", + "[5] Adrien Bardes, Jean Ponce, and Yann LeCun. VI-CReg: Variance-invariance-covariance regularization for self-supervised learning. In *ICLR*, 2022. 2", + "[6] Yingjie Cai, Xuesong Chen, Chao Zhang, Kwan-Yee Lin, Xiaogang Wang, and Hongsheng Li. Semantic scene completion via integrating instances and scene in-the-loop. In CVPR, pages 324-333, 2021. 2", + "[7] Carlos Campos, Richard Elvira, Juan J. Gómez Rodríguez, José M. M. Montiel, and Juan D Tardós. ORB-SLAM3: An accurate open-source library for visual, visual-inertial and multi-map SLAM. IEEE Trans. Robot., 37(6):1874-1890, 2021. 4, 8", + "[8] Anh-Quan Cao and Raoul de Charette. Monoscene: Monocular 3D semantic scene completion. In CVPR, pages 3981-3991, 2022. 2", + "[9] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS*2020, pages 9912-9924. 2" + ], + "bbox": [ + 107, + 436, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In ECCV, pages 132-149, 2018. 2", + "[11] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. 1, 2, 3, 4, 6, 7, 8", + "[12] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv:2003.04297 [cs.CV], 2020. 2", + "[13] Xiaokang Chen, Kwan-Yee Lin, Chen Qian, Gang Zeng, and Hongsheng Li. 3d sketch-aware semantic scene completion via semi-supervised structure prior. In CVPR, pages 4192-4201, 2020. 2", + "[14] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In CVPR, pages 9640-9649, 2021. 2", + "[15] Yuedong Chen, Chuanxia Zheng, Haofei Xu, Bohan Zhuang, Andrea Vedaldi, Tat-Jen Cham, and Jianfei Cai. MVSplat360: Feed-forward 360 scene synthesis from sparse views. In NeurIPS*2024, pages 107064-107086. 2", + "[16] Ran Cheng, Christopher Agia, Yuan Ren, Xinhai Li, and Bingbing Liu. S3CNet: A sparse semantic scene completion network for LiDAR point clouds. In CoRL, pages 2148-2161, 2020. 2", + "[17] Jang Hyun Cho, Utkarsh Mall, Kavita Bala, and Bharath Hariharan. PiCIE: Unsupervised semantic segmentation using invariance and equivariance in clustering. In CVPR, pages 16794-16804, 2021. 2, 6", + "[18] Özgün Çiçek, Ahmed Abdulkadir, Soeren S Lienkamp, Thomas Brox, and Olaf Ronneberger. 3D U-Net: Learning dense volumetric segmentation from sparse annotation. In MICCAI, pages 424–432, 2016. 1", + "[19] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The Cityscapes dataset for semantic urban scene understanding. In CVPR, pages 3213-3223, 2016. 6", + "[20] Timothee Darcet, Federico Baldassarre, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Cluster and predict latent patches for improved masked image modeling. arXiv:2502.08769 [cs.CV], 2025. 2", + "[21] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 6", + "[22] Carl Doersch, Abhinav Gupta, and Alexei A. Efros. Unsupervised visual representation learning by context prediction. In ICCV, pages 1422-1430, 2015. 2", + "[23] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth $16 \\times 16$ words: Transformers for image recognition at scale. In ICLR, 2021. 2", + "[24] Linus Ericsson, Henry Gouk, Chen Change Loy, and Timothy M. Hospedales. Self-supervised representation learning: Introduction, advances, and challenges. IEEE Trans. Signal Process., 39(3):42-62, 2022. 2" + ], + "bbox": [ + 522, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6792", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Stephanie Fu, Mark Hamilton, Laura E. Brandt, Axel Feldmann, Zhoutong Zhang, and William T. Freeman. FeatUp: A model-agnostic framework for features at any resolution. In ICLR, 2024. 4", + "[26] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The KITTI dataset. Int. J. Robot. Res., 32(11):1231-1237, 2013. 1", + "[27] Clément Godard, Oisin Mac Aodha, and Gabriel J. Brostow. Unsupervised monocular depth estimation with left-right consistency. In CVPR, pages 270-279, 2017. 4", + "[28] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, et al. Bootstrap your own latent: A new approach to self-supervised learning. In NeurIPS*2020, pages 21271-21284. 2", + "[29] Agrim Gupta, Jiajun Wu, Jia Deng, and Li Fei-Fei. Siamese Masked Autoencoders. In NeurIPS*2023, pages 40676-40693. 2", + "[30] Huy Ha and Shuran Song. Semantic abstraction: Openworld 3D scene understanding from 2D vision-language models. In CoRL, pages 643-653, 2023. 2", + "[31] Oliver Hahn, Nikita Araslanov, Simone Schaub-Meyer, and Stefan Roth. Boosting unsupervised semantic segmentation with principal mask proposals. Trans. Mach. Learn. Res., 2024. 3, 6", + "[32] Oliver Hahn, Christoph Reich, Nikita Araslanov, Daniel Cremers, Christian Rupprecht, and Stefan Roth. Scene-centric unsupervised panoptic segmentation. In CVPR, pages 24485-24495, 2025. 1", + "[33] Mark Hamilton, Zhoutong Zhang, Bharath Hariharan, Noah Snavely, and William T. Freeman. Unsupervised semantic segmentation by distilling feature correspondences. In ICLR, 2022. 1, 2, 3, 4, 5, 6, 7, 8", + "[34] Keonhee Han, Dominik Muhle, Felix Wimbauer, and Daniel Cremers. Boosting self-supervision for single-view scene completion via knowledge distillation. In CVPR, pages 9837-9847, 2024. 1, 2", + "[35] Xian-Feng Han, Hamid Laga, and Mohammed Bennamoun. Image-based 3D object reconstruction: State-of-the-art and trends in the deep learning era. IEEE Trans. Pattern Anal. Mach. Intell., 43(5):1578-1604, 2019. 2", + "[36] Robert Harb and Patrick Knobelreiter. InfoSeg: Unsupervised semantic image segmentation with mutual information maximization. In $GCPR$ , pages 18-32, 2021. 2", + "[37] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge University Press, 2003. 2", + "[38] Adrian Hayler, Felix Wimbauer, Dominik Muhle, Christian Rupprecht, and Daniel Cremers. S4C: Self-supervised semantic scene completion with neural fields. In 3DV, pages 409-420, 2024. 1, 2, 6, 7", + "[39] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9729-9738, 2020. 2", + "[40] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 2" + ], + "bbox": [ + 99, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Olivier Henaff. Data-efficient image pecognition with contrastive predictive coding. In ICML, pages 4182-4192, 2020. 2", + "[42] R. Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. In ICLR, 2019. 2", + "[43] Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3D v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. IEEE Trans. Pattern Anal. Mach. Intell., 46(12):10579-10596, 2024. 2", + "[44] Rui Huang, Songyou Peng, Ayca Takmaz, Federico Tombari, Marc Pollefeys, Shiji Song, Gao Huang, and Francis Engelmann. Segment3D: Learning fine-grained class-agnostic 3D segmentation without manual labels. In ECCV, pages 278–295, 2024. 2", + "[45] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3D semantic occupancy prediction. In CVPR, pages 9223-9232, 2023. 2", + "[46] Yuanhui Huang, Wenzhao Zheng, Borui Zhang, Jie Zhou, and Jiwen Lu. SelfOcc: Self-supervised vision-based 3D occupancy prediction. In CVPR, pages 19946-19956, 2024. 1, 2", + "[47] Joel Janai, Fatma Güney, Aseem Behl, and Andreas Geiger. Computer vision for autonomous vehicles: Problems, datasets and state of the art. Found. Trends Comput. Graph. Vis., 12(1-3):1-308, 2020. 1", + "[48] Xu Ji, Joao F. Henriques, and Andrea Vedaldi. Invariant information clustering for unsupervised image classification and segmentation. In ICCV, pages 9865-9874, 2019. 2", + "[49] Haoyi Jiang, Liu Liu, Tianheng Cheng, Xinjie Wang, Tianwei Lin, Zhizhong Su, Wenyu Liu, and Xinggang Wang. GaussTR: Foundation model-aligned Gaussian transformer for self-supervised 3D spatial understanding. arXiv:2412.13193 [cs.CV], 2024. 2", + "[50] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. LERF: Language embedded radiance fields. In ICCV, pages 19729-19739, 2023. 2", + "[51] Chanyoung Kim, Woojung Han, Dayun Ju, and Seong Jae Hwang. EAGLE: Eigen aggregation learning for object-centric unsupervised semantic segmentation. In CVPR, pages 3523-3533, 2024. 3, 6", + "[52] Diederik P. Kingma and Jimmy Lei Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 6", + "[53] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment Anything. In ICCV, pages 4015-4026, 2023. 1, 2", + "[54] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing NeRF for editing via feature field distillation. In NeurIPS*2022, pages 23311-23330. 2" + ], + "bbox": [ + 522, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6793", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[55] Alexander Koenig, Maximilian Schambach, and Johannes Otterbach. Uncovering the inner workings of STEGO for safe unsupervised semantic segmentation. In CVPRW, pages 3789-3798, 2023. 4", + "[56] Philipp Krahenbuhl and Vladlen Koltun. Efficient inference in fully connected CRFs with Gaussian edge potentials. In NIPS*2011, pages 109-117. 6", + "[57] Harold W. Kuhn. The Hungarian method for the assignment problem. Nav. Res. Logist. Q., 2:83-97, 1955. 6", + "[58] Jie Li, Yu Liu, Dong Gong, Qinfeng Shi, Xia Yuan, Chunxia Zhao, and Ian D. Reid. RGBD based dimensional decomposition residual network for 3D semantic scene completion. In CVPR, pages 7693-7702, 2019. 2", + "[59] Jie Li, Kai Han, Peng Wang, Yu Liu, and Xia Yuan. Anisotropic convolutional networks for 3D semantic scene completion. In CVPR, pages 3348-3356, 2020.", + "[60] Jie Li, Yu Liu, Xia Yuan, Chunxia Zhao, Roland Siegwart, Ian Reid, and Cesar Cadena. Depth based semantic scene completion with position importance aware loss. IEEE Robotics Autom. Lett., 5(1):219-226, 2020. 2", + "[61] Junnan Li, Pan Zhou, Caiming Xiong, and Steven Hoi. Prototypical contrastive learning of unsupervised representations. In ICLR, 2021. 2", + "[62] Pengfei Li, Yongliang Shi, Tianyu Liu, Hao Zhao, Guyue Zhou, and Ya-Qin Zhang. Semi-supervised implicit scene completion from sparse LiDAR. arXiv:2111.14798 [cs.CV], 2021. 2", + "[63] Yiming Li, Zhiding Yu, Christopher B. Choy, Chaowei Xiao, José M. Álvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. VoxFormer: Sparse voxel transformer for camera-based 3D semantic scene completion. In CVPR, pages 9087-9098, 2023. 2", + "[64] Yiming Li, Sihang Li, Xinhao Liu, Moonjun Gong, Kenan Li, Nuo Chen, Zijun Wang, Zhiheng Li, Tao Jiang, Fisher Yu, Yue Wang, Hang Zhao, Zhiding Yu, and Chen Feng. SSCBench: A large-scale 3D semantic scene completion benchmark for autonomous driving. In IROS, pages 13333-13340, 2024. 1, 2, 6", + "[65] Zhiqi Li, Zhiding Yu, David Austin, Mingsheng Fang, Shiyi Lan, Jan Kautz, and José M. Álvarez. FB-OCC: 3D occupancy prediction based on forward-backward view transformation. arXiv:2307.01492 [cs.CV], 2023. 2", + "[66] Yiyi Liao, Jun Xie, and Andreas Geiger. KITTI-360: A novel dataset and benchmarks for urban scene understanding in 2D and 3D. IEEE Trans. Pattern Anal. Mach. Intell., 45(3):3292-3310, 2023. 1, 5", + "[67] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, pages 740-755, 2024. 6", + "[68] Shice Liu, Yu Hu, Yiming Zeng, Qiankun Tang, Beibei Jin, Yinhe Han, and Xiaowei Li. See and think: Disentangling semantic scene completion. In NeurIPS*2018, pages 261-272. 2", + "[69] Stuart Lloyd. Least squares quantization in PCM. IEEE Trans. Inf. Theory, 28(2):129-137, 1982. 4, 5" + ], + "bbox": [ + 99, + 92, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[70] Zhiliang Ma and Shilong Liu. A review of 3D reconstruction techniques in civil engineering and their applications. Adv. Eng. Inform., 37:163-174, 2018. 1", + "[71] James MacQueen. Some methods for classification and analysis of multivariate observations. In Berkeley Symp. on Math. Statist. and Prob., pages 281-298, 1967. 4, 5", + "[72] Nelson Max. Optical models for direct volume rendering. IEEE Trans. Vis. Comput. Graph., 1(2):99-108, 1995. 3", + "[73] Kirill Mazur, Edgar Sucar, and Andrew J Davison. Feature-realistic neural fusion for real-time, open set scene understanding. In ICRA, pages 8201-8207, 2023. 2", + "[74] Ruihang Miao, Weizhou Liu, Mingrui Chen, Zheng Gong, Weixin Xu, Chen Hu, and Shuchang Zhou. Occdepth: A depth-aware method for 3D semantic scene completion. arXiv.2302.13540 [cs.CV], 2023. 1, 2", + "[75] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. Commun. ACM, 65(1):99-106, 2021. 2, 3", + "[76] Yue Ming, Xuyang Meng, Chunxiao Fan, and Hui Yu. Deep learning for monocular depth estimation: A review. Neurocomputing, 438:14-33, 2021. 2", + "[77] Duy Kien Nguyen, Yanghao Li, Vaibhav Aggarwal, Martin R. Oswald, Alexander Kirillov, Cees G. M. Snoek, and Xinlei Chen. R-MAE: Regions meet masked autoencoders. In ICLR, 2024. 2", + "[78] Dantong Niu, Xudong Wang, Xinyang Han, Long Lian, Roei Herzig, and Trevor Darrell. Unsupervised universal image segmentation. In CVPR, pages 22744-22754, 2024. 6, 7", + "[79] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In ECCV, pages 69-84, 2016. 2", + "[80] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. DINOv2: Learning robust visual features without supervision. Trans. Mach. Learn. Res., 2024. 2, 3, 6, 7, 8", + "[81] Martin R Oswald, Eno Töppe, Claudia Nieuwenhuis, and Daniel Cremers. A review of geometry recovery from a single image focusing on curved object reconstruction. Innovations for Shape Analysis: Models and Algorithms, pages 343-378, 2013. 2", + "[82] Mingjie Pan, Jiaming Liu, Renrui Zhang, Peixiang Huang, Xiaqi Li, Hongwei Xie, Bing Wang, Li Liu, and Shang-hang Zhang. RenderOcc: Vision-centric 3D occupancy prediction with 2D rendering supervision. In ICRA, pages 12404-12411, 2024. 2", + "[83] Songyou Peng, Kyle Genova, Chiyu “Max” Jiang, Andrea Tagliasacchi, Marc Pollefeys, and Thomas Funkhouser. OpenScene: 3D scene understanding with open vocabularies. In CVPR, pages 815–824, 2023. 2", + "[84] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In ICCV, pages 12179-12188, 2021. 6" + ], + "bbox": [ + 524, + 92, + 903, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "6794", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[85] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, et al. SAM 2: Segment anything in images and videos. arXiv:2408.00714 [cs.CV], 2024. 1", + "[86] Stephan R. Richter and Stefan Roth. Matryoshka networks: Predicting 3D geometry via nested shape layers. In CVPR, pages 1936-1944, 2018. 2", + "[87] Christoph B. Rist, David Emmerichs, Markus Enzweiler, and Dariu M. Gavrila. Semantic scene completion using local deep implicit functions on LiDAR data. IEEE Trans. Pattern Anal. Mach. Intell., 44(10):7205-7218, 2022. 2", + "[88] Luis Roldão, Raoul de Charette, and Anne Verroust-Blondet. LMSCNet: Lightweight multiscale 3D semantic completion. In 3DV, pages 111–119, 2020. 1, 2", + "[89] Luis Roldao, Raoul De Charette, and Anne Verroust-Blondet. 3D semantic scene completion: A survey. Int. J. Comput. Vis., 130(8):1978-2005, 2022. 1", + "[90] Johannes L. Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, pages 4104-4113, 2016. 2", + "[91] David Sculley. Web-scale $k$ -means clustering. In WWW, page 1177-1178, 2010. 5", + "[92] Hyun Seok Seong, WonJun Moon, SuBeen Lee, and Jae-Pil Heo. Leveraging hidden positives for unsupervised semantic segmentation. In CVPR, pages 19540-19549, 2023. 3, 6", + "[93] Nur Muhammad Mahi Shafiullah, Chris Paxton, Lerrel Pinto, Soumith Chintala, and Arthur Szlam. CLIP-Fields: Weakly supervised semantic fields for robotic memory. In ICRA Workshop on Pretraining for Robotics, 2023. 2", + "[94] William Shen, Ge Yang, Alan Yu, Jansen Wong, Leslie Pack Kaelbling, and Phillip Isola. Distilled feature fields enable few-shot language-guided manipulation. In CoRL, pages 405-424, 2023. 2", + "[95] Leon Sick, Dominik Engel, Pedro Hermosilla, and Timo Ropinski. Unsupervised semantic segmentation through depth-guided feature correlation and sampling. In CVPR, pages 3637-3646, 2024. 3, 6", + "[96] Shuran Song, Fisher Yu, Andy Zeng, Angel X. Chang, Manolis Savva, and Thomas A. Funkhouser. Semantic scene completion from a single depth image. In CVPR, pages 190-198, 2017. 1, 2", + "[97] Stanislaw Szymanowicz, Eldar Insafutdinov, Chuanxia Zheng, Dylan Campbell, Joao F. Henriques, Christian Rupprecht, and Andrea Vedaldi. Flash3D: Feed-forward generalisable 3D scene reconstruction from a single image. arXiv:2406.04343 [cs.CV], 2024. 2", + "[98] Ayca Takmaz, Elisabetta Fedele, Robert W. Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. OpenMask3D: Open-vocabulary 3D instance segmentation. In NeurIPS*2023, pages 68367-68390. 2", + "[99] Zachary Teed and Jia Deng. RAFT: Recurrent all-pairs field transforms for optical flow. In ECCV, pages 402-419, 2020. 6", + "[100] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua" + ], + "bbox": [ + 99, + 92, + 480, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Lin, and Hongyang Li. Scene as occupancy. In ICCV, pages 8372-8381, 2023. 2", + "[101] Nikolaos Tsagkas, Oisin Mac Aodha, and Chris Xiaoxuan Lu. VL-Fields: Towards language-grounded neural implicit spatial representations. In ICRA Workshop on Representations, Abstractions, and Priors for Robot Learning, 2023. 2", + "[102] Vadim Tschernezki, Iro Laina, Diane Larlus, and Andrea Vedaldi. Neural feature fusion fields: 3D distillation of self-supervised 2D image representations. In 3DV, pages 443-453, 2022. 2", + "[103] Shubham Tulsiani, Tinghui Zhou, Alexei A. Efros, and Jitendra Malik. Multi-view supervision for single-view reconstruction via differentiable ray consistency. In CVPR, 2017. 2", + "[104] Xudong Wang, Rohit Girdhar, Stella X. Yu, and Ishan Misra. Cut and learn for unsupervised object detection and instance segmentation. In CVPR, pages 3124-3134, 2023. 1, 2", + "[105] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: From error visibility to structural similarity. IEEE Trans. Image Process., 13(4): 600-612, 2004. 4", + "[106] Silvan Weder, Hermann Blum, Francis Engelmann, and Marc Pollefeys. LabelMaker: Automatic semantic label generation from RGB-D trajectories. In 3DV, pages 334-343, 2024. 2", + "[107] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In CVPR, pages 14668-14678, 2022. 2", + "[108] Felix Wimbauer, Nan Yang, Christian Rupprecht, and Daniel Cremers. Behind the scenes: Density fields for single view reconstruction. In CVPR, pages 9076-9086, 2023. 1, 2, 3, 4, 6", + "[109] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Comput. Graph. Forum, pages 641-676, 2022. 2", + "[110] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep LiDAR point cloud segmentation via learning contextual shape priors from scene completion. In AAAI, pages 3101-3109, 2021. 2", + "[111] Jiawei Yang, Boris Ivanovic, Or Litany, Xinshuo Weng, Seung Wook Kim, Boyi Li, Tong Che, Danfei Xu, Sanja Fidler, Marco Pavone, and Yue Wang. EmerneRF: Emergent spatial-temporal scene decomposition via self-supervision. In ICLR, 2024. 2", + "[112] Jiawei Yang, Katie Z Luo, Jiefeng Li, Congyue Deng, Leonidas Guibas, Dilip Krishnan, Kilian Q Weinberger, Yonglong Tian, and Yue Wang. Denoising vision transformers. In ECCV, pages 453-469, 2024. 4", + "[113] Zhuoyue Yang, Ju Dai, and Junjun Pan. 3D reconstruction from endoscopy images: A survey. Comput. Biol. Med., 175:108546, 2024. 1" + ], + "bbox": [ + 517, + 92, + 903, + 900 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "6795", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[114] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelNeRF: Neural radiance fields from one or few images. In CVPR, pages 4578-4587, 2021. 2", + "[115] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. BDD100K: A diverse driving dataset for heterogeneous multitask learning. In CVPR, pages 2633-2642, 2020. 6", + "[116] Yuanwen Yue, Anurag Das, Francis Engelmann, Siyu Tang, and Jan Eric Lenssen. Improving 2D feature representations by 3D-aware fine-tuning. In ECCV, pages 57-74, 2024. 2, 6, 7, 8", + "[117] Pingping Zhang, Wei Liu, Yinjie Lei, Huchuan Lu, and Xiaoyun Yang. Cascaded context pyramid for full-resolution 3D semantic scene completion. In ICCV, pages 7800-7809, 2019. 2", + "[118] Yunpeng Zhang, Zheng Zhu, and Dalong Du. OccFormer: Dual-path transformer for vision-based 3D semantic occupancy prediction. In ICCV, pages 9433-9443, 2023. 2", + "[119] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snively. Stereo magnification: Learning view synthesis using multiplane images. ACM Trans. Graph., 37 (4):65, 2018. 6", + "[120] Onur Özyeşil, Vladislav Voroninski, Ronen Basri, and Amit Singer. A survey of structure from motion. Acta Numer., 26:305-364, 2017. 2" + ], + "bbox": [ + 89, + 92, + 482, + 458 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "6796", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2025/Feed-Forward SceneDINO for Unsupervised Semantic Scene Completion/0cb8dcd6-d20d-4d18-a810-629c106f3cdd_model.json b/2025/Feed-Forward SceneDINO for Unsupervised Semantic Scene Completion/0cb8dcd6-d20d-4d18-a810-629c106f3cdd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b1da21a5892d574379a98f7c9c022f89d7720b37 --- /dev/null +++ b/2025/Feed-Forward SceneDINO for Unsupervised Semantic Scene Completion/0cb8dcd6-d20d-4d18-a810-629c106f3cdd_model.json @@ -0,0 +1,3174 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.132, + 0.131, + 0.867, + 0.154 + ], + "angle": 0, + "content": "Feed-Forward SceneDINO for Unsupervised Semantic Scene Completion" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.18, + 0.379, + 0.199 + ], + "angle": 0, + "content": "Aleksandar Jevtic*1" + }, + { + "type": "text", + "bbox": [ + 0.162, + 0.199, + 0.273, + 0.216 + ], + "angle": 0, + "content": "Oliver Hahn" + }, + { + "type": "text", + "bbox": [ + 0.308, + 0.199, + 0.48, + 0.218 + ], + "angle": 0, + "content": "Christian Rupprecht" + }, + { + "type": "text", + "bbox": [ + 0.136, + 0.219, + 0.216, + 0.234 + ], + "angle": 0, + "content": "\\(^{1}\\)TU Munich" + }, + { + "type": "text", + "bbox": [ + 0.237, + 0.22, + 0.332, + 0.234 + ], + "angle": 0, + "content": "\\(^{2}\\)TU Darmstadt" + }, + { + "type": "text", + "bbox": [ + 0.354, + 0.22, + 0.487, + 0.235 + ], + "angle": 0, + "content": "3University of Oxford" + }, + { + "type": "text", + "bbox": [ + 0.414, + 0.18, + 0.592, + 0.199 + ], + "angle": 0, + "content": "Christoph Reich* 1,2,4,5" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.18, + 0.643, + 0.216 + ], + "angle": 0, + "content": "Eich\\*1,2,4,5 F6 Stefan Roth 2,5,6" + }, + { + "type": "text", + "bbox": [ + 0.628, + 0.181, + 0.78, + 0.198 + ], + "angle": 0, + "content": "Felix Wimbauer1,4" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.219, + 0.635, + 0.235 + ], + "angle": 0, + "content": "4MCML 5ELIZA" + }, + { + "type": "text", + "bbox": [ + 0.679, + 0.199, + 0.836, + 0.216 + ], + "angle": 0, + "content": "Daniel Cremers \\(^{1,4,5}\\)" + }, + { + "type": "text", + "bbox": [ + 0.655, + 0.22, + 0.862, + 0.235 + ], + "angle": 0, + "content": "\\(^{6}\\)hessian.AI *equal contribution" + }, + { + "type": "text", + "bbox": [ + 0.365, + 0.237, + 0.632, + 0.251 + ], + "angle": 0, + "content": "https://visinf.github.io/scenedino" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.275, + 0.302, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.124, + 0.341, + 0.273, + 0.358 + ], + "angle": 0, + "content": "Single Input Image" + }, + { + "type": "text", + "bbox": [ + 0.306, + 0.296, + 0.388, + 0.308 + ], + "angle": 0, + "content": "SceneDINO" + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.265, + 0.6, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.44, + 0.342, + 0.571, + 0.356 + ], + "angle": 0, + "content": "3D Feature Field" + }, + { + "type": "text", + "bbox": [ + 0.602, + 0.296, + 0.701, + 0.308 + ], + "angle": 0, + "content": "Distill & Cluster" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.265, + 0.902, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.746, + 0.342, + 0.867, + 0.356 + ], + "angle": 0, + "content": "SSC Prediction" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.364, + 0.908, + 0.407 + ], + "angle": 0, + "content": "Figure 1. SceneDINO overview. Given a single input image (left), SceneDINO predicts both 3D scene geometry and 3D features in the form of a feature field (middle) in a feed-forward manner, capturing the structure and semantics of the scene. Unsupervised distillation and clustering of SceneDINO's feature space leads to unsupervised semantic scene completion predictions (right)." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.429, + 0.327, + 0.445 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.462, + 0.486, + 0.766 + ], + "angle": 0, + "content": "Semantic scene completion (SSC) aims to infer both the 3D geometry and semantics of a scene from single images. In contrast to prior work on SSC that heavily relies on expensive ground-truth annotations, we approach SSC in an unsupervised setting. Our novel method, SceneDINO, adapts techniques from self-supervised representation learning and 2D unsupervised scene understanding to SSC. Our training exclusively utilizes multi-view consistency self-supervision without any form of semantic or geometric ground truth. Given a single input image, SceneDINO infers the 3D geometry and expressive 3D DINO features in a feed-forward manner. Through a novel 3D feature distillation approach, we obtain unsupervised 3D semantics. In both 3D and 2D unsupervised scene understanding, SceneDINO reaches state-of-the-art segmentation accuracy. Linear probing our 3D features matches the segmentation accuracy of a current supervised SSC approach. Additionally, we showcase the domain generalization and multi-view consistency of SceneDINO, taking the first steps towards a strong foundation for single image 3D scene understanding." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.799, + 0.222, + 0.814 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Understanding the geometry and semantics of 3D scenes from image observations is a fundamental computer vision task with broad applications in robotics [26], autonomous driving [47, 66], medical image analysis [18, 113], and civil engineering [70]. The Semantic Scene Completion (SSC)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.43, + 0.907, + 0.566 + ], + "angle": 0, + "content": "task unifies 3D geometry and semantic prediction from limited image observations [64, 89, 96]. Recent progress in SSC has been primarily driven by utilizing supervised learning [38, 88, 96]. However, acquiring large-scale 3D annotations is highly labor-intensive [66]. While significant resources have been invested in collecting human annotations for 2D tasks [53, 85], annotating similar amounts of data in 3D remains unapproached. This motivates approaching SSC without the need for manually annotated data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.909, + 0.854 + ], + "angle": 0, + "content": "Existing SSC approaches rely on ground-truth semantic annotations and frequently utilize additional supervision from LiDAR scans [38, 46, 74, 96]. In contrast, we are the first to approach SSC in a fully unsupervised setting, i.e. without task supervision or other supervised components. In particular, we aim to approach SSC from a single image without relying on any human annotations, only learning from unlabeled multi-view images using self-supervision. This setting is extremely challenging for two reasons: first, the human-defined nature of semantic taxonomies is ambiguous, and second, a single image only provides a partial observation of the scene with many invisible areas. We take inspiration from recent advances in self-supervised learning (SSL) of 2D representations and 3D reconstruction. 2D SSL representations, such as from DINO [11], have been shown effective for 2D unsupervised scene understanding [32, 33, 104]. 3D reconstruction approaches successfully leveraged SSL from multi-view data to infer dense 3D geometry from a single image [34, 108]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.901 + ], + "angle": 0, + "content": "In this paper, we present SceneDINO, to the best of our knowledge, the first approach for unsupervised semantic scene completion. Trained using 2D SSL features" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6784" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.257 + ], + "angle": 0, + "content": "from DINO [11] and multi-view self-supervision [108], SceneDINO predicts both 3D geometry and 3D features from a single image during inference in a feed-forward manner. Our general 3D feature representations enable us to approach unsupervised 3D scene understanding. Harnessing our expressive 3D features, we propose a novel 3D feature distillation approach for obtaining unsupervised semantic predictions in 3D. While we focus on the task of unsupervised SSC, SceneDINO's features are general, offering a foundation for different 3D scene-understanding tasks by building on our 3D feature field." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.259, + 0.482, + 0.455 + ], + "angle": 0, + "content": "Specifically, we make the following contributions: (i) We introduce SceneDINO, the first approach predicting dense 3D geometry and expressive 3D features in a feedforward manner from a single image. (ii) We effectively distill SceneDINO's feature field representation in 3D, obtaining unsupervised semantic predictions. (iii) We demonstrate the first fully unsupervised SSC results. We build a simple, yet competitive unsupervised SSC baseline, lifting unsupervised 2D semantic predictions. Our SceneDINO approach outperforms this SSC baseline in unsupervised SSC as well as established 2D approaches in 2D semantic segmentation. (iv) Finally, we also showcase the domain generalization ability and multi-view consistency of SceneDINO." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.471, + 0.233, + 0.487 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.503, + 0.483, + 0.76 + ], + "angle": 0, + "content": "Single-image scene reconstruction. Estimating 3D geometry from image observations is a fundamental task in computer vision and has been studied for decades [37]. Traditional approaches, such as structure from motion [90], as well as recent neural radiance fields (NeRFs) [75], perform scene reconstruction using multiple images, as reviewed by multiple surveys [35, 109, 120]. Recently, estimating dense 3D geometry from a single image have been approached [8, 34, 81, 86, 97, 103, 108, 114]. Unlike monocular depth estimation [76], these approaches predict the depth for visible and occluded regions, reconstructing a complete scene. Behind the Scenes (BTS) [108] introduced an approach for unsupervised single-image scene reconstruction using multi-view self-supervision, which infers dense 3D geometry in a feed-forward manner. Our approach extends BTS by additionally lifting self-supervised features into 3D for unsupervised 3D scene understanding." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Semantic scene completion (SSC), also known as 3D semantic occupancy prediction, aims to jointly estimate the 3D geometry and semantics of a scene [63, 64, 96, 118]. Initial approaches used 3D semantic and geometric annotations and addressed indoor scenes [6, 13, 58-60, 68, 117], outdoor scenes with LiDAR [16, 62, 87, 88, 110], or both domains [8, 74]. Using birds-eye views has been proven effective for SSC [45, 65, 100]. To overcome the need for 3D annotations, approaches for using 2D annotations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.303 + ], + "angle": 0, + "content": "have been proposed [38, 46, 82]. While SelfOcc [46] and RenderOcc [82] use multiple inference views, S4C [38] performs single-image SSC. In particular, S4C [38] employs a supervised 2D model and lifts 2D multi-view semantic predictions into 3D. In contrast to using 2D annotations, GaussTR [49] uses 2D foundation models for SSC and multiple views during inference. However, GaussTR relies on heavily supervised foundation models, including SAM [53] and Metric3Dv2 [43], and uses weak supervision from image/text pairs. To the best of our knowledge, there is no method for approaching SSC without the need for any ground-truth annotations. Our work presents the first unsupervised SSC approach, utilizing lifted SSL features and a single RGB input image for inference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.306, + 0.905, + 0.488 + ], + "angle": 0, + "content": "Self-supervised representation learning (SSL) aims to extract general features from data without annotations, facilitating various downstream tasks such as segmentation [24]. Recent SSL methods, often based on Vision Transformers (ViTs) [23], leverage clustering [2, 9, 10, 48, 61], masked modeling [20, 29, 40, 77, 107], contrastive learning [3, 12, 14, 39, 41, 42], or negative-free [4, 5, 11, 28, 80] pretext tasks [22, 79] for large-scale training. State-of-the-art models, e.g., DINO [11], produce semantically rich, dense features, driving recent advances in 2D unsupervised scene understanding [33, 104]. We here aim to bring expressive features from DINO [11, 80] to 3D for SSC." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.49, + 0.905, + 0.778 + ], + "angle": 0, + "content": "2D-to-3D feature lifting. The expressiveness of 2D visual representations has motivated lifting 2D features into 3D [94, 111]. Existing approaches utilize multi-view 2D features for 3D feature lifting [30, 44, 50, 54, 73, 83, 93, 94, 98, 101, 102, 106, 111, 116]. Lifting 2D features is effective in various tasks, including few-shot semantic occupancy prediction [111], and refining 2D representations [116]. However, existing feature-lifting approaches fit to a single scene [50, 54, 93, 94, 101, 102, 111, 116], require RGB-D inputs [30, 44, 73, 98, 106], or work on 3D point cloud inputs [83]. The only feed-forward approaches that use RGB inputs and lift 2D features, which we are aware of, are GaussTR [49] and MVSplat360 [15]. However, both approaches utilize multiple input images during inference, and MVSplat360 [15] only predicts low-dimensional feature representations, which are not suitable for unsupervised scene understanding. In contrast, we propose the first feedforward approach for inferring lifted high-dimensional and rich 3D features using a single input image." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.905, + 0.901 + ], + "angle": 0, + "content": "2D unsupervised semantic segmentation partitions images automatically into semantically meaningful regions without any form of human annotations. Early deep learning-based methods [17, 36, 48] approach the problem via representation learning. Leveraging SSL features from DINO as an inductive prior, STEGO [33] distills the feature representation into a lower-dimensional space for unsupervised probing. Building up on STEGO, subsequent methods" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6785" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.091, + 0.348, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.103, + 0.61, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.089, + 0.885, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.26, + 0.907, + 0.345 + ], + "angle": 0, + "content": "Figure 2. SceneDINO architecture, rendering, and training. (a) Inference: Given a single input image \\(\\mathbf{I}_0\\) during inference, a 2D encoder-decoder \\(\\xi\\) produces the embedding \\(\\mathbf{E}\\) from which the local embedding \\(\\mathbf{e}_{\\mathbf{u}}\\) is interpolated. The MLP encoder \\(\\phi\\) takes in \\(\\mathbf{e}_{\\mathbf{u}}\\) and 3D position \\(\\mathbf{x}_i\\), and predicts both the density \\(\\sigma_{\\mathbf{x}_i}\\) and the 3D feature \\(f_{\\mathbf{x}_i}\\). Using a lightweight unsupervised segmentation head \\(h\\), we can obtain semantic predictions \\(p_{\\mathbf{x}_i}\\) using \\(f_{\\mathbf{x}_i}\\). (b) Rendering: Our feature field allows for volume rendering by shooting rays through it, yielding depth \\(\\hat{d}\\) and \\(\\hat{f}\\) in 2D. Color \\(c_i\\) is sampled from an another view (\\(e.g., \\mathbf{I}_1\\)) using \\(\\mathbf{u}_s\\) and rendered to obtain the reconstructed color \\(\\hat{c}\\). (c) Multi-view training: We render 2D views (features & images) from our feature field and reconstruct the training views." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.364, + 0.482, + 0.426 + ], + "angle": 0, + "content": "[31, 51, 92, 95] propose enhancements to the distillation. Our approach follows the idea of STEGO [33], extending it to 3D and integrating feature distillation using our 3D feature field to build the first unsupervised SSC approach." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.439, + 0.467, + 0.456 + ], + "angle": 0, + "content": "3. Unsupervised Semantic Scene Completion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.465, + 0.483, + 0.602 + ], + "angle": 0, + "content": "We approach semantic scene completion (SSC) without any form of manual supervision. To this end, we first describe SceneDINO, predicting 3D geometry and expressive 3D features from a single image in a feed-forward manner (Sec. 3.1), and SceneDINO's multi-view training (Sec. 3.2). Next, we present our 3D feature distillation approach to obtain unsupervised 3D semantic predictions (Sec. 3.3). An overview of our full pipeline, including inference, rendering, and multi-view self-supervision, is provided in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.603, + 0.483, + 0.729 + ], + "angle": 0, + "content": "Notation. Let \\(\\mathbf{I}_0\\in [0,1]^{3\\times \\mathrm{H}\\times \\mathrm{W}}\\) be a single RGB input image (for both training & inference) with corresponding pose \\(T_0\\in \\mathbb{R}^{4\\times 4}\\) and projection matrix \\(K_{0}\\in \\mathbb{R}^{3\\times 4}\\). For training, let \\((\\mathbf{I}_v,T_v,K_v)\\) with \\(v\\in \\{1,2,\\dots ,n\\}\\), be \\(n\\) additional views for multi-view self-supervision. Assuming a pinhole camera model, any 3D point \\(\\mathbf{x}\\in \\mathbb{R}^3\\) in world coordinates can be projected onto the image plane of view \\(v\\) and the input view \\(v = 0\\) with the perspective projection \\(\\pi_v(\\mathbf{x})\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.738, + 0.219, + 0.753 + ], + "angle": 0, + "content": "3.1. SceneDINO" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.76, + 0.483, + 0.867 + ], + "angle": 0, + "content": "Given a single input image \\(\\mathbf{I}_0\\), SceneDINO represents the dense geometric structure and features of a scene as a continuous mapping from world coordinates \\(\\mathbf{x} \\in \\mathbb{R}^3\\) to a volumetric density \\(\\sigma_{\\mathbf{x}} \\in \\mathbb{R}_{+}\\) and a feature \\(f_{\\mathbf{x}} \\in \\mathbb{R}^{\\mathrm{D}}\\). This continuous output representation is often called a feature field. While SceneDINO could represent any feature space, we aim for expressive SSL features from DINO [11, 80]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Architecture & feature field inference. Our SceneDINO architecture comprises two main parts: a 2D encoder-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.907, + 0.439 + ], + "angle": 0, + "content": "decoder \\(\\xi\\) and an MLP decoder (cf. Fig. 2a), following BTS [108]. \\(\\xi\\) takes in \\(\\mathbf{I}_0\\) and produces a per-pixel embedding \\(\\mathbf{E} \\in \\mathbb{R}^{\\mathrm{D}_{\\mathbf{E}} \\times \\mathrm{H} \\times \\mathrm{W}}\\) with \\(\\mathrm{D}_{\\mathbf{E}}\\) dimensions. Intuitively, every spatial element of \\(\\mathbf{E}\\) represents a camera ray through a pixel, capturing both local geometry and features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.44, + 0.907, + 0.561 + ], + "angle": 0, + "content": "To infer the feature at a 3D position \\(\\mathbf{x}\\), we employ a two-layer MLP decoder \\(\\phi\\) (cf. Fig. 2a). Given a position \\(\\mathbf{x}\\) within the camera frustum, we project \\(\\mathbf{x}\\) into the camera plane, obtaining the pixel location \\(\\mathbf{u} = \\pi_0(\\mathbf{x})\\). We query \\(\\mathbf{E}\\) at the position \\(\\mathbf{u}\\) using bilinear interpolation, obtaining the local embedding \\(\\mathbf{e_u}\\). Based on the embedding \\(\\mathbf{e_u}\\), the pixel position \\(\\mathbf{u}\\), and the distance \\(d_{\\mathbf{x}} \\in \\mathbb{R}_{+}\\) of \\(\\mathbf{x}\\) to the camera, we obtain the density \\(\\sigma_{\\mathbf{x}}\\) and feature prediction \\(f_{\\mathbf{x}}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.615, + 0.566, + 0.905, + 0.584 + ], + "angle": 0, + "content": "\\[\n\\left(\\sigma_ {\\mathbf {x}}, f _ {\\mathbf {x}}\\right) = \\phi \\left(\\mathbf {e} _ {\\mathbf {u}}, \\gamma (\\mathbf {u}, d _ {\\mathbf {x}})\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.588, + 0.807, + 0.604 + ], + "angle": 0, + "content": "where \\(\\gamma\\) denotes a positional encoding [75]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.606, + 0.907, + 0.758 + ], + "angle": 0, + "content": "Feature, depth & color volume rendering. SceneDINO predicts a continuous feature field from a single image. This representation can be used to render features and depth in 2D from an arbitrary viewpoint (cf. Fig. 2b), following the discretization strategy of Max et al. [72]. Given a viewpoint \\((T,K)\\), we sample L points \\(\\mathbf{x}_i\\) along the ray through pixel \\(\\mathbf{u}_r\\), with distance \\(\\delta_i\\) between \\(\\mathbf{x}_i\\) and \\(\\mathbf{x}_{i+1}\\). Based on the volumetric densities \\(\\sigma_{\\mathbf{x}_i}\\) (cf. Eq. 1), we can compute the probabilities \\(\\alpha_i\\) of the ray ending between \\(\\mathbf{x}_i\\) and \\(\\mathbf{x}_{i+1}\\), and accumulate these into \\(V_i\\), the probability of \\(\\mathbf{x}_i\\) being visible:" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.763, + 0.905, + 0.806 + ], + "angle": 0, + "content": "\\[\nV _ {i} = \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\quad \\text {w i t h} \\alpha_ {i} = 1 - \\exp (- \\sigma_ {\\mathbf {x} _ {i}} \\delta_ {i}). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.859 + ], + "angle": 0, + "content": "Using \\(V_{i}\\) and \\(\\alpha_{i}\\), we render depth \\(\\tilde{d}_{\\mathbf{u}_r}\\) and feature \\(\\tilde{f}_{\\mathbf{u}_r}\\) from the estimated features \\(f_{\\mathbf{x}_i}\\) from Eq. (1) and distances \\(d_{\\mathbf{x}_i}\\) to \\(\\mathbf{x}_i\\) onto the image plane at position \\(\\mathbf{u}_r\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.863, + 0.907, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\tilde {f} _ {\\mathbf {u} _ {r}} = \\sum_ {i = 1} ^ {\\mathrm {L}} V _ {i} \\alpha_ {i} f _ {\\mathbf {x} _ {i}} \\quad \\tilde {d} _ {\\mathbf {u} _ {r}} = \\sum_ {i = 1} ^ {\\mathrm {L}} V _ {i} \\alpha_ {i} d _ {\\mathbf {x} _ {i}}. \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6786" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.198 + ], + "angle": 0, + "content": "The differentiability of this rendering process enables us to self-supervise SceneDINO using multi-view images and their 2D feature representations (e.g., from DINO [11]). SceneDINO predicts 3D geometry and features, but does not predict color as we focus on semantic downstream tasks. To obtain color for image reconstruction during training, we follow the color sampling approach of BTS [108]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.206, + 0.314, + 0.223 + ], + "angle": 0, + "content": "3.2. 3D feature field training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.228, + 0.485, + 0.457 + ], + "angle": 0, + "content": "We train SceneDINO using multi-view self-supervision (cf. Fig. 2c), aiming to obtain an expressive and view-consistent feature field without the need for any form of manual annotations. For self-supervision, we sample \\( n + 1 \\) views \\( \\mathbf{I}_v \\) with camera parameters \\( T_v, K_v \\) from the data and obtain dense 2D target features \\( \\mathbf{F}_v \\) from a self-supervised ViT (e.g., DINO [11]). Note that the 2D features entail a resolution of \\( \\mathbf{F}_v \\in \\mathbb{R}^{\\mathrm{D} \\times \\frac{\\mathrm{H}}{p} \\times \\frac{\\mathrm{W}}{p}} \\), due to the ViT patch size \\( p \\). The set of training views and features \\( \\mathbb{V} = \\{( \\mathbf{I}_v, T_v, K_v, \\mathbf{F}_v) | v = 0, \\dots, n \\} \\) is randomly partitioned into two subsets \\( \\mathbb{V}_{\\text{source}} \\) and \\( \\mathbb{V}_{\\text{target}} \\). Training reconstructs the views \\( \\mathbb{V}_{\\text{target}} \\) using the views of \\( \\mathbb{V}_{\\text{source}} \\). In practice, we use a randomly sampled set of image patches that align with the ViT patches instead of the full image. In the following, we still refer to images and the full image resolution for the sake of brevity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.46, + 0.484, + 0.569 + ], + "angle": 0, + "content": "Image reconstruction. We aim to learn the geometry of our feature field via multi-view photometric consistency. In particular, for every image \\(\\mathbf{I}_t\\in \\mathbb{V}_{\\mathrm{target}}\\) we derive a reconstructed image \\(\\hat{\\mathbf{I}}_{t,s}\\) from every view \\(s\\) in \\(\\mathbb{V}_{\\mathrm{source}}\\) using differentiable rendering (cf. Eq. 3) and color sampling [108]. Equipped with both the reconstructed image \\(\\hat{\\mathbf{I}}_{t,s}\\) and the target image \\(\\mathbf{I}_t\\), we compute our photometric loss per view as" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.578, + 0.484, + 0.606 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {p}} = \\min _ {\\mathbf {I} _ {s} \\in \\mathbb {V} _ {\\text {s o u r c e}}} \\left(\\lambda_ {1} \\mathcal {L} _ {1} \\left(\\mathbf {I} _ {t}, \\hat {\\mathbf {I}} _ {t, s}\\right) + \\lambda_ {\\text {S S I M}} \\mathcal {L} _ {\\text {S S I M}} \\left(\\mathbf {I} _ {t}, \\hat {\\mathbf {I}} _ {t, s}\\right)\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.659 + ], + "angle": 0, + "content": "We only consider the minimum loss across the views in \\(\\mathbb{V}_{\\mathrm{source}}\\), in practice across patches. The scalars \\(\\lambda_{1}\\) and \\(\\lambda_{\\mathrm{SSIM}}\\) weight the absolute error \\(\\mathcal{L}_1\\) and the SSIM loss \\(\\mathcal{L}_{\\mathrm{SSIM}}\\) [105]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.484, + 0.737 + ], + "angle": 0, + "content": "To regularize our 3D geometry, we impose smoothness using an edge-aware smoothness loss [27]. We estimate 2D depth maps \\(\\tilde{\\mathbf{d}}_t\\) for views in \\(\\mathbb{V}_{\\mathrm{target}}\\), using Eq. 3. From \\(\\tilde{\\mathbf{d}}_t\\), we obtain the inverse mean-normalized depths \\(\\tilde{\\mathbf{d}}_t^*\\) and compute the edge-aware smoothness loss \\(\\mathcal{L}_{\\mathrm{s}}\\) per view as" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.744, + 0.483, + 0.765 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {s}} = \\left| \\nabla_ {x} \\tilde {\\mathbf {d}} _ {t} ^ {*} \\right| e ^ {- \\left| \\nabla_ {x} \\mathbf {I} _ {t} \\right|} + \\left| \\nabla_ {y} \\tilde {\\mathbf {d}} _ {t} ^ {*} \\right| e ^ {- \\left| \\nabla_ {y} \\mathbf {I} _ {t} \\right|}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.773, + 0.442, + 0.789 + ], + "angle": 0, + "content": "where \\(\\nabla_{x}\\) and \\(\\nabla_{y}\\) denote the first spatial derivatives." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.791, + 0.484, + 0.868 + ], + "angle": 0, + "content": "Feature reconstruction. We learn a multi-view consistent and expressive 3D feature field using the 2D features \\(\\mathbf{F}_t\\) from \\(\\mathbb{V}_{\\mathrm{target}}\\). As we aim to learn a high-resolution (continuous) feature field, we render 2D features using Eq. 3 at the full image resolution \\(\\hat{\\mathbf{F}}_t \\in \\mathbb{R}^{D \\times H \\times W}\\). To compensate for" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.088, + 0.906, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.232, + 0.906, + 0.302 + ], + "angle": 0, + "content": "Figure 3. 3D feature distillation. Given an input image, SceneDINO predicts a 3D feature field. 3D features \\( f_{\\mathbf{X}} \\) are sampled from the feature field. For \\( f_{\\mathbf{X}} \\), we obtain \\( f_{\\mathbf{Y}_{k\\mathrm{NN}}} \\) and \\( f_{\\mathbf{Y}_{kr\\mathrm{and}}} \\) from the feature buffer. The segmentation head \\( h \\) distills the features into a low-dimensional space and is trained using \\( \\mathcal{L}_{\\mathrm{dist}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.324, + 0.907, + 0.476 + ], + "angle": 0, + "content": "the reduced spatial dimension of \\(\\mathbf{F}_t\\), we employ the down-sampler \\(\\psi\\) proposed by Fu et al. [25] to our rendered features \\(\\hat{\\mathbf{F}}_t\\). While current 2D SSL features capture semantics, they lack multi-view consistency, i.a., due to positional encodings used in ViTs [112], leading to different features for identical visual content at two distinct positions in an image. As we aim for multi-view consistency, we compensate for this by learning a constant decomposition \\(\\overline{\\mathbf{F}} \\in \\mathbb{R}^{\\mathrm{D} \\times \\frac{\\mathrm{H}}{p} \\times \\frac{\\mathrm{W}}{p}}\\) of features induced by positional encodings. Our feature loss is defined per view as" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.487, + 0.905, + 0.505 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {f}} = 1 - \\cos - \\sin \\left(\\mathbf {F} _ {t}, \\psi \\left(\\hat {\\mathbf {F}} _ {t}\\right) + \\overline {{\\mathbf {F}}}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.515, + 0.904, + 0.53 + ], + "angle": 0, + "content": "where cos-sim is the cosine similarity between two features." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.53, + 0.905, + 0.576 + ], + "angle": 0, + "content": "As image edges correlate with semantic edges and to further impose consistency, we regularize the rendered features \\(\\hat{\\mathbf{F}}_t\\) using an edge-aware smoothness loss per view" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.584, + 0.905, + 0.604 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {f s}} = \\left| \\nabla_ {x} \\hat {\\mathbf {F}} _ {t} \\right| e ^ {- \\left| \\nabla_ {x} \\mathbf {I} _ {t} \\right|} + \\left| \\nabla_ {y} \\hat {\\mathbf {F}} _ {t} \\right| e ^ {- \\left| \\nabla_ {y} \\mathbf {I} _ {t} \\right|}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.906, + 0.674 + ], + "angle": 0, + "content": "Our final loss for training SceneDINO is a weighted sum of the photometric loss, the feature loss, and both smoothness losses, \\(\\mathcal{L}_{\\mathrm{SceneDINO}} = \\lambda_{\\mathrm{p}}\\mathcal{L}_{\\mathrm{p}} + \\lambda_{\\mathrm{s}}\\mathcal{L}_{\\mathrm{s}} + \\lambda_{\\mathrm{f}}\\mathcal{L}_{\\mathrm{f}} + \\lambda_{\\mathrm{fs}}\\mathcal{L}_{\\mathrm{fs}}\\), averaged over all pixels, features, and views." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.682, + 0.887, + 0.699 + ], + "angle": 0, + "content": "3.3. 3D feature distillation for unsupervised SSC" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.906, + 0.871 + ], + "angle": 0, + "content": "Given the expressive feature field representation, we aim to obtain unsupervised semantic predictions for SSC. While naive \\( k \\)-means [69, 71] can yield meaningful pseudo semantics, distilling features into a lower-dimensional space has been shown to be more effective in 2D semantic segmentation [33, 55]. To this end, we present a novel 3D feature distillation approach (cf. Fig. 3). We train a pointwise segmentation head \\( h \\), mapping \\( f_{\\mathbf{x}} \\in \\mathbb{R}^{\\mathrm{D}} \\) to a lower-dimensional distilled representation \\( z_{\\mathbf{x}} \\in \\mathbb{R}^{\\mathrm{K}} \\), with \\( \\mathrm{K} \\ll \\mathrm{D} \\). The resulting distilled space is clustered to obtain pseudosemantic predictions \\( p_{\\mathbf{x}} \\in [0,1]^{\\mathrm{C}} \\), with C pseudo classes." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Existing work in 2D unsupervised semantic segmentation has shown that SSL feature correspondence captures" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "1Note, camera poses can be obtained using unsupervised visual SLAM [7], strictly adhering to the fully unsupervised setting." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6787" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.198 + ], + "angle": 0, + "content": "semantic class co-occurrence [33]. This correspondence between two batches of \\(\\mathbf{N}\\) sample points \\(\\mathbf{X} = [\\mathbf{x}_1, \\dots, \\mathbf{x}_N]\\) and \\(\\mathbf{Y} = [\\mathbf{y}_1, \\dots, \\mathbf{y}_N]\\) can be expressed by pairwise feature similarity \\(S_{i,j} = \\cos -\\sin (f_{\\mathbf{x}_i}, f_{\\mathbf{y}_j}) \\in [-1, 1]\\). Similarly, we can express the correspondence in the distilled feature space by \\(S_{i,j}^h = \\cos -\\sin (h(f_{\\mathbf{x}_i}), h(f_{\\mathbf{y}_j})) \\in [-1, 1]\\). We describe the sampling of the \\(\\mathbf{x}_i\\) and \\(\\mathbf{y}_j\\) below." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.201, + 0.484, + 0.262 + ], + "angle": 0, + "content": "Feature distillation. We aim to distill features such that similar features align while dissimilar features are separated. To this end, we use the contrastive correlation loss \\(\\mathcal{L}_{\\mathrm{corr}}\\), introduced by STEGO [33] and defined as" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.272, + 0.484, + 0.305 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o r r}} \\left(f _ {\\mathbf {X}}, f _ {\\mathbf {Y}}, b\\right) = - \\sum_ {i, j} \\left(S _ {i, j} - b\\right) \\max \\left(S _ {i, j} ^ {h}, 0\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.311, + 0.483, + 0.359 + ], + "angle": 0, + "content": "where \\( f_{\\mathbf{X}} \\), \\( f_{\\mathbf{Y}} \\) are the features of the two sample batches. This loss pushes \\( S_{i,j}^{h} \\) towards 1 in case \\( S_{i,j} \\) exceeds the threshold \\( b \\). Otherwise, \\( \\mathcal{L}_{\\mathrm{corr}} \\) pushes the \\( S_{i,j}^{h} \\) below 0." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.359, + 0.484, + 0.51 + ], + "angle": 0, + "content": "The correlation loss \\(\\mathcal{L}_{\\mathrm{corr}}\\) requires informative pairs of sampled features, balancing attractive and repulsive signals. Following STEGO [33], we consider three different relations: (1) feature pairs from the same image \\((f_{\\mathbf{X}}, f_{\\mathbf{X}})\\), (2) feature pairs from an image and its \\(k\\)-nearest neighbors in feature space \\((f_{\\mathbf{X}}, f_{\\mathbf{Y}_{k\\mathrm{NN}}})\\), and (3) feature pairs from an image and a randomly sampled other image \\((f_{\\mathbf{X}}, f_{\\mathbf{Y}_{\\mathrm{rand}}})\\). Note that each pair is obtained from SceneDINO's 3D feature field, see below. Equipped with the three feature sample pairs, we compute the full distillation loss as" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.52, + 0.482, + 0.57 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {d i s t}} = \\lambda_ {\\mathrm {s e l f}} \\mathcal {L} _ {\\mathrm {c o r r}} \\left(f _ {\\mathbf {X}}, f _ {\\mathbf {X}}, b _ {\\mathrm {s e l f}}\\right) \\\\ + \\lambda_ {k \\mathrm {N N}} \\mathcal {L} _ {\\text {c o r r}} \\left(f _ {\\mathbf {X}}, f _ {\\mathbf {Y} _ {k \\mathrm {N N}}}, b _ {k \\mathrm {N N}}\\right) \\tag {9} \\\\ + \\lambda_ {\\text {r a n d}} \\mathcal {L} _ {\\text {c o r r}} \\left(f _ {\\mathbf {X}}, f _ {\\mathbf {Y} _ {\\text {r a n d}}}, b _ {\\text {r a n d}}\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.581, + 0.483, + 0.611 + ], + "angle": 0, + "content": "where \\(\\lambda_{\\mathrm{self}}\\), \\(\\lambda_{k\\mathrm{NN}}\\), and \\(\\lambda_{\\mathrm{rand}}\\) denote the scalar loss weights. \\(b_{\\mathrm{self}}\\), \\(b_{k\\mathrm{NN}}\\), and \\(b_{\\mathrm{rand}}\\) are the contrastive thresholds." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.484, + 0.734 + ], + "angle": 0, + "content": "Feature sampling in 3D. While obtaining feature pairs using 2D rendered features is straightforward [33], we aim to take advantage of our learned 3D geometry of the scene. To this end, we introduce a novel 3D feature sampling approach for the distillation loss \\(\\mathcal{L}_{\\mathrm{dist}}\\) from Eq. (9). Our goal is to sample features both similar and dissimilar in terms of the encoded semantic concept, which should capture rich semantics as well as different semantic concepts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.856 + ], + "angle": 0, + "content": "First, we obtain all G visible 3D surface points \\(\\mathbf{V} \\in \\mathbb{R}^{3 \\times \\mathrm{G}}\\) and their depth \\(d_{\\mathbf{V}} \\in \\mathbb{R}_{+}^{\\mathrm{G}}\\) from the camera. To sample points that cover different semantic concepts, we use depth as a cue and sample different depth ranges. In particular, we sort the surface points \\(\\mathbf{V}\\) based on \\(d_{\\mathbf{V}}\\). The sorted surface points \\(\\hat{\\mathbf{V}}\\) are partitioned into M equally-sized chunks; we uniformly sample a single 3D point from each chunk, resulting in M center points \\(\\mathbf{X} \\in \\mathbb{R}^{3 \\times \\mathrm{M}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Equipped with the center points \\(\\mathbf{X}\\), we aim to extract rich semantic features from the feature field. While we could just obtain the features for \\(\\mathbf{X}\\), we query positions in the" + }, + { + "type": "image", + "bbox": [ + 0.585, + 0.087, + 0.835, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.218, + 0.906, + 0.289 + ], + "angle": 0, + "content": "Figure 4. 3D feature sampling. We first sample a center point \\(\\mathbf{X}_i\\) from all visible surface points. Further points are sampled within the radius \\(r\\) around the center point \\(\\mathbf{X}_i\\). Sampled points with sufficient density are accepted; otherwise rejected. The accepted points are used to obtain the feature batch \\(f_{\\mathbf{X}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.308, + 0.906, + 0.459 + ], + "angle": 0, + "content": "neighborhood of \\(\\mathbf{X}\\) to increase semantic richness and better capture the 3D structure of the scene for distillation. In particular, for each center point, we randomly sample a point within a radius of \\(r = 0.5\\mathrm{m}\\). To account for samples falling into unoccupied regions in our feature field, we only keep samples with a sufficient density \\(\\sigma > 0.5\\). We repeat this sampling process until we obtain \\(N\\) valid samples per center point. Using these samples, we query our feature field, resulting in a feature batch \\(f_{\\mathbf{X}} \\in \\mathbb{R}^{D \\times N}\\) for each of the G center points in each scene (cf. Fig. 4)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.459, + 0.907, + 0.64 + ], + "angle": 0, + "content": "To obtain \\( f_{\\mathbf{Y}_{k\\mathrm{NN}}} \\) and \\( f_{\\mathbf{Y}_{\\mathrm{rand}}} \\), we utilize a feature buffer that efficiently stores the sampled features of multiple scenes. Given a new input image, we obtain \\( G \\) feature batches \\( f_{\\mathbf{X}} \\) as just described. For each \\( f_{\\mathbf{X}} \\), we randomly sample another feature batch from the buffer to obtain \\( f_{\\mathbf{Y}_{\\mathrm{rand}}} \\). To obtain \\( f_{\\mathbf{Y}_{k\\mathrm{NN}}} \\), we search in the feature buffer for the \\( k \\)-nearest neighbors of \\( f_{\\mathbf{X}} \\), using the average feature of each batch. From these \\( k \\)-nearest neighbors, we randomly pick a feature batch to obtain \\( f_{\\mathbf{Y}_{k\\mathrm{NN}}} \\) and compute the distillation loss \\( \\mathcal{L}_{\\mathrm{dist}} \\). After repeating this process for each of the current \\( G \\) feature batches, we add the current feature batches to the feature buffer and remove the oldest batches." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.645, + 0.906, + 0.735 + ], + "angle": 0, + "content": "Unsupervised probing. To obtain semantics, we probe the distilled K-dim. feature space using \\(k\\)-means [69, 71]. In particular, we update cluster centers \\(\\theta \\in \\mathbb{R}^{K\\times C}\\) using cosine similarity-based mini-batch \\(k\\)-means [91] during distillation. We compute \\(p_{\\mathbf{x}} = \\mathrm{softmax}(\\cos -\\mathrm{sim}(h(f_{\\mathbf{x}}),\\theta))\\) to infer C pseudo semantic class predictions." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.75, + 0.646, + 0.767 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.776, + 0.906, + 0.867 + ], + "angle": 0, + "content": "We evaluate SceneDINO on SSC and compare it to a simple unsupervised SSC baseline (Sec. 4.1). We also report results for 2D unsupervised segmentation, including domain generalization results (Sec. 4.2). Finally, we explore multiview feature consistency (Sec. 4.3) and present an analysis of SceneDINO and our 3D distillation (Sec. 4.4)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Datasets. We train using KITTI-360 [66], composed of clips from a moving vehicle equipped with cameras. For" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6788" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.304 + ], + "angle": 0, + "content": "consistency, we follow S4C [38] by sampling eight views and using the dataset's camera poses. We further provide results with estimated poses. We also show experiments for training on RealEstate10k [119], composed of monocular videos. Here, we follow the setup of BTS [108], obtaining three views. If not noted differently, we report results obtained with training on KITTI-360. For SSC and 2D semantic segmentation validation, we use the SSCBench-KITTI-360 test split [64]. Cityscapes [19] and BDD100K [115] val are used for domain generalization results. To enable evaluation in 3D and 2D, we use the 19-class taxonomy of Cityscapes and perform 2D evaluation on Cityscapes, BDD100K, and KITTI-360 on 19 classes. For SSCBench, we combine classes to adhere to the 15 SSCBench classes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.309, + 0.482, + 0.476 + ], + "angle": 0, + "content": "3D evaluation. Given our unsupervised setup, we predict pseudo-semantic classes that must be aligned with the ground truth for evaluation. We follow standard practice in 2D unsupervised semantic segmentation [17, 31, 33, 51, 92, 95] by applying Hungarian matching [57] to align our pseudo semantics. For validating the aligned semantics, we follow the standardized SSCBench [64] protocol and report both semantic performance using the mean Intersection-over-Union (mIoU) and geometric performance using IoU, precision, and recall. We report all metrics on SSCBench ranges \\(12.8\\mathrm{m}\\), \\(25.6\\mathrm{m}\\), and \\(51.2\\mathrm{m}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.481, + 0.482, + 0.587 + ], + "angle": 0, + "content": "2D evaluation. Following the established evaluation protocol in 2D unsupervised semantic segmentation [17, 31, 33, 51, 92, 95], we use the all-pixel accuracy (Acc) and mean Intersection-over-Union (mIoU) metrics. Likewise, in line with prior work, 2D segmentation predictions of all models are refined using a dense Conditional Random Field [56] before computing Acc and mIoU." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.593, + 0.482, + 0.744 + ], + "angle": 0, + "content": "Multi-view feature consistency evaluation. We aim to evaluate the multi-view consistency of our feature field. As we are not aware of any general feed-forward 3D feature field approach, we compare against 2D SSL models. To measure multi-view consistency in 2D, we use two video frames and estimate optical flow and occlusions with RAFT [99]. We backward warp 2D features of the second frame to the first. On the aligned features, we compute the feature similarity using absolute error \\((\\mathrm{L}_1)\\), the Euclidean distance \\((\\mathrm{L}_2)\\), and the cosine similarity, ignoring occlusions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Baselines. We are not aware of any existing unsupervised SSC approach. To allow for comparisons, we construct a competitive baseline for unsupervised SSC. In particular, we train the S4C approach with unsupervised semantics of the established STEGO [33] approach. For 2D semantic segmentation, we use U2Seg [78] and STEGO [33] as established unsupervised baselines. Note that U2Seg is trained on ImageNet [21] and COCO [67] using STEGO pseudo-labels. We use STEGO [33] with DINO [11] (ViT-B/8), DINOv2 [80] (ViT-B/14), and FiT3D [116] (ViT" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.089, + 0.905, + 0.16 + ], + "angle": 0, + "content": "Table 1. SSCBench-KITTI-360 results. Semantic results using mIoU and per class IoU, and geometric results using IoU, Precision, and Recall (all in %, ↑) on SSCBench-KITTI-360 test using three depth ranges. We compare our baseline S4C + STEGO to our SceneDINO. We report S4C as a 2D supervised baseline." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.165, + 0.905, + 0.416 + ], + "angle": 0, + "content": "
MethodS4C + STEGOSceneDINO (Ours)S4C
SupervisionUnsupervised2D supervision
Range12.8 m25.6 m51.2 m12.8 m25.6 m51.2 m12.8 m25.6 m51.2 m
Semantic validation
mIoU10.539.266.6010.7610.018.0016.9413.9410.19
car18.5714.099.2221.2415.9411.2122.5818.6411.49
bicycle0.010.010.010.000.000.000.000.000.00
motorcycle0.000.000.000.000.000.000.000.000.00
truck0.110.040.020.000.000.007.514.372.12
other-v.0.010.050.020.000.000.000.000.010.06
person0.010.010.010.000.000.000.000.000.00
road61.9752.4738.1551.1049.1239.8269.3861.4648.23
sidewalk18.7420.9518.2120.2622.3118.9745.0337.1228.45
building14.7524.4417.8112.3318.2714.3226.3428.4821.36
fence1.410.200.111.910.900.589.706.373.64
vegetation15.8316.5811.3031.2225.5719.8535.7828.0421.43
terrain26.499.954.1723.2618.0215.2235.0322.8815.08
pole0.080.040.040.050.050.051.230.940.65
traffic-sign0.000.000.000.000.000.001.570.830.36
other_obj.0.050.040.020.000.000.000.000.000.00
Geometric validation
IoU49.3241.0836.3949.5442.2737.6054.6445.5739.35
Precision54.0446.2341.9153.2746.1041.5959.7550.3443.59
Recall84.9578.6973.4387.6183.5979.6786.4782.7980.16
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.438, + 0.907, + 0.558 + ], + "angle": 0, + "content": "B/14) features. FiT3D offers multi-view refined DINOv2 features [116]. Note that FiT3D reports results, concatenating the refined features with DINOv2 features. We report results using both plain features only and the concatenation. We also use rendered 2D segmentations of our S4C + STEGO baseline for 2D validation. For multi-view feature consistency, we utilize DINO [11], DINOv2, and FiT3D [116] features as baselines." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.564, + 0.909, + 0.821 + ], + "angle": 0, + "content": "Implementation details. Our encoder-decoder uses a DINO-B/8 [11] backbone and a dense prediction decoder [84]. The MLP decoder \\(\\phi\\) entails two layers with 128 hidden features. As rendering features is expensive, \\(\\phi\\) predicts 64 features. We employ another MLP to upproject again to the full dimensionality \\(D = 768\\). If not stated differently, our target features are obtained from DINO-B/8 [11]. We train using a batch size of 4 and extract 32 patches of size \\(8 \\times 8\\) from each image to compute \\(\\mathcal{L}_{\\mathrm{SceneDINO}}\\). Volume rendering samples each ray at \\(L = 32\\) uniformly spaced points in inverse depth within \\([3\\mathrm{m}, 80\\mathrm{m}]\\). We train for \\(100\\mathrm{k}\\) steps using Adam [52] with a base learning rate of \\(10^{-4}\\). Training takes ca. 2 days on a single V100 GPU. We distill using a batch size of 4, 5 center points, a feature batch of size 576, and cluster with \\(K = 19\\). For kNN sampling, we use \\(k = 4\\). The feature buffer holds 256 feature batches. Refer to the supplement for more details." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.78, + 0.85 + ], + "angle": 0, + "content": "4.1.3D semantic scene completion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.901 + ], + "angle": 0, + "content": "We assess the unsupervised SSC and geometric accuracy of SceneDINO with our 3D feature distillation approach on SSCBench-KITTI-360. In particular, Tab. 1 com" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6789" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.086, + 0.91, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.234, + 0.908, + 0.278 + ], + "angle": 0, + "content": "Figure 5. Qualitative SSC comparison on KITTI-360. We show the input image, SceneDINO's feature field using the first three principal components and SSC prediction, the prediction of our baseline S4C + STEGO, and the ground truth. We only visualize surface voxels. Qualitative results show the expressiveness of our feature field and SceneDINO's capabilities to accurately reconstruct and label a scene." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.295, + 0.487, + 0.366 + ], + "angle": 0, + "content": "Table 2. 2D unsupervised semantic segmentation results on KITTI-360. Comparing SceneDINO to existing 2D methods and our S4C + STEGO 3D baseline, using Accuracy and mean IoU (in \\(\\%\\), \\(\\uparrow\\)) on the SSCBench-KITTI-360 test split. \\(\\dagger\\) denotes the use of plain FiT3D features. \\(\\ddagger\\) denotes training on ImageNet and COCO." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.37, + 0.485, + 0.476 + ], + "angle": 0, + "content": "
MethodFeaturesAccmIoU
U2Seg‡ [78]-72.8923.43
STEGO [33]DINO [11]73.3223.57
STEGO [33]DINOv2 [80]64.5424.82
STEGO [33]FiT3D [116]54.1922.29
STEGO [33]FiT3D† [116]57.2518.95
S4C [38] + STEGO [33]DINO [11]65.1621.67
SceneDINO (Ours)DINO [11]77.7425.81
" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.499, + 0.483, + 0.635 + ], + "angle": 0, + "content": "pares SceneDINO against our unsupervised SSC baseline S4C [38] + STEGO [33]. SceneDINO achieves a (semantic) mIoU of \\(8.0\\%\\) for the range of \\(51.2\\mathrm{m}\\), significantly improving over our unsupervised baseline \\((6.6\\%)\\). This demonstrates that SceneDINO effectively lifts DINO features into 3D. In terms of geometric accuracy, SceneDINO moderately improves over S4C + STEGO. Despite being fully unsupervised, SceneDINO comes within \\(2.2\\%\\) points mIoU of the 2D-supervised S4C." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.637, + 0.484, + 0.773 + ], + "angle": 0, + "content": "Figure 5 provides qualitative samples on SSCBenchKITTI-360. SceneDINO's unsupervised SSC predictions are less noisy and capture finely resolved semantics compared to S4C + STEGO. Compared to the ground truth, we observe that SceneDINO captures both the geometry and general semantics of the scene. We visualize SceneDINO's feature field (before distillation) using the first three principal components. In PCA space, we observe that our feature field captures semantically meaningful regions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.787, + 0.329, + 0.803 + ], + "angle": 0, + "content": "4.2. 2D semantic segmentation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Table 2 compares the semantic predictions of SceneDINO to recent 2D approaches and our 3D baseline. We obtain 2D semantic segmentations from SceneDINO and our S4C + STEGO baseline using semantic rendering [38]. SceneDINO with our 3D distillation approach outperforms STEGO with DINO features, an established 2D unsuper" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.295, + 0.908, + 0.38 + ], + "angle": 0, + "content": "Table 3. 2D unsupervised semantic segmentation domain generalization results. Comparing SceneDINO to existing 2D unsupervised semantic segmentation methods and S4C + STEGO 3D baseline, using Accuracy and mean IoU (in %, \\(\\dagger\\)). We train on KITTI-360 images and report domain generalization results on Cityscapes and BDD-100K val. \\(\\dagger\\) denotes plain FiT3D features." + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.384, + 0.907, + 0.505 + ], + "angle": 0, + "content": "
MethodFeaturesCityscapesBDD-100K
AccmIoUAccmIoU
U2Seg [78]-75.5718.6269.0017.99
STEGO [33]DINO [11]71.2119.4275.0221.41
STEGO [33]DINOv2 [80]68.4119.7365.7221.77
STEGO [33]FiT3D [116]66.9421.0165.9620.99
STEGO [33]FiT3D† [116]64.7617.1760.8319.09
S4C [38] + STEGO [33]DINO [11]54.8014.0444.9811.62
SceneDINO (Ours)DINO [11]73.1722.8172.2822.09
" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.523, + 0.907, + 0.629 + ], + "angle": 0, + "content": "vised semantic segmentation approach. In particular, the mIoU of SceneDINO is \\(2.24\\%\\) points higher than for STEGO (w/ DINO). Utilizing 3D refined features from FiT3D deteriorates the baseline relative to DINO, indicating that the FiT3D refinement reduces feature expressiveness. Notably, our unsupervised 3D baseline S4C + STEGO transfers significantly worse to 2D than SceneDINO." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.909, + 0.75 + ], + "angle": 0, + "content": "We also validate SceneDINO, trained on KITTI-360, on Cityscapes and BDD10K, demonstrating domain generalization. The results are reported in Tab. 3. SceneDINO outperforms all baselines in mIoU on both datasets while only falling short in Acc. S4C + STEGO falls short in generalization. We suspect this poor generalization is caused by the fact that S4C does not rely on general SSL features in the final model, while our feature field generalizes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.785, + 0.774 + ], + "angle": 0, + "content": "4.3. Multi-view feature consistency" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.909, + 0.901 + ], + "angle": 0, + "content": "We analyze the multi-view consistency of our feature field against existing 2D SSL features in Tab. 4. We report the results of SceneDINO trained on KITTI-360 and RealEstate10K. SceneDINO trained using DINO features exhibits significant improvements in multi-view feature consistency over standard DINO features. We also train SceneDINO using target features from DINOv2 [80]. Compared to standard DINOv2 and FiT3D fea" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6790" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.161 + ], + "angle": 0, + "content": "Table 4. Multi-view consistency results. Comparing multi-view consistency of SceneDINO to existing 2D SSL features, using \\( \\mathrm{L}_1 \\) distance \\( (\\downarrow) \\), \\( \\mathrm{L}_2 \\) distance \\( (\\downarrow) \\), and cosine similarity \\( (\\uparrow) \\) on KITTI-360 and RealEstate10K. We compare DINO (top) and DINOv2-based (bottom) features. \\( \\dagger \\) denotes plain FiT3D features." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.165, + 0.485, + 0.278 + ], + "angle": 0, + "content": "
MethodKITTI-360RealEstate10K
L1L2cos-simL1L2cos-sim
DINO [11]16.060.740.7014.410.660.75
SceneDINO (w/ DINO)6.450.330.935.870.280.95
DINOv2 [80]15.830.730.7014.200.660.75
FiT3D [116]22.860.810.8219.880.720.85
FiT3D† [116]7.020.330.935.670.270.95
SceneDINO (w/ DINOv2)5.240.240.964.870.220.97
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.287, + 0.485, + 0.387 + ], + "angle": 0, + "content": "Table 5. SceneDINO analysis. We analyze the role of decomposing positional encodings, the choice of downsampling features during training, the effectiveness of the feature smoothness loss, the effect of estimated camera poses, and the choice of target features. We report the mean IoU (in %, ↑) using a range of \\(51.2\\mathrm{m}\\) on SSCBench-KITTI-360 test. \\(\\Delta\\) mIoU reports the absolute difference in % points to our standard model with DINO target features." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.391, + 0.485, + 0.482 + ], + "angle": 0, + "content": "
Δ mIoUmIoUConfiguration
-1.186.82No downsampler (bilinear up. + aug.)
-1.176.83No feature smoothness loss (λfs = 0)
-0.747.26No pos. enc. decomposition
-0.127.88w/ estimated ORB-SLAM3 poses
—8.00Full framework (SceneDINO)
+1.089.08DINOv2 target features (vs. DINO)
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.502, + 0.484, + 0.581 + ], + "angle": 0, + "content": "tures, SceneDINO's feature field yields significantly better multi-view consistency. Notably, compared against plain 3D refined features of FiT3D, SceneDINO shows a better multi-view consistency on both datasets and all metrics while also offering more expressiveness (cf. Tab. 2)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.59, + 0.303, + 0.607 + ], + "angle": 0, + "content": "4.4. Analyzing SceneDINO" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.613, + 0.483, + 0.839 + ], + "angle": 0, + "content": "To understand what core components contribute to obtaining an expressive feature field of SceneDINO, we omit or replace individual components and report the results in Tab. 5. Replacing the downsampling approach with bilinear upsampling and multi-crop augmentations, similar to [1], to obtain high-resolution target features decreases the SSC mIoU by \\(1.18\\%\\). Omitting the feature smoothness loss leads to a similar mIoU drop. Abolishing the constant decomposition of positional encodings leads to an mIoU drop of \\(0.74\\%\\). Training using unsupervised camera poses estimated by ORB-SLAM3 [7] results in an insignificant mIoU drop of only \\(0.12\\%\\) over using KITTI-360 poses. Going from DINO to DINOv2 target features leads to an increased mIoU of \\(1.08\\%\\), demonstrating that SceneDINO can benefit from more expressive 2D target features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.903 + ], + "angle": 0, + "content": "In Tab. 6, we analyze our 3D distillation. Performing no distillation at all, just feature clustering, decreases mIoU by \\(1.61\\%\\). Omitting the kNN-correlation loss leads to an mIoU drop of \\(1.35\\%\\). Distilling only with center points," + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.906, + 0.16 + ], + "angle": 0, + "content": "Table 6. Feature distillation analysis. We analyze the effectiveness of distilling SceneDINO's features, the \\( k \\)-NN-correlation loss, our neighborhood sampling, and our 3D sampling approach over standard 5-crop sampling. We report the mean IoU (in %, ↑) using a range of \\( 51.2 \\mathrm{~m} \\) on SSCBench-KITTI-360 test." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.165, + 0.907, + 0.24 + ], + "angle": 0, + "content": "
Δ mIoUmIoUConfiguration
-1.616.39No distillation
-1.356.65No kNN-correlation loss (λkNN = 0)
-0.977.03No neighborhood sampling (cf. Fig. 4)
-0.477.535-crop sampling [33] (instead 3D sampling)
—8.00Full framework (SceneDINO)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.249, + 0.907, + 0.307 + ], + "angle": 0, + "content": "Table 7. Probing analysis. We analyze linear and unsupervised probing of our distilled SceneDINO features on SSCBenchKITTI-360 test using mean IoU (in %, ↑). For reference, we also report S4C (2D supervised). Linear probing uses 2D annotations." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.311, + 0.906, + 0.415 + ], + "angle": 0, + "content": "
Probing approachTarget featuresmIoU
12.8 m25.6 m51.2 m
UnsupervisedDINO [11]10.7610.018.00
DINOv2 [80]13.7611.789.08
LinearDINO [11]13.6312.079.34
DINOv2 [80]15.8513.7010.57
S4C (full training)n/a16.9413.9410.19
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.436, + 0.906, + 0.512 + ], + "angle": 0, + "content": "i.e., not performing neighborhood sampling (cf. Fig. 4), reduces mIoU by \\(0.97\\%\\). Using 5-crop feature sampling [33], instead of our proposed 3D sampling, leads to a reduced mIoU of \\(0.47\\%\\). This demonstrates the effectiveness of performing distillation in 3D using our novel approach." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.512, + 0.909, + 0.695 + ], + "angle": 0, + "content": "While focusing on unsupervised SSC, we can also linearly probe our distilled feature field (cf. Tab. 7). In particular, we train SceneDINO using different target features (DINO [11] and DINOv2 [11]), perform distillation, and probe the resulting distilled features. Using linear probing, i.e., training a single linear layer using 2D semantic labels, leads to a consistent mIoU increase over unsupervised probing. SceneDINO trained using DINOv2 target features even closes the gap to S4C, trained using 2D ground-truth semantic labels. We even surpass 2D supervised S4C slightly on the full range \\((51.2\\mathrm{m})\\), suggesting the effectiveness of SceneDINO also for weakly-supervised tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.709, + 0.634, + 0.725 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We presented SceneDINO, to our knowledge, the first approach for unsupervised semantic scene completion. Trained using multi-view images and 2D DINO features without human supervision, SceneDINO is able to predict an expressive 3D feature field using a single input image during inference. Our novel 3D distillation approach yields state-of-the-art results in unsupervised SSC. While we focus on unsupervised SSC, our multi-view feature consistency, linear probing, and domain generalization results highlight the potential of SceneDINO as a strong foundation for various 3D scene-understanding tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.958 + ], + "angle": 0, + "content": "6791" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.396 + ], + "angle": 0, + "content": "Acknowledgments. This project was partially supported by the European Research Council (ERC) Advanced Grant SIMULACRON, DFG project CR 250/26-1 \"4D-YouTube\", and GNI Project \"AICC\". This project was also partially supported by the ERC under the European Union's Horizon 2020 research and innovation programme (grant agreement No. 866008). Additionally, this work has been co-funded by the LOEWE initiative (Hesse, Germany) within the emergenCITY center [LOEWE/1/12/519/03/05.001(0016)/72] and by the Deutsche Forschungsgemeinschaft (German Research Foundation, DFG) under Germany's Excellence Strategy (EXC 3066/1 \"The Adaptive Mind\", Project No. 533717223). Christoph Reich is supported by the Konrad Zuse School of Excellence in Learning and Intelligent Systems (ELIZA) through the DAAD programme Konrad Zuse Schools of Excellence in Artificial Intelligence, sponsored by the Federal Ministry of Education and Research. Christian Rupprecht is supported by an Amazon Research Award. Finally, we acknowledge the support of the European Laboratory for Learning and Intelligent Systems (ELLIS) and thank Mateo de Mayo as well as Igor Cvišić for help with estimating camera poses." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.411, + 0.188, + 0.427 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.437, + 0.482, + 0.478 + ], + "angle": 0, + "content": "[1] Nikita Araslanov and Stefan Roth. Self-supervised augmentation consistency for adapting semantic segmentation. In CVPR, pages 15384-15394, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.481, + 0.483, + 0.522 + ], + "angle": 0, + "content": "[2] Yuki Markus Asano, Christian Rupprecht, and Andrea Vedaldi. Self-labelling via simultaneous clustering and representation learning. In ICLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.526, + 0.482, + 0.579 + ], + "angle": 0, + "content": "[3] Philip Bachman, R. Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. In NeurIPS*2019, pages 15509-15519. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.583, + 0.482, + 0.625 + ], + "angle": 0, + "content": "[4] Adrien Bardes, Jean Ponce, and Yann LeCun. VI-CRegL: Self-supervised learning of local visual features. In NeurIPS*2022, pages 8799-8810. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.628, + 0.483, + 0.669 + ], + "angle": 0, + "content": "[5] Adrien Bardes, Jean Ponce, and Yann LeCun. VI-CReg: Variance-invariance-covariance regularization for self-supervised learning. In *ICLR*, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.672, + 0.482, + 0.727 + ], + "angle": 0, + "content": "[6] Yingjie Cai, Xuesong Chen, Chao Zhang, Kwan-Yee Lin, Xiaogang Wang, and Hongsheng Li. Semantic scene completion via integrating instances and scene in-the-loop. In CVPR, pages 324-333, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.73, + 0.482, + 0.798 + ], + "angle": 0, + "content": "[7] Carlos Campos, Richard Elvira, Juan J. Gómez Rodríguez, José M. M. Montiel, and Juan D Tardós. ORB-SLAM3: An accurate open-source library for visual, visual-inertial and multi-map SLAM. IEEE Trans. Robot., 37(6):1874-1890, 2021. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.802, + 0.482, + 0.842 + ], + "angle": 0, + "content": "[8] Anh-Quan Cao and Raoul de Charette. Monoscene: Monocular 3D semantic scene completion. In CVPR, pages 3981-3991, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[9] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS*2020, pages 9912-9924. 2" + }, + { + "type": "list", + "bbox": [ + 0.108, + 0.437, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[10] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In ECCV, pages 132-149, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[11] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. 1, 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.192, + 0.904, + 0.232 + ], + "angle": 0, + "content": "[12] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv:2003.04297 [cs.CV], 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.234, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[13] Xiaokang Chen, Kwan-Yee Lin, Chen Qian, Gang Zeng, and Hongsheng Li. 3d sketch-aware semantic scene completion via semi-supervised structure prior. In CVPR, pages 4192-4201, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.289, + 0.905, + 0.329 + ], + "angle": 0, + "content": "[14] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In CVPR, pages 9640-9649, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.331, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[15] Yuedong Chen, Chuanxia Zheng, Haofei Xu, Bohan Zhuang, Andrea Vedaldi, Tat-Jen Cham, and Jianfei Cai. MVSplat360: Feed-forward 360 scene synthesis from sparse views. In NeurIPS*2024, pages 107064-107086. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.387, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[16] Ran Cheng, Christopher Agia, Yuan Ren, Xinhai Li, and Bingbing Liu. S3CNet: A sparse semantic scene completion network for LiDAR point clouds. In CoRL, pages 2148-2161, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.443, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[17] Jang Hyun Cho, Utkarsh Mall, Kavita Bala, and Bharath Hariharan. PiCIE: Unsupervised semantic segmentation using invariance and equivariance in clustering. In CVPR, pages 16794-16804, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.498, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[18] Özgün Çiçek, Ahmed Abdulkadir, Soeren S Lienkamp, Thomas Brox, and Olaf Ronneberger. 3D U-Net: Learning dense volumetric segmentation from sparse annotation. In MICCAI, pages 424–432, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.554, + 0.905, + 0.622 + ], + "angle": 0, + "content": "[19] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The Cityscapes dataset for semantic urban scene understanding. In CVPR, pages 3213-3223, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.624, + 0.905, + 0.677 + ], + "angle": 0, + "content": "[20] Timothee Darcet, Federico Baldassarre, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Cluster and predict latent patches for improved masked image modeling. arXiv:2502.08769 [cs.CV], 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.679, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[21] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.721, + 0.905, + 0.762 + ], + "angle": 0, + "content": "[22] Carl Doersch, Abhinav Gupta, and Alexei A. Efros. Unsupervised visual representation learning by context prediction. In ICCV, pages 1422-1430, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.763, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[23] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth \\(16 \\times 16\\) words: Transformers for image recognition at scale. In ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] Linus Ericsson, Henry Gouk, Chen Change Loy, and Timothy M. Hospedales. Self-supervised representation learning: Introduction, advances, and challenges. IEEE Trans. Signal Process., 39(3):42-62, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "6792" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[25] Stephanie Fu, Mark Hamilton, Laura E. Brandt, Axel Feldmann, Zhoutong Zhang, and William T. Freeman. FeatUp: A model-agnostic framework for features at any resolution. In ICLR, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.148, + 0.482, + 0.189 + ], + "angle": 0, + "content": "[26] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The KITTI dataset. Int. J. Robot. Res., 32(11):1231-1237, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.19, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[27] Clément Godard, Oisin Mac Aodha, and Gabriel J. Brostow. Unsupervised monocular depth estimation with left-right consistency. In CVPR, pages 270-279, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.233, + 0.482, + 0.301 + ], + "angle": 0, + "content": "[28] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, et al. Bootstrap your own latent: A new approach to self-supervised learning. In NeurIPS*2020, pages 21271-21284. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.302, + 0.482, + 0.342 + ], + "angle": 0, + "content": "[29] Agrim Gupta, Jiajun Wu, Jia Deng, and Li Fei-Fei. Siamese Masked Autoencoders. In NeurIPS*2023, pages 40676-40693. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.344, + 0.482, + 0.385 + ], + "angle": 0, + "content": "[30] Huy Ha and Shuran Song. Semantic abstraction: Openworld 3D scene understanding from 2D vision-language models. In CoRL, pages 643-653, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.386, + 0.482, + 0.44 + ], + "angle": 0, + "content": "[31] Oliver Hahn, Nikita Araslanov, Simone Schaub-Meyer, and Stefan Roth. Boosting unsupervised semantic segmentation with principal mask proposals. Trans. Mach. Learn. Res., 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.442, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[32] Oliver Hahn, Christoph Reich, Nikita Araslanov, Daniel Cremers, Christian Rupprecht, and Stefan Roth. Scene-centric unsupervised panoptic segmentation. In CVPR, pages 24485-24495, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.498, + 0.482, + 0.551 + ], + "angle": 0, + "content": "[33] Mark Hamilton, Zhoutong Zhang, Bharath Hariharan, Noah Snavely, and William T. Freeman. Unsupervised semantic segmentation by distilling feature correspondences. In ICLR, 2022. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.553, + 0.482, + 0.608 + ], + "angle": 0, + "content": "[34] Keonhee Han, Dominik Muhle, Felix Wimbauer, and Daniel Cremers. Boosting self-supervision for single-view scene completion via knowledge distillation. In CVPR, pages 9837-9847, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.609, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[35] Xian-Feng Han, Hamid Laga, and Mohammed Bennamoun. Image-based 3D object reconstruction: State-of-the-art and trends in the deep learning era. IEEE Trans. Pattern Anal. Mach. Intell., 43(5):1578-1604, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.665, + 0.482, + 0.706 + ], + "angle": 0, + "content": "[36] Robert Harb and Patrick Knobelreiter. InfoSeg: Unsupervised semantic image segmentation with mutual information maximization. In \\(GCPR\\), pages 18-32, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.707, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[37] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge University Press, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.748, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[38] Adrian Hayler, Felix Wimbauer, Dominik Muhle, Christian Rupprecht, and Daniel Cremers. S4C: Self-supervised semantic scene completion with neural fields. In 3DV, pages 409-420, 2024. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.804, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[39] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9729-9738, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.846, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[40] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "[41] Olivier Henaff. Data-efficient image pecognition with contrastive predictive coding. In ICML, pages 4182-4192, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.136, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[42] R. Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. In ICLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.207, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[43] Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3D v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. IEEE Trans. Pattern Anal. Mach. Intell., 46(12):10579-10596, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.291, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[44] Rui Huang, Songyou Peng, Ayca Takmaz, Federico Tombari, Marc Pollefeys, Shiji Song, Gao Huang, and Francis Engelmann. Segment3D: Learning fine-grained class-agnostic 3D segmentation without manual labels. In ECCV, pages 278–295, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.362, + 0.905, + 0.416 + ], + "angle": 0, + "content": "[45] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3D semantic occupancy prediction. In CVPR, pages 9223-9232, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.419, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[46] Yuanhui Huang, Wenzhao Zheng, Borui Zhang, Jie Zhou, and Jiwen Lu. SelfOcc: Self-supervised vision-based 3D occupancy prediction. In CVPR, pages 19946-19956, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.476, + 0.904, + 0.531 + ], + "angle": 0, + "content": "[47] Joel Janai, Fatma Güney, Aseem Behl, and Andreas Geiger. Computer vision for autonomous vehicles: Problems, datasets and state of the art. Found. Trends Comput. Graph. Vis., 12(1-3):1-308, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.532, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[48] Xu Ji, Joao F. Henriques, and Andrea Vedaldi. Invariant information clustering for unsupervised image classification and segmentation. In ICCV, pages 9865-9874, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.576, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[49] Haoyi Jiang, Liu Liu, Tianheng Cheng, Xinjie Wang, Tianwei Lin, Zhizhong Su, Wenyu Liu, and Xinggang Wang. GaussTR: Foundation model-aligned Gaussian transformer for self-supervised 3D spatial understanding. arXiv:2412.13193 [cs.CV], 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.646, + 0.905, + 0.699 + ], + "angle": 0, + "content": "[50] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. LERF: Language embedded radiance fields. In ICCV, pages 19729-19739, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.703, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[51] Chanyoung Kim, Woojung Han, Dayun Ju, and Seong Jae Hwang. EAGLE: Eigen aggregation learning for object-centric unsupervised semantic segmentation. In CVPR, pages 3523-3533, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.76, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[52] Diederik P. Kingma and Jimmy Lei Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.789, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[53] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment Anything. In ICCV, pages 4015-4026, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[54] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing NeRF for editing via feature field distillation. In NeurIPS*2022, pages 23311-23330. 2" + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "6793" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[55] Alexander Koenig, Maximilian Schambach, and Johannes Otterbach. Uncovering the inner workings of STEGO for safe unsupervised semantic segmentation. In CVPRW, pages 3789-3798, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.151, + 0.482, + 0.192 + ], + "angle": 0, + "content": "[56] Philipp Krahenbuhl and Vladlen Koltun. Efficient inference in fully connected CRFs with Gaussian edge potentials. In NIPS*2011, pages 109-117. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.194, + 0.482, + 0.223 + ], + "angle": 0, + "content": "[57] Harold W. Kuhn. The Hungarian method for the assignment problem. Nav. Res. Logist. Q., 2:83-97, 1955. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.225, + 0.482, + 0.28 + ], + "angle": 0, + "content": "[58] Jie Li, Yu Liu, Dong Gong, Qinfeng Shi, Xia Yuan, Chunxia Zhao, and Ian D. Reid. RGBD based dimensional decomposition residual network for 3D semantic scene completion. In CVPR, pages 7693-7702, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.282, + 0.482, + 0.324 + ], + "angle": 0, + "content": "[59] Jie Li, Kai Han, Peng Wang, Yu Liu, and Xia Yuan. Anisotropic convolutional networks for 3D semantic scene completion. In CVPR, pages 3348-3356, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.327, + 0.482, + 0.38 + ], + "angle": 0, + "content": "[60] Jie Li, Yu Liu, Xia Yuan, Chunxia Zhao, Roland Siegwart, Ian Reid, and Cesar Cadena. Depth based semantic scene completion with position importance aware loss. IEEE Robotics Autom. Lett., 5(1):219-226, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.384, + 0.482, + 0.424 + ], + "angle": 0, + "content": "[61] Junnan Li, Pan Zhou, Caiming Xiong, and Steven Hoi. Prototypical contrastive learning of unsupervised representations. In ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.428, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[62] Pengfei Li, Yongliang Shi, Tianyu Liu, Hao Zhao, Guyue Zhou, and Ya-Qin Zhang. Semi-supervised implicit scene completion from sparse LiDAR. arXiv:2111.14798 [cs.CV], 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.486, + 0.482, + 0.554 + ], + "angle": 0, + "content": "[63] Yiming Li, Zhiding Yu, Christopher B. Choy, Chaowei Xiao, José M. Álvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. VoxFormer: Sparse voxel transformer for camera-based 3D semantic scene completion. In CVPR, pages 9087-9098, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.557, + 0.482, + 0.639 + ], + "angle": 0, + "content": "[64] Yiming Li, Sihang Li, Xinhao Liu, Moonjun Gong, Kenan Li, Nuo Chen, Zijun Wang, Zhiheng Li, Tao Jiang, Fisher Yu, Yue Wang, Hang Zhao, Zhiding Yu, and Chen Feng. SSCBench: A large-scale 3D semantic scene completion benchmark for autonomous driving. In IROS, pages 13333-13340, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.643, + 0.482, + 0.697 + ], + "angle": 0, + "content": "[65] Zhiqi Li, Zhiding Yu, David Austin, Mingsheng Fang, Shiyi Lan, Jan Kautz, and José M. Álvarez. FB-OCC: 3D occupancy prediction based on forward-backward view transformation. arXiv:2307.01492 [cs.CV], 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.7, + 0.482, + 0.755 + ], + "angle": 0, + "content": "[66] Yiyi Liao, Jun Xie, and Andreas Geiger. KITTI-360: A novel dataset and benchmarks for urban scene understanding in 2D and 3D. IEEE Trans. Pattern Anal. Mach. Intell., 45(3):3292-3310, 2023. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.758, + 0.482, + 0.813 + ], + "angle": 0, + "content": "[67] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, pages 740-755, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.816, + 0.482, + 0.869 + ], + "angle": 0, + "content": "[68] Shice Liu, Yu Hu, Yiming Zeng, Qiankun Tang, Beibei Jin, Yinhe Han, and Xiaowei Li. See and think: Disentangling semantic scene completion. In NeurIPS*2018, pages 261-272. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[69] Stuart Lloyd. Least squares quantization in PCM. IEEE Trans. Inf. Theory, 28(2):129-137, 1982. 4, 5" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[70] Zhiliang Ma and Shilong Liu. A review of 3D reconstruction techniques in civil engineering and their applications. Adv. Eng. Inform., 37:163-174, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.137, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[71] James MacQueen. Some methods for classification and analysis of multivariate observations. In Berkeley Symp. on Math. Statist. and Prob., pages 281-298, 1967. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.181, + 0.904, + 0.208 + ], + "angle": 0, + "content": "[72] Nelson Max. Optical models for direct volume rendering. IEEE Trans. Vis. Comput. Graph., 1(2):99-108, 1995. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.211, + 0.905, + 0.252 + ], + "angle": 0, + "content": "[73] Kirill Mazur, Edgar Sucar, and Andrew J Davison. Feature-realistic neural fusion for real-time, open set scene understanding. In ICRA, pages 8201-8207, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.255, + 0.905, + 0.31 + ], + "angle": 0, + "content": "[74] Ruihang Miao, Weizhou Liu, Mingrui Chen, Zheng Gong, Weixin Xu, Chen Hu, and Shuchang Zhou. Occdepth: A depth-aware method for 3D semantic scene completion. arXiv.2302.13540 [cs.CV], 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.312, + 0.905, + 0.38 + ], + "angle": 0, + "content": "[75] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. Commun. ACM, 65(1):99-106, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.384, + 0.905, + 0.425 + ], + "angle": 0, + "content": "[76] Yue Ming, Xuyang Meng, Chunxiao Fan, and Hui Yu. Deep learning for monocular depth estimation: A review. Neurocomputing, 438:14-33, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.428, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[77] Duy Kien Nguyen, Yanghao Li, Vaibhav Aggarwal, Martin R. Oswald, Alexander Kirillov, Cees G. M. Snoek, and Xinlei Chen. R-MAE: Regions meet masked autoencoders. In ICLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.486, + 0.905, + 0.54 + ], + "angle": 0, + "content": "[78] Dantong Niu, Xudong Wang, Xinyang Han, Long Lian, Roei Herzig, and Trevor Darrell. Unsupervised universal image segmentation. In CVPR, pages 22744-22754, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.543, + 0.905, + 0.584 + ], + "angle": 0, + "content": "[79] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In ECCV, pages 69-84, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.587, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[80] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. DINOv2: Learning robust visual features without supervision. Trans. Mach. Learn. Res., 2024. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.659, + 0.905, + 0.727 + ], + "angle": 0, + "content": "[81] Martin R Oswald, Eno Töppe, Claudia Nieuwenhuis, and Daniel Cremers. A review of geometry recovery from a single image focusing on curved object reconstruction. Innovations for Shape Analysis: Models and Algorithms, pages 343-378, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.731, + 0.905, + 0.798 + ], + "angle": 0, + "content": "[82] Mingjie Pan, Jiaming Liu, Renrui Zhang, Peixiang Huang, Xiaqi Li, Hongwei Xie, Bing Wang, Li Liu, and Shang-hang Zhang. RenderOcc: Vision-centric 3D occupancy prediction with 2D rendering supervision. In ICRA, pages 12404-12411, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.802, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[83] Songyou Peng, Kyle Genova, Chiyu “Max” Jiang, Andrea Tagliasacchi, Marc Pollefeys, and Thomas Funkhouser. OpenScene: 3D scene understanding with open vocabularies. In CVPR, pages 815–824, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.86, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[84] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In ICCV, pages 12179-12188, 2021. 6" + }, + { + "type": "list", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6794" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[85] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, et al. SAM 2: Segment anything in images and videos. arXiv:2408.00714 [cs.CV], 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.164, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[86] Stephan R. Richter and Stefan Roth. Matryoshka networks: Predicting 3D geometry via nested shape layers. In CVPR, pages 1936-1944, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.207, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[87] Christoph B. Rist, David Emmerichs, Markus Enzweiler, and Dariu M. Gavrila. Semantic scene completion using local deep implicit functions on LiDAR data. IEEE Trans. Pattern Anal. Mach. Intell., 44(10):7205-7218, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.263, + 0.482, + 0.304 + ], + "angle": 0, + "content": "[88] Luis Roldão, Raoul de Charette, and Anne Verroust-Blondet. LMSCNet: Lightweight multiscale 3D semantic completion. In 3DV, pages 111–119, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.307, + 0.482, + 0.347 + ], + "angle": 0, + "content": "[89] Luis Roldao, Raoul De Charette, and Anne Verroust-Blondet. 3D semantic scene completion: A survey. Int. J. Comput. Vis., 130(8):1978-2005, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.349, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[90] Johannes L. Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, pages 4104-4113, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.392, + 0.482, + 0.419 + ], + "angle": 0, + "content": "[91] David Sculley. Web-scale \\( k \\)-means clustering. In WWW, page 1177-1178, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.421, + 0.482, + 0.474 + ], + "angle": 0, + "content": "[92] Hyun Seok Seong, WonJun Moon, SuBeen Lee, and Jae-Pil Heo. Leveraging hidden positives for unsupervised semantic segmentation. In CVPR, pages 19540-19549, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.477, + 0.482, + 0.533 + ], + "angle": 0, + "content": "[93] Nur Muhammad Mahi Shafiullah, Chris Paxton, Lerrel Pinto, Soumith Chintala, and Arthur Szlam. CLIP-Fields: Weakly supervised semantic fields for robotic memory. In ICRA Workshop on Pretraining for Robotics, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.535, + 0.482, + 0.588 + ], + "angle": 0, + "content": "[94] William Shen, Ge Yang, Alan Yu, Jansen Wong, Leslie Pack Kaelbling, and Phillip Isola. Distilled feature fields enable few-shot language-guided manipulation. In CoRL, pages 405-424, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.591, + 0.482, + 0.645 + ], + "angle": 0, + "content": "[95] Leon Sick, Dominik Engel, Pedro Hermosilla, and Timo Ropinski. Unsupervised semantic segmentation through depth-guided feature correlation and sampling. In CVPR, pages 3637-3646, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.648, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[96] Shuran Song, Fisher Yu, Andy Zeng, Angel X. Chang, Manolis Savva, and Thomas A. Funkhouser. Semantic scene completion from a single depth image. In CVPR, pages 190-198, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.704, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[97] Stanislaw Szymanowicz, Eldar Insafutdinov, Chuanxia Zheng, Dylan Campbell, Joao F. Henriques, Christian Rupprecht, and Andrea Vedaldi. Flash3D: Feed-forward generalisable 3D scene reconstruction from a single image. arXiv:2406.04343 [cs.CV], 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.775, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[98] Ayca Takmaz, Elisabetta Fedele, Robert W. Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. OpenMask3D: Open-vocabulary 3D instance segmentation. In NeurIPS*2023, pages 68367-68390. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.831, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[99] Zachary Teed and Jia Deng. RAFT: Recurrent all-pairs field transforms for optical flow. In ECCV, pages 402-419, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[100] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Lin, and Hongyang Li. Scene as occupancy. In ICCV, pages 8372-8381, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[101] Nikolaos Tsagkas, Oisin Mac Aodha, and Chris Xiaoxuan Lu. VL-Fields: Towards language-grounded neural implicit spatial representations. In ICRA Workshop on Representations, Abstractions, and Priors for Robot Learning, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.193, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[102] Vadim Tschernezki, Iro Laina, Diane Larlus, and Andrea Vedaldi. Neural feature fusion fields: 3D distillation of self-supervised 2D image representations. In 3DV, pages 443-453, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.25, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[103] Shubham Tulsiani, Tinghui Zhou, Alexei A. Efros, and Jitendra Malik. Multi-view supervision for single-view reconstruction via differentiable ray consistency. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.307, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[104] Xudong Wang, Rohit Girdhar, Stella X. Yu, and Ishan Misra. Cut and learn for unsupervised object detection and instance segmentation. In CVPR, pages 3124-3134, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.363, + 0.905, + 0.418 + ], + "angle": 0, + "content": "[105] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: From error visibility to structural similarity. IEEE Trans. Image Process., 13(4): 600-612, 2004. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.421, + 0.905, + 0.474 + ], + "angle": 0, + "content": "[106] Silvan Weder, Hermann Blum, Francis Engelmann, and Marc Pollefeys. LabelMaker: Automatic semantic label generation from RGB-D trajectories. In 3DV, pages 334-343, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.477, + 0.905, + 0.533 + ], + "angle": 0, + "content": "[107] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In CVPR, pages 14668-14678, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.535, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[108] Felix Wimbauer, Nan Yang, Christian Rupprecht, and Daniel Cremers. Behind the scenes: Density fields for single view reconstruction. In CVPR, pages 9076-9086, 2023. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.591, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[109] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Comput. Graph. Forum, pages 641-676, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.661, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[110] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep LiDAR point cloud segmentation via learning contextual shape priors from scene completion. In AAAI, pages 3101-3109, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.733, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[111] Jiawei Yang, Boris Ivanovic, Or Litany, Xinshuo Weng, Seung Wook Kim, Boyi Li, Tong Che, Danfei Xu, Sanja Fidler, Marco Pavone, and Yue Wang. EmerneRF: Emergent spatial-temporal scene decomposition via self-supervision. In ICLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[112] Jiawei Yang, Katie Z Luo, Jiefeng Li, Congyue Deng, Leonidas Guibas, Dilip Krishnan, Kilian Q Weinberger, Yonglong Tian, and Yue Wang. Denoising vision transformers. In ECCV, pages 453-469, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[113] Zhuoyue Yang, Ju Dai, and Junjun Pan. 3D reconstruction from endoscopy images: A survey. Comput. Biol. Med., 175:108546, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "6795" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.482, + 0.135 + ], + "angle": 0, + "content": "[114] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelNeRF: Neural radiance fields from one or few images. In CVPR, pages 4578-4587, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.136, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[115] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. BDD100K: A diverse driving dataset for heterogeneous multitask learning. In CVPR, pages 2633-2642, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.207, + 0.483, + 0.261 + ], + "angle": 0, + "content": "[116] Yuanwen Yue, Anurag Das, Francis Engelmann, Siyu Tang, and Jan Eric Lenssen. Improving 2D feature representations by 3D-aware fine-tuning. In ECCV, pages 57-74, 2024. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.263, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[117] Pingping Zhang, Wei Liu, Yinjie Lei, Huchuan Lu, and Xiaoyun Yang. Cascaded context pyramid for full-resolution 3D semantic scene completion. In ICCV, pages 7800-7809, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.319, + 0.482, + 0.361 + ], + "angle": 0, + "content": "[118] Yunpeng Zhang, Zheng Zhu, and Dalong Du. OccFormer: Dual-path transformer for vision-based 3D semantic occupancy prediction. In ICCV, pages 9433-9443, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.362, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[119] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snively. Stereo magnification: Learning view synthesis using multiplane images. ACM Trans. Graph., 37 (4):65, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.418, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[120] Onur Özyeşil, Vladislav Voroninski, Ronen Basri, and Amit Singer. A survey of structure from motion. Acta Numer., 26:305-364, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.483, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6796" + } + ] +] \ No newline at end of file diff --git a/2025/Feed-Forward SceneDINO for Unsupervised Semantic Scene Completion/0cb8dcd6-d20d-4d18-a810-629c106f3cdd_origin.pdf b/2025/Feed-Forward SceneDINO for Unsupervised Semantic Scene Completion/0cb8dcd6-d20d-4d18-a810-629c106f3cdd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..82e57866edf14c487f3df24e12fce83ef96d1a7e --- /dev/null +++ b/2025/Feed-Forward SceneDINO for Unsupervised Semantic Scene Completion/0cb8dcd6-d20d-4d18-a810-629c106f3cdd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3cd1302a13c6c241bc6b2ef377640a6c0e373f8aaf91160a05998af66058bc54 +size 18096293 diff --git a/2025/Feed-Forward SceneDINO for Unsupervised Semantic Scene Completion/full.md b/2025/Feed-Forward SceneDINO for Unsupervised Semantic Scene Completion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2de14378ef0f07b9b60dc2d357a6459afb0ee81b --- /dev/null +++ b/2025/Feed-Forward SceneDINO for Unsupervised Semantic Scene Completion/full.md @@ -0,0 +1,419 @@ +# Feed-Forward SceneDINO for Unsupervised Semantic Scene Completion + +Aleksandar Jevtic*1 + +Oliver Hahn + +Christian Rupprecht + +$^{1}$ TU Munich + +$^{2}$ TU Darmstadt + +3University of Oxford + +Christoph Reich* 1,2,4,5 + +Eich\*1,2,4,5 F6 Stefan Roth 2,5,6 + +Felix Wimbauer1,4 + +4MCML 5ELIZA + +Daniel Cremers $^{1,4,5}$ + +$^{6}$ hessian.AI *equal contribution + +https://visinf.github.io/scenedino + +![](images/262c286345295238ebd432bdf2ecccf1e0df6b5da244a4e6a3c87aefed1f9aa2.jpg) +Single Input Image + +SceneDINO + +![](images/bcd377efa08022c7bda5c5b9514b60b4f3f97d03895267ac41ed2035f74e99a0.jpg) +3D Feature Field +Figure 1. SceneDINO overview. Given a single input image (left), SceneDINO predicts both 3D scene geometry and 3D features in the form of a feature field (middle) in a feed-forward manner, capturing the structure and semantics of the scene. Unsupervised distillation and clustering of SceneDINO's feature space leads to unsupervised semantic scene completion predictions (right). + +Distill & Cluster + +![](images/bf9fd12e6841fe47ba904bc33b9a2eb12f5ac84965bf06620b42ef234550e8e4.jpg) +SSC Prediction + +# Abstract + +Semantic scene completion (SSC) aims to infer both the 3D geometry and semantics of a scene from single images. In contrast to prior work on SSC that heavily relies on expensive ground-truth annotations, we approach SSC in an unsupervised setting. Our novel method, SceneDINO, adapts techniques from self-supervised representation learning and 2D unsupervised scene understanding to SSC. Our training exclusively utilizes multi-view consistency self-supervision without any form of semantic or geometric ground truth. Given a single input image, SceneDINO infers the 3D geometry and expressive 3D DINO features in a feed-forward manner. Through a novel 3D feature distillation approach, we obtain unsupervised 3D semantics. In both 3D and 2D unsupervised scene understanding, SceneDINO reaches state-of-the-art segmentation accuracy. Linear probing our 3D features matches the segmentation accuracy of a current supervised SSC approach. Additionally, we showcase the domain generalization and multi-view consistency of SceneDINO, taking the first steps towards a strong foundation for single image 3D scene understanding. + +# 1. Introduction + +Understanding the geometry and semantics of 3D scenes from image observations is a fundamental computer vision task with broad applications in robotics [26], autonomous driving [47, 66], medical image analysis [18, 113], and civil engineering [70]. The Semantic Scene Completion (SSC) + +task unifies 3D geometry and semantic prediction from limited image observations [64, 89, 96]. Recent progress in SSC has been primarily driven by utilizing supervised learning [38, 88, 96]. However, acquiring large-scale 3D annotations is highly labor-intensive [66]. While significant resources have been invested in collecting human annotations for 2D tasks [53, 85], annotating similar amounts of data in 3D remains unapproached. This motivates approaching SSC without the need for manually annotated data. + +Existing SSC approaches rely on ground-truth semantic annotations and frequently utilize additional supervision from LiDAR scans [38, 46, 74, 96]. In contrast, we are the first to approach SSC in a fully unsupervised setting, i.e. without task supervision or other supervised components. In particular, we aim to approach SSC from a single image without relying on any human annotations, only learning from unlabeled multi-view images using self-supervision. This setting is extremely challenging for two reasons: first, the human-defined nature of semantic taxonomies is ambiguous, and second, a single image only provides a partial observation of the scene with many invisible areas. We take inspiration from recent advances in self-supervised learning (SSL) of 2D representations and 3D reconstruction. 2D SSL representations, such as from DINO [11], have been shown effective for 2D unsupervised scene understanding [32, 33, 104]. 3D reconstruction approaches successfully leveraged SSL from multi-view data to infer dense 3D geometry from a single image [34, 108]. + +In this paper, we present SceneDINO, to the best of our knowledge, the first approach for unsupervised semantic scene completion. Trained using 2D SSL features + +from DINO [11] and multi-view self-supervision [108], SceneDINO predicts both 3D geometry and 3D features from a single image during inference in a feed-forward manner. Our general 3D feature representations enable us to approach unsupervised 3D scene understanding. Harnessing our expressive 3D features, we propose a novel 3D feature distillation approach for obtaining unsupervised semantic predictions in 3D. While we focus on the task of unsupervised SSC, SceneDINO's features are general, offering a foundation for different 3D scene-understanding tasks by building on our 3D feature field. + +Specifically, we make the following contributions: (i) We introduce SceneDINO, the first approach predicting dense 3D geometry and expressive 3D features in a feedforward manner from a single image. (ii) We effectively distill SceneDINO's feature field representation in 3D, obtaining unsupervised semantic predictions. (iii) We demonstrate the first fully unsupervised SSC results. We build a simple, yet competitive unsupervised SSC baseline, lifting unsupervised 2D semantic predictions. Our SceneDINO approach outperforms this SSC baseline in unsupervised SSC as well as established 2D approaches in 2D semantic segmentation. (iv) Finally, we also showcase the domain generalization ability and multi-view consistency of SceneDINO. + +# 2. Related Work + +Single-image scene reconstruction. Estimating 3D geometry from image observations is a fundamental task in computer vision and has been studied for decades [37]. Traditional approaches, such as structure from motion [90], as well as recent neural radiance fields (NeRFs) [75], perform scene reconstruction using multiple images, as reviewed by multiple surveys [35, 109, 120]. Recently, estimating dense 3D geometry from a single image have been approached [8, 34, 81, 86, 97, 103, 108, 114]. Unlike monocular depth estimation [76], these approaches predict the depth for visible and occluded regions, reconstructing a complete scene. Behind the Scenes (BTS) [108] introduced an approach for unsupervised single-image scene reconstruction using multi-view self-supervision, which infers dense 3D geometry in a feed-forward manner. Our approach extends BTS by additionally lifting self-supervised features into 3D for unsupervised 3D scene understanding. + +Semantic scene completion (SSC), also known as 3D semantic occupancy prediction, aims to jointly estimate the 3D geometry and semantics of a scene [63, 64, 96, 118]. Initial approaches used 3D semantic and geometric annotations and addressed indoor scenes [6, 13, 58-60, 68, 117], outdoor scenes with LiDAR [16, 62, 87, 88, 110], or both domains [8, 74]. Using birds-eye views has been proven effective for SSC [45, 65, 100]. To overcome the need for 3D annotations, approaches for using 2D annotations + +have been proposed [38, 46, 82]. While SelfOcc [46] and RenderOcc [82] use multiple inference views, S4C [38] performs single-image SSC. In particular, S4C [38] employs a supervised 2D model and lifts 2D multi-view semantic predictions into 3D. In contrast to using 2D annotations, GaussTR [49] uses 2D foundation models for SSC and multiple views during inference. However, GaussTR relies on heavily supervised foundation models, including SAM [53] and Metric3Dv2 [43], and uses weak supervision from image/text pairs. To the best of our knowledge, there is no method for approaching SSC without the need for any ground-truth annotations. Our work presents the first unsupervised SSC approach, utilizing lifted SSL features and a single RGB input image for inference. + +Self-supervised representation learning (SSL) aims to extract general features from data without annotations, facilitating various downstream tasks such as segmentation [24]. Recent SSL methods, often based on Vision Transformers (ViTs) [23], leverage clustering [2, 9, 10, 48, 61], masked modeling [20, 29, 40, 77, 107], contrastive learning [3, 12, 14, 39, 41, 42], or negative-free [4, 5, 11, 28, 80] pretext tasks [22, 79] for large-scale training. State-of-the-art models, e.g., DINO [11], produce semantically rich, dense features, driving recent advances in 2D unsupervised scene understanding [33, 104]. We here aim to bring expressive features from DINO [11, 80] to 3D for SSC. + +2D-to-3D feature lifting. The expressiveness of 2D visual representations has motivated lifting 2D features into 3D [94, 111]. Existing approaches utilize multi-view 2D features for 3D feature lifting [30, 44, 50, 54, 73, 83, 93, 94, 98, 101, 102, 106, 111, 116]. Lifting 2D features is effective in various tasks, including few-shot semantic occupancy prediction [111], and refining 2D representations [116]. However, existing feature-lifting approaches fit to a single scene [50, 54, 93, 94, 101, 102, 111, 116], require RGB-D inputs [30, 44, 73, 98, 106], or work on 3D point cloud inputs [83]. The only feed-forward approaches that use RGB inputs and lift 2D features, which we are aware of, are GaussTR [49] and MVSplat360 [15]. However, both approaches utilize multiple input images during inference, and MVSplat360 [15] only predicts low-dimensional feature representations, which are not suitable for unsupervised scene understanding. In contrast, we propose the first feedforward approach for inferring lifted high-dimensional and rich 3D features using a single input image. + +2D unsupervised semantic segmentation partitions images automatically into semantically meaningful regions without any form of human annotations. Early deep learning-based methods [17, 36, 48] approach the problem via representation learning. Leveraging SSL features from DINO as an inductive prior, STEGO [33] distills the feature representation into a lower-dimensional space for unsupervised probing. Building up on STEGO, subsequent methods + +![](images/7a32d6794c76d05a04ce2206d89ab5f3c7736bf7547632ad0614e5dff0fc093d.jpg) +Figure 2. SceneDINO architecture, rendering, and training. (a) Inference: Given a single input image $\mathbf{I}_0$ during inference, a 2D encoder-decoder $\xi$ produces the embedding $\mathbf{E}$ from which the local embedding $\mathbf{e}_{\mathbf{u}}$ is interpolated. The MLP encoder $\phi$ takes in $\mathbf{e}_{\mathbf{u}}$ and 3D position $\mathbf{x}_i$ , and predicts both the density $\sigma_{\mathbf{x}_i}$ and the 3D feature $f_{\mathbf{x}_i}$ . Using a lightweight unsupervised segmentation head $h$ , we can obtain semantic predictions $p_{\mathbf{x}_i}$ using $f_{\mathbf{x}_i}$ . (b) Rendering: Our feature field allows for volume rendering by shooting rays through it, yielding depth $\hat{d}$ and $\hat{f}$ in 2D. Color $c_i$ is sampled from an another view ( $e.g., \mathbf{I}_1$ ) using $\mathbf{u}_s$ and rendered to obtain the reconstructed color $\hat{c}$ . (c) Multi-view training: We render 2D views (features & images) from our feature field and reconstruct the training views. + +![](images/30dd532230d021ee8257fae00817fed36b4bbce0d48b19a7f7209dd2246ec836.jpg) + +![](images/6072acde5b5bbd7b0ebfd347875ac119c96457bbc4e47eaa0ac16f0ad1b0fef0.jpg) + +[31, 51, 92, 95] propose enhancements to the distillation. Our approach follows the idea of STEGO [33], extending it to 3D and integrating feature distillation using our 3D feature field to build the first unsupervised SSC approach. + +# 3. Unsupervised Semantic Scene Completion + +We approach semantic scene completion (SSC) without any form of manual supervision. To this end, we first describe SceneDINO, predicting 3D geometry and expressive 3D features from a single image in a feed-forward manner (Sec. 3.1), and SceneDINO's multi-view training (Sec. 3.2). Next, we present our 3D feature distillation approach to obtain unsupervised 3D semantic predictions (Sec. 3.3). An overview of our full pipeline, including inference, rendering, and multi-view self-supervision, is provided in Fig. 2. + +Notation. Let $\mathbf{I}_0\in [0,1]^{3\times \mathrm{H}\times \mathrm{W}}$ be a single RGB input image (for both training & inference) with corresponding pose $T_0\in \mathbb{R}^{4\times 4}$ and projection matrix $K_{0}\in \mathbb{R}^{3\times 4}$ . For training, let $(\mathbf{I}_v,T_v,K_v)$ with $v\in \{1,2,\dots ,n\}$ , be $n$ additional views for multi-view self-supervision. Assuming a pinhole camera model, any 3D point $\mathbf{x}\in \mathbb{R}^3$ in world coordinates can be projected onto the image plane of view $v$ and the input view $v = 0$ with the perspective projection $\pi_v(\mathbf{x})$ . + +# 3.1. SceneDINO + +Given a single input image $\mathbf{I}_0$ , SceneDINO represents the dense geometric structure and features of a scene as a continuous mapping from world coordinates $\mathbf{x} \in \mathbb{R}^3$ to a volumetric density $\sigma_{\mathbf{x}} \in \mathbb{R}_{+}$ and a feature $f_{\mathbf{x}} \in \mathbb{R}^{\mathrm{D}}$ . This continuous output representation is often called a feature field. While SceneDINO could represent any feature space, we aim for expressive SSL features from DINO [11, 80]. + +Architecture & feature field inference. Our SceneDINO architecture comprises two main parts: a 2D encoder- + +decoder $\xi$ and an MLP decoder (cf. Fig. 2a), following BTS [108]. $\xi$ takes in $\mathbf{I}_0$ and produces a per-pixel embedding $\mathbf{E} \in \mathbb{R}^{\mathrm{D}_{\mathbf{E}} \times \mathrm{H} \times \mathrm{W}}$ with $\mathrm{D}_{\mathbf{E}}$ dimensions. Intuitively, every spatial element of $\mathbf{E}$ represents a camera ray through a pixel, capturing both local geometry and features. + +To infer the feature at a 3D position $\mathbf{x}$ , we employ a two-layer MLP decoder $\phi$ (cf. Fig. 2a). Given a position $\mathbf{x}$ within the camera frustum, we project $\mathbf{x}$ into the camera plane, obtaining the pixel location $\mathbf{u} = \pi_0(\mathbf{x})$ . We query $\mathbf{E}$ at the position $\mathbf{u}$ using bilinear interpolation, obtaining the local embedding $\mathbf{e_u}$ . Based on the embedding $\mathbf{e_u}$ , the pixel position $\mathbf{u}$ , and the distance $d_{\mathbf{x}} \in \mathbb{R}_{+}$ of $\mathbf{x}$ to the camera, we obtain the density $\sigma_{\mathbf{x}}$ and feature prediction $f_{\mathbf{x}}$ as + +$$ +\left(\sigma_ {\mathbf {x}}, f _ {\mathbf {x}}\right) = \phi \left(\mathbf {e} _ {\mathbf {u}}, \gamma (\mathbf {u}, d _ {\mathbf {x}})\right), \tag {1} +$$ + +where $\gamma$ denotes a positional encoding [75]. + +Feature, depth & color volume rendering. SceneDINO predicts a continuous feature field from a single image. This representation can be used to render features and depth in 2D from an arbitrary viewpoint (cf. Fig. 2b), following the discretization strategy of Max et al. [72]. Given a viewpoint $(T,K)$ , we sample L points $\mathbf{x}_i$ along the ray through pixel $\mathbf{u}_r$ , with distance $\delta_i$ between $\mathbf{x}_i$ and $\mathbf{x}_{i+1}$ . Based on the volumetric densities $\sigma_{\mathbf{x}_i}$ (cf. Eq. 1), we can compute the probabilities $\alpha_i$ of the ray ending between $\mathbf{x}_i$ and $\mathbf{x}_{i+1}$ , and accumulate these into $V_i$ , the probability of $\mathbf{x}_i$ being visible: + +$$ +V _ {i} = \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j}), \quad \text {w i t h} \alpha_ {i} = 1 - \exp (- \sigma_ {\mathbf {x} _ {i}} \delta_ {i}). \tag {2} +$$ + +Using $V_{i}$ and $\alpha_{i}$ , we render depth $\tilde{d}_{\mathbf{u}_r}$ and feature $\tilde{f}_{\mathbf{u}_r}$ from the estimated features $f_{\mathbf{x}_i}$ from Eq. (1) and distances $d_{\mathbf{x}_i}$ to $\mathbf{x}_i$ onto the image plane at position $\mathbf{u}_r$ as + +$$ +\tilde {f} _ {\mathbf {u} _ {r}} = \sum_ {i = 1} ^ {\mathrm {L}} V _ {i} \alpha_ {i} f _ {\mathbf {x} _ {i}} \quad \tilde {d} _ {\mathbf {u} _ {r}} = \sum_ {i = 1} ^ {\mathrm {L}} V _ {i} \alpha_ {i} d _ {\mathbf {x} _ {i}}. \tag {3} +$$ + +The differentiability of this rendering process enables us to self-supervise SceneDINO using multi-view images and their 2D feature representations (e.g., from DINO [11]). SceneDINO predicts 3D geometry and features, but does not predict color as we focus on semantic downstream tasks. To obtain color for image reconstruction during training, we follow the color sampling approach of BTS [108]. + +# 3.2. 3D feature field training + +We train SceneDINO using multi-view self-supervision (cf. Fig. 2c), aiming to obtain an expressive and view-consistent feature field without the need for any form of manual annotations. For self-supervision, we sample $n + 1$ views $\mathbf{I}_v$ with camera parameters $T_v, K_v$ from the data and obtain dense 2D target features $\mathbf{F}_v$ from a self-supervised ViT (e.g., DINO [11]). Note that the 2D features entail a resolution of $\mathbf{F}_v \in \mathbb{R}^{\mathrm{D} \times \frac{\mathrm{H}}{p} \times \frac{\mathrm{W}}{p}}$ , due to the ViT patch size $p$ . The set of training views and features $\mathbb{V} = \{( \mathbf{I}_v, T_v, K_v, \mathbf{F}_v) | v = 0, \dots, n \}$ is randomly partitioned into two subsets $\mathbb{V}_{\text{source}}$ and $\mathbb{V}_{\text{target}}$ . Training reconstructs the views $\mathbb{V}_{\text{target}}$ using the views of $\mathbb{V}_{\text{source}}$ . In practice, we use a randomly sampled set of image patches that align with the ViT patches instead of the full image. In the following, we still refer to images and the full image resolution for the sake of brevity. + +Image reconstruction. We aim to learn the geometry of our feature field via multi-view photometric consistency. In particular, for every image $\mathbf{I}_t\in \mathbb{V}_{\mathrm{target}}$ we derive a reconstructed image $\hat{\mathbf{I}}_{t,s}$ from every view $s$ in $\mathbb{V}_{\mathrm{source}}$ using differentiable rendering (cf. Eq. 3) and color sampling [108]. Equipped with both the reconstructed image $\hat{\mathbf{I}}_{t,s}$ and the target image $\mathbf{I}_t$ , we compute our photometric loss per view as + +$$ +\mathcal {L} _ {\mathrm {p}} = \min _ {\mathbf {I} _ {s} \in \mathbb {V} _ {\text {s o u r c e}}} \left(\lambda_ {1} \mathcal {L} _ {1} \left(\mathbf {I} _ {t}, \hat {\mathbf {I}} _ {t, s}\right) + \lambda_ {\text {S S I M}} \mathcal {L} _ {\text {S S I M}} \left(\mathbf {I} _ {t}, \hat {\mathbf {I}} _ {t, s}\right)\right). \tag {4} +$$ + +We only consider the minimum loss across the views in $\mathbb{V}_{\mathrm{source}}$ , in practice across patches. The scalars $\lambda_{1}$ and $\lambda_{\mathrm{SSIM}}$ weight the absolute error $\mathcal{L}_1$ and the SSIM loss $\mathcal{L}_{\mathrm{SSIM}}$ [105]. + +To regularize our 3D geometry, we impose smoothness using an edge-aware smoothness loss [27]. We estimate 2D depth maps $\tilde{\mathbf{d}}_t$ for views in $\mathbb{V}_{\mathrm{target}}$ , using Eq. 3. From $\tilde{\mathbf{d}}_t$ , we obtain the inverse mean-normalized depths $\tilde{\mathbf{d}}_t^*$ and compute the edge-aware smoothness loss $\mathcal{L}_{\mathrm{s}}$ per view as + +$$ +\mathcal {L} _ {\mathrm {s}} = \left| \nabla_ {x} \tilde {\mathbf {d}} _ {t} ^ {*} \right| e ^ {- \left| \nabla_ {x} \mathbf {I} _ {t} \right|} + \left| \nabla_ {y} \tilde {\mathbf {d}} _ {t} ^ {*} \right| e ^ {- \left| \nabla_ {y} \mathbf {I} _ {t} \right|}, \tag {5} +$$ + +where $\nabla_{x}$ and $\nabla_{y}$ denote the first spatial derivatives. + +Feature reconstruction. We learn a multi-view consistent and expressive 3D feature field using the 2D features $\mathbf{F}_t$ from $\mathbb{V}_{\mathrm{target}}$ . As we aim to learn a high-resolution (continuous) feature field, we render 2D features using Eq. 3 at the full image resolution $\hat{\mathbf{F}}_t \in \mathbb{R}^{D \times H \times W}$ . To compensate for + +![](images/14640c21cf9305b7701c5a14a164adcf218d304d70e5071325f4210f6cc5d0cf.jpg) +Figure 3. 3D feature distillation. Given an input image, SceneDINO predicts a 3D feature field. 3D features $f_{\mathbf{X}}$ are sampled from the feature field. For $f_{\mathbf{X}}$ , we obtain $f_{\mathbf{Y}_{k\mathrm{NN}}}$ and $f_{\mathbf{Y}_{kr\mathrm{and}}}$ from the feature buffer. The segmentation head $h$ distills the features into a low-dimensional space and is trained using $\mathcal{L}_{\mathrm{dist}}$ . + +the reduced spatial dimension of $\mathbf{F}_t$ , we employ the down-sampler $\psi$ proposed by Fu et al. [25] to our rendered features $\hat{\mathbf{F}}_t$ . While current 2D SSL features capture semantics, they lack multi-view consistency, i.a., due to positional encodings used in ViTs [112], leading to different features for identical visual content at two distinct positions in an image. As we aim for multi-view consistency, we compensate for this by learning a constant decomposition $\overline{\mathbf{F}} \in \mathbb{R}^{\mathrm{D} \times \frac{\mathrm{H}}{p} \times \frac{\mathrm{W}}{p}}$ of features induced by positional encodings. Our feature loss is defined per view as + +$$ +\mathcal {L} _ {\mathrm {f}} = 1 - \cos - \sin \left(\mathbf {F} _ {t}, \psi \left(\hat {\mathbf {F}} _ {t}\right) + \overline {{\mathbf {F}}}\right), \tag {6} +$$ + +where cos-sim is the cosine similarity between two features. + +As image edges correlate with semantic edges and to further impose consistency, we regularize the rendered features $\hat{\mathbf{F}}_t$ using an edge-aware smoothness loss per view + +$$ +\mathcal {L} _ {\mathrm {f s}} = \left| \nabla_ {x} \hat {\mathbf {F}} _ {t} \right| e ^ {- \left| \nabla_ {x} \mathbf {I} _ {t} \right|} + \left| \nabla_ {y} \hat {\mathbf {F}} _ {t} \right| e ^ {- \left| \nabla_ {y} \mathbf {I} _ {t} \right|}. \tag {7} +$$ + +Our final loss for training SceneDINO is a weighted sum of the photometric loss, the feature loss, and both smoothness losses, $\mathcal{L}_{\mathrm{SceneDINO}} = \lambda_{\mathrm{p}}\mathcal{L}_{\mathrm{p}} + \lambda_{\mathrm{s}}\mathcal{L}_{\mathrm{s}} + \lambda_{\mathrm{f}}\mathcal{L}_{\mathrm{f}} + \lambda_{\mathrm{fs}}\mathcal{L}_{\mathrm{fs}}$ , averaged over all pixels, features, and views. + +# 3.3. 3D feature distillation for unsupervised SSC + +Given the expressive feature field representation, we aim to obtain unsupervised semantic predictions for SSC. While naive $k$ -means [69, 71] can yield meaningful pseudo semantics, distilling features into a lower-dimensional space has been shown to be more effective in 2D semantic segmentation [33, 55]. To this end, we present a novel 3D feature distillation approach (cf. Fig. 3). We train a pointwise segmentation head $h$ , mapping $f_{\mathbf{x}} \in \mathbb{R}^{\mathrm{D}}$ to a lower-dimensional distilled representation $z_{\mathbf{x}} \in \mathbb{R}^{\mathrm{K}}$ , with $\mathrm{K} \ll \mathrm{D}$ . The resulting distilled space is clustered to obtain pseudosemantic predictions $p_{\mathbf{x}} \in [0,1]^{\mathrm{C}}$ , with C pseudo classes. + +Existing work in 2D unsupervised semantic segmentation has shown that SSL feature correspondence captures + +semantic class co-occurrence [33]. This correspondence between two batches of $\mathbf{N}$ sample points $\mathbf{X} = [\mathbf{x}_1, \dots, \mathbf{x}_N]$ and $\mathbf{Y} = [\mathbf{y}_1, \dots, \mathbf{y}_N]$ can be expressed by pairwise feature similarity $S_{i,j} = \cos -\sin (f_{\mathbf{x}_i}, f_{\mathbf{y}_j}) \in [-1, 1]$ . Similarly, we can express the correspondence in the distilled feature space by $S_{i,j}^h = \cos -\sin (h(f_{\mathbf{x}_i}), h(f_{\mathbf{y}_j})) \in [-1, 1]$ . We describe the sampling of the $\mathbf{x}_i$ and $\mathbf{y}_j$ below. + +Feature distillation. We aim to distill features such that similar features align while dissimilar features are separated. To this end, we use the contrastive correlation loss $\mathcal{L}_{\mathrm{corr}}$ , introduced by STEGO [33] and defined as + +$$ +\mathcal {L} _ {\text {c o r r}} \left(f _ {\mathbf {X}}, f _ {\mathbf {Y}}, b\right) = - \sum_ {i, j} \left(S _ {i, j} - b\right) \max \left(S _ {i, j} ^ {h}, 0\right), \tag {8} +$$ + +where $f_{\mathbf{X}}$ , $f_{\mathbf{Y}}$ are the features of the two sample batches. This loss pushes $S_{i,j}^{h}$ towards 1 in case $S_{i,j}$ exceeds the threshold $b$ . Otherwise, $\mathcal{L}_{\mathrm{corr}}$ pushes the $S_{i,j}^{h}$ below 0. + +The correlation loss $\mathcal{L}_{\mathrm{corr}}$ requires informative pairs of sampled features, balancing attractive and repulsive signals. Following STEGO [33], we consider three different relations: (1) feature pairs from the same image $(f_{\mathbf{X}}, f_{\mathbf{X}})$ , (2) feature pairs from an image and its $k$ -nearest neighbors in feature space $(f_{\mathbf{X}}, f_{\mathbf{Y}_{k\mathrm{NN}}})$ , and (3) feature pairs from an image and a randomly sampled other image $(f_{\mathbf{X}}, f_{\mathbf{Y}_{\mathrm{rand}}})$ . Note that each pair is obtained from SceneDINO's 3D feature field, see below. Equipped with the three feature sample pairs, we compute the full distillation loss as + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {d i s t}} = \lambda_ {\mathrm {s e l f}} \mathcal {L} _ {\mathrm {c o r r}} \left(f _ {\mathbf {X}}, f _ {\mathbf {X}}, b _ {\mathrm {s e l f}}\right) \\ + \lambda_ {k \mathrm {N N}} \mathcal {L} _ {\text {c o r r}} \left(f _ {\mathbf {X}}, f _ {\mathbf {Y} _ {k \mathrm {N N}}}, b _ {k \mathrm {N N}}\right) \tag {9} \\ + \lambda_ {\text {r a n d}} \mathcal {L} _ {\text {c o r r}} \left(f _ {\mathbf {X}}, f _ {\mathbf {Y} _ {\text {r a n d}}}, b _ {\text {r a n d}}\right), \\ \end{array} +$$ + +where $\lambda_{\mathrm{self}}$ , $\lambda_{k\mathrm{NN}}$ , and $\lambda_{\mathrm{rand}}$ denote the scalar loss weights. $b_{\mathrm{self}}$ , $b_{k\mathrm{NN}}$ , and $b_{\mathrm{rand}}$ are the contrastive thresholds. + +Feature sampling in 3D. While obtaining feature pairs using 2D rendered features is straightforward [33], we aim to take advantage of our learned 3D geometry of the scene. To this end, we introduce a novel 3D feature sampling approach for the distillation loss $\mathcal{L}_{\mathrm{dist}}$ from Eq. (9). Our goal is to sample features both similar and dissimilar in terms of the encoded semantic concept, which should capture rich semantics as well as different semantic concepts. + +First, we obtain all G visible 3D surface points $\mathbf{V} \in \mathbb{R}^{3 \times \mathrm{G}}$ and their depth $d_{\mathbf{V}} \in \mathbb{R}_{+}^{\mathrm{G}}$ from the camera. To sample points that cover different semantic concepts, we use depth as a cue and sample different depth ranges. In particular, we sort the surface points $\mathbf{V}$ based on $d_{\mathbf{V}}$ . The sorted surface points $\hat{\mathbf{V}}$ are partitioned into M equally-sized chunks; we uniformly sample a single 3D point from each chunk, resulting in M center points $\mathbf{X} \in \mathbb{R}^{3 \times \mathrm{M}}$ . + +Equipped with the center points $\mathbf{X}$ , we aim to extract rich semantic features from the feature field. While we could just obtain the features for $\mathbf{X}$ , we query positions in the + +![](images/7dd58bbc0b5b72c0032e950414ba5056d6f42570149f97ded9eebfe5bfeac687.jpg) +Figure 4. 3D feature sampling. We first sample a center point $\mathbf{X}_i$ from all visible surface points. Further points are sampled within the radius $r$ around the center point $\mathbf{X}_i$ . Sampled points with sufficient density are accepted; otherwise rejected. The accepted points are used to obtain the feature batch $f_{\mathbf{X}}$ . + +neighborhood of $\mathbf{X}$ to increase semantic richness and better capture the 3D structure of the scene for distillation. In particular, for each center point, we randomly sample a point within a radius of $r = 0.5\mathrm{m}$ . To account for samples falling into unoccupied regions in our feature field, we only keep samples with a sufficient density $\sigma > 0.5$ . We repeat this sampling process until we obtain $N$ valid samples per center point. Using these samples, we query our feature field, resulting in a feature batch $f_{\mathbf{X}} \in \mathbb{R}^{D \times N}$ for each of the G center points in each scene (cf. Fig. 4). + +To obtain $f_{\mathbf{Y}_{k\mathrm{NN}}}$ and $f_{\mathbf{Y}_{\mathrm{rand}}}$ , we utilize a feature buffer that efficiently stores the sampled features of multiple scenes. Given a new input image, we obtain $G$ feature batches $f_{\mathbf{X}}$ as just described. For each $f_{\mathbf{X}}$ , we randomly sample another feature batch from the buffer to obtain $f_{\mathbf{Y}_{\mathrm{rand}}}$ . To obtain $f_{\mathbf{Y}_{k\mathrm{NN}}}$ , we search in the feature buffer for the $k$ -nearest neighbors of $f_{\mathbf{X}}$ , using the average feature of each batch. From these $k$ -nearest neighbors, we randomly pick a feature batch to obtain $f_{\mathbf{Y}_{k\mathrm{NN}}}$ and compute the distillation loss $\mathcal{L}_{\mathrm{dist}}$ . After repeating this process for each of the current $G$ feature batches, we add the current feature batches to the feature buffer and remove the oldest batches. + +Unsupervised probing. To obtain semantics, we probe the distilled K-dim. feature space using $k$ -means [69, 71]. In particular, we update cluster centers $\theta \in \mathbb{R}^{K\times C}$ using cosine similarity-based mini-batch $k$ -means [91] during distillation. We compute $p_{\mathbf{x}} = \mathrm{softmax}(\cos -\mathrm{sim}(h(f_{\mathbf{x}}),\theta))$ to infer C pseudo semantic class predictions. + +# 4. Experiments + +We evaluate SceneDINO on SSC and compare it to a simple unsupervised SSC baseline (Sec. 4.1). We also report results for 2D unsupervised segmentation, including domain generalization results (Sec. 4.2). Finally, we explore multiview feature consistency (Sec. 4.3) and present an analysis of SceneDINO and our 3D distillation (Sec. 4.4). + +Datasets. We train using KITTI-360 [66], composed of clips from a moving vehicle equipped with cameras. For + +consistency, we follow S4C [38] by sampling eight views and using the dataset's camera poses. We further provide results with estimated poses. We also show experiments for training on RealEstate10k [119], composed of monocular videos. Here, we follow the setup of BTS [108], obtaining three views. If not noted differently, we report results obtained with training on KITTI-360. For SSC and 2D semantic segmentation validation, we use the SSCBench-KITTI-360 test split [64]. Cityscapes [19] and BDD100K [115] val are used for domain generalization results. To enable evaluation in 3D and 2D, we use the 19-class taxonomy of Cityscapes and perform 2D evaluation on Cityscapes, BDD100K, and KITTI-360 on 19 classes. For SSCBench, we combine classes to adhere to the 15 SSCBench classes. + +3D evaluation. Given our unsupervised setup, we predict pseudo-semantic classes that must be aligned with the ground truth for evaluation. We follow standard practice in 2D unsupervised semantic segmentation [17, 31, 33, 51, 92, 95] by applying Hungarian matching [57] to align our pseudo semantics. For validating the aligned semantics, we follow the standardized SSCBench [64] protocol and report both semantic performance using the mean Intersection-over-Union (mIoU) and geometric performance using IoU, precision, and recall. We report all metrics on SSCBench ranges $12.8\mathrm{m}$ , $25.6\mathrm{m}$ , and $51.2\mathrm{m}$ . + +2D evaluation. Following the established evaluation protocol in 2D unsupervised semantic segmentation [17, 31, 33, 51, 92, 95], we use the all-pixel accuracy (Acc) and mean Intersection-over-Union (mIoU) metrics. Likewise, in line with prior work, 2D segmentation predictions of all models are refined using a dense Conditional Random Field [56] before computing Acc and mIoU. + +Multi-view feature consistency evaluation. We aim to evaluate the multi-view consistency of our feature field. As we are not aware of any general feed-forward 3D feature field approach, we compare against 2D SSL models. To measure multi-view consistency in 2D, we use two video frames and estimate optical flow and occlusions with RAFT [99]. We backward warp 2D features of the second frame to the first. On the aligned features, we compute the feature similarity using absolute error $(\mathrm{L}_1)$ , the Euclidean distance $(\mathrm{L}_2)$ , and the cosine similarity, ignoring occlusions. + +Baselines. We are not aware of any existing unsupervised SSC approach. To allow for comparisons, we construct a competitive baseline for unsupervised SSC. In particular, we train the S4C approach with unsupervised semantics of the established STEGO [33] approach. For 2D semantic segmentation, we use U2Seg [78] and STEGO [33] as established unsupervised baselines. Note that U2Seg is trained on ImageNet [21] and COCO [67] using STEGO pseudo-labels. We use STEGO [33] with DINO [11] (ViT-B/8), DINOv2 [80] (ViT-B/14), and FiT3D [116] (ViT + +Table 1. SSCBench-KITTI-360 results. Semantic results using mIoU and per class IoU, and geometric results using IoU, Precision, and Recall (all in %, ↑) on SSCBench-KITTI-360 test using three depth ranges. We compare our baseline S4C + STEGO to our SceneDINO. We report S4C as a 2D supervised baseline. + +
MethodS4C + STEGOSceneDINO (Ours)S4C
SupervisionUnsupervised2D supervision
Range12.8 m25.6 m51.2 m12.8 m25.6 m51.2 m12.8 m25.6 m51.2 m
Semantic validation
mIoU10.539.266.6010.7610.018.0016.9413.9410.19
car18.5714.099.2221.2415.9411.2122.5818.6411.49
bicycle0.010.010.010.000.000.000.000.000.00
motorcycle0.000.000.000.000.000.000.000.000.00
truck0.110.040.020.000.000.007.514.372.12
other-v.0.010.050.020.000.000.000.000.010.06
person0.010.010.010.000.000.000.000.000.00
road61.9752.4738.1551.1049.1239.8269.3861.4648.23
sidewalk18.7420.9518.2120.2622.3118.9745.0337.1228.45
building14.7524.4417.8112.3318.2714.3226.3428.4821.36
fence1.410.200.111.910.900.589.706.373.64
vegetation15.8316.5811.3031.2225.5719.8535.7828.0421.43
terrain26.499.954.1723.2618.0215.2235.0322.8815.08
pole0.080.040.040.050.050.051.230.940.65
traffic-sign0.000.000.000.000.000.001.570.830.36
other_obj.0.050.040.020.000.000.000.000.000.00
Geometric validation
IoU49.3241.0836.3949.5442.2737.6054.6445.5739.35
Precision54.0446.2341.9153.2746.1041.5959.7550.3443.59
Recall84.9578.6973.4387.6183.5979.6786.4782.7980.16
+ +B/14) features. FiT3D offers multi-view refined DINOv2 features [116]. Note that FiT3D reports results, concatenating the refined features with DINOv2 features. We report results using both plain features only and the concatenation. We also use rendered 2D segmentations of our S4C + STEGO baseline for 2D validation. For multi-view feature consistency, we utilize DINO [11], DINOv2, and FiT3D [116] features as baselines. + +Implementation details. Our encoder-decoder uses a DINO-B/8 [11] backbone and a dense prediction decoder [84]. The MLP decoder $\phi$ entails two layers with 128 hidden features. As rendering features is expensive, $\phi$ predicts 64 features. We employ another MLP to upproject again to the full dimensionality $D = 768$ . If not stated differently, our target features are obtained from DINO-B/8 [11]. We train using a batch size of 4 and extract 32 patches of size $8 \times 8$ from each image to compute $\mathcal{L}_{\mathrm{SceneDINO}}$ . Volume rendering samples each ray at $L = 32$ uniformly spaced points in inverse depth within $[3\mathrm{m}, 80\mathrm{m}]$ . We train for $100\mathrm{k}$ steps using Adam [52] with a base learning rate of $10^{-4}$ . Training takes ca. 2 days on a single V100 GPU. We distill using a batch size of 4, 5 center points, a feature batch of size 576, and cluster with $K = 19$ . For kNN sampling, we use $k = 4$ . The feature buffer holds 256 feature batches. Refer to the supplement for more details. + +# 4.1.3D semantic scene completion + +We assess the unsupervised SSC and geometric accuracy of SceneDINO with our 3D feature distillation approach on SSCBench-KITTI-360. In particular, Tab. 1 com + +![](images/810d3f72e739835c1aaaf5f4c812a7b34c5c4dec26fa2b38671b11d7b415a1a0.jpg) +Figure 5. Qualitative SSC comparison on KITTI-360. We show the input image, SceneDINO's feature field using the first three principal components and SSC prediction, the prediction of our baseline S4C + STEGO, and the ground truth. We only visualize surface voxels. Qualitative results show the expressiveness of our feature field and SceneDINO's capabilities to accurately reconstruct and label a scene. + +Table 2. 2D unsupervised semantic segmentation results on KITTI-360. Comparing SceneDINO to existing 2D methods and our S4C + STEGO 3D baseline, using Accuracy and mean IoU (in $\%$ , $\uparrow$ ) on the SSCBench-KITTI-360 test split. $\dagger$ denotes the use of plain FiT3D features. $\ddagger$ denotes training on ImageNet and COCO. + +
MethodFeaturesAccmIoU
U2Seg‡ [78]-72.8923.43
STEGO [33]DINO [11]73.3223.57
STEGO [33]DINOv2 [80]64.5424.82
STEGO [33]FiT3D [116]54.1922.29
STEGO [33]FiT3D† [116]57.2518.95
S4C [38] + STEGO [33]DINO [11]65.1621.67
SceneDINO (Ours)DINO [11]77.7425.81
+ +pares SceneDINO against our unsupervised SSC baseline S4C [38] + STEGO [33]. SceneDINO achieves a (semantic) mIoU of $8.0\%$ for the range of $51.2\mathrm{m}$ , significantly improving over our unsupervised baseline $(6.6\%)$ . This demonstrates that SceneDINO effectively lifts DINO features into 3D. In terms of geometric accuracy, SceneDINO moderately improves over S4C + STEGO. Despite being fully unsupervised, SceneDINO comes within $2.2\%$ points mIoU of the 2D-supervised S4C. + +Figure 5 provides qualitative samples on SSCBenchKITTI-360. SceneDINO's unsupervised SSC predictions are less noisy and capture finely resolved semantics compared to S4C + STEGO. Compared to the ground truth, we observe that SceneDINO captures both the geometry and general semantics of the scene. We visualize SceneDINO's feature field (before distillation) using the first three principal components. In PCA space, we observe that our feature field captures semantically meaningful regions. + +# 4.2. 2D semantic segmentation + +Table 2 compares the semantic predictions of SceneDINO to recent 2D approaches and our 3D baseline. We obtain 2D semantic segmentations from SceneDINO and our S4C + STEGO baseline using semantic rendering [38]. SceneDINO with our 3D distillation approach outperforms STEGO with DINO features, an established 2D unsuper + +Table 3. 2D unsupervised semantic segmentation domain generalization results. Comparing SceneDINO to existing 2D unsupervised semantic segmentation methods and S4C + STEGO 3D baseline, using Accuracy and mean IoU (in %, $\dagger$ ). We train on KITTI-360 images and report domain generalization results on Cityscapes and BDD-100K val. $\dagger$ denotes plain FiT3D features. + +
MethodFeaturesCityscapesBDD-100K
AccmIoUAccmIoU
U2Seg [78]-75.5718.6269.0017.99
STEGO [33]DINO [11]71.2119.4275.0221.41
STEGO [33]DINOv2 [80]68.4119.7365.7221.77
STEGO [33]FiT3D [116]66.9421.0165.9620.99
STEGO [33]FiT3D† [116]64.7617.1760.8319.09
S4C [38] + STEGO [33]DINO [11]54.8014.0444.9811.62
SceneDINO (Ours)DINO [11]73.1722.8172.2822.09
+ +vised semantic segmentation approach. In particular, the mIoU of SceneDINO is $2.24\%$ points higher than for STEGO (w/ DINO). Utilizing 3D refined features from FiT3D deteriorates the baseline relative to DINO, indicating that the FiT3D refinement reduces feature expressiveness. Notably, our unsupervised 3D baseline S4C + STEGO transfers significantly worse to 2D than SceneDINO. + +We also validate SceneDINO, trained on KITTI-360, on Cityscapes and BDD10K, demonstrating domain generalization. The results are reported in Tab. 3. SceneDINO outperforms all baselines in mIoU on both datasets while only falling short in Acc. S4C + STEGO falls short in generalization. We suspect this poor generalization is caused by the fact that S4C does not rely on general SSL features in the final model, while our feature field generalizes. + +# 4.3. Multi-view feature consistency + +We analyze the multi-view consistency of our feature field against existing 2D SSL features in Tab. 4. We report the results of SceneDINO trained on KITTI-360 and RealEstate10K. SceneDINO trained using DINO features exhibits significant improvements in multi-view feature consistency over standard DINO features. We also train SceneDINO using target features from DINOv2 [80]. Compared to standard DINOv2 and FiT3D fea + +Table 4. Multi-view consistency results. Comparing multi-view consistency of SceneDINO to existing 2D SSL features, using $\mathrm{L}_1$ distance $(\downarrow)$ , $\mathrm{L}_2$ distance $(\downarrow)$ , and cosine similarity $(\uparrow)$ on KITTI-360 and RealEstate10K. We compare DINO (top) and DINOv2-based (bottom) features. $\dagger$ denotes plain FiT3D features. + +
MethodKITTI-360RealEstate10K
L1L2cos-simL1L2cos-sim
DINO [11]16.060.740.7014.410.660.75
SceneDINO (w/ DINO)6.450.330.935.870.280.95
DINOv2 [80]15.830.730.7014.200.660.75
FiT3D [116]22.860.810.8219.880.720.85
FiT3D† [116]7.020.330.935.670.270.95
SceneDINO (w/ DINOv2)5.240.240.964.870.220.97
+ +Table 5. SceneDINO analysis. We analyze the role of decomposing positional encodings, the choice of downsampling features during training, the effectiveness of the feature smoothness loss, the effect of estimated camera poses, and the choice of target features. We report the mean IoU (in %, ↑) using a range of $51.2\mathrm{m}$ on SSCBench-KITTI-360 test. $\Delta$ mIoU reports the absolute difference in % points to our standard model with DINO target features. + +
Δ mIoUmIoUConfiguration
-1.186.82No downsampler (bilinear up. + aug.)
-1.176.83No feature smoothness loss (λfs = 0)
-0.747.26No pos. enc. decomposition
-0.127.88w/ estimated ORB-SLAM3 poses
—8.00Full framework (SceneDINO)
+1.089.08DINOv2 target features (vs. DINO)
+ +tures, SceneDINO's feature field yields significantly better multi-view consistency. Notably, compared against plain 3D refined features of FiT3D, SceneDINO shows a better multi-view consistency on both datasets and all metrics while also offering more expressiveness (cf. Tab. 2). + +# 4.4. Analyzing SceneDINO + +To understand what core components contribute to obtaining an expressive feature field of SceneDINO, we omit or replace individual components and report the results in Tab. 5. Replacing the downsampling approach with bilinear upsampling and multi-crop augmentations, similar to [1], to obtain high-resolution target features decreases the SSC mIoU by $1.18\%$ . Omitting the feature smoothness loss leads to a similar mIoU drop. Abolishing the constant decomposition of positional encodings leads to an mIoU drop of $0.74\%$ . Training using unsupervised camera poses estimated by ORB-SLAM3 [7] results in an insignificant mIoU drop of only $0.12\%$ over using KITTI-360 poses. Going from DINO to DINOv2 target features leads to an increased mIoU of $1.08\%$ , demonstrating that SceneDINO can benefit from more expressive 2D target features. + +In Tab. 6, we analyze our 3D distillation. Performing no distillation at all, just feature clustering, decreases mIoU by $1.61\%$ . Omitting the kNN-correlation loss leads to an mIoU drop of $1.35\%$ . Distilling only with center points, + +Table 6. Feature distillation analysis. We analyze the effectiveness of distilling SceneDINO's features, the $k$ -NN-correlation loss, our neighborhood sampling, and our 3D sampling approach over standard 5-crop sampling. We report the mean IoU (in %, ↑) using a range of $51.2 \mathrm{~m}$ on SSCBench-KITTI-360 test. + +
Δ mIoUmIoUConfiguration
-1.616.39No distillation
-1.356.65No kNN-correlation loss (λkNN = 0)
-0.977.03No neighborhood sampling (cf. Fig. 4)
-0.477.535-crop sampling [33] (instead 3D sampling)
—8.00Full framework (SceneDINO)
+ +Table 7. Probing analysis. We analyze linear and unsupervised probing of our distilled SceneDINO features on SSCBenchKITTI-360 test using mean IoU (in %, ↑). For reference, we also report S4C (2D supervised). Linear probing uses 2D annotations. + +
Probing approachTarget featuresmIoU
12.8 m25.6 m51.2 m
UnsupervisedDINO [11]10.7610.018.00
DINOv2 [80]13.7611.789.08
LinearDINO [11]13.6312.079.34
DINOv2 [80]15.8513.7010.57
S4C (full training)n/a16.9413.9410.19
+ +i.e., not performing neighborhood sampling (cf. Fig. 4), reduces mIoU by $0.97\%$ . Using 5-crop feature sampling [33], instead of our proposed 3D sampling, leads to a reduced mIoU of $0.47\%$ . This demonstrates the effectiveness of performing distillation in 3D using our novel approach. + +While focusing on unsupervised SSC, we can also linearly probe our distilled feature field (cf. Tab. 7). In particular, we train SceneDINO using different target features (DINO [11] and DINOv2 [11]), perform distillation, and probe the resulting distilled features. Using linear probing, i.e., training a single linear layer using 2D semantic labels, leads to a consistent mIoU increase over unsupervised probing. SceneDINO trained using DINOv2 target features even closes the gap to S4C, trained using 2D ground-truth semantic labels. We even surpass 2D supervised S4C slightly on the full range $(51.2\mathrm{m})$ , suggesting the effectiveness of SceneDINO also for weakly-supervised tasks. + +# 5. Conclusion + +We presented SceneDINO, to our knowledge, the first approach for unsupervised semantic scene completion. Trained using multi-view images and 2D DINO features without human supervision, SceneDINO is able to predict an expressive 3D feature field using a single input image during inference. Our novel 3D distillation approach yields state-of-the-art results in unsupervised SSC. While we focus on unsupervised SSC, our multi-view feature consistency, linear probing, and domain generalization results highlight the potential of SceneDINO as a strong foundation for various 3D scene-understanding tasks. + +Acknowledgments. This project was partially supported by the European Research Council (ERC) Advanced Grant SIMULACRON, DFG project CR 250/26-1 "4D-YouTube", and GNI Project "AICC". This project was also partially supported by the ERC under the European Union's Horizon 2020 research and innovation programme (grant agreement No. 866008). Additionally, this work has been co-funded by the LOEWE initiative (Hesse, Germany) within the emergenCITY center [LOEWE/1/12/519/03/05.001(0016)/72] and by the Deutsche Forschungsgemeinschaft (German Research Foundation, DFG) under Germany's Excellence Strategy (EXC 3066/1 "The Adaptive Mind", Project No. 533717223). Christoph Reich is supported by the Konrad Zuse School of Excellence in Learning and Intelligent Systems (ELIZA) through the DAAD programme Konrad Zuse Schools of Excellence in Artificial Intelligence, sponsored by the Federal Ministry of Education and Research. Christian Rupprecht is supported by an Amazon Research Award. Finally, we acknowledge the support of the European Laboratory for Learning and Intelligent Systems (ELLIS) and thank Mateo de Mayo as well as Igor Cvišić for help with estimating camera poses. + +# References + +[1] Nikita Araslanov and Stefan Roth. Self-supervised augmentation consistency for adapting semantic segmentation. In CVPR, pages 15384-15394, 2021. 8 +[2] Yuki Markus Asano, Christian Rupprecht, and Andrea Vedaldi. Self-labelling via simultaneous clustering and representation learning. In ICLR, 2020. 2 +[3] Philip Bachman, R. Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. In NeurIPS*2019, pages 15509-15519. 2 +[4] Adrien Bardes, Jean Ponce, and Yann LeCun. VI-CRegL: Self-supervised learning of local visual features. In NeurIPS*2022, pages 8799-8810. 2 +[5] Adrien Bardes, Jean Ponce, and Yann LeCun. VI-CReg: Variance-invariance-covariance regularization for self-supervised learning. In *ICLR*, 2022. 2 +[6] Yingjie Cai, Xuesong Chen, Chao Zhang, Kwan-Yee Lin, Xiaogang Wang, and Hongsheng Li. Semantic scene completion via integrating instances and scene in-the-loop. In CVPR, pages 324-333, 2021. 2 +[7] Carlos Campos, Richard Elvira, Juan J. Gómez Rodríguez, José M. M. Montiel, and Juan D Tardós. ORB-SLAM3: An accurate open-source library for visual, visual-inertial and multi-map SLAM. IEEE Trans. Robot., 37(6):1874-1890, 2021. 4, 8 +[8] Anh-Quan Cao and Raoul de Charette. Monoscene: Monocular 3D semantic scene completion. In CVPR, pages 3981-3991, 2022. 2 +[9] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS*2020, pages 9912-9924. 2 + +[10] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In ECCV, pages 132-149, 2018. 2 +[11] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. 1, 2, 3, 4, 6, 7, 8 +[12] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv:2003.04297 [cs.CV], 2020. 2 +[13] Xiaokang Chen, Kwan-Yee Lin, Chen Qian, Gang Zeng, and Hongsheng Li. 3d sketch-aware semantic scene completion via semi-supervised structure prior. In CVPR, pages 4192-4201, 2020. 2 +[14] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In CVPR, pages 9640-9649, 2021. 2 +[15] Yuedong Chen, Chuanxia Zheng, Haofei Xu, Bohan Zhuang, Andrea Vedaldi, Tat-Jen Cham, and Jianfei Cai. MVSplat360: Feed-forward 360 scene synthesis from sparse views. In NeurIPS*2024, pages 107064-107086. 2 +[16] Ran Cheng, Christopher Agia, Yuan Ren, Xinhai Li, and Bingbing Liu. S3CNet: A sparse semantic scene completion network for LiDAR point clouds. In CoRL, pages 2148-2161, 2020. 2 +[17] Jang Hyun Cho, Utkarsh Mall, Kavita Bala, and Bharath Hariharan. PiCIE: Unsupervised semantic segmentation using invariance and equivariance in clustering. In CVPR, pages 16794-16804, 2021. 2, 6 +[18] Özgün Çiçek, Ahmed Abdulkadir, Soeren S Lienkamp, Thomas Brox, and Olaf Ronneberger. 3D U-Net: Learning dense volumetric segmentation from sparse annotation. In MICCAI, pages 424–432, 2016. 1 +[19] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The Cityscapes dataset for semantic urban scene understanding. In CVPR, pages 3213-3223, 2016. 6 +[20] Timothee Darcet, Federico Baldassarre, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Cluster and predict latent patches for improved masked image modeling. arXiv:2502.08769 [cs.CV], 2025. 2 +[21] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 6 +[22] Carl Doersch, Abhinav Gupta, and Alexei A. Efros. Unsupervised visual representation learning by context prediction. In ICCV, pages 1422-1430, 2015. 2 +[23] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth $16 \times 16$ words: Transformers for image recognition at scale. In ICLR, 2021. 2 +[24] Linus Ericsson, Henry Gouk, Chen Change Loy, and Timothy M. Hospedales. Self-supervised representation learning: Introduction, advances, and challenges. IEEE Trans. Signal Process., 39(3):42-62, 2022. 2 + +[25] Stephanie Fu, Mark Hamilton, Laura E. Brandt, Axel Feldmann, Zhoutong Zhang, and William T. Freeman. FeatUp: A model-agnostic framework for features at any resolution. In ICLR, 2024. 4 +[26] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The KITTI dataset. Int. J. Robot. Res., 32(11):1231-1237, 2013. 1 +[27] Clément Godard, Oisin Mac Aodha, and Gabriel J. Brostow. Unsupervised monocular depth estimation with left-right consistency. In CVPR, pages 270-279, 2017. 4 +[28] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, et al. Bootstrap your own latent: A new approach to self-supervised learning. In NeurIPS*2020, pages 21271-21284. 2 +[29] Agrim Gupta, Jiajun Wu, Jia Deng, and Li Fei-Fei. Siamese Masked Autoencoders. In NeurIPS*2023, pages 40676-40693. 2 +[30] Huy Ha and Shuran Song. Semantic abstraction: Openworld 3D scene understanding from 2D vision-language models. In CoRL, pages 643-653, 2023. 2 +[31] Oliver Hahn, Nikita Araslanov, Simone Schaub-Meyer, and Stefan Roth. Boosting unsupervised semantic segmentation with principal mask proposals. Trans. Mach. Learn. Res., 2024. 3, 6 +[32] Oliver Hahn, Christoph Reich, Nikita Araslanov, Daniel Cremers, Christian Rupprecht, and Stefan Roth. Scene-centric unsupervised panoptic segmentation. In CVPR, pages 24485-24495, 2025. 1 +[33] Mark Hamilton, Zhoutong Zhang, Bharath Hariharan, Noah Snavely, and William T. Freeman. Unsupervised semantic segmentation by distilling feature correspondences. In ICLR, 2022. 1, 2, 3, 4, 5, 6, 7, 8 +[34] Keonhee Han, Dominik Muhle, Felix Wimbauer, and Daniel Cremers. Boosting self-supervision for single-view scene completion via knowledge distillation. In CVPR, pages 9837-9847, 2024. 1, 2 +[35] Xian-Feng Han, Hamid Laga, and Mohammed Bennamoun. Image-based 3D object reconstruction: State-of-the-art and trends in the deep learning era. IEEE Trans. Pattern Anal. Mach. Intell., 43(5):1578-1604, 2019. 2 +[36] Robert Harb and Patrick Knobelreiter. InfoSeg: Unsupervised semantic image segmentation with mutual information maximization. In $GCPR$ , pages 18-32, 2021. 2 +[37] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge University Press, 2003. 2 +[38] Adrian Hayler, Felix Wimbauer, Dominik Muhle, Christian Rupprecht, and Daniel Cremers. S4C: Self-supervised semantic scene completion with neural fields. In 3DV, pages 409-420, 2024. 1, 2, 6, 7 +[39] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9729-9738, 2020. 2 +[40] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 2 + +[41] Olivier Henaff. Data-efficient image pecognition with contrastive predictive coding. In ICML, pages 4182-4192, 2020. 2 +[42] R. Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. In ICLR, 2019. 2 +[43] Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3D v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. IEEE Trans. Pattern Anal. Mach. Intell., 46(12):10579-10596, 2024. 2 +[44] Rui Huang, Songyou Peng, Ayca Takmaz, Federico Tombari, Marc Pollefeys, Shiji Song, Gao Huang, and Francis Engelmann. Segment3D: Learning fine-grained class-agnostic 3D segmentation without manual labels. In ECCV, pages 278–295, 2024. 2 +[45] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3D semantic occupancy prediction. In CVPR, pages 9223-9232, 2023. 2 +[46] Yuanhui Huang, Wenzhao Zheng, Borui Zhang, Jie Zhou, and Jiwen Lu. SelfOcc: Self-supervised vision-based 3D occupancy prediction. In CVPR, pages 19946-19956, 2024. 1, 2 +[47] Joel Janai, Fatma Güney, Aseem Behl, and Andreas Geiger. Computer vision for autonomous vehicles: Problems, datasets and state of the art. Found. Trends Comput. Graph. Vis., 12(1-3):1-308, 2020. 1 +[48] Xu Ji, Joao F. Henriques, and Andrea Vedaldi. Invariant information clustering for unsupervised image classification and segmentation. In ICCV, pages 9865-9874, 2019. 2 +[49] Haoyi Jiang, Liu Liu, Tianheng Cheng, Xinjie Wang, Tianwei Lin, Zhizhong Su, Wenyu Liu, and Xinggang Wang. GaussTR: Foundation model-aligned Gaussian transformer for self-supervised 3D spatial understanding. arXiv:2412.13193 [cs.CV], 2024. 2 +[50] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. LERF: Language embedded radiance fields. In ICCV, pages 19729-19739, 2023. 2 +[51] Chanyoung Kim, Woojung Han, Dayun Ju, and Seong Jae Hwang. EAGLE: Eigen aggregation learning for object-centric unsupervised semantic segmentation. In CVPR, pages 3523-3533, 2024. 3, 6 +[52] Diederik P. Kingma and Jimmy Lei Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 6 +[53] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment Anything. In ICCV, pages 4015-4026, 2023. 1, 2 +[54] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing NeRF for editing via feature field distillation. In NeurIPS*2022, pages 23311-23330. 2 + +[55] Alexander Koenig, Maximilian Schambach, and Johannes Otterbach. Uncovering the inner workings of STEGO for safe unsupervised semantic segmentation. In CVPRW, pages 3789-3798, 2023. 4 +[56] Philipp Krahenbuhl and Vladlen Koltun. Efficient inference in fully connected CRFs with Gaussian edge potentials. In NIPS*2011, pages 109-117. 6 +[57] Harold W. Kuhn. The Hungarian method for the assignment problem. Nav. Res. Logist. Q., 2:83-97, 1955. 6 +[58] Jie Li, Yu Liu, Dong Gong, Qinfeng Shi, Xia Yuan, Chunxia Zhao, and Ian D. Reid. RGBD based dimensional decomposition residual network for 3D semantic scene completion. In CVPR, pages 7693-7702, 2019. 2 +[59] Jie Li, Kai Han, Peng Wang, Yu Liu, and Xia Yuan. Anisotropic convolutional networks for 3D semantic scene completion. In CVPR, pages 3348-3356, 2020. +[60] Jie Li, Yu Liu, Xia Yuan, Chunxia Zhao, Roland Siegwart, Ian Reid, and Cesar Cadena. Depth based semantic scene completion with position importance aware loss. IEEE Robotics Autom. Lett., 5(1):219-226, 2020. 2 +[61] Junnan Li, Pan Zhou, Caiming Xiong, and Steven Hoi. Prototypical contrastive learning of unsupervised representations. In ICLR, 2021. 2 +[62] Pengfei Li, Yongliang Shi, Tianyu Liu, Hao Zhao, Guyue Zhou, and Ya-Qin Zhang. Semi-supervised implicit scene completion from sparse LiDAR. arXiv:2111.14798 [cs.CV], 2021. 2 +[63] Yiming Li, Zhiding Yu, Christopher B. Choy, Chaowei Xiao, José M. Álvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. VoxFormer: Sparse voxel transformer for camera-based 3D semantic scene completion. In CVPR, pages 9087-9098, 2023. 2 +[64] Yiming Li, Sihang Li, Xinhao Liu, Moonjun Gong, Kenan Li, Nuo Chen, Zijun Wang, Zhiheng Li, Tao Jiang, Fisher Yu, Yue Wang, Hang Zhao, Zhiding Yu, and Chen Feng. SSCBench: A large-scale 3D semantic scene completion benchmark for autonomous driving. In IROS, pages 13333-13340, 2024. 1, 2, 6 +[65] Zhiqi Li, Zhiding Yu, David Austin, Mingsheng Fang, Shiyi Lan, Jan Kautz, and José M. Álvarez. FB-OCC: 3D occupancy prediction based on forward-backward view transformation. arXiv:2307.01492 [cs.CV], 2023. 2 +[66] Yiyi Liao, Jun Xie, and Andreas Geiger. KITTI-360: A novel dataset and benchmarks for urban scene understanding in 2D and 3D. IEEE Trans. Pattern Anal. Mach. Intell., 45(3):3292-3310, 2023. 1, 5 +[67] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, pages 740-755, 2024. 6 +[68] Shice Liu, Yu Hu, Yiming Zeng, Qiankun Tang, Beibei Jin, Yinhe Han, and Xiaowei Li. See and think: Disentangling semantic scene completion. In NeurIPS*2018, pages 261-272. 2 +[69] Stuart Lloyd. Least squares quantization in PCM. IEEE Trans. Inf. Theory, 28(2):129-137, 1982. 4, 5 + +[70] Zhiliang Ma and Shilong Liu. A review of 3D reconstruction techniques in civil engineering and their applications. Adv. Eng. Inform., 37:163-174, 2018. 1 +[71] James MacQueen. Some methods for classification and analysis of multivariate observations. In Berkeley Symp. on Math. Statist. and Prob., pages 281-298, 1967. 4, 5 +[72] Nelson Max. Optical models for direct volume rendering. IEEE Trans. Vis. Comput. Graph., 1(2):99-108, 1995. 3 +[73] Kirill Mazur, Edgar Sucar, and Andrew J Davison. Feature-realistic neural fusion for real-time, open set scene understanding. In ICRA, pages 8201-8207, 2023. 2 +[74] Ruihang Miao, Weizhou Liu, Mingrui Chen, Zheng Gong, Weixin Xu, Chen Hu, and Shuchang Zhou. Occdepth: A depth-aware method for 3D semantic scene completion. arXiv.2302.13540 [cs.CV], 2023. 1, 2 +[75] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. Commun. ACM, 65(1):99-106, 2021. 2, 3 +[76] Yue Ming, Xuyang Meng, Chunxiao Fan, and Hui Yu. Deep learning for monocular depth estimation: A review. Neurocomputing, 438:14-33, 2021. 2 +[77] Duy Kien Nguyen, Yanghao Li, Vaibhav Aggarwal, Martin R. Oswald, Alexander Kirillov, Cees G. M. Snoek, and Xinlei Chen. R-MAE: Regions meet masked autoencoders. In ICLR, 2024. 2 +[78] Dantong Niu, Xudong Wang, Xinyang Han, Long Lian, Roei Herzig, and Trevor Darrell. Unsupervised universal image segmentation. In CVPR, pages 22744-22754, 2024. 6, 7 +[79] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In ECCV, pages 69-84, 2016. 2 +[80] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. DINOv2: Learning robust visual features without supervision. Trans. Mach. Learn. Res., 2024. 2, 3, 6, 7, 8 +[81] Martin R Oswald, Eno Töppe, Claudia Nieuwenhuis, and Daniel Cremers. A review of geometry recovery from a single image focusing on curved object reconstruction. Innovations for Shape Analysis: Models and Algorithms, pages 343-378, 2013. 2 +[82] Mingjie Pan, Jiaming Liu, Renrui Zhang, Peixiang Huang, Xiaqi Li, Hongwei Xie, Bing Wang, Li Liu, and Shang-hang Zhang. RenderOcc: Vision-centric 3D occupancy prediction with 2D rendering supervision. In ICRA, pages 12404-12411, 2024. 2 +[83] Songyou Peng, Kyle Genova, Chiyu “Max” Jiang, Andrea Tagliasacchi, Marc Pollefeys, and Thomas Funkhouser. OpenScene: 3D scene understanding with open vocabularies. In CVPR, pages 815–824, 2023. 2 +[84] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In ICCV, pages 12179-12188, 2021. 6 + +[85] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, et al. SAM 2: Segment anything in images and videos. arXiv:2408.00714 [cs.CV], 2024. 1 +[86] Stephan R. Richter and Stefan Roth. Matryoshka networks: Predicting 3D geometry via nested shape layers. In CVPR, pages 1936-1944, 2018. 2 +[87] Christoph B. Rist, David Emmerichs, Markus Enzweiler, and Dariu M. Gavrila. Semantic scene completion using local deep implicit functions on LiDAR data. IEEE Trans. Pattern Anal. Mach. Intell., 44(10):7205-7218, 2022. 2 +[88] Luis Roldão, Raoul de Charette, and Anne Verroust-Blondet. LMSCNet: Lightweight multiscale 3D semantic completion. In 3DV, pages 111–119, 2020. 1, 2 +[89] Luis Roldao, Raoul De Charette, and Anne Verroust-Blondet. 3D semantic scene completion: A survey. Int. J. Comput. Vis., 130(8):1978-2005, 2022. 1 +[90] Johannes L. Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, pages 4104-4113, 2016. 2 +[91] David Sculley. Web-scale $k$ -means clustering. In WWW, page 1177-1178, 2010. 5 +[92] Hyun Seok Seong, WonJun Moon, SuBeen Lee, and Jae-Pil Heo. Leveraging hidden positives for unsupervised semantic segmentation. In CVPR, pages 19540-19549, 2023. 3, 6 +[93] Nur Muhammad Mahi Shafiullah, Chris Paxton, Lerrel Pinto, Soumith Chintala, and Arthur Szlam. CLIP-Fields: Weakly supervised semantic fields for robotic memory. In ICRA Workshop on Pretraining for Robotics, 2023. 2 +[94] William Shen, Ge Yang, Alan Yu, Jansen Wong, Leslie Pack Kaelbling, and Phillip Isola. Distilled feature fields enable few-shot language-guided manipulation. In CoRL, pages 405-424, 2023. 2 +[95] Leon Sick, Dominik Engel, Pedro Hermosilla, and Timo Ropinski. Unsupervised semantic segmentation through depth-guided feature correlation and sampling. In CVPR, pages 3637-3646, 2024. 3, 6 +[96] Shuran Song, Fisher Yu, Andy Zeng, Angel X. Chang, Manolis Savva, and Thomas A. Funkhouser. Semantic scene completion from a single depth image. In CVPR, pages 190-198, 2017. 1, 2 +[97] Stanislaw Szymanowicz, Eldar Insafutdinov, Chuanxia Zheng, Dylan Campbell, Joao F. Henriques, Christian Rupprecht, and Andrea Vedaldi. Flash3D: Feed-forward generalisable 3D scene reconstruction from a single image. arXiv:2406.04343 [cs.CV], 2024. 2 +[98] Ayca Takmaz, Elisabetta Fedele, Robert W. Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. OpenMask3D: Open-vocabulary 3D instance segmentation. In NeurIPS*2023, pages 68367-68390. 2 +[99] Zachary Teed and Jia Deng. RAFT: Recurrent all-pairs field transforms for optical flow. In ECCV, pages 402-419, 2020. 6 +[100] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua + +Lin, and Hongyang Li. Scene as occupancy. In ICCV, pages 8372-8381, 2023. 2 +[101] Nikolaos Tsagkas, Oisin Mac Aodha, and Chris Xiaoxuan Lu. VL-Fields: Towards language-grounded neural implicit spatial representations. In ICRA Workshop on Representations, Abstractions, and Priors for Robot Learning, 2023. 2 +[102] Vadim Tschernezki, Iro Laina, Diane Larlus, and Andrea Vedaldi. Neural feature fusion fields: 3D distillation of self-supervised 2D image representations. In 3DV, pages 443-453, 2022. 2 +[103] Shubham Tulsiani, Tinghui Zhou, Alexei A. Efros, and Jitendra Malik. Multi-view supervision for single-view reconstruction via differentiable ray consistency. In CVPR, 2017. 2 +[104] Xudong Wang, Rohit Girdhar, Stella X. Yu, and Ishan Misra. Cut and learn for unsupervised object detection and instance segmentation. In CVPR, pages 3124-3134, 2023. 1, 2 +[105] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: From error visibility to structural similarity. IEEE Trans. Image Process., 13(4): 600-612, 2004. 4 +[106] Silvan Weder, Hermann Blum, Francis Engelmann, and Marc Pollefeys. LabelMaker: Automatic semantic label generation from RGB-D trajectories. In 3DV, pages 334-343, 2024. 2 +[107] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In CVPR, pages 14668-14678, 2022. 2 +[108] Felix Wimbauer, Nan Yang, Christian Rupprecht, and Daniel Cremers. Behind the scenes: Density fields for single view reconstruction. In CVPR, pages 9076-9086, 2023. 1, 2, 3, 4, 6 +[109] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Comput. Graph. Forum, pages 641-676, 2022. 2 +[110] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep LiDAR point cloud segmentation via learning contextual shape priors from scene completion. In AAAI, pages 3101-3109, 2021. 2 +[111] Jiawei Yang, Boris Ivanovic, Or Litany, Xinshuo Weng, Seung Wook Kim, Boyi Li, Tong Che, Danfei Xu, Sanja Fidler, Marco Pavone, and Yue Wang. EmerneRF: Emergent spatial-temporal scene decomposition via self-supervision. In ICLR, 2024. 2 +[112] Jiawei Yang, Katie Z Luo, Jiefeng Li, Congyue Deng, Leonidas Guibas, Dilip Krishnan, Kilian Q Weinberger, Yonglong Tian, and Yue Wang. Denoising vision transformers. In ECCV, pages 453-469, 2024. 4 +[113] Zhuoyue Yang, Ju Dai, and Junjun Pan. 3D reconstruction from endoscopy images: A survey. Comput. Biol. Med., 175:108546, 2024. 1 + +[114] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelNeRF: Neural radiance fields from one or few images. In CVPR, pages 4578-4587, 2021. 2 +[115] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. BDD100K: A diverse driving dataset for heterogeneous multitask learning. In CVPR, pages 2633-2642, 2020. 6 +[116] Yuanwen Yue, Anurag Das, Francis Engelmann, Siyu Tang, and Jan Eric Lenssen. Improving 2D feature representations by 3D-aware fine-tuning. In ECCV, pages 57-74, 2024. 2, 6, 7, 8 +[117] Pingping Zhang, Wei Liu, Yinjie Lei, Huchuan Lu, and Xiaoyun Yang. Cascaded context pyramid for full-resolution 3D semantic scene completion. In ICCV, pages 7800-7809, 2019. 2 +[118] Yunpeng Zhang, Zheng Zhu, and Dalong Du. OccFormer: Dual-path transformer for vision-based 3D semantic occupancy prediction. In ICCV, pages 9433-9443, 2023. 2 +[119] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snively. Stereo magnification: Learning view synthesis using multiplane images. ACM Trans. Graph., 37 (4):65, 2018. 6 +[120] Onur Özyeşil, Vladislav Voroninski, Ronen Basri, and Amit Singer. A survey of structure from motion. Acta Numer., 26:305-364, 2017. 2 \ No newline at end of file diff --git a/2025/Feed-Forward SceneDINO for Unsupervised Semantic Scene Completion/images.zip b/2025/Feed-Forward SceneDINO for Unsupervised Semantic Scene Completion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e970e0363ec4c2da6b0a4ed73fa9ebd549f4ae7c --- /dev/null +++ b/2025/Feed-Forward SceneDINO for Unsupervised Semantic Scene Completion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:713eefa8a06f5ecbf9843c7e1dd2b1a00d6f746fcfb6fc3b09bd5a04f5dadb96 +size 554259 diff --git a/2025/Feed-Forward SceneDINO for Unsupervised Semantic Scene Completion/layout.json b/2025/Feed-Forward SceneDINO for Unsupervised Semantic Scene Completion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5dcc8a66c9b3f647752676c740cb3a0045a49334 --- /dev/null +++ b/2025/Feed-Forward SceneDINO for Unsupervised Semantic Scene Completion/layout.json @@ -0,0 +1,14324 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 80, + 103, + 530, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 103, + 530, + 121 + ], + "spans": [ + { + "bbox": [ + 80, + 103, + 530, + 121 + ], + "type": "text", + "content": "Feed-Forward SceneDINO for Unsupervised Semantic Scene Completion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 142, + 231, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 142, + 231, + 157 + ], + "spans": [ + { + "bbox": [ + 133, + 142, + 231, + 157 + ], + "type": "text", + "content": "Aleksandar Jevtic*1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 99, + 157, + 167, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 157, + 167, + 171 + ], + "spans": [ + { + "bbox": [ + 99, + 157, + 167, + 171 + ], + "type": "text", + "content": "Oliver Hahn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 188, + 157, + 293, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 157, + 293, + 172 + ], + "spans": [ + { + "bbox": [ + 188, + 157, + 293, + 172 + ], + "type": "text", + "content": "Christian Rupprecht" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 83, + 173, + 132, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 173, + 132, + 185 + ], + "spans": [ + { + "bbox": [ + 83, + 173, + 132, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 83, + 173, + 132, + 185 + ], + "type": "text", + "content": "TU Munich" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 145, + 174, + 203, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 174, + 203, + 185 + ], + "spans": [ + { + "bbox": [ + 145, + 174, + 203, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 145, + 174, + 203, + 185 + ], + "type": "text", + "content": "TU Darmstadt" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 216, + 174, + 298, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 174, + 298, + 186 + ], + "spans": [ + { + "bbox": [ + 216, + 174, + 298, + 186 + ], + "type": "text", + "content": "3University of Oxford" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 253, + 142, + 362, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 142, + 362, + 157 + ], + "spans": [ + { + "bbox": [ + 253, + 142, + 362, + 157 + ], + "type": "text", + "content": "Christoph Reich* 1,2,4,5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 142, + 393, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 142, + 393, + 171 + ], + "spans": [ + { + "bbox": [ + 315, + 142, + 393, + 171 + ], + "type": "text", + "content": "Eich\\*1,2,4,5 F6 Stefan Roth 2,5,6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 384, + 143, + 477, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 143, + 477, + 156 + ], + "spans": [ + { + "bbox": [ + 384, + 143, + 477, + 156 + ], + "type": "text", + "content": "Felix Wimbauer1,4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 173, + 388, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 173, + 388, + 186 + ], + "spans": [ + { + "bbox": [ + 313, + 173, + 388, + 186 + ], + "type": "text", + "content": "4MCML 5ELIZA" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 415, + 157, + 511, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 157, + 511, + 171 + ], + "spans": [ + { + "bbox": [ + 415, + 157, + 511, + 171 + ], + "type": "text", + "content": "Daniel Cremers " + }, + { + "bbox": [ + 415, + 157, + 511, + 171 + ], + "type": "inline_equation", + "content": "^{1,4,5}" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 400, + 174, + 527, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 174, + 527, + 186 + ], + "spans": [ + { + "bbox": [ + 400, + 174, + 527, + 186 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 400, + 174, + 527, + 186 + ], + "type": "text", + "content": "hessian.AI *equal contribution" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 223, + 187, + 386, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 187, + 386, + 198 + ], + "spans": [ + { + "bbox": [ + 223, + 187, + 386, + 198 + ], + "type": "text", + "content": "https://visinf.github.io/scenedino" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 59, + 217, + 184, + 258 + ], + "blocks": [ + { + "bbox": [ + 59, + 217, + 184, + 258 + ], + "lines": [ + { + "bbox": [ + 59, + 217, + 184, + 258 + ], + "spans": [ + { + "bbox": [ + 59, + 217, + 184, + 258 + ], + "type": "image", + "image_path": "262c286345295238ebd432bdf2ecccf1e0df6b5da244a4e6a3c87aefed1f9aa2.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 270, + 167, + 283 + ], + "lines": [ + { + "bbox": [ + 75, + 270, + 167, + 283 + ], + "spans": [ + { + "bbox": [ + 75, + 270, + 167, + 283 + ], + "type": "text", + "content": "Single Input Image" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 187, + 234, + 237, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 234, + 237, + 243 + ], + "spans": [ + { + "bbox": [ + 187, + 234, + 237, + 243 + ], + "type": "text", + "content": "SceneDINO" + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 251, + 209, + 367, + 269 + ], + "blocks": [ + { + "bbox": [ + 251, + 209, + 367, + 269 + ], + "lines": [ + { + "bbox": [ + 251, + 209, + 367, + 269 + ], + "spans": [ + { + "bbox": [ + 251, + 209, + 367, + 269 + ], + "type": "image", + "image_path": "bcd377efa08022c7bda5c5b9514b60b4f3f97d03895267ac41ed2035f74e99a0.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 269, + 270, + 349, + 281 + ], + "lines": [ + { + "bbox": [ + 269, + 270, + 349, + 281 + ], + "spans": [ + { + "bbox": [ + 269, + 270, + 349, + 281 + ], + "type": "text", + "content": "3D Feature Field" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 288, + 555, + 322 + ], + "lines": [ + { + "bbox": [ + 54, + 288, + 555, + 322 + ], + "spans": [ + { + "bbox": [ + 54, + 288, + 555, + 322 + ], + "type": "text", + "content": "Figure 1. SceneDINO overview. Given a single input image (left), SceneDINO predicts both 3D scene geometry and 3D features in the form of a feature field (middle) in a feed-forward manner, capturing the structure and semantics of the scene. Unsupervised distillation and clustering of SceneDINO's feature space leads to unsupervised semantic scene completion predictions (right)." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 368, + 234, + 429, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 234, + 429, + 243 + ], + "spans": [ + { + "bbox": [ + 368, + 234, + 429, + 243 + ], + "type": "text", + "content": "Distill & Cluster" + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 435, + 209, + 552, + 269 + ], + "blocks": [ + { + "bbox": [ + 435, + 209, + 552, + 269 + ], + "lines": [ + { + "bbox": [ + 435, + 209, + 552, + 269 + ], + "spans": [ + { + "bbox": [ + 435, + 209, + 552, + 269 + ], + "type": "image", + "image_path": "bf9fd12e6841fe47ba904bc33b9a2eb12f5ac84965bf06620b42ef234550e8e4.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 456, + 270, + 530, + 281 + ], + "lines": [ + { + "bbox": [ + 456, + 270, + 530, + 281 + ], + "spans": [ + { + "bbox": [ + 456, + 270, + 530, + 281 + ], + "type": "text", + "content": "SSC Prediction" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 151, + 339, + 200, + 352 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 339, + 200, + 352 + ], + "spans": [ + { + "bbox": [ + 151, + 339, + 200, + 352 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 54, + 365, + 297, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 365, + 297, + 606 + ], + "spans": [ + { + "bbox": [ + 54, + 365, + 297, + 606 + ], + "type": "text", + "content": "Semantic scene completion (SSC) aims to infer both the 3D geometry and semantics of a scene from single images. In contrast to prior work on SSC that heavily relies on expensive ground-truth annotations, we approach SSC in an unsupervised setting. Our novel method, SceneDINO, adapts techniques from self-supervised representation learning and 2D unsupervised scene understanding to SSC. Our training exclusively utilizes multi-view consistency self-supervision without any form of semantic or geometric ground truth. Given a single input image, SceneDINO infers the 3D geometry and expressive 3D DINO features in a feed-forward manner. Through a novel 3D feature distillation approach, we obtain unsupervised 3D semantics. In both 3D and 2D unsupervised scene understanding, SceneDINO reaches state-of-the-art segmentation accuracy. Linear probing our 3D features matches the segmentation accuracy of a current supervised SSC approach. Additionally, we showcase the domain generalization and multi-view consistency of SceneDINO, taking the first steps towards a strong foundation for single image 3D scene understanding." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 55, + 632, + 135, + 644 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 632, + 135, + 644 + ], + "spans": [ + { + "bbox": [ + 55, + 632, + 135, + 644 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "Understanding the geometry and semantics of 3D scenes from image observations is a fundamental computer vision task with broad applications in robotics [26], autonomous driving [47, 66], medical image analysis [18, 113], and civil engineering [70]. The Semantic Scene Completion (SSC)" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 340, + 555, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 555, + 448 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 555, + 448 + ], + "type": "text", + "content": "task unifies 3D geometry and semantic prediction from limited image observations [64, 89, 96]. Recent progress in SSC has been primarily driven by utilizing supervised learning [38, 88, 96]. However, acquiring large-scale 3D annotations is highly labor-intensive [66]. While significant resources have been invested in collecting human annotations for 2D tasks [53, 85], annotating similar amounts of data in 3D remains unapproached. This motivates approaching SSC without the need for manually annotated data." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 449, + 556, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 556, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 556, + 676 + ], + "type": "text", + "content": "Existing SSC approaches rely on ground-truth semantic annotations and frequently utilize additional supervision from LiDAR scans [38, 46, 74, 96]. In contrast, we are the first to approach SSC in a fully unsupervised setting, i.e. without task supervision or other supervised components. In particular, we aim to approach SSC from a single image without relying on any human annotations, only learning from unlabeled multi-view images using self-supervision. This setting is extremely challenging for two reasons: first, the human-defined nature of semantic taxonomies is ambiguous, and second, a single image only provides a partial observation of the scene with many invisible areas. We take inspiration from recent advances in self-supervised learning (SSL) of 2D representations and 3D reconstruction. 2D SSL representations, such as from DINO [11], have been shown effective for 2D unsupervised scene understanding [32, 33, 104]. 3D reconstruction approaches successfully leveraged SSL from multi-view data to infer dense 3D geometry from a single image [34, 108]." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": "In this paper, we present SceneDINO, to the best of our knowledge, the first approach for unsupervised semantic scene completion. Trained using 2D SSL features" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6784" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "type": "text", + "content": "from DINO [11] and multi-view self-supervision [108], SceneDINO predicts both 3D geometry and 3D features from a single image during inference in a feed-forward manner. Our general 3D feature representations enable us to approach unsupervised 3D scene understanding. Harnessing our expressive 3D features, we propose a novel 3D feature distillation approach for obtaining unsupervised semantic predictions in 3D. While we focus on the task of unsupervised SSC, SceneDINO's features are general, offering a foundation for different 3D scene-understanding tasks by building on our 3D feature field." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 205, + 294, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 205, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 54, + 205, + 294, + 360 + ], + "type": "text", + "content": "Specifically, we make the following contributions: (i) We introduce SceneDINO, the first approach predicting dense 3D geometry and expressive 3D features in a feedforward manner from a single image. (ii) We effectively distill SceneDINO's feature field representation in 3D, obtaining unsupervised semantic predictions. (iii) We demonstrate the first fully unsupervised SSC results. We build a simple, yet competitive unsupervised SSC baseline, lifting unsupervised 2D semantic predictions. Our SceneDINO approach outperforms this SSC baseline in unsupervised SSC as well as established 2D approaches in 2D semantic segmentation. (iv) Finally, we also showcase the domain generalization ability and multi-view consistency of SceneDINO." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 373, + 142, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 142, + 385 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 142, + 385 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 398, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 398, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 398, + 295, + 601 + ], + "type": "text", + "content": "Single-image scene reconstruction. Estimating 3D geometry from image observations is a fundamental task in computer vision and has been studied for decades [37]. Traditional approaches, such as structure from motion [90], as well as recent neural radiance fields (NeRFs) [75], perform scene reconstruction using multiple images, as reviewed by multiple surveys [35, 109, 120]. Recently, estimating dense 3D geometry from a single image have been approached [8, 34, 81, 86, 97, 103, 108, 114]. Unlike monocular depth estimation [76], these approaches predict the depth for visible and occluded regions, reconstructing a complete scene. Behind the Scenes (BTS) [108] introduced an approach for unsupervised single-image scene reconstruction using multi-view self-supervision, which infers dense 3D geometry in a feed-forward manner. Our approach extends BTS by additionally lifting self-supervised features into 3D for unsupervised 3D scene understanding." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": "Semantic scene completion (SSC), also known as 3D semantic occupancy prediction, aims to jointly estimate the 3D geometry and semantics of a scene [63, 64, 96, 118]. Initial approaches used 3D semantic and geometric annotations and addressed indoor scenes [6, 13, 58-60, 68, 117], outdoor scenes with LiDAR [16, 62, 87, 88, 110], or both domains [8, 74]. Using birds-eye views has been proven effective for SSC [45, 65, 100]. To overcome the need for 3D annotations, approaches for using 2D annotations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "content": "have been proposed [38, 46, 82]. While SelfOcc [46] and RenderOcc [82] use multiple inference views, S4C [38] performs single-image SSC. In particular, S4C [38] employs a supervised 2D model and lifts 2D multi-view semantic predictions into 3D. In contrast to using 2D annotations, GaussTR [49] uses 2D foundation models for SSC and multiple views during inference. However, GaussTR relies on heavily supervised foundation models, including SAM [53] and Metric3Dv2 [43], and uses weak supervision from image/text pairs. To the best of our knowledge, there is no method for approaching SSC without the need for any ground-truth annotations. Our work presents the first unsupervised SSC approach, utilizing lifted SSL features and a single RGB input image for inference." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 242, + 553, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 242, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 313, + 242, + 553, + 386 + ], + "type": "text", + "content": "Self-supervised representation learning (SSL) aims to extract general features from data without annotations, facilitating various downstream tasks such as segmentation [24]. Recent SSL methods, often based on Vision Transformers (ViTs) [23], leverage clustering [2, 9, 10, 48, 61], masked modeling [20, 29, 40, 77, 107], contrastive learning [3, 12, 14, 39, 41, 42], or negative-free [4, 5, 11, 28, 80] pretext tasks [22, 79] for large-scale training. State-of-the-art models, e.g., DINO [11], produce semantically rich, dense features, driving recent advances in 2D unsupervised scene understanding [33, 104]. We here aim to bring expressive features from DINO [11, 80] to 3D for SSC." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 388, + 553, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 553, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 553, + 616 + ], + "type": "text", + "content": "2D-to-3D feature lifting. The expressiveness of 2D visual representations has motivated lifting 2D features into 3D [94, 111]. Existing approaches utilize multi-view 2D features for 3D feature lifting [30, 44, 50, 54, 73, 83, 93, 94, 98, 101, 102, 106, 111, 116]. Lifting 2D features is effective in various tasks, including few-shot semantic occupancy prediction [111], and refining 2D representations [116]. However, existing feature-lifting approaches fit to a single scene [50, 54, 93, 94, 101, 102, 111, 116], require RGB-D inputs [30, 44, 73, 98, 106], or work on 3D point cloud inputs [83]. The only feed-forward approaches that use RGB inputs and lift 2D features, which we are aware of, are GaussTR [49] and MVSplat360 [15]. However, both approaches utilize multiple input images during inference, and MVSplat360 [15] only predicts low-dimensional feature representations, which are not suitable for unsupervised scene understanding. In contrast, we propose the first feedforward approach for inferring lifted high-dimensional and rich 3D features using a single input image." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "text", + "content": "2D unsupervised semantic segmentation partitions images automatically into semantically meaningful regions without any form of human annotations. Early deep learning-based methods [17, 36, 48] approach the problem via representation learning. Leveraging SSL features from DINO as an inductive prior, STEGO [33] distills the feature representation into a lower-dimensional space for unsupervised probing. Building up on STEGO, subsequent methods" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6785" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 72, + 212, + 202 + ], + "blocks": [ + { + "bbox": [ + 65, + 72, + 212, + 202 + ], + "lines": [ + { + "bbox": [ + 65, + 72, + 212, + 202 + ], + "spans": [ + { + "bbox": [ + 65, + 72, + 212, + 202 + ], + "type": "image", + "image_path": "7a32d6794c76d05a04ce2206d89ab5f3c7736bf7547632ad0614e5dff0fc093d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 205, + 555, + 273 + ], + "lines": [ + { + "bbox": [ + 54, + 205, + 555, + 273 + ], + "spans": [ + { + "bbox": [ + 54, + 205, + 555, + 273 + ], + "type": "text", + "content": "Figure 2. SceneDINO architecture, rendering, and training. (a) Inference: Given a single input image " + }, + { + "bbox": [ + 54, + 205, + 555, + 273 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_0" + }, + { + "bbox": [ + 54, + 205, + 555, + 273 + ], + "type": "text", + "content": " during inference, a 2D encoder-decoder " + }, + { + "bbox": [ + 54, + 205, + 555, + 273 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 54, + 205, + 555, + 273 + ], + "type": "text", + "content": " produces the embedding " + }, + { + "bbox": [ + 54, + 205, + 555, + 273 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 54, + 205, + 555, + 273 + ], + "type": "text", + "content": " from which the local embedding " + }, + { + "bbox": [ + 54, + 205, + 555, + 273 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_{\\mathbf{u}}" + }, + { + "bbox": [ + 54, + 205, + 555, + 273 + ], + "type": "text", + "content": " is interpolated. The MLP encoder " + }, + { + "bbox": [ + 54, + 205, + 555, + 273 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 54, + 205, + 555, + 273 + ], + "type": "text", + "content": " takes in " + }, + { + "bbox": [ + 54, + 205, + 555, + 273 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_{\\mathbf{u}}" + }, + { + "bbox": [ + 54, + 205, + 555, + 273 + ], + "type": "text", + "content": " and 3D position " + }, + { + "bbox": [ + 54, + 205, + 555, + 273 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 54, + 205, + 555, + 273 + ], + "type": "text", + "content": ", and predicts both the density " + }, + { + "bbox": [ + 54, + 205, + 555, + 273 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathbf{x}_i}" + }, + { + "bbox": [ + 54, + 205, + 555, + 273 + ], + "type": "text", + "content": " and the 3D feature " + }, + { + "bbox": [ + 54, + 205, + 555, + 273 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{x}_i}" + }, + { + "bbox": [ + 54, + 205, + 555, + 273 + ], + "type": "text", + "content": ". Using a lightweight unsupervised segmentation head " + }, + { + "bbox": [ + 54, + 205, + 555, + 273 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 54, + 205, + 555, + 273 + ], + "type": "text", + "content": ", we can obtain semantic predictions " + }, + { + "bbox": [ + 54, + 205, + 555, + 273 + ], + "type": "inline_equation", + "content": "p_{\\mathbf{x}_i}" + }, + { + "bbox": [ + 54, + 205, + 555, + 273 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 54, + 205, + 555, + 273 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{x}_i}" + }, + { + "bbox": [ + 54, + 205, + 555, + 273 + ], + "type": "text", + "content": ". (b) Rendering: Our feature field allows for volume rendering by shooting rays through it, yielding depth " + }, + { + "bbox": [ + 54, + 205, + 555, + 273 + ], + "type": "inline_equation", + "content": "\\hat{d}" + }, + { + "bbox": [ + 54, + 205, + 555, + 273 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 205, + 555, + 273 + ], + "type": "inline_equation", + "content": "\\hat{f}" + }, + { + "bbox": [ + 54, + 205, + 555, + 273 + ], + "type": "text", + "content": " in 2D. Color " + }, + { + "bbox": [ + 54, + 205, + 555, + 273 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 54, + 205, + 555, + 273 + ], + "type": "text", + "content": " is sampled from an another view (" + }, + { + "bbox": [ + 54, + 205, + 555, + 273 + ], + "type": "inline_equation", + "content": "e.g., \\mathbf{I}_1" + }, + { + "bbox": [ + 54, + 205, + 555, + 273 + ], + "type": "text", + "content": ") using " + }, + { + "bbox": [ + 54, + 205, + 555, + 273 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_s" + }, + { + "bbox": [ + 54, + 205, + 555, + 273 + ], + "type": "text", + "content": " and rendered to obtain the reconstructed color " + }, + { + "bbox": [ + 54, + 205, + 555, + 273 + ], + "type": "inline_equation", + "content": "\\hat{c}" + }, + { + "bbox": [ + 54, + 205, + 555, + 273 + ], + "type": "text", + "content": ". (c) Multi-view training: We render 2D views (features & images) from our feature field and reconstruct the training views." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 236, + 81, + 373, + 202 + ], + "blocks": [ + { + "bbox": [ + 236, + 81, + 373, + 202 + ], + "lines": [ + { + "bbox": [ + 236, + 81, + 373, + 202 + ], + "spans": [ + { + "bbox": [ + 236, + 81, + 373, + 202 + ], + "type": "image", + "image_path": "30dd532230d021ee8257fae00817fed36b4bbce0d48b19a7f7209dd2246ec836.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 398, + 70, + 541, + 202 + ], + "blocks": [ + { + "bbox": [ + 398, + 70, + 541, + 202 + ], + "lines": [ + { + "bbox": [ + 398, + 70, + 541, + 202 + ], + "spans": [ + { + "bbox": [ + 398, + 70, + 541, + 202 + ], + "type": "image", + "image_path": "6072acde5b5bbd7b0ebfd347875ac119c96457bbc4e47eaa0ac16f0ad1b0fef0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 288, + 294, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 288, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 54, + 288, + 294, + 337 + ], + "type": "text", + "content": "[31, 51, 92, 95] propose enhancements to the distillation. Our approach follows the idea of STEGO [33], extending it to 3D and integrating feature distillation using our 3D feature field to build the first unsupervised SSC approach." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 347, + 285, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 347, + 285, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 285, + 361 + ], + "type": "text", + "content": "3. Unsupervised Semantic Scene Completion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 368, + 295, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 368, + 295, + 476 + ], + "spans": [ + { + "bbox": [ + 55, + 368, + 295, + 476 + ], + "type": "text", + "content": "We approach semantic scene completion (SSC) without any form of manual supervision. To this end, we first describe SceneDINO, predicting 3D geometry and expressive 3D features from a single image in a feed-forward manner (Sec. 3.1), and SceneDINO's multi-view training (Sec. 3.2). Next, we present our 3D feature distillation approach to obtain unsupervised 3D semantic predictions (Sec. 3.3). An overview of our full pipeline, including inference, rendering, and multi-view self-supervision, is provided in Fig. 2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 477, + 295, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 477, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 477, + 295, + 577 + ], + "type": "text", + "content": "Notation. Let " + }, + { + "bbox": [ + 55, + 477, + 295, + 577 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_0\\in [0,1]^{3\\times \\mathrm{H}\\times \\mathrm{W}}" + }, + { + "bbox": [ + 55, + 477, + 295, + 577 + ], + "type": "text", + "content": " be a single RGB input image (for both training & inference) with corresponding pose " + }, + { + "bbox": [ + 55, + 477, + 295, + 577 + ], + "type": "inline_equation", + "content": "T_0\\in \\mathbb{R}^{4\\times 4}" + }, + { + "bbox": [ + 55, + 477, + 295, + 577 + ], + "type": "text", + "content": " and projection matrix " + }, + { + "bbox": [ + 55, + 477, + 295, + 577 + ], + "type": "inline_equation", + "content": "K_{0}\\in \\mathbb{R}^{3\\times 4}" + }, + { + "bbox": [ + 55, + 477, + 295, + 577 + ], + "type": "text", + "content": ". For training, let " + }, + { + "bbox": [ + 55, + 477, + 295, + 577 + ], + "type": "inline_equation", + "content": "(\\mathbf{I}_v,T_v,K_v)" + }, + { + "bbox": [ + 55, + 477, + 295, + 577 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 477, + 295, + 577 + ], + "type": "inline_equation", + "content": "v\\in \\{1,2,\\dots ,n\\}" + }, + { + "bbox": [ + 55, + 477, + 295, + 577 + ], + "type": "text", + "content": ", be " + }, + { + "bbox": [ + 55, + 477, + 295, + 577 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 477, + 295, + 577 + ], + "type": "text", + "content": " additional views for multi-view self-supervision. Assuming a pinhole camera model, any 3D point " + }, + { + "bbox": [ + 55, + 477, + 295, + 577 + ], + "type": "inline_equation", + "content": "\\mathbf{x}\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 477, + 295, + 577 + ], + "type": "text", + "content": " in world coordinates can be projected onto the image plane of view " + }, + { + "bbox": [ + 55, + 477, + 295, + 577 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 477, + 295, + 577 + ], + "type": "text", + "content": " and the input view " + }, + { + "bbox": [ + 55, + 477, + 295, + 577 + ], + "type": "inline_equation", + "content": "v = 0" + }, + { + "bbox": [ + 55, + 477, + 295, + 577 + ], + "type": "text", + "content": " with the perspective projection " + }, + { + "bbox": [ + 55, + 477, + 295, + 577 + ], + "type": "inline_equation", + "content": "\\pi_v(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 477, + 295, + 577 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 584, + 134, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 584, + 134, + 596 + ], + "spans": [ + { + "bbox": [ + 55, + 584, + 134, + 596 + ], + "type": "text", + "content": "3.1. SceneDINO" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 601, + 295, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 601, + 295, + 686 + ], + "spans": [ + { + "bbox": [ + 54, + 601, + 295, + 686 + ], + "type": "text", + "content": "Given a single input image " + }, + { + "bbox": [ + 54, + 601, + 295, + 686 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_0" + }, + { + "bbox": [ + 54, + 601, + 295, + 686 + ], + "type": "text", + "content": ", SceneDINO represents the dense geometric structure and features of a scene as a continuous mapping from world coordinates " + }, + { + "bbox": [ + 54, + 601, + 295, + 686 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 54, + 601, + 295, + 686 + ], + "type": "text", + "content": " to a volumetric density " + }, + { + "bbox": [ + 54, + 601, + 295, + 686 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathbf{x}} \\in \\mathbb{R}_{+}" + }, + { + "bbox": [ + 54, + 601, + 295, + 686 + ], + "type": "text", + "content": " and a feature " + }, + { + "bbox": [ + 54, + 601, + 295, + 686 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{x}} \\in \\mathbb{R}^{\\mathrm{D}}" + }, + { + "bbox": [ + 54, + 601, + 295, + 686 + ], + "type": "text", + "content": ". This continuous output representation is often called a feature field. While SceneDINO could represent any feature space, we aim for expressive SSL features from DINO [11, 80]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Architecture & feature field inference. Our SceneDINO architecture comprises two main parts: a 2D encoder-" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 288, + 555, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 555, + 347 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 555, + 347 + ], + "type": "text", + "content": "decoder " + }, + { + "bbox": [ + 313, + 288, + 555, + 347 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 313, + 288, + 555, + 347 + ], + "type": "text", + "content": " and an MLP decoder (cf. Fig. 2a), following BTS [108]. " + }, + { + "bbox": [ + 313, + 288, + 555, + 347 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 313, + 288, + 555, + 347 + ], + "type": "text", + "content": " takes in " + }, + { + "bbox": [ + 313, + 288, + 555, + 347 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_0" + }, + { + "bbox": [ + 313, + 288, + 555, + 347 + ], + "type": "text", + "content": " and produces a per-pixel embedding " + }, + { + "bbox": [ + 313, + 288, + 555, + 347 + ], + "type": "inline_equation", + "content": "\\mathbf{E} \\in \\mathbb{R}^{\\mathrm{D}_{\\mathbf{E}} \\times \\mathrm{H} \\times \\mathrm{W}}" + }, + { + "bbox": [ + 313, + 288, + 555, + 347 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 288, + 555, + 347 + ], + "type": "inline_equation", + "content": "\\mathrm{D}_{\\mathbf{E}}" + }, + { + "bbox": [ + 313, + 288, + 555, + 347 + ], + "type": "text", + "content": " dimensions. Intuitively, every spatial element of " + }, + { + "bbox": [ + 313, + 288, + 555, + 347 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 313, + 288, + 555, + 347 + ], + "type": "text", + "content": " represents a camera ray through a pixel, capturing both local geometry and features." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 348, + 555, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 555, + 444 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 555, + 444 + ], + "type": "text", + "content": "To infer the feature at a 3D position " + }, + { + "bbox": [ + 313, + 348, + 555, + 444 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 348, + 555, + 444 + ], + "type": "text", + "content": ", we employ a two-layer MLP decoder " + }, + { + "bbox": [ + 313, + 348, + 555, + 444 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 348, + 555, + 444 + ], + "type": "text", + "content": " (cf. Fig. 2a). Given a position " + }, + { + "bbox": [ + 313, + 348, + 555, + 444 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 348, + 555, + 444 + ], + "type": "text", + "content": " within the camera frustum, we project " + }, + { + "bbox": [ + 313, + 348, + 555, + 444 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 348, + 555, + 444 + ], + "type": "text", + "content": " into the camera plane, obtaining the pixel location " + }, + { + "bbox": [ + 313, + 348, + 555, + 444 + ], + "type": "inline_equation", + "content": "\\mathbf{u} = \\pi_0(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 348, + 555, + 444 + ], + "type": "text", + "content": ". We query " + }, + { + "bbox": [ + 313, + 348, + 555, + 444 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 313, + 348, + 555, + 444 + ], + "type": "text", + "content": " at the position " + }, + { + "bbox": [ + 313, + 348, + 555, + 444 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 313, + 348, + 555, + 444 + ], + "type": "text", + "content": " using bilinear interpolation, obtaining the local embedding " + }, + { + "bbox": [ + 313, + 348, + 555, + 444 + ], + "type": "inline_equation", + "content": "\\mathbf{e_u}" + }, + { + "bbox": [ + 313, + 348, + 555, + 444 + ], + "type": "text", + "content": ". Based on the embedding " + }, + { + "bbox": [ + 313, + 348, + 555, + 444 + ], + "type": "inline_equation", + "content": "\\mathbf{e_u}" + }, + { + "bbox": [ + 313, + 348, + 555, + 444 + ], + "type": "text", + "content": ", the pixel position " + }, + { + "bbox": [ + 313, + 348, + 555, + 444 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 313, + 348, + 555, + 444 + ], + "type": "text", + "content": ", and the distance " + }, + { + "bbox": [ + 313, + 348, + 555, + 444 + ], + "type": "inline_equation", + "content": "d_{\\mathbf{x}} \\in \\mathbb{R}_{+}" + }, + { + "bbox": [ + 313, + 348, + 555, + 444 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 348, + 555, + 444 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 348, + 555, + 444 + ], + "type": "text", + "content": " to the camera, we obtain the density " + }, + { + "bbox": [ + 313, + 348, + 555, + 444 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathbf{x}}" + }, + { + "bbox": [ + 313, + 348, + 555, + 444 + ], + "type": "text", + "content": " and feature prediction " + }, + { + "bbox": [ + 313, + 348, + 555, + 444 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{x}}" + }, + { + "bbox": [ + 313, + 348, + 555, + 444 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 376, + 448, + 553, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 448, + 553, + 462 + ], + "spans": [ + { + "bbox": [ + 376, + 448, + 553, + 462 + ], + "type": "interline_equation", + "content": "\\left(\\sigma_ {\\mathbf {x}}, f _ {\\mathbf {x}}\\right) = \\phi \\left(\\mathbf {e} _ {\\mathbf {u}}, \\gamma (\\mathbf {u}, d _ {\\mathbf {x}})\\right), \\tag {1}", + "image_path": "e6f0498102ee58ddfecaa3e7d34742c807b9380a438066c10ac89b6381978d34.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 465, + 493, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 465, + 493, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 465, + 493, + 478 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 465, + 493, + 478 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 465, + 493, + 478 + ], + "type": "text", + "content": " denotes a positional encoding [75]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 479, + 555, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 555, + 600 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 555, + 600 + ], + "type": "text", + "content": "Feature, depth & color volume rendering. SceneDINO predicts a continuous feature field from a single image. This representation can be used to render features and depth in 2D from an arbitrary viewpoint (cf. Fig. 2b), following the discretization strategy of Max et al. [72]. Given a viewpoint " + }, + { + "bbox": [ + 313, + 479, + 555, + 600 + ], + "type": "inline_equation", + "content": "(T,K)" + }, + { + "bbox": [ + 313, + 479, + 555, + 600 + ], + "type": "text", + "content": ", we sample L points " + }, + { + "bbox": [ + 313, + 479, + 555, + 600 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 313, + 479, + 555, + 600 + ], + "type": "text", + "content": " along the ray through pixel " + }, + { + "bbox": [ + 313, + 479, + 555, + 600 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_r" + }, + { + "bbox": [ + 313, + 479, + 555, + 600 + ], + "type": "text", + "content": ", with distance " + }, + { + "bbox": [ + 313, + 479, + 555, + 600 + ], + "type": "inline_equation", + "content": "\\delta_i" + }, + { + "bbox": [ + 313, + 479, + 555, + 600 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 313, + 479, + 555, + 600 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 313, + 479, + 555, + 600 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 479, + 555, + 600 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i+1}" + }, + { + "bbox": [ + 313, + 479, + 555, + 600 + ], + "type": "text", + "content": ". Based on the volumetric densities " + }, + { + "bbox": [ + 313, + 479, + 555, + 600 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathbf{x}_i}" + }, + { + "bbox": [ + 313, + 479, + 555, + 600 + ], + "type": "text", + "content": " (cf. Eq. 1), we can compute the probabilities " + }, + { + "bbox": [ + 313, + 479, + 555, + 600 + ], + "type": "inline_equation", + "content": "\\alpha_i" + }, + { + "bbox": [ + 313, + 479, + 555, + 600 + ], + "type": "text", + "content": " of the ray ending between " + }, + { + "bbox": [ + 313, + 479, + 555, + 600 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 313, + 479, + 555, + 600 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 479, + 555, + 600 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i+1}" + }, + { + "bbox": [ + 313, + 479, + 555, + 600 + ], + "type": "text", + "content": ", and accumulate these into " + }, + { + "bbox": [ + 313, + 479, + 555, + 600 + ], + "type": "inline_equation", + "content": "V_i" + }, + { + "bbox": [ + 313, + 479, + 555, + 600 + ], + "type": "text", + "content": ", the probability of " + }, + { + "bbox": [ + 313, + 479, + 555, + 600 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 313, + 479, + 555, + 600 + ], + "type": "text", + "content": " being visible:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 323, + 604, + 553, + 638 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 604, + 553, + 638 + ], + "spans": [ + { + "bbox": [ + 323, + 604, + 553, + 638 + ], + "type": "interline_equation", + "content": "V _ {i} = \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\quad \\text {w i t h} \\alpha_ {i} = 1 - \\exp (- \\sigma_ {\\mathbf {x} _ {i}} \\delta_ {i}). \\tag {2}", + "image_path": "42fead23f4dae7a301827b5db81458e667e93f11f6ff68404feb0c1e95100faf.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 642, + 555, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 680 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 680 + ], + "type": "text", + "content": "Using " + }, + { + "bbox": [ + 313, + 642, + 555, + 680 + ], + "type": "inline_equation", + "content": "V_{i}" + }, + { + "bbox": [ + 313, + 642, + 555, + 680 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 642, + 555, + 680 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 313, + 642, + 555, + 680 + ], + "type": "text", + "content": ", we render depth " + }, + { + "bbox": [ + 313, + 642, + 555, + 680 + ], + "type": "inline_equation", + "content": "\\tilde{d}_{\\mathbf{u}_r}" + }, + { + "bbox": [ + 313, + 642, + 555, + 680 + ], + "type": "text", + "content": " and feature " + }, + { + "bbox": [ + 313, + 642, + 555, + 680 + ], + "type": "inline_equation", + "content": "\\tilde{f}_{\\mathbf{u}_r}" + }, + { + "bbox": [ + 313, + 642, + 555, + 680 + ], + "type": "text", + "content": " from the estimated features " + }, + { + "bbox": [ + 313, + 642, + 555, + 680 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{x}_i}" + }, + { + "bbox": [ + 313, + 642, + 555, + 680 + ], + "type": "text", + "content": " from Eq. (1) and distances " + }, + { + "bbox": [ + 313, + 642, + 555, + 680 + ], + "type": "inline_equation", + "content": "d_{\\mathbf{x}_i}" + }, + { + "bbox": [ + 313, + 642, + 555, + 680 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 642, + 555, + 680 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 313, + 642, + 555, + 680 + ], + "type": "text", + "content": " onto the image plane at position " + }, + { + "bbox": [ + 313, + 642, + 555, + 680 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_r" + }, + { + "bbox": [ + 313, + 642, + 555, + 680 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 344, + 683, + 555, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 683, + 555, + 716 + ], + "spans": [ + { + "bbox": [ + 344, + 683, + 555, + 716 + ], + "type": "interline_equation", + "content": "\\tilde {f} _ {\\mathbf {u} _ {r}} = \\sum_ {i = 1} ^ {\\mathrm {L}} V _ {i} \\alpha_ {i} f _ {\\mathbf {x} _ {i}} \\quad \\tilde {d} _ {\\mathbf {u} _ {r}} = \\sum_ {i = 1} ^ {\\mathrm {L}} V _ {i} \\alpha_ {i} d _ {\\mathbf {x} _ {i}}. \\tag {3}", + "image_path": "fd3fbd9e45bdd9e2905fbbb8a7b3a9fdba99e59ea5ba1ce52b2bb2bccf07684c.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6786" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": "The differentiability of this rendering process enables us to self-supervise SceneDINO using multi-view images and their 2D feature representations (e.g., from DINO [11]). SceneDINO predicts 3D geometry and features, but does not predict color as we focus on semantic downstream tasks. To obtain color for image reconstruction during training, we follow the color sampling approach of BTS [108]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 163, + 192, + 176 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 163, + 192, + 176 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 192, + 176 + ], + "type": "text", + "content": "3.2. 3D feature field training" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 180, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 180, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 296, + 361 + ], + "type": "text", + "content": "We train SceneDINO using multi-view self-supervision (cf. Fig. 2c), aiming to obtain an expressive and view-consistent feature field without the need for any form of manual annotations. For self-supervision, we sample " + }, + { + "bbox": [ + 55, + 180, + 296, + 361 + ], + "type": "inline_equation", + "content": "n + 1" + }, + { + "bbox": [ + 55, + 180, + 296, + 361 + ], + "type": "text", + "content": " views " + }, + { + "bbox": [ + 55, + 180, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_v" + }, + { + "bbox": [ + 55, + 180, + 296, + 361 + ], + "type": "text", + "content": " with camera parameters " + }, + { + "bbox": [ + 55, + 180, + 296, + 361 + ], + "type": "inline_equation", + "content": "T_v, K_v" + }, + { + "bbox": [ + 55, + 180, + 296, + 361 + ], + "type": "text", + "content": " from the data and obtain dense 2D target features " + }, + { + "bbox": [ + 55, + 180, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_v" + }, + { + "bbox": [ + 55, + 180, + 296, + 361 + ], + "type": "text", + "content": " from a self-supervised ViT (e.g., DINO [11]). Note that the 2D features entail a resolution of " + }, + { + "bbox": [ + 55, + 180, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_v \\in \\mathbb{R}^{\\mathrm{D} \\times \\frac{\\mathrm{H}}{p} \\times \\frac{\\mathrm{W}}{p}}" + }, + { + "bbox": [ + 55, + 180, + 296, + 361 + ], + "type": "text", + "content": ", due to the ViT patch size " + }, + { + "bbox": [ + 55, + 180, + 296, + 361 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 180, + 296, + 361 + ], + "type": "text", + "content": ". The set of training views and features " + }, + { + "bbox": [ + 55, + 180, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\mathbb{V} = \\{( \\mathbf{I}_v, T_v, K_v, \\mathbf{F}_v) | v = 0, \\dots, n \\}" + }, + { + "bbox": [ + 55, + 180, + 296, + 361 + ], + "type": "text", + "content": " is randomly partitioned into two subsets " + }, + { + "bbox": [ + 55, + 180, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\mathbb{V}_{\\text{source}}" + }, + { + "bbox": [ + 55, + 180, + 296, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 180, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\mathbb{V}_{\\text{target}}" + }, + { + "bbox": [ + 55, + 180, + 296, + 361 + ], + "type": "text", + "content": ". Training reconstructs the views " + }, + { + "bbox": [ + 55, + 180, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\mathbb{V}_{\\text{target}}" + }, + { + "bbox": [ + 55, + 180, + 296, + 361 + ], + "type": "text", + "content": " using the views of " + }, + { + "bbox": [ + 55, + 180, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\mathbb{V}_{\\text{source}}" + }, + { + "bbox": [ + 55, + 180, + 296, + 361 + ], + "type": "text", + "content": ". In practice, we use a randomly sampled set of image patches that align with the ViT patches instead of the full image. In the following, we still refer to images and the full image resolution for the sake of brevity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 364, + 296, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 364, + 296, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 364, + 296, + 450 + ], + "type": "text", + "content": "Image reconstruction. We aim to learn the geometry of our feature field via multi-view photometric consistency. In particular, for every image " + }, + { + "bbox": [ + 55, + 364, + 296, + 450 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_t\\in \\mathbb{V}_{\\mathrm{target}}" + }, + { + "bbox": [ + 55, + 364, + 296, + 450 + ], + "type": "text", + "content": " we derive a reconstructed image " + }, + { + "bbox": [ + 55, + 364, + 296, + 450 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}_{t,s}" + }, + { + "bbox": [ + 55, + 364, + 296, + 450 + ], + "type": "text", + "content": " from every view " + }, + { + "bbox": [ + 55, + 364, + 296, + 450 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 364, + 296, + 450 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 364, + 296, + 450 + ], + "type": "inline_equation", + "content": "\\mathbb{V}_{\\mathrm{source}}" + }, + { + "bbox": [ + 55, + 364, + 296, + 450 + ], + "type": "text", + "content": " using differentiable rendering (cf. Eq. 3) and color sampling [108]. Equipped with both the reconstructed image " + }, + { + "bbox": [ + 55, + 364, + 296, + 450 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}_{t,s}" + }, + { + "bbox": [ + 55, + 364, + 296, + 450 + ], + "type": "text", + "content": " and the target image " + }, + { + "bbox": [ + 55, + 364, + 296, + 450 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_t" + }, + { + "bbox": [ + 55, + 364, + 296, + 450 + ], + "type": "text", + "content": ", we compute our photometric loss per view as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 60, + 457, + 296, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 457, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 60, + 457, + 296, + 479 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {p}} = \\min _ {\\mathbf {I} _ {s} \\in \\mathbb {V} _ {\\text {s o u r c e}}} \\left(\\lambda_ {1} \\mathcal {L} _ {1} \\left(\\mathbf {I} _ {t}, \\hat {\\mathbf {I}} _ {t, s}\\right) + \\lambda_ {\\text {S S I M}} \\mathcal {L} _ {\\text {S S I M}} \\left(\\mathbf {I} _ {t}, \\hat {\\mathbf {I}} _ {t, s}\\right)\\right). \\tag {4}", + "image_path": "52e236ee92f1750a8b0a18800c8c7a5ba8f59cb4d7beb77fd63b5621ec717d91.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 486, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 521 + ], + "type": "text", + "content": "We only consider the minimum loss across the views in " + }, + { + "bbox": [ + 55, + 486, + 295, + 521 + ], + "type": "inline_equation", + "content": "\\mathbb{V}_{\\mathrm{source}}" + }, + { + "bbox": [ + 55, + 486, + 295, + 521 + ], + "type": "text", + "content": ", in practice across patches. The scalars " + }, + { + "bbox": [ + 55, + 486, + 295, + 521 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 55, + 486, + 295, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 486, + 295, + 521 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{SSIM}}" + }, + { + "bbox": [ + 55, + 486, + 295, + 521 + ], + "type": "text", + "content": " weight the absolute error " + }, + { + "bbox": [ + 55, + 486, + 295, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 55, + 486, + 295, + 521 + ], + "type": "text", + "content": " and the SSIM loss " + }, + { + "bbox": [ + 55, + 486, + 295, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{SSIM}}" + }, + { + "bbox": [ + 55, + 486, + 295, + 521 + ], + "type": "text", + "content": " [105]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 521, + 296, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 296, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 296, + 583 + ], + "type": "text", + "content": "To regularize our 3D geometry, we impose smoothness using an edge-aware smoothness loss [27]. We estimate 2D depth maps " + }, + { + "bbox": [ + 55, + 521, + 296, + 583 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{d}}_t" + }, + { + "bbox": [ + 55, + 521, + 296, + 583 + ], + "type": "text", + "content": " for views in " + }, + { + "bbox": [ + 55, + 521, + 296, + 583 + ], + "type": "inline_equation", + "content": "\\mathbb{V}_{\\mathrm{target}}" + }, + { + "bbox": [ + 55, + 521, + 296, + 583 + ], + "type": "text", + "content": ", using Eq. 3. From " + }, + { + "bbox": [ + 55, + 521, + 296, + 583 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{d}}_t" + }, + { + "bbox": [ + 55, + 521, + 296, + 583 + ], + "type": "text", + "content": ", we obtain the inverse mean-normalized depths " + }, + { + "bbox": [ + 55, + 521, + 296, + 583 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{d}}_t^*" + }, + { + "bbox": [ + 55, + 521, + 296, + 583 + ], + "type": "text", + "content": " and compute the edge-aware smoothness loss " + }, + { + "bbox": [ + 55, + 521, + 296, + 583 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{s}}" + }, + { + "bbox": [ + 55, + 521, + 296, + 583 + ], + "type": "text", + "content": " per view as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 88, + 589, + 295, + 605 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 589, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 88, + 589, + 295, + 605 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {s}} = \\left| \\nabla_ {x} \\tilde {\\mathbf {d}} _ {t} ^ {*} \\right| e ^ {- \\left| \\nabla_ {x} \\mathbf {I} _ {t} \\right|} + \\left| \\nabla_ {y} \\tilde {\\mathbf {d}} _ {t} ^ {*} \\right| e ^ {- \\left| \\nabla_ {y} \\mathbf {I} _ {t} \\right|}, \\tag {5}", + "image_path": "b3dfeb3a32bda66353784cf01ab529269b3c00de649ef6a57b055b193bba8b1d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 612, + 270, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 270, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 270, + 624 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 612, + 270, + 624 + ], + "type": "inline_equation", + "content": "\\nabla_{x}" + }, + { + "bbox": [ + 55, + 612, + 270, + 624 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 612, + 270, + 624 + ], + "type": "inline_equation", + "content": "\\nabla_{y}" + }, + { + "bbox": [ + 55, + 612, + 270, + 624 + ], + "type": "text", + "content": " denote the first spatial derivatives." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 626, + 296, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 626, + 296, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 626, + 296, + 687 + ], + "type": "text", + "content": "Feature reconstruction. We learn a multi-view consistent and expressive 3D feature field using the 2D features " + }, + { + "bbox": [ + 55, + 626, + 296, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_t" + }, + { + "bbox": [ + 55, + 626, + 296, + 687 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 626, + 296, + 687 + ], + "type": "inline_equation", + "content": "\\mathbb{V}_{\\mathrm{target}}" + }, + { + "bbox": [ + 55, + 626, + 296, + 687 + ], + "type": "text", + "content": ". As we aim to learn a high-resolution (continuous) feature field, we render 2D features using Eq. 3 at the full image resolution " + }, + { + "bbox": [ + 55, + 626, + 296, + 687 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_t \\in \\mathbb{R}^{D \\times H \\times W}" + }, + { + "bbox": [ + 55, + 626, + 296, + 687 + ], + "type": "text", + "content": ". To compensate for" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 69, + 554, + 177 + ], + "blocks": [ + { + "bbox": [ + 316, + 69, + 554, + 177 + ], + "lines": [ + { + "bbox": [ + 316, + 69, + 554, + 177 + ], + "spans": [ + { + "bbox": [ + 316, + 69, + 554, + 177 + ], + "type": "image", + "image_path": "14640c21cf9305b7701c5a14a164adcf218d304d70e5071325f4210f6cc5d0cf.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 183, + 554, + 239 + ], + "lines": [ + { + "bbox": [ + 313, + 183, + 554, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 183, + 554, + 239 + ], + "type": "text", + "content": "Figure 3. 3D feature distillation. Given an input image, SceneDINO predicts a 3D feature field. 3D features " + }, + { + "bbox": [ + 313, + 183, + 554, + 239 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{X}}" + }, + { + "bbox": [ + 313, + 183, + 554, + 239 + ], + "type": "text", + "content": " are sampled from the feature field. For " + }, + { + "bbox": [ + 313, + 183, + 554, + 239 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{X}}" + }, + { + "bbox": [ + 313, + 183, + 554, + 239 + ], + "type": "text", + "content": ", we obtain " + }, + { + "bbox": [ + 313, + 183, + 554, + 239 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{Y}_{k\\mathrm{NN}}}" + }, + { + "bbox": [ + 313, + 183, + 554, + 239 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 183, + 554, + 239 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{Y}_{kr\\mathrm{and}}}" + }, + { + "bbox": [ + 313, + 183, + 554, + 239 + ], + "type": "text", + "content": " from the feature buffer. The segmentation head " + }, + { + "bbox": [ + 313, + 183, + 554, + 239 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 183, + 554, + 239 + ], + "type": "text", + "content": " distills the features into a low-dimensional space and is trained using " + }, + { + "bbox": [ + 313, + 183, + 554, + 239 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{dist}}" + }, + { + "bbox": [ + 313, + 183, + 554, + 239 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 256, + 555, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 256, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 256, + 555, + 376 + ], + "type": "text", + "content": "the reduced spatial dimension of " + }, + { + "bbox": [ + 313, + 256, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_t" + }, + { + "bbox": [ + 313, + 256, + 555, + 376 + ], + "type": "text", + "content": ", we employ the down-sampler " + }, + { + "bbox": [ + 313, + 256, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 313, + 256, + 555, + 376 + ], + "type": "text", + "content": " proposed by Fu et al. [25] to our rendered features " + }, + { + "bbox": [ + 313, + 256, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_t" + }, + { + "bbox": [ + 313, + 256, + 555, + 376 + ], + "type": "text", + "content": ". While current 2D SSL features capture semantics, they lack multi-view consistency, i.a., due to positional encodings used in ViTs [112], leading to different features for identical visual content at two distinct positions in an image. As we aim for multi-view consistency, we compensate for this by learning a constant decomposition " + }, + { + "bbox": [ + 313, + 256, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{F}} \\in \\mathbb{R}^{\\mathrm{D} \\times \\frac{\\mathrm{H}}{p} \\times \\frac{\\mathrm{W}}{p}}" + }, + { + "bbox": [ + 313, + 256, + 555, + 376 + ], + "type": "text", + "content": " of features induced by positional encodings. Our feature loss is defined per view as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 361, + 385, + 553, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 385, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 361, + 385, + 553, + 399 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {f}} = 1 - \\cos - \\sin \\left(\\mathbf {F} _ {t}, \\psi \\left(\\hat {\\mathbf {F}} _ {t}\\right) + \\overline {{\\mathbf {F}}}\\right), \\tag {6}", + "image_path": "728196f9302e0043575a42763de6e8fc5f90d58dc0186d63e4f989be622e2562.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 407, + 553, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 553, + 419 + ], + "type": "text", + "content": "where cos-sim is the cosine similarity between two features." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 419, + 553, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 553, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 553, + 456 + ], + "type": "text", + "content": "As image edges correlate with semantic edges and to further impose consistency, we regularize the rendered features " + }, + { + "bbox": [ + 313, + 419, + 553, + 456 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_t" + }, + { + "bbox": [ + 313, + 419, + 553, + 456 + ], + "type": "text", + "content": " using an edge-aware smoothness loss per view" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 347, + 462, + 553, + 478 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 462, + 553, + 478 + ], + "spans": [ + { + "bbox": [ + 347, + 462, + 553, + 478 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {f s}} = \\left| \\nabla_ {x} \\hat {\\mathbf {F}} _ {t} \\right| e ^ {- \\left| \\nabla_ {x} \\mathbf {I} _ {t} \\right|} + \\left| \\nabla_ {y} \\hat {\\mathbf {F}} _ {t} \\right| e ^ {- \\left| \\nabla_ {y} \\mathbf {I} _ {t} \\right|}. \\tag {7}", + "image_path": "a24d3e8244226a9a35bf4e4a26bb2495fc96fd0a88e972721c7fb550738f18b5.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 485, + 554, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 554, + 533 + ], + "type": "text", + "content": "Our final loss for training SceneDINO is a weighted sum of the photometric loss, the feature loss, and both smoothness losses, " + }, + { + "bbox": [ + 313, + 485, + 554, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{SceneDINO}} = \\lambda_{\\mathrm{p}}\\mathcal{L}_{\\mathrm{p}} + \\lambda_{\\mathrm{s}}\\mathcal{L}_{\\mathrm{s}} + \\lambda_{\\mathrm{f}}\\mathcal{L}_{\\mathrm{f}} + \\lambda_{\\mathrm{fs}}\\mathcal{L}_{\\mathrm{fs}}" + }, + { + "bbox": [ + 313, + 485, + 554, + 533 + ], + "type": "text", + "content": ", averaged over all pixels, features, and views." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 540, + 542, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 542, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 542, + 553 + ], + "type": "text", + "content": "3.3. 3D feature distillation for unsupervised SSC" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 558, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 554, + 689 + ], + "type": "text", + "content": "Given the expressive feature field representation, we aim to obtain unsupervised semantic predictions for SSC. While naive " + }, + { + "bbox": [ + 313, + 558, + 554, + 689 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 558, + 554, + 689 + ], + "type": "text", + "content": "-means [69, 71] can yield meaningful pseudo semantics, distilling features into a lower-dimensional space has been shown to be more effective in 2D semantic segmentation [33, 55]. To this end, we present a novel 3D feature distillation approach (cf. Fig. 3). We train a pointwise segmentation head " + }, + { + "bbox": [ + 313, + 558, + 554, + 689 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 558, + 554, + 689 + ], + "type": "text", + "content": ", mapping " + }, + { + "bbox": [ + 313, + 558, + 554, + 689 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{x}} \\in \\mathbb{R}^{\\mathrm{D}}" + }, + { + "bbox": [ + 313, + 558, + 554, + 689 + ], + "type": "text", + "content": " to a lower-dimensional distilled representation " + }, + { + "bbox": [ + 313, + 558, + 554, + 689 + ], + "type": "inline_equation", + "content": "z_{\\mathbf{x}} \\in \\mathbb{R}^{\\mathrm{K}}" + }, + { + "bbox": [ + 313, + 558, + 554, + 689 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 313, + 558, + 554, + 689 + ], + "type": "inline_equation", + "content": "\\mathrm{K} \\ll \\mathrm{D}" + }, + { + "bbox": [ + 313, + 558, + 554, + 689 + ], + "type": "text", + "content": ". The resulting distilled space is clustered to obtain pseudosemantic predictions " + }, + { + "bbox": [ + 313, + 558, + 554, + 689 + ], + "type": "inline_equation", + "content": "p_{\\mathbf{x}} \\in [0,1]^{\\mathrm{C}}" + }, + { + "bbox": [ + 313, + 558, + 554, + 689 + ], + "type": "text", + "content": ", with C pseudo classes." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Existing work in 2D unsupervised semantic segmentation has shown that SSL feature correspondence captures" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "1Note, camera poses can be obtained using unsupervised visual SLAM [7], strictly adhering to the fully unsupervised setting." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6787" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": "semantic class co-occurrence [33]. This correspondence between two batches of " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{N}" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " sample points " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{X} = [\\mathbf{x}_1, \\dots, \\mathbf{x}_N]" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{Y} = [\\mathbf{y}_1, \\dots, \\mathbf{y}_N]" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " can be expressed by pairwise feature similarity " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "S_{i,j} = \\cos -\\sin (f_{\\mathbf{x}_i}, f_{\\mathbf{y}_j}) \\in [-1, 1]" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": ". Similarly, we can express the correspondence in the distilled feature space by " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "S_{i,j}^h = \\cos -\\sin (h(f_{\\mathbf{x}_i}), h(f_{\\mathbf{y}_j})) \\in [-1, 1]" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": ". We describe the sampling of the " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_j" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " below." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 159, + 296, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 159, + 296, + 207 + ], + "spans": [ + { + "bbox": [ + 55, + 159, + 296, + 207 + ], + "type": "text", + "content": "Feature distillation. We aim to distill features such that similar features align while dissimilar features are separated. To this end, we use the contrastive correlation loss " + }, + { + "bbox": [ + 55, + 159, + 296, + 207 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{corr}}" + }, + { + "bbox": [ + 55, + 159, + 296, + 207 + ], + "type": "text", + "content": ", introduced by STEGO [33] and defined as" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 215, + 296, + 241 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 215, + 296, + 241 + ], + "spans": [ + { + "bbox": [ + 69, + 215, + 296, + 241 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o r r}} \\left(f _ {\\mathbf {X}}, f _ {\\mathbf {Y}}, b\\right) = - \\sum_ {i, j} \\left(S _ {i, j} - b\\right) \\max \\left(S _ {i, j} ^ {h}, 0\\right), \\tag {8}", + "image_path": "7b3f2de7f3be1b8235eb245cc2fc0aa4b7bc4ea436482a9830471538e747d00d.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 246, + 295, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 246, + 295, + 284 + ], + "spans": [ + { + "bbox": [ + 55, + 246, + 295, + 284 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 246, + 295, + 284 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{X}}" + }, + { + "bbox": [ + 55, + 246, + 295, + 284 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 246, + 295, + 284 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{Y}}" + }, + { + "bbox": [ + 55, + 246, + 295, + 284 + ], + "type": "text", + "content": " are the features of the two sample batches. This loss pushes " + }, + { + "bbox": [ + 55, + 246, + 295, + 284 + ], + "type": "inline_equation", + "content": "S_{i,j}^{h}" + }, + { + "bbox": [ + 55, + 246, + 295, + 284 + ], + "type": "text", + "content": " towards 1 in case " + }, + { + "bbox": [ + 55, + 246, + 295, + 284 + ], + "type": "inline_equation", + "content": "S_{i,j}" + }, + { + "bbox": [ + 55, + 246, + 295, + 284 + ], + "type": "text", + "content": " exceeds the threshold " + }, + { + "bbox": [ + 55, + 246, + 295, + 284 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 55, + 246, + 295, + 284 + ], + "type": "text", + "content": ". Otherwise, " + }, + { + "bbox": [ + 55, + 246, + 295, + 284 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{corr}}" + }, + { + "bbox": [ + 55, + 246, + 295, + 284 + ], + "type": "text", + "content": " pushes the " + }, + { + "bbox": [ + 55, + 246, + 295, + 284 + ], + "type": "inline_equation", + "content": "S_{i,j}^{h}" + }, + { + "bbox": [ + 55, + 246, + 295, + 284 + ], + "type": "text", + "content": " below 0." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 284, + 296, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 284, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 296, + 403 + ], + "type": "text", + "content": "The correlation loss " + }, + { + "bbox": [ + 55, + 284, + 296, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{corr}}" + }, + { + "bbox": [ + 55, + 284, + 296, + 403 + ], + "type": "text", + "content": " requires informative pairs of sampled features, balancing attractive and repulsive signals. Following STEGO [33], we consider three different relations: (1) feature pairs from the same image " + }, + { + "bbox": [ + 55, + 284, + 296, + 403 + ], + "type": "inline_equation", + "content": "(f_{\\mathbf{X}}, f_{\\mathbf{X}})" + }, + { + "bbox": [ + 55, + 284, + 296, + 403 + ], + "type": "text", + "content": ", (2) feature pairs from an image and its " + }, + { + "bbox": [ + 55, + 284, + 296, + 403 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 284, + 296, + 403 + ], + "type": "text", + "content": "-nearest neighbors in feature space " + }, + { + "bbox": [ + 55, + 284, + 296, + 403 + ], + "type": "inline_equation", + "content": "(f_{\\mathbf{X}}, f_{\\mathbf{Y}_{k\\mathrm{NN}}})" + }, + { + "bbox": [ + 55, + 284, + 296, + 403 + ], + "type": "text", + "content": ", and (3) feature pairs from an image and a randomly sampled other image " + }, + { + "bbox": [ + 55, + 284, + 296, + 403 + ], + "type": "inline_equation", + "content": "(f_{\\mathbf{X}}, f_{\\mathbf{Y}_{\\mathrm{rand}}})" + }, + { + "bbox": [ + 55, + 284, + 296, + 403 + ], + "type": "text", + "content": ". Note that each pair is obtained from SceneDINO's 3D feature field, see below. Equipped with the three feature sample pairs, we compute the full distillation loss as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 97, + 411, + 294, + 451 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 411, + 294, + 451 + ], + "spans": [ + { + "bbox": [ + 97, + 411, + 294, + 451 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {d i s t}} = \\lambda_ {\\mathrm {s e l f}} \\mathcal {L} _ {\\mathrm {c o r r}} \\left(f _ {\\mathbf {X}}, f _ {\\mathbf {X}}, b _ {\\mathrm {s e l f}}\\right) \\\\ + \\lambda_ {k \\mathrm {N N}} \\mathcal {L} _ {\\text {c o r r}} \\left(f _ {\\mathbf {X}}, f _ {\\mathbf {Y} _ {k \\mathrm {N N}}}, b _ {k \\mathrm {N N}}\\right) \\tag {9} \\\\ + \\lambda_ {\\text {r a n d}} \\mathcal {L} _ {\\text {c o r r}} \\left(f _ {\\mathbf {X}}, f _ {\\mathbf {Y} _ {\\text {r a n d}}}, b _ {\\text {r a n d}}\\right), \\\\ \\end{array}", + "image_path": "ff8165d64e65d9a917acd3d839f028e9faf7200955b7de4221adacd4cb17bb5f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 460, + 295, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 483 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 460, + 295, + 483 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{self}}" + }, + { + "bbox": [ + 55, + 460, + 295, + 483 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 460, + 295, + 483 + ], + "type": "inline_equation", + "content": "\\lambda_{k\\mathrm{NN}}" + }, + { + "bbox": [ + 55, + 460, + 295, + 483 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 460, + 295, + 483 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{rand}}" + }, + { + "bbox": [ + 55, + 460, + 295, + 483 + ], + "type": "text", + "content": " denote the scalar loss weights. " + }, + { + "bbox": [ + 55, + 460, + 295, + 483 + ], + "type": "inline_equation", + "content": "b_{\\mathrm{self}}" + }, + { + "bbox": [ + 55, + 460, + 295, + 483 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 460, + 295, + 483 + ], + "type": "inline_equation", + "content": "b_{k\\mathrm{NN}}" + }, + { + "bbox": [ + 55, + 460, + 295, + 483 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 460, + 295, + 483 + ], + "type": "inline_equation", + "content": "b_{\\mathrm{rand}}" + }, + { + "bbox": [ + 55, + 460, + 295, + 483 + ], + "type": "text", + "content": " are the contrastive thresholds." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 486, + 296, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 296, + 581 + ], + "type": "text", + "content": "Feature sampling in 3D. While obtaining feature pairs using 2D rendered features is straightforward [33], we aim to take advantage of our learned 3D geometry of the scene. To this end, we introduce a novel 3D feature sampling approach for the distillation loss " + }, + { + "bbox": [ + 55, + 486, + 296, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{dist}}" + }, + { + "bbox": [ + 55, + 486, + 296, + 581 + ], + "type": "text", + "content": " from Eq. (9). Our goal is to sample features both similar and dissimilar in terms of the encoded semantic concept, which should capture rich semantics as well as different semantic concepts." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 296, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 677 + ], + "type": "text", + "content": "First, we obtain all G visible 3D surface points " + }, + { + "bbox": [ + 55, + 582, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{V} \\in \\mathbb{R}^{3 \\times \\mathrm{G}}" + }, + { + "bbox": [ + 55, + 582, + 296, + 677 + ], + "type": "text", + "content": " and their depth " + }, + { + "bbox": [ + 55, + 582, + 296, + 677 + ], + "type": "inline_equation", + "content": "d_{\\mathbf{V}} \\in \\mathbb{R}_{+}^{\\mathrm{G}}" + }, + { + "bbox": [ + 55, + 582, + 296, + 677 + ], + "type": "text", + "content": " from the camera. To sample points that cover different semantic concepts, we use depth as a cue and sample different depth ranges. In particular, we sort the surface points " + }, + { + "bbox": [ + 55, + 582, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 55, + 582, + 296, + 677 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 55, + 582, + 296, + 677 + ], + "type": "inline_equation", + "content": "d_{\\mathbf{V}}" + }, + { + "bbox": [ + 55, + 582, + 296, + 677 + ], + "type": "text", + "content": ". The sorted surface points " + }, + { + "bbox": [ + 55, + 582, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{V}}" + }, + { + "bbox": [ + 55, + 582, + 296, + 677 + ], + "type": "text", + "content": " are partitioned into M equally-sized chunks; we uniformly sample a single 3D point from each chunk, resulting in M center points " + }, + { + "bbox": [ + 55, + 582, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{3 \\times \\mathrm{M}}" + }, + { + "bbox": [ + 55, + 582, + 296, + 677 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Equipped with the center points " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": ", we aim to extract rich semantic features from the feature field. While we could just obtain the features for " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": ", we query positions in the" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 358, + 68, + 511, + 168 + ], + "blocks": [ + { + "bbox": [ + 358, + 68, + 511, + 168 + ], + "lines": [ + { + "bbox": [ + 358, + 68, + 511, + 168 + ], + "spans": [ + { + "bbox": [ + 358, + 68, + 511, + 168 + ], + "type": "image", + "image_path": "7dd58bbc0b5b72c0032e950414ba5056d6f42570149f97ded9eebfe5bfeac687.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 172, + 554, + 228 + ], + "lines": [ + { + "bbox": [ + 313, + 172, + 554, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 172, + 554, + 228 + ], + "type": "text", + "content": "Figure 4. 3D feature sampling. We first sample a center point " + }, + { + "bbox": [ + 313, + 172, + 554, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_i" + }, + { + "bbox": [ + 313, + 172, + 554, + 228 + ], + "type": "text", + "content": " from all visible surface points. Further points are sampled within the radius " + }, + { + "bbox": [ + 313, + 172, + 554, + 228 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 172, + 554, + 228 + ], + "type": "text", + "content": " around the center point " + }, + { + "bbox": [ + 313, + 172, + 554, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_i" + }, + { + "bbox": [ + 313, + 172, + 554, + 228 + ], + "type": "text", + "content": ". Sampled points with sufficient density are accepted; otherwise rejected. The accepted points are used to obtain the feature batch " + }, + { + "bbox": [ + 313, + 172, + 554, + 228 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{X}}" + }, + { + "bbox": [ + 313, + 172, + 554, + 228 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 243, + 554, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 554, + 363 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 554, + 363 + ], + "type": "text", + "content": "neighborhood of " + }, + { + "bbox": [ + 313, + 243, + 554, + 363 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 243, + 554, + 363 + ], + "type": "text", + "content": " to increase semantic richness and better capture the 3D structure of the scene for distillation. In particular, for each center point, we randomly sample a point within a radius of " + }, + { + "bbox": [ + 313, + 243, + 554, + 363 + ], + "type": "inline_equation", + "content": "r = 0.5\\mathrm{m}" + }, + { + "bbox": [ + 313, + 243, + 554, + 363 + ], + "type": "text", + "content": ". To account for samples falling into unoccupied regions in our feature field, we only keep samples with a sufficient density " + }, + { + "bbox": [ + 313, + 243, + 554, + 363 + ], + "type": "inline_equation", + "content": "\\sigma > 0.5" + }, + { + "bbox": [ + 313, + 243, + 554, + 363 + ], + "type": "text", + "content": ". We repeat this sampling process until we obtain " + }, + { + "bbox": [ + 313, + 243, + 554, + 363 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 243, + 554, + 363 + ], + "type": "text", + "content": " valid samples per center point. Using these samples, we query our feature field, resulting in a feature batch " + }, + { + "bbox": [ + 313, + 243, + 554, + 363 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{X}} \\in \\mathbb{R}^{D \\times N}" + }, + { + "bbox": [ + 313, + 243, + 554, + 363 + ], + "type": "text", + "content": " for each of the G center points in each scene (cf. Fig. 4)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 363, + 555, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 555, + 506 + ], + "type": "text", + "content": "To obtain " + }, + { + "bbox": [ + 313, + 363, + 555, + 506 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{Y}_{k\\mathrm{NN}}}" + }, + { + "bbox": [ + 313, + 363, + 555, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 363, + 555, + 506 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{Y}_{\\mathrm{rand}}}" + }, + { + "bbox": [ + 313, + 363, + 555, + 506 + ], + "type": "text", + "content": ", we utilize a feature buffer that efficiently stores the sampled features of multiple scenes. Given a new input image, we obtain " + }, + { + "bbox": [ + 313, + 363, + 555, + 506 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 363, + 555, + 506 + ], + "type": "text", + "content": " feature batches " + }, + { + "bbox": [ + 313, + 363, + 555, + 506 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{X}}" + }, + { + "bbox": [ + 313, + 363, + 555, + 506 + ], + "type": "text", + "content": " as just described. For each " + }, + { + "bbox": [ + 313, + 363, + 555, + 506 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{X}}" + }, + { + "bbox": [ + 313, + 363, + 555, + 506 + ], + "type": "text", + "content": ", we randomly sample another feature batch from the buffer to obtain " + }, + { + "bbox": [ + 313, + 363, + 555, + 506 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{Y}_{\\mathrm{rand}}}" + }, + { + "bbox": [ + 313, + 363, + 555, + 506 + ], + "type": "text", + "content": ". To obtain " + }, + { + "bbox": [ + 313, + 363, + 555, + 506 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{Y}_{k\\mathrm{NN}}}" + }, + { + "bbox": [ + 313, + 363, + 555, + 506 + ], + "type": "text", + "content": ", we search in the feature buffer for the " + }, + { + "bbox": [ + 313, + 363, + 555, + 506 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 363, + 555, + 506 + ], + "type": "text", + "content": "-nearest neighbors of " + }, + { + "bbox": [ + 313, + 363, + 555, + 506 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{X}}" + }, + { + "bbox": [ + 313, + 363, + 555, + 506 + ], + "type": "text", + "content": ", using the average feature of each batch. From these " + }, + { + "bbox": [ + 313, + 363, + 555, + 506 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 363, + 555, + 506 + ], + "type": "text", + "content": "-nearest neighbors, we randomly pick a feature batch to obtain " + }, + { + "bbox": [ + 313, + 363, + 555, + 506 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{Y}_{k\\mathrm{NN}}}" + }, + { + "bbox": [ + 313, + 363, + 555, + 506 + ], + "type": "text", + "content": " and compute the distillation loss " + }, + { + "bbox": [ + 313, + 363, + 555, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{dist}}" + }, + { + "bbox": [ + 313, + 363, + 555, + 506 + ], + "type": "text", + "content": ". After repeating this process for each of the current " + }, + { + "bbox": [ + 313, + 363, + 555, + 506 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 363, + 555, + 506 + ], + "type": "text", + "content": " feature batches, we add the current feature batches to the feature buffer and remove the oldest batches." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 510, + 554, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 554, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 554, + 582 + ], + "type": "text", + "content": "Unsupervised probing. To obtain semantics, we probe the distilled K-dim. feature space using " + }, + { + "bbox": [ + 313, + 510, + 554, + 582 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 510, + 554, + 582 + ], + "type": "text", + "content": "-means [69, 71]. In particular, we update cluster centers " + }, + { + "bbox": [ + 313, + 510, + 554, + 582 + ], + "type": "inline_equation", + "content": "\\theta \\in \\mathbb{R}^{K\\times C}" + }, + { + "bbox": [ + 313, + 510, + 554, + 582 + ], + "type": "text", + "content": " using cosine similarity-based mini-batch " + }, + { + "bbox": [ + 313, + 510, + 554, + 582 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 510, + 554, + 582 + ], + "type": "text", + "content": "-means [91] during distillation. We compute " + }, + { + "bbox": [ + 313, + 510, + 554, + 582 + ], + "type": "inline_equation", + "content": "p_{\\mathbf{x}} = \\mathrm{softmax}(\\cos -\\mathrm{sim}(h(f_{\\mathbf{x}}),\\theta))" + }, + { + "bbox": [ + 313, + 510, + 554, + 582 + ], + "type": "text", + "content": " to infer C pseudo semantic class predictions." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 594, + 395, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 594, + 395, + 607 + ], + "spans": [ + { + "bbox": [ + 314, + 594, + 395, + 607 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 614, + 554, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 614, + 554, + 686 + ], + "spans": [ + { + "bbox": [ + 313, + 614, + 554, + 686 + ], + "type": "text", + "content": "We evaluate SceneDINO on SSC and compare it to a simple unsupervised SSC baseline (Sec. 4.1). We also report results for 2D unsupervised segmentation, including domain generalization results (Sec. 4.2). Finally, we explore multiview feature consistency (Sec. 4.3) and present an analysis of SceneDINO and our 3D distillation (Sec. 4.4)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Datasets. We train using KITTI-360 [66], composed of clips from a moving vehicle equipped with cameras. For" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6788" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "type": "text", + "content": "consistency, we follow S4C [38] by sampling eight views and using the dataset's camera poses. We further provide results with estimated poses. We also show experiments for training on RealEstate10k [119], composed of monocular videos. Here, we follow the setup of BTS [108], obtaining three views. If not noted differently, we report results obtained with training on KITTI-360. For SSC and 2D semantic segmentation validation, we use the SSCBench-KITTI-360 test split [64]. Cityscapes [19] and BDD100K [115] val are used for domain generalization results. To enable evaluation in 3D and 2D, we use the 19-class taxonomy of Cityscapes and perform 2D evaluation on Cityscapes, BDD100K, and KITTI-360 on 19 classes. For SSCBench, we combine classes to adhere to the 15 SSCBench classes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 244, + 294, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 244, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 244, + 294, + 376 + ], + "type": "text", + "content": "3D evaluation. Given our unsupervised setup, we predict pseudo-semantic classes that must be aligned with the ground truth for evaluation. We follow standard practice in 2D unsupervised semantic segmentation [17, 31, 33, 51, 92, 95] by applying Hungarian matching [57] to align our pseudo semantics. For validating the aligned semantics, we follow the standardized SSCBench [64] protocol and report both semantic performance using the mean Intersection-over-Union (mIoU) and geometric performance using IoU, precision, and recall. We report all metrics on SSCBench ranges " + }, + { + "bbox": [ + 55, + 244, + 294, + 376 + ], + "type": "inline_equation", + "content": "12.8\\mathrm{m}" + }, + { + "bbox": [ + 55, + 244, + 294, + 376 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 244, + 294, + 376 + ], + "type": "inline_equation", + "content": "25.6\\mathrm{m}" + }, + { + "bbox": [ + 55, + 244, + 294, + 376 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 244, + 294, + 376 + ], + "type": "inline_equation", + "content": "51.2\\mathrm{m}" + }, + { + "bbox": [ + 55, + 244, + 294, + 376 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 380, + 294, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 294, + 464 + ], + "type": "text", + "content": "2D evaluation. Following the established evaluation protocol in 2D unsupervised semantic segmentation [17, 31, 33, 51, 92, 95], we use the all-pixel accuracy (Acc) and mean Intersection-over-Union (mIoU) metrics. Likewise, in line with prior work, 2D segmentation predictions of all models are refined using a dense Conditional Random Field [56] before computing Acc and mIoU." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 469, + 294, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 469, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 469, + 294, + 589 + ], + "type": "text", + "content": "Multi-view feature consistency evaluation. We aim to evaluate the multi-view consistency of our feature field. As we are not aware of any general feed-forward 3D feature field approach, we compare against 2D SSL models. To measure multi-view consistency in 2D, we use two video frames and estimate optical flow and occlusions with RAFT [99]. We backward warp 2D features of the second frame to the first. On the aligned features, we compute the feature similarity using absolute error " + }, + { + "bbox": [ + 55, + 469, + 294, + 589 + ], + "type": "inline_equation", + "content": "(\\mathrm{L}_1)" + }, + { + "bbox": [ + 55, + 469, + 294, + 589 + ], + "type": "text", + "content": ", the Euclidean distance " + }, + { + "bbox": [ + 55, + 469, + 294, + 589 + ], + "type": "inline_equation", + "content": "(\\mathrm{L}_2)" + }, + { + "bbox": [ + 55, + 469, + 294, + 589 + ], + "type": "text", + "content": ", and the cosine similarity, ignoring occlusions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "Baselines. We are not aware of any existing unsupervised SSC approach. To allow for comparisons, we construct a competitive baseline for unsupervised SSC. In particular, we train the S4C approach with unsupervised semantics of the established STEGO [33] approach. For 2D semantic segmentation, we use U2Seg [78] and STEGO [33] as established unsupervised baselines. Note that U2Seg is trained on ImageNet [21] and COCO [67] using STEGO pseudo-labels. We use STEGO [33] with DINO [11] (ViT-B/8), DINOv2 [80] (ViT-B/14), and FiT3D [116] (ViT" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 70, + 553, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 70, + 553, + 126 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 553, + 126 + ], + "type": "text", + "content": "Table 1. SSCBench-KITTI-360 results. Semantic results using mIoU and per class IoU, and geometric results using IoU, Precision, and Recall (all in %, ↑) on SSCBench-KITTI-360 test using three depth ranges. We compare our baseline S4C + STEGO to our SceneDINO. We report S4C as a 2D supervised baseline." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 315, + 130, + 553, + 329 + ], + "blocks": [ + { + "bbox": [ + 315, + 130, + 553, + 329 + ], + "lines": [ + { + "bbox": [ + 315, + 130, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 315, + 130, + 553, + 329 + ], + "type": "table", + "html": "
MethodS4C + STEGOSceneDINO (Ours)S4C
SupervisionUnsupervised2D supervision
Range12.8 m25.6 m51.2 m12.8 m25.6 m51.2 m12.8 m25.6 m51.2 m
Semantic validation
mIoU10.539.266.6010.7610.018.0016.9413.9410.19
car18.5714.099.2221.2415.9411.2122.5818.6411.49
bicycle0.010.010.010.000.000.000.000.000.00
motorcycle0.000.000.000.000.000.000.000.000.00
truck0.110.040.020.000.000.007.514.372.12
other-v.0.010.050.020.000.000.000.000.010.06
person0.010.010.010.000.000.000.000.000.00
road61.9752.4738.1551.1049.1239.8269.3861.4648.23
sidewalk18.7420.9518.2120.2622.3118.9745.0337.1228.45
building14.7524.4417.8112.3318.2714.3226.3428.4821.36
fence1.410.200.111.910.900.589.706.373.64
vegetation15.8316.5811.3031.2225.5719.8535.7828.0421.43
terrain26.499.954.1723.2618.0215.2235.0322.8815.08
pole0.080.040.040.050.050.051.230.940.65
traffic-sign0.000.000.000.000.000.001.570.830.36
other_obj.0.050.040.020.000.000.000.000.000.00
Geometric validation
IoU49.3241.0836.3949.5442.2737.6054.6445.5739.35
Precision54.0446.2341.9153.2746.1041.5959.7550.3443.59
Recall84.9578.6973.4387.6183.5979.6786.4782.7980.16
", + "image_path": "94b74e3667e9105f79b902d364b7faf83bc77eedc8cc9397174fff38e729b998.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 346, + 555, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 555, + 441 + ], + "type": "text", + "content": "B/14) features. FiT3D offers multi-view refined DINOv2 features [116]. Note that FiT3D reports results, concatenating the refined features with DINOv2 features. We report results using both plain features only and the concatenation. We also use rendered 2D segmentations of our S4C + STEGO baseline for 2D validation. For multi-view feature consistency, we utilize DINO [11], DINOv2, and FiT3D [116] features as baselines." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 446, + 556, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 446, + 556, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 446, + 556, + 650 + ], + "type": "text", + "content": "Implementation details. Our encoder-decoder uses a DINO-B/8 [11] backbone and a dense prediction decoder [84]. The MLP decoder " + }, + { + "bbox": [ + 313, + 446, + 556, + 650 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 446, + 556, + 650 + ], + "type": "text", + "content": " entails two layers with 128 hidden features. As rendering features is expensive, " + }, + { + "bbox": [ + 313, + 446, + 556, + 650 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 446, + 556, + 650 + ], + "type": "text", + "content": " predicts 64 features. We employ another MLP to upproject again to the full dimensionality " + }, + { + "bbox": [ + 313, + 446, + 556, + 650 + ], + "type": "inline_equation", + "content": "D = 768" + }, + { + "bbox": [ + 313, + 446, + 556, + 650 + ], + "type": "text", + "content": ". If not stated differently, our target features are obtained from DINO-B/8 [11]. We train using a batch size of 4 and extract 32 patches of size " + }, + { + "bbox": [ + 313, + 446, + 556, + 650 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 313, + 446, + 556, + 650 + ], + "type": "text", + "content": " from each image to compute " + }, + { + "bbox": [ + 313, + 446, + 556, + 650 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{SceneDINO}}" + }, + { + "bbox": [ + 313, + 446, + 556, + 650 + ], + "type": "text", + "content": ". Volume rendering samples each ray at " + }, + { + "bbox": [ + 313, + 446, + 556, + 650 + ], + "type": "inline_equation", + "content": "L = 32" + }, + { + "bbox": [ + 313, + 446, + 556, + 650 + ], + "type": "text", + "content": " uniformly spaced points in inverse depth within " + }, + { + "bbox": [ + 313, + 446, + 556, + 650 + ], + "type": "inline_equation", + "content": "[3\\mathrm{m}, 80\\mathrm{m}]" + }, + { + "bbox": [ + 313, + 446, + 556, + 650 + ], + "type": "text", + "content": ". We train for " + }, + { + "bbox": [ + 313, + 446, + 556, + 650 + ], + "type": "inline_equation", + "content": "100\\mathrm{k}" + }, + { + "bbox": [ + 313, + 446, + 556, + 650 + ], + "type": "text", + "content": " steps using Adam [52] with a base learning rate of " + }, + { + "bbox": [ + 313, + 446, + 556, + 650 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 313, + 446, + 556, + 650 + ], + "type": "text", + "content": ". Training takes ca. 2 days on a single V100 GPU. We distill using a batch size of 4, 5 center points, a feature batch of size 576, and cluster with " + }, + { + "bbox": [ + 313, + 446, + 556, + 650 + ], + "type": "inline_equation", + "content": "K = 19" + }, + { + "bbox": [ + 313, + 446, + 556, + 650 + ], + "type": "text", + "content": ". For kNN sampling, we use " + }, + { + "bbox": [ + 313, + 446, + 556, + 650 + ], + "type": "inline_equation", + "content": "k = 4" + }, + { + "bbox": [ + 313, + 446, + 556, + 650 + ], + "type": "text", + "content": ". The feature buffer holds 256 feature batches. Refer to the supplement for more details." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 659, + 477, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 477, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 477, + 673 + ], + "type": "text", + "content": "4.1.3D semantic scene completion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "We assess the unsupervised SSC and geometric accuracy of SceneDINO with our 3D feature distillation approach on SSCBench-KITTI-360. In particular, Tab. 1 com" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6789" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 68, + 556, + 178 + ], + "blocks": [ + { + "bbox": [ + 57, + 68, + 556, + 178 + ], + "lines": [ + { + "bbox": [ + 57, + 68, + 556, + 178 + ], + "spans": [ + { + "bbox": [ + 57, + 68, + 556, + 178 + ], + "type": "image", + "image_path": "810d3f72e739835c1aaaf5f4c812a7b34c5c4dec26fa2b38671b11d7b415a1a0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 185, + 555, + 220 + ], + "lines": [ + { + "bbox": [ + 55, + 185, + 555, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 185, + 555, + 220 + ], + "type": "text", + "content": "Figure 5. Qualitative SSC comparison on KITTI-360. We show the input image, SceneDINO's feature field using the first three principal components and SSC prediction, the prediction of our baseline S4C + STEGO, and the ground truth. We only visualize surface voxels. Qualitative results show the expressiveness of our feature field and SceneDINO's capabilities to accurately reconstruct and label a scene." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 293, + 296, + 376 + ], + "blocks": [ + { + "bbox": [ + 55, + 233, + 298, + 289 + ], + "lines": [ + { + "bbox": [ + 55, + 233, + 298, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 233, + 298, + 289 + ], + "type": "text", + "content": "Table 2. 2D unsupervised semantic segmentation results on KITTI-360. Comparing SceneDINO to existing 2D methods and our S4C + STEGO 3D baseline, using Accuracy and mean IoU (in " + }, + { + "bbox": [ + 55, + 233, + 298, + 289 + ], + "type": "inline_equation", + "content": "\\%" + }, + { + "bbox": [ + 55, + 233, + 298, + 289 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 233, + 298, + 289 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 55, + 233, + 298, + 289 + ], + "type": "text", + "content": ") on the SSCBench-KITTI-360 test split. " + }, + { + "bbox": [ + 55, + 233, + 298, + 289 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 233, + 298, + 289 + ], + "type": "text", + "content": " denotes the use of plain FiT3D features. " + }, + { + "bbox": [ + 55, + 233, + 298, + 289 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 55, + 233, + 298, + 289 + ], + "type": "text", + "content": " denotes training on ImageNet and COCO." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 293, + 296, + 376 + ], + "lines": [ + { + "bbox": [ + 56, + 293, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 296, + 376 + ], + "type": "table", + "html": "
MethodFeaturesAccmIoU
U2Seg‡ [78]-72.8923.43
STEGO [33]DINO [11]73.3223.57
STEGO [33]DINOv2 [80]64.5424.82
STEGO [33]FiT3D [116]54.1922.29
STEGO [33]FiT3D† [116]57.2518.95
S4C [38] + STEGO [33]DINO [11]65.1621.67
SceneDINO (Ours)DINO [11]77.7425.81
", + "image_path": "d57ed7dc7c538e1040e8a8fbe59df8d47fa6163efd1bda8c928cae10ed0a2317.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 395, + 295, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 395, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 53, + 395, + 295, + 502 + ], + "type": "text", + "content": "pares SceneDINO against our unsupervised SSC baseline S4C [38] + STEGO [33]. SceneDINO achieves a (semantic) mIoU of " + }, + { + "bbox": [ + 53, + 395, + 295, + 502 + ], + "type": "inline_equation", + "content": "8.0\\%" + }, + { + "bbox": [ + 53, + 395, + 295, + 502 + ], + "type": "text", + "content": " for the range of " + }, + { + "bbox": [ + 53, + 395, + 295, + 502 + ], + "type": "inline_equation", + "content": "51.2\\mathrm{m}" + }, + { + "bbox": [ + 53, + 395, + 295, + 502 + ], + "type": "text", + "content": ", significantly improving over our unsupervised baseline " + }, + { + "bbox": [ + 53, + 395, + 295, + 502 + ], + "type": "inline_equation", + "content": "(6.6\\%)" + }, + { + "bbox": [ + 53, + 395, + 295, + 502 + ], + "type": "text", + "content": ". This demonstrates that SceneDINO effectively lifts DINO features into 3D. In terms of geometric accuracy, SceneDINO moderately improves over S4C + STEGO. Despite being fully unsupervised, SceneDINO comes within " + }, + { + "bbox": [ + 53, + 395, + 295, + 502 + ], + "type": "inline_equation", + "content": "2.2\\%" + }, + { + "bbox": [ + 53, + 395, + 295, + 502 + ], + "type": "text", + "content": " points mIoU of the 2D-supervised S4C." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 504, + 296, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 504, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 504, + 296, + 612 + ], + "type": "text", + "content": "Figure 5 provides qualitative samples on SSCBenchKITTI-360. SceneDINO's unsupervised SSC predictions are less noisy and capture finely resolved semantics compared to S4C + STEGO. Compared to the ground truth, we observe that SceneDINO captures both the geometry and general semantics of the scene. We visualize SceneDINO's feature field (before distillation) using the first three principal components. In PCA space, we observe that our feature field captures semantically meaningful regions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 623, + 201, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 201, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 201, + 635 + ], + "type": "text", + "content": "4.2. 2D semantic segmentation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": "Table 2 compares the semantic predictions of SceneDINO to recent 2D approaches and our 3D baseline. We obtain 2D semantic segmentations from SceneDINO and our S4C + STEGO baseline using semantic rendering [38]. SceneDINO with our 3D distillation approach outperforms STEGO with DINO features, an established 2D unsuper" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 314, + 304, + 555, + 399 + ], + "blocks": [ + { + "bbox": [ + 313, + 233, + 555, + 300 + ], + "lines": [ + { + "bbox": [ + 313, + 233, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 233, + 555, + 300 + ], + "type": "text", + "content": "Table 3. 2D unsupervised semantic segmentation domain generalization results. Comparing SceneDINO to existing 2D unsupervised semantic segmentation methods and S4C + STEGO 3D baseline, using Accuracy and mean IoU (in %, " + }, + { + "bbox": [ + 313, + 233, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 313, + 233, + 555, + 300 + ], + "type": "text", + "content": "). We train on KITTI-360 images and report domain generalization results on Cityscapes and BDD-100K val. " + }, + { + "bbox": [ + 313, + 233, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 313, + 233, + 555, + 300 + ], + "type": "text", + "content": " denotes plain FiT3D features." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 314, + 304, + 555, + 399 + ], + "lines": [ + { + "bbox": [ + 314, + 304, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 314, + 304, + 555, + 399 + ], + "type": "table", + "html": "
MethodFeaturesCityscapesBDD-100K
AccmIoUAccmIoU
U2Seg [78]-75.5718.6269.0017.99
STEGO [33]DINO [11]71.2119.4275.0221.41
STEGO [33]DINOv2 [80]68.4119.7365.7221.77
STEGO [33]FiT3D [116]66.9421.0165.9620.99
STEGO [33]FiT3D† [116]64.7617.1760.8319.09
S4C [38] + STEGO [33]DINO [11]54.8014.0444.9811.62
SceneDINO (Ours)DINO [11]73.1722.8172.2822.09
", + "image_path": "33bbb0735d5968570bcf451a0aaad96ecfbd3730f7b9b1a210ddda94039f514e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 414, + 555, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 414, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 312, + 414, + 555, + 498 + ], + "type": "text", + "content": "vised semantic segmentation approach. In particular, the mIoU of SceneDINO is " + }, + { + "bbox": [ + 312, + 414, + 555, + 498 + ], + "type": "inline_equation", + "content": "2.24\\%" + }, + { + "bbox": [ + 312, + 414, + 555, + 498 + ], + "type": "text", + "content": " points higher than for STEGO (w/ DINO). Utilizing 3D refined features from FiT3D deteriorates the baseline relative to DINO, indicating that the FiT3D refinement reduces feature expressiveness. Notably, our unsupervised 3D baseline S4C + STEGO transfers significantly worse to 2D than SceneDINO." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 498, + 556, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 594 + ], + "type": "text", + "content": "We also validate SceneDINO, trained on KITTI-360, on Cityscapes and BDD10K, demonstrating domain generalization. The results are reported in Tab. 3. SceneDINO outperforms all baselines in mIoU on both datasets while only falling short in Acc. S4C + STEGO falls short in generalization. We suspect this poor generalization is caused by the fact that S4C does not rely on general SSL features in the final model, while our feature field generalizes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 600, + 480, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 480, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 480, + 613 + ], + "type": "text", + "content": "4.3. Multi-view feature consistency" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "type": "text", + "content": "We analyze the multi-view consistency of our feature field against existing 2D SSL features in Tab. 4. We report the results of SceneDINO trained on KITTI-360 and RealEstate10K. SceneDINO trained using DINO features exhibits significant improvements in multi-view feature consistency over standard DINO features. We also train SceneDINO using target features from DINOv2 [80]. Compared to standard DINOv2 and FiT3D fea" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6790" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 130, + 296, + 220 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 127 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 127 + ], + "type": "text", + "content": "Table 4. Multi-view consistency results. Comparing multi-view consistency of SceneDINO to existing 2D SSL features, using " + }, + { + "bbox": [ + 55, + 70, + 297, + 127 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_1" + }, + { + "bbox": [ + 55, + 70, + 297, + 127 + ], + "type": "text", + "content": " distance " + }, + { + "bbox": [ + 55, + 70, + 297, + 127 + ], + "type": "inline_equation", + "content": "(\\downarrow)" + }, + { + "bbox": [ + 55, + 70, + 297, + 127 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 70, + 297, + 127 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_2" + }, + { + "bbox": [ + 55, + 70, + 297, + 127 + ], + "type": "text", + "content": " distance " + }, + { + "bbox": [ + 55, + 70, + 297, + 127 + ], + "type": "inline_equation", + "content": "(\\downarrow)" + }, + { + "bbox": [ + 55, + 70, + 297, + 127 + ], + "type": "text", + "content": ", and cosine similarity " + }, + { + "bbox": [ + 55, + 70, + 297, + 127 + ], + "type": "inline_equation", + "content": "(\\uparrow)" + }, + { + "bbox": [ + 55, + 70, + 297, + 127 + ], + "type": "text", + "content": " on KITTI-360 and RealEstate10K. We compare DINO (top) and DINOv2-based (bottom) features. " + }, + { + "bbox": [ + 55, + 70, + 297, + 127 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 70, + 297, + 127 + ], + "type": "text", + "content": " denotes plain FiT3D features." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 130, + 296, + 220 + ], + "lines": [ + { + "bbox": [ + 56, + 130, + 296, + 220 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 296, + 220 + ], + "type": "table", + "html": "
MethodKITTI-360RealEstate10K
L1L2cos-simL1L2cos-sim
DINO [11]16.060.740.7014.410.660.75
SceneDINO (w/ DINO)6.450.330.935.870.280.95
DINOv2 [80]15.830.730.7014.200.660.75
FiT3D [116]22.860.810.8219.880.720.85
FiT3D† [116]7.020.330.935.670.270.95
SceneDINO (w/ DINOv2)5.240.240.964.870.220.97
", + "image_path": "52df232ea434c6fe491ce07cffa0c3df5c1f26eb54a4e09fdd31719287a4d2af.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 309, + 296, + 381 + ], + "blocks": [ + { + "bbox": [ + 55, + 227, + 296, + 306 + ], + "lines": [ + { + "bbox": [ + 55, + 227, + 296, + 306 + ], + "spans": [ + { + "bbox": [ + 55, + 227, + 296, + 306 + ], + "type": "text", + "content": "Table 5. SceneDINO analysis. We analyze the role of decomposing positional encodings, the choice of downsampling features during training, the effectiveness of the feature smoothness loss, the effect of estimated camera poses, and the choice of target features. We report the mean IoU (in %, ↑) using a range of " + }, + { + "bbox": [ + 55, + 227, + 296, + 306 + ], + "type": "inline_equation", + "content": "51.2\\mathrm{m}" + }, + { + "bbox": [ + 55, + 227, + 296, + 306 + ], + "type": "text", + "content": " on SSCBench-KITTI-360 test. " + }, + { + "bbox": [ + 55, + 227, + 296, + 306 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 55, + 227, + 296, + 306 + ], + "type": "text", + "content": " mIoU reports the absolute difference in % points to our standard model with DINO target features." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 309, + 296, + 381 + ], + "lines": [ + { + "bbox": [ + 56, + 309, + 296, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 296, + 381 + ], + "type": "table", + "html": "
Δ mIoUmIoUConfiguration
-1.186.82No downsampler (bilinear up. + aug.)
-1.176.83No feature smoothness loss (λfs = 0)
-0.747.26No pos. enc. decomposition
-0.127.88w/ estimated ORB-SLAM3 poses
—8.00Full framework (SceneDINO)
+1.089.08DINOv2 target features (vs. DINO)
", + "image_path": "5dbc1df1ba260bf86ff28cc7e40fae07fcb325d4ec28a632270ddb706eff2d7e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 397, + 296, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 397, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 397, + 296, + 460 + ], + "type": "text", + "content": "tures, SceneDINO's feature field yields significantly better multi-view consistency. Notably, compared against plain 3D refined features of FiT3D, SceneDINO shows a better multi-view consistency on both datasets and all metrics while also offering more expressiveness (cf. Tab. 2)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 467, + 185, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 467, + 185, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 467, + 185, + 480 + ], + "type": "text", + "content": "4.4. Analyzing SceneDINO" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 485, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 485, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 54, + 485, + 295, + 664 + ], + "type": "text", + "content": "To understand what core components contribute to obtaining an expressive feature field of SceneDINO, we omit or replace individual components and report the results in Tab. 5. Replacing the downsampling approach with bilinear upsampling and multi-crop augmentations, similar to [1], to obtain high-resolution target features decreases the SSC mIoU by " + }, + { + "bbox": [ + 54, + 485, + 295, + 664 + ], + "type": "inline_equation", + "content": "1.18\\%" + }, + { + "bbox": [ + 54, + 485, + 295, + 664 + ], + "type": "text", + "content": ". Omitting the feature smoothness loss leads to a similar mIoU drop. Abolishing the constant decomposition of positional encodings leads to an mIoU drop of " + }, + { + "bbox": [ + 54, + 485, + 295, + 664 + ], + "type": "inline_equation", + "content": "0.74\\%" + }, + { + "bbox": [ + 54, + 485, + 295, + 664 + ], + "type": "text", + "content": ". Training using unsupervised camera poses estimated by ORB-SLAM3 [7] results in an insignificant mIoU drop of only " + }, + { + "bbox": [ + 54, + 485, + 295, + 664 + ], + "type": "inline_equation", + "content": "0.12\\%" + }, + { + "bbox": [ + 54, + 485, + 295, + 664 + ], + "type": "text", + "content": " over using KITTI-360 poses. Going from DINO to DINOv2 target features leads to an increased mIoU of " + }, + { + "bbox": [ + 54, + 485, + 295, + 664 + ], + "type": "inline_equation", + "content": "1.08\\%" + }, + { + "bbox": [ + 54, + 485, + 295, + 664 + ], + "type": "text", + "content": ", demonstrating that SceneDINO can benefit from more expressive 2D target features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "In Tab. 6, we analyze our 3D distillation. Performing no distillation at all, just feature clustering, decreases mIoU by " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "1.61\\%" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": ". Omitting the kNN-correlation loss leads to an mIoU drop of " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "1.35\\%" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": ". Distilling only with center points," + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 315, + 130, + 555, + 190 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 554, + 126 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 554, + 126 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 554, + 126 + ], + "type": "text", + "content": "Table 6. Feature distillation analysis. We analyze the effectiveness of distilling SceneDINO's features, the " + }, + { + "bbox": [ + 313, + 70, + 554, + 126 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 70, + 554, + 126 + ], + "type": "text", + "content": "-NN-correlation loss, our neighborhood sampling, and our 3D sampling approach over standard 5-crop sampling. We report the mean IoU (in %, ↑) using a range of " + }, + { + "bbox": [ + 313, + 70, + 554, + 126 + ], + "type": "inline_equation", + "content": "51.2 \\mathrm{~m}" + }, + { + "bbox": [ + 313, + 70, + 554, + 126 + ], + "type": "text", + "content": " on SSCBench-KITTI-360 test." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 130, + 555, + 190 + ], + "lines": [ + { + "bbox": [ + 315, + 130, + 555, + 190 + ], + "spans": [ + { + "bbox": [ + 315, + 130, + 555, + 190 + ], + "type": "table", + "html": "
Δ mIoUmIoUConfiguration
-1.616.39No distillation
-1.356.65No kNN-correlation loss (λkNN = 0)
-0.977.03No neighborhood sampling (cf. Fig. 4)
-0.477.535-crop sampling [33] (instead 3D sampling)
—8.00Full framework (SceneDINO)
", + "image_path": "8c37d089a1e6c2802d19da28710db6f771709af7da02bb0e65c1deb53bc779fe.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 315, + 246, + 554, + 328 + ], + "blocks": [ + { + "bbox": [ + 313, + 197, + 555, + 243 + ], + "lines": [ + { + "bbox": [ + 313, + 197, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 197, + 555, + 243 + ], + "type": "text", + "content": "Table 7. Probing analysis. We analyze linear and unsupervised probing of our distilled SceneDINO features on SSCBenchKITTI-360 test using mean IoU (in %, ↑). For reference, we also report S4C (2D supervised). Linear probing uses 2D annotations." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 246, + 554, + 328 + ], + "lines": [ + { + "bbox": [ + 315, + 246, + 554, + 328 + ], + "spans": [ + { + "bbox": [ + 315, + 246, + 554, + 328 + ], + "type": "table", + "html": "
Probing approachTarget featuresmIoU
12.8 m25.6 m51.2 m
UnsupervisedDINO [11]10.7610.018.00
DINOv2 [80]13.7611.789.08
LinearDINO [11]13.6312.079.34
DINOv2 [80]15.8513.7010.57
S4C (full training)n/a16.9413.9410.19
", + "image_path": "801bc9b9593448c715c95e2c6b9091501fd7937c510b8dd4f6e6aacbe8741bdb.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 345, + 554, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 345, + 554, + 405 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 554, + 405 + ], + "type": "text", + "content": "i.e., not performing neighborhood sampling (cf. Fig. 4), reduces mIoU by " + }, + { + "bbox": [ + 313, + 345, + 554, + 405 + ], + "type": "inline_equation", + "content": "0.97\\%" + }, + { + "bbox": [ + 313, + 345, + 554, + 405 + ], + "type": "text", + "content": ". Using 5-crop feature sampling [33], instead of our proposed 3D sampling, leads to a reduced mIoU of " + }, + { + "bbox": [ + 313, + 345, + 554, + 405 + ], + "type": "inline_equation", + "content": "0.47\\%" + }, + { + "bbox": [ + 313, + 345, + 554, + 405 + ], + "type": "text", + "content": ". This demonstrates the effectiveness of performing distillation in 3D using our novel approach." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 405, + 556, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 405, + 556, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 405, + 556, + 550 + ], + "type": "text", + "content": "While focusing on unsupervised SSC, we can also linearly probe our distilled feature field (cf. Tab. 7). In particular, we train SceneDINO using different target features (DINO [11] and DINOv2 [11]), perform distillation, and probe the resulting distilled features. Using linear probing, i.e., training a single linear layer using 2D semantic labels, leads to a consistent mIoU increase over unsupervised probing. SceneDINO trained using DINOv2 target features even closes the gap to S4C, trained using 2D ground-truth semantic labels. We even surpass 2D supervised S4C slightly on the full range " + }, + { + "bbox": [ + 313, + 405, + 556, + 550 + ], + "type": "inline_equation", + "content": "(51.2\\mathrm{m})" + }, + { + "bbox": [ + 313, + 405, + 556, + 550 + ], + "type": "text", + "content": ", suggesting the effectiveness of SceneDINO also for weakly-supervised tasks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 561, + 388, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 561, + 388, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 561, + 388, + 574 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": "We presented SceneDINO, to our knowledge, the first approach for unsupervised semantic scene completion. Trained using multi-view images and 2D DINO features without human supervision, SceneDINO is able to predict an expressive 3D feature field using a single input image during inference. Our novel 3D distillation approach yields state-of-the-art results in unsupervised SSC. While we focus on unsupervised SSC, our multi-view feature consistency, linear probing, and domain generalization results highlight the potential of SceneDINO as a strong foundation for various 3D scene-understanding tasks." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "text", + "content": "6791" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 313 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 313 + ], + "type": "text", + "content": "Acknowledgments. This project was partially supported by the European Research Council (ERC) Advanced Grant SIMULACRON, DFG project CR 250/26-1 \"4D-YouTube\", and GNI Project \"AICC\". This project was also partially supported by the ERC under the European Union's Horizon 2020 research and innovation programme (grant agreement No. 866008). Additionally, this work has been co-funded by the LOEWE initiative (Hesse, Germany) within the emergenCITY center [LOEWE/1/12/519/03/05.001(0016)/72] and by the Deutsche Forschungsgemeinschaft (German Research Foundation, DFG) under Germany's Excellence Strategy (EXC 3066/1 \"The Adaptive Mind\", Project No. 533717223). Christoph Reich is supported by the Konrad Zuse School of Excellence in Learning and Intelligent Systems (ELIZA) through the DAAD programme Konrad Zuse Schools of Excellence in Artificial Intelligence, sponsored by the Federal Ministry of Education and Research. Christian Rupprecht is supported by an Amazon Research Award. Finally, we acknowledge the support of the European Laboratory for Learning and Intelligent Systems (ELLIS) and thank Mateo de Mayo as well as Igor Cvišić for help with estimating camera poses." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 325, + 115, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 325, + 115, + 338 + ], + "spans": [ + { + "bbox": [ + 56, + 325, + 115, + 338 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 346, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 66, + 346, + 294, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 346, + 294, + 378 + ], + "spans": [ + { + "bbox": [ + 66, + 346, + 294, + 378 + ], + "type": "text", + "content": "[1] Nikita Araslanov and Stefan Roth. Self-supervised augmentation consistency for adapting semantic segmentation. In CVPR, pages 15384-15394, 2021. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 380, + 295, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 380, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 66, + 380, + 295, + 413 + ], + "type": "text", + "content": "[2] Yuki Markus Asano, Christian Rupprecht, and Andrea Vedaldi. Self-labelling via simultaneous clustering and representation learning. In ICLR, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 416, + 294, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 416, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 66, + 416, + 294, + 458 + ], + "type": "text", + "content": "[3] Philip Bachman, R. Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. In NeurIPS*2019, pages 15509-15519. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 461, + 294, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 461, + 294, + 495 + ], + "spans": [ + { + "bbox": [ + 66, + 461, + 294, + 495 + ], + "type": "text", + "content": "[4] Adrien Bardes, Jean Ponce, and Yann LeCun. VI-CRegL: Self-supervised learning of local visual features. In NeurIPS*2022, pages 8799-8810. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 497, + 295, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 497, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 66, + 497, + 295, + 529 + ], + "type": "text", + "content": "[5] Adrien Bardes, Jean Ponce, and Yann LeCun. VI-CReg: Variance-invariance-covariance regularization for self-supervised learning. In *ICLR*, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 532, + 294, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 532, + 294, + 575 + ], + "spans": [ + { + "bbox": [ + 66, + 532, + 294, + 575 + ], + "type": "text", + "content": "[6] Yingjie Cai, Xuesong Chen, Chao Zhang, Kwan-Yee Lin, Xiaogang Wang, and Hongsheng Li. Semantic scene completion via integrating instances and scene in-the-loop. In CVPR, pages 324-333, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 578, + 294, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 578, + 294, + 632 + ], + "spans": [ + { + "bbox": [ + 66, + 578, + 294, + 632 + ], + "type": "text", + "content": "[7] Carlos Campos, Richard Elvira, Juan J. Gómez Rodríguez, José M. M. Montiel, and Juan D Tardós. ORB-SLAM3: An accurate open-source library for visual, visual-inertial and multi-map SLAM. IEEE Trans. Robot., 37(6):1874-1890, 2021. 4, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 635, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 635, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 66, + 635, + 294, + 666 + ], + "type": "text", + "content": "[8] Anh-Quan Cao and Raoul de Charette. Monoscene: Monocular 3D semantic scene completion. In CVPR, pages 3981-3991, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 66, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 670, + 294, + 713 + ], + "type": "text", + "content": "[9] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS*2020, pages 9912-9924. 2" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 320, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 553, + 106 + ], + "type": "text", + "content": "[10] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In ECCV, pages 132-149, 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 320, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 320, + 107, + 553, + 150 + ], + "type": "text", + "content": "[11] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. 1, 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 320, + 152, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 152, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 320, + 152, + 553, + 183 + ], + "type": "text", + "content": "[12] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv:2003.04297 [cs.CV], 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 185, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 185, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 320, + 185, + 553, + 227 + ], + "type": "text", + "content": "[13] Xiaokang Chen, Kwan-Yee Lin, Chen Qian, Gang Zeng, and Hongsheng Li. 3d sketch-aware semantic scene completion via semi-supervised structure prior. In CVPR, pages 4192-4201, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 228, + 553, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 228, + 553, + 260 + ], + "spans": [ + { + "bbox": [ + 320, + 228, + 553, + 260 + ], + "type": "text", + "content": "[14] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In CVPR, pages 9640-9649, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 262, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 262, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 320, + 262, + 553, + 304 + ], + "type": "text", + "content": "[15] Yuedong Chen, Chuanxia Zheng, Haofei Xu, Bohan Zhuang, Andrea Vedaldi, Tat-Jen Cham, and Jianfei Cai. MVSplat360: Feed-forward 360 scene synthesis from sparse views. In NeurIPS*2024, pages 107064-107086. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 306, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 306, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 320, + 306, + 553, + 348 + ], + "type": "text", + "content": "[16] Ran Cheng, Christopher Agia, Yuan Ren, Xinhai Li, and Bingbing Liu. S3CNet: A sparse semantic scene completion network for LiDAR point clouds. In CoRL, pages 2148-2161, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 350, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 350, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 320, + 350, + 553, + 392 + ], + "type": "text", + "content": "[17] Jang Hyun Cho, Utkarsh Mall, Kavita Bala, and Bharath Hariharan. PiCIE: Unsupervised semantic segmentation using invariance and equivariance in clustering. In CVPR, pages 16794-16804, 2021. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 394, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 394, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 320, + 394, + 553, + 437 + ], + "type": "text", + "content": "[18] Özgün Çiçek, Ahmed Abdulkadir, Soeren S Lienkamp, Thomas Brox, and Olaf Ronneberger. 3D U-Net: Learning dense volumetric segmentation from sparse annotation. In MICCAI, pages 424–432, 2016. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 438, + 553, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 438, + 553, + 492 + ], + "spans": [ + { + "bbox": [ + 320, + 438, + 553, + 492 + ], + "type": "text", + "content": "[19] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The Cityscapes dataset for semantic urban scene understanding. In CVPR, pages 3213-3223, 2016. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 494, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 494, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 320, + 494, + 553, + 536 + ], + "type": "text", + "content": "[20] Timothee Darcet, Federico Baldassarre, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Cluster and predict latent patches for improved masked image modeling. arXiv:2502.08769 [cs.CV], 2025. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 537, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 537, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 320, + 537, + 553, + 569 + ], + "type": "text", + "content": "[21] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 571, + 553, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 571, + 553, + 603 + ], + "spans": [ + { + "bbox": [ + 320, + 571, + 553, + 603 + ], + "type": "text", + "content": "[22] Carl Doersch, Abhinav Gupta, and Alexei A. Efros. Unsupervised visual representation learning by context prediction. In ICCV, pages 1422-1430, 2015. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 604, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 604, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 320, + 604, + 553, + 668 + ], + "type": "text", + "content": "[23] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth " + }, + { + "bbox": [ + 320, + 604, + 553, + 668 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 320, + 604, + 553, + 668 + ], + "type": "text", + "content": " words: Transformers for image recognition at scale. In ICLR, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 670, + 553, + 713 + ], + "type": "text", + "content": "[24] Linus Ericsson, Henry Gouk, Chen Change Loy, and Timothy M. Hospedales. Self-supervised representation learning: Introduction, advances, and challenges. IEEE Trans. Signal Process., 39(3):42-62, 2022. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "6792" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 294, + 115 + ], + "type": "text", + "content": "[25] Stephanie Fu, Mark Hamilton, Laura E. Brandt, Axel Feldmann, Zhoutong Zhang, and William T. Freeman. FeatUp: A model-agnostic framework for features at any resolution. In ICLR, 2024. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 117, + 294, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 117, + 294, + 149 + ], + "spans": [ + { + "bbox": [ + 61, + 117, + 294, + 149 + ], + "type": "text", + "content": "[26] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The KITTI dataset. Int. J. Robot. Res., 32(11):1231-1237, 2013. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 150, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 150, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 62, + 150, + 294, + 183 + ], + "type": "text", + "content": "[27] Clément Godard, Oisin Mac Aodha, and Gabriel J. Brostow. Unsupervised monocular depth estimation with left-right consistency. In CVPR, pages 270-279, 2017. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 184, + 294, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 184, + 294, + 238 + ], + "spans": [ + { + "bbox": [ + 62, + 184, + 294, + 238 + ], + "type": "text", + "content": "[28] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, et al. Bootstrap your own latent: A new approach to self-supervised learning. In NeurIPS*2020, pages 21271-21284. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 239, + 294, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 294, + 270 + ], + "type": "text", + "content": "[29] Agrim Gupta, Jiajun Wu, Jia Deng, and Li Fei-Fei. Siamese Masked Autoencoders. In NeurIPS*2023, pages 40676-40693. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 272, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 272, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 272, + 294, + 304 + ], + "type": "text", + "content": "[30] Huy Ha and Shuran Song. Semantic abstraction: Openworld 3D scene understanding from 2D vision-language models. In CoRL, pages 643-653, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 305, + 294, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 305, + 294, + 348 + ], + "spans": [ + { + "bbox": [ + 62, + 305, + 294, + 348 + ], + "type": "text", + "content": "[31] Oliver Hahn, Nikita Araslanov, Simone Schaub-Meyer, and Stefan Roth. Boosting unsupervised semantic segmentation with principal mask proposals. Trans. Mach. Learn. Res., 2024. 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 350, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 350, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 62, + 350, + 294, + 392 + ], + "type": "text", + "content": "[32] Oliver Hahn, Christoph Reich, Nikita Araslanov, Daniel Cremers, Christian Rupprecht, and Stefan Roth. Scene-centric unsupervised panoptic segmentation. In CVPR, pages 24485-24495, 2025. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 394, + 294, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 394, + 294, + 436 + ], + "spans": [ + { + "bbox": [ + 62, + 394, + 294, + 436 + ], + "type": "text", + "content": "[33] Mark Hamilton, Zhoutong Zhang, Bharath Hariharan, Noah Snavely, and William T. Freeman. Unsupervised semantic segmentation by distilling feature correspondences. In ICLR, 2022. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 437, + 294, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 437, + 294, + 481 + ], + "spans": [ + { + "bbox": [ + 62, + 437, + 294, + 481 + ], + "type": "text", + "content": "[34] Keonhee Han, Dominik Muhle, Felix Wimbauer, and Daniel Cremers. Boosting self-supervision for single-view scene completion via knowledge distillation. In CVPR, pages 9837-9847, 2024. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 482, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 482, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 62, + 482, + 294, + 525 + ], + "type": "text", + "content": "[35] Xian-Feng Han, Hamid Laga, and Mohammed Bennamoun. Image-based 3D object reconstruction: State-of-the-art and trends in the deep learning era. IEEE Trans. Pattern Anal. Mach. Intell., 43(5):1578-1604, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 526, + 294, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 526, + 294, + 559 + ], + "spans": [ + { + "bbox": [ + 62, + 526, + 294, + 559 + ], + "type": "text", + "content": "[36] Robert Harb and Patrick Knobelreiter. InfoSeg: Unsupervised semantic image segmentation with mutual information maximization. In " + }, + { + "bbox": [ + 62, + 526, + 294, + 559 + ], + "type": "inline_equation", + "content": "GCPR" + }, + { + "bbox": [ + 62, + 526, + 294, + 559 + ], + "type": "text", + "content": ", pages 18-32, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 559, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 559, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 62, + 559, + 294, + 591 + ], + "type": "text", + "content": "[37] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge University Press, 2003. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 592, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 592, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 62, + 592, + 294, + 635 + ], + "type": "text", + "content": "[38] Adrian Hayler, Felix Wimbauer, Dominik Muhle, Christian Rupprecht, and Daniel Cremers. S4C: Self-supervised semantic scene completion with neural fields. In 3DV, pages 409-420, 2024. 1, 2, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 636, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 636, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 62, + 636, + 294, + 669 + ], + "type": "text", + "content": "[39] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9729-9738, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "type": "text", + "content": "[40] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 320, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 553, + 105 + ], + "type": "text", + "content": "[41] Olivier Henaff. Data-efficient image pecognition with contrastive predictive coding. In ICML, pages 4182-4192, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 107, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 107, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 320, + 107, + 553, + 161 + ], + "type": "text", + "content": "[42] R. Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. In ICLR, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 163, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 163, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 320, + 163, + 553, + 228 + ], + "type": "text", + "content": "[43] Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3D v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. IEEE Trans. Pattern Anal. Mach. Intell., 46(12):10579-10596, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 230, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 230, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 320, + 230, + 553, + 285 + ], + "type": "text", + "content": "[44] Rui Huang, Songyou Peng, Ayca Takmaz, Federico Tombari, Marc Pollefeys, Shiji Song, Gao Huang, and Francis Engelmann. Segment3D: Learning fine-grained class-agnostic 3D segmentation without manual labels. In ECCV, pages 278–295, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 286, + 553, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 286, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 320, + 286, + 553, + 329 + ], + "type": "text", + "content": "[45] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3D semantic occupancy prediction. In CVPR, pages 9223-9232, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 331, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 331, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 320, + 331, + 553, + 374 + ], + "type": "text", + "content": "[46] Yuanhui Huang, Wenzhao Zheng, Borui Zhang, Jie Zhou, and Jiwen Lu. SelfOcc: Self-supervised vision-based 3D occupancy prediction. In CVPR, pages 19946-19956, 2024. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 376, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 376, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 320, + 376, + 553, + 420 + ], + "type": "text", + "content": "[47] Joel Janai, Fatma Güney, Aseem Behl, and Andreas Geiger. Computer vision for autonomous vehicles: Problems, datasets and state of the art. Found. Trends Comput. Graph. Vis., 12(1-3):1-308, 2020. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 421, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 421, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 320, + 421, + 553, + 454 + ], + "type": "text", + "content": "[48] Xu Ji, Joao F. Henriques, and Andrea Vedaldi. Invariant information clustering for unsupervised image classification and segmentation. In ICCV, pages 9865-9874, 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 456, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 456, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 320, + 456, + 553, + 510 + ], + "type": "text", + "content": "[49] Haoyi Jiang, Liu Liu, Tianheng Cheng, Xinjie Wang, Tianwei Lin, Zhizhong Su, Wenyu Liu, and Xinggang Wang. GaussTR: Foundation model-aligned Gaussian transformer for self-supervised 3D spatial understanding. arXiv:2412.13193 [cs.CV], 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 511, + 553, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 511, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 320, + 511, + 553, + 553 + ], + "type": "text", + "content": "[50] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. LERF: Language embedded radiance fields. In ICCV, pages 19729-19739, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 556, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 556, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 320, + 556, + 553, + 600 + ], + "type": "text", + "content": "[51] Chanyoung Kim, Woojung Han, Dayun Ju, and Seong Jae Hwang. EAGLE: Eigen aggregation learning for object-centric unsupervised semantic segmentation. In CVPR, pages 3523-3533, 2024. 3, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 601, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 601, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 320, + 601, + 553, + 623 + ], + "type": "text", + "content": "[52] Diederik P. Kingma and Jimmy Lei Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 320, + 624, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 624, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 320, + 624, + 553, + 678 + ], + "type": "text", + "content": "[53] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment Anything. In ICCV, pages 4015-4026, 2023. 1, 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 320, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 680, + 553, + 713 + ], + "type": "text", + "content": "[54] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing NeRF for editing via feature field distillation. In NeurIPS*2022, pages 23311-23330. 2" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6793" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "type": "text", + "content": "[55] Alexander Koenig, Maximilian Schambach, and Johannes Otterbach. Uncovering the inner workings of STEGO for safe unsupervised semantic segmentation. In CVPRW, pages 3789-3798, 2023. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 119, + 294, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 119, + 294, + 152 + ], + "spans": [ + { + "bbox": [ + 61, + 119, + 294, + 152 + ], + "type": "text", + "content": "[56] Philipp Krahenbuhl and Vladlen Koltun. Efficient inference in fully connected CRFs with Gaussian edge potentials. In NIPS*2011, pages 109-117. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 153, + 294, + 176 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 153, + 294, + 176 + ], + "spans": [ + { + "bbox": [ + 62, + 153, + 294, + 176 + ], + "type": "text", + "content": "[57] Harold W. Kuhn. The Hungarian method for the assignment problem. Nav. Res. Logist. Q., 2:83-97, 1955. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 178, + 294, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 178, + 294, + 221 + ], + "spans": [ + { + "bbox": [ + 62, + 178, + 294, + 221 + ], + "type": "text", + "content": "[58] Jie Li, Yu Liu, Dong Gong, Qinfeng Shi, Xia Yuan, Chunxia Zhao, and Ian D. Reid. RGBD based dimensional decomposition residual network for 3D semantic scene completion. In CVPR, pages 7693-7702, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 223, + 294, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 223, + 294, + 256 + ], + "spans": [ + { + "bbox": [ + 62, + 223, + 294, + 256 + ], + "type": "text", + "content": "[59] Jie Li, Kai Han, Peng Wang, Yu Liu, and Xia Yuan. Anisotropic convolutional networks for 3D semantic scene completion. In CVPR, pages 3348-3356, 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 258, + 294, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 294, + 300 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 294, + 300 + ], + "type": "text", + "content": "[60] Jie Li, Yu Liu, Xia Yuan, Chunxia Zhao, Roland Siegwart, Ian Reid, and Cesar Cadena. Depth based semantic scene completion with position importance aware loss. IEEE Robotics Autom. Lett., 5(1):219-226, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 304, + 294, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 304, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 62, + 304, + 294, + 335 + ], + "type": "text", + "content": "[61] Junnan Li, Pan Zhou, Caiming Xiong, and Steven Hoi. Prototypical contrastive learning of unsupervised representations. In ICLR, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 338, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 338, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 62, + 338, + 294, + 381 + ], + "type": "text", + "content": "[62] Pengfei Li, Yongliang Shi, Tianyu Liu, Hao Zhao, Guyue Zhou, and Ya-Qin Zhang. Semi-supervised implicit scene completion from sparse LiDAR. arXiv:2111.14798 [cs.CV], 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 384, + 294, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 384, + 294, + 438 + ], + "spans": [ + { + "bbox": [ + 62, + 384, + 294, + 438 + ], + "type": "text", + "content": "[63] Yiming Li, Zhiding Yu, Christopher B. Choy, Chaowei Xiao, José M. Álvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. VoxFormer: Sparse voxel transformer for camera-based 3D semantic scene completion. In CVPR, pages 9087-9098, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 441, + 294, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 441, + 294, + 506 + ], + "spans": [ + { + "bbox": [ + 62, + 441, + 294, + 506 + ], + "type": "text", + "content": "[64] Yiming Li, Sihang Li, Xinhao Liu, Moonjun Gong, Kenan Li, Nuo Chen, Zijun Wang, Zhiheng Li, Tao Jiang, Fisher Yu, Yue Wang, Hang Zhao, Zhiding Yu, and Chen Feng. SSCBench: A large-scale 3D semantic scene completion benchmark for autonomous driving. In IROS, pages 13333-13340, 2024. 1, 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 509, + 294, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 509, + 294, + 552 + ], + "spans": [ + { + "bbox": [ + 62, + 509, + 294, + 552 + ], + "type": "text", + "content": "[65] Zhiqi Li, Zhiding Yu, David Austin, Mingsheng Fang, Shiyi Lan, Jan Kautz, and José M. Álvarez. FB-OCC: 3D occupancy prediction based on forward-backward view transformation. arXiv:2307.01492 [cs.CV], 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 554, + 294, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 554, + 294, + 597 + ], + "spans": [ + { + "bbox": [ + 62, + 554, + 294, + 597 + ], + "type": "text", + "content": "[66] Yiyi Liao, Jun Xie, and Andreas Geiger. KITTI-360: A novel dataset and benchmarks for urban scene understanding in 2D and 3D. IEEE Trans. Pattern Anal. Mach. Intell., 45(3):3292-3310, 2023. 1, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 600, + 294, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 600, + 294, + 643 + ], + "spans": [ + { + "bbox": [ + 62, + 600, + 294, + 643 + ], + "type": "text", + "content": "[67] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, pages 740-755, 2024. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 646, + 294, + 688 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 646, + 294, + 688 + ], + "spans": [ + { + "bbox": [ + 62, + 646, + 294, + 688 + ], + "type": "text", + "content": "[68] Shice Liu, Yu Hu, Yiming Zeng, Qiankun Tang, Beibei Jin, Yinhe Han, and Xiaowei Li. See and think: Disentangling semantic scene completion. In NeurIPS*2018, pages 261-272. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 691, + 294, + 713 + ], + "type": "text", + "content": "[69] Stuart Lloyd. Least squares quantization in PCM. IEEE Trans. Inf. Theory, 28(2):129-137, 1982. 4, 5" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 321, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 321, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 321, + 73, + 553, + 106 + ], + "type": "text", + "content": "[70] Zhiliang Ma and Shilong Liu. A review of 3D reconstruction techniques in civil engineering and their applications. Adv. Eng. Inform., 37:163-174, 2018. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 108, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 108, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 321, + 108, + 553, + 140 + ], + "type": "text", + "content": "[71] James MacQueen. Some methods for classification and analysis of multivariate observations. In Berkeley Symp. on Math. Statist. and Prob., pages 281-298, 1967. 4, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 143, + 553, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 143, + 553, + 164 + ], + "spans": [ + { + "bbox": [ + 321, + 143, + 553, + 164 + ], + "type": "text", + "content": "[72] Nelson Max. Optical models for direct volume rendering. IEEE Trans. Vis. Comput. Graph., 1(2):99-108, 1995. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 167, + 553, + 199 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 167, + 553, + 199 + ], + "spans": [ + { + "bbox": [ + 321, + 167, + 553, + 199 + ], + "type": "text", + "content": "[73] Kirill Mazur, Edgar Sucar, and Andrew J Davison. Feature-realistic neural fusion for real-time, open set scene understanding. In ICRA, pages 8201-8207, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 201, + 553, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 201, + 553, + 245 + ], + "spans": [ + { + "bbox": [ + 321, + 201, + 553, + 245 + ], + "type": "text", + "content": "[74] Ruihang Miao, Weizhou Liu, Mingrui Chen, Zheng Gong, Weixin Xu, Chen Hu, and Shuchang Zhou. Occdepth: A depth-aware method for 3D semantic scene completion. arXiv.2302.13540 [cs.CV], 2023. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 247, + 553, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 247, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 321, + 247, + 553, + 300 + ], + "type": "text", + "content": "[75] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. Commun. ACM, 65(1):99-106, 2021. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 304, + 553, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 304, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 321, + 304, + 553, + 336 + ], + "type": "text", + "content": "[76] Yue Ming, Xuyang Meng, Chunxiao Fan, and Hui Yu. Deep learning for monocular depth estimation: A review. Neurocomputing, 438:14-33, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 338, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 338, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 321, + 338, + 553, + 381 + ], + "type": "text", + "content": "[77] Duy Kien Nguyen, Yanghao Li, Vaibhav Aggarwal, Martin R. Oswald, Alexander Kirillov, Cees G. M. Snoek, and Xinlei Chen. R-MAE: Regions meet masked autoencoders. In ICLR, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 384, + 553, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 384, + 553, + 427 + ], + "spans": [ + { + "bbox": [ + 321, + 384, + 553, + 427 + ], + "type": "text", + "content": "[78] Dantong Niu, Xudong Wang, Xinyang Han, Long Lian, Roei Herzig, and Trevor Darrell. Unsupervised universal image segmentation. In CVPR, pages 22744-22754, 2024. 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 430, + 553, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 430, + 553, + 462 + ], + "spans": [ + { + "bbox": [ + 321, + 430, + 553, + 462 + ], + "type": "text", + "content": "[79] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In ECCV, pages 69-84, 2016. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 321, + 464, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 464, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 321, + 464, + 553, + 520 + ], + "type": "text", + "content": "[80] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. DINOv2: Learning robust visual features without supervision. Trans. Mach. Learn. Res., 2024. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 321, + 521, + 553, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 521, + 553, + 575 + ], + "spans": [ + { + "bbox": [ + 321, + 521, + 553, + 575 + ], + "type": "text", + "content": "[81] Martin R Oswald, Eno Töppe, Claudia Nieuwenhuis, and Daniel Cremers. A review of geometry recovery from a single image focusing on curved object reconstruction. Innovations for Shape Analysis: Models and Algorithms, pages 343-378, 2013. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 321, + 578, + 553, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 578, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 321, + 578, + 553, + 632 + ], + "type": "text", + "content": "[82] Mingjie Pan, Jiaming Liu, Renrui Zhang, Peixiang Huang, Xiaqi Li, Hongwei Xie, Bing Wang, Li Liu, and Shang-hang Zhang. RenderOcc: Vision-centric 3D occupancy prediction with 2D rendering supervision. In ICRA, pages 12404-12411, 2024. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 321, + 635, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 635, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 321, + 635, + 553, + 678 + ], + "type": "text", + "content": "[83] Songyou Peng, Kyle Genova, Chiyu “Max” Jiang, Andrea Tagliasacchi, Marc Pollefeys, and Thomas Funkhouser. OpenScene: 3D scene understanding with open vocabularies. In CVPR, pages 815–824, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 321, + 681, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 681, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 321, + 681, + 553, + 712 + ], + "type": "text", + "content": "[84] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In ICCV, pages 12179-12188, 2021. 6" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6794" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 294, + 127 + ], + "type": "text", + "content": "[85] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, et al. SAM 2: Segment anything in images and videos. arXiv:2408.00714 [cs.CV], 2024. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 129, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 129, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 61, + 129, + 294, + 161 + ], + "type": "text", + "content": "[86] Stephan R. Richter and Stefan Roth. Matryoshka networks: Predicting 3D geometry via nested shape layers. In CVPR, pages 1936-1944, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 163, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 163, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 62, + 163, + 294, + 206 + ], + "type": "text", + "content": "[87] Christoph B. Rist, David Emmerichs, Markus Enzweiler, and Dariu M. Gavrila. Semantic scene completion using local deep implicit functions on LiDAR data. IEEE Trans. Pattern Anal. Mach. Intell., 44(10):7205-7218, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 208, + 294, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 208, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 62, + 208, + 294, + 240 + ], + "type": "text", + "content": "[88] Luis Roldão, Raoul de Charette, and Anne Verroust-Blondet. LMSCNet: Lightweight multiscale 3D semantic completion. In 3DV, pages 111–119, 2020. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 243, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 243, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 62, + 243, + 294, + 274 + ], + "type": "text", + "content": "[89] Luis Roldao, Raoul De Charette, and Anne Verroust-Blondet. 3D semantic scene completion: A survey. Int. J. Comput. Vis., 130(8):1978-2005, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 276, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 276, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 62, + 276, + 294, + 308 + ], + "type": "text", + "content": "[90] Johannes L. Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, pages 4104-4113, 2016. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 310, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 310, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 62, + 310, + 294, + 331 + ], + "type": "text", + "content": "[91] David Sculley. Web-scale " + }, + { + "bbox": [ + 62, + 310, + 294, + 331 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 62, + 310, + 294, + 331 + ], + "type": "text", + "content": "-means clustering. In WWW, page 1177-1178, 2010. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 333, + 294, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 333, + 294, + 375 + ], + "spans": [ + { + "bbox": [ + 62, + 333, + 294, + 375 + ], + "type": "text", + "content": "[92] Hyun Seok Seong, WonJun Moon, SuBeen Lee, and Jae-Pil Heo. Leveraging hidden positives for unsupervised semantic segmentation. In CVPR, pages 19540-19549, 2023. 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 377, + 294, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 377, + 294, + 422 + ], + "spans": [ + { + "bbox": [ + 62, + 377, + 294, + 422 + ], + "type": "text", + "content": "[93] Nur Muhammad Mahi Shafiullah, Chris Paxton, Lerrel Pinto, Soumith Chintala, and Arthur Szlam. CLIP-Fields: Weakly supervised semantic fields for robotic memory. In ICRA Workshop on Pretraining for Robotics, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 423, + 294, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 423, + 294, + 465 + ], + "spans": [ + { + "bbox": [ + 62, + 423, + 294, + 465 + ], + "type": "text", + "content": "[94] William Shen, Ge Yang, Alan Yu, Jansen Wong, Leslie Pack Kaelbling, and Phillip Isola. Distilled feature fields enable few-shot language-guided manipulation. In CoRL, pages 405-424, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 468, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 468, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 62, + 468, + 294, + 510 + ], + "type": "text", + "content": "[95] Leon Sick, Dominik Engel, Pedro Hermosilla, and Timo Ropinski. Unsupervised semantic segmentation through depth-guided feature correlation and sampling. In CVPR, pages 3637-3646, 2024. 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 513, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 513, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 513, + 294, + 555 + ], + "type": "text", + "content": "[96] Shuran Song, Fisher Yu, Andy Zeng, Angel X. Chang, Manolis Savva, and Thomas A. Funkhouser. Semantic scene completion from a single depth image. In CVPR, pages 190-198, 2017. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 557, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 557, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 557, + 294, + 612 + ], + "type": "text", + "content": "[97] Stanislaw Szymanowicz, Eldar Insafutdinov, Chuanxia Zheng, Dylan Campbell, Joao F. Henriques, Christian Rupprecht, and Andrea Vedaldi. Flash3D: Feed-forward generalisable 3D scene reconstruction from a single image. arXiv:2406.04343 [cs.CV], 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 613, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 613, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 62, + 613, + 294, + 656 + ], + "type": "text", + "content": "[98] Ayca Takmaz, Elisabetta Fedele, Robert W. Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. OpenMask3D: Open-vocabulary 3D instance segmentation. In NeurIPS*2023, pages 68367-68390. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 658, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 658, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 658, + 294, + 689 + ], + "type": "text", + "content": "[99] Zachary Teed and Jia Deng. RAFT: Recurrent all-pairs field transforms for optical flow. In ECCV, pages 402-419, 2020. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 62, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 692, + 294, + 713 + ], + "type": "text", + "content": "[100] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "type": "text", + "content": "Lin, and Hongyang Li. Scene as occupancy. In ICCV, pages 8372-8381, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 97, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 553, + 150 + ], + "type": "text", + "content": "[101] Nikolaos Tsagkas, Oisin Mac Aodha, and Chris Xiaoxuan Lu. VL-Fields: Towards language-grounded neural implicit spatial representations. In ICRA Workshop on Representations, Abstractions, and Priors for Robot Learning, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 553, + 195 + ], + "type": "text", + "content": "[102] Vadim Tschernezki, Iro Laina, Diane Larlus, and Andrea Vedaldi. Neural feature fusion fields: 3D distillation of self-supervised 2D image representations. In 3DV, pages 443-453, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 198, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 198, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 317, + 198, + 553, + 240 + ], + "type": "text", + "content": "[103] Shubham Tulsiani, Tinghui Zhou, Alexei A. Efros, and Jitendra Malik. Multi-view supervision for single-view reconstruction via differentiable ray consistency. In CVPR, 2017. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 243, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 243, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 317, + 243, + 553, + 285 + ], + "type": "text", + "content": "[104] Xudong Wang, Rohit Girdhar, Stella X. Yu, and Ishan Misra. Cut and learn for unsupervised object detection and instance segmentation. In CVPR, pages 3124-3134, 2023. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 287, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 287, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 317, + 287, + 553, + 331 + ], + "type": "text", + "content": "[105] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: From error visibility to structural similarity. IEEE Trans. Image Process., 13(4): 600-612, 2004. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 333, + 553, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 333, + 553, + 375 + ], + "spans": [ + { + "bbox": [ + 317, + 333, + 553, + 375 + ], + "type": "text", + "content": "[106] Silvan Weder, Hermann Blum, Francis Engelmann, and Marc Pollefeys. LabelMaker: Automatic semantic label generation from RGB-D trajectories. In 3DV, pages 334-343, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 377, + 553, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 377, + 553, + 422 + ], + "spans": [ + { + "bbox": [ + 317, + 377, + 553, + 422 + ], + "type": "text", + "content": "[107] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In CVPR, pages 14668-14678, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 423, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 423, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 317, + 423, + 553, + 465 + ], + "type": "text", + "content": "[108] Felix Wimbauer, Nan Yang, Christian Rupprecht, and Daniel Cremers. Behind the scenes: Density fields for single view reconstruction. In CVPR, pages 9076-9086, 2023. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 468, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 468, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 317, + 468, + 553, + 521 + ], + "type": "text", + "content": "[109] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Comput. Graph. Forum, pages 641-676, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 523, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 523, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 523, + 553, + 578 + ], + "type": "text", + "content": "[110] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep LiDAR point cloud segmentation via learning contextual shape priors from scene completion. In AAAI, pages 3101-3109, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 580, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 580, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 317, + 580, + 553, + 633 + ], + "type": "text", + "content": "[111] Jiawei Yang, Boris Ivanovic, Or Litany, Xinshuo Weng, Seung Wook Kim, Boyi Li, Tong Che, Danfei Xu, Sanja Fidler, Marco Pavone, and Yue Wang. EmerneRF: Emergent spatial-temporal scene decomposition via self-supervision. In ICLR, 2024. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 635, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 679 + ], + "type": "text", + "content": "[112] Jiawei Yang, Katie Z Luo, Jiefeng Li, Congyue Deng, Leonidas Guibas, Dilip Krishnan, Kilian Q Weinberger, Yonglong Tian, and Yue Wang. Denoising vision transformers. In ECCV, pages 453-469, 2024. 4" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "type": "text", + "content": "[113] Zhuoyue Yang, Ju Dai, and Junjun Pan. 3D reconstruction from endoscopy images: A survey. Comput. Biol. Med., 175:108546, 2024. 1" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6795" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 295, + 363 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[114] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelNeRF: Neural radiance fields from one or few images. In CVPR, pages 4578-4587, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 107, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 107, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 55, + 107, + 295, + 162 + ], + "type": "text", + "content": "[115] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. BDD100K: A diverse driving dataset for heterogeneous multitask learning. In CVPR, pages 2633-2642, 2020. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 163, + 295, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 163, + 295, + 206 + ], + "spans": [ + { + "bbox": [ + 57, + 163, + 295, + 206 + ], + "type": "text", + "content": "[116] Yuanwen Yue, Anurag Das, Francis Engelmann, Siyu Tang, and Jan Eric Lenssen. Improving 2D feature representations by 3D-aware fine-tuning. In ECCV, pages 57-74, 2024. 2, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 208, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 208, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 57, + 208, + 294, + 251 + ], + "type": "text", + "content": "[117] Pingping Zhang, Wei Liu, Yinjie Lei, Huchuan Lu, and Xiaoyun Yang. Cascaded context pyramid for full-resolution 3D semantic scene completion. In ICCV, pages 7800-7809, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 252, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 252, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 57, + 252, + 294, + 285 + ], + "type": "text", + "content": "[118] Yunpeng Zhang, Zheng Zhu, and Dalong Du. OccFormer: Dual-path transformer for vision-based 3D semantic occupancy prediction. In ICCV, pages 9433-9443, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 286, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 286, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 57, + 286, + 294, + 330 + ], + "type": "text", + "content": "[119] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snively. Stereo magnification: Learning view synthesis using multiplane images. ACM Trans. Graph., 37 (4):65, 2018. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 331, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 331, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 57, + 331, + 294, + 363 + ], + "type": "text", + "content": "[120] Onur Özyeşil, Vladislav Voroninski, Ronen Basri, and Amit Singer. A survey of structure from motion. Acta Numer., 26:305-364, 2017. 2" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6796" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Few-Shot Image Quality Assessment via Adaptation of Vision-Language Models/89cdaf16-5828-4989-b825-83012badf535_content_list.json b/2025/Few-Shot Image Quality Assessment via Adaptation of Vision-Language Models/89cdaf16-5828-4989-b825-83012badf535_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d9fc421eadc1eced4fcc9d07df81d8f748c6d73b --- /dev/null +++ b/2025/Few-Shot Image Quality Assessment via Adaptation of Vision-Language Models/89cdaf16-5828-4989-b825-83012badf535_content_list.json @@ -0,0 +1,1728 @@ +[ + { + "type": "text", + "text": "Few-Shot Image Quality Assessment via Adaptation of Vision-Language Models", + "text_level": 1, + "bbox": [ + 96, + 130, + 901, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xudong Li $^{1*}$ Zihao Huang $^{2*}$ Yan Zhang $^{1\\dagger}$ Yunhang Shen $^{3}$ Ke Li $^{3}$ Xiawu Zheng $^{1}$ Liujuan Cao $^{1}$ Rongrong Ji $^{1}$", + "bbox": [ + 200, + 179, + 795, + 217 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1 Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University, 361005, P.R. China", + "$^{2}$ Beijing Institute of Technology $^{3}$ Tencent Youtu Lab, Shanghai, China {1xd761050753, huangzihhhh, bzhy986}@gmail.com, {zhengxiawu, rrji}@xmu.edu.cn" + ], + "bbox": [ + 155, + 215, + 831, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 335, + 326, + 353 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image Quality Assessment (IQA) remains an unresolved challenge in computer vision due to complex distortions, diverse image content, and limited data availability. Existing Blind IQA (BIQA) methods largely rely on extensive human annotations, which are labor-intensive and costly due to the demanding nature of creating IQA datasets. To reduce this dependency, we propose the Gradient-Regulated MetaPrompt IQA Framework (GRMP-IQA), designed to efficiently adapt the visual-language pre-trained model, CLIP, to IQA tasks, achieving high accuracy even with limited data. GRMP-IQA consists of two core modules: (i) MetaPrompt Pre-training Module and (ii) Quality-Aware Gradient Regularization. The Meta Prompt Pre-training Module leverages a meta-learning paradigm to pre-train soft prompts with shared meta-knowledge across different distortions, enabling rapid adaptation to various IQA tasks. On the other hand, the Quality-Aware Gradient Regularization is designed to adjust the update gradients during fine-tuning, focusing the model's attention on quality-relevant features and preventing overfitting to semantic information. Extensive experiments on standard BIQA datasets demonstrate the superior performance to the state-of-the-art BIQA methods under limited data setting. Notably, utilizing just $20\\%$ of the training data, GRMP-IQA is competitive with most existing fully supervised BIQA approaches. Our code is available via https://github.com/LXDxmu/GRMP-IQA.", + "bbox": [ + 88, + 368, + 483, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 790, + 220, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the rise of the mobile internet era, the focus on computer vision has transitioned from initial concerns with compression and image processing [38] to handling user-generated content like smartphone photos and videos [6,", + "bbox": [ + 89, + 816, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a0e96b5a28fe707ed86d5577d049d86dc08d327f699e87aa4584757a3ed9bf22.jpg", + "image_caption": [ + "(b) Fine-tuning : Quality-Aware Gradient Regularization", + "Figure 1. Intuitive diagram of GRMP-IQA. (a) demonstrates using meta-learning for efficient soft prompt initialization with quality prior that enhances zero-shot generalization ( $\\uparrow$ $24\\%$ ), enabling the CLIP to adapt to BIQA tasks effectively. (b) illustrates gradient regularization during fine-tuning, which clips quality gradients aligned with semantic bias to guide the model toward quality-relevant features and reduce overly reliance on semantic content. In the t-SNE visualization, black dots represent noise samples with high semantic confidence (confidence $>0.8$ ) but incorrect quality predictions. The noticeable decrease ( $\\downarrow$ $73\\%$ ) in black dots highlights reduced semantic noise interference in IQA tasks." + ], + "image_footnote": [], + "bbox": [ + 521, + 321, + 901, + 518 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "35, 42], and lately to AI-generated content [5, 20]. This evolution has sharply increased demand for effective Blind Image Quality Assessment (BIQA) techniques, highlighting the importance of developing methodologies that can adeptly evaluate image quality without reference images. Data-driven BIQA models [34, 39, 48] based on deep neural networks have made significant progress in recent years. However, the quality scores for distorted images are often measured using the Mean Opinion Score (MOS), which is the average of multiple ratings (sometimes up to 120). As a result, acquiring a sufficient number of IQA training samples is quite labor-intensive and cost-expensive.", + "bbox": [ + 511, + 719, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution. † Corresponding authors.", + "bbox": [ + 114, + 886, + 370, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10442", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To mitigate this challenge, recent BIQA approaches [33, 37, 40, 46, 53] have leveraged large-scale image datasets through transfer learning or self-supervised learning. While these approaches improve performance, they often rely on computationally intensive pretraining or require domain-specific pretext tasks, limiting scalability. On the other hand, the vision-language model CLIP [36] has shown remarkable generalization across various downstream tasks, offering a potential alternative. Specifically, CLIP-IQA [44] demonstrates promising zero-shot performance in IQA by leveraging simple handcrafted prompts (e.g., \"Good/Bad photo\"). However, a significant performance gap remains compared to fully supervised methods [40, 45], highlighting the need to leverage limited labeled data to further close this performance gap. To bridge this gap, we introduce the concept of Few-Shot IQA, which aims to enable models to quickly generalize to new distortion types and unseen scenarios with only a small amount of labeled data (e.g., 50-200 training images). A natural way to adapt CLIP to downstream IQA tasks under this setup is through prompt tuning [11, 56, 57]. Prompt tuning techniques learn soft textual prompts (continuous embeddings) from a small labeled dataset while keeping the pretrained model parameters frozen, offering a lightweight and efficient solution.", + "bbox": [ + 89, + 90, + 483, + 454 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, as shown in Tab. 1, the direct application of the prompt tuning method CoOp [56] to IQA yields minimal improvement over simple linear probing on the CLIP image encoder, falling short of the expected benefits seen in high-level vision tasks. We attribute this limitation to a fundamental mismatch between CLIP's pretraining objectives and the specific requirements of IQA tasks. CLIP is pretrained to align text and images in a shared semantic space, with a focus on high-level semantics such as objects and scenes. In contrast, IQA demands fine-grained sensitivity to low-level distortion cues—an ability in which CLIP is relatively deficient [27, 52]. This discrepancy introduces two key challenges when applying prompt tuning to IQA tasks:", + "bbox": [ + 89, + 455, + 483, + 654 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) Sensitivity to Initialization: During prompt tuning, with CLIP's parameters frozen and limited sensitivity to distortions, the model depends heavily on the learned prompts for quality assessment. Different prompt initializations lead to distinct optimization directions [18], which may steer optimization toward features less relevant to quality, causing significant performance variability. As shown in Fig. 2, the average SRCC fluctuates considerably across different random initializations, requiring careful tuning for each IQA scenario and limiting the model's adaptability.", + "(2) Overfitting from Semantic Bias: Due to CLIP's inherent reliance on semantic information, fine-tuning with a small sample size often leads to overfitting, where the model learns spurious correlations (e.g., over-reliance on semantic features to infer quality [1, 22]) rather than true image quality indicators, which degrades IQA performance." + ], + "bbox": [ + 89, + 655, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d325e96ca932085b18e3dfbcafe7523e8fa82eb91ba75f3e9663db88f3b77ac4.jpg", + "image_caption": [ + "Figure 2. The fine-tuning results on the LIVEC dataset using CoOp [56] show that prompt tuning accuracy is highly sensitive to initial random initialization of the prompt, especially with limited training data, underscoring its critical role under data constraints." + ], + "image_footnote": [], + "bbox": [ + 516, + 87, + 730, + 199 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5e8f9120d7dfdcaaa83bd8c6fdc177716c72b5a54bc38e3f18196f9a16388e62.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 732, + 88, + 903, + 199 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As illustrated in Fig. 1, when the CLIP encoder is finetuned on a limited dataset, the latent space contains many instances (black dots) that confidently and accurately predict image semantics (with confidence levels above 0.8) but fail to reliably assess image quality.", + "bbox": [ + 511, + 280, + 906, + 354 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome these limitations, we propose Gradient-Regulated Meta-Prompt Learning for IQA (GRMP-IQA), a framework with two key components: (i) Meta Prompt Pre-training (MPP) mitigates sensitivity to prompt initialization by incorporating generalized distortion priors into the initialization of soft prompts, which enhances their adaptability to new IQA scenarios. By designing metatraining tasks with a well-annotated dataset containing diverse distortions (e.g., overexposure, blur) and optimizing prompts through bi-level gradient descent, we enable the model to learn shared meta-knowledge of quality representations across different distortions, thereby improving generalization across IQA tasks. (ii) Quality-Aware Gradient Regularization (QGR) addresses overfitting to semantic information by regulating gradient updates during fine-tuning, which balances the interplay between the quality knowledge gradient $G_{\\mathrm{qua}}$ and the semantic knowledge gradient $G_{\\mathrm{sem}}$ . By clipping $G_{\\mathrm{qua}}$ along $G_{\\mathrm{sem}}$ when their directions are overly aligned, this module suppresses semantic bias and guides the model to focus on distortion-related features, improving its ability to capture low-level quality details. Additionally, fine-tuning only CLIP's text branch can misalign quality perception between the image and text branches, limiting generalization [14]. To address this, we integrate text prompt tuning (CoOp [56]) and visual prompt tuning (VPT [11]) by jointly meta-learning initializations for both, ensuring complementary optimization to better adapt CLIP to new IQA scenarios (Tab. 5).", + "bbox": [ + 511, + 356, + 908, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 532, + 779, + 857, + 792 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a Meta Prompt Pre-training method that organizes meta tasks based on image distortion types and optimizes soft textual and visual prompts to gain shared quality meta-knowledge, enabling CLIP to rapidly adapt across various IQA scenarios.", + "- We develop a novel Quality-Aware Gradient Regularization method that clips gradients aligned with semantic" + ], + "bbox": [ + 513, + 795, + 903, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10443", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "directions to balance semantic and quality information during fine-tuning, ensuring the model prioritizes image quality while still integrating relevant semantic context.", + "bbox": [ + 102, + 90, + 480, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Extensive empirical results confirm that our approach is both effective and efficient. Notably, with just 200 data samples, our method is competitive with SOTA models on the LIVEC dataset (using $20\\%$ of the training data).", + "bbox": [ + 91, + 137, + 482, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 212, + 232, + 228 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Deep Learning Based BIQA methods", + "text_level": 1, + "bbox": [ + 89, + 238, + 410, + 255 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Early CNN-based IQA methods followed standard pretraining and fine-tuning pipelines [9, 51], while meta-learning approaches like MetaIQA [58] improved adaptability from synthetic to real-world images. However, CNNs struggle with non-local features, a gap addressed by ViT-based models [8, 13, 23, 24]. Recent methods like LIQE [52] leverage CLIP's multi-task learning, enhancing IQA through supervised fine-tuning across datasets. While effective, these methods rely heavily on extensive annotations, making them costly and time-intensive. In contrast, limited-data BIQA remains underexplored. CLIP-IQA[44] shows promise in zero-shot settings but has performance limits, highlighting the need for better use of limited labeled data. Methods like DEIQT[33] and LoDa[46] demonstrate that fine-tuning pre-trained ViTs reduces annotation needs but still underperform in few-shot scenarios. Self-supervised approaches[37, 40, 53] ease data constraints through tailored pre-training tasks but carry high computational costs. Additionally, recent studies [1, 22] explore disentangling semantic and distortion-related content for IQA—yet none fully address the dual challenges of data efficiency and computational cost.", + "bbox": [ + 89, + 261, + 482, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To bridge these gaps, our approach combines CLIP's pre-trained semantic knowledge with distortion-specific fine-tuning. By using prompt tuning to cut pre-training costs and a parameter-free regularization strategy (leveraging high- and low-level gradient correlations), it optimizes IQA more effectively with fewer labeled samples.", + "bbox": [ + 89, + 594, + 482, + 685 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Prompt Tuning", + "text_level": 1, + "bbox": [ + 89, + 696, + 243, + 713 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given the limited labeled data during training, prompt tuning is an effective approach to adapt vision-language pre-trained models for few-shot learning tasks [4, 47, 50]. CoOp [56] optimizes prompt vectors in CLIP's language branch for task adaptation but struggles with generalization to novel tasks. To address this, CoCoOp [55] introduces a lightweight meta-network to generate input-conditioned tokens, improving adaptability. Beyond textual prompts, visual prompts have also been explored for task adaptation [11]. Methods like MaPLe [14] and PromptSRC [15] incorporate trainable prompts into both language and visual branches, achieving significant performance gains across", + "bbox": [ + 89, + 719, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "base and novel tasks. Recent studies [21, 31] further construct meta-tasks based on semantic categories to initialize well-structured soft prompts, mitigating generalization decline in high-level tasks. Unlike these approaches that focus on high-level tasks, our proposed prompt tuning method is specifically designed for IQA, leveraging low-level distortion information to construct meta-tasks. This enables pretrained prompts to be distortion-sensitive and effectively capture shared quality knowledge.", + "bbox": [ + 511, + 90, + 903, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 511, + 242, + 648, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 511, + 268, + 624, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we propose the Gradient-Regulated MetaPrompt Image Quality Assessment (GRMP-IQA), which aims to adapt the CLIP for BIQA tasks with a few training samples. As depicted in Fig. 3, GRMP-IQA consists of two primary modules: (i) the Meta-Prompt Pre-training Module (MPP) and (ii) the Quality-Aware Gradient Regularization (QGR). The MPP module pre-trains visual-text prompts to acquire shared meta-knowledge on distortions, enabling quick adaptation to various IQA scenarios. The QGR module plays a key role in fine-tuning by adjusting gradient updates to prevent overfitting to semantic content.", + "bbox": [ + 511, + 290, + 903, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pre-training stage (Sec. 3.3). We randomly sample a minibatch from distortion meta-tasks, partitioning it into a support set $\\mathcal{D}^{\\mathrm{support}}$ and a query set $\\mathcal{D}^{\\mathrm{query}}$ . A bi-level gradient descent method progresses from the Inner-Loop on $\\mathcal{D}^{\\mathrm{support}}$ to the Outer-Loop on $\\mathcal{D}^{\\mathrm{query}}$ to optimize the learnable visual-textual prompts $[\\theta_T, \\theta_V]$ . These prompts are then used as initial weights for following fine-tuning.", + "bbox": [ + 511, + 458, + 903, + 563 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fine-tuning stage (Sec. 3.4). Given an input image $\\mathbf{x}$ and a hard semantic prompt $\\boldsymbol{w}_i =$ \"a photo of a [class]\", we predict the semantic class distributions $p^{\\text{sem}}(\\boldsymbol{w}_i|\\boldsymbol{x})$ and $p^{\\text{qua}}(\\boldsymbol{w}_i|\\boldsymbol{x})$ using the original semantic CLIP model and the fine-tuned CLIP model, respectively. The gradient of the KL divergence loss $\\nabla \\mathcal{L}_{kl}$ between these distributions is computed as the general semantic direction $G_{\\text{sem}}$ . Simultaneously, using soft quality prompt $\\theta_T$ and image $\\mathbf{x}$ , we compute the gradient $\\nabla \\mathcal{L}_{ce}$ of the quality loss as the quality direction $G_{\\text{qua}}$ . To refine the optimization direction of the IQA task, we adjust $G_{\\text{qua}}$ into $G_{\\text{qgr}}$ by clipping its components aligned with $G_{\\text{sem}}$ , ensuring the model focuses on image quality while reducing the impact of semantic noise.", + "bbox": [ + 511, + 564, + 905, + 761 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Visual-Text Meta-Prompt", + "text_level": 1, + "bbox": [ + 511, + 771, + 746, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visual Meta-Prompt. We utilize Deep Prompt Tuning (DPT) [56] as our Visual Meta-Prompt, with learnable parameters $\\theta_{\\mathbf{V}}$ . The input embedding for the $l$ -th layer's self-attention module in the ViT-based image encoder is denoted as $\\{\\pmb{f}^l,\\pmb{H}^l\\}$ , where $\\pmb{f}^l$ represents the classification (CLS) token, and $\\pmb{H}^{l} = \\{\\pmb{h}_{1}^{l},\\pmb{h}_{2}^{l},\\dots,\\pmb{h}_{N}^{l}\\}$ denotes the image patch embeddings. A learnable token $\\pmb{P}^{l}$ is appended to", + "bbox": [ + 511, + 794, + 903, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10444", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5dc812a037b8edb3e8e9bdb209255d9b3718d57b27accc4922a3496644d12a69.jpg", + "image_caption": [ + "(a) Pre-training stage: Meta-Prompt Pre-training Module" + ], + "image_footnote": [], + "bbox": [ + 93, + 85, + 480, + 286 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/760739a2d0f2a2c15efe37127809c8161409d5122e0d73af86cda2fd48426105.jpg", + "image_caption": [ + "(b) Fine-tuning stage: Quality-Aware Gradient Regularization", + "Figure 3. The overview of our GRMP-IQA. The core modules are the Meta-Prompt Pre-training and Quality-Aware Gradient Regularization, corresponding to two training processes. At the pre-training stage, we establish a distortion-specific meta-knowledge task for BIQA tasks, and the bi-level gradient descent is utilized to train visual-text meta-prompt $[\\theta_T,\\theta_V]$ (Sec. 3.3). These optimized prompts then serve as initial settings to efficiently adapt the CLIP model to IQA tasks with limited labels. At the few-shot fine-tuning stage, we first predict the class probability distributions $p^{sem}(w_i|x)$ and $p^{qua}(w_i|x)$ using the semantic and fine-tuned CLIP models, respectively. We then adjust IQA task gradients $G_{\\mathrm{qua}}$ by clipping gradients aligned with semantic task gradients $G_{\\mathrm{sem}}$ to generate refined gradient $G_{\\mathrm{qgr}}$ for backward updates (Sec. 3.4), mitigating the influence of semantic noise on quality predictions." + ], + "image_footnote": [], + "bbox": [ + 491, + 88, + 903, + 286 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the token sequence in each ViT layer, and the Multi-Head Self-Attention (MHSA) module processes the tokens as:", + "bbox": [ + 89, + 426, + 482, + 455 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n[ \\boldsymbol {f} ^ {l}, \\neg , \\boldsymbol {H} ^ {l} ] = \\operatorname {L a y e r} ^ {l} ([ \\boldsymbol {f} ^ {l - 1}, \\boldsymbol {P} ^ {l - 1}, \\boldsymbol {H} ^ {l - 1} ]), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 463, + 482, + 482 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the output of $P^l$ is discarded and not passed to the next layer, serving only as a set of learnable parameters.", + "bbox": [ + 89, + 489, + 482, + 520 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Text Meta-Prompt. We adopt CoOp [56] as our textual meta-prompt, with learnable parameters $\\theta_T$ . For each quality class $i$ , we construct a learnable quality prompt defined as $\\pmb{q}_i = \\{\\pmb{v}_1, \\pmb{v}_2, \\dots, \\pmb{v}_M, [\\mathrm{quality}]\\}$ . Here, [quality] $\\in$ {\"high quality\", \"low quality\"} serves as a categorical quality marker, and $\\{\\pmb{v}_m\\}_{m=1}^M$ denotes a set of $M$ learnable context vectors (with $M = 4$ in our work).", + "bbox": [ + 89, + 520, + 482, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Meta-Prompt Pre-training Module", + "text_level": 1, + "bbox": [ + 89, + 633, + 393, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During meta-learning, we aim to optimize the learnable visual-text prompt $\\theta = [\\theta_T,\\theta_V]$ . For an input image $x$ , the textual encoder $g(\\cdot)$ processes the quality prompts $q_{i}$ while the visual encoder extracts the feature vector $f$ . The probability of predicting high quality is:", + "bbox": [ + 89, + 656, + 482, + 733 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np \\left(\\boldsymbol {q} _ {\\text {h i g h}} \\mid \\boldsymbol {x}\\right) = \\frac {\\exp \\left(\\left\\langle \\boldsymbol {g} \\left(\\boldsymbol {q} _ {\\text {h i g h}}\\right) , \\boldsymbol {f} \\right\\rangle / \\tau\\right)}{\\exp \\left(\\left\\langle \\boldsymbol {g} \\left(\\boldsymbol {q} _ {\\text {h i g h}}\\right) , \\boldsymbol {f} \\right\\rangle / \\tau\\right) + \\exp \\left(\\left\\langle \\boldsymbol {g} \\left(\\boldsymbol {q} _ {\\text {l o w}}\\right) , \\boldsymbol {f} \\right\\rangle / \\tau\\right)},\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 739, + 486, + 785 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $q_{\\mathrm{high}}$ and $q_{\\mathrm{low}}$ denote high quality and low quality category prompt, $p(q_{\\mathrm{high}}|\\boldsymbol {x})$ represents the estimated probability that the quality of image $\\mathbf{x}$ is \"high quality\", $\\tau$ is a temperature parameter learned by CLIP, and $\\langle ,\\rangle$ denotes cosine similarity. The labeled quality scores are then rescaled to 0-1, denoted as $y$ , and the loss function is calculated as:", + "bbox": [ + 89, + 786, + 483, + 876 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = - (y \\log (p \\left(q _ {\\text {h i g h}} | \\boldsymbol {x}\\right)) + (1 - y) \\log (1 - p \\left(q _ {\\text {h i g h}} | \\boldsymbol {x}\\right)). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 883, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Constructing Distortion Meta-Knowledge Task. As noted by [29, 58], the ability to detect various types of image distortions is crucial for developing BIQA models with strong generalization capabilities across diverse scenarios. Moreover, the efficacy of prompt tuning is significantly dependent on the initial configuration of the prompts. This initial setup greatly affects the CLIP vision-language model's ability to swiftly adapt to different IQA scenarios. Drawing inspiration from the \"learn to learn\" ethos inherent in the deep meta-learning paradigm [43, 58], we propose an optimization-based method for effectively pre-training visual-textual prompts. These prompts incorporate shared quality insights from various image distortions, thereby enhancing CLIP's swift adaptability to IQA tasks [58]. To investigate the general rules of image distortion, we first constructed a $K_{t}$ -way distortion-specific image quality prediction task, denoted as $\\mathcal{T}_t$ . Then, it is used to build the meta-training set as $\\mathcal{D}^{\\mathrm{meta}} = \\left\\{\\mathcal{D}_{\\mathcal{T}_t}^{\\mathrm{support}}, \\mathcal{D}_{\\mathcal{T}_t}^{\\mathrm{query}}\\right\\}_{t=1}^T$ . Here, $\\mathcal{D}_{\\mathcal{T}_t}^{\\mathrm{support}}$ and $\\mathcal{D}_{\\mathcal{T}_t}^{\\mathrm{query}}$ represent the support and query sets for each task, respectively, with $T$ representing the total number of tasks. To simulate the process of prompt generalization to different distortions in BIQA scenarios, we randomly sample $k$ tasks as a mini-batch from the meta-training set, where $1 \\leq k \\leq T$ , to perform bi-level gradient optimization.", + "bbox": [ + 511, + 426, + 906, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Distortion-Aware Meta-prompt Learning. Our approach employs a bi-level gradient descent technique to bridge the learning process from the support to the query set. Specifically, it mainly consists of two optimization steps. In the Inner-Loop (the first level), we compute the gradients of the prompt parameters using the support set and apply the first update. In the Outer-Loop (the second level), we assess the", + "bbox": [ + 511, + 795, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10445", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "performance of the updated model on the query set and optimize the parameters again. This bi-level structure distills shared quality priors by training on a number of NR-IQA tasks with known distortions, enabling the meta-prompt to generalize rapidly across diverse new BIQA scenarios.", + "bbox": [ + 89, + 90, + 480, + 166 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inner-Loop. The objective of the Inner-Loop stage is to adapt the meta-prompt, denoted by $[\\pmb{\\theta}_{\\pmb{T}},\\pmb{\\theta}_{\\pmb{V}}]$ , to the $t^{th}$ support set $\\mathcal{D}_{\\mathcal{T}_t}^{\\mathrm{support}}$ within the mini-batch. During the first level of updates, we determine the loss $\\mathcal{L}(\\pmb {\\theta},\\mathcal{D}_{\\mathcal{T}_t}^{\\mathrm{support}})$ , following which the model parameters are updated on the support set using the inner learning rate $\\alpha$ , as specified by:", + "bbox": [ + 89, + 167, + 480, + 261 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\theta} _ {t} ^ {\\prime} = \\boldsymbol {\\theta} - \\alpha \\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} (\\boldsymbol {\\theta}, \\mathcal {D} _ {\\mathcal {T} _ {t}} ^ {\\text {s u p p o r t}}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 272, + 482, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Outer-Loop. In a similar vein, the second level of updates adjusts the parameters $\\theta_t^\\prime$ based on the query set $\\mathcal{D}_{\\mathcal{T}_t}^{\\mathrm{query}}$ :", + "bbox": [ + 89, + 303, + 480, + 335 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\theta} _ {t} \\leftarrow \\boldsymbol {\\theta} _ {t} ^ {\\prime} - \\alpha \\nabla_ {\\boldsymbol {\\theta} _ {t} ^ {\\prime}} \\mathcal {L} \\left(\\boldsymbol {\\theta} _ {t} ^ {\\prime}, \\mathcal {D} _ {\\mathcal {T} _ {t}} ^ {\\text {q u e r y}}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 347, + 482, + 366 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For a mini-batch of meta-tasks, this process culminates in aggregating gradients from all tasks to update the final model parameters, following the update rule:", + "bbox": [ + 89, + 377, + 482, + 422 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\theta} \\leftarrow \\boldsymbol {\\theta} - \\beta \\sum_ {t = 1} ^ {k} \\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} \\left(\\boldsymbol {\\theta} _ {t} ^ {\\prime}, \\mathcal {D} _ {\\mathcal {T} _ {t}} ^ {\\text {q u e r y}}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 436, + 482, + 476 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\beta$ represents the outer learning rate. Meta-learning effectively trains the learnable prompts $\\pmb{\\theta} = [\\pmb{\\theta}_T, \\pmb{\\theta}_V]$ , ensuring their generalization across various image distortions.", + "bbox": [ + 89, + 488, + 482, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Quality-Aware Gradient Regularization", + "text_level": 1, + "bbox": [ + 89, + 545, + 431, + 560 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We denote the original semantic CLIP model as $V^{\\mathrm{sem}}$ and the model obtained from the Meta-Prompt Pre-training Module, as $V^{\\mathrm{qua}}$ . The gradient for the IQA task is the quality gradient $G_{\\mathrm{qua}}$ , while the gradient for the semantic direction is the semantic gradient $G_{\\mathrm{sem}}$ . Recent studies [1, 53] have shown that IQA tasks can be misaligned with the high-level semantic representations of upstream tasks, leading to overfitting and reduced generalization. To prevent semantic overfitting, our Quality-aware Gradient Regularization (QGR) follows three key steps: derive a semantic direction $G_{\\mathrm{sem}}$ , compute the quality optimization direction $G_{\\mathrm{qua}}$ , and balance gradients based on their directional relationship.", + "bbox": [ + 89, + 566, + 482, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, to derive a general semantic direction $G_{\\mathrm{sem}}$ , we first design a hard prompt following [49, 52]: $\\boldsymbol{w}_i =$ \"a photo of a [class]\", where [class] represents one of nine categories: \"animal\", \"cityscape\", \"human\", \"indoor\", \"landscape\", \"night\", \"plant\", \"still-life\", \"others\". This prompt $\\boldsymbol{w}_i$ is input into the text encoders of $V^{\\mathrm{sem}}$ and $V^{\\mathrm{qua}}$ , generating text features that align with image features $f^{qua}$ from their respective visual encoders. As a result, we obtain zero-shot semantic prediction probabilities $p^{\\mathrm{sem}}(\\boldsymbol{w}_i|\\boldsymbol{x})$ and $p^{\\mathrm{qua}}(\\boldsymbol{w}_i|\\boldsymbol{x})$ . To measure the semantic alignment between", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$V^{\\mathrm{qua}}$ and $V^{\\mathrm{sem}}$ , we compute the KL divergence:", + "bbox": [ + 514, + 90, + 833, + 106 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {k l}} \\left(\\boldsymbol {V} ^ {\\text {q u a}}\\right) = - \\sum_ {i} \\boldsymbol {p} ^ {\\text {s e m}} \\left(\\boldsymbol {w} _ {i} \\mid \\boldsymbol {x}\\right) \\log \\frac {\\boldsymbol {p} ^ {\\text {q u a}} \\left(\\boldsymbol {w} _ {i} \\mid \\boldsymbol {x}\\right)}{\\boldsymbol {p} ^ {\\text {s e m}} \\left(\\boldsymbol {w} _ {i} \\mid \\boldsymbol {x}\\right)}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 114, + 903, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The gradient of $\\mathcal{L}_{\\mathrm{kl}}(V^{\\mathrm{qua}})$ , denoted as $G_{\\mathrm{sem}}$ , represents the general semantic optimization direction. Similarly, the quality direction $G_{\\mathrm{qua}}$ is obtained by computing the gradient of the cross-entropy loss $\\mathcal{L}_{\\mathrm{ce}}(V^{\\mathrm{qua}})$ , which compares the predicted probability $p(q_{\\mathrm{high}}|x)$ with the ground truth $y$ , as shown in Eq. 3. Finally, to balance the gradients between quality $G_{\\mathrm{qua}}$ and semantics $G_{\\mathrm{sem}}$ within a shared representation space, we analyze their relationship under two cases: (1) If the angle between $G_{\\mathrm{qua}}$ and $G_{\\mathrm{sem}}$ is less than 90 degrees, it indicates that the optimization directions for IQA quality knowledge and general semantic knowledge are consistent. In such cases, we clip the $G_{\\mathrm{qua}}$ along its component $G_{\\parallel}$ that parallel to the semantic direction $G_{\\mathrm{sem}}$ , to modulate the model's original quality optimization path, preventing overfitting to semantic correlations.", + "bbox": [ + 511, + 161, + 903, + 386 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(2) Conversely, if the angle between two areas of knowledge is more than 90 degrees, it means they are heading in different directions during improvement. In such situations, we don't adjust the quality gradient $G_{\\mathrm{qua}}$ , allowing the model to learn quality-aware features according to the original optimization direction. In conclusion, our gradient regularization strategy is mathematically formulated as:", + "bbox": [ + 511, + 387, + 903, + 492 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {G} _ {\\text {q g r}} = \\left\\{ \\begin{array}{l l} \\boldsymbol {G} _ {\\text {q u a}} & \\text {i f} \\boldsymbol {G} _ {\\text {q u a}} \\cdot \\boldsymbol {G} _ {\\text {s e m}} \\leq 0, \\\\ \\boldsymbol {G} _ {\\text {q u a}} - \\lambda \\frac {\\boldsymbol {G} _ {\\text {q u a}} \\cdot \\boldsymbol {G} _ {\\text {s e m}}}{\\| \\boldsymbol {G} _ {\\text {s e m}} \\| ^ {2}} \\boldsymbol {G} _ {\\text {s e m}} & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 502, + 903, + 555 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $\\lambda$ is introduced to generalize the formulation, providing flexibility in controlling the influence of general knowledge. Specifically, $\\lambda = 1$ projects $G_{\\mathrm{qua}}$ onto the orthogonal direction of $G_{\\mathrm{sem}}$ , while $\\lambda = 0$ reduces QGR to CoOp [56].", + "bbox": [ + 511, + 555, + 903, + 616 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 625, + 643, + 641 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Evaluation Protocols", + "text_level": 1, + "bbox": [ + 511, + 648, + 812, + 664 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conduct experiments on multiple BIQA datasets include LIVEC [7], KonIQ [10], LIVEFB [49], and SPAQ [3], which feature authentic distortions, and PIPAL [12], LIVE [38], and CSIQ [17], which feature synthetic distortions. LIVEC contains 1,162 mobile device images, SPAQ includes 11,125 photos from 66 smartphones, KonIQ has 10,073 images from open sources, and LIVEFB is the largest real-world dataset with 39,810 images. For synthetic distortions, LIVE and CSIQ contain 779 and 866 images with 5 and 6 types of distortions. PIPAL, a challenging dataset, includes 23,200 images with 40 types of distortions, including GAN-generated artifacts. We use Spearman's Rank Correlation Coefficient (SRCC) and the Pearson Linear Correlation Coefficient (PLCC) as metrics to quantify the monotonousness and accuracy of predictions.", + "bbox": [ + 511, + 671, + 906, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10446", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a60ac20e8b5aa59deb02574b0a961caf599086491b68eaeecb70fdae580c8852.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodLIVECKonIQCSIQLIVEPIPAL
Labels5010020050100200501002005010020050100200
HyperIQA [41]0.6480.7250.7900.6150.7100.7760.7900.8240.9090.8920.9120.9290.1020.3020.379
MetaIQA [58]0.6040.6260.6690.6180.6200.6600.7840.8490.8940.8400.8800.9190.3320.3480.371
DEIQT [33]0.6670.7180.8120.6380.6820.7540.8210.8910.9410.9200.9420.9550.3960.4100.436
MANIQA [48]0.6420.7690.7970.6520.7550.8100.7940.8470.8740.9090.9280.9570.1360.3610.470
CONTRIQUE [30]0.6950.7290.7610.7330.7940.8210.8400.9260.9400.8910.9220.9430.3790.4370.488
Re-IQA [37]0.5910.6210.7010.6850.7230.7540.8930.9070.9230.8840.8940.9290.2800.3500.431
CLIP w/ Linear Probe [36]0.6640.7210.7330.7360.7700.7820.8410.8920.9410.8960.9230.9410.2540.3030.368
CLIP w/ CoOp [44]†0.6950.7380.7460.6920.7430.762---------
LIQE [52]0.6910.7690.8100.7590.8010.8320.8380.8910.9240.9040.9340.948---
GRepQ [40]0.7600.7910.8220.8120.8360.8550.8780.9140.9410.9260.9370.9530.390†0.450†0.498†
GRMP-IQA (Ours)0.8360.8570.8750.8530.8720.8830.8930.9170.9410.9320.9430.9680.4740.5120.546
", + "bbox": [ + 94, + 88, + 906, + 284 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/03fdad7c0b62eb813b852f46a6ab86aa6a42d7039e7aa640c0e2f337a4c11a1a.jpg", + "table_caption": [ + "Table 1. SRCC performance comparison of our method with other IQA methods trained on limited labels. Bold indicates the best results, underlined marks the second-best, and the fifth-to-last through second-to-last lines show the CLIP-based IQA. † is our reproduction." + ], + "table_footnote": [], + "table_body": "
MethodLIVECSIQLIVECKonIQLIVEFBSPAQ
PLCCSRCCPLCCSRCCPLCCSRCCPLCCSRCCPLCCSRCCPLCCSRCC
Training Ratio20%20%20%20%20%20%20%20%10%10%20%20%
DEIQT [33]0.9680.9650.8850.8620.8220.7920.9080.8880.6240.5380.9120.908
LoDa [46]----0.8540.8150.9230.907----
GRMP-IQA (Ours)0.9720.9700.9580.9510.8970.8750.9310.9150.6860.6040.9250.920
Training Ratio80%80%80%80%80%80%80%80%80%80%80%80%
MetaIQA [58]0.9590.9600.9080.8990.8020.8350.8560.8870.5070.540--
CONTRIQUE [30]0.9610.9600.9550.9420.8570.8450.9060.8940.6410.5800.9190.914
DEIQT [33]0.9820.9800.9630.9460.8940.8750.9340.9210.6630.5710.9230.919
Re-IQA [37]0.9710.9700.9600.9470.8540.8400.9230.9140.7330.6450.9250.918
LIQE [52]0.9510.9700.9390.9360.9100.9040.9080.919----
CLIP-IQA+ [44]----0.8320.8050.9090.8950.5930.5750.8660.864
CDINet [54]0.9750.9770.9600.9520.8800.8650.9280.916--0.9220.919
QFM-IQM [22]0.9830.9810.9650.9540.9130.8910.9360.9220.6670.5670.9240.920
LoDa [46]0.9790.975--0.8990.8760.9440.9320.6790.5780.9280.925
GRMP-IQA (Ours)0.9830.9810.9740.9680.9160.8970.9450.9340.7040.6160.9320.927
", + "bbox": [ + 94, + 329, + 906, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Performance comparison measured by medians of SRCC and PLCC, and bold entries indicate the top two results.", + "bbox": [ + 133, + 609, + 859, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details and Setups", + "text_level": 1, + "bbox": [ + 89, + 632, + 397, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We build our model on CLIP-B/16 [36]. During pretraining, we optimize the visual-text prompt, while in finetuning, only the last four blocks of the image and text encoders are trained. For competing models, we use public results or re-train them under our setup. Each dataset is split into $80\\% / 20\\%$ for training and testing, with splits based on reference images to ensure content independence in datasets with synthetic distortions. To ensure robustness, we report the median performance across ten random splits. All experiments are run on four NVIDIA RTX3090 GPUs.", + "bbox": [ + 88, + 655, + 482, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Meta-Prompt Pre-training. We pre-train on TID2013 [32] and KADID-10K [25] dataset, which contain extensive distortion information. We set the learning rates $\\alpha$ and $\\beta$ to 1e-4 and 1e-2, and train for 50 epochs using Adam [16].", + "bbox": [ + 89, + 808, + 482, + 868 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fully Supervised Learning Setting. We randomly crop each input image into 10 patches of $224 \\times 224$ resolution", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and train the model for 9 epochs using AdamW [26]. The learning rate is $5 \\times 10^{-6}$ , with a scheduler over 9 decay epochs. The batch size is 16 for LIVEC and 128 for KonIQ. Few-Shot Learning Setting. In the few-shot setting, we follow the approach in a previous study [40] to train our GRMP-IQA model using randomly selected subsets of 50, 100, and 200 samples from the training set. The regularization weight $\\lambda$ is fixed at 5, and the training hyperparameters are set according to the schedule proposed by CoOp [56].", + "bbox": [ + 511, + 633, + 906, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Performance Comparison with SOTA", + "text_level": 1, + "bbox": [ + 511, + 784, + 836, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our method effectively acquires extensive image quality assessment knowledge, enabling it to provide powerful priors for various IQA scenarios. Tab. 1 and Tab. 2 summarize the comparative results for different experimental settings.", + "bbox": [ + 511, + 808, + 905, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Few-Shot Setting. Benefiting from the quality priors acquired during pre-training, our model achieves superior", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10447", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4bb84df46314964b4f397b9070359a0fc4cd778a3c1a3aedb34aa43f9d9f90e0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TrainingLIVEFBLIVECKonIQLIVECSIQ
TestingKonIQLIVECKonIQLIVECCSIQLIVE
DBCNN0.7160.7240.7540.7550.7580.877
HyperIAQ0.7580.7350.7720.7850.7440.926
TReS0.7130.7400.7330.7860.761-
DEIQT0.7330.7810.7440.7940.7810.932
CLIP-IQA+0.6310.6200.6970.803--
LoDa0.7630.8050.7450.811--
GRMP-IQA0.7650.7900.7820.8310.8090.935
", + "bbox": [ + 99, + 88, + 478, + 239 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/287f4534f60b502c51a7a1bb1d176688d2533ce1882baaac6ed270d69ec85c81.jpg", + "table_caption": [ + "Table 3. SRCC on the cross datasets validation. The best performances are highlighted in boldface." + ], + "table_footnote": [], + "table_body": "
ComponentLIVECKonIQ
Pre-training Meta-learningPLCCSRCCPLCCSRCC
0.8250.7960.7880.764
✓0.8230.7880.7920.761
✓0.8580.8280.8440.811
", + "bbox": [ + 99, + 287, + 478, + 393 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "performance even with limited training data. As shown in Tab. 1, our GRMP-IQA significantly outperforms the second-best model, GRepQ [40], which is specifically designed for few-shot learning. Additionally, our approach exhibits clear advantages over MetaIQA, a meta-learning-based method, underscoring the effectiveness of metaprompts for rapidly adapting CLIP to diverse IQA tasks. These results validate GRMP-IQA's strong capability in learning quality-aware feature with limited labeled samples. Fully Supervised Setting. Tab. 2 compares GRMP-IQA with other BIQA methods under full supervised setting, including self-training approaches like CONTRIQUE [30] and Re-IQA [37], as well as CLIP-based methods such as LIQE [52] and CLIP-IQA+ [44]. GRMP-IQA outperforms almost all competitors on six datasets. Achieving leading performance across these datasets is particularly challenging due to the diverse range of image content and distortion types. Notably, the proposed method achieves competitive results with the state-of-the-art (SOTA) method on several datasets while utilizing only $20\\%$ of the training data compared to the fully supervised setting.", + "bbox": [ + 89, + 441, + 482, + 760 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Generalization Capability Validation", + "text_level": 1, + "bbox": [ + 89, + 771, + 410, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate the generalization capacity of GRMP-IQA, we performed cross-dataset validation experiments, where the model was trained on one dataset and tested on others without parameter adjustments. Tab. 3 reports the SRCC results across five datasets. GRMP-IQA consistently outperforms state-of-the-art models in most cross Authentic scenarios, achieving notable improvements on LIVEC and", + "bbox": [ + 89, + 794, + 480, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/367facfc5ba92770699ca766737f0998f75c6cd6c90fe95e95dc29bdabc6125d.jpg", + "table_caption": [ + "Table 4. Ablation study on the effectiveness of meta-learning when trained with only 50 samples." + ], + "table_footnote": [], + "table_body": "
ComponentsLIVECKonIQPIPAL
MetaTextVisualPLCCSRCCPLCCSRCCPLCCSRCC
✓0.5790.5980.5920.5730.2160.203
0.6390.5890.5560.5540.3670.371
✓✓0.6990.6890.6790.6090.3230.312
✓✓0.7760.7420.7360.7010.3570.369
✓✓0.7590.7090.6220.5920.3620.396
✓✓✓0.8080.7700.7440.7130.4100.434
", + "bbox": [ + 521, + 89, + 903, + 226 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/53c3767d31b2c45804ecfa50aca6ccc0b0dcf99156eb5963dfa49c4287d0da5d.jpg", + "table_caption": [ + "Table 5. Ablation experiments with Meta-Prompt Pre-training component. The best performances are highlighted in boldface." + ], + "table_footnote": [], + "table_body": "
MethodLabels50100
DatasetPLCCSRCCPLCCSRCC
w/o QGRLIVEC0.8580.8280.8750.848
w/ QGRLIVEC0.8640.8360.8830.857
w/o QGRKonIQ0.8440.8110.8720.840
w/ QGRKonIQ0.8800.8530.8960.872
", + "bbox": [ + 521, + 273, + 903, + 388 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Ablation experiments with QGR in few-shot setting. The best performances are highlighted in boldface.", + "bbox": [ + 511, + 398, + 903, + 426 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "KonIQ dataset. Additionally, it demonstrates strong competitiveness on synthetic datasets like LIVE and CSIQ.", + "bbox": [ + 511, + 441, + 903, + 472 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 484, + 666, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Meta-Prompt Pre-training Module. This module consists of three key components: meta-learning, text meta-prompts, and visual meta-prompts. We conduct ablation studies in Tab. 5 to evaluate their zero-shot capabilities across various datasets. The baseline uses a CLIP model pre-trained on classification tasks. Row 2 evaluates meta-learning without prompts, where CLIP's visual and text encoders are fine-tuned to learn distortion knowledge across different distortions, improving IQA performance. However, tuning CLIP's weights impairs its original generalization, resulting in lower performance on KonIQ. Rows 3 and 4 assess fine-tuning only text or visual prompts during meta-learning. These strategies notably boost zero-shot performance—particularly on real-world datasets like KonIQ and LIVEC—by preserving CLIP's generalization while adapting to IQA tasks. Row 5 explores prompt tuning without meta-learning, underscoring meta-learning's role in curbing overfitting and maintaining generalization. The strongest results emerge when all components are combined, highlighting their complementary strengths.", + "bbox": [ + 509, + 507, + 906, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Meta-learning. To further investigate whether the effectiveness of our method derives from meta-learning, we conducted an ablation study. Specifically, as shown in rows 1 and 2 of Tab. 4, when using Empirical Risk Minimization without meta-learning pre-training on two synthetic dataset, the fine-tuning performance on LIVEC and KonIQ dataset", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10448", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f29ceda765abccace48421cb5381e78efb8f16936ea8a846b45e3ef3fad816a4.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 256, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4637e83833b552599c0a6fd83bb23b03105f64f0926f34abce0c511455680a46.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 258, + 88, + 477, + 209 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/96ea14da1d293d4ceed89937b68951e7d92006334a0604e584873c91385a438b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
λ50100200
PLCCSRCCPLCCSRCCPLCCSRCC
10.8500.8120.8740.8390.8930.866
30.8640.8390.8900.8650.9030.878
50.8800.8530.8960.8720.9080.883
70.8760.8510.8950.8700.9050.881
", + "bbox": [ + 98, + 287, + 478, + 397 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "was even worse than the baseline without pre-training. This indicates that additional pre-training data does not necessarily enhance performance and may even lead to overfitting [19]. In contrast, our meta-learning strategy effectively extracts generalizable quality priors from external data, significantly enhancing performance across datasets.", + "bbox": [ + 89, + 431, + 482, + 521 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Gradient Regularization. We conducted ablation studies on the QGR module. As shown in Tab. 6 and Fig. 4(a), without QGR, models are prone to overfitting under limited training data, with rapid performance saturation and a slight decline, which ultimately hinders generalization. By modulating training gradients, QGR effectively mitigates overfitting and enhances adaptability to the test dataset. To illustrate QGR's impact, we analyzed the angular difference between gradients $G_{\\mathrm{qua}}$ and $G_{\\mathrm{sem}}$ during training, as shown in Fig. 4(b). Without QGR, the angle between $G_{\\mathrm{qua}}$ and $G_{\\mathrm{sem}}$ tends toward 90 degrees, reflecting the orthogonality typical of high-dimensional random vectors [2]. In contrast, QGR introduces greater angular variation, suggesting that it encourages the model to explore more correlations between quality and semantic directions. This improves the model's ability to capture quality-related information and mitigates overfitting on limited data.", + "bbox": [ + 89, + 522, + 482, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of soft weight $\\lambda$ . On the KonIQ dataset, we conduct ablation experiments with various soft weights $\\lambda$ in Eq. 8 to examine their impact, as detailed in Tab. 7. Results show that a small $\\lambda$ diminishes the effectiveness of our QGR, whereas a large $\\lambda$ causes substantial gradient changes and decreases performance. Given our observation of the trade-off, we adopt $\\lambda = 5$ in our experiments.", + "bbox": [ + 89, + 779, + 482, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Computational Analysis. As shown in Tab. 8, we evaluate", + "bbox": [ + 89, + 885, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/87501095bc1cbad1403f79f581372fea43a009bb6e5a41b124fda62fabdce399.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 88, + 612, + 146 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/71c34f6566a07040641767070b78cedbad979b97a9875465b83d34892511aa13.jpg", + "image_caption": [ + "Best ours" + ], + "image_footnote": [], + "bbox": [ + 614, + 89, + 707, + 145 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/041b3ceeb80c7e3722a4fadce55cb5ea41bdf0d8a06bb3f176b5d9a360429e76.jpg", + "image_caption": [ + "Best DEIQT" + ], + "image_footnote": [], + "bbox": [ + 709, + 88, + 805, + 145 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c7b275e19a672c94c75dcb0d0b94938126c7b5072adf2c4b6b1752f36f1a7ce8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 807, + 89, + 901, + 145 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8633e7f956e2a7d65bdaec1d827e8142c963a84bdc4afb8365aeab8aec95dba0.jpg", + "image_caption": [ + "Best ours" + ], + "image_footnote": [], + "bbox": [ + 540, + 155, + 614, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fce79b2080d526aa9879f55f54a010bb29e8181b5170112356030a85b310bb4a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 155, + 663, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bf457257d3440694fb7015a495b185e1c367b944972727510da9b80356ad4ffa.jpg", + "image_caption": [ + "Worst DEIQT" + ], + "image_footnote": [], + "bbox": [ + 709, + 155, + 805, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a5f32b7f336a4237807150d53b768bdcaa132e49b4a4fdb05ac1a487eba338ce.jpg", + "image_caption": [ + "Best DEIQT" + ], + "image_footnote": [], + "bbox": [ + 831, + 155, + 898, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cea95d3a7e6ecbb9a031df5986daea0c42a317cece6f08ee536caded7df2750c.jpg", + "image_caption": [ + "Figure 4. (a) SRCC value during training with 50 samples. (b) Angles between $G_{qua}$ and $G_{sem}$ during training on KonIQ. Under the constraints of QGR, our method captures more correlation between quality knowledge and general semantic knowledge.", + "Worst ours", + "(a)", + "Figure 5. gMAD results between DEIQT [33] and our GRMP-IQA. (a) Fixed DEIQT at low quality. (b) Fixed DEIQT at high quality. (c) Fixed ours at low quality. (d) Fixed ours at high quality." + ], + "image_footnote": [], + "bbox": [ + 516, + 186, + 611, + 243 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d1116aec69a3babc426d88e89cc02072fa4be7f03290cd301f91af0abe9a1ea2.jpg", + "image_caption": [ + "Worst ours", + "(b)" + ], + "image_footnote": [], + "bbox": [ + 612, + 186, + 709, + 243 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c0356403e51cceb9f6c17e0327a993665552725d750b9344ac2f0077f5c804a5.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 710, + 186, + 805, + 243 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cd6d31eb3b740362ca7955d0b9ebeedbb3aaf22da2f466a261beeaaa44b24113.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 805, + 186, + 901, + 243 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/cf9c560dcaecd0bf1c051fca8981d0f619ce278dc8a7a82b127d2cd042c1c997.jpg", + "table_caption": [ + "Table 7. The ablation study about soft weight $\\lambda$ in Eq. 8." + ], + "table_footnote": [], + "table_body": "
MethodParamEfficiencyPerformance
Throughput ↑Latency ↓80% Train20% Train
Q-Align8.2B9.9 (img/sec)101.0 (ms)0.941 / 0.9400.901 / 0.903
Ours151M43.0 (img/sec)21.3 (ms)0.945 / 0.9340.931 / 0.915
", + "bbox": [ + 517, + 306, + 903, + 382 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. Comparison with Q-Align across varying training data ratio to evaluate efficiency and PLCC/SRCC on the KonIQ dataset.", + "bbox": [ + 511, + 392, + 903, + 420 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the inference latency and throughput of our method on an RTX3090 GPU, surpassing Q-Align [45] in both efficiency and accuracy. Notably, our model has only 41M learnable and 151M total parameters, far fewer than Q-Align's 8.2B.", + "bbox": [ + 511, + 430, + 903, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Qualitative Analysis", + "text_level": 1, + "bbox": [ + 511, + 500, + 705, + 515 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The gMAD competition [28] is a standard method for assessing IQA model robustness. It selects image pairs where the attacker model predicts a large quality difference, while the defender perceives them as similar. Observers then evaluate these pairs to gauge model generalization. As shown in Fig. 5, when our model acts as the defender, the attacker's selected pairs exhibit minimal perceptual quality changes. Conversely, as the attacker, our model consistently identifies pairs with significant quality differences. This highlights its strong defensive and offensive capabilities. Notably, in the second column, DEIQT misclassifies semantically similar images as having comparable quality, whereas our model accurately distinguishes their differences.", + "bbox": [ + 511, + 521, + 906, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 726, + 633, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose the GRMP-IQA framework, which generalizes well with limited data. It includes a meta-learning pre-training module that enables the CLIP model to rapidly adapt to IQA tasks and an adaptive gradient regulation module that refines gradient trajectories during fin-tuning, focusing updates on quality-aware knowledge while minimizing the negative impact of over-reliance on semantic noise. Comprehensive experiments on various BIQA datasets validate the superior generalization of our framework, especially in data-scarce scenarios.", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10449", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported by the National Science Fund for Distinguished Young Scholars (No.62025603), the National Natural Science Foundation of China (No. U21B2037, No. U22B2051, No. U23A20383, No. U21A20472, No. 62176222, No. 62176223, No. 62176226, No. 62072386, No. 62072387, No. 62072389, No. 62002305 and No. 62272401), and the Natural Science Foundation of Fujian Province of China (No. 2021J06003, No.2022J06001).", + "bbox": [ + 89, + 90, + 485, + 227 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 239, + 187, + 255 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Nithin C Babu, Vignesh Kannan, and Rajiv Soundararajan. No reference opinion unaware quality assessment of authentically distorted images. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2459-2468, 2023. 2, 3, 5", + "[2] Tony Cai, Jianqing Fan, and Tiefeng Jiang. Distributions of angles in random packing on spheres. The Journal of Machine Learning Research, 14(1):1837-1864, 2013. 8", + "[3] Yuming Fang, Hanwei Zhu, Yan Zeng, Kede Ma, and Zhou Wang. Perceptual quality assessment of smartphone photography. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3677-3686, 2020. 5", + "[4] Yuqian Fu, Yanwei Fu, and Yu-Gang Jiang. Meta-fdmixup: Cross-domain few-shot learning guided by labeled target data. In Proceedings of the 29th ACM international conference on multimedia, pages 5326-5334, 2021. 3", + "[5] Hongcheng Gao, Tianyu Pang, Chao Du, Taihang Hu, Zhijie Deng, and Min Lin. Meta-unlearning on diffusion models: Preventing relearning unlearned concepts. arXiv preprint arXiv:2410.12777, 2024. 1", + "[6] Hongcheng Gao, Jiashu Qu, Jingyi Tang, Baolong Bi, Yue Liu, Hongyu Chen, Li Liang, Li Su, and Qingming Huang. Exploring hallucination of large multimodal models in video understanding: Benchmark, analysis and mitigation. arXiv preprint arXiv:2503.19622, 2025. 1", + "[7] Deepti Ghadiyaram and Alan C Bovik. Massive online crowdsourced study of subjective and objective picture quality. IEEE Transactions on Image Processing, 25(1):372-387, 2015. 5", + "[8] S Alireza Golestaneh, Saba Dadsetan, and Kris M Kitani. No-reference image quality assessment via transformers, relative ranking, and self-consistency. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1220-1230, 2022. 3", + "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3", + "[10] Vlad Hosu, Hanhe Lin, Tamas Sziranyi, and Dietmar Saupe. Koniq-10k: An ecologically valid database for deep learning of blind image quality assessment. IEEE Transactions on Image Processing, 29:4041–4056, 2020. 5", + "[11] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Vi" + ], + "bbox": [ + 93, + 263, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 2, 3", + "[12] Gu Jinjin, Cai Haoming, Chen Haoyu, Ye Xiaoxing, Jimmy S Ren, and Dong Chao. Pipal: a large-scale image quality assessment dataset for perceptual image restoration. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI 16, pages 633-651. Springer, 2020. 5", + "[13] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5148-5157, 2021. 3", + "[14] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19113-19122, 2023. 2, 3", + "[15] Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Self-regulating prompts: Foundational model adaptation without forgetting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15190-15200, 2023. 3", + "[16] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[17] Eric Cooper Larson and Damon Michael Chandler. Most apparent distortion: full-reference image quality assessment and the role of strategy. Journal of electronic imaging, 19 (1):011006, 2010. 5", + "[18] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 2", + "[19] Aobo Li, Jinjian Wu, Yongxu Liu, and Leida Li. Bridging the synthetic-to Authentic gap: Distortion-guided unsupervised domain adaptation for blind image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28422-28431, 2024. 8", + "[20] Chunyi Li, Zicheng Zhang, Haoning Wu, Wei Sun, Xiongkuo Min, Xiaohong Liu, Guangtao Zhai, and Weisi Lin. Agiqa-3k: An open database for ai-generated image quality assessment. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 1", + "[21] Juncheng Li, Minghe Gao, Longhui Wei, Siliang Tang, Wenqiao Zhang, Mengze Li, Wei Ji, Qi Tian, Tat-Seng Chua, and Yueting Zhuang. Gradient-regulated meta-prompt learning for generalizable vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2551-2562, 2023. 3", + "[22] Xudong Li, Timin Gao, Runze Hu, Yan Zhang, Shengchuan Zhang, Xiawu Zheng, Jingyuan Zheng, Yunhang Shen, Ke Li, Yutao Liu, et al. Adaptive feature selection for no-reference image quality assessment by mitigating semantic noise sensitivity. In International Conference on Machine Learning, pages 27808-27821. PMLR, 2024. 2, 3, 6", + "[23] Xudong Li, Runze Hu, Jingyuan Zheng, Yan Zhang, Shengchuan Zhang, Xiawu Zheng, Ke Li, Yunhang Shen," + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10450", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Yutao Liu, Pingyang Dai, et al. Integrating global context contrast and local sensitivity for blind image quality assessment. In International Conference on Machine Learning, pages 27920-27941. PMLR, 2024. 3", + "[24] Xudong Li, Wenjie Nie, Yan Zhang, Runze Hu, Ke Li, Xiawu Zheng, and Liujuan Cao. Distilling spatially-heterogeneous distortion perception for blind image quality assessment. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 2344-2354, 2025. 3", + "[25] Hanhe Lin, Vlad Hosu, and Dietmar Saupe. Kadid-10k: A large-scale artificially distorted iqa database. In 2019 Eleventh International Conference on Quality of Multimedia Experience (QoMEX), pages 1-3. IEEE, 2019. 6", + "[26] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 6", + "[27] Ziwei Luo, Fredrik K Gustafsson, Zheng Zhao, Jens Sjolund, and Thomas B Schon. Controlling vision-language models for multi-task image restoration. In International Conference on Learning Representations, 2024. 2", + "[28] Kede Ma, Qingbo Wu, Zhou Wang, Zhengfang Duanmu, Hongwei Yong, Hongliang Li, and Lei Zhang. Group mad competition-a new methodology to compare objective image quality models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1664-1673, 2016. 8", + "[29] Kede Ma, Xuelin Liu, Yuming Fang, and Eero P Simoncelli. Blind image quality assessment by learning from multiple annotators. In 2019 IEEE international conference on image processing (ICIP), pages 2344-2348. IEEE, 2019. 4", + "[30] Pavan C Madhusudana, Neil Birkbeck, Yilin Wang, Balu Adsumilli, and Alan C Bovik. Image quality assessment using contrastive learning. IEEE Transactions on Image Processing, 31:4149-4161, 2022. 6, 7", + "[31] Jinyoung Park, Juyeon Ko, and Hyunwoo J Kim. Prompt learning via meta-regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26940–26950, 2024. 3", + "[32] Nikolay Ponomarenko, Lina Jin, Oleg Ieremeiev, Vladimir Lukin, Karen Egiazarian, Jaakko Astola, Benoit Vozel, Kacem Chehdi, Marco Carli, Federica Battisti, et al. Image database tid2013: Peculiarities, results and perspectives. Signal processing: Image communication, 30:57-77, 2015. 6", + "[33] Guanyi Qin, Runze Hu, Yutao Liu, Xiawu Zheng, Haotian Liu, Xiu Li, and Yan Zhang. Data-efficient image quality assessment with attention-panel decoder. In Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence, 2023. 2, 3, 6, 8", + "[34] B Qu, H Li, and W Gao. Bringing textual prompt to a-generated image quality assessment, in 2024 IEEE international conference on multimedia and expo (icme), 2024. 1", + "[35] Bowen Qu, Xiaoyu Liang, Shangkun Sun, and Wei Gao. Exploring aigc video quality: A focus on visual harmony videotext consistency and domain distribution gap. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6652-6660, 2024. 1" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 6", + "[37] Avinab Saha, Sandeep Mishra, and Alan C Bovik. Reiqa: Unsupervised learning for image quality assessment in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5846-5855, 2023. 2, 3, 6, 7", + "[38] Hamid R Sheikh, Muhammad F Sabir, and Alan C Bovik. A statistical evaluation of recent full reference image quality assessment algorithms. IEEE Transactions on image processing, 15(11):3440-3451, 2006. 1, 5", + "[39] Tianshu Song, Leida Li, Deqiang Cheng, Pengfei Chen, and Jinjian Wu. Active learning-based sample selection for label-efficient blind image quality assessment. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 1", + "[40] Suhas Srinath, Shankhanil Mitra, Shika Rao, and Rajiv Soundararajan. Learning generalizable perceptual representations for data-efficient no-reference image quality assessment. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 22-31, 2024. 2, 3, 6, 7", + "[41] Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3667-3676, 2020. 6", + "[42] Zhengzhong Tu, Yilin Wang, Neil Birkbeck, Balu Adsumilli, and Alan C Bovik. Ugc-vqa: Benchmarking blind video quality assessment for user generated content. IEEE Transactions on Image Processing, 30:4449-4464, 2021. 1", + "[43] Joaquin Vanschoren. Meta-learning: A survey. arXiv preprint arXiv:1810.03548, 2018. 4", + "[44] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 2, 3, 6, 7", + "[45] Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Yixuan Gao, Annan Wang, Erli Zhang, Wenxiu Sun, et al. Q-align: Teaching lmmss for visual scoring via discrete text-defined levels. In International Conference on Machine Learning, pages 54015-54029. PMLR, 2024. 2, 8", + "[46] Kangmin Xu, Liang Liao, Jing Xiao, Chaofeng Chen, Haoning Wu, Qiong Yan, and Weisi Lin. Boosting image quality assessment through efficient transformer adaptation with local feature enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2662-2672, 2024. 2, 3, 6", + "[47] Zhengzhuo Xu, Bowen Qu, Yiyan Qi, Sinan Du, Chengjin Xu, Chun Yuan, and Jian Guo. Chartmoe: Mixture of diversely aligned expert connector for chart understanding. arXiv preprint arXiv:2409.03277, 2024. 3", + "[48] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang." + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10451", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Maniqa: Multi-dimension attention network for no-reference image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1191-1200, 2022. 1, 6", + "[49] Zhenqiang Ying, Haoran Niu, Praful Gupta, Dhruv Mahajan, Deepti Ghadiyaram, and Alan Bovik. From patches to pictures (paq-2-piq): Mapping the perceptual space of picture quality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3575–3585, 2020. 5", + "[50] Ji Zhang, Shihan Wu, Lianli Gao, Heng Tao Shen, and Jingkuan Song. Dept: Decoupled prompt tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12924-12933, 2024. 3", + "[51] Weixia Zhang, Kede Ma, Jia Yan, Dexiang Deng, and Zhou Wang. Blind image quality assessment using a deep bilinear convolutional neural network. IEEE Transactions on Circuits and Systems for Video Technology, 30(1):36-47, 2018. 3", + "[52] Weixia Zhang, Guangtao Zhai, Ying Wei, Xiaokang Yang, and Kede Ma. Blind image quality assessment via vision-language correspondence: A multitask learning perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14071-14081, 2023. 2, 3, 5, 6, 7", + "[53] Kai Zhao, Kun Yuan, Ming Sun, Mading Li, and Xing Wen. Quality-aware pre-trained models for blind image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22302-22313, 2023. 2, 3, 5", + "[54] Limin Zheng, Yu Luo, Zihan Zhou, Jie Ling, and Guanghui Yue. Cdinet: Content distortion interaction network for blind image quality assessment. IEEE Transactions on Multimedia, 2024. 6", + "[55] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16816-16825, 2022. 3", + "[56] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 2, 3, 4, 5, 6", + "[57] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15659-15669, 2023. 2", + "[58] Hancheng Zhu, Leida Li, Jinjian Wu, Weisheng Dong, and Guangming Shi. Metaiaq: Deep meta-learning for no-reference image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14143-14152, 2020. 3, 4, 6" + ], + "bbox": [ + 91, + 90, + 483, + 824 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "10452", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Few-Shot Image Quality Assessment via Adaptation of Vision-Language Models/89cdaf16-5828-4989-b825-83012badf535_model.json b/2025/Few-Shot Image Quality Assessment via Adaptation of Vision-Language Models/89cdaf16-5828-4989-b825-83012badf535_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b1860ff64b9aaf935a8b215ab7d48691e7e9c11f --- /dev/null +++ b/2025/Few-Shot Image Quality Assessment via Adaptation of Vision-Language Models/89cdaf16-5828-4989-b825-83012badf535_model.json @@ -0,0 +1,2554 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.097, + 0.131, + 0.903, + 0.154 + ], + "angle": 0, + "content": "Few-Shot Image Quality Assessment via Adaptation of Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.201, + 0.18, + 0.796, + 0.218 + ], + "angle": 0, + "content": "Xudong Li\\(^{1*}\\) Zihao Huang\\(^{2*}\\) Yan Zhang\\(^{1\\dagger}\\) Yunhang Shen\\(^{3}\\) Ke Li\\(^{3}\\) Xiawu Zheng\\(^{1}\\) Liujuan Cao\\(^{1}\\) Rongrong Ji\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.193, + 0.217, + 0.804, + 0.252 + ], + "angle": 0, + "content": "1 Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University, 361005, P.R. China" + }, + { + "type": "text", + "bbox": [ + 0.156, + 0.251, + 0.832, + 0.287 + ], + "angle": 0, + "content": "\\(^{2}\\) Beijing Institute of Technology \\(^{3}\\) Tencent Youtu Lab, Shanghai, China {1xd761050753, huangzihhhh, bzhy986}@gmail.com, {zhengxiawu, rrji}@xmu.edu.cn" + }, + { + "type": "list", + "bbox": [ + 0.156, + 0.217, + 0.832, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.337, + 0.327, + 0.354 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.369, + 0.485, + 0.761 + ], + "angle": 0, + "content": "Image Quality Assessment (IQA) remains an unresolved challenge in computer vision due to complex distortions, diverse image content, and limited data availability. Existing Blind IQA (BIQA) methods largely rely on extensive human annotations, which are labor-intensive and costly due to the demanding nature of creating IQA datasets. To reduce this dependency, we propose the Gradient-Regulated MetaPrompt IQA Framework (GRMP-IQA), designed to efficiently adapt the visual-language pre-trained model, CLIP, to IQA tasks, achieving high accuracy even with limited data. GRMP-IQA consists of two core modules: (i) MetaPrompt Pre-training Module and (ii) Quality-Aware Gradient Regularization. The Meta Prompt Pre-training Module leverages a meta-learning paradigm to pre-train soft prompts with shared meta-knowledge across different distortions, enabling rapid adaptation to various IQA tasks. On the other hand, the Quality-Aware Gradient Regularization is designed to adjust the update gradients during fine-tuning, focusing the model's attention on quality-relevant features and preventing overfitting to semantic information. Extensive experiments on standard BIQA datasets demonstrate the superior performance to the state-of-the-art BIQA methods under limited data setting. Notably, utilizing just \\(20\\%\\) of the training data, GRMP-IQA is competitive with most existing fully supervised BIQA approaches. Our code is available via https://github.com/LXDxmu/GRMP-IQA." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.791, + 0.222, + 0.807 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.817, + 0.484, + 0.879 + ], + "angle": 0, + "content": "With the rise of the mobile internet era, the focus on computer vision has transitioned from initial concerns with compression and image processing [38] to handling user-generated content like smartphone photos and videos [6," + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.323, + 0.903, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.59, + 0.522, + 0.833, + 0.533 + ], + "angle": 0, + "content": "(b) Fine-tuning : Quality-Aware Gradient Regularization" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.544, + 0.907, + 0.697 + ], + "angle": 0, + "content": "Figure 1. Intuitive diagram of GRMP-IQA. (a) demonstrates using meta-learning for efficient soft prompt initialization with quality prior that enhances zero-shot generalization (\\(\\uparrow\\) \\(24\\%\\)), enabling the CLIP to adapt to BIQA tasks effectively. (b) illustrates gradient regularization during fine-tuning, which clips quality gradients aligned with semantic bias to guide the model toward quality-relevant features and reduce overly reliance on semantic content. In the t-SNE visualization, black dots represent noise samples with high semantic confidence (confidence \\(>0.8\\)) but incorrect quality predictions. The noticeable decrease (\\(\\downarrow\\) \\(73\\%\\)) in black dots highlights reduced semantic noise interference in IQA tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.903 + ], + "angle": 0, + "content": "35, 42], and lately to AI-generated content [5, 20]. This evolution has sharply increased demand for effective Blind Image Quality Assessment (BIQA) techniques, highlighting the importance of developing methodologies that can adeptly evaluate image quality without reference images. Data-driven BIQA models [34, 39, 48] based on deep neural networks have made significant progress in recent years. However, the quality scores for distorted images are often measured using the Mean Opinion Score (MOS), which is the average of multiple ratings (sometimes up to 120). As a result, acquiring a sufficient number of IQA training samples is quite labor-intensive and cost-expensive." + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.887, + 0.371, + 0.902 + ], + "angle": 0, + "content": "* Equal contribution. † Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10442" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.455 + ], + "angle": 0, + "content": "To mitigate this challenge, recent BIQA approaches [33, 37, 40, 46, 53] have leveraged large-scale image datasets through transfer learning or self-supervised learning. While these approaches improve performance, they often rely on computationally intensive pretraining or require domain-specific pretext tasks, limiting scalability. On the other hand, the vision-language model CLIP [36] has shown remarkable generalization across various downstream tasks, offering a potential alternative. Specifically, CLIP-IQA [44] demonstrates promising zero-shot performance in IQA by leveraging simple handcrafted prompts (e.g., \"Good/Bad photo\"). However, a significant performance gap remains compared to fully supervised methods [40, 45], highlighting the need to leverage limited labeled data to further close this performance gap. To bridge this gap, we introduce the concept of Few-Shot IQA, which aims to enable models to quickly generalize to new distortion types and unseen scenarios with only a small amount of labeled data (e.g., 50-200 training images). A natural way to adapt CLIP to downstream IQA tasks under this setup is through prompt tuning [11, 56, 57]. Prompt tuning techniques learn soft textual prompts (continuous embeddings) from a small labeled dataset while keeping the pretrained model parameters frozen, offering a lightweight and efficient solution." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.457, + 0.485, + 0.655 + ], + "angle": 0, + "content": "However, as shown in Tab. 1, the direct application of the prompt tuning method CoOp [56] to IQA yields minimal improvement over simple linear probing on the CLIP image encoder, falling short of the expected benefits seen in high-level vision tasks. We attribute this limitation to a fundamental mismatch between CLIP's pretraining objectives and the specific requirements of IQA tasks. CLIP is pretrained to align text and images in a shared semantic space, with a focus on high-level semantics such as objects and scenes. In contrast, IQA demands fine-grained sensitivity to low-level distortion cues—an ability in which CLIP is relatively deficient [27, 52]. This discrepancy introduces two key challenges when applying prompt tuning to IQA tasks:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.656, + 0.484, + 0.809 + ], + "angle": 0, + "content": "(1) Sensitivity to Initialization: During prompt tuning, with CLIP's parameters frozen and limited sensitivity to distortions, the model depends heavily on the learned prompts for quality assessment. Different prompt initializations lead to distinct optimization directions [18], which may steer optimization toward features less relevant to quality, causing significant performance variability. As shown in Fig. 2, the average SRCC fluctuates considerably across different random initializations, requiring careful tuning for each IQA scenario and limiting the model's adaptability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "(2) Overfitting from Semantic Bias: Due to CLIP's inherent reliance on semantic information, fine-tuning with a small sample size often leads to overfitting, where the model learns spurious correlations (e.g., over-reliance on semantic features to infer quality [1, 22]) rather than true image quality indicators, which degrades IQA performance." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.656, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.088, + 0.732, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.733, + 0.089, + 0.905, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.21, + 0.907, + 0.267 + ], + "angle": 0, + "content": "Figure 2. The fine-tuning results on the LIVEC dataset using CoOp [56] show that prompt tuning accuracy is highly sensitive to initial random initialization of the prompt, especially with limited training data, underscoring its critical role under data constraints." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.281, + 0.907, + 0.356 + ], + "angle": 0, + "content": "As illustrated in Fig. 1, when the CLIP encoder is finetuned on a limited dataset, the latent space contains many instances (black dots) that confidently and accurately predict image semantics (with confidence levels above 0.8) but fail to reliably assess image quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.357, + 0.909, + 0.78 + ], + "angle": 0, + "content": "To overcome these limitations, we propose Gradient-Regulated Meta-Prompt Learning for IQA (GRMP-IQA), a framework with two key components: (i) Meta Prompt Pre-training (MPP) mitigates sensitivity to prompt initialization by incorporating generalized distortion priors into the initialization of soft prompts, which enhances their adaptability to new IQA scenarios. By designing metatraining tasks with a well-annotated dataset containing diverse distortions (e.g., overexposure, blur) and optimizing prompts through bi-level gradient descent, we enable the model to learn shared meta-knowledge of quality representations across different distortions, thereby improving generalization across IQA tasks. (ii) Quality-Aware Gradient Regularization (QGR) addresses overfitting to semantic information by regulating gradient updates during fine-tuning, which balances the interplay between the quality knowledge gradient \\( G_{\\mathrm{qua}} \\) and the semantic knowledge gradient \\( G_{\\mathrm{sem}} \\). By clipping \\( G_{\\mathrm{qua}} \\) along \\( G_{\\mathrm{sem}} \\) when their directions are overly aligned, this module suppresses semantic bias and guides the model to focus on distortion-related features, improving its ability to capture low-level quality details. Additionally, fine-tuning only CLIP's text branch can misalign quality perception between the image and text branches, limiting generalization [14]. To address this, we integrate text prompt tuning (CoOp [56]) and visual prompt tuning (VPT [11]) by jointly meta-learning initializations for both, ensuring complementary optimization to better adapt CLIP to new IQA scenarios (Tab. 5)." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.78, + 0.859, + 0.794 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.796, + 0.905, + 0.87 + ], + "angle": 0, + "content": "- We propose a Meta Prompt Pre-training method that organizes meta tasks based on image distortion types and optimizes soft textual and visual prompts to gain shared quality meta-knowledge, enabling CLIP to rapidly adapt across various IQA scenarios." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.871, + 0.905, + 0.902 + ], + "angle": 0, + "content": "- We develop a novel Quality-Aware Gradient Regularization method that clips gradients aligned with semantic" + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.796, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10443" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.103, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "directions to balance semantic and quality information during fine-tuning, ensuring the model prioritizes image quality while still integrating relevant semantic context." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.138, + 0.483, + 0.198 + ], + "angle": 0, + "content": "- Extensive empirical results confirm that our approach is both effective and efficient. Notably, with just 200 data samples, our method is competitive with SOTA models on the LIVEC dataset (using \\(20\\%\\) of the training data)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.213, + 0.233, + 0.229 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.239, + 0.411, + 0.256 + ], + "angle": 0, + "content": "2.1. Deep Learning Based BIQA methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.262, + 0.483, + 0.593 + ], + "angle": 0, + "content": "Early CNN-based IQA methods followed standard pretraining and fine-tuning pipelines [9, 51], while meta-learning approaches like MetaIQA [58] improved adaptability from synthetic to real-world images. However, CNNs struggle with non-local features, a gap addressed by ViT-based models [8, 13, 23, 24]. Recent methods like LIQE [52] leverage CLIP's multi-task learning, enhancing IQA through supervised fine-tuning across datasets. While effective, these methods rely heavily on extensive annotations, making them costly and time-intensive. In contrast, limited-data BIQA remains underexplored. CLIP-IQA[44] shows promise in zero-shot settings but has performance limits, highlighting the need for better use of limited labeled data. Methods like DEIQT[33] and LoDa[46] demonstrate that fine-tuning pre-trained ViTs reduces annotation needs but still underperform in few-shot scenarios. Self-supervised approaches[37, 40, 53] ease data constraints through tailored pre-training tasks but carry high computational costs. Additionally, recent studies [1, 22] explore disentangling semantic and distortion-related content for IQA—yet none fully address the dual challenges of data efficiency and computational cost." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.595, + 0.483, + 0.686 + ], + "angle": 0, + "content": "To bridge these gaps, our approach combines CLIP's pre-trained semantic knowledge with distortion-specific fine-tuning. By using prompt tuning to cut pre-training costs and a parameter-free regularization strategy (leveraging high- and low-level gradient correlations), it optimizes IQA more effectively with fewer labeled samples." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.697, + 0.245, + 0.714 + ], + "angle": 0, + "content": "2.2. Prompt Tuning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Given the limited labeled data during training, prompt tuning is an effective approach to adapt vision-language pre-trained models for few-shot learning tasks [4, 47, 50]. CoOp [56] optimizes prompt vectors in CLIP's language branch for task adaptation but struggles with generalization to novel tasks. To address this, CoCoOp [55] introduces a lightweight meta-network to generate input-conditioned tokens, improving adaptability. Beyond textual prompts, visual prompts have also been explored for task adaptation [11]. Methods like MaPLe [14] and PromptSRC [15] incorporate trainable prompts into both language and visual branches, achieving significant performance gains across" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.228 + ], + "angle": 0, + "content": "base and novel tasks. Recent studies [21, 31] further construct meta-tasks based on semantic categories to initialize well-structured soft prompts, mitigating generalization decline in high-level tasks. Unlike these approaches that focus on high-level tasks, our proposed prompt tuning method is specifically designed for IQA, leveraging low-level distortion information to construct meta-tasks. This enables pretrained prompts to be distortion-sensitive and effectively capture shared quality knowledge." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.243, + 0.649, + 0.26 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.269, + 0.625, + 0.284 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.291, + 0.905, + 0.457 + ], + "angle": 0, + "content": "In this paper, we propose the Gradient-Regulated MetaPrompt Image Quality Assessment (GRMP-IQA), which aims to adapt the CLIP for BIQA tasks with a few training samples. As depicted in Fig. 3, GRMP-IQA consists of two primary modules: (i) the Meta-Prompt Pre-training Module (MPP) and (ii) the Quality-Aware Gradient Regularization (QGR). The MPP module pre-trains visual-text prompts to acquire shared meta-knowledge on distortions, enabling quick adaptation to various IQA scenarios. The QGR module plays a key role in fine-tuning by adjusting gradient updates to prevent overfitting to semantic content." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.459, + 0.905, + 0.564 + ], + "angle": 0, + "content": "Pre-training stage (Sec. 3.3). We randomly sample a minibatch from distortion meta-tasks, partitioning it into a support set \\(\\mathcal{D}^{\\mathrm{support}}\\) and a query set \\(\\mathcal{D}^{\\mathrm{query}}\\). A bi-level gradient descent method progresses from the Inner-Loop on \\(\\mathcal{D}^{\\mathrm{support}}\\) to the Outer-Loop on \\(\\mathcal{D}^{\\mathrm{query}}\\) to optimize the learnable visual-textual prompts \\([\\theta_T, \\theta_V]\\). These prompts are then used as initial weights for following fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.906, + 0.762 + ], + "angle": 0, + "content": "Fine-tuning stage (Sec. 3.4). Given an input image \\( \\mathbf{x} \\) and a hard semantic prompt \\( \\boldsymbol{w}_i = \\) \"a photo of a [class]\", we predict the semantic class distributions \\( p^{\\text{sem}}(\\boldsymbol{w}_i|\\boldsymbol{x}) \\) and \\( p^{\\text{qua}}(\\boldsymbol{w}_i|\\boldsymbol{x}) \\) using the original semantic CLIP model and the fine-tuned CLIP model, respectively. The gradient of the KL divergence loss \\( \\nabla \\mathcal{L}_{kl} \\) between these distributions is computed as the general semantic direction \\( G_{\\text{sem}} \\). Simultaneously, using soft quality prompt \\( \\theta_T \\) and image \\( \\mathbf{x} \\), we compute the gradient \\( \\nabla \\mathcal{L}_{ce} \\) of the quality loss as the quality direction \\( G_{\\text{qua}} \\). To refine the optimization direction of the IQA task, we adjust \\( G_{\\text{qua}} \\) into \\( G_{\\text{qgr}} \\) by clipping its components aligned with \\( G_{\\text{sem}} \\), ensuring the model focuses on image quality while reducing the impact of semantic noise." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.772, + 0.747, + 0.789 + ], + "angle": 0, + "content": "3.2. Visual-Text Meta-Prompt" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Visual Meta-Prompt. We utilize Deep Prompt Tuning (DPT) [56] as our Visual Meta-Prompt, with learnable parameters \\(\\theta_{\\mathbf{V}}\\). The input embedding for the \\(l\\)-th layer's self-attention module in the ViT-based image encoder is denoted as \\(\\{\\pmb{f}^l,\\pmb{H}^l\\}\\), where \\(\\pmb{f}^l\\) represents the classification (CLS) token, and \\(\\pmb{H}^{l} = \\{\\pmb{h}_{1}^{l},\\pmb{h}_{2}^{l},\\dots,\\pmb{h}_{N}^{l}\\}\\) denotes the image patch embeddings. A learnable token \\(\\pmb{P}^{l}\\) is appended to" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10444" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.087, + 0.482, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.29, + 0.452, + 0.305 + ], + "angle": 0, + "content": "(a) Pre-training stage: Meta-Prompt Pre-training Module" + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.089, + 0.905, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.523, + 0.29, + 0.89, + 0.305 + ], + "angle": 0, + "content": "(b) Fine-tuning stage: Quality-Aware Gradient Regularization" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.315, + 0.907, + 0.413 + ], + "angle": 0, + "content": "Figure 3. The overview of our GRMP-IQA. The core modules are the Meta-Prompt Pre-training and Quality-Aware Gradient Regularization, corresponding to two training processes. At the pre-training stage, we establish a distortion-specific meta-knowledge task for BIQA tasks, and the bi-level gradient descent is utilized to train visual-text meta-prompt \\([\\theta_T,\\theta_V]\\) (Sec. 3.3). These optimized prompts then serve as initial settings to efficiently adapt the CLIP model to IQA tasks with limited labels. At the few-shot fine-tuning stage, we first predict the class probability distributions \\(p^{sem}(w_i|x)\\) and \\(p^{qua}(w_i|x)\\) using the semantic and fine-tuned CLIP models, respectively. We then adjust IQA task gradients \\(G_{\\mathrm{qua}}\\) by clipping gradients aligned with semantic task gradients \\(G_{\\mathrm{sem}}\\) to generate refined gradient \\(G_{\\mathrm{qgr}}\\) for backward updates (Sec. 3.4), mitigating the influence of semantic noise on quality predictions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.427, + 0.483, + 0.457 + ], + "angle": 0, + "content": "the token sequence in each ViT layer, and the Multi-Head Self-Attention (MHSA) module processes the tokens as:" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.464, + 0.483, + 0.483 + ], + "angle": 0, + "content": "\\[\n[ \\boldsymbol {f} ^ {l}, \\neg , \\boldsymbol {H} ^ {l} ] = \\operatorname {L a y e r} ^ {l} ([ \\boldsymbol {f} ^ {l - 1}, \\boldsymbol {P} ^ {l - 1}, \\boldsymbol {H} ^ {l - 1} ]), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.491, + 0.483, + 0.521 + ], + "angle": 0, + "content": "where the output of \\( P^l \\) is discarded and not passed to the next layer, serving only as a set of learnable parameters." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.521, + 0.483, + 0.627 + ], + "angle": 0, + "content": "Text Meta-Prompt. We adopt CoOp [56] as our textual meta-prompt, with learnable parameters \\(\\theta_T\\). For each quality class \\(i\\), we construct a learnable quality prompt defined as \\(\\pmb{q}_i = \\{\\pmb{v}_1, \\pmb{v}_2, \\dots, \\pmb{v}_M, [\\mathrm{quality}]\\}\\). Here, [quality] \\(\\in\\) {\"high quality\", \"low quality\"} serves as a categorical quality marker, and \\(\\{\\pmb{v}_m\\}_{m=1}^M\\) denotes a set of \\(M\\) learnable context vectors (with \\(M = 4\\) in our work)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.635, + 0.395, + 0.65 + ], + "angle": 0, + "content": "3.3. Meta-Prompt Pre-training Module" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.657, + 0.483, + 0.734 + ], + "angle": 0, + "content": "During meta-learning, we aim to optimize the learnable visual-text prompt \\(\\theta = [\\theta_T,\\theta_V]\\). For an input image \\(x\\), the textual encoder \\(g(\\cdot)\\) processes the quality prompts \\(q_{i}\\) while the visual encoder extracts the feature vector \\(f\\). The probability of predicting high quality is:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.74, + 0.487, + 0.786 + ], + "angle": 0, + "content": "\\[\np \\left(\\boldsymbol {q} _ {\\text {h i g h}} \\mid \\boldsymbol {x}\\right) = \\frac {\\exp \\left(\\left\\langle \\boldsymbol {g} \\left(\\boldsymbol {q} _ {\\text {h i g h}}\\right) , \\boldsymbol {f} \\right\\rangle / \\tau\\right)}{\\exp \\left(\\left\\langle \\boldsymbol {g} \\left(\\boldsymbol {q} _ {\\text {h i g h}}\\right) , \\boldsymbol {f} \\right\\rangle / \\tau\\right) + \\exp \\left(\\left\\langle \\boldsymbol {g} \\left(\\boldsymbol {q} _ {\\text {l o w}}\\right) , \\boldsymbol {f} \\right\\rangle / \\tau\\right)},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.787, + 0.484, + 0.877 + ], + "angle": 0, + "content": "where \\(q_{\\mathrm{high}}\\) and \\(q_{\\mathrm{low}}\\) denote high quality and low quality category prompt, \\(p(q_{\\mathrm{high}}|\\boldsymbol {x})\\) represents the estimated probability that the quality of image \\(\\mathbf{x}\\) is \"high quality\", \\(\\tau\\) is a temperature parameter learned by CLIP, and \\(\\langle ,\\rangle\\) denotes cosine similarity. The labeled quality scores are then rescaled to 0-1, denoted as \\(y\\), and the loss function is calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.885, + 0.483, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = - (y \\log (p \\left(q _ {\\text {h i g h}} | \\boldsymbol {x}\\right)) + (1 - y) \\log (1 - p \\left(q _ {\\text {h i g h}} | \\boldsymbol {x}\\right)). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.427, + 0.907, + 0.794 + ], + "angle": 0, + "content": "Constructing Distortion Meta-Knowledge Task. As noted by [29, 58], the ability to detect various types of image distortions is crucial for developing BIQA models with strong generalization capabilities across diverse scenarios. Moreover, the efficacy of prompt tuning is significantly dependent on the initial configuration of the prompts. This initial setup greatly affects the CLIP vision-language model's ability to swiftly adapt to different IQA scenarios. Drawing inspiration from the \"learn to learn\" ethos inherent in the deep meta-learning paradigm [43, 58], we propose an optimization-based method for effectively pre-training visual-textual prompts. These prompts incorporate shared quality insights from various image distortions, thereby enhancing CLIP's swift adaptability to IQA tasks [58]. To investigate the general rules of image distortion, we first constructed a \\(K_{t}\\)-way distortion-specific image quality prediction task, denoted as \\(\\mathcal{T}_t\\). Then, it is used to build the meta-training set as \\(\\mathcal{D}^{\\mathrm{meta}} = \\left\\{\\mathcal{D}_{\\mathcal{T}_t}^{\\mathrm{support}}, \\mathcal{D}_{\\mathcal{T}_t}^{\\mathrm{query}}\\right\\}_{t=1}^T\\). Here, \\(\\mathcal{D}_{\\mathcal{T}_t}^{\\mathrm{support}}\\) and \\(\\mathcal{D}_{\\mathcal{T}_t}^{\\mathrm{query}}\\) represent the support and query sets for each task, respectively, with \\(T\\) representing the total number of tasks. To simulate the process of prompt generalization to different distortions in BIQA scenarios, we randomly sample \\(k\\) tasks as a mini-batch from the meta-training set, where \\(1 \\leq k \\leq T\\), to perform bi-level gradient optimization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Distortion-Aware Meta-prompt Learning. Our approach employs a bi-level gradient descent technique to bridge the learning process from the support to the query set. Specifically, it mainly consists of two optimization steps. In the Inner-Loop (the first level), we compute the gradients of the prompt parameters using the support set and apply the first update. In the Outer-Loop (the second level), we assess the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10445" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.167 + ], + "angle": 0, + "content": "performance of the updated model on the query set and optimize the parameters again. This bi-level structure distills shared quality priors by training on a number of NR-IQA tasks with known distortions, enabling the meta-prompt to generalize rapidly across diverse new BIQA scenarios." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.168, + 0.482, + 0.262 + ], + "angle": 0, + "content": "Inner-Loop. The objective of the Inner-Loop stage is to adapt the meta-prompt, denoted by \\([\\pmb{\\theta}_{\\pmb{T}},\\pmb{\\theta}_{\\pmb{V}}]\\), to the \\(t^{th}\\) support set \\(\\mathcal{D}_{\\mathcal{T}_t}^{\\mathrm{support}}\\) within the mini-batch. During the first level of updates, we determine the loss \\(\\mathcal{L}(\\pmb {\\theta},\\mathcal{D}_{\\mathcal{T}_t}^{\\mathrm{support}})\\), following which the model parameters are updated on the support set using the inner learning rate \\(\\alpha\\), as specified by:" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.273, + 0.483, + 0.293 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\theta} _ {t} ^ {\\prime} = \\boldsymbol {\\theta} - \\alpha \\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} (\\boldsymbol {\\theta}, \\mathcal {D} _ {\\mathcal {T} _ {t}} ^ {\\text {s u p p o r t}}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.304, + 0.482, + 0.336 + ], + "angle": 0, + "content": "Outer-Loop. In a similar vein, the second level of updates adjusts the parameters \\(\\theta_t^\\prime\\) based on the query set \\(\\mathcal{D}_{\\mathcal{T}_t}^{\\mathrm{query}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.348, + 0.483, + 0.367 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\theta} _ {t} \\leftarrow \\boldsymbol {\\theta} _ {t} ^ {\\prime} - \\alpha \\nabla_ {\\boldsymbol {\\theta} _ {t} ^ {\\prime}} \\mathcal {L} \\left(\\boldsymbol {\\theta} _ {t} ^ {\\prime}, \\mathcal {D} _ {\\mathcal {T} _ {t}} ^ {\\text {q u e r y}}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.378, + 0.483, + 0.424 + ], + "angle": 0, + "content": "For a mini-batch of meta-tasks, this process culminates in aggregating gradients from all tasks to update the final model parameters, following the update rule:" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.437, + 0.483, + 0.477 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\theta} \\leftarrow \\boldsymbol {\\theta} - \\beta \\sum_ {t = 1} ^ {k} \\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} \\left(\\boldsymbol {\\theta} _ {t} ^ {\\prime}, \\mathcal {D} _ {\\mathcal {T} _ {t}} ^ {\\text {q u e r y}}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.489, + 0.483, + 0.535 + ], + "angle": 0, + "content": "where \\(\\beta\\) represents the outer learning rate. Meta-learning effectively trains the learnable prompts \\(\\pmb{\\theta} = [\\pmb{\\theta}_T, \\pmb{\\theta}_V]\\), ensuring their generalization across various image distortions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.546, + 0.433, + 0.561 + ], + "angle": 0, + "content": "3.4. Quality-Aware Gradient Regularization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.483, + 0.749 + ], + "angle": 0, + "content": "We denote the original semantic CLIP model as \\( V^{\\mathrm{sem}} \\) and the model obtained from the Meta-Prompt Pre-training Module, as \\( V^{\\mathrm{qua}} \\). The gradient for the IQA task is the quality gradient \\( G_{\\mathrm{qua}} \\), while the gradient for the semantic direction is the semantic gradient \\( G_{\\mathrm{sem}} \\). Recent studies [1, 53] have shown that IQA tasks can be misaligned with the high-level semantic representations of upstream tasks, leading to overfitting and reduced generalization. To prevent semantic overfitting, our Quality-aware Gradient Regularization (QGR) follows three key steps: derive a semantic direction \\( G_{\\mathrm{sem}} \\), compute the quality optimization direction \\( G_{\\mathrm{qua}} \\), and balance gradients based on their directional relationship." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.751, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Specifically, to derive a general semantic direction \\( G_{\\mathrm{sem}} \\), we first design a hard prompt following [49, 52]: \\( \\boldsymbol{w}_i = \\) \"a photo of a [class]\", where [class] represents one of nine categories: \"animal\", \"cityscape\", \"human\", \"indoor\", \"landscape\", \"night\", \"plant\", \"still-life\", \"others\". This prompt \\( \\boldsymbol{w}_i \\) is input into the text encoders of \\( V^{\\mathrm{sem}} \\) and \\( V^{\\mathrm{qua}} \\), generating text features that align with image features \\( f^{qua} \\) from their respective visual encoders. As a result, we obtain zero-shot semantic prediction probabilities \\( p^{\\mathrm{sem}}(\\boldsymbol{w}_i|\\boldsymbol{x}) \\) and \\( p^{\\mathrm{qua}}(\\boldsymbol{w}_i|\\boldsymbol{x}) \\). To measure the semantic alignment between" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.092, + 0.834, + 0.107 + ], + "angle": 0, + "content": "\\(V^{\\mathrm{qua}}\\) and \\(V^{\\mathrm{sem}}\\), we compute the KL divergence:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.115, + 0.905, + 0.151 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {k l}} \\left(\\boldsymbol {V} ^ {\\text {q u a}}\\right) = - \\sum_ {i} \\boldsymbol {p} ^ {\\text {s e m}} \\left(\\boldsymbol {w} _ {i} \\mid \\boldsymbol {x}\\right) \\log \\frac {\\boldsymbol {p} ^ {\\text {q u a}} \\left(\\boldsymbol {w} _ {i} \\mid \\boldsymbol {x}\\right)}{\\boldsymbol {p} ^ {\\text {s e m}} \\left(\\boldsymbol {w} _ {i} \\mid \\boldsymbol {x}\\right)}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.162, + 0.905, + 0.387 + ], + "angle": 0, + "content": "The gradient of \\(\\mathcal{L}_{\\mathrm{kl}}(V^{\\mathrm{qua}})\\), denoted as \\(G_{\\mathrm{sem}}\\), represents the general semantic optimization direction. Similarly, the quality direction \\(G_{\\mathrm{qua}}\\) is obtained by computing the gradient of the cross-entropy loss \\(\\mathcal{L}_{\\mathrm{ce}}(V^{\\mathrm{qua}})\\), which compares the predicted probability \\(p(q_{\\mathrm{high}}|x)\\) with the ground truth \\(y\\), as shown in Eq. 3. Finally, to balance the gradients between quality \\(G_{\\mathrm{qua}}\\) and semantics \\(G_{\\mathrm{sem}}\\) within a shared representation space, we analyze their relationship under two cases: (1) If the angle between \\(G_{\\mathrm{qua}}\\) and \\(G_{\\mathrm{sem}}\\) is less than 90 degrees, it indicates that the optimization directions for IQA quality knowledge and general semantic knowledge are consistent. In such cases, we clip the \\(G_{\\mathrm{qua}}\\) along its component \\(G_{\\parallel}\\) that parallel to the semantic direction \\(G_{\\mathrm{sem}}\\), to modulate the model's original quality optimization path, preventing overfitting to semantic correlations." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.388, + 0.905, + 0.493 + ], + "angle": 0, + "content": "(2) Conversely, if the angle between two areas of knowledge is more than 90 degrees, it means they are heading in different directions during improvement. In such situations, we don't adjust the quality gradient \\( G_{\\mathrm{qua}} \\), allowing the model to learn quality-aware features according to the original optimization direction. In conclusion, our gradient regularization strategy is mathematically formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.503, + 0.905, + 0.556 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {G} _ {\\text {q g r}} = \\left\\{ \\begin{array}{l l} \\boldsymbol {G} _ {\\text {q u a}} & \\text {i f} \\boldsymbol {G} _ {\\text {q u a}} \\cdot \\boldsymbol {G} _ {\\text {s e m}} \\leq 0, \\\\ \\boldsymbol {G} _ {\\text {q u a}} - \\lambda \\frac {\\boldsymbol {G} _ {\\text {q u a}} \\cdot \\boldsymbol {G} _ {\\text {s e m}}}{\\| \\boldsymbol {G} _ {\\text {s e m}} \\| ^ {2}} \\boldsymbol {G} _ {\\text {s e m}} & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.556, + 0.905, + 0.617 + ], + "angle": 0, + "content": "Here, \\(\\lambda\\) is introduced to generalize the formulation, providing flexibility in controlling the influence of general knowledge. Specifically, \\(\\lambda = 1\\) projects \\(G_{\\mathrm{qua}}\\) onto the orthogonal direction of \\(G_{\\mathrm{sem}}\\), while \\(\\lambda = 0\\) reduces QGR to CoOp [56]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.626, + 0.645, + 0.642 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.65, + 0.813, + 0.665 + ], + "angle": 0, + "content": "4.1. Datasets and Evaluation Protocols" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.672, + 0.907, + 0.899 + ], + "angle": 0, + "content": "We conduct experiments on multiple BIQA datasets include LIVEC [7], KonIQ [10], LIVEFB [49], and SPAQ [3], which feature authentic distortions, and PIPAL [12], LIVE [38], and CSIQ [17], which feature synthetic distortions. LIVEC contains 1,162 mobile device images, SPAQ includes 11,125 photos from 66 smartphones, KonIQ has 10,073 images from open sources, and LIVEFB is the largest real-world dataset with 39,810 images. For synthetic distortions, LIVE and CSIQ contain 779 and 866 images with 5 and 6 types of distortions. PIPAL, a challenging dataset, includes 23,200 images with 40 types of distortions, including GAN-generated artifacts. We use Spearman's Rank Correlation Coefficient (SRCC) and the Pearson Linear Correlation Coefficient (PLCC) as metrics to quantify the monotonousness and accuracy of predictions." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10446" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.907, + 0.285 + ], + "angle": 0, + "content": "
MethodLIVECKonIQCSIQLIVEPIPAL
Labels5010020050100200501002005010020050100200
HyperIQA [41]0.6480.7250.7900.6150.7100.7760.7900.8240.9090.8920.9120.9290.1020.3020.379
MetaIQA [58]0.6040.6260.6690.6180.6200.6600.7840.8490.8940.8400.8800.9190.3320.3480.371
DEIQT [33]0.6670.7180.8120.6380.6820.7540.8210.8910.9410.9200.9420.9550.3960.4100.436
MANIQA [48]0.6420.7690.7970.6520.7550.8100.7940.8470.8740.9090.9280.9570.1360.3610.470
CONTRIQUE [30]0.6950.7290.7610.7330.7940.8210.8400.9260.9400.8910.9220.9430.3790.4370.488
Re-IQA [37]0.5910.6210.7010.6850.7230.7540.8930.9070.9230.8840.8940.9290.2800.3500.431
CLIP w/ Linear Probe [36]0.6640.7210.7330.7360.7700.7820.8410.8920.9410.8960.9230.9410.2540.3030.368
CLIP w/ CoOp [44]†0.6950.7380.7460.6920.7430.762---------
LIQE [52]0.6910.7690.8100.7590.8010.8320.8380.8910.9240.9040.9340.948---
GRepQ [40]0.7600.7910.8220.8120.8360.8550.8780.9140.9410.9260.9370.9530.390†0.450†0.498†
GRMP-IQA (Ours)0.8360.8570.8750.8530.8720.8830.8930.9170.9410.9320.9430.9680.4740.5120.546
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.296, + 0.907, + 0.325 + ], + "angle": 0, + "content": "Table 1. SRCC performance comparison of our method with other IQA methods trained on limited labels. Bold indicates the best results, underlined marks the second-best, and the fifth-to-last through second-to-last lines show the CLIP-based IQA. † is our reproduction." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.33, + 0.907, + 0.6 + ], + "angle": 0, + "content": "
MethodLIVECSIQLIVECKonIQLIVEFBSPAQ
PLCCSRCCPLCCSRCCPLCCSRCCPLCCSRCCPLCCSRCCPLCCSRCC
Training Ratio20%20%20%20%20%20%20%20%10%10%20%20%
DEIQT [33]0.9680.9650.8850.8620.8220.7920.9080.8880.6240.5380.9120.908
LoDa [46]----0.8540.8150.9230.907----
GRMP-IQA (Ours)0.9720.9700.9580.9510.8970.8750.9310.9150.6860.6040.9250.920
Training Ratio80%80%80%80%80%80%80%80%80%80%80%80%
MetaIQA [58]0.9590.9600.9080.8990.8020.8350.8560.8870.5070.540--
CONTRIQUE [30]0.9610.9600.9550.9420.8570.8450.9060.8940.6410.5800.9190.914
DEIQT [33]0.9820.9800.9630.9460.8940.8750.9340.9210.6630.5710.9230.919
Re-IQA [37]0.9710.9700.9600.9470.8540.8400.9230.9140.7330.6450.9250.918
LIQE [52]0.9510.9700.9390.9360.9100.9040.9080.919----
CLIP-IQA+ [44]----0.8320.8050.9090.8950.5930.5750.8660.864
CDINet [54]0.9750.9770.9600.9520.8800.8650.9280.916--0.9220.919
QFM-IQM [22]0.9830.9810.9650.9540.9130.8910.9360.9220.6670.5670.9240.920
LoDa [46]0.9790.975--0.8990.8760.9440.9320.6790.5780.9280.925
GRMP-IQA (Ours)0.9830.9810.9740.9680.9160.8970.9450.9340.7040.6160.9320.927
" + }, + { + "type": "table_caption", + "bbox": [ + 0.135, + 0.611, + 0.86, + 0.626 + ], + "angle": 0, + "content": "Table 2. Performance comparison measured by medians of SRCC and PLCC, and bold entries indicate the top two results." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.633, + 0.398, + 0.649 + ], + "angle": 0, + "content": "4.2. Implementation Details and Setups" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.656, + 0.483, + 0.807 + ], + "angle": 0, + "content": "We build our model on CLIP-B/16 [36]. During pretraining, we optimize the visual-text prompt, while in finetuning, only the last four blocks of the image and text encoders are trained. For competing models, we use public results or re-train them under our setup. Each dataset is split into \\(80\\% / 20\\%\\) for training and testing, with splits based on reference images to ensure content independence in datasets with synthetic distortions. To ensure robustness, we report the median performance across ten random splits. All experiments are run on four NVIDIA RTX3090 GPUs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.809, + 0.483, + 0.869 + ], + "angle": 0, + "content": "Meta-Prompt Pre-training. We pre-train on TID2013 [32] and KADID-10K [25] dataset, which contain extensive distortion information. We set the learning rates \\(\\alpha\\) and \\(\\beta\\) to 1e-4 and 1e-2, and train for 50 epochs using Adam [16]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Fully Supervised Learning Setting. We randomly crop each input image into 10 patches of \\(224 \\times 224\\) resolution" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.634, + 0.907, + 0.773 + ], + "angle": 0, + "content": "and train the model for 9 epochs using AdamW [26]. The learning rate is \\(5 \\times 10^{-6}\\), with a scheduler over 9 decay epochs. The batch size is 16 for LIVEC and 128 for KonIQ. Few-Shot Learning Setting. In the few-shot setting, we follow the approach in a previous study [40] to train our GRMP-IQA model using randomly selected subsets of 50, 100, and 200 samples from the training set. The regularization weight \\(\\lambda\\) is fixed at 5, and the training hyperparameters are set according to the schedule proposed by CoOp [56]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.785, + 0.838, + 0.802 + ], + "angle": 0, + "content": "4.3. Performance Comparison with SOTA" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.809, + 0.906, + 0.87 + ], + "angle": 0, + "content": "Our method effectively acquires extensive image quality assessment knowledge, enabling it to provide powerful priors for various IQA scenarios. Tab. 1 and Tab. 2 summarize the comparative results for different experimental settings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Few-Shot Setting. Benefiting from the quality priors acquired during pre-training, our model achieves superior" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10447" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.089, + 0.48, + 0.24 + ], + "angle": 0, + "content": "
TrainingLIVEFBLIVECKonIQLIVECSIQ
TestingKonIQLIVECKonIQLIVECCSIQLIVE
DBCNN0.7160.7240.7540.7550.7580.877
HyperIAQ0.7580.7350.7720.7850.7440.926
TReS0.7130.7400.7330.7860.761-
DEIQT0.7330.7810.7440.7940.7810.932
CLIP-IQA+0.6310.6200.6970.803--
LoDa0.7630.8050.7450.811--
GRMP-IQA0.7650.7900.7820.8310.8090.935
" + }, + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.251, + 0.482, + 0.279 + ], + "angle": 0, + "content": "Table 3. SRCC on the cross datasets validation. The best performances are highlighted in boldface." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.289, + 0.48, + 0.394 + ], + "angle": 0, + "content": "
ComponentLIVECKonIQ
Pre-training Meta-learningPLCCSRCCPLCCSRCC
0.8250.7960.7880.764
✓0.8230.7880.7920.761
✓0.8580.8280.8440.811
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.405, + 0.482, + 0.434 + ], + "angle": 0, + "content": "Table 4. Ablation study on the effectiveness of meta-learning when trained with only 50 samples." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.443, + 0.483, + 0.761 + ], + "angle": 0, + "content": "performance even with limited training data. As shown in Tab. 1, our GRMP-IQA significantly outperforms the second-best model, GRepQ [40], which is specifically designed for few-shot learning. Additionally, our approach exhibits clear advantages over MetaIQA, a meta-learning-based method, underscoring the effectiveness of metaprompts for rapidly adapting CLIP to diverse IQA tasks. These results validate GRMP-IQA's strong capability in learning quality-aware feature with limited labeled samples. Fully Supervised Setting. Tab. 2 compares GRMP-IQA with other BIQA methods under full supervised setting, including self-training approaches like CONTRIQUE [30] and Re-IQA [37], as well as CLIP-based methods such as LIQE [52] and CLIP-IQA+ [44]. GRMP-IQA outperforms almost all competitors on six datasets. Achieving leading performance across these datasets is particularly challenging due to the diverse range of image content and distortion types. Notably, the proposed method achieves competitive results with the state-of-the-art (SOTA) method on several datasets while utilizing only \\(20\\%\\) of the training data compared to the fully supervised setting." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.772, + 0.411, + 0.789 + ], + "angle": 0, + "content": "4.4. Generalization Capability Validation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.482, + 0.901 + ], + "angle": 0, + "content": "To evaluate the generalization capacity of GRMP-IQA, we performed cross-dataset validation experiments, where the model was trained on one dataset and tested on others without parameter adjustments. Tab. 3 reports the SRCC results across five datasets. GRMP-IQA consistently outperforms state-of-the-art models in most cross Authentic scenarios, achieving notable improvements on LIVEC and" + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.09, + 0.905, + 0.227 + ], + "angle": 0, + "content": "
ComponentsLIVECKonIQPIPAL
MetaTextVisualPLCCSRCCPLCCSRCCPLCCSRCC
✓0.5790.5980.5920.5730.2160.203
0.6390.5890.5560.5540.3670.371
✓✓0.6990.6890.6790.6090.3230.312
✓✓0.7760.7420.7360.7010.3570.369
✓✓0.7590.7090.6220.5920.3620.396
✓✓✓0.8080.7700.7440.7130.4100.434
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.237, + 0.905, + 0.266 + ], + "angle": 0, + "content": "Table 5. Ablation experiments with Meta-Prompt Pre-training component. The best performances are highlighted in boldface." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.275, + 0.905, + 0.389 + ], + "angle": 0, + "content": "
MethodLabels50100
DatasetPLCCSRCCPLCCSRCC
w/o QGRLIVEC0.8580.8280.8750.848
w/ QGRLIVEC0.8640.8360.8830.857
w/o QGRKonIQ0.8440.8110.8720.840
w/ QGRKonIQ0.8800.8530.8960.872
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.399, + 0.905, + 0.427 + ], + "angle": 0, + "content": "Table 6. Ablation experiments with QGR in few-shot setting. The best performances are highlighted in boldface." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.443, + 0.905, + 0.473 + ], + "angle": 0, + "content": "KonIQ dataset. Additionally, it demonstrates strong competitiveness on synthetic datasets like LIVE and CSIQ." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.485, + 0.667, + 0.501 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.508, + 0.907, + 0.81 + ], + "angle": 0, + "content": "Effect of Meta-Prompt Pre-training Module. This module consists of three key components: meta-learning, text meta-prompts, and visual meta-prompts. We conduct ablation studies in Tab. 5 to evaluate their zero-shot capabilities across various datasets. The baseline uses a CLIP model pre-trained on classification tasks. Row 2 evaluates meta-learning without prompts, where CLIP's visual and text encoders are fine-tuned to learn distortion knowledge across different distortions, improving IQA performance. However, tuning CLIP's weights impairs its original generalization, resulting in lower performance on KonIQ. Rows 3 and 4 assess fine-tuning only text or visual prompts during meta-learning. These strategies notably boost zero-shot performance—particularly on real-world datasets like KonIQ and LIVEC—by preserving CLIP's generalization while adapting to IQA tasks. Row 5 explores prompt tuning without meta-learning, underscoring meta-learning's role in curbing overfitting and maintaining generalization. The strongest results emerge when all components are combined, highlighting their complementary strengths." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Effect of Meta-learning. To further investigate whether the effectiveness of our method derives from meta-learning, we conducted an ablation study. Specifically, as shown in rows 1 and 2 of Tab. 4, when using Empirical Risk Minimization without meta-learning pre-training on two synthetic dataset, the fine-tuning performance on LIVEC and KonIQ dataset" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10448" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.258, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.179, + 0.211, + 0.193, + 0.22 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.259, + 0.089, + 0.478, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.37, + 0.211, + 0.384, + 0.22 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.225, + 0.483, + 0.281 + ], + "angle": 0, + "content": "Figure 4. (a) SRCC value during training with 50 samples. (b) Angles between \\( G_{qua} \\) and \\( G_{sem} \\) during training on KonIQ. Under the constraints of QGR, our method captures more correlation between quality knowledge and general semantic knowledge." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.289, + 0.48, + 0.398 + ], + "angle": 0, + "content": "
λ50100200
PLCCSRCCPLCCSRCCPLCCSRCC
10.8500.8120.8740.8390.8930.866
30.8640.8390.8900.8650.9030.878
50.8800.8530.8960.8720.9080.883
70.8760.8510.8950.8700.9050.881
" + }, + { + "type": "table_caption", + "bbox": [ + 0.118, + 0.408, + 0.455, + 0.423 + ], + "angle": 0, + "content": "Table 7. The ablation study about soft weight \\( \\lambda \\) in Eq. 8." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.432, + 0.483, + 0.522 + ], + "angle": 0, + "content": "was even worse than the baseline without pre-training. This indicates that additional pre-training data does not necessarily enhance performance and may even lead to overfitting [19]. In contrast, our meta-learning strategy effectively extracts generalizable quality priors from external data, significantly enhancing performance across datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.523, + 0.483, + 0.779 + ], + "angle": 0, + "content": "Effect of Gradient Regularization. We conducted ablation studies on the QGR module. As shown in Tab. 6 and Fig. 4(a), without QGR, models are prone to overfitting under limited training data, with rapid performance saturation and a slight decline, which ultimately hinders generalization. By modulating training gradients, QGR effectively mitigates overfitting and enhances adaptability to the test dataset. To illustrate QGR's impact, we analyzed the angular difference between gradients \\( G_{\\mathrm{qua}} \\) and \\( G_{\\mathrm{sem}} \\) during training, as shown in Fig. 4(b). Without QGR, the angle between \\( G_{\\mathrm{qua}} \\) and \\( G_{\\mathrm{sem}} \\) tends toward 90 degrees, reflecting the orthogonality typical of high-dimensional random vectors [2]. In contrast, QGR introduces greater angular variation, suggesting that it encourages the model to explore more correlations between quality and semantic directions. This improves the model's ability to capture quality-related information and mitigates overfitting on limited data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.885 + ], + "angle": 0, + "content": "Effect of soft weight \\(\\lambda\\). On the KonIQ dataset, we conduct ablation experiments with various soft weights \\(\\lambda\\) in Eq. 8 to examine their impact, as detailed in Tab. 7. Results show that a small \\(\\lambda\\) diminishes the effectiveness of our QGR, whereas a large \\(\\lambda\\) causes substantial gradient changes and decreases performance. Given our observation of the trade-off, we adopt \\(\\lambda = 5\\) in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Computational Analysis. As shown in Tab. 8, we evaluate" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.613, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.543, + 0.148, + 0.579, + 0.156 + ], + "angle": 0, + "content": "Best ours" + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.09, + 0.709, + 0.146 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.648, + 0.147, + 0.686, + 0.155 + ], + "angle": 0, + "content": "Best ours" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.089, + 0.806, + 0.146 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.737, + 0.147, + 0.785, + 0.154 + ], + "angle": 0, + "content": "Best DEIQT" + }, + { + "type": "image", + "bbox": [ + 0.808, + 0.09, + 0.903, + 0.146 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.834, + 0.148, + 0.881, + 0.155 + ], + "angle": 0, + "content": "Best DEIQT" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.156, + 0.616, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.179, + 0.584, + 0.188 + ], + "angle": 0, + "content": "Worst ours" + }, + { + "type": "image", + "bbox": [ + 0.644, + 0.156, + 0.664, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.645, + 0.18, + 0.688, + 0.188 + ], + "angle": 0, + "content": "Worst ours" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.156, + 0.806, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.736, + 0.18, + 0.79, + 0.188 + ], + "angle": 0, + "content": "Worst DEIQT" + }, + { + "type": "image", + "bbox": [ + 0.833, + 0.156, + 0.899, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.188, + 0.612, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.553, + 0.244, + 0.564, + 0.251 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.614, + 0.188, + 0.71, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.66, + 0.244, + 0.672, + 0.252 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.188, + 0.806, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.753, + 0.244, + 0.765, + 0.252 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.807, + 0.188, + 0.903, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.851, + 0.244, + 0.864, + 0.251 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.26, + 0.905, + 0.302 + ], + "angle": 0, + "content": "Figure 5. gMAD results between DEIQT [33] and our GRMP-IQA. (a) Fixed DEIQT at low quality. (b) Fixed DEIQT at high quality. (c) Fixed ours at low quality. (d) Fixed ours at high quality." + }, + { + "type": "table", + "bbox": [ + 0.519, + 0.308, + 0.905, + 0.383 + ], + "angle": 0, + "content": "
MethodParamEfficiencyPerformance
Throughput ↑Latency ↓80% Train20% Train
Q-Align8.2B9.9 (img/sec)101.0 (ms)0.941 / 0.9400.901 / 0.903
Ours151M43.0 (img/sec)21.3 (ms)0.945 / 0.9340.931 / 0.915
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.393, + 0.905, + 0.421 + ], + "angle": 0, + "content": "Table 8. Comparison with Q-Align across varying training data ratio to evaluate efficiency and PLCC/SRCC on the KonIQ dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.905, + 0.492 + ], + "angle": 0, + "content": "the inference latency and throughput of our method on an RTX3090 GPU, surpassing Q-Align [45] in both efficiency and accuracy. Notably, our model has only 41M learnable and 151M total parameters, far fewer than Q-Align's 8.2B." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.5, + 0.706, + 0.516 + ], + "angle": 0, + "content": "4.6. Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.907, + 0.718 + ], + "angle": 0, + "content": "The gMAD competition [28] is a standard method for assessing IQA model robustness. It selects image pairs where the attacker model predicts a large quality difference, while the defender perceives them as similar. Observers then evaluate these pairs to gauge model generalization. As shown in Fig. 5, when our model acts as the defender, the attacker's selected pairs exhibit minimal perceptual quality changes. Conversely, as the attacker, our model consistently identifies pairs with significant quality differences. This highlights its strong defensive and offensive capabilities. Notably, in the second column, DEIQT misclassifies semantically similar images as having comparable quality, whereas our model accurately distinguishes their differences." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.727, + 0.634, + 0.743 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose the GRMP-IQA framework, which generalizes well with limited data. It includes a meta-learning pre-training module that enables the CLIP model to rapidly adapt to IQA tasks and an adaptive gradient regulation module that refines gradient trajectories during fin-tuning, focusing updates on quality-aware knowledge while minimizing the negative impact of over-reliance on semantic noise. Comprehensive experiments on various BIQA datasets validate the superior generalization of our framework, especially in data-scarce scenarios." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10449" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.228 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported by the National Science Fund for Distinguished Young Scholars (No.62025603), the National Natural Science Foundation of China (No. U21B2037, No. U22B2051, No. U23A20383, No. U21A20472, No. 62176222, No. 62176223, No. 62176226, No. 62072386, No. 62072387, No. 62072389, No. 62002305 and No. 62272401), and the Natural Science Foundation of Fujian Province of China (No. 2021J06003, No.2022J06001)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.241, + 0.188, + 0.256 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.265, + 0.484, + 0.336 + ], + "angle": 0, + "content": "[1] Nithin C Babu, Vignesh Kannan, and Rajiv Soundararajan. No reference opinion unaware quality assessment of authentically distorted images. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2459-2468, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.337, + 0.484, + 0.379 + ], + "angle": 0, + "content": "[2] Tony Cai, Jianqing Fan, and Tiefeng Jiang. Distributions of angles in random packing on spheres. The Journal of Machine Learning Research, 14(1):1837-1864, 2013. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.379, + 0.483, + 0.448 + ], + "angle": 0, + "content": "[3] Yuming Fang, Hanwei Zhu, Yan Zeng, Kede Ma, and Zhou Wang. Perceptual quality assessment of smartphone photography. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3677-3686, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.45, + 0.483, + 0.506 + ], + "angle": 0, + "content": "[4] Yuqian Fu, Yanwei Fu, and Yu-Gang Jiang. Meta-fdmixup: Cross-domain few-shot learning guided by labeled target data. In Proceedings of the 29th ACM international conference on multimedia, pages 5326-5334, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.507, + 0.483, + 0.562 + ], + "angle": 0, + "content": "[5] Hongcheng Gao, Tianyu Pang, Chao Du, Taihang Hu, Zhijie Deng, and Min Lin. Meta-unlearning on diffusion models: Preventing relearning unlearned concepts. arXiv preprint arXiv:2410.12777, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.564, + 0.483, + 0.632 + ], + "angle": 0, + "content": "[6] Hongcheng Gao, Jiashu Qu, Jingyi Tang, Baolong Bi, Yue Liu, Hongyu Chen, Li Liang, Li Su, and Qingming Huang. Exploring hallucination of large multimodal models in video understanding: Benchmark, analysis and mitigation. arXiv preprint arXiv:2503.19622, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.634, + 0.483, + 0.688 + ], + "angle": 0, + "content": "[7] Deepti Ghadiyaram and Alan C Bovik. Massive online crowdsourced study of subjective and objective picture quality. IEEE Transactions on Image Processing, 25(1):372-387, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.69, + 0.483, + 0.759 + ], + "angle": 0, + "content": "[8] S Alireza Golestaneh, Saba Dadsetan, and Kris M Kitani. No-reference image quality assessment via transformers, relative ranking, and self-consistency. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1220-1230, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.761, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[10] Vlad Hosu, Hanhe Lin, Tamas Sziranyi, and Dietmar Saupe. Koniq-10k: An ecologically valid database for deep learning of blind image quality assessment. IEEE Transactions on Image Processing, 29:4041–4056, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[11] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Vi" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.265, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.907, + 0.121 + ], + "angle": 0, + "content": "sual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.907, + 0.206 + ], + "angle": 0, + "content": "[12] Gu Jinjin, Cai Haoming, Chen Haoyu, Ye Xiaoxing, Jimmy S Ren, and Dong Chao. Pipal: a large-scale image quality assessment dataset for perceptual image restoration. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI 16, pages 633-651. Springer, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.906, + 0.263 + ], + "angle": 0, + "content": "[13] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5148-5157, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.519, + 0.264, + 0.906, + 0.332 + ], + "angle": 0, + "content": "[14] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19113-19122, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.334, + 0.906, + 0.419 + ], + "angle": 0, + "content": "[15] Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Self-regulating prompts: Foundational model adaptation without forgetting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15190-15200, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.42, + 0.906, + 0.46 + ], + "angle": 0, + "content": "[16] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.461, + 0.906, + 0.517 + ], + "angle": 0, + "content": "[17] Eric Cooper Larson and Damon Michael Chandler. Most apparent distortion: full-reference image quality assessment and the role of strategy. Journal of electronic imaging, 19 (1):011006, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.906, + 0.56 + ], + "angle": 0, + "content": "[18] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.562, + 0.906, + 0.632 + ], + "angle": 0, + "content": "[19] Aobo Li, Jinjian Wu, Yongxu Liu, and Leida Li. Bridging the synthetic-to Authentic gap: Distortion-guided unsupervised domain adaptation for blind image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28422-28431, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.633, + 0.906, + 0.702 + ], + "angle": 0, + "content": "[20] Chunyi Li, Zicheng Zhang, Haoning Wu, Wei Sun, Xiongkuo Min, Xiaohong Liu, Guangtao Zhai, and Weisi Lin. Agiqa-3k: An open database for ai-generated image quality assessment. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.704, + 0.906, + 0.787 + ], + "angle": 0, + "content": "[21] Juncheng Li, Minghe Gao, Longhui Wei, Siliang Tang, Wenqiao Zhang, Mengze Li, Wei Ji, Qi Tian, Tat-Seng Chua, and Yueting Zhuang. Gradient-regulated meta-prompt learning for generalizable vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2551-2562, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.789, + 0.906, + 0.872 + ], + "angle": 0, + "content": "[22] Xudong Li, Timin Gao, Runze Hu, Yan Zhang, Shengchuan Zhang, Xiawu Zheng, Jingyuan Zheng, Yunhang Shen, Ke Li, Yutao Liu, et al. Adaptive feature selection for no-reference image quality assessment by mitigating semantic noise sensitivity. In International Conference on Machine Learning, pages 27808-27821. PMLR, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[23] Xudong Li, Runze Hu, Jingyuan Zheng, Yan Zhang, Shengchuan Zhang, Xiawu Zheng, Ke Li, Yunhang Shen," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10450" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.483, + 0.148 + ], + "angle": 0, + "content": "Yutao Liu, Pingyang Dai, et al. Integrating global context contrast and local sensitivity for blind image quality assessment. In International Conference on Machine Learning, pages 27920-27941. PMLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.483, + 0.219 + ], + "angle": 0, + "content": "[24] Xudong Li, Wenjie Nie, Yan Zhang, Runze Hu, Ke Li, Xiawu Zheng, and Liujuan Cao. Distilling spatially-heterogeneous distortion perception for blind image quality assessment. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 2344-2354, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.483, + 0.276 + ], + "angle": 0, + "content": "[25] Hanhe Lin, Vlad Hosu, and Dietmar Saupe. Kadid-10k: A large-scale artificially distorted iqa database. In 2019 Eleventh International Conference on Quality of Multimedia Experience (QoMEX), pages 1-3. IEEE, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.278, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[26] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.483, + 0.376 + ], + "angle": 0, + "content": "[27] Ziwei Luo, Fredrik K Gustafsson, Zheng Zhao, Jens Sjolund, and Thomas B Schon. Controlling vision-language models for multi-task image restoration. In International Conference on Learning Representations, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.483, + 0.459 + ], + "angle": 0, + "content": "[28] Kede Ma, Qingbo Wu, Zhou Wang, Zhengfang Duanmu, Hongwei Yong, Hongliang Li, and Lei Zhang. Group mad competition-a new methodology to compare objective image quality models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1664-1673, 2016. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.483, + 0.518 + ], + "angle": 0, + "content": "[29] Kede Ma, Xuelin Liu, Yuming Fang, and Eero P Simoncelli. Blind image quality assessment by learning from multiple annotators. In 2019 IEEE international conference on image processing (ICIP), pages 2344-2348. IEEE, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.52, + 0.483, + 0.574 + ], + "angle": 0, + "content": "[30] Pavan C Madhusudana, Neil Birkbeck, Yilin Wang, Balu Adsumilli, and Alan C Bovik. Image quality assessment using contrastive learning. IEEE Transactions on Image Processing, 31:4149-4161, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.483, + 0.631 + ], + "angle": 0, + "content": "[31] Jinyoung Park, Juyeon Ko, and Hyunwoo J Kim. Prompt learning via meta-regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26940–26950, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.483, + 0.715 + ], + "angle": 0, + "content": "[32] Nikolay Ponomarenko, Lina Jin, Oleg Ieremeiev, Vladimir Lukin, Karen Egiazarian, Jaakko Astola, Benoit Vozel, Kacem Chehdi, Marco Carli, Federica Battisti, et al. Image database tid2013: Peculiarities, results and perspectives. Signal processing: Image communication, 30:57-77, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.483, + 0.786 + ], + "angle": 0, + "content": "[33] Guanyi Qin, Runze Hu, Yutao Liu, Xiawu Zheng, Haotian Liu, Xiu Li, and Yan Zhang. Data-efficient image quality assessment with attention-panel decoder. In Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence, 2023. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.483, + 0.83 + ], + "angle": 0, + "content": "[34] B Qu, H Li, and W Gao. Bringing textual prompt to a-generated image quality assessment, in 2024 IEEE international conference on multimedia and expo (icme), 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[35] Bowen Qu, Xiaoyu Liang, Shangkun Sun, and Wei Gao. Exploring aigc video quality: A focus on visual harmony videotext consistency and domain distribution gap. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6652-6660, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.175 + ], + "angle": 0, + "content": "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.177, + 0.905, + 0.244 + ], + "angle": 0, + "content": "[37] Avinab Saha, Sandeep Mishra, and Alan C Bovik. Reiqa: Unsupervised learning for image quality assessment in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5846-5855, 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.246, + 0.905, + 0.3 + ], + "angle": 0, + "content": "[38] Hamid R Sheikh, Muhammad F Sabir, and Alan C Bovik. A statistical evaluation of recent full reference image quality assessment algorithms. IEEE Transactions on image processing, 15(11):3440-3451, 2006. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[39] Tianshu Song, Leida Li, Deqiang Cheng, Pengfei Chen, and Jinjian Wu. Active learning-based sample selection for label-efficient blind image quality assessment. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[40] Suhas Srinath, Shankhanil Mitra, Shika Rao, and Rajiv Soundararajan. Learning generalizable perceptual representations for data-efficient no-reference image quality assessment. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 22-31, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.51 + ], + "angle": 0, + "content": "[41] Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3667-3676, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[42] Zhengzhong Tu, Yilin Wang, Neil Birkbeck, Balu Adsumilli, and Alan C Bovik. Ugc-vqa: Benchmarking blind video quality assessment for user generated content. IEEE Transactions on Image Processing, 30:4449-4464, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[43] Joaquin Vanschoren. Meta-learning: A survey. arXiv preprint arXiv:1810.03548, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[44] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.905, + 0.733 + ], + "angle": 0, + "content": "[45] Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Yixuan Gao, Annan Wang, Erli Zhang, Wenxiu Sun, et al. Q-align: Teaching lmmss for visual scoring via discrete text-defined levels. In International Conference on Machine Learning, pages 54015-54029. PMLR, 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[46] Kangmin Xu, Liang Liao, Jing Xiao, Chaofeng Chen, Haoning Wu, Qiong Yan, and Weisi Lin. Boosting image quality assessment through efficient transformer adaptation with local feature enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2662-2672, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[47] Zhengzhuo Xu, Bowen Qu, Yiyan Qi, Sinan Du, Chengjin Xu, Chun Yuan, and Jian Guo. Chartmoe: Mixture of diversely aligned expert connector for chart understanding. arXiv preprint arXiv:2409.03277, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[48] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10451" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.147 + ], + "angle": 0, + "content": "Maniqa: Multi-dimension attention network for no-reference image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1191-1200, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.484, + 0.232 + ], + "angle": 0, + "content": "[49] Zhenqiang Ying, Haoran Niu, Praful Gupta, Dhruv Mahajan, Deepti Ghadiyaram, and Alan Bovik. From patches to pictures (paq-2-piq): Mapping the perceptual space of picture quality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3575–3585, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.484, + 0.289 + ], + "angle": 0, + "content": "[50] Ji Zhang, Shihan Wu, Lianli Gao, Heng Tao Shen, and Jingkuan Song. Dept: Decoupled prompt tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12924-12933, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.29, + 0.484, + 0.357 + ], + "angle": 0, + "content": "[51] Weixia Zhang, Kede Ma, Jia Yan, Dexiang Deng, and Zhou Wang. Blind image quality assessment using a deep bilinear convolutional neural network. IEEE Transactions on Circuits and Systems for Video Technology, 30(1):36-47, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.36, + 0.484, + 0.442 + ], + "angle": 0, + "content": "[52] Weixia Zhang, Guangtao Zhai, Ying Wei, Xiaokang Yang, and Kede Ma. Blind image quality assessment via vision-language correspondence: A multitask learning perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14071-14081, 2023. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.445, + 0.484, + 0.514 + ], + "angle": 0, + "content": "[53] Kai Zhao, Kun Yuan, Ming Sun, Mading Li, and Xing Wen. Quality-aware pre-trained models for blind image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22302-22313, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.516, + 0.484, + 0.57 + ], + "angle": 0, + "content": "[54] Limin Zheng, Yu Luo, Zihan Zhou, Jie Ling, and Guanghui Yue. Cdinet: Content distortion interaction network for blind image quality assessment. IEEE Transactions on Multimedia, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.572, + 0.484, + 0.64 + ], + "angle": 0, + "content": "[55] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16816-16825, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.642, + 0.484, + 0.697 + ], + "angle": 0, + "content": "[56] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.699, + 0.484, + 0.754 + ], + "angle": 0, + "content": "[57] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15659-15669, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.755, + 0.484, + 0.825 + ], + "angle": 0, + "content": "[58] Hancheng Zhu, Leida Li, Jinjian Wu, Weisheng Dong, and Guangming Shi. Metaiaq: Deep meta-learning for no-reference image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14143-14152, 2020. 3, 4, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.825 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10452" + } + ] +] \ No newline at end of file diff --git a/2025/Few-Shot Image Quality Assessment via Adaptation of Vision-Language Models/89cdaf16-5828-4989-b825-83012badf535_origin.pdf b/2025/Few-Shot Image Quality Assessment via Adaptation of Vision-Language Models/89cdaf16-5828-4989-b825-83012badf535_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d775e56d65fd7ccae7ec4e39d97645114d9fc379 --- /dev/null +++ b/2025/Few-Shot Image Quality Assessment via Adaptation of Vision-Language Models/89cdaf16-5828-4989-b825-83012badf535_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8fcb79544daeaa3f89bcfc7c3b3fc4118f6bb1a55250b29e7dbc4d4a46838fcc +size 5230371 diff --git a/2025/Few-Shot Image Quality Assessment via Adaptation of Vision-Language Models/full.md b/2025/Few-Shot Image Quality Assessment via Adaptation of Vision-Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..44074e139fa3adba584fe30726c58a6b13306831 --- /dev/null +++ b/2025/Few-Shot Image Quality Assessment via Adaptation of Vision-Language Models/full.md @@ -0,0 +1,351 @@ +# Few-Shot Image Quality Assessment via Adaptation of Vision-Language Models + +Xudong Li $^{1*}$ Zihao Huang $^{2*}$ Yan Zhang $^{1\dagger}$ Yunhang Shen $^{3}$ Ke Li $^{3}$ Xiawu Zheng $^{1}$ Liujuan Cao $^{1}$ Rongrong Ji $^{1}$ + +1 Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University, 361005, P.R. China +$^{2}$ Beijing Institute of Technology $^{3}$ Tencent Youtu Lab, Shanghai, China {1xd761050753, huangzihhhh, bzhy986}@gmail.com, {zhengxiawu, rrji}@xmu.edu.cn + +# Abstract + +Image Quality Assessment (IQA) remains an unresolved challenge in computer vision due to complex distortions, diverse image content, and limited data availability. Existing Blind IQA (BIQA) methods largely rely on extensive human annotations, which are labor-intensive and costly due to the demanding nature of creating IQA datasets. To reduce this dependency, we propose the Gradient-Regulated MetaPrompt IQA Framework (GRMP-IQA), designed to efficiently adapt the visual-language pre-trained model, CLIP, to IQA tasks, achieving high accuracy even with limited data. GRMP-IQA consists of two core modules: (i) MetaPrompt Pre-training Module and (ii) Quality-Aware Gradient Regularization. The Meta Prompt Pre-training Module leverages a meta-learning paradigm to pre-train soft prompts with shared meta-knowledge across different distortions, enabling rapid adaptation to various IQA tasks. On the other hand, the Quality-Aware Gradient Regularization is designed to adjust the update gradients during fine-tuning, focusing the model's attention on quality-relevant features and preventing overfitting to semantic information. Extensive experiments on standard BIQA datasets demonstrate the superior performance to the state-of-the-art BIQA methods under limited data setting. Notably, utilizing just $20\%$ of the training data, GRMP-IQA is competitive with most existing fully supervised BIQA approaches. Our code is available via https://github.com/LXDxmu/GRMP-IQA. + +# 1. Introduction + +With the rise of the mobile internet era, the focus on computer vision has transitioned from initial concerns with compression and image processing [38] to handling user-generated content like smartphone photos and videos [6, + +![](images/a0e96b5a28fe707ed86d5577d049d86dc08d327f699e87aa4584757a3ed9bf22.jpg) +(b) Fine-tuning : Quality-Aware Gradient Regularization +Figure 1. Intuitive diagram of GRMP-IQA. (a) demonstrates using meta-learning for efficient soft prompt initialization with quality prior that enhances zero-shot generalization ( $\uparrow$ $24\%$ ), enabling the CLIP to adapt to BIQA tasks effectively. (b) illustrates gradient regularization during fine-tuning, which clips quality gradients aligned with semantic bias to guide the model toward quality-relevant features and reduce overly reliance on semantic content. In the t-SNE visualization, black dots represent noise samples with high semantic confidence (confidence $>0.8$ ) but incorrect quality predictions. The noticeable decrease ( $\downarrow$ $73\%$ ) in black dots highlights reduced semantic noise interference in IQA tasks. + +35, 42], and lately to AI-generated content [5, 20]. This evolution has sharply increased demand for effective Blind Image Quality Assessment (BIQA) techniques, highlighting the importance of developing methodologies that can adeptly evaluate image quality without reference images. Data-driven BIQA models [34, 39, 48] based on deep neural networks have made significant progress in recent years. However, the quality scores for distorted images are often measured using the Mean Opinion Score (MOS), which is the average of multiple ratings (sometimes up to 120). As a result, acquiring a sufficient number of IQA training samples is quite labor-intensive and cost-expensive. + +To mitigate this challenge, recent BIQA approaches [33, 37, 40, 46, 53] have leveraged large-scale image datasets through transfer learning or self-supervised learning. While these approaches improve performance, they often rely on computationally intensive pretraining or require domain-specific pretext tasks, limiting scalability. On the other hand, the vision-language model CLIP [36] has shown remarkable generalization across various downstream tasks, offering a potential alternative. Specifically, CLIP-IQA [44] demonstrates promising zero-shot performance in IQA by leveraging simple handcrafted prompts (e.g., "Good/Bad photo"). However, a significant performance gap remains compared to fully supervised methods [40, 45], highlighting the need to leverage limited labeled data to further close this performance gap. To bridge this gap, we introduce the concept of Few-Shot IQA, which aims to enable models to quickly generalize to new distortion types and unseen scenarios with only a small amount of labeled data (e.g., 50-200 training images). A natural way to adapt CLIP to downstream IQA tasks under this setup is through prompt tuning [11, 56, 57]. Prompt tuning techniques learn soft textual prompts (continuous embeddings) from a small labeled dataset while keeping the pretrained model parameters frozen, offering a lightweight and efficient solution. + +However, as shown in Tab. 1, the direct application of the prompt tuning method CoOp [56] to IQA yields minimal improvement over simple linear probing on the CLIP image encoder, falling short of the expected benefits seen in high-level vision tasks. We attribute this limitation to a fundamental mismatch between CLIP's pretraining objectives and the specific requirements of IQA tasks. CLIP is pretrained to align text and images in a shared semantic space, with a focus on high-level semantics such as objects and scenes. In contrast, IQA demands fine-grained sensitivity to low-level distortion cues—an ability in which CLIP is relatively deficient [27, 52]. This discrepancy introduces two key challenges when applying prompt tuning to IQA tasks: + +(1) Sensitivity to Initialization: During prompt tuning, with CLIP's parameters frozen and limited sensitivity to distortions, the model depends heavily on the learned prompts for quality assessment. Different prompt initializations lead to distinct optimization directions [18], which may steer optimization toward features less relevant to quality, causing significant performance variability. As shown in Fig. 2, the average SRCC fluctuates considerably across different random initializations, requiring careful tuning for each IQA scenario and limiting the model's adaptability. +(2) Overfitting from Semantic Bias: Due to CLIP's inherent reliance on semantic information, fine-tuning with a small sample size often leads to overfitting, where the model learns spurious correlations (e.g., over-reliance on semantic features to infer quality [1, 22]) rather than true image quality indicators, which degrades IQA performance. + +![](images/d325e96ca932085b18e3dfbcafe7523e8fa82eb91ba75f3e9663db88f3b77ac4.jpg) +Figure 2. The fine-tuning results on the LIVEC dataset using CoOp [56] show that prompt tuning accuracy is highly sensitive to initial random initialization of the prompt, especially with limited training data, underscoring its critical role under data constraints. + +![](images/5e8f9120d7dfdcaaa83bd8c6fdc177716c72b5a54bc38e3f18196f9a16388e62.jpg) + +As illustrated in Fig. 1, when the CLIP encoder is finetuned on a limited dataset, the latent space contains many instances (black dots) that confidently and accurately predict image semantics (with confidence levels above 0.8) but fail to reliably assess image quality. + +To overcome these limitations, we propose Gradient-Regulated Meta-Prompt Learning for IQA (GRMP-IQA), a framework with two key components: (i) Meta Prompt Pre-training (MPP) mitigates sensitivity to prompt initialization by incorporating generalized distortion priors into the initialization of soft prompts, which enhances their adaptability to new IQA scenarios. By designing metatraining tasks with a well-annotated dataset containing diverse distortions (e.g., overexposure, blur) and optimizing prompts through bi-level gradient descent, we enable the model to learn shared meta-knowledge of quality representations across different distortions, thereby improving generalization across IQA tasks. (ii) Quality-Aware Gradient Regularization (QGR) addresses overfitting to semantic information by regulating gradient updates during fine-tuning, which balances the interplay between the quality knowledge gradient $G_{\mathrm{qua}}$ and the semantic knowledge gradient $G_{\mathrm{sem}}$ . By clipping $G_{\mathrm{qua}}$ along $G_{\mathrm{sem}}$ when their directions are overly aligned, this module suppresses semantic bias and guides the model to focus on distortion-related features, improving its ability to capture low-level quality details. Additionally, fine-tuning only CLIP's text branch can misalign quality perception between the image and text branches, limiting generalization [14]. To address this, we integrate text prompt tuning (CoOp [56]) and visual prompt tuning (VPT [11]) by jointly meta-learning initializations for both, ensuring complementary optimization to better adapt CLIP to new IQA scenarios (Tab. 5). + +Our contributions can be summarized as follows: + +- We propose a Meta Prompt Pre-training method that organizes meta tasks based on image distortion types and optimizes soft textual and visual prompts to gain shared quality meta-knowledge, enabling CLIP to rapidly adapt across various IQA scenarios. +- We develop a novel Quality-Aware Gradient Regularization method that clips gradients aligned with semantic + +directions to balance semantic and quality information during fine-tuning, ensuring the model prioritizes image quality while still integrating relevant semantic context. + +- Extensive empirical results confirm that our approach is both effective and efficient. Notably, with just 200 data samples, our method is competitive with SOTA models on the LIVEC dataset (using $20\%$ of the training data). + +# 2. Related Work + +# 2.1. Deep Learning Based BIQA methods + +Early CNN-based IQA methods followed standard pretraining and fine-tuning pipelines [9, 51], while meta-learning approaches like MetaIQA [58] improved adaptability from synthetic to real-world images. However, CNNs struggle with non-local features, a gap addressed by ViT-based models [8, 13, 23, 24]. Recent methods like LIQE [52] leverage CLIP's multi-task learning, enhancing IQA through supervised fine-tuning across datasets. While effective, these methods rely heavily on extensive annotations, making them costly and time-intensive. In contrast, limited-data BIQA remains underexplored. CLIP-IQA[44] shows promise in zero-shot settings but has performance limits, highlighting the need for better use of limited labeled data. Methods like DEIQT[33] and LoDa[46] demonstrate that fine-tuning pre-trained ViTs reduces annotation needs but still underperform in few-shot scenarios. Self-supervised approaches[37, 40, 53] ease data constraints through tailored pre-training tasks but carry high computational costs. Additionally, recent studies [1, 22] explore disentangling semantic and distortion-related content for IQA—yet none fully address the dual challenges of data efficiency and computational cost. + +To bridge these gaps, our approach combines CLIP's pre-trained semantic knowledge with distortion-specific fine-tuning. By using prompt tuning to cut pre-training costs and a parameter-free regularization strategy (leveraging high- and low-level gradient correlations), it optimizes IQA more effectively with fewer labeled samples. + +# 2.2. Prompt Tuning + +Given the limited labeled data during training, prompt tuning is an effective approach to adapt vision-language pre-trained models for few-shot learning tasks [4, 47, 50]. CoOp [56] optimizes prompt vectors in CLIP's language branch for task adaptation but struggles with generalization to novel tasks. To address this, CoCoOp [55] introduces a lightweight meta-network to generate input-conditioned tokens, improving adaptability. Beyond textual prompts, visual prompts have also been explored for task adaptation [11]. Methods like MaPLe [14] and PromptSRC [15] incorporate trainable prompts into both language and visual branches, achieving significant performance gains across + +base and novel tasks. Recent studies [21, 31] further construct meta-tasks based on semantic categories to initialize well-structured soft prompts, mitigating generalization decline in high-level tasks. Unlike these approaches that focus on high-level tasks, our proposed prompt tuning method is specifically designed for IQA, leveraging low-level distortion information to construct meta-tasks. This enables pretrained prompts to be distortion-sensitive and effectively capture shared quality knowledge. + +# 3. Methodology + +# 3.1. Overview + +In this paper, we propose the Gradient-Regulated MetaPrompt Image Quality Assessment (GRMP-IQA), which aims to adapt the CLIP for BIQA tasks with a few training samples. As depicted in Fig. 3, GRMP-IQA consists of two primary modules: (i) the Meta-Prompt Pre-training Module (MPP) and (ii) the Quality-Aware Gradient Regularization (QGR). The MPP module pre-trains visual-text prompts to acquire shared meta-knowledge on distortions, enabling quick adaptation to various IQA scenarios. The QGR module plays a key role in fine-tuning by adjusting gradient updates to prevent overfitting to semantic content. + +Pre-training stage (Sec. 3.3). We randomly sample a minibatch from distortion meta-tasks, partitioning it into a support set $\mathcal{D}^{\mathrm{support}}$ and a query set $\mathcal{D}^{\mathrm{query}}$ . A bi-level gradient descent method progresses from the Inner-Loop on $\mathcal{D}^{\mathrm{support}}$ to the Outer-Loop on $\mathcal{D}^{\mathrm{query}}$ to optimize the learnable visual-textual prompts $[\theta_T, \theta_V]$ . These prompts are then used as initial weights for following fine-tuning. + +Fine-tuning stage (Sec. 3.4). Given an input image $\mathbf{x}$ and a hard semantic prompt $\boldsymbol{w}_i =$ "a photo of a [class]", we predict the semantic class distributions $p^{\text{sem}}(\boldsymbol{w}_i|\boldsymbol{x})$ and $p^{\text{qua}}(\boldsymbol{w}_i|\boldsymbol{x})$ using the original semantic CLIP model and the fine-tuned CLIP model, respectively. The gradient of the KL divergence loss $\nabla \mathcal{L}_{kl}$ between these distributions is computed as the general semantic direction $G_{\text{sem}}$ . Simultaneously, using soft quality prompt $\theta_T$ and image $\mathbf{x}$ , we compute the gradient $\nabla \mathcal{L}_{ce}$ of the quality loss as the quality direction $G_{\text{qua}}$ . To refine the optimization direction of the IQA task, we adjust $G_{\text{qua}}$ into $G_{\text{qgr}}$ by clipping its components aligned with $G_{\text{sem}}$ , ensuring the model focuses on image quality while reducing the impact of semantic noise. + +# 3.2. Visual-Text Meta-Prompt + +Visual Meta-Prompt. We utilize Deep Prompt Tuning (DPT) [56] as our Visual Meta-Prompt, with learnable parameters $\theta_{\mathbf{V}}$ . The input embedding for the $l$ -th layer's self-attention module in the ViT-based image encoder is denoted as $\{\pmb{f}^l,\pmb{H}^l\}$ , where $\pmb{f}^l$ represents the classification (CLS) token, and $\pmb{H}^{l} = \{\pmb{h}_{1}^{l},\pmb{h}_{2}^{l},\dots,\pmb{h}_{N}^{l}\}$ denotes the image patch embeddings. A learnable token $\pmb{P}^{l}$ is appended to + +![](images/5dc812a037b8edb3e8e9bdb209255d9b3718d57b27accc4922a3496644d12a69.jpg) +(a) Pre-training stage: Meta-Prompt Pre-training Module + +![](images/760739a2d0f2a2c15efe37127809c8161409d5122e0d73af86cda2fd48426105.jpg) +(b) Fine-tuning stage: Quality-Aware Gradient Regularization +Figure 3. The overview of our GRMP-IQA. The core modules are the Meta-Prompt Pre-training and Quality-Aware Gradient Regularization, corresponding to two training processes. At the pre-training stage, we establish a distortion-specific meta-knowledge task for BIQA tasks, and the bi-level gradient descent is utilized to train visual-text meta-prompt $[\theta_T,\theta_V]$ (Sec. 3.3). These optimized prompts then serve as initial settings to efficiently adapt the CLIP model to IQA tasks with limited labels. At the few-shot fine-tuning stage, we first predict the class probability distributions $p^{sem}(w_i|x)$ and $p^{qua}(w_i|x)$ using the semantic and fine-tuned CLIP models, respectively. We then adjust IQA task gradients $G_{\mathrm{qua}}$ by clipping gradients aligned with semantic task gradients $G_{\mathrm{sem}}$ to generate refined gradient $G_{\mathrm{qgr}}$ for backward updates (Sec. 3.4), mitigating the influence of semantic noise on quality predictions. + +the token sequence in each ViT layer, and the Multi-Head Self-Attention (MHSA) module processes the tokens as: + +$$ +[ \boldsymbol {f} ^ {l}, \neg , \boldsymbol {H} ^ {l} ] = \operatorname {L a y e r} ^ {l} ([ \boldsymbol {f} ^ {l - 1}, \boldsymbol {P} ^ {l - 1}, \boldsymbol {H} ^ {l - 1} ]), \tag {1} +$$ + +where the output of $P^l$ is discarded and not passed to the next layer, serving only as a set of learnable parameters. + +Text Meta-Prompt. We adopt CoOp [56] as our textual meta-prompt, with learnable parameters $\theta_T$ . For each quality class $i$ , we construct a learnable quality prompt defined as $\pmb{q}_i = \{\pmb{v}_1, \pmb{v}_2, \dots, \pmb{v}_M, [\mathrm{quality}]\}$ . Here, [quality] $\in$ {"high quality", "low quality"} serves as a categorical quality marker, and $\{\pmb{v}_m\}_{m=1}^M$ denotes a set of $M$ learnable context vectors (with $M = 4$ in our work). + +# 3.3. Meta-Prompt Pre-training Module + +During meta-learning, we aim to optimize the learnable visual-text prompt $\theta = [\theta_T,\theta_V]$ . For an input image $x$ , the textual encoder $g(\cdot)$ processes the quality prompts $q_{i}$ while the visual encoder extracts the feature vector $f$ . The probability of predicting high quality is: + +$$ +p \left(\boldsymbol {q} _ {\text {h i g h}} \mid \boldsymbol {x}\right) = \frac {\exp \left(\left\langle \boldsymbol {g} \left(\boldsymbol {q} _ {\text {h i g h}}\right) , \boldsymbol {f} \right\rangle / \tau\right)}{\exp \left(\left\langle \boldsymbol {g} \left(\boldsymbol {q} _ {\text {h i g h}}\right) , \boldsymbol {f} \right\rangle / \tau\right) + \exp \left(\left\langle \boldsymbol {g} \left(\boldsymbol {q} _ {\text {l o w}}\right) , \boldsymbol {f} \right\rangle / \tau\right)}, +$$ + +where $q_{\mathrm{high}}$ and $q_{\mathrm{low}}$ denote high quality and low quality category prompt, $p(q_{\mathrm{high}}|\boldsymbol {x})$ represents the estimated probability that the quality of image $\mathbf{x}$ is "high quality", $\tau$ is a temperature parameter learned by CLIP, and $\langle ,\rangle$ denotes cosine similarity. The labeled quality scores are then rescaled to 0-1, denoted as $y$ , and the loss function is calculated as: + +$$ +\mathcal {L} = - (y \log (p \left(q _ {\text {h i g h}} | \boldsymbol {x}\right)) + (1 - y) \log (1 - p \left(q _ {\text {h i g h}} | \boldsymbol {x}\right)). \tag {3} +$$ + +Constructing Distortion Meta-Knowledge Task. As noted by [29, 58], the ability to detect various types of image distortions is crucial for developing BIQA models with strong generalization capabilities across diverse scenarios. Moreover, the efficacy of prompt tuning is significantly dependent on the initial configuration of the prompts. This initial setup greatly affects the CLIP vision-language model's ability to swiftly adapt to different IQA scenarios. Drawing inspiration from the "learn to learn" ethos inherent in the deep meta-learning paradigm [43, 58], we propose an optimization-based method for effectively pre-training visual-textual prompts. These prompts incorporate shared quality insights from various image distortions, thereby enhancing CLIP's swift adaptability to IQA tasks [58]. To investigate the general rules of image distortion, we first constructed a $K_{t}$ -way distortion-specific image quality prediction task, denoted as $\mathcal{T}_t$ . Then, it is used to build the meta-training set as $\mathcal{D}^{\mathrm{meta}} = \left\{\mathcal{D}_{\mathcal{T}_t}^{\mathrm{support}}, \mathcal{D}_{\mathcal{T}_t}^{\mathrm{query}}\right\}_{t=1}^T$ . Here, $\mathcal{D}_{\mathcal{T}_t}^{\mathrm{support}}$ and $\mathcal{D}_{\mathcal{T}_t}^{\mathrm{query}}$ represent the support and query sets for each task, respectively, with $T$ representing the total number of tasks. To simulate the process of prompt generalization to different distortions in BIQA scenarios, we randomly sample $k$ tasks as a mini-batch from the meta-training set, where $1 \leq k \leq T$ , to perform bi-level gradient optimization. + +Distortion-Aware Meta-prompt Learning. Our approach employs a bi-level gradient descent technique to bridge the learning process from the support to the query set. Specifically, it mainly consists of two optimization steps. In the Inner-Loop (the first level), we compute the gradients of the prompt parameters using the support set and apply the first update. In the Outer-Loop (the second level), we assess the + +performance of the updated model on the query set and optimize the parameters again. This bi-level structure distills shared quality priors by training on a number of NR-IQA tasks with known distortions, enabling the meta-prompt to generalize rapidly across diverse new BIQA scenarios. + +Inner-Loop. The objective of the Inner-Loop stage is to adapt the meta-prompt, denoted by $[\pmb{\theta}_{\pmb{T}},\pmb{\theta}_{\pmb{V}}]$ , to the $t^{th}$ support set $\mathcal{D}_{\mathcal{T}_t}^{\mathrm{support}}$ within the mini-batch. During the first level of updates, we determine the loss $\mathcal{L}(\pmb {\theta},\mathcal{D}_{\mathcal{T}_t}^{\mathrm{support}})$ , following which the model parameters are updated on the support set using the inner learning rate $\alpha$ , as specified by: + +$$ +\boldsymbol {\theta} _ {t} ^ {\prime} = \boldsymbol {\theta} - \alpha \nabla_ {\boldsymbol {\theta}} \mathcal {L} (\boldsymbol {\theta}, \mathcal {D} _ {\mathcal {T} _ {t}} ^ {\text {s u p p o r t}}), \tag {4} +$$ + +Outer-Loop. In a similar vein, the second level of updates adjusts the parameters $\theta_t^\prime$ based on the query set $\mathcal{D}_{\mathcal{T}_t}^{\mathrm{query}}$ : + +$$ +\boldsymbol {\theta} _ {t} \leftarrow \boldsymbol {\theta} _ {t} ^ {\prime} - \alpha \nabla_ {\boldsymbol {\theta} _ {t} ^ {\prime}} \mathcal {L} \left(\boldsymbol {\theta} _ {t} ^ {\prime}, \mathcal {D} _ {\mathcal {T} _ {t}} ^ {\text {q u e r y}}\right). \tag {5} +$$ + +For a mini-batch of meta-tasks, this process culminates in aggregating gradients from all tasks to update the final model parameters, following the update rule: + +$$ +\boldsymbol {\theta} \leftarrow \boldsymbol {\theta} - \beta \sum_ {t = 1} ^ {k} \nabla_ {\boldsymbol {\theta}} \mathcal {L} \left(\boldsymbol {\theta} _ {t} ^ {\prime}, \mathcal {D} _ {\mathcal {T} _ {t}} ^ {\text {q u e r y}}\right), \tag {6} +$$ + +where $\beta$ represents the outer learning rate. Meta-learning effectively trains the learnable prompts $\pmb{\theta} = [\pmb{\theta}_T, \pmb{\theta}_V]$ , ensuring their generalization across various image distortions. + +# 3.4. Quality-Aware Gradient Regularization + +We denote the original semantic CLIP model as $V^{\mathrm{sem}}$ and the model obtained from the Meta-Prompt Pre-training Module, as $V^{\mathrm{qua}}$ . The gradient for the IQA task is the quality gradient $G_{\mathrm{qua}}$ , while the gradient for the semantic direction is the semantic gradient $G_{\mathrm{sem}}$ . Recent studies [1, 53] have shown that IQA tasks can be misaligned with the high-level semantic representations of upstream tasks, leading to overfitting and reduced generalization. To prevent semantic overfitting, our Quality-aware Gradient Regularization (QGR) follows three key steps: derive a semantic direction $G_{\mathrm{sem}}$ , compute the quality optimization direction $G_{\mathrm{qua}}$ , and balance gradients based on their directional relationship. + +Specifically, to derive a general semantic direction $G_{\mathrm{sem}}$ , we first design a hard prompt following [49, 52]: $\boldsymbol{w}_i =$ "a photo of a [class]", where [class] represents one of nine categories: "animal", "cityscape", "human", "indoor", "landscape", "night", "plant", "still-life", "others". This prompt $\boldsymbol{w}_i$ is input into the text encoders of $V^{\mathrm{sem}}$ and $V^{\mathrm{qua}}$ , generating text features that align with image features $f^{qua}$ from their respective visual encoders. As a result, we obtain zero-shot semantic prediction probabilities $p^{\mathrm{sem}}(\boldsymbol{w}_i|\boldsymbol{x})$ and $p^{\mathrm{qua}}(\boldsymbol{w}_i|\boldsymbol{x})$ . To measure the semantic alignment between + +$V^{\mathrm{qua}}$ and $V^{\mathrm{sem}}$ , we compute the KL divergence: + +$$ +\mathcal {L} _ {\mathrm {k l}} \left(\boldsymbol {V} ^ {\text {q u a}}\right) = - \sum_ {i} \boldsymbol {p} ^ {\text {s e m}} \left(\boldsymbol {w} _ {i} \mid \boldsymbol {x}\right) \log \frac {\boldsymbol {p} ^ {\text {q u a}} \left(\boldsymbol {w} _ {i} \mid \boldsymbol {x}\right)}{\boldsymbol {p} ^ {\text {s e m}} \left(\boldsymbol {w} _ {i} \mid \boldsymbol {x}\right)}. \tag {7} +$$ + +The gradient of $\mathcal{L}_{\mathrm{kl}}(V^{\mathrm{qua}})$ , denoted as $G_{\mathrm{sem}}$ , represents the general semantic optimization direction. Similarly, the quality direction $G_{\mathrm{qua}}$ is obtained by computing the gradient of the cross-entropy loss $\mathcal{L}_{\mathrm{ce}}(V^{\mathrm{qua}})$ , which compares the predicted probability $p(q_{\mathrm{high}}|x)$ with the ground truth $y$ , as shown in Eq. 3. Finally, to balance the gradients between quality $G_{\mathrm{qua}}$ and semantics $G_{\mathrm{sem}}$ within a shared representation space, we analyze their relationship under two cases: (1) If the angle between $G_{\mathrm{qua}}$ and $G_{\mathrm{sem}}$ is less than 90 degrees, it indicates that the optimization directions for IQA quality knowledge and general semantic knowledge are consistent. In such cases, we clip the $G_{\mathrm{qua}}$ along its component $G_{\parallel}$ that parallel to the semantic direction $G_{\mathrm{sem}}$ , to modulate the model's original quality optimization path, preventing overfitting to semantic correlations. + +(2) Conversely, if the angle between two areas of knowledge is more than 90 degrees, it means they are heading in different directions during improvement. In such situations, we don't adjust the quality gradient $G_{\mathrm{qua}}$ , allowing the model to learn quality-aware features according to the original optimization direction. In conclusion, our gradient regularization strategy is mathematically formulated as: + +$$ +\boldsymbol {G} _ {\text {q g r}} = \left\{ \begin{array}{l l} \boldsymbol {G} _ {\text {q u a}} & \text {i f} \boldsymbol {G} _ {\text {q u a}} \cdot \boldsymbol {G} _ {\text {s e m}} \leq 0, \\ \boldsymbol {G} _ {\text {q u a}} - \lambda \frac {\boldsymbol {G} _ {\text {q u a}} \cdot \boldsymbol {G} _ {\text {s e m}}}{\| \boldsymbol {G} _ {\text {s e m}} \| ^ {2}} \boldsymbol {G} _ {\text {s e m}} & \text {o t h e r w i s e}, \end{array} \right. \tag {8} +$$ + +Here, $\lambda$ is introduced to generalize the formulation, providing flexibility in controlling the influence of general knowledge. Specifically, $\lambda = 1$ projects $G_{\mathrm{qua}}$ onto the orthogonal direction of $G_{\mathrm{sem}}$ , while $\lambda = 0$ reduces QGR to CoOp [56]. + +# 4. Experiments + +# 4.1. Datasets and Evaluation Protocols + +We conduct experiments on multiple BIQA datasets include LIVEC [7], KonIQ [10], LIVEFB [49], and SPAQ [3], which feature authentic distortions, and PIPAL [12], LIVE [38], and CSIQ [17], which feature synthetic distortions. LIVEC contains 1,162 mobile device images, SPAQ includes 11,125 photos from 66 smartphones, KonIQ has 10,073 images from open sources, and LIVEFB is the largest real-world dataset with 39,810 images. For synthetic distortions, LIVE and CSIQ contain 779 and 866 images with 5 and 6 types of distortions. PIPAL, a challenging dataset, includes 23,200 images with 40 types of distortions, including GAN-generated artifacts. We use Spearman's Rank Correlation Coefficient (SRCC) and the Pearson Linear Correlation Coefficient (PLCC) as metrics to quantify the monotonousness and accuracy of predictions. + +
MethodLIVECKonIQCSIQLIVEPIPAL
Labels5010020050100200501002005010020050100200
HyperIQA [41]0.6480.7250.7900.6150.7100.7760.7900.8240.9090.8920.9120.9290.1020.3020.379
MetaIQA [58]0.6040.6260.6690.6180.6200.6600.7840.8490.8940.8400.8800.9190.3320.3480.371
DEIQT [33]0.6670.7180.8120.6380.6820.7540.8210.8910.9410.9200.9420.9550.3960.4100.436
MANIQA [48]0.6420.7690.7970.6520.7550.8100.7940.8470.8740.9090.9280.9570.1360.3610.470
CONTRIQUE [30]0.6950.7290.7610.7330.7940.8210.8400.9260.9400.8910.9220.9430.3790.4370.488
Re-IQA [37]0.5910.6210.7010.6850.7230.7540.8930.9070.9230.8840.8940.9290.2800.3500.431
CLIP w/ Linear Probe [36]0.6640.7210.7330.7360.7700.7820.8410.8920.9410.8960.9230.9410.2540.3030.368
CLIP w/ CoOp [44]†0.6950.7380.7460.6920.7430.762---------
LIQE [52]0.6910.7690.8100.7590.8010.8320.8380.8910.9240.9040.9340.948---
GRepQ [40]0.7600.7910.8220.8120.8360.8550.8780.9140.9410.9260.9370.9530.390†0.450†0.498†
GRMP-IQA (Ours)0.8360.8570.8750.8530.8720.8830.8930.9170.9410.9320.9430.9680.4740.5120.546
+ +Table 1. SRCC performance comparison of our method with other IQA methods trained on limited labels. Bold indicates the best results, underlined marks the second-best, and the fifth-to-last through second-to-last lines show the CLIP-based IQA. † is our reproduction. + +
MethodLIVECSIQLIVECKonIQLIVEFBSPAQ
PLCCSRCCPLCCSRCCPLCCSRCCPLCCSRCCPLCCSRCCPLCCSRCC
Training Ratio20%20%20%20%20%20%20%20%10%10%20%20%
DEIQT [33]0.9680.9650.8850.8620.8220.7920.9080.8880.6240.5380.9120.908
LoDa [46]----0.8540.8150.9230.907----
GRMP-IQA (Ours)0.9720.9700.9580.9510.8970.8750.9310.9150.6860.6040.9250.920
Training Ratio80%80%80%80%80%80%80%80%80%80%80%80%
MetaIQA [58]0.9590.9600.9080.8990.8020.8350.8560.8870.5070.540--
CONTRIQUE [30]0.9610.9600.9550.9420.8570.8450.9060.8940.6410.5800.9190.914
DEIQT [33]0.9820.9800.9630.9460.8940.8750.9340.9210.6630.5710.9230.919
Re-IQA [37]0.9710.9700.9600.9470.8540.8400.9230.9140.7330.6450.9250.918
LIQE [52]0.9510.9700.9390.9360.9100.9040.9080.919----
CLIP-IQA+ [44]----0.8320.8050.9090.8950.5930.5750.8660.864
CDINet [54]0.9750.9770.9600.9520.8800.8650.9280.916--0.9220.919
QFM-IQM [22]0.9830.9810.9650.9540.9130.8910.9360.9220.6670.5670.9240.920
LoDa [46]0.9790.975--0.8990.8760.9440.9320.6790.5780.9280.925
GRMP-IQA (Ours)0.9830.9810.9740.9680.9160.8970.9450.9340.7040.6160.9320.927
+ +Table 2. Performance comparison measured by medians of SRCC and PLCC, and bold entries indicate the top two results. + +# 4.2. Implementation Details and Setups + +We build our model on CLIP-B/16 [36]. During pretraining, we optimize the visual-text prompt, while in finetuning, only the last four blocks of the image and text encoders are trained. For competing models, we use public results or re-train them under our setup. Each dataset is split into $80\% / 20\%$ for training and testing, with splits based on reference images to ensure content independence in datasets with synthetic distortions. To ensure robustness, we report the median performance across ten random splits. All experiments are run on four NVIDIA RTX3090 GPUs. + +Meta-Prompt Pre-training. We pre-train on TID2013 [32] and KADID-10K [25] dataset, which contain extensive distortion information. We set the learning rates $\alpha$ and $\beta$ to 1e-4 and 1e-2, and train for 50 epochs using Adam [16]. + +Fully Supervised Learning Setting. We randomly crop each input image into 10 patches of $224 \times 224$ resolution + +and train the model for 9 epochs using AdamW [26]. The learning rate is $5 \times 10^{-6}$ , with a scheduler over 9 decay epochs. The batch size is 16 for LIVEC and 128 for KonIQ. Few-Shot Learning Setting. In the few-shot setting, we follow the approach in a previous study [40] to train our GRMP-IQA model using randomly selected subsets of 50, 100, and 200 samples from the training set. The regularization weight $\lambda$ is fixed at 5, and the training hyperparameters are set according to the schedule proposed by CoOp [56]. + +# 4.3. Performance Comparison with SOTA + +Our method effectively acquires extensive image quality assessment knowledge, enabling it to provide powerful priors for various IQA scenarios. Tab. 1 and Tab. 2 summarize the comparative results for different experimental settings. + +Few-Shot Setting. Benefiting from the quality priors acquired during pre-training, our model achieves superior + +
TrainingLIVEFBLIVECKonIQLIVECSIQ
TestingKonIQLIVECKonIQLIVECCSIQLIVE
DBCNN0.7160.7240.7540.7550.7580.877
HyperIAQ0.7580.7350.7720.7850.7440.926
TReS0.7130.7400.7330.7860.761-
DEIQT0.7330.7810.7440.7940.7810.932
CLIP-IQA+0.6310.6200.6970.803--
LoDa0.7630.8050.7450.811--
GRMP-IQA0.7650.7900.7820.8310.8090.935
+ +Table 3. SRCC on the cross datasets validation. The best performances are highlighted in boldface. + +
ComponentLIVECKonIQ
Pre-training Meta-learningPLCCSRCCPLCCSRCC
0.8250.7960.7880.764
✓0.8230.7880.7920.761
✓0.8580.8280.8440.811
+ +performance even with limited training data. As shown in Tab. 1, our GRMP-IQA significantly outperforms the second-best model, GRepQ [40], which is specifically designed for few-shot learning. Additionally, our approach exhibits clear advantages over MetaIQA, a meta-learning-based method, underscoring the effectiveness of metaprompts for rapidly adapting CLIP to diverse IQA tasks. These results validate GRMP-IQA's strong capability in learning quality-aware feature with limited labeled samples. Fully Supervised Setting. Tab. 2 compares GRMP-IQA with other BIQA methods under full supervised setting, including self-training approaches like CONTRIQUE [30] and Re-IQA [37], as well as CLIP-based methods such as LIQE [52] and CLIP-IQA+ [44]. GRMP-IQA outperforms almost all competitors on six datasets. Achieving leading performance across these datasets is particularly challenging due to the diverse range of image content and distortion types. Notably, the proposed method achieves competitive results with the state-of-the-art (SOTA) method on several datasets while utilizing only $20\%$ of the training data compared to the fully supervised setting. + +# 4.4. Generalization Capability Validation + +To evaluate the generalization capacity of GRMP-IQA, we performed cross-dataset validation experiments, where the model was trained on one dataset and tested on others without parameter adjustments. Tab. 3 reports the SRCC results across five datasets. GRMP-IQA consistently outperforms state-of-the-art models in most cross Authentic scenarios, achieving notable improvements on LIVEC and + +Table 4. Ablation study on the effectiveness of meta-learning when trained with only 50 samples. + +
ComponentsLIVECKonIQPIPAL
MetaTextVisualPLCCSRCCPLCCSRCCPLCCSRCC
✓0.5790.5980.5920.5730.2160.203
0.6390.5890.5560.5540.3670.371
✓✓0.6990.6890.6790.6090.3230.312
✓✓0.7760.7420.7360.7010.3570.369
✓✓0.7590.7090.6220.5920.3620.396
✓✓✓0.8080.7700.7440.7130.4100.434
+ +Table 5. Ablation experiments with Meta-Prompt Pre-training component. The best performances are highlighted in boldface. + +
MethodLabels50100
DatasetPLCCSRCCPLCCSRCC
w/o QGRLIVEC0.8580.8280.8750.848
w/ QGRLIVEC0.8640.8360.8830.857
w/o QGRKonIQ0.8440.8110.8720.840
w/ QGRKonIQ0.8800.8530.8960.872
+ +Table 6. Ablation experiments with QGR in few-shot setting. The best performances are highlighted in boldface. + +KonIQ dataset. Additionally, it demonstrates strong competitiveness on synthetic datasets like LIVE and CSIQ. + +# 4.5. Ablation Study + +Effect of Meta-Prompt Pre-training Module. This module consists of three key components: meta-learning, text meta-prompts, and visual meta-prompts. We conduct ablation studies in Tab. 5 to evaluate their zero-shot capabilities across various datasets. The baseline uses a CLIP model pre-trained on classification tasks. Row 2 evaluates meta-learning without prompts, where CLIP's visual and text encoders are fine-tuned to learn distortion knowledge across different distortions, improving IQA performance. However, tuning CLIP's weights impairs its original generalization, resulting in lower performance on KonIQ. Rows 3 and 4 assess fine-tuning only text or visual prompts during meta-learning. These strategies notably boost zero-shot performance—particularly on real-world datasets like KonIQ and LIVEC—by preserving CLIP's generalization while adapting to IQA tasks. Row 5 explores prompt tuning without meta-learning, underscoring meta-learning's role in curbing overfitting and maintaining generalization. The strongest results emerge when all components are combined, highlighting their complementary strengths. + +Effect of Meta-learning. To further investigate whether the effectiveness of our method derives from meta-learning, we conducted an ablation study. Specifically, as shown in rows 1 and 2 of Tab. 4, when using Empirical Risk Minimization without meta-learning pre-training on two synthetic dataset, the fine-tuning performance on LIVEC and KonIQ dataset + +![](images/f29ceda765abccace48421cb5381e78efb8f16936ea8a846b45e3ef3fad816a4.jpg) +(a) + +![](images/4637e83833b552599c0a6fd83bb23b03105f64f0926f34abce0c511455680a46.jpg) +(b) + +
λ50100200
PLCCSRCCPLCCSRCCPLCCSRCC
10.8500.8120.8740.8390.8930.866
30.8640.8390.8900.8650.9030.878
50.8800.8530.8960.8720.9080.883
70.8760.8510.8950.8700.9050.881
+ +was even worse than the baseline without pre-training. This indicates that additional pre-training data does not necessarily enhance performance and may even lead to overfitting [19]. In contrast, our meta-learning strategy effectively extracts generalizable quality priors from external data, significantly enhancing performance across datasets. + +Effect of Gradient Regularization. We conducted ablation studies on the QGR module. As shown in Tab. 6 and Fig. 4(a), without QGR, models are prone to overfitting under limited training data, with rapid performance saturation and a slight decline, which ultimately hinders generalization. By modulating training gradients, QGR effectively mitigates overfitting and enhances adaptability to the test dataset. To illustrate QGR's impact, we analyzed the angular difference between gradients $G_{\mathrm{qua}}$ and $G_{\mathrm{sem}}$ during training, as shown in Fig. 4(b). Without QGR, the angle between $G_{\mathrm{qua}}$ and $G_{\mathrm{sem}}$ tends toward 90 degrees, reflecting the orthogonality typical of high-dimensional random vectors [2]. In contrast, QGR introduces greater angular variation, suggesting that it encourages the model to explore more correlations between quality and semantic directions. This improves the model's ability to capture quality-related information and mitigates overfitting on limited data. + +Effect of soft weight $\lambda$ . On the KonIQ dataset, we conduct ablation experiments with various soft weights $\lambda$ in Eq. 8 to examine their impact, as detailed in Tab. 7. Results show that a small $\lambda$ diminishes the effectiveness of our QGR, whereas a large $\lambda$ causes substantial gradient changes and decreases performance. Given our observation of the trade-off, we adopt $\lambda = 5$ in our experiments. + +Computational Analysis. As shown in Tab. 8, we evaluate + +![](images/87501095bc1cbad1403f79f581372fea43a009bb6e5a41b124fda62fabdce399.jpg) + +![](images/71c34f6566a07040641767070b78cedbad979b97a9875465b83d34892511aa13.jpg) +Best ours + +![](images/041b3ceeb80c7e3722a4fadce55cb5ea41bdf0d8a06bb3f176b5d9a360429e76.jpg) +Best DEIQT + +![](images/c7b275e19a672c94c75dcb0d0b94938126c7b5072adf2c4b6b1752f36f1a7ce8.jpg) + +![](images/8633e7f956e2a7d65bdaec1d827e8142c963a84bdc4afb8365aeab8aec95dba0.jpg) +Best ours + +![](images/fce79b2080d526aa9879f55f54a010bb29e8181b5170112356030a85b310bb4a.jpg) + +![](images/bf457257d3440694fb7015a495b185e1c367b944972727510da9b80356ad4ffa.jpg) +Worst DEIQT + +![](images/a5f32b7f336a4237807150d53b768bdcaa132e49b4a4fdb05ac1a487eba338ce.jpg) +Best DEIQT + +![](images/cea95d3a7e6ecbb9a031df5986daea0c42a317cece6f08ee536caded7df2750c.jpg) +Figure 4. (a) SRCC value during training with 50 samples. (b) Angles between $G_{qua}$ and $G_{sem}$ during training on KonIQ. Under the constraints of QGR, our method captures more correlation between quality knowledge and general semantic knowledge. +Worst ours +(a) +Figure 5. gMAD results between DEIQT [33] and our GRMP-IQA. (a) Fixed DEIQT at low quality. (b) Fixed DEIQT at high quality. (c) Fixed ours at low quality. (d) Fixed ours at high quality. + +![](images/d1116aec69a3babc426d88e89cc02072fa4be7f03290cd301f91af0abe9a1ea2.jpg) +Worst ours +(b) + +![](images/c0356403e51cceb9f6c17e0327a993665552725d750b9344ac2f0077f5c804a5.jpg) +(c) + +![](images/cd6d31eb3b740362ca7955d0b9ebeedbb3aaf22da2f466a261beeaaa44b24113.jpg) +(d) + +Table 7. The ablation study about soft weight $\lambda$ in Eq. 8. + +
MethodParamEfficiencyPerformance
Throughput ↑Latency ↓80% Train20% Train
Q-Align8.2B9.9 (img/sec)101.0 (ms)0.941 / 0.9400.901 / 0.903
Ours151M43.0 (img/sec)21.3 (ms)0.945 / 0.9340.931 / 0.915
+ +Table 8. Comparison with Q-Align across varying training data ratio to evaluate efficiency and PLCC/SRCC on the KonIQ dataset. + +the inference latency and throughput of our method on an RTX3090 GPU, surpassing Q-Align [45] in both efficiency and accuracy. Notably, our model has only 41M learnable and 151M total parameters, far fewer than Q-Align's 8.2B. + +# 4.6. Qualitative Analysis + +The gMAD competition [28] is a standard method for assessing IQA model robustness. It selects image pairs where the attacker model predicts a large quality difference, while the defender perceives them as similar. Observers then evaluate these pairs to gauge model generalization. As shown in Fig. 5, when our model acts as the defender, the attacker's selected pairs exhibit minimal perceptual quality changes. Conversely, as the attacker, our model consistently identifies pairs with significant quality differences. This highlights its strong defensive and offensive capabilities. Notably, in the second column, DEIQT misclassifies semantically similar images as having comparable quality, whereas our model accurately distinguishes their differences. + +# 5. Conclusion + +In this paper, we propose the GRMP-IQA framework, which generalizes well with limited data. It includes a meta-learning pre-training module that enables the CLIP model to rapidly adapt to IQA tasks and an adaptive gradient regulation module that refines gradient trajectories during fin-tuning, focusing updates on quality-aware knowledge while minimizing the negative impact of over-reliance on semantic noise. Comprehensive experiments on various BIQA datasets validate the superior generalization of our framework, especially in data-scarce scenarios. + +Acknowledgements. This work was supported by the National Science Fund for Distinguished Young Scholars (No.62025603), the National Natural Science Foundation of China (No. U21B2037, No. U22B2051, No. U23A20383, No. U21A20472, No. 62176222, No. 62176223, No. 62176226, No. 62072386, No. 62072387, No. 62072389, No. 62002305 and No. 62272401), and the Natural Science Foundation of Fujian Province of China (No. 2021J06003, No.2022J06001). + +# References + +[1] Nithin C Babu, Vignesh Kannan, and Rajiv Soundararajan. No reference opinion unaware quality assessment of authentically distorted images. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2459-2468, 2023. 2, 3, 5 +[2] Tony Cai, Jianqing Fan, and Tiefeng Jiang. Distributions of angles in random packing on spheres. The Journal of Machine Learning Research, 14(1):1837-1864, 2013. 8 +[3] Yuming Fang, Hanwei Zhu, Yan Zeng, Kede Ma, and Zhou Wang. Perceptual quality assessment of smartphone photography. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3677-3686, 2020. 5 +[4] Yuqian Fu, Yanwei Fu, and Yu-Gang Jiang. Meta-fdmixup: Cross-domain few-shot learning guided by labeled target data. In Proceedings of the 29th ACM international conference on multimedia, pages 5326-5334, 2021. 3 +[5] Hongcheng Gao, Tianyu Pang, Chao Du, Taihang Hu, Zhijie Deng, and Min Lin. Meta-unlearning on diffusion models: Preventing relearning unlearned concepts. arXiv preprint arXiv:2410.12777, 2024. 1 +[6] Hongcheng Gao, Jiashu Qu, Jingyi Tang, Baolong Bi, Yue Liu, Hongyu Chen, Li Liang, Li Su, and Qingming Huang. Exploring hallucination of large multimodal models in video understanding: Benchmark, analysis and mitigation. arXiv preprint arXiv:2503.19622, 2025. 1 +[7] Deepti Ghadiyaram and Alan C Bovik. Massive online crowdsourced study of subjective and objective picture quality. IEEE Transactions on Image Processing, 25(1):372-387, 2015. 5 +[8] S Alireza Golestaneh, Saba Dadsetan, and Kris M Kitani. No-reference image quality assessment via transformers, relative ranking, and self-consistency. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1220-1230, 2022. 3 +[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3 +[10] Vlad Hosu, Hanhe Lin, Tamas Sziranyi, and Dietmar Saupe. Koniq-10k: An ecologically valid database for deep learning of blind image quality assessment. IEEE Transactions on Image Processing, 29:4041–4056, 2020. 5 +[11] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Vi + +sual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 2, 3 +[12] Gu Jinjin, Cai Haoming, Chen Haoyu, Ye Xiaoxing, Jimmy S Ren, and Dong Chao. Pipal: a large-scale image quality assessment dataset for perceptual image restoration. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI 16, pages 633-651. Springer, 2020. 5 +[13] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5148-5157, 2021. 3 +[14] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19113-19122, 2023. 2, 3 +[15] Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Self-regulating prompts: Foundational model adaptation without forgetting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15190-15200, 2023. 3 +[16] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[17] Eric Cooper Larson and Damon Michael Chandler. Most apparent distortion: full-reference image quality assessment and the role of strategy. Journal of electronic imaging, 19 (1):011006, 2010. 5 +[18] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 2 +[19] Aobo Li, Jinjian Wu, Yongxu Liu, and Leida Li. Bridging the synthetic-to Authentic gap: Distortion-guided unsupervised domain adaptation for blind image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28422-28431, 2024. 8 +[20] Chunyi Li, Zicheng Zhang, Haoning Wu, Wei Sun, Xiongkuo Min, Xiaohong Liu, Guangtao Zhai, and Weisi Lin. Agiqa-3k: An open database for ai-generated image quality assessment. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 1 +[21] Juncheng Li, Minghe Gao, Longhui Wei, Siliang Tang, Wenqiao Zhang, Mengze Li, Wei Ji, Qi Tian, Tat-Seng Chua, and Yueting Zhuang. Gradient-regulated meta-prompt learning for generalizable vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2551-2562, 2023. 3 +[22] Xudong Li, Timin Gao, Runze Hu, Yan Zhang, Shengchuan Zhang, Xiawu Zheng, Jingyuan Zheng, Yunhang Shen, Ke Li, Yutao Liu, et al. Adaptive feature selection for no-reference image quality assessment by mitigating semantic noise sensitivity. In International Conference on Machine Learning, pages 27808-27821. PMLR, 2024. 2, 3, 6 +[23] Xudong Li, Runze Hu, Jingyuan Zheng, Yan Zhang, Shengchuan Zhang, Xiawu Zheng, Ke Li, Yunhang Shen, + +Yutao Liu, Pingyang Dai, et al. Integrating global context contrast and local sensitivity for blind image quality assessment. In International Conference on Machine Learning, pages 27920-27941. PMLR, 2024. 3 +[24] Xudong Li, Wenjie Nie, Yan Zhang, Runze Hu, Ke Li, Xiawu Zheng, and Liujuan Cao. Distilling spatially-heterogeneous distortion perception for blind image quality assessment. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 2344-2354, 2025. 3 +[25] Hanhe Lin, Vlad Hosu, and Dietmar Saupe. Kadid-10k: A large-scale artificially distorted iqa database. In 2019 Eleventh International Conference on Quality of Multimedia Experience (QoMEX), pages 1-3. IEEE, 2019. 6 +[26] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 6 +[27] Ziwei Luo, Fredrik K Gustafsson, Zheng Zhao, Jens Sjolund, and Thomas B Schon. Controlling vision-language models for multi-task image restoration. In International Conference on Learning Representations, 2024. 2 +[28] Kede Ma, Qingbo Wu, Zhou Wang, Zhengfang Duanmu, Hongwei Yong, Hongliang Li, and Lei Zhang. Group mad competition-a new methodology to compare objective image quality models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1664-1673, 2016. 8 +[29] Kede Ma, Xuelin Liu, Yuming Fang, and Eero P Simoncelli. Blind image quality assessment by learning from multiple annotators. In 2019 IEEE international conference on image processing (ICIP), pages 2344-2348. IEEE, 2019. 4 +[30] Pavan C Madhusudana, Neil Birkbeck, Yilin Wang, Balu Adsumilli, and Alan C Bovik. Image quality assessment using contrastive learning. IEEE Transactions on Image Processing, 31:4149-4161, 2022. 6, 7 +[31] Jinyoung Park, Juyeon Ko, and Hyunwoo J Kim. Prompt learning via meta-regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26940–26950, 2024. 3 +[32] Nikolay Ponomarenko, Lina Jin, Oleg Ieremeiev, Vladimir Lukin, Karen Egiazarian, Jaakko Astola, Benoit Vozel, Kacem Chehdi, Marco Carli, Federica Battisti, et al. Image database tid2013: Peculiarities, results and perspectives. Signal processing: Image communication, 30:57-77, 2015. 6 +[33] Guanyi Qin, Runze Hu, Yutao Liu, Xiawu Zheng, Haotian Liu, Xiu Li, and Yan Zhang. Data-efficient image quality assessment with attention-panel decoder. In Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence, 2023. 2, 3, 6, 8 +[34] B Qu, H Li, and W Gao. Bringing textual prompt to a-generated image quality assessment, in 2024 IEEE international conference on multimedia and expo (icme), 2024. 1 +[35] Bowen Qu, Xiaoyu Liang, Shangkun Sun, and Wei Gao. Exploring aigc video quality: A focus on visual harmony videotext consistency and domain distribution gap. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6652-6660, 2024. 1 + +[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 6 +[37] Avinab Saha, Sandeep Mishra, and Alan C Bovik. Reiqa: Unsupervised learning for image quality assessment in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5846-5855, 2023. 2, 3, 6, 7 +[38] Hamid R Sheikh, Muhammad F Sabir, and Alan C Bovik. A statistical evaluation of recent full reference image quality assessment algorithms. IEEE Transactions on image processing, 15(11):3440-3451, 2006. 1, 5 +[39] Tianshu Song, Leida Li, Deqiang Cheng, Pengfei Chen, and Jinjian Wu. Active learning-based sample selection for label-efficient blind image quality assessment. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 1 +[40] Suhas Srinath, Shankhanil Mitra, Shika Rao, and Rajiv Soundararajan. Learning generalizable perceptual representations for data-efficient no-reference image quality assessment. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 22-31, 2024. 2, 3, 6, 7 +[41] Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3667-3676, 2020. 6 +[42] Zhengzhong Tu, Yilin Wang, Neil Birkbeck, Balu Adsumilli, and Alan C Bovik. Ugc-vqa: Benchmarking blind video quality assessment for user generated content. IEEE Transactions on Image Processing, 30:4449-4464, 2021. 1 +[43] Joaquin Vanschoren. Meta-learning: A survey. arXiv preprint arXiv:1810.03548, 2018. 4 +[44] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 2, 3, 6, 7 +[45] Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Yixuan Gao, Annan Wang, Erli Zhang, Wenxiu Sun, et al. Q-align: Teaching lmmss for visual scoring via discrete text-defined levels. In International Conference on Machine Learning, pages 54015-54029. PMLR, 2024. 2, 8 +[46] Kangmin Xu, Liang Liao, Jing Xiao, Chaofeng Chen, Haoning Wu, Qiong Yan, and Weisi Lin. Boosting image quality assessment through efficient transformer adaptation with local feature enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2662-2672, 2024. 2, 3, 6 +[47] Zhengzhuo Xu, Bowen Qu, Yiyan Qi, Sinan Du, Chengjin Xu, Chun Yuan, and Jian Guo. Chartmoe: Mixture of diversely aligned expert connector for chart understanding. arXiv preprint arXiv:2409.03277, 2024. 3 +[48] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. + +Maniqa: Multi-dimension attention network for no-reference image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1191-1200, 2022. 1, 6 +[49] Zhenqiang Ying, Haoran Niu, Praful Gupta, Dhruv Mahajan, Deepti Ghadiyaram, and Alan Bovik. From patches to pictures (paq-2-piq): Mapping the perceptual space of picture quality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3575–3585, 2020. 5 +[50] Ji Zhang, Shihan Wu, Lianli Gao, Heng Tao Shen, and Jingkuan Song. Dept: Decoupled prompt tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12924-12933, 2024. 3 +[51] Weixia Zhang, Kede Ma, Jia Yan, Dexiang Deng, and Zhou Wang. Blind image quality assessment using a deep bilinear convolutional neural network. IEEE Transactions on Circuits and Systems for Video Technology, 30(1):36-47, 2018. 3 +[52] Weixia Zhang, Guangtao Zhai, Ying Wei, Xiaokang Yang, and Kede Ma. Blind image quality assessment via vision-language correspondence: A multitask learning perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14071-14081, 2023. 2, 3, 5, 6, 7 +[53] Kai Zhao, Kun Yuan, Ming Sun, Mading Li, and Xing Wen. Quality-aware pre-trained models for blind image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22302-22313, 2023. 2, 3, 5 +[54] Limin Zheng, Yu Luo, Zihan Zhou, Jie Ling, and Guanghui Yue. Cdinet: Content distortion interaction network for blind image quality assessment. IEEE Transactions on Multimedia, 2024. 6 +[55] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16816-16825, 2022. 3 +[56] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 2, 3, 4, 5, 6 +[57] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15659-15669, 2023. 2 +[58] Hancheng Zhu, Leida Li, Jinjian Wu, Weisheng Dong, and Guangming Shi. Metaiaq: Deep meta-learning for no-reference image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14143-14152, 2020. 3, 4, 6 \ No newline at end of file diff --git a/2025/Few-Shot Image Quality Assessment via Adaptation of Vision-Language Models/images.zip b/2025/Few-Shot Image Quality Assessment via Adaptation of Vision-Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..11bc8d6f3b258875cf3a898e9152cc53be221dec --- /dev/null +++ b/2025/Few-Shot Image Quality Assessment via Adaptation of Vision-Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e356fd11d6a4017695f8d7d1dfeca9f67e4842f0bd6bff26fa7c248e8f563890 +size 797923 diff --git a/2025/Few-Shot Image Quality Assessment via Adaptation of Vision-Language Models/layout.json b/2025/Few-Shot Image Quality Assessment via Adaptation of Vision-Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4eb9f182399f164eebc78fef9fcd523f44a39fc6 --- /dev/null +++ b/2025/Few-Shot Image Quality Assessment via Adaptation of Vision-Language Models/layout.json @@ -0,0 +1,10695 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 59, + 103, + 552, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 103, + 552, + 121 + ], + "spans": [ + { + "bbox": [ + 59, + 103, + 552, + 121 + ], + "type": "text", + "content": "Few-Shot Image Quality Assessment via Adaptation of Vision-Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 123, + 142, + 487, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 142, + 487, + 172 + ], + "spans": [ + { + "bbox": [ + 123, + 142, + 487, + 172 + ], + "type": "text", + "content": "Xudong Li" + }, + { + "bbox": [ + 123, + 142, + 487, + 172 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 123, + 142, + 487, + 172 + ], + "type": "text", + "content": " Zihao Huang" + }, + { + "bbox": [ + 123, + 142, + 487, + 172 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 123, + 142, + 487, + 172 + ], + "type": "text", + "content": " Yan Zhang" + }, + { + "bbox": [ + 123, + 142, + 487, + 172 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 123, + 142, + 487, + 172 + ], + "type": "text", + "content": " Yunhang Shen" + }, + { + "bbox": [ + 123, + 142, + 487, + 172 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 123, + 142, + 487, + 172 + ], + "type": "text", + "content": " Ke Li" + }, + { + "bbox": [ + 123, + 142, + 487, + 172 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 123, + 142, + 487, + 172 + ], + "type": "text", + "content": " Xiawu Zheng" + }, + { + "bbox": [ + 123, + 142, + 487, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 123, + 142, + 487, + 172 + ], + "type": "text", + "content": " Liujuan Cao" + }, + { + "bbox": [ + 123, + 142, + 487, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 123, + 142, + 487, + 172 + ], + "type": "text", + "content": " Rongrong Ji" + }, + { + "bbox": [ + 123, + 142, + 487, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 95, + 171, + 509, + 227 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 118, + 171, + 492, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 171, + 492, + 199 + ], + "spans": [ + { + "bbox": [ + 118, + 171, + 492, + 199 + ], + "type": "text", + "content": "1 Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University, 361005, P.R. China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 95, + 198, + 509, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 198, + 509, + 227 + ], + "spans": [ + { + "bbox": [ + 95, + 198, + 509, + 227 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 95, + 198, + 509, + 227 + ], + "type": "text", + "content": " Beijing Institute of Technology " + }, + { + "bbox": [ + 95, + 198, + 509, + 227 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 95, + 198, + 509, + 227 + ], + "type": "text", + "content": " Tencent Youtu Lab, Shanghai, China {1xd761050753, huangzihhhh, bzhy986}@gmail.com, {zhengxiawu, rrji}@xmu.edu.cn" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 152, + 266, + 200, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 266, + 200, + 280 + ], + "spans": [ + { + "bbox": [ + 152, + 266, + 200, + 280 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 292, + 296, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 292, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 54, + 292, + 296, + 602 + ], + "type": "text", + "content": "Image Quality Assessment (IQA) remains an unresolved challenge in computer vision due to complex distortions, diverse image content, and limited data availability. Existing Blind IQA (BIQA) methods largely rely on extensive human annotations, which are labor-intensive and costly due to the demanding nature of creating IQA datasets. To reduce this dependency, we propose the Gradient-Regulated MetaPrompt IQA Framework (GRMP-IQA), designed to efficiently adapt the visual-language pre-trained model, CLIP, to IQA tasks, achieving high accuracy even with limited data. GRMP-IQA consists of two core modules: (i) MetaPrompt Pre-training Module and (ii) Quality-Aware Gradient Regularization. The Meta Prompt Pre-training Module leverages a meta-learning paradigm to pre-train soft prompts with shared meta-knowledge across different distortions, enabling rapid adaptation to various IQA tasks. On the other hand, the Quality-Aware Gradient Regularization is designed to adjust the update gradients during fine-tuning, focusing the model's attention on quality-relevant features and preventing overfitting to semantic information. Extensive experiments on standard BIQA datasets demonstrate the superior performance to the state-of-the-art BIQA methods under limited data setting. Notably, utilizing just " + }, + { + "bbox": [ + 54, + 292, + 296, + 602 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 54, + 292, + 296, + 602 + ], + "type": "text", + "content": " of the training data, GRMP-IQA is competitive with most existing fully supervised BIQA approaches. Our code is available via https://github.com/LXDxmu/GRMP-IQA." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 626, + 135, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 626, + 135, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 626, + 135, + 639 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "type": "text", + "content": "With the rise of the mobile internet era, the focus on computer vision has transitioned from initial concerns with compression and image processing [38] to handling user-generated content like smartphone photos and videos [6," + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 319, + 255, + 552, + 411 + ], + "blocks": [ + { + "bbox": [ + 319, + 255, + 552, + 411 + ], + "lines": [ + { + "bbox": [ + 319, + 255, + 552, + 411 + ], + "spans": [ + { + "bbox": [ + 319, + 255, + 552, + 411 + ], + "type": "image", + "image_path": "a0e96b5a28fe707ed86d5577d049d86dc08d327f699e87aa4584757a3ed9bf22.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 413, + 509, + 422 + ], + "lines": [ + { + "bbox": [ + 361, + 413, + 509, + 422 + ], + "spans": [ + { + "bbox": [ + 361, + 413, + 509, + 422 + ], + "type": "text", + "content": "(b) Fine-tuning : Quality-Aware Gradient Regularization" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 430, + 555, + 552 + ], + "lines": [ + { + "bbox": [ + 313, + 430, + 555, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 555, + 552 + ], + "type": "text", + "content": "Figure 1. Intuitive diagram of GRMP-IQA. (a) demonstrates using meta-learning for efficient soft prompt initialization with quality prior that enhances zero-shot generalization (" + }, + { + "bbox": [ + 313, + 430, + 555, + 552 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 313, + 430, + 555, + 552 + ], + "type": "inline_equation", + "content": "24\\%" + }, + { + "bbox": [ + 313, + 430, + 555, + 552 + ], + "type": "text", + "content": "), enabling the CLIP to adapt to BIQA tasks effectively. (b) illustrates gradient regularization during fine-tuning, which clips quality gradients aligned with semantic bias to guide the model toward quality-relevant features and reduce overly reliance on semantic content. In the t-SNE visualization, black dots represent noise samples with high semantic confidence (confidence " + }, + { + "bbox": [ + 313, + 430, + 555, + 552 + ], + "type": "inline_equation", + "content": ">0.8" + }, + { + "bbox": [ + 313, + 430, + 555, + 552 + ], + "type": "text", + "content": ") but incorrect quality predictions. The noticeable decrease (" + }, + { + "bbox": [ + 313, + 430, + 555, + 552 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 313, + 430, + 555, + 552 + ], + "type": "inline_equation", + "content": "73\\%" + }, + { + "bbox": [ + 313, + 430, + 555, + 552 + ], + "type": "text", + "content": ") in black dots highlights reduced semantic noise interference in IQA tasks." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": "35, 42], and lately to AI-generated content [5, 20]. This evolution has sharply increased demand for effective Blind Image Quality Assessment (BIQA) techniques, highlighting the importance of developing methodologies that can adeptly evaluate image quality without reference images. Data-driven BIQA models [34, 39, 48] based on deep neural networks have made significant progress in recent years. However, the quality scores for distorted images are often measured using the Mean Opinion Score (MOS), which is the average of multiple ratings (sometimes up to 120). As a result, acquiring a sufficient number of IQA training samples is quite labor-intensive and cost-expensive." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 227, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 227, + 714 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 227, + 714 + ], + "type": "text", + "content": "* Equal contribution. † Corresponding authors." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10442" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 360 + ], + "type": "text", + "content": "To mitigate this challenge, recent BIQA approaches [33, 37, 40, 46, 53] have leveraged large-scale image datasets through transfer learning or self-supervised learning. While these approaches improve performance, they often rely on computationally intensive pretraining or require domain-specific pretext tasks, limiting scalability. On the other hand, the vision-language model CLIP [36] has shown remarkable generalization across various downstream tasks, offering a potential alternative. Specifically, CLIP-IQA [44] demonstrates promising zero-shot performance in IQA by leveraging simple handcrafted prompts (e.g., \"Good/Bad photo\"). However, a significant performance gap remains compared to fully supervised methods [40, 45], highlighting the need to leverage limited labeled data to further close this performance gap. To bridge this gap, we introduce the concept of Few-Shot IQA, which aims to enable models to quickly generalize to new distortion types and unseen scenarios with only a small amount of labeled data (e.g., 50-200 training images). A natural way to adapt CLIP to downstream IQA tasks under this setup is through prompt tuning [11, 56, 57]. Prompt tuning techniques learn soft textual prompts (continuous embeddings) from a small labeled dataset while keeping the pretrained model parameters frozen, offering a lightweight and efficient solution." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 361, + 296, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 296, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 296, + 518 + ], + "type": "text", + "content": "However, as shown in Tab. 1, the direct application of the prompt tuning method CoOp [56] to IQA yields minimal improvement over simple linear probing on the CLIP image encoder, falling short of the expected benefits seen in high-level vision tasks. We attribute this limitation to a fundamental mismatch between CLIP's pretraining objectives and the specific requirements of IQA tasks. CLIP is pretrained to align text and images in a shared semantic space, with a focus on high-level semantics such as objects and scenes. In contrast, IQA demands fine-grained sensitivity to low-level distortion cues—an ability in which CLIP is relatively deficient [27, 52]. This discrepancy introduces two key challenges when applying prompt tuning to IQA tasks:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 519, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 55, + 519, + 296, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 296, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 296, + 640 + ], + "type": "text", + "content": "(1) Sensitivity to Initialization: During prompt tuning, with CLIP's parameters frozen and limited sensitivity to distortions, the model depends heavily on the learned prompts for quality assessment. Different prompt initializations lead to distinct optimization directions [18], which may steer optimization toward features less relevant to quality, causing significant performance variability. As shown in Fig. 2, the average SRCC fluctuates considerably across different random initializations, requiring careful tuning for each IQA scenario and limiting the model's adaptability." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "(2) Overfitting from Semantic Bias: Due to CLIP's inherent reliance on semantic information, fine-tuning with a small sample size often leads to overfitting, where the model learns spurious correlations (e.g., over-reliance on semantic features to infer quality [1, 22]) rather than true image quality indicators, which degrades IQA performance." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 316, + 69, + 447, + 158 + ], + "blocks": [ + { + "bbox": [ + 316, + 69, + 447, + 158 + ], + "lines": [ + { + "bbox": [ + 316, + 69, + 447, + 158 + ], + "spans": [ + { + "bbox": [ + 316, + 69, + 447, + 158 + ], + "type": "image", + "image_path": "d325e96ca932085b18e3dfbcafe7523e8fa82eb91ba75f3e9663db88f3b77ac4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 166, + 555, + 211 + ], + "lines": [ + { + "bbox": [ + 313, + 166, + 555, + 211 + ], + "spans": [ + { + "bbox": [ + 313, + 166, + 555, + 211 + ], + "type": "text", + "content": "Figure 2. The fine-tuning results on the LIVEC dataset using CoOp [56] show that prompt tuning accuracy is highly sensitive to initial random initialization of the prompt, especially with limited training data, underscoring its critical role under data constraints." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 448, + 70, + 553, + 158 + ], + "blocks": [ + { + "bbox": [ + 448, + 70, + 553, + 158 + ], + "lines": [ + { + "bbox": [ + 448, + 70, + 553, + 158 + ], + "spans": [ + { + "bbox": [ + 448, + 70, + 553, + 158 + ], + "type": "image", + "image_path": "5e8f9120d7dfdcaaa83bd8c6fdc177716c72b5a54bc38e3f18196f9a16388e62.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 222, + 555, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 222, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 222, + 555, + 281 + ], + "type": "text", + "content": "As illustrated in Fig. 1, when the CLIP encoder is finetuned on a limited dataset, the latent space contains many instances (black dots) that confidently and accurately predict image semantics (with confidence levels above 0.8) but fail to reliably assess image quality." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 282, + 556, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 556, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 556, + 617 + ], + "type": "text", + "content": "To overcome these limitations, we propose Gradient-Regulated Meta-Prompt Learning for IQA (GRMP-IQA), a framework with two key components: (i) Meta Prompt Pre-training (MPP) mitigates sensitivity to prompt initialization by incorporating generalized distortion priors into the initialization of soft prompts, which enhances their adaptability to new IQA scenarios. By designing metatraining tasks with a well-annotated dataset containing diverse distortions (e.g., overexposure, blur) and optimizing prompts through bi-level gradient descent, we enable the model to learn shared meta-knowledge of quality representations across different distortions, thereby improving generalization across IQA tasks. (ii) Quality-Aware Gradient Regularization (QGR) addresses overfitting to semantic information by regulating gradient updates during fine-tuning, which balances the interplay between the quality knowledge gradient " + }, + { + "bbox": [ + 313, + 282, + 556, + 617 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{qua}}" + }, + { + "bbox": [ + 313, + 282, + 556, + 617 + ], + "type": "text", + "content": " and the semantic knowledge gradient " + }, + { + "bbox": [ + 313, + 282, + 556, + 617 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{sem}}" + }, + { + "bbox": [ + 313, + 282, + 556, + 617 + ], + "type": "text", + "content": ". By clipping " + }, + { + "bbox": [ + 313, + 282, + 556, + 617 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{qua}}" + }, + { + "bbox": [ + 313, + 282, + 556, + 617 + ], + "type": "text", + "content": " along " + }, + { + "bbox": [ + 313, + 282, + 556, + 617 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{sem}}" + }, + { + "bbox": [ + 313, + 282, + 556, + 617 + ], + "type": "text", + "content": " when their directions are overly aligned, this module suppresses semantic bias and guides the model to focus on distortion-related features, improving its ability to capture low-level quality details. Additionally, fine-tuning only CLIP's text branch can misalign quality perception between the image and text branches, limiting generalization [14]. To address this, we integrate text prompt tuning (CoOp [56]) and visual prompt tuning (VPT [11]) by jointly meta-learning initializations for both, ensuring complementary optimization to better adapt CLIP to new IQA scenarios (Tab. 5)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 326, + 617, + 525, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 617, + 525, + 628 + ], + "spans": [ + { + "bbox": [ + 326, + 617, + 525, + 628 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 630, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 314, + 630, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 630, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 314, + 630, + 553, + 689 + ], + "type": "text", + "content": "- We propose a Meta Prompt Pre-training method that organizes meta tasks based on image distortion types and optimizes soft textual and visual prompts to gain shared quality meta-knowledge, enabling CLIP to rapidly adapt across various IQA scenarios." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 689, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 314, + 689, + 553, + 714 + ], + "type": "text", + "content": "- We develop a novel Quality-Aware Gradient Regularization method that clips gradients aligned with semantic" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10443" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 294, + 108 + ], + "type": "text", + "content": "directions to balance semantic and quality information during fine-tuning, ensuring the model prioritizes image quality while still integrating relevant semantic context." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 109, + 295, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 109, + 295, + 156 + ], + "spans": [ + { + "bbox": [ + 56, + 109, + 295, + 156 + ], + "type": "text", + "content": "- Extensive empirical results confirm that our approach is both effective and efficient. Notably, with just 200 data samples, our method is competitive with SOTA models on the LIVEC dataset (using " + }, + { + "bbox": [ + 56, + 109, + 295, + 156 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 56, + 109, + 295, + 156 + ], + "type": "text", + "content": " of the training data)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 168, + 142, + 181 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 168, + 142, + 181 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 142, + 181 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 189, + 251, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 189, + 251, + 202 + ], + "spans": [ + { + "bbox": [ + 55, + 189, + 251, + 202 + ], + "type": "text", + "content": "2.1. Deep Learning Based BIQA methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 207, + 295, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 207, + 295, + 469 + ], + "spans": [ + { + "bbox": [ + 55, + 207, + 295, + 469 + ], + "type": "text", + "content": "Early CNN-based IQA methods followed standard pretraining and fine-tuning pipelines [9, 51], while meta-learning approaches like MetaIQA [58] improved adaptability from synthetic to real-world images. However, CNNs struggle with non-local features, a gap addressed by ViT-based models [8, 13, 23, 24]. Recent methods like LIQE [52] leverage CLIP's multi-task learning, enhancing IQA through supervised fine-tuning across datasets. While effective, these methods rely heavily on extensive annotations, making them costly and time-intensive. In contrast, limited-data BIQA remains underexplored. CLIP-IQA[44] shows promise in zero-shot settings but has performance limits, highlighting the need for better use of limited labeled data. Methods like DEIQT[33] and LoDa[46] demonstrate that fine-tuning pre-trained ViTs reduces annotation needs but still underperform in few-shot scenarios. Self-supervised approaches[37, 40, 53] ease data constraints through tailored pre-training tasks but carry high computational costs. Additionally, recent studies [1, 22] explore disentangling semantic and distortion-related content for IQA—yet none fully address the dual challenges of data efficiency and computational cost." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 471, + 295, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 471, + 295, + 543 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 295, + 543 + ], + "type": "text", + "content": "To bridge these gaps, our approach combines CLIP's pre-trained semantic knowledge with distortion-specific fine-tuning. By using prompt tuning to cut pre-training costs and a parameter-free regularization strategy (leveraging high- and low-level gradient correlations), it optimizes IQA more effectively with fewer labeled samples." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 552, + 149, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 149, + 565 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 149, + 565 + ], + "type": "text", + "content": "2.2. Prompt Tuning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": "Given the limited labeled data during training, prompt tuning is an effective approach to adapt vision-language pre-trained models for few-shot learning tasks [4, 47, 50]. CoOp [56] optimizes prompt vectors in CLIP's language branch for task adaptation but struggles with generalization to novel tasks. To address this, CoCoOp [55] introduces a lightweight meta-network to generate input-conditioned tokens, improving adaptability. Beyond textual prompts, visual prompts have also been explored for task adaptation [11]. Methods like MaPLe [14] and PromptSRC [15] incorporate trainable prompts into both language and visual branches, achieving significant performance gains across" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": "base and novel tasks. Recent studies [21, 31] further construct meta-tasks based on semantic categories to initialize well-structured soft prompts, mitigating generalization decline in high-level tasks. Unlike these approaches that focus on high-level tasks, our proposed prompt tuning method is specifically designed for IQA, leveraging low-level distortion information to construct meta-tasks. This enables pretrained prompts to be distortion-sensitive and effectively capture shared quality knowledge." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 192, + 397, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 397, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 397, + 205 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 213, + 382, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 382, + 224 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 382, + 224 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 230, + 553, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 230, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 230, + 553, + 361 + ], + "type": "text", + "content": "In this paper, we propose the Gradient-Regulated MetaPrompt Image Quality Assessment (GRMP-IQA), which aims to adapt the CLIP for BIQA tasks with a few training samples. As depicted in Fig. 3, GRMP-IQA consists of two primary modules: (i) the Meta-Prompt Pre-training Module (MPP) and (ii) the Quality-Aware Gradient Regularization (QGR). The MPP module pre-trains visual-text prompts to acquire shared meta-knowledge on distortions, enabling quick adaptation to various IQA scenarios. The QGR module plays a key role in fine-tuning by adjusting gradient updates to prevent overfitting to semantic content." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 363, + 553, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 553, + 446 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 553, + 446 + ], + "type": "text", + "content": "Pre-training stage (Sec. 3.3). We randomly sample a minibatch from distortion meta-tasks, partitioning it into a support set " + }, + { + "bbox": [ + 313, + 363, + 553, + 446 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{support}}" + }, + { + "bbox": [ + 313, + 363, + 553, + 446 + ], + "type": "text", + "content": " and a query set " + }, + { + "bbox": [ + 313, + 363, + 553, + 446 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{query}}" + }, + { + "bbox": [ + 313, + 363, + 553, + 446 + ], + "type": "text", + "content": ". A bi-level gradient descent method progresses from the Inner-Loop on " + }, + { + "bbox": [ + 313, + 363, + 553, + 446 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{support}}" + }, + { + "bbox": [ + 313, + 363, + 553, + 446 + ], + "type": "text", + "content": " to the Outer-Loop on " + }, + { + "bbox": [ + 313, + 363, + 553, + 446 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{query}}" + }, + { + "bbox": [ + 313, + 363, + 553, + 446 + ], + "type": "text", + "content": " to optimize the learnable visual-textual prompts " + }, + { + "bbox": [ + 313, + 363, + 553, + 446 + ], + "type": "inline_equation", + "content": "[\\theta_T, \\theta_V]" + }, + { + "bbox": [ + 313, + 363, + 553, + 446 + ], + "type": "text", + "content": ". These prompts are then used as initial weights for following fine-tuning." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 447, + 554, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 554, + 603 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 554, + 603 + ], + "type": "text", + "content": "Fine-tuning stage (Sec. 3.4). Given an input image " + }, + { + "bbox": [ + 313, + 447, + 554, + 603 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 447, + 554, + 603 + ], + "type": "text", + "content": " and a hard semantic prompt " + }, + { + "bbox": [ + 313, + 447, + 554, + 603 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}_i =" + }, + { + "bbox": [ + 313, + 447, + 554, + 603 + ], + "type": "text", + "content": " \"a photo of a [class]\", we predict the semantic class distributions " + }, + { + "bbox": [ + 313, + 447, + 554, + 603 + ], + "type": "inline_equation", + "content": "p^{\\text{sem}}(\\boldsymbol{w}_i|\\boldsymbol{x})" + }, + { + "bbox": [ + 313, + 447, + 554, + 603 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 447, + 554, + 603 + ], + "type": "inline_equation", + "content": "p^{\\text{qua}}(\\boldsymbol{w}_i|\\boldsymbol{x})" + }, + { + "bbox": [ + 313, + 447, + 554, + 603 + ], + "type": "text", + "content": " using the original semantic CLIP model and the fine-tuned CLIP model, respectively. The gradient of the KL divergence loss " + }, + { + "bbox": [ + 313, + 447, + 554, + 603 + ], + "type": "inline_equation", + "content": "\\nabla \\mathcal{L}_{kl}" + }, + { + "bbox": [ + 313, + 447, + 554, + 603 + ], + "type": "text", + "content": " between these distributions is computed as the general semantic direction " + }, + { + "bbox": [ + 313, + 447, + 554, + 603 + ], + "type": "inline_equation", + "content": "G_{\\text{sem}}" + }, + { + "bbox": [ + 313, + 447, + 554, + 603 + ], + "type": "text", + "content": ". Simultaneously, using soft quality prompt " + }, + { + "bbox": [ + 313, + 447, + 554, + 603 + ], + "type": "inline_equation", + "content": "\\theta_T" + }, + { + "bbox": [ + 313, + 447, + 554, + 603 + ], + "type": "text", + "content": " and image " + }, + { + "bbox": [ + 313, + 447, + 554, + 603 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 447, + 554, + 603 + ], + "type": "text", + "content": ", we compute the gradient " + }, + { + "bbox": [ + 313, + 447, + 554, + 603 + ], + "type": "inline_equation", + "content": "\\nabla \\mathcal{L}_{ce}" + }, + { + "bbox": [ + 313, + 447, + 554, + 603 + ], + "type": "text", + "content": " of the quality loss as the quality direction " + }, + { + "bbox": [ + 313, + 447, + 554, + 603 + ], + "type": "inline_equation", + "content": "G_{\\text{qua}}" + }, + { + "bbox": [ + 313, + 447, + 554, + 603 + ], + "type": "text", + "content": ". To refine the optimization direction of the IQA task, we adjust " + }, + { + "bbox": [ + 313, + 447, + 554, + 603 + ], + "type": "inline_equation", + "content": "G_{\\text{qua}}" + }, + { + "bbox": [ + 313, + 447, + 554, + 603 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 313, + 447, + 554, + 603 + ], + "type": "inline_equation", + "content": "G_{\\text{qgr}}" + }, + { + "bbox": [ + 313, + 447, + 554, + 603 + ], + "type": "text", + "content": " by clipping its components aligned with " + }, + { + "bbox": [ + 313, + 447, + 554, + 603 + ], + "type": "inline_equation", + "content": "G_{\\text{sem}}" + }, + { + "bbox": [ + 313, + 447, + 554, + 603 + ], + "type": "text", + "content": ", ensuring the model focuses on image quality while reducing the impact of semantic noise." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 611, + 457, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 457, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 457, + 624 + ], + "type": "text", + "content": "3.2. Visual-Text Meta-Prompt" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 629, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 553, + 714 + ], + "type": "text", + "content": "Visual Meta-Prompt. We utilize Deep Prompt Tuning (DPT) [56] as our Visual Meta-Prompt, with learnable parameters " + }, + { + "bbox": [ + 313, + 629, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathbf{V}}" + }, + { + "bbox": [ + 313, + 629, + 553, + 714 + ], + "type": "text", + "content": ". The input embedding for the " + }, + { + "bbox": [ + 313, + 629, + 553, + 714 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 629, + 553, + 714 + ], + "type": "text", + "content": "-th layer's self-attention module in the ViT-based image encoder is denoted as " + }, + { + "bbox": [ + 313, + 629, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\{\\pmb{f}^l,\\pmb{H}^l\\}" + }, + { + "bbox": [ + 313, + 629, + 553, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 629, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{f}^l" + }, + { + "bbox": [ + 313, + 629, + 553, + 714 + ], + "type": "text", + "content": " represents the classification (CLS) token, and " + }, + { + "bbox": [ + 313, + 629, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{H}^{l} = \\{\\pmb{h}_{1}^{l},\\pmb{h}_{2}^{l},\\dots,\\pmb{h}_{N}^{l}\\}" + }, + { + "bbox": [ + 313, + 629, + 553, + 714 + ], + "type": "text", + "content": " denotes the image patch embeddings. A learnable token " + }, + { + "bbox": [ + 313, + 629, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{P}^{l}" + }, + { + "bbox": [ + 313, + 629, + 553, + 714 + ], + "type": "text", + "content": " is appended to" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10444" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 68, + 294, + 227 + ], + "blocks": [ + { + "bbox": [ + 57, + 68, + 294, + 227 + ], + "lines": [ + { + "bbox": [ + 57, + 68, + 294, + 227 + ], + "spans": [ + { + "bbox": [ + 57, + 68, + 294, + 227 + ], + "type": "image", + "image_path": "5dc812a037b8edb3e8e9bdb209255d9b3718d57b27accc4922a3496644d12a69.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 229, + 276, + 241 + ], + "lines": [ + { + "bbox": [ + 69, + 229, + 276, + 241 + ], + "spans": [ + { + "bbox": [ + 69, + 229, + 276, + 241 + ], + "type": "text", + "content": "(a) Pre-training stage: Meta-Prompt Pre-training Module" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 301, + 70, + 553, + 227 + ], + "blocks": [ + { + "bbox": [ + 301, + 70, + 553, + 227 + ], + "lines": [ + { + "bbox": [ + 301, + 70, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 301, + 70, + 553, + 227 + ], + "type": "image", + "image_path": "760739a2d0f2a2c15efe37127809c8161409d5122e0d73af86cda2fd48426105.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 229, + 544, + 241 + ], + "lines": [ + { + "bbox": [ + 320, + 229, + 544, + 241 + ], + "spans": [ + { + "bbox": [ + 320, + 229, + 544, + 241 + ], + "type": "text", + "content": "(b) Fine-tuning stage: Quality-Aware Gradient Regularization" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 249, + 555, + 327 + ], + "lines": [ + { + "bbox": [ + 55, + 249, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 249, + 555, + 327 + ], + "type": "text", + "content": "Figure 3. The overview of our GRMP-IQA. The core modules are the Meta-Prompt Pre-training and Quality-Aware Gradient Regularization, corresponding to two training processes. At the pre-training stage, we establish a distortion-specific meta-knowledge task for BIQA tasks, and the bi-level gradient descent is utilized to train visual-text meta-prompt " + }, + { + "bbox": [ + 55, + 249, + 555, + 327 + ], + "type": "inline_equation", + "content": "[\\theta_T,\\theta_V]" + }, + { + "bbox": [ + 55, + 249, + 555, + 327 + ], + "type": "text", + "content": " (Sec. 3.3). These optimized prompts then serve as initial settings to efficiently adapt the CLIP model to IQA tasks with limited labels. At the few-shot fine-tuning stage, we first predict the class probability distributions " + }, + { + "bbox": [ + 55, + 249, + 555, + 327 + ], + "type": "inline_equation", + "content": "p^{sem}(w_i|x)" + }, + { + "bbox": [ + 55, + 249, + 555, + 327 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 249, + 555, + 327 + ], + "type": "inline_equation", + "content": "p^{qua}(w_i|x)" + }, + { + "bbox": [ + 55, + 249, + 555, + 327 + ], + "type": "text", + "content": " using the semantic and fine-tuned CLIP models, respectively. We then adjust IQA task gradients " + }, + { + "bbox": [ + 55, + 249, + 555, + 327 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{qua}}" + }, + { + "bbox": [ + 55, + 249, + 555, + 327 + ], + "type": "text", + "content": " by clipping gradients aligned with semantic task gradients " + }, + { + "bbox": [ + 55, + 249, + 555, + 327 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{sem}}" + }, + { + "bbox": [ + 55, + 249, + 555, + 327 + ], + "type": "text", + "content": " to generate refined gradient " + }, + { + "bbox": [ + 55, + 249, + 555, + 327 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{qgr}}" + }, + { + "bbox": [ + 55, + 249, + 555, + 327 + ], + "type": "text", + "content": " for backward updates (Sec. 3.4), mitigating the influence of semantic noise on quality predictions." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 338, + 295, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 295, + 361 + ], + "type": "text", + "content": "the token sequence in each ViT layer, and the Multi-Head Self-Attention (MHSA) module processes the tokens as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 88, + 367, + 295, + 382 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 367, + 295, + 382 + ], + "spans": [ + { + "bbox": [ + 88, + 367, + 295, + 382 + ], + "type": "interline_equation", + "content": "[ \\boldsymbol {f} ^ {l}, \\neg , \\boldsymbol {H} ^ {l} ] = \\operatorname {L a y e r} ^ {l} ([ \\boldsymbol {f} ^ {l - 1}, \\boldsymbol {P} ^ {l - 1}, \\boldsymbol {H} ^ {l - 1} ]), \\tag {1}", + "image_path": "1f489f28052f112b32da9bbfb1d1208d92c1d1776ebfa8cae321feedfa7c170f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 388, + 295, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 295, + 412 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 295, + 412 + ], + "type": "text", + "content": "where the output of " + }, + { + "bbox": [ + 55, + 388, + 295, + 412 + ], + "type": "inline_equation", + "content": "P^l" + }, + { + "bbox": [ + 55, + 388, + 295, + 412 + ], + "type": "text", + "content": " is discarded and not passed to the next layer, serving only as a set of learnable parameters." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 412, + 295, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 412, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 295, + 496 + ], + "type": "text", + "content": "Text Meta-Prompt. We adopt CoOp [56] as our textual meta-prompt, with learnable parameters " + }, + { + "bbox": [ + 55, + 412, + 295, + 496 + ], + "type": "inline_equation", + "content": "\\theta_T" + }, + { + "bbox": [ + 55, + 412, + 295, + 496 + ], + "type": "text", + "content": ". For each quality class " + }, + { + "bbox": [ + 55, + 412, + 295, + 496 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 412, + 295, + 496 + ], + "type": "text", + "content": ", we construct a learnable quality prompt defined as " + }, + { + "bbox": [ + 55, + 412, + 295, + 496 + ], + "type": "inline_equation", + "content": "\\pmb{q}_i = \\{\\pmb{v}_1, \\pmb{v}_2, \\dots, \\pmb{v}_M, [\\mathrm{quality}]\\}" + }, + { + "bbox": [ + 55, + 412, + 295, + 496 + ], + "type": "text", + "content": ". Here, [quality] " + }, + { + "bbox": [ + 55, + 412, + 295, + 496 + ], + "type": "inline_equation", + "content": "\\in" + }, + { + "bbox": [ + 55, + 412, + 295, + 496 + ], + "type": "text", + "content": " {\"high quality\", \"low quality\"} serves as a categorical quality marker, and " + }, + { + "bbox": [ + 55, + 412, + 295, + 496 + ], + "type": "inline_equation", + "content": "\\{\\pmb{v}_m\\}_{m=1}^M" + }, + { + "bbox": [ + 55, + 412, + 295, + 496 + ], + "type": "text", + "content": " denotes a set of " + }, + { + "bbox": [ + 55, + 412, + 295, + 496 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 412, + 295, + 496 + ], + "type": "text", + "content": " learnable context vectors (with " + }, + { + "bbox": [ + 55, + 412, + 295, + 496 + ], + "type": "inline_equation", + "content": "M = 4" + }, + { + "bbox": [ + 55, + 412, + 295, + 496 + ], + "type": "text", + "content": " in our work)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 502, + 241, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 241, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 241, + 514 + ], + "type": "text", + "content": "3.3. Meta-Prompt Pre-training Module" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 520, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 295, + 581 + ], + "type": "text", + "content": "During meta-learning, we aim to optimize the learnable visual-text prompt " + }, + { + "bbox": [ + 55, + 520, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\theta = [\\theta_T,\\theta_V]" + }, + { + "bbox": [ + 55, + 520, + 295, + 581 + ], + "type": "text", + "content": ". For an input image " + }, + { + "bbox": [ + 55, + 520, + 295, + 581 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 520, + 295, + 581 + ], + "type": "text", + "content": ", the textual encoder " + }, + { + "bbox": [ + 55, + 520, + 295, + 581 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 55, + 520, + 295, + 581 + ], + "type": "text", + "content": " processes the quality prompts " + }, + { + "bbox": [ + 55, + 520, + 295, + 581 + ], + "type": "inline_equation", + "content": "q_{i}" + }, + { + "bbox": [ + 55, + 520, + 295, + 581 + ], + "type": "text", + "content": " while the visual encoder extracts the feature vector " + }, + { + "bbox": [ + 55, + 520, + 295, + 581 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 520, + 295, + 581 + ], + "type": "text", + "content": ". The probability of predicting high quality is:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 586, + 298, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 298, + 622 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 298, + 622 + ], + "type": "interline_equation", + "content": "p \\left(\\boldsymbol {q} _ {\\text {h i g h}} \\mid \\boldsymbol {x}\\right) = \\frac {\\exp \\left(\\left\\langle \\boldsymbol {g} \\left(\\boldsymbol {q} _ {\\text {h i g h}}\\right) , \\boldsymbol {f} \\right\\rangle / \\tau\\right)}{\\exp \\left(\\left\\langle \\boldsymbol {g} \\left(\\boldsymbol {q} _ {\\text {h i g h}}\\right) , \\boldsymbol {f} \\right\\rangle / \\tau\\right) + \\exp \\left(\\left\\langle \\boldsymbol {g} \\left(\\boldsymbol {q} _ {\\text {l o w}}\\right) , \\boldsymbol {f} \\right\\rangle / \\tau\\right)},", + "image_path": "cd962850931676d85b7e17a415826e95ebd1b1082a24512dadcc9130abcaf05d.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 623, + 296, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 296, + 694 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 623, + 296, + 694 + ], + "type": "inline_equation", + "content": "q_{\\mathrm{high}}" + }, + { + "bbox": [ + 55, + 623, + 296, + 694 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 623, + 296, + 694 + ], + "type": "inline_equation", + "content": "q_{\\mathrm{low}}" + }, + { + "bbox": [ + 55, + 623, + 296, + 694 + ], + "type": "text", + "content": " denote high quality and low quality category prompt, " + }, + { + "bbox": [ + 55, + 623, + 296, + 694 + ], + "type": "inline_equation", + "content": "p(q_{\\mathrm{high}}|\\boldsymbol {x})" + }, + { + "bbox": [ + 55, + 623, + 296, + 694 + ], + "type": "text", + "content": " represents the estimated probability that the quality of image " + }, + { + "bbox": [ + 55, + 623, + 296, + 694 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 623, + 296, + 694 + ], + "type": "text", + "content": " is \"high quality\", " + }, + { + "bbox": [ + 55, + 623, + 296, + 694 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 623, + 296, + 694 + ], + "type": "text", + "content": " is a temperature parameter learned by CLIP, and " + }, + { + "bbox": [ + 55, + 623, + 296, + 694 + ], + "type": "inline_equation", + "content": "\\langle ,\\rangle" + }, + { + "bbox": [ + 55, + 623, + 296, + 694 + ], + "type": "text", + "content": " denotes cosine similarity. The labeled quality scores are then rescaled to 0-1, denoted as " + }, + { + "bbox": [ + 55, + 623, + 296, + 694 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 623, + 296, + 694 + ], + "type": "text", + "content": ", and the loss function is calculated as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 61, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 61, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = - (y \\log (p \\left(q _ {\\text {h i g h}} | \\boldsymbol {x}\\right)) + (1 - y) \\log (1 - p \\left(q _ {\\text {h i g h}} | \\boldsymbol {x}\\right)). \\tag {3}", + "image_path": "90f8b67f2cd09c6b59d3f6a136e7cee705ce780caa9339ccea35e33faa58a5f0.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 338, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 555, + 628 + ], + "type": "text", + "content": "Constructing Distortion Meta-Knowledge Task. As noted by [29, 58], the ability to detect various types of image distortions is crucial for developing BIQA models with strong generalization capabilities across diverse scenarios. Moreover, the efficacy of prompt tuning is significantly dependent on the initial configuration of the prompts. This initial setup greatly affects the CLIP vision-language model's ability to swiftly adapt to different IQA scenarios. Drawing inspiration from the \"learn to learn\" ethos inherent in the deep meta-learning paradigm [43, 58], we propose an optimization-based method for effectively pre-training visual-textual prompts. These prompts incorporate shared quality insights from various image distortions, thereby enhancing CLIP's swift adaptability to IQA tasks [58]. To investigate the general rules of image distortion, we first constructed a " + }, + { + "bbox": [ + 313, + 338, + 555, + 628 + ], + "type": "inline_equation", + "content": "K_{t}" + }, + { + "bbox": [ + 313, + 338, + 555, + 628 + ], + "type": "text", + "content": "-way distortion-specific image quality prediction task, denoted as " + }, + { + "bbox": [ + 313, + 338, + 555, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_t" + }, + { + "bbox": [ + 313, + 338, + 555, + 628 + ], + "type": "text", + "content": ". Then, it is used to build the meta-training set as " + }, + { + "bbox": [ + 313, + 338, + 555, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{meta}} = \\left\\{\\mathcal{D}_{\\mathcal{T}_t}^{\\mathrm{support}}, \\mathcal{D}_{\\mathcal{T}_t}^{\\mathrm{query}}\\right\\}_{t=1}^T" + }, + { + "bbox": [ + 313, + 338, + 555, + 628 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 313, + 338, + 555, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathcal{T}_t}^{\\mathrm{support}}" + }, + { + "bbox": [ + 313, + 338, + 555, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 338, + 555, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathcal{T}_t}^{\\mathrm{query}}" + }, + { + "bbox": [ + 313, + 338, + 555, + 628 + ], + "type": "text", + "content": " represent the support and query sets for each task, respectively, with " + }, + { + "bbox": [ + 313, + 338, + 555, + 628 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 338, + 555, + 628 + ], + "type": "text", + "content": " representing the total number of tasks. To simulate the process of prompt generalization to different distortions in BIQA scenarios, we randomly sample " + }, + { + "bbox": [ + 313, + 338, + 555, + 628 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 338, + 555, + 628 + ], + "type": "text", + "content": " tasks as a mini-batch from the meta-training set, where " + }, + { + "bbox": [ + 313, + 338, + 555, + 628 + ], + "type": "inline_equation", + "content": "1 \\leq k \\leq T" + }, + { + "bbox": [ + 313, + 338, + 555, + 628 + ], + "type": "text", + "content": ", to perform bi-level gradient optimization." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "content": "Distortion-Aware Meta-prompt Learning. Our approach employs a bi-level gradient descent technique to bridge the learning process from the support to the query set. Specifically, it mainly consists of two optimization steps. In the Inner-Loop (the first level), we compute the gradients of the prompt parameters using the support set and apply the first update. In the Outer-Loop (the second level), we assess the" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10445" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 132 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 132 + ], + "type": "text", + "content": "performance of the updated model on the query set and optimize the parameters again. This bi-level structure distills shared quality priors by training on a number of NR-IQA tasks with known distortions, enabling the meta-prompt to generalize rapidly across diverse new BIQA scenarios." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 294, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 294, + 207 + ], + "type": "text", + "content": "Inner-Loop. The objective of the Inner-Loop stage is to adapt the meta-prompt, denoted by " + }, + { + "bbox": [ + 55, + 133, + 294, + 207 + ], + "type": "inline_equation", + "content": "[\\pmb{\\theta}_{\\pmb{T}},\\pmb{\\theta}_{\\pmb{V}}]" + }, + { + "bbox": [ + 55, + 133, + 294, + 207 + ], + "type": "text", + "content": ", to the " + }, + { + "bbox": [ + 55, + 133, + 294, + 207 + ], + "type": "inline_equation", + "content": "t^{th}" + }, + { + "bbox": [ + 55, + 133, + 294, + 207 + ], + "type": "text", + "content": " support set " + }, + { + "bbox": [ + 55, + 133, + 294, + 207 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathcal{T}_t}^{\\mathrm{support}}" + }, + { + "bbox": [ + 55, + 133, + 294, + 207 + ], + "type": "text", + "content": " within the mini-batch. During the first level of updates, we determine the loss " + }, + { + "bbox": [ + 55, + 133, + 294, + 207 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\pmb {\\theta},\\mathcal{D}_{\\mathcal{T}_t}^{\\mathrm{support}})" + }, + { + "bbox": [ + 55, + 133, + 294, + 207 + ], + "type": "text", + "content": ", following which the model parameters are updated on the support set using the inner learning rate " + }, + { + "bbox": [ + 55, + 133, + 294, + 207 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 133, + 294, + 207 + ], + "type": "text", + "content": ", as specified by:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 216, + 295, + 232 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 216, + 295, + 232 + ], + "spans": [ + { + "bbox": [ + 115, + 216, + 295, + 232 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\theta} _ {t} ^ {\\prime} = \\boldsymbol {\\theta} - \\alpha \\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} (\\boldsymbol {\\theta}, \\mathcal {D} _ {\\mathcal {T} _ {t}} ^ {\\text {s u p p o r t}}), \\tag {4}", + "image_path": "b50510989abc00dc807297b844973d8034e667a9ac7df3b7fafe1b44cc90774f.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 240, + 294, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 240, + 294, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 294, + 266 + ], + "type": "text", + "content": "Outer-Loop. In a similar vein, the second level of updates adjusts the parameters " + }, + { + "bbox": [ + 55, + 240, + 294, + 266 + ], + "type": "inline_equation", + "content": "\\theta_t^\\prime" + }, + { + "bbox": [ + 55, + 240, + 294, + 266 + ], + "type": "text", + "content": " based on the query set " + }, + { + "bbox": [ + 55, + 240, + 294, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathcal{T}_t}^{\\mathrm{query}}" + }, + { + "bbox": [ + 55, + 240, + 294, + 266 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 111, + 275, + 295, + 290 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 275, + 295, + 290 + ], + "spans": [ + { + "bbox": [ + 111, + 275, + 295, + 290 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\theta} _ {t} \\leftarrow \\boldsymbol {\\theta} _ {t} ^ {\\prime} - \\alpha \\nabla_ {\\boldsymbol {\\theta} _ {t} ^ {\\prime}} \\mathcal {L} \\left(\\boldsymbol {\\theta} _ {t} ^ {\\prime}, \\mathcal {D} _ {\\mathcal {T} _ {t}} ^ {\\text {q u e r y}}\\right). \\tag {5}", + "image_path": "0dbe0907d1a8dd862929d9305575ecd4f722c751ac2aabe02eeaac95c041faed.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 299, + 295, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 299, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 299, + 295, + 335 + ], + "type": "text", + "content": "For a mini-batch of meta-tasks, this process culminates in aggregating gradients from all tasks to update the final model parameters, following the update rule:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 346, + 295, + 377 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 346, + 295, + 377 + ], + "spans": [ + { + "bbox": [ + 107, + 346, + 295, + 377 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\theta} \\leftarrow \\boldsymbol {\\theta} - \\beta \\sum_ {t = 1} ^ {k} \\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} \\left(\\boldsymbol {\\theta} _ {t} ^ {\\prime}, \\mathcal {D} _ {\\mathcal {T} _ {t}} ^ {\\text {q u e r y}}\\right), \\tag {6}", + "image_path": "4e07f7d427e98d20dd05b9aa46622b0459e3ed8c44612cc50aeec132c92d63d1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 387, + 295, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 387, + 295, + 423 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 295, + 423 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 387, + 295, + 423 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 387, + 295, + 423 + ], + "type": "text", + "content": " represents the outer learning rate. Meta-learning effectively trains the learnable prompts " + }, + { + "bbox": [ + 55, + 387, + 295, + 423 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta} = [\\pmb{\\theta}_T, \\pmb{\\theta}_V]" + }, + { + "bbox": [ + 55, + 387, + 295, + 423 + ], + "type": "text", + "content": ", ensuring their generalization across various image distortions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 432, + 264, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 432, + 264, + 444 + ], + "spans": [ + { + "bbox": [ + 55, + 432, + 264, + 444 + ], + "type": "text", + "content": "3.4. Quality-Aware Gradient Regularization" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 449, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 593 + ], + "type": "text", + "content": "We denote the original semantic CLIP model as " + }, + { + "bbox": [ + 55, + 449, + 295, + 593 + ], + "type": "inline_equation", + "content": "V^{\\mathrm{sem}}" + }, + { + "bbox": [ + 55, + 449, + 295, + 593 + ], + "type": "text", + "content": " and the model obtained from the Meta-Prompt Pre-training Module, as " + }, + { + "bbox": [ + 55, + 449, + 295, + 593 + ], + "type": "inline_equation", + "content": "V^{\\mathrm{qua}}" + }, + { + "bbox": [ + 55, + 449, + 295, + 593 + ], + "type": "text", + "content": ". The gradient for the IQA task is the quality gradient " + }, + { + "bbox": [ + 55, + 449, + 295, + 593 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{qua}}" + }, + { + "bbox": [ + 55, + 449, + 295, + 593 + ], + "type": "text", + "content": ", while the gradient for the semantic direction is the semantic gradient " + }, + { + "bbox": [ + 55, + 449, + 295, + 593 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{sem}}" + }, + { + "bbox": [ + 55, + 449, + 295, + 593 + ], + "type": "text", + "content": ". Recent studies [1, 53] have shown that IQA tasks can be misaligned with the high-level semantic representations of upstream tasks, leading to overfitting and reduced generalization. To prevent semantic overfitting, our Quality-aware Gradient Regularization (QGR) follows three key steps: derive a semantic direction " + }, + { + "bbox": [ + 55, + 449, + 295, + 593 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{sem}}" + }, + { + "bbox": [ + 55, + 449, + 295, + 593 + ], + "type": "text", + "content": ", compute the quality optimization direction " + }, + { + "bbox": [ + 55, + 449, + 295, + 593 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{qua}}" + }, + { + "bbox": [ + 55, + 449, + 295, + 593 + ], + "type": "text", + "content": ", and balance gradients based on their directional relationship." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "Specifically, to derive a general semantic direction " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{sem}}" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": ", we first design a hard prompt following [49, 52]: " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}_i =" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " \"a photo of a [class]\", where [class] represents one of nine categories: \"animal\", \"cityscape\", \"human\", \"indoor\", \"landscape\", \"night\", \"plant\", \"still-life\", \"others\". This prompt " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}_i" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " is input into the text encoders of " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "V^{\\mathrm{sem}}" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "V^{\\mathrm{qua}}" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": ", generating text features that align with image features " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "f^{qua}" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " from their respective visual encoders. As a result, we obtain zero-shot semantic prediction probabilities " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "p^{\\mathrm{sem}}(\\boldsymbol{w}_i|\\boldsymbol{x})" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "p^{\\mathrm{qua}}(\\boldsymbol{w}_i|\\boldsymbol{x})" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": ". To measure the semantic alignment between" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 72, + 510, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 510, + 84 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 510, + 84 + ], + "type": "inline_equation", + "content": "V^{\\mathrm{qua}}" + }, + { + "bbox": [ + 315, + 72, + 510, + 84 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 315, + 72, + 510, + 84 + ], + "type": "inline_equation", + "content": "V^{\\mathrm{sem}}" + }, + { + "bbox": [ + 315, + 72, + 510, + 84 + ], + "type": "text", + "content": ", we compute the KL divergence:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 330, + 91, + 553, + 119 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 91, + 553, + 119 + ], + "spans": [ + { + "bbox": [ + 330, + 91, + 553, + 119 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {k l}} \\left(\\boldsymbol {V} ^ {\\text {q u a}}\\right) = - \\sum_ {i} \\boldsymbol {p} ^ {\\text {s e m}} \\left(\\boldsymbol {w} _ {i} \\mid \\boldsymbol {x}\\right) \\log \\frac {\\boldsymbol {p} ^ {\\text {q u a}} \\left(\\boldsymbol {w} _ {i} \\mid \\boldsymbol {x}\\right)}{\\boldsymbol {p} ^ {\\text {s e m}} \\left(\\boldsymbol {w} _ {i} \\mid \\boldsymbol {x}\\right)}. \\tag {7}", + "image_path": "9221a9ae04a124628a32a34245530a85af407cc897a2a10fb113190346a47345.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 128, + 553, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 128, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 313, + 128, + 553, + 306 + ], + "type": "text", + "content": "The gradient of " + }, + { + "bbox": [ + 313, + 128, + 553, + 306 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{kl}}(V^{\\mathrm{qua}})" + }, + { + "bbox": [ + 313, + 128, + 553, + 306 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 313, + 128, + 553, + 306 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{sem}}" + }, + { + "bbox": [ + 313, + 128, + 553, + 306 + ], + "type": "text", + "content": ", represents the general semantic optimization direction. Similarly, the quality direction " + }, + { + "bbox": [ + 313, + 128, + 553, + 306 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{qua}}" + }, + { + "bbox": [ + 313, + 128, + 553, + 306 + ], + "type": "text", + "content": " is obtained by computing the gradient of the cross-entropy loss " + }, + { + "bbox": [ + 313, + 128, + 553, + 306 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ce}}(V^{\\mathrm{qua}})" + }, + { + "bbox": [ + 313, + 128, + 553, + 306 + ], + "type": "text", + "content": ", which compares the predicted probability " + }, + { + "bbox": [ + 313, + 128, + 553, + 306 + ], + "type": "inline_equation", + "content": "p(q_{\\mathrm{high}}|x)" + }, + { + "bbox": [ + 313, + 128, + 553, + 306 + ], + "type": "text", + "content": " with the ground truth " + }, + { + "bbox": [ + 313, + 128, + 553, + 306 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 128, + 553, + 306 + ], + "type": "text", + "content": ", as shown in Eq. 3. Finally, to balance the gradients between quality " + }, + { + "bbox": [ + 313, + 128, + 553, + 306 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{qua}}" + }, + { + "bbox": [ + 313, + 128, + 553, + 306 + ], + "type": "text", + "content": " and semantics " + }, + { + "bbox": [ + 313, + 128, + 553, + 306 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{sem}}" + }, + { + "bbox": [ + 313, + 128, + 553, + 306 + ], + "type": "text", + "content": " within a shared representation space, we analyze their relationship under two cases: (1) If the angle between " + }, + { + "bbox": [ + 313, + 128, + 553, + 306 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{qua}}" + }, + { + "bbox": [ + 313, + 128, + 553, + 306 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 128, + 553, + 306 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{sem}}" + }, + { + "bbox": [ + 313, + 128, + 553, + 306 + ], + "type": "text", + "content": " is less than 90 degrees, it indicates that the optimization directions for IQA quality knowledge and general semantic knowledge are consistent. In such cases, we clip the " + }, + { + "bbox": [ + 313, + 128, + 553, + 306 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{qua}}" + }, + { + "bbox": [ + 313, + 128, + 553, + 306 + ], + "type": "text", + "content": " along its component " + }, + { + "bbox": [ + 313, + 128, + 553, + 306 + ], + "type": "inline_equation", + "content": "G_{\\parallel}" + }, + { + "bbox": [ + 313, + 128, + 553, + 306 + ], + "type": "text", + "content": " that parallel to the semantic direction " + }, + { + "bbox": [ + 313, + 128, + 553, + 306 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{sem}}" + }, + { + "bbox": [ + 313, + 128, + 553, + 306 + ], + "type": "text", + "content": ", to modulate the model's original quality optimization path, preventing overfitting to semantic correlations." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 307, + 553, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 553, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 553, + 390 + ], + "type": "text", + "content": "(2) Conversely, if the angle between two areas of knowledge is more than 90 degrees, it means they are heading in different directions during improvement. In such situations, we don't adjust the quality gradient " + }, + { + "bbox": [ + 313, + 307, + 553, + 390 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{qua}}" + }, + { + "bbox": [ + 313, + 307, + 553, + 390 + ], + "type": "text", + "content": ", allowing the model to learn quality-aware features according to the original optimization direction. In conclusion, our gradient regularization strategy is mathematically formulated as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 398, + 553, + 440 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 398, + 553, + 440 + ], + "spans": [ + { + "bbox": [ + 321, + 398, + 553, + 440 + ], + "type": "interline_equation", + "content": "\\boldsymbol {G} _ {\\text {q g r}} = \\left\\{ \\begin{array}{l l} \\boldsymbol {G} _ {\\text {q u a}} & \\text {i f} \\boldsymbol {G} _ {\\text {q u a}} \\cdot \\boldsymbol {G} _ {\\text {s e m}} \\leq 0, \\\\ \\boldsymbol {G} _ {\\text {q u a}} - \\lambda \\frac {\\boldsymbol {G} _ {\\text {q u a}} \\cdot \\boldsymbol {G} _ {\\text {s e m}}}{\\| \\boldsymbol {G} _ {\\text {s e m}} \\| ^ {2}} \\boldsymbol {G} _ {\\text {s e m}} & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {8}", + "image_path": "2ef682fa96b5fc59dc28cdfc6ca60f64c95b00595d7a7201759910a4b3e1b423.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 440, + 553, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 440, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 440, + 553, + 488 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 440, + 553, + 488 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 440, + 553, + 488 + ], + "type": "text", + "content": " is introduced to generalize the formulation, providing flexibility in controlling the influence of general knowledge. Specifically, " + }, + { + "bbox": [ + 313, + 440, + 553, + 488 + ], + "type": "inline_equation", + "content": "\\lambda = 1" + }, + { + "bbox": [ + 313, + 440, + 553, + 488 + ], + "type": "text", + "content": " projects " + }, + { + "bbox": [ + 313, + 440, + 553, + 488 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{qua}}" + }, + { + "bbox": [ + 313, + 440, + 553, + 488 + ], + "type": "text", + "content": " onto the orthogonal direction of " + }, + { + "bbox": [ + 313, + 440, + 553, + 488 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{sem}}" + }, + { + "bbox": [ + 313, + 440, + 553, + 488 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 313, + 440, + 553, + 488 + ], + "type": "inline_equation", + "content": "\\lambda = 0" + }, + { + "bbox": [ + 313, + 440, + 553, + 488 + ], + "type": "text", + "content": " reduces QGR to CoOp [56]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 495, + 394, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 394, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 394, + 508 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 514, + 497, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 497, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 497, + 526 + ], + "type": "text", + "content": "4.1. Datasets and Evaluation Protocols" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 532, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 555, + 712 + ], + "type": "text", + "content": "We conduct experiments on multiple BIQA datasets include LIVEC [7], KonIQ [10], LIVEFB [49], and SPAQ [3], which feature authentic distortions, and PIPAL [12], LIVE [38], and CSIQ [17], which feature synthetic distortions. LIVEC contains 1,162 mobile device images, SPAQ includes 11,125 photos from 66 smartphones, KonIQ has 10,073 images from open sources, and LIVEFB is the largest real-world dataset with 39,810 images. For synthetic distortions, LIVE and CSIQ contain 779 and 866 images with 5 and 6 types of distortions. PIPAL, a challenging dataset, includes 23,200 images with 40 types of distortions, including GAN-generated artifacts. We use Spearman's Rank Correlation Coefficient (SRCC) and the Pearson Linear Correlation Coefficient (PLCC) as metrics to quantify the monotonousness and accuracy of predictions." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10446" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 555, + 225 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 555, + 225 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 555, + 225 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 555, + 225 + ], + "type": "table", + "html": "
MethodLIVECKonIQCSIQLIVEPIPAL
Labels5010020050100200501002005010020050100200
HyperIQA [41]0.6480.7250.7900.6150.7100.7760.7900.8240.9090.8920.9120.9290.1020.3020.379
MetaIQA [58]0.6040.6260.6690.6180.6200.6600.7840.8490.8940.8400.8800.9190.3320.3480.371
DEIQT [33]0.6670.7180.8120.6380.6820.7540.8210.8910.9410.9200.9420.9550.3960.4100.436
MANIQA [48]0.6420.7690.7970.6520.7550.8100.7940.8470.8740.9090.9280.9570.1360.3610.470
CONTRIQUE [30]0.6950.7290.7610.7330.7940.8210.8400.9260.9400.8910.9220.9430.3790.4370.488
Re-IQA [37]0.5910.6210.7010.6850.7230.7540.8930.9070.9230.8840.8940.9290.2800.3500.431
CLIP w/ Linear Probe [36]0.6640.7210.7330.7360.7700.7820.8410.8920.9410.8960.9230.9410.2540.3030.368
CLIP w/ CoOp [44]†0.6950.7380.7460.6920.7430.762---------
LIQE [52]0.6910.7690.8100.7590.8010.8320.8380.8910.9240.9040.9340.948---
GRepQ [40]0.7600.7910.8220.8120.8360.8550.8780.9140.9410.9260.9370.9530.390†0.450†0.498†
GRMP-IQA (Ours)0.8360.8570.8750.8530.8720.8830.8930.9170.9410.9320.9430.9680.4740.5120.546
", + "image_path": "a60ac20e8b5aa59deb02574b0a961caf599086491b68eaeecb70fdae580c8852.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 261, + 555, + 475 + ], + "blocks": [ + { + "bbox": [ + 55, + 234, + 555, + 257 + ], + "lines": [ + { + "bbox": [ + 55, + 234, + 555, + 257 + ], + "spans": [ + { + "bbox": [ + 55, + 234, + 555, + 257 + ], + "type": "text", + "content": "Table 1. SRCC performance comparison of our method with other IQA methods trained on limited labels. Bold indicates the best results, underlined marks the second-best, and the fifth-to-last through second-to-last lines show the CLIP-based IQA. † is our reproduction." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 261, + 555, + 475 + ], + "lines": [ + { + "bbox": [ + 58, + 261, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 58, + 261, + 555, + 475 + ], + "type": "table", + "html": "
MethodLIVECSIQLIVECKonIQLIVEFBSPAQ
PLCCSRCCPLCCSRCCPLCCSRCCPLCCSRCCPLCCSRCCPLCCSRCC
Training Ratio20%20%20%20%20%20%20%20%10%10%20%20%
DEIQT [33]0.9680.9650.8850.8620.8220.7920.9080.8880.6240.5380.9120.908
LoDa [46]----0.8540.8150.9230.907----
GRMP-IQA (Ours)0.9720.9700.9580.9510.8970.8750.9310.9150.6860.6040.9250.920
Training Ratio80%80%80%80%80%80%80%80%80%80%80%80%
MetaIQA [58]0.9590.9600.9080.8990.8020.8350.8560.8870.5070.540--
CONTRIQUE [30]0.9610.9600.9550.9420.8570.8450.9060.8940.6410.5800.9190.914
DEIQT [33]0.9820.9800.9630.9460.8940.8750.9340.9210.6630.5710.9230.919
Re-IQA [37]0.9710.9700.9600.9470.8540.8400.9230.9140.7330.6450.9250.918
LIQE [52]0.9510.9700.9390.9360.9100.9040.9080.919----
CLIP-IQA+ [44]----0.8320.8050.9090.8950.5930.5750.8660.864
CDINet [54]0.9750.9770.9600.9520.8800.8650.9280.916--0.9220.919
QFM-IQM [22]0.9830.9810.9650.9540.9130.8910.9360.9220.6670.5670.9240.920
LoDa [46]0.9790.975--0.8990.8760.9440.9320.6790.5780.9280.925
GRMP-IQA (Ours)0.9830.9810.9740.9680.9160.8970.9450.9340.7040.6160.9320.927
", + "image_path": "03fdad7c0b62eb813b852f46a6ab86aa6a42d7039e7aa640c0e2f337a4c11a1a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 82, + 483, + 526, + 495 + ], + "lines": [ + { + "bbox": [ + 82, + 483, + 526, + 495 + ], + "spans": [ + { + "bbox": [ + 82, + 483, + 526, + 495 + ], + "type": "text", + "content": "Table 2. Performance comparison measured by medians of SRCC and PLCC, and bold entries indicate the top two results." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 501, + 243, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 501, + 243, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 501, + 243, + 514 + ], + "type": "text", + "content": "4.2. Implementation Details and Setups" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 519, + 295, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 519, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 54, + 519, + 295, + 639 + ], + "type": "text", + "content": "We build our model on CLIP-B/16 [36]. During pretraining, we optimize the visual-text prompt, while in finetuning, only the last four blocks of the image and text encoders are trained. For competing models, we use public results or re-train them under our setup. Each dataset is split into " + }, + { + "bbox": [ + 54, + 519, + 295, + 639 + ], + "type": "inline_equation", + "content": "80\\% / 20\\%" + }, + { + "bbox": [ + 54, + 519, + 295, + 639 + ], + "type": "text", + "content": " for training and testing, with splits based on reference images to ensure content independence in datasets with synthetic distortions. To ensure robustness, we report the median performance across ten random splits. All experiments are run on four NVIDIA RTX3090 GPUs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 640, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 640, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 640, + 295, + 688 + ], + "type": "text", + "content": "Meta-Prompt Pre-training. We pre-train on TID2013 [32] and KADID-10K [25] dataset, which contain extensive distortion information. We set the learning rates " + }, + { + "bbox": [ + 55, + 640, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 640, + 295, + 688 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 640, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 640, + 295, + 688 + ], + "type": "text", + "content": " to 1e-4 and 1e-2, and train for 50 epochs using Adam [16]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Fully Supervised Learning Setting. We randomly crop each input image into 10 patches of " + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": " resolution" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 502, + 555, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 555, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 555, + 612 + ], + "type": "text", + "content": "and train the model for 9 epochs using AdamW [26]. The learning rate is " + }, + { + "bbox": [ + 313, + 502, + 555, + 612 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-6}" + }, + { + "bbox": [ + 313, + 502, + 555, + 612 + ], + "type": "text", + "content": ", with a scheduler over 9 decay epochs. The batch size is 16 for LIVEC and 128 for KonIQ. Few-Shot Learning Setting. In the few-shot setting, we follow the approach in a previous study [40] to train our GRMP-IQA model using randomly selected subsets of 50, 100, and 200 samples from the training set. The regularization weight " + }, + { + "bbox": [ + 313, + 502, + 555, + 612 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 502, + 555, + 612 + ], + "type": "text", + "content": " is fixed at 5, and the training hyperparameters are set according to the schedule proposed by CoOp [56]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 621, + 512, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 621, + 512, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 621, + 512, + 635 + ], + "type": "text", + "content": "4.3. Performance Comparison with SOTA" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 640, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 640, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 640, + 554, + 689 + ], + "type": "text", + "content": "Our method effectively acquires extensive image quality assessment knowledge, enabling it to provide powerful priors for various IQA scenarios. Tab. 1 and Tab. 2 summarize the comparative results for different experimental settings." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Few-Shot Setting. Benefiting from the quality priors acquired during pre-training, our model achieves superior" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10447" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 293, + 190 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 293, + 190 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 293, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 293, + 190 + ], + "type": "table", + "html": "
TrainingLIVEFBLIVECKonIQLIVECSIQ
TestingKonIQLIVECKonIQLIVECCSIQLIVE
DBCNN0.7160.7240.7540.7550.7580.877
HyperIAQ0.7580.7350.7720.7850.7440.926
TReS0.7130.7400.7330.7860.761-
DEIQT0.7330.7810.7440.7940.7810.932
CLIP-IQA+0.6310.6200.6970.803--
LoDa0.7630.8050.7450.811--
GRMP-IQA0.7650.7900.7820.8310.8090.935
", + "image_path": "4bb84df46314964b4f397b9070359a0fc4cd778a3c1a3aedb34aa43f9d9f90e0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 61, + 228, + 293, + 312 + ], + "blocks": [ + { + "bbox": [ + 56, + 198, + 294, + 220 + ], + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 220 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 220 + ], + "type": "text", + "content": "Table 3. SRCC on the cross datasets validation. The best performances are highlighted in boldface." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 228, + 293, + 312 + ], + "lines": [ + { + "bbox": [ + 61, + 228, + 293, + 312 + ], + "spans": [ + { + "bbox": [ + 61, + 228, + 293, + 312 + ], + "type": "table", + "html": "
ComponentLIVECKonIQ
Pre-training Meta-learningPLCCSRCCPLCCSRCC
0.8250.7960.7880.764
✓0.8230.7880.7920.761
✓0.8580.8280.8440.811
", + "image_path": "287f4534f60b502c51a7a1bb1d176688d2533ce1882baaac6ed270d69ec85c81.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 350, + 295, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 295, + 602 + ], + "type": "text", + "content": "performance even with limited training data. As shown in Tab. 1, our GRMP-IQA significantly outperforms the second-best model, GRepQ [40], which is specifically designed for few-shot learning. Additionally, our approach exhibits clear advantages over MetaIQA, a meta-learning-based method, underscoring the effectiveness of metaprompts for rapidly adapting CLIP to diverse IQA tasks. These results validate GRMP-IQA's strong capability in learning quality-aware feature with limited labeled samples. Fully Supervised Setting. Tab. 2 compares GRMP-IQA with other BIQA methods under full supervised setting, including self-training approaches like CONTRIQUE [30] and Re-IQA [37], as well as CLIP-based methods such as LIQE [52] and CLIP-IQA+ [44]. GRMP-IQA outperforms almost all competitors on six datasets. Achieving leading performance across these datasets is particularly challenging due to the diverse range of image content and distortion types. Notably, the proposed method achieves competitive results with the state-of-the-art (SOTA) method on several datasets while utilizing only " + }, + { + "bbox": [ + 55, + 350, + 295, + 602 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 55, + 350, + 295, + 602 + ], + "type": "text", + "content": " of the training data compared to the fully supervised setting." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 611, + 251, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 251, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 251, + 624 + ], + "type": "text", + "content": "4.4. Generalization Capability Validation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 294, + 713 + ], + "type": "text", + "content": "To evaluate the generalization capacity of GRMP-IQA, we performed cross-dataset validation experiments, where the model was trained on one dataset and tested on others without parameter adjustments. Tab. 3 reports the SRCC results across five datasets. GRMP-IQA consistently outperforms state-of-the-art models in most cross Authentic scenarios, achieving notable improvements on LIVEC and" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 319, + 71, + 553, + 179 + ], + "blocks": [ + { + "bbox": [ + 55, + 320, + 294, + 343 + ], + "lines": [ + { + "bbox": [ + 55, + 320, + 294, + 343 + ], + "spans": [ + { + "bbox": [ + 55, + 320, + 294, + 343 + ], + "type": "text", + "content": "Table 4. Ablation study on the effectiveness of meta-learning when trained with only 50 samples." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 71, + 553, + 179 + ], + "lines": [ + { + "bbox": [ + 319, + 71, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 319, + 71, + 553, + 179 + ], + "type": "table", + "html": "
ComponentsLIVECKonIQPIPAL
MetaTextVisualPLCCSRCCPLCCSRCCPLCCSRCC
✓0.5790.5980.5920.5730.2160.203
0.6390.5890.5560.5540.3670.371
✓✓0.6990.6890.6790.6090.3230.312
✓✓0.7760.7420.7360.7010.3570.369
✓✓0.7590.7090.6220.5920.3620.396
✓✓✓0.8080.7700.7440.7130.4100.434
", + "image_path": "367facfc5ba92770699ca766737f0998f75c6cd6c90fe95e95dc29bdabc6125d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 319, + 217, + 553, + 308 + ], + "blocks": [ + { + "bbox": [ + 313, + 187, + 553, + 210 + ], + "lines": [ + { + "bbox": [ + 313, + 187, + 553, + 210 + ], + "spans": [ + { + "bbox": [ + 313, + 187, + 553, + 210 + ], + "type": "text", + "content": "Table 5. Ablation experiments with Meta-Prompt Pre-training component. The best performances are highlighted in boldface." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 217, + 553, + 308 + ], + "lines": [ + { + "bbox": [ + 319, + 217, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 319, + 217, + 553, + 308 + ], + "type": "table", + "html": "
MethodLabels50100
DatasetPLCCSRCCPLCCSRCC
w/o QGRLIVEC0.8580.8280.8750.848
w/ QGRLIVEC0.8640.8360.8830.857
w/o QGRKonIQ0.8440.8110.8720.840
w/ QGRKonIQ0.8800.8530.8960.872
", + "image_path": "53c3767d31b2c45804ecfa50aca6ccc0b0dcf99156eb5963dfa49c4287d0da5d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 316, + 553, + 338 + ], + "lines": [ + { + "bbox": [ + 313, + 316, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 553, + 338 + ], + "type": "text", + "content": "Table 6. Ablation experiments with QGR in few-shot setting. The best performances are highlighted in boldface." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 350, + 553, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 553, + 374 + ], + "type": "text", + "content": "KonIQ dataset. Additionally, it demonstrates strong competitiveness on synthetic datasets like LIVE and CSIQ." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 384, + 408, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 408, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 408, + 396 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 402, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 402, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 312, + 402, + 555, + 641 + ], + "type": "text", + "content": "Effect of Meta-Prompt Pre-training Module. This module consists of three key components: meta-learning, text meta-prompts, and visual meta-prompts. We conduct ablation studies in Tab. 5 to evaluate their zero-shot capabilities across various datasets. The baseline uses a CLIP model pre-trained on classification tasks. Row 2 evaluates meta-learning without prompts, where CLIP's visual and text encoders are fine-tuned to learn distortion knowledge across different distortions, improving IQA performance. However, tuning CLIP's weights impairs its original generalization, resulting in lower performance on KonIQ. Rows 3 and 4 assess fine-tuning only text or visual prompts during meta-learning. These strategies notably boost zero-shot performance—particularly on real-world datasets like KonIQ and LIVEC—by preserving CLIP's generalization while adapting to IQA tasks. Row 5 explores prompt tuning without meta-learning, underscoring meta-learning's role in curbing overfitting and maintaining generalization. The strongest results emerge when all components are combined, highlighting their complementary strengths." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Effect of Meta-learning. To further investigate whether the effectiveness of our method derives from meta-learning, we conducted an ablation study. Specifically, as shown in rows 1 and 2 of Tab. 4, when using Empirical Risk Minimization without meta-learning pre-training on two synthetic dataset, the fine-tuning performance on LIVEC and KonIQ dataset" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10448" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 157, + 167 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 157, + 167 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 157, + 167 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 157, + 167 + ], + "type": "image", + "image_path": "f29ceda765abccace48421cb5381e78efb8f16936ea8a846b45e3ef3fad816a4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 109, + 167, + 118, + 174 + ], + "lines": [ + { + "bbox": [ + 109, + 167, + 118, + 174 + ], + "spans": [ + { + "bbox": [ + 109, + 167, + 118, + 174 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 158, + 70, + 292, + 166 + ], + "blocks": [ + { + "bbox": [ + 158, + 70, + 292, + 166 + ], + "lines": [ + { + "bbox": [ + 158, + 70, + 292, + 166 + ], + "spans": [ + { + "bbox": [ + 158, + 70, + 292, + 166 + ], + "type": "image", + "image_path": "4637e83833b552599c0a6fd83bb23b03105f64f0926f34abce0c511455680a46.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 226, + 167, + 235, + 174 + ], + "lines": [ + { + "bbox": [ + 226, + 167, + 235, + 174 + ], + "spans": [ + { + "bbox": [ + 226, + 167, + 235, + 174 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 60, + 228, + 293, + 315 + ], + "blocks": [ + { + "bbox": [ + 60, + 228, + 293, + 315 + ], + "lines": [ + { + "bbox": [ + 60, + 228, + 293, + 315 + ], + "spans": [ + { + "bbox": [ + 60, + 228, + 293, + 315 + ], + "type": "table", + "html": "
λ50100200
PLCCSRCCPLCCSRCCPLCCSRCC
10.8500.8120.8740.8390.8930.866
30.8640.8390.8900.8650.9030.878
50.8800.8530.8960.8720.9080.883
70.8760.8510.8950.8700.9050.881
", + "image_path": "96ea14da1d293d4ceed89937b68951e7d92006334a0604e584873c91385a438b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 342, + 295, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 295, + 413 + ], + "type": "text", + "content": "was even worse than the baseline without pre-training. This indicates that additional pre-training data does not necessarily enhance performance and may even lead to overfitting [19]. In contrast, our meta-learning strategy effectively extracts generalizable quality priors from external data, significantly enhancing performance across datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 414, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 414, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 414, + 295, + 616 + ], + "type": "text", + "content": "Effect of Gradient Regularization. We conducted ablation studies on the QGR module. As shown in Tab. 6 and Fig. 4(a), without QGR, models are prone to overfitting under limited training data, with rapid performance saturation and a slight decline, which ultimately hinders generalization. By modulating training gradients, QGR effectively mitigates overfitting and enhances adaptability to the test dataset. To illustrate QGR's impact, we analyzed the angular difference between gradients " + }, + { + "bbox": [ + 55, + 414, + 295, + 616 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{qua}}" + }, + { + "bbox": [ + 55, + 414, + 295, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 414, + 295, + 616 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{sem}}" + }, + { + "bbox": [ + 55, + 414, + 295, + 616 + ], + "type": "text", + "content": " during training, as shown in Fig. 4(b). Without QGR, the angle between " + }, + { + "bbox": [ + 55, + 414, + 295, + 616 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{qua}}" + }, + { + "bbox": [ + 55, + 414, + 295, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 414, + 295, + 616 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{sem}}" + }, + { + "bbox": [ + 55, + 414, + 295, + 616 + ], + "type": "text", + "content": " tends toward 90 degrees, reflecting the orthogonality typical of high-dimensional random vectors [2]. In contrast, QGR introduces greater angular variation, suggesting that it encourages the model to explore more correlations between quality and semantic directions. This improves the model's ability to capture quality-related information and mitigates overfitting on limited data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 617, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 700 + ], + "type": "text", + "content": "Effect of soft weight " + }, + { + "bbox": [ + 55, + 617, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 617, + 295, + 700 + ], + "type": "text", + "content": ". On the KonIQ dataset, we conduct ablation experiments with various soft weights " + }, + { + "bbox": [ + 55, + 617, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 617, + 295, + 700 + ], + "type": "text", + "content": " in Eq. 8 to examine their impact, as detailed in Tab. 7. Results show that a small " + }, + { + "bbox": [ + 55, + 617, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 617, + 295, + 700 + ], + "type": "text", + "content": " diminishes the effectiveness of our QGR, whereas a large " + }, + { + "bbox": [ + 55, + 617, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 617, + 295, + 700 + ], + "type": "text", + "content": " causes substantial gradient changes and decreases performance. Given our observation of the trade-off, we adopt " + }, + { + "bbox": [ + 55, + 617, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\lambda = 5" + }, + { + "bbox": [ + 55, + 617, + 295, + 700 + ], + "type": "text", + "content": " in our experiments." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "content": "Computational Analysis. As shown in Tab. 8, we evaluate" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 375, + 116 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 375, + 116 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 375, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 375, + 116 + ], + "type": "image", + "image_path": "87501095bc1cbad1403f79f581372fea43a009bb6e5a41b124fda62fabdce399.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 376, + 71, + 433, + 115 + ], + "blocks": [ + { + "bbox": [ + 376, + 71, + 433, + 115 + ], + "lines": [ + { + "bbox": [ + 376, + 71, + 433, + 115 + ], + "spans": [ + { + "bbox": [ + 376, + 71, + 433, + 115 + ], + "type": "image", + "image_path": "71c34f6566a07040641767070b78cedbad979b97a9875465b83d34892511aa13.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 396, + 116, + 419, + 122 + ], + "lines": [ + { + "bbox": [ + 396, + 116, + 419, + 122 + ], + "spans": [ + { + "bbox": [ + 396, + 116, + 419, + 122 + ], + "type": "text", + "content": "Best ours" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 434, + 70, + 493, + 115 + ], + "blocks": [ + { + "bbox": [ + 434, + 70, + 493, + 115 + ], + "lines": [ + { + "bbox": [ + 434, + 70, + 493, + 115 + ], + "spans": [ + { + "bbox": [ + 434, + 70, + 493, + 115 + ], + "type": "image", + "image_path": "041b3ceeb80c7e3722a4fadce55cb5ea41bdf0d8a06bb3f176b5d9a360429e76.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 451, + 116, + 480, + 121 + ], + "lines": [ + { + "bbox": [ + 451, + 116, + 480, + 121 + ], + "spans": [ + { + "bbox": [ + 451, + 116, + 480, + 121 + ], + "type": "text", + "content": "Best DEIQT" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 494, + 71, + 552, + 115 + ], + "blocks": [ + { + "bbox": [ + 494, + 71, + 552, + 115 + ], + "lines": [ + { + "bbox": [ + 494, + 71, + 552, + 115 + ], + "spans": [ + { + "bbox": [ + 494, + 71, + 552, + 115 + ], + "type": "image", + "image_path": "c7b275e19a672c94c75dcb0d0b94938126c7b5072adf2c4b6b1752f36f1a7ce8.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 331, + 123, + 376, + 140 + ], + "blocks": [ + { + "bbox": [ + 332, + 117, + 354, + 123 + ], + "lines": [ + { + "bbox": [ + 332, + 117, + 354, + 123 + ], + "spans": [ + { + "bbox": [ + 332, + 117, + 354, + 123 + ], + "type": "text", + "content": "Best ours" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 331, + 123, + 376, + 140 + ], + "lines": [ + { + "bbox": [ + 331, + 123, + 376, + 140 + ], + "spans": [ + { + "bbox": [ + 331, + 123, + 376, + 140 + ], + "type": "image", + "image_path": "8633e7f956e2a7d65bdaec1d827e8142c963a84bdc4afb8365aeab8aec95dba0.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 394, + 123, + 406, + 140 + ], + "blocks": [ + { + "bbox": [ + 394, + 123, + 406, + 140 + ], + "lines": [ + { + "bbox": [ + 394, + 123, + 406, + 140 + ], + "spans": [ + { + "bbox": [ + 394, + 123, + 406, + 140 + ], + "type": "image", + "image_path": "fce79b2080d526aa9879f55f54a010bb29e8181b5170112356030a85b310bb4a.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 434, + 123, + 493, + 148 + ], + "blocks": [ + { + "bbox": [ + 434, + 123, + 493, + 148 + ], + "lines": [ + { + "bbox": [ + 434, + 123, + 493, + 148 + ], + "spans": [ + { + "bbox": [ + 434, + 123, + 493, + 148 + ], + "type": "image", + "image_path": "bf457257d3440694fb7015a495b185e1c367b944972727510da9b80356ad4ffa.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 450, + 142, + 483, + 148 + ], + "lines": [ + { + "bbox": [ + 450, + 142, + 483, + 148 + ], + "spans": [ + { + "bbox": [ + 450, + 142, + 483, + 148 + ], + "type": "text", + "content": "Worst DEIQT" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 509, + 123, + 550, + 148 + ], + "blocks": [ + { + "bbox": [ + 510, + 117, + 539, + 122 + ], + "lines": [ + { + "bbox": [ + 510, + 117, + 539, + 122 + ], + "spans": [ + { + "bbox": [ + 510, + 117, + 539, + 122 + ], + "type": "text", + "content": "Best DEIQT" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 509, + 123, + 550, + 148 + ], + "lines": [ + { + "bbox": [ + 509, + 123, + 550, + 148 + ], + "spans": [ + { + "bbox": [ + 509, + 123, + 550, + 148 + ], + "type": "image", + "image_path": "a5f32b7f336a4237807150d53b768bdcaa132e49b4a4fdb05ac1a487eba338ce.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 316, + 148, + 374, + 193 + ], + "blocks": [ + { + "bbox": [ + 55, + 178, + 295, + 222 + ], + "lines": [ + { + "bbox": [ + 55, + 178, + 295, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 178, + 295, + 222 + ], + "type": "text", + "content": "Figure 4. (a) SRCC value during training with 50 samples. (b) Angles between " + }, + { + "bbox": [ + 55, + 178, + 295, + 222 + ], + "type": "inline_equation", + "content": "G_{qua}" + }, + { + "bbox": [ + 55, + 178, + 295, + 222 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 178, + 295, + 222 + ], + "type": "inline_equation", + "content": "G_{sem}" + }, + { + "bbox": [ + 55, + 178, + 295, + 222 + ], + "type": "text", + "content": " during training on KonIQ. Under the constraints of QGR, our method captures more correlation between quality knowledge and general semantic knowledge." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 331, + 141, + 357, + 148 + ], + "lines": [ + { + "bbox": [ + 331, + 141, + 357, + 148 + ], + "spans": [ + { + "bbox": [ + 331, + 141, + 357, + 148 + ], + "type": "text", + "content": "Worst ours" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 148, + 374, + 193 + ], + "lines": [ + { + "bbox": [ + 316, + 148, + 374, + 193 + ], + "spans": [ + { + "bbox": [ + 316, + 148, + 374, + 193 + ], + "type": "image", + "image_path": "cea95d3a7e6ecbb9a031df5986daea0c42a317cece6f08ee536caded7df2750c.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 338, + 193, + 345, + 198 + ], + "lines": [ + { + "bbox": [ + 338, + 193, + 345, + 198 + ], + "spans": [ + { + "bbox": [ + 338, + 193, + 345, + 198 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 205, + 553, + 239 + ], + "lines": [ + { + "bbox": [ + 313, + 205, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 553, + 239 + ], + "type": "text", + "content": "Figure 5. gMAD results between DEIQT [33] and our GRMP-IQA. (a) Fixed DEIQT at low quality. (b) Fixed DEIQT at high quality. (c) Fixed ours at low quality. (d) Fixed ours at high quality." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 375, + 148, + 434, + 193 + ], + "blocks": [ + { + "bbox": [ + 394, + 142, + 421, + 148 + ], + "lines": [ + { + "bbox": [ + 394, + 142, + 421, + 148 + ], + "spans": [ + { + "bbox": [ + 394, + 142, + 421, + 148 + ], + "type": "text", + "content": "Worst ours" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 375, + 148, + 434, + 193 + ], + "lines": [ + { + "bbox": [ + 375, + 148, + 434, + 193 + ], + "spans": [ + { + "bbox": [ + 375, + 148, + 434, + 193 + ], + "type": "image", + "image_path": "d1116aec69a3babc426d88e89cc02072fa4be7f03290cd301f91af0abe9a1ea2.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 193, + 411, + 199 + ], + "lines": [ + { + "bbox": [ + 403, + 193, + 411, + 199 + ], + "spans": [ + { + "bbox": [ + 403, + 193, + 411, + 199 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 435, + 148, + 493, + 193 + ], + "blocks": [ + { + "bbox": [ + 435, + 148, + 493, + 193 + ], + "lines": [ + { + "bbox": [ + 435, + 148, + 493, + 193 + ], + "spans": [ + { + "bbox": [ + 435, + 148, + 493, + 193 + ], + "type": "image", + "image_path": "c0356403e51cceb9f6c17e0327a993665552725d750b9344ac2f0077f5c804a5.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 460, + 193, + 468, + 199 + ], + "lines": [ + { + "bbox": [ + 460, + 193, + 468, + 199 + ], + "spans": [ + { + "bbox": [ + 460, + 193, + 468, + 199 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 493, + 148, + 552, + 193 + ], + "blocks": [ + { + "bbox": [ + 493, + 148, + 552, + 193 + ], + "lines": [ + { + "bbox": [ + 493, + 148, + 552, + 193 + ], + "spans": [ + { + "bbox": [ + 493, + 148, + 552, + 193 + ], + "type": "image", + "image_path": "cd6d31eb3b740362ca7955d0b9ebeedbb3aaf22da2f466a261beeaaa44b24113.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 520, + 193, + 528, + 198 + ], + "lines": [ + { + "bbox": [ + 520, + 193, + 528, + 198 + ], + "spans": [ + { + "bbox": [ + 520, + 193, + 528, + 198 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "table", + "bbox": [ + 317, + 243, + 553, + 303 + ], + "blocks": [ + { + "bbox": [ + 72, + 323, + 278, + 335 + ], + "lines": [ + { + "bbox": [ + 72, + 323, + 278, + 335 + ], + "spans": [ + { + "bbox": [ + 72, + 323, + 278, + 335 + ], + "type": "text", + "content": "Table 7. The ablation study about soft weight " + }, + { + "bbox": [ + 72, + 323, + 278, + 335 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 72, + 323, + 278, + 335 + ], + "type": "text", + "content": " in Eq. 8." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 243, + 553, + 303 + ], + "lines": [ + { + "bbox": [ + 317, + 243, + 553, + 303 + ], + "spans": [ + { + "bbox": [ + 317, + 243, + 553, + 303 + ], + "type": "table", + "html": "
MethodParamEfficiencyPerformance
Throughput ↑Latency ↓80% Train20% Train
Q-Align8.2B9.9 (img/sec)101.0 (ms)0.941 / 0.9400.901 / 0.903
Ours151M43.0 (img/sec)21.3 (ms)0.945 / 0.9340.931 / 0.915
", + "image_path": "cf9c560dcaecd0bf1c051fca8981d0f619ce278dc8a7a82b127d2cd042c1c997.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "table_body" + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 311, + 553, + 333 + ], + "lines": [ + { + "bbox": [ + 313, + 311, + 553, + 333 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 553, + 333 + ], + "type": "text", + "content": "Table 8. Comparison with Q-Align across varying training data ratio to evaluate efficiency and PLCC/SRCC on the KonIQ dataset." + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 341, + 553, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 553, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 553, + 389 + ], + "type": "text", + "content": "the inference latency and throughput of our method on an RTX3090 GPU, surpassing Q-Align [45] in both efficiency and accuracy. Notably, our model has only 41M learnable and 151M total parameters, far fewer than Q-Align's 8.2B." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 313, + 396, + 432, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 432, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 432, + 408 + ], + "type": "text", + "content": "4.6. Qualitative Analysis" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 313, + 413, + 555, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 555, + 568 + ], + "type": "text", + "content": "The gMAD competition [28] is a standard method for assessing IQA model robustness. It selects image pairs where the attacker model predicts a large quality difference, while the defender perceives them as similar. Observers then evaluate these pairs to gauge model generalization. As shown in Fig. 5, when our model acts as the defender, the attacker's selected pairs exhibit minimal perceptual quality changes. Conversely, as the attacker, our model consistently identifies pairs with significant quality differences. This highlights its strong defensive and offensive capabilities. Notably, in the second column, DEIQT misclassifies semantically similar images as having comparable quality, whereas our model accurately distinguishes their differences." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 313, + 575, + 388, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 575, + 388, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 575, + 388, + 588 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we propose the GRMP-IQA framework, which generalizes well with limited data. It includes a meta-learning pre-training module that enables the CLIP model to rapidly adapt to IQA tasks and an adaptive gradient regulation module that refines gradient trajectories during fin-tuning, focusing updates on quality-aware knowledge while minimizing the negative impact of over-reliance on semantic noise. Comprehensive experiments on various BIQA datasets validate the superior generalization of our framework, especially in data-scarce scenarios." + } + ] + } + ], + "index": 41 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10449" + } + ] + } + ], + "index": 42 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 180 + ], + "type": "text", + "content": "Acknowledgements. This work was supported by the National Science Fund for Distinguished Young Scholars (No.62025603), the National Natural Science Foundation of China (No. U21B2037, No. U22B2051, No. U23A20383, No. U21A20472, No. 62176222, No. 62176223, No. 62176226, No. 62072386, No. 62072387, No. 62072389, No. 62002305 and No. 62272401), and the Natural Science Foundation of Fujian Province of China (No. 2021J06003, No.2022J06001)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 190, + 115, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 190, + 115, + 202 + ], + "spans": [ + { + "bbox": [ + 56, + 190, + 115, + 202 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 209, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 209, + 296, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 209, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 61, + 209, + 296, + 266 + ], + "type": "text", + "content": "[1] Nithin C Babu, Vignesh Kannan, and Rajiv Soundararajan. No reference opinion unaware quality assessment of authentically distorted images. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2459-2468, 2023. 2, 3, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 266, + 296, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 266, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 62, + 266, + 296, + 300 + ], + "type": "text", + "content": "[2] Tony Cai, Jianqing Fan, and Tiefeng Jiang. Distributions of angles in random packing on spheres. The Journal of Machine Learning Research, 14(1):1837-1864, 2013. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 300, + 295, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 300, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 62, + 300, + 295, + 354 + ], + "type": "text", + "content": "[3] Yuming Fang, Hanwei Zhu, Yan Zeng, Kede Ma, and Zhou Wang. Perceptual quality assessment of smartphone photography. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3677-3686, 2020. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 356, + 295, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 356, + 295, + 400 + ], + "spans": [ + { + "bbox": [ + 62, + 356, + 295, + 400 + ], + "type": "text", + "content": "[4] Yuqian Fu, Yanwei Fu, and Yu-Gang Jiang. Meta-fdmixup: Cross-domain few-shot learning guided by labeled target data. In Proceedings of the 29th ACM international conference on multimedia, pages 5326-5334, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 401, + 295, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 401, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 62, + 401, + 295, + 445 + ], + "type": "text", + "content": "[5] Hongcheng Gao, Tianyu Pang, Chao Du, Taihang Hu, Zhijie Deng, and Min Lin. Meta-unlearning on diffusion models: Preventing relearning unlearned concepts. arXiv preprint arXiv:2410.12777, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 446, + 295, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 446, + 295, + 500 + ], + "spans": [ + { + "bbox": [ + 62, + 446, + 295, + 500 + ], + "type": "text", + "content": "[6] Hongcheng Gao, Jiashu Qu, Jingyi Tang, Baolong Bi, Yue Liu, Hongyu Chen, Li Liang, Li Su, and Qingming Huang. Exploring hallucination of large multimodal models in video understanding: Benchmark, analysis and mitigation. arXiv preprint arXiv:2503.19622, 2025. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 502, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 295, + 544 + ], + "type": "text", + "content": "[7] Deepti Ghadiyaram and Alan C Bovik. Massive online crowdsourced study of subjective and objective picture quality. IEEE Transactions on Image Processing, 25(1):372-387, 2015. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 546, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 546, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 62, + 546, + 295, + 601 + ], + "type": "text", + "content": "[8] S Alireza Golestaneh, Saba Dadsetan, and Kris M Kitani. No-reference image quality assessment via transformers, relative ranking, and self-consistency. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1220-1230, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 602, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 602, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 62, + 602, + 295, + 646 + ], + "type": "text", + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 647, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 690 + ], + "type": "text", + "content": "[10] Vlad Hosu, Hanhe Lin, Tamas Sziranyi, and Dietmar Saupe. Koniq-10k: An ecologically valid database for deep learning of blind image quality assessment. IEEE Transactions on Image Processing, 29:4041–4056, 2020. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[11] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Vi" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 335, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 555, + 95 + ], + "type": "text", + "content": "sual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 97, + 555, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 555, + 163 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 555, + 163 + ], + "type": "text", + "content": "[12] Gu Jinjin, Cai Haoming, Chen Haoyu, Ye Xiaoxing, Jimmy S Ren, and Dong Chao. Pipal: a large-scale image quality assessment dataset for perceptual image restoration. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI 16, pages 633-651. Springer, 2020. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 163, + 554, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 554, + 208 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 554, + 208 + ], + "type": "text", + "content": "[13] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5148-5157, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 209, + 554, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 554, + 262 + ], + "type": "text", + "content": "[14] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19113-19122, 2023. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 264, + 554, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 264, + 554, + 331 + ], + "spans": [ + { + "bbox": [ + 317, + 264, + 554, + 331 + ], + "type": "text", + "content": "[15] Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Self-regulating prompts: Foundational model adaptation without forgetting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15190-15200, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 332, + 554, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 332, + 554, + 364 + ], + "spans": [ + { + "bbox": [ + 317, + 332, + 554, + 364 + ], + "type": "text", + "content": "[16] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 365, + 554, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 365, + 554, + 409 + ], + "spans": [ + { + "bbox": [ + 317, + 365, + 554, + 409 + ], + "type": "text", + "content": "[17] Eric Cooper Larson and Damon Michael Chandler. Most apparent distortion: full-reference image quality assessment and the role of strategy. Journal of electronic imaging, 19 (1):011006, 2010. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 411, + 554, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 554, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 554, + 443 + ], + "type": "text", + "content": "[18] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 445, + 554, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 554, + 500 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 554, + 500 + ], + "type": "text", + "content": "[19] Aobo Li, Jinjian Wu, Yongxu Liu, and Leida Li. Bridging the synthetic-to Authentic gap: Distortion-guided unsupervised domain adaptation for blind image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28422-28431, 2024. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 501, + 554, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 501, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 501, + 554, + 555 + ], + "type": "text", + "content": "[20] Chunyi Li, Zicheng Zhang, Haoning Wu, Wei Sun, Xiongkuo Min, Xiaohong Liu, Guangtao Zhai, and Weisi Lin. Agiqa-3k: An open database for ai-generated image quality assessment. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 557, + 554, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 557, + 554, + 623 + ], + "spans": [ + { + "bbox": [ + 317, + 557, + 554, + 623 + ], + "type": "text", + "content": "[21] Juncheng Li, Minghe Gao, Longhui Wei, Siliang Tang, Wenqiao Zhang, Mengze Li, Wei Ji, Qi Tian, Tat-Seng Chua, and Yueting Zhuang. Gradient-regulated meta-prompt learning for generalizable vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2551-2562, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 624, + 554, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 624, + 554, + 690 + ], + "spans": [ + { + "bbox": [ + 317, + 624, + 554, + 690 + ], + "type": "text", + "content": "[22] Xudong Li, Timin Gao, Runze Hu, Yan Zhang, Shengchuan Zhang, Xiawu Zheng, Jingyuan Zheng, Yunhang Shen, Ke Li, Yutao Liu, et al. Adaptive feature selection for no-reference image quality assessment by mitigating semantic noise sensitivity. In International Conference on Machine Learning, pages 27808-27821. PMLR, 2024. 2, 3, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "type": "text", + "content": "[23] Xudong Li, Runze Hu, Jingyuan Zheng, Yan Zhang, Shengchuan Zhang, Xiawu Zheng, Ke Li, Yunhang Shen," + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10450" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 295, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 295, + 117 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 295, + 117 + ], + "type": "text", + "content": "Yutao Liu, Pingyang Dai, et al. Integrating global context contrast and local sensitivity for blind image quality assessment. In International Conference on Machine Learning, pages 27920-27941. PMLR, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 173 + ], + "type": "text", + "content": "[24] Xudong Li, Wenjie Nie, Yan Zhang, Runze Hu, Ke Li, Xiawu Zheng, and Liujuan Cao. Distilling spatially-heterogeneous distortion perception for blind image quality assessment. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 2344-2354, 2025. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 295, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 295, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 295, + 218 + ], + "type": "text", + "content": "[25] Hanhe Lin, Vlad Hosu, and Dietmar Saupe. Kadid-10k: A large-scale artificially distorted iqa database. In 2019 Eleventh International Conference on Quality of Multimedia Experience (QoMEX), pages 1-3. IEEE, 2019. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 220, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 220, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 57, + 220, + 294, + 252 + ], + "type": "text", + "content": "[26] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 295, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 295, + 297 + ], + "type": "text", + "content": "[27] Ziwei Luo, Fredrik K Gustafsson, Zheng Zhao, Jens Sjolund, and Thomas B Schon. Controlling vision-language models for multi-task image restoration. In International Conference on Learning Representations, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 299, + 295, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 295, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 295, + 363 + ], + "type": "text", + "content": "[28] Kede Ma, Qingbo Wu, Zhou Wang, Zhengfang Duanmu, Hongwei Yong, Hongliang Li, and Lei Zhang. Group mad competition-a new methodology to compare objective image quality models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1664-1673, 2016. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 295, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 295, + 410 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 295, + 410 + ], + "type": "text", + "content": "[29] Kede Ma, Xuelin Liu, Yuming Fang, and Eero P Simoncelli. Blind image quality assessment by learning from multiple annotators. In 2019 IEEE international conference on image processing (ICIP), pages 2344-2348. IEEE, 2019. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 295, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 295, + 454 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 295, + 454 + ], + "type": "text", + "content": "[30] Pavan C Madhusudana, Neil Birkbeck, Yilin Wang, Balu Adsumilli, and Alan C Bovik. Image quality assessment using contrastive learning. IEEE Transactions on Image Processing, 31:4149-4161, 2022. 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 456, + 295, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 295, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 295, + 499 + ], + "type": "text", + "content": "[31] Jinyoung Park, Juyeon Ko, and Hyunwoo J Kim. Prompt learning via meta-regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26940–26950, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 501, + 295, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 295, + 566 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 295, + 566 + ], + "type": "text", + "content": "[32] Nikolay Ponomarenko, Lina Jin, Oleg Ieremeiev, Vladimir Lukin, Karen Egiazarian, Jaakko Astola, Benoit Vozel, Kacem Chehdi, Marco Carli, Federica Battisti, et al. Image database tid2013: Peculiarities, results and perspectives. Signal processing: Image communication, 30:57-77, 2015. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 568, + 295, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 295, + 622 + ], + "type": "text", + "content": "[33] Guanyi Qin, Runze Hu, Yutao Liu, Xiawu Zheng, Haotian Liu, Xiu Li, and Yan Zhang. Data-efficient image quality assessment with attention-panel decoder. In Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence, 2023. 2, 3, 6, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 295, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 657 + ], + "type": "text", + "content": "[34] B Qu, H Li, and W Gao. Bringing textual prompt to a-generated image quality assessment, in 2024 IEEE international conference on multimedia and expo (icme), 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "type": "text", + "content": "[35] Bowen Qu, Xiaoyu Liang, Shangkun Sun, and Wei Gao. Exploring aigc video quality: A focus on visual harmony videotext consistency and domain distribution gap. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6652-6660, 2024. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 138 + ], + "type": "text", + "content": "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 140, + 553, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 193 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 193 + ], + "type": "text", + "content": "[37] Avinab Saha, Sandeep Mishra, and Alan C Bovik. Reiqa: Unsupervised learning for image quality assessment in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5846-5855, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 194, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 194, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 194, + 553, + 237 + ], + "type": "text", + "content": "[38] Hamid R Sheikh, Muhammad F Sabir, and Alan C Bovik. A statistical evaluation of recent full reference image quality assessment algorithms. IEEE Transactions on image processing, 15(11):3440-3451, 2006. 1, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "type": "text", + "content": "[39] Tianshu Song, Leida Li, Deqiang Cheng, Pengfei Chen, and Jinjian Wu. Active learning-based sample selection for label-efficient blind image quality assessment. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 283, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 553, + 348 + ], + "type": "text", + "content": "[40] Suhas Srinath, Shankhanil Mitra, Shika Rao, and Rajiv Soundararajan. Learning generalizable perceptual representations for data-efficient no-reference image quality assessment. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 22-31, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 350, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 403 + ], + "type": "text", + "content": "[41] Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3667-3676, 2020. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 404, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 448 + ], + "type": "text", + "content": "[42] Zhengzhong Tu, Yilin Wang, Neil Birkbeck, Balu Adsumilli, and Alan C Bovik. Ugc-vqa: Benchmarking blind video quality assessment for user generated content. IEEE Transactions on Image Processing, 30:4449-4464, 2021. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 449, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 469 + ], + "type": "text", + "content": "[43] Joaquin Vanschoren. Meta-learning: A survey. arXiv preprint arXiv:1810.03548, 2018. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 471, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 553, + 514 + ], + "type": "text", + "content": "[44] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 515, + 553, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 553, + 580 + ], + "type": "text", + "content": "[45] Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Yixuan Gao, Annan Wang, Erli Zhang, Wenxiu Sun, et al. Q-align: Teaching lmmss for visual scoring via discrete text-defined levels. In International Conference on Machine Learning, pages 54015-54029. PMLR, 2024. 2, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 582, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 646 + ], + "type": "text", + "content": "[46] Kangmin Xu, Liang Liao, Jing Xiao, Chaofeng Chen, Haoning Wu, Qiong Yan, and Weisi Lin. Boosting image quality assessment through efficient transformer adaptation with local feature enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2662-2672, 2024. 2, 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "text", + "content": "[47] Zhengzhuo Xu, Bowen Qu, Yiyan Qi, Sinan Du, Chengjin Xu, Chun Yuan, and Jian Guo. Chartmoe: Mixture of diversely aligned expert connector for chart understanding. arXiv preprint arXiv:2409.03277, 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[48] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10451" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 653 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "text", + "content": "Maniqa: Multi-dimension attention network for no-reference image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1191-1200, 2022. 1, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 183 + ], + "type": "text", + "content": "[49] Zhenqiang Ying, Haoran Niu, Praful Gupta, Dhruv Mahajan, Deepti Ghadiyaram, and Alan Bovik. From patches to pictures (paq-2-piq): Mapping the perceptual space of picture quality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3575–3585, 2020. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 296, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 296, + 228 + ], + "type": "text", + "content": "[50] Ji Zhang, Shihan Wu, Lianli Gao, Heng Tao Shen, and Jingkuan Song. Dept: Decoupled prompt tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12924-12933, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 229, + 296, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 229, + 296, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 229, + 296, + 282 + ], + "type": "text", + "content": "[51] Weixia Zhang, Kede Ma, Jia Yan, Dexiang Deng, and Zhou Wang. Blind image quality assessment using a deep bilinear convolutional neural network. IEEE Transactions on Circuits and Systems for Video Technology, 30(1):36-47, 2018. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 285, + 296, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 285, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 296, + 350 + ], + "type": "text", + "content": "[52] Weixia Zhang, Guangtao Zhai, Ying Wei, Xiaokang Yang, and Kede Ma. Blind image quality assessment via vision-language correspondence: A multitask learning perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14071-14081, 2023. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 352, + 296, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 352, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 352, + 296, + 407 + ], + "type": "text", + "content": "[53] Kai Zhao, Kun Yuan, Ming Sun, Mading Li, and Xing Wen. Quality-aware pre-trained models for blind image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22302-22313, 2023. 2, 3, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 408, + 296, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 408, + 296, + 451 + ], + "spans": [ + { + "bbox": [ + 56, + 408, + 296, + 451 + ], + "type": "text", + "content": "[54] Limin Zheng, Yu Luo, Zihan Zhou, Jie Ling, and Guanghui Yue. Cdinet: Content distortion interaction network for blind image quality assessment. IEEE Transactions on Multimedia, 2024. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 453, + 296, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 453, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 56, + 453, + 296, + 506 + ], + "type": "text", + "content": "[55] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16816-16825, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 508, + 296, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 508, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 56, + 508, + 296, + 552 + ], + "type": "text", + "content": "[56] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 553, + 296, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 553, + 296, + 597 + ], + "spans": [ + { + "bbox": [ + 56, + 553, + 296, + 597 + ], + "type": "text", + "content": "[57] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15659-15669, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 597, + 296, + 653 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 597, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 56, + 597, + 296, + 653 + ], + "type": "text", + "content": "[58] Hancheng Zhu, Leida Li, Jinjian Wu, Weisheng Dong, and Guangming Shi. Metaiaq: Deep meta-learning for no-reference image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14143-14152, 2020. 3, 4, 6" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10452" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Few-Shot Pattern Detection via Template Matching and Regression/9a1fc29b-9538-4897-943a-30b2211079d1_content_list.json b/2025/Few-Shot Pattern Detection via Template Matching and Regression/9a1fc29b-9538-4897-943a-30b2211079d1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fa5d64224abe41af3f9f265bc673aecf3165a1ec --- /dev/null +++ b/2025/Few-Shot Pattern Detection via Template Matching and Regression/9a1fc29b-9538-4897-943a-30b2211079d1_content_list.json @@ -0,0 +1,1598 @@ +[ + { + "type": "text", + "text": "Few-Shot Pattern Detection via Template Matching and Regression", + "text_level": 1, + "bbox": [ + 158, + 130, + 836, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Eunchan Jo Dahiyun Kang Sanghyun Kim Yunseon Choi Minsu Cho", + "bbox": [ + 156, + 181, + 834, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pohang University of Science and Technology (POSTECH), South Korea", + "bbox": [ + 207, + 204, + 789, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://cvlab.postech.ac.kr/research/TMR", + "bbox": [ + 316, + 224, + 674, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 273, + 326, + 289 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We address the problem of few-shot pattern detection, which aims to detect all instances of a given pattern, typically represented by a few exemplars, from an input image. Although similar problems have been studied in few-shot object counting and detection (FSCD), previous methods and their benchmarks have narrowed patterns of interest to object categories and often fail to localize non-object patterns. In this work, we propose a simple yet effective detector based on template matching and regression, dubbed TMR. While previous FSCD methods typically represent target exemplars as spatially collapsed prototypes and lose structural information, we revisit classic template matching and regression. It effectively preserves and leverages the spatial layout of exemplars through a minimalistic structure with a small number of learnable convolutional or projection layers on top of a frozen backbone. We also introduce a new dataset, dubbed RPINE, which covers a wider range of patterns than existing object-centric datasets. Our method outperforms the state-of-the-art methods on the three benchmarks, RPINE, FSCD-147, and FSCD-LVIS, and demonstrates strong generalization in cross-dataset evaluation.", + "bbox": [ + 88, + 306, + 485, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 647, + 220, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Few-shot detection aims to identify target patterns with minimal labeled examples. While significant progress has been made in few-shot object detection [7, 8, 18, 68, 69], most existing methods remain object-centric, focusing primarily on identifying object-level patterns with relatively clear boundaries. However, many real-world applications require detecting arbitrary target patterns that extend beyond objects to include structural, geometric, or abstract patterns across diverse visual data. Despite recent progress based on deep neural networks, current methods still fall short in addressing these broader pattern detection tasks. Furthermore, the object-centric design of conventional few-shot detectors may lead to performance degradation when the target object lacks clear boundaries or when occlusion and deformation cause its boundaries to become indistinct.", + "bbox": [ + 89, + 674, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5fe6ad08219b7d44758127fa1a9a891bc25f37fc8db852944122976036b023c4.jpg", + "image_caption": [ + "one-shot exemplars", + "Figure 1. Few-shot pattern detection. Given a few exemplar for each target pattern (left), the task is to detect all matching instances of each pattern (right). This example includes non-object patterns (e.g., green and blue) as well as object patterns (e.g., red)." + ], + "image_footnote": [], + "bbox": [ + 526, + 271, + 700, + 393 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2dd86fb3cd93c131095a899e5df13399f4d7197987db6270b1a06ced863f4d22.jpg", + "image_caption": [ + "ground-truth pattern detection" + ], + "image_footnote": [], + "bbox": [ + 715, + 271, + 890, + 392 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The task of few-shot pattern detection is illustrated in Fig. 1. Recent related research topics for few-shot detection include few-shot counting and detection [47], and few-shot object detection [7]. Both aim to reduce the annotation cost of object detection [12, 33, 57, 64] by learning to detect all instances of given support exemplars. Consequently, many of these methods are heavily biased on object-centric benchmarks prior [6, 16, 32, 47]. In addition, many recent approaches [7, 18, 23, 52, 68, 69] represent the support exemplars as the spatially pooled vector, often named as a prototype [60]. While this pooling strategy is effective for detecting objects, it collapses the geometric properties, such as the shape and structure of the support exemplars. As a consequence, these methods tend to underperform when detecting non-object geometric patterns such as object parts or shape-intensive elements as shown in Fig. 3.", + "bbox": [ + 511, + 488, + 908, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we revisit the classic template matching strategy and propose a simple yet effective few-shot detector for arbitrary patterns. The proposed method, dubbed template matching and regression (TMR), is designed to be aware of the structure and shape of given exemplars. Given an input image, TMR first extracts a feature map using a backbone network. It then crops a template feature from the support exemplar's bounding box using a template extraction technique based on RoIAlign [22]. This template is correlated with the image feature map to produce a template matching feature map. Using this correlation map, the", + "bbox": [ + 511, + 734, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21578", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/65eba8dd183601e29102b9e3ae5c8647d189fc49c1c40666eb235d38d7c0bc15.jpg", + "image_caption": [ + "Figure 2. Annotation examples of FSCD-147 [47] and RPINE. FSCD-147 is annotated with the repetitive object-level patterns but disregards the repetition of non-object patterns such as the egg tray pattern under the eggs in the first image. RPINE is annotated with arbitrarily noticed repeated patterns which include non-object patterns and nameless parts of objects. Plus, FSCD-147 is annotated with a single pattern, while RPINE is annotated with all existing repetitions (marked with different colors) recognized by three different annotators." + ], + "image_footnote": [], + "bbox": [ + 94, + 69, + 898, + 272 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/82a6181c922bf3664fa51dd561636a44a7fccb01cc60d33360a5bef96a04d6a6.jpg", + "image_caption": [ + "Figure 3. Few-shot detectors trained with strong object prior are often biased to objects instances and struggle to detect non-objects." + ], + "image_footnote": [], + "bbox": [ + 94, + 340, + 187, + 434 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/06d2c2087d961c1bd1a809ff91b41c64267633b76b15c576c64241e87c8bd70f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 187, + 340, + 276, + 434 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/578ef30a8b6e1c5d7eac771154410f8299ce3010aad6597453378ee5372df632.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 339, + 367, + 434 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/96b66ed01472422db90f254fc79143c02fcc0f3ba72853da1fa6d81cf924bedc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 339, + 477, + 434 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "model learns bounding box regression parameters to rectify the support exemplar's box size adaptively. This process, termed support-conditioned regression, enables the model to handle support exemplars of varying sizes more effectively. Notably, TMR consists only of a few $3 \\times 3$ and linear projections without any complicated modules such as cross-attention, commonly used in prior work [47, 52].", + "bbox": [ + 89, + 476, + 482, + 582 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although TMR is designed for general pattern detection, existing benchmarks (e.g., FSCD-147 [47], FSCD-LVIS [47]) mainly target object-level patterns, limiting comprehensive evaluation. To address this, we introduce a new dataset, Repeated Patterns IN Everywhere (RPINE), which covers diverse repeated patterns in the real world. RPINE contains images with varying degrees of objectness, from well-defined object-level patterns to non-object patterns, all annotated with bounding boxes via crowd-sourcing. Compared to FSCD datasets, RPINE provides broader coverage, including both non-object patterns and nameless parts of objects, as illustrated in Fig. 2.", + "bbox": [ + 89, + 583, + 482, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "TMR demonstrates strong performance in detecting repeated patterns, not only on RPINE but also on the FSCD benchmarks, FSCD-147 and FSCD-LVIS [47]. In particular, TMR is especially effective on RPINE, includes diverse patterns with minimal object priors. Notably, our simple architecture contributes to improved generalization across datasets. Our contribution is summarized as follows:", + "bbox": [ + 89, + 763, + 482, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We generalize the few-shot object counting and detection to a pattern detection task that does not assume objectness", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "in either the target patterns or exemplars.", + "bbox": [ + 526, + 340, + 797, + 356 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present a simple yet effective pattern detector by refining template matching, which efficiently detects coherent patterns guided by exemplars.", + "- We introduce a new densely annotated dataset, RPINE, which covers diverse repetitive patterns in the real world, ranging from object-level patterns to non-object patterns.", + "- TMR not only outperforms the state-of-the-art FSCD models on RPINE and FSCD-LVIS but also achieves strong cross-dataset generalization." + ], + "bbox": [ + 511, + 357, + 903, + 491 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 513, + 505, + 651, + 522 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Few-shot object detection (FSOD) aims to detect objects of novel classes using only a few support images of novel classes. Existing methods can be roughly categorized into two groups: finetuning based [8, 15, 63, 67, 68] and meta-learning based [7, 17-19, 24, 69, 71] approaches. Despite significant progress, finetuning methods require retraining for every new classes. For pattern detection, it is not desired to fine-tune each arbitrary pattern with indistinct object categorization. In contrast, meta-learning methods avoid class-specific fine-tuning and typically construct prototypes from support images for classifying the bounding boxes of the query image. These methods often apply global average pooling to obtain prototypes from each support exemplar, which collapses the spatial structure of the exemplar. While such spatially collapsed prototypes may effectively capture object-level bounding boxes, they struggle with non-object patterns that exhibit complex spatial structures.", + "bbox": [ + 511, + 534, + 906, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Few-shot counting (FSC) aims to count objects given a few exemplars (typically 1 to 3) within an image. Previous methods [30, 40, 55, 56, 70] have tackled this task using density-map regression without detecting the individual bounding boxes. Towards localization-based counting, few-shot counting and detection (FSCD) is proposed to combine FSC and detection. Such models include Counting-", + "bbox": [ + 511, + 795, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21579", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b992c1cab19b66227e6eb39c4f3e257111f604c62d551c81b4dcbf705283a70e.jpg", + "image_caption": [ + "Figure 4. Template Matching and Regression (TMR). A template feature map is cropped from the image feature map and then correlated with the image feature map via channel-wise template matching. The TM feature map and the image feature map are concatenated. For each feature map point, the box regressor predicts the shifting and scaling parameters of the support exemplar's size, and the presence classifier scores the presence map. Both of them consist of a $3 \\times 3$ convolution and a linear projection without any complicated layers." + ], + "image_footnote": [], + "bbox": [ + 98, + 74, + 897, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "DETR [47], SAM-C [42], DAVE [51], PseCo [23], and GeCo [52]. However, similar to the FSOD methods, these models typically generate a prototype by spatially averaging exemplar bounding boxes. This discards the spatial structure of the exemplars, potentially losing important cues and details for accurate counting and detection. In contrast, our pattern-detection method preserves the spatial structure of the exemplar for pattern matching.", + "bbox": [ + 88, + 296, + 480, + 417 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Template matching. Template matching [21, 26, 44-46, 65, 66] has been widely used from the beginning of computer vision and pattern recognition and also adopted in convolution-based neural detectors with additional regression [12, 33]. Given a 2D template, template matching identifies the matching region by sliding-window fashion.", + "bbox": [ + 88, + 420, + 482, + 511 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Repetitive pattern detection. Detecting repetitive patterns is trivial for humans but remains challenging in computer vision. Early research focused on geometrically constrained settings where patterns are nearly regular and aligned [14, 20, 31, 50]. Based on this near-regularity, early methods assumed a global repetitive structure and discovered the repetition [21, 31, 37, 38, 49] in a top-down manner. In contrast, later methods [11, 34, 61] define the smallest repeating unit, called a texel [2, 28, 35] and identify all matching subparts based on the identified texels. In our context, the texel corresponds to the given exemplar.", + "bbox": [ + 88, + 515, + 482, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Few-shot pattern detection", + "text_level": 1, + "bbox": [ + 89, + 694, + 339, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given an input image $I \\in \\mathbb{R}^{H_0 \\times W_0 \\times 3}$ , the goal of few-shot pattern detection is to predict matching patterns with a given set of support exemplars. With an abuse of notations, the model is given the set of support exemplar, $\\mathcal{E} = \\{e_1, \\dots, e_{N_s}\\}$ and aims to predict the corresponding ground-truth bounding boxes of pattern $\\mathcal{B} = \\{b_1, \\dots, b_{N_p}\\}$ , where $N_p$ and $N_s$ denotes the number of ground-truth bounding boxes of the pattern and the number of exemplars (typically referred to as the \"shot\"), respectively. The input exemplars and the output patterns are both represented as bounding boxes parameterized by their center coordinates, width, and height: $b_i, e_i \\in \\mathbb{R}^4$ .", + "bbox": [ + 89, + 718, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Template matching and regression (TMR)", + "text_level": 1, + "bbox": [ + 511, + 295, + 888, + 313 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For clarity, we primarily focus on the one-shot, single-scale pattern detection setting, where a single support exemplar is given and a single-resolution image feature map is used. However, TMR can be effortlessly extended to few-shot and multi-scale scenarios, as described in Sec.4.3 and Sec.9.", + "bbox": [ + 511, + 321, + 903, + 396 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall architecture of TMR is illustrated in Fig. 4. The input image $\\pmb{I}$ is first encoded by a backbone such as ViT [25] to extract a feature map $\\pmb{F} \\in \\mathbb{R}^{H \\times W \\times D}$ . The template feature $\\pmb{T} \\in \\mathbb{R}^{t_h \\times t_w \\times D}$ is then obtained through the template extraction process, which uses RoIAlign [22] to crop a region from the image feature map $\\pmb{F}$ based on the support exemplar's size $(s_h, s_w)$ . In RoIAlign, unlike previous fixed-size pooling methods [47, 52], the model adaptively determines the size of $\\pmb{T}$ to fit the corresponding size on $\\pmb{F}$ by rounding up the size of the exemplars on $\\pmb{F}$ as described in Sec. 8.2. This preserves the spatial alignment between $\\pmb{T}$ and $\\pmb{F}$ with translation. The image feature map $\\pmb{F}$ and template feature $\\pmb{T}$ are correlated by template matching (Sec. 4.1), which outputs the template-matching feature $F_{\\mathrm{TM}}$ . The concatenation of the feature maps $\\pmb{F}$ , $F_{\\mathrm{TM}}$ is fed to the subsequent box prediction module, which consists of a pattern box regressor and a pattern presence classifier. The pattern box regressor predicts the localization bounding boxes: $\\hat{\\pmb{B}}$ , which is parameterized by scaling and shifting factors of the given exemplar's size. The pattern presence classifier predicts the presence score of the pattern: $\\hat{\\pmb{P}}$ . A box proposal is then generated on each feature map point based on the combination of $\\hat{\\pmb{B}}$ and $\\hat{\\pmb{P}}$ (Sec. 4.2). At inference, bounding boxes with low presence scores are removed by Non-Maximum Suppression (NMS). SAM decoder [25] can be optionally applied for box refinement before NMS.", + "bbox": [ + 511, + 397, + 906, + 790 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Template matching (TM)", + "text_level": 1, + "bbox": [ + 511, + 801, + 743, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We are motivated to detect arbitrary patterns given a support exemplar of either an object or not. Non-object patterns often lack high-level semantics yet exhibit low-level structural features, thus, preserving the spatial structure of the support exemplar is crucial. Inspired by traditional tem", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21580", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "plate matching [26, 44, 65], we compute the matching score between the image and the support exemplar to detect the locations of the pattern. Specifically, template matching cross-correlates the feature map $\\mathbf{F}$ and the template feature $\\mathbf{T}$ by centering $\\mathbf{T}$ at each $(x,y)$ position in $\\mathbf{F}$ . The resultant template-matching (TM) feature $\\mathbf{F}_{\\mathrm{TM}}$ is obtained as:", + "bbox": [ + 89, + 90, + 483, + 183 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {F} _ {\\mathrm {T M}} (x, y) = \\frac {1}{t _ {w} t _ {h}} \\sum_ {x ^ {\\prime}, y ^ {\\prime}} \\boldsymbol {F} \\left(x + x ^ {\\prime} - \\lfloor \\frac {t _ {w}}{2} \\rfloor , y + y ^ {\\prime} - \\lfloor \\frac {t _ {h}}{2} \\rfloor\\right) \\boldsymbol {T} \\left(x ^ {\\prime}, y ^ {\\prime}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 189, + 486, + 239 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lfloor \\cdot \\rfloor$ denotes the floor operation used for centering $T$ at each $(x,y)$ , and $(x^{\\prime},y^{\\prime})$ ranges within the template coordinates: $\\in [0,t_w)\\times [0,t_h)$ . Note that Eq. 1 computes the correlation by channel-wise multiplication, resulting in $F_{\\mathrm{TM}}\\in \\mathbb{R}^{H\\times W\\times D}$ . Finally, the TM feature $F_{\\mathrm{TM}}$ is concatenated with the feature map $F$ : $F_{\\mathrm{P}} = [F_{\\mathrm{TM}};F]\\in \\mathbb{R}^{H\\times W\\times 2D}$ . The concatenated feature map $F_{\\mathrm{P}}$ is fed to the subsequent box prediction module.", + "bbox": [ + 89, + 241, + 483, + 361 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Support-conditioned box regression", + "text_level": 1, + "bbox": [ + 89, + 369, + 401, + 386 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our box prediction module consists of a pattern box regressor and a pattern presence classifier following the architecture of an anchor-free detection methods [64].", + "bbox": [ + 89, + 391, + 483, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The pattern box regressor $g_{\\mathrm{B}}$ consists of a $3 \\times 3$ convolutional layer followed by a linear layer and predicts the four localization parameters $(\\Delta x, \\Delta y, \\alpha_w, \\alpha_h)$ for each feature map point. Unlike the methods [47, 52] that directly regress absolute box parameters, our method performs support-conditioned regression. Bounding box parameters are predicted by scaling and shifting the support exemplar's size. This helps to dynamically adjust the exemplar's size to predict a target box size. A predicted bounding box at a feature point $(x, y)$ shifts and scales the support exemplar's size $(s_w, s_h)$ such as:", + "bbox": [ + 89, + 436, + 483, + 604 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left(x + s _ {w} \\Delta x, y + s _ {h} \\Delta y, e ^ {\\alpha_ {w}} s _ {w}, e ^ {\\alpha_ {h}} s _ {h}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 613, + 480, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The pattern presence classifier $g_{\\mathrm{P}}$ consists of a linear layer and predicts presence scores, which represent the confidence of the predicted bounding boxes at each feature map point. The aforementioned procedure of the box regressor $g_{\\mathrm{B}}$ and the presence classifier $g_{\\mathrm{P}}$ is summarized as:", + "bbox": [ + 89, + 638, + 483, + 715 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {B}} = g _ {\\mathrm {B}} \\left(\\boldsymbol {F} _ {\\mathrm {P}}\\right), \\quad \\hat {\\boldsymbol {B}} \\in \\mathbb {R} ^ {H \\times W \\times 4}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 718, + 480, + 736 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {P}} = \\sigma \\left(g _ {\\mathrm {P}} \\left(\\boldsymbol {F} _ {\\mathrm {P}}\\right)\\right), \\quad \\hat {\\boldsymbol {P}} \\in \\mathbb {R} ^ {H \\times W \\times 1}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 168, + 739, + 480, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\sigma$ denotes the sigmoid function.", + "bbox": [ + 94, + 768, + 351, + 782 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Inference", + "text_level": 1, + "bbox": [ + 89, + 787, + 200, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "At inference, we first remove bounding boxes whose presence score is lower than a threshold $\\tau$ to filter out low-confidence predictions. Afterward, we can optionally apply box localization refinement via the SAM decoder [25] as often adopted by the FSCD work [23, 42, 52]. Following [23, 52], we input the predicted box coordinates into the", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SAM prompt encoder to obtain the prompt feature. The obtained prompt feature and the image feature extracted from the SAM backbone are fed to the SAM decoder, which further refines input box coordinates. To obtain the final box prediction, $\\hat{\\mathcal{B}}$ , we apply NMS on the bounding boxes.", + "bbox": [ + 511, + 90, + 905, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "TMR is easily extended for few-shot inference without re-training. When multiple support exemplars are given, we perform the above process for each exemplar individually and then aggregate the results before applying NMS.", + "bbox": [ + 511, + 167, + 905, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.4. Learning objective", + "text_level": 1, + "bbox": [ + 511, + 234, + 694, + 251 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The training loss is composed of $\\mathcal{L}_{\\mathrm{P}}$ , which penalizes the presence score, and $\\mathcal{L}_{\\mathrm{B}}$ , which penalizes the bounding box regression. The presence loss $\\mathcal{L}_{\\mathrm{P}}$ is the binary cross-entropy loss (BCELoss) between the predicted and the ground-truth center points of the pattern bounding boxes. Instead of guiding the presence with a single pixel, we add a margin around each ground-truth center point and define the extended center point set as $\\mathcal{X}_{\\mathrm{P}}$ . We detail this in Sec. 8.3. The ground-truth presence label $P$ at a position $(x,y)$ is set to 1 if the position is in $\\mathcal{X}_{\\mathrm{P}}$ , and 0 otherwise.", + "bbox": [ + 511, + 257, + 906, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The bounding box regression loss $\\mathcal{L}_{\\mathrm{B}}$ is defined as the generalized IoU (gIoU) loss [58] between the predicted and the ground-truth bounding boxes. This loss is defined on the positions where the ground-truth patterns exist. For each center point of the pattern bounding box, the adjacent area of the center within the margin shares the corresponding bounding box parameter $B$ . The two loss functions are defined as the following:", + "bbox": [ + 511, + 409, + 905, + 529 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {P}} = \\sum_ {(x, y)} ^ {\\forall (x, y)} \\operatorname {B C E L o s s} (\\hat {\\boldsymbol {P}} (x, y), \\boldsymbol {P} (x, y)), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 535, + 905, + 580 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {B}} = \\sum_ {(x, y)} ^ {(x, y) \\in \\mathcal {X} _ {\\mathrm {P}}} \\operatorname {g I o U L o s s} (\\hat {\\boldsymbol {B}} (x, y), \\boldsymbol {B} (x, y)). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 585, + 903, + 630 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overall training loss is the sum of them: $\\mathcal{L} = \\mathcal{L}_{\\mathrm{P}} + \\mathcal{L}_{\\mathrm{B}}$", + "bbox": [ + 516, + 638, + 903, + 655 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. Proposed dataset: RPINE", + "text_level": 1, + "bbox": [ + 511, + 664, + 754, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Existing benchmarks (e.g., FSCD-147 [47], FSCD-LVIS [47]) focus on object-level patterns, limiting their use for general pattern detection. Therefore, we introduce a new pattern dataset, RPINE: Repeated Patterns In Everywhere.", + "bbox": [ + 511, + 688, + 905, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We collect images from the repetitive pattern detection literature and annotate them to contain various repetitive patterns in the wild. Images are collected from FSC-147 [47], FSCD-LVIS [47], Countbench [48], Wikiart [59], Frieze [43], and Wallpaper [36]. Tab. 1 shows where RPINE stands among related datasets. The dataset contains 4,362 images, divided into 3,925 and 435 for training and testing, respectively. As shown in Fig. 5, RPINE is annotated with multiple patterns per image. RPINE is a suitable evaluation benchmark for multi-pattern detection within an image.", + "bbox": [ + 511, + 750, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21581", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/3de0463b81516c34eace6e9a91c644bc091eadaac466ae5bafa05efed803ee39.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
datasetpattern tilingrepetitionobject biasmulti-pattern annotations per image
Wallpaper [36]near regular [38]highlow
Pascal [6], COCO [32]-lowhigh✓ (multi-category)
Wikiart [59], Frieze [9], Counting bench [48]arbitrarymidhigh
FSCD-147 [47], FSCD-LVIS [47]arbitraryhighhigh
RPINEarbitraryhighlow to high✓
", + "bbox": [ + 150, + 71, + 844, + 160 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Benchmark dataset comparison of related work. RPINE covers pattern or object repetitions in the real world. We classify 'pattern- tiling' as near-regular following [38], i.e., elements with an almost periodic lattice with minor variations in shape, color or lighting, and as arbitrary otherwise. The 'repetition' is classified based on the average number of repeated same class instances on an image: low if $\\leq 5$ , mid if $\\leq {20}$ ,and high otherwise. The 'object-bias' is high if the dataset annotations correspond to a predefined class set, and low otherwise.", + "bbox": [ + 89, + 164, + 906, + 220 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c9c8dd92228de1ad54d4c11a3f021f428bdc54e24d9c22c59d9bb909f34160e1.jpg", + "image_caption": [ + "Figure 5. GTs (blue) and TMR predictions (green) with different exemplars (red) from RPINE, which is the only dataset containing multiple patterns for each image among FSCD datasets." + ], + "image_footnote": [], + "bbox": [ + 93, + 229, + 478, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the real world, a pattern cannot be rigorously defined. We thus define the following criteria inspired by the definition of the translation symmetry [62] to minimize the subjective variance among different annotators.", + "bbox": [ + 89, + 429, + 482, + 489 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Number of patterns: max 3 different patterns are annotated per image if the image exhibits multiple patterns.", + "- Number of pattern instances: has no upperbound.", + "- Minimum Size: the width and height of a pattern instance must be at least $3\\%$ of the shorter side of the image.", + "- Appearance variance: visually similar patterns with different scales/colors/semantics/rotation angles are annotated as the same pattern.", + "- Reflection: visually similar patterns but reflection symmetric patterns are annotated in different patterns.", + "- Occlusion: if visually similar patterns are occluded from each other, the visible parts are annotated." + ], + "bbox": [ + 89, + 489, + 482, + 671 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We ask annotators to carefully draw bounding boxes on the recognized patterns by the above instructions as consistently as they can. Despite of the instruction details, pattern are not pixel-perfectly defined across annotators. Therefore, we assign three individual annotators per image and include all the annotated patterns as ground truth.", + "bbox": [ + 89, + 672, + 482, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 89, + 777, + 223, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6.1. Dataset and metrics", + "text_level": 1, + "bbox": [ + 89, + 801, + 279, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our model on RPINE as well as on the two standard FSCD benchmarks: FSCD-147 [47] and FSCDLVIS [47]. FSCD-147 contains a total of 6,135 images, with 3,659 for training, 1,286 for validation, and 1,190 for testing. FSCD-LVIS seen-split contains a total of 6,195 im", + "bbox": [ + 89, + 824, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ages, with 4,000 for training, 1,181 for validation, and 1,014 for testing, covering 372 object categories. FSCD-LVIS unseen-split contains a total of 6,201 images, with 3,959 for training and 2,242 for testing where test-time object categories are not observed during training.", + "bbox": [ + 511, + 232, + 903, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following the evaluation protocol [47, 52], we report Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) for counting. For detection, we report average precision with IoU thresholds of 0.5 and 0.75, denoted as AP50 and AP75, respectively, along with the averaged AP over IoU thresholds from 0.5 to 0.95 in increments of 0.05.", + "bbox": [ + 511, + 308, + 905, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6.2. Implementation details", + "text_level": 1, + "bbox": [ + 511, + 407, + 725, + 424 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use the pre-trained SAM-ViT/H [25] of the patch size 16 as the backbone and set it frozen during training, which returns the $64 \\times 64 \\times 256$ feature map. Due to the input patchification encapsulated in the ViT backbones [5], we find the receptive field of the raw feature map too coarse to detect small instances. We thus bilinearly interpolate the feature map resolution from $64 \\times 64$ to $128 \\times 128$ to produce a higher-resolution correlation map, which enables denser predictions and leads to better performance. (Tab. 11)", + "bbox": [ + 511, + 430, + 905, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The channel dimension is expanded to 512 via a learnable linear projection. TMR is trained with the learning rate of $10^{-4}$ with AdamW [39] and the batch size of 16 on four Nvidia RTX 3090 GPUs with 24 GB VRAM for 24 hours. The box presence threshold $\\tau$ for NMS set to 0.4, 0.3 for RPINE and FSCD-147, respectively.", + "bbox": [ + 511, + 566, + 905, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6.3. Comparison with state-of-the-art methods", + "text_level": 1, + "bbox": [ + 511, + 667, + 872, + 683 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "RPINE. To demonstrate the effectiveness of our method in counting and detecting non-object patterns, we evaluate it on RPINE. In Tab. 2, TMR surpasses the previous FSCD methods with a large margin. As the previous FSCD methods rely on prototypes for matching, they tend to struggle with non-object patterns that require an understanding of spatial details rather than semantics. In contrast, TMR effectively detects non-object patterns by incorporating spatial details in the template matching process. Note that RPINE is the only dataset equipped with multiple pattern annotations for each image in FSCD. Figure 5 demonstrates the bounding box predictions with different exemplars, where TMR predicts bounding boxes adaptively to the given exemplars. Figure 7 shows the qualitative com", + "bbox": [ + 511, + 688, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21582", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/5ca54d8ad77d20610ec7f6f8de129de41cd7dd33053da83e1a7675e86daba82e.jpg", + "table_caption": [], + "table_footnote": [ + "Table 2. One-shot pattern counting and detection results on the RPINE dataset. SD denotes box refinement with the SAM decoder. All the models are trained by the official code." + ], + "table_body": "
MethodSDMAE(↓)RMSE(↓)AP(↑)AP50(↑)AP75(↑)
C-DETR [47]9.5821.2413.8832.2010.22
SAM-C [42]✓18.7737.1418.8034.0418.74
PseCo [23]✓48.2088.1623.1844.5421.24
GeCo [52]✓9.5717.0723.3345.9321.19
TMR (ours)8.4519.8733.5964.0530.52
TMR (ours)✓8.3019.4029.6658.9425.41
", + "bbox": [ + 94, + 88, + 482, + 186 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "parisons with other FSCD methods where TMR accurately localize target patterns.", + "bbox": [ + 89, + 248, + 482, + 277 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "FSCD-LVIS and FSCD-147. We compare TMR with existing methods that are dedicated to FSCD under the FSCD setting on FSCD-LVIS and FSCD-147. The results in Tab. 3 demonstrate that TMR significantly outperforms previous state-of-the-art approaches. Notably, when evaluated on the unseen split, where test-time object categories are not observed during training, TMR surpasses prior methods. This suggests that TMR is less biased to object semantics during training potentially because TMR leverages the exemplar's spatial structure that generalizes across different categories. In addition to FSCD-LVIS, Tab. 6 compares methods on FSCD-147, where TMR performs on par with the previous methods on both one-shot and three-shot settings.", + "bbox": [ + 89, + 282, + 482, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.4. Analyses and ablation study", + "text_level": 1, + "bbox": [ + 89, + 491, + 341, + 507 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "TMR learns with less semantic object bias and generalizes well across datasets. We compare TMR and GeCo [52] in the cross-dataset scenarios by evaluating the trained models on different datasets that are unseen during training. As shown in Tab. 4, TMR presents overwhelming performances, showing its strong generalization ability. Specifically, when GeCo is trained on FSCD-147, its performance drops significantly when evaluated on different datasets compared to when tested on FSCD-147 itself. GeCo, like previous FSCD methods, relies on prototypes for both counting and detection. We also suspect this lower generalization ability than ours originates from prototype matching, which is prone to overfitting to the training object semantics. GeCo struggles when evaluated on datasets with different object semantics. In contrast, TMR utilizes the structural information for matching instead of relying on semantic-intensive prototypes and generalizes more effectively on unseen datasets.", + "bbox": [ + 89, + 517, + 482, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SAM decoder is biased to edges. Table 2 shows the negative impact of the optional box refinement using the SAM decoder [25] on RPINE. Figure 6 visualizes two representative examples when the SAM decoder (SD) degrades performance. We observe that SD tends to align box predictions with the nearest edge, which is expected given that SAM is a segmentation model. The edge-sensitive predic", + "bbox": [ + 89, + 794, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/91eaaea80cddccdecc32c8608b54c23679a510d224cb629204f6d50d001f7731.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSeenUnseen
AP(↑)AP50(↑)AP(↑)AP50(↑)
FSDetView-PB [68]2.727.571.032.89
AttRPN-PB [7]4.0811.153.157.87
C-DETR [47]4.9214.493.8511.28
DAVE [51]6.7522.514.1214.16
PseCo [23]22.3742.56--
GeCo [52]--11.4724.49
TMR (ours)27.4948.4822.7139.68
", + "bbox": [ + 521, + 88, + 895, + 227 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5644ee27fbd8fdf6806a7e8f6c1d385ac3001f2d463538429736df1d73f4641a.jpg", + "table_caption": [ + "Table 3. Three-shot counting detection-based methods on the FSCD-LVIS seen and unseen split." + ], + "table_footnote": [], + "table_body": "
TrainTestcross-evalAPAP50
GeCoTMRGeCoTMR
F-147F-14743.4244.4375.0673.83
F-LVISseen✓13.9621.2525.8737.18
RPINE✓19.4726.2138.6952.01
RPINEF-147✓36.9941.3960.3869.19
F-LVISseen✓10.0120.9217.4437.87
RPINE23.3329.6645.9358.94
", + "bbox": [ + 524, + 263, + 890, + 390 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ff1814c24ecaf2a10031653507c56e5cf60b164b830cfab7ef34ac7ae751534b.jpg", + "table_caption": [ + "Table 4. Cross-dataset comparison of GeCo [52] and TMR, where F-147, F-LVIS indicate FSCD-147 and FSCD-LVIS." + ], + "table_footnote": [], + "table_body": "
MethodOne-shotThree-shot
AP(↑)AP50(↑)AP(↑)AP50(↑)
GeCo [52]32.7169.9532.4970.51
TMR (ours)36.0171.1938.5772.61
", + "bbox": [ + 550, + 428, + 870, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 5. Comparison without optional SAM decoder on FSCD147", + "bbox": [ + 511, + 506, + 903, + 520 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tion is particularly harmful for non-typical objects, such as objects that are loosely bounded by edges or patterns with many confusing edges. The SD refinement is seemingly good at edge detection for typical object exemplars with clear boundaries, and this is why the existing FSCD methods [23, 42, 52] benefit from adopting SD for box refinement. However, arbitrary patterns are not necessarily bounded by clear edges. Hence, edge-driven refinement may be even harmful as verified on RPINE. We emphasize that SD is an additional box post-processor, and we optionally add SD to TMR to compare with the SD-based state of the arts. Table 5 compares models by taking out SD, where TMR shows powerful performance.", + "bbox": [ + 511, + 532, + 906, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effectiveness of template matching features. We verify the effectiveness of 2D template matching, which preserves the spatial structure of exemplar bounding boxes for correlation. We experiment with different input feature $F_{\\mathrm{P}}$ to the box regression module. Table 7 compares our final model (e) and its variants. The lower bound model (a) observes zero information on the exemplar. Model (b) shows that correlating with 2D templates is greatly beneficial. The final model (e) concatenates the template-matching and the image feature, providing the correlation and appearance information for pattern matching regression. The models (c,", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21583", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7f32c2d6e1e3f58e74be4164673a2c2b3d7eebd869597c36492231b4bf9fee8a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSAM decoderOne-shotThree-shot
MAE(↓)RMSE(↓)AP(↑)AP50(↑)MAE(↓)RMSE(↓)AP(↑)AP50(↑)
FSDetView-PB [68]----37.54147.0713.4132.99
AttRPN-PB [7]----32.42141.5520.9737.19
C-DETR† [47]16.99125.2219.1447.6316.79123.5622.6650.57
SAM-C† [42]✓33.17141.7735.0956.0227.97131.2427.9949.17
PseCo [23]✓14.86118.6441.6370.8713.05112.8642.9873.33
DAVE [51]11.5486.6219.4655.2710.4574.5126.8162.82
GeCo [52]✓8.1060.1643.1174.317.9154.2843.4275.06
TMR (ours)✓11.6357.4643.1571.5513.7851.8744.4373.83
", + "bbox": [ + 133, + 70, + 859, + 218 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/497e318e60bb72db0a72d6f60209b99821c155be549bcfc9839b6a4783fee8cb.jpg", + "image_caption": [ + "Figure 6. TMR without vs. with SAM decoder [25] (w/o SD vs. w/ SD) for box refinement. SAM decoder tends to align box predictions with the nearest edges. The edge-sensitive prediction is harmful for non-objects, objects that have loose boundaries, or patterns with many confusing edges. The exemplar in the first row loosely contains an arched window, but the SAM decoder snaps the box tightly its edges." + ], + "image_footnote": [], + "bbox": [ + 94, + 253, + 900, + 436 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fed2e8bb0a76fad008461f6d893415c7b3831d87a800018b817177a786b1b1fb.jpg", + "table_caption": [ + "Table 6. Comparison with few-shot detection and counting methods on FSCD-147. The one-shot performance of the models with $\\dagger$ are evaluated using the official code." + ], + "table_footnote": [], + "table_body": "
feature FPRPINEFSCD-147
APAP50APAP50
(a)F11.4424.0220.9555.59
(b)FTM32.5557.4331.9663.03
(c)F ⊕ FTM-cos29.7457.8524.7360.29
(d)F ⊕ FPM20.9447.9628.9166.93
(e)F ⊕ FTM33.5964.0536.0171.19
", + "bbox": [ + 129, + 488, + 442, + 607 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "d) replace the variants of the template matching feature $F_{\\mathrm{TM}}$ with something else. Model (c) replaces the channel-wise correlation of (e) to cosine similarity, i.e., $F_{\\mathrm{TM - cos}}$ , showing that retaining the channel dimension helps. Model (d) first performs average-pooling of the template and uses the prototype matching feature, $F_{\\mathrm{PM}}$ , for correlation. Note that this average-pooled feature is often used in previous FSC methods [51, 52, 70]. The result shows that preserving the spatial structure of exemplar bounding boxes is more effective. The representative failure cases are shown in Fig. 11 in supp. material. The pooled prototype feature loses the geometric layout of the template and struggle to detect patterns where geometric clues are crucial. Comparing (d) and (e) verifies our hypothesis that prototype-based matching relies on the object priors; Model (d) drops more significantly on RPINE, which is less object-centric than FSCD-147.", + "bbox": [ + 89, + 657, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f6bf9b3ea3fb7e22a5b4c2e95dcbf052cb2bdf98a12302d6f362ef2bcdb9aa20.jpg", + "table_caption": [ + "Table 7. Effect of template matching features (Sec. 4.1). $F_{\\mathrm{PM}}$ is the correlation with average-pooled prototype. $F_{\\mathrm{TM - cos}}$ is the cosine similarity of the template and image feature." + ], + "table_footnote": [], + "table_body": "
box regressionRPINEFSCD-147
shiftscaleAPAP50APAP50
(a)(0,0)(0,0)26.2055.2022.8157.30
(b)(Δx, Δy)(eαw, eαh)23.8851.2517.0155.14
(c)(Δx, Δy)(eαw sw, eαsh)31.2959.5235.0870.23
(d)(sw Δx, sh Δy)(eαw sw, eαsh)33.5964.0536.0171.19
", + "bbox": [ + 517, + 488, + 908, + 588 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 8. Effect of box regression methods (Sec. 4.2)", + "bbox": [ + 553, + 590, + 864, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of support-conditioned regression. Table 8 compares different variants for box regression and validates the effectiveness of our support-conditioned box regression method. As shown in (b), regressing bounding boxes without considering the support exemplar's size leads to poor localization. Notably, it even performs worse than (a), which directly uses the exemplar box without performing any regression. However, as shown in (c), incorporating the support exemplar's size when predicting the width and height leads to a significant performance gain. This improvement is further enhanced by using the exemplar's size for shifting, as demonstrated in (d). These results demonstrate the effectiveness of support-conditioned regression, guided by the exemplar's size, in achieving accurate localization.", + "bbox": [ + 511, + 619, + 906, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison on computational complexity. Table 9 compares the complexity of TMR with state-of-the-art FSCD methods, demonstrating its efficiency and effectiveness. Compared to PseCo [23], which introduces a large num", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21584", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/302d43631ad77592c7dd51b860e7557e3433e8266c0d5788ef4bda6ee871a519.jpg", + "image_caption": [ + "Figure 7. Qualitative comparison with the state-of-the-art models on RPINE (the first two images) and FSCD-147 (the last two images). More visualization examples on three datasets are included in the supplementary materials Figs 16-18." + ], + "image_footnote": [], + "bbox": [ + 155, + 74, + 841, + 398 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8af4a580d301ab6d3c4d27c34be563ca11c756e55d003deb12effb2f37b10452.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTrainable paramsTotal paramsFLOPS
PseCo [23]56.99M0.70B5.08T
GeCo [52]7.98M0.65B4.72T
TMR (ours)19.01M0.66B3.04T
", + "bbox": [ + 102, + 439, + 467, + 511 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 9. Comparison on the computational complexity", + "bbox": [ + 122, + 513, + 450, + 529 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ber of trainable parameters and FLOPs, TMR introduces only about 19M trainable parameters, which is detailed in Tab. 10. Although TMR has more trainable parameters than GeCo [52], its total parameter count remains comparable. Thanks to the simple architecture, TMR is significantly efficient in terms of FLOPs (3.04T), which is notably lower than PseCo (5.08T) and GeCo (4.72T). This reduction in computational cost not only enhances training efficiency but also results in faster inference. In contrast, the high FLOPs of PseCo and GeCo contribute to longer inference, making them less practical for real-time applications.", + "bbox": [ + 88, + 540, + 482, + 707 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Real-world application. We evaluate TMR, trained on RPINE, on scanning electron microscope (SEM) images [1], which are widely used in microprocessor inspection. Despite the domain shift, our method performs effectively and demonstrates its potential to generalize to real-world, non-object pattern scenarios (Fig. 8).", + "bbox": [ + 89, + 710, + 482, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion and discussion", + "text_level": 1, + "bbox": [ + 89, + 815, + 336, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have proposed a simple template-matching based method for few-shot pattern detection. We also introduce a new dataset with bounding box annotations that covers various patterns around the world across non-object patterns", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/868c85f38ef62351767db810bbe2fb545d09d18a3d378ae85eaeedaa7d46ebef.jpg", + "image_caption": [ + "Figure 8. Application example of TMR on the SEM dataset [1]." + ], + "image_footnote": [], + "bbox": [ + 516, + 439, + 903, + 517 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to typical objects from various image domains from nature and human-made products.", + "bbox": [ + 511, + 547, + 906, + 577 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We observed that the widely used SAM decoder improves our performance by $7\\%$ AP points on FSCD-147 by leveraging object-level edge priors, but it degrades performance on RPINE, which contains more general patterns. This suggests a promising future direction that combines the powerful pre-trained knowledge of foundation models with detection mechanisms less dependent on object-level edge priors. In parallel, exploring a lightweight, pattern-specific architecture that efficiently captures fine-grained repetitive structures without relying on object semantics could further improve generalization and efficiency.", + "bbox": [ + 511, + 579, + 908, + 746 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work was supported by Samsung Electronics Co., Ltd (IO240508-09825-01) and IITP grants (RS-2022-II220959: Few-Shot Learning of Causal Inference in Vision & Language $(50\\%)$ , RS-2022-II220113: Developing a Sustainable Collaborative Multi-modal Lifelong Learning Framework $(45\\%)$ , RS-2019-II191906: AI Graduate School Program at POSTECH $(5\\%))$ funded by Ministry of Science and ICT, Korea. We also appreciate Manjin Kim's advice on constructing RPINE.", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21585", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rossella Aversa, Mohammad Hadi Modarres, Stefano Cozzini, Regina Ciancio, and Alberto Chiusole. The first annotated set of scanning electron microscopy images for nanoscience. Scientific data, 5(1):1-10, 2018. 8", + "[2] Yunliang Cai and George Baciu. Translation symmetry detection: A repetitive pattern analysis approach. In IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 223-228, 2013. 3", + "[3] Cheng Chen, Wei Wang, John A Ozolek, and Gustavo K Rohde. A flexible and robust approach for segmenting cell nuclei from 2d microscopy images using supervised learning and template matching. Cytometry Part A, 83(5):495-507, 2013. 4", + "[4] Taco S Cohen and Max Welling. Steerable cnns. arXiv preprint arXiv:1612.08498, 2016. 4", + "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5", + "[6] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International Journal of Computer Vision (IJCV), 88:303-338, 2010. 1, 5", + "[7] Qi Fan, Wei Zhuo, Chi-Keung Tang, and Yu-Wing Tai. Few-shot object detection with attention-rpn and multi-relation detector. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4013–4022, 2020. 1, 2, 6, 7", + "[8] Zhibo Fan, Yuchen Ma, Zeming Li, and Jian Sun. Generalized few-shot object detection without forgetting. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4527-4536, 2021. 1, 2", + "[9] Christopher Funk, Seungkyu Lee, Martin R Oswald, Stavros Tsogkas, Wei Shen, Andrea Cohen, Sven Dickinson, and Yanxi Liu. 2017 iccv challenge: Detecting symmetry in the wild. In International Conference on Computer Vision Workshops (ICCVW), pages 1692-1701, 2017. 5", + "[10] Hiroyasu Furukawa, Kyle E Cordova, Michael O'Keeffe, and Omar M Yaghi. The chemistry and applications of metalorganic frameworks. Science, 341(6149):1230444, 2013. 4", + "[11] Jizhou Gao, Yin Hu, Jinze Liu, and Ruigang Yang. Unsupervised learning of high-order structural semantics from images. In Proc. IEEE International Conference on Computer Vision (ICCV), pages 2122-2129. IEEE, 2009. 3", + "[12] Ross Girshick. Fast r-cnn. In Proc. IEEE International Conference on Computer Vision (ICCV), pages 1440-1448, 2015. 1, 3", + "[13] Alice L Givan. Flow cytometry: an introduction. Flow cytometry protocols, pages 1-29, 2011. 4", + "[14] Branko Grünbaum and Geoffrey Colin Shephard. Tilings and patterns. Courier Dover Publications, 1987. 3", + "[15] Karim Guirguis, Johannes Meier, George Eskandar, Matthias Kayser, Bin Yang, and Jürgen Beyerer. Niff: Alleviating" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "forgetting in generalized few-shot object detection via neural instance feature forging. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 24193-24202, 2023. 2", + "[16] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5356-5364, 2019. 1", + "[17] Guangxing Han, Yicheng He, Shiyuan Huang, Jiawei Ma, and Shih-Fu Chang. Query adaptive few-shot object detection with heterogeneous graph convolutional networks. In Proc. IEEE International Conference on Computer Vision (ICCV), pages 3263-3272, 2021. 2", + "[18] Guangxing Han, Shiyuan Huang, Jiawei Ma, Yicheng He, and Shih-Fu Chang. Meta faster r-cnn: Towards accurate few-shot object detection with attentive feature alignment. In Proc. AAAI Conference on Artificial Intelligence (AAAI), pages 780-789, 2022. 1", + "[19] Guangxing Han, Jiawei Ma, Shiyuan Huang, Long Chen, and Shih-Fu Chang. Few-shot object detection with fully cross-transformer. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5321-5330, 2022. 2", + "[20] Junwei Han, Stephen J McKenna, and Ruixuan Wang. Regular texture analysis as statistical model selection. In Proc. European Conference on Computer Vision (ECCV), pages 242-255. Springer, 2008. 3", + "[21] James Hays, Marius Leordeanu, Alexei A Efros, and Yanxi Liu. Discovering texture regularity as a higher-order correspondence problem. In Proc. European Conference on Computer Vision (ECCV), pages 522-535. Springer, 2006. 3", + "[22] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proc. IEEE International Conference on Computer Vision (ICCV), pages 2961-2969, 2017. 1, 3", + "[23] Zhizhong Huang, Mingliang Dai, Yi Zhang, Junping Zhang, and Hongming Shan. Point segment and count: A generalized framework for object counting. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 3, 4, 6, 7, 8", + "[24] Bingyi Kang, Zhuang Liu, Xin Wang, Fisher Yu, Jiashi Feng, and Trevor Darrell. Few-shot object detection via feature reweighting. In Proc. IEEE International Conference on Computer Vision (ICCV), pages 8420-8429, 2019. 2", + "[25] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In Proc. IEEE International Conference on Computer Vision (ICCV), 2023. 3,4,5,6,7,2", + "[26] Simon Korman, Daniel Reichman, Gilad Tsur, and Shai Avidan. Fast-match: Fast affine template matching. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2331-2338, 2013. 3, 4", + "[27] Minji Lee, Hyunju Noh, and Youngsuk Kim. Diamidocarbene-derived palladium and nickel-sulfur clusters. Chemical Communications, 60(94):13867-13870, 2024. 4" + ], + "bbox": [ + 517, + 93, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21586", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Thomas Leung and Jitendra Malik. Detecting, localizing and grouping repeated scene elements from an image. In Proc. European Conference on Computer Vision (ECCV). Springer, 1996. 3", + "[29] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection, 2022. 3", + "[30] Hui Lin, Xiaopeng Hong, and Yabin Wang. Object counting: You only need to look at one. arXiv preprint arXiv:2112.05993, 2021. 2", + "[31] Hsin-Chih Lin, Ling-Ling Wang, and Shi-Nine Yang. Extracting periodicity of a regular texture based on autocorrelation functions. Pattern recognition letters, 18(5):433-443, 1997. 3", + "[32] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proc. European Conference on Computer Vision (ECCV), pages 740-755. Springer, 2014. 1, 5", + "[33] Tsung-Yi Lin, Piotr Dólár, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2117-2125, 2017. 1, 3", + "[34] Wen-Chieh Lin and Yanxi Liu. Tracking dynamic near-regular texture under occlusion and rapid movements. In Proc. European Conference on Computer Vision (ECCV), pages 44–55. Springer, 2006. 3", + "[35] Jingchen Liu and Yanxi Liu. Grasp recurring patterns from a single view. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2003-2010, 2013. 3", + "[36] Jingchen Liu, George Slota, Gang Zheng, Zhaohui Wu, Minwoo Park, Seungkyu Lee, Ingmar Rauschert, and Yanxi Liu. Symmetry detection from realworld images competition 2013: Summary and results. In IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 200-205, 2013. 4, 5", + "[37] Siying Liu, Tian-Tsong Ng, Kalyan Sunkavalli, Minh N Do, Eli Shechtman, and Nathan Carr. Patchmatch-based automatic lattice detection for near-regular textures. In Proc. IEEE International Conference on Computer Vision (ICCV), pages 181–189, 2015. 3", + "[38] Yanxi Liu, Wen-Chieh Lin, and James Hays. Near-regular texture analysis and manipulation. ACM Transactions on Graphics (TOG), 23(3):368-376, 2004. 3, 5", + "[39] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019. 5", + "[40] Erika Lu, Weidi Xie, and Andrew Zisserman. Class-agnostic counting. In Asian Conference on Computer Vision (ACCV), pages 669–684. Springer, 2019. 2", + "[41] Bronwyn A Lucas, Benjamin A Himes, Liang Xue, Timothy Grant, Julia Mahamid, and Nikolaus Grigorieff. Locating macromolecular assemblies in cells by 2d template matching with cistem. _Elife_, 10:e68946, 2021. 4", + "[42] Zhiheng Ma, Xiaopeng Hong, and Qinnan Shangguan. Can sam count anything? an empirical study on sam counting, 2023. 3, 4, 6, 7" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Massimiliano Mancini, Elisa Ricci, Barbara Caputo, and Samuel Rota Bulo. Adding new tasks to a single network with weight transformations using binary masks. In European Conference on Computer Vision Workshops (ECCVW), pages 0-0, 2018. 4", + "[44] Giovanni Marola. Using symmetry for detecting and locating objects in a picture. Computer Vision, Graphics, and Image Processing, 46(2):179-195, 1989. 3, 4", + "[45] Jiri Matas, Ondrej Chum, Martin Urban, and Tomás Pajdla. Robust wide-baseline stereo from maximally stable extremal regions. Image and vision computing, 22(10):761-767, 2004.", + "[46] Anuj Mohan, Constantine Papageorgiou, and Tomaso Poggio. Example-based object detection in images by components. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 23(4):349-361, 2001. 3", + "[47] Thanh Nguyen, Chau Pham, Khoi Nguyen, and Minh Hoai. Few-shot object counting and detection. In Proc. European Conference on Computer Vision (ECCV). Springer, 2022. 1, 2, 3, 4, 5, 6, 7", + "[48] Roni Paiss, Ariel Ephrat, Omer Tov, Shiran Zada, Inbar Mosseri, Michal Irani, and Tali Dekel. Teaching clip to count to ten. In Proc. IEEE International Conference on Computer Vision (ICCV), pages 3170-3180, 2023. 4, 5", + "[49] Minwoo Park, Kyle Brocklehurst, Robert T Collins, and Yanxi Liu. Deformed lattice detection in real-world images using mean-shift belief propagation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 31(10): 1804-1816, 2009. 3", + "[50] Mark Pauly, Niloy J Mitra, Johannes Wallner, Helmut Pottmann, and Leonidas J Guibas. Discovering structural regularity in 3d geometry. In ACM SIGGRAPH, pages 1-11, 2008. 3", + "[51] Jer Pelhan, Vitjan Zavrtanik, Matej Kristan, et al. Davea detect-and-verify paradigm for low-shot counting. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 23293-23302, 2024. 3, 6, 7", + "[52] Jer Pelhan, Alan Lukezic, Vitjan Zavrtanik, and Matej Kristan. A novel unified architecture for low-shot counting by detection and segmentation. Advances in Neural Information Processing Systems (NeurIPS), 2025. 1, 2, 3, 4, 5, 6, 7, 8", + "[53] Julien Picot, Coralie L Guerin, Caroline Le Van Kim, and Chantal M Boulanger. Flow cytometry: retrospective, fundamentals and recent instrumentation. Cytotechnology, 64: 109-130, 2012. 4", + "[54] Muhammad Farrukh Qureshi, Zohaib Mushtaq, Muhammad Zia Ur Rehman, and Ernest Nlandu Kamavuako. E2cnn: An efficient concatenated cnn for classification of surface emg extracted from upper limb. IEEE Sensors Journal, 23(8): 8989-8996, 2023. 4", + "[55] Viresh Ranjan and Minh Hoai. Vicinal counting networks. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4221-4230, 2022. 2", + "[56] Viresh Ranjan, Udbhav Sharma, Thu Nguyen, and Minh Hoai. Learning to count everything. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3394-3403, 2021. 2, 1" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21587", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[57] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 779-788, 2016. 1", + "[58] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 658-666, 2019. 4", + "[59] Babak Saleh and Ahmed Elgammal. Large-scale classification of fine-art paintings: Learning the right metric on the right feature. arXiv preprint arXiv:1505.00855, 2015. 4, 5", + "[60] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 1", + "[61] Luciano Spinello, Rudolph Triebel, Dizan Vasquez, Kai O Arras, and Roland Siegwart. Exploiting repetitive object patterns for model compression and completion. In Proc. European Conference on Computer Vision (ECCV), pages 296-309. Springer, 2010. 3", + "[62] V.J. Stenger. Timeless Reality: Symmetry, Simplicity, and Multiple Universes. Prometheus, 2009. 5", + "[63] Bo Sun, Banghuai Li, Shengcai Cai, Ye Yuan, and Chi Zhang. Fsce: Few-shot object detection via contrastive proposal encoding. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 7352-7362, 2021, 2", + "[64] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In Proc. IEEE International Conference on Computer Vision (ICCV), pages 9627-9636, 2019. 1, 4", + "[65] Du-Ming Tsai and Cheng-Huei Chiang. Rotation-invariant pattern matching using wavelet decomposition. Pattern Recognition Letters, 23(1-3):191-201, 2002. 3, 4", + "[66] Andreas Turina, Tinne Tuytelaars, and Luc Van Gool. Efficient grouping under perspective skew. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages I-I. IEEE, 2001. 3", + "[67] Xin Wang, Thomas Huang, Joseph Gonzalez, Trevor Darrell, and Fisher Yu. Frustratingly simple few-shot object detection. In Proc. International Conference on Machine Learning (ICML), pages 9919-9928. PMLR, 2020. 2", + "[68] Yang Xiao, Vincent Lepetit, and Renaud Marlet. Few-shot object detection and viewpoint estimation for objects in the wild. TPAMI, 45(3):3090-3106, 2022. 1, 2, 6, 7", + "[69] Xiaopeng Yan, Ziliang Chen, Anni Xu, Xiaoxi Wang, Xiaodan Liang, and Liang Lin. Meta r-cnn: Towards general solver for instance-level low-shot learning. In Proc. IEEE International Conference on Computer Vision (ICCV), pages 9577-9586, 2019. 1, 2", + "[70] Zhiyuan You, Kai Yang, Wenhan Luo, Xin Lu, Lei Cui, and Xinyi Le. Iterative correlation-based feature refinement for few-shot counting. arXiv preprint arXiv:2201.08959, 4, 2022. 2, 7", + "[71] Xinyu Zhang, Yuhan Liu, Yuting Wang, and Abdeslam Boularias. Detect everything with few examples. arXiv preprint arXiv:2309.12969, 2023. 2" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "21588", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Few-Shot Pattern Detection via Template Matching and Regression/9a1fc29b-9538-4897-943a-30b2211079d1_model.json b/2025/Few-Shot Pattern Detection via Template Matching and Regression/9a1fc29b-9538-4897-943a-30b2211079d1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7d0ff28542e81c19d82716efaa87c30e9b99e80d --- /dev/null +++ b/2025/Few-Shot Pattern Detection via Template Matching and Regression/9a1fc29b-9538-4897-943a-30b2211079d1_model.json @@ -0,0 +1,2477 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.16, + 0.131, + 0.838, + 0.153 + ], + "angle": 0, + "content": "Few-Shot Pattern Detection via Template Matching and Regression" + }, + { + "type": "text", + "bbox": [ + 0.157, + 0.182, + 0.836, + 0.199 + ], + "angle": 0, + "content": "Eunchan Jo Dahiyun Kang Sanghyun Kim Yunseon Choi Minsu Cho" + }, + { + "type": "text", + "bbox": [ + 0.208, + 0.205, + 0.79, + 0.223 + ], + "angle": 0, + "content": "Pohang University of Science and Technology (POSTECH), South Korea" + }, + { + "type": "text", + "bbox": [ + 0.318, + 0.225, + 0.676, + 0.239 + ], + "angle": 0, + "content": "https://cvlab.postech.ac.kr/research/TMR" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.274, + 0.327, + 0.29 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.307, + 0.486, + 0.626 + ], + "angle": 0, + "content": "We address the problem of few-shot pattern detection, which aims to detect all instances of a given pattern, typically represented by a few exemplars, from an input image. Although similar problems have been studied in few-shot object counting and detection (FSCD), previous methods and their benchmarks have narrowed patterns of interest to object categories and often fail to localize non-object patterns. In this work, we propose a simple yet effective detector based on template matching and regression, dubbed TMR. While previous FSCD methods typically represent target exemplars as spatially collapsed prototypes and lose structural information, we revisit classic template matching and regression. It effectively preserves and leverages the spatial layout of exemplars through a minimalistic structure with a small number of learnable convolutional or projection layers on top of a frozen backbone. We also introduce a new dataset, dubbed RPINE, which covers a wider range of patterns than existing object-centric datasets. Our method outperforms the state-of-the-art methods on the three benchmarks, RPINE, FSCD-147, and FSCD-LVIS, and demonstrates strong generalization in cross-dataset evaluation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.648, + 0.222, + 0.664 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.485, + 0.901 + ], + "angle": 0, + "content": "Few-shot detection aims to identify target patterns with minimal labeled examples. While significant progress has been made in few-shot object detection [7, 8, 18, 68, 69], most existing methods remain object-centric, focusing primarily on identifying object-level patterns with relatively clear boundaries. However, many real-world applications require detecting arbitrary target patterns that extend beyond objects to include structural, geometric, or abstract patterns across diverse visual data. Despite recent progress based on deep neural networks, current methods still fall short in addressing these broader pattern detection tasks. Furthermore, the object-centric design of conventional few-shot detectors may lead to performance degradation when the target object lacks clear boundaries or when occlusion and deformation cause its boundaries to become indistinct." + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.272, + 0.701, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.554, + 0.395, + 0.672, + 0.408 + ], + "angle": 0, + "content": "one-shot exemplars" + }, + { + "type": "image", + "bbox": [ + 0.717, + 0.272, + 0.892, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.715, + 0.395, + 0.894, + 0.408 + ], + "angle": 0, + "content": "ground-truth pattern detection" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.414, + 0.908, + 0.47 + ], + "angle": 0, + "content": "Figure 1. Few-shot pattern detection. Given a few exemplar for each target pattern (left), the task is to detect all matching instances of each pattern (right). This example includes non-object patterns (e.g., green and blue) as well as object patterns (e.g., red)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.489, + 0.909, + 0.732 + ], + "angle": 0, + "content": "The task of few-shot pattern detection is illustrated in Fig. 1. Recent related research topics for few-shot detection include few-shot counting and detection [47], and few-shot object detection [7]. Both aim to reduce the annotation cost of object detection [12, 33, 57, 64] by learning to detect all instances of given support exemplars. Consequently, many of these methods are heavily biased on object-centric benchmarks prior [6, 16, 32, 47]. In addition, many recent approaches [7, 18, 23, 52, 68, 69] represent the support exemplars as the spatially pooled vector, often named as a prototype [60]. While this pooling strategy is effective for detecting objects, it collapses the geometric properties, such as the shape and structure of the support exemplars. As a consequence, these methods tend to underperform when detecting non-object geometric patterns such as object parts or shape-intensive elements as shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.911, + 0.903 + ], + "angle": 0, + "content": "In this work, we revisit the classic template matching strategy and propose a simple yet effective few-shot detector for arbitrary patterns. The proposed method, dubbed template matching and regression (TMR), is designed to be aware of the structure and shape of given exemplars. Given an input image, TMR first extracts a feature map using a backbone network. It then crops a template feature from the support exemplar's bounding box using a template extraction technique based on RoIAlign [22]. This template is correlated with the image feature map to produce a template matching feature map. Using this correlation map, the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21578" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.07, + 0.9, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.277, + 0.908, + 0.335 + ], + "angle": 0, + "content": "Figure 2. Annotation examples of FSCD-147 [47] and RPINE. FSCD-147 is annotated with the repetitive object-level patterns but disregards the repetition of non-object patterns such as the egg tray pattern under the eggs in the first image. RPINE is annotated with arbitrarily noticed repeated patterns which include non-object patterns and nameless parts of objects. Plus, FSCD-147 is annotated with a single pattern, while RPINE is annotated with all existing repetitions (marked with different colors) recognized by three different annotators." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.342, + 0.188, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.189, + 0.342, + 0.277, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.34, + 0.368, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.34, + 0.478, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.439, + 0.483, + 0.468 + ], + "angle": 0, + "content": "Figure 3. Few-shot detectors trained with strong object prior are often biased to objects instances and struggle to detect non-objects." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.477, + 0.483, + 0.583 + ], + "angle": 0, + "content": "model learns bounding box regression parameters to rectify the support exemplar's box size adaptively. This process, termed support-conditioned regression, enables the model to handle support exemplars of varying sizes more effectively. Notably, TMR consists only of a few \\(3 \\times 3\\) and linear projections without any complicated modules such as cross-attention, commonly used in prior work [47, 52]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.483, + 0.764 + ], + "angle": 0, + "content": "Although TMR is designed for general pattern detection, existing benchmarks (e.g., FSCD-147 [47], FSCD-LVIS [47]) mainly target object-level patterns, limiting comprehensive evaluation. To address this, we introduce a new dataset, Repeated Patterns IN Everywhere (RPINE), which covers diverse repeated patterns in the real world. RPINE contains images with varying degrees of objectness, from well-defined object-level patterns to non-object patterns, all annotated with bounding boxes via crowd-sourcing. Compared to FSCD datasets, RPINE provides broader coverage, including both non-object patterns and nameless parts of objects, as illustrated in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.87 + ], + "angle": 0, + "content": "TMR demonstrates strong performance in detecting repeated patterns, not only on RPINE but also on the FSCD benchmarks, FSCD-147 and FSCD-LVIS [47]. In particular, TMR is especially effective on RPINE, includes diverse patterns with minimal object priors. Notably, our simple architecture contributes to improved generalization across datasets. Our contribution is summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- We generalize the few-shot object counting and detection to a pattern detection task that does not assume objectness" + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.342, + 0.799, + 0.357 + ], + "angle": 0, + "content": "in either the target patterns or exemplars." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.358, + 0.905, + 0.401 + ], + "angle": 0, + "content": "- We present a simple yet effective pattern detector by refining template matching, which efficiently detects coherent patterns guided by exemplars." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.403, + 0.905, + 0.447 + ], + "angle": 0, + "content": "- We introduce a new densely annotated dataset, RPINE, which covers diverse repetitive patterns in the real world, ranging from object-level patterns to non-object patterns." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.448, + 0.905, + 0.492 + ], + "angle": 0, + "content": "- TMR not only outperforms the state-of-the-art FSCD models on RPINE and FSCD-LVIS but also achieves strong cross-dataset generalization." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.358, + 0.905, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.506, + 0.652, + 0.523 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.907, + 0.793 + ], + "angle": 0, + "content": "Few-shot object detection (FSOD) aims to detect objects of novel classes using only a few support images of novel classes. Existing methods can be roughly categorized into two groups: finetuning based [8, 15, 63, 67, 68] and meta-learning based [7, 17-19, 24, 69, 71] approaches. Despite significant progress, finetuning methods require retraining for every new classes. For pattern detection, it is not desired to fine-tune each arbitrary pattern with indistinct object categorization. In contrast, meta-learning methods avoid class-specific fine-tuning and typically construct prototypes from support images for classifying the bounding boxes of the query image. These methods often apply global average pooling to obtain prototypes from each support exemplar, which collapses the spatial structure of the exemplar. While such spatially collapsed prototypes may effectively capture object-level bounding boxes, they struggle with non-object patterns that exhibit complex spatial structures." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Few-shot counting (FSC) aims to count objects given a few exemplars (typically 1 to 3) within an image. Previous methods [30, 40, 55, 56, 70] have tackled this task using density-map regression without detecting the individual bounding boxes. Towards localization-based counting, few-shot counting and detection (FSCD) is proposed to combine FSC and detection. Such models include Counting-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21579" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.075, + 0.898, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.229, + 0.908, + 0.288 + ], + "angle": 0, + "content": "Figure 4. Template Matching and Regression (TMR). A template feature map is cropped from the image feature map and then correlated with the image feature map via channel-wise template matching. The TM feature map and the image feature map are concatenated. For each feature map point, the box regressor predicts the shifting and scaling parameters of the support exemplar's size, and the presence classifier scores the presence map. Both of them consist of a \\(3 \\times 3\\) convolution and a linear projection without any complicated layers." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.297, + 0.482, + 0.418 + ], + "angle": 0, + "content": "DETR [47], SAM-C [42], DAVE [51], PseCo [23], and GeCo [52]. However, similar to the FSOD methods, these models typically generate a prototype by spatially averaging exemplar bounding boxes. This discards the spatial structure of the exemplars, potentially losing important cues and details for accurate counting and detection. In contrast, our pattern-detection method preserves the spatial structure of the exemplar for pattern matching." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.421, + 0.483, + 0.512 + ], + "angle": 0, + "content": "Template matching. Template matching [21, 26, 44-46, 65, 66] has been widely used from the beginning of computer vision and pattern recognition and also adopted in convolution-based neural detectors with additional regression [12, 33]. Given a 2D template, template matching identifies the matching region by sliding-window fashion." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.516, + 0.483, + 0.683 + ], + "angle": 0, + "content": "Repetitive pattern detection. Detecting repetitive patterns is trivial for humans but remains challenging in computer vision. Early research focused on geometrically constrained settings where patterns are nearly regular and aligned [14, 20, 31, 50]. Based on this near-regularity, early methods assumed a global repetitive structure and discovered the repetition [21, 31, 37, 38, 49] in a top-down manner. In contrast, later methods [11, 34, 61] define the smallest repeating unit, called a texel [2, 28, 35] and identify all matching subparts based on the identified texels. In our context, the texel corresponds to the given exemplar." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.695, + 0.34, + 0.712 + ], + "angle": 0, + "content": "3. Few-shot pattern detection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Given an input image \\(I \\in \\mathbb{R}^{H_0 \\times W_0 \\times 3}\\), the goal of few-shot pattern detection is to predict matching patterns with a given set of support exemplars. With an abuse of notations, the model is given the set of support exemplar, \\(\\mathcal{E} = \\{e_1, \\dots, e_{N_s}\\}\\) and aims to predict the corresponding ground-truth bounding boxes of pattern \\(\\mathcal{B} = \\{b_1, \\dots, b_{N_p}\\}\\), where \\(N_p\\) and \\(N_s\\) denotes the number of ground-truth bounding boxes of the pattern and the number of exemplars (typically referred to as the \"shot\"), respectively. The input exemplars and the output patterns are both represented as bounding boxes parameterized by their center coordinates, width, and height: \\(b_i, e_i \\in \\mathbb{R}^4\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.296, + 0.89, + 0.314 + ], + "angle": 0, + "content": "4. Template matching and regression (TMR)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.322, + 0.905, + 0.397 + ], + "angle": 0, + "content": "For clarity, we primarily focus on the one-shot, single-scale pattern detection setting, where a single support exemplar is given and a single-resolution image feature map is used. However, TMR can be effortlessly extended to few-shot and multi-scale scenarios, as described in Sec.4.3 and Sec.9." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.398, + 0.907, + 0.791 + ], + "angle": 0, + "content": "The overall architecture of TMR is illustrated in Fig. 4. The input image \\(\\pmb{I}\\) is first encoded by a backbone such as ViT [25] to extract a feature map \\(\\pmb{F} \\in \\mathbb{R}^{H \\times W \\times D}\\). The template feature \\(\\pmb{T} \\in \\mathbb{R}^{t_h \\times t_w \\times D}\\) is then obtained through the template extraction process, which uses RoIAlign [22] to crop a region from the image feature map \\(\\pmb{F}\\) based on the support exemplar's size \\((s_h, s_w)\\). In RoIAlign, unlike previous fixed-size pooling methods [47, 52], the model adaptively determines the size of \\(\\pmb{T}\\) to fit the corresponding size on \\(\\pmb{F}\\) by rounding up the size of the exemplars on \\(\\pmb{F}\\) as described in Sec. 8.2. This preserves the spatial alignment between \\(\\pmb{T}\\) and \\(\\pmb{F}\\) with translation. The image feature map \\(\\pmb{F}\\) and template feature \\(\\pmb{T}\\) are correlated by template matching (Sec. 4.1), which outputs the template-matching feature \\(F_{\\mathrm{TM}}\\). The concatenation of the feature maps \\(\\pmb{F}\\), \\(F_{\\mathrm{TM}}\\) is fed to the subsequent box prediction module, which consists of a pattern box regressor and a pattern presence classifier. The pattern box regressor predicts the localization bounding boxes: \\(\\hat{\\pmb{B}}\\), which is parameterized by scaling and shifting factors of the given exemplar's size. The pattern presence classifier predicts the presence score of the pattern: \\(\\hat{\\pmb{P}}\\). A box proposal is then generated on each feature map point based on the combination of \\(\\hat{\\pmb{B}}\\) and \\(\\hat{\\pmb{P}}\\) (Sec. 4.2). At inference, bounding boxes with low presence scores are removed by Non-Maximum Suppression (NMS). SAM decoder [25] can be optionally applied for box refinement before NMS." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.802, + 0.744, + 0.818 + ], + "angle": 0, + "content": "4.1. Template matching (TM)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We are motivated to detect arbitrary patterns given a support exemplar of either an object or not. Non-object patterns often lack high-level semantics yet exhibit low-level structural features, thus, preserving the spatial structure of the support exemplar is crucial. Inspired by traditional tem" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21580" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.184 + ], + "angle": 0, + "content": "plate matching [26, 44, 65], we compute the matching score between the image and the support exemplar to detect the locations of the pattern. Specifically, template matching cross-correlates the feature map \\( \\mathbf{F} \\) and the template feature \\( \\mathbf{T} \\) by centering \\( \\mathbf{T} \\) at each \\( (x,y) \\) position in \\( \\mathbf{F} \\). The resultant template-matching (TM) feature \\( \\mathbf{F}_{\\mathrm{TM}} \\) is obtained as:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.19, + 0.488, + 0.24 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {F} _ {\\mathrm {T M}} (x, y) = \\frac {1}{t _ {w} t _ {h}} \\sum_ {x ^ {\\prime}, y ^ {\\prime}} \\boldsymbol {F} \\left(x + x ^ {\\prime} - \\lfloor \\frac {t _ {w}}{2} \\rfloor , y + y ^ {\\prime} - \\lfloor \\frac {t _ {h}}{2} \\rfloor\\right) \\boldsymbol {T} \\left(x ^ {\\prime}, y ^ {\\prime}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.242, + 0.485, + 0.362 + ], + "angle": 0, + "content": "where \\(\\lfloor \\cdot \\rfloor\\) denotes the floor operation used for centering \\(T\\) at each \\((x,y)\\), and \\((x^{\\prime},y^{\\prime})\\) ranges within the template coordinates: \\(\\in [0,t_w)\\times [0,t_h)\\). Note that Eq. 1 computes the correlation by channel-wise multiplication, resulting in \\(F_{\\mathrm{TM}}\\in \\mathbb{R}^{H\\times W\\times D}\\). Finally, the TM feature \\(F_{\\mathrm{TM}}\\) is concatenated with the feature map \\(F\\): \\(F_{\\mathrm{P}} = [F_{\\mathrm{TM}};F]\\in \\mathbb{R}^{H\\times W\\times 2D}\\). The concatenated feature map \\(F_{\\mathrm{P}}\\) is fed to the subsequent box prediction module." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.371, + 0.402, + 0.387 + ], + "angle": 0, + "content": "4.2. Support-conditioned box regression" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.392, + 0.484, + 0.437 + ], + "angle": 0, + "content": "Our box prediction module consists of a pattern box regressor and a pattern presence classifier following the architecture of an anchor-free detection methods [64]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.438, + 0.484, + 0.605 + ], + "angle": 0, + "content": "The pattern box regressor \\( g_{\\mathrm{B}} \\) consists of a \\( 3 \\times 3 \\) convolutional layer followed by a linear layer and predicts the four localization parameters \\( (\\Delta x, \\Delta y, \\alpha_w, \\alpha_h) \\) for each feature map point. Unlike the methods [47, 52] that directly regress absolute box parameters, our method performs support-conditioned regression. Bounding box parameters are predicted by scaling and shifting the support exemplar's size. This helps to dynamically adjust the exemplar's size to predict a target box size. A predicted bounding box at a feature point \\( (x, y) \\) shifts and scales the support exemplar's size \\( (s_w, s_h) \\) such as:" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.614, + 0.482, + 0.63 + ], + "angle": 0, + "content": "\\[\n\\left(x + s _ {w} \\Delta x, y + s _ {h} \\Delta y, e ^ {\\alpha_ {w}} s _ {w}, e ^ {\\alpha_ {h}} s _ {h}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.64, + 0.484, + 0.716 + ], + "angle": 0, + "content": "The pattern presence classifier \\( g_{\\mathrm{P}} \\) consists of a linear layer and predicts presence scores, which represent the confidence of the predicted bounding boxes at each feature map point. The aforementioned procedure of the box regressor \\( g_{\\mathrm{B}} \\) and the presence classifier \\( g_{\\mathrm{P}} \\) is summarized as:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.719, + 0.482, + 0.737 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {B}} = g _ {\\mathrm {B}} \\left(\\boldsymbol {F} _ {\\mathrm {P}}\\right), \\quad \\hat {\\boldsymbol {B}} \\in \\mathbb {R} ^ {H \\times W \\times 4}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.169, + 0.741, + 0.482, + 0.759 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {P}} = \\sigma \\left(g _ {\\mathrm {P}} \\left(\\boldsymbol {F} _ {\\mathrm {P}}\\right)\\right), \\quad \\hat {\\boldsymbol {P}} \\in \\mathbb {R} ^ {H \\times W \\times 1}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.769, + 0.352, + 0.784 + ], + "angle": 0, + "content": "where \\(\\sigma\\) denotes the sigmoid function." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.789, + 0.201, + 0.803 + ], + "angle": 0, + "content": "4.3. Inference" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "At inference, we first remove bounding boxes whose presence score is lower than a threshold \\(\\tau\\) to filter out low-confidence predictions. Afterward, we can optionally apply box localization refinement via the SAM decoder [25] as often adopted by the FSCD work [23, 42, 52]. Following [23, 52], we input the predicted box coordinates into the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.168 + ], + "angle": 0, + "content": "SAM prompt encoder to obtain the prompt feature. The obtained prompt feature and the image feature extracted from the SAM backbone are fed to the SAM decoder, which further refines input box coordinates. To obtain the final box prediction, \\(\\hat{\\mathcal{B}}\\), we apply NMS on the bounding boxes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.168, + 0.906, + 0.228 + ], + "angle": 0, + "content": "TMR is easily extended for few-shot inference without re-training. When multiple support exemplars are given, we perform the above process for each exemplar individually and then aggregate the results before applying NMS." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.236, + 0.695, + 0.252 + ], + "angle": 0, + "content": "4.4. Learning objective" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.258, + 0.907, + 0.409 + ], + "angle": 0, + "content": "The training loss is composed of \\(\\mathcal{L}_{\\mathrm{P}}\\), which penalizes the presence score, and \\(\\mathcal{L}_{\\mathrm{B}}\\), which penalizes the bounding box regression. The presence loss \\(\\mathcal{L}_{\\mathrm{P}}\\) is the binary cross-entropy loss (BCELoss) between the predicted and the ground-truth center points of the pattern bounding boxes. Instead of guiding the presence with a single pixel, we add a margin around each ground-truth center point and define the extended center point set as \\(\\mathcal{X}_{\\mathrm{P}}\\). We detail this in Sec. 8.3. The ground-truth presence label \\(P\\) at a position \\((x,y)\\) is set to 1 if the position is in \\(\\mathcal{X}_{\\mathrm{P}}\\), and 0 otherwise." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.906, + 0.53 + ], + "angle": 0, + "content": "The bounding box regression loss \\(\\mathcal{L}_{\\mathrm{B}}\\) is defined as the generalized IoU (gIoU) loss [58] between the predicted and the ground-truth bounding boxes. This loss is defined on the positions where the ground-truth patterns exist. For each center point of the pattern bounding box, the adjacent area of the center within the margin shares the corresponding bounding box parameter \\(B\\). The two loss functions are defined as the following:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.536, + 0.906, + 0.582 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {P}} = \\sum_ {(x, y)} ^ {\\forall (x, y)} \\operatorname {B C E L o s s} (\\hat {\\boldsymbol {P}} (x, y), \\boldsymbol {P} (x, y)), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.586, + 0.905, + 0.631 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {B}} = \\sum_ {(x, y)} ^ {(x, y) \\in \\mathcal {X} _ {\\mathrm {P}}} \\operatorname {g I o U L o s s} (\\hat {\\boldsymbol {B}} (x, y), \\boldsymbol {B} (x, y)). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.64, + 0.904, + 0.656 + ], + "angle": 0, + "content": "The overall training loss is the sum of them: \\(\\mathcal{L} = \\mathcal{L}_{\\mathrm{P}} + \\mathcal{L}_{\\mathrm{B}}\\)" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.665, + 0.756, + 0.681 + ], + "angle": 0, + "content": "5. Proposed dataset: RPINE" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.906, + 0.75 + ], + "angle": 0, + "content": "Existing benchmarks (e.g., FSCD-147 [47], FSCD-LVIS [47]) focus on object-level patterns, limiting their use for general pattern detection. Therefore, we introduce a new pattern dataset, RPINE: Repeated Patterns In Everywhere." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.751, + 0.907, + 0.902 + ], + "angle": 0, + "content": "We collect images from the repetitive pattern detection literature and annotate them to contain various repetitive patterns in the wild. Images are collected from FSC-147 [47], FSCD-LVIS [47], Countbench [48], Wikiart [59], Frieze [43], and Wallpaper [36]. Tab. 1 shows where RPINE stands among related datasets. The dataset contains 4,362 images, divided into 3,925 and 435 for training and testing, respectively. As shown in Fig. 5, RPINE is annotated with multiple patterns per image. RPINE is a suitable evaluation benchmark for multi-pattern detection within an image." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21581" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.151, + 0.072, + 0.845, + 0.161 + ], + "angle": 0, + "content": "
datasetpattern tilingrepetitionobject biasmulti-pattern annotations per image
Wallpaper [36]near regular [38]highlow
Pascal [6], COCO [32]-lowhigh✓ (multi-category)
Wikiart [59], Frieze [9], Counting bench [48]arbitrarymidhigh
FSCD-147 [47], FSCD-LVIS [47]arbitraryhighhigh
RPINEarbitraryhighlow to high✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.165, + 0.907, + 0.222 + ], + "angle": 0, + "content": "Table 1. Benchmark dataset comparison of related work. RPINE covers pattern or object repetitions in the real world. We classify 'pattern- tiling' as near-regular following [38], i.e., elements with an almost periodic lattice with minor variations in shape, color or lighting, and as arbitrary otherwise. The 'repetition' is classified based on the average number of repeated same class instances on an image: low if \\( \\leq 5 \\) , mid if \\( \\leq {20} \\) ,and high otherwise. The 'object-bias' is high if the dataset annotations correspond to a predefined class set, and low otherwise." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.231, + 0.48, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.375, + 0.483, + 0.418 + ], + "angle": 0, + "content": "Figure 5. GTs (blue) and TMR predictions (green) with different exemplars (red) from RPINE, which is the only dataset containing multiple patterns for each image among FSCD datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.43, + 0.483, + 0.49 + ], + "angle": 0, + "content": "In the real world, a pattern cannot be rigorously defined. We thus define the following criteria inspired by the definition of the translation symmetry [62] to minimize the subjective variance among different annotators." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.491, + 0.483, + 0.521 + ], + "angle": 0, + "content": "- Number of patterns: max 3 different patterns are annotated per image if the image exhibits multiple patterns." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.522, + 0.433, + 0.535 + ], + "angle": 0, + "content": "- Number of pattern instances: has no upperbound." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.537, + 0.483, + 0.566 + ], + "angle": 0, + "content": "- Minimum Size: the width and height of a pattern instance must be at least \\(3\\%\\) of the shorter side of the image." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.567, + 0.483, + 0.611 + ], + "angle": 0, + "content": "- Appearance variance: visually similar patterns with different scales/colors/semantics/rotation angles are annotated as the same pattern." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.612, + 0.483, + 0.642 + ], + "angle": 0, + "content": "- Reflection: visually similar patterns but reflection symmetric patterns are annotated in different patterns." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.643, + 0.483, + 0.672 + ], + "angle": 0, + "content": "- Occlusion: if visually similar patterns are occluded from each other, the visible parts are annotated." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.491, + 0.483, + 0.672 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.764 + ], + "angle": 0, + "content": "We ask annotators to carefully draw bounding boxes on the recognized patterns by the above instructions as consistently as they can. Despite of the instruction details, pattern are not pixel-perfectly defined across annotators. Therefore, we assign three individual annotators per image and include all the annotated patterns as ground truth." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.778, + 0.224, + 0.795 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.803, + 0.281, + 0.818 + ], + "angle": 0, + "content": "6.1. Dataset and metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We evaluate our model on RPINE as well as on the two standard FSCD benchmarks: FSCD-147 [47] and FSCDLVIS [47]. FSCD-147 contains a total of 6,135 images, with 3,659 for training, 1,286 for validation, and 1,190 for testing. FSCD-LVIS seen-split contains a total of 6,195 im" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.233, + 0.905, + 0.308 + ], + "angle": 0, + "content": "ages, with 4,000 for training, 1,181 for validation, and 1,014 for testing, covering 372 object categories. FSCD-LVIS unseen-split contains a total of 6,201 images, with 3,959 for training and 2,242 for testing where test-time object categories are not observed during training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.309, + 0.906, + 0.399 + ], + "angle": 0, + "content": "Following the evaluation protocol [47, 52], we report Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) for counting. For detection, we report average precision with IoU thresholds of 0.5 and 0.75, denoted as AP50 and AP75, respectively, along with the averaged AP over IoU thresholds from 0.5 to 0.95 in increments of 0.05." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.409, + 0.727, + 0.425 + ], + "angle": 0, + "content": "6.2. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.906, + 0.567 + ], + "angle": 0, + "content": "We use the pre-trained SAM-ViT/H [25] of the patch size 16 as the backbone and set it frozen during training, which returns the \\(64 \\times 64 \\times 256\\) feature map. Due to the input patchification encapsulated in the ViT backbones [5], we find the receptive field of the raw feature map too coarse to detect small instances. We thus bilinearly interpolate the feature map resolution from \\(64 \\times 64\\) to \\(128 \\times 128\\) to produce a higher-resolution correlation map, which enables denser predictions and leads to better performance. (Tab. 11)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.906, + 0.658 + ], + "angle": 0, + "content": "The channel dimension is expanded to 512 via a learnable linear projection. TMR is trained with the learning rate of \\(10^{-4}\\) with AdamW [39] and the batch size of 16 on four Nvidia RTX 3090 GPUs with 24 GB VRAM for 24 hours. The box presence threshold \\(\\tau\\) for NMS set to 0.4, 0.3 for RPINE and FSCD-147, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.668, + 0.874, + 0.684 + ], + "angle": 0, + "content": "6.3. Comparison with state-of-the-art methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.901 + ], + "angle": 0, + "content": "RPINE. To demonstrate the effectiveness of our method in counting and detecting non-object patterns, we evaluate it on RPINE. In Tab. 2, TMR surpasses the previous FSCD methods with a large margin. As the previous FSCD methods rely on prototypes for matching, they tend to struggle with non-object patterns that require an understanding of spatial details rather than semantics. In contrast, TMR effectively detects non-object patterns by incorporating spatial details in the template matching process. Note that RPINE is the only dataset equipped with multiple pattern annotations for each image in FSCD. Figure 5 demonstrates the bounding box predictions with different exemplars, where TMR predicts bounding boxes adaptively to the given exemplars. Figure 7 shows the qualitative com" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21582" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.483, + 0.187 + ], + "angle": 0, + "content": "
MethodSDMAE(↓)RMSE(↓)AP(↑)AP50(↑)AP75(↑)
C-DETR [47]9.5821.2413.8832.2010.22
SAM-C [42]✓18.7737.1418.8034.0418.74
PseCo [23]✓48.2088.1623.1844.5421.24
GeCo [52]✓9.5717.0723.3345.9321.19
TMR (ours)8.4519.8733.5964.0530.52
TMR (ours)✓8.3019.4029.6658.9425.41
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.193, + 0.483, + 0.234 + ], + "angle": 0, + "content": "Table 2. One-shot pattern counting and detection results on the RPINE dataset. SD denotes box refinement with the SAM decoder. All the models are trained by the official code." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.249, + 0.483, + 0.279 + ], + "angle": 0, + "content": "parisons with other FSCD methods where TMR accurately localize target patterns." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.283, + 0.483, + 0.48 + ], + "angle": 0, + "content": "FSCD-LVIS and FSCD-147. We compare TMR with existing methods that are dedicated to FSCD under the FSCD setting on FSCD-LVIS and FSCD-147. The results in Tab. 3 demonstrate that TMR significantly outperforms previous state-of-the-art approaches. Notably, when evaluated on the unseen split, where test-time object categories are not observed during training, TMR surpasses prior methods. This suggests that TMR is less biased to object semantics during training potentially because TMR leverages the exemplar's spatial structure that generalizes across different categories. In addition to FSCD-LVIS, Tab. 6 compares methods on FSCD-147, where TMR performs on par with the previous methods on both one-shot and three-shot settings." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.492, + 0.342, + 0.508 + ], + "angle": 0, + "content": "6.4. Analyses and ablation study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.518, + 0.483, + 0.79 + ], + "angle": 0, + "content": "TMR learns with less semantic object bias and generalizes well across datasets. We compare TMR and GeCo [52] in the cross-dataset scenarios by evaluating the trained models on different datasets that are unseen during training. As shown in Tab. 4, TMR presents overwhelming performances, showing its strong generalization ability. Specifically, when GeCo is trained on FSCD-147, its performance drops significantly when evaluated on different datasets compared to when tested on FSCD-147 itself. GeCo, like previous FSCD methods, relies on prototypes for both counting and detection. We also suspect this lower generalization ability than ours originates from prototype matching, which is prone to overfitting to the training object semantics. GeCo struggles when evaluated on datasets with different object semantics. In contrast, TMR utilizes the structural information for matching instead of relying on semantic-intensive prototypes and generalizes more effectively on unseen datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.901 + ], + "angle": 0, + "content": "SAM decoder is biased to edges. Table 2 shows the negative impact of the optional box refinement using the SAM decoder [25] on RPINE. Figure 6 visualizes two representative examples when the SAM decoder (SD) degrades performance. We observe that SD tends to align box predictions with the nearest edge, which is expected given that SAM is a segmentation model. The edge-sensitive predic" + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.089, + 0.897, + 0.228 + ], + "angle": 0, + "content": "
MethodSeenUnseen
AP(↑)AP50(↑)AP(↑)AP50(↑)
FSDetView-PB [68]2.727.571.032.89
AttRPN-PB [7]4.0811.153.157.87
C-DETR [47]4.9214.493.8511.28
DAVE [51]6.7522.514.1214.16
PseCo [23]22.3742.56--
GeCo [52]--11.4724.49
TMR (ours)27.4948.4822.7139.68
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.232, + 0.905, + 0.261 + ], + "angle": 0, + "content": "Table 3. Three-shot counting detection-based methods on the FSCD-LVIS seen and unseen split." + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.265, + 0.892, + 0.391 + ], + "angle": 0, + "content": "
TrainTestcross-evalAPAP50
GeCoTMRGeCoTMR
F-147F-14743.4244.4375.0673.83
F-LVISseen✓13.9621.2525.8737.18
RPINE✓19.4726.2138.6952.01
RPINEF-147✓36.9941.3960.3869.19
F-LVISseen✓10.0120.9217.4437.87
RPINE23.3329.6645.9358.94
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.395, + 0.905, + 0.423 + ], + "angle": 0, + "content": "Table 4. Cross-dataset comparison of GeCo [52] and TMR, where F-147, F-LVIS indicate FSCD-147 and FSCD-LVIS." + }, + { + "type": "table", + "bbox": [ + 0.551, + 0.429, + 0.871, + 0.504 + ], + "angle": 0, + "content": "
MethodOne-shotThree-shot
AP(↑)AP50(↑)AP(↑)AP50(↑)
GeCo [52]32.7169.9532.4970.51
TMR (ours)36.0171.1938.5772.61
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.507, + 0.905, + 0.521 + ], + "angle": 0, + "content": "Table 5. Comparison without optional SAM decoder on FSCD147" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.534, + 0.907, + 0.732 + ], + "angle": 0, + "content": "tion is particularly harmful for non-typical objects, such as objects that are loosely bounded by edges or patterns with many confusing edges. The SD refinement is seemingly good at edge detection for typical object exemplars with clear boundaries, and this is why the existing FSCD methods [23, 42, 52] benefit from adopting SD for box refinement. However, arbitrary patterns are not necessarily bounded by clear edges. Hence, edge-driven refinement may be even harmful as verified on RPINE. We emphasize that SD is an additional box post-processor, and we optionally add SD to TMR to compare with the SD-based state of the arts. Table 5 compares models by taking out SD, where TMR shows powerful performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Effectiveness of template matching features. We verify the effectiveness of 2D template matching, which preserves the spatial structure of exemplar bounding boxes for correlation. We experiment with different input feature \\( F_{\\mathrm{P}} \\) to the box regression module. Table 7 compares our final model (e) and its variants. The lower bound model (a) observes zero information on the exemplar. Model (b) shows that correlating with 2D templates is greatly beneficial. The final model (e) concatenates the template-matching and the image feature, providing the correlation and appearance information for pattern matching regression. The models (c," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21583" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.135, + 0.071, + 0.861, + 0.219 + ], + "angle": 0, + "content": "
MethodSAM decoderOne-shotThree-shot
MAE(↓)RMSE(↓)AP(↑)AP50(↑)MAE(↓)RMSE(↓)AP(↑)AP50(↑)
FSDetView-PB [68]----37.54147.0713.4132.99
AttRPN-PB [7]----32.42141.5520.9737.19
C-DETR† [47]16.99125.2219.1447.6316.79123.5622.6650.57
SAM-C† [42]✓33.17141.7735.0956.0227.97131.2427.9949.17
PseCo [23]✓14.86118.6441.6370.8713.05112.8642.9873.33
DAVE [51]11.5486.6219.4655.2710.4574.5126.8162.82
GeCo [52]✓8.1060.1643.1174.317.9154.2843.4275.06
TMR (ours)✓11.6357.4643.1571.5513.7851.8744.4373.83
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.223, + 0.907, + 0.251 + ], + "angle": 0, + "content": "Table 6. Comparison with few-shot detection and counting methods on FSCD-147. The one-shot performance of the models with \\( \\dagger \\) are evaluated using the official code." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.255, + 0.901, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.44, + 0.907, + 0.484 + ], + "angle": 0, + "content": "Figure 6. TMR without vs. with SAM decoder [25] (w/o SD vs. w/ SD) for box refinement. SAM decoder tends to align box predictions with the nearest edges. The edge-sensitive prediction is harmful for non-objects, objects that have loose boundaries, or patterns with many confusing edges. The exemplar in the first row loosely contains an arched window, but the SAM decoder snaps the box tightly its edges." + }, + { + "type": "table", + "bbox": [ + 0.13, + 0.489, + 0.443, + 0.608 + ], + "angle": 0, + "content": "
feature FPRPINEFSCD-147
APAP50APAP50
(a)F11.4424.0220.9555.59
(b)FTM32.5557.4331.9663.03
(c)F ⊕ FTM-cos29.7457.8524.7360.29
(d)F ⊕ FPM20.9447.9628.9166.93
(e)F ⊕ FTM33.5964.0536.0171.19
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.611, + 0.483, + 0.655 + ], + "angle": 0, + "content": "Table 7. Effect of template matching features (Sec. 4.1). \\( F_{\\mathrm{PM}} \\) is the correlation with average-pooled prototype. \\( F_{\\mathrm{TM - cos}} \\) is the cosine similarity of the template and image feature." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.484, + 0.901 + ], + "angle": 0, + "content": "d) replace the variants of the template matching feature \\( F_{\\mathrm{TM}} \\) with something else. Model (c) replaces the channel-wise correlation of (e) to cosine similarity, i.e., \\( F_{\\mathrm{TM - cos}} \\), showing that retaining the channel dimension helps. Model (d) first performs average-pooling of the template and uses the prototype matching feature, \\( F_{\\mathrm{PM}} \\), for correlation. Note that this average-pooled feature is often used in previous FSC methods [51, 52, 70]. The result shows that preserving the spatial structure of exemplar bounding boxes is more effective. The representative failure cases are shown in Fig. 11 in supp. material. The pooled prototype feature loses the geometric layout of the template and struggle to detect patterns where geometric clues are crucial. Comparing (d) and (e) verifies our hypothesis that prototype-based matching relies on the object priors; Model (d) drops more significantly on RPINE, which is less object-centric than FSCD-147." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.489, + 0.909, + 0.589 + ], + "angle": 0, + "content": "
box regressionRPINEFSCD-147
shiftscaleAPAP50APAP50
(a)(0,0)(0,0)26.2055.2022.8157.30
(b)(Δx, Δy)(eαw, eαh)23.8851.2517.0155.14
(c)(Δx, Δy)(eαw sw, eαsh)31.2959.5235.0870.23
(d)(sw Δx, sh Δy)(eαw sw, eαsh)33.5964.0536.0171.19
" + }, + { + "type": "table_caption", + "bbox": [ + 0.554, + 0.591, + 0.866, + 0.605 + ], + "angle": 0, + "content": "Table 8. Effect of box regression methods (Sec. 4.2)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.62, + 0.907, + 0.832 + ], + "angle": 0, + "content": "Effectiveness of support-conditioned regression. Table 8 compares different variants for box regression and validates the effectiveness of our support-conditioned box regression method. As shown in (b), regressing bounding boxes without considering the support exemplar's size leads to poor localization. Notably, it even performs worse than (a), which directly uses the exemplar box without performing any regression. However, as shown in (c), incorporating the support exemplar's size when predicting the width and height leads to a significant performance gain. This improvement is further enhanced by using the exemplar's size for shifting, as demonstrated in (d). These results demonstrate the effectiveness of support-conditioned regression, guided by the exemplar's size, in achieving accurate localization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Comparison on computational complexity. Table 9 compares the complexity of TMR with state-of-the-art FSCD methods, demonstrating its efficiency and effectiveness. Compared to PseCo [23], which introduces a large num" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21584" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.156, + 0.075, + 0.842, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.403, + 0.907, + 0.434 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison with the state-of-the-art models on RPINE (the first two images) and FSCD-147 (the last two images). More visualization examples on three datasets are included in the supplementary materials Figs 16-18." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.44, + 0.468, + 0.512 + ], + "angle": 0, + "content": "
MethodTrainable paramsTotal paramsFLOPS
PseCo [23]56.99M0.70B5.08T
GeCo [52]7.98M0.65B4.72T
TMR (ours)19.01M0.66B3.04T
" + }, + { + "type": "table_caption", + "bbox": [ + 0.124, + 0.515, + 0.452, + 0.53 + ], + "angle": 0, + "content": "Table 9. Comparison on the computational complexity" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.541, + 0.483, + 0.708 + ], + "angle": 0, + "content": "ber of trainable parameters and FLOPs, TMR introduces only about 19M trainable parameters, which is detailed in Tab. 10. Although TMR has more trainable parameters than GeCo [52], its total parameter count remains comparable. Thanks to the simple architecture, TMR is significantly efficient in terms of FLOPs (3.04T), which is notably lower than PseCo (5.08T) and GeCo (4.72T). This reduction in computational cost not only enhances training efficiency but also results in faster inference. In contrast, the high FLOPs of PseCo and GeCo contribute to longer inference, making them less practical for real-time applications." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.711, + 0.483, + 0.803 + ], + "angle": 0, + "content": "Real-world application. We evaluate TMR, trained on RPINE, on scanning electron microscope (SEM) images [1], which are widely used in microprocessor inspection. Despite the domain shift, our method performs effectively and demonstrates its potential to generalize to real-world, non-object pattern scenarios (Fig. 8)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.816, + 0.337, + 0.832 + ], + "angle": 0, + "content": "7. Conclusion and discussion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.903 + ], + "angle": 0, + "content": "We have proposed a simple template-matching based method for few-shot pattern detection. We also introduce a new dataset with bounding box annotations that covers various patterns around the world across non-object patterns" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.44, + 0.905, + 0.518 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.52, + 0.521, + 0.9, + 0.536 + ], + "angle": 0, + "content": "Figure 8. Application example of TMR on the SEM dataset [1]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.548, + 0.907, + 0.578 + ], + "angle": 0, + "content": "to typical objects from various image domains from nature and human-made products." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.58, + 0.909, + 0.747 + ], + "angle": 0, + "content": "We observed that the widely used SAM decoder improves our performance by \\(7\\%\\) AP points on FSCD-147 by leveraging object-level edge priors, but it degrades performance on RPINE, which contains more general patterns. This suggests a promising future direction that combines the powerful pre-trained knowledge of foundation models with detection mechanisms less dependent on object-level edge priors. In parallel, exploring a lightweight, pattern-specific architecture that efficiently captures fine-grained repetitive structures without relying on object semantics could further improve generalization and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Acknowledgments. This work was supported by Samsung Electronics Co., Ltd (IO240508-09825-01) and IITP grants (RS-2022-II220959: Few-Shot Learning of Causal Inference in Vision & Language \\((50\\%)\\), RS-2022-II220113: Developing a Sustainable Collaborative Multi-modal Lifelong Learning Framework \\((45\\%)\\), RS-2019-II191906: AI Graduate School Program at POSTECH \\((5\\%))\\) funded by Ministry of Science and ICT, Korea. We also appreciate Manjin Kim's advice on constructing RPINE." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21585" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Rossella Aversa, Mohammad Hadi Modarres, Stefano Cozzini, Regina Ciancio, and Alberto Chiusole. The first annotated set of scanning electron microscopy images for nanoscience. Scientific data, 5(1):1-10, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.228 + ], + "angle": 0, + "content": "[2] Yunliang Cai and George Baciu. Translation symmetry detection: A repetitive pattern analysis approach. In IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 223-228, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.231, + 0.483, + 0.297 + ], + "angle": 0, + "content": "[3] Cheng Chen, Wei Wang, John A Ozolek, and Gustavo K Rohde. A flexible and robust approach for segmenting cell nuclei from 2d microscopy images using supervised learning and template matching. Cytometry Part A, 83(5):495-507, 2013. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.301, + 0.483, + 0.328 + ], + "angle": 0, + "content": "[4] Taco S Cohen and Max Welling. Steerable cnns. arXiv preprint arXiv:1612.08498, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.331, + 0.483, + 0.413 + ], + "angle": 0, + "content": "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.416, + 0.483, + 0.47 + ], + "angle": 0, + "content": "[6] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International Journal of Computer Vision (IJCV), 88:303-338, 2010. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.473, + 0.483, + 0.54 + ], + "angle": 0, + "content": "[7] Qi Fan, Wei Zhuo, Chi-Keung Tang, and Yu-Wing Tai. Few-shot object detection with attention-rpn and multi-relation detector. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4013–4022, 2020. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.544, + 0.483, + 0.599 + ], + "angle": 0, + "content": "[8] Zhibo Fan, Yuchen Ma, Zeming Li, and Jian Sun. Generalized few-shot object detection without forgetting. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4527-4536, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.601, + 0.483, + 0.669 + ], + "angle": 0, + "content": "[9] Christopher Funk, Seungkyu Lee, Martin R Oswald, Stavros Tsogkas, Wei Shen, Andrea Cohen, Sven Dickinson, and Yanxi Liu. 2017 iccv challenge: Detecting symmetry in the wild. In International Conference on Computer Vision Workshops (ICCVW), pages 1692-1701, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.672, + 0.483, + 0.712 + ], + "angle": 0, + "content": "[10] Hiroyasu Furukawa, Kyle E Cordova, Michael O'Keeffe, and Omar M Yaghi. The chemistry and applications of metalorganic frameworks. Science, 341(6149):1230444, 2013. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.715, + 0.483, + 0.769 + ], + "angle": 0, + "content": "[11] Jizhou Gao, Yin Hu, Jinze Liu, and Ruigang Yang. Unsupervised learning of high-order structural semantics from images. In Proc. IEEE International Conference on Computer Vision (ICCV), pages 2122-2129. IEEE, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.772, + 0.483, + 0.811 + ], + "angle": 0, + "content": "[12] Ross Girshick. Fast r-cnn. In Proc. IEEE International Conference on Computer Vision (ICCV), pages 1440-1448, 2015. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.814, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[13] Alice L Givan. Flow cytometry: an introduction. Flow cytometry protocols, pages 1-29, 2011. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.845, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[14] Branko Grünbaum and Geoffrey Colin Shephard. Tilings and patterns. Courier Dover Publications, 1987. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[15] Karim Guirguis, Johannes Meier, George Eskandar, Matthias Kayser, Bin Yang, and Jürgen Beyerer. Niff: Alleviating" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.094, + 0.905, + 0.147 + ], + "angle": 0, + "content": "forgetting in generalized few-shot object detection via neural instance feature forging. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 24193-24202, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[16] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5356-5364, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.206, + 0.905, + 0.273 + ], + "angle": 0, + "content": "[17] Guangxing Han, Yicheng He, Shiyuan Huang, Jiawei Ma, and Shih-Fu Chang. Query adaptive few-shot object detection with heterogeneous graph convolutional networks. In Proc. IEEE International Conference on Computer Vision (ICCV), pages 3263-3272, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.275, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[18] Guangxing Han, Shiyuan Huang, Jiawei Ma, Yicheng He, and Shih-Fu Chang. Meta faster r-cnn: Towards accurate few-shot object detection with attentive feature alignment. In Proc. AAAI Conference on Artificial Intelligence (AAAI), pages 780-789, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.345, + 0.905, + 0.411 + ], + "angle": 0, + "content": "[19] Guangxing Han, Jiawei Ma, Shiyuan Huang, Long Chen, and Shih-Fu Chang. Few-shot object detection with fully cross-transformer. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5321-5330, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.414, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[20] Junwei Han, Stephen J McKenna, and Ruixuan Wang. Regular texture analysis as statistical model selection. In Proc. European Conference on Computer Vision (ECCV), pages 242-255. Springer, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.471, + 0.905, + 0.535 + ], + "angle": 0, + "content": "[21] James Hays, Marius Leordeanu, Alexei A Efros, and Yanxi Liu. Discovering texture regularity as a higher-order correspondence problem. In Proc. European Conference on Computer Vision (ECCV), pages 522-535. Springer, 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.539, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[22] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proc. IEEE International Conference on Computer Vision (ICCV), pages 2961-2969, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.582, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[23] Zhizhong Huang, Mingliang Dai, Yi Zhang, Junping Zhang, and Hongming Shan. Point segment and count: A generalized framework for object counting. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.652, + 0.905, + 0.706 + ], + "angle": 0, + "content": "[24] Bingyi Kang, Zhuang Liu, Xin Wang, Fisher Yu, Jiashi Feng, and Trevor Darrell. Few-shot object detection via feature reweighting. In Proc. IEEE International Conference on Computer Vision (ICCV), pages 8420-8429, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.707, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[25] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In Proc. IEEE International Conference on Computer Vision (ICCV), 2023. 3,4,5,6,7,2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.79, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[26] Simon Korman, Daniel Reichman, Gilad Tsur, and Shai Avidan. Fast-match: Fast affine template matching. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2331-2338, 2013. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.847, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[27] Minji Lee, Hyunju Noh, and Youngsuk Kim. Diamidocarbene-derived palladium and nickel-sulfur clusters. Chemical Communications, 60(94):13867-13870, 2024. 4" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.094, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21586" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[28] Thomas Leung and Jitendra Malik. Detecting, localizing and grouping repeated scene elements from an image. In Proc. European Conference on Computer Vision (ECCV). Springer, 1996. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[29] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.234 + ], + "angle": 0, + "content": "[30] Hui Lin, Xiaopeng Hong, and Yabin Wang. Object counting: You only need to look at one. arXiv preprint arXiv:2112.05993, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.236, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[31] Hsin-Chih Lin, Ling-Ling Wang, and Shi-Nine Yang. Extracting periodicity of a regular texture based on autocorrelation functions. Pattern recognition letters, 18(5):433-443, 1997. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.482, + 0.361 + ], + "angle": 0, + "content": "[32] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proc. European Conference on Computer Vision (ECCV), pages 740-755. Springer, 2014. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.431 + ], + "angle": 0, + "content": "[33] Tsung-Yi Lin, Piotr Dólár, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2117-2125, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.433, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[34] Wen-Chieh Lin and Yanxi Liu. Tracking dynamic near-regular texture under occlusion and rapid movements. In Proc. European Conference on Computer Vision (ECCV), pages 44–55. Springer, 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.482, + 0.532 + ], + "angle": 0, + "content": "[35] Jingchen Liu and Yanxi Liu. Grasp recurring patterns from a single view. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2003-2010, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.482, + 0.616 + ], + "angle": 0, + "content": "[36] Jingchen Liu, George Slota, Gang Zheng, Zhaohui Wu, Minwoo Park, Seungkyu Lee, Ingmar Rauschert, and Yanxi Liu. Symmetry detection from realworld images competition 2013: Summary and results. In IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 200-205, 2013. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.617, + 0.482, + 0.686 + ], + "angle": 0, + "content": "[37] Siying Liu, Tian-Tsong Ng, Kalyan Sunkavalli, Minh N Do, Eli Shechtman, and Nathan Carr. Patchmatch-based automatic lattice detection for near-regular textures. In Proc. IEEE International Conference on Computer Vision (ICCV), pages 181–189, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.688, + 0.482, + 0.73 + ], + "angle": 0, + "content": "[38] Yanxi Liu, Wen-Chieh Lin, and James Hays. Near-regular texture analysis and manipulation. ACM Transactions on Graphics (TOG), 23(3):368-376, 2004. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.731, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[39] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.759, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[40] Erika Lu, Weidi Xie, and Andrew Zisserman. Class-agnostic counting. In Asian Conference on Computer Vision (ACCV), pages 669–684. Springer, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[41] Bronwyn A Lucas, Benjamin A Himes, Liang Xue, Timothy Grant, Julia Mahamid, and Nikolaus Grigorieff. Locating macromolecular assemblies in cells by 2d template matching with cistem. _Elife_, 10:e68946, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[42] Zhiheng Ma, Xiaopeng Hong, and Qinnan Shangguan. Can sam count anything? an empirical study on sam counting, 2023. 3, 4, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[43] Massimiliano Mancini, Elisa Ricci, Barbara Caputo, and Samuel Rota Bulo. Adding new tasks to a single network with weight transformations using binary masks. In European Conference on Computer Vision Workshops (ECCVW), pages 0-0, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[44] Giovanni Marola. Using symmetry for detecting and locating objects in a picture. Computer Vision, Graphics, and Image Processing, 46(2):179-195, 1989. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[45] Jiri Matas, Ondrej Chum, Martin Urban, and Tomás Pajdla. Robust wide-baseline stereo from maximally stable extremal regions. Image and vision computing, 22(10):761-767, 2004." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[46] Anuj Mohan, Constantine Papageorgiou, and Tomaso Poggio. Example-based object detection in images by components. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 23(4):349-361, 2001. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[47] Thanh Nguyen, Chau Pham, Khoi Nguyen, and Minh Hoai. Few-shot object counting and detection. In Proc. European Conference on Computer Vision (ECCV). Springer, 2022. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.905, + 0.433 + ], + "angle": 0, + "content": "[48] Roni Paiss, Ariel Ephrat, Omer Tov, Shiran Zada, Inbar Mosseri, Michal Irani, and Tali Dekel. Teaching clip to count to ten. In Proc. IEEE International Conference on Computer Vision (ICCV), pages 3170-3180, 2023. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[49] Minwoo Park, Kyle Brocklehurst, Robert T Collins, and Yanxi Liu. Deformed lattice detection in real-world images using mean-shift belief propagation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 31(10): 1804-1816, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[50] Mark Pauly, Niloy J Mitra, Johannes Wallner, Helmut Pottmann, and Leonidas J Guibas. Discovering structural regularity in 3d geometry. In ACM SIGGRAPH, pages 1-11, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.905, + 0.617 + ], + "angle": 0, + "content": "[51] Jer Pelhan, Vitjan Zavrtanik, Matej Kristan, et al. Davea detect-and-verify paradigm for low-shot counting. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 23293-23302, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.618, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[52] Jer Pelhan, Alan Lukezic, Vitjan Zavrtanik, and Matej Kristan. A novel unified architecture for low-shot counting by detection and segmentation. Advances in Neural Information Processing Systems (NeurIPS), 2025. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[53] Julien Picot, Coralie L Guerin, Caroline Le Van Kim, and Chantal M Boulanger. Flow cytometry: retrospective, fundamentals and recent instrumentation. Cytotechnology, 64: 109-130, 2012. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.731, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[54] Muhammad Farrukh Qureshi, Zohaib Mushtaq, Muhammad Zia Ur Rehman, and Ernest Nlandu Kamavuako. E2cnn: An efficient concatenated cnn for classification of surface emg extracted from upper limb. IEEE Sensors Journal, 23(8): 8989-8996, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[55] Viresh Ranjan and Minh Hoai. Vicinal counting networks. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4221-4230, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[56] Viresh Ranjan, Udbhav Sharma, Thu Nguyen, and Minh Hoai. Learning to count everything. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3394-3403, 2021. 2, 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21587" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[57] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 779-788, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[58] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 658-666, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[59] Babak Saleh and Ahmed Elgammal. Large-scale classification of fine-art paintings: Learning the right metric on the right feature. arXiv preprint arXiv:1505.00855, 2015. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.304 + ], + "angle": 0, + "content": "[60] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.482, + 0.374 + ], + "angle": 0, + "content": "[61] Luciano Spinello, Rudolph Triebel, Dizan Vasquez, Kai O Arras, and Roland Siegwart. Exploiting repetitive object patterns for model compression and completion. In Proc. European Conference on Computer Vision (ECCV), pages 296-309. Springer, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.375, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[62] V.J. Stenger. Timeless Reality: Symmetry, Simplicity, and Multiple Universes. Prometheus, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.482, + 0.471 + ], + "angle": 0, + "content": "[63] Bo Sun, Banghuai Li, Shengcai Cai, Ye Yuan, and Chi Zhang. Fsce: Few-shot object detection via contrastive proposal encoding. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 7352-7362, 2021, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.475, + 0.482, + 0.53 + ], + "angle": 0, + "content": "[64] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In Proc. IEEE International Conference on Computer Vision (ICCV), pages 9627-9636, 2019. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.532, + 0.482, + 0.573 + ], + "angle": 0, + "content": "[65] Du-Ming Tsai and Cheng-Huei Chiang. Rotation-invariant pattern matching using wavelet decomposition. Pattern Recognition Letters, 23(1-3):191-201, 2002. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.575, + 0.482, + 0.63 + ], + "angle": 0, + "content": "[66] Andreas Turina, Tinne Tuytelaars, and Luc Van Gool. Efficient grouping under perspective skew. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages I-I. IEEE, 2001. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.482, + 0.686 + ], + "angle": 0, + "content": "[67] Xin Wang, Thomas Huang, Joseph Gonzalez, Trevor Darrell, and Fisher Yu. Frustratingly simple few-shot object detection. In Proc. International Conference on Machine Learning (ICML), pages 9919-9928. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.688, + 0.482, + 0.729 + ], + "angle": 0, + "content": "[68] Yang Xiao, Vincent Lepetit, and Renaud Marlet. Few-shot object detection and viewpoint estimation for objects in the wild. TPAMI, 45(3):3090-3106, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.731, + 0.482, + 0.799 + ], + "angle": 0, + "content": "[69] Xiaopeng Yan, Ziliang Chen, Anni Xu, Xiaoxi Wang, Xiaodan Liang, and Liang Lin. Meta r-cnn: Towards general solver for instance-level low-shot learning. In Proc. IEEE International Conference on Computer Vision (ICCV), pages 9577-9586, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.482, + 0.855 + ], + "angle": 0, + "content": "[70] Zhiyuan You, Kai Yang, Wenhan Luo, Xin Lu, Lei Cui, and Xinyi Le. Iterative correlation-based feature refinement for few-shot counting. arXiv preprint arXiv:2201.08959, 4, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.858, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[71] Xinyu Zhang, Yuhan Liu, Yuting Wang, and Abdeslam Boularias. Detect everything with few examples. arXiv preprint arXiv:2309.12969, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "21588" + } + ] +] \ No newline at end of file diff --git a/2025/Few-Shot Pattern Detection via Template Matching and Regression/9a1fc29b-9538-4897-943a-30b2211079d1_origin.pdf b/2025/Few-Shot Pattern Detection via Template Matching and Regression/9a1fc29b-9538-4897-943a-30b2211079d1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d1c14d1caf22ed826f73ad848960707cbf69d0bd --- /dev/null +++ b/2025/Few-Shot Pattern Detection via Template Matching and Regression/9a1fc29b-9538-4897-943a-30b2211079d1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:17e63a32eb6542e29710521616fbabac627816474d8798d7fea76479d467dbd7 +size 3888875 diff --git a/2025/Few-Shot Pattern Detection via Template Matching and Regression/full.md b/2025/Few-Shot Pattern Detection via Template Matching and Regression/full.md new file mode 100644 index 0000000000000000000000000000000000000000..099674ddd4edc1b63ae83e443577bbc2633211f0 --- /dev/null +++ b/2025/Few-Shot Pattern Detection via Template Matching and Regression/full.md @@ -0,0 +1,333 @@ +# Few-Shot Pattern Detection via Template Matching and Regression + +Eunchan Jo Dahiyun Kang Sanghyun Kim Yunseon Choi Minsu Cho + +Pohang University of Science and Technology (POSTECH), South Korea + +https://cvlab.postech.ac.kr/research/TMR + +# Abstract + +We address the problem of few-shot pattern detection, which aims to detect all instances of a given pattern, typically represented by a few exemplars, from an input image. Although similar problems have been studied in few-shot object counting and detection (FSCD), previous methods and their benchmarks have narrowed patterns of interest to object categories and often fail to localize non-object patterns. In this work, we propose a simple yet effective detector based on template matching and regression, dubbed TMR. While previous FSCD methods typically represent target exemplars as spatially collapsed prototypes and lose structural information, we revisit classic template matching and regression. It effectively preserves and leverages the spatial layout of exemplars through a minimalistic structure with a small number of learnable convolutional or projection layers on top of a frozen backbone. We also introduce a new dataset, dubbed RPINE, which covers a wider range of patterns than existing object-centric datasets. Our method outperforms the state-of-the-art methods on the three benchmarks, RPINE, FSCD-147, and FSCD-LVIS, and demonstrates strong generalization in cross-dataset evaluation. + +# 1. Introduction + +Few-shot detection aims to identify target patterns with minimal labeled examples. While significant progress has been made in few-shot object detection [7, 8, 18, 68, 69], most existing methods remain object-centric, focusing primarily on identifying object-level patterns with relatively clear boundaries. However, many real-world applications require detecting arbitrary target patterns that extend beyond objects to include structural, geometric, or abstract patterns across diverse visual data. Despite recent progress based on deep neural networks, current methods still fall short in addressing these broader pattern detection tasks. Furthermore, the object-centric design of conventional few-shot detectors may lead to performance degradation when the target object lacks clear boundaries or when occlusion and deformation cause its boundaries to become indistinct. + +![](images/5fe6ad08219b7d44758127fa1a9a891bc25f37fc8db852944122976036b023c4.jpg) +one-shot exemplars +Figure 1. Few-shot pattern detection. Given a few exemplar for each target pattern (left), the task is to detect all matching instances of each pattern (right). This example includes non-object patterns (e.g., green and blue) as well as object patterns (e.g., red). + +![](images/2dd86fb3cd93c131095a899e5df13399f4d7197987db6270b1a06ced863f4d22.jpg) +ground-truth pattern detection + +The task of few-shot pattern detection is illustrated in Fig. 1. Recent related research topics for few-shot detection include few-shot counting and detection [47], and few-shot object detection [7]. Both aim to reduce the annotation cost of object detection [12, 33, 57, 64] by learning to detect all instances of given support exemplars. Consequently, many of these methods are heavily biased on object-centric benchmarks prior [6, 16, 32, 47]. In addition, many recent approaches [7, 18, 23, 52, 68, 69] represent the support exemplars as the spatially pooled vector, often named as a prototype [60]. While this pooling strategy is effective for detecting objects, it collapses the geometric properties, such as the shape and structure of the support exemplars. As a consequence, these methods tend to underperform when detecting non-object geometric patterns such as object parts or shape-intensive elements as shown in Fig. 3. + +In this work, we revisit the classic template matching strategy and propose a simple yet effective few-shot detector for arbitrary patterns. The proposed method, dubbed template matching and regression (TMR), is designed to be aware of the structure and shape of given exemplars. Given an input image, TMR first extracts a feature map using a backbone network. It then crops a template feature from the support exemplar's bounding box using a template extraction technique based on RoIAlign [22]. This template is correlated with the image feature map to produce a template matching feature map. Using this correlation map, the + +![](images/65eba8dd183601e29102b9e3ae5c8647d189fc49c1c40666eb235d38d7c0bc15.jpg) +Figure 2. Annotation examples of FSCD-147 [47] and RPINE. FSCD-147 is annotated with the repetitive object-level patterns but disregards the repetition of non-object patterns such as the egg tray pattern under the eggs in the first image. RPINE is annotated with arbitrarily noticed repeated patterns which include non-object patterns and nameless parts of objects. Plus, FSCD-147 is annotated with a single pattern, while RPINE is annotated with all existing repetitions (marked with different colors) recognized by three different annotators. + +![](images/82a6181c922bf3664fa51dd561636a44a7fccb01cc60d33360a5bef96a04d6a6.jpg) +Figure 3. Few-shot detectors trained with strong object prior are often biased to objects instances and struggle to detect non-objects. + +![](images/06d2c2087d961c1bd1a809ff91b41c64267633b76b15c576c64241e87c8bd70f.jpg) + +![](images/578ef30a8b6e1c5d7eac771154410f8299ce3010aad6597453378ee5372df632.jpg) + +![](images/96b66ed01472422db90f254fc79143c02fcc0f3ba72853da1fa6d81cf924bedc.jpg) + +model learns bounding box regression parameters to rectify the support exemplar's box size adaptively. This process, termed support-conditioned regression, enables the model to handle support exemplars of varying sizes more effectively. Notably, TMR consists only of a few $3 \times 3$ and linear projections without any complicated modules such as cross-attention, commonly used in prior work [47, 52]. + +Although TMR is designed for general pattern detection, existing benchmarks (e.g., FSCD-147 [47], FSCD-LVIS [47]) mainly target object-level patterns, limiting comprehensive evaluation. To address this, we introduce a new dataset, Repeated Patterns IN Everywhere (RPINE), which covers diverse repeated patterns in the real world. RPINE contains images with varying degrees of objectness, from well-defined object-level patterns to non-object patterns, all annotated with bounding boxes via crowd-sourcing. Compared to FSCD datasets, RPINE provides broader coverage, including both non-object patterns and nameless parts of objects, as illustrated in Fig. 2. + +TMR demonstrates strong performance in detecting repeated patterns, not only on RPINE but also on the FSCD benchmarks, FSCD-147 and FSCD-LVIS [47]. In particular, TMR is especially effective on RPINE, includes diverse patterns with minimal object priors. Notably, our simple architecture contributes to improved generalization across datasets. Our contribution is summarized as follows: + +- We generalize the few-shot object counting and detection to a pattern detection task that does not assume objectness + +in either the target patterns or exemplars. + +- We present a simple yet effective pattern detector by refining template matching, which efficiently detects coherent patterns guided by exemplars. +- We introduce a new densely annotated dataset, RPINE, which covers diverse repetitive patterns in the real world, ranging from object-level patterns to non-object patterns. +- TMR not only outperforms the state-of-the-art FSCD models on RPINE and FSCD-LVIS but also achieves strong cross-dataset generalization. + +# 2. Related work + +Few-shot object detection (FSOD) aims to detect objects of novel classes using only a few support images of novel classes. Existing methods can be roughly categorized into two groups: finetuning based [8, 15, 63, 67, 68] and meta-learning based [7, 17-19, 24, 69, 71] approaches. Despite significant progress, finetuning methods require retraining for every new classes. For pattern detection, it is not desired to fine-tune each arbitrary pattern with indistinct object categorization. In contrast, meta-learning methods avoid class-specific fine-tuning and typically construct prototypes from support images for classifying the bounding boxes of the query image. These methods often apply global average pooling to obtain prototypes from each support exemplar, which collapses the spatial structure of the exemplar. While such spatially collapsed prototypes may effectively capture object-level bounding boxes, they struggle with non-object patterns that exhibit complex spatial structures. + +Few-shot counting (FSC) aims to count objects given a few exemplars (typically 1 to 3) within an image. Previous methods [30, 40, 55, 56, 70] have tackled this task using density-map regression without detecting the individual bounding boxes. Towards localization-based counting, few-shot counting and detection (FSCD) is proposed to combine FSC and detection. Such models include Counting- + +![](images/b992c1cab19b66227e6eb39c4f3e257111f604c62d551c81b4dcbf705283a70e.jpg) +Figure 4. Template Matching and Regression (TMR). A template feature map is cropped from the image feature map and then correlated with the image feature map via channel-wise template matching. The TM feature map and the image feature map are concatenated. For each feature map point, the box regressor predicts the shifting and scaling parameters of the support exemplar's size, and the presence classifier scores the presence map. Both of them consist of a $3 \times 3$ convolution and a linear projection without any complicated layers. + +DETR [47], SAM-C [42], DAVE [51], PseCo [23], and GeCo [52]. However, similar to the FSOD methods, these models typically generate a prototype by spatially averaging exemplar bounding boxes. This discards the spatial structure of the exemplars, potentially losing important cues and details for accurate counting and detection. In contrast, our pattern-detection method preserves the spatial structure of the exemplar for pattern matching. + +Template matching. Template matching [21, 26, 44-46, 65, 66] has been widely used from the beginning of computer vision and pattern recognition and also adopted in convolution-based neural detectors with additional regression [12, 33]. Given a 2D template, template matching identifies the matching region by sliding-window fashion. + +Repetitive pattern detection. Detecting repetitive patterns is trivial for humans but remains challenging in computer vision. Early research focused on geometrically constrained settings where patterns are nearly regular and aligned [14, 20, 31, 50]. Based on this near-regularity, early methods assumed a global repetitive structure and discovered the repetition [21, 31, 37, 38, 49] in a top-down manner. In contrast, later methods [11, 34, 61] define the smallest repeating unit, called a texel [2, 28, 35] and identify all matching subparts based on the identified texels. In our context, the texel corresponds to the given exemplar. + +# 3. Few-shot pattern detection + +Given an input image $I \in \mathbb{R}^{H_0 \times W_0 \times 3}$ , the goal of few-shot pattern detection is to predict matching patterns with a given set of support exemplars. With an abuse of notations, the model is given the set of support exemplar, $\mathcal{E} = \{e_1, \dots, e_{N_s}\}$ and aims to predict the corresponding ground-truth bounding boxes of pattern $\mathcal{B} = \{b_1, \dots, b_{N_p}\}$ , where $N_p$ and $N_s$ denotes the number of ground-truth bounding boxes of the pattern and the number of exemplars (typically referred to as the "shot"), respectively. The input exemplars and the output patterns are both represented as bounding boxes parameterized by their center coordinates, width, and height: $b_i, e_i \in \mathbb{R}^4$ . + +# 4. Template matching and regression (TMR) + +For clarity, we primarily focus on the one-shot, single-scale pattern detection setting, where a single support exemplar is given and a single-resolution image feature map is used. However, TMR can be effortlessly extended to few-shot and multi-scale scenarios, as described in Sec.4.3 and Sec.9. + +The overall architecture of TMR is illustrated in Fig. 4. The input image $\pmb{I}$ is first encoded by a backbone such as ViT [25] to extract a feature map $\pmb{F} \in \mathbb{R}^{H \times W \times D}$ . The template feature $\pmb{T} \in \mathbb{R}^{t_h \times t_w \times D}$ is then obtained through the template extraction process, which uses RoIAlign [22] to crop a region from the image feature map $\pmb{F}$ based on the support exemplar's size $(s_h, s_w)$ . In RoIAlign, unlike previous fixed-size pooling methods [47, 52], the model adaptively determines the size of $\pmb{T}$ to fit the corresponding size on $\pmb{F}$ by rounding up the size of the exemplars on $\pmb{F}$ as described in Sec. 8.2. This preserves the spatial alignment between $\pmb{T}$ and $\pmb{F}$ with translation. The image feature map $\pmb{F}$ and template feature $\pmb{T}$ are correlated by template matching (Sec. 4.1), which outputs the template-matching feature $F_{\mathrm{TM}}$ . The concatenation of the feature maps $\pmb{F}$ , $F_{\mathrm{TM}}$ is fed to the subsequent box prediction module, which consists of a pattern box regressor and a pattern presence classifier. The pattern box regressor predicts the localization bounding boxes: $\hat{\pmb{B}}$ , which is parameterized by scaling and shifting factors of the given exemplar's size. The pattern presence classifier predicts the presence score of the pattern: $\hat{\pmb{P}}$ . A box proposal is then generated on each feature map point based on the combination of $\hat{\pmb{B}}$ and $\hat{\pmb{P}}$ (Sec. 4.2). At inference, bounding boxes with low presence scores are removed by Non-Maximum Suppression (NMS). SAM decoder [25] can be optionally applied for box refinement before NMS. + +# 4.1. Template matching (TM) + +We are motivated to detect arbitrary patterns given a support exemplar of either an object or not. Non-object patterns often lack high-level semantics yet exhibit low-level structural features, thus, preserving the spatial structure of the support exemplar is crucial. Inspired by traditional tem + +plate matching [26, 44, 65], we compute the matching score between the image and the support exemplar to detect the locations of the pattern. Specifically, template matching cross-correlates the feature map $\mathbf{F}$ and the template feature $\mathbf{T}$ by centering $\mathbf{T}$ at each $(x,y)$ position in $\mathbf{F}$ . The resultant template-matching (TM) feature $\mathbf{F}_{\mathrm{TM}}$ is obtained as: + +$$ +\boldsymbol {F} _ {\mathrm {T M}} (x, y) = \frac {1}{t _ {w} t _ {h}} \sum_ {x ^ {\prime}, y ^ {\prime}} \boldsymbol {F} \left(x + x ^ {\prime} - \lfloor \frac {t _ {w}}{2} \rfloor , y + y ^ {\prime} - \lfloor \frac {t _ {h}}{2} \rfloor\right) \boldsymbol {T} \left(x ^ {\prime}, y ^ {\prime}\right), \tag {1} +$$ + +where $\lfloor \cdot \rfloor$ denotes the floor operation used for centering $T$ at each $(x,y)$ , and $(x^{\prime},y^{\prime})$ ranges within the template coordinates: $\in [0,t_w)\times [0,t_h)$ . Note that Eq. 1 computes the correlation by channel-wise multiplication, resulting in $F_{\mathrm{TM}}\in \mathbb{R}^{H\times W\times D}$ . Finally, the TM feature $F_{\mathrm{TM}}$ is concatenated with the feature map $F$ : $F_{\mathrm{P}} = [F_{\mathrm{TM}};F]\in \mathbb{R}^{H\times W\times 2D}$ . The concatenated feature map $F_{\mathrm{P}}$ is fed to the subsequent box prediction module. + +# 4.2. Support-conditioned box regression + +Our box prediction module consists of a pattern box regressor and a pattern presence classifier following the architecture of an anchor-free detection methods [64]. + +The pattern box regressor $g_{\mathrm{B}}$ consists of a $3 \times 3$ convolutional layer followed by a linear layer and predicts the four localization parameters $(\Delta x, \Delta y, \alpha_w, \alpha_h)$ for each feature map point. Unlike the methods [47, 52] that directly regress absolute box parameters, our method performs support-conditioned regression. Bounding box parameters are predicted by scaling and shifting the support exemplar's size. This helps to dynamically adjust the exemplar's size to predict a target box size. A predicted bounding box at a feature point $(x, y)$ shifts and scales the support exemplar's size $(s_w, s_h)$ such as: + +$$ +\left(x + s _ {w} \Delta x, y + s _ {h} \Delta y, e ^ {\alpha_ {w}} s _ {w}, e ^ {\alpha_ {h}} s _ {h}\right). \tag {2} +$$ + +The pattern presence classifier $g_{\mathrm{P}}$ consists of a linear layer and predicts presence scores, which represent the confidence of the predicted bounding boxes at each feature map point. The aforementioned procedure of the box regressor $g_{\mathrm{B}}$ and the presence classifier $g_{\mathrm{P}}$ is summarized as: + +$$ +\hat {\boldsymbol {B}} = g _ {\mathrm {B}} \left(\boldsymbol {F} _ {\mathrm {P}}\right), \quad \hat {\boldsymbol {B}} \in \mathbb {R} ^ {H \times W \times 4}, \tag {3} +$$ + +$$ +\hat {\boldsymbol {P}} = \sigma \left(g _ {\mathrm {P}} \left(\boldsymbol {F} _ {\mathrm {P}}\right)\right), \quad \hat {\boldsymbol {P}} \in \mathbb {R} ^ {H \times W \times 1}, \tag {4} +$$ + +where $\sigma$ denotes the sigmoid function. + +# 4.3. Inference + +At inference, we first remove bounding boxes whose presence score is lower than a threshold $\tau$ to filter out low-confidence predictions. Afterward, we can optionally apply box localization refinement via the SAM decoder [25] as often adopted by the FSCD work [23, 42, 52]. Following [23, 52], we input the predicted box coordinates into the + +SAM prompt encoder to obtain the prompt feature. The obtained prompt feature and the image feature extracted from the SAM backbone are fed to the SAM decoder, which further refines input box coordinates. To obtain the final box prediction, $\hat{\mathcal{B}}$ , we apply NMS on the bounding boxes. + +TMR is easily extended for few-shot inference without re-training. When multiple support exemplars are given, we perform the above process for each exemplar individually and then aggregate the results before applying NMS. + +# 4.4. Learning objective + +The training loss is composed of $\mathcal{L}_{\mathrm{P}}$ , which penalizes the presence score, and $\mathcal{L}_{\mathrm{B}}$ , which penalizes the bounding box regression. The presence loss $\mathcal{L}_{\mathrm{P}}$ is the binary cross-entropy loss (BCELoss) between the predicted and the ground-truth center points of the pattern bounding boxes. Instead of guiding the presence with a single pixel, we add a margin around each ground-truth center point and define the extended center point set as $\mathcal{X}_{\mathrm{P}}$ . We detail this in Sec. 8.3. The ground-truth presence label $P$ at a position $(x,y)$ is set to 1 if the position is in $\mathcal{X}_{\mathrm{P}}$ , and 0 otherwise. + +The bounding box regression loss $\mathcal{L}_{\mathrm{B}}$ is defined as the generalized IoU (gIoU) loss [58] between the predicted and the ground-truth bounding boxes. This loss is defined on the positions where the ground-truth patterns exist. For each center point of the pattern bounding box, the adjacent area of the center within the margin shares the corresponding bounding box parameter $B$ . The two loss functions are defined as the following: + +$$ +\mathcal {L} _ {\mathrm {P}} = \sum_ {(x, y)} ^ {\forall (x, y)} \operatorname {B C E L o s s} (\hat {\boldsymbol {P}} (x, y), \boldsymbol {P} (x, y)), \tag {5} +$$ + +$$ +\mathcal {L} _ {\mathrm {B}} = \sum_ {(x, y)} ^ {(x, y) \in \mathcal {X} _ {\mathrm {P}}} \operatorname {g I o U L o s s} (\hat {\boldsymbol {B}} (x, y), \boldsymbol {B} (x, y)). \tag {6} +$$ + +The overall training loss is the sum of them: $\mathcal{L} = \mathcal{L}_{\mathrm{P}} + \mathcal{L}_{\mathrm{B}}$ + +# 5. Proposed dataset: RPINE + +Existing benchmarks (e.g., FSCD-147 [47], FSCD-LVIS [47]) focus on object-level patterns, limiting their use for general pattern detection. Therefore, we introduce a new pattern dataset, RPINE: Repeated Patterns In Everywhere. + +We collect images from the repetitive pattern detection literature and annotate them to contain various repetitive patterns in the wild. Images are collected from FSC-147 [47], FSCD-LVIS [47], Countbench [48], Wikiart [59], Frieze [43], and Wallpaper [36]. Tab. 1 shows where RPINE stands among related datasets. The dataset contains 4,362 images, divided into 3,925 and 435 for training and testing, respectively. As shown in Fig. 5, RPINE is annotated with multiple patterns per image. RPINE is a suitable evaluation benchmark for multi-pattern detection within an image. + +
datasetpattern tilingrepetitionobject biasmulti-pattern annotations per image
Wallpaper [36]near regular [38]highlow
Pascal [6], COCO [32]-lowhigh✓ (multi-category)
Wikiart [59], Frieze [9], Counting bench [48]arbitrarymidhigh
FSCD-147 [47], FSCD-LVIS [47]arbitraryhighhigh
RPINEarbitraryhighlow to high✓
+ +Table 1. Benchmark dataset comparison of related work. RPINE covers pattern or object repetitions in the real world. We classify 'pattern- tiling' as near-regular following [38], i.e., elements with an almost periodic lattice with minor variations in shape, color or lighting, and as arbitrary otherwise. The 'repetition' is classified based on the average number of repeated same class instances on an image: low if $\leq 5$ , mid if $\leq {20}$ ,and high otherwise. The 'object-bias' is high if the dataset annotations correspond to a predefined class set, and low otherwise. + +![](images/c9c8dd92228de1ad54d4c11a3f021f428bdc54e24d9c22c59d9bb909f34160e1.jpg) +Figure 5. GTs (blue) and TMR predictions (green) with different exemplars (red) from RPINE, which is the only dataset containing multiple patterns for each image among FSCD datasets. + +In the real world, a pattern cannot be rigorously defined. We thus define the following criteria inspired by the definition of the translation symmetry [62] to minimize the subjective variance among different annotators. + +- Number of patterns: max 3 different patterns are annotated per image if the image exhibits multiple patterns. +- Number of pattern instances: has no upperbound. +- Minimum Size: the width and height of a pattern instance must be at least $3\%$ of the shorter side of the image. +- Appearance variance: visually similar patterns with different scales/colors/semantics/rotation angles are annotated as the same pattern. +- Reflection: visually similar patterns but reflection symmetric patterns are annotated in different patterns. +- Occlusion: if visually similar patterns are occluded from each other, the visible parts are annotated. + +We ask annotators to carefully draw bounding boxes on the recognized patterns by the above instructions as consistently as they can. Despite of the instruction details, pattern are not pixel-perfectly defined across annotators. Therefore, we assign three individual annotators per image and include all the annotated patterns as ground truth. + +# 6. Experiments + +# 6.1. Dataset and metrics + +We evaluate our model on RPINE as well as on the two standard FSCD benchmarks: FSCD-147 [47] and FSCDLVIS [47]. FSCD-147 contains a total of 6,135 images, with 3,659 for training, 1,286 for validation, and 1,190 for testing. FSCD-LVIS seen-split contains a total of 6,195 im + +ages, with 4,000 for training, 1,181 for validation, and 1,014 for testing, covering 372 object categories. FSCD-LVIS unseen-split contains a total of 6,201 images, with 3,959 for training and 2,242 for testing where test-time object categories are not observed during training. + +Following the evaluation protocol [47, 52], we report Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) for counting. For detection, we report average precision with IoU thresholds of 0.5 and 0.75, denoted as AP50 and AP75, respectively, along with the averaged AP over IoU thresholds from 0.5 to 0.95 in increments of 0.05. + +# 6.2. Implementation details + +We use the pre-trained SAM-ViT/H [25] of the patch size 16 as the backbone and set it frozen during training, which returns the $64 \times 64 \times 256$ feature map. Due to the input patchification encapsulated in the ViT backbones [5], we find the receptive field of the raw feature map too coarse to detect small instances. We thus bilinearly interpolate the feature map resolution from $64 \times 64$ to $128 \times 128$ to produce a higher-resolution correlation map, which enables denser predictions and leads to better performance. (Tab. 11) + +The channel dimension is expanded to 512 via a learnable linear projection. TMR is trained with the learning rate of $10^{-4}$ with AdamW [39] and the batch size of 16 on four Nvidia RTX 3090 GPUs with 24 GB VRAM for 24 hours. The box presence threshold $\tau$ for NMS set to 0.4, 0.3 for RPINE and FSCD-147, respectively. + +# 6.3. Comparison with state-of-the-art methods + +RPINE. To demonstrate the effectiveness of our method in counting and detecting non-object patterns, we evaluate it on RPINE. In Tab. 2, TMR surpasses the previous FSCD methods with a large margin. As the previous FSCD methods rely on prototypes for matching, they tend to struggle with non-object patterns that require an understanding of spatial details rather than semantics. In contrast, TMR effectively detects non-object patterns by incorporating spatial details in the template matching process. Note that RPINE is the only dataset equipped with multiple pattern annotations for each image in FSCD. Figure 5 demonstrates the bounding box predictions with different exemplars, where TMR predicts bounding boxes adaptively to the given exemplars. Figure 7 shows the qualitative com + +
MethodSDMAE(↓)RMSE(↓)AP(↑)AP50(↑)AP75(↑)
C-DETR [47]9.5821.2413.8832.2010.22
SAM-C [42]✓18.7737.1418.8034.0418.74
PseCo [23]✓48.2088.1623.1844.5421.24
GeCo [52]✓9.5717.0723.3345.9321.19
TMR (ours)8.4519.8733.5964.0530.52
TMR (ours)✓8.3019.4029.6658.9425.41
+ +Table 2. One-shot pattern counting and detection results on the RPINE dataset. SD denotes box refinement with the SAM decoder. All the models are trained by the official code. + +parisons with other FSCD methods where TMR accurately localize target patterns. + +FSCD-LVIS and FSCD-147. We compare TMR with existing methods that are dedicated to FSCD under the FSCD setting on FSCD-LVIS and FSCD-147. The results in Tab. 3 demonstrate that TMR significantly outperforms previous state-of-the-art approaches. Notably, when evaluated on the unseen split, where test-time object categories are not observed during training, TMR surpasses prior methods. This suggests that TMR is less biased to object semantics during training potentially because TMR leverages the exemplar's spatial structure that generalizes across different categories. In addition to FSCD-LVIS, Tab. 6 compares methods on FSCD-147, where TMR performs on par with the previous methods on both one-shot and three-shot settings. + +# 6.4. Analyses and ablation study + +TMR learns with less semantic object bias and generalizes well across datasets. We compare TMR and GeCo [52] in the cross-dataset scenarios by evaluating the trained models on different datasets that are unseen during training. As shown in Tab. 4, TMR presents overwhelming performances, showing its strong generalization ability. Specifically, when GeCo is trained on FSCD-147, its performance drops significantly when evaluated on different datasets compared to when tested on FSCD-147 itself. GeCo, like previous FSCD methods, relies on prototypes for both counting and detection. We also suspect this lower generalization ability than ours originates from prototype matching, which is prone to overfitting to the training object semantics. GeCo struggles when evaluated on datasets with different object semantics. In contrast, TMR utilizes the structural information for matching instead of relying on semantic-intensive prototypes and generalizes more effectively on unseen datasets. + +SAM decoder is biased to edges. Table 2 shows the negative impact of the optional box refinement using the SAM decoder [25] on RPINE. Figure 6 visualizes two representative examples when the SAM decoder (SD) degrades performance. We observe that SD tends to align box predictions with the nearest edge, which is expected given that SAM is a segmentation model. The edge-sensitive predic + +
MethodSeenUnseen
AP(↑)AP50(↑)AP(↑)AP50(↑)
FSDetView-PB [68]2.727.571.032.89
AttRPN-PB [7]4.0811.153.157.87
C-DETR [47]4.9214.493.8511.28
DAVE [51]6.7522.514.1214.16
PseCo [23]22.3742.56--
GeCo [52]--11.4724.49
TMR (ours)27.4948.4822.7139.68
+ +Table 3. Three-shot counting detection-based methods on the FSCD-LVIS seen and unseen split. + +
TrainTestcross-evalAPAP50
GeCoTMRGeCoTMR
F-147F-14743.4244.4375.0673.83
F-LVISseen✓13.9621.2525.8737.18
RPINE✓19.4726.2138.6952.01
RPINEF-147✓36.9941.3960.3869.19
F-LVISseen✓10.0120.9217.4437.87
RPINE23.3329.6645.9358.94
+ +Table 4. Cross-dataset comparison of GeCo [52] and TMR, where F-147, F-LVIS indicate FSCD-147 and FSCD-LVIS. + +
MethodOne-shotThree-shot
AP(↑)AP50(↑)AP(↑)AP50(↑)
GeCo [52]32.7169.9532.4970.51
TMR (ours)36.0171.1938.5772.61
+ +Table 5. Comparison without optional SAM decoder on FSCD147 + +tion is particularly harmful for non-typical objects, such as objects that are loosely bounded by edges or patterns with many confusing edges. The SD refinement is seemingly good at edge detection for typical object exemplars with clear boundaries, and this is why the existing FSCD methods [23, 42, 52] benefit from adopting SD for box refinement. However, arbitrary patterns are not necessarily bounded by clear edges. Hence, edge-driven refinement may be even harmful as verified on RPINE. We emphasize that SD is an additional box post-processor, and we optionally add SD to TMR to compare with the SD-based state of the arts. Table 5 compares models by taking out SD, where TMR shows powerful performance. + +Effectiveness of template matching features. We verify the effectiveness of 2D template matching, which preserves the spatial structure of exemplar bounding boxes for correlation. We experiment with different input feature $F_{\mathrm{P}}$ to the box regression module. Table 7 compares our final model (e) and its variants. The lower bound model (a) observes zero information on the exemplar. Model (b) shows that correlating with 2D templates is greatly beneficial. The final model (e) concatenates the template-matching and the image feature, providing the correlation and appearance information for pattern matching regression. The models (c, + +
MethodSAM decoderOne-shotThree-shot
MAE(↓)RMSE(↓)AP(↑)AP50(↑)MAE(↓)RMSE(↓)AP(↑)AP50(↑)
FSDetView-PB [68]----37.54147.0713.4132.99
AttRPN-PB [7]----32.42141.5520.9737.19
C-DETR† [47]16.99125.2219.1447.6316.79123.5622.6650.57
SAM-C† [42]✓33.17141.7735.0956.0227.97131.2427.9949.17
PseCo [23]✓14.86118.6441.6370.8713.05112.8642.9873.33
DAVE [51]11.5486.6219.4655.2710.4574.5126.8162.82
GeCo [52]✓8.1060.1643.1174.317.9154.2843.4275.06
TMR (ours)✓11.6357.4643.1571.5513.7851.8744.4373.83
+ +![](images/497e318e60bb72db0a72d6f60209b99821c155be549bcfc9839b6a4783fee8cb.jpg) +Figure 6. TMR without vs. with SAM decoder [25] (w/o SD vs. w/ SD) for box refinement. SAM decoder tends to align box predictions with the nearest edges. The edge-sensitive prediction is harmful for non-objects, objects that have loose boundaries, or patterns with many confusing edges. The exemplar in the first row loosely contains an arched window, but the SAM decoder snaps the box tightly its edges. + +Table 6. Comparison with few-shot detection and counting methods on FSCD-147. The one-shot performance of the models with $\dagger$ are evaluated using the official code. + +
feature FPRPINEFSCD-147
APAP50APAP50
(a)F11.4424.0220.9555.59
(b)FTM32.5557.4331.9663.03
(c)F ⊕ FTM-cos29.7457.8524.7360.29
(d)F ⊕ FPM20.9447.9628.9166.93
(e)F ⊕ FTM33.5964.0536.0171.19
+ +d) replace the variants of the template matching feature $F_{\mathrm{TM}}$ with something else. Model (c) replaces the channel-wise correlation of (e) to cosine similarity, i.e., $F_{\mathrm{TM - cos}}$ , showing that retaining the channel dimension helps. Model (d) first performs average-pooling of the template and uses the prototype matching feature, $F_{\mathrm{PM}}$ , for correlation. Note that this average-pooled feature is often used in previous FSC methods [51, 52, 70]. The result shows that preserving the spatial structure of exemplar bounding boxes is more effective. The representative failure cases are shown in Fig. 11 in supp. material. The pooled prototype feature loses the geometric layout of the template and struggle to detect patterns where geometric clues are crucial. Comparing (d) and (e) verifies our hypothesis that prototype-based matching relies on the object priors; Model (d) drops more significantly on RPINE, which is less object-centric than FSCD-147. + +Table 7. Effect of template matching features (Sec. 4.1). $F_{\mathrm{PM}}$ is the correlation with average-pooled prototype. $F_{\mathrm{TM - cos}}$ is the cosine similarity of the template and image feature. + +
box regressionRPINEFSCD-147
shiftscaleAPAP50APAP50
(a)(0,0)(0,0)26.2055.2022.8157.30
(b)(Δx, Δy)(eαw, eαh)23.8851.2517.0155.14
(c)(Δx, Δy)(eαw sw, eαsh)31.2959.5235.0870.23
(d)(sw Δx, sh Δy)(eαw sw, eαsh)33.5964.0536.0171.19
+ +Table 8. Effect of box regression methods (Sec. 4.2) + +Effectiveness of support-conditioned regression. Table 8 compares different variants for box regression and validates the effectiveness of our support-conditioned box regression method. As shown in (b), regressing bounding boxes without considering the support exemplar's size leads to poor localization. Notably, it even performs worse than (a), which directly uses the exemplar box without performing any regression. However, as shown in (c), incorporating the support exemplar's size when predicting the width and height leads to a significant performance gain. This improvement is further enhanced by using the exemplar's size for shifting, as demonstrated in (d). These results demonstrate the effectiveness of support-conditioned regression, guided by the exemplar's size, in achieving accurate localization. + +Comparison on computational complexity. Table 9 compares the complexity of TMR with state-of-the-art FSCD methods, demonstrating its efficiency and effectiveness. Compared to PseCo [23], which introduces a large num + +![](images/302d43631ad77592c7dd51b860e7557e3433e8266c0d5788ef4bda6ee871a519.jpg) +Figure 7. Qualitative comparison with the state-of-the-art models on RPINE (the first two images) and FSCD-147 (the last two images). More visualization examples on three datasets are included in the supplementary materials Figs 16-18. + +
MethodTrainable paramsTotal paramsFLOPS
PseCo [23]56.99M0.70B5.08T
GeCo [52]7.98M0.65B4.72T
TMR (ours)19.01M0.66B3.04T
+ +Table 9. Comparison on the computational complexity + +ber of trainable parameters and FLOPs, TMR introduces only about 19M trainable parameters, which is detailed in Tab. 10. Although TMR has more trainable parameters than GeCo [52], its total parameter count remains comparable. Thanks to the simple architecture, TMR is significantly efficient in terms of FLOPs (3.04T), which is notably lower than PseCo (5.08T) and GeCo (4.72T). This reduction in computational cost not only enhances training efficiency but also results in faster inference. In contrast, the high FLOPs of PseCo and GeCo contribute to longer inference, making them less practical for real-time applications. + +Real-world application. We evaluate TMR, trained on RPINE, on scanning electron microscope (SEM) images [1], which are widely used in microprocessor inspection. Despite the domain shift, our method performs effectively and demonstrates its potential to generalize to real-world, non-object pattern scenarios (Fig. 8). + +# 7. Conclusion and discussion + +We have proposed a simple template-matching based method for few-shot pattern detection. We also introduce a new dataset with bounding box annotations that covers various patterns around the world across non-object patterns + +![](images/868c85f38ef62351767db810bbe2fb545d09d18a3d378ae85eaeedaa7d46ebef.jpg) +Figure 8. Application example of TMR on the SEM dataset [1]. + +to typical objects from various image domains from nature and human-made products. + +We observed that the widely used SAM decoder improves our performance by $7\%$ AP points on FSCD-147 by leveraging object-level edge priors, but it degrades performance on RPINE, which contains more general patterns. This suggests a promising future direction that combines the powerful pre-trained knowledge of foundation models with detection mechanisms less dependent on object-level edge priors. In parallel, exploring a lightweight, pattern-specific architecture that efficiently captures fine-grained repetitive structures without relying on object semantics could further improve generalization and efficiency. + +Acknowledgments. This work was supported by Samsung Electronics Co., Ltd (IO240508-09825-01) and IITP grants (RS-2022-II220959: Few-Shot Learning of Causal Inference in Vision & Language $(50\%)$ , RS-2022-II220113: Developing a Sustainable Collaborative Multi-modal Lifelong Learning Framework $(45\%)$ , RS-2019-II191906: AI Graduate School Program at POSTECH $(5\%))$ funded by Ministry of Science and ICT, Korea. We also appreciate Manjin Kim's advice on constructing RPINE. + +# References + +[1] Rossella Aversa, Mohammad Hadi Modarres, Stefano Cozzini, Regina Ciancio, and Alberto Chiusole. The first annotated set of scanning electron microscopy images for nanoscience. Scientific data, 5(1):1-10, 2018. 8 +[2] Yunliang Cai and George Baciu. Translation symmetry detection: A repetitive pattern analysis approach. In IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 223-228, 2013. 3 +[3] Cheng Chen, Wei Wang, John A Ozolek, and Gustavo K Rohde. A flexible and robust approach for segmenting cell nuclei from 2d microscopy images using supervised learning and template matching. Cytometry Part A, 83(5):495-507, 2013. 4 +[4] Taco S Cohen and Max Welling. Steerable cnns. arXiv preprint arXiv:1612.08498, 2016. 4 +[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5 +[6] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International Journal of Computer Vision (IJCV), 88:303-338, 2010. 1, 5 +[7] Qi Fan, Wei Zhuo, Chi-Keung Tang, and Yu-Wing Tai. Few-shot object detection with attention-rpn and multi-relation detector. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4013–4022, 2020. 1, 2, 6, 7 +[8] Zhibo Fan, Yuchen Ma, Zeming Li, and Jian Sun. Generalized few-shot object detection without forgetting. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4527-4536, 2021. 1, 2 +[9] Christopher Funk, Seungkyu Lee, Martin R Oswald, Stavros Tsogkas, Wei Shen, Andrea Cohen, Sven Dickinson, and Yanxi Liu. 2017 iccv challenge: Detecting symmetry in the wild. In International Conference on Computer Vision Workshops (ICCVW), pages 1692-1701, 2017. 5 +[10] Hiroyasu Furukawa, Kyle E Cordova, Michael O'Keeffe, and Omar M Yaghi. The chemistry and applications of metalorganic frameworks. Science, 341(6149):1230444, 2013. 4 +[11] Jizhou Gao, Yin Hu, Jinze Liu, and Ruigang Yang. Unsupervised learning of high-order structural semantics from images. In Proc. IEEE International Conference on Computer Vision (ICCV), pages 2122-2129. IEEE, 2009. 3 +[12] Ross Girshick. Fast r-cnn. In Proc. IEEE International Conference on Computer Vision (ICCV), pages 1440-1448, 2015. 1, 3 +[13] Alice L Givan. Flow cytometry: an introduction. Flow cytometry protocols, pages 1-29, 2011. 4 +[14] Branko Grünbaum and Geoffrey Colin Shephard. Tilings and patterns. Courier Dover Publications, 1987. 3 +[15] Karim Guirguis, Johannes Meier, George Eskandar, Matthias Kayser, Bin Yang, and Jürgen Beyerer. Niff: Alleviating + +forgetting in generalized few-shot object detection via neural instance feature forging. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 24193-24202, 2023. 2 +[16] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5356-5364, 2019. 1 +[17] Guangxing Han, Yicheng He, Shiyuan Huang, Jiawei Ma, and Shih-Fu Chang. Query adaptive few-shot object detection with heterogeneous graph convolutional networks. In Proc. IEEE International Conference on Computer Vision (ICCV), pages 3263-3272, 2021. 2 +[18] Guangxing Han, Shiyuan Huang, Jiawei Ma, Yicheng He, and Shih-Fu Chang. Meta faster r-cnn: Towards accurate few-shot object detection with attentive feature alignment. In Proc. AAAI Conference on Artificial Intelligence (AAAI), pages 780-789, 2022. 1 +[19] Guangxing Han, Jiawei Ma, Shiyuan Huang, Long Chen, and Shih-Fu Chang. Few-shot object detection with fully cross-transformer. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5321-5330, 2022. 2 +[20] Junwei Han, Stephen J McKenna, and Ruixuan Wang. Regular texture analysis as statistical model selection. In Proc. European Conference on Computer Vision (ECCV), pages 242-255. Springer, 2008. 3 +[21] James Hays, Marius Leordeanu, Alexei A Efros, and Yanxi Liu. Discovering texture regularity as a higher-order correspondence problem. In Proc. European Conference on Computer Vision (ECCV), pages 522-535. Springer, 2006. 3 +[22] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proc. IEEE International Conference on Computer Vision (ICCV), pages 2961-2969, 2017. 1, 3 +[23] Zhizhong Huang, Mingliang Dai, Yi Zhang, Junping Zhang, and Hongming Shan. Point segment and count: A generalized framework for object counting. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 3, 4, 6, 7, 8 +[24] Bingyi Kang, Zhuang Liu, Xin Wang, Fisher Yu, Jiashi Feng, and Trevor Darrell. Few-shot object detection via feature reweighting. In Proc. IEEE International Conference on Computer Vision (ICCV), pages 8420-8429, 2019. 2 +[25] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In Proc. IEEE International Conference on Computer Vision (ICCV), 2023. 3,4,5,6,7,2 +[26] Simon Korman, Daniel Reichman, Gilad Tsur, and Shai Avidan. Fast-match: Fast affine template matching. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2331-2338, 2013. 3, 4 +[27] Minji Lee, Hyunju Noh, and Youngsuk Kim. Diamidocarbene-derived palladium and nickel-sulfur clusters. Chemical Communications, 60(94):13867-13870, 2024. 4 + +[28] Thomas Leung and Jitendra Malik. Detecting, localizing and grouping repeated scene elements from an image. In Proc. European Conference on Computer Vision (ECCV). Springer, 1996. 3 +[29] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection, 2022. 3 +[30] Hui Lin, Xiaopeng Hong, and Yabin Wang. Object counting: You only need to look at one. arXiv preprint arXiv:2112.05993, 2021. 2 +[31] Hsin-Chih Lin, Ling-Ling Wang, and Shi-Nine Yang. Extracting periodicity of a regular texture based on autocorrelation functions. Pattern recognition letters, 18(5):433-443, 1997. 3 +[32] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proc. European Conference on Computer Vision (ECCV), pages 740-755. Springer, 2014. 1, 5 +[33] Tsung-Yi Lin, Piotr Dólár, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2117-2125, 2017. 1, 3 +[34] Wen-Chieh Lin and Yanxi Liu. Tracking dynamic near-regular texture under occlusion and rapid movements. In Proc. European Conference on Computer Vision (ECCV), pages 44–55. Springer, 2006. 3 +[35] Jingchen Liu and Yanxi Liu. Grasp recurring patterns from a single view. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2003-2010, 2013. 3 +[36] Jingchen Liu, George Slota, Gang Zheng, Zhaohui Wu, Minwoo Park, Seungkyu Lee, Ingmar Rauschert, and Yanxi Liu. Symmetry detection from realworld images competition 2013: Summary and results. In IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 200-205, 2013. 4, 5 +[37] Siying Liu, Tian-Tsong Ng, Kalyan Sunkavalli, Minh N Do, Eli Shechtman, and Nathan Carr. Patchmatch-based automatic lattice detection for near-regular textures. In Proc. IEEE International Conference on Computer Vision (ICCV), pages 181–189, 2015. 3 +[38] Yanxi Liu, Wen-Chieh Lin, and James Hays. Near-regular texture analysis and manipulation. ACM Transactions on Graphics (TOG), 23(3):368-376, 2004. 3, 5 +[39] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019. 5 +[40] Erika Lu, Weidi Xie, and Andrew Zisserman. Class-agnostic counting. In Asian Conference on Computer Vision (ACCV), pages 669–684. Springer, 2019. 2 +[41] Bronwyn A Lucas, Benjamin A Himes, Liang Xue, Timothy Grant, Julia Mahamid, and Nikolaus Grigorieff. Locating macromolecular assemblies in cells by 2d template matching with cistem. _Elife_, 10:e68946, 2021. 4 +[42] Zhiheng Ma, Xiaopeng Hong, and Qinnan Shangguan. Can sam count anything? an empirical study on sam counting, 2023. 3, 4, 6, 7 + +[43] Massimiliano Mancini, Elisa Ricci, Barbara Caputo, and Samuel Rota Bulo. Adding new tasks to a single network with weight transformations using binary masks. In European Conference on Computer Vision Workshops (ECCVW), pages 0-0, 2018. 4 +[44] Giovanni Marola. Using symmetry for detecting and locating objects in a picture. Computer Vision, Graphics, and Image Processing, 46(2):179-195, 1989. 3, 4 +[45] Jiri Matas, Ondrej Chum, Martin Urban, and Tomás Pajdla. Robust wide-baseline stereo from maximally stable extremal regions. Image and vision computing, 22(10):761-767, 2004. +[46] Anuj Mohan, Constantine Papageorgiou, and Tomaso Poggio. Example-based object detection in images by components. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 23(4):349-361, 2001. 3 +[47] Thanh Nguyen, Chau Pham, Khoi Nguyen, and Minh Hoai. Few-shot object counting and detection. In Proc. European Conference on Computer Vision (ECCV). Springer, 2022. 1, 2, 3, 4, 5, 6, 7 +[48] Roni Paiss, Ariel Ephrat, Omer Tov, Shiran Zada, Inbar Mosseri, Michal Irani, and Tali Dekel. Teaching clip to count to ten. In Proc. IEEE International Conference on Computer Vision (ICCV), pages 3170-3180, 2023. 4, 5 +[49] Minwoo Park, Kyle Brocklehurst, Robert T Collins, and Yanxi Liu. Deformed lattice detection in real-world images using mean-shift belief propagation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 31(10): 1804-1816, 2009. 3 +[50] Mark Pauly, Niloy J Mitra, Johannes Wallner, Helmut Pottmann, and Leonidas J Guibas. Discovering structural regularity in 3d geometry. In ACM SIGGRAPH, pages 1-11, 2008. 3 +[51] Jer Pelhan, Vitjan Zavrtanik, Matej Kristan, et al. Davea detect-and-verify paradigm for low-shot counting. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 23293-23302, 2024. 3, 6, 7 +[52] Jer Pelhan, Alan Lukezic, Vitjan Zavrtanik, and Matej Kristan. A novel unified architecture for low-shot counting by detection and segmentation. Advances in Neural Information Processing Systems (NeurIPS), 2025. 1, 2, 3, 4, 5, 6, 7, 8 +[53] Julien Picot, Coralie L Guerin, Caroline Le Van Kim, and Chantal M Boulanger. Flow cytometry: retrospective, fundamentals and recent instrumentation. Cytotechnology, 64: 109-130, 2012. 4 +[54] Muhammad Farrukh Qureshi, Zohaib Mushtaq, Muhammad Zia Ur Rehman, and Ernest Nlandu Kamavuako. E2cnn: An efficient concatenated cnn for classification of surface emg extracted from upper limb. IEEE Sensors Journal, 23(8): 8989-8996, 2023. 4 +[55] Viresh Ranjan and Minh Hoai. Vicinal counting networks. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4221-4230, 2022. 2 +[56] Viresh Ranjan, Udbhav Sharma, Thu Nguyen, and Minh Hoai. Learning to count everything. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3394-3403, 2021. 2, 1 + +[57] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 779-788, 2016. 1 +[58] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 658-666, 2019. 4 +[59] Babak Saleh and Ahmed Elgammal. Large-scale classification of fine-art paintings: Learning the right metric on the right feature. arXiv preprint arXiv:1505.00855, 2015. 4, 5 +[60] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 1 +[61] Luciano Spinello, Rudolph Triebel, Dizan Vasquez, Kai O Arras, and Roland Siegwart. Exploiting repetitive object patterns for model compression and completion. In Proc. European Conference on Computer Vision (ECCV), pages 296-309. Springer, 2010. 3 +[62] V.J. Stenger. Timeless Reality: Symmetry, Simplicity, and Multiple Universes. Prometheus, 2009. 5 +[63] Bo Sun, Banghuai Li, Shengcai Cai, Ye Yuan, and Chi Zhang. Fsce: Few-shot object detection via contrastive proposal encoding. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 7352-7362, 2021, 2 +[64] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In Proc. IEEE International Conference on Computer Vision (ICCV), pages 9627-9636, 2019. 1, 4 +[65] Du-Ming Tsai and Cheng-Huei Chiang. Rotation-invariant pattern matching using wavelet decomposition. Pattern Recognition Letters, 23(1-3):191-201, 2002. 3, 4 +[66] Andreas Turina, Tinne Tuytelaars, and Luc Van Gool. Efficient grouping under perspective skew. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages I-I. IEEE, 2001. 3 +[67] Xin Wang, Thomas Huang, Joseph Gonzalez, Trevor Darrell, and Fisher Yu. Frustratingly simple few-shot object detection. In Proc. International Conference on Machine Learning (ICML), pages 9919-9928. PMLR, 2020. 2 +[68] Yang Xiao, Vincent Lepetit, and Renaud Marlet. Few-shot object detection and viewpoint estimation for objects in the wild. TPAMI, 45(3):3090-3106, 2022. 1, 2, 6, 7 +[69] Xiaopeng Yan, Ziliang Chen, Anni Xu, Xiaoxi Wang, Xiaodan Liang, and Liang Lin. Meta r-cnn: Towards general solver for instance-level low-shot learning. In Proc. IEEE International Conference on Computer Vision (ICCV), pages 9577-9586, 2019. 1, 2 +[70] Zhiyuan You, Kai Yang, Wenhan Luo, Xin Lu, Lei Cui, and Xinyi Le. Iterative correlation-based feature refinement for few-shot counting. arXiv preprint arXiv:2201.08959, 4, 2022. 2, 7 +[71] Xinyu Zhang, Yuhan Liu, Yuting Wang, and Abdeslam Boularias. Detect everything with few examples. arXiv preprint arXiv:2309.12969, 2023. 2 \ No newline at end of file diff --git a/2025/Few-Shot Pattern Detection via Template Matching and Regression/images.zip b/2025/Few-Shot Pattern Detection via Template Matching and Regression/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8211abcbf06f1f19718517e1a4a90ecc79583523 --- /dev/null +++ b/2025/Few-Shot Pattern Detection via Template Matching and Regression/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ebf74cd387517228c646398230515b49e32889d9128c74f28e1407d643f83c06 +size 1000290 diff --git a/2025/Few-Shot Pattern Detection via Template Matching and Regression/layout.json b/2025/Few-Shot Pattern Detection via Template Matching and Regression/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fd78cc12ff4bc3fdf9b9a1a749ae9b848f8aa79c --- /dev/null +++ b/2025/Few-Shot Pattern Detection via Template Matching and Regression/layout.json @@ -0,0 +1,9270 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 97, + 103, + 512, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 103, + 512, + 121 + ], + "spans": [ + { + "bbox": [ + 97, + 103, + 512, + 121 + ], + "type": "text", + "content": "Few-Shot Pattern Detection via Template Matching and Regression" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 96, + 144, + 511, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 144, + 511, + 157 + ], + "spans": [ + { + "bbox": [ + 96, + 144, + 511, + 157 + ], + "type": "text", + "content": "Eunchan Jo Dahiyun Kang Sanghyun Kim Yunseon Choi Minsu Cho" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 127, + 162, + 483, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 162, + 483, + 176 + ], + "spans": [ + { + "bbox": [ + 127, + 162, + 483, + 176 + ], + "type": "text", + "content": "Pohang University of Science and Technology (POSTECH), South Korea" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 194, + 178, + 413, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 178, + 413, + 189 + ], + "spans": [ + { + "bbox": [ + 194, + 178, + 413, + 189 + ], + "type": "text", + "content": "https://cvlab.postech.ac.kr/research/TMR" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 217, + 200, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 217, + 200, + 229 + ], + "spans": [ + { + "bbox": [ + 151, + 217, + 200, + 229 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 243, + 297, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 243, + 297, + 495 + ], + "spans": [ + { + "bbox": [ + 54, + 243, + 297, + 495 + ], + "type": "text", + "content": "We address the problem of few-shot pattern detection, which aims to detect all instances of a given pattern, typically represented by a few exemplars, from an input image. Although similar problems have been studied in few-shot object counting and detection (FSCD), previous methods and their benchmarks have narrowed patterns of interest to object categories and often fail to localize non-object patterns. In this work, we propose a simple yet effective detector based on template matching and regression, dubbed TMR. While previous FSCD methods typically represent target exemplars as spatially collapsed prototypes and lose structural information, we revisit classic template matching and regression. It effectively preserves and leverages the spatial layout of exemplars through a minimalistic structure with a small number of learnable convolutional or projection layers on top of a frozen backbone. We also introduce a new dataset, dubbed RPINE, which covers a wider range of patterns than existing object-centric datasets. Our method outperforms the state-of-the-art methods on the three benchmarks, RPINE, FSCD-147, and FSCD-LVIS, and demonstrates strong generalization in cross-dataset evaluation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 513, + 135, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 513, + 135, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 513, + 135, + 525 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": "Few-shot detection aims to identify target patterns with minimal labeled examples. While significant progress has been made in few-shot object detection [7, 8, 18, 68, 69], most existing methods remain object-centric, focusing primarily on identifying object-level patterns with relatively clear boundaries. However, many real-world applications require detecting arbitrary target patterns that extend beyond objects to include structural, geometric, or abstract patterns across diverse visual data. Despite recent progress based on deep neural networks, current methods still fall short in addressing these broader pattern detection tasks. Furthermore, the object-centric design of conventional few-shot detectors may lead to performance degradation when the target object lacks clear boundaries or when occlusion and deformation cause its boundaries to become indistinct." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 322, + 215, + 429, + 312 + ], + "blocks": [ + { + "bbox": [ + 322, + 215, + 429, + 312 + ], + "lines": [ + { + "bbox": [ + 322, + 215, + 429, + 312 + ], + "spans": [ + { + "bbox": [ + 322, + 215, + 429, + 312 + ], + "type": "image", + "image_path": "5fe6ad08219b7d44758127fa1a9a891bc25f37fc8db852944122976036b023c4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 312, + 411, + 323 + ], + "lines": [ + { + "bbox": [ + 339, + 312, + 411, + 323 + ], + "spans": [ + { + "bbox": [ + 339, + 312, + 411, + 323 + ], + "type": "text", + "content": "one-shot exemplars" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 327, + 555, + 372 + ], + "lines": [ + { + "bbox": [ + 313, + 327, + 555, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 555, + 372 + ], + "type": "text", + "content": "Figure 1. Few-shot pattern detection. Given a few exemplar for each target pattern (left), the task is to detect all matching instances of each pattern (right). This example includes non-object patterns (e.g., green and blue) as well as object patterns (e.g., red)." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 438, + 215, + 545, + 311 + ], + "blocks": [ + { + "bbox": [ + 438, + 215, + 545, + 311 + ], + "lines": [ + { + "bbox": [ + 438, + 215, + 545, + 311 + ], + "spans": [ + { + "bbox": [ + 438, + 215, + 545, + 311 + ], + "type": "image", + "image_path": "2dd86fb3cd93c131095a899e5df13399f4d7197987db6270b1a06ced863f4d22.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 437, + 312, + 547, + 323 + ], + "lines": [ + { + "bbox": [ + 437, + 312, + 547, + 323 + ], + "spans": [ + { + "bbox": [ + 437, + 312, + 547, + 323 + ], + "type": "text", + "content": "ground-truth pattern detection" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 387, + 556, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 556, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 556, + 579 + ], + "type": "text", + "content": "The task of few-shot pattern detection is illustrated in Fig. 1. Recent related research topics for few-shot detection include few-shot counting and detection [47], and few-shot object detection [7]. Both aim to reduce the annotation cost of object detection [12, 33, 57, 64] by learning to detect all instances of given support exemplars. Consequently, many of these methods are heavily biased on object-centric benchmarks prior [6, 16, 32, 47]. In addition, many recent approaches [7, 18, 23, 52, 68, 69] represent the support exemplars as the spatially pooled vector, often named as a prototype [60]. While this pooling strategy is effective for detecting objects, it collapses the geometric properties, such as the shape and structure of the support exemplars. As a consequence, these methods tend to underperform when detecting non-object geometric patterns such as object parts or shape-intensive elements as shown in Fig. 3." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "type": "text", + "content": "In this work, we revisit the classic template matching strategy and propose a simple yet effective few-shot detector for arbitrary patterns. The proposed method, dubbed template matching and regression (TMR), is designed to be aware of the structure and shape of given exemplars. Given an input image, TMR first extracts a feature map using a backbone network. It then crops a template feature from the support exemplar's bounding box using a template extraction technique based on RoIAlign [22]. This template is correlated with the image feature map to produce a template matching feature map. Using this correlation map, the" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21578" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 55, + 550, + 216 + ], + "blocks": [ + { + "bbox": [ + 58, + 55, + 550, + 216 + ], + "lines": [ + { + "bbox": [ + 58, + 55, + 550, + 216 + ], + "spans": [ + { + "bbox": [ + 58, + 55, + 550, + 216 + ], + "type": "image", + "image_path": "65eba8dd183601e29102b9e3ae5c8647d189fc49c1c40666eb235d38d7c0bc15.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 219, + 555, + 265 + ], + "lines": [ + { + "bbox": [ + 55, + 219, + 555, + 265 + ], + "spans": [ + { + "bbox": [ + 55, + 219, + 555, + 265 + ], + "type": "text", + "content": "Figure 2. Annotation examples of FSCD-147 [47] and RPINE. FSCD-147 is annotated with the repetitive object-level patterns but disregards the repetition of non-object patterns such as the egg tray pattern under the eggs in the first image. RPINE is annotated with arbitrarily noticed repeated patterns which include non-object patterns and nameless parts of objects. Plus, FSCD-147 is annotated with a single pattern, while RPINE is annotated with all existing repetitions (marked with different colors) recognized by three different annotators." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 270, + 115, + 344 + ], + "blocks": [ + { + "bbox": [ + 58, + 270, + 115, + 344 + ], + "lines": [ + { + "bbox": [ + 58, + 270, + 115, + 344 + ], + "spans": [ + { + "bbox": [ + 58, + 270, + 115, + 344 + ], + "type": "image", + "image_path": "82a6181c922bf3664fa51dd561636a44a7fccb01cc60d33360a5bef96a04d6a6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 347, + 295, + 370 + ], + "lines": [ + { + "bbox": [ + 55, + 347, + 295, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 295, + 370 + ], + "type": "text", + "content": "Figure 3. Few-shot detectors trained with strong object prior are often biased to objects instances and struggle to detect non-objects." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 115, + 270, + 169, + 344 + ], + "blocks": [ + { + "bbox": [ + 115, + 270, + 169, + 344 + ], + "lines": [ + { + "bbox": [ + 115, + 270, + 169, + 344 + ], + "spans": [ + { + "bbox": [ + 115, + 270, + 169, + 344 + ], + "type": "image", + "image_path": "06d2c2087d961c1bd1a809ff91b41c64267633b76b15c576c64241e87c8bd70f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 170, + 269, + 225, + 344 + ], + "blocks": [ + { + "bbox": [ + 170, + 269, + 225, + 344 + ], + "lines": [ + { + "bbox": [ + 170, + 269, + 225, + 344 + ], + "spans": [ + { + "bbox": [ + 170, + 269, + 225, + 344 + ], + "type": "image", + "image_path": "578ef30a8b6e1c5d7eac771154410f8299ce3010aad6597453378ee5372df632.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 225, + 269, + 292, + 344 + ], + "blocks": [ + { + "bbox": [ + 225, + 269, + 292, + 344 + ], + "lines": [ + { + "bbox": [ + 225, + 269, + 292, + 344 + ], + "spans": [ + { + "bbox": [ + 225, + 269, + 292, + 344 + ], + "type": "image", + "image_path": "96b66ed01472422db90f254fc79143c02fcc0f3ba72853da1fa6d81cf924bedc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 377, + 295, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 377, + 295, + 461 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 295, + 461 + ], + "type": "text", + "content": "model learns bounding box regression parameters to rectify the support exemplar's box size adaptively. This process, termed support-conditioned regression, enables the model to handle support exemplars of varying sizes more effectively. Notably, TMR consists only of a few " + }, + { + "bbox": [ + 55, + 377, + 295, + 461 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 55, + 377, + 295, + 461 + ], + "type": "text", + "content": " and linear projections without any complicated modules such as cross-attention, commonly used in prior work [47, 52]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 462, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 605 + ], + "type": "text", + "content": "Although TMR is designed for general pattern detection, existing benchmarks (e.g., FSCD-147 [47], FSCD-LVIS [47]) mainly target object-level patterns, limiting comprehensive evaluation. To address this, we introduce a new dataset, Repeated Patterns IN Everywhere (RPINE), which covers diverse repeated patterns in the real world. RPINE contains images with varying degrees of objectness, from well-defined object-level patterns to non-object patterns, all annotated with bounding boxes via crowd-sourcing. Compared to FSCD datasets, RPINE provides broader coverage, including both non-object patterns and nameless parts of objects, as illustrated in Fig. 2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "text", + "content": "TMR demonstrates strong performance in detecting repeated patterns, not only on RPINE but also on the FSCD benchmarks, FSCD-147 and FSCD-LVIS [47]. In particular, TMR is especially effective on RPINE, includes diverse patterns with minimal object priors. Notably, our simple architecture contributes to improved generalization across datasets. Our contribution is summarized as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "- We generalize the few-shot object counting and detection to a pattern detection task that does not assume objectness" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 322, + 270, + 488, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 270, + 488, + 282 + ], + "spans": [ + { + "bbox": [ + 322, + 270, + 488, + 282 + ], + "type": "text", + "content": "in either the target patterns or exemplars." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 283, + 553, + 389 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 314, + 283, + 553, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 283, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 314, + 283, + 553, + 317 + ], + "type": "text", + "content": "- We present a simple yet effective pattern detector by refining template matching, which efficiently detects coherent patterns guided by exemplars." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 319, + 553, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 553, + 354 + ], + "type": "text", + "content": "- We introduce a new densely annotated dataset, RPINE, which covers diverse repetitive patterns in the real world, ranging from object-level patterns to non-object patterns." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 354, + 553, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 553, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 553, + 389 + ], + "type": "text", + "content": "- TMR not only outperforms the state-of-the-art FSCD models on RPINE and FSCD-LVIS but also achieves strong cross-dataset generalization." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 400, + 399, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 400, + 399, + 414 + ], + "spans": [ + { + "bbox": [ + 314, + 400, + 399, + 414 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 423, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 555, + 628 + ], + "type": "text", + "content": "Few-shot object detection (FSOD) aims to detect objects of novel classes using only a few support images of novel classes. Existing methods can be roughly categorized into two groups: finetuning based [8, 15, 63, 67, 68] and meta-learning based [7, 17-19, 24, 69, 71] approaches. Despite significant progress, finetuning methods require retraining for every new classes. For pattern detection, it is not desired to fine-tune each arbitrary pattern with indistinct object categorization. In contrast, meta-learning methods avoid class-specific fine-tuning and typically construct prototypes from support images for classifying the bounding boxes of the query image. These methods often apply global average pooling to obtain prototypes from each support exemplar, which collapses the spatial structure of the exemplar. While such spatially collapsed prototypes may effectively capture object-level bounding boxes, they struggle with non-object patterns that exhibit complex spatial structures." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "type": "text", + "content": "Few-shot counting (FSC) aims to count objects given a few exemplars (typically 1 to 3) within an image. Previous methods [30, 40, 55, 56, 70] have tackled this task using density-map regression without detecting the individual bounding boxes. Towards localization-based counting, few-shot counting and detection (FSCD) is proposed to combine FSC and detection. Such models include Counting-" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21579" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 59, + 549, + 178 + ], + "blocks": [ + { + "bbox": [ + 60, + 59, + 549, + 178 + ], + "lines": [ + { + "bbox": [ + 60, + 59, + 549, + 178 + ], + "spans": [ + { + "bbox": [ + 60, + 59, + 549, + 178 + ], + "type": "image", + "image_path": "b992c1cab19b66227e6eb39c4f3e257111f604c62d551c81b4dcbf705283a70e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 181, + 555, + 228 + ], + "lines": [ + { + "bbox": [ + 54, + 181, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 54, + 181, + 555, + 228 + ], + "type": "text", + "content": "Figure 4. Template Matching and Regression (TMR). A template feature map is cropped from the image feature map and then correlated with the image feature map via channel-wise template matching. The TM feature map and the image feature map are concatenated. For each feature map point, the box regressor predicts the shifting and scaling parameters of the support exemplar's size, and the presence classifier scores the presence map. Both of them consist of a " + }, + { + "bbox": [ + 54, + 181, + 555, + 228 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 54, + 181, + 555, + 228 + ], + "type": "text", + "content": " convolution and a linear projection without any complicated layers." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 235, + 294, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 235, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 54, + 235, + 294, + 331 + ], + "type": "text", + "content": "DETR [47], SAM-C [42], DAVE [51], PseCo [23], and GeCo [52]. However, similar to the FSOD methods, these models typically generate a prototype by spatially averaging exemplar bounding boxes. This discards the spatial structure of the exemplars, potentially losing important cues and details for accurate counting and detection. In contrast, our pattern-detection method preserves the spatial structure of the exemplar for pattern matching." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 333, + 295, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 333, + 295, + 405 + ], + "spans": [ + { + "bbox": [ + 54, + 333, + 295, + 405 + ], + "type": "text", + "content": "Template matching. Template matching [21, 26, 44-46, 65, 66] has been widely used from the beginning of computer vision and pattern recognition and also adopted in convolution-based neural detectors with additional regression [12, 33]. Given a 2D template, template matching identifies the matching region by sliding-window fashion." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 408, + 295, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 408, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 54, + 408, + 295, + 540 + ], + "type": "text", + "content": "Repetitive pattern detection. Detecting repetitive patterns is trivial for humans but remains challenging in computer vision. Early research focused on geometrically constrained settings where patterns are nearly regular and aligned [14, 20, 31, 50]. Based on this near-regularity, early methods assumed a global repetitive structure and discovered the repetition [21, 31, 37, 38, 49] in a top-down manner. In contrast, later methods [11, 34, 61] define the smallest repeating unit, called a texel [2, 28, 35] and identify all matching subparts based on the identified texels. In our context, the texel corresponds to the given exemplar." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 550, + 208, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 550, + 208, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 550, + 208, + 563 + ], + "type": "text", + "content": "3. Few-shot pattern detection" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 569, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 296, + 713 + ], + "type": "text", + "content": "Given an input image " + }, + { + "bbox": [ + 55, + 569, + 296, + 713 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{H_0 \\times W_0 \\times 3}" + }, + { + "bbox": [ + 55, + 569, + 296, + 713 + ], + "type": "text", + "content": ", the goal of few-shot pattern detection is to predict matching patterns with a given set of support exemplars. With an abuse of notations, the model is given the set of support exemplar, " + }, + { + "bbox": [ + 55, + 569, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = \\{e_1, \\dots, e_{N_s}\\}" + }, + { + "bbox": [ + 55, + 569, + 296, + 713 + ], + "type": "text", + "content": " and aims to predict the corresponding ground-truth bounding boxes of pattern " + }, + { + "bbox": [ + 55, + 569, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{B} = \\{b_1, \\dots, b_{N_p}\\}" + }, + { + "bbox": [ + 55, + 569, + 296, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 569, + 296, + 713 + ], + "type": "inline_equation", + "content": "N_p" + }, + { + "bbox": [ + 55, + 569, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 569, + 296, + 713 + ], + "type": "inline_equation", + "content": "N_s" + }, + { + "bbox": [ + 55, + 569, + 296, + 713 + ], + "type": "text", + "content": " denotes the number of ground-truth bounding boxes of the pattern and the number of exemplars (typically referred to as the \"shot\"), respectively. The input exemplars and the output patterns are both represented as bounding boxes parameterized by their center coordinates, width, and height: " + }, + { + "bbox": [ + 55, + 569, + 296, + 713 + ], + "type": "inline_equation", + "content": "b_i, e_i \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 55, + 569, + 296, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 234, + 544, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 234, + 544, + 248 + ], + "spans": [ + { + "bbox": [ + 313, + 234, + 544, + 248 + ], + "type": "text", + "content": "4. Template matching and regression (TMR)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 255, + 553, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 553, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 553, + 314 + ], + "type": "text", + "content": "For clarity, we primarily focus on the one-shot, single-scale pattern detection setting, where a single support exemplar is given and a single-resolution image feature map is used. However, TMR can be effortlessly extended to few-shot and multi-scale scenarios, as described in Sec.4.3 and Sec.9." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 315, + 555, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 555, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 555, + 626 + ], + "type": "text", + "content": "The overall architecture of TMR is illustrated in Fig. 4. The input image " + }, + { + "bbox": [ + 313, + 315, + 555, + 626 + ], + "type": "inline_equation", + "content": "\\pmb{I}" + }, + { + "bbox": [ + 313, + 315, + 555, + 626 + ], + "type": "text", + "content": " is first encoded by a backbone such as ViT [25] to extract a feature map " + }, + { + "bbox": [ + 313, + 315, + 555, + 626 + ], + "type": "inline_equation", + "content": "\\pmb{F} \\in \\mathbb{R}^{H \\times W \\times D}" + }, + { + "bbox": [ + 313, + 315, + 555, + 626 + ], + "type": "text", + "content": ". The template feature " + }, + { + "bbox": [ + 313, + 315, + 555, + 626 + ], + "type": "inline_equation", + "content": "\\pmb{T} \\in \\mathbb{R}^{t_h \\times t_w \\times D}" + }, + { + "bbox": [ + 313, + 315, + 555, + 626 + ], + "type": "text", + "content": " is then obtained through the template extraction process, which uses RoIAlign [22] to crop a region from the image feature map " + }, + { + "bbox": [ + 313, + 315, + 555, + 626 + ], + "type": "inline_equation", + "content": "\\pmb{F}" + }, + { + "bbox": [ + 313, + 315, + 555, + 626 + ], + "type": "text", + "content": " based on the support exemplar's size " + }, + { + "bbox": [ + 313, + 315, + 555, + 626 + ], + "type": "inline_equation", + "content": "(s_h, s_w)" + }, + { + "bbox": [ + 313, + 315, + 555, + 626 + ], + "type": "text", + "content": ". In RoIAlign, unlike previous fixed-size pooling methods [47, 52], the model adaptively determines the size of " + }, + { + "bbox": [ + 313, + 315, + 555, + 626 + ], + "type": "inline_equation", + "content": "\\pmb{T}" + }, + { + "bbox": [ + 313, + 315, + 555, + 626 + ], + "type": "text", + "content": " to fit the corresponding size on " + }, + { + "bbox": [ + 313, + 315, + 555, + 626 + ], + "type": "inline_equation", + "content": "\\pmb{F}" + }, + { + "bbox": [ + 313, + 315, + 555, + 626 + ], + "type": "text", + "content": " by rounding up the size of the exemplars on " + }, + { + "bbox": [ + 313, + 315, + 555, + 626 + ], + "type": "inline_equation", + "content": "\\pmb{F}" + }, + { + "bbox": [ + 313, + 315, + 555, + 626 + ], + "type": "text", + "content": " as described in Sec. 8.2. This preserves the spatial alignment between " + }, + { + "bbox": [ + 313, + 315, + 555, + 626 + ], + "type": "inline_equation", + "content": "\\pmb{T}" + }, + { + "bbox": [ + 313, + 315, + 555, + 626 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 315, + 555, + 626 + ], + "type": "inline_equation", + "content": "\\pmb{F}" + }, + { + "bbox": [ + 313, + 315, + 555, + 626 + ], + "type": "text", + "content": " with translation. The image feature map " + }, + { + "bbox": [ + 313, + 315, + 555, + 626 + ], + "type": "inline_equation", + "content": "\\pmb{F}" + }, + { + "bbox": [ + 313, + 315, + 555, + 626 + ], + "type": "text", + "content": " and template feature " + }, + { + "bbox": [ + 313, + 315, + 555, + 626 + ], + "type": "inline_equation", + "content": "\\pmb{T}" + }, + { + "bbox": [ + 313, + 315, + 555, + 626 + ], + "type": "text", + "content": " are correlated by template matching (Sec. 4.1), which outputs the template-matching feature " + }, + { + "bbox": [ + 313, + 315, + 555, + 626 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{TM}}" + }, + { + "bbox": [ + 313, + 315, + 555, + 626 + ], + "type": "text", + "content": ". The concatenation of the feature maps " + }, + { + "bbox": [ + 313, + 315, + 555, + 626 + ], + "type": "inline_equation", + "content": "\\pmb{F}" + }, + { + "bbox": [ + 313, + 315, + 555, + 626 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 315, + 555, + 626 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{TM}}" + }, + { + "bbox": [ + 313, + 315, + 555, + 626 + ], + "type": "text", + "content": " is fed to the subsequent box prediction module, which consists of a pattern box regressor and a pattern presence classifier. The pattern box regressor predicts the localization bounding boxes: " + }, + { + "bbox": [ + 313, + 315, + 555, + 626 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{B}}" + }, + { + "bbox": [ + 313, + 315, + 555, + 626 + ], + "type": "text", + "content": ", which is parameterized by scaling and shifting factors of the given exemplar's size. The pattern presence classifier predicts the presence score of the pattern: " + }, + { + "bbox": [ + 313, + 315, + 555, + 626 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{P}}" + }, + { + "bbox": [ + 313, + 315, + 555, + 626 + ], + "type": "text", + "content": ". A box proposal is then generated on each feature map point based on the combination of " + }, + { + "bbox": [ + 313, + 315, + 555, + 626 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{B}}" + }, + { + "bbox": [ + 313, + 315, + 555, + 626 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 315, + 555, + 626 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{P}}" + }, + { + "bbox": [ + 313, + 315, + 555, + 626 + ], + "type": "text", + "content": " (Sec. 4.2). At inference, bounding boxes with low presence scores are removed by Non-Maximum Suppression (NMS). SAM decoder [25] can be optionally applied for box refinement before NMS." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 635, + 455, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 455, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 455, + 647 + ], + "type": "text", + "content": "4.1. Template matching (TM)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "We are motivated to detect arbitrary patterns given a support exemplar of either an object or not. Non-object patterns often lack high-level semantics yet exhibit low-level structural features, thus, preserving the spatial structure of the support exemplar is crucial. Inspired by traditional tem" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21580" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "content": "plate matching [26, 44, 65], we compute the matching score between the image and the support exemplar to detect the locations of the pattern. Specifically, template matching cross-correlates the feature map " + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "content": " and the template feature " + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "content": " by centering " + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "content": " at each " + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "content": " position in " + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "content": ". The resultant template-matching (TM) feature " + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{TM}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "content": " is obtained as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 150, + 298, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 150, + 298, + 190 + ], + "spans": [ + { + "bbox": [ + 55, + 150, + 298, + 190 + ], + "type": "interline_equation", + "content": "\\boldsymbol {F} _ {\\mathrm {T M}} (x, y) = \\frac {1}{t _ {w} t _ {h}} \\sum_ {x ^ {\\prime}, y ^ {\\prime}} \\boldsymbol {F} \\left(x + x ^ {\\prime} - \\lfloor \\frac {t _ {w}}{2} \\rfloor , y + y ^ {\\prime} - \\lfloor \\frac {t _ {h}}{2} \\rfloor\\right) \\boldsymbol {T} \\left(x ^ {\\prime}, y ^ {\\prime}\\right), \\tag {1}", + "image_path": "93a163786c10aed82fb26c4e3f50e29388653042dd88a7ba11359110caf7b2f5.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 191, + 296, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 191, + 296, + 286 + ], + "spans": [ + { + "bbox": [ + 55, + 191, + 296, + 286 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 191, + 296, + 286 + ], + "type": "inline_equation", + "content": "\\lfloor \\cdot \\rfloor" + }, + { + "bbox": [ + 55, + 191, + 296, + 286 + ], + "type": "text", + "content": " denotes the floor operation used for centering " + }, + { + "bbox": [ + 55, + 191, + 296, + 286 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 191, + 296, + 286 + ], + "type": "text", + "content": " at each " + }, + { + "bbox": [ + 55, + 191, + 296, + 286 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 55, + 191, + 296, + 286 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 191, + 296, + 286 + ], + "type": "inline_equation", + "content": "(x^{\\prime},y^{\\prime})" + }, + { + "bbox": [ + 55, + 191, + 296, + 286 + ], + "type": "text", + "content": " ranges within the template coordinates: " + }, + { + "bbox": [ + 55, + 191, + 296, + 286 + ], + "type": "inline_equation", + "content": "\\in [0,t_w)\\times [0,t_h)" + }, + { + "bbox": [ + 55, + 191, + 296, + 286 + ], + "type": "text", + "content": ". Note that Eq. 1 computes the correlation by channel-wise multiplication, resulting in " + }, + { + "bbox": [ + 55, + 191, + 296, + 286 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{TM}}\\in \\mathbb{R}^{H\\times W\\times D}" + }, + { + "bbox": [ + 55, + 191, + 296, + 286 + ], + "type": "text", + "content": ". Finally, the TM feature " + }, + { + "bbox": [ + 55, + 191, + 296, + 286 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{TM}}" + }, + { + "bbox": [ + 55, + 191, + 296, + 286 + ], + "type": "text", + "content": " is concatenated with the feature map " + }, + { + "bbox": [ + 55, + 191, + 296, + 286 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 191, + 296, + 286 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 55, + 191, + 296, + 286 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{P}} = [F_{\\mathrm{TM}};F]\\in \\mathbb{R}^{H\\times W\\times 2D}" + }, + { + "bbox": [ + 55, + 191, + 296, + 286 + ], + "type": "text", + "content": ". The concatenated feature map " + }, + { + "bbox": [ + 55, + 191, + 296, + 286 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{P}}" + }, + { + "bbox": [ + 55, + 191, + 296, + 286 + ], + "type": "text", + "content": " is fed to the subsequent box prediction module." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 293, + 246, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 246, + 306 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 246, + 306 + ], + "type": "text", + "content": "4.2. Support-conditioned box regression" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 310, + 296, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 310, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 310, + 296, + 346 + ], + "type": "text", + "content": "Our box prediction module consists of a pattern box regressor and a pattern presence classifier following the architecture of an anchor-free detection methods [64]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 346, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 296, + 479 + ], + "type": "text", + "content": "The pattern box regressor " + }, + { + "bbox": [ + 55, + 346, + 296, + 479 + ], + "type": "inline_equation", + "content": "g_{\\mathrm{B}}" + }, + { + "bbox": [ + 55, + 346, + 296, + 479 + ], + "type": "text", + "content": " consists of a " + }, + { + "bbox": [ + 55, + 346, + 296, + 479 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 55, + 346, + 296, + 479 + ], + "type": "text", + "content": " convolutional layer followed by a linear layer and predicts the four localization parameters " + }, + { + "bbox": [ + 55, + 346, + 296, + 479 + ], + "type": "inline_equation", + "content": "(\\Delta x, \\Delta y, \\alpha_w, \\alpha_h)" + }, + { + "bbox": [ + 55, + 346, + 296, + 479 + ], + "type": "text", + "content": " for each feature map point. Unlike the methods [47, 52] that directly regress absolute box parameters, our method performs support-conditioned regression. Bounding box parameters are predicted by scaling and shifting the support exemplar's size. This helps to dynamically adjust the exemplar's size to predict a target box size. A predicted bounding box at a feature point " + }, + { + "bbox": [ + 55, + 346, + 296, + 479 + ], + "type": "inline_equation", + "content": "(x, y)" + }, + { + "bbox": [ + 55, + 346, + 296, + 479 + ], + "type": "text", + "content": " shifts and scales the support exemplar's size " + }, + { + "bbox": [ + 55, + 346, + 296, + 479 + ], + "type": "inline_equation", + "content": "(s_w, s_h)" + }, + { + "bbox": [ + 55, + 346, + 296, + 479 + ], + "type": "text", + "content": " such as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 94, + 486, + 294, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 486, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 94, + 486, + 294, + 498 + ], + "type": "interline_equation", + "content": "\\left(x + s _ {w} \\Delta x, y + s _ {h} \\Delta y, e ^ {\\alpha_ {w}} s _ {w}, e ^ {\\alpha_ {h}} s _ {h}\\right). \\tag {2}", + "image_path": "3333d20784e9547eacaa14f45495ecde2ff30b9b1cc2da7de83cbd8e3306a7d8.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 506, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 296, + 567 + ], + "type": "text", + "content": "The pattern presence classifier " + }, + { + "bbox": [ + 55, + 506, + 296, + 567 + ], + "type": "inline_equation", + "content": "g_{\\mathrm{P}}" + }, + { + "bbox": [ + 55, + 506, + 296, + 567 + ], + "type": "text", + "content": " consists of a linear layer and predicts presence scores, which represent the confidence of the predicted bounding boxes at each feature map point. The aforementioned procedure of the box regressor " + }, + { + "bbox": [ + 55, + 506, + 296, + 567 + ], + "type": "inline_equation", + "content": "g_{\\mathrm{B}}" + }, + { + "bbox": [ + 55, + 506, + 296, + 567 + ], + "type": "text", + "content": " and the presence classifier " + }, + { + "bbox": [ + 55, + 506, + 296, + 567 + ], + "type": "inline_equation", + "content": "g_{\\mathrm{P}}" + }, + { + "bbox": [ + 55, + 506, + 296, + 567 + ], + "type": "text", + "content": " is summarized as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 102, + 569, + 294, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 569, + 294, + 583 + ], + "spans": [ + { + "bbox": [ + 102, + 569, + 294, + 583 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {B}} = g _ {\\mathrm {B}} \\left(\\boldsymbol {F} _ {\\mathrm {P}}\\right), \\quad \\hat {\\boldsymbol {B}} \\in \\mathbb {R} ^ {H \\times W \\times 4}, \\tag {3}", + "image_path": "b75fc5f7287f42f2a0276cf414c5ba8c107c151e0bf72b04bf7453fcfa961358.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 103, + 586, + 294, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 586, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 103, + 586, + 294, + 601 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {P}} = \\sigma \\left(g _ {\\mathrm {P}} \\left(\\boldsymbol {F} _ {\\mathrm {P}}\\right)\\right), \\quad \\hat {\\boldsymbol {P}} \\in \\mathbb {R} ^ {H \\times W \\times 1}, \\tag {4}", + "image_path": "5335eb07db9755f32735958ea6762517a298c21b9fb3153d4440d4f48760de31.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 58, + 609, + 215, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 609, + 215, + 620 + ], + "spans": [ + { + "bbox": [ + 58, + 609, + 215, + 620 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 58, + 609, + 215, + 620 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 58, + 609, + 215, + 620 + ], + "type": "text", + "content": " denotes the sigmoid function." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 624, + 123, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 123, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 123, + 635 + ], + "type": "text", + "content": "4.3. Inference" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "At inference, we first remove bounding boxes whose presence score is lower than a threshold " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " to filter out low-confidence predictions. Afterward, we can optionally apply box localization refinement via the SAM decoder [25] as often adopted by the FSCD work [23, 42, 52]. Following [23, 52], we input the predicted box coordinates into the" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "SAM prompt encoder to obtain the prompt feature. The obtained prompt feature and the image feature extracted from the SAM backbone are fed to the SAM decoder, which further refines input box coordinates. To obtain the final box prediction, " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{B}}" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": ", we apply NMS on the bounding boxes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 133, + 554, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 554, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 554, + 180 + ], + "type": "text", + "content": "TMR is easily extended for few-shot inference without re-training. When multiple support exemplars are given, we perform the above process for each exemplar individually and then aggregate the results before applying NMS." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 186, + 425, + 199 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 186, + 425, + 199 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 425, + 199 + ], + "type": "text", + "content": "4.4. Learning objective" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 204, + 555, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 555, + 323 + ], + "type": "text", + "content": "The training loss is composed of " + }, + { + "bbox": [ + 313, + 204, + 555, + 323 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{P}}" + }, + { + "bbox": [ + 313, + 204, + 555, + 323 + ], + "type": "text", + "content": ", which penalizes the presence score, and " + }, + { + "bbox": [ + 313, + 204, + 555, + 323 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{B}}" + }, + { + "bbox": [ + 313, + 204, + 555, + 323 + ], + "type": "text", + "content": ", which penalizes the bounding box regression. The presence loss " + }, + { + "bbox": [ + 313, + 204, + 555, + 323 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{P}}" + }, + { + "bbox": [ + 313, + 204, + 555, + 323 + ], + "type": "text", + "content": " is the binary cross-entropy loss (BCELoss) between the predicted and the ground-truth center points of the pattern bounding boxes. Instead of guiding the presence with a single pixel, we add a margin around each ground-truth center point and define the extended center point set as " + }, + { + "bbox": [ + 313, + 204, + 555, + 323 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{\\mathrm{P}}" + }, + { + "bbox": [ + 313, + 204, + 555, + 323 + ], + "type": "text", + "content": ". We detail this in Sec. 8.3. The ground-truth presence label " + }, + { + "bbox": [ + 313, + 204, + 555, + 323 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 204, + 555, + 323 + ], + "type": "text", + "content": " at a position " + }, + { + "bbox": [ + 313, + 204, + 555, + 323 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 313, + 204, + 555, + 323 + ], + "type": "text", + "content": " is set to 1 if the position is in " + }, + { + "bbox": [ + 313, + 204, + 555, + 323 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{\\mathrm{P}}" + }, + { + "bbox": [ + 313, + 204, + 555, + 323 + ], + "type": "text", + "content": ", and 0 otherwise." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 324, + 554, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 554, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 554, + 419 + ], + "type": "text", + "content": "The bounding box regression loss " + }, + { + "bbox": [ + 313, + 324, + 554, + 419 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{B}}" + }, + { + "bbox": [ + 313, + 324, + 554, + 419 + ], + "type": "text", + "content": " is defined as the generalized IoU (gIoU) loss [58] between the predicted and the ground-truth bounding boxes. This loss is defined on the positions where the ground-truth patterns exist. For each center point of the pattern bounding box, the adjacent area of the center within the margin shares the corresponding bounding box parameter " + }, + { + "bbox": [ + 313, + 324, + 554, + 419 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 324, + 554, + 419 + ], + "type": "text", + "content": ". The two loss functions are defined as the following:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 344, + 424, + 554, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 424, + 554, + 460 + ], + "spans": [ + { + "bbox": [ + 344, + 424, + 554, + 460 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {P}} = \\sum_ {(x, y)} ^ {\\forall (x, y)} \\operatorname {B C E L o s s} (\\hat {\\boldsymbol {P}} (x, y), \\boldsymbol {P} (x, y)), \\tag {5}", + "image_path": "403cc07481ac3b618a2f33746226320fe2544873128eb62df16a599e7f0a3b02.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 344, + 464, + 553, + 499 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 464, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 344, + 464, + 553, + 499 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {B}} = \\sum_ {(x, y)} ^ {(x, y) \\in \\mathcal {X} _ {\\mathrm {P}}} \\operatorname {g I o U L o s s} (\\hat {\\boldsymbol {B}} (x, y), \\boldsymbol {B} (x, y)). \\tag {6}", + "image_path": "888b0a7e6717dbdf71da7cbaf3f6d9867105152218f57f7608c4e516bb2e418f.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 506, + 553, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 506, + 553, + 519 + ], + "spans": [ + { + "bbox": [ + 316, + 506, + 553, + 519 + ], + "type": "text", + "content": "The overall training loss is the sum of them: " + }, + { + "bbox": [ + 316, + 506, + 553, + 519 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathcal{L}_{\\mathrm{P}} + \\mathcal{L}_{\\mathrm{B}}" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 526, + 462, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 462, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 462, + 539 + ], + "type": "text", + "content": "5. Proposed dataset: RPINE" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 545, + 554, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 594 + ], + "type": "text", + "content": "Existing benchmarks (e.g., FSCD-147 [47], FSCD-LVIS [47]) focus on object-level patterns, limiting their use for general pattern detection. Therefore, we introduce a new pattern dataset, RPINE: Repeated Patterns In Everywhere." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": "We collect images from the repetitive pattern detection literature and annotate them to contain various repetitive patterns in the wild. Images are collected from FSC-147 [47], FSCD-LVIS [47], Countbench [48], Wikiart [59], Frieze [43], and Wallpaper [36]. Tab. 1 shows where RPINE stands among related datasets. The dataset contains 4,362 images, divided into 3,925 and 435 for training and testing, respectively. As shown in Fig. 5, RPINE is annotated with multiple patterns per image. RPINE is a suitable evaluation benchmark for multi-pattern detection within an image." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21581" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 92, + 57, + 517, + 127 + ], + "blocks": [ + { + "bbox": [ + 92, + 57, + 517, + 127 + ], + "lines": [ + { + "bbox": [ + 92, + 57, + 517, + 127 + ], + "spans": [ + { + "bbox": [ + 92, + 57, + 517, + 127 + ], + "type": "table", + "html": "
datasetpattern tilingrepetitionobject biasmulti-pattern annotations per image
Wallpaper [36]near regular [38]highlow
Pascal [6], COCO [32]-lowhigh✓ (multi-category)
Wikiart [59], Frieze [9], Counting bench [48]arbitrarymidhigh
FSCD-147 [47], FSCD-LVIS [47]arbitraryhighhigh
RPINEarbitraryhighlow to high✓
", + "image_path": "3de0463b81516c34eace6e9a91c644bc091eadaac466ae5bafa05efed803ee39.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 130, + 555, + 175 + ], + "lines": [ + { + "bbox": [ + 55, + 130, + 555, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 130, + 555, + 175 + ], + "type": "text", + "content": "Table 1. Benchmark dataset comparison of related work. RPINE covers pattern or object repetitions in the real world. We classify 'pattern- tiling' as near-regular following [38], i.e., elements with an almost periodic lattice with minor variations in shape, color or lighting, and as arbitrary otherwise. The 'repetition' is classified based on the average number of repeated same class instances on an image: low if " + }, + { + "bbox": [ + 55, + 130, + 555, + 175 + ], + "type": "inline_equation", + "content": "\\leq 5" + }, + { + "bbox": [ + 55, + 130, + 555, + 175 + ], + "type": "text", + "content": " , mid if " + }, + { + "bbox": [ + 55, + 130, + 555, + 175 + ], + "type": "inline_equation", + "content": "\\leq {20}" + }, + { + "bbox": [ + 55, + 130, + 555, + 175 + ], + "type": "text", + "content": " ,and high otherwise. The 'object-bias' is high if the dataset annotations correspond to a predefined class set, and low otherwise." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 57, + 182, + 293, + 293 + ], + "blocks": [ + { + "bbox": [ + 57, + 182, + 293, + 293 + ], + "lines": [ + { + "bbox": [ + 57, + 182, + 293, + 293 + ], + "spans": [ + { + "bbox": [ + 57, + 182, + 293, + 293 + ], + "type": "image", + "image_path": "c9c8dd92228de1ad54d4c11a3f021f428bdc54e24d9c22c59d9bb909f34160e1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 297, + 295, + 331 + ], + "lines": [ + { + "bbox": [ + 55, + 297, + 295, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 295, + 331 + ], + "type": "text", + "content": "Figure 5. GTs (blue) and TMR predictions (green) with different exemplars (red) from RPINE, which is the only dataset containing multiple patterns for each image among FSCD datasets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 340, + 295, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 295, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 295, + 388 + ], + "type": "text", + "content": "In the real world, a pattern cannot be rigorously defined. We thus define the following criteria inspired by the definition of the translation symmetry [62] to minimize the subjective variance among different annotators." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 388, + 295, + 532 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 55, + 388, + 295, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 295, + 412 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 295, + 412 + ], + "type": "text", + "content": "- Number of patterns: max 3 different patterns are annotated per image if the image exhibits multiple patterns." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 413, + 264, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 413, + 264, + 423 + ], + "spans": [ + { + "bbox": [ + 56, + 413, + 264, + 423 + ], + "type": "text", + "content": "- Number of pattern instances: has no upperbound." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 425, + 295, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 425, + 295, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 425, + 295, + 448 + ], + "type": "text", + "content": "- Minimum Size: the width and height of a pattern instance must be at least " + }, + { + "bbox": [ + 56, + 425, + 295, + 448 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 56, + 425, + 295, + 448 + ], + "type": "text", + "content": " of the shorter side of the image." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 449, + 295, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 295, + 483 + ], + "type": "text", + "content": "- Appearance variance: visually similar patterns with different scales/colors/semantics/rotation angles are annotated as the same pattern." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 484, + 295, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 484, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 56, + 484, + 295, + 508 + ], + "type": "text", + "content": "- Reflection: visually similar patterns but reflection symmetric patterns are annotated in different patterns." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 509, + 295, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 509, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 56, + 509, + 295, + 532 + ], + "type": "text", + "content": "- Occlusion: if visually similar patterns are occluded from each other, the visible parts are annotated." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 533, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 605 + ], + "type": "text", + "content": "We ask annotators to carefully draw bounding boxes on the recognized patterns by the above instructions as consistently as they can. Despite of the instruction details, pattern are not pixel-perfectly defined across annotators. Therefore, we assign three individual annotators per image and include all the annotated patterns as ground truth." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 616, + 137, + 629 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 137, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 137, + 629 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 635, + 171, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 171, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 171, + 647 + ], + "type": "text", + "content": "6.1. Dataset and metrics" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "content": "We evaluate our model on RPINE as well as on the two standard FSCD benchmarks: FSCD-147 [47] and FSCDLVIS [47]. FSCD-147 contains a total of 6,135 images, with 3,659 for training, 1,286 for validation, and 1,190 for testing. FSCD-LVIS seen-split contains a total of 6,195 im" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 184, + 553, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 184, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 184, + 553, + 243 + ], + "type": "text", + "content": "ages, with 4,000 for training, 1,181 for validation, and 1,014 for testing, covering 372 object categories. FSCD-LVIS unseen-split contains a total of 6,201 images, with 3,959 for training and 2,242 for testing where test-time object categories are not observed during training." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 244, + 554, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 244, + 554, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 244, + 554, + 316 + ], + "type": "text", + "content": "Following the evaluation protocol [47, 52], we report Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) for counting. For detection, we report average precision with IoU thresholds of 0.5 and 0.75, denoted as AP50 and AP75, respectively, along with the averaged AP over IoU thresholds from 0.5 to 0.95 in increments of 0.05." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 323, + 444, + 336 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 444, + 336 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 444, + 336 + ], + "type": "text", + "content": "6.2. Implementation details" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 341, + 554, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 554, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 554, + 449 + ], + "type": "text", + "content": "We use the pre-trained SAM-ViT/H [25] of the patch size 16 as the backbone and set it frozen during training, which returns the " + }, + { + "bbox": [ + 313, + 341, + 554, + 449 + ], + "type": "inline_equation", + "content": "64 \\times 64 \\times 256" + }, + { + "bbox": [ + 313, + 341, + 554, + 449 + ], + "type": "text", + "content": " feature map. Due to the input patchification encapsulated in the ViT backbones [5], we find the receptive field of the raw feature map too coarse to detect small instances. We thus bilinearly interpolate the feature map resolution from " + }, + { + "bbox": [ + 313, + 341, + 554, + 449 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 313, + 341, + 554, + 449 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 341, + 554, + 449 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 313, + 341, + 554, + 449 + ], + "type": "text", + "content": " to produce a higher-resolution correlation map, which enables denser predictions and leads to better performance. (Tab. 11)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 449, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 554, + 521 + ], + "type": "text", + "content": "The channel dimension is expanded to 512 via a learnable linear projection. TMR is trained with the learning rate of " + }, + { + "bbox": [ + 313, + 449, + 554, + 521 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 313, + 449, + 554, + 521 + ], + "type": "text", + "content": " with AdamW [39] and the batch size of 16 on four Nvidia RTX 3090 GPUs with 24 GB VRAM for 24 hours. The box presence threshold " + }, + { + "bbox": [ + 313, + 449, + 554, + 521 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 449, + 554, + 521 + ], + "type": "text", + "content": " for NMS set to 0.4, 0.3 for RPINE and FSCD-147, respectively." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 529, + 534, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 534, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 534, + 541 + ], + "type": "text", + "content": "6.3. Comparison with state-of-the-art methods" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": "RPINE. To demonstrate the effectiveness of our method in counting and detecting non-object patterns, we evaluate it on RPINE. In Tab. 2, TMR surpasses the previous FSCD methods with a large margin. As the previous FSCD methods rely on prototypes for matching, they tend to struggle with non-object patterns that require an understanding of spatial details rather than semantics. In contrast, TMR effectively detects non-object patterns by incorporating spatial details in the template matching process. Note that RPINE is the only dataset equipped with multiple pattern annotations for each image in FSCD. Figure 5 demonstrates the bounding box predictions with different exemplars, where TMR predicts bounding boxes adaptively to the given exemplars. Figure 7 shows the qualitative com" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21582" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 295, + 148 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 295, + 148 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 295, + 148 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 295, + 148 + ], + "type": "table", + "html": "
MethodSDMAE(↓)RMSE(↓)AP(↑)AP50(↑)AP75(↑)
C-DETR [47]9.5821.2413.8832.2010.22
SAM-C [42]✓18.7737.1418.8034.0418.74
PseCo [23]✓48.2088.1623.1844.5421.24
GeCo [52]✓9.5717.0723.3345.9321.19
TMR (ours)8.4519.8733.5964.0530.52
TMR (ours)✓8.3019.4029.6658.9425.41
", + "image_path": "5ca54d8ad77d20610ec7f6f8de129de41cd7dd33053da83e1a7675e86daba82e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 152, + 295, + 185 + ], + "lines": [ + { + "bbox": [ + 55, + 152, + 295, + 185 + ], + "spans": [ + { + "bbox": [ + 55, + 152, + 295, + 185 + ], + "type": "text", + "content": "Table 2. One-shot pattern counting and detection results on the RPINE dataset. SD denotes box refinement with the SAM decoder. All the models are trained by the official code." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 197, + 295, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 197, + 295, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 197, + 295, + 220 + ], + "type": "text", + "content": "parisons with other FSCD methods where TMR accurately localize target patterns." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 224, + 295, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 224, + 295, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 295, + 380 + ], + "type": "text", + "content": "FSCD-LVIS and FSCD-147. We compare TMR with existing methods that are dedicated to FSCD under the FSCD setting on FSCD-LVIS and FSCD-147. The results in Tab. 3 demonstrate that TMR significantly outperforms previous state-of-the-art approaches. Notably, when evaluated on the unseen split, where test-time object categories are not observed during training, TMR surpasses prior methods. This suggests that TMR is less biased to object semantics during training potentially because TMR leverages the exemplar's spatial structure that generalizes across different categories. In addition to FSCD-LVIS, Tab. 6 compares methods on FSCD-147, where TMR performs on par with the previous methods on both one-shot and three-shot settings." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 389, + 209, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 209, + 402 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 209, + 402 + ], + "type": "text", + "content": "6.4. Analyses and ablation study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 410, + 295, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 295, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 295, + 625 + ], + "type": "text", + "content": "TMR learns with less semantic object bias and generalizes well across datasets. We compare TMR and GeCo [52] in the cross-dataset scenarios by evaluating the trained models on different datasets that are unseen during training. As shown in Tab. 4, TMR presents overwhelming performances, showing its strong generalization ability. Specifically, when GeCo is trained on FSCD-147, its performance drops significantly when evaluated on different datasets compared to when tested on FSCD-147 itself. GeCo, like previous FSCD methods, relies on prototypes for both counting and detection. We also suspect this lower generalization ability than ours originates from prototype matching, which is prone to overfitting to the training object semantics. GeCo struggles when evaluated on datasets with different object semantics. In contrast, TMR utilizes the structural information for matching instead of relying on semantic-intensive prototypes and generalizes more effectively on unseen datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": "SAM decoder is biased to edges. Table 2 shows the negative impact of the optional box refinement using the SAM decoder [25] on RPINE. Figure 6 visualizes two representative examples when the SAM decoder (SD) degrades performance. We observe that SD tends to align box predictions with the nearest edge, which is expected given that SAM is a segmentation model. The edge-sensitive predic" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 319, + 70, + 548, + 180 + ], + "blocks": [ + { + "bbox": [ + 319, + 70, + 548, + 180 + ], + "lines": [ + { + "bbox": [ + 319, + 70, + 548, + 180 + ], + "spans": [ + { + "bbox": [ + 319, + 70, + 548, + 180 + ], + "type": "table", + "html": "
MethodSeenUnseen
AP(↑)AP50(↑)AP(↑)AP50(↑)
FSDetView-PB [68]2.727.571.032.89
AttRPN-PB [7]4.0811.153.157.87
C-DETR [47]4.9214.493.8511.28
DAVE [51]6.7522.514.1214.16
PseCo [23]22.3742.56--
GeCo [52]--11.4724.49
TMR (ours)27.4948.4822.7139.68
", + "image_path": "91eaaea80cddccdecc32c8608b54c23679a510d224cb629204f6d50d001f7731.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 321, + 209, + 545, + 309 + ], + "blocks": [ + { + "bbox": [ + 313, + 183, + 553, + 206 + ], + "lines": [ + { + "bbox": [ + 313, + 183, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 313, + 183, + 553, + 206 + ], + "type": "text", + "content": "Table 3. Three-shot counting detection-based methods on the FSCD-LVIS seen and unseen split." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 209, + 545, + 309 + ], + "lines": [ + { + "bbox": [ + 321, + 209, + 545, + 309 + ], + "spans": [ + { + "bbox": [ + 321, + 209, + 545, + 309 + ], + "type": "table", + "html": "
TrainTestcross-evalAPAP50
GeCoTMRGeCoTMR
F-147F-14743.4244.4375.0673.83
F-LVISseen✓13.9621.2525.8737.18
RPINE✓19.4726.2138.6952.01
RPINEF-147✓36.9941.3960.3869.19
F-LVISseen✓10.0120.9217.4437.87
RPINE23.3329.6645.9358.94
", + "image_path": "5644ee27fbd8fdf6806a7e8f6c1d385ac3001f2d463538429736df1d73f4641a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 337, + 339, + 533, + 399 + ], + "blocks": [ + { + "bbox": [ + 313, + 312, + 553, + 335 + ], + "lines": [ + { + "bbox": [ + 313, + 312, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 553, + 335 + ], + "type": "text", + "content": "Table 4. Cross-dataset comparison of GeCo [52] and TMR, where F-147, F-LVIS indicate FSCD-147 and FSCD-LVIS." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 337, + 339, + 533, + 399 + ], + "lines": [ + { + "bbox": [ + 337, + 339, + 533, + 399 + ], + "spans": [ + { + "bbox": [ + 337, + 339, + 533, + 399 + ], + "type": "table", + "html": "
MethodOne-shotThree-shot
AP(↑)AP50(↑)AP(↑)AP50(↑)
GeCo [52]32.7169.9532.4970.51
TMR (ours)36.0171.1938.5772.61
", + "image_path": "ff1814c24ecaf2a10031653507c56e5cf60b164b830cfab7ef34ac7ae751534b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 401, + 553, + 412 + ], + "lines": [ + { + "bbox": [ + 313, + 401, + 553, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 553, + 412 + ], + "type": "text", + "content": "Table 5. Comparison without optional SAM decoder on FSCD147" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 422, + 555, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 555, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 555, + 579 + ], + "type": "text", + "content": "tion is particularly harmful for non-typical objects, such as objects that are loosely bounded by edges or patterns with many confusing edges. The SD refinement is seemingly good at edge detection for typical object exemplars with clear boundaries, and this is why the existing FSCD methods [23, 42, 52] benefit from adopting SD for box refinement. However, arbitrary patterns are not necessarily bounded by clear edges. Hence, edge-driven refinement may be even harmful as verified on RPINE. We emphasize that SD is an additional box post-processor, and we optionally add SD to TMR to compare with the SD-based state of the arts. Table 5 compares models by taking out SD, where TMR shows powerful performance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "Effectiveness of template matching features. We verify the effectiveness of 2D template matching, which preserves the spatial structure of exemplar bounding boxes for correlation. We experiment with different input feature " + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{P}}" + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": " to the box regression module. Table 7 compares our final model (e) and its variants. The lower bound model (a) observes zero information on the exemplar. Model (b) shows that correlating with 2D templates is greatly beneficial. The final model (e) concatenates the template-matching and the image feature, providing the correlation and appearance information for pattern matching regression. The models (c," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21583" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 82, + 56, + 526, + 173 + ], + "blocks": [ + { + "bbox": [ + 82, + 56, + 526, + 173 + ], + "lines": [ + { + "bbox": [ + 82, + 56, + 526, + 173 + ], + "spans": [ + { + "bbox": [ + 82, + 56, + 526, + 173 + ], + "type": "table", + "html": "
MethodSAM decoderOne-shotThree-shot
MAE(↓)RMSE(↓)AP(↑)AP50(↑)MAE(↓)RMSE(↓)AP(↑)AP50(↑)
FSDetView-PB [68]----37.54147.0713.4132.99
AttRPN-PB [7]----32.42141.5520.9737.19
C-DETR† [47]16.99125.2219.1447.6316.79123.5622.6650.57
SAM-C† [42]✓33.17141.7735.0956.0227.97131.2427.9949.17
PseCo [23]✓14.86118.6441.6370.8713.05112.8642.9873.33
DAVE [51]11.5486.6219.4655.2710.4574.5126.8162.82
GeCo [52]✓8.1060.1643.1174.317.9154.2843.4275.06
TMR (ours)✓11.6357.4643.1571.5513.7851.8744.4373.83
", + "image_path": "7f32c2d6e1e3f58e74be4164673a2c2b3d7eebd869597c36492231b4bf9fee8a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 201, + 551, + 346 + ], + "blocks": [ + { + "bbox": [ + 58, + 201, + 551, + 346 + ], + "lines": [ + { + "bbox": [ + 58, + 201, + 551, + 346 + ], + "spans": [ + { + "bbox": [ + 58, + 201, + 551, + 346 + ], + "type": "image", + "image_path": "497e318e60bb72db0a72d6f60209b99821c155be549bcfc9839b6a4783fee8cb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 348, + 555, + 383 + ], + "lines": [ + { + "bbox": [ + 55, + 348, + 555, + 383 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 555, + 383 + ], + "type": "text", + "content": "Figure 6. TMR without vs. with SAM decoder [25] (w/o SD vs. w/ SD) for box refinement. SAM decoder tends to align box predictions with the nearest edges. The edge-sensitive prediction is harmful for non-objects, objects that have loose boundaries, or patterns with many confusing edges. The exemplar in the first row loosely contains an arched window, but the SAM decoder snaps the box tightly its edges." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 79, + 387, + 271, + 481 + ], + "blocks": [ + { + "bbox": [ + 55, + 176, + 555, + 198 + ], + "lines": [ + { + "bbox": [ + 55, + 176, + 555, + 198 + ], + "spans": [ + { + "bbox": [ + 55, + 176, + 555, + 198 + ], + "type": "text", + "content": "Table 6. Comparison with few-shot detection and counting methods on FSCD-147. The one-shot performance of the models with " + }, + { + "bbox": [ + 55, + 176, + 555, + 198 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 176, + 555, + 198 + ], + "type": "text", + "content": " are evaluated using the official code." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 79, + 387, + 271, + 481 + ], + "lines": [ + { + "bbox": [ + 79, + 387, + 271, + 481 + ], + "spans": [ + { + "bbox": [ + 79, + 387, + 271, + 481 + ], + "type": "table", + "html": "
feature FPRPINEFSCD-147
APAP50APAP50
(a)F11.4424.0220.9555.59
(b)FTM32.5557.4331.9663.03
(c)F ⊕ FTM-cos29.7457.8524.7360.29
(d)F ⊕ FPM20.9447.9628.9166.93
(e)F ⊕ FTM33.5964.0536.0171.19
", + "image_path": "fed2e8bb0a76fad008461f6d893415c7b3831d87a800018b817177a786b1b1fb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 521, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 296, + 713 + ], + "type": "text", + "content": "d) replace the variants of the template matching feature " + }, + { + "bbox": [ + 55, + 521, + 296, + 713 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{TM}}" + }, + { + "bbox": [ + 55, + 521, + 296, + 713 + ], + "type": "text", + "content": " with something else. Model (c) replaces the channel-wise correlation of (e) to cosine similarity, i.e., " + }, + { + "bbox": [ + 55, + 521, + 296, + 713 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{TM - cos}}" + }, + { + "bbox": [ + 55, + 521, + 296, + 713 + ], + "type": "text", + "content": ", showing that retaining the channel dimension helps. Model (d) first performs average-pooling of the template and uses the prototype matching feature, " + }, + { + "bbox": [ + 55, + 521, + 296, + 713 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{PM}}" + }, + { + "bbox": [ + 55, + 521, + 296, + 713 + ], + "type": "text", + "content": ", for correlation. Note that this average-pooled feature is often used in previous FSC methods [51, 52, 70]. The result shows that preserving the spatial structure of exemplar bounding boxes is more effective. The representative failure cases are shown in Fig. 11 in supp. material. The pooled prototype feature loses the geometric layout of the template and struggle to detect patterns where geometric clues are crucial. Comparing (d) and (e) verifies our hypothesis that prototype-based matching relies on the object priors; Model (d) drops more significantly on RPINE, which is less object-centric than FSCD-147." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 387, + 556, + 466 + ], + "blocks": [ + { + "bbox": [ + 55, + 483, + 295, + 518 + ], + "lines": [ + { + "bbox": [ + 55, + 483, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 295, + 518 + ], + "type": "text", + "content": "Table 7. Effect of template matching features (Sec. 4.1). " + }, + { + "bbox": [ + 55, + 483, + 295, + 518 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{PM}}" + }, + { + "bbox": [ + 55, + 483, + 295, + 518 + ], + "type": "text", + "content": " is the correlation with average-pooled prototype. " + }, + { + "bbox": [ + 55, + 483, + 295, + 518 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{TM - cos}}" + }, + { + "bbox": [ + 55, + 483, + 295, + 518 + ], + "type": "text", + "content": " is the cosine similarity of the template and image feature." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 387, + 556, + 466 + ], + "lines": [ + { + "bbox": [ + 317, + 387, + 556, + 466 + ], + "spans": [ + { + "bbox": [ + 317, + 387, + 556, + 466 + ], + "type": "table", + "html": "
box regressionRPINEFSCD-147
shiftscaleAPAP50APAP50
(a)(0,0)(0,0)26.2055.2022.8157.30
(b)(Δx, Δy)(eαw, eαh)23.8851.2517.0155.14
(c)(Δx, Δy)(eαw sw, eαsh)31.2959.5235.0870.23
(d)(sw Δx, sh Δy)(eαw sw, eαsh)33.5964.0536.0171.19
", + "image_path": "f6bf9b3ea3fb7e22a5b4c2e95dcbf052cb2bdf98a12302d6f362ef2bcdb9aa20.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 339, + 468, + 529, + 479 + ], + "lines": [ + { + "bbox": [ + 339, + 468, + 529, + 479 + ], + "spans": [ + { + "bbox": [ + 339, + 468, + 529, + 479 + ], + "type": "text", + "content": "Table 8. Effect of box regression methods (Sec. 4.2)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 491, + 555, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 555, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 555, + 658 + ], + "type": "text", + "content": "Effectiveness of support-conditioned regression. Table 8 compares different variants for box regression and validates the effectiveness of our support-conditioned box regression method. As shown in (b), regressing bounding boxes without considering the support exemplar's size leads to poor localization. Notably, it even performs worse than (a), which directly uses the exemplar box without performing any regression. However, as shown in (c), incorporating the support exemplar's size when predicting the width and height leads to a significant performance gain. This improvement is further enhanced by using the exemplar's size for shifting, as demonstrated in (d). These results demonstrate the effectiveness of support-conditioned regression, guided by the exemplar's size, in achieving accurate localization." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Comparison on computational complexity. Table 9 compares the complexity of TMR with state-of-the-art FSCD methods, demonstrating its efficiency and effectiveness. Compared to PseCo [23], which introduces a large num" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21584" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 95, + 59, + 515, + 316 + ], + "blocks": [ + { + "bbox": [ + 95, + 59, + 515, + 316 + ], + "lines": [ + { + "bbox": [ + 95, + 59, + 515, + 316 + ], + "spans": [ + { + "bbox": [ + 95, + 59, + 515, + 316 + ], + "type": "image", + "image_path": "302d43631ad77592c7dd51b860e7557e3433e8266c0d5788ef4bda6ee871a519.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 319, + 555, + 343 + ], + "lines": [ + { + "bbox": [ + 55, + 319, + 555, + 343 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 555, + 343 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison with the state-of-the-art models on RPINE (the first two images) and FSCD-147 (the last two images). More visualization examples on three datasets are included in the supplementary materials Figs 16-18." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 63, + 348, + 286, + 405 + ], + "blocks": [ + { + "bbox": [ + 63, + 348, + 286, + 405 + ], + "lines": [ + { + "bbox": [ + 63, + 348, + 286, + 405 + ], + "spans": [ + { + "bbox": [ + 63, + 348, + 286, + 405 + ], + "type": "table", + "html": "
MethodTrainable paramsTotal paramsFLOPS
PseCo [23]56.99M0.70B5.08T
GeCo [52]7.98M0.65B4.72T
TMR (ours)19.01M0.66B3.04T
", + "image_path": "8af4a580d301ab6d3c4d27c34be563ca11c756e55d003deb12effb2f37b10452.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 75, + 407, + 276, + 419 + ], + "lines": [ + { + "bbox": [ + 75, + 407, + 276, + 419 + ], + "spans": [ + { + "bbox": [ + 75, + 407, + 276, + 419 + ], + "type": "text", + "content": "Table 9. Comparison on the computational complexity" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 428, + 295, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 428, + 295, + 560 + ], + "spans": [ + { + "bbox": [ + 54, + 428, + 295, + 560 + ], + "type": "text", + "content": "ber of trainable parameters and FLOPs, TMR introduces only about 19M trainable parameters, which is detailed in Tab. 10. Although TMR has more trainable parameters than GeCo [52], its total parameter count remains comparable. Thanks to the simple architecture, TMR is significantly efficient in terms of FLOPs (3.04T), which is notably lower than PseCo (5.08T) and GeCo (4.72T). This reduction in computational cost not only enhances training efficiency but also results in faster inference. In contrast, the high FLOPs of PseCo and GeCo contribute to longer inference, making them less practical for real-time applications." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 563, + 295, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 295, + 635 + ], + "type": "text", + "content": "Real-world application. We evaluate TMR, trained on RPINE, on scanning electron microscope (SEM) images [1], which are widely used in microprocessor inspection. Despite the domain shift, our method performs effectively and demonstrates its potential to generalize to real-world, non-object pattern scenarios (Fig. 8)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 646, + 206, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 646, + 206, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 646, + 206, + 658 + ], + "type": "text", + "content": "7. Conclusion and discussion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "We have proposed a simple template-matching based method for few-shot pattern detection. We also introduce a new dataset with bounding box annotations that covers various patterns around the world across non-object patterns" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 348, + 553, + 410 + ], + "blocks": [ + { + "bbox": [ + 316, + 348, + 553, + 410 + ], + "lines": [ + { + "bbox": [ + 316, + 348, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 316, + 348, + 553, + 410 + ], + "type": "image", + "image_path": "868c85f38ef62351767db810bbe2fb545d09d18a3d378ae85eaeedaa7d46ebef.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 412, + 550, + 424 + ], + "lines": [ + { + "bbox": [ + 318, + 412, + 550, + 424 + ], + "spans": [ + { + "bbox": [ + 318, + 412, + 550, + 424 + ], + "type": "text", + "content": "Figure 8. Application example of TMR on the SEM dataset [1]." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 434, + 555, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 555, + 457 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 555, + 457 + ], + "type": "text", + "content": "to typical objects from various image domains from nature and human-made products." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 459, + 556, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 556, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 556, + 591 + ], + "type": "text", + "content": "We observed that the widely used SAM decoder improves our performance by " + }, + { + "bbox": [ + 313, + 459, + 556, + 591 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 313, + 459, + 556, + 591 + ], + "type": "text", + "content": " AP points on FSCD-147 by leveraging object-level edge priors, but it degrades performance on RPINE, which contains more general patterns. This suggests a promising future direction that combines the powerful pre-trained knowledge of foundation models with detection mechanisms less dependent on object-level edge priors. In parallel, exploring a lightweight, pattern-specific architecture that efficiently captures fine-grained repetitive structures without relying on object semantics could further improve generalization and efficiency." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Acknowledgments. This work was supported by Samsung Electronics Co., Ltd (IO240508-09825-01) and IITP grants (RS-2022-II220959: Few-Shot Learning of Causal Inference in Vision & Language " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "(50\\%)" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": ", RS-2022-II220113: Developing a Sustainable Collaborative Multi-modal Lifelong Learning Framework " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "(45\\%)" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": ", RS-2019-II191906: AI Graduate School Program at POSTECH " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "(5\\%))" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": " funded by Ministry of Science and ICT, Korea. We also appreciate Manjin Kim's advice on constructing RPINE." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21585" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Rossella Aversa, Mohammad Hadi Modarres, Stefano Cozzini, Regina Ciancio, and Alberto Chiusole. The first annotated set of scanning electron microscopy images for nanoscience. Scientific data, 5(1):1-10, 2018. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 180 + ], + "type": "text", + "content": "[2] Yunliang Cai and George Baciu. Translation symmetry detection: A repetitive pattern analysis approach. In IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 223-228, 2013. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 182, + 295, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 295, + 235 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 295, + 235 + ], + "type": "text", + "content": "[3] Cheng Chen, Wei Wang, John A Ozolek, and Gustavo K Rohde. A flexible and robust approach for segmenting cell nuclei from 2d microscopy images using supervised learning and template matching. Cytometry Part A, 83(5):495-507, 2013. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 238, + 295, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 238, + 295, + 259 + ], + "spans": [ + { + "bbox": [ + 62, + 238, + 295, + 259 + ], + "type": "text", + "content": "[4] Taco S Cohen and Max Welling. Steerable cnns. arXiv preprint arXiv:1612.08498, 2016. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 262, + 295, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 262, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 62, + 262, + 295, + 327 + ], + "type": "text", + "content": "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 329, + 295, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 329, + 295, + 372 + ], + "spans": [ + { + "bbox": [ + 62, + 329, + 295, + 372 + ], + "type": "text", + "content": "[6] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International Journal of Computer Vision (IJCV), 88:303-338, 2010. 1, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 374, + 295, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 374, + 295, + 427 + ], + "spans": [ + { + "bbox": [ + 62, + 374, + 295, + 427 + ], + "type": "text", + "content": "[7] Qi Fan, Wei Zhuo, Chi-Keung Tang, and Yu-Wing Tai. Few-shot object detection with attention-rpn and multi-relation detector. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4013–4022, 2020. 1, 2, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 430, + 295, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 430, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 62, + 430, + 295, + 474 + ], + "type": "text", + "content": "[8] Zhibo Fan, Yuchen Ma, Zeming Li, and Jian Sun. Generalized few-shot object detection without forgetting. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4527-4536, 2021. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 475, + 295, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 475, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 62, + 475, + 295, + 529 + ], + "type": "text", + "content": "[9] Christopher Funk, Seungkyu Lee, Martin R Oswald, Stavros Tsogkas, Wei Shen, Andrea Cohen, Sven Dickinson, and Yanxi Liu. 2017 iccv challenge: Detecting symmetry in the wild. In International Conference on Computer Vision Workshops (ICCVW), pages 1692-1701, 2017. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 532, + 295, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 532, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 57, + 532, + 295, + 563 + ], + "type": "text", + "content": "[10] Hiroyasu Furukawa, Kyle E Cordova, Michael O'Keeffe, and Omar M Yaghi. The chemistry and applications of metalorganic frameworks. Science, 341(6149):1230444, 2013. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 566, + 295, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 566, + 295, + 609 + ], + "spans": [ + { + "bbox": [ + 57, + 566, + 295, + 609 + ], + "type": "text", + "content": "[11] Jizhou Gao, Yin Hu, Jinze Liu, and Ruigang Yang. Unsupervised learning of high-order structural semantics from images. In Proc. IEEE International Conference on Computer Vision (ICCV), pages 2122-2129. IEEE, 2009. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 611, + 295, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 611, + 295, + 642 + ], + "spans": [ + { + "bbox": [ + 57, + 611, + 295, + 642 + ], + "type": "text", + "content": "[12] Ross Girshick. Fast r-cnn. In Proc. IEEE International Conference on Computer Vision (ICCV), pages 1440-1448, 2015. 1, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 644, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 644, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 644, + 295, + 667 + ], + "type": "text", + "content": "[13] Alice L Givan. Flow cytometry: an introduction. Flow cytometry protocols, pages 1-29, 2011. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 669, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 669, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 669, + 295, + 690 + ], + "type": "text", + "content": "[14] Branko Grünbaum and Geoffrey Colin Shephard. Tilings and patterns. Courier Dover Publications, 1987. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "text", + "content": "[15] Karim Guirguis, Johannes Meier, George Eskandar, Matthias Kayser, Bin Yang, and Jürgen Beyerer. Niff: Alleviating" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 74, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 116 + ], + "type": "text", + "content": "forgetting in generalized few-shot object detection via neural instance feature forging. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 24193-24202, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "type": "text", + "content": "[16] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5356-5364, 2019. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 163, + 553, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 216 + ], + "type": "text", + "content": "[17] Guangxing Han, Yicheng He, Shiyuan Huang, Jiawei Ma, and Shih-Fu Chang. Query adaptive few-shot object detection with heterogeneous graph convolutional networks. In Proc. IEEE International Conference on Computer Vision (ICCV), pages 3263-3272, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 217, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 217, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 317, + 217, + 553, + 271 + ], + "type": "text", + "content": "[18] Guangxing Han, Shiyuan Huang, Jiawei Ma, Yicheng He, and Shih-Fu Chang. Meta faster r-cnn: Towards accurate few-shot object detection with attentive feature alignment. In Proc. AAAI Conference on Artificial Intelligence (AAAI), pages 780-789, 2022. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 273, + 553, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 273, + 553, + 325 + ], + "spans": [ + { + "bbox": [ + 317, + 273, + 553, + 325 + ], + "type": "text", + "content": "[19] Guangxing Han, Jiawei Ma, Shiyuan Huang, Long Chen, and Shih-Fu Chang. Few-shot object detection with fully cross-transformer. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5321-5330, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 327, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 327, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 317, + 327, + 553, + 370 + ], + "type": "text", + "content": "[20] Junwei Han, Stephen J McKenna, and Ruixuan Wang. Regular texture analysis as statistical model selection. In Proc. European Conference on Computer Vision (ECCV), pages 242-255. Springer, 2008. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 373, + 553, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 373, + 553, + 423 + ], + "spans": [ + { + "bbox": [ + 317, + 373, + 553, + 423 + ], + "type": "text", + "content": "[21] James Hays, Marius Leordeanu, Alexei A Efros, and Yanxi Liu. Discovering texture regularity as a higher-order correspondence problem. In Proc. European Conference on Computer Vision (ECCV), pages 522-535. Springer, 2006. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 426, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 426, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 317, + 426, + 553, + 459 + ], + "type": "text", + "content": "[22] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proc. IEEE International Conference on Computer Vision (ICCV), pages 2961-2969, 2017. 1, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 460, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 553, + 514 + ], + "type": "text", + "content": "[23] Zhizhong Huang, Mingliang Dai, Yi Zhang, Junping Zhang, and Hongming Shan. Point segment and count: A generalized framework for object counting. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 516, + 553, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 516, + 553, + 559 + ], + "spans": [ + { + "bbox": [ + 317, + 516, + 553, + 559 + ], + "type": "text", + "content": "[24] Bingyi Kang, Zhuang Liu, Xin Wang, Fisher Yu, Jiashi Feng, and Trevor Darrell. Few-shot object detection via feature reweighting. In Proc. IEEE International Conference on Computer Vision (ICCV), pages 8420-8429, 2019. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 559, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 553, + 624 + ], + "type": "text", + "content": "[25] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In Proc. IEEE International Conference on Computer Vision (ICCV), 2023. 3,4,5,6,7,2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 625, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 625, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 317, + 625, + 553, + 669 + ], + "type": "text", + "content": "[26] Simon Korman, Daniel Reichman, Gilad Tsur, and Shai Avidan. Fast-match: Fast affine template matching. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2331-2338, 2013. 3, 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 670, + 553, + 712 + ], + "type": "text", + "content": "[27] Minji Lee, Hyunju Noh, and Youngsuk Kim. Diamidocarbene-derived palladium and nickel-sulfur clusters. Chemical Communications, 60(94):13867-13870, 2024. 4" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "21586" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[28] Thomas Leung and Jitendra Malik. Detecting, localizing and grouping repeated scene elements from an image. In Proc. European Conference on Computer Vision (ECCV). Springer, 1996. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "text", + "content": "[29] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "type": "text", + "content": "[30] Hui Lin, Xiaopeng Hong, and Yabin Wang. Object counting: You only need to look at one. arXiv preprint arXiv:2112.05993, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 228 + ], + "type": "text", + "content": "[31] Hsin-Chih Lin, Ling-Ling Wang, and Shi-Nine Yang. Extracting periodicity of a regular texture based on autocorrelation functions. Pattern recognition letters, 18(5):433-443, 1997. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 231, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 294, + 285 + ], + "type": "text", + "content": "[32] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proc. European Conference on Computer Vision (ECCV), pages 740-755. Springer, 2014. 1, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 286, + 294, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 341 + ], + "type": "text", + "content": "[33] Tsung-Yi Lin, Piotr Dólár, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2117-2125, 2017. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 342, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 294, + 387 + ], + "type": "text", + "content": "[34] Wen-Chieh Lin and Yanxi Liu. Tracking dynamic near-regular texture under occlusion and rapid movements. In Proc. European Conference on Computer Vision (ECCV), pages 44–55. Springer, 2006. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 294, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 421 + ], + "type": "text", + "content": "[35] Jingchen Liu and Yanxi Liu. Grasp recurring patterns from a single view. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2003-2010, 2013. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 422, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 487 + ], + "type": "text", + "content": "[36] Jingchen Liu, George Slota, Gang Zheng, Zhaohui Wu, Minwoo Park, Seungkyu Lee, Ingmar Rauschert, and Yanxi Liu. Symmetry detection from realworld images competition 2013: Summary and results. In IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 200-205, 2013. 4, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 488, + 294, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 488, + 294, + 543 + ], + "spans": [ + { + "bbox": [ + 56, + 488, + 294, + 543 + ], + "type": "text", + "content": "[37] Siying Liu, Tian-Tsong Ng, Kalyan Sunkavalli, Minh N Do, Eli Shechtman, and Nathan Carr. Patchmatch-based automatic lattice detection for near-regular textures. In Proc. IEEE International Conference on Computer Vision (ICCV), pages 181–189, 2015. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 544, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 294, + 578 + ], + "type": "text", + "content": "[38] Yanxi Liu, Wen-Chieh Lin, and James Hays. Near-regular texture analysis and manipulation. ACM Transactions on Graphics (TOG), 23(3):368-376, 2004. 3, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 578, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 578, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 578, + 294, + 600 + ], + "type": "text", + "content": "[39] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 601, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 294, + 635 + ], + "type": "text", + "content": "[40] Erika Lu, Weidi Xie, and Andrew Zisserman. Class-agnostic counting. In Asian Conference on Computer Vision (ACCV), pages 669–684. Springer, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 635, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 679 + ], + "type": "text", + "content": "[41] Bronwyn A Lucas, Benjamin A Himes, Liang Xue, Timothy Grant, Julia Mahamid, and Nikolaus Grigorieff. Locating macromolecular assemblies in cells by 2d template matching with cistem. _Elife_, 10:e68946, 2021. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "type": "text", + "content": "[42] Zhiheng Ma, Xiaopeng Hong, and Qinnan Shangguan. Can sam count anything? an empirical study on sam counting, 2023. 3, 4, 6, 7" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[43] Massimiliano Mancini, Elisa Ricci, Barbara Caputo, and Samuel Rota Bulo. Adding new tasks to a single network with weight transformations using binary masks. In European Conference on Computer Vision Workshops (ECCVW), pages 0-0, 2018. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 129, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 162 + ], + "type": "text", + "content": "[44] Giovanni Marola. Using symmetry for detecting and locating objects in a picture. Computer Vision, Graphics, and Image Processing, 46(2):179-195, 1989. 3, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "text", + "content": "[45] Jiri Matas, Ondrej Chum, Martin Urban, and Tomás Pajdla. Robust wide-baseline stereo from maximally stable extremal regions. Image and vision computing, 22(10):761-767, 2004." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 208, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 252 + ], + "type": "text", + "content": "[46] Anuj Mohan, Constantine Papageorgiou, and Tomaso Poggio. Example-based object detection in images by components. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 23(4):349-361, 2001. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "type": "text", + "content": "[47] Thanh Nguyen, Chau Pham, Khoi Nguyen, and Minh Hoai. Few-shot object counting and detection. In Proc. European Conference on Computer Vision (ECCV). Springer, 2022. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 298, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 553, + 342 + ], + "type": "text", + "content": "[48] Roni Paiss, Ariel Ephrat, Omer Tov, Shiran Zada, Inbar Mosseri, Michal Irani, and Tali Dekel. Teaching clip to count to ten. In Proc. IEEE International Conference on Computer Vision (ICCV), pages 3170-3180, 2023. 4, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 343, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 397 + ], + "type": "text", + "content": "[49] Minwoo Park, Kyle Brocklehurst, Robert T Collins, and Yanxi Liu. Deformed lattice detection in real-world images using mean-shift belief propagation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 31(10): 1804-1816, 2009. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 399, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 442 + ], + "type": "text", + "content": "[50] Mark Pauly, Niloy J Mitra, Johannes Wallner, Helmut Pottmann, and Leonidas J Guibas. Discovering structural regularity in 3d geometry. In ACM SIGGRAPH, pages 1-11, 2008. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 444, + 553, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 553, + 488 + ], + "type": "text", + "content": "[51] Jer Pelhan, Vitjan Zavrtanik, Matej Kristan, et al. Davea detect-and-verify paradigm for low-shot counting. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 23293-23302, 2024. 3, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 489, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 489, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 489, + 553, + 533 + ], + "type": "text", + "content": "[52] Jer Pelhan, Alan Lukezic, Vitjan Zavrtanik, and Matej Kristan. A novel unified architecture for low-shot counting by detection and segmentation. Advances in Neural Information Processing Systems (NeurIPS), 2025. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 534, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 553, + 578 + ], + "type": "text", + "content": "[53] Julien Picot, Coralie L Guerin, Caroline Le Van Kim, and Chantal M Boulanger. Flow cytometry: retrospective, fundamentals and recent instrumentation. Cytotechnology, 64: 109-130, 2012. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 578, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 553, + 633 + ], + "type": "text", + "content": "[54] Muhammad Farrukh Qureshi, Zohaib Mushtaq, Muhammad Zia Ur Rehman, and Ernest Nlandu Kamavuako. E2cnn: An efficient concatenated cnn for classification of surface emg extracted from upper limb. IEEE Sensors Journal, 23(8): 8989-8996, 2023. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 635, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 668 + ], + "type": "text", + "content": "[55] Viresh Ranjan and Minh Hoai. Vicinal counting networks. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4221-4230, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "text", + "content": "[56] Viresh Ranjan, Udbhav Sharma, Thu Nguyen, and Minh Hoai. Learning to count everything. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3394-3403, 2021. 2, 1" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21587" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[57] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 779-788, 2016. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "type": "text", + "content": "[58] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 658-666, 2019. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 294, + 207 + ], + "type": "text", + "content": "[59] Babak Saleh and Ahmed Elgammal. Large-scale classification of fine-art paintings: Learning the right metric on the right feature. arXiv preprint arXiv:1505.00855, 2015. 4, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 240 + ], + "type": "text", + "content": "[60] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 294, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 296 + ], + "type": "text", + "content": "[61] Luciano Spinello, Rudolph Triebel, Dizan Vasquez, Kai O Arras, and Roland Siegwart. Exploiting repetitive object patterns for model compression and completion. In Proc. European Conference on Computer Vision (ECCV), pages 296-309. Springer, 2010. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 319 + ], + "type": "text", + "content": "[62] V.J. Stenger. Timeless Reality: Symmetry, Simplicity, and Multiple Universes. Prometheus, 2009. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 320, + 294, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 294, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 294, + 373 + ], + "type": "text", + "content": "[63] Bo Sun, Banghuai Li, Shengcai Cai, Ye Yuan, and Chi Zhang. Fsce: Few-shot object detection via contrastive proposal encoding. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 7352-7362, 2021, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 376, + 294, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 419 + ], + "type": "text", + "content": "[64] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In Proc. IEEE International Conference on Computer Vision (ICCV), pages 9627-9636, 2019. 1, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 421, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 421, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 421, + 294, + 453 + ], + "type": "text", + "content": "[65] Du-Ming Tsai and Cheng-Huei Chiang. Rotation-invariant pattern matching using wavelet decomposition. Pattern Recognition Letters, 23(1-3):191-201, 2002. 3, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 455, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 455, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 294, + 498 + ], + "type": "text", + "content": "[66] Andreas Turina, Tinne Tuytelaars, and Luc Van Gool. Efficient grouping under perspective skew. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages I-I. IEEE, 2001. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 500, + 294, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 294, + 543 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 294, + 543 + ], + "type": "text", + "content": "[67] Xin Wang, Thomas Huang, Joseph Gonzalez, Trevor Darrell, and Fisher Yu. Frustratingly simple few-shot object detection. In Proc. International Conference on Machine Learning (ICML), pages 9919-9928. PMLR, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 544, + 294, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 294, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 294, + 577 + ], + "type": "text", + "content": "[68] Yang Xiao, Vincent Lepetit, and Renaud Marlet. Few-shot object detection and viewpoint estimation for objects in the wild. TPAMI, 45(3):3090-3106, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 578, + 294, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 578, + 294, + 632 + ], + "spans": [ + { + "bbox": [ + 56, + 578, + 294, + 632 + ], + "type": "text", + "content": "[69] Xiaopeng Yan, Ziliang Chen, Anni Xu, Xiaoxi Wang, Xiaodan Liang, and Liang Lin. Meta r-cnn: Towards general solver for instance-level low-shot learning. In Proc. IEEE International Conference on Computer Vision (ICCV), pages 9577-9586, 2019. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 635, + 294, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 677 + ], + "type": "text", + "content": "[70] Zhiyuan You, Kai Yang, Wenhan Luo, Xin Lu, Lei Cui, and Xinyi Le. Iterative correlation-based feature refinement for few-shot counting. arXiv preprint arXiv:2201.08959, 4, 2022. 2, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 679, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 679, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 679, + 294, + 712 + ], + "type": "text", + "content": "[71] Xinyu Zhang, Yuhan Liu, Yuting Wang, and Abdeslam Boularias. Detect everything with few examples. arXiv preprint arXiv:2309.12969, 2023. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21588" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Fewer Denoising Steps or Cheaper Per-Step Inference_ Towards Compute-Optimal Diffusion Model Deployment/98cafbf9-4954-46a4-b5af-59f0bca621f7_content_list.json b/2025/Fewer Denoising Steps or Cheaper Per-Step Inference_ Towards Compute-Optimal Diffusion Model Deployment/98cafbf9-4954-46a4-b5af-59f0bca621f7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a5538eb60cb1ac925de17f2c5d3768397be24d42 --- /dev/null +++ b/2025/Fewer Denoising Steps or Cheaper Per-Step Inference_ Towards Compute-Optimal Diffusion Model Deployment/98cafbf9-4954-46a4-b5af-59f0bca621f7_content_list.json @@ -0,0 +1,1462 @@ +[ + { + "type": "text", + "text": "Fewer Denoising Steps or Cheaper Per-Step Inference: Towards Compute-Optimal Diffusion Model Deployment", + "text_level": 1, + "bbox": [ + 176, + 130, + 818, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhenbang Du*, Yonggan Fu*, Lifu Wang*, Jiayi Qian, Xiao Luo, Yingyan (Celine) Lin \nGeorgia Institute of Technology \n{zdu89,jiayiqian,celine.lin}@gatech.edu", + "bbox": [ + 241, + 203, + 754, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion models have shown remarkable success across generative tasks, yet their high computational demands challenge deployment on resource-limited platforms. This paper investigates a critical question for compute-optimal diffusion model deployment: Under a post-training setting without fine-tuning, is it more effective to reduce the number of denoising steps or to use a cheaper per-step inference? Intuitively, reducing the number of denoising steps increases the variability of the distributions across steps, making the model more sensitive to compression. In contrast, keeping more denoising steps makes the differences smaller, preserving redundancy, and making post-training compression more feasible. To systematically examine this, we propose PostDiff, a training-free framework for accelerating pre-trained diffusion models by reducing redundancy at both the input level and module level in a post-training manner. At the input level, we propose a mixed-resolution denoising scheme based on the insight that reducing generation resolution in early denoising steps can enhance low-frequency components and improve final generation fidelity. At the module level, we employ a hybrid module caching strategy to reuse computations across denoising steps. Extensive experiments and ablation studies demonstrate that (1) PostDiff can significantly improve the fidelity-efficiency trade-off of state-of-the-art diffusion models, and (2) to boost efficiency while maintaining decent generation fidelity, reducing per-step inference cost is often more effective than reducing the number of denoising steps. Our code is available at https://github.com/GATECH-EIC/PostDiff.", + "bbox": [ + 91, + 340, + 483, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 806, + 220, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion models have rapidly advanced as a powerful tool for various generative tasks, such as image synthesis and video generation. Their success stems from the iterative de", + "bbox": [ + 89, + 832, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "noising process, which gradually refines outputs to achieve high-quality results, along with advanced model architectures. However, both the iterative nature and model complexity of diffusion models make them computationally intensive, posing a significant challenge for deployment on resource-constrained platforms, where computational resources, memory, and power are limited. Efficient deployment of diffusion models on these platforms is therefore highly desirable to broaden the accessibility of generative AI, enabling real-time applications that can operate within the strict constraints of on-device resources.", + "bbox": [ + 511, + 310, + 903, + 474 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the increasing demand for rapid diffusion model deployment, where a given diffusion model is expected to be instantly compressed and deployed without fine-tuning, a critical question arises: Under a post-training setting, is it more effective to optimize for fewer denoising steps or to reduce the per-step inference cost? This question is particularly important because each approach entails different trade-offs: Reducing the number of denoising steps introduces greater variability in feature changes between steps, which can amplify the negative impact of compression on output fidelity; Conversely, using more denoising steps reduces these step-to-step differences, preserving redundancy and making it easier to apply post-training compression without sacrificing quality. However, there is a limited exploration of how these two strategies perform in terms of fidelity-efficiency trade-offs, and investigating this could provide valuable guidelines for efficient diffusion model deployment in real-world applications.", + "bbox": [ + 511, + 477, + 903, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address this research question, we propose PostDiff, a training-free diffusion model compression framework designed to enhance the per-step inference efficiency of pretrained diffusion models at both the input level and module level in a post-training manner. Our techniques are training-free and can thus be instantly applied to emerging diffusion models for rapid deployment, enabling a comparative study of their effectiveness versus reducing denoising steps. Specifically, at the input level, we propose a mixed-resolution denoising scheme that adopts a lower generation", + "bbox": [ + 511, + 750, + 903, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 805, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 109, + 887, + 218, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3001", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "resolution in early denoising steps. As low-frequency components dominate at this stage, a reduced resolution is sufficient. We find that this reduced resolution in early denoising steps can enhance low-frequency components, thereby improving final generation fidelity and achieving a win-win in both efficiency and fidelity. At the module level, we build on existing module caching techniques, which reuse module outputs across steps, and enhance them with a hybrid caching strategy as another dimension for redundancy reduction. Our contributions are summarized as follows:", + "bbox": [ + 89, + 90, + 480, + 241 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Our work, for the first time, studies a critical question for efficient diffusion model deployment: Under a posttraining setting without fine-tuning, is it more effective to optimize for fewer denoising steps or to reduce the per-step inference cost? The answer to this question could be useful for both researchers and practitioners.", + "- To answer the above question, we propose PostDiff, a training-free framework for reducing diffusion model complexity at both the input and module levels, resulting in lower per-step inference cost. This framework enables a fair study of the relative effectiveness between fewer denoising steps and reduced per-step inference cost toward compute-optimal diffusion model deployment under a post-training setting.", + "- At the input level, we propose a mixed-resolution denoising scheme, which leverages the observation that reduced generation resolution in early denoising steps can enhance low-frequency components and final output quality, achieving a win-win in both accuracy and efficiency. In addition, to further reduce the overall computational cost at the module level, PostDiff employs a hybrid module caching method that strategically reuses the outputs of different modules on varying schedules across steps.", + "- Extensive evaluations show that PostDiff can advance the fidelity-efficiency frontier of state-of-the-art (SOTA) diffusion models, as shown in Fig. 1, and provide valuable insights into best practices for efficient diffusion model deployment. For example, we find that reducing per-step inference cost is often more effective than decreasing the number of denoising steps to improve efficiency when aiming to preserve high generation fidelity." + ], + "bbox": [ + 89, + 243, + 482, + 710 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the following sections, we illustrate and analyze the effectiveness of each technique in PostDiff in Sec. 2, and finally, we combine and apply them to SOTA diffusion models in Sec. 3 to answer our research question.", + "bbox": [ + 89, + 713, + 482, + 773 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. PostDiff: The Proposed Framework", + "text_level": 1, + "bbox": [ + 89, + 784, + 415, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we first introduce the proposed mixed-resolution denoising strategy, which reduces per-step inference cost at the input level. We begin by analyzing the motivations from previous works in Sec. 2.1, followed by an illustration of our method, an analysis of its effectiveness, and its application to examine our key question in Sec. 2.2,", + "bbox": [ + 89, + 809, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/96dcd55c2125876b6f232a70876722be8f7e2ed7d1dd6c05b3e0b978a4a821cd.jpg", + "image_caption": [ + "Figure 1. Apply our PostDiff technique to SOTA diffusion models, where “-PD” denotes the models delivered by PostDiff." + ], + "image_footnote": [], + "bbox": [ + 519, + 85, + 895, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Sec. 2.3, and Sec. 2.4, respectively. Finally, we describe how we employ and enhance module caching to improve efficiency at the module level in Sec. 2.5.", + "bbox": [ + 511, + 325, + 903, + 371 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Motivations from Previous Works", + "text_level": 1, + "bbox": [ + 511, + 380, + 808, + 395 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Low-frequency components dominate the early denoising steps. Previous works [22] find that the typical denoising process for image generation progresses through two phases: the semantics-planning phase and the fidelity-improving phase. The early semantics-planning phase focuses more on the layout and low-frequency information, starting from random noise and gradually establishing the basic structure. This indicates that a low generation resolution could be sufficient in the early denoising steps. As generation proceeds, more details, described by high-frequency information, are incrementally added, thus requiring a higher resolution.", + "bbox": [ + 511, + 402, + 903, + 583 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This insight is also partially echoed and leveraged by previous works. For example, [38] trains different diffusion models for early low-resolution denoising and later high-resolution denoising processes to avoid the propagation of inaccurate high-frequency components from early denoising steps to later steps. Additionally, cascade diffusion models [16, 40] sequentially apply different diffusion models at different denoising stages with varying generation resolutions, finding that incorporating low-frequency information from low-resolution images during high-resolution stages can enhance generation quality. Nevertheless, all these methods require multiple diffusion models, introducing significant storage and training demands.", + "bbox": [ + 511, + 584, + 905, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Input resolution is a promising dimension for reducing inference cost. High-resolution generation is a recent trend, as it provides images with more vivid details and accommodates greater information density [47]. For example, SDXL [29] generates images at 1024x1024 resolution, while PixArt- $\\Sigma$ [4] expands resolution capability to 4K. However, as resolution increases, computational demands rise dramatically, leading to significant latency [19]. As", + "bbox": [ + 511, + 780, + 905, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3002", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/136fc13924f38ecd291bd208ec099e0afa2ea63585db3dbe4ab13f47333c3b8b.jpg", + "image_caption": [ + "Figure 2. An overview of our mixed-resolution denoising strategy." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 480, + 256 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "such, recent works aim to reduce the computational load by decreasing the number of tokens in samples, using methods like token merging [1], token pruning [39], or token down-sampling [35], which effectively accelerate inference.", + "bbox": [ + 89, + 295, + 483, + 357 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Mixed-Resolution Denoising: Methodology", + "text_level": 1, + "bbox": [ + 89, + 363, + 457, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motivated by the tendency for low-frequency components in early denoising steps and the effectiveness of input resolution as a design knob for efficiency, we propose a simple yet effective post-training strategy: As shown in Fig. 2, we adopt a low generation resolution in the early denoising steps to generate low-frequency structures, and then refine high-frequency details with high generation resolution in the later denoising steps.", + "bbox": [ + 89, + 383, + 483, + 505 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Formulation. To generate an image with shape $(W,H,3)$ , we begin with Gaussian noise in the shape $(w,h,4)$ , representing the latent noise image $x_{T}$ obtained via Variational Autoencoders [30]. For simplicity, we omit the channel dimension in the remainder of this paper. The denoising process proceeds over $T$ steps, using samplers like DDIM [36] or DPM-solver [23] to obtain the denoised latent at each step. Taking DDIM [36] as an example, we compute the latent representation $x$ at step $t - 1$ as follows:", + "bbox": [ + 89, + 506, + 483, + 642 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t - 1} = \\sqrt {\\alpha_ {t - 1}} \\hat {x} _ {0} ^ {t} + x _ {t}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 230, + 652, + 482, + 669 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {x} _ {0} ^ {t} = \\frac {x _ {t} - \\sqrt {1 - \\alpha_ {t}} \\cdot \\epsilon_ {\\theta} (x _ {t} , t)}{\\sqrt {\\alpha_ {t}}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 672, + 482, + 705 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t} = \\sqrt {1 - \\alpha_ {t - 1}} \\cdot \\epsilon_ {\\theta} (x _ {t}, t), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 708, + 480, + 728 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon_{\\theta}$ denotes the model [28, 32] that predicts the added noise, and $\\{\\alpha_t\\}$ is a set of coefficients controlling the forward noise-adding process [15].", + "bbox": [ + 89, + 734, + 482, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our proposed mixed-resolution denoising strategy, we first introduce a lower-resolution latent representation, $x_{T}^{l}$ , with shape $(\\beta w, \\beta h)$ , where $0 < \\beta < 1$ is a hyperparameter that scales the resolution down. This low-resolution starting point reduces the computational load of the initial denoising phase. Another hyperparameter, $s$ , controls the transition from low to full resolution. After reaching the denoising step $t = sT$ , we compute the forecasted", + "bbox": [ + 89, + 780, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\hat{x}_0^{l,t}$ at step $t$ using Eq. (2). Then, we upscale this result to full resolution via:", + "bbox": [ + 511, + 89, + 903, + 119 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {x} _ {0} ^ {f, t} = \\operatorname {U p s a m p l e} \\left(\\hat {x} _ {0} ^ {l, t}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 629, + 128, + 903, + 148 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where Upsample refers to bilinear interpolation. Using this upscaled version, we derive the full-resolution latent $x_{t}^{f}$ at step $t$ as follows:", + "bbox": [ + 511, + 156, + 906, + 202 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t} ^ {f} = \\sqrt {\\alpha_ {t}} \\hat {x} _ {0} ^ {f, t} + \\sqrt {1 - \\alpha_ {t}} \\epsilon , \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 209, + 903, + 229 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon$ is Gaussian noise with shape $(w,h)$ . The remainder of the denoising process follows the original procedure.", + "bbox": [ + 511, + 236, + 903, + 267 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Note that we adopt this simple binary resolution strategy without bells and whistles to effectively demonstrate its concept and simplify the involved hyperparameters, thereby easing the use of our method in real-world deployment. More complex resolution schedules could potentially be proposed to further enhance fidelity-efficiency trade-offs.", + "bbox": [ + 511, + 267, + 905, + 358 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Visualization: How Low-Resolution Denoising Contributes to Boosted Final Performance", + "text_level": 1, + "bbox": [ + 511, + 364, + 905, + 396 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To study the impact of our mixed-resolution denoising strategy on final performance and understand why it works, we visualize the generation quality after each denoising step, both with and without our method. Specifically, we calculate the Clip Score [13] for $\\hat{x}_0$ predicted at each denoising step and annotate it in the case study of Fig. 3, where different rows correspond to the original full-resolution denoising process as well as our method with various choices of $s$ .", + "bbox": [ + 511, + 402, + 905, + 522 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Observations and analysis. We observe that (1) with a moderate amount of low-resolution generation steps, the Clip Score of the final generated image can be improved, achieving a win-win in generation fidelity and efficiency; (2) Regarding the evolution of the Clip Score across denoising steps, starting with a lower generation resolution results in lower Clip Scores initially, but serves as a better starting point: after switching to high generation resolution, the Clip Score increases more rapidly and gradually surpasses that of full-resolution generation. This echoes the analysis in Sec. 2.1, where reduced generation resolution in early denoising steps contributes to improved low-frequency components, thereby enhancing final generation quality; (3) The balance between low-resolution and high-resolution steps is crucial, as both low-frequency structure and high-frequency details need to be preserved to ensure high final generation quality in the last denoising step.", + "bbox": [ + 511, + 523, + 905, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This phenomenon demonstrates the underlying effectiveness of our mixed-resolution denoising strategy and is consistently observed across samples, as shown in the comprehensive evaluation in Sec. 2.4.", + "bbox": [ + 511, + 780, + 905, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.4. Mixed-Resolution Denoising: Evaluation", + "text_level": 1, + "bbox": [ + 511, + 848, + 862, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We evaluate the effectiveness of our mixed-resolution denoising strategy and examine which is more effective: re", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3003", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "cute girl, crop-top, blond hair, black glasses, stretching, with background by greg rutkowski makoto shinkai kyoto animation key art feminine mid shot", + "bbox": [ + 248, + 90, + 841, + 119 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e928864cca10772719024a02063b86d8e738c2c6c3451fbd49313102fcc912ea.jpg", + "image_caption": [ + "Figure 3. Visualize the Clip Score after each denoising step using full resolution (row 1) and mixed resolution (rows 2-5). The steps using low generation resolution are highlighted in green and those using high generation resolution are highlighted in orange." + ], + "image_footnote": [], + "bbox": [ + 106, + 122, + 898, + 398 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c35d0dfc1fecf32689f4de3fa8455c0fbac17c375acf1287ab60e6d0e8bb4c3e.jpg", + "image_caption": [ + "Figure 4. The FID-FLOPs trade-off achieved under different denoising steps and varied mixed-resolution settings. The data points with the same number of steps are annotated using the same shape." + ], + "image_footnote": [], + "bbox": [ + 96, + 449, + 478, + 693 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "duced input-level redundancy or fewer denoising steps.", + "bbox": [ + 89, + 761, + 454, + 776 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Setup. We use Stable Diffusion (SD) V1.5 (DreamShaper-7 version) [31], a popular latent diffusion model (LDM), as our model backbone and evaluate text-to-image generation performance on the MS-COCO 2014 validation dataset [21]. We vary the total denoising steps $T = \\{8,10,12,14,16,18,20\\}$ and apply our mixed-resolution denoising to each denoising step setting with varied $s = \\{0.1,0.2,0.3,0.4,0.5,0.6\\}$ . Without bells and", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "whistles, we set $\\beta = 0.5$ as our low-resolution setting, i.e., half the full resolution, and ablation studies on this are provided in the supplementary material.", + "bbox": [ + 511, + 449, + 906, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Observations and analysis. As shown in Fig. 4, we observe that (1) mixed-resolution denoising can boost achievable generation fidelity by notably reducing the optimal FID (lower is better) while simultaneously improving generation efficiency, aligning with the analysis of the role of low resolution in early denoising steps in Sec. 2.1 and Sec. 2.3; (2) With more inference steps $(T \\geq 14)$ , the optimal values of $s$ are around $0.4 \\sim 0.5$ , offering the best FID-FLOPs trade-off. For example, adopting $s = 0.5$ with 20 denoising steps can save about $37\\%$ FLOPs while achieving an FID reduction of approximately 2 compared to the original full-resolution model. (3) When aiming for relatively high generation fidelity, e.g., a $< 20$ FID, which aligns with real-world application demands, maintaining a high number of denoising steps and reducing per-step inference cost via mixed resolution is a more cost-effective strategy than reducing denoising steps. Conversely, when targeting extreme efficiency, e.g., reducing $>60\\%$ FLOPs, using fewer denoising steps becomes increasingly beneficial, as the effectiveness of mixed-resolution denoising diminishes in low-step generation with larger gaps between steps.", + "bbox": [ + 511, + 501, + 908, + 819 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Calibration for best hyperparameters. In practice, to rapidly determine the best transition step $s$ (as well as the total number of denoising steps) for a given diffusion model provided by the users, we can use a small calibration set to obtain the optimal setting. We show in the supplemen", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3004", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tary materials that performance on a small calibration set is highly correlated with that on the full evaluation set.", + "bbox": [ + 89, + 90, + 485, + 122 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.5. The Hybrid Module Caching Strategy", + "text_level": 1, + "bbox": [ + 89, + 133, + 419, + 152 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition to the resolution dimension, we further reduce model complexity at the module level by integrating existing module caching techniques [25, 42] with an enhanced hybrid caching strategy.", + "bbox": [ + 89, + 157, + 483, + 218 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Motivation. Previous works [25, 42] have found that feature maps remain highly similar across all blocks throughout the entire denoising process, indicating the potential effectiveness of caching and reusing across steps. We have further validated this through cross-step feature map distance visualization in the supplementary materials. Considering that (1) the text guidance provided by cross-attention layers primarily determines the image layout, i.e., the low-frequency structure of the image, and (2) the semantics-planning phase for layout generation occurs in the early denoising stages [22], as also discussed in Sec. 2.1, it is feasible to use the cached cross-attention feature map from early denoising steps, generated at a low resolution with our mixed-resolution denoising, for the later denoising phase, as the low-frequency information is determined early on.", + "bbox": [ + 89, + 219, + 485, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our caching strategy. We employ a hybrid module caching strategy to maximize reuse opportunities in a pretrained diffusion model. To achieve this, we integrate the following two caching mechanisms: (1) similar to DeepCache [25], we cache the deep skip branches of the U-Net, a major component of a wide range of SOTA diffusion models [17, 24, 29, 31], and reuse them for the subsequent $k$ steps. Here, $k$ controls how often the cache is updated, balancing efficiency gains and potential degradation, and is set to 2 throughout the evaluation; (2) We cache the crossattention layers at a specific denoising step and reuse them in all subsequent steps [22].", + "bbox": [ + 89, + 462, + 483, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Design choices for caching cross-attention. A major role of cross-attention is to introduce text guidance into the generation process. This is often achieved through classifier-free guidance (CFG), which provides a controllable method for incorporating conditional guidance into image generation [7, 14, 27]. Specifically, the text guidance from the model's implicit classifier is derived as follows:", + "bbox": [ + 89, + 643, + 485, + 750 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {\\theta} \\left(x _ {t}, t, w, c\\right) = \\epsilon_ {\\theta} \\left(x _ {t}, t, \\emptyset\\right) + w \\left(\\epsilon_ {\\theta} \\left(x _ {t}, t, c\\right) - \\epsilon_ {\\theta} \\left(x _ {t}, t, \\emptyset\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 779, + 488, + 810 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\varnothing$ represents the null text condition. Considering that CFG may be redundant in the later stages of inference, as the layout has already been organized in the early stages [3, 22], we abandon CFG after a certain step $m$ . When combined with our mixed-resolution denoising scheme, most", + "bbox": [ + 89, + 824, + 485, + 902 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/3a142858f42efcf3739b1b61a78a7100b4a1eb26004de2c73503e92fcd95da42.jpg", + "table_caption": [ + "Table 1. Ablation study on different caching strategies ( $m$ , Design Choice)." + ], + "table_footnote": [], + "table_body": "
SettingFLOPs (T) ↓FID ↓Clip Score ↑
Original30.42018.4230.80
Original w/o CFG15.21031.9825.52
DC17.78717.7930.75
DC+CA (5, Ave)18.7728.40
DC+CA (5, Cond)11.61018.8229.20
DC+CA (5, Uncond)20.3325.50
DC+CA (5, CFG)103.7118.22
DC+CA (10, Ave)21.1330.03
DC+CA (10, Cond)15.06121.2630.11
DC+CA (10, Uncond)21.1929.83
DC+CA (10, CFG)24.0029.12
DC+CA (15, Ave)21.7030.35
DC+CA (15, Cond)16.36021.6730.37
DC+CA (15, Uncond)21.8230.34
DC+CA (15, CFG)19.4330.48
", + "bbox": [ + 526, + 121, + 890, + 309 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CFG operations are performed under a low generation resolution, further reducing computational costs.", + "bbox": [ + 511, + 321, + 903, + 351 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on the type of guidance to be reused, there are four design choices for caching the cross-attention maps for subsequent steps:", + "bbox": [ + 511, + 352, + 906, + 396 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Ave [22]: $CA_{\\mathrm{cache}} = \\frac{1}{2}\\left(CA_t^c +CA_t^\\emptyset\\right)$", + "- Cond: $CA_{\\mathrm{cache}} = CA_{t}^{c}$", + "- Uncond: $CA_{\\mathrm{cache}} = CA_t^0$", + "- CFG: $CA_{\\mathrm{cache}} = CA_t^\\emptyset + w(CA_t^c - CA_t^\\emptyset)$" + ], + "bbox": [ + 511, + 397, + 808, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $CA_{t}^{c}$ and $CA_{t}^{\\emptyset}$ are the cross-attention conditioned on the input text and null text, respectively. Using either of these maps or a mixture of them results in varying degrees of conditioning in the subsequent generation process.", + "bbox": [ + 511, + 462, + 906, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Exploring the best cross-attention caching strategy. We empirically examine combinations of different caching strategies, including DeepCache (DC) combined with the four design choices for cross-attention caching on top of an SD V1.5 model. As shown in Tab. 1, we observe that (1) fully abandoning CFG leads to a collapse in performance, while incorporating the cross-attention caching mechanism can better balance performance and speed; (2) adding CA on top of DC further reduces computations due to redundant CFG usage in later denoising phases. Intuitively, using a smaller $m$ , i.e., abandoning CFG in earlier timesteps, reduces more FLOPs at the cost of a slight decrease in generation quality; (3) among the four design choices, \"Cond\" generally performs best, slightly surpassing \"Ave\" at larger $m$ values. Therefore, we adopt \"Cond\" as the $CA_{\\mathrm{cache}}$ choice in all subsequent experiments.", + "bbox": [ + 511, + 523, + 908, + 765 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3. PostDiff: Combine All Techniques", + "text_level": 1, + "bbox": [ + 511, + 777, + 823, + 796 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 511, + 801, + 689, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We apply our PostDiff framework, integrating mixed-resolution denoising and hybrid module caching, to several SOTA diffusion models, including SD V1.5 (DreamShaper-7 version) [31], Latent Consistency Model (LCM) [24, 36], Stable Diffusion XL (SDXL) [29], and PixArt- $\\alpha$ [5]. We", + "bbox": [ + 511, + 824, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3005", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "evaluate text-to-image generation performance on the MS-COCO 2014 validation dataset [21]. PostDiff settings are calibrated on a small set for each model, with the detailed configurations outlined as follows:", + "bbox": [ + 89, + 90, + 482, + 151 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. SD V1.5 [31]: We use the DreamShaper-v7 fine-tuned version with a final resolution of $768 \\times 768$ . Settings: $\\beta = 1/2$ , $s = 1/2$ , $w = 7.5$ , $m = 15$ .", + "2. Latent Consistency Model (LCM) [24, 36]: A distilled version of Stable Diffusion designed for few-step generation, producing $768 \\times 768$ images. Settings: $s = 1/2$ , $\\beta = 1/2$ , $w = 7.5$ , $m = 4$ . Note that since CFG is integrated into a guidance embedding [26] and the model has a few-step nature, only $CA_{\\mathrm{cache}}$ is used without DC.", + "3. Stable Diffusion XL (SDXL) [29]: A large-scale diffusion model capable of generating $1024 \\times 1024$ images. Settings: $\\beta = 3/4$ , $s = 1/5$ , $w = 5.0$ , $m = 15$ .", + "4. PixArt- $\\alpha$ [5]: A transformer-based diffusion model generating $1024 \\times 1024$ images. Due to the absence of bypass connections, only $CA_{\\mathrm{cache}}$ is used without enabling DC. Settings: $\\beta = 3/4$ , $s = 1/2$ , $w = 4.5$ , $m = 15$ ." + ], + "bbox": [ + 89, + 154, + 482, + 397 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2. Fewer Denoising Steps vs. Cheaper Per-step Inference", + "text_level": 1, + "bbox": [ + 89, + 406, + 482, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After combining the hybrid module caching with our mixed-resolution strategy to establish the complete PostDiff framework, we examine our key question by benchmarking with the FID-FLOPs/Clip Score-FLOPs trade-offs achieved by reducing denoising steps. Specifically, for SD V1.5, we use 20 denoising steps, a fixed mixed-resolution setting of $s = 1/2$ and $\\beta = 1/2$ , and vary cache settings $m$ to create the trade-off. As for PixArt- $\\alpha$ , we use a fixed 20 denoising steps with $\\beta = 3/4$ , and vary $m$ and $s$ to form the trade-off.", + "bbox": [ + 89, + 444, + 482, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Observations and insights. From Fig. 5, we observe that (1) the complete PostDiff framework, incorporating both techniques, achieves the best FID-FLOPs trade-off across all FLOPs ranges. This demonstrates that exploiting model redundancy at both the input and module levels can yield better FID-FLOPs frontiers; (2) in terms of semantic alignment measured by Clip Score, reducing per-step inference cost can achieve a comparable Clip Score-FLOPs trade-off compared to decreasing the number of denoising steps over a wide range; while in scenarios requiring extreme efficiency, decreasing the number of denoising steps can maintain better semantic alignment with a slightly higher Clip Score. Generally, reducing per-step inference cost is more effective than decreasing the number of denoising steps with better fidelity and comparable semantic alignment across a wide efficiency range.", + "bbox": [ + 89, + 580, + 482, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3. Evaluation on SOTA Diffusion Models", + "text_level": 1, + "bbox": [ + 89, + 832, + 421, + 848 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We further apply PostDiff to more diffusion models. As shown in Tab. 2, we present the evaluation results in terms of both generation quality and efficiency under (1) reduced", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c5e652a211aa5b85391ba0656d5af52224fea297dab7a5dc223b6c232d3cc68f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 88, + 709, + 234 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8e9bfa759e3976a8c6774708f5fae89e4a918fc30b330f856c87fc4ad13b0392.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 88, + 897, + 234 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/aa2e3c0349cd93b32b763868b52233031b40e3b2ec630b2fa58a7a947cfb17fb.jpg", + "image_caption": [ + "Figure 5. Comparison of FID-FLOPs/Clip Score-FLOPs trade-offs achieved by reducing per-step inference cost or the number of denoising steps on SD V1.5 and PixArt- $\\alpha$ ." + ], + "image_footnote": [], + "bbox": [ + 517, + 238, + 709, + 386 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7d3291122459d82ba4fb5a45ec4a2905e319531fb2378879f59adf4f1f4e30d1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 238, + 897, + 386 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "denoising steps or (2) maximal denoising steps with reduced model complexity via PostDiff.", + "bbox": [ + 511, + 455, + 903, + 486 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Observations and analysis. We observe that: (1) With reduced denoising steps, the improved efficiency comes at the cost of degraded generation quality, reflected in notably increased FID and reduced Clip score. In contrast, maintaining the maximum number of denoising steps while applying PostDiff better preserves generation quality, achieving lower (i.e., better) FID and comparable Clip scores compared to reducing the denoising steps. For instance, on the SD V1.5 model, PostDiff achieves a $63.14\\%$ reduction in FLOPs while improving FID by 1.8 compared to the original model. (2) PostDiff consistently demonstrates effectiveness across various diffusion models, including both large-scale and small-scale ones, LDMs and LCM, as well as U-Net-based and transformer-based models. This effectiveness is due to the broad applicability of the insights leveraged in Sec. 2.1 and Sec. 2.5. We also visualize the generation results in Fig. 6.", + "bbox": [ + 511, + 487, + 905, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Benchmark with Other Training-Free Compression Methods", + "text_level": 1, + "bbox": [ + 511, + 755, + 903, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We further benchmark PostDiff with other training-free diffusion model compression methods in Tab. 3 and Tab. 4.", + "bbox": [ + 511, + 792, + 903, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Settings. For the U-Net-based model SD V1.5, we evaluate our method against DeepCache [25], TGATE [22], ToDo [35], ToMe [1], Faster Diffusion [20], and CA-ToMe [33]. Latency is measured as the single-image generation time on an NVIDIA A5000 GPU. The cache interval", + "bbox": [ + 511, + 825, + 903, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3006", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/eaa6554f087eba3df5eac04250ae0c2396438edc38e6b54fa6fab9b840b8775b.jpg", + "table_caption": [ + "Table 2. Apply PostDiff with mixed-resolution denoising and hybrid module caching to SOTA diffusion models with varying numbers of denoising steps. Latency is measured as the single-image generation time on an NVIDIA A5000 GPU." + ], + "table_footnote": [], + "table_body": "
ModelStepsMixCacheFID ↓Clip Score ↑FLOPs (T) ↓Latency (s) ↓
SD V1.5 [31]820.6030.4112.1681.298
1617.7430.7624.3362.377
2018.4230.8030.4202.930
20✓15.6930.7819.0351.945
20✓✓16.6530.2511.1291.139
LCM [24]423.5528.643.0420.523
822.9628.726.0840.825
8✓21.0128.923.8070.666
8✓✓22.7628.583.6860.651
SDXL [29]818.0130.9247.8562.843
1614.3531.9295.7125.211
2014.1031.95119.6416.521
20✓14.3731.62109.2645.950
20✓✓14.1831.1152.6823.119
PixArt-α [5]833.0930.2134.2563.031
1629.6230.4068.5125.725
2029.1630.4185.6407.093
20✓25.9930.3866.9905.645
20✓✓25.4430.2354.7184.768
", + "bbox": [ + 114, + 127, + 883, + 467 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5545c291e943d864851e04b7180c88d34274ed3ffe347e96b030a64b1243ebe7.jpg", + "table_caption": [ + "Table 3. Benchmark with other compression methods on SD V1.5." + ], + "table_footnote": [], + "table_body": "
MethodFID ↓Clip Score ↑Latency (s) ↓
Original18.4230.802.930
DeepCache [25]17.7930.751.737
TGATE [22]19.5129.551.992
ToDo [35]18.2830.752.520
ToMe [1]17.4330.552.730
Faster Diffusion [20]17.9030.402.005
CA-ToMe [33]20.4930.412.112
Ours16.6530.251.139
", + "bbox": [ + 94, + 513, + 490, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "for DeepCache is set to 2. For TGATE, we set $m = 5$ . For ToMe, we use a merging ratio of 0.3, while ToDo, Faster Diffusion, and CA-ToMe are evaluated using their default hyperparameters from their respective papers. For the transformer-based model PixArt- $\\alpha$ , we compare Post-Diff with TGATE [22] using $m = 8$ , as well as DiTFastAttn [46] and DuCa [49] with their default settings on an NVIDIA H100 GPU, since some methods encounter out-of-memory errors on the A5000.", + "bbox": [ + 88, + 670, + 482, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Observations. Compared to previous methods that accelerate the generation process either at the module level (e.g., DeepCache, TGATE, DuCa) by reusing computations across denoising steps or at the input level (e.g., ToMe, ToDo) by compressing redundant image tokens, our method uniquely integrates both approaches, resulting in a greater", + "bbox": [ + 88, + 809, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6a2319957becacf4223674363d7e89a7677ca875e9da5082e9bdc1089f72787a.jpg", + "table_caption": [ + "Table 4. Benchmark with other training-free compression methods on PixArt-α." + ], + "table_footnote": [ + "Faster speed is achieved with FlashAttention [6]." + ], + "table_body": "
MethodFID ↓Clip Score ↑Latency (s) ↓
Original29.1630.411.752
TGATE [22]29.5429.591.275
DiTFastAttn [46]45.4630.13\\( 1.103^{\\dagger} \\)
DuCa [49]28.0830.121.724
Ours25.4430.231.382
", + "bbox": [ + 514, + 527, + 903, + 641 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "reduction in inference cost. Additionally, leveraging low-resolution images in the early stages enhances the denoising process, allowing our method to achieve the lowest FID with a competitive Clip Score across both SD V1.5 and PixArt- $\\alpha$ models.", + "bbox": [ + 511, + 671, + 906, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "These experiments highlight the benefits of leveraging redundancy at both the input and module levels. Integrating the aforementioned training-free compression techniques into PostDiff could further enhance efficiency, which we leave for future work.", + "bbox": [ + 511, + 748, + 906, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4. Related Work", + "text_level": 1, + "bbox": [ + 511, + 843, + 655, + 858 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Diffusion models. Diffusion models have emerged as a powerful approach for generative tasks, modeling the gen", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3007", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/315f9eb9be4ad714b545b7392630c906b8d3116c20b6812209d8b735fd140a07.jpg", + "image_caption": [ + "Figure 6. Visualize the generation results of different models w/o and w/ using our PostDiff." + ], + "image_footnote": [], + "bbox": [ + 116, + 88, + 895, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "erative process as a denoising sequence to enable high-quality image generation [15, 36]. Latent Diffusion Models (LDMs) [31] leverage Variational Autoencoders (VAEs) [30] to operate in a compressed latent space, reducing memory demands and enabling efficient high-resolution generation. Open-source models like Stable Diffusion V1.5 [31], Stable Diffusion XL [29], and PixArt [4, 5] have further expanded accessibility and are now applied across multiple domains [8, 10, 11]. Despite their impressive generative quality, diffusion models often suffer from long generation times, limiting their broader applicability.", + "bbox": [ + 88, + 474, + 482, + 641 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Few-Step diffusion models. A primary focus in accelerating diffusion models is reducing the number of timesteps. DDIM [36] modifies DDPM [15] by introducing non-Markovian sampling, enabling faster generation. DPM-Solver [23] applies exponential integrators to approximate solutions to diffusion ODEs, enhancing sampling efficiency. Other approaches, such as consistency models [24, 37], aim to map points on the ODE trajectory to the original sample, effectively reducing inference steps. Adversarial training is also employed to enable one-step or few-step diffusion [43-45]. However, these few-step models still incur non-trivial training/distillation cost.", + "bbox": [ + 88, + 642, + 482, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Efficient diffusion models. Efforts to accelerate diffusion models extend from data optimization to model optimization. From the data perspective, prior works have reduced computational requirements by merging similar tokens [1] or pruning redundant tokens [39]. From the model", + "bbox": [ + 88, + 824, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "perspective, recent works focus on identifying redundant feature maps across timesteps and using caching mechanisms to reuse these maps, significantly reducing computational costs [25, 42]. Additional approaches aim to slim down models through quantization to lower bit precision [12, 34, 48] or by pruning unnecessary parameters [2, 9, 41]. Additionally, several systems have been developed specifically to accelerate diffusion model inference [18, 19].", + "bbox": [ + 511, + 474, + 905, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 616, + 633, + 632 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we explore a critical question for compute-optimal diffusion model deployment: under a post-training setting, is it more effective to reduce denoising steps or decrease per-step inference cost? To investigate this, we propose PostDiff, a training-free framework that enhances diffusion model efficiency while systematically examining this trade-off. PostDiff reduces redundancy at two levels: at the input level, through a mixed-resolution denoising scheme that enhances low-frequency components in early steps, and at the module level, with a hybrid caching strategy. Extensive evaluations demonstrate that reducing per-step inference cost often achieves a better efficiency-fidelity balance than reducing denoising steps across a wide efficiency range. PostDiff not only offers a new perspective on the fidelity-efficiency trade-off for diffusion models but also provides practical insights for their deployment on resource-constrained platforms.", + "bbox": [ + 509, + 643, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3008", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported by the National Science Foundation (NSF) through CCRI funding (Award Number: 2016727), RTML funding (Award Number: 1937592), and an NSF CAREER award (Award Number: 2048183), as well as by CoCoSys, one of the seven centers in JUMP 2.0, a Semiconductor Research Corporation (SRC) program sponsored by DARPA.", + "bbox": [ + 89, + 114, + 483, + 220 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 234, + 187, + 250 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Daniel Bolya and Judy Hoffman. Token merging for fast stable diffusion. In CVPR, 2023. 3, 6, 7, 8", + "[2] Thibault Castells, Hyoung-Kyu Song, Bo-Kyeong Kim, and Shinkook Choi. LD-Pruner: Efficient pruning of latent diffusion models using task-agnostic insights. In CVPR, 2024. 8", + "[3] Angela Castillo, Jonas Kohler, Juan C Pérez, Juan Pablo Pérez, Albert Pumarola, Bernard Ghanem, Pablo Arbeláez, and Ali Thabet. Adaptive guidance: Training-free acceleration of conditional diffusion models. In AAAI, 2025. 5", + "[4] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. PixArt- $\\Sigma$ : Weak-to-strong training of diffusion transformer for 4k text-to-image generation. In ECCV, 2024. 2, 8", + "[5] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Zhongdao Wang, James T. Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. PixArt-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In ICLR, 2024. 5, 6, 7, 8", + "[6] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. FlashAttention: Fast and memory-efficient exact attention with io-awareness. In NeurIPS, 2022. 7", + "[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 5", + "[8] Zhenbang Du, Wei Feng, Haohan Wang, Yaoyu Li, Jingsen Wang, Jian Li, Zheng Zhang, Jingjing Lv, Xin Zhu, Junsheng Jin, et al. Towards reliable advertising image generation using human feedback. In ECCV, 2024. 8", + "[9] Gongfan Fang, Xinyin Ma, and Xinchao Wang. Structural pruning for diffusion models. In NeurIPS, 2023. 8", + "[10] Sicheng Gao, Xuhui Liu, Bohan Zeng, Sheng Xu, Yanjing Li, Xiaoyan Luo, Jianzhuang Liu, Xiantong Zhen, and Baochang Zhang. Implicit diffusion models for continuous super-resolution. In CVPR, 2023. 8", + "[11] He Guo, Zixuan Ye, Zhiguo Cao, and Hao Lu. In-context matting. In CVPR, 2024. 8", + "[12] Yefei He, Luping Liu, Jing Liu, Weijia Wu, Hong Zhou, and Bohan Zhuang. PTQD: Accurate post-training quantization for diffusion models. In NeurIPS, 2023. 8", + "[13] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. In EMNLP, 2021. 3", + "[14] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshop, 2021. 5" + ], + "bbox": [ + 93, + 258, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 3, 8", + "[16] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. JMLR, 2022. 2", + "[17] Bo-Kyeong Kim, Hyoung-Kyu Song, Thibault Castells, and Shinkook Choi. BK-SDM: A lightweight, fast, and cheap version of stable diffusion. In ECCV, 2024. 5", + "[18] Akio Kodaira, Chenfeng Xu, Toshiki Hazama, Takanori Yoshimoto, Kohei Ohno, Shogo Mitsuhori, Soichi Sugano, Hanying Cho, Zhijian Liu, and Kurt Keutzer. Streamdiffusion: A pipeline-level solution for real-time interactive generation. arXiv preprint, 2023. 8", + "[19] Muyang Li, Tianle Cai, Jiaxin Cao, Qinsheng Zhang, Han Cai, Junjie Bai, Yangqing Jia, Kai Li, and Song Han. Distribution: Distributed parallel inference for high-resolution diffusion models. In CVPR, 2024. 2, 8", + "[20] Senmao Li, Taihang Hu, Fahad Shahbaz Khan, Linxuan Li, Shiqi Yang, Yaxing Wang, Ming-Ming Cheng, and Jian Yang. Faster diffusion: Rethinking the role of unet encoder in diffusion models. In NeurIPS, 2024. 6, 7", + "[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 4, 6", + "[22] Haozhe Liu, Wentian Zhang, Jinheng Xie, Francesco Faccio, Mengmeng Xu, Tao Xiang, Mike Zheng Shou, Juan-Manuel Pérez-Rúa, and Jurgen Schmidhuber. Faster diffusion through temporal attention decomposition. TMLR, 2025. 2, 5, 6, 7", + "[23] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. NeurIPS, 2022. 3, 8", + "[24] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint, 2023. 5, 6, 7, 8", + "[25] Xinyin Ma, Gongfan Fang, and Xinchao Wang. Deepcache: Accelerating diffusion models for free. In CVPR, 2024. 5, 6, 7, 8", + "[26] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. In CVPR, 2023. 6", + "[27] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2021. 5", + "[28] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 3", + "[29] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024. 2, 5, 6, 7, 8" + ], + "bbox": [ + 514, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3009", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. In NeurIPS, 2019. 3, 8", + "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 4, 5, 6, 7, 8", + "[32] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 3", + "[33] Omid Saghatchian, Atiyeh Gh Moghadam, and Ahmad Nickabadi. Cached adaptive token merging: Dynamic token reduction and redundant computation elimination in diffusion model. arXiv preprint, 2025. 6, 7", + "[34] Yuzhang Shang, Zhihang Yuan, Bin Xie, Bingzhe Wu, and Yan Yan. Post-training quantization on diffusion models. In CVPR, 2023. 8", + "[35] Ethan Smith, Nayan Saxena, and Aninda Saha. Todo: Token downsampling for efficient generation of high-resolution images. In *IJCAI*, 2024. 3, 6, 7", + "[36] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 3, 5, 6, 8", + "[37] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In ICML, 2023. 8", + "[38] Jiayan Teng, Wendi Zheng, Ming Ding, Wenyi Hong, Jianqiao Wangni, Zhuoyi Yang, and Jie Tang. Relay diffusion: Unifying diffusion process across resolutions for image synthesis. In ICLR, 2024. 2", + "[39] Hongjie Wang, Difan Liu, Yan Kang, Yijun Li, Zhe Lin, Niraj K Jha, and Yuchen Liu. Attention-driven training-free efficiency enhancement of diffusion models. In CVPR, 2024. 3, 8", + "[40] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. LAVIE: High-quality video generation with cascaded latent diffusion models. IJCV, 2024. 2", + "[41] Lexington Whalen, Zhenbang Du, Haoran You, Chaojian Li, Sixu Li, and Yingyan (Celine) Lin. Early-Bird Diffusion: Investigating and leveraging timestep-aware early-bird tickets in diffusion models for efficient training. In CVPR, 2025. 8", + "[42] Felix Wimbauer, Bichen Wu, Edgar Schoenfeld, Xiaoliang Dai, Ji Hou, Zijian He, Artsiom Sanakoyeu, Peizhao Zhang, Sam Tsai, Jonas Kohler, et al. Cache me if you can: Accelerating diffusion models through block caching. In CVPR, 2024. 5, 8", + "[43] Yanwu Xu, Yang Zhao, Zhisheng Xiao, and Tingbo Hou. UFOGen: You forward once large scale text-to-image generation via diffusion gans. In CVPR, 2024. 8", + "[44] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved distribution matching distillation for fast image synthesis. In NeurIPS, 2024.", + "[45] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Frédo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In CVPR, 2024. 8" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Zhihang Yuan, Hanling Zhang, Lu Pu, Xuefei Ning, Linfeng Zhang, Tianchen Zhao, Shengen Yan, Guohao Dai, and Yu Wang. DiTFastattn: Attention compression for diffusion transformer models. In NeurIPS, 2024. 7", + "[47] Shen Zhang, Zhaowei Chen, Zhenyu Zhao, Yuhao Chen, Yao Tang, and Jiajun Liang. HiDiffusion: Unlocking higher-resolution creativity and efficiency in pretrained diffusion models. In ECCV, 2025. 2", + "[48] Tianchen Zhao, Tongcheng Fang, Enshu Liu, Rui Wan, Widyadewi Soedarmadji, Shiyao Li, Zinan Lin, Guohao Dai, Shengen Yan, Huazhong Yang, Xuefei Ning, and Yu Wang. ViDiT-Q: Efficient and accurate quantization of diffusion transformers for image and video generation. In ICLR, 2025. 8", + "[49] Chang Zou, Evelyn Zhang, Runlin Guo, Haohang Xu, Conghui He, Xuming Hu, and Linfeng Zhang. Accelerating diffusion transformers with dual feature caching. arXiv preprint, 2024. 7" + ], + "bbox": [ + 514, + 92, + 903, + 345 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3010", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Fewer Denoising Steps or Cheaper Per-Step Inference_ Towards Compute-Optimal Diffusion Model Deployment/98cafbf9-4954-46a4-b5af-59f0bca621f7_model.json b/2025/Fewer Denoising Steps or Cheaper Per-Step Inference_ Towards Compute-Optimal Diffusion Model Deployment/98cafbf9-4954-46a4-b5af-59f0bca621f7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c9481e21eee6e8dc2ca275e78fa6c78fa5aba59c --- /dev/null +++ b/2025/Fewer Denoising Steps or Cheaper Per-Step Inference_ Towards Compute-Optimal Diffusion Model Deployment/98cafbf9-4954-46a4-b5af-59f0bca621f7_model.json @@ -0,0 +1,2123 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.807, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.178, + 0.131, + 0.82, + 0.177 + ], + "angle": 0, + "content": "Fewer Denoising Steps or Cheaper Per-Step Inference: Towards Compute-Optimal Diffusion Model Deployment" + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.204, + 0.756, + 0.277 + ], + "angle": 0, + "content": "Zhenbang Du*, Yonggan Fu*, Lifu Wang*, Jiayi Qian, Xiao Luo, Yingyan (Celine) Lin \nGeorgia Institute of Technology \n{zdu89,jiayiqian,celine.lin}@gatech.edu" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.342, + 0.485, + 0.78 + ], + "angle": 0, + "content": "Diffusion models have shown remarkable success across generative tasks, yet their high computational demands challenge deployment on resource-limited platforms. This paper investigates a critical question for compute-optimal diffusion model deployment: Under a post-training setting without fine-tuning, is it more effective to reduce the number of denoising steps or to use a cheaper per-step inference? Intuitively, reducing the number of denoising steps increases the variability of the distributions across steps, making the model more sensitive to compression. In contrast, keeping more denoising steps makes the differences smaller, preserving redundancy, and making post-training compression more feasible. To systematically examine this, we propose PostDiff, a training-free framework for accelerating pre-trained diffusion models by reducing redundancy at both the input level and module level in a post-training manner. At the input level, we propose a mixed-resolution denoising scheme based on the insight that reducing generation resolution in early denoising steps can enhance low-frequency components and improve final generation fidelity. At the module level, we employ a hybrid module caching strategy to reuse computations across denoising steps. Extensive experiments and ablation studies demonstrate that (1) PostDiff can significantly improve the fidelity-efficiency trade-off of state-of-the-art diffusion models, and (2) to boost efficiency while maintaining decent generation fidelity, reducing per-step inference cost is often more effective than reducing the number of denoising steps. Our code is available at https://github.com/GATECH-EIC/PostDiff." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.808, + 0.222, + 0.823 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.833, + 0.483, + 0.879 + ], + "angle": 0, + "content": "Diffusion models have rapidly advanced as a powerful tool for various generative tasks, such as image synthesis and video generation. Their success stems from the iterative de" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.311, + 0.905, + 0.476 + ], + "angle": 0, + "content": "noising process, which gradually refines outputs to achieve high-quality results, along with advanced model architectures. However, both the iterative nature and model complexity of diffusion models make them computationally intensive, posing a significant challenge for deployment on resource-constrained platforms, where computational resources, memory, and power are limited. Efficient deployment of diffusion models on these platforms is therefore highly desirable to broaden the accessibility of generative AI, enabling real-time applications that can operate within the strict constraints of on-device resources." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.905, + 0.75 + ], + "angle": 0, + "content": "With the increasing demand for rapid diffusion model deployment, where a given diffusion model is expected to be instantly compressed and deployed without fine-tuning, a critical question arises: Under a post-training setting, is it more effective to optimize for fewer denoising steps or to reduce the per-step inference cost? This question is particularly important because each approach entails different trade-offs: Reducing the number of denoising steps introduces greater variability in feature changes between steps, which can amplify the negative impact of compression on output fidelity; Conversely, using more denoising steps reduces these step-to-step differences, preserving redundancy and making it easier to apply post-training compression without sacrificing quality. However, there is a limited exploration of how these two strategies perform in terms of fidelity-efficiency trade-offs, and investigating this could provide valuable guidelines for efficient diffusion model deployment in real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.905, + 0.901 + ], + "angle": 0, + "content": "To address this research question, we propose PostDiff, a training-free diffusion model compression framework designed to enhance the per-step inference efficiency of pretrained diffusion models at both the input level and module level in a post-training manner. Our techniques are training-free and can thus be instantly applied to emerging diffusion models for rapid deployment, enabling a comparative study of their effectiveness versus reducing denoising steps. Specifically, at the input level, we propose a mixed-resolution denoising scheme that adopts a lower generation" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.22, + 0.9 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3001" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.242 + ], + "angle": 0, + "content": "resolution in early denoising steps. As low-frequency components dominate at this stage, a reduced resolution is sufficient. We find that this reduced resolution in early denoising steps can enhance low-frequency components, thereby improving final generation fidelity and achieving a win-win in both efficiency and fidelity. At the module level, we build on existing module caching techniques, which reuse module outputs across steps, and enhance them with a hybrid caching strategy as another dimension for redundancy reduction. Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.244, + 0.483, + 0.334 + ], + "angle": 0, + "content": "- Our work, for the first time, studies a critical question for efficient diffusion model deployment: Under a posttraining setting without fine-tuning, is it more effective to optimize for fewer denoising steps or to reduce the per-step inference cost? The answer to this question could be useful for both researchers and practitioners." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.335, + 0.483, + 0.456 + ], + "angle": 0, + "content": "- To answer the above question, we propose PostDiff, a training-free framework for reducing diffusion model complexity at both the input and module levels, resulting in lower per-step inference cost. This framework enables a fair study of the relative effectiveness between fewer denoising steps and reduced per-step inference cost toward compute-optimal diffusion model deployment under a post-training setting." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.456, + 0.483, + 0.591 + ], + "angle": 0, + "content": "- At the input level, we propose a mixed-resolution denoising scheme, which leverages the observation that reduced generation resolution in early denoising steps can enhance low-frequency components and final output quality, achieving a win-win in both accuracy and efficiency. In addition, to further reduce the overall computational cost at the module level, PostDiff employs a hybrid module caching method that strategically reuses the outputs of different modules on varying schedules across steps." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.591, + 0.483, + 0.712 + ], + "angle": 0, + "content": "- Extensive evaluations show that PostDiff can advance the fidelity-efficiency frontier of state-of-the-art (SOTA) diffusion models, as shown in Fig. 1, and provide valuable insights into best practices for efficient diffusion model deployment. For example, we find that reducing per-step inference cost is often more effective than decreasing the number of denoising steps to improve efficiency when aiming to preserve high generation fidelity." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.244, + 0.483, + 0.712 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.714, + 0.483, + 0.775 + ], + "angle": 0, + "content": "In the following sections, we illustrate and analyze the effectiveness of each technique in PostDiff in Sec. 2, and finally, we combine and apply them to SOTA diffusion models in Sec. 3 to answer our research question." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.785, + 0.416, + 0.803 + ], + "angle": 0, + "content": "2. PostDiff: The Proposed Framework" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In this section, we first introduce the proposed mixed-resolution denoising strategy, which reduces per-step inference cost at the input level. We begin by analyzing the motivations from previous works in Sec. 2.1, followed by an illustration of our method, an analysis of its effectiveness, and its application to examine our key question in Sec. 2.2," + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.086, + 0.897, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.29, + 0.905, + 0.319 + ], + "angle": 0, + "content": "Figure 1. Apply our PostDiff technique to SOTA diffusion models, where “-PD” denotes the models delivered by PostDiff." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.326, + 0.905, + 0.372 + ], + "angle": 0, + "content": "Sec. 2.3, and Sec. 2.4, respectively. Finally, we describe how we employ and enhance module caching to improve efficiency at the module level in Sec. 2.5." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.381, + 0.81, + 0.396 + ], + "angle": 0, + "content": "2.1. Motivations from Previous Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.403, + 0.905, + 0.584 + ], + "angle": 0, + "content": "Low-frequency components dominate the early denoising steps. Previous works [22] find that the typical denoising process for image generation progresses through two phases: the semantics-planning phase and the fidelity-improving phase. The early semantics-planning phase focuses more on the layout and low-frequency information, starting from random noise and gradually establishing the basic structure. This indicates that a low generation resolution could be sufficient in the early denoising steps. As generation proceeds, more details, described by high-frequency information, are incrementally added, thus requiring a higher resolution." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.585, + 0.906, + 0.78 + ], + "angle": 0, + "content": "This insight is also partially echoed and leveraged by previous works. For example, [38] trains different diffusion models for early low-resolution denoising and later high-resolution denoising processes to avoid the propagation of inaccurate high-frequency components from early denoising steps to later steps. Additionally, cascade diffusion models [16, 40] sequentially apply different diffusion models at different denoising stages with varying generation resolutions, finding that incorporating low-frequency information from low-resolution images during high-resolution stages can enhance generation quality. Nevertheless, all these methods require multiple diffusion models, introducing significant storage and training demands." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Input resolution is a promising dimension for reducing inference cost. High-resolution generation is a recent trend, as it provides images with more vivid details and accommodates greater information density [47]. For example, SDXL [29] generates images at 1024x1024 resolution, while PixArt- \\(\\Sigma\\) [4] expands resolution capability to 4K. However, as resolution increases, computational demands rise dramatically, leading to significant latency [19]. As" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3002" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.482, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.27, + 0.483, + 0.287 + ], + "angle": 0, + "content": "Figure 2. An overview of our mixed-resolution denoising strategy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.296, + 0.484, + 0.358 + ], + "angle": 0, + "content": "such, recent works aim to reduce the computational load by decreasing the number of tokens in samples, using methods like token merging [1], token pruning [39], or token down-sampling [35], which effectively accelerate inference." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.364, + 0.458, + 0.381 + ], + "angle": 0, + "content": "2.2. Mixed-Resolution Denoising: Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.385, + 0.484, + 0.506 + ], + "angle": 0, + "content": "Motivated by the tendency for low-frequency components in early denoising steps and the effectiveness of input resolution as a design knob for efficiency, we propose a simple yet effective post-training strategy: As shown in Fig. 2, we adopt a low generation resolution in the early denoising steps to generate low-frequency structures, and then refine high-frequency details with high generation resolution in the later denoising steps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.507, + 0.484, + 0.643 + ], + "angle": 0, + "content": "Formulation. To generate an image with shape \\((W,H,3)\\), we begin with Gaussian noise in the shape \\((w,h,4)\\), representing the latent noise image \\(x_{T}\\) obtained via Variational Autoencoders [30]. For simplicity, we omit the channel dimension in the remainder of this paper. The denoising process proceeds over \\(T\\) steps, using samplers like DDIM [36] or DPM-solver [23] to obtain the denoised latent at each step. Taking DDIM [36] as an example, we compute the latent representation \\(x\\) at step \\(t - 1\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.232, + 0.653, + 0.483, + 0.67 + ], + "angle": 0, + "content": "\\[\nx _ {t - 1} = \\sqrt {\\alpha_ {t - 1}} \\hat {x} _ {0} ^ {t} + x _ {t}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.673, + 0.483, + 0.707 + ], + "angle": 0, + "content": "\\[\n\\hat {x} _ {0} ^ {t} = \\frac {x _ {t} - \\sqrt {1 - \\alpha_ {t}} \\cdot \\epsilon_ {\\theta} (x _ {t} , t)}{\\sqrt {\\alpha_ {t}}}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.709, + 0.482, + 0.729 + ], + "angle": 0, + "content": "\\[\nx _ {t} = \\sqrt {1 - \\alpha_ {t - 1}} \\cdot \\epsilon_ {\\theta} (x _ {t}, t), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.78 + ], + "angle": 0, + "content": "where \\(\\epsilon_{\\theta}\\) denotes the model [28, 32] that predicts the added noise, and \\(\\{\\alpha_t\\}\\) is a set of coefficients controlling the forward noise-adding process [15]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In our proposed mixed-resolution denoising strategy, we first introduce a lower-resolution latent representation, \\( x_{T}^{l} \\), with shape \\( (\\beta w, \\beta h) \\), where \\( 0 < \\beta < 1 \\) is a hyperparameter that scales the resolution down. This low-resolution starting point reduces the computational load of the initial denoising phase. Another hyperparameter, \\( s \\), controls the transition from low to full resolution. After reaching the denoising step \\( t = sT \\), we compute the forecasted" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.09, + 0.905, + 0.121 + ], + "angle": 0, + "content": "\\(\\hat{x}_0^{l,t}\\) at step \\(t\\) using Eq. (2). Then, we upscale this result to full resolution via:" + }, + { + "type": "equation", + "bbox": [ + 0.63, + 0.13, + 0.905, + 0.15 + ], + "angle": 0, + "content": "\\[\n\\hat {x} _ {0} ^ {f, t} = \\operatorname {U p s a m p l e} \\left(\\hat {x} _ {0} ^ {l, t}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.157, + 0.907, + 0.203 + ], + "angle": 0, + "content": "where Upsample refers to bilinear interpolation. Using this upscaled version, we derive the full-resolution latent \\( x_{t}^{f} \\) at step \\( t \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.615, + 0.21, + 0.905, + 0.231 + ], + "angle": 0, + "content": "\\[\nx _ {t} ^ {f} = \\sqrt {\\alpha_ {t}} \\hat {x} _ {0} ^ {f, t} + \\sqrt {1 - \\alpha_ {t}} \\epsilon , \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.237, + 0.905, + 0.268 + ], + "angle": 0, + "content": "where \\(\\epsilon\\) is Gaussian noise with shape \\((w,h)\\). The remainder of the denoising process follows the original procedure." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.268, + 0.906, + 0.359 + ], + "angle": 0, + "content": "Note that we adopt this simple binary resolution strategy without bells and whistles to effectively demonstrate its concept and simplify the involved hyperparameters, thereby easing the use of our method in real-world deployment. More complex resolution schedules could potentially be proposed to further enhance fidelity-efficiency trade-offs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.366, + 0.906, + 0.397 + ], + "angle": 0, + "content": "2.3. Visualization: How Low-Resolution Denoising Contributes to Boosted Final Performance" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.403, + 0.906, + 0.523 + ], + "angle": 0, + "content": "To study the impact of our mixed-resolution denoising strategy on final performance and understand why it works, we visualize the generation quality after each denoising step, both with and without our method. Specifically, we calculate the Clip Score [13] for \\(\\hat{x}_0\\) predicted at each denoising step and annotate it in the case study of Fig. 3, where different rows correspond to the original full-resolution denoising process as well as our method with various choices of \\(s\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.906, + 0.78 + ], + "angle": 0, + "content": "Observations and analysis. We observe that (1) with a moderate amount of low-resolution generation steps, the Clip Score of the final generated image can be improved, achieving a win-win in generation fidelity and efficiency; (2) Regarding the evolution of the Clip Score across denoising steps, starting with a lower generation resolution results in lower Clip Scores initially, but serves as a better starting point: after switching to high generation resolution, the Clip Score increases more rapidly and gradually surpasses that of full-resolution generation. This echoes the analysis in Sec. 2.1, where reduced generation resolution in early denoising steps contributes to improved low-frequency components, thereby enhancing final generation quality; (3) The balance between low-resolution and high-resolution steps is crucial, as both low-frequency structure and high-frequency details need to be preserved to ensure high final generation quality in the last denoising step." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.906, + 0.84 + ], + "angle": 0, + "content": "This phenomenon demonstrates the underlying effectiveness of our mixed-resolution denoising strategy and is consistently observed across samples, as shown in the comprehensive evaluation in Sec. 2.4." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.864, + 0.866 + ], + "angle": 0, + "content": "2.4. Mixed-Resolution Denoising: Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We evaluate the effectiveness of our mixed-resolution denoising strategy and examine which is more effective: re" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3003" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.249, + 0.092, + 0.843, + 0.121 + ], + "angle": 0, + "content": "cute girl, crop-top, blond hair, black glasses, stretching, with background by greg rutkowski makoto shinkai kyoto animation key art feminine mid shot" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.123, + 0.9, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.408, + 0.907, + 0.438 + ], + "angle": 0, + "content": "Figure 3. Visualize the Clip Score after each denoising step using full resolution (row 1) and mixed resolution (rows 2-5). The steps using low generation resolution are highlighted in green and those using high generation resolution are highlighted in orange." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.45, + 0.48, + 0.694 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.706, + 0.483, + 0.749 + ], + "angle": 0, + "content": "Figure 4. The FID-FLOPs trade-off achieved under different denoising steps and varied mixed-resolution settings. The data points with the same number of steps are annotated using the same shape." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.762, + 0.455, + 0.777 + ], + "angle": 0, + "content": "duced input-level redundancy or fewer denoising steps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Setup. We use Stable Diffusion (SD) V1.5 (DreamShaper-7 version) [31], a popular latent diffusion model (LDM), as our model backbone and evaluate text-to-image generation performance on the MS-COCO 2014 validation dataset [21]. We vary the total denoising steps \\( T = \\{8,10,12,14,16,18,20\\} \\) and apply our mixed-resolution denoising to each denoising step setting with varied \\( s = \\{0.1,0.2,0.3,0.4,0.5,0.6\\} \\). Without bells and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.45, + 0.907, + 0.496 + ], + "angle": 0, + "content": "whistles, we set \\(\\beta = 0.5\\) as our low-resolution setting, i.e., half the full resolution, and ablation studies on this are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.502, + 0.909, + 0.82 + ], + "angle": 0, + "content": "Observations and analysis. As shown in Fig. 4, we observe that (1) mixed-resolution denoising can boost achievable generation fidelity by notably reducing the optimal FID (lower is better) while simultaneously improving generation efficiency, aligning with the analysis of the role of low resolution in early denoising steps in Sec. 2.1 and Sec. 2.3; (2) With more inference steps \\((T \\geq 14)\\), the optimal values of \\(s\\) are around \\(0.4 \\sim 0.5\\), offering the best FID-FLOPs trade-off. For example, adopting \\(s = 0.5\\) with 20 denoising steps can save about \\(37\\%\\) FLOPs while achieving an FID reduction of approximately 2 compared to the original full-resolution model. (3) When aiming for relatively high generation fidelity, e.g., a \\(< 20\\) FID, which aligns with real-world application demands, maintaining a high number of denoising steps and reducing per-step inference cost via mixed resolution is a more cost-effective strategy than reducing denoising steps. Conversely, when targeting extreme efficiency, e.g., reducing \\(>60\\%\\) FLOPs, using fewer denoising steps becomes increasingly beneficial, as the effectiveness of mixed-resolution denoising diminishes in low-step generation with larger gaps between steps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Calibration for best hyperparameters. In practice, to rapidly determine the best transition step \\( s \\) (as well as the total number of denoising steps) for a given diffusion model provided by the users, we can use a small calibration set to obtain the optimal setting. We show in the supplemen" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3004" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.123 + ], + "angle": 0, + "content": "tary materials that performance on a small calibration set is highly correlated with that on the full evaluation set." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.134, + 0.421, + 0.153 + ], + "angle": 0, + "content": "2.5. The Hybrid Module Caching Strategy" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.158, + 0.485, + 0.219 + ], + "angle": 0, + "content": "In addition to the resolution dimension, we further reduce model complexity at the module level by integrating existing module caching techniques [25, 42] with an enhanced hybrid caching strategy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.22, + 0.486, + 0.461 + ], + "angle": 0, + "content": "Motivation. Previous works [25, 42] have found that feature maps remain highly similar across all blocks throughout the entire denoising process, indicating the potential effectiveness of caching and reusing across steps. We have further validated this through cross-step feature map distance visualization in the supplementary materials. Considering that (1) the text guidance provided by cross-attention layers primarily determines the image layout, i.e., the low-frequency structure of the image, and (2) the semantics-planning phase for layout generation occurs in the early denoising stages [22], as also discussed in Sec. 2.1, it is feasible to use the cached cross-attention feature map from early denoising steps, generated at a low resolution with our mixed-resolution denoising, for the later denoising phase, as the low-frequency information is determined early on." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.463, + 0.485, + 0.644 + ], + "angle": 0, + "content": "Our caching strategy. We employ a hybrid module caching strategy to maximize reuse opportunities in a pretrained diffusion model. To achieve this, we integrate the following two caching mechanisms: (1) similar to DeepCache [25], we cache the deep skip branches of the U-Net, a major component of a wide range of SOTA diffusion models [17, 24, 29, 31], and reuse them for the subsequent \\(k\\) steps. Here, \\(k\\) controls how often the cache is updated, balancing efficiency gains and potential degradation, and is set to 2 throughout the evaluation; (2) We cache the crossattention layers at a specific denoising step and reuse them in all subsequent steps [22]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.486, + 0.751 + ], + "angle": 0, + "content": "Design choices for caching cross-attention. A major role of cross-attention is to introduce text guidance into the generation process. This is often achieved through classifier-free guidance (CFG), which provides a controllable method for incorporating conditional guidance into image generation [7, 14, 27]. Specifically, the text guidance from the model's implicit classifier is derived as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.78, + 0.489, + 0.811 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {\\theta} \\left(x _ {t}, t, w, c\\right) = \\epsilon_ {\\theta} \\left(x _ {t}, t, \\emptyset\\right) + w \\left(\\epsilon_ {\\theta} \\left(x _ {t}, t, c\\right) - \\epsilon_ {\\theta} \\left(x _ {t}, t, \\emptyset\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.486, + 0.903 + ], + "angle": 0, + "content": "where \\(\\varnothing\\) represents the null text condition. Considering that CFG may be redundant in the later stages of inference, as the layout has already been organized in the early stages [3, 22], we abandon CFG after a certain step \\(m\\). When combined with our mixed-resolution denoising scheme, most" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.907, + 0.117 + ], + "angle": 0, + "content": "Table 1. Ablation study on different caching strategies (\\(m\\), Design Choice)." + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.122, + 0.892, + 0.31 + ], + "angle": 0, + "content": "
SettingFLOPs (T) ↓FID ↓Clip Score ↑
Original30.42018.4230.80
Original w/o CFG15.21031.9825.52
DC17.78717.7930.75
DC+CA (5, Ave)18.7728.40
DC+CA (5, Cond)11.61018.8229.20
DC+CA (5, Uncond)20.3325.50
DC+CA (5, CFG)103.7118.22
DC+CA (10, Ave)21.1330.03
DC+CA (10, Cond)15.06121.2630.11
DC+CA (10, Uncond)21.1929.83
DC+CA (10, CFG)24.0029.12
DC+CA (15, Ave)21.7030.35
DC+CA (15, Cond)16.36021.6730.37
DC+CA (15, Uncond)21.8230.34
DC+CA (15, CFG)19.4330.48
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.322, + 0.905, + 0.352 + ], + "angle": 0, + "content": "CFG operations are performed under a low generation resolution, further reducing computational costs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.353, + 0.907, + 0.397 + ], + "angle": 0, + "content": "Based on the type of guidance to be reused, there are four design choices for caching the cross-attention maps for subsequent steps:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.398, + 0.783, + 0.417 + ], + "angle": 0, + "content": "- Ave [22]: \\(CA_{\\mathrm{cache}} = \\frac{1}{2}\\left(CA_t^c +CA_t^\\emptyset\\right)\\)" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.416, + 0.686, + 0.43 + ], + "angle": 0, + "content": "- Cond: \\(CA_{\\mathrm{cache}} = CA_{t}^{c}\\)" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.43, + 0.703, + 0.445 + ], + "angle": 0, + "content": "- Uncond: \\(CA_{\\mathrm{cache}} = CA_t^0\\)" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.444, + 0.81, + 0.462 + ], + "angle": 0, + "content": "- CFG: \\(CA_{\\mathrm{cache}} = CA_t^\\emptyset + w(CA_t^c - CA_t^\\emptyset)\\)" + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.398, + 0.81, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.463, + 0.907, + 0.524 + ], + "angle": 0, + "content": "Here, \\(CA_{t}^{c}\\) and \\(CA_{t}^{\\emptyset}\\) are the cross-attention conditioned on the input text and null text, respectively. Using either of these maps or a mixture of them results in varying degrees of conditioning in the subsequent generation process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.909, + 0.766 + ], + "angle": 0, + "content": "Exploring the best cross-attention caching strategy. We empirically examine combinations of different caching strategies, including DeepCache (DC) combined with the four design choices for cross-attention caching on top of an SD V1.5 model. As shown in Tab. 1, we observe that (1) fully abandoning CFG leads to a collapse in performance, while incorporating the cross-attention caching mechanism can better balance performance and speed; (2) adding CA on top of DC further reduces computations due to redundant CFG usage in later denoising phases. Intuitively, using a smaller \\( m \\), i.e., abandoning CFG in earlier timesteps, reduces more FLOPs at the cost of a slight decrease in generation quality; (3) among the four design choices, \"Cond\" generally performs best, slightly surpassing \"Ave\" at larger \\( m \\) values. Therefore, we adopt \"Cond\" as the \\( CA_{\\mathrm{cache}} \\) choice in all subsequent experiments." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.778, + 0.825, + 0.797 + ], + "angle": 0, + "content": "3. PostDiff: Combine All Techniques" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.803, + 0.691, + 0.82 + ], + "angle": 0, + "content": "3.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We apply our PostDiff framework, integrating mixed-resolution denoising and hybrid module caching, to several SOTA diffusion models, including SD V1.5 (DreamShaper-7 version) [31], Latent Consistency Model (LCM) [24, 36], Stable Diffusion XL (SDXL) [29], and PixArt-\\(\\alpha\\) [5]. We" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3005" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.152 + ], + "angle": 0, + "content": "evaluate text-to-image generation performance on the MS-COCO 2014 validation dataset [21]. PostDiff settings are calibrated on a small set for each model, with the detailed configurations outlined as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.155, + 0.483, + 0.2 + ], + "angle": 0, + "content": "1. SD V1.5 [31]: We use the DreamShaper-v7 fine-tuned version with a final resolution of \\( 768 \\times 768 \\). Settings: \\( \\beta = 1/2 \\), \\( s = 1/2 \\), \\( w = 7.5 \\), \\( m = 15 \\)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.2, + 0.483, + 0.291 + ], + "angle": 0, + "content": "2. Latent Consistency Model (LCM) [24, 36]: A distilled version of Stable Diffusion designed for few-step generation, producing \\(768 \\times 768\\) images. Settings: \\(s = 1/2\\), \\(\\beta = 1/2\\), \\(w = 7.5\\), \\(m = 4\\). Note that since CFG is integrated into a guidance embedding [26] and the model has a few-step nature, only \\(CA_{\\mathrm{cache}}\\) is used without DC." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.291, + 0.482, + 0.336 + ], + "angle": 0, + "content": "3. Stable Diffusion XL (SDXL) [29]: A large-scale diffusion model capable of generating \\(1024 \\times 1024\\) images. Settings: \\(\\beta = 3/4\\), \\(s = 1/5\\), \\(w = 5.0\\), \\(m = 15\\)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.336, + 0.482, + 0.398 + ], + "angle": 0, + "content": "4. PixArt-\\(\\alpha\\) [5]: A transformer-based diffusion model generating \\(1024 \\times 1024\\) images. Due to the absence of bypass connections, only \\(CA_{\\mathrm{cache}}\\) is used without enabling DC. Settings: \\(\\beta = 3/4\\), \\(s = 1/2\\), \\(w = 4.5\\), \\(m = 15\\)." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.155, + 0.483, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.407, + 0.483, + 0.438 + ], + "angle": 0, + "content": "3.2. Fewer Denoising Steps vs. Cheaper Per-step Inference" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.445, + 0.483, + 0.581 + ], + "angle": 0, + "content": "After combining the hybrid module caching with our mixed-resolution strategy to establish the complete PostDiff framework, we examine our key question by benchmarking with the FID-FLOPs/Clip Score-FLOPs trade-offs achieved by reducing denoising steps. Specifically, for SD V1.5, we use 20 denoising steps, a fixed mixed-resolution setting of \\( s = 1/2 \\) and \\( \\beta = 1/2 \\), and vary cache settings \\( m \\) to create the trade-off. As for PixArt- \\( \\alpha \\), we use a fixed 20 denoising steps with \\( \\beta = 3/4 \\), and vary \\( m \\) and \\( s \\) to form the trade-off." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.483, + 0.824 + ], + "angle": 0, + "content": "Observations and insights. From Fig. 5, we observe that (1) the complete PostDiff framework, incorporating both techniques, achieves the best FID-FLOPs trade-off across all FLOPs ranges. This demonstrates that exploiting model redundancy at both the input and module levels can yield better FID-FLOPs frontiers; (2) in terms of semantic alignment measured by Clip Score, reducing per-step inference cost can achieve a comparable Clip Score-FLOPs trade-off compared to decreasing the number of denoising steps over a wide range; while in scenarios requiring extreme efficiency, decreasing the number of denoising steps can maintain better semantic alignment with a slightly higher Clip Score. Generally, reducing per-step inference cost is more effective than decreasing the number of denoising steps with better fidelity and comparable semantic alignment across a wide efficiency range." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.422, + 0.849 + ], + "angle": 0, + "content": "3.3. Evaluation on SOTA Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We further apply PostDiff to more diffusion models. As shown in Tab. 2, we present the evaluation results in terms of both generation quality and efficiency under (1) reduced" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.089, + 0.71, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.089, + 0.898, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.239, + 0.71, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.239, + 0.898, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.399, + 0.907, + 0.441 + ], + "angle": 0, + "content": "Figure 5. Comparison of FID-FLOPs/Clip Score-FLOPs trade-offs achieved by reducing per-step inference cost or the number of denoising steps on SD V1.5 and PixArt-\\(\\alpha\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.457, + 0.905, + 0.487 + ], + "angle": 0, + "content": "denoising steps or (2) maximal denoising steps with reduced model complexity via PostDiff." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.488, + 0.906, + 0.745 + ], + "angle": 0, + "content": "Observations and analysis. We observe that: (1) With reduced denoising steps, the improved efficiency comes at the cost of degraded generation quality, reflected in notably increased FID and reduced Clip score. In contrast, maintaining the maximum number of denoising steps while applying PostDiff better preserves generation quality, achieving lower (i.e., better) FID and comparable Clip scores compared to reducing the denoising steps. For instance, on the SD V1.5 model, PostDiff achieves a \\(63.14\\%\\) reduction in FLOPs while improving FID by 1.8 compared to the original model. (2) PostDiff consistently demonstrates effectiveness across various diffusion models, including both large-scale and small-scale ones, LDMs and LCM, as well as U-Net-based and transformer-based models. This effectiveness is due to the broad applicability of the insights leveraged in Sec. 2.1 and Sec. 2.5. We also visualize the generation results in Fig. 6." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.756, + 0.905, + 0.788 + ], + "angle": 0, + "content": "3.4. Benchmark with Other Training-Free Compression Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.794, + 0.905, + 0.824 + ], + "angle": 0, + "content": "We further benchmark PostDiff with other training-free diffusion model compression methods in Tab. 3 and Tab. 4." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.9 + ], + "angle": 0, + "content": "Settings. For the U-Net-based model SD V1.5, we evaluate our method against DeepCache [25], TGATE [22], ToDo [35], ToMe [1], Faster Diffusion [20], and CA-ToMe [33]. Latency is measured as the single-image generation time on an NVIDIA A5000 GPU. The cache interval" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3006" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 2. Apply PostDiff with mixed-resolution denoising and hybrid module caching to SOTA diffusion models with varying numbers of denoising steps. Latency is measured as the single-image generation time on an NVIDIA A5000 GPU." + }, + { + "type": "table", + "bbox": [ + 0.116, + 0.128, + 0.885, + 0.468 + ], + "angle": 0, + "content": "
ModelStepsMixCacheFID ↓Clip Score ↑FLOPs (T) ↓Latency (s) ↓
SD V1.5 [31]820.6030.4112.1681.298
1617.7430.7624.3362.377
2018.4230.8030.4202.930
20✓15.6930.7819.0351.945
20✓✓16.6530.2511.1291.139
LCM [24]423.5528.643.0420.523
822.9628.726.0840.825
8✓21.0128.923.8070.666
8✓✓22.7628.583.6860.651
SDXL [29]818.0130.9247.8562.843
1614.3531.9295.7125.211
2014.1031.95119.6416.521
20✓14.3731.62109.2645.950
20✓✓14.1831.1152.6823.119
PixArt-α [5]833.0930.2134.2563.031
1629.6230.4068.5125.725
2029.1630.4185.6407.093
20✓25.9930.3866.9905.645
20✓✓25.4430.2354.7184.768
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.489, + 0.483, + 0.503 + ], + "angle": 0, + "content": "Table 3. Benchmark with other compression methods on SD V1.5." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.514, + 0.491, + 0.653 + ], + "angle": 0, + "content": "
MethodFID ↓Clip Score ↑Latency (s) ↓
Original18.4230.802.930
DeepCache [25]17.7930.751.737
TGATE [22]19.5129.551.992
ToDo [35]18.2830.752.520
ToMe [1]17.4330.552.730
Faster Diffusion [20]17.9030.402.005
CA-ToMe [33]20.4930.412.112
Ours16.6530.251.139
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.671, + 0.483, + 0.807 + ], + "angle": 0, + "content": "for DeepCache is set to 2. For TGATE, we set \\(m = 5\\). For ToMe, we use a merging ratio of 0.3, while ToDo, Faster Diffusion, and CA-ToMe are evaluated using their default hyperparameters from their respective papers. For the transformer-based model PixArt-\\(\\alpha\\), we compare Post-Diff with TGATE [22] using \\(m = 8\\), as well as DiTFastAttn [46] and DuCa [49] with their default settings on an NVIDIA H100 GPU, since some methods encounter out-of-memory errors on the A5000." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.81, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Observations. Compared to previous methods that accelerate the generation process either at the module level (e.g., DeepCache, TGATE, DuCa) by reusing computations across denoising steps or at the input level (e.g., ToMe, ToDo) by compressing redundant image tokens, our method uniquely integrates both approaches, resulting in a greater" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.489, + 0.907, + 0.516 + ], + "angle": 0, + "content": "Table 4. Benchmark with other training-free compression methods on PixArt-α." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.528, + 0.905, + 0.642 + ], + "angle": 0, + "content": "
MethodFID ↓Clip Score ↑Latency (s) ↓
Original29.1630.411.752
TGATE [22]29.5429.591.275
DiTFastAttn [46]45.4630.13\\( 1.103^{\\dagger} \\)
DuCa [49]28.0830.121.724
Ours25.4430.231.382
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.526, + 0.642, + 0.782, + 0.655 + ], + "angle": 0, + "content": "Faster speed is achieved with FlashAttention [6]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.672, + 0.907, + 0.746 + ], + "angle": 0, + "content": "reduction in inference cost. Additionally, leveraging low-resolution images in the early stages enhances the denoising process, allowing our method to achieve the lowest FID with a competitive Clip Score across both SD V1.5 and PixArt-\\(\\alpha\\) models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.749, + 0.907, + 0.824 + ], + "angle": 0, + "content": "These experiments highlight the benefits of leveraging redundancy at both the input and module levels. Integrating the aforementioned training-free compression techniques into PostDiff could further enhance efficiency, which we leave for future work." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.844, + 0.656, + 0.859 + ], + "angle": 0, + "content": "4. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Diffusion models. Diffusion models have emerged as a powerful approach for generative tasks, modeling the gen" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3007" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.089, + 0.896, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.447, + 0.772, + 0.461 + ], + "angle": 0, + "content": "Figure 6. Visualize the generation results of different models w/o and w/ using our PostDiff." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.475, + 0.483, + 0.642 + ], + "angle": 0, + "content": "erative process as a denoising sequence to enable high-quality image generation [15, 36]. Latent Diffusion Models (LDMs) [31] leverage Variational Autoencoders (VAEs) [30] to operate in a compressed latent space, reducing memory demands and enabling efficient high-resolution generation. Open-source models like Stable Diffusion V1.5 [31], Stable Diffusion XL [29], and PixArt [4, 5] have further expanded accessibility and are now applied across multiple domains [8, 10, 11]. Despite their impressive generative quality, diffusion models often suffer from long generation times, limiting their broader applicability." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.643, + 0.483, + 0.824 + ], + "angle": 0, + "content": "Few-Step diffusion models. A primary focus in accelerating diffusion models is reducing the number of timesteps. DDIM [36] modifies DDPM [15] by introducing non-Markovian sampling, enabling faster generation. DPM-Solver [23] applies exponential integrators to approximate solutions to diffusion ODEs, enhancing sampling efficiency. Other approaches, such as consistency models [24, 37], aim to map points on the ODE trajectory to the original sample, effectively reducing inference steps. Adversarial training is also employed to enable one-step or few-step diffusion [43-45]. However, these few-step models still incur non-trivial training/distillation cost." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.825, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Efficient diffusion models. Efforts to accelerate diffusion models extend from data optimization to model optimization. From the data perspective, prior works have reduced computational requirements by merging similar tokens [1] or pruning redundant tokens [39]. From the model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.475, + 0.906, + 0.598 + ], + "angle": 0, + "content": "perspective, recent works focus on identifying redundant feature maps across timesteps and using caching mechanisms to reuse these maps, significantly reducing computational costs [25, 42]. Additional approaches aim to slim down models through quantization to lower bit precision [12, 34, 48] or by pruning unnecessary parameters [2, 9, 41]. Additionally, several systems have been developed specifically to accelerate diffusion model inference [18, 19]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.617, + 0.634, + 0.633 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.644, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In this work, we explore a critical question for compute-optimal diffusion model deployment: under a post-training setting, is it more effective to reduce denoising steps or decrease per-step inference cost? To investigate this, we propose PostDiff, a training-free framework that enhances diffusion model efficiency while systematically examining this trade-off. PostDiff reduces redundancy at two levels: at the input level, through a mixed-resolution denoising scheme that enhances low-frequency components in early steps, and at the module level, with a hybrid caching strategy. Extensive evaluations demonstrate that reducing per-step inference cost often achieves a better efficiency-fidelity balance than reducing denoising steps across a wide efficiency range. PostDiff not only offers a new perspective on the fidelity-efficiency trade-off for diffusion models but also provides practical insights for their deployment on resource-constrained platforms." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3008" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.222 + ], + "angle": 0, + "content": "This work is supported by the National Science Foundation (NSF) through CCRI funding (Award Number: 2016727), RTML funding (Award Number: 1937592), and an NSF CAREER award (Award Number: 2048183), as well as by CoCoSys, one of the seven centers in JUMP 2.0, a Semiconductor Research Corporation (SRC) program sponsored by DARPA." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.235, + 0.188, + 0.251 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.26, + 0.483, + 0.288 + ], + "angle": 0, + "content": "[1] Daniel Bolya and Judy Hoffman. Token merging for fast stable diffusion. In CVPR, 2023. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.29, + 0.484, + 0.344 + ], + "angle": 0, + "content": "[2] Thibault Castells, Hyoung-Kyu Song, Bo-Kyeong Kim, and Shinkook Choi. LD-Pruner: Efficient pruning of latent diffusion models using task-agnostic insights. In CVPR, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.346, + 0.483, + 0.401 + ], + "angle": 0, + "content": "[3] Angela Castillo, Jonas Kohler, Juan C Pérez, Juan Pablo Pérez, Albert Pumarola, Bernard Ghanem, Pablo Arbeláez, and Ali Thabet. Adaptive guidance: Training-free acceleration of conditional diffusion models. In AAAI, 2025. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.403, + 0.484, + 0.471 + ], + "angle": 0, + "content": "[4] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. PixArt- \\(\\Sigma\\) : Weak-to-strong training of diffusion transformer for 4k text-to-image generation. In ECCV, 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.473, + 0.484, + 0.543 + ], + "angle": 0, + "content": "[5] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Zhongdao Wang, James T. Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. PixArt-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In ICLR, 2024. 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.545, + 0.483, + 0.585 + ], + "angle": 0, + "content": "[6] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. FlashAttention: Fast and memory-efficient exact attention with io-awareness. In NeurIPS, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.587, + 0.483, + 0.614 + ], + "angle": 0, + "content": "[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.616, + 0.484, + 0.671 + ], + "angle": 0, + "content": "[8] Zhenbang Du, Wei Feng, Haohan Wang, Yaoyu Li, Jingsen Wang, Jian Li, Zheng Zhang, Jingjing Lv, Xin Zhu, Junsheng Jin, et al. Towards reliable advertising image generation using human feedback. In ECCV, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.673, + 0.483, + 0.701 + ], + "angle": 0, + "content": "[9] Gongfan Fang, Xinyin Ma, and Xinchao Wang. Structural pruning for diffusion models. In NeurIPS, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.702, + 0.483, + 0.757 + ], + "angle": 0, + "content": "[10] Sicheng Gao, Xuhui Liu, Bohan Zeng, Sheng Xu, Yanjing Li, Xiaoyan Luo, Jianzhuang Liu, Xiantong Zhen, and Baochang Zhang. Implicit diffusion models for continuous super-resolution. In CVPR, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.758, + 0.483, + 0.786 + ], + "angle": 0, + "content": "[11] He Guo, Zixuan Ye, Zhiguo Cao, and Hao Lu. In-context matting. In CVPR, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.788, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[12] Yefei He, Luping Liu, Jing Liu, Weijia Wu, Hong Zhou, and Bohan Zhuang. PTQD: Accurate post-training quantization for diffusion models. In NeurIPS, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.831, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[13] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. In EMNLP, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[14] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshop, 2021. 5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.26, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.163 + ], + "angle": 0, + "content": "[16] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. JMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.206 + ], + "angle": 0, + "content": "[17] Bo-Kyeong Kim, Hyoung-Kyu Song, Thibault Castells, and Shinkook Choi. BK-SDM: A lightweight, fast, and cheap version of stable diffusion. In ECCV, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.208, + 0.905, + 0.277 + ], + "angle": 0, + "content": "[18] Akio Kodaira, Chenfeng Xu, Toshiki Hazama, Takanori Yoshimoto, Kohei Ohno, Shogo Mitsuhori, Soichi Sugano, Hanying Cho, Zhijian Liu, and Kurt Keutzer. Streamdiffusion: A pipeline-level solution for real-time interactive generation. arXiv preprint, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.279, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[19] Muyang Li, Tianle Cai, Jiaxin Cao, Qinsheng Zhang, Han Cai, Junjie Bai, Yangqing Jia, Kai Li, and Song Han. Distribution: Distributed parallel inference for high-resolution diffusion models. In CVPR, 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.39 + ], + "angle": 0, + "content": "[20] Senmao Li, Taihang Hu, Fahad Shahbaz Khan, Linxuan Li, Shiqi Yang, Yaxing Wang, Ming-Ming Cheng, and Jian Yang. Faster diffusion: Rethinking the role of unet encoder in diffusion models. In NeurIPS, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[22] Haozhe Liu, Wentian Zhang, Jinheng Xie, Francesco Faccio, Mengmeng Xu, Tao Xiang, Mike Zheng Shou, Juan-Manuel Pérez-Rúa, and Jurgen Schmidhuber. Faster diffusion through temporal attention decomposition. TMLR, 2025. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[23] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. NeurIPS, 2022. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[24] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint, 2023. 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[25] Xinyin Ma, Gongfan Fang, and Xinchao Wang. Deepcache: Accelerating diffusion models for free. In CVPR, 2024. 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.676, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[26] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. In CVPR, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.733, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[27] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.802, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[28] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[29] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024. 2, 5, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3009" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.133 + ], + "angle": 0, + "content": "[30] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. In NeurIPS, 2019. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.194, + 0.482, + 0.234 + ], + "angle": 0, + "content": "[32] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.238, + 0.482, + 0.293 + ], + "angle": 0, + "content": "[33] Omid Saghatchian, Atiyeh Gh Moghadam, and Ahmad Nickabadi. Cached adaptive token merging: Dynamic token reduction and redundant computation elimination in diffusion model. arXiv preprint, 2025. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.296, + 0.482, + 0.336 + ], + "angle": 0, + "content": "[34] Yuzhang Shang, Zhihang Yuan, Bin Xie, Bingzhe Wu, and Yan Yan. Post-training quantization on diffusion models. In CVPR, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.339, + 0.482, + 0.38 + ], + "angle": 0, + "content": "[35] Ethan Smith, Nayan Saxena, and Aninda Saha. Todo: Token downsampling for efficient generation of high-resolution images. In *IJCAI*, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.383, + 0.482, + 0.41 + ], + "angle": 0, + "content": "[36] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.482, + 0.44 + ], + "angle": 0, + "content": "[37] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In ICML, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.443, + 0.482, + 0.497 + ], + "angle": 0, + "content": "[38] Jiayan Teng, Wendi Zheng, Ming Ding, Wenyi Hong, Jianqiao Wangni, Zhuoyi Yang, and Jie Tang. Relay diffusion: Unifying diffusion process across resolutions for image synthesis. In ICLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.5, + 0.482, + 0.554 + ], + "angle": 0, + "content": "[39] Hongjie Wang, Difan Liu, Yan Kang, Yijun Li, Zhe Lin, Niraj K Jha, and Yuchen Liu. Attention-driven training-free efficiency enhancement of diffusion models. In CVPR, 2024. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.557, + 0.482, + 0.612 + ], + "angle": 0, + "content": "[40] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. LAVIE: High-quality video generation with cascaded latent diffusion models. IJCV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.615, + 0.482, + 0.67 + ], + "angle": 0, + "content": "[41] Lexington Whalen, Zhenbang Du, Haoran You, Chaojian Li, Sixu Li, and Yingyan (Celine) Lin. Early-Bird Diffusion: Investigating and leveraging timestep-aware early-bird tickets in diffusion models for efficient training. In CVPR, 2025. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.673, + 0.482, + 0.741 + ], + "angle": 0, + "content": "[42] Felix Wimbauer, Bichen Wu, Edgar Schoenfeld, Xiaoliang Dai, Ji Hou, Zijian He, Artsiom Sanakoyeu, Peizhao Zhang, Sam Tsai, Jonas Kohler, et al. Cache me if you can: Accelerating diffusion models through block caching. In CVPR, 2024. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.744, + 0.482, + 0.785 + ], + "angle": 0, + "content": "[43] Yanwu Xu, Yang Zhao, Zhisheng Xiao, and Tingbo Hou. UFOGen: You forward once large scale text-to-image generation via diffusion gans. In CVPR, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.482, + 0.842 + ], + "angle": 0, + "content": "[44] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved distribution matching distillation for fast image synthesis. In NeurIPS, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[45] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Frédo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In CVPR, 2024. 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[46] Zhihang Yuan, Hanling Zhang, Lu Pu, Xuefei Ning, Linfeng Zhang, Tianchen Zhao, Shengen Yan, Guohao Dai, and Yu Wang. DiTFastattn: Attention compression for diffusion transformer models. In NeurIPS, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.149, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[47] Shen Zhang, Zhaowei Chen, Zhenyu Zhao, Yuhao Chen, Yao Tang, and Jiajun Liang. HiDiffusion: Unlocking higher-resolution creativity and efficiency in pretrained diffusion models. In ECCV, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.206, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[48] Tianchen Zhao, Tongcheng Fang, Enshu Liu, Rui Wan, Widyadewi Soedarmadji, Shiyao Li, Zinan Lin, Guohao Dai, Shengen Yan, Huazhong Yang, Xuefei Ning, and Yu Wang. ViDiT-Q: Efficient and accurate quantization of diffusion transformers for image and video generation. In ICLR, 2025. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.29, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[49] Chang Zou, Evelyn Zhang, Runlin Guo, Haohang Xu, Conghui He, Xuming Hu, and Linfeng Zhang. Accelerating diffusion transformers with dual feature caching. arXiv preprint, 2024. 7" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3010" + } + ] +] \ No newline at end of file diff --git a/2025/Fewer Denoising Steps or Cheaper Per-Step Inference_ Towards Compute-Optimal Diffusion Model Deployment/98cafbf9-4954-46a4-b5af-59f0bca621f7_origin.pdf b/2025/Fewer Denoising Steps or Cheaper Per-Step Inference_ Towards Compute-Optimal Diffusion Model Deployment/98cafbf9-4954-46a4-b5af-59f0bca621f7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4d8c0b955237fd68fdded8c89284b51853c50776 --- /dev/null +++ b/2025/Fewer Denoising Steps or Cheaper Per-Step Inference_ Towards Compute-Optimal Diffusion Model Deployment/98cafbf9-4954-46a4-b5af-59f0bca621f7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3f146cff73f4638dd6f8c883b6893ce77b7e23756f863d2a468c118f1ea9a640 +size 13205658 diff --git a/2025/Fewer Denoising Steps or Cheaper Per-Step Inference_ Towards Compute-Optimal Diffusion Model Deployment/full.md b/2025/Fewer Denoising Steps or Cheaper Per-Step Inference_ Towards Compute-Optimal Diffusion Model Deployment/full.md new file mode 100644 index 0000000000000000000000000000000000000000..db6a45e29a39416de74e74ccd21d3ac9e0558710 --- /dev/null +++ b/2025/Fewer Denoising Steps or Cheaper Per-Step Inference_ Towards Compute-Optimal Diffusion Model Deployment/full.md @@ -0,0 +1,295 @@ +# Fewer Denoising Steps or Cheaper Per-Step Inference: Towards Compute-Optimal Diffusion Model Deployment + +Zhenbang Du*, Yonggan Fu*, Lifu Wang*, Jiayi Qian, Xiao Luo, Yingyan (Celine) Lin +Georgia Institute of Technology +{zdu89,jiayiqian,celine.lin}@gatech.edu + +# Abstract + +Diffusion models have shown remarkable success across generative tasks, yet their high computational demands challenge deployment on resource-limited platforms. This paper investigates a critical question for compute-optimal diffusion model deployment: Under a post-training setting without fine-tuning, is it more effective to reduce the number of denoising steps or to use a cheaper per-step inference? Intuitively, reducing the number of denoising steps increases the variability of the distributions across steps, making the model more sensitive to compression. In contrast, keeping more denoising steps makes the differences smaller, preserving redundancy, and making post-training compression more feasible. To systematically examine this, we propose PostDiff, a training-free framework for accelerating pre-trained diffusion models by reducing redundancy at both the input level and module level in a post-training manner. At the input level, we propose a mixed-resolution denoising scheme based on the insight that reducing generation resolution in early denoising steps can enhance low-frequency components and improve final generation fidelity. At the module level, we employ a hybrid module caching strategy to reuse computations across denoising steps. Extensive experiments and ablation studies demonstrate that (1) PostDiff can significantly improve the fidelity-efficiency trade-off of state-of-the-art diffusion models, and (2) to boost efficiency while maintaining decent generation fidelity, reducing per-step inference cost is often more effective than reducing the number of denoising steps. Our code is available at https://github.com/GATECH-EIC/PostDiff. + +# 1. Introduction + +Diffusion models have rapidly advanced as a powerful tool for various generative tasks, such as image synthesis and video generation. Their success stems from the iterative de + +noising process, which gradually refines outputs to achieve high-quality results, along with advanced model architectures. However, both the iterative nature and model complexity of diffusion models make them computationally intensive, posing a significant challenge for deployment on resource-constrained platforms, where computational resources, memory, and power are limited. Efficient deployment of diffusion models on these platforms is therefore highly desirable to broaden the accessibility of generative AI, enabling real-time applications that can operate within the strict constraints of on-device resources. + +With the increasing demand for rapid diffusion model deployment, where a given diffusion model is expected to be instantly compressed and deployed without fine-tuning, a critical question arises: Under a post-training setting, is it more effective to optimize for fewer denoising steps or to reduce the per-step inference cost? This question is particularly important because each approach entails different trade-offs: Reducing the number of denoising steps introduces greater variability in feature changes between steps, which can amplify the negative impact of compression on output fidelity; Conversely, using more denoising steps reduces these step-to-step differences, preserving redundancy and making it easier to apply post-training compression without sacrificing quality. However, there is a limited exploration of how these two strategies perform in terms of fidelity-efficiency trade-offs, and investigating this could provide valuable guidelines for efficient diffusion model deployment in real-world applications. + +To address this research question, we propose PostDiff, a training-free diffusion model compression framework designed to enhance the per-step inference efficiency of pretrained diffusion models at both the input level and module level in a post-training manner. Our techniques are training-free and can thus be instantly applied to emerging diffusion models for rapid deployment, enabling a comparative study of their effectiveness versus reducing denoising steps. Specifically, at the input level, we propose a mixed-resolution denoising scheme that adopts a lower generation + +resolution in early denoising steps. As low-frequency components dominate at this stage, a reduced resolution is sufficient. We find that this reduced resolution in early denoising steps can enhance low-frequency components, thereby improving final generation fidelity and achieving a win-win in both efficiency and fidelity. At the module level, we build on existing module caching techniques, which reuse module outputs across steps, and enhance them with a hybrid caching strategy as another dimension for redundancy reduction. Our contributions are summarized as follows: + +- Our work, for the first time, studies a critical question for efficient diffusion model deployment: Under a posttraining setting without fine-tuning, is it more effective to optimize for fewer denoising steps or to reduce the per-step inference cost? The answer to this question could be useful for both researchers and practitioners. +- To answer the above question, we propose PostDiff, a training-free framework for reducing diffusion model complexity at both the input and module levels, resulting in lower per-step inference cost. This framework enables a fair study of the relative effectiveness between fewer denoising steps and reduced per-step inference cost toward compute-optimal diffusion model deployment under a post-training setting. +- At the input level, we propose a mixed-resolution denoising scheme, which leverages the observation that reduced generation resolution in early denoising steps can enhance low-frequency components and final output quality, achieving a win-win in both accuracy and efficiency. In addition, to further reduce the overall computational cost at the module level, PostDiff employs a hybrid module caching method that strategically reuses the outputs of different modules on varying schedules across steps. +- Extensive evaluations show that PostDiff can advance the fidelity-efficiency frontier of state-of-the-art (SOTA) diffusion models, as shown in Fig. 1, and provide valuable insights into best practices for efficient diffusion model deployment. For example, we find that reducing per-step inference cost is often more effective than decreasing the number of denoising steps to improve efficiency when aiming to preserve high generation fidelity. + +In the following sections, we illustrate and analyze the effectiveness of each technique in PostDiff in Sec. 2, and finally, we combine and apply them to SOTA diffusion models in Sec. 3 to answer our research question. + +# 2. PostDiff: The Proposed Framework + +In this section, we first introduce the proposed mixed-resolution denoising strategy, which reduces per-step inference cost at the input level. We begin by analyzing the motivations from previous works in Sec. 2.1, followed by an illustration of our method, an analysis of its effectiveness, and its application to examine our key question in Sec. 2.2, + +![](images/96dcd55c2125876b6f232a70876722be8f7e2ed7d1dd6c05b3e0b978a4a821cd.jpg) +Figure 1. Apply our PostDiff technique to SOTA diffusion models, where “-PD” denotes the models delivered by PostDiff. + +Sec. 2.3, and Sec. 2.4, respectively. Finally, we describe how we employ and enhance module caching to improve efficiency at the module level in Sec. 2.5. + +# 2.1. Motivations from Previous Works + +Low-frequency components dominate the early denoising steps. Previous works [22] find that the typical denoising process for image generation progresses through two phases: the semantics-planning phase and the fidelity-improving phase. The early semantics-planning phase focuses more on the layout and low-frequency information, starting from random noise and gradually establishing the basic structure. This indicates that a low generation resolution could be sufficient in the early denoising steps. As generation proceeds, more details, described by high-frequency information, are incrementally added, thus requiring a higher resolution. + +This insight is also partially echoed and leveraged by previous works. For example, [38] trains different diffusion models for early low-resolution denoising and later high-resolution denoising processes to avoid the propagation of inaccurate high-frequency components from early denoising steps to later steps. Additionally, cascade diffusion models [16, 40] sequentially apply different diffusion models at different denoising stages with varying generation resolutions, finding that incorporating low-frequency information from low-resolution images during high-resolution stages can enhance generation quality. Nevertheless, all these methods require multiple diffusion models, introducing significant storage and training demands. + +Input resolution is a promising dimension for reducing inference cost. High-resolution generation is a recent trend, as it provides images with more vivid details and accommodates greater information density [47]. For example, SDXL [29] generates images at 1024x1024 resolution, while PixArt- $\Sigma$ [4] expands resolution capability to 4K. However, as resolution increases, computational demands rise dramatically, leading to significant latency [19]. As + +![](images/136fc13924f38ecd291bd208ec099e0afa2ea63585db3dbe4ab13f47333c3b8b.jpg) +Figure 2. An overview of our mixed-resolution denoising strategy. + +such, recent works aim to reduce the computational load by decreasing the number of tokens in samples, using methods like token merging [1], token pruning [39], or token down-sampling [35], which effectively accelerate inference. + +# 2.2. Mixed-Resolution Denoising: Methodology + +Motivated by the tendency for low-frequency components in early denoising steps and the effectiveness of input resolution as a design knob for efficiency, we propose a simple yet effective post-training strategy: As shown in Fig. 2, we adopt a low generation resolution in the early denoising steps to generate low-frequency structures, and then refine high-frequency details with high generation resolution in the later denoising steps. + +Formulation. To generate an image with shape $(W,H,3)$ , we begin with Gaussian noise in the shape $(w,h,4)$ , representing the latent noise image $x_{T}$ obtained via Variational Autoencoders [30]. For simplicity, we omit the channel dimension in the remainder of this paper. The denoising process proceeds over $T$ steps, using samplers like DDIM [36] or DPM-solver [23] to obtain the denoised latent at each step. Taking DDIM [36] as an example, we compute the latent representation $x$ at step $t - 1$ as follows: + +$$ +x _ {t - 1} = \sqrt {\alpha_ {t - 1}} \hat {x} _ {0} ^ {t} + x _ {t}, \tag {1} +$$ + +$$ +\hat {x} _ {0} ^ {t} = \frac {x _ {t} - \sqrt {1 - \alpha_ {t}} \cdot \epsilon_ {\theta} (x _ {t} , t)}{\sqrt {\alpha_ {t}}}, \tag {2} +$$ + +$$ +x _ {t} = \sqrt {1 - \alpha_ {t - 1}} \cdot \epsilon_ {\theta} (x _ {t}, t), \tag {3} +$$ + +where $\epsilon_{\theta}$ denotes the model [28, 32] that predicts the added noise, and $\{\alpha_t\}$ is a set of coefficients controlling the forward noise-adding process [15]. + +In our proposed mixed-resolution denoising strategy, we first introduce a lower-resolution latent representation, $x_{T}^{l}$ , with shape $(\beta w, \beta h)$ , where $0 < \beta < 1$ is a hyperparameter that scales the resolution down. This low-resolution starting point reduces the computational load of the initial denoising phase. Another hyperparameter, $s$ , controls the transition from low to full resolution. After reaching the denoising step $t = sT$ , we compute the forecasted + +$\hat{x}_0^{l,t}$ at step $t$ using Eq. (2). Then, we upscale this result to full resolution via: + +$$ +\hat {x} _ {0} ^ {f, t} = \operatorname {U p s a m p l e} \left(\hat {x} _ {0} ^ {l, t}\right), \tag {4} +$$ + +where Upsample refers to bilinear interpolation. Using this upscaled version, we derive the full-resolution latent $x_{t}^{f}$ at step $t$ as follows: + +$$ +x _ {t} ^ {f} = \sqrt {\alpha_ {t}} \hat {x} _ {0} ^ {f, t} + \sqrt {1 - \alpha_ {t}} \epsilon , \tag {5} +$$ + +where $\epsilon$ is Gaussian noise with shape $(w,h)$ . The remainder of the denoising process follows the original procedure. + +Note that we adopt this simple binary resolution strategy without bells and whistles to effectively demonstrate its concept and simplify the involved hyperparameters, thereby easing the use of our method in real-world deployment. More complex resolution schedules could potentially be proposed to further enhance fidelity-efficiency trade-offs. + +# 2.3. Visualization: How Low-Resolution Denoising Contributes to Boosted Final Performance + +To study the impact of our mixed-resolution denoising strategy on final performance and understand why it works, we visualize the generation quality after each denoising step, both with and without our method. Specifically, we calculate the Clip Score [13] for $\hat{x}_0$ predicted at each denoising step and annotate it in the case study of Fig. 3, where different rows correspond to the original full-resolution denoising process as well as our method with various choices of $s$ . + +Observations and analysis. We observe that (1) with a moderate amount of low-resolution generation steps, the Clip Score of the final generated image can be improved, achieving a win-win in generation fidelity and efficiency; (2) Regarding the evolution of the Clip Score across denoising steps, starting with a lower generation resolution results in lower Clip Scores initially, but serves as a better starting point: after switching to high generation resolution, the Clip Score increases more rapidly and gradually surpasses that of full-resolution generation. This echoes the analysis in Sec. 2.1, where reduced generation resolution in early denoising steps contributes to improved low-frequency components, thereby enhancing final generation quality; (3) The balance between low-resolution and high-resolution steps is crucial, as both low-frequency structure and high-frequency details need to be preserved to ensure high final generation quality in the last denoising step. + +This phenomenon demonstrates the underlying effectiveness of our mixed-resolution denoising strategy and is consistently observed across samples, as shown in the comprehensive evaluation in Sec. 2.4. + +# 2.4. Mixed-Resolution Denoising: Evaluation + +We evaluate the effectiveness of our mixed-resolution denoising strategy and examine which is more effective: re + +cute girl, crop-top, blond hair, black glasses, stretching, with background by greg rutkowski makoto shinkai kyoto animation key art feminine mid shot + +![](images/e928864cca10772719024a02063b86d8e738c2c6c3451fbd49313102fcc912ea.jpg) +Figure 3. Visualize the Clip Score after each denoising step using full resolution (row 1) and mixed resolution (rows 2-5). The steps using low generation resolution are highlighted in green and those using high generation resolution are highlighted in orange. + +![](images/c35d0dfc1fecf32689f4de3fa8455c0fbac17c375acf1287ab60e6d0e8bb4c3e.jpg) +Figure 4. The FID-FLOPs trade-off achieved under different denoising steps and varied mixed-resolution settings. The data points with the same number of steps are annotated using the same shape. + +duced input-level redundancy or fewer denoising steps. + +Setup. We use Stable Diffusion (SD) V1.5 (DreamShaper-7 version) [31], a popular latent diffusion model (LDM), as our model backbone and evaluate text-to-image generation performance on the MS-COCO 2014 validation dataset [21]. We vary the total denoising steps $T = \{8,10,12,14,16,18,20\}$ and apply our mixed-resolution denoising to each denoising step setting with varied $s = \{0.1,0.2,0.3,0.4,0.5,0.6\}$ . Without bells and + +whistles, we set $\beta = 0.5$ as our low-resolution setting, i.e., half the full resolution, and ablation studies on this are provided in the supplementary material. + +Observations and analysis. As shown in Fig. 4, we observe that (1) mixed-resolution denoising can boost achievable generation fidelity by notably reducing the optimal FID (lower is better) while simultaneously improving generation efficiency, aligning with the analysis of the role of low resolution in early denoising steps in Sec. 2.1 and Sec. 2.3; (2) With more inference steps $(T \geq 14)$ , the optimal values of $s$ are around $0.4 \sim 0.5$ , offering the best FID-FLOPs trade-off. For example, adopting $s = 0.5$ with 20 denoising steps can save about $37\%$ FLOPs while achieving an FID reduction of approximately 2 compared to the original full-resolution model. (3) When aiming for relatively high generation fidelity, e.g., a $< 20$ FID, which aligns with real-world application demands, maintaining a high number of denoising steps and reducing per-step inference cost via mixed resolution is a more cost-effective strategy than reducing denoising steps. Conversely, when targeting extreme efficiency, e.g., reducing $>60\%$ FLOPs, using fewer denoising steps becomes increasingly beneficial, as the effectiveness of mixed-resolution denoising diminishes in low-step generation with larger gaps between steps. + +Calibration for best hyperparameters. In practice, to rapidly determine the best transition step $s$ (as well as the total number of denoising steps) for a given diffusion model provided by the users, we can use a small calibration set to obtain the optimal setting. We show in the supplemen + +tary materials that performance on a small calibration set is highly correlated with that on the full evaluation set. + +# 2.5. The Hybrid Module Caching Strategy + +In addition to the resolution dimension, we further reduce model complexity at the module level by integrating existing module caching techniques [25, 42] with an enhanced hybrid caching strategy. + +Motivation. Previous works [25, 42] have found that feature maps remain highly similar across all blocks throughout the entire denoising process, indicating the potential effectiveness of caching and reusing across steps. We have further validated this through cross-step feature map distance visualization in the supplementary materials. Considering that (1) the text guidance provided by cross-attention layers primarily determines the image layout, i.e., the low-frequency structure of the image, and (2) the semantics-planning phase for layout generation occurs in the early denoising stages [22], as also discussed in Sec. 2.1, it is feasible to use the cached cross-attention feature map from early denoising steps, generated at a low resolution with our mixed-resolution denoising, for the later denoising phase, as the low-frequency information is determined early on. + +Our caching strategy. We employ a hybrid module caching strategy to maximize reuse opportunities in a pretrained diffusion model. To achieve this, we integrate the following two caching mechanisms: (1) similar to DeepCache [25], we cache the deep skip branches of the U-Net, a major component of a wide range of SOTA diffusion models [17, 24, 29, 31], and reuse them for the subsequent $k$ steps. Here, $k$ controls how often the cache is updated, balancing efficiency gains and potential degradation, and is set to 2 throughout the evaluation; (2) We cache the crossattention layers at a specific denoising step and reuse them in all subsequent steps [22]. + +Design choices for caching cross-attention. A major role of cross-attention is to introduce text guidance into the generation process. This is often achieved through classifier-free guidance (CFG), which provides a controllable method for incorporating conditional guidance into image generation [7, 14, 27]. Specifically, the text guidance from the model's implicit classifier is derived as follows: + +$$ +\epsilon_ {\theta} \left(x _ {t}, t, w, c\right) = \epsilon_ {\theta} \left(x _ {t}, t, \emptyset\right) + w \left(\epsilon_ {\theta} \left(x _ {t}, t, c\right) - \epsilon_ {\theta} \left(x _ {t}, t, \emptyset\right)\right), \tag {6} +$$ + +where $\varnothing$ represents the null text condition. Considering that CFG may be redundant in the later stages of inference, as the layout has already been organized in the early stages [3, 22], we abandon CFG after a certain step $m$ . When combined with our mixed-resolution denoising scheme, most + +Table 1. Ablation study on different caching strategies ( $m$ , Design Choice). + +
SettingFLOPs (T) ↓FID ↓Clip Score ↑
Original30.42018.4230.80
Original w/o CFG15.21031.9825.52
DC17.78717.7930.75
DC+CA (5, Ave)18.7728.40
DC+CA (5, Cond)11.61018.8229.20
DC+CA (5, Uncond)20.3325.50
DC+CA (5, CFG)103.7118.22
DC+CA (10, Ave)21.1330.03
DC+CA (10, Cond)15.06121.2630.11
DC+CA (10, Uncond)21.1929.83
DC+CA (10, CFG)24.0029.12
DC+CA (15, Ave)21.7030.35
DC+CA (15, Cond)16.36021.6730.37
DC+CA (15, Uncond)21.8230.34
DC+CA (15, CFG)19.4330.48
+ +CFG operations are performed under a low generation resolution, further reducing computational costs. + +Based on the type of guidance to be reused, there are four design choices for caching the cross-attention maps for subsequent steps: + +- Ave [22]: $CA_{\mathrm{cache}} = \frac{1}{2}\left(CA_t^c +CA_t^\emptyset\right)$ +- Cond: $CA_{\mathrm{cache}} = CA_{t}^{c}$ +- Uncond: $CA_{\mathrm{cache}} = CA_t^0$ +- CFG: $CA_{\mathrm{cache}} = CA_t^\emptyset + w(CA_t^c - CA_t^\emptyset)$ + +Here, $CA_{t}^{c}$ and $CA_{t}^{\emptyset}$ are the cross-attention conditioned on the input text and null text, respectively. Using either of these maps or a mixture of them results in varying degrees of conditioning in the subsequent generation process. + +Exploring the best cross-attention caching strategy. We empirically examine combinations of different caching strategies, including DeepCache (DC) combined with the four design choices for cross-attention caching on top of an SD V1.5 model. As shown in Tab. 1, we observe that (1) fully abandoning CFG leads to a collapse in performance, while incorporating the cross-attention caching mechanism can better balance performance and speed; (2) adding CA on top of DC further reduces computations due to redundant CFG usage in later denoising phases. Intuitively, using a smaller $m$ , i.e., abandoning CFG in earlier timesteps, reduces more FLOPs at the cost of a slight decrease in generation quality; (3) among the four design choices, "Cond" generally performs best, slightly surpassing "Ave" at larger $m$ values. Therefore, we adopt "Cond" as the $CA_{\mathrm{cache}}$ choice in all subsequent experiments. + +# 3. PostDiff: Combine All Techniques + +# 3.1. Experiment Setup + +We apply our PostDiff framework, integrating mixed-resolution denoising and hybrid module caching, to several SOTA diffusion models, including SD V1.5 (DreamShaper-7 version) [31], Latent Consistency Model (LCM) [24, 36], Stable Diffusion XL (SDXL) [29], and PixArt- $\alpha$ [5]. We + +evaluate text-to-image generation performance on the MS-COCO 2014 validation dataset [21]. PostDiff settings are calibrated on a small set for each model, with the detailed configurations outlined as follows: + +1. SD V1.5 [31]: We use the DreamShaper-v7 fine-tuned version with a final resolution of $768 \times 768$ . Settings: $\beta = 1/2$ , $s = 1/2$ , $w = 7.5$ , $m = 15$ . +2. Latent Consistency Model (LCM) [24, 36]: A distilled version of Stable Diffusion designed for few-step generation, producing $768 \times 768$ images. Settings: $s = 1/2$ , $\beta = 1/2$ , $w = 7.5$ , $m = 4$ . Note that since CFG is integrated into a guidance embedding [26] and the model has a few-step nature, only $CA_{\mathrm{cache}}$ is used without DC. +3. Stable Diffusion XL (SDXL) [29]: A large-scale diffusion model capable of generating $1024 \times 1024$ images. Settings: $\beta = 3/4$ , $s = 1/5$ , $w = 5.0$ , $m = 15$ . +4. PixArt- $\alpha$ [5]: A transformer-based diffusion model generating $1024 \times 1024$ images. Due to the absence of bypass connections, only $CA_{\mathrm{cache}}$ is used without enabling DC. Settings: $\beta = 3/4$ , $s = 1/2$ , $w = 4.5$ , $m = 15$ . + +# 3.2. Fewer Denoising Steps vs. Cheaper Per-step Inference + +After combining the hybrid module caching with our mixed-resolution strategy to establish the complete PostDiff framework, we examine our key question by benchmarking with the FID-FLOPs/Clip Score-FLOPs trade-offs achieved by reducing denoising steps. Specifically, for SD V1.5, we use 20 denoising steps, a fixed mixed-resolution setting of $s = 1/2$ and $\beta = 1/2$ , and vary cache settings $m$ to create the trade-off. As for PixArt- $\alpha$ , we use a fixed 20 denoising steps with $\beta = 3/4$ , and vary $m$ and $s$ to form the trade-off. + +Observations and insights. From Fig. 5, we observe that (1) the complete PostDiff framework, incorporating both techniques, achieves the best FID-FLOPs trade-off across all FLOPs ranges. This demonstrates that exploiting model redundancy at both the input and module levels can yield better FID-FLOPs frontiers; (2) in terms of semantic alignment measured by Clip Score, reducing per-step inference cost can achieve a comparable Clip Score-FLOPs trade-off compared to decreasing the number of denoising steps over a wide range; while in scenarios requiring extreme efficiency, decreasing the number of denoising steps can maintain better semantic alignment with a slightly higher Clip Score. Generally, reducing per-step inference cost is more effective than decreasing the number of denoising steps with better fidelity and comparable semantic alignment across a wide efficiency range. + +# 3.3. Evaluation on SOTA Diffusion Models + +We further apply PostDiff to more diffusion models. As shown in Tab. 2, we present the evaluation results in terms of both generation quality and efficiency under (1) reduced + +![](images/c5e652a211aa5b85391ba0656d5af52224fea297dab7a5dc223b6c232d3cc68f.jpg) + +![](images/8e9bfa759e3976a8c6774708f5fae89e4a918fc30b330f856c87fc4ad13b0392.jpg) + +![](images/aa2e3c0349cd93b32b763868b52233031b40e3b2ec630b2fa58a7a947cfb17fb.jpg) +Figure 5. Comparison of FID-FLOPs/Clip Score-FLOPs trade-offs achieved by reducing per-step inference cost or the number of denoising steps on SD V1.5 and PixArt- $\alpha$ . + +![](images/7d3291122459d82ba4fb5a45ec4a2905e319531fb2378879f59adf4f1f4e30d1.jpg) + +denoising steps or (2) maximal denoising steps with reduced model complexity via PostDiff. + +Observations and analysis. We observe that: (1) With reduced denoising steps, the improved efficiency comes at the cost of degraded generation quality, reflected in notably increased FID and reduced Clip score. In contrast, maintaining the maximum number of denoising steps while applying PostDiff better preserves generation quality, achieving lower (i.e., better) FID and comparable Clip scores compared to reducing the denoising steps. For instance, on the SD V1.5 model, PostDiff achieves a $63.14\%$ reduction in FLOPs while improving FID by 1.8 compared to the original model. (2) PostDiff consistently demonstrates effectiveness across various diffusion models, including both large-scale and small-scale ones, LDMs and LCM, as well as U-Net-based and transformer-based models. This effectiveness is due to the broad applicability of the insights leveraged in Sec. 2.1 and Sec. 2.5. We also visualize the generation results in Fig. 6. + +# 3.4. Benchmark with Other Training-Free Compression Methods + +We further benchmark PostDiff with other training-free diffusion model compression methods in Tab. 3 and Tab. 4. + +Settings. For the U-Net-based model SD V1.5, we evaluate our method against DeepCache [25], TGATE [22], ToDo [35], ToMe [1], Faster Diffusion [20], and CA-ToMe [33]. Latency is measured as the single-image generation time on an NVIDIA A5000 GPU. The cache interval + +Table 2. Apply PostDiff with mixed-resolution denoising and hybrid module caching to SOTA diffusion models with varying numbers of denoising steps. Latency is measured as the single-image generation time on an NVIDIA A5000 GPU. + +
ModelStepsMixCacheFID ↓Clip Score ↑FLOPs (T) ↓Latency (s) ↓
SD V1.5 [31]820.6030.4112.1681.298
1617.7430.7624.3362.377
2018.4230.8030.4202.930
20✓15.6930.7819.0351.945
20✓✓16.6530.2511.1291.139
LCM [24]423.5528.643.0420.523
822.9628.726.0840.825
8✓21.0128.923.8070.666
8✓✓22.7628.583.6860.651
SDXL [29]818.0130.9247.8562.843
1614.3531.9295.7125.211
2014.1031.95119.6416.521
20✓14.3731.62109.2645.950
20✓✓14.1831.1152.6823.119
PixArt-α [5]833.0930.2134.2563.031
1629.6230.4068.5125.725
2029.1630.4185.6407.093
20✓25.9930.3866.9905.645
20✓✓25.4430.2354.7184.768
+ +Table 3. Benchmark with other compression methods on SD V1.5. + +
MethodFID ↓Clip Score ↑Latency (s) ↓
Original18.4230.802.930
DeepCache [25]17.7930.751.737
TGATE [22]19.5129.551.992
ToDo [35]18.2830.752.520
ToMe [1]17.4330.552.730
Faster Diffusion [20]17.9030.402.005
CA-ToMe [33]20.4930.412.112
Ours16.6530.251.139
+ +for DeepCache is set to 2. For TGATE, we set $m = 5$ . For ToMe, we use a merging ratio of 0.3, while ToDo, Faster Diffusion, and CA-ToMe are evaluated using their default hyperparameters from their respective papers. For the transformer-based model PixArt- $\alpha$ , we compare Post-Diff with TGATE [22] using $m = 8$ , as well as DiTFastAttn [46] and DuCa [49] with their default settings on an NVIDIA H100 GPU, since some methods encounter out-of-memory errors on the A5000. + +Observations. Compared to previous methods that accelerate the generation process either at the module level (e.g., DeepCache, TGATE, DuCa) by reusing computations across denoising steps or at the input level (e.g., ToMe, ToDo) by compressing redundant image tokens, our method uniquely integrates both approaches, resulting in a greater + +Table 4. Benchmark with other training-free compression methods on PixArt-α. + +
MethodFID ↓Clip Score ↑Latency (s) ↓
Original29.1630.411.752
TGATE [22]29.5429.591.275
DiTFastAttn [46]45.4630.13\( 1.103^{\dagger} \)
DuCa [49]28.0830.121.724
Ours25.4430.231.382
+ +Faster speed is achieved with FlashAttention [6]. + +reduction in inference cost. Additionally, leveraging low-resolution images in the early stages enhances the denoising process, allowing our method to achieve the lowest FID with a competitive Clip Score across both SD V1.5 and PixArt- $\alpha$ models. + +These experiments highlight the benefits of leveraging redundancy at both the input and module levels. Integrating the aforementioned training-free compression techniques into PostDiff could further enhance efficiency, which we leave for future work. + +# 4. Related Work + +Diffusion models. Diffusion models have emerged as a powerful approach for generative tasks, modeling the gen + +![](images/315f9eb9be4ad714b545b7392630c906b8d3116c20b6812209d8b735fd140a07.jpg) +Figure 6. Visualize the generation results of different models w/o and w/ using our PostDiff. + +erative process as a denoising sequence to enable high-quality image generation [15, 36]. Latent Diffusion Models (LDMs) [31] leverage Variational Autoencoders (VAEs) [30] to operate in a compressed latent space, reducing memory demands and enabling efficient high-resolution generation. Open-source models like Stable Diffusion V1.5 [31], Stable Diffusion XL [29], and PixArt [4, 5] have further expanded accessibility and are now applied across multiple domains [8, 10, 11]. Despite their impressive generative quality, diffusion models often suffer from long generation times, limiting their broader applicability. + +Few-Step diffusion models. A primary focus in accelerating diffusion models is reducing the number of timesteps. DDIM [36] modifies DDPM [15] by introducing non-Markovian sampling, enabling faster generation. DPM-Solver [23] applies exponential integrators to approximate solutions to diffusion ODEs, enhancing sampling efficiency. Other approaches, such as consistency models [24, 37], aim to map points on the ODE trajectory to the original sample, effectively reducing inference steps. Adversarial training is also employed to enable one-step or few-step diffusion [43-45]. However, these few-step models still incur non-trivial training/distillation cost. + +Efficient diffusion models. Efforts to accelerate diffusion models extend from data optimization to model optimization. From the data perspective, prior works have reduced computational requirements by merging similar tokens [1] or pruning redundant tokens [39]. From the model + +perspective, recent works focus on identifying redundant feature maps across timesteps and using caching mechanisms to reuse these maps, significantly reducing computational costs [25, 42]. Additional approaches aim to slim down models through quantization to lower bit precision [12, 34, 48] or by pruning unnecessary parameters [2, 9, 41]. Additionally, several systems have been developed specifically to accelerate diffusion model inference [18, 19]. + +# 5. Conclusion + +In this work, we explore a critical question for compute-optimal diffusion model deployment: under a post-training setting, is it more effective to reduce denoising steps or decrease per-step inference cost? To investigate this, we propose PostDiff, a training-free framework that enhances diffusion model efficiency while systematically examining this trade-off. PostDiff reduces redundancy at two levels: at the input level, through a mixed-resolution denoising scheme that enhances low-frequency components in early steps, and at the module level, with a hybrid caching strategy. Extensive evaluations demonstrate that reducing per-step inference cost often achieves a better efficiency-fidelity balance than reducing denoising steps across a wide efficiency range. PostDiff not only offers a new perspective on the fidelity-efficiency trade-off for diffusion models but also provides practical insights for their deployment on resource-constrained platforms. + +# Acknowledgment + +This work is supported by the National Science Foundation (NSF) through CCRI funding (Award Number: 2016727), RTML funding (Award Number: 1937592), and an NSF CAREER award (Award Number: 2048183), as well as by CoCoSys, one of the seven centers in JUMP 2.0, a Semiconductor Research Corporation (SRC) program sponsored by DARPA. + +# References + +[1] Daniel Bolya and Judy Hoffman. Token merging for fast stable diffusion. In CVPR, 2023. 3, 6, 7, 8 +[2] Thibault Castells, Hyoung-Kyu Song, Bo-Kyeong Kim, and Shinkook Choi. LD-Pruner: Efficient pruning of latent diffusion models using task-agnostic insights. In CVPR, 2024. 8 +[3] Angela Castillo, Jonas Kohler, Juan C Pérez, Juan Pablo Pérez, Albert Pumarola, Bernard Ghanem, Pablo Arbeláez, and Ali Thabet. Adaptive guidance: Training-free acceleration of conditional diffusion models. In AAAI, 2025. 5 +[4] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. PixArt- $\Sigma$ : Weak-to-strong training of diffusion transformer for 4k text-to-image generation. In ECCV, 2024. 2, 8 +[5] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Zhongdao Wang, James T. Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. PixArt-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In ICLR, 2024. 5, 6, 7, 8 +[6] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. FlashAttention: Fast and memory-efficient exact attention with io-awareness. In NeurIPS, 2022. 7 +[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 5 +[8] Zhenbang Du, Wei Feng, Haohan Wang, Yaoyu Li, Jingsen Wang, Jian Li, Zheng Zhang, Jingjing Lv, Xin Zhu, Junsheng Jin, et al. Towards reliable advertising image generation using human feedback. In ECCV, 2024. 8 +[9] Gongfan Fang, Xinyin Ma, and Xinchao Wang. Structural pruning for diffusion models. In NeurIPS, 2023. 8 +[10] Sicheng Gao, Xuhui Liu, Bohan Zeng, Sheng Xu, Yanjing Li, Xiaoyan Luo, Jianzhuang Liu, Xiantong Zhen, and Baochang Zhang. Implicit diffusion models for continuous super-resolution. In CVPR, 2023. 8 +[11] He Guo, Zixuan Ye, Zhiguo Cao, and Hao Lu. In-context matting. In CVPR, 2024. 8 +[12] Yefei He, Luping Liu, Jing Liu, Weijia Wu, Hong Zhou, and Bohan Zhuang. PTQD: Accurate post-training quantization for diffusion models. In NeurIPS, 2023. 8 +[13] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. In EMNLP, 2021. 3 +[14] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshop, 2021. 5 + +[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 3, 8 +[16] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. JMLR, 2022. 2 +[17] Bo-Kyeong Kim, Hyoung-Kyu Song, Thibault Castells, and Shinkook Choi. BK-SDM: A lightweight, fast, and cheap version of stable diffusion. In ECCV, 2024. 5 +[18] Akio Kodaira, Chenfeng Xu, Toshiki Hazama, Takanori Yoshimoto, Kohei Ohno, Shogo Mitsuhori, Soichi Sugano, Hanying Cho, Zhijian Liu, and Kurt Keutzer. Streamdiffusion: A pipeline-level solution for real-time interactive generation. arXiv preprint, 2023. 8 +[19] Muyang Li, Tianle Cai, Jiaxin Cao, Qinsheng Zhang, Han Cai, Junjie Bai, Yangqing Jia, Kai Li, and Song Han. Distribution: Distributed parallel inference for high-resolution diffusion models. In CVPR, 2024. 2, 8 +[20] Senmao Li, Taihang Hu, Fahad Shahbaz Khan, Linxuan Li, Shiqi Yang, Yaxing Wang, Ming-Ming Cheng, and Jian Yang. Faster diffusion: Rethinking the role of unet encoder in diffusion models. In NeurIPS, 2024. 6, 7 +[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 4, 6 +[22] Haozhe Liu, Wentian Zhang, Jinheng Xie, Francesco Faccio, Mengmeng Xu, Tao Xiang, Mike Zheng Shou, Juan-Manuel Pérez-Rúa, and Jurgen Schmidhuber. Faster diffusion through temporal attention decomposition. TMLR, 2025. 2, 5, 6, 7 +[23] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. NeurIPS, 2022. 3, 8 +[24] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint, 2023. 5, 6, 7, 8 +[25] Xinyin Ma, Gongfan Fang, and Xinchao Wang. Deepcache: Accelerating diffusion models for free. In CVPR, 2024. 5, 6, 7, 8 +[26] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. In CVPR, 2023. 6 +[27] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2021. 5 +[28] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 3 +[29] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024. 2, 5, 6, 7, 8 + +[30] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. In NeurIPS, 2019. 3, 8 +[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 4, 5, 6, 7, 8 +[32] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 3 +[33] Omid Saghatchian, Atiyeh Gh Moghadam, and Ahmad Nickabadi. Cached adaptive token merging: Dynamic token reduction and redundant computation elimination in diffusion model. arXiv preprint, 2025. 6, 7 +[34] Yuzhang Shang, Zhihang Yuan, Bin Xie, Bingzhe Wu, and Yan Yan. Post-training quantization on diffusion models. In CVPR, 2023. 8 +[35] Ethan Smith, Nayan Saxena, and Aninda Saha. Todo: Token downsampling for efficient generation of high-resolution images. In *IJCAI*, 2024. 3, 6, 7 +[36] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 3, 5, 6, 8 +[37] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In ICML, 2023. 8 +[38] Jiayan Teng, Wendi Zheng, Ming Ding, Wenyi Hong, Jianqiao Wangni, Zhuoyi Yang, and Jie Tang. Relay diffusion: Unifying diffusion process across resolutions for image synthesis. In ICLR, 2024. 2 +[39] Hongjie Wang, Difan Liu, Yan Kang, Yijun Li, Zhe Lin, Niraj K Jha, and Yuchen Liu. Attention-driven training-free efficiency enhancement of diffusion models. In CVPR, 2024. 3, 8 +[40] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. LAVIE: High-quality video generation with cascaded latent diffusion models. IJCV, 2024. 2 +[41] Lexington Whalen, Zhenbang Du, Haoran You, Chaojian Li, Sixu Li, and Yingyan (Celine) Lin. Early-Bird Diffusion: Investigating and leveraging timestep-aware early-bird tickets in diffusion models for efficient training. In CVPR, 2025. 8 +[42] Felix Wimbauer, Bichen Wu, Edgar Schoenfeld, Xiaoliang Dai, Ji Hou, Zijian He, Artsiom Sanakoyeu, Peizhao Zhang, Sam Tsai, Jonas Kohler, et al. Cache me if you can: Accelerating diffusion models through block caching. In CVPR, 2024. 5, 8 +[43] Yanwu Xu, Yang Zhao, Zhisheng Xiao, and Tingbo Hou. UFOGen: You forward once large scale text-to-image generation via diffusion gans. In CVPR, 2024. 8 +[44] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved distribution matching distillation for fast image synthesis. In NeurIPS, 2024. +[45] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Frédo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In CVPR, 2024. 8 + +[46] Zhihang Yuan, Hanling Zhang, Lu Pu, Xuefei Ning, Linfeng Zhang, Tianchen Zhao, Shengen Yan, Guohao Dai, and Yu Wang. DiTFastattn: Attention compression for diffusion transformer models. In NeurIPS, 2024. 7 +[47] Shen Zhang, Zhaowei Chen, Zhenyu Zhao, Yuhao Chen, Yao Tang, and Jiajun Liang. HiDiffusion: Unlocking higher-resolution creativity and efficiency in pretrained diffusion models. In ECCV, 2025. 2 +[48] Tianchen Zhao, Tongcheng Fang, Enshu Liu, Rui Wan, Widyadewi Soedarmadji, Shiyao Li, Zinan Lin, Guohao Dai, Shengen Yan, Huazhong Yang, Xuefei Ning, and Yu Wang. ViDiT-Q: Efficient and accurate quantization of diffusion transformers for image and video generation. In ICLR, 2025. 8 +[49] Chang Zou, Evelyn Zhang, Runlin Guo, Haohang Xu, Conghui He, Xuming Hu, and Linfeng Zhang. Accelerating diffusion transformers with dual feature caching. arXiv preprint, 2024. 7 \ No newline at end of file diff --git a/2025/Fewer Denoising Steps or Cheaper Per-Step Inference_ Towards Compute-Optimal Diffusion Model Deployment/images.zip b/2025/Fewer Denoising Steps or Cheaper Per-Step Inference_ Towards Compute-Optimal Diffusion Model Deployment/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b3ab319087ac6e435b5b7a3a4f68602f6599fcd5 --- /dev/null +++ b/2025/Fewer Denoising Steps or Cheaper Per-Step Inference_ Towards Compute-Optimal Diffusion Model Deployment/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:860f22238c25184c691cc58f0e86b04979161a5c42b8f69594f5bec9d3e2aff0 +size 770854 diff --git a/2025/Fewer Denoising Steps or Cheaper Per-Step Inference_ Towards Compute-Optimal Diffusion Model Deployment/layout.json b/2025/Fewer Denoising Steps or Cheaper Per-Step Inference_ Towards Compute-Optimal Diffusion Model Deployment/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..319f447c790fcaaa5627b2d8186c0c540a27d786 --- /dev/null +++ b/2025/Fewer Denoising Steps or Cheaper Per-Step Inference_ Towards Compute-Optimal Diffusion Model Deployment/layout.json @@ -0,0 +1,8102 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 108, + 103, + 501, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 103, + 501, + 140 + ], + "spans": [ + { + "bbox": [ + 108, + 103, + 501, + 140 + ], + "type": "text", + "content": "Fewer Denoising Steps or Cheaper Per-Step Inference: Towards Compute-Optimal Diffusion Model Deployment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 148, + 161, + 462, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 161, + 462, + 219 + ], + "spans": [ + { + "bbox": [ + 148, + 161, + 462, + 219 + ], + "type": "text", + "content": "Zhenbang Du*, Yonggan Fu*, Lifu Wang*, Jiayi Qian, Xiao Luo, Yingyan (Celine) Lin \nGeorgia Institute of Technology \n{zdu89,jiayiqian,celine.lin}@gatech.edu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 270, + 296, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 270, + 296, + 617 + ], + "spans": [ + { + "bbox": [ + 56, + 270, + 296, + 617 + ], + "type": "text", + "content": "Diffusion models have shown remarkable success across generative tasks, yet their high computational demands challenge deployment on resource-limited platforms. This paper investigates a critical question for compute-optimal diffusion model deployment: Under a post-training setting without fine-tuning, is it more effective to reduce the number of denoising steps or to use a cheaper per-step inference? Intuitively, reducing the number of denoising steps increases the variability of the distributions across steps, making the model more sensitive to compression. In contrast, keeping more denoising steps makes the differences smaller, preserving redundancy, and making post-training compression more feasible. To systematically examine this, we propose PostDiff, a training-free framework for accelerating pre-trained diffusion models by reducing redundancy at both the input level and module level in a post-training manner. At the input level, we propose a mixed-resolution denoising scheme based on the insight that reducing generation resolution in early denoising steps can enhance low-frequency components and improve final generation fidelity. At the module level, we employ a hybrid module caching strategy to reuse computations across denoising steps. Extensive experiments and ablation studies demonstrate that (1) PostDiff can significantly improve the fidelity-efficiency trade-off of state-of-the-art diffusion models, and (2) to boost efficiency while maintaining decent generation fidelity, reducing per-step inference cost is often more effective than reducing the number of denoising steps. Our code is available at https://github.com/GATECH-EIC/PostDiff." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 639, + 135, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 639, + 135, + 651 + ], + "spans": [ + { + "bbox": [ + 56, + 639, + 135, + 651 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 659, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 295, + 696 + ], + "type": "text", + "content": "Diffusion models have rapidly advanced as a powerful tool for various generative tasks, such as image synthesis and video generation. Their success stems from the iterative de" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 246, + 553, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 246, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 553, + 376 + ], + "type": "text", + "content": "noising process, which gradually refines outputs to achieve high-quality results, along with advanced model architectures. However, both the iterative nature and model complexity of diffusion models make them computationally intensive, posing a significant challenge for deployment on resource-constrained platforms, where computational resources, memory, and power are limited. Efficient deployment of diffusion models on these platforms is therefore highly desirable to broaden the accessibility of generative AI, enabling real-time applications that can operate within the strict constraints of on-device resources." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 378, + 553, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 553, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 553, + 594 + ], + "type": "text", + "content": "With the increasing demand for rapid diffusion model deployment, where a given diffusion model is expected to be instantly compressed and deployed without fine-tuning, a critical question arises: Under a post-training setting, is it more effective to optimize for fewer denoising steps or to reduce the per-step inference cost? This question is particularly important because each approach entails different trade-offs: Reducing the number of denoising steps introduces greater variability in feature changes between steps, which can amplify the negative impact of compression on output fidelity; Conversely, using more denoising steps reduces these step-to-step differences, preserving redundancy and making it easier to apply post-training compression without sacrificing quality. However, there is a limited exploration of how these two strategies perform in terms of fidelity-efficiency trade-offs, and investigating this could provide valuable guidelines for efficient diffusion model deployment in real-world applications." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 594, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 553, + 713 + ], + "type": "text", + "content": "To address this research question, we propose PostDiff, a training-free diffusion model compression framework designed to enhance the per-step inference efficiency of pretrained diffusion models at both the input level and module level in a post-training manner. Our techniques are training-free and can thus be instantly applied to emerging diffusion models for rapid deployment, enabling a comparative study of their effectiveness versus reducing denoising steps. Specifically, at the input level, we propose a mixed-resolution denoising scheme that adopts a lower generation" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 134, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 134, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 134, + 712 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3001" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": "resolution in early denoising steps. As low-frequency components dominate at this stage, a reduced resolution is sufficient. We find that this reduced resolution in early denoising steps can enhance low-frequency components, thereby improving final generation fidelity and achieving a win-win in both efficiency and fidelity. At the module level, we build on existing module caching techniques, which reuse module outputs across steps, and enhance them with a hybrid caching strategy as another dimension for redundancy reduction. Our contributions are summarized as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 193, + 295, + 563 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 193, + 295, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 193, + 295, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 193, + 295, + 264 + ], + "type": "text", + "content": "- Our work, for the first time, studies a critical question for efficient diffusion model deployment: Under a posttraining setting without fine-tuning, is it more effective to optimize for fewer denoising steps or to reduce the per-step inference cost? The answer to this question could be useful for both researchers and practitioners." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 265, + 295, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 265, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 295, + 361 + ], + "type": "text", + "content": "- To answer the above question, we propose PostDiff, a training-free framework for reducing diffusion model complexity at both the input and module levels, resulting in lower per-step inference cost. This framework enables a fair study of the relative effectiveness between fewer denoising steps and reduced per-step inference cost toward compute-optimal diffusion model deployment under a post-training setting." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 361, + 295, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 295, + 468 + ], + "type": "text", + "content": "- At the input level, we propose a mixed-resolution denoising scheme, which leverages the observation that reduced generation resolution in early denoising steps can enhance low-frequency components and final output quality, achieving a win-win in both accuracy and efficiency. In addition, to further reduce the overall computational cost at the module level, PostDiff employs a hybrid module caching method that strategically reuses the outputs of different modules on varying schedules across steps." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 468, + 295, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 295, + 563 + ], + "type": "text", + "content": "- Extensive evaluations show that PostDiff can advance the fidelity-efficiency frontier of state-of-the-art (SOTA) diffusion models, as shown in Fig. 1, and provide valuable insights into best practices for efficient diffusion model deployment. For example, we find that reducing per-step inference cost is often more effective than decreasing the number of denoising steps to improve efficiency when aiming to preserve high generation fidelity." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 565, + 295, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 565, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 565, + 295, + 613 + ], + "type": "text", + "content": "In the following sections, we illustrate and analyze the effectiveness of each technique in PostDiff in Sec. 2, and finally, we combine and apply them to SOTA diffusion models in Sec. 3 to answer our research question." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 621, + 254, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 621, + 254, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 621, + 254, + 635 + ], + "type": "text", + "content": "2. PostDiff: The Proposed Framework" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "content": "In this section, we first introduce the proposed mixed-resolution denoising strategy, which reduces per-step inference cost at the input level. We begin by analyzing the motivations from previous works in Sec. 2.1, followed by an illustration of our method, an analysis of its effectiveness, and its application to examine our key question in Sec. 2.2," + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 318, + 68, + 548, + 228 + ], + "blocks": [ + { + "bbox": [ + 318, + 68, + 548, + 228 + ], + "lines": [ + { + "bbox": [ + 318, + 68, + 548, + 228 + ], + "spans": [ + { + "bbox": [ + 318, + 68, + 548, + 228 + ], + "type": "image", + "image_path": "96dcd55c2125876b6f232a70876722be8f7e2ed7d1dd6c05b3e0b978a4a821cd.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 229, + 553, + 252 + ], + "lines": [ + { + "bbox": [ + 313, + 229, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 229, + 553, + 252 + ], + "type": "text", + "content": "Figure 1. Apply our PostDiff technique to SOTA diffusion models, where “-PD” denotes the models delivered by PostDiff." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 258, + 553, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 553, + 294 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 553, + 294 + ], + "type": "text", + "content": "Sec. 2.3, and Sec. 2.4, respectively. Finally, we describe how we employ and enhance module caching to improve efficiency at the module level in Sec. 2.5." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 301, + 495, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 495, + 313 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 495, + 313 + ], + "type": "text", + "content": "2.1. Motivations from Previous Works" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 319, + 553, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 553, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 553, + 462 + ], + "type": "text", + "content": "Low-frequency components dominate the early denoising steps. Previous works [22] find that the typical denoising process for image generation progresses through two phases: the semantics-planning phase and the fidelity-improving phase. The early semantics-planning phase focuses more on the layout and low-frequency information, starting from random noise and gradually establishing the basic structure. This indicates that a low generation resolution could be sufficient in the early denoising steps. As generation proceeds, more details, described by high-frequency information, are incrementally added, thus requiring a higher resolution." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 463, + 554, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 554, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 554, + 617 + ], + "type": "text", + "content": "This insight is also partially echoed and leveraged by previous works. For example, [38] trains different diffusion models for early low-resolution denoising and later high-resolution denoising processes to avoid the propagation of inaccurate high-frequency components from early denoising steps to later steps. Additionally, cascade diffusion models [16, 40] sequentially apply different diffusion models at different denoising stages with varying generation resolutions, finding that incorporating low-frequency information from low-resolution images during high-resolution stages can enhance generation quality. Nevertheless, all these methods require multiple diffusion models, introducing significant storage and training demands." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 618, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 554, + 714 + ], + "type": "text", + "content": "Input resolution is a promising dimension for reducing inference cost. High-resolution generation is a recent trend, as it provides images with more vivid details and accommodates greater information density [47]. For example, SDXL [29] generates images at 1024x1024 resolution, while PixArt- " + }, + { + "bbox": [ + 313, + 618, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 313, + 618, + 554, + 714 + ], + "type": "text", + "content": " [4] expands resolution capability to 4K. However, as resolution increases, computational demands rise dramatically, leading to significant latency [19]. As" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3002" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 294, + 203 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 294, + 203 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 294, + 203 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 294, + 203 + ], + "type": "image", + "image_path": "136fc13924f38ecd291bd208ec099e0afa2ea63585db3dbe4ab13f47333c3b8b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 213, + 295, + 227 + ], + "lines": [ + { + "bbox": [ + 55, + 213, + 295, + 227 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 295, + 227 + ], + "type": "text", + "content": "Figure 2. An overview of our mixed-resolution denoising strategy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 234, + 296, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 234, + 296, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 234, + 296, + 283 + ], + "type": "text", + "content": "such, recent works aim to reduce the computational load by decreasing the number of tokens in samples, using methods like token merging [1], token pruning [39], or token down-sampling [35], which effectively accelerate inference." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 288, + 280, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 280, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 280, + 301 + ], + "type": "text", + "content": "2.2. Mixed-Resolution Denoising: Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 304, + 296, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 296, + 400 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 296, + 400 + ], + "type": "text", + "content": "Motivated by the tendency for low-frequency components in early denoising steps and the effectiveness of input resolution as a design knob for efficiency, we propose a simple yet effective post-training strategy: As shown in Fig. 2, we adopt a low generation resolution in the early denoising steps to generate low-frequency structures, and then refine high-frequency details with high generation resolution in the later denoising steps." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 401, + 296, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 296, + 509 + ], + "type": "text", + "content": "Formulation. To generate an image with shape " + }, + { + "bbox": [ + 55, + 401, + 296, + 509 + ], + "type": "inline_equation", + "content": "(W,H,3)" + }, + { + "bbox": [ + 55, + 401, + 296, + 509 + ], + "type": "text", + "content": ", we begin with Gaussian noise in the shape " + }, + { + "bbox": [ + 55, + 401, + 296, + 509 + ], + "type": "inline_equation", + "content": "(w,h,4)" + }, + { + "bbox": [ + 55, + 401, + 296, + 509 + ], + "type": "text", + "content": ", representing the latent noise image " + }, + { + "bbox": [ + 55, + 401, + 296, + 509 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 55, + 401, + 296, + 509 + ], + "type": "text", + "content": " obtained via Variational Autoencoders [30]. For simplicity, we omit the channel dimension in the remainder of this paper. The denoising process proceeds over " + }, + { + "bbox": [ + 55, + 401, + 296, + 509 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 401, + 296, + 509 + ], + "type": "text", + "content": " steps, using samplers like DDIM [36] or DPM-solver [23] to obtain the denoised latent at each step. Taking DDIM [36] as an example, we compute the latent representation " + }, + { + "bbox": [ + 55, + 401, + 296, + 509 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 401, + 296, + 509 + ], + "type": "text", + "content": " at step " + }, + { + "bbox": [ + 55, + 401, + 296, + 509 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 55, + 401, + 296, + 509 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 517, + 295, + 530 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 517, + 295, + 530 + ], + "spans": [ + { + "bbox": [ + 141, + 517, + 295, + 530 + ], + "type": "interline_equation", + "content": "x _ {t - 1} = \\sqrt {\\alpha_ {t - 1}} \\hat {x} _ {0} ^ {t} + x _ {t}, \\tag {1}", + "image_path": "c1e7c7ebec5b58c3080dd601afb4088f2448480bcfc84f6fb5b30ace88826ad7.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 111, + 533, + 295, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 533, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 111, + 533, + 295, + 559 + ], + "type": "interline_equation", + "content": "\\hat {x} _ {0} ^ {t} = \\frac {x _ {t} - \\sqrt {1 - \\alpha_ {t}} \\cdot \\epsilon_ {\\theta} (x _ {t} , t)}{\\sqrt {\\alpha_ {t}}}, \\tag {2}", + "image_path": "0536a4bfe3a07e972d85f64b2f2ef085c14893ec515bbfef9539d3b2a2983183.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 123, + 561, + 294, + 577 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 561, + 294, + 577 + ], + "spans": [ + { + "bbox": [ + 123, + 561, + 294, + 577 + ], + "type": "interline_equation", + "content": "x _ {t} = \\sqrt {1 - \\alpha_ {t - 1}} \\cdot \\epsilon_ {\\theta} (x _ {t}, t), \\tag {3}", + "image_path": "9aa9b213b1056f326dc963108c88bddf2841480d85e41a66da9c7caff9cd2b98.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 582, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 617 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 582, + 295, + 617 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 55, + 582, + 295, + 617 + ], + "type": "text", + "content": " denotes the model [28, 32] that predicts the added noise, and " + }, + { + "bbox": [ + 55, + 582, + 295, + 617 + ], + "type": "inline_equation", + "content": "\\{\\alpha_t\\}" + }, + { + "bbox": [ + 55, + 582, + 295, + 617 + ], + "type": "text", + "content": " is a set of coefficients controlling the forward noise-adding process [15]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": "In our proposed mixed-resolution denoising strategy, we first introduce a lower-resolution latent representation, " + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "inline_equation", + "content": "x_{T}^{l}" + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": ", with shape " + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "inline_equation", + "content": "(\\beta w, \\beta h)" + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "inline_equation", + "content": "0 < \\beta < 1" + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": " is a hyperparameter that scales the resolution down. This low-resolution starting point reduces the computational load of the initial denoising phase. Another hyperparameter, " + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": ", controls the transition from low to full resolution. After reaching the denoising step " + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "inline_equation", + "content": "t = sT" + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": ", we compute the forecasted" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 71, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 71, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 553, + 95 + ], + "type": "inline_equation", + "content": "\\hat{x}_0^{l,t}" + }, + { + "bbox": [ + 313, + 71, + 553, + 95 + ], + "type": "text", + "content": " at step " + }, + { + "bbox": [ + 313, + 71, + 553, + 95 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 71, + 553, + 95 + ], + "type": "text", + "content": " using Eq. (2). Then, we upscale this result to full resolution via:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 385, + 102, + 553, + 118 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 102, + 553, + 118 + ], + "spans": [ + { + "bbox": [ + 385, + 102, + 553, + 118 + ], + "type": "interline_equation", + "content": "\\hat {x} _ {0} ^ {f, t} = \\operatorname {U p s a m p l e} \\left(\\hat {x} _ {0} ^ {l, t}\\right), \\tag {4}", + "image_path": "357d7c11bb6ce365b86a756c44eb72a93f0a3551432ac3065d2fb9a067bff3e2.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 124, + 555, + 160 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 124, + 555, + 160 + ], + "spans": [ + { + "bbox": [ + 313, + 124, + 555, + 160 + ], + "type": "text", + "content": "where Upsample refers to bilinear interpolation. Using this upscaled version, we derive the full-resolution latent " + }, + { + "bbox": [ + 313, + 124, + 555, + 160 + ], + "type": "inline_equation", + "content": "x_{t}^{f}" + }, + { + "bbox": [ + 313, + 124, + 555, + 160 + ], + "type": "text", + "content": " at step " + }, + { + "bbox": [ + 313, + 124, + 555, + 160 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 124, + 555, + 160 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 376, + 166, + 553, + 182 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 166, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 376, + 166, + 553, + 182 + ], + "type": "interline_equation", + "content": "x _ {t} ^ {f} = \\sqrt {\\alpha_ {t}} \\hat {x} _ {0} ^ {f, t} + \\sqrt {1 - \\alpha_ {t}} \\epsilon , \\tag {5}", + "image_path": "398888dc099477fec3a08f36f2c8fe2c9cc5aa9b9b2cb1c566cee552d4d67eb7.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 187, + 553, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 187, + 553, + 212 + ], + "spans": [ + { + "bbox": [ + 313, + 187, + 553, + 212 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 187, + 553, + 212 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 187, + 553, + 212 + ], + "type": "text", + "content": " is Gaussian noise with shape " + }, + { + "bbox": [ + 313, + 187, + 553, + 212 + ], + "type": "inline_equation", + "content": "(w,h)" + }, + { + "bbox": [ + 313, + 187, + 553, + 212 + ], + "type": "text", + "content": ". The remainder of the denoising process follows the original procedure." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 212, + 554, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 212, + 554, + 284 + ], + "spans": [ + { + "bbox": [ + 313, + 212, + 554, + 284 + ], + "type": "text", + "content": "Note that we adopt this simple binary resolution strategy without bells and whistles to effectively demonstrate its concept and simplify the involved hyperparameters, thereby easing the use of our method in real-world deployment. More complex resolution schedules could potentially be proposed to further enhance fidelity-efficiency trade-offs." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 289, + 554, + 314 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 554, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 554, + 314 + ], + "type": "text", + "content": "2.3. Visualization: How Low-Resolution Denoising Contributes to Boosted Final Performance" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 319, + 554, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 554, + 414 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 554, + 414 + ], + "type": "text", + "content": "To study the impact of our mixed-resolution denoising strategy on final performance and understand why it works, we visualize the generation quality after each denoising step, both with and without our method. Specifically, we calculate the Clip Score [13] for " + }, + { + "bbox": [ + 313, + 319, + 554, + 414 + ], + "type": "inline_equation", + "content": "\\hat{x}_0" + }, + { + "bbox": [ + 313, + 319, + 554, + 414 + ], + "type": "text", + "content": " predicted at each denoising step and annotate it in the case study of Fig. 3, where different rows correspond to the original full-resolution denoising process as well as our method with various choices of " + }, + { + "bbox": [ + 313, + 319, + 554, + 414 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 319, + 554, + 414 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 415, + 554, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 554, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 554, + 617 + ], + "type": "text", + "content": "Observations and analysis. We observe that (1) with a moderate amount of low-resolution generation steps, the Clip Score of the final generated image can be improved, achieving a win-win in generation fidelity and efficiency; (2) Regarding the evolution of the Clip Score across denoising steps, starting with a lower generation resolution results in lower Clip Scores initially, but serves as a better starting point: after switching to high generation resolution, the Clip Score increases more rapidly and gradually surpasses that of full-resolution generation. This echoes the analysis in Sec. 2.1, where reduced generation resolution in early denoising steps contributes to improved low-frequency components, thereby enhancing final generation quality; (3) The balance between low-resolution and high-resolution steps is crucial, as both low-frequency structure and high-frequency details need to be preserved to ensure high final generation quality in the last denoising step." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 618, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 554, + 665 + ], + "type": "text", + "content": "This phenomenon demonstrates the underlying effectiveness of our mixed-resolution denoising strategy and is consistently observed across samples, as shown in the comprehensive evaluation in Sec. 2.4." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 672, + 528, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 528, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 528, + 685 + ], + "type": "text", + "content": "2.4. Mixed-Resolution Denoising: Evaluation" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "We evaluate the effectiveness of our mixed-resolution denoising strategy and examine which is more effective: re" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3003" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 152, + 72, + 515, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 72, + 515, + 95 + ], + "spans": [ + { + "bbox": [ + 152, + 72, + 515, + 95 + ], + "type": "text", + "content": "cute girl, crop-top, blond hair, black glasses, stretching, with background by greg rutkowski makoto shinkai kyoto animation key art feminine mid shot" + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 65, + 97, + 550, + 316 + ], + "blocks": [ + { + "bbox": [ + 65, + 97, + 550, + 316 + ], + "lines": [ + { + "bbox": [ + 65, + 97, + 550, + 316 + ], + "spans": [ + { + "bbox": [ + 65, + 97, + 550, + 316 + ], + "type": "image", + "image_path": "e928864cca10772719024a02063b86d8e738c2c6c3451fbd49313102fcc912ea.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 323, + 555, + 346 + ], + "lines": [ + { + "bbox": [ + 55, + 323, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 555, + 346 + ], + "type": "text", + "content": "Figure 3. Visualize the Clip Score after each denoising step using full resolution (row 1) and mixed resolution (rows 2-5). The steps using low generation resolution are highlighted in green and those using high generation resolution are highlighted in orange." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 59, + 356, + 293, + 549 + ], + "blocks": [ + { + "bbox": [ + 59, + 356, + 293, + 549 + ], + "lines": [ + { + "bbox": [ + 59, + 356, + 293, + 549 + ], + "spans": [ + { + "bbox": [ + 59, + 356, + 293, + 549 + ], + "type": "image", + "image_path": "c35d0dfc1fecf32689f4de3fa8455c0fbac17c375acf1287ab60e6d0e8bb4c3e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 559, + 295, + 593 + ], + "lines": [ + { + "bbox": [ + 55, + 559, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 295, + 593 + ], + "type": "text", + "content": "Figure 4. The FID-FLOPs trade-off achieved under different denoising steps and varied mixed-resolution settings. The data points with the same number of steps are annotated using the same shape." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 603, + 278, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 603, + 278, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 603, + 278, + 615 + ], + "type": "text", + "content": "duced input-level redundancy or fewer denoising steps." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "Setup. We use Stable Diffusion (SD) V1.5 (DreamShaper-7 version) [31], a popular latent diffusion model (LDM), as our model backbone and evaluate text-to-image generation performance on the MS-COCO 2014 validation dataset [21]. We vary the total denoising steps " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "T = \\{8,10,12,14,16,18,20\\}" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": " and apply our mixed-resolution denoising to each denoising step setting with varied " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "s = \\{0.1,0.2,0.3,0.4,0.5,0.6\\}" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": ". Without bells and" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 356, + 555, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 356, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 555, + 392 + ], + "type": "text", + "content": "whistles, we set " + }, + { + "bbox": [ + 313, + 356, + 555, + 392 + ], + "type": "inline_equation", + "content": "\\beta = 0.5" + }, + { + "bbox": [ + 313, + 356, + 555, + 392 + ], + "type": "text", + "content": " as our low-resolution setting, i.e., half the full resolution, and ablation studies on this are provided in the supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 397, + 556, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 556, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 556, + 649 + ], + "type": "text", + "content": "Observations and analysis. As shown in Fig. 4, we observe that (1) mixed-resolution denoising can boost achievable generation fidelity by notably reducing the optimal FID (lower is better) while simultaneously improving generation efficiency, aligning with the analysis of the role of low resolution in early denoising steps in Sec. 2.1 and Sec. 2.3; (2) With more inference steps " + }, + { + "bbox": [ + 313, + 397, + 556, + 649 + ], + "type": "inline_equation", + "content": "(T \\geq 14)" + }, + { + "bbox": [ + 313, + 397, + 556, + 649 + ], + "type": "text", + "content": ", the optimal values of " + }, + { + "bbox": [ + 313, + 397, + 556, + 649 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 397, + 556, + 649 + ], + "type": "text", + "content": " are around " + }, + { + "bbox": [ + 313, + 397, + 556, + 649 + ], + "type": "inline_equation", + "content": "0.4 \\sim 0.5" + }, + { + "bbox": [ + 313, + 397, + 556, + 649 + ], + "type": "text", + "content": ", offering the best FID-FLOPs trade-off. For example, adopting " + }, + { + "bbox": [ + 313, + 397, + 556, + 649 + ], + "type": "inline_equation", + "content": "s = 0.5" + }, + { + "bbox": [ + 313, + 397, + 556, + 649 + ], + "type": "text", + "content": " with 20 denoising steps can save about " + }, + { + "bbox": [ + 313, + 397, + 556, + 649 + ], + "type": "inline_equation", + "content": "37\\%" + }, + { + "bbox": [ + 313, + 397, + 556, + 649 + ], + "type": "text", + "content": " FLOPs while achieving an FID reduction of approximately 2 compared to the original full-resolution model. (3) When aiming for relatively high generation fidelity, e.g., a " + }, + { + "bbox": [ + 313, + 397, + 556, + 649 + ], + "type": "inline_equation", + "content": "< 20" + }, + { + "bbox": [ + 313, + 397, + 556, + 649 + ], + "type": "text", + "content": " FID, which aligns with real-world application demands, maintaining a high number of denoising steps and reducing per-step inference cost via mixed resolution is a more cost-effective strategy than reducing denoising steps. Conversely, when targeting extreme efficiency, e.g., reducing " + }, + { + "bbox": [ + 313, + 397, + 556, + 649 + ], + "type": "inline_equation", + "content": ">60\\%" + }, + { + "bbox": [ + 313, + 397, + 556, + 649 + ], + "type": "text", + "content": " FLOPs, using fewer denoising steps becomes increasingly beneficial, as the effectiveness of mixed-resolution denoising diminishes in low-step generation with larger gaps between steps." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "Calibration for best hyperparameters. In practice, to rapidly determine the best transition step " + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": " (as well as the total number of denoising steps) for a given diffusion model provided by the users, we can use a small calibration set to obtain the optimal setting. We show in the supplemen" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3004" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "content": "tary materials that performance on a small calibration set is highly correlated with that on the full evaluation set." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 106, + 257, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 106, + 257, + 121 + ], + "spans": [ + { + "bbox": [ + 55, + 106, + 257, + 121 + ], + "type": "text", + "content": "2.5. The Hybrid Module Caching Strategy" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 125, + 296, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 125, + 296, + 173 + ], + "spans": [ + { + "bbox": [ + 55, + 125, + 296, + 173 + ], + "type": "text", + "content": "In addition to the resolution dimension, we further reduce model complexity at the module level by integrating existing module caching techniques [25, 42] with an enhanced hybrid caching strategy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 174, + 297, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 174, + 297, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 174, + 297, + 365 + ], + "type": "text", + "content": "Motivation. Previous works [25, 42] have found that feature maps remain highly similar across all blocks throughout the entire denoising process, indicating the potential effectiveness of caching and reusing across steps. We have further validated this through cross-step feature map distance visualization in the supplementary materials. Considering that (1) the text guidance provided by cross-attention layers primarily determines the image layout, i.e., the low-frequency structure of the image, and (2) the semantics-planning phase for layout generation occurs in the early denoising stages [22], as also discussed in Sec. 2.1, it is feasible to use the cached cross-attention feature map from early denoising steps, generated at a low resolution with our mixed-resolution denoising, for the later denoising phase, as the low-frequency information is determined early on." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 366, + 296, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 296, + 510 + ], + "type": "text", + "content": "Our caching strategy. We employ a hybrid module caching strategy to maximize reuse opportunities in a pretrained diffusion model. To achieve this, we integrate the following two caching mechanisms: (1) similar to DeepCache [25], we cache the deep skip branches of the U-Net, a major component of a wide range of SOTA diffusion models [17, 24, 29, 31], and reuse them for the subsequent " + }, + { + "bbox": [ + 55, + 366, + 296, + 510 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 366, + 296, + 510 + ], + "type": "text", + "content": " steps. Here, " + }, + { + "bbox": [ + 55, + 366, + 296, + 510 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 366, + 296, + 510 + ], + "type": "text", + "content": " controls how often the cache is updated, balancing efficiency gains and potential degradation, and is set to 2 throughout the evaluation; (2) We cache the crossattention layers at a specific denoising step and reuse them in all subsequent steps [22]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 510, + 297, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 297, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 297, + 594 + ], + "type": "text", + "content": "Design choices for caching cross-attention. A major role of cross-attention is to introduce text guidance into the generation process. This is often achieved through classifier-free guidance (CFG), which provides a controllable method for incorporating conditional guidance into image generation [7, 14, 27]. Specifically, the text guidance from the model's implicit classifier is derived as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 617, + 299, + 642 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 299, + 642 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 299, + 642 + ], + "type": "interline_equation", + "content": "\\epsilon_ {\\theta} \\left(x _ {t}, t, w, c\\right) = \\epsilon_ {\\theta} \\left(x _ {t}, t, \\emptyset\\right) + w \\left(\\epsilon_ {\\theta} \\left(x _ {t}, t, c\\right) - \\epsilon_ {\\theta} \\left(x _ {t}, t, \\emptyset\\right)\\right), \\tag {6}", + "image_path": "4ca8dbb6b083c6a3100a6e587f8fd78cff892ff9ef419cdfb432eebd737d6c75.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 653, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 297, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 653, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\varnothing" + }, + { + "bbox": [ + 55, + 653, + 297, + 715 + ], + "type": "text", + "content": " represents the null text condition. Considering that CFG may be redundant in the later stages of inference, as the layout has already been organized in the early stages [3, 22], we abandon CFG after a certain step " + }, + { + "bbox": [ + 55, + 653, + 297, + 715 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 653, + 297, + 715 + ], + "type": "text", + "content": ". When combined with our mixed-resolution denoising scheme, most" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 322, + 96, + 545, + 245 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 555, + 92 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 92 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 92 + ], + "type": "text", + "content": "Table 1. Ablation study on different caching strategies (" + }, + { + "bbox": [ + 313, + 71, + 555, + 92 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 71, + 555, + 92 + ], + "type": "text", + "content": ", Design Choice)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 322, + 96, + 545, + 245 + ], + "lines": [ + { + "bbox": [ + 322, + 96, + 545, + 245 + ], + "spans": [ + { + "bbox": [ + 322, + 96, + 545, + 245 + ], + "type": "table", + "html": "
SettingFLOPs (T) ↓FID ↓Clip Score ↑
Original30.42018.4230.80
Original w/o CFG15.21031.9825.52
DC17.78717.7930.75
DC+CA (5, Ave)18.7728.40
DC+CA (5, Cond)11.61018.8229.20
DC+CA (5, Uncond)20.3325.50
DC+CA (5, CFG)103.7118.22
DC+CA (10, Ave)21.1330.03
DC+CA (10, Cond)15.06121.2630.11
DC+CA (10, Uncond)21.1929.83
DC+CA (10, CFG)24.0029.12
DC+CA (15, Ave)21.7030.35
DC+CA (15, Cond)16.36021.6730.37
DC+CA (15, Uncond)21.8230.34
DC+CA (15, CFG)19.4330.48
", + "image_path": "3a142858f42efcf3739b1b61a78a7100b4a1eb26004de2c73503e92fcd95da42.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 255, + 553, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 553, + 278 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 553, + 278 + ], + "type": "text", + "content": "CFG operations are performed under a low generation resolution, further reducing computational costs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 279, + 555, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 555, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 555, + 314 + ], + "type": "text", + "content": "Based on the type of guidance to be reused, there are four design choices for caching the cross-attention maps for subsequent steps:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 315, + 495, + 365 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 313, + 315, + 479, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 479, + 330 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 479, + 330 + ], + "type": "text", + "content": "- Ave [22]: " + }, + { + "bbox": [ + 313, + 315, + 479, + 330 + ], + "type": "inline_equation", + "content": "CA_{\\mathrm{cache}} = \\frac{1}{2}\\left(CA_t^c +CA_t^\\emptyset\\right)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 329, + 419, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 329, + 419, + 340 + ], + "spans": [ + { + "bbox": [ + 314, + 329, + 419, + 340 + ], + "type": "text", + "content": "- Cond: " + }, + { + "bbox": [ + 314, + 329, + 419, + 340 + ], + "type": "inline_equation", + "content": "CA_{\\mathrm{cache}} = CA_{t}^{c}" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 340, + 430, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 340, + 430, + 352 + ], + "spans": [ + { + "bbox": [ + 314, + 340, + 430, + 352 + ], + "type": "text", + "content": "- Uncond: " + }, + { + "bbox": [ + 314, + 340, + 430, + 352 + ], + "type": "inline_equation", + "content": "CA_{\\mathrm{cache}} = CA_t^0" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 351, + 495, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 351, + 495, + 365 + ], + "spans": [ + { + "bbox": [ + 314, + 351, + 495, + 365 + ], + "type": "text", + "content": "- CFG: " + }, + { + "bbox": [ + 314, + 351, + 495, + 365 + ], + "type": "inline_equation", + "content": "CA_{\\mathrm{cache}} = CA_t^\\emptyset + w(CA_t^c - CA_t^\\emptyset)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 366, + 555, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 555, + 415 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 366, + 555, + 415 + ], + "type": "inline_equation", + "content": "CA_{t}^{c}" + }, + { + "bbox": [ + 313, + 366, + 555, + 415 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 366, + 555, + 415 + ], + "type": "inline_equation", + "content": "CA_{t}^{\\emptyset}" + }, + { + "bbox": [ + 313, + 366, + 555, + 415 + ], + "type": "text", + "content": " are the cross-attention conditioned on the input text and null text, respectively. Using either of these maps or a mixture of them results in varying degrees of conditioning in the subsequent generation process." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 415, + 556, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 556, + 606 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 556, + 606 + ], + "type": "text", + "content": "Exploring the best cross-attention caching strategy. We empirically examine combinations of different caching strategies, including DeepCache (DC) combined with the four design choices for cross-attention caching on top of an SD V1.5 model. As shown in Tab. 1, we observe that (1) fully abandoning CFG leads to a collapse in performance, while incorporating the cross-attention caching mechanism can better balance performance and speed; (2) adding CA on top of DC further reduces computations due to redundant CFG usage in later denoising phases. Intuitively, using a smaller " + }, + { + "bbox": [ + 313, + 415, + 556, + 606 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 415, + 556, + 606 + ], + "type": "text", + "content": ", i.e., abandoning CFG in earlier timesteps, reduces more FLOPs at the cost of a slight decrease in generation quality; (3) among the four design choices, \"Cond\" generally performs best, slightly surpassing \"Ave\" at larger " + }, + { + "bbox": [ + 313, + 415, + 556, + 606 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 415, + 556, + 606 + ], + "type": "text", + "content": " values. Therefore, we adopt \"Cond\" as the " + }, + { + "bbox": [ + 313, + 415, + 556, + 606 + ], + "type": "inline_equation", + "content": "CA_{\\mathrm{cache}}" + }, + { + "bbox": [ + 313, + 415, + 556, + 606 + ], + "type": "text", + "content": " choice in all subsequent experiments." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 616, + 504, + 631 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 504, + 631 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 504, + 631 + ], + "type": "text", + "content": "3. PostDiff: Combine All Techniques" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 635, + 422, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 422, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 422, + 649 + ], + "type": "text", + "content": "3.1. Experiment Setup" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "content": "We apply our PostDiff framework, integrating mixed-resolution denoising and hybrid module caching, to several SOTA diffusion models, including SD V1.5 (DreamShaper-7 version) [31], Latent Consistency Model (LCM) [24, 36], Stable Diffusion XL (SDXL) [29], and PixArt-" + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "content": " [5]. We" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3005" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "content": "evaluate text-to-image generation performance on the MS-COCO 2014 validation dataset [21]. PostDiff settings are calibrated on a small set for each model, with the detailed configurations outlined as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 122, + 295, + 315 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 56, + 122, + 295, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 122, + 295, + 158 + ], + "spans": [ + { + "bbox": [ + 56, + 122, + 295, + 158 + ], + "type": "text", + "content": "1. SD V1.5 [31]: We use the DreamShaper-v7 fine-tuned version with a final resolution of " + }, + { + "bbox": [ + 56, + 122, + 295, + 158 + ], + "type": "inline_equation", + "content": "768 \\times 768" + }, + { + "bbox": [ + 56, + 122, + 295, + 158 + ], + "type": "text", + "content": ". Settings: " + }, + { + "bbox": [ + 56, + 122, + 295, + 158 + ], + "type": "inline_equation", + "content": "\\beta = 1/2" + }, + { + "bbox": [ + 56, + 122, + 295, + 158 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 122, + 295, + 158 + ], + "type": "inline_equation", + "content": "s = 1/2" + }, + { + "bbox": [ + 56, + 122, + 295, + 158 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 122, + 295, + 158 + ], + "type": "inline_equation", + "content": "w = 7.5" + }, + { + "bbox": [ + 56, + 122, + 295, + 158 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 122, + 295, + 158 + ], + "type": "inline_equation", + "content": "m = 15" + }, + { + "bbox": [ + 56, + 122, + 295, + 158 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 158, + 295, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 158, + 295, + 230 + ], + "spans": [ + { + "bbox": [ + 55, + 158, + 295, + 230 + ], + "type": "text", + "content": "2. Latent Consistency Model (LCM) [24, 36]: A distilled version of Stable Diffusion designed for few-step generation, producing " + }, + { + "bbox": [ + 55, + 158, + 295, + 230 + ], + "type": "inline_equation", + "content": "768 \\times 768" + }, + { + "bbox": [ + 55, + 158, + 295, + 230 + ], + "type": "text", + "content": " images. Settings: " + }, + { + "bbox": [ + 55, + 158, + 295, + 230 + ], + "type": "inline_equation", + "content": "s = 1/2" + }, + { + "bbox": [ + 55, + 158, + 295, + 230 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 158, + 295, + 230 + ], + "type": "inline_equation", + "content": "\\beta = 1/2" + }, + { + "bbox": [ + 55, + 158, + 295, + 230 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 158, + 295, + 230 + ], + "type": "inline_equation", + "content": "w = 7.5" + }, + { + "bbox": [ + 55, + 158, + 295, + 230 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 158, + 295, + 230 + ], + "type": "inline_equation", + "content": "m = 4" + }, + { + "bbox": [ + 55, + 158, + 295, + 230 + ], + "type": "text", + "content": ". Note that since CFG is integrated into a guidance embedding [26] and the model has a few-step nature, only " + }, + { + "bbox": [ + 55, + 158, + 295, + 230 + ], + "type": "inline_equation", + "content": "CA_{\\mathrm{cache}}" + }, + { + "bbox": [ + 55, + 158, + 295, + 230 + ], + "type": "text", + "content": " is used without DC." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 230, + 294, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 230, + 294, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 230, + 294, + 266 + ], + "type": "text", + "content": "3. Stable Diffusion XL (SDXL) [29]: A large-scale diffusion model capable of generating " + }, + { + "bbox": [ + 55, + 230, + 294, + 266 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 55, + 230, + 294, + 266 + ], + "type": "text", + "content": " images. Settings: " + }, + { + "bbox": [ + 55, + 230, + 294, + 266 + ], + "type": "inline_equation", + "content": "\\beta = 3/4" + }, + { + "bbox": [ + 55, + 230, + 294, + 266 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 230, + 294, + 266 + ], + "type": "inline_equation", + "content": "s = 1/5" + }, + { + "bbox": [ + 55, + 230, + 294, + 266 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 230, + 294, + 266 + ], + "type": "inline_equation", + "content": "w = 5.0" + }, + { + "bbox": [ + 55, + 230, + 294, + 266 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 230, + 294, + 266 + ], + "type": "inline_equation", + "content": "m = 15" + }, + { + "bbox": [ + 55, + 230, + 294, + 266 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 266, + 294, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 294, + 315 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 294, + 315 + ], + "type": "text", + "content": "4. PixArt-" + }, + { + "bbox": [ + 55, + 266, + 294, + 315 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 266, + 294, + 315 + ], + "type": "text", + "content": " [5]: A transformer-based diffusion model generating " + }, + { + "bbox": [ + 55, + 266, + 294, + 315 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 55, + 266, + 294, + 315 + ], + "type": "text", + "content": " images. Due to the absence of bypass connections, only " + }, + { + "bbox": [ + 55, + 266, + 294, + 315 + ], + "type": "inline_equation", + "content": "CA_{\\mathrm{cache}}" + }, + { + "bbox": [ + 55, + 266, + 294, + 315 + ], + "type": "text", + "content": " is used without enabling DC. Settings: " + }, + { + "bbox": [ + 55, + 266, + 294, + 315 + ], + "type": "inline_equation", + "content": "\\beta = 3/4" + }, + { + "bbox": [ + 55, + 266, + 294, + 315 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 266, + 294, + 315 + ], + "type": "inline_equation", + "content": "s = 1/2" + }, + { + "bbox": [ + 55, + 266, + 294, + 315 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 266, + 294, + 315 + ], + "type": "inline_equation", + "content": "w = 4.5" + }, + { + "bbox": [ + 55, + 266, + 294, + 315 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 266, + 294, + 315 + ], + "type": "inline_equation", + "content": "m = 15" + }, + { + "bbox": [ + 55, + 266, + 294, + 315 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 322, + 295, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 322, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 322, + 295, + 346 + ], + "type": "text", + "content": "3.2. Fewer Denoising Steps vs. Cheaper Per-step Inference" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 352, + 295, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 352, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 295, + 460 + ], + "type": "text", + "content": "After combining the hybrid module caching with our mixed-resolution strategy to establish the complete PostDiff framework, we examine our key question by benchmarking with the FID-FLOPs/Clip Score-FLOPs trade-offs achieved by reducing denoising steps. Specifically, for SD V1.5, we use 20 denoising steps, a fixed mixed-resolution setting of " + }, + { + "bbox": [ + 55, + 352, + 295, + 460 + ], + "type": "inline_equation", + "content": "s = 1/2" + }, + { + "bbox": [ + 55, + 352, + 295, + 460 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 352, + 295, + 460 + ], + "type": "inline_equation", + "content": "\\beta = 1/2" + }, + { + "bbox": [ + 55, + 352, + 295, + 460 + ], + "type": "text", + "content": ", and vary cache settings " + }, + { + "bbox": [ + 55, + 352, + 295, + 460 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 352, + 295, + 460 + ], + "type": "text", + "content": " to create the trade-off. As for PixArt- " + }, + { + "bbox": [ + 55, + 352, + 295, + 460 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 352, + 295, + 460 + ], + "type": "text", + "content": ", we use a fixed 20 denoising steps with " + }, + { + "bbox": [ + 55, + 352, + 295, + 460 + ], + "type": "inline_equation", + "content": "\\beta = 3/4" + }, + { + "bbox": [ + 55, + 352, + 295, + 460 + ], + "type": "text", + "content": ", and vary " + }, + { + "bbox": [ + 55, + 352, + 295, + 460 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 352, + 295, + 460 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 352, + 295, + 460 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 352, + 295, + 460 + ], + "type": "text", + "content": " to form the trade-off." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 460, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 652 + ], + "type": "text", + "content": "Observations and insights. From Fig. 5, we observe that (1) the complete PostDiff framework, incorporating both techniques, achieves the best FID-FLOPs trade-off across all FLOPs ranges. This demonstrates that exploiting model redundancy at both the input and module levels can yield better FID-FLOPs frontiers; (2) in terms of semantic alignment measured by Clip Score, reducing per-step inference cost can achieve a comparable Clip Score-FLOPs trade-off compared to decreasing the number of denoising steps over a wide range; while in scenarios requiring extreme efficiency, decreasing the number of denoising steps can maintain better semantic alignment with a slightly higher Clip Score. Generally, reducing per-step inference cost is more effective than decreasing the number of denoising steps with better fidelity and comparable semantic alignment across a wide efficiency range." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 659, + 258, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 258, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 258, + 672 + ], + "type": "text", + "content": "3.3. Evaluation on SOTA Diffusion Models" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "We further apply PostDiff to more diffusion models. As shown in Tab. 2, we present the evaluation results in terms of both generation quality and efficiency under (1) reduced" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 70, + 434, + 186 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 434, + 186 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 434, + 186 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 434, + 186 + ], + "type": "image", + "image_path": "c5e652a211aa5b85391ba0656d5af52224fea297dab7a5dc223b6c232d3cc68f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 434, + 70, + 549, + 186 + ], + "blocks": [ + { + "bbox": [ + 434, + 70, + 549, + 186 + ], + "lines": [ + { + "bbox": [ + 434, + 70, + 549, + 186 + ], + "spans": [ + { + "bbox": [ + 434, + 70, + 549, + 186 + ], + "type": "image", + "image_path": "8e9bfa759e3976a8c6774708f5fae89e4a918fc30b330f856c87fc4ad13b0392.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 317, + 189, + 434, + 306 + ], + "blocks": [ + { + "bbox": [ + 317, + 189, + 434, + 306 + ], + "lines": [ + { + "bbox": [ + 317, + 189, + 434, + 306 + ], + "spans": [ + { + "bbox": [ + 317, + 189, + 434, + 306 + ], + "type": "image", + "image_path": "aa2e3c0349cd93b32b763868b52233031b40e3b2ec630b2fa58a7a947cfb17fb.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 316, + 555, + 349 + ], + "lines": [ + { + "bbox": [ + 313, + 316, + 555, + 349 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 555, + 349 + ], + "type": "text", + "content": "Figure 5. Comparison of FID-FLOPs/Clip Score-FLOPs trade-offs achieved by reducing per-step inference cost or the number of denoising steps on SD V1.5 and PixArt-" + }, + { + "bbox": [ + 313, + 316, + 555, + 349 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 316, + 555, + 349 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 434, + 189, + 549, + 306 + ], + "blocks": [ + { + "bbox": [ + 434, + 189, + 549, + 306 + ], + "lines": [ + { + "bbox": [ + 434, + 189, + 549, + 306 + ], + "spans": [ + { + "bbox": [ + 434, + 189, + 549, + 306 + ], + "type": "image", + "image_path": "7d3291122459d82ba4fb5a45ec4a2905e319531fb2378879f59adf4f1f4e30d1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 361, + 553, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 553, + 385 + ], + "type": "text", + "content": "denoising steps or (2) maximal denoising steps with reduced model complexity via PostDiff." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 386, + 554, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 386, + 554, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 554, + 590 + ], + "type": "text", + "content": "Observations and analysis. We observe that: (1) With reduced denoising steps, the improved efficiency comes at the cost of degraded generation quality, reflected in notably increased FID and reduced Clip score. In contrast, maintaining the maximum number of denoising steps while applying PostDiff better preserves generation quality, achieving lower (i.e., better) FID and comparable Clip scores compared to reducing the denoising steps. For instance, on the SD V1.5 model, PostDiff achieves a " + }, + { + "bbox": [ + 313, + 386, + 554, + 590 + ], + "type": "inline_equation", + "content": "63.14\\%" + }, + { + "bbox": [ + 313, + 386, + 554, + 590 + ], + "type": "text", + "content": " reduction in FLOPs while improving FID by 1.8 compared to the original model. (2) PostDiff consistently demonstrates effectiveness across various diffusion models, including both large-scale and small-scale ones, LDMs and LCM, as well as U-Net-based and transformer-based models. This effectiveness is due to the broad applicability of the insights leveraged in Sec. 2.1 and Sec. 2.5. We also visualize the generation results in Fig. 6." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 598, + 553, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 598, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 598, + 553, + 624 + ], + "type": "text", + "content": "3.4. Benchmark with Other Training-Free Compression Methods" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 628, + 553, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 553, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 553, + 652 + ], + "type": "text", + "content": "We further benchmark PostDiff with other training-free diffusion model compression methods in Tab. 3 and Tab. 4." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 654, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 712 + ], + "type": "text", + "content": "Settings. For the U-Net-based model SD V1.5, we evaluate our method against DeepCache [25], TGATE [22], ToDo [35], ToMe [1], Faster Diffusion [20], and CA-ToMe [33]. Latency is measured as the single-image generation time on an NVIDIA A5000 GPU. The cache interval" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3006" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 101, + 541, + 370 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 2. Apply PostDiff with mixed-resolution denoising and hybrid module caching to SOTA diffusion models with varying numbers of denoising steps. Latency is measured as the single-image generation time on an NVIDIA A5000 GPU." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 101, + 541, + 370 + ], + "lines": [ + { + "bbox": [ + 70, + 101, + 541, + 370 + ], + "spans": [ + { + "bbox": [ + 70, + 101, + 541, + 370 + ], + "type": "table", + "html": "
ModelStepsMixCacheFID ↓Clip Score ↑FLOPs (T) ↓Latency (s) ↓
SD V1.5 [31]820.6030.4112.1681.298
1617.7430.7624.3362.377
2018.4230.8030.4202.930
20✓15.6930.7819.0351.945
20✓✓16.6530.2511.1291.139
LCM [24]423.5528.643.0420.523
822.9628.726.0840.825
8✓21.0128.923.8070.666
8✓✓22.7628.583.6860.651
SDXL [29]818.0130.9247.8562.843
1614.3531.9295.7125.211
2014.1031.95119.6416.521
20✓14.3731.62109.2645.950
20✓✓14.1831.1152.6823.119
PixArt-α [5]833.0930.2134.2563.031
1629.6230.4068.5125.725
2029.1630.4185.6407.093
20✓25.9930.3866.9905.645
20✓✓25.4430.2354.7184.768
", + "image_path": "eaa6554f087eba3df5eac04250ae0c2396438edc38e6b54fa6fab9b840b8775b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 58, + 407, + 300, + 517 + ], + "blocks": [ + { + "bbox": [ + 55, + 387, + 295, + 398 + ], + "lines": [ + { + "bbox": [ + 55, + 387, + 295, + 398 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 295, + 398 + ], + "type": "text", + "content": "Table 3. Benchmark with other compression methods on SD V1.5." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 407, + 300, + 517 + ], + "lines": [ + { + "bbox": [ + 58, + 407, + 300, + 517 + ], + "spans": [ + { + "bbox": [ + 58, + 407, + 300, + 517 + ], + "type": "table", + "html": "
MethodFID ↓Clip Score ↑Latency (s) ↓
Original18.4230.802.930
DeepCache [25]17.7930.751.737
TGATE [22]19.5129.551.992
ToDo [35]18.2830.752.520
ToMe [1]17.4330.552.730
Faster Diffusion [20]17.9030.402.005
CA-ToMe [33]20.4930.412.112
Ours16.6530.251.139
", + "image_path": "5545c291e943d864851e04b7180c88d34274ed3ffe347e96b030a64b1243ebe7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 531, + 295, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 531, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 54, + 531, + 295, + 639 + ], + "type": "text", + "content": "for DeepCache is set to 2. For TGATE, we set " + }, + { + "bbox": [ + 54, + 531, + 295, + 639 + ], + "type": "inline_equation", + "content": "m = 5" + }, + { + "bbox": [ + 54, + 531, + 295, + 639 + ], + "type": "text", + "content": ". For ToMe, we use a merging ratio of 0.3, while ToDo, Faster Diffusion, and CA-ToMe are evaluated using their default hyperparameters from their respective papers. For the transformer-based model PixArt-" + }, + { + "bbox": [ + 54, + 531, + 295, + 639 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 54, + 531, + 295, + 639 + ], + "type": "text", + "content": ", we compare Post-Diff with TGATE [22] using " + }, + { + "bbox": [ + 54, + 531, + 295, + 639 + ], + "type": "inline_equation", + "content": "m = 8" + }, + { + "bbox": [ + 54, + 531, + 295, + 639 + ], + "type": "text", + "content": ", as well as DiTFastAttn [46] and DuCa [49] with their default settings on an NVIDIA H100 GPU, since some methods encounter out-of-memory errors on the A5000." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 641, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 641, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 641, + 295, + 714 + ], + "type": "text", + "content": "Observations. Compared to previous methods that accelerate the generation process either at the module level (e.g., DeepCache, TGATE, DuCa) by reusing computations across denoising steps or at the input level (e.g., ToMe, ToDo) by compressing redundant image tokens, our method uniquely integrates both approaches, resulting in a greater" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 315, + 418, + 553, + 508 + ], + "blocks": [ + { + "bbox": [ + 313, + 387, + 555, + 408 + ], + "lines": [ + { + "bbox": [ + 313, + 387, + 555, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 555, + 408 + ], + "type": "text", + "content": "Table 4. Benchmark with other training-free compression methods on PixArt-α." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 418, + 553, + 508 + ], + "lines": [ + { + "bbox": [ + 315, + 418, + 553, + 508 + ], + "spans": [ + { + "bbox": [ + 315, + 418, + 553, + 508 + ], + "type": "table", + "html": "
MethodFID ↓Clip Score ↑Latency (s) ↓
Original29.1630.411.752
TGATE [22]29.5429.591.275
DiTFastAttn [46]45.4630.13\\( 1.103^{\\dagger} \\)
DuCa [49]28.0830.121.724
Ours25.4430.231.382
", + "image_path": "6a2319957becacf4223674363d7e89a7677ca875e9da5082e9bdc1089f72787a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 321, + 508, + 478, + 518 + ], + "lines": [ + { + "bbox": [ + 321, + 508, + 478, + 518 + ], + "spans": [ + { + "bbox": [ + 321, + 508, + 478, + 518 + ], + "type": "text", + "content": "Faster speed is achieved with FlashAttention [6]." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 532, + 555, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 555, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 555, + 590 + ], + "type": "text", + "content": "reduction in inference cost. Additionally, leveraging low-resolution images in the early stages enhances the denoising process, allowing our method to achieve the lowest FID with a competitive Clip Score across both SD V1.5 and PixArt-" + }, + { + "bbox": [ + 313, + 532, + 555, + 590 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 532, + 555, + 590 + ], + "type": "text", + "content": " models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 593, + 555, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 593, + 555, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 593, + 555, + 652 + ], + "type": "text", + "content": "These experiments highlight the benefits of leveraging redundancy at both the input and module levels. Integrating the aforementioned training-free compression techniques into PostDiff could further enhance efficiency, which we leave for future work." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 668, + 401, + 680 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 668, + 401, + 680 + ], + "spans": [ + { + "bbox": [ + 313, + 668, + 401, + 680 + ], + "type": "text", + "content": "4. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "Diffusion models. Diffusion models have emerged as a powerful approach for generative tasks, modeling the gen" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3007" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 70, + 548, + 346 + ], + "blocks": [ + { + "bbox": [ + 71, + 70, + 548, + 346 + ], + "lines": [ + { + "bbox": [ + 71, + 70, + 548, + 346 + ], + "spans": [ + { + "bbox": [ + 71, + 70, + 548, + 346 + ], + "type": "image", + "image_path": "315f9eb9be4ad714b545b7392630c906b8d3116c20b6812209d8b735fd140a07.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 354, + 472, + 365 + ], + "lines": [ + { + "bbox": [ + 137, + 354, + 472, + 365 + ], + "spans": [ + { + "bbox": [ + 137, + 354, + 472, + 365 + ], + "type": "text", + "content": "Figure 6. Visualize the generation results of different models w/o and w/ using our PostDiff." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 376, + 295, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 376, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 54, + 376, + 295, + 508 + ], + "type": "text", + "content": "erative process as a denoising sequence to enable high-quality image generation [15, 36]. Latent Diffusion Models (LDMs) [31] leverage Variational Autoencoders (VAEs) [30] to operate in a compressed latent space, reducing memory demands and enabling efficient high-resolution generation. Open-source models like Stable Diffusion V1.5 [31], Stable Diffusion XL [29], and PixArt [4, 5] have further expanded accessibility and are now applied across multiple domains [8, 10, 11]. Despite their impressive generative quality, diffusion models often suffer from long generation times, limiting their broader applicability." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 509, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 509, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 54, + 509, + 295, + 652 + ], + "type": "text", + "content": "Few-Step diffusion models. A primary focus in accelerating diffusion models is reducing the number of timesteps. DDIM [36] modifies DDPM [15] by introducing non-Markovian sampling, enabling faster generation. DPM-Solver [23] applies exponential integrators to approximate solutions to diffusion ODEs, enhancing sampling efficiency. Other approaches, such as consistency models [24, 37], aim to map points on the ODE trajectory to the original sample, effectively reducing inference steps. Adversarial training is also employed to enable one-step or few-step diffusion [43-45]. However, these few-step models still incur non-trivial training/distillation cost." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 653, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 653, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 653, + 295, + 715 + ], + "type": "text", + "content": "Efficient diffusion models. Efforts to accelerate diffusion models extend from data optimization to model optimization. From the data perspective, prior works have reduced computational requirements by merging similar tokens [1] or pruning redundant tokens [39]. From the model" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 376, + 554, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 554, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 554, + 473 + ], + "type": "text", + "content": "perspective, recent works focus on identifying redundant feature maps across timesteps and using caching mechanisms to reuse these maps, significantly reducing computational costs [25, 42]. Additional approaches aim to slim down models through quantization to lower bit precision [12, 34, 48] or by pruning unnecessary parameters [2, 9, 41]. Additionally, several systems have been developed specifically to accelerate diffusion model inference [18, 19]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 488, + 388, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 488, + 388, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 488, + 388, + 501 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 312, + 510, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 510, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 510, + 555, + 715 + ], + "type": "text", + "content": "In this work, we explore a critical question for compute-optimal diffusion model deployment: under a post-training setting, is it more effective to reduce denoising steps or decrease per-step inference cost? To investigate this, we propose PostDiff, a training-free framework that enhances diffusion model efficiency while systematically examining this trade-off. PostDiff reduces redundancy at two levels: at the input level, through a mixed-resolution denoising scheme that enhances low-frequency components in early steps, and at the module level, with a hybrid caching strategy. Extensive evaluations demonstrate that reducing per-step inference cost often achieves a better efficiency-fidelity balance than reducing denoising steps across a wide efficiency range. PostDiff not only offers a new perspective on the fidelity-efficiency trade-off for diffusion models but also provides practical insights for their deployment on resource-constrained platforms." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3008" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "type": "text", + "content": "This work is supported by the National Science Foundation (NSF) through CCRI funding (Award Number: 2016727), RTML funding (Award Number: 1937592), and an NSF CAREER award (Award Number: 2048183), as well as by CoCoSys, one of the seven centers in JUMP 2.0, a Semiconductor Research Corporation (SRC) program sponsored by DARPA." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 186, + 115, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 115, + 198 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 115, + 198 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 205, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 205, + 295, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 205, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 61, + 205, + 295, + 228 + ], + "type": "text", + "content": "[1] Daniel Bolya and Judy Hoffman. Token merging for fast stable diffusion. In CVPR, 2023. 3, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 229, + 296, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 229, + 296, + 272 + ], + "spans": [ + { + "bbox": [ + 61, + 229, + 296, + 272 + ], + "type": "text", + "content": "[2] Thibault Castells, Hyoung-Kyu Song, Bo-Kyeong Kim, and Shinkook Choi. LD-Pruner: Efficient pruning of latent diffusion models using task-agnostic insights. In CVPR, 2024. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 274, + 295, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 274, + 295, + 317 + ], + "spans": [ + { + "bbox": [ + 61, + 274, + 295, + 317 + ], + "type": "text", + "content": "[3] Angela Castillo, Jonas Kohler, Juan C Pérez, Juan Pablo Pérez, Albert Pumarola, Bernard Ghanem, Pablo Arbeláez, and Ali Thabet. Adaptive guidance: Training-free acceleration of conditional diffusion models. In AAAI, 2025. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 319, + 296, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 319, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 62, + 319, + 296, + 373 + ], + "type": "text", + "content": "[4] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. PixArt- " + }, + { + "bbox": [ + 62, + 319, + 296, + 373 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 62, + 319, + 296, + 373 + ], + "type": "text", + "content": " : Weak-to-strong training of diffusion transformer for 4k text-to-image generation. In ECCV, 2024. 2, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 374, + 296, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 374, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 374, + 296, + 430 + ], + "type": "text", + "content": "[5] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Zhongdao Wang, James T. Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. PixArt-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In ICLR, 2024. 5, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 431, + 295, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 431, + 295, + 463 + ], + "spans": [ + { + "bbox": [ + 62, + 431, + 295, + 463 + ], + "type": "text", + "content": "[6] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. FlashAttention: Fast and memory-efficient exact attention with io-awareness. In NeurIPS, 2022. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 464, + 295, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 464, + 295, + 486 + ], + "spans": [ + { + "bbox": [ + 62, + 464, + 295, + 486 + ], + "type": "text", + "content": "[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 487, + 296, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 487, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 62, + 487, + 296, + 531 + ], + "type": "text", + "content": "[8] Zhenbang Du, Wei Feng, Haohan Wang, Yaoyu Li, Jingsen Wang, Jian Li, Zheng Zhang, Jingjing Lv, Xin Zhu, Junsheng Jin, et al. Towards reliable advertising image generation using human feedback. In ECCV, 2024. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 533, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 533, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 533, + 295, + 555 + ], + "type": "text", + "content": "[9] Gongfan Fang, Xinyin Ma, and Xinchao Wang. Structural pruning for diffusion models. In NeurIPS, 2023. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 555, + 295, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 555, + 295, + 599 + ], + "spans": [ + { + "bbox": [ + 57, + 555, + 295, + 599 + ], + "type": "text", + "content": "[10] Sicheng Gao, Xuhui Liu, Bohan Zeng, Sheng Xu, Yanjing Li, Xiaoyan Luo, Jianzhuang Liu, Xiantong Zhen, and Baochang Zhang. Implicit diffusion models for continuous super-resolution. In CVPR, 2023. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 600, + 295, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 600, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 57, + 600, + 295, + 622 + ], + "type": "text", + "content": "[11] He Guo, Zixuan Ye, Zhiguo Cao, and Hao Lu. In-context matting. In CVPR, 2024. 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 624, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 656 + ], + "type": "text", + "content": "[12] Yefei He, Luping Liu, Jing Liu, Weijia Wu, Hong Zhou, and Bohan Zhuang. PTQD: Accurate post-training quantization for diffusion models. In NeurIPS, 2023. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 658, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 689 + ], + "type": "text", + "content": "[13] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. In EMNLP, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[14] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshop, 2021. 5" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "text", + "content": "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 3, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 96, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 129 + ], + "type": "text", + "content": "[16] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. JMLR, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 130, + 553, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 163 + ], + "type": "text", + "content": "[17] Bo-Kyeong Kim, Hyoung-Kyu Song, Thibault Castells, and Shinkook Choi. BK-SDM: A lightweight, fast, and cheap version of stable diffusion. In ECCV, 2024. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 164, + 553, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 164, + 553, + 219 + ], + "spans": [ + { + "bbox": [ + 317, + 164, + 553, + 219 + ], + "type": "text", + "content": "[18] Akio Kodaira, Chenfeng Xu, Toshiki Hazama, Takanori Yoshimoto, Kohei Ohno, Shogo Mitsuhori, Soichi Sugano, Hanying Cho, Zhijian Liu, and Kurt Keutzer. Streamdiffusion: A pipeline-level solution for real-time interactive generation. arXiv preprint, 2023. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "type": "text", + "content": "[19] Muyang Li, Tianle Cai, Jiaxin Cao, Qinsheng Zhang, Han Cai, Junjie Bai, Yangqing Jia, Kai Li, and Song Han. Distribution: Distributed parallel inference for high-resolution diffusion models. In CVPR, 2024. 2, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "text", + "content": "[20] Senmao Li, Taihang Hu, Fahad Shahbaz Khan, Linxuan Li, Shiqi Yang, Yaxing Wang, Ming-Ming Cheng, and Jian Yang. Faster diffusion: Rethinking the role of unet encoder in diffusion models. In NeurIPS, 2024. 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 310, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 553, + 354 + ], + "type": "text", + "content": "[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 4, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 355, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 553, + 409 + ], + "type": "text", + "content": "[22] Haozhe Liu, Wentian Zhang, Jinheng Xie, Francesco Faccio, Mengmeng Xu, Tao Xiang, Mike Zheng Shou, Juan-Manuel Pérez-Rúa, and Jurgen Schmidhuber. Faster diffusion through temporal attention decomposition. TMLR, 2025. 2, 5, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 411, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 453 + ], + "type": "text", + "content": "[23] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. NeurIPS, 2022. 3, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 456, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 498 + ], + "type": "text", + "content": "[24] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint, 2023. 5, 6, 7, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 501, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 553, + 533 + ], + "type": "text", + "content": "[25] Xinyin Ma, Gongfan Fang, and Xinchao Wang. Deepcache: Accelerating diffusion models for free. In CVPR, 2024. 5, 6, 7, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 535, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 535, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 535, + 553, + 578 + ], + "type": "text", + "content": "[26] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. In CVPR, 2023. 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 580, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 580, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 580, + 553, + 633 + ], + "type": "text", + "content": "[27] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2021. 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 315, + 635, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 635, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 315, + 635, + 553, + 657 + ], + "type": "text", + "content": "[28] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[29] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3009" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "text", + "content": "[30] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. In NeurIPS, 2019. 3, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "text", + "content": "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 153, + 294, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 153, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 153, + 294, + 185 + ], + "type": "text", + "content": "[32] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 188, + 294, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 188, + 294, + 232 + ], + "spans": [ + { + "bbox": [ + 56, + 188, + 294, + 232 + ], + "type": "text", + "content": "[33] Omid Saghatchian, Atiyeh Gh Moghadam, and Ahmad Nickabadi. Cached adaptive token merging: Dynamic token reduction and redundant computation elimination in diffusion model. arXiv preprint, 2025. 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 234, + 294, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 234, + 294, + 266 + ], + "spans": [ + { + "bbox": [ + 56, + 234, + 294, + 266 + ], + "type": "text", + "content": "[34] Yuzhang Shang, Zhihang Yuan, Bin Xie, Bingzhe Wu, and Yan Yan. Post-training quantization on diffusion models. In CVPR, 2023. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 268, + 294, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 268, + 294, + 300 + ], + "spans": [ + { + "bbox": [ + 56, + 268, + 294, + 300 + ], + "type": "text", + "content": "[35] Ethan Smith, Nayan Saxena, and Aninda Saha. Todo: Token downsampling for efficient generation of high-resolution images. In *IJCAI*, 2024. 3, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 303, + 294, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 303, + 294, + 324 + ], + "spans": [ + { + "bbox": [ + 56, + 303, + 294, + 324 + ], + "type": "text", + "content": "[36] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 3, 5, 6, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 327, + 294, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 348 + ], + "type": "text", + "content": "[37] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In ICML, 2023. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 350, + 294, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 294, + 393 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 294, + 393 + ], + "type": "text", + "content": "[38] Jiayan Teng, Wendi Zheng, Ming Ding, Wenyi Hong, Jianqiao Wangni, Zhuoyi Yang, and Jie Tang. Relay diffusion: Unifying diffusion process across resolutions for image synthesis. In ICLR, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 396, + 294, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 396, + 294, + 438 + ], + "spans": [ + { + "bbox": [ + 56, + 396, + 294, + 438 + ], + "type": "text", + "content": "[39] Hongjie Wang, Difan Liu, Yan Kang, Yijun Li, Zhe Lin, Niraj K Jha, and Yuchen Liu. Attention-driven training-free efficiency enhancement of diffusion models. In CVPR, 2024. 3, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 441, + 294, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 441, + 294, + 484 + ], + "spans": [ + { + "bbox": [ + 56, + 441, + 294, + 484 + ], + "type": "text", + "content": "[40] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. LAVIE: High-quality video generation with cascaded latent diffusion models. IJCV, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 487, + 294, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 487, + 294, + 530 + ], + "spans": [ + { + "bbox": [ + 56, + 487, + 294, + 530 + ], + "type": "text", + "content": "[41] Lexington Whalen, Zhenbang Du, Haoran You, Chaojian Li, Sixu Li, and Yingyan (Celine) Lin. Early-Bird Diffusion: Investigating and leveraging timestep-aware early-bird tickets in diffusion models for efficient training. In CVPR, 2025. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 533, + 294, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 533, + 294, + 586 + ], + "spans": [ + { + "bbox": [ + 56, + 533, + 294, + 586 + ], + "type": "text", + "content": "[42] Felix Wimbauer, Bichen Wu, Edgar Schoenfeld, Xiaoliang Dai, Ji Hou, Zijian He, Artsiom Sanakoyeu, Peizhao Zhang, Sam Tsai, Jonas Kohler, et al. Cache me if you can: Accelerating diffusion models through block caching. In CVPR, 2024. 5, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 589, + 294, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 589, + 294, + 621 + ], + "spans": [ + { + "bbox": [ + 56, + 589, + 294, + 621 + ], + "type": "text", + "content": "[43] Yanwu Xu, Yang Zhao, Zhisheng Xiao, and Tingbo Hou. UFOGen: You forward once large scale text-to-image generation via diffusion gans. In CVPR, 2024. 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 624, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 666 + ], + "type": "text", + "content": "[44] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved distribution matching distillation for fast image synthesis. In NeurIPS, 2024." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[45] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Frédo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In CVPR, 2024. 8" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 553, + 274 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 315, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 553, + 116 + ], + "type": "text", + "content": "[46] Zhihang Yuan, Hanling Zhang, Lu Pu, Xuefei Ning, Linfeng Zhang, Tianchen Zhao, Shengen Yan, Guohao Dai, and Yu Wang. DiTFastattn: Attention compression for diffusion transformer models. In NeurIPS, 2024. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 315, + 118, + 553, + 161 + ], + "type": "text", + "content": "[47] Shen Zhang, Zhaowei Chen, Zhenyu Zhao, Yuhao Chen, Yao Tang, and Jiajun Liang. HiDiffusion: Unlocking higher-resolution creativity and efficiency in pretrained diffusion models. In ECCV, 2025. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 163, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 163, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 315, + 163, + 553, + 227 + ], + "type": "text", + "content": "[48] Tianchen Zhao, Tongcheng Fang, Enshu Liu, Rui Wan, Widyadewi Soedarmadji, Shiyao Li, Zinan Lin, Guohao Dai, Shengen Yan, Huazhong Yang, Xuefei Ning, and Yu Wang. ViDiT-Q: Efficient and accurate quantization of diffusion transformers for image and video generation. In ICLR, 2025. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 229, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 229, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 315, + 229, + 553, + 274 + ], + "type": "text", + "content": "[49] Chang Zou, Evelyn Zhang, Runlin Guo, Haohang Xu, Conghui He, Xuming Hu, and Linfeng Zhang. Accelerating diffusion transformers with dual feature caching. arXiv preprint, 2024. 7" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3010" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/FiVE-Bench_ A Fine-grained Video Editing Benchmark for Evaluating Emerging Diffusion and Rectified Flow Models/61b80164-0e5a-42b8-b65d-d1598a219bbd_content_list.json b/2025/FiVE-Bench_ A Fine-grained Video Editing Benchmark for Evaluating Emerging Diffusion and Rectified Flow Models/61b80164-0e5a-42b8-b65d-d1598a219bbd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..95c718da746777e5d7d25c85c9742f4e61ec5016 --- /dev/null +++ b/2025/FiVE-Bench_ A Fine-grained Video Editing Benchmark for Evaluating Emerging Diffusion and Rectified Flow Models/61b80164-0e5a-42b8-b65d-d1598a219bbd_content_list.json @@ -0,0 +1,1633 @@ +[ + { + "type": "text", + "text": "FiVE-Bench: A Fine-grained Video Editing Benchmark for Evaluating Emerging Diffusion and Rectified Flow Models", + "text_level": 1, + "bbox": [ + 176, + 123, + 821, + 189 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Minghan Li $^{1*}$ , Chenxi Xie $^{2*}$ , Yichen Wu $^{1,3}$ , Lei Zhang $^{2}$ and Mengyu Wang $^{1\\dagger}$ $^{1}$ Harvard University", + "bbox": [ + 194, + 214, + 799, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Hong Kong Polytechnic University, $^{3}$ City University of Hong Kong", + "bbox": [ + 228, + 252, + 766, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{mili4, mengyu_wang}@meei.harvard.edu", + "bbox": [ + 333, + 271, + 656, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "FiVE-Dataset", + "text_level": 1, + "bbox": [ + 189, + 295, + 267, + 306 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/824625187b3e53d65c739fdb8bb3993b55521516edb3e65a20d335514534c250.jpg", + "image_caption": [ + "Figure 1. Overview of the proposed FiVE-Bench, with editing results from Wan-Edit and corresponding FiVE-Acc scores." + ], + "image_footnote": [], + "bbox": [ + 94, + 309, + 372, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wan-Edit Results and FiVE-Evaluation", + "text_level": 1, + "bbox": [ + 544, + 295, + 754, + 306 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0ef43fd5061b1234cc2b78d174f485e754cf24259bd6c7c50c28b0caace1be8a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 380, + 309, + 901, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 599, + 326, + 614 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Numerous text-to-video (T2V) editing methods have emerged recently, but the lack of a standardized benchmark for fair evaluation has led to inconsistent claims and an inability to assess model sensitivity to hyperparameters. Fine-grained video editing is crucial for enabling precise, object-level modifications while maintaining context and temporal consistency. To address this, we introduce FiVE-Bench, a Fine-grained Video Editing Benchmark for evaluating emerging diffusion and rectified flow models. Our benchmark includes 74 real-world videos and 26 generated videos, featuring 6 fine-grained editing types, 420 object-level editing prompt pairs, and their corresponding masks. Additionally, we adapt the latest rectified flow (RF) T2V generation models—Pyramid-Flow [14] and Wan2.1 [40]—by introducing FlowEdit [18], resulting in training-free and inversion-free video editing mod", + "bbox": [ + 86, + 632, + 485, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "els Pyramid-Edit and Wan-Edit. We compare five diffusion methods with our two RF methods on the proposed FiVE-Bench, evaluating them across 15 metrics. These metrics include background preservation, text-video similarity, temporal consistency, and generated video quality. To further enhance object-level evaluation, we introduce FiVE-Acc, a novel metric leveraging Vision-Language Models (VLMs) to assess the success of fine-grained video editing. Experimental results demonstrate that RF-based editing significantly outperforms diffusion-based methods, with WanEdit achieving the best overall performance and exhibiting the least sensitivity to hyperparameters. More video demo available on the website: https://sites.google.com/view/five-benchmark.", + "bbox": [ + 511, + 601, + 906, + 811 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 844, + 645, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in text-to-image (T2I) and text-to-video (T2V) generation [1, 14, 24, 26, 31, 36, 40] have", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution, † Corresponding author.", + "bbox": [ + 107, + 886, + 349, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16672", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "led to the emergence of numerous image and video editing methods [4, 9, 10, 46], enabling users to create and modify video content with unprecedented flexibility. These methods leverage powerful generative models, such as diffusion models [6, 11, 12, 35] and Rectified Flow (RF) [8, 23] models, to synthesize and edit videos based on textual prompts.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, despite significant progress in these fields, the lack of a standardized benchmark for fair and comprehensive evaluation has still become a major bottleneck hindering further development. Currently, the video editing field has only one publicly available dataset, TGVE [47] /TGVE+ [34] (see Table 1), which includes 76 videos and 6 editing types, such as style transfer, background replacement, single-object and multi-object transformations, as well as object addition and removal. While TGVE+ aims to cover as many editing types as possible, its evaluation within each editing type is not comprehensive, particularly in supporting fine-grained video editing. Fine-grained video editing requires precise object-level modifications while preserving the overall context and temporal consistency of the video, posing higher demands on model performance.", + "bbox": [ + 89, + 184, + 482, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Due to the lack of a standardized benchmark and unified evaluation framework for fine-grained video editing, existing methods [9, 21, 28, 50] often rely on self-collected small-scale video datasets to validate their effectiveness. This approach not only makes it difficult to objectively compare the performance of different methods but also fails to comprehensively assess model robustness to hyperparameters or determine their suitability for real-world applications. Therefore, building a comprehensive and challenging benchmark for fine-grained video editing has become an urgent need to advance the field.", + "bbox": [ + 89, + 412, + 482, + 578 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we introduce FiVE-Bench (Fine-grained Video Editing Benchmark), a comprehensive benchmark designed to evaluate emerging diffusion and rectified flow models for fine-grained video editing. FiVE-Bench includes 74 real-world videos and 26 generated videos, covering a diverse range of scenes and editing scenarios. It features 6 fine-grained editing types, 420 object-level editing prompt pairs, and their corresponding masks, providing a rich and challenging benchmark for video editing. Additionally, we propose FiVE-Acc, a novel evaluation metric that leverages Vision-Language Models (VLMs) to measure the success rate of fine-grained object-level editing. Unlike traditional similarity-based scores, FiVE-Acc focuses on evaluating whether the intended edit was successfully achieved, with particular emphasis on semantic alignment of the edited object.", + "bbox": [ + 89, + 580, + 482, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, many image editing methods [5, 18, 32, 41] based on the latest rectified flow T2I generation model Flux [19] have achieved state-of-the-art editing results and fidelity, such as RF-Inversion [32], RF-Solver [41], and FlowEdit [18]. However, to the best of our knowledge,", + "bbox": [ + 89, + 824, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "there are no video editing methods based on RF T2V generation models, primarily due to the relatively slow development of T2V generation models. Fortunately, the recent release of two state-of-the-art RF-based T2V generation models—Pyramid-Flow [14] and Wan2.1 [40]—has opened new possibilities for RF video editing.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building on the latest advancements, we adapt Pyramid-Flow and Wan2.1 by introducing FlowEdit [18]. This adaptation facilitates training-free and inversion-free video editing models, Pyramid-Edit and Wan-Edit, which leverage the strong temporal consistency inherent in video generation models without requiring additional attention mechanisms. Specifically, Pyramid-Flow employs a multi-resolution and temporally autoregressive model architecture across time steps. To adapt to this unique design, we modified FlowEdit to support multi-resolution processing, enabling seamless integration with the Pyramid-Flow framework, leading to Pyramid-Edit. On the other hand, Wan2.1 utilizes the MMDiT architecture, which unfolds temporal frames into sequences and processes them via self-attention. Since its design is fundamentally similar to image generation models, Wan2.1 can directly adopt FlowEdit without any additional modifications, giving rise to Wan-Edit.", + "bbox": [ + 511, + 183, + 903, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To show the effectiveness of the proposed RF-based editing methods, we compare them with five diffusion-based editing methods on the built FiVE-Bench, evaluating them across fifteen metrics, including background preservation, edited text-video similarity, temporal consistency, and generated video quality. Our experiments demonstrate that RF-based editing methods significantly outperform diffusion-based methods across multiple metrics. Notably, Wan-Edit achieves the best overall performance, exhibiting superior editing quality and the least sensitivity to hyperparameters. These findings highlight the potential of RF-based models for fine-grained video editing and provide valuable insights for future research in this area.", + "bbox": [ + 511, + 440, + 903, + 635 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, this work makes four key contributions:", + "bbox": [ + 535, + 638, + 880, + 652 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce FiVE-Bench, a comprehensive benchmark for fine-grained video editing, featuring diverse videos, editing types, and evaluation metrics.", + "- We propose FiVE-Acc, a novel metric leveraging VLMs to assess the success of fine-grained editing.", + "- We propose the first two RF-based T2V editing models: Pyramid-Edit and Wan-Edit.", + "- We evaluate 7 video editing methods across 15 metrics, showing that RF-based editing consistently outperforms diffusion-based approaches." + ], + "bbox": [ + 513, + 652, + 903, + 804 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 821, + 653, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Diffusion and RF Inversion", + "text_level": 1, + "bbox": [ + 513, + 847, + 758, + 861 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion inversion. Diffusion models (DMs) drive advances in inversion techniques for better control over gen", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16673", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/d3f7bf5afa3ed7715ed2242eda301ebf756e753978f74b45840d056e1c475f49.jpg", + "table_caption": [ + "Table 1. Statistics of existing video editing datasets and benchmarks." + ], + "table_footnote": [], + "table_body": "
DatasetUsageNum. ImagesNum. VideosFrames Per VideoNum. PromptsGen. VideosSrc. obj. WordsEdit obj. WordsObj. Instruct.Src. obj. MasksType of Edited Prompts
VIVID-10MTrain672K73.7K3010MXXX✓✓object, add, remove
Señorita-2MTrain0388.9K33~642MXXX✓Xstyle, object, multiple, color, add, remove, motion
TGVEEval07632304XXXXXstyle, object, bg, multiple
TGVE+Eval076321417XXXXXstyle, object, bg, multiple, color, add, remove
VIVID-10MEval06433~64852XXX✓✓object, add, remove
FiVE (Ours)Eval010035 ~ 126420✓✓✓✓✓object (w&w/o non-rigid), color, material, add, remove
", + "bbox": [ + 99, + 107, + 903, + 219 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "erated samples. DDIM [35] accelerates sampling but struggles with fine-grained reconstructions due to nonlinearities and score estimation errors. To improve accuracy, fine-tuning and optimization-based methods [7, 15, 25, 39, 45] have been proposed. While these advancements significantly enhance the fidelity and accuracy of DDIM inversion, the associated computational costs and resource demands remain substantial limiting factors. Although diffusion-based methods have achieved outstanding performance, recent studies [8, 14, 19, 32] suggest that Rectified Flow models (RFs) hold significant potential to surpass DMs in certain applications. Notably, the inversion and editing capabilities of RF models have been explored in image editing [5, 18, 32, 41, 48], demonstrating their efficiency and effectiveness. Building on these strengths, this work extends their potential in video editing.", + "bbox": [ + 88, + 244, + 485, + 487 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Video Editing", + "text_level": 1, + "bbox": [ + 89, + 500, + 235, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video editing techniques can generally be categorized into three types based on their training: training-free methods, one/few-shot finetuned methods [37, 43, 46, 55], and massive data finetuned methods. Training-free models [17, 21, 28, 50], enable task execution without retraining. TokenFlow [9] utilizes pre-trained models for efficient video synthesis, while DMT [51] transfers motion via diffusion. Other models like Pixel2Video [3], Render-A-Video [49], and Text2Video-zero [16] enhance accessibility for real-time video editing. One/few-shot fine-tuned models adapt pre-trained models for video editing with minimal fine-tuning. Methods like Tune-A-Video [46] and MotionEditor [37] achieve strong results using limited annotated samples, balancing customization and efficiency for task-specific editing without extensive retraining.", + "bbox": [ + 91, + 522, + 482, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Massive data fine-tuned models undergo extensive training on large datasets for high-quality, versatile video editing [29]. InstructVid2Vid [29] enables complex edits via natural language instructions, while EffiVED [53] refines broad datasets into high-quality subsets for efficient editing. SF-V [54] introduces adversarial training for video synthesis. Despite their precision and adaptability, these methods are computationally expensive, limiting practicality in resource-constrained settings. These methods focus on learning video motion, which may compromise back", + "bbox": [ + 89, + 750, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ground preservation. Recently, VideoGrain [50] has emphasized multi-granularity video editing rather than fine-grained edits, aiming to retain background details while modifying target objects. This work aims to address this issue and advance the field.", + "bbox": [ + 511, + 244, + 905, + 321 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. FiVE Benchmark and FiVE-Acc Metric", + "text_level": 1, + "bbox": [ + 511, + 335, + 870, + 352 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. FiVE Benchmark", + "text_level": 1, + "bbox": [ + 511, + 362, + 684, + 377 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We present the FiVE benchmark, which includes the FiVE video dataset and six fine-grained editing tasks. Additional details can be found in the Appendix.", + "bbox": [ + 511, + 385, + 905, + 430 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Videos in FiVE-Bench. We meticulously curate 74 real-world videos suitable for fine-grained editing from the DAVIS [27] dataset. Consecutive frames are extracted at every 8-frame interval, and the GPT-4o [13] is employed to generate formatted captions for these videos. These captions encompass details pertaining to object category, action, background, and camera movement. Additionally, during the annotation process, we document instances of object deformation (such as limb movements in humans or animals), which aids in differentiating the complexity of editing tasks during evaluation. To further enhance the diversity of our benchmark, we generate 26 highly realistic synthetic videos using the T2V model [40]. These synthetic videos not only expand the range of video categories but also enable a comparative analysis of editing performance across real and synthetic videos.", + "bbox": [ + 511, + 431, + 906, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Six Fine-grained Video Editing Tasks. FiVE-Bench comprises six fine-grained video editing tasks, totaling 420 high-quality editing prompt pairs. It primarily includes four object-targeted editing tasks of increasing complexity: object substitution without non-rigid deformation, object substitution with non-rigid deformation, color alteration, and material modification. For each task, we use GPT-4o [13] to generate four source-target prompt pairs per video in the benchmark, resulting in 400 high-quality editing pairs. To further enhance the diversity of editing types within the benchmark, we select 10 videos each for the add and remove tasks and design corresponding source-target prompt pairs. Additionally, we explicitly provide both object-related source words and edited words in the source and edited text prompts, respectively, denoted as", + "bbox": [ + 509, + 674, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16674", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/64eeee1923f5e9435008803cc18b46ffb3f030ad222c11dd33d8538f81b93101.jpg", + "image_caption": [ + "Figure 2. The VLM-based FiVE accuracy (FiVE-Acc) evaluation pipeline." + ], + "image_footnote": [], + "bbox": [ + 98, + 89, + 903, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "\"src. obj. words\" and \"edit obj. words\" in Table 1. To accommodate different editing methods, we use GPT-4o [13] to generate instruction prompts (\"Obj. Instr.\" in Table 1) for each source-target pair, ensuring compatibility with models like InstructPix2Pix [2]. For each video, we employ SAM2 [20, 30] to generate masks of the edited regions, enabling the evaluation of background preservation metrics.", + "bbox": [ + 88, + 282, + 482, + 390 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. FiVE-Acc Metric", + "text_level": 1, + "bbox": [ + 89, + 402, + 261, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To evaluate the accuracy of fine-grained video editing, we propose FiVE accuracy (FiVE-Acc) evaluation, a framework assessing how precisely video editing models modify the target object. Fig. 2 presents the complete pipeline, comprising video editing and accuracy evaluation.", + "bbox": [ + 89, + 426, + 483, + 502 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Video editing outputs. Given a source video, a video editing model is prompted with a source prompt, describing the initial scene, and a target prompt, specifying the desired modification. The model generates a target video, where the intended transformation is applied, as shown in Fig. 2a.", + "bbox": [ + 89, + 503, + 483, + 579 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "FiVE-Acc evaluation. To quantitatively evaluate the fidelity of video editing, we employ a Vision-Language Model (VLM) (e.g., Qwen-2.5-VL) to analyze the edited video. The evaluation includes two types of questions:", + "bbox": [ + 89, + 580, + 483, + 641 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Yes/No Questions: The model is asked whether the source and target objects are present in the target video. Editing succeeds only if the model answers \"No\" for the source object and \"Yes\" for the target, leading to the calculation of $YN$ -Acc. The source object question ensures the model both removes the source and adds the target, refining accuracy evaluation.", + "- Multi-choice Question: It evaluates whether the VLM recognizes the target (e.g., flamingo) or source object (e.g., black swan) in the edited video. The model selects between the two objects to compute MC-Acc, measuring its recognition accuracy." + ], + "bbox": [ + 89, + 642, + 482, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the two obtained accuracies, we further calculate the union and intersection accuracies to derive $\\cup$ -Acc and $\\cap$ -Acc. These offer a more detailed evaluation: the former reflects the model's overall editing success accuracy, while the latter highlights the model's high-quality editing", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "success accuracy. $FiVE - Acc$ is the average of four components: $FiVE - Acc = 0.25 \\times (YN - Acc + MC - Acc + \\cup - Acc + \\cap - Acc)$ . It uses the VLM's recognition ability to evaluate the alignment between the edited object and its real data distribution. To the best of our knowledge, we are the first to introduce accuracy evaluation in fine-grained T2V editing.", + "bbox": [ + 511, + 282, + 906, + 376 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Methods: Pyramid-Edit and Wan-Edit", + "text_level": 1, + "bbox": [ + 511, + 388, + 862, + 406 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Preliminary", + "text_level": 1, + "bbox": [ + 511, + 412, + 643, + 430 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Rectified Flow (RF) Models [22, 23]. Let $q_{0}$ denote the source distribution (i.e., standard Gaussian $\\mathcal{N}(0,I)$ ), $p_{0}$ refer to the target distribution (i.e., the distribution over images), and $v_{t}(\\cdot)$ is the time-varying vector field. RF models [22, 23] aim to transform $q_{0}$ progressively into $p_{0}$ by solving the following ordinary differential equation (ODE):", + "bbox": [ + 511, + 435, + 905, + 527 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nd X _ {t} = v _ {t} \\left(X _ {t}\\right) d t, \\quad t \\in [ 0, 1 ]. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 532, + 903, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, the vector field $v_{t}(X_{t}) = f(X_{t},c,1 - t;\\varphi)$ , where $f(\\cdot)$ is a neural network parameterized by $\\varphi$ and $c$ is the guided text. Given the ODE in Eq. (1), we start with $X_0\\sim q_0$ from the source distribution and integrate over $t:0\\to 1$ to reach the endpoint $X_{1}\\sim p_{0}$ , which is sampled from the target distribution. The forward process of RF assumes the linear path between the two states: $X_{t} = (1 - t)X_{0} + tX_{1}$ , then we can derive the vector field of $v_{t}(\\cdot):u_{t}(X_{t}) = X_{1} - X_{0}$ . Then, $v_{t}(X_{t})$ is used to approximate the network $f(X_{t},c,t;\\varphi)$ by minimizing the loss function $\\mathcal{L}_{\\varphi} = \\mathbb{E}_{t\\sim \\mathcal{U}[0,1],Y_t}[\\| v_t(X_t) - f(X_t,c,t;\\varphi)\\| _2^2 ]$ . RF ensures that their sampling paths are relatively straight, enabling the use of a small number of discretization steps.", + "bbox": [ + 511, + 553, + 906, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "RF-based Image Editing: FlowEdit [18]. Image editing involves taking a source image $X_{1}^{src}$ and a source text prompt $c^{src}$ , and allowing the user to provide a target text prompt $c^{tgt}$ for targeted fine-grained editing to ultimately yield the edited image $X_{1}^{tgt}$ . For example, the source text prompt might be \"A black swan drifting,\" while the target text prompt could be \"A flamingo drifting.\" To simplify notation, we denote the vector fields corresponding to the source text prompt and the target text prompt as $v_{t}(X_{t}^{src})$ and $v_{t}(X_{t}^{tgt})$ , respectively. Their sampling process in RF", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16675", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/36f1d30bd80c21ddad15542befab8a480f8acaf33ffe63c46e1f5ac8296a5f5f.jpg", + "image_caption": [ + "(a) Pyramid-Edit" + ], + "image_footnote": [], + "bbox": [ + 98, + 89, + 472, + 383 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4bda0b0d40e70e4a2de0cc6e2e50891490ea6bbd0710f82eb3d74cd493d3426e.jpg", + "image_caption": [ + "(b) Wan-Edit", + "Figure 3. Two RF-based inversion-free video editing models: Pyramid-Edit and Wan-Edit. Pyramid-Edit is a multi-resolution temporal autoregressive architecture, while Wan-Edit (DiT architecture) treats temporal frames as a sequence and processing them simultaneously." + ], + "image_footnote": [], + "bbox": [ + 516, + 88, + 880, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "follows: $dX_{t}^{src} = v_{t}(X_{t}^{src})dt$ , $dX_{t}^{tgt} = v_{t}(X_{t}^{tgt})dt$ .", + "bbox": [ + 89, + 459, + 423, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "FlowEdit operates inversion-free, mapping source modes to the nearest target modes. It assumes target images follow straight trajectories, allowing latent interpolation. Thus, the nearest target mode is derived from the difference between source and target vector fields as", + "bbox": [ + 89, + 474, + 483, + 551 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nX _ {t} ^ {e d i t} = X _ {1} ^ {s r c} + X _ {t} ^ {t g t} - X _ {t} ^ {s r c}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 554, + 480, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here $X_{t}^{edit} = X_{t}^{tgt}$ , if $t \\to 1$ . A more straightforward explanation is: $X_{t}^{edit}$ evolves within the target distribution, continuously transitioning through the clean latent space across all timesteps, (i.e. $X_{t}^{edit}$ gradually transforms from a black swan to a flamingo as the timestep progresses). Therefore, we can obtain $X_{1}^{tgt}$ simply by solving $X_{t}^{edit}$ :", + "bbox": [ + 89, + 573, + 483, + 665 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} X _ {t} ^ {e d i t} = X _ {t - 1} ^ {e d i t} + d X _ {t} ^ {t g t} - d X _ {t} ^ {s r c} \\\\ \\approx X _ {t - 1} ^ {e d i t} + \\mathbb {E} \\left[ v _ {t} \\left(\\hat {X} _ {t} ^ {t g t}\\right) - v _ {t} \\left(X _ {t} ^ {s r c}\\right) \\mid X _ {1} ^ {s r c} \\right] d t \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 667, + 482, + 707 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $X_{t}^{tgt}$ is approximated as $\\hat{X}_{t}^{tgt} = X_{t - 1}^{edit} + X_{t}^{src} - X_{1}^{src}$ . To obtain the edited image, $X_{t}^{edit}$ is iteratively updated from $\\epsilon$ to 1, where $\\epsilon \\in [0,1)$ represents skipped timesteps.", + "bbox": [ + 89, + 710, + 483, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. RF-based Video Editing", + "text_level": 1, + "bbox": [ + 89, + 763, + 313, + 780 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This section introduces Pyramid-Edit, followed by a discussion of Wan-Edit, highlighting key features and differences.", + "bbox": [ + 89, + 785, + 483, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.1. Pyramid-Edit", + "text_level": 1, + "bbox": [ + 89, + 821, + 235, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pyramid-Flow [14] employs a multi-resolution scheme across timesteps and a temporally autoregressive architecture to effectively handle the spatial-temporal complexity in video generative modeling. It decomposes the flow", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "into segments over $K$ timestep windows within the interval [0,1]. Each window interpolates between successive solutions, reducing redundant computations in the earlier steps. For example, in the sampling process, within the $k$ -th window $[s_k, e_k]$ , it defines the rescaled timestep $t_k = (t - s_k) / (e_k - s_k)$ and the corresponding flow is", + "bbox": [ + 511, + 460, + 906, + 551 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nX _ {t _ {k}} = \\left(1 - t _ {k}\\right) \\bar {\\mathbf {x}} _ {s _ {k}} + t _ {k} \\bar {\\mathbf {x}} _ {e _ {k}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 564, + 906, + 580 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\mathbf {x}} _ {s _ {k}} = \\left(1 - s _ {k}\\right) X _ {0} + s _ {k} U p (D o w n \\left(X _ {1}, 2 ^ {k + 1}\\right)), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 582, + 905, + 599 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\mathbf {x}} _ {e _ {k}} = \\left(1 - e _ {k}\\right) X _ {0} + e _ {k} D o w n \\left(X _ {1}, 2 ^ {k}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 602, + 903, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\text{Down}(\\cdot, \\cdot)$ and $\\text{Up}(\\cdot)$ represent the down-sampling and up-sampling, respectively. Meanwhile, the vector field in the $k$ -th window is defined as $v_{t_k}(X_{t_k}) = \\bar{\\mathbf{x}}_{e_k} - \\bar{\\mathbf{x}}_{s_k}$ . To maintain continuity in the probability path within the pyramid structure, corrective noise is added at transition points between stages, as follows:", + "bbox": [ + 511, + 631, + 906, + 722 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nX _ {s _ {k - 1}} = \\frac {s _ {k - 1}}{e _ {k}} U p \\left(X _ {e _ {k}}\\right) + \\alpha \\mathbf {n} ^ {\\prime}, s. t. \\mathbf {n} _ {k - 1} \\sim \\mathcal {N} \\left(0, \\boldsymbol {\\Sigma} ^ {\\prime}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 733, + 906, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the rescaling coefficient $s_{k-1} / e_k$ aligns the means of these distributions, the corrective noise $\\mathbf{n}_{k-1}'$ weighted by $\\alpha$ aligns their covariance matrices, and the covariance matrix $\\boldsymbol{\\Sigma}'$ is related to the upsampling function.", + "bbox": [ + 511, + 770, + 905, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pyramid-Edit. The autoregressive architecture allows Pyramid-Edit to process frames sequentially, where previous frames serve as conditional information. However, the multi-resolution scheme across timesteps hinders direct use of FlowEdit. To address this, Pyramid-Edit integrates", + "bbox": [ + 511, + 830, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16676", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "FlowEdit (Eq. (3)) into each window as follows:", + "bbox": [ + 89, + 90, + 413, + 104 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nX _ {t _ {k}} ^ {e d i t} = X _ {t _ {k} - 1} ^ {e d i t} + \\mathbb {E} [ v _ {t _ {k}} (\\hat {X} _ {t _ {k}} ^ {t g t}) - v _ {t _ {k}} (X _ {t _ {k}} ^ {s r c}) | \\bar {\\mathbf {x}} _ {e _ {k}} ^ {s r c} ] d t, \\qquad (8)\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 113, + 482, + 132 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\hat{X}_{t_k}^{tgt} = X_{t_k - 1}^{edit} + X_{t_k}^{src} - \\bar{\\mathbf{x}}_{e_k}^{src}$ , and the target distribution is the end point of each window $\\bar{\\mathbf{x}}_{e_k}^{src}$ .", + "bbox": [ + 88, + 138, + 482, + 169 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The workflow proceeds as illustrated in Fig. 3a and is structured as follows: starting from noise and the lowest-resolution latent, $X_{t_0}^{edit}$ is initialized as $\\bar{\\mathbf{x}}_{e_0}^{src}$ , Eq. (8) is used to reconstruct and edit low-frequency information. At the end of the first stage, the reconstructed latent and the edited latent are upsampled to a higher resolution and passed through corrective noise as described in Eq. (7) to obtain the starting point for the next stage, $X_{s_1}^{edit}$ and $X_{s_1}^{src}$ . However, FlowEdit operates within the target distribution, meaning we require the end point for this stage, $X_{e_1}^{edit}$ and $X_{e_1}^{src}$ . The latter, $X_{e_1}^{src}$ , can be obtained through interpolation between $X_0$ and $X_1^{src}$ , i.e. $(X_{e_1}^{src} = \\bar{\\mathbf{x}}_{e_1}^{src})$ . To estimate the edited endpoint $X_{e_1}^{edit}$ , we first compute the difference between the start and end points of the source prompt in this stage. This difference is then added to the edited start latent variable, providing an estimate of $X_{e_1}^{edit}$ :", + "bbox": [ + 89, + 169, + 482, + 393 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\mathbf {x}} _ {e _ {1}} ^ {s r c} = X _ {s _ {1}} ^ {s r c} + \\bigl (\\bar {\\mathbf {x}} _ {e _ {1}} ^ {s r c} - X _ {s _ {1}} ^ {s r c} \\bigr), X _ {e _ {1}} ^ {\\mathrm {e d i t}} = X _ {s _ {1}} ^ {\\mathrm {e d i t}} + \\bigl (\\bar {\\mathbf {x}} _ {e _ {1}} ^ {s r c} - X _ {s _ {1}} ^ {s r c} \\bigr).\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 401, + 472, + 420 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This step ensures that the edited and reconstructed latents remain aligned with the source image distribution while incorporating the desired modifications, preserving the integrity of the generated features. This is key to the success of FlowEdit under the pyramid structure.", + "bbox": [ + 89, + 428, + 483, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Another key point is that, starting from the second frame, Pyramid-Edit respectively incorporates the source and edited historical information from previous frames (e.g., the $(i - 1)$ -th frame) as a conditioning factor within its noise-estimation network. This allows the model to leverage temporal dependencies from both the source and edited frames, enhancing the temporal consistency. The source and edited sampling processes can be simplified as follows:", + "bbox": [ + 89, + 503, + 483, + 625 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\underbrace {\\ldots \\rightarrow D o w n (X _ {1} ^ {s r c , i - 2} , 2 ^ {k + 1}) \\rightarrow D o w n (X _ {1} ^ {s r c , i - 1} , 2 ^ {k})} _ {\\mathrm {S o u r c e H i s t y o r c o n d i t i o n}} \\rightarrow X _ {t _ {k}} ^ {s r c , i},\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 632, + 475, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For generating subsequent frames, simply introduce the superscripts $i$ as indices for the historical latents in Eq. (8).", + "bbox": [ + 89, + 708, + 482, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.2. Wan-Edit", + "text_level": 1, + "bbox": [ + 89, + 746, + 205, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Wan2.1 [40] is a state-of-the-art T2V generation model built on the mainstream DiT framework [26], achieving significant advancements through innovations such as a novel 3D VAE, scalable training strategies, large-scale data curation. These contributions enable Wan2.1 to generate high-quality, temporally consistent videos with improved efficiency and scalability. We focus on adapting its core framework for video editing, leveraging its strengths in temporal modeling and high-fidelity generation.", + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared to Pyramid Flow, Wan2.1 adopts a much simpler architecture. It consists of 30 WanAttentionBlock layers, each integrating vision self-attention, text-to-vision cross-attention, and feed-forward mechanisms. Multiframe video inputs are encoded into a 3D vision latent by the 3D VAE encoder. These 3D vision latents are then flattened into a sequence and fed into the self-attention layers to encode temporal consistency. Finally, the sequence interacts with text tokens through cross-attention to achieve T2V generation. The self-attention mechanism captures intra-frame dependencies, while the cross-attention module integrates text tokens to enable text-guided video generation. This joint modeling of visual and textual information ensures high-quality, semantically aligned video outputs.", + "bbox": [ + 511, + 90, + 903, + 301 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Wan-Edit. This elegant and streamlined architecture closely resembles that of image generation models, enabling FlowEdit to be directly integrated without significant modifications. For multi-frame inputs from $i_1$ to $i_2$ , denote as $\\mathbf{X}_1^{src} = [X_1^{src,i_1},\\dots ,X_1^{src,i_2}]$ , the editing process handles all frames simultaneously by extending FlowEdit to the entire sequence, ensuring temporal consistency and efficient processing. The process is defined as follows:", + "bbox": [ + 511, + 301, + 905, + 422 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nd \\mathbf {X} _ {t} ^ {\\text {e d i t}} = \\mathbb {E} \\left[ v _ {t} ^ {\\Delta} \\left(\\mathbf {X} _ {t} ^ {\\text {s r c}}, \\mathbf {X} _ {t} ^ {\\text {e d i t}} + \\mathbf {X} _ {t} ^ {\\text {s r c}} - \\mathbf {X} _ {1} ^ {\\text {s r c}}\\right) \\mid \\mathbf {X} _ {1} ^ {\\text {s r c}} \\right] d t \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 430, + 903, + 449 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This allows for smooth and coherent edits across all frames, preserving the temporal dynamics of the video.", + "bbox": [ + 511, + 455, + 903, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 513, + 497, + 645, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline methods. We compare our RF-based editing methods with five diffusion-based models: TokenFlow [9], DMT [51], VidToMe [21], AnyV2V [17], and VideoGrain [50]. Diffusion-based models require inversion before editing, while our RF-based methods are inversion-free. DMT relies on optimization during editing, and VideoGrain uses depth maps to assist editing.", + "bbox": [ + 511, + 523, + 905, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. We fairly evaluate all models on our FiVE-Bench using 15 metrics. The nine commonly used metrics in Table 2 cover six aspects: structure distance [38], background preservation (PSNR, LPIPS [52], MSE, and SSIM [42] outside the editing mask), edit prompt-image consistency (CLIPSIM [44] for the full image and masked regions), image quality assessment (NIQE [33]), temporal consistency (motion fidelity score [51]) and running time. Additionally, our proposed FiVE-Acc introduces five metrics in Table. 3 to evaluate the accuracy of successful object editing.", + "bbox": [ + 511, + 628, + 905, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. Pyramid-Edit uses the 384P Pyramid-Flow model, and Wan-Edit on the Wan2.1 1.3B model. Pyramid-Edit uses 20 timesteps for the first frame and 10 for subsequent frames, while Wan-Edit uses 50 timesteps. All experiments were run on a single H100 GPU.", + "bbox": [ + 511, + 795, + 905, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Due to space limitations, see the Appendix for details on comparison methods, metrics, and implementation.", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16677", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fbe2db0efb4527685344ac157c73a7e4a8917db339c0f5dc7c66d85011d55a7c.jpg", + "table_caption": [ + "Table 2. Comparison of diffusion- and flow-based video editing methods on our proposed FiVE-Bench. * and † denote methods that require optimization and depth/segmentation maps, respectively." + ], + "table_footnote": [], + "table_body": "
MethodsStructure Dist.×103↓Background PreservationText AlignmentIQA NIQE↓Motion Fidelity S.×102↑Time (s) Per Frame↓
PSNR↑LPIPS×103↓MSE×104↓SSIM×102↑CLIPS↑CLIPS-edit ↑
Source Videos0.00∞0.000.00100.0024.5919.876.3393.76-
TokenFlow [9]35.6219.06263.61138.6572.5126.4621.154.0189.008.04
DMT* [51]85.9514.71404.60372.7851.6426.6621.445.2482.3025.98
VidToMe [21]22.3721.15263.9188.7570.6926.8421.054.6890.063.25
AnyV2V [17]71.3615.90348.59342.9750.7724.8919.725.0460.366.11
VideoGrain† [50]12.4027.05185.2125.1079.1325.6920.314.0888.5727.12
RFs (Ours)Pyramid-Edit28.6520.84276.5995.6371.7226.8220.205.4880.591.44
Wan-Edit12.5325.5794.6141.8482.5526.3921.236.5489.433.07
", + "bbox": [ + 102, + 121, + 900, + 277 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b3e95b30cb72bb11c9e104ddf2c0939bc798116d3699e837978b6a30233ee0f1.jpg", + "table_caption": [ + "Table 3. Comparison of diffusion- and flow-based video editing methods on the FiVE-Bench using FiVE-Acc metrics." + ], + "table_footnote": [], + "table_body": "
MethodYN-AccMC-AccU-Acc∩-AccFiVE-Acc↑
TokenFlow [9]19.3635.5136.6818.1827.43
DMT* [51]34.7862.0662.9833.8648.42
VidToMe [21]20.0333.5036.2017.3426.77
AnyV2V [17]30.6245.4248.9627.0938.02
VideoGrain† [50]30.5043.9744.3030.1737.23
Pyramid-Edit33.6754.0156.3631.3143.84
Wan-Edit41.4152.5355.7238.2246.97
", + "bbox": [ + 102, + 333, + 470, + 454 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Quantitative Comparison", + "text_level": 1, + "bbox": [ + 89, + 482, + 323, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison on Common Metrics. From the results, we observe that VideoGrain achieves the best performance among diffusion models, in terms of Structure Distance (12.40) and Background Preservation (PSNR: 27.05, MSE:25.10). Meanwhile, Wan-Edit, our proposed method, outperforms in the other two Background Preservation metrics (LPIPS: 94.61, SSIM: 82.55). Overall, VideoGrain and Wan-Edit perform similarly, but VideoGrain requires depth map assistance for editing and is extremely slow. In contrast, the inversion-free Wan-Edit demonstrates the highest efficiency with the lowest Time per Frame (1.44 seconds). These findings highlight the strengths and trade-offs of each method, offering valuable insights for selecting the appropriate approach based on specific application needs, and greatly contribute to the advancement of video editing.", + "bbox": [ + 88, + 506, + 482, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison on FiVE-Acc Metrics. Our evaluation of diffusion- and flow-based video editing methods on FiVE-Acc metrics reveals key findings. DMT, an optimization-based method, excels in MC-Acc (62.06%), U-Acc (62.98%), and overall FiVE-Acc (48.42%), highlighting its strength in fine-grained text alignment. This is due to its iterative optimization process, which prioritizes text fidelity, sometimes at the cost of background preservation. In contrast, Wan-Edit performs well in YN-Acc (41.41%) and $\\cap$ -Acc (38.22%). Pyramid-Edit also shows competitive results, especially in MC-Acc (54.01%) and U", + "bbox": [ + 89, + 734, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3313668fe285f9d8f3fe058aff6f68567e24c434319e1f6e4a0d47d16cd01527.jpg", + "table_caption": [ + "Table 4. Human vs. VLM model evaluation on FiVE-Acc metrics." + ], + "table_footnote": [], + "table_body": "
EvaluatorYN-AccMC-Acc∪-Acc∩-AccFiVE-Acc↑
Qwen-2.5-VL41.4152.5355.7238.2246.97
Human44.3750.3151.9842.7047.34
Δ (Diff)+2.96-2.22-3.74+4.48+0.37
", + "bbox": [ + 535, + 319, + 890, + 388 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Acc (56.36%), emphasizing the importance of fine-grained alignment. U-Acc and $\\cap$ -Acc metrics offer insights into task consistency, emphasizing the need for VLM-based evaluation to ensure visual and semantic coherence in editing.", + "bbox": [ + 511, + 414, + 903, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Overall, VideoGrain and Wan-Edit strike the best balance across metrics for fine-grained video editing. Wan-Edit outperforms VideoGrain in speed, offering faster and more stable results due to enhanced RF video generation and inversion models.", + "bbox": [ + 511, + 474, + 905, + 549 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Human Preference Validation for FiVE-Acc. We conducted a human study on 16 randomly sampled videos with 64 prompts, based on Wan-Edit editing outputs. Table 4 shows that human scores closely match Qwen-2.5-VL's results (47.34 vs. 46.97), confirming FiVE-Acc aligns well with human perception and offers consistent evaluation.", + "bbox": [ + 511, + 550, + 905, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Qualitative Comparison", + "text_level": 1, + "bbox": [ + 511, + 651, + 735, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Fig. 4, analyzing by edit type, non-rigid transformations and color changes are the easiest fine-grained edits, successfully handled by almost all methods. Object addition and rigid transformations are only successfully performed by the optimization-based DMT and our Wan-Edit. All methods fail at object removal, making it the most challenging task. From the perspective of background preservation in fine-grained editing, DMT and VidToME introduce significant alterations, while the recently proposed VideoGrain and the RF-based Pyramid-Edit and Wan-Edit largely retain the original background. The qualitative analysis further validates the quantitative findings: DMT achieves the best text alignment but disrupts the background, while VideoGrain and Wan-Edit offer the best balance across all metrics. Please refer to Appendix for details.", + "bbox": [ + 509, + 674, + 906, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16678", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Edit1 Object (w/o non-rigid deform): Woman $\\rightarrow$ Man", + "bbox": [ + 106, + 90, + 382, + 101 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d51bc72ab52e0991c3f6613e73093db934209f0d905111b10dd0ad72cc5eb293.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 102, + 104, + 895, + 215 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2249861f1ef7256f5797f79bfde2c1e68c1c9201a597c0cca2fcf0c230bf1a1b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 102, + 215, + 895, + 315 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Edit3 Color: black dress $\\rightarrow$ red dress", + "bbox": [ + 106, + 316, + 294, + 325 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c5fcfe4470254a9f0e1cd6ab8f3a6d25ede2b922bc75fbb2e84ea6404bedd489.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 102, + 327, + 895, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Edit4 Material: A woman $\\rightarrow$ A porcelain woman", + "bbox": [ + 106, + 428, + 349, + 438 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ae21f1ee99f227fd8b908a2e08fb78d8d8b23aea32d714737009a142f27a5871.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 102, + 440, + 895, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "EditS Add: A woman $\\rightarrow$ A woman followed closely by a dog", + "bbox": [ + 104, + 558, + 413, + 570 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/844b9bde26cec22be6b4247f746b8092d15081a3412e02d36987849bf727070a.jpg", + "image_caption": [ + "Figure 4. Editing results across six editing types and five comparison methods." + ], + "image_footnote": [], + "bbox": [ + 101, + 571, + 895, + 685 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and Future Work. This work simply adapts image editing to video editing, which leads to room for improvement in the quality of Pyramid-Edit. In future work, we aim to further explore advancements in RF models for video editing. Additionally, handling challenging cases in the benchmark, such as large motions or object removal, will be an important direction for improvement.", + "bbox": [ + 89, + 728, + 482, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 89, + 848, + 209, + 863 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce FiVE-Bench, a benchmark for fine-grained video editing, and propose the VLM-based FiVE-Acc met", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ric, which evaluates the accuracy of object-level editing success. Additionally, we adapt two RF-based video editing methods, Pyramid-Edit and Wan-Edit. To the best of our knowledge, this is the first comprehensive quantitative and qualitative comparison of emerging diffusion-based and flow-based models in the video editing community. FiVE-Bench and FiVE-Acc provide a standardized framework for fine-grained video editing evaluation, guiding future advancements in efficient and high-fidelity video editing solutions. We hope this benchmark drives innovation in both research and real-world applications.", + "bbox": [ + 511, + 728, + 906, + 895 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16679", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 1", + "[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In IEEE Conf. Comput. Vis. Pattern Recog., pages 18392-18402, 2023. 4", + "[3] Duygu Ceylan, Chun-Hao P Huang, and Niloy J Mitra. Pix2video: Video editing using image diffusion. In Int. Conf. Comput. Vis., pages 23206-23217, 2023. 3", + "[4] Weifeng Chen, Jie Wu, Pan Xie, Hefeng Wu, Jiashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-a-video: Controllable text-to-video generation with diffusion models. arXiv preprint arXiv:2305.13840, 2023. 2", + "[5] Yingying Deng, Xiangyu He, Changwang Mei, Peisong Wang, and Fan Tang. Fireflow: Fast inversion of rectified flow for image semantic editing. arXiv preprint arXiv:2412.07517, 2024. 2, 3", + "[6] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Adv. Neural Inform. Process. Syst., 34:8780-8794, 2021. 2", + "[7] Wenkai Dong, Song Xue, Xiaoyue Duan, and Shumin Han. Prompt tuning inversion for text-driven image editing using diffusion models. In Int. Conf. Comput. Vis., pages 7430-7440, 2023. 3", + "[8] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Int. Conf. on Machine Learning, 2024. 2, 3", + "[9] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. In Int. Conf. Learn. Represent., 2024. 2, 3, 6, 7", + "[10] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff:Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 2", + "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Adv. Neural Inform. Process. Syst., 33:6840-6851, 2020. 2", + "[12] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2", + "[13] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 3, 4", + "[14] Yang Jin, Zhicheng Sun, Ningyuan Li, Kun Xu, Hao Jiang, Nan Zhuang, Quzhe Huang, Yang Song, Yadong Mu, and Zhouchen Lin. Pyramidal flow matching for efficient video generative modeling. arXiv preprint arXiv:2410.05954, 2024. 1, 2, 3, 5" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Xuan Ju, Ailing Zeng, Yuxuan Bian, Shaoteng Liu, and Qiang Xu. Pnp inversion: Boosting diffusion-based editing with 3 lines of code. International Conference on Learning Representations (ICLR), 2024. 3", + "[16] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439, 2023. 3", + "[17] Max Ku, Cong Wei, Weiming Ren, Harry Yang, and Wenhu Chen. Anyv2v: A tuning-free framework for any video-to-video editing tasks. arXiv preprint arXiv:2403.14468, 2024. 3, 6, 7", + "[18] Vladimir Kulikov, Matan Kleiner, Inbar Huberman-Spiegelglas, and Tomer Michaeli. Flowedit: Inversion-free text-based editing using pre-trained flow models. arXiv preprint arXiv:2412.08629, 2024. 1, 2, 3, 4", + "[19] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux.2,3", + "[20] Minghan Li, Shuai Li, Xindong Zhang, and Lei Zhang. Univs: Unified and universal video segmentation with prompts as queries. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3227-3238, 2024. 4", + "[21] Xirui Li, Chao Ma, Xiaokang Yang, and Ming-Hsuan Yang. Vidthome: Video token merging for zero-shot video editing. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7486-7495, 2024. 2, 3, 6, 7", + "[22] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. In Int. Conf. Learn. Represent., 2023. 4", + "[23] Xingchao Liu, Chengyue Gong, et al. Flow straight and fast: Learning to generate and transfer data with rectified flow. In Int. Conf. Learn. Represent., 2023. 2, 4", + "[24] Guoqing Ma, Haoyang Huang, Kun Yan, Liangyu Chen, Nan Duan, Shengming Yin, Changyi Wan, Ranchen Ming, Xiaoniu Song, Xing Chen, et al. Step-video-t2v technical report: The practice, challenges, and future of video foundation model. arXiv preprint arXiv:2502.10248, 2025. 1", + "[25] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6038–6047, 2023. 3", + "[26] OpenAI. Video generation models as world simulators, 2024. 1, 6", + "[27] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. 3", + "[28] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. arXiv preprint arXiv:2303.09535, 2023. 2, 3", + "[29] Bosheng Qin, Juncheng Li, Siliang Tang, Tat-Seng Chua, and Yueting Zhuang. Instructvid2vid: Controllable video editing with natural language instructions. In 2024 IEEE" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16680", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2024. 3", + "[30] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dólar, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos, 2024. 4", + "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10684-10695, 2022. 1", + "[32] Litu Rout, Yujia Chen, Nataniel Ruiz, Constantine Caramanis, Sanjay Shakkottai, and Wen-Sheng Chu. Semantic image inversion and editing using rectified stochastic differential equations. arXiv preprint arXiv:2410.10792, 2024. 2, 3", + "[33] Michele A Saad and Alan C Bovik. Blind quality assessment of videos using a model of natural scene statistics and motion coherency. In 2012 Conference Record of the Forty Sixth Asilomar Conference on Signals, Systems and Computers (ASILOMAR), pages 332-336. IEEE, 2012. 6", + "[34] Uriel Singer, Amit Zohar, Yuval Kirstain, Shelly Sheynin, Adam Polyak, Devi Parikh, and Yaniv Taigman. Video editing via factorized diffusion distillation. In Eur. Conf. Comput. Vis., pages 450-466. Springer, 2024. 2", + "[35] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 2, 3", + "[36] Spencer Sterling. Zeroscope. https://huggingface.co/cerspense/zeroscope_v2_576w, 2023.1", + "[37] Shuyuan Tu, Qi Dai, Zhi-Qi Cheng, Han Hu, Xintong Han, Zuxuan Wu, and Yu-Gang Jiang. Motioneditor: Editing video motion via content-aware diffusion. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7882-7891, 2024. 3", + "[38] Narek Tumanyan, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Splicing vit features for semantic appearance transfer. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10748-10757, 2022. 6", + "[39] Bram Wallace, Akash Gokul, and Nikhil Naik. Edict: Exact diffusion inversion via coupled transformations. In IEEE Conf. Comput. Vis. Pattern Recog., pages 22532-22541, 2023. 3", + "[40] Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, et al. Wan: Open and advanced large-scale video generative models. arXiv preprint arXiv:2503.20314, 2025. 1, 2, 3, 6", + "[41] Jiangshan Wang, Junfu Pu, Zhongang Qi, Jiayi Guo, Yue Ma, Nisha Huang, Yuxin Chen, Xiu Li, and Ying Shan. Taming rectified flow for inversion and editing. arXiv preprint arXiv:2411.04746, 2024. 2, 3", + "[42] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process., 13(4): 600-612, 2004. 6" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Yujie Wei, Shiwei Zhang, Zhiwu Qing, Hangjie Yuan, Zhiheng Liu, Yu Liu, Yingya Zhang, Jingren Zhou, and Hongming Shan. Dreamvideo: Composing your dream videos with customized subject and motion. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6537-6549, 2024. 3", + "[44] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. GODIVA: Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806, 2021. 6", + "[45] Chen Henry Wu and Fernando De la Torre. A latent space of stochastic diffusion models for zero-shot image editing and guidance. In Int. Conf. Comput. Vis., pages 7378-7387, 2023. 3", + "[46] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Int. Conf. Comput. Vis., pages 7623-7633, 2023. 2, 3", + "[47] Jay Zhangjie Wu, Xiuyu Li, Difei Gao, Zhen Dong, Jinbin Bai, Aishani Singh, Xiaoyu Xiang, Youzeng Li, Zuwei Huang, Yuanxi Sun, et al. Cvpr 2023 text guided video editing competition. arXiv preprint arXiv:2310.16003, 2023. 2", + "[48] Chenxi Xie, Minghan Li, Shuai Li, Yuhui Wu, Qiaosi Yi, and Lei Zhang. Dnaedit: Direct noise alignment for text-guided rectified flow editing. arXiv preprint arXiv:2506.01430, 2025.3", + "[49] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. In SIGGRAPH Asia 2023 Conference Papers, pages 1-11, 2023. 3", + "[50] Xiangpeng Yang, Linchao Zhu, Hehe Fan, and Yi Yang. Videograin: Modulating space-time attention for multigrained video editing. arXiv preprint arXiv:2502.17258, 2025. 2, 3, 6, 7", + "[51] Danah Yatim, Rafail Fridman, Omer Bar-Tal, Yoni Kasten, and Tali Dekel. Space-time diffusion features for zero-shot text-driven motion transfer. In IEEE Conf. Comput. Vis. Pattern Recog., pages 8466-8476, 2024. 3, 6, 7", + "[52] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conf. Comput. Vis. Pattern Recog., pages 586-595, 2018. 6", + "[53] Zhenghao Zhang, Zuozhuo Dai, Long Qin, and Weizhi Wang. Effived: Efficient video editing via text-instruction diffusion models. arXiv preprint arXiv:2403.11568, 2024. 3", + "[54] Zhixing Zhang, Yanyu Li, Yushu Wu, Yanwu Xu, Anil Kag, Ivan Skorokhodov, Willi Menapace, Aliaksandr Siarohin, Junli Cao, Dimitris Metaxas, et al. Sf-v: Single forward video generation model. arXiv preprint arXiv:2406.04324, 2024. 3", + "[55] Rui Zhao, Yuchao Gu, Jay Zhangjie Wu, David Junhao Zhang, Jia-Wei Liu, Weijia Wu, Jussi Keppo, and Mike Zheng Shou. Motiondirector: Motion customization of text-to-video diffusion models. In Eur. Conf. Comput. Vis., pages 273–290. Springer, 2025. 3" + ], + "bbox": [ + 516, + 90, + 903, + 869 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16681", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/FiVE-Bench_ A Fine-grained Video Editing Benchmark for Evaluating Emerging Diffusion and Rectified Flow Models/61b80164-0e5a-42b8-b65d-d1598a219bbd_model.json b/2025/FiVE-Bench_ A Fine-grained Video Editing Benchmark for Evaluating Emerging Diffusion and Rectified Flow Models/61b80164-0e5a-42b8-b65d-d1598a219bbd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a49adc5ca932bb919e825077005f0d12daf59602 --- /dev/null +++ b/2025/FiVE-Bench_ A Fine-grained Video Editing Benchmark for Evaluating Emerging Diffusion and Rectified Flow Models/61b80164-0e5a-42b8-b65d-d1598a219bbd_model.json @@ -0,0 +1,2288 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.177, + 0.124, + 0.823, + 0.19 + ], + "angle": 0, + "content": "FiVE-Bench: A Fine-grained Video Editing Benchmark for Evaluating Emerging Diffusion and Rectified Flow Models" + }, + { + "type": "text", + "bbox": [ + 0.196, + 0.215, + 0.8, + 0.252 + ], + "angle": 0, + "content": "Minghan Li\\(^{1*}\\), Chenxi Xie\\(^{2*}\\), Yichen Wu\\(^{1,3}\\), Lei Zhang\\(^{2}\\) and Mengyu Wang\\(^{1\\dagger}\\) \n\\(^{1}\\)Harvard University" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.253, + 0.767, + 0.27 + ], + "angle": 0, + "content": "\\(^{2}\\)Hong Kong Polytechnic University, \\(^{3}\\)City University of Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.334, + 0.272, + 0.658, + 0.287 + ], + "angle": 0, + "content": "{mili4, mengyu_wang}@meei.harvard.edu" + }, + { + "type": "title", + "bbox": [ + 0.191, + 0.296, + 0.269, + 0.308 + ], + "angle": 0, + "content": "FiVE-Dataset" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.31, + 0.373, + 0.568 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.545, + 0.296, + 0.755, + 0.308 + ], + "angle": 0, + "content": "Wan-Edit Results and FiVE-Evaluation" + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.31, + 0.903, + 0.567 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.13, + 0.573, + 0.867, + 0.588 + ], + "angle": 0, + "content": "Figure 1. Overview of the proposed FiVE-Bench, with editing results from Wan-Edit and corresponding FiVE-Acc scores." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.601, + 0.327, + 0.616 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.633, + 0.486, + 0.875 + ], + "angle": 0, + "content": "Numerous text-to-video (T2V) editing methods have emerged recently, but the lack of a standardized benchmark for fair evaluation has led to inconsistent claims and an inability to assess model sensitivity to hyperparameters. Fine-grained video editing is crucial for enabling precise, object-level modifications while maintaining context and temporal consistency. To address this, we introduce FiVE-Bench, a Fine-grained Video Editing Benchmark for evaluating emerging diffusion and rectified flow models. Our benchmark includes 74 real-world videos and 26 generated videos, featuring 6 fine-grained editing types, 420 object-level editing prompt pairs, and their corresponding masks. Additionally, we adapt the latest rectified flow (RF) T2V generation models—Pyramid-Flow [14] and Wan2.1 [40]—by introducing FlowEdit [18], resulting in training-free and inversion-free video editing mod" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.602, + 0.907, + 0.813 + ], + "angle": 0, + "content": "els Pyramid-Edit and Wan-Edit. We compare five diffusion methods with our two RF methods on the proposed FiVE-Bench, evaluating them across 15 metrics. These metrics include background preservation, text-video similarity, temporal consistency, and generated video quality. To further enhance object-level evaluation, we introduce FiVE-Acc, a novel metric leveraging Vision-Language Models (VLMs) to assess the success of fine-grained video editing. Experimental results demonstrate that RF-based editing significantly outperforms diffusion-based methods, with WanEdit achieving the best overall performance and exhibiting the least sensitivity to hyperparameters. More video demo available on the website: https://sites.google.com/view/five-benchmark." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.845, + 0.646, + 0.86 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Recent advancements in text-to-image (T2I) and text-to-video (T2V) generation [1, 14, 24, 26, 31, 36, 40] have" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.351, + 0.901 + ], + "angle": 0, + "content": "*Equal contribution, † Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16672" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.183 + ], + "angle": 0, + "content": "led to the emergence of numerous image and video editing methods [4, 9, 10, 46], enabling users to create and modify video content with unprecedented flexibility. These methods leverage powerful generative models, such as diffusion models [6, 11, 12, 35] and Rectified Flow (RF) [8, 23] models, to synthesize and edit videos based on textual prompts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.185, + 0.483, + 0.412 + ], + "angle": 0, + "content": "However, despite significant progress in these fields, the lack of a standardized benchmark for fair and comprehensive evaluation has still become a major bottleneck hindering further development. Currently, the video editing field has only one publicly available dataset, TGVE [47] /TGVE+ [34] (see Table 1), which includes 76 videos and 6 editing types, such as style transfer, background replacement, single-object and multi-object transformations, as well as object addition and removal. While TGVE+ aims to cover as many editing types as possible, its evaluation within each editing type is not comprehensive, particularly in supporting fine-grained video editing. Fine-grained video editing requires precise object-level modifications while preserving the overall context and temporal consistency of the video, posing higher demands on model performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.413, + 0.483, + 0.579 + ], + "angle": 0, + "content": "Due to the lack of a standardized benchmark and unified evaluation framework for fine-grained video editing, existing methods [9, 21, 28, 50] often rely on self-collected small-scale video datasets to validate their effectiveness. This approach not only makes it difficult to objectively compare the performance of different methods but also fails to comprehensively assess model robustness to hyperparameters or determine their suitability for real-world applications. Therefore, building a comprehensive and challenging benchmark for fine-grained video editing has become an urgent need to advance the field." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.483, + 0.824 + ], + "angle": 0, + "content": "To address these challenges, we introduce FiVE-Bench (Fine-grained Video Editing Benchmark), a comprehensive benchmark designed to evaluate emerging diffusion and rectified flow models for fine-grained video editing. FiVE-Bench includes 74 real-world videos and 26 generated videos, covering a diverse range of scenes and editing scenarios. It features 6 fine-grained editing types, 420 object-level editing prompt pairs, and their corresponding masks, providing a rich and challenging benchmark for video editing. Additionally, we propose FiVE-Acc, a novel evaluation metric that leverages Vision-Language Models (VLMs) to measure the success rate of fine-grained object-level editing. Unlike traditional similarity-based scores, FiVE-Acc focuses on evaluating whether the intended edit was successfully achieved, with particular emphasis on semantic alignment of the edited object." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Recently, many image editing methods [5, 18, 32, 41] based on the latest rectified flow T2I generation model Flux [19] have achieved state-of-the-art editing results and fidelity, such as RF-Inversion [32], RF-Solver [41], and FlowEdit [18]. However, to the best of our knowledge," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "there are no video editing methods based on RF T2V generation models, primarily due to the relatively slow development of T2V generation models. Fortunately, the recent release of two state-of-the-art RF-based T2V generation models—Pyramid-Flow [14] and Wan2.1 [40]—has opened new possibilities for RF video editing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.184, + 0.905, + 0.44 + ], + "angle": 0, + "content": "Building on the latest advancements, we adapt Pyramid-Flow and Wan2.1 by introducing FlowEdit [18]. This adaptation facilitates training-free and inversion-free video editing models, Pyramid-Edit and Wan-Edit, which leverage the strong temporal consistency inherent in video generation models without requiring additional attention mechanisms. Specifically, Pyramid-Flow employs a multi-resolution and temporally autoregressive model architecture across time steps. To adapt to this unique design, we modified FlowEdit to support multi-resolution processing, enabling seamless integration with the Pyramid-Flow framework, leading to Pyramid-Edit. On the other hand, Wan2.1 utilizes the MMDiT architecture, which unfolds temporal frames into sequences and processes them via self-attention. Since its design is fundamentally similar to image generation models, Wan2.1 can directly adopt FlowEdit without any additional modifications, giving rise to Wan-Edit." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.441, + 0.905, + 0.636 + ], + "angle": 0, + "content": "To show the effectiveness of the proposed RF-based editing methods, we compare them with five diffusion-based editing methods on the built FiVE-Bench, evaluating them across fifteen metrics, including background preservation, edited text-video similarity, temporal consistency, and generated video quality. Our experiments demonstrate that RF-based editing methods significantly outperform diffusion-based methods across multiple metrics. Notably, Wan-Edit achieves the best overall performance, exhibiting superior editing quality and the least sensitivity to hyperparameters. These findings highlight the potential of RF-based models for fine-grained video editing and provide valuable insights for future research in this area." + }, + { + "type": "text", + "bbox": [ + 0.536, + 0.639, + 0.882, + 0.653 + ], + "angle": 0, + "content": "In summary, this work makes four key contributions:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.654, + 0.905, + 0.699 + ], + "angle": 0, + "content": "- We introduce FiVE-Bench, a comprehensive benchmark for fine-grained video editing, featuring diverse videos, editing types, and evaluation metrics." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.7, + 0.905, + 0.73 + ], + "angle": 0, + "content": "- We propose FiVE-Acc, a novel metric leveraging VLMs to assess the success of fine-grained editing." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.731, + 0.904, + 0.759 + ], + "angle": 0, + "content": "- We propose the first two RF-based T2V editing models: Pyramid-Edit and Wan-Edit." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.76, + 0.905, + 0.805 + ], + "angle": 0, + "content": "- We evaluate 7 video editing methods across 15 metrics, showing that RF-based editing consistently outperforms diffusion-based approaches." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.654, + 0.905, + 0.805 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.822, + 0.655, + 0.837 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.848, + 0.759, + 0.862 + ], + "angle": 0, + "content": "2.1. Diffusion and RF Inversion" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Diffusion inversion. Diffusion models (DMs) drive advances in inversion techniques for better control over gen" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "16673" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.293, + 0.09, + 0.703, + 0.104 + ], + "angle": 0, + "content": "Table 1. Statistics of existing video editing datasets and benchmarks." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.108, + 0.905, + 0.22 + ], + "angle": 0, + "content": "
DatasetUsageNum. ImagesNum. VideosFrames Per VideoNum. PromptsGen. VideosSrc. obj. WordsEdit obj. WordsObj. Instruct.Src. obj. MasksType of Edited Prompts
VIVID-10MTrain672K73.7K3010MXXX✓✓object, add, remove
Señorita-2MTrain0388.9K33~642MXXX✓Xstyle, object, multiple, color, add, remove, motion
TGVEEval07632304XXXXXstyle, object, bg, multiple
TGVE+Eval076321417XXXXXstyle, object, bg, multiple, color, add, remove
VIVID-10MEval06433~64852XXX✓✓object, add, remove
FiVE (Ours)Eval010035 ~ 126420✓✓✓✓✓object (w&w/o non-rigid), color, material, add, remove
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.246, + 0.486, + 0.488 + ], + "angle": 0, + "content": "erated samples. DDIM [35] accelerates sampling but struggles with fine-grained reconstructions due to nonlinearities and score estimation errors. To improve accuracy, fine-tuning and optimization-based methods [7, 15, 25, 39, 45] have been proposed. While these advancements significantly enhance the fidelity and accuracy of DDIM inversion, the associated computational costs and resource demands remain substantial limiting factors. Although diffusion-based methods have achieved outstanding performance, recent studies [8, 14, 19, 32] suggest that Rectified Flow models (RFs) hold significant potential to surpass DMs in certain applications. Notably, the inversion and editing capabilities of RF models have been explored in image editing [5, 18, 32, 41, 48], demonstrating their efficiency and effectiveness. Building on these strengths, this work extends their potential in video editing." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.5, + 0.236, + 0.515 + ], + "angle": 0, + "content": "2.2. Video Editing" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.523, + 0.483, + 0.749 + ], + "angle": 0, + "content": "Video editing techniques can generally be categorized into three types based on their training: training-free methods, one/few-shot finetuned methods [37, 43, 46, 55], and massive data finetuned methods. Training-free models [17, 21, 28, 50], enable task execution without retraining. TokenFlow [9] utilizes pre-trained models for efficient video synthesis, while DMT [51] transfers motion via diffusion. Other models like Pixel2Video [3], Render-A-Video [49], and Text2Video-zero [16] enhance accessibility for real-time video editing. One/few-shot fine-tuned models adapt pre-trained models for video editing with minimal fine-tuning. Methods like Tune-A-Video [46] and MotionEditor [37] achieve strong results using limited annotated samples, balancing customization and efficiency for task-specific editing without extensive retraining." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Massive data fine-tuned models undergo extensive training on large datasets for high-quality, versatile video editing [29]. InstructVid2Vid [29] enables complex edits via natural language instructions, while EffiVED [53] refines broad datasets into high-quality subsets for efficient editing. SF-V [54] introduces adversarial training for video synthesis. Despite their precision and adaptability, these methods are computationally expensive, limiting practicality in resource-constrained settings. These methods focus on learning video motion, which may compromise back" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.246, + 0.906, + 0.322 + ], + "angle": 0, + "content": "ground preservation. Recently, VideoGrain [50] has emphasized multi-granularity video editing rather than fine-grained edits, aiming to retain background details while modifying target objects. This work aims to address this issue and advance the field." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.337, + 0.871, + 0.353 + ], + "angle": 0, + "content": "3. FiVE Benchmark and FiVE-Acc Metric" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.363, + 0.686, + 0.378 + ], + "angle": 0, + "content": "3.1. FiVE Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.386, + 0.906, + 0.431 + ], + "angle": 0, + "content": "We present the FiVE benchmark, which includes the FiVE video dataset and six fine-grained editing tasks. Additional details can be found in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.432, + 0.907, + 0.674 + ], + "angle": 0, + "content": "Videos in FiVE-Bench. We meticulously curate 74 real-world videos suitable for fine-grained editing from the DAVIS [27] dataset. Consecutive frames are extracted at every 8-frame interval, and the GPT-4o [13] is employed to generate formatted captions for these videos. These captions encompass details pertaining to object category, action, background, and camera movement. Additionally, during the annotation process, we document instances of object deformation (such as limb movements in humans or animals), which aids in differentiating the complexity of editing tasks during evaluation. To further enhance the diversity of our benchmark, we generate 26 highly realistic synthetic videos using the T2V model [40]. These synthetic videos not only expand the range of video categories but also enable a comparative analysis of editing performance across real and synthetic videos." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.675, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Six Fine-grained Video Editing Tasks. FiVE-Bench comprises six fine-grained video editing tasks, totaling 420 high-quality editing prompt pairs. It primarily includes four object-targeted editing tasks of increasing complexity: object substitution without non-rigid deformation, object substitution with non-rigid deformation, color alteration, and material modification. For each task, we use GPT-4o [13] to generate four source-target prompt pairs per video in the benchmark, resulting in 400 high-quality editing pairs. To further enhance the diversity of editing types within the benchmark, we select 10 videos each for the add and remove tasks and design corresponding source-target prompt pairs. Additionally, we explicitly provide both object-related source words and edited words in the source and edited text prompts, respectively, denoted as" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16674" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.09, + 0.904, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.276, + 0.243, + 0.72, + 0.258 + ], + "angle": 0, + "content": "Figure 2. The VLM-based FiVE accuracy (FiVE-Acc) evaluation pipeline." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.284, + 0.483, + 0.391 + ], + "angle": 0, + "content": "\"src. obj. words\" and \"edit obj. words\" in Table 1. To accommodate different editing methods, we use GPT-4o [13] to generate instruction prompts (\"Obj. Instr.\" in Table 1) for each source-target pair, ensuring compatibility with models like InstructPix2Pix [2]. For each video, we employ SAM2 [20, 30] to generate masks of the edited regions, enabling the evaluation of background preservation metrics." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.403, + 0.262, + 0.419 + ], + "angle": 0, + "content": "3.2. FiVE-Acc Metric" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.427, + 0.484, + 0.503 + ], + "angle": 0, + "content": "To evaluate the accuracy of fine-grained video editing, we propose FiVE accuracy (FiVE-Acc) evaluation, a framework assessing how precisely video editing models modify the target object. Fig. 2 presents the complete pipeline, comprising video editing and accuracy evaluation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.504, + 0.484, + 0.58 + ], + "angle": 0, + "content": "Video editing outputs. Given a source video, a video editing model is prompted with a source prompt, describing the initial scene, and a target prompt, specifying the desired modification. The model generates a target video, where the intended transformation is applied, as shown in Fig. 2a." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.581, + 0.484, + 0.642 + ], + "angle": 0, + "content": "FiVE-Acc evaluation. To quantitatively evaluate the fidelity of video editing, we employ a Vision-Language Model (VLM) (e.g., Qwen-2.5-VL) to analyze the edited video. The evaluation includes two types of questions:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.643, + 0.483, + 0.748 + ], + "angle": 0, + "content": "- Yes/No Questions: The model is asked whether the source and target objects are present in the target video. Editing succeeds only if the model answers \"No\" for the source object and \"Yes\" for the target, leading to the calculation of \\(YN\\)-Acc. The source object question ensures the model both removes the source and adds the target, refining accuracy evaluation." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.749, + 0.483, + 0.825 + ], + "angle": 0, + "content": "- Multi-choice Question: It evaluates whether the VLM recognizes the target (e.g., flamingo) or source object (e.g., black swan) in the edited video. The model selects between the two objects to compute MC-Acc, measuring its recognition accuracy." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.643, + 0.483, + 0.825 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Based on the two obtained accuracies, we further calculate the union and intersection accuracies to derive \\(\\cup\\)-Acc and \\(\\cap\\)-Acc. These offer a more detailed evaluation: the former reflects the model's overall editing success accuracy, while the latter highlights the model's high-quality editing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.284, + 0.907, + 0.377 + ], + "angle": 0, + "content": "success accuracy. \\(FiVE - Acc\\) is the average of four components: \\(FiVE - Acc = 0.25 \\times (YN - Acc + MC - Acc + \\cup - Acc + \\cap - Acc)\\). It uses the VLM's recognition ability to evaluate the alignment between the edited object and its real data distribution. To the best of our knowledge, we are the first to introduce accuracy evaluation in fine-grained T2V editing." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.389, + 0.863, + 0.407 + ], + "angle": 0, + "content": "4. Methods: Pyramid-Edit and Wan-Edit" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.414, + 0.644, + 0.431 + ], + "angle": 0, + "content": "4.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.436, + 0.906, + 0.528 + ], + "angle": 0, + "content": "Rectified Flow (RF) Models [22, 23]. Let \\( q_{0} \\) denote the source distribution (i.e., standard Gaussian \\( \\mathcal{N}(0,I) \\)), \\( p_{0} \\) refer to the target distribution (i.e., the distribution over images), and \\( v_{t}(\\cdot) \\) is the time-varying vector field. RF models [22, 23] aim to transform \\( q_{0} \\) progressively into \\( p_{0} \\) by solving the following ordinary differential equation (ODE):" + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.534, + 0.905, + 0.549 + ], + "angle": 0, + "content": "\\[\nd X _ {t} = v _ {t} \\left(X _ {t}\\right) d t, \\quad t \\in [ 0, 1 ]. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.554, + 0.907, + 0.749 + ], + "angle": 0, + "content": "Here, the vector field \\( v_{t}(X_{t}) = f(X_{t},c,1 - t;\\varphi) \\), where \\( f(\\cdot) \\) is a neural network parameterized by \\( \\varphi \\) and \\( c \\) is the guided text. Given the ODE in Eq. (1), we start with \\( X_0\\sim q_0 \\) from the source distribution and integrate over \\( t:0\\to 1 \\) to reach the endpoint \\( X_{1}\\sim p_{0} \\), which is sampled from the target distribution. The forward process of RF assumes the linear path between the two states: \\( X_{t} = (1 - t)X_{0} + tX_{1} \\), then we can derive the vector field of \\( v_{t}(\\cdot):u_{t}(X_{t}) = X_{1} - X_{0} \\). Then, \\( v_{t}(X_{t}) \\) is used to approximate the network \\( f(X_{t},c,t;\\varphi) \\) by minimizing the loss function \\( \\mathcal{L}_{\\varphi} = \\mathbb{E}_{t\\sim \\mathcal{U}[0,1],Y_t}[\\| v_t(X_t) - f(X_t,c,t;\\varphi)\\| _2^2 ] \\). RF ensures that their sampling paths are relatively straight, enabling the use of a small number of discretization steps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "RF-based Image Editing: FlowEdit [18]. Image editing involves taking a source image \\( X_{1}^{src} \\) and a source text prompt \\( c^{src} \\), and allowing the user to provide a target text prompt \\( c^{tgt} \\) for targeted fine-grained editing to ultimately yield the edited image \\( X_{1}^{tgt} \\). For example, the source text prompt might be \"A black swan drifting,\" while the target text prompt could be \"A flamingo drifting.\" To simplify notation, we denote the vector fields corresponding to the source text prompt and the target text prompt as \\( v_{t}(X_{t}^{src}) \\) and \\( v_{t}(X_{t}^{tgt}) \\), respectively. Their sampling process in RF" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16675" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.09, + 0.473, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.243, + 0.388, + 0.334, + 0.401 + ], + "angle": 0, + "content": "(a) Pyramid-Edit" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.882, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.67, + 0.388, + 0.741, + 0.4 + ], + "angle": 0, + "content": "(b) Wan-Edit" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.406, + 0.907, + 0.435 + ], + "angle": 0, + "content": "Figure 3. Two RF-based inversion-free video editing models: Pyramid-Edit and Wan-Edit. Pyramid-Edit is a multi-resolution temporal autoregressive architecture, while Wan-Edit (DiT architecture) treats temporal frames as a sequence and processing them simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.46, + 0.424, + 0.476 + ], + "angle": 0, + "content": "follows: \\(dX_{t}^{src} = v_{t}(X_{t}^{src})dt\\), \\(dX_{t}^{tgt} = v_{t}(X_{t}^{tgt})dt\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.476, + 0.484, + 0.552 + ], + "angle": 0, + "content": "FlowEdit operates inversion-free, mapping source modes to the nearest target modes. It assumes target images follow straight trajectories, allowing latent interpolation. Thus, the nearest target mode is derived from the difference between source and target vector fields as" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.555, + 0.482, + 0.572 + ], + "angle": 0, + "content": "\\[\nX _ {t} ^ {e d i t} = X _ {1} ^ {s r c} + X _ {t} ^ {t g t} - X _ {t} ^ {s r c}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.574, + 0.484, + 0.666 + ], + "angle": 0, + "content": "Here \\( X_{t}^{edit} = X_{t}^{tgt} \\), if \\( t \\to 1 \\). A more straightforward explanation is: \\( X_{t}^{edit} \\) evolves within the target distribution, continuously transitioning through the clean latent space across all timesteps, (i.e. \\( X_{t}^{edit} \\) gradually transforms from a black swan to a flamingo as the timestep progresses). Therefore, we can obtain \\( X_{1}^{tgt} \\) simply by solving \\( X_{t}^{edit} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.669, + 0.483, + 0.708 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} X _ {t} ^ {e d i t} = X _ {t - 1} ^ {e d i t} + d X _ {t} ^ {t g t} - d X _ {t} ^ {s r c} \\\\ \\approx X _ {t - 1} ^ {e d i t} + \\mathbb {E} \\left[ v _ {t} \\left(\\hat {X} _ {t} ^ {t g t}\\right) - v _ {t} \\left(X _ {t} ^ {s r c}\\right) \\mid X _ {1} ^ {s r c} \\right] d t \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.711, + 0.484, + 0.759 + ], + "angle": 0, + "content": "where \\( X_{t}^{tgt} \\) is approximated as \\( \\hat{X}_{t}^{tgt} = X_{t - 1}^{edit} + X_{t}^{src} - X_{1}^{src} \\). To obtain the edited image, \\( X_{t}^{edit} \\) is iteratively updated from \\( \\epsilon \\) to 1, where \\( \\epsilon \\in [0,1) \\) represents skipped timesteps." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.765, + 0.314, + 0.781 + ], + "angle": 0, + "content": "4.2. RF-based Video Editing" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.786, + 0.484, + 0.817 + ], + "angle": 0, + "content": "This section introduces Pyramid-Edit, followed by a discussion of Wan-Edit, highlighting key features and differences." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.822, + 0.236, + 0.836 + ], + "angle": 0, + "content": "4.2.1. Pyramid-Edit" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Pyramid-Flow [14] employs a multi-resolution scheme across timesteps and a temporally autoregressive architecture to effectively handle the spatial-temporal complexity in video generative modeling. It decomposes the flow" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.461, + 0.907, + 0.553 + ], + "angle": 0, + "content": "into segments over \\(K\\) timestep windows within the interval [0,1]. Each window interpolates between successive solutions, reducing redundant computations in the earlier steps. For example, in the sampling process, within the \\(k\\)-th window \\([s_k, e_k]\\), it defines the rescaled timestep \\(t_k = (t - s_k) / (e_k - s_k)\\) and the corresponding flow is" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.565, + 0.907, + 0.581 + ], + "angle": 0, + "content": "\\[\nX _ {t _ {k}} = \\left(1 - t _ {k}\\right) \\bar {\\mathbf {x}} _ {s _ {k}} + t _ {k} \\bar {\\mathbf {x}} _ {e _ {k}}, \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.583, + 0.906, + 0.6 + ], + "angle": 0, + "content": "\\[\n\\bar {\\mathbf {x}} _ {s _ {k}} = \\left(1 - s _ {k}\\right) X _ {0} + s _ {k} U p (D o w n \\left(X _ {1}, 2 ^ {k + 1}\\right)), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.603, + 0.905, + 0.62 + ], + "angle": 0, + "content": "\\[\n\\bar {\\mathbf {x}} _ {e _ {k}} = \\left(1 - e _ {k}\\right) X _ {0} + e _ {k} D o w n \\left(X _ {1}, 2 ^ {k}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.632, + 0.907, + 0.723 + ], + "angle": 0, + "content": "where \\( \\text{Down}(\\cdot, \\cdot) \\) and \\( \\text{Up}(\\cdot) \\) represent the down-sampling and up-sampling, respectively. Meanwhile, the vector field in the \\( k \\)-th window is defined as \\( v_{t_k}(X_{t_k}) = \\bar{\\mathbf{x}}_{e_k} - \\bar{\\mathbf{x}}_{s_k} \\). To maintain continuity in the probability path within the pyramid structure, corrective noise is added at transition points between stages, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.734, + 0.907, + 0.76 + ], + "angle": 0, + "content": "\\[\nX _ {s _ {k - 1}} = \\frac {s _ {k - 1}}{e _ {k}} U p \\left(X _ {e _ {k}}\\right) + \\alpha \\mathbf {n} ^ {\\prime}, s. t. \\mathbf {n} _ {k - 1} \\sim \\mathcal {N} \\left(0, \\boldsymbol {\\Sigma} ^ {\\prime}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.771, + 0.906, + 0.831 + ], + "angle": 0, + "content": "where the rescaling coefficient \\( s_{k-1} / e_k \\) aligns the means of these distributions, the corrective noise \\( \\mathbf{n}_{k-1}' \\) weighted by \\( \\alpha \\) aligns their covariance matrices, and the covariance matrix \\( \\boldsymbol{\\Sigma}' \\) is related to the upsampling function." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.831, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Pyramid-Edit. The autoregressive architecture allows Pyramid-Edit to process frames sequentially, where previous frames serve as conditional information. However, the multi-resolution scheme across timesteps hinders direct use of FlowEdit. To address this, Pyramid-Edit integrates" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16676" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.414, + 0.106 + ], + "angle": 0, + "content": "FlowEdit (Eq. (3)) into each window as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.114, + 0.483, + 0.133 + ], + "angle": 0, + "content": "\\[\nX _ {t _ {k}} ^ {e d i t} = X _ {t _ {k} - 1} ^ {e d i t} + \\mathbb {E} [ v _ {t _ {k}} (\\hat {X} _ {t _ {k}} ^ {t g t}) - v _ {t _ {k}} (X _ {t _ {k}} ^ {s r c}) | \\bar {\\mathbf {x}} _ {e _ {k}} ^ {s r c} ] d t, \\qquad (8)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.14, + 0.483, + 0.17 + ], + "angle": 0, + "content": "where \\(\\hat{X}_{t_k}^{tgt} = X_{t_k - 1}^{edit} + X_{t_k}^{src} - \\bar{\\mathbf{x}}_{e_k}^{src}\\), and the target distribution is the end point of each window \\(\\bar{\\mathbf{x}}_{e_k}^{src}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.17, + 0.483, + 0.395 + ], + "angle": 0, + "content": "The workflow proceeds as illustrated in Fig. 3a and is structured as follows: starting from noise and the lowest-resolution latent, \\( X_{t_0}^{edit} \\) is initialized as \\( \\bar{\\mathbf{x}}_{e_0}^{src} \\), Eq. (8) is used to reconstruct and edit low-frequency information. At the end of the first stage, the reconstructed latent and the edited latent are upsampled to a higher resolution and passed through corrective noise as described in Eq. (7) to obtain the starting point for the next stage, \\( X_{s_1}^{edit} \\) and \\( X_{s_1}^{src} \\). However, FlowEdit operates within the target distribution, meaning we require the end point for this stage, \\( X_{e_1}^{edit} \\) and \\( X_{e_1}^{src} \\). The latter, \\( X_{e_1}^{src} \\), can be obtained through interpolation between \\( X_0 \\) and \\( X_1^{src} \\), i.e. \\( (X_{e_1}^{src} = \\bar{\\mathbf{x}}_{e_1}^{src}) \\). To estimate the edited endpoint \\( X_{e_1}^{edit} \\), we first compute the difference between the start and end points of the source prompt in this stage. This difference is then added to the edited start latent variable, providing an estimate of \\( X_{e_1}^{edit} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.402, + 0.473, + 0.421 + ], + "angle": 0, + "content": "\\[\n\\bar {\\mathbf {x}} _ {e _ {1}} ^ {s r c} = X _ {s _ {1}} ^ {s r c} + \\bigl (\\bar {\\mathbf {x}} _ {e _ {1}} ^ {s r c} - X _ {s _ {1}} ^ {s r c} \\bigr), X _ {e _ {1}} ^ {\\mathrm {e d i t}} = X _ {s _ {1}} ^ {\\mathrm {e d i t}} + \\bigl (\\bar {\\mathbf {x}} _ {e _ {1}} ^ {s r c} - X _ {s _ {1}} ^ {s r c} \\bigr).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.429, + 0.484, + 0.504 + ], + "angle": 0, + "content": "This step ensures that the edited and reconstructed latents remain aligned with the source image distribution while incorporating the desired modifications, preserving the integrity of the generated features. This is key to the success of FlowEdit under the pyramid structure." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.505, + 0.484, + 0.626 + ], + "angle": 0, + "content": "Another key point is that, starting from the second frame, Pyramid-Edit respectively incorporates the source and edited historical information from previous frames (e.g., the \\((i - 1)\\)-th frame) as a conditioning factor within its noise-estimation network. This allows the model to leverage temporal dependencies from both the source and edited frames, enhancing the temporal consistency. The source and edited sampling processes can be simplified as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.633, + 0.477, + 0.699 + ], + "angle": 0, + "content": "\\[\n\\underbrace {\\ldots \\rightarrow D o w n (X _ {1} ^ {s r c , i - 2} , 2 ^ {k + 1}) \\rightarrow D o w n (X _ {1} ^ {s r c , i - 1} , 2 ^ {k})} _ {\\mathrm {S o u r c e H i s t y o r c o n d i t i o n}} \\rightarrow X _ {t _ {k}} ^ {s r c , i},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.709, + 0.483, + 0.74 + ], + "angle": 0, + "content": "For generating subsequent frames, simply introduce the superscripts \\(i\\) as indices for the historical latents in Eq. (8)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.747, + 0.206, + 0.76 + ], + "angle": 0, + "content": "4.2.2. Wan-Edit" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Wan2.1 [40] is a state-of-the-art T2V generation model built on the mainstream DiT framework [26], achieving significant advancements through innovations such as a novel 3D VAE, scalable training strategies, large-scale data curation. These contributions enable Wan2.1 to generate high-quality, temporally consistent videos with improved efficiency and scalability. We focus on adapting its core framework for video editing, leveraging its strengths in temporal modeling and high-fidelity generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.303 + ], + "angle": 0, + "content": "Compared to Pyramid Flow, Wan2.1 adopts a much simpler architecture. It consists of 30 WanAttentionBlock layers, each integrating vision self-attention, text-to-vision cross-attention, and feed-forward mechanisms. Multiframe video inputs are encoded into a 3D vision latent by the 3D VAE encoder. These 3D vision latents are then flattened into a sequence and fed into the self-attention layers to encode temporal consistency. Finally, the sequence interacts with text tokens through cross-attention to achieve T2V generation. The self-attention mechanism captures intra-frame dependencies, while the cross-attention module integrates text tokens to enable text-guided video generation. This joint modeling of visual and textual information ensures high-quality, semantically aligned video outputs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.303, + 0.906, + 0.424 + ], + "angle": 0, + "content": "Wan-Edit. This elegant and streamlined architecture closely resembles that of image generation models, enabling FlowEdit to be directly integrated without significant modifications. For multi-frame inputs from \\( i_1 \\) to \\( i_2 \\), denote as \\( \\mathbf{X}_1^{src} = [X_1^{src,i_1},\\dots ,X_1^{src,i_2}] \\), the editing process handles all frames simultaneously by extending FlowEdit to the entire sequence, ensuring temporal consistency and efficient processing. The process is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.431, + 0.905, + 0.45 + ], + "angle": 0, + "content": "\\[\nd \\mathbf {X} _ {t} ^ {\\text {e d i t}} = \\mathbb {E} \\left[ v _ {t} ^ {\\Delta} \\left(\\mathbf {X} _ {t} ^ {\\text {s r c}}, \\mathbf {X} _ {t} ^ {\\text {e d i t}} + \\mathbf {X} _ {t} ^ {\\text {s r c}} - \\mathbf {X} _ {1} ^ {\\text {s r c}}\\right) \\mid \\mathbf {X} _ {1} ^ {\\text {s r c}} \\right] d t \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.456, + 0.905, + 0.486 + ], + "angle": 0, + "content": "This allows for smooth and coherent edits across all frames, preserving the temporal dynamics of the video." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.498, + 0.646, + 0.515 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.906, + 0.629 + ], + "angle": 0, + "content": "Baseline methods. We compare our RF-based editing methods with five diffusion-based models: TokenFlow [9], DMT [51], VidToMe [21], AnyV2V [17], and VideoGrain [50]. Diffusion-based models require inversion before editing, while our RF-based methods are inversion-free. DMT relies on optimization during editing, and VideoGrain uses depth maps to assist editing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.906, + 0.795 + ], + "angle": 0, + "content": "Evaluation metrics. We fairly evaluate all models on our FiVE-Bench using 15 metrics. The nine commonly used metrics in Table 2 cover six aspects: structure distance [38], background preservation (PSNR, LPIPS [52], MSE, and SSIM [42] outside the editing mask), edit prompt-image consistency (CLIPSIM [44] for the full image and masked regions), image quality assessment (NIQE [33]), temporal consistency (motion fidelity score [51]) and running time. Additionally, our proposed FiVE-Acc introduces five metrics in Table. 3 to evaluate the accuracy of successful object editing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.871 + ], + "angle": 0, + "content": "Implementation details. Pyramid-Edit uses the 384P Pyramid-Flow model, and Wan-Edit on the Wan2.1 1.3B model. Pyramid-Edit uses 20 timesteps for the first frame and 10 for subsequent frames, while Wan-Edit uses 50 timesteps. All experiments were run on a single H100 GPU." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Due to space limitations, see the Appendix for details on comparison methods, metrics, and implementation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16677" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 2. Comparison of diffusion- and flow-based video editing methods on our proposed FiVE-Bench. * and † denote methods that require optimization and depth/segmentation maps, respectively." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.122, + 0.901, + 0.279 + ], + "angle": 0, + "content": "
MethodsStructure Dist.×103↓Background PreservationText AlignmentIQA NIQE↓Motion Fidelity S.×102↑Time (s) Per Frame↓
PSNR↑LPIPS×103↓MSE×104↓SSIM×102↑CLIPS↑CLIPS-edit ↑
Source Videos0.00∞0.000.00100.0024.5919.876.3393.76-
TokenFlow [9]35.6219.06263.61138.6572.5126.4621.154.0189.008.04
DMT* [51]85.9514.71404.60372.7851.6426.6621.445.2482.3025.98
VidToMe [21]22.3721.15263.9188.7570.6926.8421.054.6890.063.25
AnyV2V [17]71.3615.90348.59342.9750.7724.8919.725.0460.366.11
VideoGrain† [50]12.4027.05185.2125.1079.1325.6920.314.0888.5727.12
RFs (Ours)Pyramid-Edit28.6520.84276.5995.6371.7226.8220.205.4880.591.44
Wan-Edit12.5325.5794.6141.8482.5526.3921.236.5489.433.07
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.301, + 0.483, + 0.33 + ], + "angle": 0, + "content": "Table 3. Comparison of diffusion- and flow-based video editing methods on the FiVE-Bench using FiVE-Acc metrics." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.334, + 0.472, + 0.455 + ], + "angle": 0, + "content": "
MethodYN-AccMC-AccU-Acc∩-AccFiVE-Acc↑
TokenFlow [9]19.3635.5136.6818.1827.43
DMT* [51]34.7862.0662.9833.8648.42
VidToMe [21]20.0333.5036.2017.3426.77
AnyV2V [17]30.6245.4248.9627.0938.02
VideoGrain† [50]30.5043.9744.3030.1737.23
Pyramid-Edit33.6754.0156.3631.3143.84
Wan-Edit41.4152.5355.7238.2246.97
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.483, + 0.325, + 0.5 + ], + "angle": 0, + "content": "5.1. Quantitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.507, + 0.483, + 0.734 + ], + "angle": 0, + "content": "Comparison on Common Metrics. From the results, we observe that VideoGrain achieves the best performance among diffusion models, in terms of Structure Distance (12.40) and Background Preservation (PSNR: 27.05, MSE:25.10). Meanwhile, Wan-Edit, our proposed method, outperforms in the other two Background Preservation metrics (LPIPS: 94.61, SSIM: 82.55). Overall, VideoGrain and Wan-Edit perform similarly, but VideoGrain requires depth map assistance for editing and is extremely slow. In contrast, the inversion-free Wan-Edit demonstrates the highest efficiency with the lowest Time per Frame (1.44 seconds). These findings highlight the strengths and trade-offs of each method, offering valuable insights for selecting the appropriate approach based on specific application needs, and greatly contribute to the advancement of video editing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Comparison on FiVE-Acc Metrics. Our evaluation of diffusion- and flow-based video editing methods on FiVE-Acc metrics reveals key findings. DMT, an optimization-based method, excels in MC-Acc (62.06%), U-Acc (62.98%), and overall FiVE-Acc (48.42%), highlighting its strength in fine-grained text alignment. This is due to its iterative optimization process, which prioritizes text fidelity, sometimes at the cost of background preservation. In contrast, Wan-Edit performs well in YN-Acc (41.41%) and \\(\\cap\\)-Acc (38.22%). Pyramid-Edit also shows competitive results, especially in MC-Acc (54.01%) and U" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.301, + 0.905, + 0.315 + ], + "angle": 0, + "content": "Table 4. Human vs. VLM model evaluation on FiVE-Acc metrics." + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.32, + 0.892, + 0.39 + ], + "angle": 0, + "content": "
EvaluatorYN-AccMC-Acc∪-Acc∩-AccFiVE-Acc↑
Qwen-2.5-VL41.4152.5355.7238.2246.97
Human44.3750.3151.9842.7047.34
Δ (Diff)+2.96-2.22-3.74+4.48+0.37
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.415, + 0.905, + 0.476 + ], + "angle": 0, + "content": "Acc (56.36%), emphasizing the importance of fine-grained alignment. U-Acc and \\(\\cap\\)-Acc metrics offer insights into task consistency, emphasizing the need for VLM-based evaluation to ensure visual and semantic coherence in editing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.476, + 0.906, + 0.55 + ], + "angle": 0, + "content": "Overall, VideoGrain and Wan-Edit strike the best balance across metrics for fine-grained video editing. Wan-Edit outperforms VideoGrain in speed, offering faster and more stable results due to enhanced RF video generation and inversion models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.551, + 0.906, + 0.643 + ], + "angle": 0, + "content": "Human Preference Validation for FiVE-Acc. We conducted a human study on 16 randomly sampled videos with 64 prompts, based on Wan-Edit editing outputs. Table 4 shows that human scores closely match Qwen-2.5-VL's results (47.34 vs. 46.97), confirming FiVE-Acc aligns well with human perception and offers consistent evaluation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.652, + 0.736, + 0.669 + ], + "angle": 0, + "content": "5.2. Qualitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.675, + 0.907, + 0.902 + ], + "angle": 0, + "content": "In Fig. 4, analyzing by edit type, non-rigid transformations and color changes are the easiest fine-grained edits, successfully handled by almost all methods. Object addition and rigid transformations are only successfully performed by the optimization-based DMT and our Wan-Edit. All methods fail at object removal, making it the most challenging task. From the perspective of background preservation in fine-grained editing, DMT and VidToME introduce significant alterations, while the recently proposed VideoGrain and the RF-based Pyramid-Edit and Wan-Edit largely retain the original background. The qualitative analysis further validates the quantitative findings: DMT achieves the best text alignment but disrupts the background, while VideoGrain and Wan-Edit offer the best balance across all metrics. Please refer to Appendix for details." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16678" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.107, + 0.091, + 0.383, + 0.102 + ], + "angle": 0, + "content": "Edit1 Object (w/o non-rigid deform): Woman \\(\\rightarrow\\) Man" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.105, + 0.896, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.217, + 0.896, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.317, + 0.295, + 0.326 + ], + "angle": 0, + "content": "Edit3 Color: black dress \\(\\rightarrow\\) red dress" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.328, + 0.896, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.429, + 0.351, + 0.439 + ], + "angle": 0, + "content": "Edit4 Material: A woman \\(\\rightarrow\\) A porcelain woman" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.441, + 0.896, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.559, + 0.414, + 0.571 + ], + "angle": 0, + "content": "EditS Add: A woman \\(\\rightarrow\\) A woman followed closely by a dog" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.572, + 0.896, + 0.686 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.264, + 0.688, + 0.731, + 0.703 + ], + "angle": 0, + "content": "Figure 4. Editing results across six editing types and five comparison methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.729, + 0.483, + 0.837 + ], + "angle": 0, + "content": "Limitations and Future Work. This work simply adapts image editing to video editing, which leads to room for improvement in the quality of Pyramid-Edit. In future work, we aim to further explore advancements in RF models for video editing. Additionally, handling challenging cases in the benchmark, such as large motions or object removal, will be an important direction for improvement." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.21, + 0.864 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We introduce FiVE-Bench, a benchmark for fine-grained video editing, and propose the VLM-based FiVE-Acc met" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.729, + 0.907, + 0.896 + ], + "angle": 0, + "content": "ric, which evaluates the accuracy of object-level editing success. Additionally, we adapt two RF-based video editing methods, Pyramid-Edit and Wan-Edit. To the best of our knowledge, this is the first comprehensive quantitative and qualitative comparison of emerging diffusion-based and flow-based models in the video editing community. FiVE-Bench and FiVE-Acc provide a standardized framework for fine-grained video editing evaluation, guiding future advancements in efficient and high-fidelity video editing solutions. We hope this benchmark drives innovation in both research and real-world applications." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16679" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.185, + 0.484, + 0.24 + ], + "angle": 0, + "content": "[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In IEEE Conf. Comput. Vis. Pattern Recog., pages 18392-18402, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.242, + 0.483, + 0.283 + ], + "angle": 0, + "content": "[3] Duygu Ceylan, Chun-Hao P Huang, and Niloy J Mitra. Pix2video: Video editing using image diffusion. In Int. Conf. Comput. Vis., pages 23206-23217, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.284, + 0.482, + 0.338 + ], + "angle": 0, + "content": "[4] Weifeng Chen, Jie Wu, Pan Xie, Hefeng Wu, Jiashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-a-video: Controllable text-to-video generation with diffusion models. arXiv preprint arXiv:2305.13840, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.34, + 0.483, + 0.395 + ], + "angle": 0, + "content": "[5] Yingying Deng, Xiangyu He, Changwang Mei, Peisong Wang, and Fan Tang. Fireflow: Fast inversion of rectified flow for image semantic editing. arXiv preprint arXiv:2412.07517, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.397, + 0.482, + 0.437 + ], + "angle": 0, + "content": "[6] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Adv. Neural Inform. Process. Syst., 34:8780-8794, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.439, + 0.483, + 0.493 + ], + "angle": 0, + "content": "[7] Wenkai Dong, Song Xue, Xiaoyue Duan, and Shumin Han. Prompt tuning inversion for text-driven image editing using diffusion models. In Int. Conf. Comput. Vis., pages 7430-7440, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.495, + 0.482, + 0.564 + ], + "angle": 0, + "content": "[8] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Int. Conf. on Machine Learning, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.482, + 0.606 + ], + "angle": 0, + "content": "[9] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. In Int. Conf. Learn. Represent., 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.607, + 0.482, + 0.662 + ], + "angle": 0, + "content": "[10] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff:Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.663, + 0.482, + 0.704 + ], + "angle": 0, + "content": "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Adv. Neural Inform. Process. Syst., 33:6840-6851, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.705, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[12] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[13] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[14] Yang Jin, Zhicheng Sun, Ningyuan Li, Kun Xu, Hao Jiang, Nan Zhuang, Quzhe Huang, Yang Song, Yadong Mu, and Zhouchen Lin. Pyramidal flow matching for efficient video generative modeling. arXiv preprint arXiv:2410.05954, 2024. 1, 2, 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[15] Xuan Ju, Ailing Zeng, Yuxuan Bian, Shaoteng Liu, and Qiang Xu. Pnp inversion: Boosting diffusion-based editing with 3 lines of code. International Conference on Learning Representations (ICLR), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[16] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.221, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[17] Max Ku, Cong Wei, Weiming Ren, Harry Yang, and Wenhu Chen. Anyv2v: A tuning-free framework for any video-to-video editing tasks. arXiv preprint arXiv:2403.14468, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[18] Vladimir Kulikov, Matan Kleiner, Inbar Huberman-Spiegelglas, and Tomer Michaeli. Flowedit: Inversion-free text-based editing using pre-trained flow models. arXiv preprint arXiv:2412.08629, 2024. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.904, + 0.36 + ], + "angle": 0, + "content": "[19] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux.2,3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[20] Minghan Li, Shuai Li, Xindong Zhang, and Lei Zhang. Univs: Unified and universal video segmentation with prompts as queries. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3227-3238, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[21] Xirui Li, Chao Ma, Xiaokang Yang, and Ming-Hsuan Yang. Vidthome: Video token merging for zero-shot video editing. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7486-7495, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[22] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. In Int. Conf. Learn. Represent., 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[23] Xingchao Liu, Chengyue Gong, et al. Flow straight and fast: Learning to generate and transfer data with rectified flow. In Int. Conf. Learn. Represent., 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.562, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[24] Guoqing Ma, Haoyang Huang, Kun Yan, Liangyu Chen, Nan Duan, Shengming Yin, Changyi Wan, Ranchen Ming, Xiaoniu Song, Xing Chen, et al. Step-video-t2v technical report: The practice, challenges, and future of video foundation model. arXiv preprint arXiv:2502.10248, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[25] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6038–6047, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.904, + 0.716 + ], + "angle": 0, + "content": "[26] OpenAI. Video generation models as world simulators, 2024. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.718, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[27] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[28] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. arXiv preprint arXiv:2303.09535, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[29] Bosheng Qin, Juncheng Li, Siliang Tang, Tat-Seng Chua, and Yueting Zhuang. Instructvid2vid: Controllable video editing with natural language instructions. In 2024 IEEE" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16680" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.217 + ], + "angle": 0, + "content": "[30] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dólar, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.482, + 0.276 + ], + "angle": 0, + "content": "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10684-10695, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.344 + ], + "angle": 0, + "content": "[32] Litu Rout, Yujia Chen, Nataniel Ruiz, Constantine Caramanis, Sanjay Shakkottai, and Wen-Sheng Chu. Semantic image inversion and editing using rectified stochastic differential equations. arXiv preprint arXiv:2410.10792, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.347, + 0.482, + 0.416 + ], + "angle": 0, + "content": "[33] Michele A Saad and Alan C Bovik. Blind quality assessment of videos using a model of natural scene statistics and motion coherency. In 2012 Conference Record of the Forty Sixth Asilomar Conference on Signals, Systems and Computers (ASILOMAR), pages 332-336. IEEE, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.418, + 0.482, + 0.474 + ], + "angle": 0, + "content": "[34] Uriel Singer, Amit Zohar, Yuval Kirstain, Shelly Sheynin, Adam Polyak, Devi Parikh, and Yaniv Taigman. Video editing via factorized diffusion distillation. In Eur. Conf. Comput. Vis., pages 450-466. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[35] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.48, + 0.546 + ], + "angle": 0, + "content": "[36] Spencer Sterling. Zeroscope. https://huggingface.co/cerspense/zeroscope_v2_576w, 2023.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.603 + ], + "angle": 0, + "content": "[37] Shuyuan Tu, Qi Dai, Zhi-Qi Cheng, Han Hu, Xintong Han, Zuxuan Wu, and Yu-Gang Jiang. Motioneditor: Editing video motion via content-aware diffusion. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7882-7891, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.605, + 0.482, + 0.659 + ], + "angle": 0, + "content": "[38] Narek Tumanyan, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Splicing vit features for semantic appearance transfer. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10748-10757, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[39] Bram Wallace, Akash Gokul, and Nikhil Naik. Edict: Exact diffusion inversion via coupled transformations. In IEEE Conf. Comput. Vis. Pattern Recog., pages 22532-22541, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[40] Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, et al. Wan: Open and advanced large-scale video generative models. arXiv preprint arXiv:2503.20314, 2025. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[41] Jiangshan Wang, Junfu Pu, Zhongang Qi, Jiayi Guo, Yue Ma, Nisha Huang, Yuxin Chen, Xiu Li, and Ying Shan. Taming rectified flow for inversion and editing. arXiv preprint arXiv:2411.04746, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[42] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process., 13(4): 600-612, 2004. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[43] Yujie Wei, Shiwei Zhang, Zhiwu Qing, Hangjie Yuan, Zhiheng Liu, Yu Liu, Yingya Zhang, Jingren Zhou, and Hongming Shan. Dreamvideo: Composing your dream videos with customized subject and motion. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6537-6549, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[44] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. GODIVA: Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[45] Chen Henry Wu and Fernando De la Torre. A latent space of stochastic diffusion models for zero-shot image editing and guidance. In Int. Conf. Comput. Vis., pages 7378-7387, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.345 + ], + "angle": 0, + "content": "[46] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Int. Conf. Comput. Vis., pages 7623-7633, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.347, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[47] Jay Zhangjie Wu, Xiuyu Li, Difei Gao, Zhen Dong, Jinbin Bai, Aishani Singh, Xiaoyu Xiang, Youzeng Li, Zuwei Huang, Yuanxi Sun, et al. Cvpr 2023 text guided video editing competition. arXiv preprint arXiv:2310.16003, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.404, + 0.905, + 0.458 + ], + "angle": 0, + "content": "[48] Chenxi Xie, Minghan Li, Shuai Li, Yuhui Wu, Qiaosi Yi, and Lei Zhang. Dnaedit: Direct noise alignment for text-guided rectified flow editing. arXiv preprint arXiv:2506.01430, 2025.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.46, + 0.905, + 0.515 + ], + "angle": 0, + "content": "[49] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. In SIGGRAPH Asia 2023 Conference Papers, pages 1-11, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.517, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[50] Xiangpeng Yang, Linchao Zhu, Hehe Fan, and Yi Yang. Videograin: Modulating space-time attention for multigrained video editing. arXiv preprint arXiv:2502.17258, 2025. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.574, + 0.905, + 0.628 + ], + "angle": 0, + "content": "[51] Danah Yatim, Rafail Fridman, Omer Bar-Tal, Yoni Kasten, and Tali Dekel. Space-time diffusion features for zero-shot text-driven motion transfer. In IEEE Conf. Comput. Vis. Pattern Recog., pages 8466-8476, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.63, + 0.905, + 0.685 + ], + "angle": 0, + "content": "[52] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conf. Comput. Vis. Pattern Recog., pages 586-595, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.687, + 0.905, + 0.728 + ], + "angle": 0, + "content": "[53] Zhenghao Zhang, Zuozhuo Dai, Long Qin, and Weizhi Wang. Effived: Efficient video editing via text-instruction diffusion models. arXiv preprint arXiv:2403.11568, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.729, + 0.905, + 0.797 + ], + "angle": 0, + "content": "[54] Zhixing Zhang, Yanyu Li, Yushu Wu, Yanwu Xu, Anil Kag, Ivan Skorokhodov, Willi Menapace, Aliaksandr Siarohin, Junli Cao, Dimitris Metaxas, et al. Sf-v: Single forward video generation model. arXiv preprint arXiv:2406.04324, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.8, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[55] Rui Zhao, Yuchao Gu, Jay Zhangjie Wu, David Junhao Zhang, Jia-Wei Liu, Weijia Wu, Jussi Keppo, and Mike Zheng Shou. Motiondirector: Motion customization of text-to-video diffusion models. In Eur. Conf. Comput. Vis., pages 273–290. Springer, 2025. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.87 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16681" + } + ] +] \ No newline at end of file diff --git a/2025/FiVE-Bench_ A Fine-grained Video Editing Benchmark for Evaluating Emerging Diffusion and Rectified Flow Models/61b80164-0e5a-42b8-b65d-d1598a219bbd_origin.pdf b/2025/FiVE-Bench_ A Fine-grained Video Editing Benchmark for Evaluating Emerging Diffusion and Rectified Flow Models/61b80164-0e5a-42b8-b65d-d1598a219bbd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..35c97a02a122520c160eb9b450cf1006ee0abf6b --- /dev/null +++ b/2025/FiVE-Bench_ A Fine-grained Video Editing Benchmark for Evaluating Emerging Diffusion and Rectified Flow Models/61b80164-0e5a-42b8-b65d-d1598a219bbd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a06e013479a2800950966e641f486b4c7526ea5674729b38ee0337c861f45790 +size 12600320 diff --git a/2025/FiVE-Bench_ A Fine-grained Video Editing Benchmark for Evaluating Emerging Diffusion and Rectified Flow Models/full.md b/2025/FiVE-Bench_ A Fine-grained Video Editing Benchmark for Evaluating Emerging Diffusion and Rectified Flow Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..aed4f57464fff45e8624fc1cd37c9cf788690e4c --- /dev/null +++ b/2025/FiVE-Bench_ A Fine-grained Video Editing Benchmark for Evaluating Emerging Diffusion and Rectified Flow Models/full.md @@ -0,0 +1,333 @@ +# FiVE-Bench: A Fine-grained Video Editing Benchmark for Evaluating Emerging Diffusion and Rectified Flow Models + +Minghan Li $^{1*}$ , Chenxi Xie $^{2*}$ , Yichen Wu $^{1,3}$ , Lei Zhang $^{2}$ and Mengyu Wang $^{1\dagger}$ $^{1}$ Harvard University + +$^{2}$ Hong Kong Polytechnic University, $^{3}$ City University of Hong Kong + +{mili4, mengyu_wang}@meei.harvard.edu + +# FiVE-Dataset + +![](images/824625187b3e53d65c739fdb8bb3993b55521516edb3e65a20d335514534c250.jpg) +Figure 1. Overview of the proposed FiVE-Bench, with editing results from Wan-Edit and corresponding FiVE-Acc scores. + +# Wan-Edit Results and FiVE-Evaluation + +![](images/0ef43fd5061b1234cc2b78d174f485e754cf24259bd6c7c50c28b0caace1be8a.jpg) + +# Abstract + +Numerous text-to-video (T2V) editing methods have emerged recently, but the lack of a standardized benchmark for fair evaluation has led to inconsistent claims and an inability to assess model sensitivity to hyperparameters. Fine-grained video editing is crucial for enabling precise, object-level modifications while maintaining context and temporal consistency. To address this, we introduce FiVE-Bench, a Fine-grained Video Editing Benchmark for evaluating emerging diffusion and rectified flow models. Our benchmark includes 74 real-world videos and 26 generated videos, featuring 6 fine-grained editing types, 420 object-level editing prompt pairs, and their corresponding masks. Additionally, we adapt the latest rectified flow (RF) T2V generation models—Pyramid-Flow [14] and Wan2.1 [40]—by introducing FlowEdit [18], resulting in training-free and inversion-free video editing mod + +els Pyramid-Edit and Wan-Edit. We compare five diffusion methods with our two RF methods on the proposed FiVE-Bench, evaluating them across 15 metrics. These metrics include background preservation, text-video similarity, temporal consistency, and generated video quality. To further enhance object-level evaluation, we introduce FiVE-Acc, a novel metric leveraging Vision-Language Models (VLMs) to assess the success of fine-grained video editing. Experimental results demonstrate that RF-based editing significantly outperforms diffusion-based methods, with WanEdit achieving the best overall performance and exhibiting the least sensitivity to hyperparameters. More video demo available on the website: https://sites.google.com/view/five-benchmark. + +# 1. Introduction + +Recent advancements in text-to-image (T2I) and text-to-video (T2V) generation [1, 14, 24, 26, 31, 36, 40] have + +led to the emergence of numerous image and video editing methods [4, 9, 10, 46], enabling users to create and modify video content with unprecedented flexibility. These methods leverage powerful generative models, such as diffusion models [6, 11, 12, 35] and Rectified Flow (RF) [8, 23] models, to synthesize and edit videos based on textual prompts. + +However, despite significant progress in these fields, the lack of a standardized benchmark for fair and comprehensive evaluation has still become a major bottleneck hindering further development. Currently, the video editing field has only one publicly available dataset, TGVE [47] /TGVE+ [34] (see Table 1), which includes 76 videos and 6 editing types, such as style transfer, background replacement, single-object and multi-object transformations, as well as object addition and removal. While TGVE+ aims to cover as many editing types as possible, its evaluation within each editing type is not comprehensive, particularly in supporting fine-grained video editing. Fine-grained video editing requires precise object-level modifications while preserving the overall context and temporal consistency of the video, posing higher demands on model performance. + +Due to the lack of a standardized benchmark and unified evaluation framework for fine-grained video editing, existing methods [9, 21, 28, 50] often rely on self-collected small-scale video datasets to validate their effectiveness. This approach not only makes it difficult to objectively compare the performance of different methods but also fails to comprehensively assess model robustness to hyperparameters or determine their suitability for real-world applications. Therefore, building a comprehensive and challenging benchmark for fine-grained video editing has become an urgent need to advance the field. + +To address these challenges, we introduce FiVE-Bench (Fine-grained Video Editing Benchmark), a comprehensive benchmark designed to evaluate emerging diffusion and rectified flow models for fine-grained video editing. FiVE-Bench includes 74 real-world videos and 26 generated videos, covering a diverse range of scenes and editing scenarios. It features 6 fine-grained editing types, 420 object-level editing prompt pairs, and their corresponding masks, providing a rich and challenging benchmark for video editing. Additionally, we propose FiVE-Acc, a novel evaluation metric that leverages Vision-Language Models (VLMs) to measure the success rate of fine-grained object-level editing. Unlike traditional similarity-based scores, FiVE-Acc focuses on evaluating whether the intended edit was successfully achieved, with particular emphasis on semantic alignment of the edited object. + +Recently, many image editing methods [5, 18, 32, 41] based on the latest rectified flow T2I generation model Flux [19] have achieved state-of-the-art editing results and fidelity, such as RF-Inversion [32], RF-Solver [41], and FlowEdit [18]. However, to the best of our knowledge, + +there are no video editing methods based on RF T2V generation models, primarily due to the relatively slow development of T2V generation models. Fortunately, the recent release of two state-of-the-art RF-based T2V generation models—Pyramid-Flow [14] and Wan2.1 [40]—has opened new possibilities for RF video editing. + +Building on the latest advancements, we adapt Pyramid-Flow and Wan2.1 by introducing FlowEdit [18]. This adaptation facilitates training-free and inversion-free video editing models, Pyramid-Edit and Wan-Edit, which leverage the strong temporal consistency inherent in video generation models without requiring additional attention mechanisms. Specifically, Pyramid-Flow employs a multi-resolution and temporally autoregressive model architecture across time steps. To adapt to this unique design, we modified FlowEdit to support multi-resolution processing, enabling seamless integration with the Pyramid-Flow framework, leading to Pyramid-Edit. On the other hand, Wan2.1 utilizes the MMDiT architecture, which unfolds temporal frames into sequences and processes them via self-attention. Since its design is fundamentally similar to image generation models, Wan2.1 can directly adopt FlowEdit without any additional modifications, giving rise to Wan-Edit. + +To show the effectiveness of the proposed RF-based editing methods, we compare them with five diffusion-based editing methods on the built FiVE-Bench, evaluating them across fifteen metrics, including background preservation, edited text-video similarity, temporal consistency, and generated video quality. Our experiments demonstrate that RF-based editing methods significantly outperform diffusion-based methods across multiple metrics. Notably, Wan-Edit achieves the best overall performance, exhibiting superior editing quality and the least sensitivity to hyperparameters. These findings highlight the potential of RF-based models for fine-grained video editing and provide valuable insights for future research in this area. + +In summary, this work makes four key contributions: + +- We introduce FiVE-Bench, a comprehensive benchmark for fine-grained video editing, featuring diverse videos, editing types, and evaluation metrics. +- We propose FiVE-Acc, a novel metric leveraging VLMs to assess the success of fine-grained editing. +- We propose the first two RF-based T2V editing models: Pyramid-Edit and Wan-Edit. +- We evaluate 7 video editing methods across 15 metrics, showing that RF-based editing consistently outperforms diffusion-based approaches. + +# 2. Related Work + +# 2.1. Diffusion and RF Inversion + +Diffusion inversion. Diffusion models (DMs) drive advances in inversion techniques for better control over gen + +Table 1. Statistics of existing video editing datasets and benchmarks. + +
DatasetUsageNum. ImagesNum. VideosFrames Per VideoNum. PromptsGen. VideosSrc. obj. WordsEdit obj. WordsObj. Instruct.Src. obj. MasksType of Edited Prompts
VIVID-10MTrain672K73.7K3010MXXX✓✓object, add, remove
Señorita-2MTrain0388.9K33~642MXXX✓Xstyle, object, multiple, color, add, remove, motion
TGVEEval07632304XXXXXstyle, object, bg, multiple
TGVE+Eval076321417XXXXXstyle, object, bg, multiple, color, add, remove
VIVID-10MEval06433~64852XXX✓✓object, add, remove
FiVE (Ours)Eval010035 ~ 126420✓✓✓✓✓object (w&w/o non-rigid), color, material, add, remove
+ +erated samples. DDIM [35] accelerates sampling but struggles with fine-grained reconstructions due to nonlinearities and score estimation errors. To improve accuracy, fine-tuning and optimization-based methods [7, 15, 25, 39, 45] have been proposed. While these advancements significantly enhance the fidelity and accuracy of DDIM inversion, the associated computational costs and resource demands remain substantial limiting factors. Although diffusion-based methods have achieved outstanding performance, recent studies [8, 14, 19, 32] suggest that Rectified Flow models (RFs) hold significant potential to surpass DMs in certain applications. Notably, the inversion and editing capabilities of RF models have been explored in image editing [5, 18, 32, 41, 48], demonstrating their efficiency and effectiveness. Building on these strengths, this work extends their potential in video editing. + +# 2.2. Video Editing + +Video editing techniques can generally be categorized into three types based on their training: training-free methods, one/few-shot finetuned methods [37, 43, 46, 55], and massive data finetuned methods. Training-free models [17, 21, 28, 50], enable task execution without retraining. TokenFlow [9] utilizes pre-trained models for efficient video synthesis, while DMT [51] transfers motion via diffusion. Other models like Pixel2Video [3], Render-A-Video [49], and Text2Video-zero [16] enhance accessibility for real-time video editing. One/few-shot fine-tuned models adapt pre-trained models for video editing with minimal fine-tuning. Methods like Tune-A-Video [46] and MotionEditor [37] achieve strong results using limited annotated samples, balancing customization and efficiency for task-specific editing without extensive retraining. + +Massive data fine-tuned models undergo extensive training on large datasets for high-quality, versatile video editing [29]. InstructVid2Vid [29] enables complex edits via natural language instructions, while EffiVED [53] refines broad datasets into high-quality subsets for efficient editing. SF-V [54] introduces adversarial training for video synthesis. Despite their precision and adaptability, these methods are computationally expensive, limiting practicality in resource-constrained settings. These methods focus on learning video motion, which may compromise back + +ground preservation. Recently, VideoGrain [50] has emphasized multi-granularity video editing rather than fine-grained edits, aiming to retain background details while modifying target objects. This work aims to address this issue and advance the field. + +# 3. FiVE Benchmark and FiVE-Acc Metric + +# 3.1. FiVE Benchmark + +We present the FiVE benchmark, which includes the FiVE video dataset and six fine-grained editing tasks. Additional details can be found in the Appendix. + +Videos in FiVE-Bench. We meticulously curate 74 real-world videos suitable for fine-grained editing from the DAVIS [27] dataset. Consecutive frames are extracted at every 8-frame interval, and the GPT-4o [13] is employed to generate formatted captions for these videos. These captions encompass details pertaining to object category, action, background, and camera movement. Additionally, during the annotation process, we document instances of object deformation (such as limb movements in humans or animals), which aids in differentiating the complexity of editing tasks during evaluation. To further enhance the diversity of our benchmark, we generate 26 highly realistic synthetic videos using the T2V model [40]. These synthetic videos not only expand the range of video categories but also enable a comparative analysis of editing performance across real and synthetic videos. + +Six Fine-grained Video Editing Tasks. FiVE-Bench comprises six fine-grained video editing tasks, totaling 420 high-quality editing prompt pairs. It primarily includes four object-targeted editing tasks of increasing complexity: object substitution without non-rigid deformation, object substitution with non-rigid deformation, color alteration, and material modification. For each task, we use GPT-4o [13] to generate four source-target prompt pairs per video in the benchmark, resulting in 400 high-quality editing pairs. To further enhance the diversity of editing types within the benchmark, we select 10 videos each for the add and remove tasks and design corresponding source-target prompt pairs. Additionally, we explicitly provide both object-related source words and edited words in the source and edited text prompts, respectively, denoted as + +![](images/64eeee1923f5e9435008803cc18b46ffb3f030ad222c11dd33d8538f81b93101.jpg) +Figure 2. The VLM-based FiVE accuracy (FiVE-Acc) evaluation pipeline. + +"src. obj. words" and "edit obj. words" in Table 1. To accommodate different editing methods, we use GPT-4o [13] to generate instruction prompts ("Obj. Instr." in Table 1) for each source-target pair, ensuring compatibility with models like InstructPix2Pix [2]. For each video, we employ SAM2 [20, 30] to generate masks of the edited regions, enabling the evaluation of background preservation metrics. + +# 3.2. FiVE-Acc Metric + +To evaluate the accuracy of fine-grained video editing, we propose FiVE accuracy (FiVE-Acc) evaluation, a framework assessing how precisely video editing models modify the target object. Fig. 2 presents the complete pipeline, comprising video editing and accuracy evaluation. + +Video editing outputs. Given a source video, a video editing model is prompted with a source prompt, describing the initial scene, and a target prompt, specifying the desired modification. The model generates a target video, where the intended transformation is applied, as shown in Fig. 2a. + +FiVE-Acc evaluation. To quantitatively evaluate the fidelity of video editing, we employ a Vision-Language Model (VLM) (e.g., Qwen-2.5-VL) to analyze the edited video. The evaluation includes two types of questions: + +- Yes/No Questions: The model is asked whether the source and target objects are present in the target video. Editing succeeds only if the model answers "No" for the source object and "Yes" for the target, leading to the calculation of $YN$ -Acc. The source object question ensures the model both removes the source and adds the target, refining accuracy evaluation. +- Multi-choice Question: It evaluates whether the VLM recognizes the target (e.g., flamingo) or source object (e.g., black swan) in the edited video. The model selects between the two objects to compute MC-Acc, measuring its recognition accuracy. + +Based on the two obtained accuracies, we further calculate the union and intersection accuracies to derive $\cup$ -Acc and $\cap$ -Acc. These offer a more detailed evaluation: the former reflects the model's overall editing success accuracy, while the latter highlights the model's high-quality editing + +success accuracy. $FiVE - Acc$ is the average of four components: $FiVE - Acc = 0.25 \times (YN - Acc + MC - Acc + \cup - Acc + \cap - Acc)$ . It uses the VLM's recognition ability to evaluate the alignment between the edited object and its real data distribution. To the best of our knowledge, we are the first to introduce accuracy evaluation in fine-grained T2V editing. + +# 4. Methods: Pyramid-Edit and Wan-Edit + +# 4.1. Preliminary + +Rectified Flow (RF) Models [22, 23]. Let $q_{0}$ denote the source distribution (i.e., standard Gaussian $\mathcal{N}(0,I)$ ), $p_{0}$ refer to the target distribution (i.e., the distribution over images), and $v_{t}(\cdot)$ is the time-varying vector field. RF models [22, 23] aim to transform $q_{0}$ progressively into $p_{0}$ by solving the following ordinary differential equation (ODE): + +$$ +d X _ {t} = v _ {t} \left(X _ {t}\right) d t, \quad t \in [ 0, 1 ]. \tag {1} +$$ + +Here, the vector field $v_{t}(X_{t}) = f(X_{t},c,1 - t;\varphi)$ , where $f(\cdot)$ is a neural network parameterized by $\varphi$ and $c$ is the guided text. Given the ODE in Eq. (1), we start with $X_0\sim q_0$ from the source distribution and integrate over $t:0\to 1$ to reach the endpoint $X_{1}\sim p_{0}$ , which is sampled from the target distribution. The forward process of RF assumes the linear path between the two states: $X_{t} = (1 - t)X_{0} + tX_{1}$ , then we can derive the vector field of $v_{t}(\cdot):u_{t}(X_{t}) = X_{1} - X_{0}$ . Then, $v_{t}(X_{t})$ is used to approximate the network $f(X_{t},c,t;\varphi)$ by minimizing the loss function $\mathcal{L}_{\varphi} = \mathbb{E}_{t\sim \mathcal{U}[0,1],Y_t}[\| v_t(X_t) - f(X_t,c,t;\varphi)\| _2^2 ]$ . RF ensures that their sampling paths are relatively straight, enabling the use of a small number of discretization steps. + +RF-based Image Editing: FlowEdit [18]. Image editing involves taking a source image $X_{1}^{src}$ and a source text prompt $c^{src}$ , and allowing the user to provide a target text prompt $c^{tgt}$ for targeted fine-grained editing to ultimately yield the edited image $X_{1}^{tgt}$ . For example, the source text prompt might be "A black swan drifting," while the target text prompt could be "A flamingo drifting." To simplify notation, we denote the vector fields corresponding to the source text prompt and the target text prompt as $v_{t}(X_{t}^{src})$ and $v_{t}(X_{t}^{tgt})$ , respectively. Their sampling process in RF + +![](images/36f1d30bd80c21ddad15542befab8a480f8acaf33ffe63c46e1f5ac8296a5f5f.jpg) +(a) Pyramid-Edit + +![](images/4bda0b0d40e70e4a2de0cc6e2e50891490ea6bbd0710f82eb3d74cd493d3426e.jpg) +(b) Wan-Edit +Figure 3. Two RF-based inversion-free video editing models: Pyramid-Edit and Wan-Edit. Pyramid-Edit is a multi-resolution temporal autoregressive architecture, while Wan-Edit (DiT architecture) treats temporal frames as a sequence and processing them simultaneously. + +follows: $dX_{t}^{src} = v_{t}(X_{t}^{src})dt$ , $dX_{t}^{tgt} = v_{t}(X_{t}^{tgt})dt$ . + +FlowEdit operates inversion-free, mapping source modes to the nearest target modes. It assumes target images follow straight trajectories, allowing latent interpolation. Thus, the nearest target mode is derived from the difference between source and target vector fields as + +$$ +X _ {t} ^ {e d i t} = X _ {1} ^ {s r c} + X _ {t} ^ {t g t} - X _ {t} ^ {s r c}. \tag {2} +$$ + +Here $X_{t}^{edit} = X_{t}^{tgt}$ , if $t \to 1$ . A more straightforward explanation is: $X_{t}^{edit}$ evolves within the target distribution, continuously transitioning through the clean latent space across all timesteps, (i.e. $X_{t}^{edit}$ gradually transforms from a black swan to a flamingo as the timestep progresses). Therefore, we can obtain $X_{1}^{tgt}$ simply by solving $X_{t}^{edit}$ : + +$$ +\begin{array}{l} X _ {t} ^ {e d i t} = X _ {t - 1} ^ {e d i t} + d X _ {t} ^ {t g t} - d X _ {t} ^ {s r c} \\ \approx X _ {t - 1} ^ {e d i t} + \mathbb {E} \left[ v _ {t} \left(\hat {X} _ {t} ^ {t g t}\right) - v _ {t} \left(X _ {t} ^ {s r c}\right) \mid X _ {1} ^ {s r c} \right] d t \tag {3} \\ \end{array} +$$ + +where $X_{t}^{tgt}$ is approximated as $\hat{X}_{t}^{tgt} = X_{t - 1}^{edit} + X_{t}^{src} - X_{1}^{src}$ . To obtain the edited image, $X_{t}^{edit}$ is iteratively updated from $\epsilon$ to 1, where $\epsilon \in [0,1)$ represents skipped timesteps. + +# 4.2. RF-based Video Editing + +This section introduces Pyramid-Edit, followed by a discussion of Wan-Edit, highlighting key features and differences. + +# 4.2.1. Pyramid-Edit + +Pyramid-Flow [14] employs a multi-resolution scheme across timesteps and a temporally autoregressive architecture to effectively handle the spatial-temporal complexity in video generative modeling. It decomposes the flow + +into segments over $K$ timestep windows within the interval [0,1]. Each window interpolates between successive solutions, reducing redundant computations in the earlier steps. For example, in the sampling process, within the $k$ -th window $[s_k, e_k]$ , it defines the rescaled timestep $t_k = (t - s_k) / (e_k - s_k)$ and the corresponding flow is + +$$ +X _ {t _ {k}} = \left(1 - t _ {k}\right) \bar {\mathbf {x}} _ {s _ {k}} + t _ {k} \bar {\mathbf {x}} _ {e _ {k}}, \tag {4} +$$ + +$$ +\bar {\mathbf {x}} _ {s _ {k}} = \left(1 - s _ {k}\right) X _ {0} + s _ {k} U p (D o w n \left(X _ {1}, 2 ^ {k + 1}\right)), \tag {5} +$$ + +$$ +\bar {\mathbf {x}} _ {e _ {k}} = \left(1 - e _ {k}\right) X _ {0} + e _ {k} D o w n \left(X _ {1}, 2 ^ {k}\right), \tag {6} +$$ + +where $\text{Down}(\cdot, \cdot)$ and $\text{Up}(\cdot)$ represent the down-sampling and up-sampling, respectively. Meanwhile, the vector field in the $k$ -th window is defined as $v_{t_k}(X_{t_k}) = \bar{\mathbf{x}}_{e_k} - \bar{\mathbf{x}}_{s_k}$ . To maintain continuity in the probability path within the pyramid structure, corrective noise is added at transition points between stages, as follows: + +$$ +X _ {s _ {k - 1}} = \frac {s _ {k - 1}}{e _ {k}} U p \left(X _ {e _ {k}}\right) + \alpha \mathbf {n} ^ {\prime}, s. t. \mathbf {n} _ {k - 1} \sim \mathcal {N} \left(0, \boldsymbol {\Sigma} ^ {\prime}\right), \tag {7} +$$ + +where the rescaling coefficient $s_{k-1} / e_k$ aligns the means of these distributions, the corrective noise $\mathbf{n}_{k-1}'$ weighted by $\alpha$ aligns their covariance matrices, and the covariance matrix $\boldsymbol{\Sigma}'$ is related to the upsampling function. + +Pyramid-Edit. The autoregressive architecture allows Pyramid-Edit to process frames sequentially, where previous frames serve as conditional information. However, the multi-resolution scheme across timesteps hinders direct use of FlowEdit. To address this, Pyramid-Edit integrates + +FlowEdit (Eq. (3)) into each window as follows: + +$$ +X _ {t _ {k}} ^ {e d i t} = X _ {t _ {k} - 1} ^ {e d i t} + \mathbb {E} [ v _ {t _ {k}} (\hat {X} _ {t _ {k}} ^ {t g t}) - v _ {t _ {k}} (X _ {t _ {k}} ^ {s r c}) | \bar {\mathbf {x}} _ {e _ {k}} ^ {s r c} ] d t, \qquad (8) +$$ + +where $\hat{X}_{t_k}^{tgt} = X_{t_k - 1}^{edit} + X_{t_k}^{src} - \bar{\mathbf{x}}_{e_k}^{src}$ , and the target distribution is the end point of each window $\bar{\mathbf{x}}_{e_k}^{src}$ . + +The workflow proceeds as illustrated in Fig. 3a and is structured as follows: starting from noise and the lowest-resolution latent, $X_{t_0}^{edit}$ is initialized as $\bar{\mathbf{x}}_{e_0}^{src}$ , Eq. (8) is used to reconstruct and edit low-frequency information. At the end of the first stage, the reconstructed latent and the edited latent are upsampled to a higher resolution and passed through corrective noise as described in Eq. (7) to obtain the starting point for the next stage, $X_{s_1}^{edit}$ and $X_{s_1}^{src}$ . However, FlowEdit operates within the target distribution, meaning we require the end point for this stage, $X_{e_1}^{edit}$ and $X_{e_1}^{src}$ . The latter, $X_{e_1}^{src}$ , can be obtained through interpolation between $X_0$ and $X_1^{src}$ , i.e. $(X_{e_1}^{src} = \bar{\mathbf{x}}_{e_1}^{src})$ . To estimate the edited endpoint $X_{e_1}^{edit}$ , we first compute the difference between the start and end points of the source prompt in this stage. This difference is then added to the edited start latent variable, providing an estimate of $X_{e_1}^{edit}$ : + +$$ +\bar {\mathbf {x}} _ {e _ {1}} ^ {s r c} = X _ {s _ {1}} ^ {s r c} + \bigl (\bar {\mathbf {x}} _ {e _ {1}} ^ {s r c} - X _ {s _ {1}} ^ {s r c} \bigr), X _ {e _ {1}} ^ {\mathrm {e d i t}} = X _ {s _ {1}} ^ {\mathrm {e d i t}} + \bigl (\bar {\mathbf {x}} _ {e _ {1}} ^ {s r c} - X _ {s _ {1}} ^ {s r c} \bigr). +$$ + +This step ensures that the edited and reconstructed latents remain aligned with the source image distribution while incorporating the desired modifications, preserving the integrity of the generated features. This is key to the success of FlowEdit under the pyramid structure. + +Another key point is that, starting from the second frame, Pyramid-Edit respectively incorporates the source and edited historical information from previous frames (e.g., the $(i - 1)$ -th frame) as a conditioning factor within its noise-estimation network. This allows the model to leverage temporal dependencies from both the source and edited frames, enhancing the temporal consistency. The source and edited sampling processes can be simplified as follows: + +$$ +\underbrace {\ldots \rightarrow D o w n (X _ {1} ^ {s r c , i - 2} , 2 ^ {k + 1}) \rightarrow D o w n (X _ {1} ^ {s r c , i - 1} , 2 ^ {k})} _ {\mathrm {S o u r c e H i s t y o r c o n d i t i o n}} \rightarrow X _ {t _ {k}} ^ {s r c , i}, +$$ + +For generating subsequent frames, simply introduce the superscripts $i$ as indices for the historical latents in Eq. (8). + +# 4.2.2. Wan-Edit + +Wan2.1 [40] is a state-of-the-art T2V generation model built on the mainstream DiT framework [26], achieving significant advancements through innovations such as a novel 3D VAE, scalable training strategies, large-scale data curation. These contributions enable Wan2.1 to generate high-quality, temporally consistent videos with improved efficiency and scalability. We focus on adapting its core framework for video editing, leveraging its strengths in temporal modeling and high-fidelity generation. + +Compared to Pyramid Flow, Wan2.1 adopts a much simpler architecture. It consists of 30 WanAttentionBlock layers, each integrating vision self-attention, text-to-vision cross-attention, and feed-forward mechanisms. Multiframe video inputs are encoded into a 3D vision latent by the 3D VAE encoder. These 3D vision latents are then flattened into a sequence and fed into the self-attention layers to encode temporal consistency. Finally, the sequence interacts with text tokens through cross-attention to achieve T2V generation. The self-attention mechanism captures intra-frame dependencies, while the cross-attention module integrates text tokens to enable text-guided video generation. This joint modeling of visual and textual information ensures high-quality, semantically aligned video outputs. + +Wan-Edit. This elegant and streamlined architecture closely resembles that of image generation models, enabling FlowEdit to be directly integrated without significant modifications. For multi-frame inputs from $i_1$ to $i_2$ , denote as $\mathbf{X}_1^{src} = [X_1^{src,i_1},\dots ,X_1^{src,i_2}]$ , the editing process handles all frames simultaneously by extending FlowEdit to the entire sequence, ensuring temporal consistency and efficient processing. The process is defined as follows: + +$$ +d \mathbf {X} _ {t} ^ {\text {e d i t}} = \mathbb {E} \left[ v _ {t} ^ {\Delta} \left(\mathbf {X} _ {t} ^ {\text {s r c}}, \mathbf {X} _ {t} ^ {\text {e d i t}} + \mathbf {X} _ {t} ^ {\text {s r c}} - \mathbf {X} _ {1} ^ {\text {s r c}}\right) \mid \mathbf {X} _ {1} ^ {\text {s r c}} \right] d t \tag {9} +$$ + +This allows for smooth and coherent edits across all frames, preserving the temporal dynamics of the video. + +# 5. Experiments + +Baseline methods. We compare our RF-based editing methods with five diffusion-based models: TokenFlow [9], DMT [51], VidToMe [21], AnyV2V [17], and VideoGrain [50]. Diffusion-based models require inversion before editing, while our RF-based methods are inversion-free. DMT relies on optimization during editing, and VideoGrain uses depth maps to assist editing. + +Evaluation metrics. We fairly evaluate all models on our FiVE-Bench using 15 metrics. The nine commonly used metrics in Table 2 cover six aspects: structure distance [38], background preservation (PSNR, LPIPS [52], MSE, and SSIM [42] outside the editing mask), edit prompt-image consistency (CLIPSIM [44] for the full image and masked regions), image quality assessment (NIQE [33]), temporal consistency (motion fidelity score [51]) and running time. Additionally, our proposed FiVE-Acc introduces five metrics in Table. 3 to evaluate the accuracy of successful object editing. + +Implementation details. Pyramid-Edit uses the 384P Pyramid-Flow model, and Wan-Edit on the Wan2.1 1.3B model. Pyramid-Edit uses 20 timesteps for the first frame and 10 for subsequent frames, while Wan-Edit uses 50 timesteps. All experiments were run on a single H100 GPU. + +Due to space limitations, see the Appendix for details on comparison methods, metrics, and implementation. + +Table 2. Comparison of diffusion- and flow-based video editing methods on our proposed FiVE-Bench. * and † denote methods that require optimization and depth/segmentation maps, respectively. + +
MethodsStructure Dist.×103↓Background PreservationText AlignmentIQA NIQE↓Motion Fidelity S.×102↑Time (s) Per Frame↓
PSNR↑LPIPS×103↓MSE×104↓SSIM×102↑CLIPS↑CLIPS-edit ↑
Source Videos0.00∞0.000.00100.0024.5919.876.3393.76-
TokenFlow [9]35.6219.06263.61138.6572.5126.4621.154.0189.008.04
DMT* [51]85.9514.71404.60372.7851.6426.6621.445.2482.3025.98
VidToMe [21]22.3721.15263.9188.7570.6926.8421.054.6890.063.25
AnyV2V [17]71.3615.90348.59342.9750.7724.8919.725.0460.366.11
VideoGrain† [50]12.4027.05185.2125.1079.1325.6920.314.0888.5727.12
RFs (Ours)Pyramid-Edit28.6520.84276.5995.6371.7226.8220.205.4880.591.44
Wan-Edit12.5325.5794.6141.8482.5526.3921.236.5489.433.07
+ +Table 3. Comparison of diffusion- and flow-based video editing methods on the FiVE-Bench using FiVE-Acc metrics. + +
MethodYN-AccMC-AccU-Acc∩-AccFiVE-Acc↑
TokenFlow [9]19.3635.5136.6818.1827.43
DMT* [51]34.7862.0662.9833.8648.42
VidToMe [21]20.0333.5036.2017.3426.77
AnyV2V [17]30.6245.4248.9627.0938.02
VideoGrain† [50]30.5043.9744.3030.1737.23
Pyramid-Edit33.6754.0156.3631.3143.84
Wan-Edit41.4152.5355.7238.2246.97
+ +# 5.1. Quantitative Comparison + +Comparison on Common Metrics. From the results, we observe that VideoGrain achieves the best performance among diffusion models, in terms of Structure Distance (12.40) and Background Preservation (PSNR: 27.05, MSE:25.10). Meanwhile, Wan-Edit, our proposed method, outperforms in the other two Background Preservation metrics (LPIPS: 94.61, SSIM: 82.55). Overall, VideoGrain and Wan-Edit perform similarly, but VideoGrain requires depth map assistance for editing and is extremely slow. In contrast, the inversion-free Wan-Edit demonstrates the highest efficiency with the lowest Time per Frame (1.44 seconds). These findings highlight the strengths and trade-offs of each method, offering valuable insights for selecting the appropriate approach based on specific application needs, and greatly contribute to the advancement of video editing. + +Comparison on FiVE-Acc Metrics. Our evaluation of diffusion- and flow-based video editing methods on FiVE-Acc metrics reveals key findings. DMT, an optimization-based method, excels in MC-Acc (62.06%), U-Acc (62.98%), and overall FiVE-Acc (48.42%), highlighting its strength in fine-grained text alignment. This is due to its iterative optimization process, which prioritizes text fidelity, sometimes at the cost of background preservation. In contrast, Wan-Edit performs well in YN-Acc (41.41%) and $\cap$ -Acc (38.22%). Pyramid-Edit also shows competitive results, especially in MC-Acc (54.01%) and U + +Table 4. Human vs. VLM model evaluation on FiVE-Acc metrics. + +
EvaluatorYN-AccMC-Acc∪-Acc∩-AccFiVE-Acc↑
Qwen-2.5-VL41.4152.5355.7238.2246.97
Human44.3750.3151.9842.7047.34
Δ (Diff)+2.96-2.22-3.74+4.48+0.37
+ +Acc (56.36%), emphasizing the importance of fine-grained alignment. U-Acc and $\cap$ -Acc metrics offer insights into task consistency, emphasizing the need for VLM-based evaluation to ensure visual and semantic coherence in editing. + +Overall, VideoGrain and Wan-Edit strike the best balance across metrics for fine-grained video editing. Wan-Edit outperforms VideoGrain in speed, offering faster and more stable results due to enhanced RF video generation and inversion models. + +Human Preference Validation for FiVE-Acc. We conducted a human study on 16 randomly sampled videos with 64 prompts, based on Wan-Edit editing outputs. Table 4 shows that human scores closely match Qwen-2.5-VL's results (47.34 vs. 46.97), confirming FiVE-Acc aligns well with human perception and offers consistent evaluation. + +# 5.2. Qualitative Comparison + +In Fig. 4, analyzing by edit type, non-rigid transformations and color changes are the easiest fine-grained edits, successfully handled by almost all methods. Object addition and rigid transformations are only successfully performed by the optimization-based DMT and our Wan-Edit. All methods fail at object removal, making it the most challenging task. From the perspective of background preservation in fine-grained editing, DMT and VidToME introduce significant alterations, while the recently proposed VideoGrain and the RF-based Pyramid-Edit and Wan-Edit largely retain the original background. The qualitative analysis further validates the quantitative findings: DMT achieves the best text alignment but disrupts the background, while VideoGrain and Wan-Edit offer the best balance across all metrics. Please refer to Appendix for details. + +Edit1 Object (w/o non-rigid deform): Woman $\rightarrow$ Man + +![](images/d51bc72ab52e0991c3f6613e73093db934209f0d905111b10dd0ad72cc5eb293.jpg) + +![](images/2249861f1ef7256f5797f79bfde2c1e68c1c9201a597c0cca2fcf0c230bf1a1b.jpg) + +Edit3 Color: black dress $\rightarrow$ red dress + +![](images/c5fcfe4470254a9f0e1cd6ab8f3a6d25ede2b922bc75fbb2e84ea6404bedd489.jpg) + +Edit4 Material: A woman $\rightarrow$ A porcelain woman + +![](images/ae21f1ee99f227fd8b908a2e08fb78d8d8b23aea32d714737009a142f27a5871.jpg) + +EditS Add: A woman $\rightarrow$ A woman followed closely by a dog + +![](images/844b9bde26cec22be6b4247f746b8092d15081a3412e02d36987849bf727070a.jpg) +Figure 4. Editing results across six editing types and five comparison methods. + +Limitations and Future Work. This work simply adapts image editing to video editing, which leads to room for improvement in the quality of Pyramid-Edit. In future work, we aim to further explore advancements in RF models for video editing. Additionally, handling challenging cases in the benchmark, such as large motions or object removal, will be an important direction for improvement. + +# 6. Conclusion + +We introduce FiVE-Bench, a benchmark for fine-grained video editing, and propose the VLM-based FiVE-Acc met + +ric, which evaluates the accuracy of object-level editing success. Additionally, we adapt two RF-based video editing methods, Pyramid-Edit and Wan-Edit. To the best of our knowledge, this is the first comprehensive quantitative and qualitative comparison of emerging diffusion-based and flow-based models in the video editing community. FiVE-Bench and FiVE-Acc provide a standardized framework for fine-grained video editing evaluation, guiding future advancements in efficient and high-fidelity video editing solutions. We hope this benchmark drives innovation in both research and real-world applications. + +# References + +[1] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 1 +[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In IEEE Conf. Comput. Vis. Pattern Recog., pages 18392-18402, 2023. 4 +[3] Duygu Ceylan, Chun-Hao P Huang, and Niloy J Mitra. Pix2video: Video editing using image diffusion. In Int. Conf. Comput. Vis., pages 23206-23217, 2023. 3 +[4] Weifeng Chen, Jie Wu, Pan Xie, Hefeng Wu, Jiashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-a-video: Controllable text-to-video generation with diffusion models. arXiv preprint arXiv:2305.13840, 2023. 2 +[5] Yingying Deng, Xiangyu He, Changwang Mei, Peisong Wang, and Fan Tang. Fireflow: Fast inversion of rectified flow for image semantic editing. arXiv preprint arXiv:2412.07517, 2024. 2, 3 +[6] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Adv. Neural Inform. Process. Syst., 34:8780-8794, 2021. 2 +[7] Wenkai Dong, Song Xue, Xiaoyue Duan, and Shumin Han. Prompt tuning inversion for text-driven image editing using diffusion models. In Int. Conf. Comput. Vis., pages 7430-7440, 2023. 3 +[8] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Int. Conf. on Machine Learning, 2024. 2, 3 +[9] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. In Int. Conf. Learn. Represent., 2024. 2, 3, 6, 7 +[10] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff:Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 2 +[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Adv. Neural Inform. Process. Syst., 33:6840-6851, 2020. 2 +[12] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2 +[13] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 3, 4 +[14] Yang Jin, Zhicheng Sun, Ningyuan Li, Kun Xu, Hao Jiang, Nan Zhuang, Quzhe Huang, Yang Song, Yadong Mu, and Zhouchen Lin. Pyramidal flow matching for efficient video generative modeling. arXiv preprint arXiv:2410.05954, 2024. 1, 2, 3, 5 + +[15] Xuan Ju, Ailing Zeng, Yuxuan Bian, Shaoteng Liu, and Qiang Xu. Pnp inversion: Boosting diffusion-based editing with 3 lines of code. International Conference on Learning Representations (ICLR), 2024. 3 +[16] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439, 2023. 3 +[17] Max Ku, Cong Wei, Weiming Ren, Harry Yang, and Wenhu Chen. Anyv2v: A tuning-free framework for any video-to-video editing tasks. arXiv preprint arXiv:2403.14468, 2024. 3, 6, 7 +[18] Vladimir Kulikov, Matan Kleiner, Inbar Huberman-Spiegelglas, and Tomer Michaeli. Flowedit: Inversion-free text-based editing using pre-trained flow models. arXiv preprint arXiv:2412.08629, 2024. 1, 2, 3, 4 +[19] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux.2,3 +[20] Minghan Li, Shuai Li, Xindong Zhang, and Lei Zhang. Univs: Unified and universal video segmentation with prompts as queries. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3227-3238, 2024. 4 +[21] Xirui Li, Chao Ma, Xiaokang Yang, and Ming-Hsuan Yang. Vidthome: Video token merging for zero-shot video editing. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7486-7495, 2024. 2, 3, 6, 7 +[22] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. In Int. Conf. Learn. Represent., 2023. 4 +[23] Xingchao Liu, Chengyue Gong, et al. Flow straight and fast: Learning to generate and transfer data with rectified flow. In Int. Conf. Learn. Represent., 2023. 2, 4 +[24] Guoqing Ma, Haoyang Huang, Kun Yan, Liangyu Chen, Nan Duan, Shengming Yin, Changyi Wan, Ranchen Ming, Xiaoniu Song, Xing Chen, et al. Step-video-t2v technical report: The practice, challenges, and future of video foundation model. arXiv preprint arXiv:2502.10248, 2025. 1 +[25] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6038–6047, 2023. 3 +[26] OpenAI. Video generation models as world simulators, 2024. 1, 6 +[27] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. 3 +[28] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. arXiv preprint arXiv:2303.09535, 2023. 2, 3 +[29] Bosheng Qin, Juncheng Li, Siliang Tang, Tat-Seng Chua, and Yueting Zhuang. Instructvid2vid: Controllable video editing with natural language instructions. In 2024 IEEE + +International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2024. 3 +[30] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dólar, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos, 2024. 4 +[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10684-10695, 2022. 1 +[32] Litu Rout, Yujia Chen, Nataniel Ruiz, Constantine Caramanis, Sanjay Shakkottai, and Wen-Sheng Chu. Semantic image inversion and editing using rectified stochastic differential equations. arXiv preprint arXiv:2410.10792, 2024. 2, 3 +[33] Michele A Saad and Alan C Bovik. Blind quality assessment of videos using a model of natural scene statistics and motion coherency. In 2012 Conference Record of the Forty Sixth Asilomar Conference on Signals, Systems and Computers (ASILOMAR), pages 332-336. IEEE, 2012. 6 +[34] Uriel Singer, Amit Zohar, Yuval Kirstain, Shelly Sheynin, Adam Polyak, Devi Parikh, and Yaniv Taigman. Video editing via factorized diffusion distillation. In Eur. Conf. Comput. Vis., pages 450-466. Springer, 2024. 2 +[35] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 2, 3 +[36] Spencer Sterling. Zeroscope. https://huggingface.co/cerspense/zeroscope_v2_576w, 2023.1 +[37] Shuyuan Tu, Qi Dai, Zhi-Qi Cheng, Han Hu, Xintong Han, Zuxuan Wu, and Yu-Gang Jiang. Motioneditor: Editing video motion via content-aware diffusion. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7882-7891, 2024. 3 +[38] Narek Tumanyan, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Splicing vit features for semantic appearance transfer. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10748-10757, 2022. 6 +[39] Bram Wallace, Akash Gokul, and Nikhil Naik. Edict: Exact diffusion inversion via coupled transformations. In IEEE Conf. Comput. Vis. Pattern Recog., pages 22532-22541, 2023. 3 +[40] Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, et al. Wan: Open and advanced large-scale video generative models. arXiv preprint arXiv:2503.20314, 2025. 1, 2, 3, 6 +[41] Jiangshan Wang, Junfu Pu, Zhongang Qi, Jiayi Guo, Yue Ma, Nisha Huang, Yuxin Chen, Xiu Li, and Ying Shan. Taming rectified flow for inversion and editing. arXiv preprint arXiv:2411.04746, 2024. 2, 3 +[42] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process., 13(4): 600-612, 2004. 6 + +[43] Yujie Wei, Shiwei Zhang, Zhiwu Qing, Hangjie Yuan, Zhiheng Liu, Yu Liu, Yingya Zhang, Jingren Zhou, and Hongming Shan. Dreamvideo: Composing your dream videos with customized subject and motion. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6537-6549, 2024. 3 +[44] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. GODIVA: Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806, 2021. 6 +[45] Chen Henry Wu and Fernando De la Torre. A latent space of stochastic diffusion models for zero-shot image editing and guidance. In Int. Conf. Comput. Vis., pages 7378-7387, 2023. 3 +[46] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Int. Conf. Comput. Vis., pages 7623-7633, 2023. 2, 3 +[47] Jay Zhangjie Wu, Xiuyu Li, Difei Gao, Zhen Dong, Jinbin Bai, Aishani Singh, Xiaoyu Xiang, Youzeng Li, Zuwei Huang, Yuanxi Sun, et al. Cvpr 2023 text guided video editing competition. arXiv preprint arXiv:2310.16003, 2023. 2 +[48] Chenxi Xie, Minghan Li, Shuai Li, Yuhui Wu, Qiaosi Yi, and Lei Zhang. Dnaedit: Direct noise alignment for text-guided rectified flow editing. arXiv preprint arXiv:2506.01430, 2025.3 +[49] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. In SIGGRAPH Asia 2023 Conference Papers, pages 1-11, 2023. 3 +[50] Xiangpeng Yang, Linchao Zhu, Hehe Fan, and Yi Yang. Videograin: Modulating space-time attention for multigrained video editing. arXiv preprint arXiv:2502.17258, 2025. 2, 3, 6, 7 +[51] Danah Yatim, Rafail Fridman, Omer Bar-Tal, Yoni Kasten, and Tali Dekel. Space-time diffusion features for zero-shot text-driven motion transfer. In IEEE Conf. Comput. Vis. Pattern Recog., pages 8466-8476, 2024. 3, 6, 7 +[52] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conf. Comput. Vis. Pattern Recog., pages 586-595, 2018. 6 +[53] Zhenghao Zhang, Zuozhuo Dai, Long Qin, and Weizhi Wang. Effived: Efficient video editing via text-instruction diffusion models. arXiv preprint arXiv:2403.11568, 2024. 3 +[54] Zhixing Zhang, Yanyu Li, Yushu Wu, Yanwu Xu, Anil Kag, Ivan Skorokhodov, Willi Menapace, Aliaksandr Siarohin, Junli Cao, Dimitris Metaxas, et al. Sf-v: Single forward video generation model. arXiv preprint arXiv:2406.04324, 2024. 3 +[55] Rui Zhao, Yuchao Gu, Jay Zhangjie Wu, David Junhao Zhang, Jia-Wei Liu, Weijia Wu, Jussi Keppo, and Mike Zheng Shou. Motiondirector: Motion customization of text-to-video diffusion models. In Eur. Conf. Comput. Vis., pages 273–290. Springer, 2025. 3 \ No newline at end of file diff --git a/2025/FiVE-Bench_ A Fine-grained Video Editing Benchmark for Evaluating Emerging Diffusion and Rectified Flow Models/images.zip b/2025/FiVE-Bench_ A Fine-grained Video Editing Benchmark for Evaluating Emerging Diffusion and Rectified Flow Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..aed1ab782e77be8cb55d472e865a3090bc8847a3 --- /dev/null +++ b/2025/FiVE-Bench_ A Fine-grained Video Editing Benchmark for Evaluating Emerging Diffusion and Rectified Flow Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9faf581da4b9e5d518ad44f22c3b68f16641faa3e8b7078f0f7ab555d107f6e3 +size 1000620 diff --git a/2025/FiVE-Bench_ A Fine-grained Video Editing Benchmark for Evaluating Emerging Diffusion and Rectified Flow Models/layout.json b/2025/FiVE-Bench_ A Fine-grained Video Editing Benchmark for Evaluating Emerging Diffusion and Rectified Flow Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..da98fd4e2852510adefa54f274b5e252a696f176 --- /dev/null +++ b/2025/FiVE-Bench_ A Fine-grained Video Editing Benchmark for Evaluating Emerging Diffusion and Rectified Flow Models/layout.json @@ -0,0 +1,8703 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 108, + 98, + 503, + 150 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 98, + 503, + 150 + ], + "spans": [ + { + "bbox": [ + 108, + 98, + 503, + 150 + ], + "type": "text", + "content": "FiVE-Bench: A Fine-grained Video Editing Benchmark for Evaluating Emerging Diffusion and Rectified Flow Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 119, + 170, + 489, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 170, + 489, + 199 + ], + "spans": [ + { + "bbox": [ + 119, + 170, + 489, + 199 + ], + "type": "text", + "content": "Minghan Li" + }, + { + "bbox": [ + 119, + 170, + 489, + 199 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 119, + 170, + 489, + 199 + ], + "type": "text", + "content": ", Chenxi Xie" + }, + { + "bbox": [ + 119, + 170, + 489, + 199 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 119, + 170, + 489, + 199 + ], + "type": "text", + "content": ", Yichen Wu" + }, + { + "bbox": [ + 119, + 170, + 489, + 199 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 119, + 170, + 489, + 199 + ], + "type": "text", + "content": ", Lei Zhang" + }, + { + "bbox": [ + 119, + 170, + 489, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 119, + 170, + 489, + 199 + ], + "type": "text", + "content": " and Mengyu Wang" + }, + { + "bbox": [ + 119, + 170, + 489, + 199 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 119, + 170, + 489, + 199 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 119, + 170, + 489, + 199 + ], + "type": "text", + "content": "Harvard University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 140, + 200, + 469, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 200, + 469, + 213 + ], + "spans": [ + { + "bbox": [ + 140, + 200, + 469, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 140, + 200, + 469, + 213 + ], + "type": "text", + "content": "Hong Kong Polytechnic University, " + }, + { + "bbox": [ + 140, + 200, + 469, + 213 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 140, + 200, + 469, + 213 + ], + "type": "text", + "content": "City University of Hong Kong" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 204, + 215, + 402, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 215, + 402, + 227 + ], + "spans": [ + { + "bbox": [ + 204, + 215, + 402, + 227 + ], + "type": "text", + "content": "{mili4, mengyu_wang}@meei.harvard.edu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 234, + 164, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 234, + 164, + 243 + ], + "spans": [ + { + "bbox": [ + 116, + 234, + 164, + 243 + ], + "type": "text", + "content": "FiVE-Dataset" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 58, + 245, + 228, + 449 + ], + "blocks": [ + { + "bbox": [ + 58, + 245, + 228, + 449 + ], + "lines": [ + { + "bbox": [ + 58, + 245, + 228, + 449 + ], + "spans": [ + { + "bbox": [ + 58, + 245, + 228, + 449 + ], + "type": "image", + "image_path": "824625187b3e53d65c739fdb8bb3993b55521516edb3e65a20d335514534c250.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 79, + 453, + 530, + 465 + ], + "lines": [ + { + "bbox": [ + 79, + 453, + 530, + 465 + ], + "spans": [ + { + "bbox": [ + 79, + 453, + 530, + 465 + ], + "type": "text", + "content": "Figure 1. Overview of the proposed FiVE-Bench, with editing results from Wan-Edit and corresponding FiVE-Acc scores." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 333, + 234, + 462, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 234, + 462, + 243 + ], + "spans": [ + { + "bbox": [ + 333, + 234, + 462, + 243 + ], + "type": "text", + "content": "Wan-Edit Results and FiVE-Evaluation" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 233, + 245, + 552, + 449 + ], + "blocks": [ + { + "bbox": [ + 233, + 245, + 552, + 449 + ], + "lines": [ + { + "bbox": [ + 233, + 245, + 552, + 449 + ], + "spans": [ + { + "bbox": [ + 233, + 245, + 552, + 449 + ], + "type": "image", + "image_path": "0ef43fd5061b1234cc2b78d174f485e754cf24259bd6c7c50c28b0caace1be8a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 151, + 475, + 200, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 475, + 200, + 487 + ], + "spans": [ + { + "bbox": [ + 151, + 475, + 200, + 487 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 501, + 297, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 501, + 297, + 693 + ], + "spans": [ + { + "bbox": [ + 53, + 501, + 297, + 693 + ], + "type": "text", + "content": "Numerous text-to-video (T2V) editing methods have emerged recently, but the lack of a standardized benchmark for fair evaluation has led to inconsistent claims and an inability to assess model sensitivity to hyperparameters. Fine-grained video editing is crucial for enabling precise, object-level modifications while maintaining context and temporal consistency. To address this, we introduce FiVE-Bench, a Fine-grained Video Editing Benchmark for evaluating emerging diffusion and rectified flow models. Our benchmark includes 74 real-world videos and 26 generated videos, featuring 6 fine-grained editing types, 420 object-level editing prompt pairs, and their corresponding masks. Additionally, we adapt the latest rectified flow (RF) T2V generation models—Pyramid-Flow [14] and Wan2.1 [40]—by introducing FlowEdit [18], resulting in training-free and inversion-free video editing mod" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 476, + 555, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 476, + 555, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 476, + 555, + 643 + ], + "type": "text", + "content": "els Pyramid-Edit and Wan-Edit. We compare five diffusion methods with our two RF methods on the proposed FiVE-Bench, evaluating them across 15 metrics. These metrics include background preservation, text-video similarity, temporal consistency, and generated video quality. To further enhance object-level evaluation, we introduce FiVE-Acc, a novel metric leveraging Vision-Language Models (VLMs) to assess the success of fine-grained video editing. Experimental results demonstrate that RF-based editing significantly outperforms diffusion-based methods, with WanEdit achieving the best overall performance and exhibiting the least sensitivity to hyperparameters. More video demo available on the website: https://sites.google.com/view/five-benchmark." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 669, + 395, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 669, + 395, + 681 + ], + "spans": [ + { + "bbox": [ + 314, + 669, + 395, + 681 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Recent advancements in text-to-image (T2I) and text-to-video (T2V) generation [1, 14, 24, 26, 31, 36, 40] have" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 214, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 214, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 214, + 713 + ], + "type": "text", + "content": "*Equal contribution, † Corresponding author." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16672" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "led to the emergence of numerous image and video editing methods [4, 9, 10, 46], enabling users to create and modify video content with unprecedented flexibility. These methods leverage powerful generative models, such as diffusion models [6, 11, 12, 35] and Rectified Flow (RF) [8, 23] models, to synthesize and edit videos based on textual prompts." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 146, + 295, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 146, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 146, + 295, + 326 + ], + "type": "text", + "content": "However, despite significant progress in these fields, the lack of a standardized benchmark for fair and comprehensive evaluation has still become a major bottleneck hindering further development. Currently, the video editing field has only one publicly available dataset, TGVE [47] /TGVE+ [34] (see Table 1), which includes 76 videos and 6 editing types, such as style transfer, background replacement, single-object and multi-object transformations, as well as object addition and removal. While TGVE+ aims to cover as many editing types as possible, its evaluation within each editing type is not comprehensive, particularly in supporting fine-grained video editing. Fine-grained video editing requires precise object-level modifications while preserving the overall context and temporal consistency of the video, posing higher demands on model performance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 327, + 295, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 295, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 295, + 458 + ], + "type": "text", + "content": "Due to the lack of a standardized benchmark and unified evaluation framework for fine-grained video editing, existing methods [9, 21, 28, 50] often rely on self-collected small-scale video datasets to validate their effectiveness. This approach not only makes it difficult to objectively compare the performance of different methods but also fails to comprehensively assess model robustness to hyperparameters or determine their suitability for real-world applications. Therefore, building a comprehensive and challenging benchmark for fine-grained video editing has become an urgent need to advance the field." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 460, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 652 + ], + "type": "text", + "content": "To address these challenges, we introduce FiVE-Bench (Fine-grained Video Editing Benchmark), a comprehensive benchmark designed to evaluate emerging diffusion and rectified flow models for fine-grained video editing. FiVE-Bench includes 74 real-world videos and 26 generated videos, covering a diverse range of scenes and editing scenarios. It features 6 fine-grained editing types, 420 object-level editing prompt pairs, and their corresponding masks, providing a rich and challenging benchmark for video editing. Additionally, we propose FiVE-Acc, a novel evaluation metric that leverages Vision-Language Models (VLMs) to measure the success rate of fine-grained object-level editing. Unlike traditional similarity-based scores, FiVE-Acc focuses on evaluating whether the intended edit was successfully achieved, with particular emphasis on semantic alignment of the edited object." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "content": "Recently, many image editing methods [5, 18, 32, 41] based on the latest rectified flow T2I generation model Flux [19] have achieved state-of-the-art editing results and fidelity, such as RF-Inversion [32], RF-Solver [41], and FlowEdit [18]. However, to the best of our knowledge," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "there are no video editing methods based on RF T2V generation models, primarily due to the relatively slow development of T2V generation models. Fortunately, the recent release of two state-of-the-art RF-based T2V generation models—Pyramid-Flow [14] and Wan2.1 [40]—has opened new possibilities for RF video editing." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 145, + 553, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 145, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 145, + 553, + 348 + ], + "type": "text", + "content": "Building on the latest advancements, we adapt Pyramid-Flow and Wan2.1 by introducing FlowEdit [18]. This adaptation facilitates training-free and inversion-free video editing models, Pyramid-Edit and Wan-Edit, which leverage the strong temporal consistency inherent in video generation models without requiring additional attention mechanisms. Specifically, Pyramid-Flow employs a multi-resolution and temporally autoregressive model architecture across time steps. To adapt to this unique design, we modified FlowEdit to support multi-resolution processing, enabling seamless integration with the Pyramid-Flow framework, leading to Pyramid-Edit. On the other hand, Wan2.1 utilizes the MMDiT architecture, which unfolds temporal frames into sequences and processes them via self-attention. Since its design is fundamentally similar to image generation models, Wan2.1 can directly adopt FlowEdit without any additional modifications, giving rise to Wan-Edit." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 349, + 553, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 553, + 503 + ], + "type": "text", + "content": "To show the effectiveness of the proposed RF-based editing methods, we compare them with five diffusion-based editing methods on the built FiVE-Bench, evaluating them across fifteen metrics, including background preservation, edited text-video similarity, temporal consistency, and generated video quality. Our experiments demonstrate that RF-based editing methods significantly outperform diffusion-based methods across multiple metrics. Notably, Wan-Edit achieves the best overall performance, exhibiting superior editing quality and the least sensitivity to hyperparameters. These findings highlight the potential of RF-based models for fine-grained video editing and provide valuable insights for future research in this area." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 328, + 506, + 539, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 506, + 539, + 517 + ], + "spans": [ + { + "bbox": [ + 328, + 506, + 539, + 517 + ], + "type": "text", + "content": "In summary, this work makes four key contributions:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 517, + 553, + 637 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 314, + 517, + 553, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 517, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 314, + 517, + 553, + 553 + ], + "type": "text", + "content": "- We introduce FiVE-Bench, a comprehensive benchmark for fine-grained video editing, featuring diverse videos, editing types, and evaluation metrics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 554, + 553, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 554, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 314, + 554, + 553, + 578 + ], + "type": "text", + "content": "- We propose FiVE-Acc, a novel metric leveraging VLMs to assess the success of fine-grained editing." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 578, + 553, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 578, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 314, + 578, + 553, + 601 + ], + "type": "text", + "content": "- We propose the first two RF-based T2V editing models: Pyramid-Edit and Wan-Edit." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 601, + 553, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 601, + 553, + 637 + ], + "spans": [ + { + "bbox": [ + 314, + 601, + 553, + 637 + ], + "type": "text", + "content": "- We evaluate 7 video editing methods across 15 metrics, showing that RF-based editing consistently outperforms diffusion-based approaches." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 651, + 400, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 651, + 400, + 662 + ], + "spans": [ + { + "bbox": [ + 314, + 651, + 400, + 662 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 671, + 464, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 671, + 464, + 682 + ], + "spans": [ + { + "bbox": [ + 314, + 671, + 464, + 682 + ], + "type": "text", + "content": "2.1. Diffusion and RF Inversion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Diffusion inversion. Diffusion models (DMs) drive advances in inversion techniques for better control over gen" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16673" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 85, + 553, + 174 + ], + "blocks": [ + { + "bbox": [ + 179, + 71, + 430, + 82 + ], + "lines": [ + { + "bbox": [ + 179, + 71, + 430, + 82 + ], + "spans": [ + { + "bbox": [ + 179, + 71, + 430, + 82 + ], + "type": "text", + "content": "Table 1. Statistics of existing video editing datasets and benchmarks." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 85, + 553, + 174 + ], + "lines": [ + { + "bbox": [ + 61, + 85, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 61, + 85, + 553, + 174 + ], + "type": "table", + "html": "
DatasetUsageNum. ImagesNum. VideosFrames Per VideoNum. PromptsGen. VideosSrc. obj. WordsEdit obj. WordsObj. Instruct.Src. obj. MasksType of Edited Prompts
VIVID-10MTrain672K73.7K3010MXXX✓✓object, add, remove
Señorita-2MTrain0388.9K33~642MXXX✓Xstyle, object, multiple, color, add, remove, motion
TGVEEval07632304XXXXXstyle, object, bg, multiple
TGVE+Eval076321417XXXXXstyle, object, bg, multiple, color, add, remove
VIVID-10MEval06433~64852XXX✓✓object, add, remove
FiVE (Ours)Eval010035 ~ 126420✓✓✓✓✓object (w&w/o non-rigid), color, material, add, remove
", + "image_path": "d3f7bf5afa3ed7715ed2242eda301ebf756e753978f74b45840d056e1c475f49.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 194, + 297, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 194, + 297, + 386 + ], + "spans": [ + { + "bbox": [ + 54, + 194, + 297, + 386 + ], + "type": "text", + "content": "erated samples. DDIM [35] accelerates sampling but struggles with fine-grained reconstructions due to nonlinearities and score estimation errors. To improve accuracy, fine-tuning and optimization-based methods [7, 15, 25, 39, 45] have been proposed. While these advancements significantly enhance the fidelity and accuracy of DDIM inversion, the associated computational costs and resource demands remain substantial limiting factors. Although diffusion-based methods have achieved outstanding performance, recent studies [8, 14, 19, 32] suggest that Rectified Flow models (RFs) hold significant potential to surpass DMs in certain applications. Notably, the inversion and editing capabilities of RF models have been explored in image editing [5, 18, 32, 41, 48], demonstrating their efficiency and effectiveness. Building on these strengths, this work extends their potential in video editing." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 396, + 144, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 144, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 144, + 407 + ], + "type": "text", + "content": "2.2. Video Editing" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 414, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 414, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 56, + 414, + 295, + 593 + ], + "type": "text", + "content": "Video editing techniques can generally be categorized into three types based on their training: training-free methods, one/few-shot finetuned methods [37, 43, 46, 55], and massive data finetuned methods. Training-free models [17, 21, 28, 50], enable task execution without retraining. TokenFlow [9] utilizes pre-trained models for efficient video synthesis, while DMT [51] transfers motion via diffusion. Other models like Pixel2Video [3], Render-A-Video [49], and Text2Video-zero [16] enhance accessibility for real-time video editing. One/few-shot fine-tuned models adapt pre-trained models for video editing with minimal fine-tuning. Methods like Tune-A-Video [46] and MotionEditor [37] achieve strong results using limited annotated samples, balancing customization and efficiency for task-specific editing without extensive retraining." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": "Massive data fine-tuned models undergo extensive training on large datasets for high-quality, versatile video editing [29]. InstructVid2Vid [29] enables complex edits via natural language instructions, while EffiVED [53] refines broad datasets into high-quality subsets for efficient editing. SF-V [54] introduces adversarial training for video synthesis. Despite their precision and adaptability, these methods are computationally expensive, limiting practicality in resource-constrained settings. These methods focus on learning video motion, which may compromise back" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 194, + 554, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 554, + 255 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 554, + 255 + ], + "type": "text", + "content": "ground preservation. Recently, VideoGrain [50] has emphasized multi-granularity video editing rather than fine-grained edits, aiming to retain background details while modifying target objects. This work aims to address this issue and advance the field." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 266, + 533, + 279 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 533, + 279 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 533, + 279 + ], + "type": "text", + "content": "3. FiVE Benchmark and FiVE-Acc Metric" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 287, + 419, + 299 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 287, + 419, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 287, + 419, + 299 + ], + "type": "text", + "content": "3.1. FiVE Benchmark" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 305, + 554, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 305, + 554, + 341 + ], + "spans": [ + { + "bbox": [ + 313, + 305, + 554, + 341 + ], + "type": "text", + "content": "We present the FiVE benchmark, which includes the FiVE video dataset and six fine-grained editing tasks. Additional details can be found in the Appendix." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 342, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 533 + ], + "type": "text", + "content": "Videos in FiVE-Bench. We meticulously curate 74 real-world videos suitable for fine-grained editing from the DAVIS [27] dataset. Consecutive frames are extracted at every 8-frame interval, and the GPT-4o [13] is employed to generate formatted captions for these videos. These captions encompass details pertaining to object category, action, background, and camera movement. Additionally, during the annotation process, we document instances of object deformation (such as limb movements in humans or animals), which aids in differentiating the complexity of editing tasks during evaluation. To further enhance the diversity of our benchmark, we generate 26 highly realistic synthetic videos using the T2V model [40]. These synthetic videos not only expand the range of video categories but also enable a comparative analysis of editing performance across real and synthetic videos." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 534, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 534, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 534, + 556, + 714 + ], + "type": "text", + "content": "Six Fine-grained Video Editing Tasks. FiVE-Bench comprises six fine-grained video editing tasks, totaling 420 high-quality editing prompt pairs. It primarily includes four object-targeted editing tasks of increasing complexity: object substitution without non-rigid deformation, object substitution with non-rigid deformation, color alteration, and material modification. For each task, we use GPT-4o [13] to generate four source-target prompt pairs per video in the benchmark, resulting in 400 high-quality editing pairs. To further enhance the diversity of editing types within the benchmark, we select 10 videos each for the add and remove tasks and design corresponding source-target prompt pairs. Additionally, we explicitly provide both object-related source words and edited words in the source and edited text prompts, respectively, denoted as" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16674" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 71, + 553, + 186 + ], + "blocks": [ + { + "bbox": [ + 60, + 71, + 553, + 186 + ], + "lines": [ + { + "bbox": [ + 60, + 71, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 60, + 71, + 553, + 186 + ], + "type": "image", + "image_path": "64eeee1923f5e9435008803cc18b46ffb3f030ad222c11dd33d8538f81b93101.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 168, + 192, + 440, + 204 + ], + "lines": [ + { + "bbox": [ + 168, + 192, + 440, + 204 + ], + "spans": [ + { + "bbox": [ + 168, + 192, + 440, + 204 + ], + "type": "text", + "content": "Figure 2. The VLM-based FiVE accuracy (FiVE-Acc) evaluation pipeline." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 224, + 295, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 224, + 295, + 309 + ], + "spans": [ + { + "bbox": [ + 54, + 224, + 295, + 309 + ], + "type": "text", + "content": "\"src. obj. words\" and \"edit obj. words\" in Table 1. To accommodate different editing methods, we use GPT-4o [13] to generate instruction prompts (\"Obj. Instr.\" in Table 1) for each source-target pair, ensuring compatibility with models like InstructPix2Pix [2]. For each video, we employ SAM2 [20, 30] to generate masks of the edited regions, enabling the evaluation of background preservation metrics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 319, + 160, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 160, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 160, + 331 + ], + "type": "text", + "content": "3.2. FiVE-Acc Metric" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 338, + 296, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 296, + 398 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 296, + 398 + ], + "type": "text", + "content": "To evaluate the accuracy of fine-grained video editing, we propose FiVE accuracy (FiVE-Acc) evaluation, a framework assessing how precisely video editing models modify the target object. Fig. 2 presents the complete pipeline, comprising video editing and accuracy evaluation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 399, + 296, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 296, + 459 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 296, + 459 + ], + "type": "text", + "content": "Video editing outputs. Given a source video, a video editing model is prompted with a source prompt, describing the initial scene, and a target prompt, specifying the desired modification. The model generates a target video, where the intended transformation is applied, as shown in Fig. 2a." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 460, + 296, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 296, + 508 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 296, + 508 + ], + "type": "text", + "content": "FiVE-Acc evaluation. To quantitatively evaluate the fidelity of video editing, we employ a Vision-Language Model (VLM) (e.g., Qwen-2.5-VL) to analyze the edited video. The evaluation includes two types of questions:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 509, + 295, + 653 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 55, + 509, + 295, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 592 + ], + "type": "text", + "content": "- Yes/No Questions: The model is asked whether the source and target objects are present in the target video. Editing succeeds only if the model answers \"No\" for the source object and \"Yes\" for the target, leading to the calculation of " + }, + { + "bbox": [ + 55, + 509, + 295, + 592 + ], + "type": "inline_equation", + "content": "YN" + }, + { + "bbox": [ + 55, + 509, + 295, + 592 + ], + "type": "text", + "content": "-Acc. The source object question ensures the model both removes the source and adds the target, refining accuracy evaluation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 593, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 593, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 593, + 295, + 653 + ], + "type": "text", + "content": "- Multi-choice Question: It evaluates whether the VLM recognizes the target (e.g., flamingo) or source object (e.g., black swan) in the edited video. The model selects between the two objects to compute MC-Acc, measuring its recognition accuracy." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "Based on the two obtained accuracies, we further calculate the union and intersection accuracies to derive " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\cup" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "-Acc and " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\cap" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "-Acc. These offer a more detailed evaluation: the former reflects the model's overall editing success accuracy, while the latter highlights the model's high-quality editing" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 224, + 555, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 224, + 555, + 298 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 555, + 298 + ], + "type": "text", + "content": "success accuracy. " + }, + { + "bbox": [ + 313, + 224, + 555, + 298 + ], + "type": "inline_equation", + "content": "FiVE - Acc" + }, + { + "bbox": [ + 313, + 224, + 555, + 298 + ], + "type": "text", + "content": " is the average of four components: " + }, + { + "bbox": [ + 313, + 224, + 555, + 298 + ], + "type": "inline_equation", + "content": "FiVE - Acc = 0.25 \\times (YN - Acc + MC - Acc + \\cup - Acc + \\cap - Acc)" + }, + { + "bbox": [ + 313, + 224, + 555, + 298 + ], + "type": "text", + "content": ". It uses the VLM's recognition ability to evaluate the alignment between the edited object and its real data distribution. To the best of our knowledge, we are the first to introduce accuracy evaluation in fine-grained T2V editing." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 308, + 528, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 528, + 322 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 528, + 322 + ], + "type": "text", + "content": "4. Methods: Pyramid-Edit and Wan-Edit" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 327, + 394, + 341 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 394, + 341 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 394, + 341 + ], + "type": "text", + "content": "4.1. Preliminary" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 345, + 554, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 345, + 554, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 554, + 418 + ], + "type": "text", + "content": "Rectified Flow (RF) Models [22, 23]. Let " + }, + { + "bbox": [ + 313, + 345, + 554, + 418 + ], + "type": "inline_equation", + "content": "q_{0}" + }, + { + "bbox": [ + 313, + 345, + 554, + 418 + ], + "type": "text", + "content": " denote the source distribution (i.e., standard Gaussian " + }, + { + "bbox": [ + 313, + 345, + 554, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,I)" + }, + { + "bbox": [ + 313, + 345, + 554, + 418 + ], + "type": "text", + "content": "), " + }, + { + "bbox": [ + 313, + 345, + 554, + 418 + ], + "type": "inline_equation", + "content": "p_{0}" + }, + { + "bbox": [ + 313, + 345, + 554, + 418 + ], + "type": "text", + "content": " refer to the target distribution (i.e., the distribution over images), and " + }, + { + "bbox": [ + 313, + 345, + 554, + 418 + ], + "type": "inline_equation", + "content": "v_{t}(\\cdot)" + }, + { + "bbox": [ + 313, + 345, + 554, + 418 + ], + "type": "text", + "content": " is the time-varying vector field. RF models [22, 23] aim to transform " + }, + { + "bbox": [ + 313, + 345, + 554, + 418 + ], + "type": "inline_equation", + "content": "q_{0}" + }, + { + "bbox": [ + 313, + 345, + 554, + 418 + ], + "type": "text", + "content": " progressively into " + }, + { + "bbox": [ + 313, + 345, + 554, + 418 + ], + "type": "inline_equation", + "content": "p_{0}" + }, + { + "bbox": [ + 313, + 345, + 554, + 418 + ], + "type": "text", + "content": " by solving the following ordinary differential equation (ODE):" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 378, + 422, + 553, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 422, + 553, + 434 + ], + "spans": [ + { + "bbox": [ + 378, + 422, + 553, + 434 + ], + "type": "interline_equation", + "content": "d X _ {t} = v _ {t} \\left(X _ {t}\\right) d t, \\quad t \\in [ 0, 1 ]. \\tag {1}", + "image_path": "a4d0e124d6f80f8a6c039261c1273801b459a1731dbd2a739ec367646e34f2f6.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 438, + 555, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 555, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 555, + 593 + ], + "type": "text", + "content": "Here, the vector field " + }, + { + "bbox": [ + 313, + 438, + 555, + 593 + ], + "type": "inline_equation", + "content": "v_{t}(X_{t}) = f(X_{t},c,1 - t;\\varphi)" + }, + { + "bbox": [ + 313, + 438, + 555, + 593 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 438, + 555, + 593 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 313, + 438, + 555, + 593 + ], + "type": "text", + "content": " is a neural network parameterized by " + }, + { + "bbox": [ + 313, + 438, + 555, + 593 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 313, + 438, + 555, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 438, + 555, + 593 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 438, + 555, + 593 + ], + "type": "text", + "content": " is the guided text. Given the ODE in Eq. (1), we start with " + }, + { + "bbox": [ + 313, + 438, + 555, + 593 + ], + "type": "inline_equation", + "content": "X_0\\sim q_0" + }, + { + "bbox": [ + 313, + 438, + 555, + 593 + ], + "type": "text", + "content": " from the source distribution and integrate over " + }, + { + "bbox": [ + 313, + 438, + 555, + 593 + ], + "type": "inline_equation", + "content": "t:0\\to 1" + }, + { + "bbox": [ + 313, + 438, + 555, + 593 + ], + "type": "text", + "content": " to reach the endpoint " + }, + { + "bbox": [ + 313, + 438, + 555, + 593 + ], + "type": "inline_equation", + "content": "X_{1}\\sim p_{0}" + }, + { + "bbox": [ + 313, + 438, + 555, + 593 + ], + "type": "text", + "content": ", which is sampled from the target distribution. The forward process of RF assumes the linear path between the two states: " + }, + { + "bbox": [ + 313, + 438, + 555, + 593 + ], + "type": "inline_equation", + "content": "X_{t} = (1 - t)X_{0} + tX_{1}" + }, + { + "bbox": [ + 313, + 438, + 555, + 593 + ], + "type": "text", + "content": ", then we can derive the vector field of " + }, + { + "bbox": [ + 313, + 438, + 555, + 593 + ], + "type": "inline_equation", + "content": "v_{t}(\\cdot):u_{t}(X_{t}) = X_{1} - X_{0}" + }, + { + "bbox": [ + 313, + 438, + 555, + 593 + ], + "type": "text", + "content": ". Then, " + }, + { + "bbox": [ + 313, + 438, + 555, + 593 + ], + "type": "inline_equation", + "content": "v_{t}(X_{t})" + }, + { + "bbox": [ + 313, + 438, + 555, + 593 + ], + "type": "text", + "content": " is used to approximate the network " + }, + { + "bbox": [ + 313, + 438, + 555, + 593 + ], + "type": "inline_equation", + "content": "f(X_{t},c,t;\\varphi)" + }, + { + "bbox": [ + 313, + 438, + 555, + 593 + ], + "type": "text", + "content": " by minimizing the loss function " + }, + { + "bbox": [ + 313, + 438, + 555, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\varphi} = \\mathbb{E}_{t\\sim \\mathcal{U}[0,1],Y_t}[\\| v_t(X_t) - f(X_t,c,t;\\varphi)\\| _2^2 ]" + }, + { + "bbox": [ + 313, + 438, + 555, + 593 + ], + "type": "text", + "content": ". RF ensures that their sampling paths are relatively straight, enabling the use of a small number of discretization steps." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "RF-based Image Editing: FlowEdit [18]. Image editing involves taking a source image " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "X_{1}^{src}" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": " and a source text prompt " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "c^{src}" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": ", and allowing the user to provide a target text prompt " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "c^{tgt}" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": " for targeted fine-grained editing to ultimately yield the edited image " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "X_{1}^{tgt}" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": ". For example, the source text prompt might be \"A black swan drifting,\" while the target text prompt could be \"A flamingo drifting.\" To simplify notation, we denote the vector fields corresponding to the source text prompt and the target text prompt as " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "v_{t}(X_{t}^{src})" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "v_{t}(X_{t}^{tgt})" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": ", respectively. Their sampling process in RF" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16675" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 71, + 289, + 304 + ], + "blocks": [ + { + "bbox": [ + 60, + 71, + 289, + 304 + ], + "lines": [ + { + "bbox": [ + 60, + 71, + 289, + 304 + ], + "spans": [ + { + "bbox": [ + 60, + 71, + 289, + 304 + ], + "type": "image", + "image_path": "36f1d30bd80c21ddad15542befab8a480f8acaf33ffe63c46e1f5ac8296a5f5f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 148, + 307, + 204, + 317 + ], + "lines": [ + { + "bbox": [ + 148, + 307, + 204, + 317 + ], + "spans": [ + { + "bbox": [ + 148, + 307, + 204, + 317 + ], + "type": "text", + "content": "(a) Pyramid-Edit" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 539, + 304 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 539, + 304 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 539, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 539, + 304 + ], + "type": "image", + "image_path": "4bda0b0d40e70e4a2de0cc6e2e50891490ea6bbd0710f82eb3d74cd493d3426e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 307, + 453, + 316 + ], + "lines": [ + { + "bbox": [ + 410, + 307, + 453, + 316 + ], + "spans": [ + { + "bbox": [ + 410, + 307, + 453, + 316 + ], + "type": "text", + "content": "(b) Wan-Edit" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 321, + 555, + 344 + ], + "lines": [ + { + "bbox": [ + 55, + 321, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 321, + 555, + 344 + ], + "type": "text", + "content": "Figure 3. Two RF-based inversion-free video editing models: Pyramid-Edit and Wan-Edit. Pyramid-Edit is a multi-resolution temporal autoregressive architecture, while Wan-Edit (DiT architecture) treats temporal frames as a sequence and processing them simultaneously." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 364, + 259, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 364, + 259, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 364, + 259, + 376 + ], + "type": "text", + "content": "follows: " + }, + { + "bbox": [ + 55, + 364, + 259, + 376 + ], + "type": "inline_equation", + "content": "dX_{t}^{src} = v_{t}(X_{t}^{src})dt" + }, + { + "bbox": [ + 55, + 364, + 259, + 376 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 364, + 259, + 376 + ], + "type": "inline_equation", + "content": "dX_{t}^{tgt} = v_{t}(X_{t}^{tgt})dt" + }, + { + "bbox": [ + 55, + 364, + 259, + 376 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 376, + 296, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 296, + 437 + ], + "type": "text", + "content": "FlowEdit operates inversion-free, mapping source modes to the nearest target modes. It assumes target images follow straight trajectories, allowing latent interpolation. Thus, the nearest target mode is derived from the difference between source and target vector fields as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 439, + 294, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 439, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 116, + 439, + 294, + 453 + ], + "type": "interline_equation", + "content": "X _ {t} ^ {e d i t} = X _ {1} ^ {s r c} + X _ {t} ^ {t g t} - X _ {t} ^ {s r c}. \\tag {2}", + "image_path": "3191abdc419c9d1c493a21c8003af5beea0dca64f3d0d4eedd24a50c346b38c6.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 454, + 296, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 454, + 296, + 527 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 296, + 527 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 55, + 454, + 296, + 527 + ], + "type": "inline_equation", + "content": "X_{t}^{edit} = X_{t}^{tgt}" + }, + { + "bbox": [ + 55, + 454, + 296, + 527 + ], + "type": "text", + "content": ", if " + }, + { + "bbox": [ + 55, + 454, + 296, + 527 + ], + "type": "inline_equation", + "content": "t \\to 1" + }, + { + "bbox": [ + 55, + 454, + 296, + 527 + ], + "type": "text", + "content": ". A more straightforward explanation is: " + }, + { + "bbox": [ + 55, + 454, + 296, + 527 + ], + "type": "inline_equation", + "content": "X_{t}^{edit}" + }, + { + "bbox": [ + 55, + 454, + 296, + 527 + ], + "type": "text", + "content": " evolves within the target distribution, continuously transitioning through the clean latent space across all timesteps, (i.e. " + }, + { + "bbox": [ + 55, + 454, + 296, + 527 + ], + "type": "inline_equation", + "content": "X_{t}^{edit}" + }, + { + "bbox": [ + 55, + 454, + 296, + 527 + ], + "type": "text", + "content": " gradually transforms from a black swan to a flamingo as the timestep progresses). Therefore, we can obtain " + }, + { + "bbox": [ + 55, + 454, + 296, + 527 + ], + "type": "inline_equation", + "content": "X_{1}^{tgt}" + }, + { + "bbox": [ + 55, + 454, + 296, + 527 + ], + "type": "text", + "content": " simply by solving " + }, + { + "bbox": [ + 55, + 454, + 296, + 527 + ], + "type": "inline_equation", + "content": "X_{t}^{edit}" + }, + { + "bbox": [ + 55, + 454, + 296, + 527 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 80, + 529, + 295, + 560 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 529, + 295, + 560 + ], + "spans": [ + { + "bbox": [ + 80, + 529, + 295, + 560 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} X _ {t} ^ {e d i t} = X _ {t - 1} ^ {e d i t} + d X _ {t} ^ {t g t} - d X _ {t} ^ {s r c} \\\\ \\approx X _ {t - 1} ^ {e d i t} + \\mathbb {E} \\left[ v _ {t} \\left(\\hat {X} _ {t} ^ {t g t}\\right) - v _ {t} \\left(X _ {t} ^ {s r c}\\right) \\mid X _ {1} ^ {s r c} \\right] d t \\tag {3} \\\\ \\end{array}", + "image_path": "ec6b86efade775feb9d950ebf5f727eab42c36e5349ff0e43fff265b50b30f0a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 563, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 296, + 601 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 563, + 296, + 601 + ], + "type": "inline_equation", + "content": "X_{t}^{tgt}" + }, + { + "bbox": [ + 55, + 563, + 296, + 601 + ], + "type": "text", + "content": " is approximated as " + }, + { + "bbox": [ + 55, + 563, + 296, + 601 + ], + "type": "inline_equation", + "content": "\\hat{X}_{t}^{tgt} = X_{t - 1}^{edit} + X_{t}^{src} - X_{1}^{src}" + }, + { + "bbox": [ + 55, + 563, + 296, + 601 + ], + "type": "text", + "content": ". To obtain the edited image, " + }, + { + "bbox": [ + 55, + 563, + 296, + 601 + ], + "type": "inline_equation", + "content": "X_{t}^{edit}" + }, + { + "bbox": [ + 55, + 563, + 296, + 601 + ], + "type": "text", + "content": " is iteratively updated from " + }, + { + "bbox": [ + 55, + 563, + 296, + 601 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 563, + 296, + 601 + ], + "type": "text", + "content": " to 1, where " + }, + { + "bbox": [ + 55, + 563, + 296, + 601 + ], + "type": "inline_equation", + "content": "\\epsilon \\in [0,1)" + }, + { + "bbox": [ + 55, + 563, + 296, + 601 + ], + "type": "text", + "content": " represents skipped timesteps." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 605, + 192, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 192, + 618 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 192, + 618 + ], + "type": "text", + "content": "4.2. RF-based Video Editing" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 622, + 296, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 622, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 622, + 296, + 647 + ], + "type": "text", + "content": "This section introduces Pyramid-Edit, followed by a discussion of Wan-Edit, highlighting key features and differences." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 651, + 144, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 144, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 144, + 662 + ], + "type": "text", + "content": "4.2.1. Pyramid-Edit" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Pyramid-Flow [14] employs a multi-resolution scheme across timesteps and a temporally autoregressive architecture to effectively handle the spatial-temporal complexity in video generative modeling. It decomposes the flow" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 365, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 437 + ], + "type": "text", + "content": "into segments over " + }, + { + "bbox": [ + 313, + 365, + 555, + 437 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 365, + 555, + 437 + ], + "type": "text", + "content": " timestep windows within the interval [0,1]. Each window interpolates between successive solutions, reducing redundant computations in the earlier steps. For example, in the sampling process, within the " + }, + { + "bbox": [ + 313, + 365, + 555, + 437 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 365, + 555, + 437 + ], + "type": "text", + "content": "-th window " + }, + { + "bbox": [ + 313, + 365, + 555, + 437 + ], + "type": "inline_equation", + "content": "[s_k, e_k]" + }, + { + "bbox": [ + 313, + 365, + 555, + 437 + ], + "type": "text", + "content": ", it defines the rescaled timestep " + }, + { + "bbox": [ + 313, + 365, + 555, + 437 + ], + "type": "inline_equation", + "content": "t_k = (t - s_k) / (e_k - s_k)" + }, + { + "bbox": [ + 313, + 365, + 555, + 437 + ], + "type": "text", + "content": " and the corresponding flow is" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 344, + 447, + 555, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 447, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 344, + 447, + 555, + 460 + ], + "type": "interline_equation", + "content": "X _ {t _ {k}} = \\left(1 - t _ {k}\\right) \\bar {\\mathbf {x}} _ {s _ {k}} + t _ {k} \\bar {\\mathbf {x}} _ {e _ {k}}, \\tag {4}", + "image_path": "55a6fc39df6e9e80ea3c673b6fcf1dd45acc743a32dfac3de6095ad410e4336b.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 345, + 461, + 554, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 461, + 554, + 475 + ], + "spans": [ + { + "bbox": [ + 345, + 461, + 554, + 475 + ], + "type": "interline_equation", + "content": "\\bar {\\mathbf {x}} _ {s _ {k}} = \\left(1 - s _ {k}\\right) X _ {0} + s _ {k} U p (D o w n \\left(X _ {1}, 2 ^ {k + 1}\\right)), \\tag {5}", + "image_path": "5320c9f099c732991fd164d6a278dc9c0b2ffe022895a01b207191509d8623e1.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 345, + 477, + 553, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 477, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 345, + 477, + 553, + 491 + ], + "type": "interline_equation", + "content": "\\bar {\\mathbf {x}} _ {e _ {k}} = \\left(1 - e _ {k}\\right) X _ {0} + e _ {k} D o w n \\left(X _ {1}, 2 ^ {k}\\right), \\tag {6}", + "image_path": "e11569ea7889b08fa1689b475792a60d5306e9b5c1b6027751355b0eca262502.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 500, + 555, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 500, + 555, + 572 + ], + "spans": [ + { + "bbox": [ + 313, + 500, + 555, + 572 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 500, + 555, + 572 + ], + "type": "inline_equation", + "content": "\\text{Down}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 313, + 500, + 555, + 572 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 500, + 555, + 572 + ], + "type": "inline_equation", + "content": "\\text{Up}(\\cdot)" + }, + { + "bbox": [ + 313, + 500, + 555, + 572 + ], + "type": "text", + "content": " represent the down-sampling and up-sampling, respectively. Meanwhile, the vector field in the " + }, + { + "bbox": [ + 313, + 500, + 555, + 572 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 500, + 555, + 572 + ], + "type": "text", + "content": "-th window is defined as " + }, + { + "bbox": [ + 313, + 500, + 555, + 572 + ], + "type": "inline_equation", + "content": "v_{t_k}(X_{t_k}) = \\bar{\\mathbf{x}}_{e_k} - \\bar{\\mathbf{x}}_{s_k}" + }, + { + "bbox": [ + 313, + 500, + 555, + 572 + ], + "type": "text", + "content": ". To maintain continuity in the probability path within the pyramid structure, corrective noise is added at transition points between stages, as follows:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 326, + 581, + 555, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 581, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 326, + 581, + 555, + 601 + ], + "type": "interline_equation", + "content": "X _ {s _ {k - 1}} = \\frac {s _ {k - 1}}{e _ {k}} U p \\left(X _ {e _ {k}}\\right) + \\alpha \\mathbf {n} ^ {\\prime}, s. t. \\mathbf {n} _ {k - 1} \\sim \\mathcal {N} \\left(0, \\boldsymbol {\\Sigma} ^ {\\prime}\\right), \\tag {7}", + "image_path": "7244a8e2956da8353755908efbde898ed2d87e5235666b78db5e343a5859e5a8.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 610, + 554, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 610, + 554, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 610, + 554, + 658 + ], + "type": "text", + "content": "where the rescaling coefficient " + }, + { + "bbox": [ + 313, + 610, + 554, + 658 + ], + "type": "inline_equation", + "content": "s_{k-1} / e_k" + }, + { + "bbox": [ + 313, + 610, + 554, + 658 + ], + "type": "text", + "content": " aligns the means of these distributions, the corrective noise " + }, + { + "bbox": [ + 313, + 610, + 554, + 658 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_{k-1}'" + }, + { + "bbox": [ + 313, + 610, + 554, + 658 + ], + "type": "text", + "content": " weighted by " + }, + { + "bbox": [ + 313, + 610, + 554, + 658 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 610, + 554, + 658 + ], + "type": "text", + "content": " aligns their covariance matrices, and the covariance matrix " + }, + { + "bbox": [ + 313, + 610, + 554, + 658 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Sigma}'" + }, + { + "bbox": [ + 313, + 610, + 554, + 658 + ], + "type": "text", + "content": " is related to the upsampling function." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 658, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 555, + 714 + ], + "type": "text", + "content": "Pyramid-Edit. The autoregressive architecture allows Pyramid-Edit to process frames sequentially, where previous frames serve as conditional information. However, the multi-resolution scheme across timesteps hinders direct use of FlowEdit. To address this, Pyramid-Edit integrates" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16676" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 253, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 253, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 253, + 83 + ], + "type": "text", + "content": "FlowEdit (Eq. (3)) into each window as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 90, + 295, + 105 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 90, + 295, + 105 + ], + "spans": [ + { + "bbox": [ + 69, + 90, + 295, + 105 + ], + "type": "interline_equation", + "content": "X _ {t _ {k}} ^ {e d i t} = X _ {t _ {k} - 1} ^ {e d i t} + \\mathbb {E} [ v _ {t _ {k}} (\\hat {X} _ {t _ {k}} ^ {t g t}) - v _ {t _ {k}} (X _ {t _ {k}} ^ {s r c}) | \\bar {\\mathbf {x}} _ {e _ {k}} ^ {s r c} ] d t, \\qquad (8)", + "image_path": "d87ff81de50be24349ba22e562031f4977db02bf0c12241b39c48a3d73ebbcb5.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 110, + 295, + 134 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 110, + 295, + 134 + ], + "spans": [ + { + "bbox": [ + 54, + 110, + 295, + 134 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 54, + 110, + 295, + 134 + ], + "type": "inline_equation", + "content": "\\hat{X}_{t_k}^{tgt} = X_{t_k - 1}^{edit} + X_{t_k}^{src} - \\bar{\\mathbf{x}}_{e_k}^{src}" + }, + { + "bbox": [ + 54, + 110, + 295, + 134 + ], + "type": "text", + "content": ", and the target distribution is the end point of each window " + }, + { + "bbox": [ + 54, + 110, + 295, + 134 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{x}}_{e_k}^{src}" + }, + { + "bbox": [ + 54, + 110, + 295, + 134 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 134, + 295, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 134, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 134, + 295, + 312 + ], + "type": "text", + "content": "The workflow proceeds as illustrated in Fig. 3a and is structured as follows: starting from noise and the lowest-resolution latent, " + }, + { + "bbox": [ + 55, + 134, + 295, + 312 + ], + "type": "inline_equation", + "content": "X_{t_0}^{edit}" + }, + { + "bbox": [ + 55, + 134, + 295, + 312 + ], + "type": "text", + "content": " is initialized as " + }, + { + "bbox": [ + 55, + 134, + 295, + 312 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{x}}_{e_0}^{src}" + }, + { + "bbox": [ + 55, + 134, + 295, + 312 + ], + "type": "text", + "content": ", Eq. (8) is used to reconstruct and edit low-frequency information. At the end of the first stage, the reconstructed latent and the edited latent are upsampled to a higher resolution and passed through corrective noise as described in Eq. (7) to obtain the starting point for the next stage, " + }, + { + "bbox": [ + 55, + 134, + 295, + 312 + ], + "type": "inline_equation", + "content": "X_{s_1}^{edit}" + }, + { + "bbox": [ + 55, + 134, + 295, + 312 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 134, + 295, + 312 + ], + "type": "inline_equation", + "content": "X_{s_1}^{src}" + }, + { + "bbox": [ + 55, + 134, + 295, + 312 + ], + "type": "text", + "content": ". However, FlowEdit operates within the target distribution, meaning we require the end point for this stage, " + }, + { + "bbox": [ + 55, + 134, + 295, + 312 + ], + "type": "inline_equation", + "content": "X_{e_1}^{edit}" + }, + { + "bbox": [ + 55, + 134, + 295, + 312 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 134, + 295, + 312 + ], + "type": "inline_equation", + "content": "X_{e_1}^{src}" + }, + { + "bbox": [ + 55, + 134, + 295, + 312 + ], + "type": "text", + "content": ". The latter, " + }, + { + "bbox": [ + 55, + 134, + 295, + 312 + ], + "type": "inline_equation", + "content": "X_{e_1}^{src}" + }, + { + "bbox": [ + 55, + 134, + 295, + 312 + ], + "type": "text", + "content": ", can be obtained through interpolation between " + }, + { + "bbox": [ + 55, + 134, + 295, + 312 + ], + "type": "inline_equation", + "content": "X_0" + }, + { + "bbox": [ + 55, + 134, + 295, + 312 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 134, + 295, + 312 + ], + "type": "inline_equation", + "content": "X_1^{src}" + }, + { + "bbox": [ + 55, + 134, + 295, + 312 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 55, + 134, + 295, + 312 + ], + "type": "inline_equation", + "content": "(X_{e_1}^{src} = \\bar{\\mathbf{x}}_{e_1}^{src})" + }, + { + "bbox": [ + 55, + 134, + 295, + 312 + ], + "type": "text", + "content": ". To estimate the edited endpoint " + }, + { + "bbox": [ + 55, + 134, + 295, + 312 + ], + "type": "inline_equation", + "content": "X_{e_1}^{edit}" + }, + { + "bbox": [ + 55, + 134, + 295, + 312 + ], + "type": "text", + "content": ", we first compute the difference between the start and end points of the source prompt in this stage. This difference is then added to the edited start latent variable, providing an estimate of " + }, + { + "bbox": [ + 55, + 134, + 295, + 312 + ], + "type": "inline_equation", + "content": "X_{e_1}^{edit}" + }, + { + "bbox": [ + 55, + 134, + 295, + 312 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 318, + 289, + 333 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 318, + 289, + 333 + ], + "spans": [ + { + "bbox": [ + 61, + 318, + 289, + 333 + ], + "type": "interline_equation", + "content": "\\bar {\\mathbf {x}} _ {e _ {1}} ^ {s r c} = X _ {s _ {1}} ^ {s r c} + \\bigl (\\bar {\\mathbf {x}} _ {e _ {1}} ^ {s r c} - X _ {s _ {1}} ^ {s r c} \\bigr), X _ {e _ {1}} ^ {\\mathrm {e d i t}} = X _ {s _ {1}} ^ {\\mathrm {e d i t}} + \\bigl (\\bar {\\mathbf {x}} _ {e _ {1}} ^ {s r c} - X _ {s _ {1}} ^ {s r c} \\bigr).", + "image_path": "fb66fd545a6319b3f7232be2efb506abf766b68177ca0a8a51ef5900ed31c13d.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 339, + 296, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 339, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 339, + 296, + 399 + ], + "type": "text", + "content": "This step ensures that the edited and reconstructed latents remain aligned with the source image distribution while incorporating the desired modifications, preserving the integrity of the generated features. This is key to the success of FlowEdit under the pyramid structure." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 399, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 296, + 495 + ], + "type": "text", + "content": "Another key point is that, starting from the second frame, Pyramid-Edit respectively incorporates the source and edited historical information from previous frames (e.g., the " + }, + { + "bbox": [ + 55, + 399, + 296, + 495 + ], + "type": "inline_equation", + "content": "(i - 1)" + }, + { + "bbox": [ + 55, + 399, + 296, + 495 + ], + "type": "text", + "content": "-th frame) as a conditioning factor within its noise-estimation network. This allows the model to leverage temporal dependencies from both the source and edited frames, enhancing the temporal consistency. The source and edited sampling processes can be simplified as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 501, + 291, + 553 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 501, + 291, + 553 + ], + "spans": [ + { + "bbox": [ + 61, + 501, + 291, + 553 + ], + "type": "interline_equation", + "content": "\\underbrace {\\ldots \\rightarrow D o w n (X _ {1} ^ {s r c , i - 2} , 2 ^ {k + 1}) \\rightarrow D o w n (X _ {1} ^ {s r c , i - 1} , 2 ^ {k})} _ {\\mathrm {S o u r c e H i s t y o r c o n d i t i o n}} \\rightarrow X _ {t _ {k}} ^ {s r c , i},", + "image_path": "1a0322285dd1aba614a7fafcc7575274e4929edbdbd905079e2ca928801a0710.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 561, + 295, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 561, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 561, + 295, + 586 + ], + "type": "text", + "content": "For generating subsequent frames, simply introduce the superscripts " + }, + { + "bbox": [ + 55, + 561, + 295, + 586 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 561, + 295, + 586 + ], + "type": "text", + "content": " as indices for the historical latents in Eq. (8)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 591, + 126, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 591, + 126, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 591, + 126, + 601 + ], + "type": "text", + "content": "4.2.2. Wan-Edit" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": "Wan2.1 [40] is a state-of-the-art T2V generation model built on the mainstream DiT framework [26], achieving significant advancements through innovations such as a novel 3D VAE, scalable training strategies, large-scale data curation. These contributions enable Wan2.1 to generate high-quality, temporally consistent videos with improved efficiency and scalability. We focus on adapting its core framework for video editing, leveraging its strengths in temporal modeling and high-fidelity generation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "content": "Compared to Pyramid Flow, Wan2.1 adopts a much simpler architecture. It consists of 30 WanAttentionBlock layers, each integrating vision self-attention, text-to-vision cross-attention, and feed-forward mechanisms. Multiframe video inputs are encoded into a 3D vision latent by the 3D VAE encoder. These 3D vision latents are then flattened into a sequence and fed into the self-attention layers to encode temporal consistency. Finally, the sequence interacts with text tokens through cross-attention to achieve T2V generation. The self-attention mechanism captures intra-frame dependencies, while the cross-attention module integrates text tokens to enable text-guided video generation. This joint modeling of visual and textual information ensures high-quality, semantically aligned video outputs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 239, + 554, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 554, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 554, + 335 + ], + "type": "text", + "content": "Wan-Edit. This elegant and streamlined architecture closely resembles that of image generation models, enabling FlowEdit to be directly integrated without significant modifications. For multi-frame inputs from " + }, + { + "bbox": [ + 313, + 239, + 554, + 335 + ], + "type": "inline_equation", + "content": "i_1" + }, + { + "bbox": [ + 313, + 239, + 554, + 335 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 239, + 554, + 335 + ], + "type": "inline_equation", + "content": "i_2" + }, + { + "bbox": [ + 313, + 239, + 554, + 335 + ], + "type": "text", + "content": ", denote as " + }, + { + "bbox": [ + 313, + 239, + 554, + 335 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_1^{src} = [X_1^{src,i_1},\\dots ,X_1^{src,i_2}]" + }, + { + "bbox": [ + 313, + 239, + 554, + 335 + ], + "type": "text", + "content": ", the editing process handles all frames simultaneously by extending FlowEdit to the entire sequence, ensuring temporal consistency and efficient processing. The process is defined as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 321, + 341, + 553, + 356 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 341, + 553, + 356 + ], + "spans": [ + { + "bbox": [ + 321, + 341, + 553, + 356 + ], + "type": "interline_equation", + "content": "d \\mathbf {X} _ {t} ^ {\\text {e d i t}} = \\mathbb {E} \\left[ v _ {t} ^ {\\Delta} \\left(\\mathbf {X} _ {t} ^ {\\text {s r c}}, \\mathbf {X} _ {t} ^ {\\text {e d i t}} + \\mathbf {X} _ {t} ^ {\\text {s r c}} - \\mathbf {X} _ {1} ^ {\\text {s r c}}\\right) \\mid \\mathbf {X} _ {1} ^ {\\text {s r c}} \\right] d t \\tag {9}", + "image_path": "634c45ba7e91fb0fea47ddbf436ce8da83f4e75ede84eb66da4106f56d9c95eb.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 361, + 553, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 553, + 384 + ], + "type": "text", + "content": "This allows for smooth and coherent edits across all frames, preserving the temporal dynamics of the video." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 394, + 395, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 394, + 395, + 407 + ], + "spans": [ + { + "bbox": [ + 314, + 394, + 395, + 407 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 415, + 554, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 554, + 498 + ], + "type": "text", + "content": "Baseline methods. We compare our RF-based editing methods with five diffusion-based models: TokenFlow [9], DMT [51], VidToMe [21], AnyV2V [17], and VideoGrain [50]. Diffusion-based models require inversion before editing, while our RF-based methods are inversion-free. DMT relies on optimization during editing, and VideoGrain uses depth maps to assist editing." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 498, + 554, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 629 + ], + "type": "text", + "content": "Evaluation metrics. We fairly evaluate all models on our FiVE-Bench using 15 metrics. The nine commonly used metrics in Table 2 cover six aspects: structure distance [38], background preservation (PSNR, LPIPS [52], MSE, and SSIM [42] outside the editing mask), edit prompt-image consistency (CLIPSIM [44] for the full image and masked regions), image quality assessment (NIQE [33]), temporal consistency (motion fidelity score [51]) and running time. Additionally, our proposed FiVE-Acc introduces five metrics in Table. 3 to evaluate the accuracy of successful object editing." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 630, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 689 + ], + "type": "text", + "content": "Implementation details. Pyramid-Edit uses the 384P Pyramid-Flow model, and Wan-Edit on the Wan2.1 1.3B model. Pyramid-Edit uses 20 timesteps for the first frame and 10 for subsequent frames, while Wan-Edit uses 50 timesteps. All experiments were run on a single H100 GPU." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Due to space limitations, see the Appendix for details on comparison methods, metrics, and implementation." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16677" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 96, + 551, + 220 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 2. Comparison of diffusion- and flow-based video editing methods on our proposed FiVE-Bench. * and † denote methods that require optimization and depth/segmentation maps, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 96, + 551, + 220 + ], + "lines": [ + { + "bbox": [ + 63, + 96, + 551, + 220 + ], + "spans": [ + { + "bbox": [ + 63, + 96, + 551, + 220 + ], + "type": "table", + "html": "
MethodsStructure Dist.×103↓Background PreservationText AlignmentIQA NIQE↓Motion Fidelity S.×102↑Time (s) Per Frame↓
PSNR↑LPIPS×103↓MSE×104↓SSIM×102↑CLIPS↑CLIPS-edit ↑
Source Videos0.00∞0.000.00100.0024.5919.876.3393.76-
TokenFlow [9]35.6219.06263.61138.6572.5126.4621.154.0189.008.04
DMT* [51]85.9514.71404.60372.7851.6426.6621.445.2482.3025.98
VidToMe [21]22.3721.15263.9188.7570.6926.8421.054.6890.063.25
AnyV2V [17]71.3615.90348.59342.9750.7724.8919.725.0460.366.11
VideoGrain† [50]12.4027.05185.2125.1079.1325.6920.314.0888.5727.12
RFs (Ours)Pyramid-Edit28.6520.84276.5995.6371.7226.8220.205.4880.591.44
Wan-Edit12.5325.5794.6141.8482.5526.3921.236.5489.433.07
", + "image_path": "fbe2db0efb4527685344ac157c73a7e4a8917db339c0f5dc7c66d85011d55a7c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 63, + 264, + 288, + 360 + ], + "blocks": [ + { + "bbox": [ + 55, + 238, + 295, + 261 + ], + "lines": [ + { + "bbox": [ + 55, + 238, + 295, + 261 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 295, + 261 + ], + "type": "text", + "content": "Table 3. Comparison of diffusion- and flow-based video editing methods on the FiVE-Bench using FiVE-Acc metrics." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 264, + 288, + 360 + ], + "lines": [ + { + "bbox": [ + 63, + 264, + 288, + 360 + ], + "spans": [ + { + "bbox": [ + 63, + 264, + 288, + 360 + ], + "type": "table", + "html": "
MethodYN-AccMC-AccU-Acc∩-AccFiVE-Acc↑
TokenFlow [9]19.3635.5136.6818.1827.43
DMT* [51]34.7862.0662.9833.8648.42
VidToMe [21]20.0333.5036.2017.3426.77
AnyV2V [17]30.6245.4248.9627.0938.02
VideoGrain† [50]30.5043.9744.3030.1737.23
Pyramid-Edit33.6754.0156.3631.3143.84
Wan-Edit41.4152.5355.7238.2246.97
", + "image_path": "b3e95b30cb72bb11c9e104ddf2c0939bc798116d3699e837978b6a30233ee0f1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 382, + 198, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 382, + 198, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 382, + 198, + 396 + ], + "type": "text", + "content": "5.1. Quantitative Comparison" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 401, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 401, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 54, + 401, + 295, + 581 + ], + "type": "text", + "content": "Comparison on Common Metrics. From the results, we observe that VideoGrain achieves the best performance among diffusion models, in terms of Structure Distance (12.40) and Background Preservation (PSNR: 27.05, MSE:25.10). Meanwhile, Wan-Edit, our proposed method, outperforms in the other two Background Preservation metrics (LPIPS: 94.61, SSIM: 82.55). Overall, VideoGrain and Wan-Edit perform similarly, but VideoGrain requires depth map assistance for editing and is extremely slow. In contrast, the inversion-free Wan-Edit demonstrates the highest efficiency with the lowest Time per Frame (1.44 seconds). These findings highlight the strengths and trade-offs of each method, offering valuable insights for selecting the appropriate approach based on specific application needs, and greatly contribute to the advancement of video editing." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": "Comparison on FiVE-Acc Metrics. Our evaluation of diffusion- and flow-based video editing methods on FiVE-Acc metrics reveals key findings. DMT, an optimization-based method, excels in MC-Acc (62.06%), U-Acc (62.98%), and overall FiVE-Acc (48.42%), highlighting its strength in fine-grained text alignment. This is due to its iterative optimization process, which prioritizes text fidelity, sometimes at the cost of background preservation. In contrast, Wan-Edit performs well in YN-Acc (41.41%) and " + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\cap" + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": "-Acc (38.22%). Pyramid-Edit also shows competitive results, especially in MC-Acc (54.01%) and U" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 328, + 253, + 545, + 308 + ], + "blocks": [ + { + "bbox": [ + 314, + 238, + 553, + 249 + ], + "lines": [ + { + "bbox": [ + 314, + 238, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 314, + 238, + 553, + 249 + ], + "type": "text", + "content": "Table 4. Human vs. VLM model evaluation on FiVE-Acc metrics." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 253, + 545, + 308 + ], + "lines": [ + { + "bbox": [ + 328, + 253, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 328, + 253, + 545, + 308 + ], + "type": "table", + "html": "
EvaluatorYN-AccMC-Acc∪-Acc∩-AccFiVE-Acc↑
Qwen-2.5-VL41.4152.5355.7238.2246.97
Human44.3750.3151.9842.7047.34
Δ (Diff)+2.96-2.22-3.74+4.48+0.37
", + "image_path": "3313668fe285f9d8f3fe058aff6f68567e24c434319e1f6e4a0d47d16cd01527.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 328, + 553, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 328, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 328, + 553, + 376 + ], + "type": "text", + "content": "Acc (56.36%), emphasizing the importance of fine-grained alignment. U-Acc and " + }, + { + "bbox": [ + 313, + 328, + 553, + 376 + ], + "type": "inline_equation", + "content": "\\cap" + }, + { + "bbox": [ + 313, + 328, + 553, + 376 + ], + "type": "text", + "content": "-Acc metrics offer insights into task consistency, emphasizing the need for VLM-based evaluation to ensure visual and semantic coherence in editing." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 376, + 554, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 554, + 435 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 554, + 435 + ], + "type": "text", + "content": "Overall, VideoGrain and Wan-Edit strike the best balance across metrics for fine-grained video editing. Wan-Edit outperforms VideoGrain in speed, offering faster and more stable results due to enhanced RF video generation and inversion models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 436, + 554, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 554, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 554, + 509 + ], + "type": "text", + "content": "Human Preference Validation for FiVE-Acc. We conducted a human study on 16 randomly sampled videos with 64 prompts, based on Wan-Edit editing outputs. Table 4 shows that human scores closely match Qwen-2.5-VL's results (47.34 vs. 46.97), confirming FiVE-Acc aligns well with human perception and offers consistent evaluation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 516, + 450, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 516, + 450, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 516, + 450, + 529 + ], + "type": "text", + "content": "5.2. Qualitative Comparison" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 534, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 534, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 534, + 555, + 714 + ], + "type": "text", + "content": "In Fig. 4, analyzing by edit type, non-rigid transformations and color changes are the easiest fine-grained edits, successfully handled by almost all methods. Object addition and rigid transformations are only successfully performed by the optimization-based DMT and our Wan-Edit. All methods fail at object removal, making it the most challenging task. From the perspective of background preservation in fine-grained editing, DMT and VidToME introduce significant alterations, while the recently proposed VideoGrain and the RF-based Pyramid-Edit and Wan-Edit largely retain the original background. The qualitative analysis further validates the quantitative findings: DMT achieves the best text alignment but disrupts the background, while VideoGrain and Wan-Edit offer the best balance across all metrics. Please refer to Appendix for details." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16678" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 65, + 72, + 234, + 80 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 72, + 234, + 80 + ], + "spans": [ + { + "bbox": [ + 65, + 72, + 234, + 80 + ], + "type": "text", + "content": "Edit1 Object (w/o non-rigid deform): Woman " + }, + { + "bbox": [ + 65, + 72, + 234, + 80 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 65, + 72, + 234, + 80 + ], + "type": "text", + "content": " Man" + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 63, + 83, + 548, + 171 + ], + "blocks": [ + { + "bbox": [ + 63, + 83, + 548, + 171 + ], + "lines": [ + { + "bbox": [ + 63, + 83, + 548, + 171 + ], + "spans": [ + { + "bbox": [ + 63, + 83, + 548, + 171 + ], + "type": "image", + "image_path": "d51bc72ab52e0991c3f6613e73093db934209f0d905111b10dd0ad72cc5eb293.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 63, + 171, + 548, + 250 + ], + "blocks": [ + { + "bbox": [ + 63, + 171, + 548, + 250 + ], + "lines": [ + { + "bbox": [ + 63, + 171, + 548, + 250 + ], + "spans": [ + { + "bbox": [ + 63, + 171, + 548, + 250 + ], + "type": "image", + "image_path": "2249861f1ef7256f5797f79bfde2c1e68c1c9201a597c0cca2fcf0c230bf1a1b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 65, + 251, + 180, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 251, + 180, + 258 + ], + "spans": [ + { + "bbox": [ + 65, + 251, + 180, + 258 + ], + "type": "text", + "content": "Edit3 Color: black dress " + }, + { + "bbox": [ + 65, + 251, + 180, + 258 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 65, + 251, + 180, + 258 + ], + "type": "text", + "content": " red dress" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 63, + 259, + 548, + 338 + ], + "blocks": [ + { + "bbox": [ + 63, + 259, + 548, + 338 + ], + "lines": [ + { + "bbox": [ + 63, + 259, + 548, + 338 + ], + "spans": [ + { + "bbox": [ + 63, + 259, + 548, + 338 + ], + "type": "image", + "image_path": "c5fcfe4470254a9f0e1cd6ab8f3a6d25ede2b922bc75fbb2e84ea6404bedd489.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 65, + 339, + 214, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 339, + 214, + 347 + ], + "spans": [ + { + "bbox": [ + 65, + 339, + 214, + 347 + ], + "type": "text", + "content": "Edit4 Material: A woman " + }, + { + "bbox": [ + 65, + 339, + 214, + 347 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 65, + 339, + 214, + 347 + ], + "type": "text", + "content": " A porcelain woman" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 63, + 349, + 548, + 437 + ], + "blocks": [ + { + "bbox": [ + 63, + 349, + 548, + 437 + ], + "lines": [ + { + "bbox": [ + 63, + 349, + 548, + 437 + ], + "spans": [ + { + "bbox": [ + 63, + 349, + 548, + 437 + ], + "type": "image", + "image_path": "ae21f1ee99f227fd8b908a2e08fb78d8d8b23aea32d714737009a142f27a5871.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 64, + 442, + 253, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 442, + 253, + 452 + ], + "spans": [ + { + "bbox": [ + 64, + 442, + 253, + 452 + ], + "type": "text", + "content": "EditS Add: A woman " + }, + { + "bbox": [ + 64, + 442, + 253, + 452 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 64, + 442, + 253, + 452 + ], + "type": "text", + "content": " A woman followed closely by a dog" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 62, + 453, + 548, + 543 + ], + "blocks": [ + { + "bbox": [ + 62, + 453, + 548, + 543 + ], + "lines": [ + { + "bbox": [ + 62, + 453, + 548, + 543 + ], + "spans": [ + { + "bbox": [ + 62, + 453, + 548, + 543 + ], + "type": "image", + "image_path": "844b9bde26cec22be6b4247f746b8092d15081a3412e02d36987849bf727070a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 161, + 544, + 447, + 556 + ], + "lines": [ + { + "bbox": [ + 161, + 544, + 447, + 556 + ], + "spans": [ + { + "bbox": [ + 161, + 544, + 447, + 556 + ], + "type": "text", + "content": "Figure 4. Editing results across six editing types and five comparison methods." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 577, + 295, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 577, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 577, + 295, + 662 + ], + "type": "text", + "content": "Limitations and Future Work. This work simply adapts image editing to video editing, which leads to room for improvement in the quality of Pyramid-Edit. In future work, we aim to further explore advancements in RF models for video editing. Additionally, handling challenging cases in the benchmark, such as large motions or object removal, will be an important direction for improvement." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 672, + 128, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 128, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 128, + 684 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "We introduce FiVE-Bench, a benchmark for fine-grained video editing, and propose the VLM-based FiVE-Acc met" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 577, + 555, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 577, + 555, + 709 + ], + "spans": [ + { + "bbox": [ + 313, + 577, + 555, + 709 + ], + "type": "text", + "content": "ric, which evaluates the accuracy of object-level editing success. Additionally, we adapt two RF-based video editing methods, Pyramid-Edit and Wan-Edit. To the best of our knowledge, this is the first comprehensive quantitative and qualitative comparison of emerging diffusion-based and flow-based models in the video editing community. FiVE-Bench and FiVE-Acc provide a standardized framework for fine-grained video editing evaluation, guiding future advancements in efficient and high-fidelity video editing solutions. We hope this benchmark drives innovation in both research and real-world applications." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16679" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 146, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 146, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 296, + 190 + ], + "type": "text", + "content": "[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In IEEE Conf. Comput. Vis. Pattern Recog., pages 18392-18402, 2023. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 191, + 295, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 191, + 295, + 224 + ], + "spans": [ + { + "bbox": [ + 62, + 191, + 295, + 224 + ], + "type": "text", + "content": "[3] Duygu Ceylan, Chun-Hao P Huang, and Niloy J Mitra. Pix2video: Video editing using image diffusion. In Int. Conf. Comput. Vis., pages 23206-23217, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 224, + 294, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 224, + 294, + 267 + ], + "spans": [ + { + "bbox": [ + 62, + 224, + 294, + 267 + ], + "type": "text", + "content": "[4] Weifeng Chen, Jie Wu, Pan Xie, Hefeng Wu, Jiashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-a-video: Controllable text-to-video generation with diffusion models. arXiv preprint arXiv:2305.13840, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 269, + 295, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 269, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 269, + 295, + 312 + ], + "type": "text", + "content": "[5] Yingying Deng, Xiangyu He, Changwang Mei, Peisong Wang, and Fan Tang. Fireflow: Fast inversion of rectified flow for image semantic editing. arXiv preprint arXiv:2412.07517, 2024. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 314, + 294, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 314, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 314, + 294, + 346 + ], + "type": "text", + "content": "[6] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Adv. Neural Inform. Process. Syst., 34:8780-8794, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 347, + 295, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 347, + 295, + 390 + ], + "spans": [ + { + "bbox": [ + 62, + 347, + 295, + 390 + ], + "type": "text", + "content": "[7] Wenkai Dong, Song Xue, Xiaoyue Duan, and Shumin Han. Prompt tuning inversion for text-driven image editing using diffusion models. In Int. Conf. Comput. Vis., pages 7430-7440, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 392, + 294, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 294, + 446 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 294, + 446 + ], + "type": "text", + "content": "[8] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Int. Conf. on Machine Learning, 2024. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 447, + 294, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 294, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 294, + 479 + ], + "type": "text", + "content": "[9] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. In Int. Conf. Learn. Represent., 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 480, + 294, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 480, + 294, + 524 + ], + "spans": [ + { + "bbox": [ + 57, + 480, + 294, + 524 + ], + "type": "text", + "content": "[10] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff:Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 525, + 294, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 525, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 57, + 525, + 294, + 557 + ], + "type": "text", + "content": "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Adv. Neural Inform. Process. Syst., 33:6840-6851, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 558, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 558, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 57, + 558, + 294, + 612 + ], + "type": "text", + "content": "[12] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 614, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 294, + 658 + ], + "type": "text", + "content": "[13] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 3, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 294, + 712 + ], + "type": "text", + "content": "[14] Yang Jin, Zhicheng Sun, Ningyuan Li, Kun Xu, Hao Jiang, Nan Zhuang, Quzhe Huang, Yang Song, Yadong Mu, and Zhouchen Lin. Pyramidal flow matching for efficient video generative modeling. arXiv preprint arXiv:2410.05954, 2024. 1, 2, 3, 5" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[15] Xuan Ju, Ailing Zeng, Yuxuan Bian, Shaoteng Liu, and Qiang Xu. Pnp inversion: Boosting diffusion-based editing with 3 lines of code. International Conference on Learning Representations (ICLR), 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 118, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 173 + ], + "type": "text", + "content": "[16] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 175, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 553, + 217 + ], + "type": "text", + "content": "[17] Max Ku, Cong Wei, Weiming Ren, Harry Yang, and Wenhu Chen. Anyv2v: A tuning-free framework for any video-to-video editing tasks. arXiv preprint arXiv:2403.14468, 2024. 3, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "type": "text", + "content": "[18] Vladimir Kulikov, Matan Kleiner, Inbar Huberman-Spiegelglas, and Tomer Michaeli. Flowedit: Inversion-free text-based editing using pre-trained flow models. arXiv preprint arXiv:2412.08629, 2024. 1, 2, 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 264, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 285 + ], + "type": "text", + "content": "[19] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux.2,3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "type": "text", + "content": "[20] Minghan Li, Shuai Li, Xindong Zhang, and Lei Zhang. Univs: Unified and universal video segmentation with prompts as queries. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3227-3238, 2024. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 332, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 374 + ], + "type": "text", + "content": "[21] Xirui Li, Chao Ma, Xiaokang Yang, and Ming-Hsuan Yang. Vidthome: Video token merging for zero-shot video editing. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7486-7495, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 376, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 409 + ], + "type": "text", + "content": "[22] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. In Int. Conf. Learn. Represent., 2023. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 411, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 443 + ], + "type": "text", + "content": "[23] Xingchao Liu, Chengyue Gong, et al. Flow straight and fast: Learning to generate and transfer data with rectified flow. In Int. Conf. Learn. Represent., 2023. 2, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 445, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 553, + 498 + ], + "type": "text", + "content": "[24] Guoqing Ma, Haoyang Huang, Kun Yan, Liangyu Chen, Nan Duan, Shengming Yin, Changyi Wan, Ranchen Ming, Xiaoniu Song, Xing Chen, et al. Step-video-t2v technical report: The practice, challenges, and future of video foundation model. arXiv preprint arXiv:2502.10248, 2025. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 500, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 553, + 544 + ], + "type": "text", + "content": "[25] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6038–6047, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 545, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 567 + ], + "type": "text", + "content": "[26] OpenAI. Video generation models as world simulators, 2024. 1, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 568, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 568, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 317, + 568, + 553, + 633 + ], + "type": "text", + "content": "[27] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 635, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 679 + ], + "type": "text", + "content": "[28] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. arXiv preprint arXiv:2303.09535, 2023. 2, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[29] Bosheng Qin, Juncheng Li, Siliang Tang, Tat-Seng Chua, and Yueting Zhuang. Instructvid2vid: Controllable video editing with natural language instructions. In 2024 IEEE" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16680" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 171 + ], + "type": "text", + "content": "[30] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dólar, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos, 2024. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 294, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 294, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 294, + 218 + ], + "type": "text", + "content": "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10684-10695, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 272 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 272 + ], + "type": "text", + "content": "[32] Litu Rout, Yujia Chen, Nataniel Ruiz, Constantine Caramanis, Sanjay Shakkottai, and Wen-Sheng Chu. Semantic image inversion and editing using rectified stochastic differential equations. arXiv preprint arXiv:2410.10792, 2024. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 274, + 294, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 274, + 294, + 329 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 294, + 329 + ], + "type": "text", + "content": "[33] Michele A Saad and Alan C Bovik. Blind quality assessment of videos using a model of natural scene statistics and motion coherency. In 2012 Conference Record of the Forty Sixth Asilomar Conference on Signals, Systems and Computers (ASILOMAR), pages 332-336. IEEE, 2012. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 294, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 375 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 375 + ], + "type": "text", + "content": "[34] Uriel Singer, Amit Zohar, Yuval Kirstain, Shelly Sheynin, Adam Polyak, Devi Parikh, and Yaniv Taigman. Video editing via factorized diffusion distillation. In Eur. Conf. Comput. Vis., pages 450-466. Springer, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 409 + ], + "type": "text", + "content": "[35] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 293, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 293, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 293, + 432 + ], + "type": "text", + "content": "[36] Spencer Sterling. Zeroscope. https://huggingface.co/cerspense/zeroscope_v2_576w, 2023.1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 434, + 294, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 477 + ], + "type": "text", + "content": "[37] Shuyuan Tu, Qi Dai, Zhi-Qi Cheng, Han Hu, Xintong Han, Zuxuan Wu, and Yu-Gang Jiang. Motioneditor: Editing video motion via content-aware diffusion. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7882-7891, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 479, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 294, + 521 + ], + "type": "text", + "content": "[38] Narek Tumanyan, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Splicing vit features for semantic appearance transfer. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10748-10757, 2022. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 523, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 566 + ], + "type": "text", + "content": "[39] Bram Wallace, Akash Gokul, and Nikhil Naik. Edict: Exact diffusion inversion via coupled transformations. In IEEE Conf. Comput. Vis. Pattern Recog., pages 22532-22541, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 568, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 622 + ], + "type": "text", + "content": "[40] Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, et al. Wan: Open and advanced large-scale video generative models. arXiv preprint arXiv:2503.20314, 2025. 1, 2, 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "type": "text", + "content": "[41] Jiangshan Wang, Junfu Pu, Zhongang Qi, Jiayi Guo, Yue Ma, Nisha Huang, Yuxin Chen, Xiu Li, and Ying Shan. Taming rectified flow for inversion and editing. arXiv preprint arXiv:2411.04746, 2024. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[42] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process., 13(4): 600-612, 2004. 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 689 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "type": "text", + "content": "[43] Yujie Wei, Shiwei Zhang, Zhiwu Qing, Hangjie Yuan, Zhiheng Liu, Yu Liu, Yingya Zhang, Jingren Zhou, and Hongming Shan. Dreamvideo: Composing your dream videos with customized subject and motion. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6537-6549, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "text", + "content": "[44] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. GODIVA: Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806, 2021. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 174, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 553, + 217 + ], + "type": "text", + "content": "[45] Chen Henry Wu and Fernando De la Torre. A latent space of stochastic diffusion models for zero-shot image editing and guidance. In Int. Conf. Comput. Vis., pages 7378-7387, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 219, + 553, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 273 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 273 + ], + "type": "text", + "content": "[46] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Int. Conf. Comput. Vis., pages 7623-7633, 2023. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 274, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 553, + 319 + ], + "type": "text", + "content": "[47] Jay Zhangjie Wu, Xiuyu Li, Difei Gao, Zhen Dong, Jinbin Bai, Aishani Singh, Xiaoyu Xiang, Youzeng Li, Zuwei Huang, Yuanxi Sun, et al. Cvpr 2023 text guided video editing competition. arXiv preprint arXiv:2310.16003, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 319, + 553, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 362 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 362 + ], + "type": "text", + "content": "[48] Chenxi Xie, Minghan Li, Shuai Li, Yuhui Wu, Qiaosi Yi, and Lei Zhang. Dnaedit: Direct noise alignment for text-guided rectified flow editing. arXiv preprint arXiv:2506.01430, 2025.3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 364, + 553, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 364, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 316, + 364, + 553, + 407 + ], + "type": "text", + "content": "[49] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. In SIGGRAPH Asia 2023 Conference Papers, pages 1-11, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 409, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 409, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 409, + 553, + 453 + ], + "type": "text", + "content": "[50] Xiangpeng Yang, Linchao Zhu, Hehe Fan, and Yi Yang. Videograin: Modulating space-time attention for multigrained video editing. arXiv preprint arXiv:2502.17258, 2025. 2, 3, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 454, + 553, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 454, + 553, + 497 + ], + "spans": [ + { + "bbox": [ + 316, + 454, + 553, + 497 + ], + "type": "text", + "content": "[51] Danah Yatim, Rafail Fridman, Omer Bar-Tal, Yoni Kasten, and Tali Dekel. Space-time diffusion features for zero-shot text-driven motion transfer. In IEEE Conf. Comput. Vis. Pattern Recog., pages 8466-8476, 2024. 3, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 498, + 553, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 553, + 542 + ], + "type": "text", + "content": "[52] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conf. Comput. Vis. Pattern Recog., pages 586-595, 2018. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 544, + 553, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 553, + 576 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 553, + 576 + ], + "type": "text", + "content": "[53] Zhenghao Zhang, Zuozhuo Dai, Long Qin, and Weizhi Wang. Effived: Efficient video editing via text-instruction diffusion models. arXiv preprint arXiv:2403.11568, 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 577, + 553, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 577, + 553, + 631 + ], + "spans": [ + { + "bbox": [ + 316, + 577, + 553, + 631 + ], + "type": "text", + "content": "[54] Zhixing Zhang, Yanyu Li, Yushu Wu, Yanwu Xu, Anil Kag, Ivan Skorokhodov, Willi Menapace, Aliaksandr Siarohin, Junli Cao, Dimitris Metaxas, et al. Sf-v: Single forward video generation model. arXiv preprint arXiv:2406.04324, 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 633, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 633, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 633, + 553, + 689 + ], + "type": "text", + "content": "[55] Rui Zhao, Yuchao Gu, Jay Zhangjie Wu, David Junhao Zhang, Jia-Wei Liu, Weijia Wu, Jussi Keppo, and Mike Zheng Shou. Motiondirector: Motion customization of text-to-video diffusion models. In Eur. Conf. Comput. Vis., pages 273–290. Springer, 2025. 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16681" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/FiffDepth_ Feed-forward Transformation of Diffusion-Based Generators for Detailed Depth Estimation/3c8dc7fb-98b2-4d77-857e-35d78d6eaaf6_content_list.json b/2025/FiffDepth_ Feed-forward Transformation of Diffusion-Based Generators for Detailed Depth Estimation/3c8dc7fb-98b2-4d77-857e-35d78d6eaaf6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f9cb28493c1bda18eef9ad61ee66b269a1f94d5d --- /dev/null +++ b/2025/FiffDepth_ Feed-forward Transformation of Diffusion-Based Generators for Detailed Depth Estimation/3c8dc7fb-98b2-4d77-857e-35d78d6eaaf6_content_list.json @@ -0,0 +1,2041 @@ +[ + { + "type": "text", + "text": "FiffDepth: Feed-forward Transformation of Diffusion-Based Generators for Detailed Depth Estimation", + "text_level": 1, + "bbox": [ + 117, + 128, + 880, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yunpeng Bai Qixing Huang", + "bbox": [ + 364, + 204, + 630, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The University of Texas at Austin", + "bbox": [ + 361, + 222, + 630, + 239 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2936be9c03d6170cca4e7ff8b69beca682c98a4a7ad8844ccbb4a5115999ca38.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 130, + 290, + 305, + 369 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7fd8549e482b0256e8162492758759437c39e715d5599b4dee427bb8d00d32e0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 130, + 369, + 303, + 469 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c584bded786f54295d81edc2f2e3b04cd8cd41ee241ee17ce1525f5e078a9833.jpg", + "image_caption": [ + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 130, + 469, + 303, + 547 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9ccce071ad3f046138e947a5a88d9547699f897cb26bd5058e5142ed7da6a4e3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 291, + 491, + 368 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1c2c1ec4c1b93b5076006b7e1758325bd37fec7f87565d5a5ab84fc04bbfa9f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 368, + 491, + 469 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/594cfb007bb8a09bc0e1c526a58c1de3808211b0dad9057ae2fc7d99f7e82653.jpg", + "image_caption": [ + "FifDepth" + ], + "image_footnote": [], + "bbox": [ + 316, + 469, + 488, + 547 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ed2630d94aafe8f960a30cd44ed5a0f96f95d849255669835d466cd3248d7c46.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 291, + 676, + 368 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4e83a5ccbcfc041e27e9dde36a21799f2082c2e84fe5f456404a6e12d54eef4d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 368, + 676, + 468 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7573848849c9e461db6024935d95e2c8d09a87903046bf246f59aff950a93011.jpg", + "image_caption": [ + "Depth Anything v2", + "Figure 1. Compared to other methods, our model achieves more accurate details and better generalization in depth estimation. The final row shows the point cloud generated from the estimated depth results, and the corresponding depth map can be referenced in Figure 5." + ], + "image_footnote": [], + "bbox": [ + 503, + 469, + 671, + 546 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1d6c6030d5726e9c8b84718e5544235beab47812d087f2836fbb1c1629fea4db.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 291, + 862, + 368 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a0f45a70f9ff381299db6c9bb83842889e3a7847178667cf63897b1d4b6bc9b8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 369, + 862, + 469 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/18c80cee368663e70893ce5fef6bf4023d4c22bb42c7c4ea6836f27788de9e25.jpg", + "image_caption": [ + "Depth Pro" + ], + "image_footnote": [], + "bbox": [ + 687, + 469, + 854, + 546 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 617, + 326, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Monocular Depth Estimation (MDE) is a fundamental 3D vision problem with numerous applications such as 3D scene reconstruction, autonomous navigation, and AI content creation. However, robust and generalizable MDE remains challenging due to limited real-world labeled data and distribution gaps between synthetic datasets and real data. Existing methods often struggle with real-world test data with low efficiency, reduced accuracy, and lack of detail. To address these issues, we propose an efficient MDE approach named FiftDepth. The key feature of FiftDepth is its use of diffusion priors. It transforms diffusion-based image generators into a feed-forward architecture for detailed depth estimation. FiftDepth preserves key generative features and integrates the strong generalization capabilities of models like DINOv2. Through benchmark evaluations, we demonstrate that FiftDepth achieves exceptional", + "bbox": [ + 88, + 660, + 485, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "accuracy, stability, and fine-grained detail, offering significant improvements in MDE performance against state-of-the-art MDE approaches. The paper's source code is available here: https://yunpeng1998.github.io/FiffDepth/", + "bbox": [ + 511, + 618, + 908, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 724, + 645, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Monocular Depth Estimation (MDE) is a fundamental 3D vision problem with numerous applications in 3D scene reconstruction [46], autonomous navigation [47], and more recently in generating AI-based content [2, 41]. MDE has made significant progress in the era of deep learning [21, 34, 35, 52]. Neural networks trained from paired datasets of images and pixel depth exhibit encouraging results and often outperform non-deep learning based counterparts that build on monocular depth cues. Despite significant progress, fundamental challenges remain in efficiency,", + "bbox": [ + 509, + 750, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6023", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "accuracy, and generalization on diverse in-the-wild data. This is because 1) real depth datasets are usually noisy, and 2) while synthetic data can be used, there exist domain gaps between synthetic datasets and diverse in-the-wild data.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, current MDE research relies primarily on synthetic data due to its high-quality annotations and controlled environments. However, the scale and variety of synthetic datasets remain insufficient for comprehensive training. To address this, synthetic-to-real transfer techniques and the utilization of pre-trained models have emerged as viable solutions. Among pre-trained models, generative networks [37] preserve intricate image details more effectively than feed-forward networks (FFNs) like DINOv2 [31], thus holding greater promise for dense prediction models. However, generative models, while detail-rich, often fall short in synthetic-to-real transfer due to their limited generalization capabilities outside the training domain.", + "bbox": [ + 89, + 154, + 482, + 351 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous studies [15, 21] have adopted pre-trained diffusion models, where the main idea is to directly finetuning pre-trained RGB image diffusion models into depth map diffusion generation models conditioned on images. However, this method may not be ideal, as dense prediction models require certainty over diversity. The introduction of any noise or uncertainty during the generation process by these methods is sub-optimal. In contrast, we observe that simply using the denoising diffusion module in a feedforward manner yields better and more stable results. This method capitalizes on extending the trajectories of image diffusion models into the depth domain, representing a significant advancement in both accuracy and efficiency for generative model-based depth estimation methods.", + "bbox": [ + 89, + 354, + 482, + 565 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, we optimize diffusion trajectories for MDE tasks. To enable the diffusion model to better retain certain detailed generative features when fine-tuned into an MDE model, we preserve the original generative training trajectory while training the model for depth prediction, aiming to maintain the detailed features of the original generative model as much as possible. Furthermore, recognizing the limitations of fintuned diffusion models in maintaining robustness in diverse real-world images—with inaccuracies in predicted depth occurring mainly in the low-frequency components—we leverage the strengths of a DINOv2 [31] based model, which excels at predicting accurate low-frequency depth despite its reduced fine detail. To address this issue, we use the diffusion model itself to learn a filter that refines its inaccurate predictions, producing low-frequency outputs with a detail level similar to that of DINOv2's predictions, thereby matching DINOv2's results and optimizing the low-frequency component of our output. This approach allows us to incorporate a large amount of real image data for training without sacrificing detail preservation of generative models, while simultaneously leveraging DINOv2's strong generalization capabilities to enhance", + "bbox": [ + 89, + 568, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the overall robustness and precision of our MDE models.", + "bbox": [ + 511, + 90, + 888, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, the contributions of our work are as follows: 1) We propose an improved approach for transforming generative models into dense prediction models, specifically for depth prediction tasks, by leveraging diffusion model trajectories in a more stable, feed-forward manner; 2) We introduce a novel distillation method that transfers the robust generalization capabilities of models like DINOv2 to diffusion backbones; 3) Our method demonstrates higher stability, accuracy, and efficiency in depth estimation compared to other approaches based on generative models; 4) Compared to other FFN models, our approach achieves more detailed prediction results, marking a significant advancement in the field of MDE.", + "bbox": [ + 511, + 107, + 903, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 319, + 661, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Depth Estimation", + "text_level": 1, + "bbox": [ + 513, + 344, + 684, + 359 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Depth estimation has always been a widely researched topic, with numerous studies conducted in the past including single-image [1, 13, 14, 24, 28, 32, 50] and video depth estimation [10, 23, 43, 48, 54]. Early efforts in image depth estimation, such as DIW [8] and OASIS [9], focused on predicting relative (ordinal) depth. Subsequent approaches, including MegaDepth [27] and DiverseDepth [56], used extensive collections of photographs from the Internet to develop models that adapt to unseen data, while MiDaS [34] improved generalization by incorporating a diverse range of datasets during training. Recent advances, including DPT [35] and Omnidata [12], adopted transformer-based architectures to improve depth estimation performance.", + "bbox": [ + 511, + 367, + 903, + 564 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, due to limitations in models and data, these methods exhibit very limited generalization in various open-world scenes. Recently, some efforts [15, 21, 51, 52] in image depth estimation have made open-world depth estimation feasible by leveraging the power of vast amounts of unlabeled image data and the capabilities of pre-trained generative models. Although these methods have achieved remarkable progress in the field of depth estimation, challenges such as low efficiency, limited generalization, and insufficient detail preservation remain.", + "bbox": [ + 511, + 564, + 905, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Diffusion Models as Representation Learner", + "text_level": 1, + "bbox": [ + 511, + 726, + 888, + 742 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The diffusion model [3, 37] training process strongly resembles that of denoising autoencoders (DAE) [4, 19, 45], as both are designed to recover clean images from noise-corrupted inputs. Recent studies have shown that semantic image representations learned from diffusion models can be effectively used for various downstream recognition tasks, such as correspondence [59], semantic segmentation [61], and keypoint detection [53]. Notably, the features extracted from diffusion models tend to preserve more intricate details, which has prompted the adoption of pre", + "bbox": [ + 511, + 750, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6024", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4fd3a713f38948f876a77c37a1f7f9105f2631393f5d035b62e6bba210784ae9.jpg", + "image_caption": [ + "Figure 2. Overview of the proposed method. To simplify the representation, all the images we used above correspond to the respective latents. We transform the pre-trained diffusion model into a feed-forward approach for depth prediction, using only the result at $t = 0$ as the output during inference. During training, at $t = 0$ , we use synthetic data to ensure detailed results, while at $t = -1$ , we leverage pseudo-labels generated by DINOv2 for supervision." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 410, + 387 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/00044d553bde24e5bd68df7ff62ded032f3e934065480a77c66d061cade33c36.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 88, + 903, + 388 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "trained diffusion models for dense prediction tasks. Recent studies, including Marigold [21] and GeoWizard [15], rely on the standard diffusion framework and pre-trained parameters to perform dense prediction tasks. Emerging approaches [18, 49, 55] attempt to bypass the stochastic phase of diffusion models by employing deterministic frameworks. However, these adaptations lack deeper exploration of the model's potential, often leading to suboptimal performance and the need for additional post-processing to refine results. Moreover, diffusion-based techniques generally exhibit limited generalization capabilities. For instance, BetterDepth [60] incorporates external depth priors as inputs but remains a stochastic framework and heavily depends on the quality of other models.", + "bbox": [ + 88, + 486, + 485, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 715, + 183, + 729 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview: Feed-forward Transformation of Diffusion Models", + "text_level": 1, + "bbox": [ + 89, + 741, + 483, + 771 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fundamentally, generative models construct mappings between a latent space and the ambient data space. These mappings align closely with the needs of depth estimation and other visual recognition tasks, where precise mappings from image data to the corresponding labels are essential. In these tasks, the scarcity of labeled data often limits the precision of the trained models. In the context of learning from a collection of unlabeled data instances, advanced gen", + "bbox": [ + 89, + 779, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "erative models, trained on massive datasets, are capable of learning robust mappings. They exhibit great promise for transferring knowledge to other visual prediction applications [53, 59, 61]. Our work builds on this approach, further exploring its application in depth estimation.", + "bbox": [ + 511, + 486, + 906, + 564 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, we finetune a well-trained Stable Diffusion (SD) [37] model to construct our model. A diffusion process constructs multiple intermediate states by progressively adding noise to the data $\\mathbf{x}_0$ , defined as $\\mathbf{x}_t = \\sqrt{\\bar{\\alpha}_t}\\mathbf{x}_0 + \\sqrt{1 - \\bar{\\alpha}_t}\\epsilon$ , where $\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ , and $\\bar{\\alpha}_t \\coloneqq \\prod_{s=1}^{t} 1 - \\beta_s$ with noise schedule $\\{\\beta_1, \\ldots, \\beta_T\\}$ . Then, diffusion generative model gradually learns the mapping between two distributions by denoising each step, denoted as $x_T \\to x_{T-1} \\to \\dots \\to x_0$ . The diffusion model is primarily a denoising model $\\epsilon_\\theta$ that follows a loss function $\\mathcal{L}_{\\mathrm{simple}} = \\mathbb{E}_{x_0,\\epsilon,t} \\left[ \\| \\epsilon - \\epsilon_\\theta(x_t,t) \\|^2 \\right]$ .", + "bbox": [ + 511, + 571, + 908, + 741 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In depth estimation, the mapping from visual images to depth labels should ideally be deterministic. By leveraging a robust, pre-trained mapping within the generative model, there is no need to decompose the image-label mapping into multiple steps during training. Instead, we can extend the mapping trajectories of the existing diffusion model directly into the depth domain by adapting the learned diffusion process to act as a deterministic one-step feed-forward network. Since our MDE model is built upon the diffusion trajectory for its extension, we set the time step input as", + "bbox": [ + 511, + 750, + 910, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6025", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$t = 0$ in this feed-forward step.", + "bbox": [ + 89, + 90, + 299, + 104 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {d} _ {0} = \\hat {\\boldsymbol {\\epsilon}} _ {\\theta} \\left(\\mathbf {x} _ {0}, t = 0\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 119, + 482, + 135 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since we use diffusion model's parameters to construct our network, we also use $\\hat{\\epsilon}_{\\theta}$ to represent our network model here. The input $\\mathbf{x}_0$ is the latent representation of RGB image, and $\\mathbf{d}_0$ is the latent representation of \"depth image,\" following Marigold's encoding approach. The depth latent $\\mathbf{d}_0$ can be reconstructed into a depth map using the VAE of SD with negligible error. While there have been preliminary attempts [18, 49] to use similar approaches, they remain in a nascent stage and lack the precision, robustness, and richness in detail needed for effective depth estimation. In the following sections, we will delve into each step of our method, detailing how we adapt the diffusion model to enhance accuracy and detail in depth estimation tasks, such as the key technical contributions of preserving diffusion trajectories and improving synthetic-to-real robustness, as discussed in Section 3.2 and Section 3.3, respectively. The overall workflow of the method is illustrated in Figure 2.", + "bbox": [ + 89, + 148, + 483, + 404 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Keeping Diffusion Trajectories", + "text_level": 1, + "bbox": [ + 89, + 415, + 362, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since our approach leverages the trajectory of the diffusion model, it is crucial to prevent degradation of this trajectory during training. To achieve this, when fine-tuning the diffusion model to transition it into a feed-forward depth estimator, we simultaneously maintain the feed-forward step along with the preceding denoising training steps from the original diffusion model. While this trajectory was initially developed for image generation, directly applying it as-is does not facilitate an optimal transition to the depth domain. Thus, instead of predicting purely image-based latents, we modify the target latent to be a blend of image and depth representations.", + "bbox": [ + 89, + 436, + 483, + 619 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {b} _ {0} = \\gamma \\mathbf {x} _ {0} + (1 - \\gamma) \\mathbf {d} _ {0}, \\\\ \\mathbf {b} _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {b} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\boldsymbol {\\epsilon}, t \\in \\{1, \\dots , T \\}. \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 631, + 480, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\mathbf{b}_0$ represents the blended latent. $\\gamma$ controls the balance between the image and depth latents. In this process, we also use v-prediction re-parameterization approach [38] to define the training objective:", + "bbox": [ + 89, + 676, + 483, + 736 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {v} _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {\\epsilon} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\mathbf {b} _ {0}, \\\\ L _ {k} = \\left\\| \\mathbf {v} _ {t} - \\hat {\\boldsymbol {\\epsilon}} _ {\\theta} \\left(\\mathbf {b} _ {t}, t\\right) \\right\\| _ {2} ^ {2}, t \\in \\{1, \\dots , T \\}. \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 750, + 480, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Intuitively, this approach forces the diffusion model to preserve the shared features between the image generation task and the depth estimation task, which are captured in the blended training target. Therefore, it allows the diffusion model to adapt more naturally to depth estimation while retaining essential generative features. Consequently, during fine-tuning, the model maintains features that enhance the", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "accuracy and detail of depth predictions. This part is used exclusively during training. At inference time, our model functions as a fully deterministic framework.", + "bbox": [ + 511, + 90, + 905, + 136 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Learnable Filter Distillation", + "text_level": 1, + "bbox": [ + 511, + 148, + 764, + 162 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following previous methods, the above training process only uses synthetic data because it provides high-quality depth Ground Truth. However, this reliance limits both Marigold [21] and our approach, as SD-based MDE models trained solely on synthetic datasets often struggle to generalize well to in-the-wild data. Nevertheless, in this case, the predictions produced by the SD-based MDE model still retain the necessary details—precisely those details we expect to preserve in the final output. Therefore, enhancing the model's robustness essentially means improving the accuracy of the low-frequency components in the model's output on real image.", + "bbox": [ + 511, + 172, + 906, + 352 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by prior work [52], we observe that DINOv2 [31], trained on synthetic data, can generalize effectively to real-world images. However, its depth estimates often lack the necessary fine details—in other words, it can accurately predict the low-frequency depth components for massive real images but misses the high-frequency details. This characteristic aligns well with our needs, so we attempt to leverage the abundance of labels generated by the DINOv2 model to enhance our model's robustness. A straightforward method, as used in Depth Anything v2 [52], is to use a fintuned DINOv2-G depth model to generate pseudo labels, expanding the training set. However, since DINOv2's depth predictions lack fine detail, directly incorporating these pseudo labels on the feed-forward output $\\mathbf{d}_0$ risks ignoring the detailed features inherent to our model.", + "bbox": [ + 511, + 354, + 908, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this, we propose learning a filter, denoted as $F(d_0)$ that processes our output to remove high-frequency details (e.g., fine details). This filtering produces results at a detail level similar to that of DINOv2, allowing us to focus supervision solely on the less accurate low-frequency components. The filter $F$ is obtained through learning. In fact, we observed that the fine-tuned SD model itself already serves as an effective filter learner because the diffusion model is inherently designed to model the subtle differences across different time steps. Therefore, we directly apply an additional SD step on the feed-forward output $\\mathbf{d}_0$ to simulate this filter. To maintain consistency with the inputs and concepts used in the diffusion model, we represent this filter $F: t_0 \\rightarrow t_{-1}$ as the process that transforms the output from $t_0$ to $t_{-1}$ :", + "bbox": [ + 511, + 582, + 908, + 809 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {d} _ {- 1} = \\hat {\\mathbf {e}} _ {\\theta} \\left(\\mathbf {d} _ {0}, t = - 1\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 828, + 903, + 843 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "At this stage, we can use the labels predicted by DINOv2 to supervise $\\mathbf{d}_{-1}$ , allowing us to transfer DINOv2's robustness without interfering with the detailed features in $\\mathbf{d}_0$ . In", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6026", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ad53793f8cce5cd1662bd934afa7d02c492e3471a6e3070ae8af51887fcebe03.jpg", + "image_caption": [ + "Figure 3. Filter learning. We use a learnable filter to map our results to detail levels similar to DINOv2's, matching its outputs and thereby transferring DINOv2's generalization capabilities to our model without compromising our inherent details." + ], + "image_footnote": [], + "bbox": [ + 91, + 93, + 483, + 256 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "this process, we use real-world image data for $\\mathbf{x}_0$ . Figure 3 illustrates the effects of Eq. (4).", + "bbox": [ + 89, + 351, + 482, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Final Objective", + "text_level": 1, + "bbox": [ + 89, + 388, + 246, + 406 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For $\\mathbf{d}_0$ and $\\mathbf{d}_{-1}$ , we follow the method of MiDaS [34] using MAE loss $L_{\\mathrm{MAE}}$ and gradient matching loss $L_{\\mathrm{GM}}$ as depth loss. However, unlike the original approach, we apply these losses in the latent space of Stable Diffusion instead:", + "bbox": [ + 89, + 411, + 483, + 472 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {M A E}} \\left(\\mathbf {d}, \\mathbf {d} ^ {*}\\right) = \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\left| \\mathbf {d} _ {i} - \\mathbf {d} _ {i} ^ {*} \\right|, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 483, + 482, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{d}$ represents the ground truth latent, and $\\mathbf{d}^*$ is the model's predicted value. $M$ denotes the number of pixels in the depth latent.", + "bbox": [ + 89, + 527, + 483, + 574 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {G M}} \\left(\\mathbf {d}, \\mathbf {d} ^ {*}\\right) = \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\left(| \\nabla_ {x} R _ {i} | + | \\nabla_ {y} R _ {i} |\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 590, + 482, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $R_{i} = \\mathbf{d}_{i} - \\mathbf{d}_{i}^{*}$ . Hence, the final objective is expressed as a weighted sum of the losses $L_{\\mathrm{MAE}}$ , $L_{\\mathrm{GM}}$ , and $L_{k}$ :", + "bbox": [ + 89, + 638, + 482, + 671 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {\\text {f i n a l}} = \\sum_ {t \\in \\{- 1, 0 \\}} \\left(\\lambda_ {\\mathrm {M A E}} L _ {\\mathrm {M A E}} \\left(\\mathbf {d} _ {t}, \\mathbf {d} _ {t} ^ {*}\\right) \\right. \\\\ + \\lambda_ {\\mathrm {G M}} L _ {\\mathrm {G M}} \\left(\\mathbf {d} _ {t}, \\mathbf {d} _ {t} ^ {*}\\right)) + \\lambda_ {k} L _ {k}. \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 689, + 482, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\lambda_{\\mathrm{MAE}}$ $\\lambda_{\\mathrm{GM}}$ , and $\\lambda_{k}$ are the weighting factors for their respective loss term. $\\mathbf{d}_0^*$ represents the ground truth (GT) from the synthetic dataset, and $\\mathbf{d}_{-1}^{*}$ represents the pseudo label generated by DINOv2 for real images.", + "bbox": [ + 89, + 751, + 483, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 823, + 223, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 848, + 307, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During the training process, we preserve the diffusion trajectory while following the original DDPM noise sched", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "uler [20] using 1000 diffusion steps. To better leverage pre-trained models, we use the Depth Anything V2-Large model as the DINO v2 model for supervision. Our training dataset consists of two parts. For training at $t = 0$ and during trajectory retention, we follow previous approaches and use two synthetic datasets, Hypersim [36] and Virtual KITTI [7], which cover both indoor and outdoor scenes, with a total of 74K images. For training at $t = -1$ , we use real-world data from the LAION-Art dataset, a subset of LAION-5B [40] containing 8 million samples. However, we observed that training with only 0.2 million samples was sufficient. Synthetic data and real data each account for half of each batch. The parameters are set as follows: $\\gamma = 0.5$ , $\\lambda_{\\mathrm{MAE}} = 1$ , $\\lambda_{\\mathrm{GM}} = 0.5$ , $\\lambda_{k} = 0.2$ .", + "bbox": [ + 511, + 90, + 906, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Comparison", + "text_level": 1, + "bbox": [ + 511, + 316, + 645, + 332 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Zero-shot affine-invariant depth. For the evaluation of affine-invariant depth, we use the same datasets and evaluation protocol as Marigold. These datasets include NYUv2 [42], ScanNet [11], KITTI [16], ETH3D [39], and DIODE [44]. We compared FiftDepth with 14 methods that produce affine-invariant depth maps/disparities, all claiming zero-shot generalization capabilities. These include the earlier methods [12, 34, 35, 56-58], as well as the more recent ones [15, 17, 18, 21, 29, 49, 51, 52]. As shown in Table 1, FiftDepth achieves the best or state-of-the-art comparable results in most test scenarios. For visualization results, please refer to Figure 4. Our method not only accurately predicts the relative depth relationships but also excels in identifying and predicting depth for very fine objects. We also evaluate our method on the DA-2K introduced by Depth Anything v2. On this dataset, our method also performs comparably to Depth Anything v2.", + "bbox": [ + 511, + 339, + 906, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compared to methods like Depth Anything, which rely on massive training datasets, our model achieves comparable generalization to DAv2 while being trained on only a small amount of real data. To validate the generalization capability of our method, we present some test examples from special scenarios in Figure 5, including games, artworks, AI-generated content, and movies. Our method shows comparable generalization to DAv2 while preserving more details. In contrast, other methods' results are unsatisfactory in both generalization and detail preservation.", + "bbox": [ + 511, + 597, + 905, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Zero-shot boundaries. To further demonstrate the accuracy of our method in predicting fine structures, we also employ the Zero-shot Boundaries Metric introduced in the recent work Depth Pro [5] to evaluate boundary sharpness. Following Depth Pro, we compute the depth average boundary F1 score for datasets with ground truth and the boundary recall (R) for datasets with matting or segmentation annotations. The former datasets include Sintel [6], Spring [30], and iBims [22], while the latter include AM-2k [26], P3M-10k [25], and DIS-5k [33]. For details on the boundaries", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6027", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/99706824d0fa13b46374821df693e32d49b516cec9a11f7cc101d7f132978c97.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTraining DataNYUv2KITTIETH3DScanNetDIODE-FullDA-2K
AbsRel ↓δ1 ↑AbsRel ↓δ1 ↑AbsRel ↓δ1 ↑AbsRel ↓δ1 ↑AbsRel ↓δ1 ↑Acc (%)
DiverseDepth320K11.787.519.070.422.869.410.988.237.663.179.3
MiDaS2M11.188.523.663.018.475.212.184.633.271.580.6
LeReS354K9.091.614.978.417.177.79.191.727.176.681.1
Omnidata v212.2M7.494.514.983.516.677.87.593.633.974.276.8
HDN300K6.994.811.586.712.183.38.093.924.678.085.7
DPT1.4M9.890.310.090.17.894.68.293.418.275.883.2
Marigold74K*5.596.49.991.66.496.06.495.130.877.386.8
e2e-ft74K*5.296.69.691.96.495.95.896.230.277.983.6
DepthFM74K*6.595.68.393.47.895.96.894.924.574.185.8
GenPercept74K*5.696.013.084.27.095.66.296.130.777.685.1
Lotus-D59K*5.396.78.192.86.595.35.896.329.978.186.8
Lotus-G59K*5.496.68.587.76.296.16.096.029.478.586.2
GeoWizard280K*5.296.69.792.16.496.16.195.329.779.288.1
DepthAnything v1-L62.6M*4.398.17.694.712.788.24.298.027.775.988.5
DepthAnything v2-L62.6M*4.597.97.494.613.186.54.297.826.275.497.1
FiffDepth (Ours)274K*4.497.87.393.57.197.24.297.923.978.197.1
", + "bbox": [ + 91, + 88, + 883, + 366 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparison with other affine-invariant depth estimators on several zero-shot benchmarks. We use AbsRel (absolute relative error: $|d^{*} - d| / d$ ) and $\\delta_{1}$ (percentage of max $(d^{*} / d, d / d^{*}) < 1.25$ ). All metrics are reported as percentages; bold numbers are the best, underscored second best. Methods marked with an asterisk (*) utilize pre-trained models.", + "bbox": [ + 89, + 376, + 906, + 417 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b8e003d99145206d23b58668055b5a806445f0da1433bd367aba91c34e6f60f8.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison across different datasets. Our method is capable of predicting the depth of various fine objects, such as lampposts, railings, and chair legs." + ], + "image_footnote": [], + "bbox": [ + 96, + 426, + 908, + 851 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6028", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f06461048ea5d2948f3a0da1823c05df844e90024b3f8b1447475d944e66f98b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 87, + 344, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f02883b984eab1ab1cb4036f2059aa36d41cafbb6b7321c5a6a585f0566cd21b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 366, + 87, + 524, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d93e4784913ffcee796906c973888065ca5e137f5f8fbe87a26c8b911cded3ec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 550, + 88, + 763, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d37fb4dc68b4c71d09dd3dd09929f52b4286db9a2a81d69cd0538ac7417d0a99.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 787, + 88, + 903, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e68d29acd39c2a95b7584686916156b059ed8d0be4d09b2991edf0330ced5ace.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 184, + 341, + 281 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9ad3aabcad67d77ab6c73451084c127d247dd955ef0b3ddf5c5c0b41766f2073.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 186, + 524, + 280 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e55bd8ffe43077539fca55c09cda96734f9c49e5c1c106592d0e106038943c6b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 550, + 186, + 763, + 280 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/96f7b1e29b1b75219063e14fdb4e1f11bbd55f6e733bcac85d1b9398ac627601.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 787, + 186, + 903, + 280 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8a32d8e070a3cec2dda0c6d6bf92112422970d2e45c102ccc7b1831a14aeb023.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 284, + 341, + 378 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7cbfff5359cce5480dda847cc364eaf4374e459e6605e1f14a42a1e38deeaa9d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 284, + 524, + 378 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f15d998a2f674f41ad93b904533e334c6f8fa9905da0f9c431b9a8027bb75086.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 550, + 284, + 763, + 378 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/65b15ad3d14dcfdf37f169530e387220c0a0f4be1db9c162bc43b40135ded43b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 787, + 284, + 903, + 378 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5b0136e4a70814225ff9ad68e12d98fd6cdd47a2533a82d83106c1c9cf2ff34d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 382, + 341, + 477 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d23b97a3d5053b208c1695cff1d216aa44c759e877dae29b50cf0e70edc21e3a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 382, + 524, + 477 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b01e457706727a58340b4782b335d9894c48b283b65ef0613003ad06ba0e3861.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 550, + 382, + 763, + 477 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7dbdaec09dc5d19332391d91976db620033167d5201946c6543662e99d052c14.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 787, + 382, + 903, + 477 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7c6dd9c5e0d37fa753389b59a2429bf602a835309e8bad16363da928abb837da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 479, + 344, + 575 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/55e6dbca66a3508f40a4446e71f825e3c86c9546b39cffd8090a46ebafc322a0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 479, + 524, + 575 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/352cae2494fb7f7f5a3064623cb9ad8fad88c60768efc92c343aa5d31f1e356f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 550, + 479, + 763, + 575 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7e12cb90c72abd1f0ca518327f5992f2e1619ef7e28f77df30c38f6c9d24af84.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 787, + 479, + 903, + 575 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9ea1be9da8d3f3619bb32639cd42d42f504dc2ab8bb70cffcf3987ab8dbf565b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 578, + 341, + 674 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f5c295656393be322a38e937da6b92b3df19fa51d55f639e31d3fbc2433a1338.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 578, + 524, + 674 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/631c684044426f78ee5948c8e54c73fcf313555544667f9e23b799a068628b18.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 550, + 578, + 763, + 674 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5574e4f75cd99d2ddf79fe59fa44d4a1de17aafeca7aee5ae327a3704bbfff8f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 787, + 578, + 903, + 674 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a096fcd922912a6b0cd0fded11d29e19d162df8e2658f0286894bf77f89ebb48.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 676, + 341, + 773 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a953e849b95fd4a0cf76f04ebcaf4e625758c85498c036f63154c6f025797bbd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 676, + 524, + 773 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/14668697365c19fbe1a5240ad5f67ee20ad025640d6d803f360482e44da0d575.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 550, + 676, + 763, + 773 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/93d5e58e6b876b030d6748806502b26f564d44495ed94a535cb1372eaa3a0c61.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 787, + 676, + 903, + 773 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3e9a5d12703748d56e89df067705455069a798bf2fa74705a1df6891334216b2.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison on special scenarios. In the special scenarios of games, artworks, AI-generated content, and movies, our method demonstrates strong generalization capability and the ability to predict detailed depth." + ], + "image_footnote": [], + "bbox": [ + 99, + 776, + 344, + 871 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0cb53a9ded46b4334bd4a68fcd10f4e388dcb382db6c0f2e4e687f25ed60b8b0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 776, + 524, + 871 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1007fe4e7e5f20e0f07830c3f573be5c74c82b824042b410f2c549c10c02ff71.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 550, + 773, + 763, + 871 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dfb46957cde7b660ae50ea55cbc9c2e15d7168ccdab2475c6d7bde1d8bd0151e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 787, + 776, + 903, + 871 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6029", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/db957fbc803541c18c65c0ab7824d0e348c83233adf0f8b163ccff6ee28da947.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 93, + 88, + 225, + 159 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4408792880f35668c66f7a1773d948ceb3ccdb98235364b8f9feb83ce3b51363.jpg", + "image_caption": [ + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 93, + 159, + 227, + 231 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7cec304557161c7139f9dad19e333aa9d0b0889d7adf4e89baf4594d2f0c7293.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 227, + 88, + 362, + 159 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a55904f4b059b258a9ec9d50b06981f63f4f8fcb643adea9ced92af0113eb18a.jpg", + "image_caption": [ + "Alpha Matte" + ], + "image_footnote": [], + "bbox": [ + 227, + 159, + 362, + 231 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/85a2bf0fb2277c68cdf14851d8701583ed6dd48677df072472a1ee31d230754c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 88, + 496, + 159 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4f7f9728d41f27fc60491e57903b1c9948e658fa69a86047dcb9e3a382dcd469.jpg", + "image_caption": [ + "FifDepth", + "Figure 6. Boundary visualization comparison. These samples are from the AM-2k dataset." + ], + "image_footnote": [], + "bbox": [ + 364, + 159, + 496, + 231 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3c8e678e149d1839280e615bdb6afb36d945b59824e59343e53860876a6cb6db.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 88, + 633, + 159 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b23fe43b85fe0a0494b3d00e83e6d84f4a44129720dab601c190bcf02d3525b5.jpg", + "image_caption": [ + "Depth Pro" + ], + "image_footnote": [], + "bbox": [ + 500, + 159, + 633, + 231 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ed56d1b4aaace17d2a80e4834450a3833a9df91728d4161330364119eba2c64a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 88, + 767, + 159 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d820bf91a99cda0ebb8c845a6d48d36e17e9aa3eb44f650a5029d52a11b3fd55.jpg", + "image_caption": [ + "Depth Any. v2" + ], + "image_footnote": [], + "bbox": [ + 635, + 159, + 767, + 231 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e4403e627c42001ab1b89e07a197446a4e327d306ef317c293eceaf7311598ac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 769, + 88, + 903, + 159 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e0e519a1d3d2c85aaf2bc92d9a7274adf247accebe061d20beb792f37289fd62.jpg", + "image_caption": [ + "DepthFM" + ], + "image_footnote": [], + "bbox": [ + 769, + 159, + 903, + 231 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/26061047379f50cacf4fa4e1bc1cefa0c104a14f97a81db0c31de9ac5c7c95e6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSintel F1↑Spring F1↑iBims F1↑AM R↑P3M R↑DIS R↑
DepthAnything v20.2280.0560.1110.1070.1310.056
Depth Pro0.4090.0790.1760.1730.1680.077
FiffDepth (Ours)0.4230.0860.1890.1760.1790.091
", + "bbox": [ + 178, + 272, + 820, + 337 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/28b816a245926acf8c23a4ff461688dc1574c7d8fec326d75bf874374e56612b.jpg", + "table_caption": [ + "Table 2. Zero-shot boundary accuracy. We provide the F1 score for datasets containing ground-truth depth and boundary recall (R) for those with matting or segmentation labels." + ], + "table_footnote": [], + "table_body": "
MethodMarigoldMarigold (LCM)GeoWizardDepthFMDepthAnything v2-LDepth ProOurs
Time (s)1031.7190.390.0260.230.092
", + "bbox": [ + 114, + 378, + 883, + 416 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Running time comparison. We perform inference on ${100512} \\times {512}$ images using these methods and report the average time.", + "bbox": [ + 91, + 426, + 901, + 441 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/152617c8d3931740ac93ee795d9a59ecddd9329bd3f4bc2c6cca17915c4c4052.jpg", + "image_caption": [ + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 93, + 455, + 189, + 502 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3319f3671586c6464afb2e2cad2ef05aedfe65842b41d1c2497d96a7001f9c09.jpg", + "image_caption": [ + "w/o blend" + ], + "image_footnote": [], + "bbox": [ + 191, + 455, + 287, + 502 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2ade62655ef2de54b797386ae8476937bfba6c2503ec0df92a1b27166a12770a.jpg", + "image_caption": [ + "w/o $L_{k}$" + ], + "image_footnote": [], + "bbox": [ + 289, + 455, + 383, + 502 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0e97e459a08548dccd2934beb21bb88f8226d3f93d1731f8f7879a9bff50ae5f.jpg", + "image_caption": [ + "FiffDepth Output" + ], + "image_footnote": [], + "bbox": [ + 385, + 455, + 480, + 502 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cb318011e50e35e791049b038982c206117c2b031f64ab0749f606c0dca4ff75.jpg", + "image_caption": [ + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 93, + 516, + 189, + 569 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5a0211de8ca57da8628960965f017df04f6b749c95454e0fed8d50c0b4e7dc5d.jpg", + "image_caption": [ + "w/o DINO" + ], + "image_footnote": [], + "bbox": [ + 191, + 516, + 287, + 569 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/976d74704298fc43877737e786a6547ccc0a9c7fa278851b2626b0ce4c6c9e27.jpg", + "image_caption": [ + "DINO on $d_0$" + ], + "image_footnote": [], + "bbox": [ + 289, + 516, + 383, + 569 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0d3ffd45d5694eef30e5b07fee1d6b9e795531cd0f168a4b47e94c9b36a587cb.jpg", + "image_caption": [ + "FiffDepth Output", + "Figure 7. Ablation studies. The generalization capability and depth details of the method are affected when some essential components are missing." + ], + "image_footnote": [], + "bbox": [ + 385, + 516, + 480, + 570 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "metric and its computation, please refer to the Depth Pro paper for further details. Quantitative comparisons in Table 2 demonstrate that our method surpasses Depth Pro and other approaches in boundary prediction. Additionally, the visual results in Figure 6 further validate that our method predicts more accurate boundaries.", + "bbox": [ + 89, + 656, + 482, + 744 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Running time. Generative MDE models adopt the diffusion paradigm, and the resulting instability necessitates test-time assembly, leading to a lengthy inference time. In contrast, our feed-forward approach provides significant efficiency advantages. We evaluate the average inference time for a $512 \\times 512$ image on an NVIDIA Titan RTX GPU. As shown in Table 3, our method significantly outperforms other generative approaches in terms of efficiency and achieves performance comparable to DAv2. We tested these methods using their default settings.", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation studies. We conduct ablation studies to validate components of our method. Keeping the diffusion trajectory but predicting purely image latents affects the relative depth relationships between objects (1st row, 1st result in Fig. 7). Without keeping trajectory, some details are lost (1st row, 2nd result in Fig. 7). Omitting DINO supervision impacts the model's generalization ability (2nd row, 1st result in Fig. 7). Using DINO supervision at $d_{0}$ also reduces details (2nd row, 2nd result in Fig. 7).", + "bbox": [ + 511, + 458, + 906, + 594 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 632, + 633, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we transform diffusion models into stable, feed-forward depth estimators, achieving significant improvements in accuracy and efficiency over generative model-based methods. By combining the detail preservation of generative models with the robust generalization of FFN models like DINOv2, our hybrid approach bridges the synthetic-to-real gap, enhancing stability, predictability, and resolution in MDE for diverse real-world scenarios.", + "bbox": [ + 511, + 665, + 906, + 785 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. Q. Huang acknowledges the support of NSF-2047677, NSF-2413161, NSF-2504906, NSF-2515626, and GIFTs from Adobe and Google. This research has been supported by computing support on the Vista GPU Cluster through the Center for Generative AI (CGAI) and the Texas Advanced Computing Center (TACC) at UT Austin.", + "bbox": [ + 511, + 795, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6030", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Shubhra Aich, Jean Marie Uwabeza Vianney, Md Amirul Islam, and Mannat Kaur Bingbing Liu. Bidirectional attention network for monocular depth estimation. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 11746-11752. IEEE, 2021. 2", + "[2] Yunpeng Bai, Yanbo Fan, Xuan Wang, Yong Zhang, Jingxiang Sun, Chun Yuan, and Ying Shan. High-fidelity facial avatar reconstruction from monocular video with generative priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4541-4551, 2023. 1", + "[3] Yunpeng Bai, Xintao Wang, Yan-Pei Cao, Yixiao Ge, Chun Yuan, and Ying Shan. Dreamdiffusion: High-quality eegt-to-image generation with temporal masked signal modeling and clip alignment. In European Conference on Computer Vision, pages 472-488. Springer, 2024. 2", + "[4] Yoshua Bengio, Li Yao, Guillaume Alain, and Pascal Vincent. Generalized denoising auto-encoders as generative models. Advances in neural information processing systems, 26, 2013. 2", + "[5] Aleksei Bochkovskii, Amaël Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. arXiv preprint arXiv:2410.02073, 2024. 5", + "[6] Daniel J Butler, Jonas Wulff, Garrett B Stanley, and Michael J Black. A naturalistic open source movie for optical flow evaluation. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part VI 12, pages 611-625. Springer, 2012. 5", + "[7] Yohann Cabon, Naila Murray, and Martin Humenberger. Virtual katii 2. arXiv preprint arXiv:2001.10773, 2020. 5", + "[8] Weifeng Chen, Zhao Fu, Dawei Yang, and Jia Deng. Single-image depth perception in the wild. Advances in neural information processing systems, 29, 2016. 2", + "[9] Weifeng Chen, Shengyi Qian, David Fan, Noriyuki Kojima, Max Hamilton, and Jia Deng. Oasis: A large-scale dataset for single image 3d in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 679-688, 2020. 2", + "[10] Yuhua Chen, Cordelia Schmid, and Cristian Sminchisescu. Self-supervised learning with geometric constraints in monocular video: Connecting flow, depth, and camera. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7063-7072, 2019. 2", + "[11] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 5", + "[12] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multitask mid-level vision datasets from 3d scans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10786-10796, 2021. 2, 5" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. Advances in neural information processing systems, 27, 2014. 2", + "[14] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Batmanghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2002-2011, 2018. 2", + "[15] Xiao Fu, Wei Yin, Mu Hu, Kaixuan Wang, Yuexin Ma, Ping Tan, Shaojie Shen, Dahua Lin, and Xiaoxiao Long. Geowizard: Unleashing the diffusion priors for 3d geometry estimation from a single image. arXiv preprint arXiv:2403.12013, 2024. 2, 3, 5", + "[16] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 5", + "[17] Ming Gui, Johannes S Fischer, Ulrich Prestel, Pingchuan Ma, Dmytro Kotovenko, Olga Grebenkova, Stefan Andreas Baumann, Vincent Tao Hu, and Björn Ommer. Depthfm: Fast monocular depth estimation with flow matching. arXiv preprint arXiv:2403.13788, 2024. 5", + "[18] Jing He, Haodong Li, Wei Yin, Yixun Liang, Leheng Li, Kaiqiang Zhou, Hongbo Liu, Bingbing Liu, and YingCong Chen. Lotus: Diffusion-based visual foundation model for high-quality dense prediction. arXiv preprint arXiv:2409.18124, 2024. 3, 4, 5", + "[19] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2", + "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 5", + "[21] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 3, 4, 5", + "[22] Tobias Koch, Lukas Liebel, Friedrich Fraundorfer, and Marco Korner. Evaluation of cnn-based single-image depth estimation methods. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, pages 0–0, 2018. 5", + "[23] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1611-1621, 2021. 2", + "[24] Jin Han Lee, Myung-Kyu Han, Dong Wook Ko, and Il Hong Suh. From big to small: Multi-scale local planar guidance for monocular depth estimation. arXiv preprint arXiv:1907.10326, 2019. 2", + "[25] Jizhizi Li, Sihan Ma, Jing Zhang, and Dacheng Tao. Privacy-preserving portrait matting. In Proceedings of the 29th ACM" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6031", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "international conference on multimedia, pages 3501-3509, 2021. 5", + "[26] Jizhizi Li, Jing Zhang, Stephen J Maybank, and Dacheng Tao. Bridging composite and real: towards end-to-end deep image matting. International Journal of Computer Vision, 130(2):246-266, 2022. 5", + "[27] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2041-2050, 2018. 2", + "[28] Zhenyu Li, Zehui Chen, Xianming Liu, and Junjun Jiang. Depthformer: Exploiting long-range correlation and local information for accurate monocular depth estimation. Machine Intelligence Research, 20(6):837-854, 2023. 2", + "[29] Gonzalo Martin Garcia, Karim Abou Zeid, Christian Schmidt, Daan de Geus, Alexander Hermans, and Bastian Leibe. Fine-tuning image-conditional diffusion models is easier than you think. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2025. 5", + "[30] Lukas Mehl, Jenny Schmalfuss, Azin Jahedi, Yaroslava Nalivayko, and Andres Bruhn. Spring: A high-resolution high-detail dataset and benchmark for scene flow, optical flow and stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4981-4991, 2023. 5", + "[31] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 4", + "[32] Vaishakh Patil, Christos Sakaridis, Alexander Liniger, and Luc Van Gool. P3depth: Monocular depth estimation with a piecewise planarity prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1610-1621, 2022. 2", + "[33] Xuebin Qin, Hang Dai, Xiaobin Hu, Deng-Ping Fan, Ling Shao, and Luc Van Gool. Highly accurate dichotomous image segmentation. In European Conference on Computer Vision, pages 38-56. Springer, 2022. 5", + "[34] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 44(3):1623-1637, 2020. 1, 2, 5", + "[35] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021. 1, 2, 5", + "[36] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10912-10922, 2021. 5", + "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3", + "[38] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 4", + "[39] Thomas Schops, Johannes L Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3260-3269, 2017. 5", + "[40] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 5", + "[41] Jaidev Shriram, Alex Trevithick, Lingjie Liu, and Ravi Ramamoorthi. Realmdreamer: Text-driven 3d scene generation with inpainting and depth diffusion. arXiv preprint arXiv:2404.07199, 2024. 1", + "[42] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12, pages 746-760. Springer, 2012. 5", + "[43] Zachary Teed and Jia Deng. Deepv2d: Video to depth with differentiable structure from motion. arXiv preprint arXiv:1812.04605, 2018. 2", + "[44] Igor Vasiljevic, Nick Kolkin, Shanyi Zhang, Ruotian Luo, Haochen Wang, Falcon Z Dai, Andrea F Daniele, Mohammadreza Mostajabi, Steven Basart, Matthew R Walter, et al. Diode: A dense indoor and outdoor depth dataset. arXiv preprint arXiv:1908.00463, 2019. 5", + "[45] Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. In Proceedings of the 25th international conference on Machine learning, pages 1096-1103, 2008. 2", + "[46] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9065-9076, 2023. 1", + "[47] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudolidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8445-8453, 2019. 1", + "[48] Yiran Wang, Min Shi, Jiaqi Li, Zihao Huang, Zhiguo Cao, Jianming Zhang, Ke Xian, and Guosheng Lin. Neural video depth stabilizer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9466-9476, 2023. 2" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6032", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Guangkai Xu, Yongtao Ge, Mingyu Liu, Chengxiang Fan, Kangyang Xie, Zhiyue Zhao, Hao Chen, and Chunhua Shen. Diffusion models trained with large data are transferable visual models. arXiv preprint arXiv:2403.06090, 2024. 3, 4, 5", + "[50] Guanglei Yang, Hao Tang, Mingli Ding, Nicu Sebe, and Elisa Ricci. Transformer-based attention networks for continuous pixel-wise prediction. In Proceedings of the IEEE/CVF International Conference on Computer vision, pages 16269-16279, 2021. 2", + "[51] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 2, 5", + "[52] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024. 1, 2, 4, 5", + "[53] Xingyi Yang and Xinchao Wang. Diffusion model as representation learner. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18938-18949, 2023. 2, 3", + "[54] Rajeev Yasarla, Hong Cai, Jisoo Jeong, Yunxiao Shi, Risheek Garrepalli, and Fatih Porikli. Mamo: Leveraging memory and attention for monocular video depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8754-8764, 2023. 2", + "[55] Chongjie Ye, Lingteng Qiu, Xiaodong Gu, Qi Zuo, Yushuang Wu, Zilong Dong, Liefeng Bo, Yuliang Xiu, and Xiaoguang Han. Stablenormal: Reducing diffusion variance for stable and sharp normal. arXiv preprint arXiv:2406.16864, 2024. 3", + "[56] Wei Yin, Xinlong Wang, Chunhua Shen, Yifan Liu, Zhi Tian, Songcen Xu, Changming Sun, and Dou Renyin. Diversedepth: Affine-invariant depth prediction using diverse data. arXiv preprint arXiv:2002.00569, 2020. 2, 5", + "[57] Wei Yin, Jianming Zhang, Oliver Wang, Simon Niklaus, Long Mai, Simon Chen, and Chunhua Shen. Learning to recover 3d scene shape from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 204-213, 2021.", + "[58] Chi Zhang, Wei Yin, Billzb Wang, Gang Yu, Bin Fu, and Chunhua Shen. Hierarchical normalization for robust monocular depth estimation. Advances in Neural Information Processing Systems, 35:14128-14139, 2022. 5", + "[59] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence. Advances in Neural Information Processing Systems, 36, 2024. 2, 3", + "[60] Xiang Zhang, Bingxin Ke, Hayko Riemenschneider, Nando Metzger, Anton Obukhov, Markus Gross, Konrad Schindler, and Christopher Schroers. Betterdepth: Plug-and-play diffusion refiner for zero-shot monocular depth estimation. arXiv preprint arXiv:2407.17952, 2024. 3", + "[61] Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing text-to-image diffu" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "sion models for visual perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5729-5739, 2023. 2, 3", + "bbox": [ + 545, + 90, + 903, + 133 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "6033", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/FiffDepth_ Feed-forward Transformation of Diffusion-Based Generators for Detailed Depth Estimation/3c8dc7fb-98b2-4d77-857e-35d78d6eaaf6_model.json b/2025/FiffDepth_ Feed-forward Transformation of Diffusion-Based Generators for Detailed Depth Estimation/3c8dc7fb-98b2-4d77-857e-35d78d6eaaf6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..54ffe28310beb505e5568cd65075255ce52e9a9c --- /dev/null +++ b/2025/FiffDepth_ Feed-forward Transformation of Diffusion-Based Generators for Detailed Depth Estimation/3c8dc7fb-98b2-4d77-857e-35d78d6eaaf6_model.json @@ -0,0 +1,2752 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.118, + 0.13, + 0.882, + 0.177 + ], + "angle": 0, + "content": "FiffDepth: Feed-forward Transformation of Diffusion-Based Generators for Detailed Depth Estimation" + }, + { + "type": "text", + "bbox": [ + 0.365, + 0.205, + 0.631, + 0.223 + ], + "angle": 0, + "content": "Yunpeng Bai Qixing Huang" + }, + { + "type": "text", + "bbox": [ + 0.362, + 0.223, + 0.631, + 0.24 + ], + "angle": 0, + "content": "The University of Texas at Austin" + }, + { + "type": "image", + "bbox": [ + 0.132, + 0.291, + 0.306, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.132, + 0.37, + 0.305, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.132, + 0.47, + 0.305, + 0.548 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.179, + 0.551, + 0.26, + 0.565 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.292, + 0.492, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.369, + 0.492, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.47, + 0.49, + 0.548 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.372, + 0.55, + 0.438, + 0.565 + ], + "angle": 0, + "content": "FifDepth" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.292, + 0.678, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.369, + 0.678, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.47, + 0.672, + 0.547 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.55, + 0.654, + 0.565 + ], + "angle": 0, + "content": "Depth Anything v2" + }, + { + "type": "image", + "bbox": [ + 0.689, + 0.292, + 0.864, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.689, + 0.37, + 0.863, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.689, + 0.47, + 0.855, + 0.547 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.741, + 0.55, + 0.81, + 0.565 + ], + "angle": 0, + "content": "Depth Pro" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.57, + 0.908, + 0.599 + ], + "angle": 0, + "content": "Figure 1. Compared to other methods, our model achieves more accurate details and better generalization in depth estimation. The final row shows the point cloud generated from the estimated depth results, and the corresponding depth map can be referenced in Figure 5." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.618, + 0.327, + 0.634 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.661, + 0.486, + 0.904 + ], + "angle": 0, + "content": "Monocular Depth Estimation (MDE) is a fundamental 3D vision problem with numerous applications such as 3D scene reconstruction, autonomous navigation, and AI content creation. However, robust and generalizable MDE remains challenging due to limited real-world labeled data and distribution gaps between synthetic datasets and real data. Existing methods often struggle with real-world test data with low efficiency, reduced accuracy, and lack of detail. To address these issues, we propose an efficient MDE approach named FiftDepth. The key feature of FiftDepth is its use of diffusion priors. It transforms diffusion-based image generators into a feed-forward architecture for detailed depth estimation. FiftDepth preserves key generative features and integrates the strong generalization capabilities of models like DINOv2. Through benchmark evaluations, we demonstrate that FiftDepth achieves exceptional" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.619, + 0.909, + 0.695 + ], + "angle": 0, + "content": "accuracy, stability, and fine-grained detail, offering significant improvements in MDE performance against state-of-the-art MDE approaches. The paper's source code is available here: https://yunpeng1998.github.io/FiffDepth/" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.725, + 0.646, + 0.74 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.751, + 0.907, + 0.904 + ], + "angle": 0, + "content": "Monocular Depth Estimation (MDE) is a fundamental 3D vision problem with numerous applications in 3D scene reconstruction [46], autonomous navigation [47], and more recently in generating AI-based content [2, 41]. MDE has made significant progress in the era of deep learning [21, 34, 35, 52]. Neural networks trained from paired datasets of images and pixel depth exhibit encouraging results and often outperform non-deep learning based counterparts that build on monocular depth cues. Despite significant progress, fundamental challenges remain in efficiency," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6023" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "accuracy, and generalization on diverse in-the-wild data. This is because 1) real depth datasets are usually noisy, and 2) while synthetic data can be used, there exist domain gaps between synthetic datasets and diverse in-the-wild data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.155, + 0.483, + 0.352 + ], + "angle": 0, + "content": "Specifically, current MDE research relies primarily on synthetic data due to its high-quality annotations and controlled environments. However, the scale and variety of synthetic datasets remain insufficient for comprehensive training. To address this, synthetic-to-real transfer techniques and the utilization of pre-trained models have emerged as viable solutions. Among pre-trained models, generative networks [37] preserve intricate image details more effectively than feed-forward networks (FFNs) like DINOv2 [31], thus holding greater promise for dense prediction models. However, generative models, while detail-rich, often fall short in synthetic-to-real transfer due to their limited generalization capabilities outside the training domain." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.355, + 0.483, + 0.566 + ], + "angle": 0, + "content": "Previous studies [15, 21] have adopted pre-trained diffusion models, where the main idea is to directly finetuning pre-trained RGB image diffusion models into depth map diffusion generation models conditioned on images. However, this method may not be ideal, as dense prediction models require certainty over diversity. The introduction of any noise or uncertainty during the generation process by these methods is sub-optimal. In contrast, we observe that simply using the denoising diffusion module in a feedforward manner yields better and more stable results. This method capitalizes on extending the trajectories of image diffusion models into the depth domain, representing a significant advancement in both accuracy and efficiency for generative model-based depth estimation methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Specifically, we optimize diffusion trajectories for MDE tasks. To enable the diffusion model to better retain certain detailed generative features when fine-tuned into an MDE model, we preserve the original generative training trajectory while training the model for depth prediction, aiming to maintain the detailed features of the original generative model as much as possible. Furthermore, recognizing the limitations of fintuned diffusion models in maintaining robustness in diverse real-world images—with inaccuracies in predicted depth occurring mainly in the low-frequency components—we leverage the strengths of a DINOv2 [31] based model, which excels at predicting accurate low-frequency depth despite its reduced fine detail. To address this issue, we use the diffusion model itself to learn a filter that refines its inaccurate predictions, producing low-frequency outputs with a detail level similar to that of DINOv2's predictions, thereby matching DINOv2's results and optimizing the low-frequency component of our output. This approach allows us to incorporate a large amount of real image data for training without sacrificing detail preservation of generative models, while simultaneously leveraging DINOv2's strong generalization capabilities to enhance" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.889, + 0.107 + ], + "angle": 0, + "content": "the overall robustness and precision of our MDE models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.108, + 0.905, + 0.303 + ], + "angle": 0, + "content": "In summary, the contributions of our work are as follows: 1) We propose an improved approach for transforming generative models into dense prediction models, specifically for depth prediction tasks, by leveraging diffusion model trajectories in a more stable, feed-forward manner; 2) We introduce a novel distillation method that transfers the robust generalization capabilities of models like DINOv2 to diffusion backbones; 3) Our method demonstrates higher stability, accuracy, and efficiency in depth estimation compared to other approaches based on generative models; 4) Compared to other FFN models, our approach achieves more detailed prediction results, marking a significant advancement in the field of MDE." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.32, + 0.663, + 0.334 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.345, + 0.686, + 0.361 + ], + "angle": 0, + "content": "2.1. Depth Estimation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.368, + 0.905, + 0.565 + ], + "angle": 0, + "content": "Depth estimation has always been a widely researched topic, with numerous studies conducted in the past including single-image [1, 13, 14, 24, 28, 32, 50] and video depth estimation [10, 23, 43, 48, 54]. Early efforts in image depth estimation, such as DIW [8] and OASIS [9], focused on predicting relative (ordinal) depth. Subsequent approaches, including MegaDepth [27] and DiverseDepth [56], used extensive collections of photographs from the Internet to develop models that adapt to unseen data, while MiDaS [34] improved generalization by incorporating a diverse range of datasets during training. Recent advances, including DPT [35] and Omnidata [12], adopted transformer-based architectures to improve depth estimation performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.906, + 0.716 + ], + "angle": 0, + "content": "However, due to limitations in models and data, these methods exhibit very limited generalization in various open-world scenes. Recently, some efforts [15, 21, 51, 52] in image depth estimation have made open-world depth estimation feasible by leveraging the power of vast amounts of unlabeled image data and the capabilities of pre-trained generative models. Although these methods have achieved remarkable progress in the field of depth estimation, challenges such as low efficiency, limited generalization, and insufficient detail preservation remain." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.727, + 0.889, + 0.743 + ], + "angle": 0, + "content": "2.2. Diffusion Models as Representation Learner" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.905, + 0.901 + ], + "angle": 0, + "content": "The diffusion model [3, 37] training process strongly resembles that of denoising autoencoders (DAE) [4, 19, 45], as both are designed to recover clean images from noise-corrupted inputs. Recent studies have shown that semantic image representations learned from diffusion models can be effectively used for various downstream recognition tasks, such as correspondence [59], semantic segmentation [61], and keypoint detection [53]. Notably, the features extracted from diffusion models tend to preserve more intricate details, which has prompted the adoption of pre" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6024" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.088, + 0.411, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.089, + 0.904, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.404, + 0.908, + 0.463 + ], + "angle": 0, + "content": "Figure 2. Overview of the proposed method. To simplify the representation, all the images we used above correspond to the respective latents. We transform the pre-trained diffusion model into a feed-forward approach for depth prediction, using only the result at \\( t = 0 \\) as the output during inference. During training, at \\( t = 0 \\), we use synthetic data to ensure detailed results, while at \\( t = -1 \\), we leverage pseudo-labels generated by DINOv2 for supervision." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.487, + 0.486, + 0.699 + ], + "angle": 0, + "content": "trained diffusion models for dense prediction tasks. Recent studies, including Marigold [21] and GeoWizard [15], rely on the standard diffusion framework and pre-trained parameters to perform dense prediction tasks. Emerging approaches [18, 49, 55] attempt to bypass the stochastic phase of diffusion models by employing deterministic frameworks. However, these adaptations lack deeper exploration of the model's potential, often leading to suboptimal performance and the need for additional post-processing to refine results. Moreover, diffusion-based techniques generally exhibit limited generalization capabilities. For instance, BetterDepth [60] incorporates external depth priors as inputs but remains a stochastic framework and heavily depends on the quality of other models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.716, + 0.184, + 0.731 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.742, + 0.485, + 0.772 + ], + "angle": 0, + "content": "3.1. Overview: Feed-forward Transformation of Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Fundamentally, generative models construct mappings between a latent space and the ambient data space. These mappings align closely with the needs of depth estimation and other visual recognition tasks, where precise mappings from image data to the corresponding labels are essential. In these tasks, the scarcity of labeled data often limits the precision of the trained models. In the context of learning from a collection of unlabeled data instances, advanced gen" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.487, + 0.907, + 0.565 + ], + "angle": 0, + "content": "erative models, trained on massive datasets, are capable of learning robust mappings. They exhibit great promise for transferring knowledge to other visual prediction applications [53, 59, 61]. Our work builds on this approach, further exploring its application in depth estimation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.573, + 0.909, + 0.742 + ], + "angle": 0, + "content": "Specifically, we finetune a well-trained Stable Diffusion (SD) [37] model to construct our model. A diffusion process constructs multiple intermediate states by progressively adding noise to the data \\(\\mathbf{x}_0\\), defined as \\(\\mathbf{x}_t = \\sqrt{\\bar{\\alpha}_t}\\mathbf{x}_0 + \\sqrt{1 - \\bar{\\alpha}_t}\\epsilon\\), where \\(\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\), and \\(\\bar{\\alpha}_t \\coloneqq \\prod_{s=1}^{t} 1 - \\beta_s\\) with noise schedule \\(\\{\\beta_1, \\ldots, \\beta_T\\}\\). Then, diffusion generative model gradually learns the mapping between two distributions by denoising each step, denoted as \\(x_T \\to x_{T-1} \\to \\dots \\to x_0\\). The diffusion model is primarily a denoising model \\(\\epsilon_\\theta\\) that follows a loss function \\(\\mathcal{L}_{\\mathrm{simple}} = \\mathbb{E}_{x_0,\\epsilon,t} \\left[ \\| \\epsilon - \\epsilon_\\theta(x_t,t) \\|^2 \\right]\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.911, + 0.903 + ], + "angle": 0, + "content": "In depth estimation, the mapping from visual images to depth labels should ideally be deterministic. By leveraging a robust, pre-trained mapping within the generative model, there is no need to decompose the image-label mapping into multiple steps during training. Instead, we can extend the mapping trajectories of the existing diffusion model directly into the depth domain by adapting the learned diffusion process to act as a deterministic one-step feed-forward network. Since our MDE model is built upon the diffusion trajectory for its extension, we set the time step input as" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6025" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.3, + 0.106 + ], + "angle": 0, + "content": "\\(t = 0\\) in this feed-forward step." + }, + { + "type": "equation", + "bbox": [ + 0.216, + 0.12, + 0.483, + 0.136 + ], + "angle": 0, + "content": "\\[\n\\mathbf {d} _ {0} = \\hat {\\boldsymbol {\\epsilon}} _ {\\theta} \\left(\\mathbf {x} _ {0}, t = 0\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.149, + 0.484, + 0.405 + ], + "angle": 0, + "content": "Since we use diffusion model's parameters to construct our network, we also use \\(\\hat{\\epsilon}_{\\theta}\\) to represent our network model here. The input \\(\\mathbf{x}_0\\) is the latent representation of RGB image, and \\(\\mathbf{d}_0\\) is the latent representation of \"depth image,\" following Marigold's encoding approach. The depth latent \\(\\mathbf{d}_0\\) can be reconstructed into a depth map using the VAE of SD with negligible error. While there have been preliminary attempts [18, 49] to use similar approaches, they remain in a nascent stage and lack the precision, robustness, and richness in detail needed for effective depth estimation. In the following sections, we will delve into each step of our method, detailing how we adapt the diffusion model to enhance accuracy and detail in depth estimation tasks, such as the key technical contributions of preserving diffusion trajectories and improving synthetic-to-real robustness, as discussed in Section 3.2 and Section 3.3, respectively. The overall workflow of the method is illustrated in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.416, + 0.364, + 0.433 + ], + "angle": 0, + "content": "3.2. Keeping Diffusion Trajectories" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.438, + 0.484, + 0.62 + ], + "angle": 0, + "content": "Since our approach leverages the trajectory of the diffusion model, it is crucial to prevent degradation of this trajectory during training. To achieve this, when fine-tuning the diffusion model to transition it into a feed-forward depth estimator, we simultaneously maintain the feed-forward step along with the preceding denoising training steps from the original diffusion model. While this trajectory was initially developed for image generation, directly applying it as-is does not facilitate an optimal transition to the depth domain. Thus, instead of predicting purely image-based latents, we modify the target latent to be a blend of image and depth representations." + }, + { + "type": "equation", + "bbox": [ + 0.147, + 0.632, + 0.482, + 0.669 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {b} _ {0} = \\gamma \\mathbf {x} _ {0} + (1 - \\gamma) \\mathbf {d} _ {0}, \\\\ \\mathbf {b} _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {b} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\boldsymbol {\\epsilon}, t \\in \\{1, \\dots , T \\}. \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.677, + 0.484, + 0.737 + ], + "angle": 0, + "content": "Here, \\(\\mathbf{b}_0\\) represents the blended latent. \\(\\gamma\\) controls the balance between the image and depth latents. In this process, we also use v-prediction re-parameterization approach [38] to define the training objective:" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.75, + 0.482, + 0.789 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {v} _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {\\epsilon} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\mathbf {b} _ {0}, \\\\ L _ {k} = \\left\\| \\mathbf {v} _ {t} - \\hat {\\boldsymbol {\\epsilon}} _ {\\theta} \\left(\\mathbf {b} _ {t}, t\\right) \\right\\| _ {2} ^ {2}, t \\in \\{1, \\dots , T \\}. \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Intuitively, this approach forces the diffusion model to preserve the shared features between the image generation task and the depth estimation task, which are captured in the blended training target. Therefore, it allows the diffusion model to adapt more naturally to depth estimation while retaining essential generative features. Consequently, during fine-tuning, the model maintains features that enhance the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.137 + ], + "angle": 0, + "content": "accuracy and detail of depth predictions. This part is used exclusively during training. At inference time, our model functions as a fully deterministic framework." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.15, + 0.766, + 0.164 + ], + "angle": 0, + "content": "3.3. Learnable Filter Distillation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.173, + 0.907, + 0.353 + ], + "angle": 0, + "content": "Following previous methods, the above training process only uses synthetic data because it provides high-quality depth Ground Truth. However, this reliance limits both Marigold [21] and our approach, as SD-based MDE models trained solely on synthetic datasets often struggle to generalize well to in-the-wild data. Nevertheless, in this case, the predictions produced by the SD-based MDE model still retain the necessary details—precisely those details we expect to preserve in the final output. Therefore, enhancing the model's robustness essentially means improving the accuracy of the low-frequency components in the model's output on real image." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.355, + 0.909, + 0.581 + ], + "angle": 0, + "content": "Inspired by prior work [52], we observe that DINOv2 [31], trained on synthetic data, can generalize effectively to real-world images. However, its depth estimates often lack the necessary fine details—in other words, it can accurately predict the low-frequency depth components for massive real images but misses the high-frequency details. This characteristic aligns well with our needs, so we attempt to leverage the abundance of labels generated by the DINOv2 model to enhance our model's robustness. A straightforward method, as used in Depth Anything v2 [52], is to use a fintuned DINOv2-G depth model to generate pseudo labels, expanding the training set. However, since DINOv2's depth predictions lack fine detail, directly incorporating these pseudo labels on the feed-forward output \\(\\mathbf{d}_0\\) risks ignoring the detailed features inherent to our model." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.583, + 0.909, + 0.81 + ], + "angle": 0, + "content": "To address this, we propose learning a filter, denoted as \\( F(d_0) \\) that processes our output to remove high-frequency details (e.g., fine details). This filtering produces results at a detail level similar to that of DINOv2, allowing us to focus supervision solely on the less accurate low-frequency components. The filter \\( F \\) is obtained through learning. In fact, we observed that the fine-tuned SD model itself already serves as an effective filter learner because the diffusion model is inherently designed to model the subtle differences across different time steps. Therefore, we directly apply an additional SD step on the feed-forward output \\( \\mathbf{d}_0 \\) to simulate this filter. To maintain consistency with the inputs and concepts used in the diffusion model, we represent this filter \\( F: t_0 \\rightarrow t_{-1} \\) as the process that transforms the output from \\( t_0 \\) to \\( t_{-1} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.829, + 0.905, + 0.844 + ], + "angle": 0, + "content": "\\[\n\\mathbf {d} _ {- 1} = \\hat {\\mathbf {e}} _ {\\theta} \\left(\\mathbf {d} _ {0}, t = - 1\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.901 + ], + "angle": 0, + "content": "At this stage, we can use the labels predicted by DINOv2 to supervise \\(\\mathbf{d}_{-1}\\), allowing us to transfer DINOv2's robustness without interfering with the detailed features in \\(\\mathbf{d}_0\\). In" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6026" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.094, + 0.485, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.273, + 0.485, + 0.332 + ], + "angle": 0, + "content": "Figure 3. Filter learning. We use a learnable filter to map our results to detail levels similar to DINOv2's, matching its outputs and thereby transferring DINOv2's generalization capabilities to our model without compromising our inherent details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.352, + 0.483, + 0.385 + ], + "angle": 0, + "content": "this process, we use real-world image data for \\(\\mathbf{x}_0\\). Figure 3 illustrates the effects of Eq. (4)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.39, + 0.248, + 0.407 + ], + "angle": 0, + "content": "3.4. Final Objective" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.412, + 0.484, + 0.473 + ], + "angle": 0, + "content": "For \\(\\mathbf{d}_0\\) and \\(\\mathbf{d}_{-1}\\), we follow the method of MiDaS [34] using MAE loss \\(L_{\\mathrm{MAE}}\\) and gradient matching loss \\(L_{\\mathrm{GM}}\\) as depth loss. However, unlike the original approach, we apply these losses in the latent space of Stable Diffusion instead:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.484, + 0.483, + 0.525 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {M A E}} \\left(\\mathbf {d}, \\mathbf {d} ^ {*}\\right) = \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\left| \\mathbf {d} _ {i} - \\mathbf {d} _ {i} ^ {*} \\right|, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.528, + 0.484, + 0.575 + ], + "angle": 0, + "content": "where \\(\\mathbf{d}\\) represents the ground truth latent, and \\(\\mathbf{d}^*\\) is the model's predicted value. \\(M\\) denotes the number of pixels in the depth latent." + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.592, + 0.483, + 0.635 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {G M}} \\left(\\mathbf {d}, \\mathbf {d} ^ {*}\\right) = \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\left(| \\nabla_ {x} R _ {i} | + | \\nabla_ {y} R _ {i} |\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.64, + 0.483, + 0.672 + ], + "angle": 0, + "content": "where \\(R_{i} = \\mathbf{d}_{i} - \\mathbf{d}_{i}^{*}\\). Hence, the final objective is expressed as a weighted sum of the losses \\(L_{\\mathrm{MAE}}\\), \\(L_{\\mathrm{GM}}\\), and \\(L_{k}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.118, + 0.69, + 0.483, + 0.745 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {\\text {f i n a l}} = \\sum_ {t \\in \\{- 1, 0 \\}} \\left(\\lambda_ {\\mathrm {M A E}} L _ {\\mathrm {M A E}} \\left(\\mathbf {d} _ {t}, \\mathbf {d} _ {t} ^ {*}\\right) \\right. \\\\ + \\lambda_ {\\mathrm {G M}} L _ {\\mathrm {G M}} \\left(\\mathbf {d} _ {t}, \\mathbf {d} _ {t} ^ {*}\\right)) + \\lambda_ {k} L _ {k}. \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.752, + 0.484, + 0.813 + ], + "angle": 0, + "content": "\\(\\lambda_{\\mathrm{MAE}}\\) \\(\\lambda_{\\mathrm{GM}}\\), and \\(\\lambda_{k}\\) are the weighting factors for their respective loss term. \\(\\mathbf{d}_0^*\\) represents the ground truth (GT) from the synthetic dataset, and \\(\\mathbf{d}_{-1}^{*}\\) represents the pseudo label generated by DINOv2 for real images." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.824, + 0.224, + 0.841 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.308, + 0.866 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "During the training process, we preserve the diffusion trajectory while following the original DDPM noise sched" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.304 + ], + "angle": 0, + "content": "uler [20] using 1000 diffusion steps. To better leverage pre-trained models, we use the Depth Anything V2-Large model as the DINO v2 model for supervision. Our training dataset consists of two parts. For training at \\( t = 0 \\) and during trajectory retention, we follow previous approaches and use two synthetic datasets, Hypersim [36] and Virtual KITTI [7], which cover both indoor and outdoor scenes, with a total of 74K images. For training at \\( t = -1 \\), we use real-world data from the LAION-Art dataset, a subset of LAION-5B [40] containing 8 million samples. However, we observed that training with only 0.2 million samples was sufficient. Synthetic data and real data each account for half of each batch. The parameters are set as follows: \\( \\gamma = 0.5 \\), \\( \\lambda_{\\mathrm{MAE}} = 1 \\), \\( \\lambda_{\\mathrm{GM}} = 0.5 \\), \\( \\lambda_{k} = 0.2 \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.317, + 0.646, + 0.333 + ], + "angle": 0, + "content": "4.2. Comparison" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.34, + 0.907, + 0.597 + ], + "angle": 0, + "content": "Zero-shot affine-invariant depth. For the evaluation of affine-invariant depth, we use the same datasets and evaluation protocol as Marigold. These datasets include NYUv2 [42], ScanNet [11], KITTI [16], ETH3D [39], and DIODE [44]. We compared FiftDepth with 14 methods that produce affine-invariant depth maps/disparities, all claiming zero-shot generalization capabilities. These include the earlier methods [12, 34, 35, 56-58], as well as the more recent ones [15, 17, 18, 21, 29, 49, 51, 52]. As shown in Table 1, FiftDepth achieves the best or state-of-the-art comparable results in most test scenarios. For visualization results, please refer to Figure 4. Our method not only accurately predicts the relative depth relationships but also excels in identifying and predicting depth for very fine objects. We also evaluate our method on the DA-2K introduced by Depth Anything v2. On this dataset, our method also performs comparably to Depth Anything v2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.906, + 0.749 + ], + "angle": 0, + "content": "Compared to methods like Depth Anything, which rely on massive training datasets, our model achieves comparable generalization to DAv2 while being trained on only a small amount of real data. To validate the generalization capability of our method, we present some test examples from special scenarios in Figure 5, including games, artworks, AI-generated content, and movies. Our method shows comparable generalization to DAv2 while preserving more details. In contrast, other methods' results are unsatisfactory in both generalization and detail preservation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Zero-shot boundaries. To further demonstrate the accuracy of our method in predicting fine structures, we also employ the Zero-shot Boundaries Metric introduced in the recent work Depth Pro [5] to evaluate boundary sharpness. Following Depth Pro, we compute the depth average boundary F1 score for datasets with ground truth and the boundary recall (R) for datasets with matting or segmentation annotations. The former datasets include Sintel [6], Spring [30], and iBims [22], while the latter include AM-2k [26], P3M-10k [25], and DIS-5k [33]. For details on the boundaries" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6027" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.885, + 0.367 + ], + "angle": 0, + "content": "
MethodTraining DataNYUv2KITTIETH3DScanNetDIODE-FullDA-2K
AbsRel ↓δ1 ↑AbsRel ↓δ1 ↑AbsRel ↓δ1 ↑AbsRel ↓δ1 ↑AbsRel ↓δ1 ↑Acc (%)
DiverseDepth320K11.787.519.070.422.869.410.988.237.663.179.3
MiDaS2M11.188.523.663.018.475.212.184.633.271.580.6
LeReS354K9.091.614.978.417.177.79.191.727.176.681.1
Omnidata v212.2M7.494.514.983.516.677.87.593.633.974.276.8
HDN300K6.994.811.586.712.183.38.093.924.678.085.7
DPT1.4M9.890.310.090.17.894.68.293.418.275.883.2
Marigold74K*5.596.49.991.66.496.06.495.130.877.386.8
e2e-ft74K*5.296.69.691.96.495.95.896.230.277.983.6
DepthFM74K*6.595.68.393.47.895.96.894.924.574.185.8
GenPercept74K*5.696.013.084.27.095.66.296.130.777.685.1
Lotus-D59K*5.396.78.192.86.595.35.896.329.978.186.8
Lotus-G59K*5.496.68.587.76.296.16.096.029.478.586.2
GeoWizard280K*5.296.69.792.16.496.16.195.329.779.288.1
DepthAnything v1-L62.6M*4.398.17.694.712.788.24.298.027.775.988.5
DepthAnything v2-L62.6M*4.597.97.494.613.186.54.297.826.275.497.1
FiffDepth (Ours)274K*4.497.87.393.57.197.24.297.923.978.197.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.377, + 0.907, + 0.419 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison with other affine-invariant depth estimators on several zero-shot benchmarks. We use AbsRel (absolute relative error: \\( |d^{*} - d| / d \\)) and \\( \\delta_{1} \\) (percentage of max \\( (d^{*} / d, d / d^{*}) < 1.25 \\)). All metrics are reported as percentages; bold numbers are the best, underscored second best. Methods marked with an asterisk (*) utilize pre-trained models." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.428, + 0.909, + 0.852 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.868, + 0.907, + 0.896 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison across different datasets. Our method is capable of predicting the depth of various fine objects, such as lampposts, railings, and chair legs." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6028" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.088, + 0.345, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.088, + 0.526, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.551, + 0.089, + 0.764, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.789, + 0.089, + 0.905, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.185, + 0.343, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.187, + 0.526, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.187, + 0.764, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.789, + 0.187, + 0.905, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.285, + 0.343, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.285, + 0.526, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.285, + 0.764, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.789, + 0.285, + 0.905, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.383, + 0.343, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.383, + 0.526, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.383, + 0.764, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.789, + 0.383, + 0.905, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.481, + 0.345, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.481, + 0.526, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.481, + 0.764, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.789, + 0.481, + 0.905, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.579, + 0.343, + 0.675 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.579, + 0.526, + 0.675 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.579, + 0.764, + 0.675 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.789, + 0.579, + 0.905, + 0.675 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.678, + 0.343, + 0.774 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.678, + 0.526, + 0.774 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.678, + 0.764, + 0.774 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.789, + 0.678, + 0.905, + 0.774 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.777, + 0.345, + 0.872 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.777, + 0.526, + 0.872 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.775, + 0.764, + 0.872 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.789, + 0.777, + 0.905, + 0.872 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.882, + 0.905, + 0.91 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison on special scenarios. In the special scenarios of games, artworks, AI-generated content, and movies, our method demonstrates strong generalization capability and the ability to predict detailed depth." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6029" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.227, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.16, + 0.228, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.122, + 0.235, + 0.2, + 0.248 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.089, + 0.363, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.16, + 0.363, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.256, + 0.234, + 0.337, + 0.248 + ], + "angle": 0, + "content": "Alpha Matte" + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.089, + 0.498, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.16, + 0.498, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.4, + 0.234, + 0.463, + 0.248 + ], + "angle": 0, + "content": "FifDepth" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.089, + 0.634, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.16, + 0.634, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.535, + 0.234, + 0.6, + 0.248 + ], + "angle": 0, + "content": "Depth Pro" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.089, + 0.769, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.16, + 0.769, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.656, + 0.234, + 0.748, + 0.248 + ], + "angle": 0, + "content": "Depth Any. v2" + }, + { + "type": "image", + "bbox": [ + 0.77, + 0.089, + 0.905, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.77, + 0.16, + 0.905, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.807, + 0.234, + 0.869, + 0.248 + ], + "angle": 0, + "content": "DepthFM" + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.255, + 0.773, + 0.269 + ], + "angle": 0, + "content": "Figure 6. Boundary visualization comparison. These samples are from the AM-2k dataset." + }, + { + "type": "table", + "bbox": [ + 0.179, + 0.273, + 0.821, + 0.338 + ], + "angle": 0, + "content": "
MethodSintel F1↑Spring F1↑iBims F1↑AM R↑P3M R↑DIS R↑
DepthAnything v20.2280.0560.1110.1070.1310.056
Depth Pro0.4090.0790.1760.1730.1680.077
FiffDepth (Ours)0.4230.0860.1890.1760.1790.091
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.348, + 0.907, + 0.378 + ], + "angle": 0, + "content": "Table 2. Zero-shot boundary accuracy. We provide the F1 score for datasets containing ground-truth depth and boundary recall (R) for those with matting or segmentation labels." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.38, + 0.884, + 0.417 + ], + "angle": 0, + "content": "
MethodMarigoldMarigold (LCM)GeoWizardDepthFMDepthAnything v2-LDepth ProOurs
Time (s)1031.7190.390.0260.230.092
" + }, + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.427, + 0.903, + 0.442 + ], + "angle": 0, + "content": "Table 3. Running time comparison. We perform inference on \\( {100512} \\times {512} \\) images using these methods and report the average time." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.456, + 0.19, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.504, + 0.171, + 0.515 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.456, + 0.288, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.504, + 0.263, + 0.514 + ], + "angle": 0, + "content": "w/o blend" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.456, + 0.385, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.319, + 0.504, + 0.352, + 0.515 + ], + "angle": 0, + "content": "w/o \\(L_{k}\\)" + }, + { + "type": "image", + "bbox": [ + 0.386, + 0.456, + 0.482, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.393, + 0.504, + 0.474, + 0.515 + ], + "angle": 0, + "content": "FiffDepth Output" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.517, + 0.19, + 0.57 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.572, + 0.171, + 0.583 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.517, + 0.288, + 0.57 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.572, + 0.264, + 0.583 + ], + "angle": 0, + "content": "w/o DINO" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.517, + 0.385, + 0.57 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.307, + 0.572, + 0.366, + 0.583 + ], + "angle": 0, + "content": "DINO on \\(d_0\\)" + }, + { + "type": "image", + "bbox": [ + 0.386, + 0.517, + 0.482, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.393, + 0.572, + 0.474, + 0.583 + ], + "angle": 0, + "content": "FiffDepth Output" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.6, + 0.483, + 0.641 + ], + "angle": 0, + "content": "Figure 7. Ablation studies. The generalization capability and depth details of the method are affected when some essential components are missing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.657, + 0.483, + 0.746 + ], + "angle": 0, + "content": "metric and its computation, please refer to the Depth Pro paper for further details. Quantitative comparisons in Table 2 demonstrate that our method surpasses Depth Pro and other approaches in boundary prediction. Additionally, the visual results in Figure 6 further validate that our method predicts more accurate boundaries." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Running time. Generative MDE models adopt the diffusion paradigm, and the resulting instability necessitates test-time assembly, leading to a lengthy inference time. In contrast, our feed-forward approach provides significant efficiency advantages. We evaluate the average inference time for a \\(512 \\times 512\\) image on an NVIDIA Titan RTX GPU. As shown in Table 3, our method significantly outperforms other generative approaches in terms of efficiency and achieves performance comparable to DAv2. We tested these methods using their default settings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.459, + 0.907, + 0.595 + ], + "angle": 0, + "content": "Ablation studies. We conduct ablation studies to validate components of our method. Keeping the diffusion trajectory but predicting purely image latents affects the relative depth relationships between objects (1st row, 1st result in Fig. 7). Without keeping trajectory, some details are lost (1st row, 2nd result in Fig. 7). Omitting DINO supervision impacts the model's generalization ability (2nd row, 1st result in Fig. 7). Using DINO supervision at \\( d_{0} \\) also reduces details (2nd row, 2nd result in Fig. 7)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.633, + 0.634, + 0.648 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.666, + 0.907, + 0.786 + ], + "angle": 0, + "content": "In this work, we transform diffusion models into stable, feed-forward depth estimators, achieving significant improvements in accuracy and efficiency over generative model-based methods. By combining the detail preservation of generative models with the robust generalization of FFN models like DINOv2, our hybrid approach bridges the synthetic-to-real gap, enhancing stability, predictability, and resolution in MDE for diverse real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. Q. Huang acknowledges the support of NSF-2047677, NSF-2413161, NSF-2504906, NSF-2515626, and GIFTs from Adobe and Google. This research has been supported by computing support on the Vista GPU Cluster through the Center for Generative AI (CGAI) and the Texas Advanced Computing Center (TACC) at UT Austin." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6030" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Shubhra Aich, Jean Marie Uwabeza Vianney, Md Amirul Islam, and Mannat Kaur Bingbing Liu. Bidirectional attention network for monocular depth estimation. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 11746-11752. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.186, + 0.484, + 0.268 + ], + "angle": 0, + "content": "[2] Yunpeng Bai, Yanbo Fan, Xuan Wang, Yong Zhang, Jingxiang Sun, Chun Yuan, and Ying Shan. High-fidelity facial avatar reconstruction from monocular video with generative priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4541-4551, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.271, + 0.483, + 0.338 + ], + "angle": 0, + "content": "[3] Yunpeng Bai, Xintao Wang, Yan-Pei Cao, Yixiao Ge, Chun Yuan, and Ying Shan. Dreamdiffusion: High-quality eegt-to-image generation with temporal masked signal modeling and clip alignment. In European Conference on Computer Vision, pages 472-488. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.341, + 0.483, + 0.394 + ], + "angle": 0, + "content": "[4] Yoshua Bengio, Li Yao, Guillaume Alain, and Pascal Vincent. Generalized denoising auto-encoders as generative models. Advances in neural information processing systems, 26, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.397, + 0.482, + 0.464 + ], + "angle": 0, + "content": "[5] Aleksei Bochkovskii, Amaël Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. arXiv preprint arXiv:2410.02073, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.467, + 0.482, + 0.548 + ], + "angle": 0, + "content": "[6] Daniel J Butler, Jonas Wulff, Garrett B Stanley, and Michael J Black. A naturalistic open source movie for optical flow evaluation. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part VI 12, pages 611-625. Springer, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.482, + 0.577 + ], + "angle": 0, + "content": "[7] Yohann Cabon, Naila Murray, and Martin Humenberger. Virtual katii 2. arXiv preprint arXiv:2001.10773, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.58, + 0.482, + 0.62 + ], + "angle": 0, + "content": "[8] Weifeng Chen, Zhao Fu, Dawei Yang, and Jia Deng. Single-image depth perception in the wild. Advances in neural information processing systems, 29, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.622, + 0.483, + 0.69 + ], + "angle": 0, + "content": "[9] Weifeng Chen, Shengyi Qian, David Fan, Noriyuki Kojima, Max Hamilton, and Jia Deng. Oasis: A large-scale dataset for single image 3d in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 679-688, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.692, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[10] Yuhua Chen, Cordelia Schmid, and Cristian Sminchisescu. Self-supervised learning with geometric constraints in monocular video: Connecting flow, depth, and camera. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7063-7072, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.763, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[11] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.833, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[12] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multitask mid-level vision datasets from 3d scans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10786-10796, 2021. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[13] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. Advances in neural information processing systems, 27, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[14] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Batmanghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2002-2011, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.221, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[15] Xiao Fu, Wei Yin, Mu Hu, Kaixuan Wang, Yuexin Ma, Ping Tan, Shaojie Shen, Dahua Lin, and Xiaoxiao Long. Geowizard: Unleashing the diffusion priors for 3d geometry estimation from a single image. arXiv preprint arXiv:2403.12013, 2024. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.293, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[16] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.349, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[17] Ming Gui, Johannes S Fischer, Ulrich Prestel, Pingchuan Ma, Dmytro Kotovenko, Olga Grebenkova, Stefan Andreas Baumann, Vincent Tao Hu, and Björn Ommer. Depthfm: Fast monocular depth estimation with flow matching. arXiv preprint arXiv:2403.13788, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.42, + 0.905, + 0.487 + ], + "angle": 0, + "content": "[18] Jing He, Haodong Li, Wei Yin, Yixun Liang, Leheng Li, Kaiqiang Zhou, Hongbo Liu, Bingbing Liu, and YingCong Chen. Lotus: Diffusion-based visual foundation model for high-quality dense prediction. arXiv preprint arXiv:2409.18124, 2024. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.49, + 0.905, + 0.558 + ], + "angle": 0, + "content": "[19] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.562, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.605, + 0.905, + 0.685 + ], + "angle": 0, + "content": "[21] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.689, + 0.905, + 0.756 + ], + "angle": 0, + "content": "[22] Tobias Koch, Lukas Liebel, Friedrich Fraundorfer, and Marco Korner. Evaluation of cnn-based single-image depth estimation methods. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, pages 0–0, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.76, + 0.907, + 0.814 + ], + "angle": 0, + "content": "[23] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1611-1621, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.817, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[24] Jin Han Lee, Myung-Kyu Han, Dong Wook Ko, and Il Hong Suh. From big to small: Multi-scale local planar guidance for monocular depth estimation. arXiv preprint arXiv:1907.10326, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Jizhizi Li, Sihan Ma, Jing Zhang, and Dacheng Tao. Privacy-preserving portrait matting. In Proceedings of the 29th ACM" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "6031" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "international conference on multimedia, pages 3501-3509, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[26] Jizhizi Li, Jing Zhang, Stephen J Maybank, and Dacheng Tao. Bridging composite and real: towards end-to-end deep image matting. International Journal of Computer Vision, 130(2):246-266, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.234 + ], + "angle": 0, + "content": "[27] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2041-2050, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.236, + 0.483, + 0.291 + ], + "angle": 0, + "content": "[28] Zhenyu Li, Zehui Chen, Xianming Liu, and Junjun Jiang. Depthformer: Exploiting long-range correlation and local information for accurate monocular depth estimation. Machine Intelligence Research, 20(6):837-854, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.293, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[29] Gonzalo Martin Garcia, Karim Abou Zeid, Christian Schmidt, Daan de Geus, Alexander Hermans, and Bastian Leibe. Fine-tuning image-conditional diffusion models is easier than you think. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2025. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[30] Lukas Mehl, Jenny Schmalfuss, Azin Jahedi, Yaroslava Nalivayko, and Andres Bruhn. Spring: A high-resolution high-detail dataset and benchmark for scene flow, optical flow and stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4981-4991, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.531 + ], + "angle": 0, + "content": "[31] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.483, + 0.602 + ], + "angle": 0, + "content": "[32] Vaishakh Patil, Christos Sakaridis, Alexander Liniger, and Luc Van Gool. P3depth: Monocular depth estimation with a piecewise planarity prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1610-1621, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.659 + ], + "angle": 0, + "content": "[33] Xuebin Qin, Hang Dai, Xiaobin Hu, Deng-Ping Fan, Ling Shao, and Luc Van Gool. Highly accurate dichotomous image segmentation. In European Conference on Computer Vision, pages 38-56. Springer, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.73 + ], + "angle": 0, + "content": "[34] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 44(3):1623-1637, 2020. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.483, + 0.787 + ], + "angle": 0, + "content": "[35] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[36] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10912-10922, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.135 + ], + "angle": 0, + "content": "synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.906, + 0.177 + ], + "angle": 0, + "content": "[38] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.906, + 0.262 + ], + "angle": 0, + "content": "[39] Thomas Schops, Johannes L Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3260-3269, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.906, + 0.347 + ], + "angle": 0, + "content": "[40] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.906, + 0.403 + ], + "angle": 0, + "content": "[41] Jaidev Shriram, Alex Trevithick, Lingjie Liu, and Ravi Ramamoorthi. Realmdreamer: Text-driven 3d scene generation with inpainting and depth diffusion. arXiv preprint arXiv:2404.07199, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.906, + 0.489 + ], + "angle": 0, + "content": "[42] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12, pages 746-760. Springer, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.906, + 0.532 + ], + "angle": 0, + "content": "[43] Zachary Teed and Jia Deng. Deepv2d: Video to depth with differentiable structure from motion. arXiv preprint arXiv:1812.04605, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.534, + 0.906, + 0.603 + ], + "angle": 0, + "content": "[44] Igor Vasiljevic, Nick Kolkin, Shanyi Zhang, Ruotian Luo, Haochen Wang, Falcon Z Dai, Andrea F Daniele, Mohammadreza Mostajabi, Steven Basart, Matthew R Walter, et al. Diode: A dense indoor and outdoor depth dataset. arXiv preprint arXiv:1908.00463, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.605, + 0.906, + 0.673 + ], + "angle": 0, + "content": "[45] Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. In Proceedings of the 25th international conference on Machine learning, pages 1096-1103, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.676, + 0.906, + 0.744 + ], + "angle": 0, + "content": "[46] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9065-9076, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.747, + 0.907, + 0.83 + ], + "angle": 0, + "content": "[47] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudolidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8445-8453, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.906, + 0.899 + ], + "angle": 0, + "content": "[48] Yiran Wang, Min Shi, Jiaqi Li, Zihao Huang, Zhiguo Cao, Jianming Zhang, Ke Xian, and Guosheng Lin. Neural video depth stabilizer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9466-9476, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6032" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.16 + ], + "angle": 0, + "content": "[49] Guangkai Xu, Yongtao Ge, Mingyu Liu, Chengxiang Fan, Kangyang Xie, Zhiyue Zhao, Hao Chen, and Chunhua Shen. Diffusion models trained with large data are transferable visual models. arXiv preprint arXiv:2403.06090, 2024. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.161, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[50] Guanglei Yang, Hao Tang, Mingli Ding, Nicu Sebe, and Elisa Ricci. Transformer-based attention networks for continuous pixel-wise prediction. In Proceedings of the IEEE/CVF International Conference on Computer vision, pages 16269-16279, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.232, + 0.483, + 0.302 + ], + "angle": 0, + "content": "[51] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.482, + 0.355 + ], + "angle": 0, + "content": "[52] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024. 1, 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.357, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[53] Xingyi Yang and Xinchao Wang. Diffusion model as representation learner. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18938-18949, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.482, + 0.483 + ], + "angle": 0, + "content": "[54] Rajeev Yasarla, Hong Cai, Jisoo Jeong, Yunxiao Shi, Risheek Garrepalli, and Fatih Porikli. Mamo: Leveraging memory and attention for monocular video depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8754-8764, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[55] Chongjie Ye, Lingteng Qiu, Xiaodong Gu, Qi Zuo, Yushuang Wu, Zilong Dong, Liefeng Bo, Yuliang Xiu, and Xiaoguang Han. Stablenormal: Reducing diffusion variance for stable and sharp normal. arXiv preprint arXiv:2406.16864, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.609 + ], + "angle": 0, + "content": "[56] Wei Yin, Xinlong Wang, Chunhua Shen, Yifan Liu, Zhi Tian, Songcen Xu, Changming Sun, and Dou Renyin. Diversedepth: Affine-invariant depth prediction using diverse data. arXiv preprint arXiv:2002.00569, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.608, + 0.482, + 0.678 + ], + "angle": 0, + "content": "[57] Wei Yin, Jianming Zhang, Oliver Wang, Simon Niklaus, Long Mai, Simon Chen, and Chunhua Shen. Learning to recover 3d scene shape from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 204-213, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.734 + ], + "angle": 0, + "content": "[58] Chi Zhang, Wei Yin, Billzb Wang, Gang Yu, Bin Fu, and Chunhua Shen. Hierarchical normalization for robust monocular depth estimation. Advances in Neural Information Processing Systems, 35:14128-14139, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[59] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence. Advances in Neural Information Processing Systems, 36, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.873 + ], + "angle": 0, + "content": "[60] Xiang Zhang, Bingxin Ke, Hayko Riemenschneider, Nando Metzger, Anton Obukhov, Markus Gross, Konrad Schindler, and Christopher Schroers. Betterdepth: Plug-and-play diffusion refiner for zero-shot monocular depth estimation. arXiv preprint arXiv:2407.17952, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[61] Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing text-to-image diffu" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.547, + 0.092, + 0.905, + 0.135 + ], + "angle": 0, + "content": "sion models for visual perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5729-5739, 2023. 2, 3" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6033" + } + ] +] \ No newline at end of file diff --git a/2025/FiffDepth_ Feed-forward Transformation of Diffusion-Based Generators for Detailed Depth Estimation/3c8dc7fb-98b2-4d77-857e-35d78d6eaaf6_origin.pdf b/2025/FiffDepth_ Feed-forward Transformation of Diffusion-Based Generators for Detailed Depth Estimation/3c8dc7fb-98b2-4d77-857e-35d78d6eaaf6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..27f40750311b1c2448365b34d13711bd44b180fb --- /dev/null +++ b/2025/FiffDepth_ Feed-forward Transformation of Diffusion-Based Generators for Detailed Depth Estimation/3c8dc7fb-98b2-4d77-857e-35d78d6eaaf6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9446462b12ee7b5e997b1ccaf7e5cd3c2a72839922c9431bbf9bc99b14c32ec6 +size 24788663 diff --git a/2025/FiffDepth_ Feed-forward Transformation of Diffusion-Based Generators for Detailed Depth Estimation/full.md b/2025/FiffDepth_ Feed-forward Transformation of Diffusion-Based Generators for Detailed Depth Estimation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..55fb9cb298d2fbd9b675e17c1c309b0c797f43d8 --- /dev/null +++ b/2025/FiffDepth_ Feed-forward Transformation of Diffusion-Based Generators for Detailed Depth Estimation/full.md @@ -0,0 +1,390 @@ +# FiffDepth: Feed-forward Transformation of Diffusion-Based Generators for Detailed Depth Estimation + +Yunpeng Bai Qixing Huang + +The University of Texas at Austin + +![](images/2936be9c03d6170cca4e7ff8b69beca682c98a4a7ad8844ccbb4a5115999ca38.jpg) + +![](images/7fd8549e482b0256e8162492758759437c39e715d5599b4dee427bb8d00d32e0.jpg) + +![](images/c584bded786f54295d81edc2f2e3b04cd8cd41ee241ee17ce1525f5e078a9833.jpg) +Input Image + +![](images/9ccce071ad3f046138e947a5a88d9547699f897cb26bd5058e5142ed7da6a4e3.jpg) + +![](images/1c2c1ec4c1b93b5076006b7e1758325bd37fec7f87565d5a5ab84fc04bbfa9f6.jpg) + +![](images/594cfb007bb8a09bc0e1c526a58c1de3808211b0dad9057ae2fc7d99f7e82653.jpg) +FifDepth + +![](images/ed2630d94aafe8f960a30cd44ed5a0f96f95d849255669835d466cd3248d7c46.jpg) + +![](images/4e83a5ccbcfc041e27e9dde36a21799f2082c2e84fe5f456404a6e12d54eef4d.jpg) + +![](images/7573848849c9e461db6024935d95e2c8d09a87903046bf246f59aff950a93011.jpg) +Depth Anything v2 +Figure 1. Compared to other methods, our model achieves more accurate details and better generalization in depth estimation. The final row shows the point cloud generated from the estimated depth results, and the corresponding depth map can be referenced in Figure 5. + +![](images/1d6c6030d5726e9c8b84718e5544235beab47812d087f2836fbb1c1629fea4db.jpg) + +![](images/a0f45a70f9ff381299db6c9bb83842889e3a7847178667cf63897b1d4b6bc9b8.jpg) + +![](images/18c80cee368663e70893ce5fef6bf4023d4c22bb42c7c4ea6836f27788de9e25.jpg) +Depth Pro + +# Abstract + +Monocular Depth Estimation (MDE) is a fundamental 3D vision problem with numerous applications such as 3D scene reconstruction, autonomous navigation, and AI content creation. However, robust and generalizable MDE remains challenging due to limited real-world labeled data and distribution gaps between synthetic datasets and real data. Existing methods often struggle with real-world test data with low efficiency, reduced accuracy, and lack of detail. To address these issues, we propose an efficient MDE approach named FiftDepth. The key feature of FiftDepth is its use of diffusion priors. It transforms diffusion-based image generators into a feed-forward architecture for detailed depth estimation. FiftDepth preserves key generative features and integrates the strong generalization capabilities of models like DINOv2. Through benchmark evaluations, we demonstrate that FiftDepth achieves exceptional + +accuracy, stability, and fine-grained detail, offering significant improvements in MDE performance against state-of-the-art MDE approaches. The paper's source code is available here: https://yunpeng1998.github.io/FiffDepth/ + +# 1. Introduction + +Monocular Depth Estimation (MDE) is a fundamental 3D vision problem with numerous applications in 3D scene reconstruction [46], autonomous navigation [47], and more recently in generating AI-based content [2, 41]. MDE has made significant progress in the era of deep learning [21, 34, 35, 52]. Neural networks trained from paired datasets of images and pixel depth exhibit encouraging results and often outperform non-deep learning based counterparts that build on monocular depth cues. Despite significant progress, fundamental challenges remain in efficiency, + +accuracy, and generalization on diverse in-the-wild data. This is because 1) real depth datasets are usually noisy, and 2) while synthetic data can be used, there exist domain gaps between synthetic datasets and diverse in-the-wild data. + +Specifically, current MDE research relies primarily on synthetic data due to its high-quality annotations and controlled environments. However, the scale and variety of synthetic datasets remain insufficient for comprehensive training. To address this, synthetic-to-real transfer techniques and the utilization of pre-trained models have emerged as viable solutions. Among pre-trained models, generative networks [37] preserve intricate image details more effectively than feed-forward networks (FFNs) like DINOv2 [31], thus holding greater promise for dense prediction models. However, generative models, while detail-rich, often fall short in synthetic-to-real transfer due to their limited generalization capabilities outside the training domain. + +Previous studies [15, 21] have adopted pre-trained diffusion models, where the main idea is to directly finetuning pre-trained RGB image diffusion models into depth map diffusion generation models conditioned on images. However, this method may not be ideal, as dense prediction models require certainty over diversity. The introduction of any noise or uncertainty during the generation process by these methods is sub-optimal. In contrast, we observe that simply using the denoising diffusion module in a feedforward manner yields better and more stable results. This method capitalizes on extending the trajectories of image diffusion models into the depth domain, representing a significant advancement in both accuracy and efficiency for generative model-based depth estimation methods. + +Specifically, we optimize diffusion trajectories for MDE tasks. To enable the diffusion model to better retain certain detailed generative features when fine-tuned into an MDE model, we preserve the original generative training trajectory while training the model for depth prediction, aiming to maintain the detailed features of the original generative model as much as possible. Furthermore, recognizing the limitations of fintuned diffusion models in maintaining robustness in diverse real-world images—with inaccuracies in predicted depth occurring mainly in the low-frequency components—we leverage the strengths of a DINOv2 [31] based model, which excels at predicting accurate low-frequency depth despite its reduced fine detail. To address this issue, we use the diffusion model itself to learn a filter that refines its inaccurate predictions, producing low-frequency outputs with a detail level similar to that of DINOv2's predictions, thereby matching DINOv2's results and optimizing the low-frequency component of our output. This approach allows us to incorporate a large amount of real image data for training without sacrificing detail preservation of generative models, while simultaneously leveraging DINOv2's strong generalization capabilities to enhance + +the overall robustness and precision of our MDE models. + +In summary, the contributions of our work are as follows: 1) We propose an improved approach for transforming generative models into dense prediction models, specifically for depth prediction tasks, by leveraging diffusion model trajectories in a more stable, feed-forward manner; 2) We introduce a novel distillation method that transfers the robust generalization capabilities of models like DINOv2 to diffusion backbones; 3) Our method demonstrates higher stability, accuracy, and efficiency in depth estimation compared to other approaches based on generative models; 4) Compared to other FFN models, our approach achieves more detailed prediction results, marking a significant advancement in the field of MDE. + +# 2. Related Works + +# 2.1. Depth Estimation + +Depth estimation has always been a widely researched topic, with numerous studies conducted in the past including single-image [1, 13, 14, 24, 28, 32, 50] and video depth estimation [10, 23, 43, 48, 54]. Early efforts in image depth estimation, such as DIW [8] and OASIS [9], focused on predicting relative (ordinal) depth. Subsequent approaches, including MegaDepth [27] and DiverseDepth [56], used extensive collections of photographs from the Internet to develop models that adapt to unseen data, while MiDaS [34] improved generalization by incorporating a diverse range of datasets during training. Recent advances, including DPT [35] and Omnidata [12], adopted transformer-based architectures to improve depth estimation performance. + +However, due to limitations in models and data, these methods exhibit very limited generalization in various open-world scenes. Recently, some efforts [15, 21, 51, 52] in image depth estimation have made open-world depth estimation feasible by leveraging the power of vast amounts of unlabeled image data and the capabilities of pre-trained generative models. Although these methods have achieved remarkable progress in the field of depth estimation, challenges such as low efficiency, limited generalization, and insufficient detail preservation remain. + +# 2.2. Diffusion Models as Representation Learner + +The diffusion model [3, 37] training process strongly resembles that of denoising autoencoders (DAE) [4, 19, 45], as both are designed to recover clean images from noise-corrupted inputs. Recent studies have shown that semantic image representations learned from diffusion models can be effectively used for various downstream recognition tasks, such as correspondence [59], semantic segmentation [61], and keypoint detection [53]. Notably, the features extracted from diffusion models tend to preserve more intricate details, which has prompted the adoption of pre + +![](images/4fd3a713f38948f876a77c37a1f7f9105f2631393f5d035b62e6bba210784ae9.jpg) +Figure 2. Overview of the proposed method. To simplify the representation, all the images we used above correspond to the respective latents. We transform the pre-trained diffusion model into a feed-forward approach for depth prediction, using only the result at $t = 0$ as the output during inference. During training, at $t = 0$ , we use synthetic data to ensure detailed results, while at $t = -1$ , we leverage pseudo-labels generated by DINOv2 for supervision. + +![](images/00044d553bde24e5bd68df7ff62ded032f3e934065480a77c66d061cade33c36.jpg) + +trained diffusion models for dense prediction tasks. Recent studies, including Marigold [21] and GeoWizard [15], rely on the standard diffusion framework and pre-trained parameters to perform dense prediction tasks. Emerging approaches [18, 49, 55] attempt to bypass the stochastic phase of diffusion models by employing deterministic frameworks. However, these adaptations lack deeper exploration of the model's potential, often leading to suboptimal performance and the need for additional post-processing to refine results. Moreover, diffusion-based techniques generally exhibit limited generalization capabilities. For instance, BetterDepth [60] incorporates external depth priors as inputs but remains a stochastic framework and heavily depends on the quality of other models. + +# 3. Method + +# 3.1. Overview: Feed-forward Transformation of Diffusion Models + +Fundamentally, generative models construct mappings between a latent space and the ambient data space. These mappings align closely with the needs of depth estimation and other visual recognition tasks, where precise mappings from image data to the corresponding labels are essential. In these tasks, the scarcity of labeled data often limits the precision of the trained models. In the context of learning from a collection of unlabeled data instances, advanced gen + +erative models, trained on massive datasets, are capable of learning robust mappings. They exhibit great promise for transferring knowledge to other visual prediction applications [53, 59, 61]. Our work builds on this approach, further exploring its application in depth estimation. + +Specifically, we finetune a well-trained Stable Diffusion (SD) [37] model to construct our model. A diffusion process constructs multiple intermediate states by progressively adding noise to the data $\mathbf{x}_0$ , defined as $\mathbf{x}_t = \sqrt{\bar{\alpha}_t}\mathbf{x}_0 + \sqrt{1 - \bar{\alpha}_t}\epsilon$ , where $\epsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ , and $\bar{\alpha}_t \coloneqq \prod_{s=1}^{t} 1 - \beta_s$ with noise schedule $\{\beta_1, \ldots, \beta_T\}$ . Then, diffusion generative model gradually learns the mapping between two distributions by denoising each step, denoted as $x_T \to x_{T-1} \to \dots \to x_0$ . The diffusion model is primarily a denoising model $\epsilon_\theta$ that follows a loss function $\mathcal{L}_{\mathrm{simple}} = \mathbb{E}_{x_0,\epsilon,t} \left[ \| \epsilon - \epsilon_\theta(x_t,t) \|^2 \right]$ . + +In depth estimation, the mapping from visual images to depth labels should ideally be deterministic. By leveraging a robust, pre-trained mapping within the generative model, there is no need to decompose the image-label mapping into multiple steps during training. Instead, we can extend the mapping trajectories of the existing diffusion model directly into the depth domain by adapting the learned diffusion process to act as a deterministic one-step feed-forward network. Since our MDE model is built upon the diffusion trajectory for its extension, we set the time step input as + +$t = 0$ in this feed-forward step. + +$$ +\mathbf {d} _ {0} = \hat {\boldsymbol {\epsilon}} _ {\theta} \left(\mathbf {x} _ {0}, t = 0\right). \tag {1} +$$ + +Since we use diffusion model's parameters to construct our network, we also use $\hat{\epsilon}_{\theta}$ to represent our network model here. The input $\mathbf{x}_0$ is the latent representation of RGB image, and $\mathbf{d}_0$ is the latent representation of "depth image," following Marigold's encoding approach. The depth latent $\mathbf{d}_0$ can be reconstructed into a depth map using the VAE of SD with negligible error. While there have been preliminary attempts [18, 49] to use similar approaches, they remain in a nascent stage and lack the precision, robustness, and richness in detail needed for effective depth estimation. In the following sections, we will delve into each step of our method, detailing how we adapt the diffusion model to enhance accuracy and detail in depth estimation tasks, such as the key technical contributions of preserving diffusion trajectories and improving synthetic-to-real robustness, as discussed in Section 3.2 and Section 3.3, respectively. The overall workflow of the method is illustrated in Figure 2. + +# 3.2. Keeping Diffusion Trajectories + +Since our approach leverages the trajectory of the diffusion model, it is crucial to prevent degradation of this trajectory during training. To achieve this, when fine-tuning the diffusion model to transition it into a feed-forward depth estimator, we simultaneously maintain the feed-forward step along with the preceding denoising training steps from the original diffusion model. While this trajectory was initially developed for image generation, directly applying it as-is does not facilitate an optimal transition to the depth domain. Thus, instead of predicting purely image-based latents, we modify the target latent to be a blend of image and depth representations. + +$$ +\begin{array}{l} \mathbf {b} _ {0} = \gamma \mathbf {x} _ {0} + (1 - \gamma) \mathbf {d} _ {0}, \\ \mathbf {b} _ {t} = \sqrt {\bar {\alpha} _ {t}} \mathbf {b} _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \boldsymbol {\epsilon}, t \in \{1, \dots , T \}. \tag {2} \\ \end{array} +$$ + +Here, $\mathbf{b}_0$ represents the blended latent. $\gamma$ controls the balance between the image and depth latents. In this process, we also use v-prediction re-parameterization approach [38] to define the training objective: + +$$ +\begin{array}{l} \mathbf {v} _ {t} = \sqrt {\bar {\alpha} _ {t}} \boldsymbol {\epsilon} - \sqrt {1 - \bar {\alpha} _ {t}} \mathbf {b} _ {0}, \\ L _ {k} = \left\| \mathbf {v} _ {t} - \hat {\boldsymbol {\epsilon}} _ {\theta} \left(\mathbf {b} _ {t}, t\right) \right\| _ {2} ^ {2}, t \in \{1, \dots , T \}. \tag {3} \\ \end{array} +$$ + +Intuitively, this approach forces the diffusion model to preserve the shared features between the image generation task and the depth estimation task, which are captured in the blended training target. Therefore, it allows the diffusion model to adapt more naturally to depth estimation while retaining essential generative features. Consequently, during fine-tuning, the model maintains features that enhance the + +accuracy and detail of depth predictions. This part is used exclusively during training. At inference time, our model functions as a fully deterministic framework. + +# 3.3. Learnable Filter Distillation + +Following previous methods, the above training process only uses synthetic data because it provides high-quality depth Ground Truth. However, this reliance limits both Marigold [21] and our approach, as SD-based MDE models trained solely on synthetic datasets often struggle to generalize well to in-the-wild data. Nevertheless, in this case, the predictions produced by the SD-based MDE model still retain the necessary details—precisely those details we expect to preserve in the final output. Therefore, enhancing the model's robustness essentially means improving the accuracy of the low-frequency components in the model's output on real image. + +Inspired by prior work [52], we observe that DINOv2 [31], trained on synthetic data, can generalize effectively to real-world images. However, its depth estimates often lack the necessary fine details—in other words, it can accurately predict the low-frequency depth components for massive real images but misses the high-frequency details. This characteristic aligns well with our needs, so we attempt to leverage the abundance of labels generated by the DINOv2 model to enhance our model's robustness. A straightforward method, as used in Depth Anything v2 [52], is to use a fintuned DINOv2-G depth model to generate pseudo labels, expanding the training set. However, since DINOv2's depth predictions lack fine detail, directly incorporating these pseudo labels on the feed-forward output $\mathbf{d}_0$ risks ignoring the detailed features inherent to our model. + +To address this, we propose learning a filter, denoted as $F(d_0)$ that processes our output to remove high-frequency details (e.g., fine details). This filtering produces results at a detail level similar to that of DINOv2, allowing us to focus supervision solely on the less accurate low-frequency components. The filter $F$ is obtained through learning. In fact, we observed that the fine-tuned SD model itself already serves as an effective filter learner because the diffusion model is inherently designed to model the subtle differences across different time steps. Therefore, we directly apply an additional SD step on the feed-forward output $\mathbf{d}_0$ to simulate this filter. To maintain consistency with the inputs and concepts used in the diffusion model, we represent this filter $F: t_0 \rightarrow t_{-1}$ as the process that transforms the output from $t_0$ to $t_{-1}$ : + +$$ +\mathbf {d} _ {- 1} = \hat {\mathbf {e}} _ {\theta} \left(\mathbf {d} _ {0}, t = - 1\right). \tag {4} +$$ + +At this stage, we can use the labels predicted by DINOv2 to supervise $\mathbf{d}_{-1}$ , allowing us to transfer DINOv2's robustness without interfering with the detailed features in $\mathbf{d}_0$ . In + +![](images/ad53793f8cce5cd1662bd934afa7d02c492e3471a6e3070ae8af51887fcebe03.jpg) +Figure 3. Filter learning. We use a learnable filter to map our results to detail levels similar to DINOv2's, matching its outputs and thereby transferring DINOv2's generalization capabilities to our model without compromising our inherent details. + +this process, we use real-world image data for $\mathbf{x}_0$ . Figure 3 illustrates the effects of Eq. (4). + +# 3.4. Final Objective + +For $\mathbf{d}_0$ and $\mathbf{d}_{-1}$ , we follow the method of MiDaS [34] using MAE loss $L_{\mathrm{MAE}}$ and gradient matching loss $L_{\mathrm{GM}}$ as depth loss. However, unlike the original approach, we apply these losses in the latent space of Stable Diffusion instead: + +$$ +L _ {\mathrm {M A E}} \left(\mathbf {d}, \mathbf {d} ^ {*}\right) = \frac {1}{M} \sum_ {i = 1} ^ {M} \left| \mathbf {d} _ {i} - \mathbf {d} _ {i} ^ {*} \right|, \tag {5} +$$ + +where $\mathbf{d}$ represents the ground truth latent, and $\mathbf{d}^*$ is the model's predicted value. $M$ denotes the number of pixels in the depth latent. + +$$ +\mathcal {L} _ {\mathrm {G M}} \left(\mathbf {d}, \mathbf {d} ^ {*}\right) = \frac {1}{M} \sum_ {i = 1} ^ {M} \left(| \nabla_ {x} R _ {i} | + | \nabla_ {y} R _ {i} |\right), \tag {6} +$$ + +where $R_{i} = \mathbf{d}_{i} - \mathbf{d}_{i}^{*}$ . Hence, the final objective is expressed as a weighted sum of the losses $L_{\mathrm{MAE}}$ , $L_{\mathrm{GM}}$ , and $L_{k}$ : + +$$ +\begin{array}{l} L _ {\text {f i n a l}} = \sum_ {t \in \{- 1, 0 \}} \left(\lambda_ {\mathrm {M A E}} L _ {\mathrm {M A E}} \left(\mathbf {d} _ {t}, \mathbf {d} _ {t} ^ {*}\right) \right. \\ + \lambda_ {\mathrm {G M}} L _ {\mathrm {G M}} \left(\mathbf {d} _ {t}, \mathbf {d} _ {t} ^ {*}\right)) + \lambda_ {k} L _ {k}. \tag {7} \\ \end{array} +$$ + +$\lambda_{\mathrm{MAE}}$ $\lambda_{\mathrm{GM}}$ , and $\lambda_{k}$ are the weighting factors for their respective loss term. $\mathbf{d}_0^*$ represents the ground truth (GT) from the synthetic dataset, and $\mathbf{d}_{-1}^{*}$ represents the pseudo label generated by DINOv2 for real images. + +# 4. Experiments + +# 4.1. Implementation Details + +During the training process, we preserve the diffusion trajectory while following the original DDPM noise sched + +uler [20] using 1000 diffusion steps. To better leverage pre-trained models, we use the Depth Anything V2-Large model as the DINO v2 model for supervision. Our training dataset consists of two parts. For training at $t = 0$ and during trajectory retention, we follow previous approaches and use two synthetic datasets, Hypersim [36] and Virtual KITTI [7], which cover both indoor and outdoor scenes, with a total of 74K images. For training at $t = -1$ , we use real-world data from the LAION-Art dataset, a subset of LAION-5B [40] containing 8 million samples. However, we observed that training with only 0.2 million samples was sufficient. Synthetic data and real data each account for half of each batch. The parameters are set as follows: $\gamma = 0.5$ , $\lambda_{\mathrm{MAE}} = 1$ , $\lambda_{\mathrm{GM}} = 0.5$ , $\lambda_{k} = 0.2$ . + +# 4.2. Comparison + +Zero-shot affine-invariant depth. For the evaluation of affine-invariant depth, we use the same datasets and evaluation protocol as Marigold. These datasets include NYUv2 [42], ScanNet [11], KITTI [16], ETH3D [39], and DIODE [44]. We compared FiftDepth with 14 methods that produce affine-invariant depth maps/disparities, all claiming zero-shot generalization capabilities. These include the earlier methods [12, 34, 35, 56-58], as well as the more recent ones [15, 17, 18, 21, 29, 49, 51, 52]. As shown in Table 1, FiftDepth achieves the best or state-of-the-art comparable results in most test scenarios. For visualization results, please refer to Figure 4. Our method not only accurately predicts the relative depth relationships but also excels in identifying and predicting depth for very fine objects. We also evaluate our method on the DA-2K introduced by Depth Anything v2. On this dataset, our method also performs comparably to Depth Anything v2. + +Compared to methods like Depth Anything, which rely on massive training datasets, our model achieves comparable generalization to DAv2 while being trained on only a small amount of real data. To validate the generalization capability of our method, we present some test examples from special scenarios in Figure 5, including games, artworks, AI-generated content, and movies. Our method shows comparable generalization to DAv2 while preserving more details. In contrast, other methods' results are unsatisfactory in both generalization and detail preservation. + +Zero-shot boundaries. To further demonstrate the accuracy of our method in predicting fine structures, we also employ the Zero-shot Boundaries Metric introduced in the recent work Depth Pro [5] to evaluate boundary sharpness. Following Depth Pro, we compute the depth average boundary F1 score for datasets with ground truth and the boundary recall (R) for datasets with matting or segmentation annotations. The former datasets include Sintel [6], Spring [30], and iBims [22], while the latter include AM-2k [26], P3M-10k [25], and DIS-5k [33]. For details on the boundaries + +
MethodTraining DataNYUv2KITTIETH3DScanNetDIODE-FullDA-2K
AbsRel ↓δ1 ↑AbsRel ↓δ1 ↑AbsRel ↓δ1 ↑AbsRel ↓δ1 ↑AbsRel ↓δ1 ↑Acc (%)
DiverseDepth320K11.787.519.070.422.869.410.988.237.663.179.3
MiDaS2M11.188.523.663.018.475.212.184.633.271.580.6
LeReS354K9.091.614.978.417.177.79.191.727.176.681.1
Omnidata v212.2M7.494.514.983.516.677.87.593.633.974.276.8
HDN300K6.994.811.586.712.183.38.093.924.678.085.7
DPT1.4M9.890.310.090.17.894.68.293.418.275.883.2
Marigold74K*5.596.49.991.66.496.06.495.130.877.386.8
e2e-ft74K*5.296.69.691.96.495.95.896.230.277.983.6
DepthFM74K*6.595.68.393.47.895.96.894.924.574.185.8
GenPercept74K*5.696.013.084.27.095.66.296.130.777.685.1
Lotus-D59K*5.396.78.192.86.595.35.896.329.978.186.8
Lotus-G59K*5.496.68.587.76.296.16.096.029.478.586.2
GeoWizard280K*5.296.69.792.16.496.16.195.329.779.288.1
DepthAnything v1-L62.6M*4.398.17.694.712.788.24.298.027.775.988.5
DepthAnything v2-L62.6M*4.597.97.494.613.186.54.297.826.275.497.1
FiffDepth (Ours)274K*4.497.87.393.57.197.24.297.923.978.197.1
+ +Table 1. Quantitative comparison with other affine-invariant depth estimators on several zero-shot benchmarks. We use AbsRel (absolute relative error: $|d^{*} - d| / d$ ) and $\delta_{1}$ (percentage of max $(d^{*} / d, d / d^{*}) < 1.25$ ). All metrics are reported as percentages; bold numbers are the best, underscored second best. Methods marked with an asterisk (*) utilize pre-trained models. + +![](images/b8e003d99145206d23b58668055b5a806445f0da1433bd367aba91c34e6f60f8.jpg) +Figure 4. Qualitative comparison across different datasets. Our method is capable of predicting the depth of various fine objects, such as lampposts, railings, and chair legs. + +![](images/f06461048ea5d2948f3a0da1823c05df844e90024b3f8b1447475d944e66f98b.jpg) + +![](images/f02883b984eab1ab1cb4036f2059aa36d41cafbb6b7321c5a6a585f0566cd21b.jpg) + +![](images/d93e4784913ffcee796906c973888065ca5e137f5f8fbe87a26c8b911cded3ec.jpg) + +![](images/d37fb4dc68b4c71d09dd3dd09929f52b4286db9a2a81d69cd0538ac7417d0a99.jpg) + +![](images/e68d29acd39c2a95b7584686916156b059ed8d0be4d09b2991edf0330ced5ace.jpg) + +![](images/9ad3aabcad67d77ab6c73451084c127d247dd955ef0b3ddf5c5c0b41766f2073.jpg) + +![](images/e55bd8ffe43077539fca55c09cda96734f9c49e5c1c106592d0e106038943c6b.jpg) + +![](images/96f7b1e29b1b75219063e14fdb4e1f11bbd55f6e733bcac85d1b9398ac627601.jpg) + +![](images/8a32d8e070a3cec2dda0c6d6bf92112422970d2e45c102ccc7b1831a14aeb023.jpg) + +![](images/7cbfff5359cce5480dda847cc364eaf4374e459e6605e1f14a42a1e38deeaa9d.jpg) + +![](images/f15d998a2f674f41ad93b904533e334c6f8fa9905da0f9c431b9a8027bb75086.jpg) + +![](images/65b15ad3d14dcfdf37f169530e387220c0a0f4be1db9c162bc43b40135ded43b.jpg) + +![](images/5b0136e4a70814225ff9ad68e12d98fd6cdd47a2533a82d83106c1c9cf2ff34d.jpg) + +![](images/d23b97a3d5053b208c1695cff1d216aa44c759e877dae29b50cf0e70edc21e3a.jpg) + +![](images/b01e457706727a58340b4782b335d9894c48b283b65ef0613003ad06ba0e3861.jpg) + +![](images/7dbdaec09dc5d19332391d91976db620033167d5201946c6543662e99d052c14.jpg) + +![](images/7c6dd9c5e0d37fa753389b59a2429bf602a835309e8bad16363da928abb837da.jpg) + +![](images/55e6dbca66a3508f40a4446e71f825e3c86c9546b39cffd8090a46ebafc322a0.jpg) + +![](images/352cae2494fb7f7f5a3064623cb9ad8fad88c60768efc92c343aa5d31f1e356f.jpg) + +![](images/7e12cb90c72abd1f0ca518327f5992f2e1619ef7e28f77df30c38f6c9d24af84.jpg) + +![](images/9ea1be9da8d3f3619bb32639cd42d42f504dc2ab8bb70cffcf3987ab8dbf565b.jpg) + +![](images/f5c295656393be322a38e937da6b92b3df19fa51d55f639e31d3fbc2433a1338.jpg) + +![](images/631c684044426f78ee5948c8e54c73fcf313555544667f9e23b799a068628b18.jpg) + +![](images/5574e4f75cd99d2ddf79fe59fa44d4a1de17aafeca7aee5ae327a3704bbfff8f.jpg) + +![](images/a096fcd922912a6b0cd0fded11d29e19d162df8e2658f0286894bf77f89ebb48.jpg) + +![](images/a953e849b95fd4a0cf76f04ebcaf4e625758c85498c036f63154c6f025797bbd.jpg) + +![](images/14668697365c19fbe1a5240ad5f67ee20ad025640d6d803f360482e44da0d575.jpg) + +![](images/93d5e58e6b876b030d6748806502b26f564d44495ed94a535cb1372eaa3a0c61.jpg) + +![](images/3e9a5d12703748d56e89df067705455069a798bf2fa74705a1df6891334216b2.jpg) +Figure 5. Qualitative comparison on special scenarios. In the special scenarios of games, artworks, AI-generated content, and movies, our method demonstrates strong generalization capability and the ability to predict detailed depth. + +![](images/0cb53a9ded46b4334bd4a68fcd10f4e388dcb382db6c0f2e4e687f25ed60b8b0.jpg) + +![](images/1007fe4e7e5f20e0f07830c3f573be5c74c82b824042b410f2c549c10c02ff71.jpg) + +![](images/dfb46957cde7b660ae50ea55cbc9c2e15d7168ccdab2475c6d7bde1d8bd0151e.jpg) + +![](images/db957fbc803541c18c65c0ab7824d0e348c83233adf0f8b163ccff6ee28da947.jpg) + +![](images/4408792880f35668c66f7a1773d948ceb3ccdb98235364b8f9feb83ce3b51363.jpg) +Input Image + +![](images/7cec304557161c7139f9dad19e333aa9d0b0889d7adf4e89baf4594d2f0c7293.jpg) + +![](images/a55904f4b059b258a9ec9d50b06981f63f4f8fcb643adea9ced92af0113eb18a.jpg) +Alpha Matte + +![](images/85a2bf0fb2277c68cdf14851d8701583ed6dd48677df072472a1ee31d230754c.jpg) + +![](images/4f7f9728d41f27fc60491e57903b1c9948e658fa69a86047dcb9e3a382dcd469.jpg) +FifDepth +Figure 6. Boundary visualization comparison. These samples are from the AM-2k dataset. + +![](images/3c8e678e149d1839280e615bdb6afb36d945b59824e59343e53860876a6cb6db.jpg) + +![](images/b23fe43b85fe0a0494b3d00e83e6d84f4a44129720dab601c190bcf02d3525b5.jpg) +Depth Pro + +![](images/ed56d1b4aaace17d2a80e4834450a3833a9df91728d4161330364119eba2c64a.jpg) + +![](images/d820bf91a99cda0ebb8c845a6d48d36e17e9aa3eb44f650a5029d52a11b3fd55.jpg) +Depth Any. v2 + +![](images/e4403e627c42001ab1b89e07a197446a4e327d306ef317c293eceaf7311598ac.jpg) + +![](images/e0e519a1d3d2c85aaf2bc92d9a7274adf247accebe061d20beb792f37289fd62.jpg) +DepthFM + +
MethodSintel F1↑Spring F1↑iBims F1↑AM R↑P3M R↑DIS R↑
DepthAnything v20.2280.0560.1110.1070.1310.056
Depth Pro0.4090.0790.1760.1730.1680.077
FiffDepth (Ours)0.4230.0860.1890.1760.1790.091
+ +Table 2. Zero-shot boundary accuracy. We provide the F1 score for datasets containing ground-truth depth and boundary recall (R) for those with matting or segmentation labels. + +
MethodMarigoldMarigold (LCM)GeoWizardDepthFMDepthAnything v2-LDepth ProOurs
Time (s)1031.7190.390.0260.230.092
+ +Table 3. Running time comparison. We perform inference on ${100512} \times {512}$ images using these methods and report the average time. + +![](images/152617c8d3931740ac93ee795d9a59ecddd9329bd3f4bc2c6cca17915c4c4052.jpg) +Input Image + +![](images/3319f3671586c6464afb2e2cad2ef05aedfe65842b41d1c2497d96a7001f9c09.jpg) +w/o blend + +![](images/2ade62655ef2de54b797386ae8476937bfba6c2503ec0df92a1b27166a12770a.jpg) +w/o $L_{k}$ + +![](images/0e97e459a08548dccd2934beb21bb88f8226d3f93d1731f8f7879a9bff50ae5f.jpg) +FiffDepth Output + +![](images/cb318011e50e35e791049b038982c206117c2b031f64ab0749f606c0dca4ff75.jpg) +Input Image + +![](images/5a0211de8ca57da8628960965f017df04f6b749c95454e0fed8d50c0b4e7dc5d.jpg) +w/o DINO + +![](images/976d74704298fc43877737e786a6547ccc0a9c7fa278851b2626b0ce4c6c9e27.jpg) +DINO on $d_0$ + +![](images/0d3ffd45d5694eef30e5b07fee1d6b9e795531cd0f168a4b47e94c9b36a587cb.jpg) +FiffDepth Output +Figure 7. Ablation studies. The generalization capability and depth details of the method are affected when some essential components are missing. + +metric and its computation, please refer to the Depth Pro paper for further details. Quantitative comparisons in Table 2 demonstrate that our method surpasses Depth Pro and other approaches in boundary prediction. Additionally, the visual results in Figure 6 further validate that our method predicts more accurate boundaries. + +Running time. Generative MDE models adopt the diffusion paradigm, and the resulting instability necessitates test-time assembly, leading to a lengthy inference time. In contrast, our feed-forward approach provides significant efficiency advantages. We evaluate the average inference time for a $512 \times 512$ image on an NVIDIA Titan RTX GPU. As shown in Table 3, our method significantly outperforms other generative approaches in terms of efficiency and achieves performance comparable to DAv2. We tested these methods using their default settings. + +Ablation studies. We conduct ablation studies to validate components of our method. Keeping the diffusion trajectory but predicting purely image latents affects the relative depth relationships between objects (1st row, 1st result in Fig. 7). Without keeping trajectory, some details are lost (1st row, 2nd result in Fig. 7). Omitting DINO supervision impacts the model's generalization ability (2nd row, 1st result in Fig. 7). Using DINO supervision at $d_{0}$ also reduces details (2nd row, 2nd result in Fig. 7). + +# 5. Conclusion + +In this work, we transform diffusion models into stable, feed-forward depth estimators, achieving significant improvements in accuracy and efficiency over generative model-based methods. By combining the detail preservation of generative models with the robust generalization of FFN models like DINOv2, our hybrid approach bridges the synthetic-to-real gap, enhancing stability, predictability, and resolution in MDE for diverse real-world scenarios. + +Acknowledgements. Q. Huang acknowledges the support of NSF-2047677, NSF-2413161, NSF-2504906, NSF-2515626, and GIFTs from Adobe and Google. This research has been supported by computing support on the Vista GPU Cluster through the Center for Generative AI (CGAI) and the Texas Advanced Computing Center (TACC) at UT Austin. + +# References + +[1] Shubhra Aich, Jean Marie Uwabeza Vianney, Md Amirul Islam, and Mannat Kaur Bingbing Liu. Bidirectional attention network for monocular depth estimation. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 11746-11752. IEEE, 2021. 2 +[2] Yunpeng Bai, Yanbo Fan, Xuan Wang, Yong Zhang, Jingxiang Sun, Chun Yuan, and Ying Shan. High-fidelity facial avatar reconstruction from monocular video with generative priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4541-4551, 2023. 1 +[3] Yunpeng Bai, Xintao Wang, Yan-Pei Cao, Yixiao Ge, Chun Yuan, and Ying Shan. Dreamdiffusion: High-quality eegt-to-image generation with temporal masked signal modeling and clip alignment. In European Conference on Computer Vision, pages 472-488. Springer, 2024. 2 +[4] Yoshua Bengio, Li Yao, Guillaume Alain, and Pascal Vincent. Generalized denoising auto-encoders as generative models. Advances in neural information processing systems, 26, 2013. 2 +[5] Aleksei Bochkovskii, Amaël Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. arXiv preprint arXiv:2410.02073, 2024. 5 +[6] Daniel J Butler, Jonas Wulff, Garrett B Stanley, and Michael J Black. A naturalistic open source movie for optical flow evaluation. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part VI 12, pages 611-625. Springer, 2012. 5 +[7] Yohann Cabon, Naila Murray, and Martin Humenberger. Virtual katii 2. arXiv preprint arXiv:2001.10773, 2020. 5 +[8] Weifeng Chen, Zhao Fu, Dawei Yang, and Jia Deng. Single-image depth perception in the wild. Advances in neural information processing systems, 29, 2016. 2 +[9] Weifeng Chen, Shengyi Qian, David Fan, Noriyuki Kojima, Max Hamilton, and Jia Deng. Oasis: A large-scale dataset for single image 3d in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 679-688, 2020. 2 +[10] Yuhua Chen, Cordelia Schmid, and Cristian Sminchisescu. Self-supervised learning with geometric constraints in monocular video: Connecting flow, depth, and camera. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7063-7072, 2019. 2 +[11] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 5 +[12] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multitask mid-level vision datasets from 3d scans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10786-10796, 2021. 2, 5 + +[13] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. Advances in neural information processing systems, 27, 2014. 2 +[14] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Batmanghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2002-2011, 2018. 2 +[15] Xiao Fu, Wei Yin, Mu Hu, Kaixuan Wang, Yuexin Ma, Ping Tan, Shaojie Shen, Dahua Lin, and Xiaoxiao Long. Geowizard: Unleashing the diffusion priors for 3d geometry estimation from a single image. arXiv preprint arXiv:2403.12013, 2024. 2, 3, 5 +[16] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 5 +[17] Ming Gui, Johannes S Fischer, Ulrich Prestel, Pingchuan Ma, Dmytro Kotovenko, Olga Grebenkova, Stefan Andreas Baumann, Vincent Tao Hu, and Björn Ommer. Depthfm: Fast monocular depth estimation with flow matching. arXiv preprint arXiv:2403.13788, 2024. 5 +[18] Jing He, Haodong Li, Wei Yin, Yixun Liang, Leheng Li, Kaiqiang Zhou, Hongbo Liu, Bingbing Liu, and YingCong Chen. Lotus: Diffusion-based visual foundation model for high-quality dense prediction. arXiv preprint arXiv:2409.18124, 2024. 3, 4, 5 +[19] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2 +[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 5 +[21] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 3, 4, 5 +[22] Tobias Koch, Lukas Liebel, Friedrich Fraundorfer, and Marco Korner. Evaluation of cnn-based single-image depth estimation methods. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, pages 0–0, 2018. 5 +[23] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1611-1621, 2021. 2 +[24] Jin Han Lee, Myung-Kyu Han, Dong Wook Ko, and Il Hong Suh. From big to small: Multi-scale local planar guidance for monocular depth estimation. arXiv preprint arXiv:1907.10326, 2019. 2 +[25] Jizhizi Li, Sihan Ma, Jing Zhang, and Dacheng Tao. Privacy-preserving portrait matting. In Proceedings of the 29th ACM + +international conference on multimedia, pages 3501-3509, 2021. 5 +[26] Jizhizi Li, Jing Zhang, Stephen J Maybank, and Dacheng Tao. Bridging composite and real: towards end-to-end deep image matting. International Journal of Computer Vision, 130(2):246-266, 2022. 5 +[27] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2041-2050, 2018. 2 +[28] Zhenyu Li, Zehui Chen, Xianming Liu, and Junjun Jiang. Depthformer: Exploiting long-range correlation and local information for accurate monocular depth estimation. Machine Intelligence Research, 20(6):837-854, 2023. 2 +[29] Gonzalo Martin Garcia, Karim Abou Zeid, Christian Schmidt, Daan de Geus, Alexander Hermans, and Bastian Leibe. Fine-tuning image-conditional diffusion models is easier than you think. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2025. 5 +[30] Lukas Mehl, Jenny Schmalfuss, Azin Jahedi, Yaroslava Nalivayko, and Andres Bruhn. Spring: A high-resolution high-detail dataset and benchmark for scene flow, optical flow and stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4981-4991, 2023. 5 +[31] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 4 +[32] Vaishakh Patil, Christos Sakaridis, Alexander Liniger, and Luc Van Gool. P3depth: Monocular depth estimation with a piecewise planarity prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1610-1621, 2022. 2 +[33] Xuebin Qin, Hang Dai, Xiaobin Hu, Deng-Ping Fan, Ling Shao, and Luc Van Gool. Highly accurate dichotomous image segmentation. In European Conference on Computer Vision, pages 38-56. Springer, 2022. 5 +[34] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 44(3):1623-1637, 2020. 1, 2, 5 +[35] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021. 1, 2, 5 +[36] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10912-10922, 2021. 5 +[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image + +synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3 +[38] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 4 +[39] Thomas Schops, Johannes L Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3260-3269, 2017. 5 +[40] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 5 +[41] Jaidev Shriram, Alex Trevithick, Lingjie Liu, and Ravi Ramamoorthi. Realmdreamer: Text-driven 3d scene generation with inpainting and depth diffusion. arXiv preprint arXiv:2404.07199, 2024. 1 +[42] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12, pages 746-760. Springer, 2012. 5 +[43] Zachary Teed and Jia Deng. Deepv2d: Video to depth with differentiable structure from motion. arXiv preprint arXiv:1812.04605, 2018. 2 +[44] Igor Vasiljevic, Nick Kolkin, Shanyi Zhang, Ruotian Luo, Haochen Wang, Falcon Z Dai, Andrea F Daniele, Mohammadreza Mostajabi, Steven Basart, Matthew R Walter, et al. Diode: A dense indoor and outdoor depth dataset. arXiv preprint arXiv:1908.00463, 2019. 5 +[45] Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. In Proceedings of the 25th international conference on Machine learning, pages 1096-1103, 2008. 2 +[46] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9065-9076, 2023. 1 +[47] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudolidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8445-8453, 2019. 1 +[48] Yiran Wang, Min Shi, Jiaqi Li, Zihao Huang, Zhiguo Cao, Jianming Zhang, Ke Xian, and Guosheng Lin. Neural video depth stabilizer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9466-9476, 2023. 2 + +[49] Guangkai Xu, Yongtao Ge, Mingyu Liu, Chengxiang Fan, Kangyang Xie, Zhiyue Zhao, Hao Chen, and Chunhua Shen. Diffusion models trained with large data are transferable visual models. arXiv preprint arXiv:2403.06090, 2024. 3, 4, 5 +[50] Guanglei Yang, Hao Tang, Mingli Ding, Nicu Sebe, and Elisa Ricci. Transformer-based attention networks for continuous pixel-wise prediction. In Proceedings of the IEEE/CVF International Conference on Computer vision, pages 16269-16279, 2021. 2 +[51] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 2, 5 +[52] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024. 1, 2, 4, 5 +[53] Xingyi Yang and Xinchao Wang. Diffusion model as representation learner. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18938-18949, 2023. 2, 3 +[54] Rajeev Yasarla, Hong Cai, Jisoo Jeong, Yunxiao Shi, Risheek Garrepalli, and Fatih Porikli. Mamo: Leveraging memory and attention for monocular video depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8754-8764, 2023. 2 +[55] Chongjie Ye, Lingteng Qiu, Xiaodong Gu, Qi Zuo, Yushuang Wu, Zilong Dong, Liefeng Bo, Yuliang Xiu, and Xiaoguang Han. Stablenormal: Reducing diffusion variance for stable and sharp normal. arXiv preprint arXiv:2406.16864, 2024. 3 +[56] Wei Yin, Xinlong Wang, Chunhua Shen, Yifan Liu, Zhi Tian, Songcen Xu, Changming Sun, and Dou Renyin. Diversedepth: Affine-invariant depth prediction using diverse data. arXiv preprint arXiv:2002.00569, 2020. 2, 5 +[57] Wei Yin, Jianming Zhang, Oliver Wang, Simon Niklaus, Long Mai, Simon Chen, and Chunhua Shen. Learning to recover 3d scene shape from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 204-213, 2021. +[58] Chi Zhang, Wei Yin, Billzb Wang, Gang Yu, Bin Fu, and Chunhua Shen. Hierarchical normalization for robust monocular depth estimation. Advances in Neural Information Processing Systems, 35:14128-14139, 2022. 5 +[59] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence. Advances in Neural Information Processing Systems, 36, 2024. 2, 3 +[60] Xiang Zhang, Bingxin Ke, Hayko Riemenschneider, Nando Metzger, Anton Obukhov, Markus Gross, Konrad Schindler, and Christopher Schroers. Betterdepth: Plug-and-play diffusion refiner for zero-shot monocular depth estimation. arXiv preprint arXiv:2407.17952, 2024. 3 +[61] Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing text-to-image diffu + +sion models for visual perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5729-5739, 2023. 2, 3 \ No newline at end of file diff --git a/2025/FiffDepth_ Feed-forward Transformation of Diffusion-Based Generators for Detailed Depth Estimation/images.zip b/2025/FiffDepth_ Feed-forward Transformation of Diffusion-Based Generators for Detailed Depth Estimation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0a9989ceb1d931acedc5b66f1bb036a73f98a543 --- /dev/null +++ b/2025/FiffDepth_ Feed-forward Transformation of Diffusion-Based Generators for Detailed Depth Estimation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e216fd5045618fa91ed514b6fe34c11c57dbb1af6c71a4c0aeaa782279799928 +size 1013825 diff --git a/2025/FiffDepth_ Feed-forward Transformation of Diffusion-Based Generators for Detailed Depth Estimation/layout.json b/2025/FiffDepth_ Feed-forward Transformation of Diffusion-Based Generators for Detailed Depth Estimation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7afed121df437b70e3920d62c5e05ea8c94a4637 --- /dev/null +++ b/2025/FiffDepth_ Feed-forward Transformation of Diffusion-Based Generators for Detailed Depth Estimation/layout.json @@ -0,0 +1,10188 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 72, + 102, + 539, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 102, + 539, + 140 + ], + "spans": [ + { + "bbox": [ + 72, + 102, + 539, + 140 + ], + "type": "text", + "content": "FiffDepth: Feed-forward Transformation of Diffusion-Based Generators for Detailed Depth Estimation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 223, + 162, + 386, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 162, + 386, + 176 + ], + "spans": [ + { + "bbox": [ + 223, + 162, + 386, + 176 + ], + "type": "text", + "content": "Yunpeng Bai Qixing Huang" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 221, + 176, + 386, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 176, + 386, + 190 + ], + "spans": [ + { + "bbox": [ + 221, + 176, + 386, + 190 + ], + "type": "text", + "content": "The University of Texas at Austin" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 80, + 230, + 187, + 293 + ], + "blocks": [ + { + "bbox": [ + 80, + 230, + 187, + 293 + ], + "lines": [ + { + "bbox": [ + 80, + 230, + 187, + 293 + ], + "spans": [ + { + "bbox": [ + 80, + 230, + 187, + 293 + ], + "type": "image", + "image_path": "2936be9c03d6170cca4e7ff8b69beca682c98a4a7ad8844ccbb4a5115999ca38.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 80, + 293, + 186, + 372 + ], + "blocks": [ + { + "bbox": [ + 80, + 293, + 186, + 372 + ], + "lines": [ + { + "bbox": [ + 80, + 293, + 186, + 372 + ], + "spans": [ + { + "bbox": [ + 80, + 293, + 186, + 372 + ], + "type": "image", + "image_path": "7fd8549e482b0256e8162492758759437c39e715d5599b4dee427bb8d00d32e0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 80, + 372, + 186, + 434 + ], + "blocks": [ + { + "bbox": [ + 80, + 372, + 186, + 434 + ], + "lines": [ + { + "bbox": [ + 80, + 372, + 186, + 434 + ], + "spans": [ + { + "bbox": [ + 80, + 372, + 186, + 434 + ], + "type": "image", + "image_path": "c584bded786f54295d81edc2f2e3b04cd8cd41ee241ee17ce1525f5e078a9833.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 109, + 436, + 159, + 447 + ], + "lines": [ + { + "bbox": [ + 109, + 436, + 159, + 447 + ], + "spans": [ + { + "bbox": [ + 109, + 436, + 159, + 447 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 194, + 231, + 301, + 292 + ], + "blocks": [ + { + "bbox": [ + 194, + 231, + 301, + 292 + ], + "lines": [ + { + "bbox": [ + 194, + 231, + 301, + 292 + ], + "spans": [ + { + "bbox": [ + 194, + 231, + 301, + 292 + ], + "type": "image", + "image_path": "9ccce071ad3f046138e947a5a88d9547699f897cb26bd5058e5142ed7da6a4e3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 194, + 292, + 301, + 372 + ], + "blocks": [ + { + "bbox": [ + 194, + 292, + 301, + 372 + ], + "lines": [ + { + "bbox": [ + 194, + 292, + 301, + 372 + ], + "spans": [ + { + "bbox": [ + 194, + 292, + 301, + 372 + ], + "type": "image", + "image_path": "1c2c1ec4c1b93b5076006b7e1758325bd37fec7f87565d5a5ab84fc04bbfa9f6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 194, + 372, + 299, + 434 + ], + "blocks": [ + { + "bbox": [ + 194, + 372, + 299, + 434 + ], + "lines": [ + { + "bbox": [ + 194, + 372, + 299, + 434 + ], + "spans": [ + { + "bbox": [ + 194, + 372, + 299, + 434 + ], + "type": "image", + "image_path": "594cfb007bb8a09bc0e1c526a58c1de3808211b0dad9057ae2fc7d99f7e82653.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 435, + 268, + 447 + ], + "lines": [ + { + "bbox": [ + 227, + 435, + 268, + 447 + ], + "spans": [ + { + "bbox": [ + 227, + 435, + 268, + 447 + ], + "type": "text", + "content": "FifDepth" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 308, + 231, + 414, + 292 + ], + "blocks": [ + { + "bbox": [ + 308, + 231, + 414, + 292 + ], + "lines": [ + { + "bbox": [ + 308, + 231, + 414, + 292 + ], + "spans": [ + { + "bbox": [ + 308, + 231, + 414, + 292 + ], + "type": "image", + "image_path": "ed2630d94aafe8f960a30cd44ed5a0f96f95d849255669835d466cd3248d7c46.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 308, + 292, + 414, + 371 + ], + "blocks": [ + { + "bbox": [ + 308, + 292, + 414, + 371 + ], + "lines": [ + { + "bbox": [ + 308, + 292, + 414, + 371 + ], + "spans": [ + { + "bbox": [ + 308, + 292, + 414, + 371 + ], + "type": "image", + "image_path": "4e83a5ccbcfc041e27e9dde36a21799f2082c2e84fe5f456404a6e12d54eef4d.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 308, + 372, + 411, + 433 + ], + "blocks": [ + { + "bbox": [ + 308, + 372, + 411, + 433 + ], + "lines": [ + { + "bbox": [ + 308, + 372, + 411, + 433 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 411, + 433 + ], + "type": "image", + "image_path": "7573848849c9e461db6024935d95e2c8d09a87903046bf246f59aff950a93011.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 435, + 400, + 447 + ], + "lines": [ + { + "bbox": [ + 322, + 435, + 400, + 447 + ], + "spans": [ + { + "bbox": [ + 322, + 435, + 400, + 447 + ], + "type": "text", + "content": "Depth Anything v2" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 451, + 555, + 474 + ], + "lines": [ + { + "bbox": [ + 55, + 451, + 555, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 555, + 474 + ], + "type": "text", + "content": "Figure 1. Compared to other methods, our model achieves more accurate details and better generalization in depth estimation. The final row shows the point cloud generated from the estimated depth results, and the corresponding depth map can be referenced in Figure 5." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 421, + 231, + 528, + 292 + ], + "blocks": [ + { + "bbox": [ + 421, + 231, + 528, + 292 + ], + "lines": [ + { + "bbox": [ + 421, + 231, + 528, + 292 + ], + "spans": [ + { + "bbox": [ + 421, + 231, + 528, + 292 + ], + "type": "image", + "image_path": "1d6c6030d5726e9c8b84718e5544235beab47812d087f2836fbb1c1629fea4db.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 421, + 293, + 528, + 372 + ], + "blocks": [ + { + "bbox": [ + 421, + 293, + 528, + 372 + ], + "lines": [ + { + "bbox": [ + 421, + 293, + 528, + 372 + ], + "spans": [ + { + "bbox": [ + 421, + 293, + 528, + 372 + ], + "type": "image", + "image_path": "a0f45a70f9ff381299db6c9bb83842889e3a7847178667cf63897b1d4b6bc9b8.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 421, + 372, + 523, + 433 + ], + "blocks": [ + { + "bbox": [ + 421, + 372, + 523, + 433 + ], + "lines": [ + { + "bbox": [ + 421, + 372, + 523, + 433 + ], + "spans": [ + { + "bbox": [ + 421, + 372, + 523, + 433 + ], + "type": "image", + "image_path": "18c80cee368663e70893ce5fef6bf4023d4c22bb42c7c4ea6836f27788de9e25.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 453, + 435, + 495, + 447 + ], + "lines": [ + { + "bbox": [ + 453, + 435, + 495, + 447 + ], + "spans": [ + { + "bbox": [ + 453, + 435, + 495, + 447 + ], + "type": "text", + "content": "Depth Pro" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 152, + 489, + 200, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 489, + 200, + 502 + ], + "spans": [ + { + "bbox": [ + 152, + 489, + 200, + 502 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 54, + 523, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 523, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 523, + 297, + 715 + ], + "type": "text", + "content": "Monocular Depth Estimation (MDE) is a fundamental 3D vision problem with numerous applications such as 3D scene reconstruction, autonomous navigation, and AI content creation. However, robust and generalizable MDE remains challenging due to limited real-world labeled data and distribution gaps between synthetic datasets and real data. Existing methods often struggle with real-world test data with low efficiency, reduced accuracy, and lack of detail. To address these issues, we propose an efficient MDE approach named FiftDepth. The key feature of FiftDepth is its use of diffusion priors. It transforms diffusion-based image generators into a feed-forward architecture for detailed depth estimation. FiftDepth preserves key generative features and integrates the strong generalization capabilities of models like DINOv2. Through benchmark evaluations, we demonstrate that FiftDepth achieves exceptional" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 490, + 556, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 490, + 556, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 490, + 556, + 550 + ], + "type": "text", + "content": "accuracy, stability, and fine-grained detail, offering significant improvements in MDE performance against state-of-the-art MDE approaches. The paper's source code is available here: https://yunpeng1998.github.io/FiffDepth/" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 314, + 574, + 395, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 574, + 395, + 586 + ], + "spans": [ + { + "bbox": [ + 314, + 574, + 395, + 586 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 594, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 594, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 594, + 555, + 715 + ], + "type": "text", + "content": "Monocular Depth Estimation (MDE) is a fundamental 3D vision problem with numerous applications in 3D scene reconstruction [46], autonomous navigation [47], and more recently in generating AI-based content [2, 41]. MDE has made significant progress in the era of deep learning [21, 34, 35, 52]. Neural networks trained from paired datasets of images and pixel depth exhibit encouraging results and often outperform non-deep learning based counterparts that build on monocular depth cues. Despite significant progress, fundamental challenges remain in efficiency," + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6023" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "accuracy, and generalization on diverse in-the-wild data. This is because 1) real depth datasets are usually noisy, and 2) while synthetic data can be used, there exist domain gaps between synthetic datasets and diverse in-the-wild data." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 122, + 295, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 122, + 295, + 278 + ], + "spans": [ + { + "bbox": [ + 55, + 122, + 295, + 278 + ], + "type": "text", + "content": "Specifically, current MDE research relies primarily on synthetic data due to its high-quality annotations and controlled environments. However, the scale and variety of synthetic datasets remain insufficient for comprehensive training. To address this, synthetic-to-real transfer techniques and the utilization of pre-trained models have emerged as viable solutions. Among pre-trained models, generative networks [37] preserve intricate image details more effectively than feed-forward networks (FFNs) like DINOv2 [31], thus holding greater promise for dense prediction models. However, generative models, while detail-rich, often fall short in synthetic-to-real transfer due to their limited generalization capabilities outside the training domain." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 281, + 295, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 295, + 448 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 295, + 448 + ], + "type": "text", + "content": "Previous studies [15, 21] have adopted pre-trained diffusion models, where the main idea is to directly finetuning pre-trained RGB image diffusion models into depth map diffusion generation models conditioned on images. However, this method may not be ideal, as dense prediction models require certainty over diversity. The introduction of any noise or uncertainty during the generation process by these methods is sub-optimal. In contrast, we observe that simply using the denoising diffusion module in a feedforward manner yields better and more stable results. This method capitalizes on extending the trajectories of image diffusion models into the depth domain, representing a significant advancement in both accuracy and efficiency for generative model-based depth estimation methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "type": "text", + "content": "Specifically, we optimize diffusion trajectories for MDE tasks. To enable the diffusion model to better retain certain detailed generative features when fine-tuned into an MDE model, we preserve the original generative training trajectory while training the model for depth prediction, aiming to maintain the detailed features of the original generative model as much as possible. Furthermore, recognizing the limitations of fintuned diffusion models in maintaining robustness in diverse real-world images—with inaccuracies in predicted depth occurring mainly in the low-frequency components—we leverage the strengths of a DINOv2 [31] based model, which excels at predicting accurate low-frequency depth despite its reduced fine detail. To address this issue, we use the diffusion model itself to learn a filter that refines its inaccurate predictions, producing low-frequency outputs with a detail level similar to that of DINOv2's predictions, thereby matching DINOv2's results and optimizing the low-frequency component of our output. This approach allows us to incorporate a large amount of real image data for training without sacrificing detail preservation of generative models, while simultaneously leveraging DINOv2's strong generalization capabilities to enhance" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 544, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 544, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 544, + 84 + ], + "type": "text", + "content": "the overall robustness and precision of our MDE models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 85, + 553, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 85, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 85, + 553, + 239 + ], + "type": "text", + "content": "In summary, the contributions of our work are as follows: 1) We propose an improved approach for transforming generative models into dense prediction models, specifically for depth prediction tasks, by leveraging diffusion model trajectories in a more stable, feed-forward manner; 2) We introduce a novel distillation method that transfers the robust generalization capabilities of models like DINOv2 to diffusion backbones; 3) Our method demonstrates higher stability, accuracy, and efficiency in depth estimation compared to other approaches based on generative models; 4) Compared to other FFN models, our approach achieves more detailed prediction results, marking a significant advancement in the field of MDE." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 253, + 405, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 253, + 405, + 264 + ], + "spans": [ + { + "bbox": [ + 314, + 253, + 405, + 264 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 273, + 419, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 273, + 419, + 285 + ], + "spans": [ + { + "bbox": [ + 314, + 273, + 419, + 285 + ], + "type": "text", + "content": "2.1. Depth Estimation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 291, + 553, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 291, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 553, + 447 + ], + "type": "text", + "content": "Depth estimation has always been a widely researched topic, with numerous studies conducted in the past including single-image [1, 13, 14, 24, 28, 32, 50] and video depth estimation [10, 23, 43, 48, 54]. Early efforts in image depth estimation, such as DIW [8] and OASIS [9], focused on predicting relative (ordinal) depth. Subsequent approaches, including MegaDepth [27] and DiverseDepth [56], used extensive collections of photographs from the Internet to develop models that adapt to unseen data, while MiDaS [34] improved generalization by incorporating a diverse range of datasets during training. Recent advances, including DPT [35] and Omnidata [12], adopted transformer-based architectures to improve depth estimation performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 447, + 554, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 554, + 567 + ], + "type": "text", + "content": "However, due to limitations in models and data, these methods exhibit very limited generalization in various open-world scenes. Recently, some efforts [15, 21, 51, 52] in image depth estimation have made open-world depth estimation feasible by leveraging the power of vast amounts of unlabeled image data and the capabilities of pre-trained generative models. Although these methods have achieved remarkable progress in the field of depth estimation, challenges such as low efficiency, limited generalization, and insufficient detail preservation remain." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 575, + 544, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 575, + 544, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 575, + 544, + 588 + ], + "type": "text", + "content": "2.2. Diffusion Models as Representation Learner" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 594, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 553, + 713 + ], + "type": "text", + "content": "The diffusion model [3, 37] training process strongly resembles that of denoising autoencoders (DAE) [4, 19, 45], as both are designed to recover clean images from noise-corrupted inputs. Recent studies have shown that semantic image representations learned from diffusion models can be effectively used for various downstream recognition tasks, such as correspondence [59], semantic segmentation [61], and keypoint detection [53]. Notably, the features extracted from diffusion models tend to preserve more intricate details, which has prompted the adoption of pre" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6024" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 251, + 307 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 251, + 307 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 251, + 307 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 251, + 307 + ], + "type": "image", + "image_path": "4fd3a713f38948f876a77c37a1f7f9105f2631393f5d035b62e6bba210784ae9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 319, + 555, + 366 + ], + "lines": [ + { + "bbox": [ + 54, + 319, + 555, + 366 + ], + "spans": [ + { + "bbox": [ + 54, + 319, + 555, + 366 + ], + "type": "text", + "content": "Figure 2. Overview of the proposed method. To simplify the representation, all the images we used above correspond to the respective latents. We transform the pre-trained diffusion model into a feed-forward approach for depth prediction, using only the result at " + }, + { + "bbox": [ + 54, + 319, + 555, + 366 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 54, + 319, + 555, + 366 + ], + "type": "text", + "content": " as the output during inference. During training, at " + }, + { + "bbox": [ + 54, + 319, + 555, + 366 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 54, + 319, + 555, + 366 + ], + "type": "text", + "content": ", we use synthetic data to ensure detailed results, while at " + }, + { + "bbox": [ + 54, + 319, + 555, + 366 + ], + "type": "inline_equation", + "content": "t = -1" + }, + { + "bbox": [ + 54, + 319, + 555, + 366 + ], + "type": "text", + "content": ", we leverage pseudo-labels generated by DINOv2 for supervision." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 253, + 70, + 553, + 308 + ], + "blocks": [ + { + "bbox": [ + 253, + 70, + 553, + 308 + ], + "lines": [ + { + "bbox": [ + 253, + 70, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 253, + 70, + 553, + 308 + ], + "type": "image", + "image_path": "00044d553bde24e5bd68df7ff62ded032f3e934065480a77c66d061cade33c36.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 385, + 297, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 385, + 297, + 553 + ], + "spans": [ + { + "bbox": [ + 54, + 385, + 297, + 553 + ], + "type": "text", + "content": "trained diffusion models for dense prediction tasks. Recent studies, including Marigold [21] and GeoWizard [15], rely on the standard diffusion framework and pre-trained parameters to perform dense prediction tasks. Emerging approaches [18, 49, 55] attempt to bypass the stochastic phase of diffusion models by employing deterministic frameworks. However, these adaptations lack deeper exploration of the model's potential, often leading to suboptimal performance and the need for additional post-processing to refine results. Moreover, diffusion-based techniques generally exhibit limited generalization capabilities. For instance, BetterDepth [60] incorporates external depth priors as inputs but remains a stochastic framework and heavily depends on the quality of other models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 567, + 112, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 112, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 112, + 578 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 587, + 296, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 296, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 296, + 611 + ], + "type": "text", + "content": "3.1. Overview: Feed-forward Transformation of Diffusion Models" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": "Fundamentally, generative models construct mappings between a latent space and the ambient data space. These mappings align closely with the needs of depth estimation and other visual recognition tasks, where precise mappings from image data to the corresponding labels are essential. In these tasks, the scarcity of labeled data often limits the precision of the trained models. In the context of learning from a collection of unlabeled data instances, advanced gen" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 385, + 555, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 385, + 555, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 555, + 447 + ], + "type": "text", + "content": "erative models, trained on massive datasets, are capable of learning robust mappings. They exhibit great promise for transferring knowledge to other visual prediction applications [53, 59, 61]. Our work builds on this approach, further exploring its application in depth estimation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 453, + 556, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 556, + 587 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 556, + 587 + ], + "type": "text", + "content": "Specifically, we finetune a well-trained Stable Diffusion (SD) [37] model to construct our model. A diffusion process constructs multiple intermediate states by progressively adding noise to the data " + }, + { + "bbox": [ + 313, + 453, + 556, + 587 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 313, + 453, + 556, + 587 + ], + "type": "text", + "content": ", defined as " + }, + { + "bbox": [ + 313, + 453, + 556, + 587 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t = \\sqrt{\\bar{\\alpha}_t}\\mathbf{x}_0 + \\sqrt{1 - \\bar{\\alpha}_t}\\epsilon" + }, + { + "bbox": [ + 313, + 453, + 556, + 587 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 453, + 556, + 587 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 313, + 453, + 556, + 587 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 453, + 556, + 587 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t \\coloneqq \\prod_{s=1}^{t} 1 - \\beta_s" + }, + { + "bbox": [ + 313, + 453, + 556, + 587 + ], + "type": "text", + "content": " with noise schedule " + }, + { + "bbox": [ + 313, + 453, + 556, + 587 + ], + "type": "inline_equation", + "content": "\\{\\beta_1, \\ldots, \\beta_T\\}" + }, + { + "bbox": [ + 313, + 453, + 556, + 587 + ], + "type": "text", + "content": ". Then, diffusion generative model gradually learns the mapping between two distributions by denoising each step, denoted as " + }, + { + "bbox": [ + 313, + 453, + 556, + 587 + ], + "type": "inline_equation", + "content": "x_T \\to x_{T-1} \\to \\dots \\to x_0" + }, + { + "bbox": [ + 313, + 453, + 556, + 587 + ], + "type": "text", + "content": ". The diffusion model is primarily a denoising model " + }, + { + "bbox": [ + 313, + 453, + 556, + 587 + ], + "type": "inline_equation", + "content": "\\epsilon_\\theta" + }, + { + "bbox": [ + 313, + 453, + 556, + 587 + ], + "type": "text", + "content": " that follows a loss function " + }, + { + "bbox": [ + 313, + 453, + 556, + 587 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{simple}} = \\mathbb{E}_{x_0,\\epsilon,t} \\left[ \\| \\epsilon - \\epsilon_\\theta(x_t,t) \\|^2 \\right]" + }, + { + "bbox": [ + 313, + 453, + 556, + 587 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "content": "In depth estimation, the mapping from visual images to depth labels should ideally be deterministic. By leveraging a robust, pre-trained mapping within the generative model, there is no need to decompose the image-label mapping into multiple steps during training. Instead, we can extend the mapping trajectories of the existing diffusion model directly into the depth domain by adapting the learned diffusion process to act as a deterministic one-step feed-forward network. Since our MDE model is built upon the diffusion trajectory for its extension, we set the time step input as" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6025" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 183, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 183, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 183, + 83 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 55, + 72, + 183, + 83 + ], + "type": "text", + "content": " in this feed-forward step." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 95, + 295, + 107 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 95, + 295, + 107 + ], + "spans": [ + { + "bbox": [ + 132, + 95, + 295, + 107 + ], + "type": "interline_equation", + "content": "\\mathbf {d} _ {0} = \\hat {\\boldsymbol {\\epsilon}} _ {\\theta} \\left(\\mathbf {x} _ {0}, t = 0\\right). \\tag {1}", + "image_path": "c1ccfb3f36e2a325a06b16ae91ba97bdc5d00b28a22a54722e3c0713a093f24d.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 118, + 296, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 118, + 296, + 320 + ], + "spans": [ + { + "bbox": [ + 55, + 118, + 296, + 320 + ], + "type": "text", + "content": "Since we use diffusion model's parameters to construct our network, we also use " + }, + { + "bbox": [ + 55, + 118, + 296, + 320 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_{\\theta}" + }, + { + "bbox": [ + 55, + 118, + 296, + 320 + ], + "type": "text", + "content": " to represent our network model here. The input " + }, + { + "bbox": [ + 55, + 118, + 296, + 320 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 55, + 118, + 296, + 320 + ], + "type": "text", + "content": " is the latent representation of RGB image, and " + }, + { + "bbox": [ + 55, + 118, + 296, + 320 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_0" + }, + { + "bbox": [ + 55, + 118, + 296, + 320 + ], + "type": "text", + "content": " is the latent representation of \"depth image,\" following Marigold's encoding approach. The depth latent " + }, + { + "bbox": [ + 55, + 118, + 296, + 320 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_0" + }, + { + "bbox": [ + 55, + 118, + 296, + 320 + ], + "type": "text", + "content": " can be reconstructed into a depth map using the VAE of SD with negligible error. While there have been preliminary attempts [18, 49] to use similar approaches, they remain in a nascent stage and lack the precision, robustness, and richness in detail needed for effective depth estimation. In the following sections, we will delve into each step of our method, detailing how we adapt the diffusion model to enhance accuracy and detail in depth estimation tasks, such as the key technical contributions of preserving diffusion trajectories and improving synthetic-to-real robustness, as discussed in Section 3.2 and Section 3.3, respectively. The overall workflow of the method is illustrated in Figure 2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 329, + 222, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 329, + 222, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 329, + 222, + 342 + ], + "type": "text", + "content": "3.2. Keeping Diffusion Trajectories" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 346, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 296, + 491 + ], + "type": "text", + "content": "Since our approach leverages the trajectory of the diffusion model, it is crucial to prevent degradation of this trajectory during training. To achieve this, when fine-tuning the diffusion model to transition it into a feed-forward depth estimator, we simultaneously maintain the feed-forward step along with the preceding denoising training steps from the original diffusion model. While this trajectory was initially developed for image generation, directly applying it as-is does not facilitate an optimal transition to the depth domain. Thus, instead of predicting purely image-based latents, we modify the target latent to be a blend of image and depth representations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 500, + 294, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 500, + 294, + 529 + ], + "spans": [ + { + "bbox": [ + 89, + 500, + 294, + 529 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {b} _ {0} = \\gamma \\mathbf {x} _ {0} + (1 - \\gamma) \\mathbf {d} _ {0}, \\\\ \\mathbf {b} _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {b} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\boldsymbol {\\epsilon}, t \\in \\{1, \\dots , T \\}. \\tag {2} \\\\ \\end{array}", + "image_path": "72a100c2094bb963cbccb56c329f5cd092920fed58c3e5e8ad18093867f4bd09.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 536, + 296, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 536, + 296, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 296, + 583 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 536, + 296, + 583 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_0" + }, + { + "bbox": [ + 55, + 536, + 296, + 583 + ], + "type": "text", + "content": " represents the blended latent. " + }, + { + "bbox": [ + 55, + 536, + 296, + 583 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 536, + 296, + 583 + ], + "type": "text", + "content": " controls the balance between the image and depth latents. In this process, we also use v-prediction re-parameterization approach [38] to define the training objective:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 92, + 594, + 294, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 594, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 92, + 594, + 294, + 624 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {v} _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {\\epsilon} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\mathbf {b} _ {0}, \\\\ L _ {k} = \\left\\| \\mathbf {v} _ {t} - \\hat {\\boldsymbol {\\epsilon}} _ {\\theta} \\left(\\mathbf {b} _ {t}, t\\right) \\right\\| _ {2} ^ {2}, t \\in \\{1, \\dots , T \\}. \\tag {3} \\\\ \\end{array}", + "image_path": "867502a775258c9d39c2507857718022aa2a35d29dc42de3ee8da0650cd20d95.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "Intuitively, this approach forces the diffusion model to preserve the shared features between the image generation task and the depth estimation task, which are captured in the blended training target. Therefore, it allows the diffusion model to adapt more naturally to depth estimation while retaining essential generative features. Consequently, during fine-tuning, the model maintains features that enhance the" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "text", + "content": "accuracy and detail of depth predictions. This part is used exclusively during training. At inference time, our model functions as a fully deterministic framework." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 118, + 468, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 118, + 468, + 129 + ], + "spans": [ + { + "bbox": [ + 313, + 118, + 468, + 129 + ], + "type": "text", + "content": "3.3. Learnable Filter Distillation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 137, + 555, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 137, + 555, + 279 + ], + "spans": [ + { + "bbox": [ + 313, + 137, + 555, + 279 + ], + "type": "text", + "content": "Following previous methods, the above training process only uses synthetic data because it provides high-quality depth Ground Truth. However, this reliance limits both Marigold [21] and our approach, as SD-based MDE models trained solely on synthetic datasets often struggle to generalize well to in-the-wild data. Nevertheless, in this case, the predictions produced by the SD-based MDE model still retain the necessary details—precisely those details we expect to preserve in the final output. Therefore, enhancing the model's robustness essentially means improving the accuracy of the low-frequency components in the model's output on real image." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 281, + 556, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 556, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 556, + 460 + ], + "type": "text", + "content": "Inspired by prior work [52], we observe that DINOv2 [31], trained on synthetic data, can generalize effectively to real-world images. However, its depth estimates often lack the necessary fine details—in other words, it can accurately predict the low-frequency depth components for massive real images but misses the high-frequency details. This characteristic aligns well with our needs, so we attempt to leverage the abundance of labels generated by the DINOv2 model to enhance our model's robustness. A straightforward method, as used in Depth Anything v2 [52], is to use a fintuned DINOv2-G depth model to generate pseudo labels, expanding the training set. However, since DINOv2's depth predictions lack fine detail, directly incorporating these pseudo labels on the feed-forward output " + }, + { + "bbox": [ + 313, + 281, + 556, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_0" + }, + { + "bbox": [ + 313, + 281, + 556, + 460 + ], + "type": "text", + "content": " risks ignoring the detailed features inherent to our model." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 461, + 556, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 556, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 556, + 641 + ], + "type": "text", + "content": "To address this, we propose learning a filter, denoted as " + }, + { + "bbox": [ + 313, + 461, + 556, + 641 + ], + "type": "inline_equation", + "content": "F(d_0)" + }, + { + "bbox": [ + 313, + 461, + 556, + 641 + ], + "type": "text", + "content": " that processes our output to remove high-frequency details (e.g., fine details). This filtering produces results at a detail level similar to that of DINOv2, allowing us to focus supervision solely on the less accurate low-frequency components. The filter " + }, + { + "bbox": [ + 313, + 461, + 556, + 641 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 461, + 556, + 641 + ], + "type": "text", + "content": " is obtained through learning. In fact, we observed that the fine-tuned SD model itself already serves as an effective filter learner because the diffusion model is inherently designed to model the subtle differences across different time steps. Therefore, we directly apply an additional SD step on the feed-forward output " + }, + { + "bbox": [ + 313, + 461, + 556, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_0" + }, + { + "bbox": [ + 313, + 461, + 556, + 641 + ], + "type": "text", + "content": " to simulate this filter. To maintain consistency with the inputs and concepts used in the diffusion model, we represent this filter " + }, + { + "bbox": [ + 313, + 461, + 556, + 641 + ], + "type": "inline_equation", + "content": "F: t_0 \\rightarrow t_{-1}" + }, + { + "bbox": [ + 313, + 461, + 556, + 641 + ], + "type": "text", + "content": " as the process that transforms the output from " + }, + { + "bbox": [ + 313, + 461, + 556, + 641 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 313, + 461, + 556, + 641 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 461, + 556, + 641 + ], + "type": "inline_equation", + "content": "t_{-1}" + }, + { + "bbox": [ + 313, + 461, + 556, + 641 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 383, + 656, + 553, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 656, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 383, + 656, + 553, + 668 + ], + "type": "interline_equation", + "content": "\\mathbf {d} _ {- 1} = \\hat {\\mathbf {e}} _ {\\theta} \\left(\\mathbf {d} _ {0}, t = - 1\\right). \\tag {4}", + "image_path": "10dac6857d9842e5b915436d9f814fc708abf1687514e2a885b6134060136966.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "At this stage, we can use the labels predicted by DINOv2 to supervise " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{-1}" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": ", allowing us to transfer DINOv2's robustness without interfering with the detailed features in " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_0" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": ". In" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6026" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 74, + 296, + 203 + ], + "blocks": [ + { + "bbox": [ + 56, + 74, + 296, + 203 + ], + "lines": [ + { + "bbox": [ + 56, + 74, + 296, + 203 + ], + "spans": [ + { + "bbox": [ + 56, + 74, + 296, + 203 + ], + "type": "image", + "image_path": "ad53793f8cce5cd1662bd934afa7d02c492e3471a6e3070ae8af51887fcebe03.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 216, + 296, + 262 + ], + "lines": [ + { + "bbox": [ + 55, + 216, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 296, + 262 + ], + "type": "text", + "content": "Figure 3. Filter learning. We use a learnable filter to map our results to detail levels similar to DINOv2's, matching its outputs and thereby transferring DINOv2's generalization capabilities to our model without compromising our inherent details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 278, + 295, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 278, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 295, + 304 + ], + "type": "text", + "content": "this process, we use real-world image data for " + }, + { + "bbox": [ + 55, + 278, + 295, + 304 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 55, + 278, + 295, + 304 + ], + "type": "text", + "content": ". Figure 3 illustrates the effects of Eq. (4)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 308, + 151, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 308, + 151, + 322 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 151, + 322 + ], + "type": "text", + "content": "3.4. Final Objective" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 326, + 296, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 326, + 296, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 296, + 374 + ], + "type": "text", + "content": "For " + }, + { + "bbox": [ + 55, + 326, + 296, + 374 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_0" + }, + { + "bbox": [ + 55, + 326, + 296, + 374 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 326, + 296, + 374 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{-1}" + }, + { + "bbox": [ + 55, + 326, + 296, + 374 + ], + "type": "text", + "content": ", we follow the method of MiDaS [34] using MAE loss " + }, + { + "bbox": [ + 55, + 326, + 296, + 374 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{MAE}}" + }, + { + "bbox": [ + 55, + 326, + 296, + 374 + ], + "type": "text", + "content": " and gradient matching loss " + }, + { + "bbox": [ + 55, + 326, + 296, + 374 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{GM}}" + }, + { + "bbox": [ + 55, + 326, + 296, + 374 + ], + "type": "text", + "content": " as depth loss. However, unlike the original approach, we apply these losses in the latent space of Stable Diffusion instead:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 383, + 295, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 383, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 104, + 383, + 295, + 415 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {M A E}} \\left(\\mathbf {d}, \\mathbf {d} ^ {*}\\right) = \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\left| \\mathbf {d} _ {i} - \\mathbf {d} _ {i} ^ {*} \\right|, \\tag {5}", + "image_path": "231fdbff2b80d4a7fb138cbfd4850142373a880116408b31b67f36383b660c06.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 418, + 296, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 296, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 296, + 455 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 418, + 296, + 455 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 55, + 418, + 296, + 455 + ], + "type": "text", + "content": " represents the ground truth latent, and " + }, + { + "bbox": [ + 55, + 418, + 296, + 455 + ], + "type": "inline_equation", + "content": "\\mathbf{d}^*" + }, + { + "bbox": [ + 55, + 418, + 296, + 455 + ], + "type": "text", + "content": " is the model's predicted value. " + }, + { + "bbox": [ + 55, + 418, + 296, + 455 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 418, + 296, + 455 + ], + "type": "text", + "content": " denotes the number of pixels in the depth latent." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 85, + 468, + 295, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 468, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 85, + 468, + 295, + 502 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {G M}} \\left(\\mathbf {d}, \\mathbf {d} ^ {*}\\right) = \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\left(| \\nabla_ {x} R _ {i} | + | \\nabla_ {y} R _ {i} |\\right), \\tag {6}", + "image_path": "383c38289580d17c6755001dda2580261a35e781ea374b5dba843522c7323478.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 506, + 295, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 295, + 532 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 506, + 295, + 532 + ], + "type": "inline_equation", + "content": "R_{i} = \\mathbf{d}_{i} - \\mathbf{d}_{i}^{*}" + }, + { + "bbox": [ + 55, + 506, + 295, + 532 + ], + "type": "text", + "content": ". Hence, the final objective is expressed as a weighted sum of the losses " + }, + { + "bbox": [ + 55, + 506, + 295, + 532 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{MAE}}" + }, + { + "bbox": [ + 55, + 506, + 295, + 532 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 506, + 295, + 532 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{GM}}" + }, + { + "bbox": [ + 55, + 506, + 295, + 532 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 506, + 295, + 532 + ], + "type": "inline_equation", + "content": "L_{k}" + }, + { + "bbox": [ + 55, + 506, + 295, + 532 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 72, + 546, + 295, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 546, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 72, + 546, + 295, + 590 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {\\text {f i n a l}} = \\sum_ {t \\in \\{- 1, 0 \\}} \\left(\\lambda_ {\\mathrm {M A E}} L _ {\\mathrm {M A E}} \\left(\\mathbf {d} _ {t}, \\mathbf {d} _ {t} ^ {*}\\right) \\right. \\\\ + \\lambda_ {\\mathrm {G M}} L _ {\\mathrm {G M}} \\left(\\mathbf {d} _ {t}, \\mathbf {d} _ {t} ^ {*}\\right)) + \\lambda_ {k} L _ {k}. \\tag {7} \\\\ \\end{array}", + "image_path": "780269963a05c13f2902875c4c61b07e094dfb779adb1c023493eb45f7d8c31b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 595, + 296, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 595, + 296, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 595, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{MAE}}" + }, + { + "bbox": [ + 55, + 595, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{GM}}" + }, + { + "bbox": [ + 55, + 595, + 296, + 643 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 595, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\lambda_{k}" + }, + { + "bbox": [ + 55, + 595, + 296, + 643 + ], + "type": "text", + "content": " are the weighting factors for their respective loss term. " + }, + { + "bbox": [ + 55, + 595, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_0^*" + }, + { + "bbox": [ + 55, + 595, + 296, + 643 + ], + "type": "text", + "content": " represents the ground truth (GT) from the synthetic dataset, and " + }, + { + "bbox": [ + 55, + 595, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{-1}^{*}" + }, + { + "bbox": [ + 55, + 595, + 296, + 643 + ], + "type": "text", + "content": " represents the pseudo label generated by DINOv2 for real images." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 652, + 137, + 666 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 652, + 137, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 652, + 137, + 666 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 672, + 188, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 188, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 188, + 685 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "During the training process, we preserve the diffusion trajectory while following the original DDPM noise sched" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": "uler [20] using 1000 diffusion steps. To better leverage pre-trained models, we use the Depth Anything V2-Large model as the DINO v2 model for supervision. Our training dataset consists of two parts. For training at " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": " and during trajectory retention, we follow previous approaches and use two synthetic datasets, Hypersim [36] and Virtual KITTI [7], which cover both indoor and outdoor scenes, with a total of 74K images. For training at " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "t = -1" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": ", we use real-world data from the LAION-Art dataset, a subset of LAION-5B [40] containing 8 million samples. However, we observed that training with only 0.2 million samples was sufficient. Synthetic data and real data each account for half of each batch. The parameters are set as follows: " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "\\gamma = 0.5" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{MAE}} = 1" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{GM}} = 0.5" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "\\lambda_{k} = 0.2" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 251, + 395, + 263 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 395, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 395, + 263 + ], + "type": "text", + "content": "4.2. Comparison" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 269, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 269, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 269, + 555, + 472 + ], + "type": "text", + "content": "Zero-shot affine-invariant depth. For the evaluation of affine-invariant depth, we use the same datasets and evaluation protocol as Marigold. These datasets include NYUv2 [42], ScanNet [11], KITTI [16], ETH3D [39], and DIODE [44]. We compared FiftDepth with 14 methods that produce affine-invariant depth maps/disparities, all claiming zero-shot generalization capabilities. These include the earlier methods [12, 34, 35, 56-58], as well as the more recent ones [15, 17, 18, 21, 29, 49, 51, 52]. As shown in Table 1, FiftDepth achieves the best or state-of-the-art comparable results in most test scenarios. For visualization results, please refer to Figure 4. Our method not only accurately predicts the relative depth relationships but also excels in identifying and predicting depth for very fine objects. We also evaluate our method on the DA-2K introduced by Depth Anything v2. On this dataset, our method also performs comparably to Depth Anything v2." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 473, + 554, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 554, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 554, + 593 + ], + "type": "text", + "content": "Compared to methods like Depth Anything, which rely on massive training datasets, our model achieves comparable generalization to DAv2 while being trained on only a small amount of real data. To validate the generalization capability of our method, we present some test examples from special scenarios in Figure 5, including games, artworks, AI-generated content, and movies. Our method shows comparable generalization to DAv2 while preserving more details. In contrast, other methods' results are unsatisfactory in both generalization and detail preservation." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "Zero-shot boundaries. To further demonstrate the accuracy of our method in predicting fine structures, we also employ the Zero-shot Boundaries Metric introduced in the recent work Depth Pro [5] to evaluate boundary sharpness. Following Depth Pro, we compute the depth average boundary F1 score for datasets with ground truth and the boundary recall (R) for datasets with matting or segmentation annotations. The former datasets include Sintel [6], Spring [30], and iBims [22], while the latter include AM-2k [26], P3M-10k [25], and DIS-5k [33]. For details on the boundaries" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6027" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 541, + 290 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 541, + 290 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 541, + 290 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 541, + 290 + ], + "type": "table", + "html": "
MethodTraining DataNYUv2KITTIETH3DScanNetDIODE-FullDA-2K
AbsRel ↓δ1 ↑AbsRel ↓δ1 ↑AbsRel ↓δ1 ↑AbsRel ↓δ1 ↑AbsRel ↓δ1 ↑Acc (%)
DiverseDepth320K11.787.519.070.422.869.410.988.237.663.179.3
MiDaS2M11.188.523.663.018.475.212.184.633.271.580.6
LeReS354K9.091.614.978.417.177.79.191.727.176.681.1
Omnidata v212.2M7.494.514.983.516.677.87.593.633.974.276.8
HDN300K6.994.811.586.712.183.38.093.924.678.085.7
DPT1.4M9.890.310.090.17.894.68.293.418.275.883.2
Marigold74K*5.596.49.991.66.496.06.495.130.877.386.8
e2e-ft74K*5.296.69.691.96.495.95.896.230.277.983.6
DepthFM74K*6.595.68.393.47.895.96.894.924.574.185.8
GenPercept74K*5.696.013.084.27.095.66.296.130.777.685.1
Lotus-D59K*5.396.78.192.86.595.35.896.329.978.186.8
Lotus-G59K*5.496.68.587.76.296.16.096.029.478.586.2
GeoWizard280K*5.296.69.792.16.496.16.195.329.779.288.1
DepthAnything v1-L62.6M*4.398.17.694.712.788.24.298.027.775.988.5
DepthAnything v2-L62.6M*4.597.97.494.613.186.54.297.826.275.497.1
FiffDepth (Ours)274K*4.497.87.393.57.197.24.297.923.978.197.1
", + "image_path": "99706824d0fa13b46374821df693e32d49b516cec9a11f7cc101d7f132978c97.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 298, + 555, + 331 + ], + "lines": [ + { + "bbox": [ + 55, + 298, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 298, + 555, + 331 + ], + "type": "text", + "content": "Table 1. Quantitative comparison with other affine-invariant depth estimators on several zero-shot benchmarks. We use AbsRel (absolute relative error: " + }, + { + "bbox": [ + 55, + 298, + 555, + 331 + ], + "type": "inline_equation", + "content": "|d^{*} - d| / d" + }, + { + "bbox": [ + 55, + 298, + 555, + 331 + ], + "type": "text", + "content": ") and " + }, + { + "bbox": [ + 55, + 298, + 555, + 331 + ], + "type": "inline_equation", + "content": "\\delta_{1}" + }, + { + "bbox": [ + 55, + 298, + 555, + 331 + ], + "type": "text", + "content": " (percentage of max " + }, + { + "bbox": [ + 55, + 298, + 555, + 331 + ], + "type": "inline_equation", + "content": "(d^{*} / d, d / d^{*}) < 1.25" + }, + { + "bbox": [ + 55, + 298, + 555, + 331 + ], + "type": "text", + "content": "). All metrics are reported as percentages; bold numbers are the best, underscored second best. Methods marked with an asterisk (*) utilize pre-trained models." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 59, + 338, + 556, + 674 + ], + "blocks": [ + { + "bbox": [ + 59, + 338, + 556, + 674 + ], + "lines": [ + { + "bbox": [ + 59, + 338, + 556, + 674 + ], + "spans": [ + { + "bbox": [ + 59, + 338, + 556, + 674 + ], + "type": "image", + "image_path": "b8e003d99145206d23b58668055b5a806445f0da1433bd367aba91c34e6f60f8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 687, + 555, + 709 + ], + "lines": [ + { + "bbox": [ + 55, + 687, + 555, + 709 + ], + "spans": [ + { + "bbox": [ + 55, + 687, + 555, + 709 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison across different datasets. Our method is capable of predicting the depth of various fine objects, such as lampposts, railings, and chair legs." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6028" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 69, + 211, + 144 + ], + "blocks": [ + { + "bbox": [ + 61, + 69, + 211, + 144 + ], + "lines": [ + { + "bbox": [ + 61, + 69, + 211, + 144 + ], + "spans": [ + { + "bbox": [ + 61, + 69, + 211, + 144 + ], + "type": "image", + "image_path": "f06461048ea5d2948f3a0da1823c05df844e90024b3f8b1447475d944e66f98b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 224, + 69, + 321, + 144 + ], + "blocks": [ + { + "bbox": [ + 224, + 69, + 321, + 144 + ], + "lines": [ + { + "bbox": [ + 224, + 69, + 321, + 144 + ], + "spans": [ + { + "bbox": [ + 224, + 69, + 321, + 144 + ], + "type": "image", + "image_path": "f02883b984eab1ab1cb4036f2059aa36d41cafbb6b7321c5a6a585f0566cd21b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 337, + 70, + 467, + 144 + ], + "blocks": [ + { + "bbox": [ + 337, + 70, + 467, + 144 + ], + "lines": [ + { + "bbox": [ + 337, + 70, + 467, + 144 + ], + "spans": [ + { + "bbox": [ + 337, + 70, + 467, + 144 + ], + "type": "image", + "image_path": "d93e4784913ffcee796906c973888065ca5e137f5f8fbe87a26c8b911cded3ec.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 482, + 70, + 553, + 144 + ], + "blocks": [ + { + "bbox": [ + 482, + 70, + 553, + 144 + ], + "lines": [ + { + "bbox": [ + 482, + 70, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 482, + 70, + 553, + 144 + ], + "type": "image", + "image_path": "d37fb4dc68b4c71d09dd3dd09929f52b4286db9a2a81d69cd0538ac7417d0a99.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 61, + 146, + 209, + 223 + ], + "blocks": [ + { + "bbox": [ + 61, + 146, + 209, + 223 + ], + "lines": [ + { + "bbox": [ + 61, + 146, + 209, + 223 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 209, + 223 + ], + "type": "image", + "image_path": "e68d29acd39c2a95b7584686916156b059ed8d0be4d09b2991edf0330ced5ace.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 225, + 148, + 321, + 222 + ], + "blocks": [ + { + "bbox": [ + 225, + 148, + 321, + 222 + ], + "lines": [ + { + "bbox": [ + 225, + 148, + 321, + 222 + ], + "spans": [ + { + "bbox": [ + 225, + 148, + 321, + 222 + ], + "type": "image", + "image_path": "9ad3aabcad67d77ab6c73451084c127d247dd955ef0b3ddf5c5c0b41766f2073.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 337, + 148, + 467, + 222 + ], + "blocks": [ + { + "bbox": [ + 337, + 148, + 467, + 222 + ], + "lines": [ + { + "bbox": [ + 337, + 148, + 467, + 222 + ], + "spans": [ + { + "bbox": [ + 337, + 148, + 467, + 222 + ], + "type": "image", + "image_path": "e55bd8ffe43077539fca55c09cda96734f9c49e5c1c106592d0e106038943c6b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 482, + 148, + 553, + 222 + ], + "blocks": [ + { + "bbox": [ + 482, + 148, + 553, + 222 + ], + "lines": [ + { + "bbox": [ + 482, + 148, + 553, + 222 + ], + "spans": [ + { + "bbox": [ + 482, + 148, + 553, + 222 + ], + "type": "image", + "image_path": "96f7b1e29b1b75219063e14fdb4e1f11bbd55f6e733bcac85d1b9398ac627601.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 61, + 225, + 209, + 300 + ], + "blocks": [ + { + "bbox": [ + 61, + 225, + 209, + 300 + ], + "lines": [ + { + "bbox": [ + 61, + 225, + 209, + 300 + ], + "spans": [ + { + "bbox": [ + 61, + 225, + 209, + 300 + ], + "type": "image", + "image_path": "8a32d8e070a3cec2dda0c6d6bf92112422970d2e45c102ccc7b1831a14aeb023.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 225, + 225, + 321, + 300 + ], + "blocks": [ + { + "bbox": [ + 225, + 225, + 321, + 300 + ], + "lines": [ + { + "bbox": [ + 225, + 225, + 321, + 300 + ], + "spans": [ + { + "bbox": [ + 225, + 225, + 321, + 300 + ], + "type": "image", + "image_path": "7cbfff5359cce5480dda847cc364eaf4374e459e6605e1f14a42a1e38deeaa9d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 337, + 225, + 467, + 300 + ], + "blocks": [ + { + "bbox": [ + 337, + 225, + 467, + 300 + ], + "lines": [ + { + "bbox": [ + 337, + 225, + 467, + 300 + ], + "spans": [ + { + "bbox": [ + 337, + 225, + 467, + 300 + ], + "type": "image", + "image_path": "f15d998a2f674f41ad93b904533e334c6f8fa9905da0f9c431b9a8027bb75086.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 482, + 225, + 553, + 300 + ], + "blocks": [ + { + "bbox": [ + 482, + 225, + 553, + 300 + ], + "lines": [ + { + "bbox": [ + 482, + 225, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 482, + 225, + 553, + 300 + ], + "type": "image", + "image_path": "65b15ad3d14dcfdf37f169530e387220c0a0f4be1db9c162bc43b40135ded43b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 61, + 303, + 209, + 378 + ], + "blocks": [ + { + "bbox": [ + 61, + 303, + 209, + 378 + ], + "lines": [ + { + "bbox": [ + 61, + 303, + 209, + 378 + ], + "spans": [ + { + "bbox": [ + 61, + 303, + 209, + 378 + ], + "type": "image", + "image_path": "5b0136e4a70814225ff9ad68e12d98fd6cdd47a2533a82d83106c1c9cf2ff34d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 225, + 303, + 321, + 378 + ], + "blocks": [ + { + "bbox": [ + 225, + 303, + 321, + 378 + ], + "lines": [ + { + "bbox": [ + 225, + 303, + 321, + 378 + ], + "spans": [ + { + "bbox": [ + 225, + 303, + 321, + 378 + ], + "type": "image", + "image_path": "d23b97a3d5053b208c1695cff1d216aa44c759e877dae29b50cf0e70edc21e3a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 337, + 303, + 467, + 378 + ], + "blocks": [ + { + "bbox": [ + 337, + 303, + 467, + 378 + ], + "lines": [ + { + "bbox": [ + 337, + 303, + 467, + 378 + ], + "spans": [ + { + "bbox": [ + 337, + 303, + 467, + 378 + ], + "type": "image", + "image_path": "b01e457706727a58340b4782b335d9894c48b283b65ef0613003ad06ba0e3861.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 482, + 303, + 553, + 378 + ], + "blocks": [ + { + "bbox": [ + 482, + 303, + 553, + 378 + ], + "lines": [ + { + "bbox": [ + 482, + 303, + 553, + 378 + ], + "spans": [ + { + "bbox": [ + 482, + 303, + 553, + 378 + ], + "type": "image", + "image_path": "7dbdaec09dc5d19332391d91976db620033167d5201946c6543662e99d052c14.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 61, + 380, + 211, + 456 + ], + "blocks": [ + { + "bbox": [ + 61, + 380, + 211, + 456 + ], + "lines": [ + { + "bbox": [ + 61, + 380, + 211, + 456 + ], + "spans": [ + { + "bbox": [ + 61, + 380, + 211, + 456 + ], + "type": "image", + "image_path": "7c6dd9c5e0d37fa753389b59a2429bf602a835309e8bad16363da928abb837da.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 225, + 380, + 321, + 456 + ], + "blocks": [ + { + "bbox": [ + 225, + 380, + 321, + 456 + ], + "lines": [ + { + "bbox": [ + 225, + 380, + 321, + 456 + ], + "spans": [ + { + "bbox": [ + 225, + 380, + 321, + 456 + ], + "type": "image", + "image_path": "55e6dbca66a3508f40a4446e71f825e3c86c9546b39cffd8090a46ebafc322a0.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 337, + 380, + 467, + 456 + ], + "blocks": [ + { + "bbox": [ + 337, + 380, + 467, + 456 + ], + "lines": [ + { + "bbox": [ + 337, + 380, + 467, + 456 + ], + "spans": [ + { + "bbox": [ + 337, + 380, + 467, + 456 + ], + "type": "image", + "image_path": "352cae2494fb7f7f5a3064623cb9ad8fad88c60768efc92c343aa5d31f1e356f.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 482, + 380, + 553, + 456 + ], + "blocks": [ + { + "bbox": [ + 482, + 380, + 553, + 456 + ], + "lines": [ + { + "bbox": [ + 482, + 380, + 553, + 456 + ], + "spans": [ + { + "bbox": [ + 482, + 380, + 553, + 456 + ], + "type": "image", + "image_path": "7e12cb90c72abd1f0ca518327f5992f2e1619ef7e28f77df30c38f6c9d24af84.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 61, + 458, + 209, + 534 + ], + "blocks": [ + { + "bbox": [ + 61, + 458, + 209, + 534 + ], + "lines": [ + { + "bbox": [ + 61, + 458, + 209, + 534 + ], + "spans": [ + { + "bbox": [ + 61, + 458, + 209, + 534 + ], + "type": "image", + "image_path": "9ea1be9da8d3f3619bb32639cd42d42f504dc2ab8bb70cffcf3987ab8dbf565b.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 225, + 458, + 321, + 534 + ], + "blocks": [ + { + "bbox": [ + 225, + 458, + 321, + 534 + ], + "lines": [ + { + "bbox": [ + 225, + 458, + 321, + 534 + ], + "spans": [ + { + "bbox": [ + 225, + 458, + 321, + 534 + ], + "type": "image", + "image_path": "f5c295656393be322a38e937da6b92b3df19fa51d55f639e31d3fbc2433a1338.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 337, + 458, + 467, + 534 + ], + "blocks": [ + { + "bbox": [ + 337, + 458, + 467, + 534 + ], + "lines": [ + { + "bbox": [ + 337, + 458, + 467, + 534 + ], + "spans": [ + { + "bbox": [ + 337, + 458, + 467, + 534 + ], + "type": "image", + "image_path": "631c684044426f78ee5948c8e54c73fcf313555544667f9e23b799a068628b18.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 482, + 458, + 553, + 534 + ], + "blocks": [ + { + "bbox": [ + 482, + 458, + 553, + 534 + ], + "lines": [ + { + "bbox": [ + 482, + 458, + 553, + 534 + ], + "spans": [ + { + "bbox": [ + 482, + 458, + 553, + 534 + ], + "type": "image", + "image_path": "5574e4f75cd99d2ddf79fe59fa44d4a1de17aafeca7aee5ae327a3704bbfff8f.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 61, + 536, + 209, + 613 + ], + "blocks": [ + { + "bbox": [ + 61, + 536, + 209, + 613 + ], + "lines": [ + { + "bbox": [ + 61, + 536, + 209, + 613 + ], + "spans": [ + { + "bbox": [ + 61, + 536, + 209, + 613 + ], + "type": "image", + "image_path": "a096fcd922912a6b0cd0fded11d29e19d162df8e2658f0286894bf77f89ebb48.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 225, + 536, + 321, + 613 + ], + "blocks": [ + { + "bbox": [ + 225, + 536, + 321, + 613 + ], + "lines": [ + { + "bbox": [ + 225, + 536, + 321, + 613 + ], + "spans": [ + { + "bbox": [ + 225, + 536, + 321, + 613 + ], + "type": "image", + "image_path": "a953e849b95fd4a0cf76f04ebcaf4e625758c85498c036f63154c6f025797bbd.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 337, + 536, + 467, + 613 + ], + "blocks": [ + { + "bbox": [ + 337, + 536, + 467, + 613 + ], + "lines": [ + { + "bbox": [ + 337, + 536, + 467, + 613 + ], + "spans": [ + { + "bbox": [ + 337, + 536, + 467, + 613 + ], + "type": "image", + "image_path": "14668697365c19fbe1a5240ad5f67ee20ad025640d6d803f360482e44da0d575.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 482, + 536, + 553, + 613 + ], + "blocks": [ + { + "bbox": [ + 482, + 536, + 553, + 613 + ], + "lines": [ + { + "bbox": [ + 482, + 536, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 482, + 536, + 553, + 613 + ], + "type": "image", + "image_path": "93d5e58e6b876b030d6748806502b26f564d44495ed94a535cb1372eaa3a0c61.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 61, + 615, + 211, + 690 + ], + "blocks": [ + { + "bbox": [ + 61, + 615, + 211, + 690 + ], + "lines": [ + { + "bbox": [ + 61, + 615, + 211, + 690 + ], + "spans": [ + { + "bbox": [ + 61, + 615, + 211, + 690 + ], + "type": "image", + "image_path": "3e9a5d12703748d56e89df067705455069a798bf2fa74705a1df6891334216b2.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 698, + 553, + 720 + ], + "lines": [ + { + "bbox": [ + 55, + 698, + 553, + 720 + ], + "spans": [ + { + "bbox": [ + 55, + 698, + 553, + 720 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison on special scenarios. In the special scenarios of games, artworks, AI-generated content, and movies, our method demonstrates strong generalization capability and the ability to predict detailed depth." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 225, + 615, + 321, + 690 + ], + "blocks": [ + { + "bbox": [ + 225, + 615, + 321, + 690 + ], + "lines": [ + { + "bbox": [ + 225, + 615, + 321, + 690 + ], + "spans": [ + { + "bbox": [ + 225, + 615, + 321, + 690 + ], + "type": "image", + "image_path": "0cb53a9ded46b4334bd4a68fcd10f4e388dcb382db6c0f2e4e687f25ed60b8b0.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 337, + 613, + 467, + 690 + ], + "blocks": [ + { + "bbox": [ + 337, + 613, + 467, + 690 + ], + "lines": [ + { + "bbox": [ + 337, + 613, + 467, + 690 + ], + "spans": [ + { + "bbox": [ + 337, + 613, + 467, + 690 + ], + "type": "image", + "image_path": "1007fe4e7e5f20e0f07830c3f573be5c74c82b824042b410f2c549c10c02ff71.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 482, + 615, + 553, + 690 + ], + "blocks": [ + { + "bbox": [ + 482, + 615, + 553, + 690 + ], + "lines": [ + { + "bbox": [ + 482, + 615, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 482, + 615, + 553, + 690 + ], + "type": "image", + "image_path": "dfb46957cde7b660ae50ea55cbc9c2e15d7168ccdab2475c6d7bde1d8bd0151e.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6029" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 138, + 126 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 138, + 126 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 138, + 126 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 138, + 126 + ], + "type": "image", + "image_path": "db957fbc803541c18c65c0ab7824d0e348c83233adf0f8b163ccff6ee28da947.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 57, + 126, + 139, + 183 + ], + "blocks": [ + { + "bbox": [ + 57, + 126, + 139, + 183 + ], + "lines": [ + { + "bbox": [ + 57, + 126, + 139, + 183 + ], + "spans": [ + { + "bbox": [ + 57, + 126, + 139, + 183 + ], + "type": "image", + "image_path": "4408792880f35668c66f7a1773d948ceb3ccdb98235364b8f9feb83ce3b51363.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 186, + 122, + 196 + ], + "lines": [ + { + "bbox": [ + 74, + 186, + 122, + 196 + ], + "spans": [ + { + "bbox": [ + 74, + 186, + 122, + 196 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 139, + 70, + 222, + 126 + ], + "blocks": [ + { + "bbox": [ + 139, + 70, + 222, + 126 + ], + "lines": [ + { + "bbox": [ + 139, + 70, + 222, + 126 + ], + "spans": [ + { + "bbox": [ + 139, + 70, + 222, + 126 + ], + "type": "image", + "image_path": "7cec304557161c7139f9dad19e333aa9d0b0889d7adf4e89baf4594d2f0c7293.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 139, + 126, + 222, + 183 + ], + "blocks": [ + { + "bbox": [ + 139, + 126, + 222, + 183 + ], + "lines": [ + { + "bbox": [ + 139, + 126, + 222, + 183 + ], + "spans": [ + { + "bbox": [ + 139, + 126, + 222, + 183 + ], + "type": "image", + "image_path": "a55904f4b059b258a9ec9d50b06981f63f4f8fcb643adea9ced92af0113eb18a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 185, + 206, + 196 + ], + "lines": [ + { + "bbox": [ + 156, + 185, + 206, + 196 + ], + "spans": [ + { + "bbox": [ + 156, + 185, + 206, + 196 + ], + "type": "text", + "content": "Alpha Matte" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 223, + 70, + 304, + 126 + ], + "blocks": [ + { + "bbox": [ + 223, + 70, + 304, + 126 + ], + "lines": [ + { + "bbox": [ + 223, + 70, + 304, + 126 + ], + "spans": [ + { + "bbox": [ + 223, + 70, + 304, + 126 + ], + "type": "image", + "image_path": "85a2bf0fb2277c68cdf14851d8701583ed6dd48677df072472a1ee31d230754c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 223, + 126, + 304, + 183 + ], + "blocks": [ + { + "bbox": [ + 223, + 126, + 304, + 183 + ], + "lines": [ + { + "bbox": [ + 223, + 126, + 304, + 183 + ], + "spans": [ + { + "bbox": [ + 223, + 126, + 304, + 183 + ], + "type": "image", + "image_path": "4f7f9728d41f27fc60491e57903b1c9948e658fa69a86047dcb9e3a382dcd469.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 244, + 185, + 283, + 196 + ], + "lines": [ + { + "bbox": [ + 244, + 185, + 283, + 196 + ], + "spans": [ + { + "bbox": [ + 244, + 185, + 283, + 196 + ], + "type": "text", + "content": "FifDepth" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 136, + 201, + 473, + 213 + ], + "lines": [ + { + "bbox": [ + 136, + 201, + 473, + 213 + ], + "spans": [ + { + "bbox": [ + 136, + 201, + 473, + 213 + ], + "type": "text", + "content": "Figure 6. Boundary visualization comparison. These samples are from the AM-2k dataset." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 306, + 70, + 388, + 126 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 388, + 126 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 388, + 126 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 388, + 126 + ], + "type": "image", + "image_path": "3c8e678e149d1839280e615bdb6afb36d945b59824e59343e53860876a6cb6db.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 306, + 126, + 388, + 183 + ], + "blocks": [ + { + "bbox": [ + 306, + 126, + 388, + 183 + ], + "lines": [ + { + "bbox": [ + 306, + 126, + 388, + 183 + ], + "spans": [ + { + "bbox": [ + 306, + 126, + 388, + 183 + ], + "type": "image", + "image_path": "b23fe43b85fe0a0494b3d00e83e6d84f4a44129720dab601c190bcf02d3525b5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 327, + 185, + 367, + 196 + ], + "lines": [ + { + "bbox": [ + 327, + 185, + 367, + 196 + ], + "spans": [ + { + "bbox": [ + 327, + 185, + 367, + 196 + ], + "type": "text", + "content": "Depth Pro" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 389, + 70, + 470, + 126 + ], + "blocks": [ + { + "bbox": [ + 389, + 70, + 470, + 126 + ], + "lines": [ + { + "bbox": [ + 389, + 70, + 470, + 126 + ], + "spans": [ + { + "bbox": [ + 389, + 70, + 470, + 126 + ], + "type": "image", + "image_path": "ed56d1b4aaace17d2a80e4834450a3833a9df91728d4161330364119eba2c64a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 389, + 126, + 470, + 183 + ], + "blocks": [ + { + "bbox": [ + 389, + 126, + 470, + 183 + ], + "lines": [ + { + "bbox": [ + 389, + 126, + 470, + 183 + ], + "spans": [ + { + "bbox": [ + 389, + 126, + 470, + 183 + ], + "type": "image", + "image_path": "d820bf91a99cda0ebb8c845a6d48d36e17e9aa3eb44f650a5029d52a11b3fd55.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 401, + 185, + 457, + 196 + ], + "lines": [ + { + "bbox": [ + 401, + 185, + 457, + 196 + ], + "spans": [ + { + "bbox": [ + 401, + 185, + 457, + 196 + ], + "type": "text", + "content": "Depth Any. v2" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 471, + 70, + 553, + 126 + ], + "blocks": [ + { + "bbox": [ + 471, + 70, + 553, + 126 + ], + "lines": [ + { + "bbox": [ + 471, + 70, + 553, + 126 + ], + "spans": [ + { + "bbox": [ + 471, + 70, + 553, + 126 + ], + "type": "image", + "image_path": "e4403e627c42001ab1b89e07a197446a4e327d306ef317c293eceaf7311598ac.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 471, + 126, + 553, + 183 + ], + "blocks": [ + { + "bbox": [ + 471, + 126, + 553, + 183 + ], + "lines": [ + { + "bbox": [ + 471, + 126, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 471, + 126, + 553, + 183 + ], + "type": "image", + "image_path": "e0e519a1d3d2c85aaf2bc92d9a7274adf247accebe061d20beb792f37289fd62.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 493, + 185, + 531, + 196 + ], + "lines": [ + { + "bbox": [ + 493, + 185, + 531, + 196 + ], + "spans": [ + { + "bbox": [ + 493, + 185, + 531, + 196 + ], + "type": "text", + "content": "DepthFM" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 109, + 216, + 502, + 267 + ], + "blocks": [ + { + "bbox": [ + 109, + 216, + 502, + 267 + ], + "lines": [ + { + "bbox": [ + 109, + 216, + 502, + 267 + ], + "spans": [ + { + "bbox": [ + 109, + 216, + 502, + 267 + ], + "type": "table", + "html": "
MethodSintel F1↑Spring F1↑iBims F1↑AM R↑P3M R↑DIS R↑
DepthAnything v20.2280.0560.1110.1070.1310.056
Depth Pro0.4090.0790.1760.1730.1680.077
FiffDepth (Ours)0.4230.0860.1890.1760.1790.091
", + "image_path": "26061047379f50cacf4fa4e1bc1cefa0c104a14f97a81db0c31de9ac5c7c95e6.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_body" + } + ], + "index": 19 + }, + { + "type": "table", + "bbox": [ + 70, + 300, + 541, + 330 + ], + "blocks": [ + { + "bbox": [ + 55, + 275, + 555, + 299 + ], + "lines": [ + { + "bbox": [ + 55, + 275, + 555, + 299 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 555, + 299 + ], + "type": "text", + "content": "Table 2. Zero-shot boundary accuracy. We provide the F1 score for datasets containing ground-truth depth and boundary recall (R) for those with matting or segmentation labels." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 300, + 541, + 330 + ], + "lines": [ + { + "bbox": [ + 70, + 300, + 541, + 330 + ], + "spans": [ + { + "bbox": [ + 70, + 300, + 541, + 330 + ], + "type": "table", + "html": "
MethodMarigoldMarigold (LCM)GeoWizardDepthFMDepthAnything v2-LDepth ProOurs
Time (s)1031.7190.390.0260.230.092
", + "image_path": "28b816a245926acf8c23a4ff461688dc1574c7d8fec326d75bf874374e56612b.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 56, + 338, + 552, + 350 + ], + "lines": [ + { + "bbox": [ + 56, + 338, + 552, + 350 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 552, + 350 + ], + "type": "text", + "content": "Table 3. Running time comparison. We perform inference on " + }, + { + "bbox": [ + 56, + 338, + 552, + 350 + ], + "type": "inline_equation", + "content": "{100512} \\times {512}" + }, + { + "bbox": [ + 56, + 338, + 552, + 350 + ], + "type": "text", + "content": " images using these methods and report the average time." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 57, + 361, + 116, + 398 + ], + "blocks": [ + { + "bbox": [ + 57, + 361, + 116, + 398 + ], + "lines": [ + { + "bbox": [ + 57, + 361, + 116, + 398 + ], + "spans": [ + { + "bbox": [ + 57, + 361, + 116, + 398 + ], + "type": "image", + "image_path": "152617c8d3931740ac93ee795d9a59ecddd9329bd3f4bc2c6cca17915c4c4052.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 399, + 104, + 407 + ], + "lines": [ + { + "bbox": [ + 69, + 399, + 104, + 407 + ], + "spans": [ + { + "bbox": [ + 69, + 399, + 104, + 407 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 117, + 361, + 176, + 398 + ], + "blocks": [ + { + "bbox": [ + 117, + 361, + 176, + 398 + ], + "lines": [ + { + "bbox": [ + 117, + 361, + 176, + 398 + ], + "spans": [ + { + "bbox": [ + 117, + 361, + 176, + 398 + ], + "type": "image", + "image_path": "3319f3671586c6464afb2e2cad2ef05aedfe65842b41d1c2497d96a7001f9c09.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 399, + 160, + 407 + ], + "lines": [ + { + "bbox": [ + 131, + 399, + 160, + 407 + ], + "spans": [ + { + "bbox": [ + 131, + 399, + 160, + 407 + ], + "type": "text", + "content": "w/o blend" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 177, + 361, + 235, + 398 + ], + "blocks": [ + { + "bbox": [ + 177, + 361, + 235, + 398 + ], + "lines": [ + { + "bbox": [ + 177, + 361, + 235, + 398 + ], + "spans": [ + { + "bbox": [ + 177, + 361, + 235, + 398 + ], + "type": "image", + "image_path": "2ade62655ef2de54b797386ae8476937bfba6c2503ec0df92a1b27166a12770a.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 195, + 399, + 215, + 407 + ], + "lines": [ + { + "bbox": [ + 195, + 399, + 215, + 407 + ], + "spans": [ + { + "bbox": [ + 195, + 399, + 215, + 407 + ], + "type": "text", + "content": "w/o " + }, + { + "bbox": [ + 195, + 399, + 215, + 407 + ], + "type": "inline_equation", + "content": "L_{k}" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 236, + 361, + 294, + 398 + ], + "blocks": [ + { + "bbox": [ + 236, + 361, + 294, + 398 + ], + "lines": [ + { + "bbox": [ + 236, + 361, + 294, + 398 + ], + "spans": [ + { + "bbox": [ + 236, + 361, + 294, + 398 + ], + "type": "image", + "image_path": "0e97e459a08548dccd2934beb21bb88f8226d3f93d1731f8f7879a9bff50ae5f.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 240, + 399, + 290, + 407 + ], + "lines": [ + { + "bbox": [ + 240, + 399, + 290, + 407 + ], + "spans": [ + { + "bbox": [ + 240, + 399, + 290, + 407 + ], + "type": "text", + "content": "FiffDepth Output" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 57, + 409, + 116, + 451 + ], + "blocks": [ + { + "bbox": [ + 57, + 409, + 116, + 451 + ], + "lines": [ + { + "bbox": [ + 57, + 409, + 116, + 451 + ], + "spans": [ + { + "bbox": [ + 57, + 409, + 116, + 451 + ], + "type": "image", + "image_path": "cb318011e50e35e791049b038982c206117c2b031f64ab0749f606c0dca4ff75.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 453, + 104, + 461 + ], + "lines": [ + { + "bbox": [ + 69, + 453, + 104, + 461 + ], + "spans": [ + { + "bbox": [ + 69, + 453, + 104, + 461 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 117, + 409, + 176, + 451 + ], + "blocks": [ + { + "bbox": [ + 117, + 409, + 176, + 451 + ], + "lines": [ + { + "bbox": [ + 117, + 409, + 176, + 451 + ], + "spans": [ + { + "bbox": [ + 117, + 409, + 176, + 451 + ], + "type": "image", + "image_path": "5a0211de8ca57da8628960965f017df04f6b749c95454e0fed8d50c0b4e7dc5d.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 453, + 161, + 461 + ], + "lines": [ + { + "bbox": [ + 130, + 453, + 161, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 161, + 461 + ], + "type": "text", + "content": "w/o DINO" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 177, + 409, + 235, + 451 + ], + "blocks": [ + { + "bbox": [ + 177, + 409, + 235, + 451 + ], + "lines": [ + { + "bbox": [ + 177, + 409, + 235, + 451 + ], + "spans": [ + { + "bbox": [ + 177, + 409, + 235, + 451 + ], + "type": "image", + "image_path": "976d74704298fc43877737e786a6547ccc0a9c7fa278851b2626b0ce4c6c9e27.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 187, + 453, + 223, + 461 + ], + "lines": [ + { + "bbox": [ + 187, + 453, + 223, + 461 + ], + "spans": [ + { + "bbox": [ + 187, + 453, + 223, + 461 + ], + "type": "text", + "content": "DINO on " + }, + { + "bbox": [ + 187, + 453, + 223, + 461 + ], + "type": "inline_equation", + "content": "d_0" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 236, + 409, + 294, + 452 + ], + "blocks": [ + { + "bbox": [ + 236, + 409, + 294, + 452 + ], + "lines": [ + { + "bbox": [ + 236, + 409, + 294, + 452 + ], + "spans": [ + { + "bbox": [ + 236, + 409, + 294, + 452 + ], + "type": "image", + "image_path": "0d3ffd45d5694eef30e5b07fee1d6b9e795531cd0f168a4b47e94c9b36a587cb.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 240, + 453, + 290, + 461 + ], + "lines": [ + { + "bbox": [ + 240, + 453, + 290, + 461 + ], + "spans": [ + { + "bbox": [ + 240, + 453, + 290, + 461 + ], + "type": "text", + "content": "FiffDepth Output" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 475, + 295, + 507 + ], + "lines": [ + { + "bbox": [ + 55, + 475, + 295, + 507 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 295, + 507 + ], + "type": "text", + "content": "Figure 7. Ablation studies. The generalization capability and depth details of the method are affected when some essential components are missing." + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "bbox": [ + 55, + 520, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 295, + 590 + ], + "type": "text", + "content": "metric and its computation, please refer to the Depth Pro paper for further details. Quantitative comparisons in Table 2 demonstrate that our method surpasses Depth Pro and other approaches in boundary prediction. Additionally, the visual results in Figure 6 further validate that our method predicts more accurate boundaries." + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "Running time. Generative MDE models adopt the diffusion paradigm, and the resulting instability necessitates test-time assembly, leading to a lengthy inference time. In contrast, our feed-forward approach provides significant efficiency advantages. We evaluate the average inference time for a " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " image on an NVIDIA Titan RTX GPU. As shown in Table 3, our method significantly outperforms other generative approaches in terms of efficiency and achieves performance comparable to DAv2. We tested these methods using their default settings." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 313, + 363, + 555, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 555, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 555, + 471 + ], + "type": "text", + "content": "Ablation studies. We conduct ablation studies to validate components of our method. Keeping the diffusion trajectory but predicting purely image latents affects the relative depth relationships between objects (1st row, 1st result in Fig. 7). Without keeping trajectory, some details are lost (1st row, 2nd result in Fig. 7). Omitting DINO supervision impacts the model's generalization ability (2nd row, 1st result in Fig. 7). Using DINO supervision at " + }, + { + "bbox": [ + 313, + 363, + 555, + 471 + ], + "type": "inline_equation", + "content": "d_{0}" + }, + { + "bbox": [ + 313, + 363, + 555, + 471 + ], + "type": "text", + "content": " also reduces details (2nd row, 2nd result in Fig. 7)." + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 314, + 501, + 388, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 501, + 388, + 513 + ], + "spans": [ + { + "bbox": [ + 314, + 501, + 388, + 513 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 313, + 527, + 555, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 527, + 555, + 622 + ], + "spans": [ + { + "bbox": [ + 313, + 527, + 555, + 622 + ], + "type": "text", + "content": "In this work, we transform diffusion models into stable, feed-forward depth estimators, achieving significant improvements in accuracy and efficiency over generative model-based methods. By combining the detail preservation of generative models with the robust generalization of FFN models like DINOv2, our hybrid approach bridges the synthetic-to-real gap, enhancing stability, predictability, and resolution in MDE for diverse real-world scenarios." + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "content": "Acknowledgements. Q. Huang acknowledges the support of NSF-2047677, NSF-2413161, NSF-2504906, NSF-2515626, and GIFTs from Adobe and Google. This research has been supported by computing support on the Vista GPU Cluster through the Center for Generative AI (CGAI) and the Texas Advanced Computing Center (TACC) at UT Austin." + } + ] + } + ], + "index": 45 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6030" + } + ] + } + ], + "index": 46 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Shubhra Aich, Jean Marie Uwabeza Vianney, Md Amirul Islam, and Mannat Kaur Bingbing Liu. Bidirectional attention network for monocular depth estimation. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 11746-11752. IEEE, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 147, + 296, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 147, + 296, + 212 + ], + "spans": [ + { + "bbox": [ + 61, + 147, + 296, + 212 + ], + "type": "text", + "content": "[2] Yunpeng Bai, Yanbo Fan, Xuan Wang, Yong Zhang, Jingxiang Sun, Chun Yuan, and Ying Shan. High-fidelity facial avatar reconstruction from monocular video with generative priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4541-4551, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 214, + 295, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 214, + 295, + 267 + ], + "spans": [ + { + "bbox": [ + 62, + 214, + 295, + 267 + ], + "type": "text", + "content": "[3] Yunpeng Bai, Xintao Wang, Yan-Pei Cao, Yixiao Ge, Chun Yuan, and Ying Shan. Dreamdiffusion: High-quality eegt-to-image generation with temporal masked signal modeling and clip alignment. In European Conference on Computer Vision, pages 472-488. Springer, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 270, + 295, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 270, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 270, + 295, + 312 + ], + "type": "text", + "content": "[4] Yoshua Bengio, Li Yao, Guillaume Alain, and Pascal Vincent. Generalized denoising auto-encoders as generative models. Advances in neural information processing systems, 26, 2013. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 314, + 294, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 314, + 294, + 367 + ], + "spans": [ + { + "bbox": [ + 62, + 314, + 294, + 367 + ], + "type": "text", + "content": "[5] Aleksei Bochkovskii, Amaël Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. arXiv preprint arXiv:2410.02073, 2024. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 369, + 294, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 294, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 294, + 434 + ], + "type": "text", + "content": "[6] Daniel J Butler, Jonas Wulff, Garrett B Stanley, and Michael J Black. A naturalistic open source movie for optical flow evaluation. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part VI 12, pages 611-625. Springer, 2012. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 436, + 294, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 294, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 294, + 456 + ], + "type": "text", + "content": "[7] Yohann Cabon, Naila Murray, and Martin Humenberger. Virtual katii 2. arXiv preprint arXiv:2001.10773, 2020. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 459, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 459, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 459, + 294, + 491 + ], + "type": "text", + "content": "[8] Weifeng Chen, Zhao Fu, Dawei Yang, and Jia Deng. Single-image depth perception in the wild. Advances in neural information processing systems, 29, 2016. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 492, + 295, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 492, + 295, + 546 + ], + "spans": [ + { + "bbox": [ + 62, + 492, + 295, + 546 + ], + "type": "text", + "content": "[9] Weifeng Chen, Shengyi Qian, David Fan, Noriyuki Kojima, Max Hamilton, and Jia Deng. Oasis: A large-scale dataset for single image 3d in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 679-688, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 548, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 548, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 57, + 548, + 294, + 602 + ], + "type": "text", + "content": "[10] Yuhua Chen, Cordelia Schmid, and Cristian Sminchisescu. Self-supervised learning with geometric constraints in monocular video: Connecting flow, depth, and camera. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7063-7072, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 604, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 604, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 57, + 604, + 294, + 658 + ], + "type": "text", + "content": "[11] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 659, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 659, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 659, + 294, + 713 + ], + "type": "text", + "content": "[12] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multitask mid-level vision datasets from 3d scans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10786-10796, 2021. 2, 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[13] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. Advances in neural information processing systems, 27, 2014. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 119, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 173 + ], + "type": "text", + "content": "[14] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Batmanghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2002-2011, 2018. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 175, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 553, + 228 + ], + "type": "text", + "content": "[15] Xiao Fu, Wei Yin, Mu Hu, Kaixuan Wang, Yuexin Ma, Ping Tan, Shaojie Shen, Dahua Lin, and Xiaoxiao Long. Geowizard: Unleashing the diffusion priors for 3d geometry estimation from a single image. arXiv preprint arXiv:2403.12013, 2024. 2, 3, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 232, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 232, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 553, + 274 + ], + "type": "text", + "content": "[16] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 276, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 276, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 317, + 276, + 553, + 330 + ], + "type": "text", + "content": "[17] Ming Gui, Johannes S Fischer, Ulrich Prestel, Pingchuan Ma, Dmytro Kotovenko, Olga Grebenkova, Stefan Andreas Baumann, Vincent Tao Hu, and Björn Ommer. Depthfm: Fast monocular depth estimation with flow matching. arXiv preprint arXiv:2403.13788, 2024. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 332, + 553, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 332, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 317, + 332, + 553, + 385 + ], + "type": "text", + "content": "[18] Jing He, Haodong Li, Wei Yin, Yixun Liang, Leheng Li, Kaiqiang Zhou, Hongbo Liu, Bingbing Liu, and YingCong Chen. Lotus: Diffusion-based visual foundation model for high-quality dense prediction. arXiv preprint arXiv:2409.18124, 2024. 3, 4, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 388, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 388, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 317, + 388, + 553, + 441 + ], + "type": "text", + "content": "[19] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 445, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 553, + 476 + ], + "type": "text", + "content": "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 479, + 553, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 479, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 317, + 479, + 553, + 542 + ], + "type": "text", + "content": "[21] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 545, + 553, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 545, + 553, + 598 + ], + "spans": [ + { + "bbox": [ + 317, + 545, + 553, + 598 + ], + "type": "text", + "content": "[22] Tobias Koch, Lukas Liebel, Friedrich Fraundorfer, and Marco Korner. Evaluation of cnn-based single-image depth estimation methods. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, pages 0–0, 2018. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 601, + 555, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 601, + 555, + 644 + ], + "spans": [ + { + "bbox": [ + 317, + 601, + 555, + 644 + ], + "type": "text", + "content": "[23] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1611-1621, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "type": "text", + "content": "[24] Jin Han Lee, Myung-Kyu Han, Dong Wook Ko, and Il Hong Suh. From big to small: Multi-scale local planar guidance for monocular depth estimation. arXiv preprint arXiv:1907.10326, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[25] Jizhizi Li, Sihan Ma, Jing Zhang, and Dacheng Tao. Privacy-preserving portrait matting. In Proceedings of the 29th ACM" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "6031" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "international conference on multimedia, pages 3501-3509, 2021. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[26] Jizhizi Li, Jing Zhang, Stephen J Maybank, and Dacheng Tao. Bridging composite and real: towards end-to-end deep image matting. International Journal of Computer Vision, 130(2):246-266, 2022. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 185 + ], + "type": "text", + "content": "[27] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2041-2050, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 295, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 295, + 230 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 295, + 230 + ], + "type": "text", + "content": "[28] Zhenyu Li, Zehui Chen, Xianming Liu, and Junjun Jiang. Depthformer: Exploiting long-range correlation and local information for accurate monocular depth estimation. Machine Intelligence Research, 20(6):837-854, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 232, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 294, + 297 + ], + "type": "text", + "content": "[29] Gonzalo Martin Garcia, Karim Abou Zeid, Christian Schmidt, Daan de Geus, Alexander Hermans, and Bastian Leibe. Fine-tuning image-conditional diffusion models is easier than you think. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2025. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 299, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 294, + 363 + ], + "type": "text", + "content": "[30] Lukas Mehl, Jenny Schmalfuss, Azin Jahedi, Yaroslava Nalivayko, and Andres Bruhn. Spring: A high-resolution high-detail dataset and benchmark for scene flow, optical flow and stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4981-4991, 2023. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 294, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 420 + ], + "type": "text", + "content": "[31] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 295, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 295, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 295, + 476 + ], + "type": "text", + "content": "[32] Vaishakh Patil, Christos Sakaridis, Alexander Liniger, and Luc Van Gool. P3depth: Monocular depth estimation with a piecewise planarity prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1610-1621, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 478, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 521 + ], + "type": "text", + "content": "[33] Xuebin Qin, Hang Dai, Xiaobin Hu, Deng-Ping Fan, Ling Shao, and Luc Van Gool. Highly accurate dichotomous image segmentation. In European Conference on Computer Vision, pages 38-56. Springer, 2022. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "type": "text", + "content": "[34] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 44(3):1623-1637, 2020. 1, 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 295, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 295, + 623 + ], + "type": "text", + "content": "[35] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021. 1, 2, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 689 + ], + "type": "text", + "content": "[36] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10912-10922, 2021. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "text", + "content": "synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 108, + 554, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 554, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 554, + 140 + ], + "type": "text", + "content": "[38] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 141, + 554, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 554, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 554, + 207 + ], + "type": "text", + "content": "[39] Thomas Schops, Johannes L Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3260-3269, 2017. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 209, + 554, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 554, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 554, + 274 + ], + "type": "text", + "content": "[40] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 276, + 554, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 554, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 554, + 319 + ], + "type": "text", + "content": "[41] Jaidev Shriram, Alex Trevithick, Lingjie Liu, and Ravi Ramamoorthi. Realmdreamer: Text-driven 3d scene generation with inpainting and depth diffusion. arXiv preprint arXiv:2404.07199, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 321, + 554, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 554, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 554, + 387 + ], + "type": "text", + "content": "[42] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12, pages 746-760. Springer, 2012. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 388, + 554, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 554, + 421 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 554, + 421 + ], + "type": "text", + "content": "[43] Zachary Teed and Jia Deng. Deepv2d: Video to depth with differentiable structure from motion. arXiv preprint arXiv:1812.04605, 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 422, + 554, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 554, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 554, + 477 + ], + "type": "text", + "content": "[44] Igor Vasiljevic, Nick Kolkin, Shanyi Zhang, Ruotian Luo, Haochen Wang, Falcon Z Dai, Andrea F Daniele, Mohammadreza Mostajabi, Steven Basart, Matthew R Walter, et al. Diode: A dense indoor and outdoor depth dataset. arXiv preprint arXiv:1908.00463, 2019. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 479, + 554, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 479, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 479, + 554, + 533 + ], + "type": "text", + "content": "[45] Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. In Proceedings of the 25th international conference on Machine learning, pages 1096-1103, 2008. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 535, + 554, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 535, + 554, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 535, + 554, + 589 + ], + "type": "text", + "content": "[46] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9065-9076, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 591, + 555, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 591, + 555, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 591, + 555, + 657 + ], + "type": "text", + "content": "[47] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudolidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8445-8453, 2019. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 658, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 554, + 712 + ], + "type": "text", + "content": "[48] Yiran Wang, Min Shi, Jiaqi Li, Zihao Huang, Zhiguo Cao, Jianming Zhang, Ke Xian, and Guosheng Lin. Neural video depth stabilizer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9466-9476, 2023. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6032" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "type": "text", + "content": "[49] Guangkai Xu, Yongtao Ge, Mingyu Liu, Chengxiang Fan, Kangyang Xie, Zhiyue Zhao, Hao Chen, and Chunhua Shen. Diffusion models trained with large data are transferable visual models. arXiv preprint arXiv:2403.06090, 2024. 3, 4, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 127, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 127, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 127, + 294, + 183 + ], + "type": "text", + "content": "[50] Guanglei Yang, Hao Tang, Mingli Ding, Nicu Sebe, and Elisa Ricci. Transformer-based attention networks for continuous pixel-wise prediction. In Proceedings of the IEEE/CVF International Conference on Computer vision, pages 16269-16279, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 183, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 183, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 183, + 295, + 239 + ], + "type": "text", + "content": "[51] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 239, + 294, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 281 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 281 + ], + "type": "text", + "content": "[52] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024. 1, 2, 4, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 282, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 282, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 282, + 294, + 326 + ], + "type": "text", + "content": "[53] Xingyi Yang and Xinchao Wang. Diffusion model as representation learner. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18938-18949, 2023. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 294, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 382 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 382 + ], + "type": "text", + "content": "[54] Rajeev Yasarla, Hong Cai, Jisoo Jeong, Yunxiao Shi, Risheek Garrepalli, and Fatih Porikli. Mamo: Leveraging memory and attention for monocular video depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8754-8764, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 383, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 437 + ], + "type": "text", + "content": "[55] Chongjie Ye, Lingteng Qiu, Xiaodong Gu, Qi Zuo, Yushuang Wu, Zilong Dong, Liefeng Bo, Yuliang Xiu, and Xiaoguang Han. Stablenormal: Reducing diffusion variance for stable and sharp normal. arXiv preprint arXiv:2406.16864, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 437, + 294, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 482 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 482 + ], + "type": "text", + "content": "[56] Wei Yin, Xinlong Wang, Chunhua Shen, Yifan Liu, Zhi Tian, Songcen Xu, Changming Sun, and Dou Renyin. Diversedepth: Affine-invariant depth prediction using diverse data. arXiv preprint arXiv:2002.00569, 2020. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 481, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 481, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 481, + 294, + 536 + ], + "type": "text", + "content": "[57] Wei Yin, Jianming Zhang, Oliver Wang, Simon Niklaus, Long Mai, Simon Chen, and Chunhua Shen. Learning to recover 3d scene shape from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 204-213, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 537, + 294, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 581 + ], + "type": "text", + "content": "[58] Chi Zhang, Wei Yin, Billzb Wang, Gang Yu, Bin Fu, and Chunhua Shen. Hierarchical normalization for robust monocular depth estimation. Advances in Neural Information Processing Systems, 35:14128-14139, 2022. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 582, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 294, + 635 + ], + "type": "text", + "content": "[59] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence. Advances in Neural Information Processing Systems, 36, 2024. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 636, + 294, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 691 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 691 + ], + "type": "text", + "content": "[60] Xiang Zhang, Bingxin Ke, Hayko Riemenschneider, Nando Metzger, Anton Obukhov, Markus Gross, Konrad Schindler, and Christopher Schroers. Betterdepth: Plug-and-play diffusion refiner for zero-shot monocular depth estimation. arXiv preprint arXiv:2407.17952, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[61] Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing text-to-image diffu" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 334, + 72, + 553, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 334, + 72, + 553, + 106 + ], + "type": "text", + "content": "sion models for visual perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5729-5739, 2023. 2, 3" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6033" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/FinMMR_ Make Financial Numerical Reasoning More Multimodal, Comprehensive, and Challenging/9bdd874d-4c07-4812-a904-cd4ae1c8c840_content_list.json b/2025/FinMMR_ Make Financial Numerical Reasoning More Multimodal, Comprehensive, and Challenging/9bdd874d-4c07-4812-a904-cd4ae1c8c840_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8320904b7861c2513e7d0ae4d33f88e61d721fc8 --- /dev/null +++ b/2025/FinMMR_ Make Financial Numerical Reasoning More Multimodal, Comprehensive, and Challenging/9bdd874d-4c07-4812-a904-cd4ae1c8c840_content_list.json @@ -0,0 +1,2094 @@ +[ + { + "type": "text", + "text": "$\\mathcal{F}_{\\mathsf{M}}$ FinMMR: Make Financial Numerical Reasoning More Multimodal, Comprehensive, and Challenging", + "text_level": 1, + "bbox": [ + 233, + 128, + 769, + 185 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6d3fed1f0bcad87000c97473484ddeecc534d3a4c21ada62bf3c83cc519932b9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 210, + 901, + 351 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6144a90463c441bce03a3a2907a21b545073c0b3c5b2f67d5224c79ff4245c54.jpg", + "image_caption": [ + "Figure 1. Overview of FinMMR. FinMMR presents three challenges: (1) Visual Perception: 8.7K financial images of 14 categories; (2) Knowledge Reasoning: 4.3K financial questions of 14 subdomains; (3) Numerical Computation: multi-step precise calculations." + ], + "image_footnote": [], + "bbox": [ + 94, + 357, + 297, + 527 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f24d79bb435f65d9e7c00c37f347d57d67a992fd34dfd9ffccd047a1b42f51e7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 300, + 357, + 542, + 527 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/01dee61a627ed956832cce85fa103ad5b11d1ca2ad78aa95efae85d2e2a430fa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 542, + 358, + 901, + 527 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 580, + 326, + 597 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present FinMMR, a novel bilingual multimodal benchmark tailored to evaluate the reasoning capabilities of multimodal large language models (MLLMs) in financial numerical reasoning tasks. Compared to existing benchmarks, our work introduces three significant advancements. (1) Multimodality: We meticulously transform existing financial reasoning benchmarks, and construct novel questions from the latest Chinese financial research reports. FinMMR comprises 4.3K questions and 8.7K images spanning 14 categories, including tables, bar charts, and ownership structure charts. (2) Comprehensiveness: FinMMR encompasses 14 financial subdomains, including corporate finance, banking, and industry analysis, significantly exceeding existing benchmarks in financial domain knowledge breadth. (3) Challenge: Models are required to perform multi-step precise numerical reasoning by integrating financial knowledge with the understanding of complex financial", + "bbox": [ + 86, + 614, + 485, + 872 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "images and text. The best-performing MLLM achieves only $53.0\\%$ accuracy on Hard problems. We believe that FinMMR will drive advancements in enhancing the reasoning capabilities of MLLMs in real-world scenarios.", + "bbox": [ + 511, + 583, + 906, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 676, + 643, + 691 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, large reasoning models (LRMs) [21, 47, 48, 54, 56, 62], show powerful reasoning capabilities over multi-step reasoning tasks, with train-time scaling and test-time scaling [31, 45]. These reasoning models are proficient in code [11, 29], math [35, 41], and science [59]. Multimodal large language models (MLLMs) [5, 19, 46] also exhibit greater performance on multimodal reasoning [39, 65].", + "bbox": [ + 509, + 702, + 906, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite significant advancements, there remains a notable gap in understanding the practical applicability of MLLMs in numerical reasoning within real-world scenarios, particularly in high-stakes fields such as finance and healthcare. As shown in Fig. 1, financial analysts in their daily work are required to read visually enriched finan", + "bbox": [ + 509, + 810, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3245", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rich Images Challenge Visual Perception", + "text_level": 1, + "bbox": [ + 122, + 95, + 320, + 125 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "问题：计算2016年至2024年Q2期间，上海银行和上市城商行的个人住房贷款年均复合增长率之差，结果以百分比表示，保留两位小数。(Calculate the difference in the compound annual growth rate (CAGR) of personal housing loans between Bank of Shanghai and listed city commercial banks from 2016 to Q2 2024. Present the result as a percentage rounded to two decimal places.)", + "bbox": [ + 114, + 136, + 339, + 215 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bd2ebb0c3cb9c81d16470e42bf5637c0afce94ef8fdb3223377f345ac18a68b5.jpg", + "image_caption": [ + "Answer: 2.17" + ], + "image_footnote": [], + "bbox": [ + 140, + 220, + 310, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Compound Average Growth Rate (CAGR)", + "text_level": 1, + "bbox": [ + 130, + 325, + 310, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "问题：请计算惠济新生对日照公司的间接持股比例，结果以百分比表示，保留两位小数。(Please calculate the indirect shareholding ratio of Huiji Xinsheng in Rizhao Company. Present the result as a percentage, rounded to two decimal places.)", + "bbox": [ + 112, + 335, + 344, + 386 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2c06b20ae45c66588922b3929d805d89c940c6b68154019b66b39aa3f0785f0a.jpg", + "image_caption": [ + "Answer: 13.65" + ], + "image_footnote": [], + "bbox": [ + 150, + 391, + 303, + 476 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Indirect Shareholding Ratio", + "text_level": 1, + "bbox": [ + 156, + 506, + 279, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "问题：请计算2022年第四季度和2023年第一季度的总资本开支，并将其与2021年第四季度的资本开支进行比较，计算下降百分比，结果保留两位小数。", + "bbox": [ + 112, + 518, + 344, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(Calculate the total capital expenditure for Q4 2022 and Q1 2023 combined, and compare it to the capital expenditure for Q4 2021. Calculate the percentage decrease, rounded to two decimal places.)", + "bbox": [ + 112, + 547, + 344, + 587 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1941a902cf9631c3e2f1f6cdf3f7c0af2ac31eecc1e88019fce8aa05dc5eda03.jpg", + "image_caption": [ + "Answer: -110.36", + "Capital Expenditure", + "Figure 2. Sampled FinMMR examples with two language (i.e., English and Chinese), rich images and different knowledge. The questions and images need expert-level visual perception, knowledge reasoning and numerical computation." + ], + "image_footnote": [], + "bbox": [ + 117, + 590, + 336, + 648 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Comprehensive Subdomains Challenge Knowledge Reasoning", + "text_level": 1, + "bbox": [ + 380, + 95, + 607, + 125 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Q: Table 19.3 shows a book balance sheet for the Wishing Well Motel chain. The company's long-term debt is secured by its real estate assets, but it also uses short-term bank loans as a permanent source of financing. It pays $10\\%$ interest on the bank debt and $9\\%$ interest on the secured debt. Wishing Well has 10 million shares of stock outstanding, trading at $\\$ 90$ per share. The expected return on Wishing Well's common stock is $18\\%$ . Calculate Wishing Well's WACC. Assume that the book and market values of Wishing Well's debt are the same. The marginal tax rate is $21\\%$ . Answer as a percentage to single decimal place.", + "bbox": [ + 380, + 136, + 609, + 248 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": ":", + "text_level": 1, + "bbox": [ + 380, + 250, + 421, + 258 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/f3693169345467c5dd0559a9a553a3dacece01a3098c7f498e0572128f1d1ac5.jpg", + "table_caption": [], + "table_footnote": [ + "TABLE 19.3 Book balance sheet for Wishing Well Inc. (figures in $ millions)" + ], + "table_body": "
Cash and marketable securities100Bank loan280
Accounts receivable200Accounts payable120
Inventory50Current liabilities400
Current assets350
Real estate2,100Long-term debt1,800
Other assets150Equity400
Total2,600Total2,600
", + "bbox": [ + 385, + 265, + 606, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Weighted Average Cost", + "text_level": 1, + "bbox": [ + 397, + 335, + 504, + 343 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Q: For each of the investments shown in the following table, calculate the rate of return earned over the unspecified time period. What is the rate of return for Investment A? Answer as a percentage to the nearest integer.", + "bbox": [ + 380, + 349, + 606, + 396 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": ":", + "text_level": 1, + "bbox": [ + 380, + 400, + 421, + 409 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/deea4bed290fc9743d21fcb1d6cf83389fb6841bdaf530acccb4cde1bb404345.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
InvestmentCash flow during periodBeginning-of-period valueEnd-of-period value
A$-2,800$23,400$20,100
B16,000225,000324,000
C7006,5008,000
D3,58036,60046,500
E-50062,70052,800
", + "bbox": [ + 393, + 421, + 599, + 479 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Rate of Return (RoR)", + "text_level": 1, + "bbox": [ + 439, + 497, + 532, + 505 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Q: Ricky is considering purchasing an apartment costing \\( \\\\( {700},{000} \\) . He will pay a 30% down payment and take out a mortgage for the remainder. Since he just got married and wants to save some money for future use, he will choose the plan with the lowest monthly payment. After visiting several banks, he received the following mortgage offers: What is the monthly payment for Bank A? Answer to two decimal places.", + "bbox": [ + 377, + 508, + 609, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": ":", + "text_level": 1, + "bbox": [ + 379, + 582, + 416, + 590 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/e0b2f16f8afad2b055de20658573bdc7e1ed2ba9c07e7edbe3012dd267e4874e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BankInterest rateTerm (years)
A3.5%15
B320
C425
D4.518
", + "bbox": [ + 385, + 594, + 607, + 674 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Complex Formulas Challenge Numerical Computation", + "text_level": 1, + "bbox": [ + 645, + 95, + 883, + 125 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Q: What is the total weighted average Cash and Payment-In-Kind (PIK) interest rate payable under the subordinated and senior notes portfolio in the year 2023?", + "bbox": [ + 651, + 136, + 877, + 176 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": ":", + "text_level": 1, + "bbox": [ + 653, + 179, + 694, + 188 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/071fb689a0aaef17aff21a3b43919fb6d114e822259edf1f903d29f34753a1c4.jpg", + "table_caption": [ + "Monthly Payment" + ], + "table_footnote": [], + "table_body": "
EUR
In Mio €201820172016%2015
Deutschland3.498.5003.498.5003.498.500100%100%
USA1.200.0001.200.0001.200.000--
Nordafrika, Naher Osten1.200.0001.200.0001.200.000--
Asien-Pazifik1.200.0001.200.0001.200.000--
davon1.200.0001.200.0001.200.000--
davon1.200.0001.200.0001.200.000--
davon1.200.0001.200.001.200.000--
davon1.200.0001.200.0001.200.000--
davon1.200.0001.200.0001.200.000--
deutschland3.498.5003.498.5003.498.500100%100%
davon1.200.0001.200.0001.200.000--
", + "bbox": [ + 661, + 191, + 870, + 234 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": ":", + "text_level": 1, + "bbox": [ + 653, + 247, + 702, + 256 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/ddf3c43dd63471cdf2c551056206f8665b1b489ee3598c16488fda12ee8959c6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
[average = (investmenti × cashi) / ∑investmenti] + (investmenti × PIKi) / ∑investmenti]
", + "bbox": [ + 661, + 258, + 870, + 284 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Total Weighted Average Cash and", + "text_level": 1, + "bbox": [ + 687, + 305, + 846, + 313 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Payment-In-Kind (PIK) Interest Rate", + "text_level": 1, + "bbox": [ + 681, + 314, + 849, + 323 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Q: John oversees a fund, with the returns for the first three years displayed below: What will be the holding period return (expressed as a percentage)? Answer to three decimal places.", + "bbox": [ + 651, + 327, + 880, + 366 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": ":", + "text_level": 1, + "bbox": [ + 653, + 369, + 694, + 378 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/2b31eb1cea01af1f8b62652401cdbd73b9636e36bdb39fc7386c6513f67f6ba8.jpg", + "table_caption": [ + "Answer: 12.0" + ], + "table_footnote": [], + "table_body": "
YearInvestmentReturn
1$ 50012%
2$ 6005%
3$ 10001%
", + "bbox": [ + 666, + 383, + 867, + 430 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": ":", + "text_level": 1, + "bbox": [ + 653, + 438, + 702, + 446 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ H P R = \\left(\\frac {\\sum \\left(\\text {i n v e s t m e n t} _ {t} \\times (1 + \\text {r e t u r n} _ {t})\\right)}{\\sum \\text {i n v e s t m e n t}} - 1\\right) \\times 1 0 0 \\right]\n$$\n", + "text_format": "latex", + "bbox": [ + 676, + 450, + 852, + 476 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Answer: 4.762", + "text_level": 1, + "bbox": [ + 808, + 482, + 874, + 489 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Holding Period Return (HPR)", + "text_level": 1, + "bbox": [ + 702, + 497, + 830, + 506 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Q: What is the average quarterly share price in 2021 assuming each quarter had a completely uniform price distribution in high and low bid prices, measured in dollars?", + "bbox": [ + 655, + 508, + 870, + 550 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": ":", + "text_level": 1, + "bbox": [ + 655, + 555, + 696, + 564 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/69819e936c14322d348f4d9eeeb40f6090f6291b91df1508612a22b649c9831a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Quarter EndedHigh BidLow Bid
October 31, 2021 (1)$4.51$3.99
July 31, 2021$8.47$2.30
April 30, 2021$13.00$3.80
January 31, 2021$14.50$0.015
October 31, 2020 (2)$0.199$0.01
", + "bbox": [ + 668, + 565, + 867, + 616 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": ":", + "text_level": 1, + "bbox": [ + 656, + 623, + 702, + 631 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n[ \\text {A v e r a g e} = \\frac {1}{4} \\sum_ {i = 1} ^ {4} \\frac {\\text {H i g h} _ {i} + \\text {L o w} _ {i}}{2} ]\n$$\n", + "text_format": "latex", + "bbox": [ + 687, + 633, + 826, + 659 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Answer: 6.32", + "text_level": 1, + "bbox": [ + 812, + 660, + 872, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Average Quarterly Price", + "text_level": 1, + "bbox": [ + 715, + 675, + 823, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "cial documents, extract key financial indicators from tables, images, and texts, and perform multi-step precise numerical calculations to support professional decision-making. Similarly, to achieve expert artificial general intelligence (AGI) [8, 25, 42, 43, 65], MLLMs are expected to comprehend complex domain-specific images akin to human experts, and apply domain knowledge to perform accurate numerical reasoning. This raises the question: Can current MLLMs seamlessly integrate vision and text to perform", + "bbox": [ + 89, + 758, + 483, + 895 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "domain-specific complex reasoning, matching the proficiency of LRMs in pure text-based tasks?", + "bbox": [ + 511, + 758, + 903, + 789 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, we choose the financial domain to evaluate the complex reasoning capabilities of MLLMs, where precision and transparent reasoning are paramount [32]. Existing numerical reasoning benchmarks for finance are limited in their text-based reasoning, coverage of specific financial knowledge, and complexity of reasoning [14, 15, 32, 67, 69]. FAMMA [63] is mainly modelled after textbooks", + "bbox": [ + 511, + 795, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3246", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and CFA exam questions, MathVista [39] does not involve the application of financial knowledge, MMMU [65] and MMMU-Pro [66] are restricted to multiple-choice questions, diverging from authentic financial problem-solving. The lack of high-quality, knowledge-intensive multimodal financial numerical reasoning benchmarks makes it challenging to objectively evaluate the actual reasoning capabilities of MLLMs and analyze their shortcomings.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Therefore, we propose FinMMR, a bilingual multimodal numerical reasoning benchmark designed to evaluate the reasoning capabilities of MLLMs in the finance domain. The benchmark comprises 4.3K problems, covering 14 financial subdomains (e.g., corporate finance and industry analysis), with 8.7K images derived from 14 categories (e.g., tables and ownership structure charts). Each problem includes rich images, an unambiguous question, a Python-formatted solution, and a precise answer.", + "bbox": [ + 89, + 213, + 480, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For multimodality, without representing financial tables as structured text, FinMMR represent all tables, charts, and diagrams as images. For comprehensiveness, FinMMR covers 14 financial subdomains and two languages (i.e., English and Chinese), demanding domain knowledge such as option pricing and portfolio management. For challenge, FinMMR focus on multi-step numerical reasoning, requiring models to provide exact numerical answers under strict evaluation criteria (emphasizing units, percentages, and decimal places). Furthermore, we mix each Chinese questions with two distractor images that are contextually adjacent to the ground images, approaching real-world multimodal reasoning scenarios.", + "bbox": [ + 89, + 352, + 482, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We evaluate 15 state-of-the-art (SOTA) MLLMs [18-21, 46-48, 56], utilizing Chain-of-Thought (CoT) [60] and Program-of-Thought (PoT) [13]. The experimental results on FinMMR reveals three key findings:", + "bbox": [ + 89, + 550, + 482, + 611 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- MLLMs Face Significant Challenges in Domain-Specific Multimodal Numerical Reasoning: All evaluated models underperform on FinMMR. On the Hard test set, the best-performing model, Claude 3.7 Sonnet with 64K extended thinking, achieves only $53.00\\%$ accuracy, while OpenAI o1 achieves merely $48.40\\%$ . Through error analysis, we identify that visual perception, knowledge reasoning, and numerical computation collectively pose challenges to MLLMs. Current MLLMs still struggle with complex multimodal reasoning tasks in specialized domains, compared to text-based reasoning.", + "- Better Synergy Between Visual Perception and Complex Reasoning Is Needed: Distracting images lead to a greater than $10\\%$ drop in accuracy for Qwen2.5-VL-72B compared to ground images alone, indicating that irrelevant visual information severely impacts multimodal reasoning. By decoupling visual filtering and reasoning, Qwen2.5-VL-72B improved from $64.73\\%$ to $71.56\\%$ . Combining MLLMs with LRMs, by efficiently" + ], + "bbox": [ + 89, + 613, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Parsing visual information into structured text and leveraging the LRMs' text-based reasoning capabilities, can also yield better performance. The combination of GPT4o and DeepSeek-R1 achieves $86.72\\%$ accuracy on 1,160 tabular QA instances, outperforming Claude 3.7 Sonnet $(83.53\\%)$ .", + "bbox": [ + 526, + 90, + 903, + 180 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Refined Structured Domain Knowledge Enhances MLLMs' Complex Reasoning: Leading MLLMs lack sufficient experience in applying rich domain knowledge when solving complex reasoning tasks. By annotating structured financial functions and leveraging the model's ability to generate retrieval questions and make judgments, knowledge augmentation can significantly improve MLLMs' performance. MLLMs achieve absolute improvements ranging from 2.76 to 4.31 percentage points, allowing weaker models to approach SOTA performance, while SOTA models exhibit further gains.", + "bbox": [ + 513, + 181, + 903, + 347 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "These findings highlight the bottlenecks of MLLMs in complex multimodal reasoning tasks in expert domains closer to real-world scenarios. They emphasize the need for continuous improvements in three key areas: more intricate visual perception, more specialized knowledge reasoning, and more accurate numerical computation. Alternatively, leveraging tools or model combinations can help achieve a balance between performance and cost, enabling MLLMs to perform expert-level reasoning tasks like human experts.", + "bbox": [ + 511, + 348, + 903, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 500, + 653, + 515 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. LRMs and MLLMs", + "text_level": 1, + "bbox": [ + 513, + 525, + 700, + 540 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "By integrating train-time scaling and test-time scaling [31, 45], LRMs have demonstrated remarkable reasoning capabilities [62]. However, most LRMs are limited to handling text-based problems. The growing demand for solving real-world tasks has spurred the development of MLLMs [5, 19, 55] and benchmarks designed to evaluate the perception and reasoning abilities of MLLMs [10, 23, 27, 30, 33, 36, 38, 64-66]. For instance, MME-CoT [30] evaluates models' space-time understanding, while EMMA [27] focuses STEM subjects. Following this trend, domain-specific benchmarks which require deep domain expertise have emerged, such as MathVista [39] for mathematics and GMAI-MMBench [12] for medicine. Yet, financial reasoning remains an unexplored area in the current landscape of MLLM benchmarks.", + "bbox": [ + 511, + 547, + 906, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Financial Benchmarks", + "text_level": 1, + "bbox": [ + 513, + 787, + 723, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The financial domain presents a distinct and more formidable set of challenges for model evaluation, which arise from its inherent complexity, information density, and dependence on expertise. The majority of existing text-only financial numerical reasoning benchmarks [14, 15, 32, 67-69] are constrained by limitations such as sub-optimal an", + "bbox": [ + 511, + 810, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3247", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/dc0f3a27e22ff74eb53812730eabd607fbc90cff709eb466a4850619c94679b8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PropertyValue
# Total Questions4,300
# Easy/Medium/Hard1,300/1,500/1,500
# Validation/Test900/3,400
# Chinese/English2,150/2,150
# Operators (Easy/Medium/Hard)1.75/2.97/5.34
# Lines of Code (Easy/Medium/Hard)4.14/5.06/7.34
# Parentheses (Easy/Medium/Hard)0.88/3.11/4.25
# Difficulty (Easy/Medium/Hard)1.93/2.96/3.79
", + "bbox": [ + 91, + 88, + 501, + 247 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Key statistics of FinMMR (Avg values of three subsets).", + "bbox": [ + 91, + 257, + 478, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "notation quality, narrow domain knowledge coverage, and overly simplistic reasoning tasks. Although FinanceReasoning [53] offers complex tasks with rich knowledge and high-quality annotations, its text-only modality limits multimodal reasoning evaluation.", + "bbox": [ + 89, + 300, + 482, + 376 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recent multimodal financial benchmarks have sought to bridge this gap but still possess limitations. FAMMA [63] being sourced from textbooks and examinations does not mirror the real-world tasks. FinMME [40] uses a multiple-choice format, which may overestimate model reasoning due to guesswork. MME-Finance [24] is constrained by coarse annotations and an isolated assessment of domain knowledge, limiting holistic evaluation of real-world financial reasoning.", + "bbox": [ + 89, + 377, + 483, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. FinMMR Benchmark", + "text_level": 1, + "bbox": [ + 89, + 530, + 297, + 546 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Overview of FinMMR", + "text_level": 1, + "bbox": [ + 89, + 555, + 297, + 570 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We introduce FinMMR, a bilingual (English and Chinese) multimodal benchmark for evaluating the reasoning capabilities of MLLMs in financial numerical reasoning tasks. FinMMR consists of 4,300 questions covering 14 financial subdomains (e.g., corporate finance, industry analysis, financial markets, asset management). The key statistics are summarized in Tab. 1, and the composition of sub-domains and images is illustrated in Fig. 1. As illustrated in Fig. 2, FinMMR introduces three key challenges:", + "bbox": [ + 89, + 578, + 482, + 715 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Rich Images Challenge Visual Perception: FinMMR comprises 8.7K images from 14 categories (e.g., bar charts, line charts, ownership structure charts). MLLMs must identify relevant images among distractors and extract critical information from the correct images.", + "- Comprehensive Subdomains Challenge Knowledge Reasoning: MLLMs need to flexibly apply diverse domain-specific financial knowledge from 14 subdomains to solve multi-step reasoning tasks.", + "- Complex Formulas Challenge Numerical Computation: All questions require precise numerical answers, eliminating the potential bias from lucky guesses that" + ], + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "could occur in multiple-choice formats.", + "bbox": [ + 527, + 90, + 790, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Data Curation Process", + "text_level": 1, + "bbox": [ + 513, + 119, + 723, + 135 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first curated a subset of questions from public text-based financial reasoning benchmarks and systematically transformed them into multimodal problems using a unified standard. Subsequently, we constructed a novel multimodal Chinese Research Report Question Answering (CR-RQA) benchmark from scratch, merging two data sources into FinMMR. Each question is accompanied by an executable Python solution, yields a numerical answer and demonstrates a clear reasoning chain.", + "bbox": [ + 511, + 143, + 906, + 279 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Update to Public Benchmarks We re-announce 124 and 163 financial questions from MMMU [65] and MMMU-Pro [66], respectively. Following rigorous verification, these questions were directly incorporated into our dataset. Furthermore, we extracted 77, 288, and 795 questions from FinanceMath [67], CodeTAT-QA [32], and CodeFinQA [32], respectively. From DocMath-Eval [68], we further obtained 703 questions from its four subsets. For each question, we rendered any tabular data as images while removing the corresponding table information from the text, ensuring that MLLMs cannot rely solely on textual content.", + "bbox": [ + 511, + 279, + 906, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Building a Novel Dataset from Scratch We collect 90 research reports, all of which are obtained through authorized access and cover diverse topics such as industry research, macroeconomic analysis, and strategy research. We use 360LayoutAnalysis [1] to extract informative images and discard those lacking explicit numerical data, reducing ambiguity. For each retained image, we prompt Qwen-VLMax [6] to formulate questions requiring multiple reasoning steps or complex calculations. Each question is accompanied by an executable Python solution and a definitive numerical answer.", + "bbox": [ + 511, + 446, + 905, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Furthermore, we introduce distractor images sourced from the same reports adjacent to ground images to challenge MLLMs in extracting relevant numerical information from structured, densely packed visuals.", + "bbox": [ + 511, + 613, + 905, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Data Quality Assurance This process ensured that every question was clearly written, featured a detailed reasoning solution, and included an accurate final answer. The annotators included 16 graduate students in finance and two experts holding CFA certifications. With the aid of LLMs, this meticulous verification process spanned three months, culminating in a dataset that meets high standards of clarity and correctness.", + "bbox": [ + 511, + 674, + 905, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dataset Splitting and Release To classify the problems by difficulty, we employ a heuristic algorithm that takes into account the number of operators $(o)$ , code lines $(l)$ and parentheses pairs $(p)$ in the Python solution. Specifically, the difficulty of reasoning $rc$ of a problem is defined as:", + "bbox": [ + 511, + 795, + 905, + 872 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nr c = \\ln (\\max (o, 1)) + \\ln (\\max (l + p, 1)) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 885, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3248", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/36bf93ff9009eed725d220edcdd51597f18cfd498d2cf7ea7c772bb200c376b8.jpg", + "image_caption": [ + "Figure 3. The comparison between finance-related benchmarks. These benchmarks vary in size, domain coverage, modalities, and question type, with some focusing on text-only data while others include images. Each axis has scale labels with varying ranges to measure the number of questions from each benchmark across different subdomains. In the Modalities, T means text input, I means Images input, P.I. means pure images input. In the Question Type, MC means multi-choice answer, NUM means numerical answer." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 395, + 257 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/41f33f9be512944cdddffd9bd2c6513d27bdc58f56a962c25a2802bf3828c27c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BenchmarkSize (Fin)Domain CoverageModalitiesQuestion Type
MMMU [65]11,550 (1,603)10T+IMC
MMMU-Pro [66]1,730 (286)10T+I, P.I.MC
FinanceMath [67]1,200 (1,200)10TNUM
CodeTAT-QA [32]3,144 (3,144)6TNUM
CodeFinQA [32]5,463 (5,463)13TNUM
DocMath-Eval [68]4,000 (4,000)12TNUM
FAMMA [63]1,758 (1,758)8T+IMC, NUM
CRRQA (ours)2,150 (2,150)13T+I, P.I.NUM
FinMMR (ours)4,300 (4,300)14T+I, P.I.NUM
", + "bbox": [ + 418, + 88, + 913, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "FinMMR is classified as Easy (1,300 examples), Medium (1,500 examples) and Hard (1,500 examples) based on this formula, with each level randomly split into test and validation sets. All questions are publicly available, while only the 300 validation answers per level are released, while test answers remain private to prevent data leakage [22, 50, 52]. We maintain an online evaluation platform that enables researchers to assess their models.", + "bbox": [ + 89, + 351, + 483, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Comparisons with Existing Benchmarks", + "text_level": 1, + "bbox": [ + 89, + 484, + 436, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As illustrated in Fig. 3, previous work has studied multidiscipline multimodal reasoning [65, 66], general mathematical reasoning [39] or text-based financial QA [32, 67, 68]. FinMMR focus on multimodal financial numerical reasoning, curating 4,300 questions requiring a deep understanding of domain-specific images (e.g., earnings reports, candlestick charts). To mimic real-world scenario, we deliberately incorporate 3,938 distractors into 2,150 questions to rigorously evaluate MLLMs' visual perception capability. Compared to existing financial benchmarks, they suffer from narrow domain coverage [14, 69]. FinMMR encompasses 14 financial subdomains and 14 image categories, comprehensively evaluating MLLMs' domain-specific reasoning capabilities.", + "bbox": [ + 88, + 506, + 482, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Evaluation System", + "text_level": 1, + "bbox": [ + 89, + 734, + 272, + 752 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To facilitate the evaluation of complex reasoning on FinMMR, we established a dedicated evaluation system. All MLLMs evaluated were accessed through official APIs.", + "bbox": [ + 89, + 760, + 482, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Multimodal Large Language Models", + "text_level": 1, + "bbox": [ + 89, + 816, + 408, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We systematically evaluate the multimodal reasoning capabilities of 15 recent MLLMs under the zero-shot setting. The evaluated MLLMs are: OpenAI o1 [46], GPT-4o [44], Claude 3.7 Sonnet (including thinking mode) [5],", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Gemini 2.0 Flash Thinking [19], Gemini 2.0 Pro [20], Gemini 2.0 Flash [18], InternVL2.5-78B [16], Grok 2 Vision [61], Pixtral Large [2], Qwen2.5-VL-72B [7], QVQ72B-Preview [55], Qwen-Omni-Turbo [57], Llama 4 Maverick [4], Gemma 3 27B [17], and Mistral Small 3.1 [3].", + "bbox": [ + 511, + 351, + 905, + 429 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Prompting Methods", + "text_level": 1, + "bbox": [ + 511, + 436, + 705, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our evaluation adopts CoT [60], PoT [13] and IO (without any external prompts) prompting methods. Due to budget constraints, we report OpenAI o1 performance on the Hard subset only. Detailed prompts can be found in the Appendix.", + "bbox": [ + 511, + 459, + 905, + 536 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Answer Extraction and Evaluation", + "text_level": 1, + "bbox": [ + 511, + 547, + 816, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following Zhao et al. [67], we extract answers based on the prompting methods. For CoT and IO outputs, we employ GPT-4o-mini to extract numerical answers. For PoT, we run the generated program for numerical results. Finally, we conduct a strict accuracy evaluation, comparing numerical results with ground truth and deeming the results accurate within a stringent error tolerance of $0.2\\%$ .", + "bbox": [ + 511, + 569, + 905, + 675 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 690, + 645, + 707 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We answer the following research questions (RQs): RQ1: Are MLLMs multimodal reasoners with extended thinking and PoT prompts? RQ2: What are the primary challenges facing MLLMs? RQ3: How can the visual perception difficulties of MLLMs be mitigated? RQ4: How can the knowledge reasoning capabilities of MLLMs be enhanced? RQ5: How can the numerical computation abilities of MLLMs be compensated for?", + "bbox": [ + 511, + 715, + 905, + 837 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Main Results (RQ1)", + "text_level": 1, + "bbox": [ + 511, + 847, + 705, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The performance of the MLLMs evaluated using different prompting methods on FinMMR is shown in Tab. 2.", + "bbox": [ + 511, + 869, + 905, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3249", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a4ae46a275dd9df3d6778cf1390f937dddc281f452ae7eeaec7c3f21f12728ac.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelSizeExtended thinkingHardMediumEasyAvg.Token (M)
IOCoTPoTCoTPoTCoTPoTCoTPoTCoTPoT
Proprietary MLLMs
Claude 3.7 Sonnet✓ (64K)53.0051.0051.4062.5062.1778.5078.5064.0064.028.5111.25
Claude 3.7 SonnetX49.8050.8048.5062.2558.8377.0076.9263.3561.420.990.89
OpenAI o1✓48.0048.4044.70------2.522.12
GPT-4oX-45.4047.8063.3359.9278.0076.0062.2461.240.850.41
Gemini 2.0 ProX-46.5047.3060.5857.9275.5075.6760.8660.300.850.45
Gemini 2.0 Flash Thinking✓-46.0046.0060.7556.5877.1774.1761.3158.921.300.48
Gemini 2.0 FlashX-44.4045.9057.8353.4274.9273.7559.0557.690.790.43
Grok 2 VisionX-27.8025.5041.5035.8373.0872.8347.4644.721.130.60
Qwen-Omni-TurboX-17.5027.3035.8348.0057.5061.6736.9445.660.900.42
Open-source MLLMs
Llama 4 Maverick17BX-48.7047.8063.2559.1777.8377.8363.2661.600.880.47
Qwen2.5-VL-72B72BX-43.3046.2063.4264.1777.4275.8361.3862.071.050.44
InternVL2.5-78B78BX-37.4044.0060.5061.1770.9270.5856.2758.58--
QVQ-72B-Preview72B✓43.3040.306.2055.679.6775.4212.4257.139.435.435.70
Pixtral Large124BX-19.7025.0039.8339.7570.0070.1743.1844.971.150.75
Gemma 3 27B27BX-23.4022.3045.1736.4269.0861.5845.8840.100.970.47
Mistral Small 3.124BX-19.7015.2038.4229.7567.6749.4241.9331.461.150.60
", + "bbox": [ + 98, + 88, + 906, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Results of different models using IO, CoT and PoT prompting methods on the test set of FinMMR. We use the best Accuracy on the Hard subset as the ranking indicator of model performance. The results underscore the superior performance of reasoning-enhanced MLLM (i.e., Claude 3.7 Sonnet with 64K extended thinking) with PoT in complex multimodal numerical reasoning task.", + "bbox": [ + 89, + 361, + 906, + 405 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Challenges of MLLMs in Domain-Specific Complex Numerical Reasoning As the difficulty increases, the average accuracy shows a continuous and significant decline. In CoT setting, the average accuracy rates on the Easy, Medium, and Hard subsets are $73.33\\%$ , $54.05\\%$ , and $38.14\\%$ , respectively. The currently best-performing model (i.e., Claude 3.7 Sonnet with 64K extended thinking) consistently demonstrates superior performance across all difficulty subsets using CoT prompting method. However, its accuracy on the Hard subset remains below the $60\\%$ passing threshold in both prompting methods. On the overall test set, Claude 3.7 Sonnet achieves only $64.02\\%$ accuracy. These results highlight the challenging nature and rigorous standards of FinMMR, effectively assessing the limits of MLLMs' reasoning capabilities and the disparities among models compared to previous benchmarks.", + "bbox": [ + 88, + 429, + 480, + 671 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Does extended thinking help? Reasoning-enhanced models show consistent improvements, compared with non-reasoning-enhanced MLLMs. Claude 3.7 Sonnet with 64K extended thinking achieves 2.20 percentage points higher accuracy than the model without extended thinking (53.00% vs. 50.80%) on the Hard subset. This enhancement comes at the cost of using nearly 12 times more tokens for intricate thinking (4.06M vs. 0.34M). This trend also persists in the Gemini 2.0 Flash series.", + "bbox": [ + 88, + 672, + 480, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We observe that QVQ-72B-Preview lose basic code generation capabilities due to the reinforcement learning of text-based long thought, which is likely attributed to biases in training strategies and training data. On the Hard subset, this model achieves a code execution success rate of only $10.90\\%$ , resulting in an accuracy of merely $6.20\\%$ in", + "bbox": [ + 88, + 810, + 480, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "PoT setting, significantly lower than the $40.30\\%$ accuracy achieved with CoT. This finding highlights the importance of maintaining foundational capabilities, such as programming, while enhancing the reasoning abilities of MLLMs, to avoid rendering them ineffective in other basic tasks.", + "bbox": [ + 511, + 429, + 903, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Does PoT help? Experimental results strongly validate the superiority of PoT over CoT in numerical reasoning tasks, especially on the Hard subset. PoT (not including QVQ-72B-Preview) achieves a mean accuracy of $37.64\\%$ versus $36.20\\%$ for CoT, representing an improvement of 1.44 percentage points. Furthermore, PoT encourages MLLMs to leverage structured code generation to reduce token consumption during reasoning. Under similar or lower token usage, PoT achieves similar or greater accuracy. For instance, GPT-4o achieves a 2.40 percentage points improvement in accuracy over CoT while consuming significantly fewer tokens in PoT setting. Similarly, Qwen2.5-VL-72B demonstrates the most pronounced efficiency gains: PoT improves accuracy to $64.17\\%$ from $63.42\\%$ while reducing token consumption by $58.88\\%$ (153K vs. 373K) on the Medium subset. When addressing complex numerical reasoning problems, PoT avoids precise numerical calculations by utilizing external tools (i.e., Python interpreter) and reduces the need for repetitive text-based reasoning, which is beneficial for most MLLMs.", + "bbox": [ + 511, + 506, + 903, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "However, we also observe that for certain reasoning-enhanced models (i.e., Claude 3.7 Sonnet with 64K extended thinking and OpenAI o1), due to the enforced requirement for long thought, they still engage in extensive text-based reasoning before generating code even in PoT setting, resulting in exceptionally high token consumption", + "bbox": [ + 511, + 810, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3250", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/70fcafdd75c30a4b18a32ddcf90e0832d04c8c27aaff09172eecffd3558cac18.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SubsetTestDegradationValidationDegradation
Ground Images (%)Distractor Images (%)Ground Images (%)Distractor Images (%)
Hard57.1847.23↓9.9556.7445.58↓11.16
Medium73.0161.36↓11.6577.7864.73↓12.35
Easy61.5953.64↓7.9560.6151.52↓9.09
The improvement achieved by the filtering-reasoning pipeline on the Medium validation set: 64.73 → 71.56 ↑6.83
", + "bbox": [ + 94, + 88, + 906, + 189 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Degradation of Qwen2.5-VL-72B on all subsets due to distractor images and improvement achieved by the filtering-reasoning pipeline on the Medium validation set in PoT setting.", + "bbox": [ + 89, + 199, + 906, + 227 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "on the Hard subset (4.48M and 2.12M), which is more than 10 times that of other MLLMs. To further investigate this, we added an IO baseline without any external prompts for reasoning-enhanced models on the Hard subset. The IO group achieved the highest accuracy, which we attribute to the comprehensive built-in system prompts embedded in the tested proprietary models. This highlights the need for future research to balance reasoning performance with the control of inefficient and redundant token generation, aiming to achieve a good trade-off between performance and cost, as well as to investigate whether PoT prompting can yield significant performance gains on open-source reasoning-enhanced models.", + "bbox": [ + 88, + 253, + 480, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Error Analysis (RQ2)", + "text_level": 1, + "bbox": [ + 89, + 462, + 294, + 478 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To better analyze the capabilities and limitations of MLLMs on FinMMR, we conduct a detailed error analysis for the Claude 3.7 Sonnet with 64K extended thinking in PoT setting. Error analysis is based on 100 sampled failure cases, which we categorize into three main error types, some of which involve compound errors. More details of error cases are provided in the Appendix.", + "bbox": [ + 89, + 484, + 482, + 590 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Visual Perception Error (30/100): The model incorrectly perceives, identifies, or interprets visual information from images, or mistakenly recognizes incorrect data, subsequently causing errors in calculations, broken reasoning chains, or incorrect conclusions.", + "- Knowledge Reasoning Error (38/100): Due to insufficient domain-specific knowledge, the model exhibits logical confusion or conceptual misunderstandings during reasoning, leading to incorrect answers.", + "- Numerical Computation Error (32/100): In problems involving mathematical operations and numerical reasoning, the model produces significant deviations from the correct answers due to errors in the calculation steps, precision control, or numerical hallucination." + ], + "bbox": [ + 89, + 594, + 482, + 805 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Visual Filtering for Reasoning (RQ3)", + "text_level": 1, + "bbox": [ + 89, + 816, + 410, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Tab. 3, when processing multi-image inputs containing distractor images, Qwen2.5-VL-72B demonstrates significantly lower accuracy across all difficulty levels compared to ground images scenarios. This finding", + "bbox": [ + 89, + 839, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "aligns with conclusions from previous work [37, 51], indicating that irrelevant visual information substantially interferes with MLLMs' reasoning capabilities. In particular, the Medium validation set exhibits the most pronounced performance drop (77.78% ground images vs. 64.73% distractor images), attributed to two key characteristics: (1) moderate complexity making visual perception quality the performance bottleneck; (2) semantic relevance between distractors and questions increasing visual filtering difficulty. To address distractor image interference, we propose a two-stage multimodal reasoning pipeline:", + "bbox": [ + 511, + 253, + 903, + 419 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Visual Filtering: We first instruct MLLM to analyze the set of images $\\mathcal{I}$ and the question $q$ , assessing the relevance of the image (relevant/irrelevant). Irrelevant images are excluded from subsequent reasoning.", + "- Enhanced Reasoning: Then, the filtered subset $\\mathcal{I}'$ and the question $q$ are input into the MLLM for the final reasoning. The system automatically reverts to the original inputs $\\mathcal{I}$ if all images are mistakenly filtered." + ], + "bbox": [ + 513, + 421, + 903, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Does the pipeline help? As illustrated in Tab. 3, we evaluate Qwen2.5-VL-72B on the 207 English questions with distractor images of the Medium validation set. Our method achieves an overall accuracy of $71.56\\%$ , representing a 6.83 percentage points improvement over direct reasoning. This result is only 6.22 percentage points away from the ideal accuracy of the ground images scenarios $(77.78\\%)$ . Detailed analysis reveals $73.43\\%$ (152/207) ground image recognition accuracy during filtering. When correctly identified, the accuracy of these problems increases to $81.58\\%$ (124/152). This finding underscores the necessity to enhance the ability of MLLMs to filter out irrelevant image information, thereby strengthening their robustness in reasoning within more complex real-world scenarios.", + "bbox": [ + 511, + 542, + 906, + 755 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Knowledge Augmentation (RQ4)", + "text_level": 1, + "bbox": [ + 511, + 767, + 802, + 782 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To enhance the understanding and application capabilities of financial knowledge of MLLMs, we explore a method of refined knowledge augmentation to improve the performance of MLLM in domain-specific reasoning tasks.", + "bbox": [ + 511, + 790, + 903, + 851 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "- Function Library Construction: We annotate a comprehensive financial function library containing 3,133 Python functions from financial encyclopedia. Each func", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3251", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bcbd361d99c7192321ef46924bd2a5f196041c77b4b8b684e226f6ca454c3bad.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SettingPoTRAG with PoT
Gemini 2.0 Flash Thinking78.7183.02 (+4.31)
GPT-4o80.6083.62 (+3.02)
Claude 3.7 Sonnet81.2185.43 (+4.22)
Claude 3.7 Sonnet (64K)83.5386.29 (+2.76)
", + "bbox": [ + 96, + 88, + 478, + 183 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Improvements of different MLLMs with knowledge augmentation on the 1,160 instances of FinMMR in PoT setting.", + "bbox": [ + 89, + 191, + 482, + 223 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tion includes precise functional descriptions, parameter explanations, and step-by-step implementation code.", + "bbox": [ + 102, + 247, + 482, + 277 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- MLLM-Instructed Knowledge Retrieval: In financial problems with hybrid contexts, using short questions or full contexts for retrieval often fails to retrieve directly relevant knowledge [9, 49]. We observed that powerful MLLMs (e.g., GPT-40) can effectively summarize rich semantic information from contexts. Therefore, we first prompt the MLLM to generate precise retrieval queries based on the context [34, 58]. Then we use Contriever [28] to retrieve relevant financial Python functions, based on the semantic similarity between the refined queries and functional descriptions.", + "- MLLM as Retrieval Judge: Recent studies have shown that models are capable of judging the relevance of candidates retrieved for the question [26]. In this setting, we first retrieved the Top-30 financial functions and then prompted the MLLM to select the Top-3 functions most useful to answer the question, if any." + ], + "bbox": [ + 89, + 277, + 483, + 536 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Does knowledge augmentation help? As shown in Tab. 4, all evaluated MLLMs enhanced with financial function knowledge achieved significant improvements, ranging from 2.76 to 4.31 percentage points. Leveraging the improved retrieval efficiency enabled by MLLM-Instructed Knowledge Retrieval and MLLM as Retrieval Judge, the knowledge augmentation approach achieves greater performance, boosting the accuracy of Claude 3.7 Sonnet with 64K thinking to $86.29\\%$ . Notably, Gemini 2.0 Flash Thinking, which has relatively weaker reasoning capabilities, also improved from $78.71\\%$ to $83.02\\%$ , approaching the performance of Claude 3.7 Sonnet $(83.53\\%)$ without knowledge augmentation. The results further illustrate that refined domain-specific reasoning knowledge can significantly enhance the performance of MLLMs in complex reasoning tasks within expert domains.", + "bbox": [ + 88, + 537, + 482, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Visual Parser with Reasoner (RQ5)", + "text_level": 1, + "bbox": [ + 89, + 787, + 397, + 804 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In complex multimodal numerical reasoning tasks, single models often struggle to simultaneously achieve visual perception, knowledge reasoning, and numerical computation. To investigate the potential of model collaboration, we instruct the MLLM to act as the Visual Parser, responsible for carefully converting images into structured textual data.", + "bbox": [ + 89, + 809, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7d6557cfa9525f56834045e280e4a6b9d3f2d22c08400430bd46e0c9f475df53.jpg", + "image_caption": [ + "Figure 4. Results of model combinations and individual models." + ], + "image_footnote": [], + "bbox": [ + 514, + 88, + 905, + 237 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Then, an LRM acts as the Reasoner, performing multi-step numerical reasoning based on the textual context.", + "bbox": [ + 511, + 287, + 903, + 318 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Specifically, we filter 1,160 tabular QA instances from FinMMR and utilize GPT-4o as the Visual Parser, instructing it to separate table headers or cells with vertical bars (|) and rows with newlines. For the Reasoner component, in addition to GPT-4o, we evaluate several LRMs (i.e., Claude 3.7 Sonnet, Gemini 2.0 Flash Thinking, DeepSeek-R1, and OpenAI o1).", + "bbox": [ + 511, + 319, + 903, + 424 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Does model collaboration help? As shown in Fig. 4, the structured data from GPT-4o's visual parsing significantly enhances downstream reasoning. The individual model (i.e., GPT-4o with PoT) achieves an accuracy of $80.60\\%$ , while the combination of models improves the accuracy to $86.72\\%$ (i.e., DeepSeek-R1 serving as the Reasoner with PoT). Performance variance emerges across reasoning-enhanced models using identical visual inputs. Claude 3.7 Sonnet reaches $85.52\\%$ , outperforming Gemini 2.0 Flash Thinking $(82.24\\%)$ , confirming the decisive impact of text-based reasoning capabilities. This evidences that model collaboration effectively compensates for individual model limitations through complementary strengths.", + "bbox": [ + 511, + 424, + 905, + 635 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 648, + 633, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce FinMMR, a multimodal, comprehensive, and challenging benchmark for evaluating the financial numerical reasoning capabilities of MLLMs. FinMMR challenges MLLMs' intricate visual perception, specialized knowledge reasoning, and accurate numerical computation through its rich images, comprehensive subdomains, and complex formulas embedded in each multimodal financial question. The evaluation results reveal that 15 proprietary and open-source MLLMs still struggle with complex multimodal reasoning tasks in specialized domains. FinMMR highlights the bottlenecks of MLLMs and the need for continuous improvements, including reasoning-enhanced training, tool use, refined structured knowledge augmentation and model combinations, allowing models to perform expert-level reasoning tasks closer to the real-world scenarios.", + "bbox": [ + 511, + 674, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3252", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported by the National Natural Science Foundation of China (Grant Nos. 62176026, 62473271), the Beijing Natural Science Foundation (Grant No. QY24214), the BUPT Innovation and Entrepreneurship Support Program (Grant Nos. 2025-YC-A033, 2025-YC-A042), and data support from Hithink RoyalFlush Information Network Co., Ltd. This work is also supported by the Engineering Research Center of Information Networks, Ministry of Education, China. We would also like to thank the anonymous reviewers and area chairs for constructive discussions and feedback.", + "bbox": [ + 89, + 113, + 485, + 267 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 295, + 187, + 311 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] 360AILAB-NLP. 360layoutanalysis. https://github.com/360AILAB-NLP/360LayoutAnalysis, 2024. Accessed: 2024-06-15. 4", + "[2] Mistral AI. Pixtral large. https://mistral.ai/news/pixtral-large, 2024. Accessed: 2024-11-18. 5", + "[3] Mistral AI. Mistral small 3.1. https://mistral.ai/news/mistral-small-3-1, 2025. Accessed: 2025-03-17. 5", + "[4] AI@Meta. The llama 4 herd: The beginning of a new era of natively multimodal ai innovation. https://ai.meta.com/blog/llama-4-multimodal-intelligence, 2025. Accessed: 2025-04-05. 5", + "[5] Anthropic. Claude 3.7 sonnet and claude code. https://www.anthropic.com/news/claude-3-7-sonnet, 2025. Accessed: 2025-02-25. 1, 3, 5", + "[6] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint, 2023. 4", + "[7] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zhenen Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Yang, Haiyang Xu, and Junyang Lin. Qwen2.5-vl technical report. arXiv preprint, 2025. 5", + "[8] Sébastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, Harsha Nori, Hamid Palangi, Marco Tulio Ribeiro, and Yi Zhang. Sparks of artificial general intelligence: Early experiments with gpt-4. arXiv preprint, 2023. 2", + "[9] Liang Chen, Yang Deng, Yatao Bian, Zeyu Qin, Bingzhe Wu, Tat-Seng Chua, and Kam-Fai Wong. Beyond factuality: A comprehensive evaluation of large language models as knowledge generators. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 6325-6341, Singapore, 2023. Association for Computational Linguistics. 8" + ], + "bbox": [ + 99, + 320, + 480, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, and Feng Zhao. Are we on the right way for evaluating large vision-language models? In Advances in Neural Information Processing Systems, pages 27056-27087. Curran Associates, Inc., 2024. 3", + "[11] Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott Gray, Nick Ryder, Mikhail Pavlov, Alethea Power, Lukasz Kaiser, Mohammad Bavarian, Clemens Winter, Philippe Tillet, Felipe Petroski Such, Dave Cummings, Matthias Plappert, Fotios Chantzis, Elizabeth Barnes, Ariel Herbert-Voss, William Hebgen Guss, Alex Nichol, Alex Paino, Nikolas Tezak, Jie Tang, Igor Babuschkin, Suchir Balaji, Shantanu Jain, William Saunders, Christopher Hesse, Andrew N. Carr, Jan Leike, Josh Achiam, Vedant Misra, Evan Morikawa, Alec Radford, Matthew Knight, Miles Brundage, Mira Murati, Katie Mayer, Peter Welinder, Bob McGrew, Dario Amodei, Sam McCandlish, Ilya Sutskever, and Wojciech Zaremba. Evaluating large language models trained on code. arXiv preprint, 2021. 1", + "[12] Pengcheng Chen, Jin Ye, Guoan Wang, Yanjun Li, Zhongying Deng, Wei Li, Tianbin Li, Haodong Duan, Ziyan Huang, Yanzhou Su, Benyou Wang, Shaoting Zhang, Bin Fu, Jianfei Cai, Bohan Zhuang, Eric J Seibel, Yu Qiao, and Junjun He. Gmai-mmbench: A comprehensive multimodal evaluation benchmark towards general medical ai. In Advances in Neural Information Processing Systems, pages 94327-94427. Curran Associates, Inc., 2024. 3", + "[13] Wenhu Chen, Xueguang Ma, Xinyi Wang, and William W. Cohen. Program of thoughts prompting: Disentangling computation from reasoning for numerical reasoning tasks. Transactions on Machine Learning Research, 2023. 3, 5", + "[14] Zhiyu Chen, Wenhu Chen, Charese Smiley, Sameena Shah, Iana Borova, Dylan Langdon, Reema Moussa, Matt Beane, Ting-Hao Huang, Bryan Routledge, and William Yang Wang. FinQA: A dataset of numerical reasoning over financial data. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3697-3711, Online and Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 2, 3, 5", + "[15] Zhiyu Chen, Shiyang Li, Charese Smiley, Zhiqiang Ma, Sameena Shah, and William Yang Wang. ConvFinQA: Exploring the chain of numerical reasoning in conversational finance question answering. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 6279-6292, Abu Dhabi, United Arab Emirates, 2022. Association for Computational Linguistics. 2, 3", + "[16] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, Lixin Gu, Xuehui Wang, Qingyun Li, Yimin Ren, Zixuan Chen, Jiapeng Luo, Jiahao Wang, Tan Jiang, Bo Wang, Conghui He, Botian Shi, Xingcheng Zhang, Han Lv, Yi Wang, Wenqi Shao, Pei Chu, Zhongying Tu, Tong He, Zhiyong Wu, Huipeng Deng, Jiaye Ge, Kai Chen, Kaipeng" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3253", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Zhang, Limin Wang, Min Dou, Lewei Lu, Xizhou Zhu, Tong Lu, Dahua Lin, Yu Qiao, Jifeng Dai, and Wenhai Wang. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint, 2025. 5", + "[17] Google DeepMind. Introducing gemma 3: The most capable model you can run on a singlegpu or tpu. https://blog.google/technology/developers/gemma-3, 2025. Accessed: 2025-03-12. 5", + "[18] Google DeepMind. Gemini 2.0: Flash, flash-lite and pro. https://developers.googleblog.com/en/gemini-2-family-expands, 2025. Accessed: 2025-02-05. 3, 5", + "[19] Google DeepMind. Gemini 2.0 flash thinking. https://ai.google.dev/gemini-api/docs/thinking, 2025. Accessed: 2025-02-05. 1, 3, 5", + "[20] Google DeepMind. Gemini 2.0. https://cloud.google.com/vertex-ai/generative-ai/docs/gemini-v2#2.0-pro, 2025. Accessed: 2025-02-05. 5", + "[21] DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, Aixin Liu, Bing Xue, Bingxuan Wang, Bochao Wu, Bei Feng, Chengda Lu, Chenggang Zhang, Chengqi Deng, Chenyu Zhang, Chong Ruan, Damai Dai, Deli Chen, Dongjie Ji, Erhang Li, Fangyun Lin, Fucong Dai, Fuli Luo, Guangbo Hao, Guanting Chen, Guowei Li, H. Zhang, Han Bao, Hanwei Xu, Haocheng Wang, Honghui Ding, Huajian Xin, Huazuo Gao, Hui Qu, Hui Li, Jianzhong Guo, Jiashi Li, Jiawei Wang, Jingchang Chen, Jingyang Yuan, Junjie Qiu, Junlong Li, J. L. Cai, Jiaqi Ni, Jian Liang, Jin Chen, Kai Dong, Kai Hu, Kaige Gao, Kang Guan, Kexin Huang, Kuai Yu, Lean Wang, Lecong Zhang, Liang Zhao, Litong Wang, Liyue Zhang, Lei Xu, Leyi Xia, Mingchuan Zhang, Minghua Zhang, Minghui Tang, Meng Li, Miaojun Wang, Mingming Li, Ning Tian, Panpan Huang, Peng Zhang, Qiancheng Wang, Qinyu Chen, Qiushi Du, Ruiqi Ge, Ruisong Zhang, Ruizhe Pan, Runji Wang, R. J. Chen, R. L. Jin, Ruyi Chen, Shanghao Lu, Shangyan Zhou, Shanhuang Chen, Shengfeng Ye, Shiyu Wang, Shuiping Yu, Shunfeng Zhou, Shuting Pan S. S. Li Shuang Zhou Shaoqing Wu Shengfeng Ye Tao Yun Tian Pei Tianyu Sun T. Wang Wangding Zeng Wanjia Zhao Wen Liu Wenfeng Liang Wenjun Gao Wenqin Yu Wentao Zhang W. L. Xiao Wei An Xiaodong Liu Xiaohan Wang Xiaokang Chen Xiaotao Nie Xin Cheng Xin Liu Xin Xie Xingchao Liu Xinyu Yang Xinyuan Li Xuecheng Su Xuheng Lin X.Q.Li Xiangyue Jin Xiaojin Shen Xiaosha Chen Xiaowen Sun Xiaoxiang Wang Xinnan Song Xinyi Zhou Xianzu Wang Xinxia Shan Y.K. Li Y.Q. WangY.X.Wei,Yang Zhang Yanhong Xu Yao Li Yao Zhao Yaofeng Sun Yaohui Wang Yi Yu Yichao Zhang Yifan Shi Yiliang Xiong Ying He Yishi Piao Yisong Wang Yixuan Tan Yiyang Ma Yiyuan Liu Yongqiang Guo Yuan OuYuduan Wang Yue Gong Yuheng Zou Yujia HeYunfan Xiong Yuxiang Luo Yuxiang You Yuxuan Liu Yuyang ZhouY.X.ZhuYanhong Xu Yanping Huang Yaohui Li Yi ZhengYuchen ZhuYunxian Ma Ying Tang Yukun" + ], + "bbox": [ + 91, + 90, + 482, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Zha, Yuting Yan, Z. Z. Ren, Zehui Ren, Zhangli Sha, Zhe Fu, Zhean Xu, Zhenda Xie, Zhengyan Zhang, Zhenen Hao, Zhicheng Ma, Zhigang Yan, Zhiyu Wu, Zihui Gu, Zijia Zhu, Zijun Liu, Zilin Li, Ziwei Xie, Ziyang Song, Zizheng Pan, Zhen Huang, Zhipeng Xu, Zhongyu Zhang, and Zhen Zhang. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint, 2025. 1, 3", + "[22] Chunyuan Deng, Yilun Zhao, Xiangru Tang, Mark Gerstein, and Arman Cohan. Investigating data contamination in modern benchmarks for large language models. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 8706-8719, Mexico City, Mexico, 2024. Association for Computational Linguistics. 5", + "[23] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, Yunsheng Wu, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint, 2024. 3", + "[24] Ziliang Gan, Yu Lu, Dong Zhang, Haohan Li, Che Liu, Jian Liu, Ji Liu, Haipang Wu, Chaoyou Fu, Zenglin Xu, Rongjunchen Zhang, and Yong Dai. Mme-finance: A multimodal finance benchmark for expert-level understanding and reasoning. arXiv preprint, 2024. 4", + "[25] Yingqiang Ge, Wenyue Hua, Kai Mei, jianchao ji, Juntao Tan, Shuyuan Xu, Zelong Li, and Yongfeng Zhang. Opennagi: When llm meets domain experts. In Advances in Neural Information Processing Systems, pages 5539-5568. Curran Associates, Inc., 2023. 2", + "[26] Jian Guan, Wei Wu, Zujie Wen, Peng Xu, Hongning Wang, and Minlie Huang. Amor: A recipe for building adaptable modular knowledge agents through process feedback. In Advances in Neural Information Processing Systems, pages 126118-126148. Curran Associates, Inc., 2024. 8", + "[27] Yunzhuo Hao, Jiawei Gu, Huichen Will Wang, Linjie Li, Zhengyuan Yang, Lijuan Wang, and Yu Cheng. Can MLLMs reason in multimodality? EMMA: An enhanced multimodal reasoning benchmark. In *Forty-second International Conference on Machine Learning*, 2025. 3", + "[28] Gautier Izacard, Mathilde Caron, Lucas Hosseini, Sebastian Riedel, Piotr Bojanowski, Armand Joulin, and Edouard Grave. Unsupervised dense information retrieval with contrastive learning. Transactions on Machine Learning Research, 2022. 8", + "[29] Naman Jain, King Han, Alex Gu, Wen-Ding Li, Fanjia Yan, Tianjun Zhang, Sida Wang, Armando Solar-Lezama, Koushik Sen, and Ion Stoica. Livecodebench: Holistic and contamination free evaluation of large language models for code. In The Thirteenth International Conference on Learning Representations, 2025. 1", + "[30] Dongzhi Jiang, Renrui Zhang, Ziyu Guo, Yanwei Li, Yu Qi, Xinyan Chen, Liuhui Wang, Jianhan Jin, Claire Guo, Shen Yan, Bo Zhang, Chaoyou Fu, Peng Gao, and Hongsheng Li. MME-cot: Benchmarking chain-of-thought in large multimodal models for reasoning quality, robustness, and efficiency. In Forty-second International Conference on Machine Learning, 2025. 3" + ], + "bbox": [ + 516, + 90, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3254", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint, 2020. 1, 3", + "[32] Michael Krumdick, Rik Koncel-Kedziorski, Viet Dac Lai, Varshini Reddy, Charles Lovering, and Chris Tanner. BizBench: A quantitative reasoning benchmark for business and finance. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8309-8332, Bangkok, Thailand, 2024. Association for Computational Linguistics. 2, 3, 4, 5", + "[33] Bohao Li, Yuying Ge, Yixiao Ge, Guangzhi Wang, Rui Wang, Ruimao Zhang, and Ying Shan. Seed-bench: Benchmarking multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13299-13308, 2024. 3", + "[34] Xiaoxi Li, Guanting Dong, Jiajie Jin, Yuyao Zhang, Yujia Zhou, Yutao Zhu, Peitian Zhang, and Zhicheng Dou. Searcho1: Agentic search-enhanced large reasoning models. arXiv preprint, 2025. 8", + "[35] Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let's verify step by step. In The Twelfth International Conference on Learning Representations, 2024. 1", + "[36] Fuxiao Liu, Xiaoyang Wang, Wenlin Yao, Jianshu Chen, Kaiqiang Song, Sangwoo Cho, Yaser Yacoob, and Dong Yu. MMC: Advancing multimodal chart understanding with large-scale instruction tuning. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 1287-1310, Mexico City, Mexico, 2024. Association for Computational Linguistics. 3", + "[37] Haowei Liu, Xi Zhang, Haiyang Xu, Yaya Shi, Chaoya Jiang, Ming Yan, Ji Zhang, Fei Huang, Chunfeng Yuan, Bing Li, and Weiming Hu. MIBench: Evaluating multimodal large language models over multiple images. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 22417-22428, Miami, Florida, USA, 2024. Association for Computational Linguistics. 7", + "[38] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin. Mmbench: Is your multi-modal model an all-around player? In Computer Vision - ECCV 2024, pages 216-233, Cham, 2025. Springer Nature Switzerland. 3", + "[39] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating math reasoning in visual contexts with gpt-4v, bard, and other large multimodal models. In The Twelfth International Conference on Learning Representations, 2024. 1, 3, 5", + "[40] Junyu Luo, Zhizhuo Kou, Liming Yang, Xiao Luo, Jinsheng Huang, Zhiping Xiao, Jingshu Peng, Chengzhong Liu, Ji-aming Ji, Xuanzhe Liu, Sirui Han, Ming Zhang, and Yike Guo. FinMME: Benchmark dataset for financial multi-modal reasoning evaluation. In Proceedings of the 63rd Annual" + ], + "bbox": [ + 89, + 92, + 482, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 29465-29489, Vienna, Austria, 2025. Association for Computational Linguistics. 4", + "[41] Yujun Mao, Yoon Kim, and Yilun Zhou. CHAMP: A competition-level dataset for fine-grained analyses of LLMs' mathematical reasoning capabilities. In Findings of the Association for Computational Linguistics: ACL 2024, pages 13256-13274, Bangkok, Thailand, 2024. Association for Computational Linguistics. 1", + "[42] Grégoire Mialon, Clémentine Fourrier, Thomas Wolf, Yann LeCun, and Thomas Scialom. GAIA: a benchmark for general AI assistants. In The Twelfth International Conference on Learning Representations, 2024. 2", + "[43] Meredith Ringel Morris, Jascha Sohl-Dickstein, Noah Fiedel, Tris Warkentin, Allan Dafoe, Aleksandra Faust, Clement Farabet, and Shane Legg. Position: Levels of AGI for operationalizing progress on the path to AGI. In *Forty-first International Conference on Machine Learning*, 2024. 2", + "[44] OpenAI. Hello gpt-4o. https://openai.com/index/hello-gpt-4o, 2024. Accessed: 2024-05-13. 5", + "[45] OpenAI. Learning to reason with llms. https://openai.com/index/learning-to-reason-with-llms, 2024. Accessed: 2024-09-12. 1, 3", + "[46] OpenAI. Openai o1 system card. https://openai.com/index/openai-o1-system-card, 2024. Accessed: 2024-12-05. 1, 3, 5", + "[47] OpenAI. Openai o1-mini. https://openai.com/index/openai-ol-mini-advancing-cost-efficient-reasoning, 2024. Accessed: 2024-09-12. 1", + "[48] OpenAI. Openai o3-mini system card. https://openai.com/index/o3-mini-system-card, 2025. Accessed: 2025-01-31. 1, 3", + "[49] Baolin Peng, Michel Galley, Pengcheng He, Hao Cheng, Yujia Xie, Yu Hu, Qiuyuan Huang, Lars Liden, Zhou Yu, Weizhu Chen, and Jianfeng Gao. Check your facts and try again: Improving large language models with external knowledge and automated feedback. arXiv preprint, 2023. 8", + "[50] Oscar Sainz, Jon Campos, Iker García-Ferrero, Julien Etxaniz, Oier Lopez de Lacalle, and Eneko Agirre. NLP evaluation in trouble: On the need to measure LLM data contamination for each benchmark. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 10776-10787, Singapore, 2023. Association for Computational Linguistics. 5", + "[51] Aditya Sharma, Michael Saxon, and William Yang Wang. Losing visual needles in image haystacks: Vision language models are easily distracted in short and long contexts. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 5429-5451, Miami, Florida, USA, 2024. Association for Computational Linguistics. 7", + "[52] Weijia Shi, Anirudh Ajith, Mengzhou Xia, Yangsibo Huang, Daogao Liu, Terra Blevins, Danqi Chen, and Luke Zettlemoyer. Detecting pretraining data from large language models. In The Twelfth International Conference on Learning Representations, 2024. 5" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "3255", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Zichen Tang, Haihong E, Ziyan Ma, Haoyang He, Jiacheng Liu, Zhongjun Yang, Zihua Rong, Rongjin Li, Kun Ji, Qing Huang, Xinyang Hu, Yang Liu, and Qianhe Zheng. FinanceReasoning: Benchmarking financial numerical reasoning more credible, comprehensive and challenging. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15721-15749, Vienna, Austria, 2025. Association for Computational Linguistics. 4", + "[54] Kimi Team, Angang Du, Bofei Gao, Bowei Xing, Changjiu Jiang, Cheng Chen, Cheng Li, Chenjun Xiao, Chenzhuang Du, Chonghua Liao, Chuning Tang, Congcong Wang, Dehao Zhang, Enming Yuan, Enzhe Lu, Fengxiang Tang, Flood Sung, Guangda Wei, Guokun Lai, Haiqing Guo, Han Zhu, Hao Ding, Hao Hu, Hao Yang, Hao Zhang, Haotian Yao, Haotian Zhao, Haoyu Lu, Haoze Li, Haozhen Yu, Hongcheng Gao, Huabin Zheng, Huan Yuan, Jia Chen, Jianhang Guo, Jianlin Su, Jianzhou Wang, Jie Zhao, Jin Zhang, Jingyuan Liu, Junjie Yan, Junyan Wu, Lidong Shi, Ling Ye, Longhui Yu, Mengnan Dong, Neo Zhang, Ningchen Ma, Qiwei Pan, Qucheng Gong, Shaowei Liu, Shengling Ma, Shupeng Wei, Sihan Cao, Siying Huang, Tao Jiang, Weihao Gao, Weimin Xiong, Weiran He, Weixiao Huang, Weixin Xu, Wenhao Wu, Wenyang He, Xianghui Wei, Xianqing Jia, Xingzhe Wu, Xinran Xu, Xinxing Zu, Xinyu Zhou, Xuehai Pan, Y. Charles, Yang Li, Yangyang Hu, Yangyang Liu, Yanru Chen, Yejie Wang, Yibo Liu, Yidao Qin, Yifeng Liu, Ying Yang, Yiping Bao, Yulun Du, Yuxin Wu, Yuzhi Wang, Zaida Zhou, Zhaoji Wang, Zhaowei Li, Zhen Zhu, Zheng Zhang, Zhexu Wang, Zhilin Yang, Zhiqi Huang, Zihao Huang, Ziyao Xu, Zonghan Yang, and Zongyu Lin. Kimi k1.5: Scaling reinforcement learning with llms. arXiv preprint, 2025. 1", + "[55] Qwen Team. Qvq: To see the world with wisdom. https://qwenlm.github.io/blog/qvq-72b-preview, 2024. Accessed: 2024-12-25. 3, 5", + "[56] Qwen Team. Qwq: Reflect deeply on the boundaries of the unknown. https://qwenlm.github.io/blog/qwq-32b-preview, 2024. Accessed: 2024-11-28. 1, 3", + "[57] Qwen Team. qwen-omni. https://help.aliyun.com/zh/model-studio/user-guide/qwen-omni, 2025. Accessed: 2025-03-26. 5", + "[58] Prakhar Verma, Sukruta Prakash Midigeshi, Gaurav Sinha, Arno Solin, Nagarajan Natarajan, and Amit Sharma. Plan*RAG: Efficient test-time planning for retrieval augmented generation. In Workshop on Reasoning and Planning for Large Language Models, 2025. 8", + "[59] Xiaoxuan Wang, Ziniu Hu, Pan Lu, Yanqiao Zhu, Jieyu Zhang, Satyen Subramaniam, Arjun R Loomba, Shichang Zhang, Yizhou Sun, and Wei Wang. Scibench: Evaluating college-level scientific problem-solving abilities of large language models. In *Forty-first International Conference on Machine Learning*, 2024. 1", + "[60] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian richter, Fei Xia, Ed Chi, Quoc V Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models. In Advances in Neural Information Pro" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "cessing Systems, pages 24824-24837. Curran Associates, Inc., 2022. 3, 5", + "[61] xAI. Models. https://docs.x.ai/docs/models? cluster=us-east-1#model-constraints, 2024. Accessed: 2024-12-12. 5", + "[62] Fengli Xu, Qianyue Hao, Zefang Zong, Jingwei Wang, Yunke Zhang, Jingyi Wang, Xiaochong Lan, Jiahui Gong, Tianjian Ouyang, Fanjin Meng, Chenyang Shao, Yuwei Yan, Qinglong Yang, Yiwen Song, Sijian Ren, Xinyuan Hu, Yu Li, Jie Feng, Chen Gao, and Yong Li. Towards large reasoning models: A survey of reinforced reasoning with large language models. arXiv preprint, 2025. 1, 3", + "[63] Siqiao Xue, Tingting Chen, Fan Zhou, Qingyang Dai, Zhixuan Chu, and Hongyuan Mei. Famma: A benchmark for financial domain multilingual multimodal question answering. arXiv preprint, 2024. 2, 4, 5", + "[64] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. MM-vet: Evaluating large multimodal models for integrated capabilities. In *Forty-first International Conference on Machine Learning*, 2024. 3", + "[65] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Ren-liang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9556-9567, 2024. 1, 2, 3, 4, 5", + "[66] Xiang Yue, Tianyu Zheng, Yuansheng Ni, Yubo Wang, Kai Zhang, Shengbang Tong, Yuxuan Sun, Botao Yu, Ge Zhang, Huan Sun, Yu Su, Wenhu Chen, and Graham Neubig. MMMU-pro: A more robust multi-discipline multimodal understanding benchmark. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15134-15186, Vienna, Austria, 2025. Association for Computational Linguistics. 3, 4, 5", + "[67] Yilun Zhao, Hongjun Liu, Yitao Long, Rui Zhang, Chen Zhao, and Arman Cohan. Financemath: Knowledge-intensive math reasoning in finance domains. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12841-12858, Bangkok, Thailand, 2024. Association for Computational Linguistics. 2, 3, 4, 5", + "[68] Yilun Zhao, Yitao Long, Hongjun Liu, Ryo Kamoi, Linyong Nan, Lyuhao Chen, Yixin Liu, Xiangru Tang, Rui Zhang, and Arman Cohan. DocMath-eval: Evaluating math reasoning capabilities of LLMs in understanding long and specialized documents. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 16103-16120, Bangkok, Thailand, 2024. Association for Computational Linguistics. 4, 5", + "[69] Fengbin Zhu, Wenqiang Lei, Youcheng Huang, Chao Wang, Shuo Zhang, Jiancheng Lv, Fuli Feng, and Tat-Seng Chua. TAT-QA: A question answering benchmark on a hybrid of" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "3256", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "tabular and textual content in finance. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 3277-3287, Online, 2021. Association for Computational Linguistics. 2, 3, 5", + "bbox": [ + 122, + 90, + 486, + 176 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "3257", + "bbox": [ + 482, + 944, + 516, + 957 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2025/FinMMR_ Make Financial Numerical Reasoning More Multimodal, Comprehensive, and Challenging/9bdd874d-4c07-4812-a904-cd4ae1c8c840_model.json b/2025/FinMMR_ Make Financial Numerical Reasoning More Multimodal, Comprehensive, and Challenging/9bdd874d-4c07-4812-a904-cd4ae1c8c840_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2ee1c26411f23fb71c4772241881d6c526426110 --- /dev/null +++ b/2025/FinMMR_ Make Financial Numerical Reasoning More Multimodal, Comprehensive, and Challenging/9bdd874d-4c07-4812-a904-cd4ae1c8c840_model.json @@ -0,0 +1,2943 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.13, + 0.77, + 0.186 + ], + "angle": 0, + "content": "\\(\\mathcal{F}_{\\mathsf{M}}\\) FinMMR: Make Financial Numerical Reasoning More Multimodal, Comprehensive, and Challenging" + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.212, + 0.903, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.358, + 0.299, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.358, + 0.543, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.544, + 0.359, + 0.903, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.541, + 0.907, + 0.571 + ], + "angle": 0, + "content": "Figure 1. Overview of FinMMR. FinMMR presents three challenges: (1) Visual Perception: 8.7K financial images of 14 categories; (2) Knowledge Reasoning: 4.3K financial questions of 14 subdomains; (3) Numerical Computation: multi-step precise calculations." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.582, + 0.327, + 0.598 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.615, + 0.486, + 0.873 + ], + "angle": 0, + "content": "We present FinMMR, a novel bilingual multimodal benchmark tailored to evaluate the reasoning capabilities of multimodal large language models (MLLMs) in financial numerical reasoning tasks. Compared to existing benchmarks, our work introduces three significant advancements. (1) Multimodality: We meticulously transform existing financial reasoning benchmarks, and construct novel questions from the latest Chinese financial research reports. FinMMR comprises 4.3K questions and 8.7K images spanning 14 categories, including tables, bar charts, and ownership structure charts. (2) Comprehensiveness: FinMMR encompasses 14 financial subdomains, including corporate finance, banking, and industry analysis, significantly exceeding existing benchmarks in financial domain knowledge breadth. (3) Challenge: Models are required to perform multi-step precise numerical reasoning by integrating financial knowledge with the understanding of complex financial" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.584, + 0.907, + 0.645 + ], + "angle": 0, + "content": "images and text. The best-performing MLLM achieves only \\(53.0\\%\\) accuracy on Hard problems. We believe that FinMMR will drive advancements in enhancing the reasoning capabilities of MLLMs in real-world scenarios." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.677, + 0.645, + 0.692 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.703, + 0.907, + 0.809 + ], + "angle": 0, + "content": "Recently, large reasoning models (LRMs) [21, 47, 48, 54, 56, 62], show powerful reasoning capabilities over multi-step reasoning tasks, with train-time scaling and test-time scaling [31, 45]. These reasoning models are proficient in code [11, 29], math [35, 41], and science [59]. Multimodal large language models (MLLMs) [5, 19, 46] also exhibit greater performance on multimodal reasoning [39, 65]." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Despite significant advancements, there remains a notable gap in understanding the practical applicability of MLLMs in numerical reasoning within real-world scenarios, particularly in high-stakes fields such as finance and healthcare. As shown in Fig. 1, financial analysts in their daily work are required to read visually enriched finan" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3245" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.124, + 0.097, + 0.321, + 0.126 + ], + "angle": 0, + "content": "Rich Images Challenge Visual Perception" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.137, + 0.34, + 0.216 + ], + "angle": 0, + "content": "问题：计算2016年至2024年Q2期间，上海银行和上市城商行的个人住房贷款年均复合增长率之差，结果以百分比表示，保留两位小数。(Calculate the difference in the compound annual growth rate (CAGR) of personal housing loans between Bank of Shanghai and listed city commercial banks from 2016 to Q2 2024. Present the result as a percentage rounded to two decimal places.)" + }, + { + "type": "image", + "bbox": [ + 0.142, + 0.221, + 0.312, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.278, + 0.311, + 0.336, + 0.319 + ], + "angle": 0, + "content": "Answer: 2.17" + }, + { + "type": "title", + "bbox": [ + 0.132, + 0.326, + 0.312, + 0.334 + ], + "angle": 0, + "content": "Compound Average Growth Rate (CAGR)" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.337, + 0.345, + 0.387 + ], + "angle": 0, + "content": "问题：请计算惠济新生对日照公司的间接持股比例，结果以百分比表示，保留两位小数。(Please calculate the indirect shareholding ratio of Huiji Xinsheng in Rizhao Company. Present the result as a percentage, rounded to two decimal places.)" + }, + { + "type": "image", + "bbox": [ + 0.151, + 0.392, + 0.305, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.273, + 0.492, + 0.338, + 0.501 + ], + "angle": 0, + "content": "Answer: 13.65" + }, + { + "type": "title", + "bbox": [ + 0.158, + 0.507, + 0.281, + 0.515 + ], + "angle": 0, + "content": "Indirect Shareholding Ratio" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.519, + 0.345, + 0.548 + ], + "angle": 0, + "content": "问题：请计算2022年第四季度和2023年第一季度的总资本开支，并将其与2021年第四季度的资本开支进行比较，计算下降百分比，结果保留两位小数。" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.549, + 0.345, + 0.588 + ], + "angle": 0, + "content": "(Calculate the total capital expenditure for Q4 2022 and Q1 2023 combined, and compare it to the capital expenditure for Q4 2021. Calculate the percentage decrease, rounded to two decimal places.)" + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.591, + 0.337, + 0.649 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.266, + 0.661, + 0.341, + 0.669 + ], + "angle": 0, + "content": "Answer: -110.36" + }, + { + "type": "image_caption", + "bbox": [ + 0.177, + 0.678, + 0.267, + 0.687 + ], + "angle": 0, + "content": "Capital Expenditure" + }, + { + "type": "title", + "bbox": [ + 0.382, + 0.097, + 0.609, + 0.126 + ], + "angle": 0, + "content": "Comprehensive Subdomains Challenge Knowledge Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.381, + 0.137, + 0.611, + 0.249 + ], + "angle": 0, + "content": "Q: Table 19.3 shows a book balance sheet for the Wishing Well Motel chain. The company's long-term debt is secured by its real estate assets, but it also uses short-term bank loans as a permanent source of financing. It pays \\(10\\%\\) interest on the bank debt and \\(9\\%\\) interest on the secured debt. Wishing Well has 10 million shares of stock outstanding, trading at \\(\\$ 90\\)per share. The expected return on Wishing Well's common stock is \\(18\\%\\). Calculate Wishing Well's WACC. Assume that the book and market values of Wishing Well's debt are the same. The marginal tax rate is \\(21\\%\\). Answer as a percentage to single decimal place." + }, + { + "type": "title", + "bbox": [ + 0.382, + 0.251, + 0.422, + 0.26 + ], + "angle": 0, + "content": ":" + }, + { + "type": "table", + "bbox": [ + 0.386, + 0.266, + 0.607, + 0.318 + ], + "angle": 0, + "content": "
Cash and marketable securities100Bank loan280
Accounts receivable200Accounts payable120
Inventory50Current liabilities400
Current assets350
Real estate2,100Long-term debt1,800
Other assets150Equity400
Total2,600Total2,600
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.384, + 0.321, + 0.597, + 0.329 + ], + "angle": 0, + "content": "TABLE 19.3 Book balance sheet for Wishing Well Inc. (figures in $ millions)" + }, + { + "type": "title", + "bbox": [ + 0.398, + 0.336, + 0.506, + 0.344 + ], + "angle": 0, + "content": "Weighted Average Cost" + }, + { + "type": "text", + "bbox": [ + 0.381, + 0.351, + 0.607, + 0.397 + ], + "angle": 0, + "content": "Q: For each of the investments shown in the following table, calculate the rate of return earned over the unspecified time period. What is the rate of return for Investment A? Answer as a percentage to the nearest integer." + }, + { + "type": "title", + "bbox": [ + 0.382, + 0.401, + 0.422, + 0.41 + ], + "angle": 0, + "content": ":" + }, + { + "type": "table", + "bbox": [ + 0.395, + 0.422, + 0.6, + 0.48 + ], + "angle": 0, + "content": "
InvestmentCash flow during periodBeginning-of-period valueEnd-of-period value
A$-2,800$23,400$20,100
B16,000225,000324,000
C7006,5008,000
D3,58036,60046,500
E-50062,70052,800
" + }, + { + "type": "title", + "bbox": [ + 0.441, + 0.498, + 0.534, + 0.506 + ], + "angle": 0, + "content": "Rate of Return (RoR)" + }, + { + "type": "text", + "bbox": [ + 0.379, + 0.509, + 0.611, + 0.582 + ], + "angle": 0, + "content": "Q: Ricky is considering purchasing an apartment costing \\( \\\\( {700},{000} \\) . He will pay a 30% down payment and take out a mortgage for the remainder. Since he just got married and wants to save some money for future use, he will choose the plan with the lowest monthly payment. After visiting several banks, he received the following mortgage offers: What is the monthly payment for Bank A? Answer to two decimal places." + }, + { + "type": "title", + "bbox": [ + 0.38, + 0.583, + 0.418, + 0.592 + ], + "angle": 0, + "content": ":" + }, + { + "type": "table", + "bbox": [ + 0.386, + 0.595, + 0.608, + 0.675 + ], + "angle": 0, + "content": "
BankInterest rateTerm (years)
A3.5%15
B320
C425
D4.518
" + }, + { + "type": "table_caption", + "bbox": [ + 0.452, + 0.679, + 0.533, + 0.688 + ], + "angle": 0, + "content": "Monthly Payment" + }, + { + "type": "title", + "bbox": [ + 0.647, + 0.097, + 0.885, + 0.125 + ], + "angle": 0, + "content": "Complex Formulas Challenge Numerical Computation" + }, + { + "type": "text", + "bbox": [ + 0.653, + 0.137, + 0.878, + 0.177 + ], + "angle": 0, + "content": "Q: What is the total weighted average Cash and Payment-In-Kind (PIK) interest rate payable under the subordinated and senior notes portfolio in the year 2023?" + }, + { + "type": "title", + "bbox": [ + 0.655, + 0.18, + 0.696, + 0.189 + ], + "angle": 0, + "content": ":" + }, + { + "type": "table", + "bbox": [ + 0.663, + 0.192, + 0.871, + 0.236 + ], + "angle": 0, + "content": "
EUR
In Mio €201820172016%2015
Deutschland3.498.5003.498.5003.498.500100%100%
USA1.200.0001.200.0001.200.000--
Nordafrika, Naher Osten1.200.0001.200.0001.200.000--
Asien-Pazifik1.200.0001.200.0001.200.000--
davon1.200.0001.200.0001.200.000--
davon1.200.0001.200.0001.200.000--
davon1.200.0001.200.001.200.000--
davon1.200.0001.200.0001.200.000--
davon1.200.0001.200.0001.200.000--
deutschland3.498.5003.498.5003.498.500100%100%
davon1.200.0001.200.0001.200.000--
" + }, + { + "type": "title", + "bbox": [ + 0.655, + 0.248, + 0.703, + 0.257 + ], + "angle": 0, + "content": ":" + }, + { + "type": "table", + "bbox": [ + 0.662, + 0.259, + 0.872, + 0.285 + ], + "angle": 0, + "content": "
[average = (investmenti × cashi) / ∑investmenti] + (investmenti × PIKi) / ∑investmenti]
" + }, + { + "type": "table_caption", + "bbox": [ + 0.813, + 0.291, + 0.872, + 0.3 + ], + "angle": 0, + "content": "Answer: 12.0" + }, + { + "type": "title", + "bbox": [ + 0.689, + 0.306, + 0.848, + 0.314 + ], + "angle": 0, + "content": "Total Weighted Average Cash and" + }, + { + "type": "title", + "bbox": [ + 0.682, + 0.315, + 0.85, + 0.324 + ], + "angle": 0, + "content": "Payment-In-Kind (PIK) Interest Rate" + }, + { + "type": "text", + "bbox": [ + 0.653, + 0.328, + 0.881, + 0.367 + ], + "angle": 0, + "content": "Q: John oversees a fund, with the returns for the first three years displayed below: What will be the holding period return (expressed as a percentage)? Answer to three decimal places." + }, + { + "type": "title", + "bbox": [ + 0.655, + 0.371, + 0.696, + 0.38 + ], + "angle": 0, + "content": ":" + }, + { + "type": "table", + "bbox": [ + 0.667, + 0.385, + 0.868, + 0.431 + ], + "angle": 0, + "content": "
YearInvestmentReturn
1$ 50012%
2$ 6005%
3$ 10001%
" + }, + { + "type": "title", + "bbox": [ + 0.655, + 0.439, + 0.703, + 0.448 + ], + "angle": 0, + "content": ":" + }, + { + "type": "equation", + "bbox": [ + 0.677, + 0.452, + 0.853, + 0.477 + ], + "angle": 0, + "content": "\\[\n\\left[ H P R = \\left(\\frac {\\sum \\left(\\text {i n v e s t m e n t} _ {t} \\times (1 + \\text {r e t u r n} _ {t})\\right)}{\\sum \\text {i n v e s t m e n t}} - 1\\right) \\times 1 0 0 \\right]\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.81, + 0.483, + 0.875, + 0.491 + ], + "angle": 0, + "content": "Answer: 4.762" + }, + { + "type": "title", + "bbox": [ + 0.704, + 0.498, + 0.831, + 0.507 + ], + "angle": 0, + "content": "Holding Period Return (HPR)" + }, + { + "type": "text", + "bbox": [ + 0.656, + 0.51, + 0.871, + 0.551 + ], + "angle": 0, + "content": "Q: What is the average quarterly share price in 2021 assuming each quarter had a completely uniform price distribution in high and low bid prices, measured in dollars?" + }, + { + "type": "title", + "bbox": [ + 0.656, + 0.556, + 0.697, + 0.565 + ], + "angle": 0, + "content": ":" + }, + { + "type": "table", + "bbox": [ + 0.669, + 0.566, + 0.869, + 0.617 + ], + "angle": 0, + "content": "
Quarter EndedHigh BidLow Bid
October 31, 2021 (1)$4.51$3.99
July 31, 2021$8.47$2.30
April 30, 2021$13.00$3.80
January 31, 2021$14.50$0.015
October 31, 2020 (2)$0.199$0.01
" + }, + { + "type": "title", + "bbox": [ + 0.657, + 0.624, + 0.704, + 0.632 + ], + "angle": 0, + "content": ":" + }, + { + "type": "equation", + "bbox": [ + 0.688, + 0.634, + 0.827, + 0.66 + ], + "angle": 0, + "content": "\\[\n[ \\text {A v e r a g e} = \\frac {1}{4} \\sum_ {i = 1} ^ {4} \\frac {\\text {H i g h} _ {i} + \\text {L o w} _ {i}}{2} ]\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.813, + 0.661, + 0.873, + 0.669 + ], + "angle": 0, + "content": "Answer: 6.32" + }, + { + "type": "title", + "bbox": [ + 0.716, + 0.676, + 0.824, + 0.685 + ], + "angle": 0, + "content": "Average Quarterly Price" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.705, + 0.907, + 0.734 + ], + "angle": 0, + "content": "Figure 2. Sampled FinMMR examples with two language (i.e., English and Chinese), rich images and different knowledge. The questions and images need expert-level visual perception, knowledge reasoning and numerical computation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.76, + 0.485, + 0.896 + ], + "angle": 0, + "content": "cial documents, extract key financial indicators from tables, images, and texts, and perform multi-step precise numerical calculations to support professional decision-making. Similarly, to achieve expert artificial general intelligence (AGI) [8, 25, 42, 43, 65], MLLMs are expected to comprehend complex domain-specific images akin to human experts, and apply domain knowledge to perform accurate numerical reasoning. This raises the question: Can current MLLMs seamlessly integrate vision and text to perform" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.76, + 0.905, + 0.79 + ], + "angle": 0, + "content": "domain-specific complex reasoning, matching the proficiency of LRMs in pure text-based tasks?" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Specifically, we choose the financial domain to evaluate the complex reasoning capabilities of MLLMs, where precision and transparent reasoning are paramount [32]. Existing numerical reasoning benchmarks for finance are limited in their text-based reasoning, coverage of specific financial knowledge, and complexity of reasoning [14, 15, 32, 67, 69]. FAMMA [63] is mainly modelled after textbooks" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3246" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "and CFA exam questions, MathVista [39] does not involve the application of financial knowledge, MMMU [65] and MMMU-Pro [66] are restricted to multiple-choice questions, diverging from authentic financial problem-solving. The lack of high-quality, knowledge-intensive multimodal financial numerical reasoning benchmarks makes it challenging to objectively evaluate the actual reasoning capabilities of MLLMs and analyze their shortcomings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.214, + 0.482, + 0.35 + ], + "angle": 0, + "content": "Therefore, we propose FinMMR, a bilingual multimodal numerical reasoning benchmark designed to evaluate the reasoning capabilities of MLLMs in the finance domain. The benchmark comprises 4.3K problems, covering 14 financial subdomains (e.g., corporate finance and industry analysis), with 8.7K images derived from 14 categories (e.g., tables and ownership structure charts). Each problem includes rich images, an unambiguous question, a Python-formatted solution, and a precise answer." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.353, + 0.483, + 0.549 + ], + "angle": 0, + "content": "For multimodality, without representing financial tables as structured text, FinMMR represent all tables, charts, and diagrams as images. For comprehensiveness, FinMMR covers 14 financial subdomains and two languages (i.e., English and Chinese), demanding domain knowledge such as option pricing and portfolio management. For challenge, FinMMR focus on multi-step numerical reasoning, requiring models to provide exact numerical answers under strict evaluation criteria (emphasizing units, percentages, and decimal places). Furthermore, we mix each Chinese questions with two distractor images that are contextually adjacent to the ground images, approaching real-world multimodal reasoning scenarios." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.483, + 0.612 + ], + "angle": 0, + "content": "We evaluate 15 state-of-the-art (SOTA) MLLMs [18-21, 46-48, 56], utilizing Chain-of-Thought (CoT) [60] and Program-of-Thought (PoT) [13]. The experimental results on FinMMR reveals three key findings:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.614, + 0.483, + 0.78 + ], + "angle": 0, + "content": "- MLLMs Face Significant Challenges in Domain-Specific Multimodal Numerical Reasoning: All evaluated models underperform on FinMMR. On the Hard test set, the best-performing model, Claude 3.7 Sonnet with 64K extended thinking, achieves only \\(53.00\\%\\) accuracy, while OpenAI o1 achieves merely \\(48.40\\%\\). Through error analysis, we identify that visual perception, knowledge reasoning, and numerical computation collectively pose challenges to MLLMs. Current MLLMs still struggle with complex multimodal reasoning tasks in specialized domains, compared to text-based reasoning." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.781, + 0.483, + 0.902 + ], + "angle": 0, + "content": "- Better Synergy Between Visual Perception and Complex Reasoning Is Needed: Distracting images lead to a greater than \\(10\\%\\) drop in accuracy for Qwen2.5-VL-72B compared to ground images alone, indicating that irrelevant visual information severely impacts multimodal reasoning. By decoupling visual filtering and reasoning, Qwen2.5-VL-72B improved from \\(64.73\\%\\) to \\(71.56\\%\\). Combining MLLMs with LRMs, by efficiently" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.614, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.905, + 0.181 + ], + "angle": 0, + "content": "Parsing visual information into structured text and leveraging the LRMs' text-based reasoning capabilities, can also yield better performance. The combination of GPT4o and DeepSeek-R1 achieves \\(86.72\\%\\) accuracy on 1,160 tabular QA instances, outperforming Claude 3.7 Sonnet \\((83.53\\%)\\)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.183, + 0.905, + 0.348 + ], + "angle": 0, + "content": "- Refined Structured Domain Knowledge Enhances MLLMs' Complex Reasoning: Leading MLLMs lack sufficient experience in applying rich domain knowledge when solving complex reasoning tasks. By annotating structured financial functions and leveraging the model's ability to generate retrieval questions and make judgments, knowledge augmentation can significantly improve MLLMs' performance. MLLMs achieve absolute improvements ranging from 2.76 to 4.31 percentage points, allowing weaker models to approach SOTA performance, while SOTA models exhibit further gains." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.349, + 0.905, + 0.485 + ], + "angle": 0, + "content": "These findings highlight the bottlenecks of MLLMs in complex multimodal reasoning tasks in expert domains closer to real-world scenarios. They emphasize the need for continuous improvements in three key areas: more intricate visual perception, more specialized knowledge reasoning, and more accurate numerical computation. Alternatively, leveraging tools or model combinations can help achieve a balance between performance and cost, enabling MLLMs to perform expert-level reasoning tasks like human experts." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.501, + 0.655, + 0.516 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.526, + 0.702, + 0.541 + ], + "angle": 0, + "content": "2.1. LRMs and MLLMs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.549, + 0.907, + 0.775 + ], + "angle": 0, + "content": "By integrating train-time scaling and test-time scaling [31, 45], LRMs have demonstrated remarkable reasoning capabilities [62]. However, most LRMs are limited to handling text-based problems. The growing demand for solving real-world tasks has spurred the development of MLLMs [5, 19, 55] and benchmarks designed to evaluate the perception and reasoning abilities of MLLMs [10, 23, 27, 30, 33, 36, 38, 64-66]. For instance, MME-CoT [30] evaluates models' space-time understanding, while EMMA [27] focuses STEM subjects. Following this trend, domain-specific benchmarks which require deep domain expertise have emerged, such as MathVista [39] for mathematics and GMAI-MMBench [12] for medicine. Yet, financial reasoning remains an unexplored area in the current landscape of MLLM benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.788, + 0.724, + 0.802 + ], + "angle": 0, + "content": "2.2. Financial Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.905, + 0.901 + ], + "angle": 0, + "content": "The financial domain presents a distinct and more formidable set of challenges for model evaluation, which arise from its inherent complexity, information density, and dependence on expertise. The majority of existing text-only financial numerical reasoning benchmarks [14, 15, 32, 67-69] are constrained by limitations such as sub-optimal an" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3247" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.502, + 0.248 + ], + "angle": 0, + "content": "
PropertyValue
# Total Questions4,300
# Easy/Medium/Hard1,300/1,500/1,500
# Validation/Test900/3,400
# Chinese/English2,150/2,150
# Operators (Easy/Medium/Hard)1.75/2.97/5.34
# Lines of Code (Easy/Medium/Hard)4.14/5.06/7.34
# Parentheses (Easy/Medium/Hard)0.88/3.11/4.25
# Difficulty (Easy/Medium/Hard)1.93/2.96/3.79
" + }, + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.258, + 0.48, + 0.273 + ], + "angle": 0, + "content": "Table 1. Key statistics of FinMMR (Avg values of three subsets)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.301, + 0.483, + 0.377 + ], + "angle": 0, + "content": "notation quality, narrow domain knowledge coverage, and overly simplistic reasoning tasks. Although FinanceReasoning [53] offers complex tasks with rich knowledge and high-quality annotations, its text-only modality limits multimodal reasoning evaluation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.378, + 0.484, + 0.514 + ], + "angle": 0, + "content": "Recent multimodal financial benchmarks have sought to bridge this gap but still possess limitations. FAMMA [63] being sourced from textbooks and examinations does not mirror the real-world tasks. FinMME [40] uses a multiple-choice format, which may overestimate model reasoning due to guesswork. MME-Finance [24] is constrained by coarse annotations and an isolated assessment of domain knowledge, limiting holistic evaluation of real-world financial reasoning." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.531, + 0.298, + 0.547 + ], + "angle": 0, + "content": "3. FinMMR Benchmark" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.556, + 0.299, + 0.571 + ], + "angle": 0, + "content": "3.1. Overview of FinMMR" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.579, + 0.483, + 0.716 + ], + "angle": 0, + "content": "We introduce FinMMR, a bilingual (English and Chinese) multimodal benchmark for evaluating the reasoning capabilities of MLLMs in financial numerical reasoning tasks. FinMMR consists of 4,300 questions covering 14 financial subdomains (e.g., corporate finance, industry analysis, financial markets, asset management). The key statistics are summarized in Tab. 1, and the composition of sub-domains and images is illustrated in Fig. 1. As illustrated in Fig. 2, FinMMR introduces three key challenges:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.72, + 0.483, + 0.795 + ], + "angle": 0, + "content": "- Rich Images Challenge Visual Perception: FinMMR comprises 8.7K images from 14 categories (e.g., bar charts, line charts, ownership structure charts). MLLMs must identify relevant images among distractors and extract critical information from the correct images." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.796, + 0.483, + 0.856 + ], + "angle": 0, + "content": "- Comprehensive Subdomains Challenge Knowledge Reasoning: MLLMs need to flexibly apply diverse domain-specific financial knowledge from 14 subdomains to solve multi-step reasoning tasks." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "- Complex Formulas Challenge Numerical Computation: All questions require precise numerical answers, eliminating the potential bias from lucky guesses that" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.72, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.528, + 0.092, + 0.792, + 0.107 + ], + "angle": 0, + "content": "could occur in multiple-choice formats." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.121, + 0.725, + 0.136 + ], + "angle": 0, + "content": "3.2. Data Curation Process" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.144, + 0.907, + 0.28 + ], + "angle": 0, + "content": "We first curated a subset of questions from public text-based financial reasoning benchmarks and systematically transformed them into multimodal problems using a unified standard. Subsequently, we constructed a novel multimodal Chinese Research Report Question Answering (CR-RQA) benchmark from scratch, merging two data sources into FinMMR. Each question is accompanied by an executable Python solution, yields a numerical answer and demonstrates a clear reasoning chain." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.28, + 0.907, + 0.447 + ], + "angle": 0, + "content": "Update to Public Benchmarks We re-announce 124 and 163 financial questions from MMMU [65] and MMMU-Pro [66], respectively. Following rigorous verification, these questions were directly incorporated into our dataset. Furthermore, we extracted 77, 288, and 795 questions from FinanceMath [67], CodeTAT-QA [32], and CodeFinQA [32], respectively. From DocMath-Eval [68], we further obtained 703 questions from its four subsets. For each question, we rendered any tabular data as images while removing the corresponding table information from the text, ensuring that MLLMs cannot rely solely on textual content." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.448, + 0.906, + 0.612 + ], + "angle": 0, + "content": "Building a Novel Dataset from Scratch We collect 90 research reports, all of which are obtained through authorized access and cover diverse topics such as industry research, macroeconomic analysis, and strategy research. We use 360LayoutAnalysis [1] to extract informative images and discard those lacking explicit numerical data, reducing ambiguity. For each retained image, we prompt Qwen-VLMax [6] to formulate questions requiring multiple reasoning steps or complex calculations. Each question is accompanied by an executable Python solution and a definitive numerical answer." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.906, + 0.675 + ], + "angle": 0, + "content": "Furthermore, we introduce distractor images sourced from the same reports adjacent to ground images to challenge MLLMs in extracting relevant numerical information from structured, densely packed visuals." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.906, + 0.795 + ], + "angle": 0, + "content": "Data Quality Assurance This process ensured that every question was clearly written, featured a detailed reasoning solution, and included an accurate final answer. The annotators included 16 graduate students in finance and two experts holding CFA certifications. With the aid of LLMs, this meticulous verification process spanned three months, culminating in a dataset that meets high standards of clarity and correctness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.873 + ], + "angle": 0, + "content": "Dataset Splitting and Release To classify the problems by difficulty, we employ a heuristic algorithm that takes into account the number of operators \\((o)\\), code lines \\((l)\\) and parentheses pairs \\((p)\\) in the Python solution. Specifically, the difficulty of reasoning \\(rc\\) of a problem is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.886, + 0.907, + 0.903 + ], + "angle": 0, + "content": "\\[\nr c = \\ln (\\max (o, 1)) + \\ln (\\max (l + p, 1)) \\tag {1}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3248" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.396, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.419, + 0.089, + 0.915, + 0.259 + ], + "angle": 0, + "content": "
BenchmarkSize (Fin)Domain CoverageModalitiesQuestion Type
MMMU [65]11,550 (1,603)10T+IMC
MMMU-Pro [66]1,730 (286)10T+I, P.I.MC
FinanceMath [67]1,200 (1,200)10TNUM
CodeTAT-QA [32]3,144 (3,144)6TNUM
CodeFinQA [32]5,463 (5,463)13TNUM
DocMath-Eval [68]4,000 (4,000)12TNUM
FAMMA [63]1,758 (1,758)8T+IMC, NUM
CRRQA (ours)2,150 (2,150)13T+I, P.I.NUM
FinMMR (ours)4,300 (4,300)14T+I, P.I.NUM
" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.269, + 0.908, + 0.327 + ], + "angle": 0, + "content": "Figure 3. The comparison between finance-related benchmarks. These benchmarks vary in size, domain coverage, modalities, and question type, with some focusing on text-only data while others include images. Each axis has scale labels with varying ranges to measure the number of questions from each benchmark across different subdomains. In the Modalities, T means text input, I means Images input, P.I. means pure images input. In the Question Type, MC means multi-choice answer, NUM means numerical answer." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.352, + 0.484, + 0.472 + ], + "angle": 0, + "content": "FinMMR is classified as Easy (1,300 examples), Medium (1,500 examples) and Hard (1,500 examples) based on this formula, with each level randomly split into test and validation sets. All questions are publicly available, while only the 300 validation answers per level are released, while test answers remain private to prevent data leakage [22, 50, 52]. We maintain an online evaluation platform that enables researchers to assess their models." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.485, + 0.437, + 0.501 + ], + "angle": 0, + "content": "3.3. Comparisons with Existing Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.507, + 0.483, + 0.719 + ], + "angle": 0, + "content": "As illustrated in Fig. 3, previous work has studied multidiscipline multimodal reasoning [65, 66], general mathematical reasoning [39] or text-based financial QA [32, 67, 68]. FinMMR focus on multimodal financial numerical reasoning, curating 4,300 questions requiring a deep understanding of domain-specific images (e.g., earnings reports, candlestick charts). To mimic real-world scenario, we deliberately incorporate 3,938 distractors into 2,150 questions to rigorously evaluate MLLMs' visual perception capability. Compared to existing financial benchmarks, they suffer from narrow domain coverage [14, 69]. FinMMR encompasses 14 financial subdomains and 14 image categories, comprehensively evaluating MLLMs' domain-specific reasoning capabilities." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.735, + 0.273, + 0.753 + ], + "angle": 0, + "content": "4. Evaluation System" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.761, + 0.483, + 0.807 + ], + "angle": 0, + "content": "To facilitate the evaluation of complex reasoning on FinMMR, we established a dedicated evaluation system. All MLLMs evaluated were accessed through official APIs." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.818, + 0.41, + 0.835 + ], + "angle": 0, + "content": "4.1. Multimodal Large Language Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.903 + ], + "angle": 0, + "content": "We systematically evaluate the multimodal reasoning capabilities of 15 recent MLLMs under the zero-shot setting. The evaluated MLLMs are: OpenAI o1 [46], GPT-4o [44], Claude 3.7 Sonnet (including thinking mode) [5]," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.352, + 0.906, + 0.43 + ], + "angle": 0, + "content": "Gemini 2.0 Flash Thinking [19], Gemini 2.0 Pro [20], Gemini 2.0 Flash [18], InternVL2.5-78B [16], Grok 2 Vision [61], Pixtral Large [2], Qwen2.5-VL-72B [7], QVQ72B-Preview [55], Qwen-Omni-Turbo [57], Llama 4 Maverick [4], Gemma 3 27B [17], and Mistral Small 3.1 [3]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.438, + 0.706, + 0.455 + ], + "angle": 0, + "content": "4.2. Prompting Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.46, + 0.906, + 0.537 + ], + "angle": 0, + "content": "Our evaluation adopts CoT [60], PoT [13] and IO (without any external prompts) prompting methods. Due to budget constraints, we report OpenAI o1 performance on the Hard subset only. Detailed prompts can be found in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.548, + 0.817, + 0.563 + ], + "angle": 0, + "content": "4.3. Answer Extraction and Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.57, + 0.906, + 0.676 + ], + "angle": 0, + "content": "Following Zhao et al. [67], we extract answers based on the prompting methods. For CoT and IO outputs, we employ GPT-4o-mini to extract numerical answers. For PoT, we run the generated program for numerical results. Finally, we conduct a strict accuracy evaluation, comparing numerical results with ground truth and deeming the results accurate within a stringent error tolerance of \\(0.2\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.691, + 0.646, + 0.708 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.717, + 0.906, + 0.838 + ], + "angle": 0, + "content": "We answer the following research questions (RQs): RQ1: Are MLLMs multimodal reasoners with extended thinking and PoT prompts? RQ2: What are the primary challenges facing MLLMs? RQ3: How can the visual perception difficulties of MLLMs be mitigated? RQ4: How can the knowledge reasoning capabilities of MLLMs be enhanced? RQ5: How can the numerical computation abilities of MLLMs be compensated for?" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.706, + 0.865 + ], + "angle": 0, + "content": "5.1. Main Results (RQ1)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.903 + ], + "angle": 0, + "content": "The performance of the MLLMs evaluated using different prompting methods on FinMMR is shown in Tab. 2." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3249" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.099, + 0.089, + 0.908, + 0.351 + ], + "angle": 0, + "content": "
ModelSizeExtended thinkingHardMediumEasyAvg.Token (M)
IOCoTPoTCoTPoTCoTPoTCoTPoTCoTPoT
Proprietary MLLMs
Claude 3.7 Sonnet✓ (64K)53.0051.0051.4062.5062.1778.5078.5064.0064.028.5111.25
Claude 3.7 SonnetX49.8050.8048.5062.2558.8377.0076.9263.3561.420.990.89
OpenAI o1✓48.0048.4044.70------2.522.12
GPT-4oX-45.4047.8063.3359.9278.0076.0062.2461.240.850.41
Gemini 2.0 ProX-46.5047.3060.5857.9275.5075.6760.8660.300.850.45
Gemini 2.0 Flash Thinking✓-46.0046.0060.7556.5877.1774.1761.3158.921.300.48
Gemini 2.0 FlashX-44.4045.9057.8353.4274.9273.7559.0557.690.790.43
Grok 2 VisionX-27.8025.5041.5035.8373.0872.8347.4644.721.130.60
Qwen-Omni-TurboX-17.5027.3035.8348.0057.5061.6736.9445.660.900.42
Open-source MLLMs
Llama 4 Maverick17BX-48.7047.8063.2559.1777.8377.8363.2661.600.880.47
Qwen2.5-VL-72B72BX-43.3046.2063.4264.1777.4275.8361.3862.071.050.44
InternVL2.5-78B78BX-37.4044.0060.5061.1770.9270.5856.2758.58--
QVQ-72B-Preview72B✓43.3040.306.2055.679.6775.4212.4257.139.435.435.70
Pixtral Large124BX-19.7025.0039.8339.7570.0070.1743.1844.971.150.75
Gemma 3 27B27BX-23.4022.3045.1736.4269.0861.5845.8840.100.970.47
Mistral Small 3.124BX-19.7015.2038.4229.7567.6749.4241.9331.461.150.60
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.362, + 0.907, + 0.406 + ], + "angle": 0, + "content": "Table 2. Results of different models using IO, CoT and PoT prompting methods on the test set of FinMMR. We use the best Accuracy on the Hard subset as the ranking indicator of model performance. The results underscore the superior performance of reasoning-enhanced MLLM (i.e., Claude 3.7 Sonnet with 64K extended thinking) with PoT in complex multimodal numerical reasoning task." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.43, + 0.482, + 0.672 + ], + "angle": 0, + "content": "Challenges of MLLMs in Domain-Specific Complex Numerical Reasoning As the difficulty increases, the average accuracy shows a continuous and significant decline. In CoT setting, the average accuracy rates on the Easy, Medium, and Hard subsets are \\(73.33\\%\\), \\(54.05\\%\\), and \\(38.14\\%\\), respectively. The currently best-performing model (i.e., Claude 3.7 Sonnet with 64K extended thinking) consistently demonstrates superior performance across all difficulty subsets using CoT prompting method. However, its accuracy on the Hard subset remains below the \\(60\\%\\) passing threshold in both prompting methods. On the overall test set, Claude 3.7 Sonnet achieves only \\(64.02\\%\\) accuracy. These results highlight the challenging nature and rigorous standards of FinMMR, effectively assessing the limits of MLLMs' reasoning capabilities and the disparities among models compared to previous benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.673, + 0.482, + 0.809 + ], + "angle": 0, + "content": "Does extended thinking help? Reasoning-enhanced models show consistent improvements, compared with non-reasoning-enhanced MLLMs. Claude 3.7 Sonnet with 64K extended thinking achieves 2.20 percentage points higher accuracy than the model without extended thinking (53.00% vs. 50.80%) on the Hard subset. This enhancement comes at the cost of using nearly 12 times more tokens for intricate thinking (4.06M vs. 0.34M). This trend also persists in the Gemini 2.0 Flash series." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.811, + 0.482, + 0.902 + ], + "angle": 0, + "content": "We observe that QVQ-72B-Preview lose basic code generation capabilities due to the reinforcement learning of text-based long thought, which is likely attributed to biases in training strategies and training data. On the Hard subset, this model achieves a code execution success rate of only \\(10.90\\%\\), resulting in an accuracy of merely \\(6.20\\%\\) in" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.43, + 0.905, + 0.505 + ], + "angle": 0, + "content": "PoT setting, significantly lower than the \\(40.30\\%\\) accuracy achieved with CoT. This finding highlights the importance of maintaining foundational capabilities, such as programming, while enhancing the reasoning abilities of MLLMs, to avoid rendering them ineffective in other basic tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.905, + 0.809 + ], + "angle": 0, + "content": "Does PoT help? Experimental results strongly validate the superiority of PoT over CoT in numerical reasoning tasks, especially on the Hard subset. PoT (not including QVQ-72B-Preview) achieves a mean accuracy of \\(37.64\\%\\) versus \\(36.20\\%\\) for CoT, representing an improvement of 1.44 percentage points. Furthermore, PoT encourages MLLMs to leverage structured code generation to reduce token consumption during reasoning. Under similar or lower token usage, PoT achieves similar or greater accuracy. For instance, GPT-4o achieves a 2.40 percentage points improvement in accuracy over CoT while consuming significantly fewer tokens in PoT setting. Similarly, Qwen2.5-VL-72B demonstrates the most pronounced efficiency gains: PoT improves accuracy to \\(64.17\\%\\) from \\(63.42\\%\\) while reducing token consumption by \\(58.88\\%\\) (153K vs. 373K) on the Medium subset. When addressing complex numerical reasoning problems, PoT avoids precise numerical calculations by utilizing external tools (i.e., Python interpreter) and reduces the need for repetitive text-based reasoning, which is beneficial for most MLLMs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.905, + 0.901 + ], + "angle": 0, + "content": "However, we also observe that for certain reasoning-enhanced models (i.e., Claude 3.7 Sonnet with 64K extended thinking and OpenAI o1), due to the enforced requirement for long thought, they still engage in extensive text-based reasoning before generating code even in PoT setting, resulting in exceptionally high token consumption" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3250" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.907, + 0.19 + ], + "angle": 0, + "content": "
SubsetTestDegradationValidationDegradation
Ground Images (%)Distractor Images (%)Ground Images (%)Distractor Images (%)
Hard57.1847.23↓9.9556.7445.58↓11.16
Medium73.0161.36↓11.6577.7864.73↓12.35
Easy61.5953.64↓7.9560.6151.52↓9.09
The improvement achieved by the filtering-reasoning pipeline on the Medium validation set: 64.73 → 71.56 ↑6.83
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.2, + 0.907, + 0.228 + ], + "angle": 0, + "content": "Table 3. Degradation of Qwen2.5-VL-72B on all subsets due to distractor images and improvement achieved by the filtering-reasoning pipeline on the Medium validation set in PoT setting." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.255, + 0.482, + 0.452 + ], + "angle": 0, + "content": "on the Hard subset (4.48M and 2.12M), which is more than 10 times that of other MLLMs. To further investigate this, we added an IO baseline without any external prompts for reasoning-enhanced models on the Hard subset. The IO group achieved the highest accuracy, which we attribute to the comprehensive built-in system prompts embedded in the tested proprietary models. This highlights the need for future research to balance reasoning performance with the control of inefficient and redundant token generation, aiming to achieve a good trade-off between performance and cost, as well as to investigate whether PoT prompting can yield significant performance gains on open-source reasoning-enhanced models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.463, + 0.295, + 0.479 + ], + "angle": 0, + "content": "5.2. Error Analysis (RQ2)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.485, + 0.483, + 0.591 + ], + "angle": 0, + "content": "To better analyze the capabilities and limitations of MLLMs on FinMMR, we conduct a detailed error analysis for the Claude 3.7 Sonnet with 64K extended thinking in PoT setting. Error analysis is based on 100 sampled failure cases, which we categorize into three main error types, some of which involve compound errors. More details of error cases are provided in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.595, + 0.483, + 0.669 + ], + "angle": 0, + "content": "- Visual Perception Error (30/100): The model incorrectly perceives, identifies, or interprets visual information from images, or mistakenly recognizes incorrect data, subsequently causing errors in calculations, broken reasoning chains, or incorrect conclusions." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.67, + 0.483, + 0.731 + ], + "angle": 0, + "content": "- Knowledge Reasoning Error (38/100): Due to insufficient domain-specific knowledge, the model exhibits logical confusion or conceptual misunderstandings during reasoning, leading to incorrect answers." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.731, + 0.483, + 0.806 + ], + "angle": 0, + "content": "- Numerical Computation Error (32/100): In problems involving mathematical operations and numerical reasoning, the model produces significant deviations from the correct answers due to errors in the calculation steps, precision control, or numerical hallucination." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.595, + 0.483, + 0.806 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.411, + 0.835 + ], + "angle": 0, + "content": "5.3. Visual Filtering for Reasoning (RQ3)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "As shown in Tab. 3, when processing multi-image inputs containing distractor images, Qwen2.5-VL-72B demonstrates significantly lower accuracy across all difficulty levels compared to ground images scenarios. This finding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.255, + 0.905, + 0.42 + ], + "angle": 0, + "content": "aligns with conclusions from previous work [37, 51], indicating that irrelevant visual information substantially interferes with MLLMs' reasoning capabilities. In particular, the Medium validation set exhibits the most pronounced performance drop (77.78% ground images vs. 64.73% distractor images), attributed to two key characteristics: (1) moderate complexity making visual perception quality the performance bottleneck; (2) semantic relevance between distractors and questions increasing visual filtering difficulty. To address distractor image interference, we propose a two-stage multimodal reasoning pipeline:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.422, + 0.905, + 0.481 + ], + "angle": 0, + "content": "- Visual Filtering: We first instruct MLLM to analyze the set of images \\(\\mathcal{I}\\) and the question \\(q\\), assessing the relevance of the image (relevant/irrelevant). Irrelevant images are excluded from subsequent reasoning." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.482, + 0.905, + 0.543 + ], + "angle": 0, + "content": "- Enhanced Reasoning: Then, the filtered subset \\(\\mathcal{I}'\\) and the question \\(q\\) are input into the MLLM for the final reasoning. The system automatically reverts to the original inputs \\(\\mathcal{I}\\) if all images are mistakenly filtered." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.422, + 0.905, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.544, + 0.907, + 0.756 + ], + "angle": 0, + "content": "Does the pipeline help? As illustrated in Tab. 3, we evaluate Qwen2.5-VL-72B on the 207 English questions with distractor images of the Medium validation set. Our method achieves an overall accuracy of \\(71.56\\%\\), representing a 6.83 percentage points improvement over direct reasoning. This result is only 6.22 percentage points away from the ideal accuracy of the ground images scenarios \\((77.78\\%)\\). Detailed analysis reveals \\(73.43\\%\\) (152/207) ground image recognition accuracy during filtering. When correctly identified, the accuracy of these problems increases to \\(81.58\\%\\) (124/152). This finding underscores the necessity to enhance the ability of MLLMs to filter out irrelevant image information, thereby strengthening their robustness in reasoning within more complex real-world scenarios." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.768, + 0.803, + 0.784 + ], + "angle": 0, + "content": "5.4. Knowledge Augmentation (RQ4)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.791, + 0.905, + 0.852 + ], + "angle": 0, + "content": "To enhance the understanding and application capabilities of financial knowledge of MLLMs, we explore a method of refined knowledge augmentation to improve the performance of MLLM in domain-specific reasoning tasks." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "- Function Library Construction: We annotate a comprehensive financial function library containing 3,133 Python functions from financial encyclopedia. Each func" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3251" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.089, + 0.48, + 0.184 + ], + "angle": 0, + "content": "
SettingPoTRAG with PoT
Gemini 2.0 Flash Thinking78.7183.02 (+4.31)
GPT-4o80.6083.62 (+3.02)
Claude 3.7 Sonnet81.2185.43 (+4.22)
Claude 3.7 Sonnet (64K)83.5386.29 (+2.76)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.193, + 0.483, + 0.224 + ], + "angle": 0, + "content": "Table 4. Improvements of different MLLMs with knowledge augmentation on the 1,160 instances of FinMMR in PoT setting." + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.248, + 0.483, + 0.278 + ], + "angle": 0, + "content": "tion includes precise functional descriptions, parameter explanations, and step-by-step implementation code." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.279, + 0.483, + 0.445 + ], + "angle": 0, + "content": "- MLLM-Instructed Knowledge Retrieval: In financial problems with hybrid contexts, using short questions or full contexts for retrieval often fails to retrieve directly relevant knowledge [9, 49]. We observed that powerful MLLMs (e.g., GPT-40) can effectively summarize rich semantic information from contexts. Therefore, we first prompt the MLLM to generate precise retrieval queries based on the context [34, 58]. Then we use Contriever [28] to retrieve relevant financial Python functions, based on the semantic similarity between the refined queries and functional descriptions." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.445, + 0.484, + 0.537 + ], + "angle": 0, + "content": "- MLLM as Retrieval Judge: Recent studies have shown that models are capable of judging the relevance of candidates retrieved for the question [26]. In this setting, we first retrieved the Top-30 financial functions and then prompted the MLLM to select the Top-3 functions most useful to answer the question, if any." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.279, + 0.484, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.538, + 0.483, + 0.78 + ], + "angle": 0, + "content": "Does knowledge augmentation help? As shown in Tab. 4, all evaluated MLLMs enhanced with financial function knowledge achieved significant improvements, ranging from 2.76 to 4.31 percentage points. Leveraging the improved retrieval efficiency enabled by MLLM-Instructed Knowledge Retrieval and MLLM as Retrieval Judge, the knowledge augmentation approach achieves greater performance, boosting the accuracy of Claude 3.7 Sonnet with 64K thinking to \\(86.29\\%\\). Notably, Gemini 2.0 Flash Thinking, which has relatively weaker reasoning capabilities, also improved from \\(78.71\\%\\) to \\(83.02\\%\\), approaching the performance of Claude 3.7 Sonnet \\((83.53\\%)\\) without knowledge augmentation. The results further illustrate that refined domain-specific reasoning knowledge can significantly enhance the performance of MLLMs in complex reasoning tasks within expert domains." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.788, + 0.398, + 0.805 + ], + "angle": 0, + "content": "5.5. Visual Parser with Reasoner (RQ5)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In complex multimodal numerical reasoning tasks, single models often struggle to simultaneously achieve visual perception, knowledge reasoning, and numerical computation. To investigate the potential of model collaboration, we instruct the MLLM to act as the Visual Parser, responsible for carefully converting images into structured textual data." + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.089, + 0.906, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.248, + 0.9, + 0.263 + ], + "angle": 0, + "content": "Figure 4. Results of model combinations and individual models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.289, + 0.905, + 0.319 + ], + "angle": 0, + "content": "Then, an LRM acts as the Reasoner, performing multi-step numerical reasoning based on the textual context." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.32, + 0.905, + 0.425 + ], + "angle": 0, + "content": "Specifically, we filter 1,160 tabular QA instances from FinMMR and utilize GPT-4o as the Visual Parser, instructing it to separate table headers or cells with vertical bars (|) and rows with newlines. For the Reasoner component, in addition to GPT-4o, we evaluate several LRMs (i.e., Claude 3.7 Sonnet, Gemini 2.0 Flash Thinking, DeepSeek-R1, and OpenAI o1)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.425, + 0.906, + 0.636 + ], + "angle": 0, + "content": "Does model collaboration help? As shown in Fig. 4, the structured data from GPT-4o's visual parsing significantly enhances downstream reasoning. The individual model (i.e., GPT-4o with PoT) achieves an accuracy of \\(80.60\\%\\), while the combination of models improves the accuracy to \\(86.72\\%\\) (i.e., DeepSeek-R1 serving as the Reasoner with PoT). Performance variance emerges across reasoning-enhanced models using identical visual inputs. Claude 3.7 Sonnet reaches \\(85.52\\%\\), outperforming Gemini 2.0 Flash Thinking \\((82.24\\%)\\), confirming the decisive impact of text-based reasoning capabilities. This evidences that model collaboration effectively compensates for individual model limitations through complementary strengths." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.649, + 0.634, + 0.665 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We introduce FinMMR, a multimodal, comprehensive, and challenging benchmark for evaluating the financial numerical reasoning capabilities of MLLMs. FinMMR challenges MLLMs' intricate visual perception, specialized knowledge reasoning, and accurate numerical computation through its rich images, comprehensive subdomains, and complex formulas embedded in each multimodal financial question. The evaluation results reveal that 15 proprietary and open-source MLLMs still struggle with complex multimodal reasoning tasks in specialized domains. FinMMR highlights the bottlenecks of MLLMs and the need for continuous improvements, including reasoning-enhanced training, tool use, refined structured knowledge augmentation and model combinations, allowing models to perform expert-level reasoning tasks closer to the real-world scenarios." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3252" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.486, + 0.268 + ], + "angle": 0, + "content": "This work is supported by the National Natural Science Foundation of China (Grant Nos. 62176026, 62473271), the Beijing Natural Science Foundation (Grant No. QY24214), the BUPT Innovation and Entrepreneurship Support Program (Grant Nos. 2025-YC-A033, 2025-YC-A042), and data support from Hithink RoyalFlush Information Network Co., Ltd. This work is also supported by the Engineering Research Center of Information Networks, Ministry of Education, China. We would also like to thank the anonymous reviewers and area chairs for constructive discussions and feedback." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.296, + 0.188, + 0.312 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.321, + 0.482, + 0.362 + ], + "angle": 0, + "content": "[1] 360AILAB-NLP. 360layoutanalysis. https://github.com/360AILAB-NLP/360LayoutAnalysis, 2024. Accessed: 2024-06-15. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.364, + 0.482, + 0.405 + ], + "angle": 0, + "content": "[2] Mistral AI. Pixtral large. https://mistral.ai/news/pixtral-large, 2024. Accessed: 2024-11-18. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.408, + 0.482, + 0.448 + ], + "angle": 0, + "content": "[3] Mistral AI. Mistral small 3.1. https://mistral.ai/news/mistral-small-3-1, 2025. Accessed: 2025-03-17. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.451, + 0.482, + 0.505 + ], + "angle": 0, + "content": "[4] AI@Meta. The llama 4 herd: The beginning of a new era of natively multimodal ai innovation. https://ai.meta.com/blog/llama-4-multimodal-intelligence, 2025. Accessed: 2025-04-05. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.508, + 0.482, + 0.548 + ], + "angle": 0, + "content": "[5] Anthropic. Claude 3.7 sonnet and claude code. https://www.anthropic.com/news/claude-3-7-sonnet, 2025. Accessed: 2025-02-25. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.482, + 0.619 + ], + "angle": 0, + "content": "[6] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.622, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[7] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zhenen Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Yang, Haiyang Xu, and Junyang Lin. Qwen2.5-vl technical report. arXiv preprint, 2025. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.72, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[8] Sébastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, Harsha Nori, Hamid Palangi, Marco Tulio Ribeiro, and Yi Zhang. Sparks of artificial general intelligence: Early experiments with gpt-4. arXiv preprint, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.805, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[9] Liang Chen, Yang Deng, Yatao Bian, Zeyu Qin, Bingzhe Wu, Tat-Seng Chua, and Kam-Fai Wong. Beyond factuality: A comprehensive evaluation of large language models as knowledge generators. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 6325-6341, Singapore, 2023. Association for Computational Linguistics. 8" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.321, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.175 + ], + "angle": 0, + "content": "[10] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, and Feng Zhao. Are we on the right way for evaluating large vision-language models? In Advances in Neural Information Processing Systems, pages 27056-27087. Curran Associates, Inc., 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.905, + 0.426 + ], + "angle": 0, + "content": "[11] Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott Gray, Nick Ryder, Mikhail Pavlov, Alethea Power, Lukasz Kaiser, Mohammad Bavarian, Clemens Winter, Philippe Tillet, Felipe Petroski Such, Dave Cummings, Matthias Plappert, Fotios Chantzis, Elizabeth Barnes, Ariel Herbert-Voss, William Hebgen Guss, Alex Nichol, Alex Paino, Nikolas Tezak, Jie Tang, Igor Babuschkin, Suchir Balaji, Shantanu Jain, William Saunders, Christopher Hesse, Andrew N. Carr, Jan Leike, Josh Achiam, Vedant Misra, Evan Morikawa, Alec Radford, Matthew Knight, Miles Brundage, Mira Murati, Katie Mayer, Peter Welinder, Bob McGrew, Dario Amodei, Sam McCandlish, Ilya Sutskever, and Wojciech Zaremba. Evaluating large language models trained on code. arXiv preprint, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.428, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[12] Pengcheng Chen, Jin Ye, Guoan Wang, Yanjun Li, Zhongying Deng, Wei Li, Tianbin Li, Haodong Duan, Ziyan Huang, Yanzhou Su, Benyou Wang, Shaoting Zhang, Bin Fu, Jianfei Cai, Bohan Zhuang, Eric J Seibel, Yu Qiao, and Junjun He. Gmai-mmbench: A comprehensive multimodal evaluation benchmark towards general medical ai. In Advances in Neural Information Processing Systems, pages 94327-94427. Curran Associates, Inc., 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.539, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[13] Wenhu Chen, Xueguang Ma, Xinyi Wang, and William W. Cohen. Program of thoughts prompting: Disentangling computation from reasoning for numerical reasoning tasks. Transactions on Machine Learning Research, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.596, + 0.905, + 0.706 + ], + "angle": 0, + "content": "[14] Zhiyu Chen, Wenhu Chen, Charese Smiley, Sameena Shah, Iana Borova, Dylan Langdon, Reema Moussa, Matt Beane, Ting-Hao Huang, Bryan Routledge, and William Yang Wang. FinQA: A dataset of numerical reasoning over financial data. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3697-3711, Online and Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.707, + 0.905, + 0.803 + ], + "angle": 0, + "content": "[15] Zhiyu Chen, Shiyang Li, Charese Smiley, Zhiqiang Ma, Sameena Shah, and William Yang Wang. ConvFinQA: Exploring the chain of numerical reasoning in conversational finance question answering. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 6279-6292, Abu Dhabi, United Arab Emirates, 2022. Association for Computational Linguistics. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.805, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[16] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, Lixin Gu, Xuehui Wang, Qingyun Li, Yimin Ren, Zixuan Chen, Jiapeng Luo, Jiahao Wang, Tan Jiang, Bo Wang, Conghui He, Botian Shi, Xingcheng Zhang, Han Lv, Yi Wang, Wenqi Shao, Pei Chu, Zhongying Tu, Tong He, Zhiyong Wu, Huipeng Deng, Jiaye Ge, Kai Chen, Kaipeng" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3253" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.483, + 0.162 + ], + "angle": 0, + "content": "Zhang, Limin Wang, Min Dou, Lewei Lu, Xizhou Zhu, Tong Lu, Dahua Lin, Yu Qiao, Jifeng Dai, and Wenhai Wang. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint, 2025. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[17] Google DeepMind. Introducing gemma 3: The most capable model you can run on a singlegpu or tpu. https://blog.google/technology/developers/gemma-3, 2025. Accessed: 2025-03-12. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.482, + 0.274 + ], + "angle": 0, + "content": "[18] Google DeepMind. Gemini 2.0: Flash, flash-lite and pro. https://developers.googleblog.com/en/gemini-2-family-expands, 2025. Accessed: 2025-02-05. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.276, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[19] Google DeepMind. Gemini 2.0 flash thinking. https://ai.google.dev/gemini-api/docs/thinking, 2025. Accessed: 2025-02-05. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.318, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[20] Google DeepMind. Gemini 2.0. https://cloud.google.com/vertex-ai/generative-ai/docs/gemini-v2#2.0-pro, 2025. Accessed: 2025-02-05. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.361, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[21] DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, Aixin Liu, Bing Xue, Bingxuan Wang, Bochao Wu, Bei Feng, Chengda Lu, Chenggang Zhang, Chengqi Deng, Chenyu Zhang, Chong Ruan, Damai Dai, Deli Chen, Dongjie Ji, Erhang Li, Fangyun Lin, Fucong Dai, Fuli Luo, Guangbo Hao, Guanting Chen, Guowei Li, H. Zhang, Han Bao, Hanwei Xu, Haocheng Wang, Honghui Ding, Huajian Xin, Huazuo Gao, Hui Qu, Hui Li, Jianzhong Guo, Jiashi Li, Jiawei Wang, Jingchang Chen, Jingyang Yuan, Junjie Qiu, Junlong Li, J. L. Cai, Jiaqi Ni, Jian Liang, Jin Chen, Kai Dong, Kai Hu, Kaige Gao, Kang Guan, Kexin Huang, Kuai Yu, Lean Wang, Lecong Zhang, Liang Zhao, Litong Wang, Liyue Zhang, Lei Xu, Leyi Xia, Mingchuan Zhang, Minghua Zhang, Minghui Tang, Meng Li, Miaojun Wang, Mingming Li, Ning Tian, Panpan Huang, Peng Zhang, Qiancheng Wang, Qinyu Chen, Qiushi Du, Ruiqi Ge, Ruisong Zhang, Ruizhe Pan, Runji Wang, R. J. Chen, R. L. Jin, Ruyi Chen, Shanghao Lu, Shangyan Zhou, Shanhuang Chen, Shengfeng Ye, Shiyu Wang, Shuiping Yu, Shunfeng Zhou, Shuting Pan S. S. Li Shuang Zhou Shaoqing Wu Shengfeng Ye Tao Yun Tian Pei Tianyu Sun T. Wang Wangding Zeng Wanjia Zhao Wen Liu Wenfeng Liang Wenjun Gao Wenqin Yu Wentao Zhang W. L. Xiao Wei An Xiaodong Liu Xiaohan Wang Xiaokang Chen Xiaotao Nie Xin Cheng Xin Liu Xin Xie Xingchao Liu Xinyu Yang Xinyuan Li Xuecheng Su Xuheng Lin X.Q.Li Xiangyue Jin Xiaojin Shen Xiaosha Chen Xiaowen Sun Xiaoxiang Wang Xinnan Song Xinyi Zhou Xianzu Wang Xinxia Shan Y.K. Li Y.Q. WangY.X.Wei,Yang Zhang Yanhong Xu Yao Li Yao Zhao Yaofeng Sun Yaohui Wang Yi Yu Yichao Zhang Yifan Shi Yiliang Xiong Ying He Yishi Piao Yisong Wang Yixuan Tan Yiyang Ma Yiyuan Liu Yongqiang Guo Yuan OuYuduan Wang Yue Gong Yuheng Zou Yujia HeYunfan Xiong Yuxiang Luo Yuxiang You Yuxuan Liu Yuyang ZhouY.X.ZhuYanhong Xu Yanping Huang Yaohui Li Yi ZhengYuchen ZhuYunxian Ma Ying Tang Yukun" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.905, + 0.189 + ], + "angle": 0, + "content": "Zha, Yuting Yan, Z. Z. Ren, Zehui Ren, Zhangli Sha, Zhe Fu, Zhean Xu, Zhenda Xie, Zhengyan Zhang, Zhenen Hao, Zhicheng Ma, Zhigang Yan, Zhiyu Wu, Zihui Gu, Zijia Zhu, Zijun Liu, Zilin Li, Ziwei Xie, Ziyang Song, Zizheng Pan, Zhen Huang, Zhipeng Xu, Zhongyu Zhang, and Zhen Zhang. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint, 2025. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.19, + 0.907, + 0.3 + ], + "angle": 0, + "content": "[22] Chunyuan Deng, Yilun Zhao, Xiangru Tang, Mark Gerstein, and Arman Cohan. Investigating data contamination in modern benchmarks for large language models. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 8706-8719, Mexico City, Mexico, 2024. Association for Computational Linguistics. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.301, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[23] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, Yunsheng Wu, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.371, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[24] Ziliang Gan, Yu Lu, Dong Zhang, Haohan Li, Che Liu, Jian Liu, Ji Liu, Haipang Wu, Chaoyou Fu, Zenglin Xu, Rongjunchen Zhang, and Yong Dai. Mme-finance: A multimodal finance benchmark for expert-level understanding and reasoning. arXiv preprint, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.441, + 0.905, + 0.509 + ], + "angle": 0, + "content": "[25] Yingqiang Ge, Wenyue Hua, Kai Mei, jianchao ji, Juntao Tan, Shuyuan Xu, Zelong Li, and Yongfeng Zhang. Opennagi: When llm meets domain experts. In Advances in Neural Information Processing Systems, pages 5539-5568. Curran Associates, Inc., 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.579 + ], + "angle": 0, + "content": "[26] Jian Guan, Wei Wu, Zujie Wen, Peng Xu, Hongning Wang, and Minlie Huang. Amor: A recipe for building adaptable modular knowledge agents through process feedback. In Advances in Neural Information Processing Systems, pages 126118-126148. Curran Associates, Inc., 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[27] Yunzhuo Hao, Jiawei Gu, Huichen Will Wang, Linjie Li, Zhengyuan Yang, Lijuan Wang, and Yu Cheng. Can MLLMs reason in multimodality? EMMA: An enhanced multimodal reasoning benchmark. In *Forty-second International Conference on Machine Learning*, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[28] Gautier Izacard, Mathilde Caron, Lucas Hosseini, Sebastian Riedel, Piotr Bojanowski, Armand Joulin, and Edouard Grave. Unsupervised dense information retrieval with contrastive learning. Transactions on Machine Learning Research, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.72, + 0.905, + 0.803 + ], + "angle": 0, + "content": "[29] Naman Jain, King Han, Alex Gu, Wen-Ding Li, Fanjia Yan, Tianjun Zhang, Sida Wang, Armando Solar-Lezama, Koushik Sen, and Ion Stoica. Livecodebench: Holistic and contamination free evaluation of large language models for code. In The Thirteenth International Conference on Learning Representations, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[30] Dongzhi Jiang, Renrui Zhang, Ziyu Guo, Yanwei Li, Yu Qi, Xinyan Chen, Liuhui Wang, Jianhan Jin, Claire Guo, Shen Yan, Bo Zhang, Chaoyou Fu, Peng Gao, and Hongsheng Li. MME-cot: Benchmarking chain-of-thought in large multimodal models for reasoning quality, robustness, and efficiency. In Forty-second International Conference on Machine Learning, 2025. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3254" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[31] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.149, + 0.482, + 0.244 + ], + "angle": 0, + "content": "[32] Michael Krumdick, Rik Koncel-Kedziorski, Viet Dac Lai, Varshini Reddy, Charles Lovering, and Chris Tanner. BizBench: A quantitative reasoning benchmark for business and finance. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8309-8332, Bangkok, Thailand, 2024. Association for Computational Linguistics. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.247, + 0.483, + 0.315 + ], + "angle": 0, + "content": "[33] Bohao Li, Yuying Ge, Yixiao Ge, Guangzhi Wang, Rui Wang, Ruimao Zhang, and Ying Shan. Seed-bench: Benchmarking multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13299-13308, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.317, + 0.483, + 0.371 + ], + "angle": 0, + "content": "[34] Xiaoxi Li, Guanting Dong, Jiajie Jin, Yuyao Zhang, Yujia Zhou, Yutao Zhu, Peitian Zhang, and Zhicheng Dou. Searcho1: Agentic search-enhanced large reasoning models. arXiv preprint, 2025. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.373, + 0.482, + 0.44 + ], + "angle": 0, + "content": "[35] Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let's verify step by step. In The Twelfth International Conference on Learning Representations, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.442, + 0.483, + 0.566 + ], + "angle": 0, + "content": "[36] Fuxiao Liu, Xiaoyang Wang, Wenlin Yao, Jianshu Chen, Kaiqiang Song, Sangwoo Cho, Yaser Yacoob, and Dong Yu. MMC: Advancing multimodal chart understanding with large-scale instruction tuning. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 1287-1310, Mexico City, Mexico, 2024. Association for Computational Linguistics. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.568, + 0.483, + 0.663 + ], + "angle": 0, + "content": "[37] Haowei Liu, Xi Zhang, Haiyang Xu, Yaya Shi, Chaoya Jiang, Ming Yan, Ji Zhang, Fei Huang, Chunfeng Yuan, Bing Li, and Weiming Hu. MIBench: Evaluating multimodal large language models over multiple images. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 22417-22428, Miami, Florida, USA, 2024. Association for Computational Linguistics. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.665, + 0.483, + 0.746 + ], + "angle": 0, + "content": "[38] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin. Mmbench: Is your multi-modal model an all-around player? In Computer Vision - ECCV 2024, pages 216-233, Cham, 2025. Springer Nature Switzerland. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.748, + 0.483, + 0.83 + ], + "angle": 0, + "content": "[39] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating math reasoning in visual contexts with gpt-4v, bard, and other large multimodal models. In The Twelfth International Conference on Learning Representations, 2024. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[40] Junyu Luo, Zhizhuo Kou, Liming Yang, Xiao Luo, Jinsheng Huang, Zhiping Xiao, Jingshu Peng, Chengzhong Liu, Ji-aming Ji, Xuanzhe Liu, Sirui Han, Ming Zhang, and Yike Guo. FinMME: Benchmark dataset for financial multi-modal reasoning evaluation. In Proceedings of the 63rd Annual" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 29465-29489, Vienna, Austria, 2025. Association for Computational Linguistics. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[41] Yujun Mao, Yoon Kim, and Yilun Zhou. CHAMP: A competition-level dataset for fine-grained analyses of LLMs' mathematical reasoning capabilities. In Findings of the Association for Computational Linguistics: ACL 2024, pages 13256-13274, Bangkok, Thailand, 2024. Association for Computational Linguistics. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.905, + 0.273 + ], + "angle": 0, + "content": "[42] Grégoire Mialon, Clémentine Fourrier, Thomas Wolf, Yann LeCun, and Thomas Scialom. GAIA: a benchmark for general AI assistants. In The Twelfth International Conference on Learning Representations, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.904, + 0.354 + ], + "angle": 0, + "content": "[43] Meredith Ringel Morris, Jascha Sohl-Dickstein, Noah Fiedel, Tris Warkentin, Allan Dafoe, Aleksandra Faust, Clement Farabet, and Shane Legg. Position: Levels of AGI for operationalizing progress on the path to AGI. In *Forty-first International Conference on Machine Learning*, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.357, + 0.904, + 0.396 + ], + "angle": 0, + "content": "[44] OpenAI. Hello gpt-4o. https://openai.com/index/hello-gpt-4o, 2024. Accessed: 2024-05-13. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.4, + 0.904, + 0.44 + ], + "angle": 0, + "content": "[45] OpenAI. Learning to reason with llms. https://openai.com/index/learning-to-reason-with-llms, 2024. Accessed: 2024-09-12. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.904, + 0.482 + ], + "angle": 0, + "content": "[46] OpenAI. Openai o1 system card. https://openai.com/index/openai-o1-system-card, 2024. Accessed: 2024-12-05. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.484, + 0.904, + 0.536 + ], + "angle": 0, + "content": "[47] OpenAI. Openai o1-mini. https://openai.com/index/openai-ol-mini-advancing-cost-efficient-reasoning, 2024. Accessed: 2024-09-12. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.54, + 0.905, + 0.579 + ], + "angle": 0, + "content": "[48] OpenAI. Openai o3-mini system card. https://openai.com/index/o3-mini-system-card, 2025. Accessed: 2025-01-31. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.582, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[49] Baolin Peng, Michel Galley, Pengcheng He, Hao Cheng, Yujia Xie, Yu Hu, Qiuyuan Huang, Lars Liden, Zhou Yu, Weizhu Chen, and Jianfeng Gao. Check your facts and try again: Improving large language models with external knowledge and automated feedback. arXiv preprint, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.652, + 0.905, + 0.747 + ], + "angle": 0, + "content": "[50] Oscar Sainz, Jon Campos, Iker García-Ferrero, Julien Etxaniz, Oier Lopez de Lacalle, and Eneko Agirre. NLP evaluation in trouble: On the need to measure LLM data contamination for each benchmark. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 10776-10787, Singapore, 2023. Association for Computational Linguistics. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.749, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[51] Aditya Sharma, Michael Saxon, and William Yang Wang. Losing visual needles in image haystacks: Vision language models are easily distracted in short and long contexts. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 5429-5451, Miami, Florida, USA, 2024. Association for Computational Linguistics. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[52] Weijia Shi, Anirudh Ajith, Mengzhou Xia, Yangsibo Huang, Daogao Liu, Terra Blevins, Danqi Chen, and Luke Zettlemoyer. Detecting pretraining data from large language models. In The Twelfth International Conference on Learning Representations, 2024. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3255" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.219 + ], + "angle": 0, + "content": "[53] Zichen Tang, Haihong E, Ziyan Ma, Haoyang He, Jiacheng Liu, Zhongjun Yang, Zihua Rong, Rongjin Li, Kun Ji, Qing Huang, Xinyang Hu, Yang Liu, and Qianhe Zheng. FinanceReasoning: Benchmarking financial numerical reasoning more credible, comprehensive and challenging. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15721-15749, Vienna, Austria, 2025. Association for Computational Linguistics. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.486, + 0.552 + ], + "angle": 0, + "content": "[54] Kimi Team, Angang Du, Bofei Gao, Bowei Xing, Changjiu Jiang, Cheng Chen, Cheng Li, Chenjun Xiao, Chenzhuang Du, Chonghua Liao, Chuning Tang, Congcong Wang, Dehao Zhang, Enming Yuan, Enzhe Lu, Fengxiang Tang, Flood Sung, Guangda Wei, Guokun Lai, Haiqing Guo, Han Zhu, Hao Ding, Hao Hu, Hao Yang, Hao Zhang, Haotian Yao, Haotian Zhao, Haoyu Lu, Haoze Li, Haozhen Yu, Hongcheng Gao, Huabin Zheng, Huan Yuan, Jia Chen, Jianhang Guo, Jianlin Su, Jianzhou Wang, Jie Zhao, Jin Zhang, Jingyuan Liu, Junjie Yan, Junyan Wu, Lidong Shi, Ling Ye, Longhui Yu, Mengnan Dong, Neo Zhang, Ningchen Ma, Qiwei Pan, Qucheng Gong, Shaowei Liu, Shengling Ma, Shupeng Wei, Sihan Cao, Siying Huang, Tao Jiang, Weihao Gao, Weimin Xiong, Weiran He, Weixiao Huang, Weixin Xu, Wenhao Wu, Wenyang He, Xianghui Wei, Xianqing Jia, Xingzhe Wu, Xinran Xu, Xinxing Zu, Xinyu Zhou, Xuehai Pan, Y. Charles, Yang Li, Yangyang Hu, Yangyang Liu, Yanru Chen, Yejie Wang, Yibo Liu, Yidao Qin, Yifeng Liu, Ying Yang, Yiping Bao, Yulun Du, Yuxin Wu, Yuzhi Wang, Zaida Zhou, Zhaoji Wang, Zhaowei Li, Zhen Zhu, Zheng Zhang, Zhexu Wang, Zhilin Yang, Zhiqi Huang, Zihao Huang, Ziyao Xu, Zonghan Yang, and Zongyu Lin. Kimi k1.5: Scaling reinforcement learning with llms. arXiv preprint, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.555, + 0.484, + 0.597 + ], + "angle": 0, + "content": "[55] Qwen Team. Qvq: To see the world with wisdom. https://qwenlm.github.io/blog/qvq-72b-preview, 2024. Accessed: 2024-12-25. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.599, + 0.484, + 0.641 + ], + "angle": 0, + "content": "[56] Qwen Team. Qwq: Reflect deeply on the boundaries of the unknown. https://qwenlm.github.io/blog/qwq-32b-preview, 2024. Accessed: 2024-11-28. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.643, + 0.482, + 0.685 + ], + "angle": 0, + "content": "[57] Qwen Team. qwen-omni. https://help.aliyun.com/zh/model-studio/user-guide/qwen-omni, 2025. Accessed: 2025-03-26. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.687, + 0.484, + 0.757 + ], + "angle": 0, + "content": "[58] Prakhar Verma, Sukruta Prakash Midigeshi, Gaurav Sinha, Arno Solin, Nagarajan Natarajan, and Amit Sharma. Plan*RAG: Efficient test-time planning for retrieval augmented generation. In Workshop on Reasoning and Planning for Large Language Models, 2025. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.759, + 0.484, + 0.843 + ], + "angle": 0, + "content": "[59] Xiaoxuan Wang, Ziniu Hu, Pan Lu, Yanqiao Zhu, Jieyu Zhang, Satyen Subramaniam, Arjun R Loomba, Shichang Zhang, Yizhou Sun, and Wei Wang. Scibench: Evaluating college-level scientific problem-solving abilities of large language models. In *Forty-first International Conference on Machine Learning*, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[60] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian richter, Fei Xia, Ed Chi, Quoc V Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models. In Advances in Neural Information Pro" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "cessing Systems, pages 24824-24837. Curran Associates, Inc., 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[61] xAI. Models. https://docs.x.ai/docs/models? cluster=us-east-1#model-constraints, 2024. Accessed: 2024-12-12. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[62] Fengli Xu, Qianyue Hao, Zefang Zong, Jingwei Wang, Yunke Zhang, Jingyi Wang, Xiaochong Lan, Jiahui Gong, Tianjian Ouyang, Fanjin Meng, Chenyang Shao, Yuwei Yan, Qinglong Yang, Yiwen Song, Sijian Ren, Xinyuan Hu, Yu Li, Jie Feng, Chen Gao, and Yong Li. Towards large reasoning models: A survey of reinforced reasoning with large language models. arXiv preprint, 2025. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.316 + ], + "angle": 0, + "content": "[63] Siqiao Xue, Tingting Chen, Fan Zhou, Qingyang Dai, Zhixuan Chu, and Hongyuan Mei. Famma: A benchmark for financial domain multilingual multimodal question answering. arXiv preprint, 2024. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.905, + 0.386 + ], + "angle": 0, + "content": "[64] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. MM-vet: Evaluating large multimodal models for integrated capabilities. In *Forty-first International Conference on Machine Learning*, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[65] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Ren-liang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9556-9567, 2024. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.905, + 0.649 + ], + "angle": 0, + "content": "[66] Xiang Yue, Tianyu Zheng, Yuansheng Ni, Yubo Wang, Kai Zhang, Shengbang Tong, Yuxuan Sun, Botao Yu, Ge Zhang, Huan Sun, Yu Su, Wenhu Chen, and Graham Neubig. MMMU-pro: A more robust multi-discipline multimodal understanding benchmark. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15134-15186, Vienna, Austria, 2025. Association for Computational Linguistics. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.65, + 0.905, + 0.747 + ], + "angle": 0, + "content": "[67] Yilun Zhao, Hongjun Liu, Yitao Long, Rui Zhang, Chen Zhao, and Arman Cohan. Financemath: Knowledge-intensive math reasoning in finance domains. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12841-12858, Bangkok, Thailand, 2024. Association for Computational Linguistics. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.747, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[68] Yilun Zhao, Yitao Long, Hongjun Liu, Ryo Kamoi, Linyong Nan, Lyuhao Chen, Yixin Liu, Xiangru Tang, Rui Zhang, and Arman Cohan. DocMath-eval: Evaluating math reasoning capabilities of LLMs in understanding long and specialized documents. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 16103-16120, Bangkok, Thailand, 2024. Association for Computational Linguistics. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[69] Fengbin Zhu, Wenqiang Lei, Youcheng Huang, Chao Wang, Shuo Zhang, Jiancheng Lv, Fuli Feng, and Tat-Seng Chua. TAT-QA: A question answering benchmark on a hybrid of" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "3256" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.124, + 0.092, + 0.487, + 0.177 + ], + "angle": 0, + "content": "tabular and textual content in finance. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 3277-3287, Online, 2021. Association for Computational Linguistics. 2, 3, 5" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.958 + ], + "angle": 0, + "content": "3257" + } + ] +] \ No newline at end of file diff --git a/2025/FinMMR_ Make Financial Numerical Reasoning More Multimodal, Comprehensive, and Challenging/9bdd874d-4c07-4812-a904-cd4ae1c8c840_origin.pdf b/2025/FinMMR_ Make Financial Numerical Reasoning More Multimodal, Comprehensive, and Challenging/9bdd874d-4c07-4812-a904-cd4ae1c8c840_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..45212aec1a24d22cf9ba7561cd4aca71443999fc --- /dev/null +++ b/2025/FinMMR_ Make Financial Numerical Reasoning More Multimodal, Comprehensive, and Challenging/9bdd874d-4c07-4812-a904-cd4ae1c8c840_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7387787a20c96a16799a4cf1262bc63a852a5a0853c2e4ee5b34360bfadd2ffb +size 4337030 diff --git a/2025/FinMMR_ Make Financial Numerical Reasoning More Multimodal, Comprehensive, and Challenging/full.md b/2025/FinMMR_ Make Financial Numerical Reasoning More Multimodal, Comprehensive, and Challenging/full.md new file mode 100644 index 0000000000000000000000000000000000000000..35b52b47727d5aa8aa0f9bb22c6c9276ce8496d4 --- /dev/null +++ b/2025/FinMMR_ Make Financial Numerical Reasoning More Multimodal, Comprehensive, and Challenging/full.md @@ -0,0 +1,397 @@ +# $\mathcal{F}_{\mathsf{M}}$ FinMMR: Make Financial Numerical Reasoning More Multimodal, Comprehensive, and Challenging + +![](images/6d3fed1f0bcad87000c97473484ddeecc534d3a4c21ada62bf3c83cc519932b9.jpg) + +![](images/6144a90463c441bce03a3a2907a21b545073c0b3c5b2f67d5224c79ff4245c54.jpg) +Figure 1. Overview of FinMMR. FinMMR presents three challenges: (1) Visual Perception: 8.7K financial images of 14 categories; (2) Knowledge Reasoning: 4.3K financial questions of 14 subdomains; (3) Numerical Computation: multi-step precise calculations. + +![](images/f24d79bb435f65d9e7c00c37f347d57d67a992fd34dfd9ffccd047a1b42f51e7.jpg) + +![](images/01dee61a627ed956832cce85fa103ad5b11d1ca2ad78aa95efae85d2e2a430fa.jpg) + +# Abstract + +We present FinMMR, a novel bilingual multimodal benchmark tailored to evaluate the reasoning capabilities of multimodal large language models (MLLMs) in financial numerical reasoning tasks. Compared to existing benchmarks, our work introduces three significant advancements. (1) Multimodality: We meticulously transform existing financial reasoning benchmarks, and construct novel questions from the latest Chinese financial research reports. FinMMR comprises 4.3K questions and 8.7K images spanning 14 categories, including tables, bar charts, and ownership structure charts. (2) Comprehensiveness: FinMMR encompasses 14 financial subdomains, including corporate finance, banking, and industry analysis, significantly exceeding existing benchmarks in financial domain knowledge breadth. (3) Challenge: Models are required to perform multi-step precise numerical reasoning by integrating financial knowledge with the understanding of complex financial + +images and text. The best-performing MLLM achieves only $53.0\%$ accuracy on Hard problems. We believe that FinMMR will drive advancements in enhancing the reasoning capabilities of MLLMs in real-world scenarios. + +# 1. Introduction + +Recently, large reasoning models (LRMs) [21, 47, 48, 54, 56, 62], show powerful reasoning capabilities over multi-step reasoning tasks, with train-time scaling and test-time scaling [31, 45]. These reasoning models are proficient in code [11, 29], math [35, 41], and science [59]. Multimodal large language models (MLLMs) [5, 19, 46] also exhibit greater performance on multimodal reasoning [39, 65]. + +Despite significant advancements, there remains a notable gap in understanding the practical applicability of MLLMs in numerical reasoning within real-world scenarios, particularly in high-stakes fields such as finance and healthcare. As shown in Fig. 1, financial analysts in their daily work are required to read visually enriched finan + +# Rich Images Challenge Visual Perception + +问题：计算2016年至2024年Q2期间，上海银行和上市城商行的个人住房贷款年均复合增长率之差，结果以百分比表示，保留两位小数。(Calculate the difference in the compound annual growth rate (CAGR) of personal housing loans between Bank of Shanghai and listed city commercial banks from 2016 to Q2 2024. Present the result as a percentage rounded to two decimal places.) + +![](images/bd2ebb0c3cb9c81d16470e42bf5637c0afce94ef8fdb3223377f345ac18a68b5.jpg) +Answer: 2.17 + +# Compound Average Growth Rate (CAGR) + +问题：请计算惠济新生对日照公司的间接持股比例，结果以百分比表示，保留两位小数。(Please calculate the indirect shareholding ratio of Huiji Xinsheng in Rizhao Company. Present the result as a percentage, rounded to two decimal places.) + +![](images/2c06b20ae45c66588922b3929d805d89c940c6b68154019b66b39aa3f0785f0a.jpg) +Answer: 13.65 + +# Indirect Shareholding Ratio + +问题：请计算2022年第四季度和2023年第一季度的总资本开支，并将其与2021年第四季度的资本开支进行比较，计算下降百分比，结果保留两位小数。 + +(Calculate the total capital expenditure for Q4 2022 and Q1 2023 combined, and compare it to the capital expenditure for Q4 2021. Calculate the percentage decrease, rounded to two decimal places.) + +![](images/1941a902cf9631c3e2f1f6cdf3f7c0af2ac31eecc1e88019fce8aa05dc5eda03.jpg) +Answer: -110.36 +Capital Expenditure +Figure 2. Sampled FinMMR examples with two language (i.e., English and Chinese), rich images and different knowledge. The questions and images need expert-level visual perception, knowledge reasoning and numerical computation. + +# Comprehensive Subdomains Challenge Knowledge Reasoning + +Q: Table 19.3 shows a book balance sheet for the Wishing Well Motel chain. The company's long-term debt is secured by its real estate assets, but it also uses short-term bank loans as a permanent source of financing. It pays $10\%$ interest on the bank debt and $9\%$ interest on the secured debt. Wishing Well has 10 million shares of stock outstanding, trading at $\$ 90$ per share. The expected return on Wishing Well's common stock is $18\%$ . Calculate Wishing Well's WACC. Assume that the book and market values of Wishing Well's debt are the same. The marginal tax rate is $21\%$ . Answer as a percentage to single decimal place. + +# : + +
Cash and marketable securities100Bank loan280
Accounts receivable200Accounts payable120
Inventory50Current liabilities400
Current assets350
Real estate2,100Long-term debt1,800
Other assets150Equity400
Total2,600Total2,600
+ +TABLE 19.3 Book balance sheet for Wishing Well Inc. (figures in $ millions) + +# Weighted Average Cost + +Q: For each of the investments shown in the following table, calculate the rate of return earned over the unspecified time period. What is the rate of return for Investment A? Answer as a percentage to the nearest integer. + +# : + +
InvestmentCash flow during periodBeginning-of-period valueEnd-of-period value
A$-2,800$23,400$20,100
B16,000225,000324,000
C7006,5008,000
D3,58036,60046,500
E-50062,70052,800
+ +# Rate of Return (RoR) + +Q: Ricky is considering purchasing an apartment costing \( \\( {700},{000} \) . He will pay a 30% down payment and take out a mortgage for the remainder. Since he just got married and wants to save some money for future use, he will choose the plan with the lowest monthly payment. After visiting several banks, he received the following mortgage offers: What is the monthly payment for Bank A? Answer to two decimal places. + +# : + +
BankInterest rateTerm (years)
A3.5%15
B320
C425
D4.518
+ +# Complex Formulas Challenge Numerical Computation + +Q: What is the total weighted average Cash and Payment-In-Kind (PIK) interest rate payable under the subordinated and senior notes portfolio in the year 2023? + +# : + +Monthly Payment + +
EUR
In Mio €201820172016%2015
Deutschland3.498.5003.498.5003.498.500100%100%
USA1.200.0001.200.0001.200.000--
Nordafrika, Naher Osten1.200.0001.200.0001.200.000--
Asien-Pazifik1.200.0001.200.0001.200.000--
davon1.200.0001.200.0001.200.000--
davon1.200.0001.200.0001.200.000--
davon1.200.0001.200.001.200.000--
davon1.200.0001.200.0001.200.000--
davon1.200.0001.200.0001.200.000--
deutschland3.498.5003.498.5003.498.500100%100%
davon1.200.0001.200.0001.200.000--
+ +# : + +
[average = (investmenti × cashi) / ∑investmenti] + (investmenti × PIKi) / ∑investmenti]
+ +# Total Weighted Average Cash and + +# Payment-In-Kind (PIK) Interest Rate + +Q: John oversees a fund, with the returns for the first three years displayed below: What will be the holding period return (expressed as a percentage)? Answer to three decimal places. + +# : + +Answer: 12.0 + +
YearInvestmentReturn
1$ 50012%
2$ 6005%
3$ 10001%
+ +# : + +$$ +\left[ H P R = \left(\frac {\sum \left(\text {i n v e s t m e n t} _ {t} \times (1 + \text {r e t u r n} _ {t})\right)}{\sum \text {i n v e s t m e n t}} - 1\right) \times 1 0 0 \right] +$$ + +# Answer: 4.762 + +# Holding Period Return (HPR) + +Q: What is the average quarterly share price in 2021 assuming each quarter had a completely uniform price distribution in high and low bid prices, measured in dollars? + +# : + +
Quarter EndedHigh BidLow Bid
October 31, 2021 (1)$4.51$3.99
July 31, 2021$8.47$2.30
April 30, 2021$13.00$3.80
January 31, 2021$14.50$0.015
October 31, 2020 (2)$0.199$0.01
+ +# : + +$$ +[ \text {A v e r a g e} = \frac {1}{4} \sum_ {i = 1} ^ {4} \frac {\text {H i g h} _ {i} + \text {L o w} _ {i}}{2} ] +$$ + +# Answer: 6.32 + +# Average Quarterly Price + +cial documents, extract key financial indicators from tables, images, and texts, and perform multi-step precise numerical calculations to support professional decision-making. Similarly, to achieve expert artificial general intelligence (AGI) [8, 25, 42, 43, 65], MLLMs are expected to comprehend complex domain-specific images akin to human experts, and apply domain knowledge to perform accurate numerical reasoning. This raises the question: Can current MLLMs seamlessly integrate vision and text to perform + +domain-specific complex reasoning, matching the proficiency of LRMs in pure text-based tasks? + +Specifically, we choose the financial domain to evaluate the complex reasoning capabilities of MLLMs, where precision and transparent reasoning are paramount [32]. Existing numerical reasoning benchmarks for finance are limited in their text-based reasoning, coverage of specific financial knowledge, and complexity of reasoning [14, 15, 32, 67, 69]. FAMMA [63] is mainly modelled after textbooks + +and CFA exam questions, MathVista [39] does not involve the application of financial knowledge, MMMU [65] and MMMU-Pro [66] are restricted to multiple-choice questions, diverging from authentic financial problem-solving. The lack of high-quality, knowledge-intensive multimodal financial numerical reasoning benchmarks makes it challenging to objectively evaluate the actual reasoning capabilities of MLLMs and analyze their shortcomings. + +Therefore, we propose FinMMR, a bilingual multimodal numerical reasoning benchmark designed to evaluate the reasoning capabilities of MLLMs in the finance domain. The benchmark comprises 4.3K problems, covering 14 financial subdomains (e.g., corporate finance and industry analysis), with 8.7K images derived from 14 categories (e.g., tables and ownership structure charts). Each problem includes rich images, an unambiguous question, a Python-formatted solution, and a precise answer. + +For multimodality, without representing financial tables as structured text, FinMMR represent all tables, charts, and diagrams as images. For comprehensiveness, FinMMR covers 14 financial subdomains and two languages (i.e., English and Chinese), demanding domain knowledge such as option pricing and portfolio management. For challenge, FinMMR focus on multi-step numerical reasoning, requiring models to provide exact numerical answers under strict evaluation criteria (emphasizing units, percentages, and decimal places). Furthermore, we mix each Chinese questions with two distractor images that are contextually adjacent to the ground images, approaching real-world multimodal reasoning scenarios. + +We evaluate 15 state-of-the-art (SOTA) MLLMs [18-21, 46-48, 56], utilizing Chain-of-Thought (CoT) [60] and Program-of-Thought (PoT) [13]. The experimental results on FinMMR reveals three key findings: + +- MLLMs Face Significant Challenges in Domain-Specific Multimodal Numerical Reasoning: All evaluated models underperform on FinMMR. On the Hard test set, the best-performing model, Claude 3.7 Sonnet with 64K extended thinking, achieves only $53.00\%$ accuracy, while OpenAI o1 achieves merely $48.40\%$ . Through error analysis, we identify that visual perception, knowledge reasoning, and numerical computation collectively pose challenges to MLLMs. Current MLLMs still struggle with complex multimodal reasoning tasks in specialized domains, compared to text-based reasoning. +- Better Synergy Between Visual Perception and Complex Reasoning Is Needed: Distracting images lead to a greater than $10\%$ drop in accuracy for Qwen2.5-VL-72B compared to ground images alone, indicating that irrelevant visual information severely impacts multimodal reasoning. By decoupling visual filtering and reasoning, Qwen2.5-VL-72B improved from $64.73\%$ to $71.56\%$ . Combining MLLMs with LRMs, by efficiently + +Parsing visual information into structured text and leveraging the LRMs' text-based reasoning capabilities, can also yield better performance. The combination of GPT4o and DeepSeek-R1 achieves $86.72\%$ accuracy on 1,160 tabular QA instances, outperforming Claude 3.7 Sonnet $(83.53\%)$ . + +- Refined Structured Domain Knowledge Enhances MLLMs' Complex Reasoning: Leading MLLMs lack sufficient experience in applying rich domain knowledge when solving complex reasoning tasks. By annotating structured financial functions and leveraging the model's ability to generate retrieval questions and make judgments, knowledge augmentation can significantly improve MLLMs' performance. MLLMs achieve absolute improvements ranging from 2.76 to 4.31 percentage points, allowing weaker models to approach SOTA performance, while SOTA models exhibit further gains. + +These findings highlight the bottlenecks of MLLMs in complex multimodal reasoning tasks in expert domains closer to real-world scenarios. They emphasize the need for continuous improvements in three key areas: more intricate visual perception, more specialized knowledge reasoning, and more accurate numerical computation. Alternatively, leveraging tools or model combinations can help achieve a balance between performance and cost, enabling MLLMs to perform expert-level reasoning tasks like human experts. + +# 2. Related Work + +# 2.1. LRMs and MLLMs + +By integrating train-time scaling and test-time scaling [31, 45], LRMs have demonstrated remarkable reasoning capabilities [62]. However, most LRMs are limited to handling text-based problems. The growing demand for solving real-world tasks has spurred the development of MLLMs [5, 19, 55] and benchmarks designed to evaluate the perception and reasoning abilities of MLLMs [10, 23, 27, 30, 33, 36, 38, 64-66]. For instance, MME-CoT [30] evaluates models' space-time understanding, while EMMA [27] focuses STEM subjects. Following this trend, domain-specific benchmarks which require deep domain expertise have emerged, such as MathVista [39] for mathematics and GMAI-MMBench [12] for medicine. Yet, financial reasoning remains an unexplored area in the current landscape of MLLM benchmarks. + +# 2.2. Financial Benchmarks + +The financial domain presents a distinct and more formidable set of challenges for model evaluation, which arise from its inherent complexity, information density, and dependence on expertise. The majority of existing text-only financial numerical reasoning benchmarks [14, 15, 32, 67-69] are constrained by limitations such as sub-optimal an + +
PropertyValue
# Total Questions4,300
# Easy/Medium/Hard1,300/1,500/1,500
# Validation/Test900/3,400
# Chinese/English2,150/2,150
# Operators (Easy/Medium/Hard)1.75/2.97/5.34
# Lines of Code (Easy/Medium/Hard)4.14/5.06/7.34
# Parentheses (Easy/Medium/Hard)0.88/3.11/4.25
# Difficulty (Easy/Medium/Hard)1.93/2.96/3.79
+ +Table 1. Key statistics of FinMMR (Avg values of three subsets). + +notation quality, narrow domain knowledge coverage, and overly simplistic reasoning tasks. Although FinanceReasoning [53] offers complex tasks with rich knowledge and high-quality annotations, its text-only modality limits multimodal reasoning evaluation. + +Recent multimodal financial benchmarks have sought to bridge this gap but still possess limitations. FAMMA [63] being sourced from textbooks and examinations does not mirror the real-world tasks. FinMME [40] uses a multiple-choice format, which may overestimate model reasoning due to guesswork. MME-Finance [24] is constrained by coarse annotations and an isolated assessment of domain knowledge, limiting holistic evaluation of real-world financial reasoning. + +# 3. FinMMR Benchmark + +# 3.1. Overview of FinMMR + +We introduce FinMMR, a bilingual (English and Chinese) multimodal benchmark for evaluating the reasoning capabilities of MLLMs in financial numerical reasoning tasks. FinMMR consists of 4,300 questions covering 14 financial subdomains (e.g., corporate finance, industry analysis, financial markets, asset management). The key statistics are summarized in Tab. 1, and the composition of sub-domains and images is illustrated in Fig. 1. As illustrated in Fig. 2, FinMMR introduces three key challenges: + +- Rich Images Challenge Visual Perception: FinMMR comprises 8.7K images from 14 categories (e.g., bar charts, line charts, ownership structure charts). MLLMs must identify relevant images among distractors and extract critical information from the correct images. +- Comprehensive Subdomains Challenge Knowledge Reasoning: MLLMs need to flexibly apply diverse domain-specific financial knowledge from 14 subdomains to solve multi-step reasoning tasks. +- Complex Formulas Challenge Numerical Computation: All questions require precise numerical answers, eliminating the potential bias from lucky guesses that + +could occur in multiple-choice formats. + +# 3.2. Data Curation Process + +We first curated a subset of questions from public text-based financial reasoning benchmarks and systematically transformed them into multimodal problems using a unified standard. Subsequently, we constructed a novel multimodal Chinese Research Report Question Answering (CR-RQA) benchmark from scratch, merging two data sources into FinMMR. Each question is accompanied by an executable Python solution, yields a numerical answer and demonstrates a clear reasoning chain. + +Update to Public Benchmarks We re-announce 124 and 163 financial questions from MMMU [65] and MMMU-Pro [66], respectively. Following rigorous verification, these questions were directly incorporated into our dataset. Furthermore, we extracted 77, 288, and 795 questions from FinanceMath [67], CodeTAT-QA [32], and CodeFinQA [32], respectively. From DocMath-Eval [68], we further obtained 703 questions from its four subsets. For each question, we rendered any tabular data as images while removing the corresponding table information from the text, ensuring that MLLMs cannot rely solely on textual content. + +Building a Novel Dataset from Scratch We collect 90 research reports, all of which are obtained through authorized access and cover diverse topics such as industry research, macroeconomic analysis, and strategy research. We use 360LayoutAnalysis [1] to extract informative images and discard those lacking explicit numerical data, reducing ambiguity. For each retained image, we prompt Qwen-VLMax [6] to formulate questions requiring multiple reasoning steps or complex calculations. Each question is accompanied by an executable Python solution and a definitive numerical answer. + +Furthermore, we introduce distractor images sourced from the same reports adjacent to ground images to challenge MLLMs in extracting relevant numerical information from structured, densely packed visuals. + +Data Quality Assurance This process ensured that every question was clearly written, featured a detailed reasoning solution, and included an accurate final answer. The annotators included 16 graduate students in finance and two experts holding CFA certifications. With the aid of LLMs, this meticulous verification process spanned three months, culminating in a dataset that meets high standards of clarity and correctness. + +Dataset Splitting and Release To classify the problems by difficulty, we employ a heuristic algorithm that takes into account the number of operators $(o)$ , code lines $(l)$ and parentheses pairs $(p)$ in the Python solution. Specifically, the difficulty of reasoning $rc$ of a problem is defined as: + +$$ +r c = \ln (\max (o, 1)) + \ln (\max (l + p, 1)) \tag {1} +$$ + +![](images/36bf93ff9009eed725d220edcdd51597f18cfd498d2cf7ea7c772bb200c376b8.jpg) +Figure 3. The comparison between finance-related benchmarks. These benchmarks vary in size, domain coverage, modalities, and question type, with some focusing on text-only data while others include images. Each axis has scale labels with varying ranges to measure the number of questions from each benchmark across different subdomains. In the Modalities, T means text input, I means Images input, P.I. means pure images input. In the Question Type, MC means multi-choice answer, NUM means numerical answer. + +
BenchmarkSize (Fin)Domain CoverageModalitiesQuestion Type
MMMU [65]11,550 (1,603)10T+IMC
MMMU-Pro [66]1,730 (286)10T+I, P.I.MC
FinanceMath [67]1,200 (1,200)10TNUM
CodeTAT-QA [32]3,144 (3,144)6TNUM
CodeFinQA [32]5,463 (5,463)13TNUM
DocMath-Eval [68]4,000 (4,000)12TNUM
FAMMA [63]1,758 (1,758)8T+IMC, NUM
CRRQA (ours)2,150 (2,150)13T+I, P.I.NUM
FinMMR (ours)4,300 (4,300)14T+I, P.I.NUM
+ +FinMMR is classified as Easy (1,300 examples), Medium (1,500 examples) and Hard (1,500 examples) based on this formula, with each level randomly split into test and validation sets. All questions are publicly available, while only the 300 validation answers per level are released, while test answers remain private to prevent data leakage [22, 50, 52]. We maintain an online evaluation platform that enables researchers to assess their models. + +# 3.3. Comparisons with Existing Benchmarks + +As illustrated in Fig. 3, previous work has studied multidiscipline multimodal reasoning [65, 66], general mathematical reasoning [39] or text-based financial QA [32, 67, 68]. FinMMR focus on multimodal financial numerical reasoning, curating 4,300 questions requiring a deep understanding of domain-specific images (e.g., earnings reports, candlestick charts). To mimic real-world scenario, we deliberately incorporate 3,938 distractors into 2,150 questions to rigorously evaluate MLLMs' visual perception capability. Compared to existing financial benchmarks, they suffer from narrow domain coverage [14, 69]. FinMMR encompasses 14 financial subdomains and 14 image categories, comprehensively evaluating MLLMs' domain-specific reasoning capabilities. + +# 4. Evaluation System + +To facilitate the evaluation of complex reasoning on FinMMR, we established a dedicated evaluation system. All MLLMs evaluated were accessed through official APIs. + +# 4.1. Multimodal Large Language Models + +We systematically evaluate the multimodal reasoning capabilities of 15 recent MLLMs under the zero-shot setting. The evaluated MLLMs are: OpenAI o1 [46], GPT-4o [44], Claude 3.7 Sonnet (including thinking mode) [5], + +Gemini 2.0 Flash Thinking [19], Gemini 2.0 Pro [20], Gemini 2.0 Flash [18], InternVL2.5-78B [16], Grok 2 Vision [61], Pixtral Large [2], Qwen2.5-VL-72B [7], QVQ72B-Preview [55], Qwen-Omni-Turbo [57], Llama 4 Maverick [4], Gemma 3 27B [17], and Mistral Small 3.1 [3]. + +# 4.2. Prompting Methods + +Our evaluation adopts CoT [60], PoT [13] and IO (without any external prompts) prompting methods. Due to budget constraints, we report OpenAI o1 performance on the Hard subset only. Detailed prompts can be found in the Appendix. + +# 4.3. Answer Extraction and Evaluation + +Following Zhao et al. [67], we extract answers based on the prompting methods. For CoT and IO outputs, we employ GPT-4o-mini to extract numerical answers. For PoT, we run the generated program for numerical results. Finally, we conduct a strict accuracy evaluation, comparing numerical results with ground truth and deeming the results accurate within a stringent error tolerance of $0.2\%$ . + +# 5. Experiments + +We answer the following research questions (RQs): RQ1: Are MLLMs multimodal reasoners with extended thinking and PoT prompts? RQ2: What are the primary challenges facing MLLMs? RQ3: How can the visual perception difficulties of MLLMs be mitigated? RQ4: How can the knowledge reasoning capabilities of MLLMs be enhanced? RQ5: How can the numerical computation abilities of MLLMs be compensated for? + +# 5.1. Main Results (RQ1) + +The performance of the MLLMs evaluated using different prompting methods on FinMMR is shown in Tab. 2. + +
ModelSizeExtended thinkingHardMediumEasyAvg.Token (M)
IOCoTPoTCoTPoTCoTPoTCoTPoTCoTPoT
Proprietary MLLMs
Claude 3.7 Sonnet✓ (64K)53.0051.0051.4062.5062.1778.5078.5064.0064.028.5111.25
Claude 3.7 SonnetX49.8050.8048.5062.2558.8377.0076.9263.3561.420.990.89
OpenAI o1✓48.0048.4044.70------2.522.12
GPT-4oX-45.4047.8063.3359.9278.0076.0062.2461.240.850.41
Gemini 2.0 ProX-46.5047.3060.5857.9275.5075.6760.8660.300.850.45
Gemini 2.0 Flash Thinking✓-46.0046.0060.7556.5877.1774.1761.3158.921.300.48
Gemini 2.0 FlashX-44.4045.9057.8353.4274.9273.7559.0557.690.790.43
Grok 2 VisionX-27.8025.5041.5035.8373.0872.8347.4644.721.130.60
Qwen-Omni-TurboX-17.5027.3035.8348.0057.5061.6736.9445.660.900.42
Open-source MLLMs
Llama 4 Maverick17BX-48.7047.8063.2559.1777.8377.8363.2661.600.880.47
Qwen2.5-VL-72B72BX-43.3046.2063.4264.1777.4275.8361.3862.071.050.44
InternVL2.5-78B78BX-37.4044.0060.5061.1770.9270.5856.2758.58--
QVQ-72B-Preview72B✓43.3040.306.2055.679.6775.4212.4257.139.435.435.70
Pixtral Large124BX-19.7025.0039.8339.7570.0070.1743.1844.971.150.75
Gemma 3 27B27BX-23.4022.3045.1736.4269.0861.5845.8840.100.970.47
Mistral Small 3.124BX-19.7015.2038.4229.7567.6749.4241.9331.461.150.60
+ +Table 2. Results of different models using IO, CoT and PoT prompting methods on the test set of FinMMR. We use the best Accuracy on the Hard subset as the ranking indicator of model performance. The results underscore the superior performance of reasoning-enhanced MLLM (i.e., Claude 3.7 Sonnet with 64K extended thinking) with PoT in complex multimodal numerical reasoning task. + +Challenges of MLLMs in Domain-Specific Complex Numerical Reasoning As the difficulty increases, the average accuracy shows a continuous and significant decline. In CoT setting, the average accuracy rates on the Easy, Medium, and Hard subsets are $73.33\%$ , $54.05\%$ , and $38.14\%$ , respectively. The currently best-performing model (i.e., Claude 3.7 Sonnet with 64K extended thinking) consistently demonstrates superior performance across all difficulty subsets using CoT prompting method. However, its accuracy on the Hard subset remains below the $60\%$ passing threshold in both prompting methods. On the overall test set, Claude 3.7 Sonnet achieves only $64.02\%$ accuracy. These results highlight the challenging nature and rigorous standards of FinMMR, effectively assessing the limits of MLLMs' reasoning capabilities and the disparities among models compared to previous benchmarks. + +Does extended thinking help? Reasoning-enhanced models show consistent improvements, compared with non-reasoning-enhanced MLLMs. Claude 3.7 Sonnet with 64K extended thinking achieves 2.20 percentage points higher accuracy than the model without extended thinking (53.00% vs. 50.80%) on the Hard subset. This enhancement comes at the cost of using nearly 12 times more tokens for intricate thinking (4.06M vs. 0.34M). This trend also persists in the Gemini 2.0 Flash series. + +We observe that QVQ-72B-Preview lose basic code generation capabilities due to the reinforcement learning of text-based long thought, which is likely attributed to biases in training strategies and training data. On the Hard subset, this model achieves a code execution success rate of only $10.90\%$ , resulting in an accuracy of merely $6.20\%$ in + +PoT setting, significantly lower than the $40.30\%$ accuracy achieved with CoT. This finding highlights the importance of maintaining foundational capabilities, such as programming, while enhancing the reasoning abilities of MLLMs, to avoid rendering them ineffective in other basic tasks. + +Does PoT help? Experimental results strongly validate the superiority of PoT over CoT in numerical reasoning tasks, especially on the Hard subset. PoT (not including QVQ-72B-Preview) achieves a mean accuracy of $37.64\%$ versus $36.20\%$ for CoT, representing an improvement of 1.44 percentage points. Furthermore, PoT encourages MLLMs to leverage structured code generation to reduce token consumption during reasoning. Under similar or lower token usage, PoT achieves similar or greater accuracy. For instance, GPT-4o achieves a 2.40 percentage points improvement in accuracy over CoT while consuming significantly fewer tokens in PoT setting. Similarly, Qwen2.5-VL-72B demonstrates the most pronounced efficiency gains: PoT improves accuracy to $64.17\%$ from $63.42\%$ while reducing token consumption by $58.88\%$ (153K vs. 373K) on the Medium subset. When addressing complex numerical reasoning problems, PoT avoids precise numerical calculations by utilizing external tools (i.e., Python interpreter) and reduces the need for repetitive text-based reasoning, which is beneficial for most MLLMs. + +However, we also observe that for certain reasoning-enhanced models (i.e., Claude 3.7 Sonnet with 64K extended thinking and OpenAI o1), due to the enforced requirement for long thought, they still engage in extensive text-based reasoning before generating code even in PoT setting, resulting in exceptionally high token consumption + +
SubsetTestDegradationValidationDegradation
Ground Images (%)Distractor Images (%)Ground Images (%)Distractor Images (%)
Hard57.1847.23↓9.9556.7445.58↓11.16
Medium73.0161.36↓11.6577.7864.73↓12.35
Easy61.5953.64↓7.9560.6151.52↓9.09
The improvement achieved by the filtering-reasoning pipeline on the Medium validation set: 64.73 → 71.56 ↑6.83
+ +Table 3. Degradation of Qwen2.5-VL-72B on all subsets due to distractor images and improvement achieved by the filtering-reasoning pipeline on the Medium validation set in PoT setting. + +on the Hard subset (4.48M and 2.12M), which is more than 10 times that of other MLLMs. To further investigate this, we added an IO baseline without any external prompts for reasoning-enhanced models on the Hard subset. The IO group achieved the highest accuracy, which we attribute to the comprehensive built-in system prompts embedded in the tested proprietary models. This highlights the need for future research to balance reasoning performance with the control of inefficient and redundant token generation, aiming to achieve a good trade-off between performance and cost, as well as to investigate whether PoT prompting can yield significant performance gains on open-source reasoning-enhanced models. + +# 5.2. Error Analysis (RQ2) + +To better analyze the capabilities and limitations of MLLMs on FinMMR, we conduct a detailed error analysis for the Claude 3.7 Sonnet with 64K extended thinking in PoT setting. Error analysis is based on 100 sampled failure cases, which we categorize into three main error types, some of which involve compound errors. More details of error cases are provided in the Appendix. + +- Visual Perception Error (30/100): The model incorrectly perceives, identifies, or interprets visual information from images, or mistakenly recognizes incorrect data, subsequently causing errors in calculations, broken reasoning chains, or incorrect conclusions. +- Knowledge Reasoning Error (38/100): Due to insufficient domain-specific knowledge, the model exhibits logical confusion or conceptual misunderstandings during reasoning, leading to incorrect answers. +- Numerical Computation Error (32/100): In problems involving mathematical operations and numerical reasoning, the model produces significant deviations from the correct answers due to errors in the calculation steps, precision control, or numerical hallucination. + +# 5.3. Visual Filtering for Reasoning (RQ3) + +As shown in Tab. 3, when processing multi-image inputs containing distractor images, Qwen2.5-VL-72B demonstrates significantly lower accuracy across all difficulty levels compared to ground images scenarios. This finding + +aligns with conclusions from previous work [37, 51], indicating that irrelevant visual information substantially interferes with MLLMs' reasoning capabilities. In particular, the Medium validation set exhibits the most pronounced performance drop (77.78% ground images vs. 64.73% distractor images), attributed to two key characteristics: (1) moderate complexity making visual perception quality the performance bottleneck; (2) semantic relevance between distractors and questions increasing visual filtering difficulty. To address distractor image interference, we propose a two-stage multimodal reasoning pipeline: + +- Visual Filtering: We first instruct MLLM to analyze the set of images $\mathcal{I}$ and the question $q$ , assessing the relevance of the image (relevant/irrelevant). Irrelevant images are excluded from subsequent reasoning. +- Enhanced Reasoning: Then, the filtered subset $\mathcal{I}'$ and the question $q$ are input into the MLLM for the final reasoning. The system automatically reverts to the original inputs $\mathcal{I}$ if all images are mistakenly filtered. + +Does the pipeline help? As illustrated in Tab. 3, we evaluate Qwen2.5-VL-72B on the 207 English questions with distractor images of the Medium validation set. Our method achieves an overall accuracy of $71.56\%$ , representing a 6.83 percentage points improvement over direct reasoning. This result is only 6.22 percentage points away from the ideal accuracy of the ground images scenarios $(77.78\%)$ . Detailed analysis reveals $73.43\%$ (152/207) ground image recognition accuracy during filtering. When correctly identified, the accuracy of these problems increases to $81.58\%$ (124/152). This finding underscores the necessity to enhance the ability of MLLMs to filter out irrelevant image information, thereby strengthening their robustness in reasoning within more complex real-world scenarios. + +# 5.4. Knowledge Augmentation (RQ4) + +To enhance the understanding and application capabilities of financial knowledge of MLLMs, we explore a method of refined knowledge augmentation to improve the performance of MLLM in domain-specific reasoning tasks. + +- Function Library Construction: We annotate a comprehensive financial function library containing 3,133 Python functions from financial encyclopedia. Each func + +
SettingPoTRAG with PoT
Gemini 2.0 Flash Thinking78.7183.02 (+4.31)
GPT-4o80.6083.62 (+3.02)
Claude 3.7 Sonnet81.2185.43 (+4.22)
Claude 3.7 Sonnet (64K)83.5386.29 (+2.76)
+ +Table 4. Improvements of different MLLMs with knowledge augmentation on the 1,160 instances of FinMMR in PoT setting. + +tion includes precise functional descriptions, parameter explanations, and step-by-step implementation code. + +- MLLM-Instructed Knowledge Retrieval: In financial problems with hybrid contexts, using short questions or full contexts for retrieval often fails to retrieve directly relevant knowledge [9, 49]. We observed that powerful MLLMs (e.g., GPT-40) can effectively summarize rich semantic information from contexts. Therefore, we first prompt the MLLM to generate precise retrieval queries based on the context [34, 58]. Then we use Contriever [28] to retrieve relevant financial Python functions, based on the semantic similarity between the refined queries and functional descriptions. +- MLLM as Retrieval Judge: Recent studies have shown that models are capable of judging the relevance of candidates retrieved for the question [26]. In this setting, we first retrieved the Top-30 financial functions and then prompted the MLLM to select the Top-3 functions most useful to answer the question, if any. + +Does knowledge augmentation help? As shown in Tab. 4, all evaluated MLLMs enhanced with financial function knowledge achieved significant improvements, ranging from 2.76 to 4.31 percentage points. Leveraging the improved retrieval efficiency enabled by MLLM-Instructed Knowledge Retrieval and MLLM as Retrieval Judge, the knowledge augmentation approach achieves greater performance, boosting the accuracy of Claude 3.7 Sonnet with 64K thinking to $86.29\%$ . Notably, Gemini 2.0 Flash Thinking, which has relatively weaker reasoning capabilities, also improved from $78.71\%$ to $83.02\%$ , approaching the performance of Claude 3.7 Sonnet $(83.53\%)$ without knowledge augmentation. The results further illustrate that refined domain-specific reasoning knowledge can significantly enhance the performance of MLLMs in complex reasoning tasks within expert domains. + +# 5.5. Visual Parser with Reasoner (RQ5) + +In complex multimodal numerical reasoning tasks, single models often struggle to simultaneously achieve visual perception, knowledge reasoning, and numerical computation. To investigate the potential of model collaboration, we instruct the MLLM to act as the Visual Parser, responsible for carefully converting images into structured textual data. + +![](images/7d6557cfa9525f56834045e280e4a6b9d3f2d22c08400430bd46e0c9f475df53.jpg) +Figure 4. Results of model combinations and individual models. + +Then, an LRM acts as the Reasoner, performing multi-step numerical reasoning based on the textual context. + +Specifically, we filter 1,160 tabular QA instances from FinMMR and utilize GPT-4o as the Visual Parser, instructing it to separate table headers or cells with vertical bars (|) and rows with newlines. For the Reasoner component, in addition to GPT-4o, we evaluate several LRMs (i.e., Claude 3.7 Sonnet, Gemini 2.0 Flash Thinking, DeepSeek-R1, and OpenAI o1). + +Does model collaboration help? As shown in Fig. 4, the structured data from GPT-4o's visual parsing significantly enhances downstream reasoning. The individual model (i.e., GPT-4o with PoT) achieves an accuracy of $80.60\%$ , while the combination of models improves the accuracy to $86.72\%$ (i.e., DeepSeek-R1 serving as the Reasoner with PoT). Performance variance emerges across reasoning-enhanced models using identical visual inputs. Claude 3.7 Sonnet reaches $85.52\%$ , outperforming Gemini 2.0 Flash Thinking $(82.24\%)$ , confirming the decisive impact of text-based reasoning capabilities. This evidences that model collaboration effectively compensates for individual model limitations through complementary strengths. + +# 6. Conclusion + +We introduce FinMMR, a multimodal, comprehensive, and challenging benchmark for evaluating the financial numerical reasoning capabilities of MLLMs. FinMMR challenges MLLMs' intricate visual perception, specialized knowledge reasoning, and accurate numerical computation through its rich images, comprehensive subdomains, and complex formulas embedded in each multimodal financial question. The evaluation results reveal that 15 proprietary and open-source MLLMs still struggle with complex multimodal reasoning tasks in specialized domains. FinMMR highlights the bottlenecks of MLLMs and the need for continuous improvements, including reasoning-enhanced training, tool use, refined structured knowledge augmentation and model combinations, allowing models to perform expert-level reasoning tasks closer to the real-world scenarios. + +# Acknowledgements + +This work is supported by the National Natural Science Foundation of China (Grant Nos. 62176026, 62473271), the Beijing Natural Science Foundation (Grant No. QY24214), the BUPT Innovation and Entrepreneurship Support Program (Grant Nos. 2025-YC-A033, 2025-YC-A042), and data support from Hithink RoyalFlush Information Network Co., Ltd. This work is also supported by the Engineering Research Center of Information Networks, Ministry of Education, China. We would also like to thank the anonymous reviewers and area chairs for constructive discussions and feedback. + +# References + +[1] 360AILAB-NLP. 360layoutanalysis. https://github.com/360AILAB-NLP/360LayoutAnalysis, 2024. Accessed: 2024-06-15. 4 +[2] Mistral AI. Pixtral large. https://mistral.ai/news/pixtral-large, 2024. Accessed: 2024-11-18. 5 +[3] Mistral AI. Mistral small 3.1. https://mistral.ai/news/mistral-small-3-1, 2025. Accessed: 2025-03-17. 5 +[4] AI@Meta. The llama 4 herd: The beginning of a new era of natively multimodal ai innovation. https://ai.meta.com/blog/llama-4-multimodal-intelligence, 2025. Accessed: 2025-04-05. 5 +[5] Anthropic. Claude 3.7 sonnet and claude code. https://www.anthropic.com/news/claude-3-7-sonnet, 2025. Accessed: 2025-02-25. 1, 3, 5 +[6] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint, 2023. 4 +[7] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zhenen Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Yang, Haiyang Xu, and Junyang Lin. Qwen2.5-vl technical report. arXiv preprint, 2025. 5 +[8] Sébastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, Harsha Nori, Hamid Palangi, Marco Tulio Ribeiro, and Yi Zhang. Sparks of artificial general intelligence: Early experiments with gpt-4. arXiv preprint, 2023. 2 +[9] Liang Chen, Yang Deng, Yatao Bian, Zeyu Qin, Bingzhe Wu, Tat-Seng Chua, and Kam-Fai Wong. Beyond factuality: A comprehensive evaluation of large language models as knowledge generators. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 6325-6341, Singapore, 2023. Association for Computational Linguistics. 8 + +[10] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, and Feng Zhao. Are we on the right way for evaluating large vision-language models? In Advances in Neural Information Processing Systems, pages 27056-27087. Curran Associates, Inc., 2024. 3 +[11] Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott Gray, Nick Ryder, Mikhail Pavlov, Alethea Power, Lukasz Kaiser, Mohammad Bavarian, Clemens Winter, Philippe Tillet, Felipe Petroski Such, Dave Cummings, Matthias Plappert, Fotios Chantzis, Elizabeth Barnes, Ariel Herbert-Voss, William Hebgen Guss, Alex Nichol, Alex Paino, Nikolas Tezak, Jie Tang, Igor Babuschkin, Suchir Balaji, Shantanu Jain, William Saunders, Christopher Hesse, Andrew N. Carr, Jan Leike, Josh Achiam, Vedant Misra, Evan Morikawa, Alec Radford, Matthew Knight, Miles Brundage, Mira Murati, Katie Mayer, Peter Welinder, Bob McGrew, Dario Amodei, Sam McCandlish, Ilya Sutskever, and Wojciech Zaremba. Evaluating large language models trained on code. arXiv preprint, 2021. 1 +[12] Pengcheng Chen, Jin Ye, Guoan Wang, Yanjun Li, Zhongying Deng, Wei Li, Tianbin Li, Haodong Duan, Ziyan Huang, Yanzhou Su, Benyou Wang, Shaoting Zhang, Bin Fu, Jianfei Cai, Bohan Zhuang, Eric J Seibel, Yu Qiao, and Junjun He. Gmai-mmbench: A comprehensive multimodal evaluation benchmark towards general medical ai. In Advances in Neural Information Processing Systems, pages 94327-94427. Curran Associates, Inc., 2024. 3 +[13] Wenhu Chen, Xueguang Ma, Xinyi Wang, and William W. Cohen. Program of thoughts prompting: Disentangling computation from reasoning for numerical reasoning tasks. Transactions on Machine Learning Research, 2023. 3, 5 +[14] Zhiyu Chen, Wenhu Chen, Charese Smiley, Sameena Shah, Iana Borova, Dylan Langdon, Reema Moussa, Matt Beane, Ting-Hao Huang, Bryan Routledge, and William Yang Wang. FinQA: A dataset of numerical reasoning over financial data. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3697-3711, Online and Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 2, 3, 5 +[15] Zhiyu Chen, Shiyang Li, Charese Smiley, Zhiqiang Ma, Sameena Shah, and William Yang Wang. ConvFinQA: Exploring the chain of numerical reasoning in conversational finance question answering. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 6279-6292, Abu Dhabi, United Arab Emirates, 2022. Association for Computational Linguistics. 2, 3 +[16] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, Lixin Gu, Xuehui Wang, Qingyun Li, Yimin Ren, Zixuan Chen, Jiapeng Luo, Jiahao Wang, Tan Jiang, Bo Wang, Conghui He, Botian Shi, Xingcheng Zhang, Han Lv, Yi Wang, Wenqi Shao, Pei Chu, Zhongying Tu, Tong He, Zhiyong Wu, Huipeng Deng, Jiaye Ge, Kai Chen, Kaipeng + +Zhang, Limin Wang, Min Dou, Lewei Lu, Xizhou Zhu, Tong Lu, Dahua Lin, Yu Qiao, Jifeng Dai, and Wenhai Wang. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint, 2025. 5 +[17] Google DeepMind. Introducing gemma 3: The most capable model you can run on a singlegpu or tpu. https://blog.google/technology/developers/gemma-3, 2025. Accessed: 2025-03-12. 5 +[18] Google DeepMind. Gemini 2.0: Flash, flash-lite and pro. https://developers.googleblog.com/en/gemini-2-family-expands, 2025. Accessed: 2025-02-05. 3, 5 +[19] Google DeepMind. Gemini 2.0 flash thinking. https://ai.google.dev/gemini-api/docs/thinking, 2025. Accessed: 2025-02-05. 1, 3, 5 +[20] Google DeepMind. Gemini 2.0. https://cloud.google.com/vertex-ai/generative-ai/docs/gemini-v2#2.0-pro, 2025. Accessed: 2025-02-05. 5 +[21] DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, Aixin Liu, Bing Xue, Bingxuan Wang, Bochao Wu, Bei Feng, Chengda Lu, Chenggang Zhang, Chengqi Deng, Chenyu Zhang, Chong Ruan, Damai Dai, Deli Chen, Dongjie Ji, Erhang Li, Fangyun Lin, Fucong Dai, Fuli Luo, Guangbo Hao, Guanting Chen, Guowei Li, H. Zhang, Han Bao, Hanwei Xu, Haocheng Wang, Honghui Ding, Huajian Xin, Huazuo Gao, Hui Qu, Hui Li, Jianzhong Guo, Jiashi Li, Jiawei Wang, Jingchang Chen, Jingyang Yuan, Junjie Qiu, Junlong Li, J. L. Cai, Jiaqi Ni, Jian Liang, Jin Chen, Kai Dong, Kai Hu, Kaige Gao, Kang Guan, Kexin Huang, Kuai Yu, Lean Wang, Lecong Zhang, Liang Zhao, Litong Wang, Liyue Zhang, Lei Xu, Leyi Xia, Mingchuan Zhang, Minghua Zhang, Minghui Tang, Meng Li, Miaojun Wang, Mingming Li, Ning Tian, Panpan Huang, Peng Zhang, Qiancheng Wang, Qinyu Chen, Qiushi Du, Ruiqi Ge, Ruisong Zhang, Ruizhe Pan, Runji Wang, R. J. Chen, R. L. Jin, Ruyi Chen, Shanghao Lu, Shangyan Zhou, Shanhuang Chen, Shengfeng Ye, Shiyu Wang, Shuiping Yu, Shunfeng Zhou, Shuting Pan S. S. Li Shuang Zhou Shaoqing Wu Shengfeng Ye Tao Yun Tian Pei Tianyu Sun T. Wang Wangding Zeng Wanjia Zhao Wen Liu Wenfeng Liang Wenjun Gao Wenqin Yu Wentao Zhang W. L. Xiao Wei An Xiaodong Liu Xiaohan Wang Xiaokang Chen Xiaotao Nie Xin Cheng Xin Liu Xin Xie Xingchao Liu Xinyu Yang Xinyuan Li Xuecheng Su Xuheng Lin X.Q.Li Xiangyue Jin Xiaojin Shen Xiaosha Chen Xiaowen Sun Xiaoxiang Wang Xinnan Song Xinyi Zhou Xianzu Wang Xinxia Shan Y.K. Li Y.Q. WangY.X.Wei,Yang Zhang Yanhong Xu Yao Li Yao Zhao Yaofeng Sun Yaohui Wang Yi Yu Yichao Zhang Yifan Shi Yiliang Xiong Ying He Yishi Piao Yisong Wang Yixuan Tan Yiyang Ma Yiyuan Liu Yongqiang Guo Yuan OuYuduan Wang Yue Gong Yuheng Zou Yujia HeYunfan Xiong Yuxiang Luo Yuxiang You Yuxuan Liu Yuyang ZhouY.X.ZhuYanhong Xu Yanping Huang Yaohui Li Yi ZhengYuchen ZhuYunxian Ma Ying Tang Yukun + +Zha, Yuting Yan, Z. Z. Ren, Zehui Ren, Zhangli Sha, Zhe Fu, Zhean Xu, Zhenda Xie, Zhengyan Zhang, Zhenen Hao, Zhicheng Ma, Zhigang Yan, Zhiyu Wu, Zihui Gu, Zijia Zhu, Zijun Liu, Zilin Li, Ziwei Xie, Ziyang Song, Zizheng Pan, Zhen Huang, Zhipeng Xu, Zhongyu Zhang, and Zhen Zhang. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint, 2025. 1, 3 +[22] Chunyuan Deng, Yilun Zhao, Xiangru Tang, Mark Gerstein, and Arman Cohan. Investigating data contamination in modern benchmarks for large language models. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 8706-8719, Mexico City, Mexico, 2024. Association for Computational Linguistics. 5 +[23] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, Yunsheng Wu, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint, 2024. 3 +[24] Ziliang Gan, Yu Lu, Dong Zhang, Haohan Li, Che Liu, Jian Liu, Ji Liu, Haipang Wu, Chaoyou Fu, Zenglin Xu, Rongjunchen Zhang, and Yong Dai. Mme-finance: A multimodal finance benchmark for expert-level understanding and reasoning. arXiv preprint, 2024. 4 +[25] Yingqiang Ge, Wenyue Hua, Kai Mei, jianchao ji, Juntao Tan, Shuyuan Xu, Zelong Li, and Yongfeng Zhang. Opennagi: When llm meets domain experts. In Advances in Neural Information Processing Systems, pages 5539-5568. Curran Associates, Inc., 2023. 2 +[26] Jian Guan, Wei Wu, Zujie Wen, Peng Xu, Hongning Wang, and Minlie Huang. Amor: A recipe for building adaptable modular knowledge agents through process feedback. In Advances in Neural Information Processing Systems, pages 126118-126148. Curran Associates, Inc., 2024. 8 +[27] Yunzhuo Hao, Jiawei Gu, Huichen Will Wang, Linjie Li, Zhengyuan Yang, Lijuan Wang, and Yu Cheng. Can MLLMs reason in multimodality? EMMA: An enhanced multimodal reasoning benchmark. In *Forty-second International Conference on Machine Learning*, 2025. 3 +[28] Gautier Izacard, Mathilde Caron, Lucas Hosseini, Sebastian Riedel, Piotr Bojanowski, Armand Joulin, and Edouard Grave. Unsupervised dense information retrieval with contrastive learning. Transactions on Machine Learning Research, 2022. 8 +[29] Naman Jain, King Han, Alex Gu, Wen-Ding Li, Fanjia Yan, Tianjun Zhang, Sida Wang, Armando Solar-Lezama, Koushik Sen, and Ion Stoica. Livecodebench: Holistic and contamination free evaluation of large language models for code. In The Thirteenth International Conference on Learning Representations, 2025. 1 +[30] Dongzhi Jiang, Renrui Zhang, Ziyu Guo, Yanwei Li, Yu Qi, Xinyan Chen, Liuhui Wang, Jianhan Jin, Claire Guo, Shen Yan, Bo Zhang, Chaoyou Fu, Peng Gao, and Hongsheng Li. MME-cot: Benchmarking chain-of-thought in large multimodal models for reasoning quality, robustness, and efficiency. In Forty-second International Conference on Machine Learning, 2025. 3 + +[31] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint, 2020. 1, 3 +[32] Michael Krumdick, Rik Koncel-Kedziorski, Viet Dac Lai, Varshini Reddy, Charles Lovering, and Chris Tanner. BizBench: A quantitative reasoning benchmark for business and finance. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8309-8332, Bangkok, Thailand, 2024. Association for Computational Linguistics. 2, 3, 4, 5 +[33] Bohao Li, Yuying Ge, Yixiao Ge, Guangzhi Wang, Rui Wang, Ruimao Zhang, and Ying Shan. Seed-bench: Benchmarking multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13299-13308, 2024. 3 +[34] Xiaoxi Li, Guanting Dong, Jiajie Jin, Yuyao Zhang, Yujia Zhou, Yutao Zhu, Peitian Zhang, and Zhicheng Dou. Searcho1: Agentic search-enhanced large reasoning models. arXiv preprint, 2025. 8 +[35] Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let's verify step by step. In The Twelfth International Conference on Learning Representations, 2024. 1 +[36] Fuxiao Liu, Xiaoyang Wang, Wenlin Yao, Jianshu Chen, Kaiqiang Song, Sangwoo Cho, Yaser Yacoob, and Dong Yu. MMC: Advancing multimodal chart understanding with large-scale instruction tuning. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 1287-1310, Mexico City, Mexico, 2024. Association for Computational Linguistics. 3 +[37] Haowei Liu, Xi Zhang, Haiyang Xu, Yaya Shi, Chaoya Jiang, Ming Yan, Ji Zhang, Fei Huang, Chunfeng Yuan, Bing Li, and Weiming Hu. MIBench: Evaluating multimodal large language models over multiple images. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 22417-22428, Miami, Florida, USA, 2024. Association for Computational Linguistics. 7 +[38] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin. Mmbench: Is your multi-modal model an all-around player? In Computer Vision - ECCV 2024, pages 216-233, Cham, 2025. Springer Nature Switzerland. 3 +[39] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating math reasoning in visual contexts with gpt-4v, bard, and other large multimodal models. In The Twelfth International Conference on Learning Representations, 2024. 1, 3, 5 +[40] Junyu Luo, Zhizhuo Kou, Liming Yang, Xiao Luo, Jinsheng Huang, Zhiping Xiao, Jingshu Peng, Chengzhong Liu, Ji-aming Ji, Xuanzhe Liu, Sirui Han, Ming Zhang, and Yike Guo. FinMME: Benchmark dataset for financial multi-modal reasoning evaluation. In Proceedings of the 63rd Annual + +Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 29465-29489, Vienna, Austria, 2025. Association for Computational Linguistics. 4 +[41] Yujun Mao, Yoon Kim, and Yilun Zhou. CHAMP: A competition-level dataset for fine-grained analyses of LLMs' mathematical reasoning capabilities. In Findings of the Association for Computational Linguistics: ACL 2024, pages 13256-13274, Bangkok, Thailand, 2024. Association for Computational Linguistics. 1 +[42] Grégoire Mialon, Clémentine Fourrier, Thomas Wolf, Yann LeCun, and Thomas Scialom. GAIA: a benchmark for general AI assistants. In The Twelfth International Conference on Learning Representations, 2024. 2 +[43] Meredith Ringel Morris, Jascha Sohl-Dickstein, Noah Fiedel, Tris Warkentin, Allan Dafoe, Aleksandra Faust, Clement Farabet, and Shane Legg. Position: Levels of AGI for operationalizing progress on the path to AGI. In *Forty-first International Conference on Machine Learning*, 2024. 2 +[44] OpenAI. Hello gpt-4o. https://openai.com/index/hello-gpt-4o, 2024. Accessed: 2024-05-13. 5 +[45] OpenAI. Learning to reason with llms. https://openai.com/index/learning-to-reason-with-llms, 2024. Accessed: 2024-09-12. 1, 3 +[46] OpenAI. Openai o1 system card. https://openai.com/index/openai-o1-system-card, 2024. Accessed: 2024-12-05. 1, 3, 5 +[47] OpenAI. Openai o1-mini. https://openai.com/index/openai-ol-mini-advancing-cost-efficient-reasoning, 2024. Accessed: 2024-09-12. 1 +[48] OpenAI. Openai o3-mini system card. https://openai.com/index/o3-mini-system-card, 2025. Accessed: 2025-01-31. 1, 3 +[49] Baolin Peng, Michel Galley, Pengcheng He, Hao Cheng, Yujia Xie, Yu Hu, Qiuyuan Huang, Lars Liden, Zhou Yu, Weizhu Chen, and Jianfeng Gao. Check your facts and try again: Improving large language models with external knowledge and automated feedback. arXiv preprint, 2023. 8 +[50] Oscar Sainz, Jon Campos, Iker García-Ferrero, Julien Etxaniz, Oier Lopez de Lacalle, and Eneko Agirre. NLP evaluation in trouble: On the need to measure LLM data contamination for each benchmark. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 10776-10787, Singapore, 2023. Association for Computational Linguistics. 5 +[51] Aditya Sharma, Michael Saxon, and William Yang Wang. Losing visual needles in image haystacks: Vision language models are easily distracted in short and long contexts. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 5429-5451, Miami, Florida, USA, 2024. Association for Computational Linguistics. 7 +[52] Weijia Shi, Anirudh Ajith, Mengzhou Xia, Yangsibo Huang, Daogao Liu, Terra Blevins, Danqi Chen, and Luke Zettlemoyer. Detecting pretraining data from large language models. In The Twelfth International Conference on Learning Representations, 2024. 5 + +[53] Zichen Tang, Haihong E, Ziyan Ma, Haoyang He, Jiacheng Liu, Zhongjun Yang, Zihua Rong, Rongjin Li, Kun Ji, Qing Huang, Xinyang Hu, Yang Liu, and Qianhe Zheng. FinanceReasoning: Benchmarking financial numerical reasoning more credible, comprehensive and challenging. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15721-15749, Vienna, Austria, 2025. Association for Computational Linguistics. 4 +[54] Kimi Team, Angang Du, Bofei Gao, Bowei Xing, Changjiu Jiang, Cheng Chen, Cheng Li, Chenjun Xiao, Chenzhuang Du, Chonghua Liao, Chuning Tang, Congcong Wang, Dehao Zhang, Enming Yuan, Enzhe Lu, Fengxiang Tang, Flood Sung, Guangda Wei, Guokun Lai, Haiqing Guo, Han Zhu, Hao Ding, Hao Hu, Hao Yang, Hao Zhang, Haotian Yao, Haotian Zhao, Haoyu Lu, Haoze Li, Haozhen Yu, Hongcheng Gao, Huabin Zheng, Huan Yuan, Jia Chen, Jianhang Guo, Jianlin Su, Jianzhou Wang, Jie Zhao, Jin Zhang, Jingyuan Liu, Junjie Yan, Junyan Wu, Lidong Shi, Ling Ye, Longhui Yu, Mengnan Dong, Neo Zhang, Ningchen Ma, Qiwei Pan, Qucheng Gong, Shaowei Liu, Shengling Ma, Shupeng Wei, Sihan Cao, Siying Huang, Tao Jiang, Weihao Gao, Weimin Xiong, Weiran He, Weixiao Huang, Weixin Xu, Wenhao Wu, Wenyang He, Xianghui Wei, Xianqing Jia, Xingzhe Wu, Xinran Xu, Xinxing Zu, Xinyu Zhou, Xuehai Pan, Y. Charles, Yang Li, Yangyang Hu, Yangyang Liu, Yanru Chen, Yejie Wang, Yibo Liu, Yidao Qin, Yifeng Liu, Ying Yang, Yiping Bao, Yulun Du, Yuxin Wu, Yuzhi Wang, Zaida Zhou, Zhaoji Wang, Zhaowei Li, Zhen Zhu, Zheng Zhang, Zhexu Wang, Zhilin Yang, Zhiqi Huang, Zihao Huang, Ziyao Xu, Zonghan Yang, and Zongyu Lin. Kimi k1.5: Scaling reinforcement learning with llms. arXiv preprint, 2025. 1 +[55] Qwen Team. Qvq: To see the world with wisdom. https://qwenlm.github.io/blog/qvq-72b-preview, 2024. Accessed: 2024-12-25. 3, 5 +[56] Qwen Team. Qwq: Reflect deeply on the boundaries of the unknown. https://qwenlm.github.io/blog/qwq-32b-preview, 2024. Accessed: 2024-11-28. 1, 3 +[57] Qwen Team. qwen-omni. https://help.aliyun.com/zh/model-studio/user-guide/qwen-omni, 2025. Accessed: 2025-03-26. 5 +[58] Prakhar Verma, Sukruta Prakash Midigeshi, Gaurav Sinha, Arno Solin, Nagarajan Natarajan, and Amit Sharma. Plan*RAG: Efficient test-time planning for retrieval augmented generation. In Workshop on Reasoning and Planning for Large Language Models, 2025. 8 +[59] Xiaoxuan Wang, Ziniu Hu, Pan Lu, Yanqiao Zhu, Jieyu Zhang, Satyen Subramaniam, Arjun R Loomba, Shichang Zhang, Yizhou Sun, and Wei Wang. Scibench: Evaluating college-level scientific problem-solving abilities of large language models. In *Forty-first International Conference on Machine Learning*, 2024. 1 +[60] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian richter, Fei Xia, Ed Chi, Quoc V Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models. In Advances in Neural Information Pro + +cessing Systems, pages 24824-24837. Curran Associates, Inc., 2022. 3, 5 +[61] xAI. Models. https://docs.x.ai/docs/models? cluster=us-east-1#model-constraints, 2024. Accessed: 2024-12-12. 5 +[62] Fengli Xu, Qianyue Hao, Zefang Zong, Jingwei Wang, Yunke Zhang, Jingyi Wang, Xiaochong Lan, Jiahui Gong, Tianjian Ouyang, Fanjin Meng, Chenyang Shao, Yuwei Yan, Qinglong Yang, Yiwen Song, Sijian Ren, Xinyuan Hu, Yu Li, Jie Feng, Chen Gao, and Yong Li. Towards large reasoning models: A survey of reinforced reasoning with large language models. arXiv preprint, 2025. 1, 3 +[63] Siqiao Xue, Tingting Chen, Fan Zhou, Qingyang Dai, Zhixuan Chu, and Hongyuan Mei. Famma: A benchmark for financial domain multilingual multimodal question answering. arXiv preprint, 2024. 2, 4, 5 +[64] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. MM-vet: Evaluating large multimodal models for integrated capabilities. In *Forty-first International Conference on Machine Learning*, 2024. 3 +[65] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Ren-liang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9556-9567, 2024. 1, 2, 3, 4, 5 +[66] Xiang Yue, Tianyu Zheng, Yuansheng Ni, Yubo Wang, Kai Zhang, Shengbang Tong, Yuxuan Sun, Botao Yu, Ge Zhang, Huan Sun, Yu Su, Wenhu Chen, and Graham Neubig. MMMU-pro: A more robust multi-discipline multimodal understanding benchmark. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15134-15186, Vienna, Austria, 2025. Association for Computational Linguistics. 3, 4, 5 +[67] Yilun Zhao, Hongjun Liu, Yitao Long, Rui Zhang, Chen Zhao, and Arman Cohan. Financemath: Knowledge-intensive math reasoning in finance domains. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12841-12858, Bangkok, Thailand, 2024. Association for Computational Linguistics. 2, 3, 4, 5 +[68] Yilun Zhao, Yitao Long, Hongjun Liu, Ryo Kamoi, Linyong Nan, Lyuhao Chen, Yixin Liu, Xiangru Tang, Rui Zhang, and Arman Cohan. DocMath-eval: Evaluating math reasoning capabilities of LLMs in understanding long and specialized documents. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 16103-16120, Bangkok, Thailand, 2024. Association for Computational Linguistics. 4, 5 +[69] Fengbin Zhu, Wenqiang Lei, Youcheng Huang, Chao Wang, Shuo Zhang, Jiancheng Lv, Fuli Feng, and Tat-Seng Chua. TAT-QA: A question answering benchmark on a hybrid of + +tabular and textual content in finance. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 3277-3287, Online, 2021. Association for Computational Linguistics. 2, 3, 5 \ No newline at end of file diff --git a/2025/FinMMR_ Make Financial Numerical Reasoning More Multimodal, Comprehensive, and Challenging/images.zip b/2025/FinMMR_ Make Financial Numerical Reasoning More Multimodal, Comprehensive, and Challenging/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..bd559a9f19dc85fdc960ffdb5397c7477118ef0e --- /dev/null +++ b/2025/FinMMR_ Make Financial Numerical Reasoning More Multimodal, Comprehensive, and Challenging/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bd314a253ee4ea621a36d7af5d7d6b839510264cb82e954d9d24026c5984be69 +size 598838 diff --git a/2025/FinMMR_ Make Financial Numerical Reasoning More Multimodal, Comprehensive, and Challenging/layout.json b/2025/FinMMR_ Make Financial Numerical Reasoning More Multimodal, Comprehensive, and Challenging/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..429d7fcb96157f041c9a5fa29c24c515d9d5d497 --- /dev/null +++ b/2025/FinMMR_ Make Financial Numerical Reasoning More Multimodal, Comprehensive, and Challenging/layout.json @@ -0,0 +1,9884 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 143, + 102, + 471, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 102, + 471, + 147 + ], + "spans": [ + { + "bbox": [ + 143, + 102, + 471, + 147 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathsf{M}}" + }, + { + "bbox": [ + 143, + 102, + 471, + 147 + ], + "type": "text", + "content": " FinMMR: Make Financial Numerical Reasoning More Multimodal, Comprehensive, and Challenging" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 61, + 167, + 552, + 278 + ], + "blocks": [ + { + "bbox": [ + 61, + 167, + 552, + 278 + ], + "lines": [ + { + "bbox": [ + 61, + 167, + 552, + 278 + ], + "spans": [ + { + "bbox": [ + 61, + 167, + 552, + 278 + ], + "type": "image", + "image_path": "6d3fed1f0bcad87000c97473484ddeecc534d3a4c21ada62bf3c83cc519932b9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 58, + 283, + 182, + 418 + ], + "blocks": [ + { + "bbox": [ + 58, + 283, + 182, + 418 + ], + "lines": [ + { + "bbox": [ + 58, + 283, + 182, + 418 + ], + "spans": [ + { + "bbox": [ + 58, + 283, + 182, + 418 + ], + "type": "image", + "image_path": "6144a90463c441bce03a3a2907a21b545073c0b3c5b2f67d5224c79ff4245c54.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 428, + 555, + 452 + ], + "lines": [ + { + "bbox": [ + 55, + 428, + 555, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 428, + 555, + 452 + ], + "type": "text", + "content": "Figure 1. Overview of FinMMR. FinMMR presents three challenges: (1) Visual Perception: 8.7K financial images of 14 categories; (2) Knowledge Reasoning: 4.3K financial questions of 14 subdomains; (3) Numerical Computation: multi-step precise calculations." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 184, + 283, + 332, + 418 + ], + "blocks": [ + { + "bbox": [ + 184, + 283, + 332, + 418 + ], + "lines": [ + { + "bbox": [ + 184, + 283, + 332, + 418 + ], + "spans": [ + { + "bbox": [ + 184, + 283, + 332, + 418 + ], + "type": "image", + "image_path": "f24d79bb435f65d9e7c00c37f347d57d67a992fd34dfd9ffccd047a1b42f51e7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 332, + 284, + 552, + 418 + ], + "blocks": [ + { + "bbox": [ + 332, + 284, + 552, + 418 + ], + "lines": [ + { + "bbox": [ + 332, + 284, + 552, + 418 + ], + "spans": [ + { + "bbox": [ + 332, + 284, + 552, + 418 + ], + "type": "image", + "image_path": "01dee61a627ed956832cce85fa103ad5b11d1ca2ad78aa95efae85d2e2a430fa.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 460, + 200, + 473 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 460, + 200, + 473 + ], + "spans": [ + { + "bbox": [ + 151, + 460, + 200, + 473 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 487, + 297, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 487, + 297, + 691 + ], + "spans": [ + { + "bbox": [ + 53, + 487, + 297, + 691 + ], + "type": "text", + "content": "We present FinMMR, a novel bilingual multimodal benchmark tailored to evaluate the reasoning capabilities of multimodal large language models (MLLMs) in financial numerical reasoning tasks. Compared to existing benchmarks, our work introduces three significant advancements. (1) Multimodality: We meticulously transform existing financial reasoning benchmarks, and construct novel questions from the latest Chinese financial research reports. FinMMR comprises 4.3K questions and 8.7K images spanning 14 categories, including tables, bar charts, and ownership structure charts. (2) Comprehensiveness: FinMMR encompasses 14 financial subdomains, including corporate finance, banking, and industry analysis, significantly exceeding existing benchmarks in financial domain knowledge breadth. (3) Challenge: Models are required to perform multi-step precise numerical reasoning by integrating financial knowledge with the understanding of complex financial" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 462, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 510 + ], + "type": "text", + "content": "images and text. The best-performing MLLM achieves only " + }, + { + "bbox": [ + 313, + 462, + 555, + 510 + ], + "type": "inline_equation", + "content": "53.0\\%" + }, + { + "bbox": [ + 313, + 462, + 555, + 510 + ], + "type": "text", + "content": " accuracy on Hard problems. We believe that FinMMR will drive advancements in enhancing the reasoning capabilities of MLLMs in real-world scenarios." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 536, + 394, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 536, + 394, + 548 + ], + "spans": [ + { + "bbox": [ + 314, + 536, + 394, + 548 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 556, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 556, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 312, + 556, + 555, + 640 + ], + "type": "text", + "content": "Recently, large reasoning models (LRMs) [21, 47, 48, 54, 56, 62], show powerful reasoning capabilities over multi-step reasoning tasks, with train-time scaling and test-time scaling [31, 45]. These reasoning models are proficient in code [11, 29], math [35, 41], and science [59]. Multimodal large language models (MLLMs) [5, 19, 46] also exhibit greater performance on multimodal reasoning [39, 65]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 642, + 556, + 714 + ], + "type": "text", + "content": "Despite significant advancements, there remains a notable gap in understanding the practical applicability of MLLMs in numerical reasoning within real-world scenarios, particularly in high-stakes fields such as finance and healthcare. As shown in Fig. 1, financial analysts in their daily work are required to read visually enriched finan" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3245" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 75, + 76, + 196, + 99 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 76, + 196, + 99 + ], + "spans": [ + { + "bbox": [ + 75, + 76, + 196, + 99 + ], + "type": "text", + "content": "Rich Images Challenge Visual Perception" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 70, + 108, + 208, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 108, + 208, + 171 + ], + "spans": [ + { + "bbox": [ + 70, + 108, + 208, + 171 + ], + "type": "text", + "content": "问题：计算2016年至2024年Q2期间，上海银行和上市城商行的个人住房贷款年均复合增长率之差，结果以百分比表示，保留两位小数。(Calculate the difference in the compound annual growth rate (CAGR) of personal housing loans between Bank of Shanghai and listed city commercial banks from 2016 to Q2 2024. Present the result as a percentage rounded to two decimal places.)" + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 86, + 175, + 190, + 240 + ], + "blocks": [ + { + "bbox": [ + 86, + 175, + 190, + 240 + ], + "lines": [ + { + "bbox": [ + 86, + 175, + 190, + 240 + ], + "spans": [ + { + "bbox": [ + 86, + 175, + 190, + 240 + ], + "type": "image", + "image_path": "bd2ebb0c3cb9c81d16470e42bf5637c0afce94ef8fdb3223377f345ac18a68b5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 170, + 246, + 205, + 252 + ], + "lines": [ + { + "bbox": [ + 170, + 246, + 205, + 252 + ], + "spans": [ + { + "bbox": [ + 170, + 246, + 205, + 252 + ], + "type": "text", + "content": "Answer: 2.17" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 258, + 190, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 258, + 190, + 264 + ], + "spans": [ + { + "bbox": [ + 80, + 258, + 190, + 264 + ], + "type": "text", + "content": "Compound Average Growth Rate (CAGR)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 266, + 211, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 266, + 211, + 306 + ], + "spans": [ + { + "bbox": [ + 69, + 266, + 211, + 306 + ], + "type": "text", + "content": "问题：请计算惠济新生对日照公司的间接持股比例，结果以百分比表示，保留两位小数。(Please calculate the indirect shareholding ratio of Huiji Xinsheng in Rizhao Company. Present the result as a percentage, rounded to two decimal places.)" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 92, + 310, + 186, + 377 + ], + "blocks": [ + { + "bbox": [ + 92, + 310, + 186, + 377 + ], + "lines": [ + { + "bbox": [ + 92, + 310, + 186, + 377 + ], + "spans": [ + { + "bbox": [ + 92, + 310, + 186, + 377 + ], + "type": "image", + "image_path": "2c06b20ae45c66588922b3929d805d89c940c6b68154019b66b39aa3f0785f0a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 167, + 389, + 206, + 396 + ], + "lines": [ + { + "bbox": [ + 167, + 389, + 206, + 396 + ], + "spans": [ + { + "bbox": [ + 167, + 389, + 206, + 396 + ], + "type": "text", + "content": "Answer: 13.65" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 96, + 401, + 171, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 401, + 171, + 407 + ], + "spans": [ + { + "bbox": [ + 96, + 401, + 171, + 407 + ], + "type": "text", + "content": "Indirect Shareholding Ratio" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 411, + 211, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 411, + 211, + 434 + ], + "spans": [ + { + "bbox": [ + 69, + 411, + 211, + 434 + ], + "type": "text", + "content": "问题：请计算2022年第四季度和2023年第一季度的总资本开支，并将其与2021年第四季度的资本开支进行比较，计算下降百分比，结果保留两位小数。" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 434, + 211, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 434, + 211, + 465 + ], + "spans": [ + { + "bbox": [ + 69, + 434, + 211, + 465 + ], + "type": "text", + "content": "(Calculate the total capital expenditure for Q4 2022 and Q1 2023 combined, and compare it to the capital expenditure for Q4 2021. Calculate the percentage decrease, rounded to two decimal places.)" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 72, + 468, + 206, + 514 + ], + "blocks": [ + { + "bbox": [ + 72, + 468, + 206, + 514 + ], + "lines": [ + { + "bbox": [ + 72, + 468, + 206, + 514 + ], + "spans": [ + { + "bbox": [ + 72, + 468, + 206, + 514 + ], + "type": "image", + "image_path": "1941a902cf9631c3e2f1f6cdf3f7c0af2ac31eecc1e88019fce8aa05dc5eda03.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 523, + 208, + 529 + ], + "lines": [ + { + "bbox": [ + 162, + 523, + 208, + 529 + ], + "spans": [ + { + "bbox": [ + 162, + 523, + 208, + 529 + ], + "type": "text", + "content": "Answer: -110.36" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 108, + 536, + 163, + 544 + ], + "lines": [ + { + "bbox": [ + 108, + 536, + 163, + 544 + ], + "spans": [ + { + "bbox": [ + 108, + 536, + 163, + 544 + ], + "type": "text", + "content": "Capital Expenditure" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 558, + 555, + 581 + ], + "lines": [ + { + "bbox": [ + 55, + 558, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 555, + 581 + ], + "type": "text", + "content": "Figure 2. Sampled FinMMR examples with two language (i.e., English and Chinese), rich images and different knowledge. The questions and images need expert-level visual perception, knowledge reasoning and numerical computation." + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 233, + 76, + 372, + 99 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 76, + 372, + 99 + ], + "spans": [ + { + "bbox": [ + 233, + 76, + 372, + 99 + ], + "type": "text", + "content": "Comprehensive Subdomains Challenge Knowledge Reasoning" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 233, + 108, + 373, + 197 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 108, + 373, + 197 + ], + "spans": [ + { + "bbox": [ + 233, + 108, + 373, + 197 + ], + "type": "text", + "content": "Q: Table 19.3 shows a book balance sheet for the Wishing Well Motel chain. The company's long-term debt is secured by its real estate assets, but it also uses short-term bank loans as a permanent source of financing. It pays " + }, + { + "bbox": [ + 233, + 108, + 373, + 197 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 233, + 108, + 373, + 197 + ], + "type": "text", + "content": " interest on the bank debt and " + }, + { + "bbox": [ + 233, + 108, + 373, + 197 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 233, + 108, + 373, + 197 + ], + "type": "text", + "content": " interest on the secured debt. Wishing Well has 10 million shares of stock outstanding, trading at " + }, + { + "bbox": [ + 233, + 108, + 373, + 197 + ], + "type": "inline_equation", + "content": "\\$ 90" + }, + { + "bbox": [ + 233, + 108, + 373, + 197 + ], + "type": "text", + "content": "per share. The expected return on Wishing Well's common stock is " + }, + { + "bbox": [ + 233, + 108, + 373, + 197 + ], + "type": "inline_equation", + "content": "18\\%" + }, + { + "bbox": [ + 233, + 108, + 373, + 197 + ], + "type": "text", + "content": ". Calculate Wishing Well's WACC. Assume that the book and market values of Wishing Well's debt are the same. The marginal tax rate is " + }, + { + "bbox": [ + 233, + 108, + 373, + 197 + ], + "type": "inline_equation", + "content": "21\\%" + }, + { + "bbox": [ + 233, + 108, + 373, + 197 + ], + "type": "text", + "content": ". Answer as a percentage to single decimal place." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 233, + 198, + 258, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 198, + 258, + 205 + ], + "spans": [ + { + "bbox": [ + 233, + 198, + 258, + 205 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 236, + 210, + 371, + 251 + ], + "blocks": [ + { + "bbox": [ + 236, + 210, + 371, + 251 + ], + "lines": [ + { + "bbox": [ + 236, + 210, + 371, + 251 + ], + "spans": [ + { + "bbox": [ + 236, + 210, + 371, + 251 + ], + "type": "table", + "html": "
Cash and marketable securities100Bank loan280
Accounts receivable200Accounts payable120
Inventory50Current liabilities400
Current assets350
Real estate2,100Long-term debt1,800
Other assets150Equity400
Total2,600Total2,600
", + "image_path": "f3693169345467c5dd0559a9a553a3dacece01a3098c7f498e0572128f1d1ac5.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 235, + 254, + 365, + 260 + ], + "lines": [ + { + "bbox": [ + 235, + 254, + 365, + 260 + ], + "spans": [ + { + "bbox": [ + 235, + 254, + 365, + 260 + ], + "type": "text", + "content": "TABLE 19.3 Book balance sheet for Wishing Well Inc. (figures in $ millions)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 17 + }, + { + "bbox": [ + 243, + 266, + 309, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 266, + 309, + 272 + ], + "spans": [ + { + "bbox": [ + 243, + 266, + 309, + 272 + ], + "type": "text", + "content": "Weighted Average Cost" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 233, + 277, + 371, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 277, + 371, + 314 + ], + "spans": [ + { + "bbox": [ + 233, + 277, + 371, + 314 + ], + "type": "text", + "content": "Q: For each of the investments shown in the following table, calculate the rate of return earned over the unspecified time period. What is the rate of return for Investment A? Answer as a percentage to the nearest integer." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 233, + 317, + 258, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 317, + 258, + 324 + ], + "spans": [ + { + "bbox": [ + 233, + 317, + 258, + 324 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 21 + }, + { + "type": "table", + "bbox": [ + 241, + 334, + 367, + 380 + ], + "blocks": [ + { + "bbox": [ + 241, + 334, + 367, + 380 + ], + "lines": [ + { + "bbox": [ + 241, + 334, + 367, + 380 + ], + "spans": [ + { + "bbox": [ + 241, + 334, + 367, + 380 + ], + "type": "table", + "html": "
InvestmentCash flow during periodBeginning-of-period valueEnd-of-period value
A$-2,800$23,400$20,100
B16,000225,000324,000
C7006,5008,000
D3,58036,60046,500
E-50062,70052,800
", + "image_path": "deea4bed290fc9743d21fcb1d6cf83389fb6841bdaf530acccb4cde1bb404345.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "table_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 269, + 394, + 326, + 400 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 269, + 394, + 326, + 400 + ], + "spans": [ + { + "bbox": [ + 269, + 394, + 326, + 400 + ], + "type": "text", + "content": "Rate of Return (RoR)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 231, + 403, + 373, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 403, + 373, + 460 + ], + "spans": [ + { + "bbox": [ + 231, + 403, + 373, + 460 + ], + "type": "text", + "content": "Q: Ricky is considering purchasing an apartment costing \\( \\\\( {700},{000} \\) . He will pay a 30% down payment and take out a mortgage for the remainder. Since he just got married and wants to save some money for future use, he will choose the plan with the lowest monthly payment. After visiting several banks, he received the following mortgage offers: What is the monthly payment for Bank A? Answer to two decimal places." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 232, + 461, + 255, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 461, + 255, + 468 + ], + "spans": [ + { + "bbox": [ + 232, + 461, + 255, + 468 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 25 + }, + { + "type": "table", + "bbox": [ + 236, + 471, + 372, + 534 + ], + "blocks": [ + { + "bbox": [ + 236, + 471, + 372, + 534 + ], + "lines": [ + { + "bbox": [ + 236, + 471, + 372, + 534 + ], + "spans": [ + { + "bbox": [ + 236, + 471, + 372, + 534 + ], + "type": "table", + "html": "
BankInterest rateTerm (years)
A3.5%15
B320
C425
D4.518
", + "image_path": "e0b2f16f8afad2b055de20658573bdc7e1ed2ba9c07e7edbe3012dd267e4874e.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "table_body" + } + ], + "index": 26 + }, + { + "bbox": [ + 395, + 76, + 541, + 99 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 76, + 541, + 99 + ], + "spans": [ + { + "bbox": [ + 395, + 76, + 541, + 99 + ], + "type": "text", + "content": "Complex Formulas Challenge Numerical Computation" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 399, + 108, + 537, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 108, + 537, + 140 + ], + "spans": [ + { + "bbox": [ + 399, + 108, + 537, + 140 + ], + "type": "text", + "content": "Q: What is the total weighted average Cash and Payment-In-Kind (PIK) interest rate payable under the subordinated and senior notes portfolio in the year 2023?" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 400, + 142, + 425, + 149 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 142, + 425, + 149 + ], + "spans": [ + { + "bbox": [ + 400, + 142, + 425, + 149 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 30 + }, + { + "type": "table", + "bbox": [ + 405, + 152, + 533, + 186 + ], + "blocks": [ + { + "bbox": [ + 276, + 537, + 326, + 544 + ], + "lines": [ + { + "bbox": [ + 276, + 537, + 326, + 544 + ], + "spans": [ + { + "bbox": [ + 276, + 537, + 326, + 544 + ], + "type": "text", + "content": "Monthly Payment" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 405, + 152, + 533, + 186 + ], + "lines": [ + { + "bbox": [ + 405, + 152, + 533, + 186 + ], + "spans": [ + { + "bbox": [ + 405, + 152, + 533, + 186 + ], + "type": "table", + "html": "
EUR
In Mio €201820172016%2015
Deutschland3.498.5003.498.5003.498.500100%100%
USA1.200.0001.200.0001.200.000--
Nordafrika, Naher Osten1.200.0001.200.0001.200.000--
Asien-Pazifik1.200.0001.200.0001.200.000--
davon1.200.0001.200.0001.200.000--
davon1.200.0001.200.0001.200.000--
davon1.200.0001.200.001.200.000--
davon1.200.0001.200.0001.200.000--
davon1.200.0001.200.0001.200.000--
deutschland3.498.5003.498.5003.498.500100%100%
davon1.200.0001.200.0001.200.000--
", + "image_path": "071fb689a0aaef17aff21a3b43919fb6d114e822259edf1f903d29f34753a1c4.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "table_body" + } + ], + "index": 31 + }, + { + "bbox": [ + 400, + 196, + 430, + 203 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 196, + 430, + 203 + ], + "spans": [ + { + "bbox": [ + 400, + 196, + 430, + 203 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 32 + }, + { + "type": "table", + "bbox": [ + 405, + 205, + 533, + 225 + ], + "blocks": [ + { + "bbox": [ + 405, + 205, + 533, + 225 + ], + "lines": [ + { + "bbox": [ + 405, + 205, + 533, + 225 + ], + "spans": [ + { + "bbox": [ + 405, + 205, + 533, + 225 + ], + "type": "table", + "html": "
[average = (investmenti × cashi) / ∑investmenti] + (investmenti × PIKi) / ∑investmenti]
", + "image_path": "ddf3c43dd63471cdf2c551056206f8665b1b489ee3598c16488fda12ee8959c6.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "table_body" + } + ], + "index": 33 + }, + { + "bbox": [ + 421, + 242, + 518, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 421, + 242, + 518, + 248 + ], + "spans": [ + { + "bbox": [ + 421, + 242, + 518, + 248 + ], + "type": "text", + "content": "Total Weighted Average Cash and" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 417, + 249, + 520, + 256 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 417, + 249, + 520, + 256 + ], + "spans": [ + { + "bbox": [ + 417, + 249, + 520, + 256 + ], + "type": "text", + "content": "Payment-In-Kind (PIK) Interest Rate" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 399, + 259, + 539, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 259, + 539, + 290 + ], + "spans": [ + { + "bbox": [ + 399, + 259, + 539, + 290 + ], + "type": "text", + "content": "Q: John oversees a fund, with the returns for the first three years displayed below: What will be the holding period return (expressed as a percentage)? Answer to three decimal places." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 400, + 293, + 425, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 293, + 425, + 300 + ], + "spans": [ + { + "bbox": [ + 400, + 293, + 425, + 300 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 38 + }, + { + "type": "table", + "bbox": [ + 408, + 304, + 531, + 341 + ], + "blocks": [ + { + "bbox": [ + 497, + 230, + 533, + 237 + ], + "lines": [ + { + "bbox": [ + 497, + 230, + 533, + 237 + ], + "spans": [ + { + "bbox": [ + 497, + 230, + 533, + 237 + ], + "type": "text", + "content": "Answer: 12.0" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 408, + 304, + 531, + 341 + ], + "lines": [ + { + "bbox": [ + 408, + 304, + 531, + 341 + ], + "spans": [ + { + "bbox": [ + 408, + 304, + 531, + 341 + ], + "type": "table", + "html": "
YearInvestmentReturn
1$ 50012%
2$ 6005%
3$ 10001%
", + "image_path": "2b31eb1cea01af1f8b62652401cdbd73b9636e36bdb39fc7386c6513f67f6ba8.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "table_body" + } + ], + "index": 39 + }, + { + "bbox": [ + 400, + 347, + 430, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 347, + 430, + 354 + ], + "spans": [ + { + "bbox": [ + 400, + 347, + 430, + 354 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 414, + 357, + 522, + 377 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 414, + 357, + 522, + 377 + ], + "spans": [ + { + "bbox": [ + 414, + 357, + 522, + 377 + ], + "type": "interline_equation", + "content": "\\left[ H P R = \\left(\\frac {\\sum \\left(\\text {i n v e s t m e n t} _ {t} \\times (1 + \\text {r e t u r n} _ {t})\\right)}{\\sum \\text {i n v e s t m e n t}} - 1\\right) \\times 1 0 0 \\right]", + "image_path": "59d9c12d1c323d81cab4b8303878a560a876703e361aa9364d3f422da919b474.jpg" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 495, + 382, + 535, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 495, + 382, + 535, + 388 + ], + "spans": [ + { + "bbox": [ + 495, + 382, + 535, + 388 + ], + "type": "text", + "content": "Answer: 4.762" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 430, + 394, + 508, + 401 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 430, + 394, + 508, + 401 + ], + "spans": [ + { + "bbox": [ + 430, + 394, + 508, + 401 + ], + "type": "text", + "content": "Holding Period Return (HPR)" + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 401, + 403, + 533, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 401, + 403, + 533, + 436 + ], + "spans": [ + { + "bbox": [ + 401, + 403, + 533, + 436 + ], + "type": "text", + "content": "Q: What is the average quarterly share price in 2021 assuming each quarter had a completely uniform price distribution in high and low bid prices, measured in dollars?" + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 401, + 440, + 426, + 447 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 401, + 440, + 426, + 447 + ], + "spans": [ + { + "bbox": [ + 401, + 440, + 426, + 447 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 45 + }, + { + "type": "table", + "bbox": [ + 409, + 448, + 531, + 488 + ], + "blocks": [ + { + "bbox": [ + 409, + 448, + 531, + 488 + ], + "lines": [ + { + "bbox": [ + 409, + 448, + 531, + 488 + ], + "spans": [ + { + "bbox": [ + 409, + 448, + 531, + 488 + ], + "type": "table", + "html": "
Quarter EndedHigh BidLow Bid
October 31, 2021 (1)$4.51$3.99
July 31, 2021$8.47$2.30
April 30, 2021$13.00$3.80
January 31, 2021$14.50$0.015
October 31, 2020 (2)$0.199$0.01
", + "image_path": "69819e936c14322d348f4d9eeeb40f6090f6291b91df1508612a22b649c9831a.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "table_body" + } + ], + "index": 46 + }, + { + "bbox": [ + 402, + 494, + 430, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 494, + 430, + 500 + ], + "spans": [ + { + "bbox": [ + 402, + 494, + 430, + 500 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 421, + 502, + 506, + 522 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 421, + 502, + 506, + 522 + ], + "spans": [ + { + "bbox": [ + 421, + 502, + 506, + 522 + ], + "type": "interline_equation", + "content": "[ \\text {A v e r a g e} = \\frac {1}{4} \\sum_ {i = 1} ^ {4} \\frac {\\text {H i g h} _ {i} + \\text {L o w} _ {i}}{2} ]", + "image_path": "53f6a7edcc05163720b773581c22e0510df714a51f52cde6d37e35a5775ed4d4.jpg" + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 497, + 523, + 534, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 497, + 523, + 534, + 529 + ], + "spans": [ + { + "bbox": [ + 497, + 523, + 534, + 529 + ], + "type": "text", + "content": "Answer: 6.32" + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 438, + 535, + 504, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 438, + 535, + 504, + 542 + ], + "spans": [ + { + "bbox": [ + 438, + 535, + 504, + 542 + ], + "type": "text", + "content": "Average Quarterly Price" + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 55, + 601, + 296, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 601, + 296, + 709 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 296, + 709 + ], + "type": "text", + "content": "cial documents, extract key financial indicators from tables, images, and texts, and perform multi-step precise numerical calculations to support professional decision-making. Similarly, to achieve expert artificial general intelligence (AGI) [8, 25, 42, 43, 65], MLLMs are expected to comprehend complex domain-specific images akin to human experts, and apply domain knowledge to perform accurate numerical reasoning. This raises the question: Can current MLLMs seamlessly integrate vision and text to perform" + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 313, + 601, + 553, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 601, + 553, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 601, + 553, + 625 + ], + "type": "text", + "content": "domain-specific complex reasoning, matching the proficiency of LRMs in pure text-based tasks?" + } + ] + } + ], + "index": 53 + }, + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "content": "Specifically, we choose the financial domain to evaluate the complex reasoning capabilities of MLLMs, where precision and transparent reasoning are paramount [32]. Existing numerical reasoning benchmarks for finance are limited in their text-based reasoning, coverage of specific financial knowledge, and complexity of reasoning [14, 15, 32, 67, 69]. FAMMA [63] is mainly modelled after textbooks" + } + ] + } + ], + "index": 54 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3246" + } + ] + } + ], + "index": 55 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "and CFA exam questions, MathVista [39] does not involve the application of financial knowledge, MMMU [65] and MMMU-Pro [66] are restricted to multiple-choice questions, diverging from authentic financial problem-solving. The lack of high-quality, knowledge-intensive multimodal financial numerical reasoning benchmarks makes it challenging to objectively evaluate the actual reasoning capabilities of MLLMs and analyze their shortcomings." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 169, + 294, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 169, + 294, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 169, + 294, + 277 + ], + "type": "text", + "content": "Therefore, we propose FinMMR, a bilingual multimodal numerical reasoning benchmark designed to evaluate the reasoning capabilities of MLLMs in the finance domain. The benchmark comprises 4.3K problems, covering 14 financial subdomains (e.g., corporate finance and industry analysis), with 8.7K images derived from 14 categories (e.g., tables and ownership structure charts). Each problem includes rich images, an unambiguous question, a Python-formatted solution, and a precise answer." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 279, + 295, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 279, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 295, + 434 + ], + "type": "text", + "content": "For multimodality, without representing financial tables as structured text, FinMMR represent all tables, charts, and diagrams as images. For comprehensiveness, FinMMR covers 14 financial subdomains and two languages (i.e., English and Chinese), demanding domain knowledge such as option pricing and portfolio management. For challenge, FinMMR focus on multi-step numerical reasoning, requiring models to provide exact numerical answers under strict evaluation criteria (emphasizing units, percentages, and decimal places). Furthermore, we mix each Chinese questions with two distractor images that are contextually adjacent to the ground images, approaching real-world multimodal reasoning scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 436, + 295, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 295, + 484 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 295, + 484 + ], + "type": "text", + "content": "We evaluate 15 state-of-the-art (SOTA) MLLMs [18-21, 46-48, 56], utilizing Chain-of-Thought (CoT) [60] and Program-of-Thought (PoT) [13]. The experimental results on FinMMR reveals three key findings:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 486, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 486, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 617 + ], + "type": "text", + "content": "- MLLMs Face Significant Challenges in Domain-Specific Multimodal Numerical Reasoning: All evaluated models underperform on FinMMR. On the Hard test set, the best-performing model, Claude 3.7 Sonnet with 64K extended thinking, achieves only " + }, + { + "bbox": [ + 55, + 486, + 295, + 617 + ], + "type": "inline_equation", + "content": "53.00\\%" + }, + { + "bbox": [ + 55, + 486, + 295, + 617 + ], + "type": "text", + "content": " accuracy, while OpenAI o1 achieves merely " + }, + { + "bbox": [ + 55, + 486, + 295, + 617 + ], + "type": "inline_equation", + "content": "48.40\\%" + }, + { + "bbox": [ + 55, + 486, + 295, + 617 + ], + "type": "text", + "content": ". Through error analysis, we identify that visual perception, knowledge reasoning, and numerical computation collectively pose challenges to MLLMs. Current MLLMs still struggle with complex multimodal reasoning tasks in specialized domains, compared to text-based reasoning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": "- Better Synergy Between Visual Perception and Complex Reasoning Is Needed: Distracting images lead to a greater than " + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": " drop in accuracy for Qwen2.5-VL-72B compared to ground images alone, indicating that irrelevant visual information severely impacts multimodal reasoning. By decoupling visual filtering and reasoning, Qwen2.5-VL-72B improved from " + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "inline_equation", + "content": "64.73\\%" + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "inline_equation", + "content": "71.56\\%" + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": ". Combining MLLMs with LRMs, by efficiently" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 322, + 72, + 553, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 553, + 143 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 553, + 143 + ], + "type": "text", + "content": "Parsing visual information into structured text and leveraging the LRMs' text-based reasoning capabilities, can also yield better performance. The combination of GPT4o and DeepSeek-R1 achieves " + }, + { + "bbox": [ + 322, + 72, + 553, + 143 + ], + "type": "inline_equation", + "content": "86.72\\%" + }, + { + "bbox": [ + 322, + 72, + 553, + 143 + ], + "type": "text", + "content": " accuracy on 1,160 tabular QA instances, outperforming Claude 3.7 Sonnet " + }, + { + "bbox": [ + 322, + 72, + 553, + 143 + ], + "type": "inline_equation", + "content": "(83.53\\%)" + }, + { + "bbox": [ + 322, + 72, + 553, + 143 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 144, + 553, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 144, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 314, + 144, + 553, + 275 + ], + "type": "text", + "content": "- Refined Structured Domain Knowledge Enhances MLLMs' Complex Reasoning: Leading MLLMs lack sufficient experience in applying rich domain knowledge when solving complex reasoning tasks. By annotating structured financial functions and leveraging the model's ability to generate retrieval questions and make judgments, knowledge augmentation can significantly improve MLLMs' performance. MLLMs achieve absolute improvements ranging from 2.76 to 4.31 percentage points, allowing weaker models to approach SOTA performance, while SOTA models exhibit further gains." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 276, + 553, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 276, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 553, + 384 + ], + "type": "text", + "content": "These findings highlight the bottlenecks of MLLMs in complex multimodal reasoning tasks in expert domains closer to real-world scenarios. They emphasize the need for continuous improvements in three key areas: more intricate visual perception, more specialized knowledge reasoning, and more accurate numerical computation. Alternatively, leveraging tools or model combinations can help achieve a balance between performance and cost, enabling MLLMs to perform expert-level reasoning tasks like human experts." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 396, + 400, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 396, + 400, + 408 + ], + "spans": [ + { + "bbox": [ + 314, + 396, + 400, + 408 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 416, + 429, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 416, + 429, + 428 + ], + "spans": [ + { + "bbox": [ + 314, + 416, + 429, + 428 + ], + "type": "text", + "content": "2.1. LRMs and MLLMs" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 434, + 555, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 555, + 613 + ], + "type": "text", + "content": "By integrating train-time scaling and test-time scaling [31, 45], LRMs have demonstrated remarkable reasoning capabilities [62]. However, most LRMs are limited to handling text-based problems. The growing demand for solving real-world tasks has spurred the development of MLLMs [5, 19, 55] and benchmarks designed to evaluate the perception and reasoning abilities of MLLMs [10, 23, 27, 30, 33, 36, 38, 64-66]. For instance, MME-CoT [30] evaluates models' space-time understanding, while EMMA [27] focuses STEM subjects. Following this trend, domain-specific benchmarks which require deep domain expertise have emerged, such as MathVista [39] for mathematics and GMAI-MMBench [12] for medicine. Yet, financial reasoning remains an unexplored area in the current landscape of MLLM benchmarks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 624, + 443, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 624, + 443, + 635 + ], + "spans": [ + { + "bbox": [ + 314, + 624, + 443, + 635 + ], + "type": "text", + "content": "2.2. Financial Benchmarks" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "content": "The financial domain presents a distinct and more formidable set of challenges for model evaluation, which arise from its inherent complexity, information density, and dependence on expertise. The majority of existing text-only financial numerical reasoning benchmarks [14, 15, 32, 67-69] are constrained by limitations such as sub-optimal an" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3247" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 307, + 196 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 307, + 196 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 307, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 307, + 196 + ], + "type": "table", + "html": "
PropertyValue
# Total Questions4,300
# Easy/Medium/Hard1,300/1,500/1,500
# Validation/Test900/3,400
# Chinese/English2,150/2,150
# Operators (Easy/Medium/Hard)1.75/2.97/5.34
# Lines of Code (Easy/Medium/Hard)4.14/5.06/7.34
# Parentheses (Easy/Medium/Hard)0.88/3.11/4.25
# Difficulty (Easy/Medium/Hard)1.93/2.96/3.79
", + "image_path": "dc0f3a27e22ff74eb53812730eabd607fbc90cff709eb466a4850619c94679b8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 204, + 293, + 216 + ], + "lines": [ + { + "bbox": [ + 56, + 204, + 293, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 204, + 293, + 216 + ], + "type": "text", + "content": "Table 1. Key statistics of FinMMR (Avg values of three subsets)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 238, + 295, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 295, + 298 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 295, + 298 + ], + "type": "text", + "content": "notation quality, narrow domain knowledge coverage, and overly simplistic reasoning tasks. Although FinanceReasoning [53] offers complex tasks with rich knowledge and high-quality annotations, its text-only modality limits multimodal reasoning evaluation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 299, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 299, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 299, + 296, + 407 + ], + "type": "text", + "content": "Recent multimodal financial benchmarks have sought to bridge this gap but still possess limitations. FAMMA [63] being sourced from textbooks and examinations does not mirror the real-world tasks. FinMME [40] uses a multiple-choice format, which may overestimate model reasoning due to guesswork. MME-Finance [24] is constrained by coarse annotations and an isolated assessment of domain knowledge, limiting holistic evaluation of real-world financial reasoning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 420, + 182, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 420, + 182, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 420, + 182, + 433 + ], + "type": "text", + "content": "3. FinMMR Benchmark" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 440, + 182, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 440, + 182, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 440, + 182, + 452 + ], + "type": "text", + "content": "3.1. Overview of FinMMR" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 458, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 458, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 458, + 295, + 567 + ], + "type": "text", + "content": "We introduce FinMMR, a bilingual (English and Chinese) multimodal benchmark for evaluating the reasoning capabilities of MLLMs in financial numerical reasoning tasks. FinMMR consists of 4,300 questions covering 14 financial subdomains (e.g., corporate finance, industry analysis, financial markets, asset management). The key statistics are summarized in Tab. 1, and the composition of sub-domains and images is illustrated in Fig. 1. As illustrated in Fig. 2, FinMMR introduces three key challenges:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "text", + "content": "- Rich Images Challenge Visual Perception: FinMMR comprises 8.7K images from 14 categories (e.g., bar charts, line charts, ownership structure charts). MLLMs must identify relevant images among distractors and extract critical information from the correct images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 630, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 677 + ], + "type": "text", + "content": "- Comprehensive Subdomains Challenge Knowledge Reasoning: MLLMs need to flexibly apply diverse domain-specific financial knowledge from 14 subdomains to solve multi-step reasoning tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "- Complex Formulas Challenge Numerical Computation: All questions require precise numerical answers, eliminating the potential bias from lucky guesses that" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 323, + 72, + 484, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 72, + 484, + 84 + ], + "spans": [ + { + "bbox": [ + 323, + 72, + 484, + 84 + ], + "type": "text", + "content": "could occur in multiple-choice formats." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 95, + 443, + 107 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 95, + 443, + 107 + ], + "spans": [ + { + "bbox": [ + 314, + 95, + 443, + 107 + ], + "type": "text", + "content": "3.2. Data Curation Process" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 114, + 555, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 114, + 555, + 221 + ], + "spans": [ + { + "bbox": [ + 313, + 114, + 555, + 221 + ], + "type": "text", + "content": "We first curated a subset of questions from public text-based financial reasoning benchmarks and systematically transformed them into multimodal problems using a unified standard. Subsequently, we constructed a novel multimodal Chinese Research Report Question Answering (CR-RQA) benchmark from scratch, merging two data sources into FinMMR. Each question is accompanied by an executable Python solution, yields a numerical answer and demonstrates a clear reasoning chain." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 221, + 555, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 221, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 221, + 555, + 354 + ], + "type": "text", + "content": "Update to Public Benchmarks We re-announce 124 and 163 financial questions from MMMU [65] and MMMU-Pro [66], respectively. Following rigorous verification, these questions were directly incorporated into our dataset. Furthermore, we extracted 77, 288, and 795 questions from FinanceMath [67], CodeTAT-QA [32], and CodeFinQA [32], respectively. From DocMath-Eval [68], we further obtained 703 questions from its four subsets. For each question, we rendered any tabular data as images while removing the corresponding table information from the text, ensuring that MLLMs cannot rely solely on textual content." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 354, + 554, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 554, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 554, + 484 + ], + "type": "text", + "content": "Building a Novel Dataset from Scratch We collect 90 research reports, all of which are obtained through authorized access and cover diverse topics such as industry research, macroeconomic analysis, and strategy research. We use 360LayoutAnalysis [1] to extract informative images and discard those lacking explicit numerical data, reducing ambiguity. For each retained image, we prompt Qwen-VLMax [6] to formulate questions requiring multiple reasoning steps or complex calculations. Each question is accompanied by an executable Python solution and a definitive numerical answer." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 486, + 554, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 554, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 554, + 534 + ], + "type": "text", + "content": "Furthermore, we introduce distractor images sourced from the same reports adjacent to ground images to challenge MLLMs in extracting relevant numerical information from structured, densely packed visuals." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 534, + 554, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 554, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 554, + 629 + ], + "type": "text", + "content": "Data Quality Assurance This process ensured that every question was clearly written, featured a detailed reasoning solution, and included an accurate final answer. The annotators included 16 graduate students in finance and two experts holding CFA certifications. With the aid of LLMs, this meticulous verification process spanned three months, culminating in a dataset that meets high standards of clarity and correctness." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 630, + 554, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 691 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 691 + ], + "type": "text", + "content": "Dataset Splitting and Release To classify the problems by difficulty, we employ a heuristic algorithm that takes into account the number of operators " + }, + { + "bbox": [ + 313, + 630, + 554, + 691 + ], + "type": "inline_equation", + "content": "(o)" + }, + { + "bbox": [ + 313, + 630, + 554, + 691 + ], + "type": "text", + "content": ", code lines " + }, + { + "bbox": [ + 313, + 630, + 554, + 691 + ], + "type": "inline_equation", + "content": "(l)" + }, + { + "bbox": [ + 313, + 630, + 554, + 691 + ], + "type": "text", + "content": " and parentheses pairs " + }, + { + "bbox": [ + 313, + 630, + 554, + 691 + ], + "type": "inline_equation", + "content": "(p)" + }, + { + "bbox": [ + 313, + 630, + 554, + 691 + ], + "type": "text", + "content": " in the Python solution. Specifically, the difficulty of reasoning " + }, + { + "bbox": [ + 313, + 630, + 554, + 691 + ], + "type": "inline_equation", + "content": "rc" + }, + { + "bbox": [ + 313, + 630, + 554, + 691 + ], + "type": "text", + "content": " of a problem is defined as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 350, + 701, + 555, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 701, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 350, + 701, + 555, + 715 + ], + "type": "interline_equation", + "content": "r c = \\ln (\\max (o, 1)) + \\ln (\\max (l + p, 1)) \\tag {1}", + "image_path": "2949bf6b8d1bace0b80cd8901daa5b802714cc1bf957ddbd0343e7bb93dee65b.jpg" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3248" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 242, + 204 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 242, + 204 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 242, + 204 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 242, + 204 + ], + "type": "image", + "image_path": "36bf93ff9009eed725d220edcdd51597f18cfd498d2cf7ea7c772bb200c376b8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 213, + 555, + 258 + ], + "lines": [ + { + "bbox": [ + 54, + 213, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 54, + 213, + 555, + 258 + ], + "type": "text", + "content": "Figure 3. The comparison between finance-related benchmarks. These benchmarks vary in size, domain coverage, modalities, and question type, with some focusing on text-only data while others include images. Each axis has scale labels with varying ranges to measure the number of questions from each benchmark across different subdomains. In the Modalities, T means text input, I means Images input, P.I. means pure images input. In the Question Type, MC means multi-choice answer, NUM means numerical answer." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 256, + 70, + 559, + 205 + ], + "blocks": [ + { + "bbox": [ + 256, + 70, + 559, + 205 + ], + "lines": [ + { + "bbox": [ + 256, + 70, + 559, + 205 + ], + "spans": [ + { + "bbox": [ + 256, + 70, + 559, + 205 + ], + "type": "table", + "html": "
BenchmarkSize (Fin)Domain CoverageModalitiesQuestion Type
MMMU [65]11,550 (1,603)10T+IMC
MMMU-Pro [66]1,730 (286)10T+I, P.I.MC
FinanceMath [67]1,200 (1,200)10TNUM
CodeTAT-QA [32]3,144 (3,144)6TNUM
CodeFinQA [32]5,463 (5,463)13TNUM
DocMath-Eval [68]4,000 (4,000)12TNUM
FAMMA [63]1,758 (1,758)8T+IMC, NUM
CRRQA (ours)2,150 (2,150)13T+I, P.I.NUM
FinMMR (ours)4,300 (4,300)14T+I, P.I.NUM
", + "image_path": "41f33f9be512944cdddffd9bd2c6513d27bdc58f56a962c25a2802bf3828c27c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 278, + 296, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 278, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 296, + 373 + ], + "type": "text", + "content": "FinMMR is classified as Easy (1,300 examples), Medium (1,500 examples) and Hard (1,500 examples) based on this formula, with each level randomly split into test and validation sets. All questions are publicly available, while only the 300 validation answers per level are released, while test answers remain private to prevent data leakage [22, 50, 52]. We maintain an online evaluation platform that enables researchers to assess their models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 384, + 267, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 267, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 267, + 396 + ], + "type": "text", + "content": "3.3. Comparisons with Existing Benchmarks" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 401, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 401, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 54, + 401, + 295, + 569 + ], + "type": "text", + "content": "As illustrated in Fig. 3, previous work has studied multidiscipline multimodal reasoning [65, 66], general mathematical reasoning [39] or text-based financial QA [32, 67, 68]. FinMMR focus on multimodal financial numerical reasoning, curating 4,300 questions requiring a deep understanding of domain-specific images (e.g., earnings reports, candlestick charts). To mimic real-world scenario, we deliberately incorporate 3,938 distractors into 2,150 questions to rigorously evaluate MLLMs' visual perception capability. Compared to existing financial benchmarks, they suffer from narrow domain coverage [14, 69]. FinMMR encompasses 14 financial subdomains and 14 image categories, comprehensively evaluating MLLMs' domain-specific reasoning capabilities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 167, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 167, + 596 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 167, + 596 + ], + "type": "text", + "content": "4. Evaluation System" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 602, + 295, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 602, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 602, + 295, + 639 + ], + "type": "text", + "content": "To facilitate the evaluation of complex reasoning on FinMMR, we established a dedicated evaluation system. All MLLMs evaluated were accessed through official APIs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 647, + 250, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 250, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 250, + 661 + ], + "type": "text", + "content": "4.1. Multimodal Large Language Models" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "We systematically evaluate the multimodal reasoning capabilities of 15 recent MLLMs under the zero-shot setting. The evaluated MLLMs are: OpenAI o1 [46], GPT-4o [44], Claude 3.7 Sonnet (including thinking mode) [5]," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 278, + 554, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 278, + 554, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 554, + 340 + ], + "type": "text", + "content": "Gemini 2.0 Flash Thinking [19], Gemini 2.0 Pro [20], Gemini 2.0 Flash [18], InternVL2.5-78B [16], Grok 2 Vision [61], Pixtral Large [2], Qwen2.5-VL-72B [7], QVQ72B-Preview [55], Qwen-Omni-Turbo [57], Llama 4 Maverick [4], Gemma 3 27B [17], and Mistral Small 3.1 [3]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 346, + 432, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 432, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 432, + 360 + ], + "type": "text", + "content": "4.2. Prompting Methods" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 364, + 554, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 364, + 554, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 364, + 554, + 425 + ], + "type": "text", + "content": "Our evaluation adopts CoT [60], PoT [13] and IO (without any external prompts) prompting methods. Due to budget constraints, we report OpenAI o1 performance on the Hard subset only. Detailed prompts can be found in the Appendix." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 434, + 500, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 500, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 500, + 445 + ], + "type": "text", + "content": "4.3. Answer Extraction and Evaluation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 451, + 554, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 554, + 535 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 554, + 535 + ], + "type": "text", + "content": "Following Zhao et al. [67], we extract answers based on the prompting methods. For CoT and IO outputs, we employ GPT-4o-mini to extract numerical answers. For PoT, we run the generated program for numerical results. Finally, we conduct a strict accuracy evaluation, comparing numerical results with ground truth and deeming the results accurate within a stringent error tolerance of " + }, + { + "bbox": [ + 313, + 451, + 554, + 535 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 313, + 451, + 554, + 535 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 547, + 395, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 547, + 395, + 560 + ], + "spans": [ + { + "bbox": [ + 313, + 547, + 395, + 560 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 567, + 554, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 554, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 554, + 663 + ], + "type": "text", + "content": "We answer the following research questions (RQs): RQ1: Are MLLMs multimodal reasoners with extended thinking and PoT prompts? RQ2: What are the primary challenges facing MLLMs? RQ3: How can the visual perception difficulties of MLLMs be mitigated? RQ4: How can the knowledge reasoning capabilities of MLLMs be enhanced? RQ5: How can the numerical computation abilities of MLLMs be compensated for?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 671, + 432, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 432, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 432, + 685 + ], + "type": "text", + "content": "5.1. Main Results (RQ1)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "text", + "content": "The performance of the MLLMs evaluated using different prompting methods on FinMMR is shown in Tab. 2." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3249" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 70, + 555, + 277 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 555, + 277 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 555, + 277 + ], + "type": "table", + "html": "
ModelSizeExtended thinkingHardMediumEasyAvg.Token (M)
IOCoTPoTCoTPoTCoTPoTCoTPoTCoTPoT
Proprietary MLLMs
Claude 3.7 Sonnet✓ (64K)53.0051.0051.4062.5062.1778.5078.5064.0064.028.5111.25
Claude 3.7 SonnetX49.8050.8048.5062.2558.8377.0076.9263.3561.420.990.89
OpenAI o1✓48.0048.4044.70------2.522.12
GPT-4oX-45.4047.8063.3359.9278.0076.0062.2461.240.850.41
Gemini 2.0 ProX-46.5047.3060.5857.9275.5075.6760.8660.300.850.45
Gemini 2.0 Flash Thinking✓-46.0046.0060.7556.5877.1774.1761.3158.921.300.48
Gemini 2.0 FlashX-44.4045.9057.8353.4274.9273.7559.0557.690.790.43
Grok 2 VisionX-27.8025.5041.5035.8373.0872.8347.4644.721.130.60
Qwen-Omni-TurboX-17.5027.3035.8348.0057.5061.6736.9445.660.900.42
Open-source MLLMs
Llama 4 Maverick17BX-48.7047.8063.2559.1777.8377.8363.2661.600.880.47
Qwen2.5-VL-72B72BX-43.3046.2063.4264.1777.4275.8361.3862.071.050.44
InternVL2.5-78B78BX-37.4044.0060.5061.1770.9270.5856.2758.58--
QVQ-72B-Preview72B✓43.3040.306.2055.679.6775.4212.4257.139.435.435.70
Pixtral Large124BX-19.7025.0039.8339.7570.0070.1743.1844.971.150.75
Gemma 3 27B27BX-23.4022.3045.1736.4269.0861.5845.8840.100.970.47
Mistral Small 3.124BX-19.7015.2038.4229.7567.6749.4241.9331.461.150.60
", + "image_path": "a4ae46a275dd9df3d6778cf1390f937dddc281f452ae7eeaec7c3f21f12728ac.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 286, + 555, + 321 + ], + "lines": [ + { + "bbox": [ + 55, + 286, + 555, + 321 + ], + "spans": [ + { + "bbox": [ + 55, + 286, + 555, + 321 + ], + "type": "text", + "content": "Table 2. Results of different models using IO, CoT and PoT prompting methods on the test set of FinMMR. We use the best Accuracy on the Hard subset as the ranking indicator of model performance. The results underscore the superior performance of reasoning-enhanced MLLM (i.e., Claude 3.7 Sonnet with 64K extended thinking) with PoT in complex multimodal numerical reasoning task." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 340, + 294, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 340, + 294, + 532 + ], + "spans": [ + { + "bbox": [ + 54, + 340, + 294, + 532 + ], + "type": "text", + "content": "Challenges of MLLMs in Domain-Specific Complex Numerical Reasoning As the difficulty increases, the average accuracy shows a continuous and significant decline. In CoT setting, the average accuracy rates on the Easy, Medium, and Hard subsets are " + }, + { + "bbox": [ + 54, + 340, + 294, + 532 + ], + "type": "inline_equation", + "content": "73.33\\%" + }, + { + "bbox": [ + 54, + 340, + 294, + 532 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 340, + 294, + 532 + ], + "type": "inline_equation", + "content": "54.05\\%" + }, + { + "bbox": [ + 54, + 340, + 294, + 532 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 340, + 294, + 532 + ], + "type": "inline_equation", + "content": "38.14\\%" + }, + { + "bbox": [ + 54, + 340, + 294, + 532 + ], + "type": "text", + "content": ", respectively. The currently best-performing model (i.e., Claude 3.7 Sonnet with 64K extended thinking) consistently demonstrates superior performance across all difficulty subsets using CoT prompting method. However, its accuracy on the Hard subset remains below the " + }, + { + "bbox": [ + 54, + 340, + 294, + 532 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 54, + 340, + 294, + 532 + ], + "type": "text", + "content": " passing threshold in both prompting methods. On the overall test set, Claude 3.7 Sonnet achieves only " + }, + { + "bbox": [ + 54, + 340, + 294, + 532 + ], + "type": "inline_equation", + "content": "64.02\\%" + }, + { + "bbox": [ + 54, + 340, + 294, + 532 + ], + "type": "text", + "content": " accuracy. These results highlight the challenging nature and rigorous standards of FinMMR, effectively assessing the limits of MLLMs' reasoning capabilities and the disparities among models compared to previous benchmarks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 533, + 294, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 533, + 294, + 640 + ], + "spans": [ + { + "bbox": [ + 54, + 533, + 294, + 640 + ], + "type": "text", + "content": "Does extended thinking help? Reasoning-enhanced models show consistent improvements, compared with non-reasoning-enhanced MLLMs. Claude 3.7 Sonnet with 64K extended thinking achieves 2.20 percentage points higher accuracy than the model without extended thinking (53.00% vs. 50.80%) on the Hard subset. This enhancement comes at the cost of using nearly 12 times more tokens for intricate thinking (4.06M vs. 0.34M). This trend also persists in the Gemini 2.0 Flash series." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 642, + 294, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 642, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 642, + 294, + 714 + ], + "type": "text", + "content": "We observe that QVQ-72B-Preview lose basic code generation capabilities due to the reinforcement learning of text-based long thought, which is likely attributed to biases in training strategies and training data. On the Hard subset, this model achieves a code execution success rate of only " + }, + { + "bbox": [ + 54, + 642, + 294, + 714 + ], + "type": "inline_equation", + "content": "10.90\\%" + }, + { + "bbox": [ + 54, + 642, + 294, + 714 + ], + "type": "text", + "content": ", resulting in an accuracy of merely " + }, + { + "bbox": [ + 54, + 642, + 294, + 714 + ], + "type": "inline_equation", + "content": "6.20\\%" + }, + { + "bbox": [ + 54, + 642, + 294, + 714 + ], + "type": "text", + "content": " in" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 340, + 553, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 553, + 399 + ], + "type": "text", + "content": "PoT setting, significantly lower than the " + }, + { + "bbox": [ + 313, + 340, + 553, + 399 + ], + "type": "inline_equation", + "content": "40.30\\%" + }, + { + "bbox": [ + 313, + 340, + 553, + 399 + ], + "type": "text", + "content": " accuracy achieved with CoT. This finding highlights the importance of maintaining foundational capabilities, such as programming, while enhancing the reasoning abilities of MLLMs, to avoid rendering them ineffective in other basic tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 401, + 553, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 553, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 553, + 640 + ], + "type": "text", + "content": "Does PoT help? Experimental results strongly validate the superiority of PoT over CoT in numerical reasoning tasks, especially on the Hard subset. PoT (not including QVQ-72B-Preview) achieves a mean accuracy of " + }, + { + "bbox": [ + 313, + 401, + 553, + 640 + ], + "type": "inline_equation", + "content": "37.64\\%" + }, + { + "bbox": [ + 313, + 401, + 553, + 640 + ], + "type": "text", + "content": " versus " + }, + { + "bbox": [ + 313, + 401, + 553, + 640 + ], + "type": "inline_equation", + "content": "36.20\\%" + }, + { + "bbox": [ + 313, + 401, + 553, + 640 + ], + "type": "text", + "content": " for CoT, representing an improvement of 1.44 percentage points. Furthermore, PoT encourages MLLMs to leverage structured code generation to reduce token consumption during reasoning. Under similar or lower token usage, PoT achieves similar or greater accuracy. For instance, GPT-4o achieves a 2.40 percentage points improvement in accuracy over CoT while consuming significantly fewer tokens in PoT setting. Similarly, Qwen2.5-VL-72B demonstrates the most pronounced efficiency gains: PoT improves accuracy to " + }, + { + "bbox": [ + 313, + 401, + 553, + 640 + ], + "type": "inline_equation", + "content": "64.17\\%" + }, + { + "bbox": [ + 313, + 401, + 553, + 640 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 401, + 553, + 640 + ], + "type": "inline_equation", + "content": "63.42\\%" + }, + { + "bbox": [ + 313, + 401, + 553, + 640 + ], + "type": "text", + "content": " while reducing token consumption by " + }, + { + "bbox": [ + 313, + 401, + 553, + 640 + ], + "type": "inline_equation", + "content": "58.88\\%" + }, + { + "bbox": [ + 313, + 401, + 553, + 640 + ], + "type": "text", + "content": " (153K vs. 373K) on the Medium subset. When addressing complex numerical reasoning problems, PoT avoids precise numerical calculations by utilizing external tools (i.e., Python interpreter) and reduces the need for repetitive text-based reasoning, which is beneficial for most MLLMs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "content": "However, we also observe that for certain reasoning-enhanced models (i.e., Claude 3.7 Sonnet with 64K extended thinking and OpenAI o1), due to the enforced requirement for long thought, they still engage in extensive text-based reasoning before generating code even in PoT setting, resulting in exceptionally high token consumption" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3250" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 555, + 150 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 555, + 150 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 555, + 150 + ], + "type": "table", + "html": "
SubsetTestDegradationValidationDegradation
Ground Images (%)Distractor Images (%)Ground Images (%)Distractor Images (%)
Hard57.1847.23↓9.9556.7445.58↓11.16
Medium73.0161.36↓11.6577.7864.73↓12.35
Easy61.5953.64↓7.9560.6151.52↓9.09
The improvement achieved by the filtering-reasoning pipeline on the Medium validation set: 64.73 → 71.56 ↑6.83
", + "image_path": "70fcafdd75c30a4b18a32ddcf90e0832d04c8c27aaff09172eecffd3558cac18.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 158, + 555, + 180 + ], + "lines": [ + { + "bbox": [ + 55, + 158, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 158, + 555, + 180 + ], + "type": "text", + "content": "Table 3. Degradation of Qwen2.5-VL-72B on all subsets due to distractor images and improvement achieved by the filtering-reasoning pipeline on the Medium validation set in PoT setting." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 201, + 294, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 201, + 294, + 357 + ], + "spans": [ + { + "bbox": [ + 54, + 201, + 294, + 357 + ], + "type": "text", + "content": "on the Hard subset (4.48M and 2.12M), which is more than 10 times that of other MLLMs. To further investigate this, we added an IO baseline without any external prompts for reasoning-enhanced models on the Hard subset. The IO group achieved the highest accuracy, which we attribute to the comprehensive built-in system prompts embedded in the tested proprietary models. This highlights the need for future research to balance reasoning performance with the control of inefficient and redundant token generation, aiming to achieve a good trade-off between performance and cost, as well as to investigate whether PoT prompting can yield significant performance gains on open-source reasoning-enhanced models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 366, + 180, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 180, + 379 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 180, + 379 + ], + "type": "text", + "content": "5.2. Error Analysis (RQ2)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 384, + 295, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 295, + 468 + ], + "type": "text", + "content": "To better analyze the capabilities and limitations of MLLMs on FinMMR, we conduct a detailed error analysis for the Claude 3.7 Sonnet with 64K extended thinking in PoT setting. Error analysis is based on 100 sampled failure cases, which we categorize into three main error types, some of which involve compound errors. More details of error cases are provided in the Appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 471, + 295, + 638 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 55, + 471, + 295, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 471, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 295, + 529 + ], + "type": "text", + "content": "- Visual Perception Error (30/100): The model incorrectly perceives, identifies, or interprets visual information from images, or mistakenly recognizes incorrect data, subsequently causing errors in calculations, broken reasoning chains, or incorrect conclusions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 530, + 295, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 530, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 530, + 295, + 578 + ], + "type": "text", + "content": "- Knowledge Reasoning Error (38/100): Due to insufficient domain-specific knowledge, the model exhibits logical confusion or conceptual misunderstandings during reasoning, leading to incorrect answers." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 578, + 295, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 295, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 295, + 638 + ], + "type": "text", + "content": "- Numerical Computation Error (32/100): In problems involving mathematical operations and numerical reasoning, the model produces significant deviations from the correct answers due to errors in the calculation steps, precision control, or numerical hallucination." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 647, + 251, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 251, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 251, + 661 + ], + "type": "text", + "content": "5.3. Visual Filtering for Reasoning (RQ3)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": "As shown in Tab. 3, when processing multi-image inputs containing distractor images, Qwen2.5-VL-72B demonstrates significantly lower accuracy across all difficulty levels compared to ground images scenarios. This finding" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 201, + 553, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 201, + 553, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 553, + 332 + ], + "type": "text", + "content": "aligns with conclusions from previous work [37, 51], indicating that irrelevant visual information substantially interferes with MLLMs' reasoning capabilities. In particular, the Medium validation set exhibits the most pronounced performance drop (77.78% ground images vs. 64.73% distractor images), attributed to two key characteristics: (1) moderate complexity making visual perception quality the performance bottleneck; (2) semantic relevance between distractors and questions increasing visual filtering difficulty. To address distractor image interference, we propose a two-stage multimodal reasoning pipeline:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 334, + 553, + 430 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 314, + 334, + 553, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 334, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 314, + 334, + 553, + 380 + ], + "type": "text", + "content": "- Visual Filtering: We first instruct MLLM to analyze the set of images " + }, + { + "bbox": [ + 314, + 334, + 553, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 314, + 334, + 553, + 380 + ], + "type": "text", + "content": " and the question " + }, + { + "bbox": [ + 314, + 334, + 553, + 380 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 314, + 334, + 553, + 380 + ], + "type": "text", + "content": ", assessing the relevance of the image (relevant/irrelevant). Irrelevant images are excluded from subsequent reasoning." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 381, + 553, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 381, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 314, + 381, + 553, + 430 + ], + "type": "text", + "content": "- Enhanced Reasoning: Then, the filtered subset " + }, + { + "bbox": [ + 314, + 381, + 553, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{I}'" + }, + { + "bbox": [ + 314, + 381, + 553, + 430 + ], + "type": "text", + "content": " and the question " + }, + { + "bbox": [ + 314, + 381, + 553, + 430 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 314, + 381, + 553, + 430 + ], + "type": "text", + "content": " are input into the MLLM for the final reasoning. The system automatically reverts to the original inputs " + }, + { + "bbox": [ + 314, + 381, + 553, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 314, + 381, + 553, + 430 + ], + "type": "text", + "content": " if all images are mistakenly filtered." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 430, + 555, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 555, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 555, + 598 + ], + "type": "text", + "content": "Does the pipeline help? As illustrated in Tab. 3, we evaluate Qwen2.5-VL-72B on the 207 English questions with distractor images of the Medium validation set. Our method achieves an overall accuracy of " + }, + { + "bbox": [ + 313, + 430, + 555, + 598 + ], + "type": "inline_equation", + "content": "71.56\\%" + }, + { + "bbox": [ + 313, + 430, + 555, + 598 + ], + "type": "text", + "content": ", representing a 6.83 percentage points improvement over direct reasoning. This result is only 6.22 percentage points away from the ideal accuracy of the ground images scenarios " + }, + { + "bbox": [ + 313, + 430, + 555, + 598 + ], + "type": "inline_equation", + "content": "(77.78\\%)" + }, + { + "bbox": [ + 313, + 430, + 555, + 598 + ], + "type": "text", + "content": ". Detailed analysis reveals " + }, + { + "bbox": [ + 313, + 430, + 555, + 598 + ], + "type": "inline_equation", + "content": "73.43\\%" + }, + { + "bbox": [ + 313, + 430, + 555, + 598 + ], + "type": "text", + "content": " (152/207) ground image recognition accuracy during filtering. When correctly identified, the accuracy of these problems increases to " + }, + { + "bbox": [ + 313, + 430, + 555, + 598 + ], + "type": "inline_equation", + "content": "81.58\\%" + }, + { + "bbox": [ + 313, + 430, + 555, + 598 + ], + "type": "text", + "content": " (124/152). This finding underscores the necessity to enhance the ability of MLLMs to filter out irrelevant image information, thereby strengthening their robustness in reasoning within more complex real-world scenarios." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 608, + 491, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 608, + 491, + 620 + ], + "spans": [ + { + "bbox": [ + 313, + 608, + 491, + 620 + ], + "type": "text", + "content": "5.4. Knowledge Augmentation (RQ4)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 626, + 553, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 626, + 553, + 674 + ], + "spans": [ + { + "bbox": [ + 313, + 626, + 553, + 674 + ], + "type": "text", + "content": "To enhance the understanding and application capabilities of financial knowledge of MLLMs, we explore a method of refined knowledge augmentation to improve the performance of MLLM in domain-specific reasoning tasks." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "- Function Library Construction: We annotate a comprehensive financial function library containing 3,133 Python functions from financial encyclopedia. Each func" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3251" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 293, + 145 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 293, + 145 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 293, + 145 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 293, + 145 + ], + "type": "table", + "html": "
SettingPoTRAG with PoT
Gemini 2.0 Flash Thinking78.7183.02 (+4.31)
GPT-4o80.6083.62 (+3.02)
Claude 3.7 Sonnet81.2185.43 (+4.22)
Claude 3.7 Sonnet (64K)83.5386.29 (+2.76)
", + "image_path": "bcbd361d99c7192321ef46924bd2a5f196041c77b4b8b684e226f6ca454c3bad.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 152, + 295, + 177 + ], + "lines": [ + { + "bbox": [ + 55, + 152, + 295, + 177 + ], + "spans": [ + { + "bbox": [ + 55, + 152, + 295, + 177 + ], + "type": "text", + "content": "Table 4. Improvements of different MLLMs with knowledge augmentation on the 1,160 instances of FinMMR in PoT setting." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 63, + 196, + 295, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 196, + 295, + 220 + ], + "spans": [ + { + "bbox": [ + 63, + 196, + 295, + 220 + ], + "type": "text", + "content": "tion includes precise functional descriptions, parameter explanations, and step-by-step implementation code." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 220, + 296, + 425 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 220, + 295, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 220, + 295, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 295, + 352 + ], + "type": "text", + "content": "- MLLM-Instructed Knowledge Retrieval: In financial problems with hybrid contexts, using short questions or full contexts for retrieval often fails to retrieve directly relevant knowledge [9, 49]. We observed that powerful MLLMs (e.g., GPT-40) can effectively summarize rich semantic information from contexts. Therefore, we first prompt the MLLM to generate precise retrieval queries based on the context [34, 58]. Then we use Contriever [28] to retrieve relevant financial Python functions, based on the semantic similarity between the refined queries and functional descriptions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 352, + 296, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 352, + 296, + 425 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 296, + 425 + ], + "type": "text", + "content": "- MLLM as Retrieval Judge: Recent studies have shown that models are capable of judging the relevance of candidates retrieved for the question [26]. In this setting, we first retrieved the Top-30 financial functions and then prompted the MLLM to select the Top-3 functions most useful to answer the question, if any." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 54, + 426, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 426, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 54, + 426, + 295, + 617 + ], + "type": "text", + "content": "Does knowledge augmentation help? As shown in Tab. 4, all evaluated MLLMs enhanced with financial function knowledge achieved significant improvements, ranging from 2.76 to 4.31 percentage points. Leveraging the improved retrieval efficiency enabled by MLLM-Instructed Knowledge Retrieval and MLLM as Retrieval Judge, the knowledge augmentation approach achieves greater performance, boosting the accuracy of Claude 3.7 Sonnet with 64K thinking to " + }, + { + "bbox": [ + 54, + 426, + 295, + 617 + ], + "type": "inline_equation", + "content": "86.29\\%" + }, + { + "bbox": [ + 54, + 426, + 295, + 617 + ], + "type": "text", + "content": ". Notably, Gemini 2.0 Flash Thinking, which has relatively weaker reasoning capabilities, also improved from " + }, + { + "bbox": [ + 54, + 426, + 295, + 617 + ], + "type": "inline_equation", + "content": "78.71\\%" + }, + { + "bbox": [ + 54, + 426, + 295, + 617 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 54, + 426, + 295, + 617 + ], + "type": "inline_equation", + "content": "83.02\\%" + }, + { + "bbox": [ + 54, + 426, + 295, + 617 + ], + "type": "text", + "content": ", approaching the performance of Claude 3.7 Sonnet " + }, + { + "bbox": [ + 54, + 426, + 295, + 617 + ], + "type": "inline_equation", + "content": "(83.53\\%)" + }, + { + "bbox": [ + 54, + 426, + 295, + 617 + ], + "type": "text", + "content": " without knowledge augmentation. The results further illustrate that refined domain-specific reasoning knowledge can significantly enhance the performance of MLLMs in complex reasoning tasks within expert domains." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 624, + 243, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 243, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 243, + 637 + ], + "type": "text", + "content": "5.5. Visual Parser with Reasoner (RQ5)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "content": "In complex multimodal numerical reasoning tasks, single models often struggle to simultaneously achieve visual perception, knowledge reasoning, and numerical computation. To investigate the potential of model collaboration, we instruct the MLLM to act as the Visual Parser, responsible for carefully converting images into structured textual data." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 315, + 70, + 554, + 188 + ], + "blocks": [ + { + "bbox": [ + 315, + 70, + 554, + 188 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 554, + 188 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 554, + 188 + ], + "type": "image", + "image_path": "7d6557cfa9525f56834045e280e4a6b9d3f2d22c08400430bd46e0c9f475df53.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 196, + 550, + 208 + ], + "lines": [ + { + "bbox": [ + 317, + 196, + 550, + 208 + ], + "spans": [ + { + "bbox": [ + 317, + 196, + 550, + 208 + ], + "type": "text", + "content": "Figure 4. Results of model combinations and individual models." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 228, + 553, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 553, + 252 + ], + "type": "text", + "content": "Then, an LRM acts as the Reasoner, performing multi-step numerical reasoning based on the textual context." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 253, + 553, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 553, + 336 + ], + "type": "text", + "content": "Specifically, we filter 1,160 tabular QA instances from FinMMR and utilize GPT-4o as the Visual Parser, instructing it to separate table headers or cells with vertical bars (|) and rows with newlines. For the Reasoner component, in addition to GPT-4o, we evaluate several LRMs (i.e., Claude 3.7 Sonnet, Gemini 2.0 Flash Thinking, DeepSeek-R1, and OpenAI o1)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 336, + 554, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 554, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 554, + 503 + ], + "type": "text", + "content": "Does model collaboration help? As shown in Fig. 4, the structured data from GPT-4o's visual parsing significantly enhances downstream reasoning. The individual model (i.e., GPT-4o with PoT) achieves an accuracy of " + }, + { + "bbox": [ + 313, + 336, + 554, + 503 + ], + "type": "inline_equation", + "content": "80.60\\%" + }, + { + "bbox": [ + 313, + 336, + 554, + 503 + ], + "type": "text", + "content": ", while the combination of models improves the accuracy to " + }, + { + "bbox": [ + 313, + 336, + 554, + 503 + ], + "type": "inline_equation", + "content": "86.72\\%" + }, + { + "bbox": [ + 313, + 336, + 554, + 503 + ], + "type": "text", + "content": " (i.e., DeepSeek-R1 serving as the Reasoner with PoT). Performance variance emerges across reasoning-enhanced models using identical visual inputs. Claude 3.7 Sonnet reaches " + }, + { + "bbox": [ + 313, + 336, + 554, + 503 + ], + "type": "inline_equation", + "content": "85.52\\%" + }, + { + "bbox": [ + 313, + 336, + 554, + 503 + ], + "type": "text", + "content": ", outperforming Gemini 2.0 Flash Thinking " + }, + { + "bbox": [ + 313, + 336, + 554, + 503 + ], + "type": "inline_equation", + "content": "(82.24\\%)" + }, + { + "bbox": [ + 313, + 336, + 554, + 503 + ], + "type": "text", + "content": ", confirming the decisive impact of text-based reasoning capabilities. This evidences that model collaboration effectively compensates for individual model limitations through complementary strengths." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 514, + 388, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 514, + 388, + 526 + ], + "spans": [ + { + "bbox": [ + 314, + 514, + 388, + 526 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "We introduce FinMMR, a multimodal, comprehensive, and challenging benchmark for evaluating the financial numerical reasoning capabilities of MLLMs. FinMMR challenges MLLMs' intricate visual perception, specialized knowledge reasoning, and accurate numerical computation through its rich images, comprehensive subdomains, and complex formulas embedded in each multimodal financial question. The evaluation results reveal that 15 proprietary and open-source MLLMs still struggle with complex multimodal reasoning tasks in specialized domains. FinMMR highlights the bottlenecks of MLLMs and the need for continuous improvements, including reasoning-enhanced training, tool use, refined structured knowledge augmentation and model combinations, allowing models to perform expert-level reasoning tasks closer to the real-world scenarios." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3252" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 297, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 212 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 212 + ], + "type": "text", + "content": "This work is supported by the National Natural Science Foundation of China (Grant Nos. 62176026, 62473271), the Beijing Natural Science Foundation (Grant No. QY24214), the BUPT Innovation and Entrepreneurship Support Program (Grant Nos. 2025-YC-A033, 2025-YC-A042), and data support from Hithink RoyalFlush Information Network Co., Ltd. This work is also supported by the Engineering Research Center of Information Networks, Ministry of Education, China. We would also like to thank the anonymous reviewers and area chairs for constructive discussions and feedback." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 234, + 115, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 234, + 115, + 247 + ], + "spans": [ + { + "bbox": [ + 56, + 234, + 115, + 247 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 254, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 254, + 294, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 254, + 294, + 286 + ], + "spans": [ + { + "bbox": [ + 61, + 254, + 294, + 286 + ], + "type": "text", + "content": "[1] 360AILAB-NLP. 360layoutanalysis. https://github.com/360AILAB-NLP/360LayoutAnalysis, 2024. Accessed: 2024-06-15. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 288, + 294, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 288, + 294, + 320 + ], + "spans": [ + { + "bbox": [ + 61, + 288, + 294, + 320 + ], + "type": "text", + "content": "[2] Mistral AI. Pixtral large. https://mistral.ai/news/pixtral-large, 2024. Accessed: 2024-11-18. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 323, + 294, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 323, + 294, + 354 + ], + "spans": [ + { + "bbox": [ + 62, + 323, + 294, + 354 + ], + "type": "text", + "content": "[3] Mistral AI. Mistral small 3.1. https://mistral.ai/news/mistral-small-3-1, 2025. Accessed: 2025-03-17. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 357, + 294, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 357, + 294, + 399 + ], + "spans": [ + { + "bbox": [ + 62, + 357, + 294, + 399 + ], + "type": "text", + "content": "[4] AI@Meta. The llama 4 herd: The beginning of a new era of natively multimodal ai innovation. https://ai.meta.com/blog/llama-4-multimodal-intelligence, 2025. Accessed: 2025-04-05. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 402, + 294, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 402, + 294, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 402, + 294, + 434 + ], + "type": "text", + "content": "[5] Anthropic. Claude 3.7 sonnet and claude code. https://www.anthropic.com/news/claude-3-7-sonnet, 2025. Accessed: 2025-02-25. 1, 3, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 436, + 294, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 294, + 490 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 294, + 490 + ], + "type": "text", + "content": "[6] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint, 2023. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 492, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 492, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 62, + 492, + 294, + 568 + ], + "type": "text", + "content": "[7] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zhenen Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Yang, Haiyang Xu, and Junyang Lin. Qwen2.5-vl technical report. arXiv preprint, 2025. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 570, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 570, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 62, + 570, + 294, + 635 + ], + "type": "text", + "content": "[8] Sébastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, Harsha Nori, Hamid Palangi, Marco Tulio Ribeiro, and Yi Zhang. Sparks of artificial general intelligence: Early experiments with gpt-4. arXiv preprint, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 637, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 637, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 637, + 294, + 713 + ], + "type": "text", + "content": "[9] Liang Chen, Yang Deng, Yatao Bian, Zeyu Qin, Bingzhe Wu, Tat-Seng Chua, and Kam-Fai Wong. Beyond factuality: A comprehensive evaluation of large language models as knowledge generators. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 6325-6341, Singapore, 2023. Association for Computational Linguistics. 8" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 138 + ], + "type": "text", + "content": "[10] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, and Feng Zhao. Are we on the right way for evaluating large vision-language models? In Advances in Neural Information Processing Systems, pages 27056-27087. Curran Associates, Inc., 2024. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 140, + 553, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 337 + ], + "type": "text", + "content": "[11] Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott Gray, Nick Ryder, Mikhail Pavlov, Alethea Power, Lukasz Kaiser, Mohammad Bavarian, Clemens Winter, Philippe Tillet, Felipe Petroski Such, Dave Cummings, Matthias Plappert, Fotios Chantzis, Elizabeth Barnes, Ariel Herbert-Voss, William Hebgen Guss, Alex Nichol, Alex Paino, Nikolas Tezak, Jie Tang, Igor Babuschkin, Suchir Balaji, Shantanu Jain, William Saunders, Christopher Hesse, Andrew N. Carr, Jan Leike, Josh Achiam, Vedant Misra, Evan Morikawa, Alec Radford, Matthew Knight, Miles Brundage, Mira Murati, Katie Mayer, Peter Welinder, Bob McGrew, Dario Amodei, Sam McCandlish, Ilya Sutskever, and Wojciech Zaremba. Evaluating large language models trained on code. arXiv preprint, 2021. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 338, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 338, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 317, + 338, + 553, + 425 + ], + "type": "text", + "content": "[12] Pengcheng Chen, Jin Ye, Guoan Wang, Yanjun Li, Zhongying Deng, Wei Li, Tianbin Li, Haodong Duan, Ziyan Huang, Yanzhou Su, Benyou Wang, Shaoting Zhang, Bin Fu, Jianfei Cai, Bohan Zhuang, Eric J Seibel, Yu Qiao, and Junjun He. Gmai-mmbench: A comprehensive multimodal evaluation benchmark towards general medical ai. In Advances in Neural Information Processing Systems, pages 94327-94427. Curran Associates, Inc., 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 426, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 426, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 317, + 426, + 553, + 470 + ], + "type": "text", + "content": "[13] Wenhu Chen, Xueguang Ma, Xinyi Wang, and William W. Cohen. Program of thoughts prompting: Disentangling computation from reasoning for numerical reasoning tasks. Transactions on Machine Learning Research, 2023. 3, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 472, + 553, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 553, + 559 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 553, + 559 + ], + "type": "text", + "content": "[14] Zhiyu Chen, Wenhu Chen, Charese Smiley, Sameena Shah, Iana Borova, Dylan Langdon, Reema Moussa, Matt Beane, Ting-Hao Huang, Bryan Routledge, and William Yang Wang. FinQA: A dataset of numerical reasoning over financial data. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3697-3711, Online and Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 2, 3, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 559, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 553, + 635 + ], + "type": "text", + "content": "[15] Zhiyu Chen, Shiyang Li, Charese Smiley, Zhiqiang Ma, Sameena Shah, and William Yang Wang. ConvFinQA: Exploring the chain of numerical reasoning in conversational finance question answering. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 6279-6292, Abu Dhabi, United Arab Emirates, 2022. Association for Computational Linguistics. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 637, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 637, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 637, + 553, + 713 + ], + "type": "text", + "content": "[16] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, Lixin Gu, Xuehui Wang, Qingyun Li, Yimin Ren, Zixuan Chen, Jiapeng Luo, Jiahao Wang, Tan Jiang, Bo Wang, Conghui He, Botian Shi, Xingcheng Zhang, Han Lv, Yi Wang, Wenqi Shao, Pei Chu, Zhongying Tu, Tong He, Zhiyong Wu, Huipeng Deng, Jiaye Ge, Kai Chen, Kaipeng" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3253" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 76, + 72, + 295, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 295, + 128 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 295, + 128 + ], + "type": "text", + "content": "Zhang, Limin Wang, Min Dou, Lewei Lu, Xizhou Zhu, Tong Lu, Dahua Lin, Yu Qiao, Jifeng Dai, and Wenhai Wang. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint, 2025. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "type": "text", + "content": "[17] Google DeepMind. Introducing gemma 3: The most capable model you can run on a singlegpu or tpu. https://blog.google/technology/developers/gemma-3, 2025. Accessed: 2025-03-12. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 173, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 294, + 217 + ], + "type": "text", + "content": "[18] Google DeepMind. Gemini 2.0: Flash, flash-lite and pro. https://developers.googleblog.com/en/gemini-2-family-expands, 2025. Accessed: 2025-02-05. 3, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 218, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 218, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 218, + 294, + 251 + ], + "type": "text", + "content": "[19] Google DeepMind. Gemini 2.0 flash thinking. https://ai.google.dev/gemini-api/docs/thinking, 2025. Accessed: 2025-02-05. 1, 3, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 251, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 294, + 285 + ], + "type": "text", + "content": "[20] Google DeepMind. Gemini 2.0. https://cloud.google.com/vertex-ai/generative-ai/docs/gemini-v2#2.0-pro, 2025. Accessed: 2025-02-05. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 285, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 285, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 295, + 714 + ], + "type": "text", + "content": "[21] DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, Aixin Liu, Bing Xue, Bingxuan Wang, Bochao Wu, Bei Feng, Chengda Lu, Chenggang Zhang, Chengqi Deng, Chenyu Zhang, Chong Ruan, Damai Dai, Deli Chen, Dongjie Ji, Erhang Li, Fangyun Lin, Fucong Dai, Fuli Luo, Guangbo Hao, Guanting Chen, Guowei Li, H. Zhang, Han Bao, Hanwei Xu, Haocheng Wang, Honghui Ding, Huajian Xin, Huazuo Gao, Hui Qu, Hui Li, Jianzhong Guo, Jiashi Li, Jiawei Wang, Jingchang Chen, Jingyang Yuan, Junjie Qiu, Junlong Li, J. L. Cai, Jiaqi Ni, Jian Liang, Jin Chen, Kai Dong, Kai Hu, Kaige Gao, Kang Guan, Kexin Huang, Kuai Yu, Lean Wang, Lecong Zhang, Liang Zhao, Litong Wang, Liyue Zhang, Lei Xu, Leyi Xia, Mingchuan Zhang, Minghua Zhang, Minghui Tang, Meng Li, Miaojun Wang, Mingming Li, Ning Tian, Panpan Huang, Peng Zhang, Qiancheng Wang, Qinyu Chen, Qiushi Du, Ruiqi Ge, Ruisong Zhang, Ruizhe Pan, Runji Wang, R. J. Chen, R. L. Jin, Ruyi Chen, Shanghao Lu, Shangyan Zhou, Shanhuang Chen, Shengfeng Ye, Shiyu Wang, Shuiping Yu, Shunfeng Zhou, Shuting Pan S. S. Li Shuang Zhou Shaoqing Wu Shengfeng Ye Tao Yun Tian Pei Tianyu Sun T. Wang Wangding Zeng Wanjia Zhao Wen Liu Wenfeng Liang Wenjun Gao Wenqin Yu Wentao Zhang W. L. Xiao Wei An Xiaodong Liu Xiaohan Wang Xiaokang Chen Xiaotao Nie Xin Cheng Xin Liu Xin Xie Xingchao Liu Xinyu Yang Xinyuan Li Xuecheng Su Xuheng Lin X.Q.Li Xiangyue Jin Xiaojin Shen Xiaosha Chen Xiaowen Sun Xiaoxiang Wang Xinnan Song Xinyi Zhou Xianzu Wang Xinxia Shan Y.K. Li Y.Q. WangY.X.Wei,Yang Zhang Yanhong Xu Yao Li Yao Zhao Yaofeng Sun Yaohui Wang Yi Yu Yichao Zhang Yifan Shi Yiliang Xiong Ying He Yishi Piao Yisong Wang Yixuan Tan Yiyang Ma Yiyuan Liu Yongqiang Guo Yuan OuYuduan Wang Yue Gong Yuheng Zou Yujia HeYunfan Xiong Yuxiang Luo Yuxiang You Yuxuan Liu Yuyang ZhouY.X.ZhuYanhong Xu Yanping Huang Yaohui Li Yi ZhengYuchen ZhuYunxian Ma Ying Tang Yukun" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 333, + 72, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 553, + 149 + ], + "type": "text", + "content": "Zha, Yuting Yan, Z. Z. Ren, Zehui Ren, Zhangli Sha, Zhe Fu, Zhean Xu, Zhenda Xie, Zhengyan Zhang, Zhenen Hao, Zhicheng Ma, Zhigang Yan, Zhiyu Wu, Zihui Gu, Zijia Zhu, Zijun Liu, Zilin Li, Ziwei Xie, Ziyang Song, Zizheng Pan, Zhen Huang, Zhipeng Xu, Zhongyu Zhang, and Zhen Zhang. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint, 2025. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 316, + 150, + 555, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 150, + 555, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 150, + 555, + 237 + ], + "type": "text", + "content": "[22] Chunyuan Deng, Yilun Zhao, Xiangru Tang, Mark Gerstein, and Arman Cohan. Investigating data contamination in modern benchmarks for large language models. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 8706-8719, Mexico City, Mexico, 2024. Association for Computational Linguistics. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 316, + 238, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 238, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 238, + 553, + 293 + ], + "type": "text", + "content": "[23] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, Yunsheng Wu, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 316, + 293, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 293, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 293, + 553, + 348 + ], + "type": "text", + "content": "[24] Ziliang Gan, Yu Lu, Dong Zhang, Haohan Li, Che Liu, Jian Liu, Ji Liu, Haipang Wu, Chaoyou Fu, Zenglin Xu, Rongjunchen Zhang, and Yong Dai. Mme-finance: A multimodal finance benchmark for expert-level understanding and reasoning. arXiv preprint, 2024. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 316, + 349, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 349, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 349, + 553, + 403 + ], + "type": "text", + "content": "[25] Yingqiang Ge, Wenyue Hua, Kai Mei, jianchao ji, Juntao Tan, Shuyuan Xu, Zelong Li, and Yongfeng Zhang. Opennagi: When llm meets domain experts. In Advances in Neural Information Processing Systems, pages 5539-5568. Curran Associates, Inc., 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 404, + 553, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 458 + ], + "type": "text", + "content": "[26] Jian Guan, Wei Wu, Zujie Wen, Peng Xu, Hongning Wang, and Minlie Huang. Amor: A recipe for building adaptable modular knowledge agents through process feedback. In Advances in Neural Information Processing Systems, pages 126118-126148. Curran Associates, Inc., 2024. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 460, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 514 + ], + "type": "text", + "content": "[27] Yunzhuo Hao, Jiawei Gu, Huichen Will Wang, Linjie Li, Zhengyuan Yang, Lijuan Wang, and Yu Cheng. Can MLLMs reason in multimodality? EMMA: An enhanced multimodal reasoning benchmark. In *Forty-second International Conference on Machine Learning*, 2025. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 515, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 553, + 568 + ], + "type": "text", + "content": "[28] Gautier Izacard, Mathilde Caron, Lucas Hosseini, Sebastian Riedel, Piotr Bojanowski, Armand Joulin, and Edouard Grave. Unsupervised dense information retrieval with contrastive learning. Transactions on Machine Learning Research, 2022. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 570, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 570, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 570, + 553, + 635 + ], + "type": "text", + "content": "[29] Naman Jain, King Han, Alex Gu, Wen-Ding Li, Fanjia Yan, Tianjun Zhang, Sida Wang, Armando Solar-Lezama, Koushik Sen, and Ion Stoica. Livecodebench: Holistic and contamination free evaluation of large language models for code. In The Thirteenth International Conference on Learning Representations, 2025. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 636, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 713 + ], + "type": "text", + "content": "[30] Dongzhi Jiang, Renrui Zhang, Ziyu Guo, Yanwei Li, Yu Qi, Xinyan Chen, Liuhui Wang, Jianhan Jin, Claire Guo, Shen Yan, Bo Zhang, Chaoyou Fu, Peng Gao, and Hongsheng Li. MME-cot: Benchmarking chain-of-thought in large multimodal models for reasoning quality, robustness, and efficiency. In Forty-second International Conference on Machine Learning, 2025. 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3254" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[31] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint, 2020. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 118, + 294, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 118, + 294, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 118, + 294, + 193 + ], + "type": "text", + "content": "[32] Michael Krumdick, Rik Koncel-Kedziorski, Viet Dac Lai, Varshini Reddy, Charles Lovering, and Chris Tanner. BizBench: A quantitative reasoning benchmark for business and finance. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8309-8332, Bangkok, Thailand, 2024. Association for Computational Linguistics. 2, 3, 4, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 195, + 295, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 195, + 295, + 249 + ], + "spans": [ + { + "bbox": [ + 57, + 195, + 295, + 249 + ], + "type": "text", + "content": "[33] Bohao Li, Yuying Ge, Yixiao Ge, Guangzhi Wang, Rui Wang, Ruimao Zhang, and Ying Shan. Seed-bench: Benchmarking multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13299-13308, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 251, + 295, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 251, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 57, + 251, + 295, + 293 + ], + "type": "text", + "content": "[34] Xiaoxi Li, Guanting Dong, Jiajie Jin, Yuyao Zhang, Yujia Zhou, Yutao Zhu, Peitian Zhang, and Zhicheng Dou. Searcho1: Agentic search-enhanced large reasoning models. arXiv preprint, 2025. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 295, + 294, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 295, + 294, + 348 + ], + "spans": [ + { + "bbox": [ + 57, + 295, + 294, + 348 + ], + "type": "text", + "content": "[35] Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let's verify step by step. In The Twelfth International Conference on Learning Representations, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 350, + 295, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 350, + 295, + 448 + ], + "spans": [ + { + "bbox": [ + 57, + 350, + 295, + 448 + ], + "type": "text", + "content": "[36] Fuxiao Liu, Xiaoyang Wang, Wenlin Yao, Jianshu Chen, Kaiqiang Song, Sangwoo Cho, Yaser Yacoob, and Dong Yu. MMC: Advancing multimodal chart understanding with large-scale instruction tuning. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 1287-1310, Mexico City, Mexico, 2024. Association for Computational Linguistics. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 449, + 295, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 449, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 57, + 449, + 295, + 525 + ], + "type": "text", + "content": "[37] Haowei Liu, Xi Zhang, Haiyang Xu, Yaya Shi, Chaoya Jiang, Ming Yan, Ji Zhang, Fei Huang, Chunfeng Yuan, Bing Li, and Weiming Hu. MIBench: Evaluating multimodal large language models over multiple images. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 22417-22428, Miami, Florida, USA, 2024. Association for Computational Linguistics. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 526, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 526, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 526, + 295, + 590 + ], + "type": "text", + "content": "[38] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin. Mmbench: Is your multi-modal model an all-around player? In Computer Vision - ECCV 2024, pages 216-233, Cham, 2025. Springer Nature Switzerland. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 592, + 295, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 592, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 57, + 592, + 295, + 657 + ], + "type": "text", + "content": "[39] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating math reasoning in visual contexts with gpt-4v, bard, and other large multimodal models. In The Twelfth International Conference on Learning Representations, 2024. 1, 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "type": "text", + "content": "[40] Junyu Luo, Zhizhuo Kou, Liming Yang, Xiao Luo, Jinsheng Huang, Zhiping Xiao, Jingshu Peng, Chengzhong Liu, Ji-aming Ji, Xuanzhe Liu, Sirui Han, Ming Zhang, and Yike Guo. FinMME: Benchmark dataset for financial multi-modal reasoning evaluation. In Proceedings of the 63rd Annual" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 29465-29489, Vienna, Austria, 2025. Association for Computational Linguistics. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 107, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 171 + ], + "type": "text", + "content": "[41] Yujun Mao, Yoon Kim, and Yilun Zhou. CHAMP: A competition-level dataset for fine-grained analyses of LLMs' mathematical reasoning capabilities. In Findings of the Association for Computational Linguistics: ACL 2024, pages 13256-13274, Bangkok, Thailand, 2024. Association for Computational Linguistics. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 173, + 553, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 553, + 216 + ], + "type": "text", + "content": "[42] Grégoire Mialon, Clémentine Fourrier, Thomas Wolf, Yann LeCun, and Thomas Scialom. GAIA: a benchmark for general AI assistants. In The Twelfth International Conference on Learning Representations, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 217, + 553, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 280 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 280 + ], + "type": "text", + "content": "[43] Meredith Ringel Morris, Jascha Sohl-Dickstein, Noah Fiedel, Tris Warkentin, Allan Dafoe, Aleksandra Faust, Clement Farabet, and Shane Legg. Position: Levels of AGI for operationalizing progress on the path to AGI. In *Forty-first International Conference on Machine Learning*, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 282, + 553, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 282, + 553, + 313 + ], + "spans": [ + { + "bbox": [ + 316, + 282, + 553, + 313 + ], + "type": "text", + "content": "[44] OpenAI. Hello gpt-4o. https://openai.com/index/hello-gpt-4o, 2024. Accessed: 2024-05-13. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 316, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 316, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 317, + 316, + 553, + 348 + ], + "type": "text", + "content": "[45] OpenAI. Learning to reason with llms. https://openai.com/index/learning-to-reason-with-llms, 2024. Accessed: 2024-09-12. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 350, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 381 + ], + "type": "text", + "content": "[46] OpenAI. Openai o1 system card. https://openai.com/index/openai-o1-system-card, 2024. Accessed: 2024-12-05. 1, 3, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 383, + 553, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 383, + 553, + 424 + ], + "spans": [ + { + "bbox": [ + 317, + 383, + 553, + 424 + ], + "type": "text", + "content": "[47] OpenAI. Openai o1-mini. https://openai.com/index/openai-ol-mini-advancing-cost-efficient-reasoning, 2024. Accessed: 2024-09-12. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 427, + 553, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 427, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 316, + 427, + 553, + 458 + ], + "type": "text", + "content": "[48] OpenAI. Openai o3-mini system card. https://openai.com/index/o3-mini-system-card, 2025. Accessed: 2025-01-31. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 460, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 553, + 514 + ], + "type": "text", + "content": "[49] Baolin Peng, Michel Galley, Pengcheng He, Hao Cheng, Yujia Xie, Yu Hu, Qiuyuan Huang, Lars Liden, Zhou Yu, Weizhu Chen, and Jianfeng Gao. Check your facts and try again: Improving large language models with external knowledge and automated feedback. arXiv preprint, 2023. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 516, + 553, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 516, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 317, + 516, + 553, + 591 + ], + "type": "text", + "content": "[50] Oscar Sainz, Jon Campos, Iker García-Ferrero, Julien Etxaniz, Oier Lopez de Lacalle, and Eneko Agirre. NLP evaluation in trouble: On the need to measure LLM data contamination for each benchmark. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 10776-10787, Singapore, 2023. Association for Computational Linguistics. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 593, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 593, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 317, + 593, + 553, + 658 + ], + "type": "text", + "content": "[51] Aditya Sharma, Michael Saxon, and William Yang Wang. Losing visual needles in image haystacks: Vision language models are easily distracted in short and long contexts. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 5429-5451, Miami, Florida, USA, 2024. Association for Computational Linguistics. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[52] Weijia Shi, Anirudh Ajith, Mengzhou Xia, Yangsibo Huang, Daogao Liu, Terra Blevins, Danqi Chen, and Luke Zettlemoyer. Detecting pretraining data from large language models. In The Twelfth International Conference on Learning Representations, 2024. 5" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3255" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 173 + ], + "type": "text", + "content": "[53] Zichen Tang, Haihong E, Ziyan Ma, Haoyang He, Jiacheng Liu, Zhongjun Yang, Zihua Rong, Rongjin Li, Kun Ji, Qing Huang, Xinyang Hu, Yang Liu, and Qianhe Zheng. FinanceReasoning: Benchmarking financial numerical reasoning more credible, comprehensive and challenging. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15721-15749, Vienna, Austria, 2025. Association for Computational Linguistics. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 174, + 297, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 297, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 297, + 437 + ], + "type": "text", + "content": "[54] Kimi Team, Angang Du, Bofei Gao, Bowei Xing, Changjiu Jiang, Cheng Chen, Cheng Li, Chenjun Xiao, Chenzhuang Du, Chonghua Liao, Chuning Tang, Congcong Wang, Dehao Zhang, Enming Yuan, Enzhe Lu, Fengxiang Tang, Flood Sung, Guangda Wei, Guokun Lai, Haiqing Guo, Han Zhu, Hao Ding, Hao Hu, Hao Yang, Hao Zhang, Haotian Yao, Haotian Zhao, Haoyu Lu, Haoze Li, Haozhen Yu, Hongcheng Gao, Huabin Zheng, Huan Yuan, Jia Chen, Jianhang Guo, Jianlin Su, Jianzhou Wang, Jie Zhao, Jin Zhang, Jingyuan Liu, Junjie Yan, Junyan Wu, Lidong Shi, Ling Ye, Longhui Yu, Mengnan Dong, Neo Zhang, Ningchen Ma, Qiwei Pan, Qucheng Gong, Shaowei Liu, Shengling Ma, Shupeng Wei, Sihan Cao, Siying Huang, Tao Jiang, Weihao Gao, Weimin Xiong, Weiran He, Weixiao Huang, Weixin Xu, Wenhao Wu, Wenyang He, Xianghui Wei, Xianqing Jia, Xingzhe Wu, Xinran Xu, Xinxing Zu, Xinyu Zhou, Xuehai Pan, Y. Charles, Yang Li, Yangyang Hu, Yangyang Liu, Yanru Chen, Yejie Wang, Yibo Liu, Yidao Qin, Yifeng Liu, Ying Yang, Yiping Bao, Yulun Du, Yuxin Wu, Yuzhi Wang, Zaida Zhou, Zhaoji Wang, Zhaowei Li, Zhen Zhu, Zheng Zhang, Zhexu Wang, Zhilin Yang, Zhiqi Huang, Zihao Huang, Ziyao Xu, Zonghan Yang, and Zongyu Lin. Kimi k1.5: Scaling reinforcement learning with llms. arXiv preprint, 2025. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 439, + 296, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 439, + 296, + 472 + ], + "spans": [ + { + "bbox": [ + 56, + 439, + 296, + 472 + ], + "type": "text", + "content": "[55] Qwen Team. Qvq: To see the world with wisdom. https://qwenlm.github.io/blog/qvq-72b-preview, 2024. Accessed: 2024-12-25. 3, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 474, + 296, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 474, + 296, + 507 + ], + "spans": [ + { + "bbox": [ + 56, + 474, + 296, + 507 + ], + "type": "text", + "content": "[56] Qwen Team. Qwq: Reflect deeply on the boundaries of the unknown. https://qwenlm.github.io/blog/qwq-32b-preview, 2024. Accessed: 2024-11-28. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 509, + 294, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 509, + 294, + 542 + ], + "spans": [ + { + "bbox": [ + 56, + 509, + 294, + 542 + ], + "type": "text", + "content": "[57] Qwen Team. qwen-omni. https://help.aliyun.com/zh/model-studio/user-guide/qwen-omni, 2025. Accessed: 2025-03-26. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 544, + 296, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 296, + 599 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 296, + 599 + ], + "type": "text", + "content": "[58] Prakhar Verma, Sukruta Prakash Midigeshi, Gaurav Sinha, Arno Solin, Nagarajan Natarajan, and Amit Sharma. Plan*RAG: Efficient test-time planning for retrieval augmented generation. In Workshop on Reasoning and Planning for Large Language Models, 2025. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 601, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 296, + 667 + ], + "type": "text", + "content": "[59] Xiaoxuan Wang, Ziniu Hu, Pan Lu, Yanqiao Zhu, Jieyu Zhang, Satyen Subramaniam, Arjun R Loomba, Shichang Zhang, Yizhou Sun, and Wei Wang. Scibench: Evaluating college-level scientific problem-solving abilities of large language models. In *Forty-first International Conference on Machine Learning*, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 669, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 296, + 713 + ], + "type": "text", + "content": "[60] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian richter, Fei Xia, Ed Chi, Quoc V Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models. In Advances in Neural Information Pro" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "cessing Systems, pages 24824-24837. Curran Associates, Inc., 2022. 3, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "type": "text", + "content": "[61] xAI. Models. https://docs.x.ai/docs/models? cluster=us-east-1#model-constraints, 2024. Accessed: 2024-12-12. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 316, + 129, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 205 + ], + "type": "text", + "content": "[62] Fengli Xu, Qianyue Hao, Zefang Zong, Jingwei Wang, Yunke Zhang, Jingyi Wang, Xiaochong Lan, Jiahui Gong, Tianjian Ouyang, Fanjin Meng, Chenyang Shao, Yuwei Yan, Qinglong Yang, Yiwen Song, Sijian Ren, Xinyuan Hu, Yu Li, Jie Feng, Chen Gao, and Yong Li. Towards large reasoning models: A survey of reinforced reasoning with large language models. arXiv preprint, 2025. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 206, + 553, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 250 + ], + "type": "text", + "content": "[63] Siqiao Xue, Tingting Chen, Fan Zhou, Qingyang Dai, Zhixuan Chu, and Hongyuan Mei. Famma: A benchmark for financial domain multilingual multimodal question answering. arXiv preprint, 2024. 2, 4, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 251, + 553, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 305 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 305 + ], + "type": "text", + "content": "[64] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. MM-vet: Evaluating large multimodal models for integrated capabilities. In *Forty-first International Conference on Machine Learning*, 2024. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 305, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 553, + 415 + ], + "type": "text", + "content": "[65] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Ren-liang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9556-9567, 2024. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 415, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 514 + ], + "type": "text", + "content": "[66] Xiang Yue, Tianyu Zheng, Yuansheng Ni, Yubo Wang, Kai Zhang, Shengbang Tong, Yuxuan Sun, Botao Yu, Ge Zhang, Huan Sun, Yu Su, Wenhu Chen, and Graham Neubig. MMMU-pro: A more robust multi-discipline multimodal understanding benchmark. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15134-15186, Vienna, Austria, 2025. Association for Computational Linguistics. 3, 4, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 514, + 553, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 514, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 316, + 514, + 553, + 591 + ], + "type": "text", + "content": "[67] Yilun Zhao, Hongjun Liu, Yitao Long, Rui Zhang, Chen Zhao, and Arman Cohan. Financemath: Knowledge-intensive math reasoning in finance domains. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12841-12858, Bangkok, Thailand, 2024. Association for Computational Linguistics. 2, 3, 4, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 591, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 591, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 591, + 553, + 679 + ], + "type": "text", + "content": "[68] Yilun Zhao, Yitao Long, Hongjun Liu, Ryo Kamoi, Linyong Nan, Lyuhao Chen, Yixin Liu, Xiangru Tang, Rui Zhang, and Arman Cohan. DocMath-eval: Evaluating math reasoning capabilities of LLMs in understanding long and specialized documents. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 16103-16120, Bangkok, Thailand, 2024. Association for Computational Linguistics. 4, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[69] Fengbin Zhu, Wenqiang Lei, Youcheng Huang, Chao Wang, Shuo Zhang, Jiancheng Lv, Fuli Feng, and Tat-Seng Chua. TAT-QA: A question answering benchmark on a hybrid of" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3256" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 75, + 72, + 298, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 298, + 140 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 298, + 140 + ], + "type": "text", + "content": "tabular and textual content in finance. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 3277-3287, Online, 2021. Association for Computational Linguistics. 2, 3, 5" + } + ] + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 758 + ], + "type": "text", + "content": "3257" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file